id
string | image
string | input
string | system
string | output
string | boxes
sequence | box_texts
sequence | detail
dict |
|---|---|---|---|---|---|---|---|
1811.01054-zh-3 | 1811.01054-zh-3.jpg | 影响稍后介绍的边界中的常量项。例如，假设$\overline{\mathrm{E}}$2 对任何 $q(i)\leq\infty$ 和$Q_\sigma(i)=1$ 的 ReLU 成立，对任何$q(i)>0$和$Q_\sigma(i)=1/(2+2e^{q(i)})$的 sigmoid 成立。如[2]所示，GAN 的实际训练是在参数位于紧凑空间的神经网络上进行的。因此，我们考虑以下两个紧凑参数集如[5,8,22,24],
$$\mathcal{W}_{1,\infty}:=\prod_{i=1}^{d-1}\left\{\boldsymbol{W}_{i}\in\mathbb{R}^{n_{i}\times n_{i+1}}:\|\boldsymbol{W}_{i}\|_{1,\infty}\leq M_{1,\infty}(i)\right\}\times\left\{\boldsymbol{w}_{d}\in\mathbb{R}^{n_{d}}:\|\boldsymbol{w}_{d}\|_{1}\leq M_{1,\infty}(d)\right\}\\\mathcal{W}_{F}:=\prod_{i=1}^{d-1}\left\{\boldsymbol{W}_{i}\in\mathbb{R}^{n_{i}\times n_{i+1}}:\|\boldsymbol{W}_{i}\|_{F}\leq M_{F}(i)\right\}\times\left\{\boldsymbol{w}_{d}\in\mathbb{R}^{n_{d}}:\|\boldsymbol{w}_{d}\|\leq M_{F}(d)\right\}.$$
## 2.2 Minimax Estimation Problem
在本文中，我们研究了定义如下的极小极大估计问题。假设$\mathcal{P}$是感兴趣的 Borel 概率度量的子集。令$\hat{d}(n,m)$表示通过使用分别生成 i 的样本$\{x_i\}_{i=1}^n$和$\{y_j\}_{j=1}^m$构造的神经网络距离$d_{\mathcal{F}_{nn}}(\mu,\nu)$ 的任何估计器。i。d。由$\mu,\nu\in\mathcal{P}$。我们的目标是首先找到估计误差的下界$\dot{C_{l}} ( \dot{\mathcal{P} } , n, m)$ , 使得
$$\inf\limits_{\hat{d}(n,m)}\sup\limits_{\mu,\nu\in\mathcal{P}}\mathbb{P}\left\{|d_{\mathcal{F}_{nn}}(\mu,\nu)-\hat{d}(n,m)|\geq C_{l}(\mathcal{P},n,m)\right\}>0,$$
(6)
,其中$\mathbb{P}$是相对于随机样本$\{x_i\}_{i=1}^n$和$\{\boldsymbol{y}_i\}_{i=1}^m$的概率度量。然后，我们将重点放在经验估$计器d_{\mathcal{F}_{nn}}(\hat{\mu}_{n},\hat{\nu}_{m})$上$,\tilde\text{并有兴趣在估计误差上找到一个上限}C_{u}(\mathcal{P},n,m)$ ,使得对于任何任意小的$\delta > 0$ ,
(7)
$$\sup_{\mu,\nu\in\mathcal{P}}\mathbb{P}\left\{\left|d_{\mathcal{F}_{nn}}(\mu,\nu)-d_{\mathcal{F}_{nn}}(\hat{\mu}_{n},\hat{\nu}_{m})\right|\leq C_{u}(\mathcal{P},n,m)\right\}>1-\delta.$$
显然，如果(7)的左侧定义为与 MAXTHX\_1 相同的极小极大意义
可以看出，极小极大估计问题是相对于$\mu$和$\nu$所属的分布的集合$\mathcal{P}$定义的。在本文中，我们考虑$\mathbb{R}^h$上所有亚高斯分布的集合。我们进一步将集合分为两个子集，分别进行分析，对于这两个子集，技术工具有很大的不同。第一组$\mathcal{P}_\mathrm{uB}$包含所有具有无界支持的亚高斯分布， 以及有界均值和方差。具体来说，我们假设存在$\tau>0$和$\Gamma_\mathrm{uB}>0$,因此对于任何概率度量$\mu \in \mathcal{P} _\text{uB 和 任 何 向 量 }a\in \mathbb{R} ^h$ ,
$\mathbb{E}_{\boldsymbol{x}\sim\mu}e^{\boldsymbol{a}^T(\boldsymbol{x}-\mathbb{E}(\boldsymbol{x}))}\leq e^{\|\boldsymbol{a}\|^2\tau^2/2}$with$0< \tau , \| \mathbb{E} ( \boldsymbol{x}) \| \leq \Gamma _\mathrm{uB}.$
(8)
分布的第二类$\mathcal{P}_\mathrm{B}$包含具有有限支持$\mathcal{X} = \{ x: \| x\| \leq \Gamma _\mathrm{B} \} \subset \mathbb{R} ^h$的所有亚高斯分布 (请注意， 该集合实际上包括具有有限支持的所有分布)。这两个互斥的类涵盖了实践中的大多数概率分布
## 3 $\textbf{Main Results}$
## 3.1 Minimax Lower Bound
我们首先为具有无界支持的亚高斯分布类$\mathcal{P}_\mathrm{uB}$开发以下 minimax 下界。
Theorem 1 (unbounded-support sub-Gaussian class $\mathcal{P} _{\mathbf{uB}}) . \textit{ Let F}_{nn}\textit{be the set of neural networks}$ $definedby( 4) . Forthe$ parameter set $W_{F}$ $in( 5) , if$ $\sqrt {m^{- 1}+ n^{- 1}}< \sqrt {3}q( 1) / ( 2M_{F}( 1) \Gamma _{\mathrm{uB}})$ $, then$
$$\inf\limits_{\hat{d}(n,m)}\sup\limits_{\mu,\nu\in\mathcal{P}_{uB}}\mathbb{P}\left\{\left|d_{\mathcal{F}_{nn}}(\mu,\nu)-\hat{d}(n,m)\right|\geq C(\mathcal{P}_{uB})\max\left(n^{-1/2},\:m^{-1/2}\right)\right\}\geq\frac{1}{4},$$
(9)
where
$$C(\mathcal{P}_{\mathrm{uB}})=\frac{\sqrt{3}}{6}M_{F}(1)M_{F}(d)\Gamma_{\mathrm{uB}}\left(1-\Phi\left(\frac{q(1)}{2M_{F}(1)\Gamma_{\mathrm{uB}}}\right)\right)\prod_{i=2}^{d-1}\Omega(i)\prod_{i=1}^{d-1}Q_{\sigma}(i),$$
(10)
$and\Phi(\cdot)isthecumulativedistributionfunction(CDF)ofthestandardGaussiandistributionand$
the constants $\Omega ( i) , i= 2, 3, . . . , d- 1\textit{are given by the following recursion}$
$\Omega(2)=\operatorname*{min}\left\{M_{F}(2),\:q(2)/\sigma_{1}(q(1))\right\}$,
$\Omega ( i) = \operatorname* { min} \left \{ M_{F}( i) , \:q( i) / \sigma _{i- 1}( \Omega ( i- 1) \cdots \Omega ( 2) \sigma _{1}( q( 1) ) ) \right \}$ for $i= 3, 4, . . . , d- 1.$ (11)
The same result holds for the parameter set Wi, by replacing MF(i) in (10) with M $M_{1, \infty }( i)$ .
4 | [
[
"0.1748",
"0.09216",
"0.8228",
"0.09155",
"0.8228",
"0.1061",
"0.1748",
"0.1067"
],
[
"0.1772",
"0.108",
"0.8228",
"0.108",
"0.8228",
"0.1225",
"0.1772",
"0.1225"
],
[
"0.1765",
"0.1225",
"0.822",
"0.1225",
"0.822",
"0.1364",
"0.1765",
"0.1364"
],
[
"0.1765",
"0.1364",
"0.2957",
"0.1364",
"0.2957",
"0.1497",
"0.1765",
"0.1497"
],
[
"0.259",
"0.1559",
"0.2695",
"0.1559",
"0.2695",
"0.1604",
"0.259",
"0.1604"
],
[
"0.1847",
"0.1686",
"0.2524",
"0.1686",
"0.2524",
"0.1799",
"0.1847",
"0.1799"
],
[
"0.2524",
"0.1642",
"0.8164",
"0.1642",
"0.8164",
"0.1818",
"0.2524",
"0.1818"
],
[
"0.259",
"0.1818",
"0.2712",
"0.1849",
"0.2695",
"0.1925",
"0.2566",
"0.1895"
],
[
"0.2695",
"0.1863",
"0.2737",
"0.1863",
"0.2737",
"0.1913",
"0.2695",
"0.1913"
],
[
"0.255",
"0.202",
"0.2737",
"0.1989",
"0.2769",
"0.2058",
"0.2573",
"0.209"
],
[
"0.1936",
"0.209",
"0.2566",
"0.2108",
"0.2566",
"0.226",
"0.1936",
"0.2235"
],
[
"0.25",
"0.2084",
"0.8267",
"0.207",
"0.8267",
"0.2279",
"0.25",
"0.2291"
],
[
"0.2532",
"0.2247",
"0.2778",
"0.2267",
"0.2769",
"0.2405",
"0.2507",
"0.238"
],
[
"0.1748",
"0.2524",
"0.4224",
"0.2524",
"0.4224",
"0.2664",
"0.1748",
"0.2664"
],
[
"0.1748",
"0.2778",
"0.822",
"0.2778",
"0.822",
"0.2917",
"0.1748",
"0.2917"
],
[
"0.1765",
"0.2942",
"0.822",
"0.2942",
"0.822",
"0.3093",
"0.1765",
"0.3093"
],
[
"0.1765",
"0.31",
"0.822",
"0.31",
"0.822",
"0.3245",
"0.1765",
"0.3245"
],
[
"0.1748",
"0.3245",
"0.3105",
"0.3245",
"0.3105",
"0.3384",
"0.1748",
"0.3384"
],
[
"0.3015",
"0.3452",
"0.7036",
"0.3452",
"0.7036",
"0.363",
"0.3015",
"0.363"
],
[
"0.8037",
"0.348",
"0.824",
"0.348",
"0.824",
"0.361",
"0.8037",
"0.361"
],
[
"0.2942",
"0.3599",
"0.3743",
"0.3599",
"0.3743",
"0.3706",
"0.2942",
"0.3706"
],
[
"0.1912",
"0.3757",
"0.822",
"0.3757",
"0.822",
"0.3901",
"0.1912",
"0.3901"
],
[
"0.1765",
"0.3894",
"0.8228",
"0.3894",
"0.8228",
"0.404",
"0.1765",
"0.404"
],
[
"0.1772",
"0.4048",
"0.2778",
"0.4048",
"0.2778",
"0.4167",
"0.1772",
"0.4167"
],
[
"0.2893",
"0.4216",
"0.716",
"0.4216",
"0.716",
"0.4363",
"0.2893",
"0.4363"
],
[
"0.8037",
"0.4216",
"0.826",
"0.4216",
"0.826",
"0.435",
"0.8037",
"0.435"
],
[
"0.2852",
"0.4363",
"0.3145",
"0.4363",
"0.3145",
"0.4443",
"0.2852",
"0.4443"
],
[
"0.1765",
"0.4495",
"0.6396",
"0.4495",
"0.6396",
"0.4634",
"0.1765",
"0.4634"
],
[
"0.179",
"0.4702",
"0.822",
"0.4702",
"0.822",
"0.4841",
"0.179",
"0.4841"
],
[
"0.1765",
"0.483",
"0.8228",
"0.4841",
"0.8228",
"0.4988",
"0.1765",
"0.4976"
],
[
"0.1772",
"0.4988",
"0.822",
"0.4988",
"0.822",
"0.512",
"0.1772",
"0.512"
],
[
"0.1748",
"0.511",
"0.8228",
"0.512",
"0.8228",
"0.5264",
"0.1748",
"0.526"
],
[
"0.1732",
"0.526",
"0.3823",
"0.5244",
"0.3823",
"0.539",
"0.1732",
"0.5405"
],
[
"0.2957",
"0.5405",
"0.6978",
"0.5415",
"0.6978",
"0.563",
"0.2957",
"0.5625"
],
[
"0.8037",
"0.547",
"0.826",
"0.547",
"0.826",
"0.56",
"0.8037",
"0.56"
],
[
"0.1772",
"0.5664",
"0.822",
"0.5664",
"0.822",
"0.58",
"0.1772",
"0.58"
],
[
"0.1772",
"0.5806",
"0.8228",
"0.5806",
"0.8228",
"0.5947",
"0.1772",
"0.5947"
],
[
"0.1732",
"0.5947",
"0.21",
"0.5947",
"0.21",
"0.61",
"0.1732",
"0.61"
],
[
"0.1724",
"0.6255",
"0.3218",
"0.6274",
"0.3218",
"0.6426",
"0.1724",
"0.64"
],
[
"0.1732",
"0.656",
"0.2157",
"0.656",
"0.2157",
"0.6704",
"0.1732",
"0.6704"
],
[
"0.2133",
"0.657",
"0.38",
"0.657",
"0.38",
"0.6704",
"0.2133",
"0.6704"
],
[
"0.1765",
"0.683",
"0.6733",
"0.683",
"0.6733",
"0.6963",
"0.1765",
"0.6963"
],
[
"0.1765",
"0.6997",
"0.822",
"0.6997",
"0.822",
"0.714",
"0.1765",
"0.714"
],
[
"0.1765",
"0.7153",
"0.8203",
"0.7153",
"0.8203",
"0.729",
"0.1765",
"0.729"
],
[
"0.2296",
"0.7266",
"0.7793",
"0.733",
"0.7793",
"0.7646",
"0.2296",
"0.7583"
],
[
"0.777",
"0.7393",
"0.822",
"0.7393",
"0.822",
"0.754",
"0.777",
"0.754"
],
[
"0.2296",
"0.75",
"0.33",
"0.752",
"0.33",
"0.765",
"0.2296",
"0.763"
],
[
"0.1748",
"0.768",
"0.2173",
"0.768",
"0.2173",
"0.78",
"0.1748",
"0.78"
],
[
"0.3186",
"0.782",
"0.619",
"0.782",
"0.619",
"0.8257",
"0.3186",
"0.8257"
],
[
"0.6953",
"0.786",
"0.71",
"0.786",
"0.71",
"0.7905",
"0.6953",
"0.7905"
],
[
"0.2998",
"0.7954",
"0.308",
"0.7905",
"0.312",
"0.7954",
"0.3057",
"0.8"
],
[
"0.3096",
"0.787",
"0.339",
"0.787",
"0.339",
"0.8",
"0.3096",
"0.8"
],
[
"0.2378",
"0.797",
"0.3057",
"0.797",
"0.3057",
"0.808",
"0.2378",
"0.808"
],
[
"0.626",
"0.7954",
"0.767",
"0.7954",
"0.767",
"0.81",
"0.626",
"0.81"
],
[
"0.763",
"0.799",
"0.789",
"0.799",
"0.789",
"0.8066",
"0.763",
"0.8066"
],
[
"0.79",
"0.797",
"0.8228",
"0.797",
"0.8228",
"0.8096",
"0.79",
"0.8096"
],
[
"0.3105",
"0.8047",
"0.3325",
"0.8047",
"0.3325",
"0.813",
"0.3105",
"0.813"
],
[
"0.6235",
"0.803",
"0.63",
"0.803",
"0.63",
"0.806",
"0.6235",
"0.806"
],
[
"0.6396",
"0.8164",
"0.653",
"0.8164",
"0.653",
"0.821",
"0.6396",
"0.821"
],
[
"0.6914",
"0.813",
"0.7114",
"0.813",
"0.7114",
"0.8237",
"0.6914",
"0.8237"
],
[
"0.1765",
"0.827",
"0.824",
"0.827",
"0.824",
"0.8403",
"0.1765",
"0.8403"
],
[
"0.1748",
"0.8413",
"0.6567",
"0.8413",
"0.6567",
"0.855",
"0.1748",
"0.855"
],
[
"0.2018",
"0.86",
"0.468",
"0.86",
"0.468",
"0.8735",
"0.2018",
"0.8735"
],
[
"0.2059",
"0.88",
"0.824",
"0.88",
"0.824",
"0.893",
"0.2059",
"0.893"
],
[
"0.1732",
"0.8984",
"0.7876",
"0.8984",
"0.7876",
"0.912",
"0.1732",
"0.912"
],
[
"0.4902",
"0.936",
"0.5073",
"0.936",
"0.5073",
"0.9497",
"0.4902",
"0.9497"
]
] | [
"影响稍后介绍的边界中的常量项。例如，假设E2对任何g(i）≤8o和Q。（）=1的ReLU",
"成立，对任何q(i)>0和Q。(i)=1/(2+2e9(i)）的sigmoid成立。如[2]所示，GAN的实",
"际训练是在参数位于紧凑空间的神经网络上进行的。因此，我们考虑以下两个紧参数集",
"如[5,8,22,24],",
"",
"Wi.oo : =",
"II (W, e IRni xni+1 : IlWill, ≤ Mi,(i)) × [wd E Rnd : llwall ≤ Mi,(d) ,",
"",
"",
"d-",
"WF:=",
"I (W; E Rn+xni+ : IWill ≤ Mr(i) ×(wa E Rnd : llwall ≤Mr(d). (5)",
"=1",
"2.2MinimaxEstimationProblem",
"在本文中，我们研究了定义如下的极小极大估计问题。假设P是感兴趣的Borel概率度量",
"的子集。令d(n,m）表示通过使用分别生成i的样本(ai}n=i和(yi}=i构造的神经网络距",
"离dFn(μ,v)的任何估计器。i。d。由μ,VEP。我们的目标是首先找到估计误差的下界",
"Ci(P,n,m)，使得",
"infsup PIdFnn (μ,u) -d(n,m)I ≥Ci(P,n, m)/>0,",
"(6)",
"d(n,m)μ,VEP",
"其中P是相对于随机样本[α,}\"-1和{yi}\"1的概率度量。然后，我们将重点放在经验估",
"计器dFnn（in，Dm）上，并有兴趣在估计误差上找到一个上限Cu(P,n,m），使得对于任何任",
"意小的>0，",
"sup P[ldFnn(μ,v) -dFnn (in,Vm)/ ≤Cu(P,n,m)) > 1 -d.",
"(7)",
"H,VE",
"显然，如果（7）的左侧定义为与MAXTHX_1相同的极小极大意义",
"可以看出，极小极大估计问题是相对于μ和V所属的分布的集合P定义的。在本文中，我",
"们考虑R上所有亚高斯分布的集合。我们进一步将集合分为两个子集，分别进行分析，对",
"于这两个子集，技术工具有很大的不同。第一组PuB包含所有具有无界支持的亚高斯分布",
"以及有界均值和方差。具体来说，我们假设存在>0和IuB>0，因此对于任何概率度量",
"μEPuB和任何向量aERh,",
"Ea~μ ea (a-E(am) ≤ellP- /2 with 0 < T, IE(α)l ≤ TuB.",
"(8)",
"分布的第二类PB包含具有有限支持=(a：al≤I}CRh的所有亚高斯分布（请注意",
"该集合实际上包括具有有限支持的所有分布）。这两个互斥的类涵盖了实践中的大多数概率",
"分布",
"3MainResults",
"3.1",
"Minimax Lower Bound",
"我们首先为具有无界支持的亚高斯分布类PuB开发以下minimax下界。",
"Theorem 1 (unbounded-support sub-Gaussian class PuB ). Let Fnn be the set of neural networks",
"defined by (4). For the parameter set WF in (5), if Vm-1 + n-1 < V3q(1)/(2MF(1)FuB), ther",
"imf sup, [/d. (μ,) -d(nm) ≥ (Pa) max (n-1/2, m-1/) ≥ 1",
" (9)",
"i(n,m)μ,VEPuB",
"where",
"Mra)Mr( (1--(2r)",
"1-",
"",
"V3",
"C(PuB) =",
"/ 2(i) // Q.(i),",
"",
"(10)",
"6",
"",
"",
"2=1",
"and () is the cumulative distribution function (CDF) of the standard Gaussian distribution and",
"the constants 2(i),i= 2, 3,..., d -1 are given by the following recursion",
"2(2) =min [MF(2), q(2) /oi(q(1)) / ,",
"2(i) =min /MF(i), q(℃) /oi-1(2(i - 1) ..2(2)ai(q(1)))1 for i = 3, 4, .., d - 1. (11)",
"The same result holds for the parameter set Wi,oo by replacing M (i) in (1O) with Mi,oo (i)",
"4"
] | {
"arxiv_id": "1811.01054",
"language": "zh",
"page_idx": 3
} |
||
1811.01054-zh-4 | 1811.01054-zh-4.jpg | 定理 1意味着$d_\mathcal{F}_{nn}(\mu,\nu)$不能由类$\mathcal{P}_\mathrm{uB}$上的任何估计器以比$\max\left(n^{-1/2},m^{-1/2}\right)$快的速率进行估计。定理 1 的证明基于 LeCam 方法。在$\mathbb{E}[26]$ 中也使用了这种技术来推导用于估计MMD 的 minimax 下界。但是，我们的技术发展与$\overline{\mathbb{E}}[26]$ 中的技术发展有很大不同。具体来说，LeCam 方法的一个主要步骤是降低由于两个假设分布而产生的差异。在[26]中的 MMD 情况下，对于选定的分布，MMD 可以用封闭形式表示。因此，LeCam 方法的下界可以基于MMD 的这种封闭形式得出。作为比较，神经网络距离对于所选分布没有封闭形式的表达式。因此，我们的推导涉及到神经网络函数相对于两个相应分布的期望差异的下限。这种发展需要大量的机器来处理神经网络的复杂多层结构。有关更多详细信息，请参见附录 A.1
对于一般的神经网络，$C(\mathcal{P}_\mathrm{uB})$采用 (10) 中的复杂形式。接下来，我们专门研究 ReLU 网络，
以说明该常数如何取决于神经网络参数。
Corollary 1. Under the setting of Theorem $l,suppose$ each activation function is ReLU, i.e.,
$\sigma _{i}( z) = \operatorname* { max} \{ 0, z\} $, $i= 1, 2, . . . , d- 1.$ For the parameter set $W_{F}$ and all $m, n\geq 1$ $, we$ have
$$\inf\limits_{\hat{d}(n,m)}\sup\limits_{\mu,\nu\in\mathcal{P}_{\text{uB}}}\mathbb{P}\left\{\left|d_{\mathcal{F}_{nn}}(\mu,\nu)-\hat{d}(n,m)\right|\geq0.08\:\Gamma_{\text{uB}}\prod\limits_{i=1}^{d}M_{F}(i)\max\left(n^{-1/2},m^{-1/2}\right)\right\}\geq\frac{1}{4}.$$
The same result holds for the parameter set $\mathcal{W}_{1,\infty}\textit{by replacing M}_F(i)$ with $M_{1,\infty}(i)$
接下来，我们为具有有限支持的分发类$\mathcal{P}_\mathrm{B}$提供 minimax 下界。证明 (请参见附录EB)也
基于 LeCam 方法，但是构造的分布具有有界支持集，这与定理 1 不同。
Theorem 2 (bounded-support class $P_{\mathrm{B} }) . Let$ $F_{nn}$ be the set of neural networks defined by(4).
For the parameter set $W_{F}$ $, we$ have
(12)
$$\inf\limits_{\hat{d}(n,m)}\sup\limits_{\mu,\nu\in\mathcal{P}_{\text{B}}}\mathbb{P}\left\{\left|d_{\mathcal{F}_{n\:n}}(\mu,\nu)-\hat{d}(n,m)\right|\geq C(\mathcal{P}_{\text{B}})\max\left(n^{-1/2},m^{-1/2}\right)\right\}\geq\frac{1}{4},$$
where
$C( \mathcal{P} _{\mathbf{B} }) = 0. 17\left ( M_{F}( d) \sigma _{d- 1}( \cdots \sigma _{1}( M_{F}( 1) \Gamma _{\mathbf{B} }) ) - M_{F}( d) \sigma _{d- 1}( \cdots \sigma _{1}( - M_{F}( 1) \Gamma _{\mathbf{B} }) ) \right ) $, (13)
where all constants $M_{F}(i),i=1,2,...,d$ in the second term of the right side of (13) have negative signs. The same result holds for the parameter set W $_{1,\infty}\textit{by replacing M}_{F}(i)$ in (13) with $M_{1,\infty}(i)$ Corollary 2. Under the setting of Theorem 2,suppose that each activation function is ReLU. For the parameter set WF , we have,
$$\inf_{\hat{d}(n,m)}\sup_{\mu,\nu\in\mathcal{P}_{\mathrm{B}}}\mathbb{P}\left\{\left|d_{\mathcal{F}_{n\:n}}(\mu,\nu)-\hat{d}(n,m)\right|\geq0.17\:\Gamma_{\mathrm{B}}\prod_{i=1}^{d}M_{F}(i)\max\left(n^{-1/2},\:m^{-1/2}\right)\right\}\geq\frac{1}{4}.$$
The same result holds for the parameter set W$_{1,\infty}\textit{by replacing M}_{F}(i)$ with $M_{1,\infty}(i)$
3.2 Rademacher Complexity-based Upper Bound
在本小节中，我们提供了$\left|d_\mathcal{F}_{nn}(\mu,\nu)-d_\mathcal{F}_{nn}(\hat{\mu}_n,\hat{\nu}_m)\right|$的上限，该上限用作极小极大估计误差的上限。我们的主要技术发展在于推导无界支持子高斯类$P_\mathrm{uB}$的界，这需要许多新的技术发展。我们在 3.4 节中讨论它的证明。
Theorem 3 (unbounded-support sub-Gaussian class $P_{\mathrm{uB}}) . \textit{ Let F}_{nn}\textit{be the set of neural networks}$ $definedby(4),andsupposethattwodistributions\mu,\nu\in\mathcal{P}_{\mathrm{uB}}and\hat{\mu}_{n},\hat{\nu}_{m}$ are their empirical measures. For a constant $\delta>0$ satisfying $\sqrt6h\min\{n,m\}\sqrt{m^{-1}+n^{-1}}\geq4\sqrt{\log(1/\delta)},we$ have $(\mathbf{I})$ Ifthe parameter set is $W_F$ and each activation function satisfies $\sigma _i( \alpha x) = \alpha \sigma _i( x) \textit{for all }\alpha > 0$ $(e.g.,ReLUorleakyReLU),thenwithprobabilityatleast1-\delta overtherandomnessof\hat{\mu}_{n}and\hat{\nu}_{m}$,
$$\begin{aligned}|d_{\mathcal{F}_{nn}}(\mu,\nu)&-d_{\mathcal{F}_{nn}}(\hat{\mu}_{n},\hat{\nu}_{m})|\\&\leq2\Gamma_{\mathrm{uB}}\prod_{i=1}^{d}M_{F}(i)\prod_{i=1}^{d-1}L_{i}\left(\sqrt{6d\log2+5h/4}+\sqrt{2h\log(1/\delta)}\right)\left(n^{-1/2}+m^{-1/2}\right).\end{aligned}$$
5 | [
[
"0.1748",
"0.0903",
"0.824",
"0.089",
"0.824",
"0.1067",
"0.1748",
"0.108"
],
[
"0.1765",
"0.108",
"0.8203",
"0.108",
"0.8203",
"0.1225",
"0.1765",
"0.1225"
],
[
"0.1765",
"0.1225",
"0.8228",
"0.1225",
"0.8228",
"0.1364",
"0.1765",
"0.1364"
],
[
"0.179",
"0.1383",
"0.822",
"0.1383",
"0.822",
"0.1483",
"0.179",
"0.1483"
],
[
"0.1772",
"0.1509",
"0.8203",
"0.1509",
"0.8203",
"0.1642",
"0.1772",
"0.1642"
],
[
"0.1765",
"0.1642",
"0.822",
"0.1642",
"0.822",
"0.178",
"0.1765",
"0.178"
],
[
"0.1765",
"0.178",
"0.8228",
"0.178",
"0.8228",
"0.1913",
"0.1765",
"0.1913"
],
[
"0.1765",
"0.1901",
"0.8013",
"0.1901",
"0.8013",
"0.2039",
"0.1765",
"0.2039"
],
[
"0.1772",
"0.2122",
"0.822",
"0.2122",
"0.822",
"0.226",
"0.1772",
"0.226"
],
[
"0.1772",
"0.2267",
"0.4631",
"0.2267",
"0.4631",
"0.2405",
"0.1772",
"0.2405"
],
[
"0.1765",
"0.2456",
"0.822",
"0.2456",
"0.822",
"0.26",
"0.1765",
"0.26"
],
[
"0.1748",
"0.26",
"0.796",
"0.26",
"0.796",
"0.2734",
"0.1748",
"0.2734"
],
[
"0.1978",
"0.2942",
"0.2957",
"0.2986",
"0.2942",
"0.3125",
"0.1969",
"0.308"
],
[
"0.3137",
"0.2954",
"0.653",
"0.2942",
"0.653",
"0.3125",
"0.3137",
"0.3132"
],
[
"0.6567",
"0.2935",
"0.7607",
"0.2917",
"0.7607",
"0.3093",
"0.6567",
"0.3113"
],
[
"0.1887",
"0.31",
"0.2876",
"0.31",
"0.2876",
"0.3213",
"0.1887",
"0.3213"
],
[
"0.1732",
"0.332",
"0.7363",
"0.334",
"0.7363",
"0.3484",
"0.1732",
"0.3467"
],
[
"0.1765",
"0.3623",
"0.8228",
"0.3623",
"0.8228",
"0.3757",
"0.1765",
"0.3757"
],
[
"0.1772",
"0.3757",
"0.681",
"0.3757",
"0.681",
"0.3901",
"0.1772",
"0.3901"
],
[
"0.1772",
"0.3958",
"0.822",
"0.3958",
"0.822",
"0.4092",
"0.1772",
"0.4092"
],
[
"0.1724",
"0.4077",
"0.4126",
"0.4092",
"0.4126",
"0.4236",
"0.1724",
"0.4224"
],
[
"0.2443",
"0.44",
"0.3447",
"0.44",
"0.3447",
"0.452",
"0.2443",
"0.452"
],
[
"0.3472",
"0.4382",
"0.7466",
"0.4363",
"0.7466",
"0.4539",
"0.3472",
"0.4558"
],
[
"0.7173",
"0.443",
"0.7646",
"0.443",
"0.7646",
"0.4583",
"0.7173",
"0.4583"
],
[
"0.767",
"0.44",
"0.822",
"0.44",
"0.822",
"0.455",
"0.767",
"0.455"
],
[
"0.2345",
"0.4526",
"0.3162",
"0.4526",
"0.3162",
"0.464",
"0.2345",
"0.464"
],
[
"0.1748",
"0.4734",
"0.2173",
"0.4734",
"0.2173",
"0.488",
"0.1748",
"0.488"
],
[
"0.2009",
"0.4976",
"0.824",
"0.4976",
"0.824",
"0.512",
"0.2009",
"0.512"
],
[
"0.1765",
"0.5195",
"0.822",
"0.5215",
"0.822",
"0.536",
"0.1765",
"0.534"
],
[
"0.1748",
"0.535",
"0.824",
"0.535",
"0.824",
"0.55",
"0.1748",
"0.55"
],
[
"0.1732",
"0.5674",
"0.8228",
"0.5684",
"0.8228",
"0.5835",
"0.1732",
"0.582"
],
[
"0.1748",
"0.584",
"0.3938",
"0.584",
"0.3938",
"0.5977",
"0.1748",
"0.5977"
],
[
"0.2018",
"0.622",
"0.2957",
"0.622",
"0.2957",
"0.632",
"0.2018",
"0.632"
],
[
"0.3015",
"0.616",
"0.7344",
"0.6123",
"0.7344",
"0.636",
"0.3015",
"0.64"
],
[
"0.1912",
"0.6333",
"0.2778",
"0.6333",
"0.2778",
"0.6445",
"0.1912",
"0.6445"
],
[
"0.1732",
"0.656",
"0.738",
"0.657",
"0.738",
"0.672",
"0.1732",
"0.6704"
],
[
"0.1748",
"0.6895",
"0.5376",
"0.6895",
"0.5376",
"0.7026",
"0.1748",
"0.7026"
],
[
"0.1765",
"0.7153",
"0.822",
"0.7153",
"0.822",
"0.73",
"0.1765",
"0.73"
],
[
"0.1748",
"0.728",
"0.822",
"0.729",
"0.822",
"0.742",
"0.1748",
"0.741"
],
[
"0.1765",
"0.742",
"0.46",
"0.742",
"0.46",
"0.757",
"0.1765",
"0.757"
],
[
"0.1772",
"0.762",
"0.822",
"0.762",
"0.822",
"0.776",
"0.1772",
"0.776"
],
[
"0.1748",
"0.7744",
"0.8228",
"0.776",
"0.8228",
"0.7905",
"0.1748",
"0.789"
],
[
"0.1748",
"0.7915",
"0.822",
"0.7915",
"0.822",
"0.806",
"0.1748",
"0.806"
],
[
"0.1765",
"0.8125",
"0.824",
"0.8125",
"0.824",
"0.827",
"0.1765",
"0.827"
],
[
"0.1732",
"0.8257",
"0.8203",
"0.8257",
"0.8203",
"0.8403",
"0.1732",
"0.8403"
],
[
"0.1912",
"0.8486",
"0.3857",
"0.8496",
"0.3857",
"0.8643",
"0.1912",
"0.863"
],
[
"0.2532",
"0.879",
"0.8037",
"0.877",
"0.8037",
"0.898",
"0.2532",
"0.8994"
],
[
"0.3096",
"0.8735",
"0.3162",
"0.8726",
"0.3162",
"0.8755",
"0.3105",
"0.877"
],
[
"0.304",
"0.8994",
"0.326",
"0.8994",
"0.326",
"0.9077",
"0.304",
"0.9077"
],
[
"0.3774",
"0.9014",
"0.3896",
"0.9014",
"0.3896",
"0.906",
"0.3774",
"0.906"
],
[
"0.4944",
"0.9375",
"0.5063",
"0.9375",
"0.5063",
"0.9487",
"0.4944",
"0.9487"
]
] | [
"定理1意味着dFn（μ,）不能由类PuB上的任何估计器以比max（n-1/2,m-1/2）快的速率",
"进行估计。定理1的证明基于LeCam方法。在E[26]中也使用了这种技术来推导用于估计",
"MMD的minimax下界。但是，我们的技术发展与E[26]中的技术发展有很大不同。具体来",
"说，LeCam方法的一个主要步骤是降低由于两个假设分布而产生的差异。在[261中的MMD",
"情况下，对于选定的分布，MMD可以用封闭形式表示。因此，LeCam方法的下界可以基于",
"MMD的这种封闭形式得出。作为比较，神经网络距离对于所选分布没有封闭形式的表达",
"式。因此，我们的推导涉及到神经网络函数相对于两个相应分布的期望差异的下限。这种发",
"展需要大量的机器来处理神经网络的复杂多层结构。有关更多详细信息，请参见附录A.1",
"对于一般的神经网络，C(PuB）采用（10）中的复杂形式。接下来，我们专门研究ReLU网络",
"以说明该常数如何取决于神经网络参数。",
"Corollary 1. Under the setting of Theorem I , suppose each activation function is ReLU, i.e.,",
"o;(z) = max[0, z), i = 1, 2, ..., d - 1. For the parameter set WF and all m, n ≥ 1, we have",
"inf.sup P",
"dfnn (μ, v) - d(n, m)≥0.08TuB II MF(i)max (",
"n-1/2, m-1/2)",
"d(n,m)μ,VePuB",
"The same result holds for the parameter set Wi,oo by replacing Mr (i) with Mi,co(i)",
"接下来，我们为具有有限支持的分发类PB提供minimax下界。证明（请参见附录EB）也",
"基于LeCam方法，但是构造的分布具有有界支持集，这与定理1不同。",
"Theorem 2 (bounded-support class P ). Let Fnn be the set of neural networks defined by (4 ).",
"For the parameter set We, we have",
"infsup P",
"dfFm(μ,v) - d(n,m)≥C(Pg)max (n-1/2, m-1/2) / ≥",
"1≥4",
" (12)",
"d(n,m)μ,VEP",
"where",
"C(PB) = 0.17(MF(d)od-1(..:αi(MF(1)FB)) -MF(d)αd-1(..0i(-MF(1)B))),(13)",
"where all constants MF(i),i = 1, 2, .., d in the second term of the right side of (13) have negative",
"signs. The same result holds for the parameter set Wi,oo by replacing Mr(i) in (13) with Mi,oo(i)",
"Corollary2.Under the setting of Theorem 2,suppose that each activation function is ReLU.For",
"the parameter set Wr, we have,",
"infsupP",
"d. (μu,v) -a(n, m)| ≥0.17IsIIMr(i)max (n-1/2, m-1",
"d(n,m) μ,VEPB",
"The same result holds for the parameter set Wi,oo by replacing Mr(i) with Mi,oo(i)",
"3.2Rademacher Complexity-based UpperBound",
"在本小节中，我们提供了ldFnn（μu,v）一dFnn（in，m）的上限，该上限用作极小极大估计误",
"差的上限。我们的主要技术发展在于推导无界支持子高斯类PuB的界，这需要许多新的技",
"术发展。我们在3.4节中讨论它的证明。",
"Theorem 3 (unbounded-support sub-Gaussian class Pu ). Let Fnn be the set of neural networks",
"defined by ( 4 ), and suppose that two distributions μ, v E PuB and jin, Dm are their empirical",
"measures. For a constant d > 0 satisfying V6h min(n, m) Vm-1 + n-1 ≥ 4/log(1/3) , we have",
"() If the parameter set is W and each activation function satisfies o;(αr) = ao;(c) for all α > 0",
"(e.g., ReLU or leaky ReLU), then with probability at least 1 - d over the randomness of jpn and im :",
"[dFun(μ,v)-dFnn (in,Dm)l",
"≤2FuBIIMr(i) IIL; (V6dlog2+ 5h/4+V2hlog(1/8) (n-1/2 +m-1/2)",
"F",
"=1",
"",
"5"
] | {
"arxiv_id": "1811.01054",
"language": "zh",
"page_idx": 4
} |
||
2007.15619-zh-0 | 2007.15619-zh-0.jpg | 基于人工智能的监测 &响应系统，用于东南亚医院对 COVID-19 的准备
$\begin{array}{c}{\mathrm{Tushar~Goswamy^{1*},~Naishadh~Parmar^{1},~Ayush~Gupta^{1},~Raunak~Shah^{1},~Vatsalya}}\\{\mathrm{Tandon^{1},~Varun~Goyal^{1},~Sanyog~Gupta^{1},~Karishma~Laud^{1},~Shivam~Gupta^{1},~Sudhanshu}}\end{array}$
Mishr$a^{1}$, Ashutosh Mod$i^{1}$,
$^{1}$Indian Institute of Technology (IIT) Kanpur, Kanpur, India
$\{$ tgoswamy, naishadh, ayushgup, raunaks, vatsalt, varung, sanyog, kslaud, guptash, sdhanshu,
ashutoshm $\}$ @iitk.ac.in
## Abstract
在这项研究中，我们使用用户的 Twitter 数据来研究由于 COVID-19 大流行而导致的住院人数激增。在本研究的范围内，我们将工作重点放在印度，印度尼西亚和孟加拉国，并计划将这种方法扩展到东南亚的其他地区。我们的研究旨在确定医院人满为患的事件，呼吸机等关键设备的短缺以及缺乏可用的 ICU 病房。该系统可以帮助了解这些国家的医疗机构的医疗准备水平以及大流行蔓延时医院的负担。系统管道包括在粒度级别上刮擦历史推文以获得语料库，使用自然语言处理工具处理语料库，从处理后的数据中计算信号，最后通过比较地面报告和公告来评估结果。我们已经部署了神经翻译模型来解释区域语言的使用
通过这项研究，我们对人工智能社区的主要贡献是展示了基于 NLP 的 Twitter 模型在监测 COVID-19 大流行对卫生设施造成的负担方面的应用。据我们所知，这是同类产品中的第一种也是唯一的方法，它可以根据Twitter 数据准确地检测受灾最严重地区的趋势。我们正在与世卫组织东南亚区域办事处(WHO-SEARO)的成员密切合作，研究和监测我们模型的信号，旨在帮助他们监测这些国家的情况，并确定哪些地区正面临着由于大流行而导致的资源短缺。因此，公共卫生组织可以使用我们的模型向模型确定的地区的当局建议适当的行动
This research paper proposes a COVID-19 monitoring and response system to identify the surge in the volume of patients at hospitals and shortage of critical equipment like ventilators in South-east Asian countries, to understand the burden on health facilities. This can help authorities in these regions with resource planning measures to redirect resources to the regions identified by the model. Due to the lack of publicly available data on the influx of patients in hospitals, or the shortage of equipment, ICU units or hospital beds that regions in these countries might be facing, we leverage Twitter data for gleaning this information. The approach has yielded accurate results for states in India, and we are working on validating the model for the remaining countries so that it can serve as a reliable tool for authorities to monitor the burden on hospitals.
## 1 Introduction
2 Data Extraction and Pre-processing
2.1 Natural Language Processing for Tweets
Historical Tweet Extraction
我们使用 getOldTweets3API [Mottl, Updated 2019] 从Twitter 网站抓取和提取历史推文。不幸的是，Twitter 有一些限制，由于这些限制，我们无法访问从刮擦之日起超过 7 天的所有推文。这导致了数据中的误导性峰值(图。1)。为了解决这个问题，我们使用整个峰值的变化因子来缩放推文。Data Cleaning
像 Twitter 和 Facebook 这样的社交媒体网站鼓励用户频繁表达他们的想法，观点和生活的随机细节。印度拥有全球第八大 Twitter 用户群，拥有 13。1500 万用户并在增长，其次是印度尼西亚，拥有 11。800 万用户[Statista, 2020]。这凸显了从这些国家数百万用户发布的推文中获得有用见解的潜力。推文和状态更新的范围从重大事件到无聊的评论。大多数消息包含很少的信息价值，但数百万条消息的聚合可以产生有价值的知识。Twitter 用户经常公开表达有关医院人满为患，他们或其亲属因设备短缺而面临的困难以及大流行引起的其他问题的个人经历，这可以帮助了解情况的实际情况。先前的研究使用有关症状[Paul and Dredze, 2011] 的推文研究了 Twitter 趋势与流感发生率之间的相关性。统计技术已用于使用Twitter 数据[Santillana et al., 2015] 预测流感发病率。在 2012 年 [Broniatowski et al., 2013] 流感流行期间，美国已在地方一级监测了流感发病率。同样，Signorini 等人[Signorini et al.,2011]研究了 Twitter 数据与 H1N1病例之间的相关性，以追踪感染
为了消除数据中的噪声并提取重要信息，我们对 tweet
语料库执行了以下操作：
$\bullet$ Removing Website Links: To prevent the same in-
formation from being captured twice.
$\bullet$ Removing non-ASCII characters: To eliminate noise
and focus on relevant keywords only
$\bullet$ Removing Stopwords: Removed words like `is',“an',
“the’ to focus on hospital-related words in the fre-
quency analysis
$^{*}$Contact Author | [
[
"0.1316",
"0.11426",
"0.8667",
"0.11426",
"0.8667",
"0.1288",
"0.1316",
"0.1288"
],
[
"0.1478",
"0.1433",
"0.8486",
"0.1445",
"0.8486",
"0.1622",
"0.1478",
"0.161"
],
[
"0.1177",
"0.161",
"0.8784",
"0.161",
"0.8784",
"0.1787",
"0.1177",
"0.1787"
],
[
"0.3782",
"0.1787",
"0.614",
"0.1787",
"0.614",
"0.1925",
"0.3782",
"0.1925"
],
[
"0.2402",
"0.1957",
"0.755",
"0.1976",
"0.755",
"0.2153",
"0.2402",
"0.2134"
],
[
"0.10455",
"0.2184",
"0.892",
"0.2184",
"0.892",
"0.2329",
"0.10455",
"0.2329"
],
[
"0.3987",
"0.2349",
"0.5986",
"0.2349",
"0.5986",
"0.2488",
"0.3987",
"0.2488"
],
[
"0.2484",
"0.2715",
"0.3245",
"0.2734",
"0.3245",
"0.289",
"0.2484",
"0.2874"
],
[
"0.5312",
"0.2747",
"0.911",
"0.2747",
"0.911",
"0.2886",
"0.5312",
"0.2886"
],
[
"0.5156",
"0.2886",
"0.911",
"0.2886",
"0.911",
"0.3018",
"0.5156",
"0.3018"
],
[
"0.1177",
"0.2986",
"0.452",
"0.2986",
"0.452",
"0.3132",
"0.1177",
"0.3132"
],
[
"0.5156",
"0.3037",
"0.91",
"0.3037",
"0.91",
"0.3145",
"0.5156",
"0.3145"
],
[
"0.1177",
"0.3132",
"0.452",
"0.3125",
"0.452",
"0.327",
"0.1177",
"0.3276"
],
[
"0.5156",
"0.3164",
"0.9087",
"0.3164",
"0.9087",
"0.33",
"0.5156",
"0.33"
],
[
"0.11926",
"0.3276",
"0.4534",
"0.3276",
"0.4534",
"0.3416",
"0.11926",
"0.3416"
],
[
"0.513",
"0.3289",
"0.91",
"0.33",
"0.91",
"0.3435",
"0.513",
"0.3423"
],
[
"0.1217",
"0.3416",
"0.452",
"0.3416",
"0.452",
"0.3547",
"0.1217",
"0.3547"
],
[
"0.5146",
"0.3435",
"0.911",
"0.3435",
"0.911",
"0.3567",
"0.5146",
"0.3567"
],
[
"0.1209",
"0.3567",
"0.451",
"0.3567",
"0.451",
"0.3674",
"0.1209",
"0.3674"
],
[
"0.5146",
"0.356",
"0.911",
"0.358",
"0.911",
"0.3718",
"0.5146",
"0.3694"
],
[
"0.11926",
"0.3694",
"0.4478",
"0.3694",
"0.4478",
"0.3801",
"0.11926",
"0.3801"
],
[
"0.5146",
"0.3694",
"0.911",
"0.3706",
"0.911",
"0.385",
"0.5146",
"0.3838"
],
[
"0.1217",
"0.3838",
"0.4478",
"0.3838",
"0.4478",
"0.3945",
"0.1217",
"0.3945"
],
[
"0.5146",
"0.3838",
"0.91",
"0.3838",
"0.91",
"0.3977",
"0.5146",
"0.3977"
],
[
"0.11926",
"0.3977",
"0.452",
"0.3977",
"0.452",
"0.411",
"0.11926",
"0.411"
],
[
"0.5146",
"0.3997",
"0.91",
"0.3997",
"0.91",
"0.4136",
"0.5146",
"0.4136"
],
[
"0.11926",
"0.411",
"0.452",
"0.4104",
"0.452",
"0.4236",
"0.11926",
"0.4248"
],
[
"0.5156",
"0.4136",
"0.911",
"0.4136",
"0.911",
"0.4268",
"0.5156",
"0.4268"
],
[
"0.1209",
"0.4248",
"0.4534",
"0.4248",
"0.4534",
"0.4382",
"0.1209",
"0.4382"
],
[
"0.5146",
"0.4268",
"0.6626",
"0.4268",
"0.6626",
"0.44",
"0.5146",
"0.44"
],
[
"0.11926",
"0.4395",
"0.451",
"0.4395",
"0.451",
"0.4526",
"0.11926",
"0.4526"
],
[
"0.5293",
"0.4414",
"0.911",
"0.4414",
"0.911",
"0.455",
"0.5293",
"0.455"
],
[
"0.11926",
"0.452",
"0.452",
"0.4526",
"0.452",
"0.4666",
"0.11926",
"0.4653"
],
[
"0.5146",
"0.4539",
"0.911",
"0.4526",
"0.911",
"0.4673",
"0.5146",
"0.4685"
],
[
"0.11926",
"0.4666",
"0.451",
"0.4666",
"0.451",
"0.4797",
"0.11926",
"0.4797"
],
[
"0.513",
"0.4666",
"0.911",
"0.4673",
"0.911",
"0.4817",
"0.513",
"0.481"
],
[
"0.1209",
"0.4797",
"0.452",
"0.4797",
"0.452",
"0.4937",
"0.1209",
"0.4937"
],
[
"0.5146",
"0.481",
"0.9126",
"0.4817",
"0.9126",
"0.4968",
"0.5146",
"0.4956"
],
[
"0.1209",
"0.4937",
"0.451",
"0.4937",
"0.451",
"0.507",
"0.1209",
"0.507"
],
[
"0.5146",
"0.4944",
"0.911",
"0.4968",
"0.911",
"0.5103",
"0.5146",
"0.5083"
],
[
"0.11926",
"0.5083",
"0.452",
"0.5083",
"0.452",
"0.5215",
"0.11926",
"0.5215"
],
[
"0.5156",
"0.5103",
"0.9126",
"0.5103",
"0.9126",
"0.5234",
"0.5156",
"0.5234"
],
[
"0.11926",
"0.5215",
"0.452",
"0.5215",
"0.452",
"0.535",
"0.11926",
"0.535"
],
[
"0.5156",
"0.5234",
"0.911",
"0.5234",
"0.911",
"0.537",
"0.5156",
"0.537"
],
[
"0.1365",
"0.534",
"0.452",
"0.535",
"0.452",
"0.549",
"0.1365",
"0.5474"
],
[
"0.5156",
"0.537",
"0.91",
"0.537",
"0.91",
"0.5503",
"0.5156",
"0.5503"
],
[
"0.11926",
"0.55",
"0.3203",
"0.55",
"0.3203",
"0.563",
"0.11926",
"0.563"
],
[
"0.5156",
"0.5503",
"0.911",
"0.5503",
"0.911",
"0.5645",
"0.5156",
"0.5645"
],
[
"0.5156",
"0.565",
"0.9062",
"0.565",
"0.9062",
"0.579",
"0.5156",
"0.579"
],
[
"0.08496",
"0.5903",
"0.1168",
"0.5903",
"0.1168",
"0.6016",
"0.08496",
"0.6016"
],
[
"0.11926",
"0.5903",
"0.2296",
"0.5903",
"0.2296",
"0.604",
"0.11926",
"0.604"
],
[
"0.5107",
"0.5913",
"0.851",
"0.5933",
"0.851",
"0.6094",
"0.5107",
"0.607"
],
[
"0.08826",
"0.613",
"0.486",
"0.613",
"0.486",
"0.627",
"0.08826",
"0.627"
],
[
"0.511",
"0.613",
"0.876",
"0.613",
"0.876",
"0.6274",
"0.511",
"0.6274"
],
[
"0.08905",
"0.629",
"0.4836",
"0.629",
"0.4836",
"0.6396",
"0.08905",
"0.6396"
],
[
"0.5186",
"0.631",
"0.7173",
"0.631",
"0.7173",
"0.6445",
"0.5186",
"0.6445"
],
[
"0.08905",
"0.64",
"0.4836",
"0.64",
"0.4836",
"0.654",
"0.08905",
"0.654"
],
[
"0.5146",
"0.646",
"0.911",
"0.646",
"0.911",
"0.66",
"0.5146",
"0.66"
],
[
"0.08826",
"0.655",
"0.4795",
"0.655",
"0.4795",
"0.6685",
"0.08826",
"0.6685"
],
[
"0.5146",
"0.66",
"0.9087",
"0.66",
"0.9087",
"0.674",
"0.5146",
"0.674"
],
[
"0.08905",
"0.6704",
"0.4836",
"0.6704",
"0.4836",
"0.681",
"0.08905",
"0.681"
],
[
"0.5146",
"0.675",
"0.9087",
"0.675",
"0.9087",
"0.688",
"0.5146",
"0.688"
],
[
"0.08826",
"0.683",
"0.4844",
"0.683",
"0.4844",
"0.6963",
"0.08826",
"0.6963"
],
[
"0.5146",
"0.688",
"0.9126",
"0.688",
"0.9126",
"0.702",
"0.5146",
"0.702"
],
[
"0.08826",
"0.6963",
"0.4844",
"0.696",
"0.4844",
"0.7104",
"0.08826",
"0.711"
],
[
"0.5146",
"0.6997",
"0.9126",
"0.7007",
"0.9126",
"0.7153",
"0.5146",
"0.714"
],
[
"0.0907",
"0.712",
"0.4836",
"0.712",
"0.4836",
"0.7227",
"0.0907",
"0.7227"
],
[
"0.513",
"0.714",
"0.638",
"0.7153",
"0.638",
"0.73",
"0.513",
"0.729"
],
[
"0.08826",
"0.7246",
"0.4844",
"0.7246",
"0.4844",
"0.738",
"0.08826",
"0.738"
],
[
"0.08905",
"0.738",
"0.4844",
"0.738",
"0.4844",
"0.752",
"0.08905",
"0.752"
],
[
"0.5156",
"0.735",
"0.6216",
"0.7363",
"0.6216",
"0.7505",
"0.5156",
"0.75"
],
[
"0.08826",
"0.7524",
"0.4844",
"0.7524",
"0.4844",
"0.7666",
"0.08826",
"0.7666"
],
[
"0.513",
"0.75",
"0.911",
"0.7505",
"0.911",
"0.765",
"0.513",
"0.7646"
],
[
"0.08905",
"0.7666",
"0.4844",
"0.7666",
"0.4844",
"0.78",
"0.08905",
"0.78"
],
[
"0.5146",
"0.765",
"0.683",
"0.765",
"0.683",
"0.779",
"0.5146",
"0.779"
],
[
"0.08826",
"0.78",
"0.4844",
"0.78",
"0.4844",
"0.7935",
"0.08826",
"0.7935"
],
[
"0.527",
"0.784",
"0.911",
"0.784",
"0.911",
"0.7974",
"0.527",
"0.7974"
],
[
"0.08826",
"0.7935",
"0.4844",
"0.7935",
"0.4844",
"0.8066",
"0.08826",
"0.8066"
],
[
"0.544",
"0.7974",
"0.812",
"0.7974",
"0.812",
"0.811",
"0.544",
"0.811"
],
[
"0.08905",
"0.8096",
"0.482",
"0.8096",
"0.482",
"0.8193",
"0.08905",
"0.8193"
],
[
"0.527",
"0.815",
"0.9126",
"0.815",
"0.9126",
"0.83",
"0.527",
"0.83"
],
[
"0.08826",
"0.8213",
"0.4844",
"0.8213",
"0.4844",
"0.8354",
"0.08826",
"0.8354"
],
[
"0.5415",
"0.8276",
"0.8057",
"0.83",
"0.8057",
"0.8447",
"0.5415",
"0.8423"
],
[
"0.08826",
"0.8354",
"0.4844",
"0.8354",
"0.4844",
"0.8486",
"0.08826",
"0.8486"
],
[
"0.0866",
"0.8496",
"0.3162",
"0.8496",
"0.3162",
"0.863",
"0.0866",
"0.863"
],
[
"0.528",
"0.8477",
"0.9126",
"0.8477",
"0.9126",
"0.8613",
"0.528",
"0.8613"
],
[
"0.546",
"0.8623",
"0.91",
"0.8623",
"0.91",
"0.8755",
"0.546",
"0.8755"
],
[
"0.1095",
"0.8755",
"0.2197",
"0.8755",
"0.2197",
"0.8896",
"0.1095",
"0.8896"
],
[
"0.546",
"0.8755",
"0.661",
"0.8755",
"0.661",
"0.8896",
"0.546",
"0.8896"
]
] | [
"基于人工智能的监测&响应系统，用于东南亚医院对COVID-19的准备",
"Tushar Goswamyl*, Naishadh Parmarl, Ayush Guptal, Raunak Shah', Vatsalya",
"Tandon', Varun Goyall, Sanyog Gupta', Karishma Laud', Shivam Gupta', Sudhanshu",
"Mishral,Ashutosh Modil.",
"1Indian Institute of Technology (IIT) Kanpur, Kanpur, India",
"{ tgoswamy, naishadh, ayushgup, raunaks, vatsalt, varung, sanyog, kslaud, guptash, sdhanshu",
"ashutoshm 1 @iitk.ac.in",
"Abstract",
"在这项研究中，我们使用用户的Twitter数据来研究",
"由于COVID-19大流行而导致的住院人数激增。在本研",
"This research paper proposes a COVID-19",
"究的范围内，我们将工作重点放在印度，印度尼西亚和",
"monitoring and response system to identify the",
"孟加拉国，并计划将这种方法扩展到东南亚的其他地区",
"surge in the volume of patients at hospitals",
"我们的研究旨在确定医院人满为患的事件，呼吸机等关",
"and shortage of critical equipment like venti-",
"键设备的短缺以及缺乏可用的ICU病房。该系统可以帮",
"lators in South-east Asian countries,to under-",
"助了解这些国家的医疗机构的医疗准备水平以及大流行",
"stand the burden on healthfacilities.This cal",
"蔓延时医院的负担。系统管道包括在粒度级别上刮擦历",
"help authorities in theseregions with resource",
"史推文以获得语料库，使用自然语言处理工具处理语料",
"planning measures to redirect resources to the",
"库，从处理后的数据中计算信号，最后通过比较地面报",
"regions identified by the model.Due to the",
"告和公告来评估结果。我们已经部署了神经翻译模型来",
"lack of publicly available data on the influx of",
"解释区域语言的使用",
"patients in hospitals, or the shortage of equip-",
"通过这项研究，我们对人工智能社区的主要贡献是展",
"ment, ICU units or hospital beds that regions",
"示了基于NLP的Twitter模型在监测COVID-19大流",
"in these countries might be facing, we leverage",
"行对卫生设施造成的负担方面的应用。据我们所知，这",
"Twitter data for gleaning this information. The",
"是同类产品中的第一种也是唯一的方法，它可以根据",
"approach has yielded accurate results for states",
"Twitter数据准确地检测受灾最严重地区的趋势。我们正",
"in India, and we are working on validating the",
"在与世卫组织东南亚区域办事处（WHO-SEARO）的成",
"model for the remaining countries so that it can",
"员密切合作，研究和监测我们模型的信号，旨在帮助他",
"rve as a reliable tool for authorities to moni-",
"们监测这些国家的情况，并确定哪些地区正面临着由于",
"tor the burden on hospitals.",
"大流行而导致的资源短缺。因此，公共卫生组织可以使",
"用我们的模型向模型确定的地区的当局建议适当的行动",
"1",
"Introduction",
"2Data Extraction and Pre-processing",
"像Twitter和Facebook这样的社交媒体网站鼓励用户",
"2.1Natural Language Processing for Tweets",
"频繁表达他们的想法，观点和生活的随机细节。印度拥有",
"Historical Tweet Extraction",
"全球第八大Twitter用户群，拥有13。1500万用户并在",
"我们使用getOldTweets3API[Mottl,Updated2019]从",
"增长，其次是印度尼西亚，拥有11。800万用户[Statista",
"Twitter网站抓取和提取历史推文。不幸的是，Twitter",
"2020。这凸显了从这些国家数百万用户发布的推文中获",
"有一些限制，由于这些限制，我们无法访问从刮擦之日",
"得有用见解的潜力。推文和状态更新的范围从重大事件",
"起超过7天的所有推文。这导致了数据中的误导性峰值",
"到无聊的评论。大多数消息包含很少的信息价值，但数",
"（图。1）。为了解决这个问题，我们使用整个峰值的变化",
"百万条消息的聚合可以产生有价值的知识。Twitter用户",
"因子来缩放推文。",
"经常公开表达有关医院人满为患，他们或其亲属因设备",
"短缺而面临的困难以及大流行引起的其他问题的个人经",
"Data Cleaning",
"历，这可以帮助了解情况的实际情况。先前的研究使用",
"为了消除数据中的噪声并提取重要信息，我们对tweet",
"有关症状[Pauland Dredze,2011] 的推文研究了Twitter",
"语料库执行了以下操作：",
"趋势与流感发生率之间的相关性。统计技术已用于使用",
": Removing Website Links: To prevent the same in-",
"Twitter数据[Santillanaetal.,2015]预测流感发病率。",
"formation from being captured twice.",
"在2012年Broniatowskietal.2013流感流行期间，美",
": Removing non-ASCII characters: To eliminate noise",
"国已在地方一级监测了流感发病率。同样，Signorini等",
"and focus on relevant keywords only",
"人[Signorinietal.,2011]研究了Twitter数据与H1N1",
"病例之间的相关性，以追踪感染",
". Removing Stopwords: Removed words like is', ‘an',",
"‘the' to focus on hospital-related words in the fre-",
"*Contact Author",
"quency analysis"
] | {
"arxiv_id": "2007.15619",
"language": "zh",
"page_idx": 0
} |
||
2007.15619-zh-1 | 2007.15619-zh-1.jpg | $\bullet$ Tokenisation: We utilized the NLTK TweetTok-
enizer API [Loper and Bird, 2002] to tokenize tweets. This was done to aid the keyword calculation process in subsequent steps.
$\bullet$ Lemmatisation: Implemented lemmatization on the
tokens obtained for each tweet to convert the higher
form of each word to their base forms.
我们观察到印尼语 tweet 被大量代码混合为印尼语 Bahasa 和英语。因此，我们实现了 Barik 等人描述的管道的修改版本。[Barik et al., 2019] 在计算分数之前对印度尼西亚的推文进行归一化和处理。对于来自孟加拉国的推文，大多数推文没有经过编码混合，而是使用罗马英语脚本或孟加拉语脚本。因此，我们使用上面提到的同一组操作处理英语推文，并为 Bangla 推文实现了标记化和规范化。
Keyword Selection
为了人围与我们的分析最相关并且可以为趋势产生准确信号的关键字，我们首先创建了与研究相关的常用词的语料库，例如“医院”,“ICU”等。然后使用潜在狄利克雷分配[Blei et al.,2003] 应用主题建模，以查找与我们的初始语料库相似类别的单词。主题建模根据相似词的使用情况以及它们的权重提供相似词的集群，以指示集群中的词的关联程度。我们还进行了 n-gram 分析， 以查找这些关键字在我们的语料库中的频率。然后使用Word2Vec [Mikolov et al., 2013] 查找与这些关键字最相似的单词。它允许我们通过考虑单词的词汇和语义特征来为词汇中的所有单词创建向量表示。研究了所有关键字的上下文，以通过避免不相关的单词/短语来最大程度地减少语料库中的噪音，同时确保捕获关键信号。最后根据上述方法，我们为印度，印度尼西亚和孟加拉国列出了以下关键字：
$\bullet$ India: ‘Hospital’,“Medical College’,“Beds’,“ICU”,
$`$Shortage"
$\bullet$ Indonesia: Hospital', Medical', Beds'. ICU"
$`$Shortage',“Rumah Sakit',“Tempat Tidur',“Over kapasitas',~Penuh',~`IGD',~“UGD',~`Fasilitas~Kese- $\hom$,
$\bullet$Bangladesh: Hospital',“Medical',“Beds',‘ICU',
$`\overset{`}{\operatorname*{Shortage',~`kendra'}}$
3
Twitter Score Calculation
3.1 Filtering
自 2020 年 3 月 1 日起，我们收集了印度所有 29 个州
印度尼西亚的 34 个省以及整个孟加拉国的推文，使用
${\text{关键字：'corona', 'covid'和'hospital'}}$
### 3.2 Score Calculation
我们针对该模型进行了不同的分数组合实验，最后根据
将使用该模型的公共卫生机构的要求人围以下内容：
Twitter Word Count/Day
我们通过计算某个地区人围关键字的每日计数来获得Twitter 单词计数/天数图。它旨在捕捉医院过度拥挤以及床位和关键设备短缺的事件
Twitter Volume/Day
Twitter 音量/日得分是根据过滤后的推文中所有单词的计数来计算的。这表明该地区与 COVID-19 大流行相关的推文数量趋势
4 Data Adjustment
4.1 Adjusting the peak
我们在上一节中提到的两个图中都发现了一个突然的峰值。经过彻底的分析和观察趋势，通过重新抓取一周的数据，我们发现，如果我们再次抓取数据，峰值会移动一天，并且总是发生在从抓取之日起的第 7 个历史日。这可以归因于 Twitter 可能对访问历史推文施加的限制。为了克服这个问题，我们使用跨峰值的值的比率对超过7天的历史推文进行了归一化。这是这样做的，因为最近 7 天的推文被刮掉了，这个问题只出现在从刮掉日期开始超过 7 天的推文中。德里的原始和调整后的地块可以在图中看到。1
<FigureHere>
Figure 1: Twitter Word Count/Day plot before and after
adjustment for Delhi
4.2 Data Modelling
当我们直接绘制数据时，它会拾取数据中的噪声，这可以看作是随机波动。这可能会在分析中产生误导，因此我们通过统计技术“平滑”数据。我们尝试了以下平滑技术，并列出了与阳性病例数据相关性最高的方法：
$\bullet$ Moving Averages: We successively plot the aver-
age of $n$ -days (which is the window size) to get a smoother curve which captures the overall trend better. Different values of $n$ ( 3/ 4/ 5 etc. ) are tested, and the one which gave the highest correlation with the positive cases is selected [Wau, Updated 2019]
$\bullet$Exponential Smoothing: we assign a higher weight
to the most recent observation when computing the moving averages. Under this approach, the most recent observation should get a little more weight than the 2nd most recent, and the 2nd most recent should get a little more weight than the 3rd most recent, and so on [ Nau, Updated 2019]
$\bullet$ Holt's Linear Exponential Smoothing: This tech-
nique includes two smoothing constants, one for the $\bar{\text{level and one for the trend. Two equations, one for}}$ an estimate of the local level, and the local trend's | [
[
"0.10297",
"0.0682",
"0.4844",
"0.0682",
"0.4844",
"0.0821",
"0.10297",
"0.0821"
],
[
"0.5195",
"0.07135",
"0.6685",
"0.07135",
"0.6685",
"0.0821",
"0.5195",
"0.0821"
],
[
"0.1177",
"0.0833",
"0.4878",
"0.0833",
"0.4878",
"0.0966",
"0.1177",
"0.0966"
],
[
"0.513",
"0.0852",
"0.91",
"0.0852",
"0.91",
"0.0985",
"0.513",
"0.0985"
],
[
"0.11926",
"0.09784",
"0.4844",
"0.09784",
"0.4844",
"0.1111",
"0.11926",
"0.1111"
],
[
"0.5146",
"0.09973",
"0.91",
"0.09973",
"0.91",
"0.113",
"0.5146",
"0.113"
],
[
"0.11926",
"0.11237",
"0.3652",
"0.11237",
"0.3652",
"0.1256",
"0.11926",
"0.1256"
],
[
"0.5146",
"0.113",
"0.63",
"0.113",
"0.63",
"0.127",
"0.5146",
"0.127"
],
[
"0.10297",
"0.1332",
"0.4844",
"0.1332",
"0.4844",
"0.1433",
"0.10297",
"0.1433"
],
[
"0.513",
"0.1445",
"0.5513",
"0.1445",
"0.5513",
"0.1559",
"0.513",
"0.1559"
],
[
"0.5435",
"0.1414",
"0.6953",
"0.1427",
"0.6953",
"0.1572",
"0.5435",
"0.1559"
],
[
"0.11926",
"0.1477",
"0.486",
"0.1477",
"0.486",
"0.1578",
"0.11926",
"0.1578"
],
[
"0.1177",
"0.1578",
"0.3938",
"0.1604",
"0.3938",
"0.1736",
"0.1177",
"0.1718"
],
[
"0.511",
"0.1635",
"0.71",
"0.1642",
"0.71",
"0.1787",
"0.511",
"0.178"
],
[
"0.08826",
"0.1799",
"0.4836",
"0.1799",
"0.4836",
"0.1932",
"0.08826",
"0.1932"
],
[
"0.5146",
"0.1818",
"0.9126",
"0.1818",
"0.9126",
"0.1957",
"0.5146",
"0.1957"
],
[
"0.08826",
"0.1932",
"0.486",
"0.1932",
"0.486",
"0.207",
"0.08826",
"0.207"
],
[
"0.513",
"0.1957",
"0.911",
"0.1963",
"0.911",
"0.2108",
"0.513",
"0.2102"
],
[
"0.08905",
"0.209",
"0.4836",
"0.209",
"0.4836",
"0.2197",
"0.08905",
"0.2197"
],
[
"0.511",
"0.2108",
"0.91",
"0.2108",
"0.91",
"0.2247",
"0.511",
"0.2247"
],
[
"0.08826",
"0.2216",
"0.4844",
"0.2216",
"0.4844",
"0.2349",
"0.08826",
"0.2349"
],
[
"0.527",
"0.2247",
"0.9126",
"0.2247",
"0.9126",
"0.238",
"0.527",
"0.238"
],
[
"0.08826",
"0.2349",
"0.4844",
"0.2349",
"0.4844",
"0.2488",
"0.08826",
"0.2488"
],
[
"0.513",
"0.2374",
"0.911",
"0.238",
"0.911",
"0.2524",
"0.513",
"0.252"
],
[
"0.0866",
"0.2488",
"0.4844",
"0.2488",
"0.4844",
"0.262",
"0.0866",
"0.262"
],
[
"0.5156",
"0.2524",
"0.911",
"0.2524",
"0.911",
"0.2664",
"0.5156",
"0.2664"
],
[
"0.1013",
"0.2632",
"0.486",
"0.2632",
"0.486",
"0.2766",
"0.1013",
"0.2766"
],
[
"0.5146",
"0.2676",
"0.9126",
"0.2676",
"0.9126",
"0.281",
"0.5146",
"0.281"
],
[
"0.08905",
"0.2766",
"0.1765",
"0.2766",
"0.1765",
"0.2886",
"0.08905",
"0.2886"
],
[
"0.5146",
"0.2798",
"0.911",
"0.2798",
"0.911",
"0.2942",
"0.5146",
"0.2942"
],
[
"0.5146",
"0.2935",
"0.911",
"0.2935",
"0.911",
"0.307",
"0.5146",
"0.307"
],
[
"0.09235",
"0.3005",
"0.2263",
"0.3005",
"0.2263",
"0.3145",
"0.09235",
"0.3145"
],
[
"0.5146",
"0.308",
"0.638",
"0.308",
"0.638",
"0.3213",
"0.5146",
"0.3213"
],
[
"0.08905",
"0.3164",
"0.486",
"0.3164",
"0.486",
"0.3308",
"0.08905",
"0.3308"
],
[
"0.08826",
"0.3308",
"0.4844",
"0.3308",
"0.4844",
"0.3447",
"0.08826",
"0.3447"
],
[
"0.08826",
"0.3447",
"0.4844",
"0.3447",
"0.4844",
"0.3591",
"0.08826",
"0.3591"
],
[
"0.08905",
"0.3591",
"0.4844",
"0.3591",
"0.4844",
"0.3726",
"0.08905",
"0.3726"
],
[
"0.5483",
"0.363",
"0.5645",
"0.363",
"0.5645",
"0.3674",
"0.5483",
"0.3674"
],
[
"0.5703",
"0.363",
"0.6284",
"0.363",
"0.6284",
"0.3674",
"0.5703",
"0.3674"
],
[
"0.08826",
"0.3726",
"0.4844",
"0.3726",
"0.4844",
"0.3865",
"0.08826",
"0.3865"
],
[
"0.5474",
"0.3726",
"0.562",
"0.377",
"0.558",
"0.3865",
"0.5435",
"0.3818"
],
[
"0.08826",
"0.387",
"0.4844",
"0.387",
"0.4844",
"0.401",
"0.08826",
"0.401"
],
[
"0.08826",
"0.401",
"0.4836",
"0.401",
"0.4836",
"0.414",
"0.08826",
"0.414"
],
[
"0.08826",
"0.414",
"0.4844",
"0.414",
"0.4844",
"0.428",
"0.08826",
"0.428"
],
[
"0.0866",
"0.4268",
"0.4844",
"0.428",
"0.4844",
"0.4426",
"0.0866",
"0.4414"
],
[
"0.08826",
"0.4426",
"0.4844",
"0.4426",
"0.4844",
"0.4558",
"0.08826",
"0.4558"
],
[
"0.08826",
"0.4558",
"0.486",
"0.4558",
"0.486",
"0.4697",
"0.08826",
"0.4697"
],
[
"0.08826",
"0.4697",
"0.4844",
"0.4697",
"0.4844",
"0.483",
"0.08826",
"0.483"
],
[
"0.6216",
"0.4702",
"0.634",
"0.4702",
"0.634",
"0.4797",
"0.6216",
"0.4797"
],
[
"0.8013",
"0.4717",
"0.8145",
"0.4717",
"0.8145",
"0.4797",
"0.8013",
"0.4797"
],
[
"0.863",
"0.473",
"0.8716",
"0.473",
"0.8716",
"0.4778",
"0.863",
"0.4778"
],
[
"0.74",
"0.4734",
"0.751",
"0.4734",
"0.751",
"0.4778",
"0.74",
"0.4778"
],
[
"0.683",
"0.4734",
"0.6914",
"0.4734",
"0.6914",
"0.4792",
"0.683",
"0.4792"
],
[
"0.0866",
"0.4817",
"0.4844",
"0.483",
"0.4844",
"0.4976",
"0.0866",
"0.4968"
],
[
"0.7036",
"0.4766",
"0.7305",
"0.4766",
"0.7305",
"0.4849",
"0.7036",
"0.4849"
],
[
"0.08826",
"0.4976",
"0.4844",
"0.4976",
"0.4844",
"0.511",
"0.08826",
"0.511"
],
[
"0.511",
"0.5005",
"0.911",
"0.5005",
"0.911",
"0.5146",
"0.511",
"0.5146"
],
[
"0.08905",
"0.513",
"0.2091",
"0.513",
"0.2091",
"0.5234",
"0.08905",
"0.5234"
],
[
"0.511",
"0.513",
"0.653",
"0.513",
"0.653",
"0.5264",
"0.511",
"0.5264"
],
[
"0.10046",
"0.531",
"0.4844",
"0.531",
"0.4844",
"0.5444",
"0.10046",
"0.5444"
],
[
"0.1209",
"0.547",
"0.1929",
"0.547",
"0.1929",
"0.558",
"0.1209",
"0.558"
],
[
"0.511",
"0.56",
"0.6846",
"0.5625",
"0.6846",
"0.577",
"0.511",
"0.5747"
],
[
"0.1013",
"0.5674",
"0.4429",
"0.5674",
"0.4429",
"0.5776",
"0.1013",
"0.5776"
],
[
"0.437",
"0.5664",
"0.4844",
"0.5664",
"0.4844",
"0.5776",
"0.437",
"0.5776"
],
[
"0.1177",
"0.58",
"0.4878",
"0.58",
"0.4878",
"0.5933",
"0.1177",
"0.5933"
],
[
"0.5156",
"0.5806",
"0.911",
"0.5806",
"0.911",
"0.5947",
"0.5156",
"0.5947"
],
[
"0.1177",
"0.5933",
"0.4844",
"0.5933",
"0.4844",
"0.607",
"0.1177",
"0.607"
],
[
"0.5146",
"0.5947",
"0.911",
"0.5947",
"0.911",
"0.608",
"0.5146",
"0.608"
],
[
"0.11926",
"0.6094",
"0.165",
"0.6094",
"0.165",
"0.6206",
"0.11926",
"0.6206"
],
[
"0.5146",
"0.607",
"0.911",
"0.607",
"0.911",
"0.622",
"0.5146",
"0.622"
],
[
"0.10046",
"0.6274",
"0.486",
"0.6274",
"0.486",
"0.6416",
"0.10046",
"0.6416"
],
[
"0.513",
"0.6226",
"0.8784",
"0.6226",
"0.8784",
"0.636",
"0.513",
"0.636"
],
[
"0.11926",
"0.64",
"0.259",
"0.6416",
"0.259",
"0.656",
"0.11926",
"0.655"
],
[
"0.527",
"0.6426",
"0.91",
"0.6426",
"0.91",
"0.657",
"0.527",
"0.657"
],
[
"0.544",
"0.657",
"0.911",
"0.657",
"0.911",
"0.6704",
"0.544",
"0.6704"
],
[
"0.08496",
"0.675",
"0.12335",
"0.675",
"0.12335",
"0.688",
"0.08496",
"0.688"
],
[
"0.1209",
"0.675",
"0.339",
"0.675",
"0.339",
"0.688",
"0.1209",
"0.688"
],
[
"0.546",
"0.6724",
"0.911",
"0.6724",
"0.911",
"0.683",
"0.546",
"0.683"
],
[
"0.544",
"0.685",
"0.91",
"0.685",
"0.91",
"0.6987",
"0.544",
"0.6987"
],
[
"0.0866",
"0.6987",
"0.1177",
"0.6987",
"0.1177",
"0.7104",
"0.0866",
"0.7104"
],
[
"0.1259",
"0.696",
"0.2035",
"0.6987",
"0.2018",
"0.7134",
"0.125",
"0.7104"
],
[
"0.546",
"0.6987",
"0.9126",
"0.6987",
"0.9126",
"0.712",
"0.546",
"0.712"
],
[
"0.546",
"0.712",
"0.9004",
"0.712",
"0.9004",
"0.726",
"0.546",
"0.726"
],
[
"0.1168",
"0.6997",
"0.1316",
"0.6997",
"0.1316",
"0.7085",
"0.1168",
"0.7085"
],
[
"0.0907",
"0.7197",
"0.482",
"0.7197",
"0.482",
"0.73",
"0.0907",
"0.73"
],
[
"0.08826",
"0.732",
"0.4844",
"0.732",
"0.4844",
"0.7456",
"0.08826",
"0.7456"
],
[
"0.5254",
"0.73",
"0.9126",
"0.731",
"0.9126",
"0.7456",
"0.5254",
"0.744"
],
[
"0.08826",
"0.7456",
"0.3586",
"0.7456",
"0.3586",
"0.759",
"0.08826",
"0.759"
],
[
"0.546",
"0.7456",
"0.9126",
"0.7456",
"0.9126",
"0.759",
"0.546",
"0.759"
],
[
"0.544",
"0.759",
"0.911",
"0.7583",
"0.911",
"0.773",
"0.544",
"0.7734"
],
[
"0.0866",
"0.7734",
"0.1136",
"0.7734",
"0.1136",
"0.785",
"0.0866",
"0.785"
],
[
"0.1316",
"0.773",
"0.273",
"0.773",
"0.273",
"0.786",
"0.1316",
"0.786"
],
[
"0.5474",
"0.7744",
"0.911",
"0.7744",
"0.911",
"0.785",
"0.5474",
"0.785"
],
[
"0.5474",
"0.787",
"0.911",
"0.787",
"0.911",
"0.8003",
"0.5474",
"0.8003"
],
[
"0.1168",
"0.776",
"0.1316",
"0.776",
"0.1316",
"0.7827",
"0.1168",
"0.7827"
],
[
"0.08826",
"0.7935",
"0.4844",
"0.7935",
"0.4844",
"0.8066",
"0.08826",
"0.8066"
],
[
"0.0866",
"0.806",
"0.452",
"0.8066",
"0.452",
"0.8213",
"0.0866",
"0.821"
],
[
"0.5474",
"0.803",
"0.911",
"0.803",
"0.911",
"0.813",
"0.5474",
"0.813"
],
[
"0.544",
"0.8164",
"0.8203",
"0.815",
"0.8203",
"0.827",
"0.544",
"0.8276"
],
[
"0.0907",
"0.83",
"0.2769",
"0.832",
"0.2769",
"0.845",
"0.0907",
"0.8433"
],
[
"0.528",
"0.834",
"0.9087",
"0.834",
"0.9087",
"0.8477",
"0.528",
"0.8477"
],
[
"0.0866",
"0.8477",
"0.4836",
"0.8477",
"0.4836",
"0.8613",
"0.0866",
"0.8613"
],
[
"0.546",
"0.8477",
"0.9126",
"0.8477",
"0.9126",
"0.8613",
"0.546",
"0.8613"
],
[
"0.08826",
"0.8623",
"0.4844",
"0.8623",
"0.4844",
"0.8755",
"0.08826",
"0.8755"
],
[
"0.546",
"0.8623",
"0.9126",
"0.8623",
"0.9126",
"0.8755",
"0.546",
"0.8755"
],
[
"0.08496",
"0.8755",
"0.2998",
"0.8755",
"0.2998",
"0.8896",
"0.08496",
"0.8896"
],
[
"0.544",
"0.8755",
"0.914",
"0.8755",
"0.914",
"0.8896",
"0.544",
"0.8896"
]
] | [
". Tokenisation:We utilized the NLTK TweetTok-",
"TwitterVolume/Day",
"enizer API [Loper and Bird, 2002] to tokenize",
"Twitter音量/日得分是根据过滤后的推文中所有单词的",
"tweets. This was done to aid the keyword calcu-",
"计数来计算的。这表明该地区与COVID-19大流行相关",
"lation process in subsequent steps.",
"的推文数量趋势",
".Lemmatisation:Implementedlemmatizationonthe",
"4",
"Data Adjustment",
"tokensobtainedforeachtweetto convertthehigher",
"form of each word to their base forms.",
"4.1Adjusting thepeak",
"我们观察到印尼语tweet被大量代码混合为印尼语Ba-",
"我们在上一节中提到的两个图中都发现了二一个突然的峰",
"hasa和英语。因此，我们实现了Barik等人描述的管道",
"值。经过彻底的分析和观察趋势，通过重新抓取一周的",
"的修改版本。Bariketal.，2019在计算分数之前对印度",
"数据，我们发现，如果我们再次抓取数据，峰值会移动",
"尼西亚的推文进行归一化和处理。对于来自孟加拉国的",
"天，并且总是发生在从抓取之日起的第7个历史日。",
"推文，大多数推文没有经过编码混合，而是使用罗马英",
"这可以归因于Twitter可能对访问历史推文施加的限制。",
"语脚本或孟加拉语脚本。因此，我们使用上面提到的同",
"为了克服这个问题，我们使用跨峰值的值的比率对超过",
"组操作处理英语推文，并为Bangla推文实现了标记",
"7天的历史推文进行了归一化。这是这样做的，因为最",
"化和规范化。",
"近7天的推文被刮掉了，这个问题只出现在从刮掉日期",
"开始超过7天的推文中。德里的原始和调整后的地块可",
"Keyword Selection",
"以在图中看到。1",
"为了入围与我们的分析最相关并且可以为趋势产生准确",
"信号的关键字，我们首先创建了与研究相关的常用词的",
"语料库，例如“医院”，“ICU”等。然后使用潜在狄利",
"克雷分配[Bleietal.，2003]应用主题建模，以查找与",
"1750:",
"Baforedietm",
"我们的初始语料库相似类别的单词。主题建模根据相似",
"1500",
"词的使用情况以及它们的权重提供相似词的集群，以指",
"示集群中的词的关联程度。我们还进行了n-gram分析",
"以查找这些关键字在我们的语料库中的频率。然后使用",
"Word2Vec[Mikolovetal.,2013]查找与这些关键字最相",
"似的单词。它允许我们通过考虑单词的词汇和语义特征",
"来为词汇中的所有单词创建向量表示。研究了所有关键",
"字的上下文，以通过避免不相关的单词/短语来最大程度",
"20",
"80",
"100",
"60",
"40",
"地减少语料库中的噪音，同时确保捕获关键信号。最后",
"Day Inde:",
"根据上述方法，我们为印度，印度尼西亚和孟加拉国列",
"Figure 1: Twitter Word Count/Day plot before and after",
"出了以下关键字：",
"adjustment for Delhi",
". India:\"Hospital',\"Medical College',“Beds',“ICU'",
"‘Shortage'",
"4.2Data Modelling",
".Indonesia:Hospital',Medical'Beds.",
"\"ICU\"",
"\"Shortage', \"Rumah Sakit', \"Tempat Tidur,\"Over",
"当我们直接绘制数据时，它会拾取数据中的噪声，这可",
"kapasitas',“Penuh',‘IGD',“UGD',“Fasilitas Kese-",
"以看作是随机波动。这可能会在分析中产生误导，因此",
"hatan",
"我们通过统计技术“平滑”数据。我们尝试了以下平滑",
":Bangladesh:Hospital', \"Medical',Beds', \"ICU',",
"技术，并列出了与阳性病例数据相关性最高的方法：",
"‘Shortage',‘kendra",
": Moving Averages: We successively plot the aver-",
"age of n -days (which is the window size) to get",
"3",
"Twitter Score Calculation",
"a smoother curve which captures the overall trend",
"better. Different values of n (3/4/5 etc.) are tested.",
"3.1",
"Filtering",
"and the one which gave the highest correlation with",
"the positive cases is selected [Nau, Updated 2019]",
"",
"自2020年3月1日起，我们收集了印度所有29个州",
"印度尼西亚的34个省以及整个孟加拉国的推文，使用",
". Exponential Smoothing: we assign a higher weight",
"关键字：\"corona\"，\"covid'和\"hospital”",
"to the most recent observation when computing the",
"moving averages. Under this approach, the most",
"3.2",
"Score Calculation",
"recent observation should get a littlemore weight",
"than the 2nd most recent, and the 2nd most recent",
"",
"我们针对该模型进行了不同的分数组合实验，最后根据",
"将使用该模型的公共卫生机构的要求人围以下内容：",
"should get a littlemore weight than the3rd most",
"recent,and soonNau,Updated2019",
"TwitterWordCount/Day",
": Holt's Linear Exponential Smoothing: This tech",
"我们通过计算某个地区人围关键字的每日计数来获得",
"nique includes two smoothing constants, one for the",
"Twitter单词计数/天数图。它旨在捕捉医院过度拥挤以",
"level and one for the trend. Two equations, one for",
"及床位和关键设备短缺的事件",
"an estimate of the local level, and the local trend's"
] | {
"arxiv_id": "2007.15619",
"language": "zh",
"page_idx": 1
} |
||
2007.15619-zh-3 | 2007.15619-zh-3.jpg | <FigureHere>
Figure 6: Twitter Word Count/Day Plot for Kerala
地捕捉到该州没有发生任何过度拥挤或关键设备短缺的
情况
### 6 Conclusion
从文献综述和获得的结果中，我们可以得出结论，从Twitter 数据中获得的信息可以提供有关疾病传播及其对医疗保健系统影响的有用见解。Twitter 可以提供有关医疗设施负担的基本现实的趋势，这可能不会在政府的官方报告中得到体现。。我们发现信号和峰值增加， 这与 COVID-19 病例数量的增加以及新闻报道证实的医院过度拥挤的发生率是一致的。因此，研究人员和流行病学家可以通过使用 Twitter 数据模型来扩展用于监测 COVID-19 大流行的方法范围，如本文所述。但是， Twitter 无法提供所有答案，并且对于某些类型的信息可能并不可靠。该模型的一个显着局限性是，社交媒体是用户可以自由发布任何内容的平台，因此，无法验证任何个人推文的声明。因此，我们所依赖的假设是，如果成千上万的人在推特上发布事件，这是真实的，值得报道。但是，这些需要由可信赖的来源进行验证，例如经过验证的新闻文章，以建立 Twitter 数据报告的声明。
References
[Barik et al., 2019] Anab Maulana Barik, Rahmad Ma-
hendra, and Mirna Adriani.
Normalization of
indonesian-english code-mixed twitter data. In W-
NUT@ EMNLP, 2019.
[Biswas, 2020] Soutik Biswas. Coronavirus: How india's kerala state'flattened the curve'. BBC News, April 2020.
[Blei et al.,2003] David M. Blei, Andrew Y. Ng, and Michael I. Jordan. Latent dirichlet allocation. J. Mach. Learn. Res. , 3( null) : 993- 1022, March 2003.
[Broniatowski et al., 2013] David A. Broniatowski, Michael J. Paul, and Mark Dredze. National and local infuenza surveillance through twitter: An analysis of the 2012-2013 influenza epidemic. PLOS ONE, 8( 12) , 12 2013.
[Dutt, 2020] Anonna Dutt. 20% beds in 117 private hos-
pitals to be reserved for covid-19 surge. Hindustan
Times, May 2020.
[Faleiro, 2020] Sonia Faleiro. What the world can learn
from kerala about how to fight covid-19. MIT Tech-
nology Review, April 2020.
[Jha, 2020] Durgesh Jha. Covid beds running out in
delhi private hospitals. Times of India, May 2020.
[Lalwani, 2020] Vijayta Lalwani.
$`$please help$°:$As
coronavirus cases soar in delhi, patients are struggling
to find hospital beds. Scroll India, June 2020.
[Loper and Bird, 2002] Edward Loper and Steven Bird. Nltk: The natural language toolkit. In In Proceedings of the ACL Workshop on Effective Tools and Methodologies for Teaching Natural Language Processing and Computational Linguistics. Philadelphia: Association for Computational Linguistics, 2002.
[Mikolov et al., 2013] Tomas Mikolov, Kai Chen, Greg
Corrado, and Jeffrey Dean. Efficient estimation of
word representations in vector space, 2013.
[Mottl, Updated 2019] Dmitry Mottl. Getoldtweets3.
${\mathrm{Updated:~2019.}}$
[Nau, Updated 2019] Robert Nau. Statistical forecast-
ing: notes on regression and time series analysis. Up-
${\mathrm{dated:~2019.}}$
[Paul and Dredze, 2011] Michael Paul and Mark Dredze. You are what you tweet: Analyzing twitter for public health. 2011.
[Rawat,2020] Mukesh Rawat. Coronavirus in india:
Tracking country's frst 50 covid-19 cases; what num-
bers tell. India Today, March 2020.
[Roy and Babu, 2020] Saikat Roy and Suresh Babu.
$\mathrm{Dealing~with~covid- 19~pandemic: ~Why~the~}$kerala $\mathrm{model}^{,}$ is working. The Indian Express, May 2020.
[Santillana et al., 2015] Mauricio Santillana, André T.
Nguyen, Mark Dredze, Michael J. Paul, Elaine O. Nsoesie, and John S. Brownstein. Combining search, social media, and traditional data sources to improve influenza surveillance. PLOS Computational Biology, 11(10){:}1{-}15, 10 2015.
[Sibtain, 2020] Qamar Sibtain. Coronavirus: Doctors
clash at overcrowded lnjp hospital. India Today, April
2020.
[Signorini et al., 2011] Alessio Signorini, Alberto Maria
Segre, and Philip M. Polgreen. The use of twitter to track levels of disease activity and public concern in the u.s. during the influenza a hln1 pandemic. PLOS ONE, 6( 5) : 1- 10, 05 2011.
[Staf, 2020] The Wire Staf. Mumbai runs out of hospital
beds for suspected covid-19 patients, starts a‘waitlist’
.The Wire, May 2020. [Statista, 2020] Statista.
Leading countries based
on number of Twitter users as of April 2020.
https://www.statista.com/statistics/242606/
number-of-active-twitter-users-in-selected-countries/,
April 2020.
[Tare, 2020] Kiran Tare. Mumbai: Viral video shows bodies of coronavirus victims lying next to patients at sion hospital. India Today, May 2020. | [
[
"0.5146",
"0.06757",
"0.914",
"0.06946",
"0.914",
"0.084",
"0.5146",
"0.0821"
],
[
"0.5293",
"0.0821",
"0.886",
"0.0833",
"0.886",
"0.09784",
"0.5293",
"0.0966"
],
[
"0.5146",
"0.1016",
"0.9126",
"0.1016",
"0.9126",
"0.11554",
"0.5146",
"0.11554"
],
[
"0.5293",
"0.11554",
"0.914",
"0.11615",
"0.914",
"0.1307",
"0.5293",
"0.13"
],
[
"0.5293",
"0.13",
"0.859",
"0.13",
"0.859",
"0.1445",
"0.5293",
"0.1445"
],
[
"0.5156",
"0.1483",
"0.911",
"0.1483",
"0.911",
"0.1622",
"0.5156",
"0.1622"
],
[
"0.528",
"0.161",
"0.9126",
"0.1622",
"0.9126",
"0.1768",
"0.528",
"0.1755"
],
[
"0.3757",
"0.1749",
"0.464",
"0.1749",
"0.464",
"0.1863",
"0.3757",
"0.1863"
],
[
"0.5293",
"0.1755",
"0.91",
"0.1755",
"0.91",
"0.1895",
"0.5293",
"0.1895"
],
[
"0.5293",
"0.1901",
"0.9126",
"0.1901",
"0.9126",
"0.2039",
"0.5293",
"0.2039"
],
[
"0.5293",
"0.2026",
"0.914",
"0.2039",
"0.914",
"0.2184",
"0.5293",
"0.2172"
],
[
"0.11273",
"0.2153",
"0.4575",
"0.214",
"0.4575",
"0.2279",
"0.11273",
"0.2291"
],
[
"0.5293",
"0.2166",
"0.7935",
"0.2172",
"0.7935",
"0.2317",
"0.5293",
"0.2311"
],
[
"0.5146",
"0.2349",
"0.9126",
"0.2361",
"0.9126",
"0.2507",
"0.5146",
"0.25"
],
[
"0.08826",
"0.2507",
"0.4844",
"0.2507",
"0.4844",
"0.2644",
"0.08826",
"0.2644"
],
[
"0.5312",
"0.2507",
"0.914",
"0.2507",
"0.914",
"0.2644",
"0.5312",
"0.2644"
],
[
"0.0866",
"0.2651",
"0.12335",
"0.2651",
"0.12335",
"0.281",
"0.0866",
"0.281"
],
[
"0.5293",
"0.2651",
"0.8384",
"0.2651",
"0.8384",
"0.279",
"0.5293",
"0.279"
],
[
"0.5156",
"0.2827",
"0.91",
"0.2827",
"0.91",
"0.2966",
"0.5156",
"0.2966"
],
[
"0.0825",
"0.2966",
"0.10297",
"0.2917",
"0.10864",
"0.3037",
"0.08826",
"0.3093"
],
[
"0.11926",
"0.2922",
"0.2157",
"0.2942",
"0.2157",
"0.3093",
"0.11926",
"0.307"
],
[
"0.5293",
"0.2974",
"0.6406",
"0.2974",
"0.6406",
"0.3113",
"0.5293",
"0.3113"
],
[
"0.08826",
"0.3145",
"0.486",
"0.3145",
"0.486",
"0.3289",
"0.08826",
"0.3289"
],
[
"0.5146",
"0.3145",
"0.911",
"0.3145",
"0.911",
"0.3276",
"0.5146",
"0.3276"
],
[
"0.0866",
"0.3289",
"0.486",
"0.3289",
"0.486",
"0.3435",
"0.0866",
"0.3435"
],
[
"0.528",
"0.33",
"0.9126",
"0.33",
"0.9126",
"0.3447",
"0.528",
"0.3447"
],
[
"0.08826",
"0.3423",
"0.4844",
"0.3423",
"0.4844",
"0.356",
"0.08826",
"0.356"
],
[
"0.5312",
"0.3447",
"0.621",
"0.3447",
"0.621",
"0.356",
"0.5312",
"0.356"
],
[
"0.08826",
"0.3547",
"0.4844",
"0.3567",
"0.4844",
"0.3706",
"0.08826",
"0.3687"
],
[
"0.5156",
"0.361",
"0.914",
"0.361",
"0.914",
"0.3757",
"0.5156",
"0.3757"
],
[
"0.08826",
"0.3706",
"0.4836",
"0.3706",
"0.4836",
"0.3838",
"0.08826",
"0.3838"
],
[
"0.5293",
"0.377",
"0.914",
"0.377",
"0.914",
"0.3901",
"0.5293",
"0.3901"
],
[
"0.08826",
"0.3833",
"0.4836",
"0.3833",
"0.4836",
"0.3965",
"0.08826",
"0.3965"
],
[
"0.5293",
"0.3901",
"0.701",
"0.3901",
"0.701",
"0.404",
"0.5293",
"0.404"
],
[
"0.08826",
"0.3965",
"0.4844",
"0.3977",
"0.4844",
"0.4124",
"0.08826",
"0.411"
],
[
"0.08826",
"0.4124",
"0.4844",
"0.4124",
"0.4844",
"0.4255",
"0.08826",
"0.4255"
],
[
"0.5156",
"0.4092",
"0.9126",
"0.4092",
"0.9126",
"0.4224",
"0.5156",
"0.4224"
],
[
"0.08826",
"0.4255",
"0.482",
"0.4255",
"0.482",
"0.4395",
"0.08826",
"0.4395"
],
[
"0.5317",
"0.4248",
"0.9087",
"0.4248",
"0.9087",
"0.435",
"0.5317",
"0.435"
],
[
"0.08826",
"0.4395",
"0.4844",
"0.4395",
"0.4844",
"0.4539",
"0.08826",
"0.4539"
],
[
"0.528",
"0.4375",
"0.786",
"0.4375",
"0.786",
"0.4507",
"0.528",
"0.4507"
],
[
"0.08905",
"0.4539",
"0.4844",
"0.4539",
"0.4844",
"0.4673",
"0.08905",
"0.4673"
],
[
"0.5146",
"0.4558",
"0.9126",
"0.4558",
"0.9126",
"0.4697",
"0.5146",
"0.4697"
],
[
"0.08905",
"0.4673",
"0.4844",
"0.4673",
"0.4844",
"0.481",
"0.08905",
"0.481"
],
[
"0.5293",
"0.4697",
"0.914",
"0.4697",
"0.914",
"0.483",
"0.5293",
"0.483"
],
[
"0.08826",
"0.481",
"0.4844",
"0.481",
"0.4844",
"0.4956",
"0.08826",
"0.4956"
],
[
"0.5293",
"0.4841",
"0.898",
"0.4841",
"0.898",
"0.4976",
"0.5293",
"0.4976"
],
[
"0.08905",
"0.4956",
"0.4844",
"0.4956",
"0.4844",
"0.509",
"0.08905",
"0.509"
],
[
"0.5146",
"0.502",
"0.911",
"0.502",
"0.911",
"0.515",
"0.5146",
"0.515"
],
[
"0.08826",
"0.509",
"0.4844",
"0.509",
"0.4844",
"0.523",
"0.08826",
"0.523"
],
[
"0.0866",
"0.5215",
"0.4697",
"0.521",
"0.4697",
"0.535",
"0.0866",
"0.536"
],
[
"0.5312",
"0.5166",
"0.91",
"0.5166",
"0.91",
"0.53",
"0.5312",
"0.53"
],
[
"0.5293",
"0.53",
"0.9126",
"0.53",
"0.9126",
"0.5435",
"0.5293",
"0.5435"
],
[
"0.5293",
"0.5435",
"0.9126",
"0.5435",
"0.9126",
"0.5566",
"0.5293",
"0.5566"
],
[
"0.08496",
"0.5503",
"0.1772",
"0.553",
"0.1772",
"0.5674",
"0.08496",
"0.565"
],
[
"0.528",
"0.5547",
"0.9126",
"0.558",
"0.9126",
"0.5728",
"0.528",
"0.5693"
],
[
"0.0866",
"0.5713",
"0.4836",
"0.5713",
"0.4836",
"0.5854",
"0.0866",
"0.5854"
],
[
"0.5312",
"0.5713",
"0.687",
"0.5713",
"0.687",
"0.5854",
"0.5312",
"0.5854"
],
[
"0.1013",
"0.586",
"0.3342",
"0.586",
"0.3342",
"0.5996",
"0.1013",
"0.5996"
],
[
"0.353",
"0.586",
"0.4878",
"0.586",
"0.4878",
"0.5996",
"0.353",
"0.5996"
],
[
"0.5146",
"0.5884",
"0.9126",
"0.5903",
"0.9126",
"0.605",
"0.5146",
"0.603"
],
[
"0.10455",
"0.5996",
"0.486",
"0.5996",
"0.486",
"0.613",
"0.10455",
"0.613"
],
[
"0.528",
"0.603",
"0.914",
"0.604",
"0.914",
"0.6187",
"0.528",
"0.617"
],
[
"0.10455",
"0.613",
"0.2615",
"0.613",
"0.2615",
"0.627",
"0.10455",
"0.627"
],
[
"0.5312",
"0.616",
"0.569",
"0.6187",
"0.568",
"0.631",
"0.5293",
"0.629"
],
[
"0.0866",
"0.632",
"0.486",
"0.632",
"0.486",
"0.646",
"0.0866",
"0.646"
],
[
"0.5156",
"0.637",
"0.9126",
"0.637",
"0.9126",
"0.651",
"0.5156",
"0.651"
],
[
"0.10297",
"0.6465",
"0.486",
"0.6465",
"0.486",
"0.661",
"0.10297",
"0.661"
],
[
"0.5293",
"0.6494",
"0.914",
"0.6494",
"0.914",
"0.664",
"0.5293",
"0.664"
],
[
"0.10455",
"0.659",
"0.1438",
"0.661",
"0.1422",
"0.674",
"0.10297",
"0.672"
],
[
"0.5293",
"0.664",
"0.9126",
"0.664",
"0.9126",
"0.678",
"0.5293",
"0.678"
],
[
"0.0866",
"0.678",
"0.486",
"0.6787",
"0.486",
"0.6934",
"0.0866",
"0.6924"
],
[
"0.528",
"0.678",
"0.9126",
"0.677",
"0.9126",
"0.6914",
"0.528",
"0.6924"
],
[
"0.10455",
"0.6924",
"0.4558",
"0.6924",
"0.4558",
"0.7056",
"0.10455",
"0.7056"
],
[
"0.4656",
"0.6963",
"0.4836",
"0.6963",
"0.4836",
"0.704",
"0.4656",
"0.704"
],
[
"0.5312",
"0.6924",
"0.714",
"0.6924",
"0.714",
"0.7056",
"0.5312",
"0.7056"
],
[
"0.10455",
"0.707",
"0.4592",
"0.707",
"0.4592",
"0.72",
"0.10455",
"0.72"
],
[
"0.5146",
"0.709",
"0.9126",
"0.7104",
"0.9126",
"0.7246",
"0.5146",
"0.7236"
],
[
"0.0866",
"0.7246",
"0.3743",
"0.7246",
"0.3743",
"0.738",
"0.0866",
"0.738"
],
[
"0.3823",
"0.7266",
"0.4836",
"0.7266",
"0.4836",
"0.738",
"0.3823",
"0.738"
],
[
"0.528",
"0.7236",
"0.911",
"0.7246",
"0.911",
"0.7393",
"0.528",
"0.738"
],
[
"0.3694",
"0.731",
"0.3857",
"0.728",
"0.3882",
"0.7344",
"0.3718",
"0.7373"
],
[
"0.1054",
"0.741",
"0.486",
"0.741",
"0.486",
"0.752",
"0.1054",
"0.752"
],
[
"0.5317",
"0.7373",
"0.697",
"0.738",
"0.697",
"0.7524",
"0.5317",
"0.752"
],
[
"0.10297",
"0.754",
"0.4558",
"0.754",
"0.4558",
"0.7676",
"0.10297",
"0.7676"
],
[
"0.4617",
"0.755",
"0.4844",
"0.755",
"0.4844",
"0.7666",
"0.4617",
"0.7666"
],
[
"0.5156",
"0.757",
"0.6953",
"0.757",
"0.6953",
"0.771",
"0.5156",
"0.771"
],
[
"0.716",
"0.756",
"0.9126",
"0.757",
"0.9126",
"0.7715",
"0.716",
"0.771"
],
[
"0.10297",
"0.7676",
"0.486",
"0.7666",
"0.486",
"0.781",
"0.10297",
"0.782"
],
[
"0.5293",
"0.7715",
"0.914",
"0.7715",
"0.914",
"0.785",
"0.5293",
"0.785"
],
[
"0.10297",
"0.78",
"0.2573",
"0.781",
"0.2573",
"0.7954",
"0.10297",
"0.7944"
],
[
"0.528",
"0.785",
"0.851",
"0.784",
"0.851",
"0.799",
"0.528",
"0.8"
],
[
"0.08496",
"0.799",
"0.4844",
"0.8",
"0.4844",
"0.8145",
"0.08496",
"0.813"
],
[
"0.5293",
"0.8",
"0.914",
"0.8",
"0.914",
"0.813",
"0.5293",
"0.813"
],
[
"0.10455",
"0.815",
"0.486",
"0.815",
"0.486",
"0.829",
"0.10455",
"0.829"
],
[
"0.5293",
"0.813",
"0.613",
"0.811",
"0.613",
"0.8257",
"0.5293",
"0.8276"
],
[
"0.1054",
"0.829",
"0.2345",
"0.829",
"0.2345",
"0.8423",
"0.1054",
"0.8423"
],
[
"0.513",
"0.83",
"0.914",
"0.8306",
"0.914",
"0.845",
"0.513",
"0.8447"
],
[
"0.08496",
"0.845",
"0.486",
"0.8467",
"0.486",
"0.8613",
"0.08496",
"0.86"
],
[
"0.528",
"0.8447",
"0.9126",
"0.845",
"0.9126",
"0.86",
"0.528",
"0.8594"
],
[
"0.10297",
"0.8613",
"0.486",
"0.8613",
"0.486",
"0.875",
"0.10297",
"0.875"
],
[
"0.5293",
"0.86",
"0.8037",
"0.86",
"0.8037",
"0.8735",
"0.5293",
"0.8735"
],
[
"0.10297",
"0.8755",
"0.298",
"0.8755",
"0.298",
"0.8896",
"0.10297",
"0.8896"
]
] | [
"[Jha, 2020] Durgesh Jha. Covid beds running out in",
"delhi private hospitals. Times of India, May 2020",
"[Lalwani, 2020] Vijayta Lalwani.please help':As",
"coronavirus cases soar in delhi, patients are struggling",
"to find hospital beds. Scroll India, June 2020.",
"[Loper and Bird, 2002] Edward Loper and Steven Bird.",
"Nltk: The natural language toolkit. In In Proceedings",
"AA",
"of the ACL Workshop on Effective Tools and Method-",
"ologies for Teaching Natural Language Processing and",
"Computational Linguistics.Philadelphia:Association",
"Figure 6: Twitter Word Count/Day Plot for Kerala",
"for Computational Linguistics, 2002.",
"[Mikolov et al., 2013] Tomas Mikolov, Kai Chen, Greg",
"地捕捉到该州没有发生任何过度拥挤或关键设备短缺的",
"Corrado, and Jeffrey Dean. Efficient estimation of",
"情况",
"word representations in vector space, 2013.",
"[Mottl, Updated 2019] Dmitry Mottl.Getoldtweets3.",
"6",
"Conclusion",
"Updated:2019.",
"从文献综述和获得的结果中，我们可以得出结论，从",
"[Nau.Updated 2019] Robert Nau. Statistical forecast-",
"Twitter数据中获得的信息可以提供有关疾病传播及其",
"ing: notes on regression and time series analysis. Up-",
"对医疗保健系统影响的有用见解。Twitter可以提供有",
"dated: 2019.",
"关医疗设施负担的基本现实的趋势，这可能不会在政府",
"[Paul and Dredze, 201i] Michael Pauland Mark",
"的官方报告中得到体现。。我们发现信号和峰值增加",
"Dredze. You are what you tweet: Analyzing twitter",
"这与COVID-19病例数量的增加以及新闻报道证实的",
"for public health. 2011.",
"医院过度拥挤的发生率是一致的。因此，研究人员和流",
"行病学家可以通过使用Twitter数据模型来扩展用于监",
"[Rawat, 2020] Mukesh Rawat.Coronavirus in india:",
"测COVID-19大流行的方法范围，如本文所述。但是",
"Tracking country'sfirst 50 covid-19cases:what num-",
"Twitter无法提供所有答案，并且对于某些类型的信息可",
"bers tell. India Today, March 2020.",
"能并不可靠。该模型的一个显着局限性是，社交媒体是",
"[Roy and Babu, 2020] Saikat Roy and Suresh Babu.",
"用户可以自由发布任何内容的平台，因此，无法验证任",
"Dealing with covid-i9 pandemic: Why the kerala",
"何个人推文的声明。因此，我们所依赖的假设是，如果",
"model’ is working. The Indian Express, May 2020.",
"成千上万的人在推特上发布事件，这是真实的，值得报",
"[Santillana etal.,2015] MauricioSantillana, AndreT.",
"道。但是，这些需要由可信赖的来源进行验证，例如经",
"过验证的新闻文章，以建立Twitter数据报告的声明。",
"Nguyen, Mark Dredze, Michael J. Paul, Elaine O.",
"Nsoesie, and John S. Brownstein. Combining search,",
"social media, and traditional data sources to improve",
"References",
"inffuenza surveillance.PLOS Computational Biology,",
"[Barik et al., 2019] Anab Maulana Barik, Rahmad Ma-",
"11(10):1-15, 10 2015.",
"hendra, and Mirna Adriani.",
"Normalization of",
"[Sibtain, 2020] Qamar Sibtain. Coronavirus: Doctors",
"indonesian-english code-mixed twitter data.In W-",
"clash at overcrowded Injp hospital. India Today, April",
"NUT@EMNLP,2019.",
"2020.",
"[Biswas, 202o] Soutik Biswas. Coronavirus: How india's",
"[Signorini et al., 201i] Alessio Signorini, Alberto Maria",
"kerala state 'fattened the curve: BBC News, April",
"Segre, and Philip M. Polgreen. The use of twitter to",
"2020.",
"track levels of disease activity and public concern in",
"[Blei et al., 2003] David M. Blei, Andrew Y.Ng, and",
"the u.s. during the influenza a hlnl pandemic. PLOS",
"Michael I.Jordan. Latent dirichlet allocation.",
"",
"ONE, 6(5):1-10, 05 2011.",
"Mach. Learn. Res., 3(null):993-1022, March 2003.",
"[Staf, 2020] The Wire Staf. Mumbai runs out of hospital",
"[Broniatowski et al.. 2013] DavidA.",
"Broniatowski",
"beds for suspected covid-19 patients, starts a ‘waitlist",
"",
"Michael J. Paul, and Mark Dredze.National and",
".The Wire, May 2020.",
"local influenza surveillance through twitter:",
"An",
"[Statista, 2020] Statista.",
"Leading countries based",
"analysis of the 2012-2013 influenza epidemic. PLOS",
"on number of Twitter users as of April 2020.",
"ONE, 8(12), 12 2013.",
"https://www.statista.com/statistics/242606/",
"[Dutt, 2020] Anonna Dutt. 20% beds in 117 private hos-",
"number-of-active-twitter-users-in-selected-countries/",
"pitals to be reserved for covid-19 surge.Hindustan",
"April 2020.",
"Times, May 2020.",
"[Tare, 2020] Kiran Tare. Mumbai: Viral video shows",
"[Faleiro, 2020] Sonia Faleiro. What the world can learn",
"bodies of coronavirus victims lying next to patients at",
"from kerala about how to fight covid-19. MIT Tech-",
"sion hospital. India Today, May 2020.",
"nology Review, April 2020."
] | {
"arxiv_id": "2007.15619",
"language": "zh",
"page_idx": 3
} |
||
2304.13445-zh-0 | 2304.13445-zh-0.jpg | Neural-PBIR Reconstruction of Shape, Material, and Illumination
Cheng Sun $^{1,2*}$
Kai Yan $^{3}$
Guangyan Cai $^{1,3*}$
Zhengqin Li$^{1}$
Cheng Zhang $^{1}$
Zhao Dong 1
Carl Marshall 1
Jia-Bin Huang $^{1,4}$
Shuang Zhao 3
$^{1}$Meta RLR
$^{2}$ National Tsing Hua University
$^3$ University of California, Irvine
$^{4}$University of Maryland, College Park
$^*$ The authors contribute equally to this paper
<FigureHere>
(a) Input images (b1) Albedo (b2) Roughness (b3) Shape&lighting (c1) Novel view (c2) Relit 1
(c3) Relit 2
Figure 1: Neural-PBIR recovers high-fidelity material (b1,2), shape and lighting (b3), enabling realistic re-
rendering (c1-3).
$\operatorname{Abstract}$
Reconstructing the shape and spatially varying surface appearances of a physical-world object as well as its surrounding illumination based on 2D images (e.g., photographs) of the object has been a long-standing problem in computer vision and graphics. In this paper, we introduce an accurate and highly efficient object reconstruction pipeline combining neural based object reconstruction and physics-based inverse rendering (PBIR) .Our pipeline firstly leverages a neural SDF based shape reconstruction to produce high-quality but potentially imperfect object shape. Then, we introduce a neural material and lighting distillation stage to achieve high-quality predictions for material and illumination. In the last stage, initialized by the neural predictions, we perform PBIR to refine the initial results and obtain the fnal high-quality reconstruction of object shape, material, and illumination. Experimental results demonstrate our pipeline significantly outperforms existing methods quality-wise and performancewise. Code: https://neural-pbir.github.io/
材料采集方法通常假设已知或简单的几何形状(例如。g。,平面), 具有高度控制的照明[25,38,47],通常用光舞台或龙门架创建。当这种捕获条件不可用时，这大大限制了它们的实用性
最近，新技术的出现使我们能够从物体的 2D 图像中共同重建形状，材料和照明。在高层次上，这些技术可以分为两类。神经重建方法将对象的外观编码到多层感知器 (MLP) 中，并通过可区分的体积光线跟踪最小化来自不同视图的渲染误差来优化网络。NeRF[21] 重建了基于密度场的辐射场，该辐射场允许高质量的视图合成，但不能重新照明。一系列方法[34,39,27] 从有符号距离函数计算密度场，以实现高质量的几何重建。最近的作品[44,45,4,5,23,12,46]寻求从输入图像中完全分解形状，材料和照明。但是，由于体积光线追踪和神经渲染的计算成本很高，这些方法需要花费数小时[23,12]或一天[4,46]来运行，并且通常无法对更复杂的间接照明[44,45,4,5,23,12]进行建模， 从而导致阴影和渗色被烘焙到材料反射率中。几种新方法[31,22,8]使用混合体积表示和有效的 MLP 显着降低了辐射场重建的计算成本。我们是最早采用这些新技术来有效地联合恢复几何形状，材料和照明的公司之一
相反，基于物理的逆渲染(PBIR)[7,19,26,1]通过计算图像外观相对于场景参数的无偏梯度来优化形状， 材质和照明。利用基于物理的可区分渲染器[42,13,18] ,最先进的 PBIR 管道可以有效地处理复杂的光传输效果，例如软阴影和互反射。通过基于体积的神经渲染无法轻松处理这种复杂的光传输效果。另一方面，由于PBIR 方法依赖于基于梯度的优化来细化内在成分，因此它们可能容易出现局部极小值和过拟合。因此，它们可能需要良好的初始化以实现最佳的重建质量
# 1. Introduction
从图像中重建几何，材料反射率和照明，也称为逆渲染，是计算机视觉和图形学中的长期挑战。传统上， 三个内在成分的获取主要是独立研究的。例如，多视图立体 (MVS) [29,30,10]和飞行时间[48]方法通常基于漫反射率假设，仅关注恢复对象几何形状。经典的 | [
[
"0.152",
"0.132",
"0.818",
"0.1332",
"0.818",
"0.1509",
"0.152",
"0.1497"
],
[
"0.1209",
"0.1812",
"0.2507",
"0.1768",
"0.2524",
"0.1989",
"0.1217",
"0.2026"
],
[
"0.2876",
"0.1812",
"0.451",
"0.178",
"0.451",
"0.1989",
"0.2876",
"0.202"
],
[
"0.6436",
"0.1812",
"0.739",
"0.1787",
"0.739",
"0.1976",
"0.6436",
"0.1995"
],
[
"0.774",
"0.1812",
"0.902",
"0.1831",
"0.902",
"0.2008",
"0.774",
"0.1989"
],
[
"0.3635",
"0.1989",
"0.5195",
"0.1963",
"0.5195",
"0.2153",
"0.3635",
"0.2172"
],
[
"0.4878",
"0.1818",
"0.6094",
"0.1818",
"0.6094",
"0.1995",
"0.4878",
"0.1995"
],
[
"0.5557",
"0.1995",
"0.693",
"0.1976",
"0.693",
"0.2153",
"0.5557",
"0.2172"
],
[
"0.7236",
"0.1989",
"0.824",
"0.2008",
"0.824",
"0.2153",
"0.7236",
"0.2134"
],
[
"0.08826",
"0.2172",
"0.1765",
"0.2197",
"0.1765",
"0.2343",
"0.08826",
"0.2317"
],
[
"0.1929",
"0.2008",
"0.326",
"0.2008",
"0.326",
"0.214",
"0.1929",
"0.214"
],
[
"0.1936",
"0.2172",
"0.4182",
"0.2184",
"0.4182",
"0.2329",
"0.1936",
"0.2317"
],
[
"0.4346",
"0.2197",
"0.66",
"0.2197",
"0.66",
"0.2329",
"0.4346",
"0.2329"
],
[
"0.6733",
"0.2197",
"0.937",
"0.2197",
"0.937",
"0.2329",
"0.6733",
"0.2329"
],
[
"0.3635",
"0.2361",
"0.669",
"0.238",
"0.669",
"0.252",
"0.3635",
"0.25"
],
[
"0.08496",
"0.363",
"0.2157",
"0.363",
"0.2157",
"0.377",
"0.08496",
"0.377"
],
[
"0.21",
"0.363",
"0.6846",
"0.363",
"0.6846",
"0.377",
"0.21",
"0.377"
],
[
"0.6978",
"0.3643",
"0.7812",
"0.3643",
"0.7812",
"0.3757",
"0.6978",
"0.3757"
],
[
"0.8096",
"0.363",
"0.892",
"0.361",
"0.892",
"0.375",
"0.8096",
"0.377"
],
[
"0.0809",
"0.385",
"0.8896",
"0.385",
"0.8896",
"0.3984",
"0.0809",
"0.3984"
],
[
"0.683",
"0.3674",
"0.701",
"0.3674",
"0.701",
"0.3718",
"0.683",
"0.3718"
],
[
"0.0801",
"0.401",
"0.1978",
"0.401",
"0.1978",
"0.414",
"0.0801",
"0.414"
],
[
"0.2378",
"0.4255",
"0.312",
"0.4255",
"0.312",
"0.44",
"0.2378",
"0.44"
],
[
"0.505",
"0.428",
"0.8906",
"0.428",
"0.8906",
"0.4414",
"0.505",
"0.4414"
],
[
"0.5034",
"0.443",
"0.8896",
"0.443",
"0.8896",
"0.457",
"0.5034",
"0.457"
],
[
"0.0972",
"0.459",
"0.4656",
"0.4602",
"0.4656",
"0.4749",
"0.0972",
"0.4734"
],
[
"0.505",
"0.4583",
"0.892",
"0.4583",
"0.892",
"0.4717",
"0.505",
"0.4717"
],
[
"0.0809",
"0.476",
"0.4673",
"0.476",
"0.4673",
"0.4893",
"0.0809",
"0.4893"
],
[
"0.5024",
"0.4734",
"0.6665",
"0.4734",
"0.6665",
"0.4873",
"0.5024",
"0.4873"
],
[
"0.0809",
"0.4912",
"0.4673",
"0.4912",
"0.4673",
"0.505",
"0.0809",
"0.505"
],
[
"0.5215",
"0.4912",
"0.8906",
"0.4912",
"0.8906",
"0.505",
"0.5215",
"0.505"
],
[
"0.0809",
"0.507",
"0.468",
"0.507",
"0.468",
"0.521",
"0.0809",
"0.521"
],
[
"0.5034",
"0.505",
"0.8906",
"0.506",
"0.8906",
"0.521",
"0.5034",
"0.5195"
],
[
"0.0809",
"0.5215",
"0.468",
"0.5215",
"0.468",
"0.535",
"0.0809",
"0.535"
],
[
"0.505",
"0.5215",
"0.8896",
"0.5215",
"0.8896",
"0.535",
"0.505",
"0.535"
],
[
"0.0801",
"0.535",
"0.4673",
"0.534",
"0.4673",
"0.549",
"0.0801",
"0.55"
],
[
"0.5034",
"0.536",
"0.8906",
"0.536",
"0.8906",
"0.55",
"0.5034",
"0.55"
],
[
"0.0784",
"0.5503",
"0.4673",
"0.55",
"0.4673",
"0.5645",
"0.0784",
"0.565"
],
[
"0.5034",
"0.5503",
"0.888",
"0.5503",
"0.888",
"0.5645",
"0.5034",
"0.5645"
],
[
"0.0809",
"0.5664",
"0.468",
"0.5664",
"0.468",
"0.58",
"0.0809",
"0.58"
],
[
"0.505",
"0.5674",
"0.8896",
"0.5674",
"0.8896",
"0.5806",
"0.505",
"0.5806"
],
[
"0.0809",
"0.5806",
"0.468",
"0.5806",
"0.468",
"0.595",
"0.0809",
"0.595"
],
[
"0.505",
"0.582",
"0.888",
"0.582",
"0.888",
"0.595",
"0.505",
"0.595"
],
[
"0.0809",
"0.5967",
"0.468",
"0.5967",
"0.468",
"0.61",
"0.0809",
"0.61"
],
[
"0.505",
"0.5967",
"0.8906",
"0.5967",
"0.8906",
"0.6113",
"0.505",
"0.6113"
],
[
"0.0809",
"0.6113",
"0.4656",
"0.6113",
"0.4656",
"0.625",
"0.0809",
"0.625"
],
[
"0.505",
"0.6123",
"0.8896",
"0.6123",
"0.8896",
"0.6255",
"0.505",
"0.6255"
],
[
"0.0801",
"0.627",
"0.468",
"0.627",
"0.468",
"0.64",
"0.0801",
"0.64"
],
[
"0.505",
"0.627",
"0.8906",
"0.627",
"0.8906",
"0.64",
"0.505",
"0.64"
],
[
"0.0809",
"0.6426",
"0.468",
"0.6426",
"0.468",
"0.656",
"0.0809",
"0.656"
],
[
"0.505",
"0.6426",
"0.892",
"0.6426",
"0.892",
"0.657",
"0.505",
"0.657"
],
[
"0.0801",
"0.657",
"0.468",
"0.657",
"0.468",
"0.6704",
"0.0801",
"0.6704"
],
[
"0.505",
"0.6577",
"0.892",
"0.6577",
"0.892",
"0.672",
"0.505",
"0.672"
],
[
"0.0801",
"0.6724",
"0.4673",
"0.6724",
"0.4673",
"0.686",
"0.0801",
"0.686"
],
[
"0.5034",
"0.6724",
"0.8906",
"0.6724",
"0.8906",
"0.686",
"0.5034",
"0.686"
],
[
"0.0809",
"0.6875",
"0.468",
"0.6875",
"0.468",
"0.7007",
"0.0809",
"0.7007"
],
[
"0.505",
"0.6875",
"0.8906",
"0.6875",
"0.8906",
"0.7007",
"0.505",
"0.7007"
],
[
"0.0801",
"0.702",
"0.468",
"0.7007",
"0.468",
"0.7153",
"0.0801",
"0.7163"
],
[
"0.5034",
"0.702",
"0.8906",
"0.7007",
"0.8906",
"0.7153",
"0.5034",
"0.7163"
],
[
"0.0809",
"0.7163",
"0.4656",
"0.7163",
"0.4656",
"0.73",
"0.0809",
"0.73"
],
[
"0.5034",
"0.7173",
"0.8906",
"0.7173",
"0.8906",
"0.731",
"0.5034",
"0.731"
],
[
"0.0809",
"0.7344",
"0.464",
"0.7344",
"0.464",
"0.744",
"0.0809",
"0.744"
],
[
"0.505",
"0.733",
"0.8896",
"0.733",
"0.8896",
"0.747",
"0.505",
"0.747"
],
[
"0.0809",
"0.7476",
"0.384",
"0.7476",
"0.384",
"0.761",
"0.0809",
"0.761"
],
[
"0.505",
"0.7485",
"0.54",
"0.7485",
"0.54",
"0.7603",
"0.505",
"0.7603"
],
[
"0.524",
"0.765",
"0.8906",
"0.765",
"0.8906",
"0.78",
"0.524",
"0.78"
],
[
"0.5063",
"0.7827",
"0.8896",
"0.7827",
"0.8896",
"0.7935",
"0.5063",
"0.7935"
],
[
"0.0801",
"0.8",
"0.2173",
"0.8",
"0.2173",
"0.813",
"0.0801",
"0.813"
],
[
"0.5024",
"0.797",
"0.888",
"0.797",
"0.888",
"0.81",
"0.5024",
"0.81"
],
[
"0.5195",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.5195",
"0.8247"
],
[
"0.1013",
"0.8257",
"0.468",
"0.8257",
"0.468",
"0.8403",
"0.1013",
"0.8403"
],
[
"0.5034",
"0.8247",
"0.8906",
"0.8257",
"0.8906",
"0.8403",
"0.5034",
"0.839"
],
[
"0.0809",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0809",
"0.855"
],
[
"0.505",
"0.8413",
"0.8896",
"0.8413",
"0.8896",
"0.855",
"0.505",
"0.855"
],
[
"0.08417",
"0.857",
"0.468",
"0.857",
"0.468",
"0.8706",
"0.08417",
"0.8706"
],
[
"0.5034",
"0.857",
"0.8896",
"0.857",
"0.8896",
"0.8706",
"0.5034",
"0.8706"
],
[
"0.0825",
"0.872",
"0.468",
"0.872",
"0.468",
"0.8853",
"0.0825",
"0.8853"
],
[
"0.5034",
"0.8706",
"0.8906",
"0.872",
"0.8906",
"0.886",
"0.5034",
"0.8853"
],
[
"0.0809",
"0.887",
"0.4673",
"0.887",
"0.4673",
"0.901",
"0.0809",
"0.901"
],
[
"0.505",
"0.887",
"0.831",
"0.887",
"0.831",
"0.901",
"0.505",
"0.901"
]
] | [
"Neural-PBIR Reconstruction of Shape, Material, and Illumination",
"Cheng Sun 1,2*",
"Guangyan Cai 1,3*",
"Kai Yan 3",
"Cheng Zhang 1",
"Jia-Bin Huang 1,4",
"Zhengqin Li 1",
"Shuang Zhao 3",
" Zhao Dong",
"1 Meta RLR",
"Carl Marshall 1",
"2National Tsing Hua University",
"3 University of California, Irvine",
"4 University of Maryland, College Park",
"The authors contribute equally to this paper",
"(a) Input images",
"(b1) Albedo(b2) Roughness (b3) Shape& lighting (c1) Novel view",
"(c2) Relit 1",
"(c3) Relit 2",
"Figure 1:Neural-PBIR recovers high-fidelity material (bl,2), shape and lighting (b3), enabling realistic re-",
"",
"rendering (cl-3)",
"Abstract",
"材料采集方法通常假设已知或简单的几何形状（例如。",
"g。，平面)，具有高度控制的照明[25，38，47]，通常用",
"Reconstructing the shape and spatially varying sur-",
"光舞台或龙门架创建。当这种捕获条件不可用时，这大",
"face appearances of a physical-world object as well as",
"大限制了它们的实用性",
"its surrounding illumination based on 2D images (e.g.,",
"最近，新技术的出现使我们能够从物体的2D图像",
"photographs) of the object has been a long-standing",
"中共同重建形状，材料和照明。在高层次上，这些技术",
"problem in computer vision and graphics. In this pa-",
"可以分为两类。神经重建方法将对象的外观编码到多",
"per, we introduce an accurate and highly efficient ob-",
"层感知器（MLP）中，并通过可区分的体积光线跟踪最",
"ject reconstruction pipeline combining neural based ob-",
"小化来自不同视图的谊染误差来优化网络。NeRF21",
"ject reconstruction and physics-based inverse rendering",
"重建了基于密度场的辐射场，该辐射场允许高质量的",
"(PBIR) . Our pipeline firstly leverages a neural SDF",
"视图合成，但不能重新照明。一系列方法[34，39，27",
"based shape reconstruction to produce high-quality but",
"从有符号距离函数计算密度场，以实现高质量的几何",
"potentially imperfect object shape.Then, we intro-",
"重建。最近的作品「44，45，4，5，23，12，46】寻求从输",
"duce a neural material and lighting distillation stage",
"图像中完全分解形状，材料和照明。但是，由于体积光",
"to achieve high-quality predictions for material and il-",
"线追踪和神经染的计算成本很高，这些方法需要花",
"lumination. In the last stage,initialized by the neural",
"费数小时[23，12]或一天4，46】来运行，并且通常无",
"predictions, we perform PBIR to refine the initial re-",
"法对更复杂的间接照明「44，45，4，5，23，121进行建模，",
"sults and obtain the final high-quality reconstruction of",
"从而导致阴影和渗色被烘焙到材料反射率中。几种新",
"object shape, material, and illumination. Experimental",
"方法[31，22，8]使用混合体积表示和有效的MLP显着",
"results demonstrate our pipeline significantly outper-",
"降低了辐射场重建的计算成本。我们是最早采用这些",
"formsexistingmethodsquality-wiseandperformance",
"新技术来有效地联合恢复几何形状，材料和照明的公",
"wise. Code: https://neural-pbir.github.io/",
"司之",
"相反，基于物理的逆染（PBIR）[7，19,26,11通过",
"计算图像外观相对于场景参数的无偏梯度来优化形状，",
"1.Introduction",
"材质和照明。利用基于物理的可区分染器「42，13,18",
"最先进的PBIR管道可以有效地处理复杂的光传输",
"从图像中重建几何，材料反射率和照明，也称为逆",
"效果，例如软阴影和互反射。通过基于体积的神经染",
"谊染，是计算机视觉和图形学中的长期挑战。传统上",
"无法轻松处理这种复杂的光传输效果。另一方面，由于",
"三个内在成分的获取主要是独立研究的。例如，多视图",
"PBIR方法依赖于基于梯度的优化来细化内在成分，因",
"立体（MVS）29，30，10]和飞行时间【48]方法通常基",
"此它们可能容易出现局部极小值和过拟合。因此，它们",
"于漫反射率假设，仅关注恢复对象几何形状。经典的",
"可能需要良好的初始化以实现最佳的重建质量"
] | {
"arxiv_id": "2304.13445",
"language": "zh",
"page_idx": 0
} |
||
2304.13445-zh-1 | 2304.13445-zh-1.jpg | <FigureHere>
在本文中，我们提出了一种高效，准确的逆渲染流水线，具有神经重建和 PBIR 方法的优点。我们的管道尝试根据在静态但任意照明下捕获的对象的多个图像来估计几何形状，空间变化的材料反射率和 HDR 环境图。如图所示。2,我们的管道由三个阶段组成。在第一阶段，我们提出了一种用于快速神经 SDF 和辐射场重建的基于混合神经体积的方法，该方法实现了最先进的几何精度和效率。在下一阶段，基于重建的几何形状和辐射场，我们设计了一种有效的优化方法，通过拟合表面光场来提炼材料和照明。我们的方法依赖于辐射场来处理可见性和间接照明，但与最近的一些工作相比，避免了昂贵的体积光线跟踪到显着的计算成本。最后，我们使用先进的 PBIR 框架[42,13]来共同完善几何、材料和照明。请注意，我们的 PBIR 框架以物理上正确和无偏的方式对复杂的光传输效果(如可见性， 遮挡，柔和阴影和相互反射)进行建模，同时仍然比最近的逆渲染方法快得多
Figure 2: Our pipeline for joint shape, material, and lighting estimation.
[9] 和曼哈顿对齐$\mathbb{E}[11]$ 来提高质量。这些方法使用大型 MLP 作为其体积表示，但是每个场景的优化速度很慢(许多小时)。最近的进展[31,40,22,8]通过使用显式网格显示出极大的优化加速而不损失质量。不幸的是，使用显式体积来建模 SDF 会导致颠簸和嘈杂的表面 [41] 。SfM 稀疏点云$\mathbb{E}[43,9]$ 或单眼深度/正常$\mathbb{E}[41]$ 的监督可能会减轻难度，但超出了我们的范围。我们为显式 SDF 网格优化提出了简单有效的正则化损失，在不使用外部先验的情况下实现了快速和高质量的表面
Material and lighting estimation. 几种神经重建方法[44,45,4,5,23,12,46]采用与我们相同的设置，可以从多个图像中同时重建几何，材料和照明。较早的工作 [45] 直接从固定的几何形状优化了低分辨率的环境图和材料，从而导致了由该问题的高度不适定性引起的嘈杂的照明重建。最近的方法使用球形高斯[44,4,46] 或预滤波近似[5,23]的混合物对照明进行建模，并使用球形高斯瓣 [44] 或低维潜在代码[4,5,46]约束材料反射，以提高重建质量。由于基于 MLP 的神经渲染的高计算成本，它们中的大多数仅考虑无遮挡的直接照明[44,4,23]或使用固定阴影贴图[45,5] 。Nvdiffrecmc [12]使用带有去噪器的可微分光线跟踪器对阴影进行建模，该去噪器可显着减少样本数量，但仍无法对相互反射进行建模。MII [46]训练另一个网络来预测来自辐射场的可见性和相互反射，这需要几个小时。相反，我们的混合，高度优化的神经 SDF 框架使我们能够在不到 10 分钟的时间内有效地提取几何形状，材料，照明， 可见性和相互反射。我们先进的 PBIR 框架首次允许以物理无偏的方式建模的直接和间接照明对照明，材料和几何形状进行整体优化
Physics-based inverse rendering. 与基于神经的方法不同，最近的几部作品[19,12,7]利用图形界开发的经典图像公式化模型，并尝试使用基于梯度的优化来反转此过程，并使用基于物理的可微分渲染器计算梯度。这些方法可以在优化过程中处理复杂的光传输效应，但容易出现局部极小值和过拟合。在我们的工作中，我们不是从头开始优化，而是将 PBIR 作为改进
具体而言，我们的贡献包括以下内容：
$\bullet$ A hybrid volume representation for fast and accu-
rate geometry reconstruction.
$\bullet$ A efficient optimization scheme to distill high-
quality initial material and lighting estimation from
the reconstructed geometry and radiance field
$\bullet$ An advanced PBIR framework that jointly opti-
mizes materials, lighting and geometry with visibility and interreflection handled in a physically unbiased way.
$\bullet$ A end-to-end pipeline that achieves state-of-the-art
geometry, material and lighting estimation that en-
ables realistic view synthesis and relighting.
# 2. Related Works
Volumetric surface reconstruction.
静态场景的基于
体积的表面重建的最新进展显示了高质量和可靠的结果。NeuS [34] 和 VolSDF [39] 用有符号距离值代替NeRF [21] 的密度预测，提出了一种无偏和遮挡感知的体绘制，取得了可喜的结果。后续工作通过引入正则化损失来促进表面平滑度$\mathbb{E}[43]$ ,多视图一致性$\mathbb{E}$
2 | [
[
"0.1447",
"0.1016",
"0.3015",
"0.1016",
"0.3015",
"0.11237",
"0.1447",
"0.11237"
],
[
"0.3857",
"0.1029",
"0.581",
"0.1029",
"0.581",
"0.113",
"0.3857",
"0.113"
],
[
"0.6567",
"0.09973",
"0.8267",
"0.101",
"0.8267",
"0.11426",
"0.6567",
"0.113"
],
[
"0.29",
"0.137",
"0.308",
"0.137",
"0.308",
"0.1465",
"0.29",
"0.1465"
],
[
"0.789",
"0.1364",
"0.812",
"0.1364",
"0.812",
"0.1477",
"0.789",
"0.1477"
],
[
"0.2778",
"0.1483",
"0.3203",
"0.1483",
"0.3203",
"0.159",
"0.2778",
"0.159"
],
[
"0.7686",
"0.1547",
"0.822",
"0.1572",
"0.8203",
"0.1724",
"0.767",
"0.1704"
],
[
"0.7983",
"0.1528",
"0.8164",
"0.1477",
"0.8203",
"0.1559",
"0.8022",
"0.1604"
],
[
"0.513",
"0.1578",
"0.5435",
"0.1578",
"0.5435",
"0.1692",
"0.513",
"0.1692"
],
[
"0.3308",
"0.1572",
"0.3896",
"0.1572",
"0.3896",
"0.1724",
"0.3308",
"0.1724"
],
[
"0.5376",
"0.1604",
"0.5605",
"0.161",
"0.5596",
"0.1724",
"0.536",
"0.1718"
],
[
"0.589",
"0.1635",
"0.638",
"0.1635",
"0.638",
"0.178",
"0.589",
"0.178"
],
[
"0.2769",
"0.1692",
"0.3218",
"0.1718",
"0.3203",
"0.1831",
"0.2754",
"0.1812"
],
[
"0.773",
"0.1718",
"0.819",
"0.1718",
"0.819",
"0.1818",
"0.773",
"0.1818"
],
[
"0.5146",
"0.1749",
"0.5596",
"0.1749",
"0.5596",
"0.1863",
"0.5146",
"0.1863"
],
[
"0.29",
"0.161",
"0.3218",
"0.161",
"0.3218",
"0.1692",
"0.29",
"0.1692"
],
[
"0.1405",
"0.1913",
"0.179",
"0.1913",
"0.179",
"0.1957",
"0.1405",
"0.1957"
],
[
"0.511",
"0.1925",
"0.562",
"0.1925",
"0.562",
"0.207",
"0.511",
"0.207"
],
[
"0.7715",
"0.1901",
"0.822",
"0.1925",
"0.8203",
"0.209",
"0.7695",
"0.207"
],
[
"0.429",
"0.2052",
"0.4592",
"0.2052",
"0.4592",
"0.2166",
"0.429",
"0.2166"
],
[
"0.687",
"0.2052",
"0.7134",
"0.2052",
"0.7134",
"0.2134",
"0.687",
"0.2134"
],
[
"0.7812",
"0.2108",
"0.8105",
"0.2108",
"0.8105",
"0.2229",
"0.7812",
"0.2229"
],
[
"0.523",
"0.2122",
"0.5513",
"0.2122",
"0.5513",
"0.2235",
"0.523",
"0.2235"
],
[
"0.267",
"0.2108",
"0.335",
"0.2108",
"0.335",
"0.2216",
"0.267",
"0.2216"
],
[
"0.125",
"0.2216",
"0.152",
"0.2235",
"0.1503",
"0.2343",
"0.12335",
"0.2317"
],
[
"0.701",
"0.2203",
"0.7344",
"0.2203",
"0.7344",
"0.2424",
"0.701",
"0.2424"
],
[
"0.125",
"0.2343",
"0.156",
"0.2343",
"0.156",
"0.2424",
"0.125",
"0.2424"
],
[
"0.3308",
"0.2361",
"0.38",
"0.2361",
"0.38",
"0.252",
"0.3308",
"0.252"
],
[
"0.4355",
"0.2437",
"0.501",
"0.2474",
"0.5",
"0.262",
"0.4346",
"0.2583"
],
[
"0.789",
"0.2437",
"0.834",
"0.2437",
"0.834",
"0.2551",
"0.789",
"0.2551"
],
[
"0.544",
"0.2443",
"0.5884",
"0.2443",
"0.5884",
"0.2551",
"0.544",
"0.2551"
],
[
"0.1772",
"0.2443",
"0.2279",
"0.2443",
"0.2279",
"0.2588",
"0.1772",
"0.2588"
],
[
"0.6855",
"0.25",
"0.763",
"0.254",
"0.761",
"0.2651",
"0.6846",
"0.2615"
],
[
"0.6724",
"0.2588",
"0.6787",
"0.2588",
"0.6787",
"0.262",
"0.6724",
"0.262"
],
[
"0.721",
"0.2715",
"0.7256",
"0.2715",
"0.7256",
"0.2766",
"0.721",
"0.2766"
],
[
"0.2263",
"0.2974",
"0.7427",
"0.2974",
"0.7427",
"0.3113",
"0.2263",
"0.3113"
],
[
"0.10046",
"0.3245",
"0.468",
"0.3245",
"0.468",
"0.3384",
"0.10046",
"0.3384"
],
[
"0.505",
"0.3245",
"0.8896",
"0.3245",
"0.8896",
"0.3384",
"0.505",
"0.3384"
],
[
"0.0825",
"0.339",
"0.468",
"0.339",
"0.468",
"0.3528",
"0.0825",
"0.3528"
],
[
"0.5024",
"0.3372",
"0.8906",
"0.3384",
"0.8906",
"0.3528",
"0.5024",
"0.3516"
],
[
"0.0825",
"0.3547",
"0.468",
"0.3547",
"0.468",
"0.3687",
"0.0825",
"0.3687"
],
[
"0.5034",
"0.3547",
"0.8906",
"0.3547",
"0.8906",
"0.3687",
"0.5034",
"0.3687"
],
[
"0.0825",
"0.3694",
"0.468",
"0.3694",
"0.468",
"0.3833",
"0.0825",
"0.3833"
],
[
"0.505",
"0.3694",
"0.8906",
"0.3694",
"0.8906",
"0.3833",
"0.505",
"0.3833"
],
[
"0.0825",
"0.385",
"0.4673",
"0.385",
"0.4673",
"0.3984",
"0.0825",
"0.3984"
],
[
"0.5034",
"0.385",
"0.8906",
"0.385",
"0.8906",
"0.3984",
"0.5034",
"0.3984"
],
[
"0.0907",
"0.3997",
"0.4673",
"0.3997",
"0.4673",
"0.4136",
"0.0907",
"0.4136"
],
[
"0.5034",
"0.3997",
"0.8896",
"0.3997",
"0.8896",
"0.4136",
"0.5034",
"0.4136"
],
[
"0.08417",
"0.4167",
"0.4673",
"0.4167",
"0.4673",
"0.4268",
"0.08417",
"0.4268"
],
[
"0.505",
"0.414",
"0.8906",
"0.414",
"0.8906",
"0.428",
"0.505",
"0.428"
],
[
"0.0825",
"0.43",
"0.4673",
"0.43",
"0.4673",
"0.443",
"0.0825",
"0.443"
],
[
"0.505",
"0.43",
"0.8896",
"0.43",
"0.8896",
"0.443",
"0.505",
"0.443"
],
[
"0.0809",
"0.4443",
"0.4673",
"0.4443",
"0.4673",
"0.4583",
"0.0809",
"0.4583"
],
[
"0.5034",
"0.443",
"0.8906",
"0.4443",
"0.8906",
"0.459",
"0.5034",
"0.4583"
],
[
"0.0825",
"0.4602",
"0.4673",
"0.4602",
"0.4673",
"0.4734",
"0.0825",
"0.4734"
],
[
"0.505",
"0.461",
"0.5864",
"0.461",
"0.5864",
"0.473",
"0.505",
"0.473"
],
[
"0.0825",
"0.476",
"0.4673",
"0.476",
"0.4673",
"0.4893",
"0.0825",
"0.4893"
],
[
"0.5073",
"0.483",
"0.7534",
"0.4841",
"0.7534",
"0.4988",
"0.5073",
"0.4976"
],
[
"0.0825",
"0.4905",
"0.468",
"0.4905",
"0.468",
"0.504",
"0.0825",
"0.504"
],
[
"0.7715",
"0.4849",
"0.8896",
"0.4849",
"0.8896",
"0.4988",
"0.7715",
"0.4988"
],
[
"0.0809",
"0.505",
"0.468",
"0.505",
"0.468",
"0.5195",
"0.0809",
"0.5195"
],
[
"0.505",
"0.5",
"0.8896",
"0.5",
"0.8896",
"0.513",
"0.505",
"0.513"
],
[
"0.0825",
"0.521",
"0.468",
"0.521",
"0.468",
"0.534",
"0.0825",
"0.534"
],
[
"0.505",
"0.515",
"0.8906",
"0.515",
"0.8906",
"0.529",
"0.505",
"0.529"
],
[
"0.505",
"0.53",
"0.8906",
"0.53",
"0.8906",
"0.5435",
"0.505",
"0.5435"
],
[
"0.08417",
"0.536",
"0.4673",
"0.536",
"0.4673",
"0.55",
"0.08417",
"0.55"
],
[
"0.0809",
"0.5503",
"0.468",
"0.5503",
"0.468",
"0.5645",
"0.0809",
"0.5645"
],
[
"0.505",
"0.5454",
"0.8906",
"0.5454",
"0.8906",
"0.5596",
"0.505",
"0.5596"
],
[
"0.0809",
"0.565",
"0.2443",
"0.565",
"0.2443",
"0.579",
"0.0809",
"0.579"
],
[
"0.505",
"0.56",
"0.888",
"0.56",
"0.888",
"0.5737",
"0.505",
"0.5737"
],
[
"0.0989",
"0.5806",
"0.3694",
"0.5806",
"0.3694",
"0.5947",
"0.0989",
"0.5947"
],
[
"0.505",
"0.5757",
"0.8906",
"0.5757",
"0.8906",
"0.589",
"0.505",
"0.589"
],
[
"0.08417",
"0.5967",
"0.4673",
"0.5967",
"0.4673",
"0.61",
"0.08417",
"0.61"
],
[
"0.505",
"0.5913",
"0.8906",
"0.5913",
"0.8906",
"0.605",
"0.505",
"0.605"
],
[
"0.10046",
"0.61",
"0.312",
"0.61",
"0.312",
"0.6235",
"0.10046",
"0.6235"
],
[
"0.5034",
"0.605",
"0.8906",
"0.604",
"0.8906",
"0.6187",
"0.5034",
"0.619"
],
[
"0.08417",
"0.6235",
"0.4656",
"0.6235",
"0.4656",
"0.637",
"0.08417",
"0.637"
],
[
"0.505",
"0.6206",
"0.8896",
"0.6206",
"0.8896",
"0.634",
"0.505",
"0.634"
],
[
"0.10046",
"0.64",
"0.4673",
"0.6396",
"0.4673",
"0.653",
"0.10046",
"0.654"
],
[
"0.5073",
"0.637",
"0.8896",
"0.637",
"0.8896",
"0.648",
"0.5073",
"0.648"
],
[
"0.10046",
"0.655",
"0.4355",
"0.655",
"0.4355",
"0.6694",
"0.10046",
"0.6694"
],
[
"0.5063",
"0.651",
"0.8896",
"0.651",
"0.8896",
"0.664",
"0.5063",
"0.664"
],
[
"0.0825",
"0.6685",
"0.4673",
"0.6685",
"0.4673",
"0.6816",
"0.0825",
"0.6816"
],
[
"0.505",
"0.6665",
"0.8896",
"0.6665",
"0.8896",
"0.6797",
"0.505",
"0.6797"
],
[
"0.0989",
"0.6846",
"0.4673",
"0.683",
"0.4673",
"0.6963",
"0.0989",
"0.6978"
],
[
"0.505",
"0.681",
"0.8896",
"0.681",
"0.8896",
"0.6943",
"0.505",
"0.6943"
],
[
"0.1013",
"0.6997",
"0.4673",
"0.6997",
"0.4673",
"0.7134",
"0.1013",
"0.7134"
],
[
"0.505",
"0.6963",
"0.8906",
"0.6963",
"0.8906",
"0.7104",
"0.505",
"0.7104"
],
[
"0.1013",
"0.7163",
"0.1683",
"0.7163",
"0.1683",
"0.7266",
"0.1013",
"0.7266"
],
[
"0.505",
"0.711",
"0.8906",
"0.711",
"0.8906",
"0.7246",
"0.505",
"0.7246"
],
[
"0.08417",
"0.729",
"0.468",
"0.729",
"0.468",
"0.742",
"0.08417",
"0.742"
],
[
"0.505",
"0.7266",
"0.888",
"0.7266",
"0.888",
"0.7407",
"0.505",
"0.7407"
],
[
"0.0989",
"0.744",
"0.464",
"0.7437",
"0.464",
"0.757",
"0.0989",
"0.7583"
],
[
"0.5063",
"0.7437",
"0.8896",
"0.7437",
"0.8896",
"0.754",
"0.5063",
"0.754"
],
[
"0.10046",
"0.7583",
"0.4126",
"0.759",
"0.4126",
"0.773",
"0.10046",
"0.7715"
],
[
"0.505",
"0.757",
"0.8906",
"0.757",
"0.8906",
"0.771",
"0.505",
"0.771"
],
[
"0.5024",
"0.771",
"0.7017",
"0.7715",
"0.7017",
"0.786",
"0.5024",
"0.785"
],
[
"0.0801",
"0.787",
"0.2361",
"0.787",
"0.2361",
"0.8003",
"0.0801",
"0.8003"
],
[
"0.5073",
"0.7944",
"0.759",
"0.7954",
"0.759",
"0.81",
"0.5073",
"0.8096"
],
[
"0.7524",
"0.797",
"0.8906",
"0.797",
"0.8906",
"0.81",
"0.7524",
"0.81"
],
[
"0.3506",
"0.811",
"0.4673",
"0.811",
"0.4673",
"0.8247",
"0.3506",
"0.8247"
],
[
"0.505",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.505",
"0.8247"
],
[
"0.3472",
"0.815",
"0.353",
"0.815",
"0.353",
"0.821",
"0.3472",
"0.821"
],
[
"0.08826",
"0.8125",
"0.3506",
"0.8125",
"0.3506",
"0.8228",
"0.08826",
"0.8228"
],
[
"0.0825",
"0.827",
"0.4673",
"0.827",
"0.4673",
"0.8403",
"0.0825",
"0.8403"
],
[
"0.505",
"0.8257",
"0.8906",
"0.8257",
"0.8906",
"0.839",
"0.505",
"0.839"
],
[
"0.0825",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0825",
"0.855"
],
[
"0.505",
"0.8413",
"0.8906",
"0.8413",
"0.8906",
"0.855",
"0.505",
"0.855"
],
[
"0.0801",
"0.855",
"0.4673",
"0.856",
"0.4673",
"0.8706",
"0.0801",
"0.869"
],
[
"0.5034",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.5034",
"0.8706"
],
[
"0.0809",
"0.872",
"0.468",
"0.872",
"0.468",
"0.8853",
"0.0809",
"0.8853"
],
[
"0.5034",
"0.872",
"0.8906",
"0.872",
"0.8906",
"0.8853",
"0.5034",
"0.8853"
],
[
"0.0809",
"0.886",
"0.4697",
"0.8853",
"0.4697",
"0.8994",
"0.0809",
"0.901"
],
[
"0.5024",
"0.8853",
"0.8906",
"0.886",
"0.8906",
"0.901",
"0.5024",
"0.8994"
],
[
"0.4805",
"0.925",
"0.4927",
"0.925",
"0.4927",
"0.9375",
"0.4805",
"0.9375"
]
] | [
"NeuralSurfaceReconstruction",
"Neural Material&LightingDistillation",
"Physics-based Inverse Rendering",
"SDF",
"",
"吧",
"",
"",
"",
"Extract",
"",
"Refine",
"",
"Materials",
"Materials",
"",
"",
"",
"",
"Mesh",
"Mesh",
"Light",
"Light",
"Radiance Field",
"Targe",
"1",
"Images",
"Distill",
"Coarse Renderin",
"Optimize",
"Optimize",
"m",
"rentiable Rendering",
"T",
"",
"Figure 2: Our pipeline for joint shape, material, and lighting estimation.",
"在本文中，我们提出了一种高效，准确的逆染流",
"[9]和曼哈顿对齐E[11]来提高质量。这些方法使用大",
"水线，具有神经重建和PBIR方法的优点。我们的管道",
"型MLP作为其体积表示，但是每个场景的优化速度",
"尝试根据在静态但任意照明下捕获的对象的多个图像",
"很慢（许多小时）。最近的进展[31，40，22，8]通过使",
"来估计几何形状，空间变化的材料反射率和HDR环境",
"用显式网格显示出极大的优化加速而不损失质量。不",
"图。如图所示。2，我们的管道由三个阶段组成。在第",
"幸的是，使用显式体积来建模SDF会导致颠簸和噜杂",
"一阶段，我们提出了一种用于快速神经SDF和辐射场",
"的表面【41]。SfM稀疏点云E[43，9]或单眼深度/正",
"重建的基于混合神经体积的方法，该方法实现了最先",
"常E41]的监督可能会减轻难度，但超出了我们的范",
"进的儿何精度和效率。在下一阶段，基于重建的几何形",
"围。我们为显式SDF网格优化提出了简单有效的正则",
"状和辐射场，我们设计了一种有效的优化方法，通过拟",
"化损失，在不使用外部先验的情况下实现了快速和高",
"合表面光场来提炼材料和照明。我们的方法依赖于辐",
"质量的表面",
"射场来处理可见性和间接照明，但与最近的一些工作",
"Material and lighting estimation.",
"相比，避免了昂贵的体积光线跟踪到显着的计算成本。",
"儿种神经重建方",
"最后，我们使用先进的PBIR框架【42，13]来共同完善",
"法「44，45，4，5，23，12，46]采用与我们相同的设置，可",
"几何、材料和照明。请注意，我们的PBIR框架以物理",
"以从多个图像中同时重建几何，材料和照明。较早的工",
"作45】直接从固定的几何形状优化了低分辨率的环境",
"上正确和无偏的方式对复杂的光传输效果（如可见性，",
"遮挡，柔和阴影和相互反射）进行建模，同时仍然比最",
"图和材料，从而导致了由该问题的高度不适定性引起的",
"近的逆谊染方法快得多",
"噜杂的照明重建。最近的方法使用球形高斯「44，4，46",
"具体而言，我们的贡献包括以下内容：",
"或预滤波近似[5，231的混合物对照明进行建模，并使",
": A hybrid volume representation for fast and accu-",
"用球形高斯瓣【44]或低维潜在代码【4，5，46]约束材料",
"rate geometry reconstruction.",
"反射，以提高重建质量。由于基于MLP的神经染的",
". A efficient optimization scheme to distill high-",
"高计算成本，它们中的大多数仅考虑无遮挡的直接照明",
"quality initial material and lighting estimation from",
"44，4，23或使用固定阴影贴图45，5。Nvdiffrecmc",
"the reconstructed geometry and radiance field.",
"[12】使用带有去噪器的可微分光线跟踪器对阴影进行",
": An advanced PBIR framework that jointly opti-",
"建模，该去噪器可显着减少样本数量，但仍无法对相互",
"mizes materials, lighting and geometry with visibil-",
"反射进行建模。MII「46]训练另一个网络来预测来自辐",
"ity and interreflection handled in a physically unbi-",
"射场的可见性和相互反射，这需要几个小时。相反，我",
"ased way",
"们的混合，高度优化的神经SDF框架使我们能够在不",
": A end-to-end pipeline that achieves state-of-the-art",
"到10分钟的时间内有效地提取几何形状，材料，照明",
"geometry, material and lighting estimation that en.",
"可见性和相互反射。我们先进的PBIR框架首次充许",
"ables realistic view synthesis and relighting",
"以物理无偏的方式建模的直接和间接照明对照明，材",
"料和几何形状进行整体优化",
"2. Related Works",
"Physics-based inverse rendering.",
"与基于神经的方法",
"静态场景的基于",
"不同，最近的几部作品[19，12，71利用图形界开发的",
"",
"Volumetric surfacereconstruction.",
"体积的表面重建的最新进展显示了高质量和可靠的结",
"经典图像公式化模型，并尝试使用基于梯度的优化来",
"果。NeuS[34]和VolSDF[39]用有符号距离值代替",
"反转此过程，并使用基于物理的可微分谊染器计算梯",
"NeRF「21]的密度预测，提出了一种无偏和遮挡感知",
"度。这些方法可以在优化过程中处理复杂的光传输效",
"的体绘制，取得了可喜的结果。后续工作通过引入正",
"应，但容易出现局部极小值和过拟合。在我们的工作",
"则化损失来促进表面平滑度E[43]，多视图一致性E",
"中，我们不是从头开始优化，而是将PBIR作为改进",
"2"
] | {
"arxiv_id": "2304.13445",
"language": "zh",
"page_idx": 1
} |
||
2304.13445-zh-2 | 2304.13445-zh-2.jpg | 阶段来提高鲁棒性
## 3. Our Method
在固定的未知照明(具有已知的相机参数)下提供不透明物体的多视图图像，我们的技术重建了物体的形状和反射率以及照明条件
如 Fig. 2 所示，我们的管道由三个主要阶段组成。第一阶段 (Sec. 3.1 ) 是快速而精确的表面重建步骤，将直接 SDF 网格优化引入 NeuS $\mathbb{E}[34]$。与该表面相关的是过度拟合的辐射场，该辐射场不能完全模拟物体的表面反射率。我们的第二阶段 (Sec. 3.2 ) 是一种有效的神经蒸馏方法，可将辐射场转换为基于物理的反射率$\boxed{\mathrm{F}[6]}$ 和照明模型。最后，我们的第三阶段 (Sec. 3.3 )利用基于物理的逆渲染(PBIR)来进一步完善前两个阶段重建的对象几何形状和反射率。此阶段利用基于物理的可区分渲染来捕获全局照明 (GI) 效果，例如柔和阴影和相互反射
# 3.1. Neural Surface Reconstruction
将固定照明(并具有已知的相机参数)下的不透明物体的多个图像作为输入，我们方法的第一阶段会对物体表面进行详细的重建。为此，我们优化了物体表面和最能描述输人图像的输出辐射场。。在此阶段，为了提高效率和鲁棒性，我们将物体表面表示为有符号距离场 (SDF) $S(\boldsymbol{x})$的零水平集$\{ x\in \mathbb{R} ^3\mid S( \boldsymbol{x}) = 0\}$ , 并将辐射场$L_\mathrm{o} ( \boldsymbol{x}, - \boldsymbol{v})$ $( \dot {\text{对 于 任 何 位 置 }x}\in \mathbb{R} ^3$和观察方向$v\in\mathbb{S}^2$ )表示为非基于物理的通用函数逼近器
Unbiased volume rendering with SDF. 为了计算像素的颜色，我们首先沿着相应的相机射线对$N$查询点$\{\boldsymbol{x}_{i}=\boldsymbol{o}+t_{i}\boldsymbol{v}\}_{i=1}^{N}($带有$0<t_1<t_2<\ldots<t_{N})$进行采样，该射线起源于具有观察方向$v$的相机位置$o$ 。然后，我们查询点符号距离$S(x_i)$和辐射度$L_\mathrm{o}(\boldsymbol{x}_i,\boldsymbol{v})$ 的场景表示。在 NeuS [34] 的无偏渲染之后，我们通过
$$\alpha_i=\max\left(0,\:\frac{\sigma(S(\boldsymbol{x}_i))-\sigma(S(\boldsymbol{x}_{i+1}))}{\sigma(S(\boldsymbol{x}_i))}\right)\:,$$
(1,
将所有$1\leq i\leq N$的查询符号距离激活为 alpha, 其中$\sigma$表示 Sigmoid 函数。然后，通过查询的 alpha 和辐射值
$$C=\sum_{i=1}^NT_i\:\alpha_i\:L_\text{o}(\boldsymbol{x}_i,-\boldsymbol{v})\quad\text{where}\:T_i=\prod_{j=1}^{i-1}(1-\alpha_j)\:.$$
的 alpha 混合来计算像素颜色$C$
Voxel-based scene representation. NeuS 使用大型MLP 对 SDF $S$和辐射场$L_\mathrm{o}$进行建模。不幸的是，由于此 MLP 的查询成本很高，因此 NeuS 的优化过程可能非常耗时
为了获得更好的性能，我们通过设置
$$\begin{aligned}S(\boldsymbol{x})&=\mathrm{interp}(\boldsymbol{x},\:V^{(\mathrm{sdf})})\:,\\L_{\mathrm{o}}(\boldsymbol{x},-\boldsymbol{v})&=\mathrm{MLP}\left(\mathrm{interp}(\boldsymbol{x},\:V^{(\mathrm{feat})}),\:\boldsymbol{v}\right)\:,\end{aligned}$$
(3)
来适应 DVGO $\mathbb{E}[31]$中基于密集网格的场景表示，其中 interp() 表示三线性插值，$V^\mathrm{(sdf)}$是密集的 SDF 网格，$V^{(\mathrm{feat})}$是密集的特征网格，而 XMATHISLLP 激活层
在实践中，我们使用两组$V^\mathrm{(sdf)}$和$V^\mathrm{(feat)}$网格对前景进行建模 (i。e。,感兴趣的对象)和具有不同分辨率的背景。有关更多详细信息，请参阅补充资料
Adaptive Huber loss. 我们引人了 Huber 损失，以减少镜面高光的影响，这些高光可能会在重建表面中引起颠簸的伪影：
$$\mathcal{L}_{\text{photo}}=\sum_r\begin{cases}(\mathcal{I}[r]-C[r])^2,&(|\mathcal{I}[r]-C[r]|\le t)\\2t\:|\mathcal{I}[r]-C[r]|-t^2,&\text{(otherwise)}\end{cases}$$
(4)
其中：$\mathcal{I}[r]$和$C[r]$分别表示$r$的颜色目标和渲染图像
的第 3 个像素
直观上，$\mathcal{L}_\mathrm{photo}$的梯度的无穷大范数被 2$t$钳位，以防止明亮的像素 (e.g.,那些表现出镜面高光的像素) 主导优化。实际上，我们不是调整常数$t$ ,而是在每次迭代中通过$|\mathcal{I}[r]-C[r]|$中值的运行平均值自适应地更新$t$ ,以钳制大约一半像素的梯度
Laplacian regularization. 为了进一步提高我们方法的鲁棒性并减少潜在的伪像，我们使用拉普拉斯损失对 SDF 网格$V^{(\mathrm{sdf})}$进行正则化：
$$\mathcal{L}_{\mathrm{lap}}=\sum_{u}\left(\sum_{u'\in N(u)}\left(V^{\mathrm{(sdf)}}[u']-V^{\mathrm{(sdf)}}[u]\right)\right)^{2}\:,$$
(5)
其中$V^{(\mathrm{sdf})}[u]$表示存储在网格点$u$处的 SDF 值，$N(u)$
表示 MAXX 的六个直接邻居
Training. 继 DVGO $\mathbb{E}[31]$之后，我们使用渐进式网格缩放来提高效率和更连贯的结果。我们的实验还表明，使用每点 RGB 损失$\mathcal{L}_\mathrm{pp\_rgb}$可以提高收敛速度和质量：
(6,
$$\mathcal{L}_{\mathrm{pp_rgb}}=\sum_{r,i}T_{i}^{(r)}\alpha_{i}^{(r)}\left|L_{\mathrm{o}}\left(\boldsymbol{x}_{i}^{(r)},-\boldsymbol{v}^{(r)}\right)-\mathcal{I}[r]\right|\:,$$
其中$\alpha_i^{(r)}$和$T_i^{(r)}$是使用 Eqs.(1)and (2)为每个像素$r$计算的。总之，为了在 Eq.(3)中训练我们的 SDF 和特征网格$V^{(\mathrm{sdf})}$和$V^{(\mathrm{feat})}$,我们最小化目标：
$$\mathcal{L}_{\mathrm{surf}}=\mathcal{L}_{\mathrm{photo}}+w_{\mathrm{lap}}\mathcal{L}_{\mathrm{lap}}+w_{\mathrm{pp_rgb}}\mathcal{L}_{\mathrm{pp_rgb}}\:.$$
(7)
Postprocessing. 可选地，我们的表面重建阶段为每个输人图像生成一个抗锯齿遮罩，该遮罩使用网格光栅化和 alpha 遮罩将对象与背景分离$\mathbb{E}[28]$ 。生成的遮罩可用于我们的基于物理的逆渲染阶段 (Sec. 3.3 ), 以促进对象形状的细化
3.2. Neural material and lighting distillation
从表面重建阶段 (Sec. 3.1 ) 提供了优化的物体表面和出射辐射场，我们管道的第二阶段的目标是获得表面反射率和光照条件的初始估计。为此，我们利用辐射场$L_{\mathrm{o}}$作为教师模型，将学习到的表面颜色提取到基于
3 | [
[
"0.0809",
"0.09216",
"0.2133",
"0.09216",
"0.2133",
"0.1061",
"0.0809",
"0.1061"
],
[
"0.5107",
"0.09216",
"0.8906",
"0.09216",
"0.8906",
"0.1061",
"0.5107",
"0.1061"
],
[
"0.5034",
"0.1048",
"0.8906",
"0.1048",
"0.8906",
"0.1225",
"0.5034",
"0.1225"
],
[
"0.0768",
"0.1225",
"0.2157",
"0.1218",
"0.2157",
"0.1364",
"0.0768",
"0.137"
],
[
"0.5034",
"0.1218",
"0.892",
"0.1218",
"0.892",
"0.1364",
"0.5034",
"0.1364"
],
[
"0.5034",
"0.1395",
"0.5376",
"0.1395",
"0.5376",
"0.1509",
"0.5034",
"0.1509"
],
[
"0.10046",
"0.1483",
"0.468",
"0.1483",
"0.468",
"0.1622",
"0.10046",
"0.1622"
],
[
"0.5195",
"0.1497",
"0.8906",
"0.1509",
"0.8906",
"0.1686",
"0.5195",
"0.1674"
],
[
"0.0825",
"0.1635",
"0.468",
"0.1635",
"0.468",
"0.1768",
"0.0825",
"0.1768"
],
[
"0.5063",
"0.1692",
"0.8896",
"0.1692",
"0.8896",
"0.1799",
"0.5063",
"0.1799"
],
[
"0.0809",
"0.1787",
"0.2754",
"0.1787",
"0.2754",
"0.1925",
"0.0809",
"0.1925"
],
[
"0.505",
"0.1831",
"0.83",
"0.1831",
"0.83",
"0.1963",
"0.505",
"0.1963"
],
[
"0.1013",
"0.1932",
"0.4673",
"0.1932",
"0.4673",
"0.207",
"0.1013",
"0.207"
],
[
"0.509",
"0.202",
"0.8896",
"0.2026",
"0.8896",
"0.2172",
"0.509",
"0.2166"
],
[
"0.0948",
"0.2108",
"0.4673",
"0.2108",
"0.4673",
"0.2216",
"0.0948",
"0.2216"
],
[
"0.5034",
"0.2172",
"0.888",
"0.2172",
"0.888",
"0.2311",
"0.5034",
"0.2311"
],
[
"0.0825",
"0.2247",
"0.468",
"0.2247",
"0.468",
"0.238",
"0.0825",
"0.238"
],
[
"0.5034",
"0.2311",
"0.626",
"0.2311",
"0.626",
"0.2443",
"0.5034",
"0.2443"
],
[
"0.0825",
"0.2412",
"0.4656",
"0.2412",
"0.4656",
"0.252",
"0.0825",
"0.252"
],
[
"0.0809",
"0.2551",
"0.468",
"0.2551",
"0.468",
"0.2683",
"0.0809",
"0.2683"
],
[
"0.6055",
"0.2568",
"0.876",
"0.2568",
"0.876",
"0.2708",
"0.6055",
"0.2708"
],
[
"0.5156",
"0.2651",
"0.6006",
"0.262",
"0.601",
"0.2766",
"0.517",
"0.2798"
],
[
"0.0825",
"0.2715",
"0.4656",
"0.2715",
"0.4656",
"0.2822",
"0.0825",
"0.2822"
],
[
"0.5923",
"0.2715",
"0.834",
"0.2708",
"0.834",
"0.2886",
"0.5923",
"0.289"
],
[
"0.0809",
"0.2854",
"0.468",
"0.2854",
"0.468",
"0.2986",
"0.0809",
"0.2986"
],
[
"0.584",
"0.2822",
"0.597",
"0.2822",
"0.597",
"0.286",
"0.584",
"0.286"
],
[
"0.871",
"0.289",
"0.8906",
"0.289",
"0.8906",
"0.3018",
"0.871",
"0.3018"
],
[
"0.0825",
"0.2998",
"0.468",
"0.2998",
"0.468",
"0.3132",
"0.0825",
"0.3132"
],
[
"0.509",
"0.3037",
"0.8906",
"0.3037",
"0.8906",
"0.3176",
"0.509",
"0.3176"
],
[
"0.08417",
"0.3164",
"0.4673",
"0.3164",
"0.4673",
"0.327",
"0.08417",
"0.327"
],
[
"0.505",
"0.3193",
"0.6045",
"0.3193",
"0.6045",
"0.3308",
"0.505",
"0.3308"
],
[
"0.08417",
"0.33",
"0.468",
"0.33",
"0.468",
"0.3435",
"0.08417",
"0.3435"
],
[
"0.524",
"0.334",
"0.8906",
"0.334",
"0.8906",
"0.3484",
"0.524",
"0.3484"
],
[
"0.0809",
"0.3452",
"0.2279",
"0.3452",
"0.2279",
"0.3591",
"0.0809",
"0.3591"
],
[
"0.5034",
"0.3499",
"0.8906",
"0.3499",
"0.8906",
"0.363",
"0.5034",
"0.363"
],
[
"0.505",
"0.363",
"0.8906",
"0.363",
"0.8906",
"0.3774",
"0.505",
"0.3774"
],
[
"0.0809",
"0.3718",
"0.3545",
"0.3718",
"0.3545",
"0.385",
"0.0809",
"0.385"
],
[
"0.5034",
"0.379",
"0.8896",
"0.379",
"0.8896",
"0.3926",
"0.5034",
"0.3926"
],
[
"0.10046",
"0.3958",
"0.468",
"0.3958",
"0.468",
"0.4092",
"0.10046",
"0.4092"
],
[
"0.5034",
"0.3933",
"0.749",
"0.3933",
"0.749",
"0.4072",
"0.5034",
"0.4072"
],
[
"0.0809",
"0.4104",
"0.4673",
"0.4104",
"0.4673",
"0.4236",
"0.0809",
"0.4236"
],
[
"0.5063",
"0.4124",
"0.8906",
"0.4136",
"0.8906",
"0.428",
"0.5063",
"0.4268"
],
[
"0.0809",
"0.4255",
"0.468",
"0.4255",
"0.468",
"0.4395",
"0.0809",
"0.4395"
],
[
"0.505",
"0.428",
"0.8896",
"0.428",
"0.8896",
"0.4414",
"0.505",
"0.4414"
],
[
"0.0809",
"0.44",
"0.4656",
"0.44",
"0.4656",
"0.4539",
"0.0809",
"0.4539"
],
[
"0.501",
"0.4382",
"0.738",
"0.4395",
"0.738",
"0.457",
"0.501",
"0.4558"
],
[
"0.0801",
"0.455",
"0.468",
"0.4539",
"0.468",
"0.4673",
"0.0801",
"0.4685"
],
[
"0.0809",
"0.4685",
"0.4592",
"0.4697",
"0.4592",
"0.4841",
"0.0809",
"0.483"
],
[
"0.455",
"0.476",
"0.464",
"0.476",
"0.464",
"0.481",
"0.455",
"0.481"
],
[
"0.5337",
"0.476",
"0.584",
"0.476",
"0.584",
"0.4873",
"0.5337",
"0.4873"
],
[
"0.6216",
"0.4702",
"0.7983",
"0.4685",
"0.7983",
"0.4893",
"0.6216",
"0.4912"
],
[
"0.584",
"0.4766",
"0.5923",
"0.483",
"0.585",
"0.4873",
"0.5786",
"0.4817"
],
[
"0.8613",
"0.4766",
"0.884",
"0.4766",
"0.884",
"0.4849",
"0.8613",
"0.4849"
],
[
"0.0809",
"0.4849",
"0.468",
"0.4849",
"0.468",
"0.4988",
"0.0809",
"0.4988"
],
[
"0.6113",
"0.4849",
"0.66",
"0.4893",
"0.6587",
"0.5005",
"0.6094",
"0.4968"
],
[
"0.0809",
"0.5005",
"0.4468",
"0.5005",
"0.4468",
"0.5146",
"0.0809",
"0.5146"
],
[
"0.5073",
"0.5103",
"0.8896",
"0.512",
"0.8896",
"0.5264",
"0.5073",
"0.5244"
],
[
"0.0866",
"0.5215",
"0.4673",
"0.5215",
"0.4673",
"0.535",
"0.0866",
"0.535"
],
[
"0.5024",
"0.526",
"0.714",
"0.5264",
"0.714",
"0.5415",
"0.5024",
"0.5405"
],
[
"0.0825",
"0.535",
"0.4673",
"0.535",
"0.4673",
"0.549",
"0.0825",
"0.549"
],
[
"0.511",
"0.549",
"0.6006",
"0.549",
"0.6006",
"0.56",
"0.511",
"0.56"
],
[
"0.5933",
"0.547",
"0.8906",
"0.547",
"0.8906",
"0.56",
"0.5933",
"0.56"
],
[
"0.2059",
"0.55",
"0.3652",
"0.552",
"0.3652",
"0.5625",
"0.2059",
"0.56"
],
[
"0.08417",
"0.552",
"0.1978",
"0.55",
"0.1978",
"0.56",
"0.08417",
"0.5625"
],
[
"0.3896",
"0.553",
"0.4673",
"0.55",
"0.4673",
"0.561",
"0.3896",
"0.5645"
],
[
"0.0825",
"0.5625",
"0.468",
"0.5625",
"0.468",
"0.5757",
"0.0825",
"0.5757"
],
[
"0.505",
"0.5625",
"0.8906",
"0.5625",
"0.8906",
"0.5757",
"0.505",
"0.5757"
],
[
"0.0809",
"0.577",
"0.4673",
"0.577",
"0.4673",
"0.5903",
"0.0809",
"0.5903"
],
[
"0.5024",
"0.5737",
"0.892",
"0.5747",
"0.892",
"0.5923",
"0.5024",
"0.5913"
],
[
"0.0809",
"0.5903",
"0.4673",
"0.5903",
"0.4673",
"0.604",
"0.0809",
"0.604"
],
[
"0.505",
"0.5903",
"0.5596",
"0.5903",
"0.5596",
"0.6016",
"0.505",
"0.6016"
],
[
"0.0809",
"0.605",
"0.10046",
"0.605",
"0.10046",
"0.617",
"0.0809",
"0.617"
],
[
"0.2141",
"0.617",
"0.4102",
"0.616",
"0.4102",
"0.6353",
"0.2141",
"0.636"
],
[
"0.528",
"0.613",
"0.8896",
"0.613",
"0.8896",
"0.631",
"0.528",
"0.631"
],
[
"0.1381",
"0.625",
"0.2051",
"0.6226",
"0.2051",
"0.637",
"0.1381",
"0.6396"
],
[
"0.4019",
"0.625",
"0.468",
"0.625",
"0.468",
"0.6396",
"0.4019",
"0.6396"
],
[
"0.6045",
"0.629",
"0.6177",
"0.634",
"0.613",
"0.6416",
"0.5986",
"0.636"
],
[
"0.282",
"0.6353",
"0.3406",
"0.6353",
"0.3406",
"0.6465",
"0.282",
"0.6465"
],
[
"0.08826",
"0.654",
"0.4673",
"0.654",
"0.4673",
"0.664",
"0.08826",
"0.664"
],
[
"0.5156",
"0.653",
"0.8906",
"0.653",
"0.8906",
"0.6704",
"0.5156",
"0.6704"
],
[
"0.0825",
"0.6665",
"0.4673",
"0.6665",
"0.4673",
"0.6797",
"0.0825",
"0.6797"
],
[
"0.5024",
"0.6665",
"0.8906",
"0.6675",
"0.8906",
"0.685",
"0.5024",
"0.6846"
],
[
"0.0825",
"0.681",
"0.1316",
"0.681",
"0.1316",
"0.6924",
"0.0825",
"0.6924"
],
[
"0.5024",
"0.6797",
"0.8506",
"0.6797",
"0.8506",
"0.7007",
"0.5024",
"0.7007"
],
[
"0.1323",
"0.705",
"0.1462",
"0.705",
"0.1462",
"0.7134",
"0.1323",
"0.7134"
],
[
"0.3513",
"0.7056",
"0.3694",
"0.7056",
"0.3694",
"0.714",
"0.3513",
"0.714"
],
[
"0.546",
"0.707",
"0.892",
"0.709",
"0.892",
"0.726",
"0.546",
"0.7236"
],
[
"0.1259",
"0.7163",
"0.3472",
"0.7163",
"0.3472",
"0.731",
"0.1259",
"0.731"
],
[
"0.3472",
"0.7163",
"0.4673",
"0.7163",
"0.4673",
"0.731",
"0.3472",
"0.731"
],
[
"0.1274",
"0.729",
"0.1503",
"0.729",
"0.1503",
"0.741",
"0.1274",
"0.741"
],
[
"0.0972",
"0.7173",
"0.1323",
"0.7173",
"0.1323",
"0.729",
"0.0972",
"0.729"
],
[
"0.3489",
"0.731",
"0.3708",
"0.731",
"0.3708",
"0.7437",
"0.3489",
"0.7437"
],
[
"0.5107",
"0.7344",
"0.6274",
"0.7344",
"0.6274",
"0.7476",
"0.5107",
"0.7476"
],
[
"0.6177",
"0.733",
"0.8896",
"0.733",
"0.8896",
"0.747",
"0.6177",
"0.747"
],
[
"0.505",
"0.7485",
"0.8906",
"0.7485",
"0.8906",
"0.762",
"0.505",
"0.762"
],
[
"0.0866",
"0.756",
"0.3186",
"0.756",
"0.3186",
"0.7695",
"0.0866",
"0.7695"
],
[
"0.505",
"0.7646",
"0.8906",
"0.7646",
"0.8906",
"0.778",
"0.505",
"0.778"
],
[
"0.08496",
"0.776",
"0.3308",
"0.7764",
"0.3308",
"0.7915",
"0.08496",
"0.7905"
],
[
"0.3513",
"0.7764",
"0.468",
"0.7764",
"0.468",
"0.7905",
"0.3513",
"0.7905"
],
[
"0.5034",
"0.779",
"0.8906",
"0.779",
"0.8906",
"0.7925",
"0.5034",
"0.7925"
],
[
"0.0801",
"0.7905",
"0.4673",
"0.7925",
"0.4673",
"0.8066",
"0.0801",
"0.8047"
],
[
"0.5034",
"0.7944",
"0.652",
"0.7944",
"0.652",
"0.808",
"0.5034",
"0.808"
],
[
"0.08417",
"0.8066",
"0.4673",
"0.8066",
"0.4673",
"0.821",
"0.08417",
"0.821"
],
[
"0.0825",
"0.8237",
"0.1642",
"0.8237",
"0.1642",
"0.8354",
"0.0825",
"0.8354"
],
[
"0.5034",
"0.8184",
"0.846",
"0.8184",
"0.846",
"0.832",
"0.5034",
"0.832"
],
[
"0.0989",
"0.837",
"0.361",
"0.836",
"0.361",
"0.851",
"0.0989",
"0.8516"
],
[
"0.524",
"0.8413",
"0.8906",
"0.8413",
"0.8906",
"0.855",
"0.524",
"0.855"
],
[
"0.505",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.505",
"0.8706"
],
[
"0.1666",
"0.863",
"0.3367",
"0.863",
"0.3367",
"0.8774",
"0.1666",
"0.8774"
],
[
"0.505",
"0.8726",
"0.8906",
"0.8726",
"0.8906",
"0.886",
"0.505",
"0.886"
],
[
"0.4478",
"0.879",
"0.4656",
"0.879",
"0.4656",
"0.888",
"0.4478",
"0.888"
],
[
"0.1316",
"0.8853",
"0.4182",
"0.8853",
"0.4182",
"0.8994",
"0.1316",
"0.8994"
],
[
"0.505",
"0.887",
"0.8896",
"0.887",
"0.8896",
"0.901",
"0.505",
"0.901"
],
[
"0.4795",
"0.9253",
"0.4927",
"0.9253",
"0.4927",
"0.9375",
"0.4795",
"0.9375"
]
] | [
"阶段来提高鲁棒性",
"来适应DVGOE311中基于密集网格的场景表示，其",
"中interp(）表示三线性插值，V(sdf）是密集的SDF网",
"3. Our Method",
"格，V(feat）是密集的特征网格，而XMATHISLLP激",
"活层",
"在固定的未知照明（具有已知的相机参数）下提供",
"在实践中，我们使用两组V(sdf）和V(feat）网格对前",
"不透明物体的多视图图像，我们的技术重建了物体的",
"景进行建模（i。e。，感兴趣的对象）和具有不同分辨率",
"形状和反射率以及照明条件",
"的背景。有关更多详细信息，请参阅补充资料",
"如Fig.2所示，我们的管道由三个主要阶段组成。第",
"AdaptiveHuberloss．我们引人了Huber损失，以",
"-阶段（Sec.3.1）是快速而精确的表面重建步骤，将",
"减少镜面高光的影响，这些高光可能会在重建表面中",
"直接SDF网格优化引入NeuSE[34l。与该表面相关",
"引起颠簸的伪影：",
"的是过度拟合的辐射场，该辐射场不能完全模拟物体",
"的表面反射率。我们的第二阶段（Sec.3.2）是一种有效",
"[([r] -C[r]),(([r] -C[r]l ≤t)",
"Cphoto =",
"的神经蒸馏方法，可将辐射场转换为基于物理的反射",
"{2t [Z[r] - C[r] - t2, (otherwise)",
"率E[6]和照明模型。最后，我们的第三阶段（Sec.3.3",
"",
"(4)",
"）利用基于物理的逆渣染（PBIR）来进一步完善前两",
"其中：工[r]和C[rl分别表示r的颜色目标和谊染图像",
"个阶段重建的对象儿何形状和反射率。此阶段利用基",
"的第3个像素",
"于物理的可区分染来捕获全局照明（GI）效果，例如",
"直观上，Cphoto的梯度的无穷大范数被2t钳位，以",
"柔和阴影和相互反射",
"防止明亮的像素（e.g.，那些表现出镜面高光的像素）",
"主导优化。实际上，我们不是调整常数t，而是在每次",
"3.1. Neural Surface Reconstruction",
"送代中通过[[r]-C[rl|中值的运行平均值自适应地更",
"将固定照明（并具有已知的相机参数）下的不透明",
"新t，以钳制大约一半像素的梯度",
"物体的多个图像作为输入，我们方法的第一阶段会对",
"Laplacianregularization：，为了进一步提高我们方法",
"物体表面进行详细的重建。为此，我们优化了物体表面",
"的鲁棒性并减少潜在的伪像，我们使用拉普拉斯损失",
"和最能描述输入图像的输出辐射场。。在此阶段，为了",
"对SDF网格V(sdt)进行正则化：",
"提高效率和鲁棒性，我们将物体表面表示为有符号距",
"离场（SDF）S(α）的零水平集(αR3「S(α)=0)",
"",
"Clap=",
"Z (v(sd)[u'] -V(ad)[",
"7",
"75",
"并将辐射场L。(a,-u）（对于任何位置aEIR3和观察",
"UENu",
"方向ES2）表示为非基于物理的通用函数逼近器",
"其中V(sdf)[ul表示存储在网格点u处的SDF值，N(u)",
"Unbiased volumerenderingwithSDF.为了计算像",
"表示MAXX的六个直接邻居",
"素的颜色，我们首先沿着相应的相机射线对N查询点",
"Training.",
"继DVGOE[31]之后：我们使用渐进式",
"（带有0<t1<t2<",
"Ri=o+tulN",
"<tN）进",
"行采样，该射线起源于具有观察方向的相机位置o。",
"网格缩放来提高效率和更连贯的结果。我们的实验还",
"然后，我们查询点符号距离S(ai）和辐射度L。(ai，u)",
"表明，使用每点RGB损失Cpp_rgb可以提高收敛速度",
"的场景表示。在NeuS[34]的无偏谊染之后，我们通",
"和质量：",
"过",
"0. (S(a:)-o(S(ai+1))",
"Cpp_rgb =T()α() |L(α(n), -u()) -[, (6)",
"Qi = max",
"， (1)",
"1",
"a(s(a)",
"将所有1<i<N的香询符号距离激活为alpha，其",
"其中α）和T(）是使用Eqs.（1)and(2)为每个像",
"中表示Sigmoid函数。然后，通过查询的alpha和",
"素r计算的。总之，为了在Eq.（3）中训练我们的SDF",
"辐射值",
"和特征网格V(sdf)和V(feat），我们最小化目标：",
"N",
"-1",
"Csurf = Cphoto+Wlap Clap+Wpp_rgb Cpp_rgb .(7)",
"TQi Lo(ai,-)where T =",
"(1-Q) (2)",
"二",
"C=",
"j=1",
"Postprocessing.",
"可选地，我们的表面重建阶段为每",
"个输入图像生成一个抗锯齿遮罩，该遮罩使用网格光",
"的alpha混合来计算像素颜色C",
"栅化和alpha遮罩将对象与背景分离E[28]。生成的遮",
"Voxel-based scene representation.",
"NeuS使用大型",
"罩可用于我们的基于物理的逆疽染阶段（Sec.3.3），以",
"MLP对SDFS和辐射场L。进行建模。不幸的是，由",
"促进对象形状的细化",
"于此MLP的查询成本很高，因此NeuS的优化过程可",
"能非常耗时",
"3.2. Neural material and lighting distillation",
"为了获得更好的性能，我们通过设置",
"从表面重建阶段（Sec.3.1）提供了优化的物体表面",
"和出射辐射场，我们管道的第二阶段的目标是获得表",
"S(α) =interp(a, V(sdf) ,",
"面反射率和光照条件的初始估计。为此，我们利用辐射",
"(3)",
"Lo(,-u) = MLP (interp(α, V(feat), ),",
"场L。作为教师模型，将学习到的表面颜色提取到基于",
"3"
] | {
"arxiv_id": "2304.13445",
"language": "zh",
"page_idx": 2
} |
||
1704.02497-zh-0 | 1704.02497-zh-0.jpg | 关于语义变化的线性：通过动态图模型研究含义变化
Steffen Eger
Ubiquitous Knowledge Processing Lab
Alexander Mehler Text Technology Lab
Department of Computer Science Department of Computer Science
Technische Universität Darmstadt Goethe-Universität Frankfurt am Main
## $\operatorname{Abstract}$
We consider two graph models of semantic change. The first is a timeseries model that relates embedding vectors from one time period to embedding vectors of previous time periods. In the second, we construct one graph for each word: nodes in this graph correspond to time points and edge weights to the similarity of the word's meaning across two time points. We apply our two models to corpora across three different languages. We find that semantic change is linear in two senses. Firstly, today's embedding vectors ( =meaning) of words can be derived as linear combinations of embedding vectors of their neighbors in previous time periods. Secondly, selfsimilarity of words decays linearly in time. We consider both findings as new laws/hypotheses of semantic change.
## 1 Introduction
含义不统一，既不跨空间，也不跨时间。在整个空间中，不同的语言倾向于对相应的术语表现出不同的多义关联 (Eger et al., 2015; Kulkarni et al., 2015b) 。随着时间的推移，已经记录了几个著名的英语意义变化的例子。例如，在 1970 年代，“同性恋”一词的含义已从20 th 世纪初的“快乐”的形容词含义转变为“同性恋 (Kulkarni et al.,2015a)”的当前含义。。同样，技术进步也导致了术语的语义扩展，例如传输，鼠标或苹果
在这项工作中，我们考虑了语义变化的两个
图模型。我们的第一个模型是一个动态模型，分析将一个单词在一个时间段内的语义与该
因为其基础范式是 (时间)系列图。系列图中
的每个节点对应一个单词，与该单词相关联的是语义 embedding 向量。然后，我们询问如何根据先前时间段中相邻单词的嵌入向量来预测一个时间段(图)中的嵌入向量。特别是， 我们假设存在一种线性函数关系，该关系将单词的今天含义与过去的邻居含义耦合在一起。在估计该模型的系数时，我们发现线性形式看起来确实非常合理。然后，这种功能形式使我们能够解决其他问题，例如单词之间的负面关系-表示随时间的语义差异-以及对未来的预测。我们称我们的第二个图模型为时间索引自相似图。在这些图中，每个节点对应一个时间点，两个时间点之间的链接表示所考虑的两个时间点之间特定单词的语义相似性。对这些图的分析表明，大多数单词都遵循线性语义“衰减”定律：语义自相似性随时间线性降低
在我们的工作中，我们通过从上下文预测神经网络体系结构派生的词嵌入来捕获语义， 这些词嵌人已成为分布式语义建模 (Baroni et al.,2014)中的最新技术。但是，我们的方法和结果部分独立于这种表示，因为我们采用结构主义方法：我们通过对单词的语义相似关系对单词的含义进行建模来得出新的“二阶嵌${\text{人”。词汇 }(\mathrm{de~Saussure,~1916;~Rieger,~2003)~}}.$ 因此，未来的研究原则上可以通过任何其他能够在词汇单元之间产生语义相似度值的方法来替代这里考虑的语义的深度学习架构
这项工作的结构如下。在 82 中，我们讨论了相关工作。在 §3.1 和 3.2 中，我们分别正式介绍了概述的两个图模型。在 §4 中，我们详细介绍了我们的实验，在 §5 中，我们总结了
## 2 Related work
从广义上讲，可以区分两种最近的 NLP 方法来进行含义更改分析。一方面，粗粒度趋势单词在前一个时间段(Jatowt and Duh, 2014; | [
[
"0.2241",
"0.0974",
"0.775",
"0.0974",
"0.775",
"0.111",
"0.2241",
"0.111"
],
[
"0.2822",
"0.1484",
"0.388",
"0.1484",
"0.388",
"0.1633",
"0.2822",
"0.1633"
],
[
"0.5884",
"0.1478",
"0.744",
"0.1478",
"0.744",
"0.1609",
"0.5884",
"0.1609"
],
[
"0.1704",
"0.165",
"0.503",
"0.165",
"0.503",
"0.1799",
"0.1704",
"0.1799"
],
[
"0.5728",
"0.165",
"0.7573",
"0.165",
"0.7573",
"0.1799",
"0.5728",
"0.1799"
],
[
"0.1914",
"0.1818",
"0.4836",
"0.1818",
"0.4836",
"0.1953",
"0.1914",
"0.1953"
],
[
"0.5225",
"0.1818",
"0.81",
"0.1818",
"0.81",
"0.1953",
"0.5225",
"0.1953"
],
[
"0.189",
"0.1965",
"0.833",
"0.1978",
"0.833",
"0.212",
"0.189",
"0.2114"
],
[
"0.2627",
"0.275",
"0.3435",
"0.2766",
"0.3435",
"0.2922",
"0.2627",
"0.2903"
],
[
"0.511",
"0.2756",
"0.876",
"0.2756",
"0.876",
"0.2893",
"0.511",
"0.2893"
],
[
"0.513",
"0.2922",
"0.876",
"0.2922",
"0.876",
"0.306",
"0.513",
"0.306"
],
[
"0.1486",
"0.3027",
"0.455",
"0.3027",
"0.455",
"0.3164",
"0.1486",
"0.3164"
],
[
"0.511",
"0.3088",
"0.8774",
"0.3088",
"0.8774",
"0.3225",
"0.511",
"0.3225"
],
[
"0.1486",
"0.3196",
"0.4534",
"0.3196",
"0.4534",
"0.3333",
"0.1486",
"0.3333"
],
[
"0.5146",
"0.3267",
"0.8774",
"0.3267",
"0.8774",
"0.3403",
"0.5146",
"0.3403"
],
[
"0.147",
"0.3342",
"0.4568",
"0.335",
"0.4568",
"0.3499",
"0.147",
"0.3486"
],
[
"0.511",
"0.3416",
"0.8774",
"0.3416",
"0.8774",
"0.3552",
"0.511",
"0.3552"
],
[
"0.1486",
"0.3516",
"0.4534",
"0.3516",
"0.4534",
"0.3652",
"0.1486",
"0.3652"
],
[
"0.513",
"0.357",
"0.8774",
"0.357",
"0.8774",
"0.3706",
"0.513",
"0.3706"
],
[
"0.147",
"0.3677",
"0.4534",
"0.3677",
"0.4534",
"0.3813",
"0.147",
"0.3813"
],
[
"0.513",
"0.3748",
"0.8774",
"0.3748",
"0.8774",
"0.3884",
"0.513",
"0.3884"
],
[
"0.1486",
"0.3843",
"0.4568",
"0.3843",
"0.4568",
"0.398",
"0.1486",
"0.398"
],
[
"0.513",
"0.3901",
"0.8774",
"0.3901",
"0.8774",
"0.4038",
"0.513",
"0.4038"
],
[
"0.147",
"0.3997",
"0.4568",
"0.3984",
"0.4568",
"0.4133",
"0.147",
"0.4146"
],
[
"0.511",
"0.405",
"0.8774",
"0.4062",
"0.8774",
"0.4204",
"0.511",
"0.4192"
],
[
"0.147",
"0.415",
"0.4568",
"0.4146",
"0.4568",
"0.4287",
"0.147",
"0.43"
],
[
"0.5146",
"0.4229",
"0.8774",
"0.4229",
"0.8774",
"0.4358",
"0.5146",
"0.4358"
],
[
"0.1486",
"0.4316",
"0.4568",
"0.4316",
"0.4568",
"0.4453",
"0.1486",
"0.4453"
],
[
"0.513",
"0.4382",
"0.876",
"0.4382",
"0.876",
"0.452",
"0.513",
"0.452"
],
[
"0.147",
"0.4465",
"0.455",
"0.4478",
"0.455",
"0.462",
"0.147",
"0.4614"
],
[
"0.513",
"0.4531",
"0.8774",
"0.4536",
"0.8774",
"0.4685",
"0.513",
"0.4673"
],
[
"0.1486",
"0.4644",
"0.4568",
"0.4644",
"0.4568",
"0.478",
"0.1486",
"0.478"
],
[
"0.5156",
"0.4727",
"0.876",
"0.4727",
"0.876",
"0.4827",
"0.5156",
"0.4827"
],
[
"0.1486",
"0.4797",
"0.4575",
"0.4797",
"0.4575",
"0.4946",
"0.1486",
"0.4946"
],
[
"0.513",
"0.4868",
"0.8774",
"0.4868",
"0.8774",
"0.5005",
"0.513",
"0.5005"
],
[
"0.147",
"0.4963",
"0.4575",
"0.4963",
"0.4575",
"0.511",
"0.147",
"0.511"
],
[
"0.511",
"0.502",
"0.8774",
"0.503",
"0.8774",
"0.517",
"0.511",
"0.5166"
],
[
"0.1453",
"0.511",
"0.4575",
"0.5127",
"0.4575",
"0.527",
"0.1453",
"0.5254"
],
[
"0.511",
"0.5186",
"0.8447",
"0.517",
"0.8447",
"0.532",
"0.511",
"0.533"
],
[
"0.1486",
"0.5293",
"0.455",
"0.5293",
"0.455",
"0.542",
"0.1486",
"0.542"
],
[
"0.531",
"0.535",
"0.876",
"0.535",
"0.876",
"0.549",
"0.531",
"0.549"
],
[
"0.1486",
"0.5444",
"0.455",
"0.5444",
"0.455",
"0.558",
"0.1486",
"0.558"
],
[
"0.513",
"0.552",
"0.876",
"0.552",
"0.876",
"0.5654",
"0.513",
"0.5654"
],
[
"0.147",
"0.561",
"0.4568",
"0.561",
"0.4568",
"0.5747",
"0.147",
"0.5747"
],
[
"0.513",
"0.5684",
"0.8774",
"0.5684",
"0.8774",
"0.582",
"0.513",
"0.582"
],
[
"0.147",
"0.5767",
"0.455",
"0.576",
"0.455",
"0.5903",
"0.147",
"0.5913"
],
[
"0.513",
"0.585",
"0.8774",
"0.585",
"0.8774",
"0.5986",
"0.513",
"0.5986"
],
[
"0.1486",
"0.593",
"0.4568",
"0.593",
"0.4568",
"0.6055",
"0.1486",
"0.6055"
],
[
"0.511",
"0.5996",
"0.876",
"0.5996",
"0.876",
"0.6133",
"0.511",
"0.6133"
],
[
"0.1486",
"0.6094",
"0.4568",
"0.6094",
"0.4568",
"0.6226",
"0.1486",
"0.6226"
],
[
"0.513",
"0.616",
"0.8774",
"0.616",
"0.8774",
"0.63",
"0.513",
"0.63"
],
[
"0.1453",
"0.6235",
"0.4417",
"0.6245",
"0.4417",
"0.6396",
"0.1453",
"0.638"
],
[
"0.5146",
"0.633",
"0.8774",
"0.633",
"0.8774",
"0.6465",
"0.5146",
"0.6465"
],
[
"0.1436",
"0.6494",
"0.2644",
"0.6494",
"0.2644",
"0.6646",
"0.1436",
"0.6646"
],
[
"0.5146",
"0.6484",
"0.879",
"0.6484",
"0.879",
"0.662",
"0.5146",
"0.662"
],
[
"0.513",
"0.663",
"0.8774",
"0.663",
"0.8774",
"0.677",
"0.513",
"0.677"
],
[
"0.1184",
"0.652",
"0.1511",
"0.652",
"0.1511",
"0.663",
"0.1184",
"0.663"
],
[
"0.12177",
"0.675",
"0.4878",
"0.675",
"0.4878",
"0.6895",
"0.12177",
"0.6895"
],
[
"0.511",
"0.6797",
"0.8774",
"0.6797",
"0.8774",
"0.6934",
"0.511",
"0.6934"
],
[
"0.12006",
"0.6914",
"0.486",
"0.6914",
"0.486",
"0.705",
"0.12006",
"0.705"
],
[
"0.511",
"0.697",
"0.8237",
"0.697",
"0.8237",
"0.7104",
"0.511",
"0.7104"
],
[
"0.1184",
"0.706",
"0.4836",
"0.706",
"0.4836",
"0.7197",
"0.1184",
"0.7197"
],
[
"0.529",
"0.7144",
"0.8774",
"0.7144",
"0.8774",
"0.728",
"0.529",
"0.728"
],
[
"0.1184",
"0.7236",
"0.4844",
"0.7236",
"0.4844",
"0.737",
"0.1184",
"0.737"
],
[
"0.511",
"0.7285",
"0.8774",
"0.7285",
"0.8774",
"0.742",
"0.511",
"0.742"
],
[
"0.12006",
"0.7393",
"0.486",
"0.7393",
"0.486",
"0.753",
"0.12006",
"0.753"
],
[
"0.513",
"0.745",
"0.8774",
"0.745",
"0.8774",
"0.759",
"0.513",
"0.759"
],
[
"0.12006",
"0.755",
"0.4878",
"0.755",
"0.4878",
"0.7695",
"0.12006",
"0.7695"
],
[
"0.513",
"0.763",
"0.8716",
"0.763",
"0.8716",
"0.777",
"0.513",
"0.777"
],
[
"0.1184",
"0.7695",
"0.486",
"0.771",
"0.486",
"0.785",
"0.1184",
"0.7837"
],
[
"0.12006",
"0.788",
"0.4844",
"0.788",
"0.4844",
"0.802",
"0.12006",
"0.802"
],
[
"0.51",
"0.7944",
"0.55",
"0.7944",
"0.55",
"0.8047",
"0.51",
"0.8047"
],
[
"0.546",
"0.792",
"0.665",
"0.792",
"0.665",
"0.807",
"0.546",
"0.807"
],
[
"0.1234",
"0.804",
"0.4836",
"0.804",
"0.4836",
"0.818",
"0.1234",
"0.818"
],
[
"0.12177",
"0.8193",
"0.34",
"0.8193",
"0.34",
"0.8345",
"0.12177",
"0.8345"
],
[
"0.513",
"0.8184",
"0.879",
"0.8193",
"0.879",
"0.8345",
"0.513",
"0.833"
],
[
"0.1377",
"0.836",
"0.4844",
"0.836",
"0.4844",
"0.8496",
"0.1377",
"0.8496"
],
[
"0.513",
"0.836",
"0.8774",
"0.836",
"0.8774",
"0.8496",
"0.513",
"0.8496"
],
[
"0.12006",
"0.8516",
"0.4895",
"0.8516",
"0.4895",
"0.865",
"0.12006",
"0.865"
],
[
"0.513",
"0.8516",
"0.879",
"0.8516",
"0.879",
"0.865",
"0.513",
"0.865"
],
[
"0.1184",
"0.868",
"0.4836",
"0.868",
"0.4836",
"0.882",
"0.1184",
"0.882"
],
[
"0.513",
"0.868",
"0.879",
"0.868",
"0.879",
"0.882",
"0.513",
"0.882"
]
] | [
"关于语义变化的线性：通过动态图模型研究含义变化",
"Steffen Eger",
"AlexanderMehler",
"Ubiquitous Knowledge Processing Lab",
"Text Technology Lab",
"Department of Computer Science",
"Department of Computer Science",
"Technische Universitat Darmstadt Goethe-Universitat Frankfurt am Main",
"Abstract",
"的每个节点对应一个单词，与该单词相关联的",
"是语义embedding向量。然后，我们询问如",
"We consider two graph models of se-",
"何根据先前时间段中相邻单词的嵌入向量来",
"mantic change. The first is a time-",
"预测一个时间段（图）中的嵌入向量。特别是，",
"series model that relates embedding",
"我们假设存在一种线性函数关系，该关系将单",
"vectors from one time period to em-",
"词的今天含义与过去的邻居含义耦合在一起。",
"bedding vectors of previous time peri",
"在估计该模型的系数时，我们发现线性形式",
"ods. In the second, we construct one",
"看起来确实非常合理。然后，这种功能形式使",
"graph for each word: nodes in this",
"我们能够解决其他问题，例如单词之间的负面",
"graph correspond to time points and",
"关系-表示随时间的语义差异-以及对未来的预",
"edge weights to the similarity of the",
"测。我们称我们的第二个图模型为时间索引自",
"word's meaning across two time points.",
"相似图。在这些图中，每个节点对应一个时间",
"We apply our two models to corpora",
"点，两个时间点之间的链接表示所考虑的两个",
"across three different languages. We",
"时间点之间特定单词的语义相似性。对这些图",
"find that semantic change is linear in",
"的分析表明，大多数单词都遵循线性语义“衰",
"two senses. Firstly, today's embedding",
"减”定律：语义自相似性随时间线性降低",
"vectors (= meaning) of words can be",
"在我们的工作中，我们通过从上下文预测",
"derived as linear combinations of em-",
"神经网络体系结构派生的词嵌入来捕获语义",
"bedding vectors of their neighbors in",
"这些词嵌已成为分布式语义建模（Baroniet",
"previous time periods. Secondly, self-",
"al.，2014）中的最新技术。但是，我们的方法",
"similarity of words decays linearly in",
"和结果部分独立于这种表示，因为我们采用",
"time. We consider both findings as new",
"结构主义方法：我们通过对单词的语义相似关",
"laws/hypotheses of semantic change.",
"系对单词的含义进行建模来得出新的“二阶嵌",
"Introduction",
"人”。词汇(deSaussure,1916Rieger,2003)。",
"因此，未来的研究原则上可以通过任何其他能",
"1",
"含义不统一，既不跨空间，也不跨时间。在",
"够在词汇单元之间产生语义相似度值的方法",
"整个空间中，不同的语言倾向于对相应的术",
"来替代这里考虑的语义的深度学习架构",
"语表现出不同的多义关联（Egeretal.，2015",
"这项工作的结构如下。在S2中，我们讨论",
"Kulkarnietal.，2015b）。随着时间的推移，已",
"了相关工作。在63.1和3.2中，我们分别正式",
"经记录了几个著名的英语意义变化的例子。例",
"介绍了概述的两个图模型。在S4中，我们详",
"如，在1970年代，“同性恋”一词的含义已从",
"细介绍了我们的实验，在$5中，我们总结了",
"20th世纪初的“快乐”的形容词含义转变为",
"“同性恋（Kulkarnietal.，2015a）”的当前含",
"2",
"Related work",
"义。。同样，技术进步也导致了术语的语义扩",
"展，例如传输，鼠标或苹果",
"从广义上讲，可以区分两种最近的NLP方法",
"在这项工作中，我们考虑了语义变化的两个",
"来进行含义更改分析。一方面，粗粒度趋势",
"图模型。我们的第一个模型是一个动态模型，",
"分析将一个单词在一个时间段内的语义与该",
"因为其基础范式是（时间）系列图。系列图中",
"单词在前一个时间段（JatowtandDuh，2014;"
] | {
"arxiv_id": "1704.02497",
"language": "zh",
"page_idx": 0
} |
||
1704.02497-zh-1 | 1704.02497-zh-1.jpg | 在哪些方面发生了变化 (e。g。, 语义扩展或缩小),但只是旨在捕捉是否发生了意义变化。相反，更细粒度的分析通常会在语料库中对单词进行意义标记，然后研究相应含义分布 (Rohrdantz et al., 2011; Mitra et al., 2014; P$\" {o}$litzetal. , 2015; Zhang et al. , 2015) 的变化。可以通过对单词的上下文向量进行聚类来实现意义标记 (Huang et al., 2012; Chen et al., 2014; Neelakantan et al., 2014) 或通过应用基于 LDA 的技术，其中单词上下文扮演文档的角色，单词意义扮演主题(Rohrdantz et al., 2011; Lau et al., 2012) 的角色。最后，有一些研究可以测试特定的含义变化假设，例如相似的单词是否会随着时间的推移而出现含义差异(根据“差异定律”) (Xu and Kemp, 2015) 以及旨在跨时间检测相应术语的论文(具有相似含义的单词/两个时间段中的角色，但词汇形式可能不同)(Zhang et al., 2015)
## 3 Graph models
让$V=\{w_1,\ldots,w_{|V|}\}$成为所有时间段$t\in\mathcal{T}$ 中所有单词的通用词汇(交集)。这里，$\mathcal{T}$是一组时间索引。将时间周期$t$处的单词$w_i$的嵌入表示为$\boldsymbol w_i(t)\in\mathbb{R}^d$ 。由于两个不同时间段$s,t$的嵌入$\boldsymbol w_i(s),\boldsymbol w_i(t)$通常是不可比较的， 因为它们可能位于不同的坐标系中，因此我们考虑向量$\tilde{\boldsymbol{w}}_i(t)=$
$$\begin{pmatrix}\sin(\boldsymbol w_i(t),\boldsymbol w_1(t)),\dots,\sin(\boldsymbol w_i(t),\boldsymbol w_{|V|}(t))\end{pmatrix},$$
,其中每个向量都位于 $\mathbb{R}^|V|$ 中，并且其中 sim 是类似函数，例如余弦。我们注意到我们对$\tilde{\boldsymbol{w}}_i(t)$的结构主义定义并不是没有问题的，因为根据我们的假设，$\boldsymbol w_1(t),\ldots,\boldsymbol{w}_{|V|}(t)$在$t$上往往是不同的，因此随着时间的推移，这些“参考点”是不一致的。但是，由于我们可能假设至少几个单词的含义随着时间的推移是稳定的，因此我们强烈希望向量$\tilde{\boldsymbol{w}}_i(t)$适合我们分析含义变化的任务。$^{1}$为方便起见，在本工作的其余部分中，我们不区分$\boldsymbol w_i(t)$和$\tilde{\boldsymbol{w}}_i(t)$
$^{1}$An alternative to our second-order embeddings is to project vectors from different time periods in a common space (Mikolov et al., 2013a; Faruqui and Dyer, 2014) , which requires to find corresponding terms across time. Further, one could also consider a $`$core' vocabulary of semantically stable words, e.g., in the spirit of Swadesh (1952) , instead of using all vocabu- ${\tilde{\text{lary words as reference}}}.$
Kulkarni et al., 2015a) 中的含义进行比较。这 3.1 A linear model of semantic change
种粗粒度的模型本身并没有具体说明一个词 我们假设并随后测试了以下含义动态模型，该
模型描述了单词$w_i$随时间变化的含义：
(2)
$$\boldsymbol{w}_i(t)=\sum_{n=1}^p\sum_{w_j\in V\cap N(w_i)}\alpha_{w_j}^n\boldsymbol{w}_j(t-n)$$
,其中$\alpha_w_j^n\in\mathbb{R}$ ,对于$n=1,\ldots,p$,是含义向量的系数。。集合$N(w_i)\subseteq V$表示单词$w_i$ 的“邻居”集合。$^{2}$这个模型说，一个单词 $w_i$ 在某个时间$t$的含义是通过参考其在以前时间段内的“邻居”的含义来确定的，并且潜在的功能关系是线性的
我们注意到，方程式$\mathbb{E}(2)$ 描述的模型是时间序列模型，尤其是具有特殊结构的向量自回归(VAR)模型。该模型也可以在所谓的“意见动态模型”(Golub and Jackson, 2010; Acemoglu and Ozdaglar, 2011; Eger, 2016) 的社会经济背景下看到。。假设代理位于网络结构中，并根据其与其他代理的关系不断更新其意见/信念/行动。模型(2)将多维嵌入向量替换为一维意见。
3.2 Time-indexed self-similarity graphs
我们通过考虑$V$中每个单词$w$的完全连接图$G(w)$来跟踪含义更改。$G(w)$的节点是时间索引$\mathcal{T}$,并且在任何两个$s,t\in\mathcal{T}$之间存在无向链接，其权重由$\sin(\boldsymbol{w}(s),\boldsymbol{w}(t))$ 给出。我们将图称为$G(w)$时间索引自相似性(TISS) 图，因为它们表示给定单词在不同时间段内与其自身的(语义)相似性。特别感兴趣的可能在于这些图中的薄弱环节，因为它们表明两个不同时间段之间的相似性较低，即。e。,语义随时间变化
### 4 Experiments
数据作为英语的语料库，我们使用历史美国人语料库 (COHA)。$^3$这涵盖了从1810到2000 时间段的文本。我们提取了两个切片： 1900-2000 年和 1810-2000 年。对于这两个切片，每个时间段$t$都是一个十年，e。g。$,\mathcal{T}=$ $\{ 1810, 1820, 1830, \ldots \}$ $。^{4}$对于每个切片，我们只保留与词类名词，形容词和动词相关联的单词。出于计算和估计的目的，我们也只
${^{2}\text{We also con}}$strain the vectors $\boldsymbol w_i(t)$ ,for all $w_i\in V$ , to contain non-zero entries only for words in $N( w_{i})$ .
$^3$http://corpus.byu.edu/coha/.
$\begin{array}{cc}{^{4}\text{Each time period contains texts that were written}}\\{^{4}\text{Each time period contains texts that were written}}\end{array}$
in that decade.
2 | [
[
"0.12006",
"0.0885",
"0.4895",
"0.0885",
"0.4895",
"0.10156",
"0.12006",
"0.10156"
],
[
"0.5054",
"0.0861",
"0.839",
"0.0885",
"0.839",
"0.10156",
"0.5054",
"0.0998"
],
[
"0.12177",
"0.10394",
"0.486",
"0.10394",
"0.486",
"0.11755",
"0.12177",
"0.11755"
],
[
"0.511",
"0.10925",
"0.8774",
"0.10925",
"0.8774",
"0.1223",
"0.511",
"0.1223"
],
[
"0.1184",
"0.11816",
"0.486",
"0.1194",
"0.486",
"0.1342",
"0.1184",
"0.133"
],
[
"0.513",
"0.1247",
"0.825",
"0.1247",
"0.825",
"0.1383",
"0.513",
"0.1383"
],
[
"0.12006",
"0.136",
"0.4844",
"0.136",
"0.4844",
"0.1497",
"0.12006",
"0.1497"
],
[
"0.12177",
"0.152",
"0.4836",
"0.152",
"0.4836",
"0.165",
"0.12177",
"0.165"
],
[
"0.534",
"0.1567",
"0.881",
"0.1567",
"0.881",
"0.1818",
"0.534",
"0.1818"
],
[
"0.613",
"0.1567",
"0.6206",
"0.1567",
"0.6206",
"0.1603",
"0.613",
"0.1603"
],
[
"0.12006",
"0.1687",
"0.4836",
"0.1687",
"0.4836",
"0.1818",
"0.12006",
"0.1818"
],
[
"0.1184",
"0.1841",
"0.486",
"0.1853",
"0.486",
"0.199",
"0.1184",
"0.1978"
],
[
"0.6035",
"0.1818",
"0.723",
"0.1818",
"0.723",
"0.1953",
"0.6035",
"0.1953"
],
[
"0.1184",
"0.1996",
"0.4844",
"0.2019",
"0.4844",
"0.2156",
"0.1184",
"0.2131"
],
[
"0.513",
"0.209",
"0.879",
"0.2102",
"0.879",
"0.2245",
"0.513",
"0.2239"
],
[
"0.12177",
"0.2185",
"0.4836",
"0.2185",
"0.4836",
"0.2286",
"0.12177",
"0.2286"
],
[
"0.513",
"0.2268",
"0.879",
"0.228",
"0.879",
"0.2411",
"0.513",
"0.2405"
],
[
"0.12006",
"0.2328",
"0.4844",
"0.2328",
"0.4844",
"0.2477",
"0.12006",
"0.2477"
],
[
"0.51",
"0.2423",
"0.879",
"0.2434",
"0.879",
"0.2583",
"0.51",
"0.257"
],
[
"0.12006",
"0.2488",
"0.486",
"0.2488",
"0.486",
"0.2625",
"0.12006",
"0.2625"
],
[
"0.12006",
"0.2654",
"0.4844",
"0.2654",
"0.4844",
"0.279",
"0.12006",
"0.279"
],
[
"0.5146",
"0.2612",
"0.8774",
"0.2612",
"0.8774",
"0.2715",
"0.5146",
"0.2715"
],
[
"0.12006",
"0.2808",
"0.4878",
"0.2808",
"0.4878",
"0.2944",
"0.12006",
"0.2944"
],
[
"0.513",
"0.2756",
"0.879",
"0.2756",
"0.879",
"0.2893",
"0.513",
"0.2893"
],
[
"0.11755",
"0.2944",
"0.4878",
"0.2935",
"0.4878",
"0.3113",
"0.11755",
"0.3123"
],
[
"0.513",
"0.2915",
"0.678",
"0.2915",
"0.678",
"0.3047",
"0.513",
"0.3047"
],
[
"0.5283",
"0.3071",
"0.8774",
"0.308",
"0.8774",
"0.3225",
"0.5283",
"0.3218"
],
[
"0.12177",
"0.3135",
"0.4844",
"0.3135",
"0.4844",
"0.3267",
"0.12177",
"0.3267"
],
[
"0.12006",
"0.3289",
"0.4844",
"0.3289",
"0.4844",
"0.3425",
"0.12006",
"0.3425"
],
[
"0.511",
"0.3237",
"0.875",
"0.3237",
"0.875",
"0.3374",
"0.511",
"0.3374"
],
[
"0.12006",
"0.3445",
"0.4844",
"0.3445",
"0.4844",
"0.3582",
"0.12006",
"0.3582"
],
[
"0.513",
"0.3403",
"0.8774",
"0.3403",
"0.8774",
"0.354",
"0.513",
"0.354"
],
[
"0.513",
"0.3557",
"0.879",
"0.3557",
"0.879",
"0.3694",
"0.513",
"0.3694"
],
[
"0.12177",
"0.3635",
"0.4836",
"0.3635",
"0.4836",
"0.3735",
"0.12177",
"0.3735"
],
[
"0.511",
"0.3718",
"0.8774",
"0.3718",
"0.8774",
"0.3855",
"0.511",
"0.3855"
],
[
"0.12006",
"0.3777",
"0.4878",
"0.3777",
"0.4878",
"0.3914",
"0.12006",
"0.3914"
],
[
"0.513",
"0.3884",
"0.881",
"0.3884",
"0.881",
"0.402",
"0.513",
"0.402"
],
[
"0.12006",
"0.3943",
"0.4072",
"0.3943",
"0.4072",
"0.408",
"0.12006",
"0.408"
],
[
"0.513",
"0.4026",
"0.879",
"0.4038",
"0.879",
"0.4187",
"0.513",
"0.4175"
],
[
"0.1503",
"0.4204",
"0.2795",
"0.4204",
"0.2795",
"0.4353",
"0.1503",
"0.4353"
],
[
"0.5146",
"0.4204",
"0.8774",
"0.4204",
"0.8774",
"0.434",
"0.5146",
"0.434"
],
[
"0.51",
"0.4353",
"0.635",
"0.4358",
"0.635",
"0.4507",
"0.51",
"0.4495"
],
[
"0.12006",
"0.4229",
"0.157",
"0.4229",
"0.157",
"0.434",
"0.12006",
"0.434"
],
[
"0.12006",
"0.4465",
"0.4844",
"0.4465",
"0.4844",
"0.4614",
"0.12006",
"0.4614"
],
[
"0.12177",
"0.462",
"0.4878",
"0.462",
"0.4878",
"0.4756",
"0.12177",
"0.4756"
],
[
"0.509",
"0.462",
"0.843",
"0.4644",
"0.843",
"0.4785",
"0.509",
"0.4768"
],
[
"0.1351",
"0.4785",
"0.4844",
"0.4785",
"0.4844",
"0.4922",
"0.1351",
"0.4922"
],
[
"0.513",
"0.485",
"0.8774",
"0.485",
"0.8774",
"0.4988",
"0.513",
"0.4988"
],
[
"0.12006",
"0.4946",
"0.4844",
"0.4946",
"0.4844",
"0.5083",
"0.12006",
"0.5083"
],
[
"0.511",
"0.5005",
"0.876",
"0.5005",
"0.876",
"0.514",
"0.511",
"0.514"
],
[
"0.1184",
"0.511",
"0.486",
"0.511",
"0.486",
"0.525",
"0.1184",
"0.525"
],
[
"0.513",
"0.517",
"0.8774",
"0.517",
"0.8774",
"0.531",
"0.513",
"0.531"
],
[
"0.12177",
"0.527",
"0.4878",
"0.527",
"0.4878",
"0.5405",
"0.12177",
"0.5405"
],
[
"0.5146",
"0.533",
"0.8774",
"0.533",
"0.8774",
"0.5464",
"0.5146",
"0.5464"
],
[
"0.12006",
"0.5435",
"0.2612",
"0.5435",
"0.2612",
"0.557",
"0.12006",
"0.557"
],
[
"0.513",
"0.549",
"0.879",
"0.5474",
"0.879",
"0.561",
"0.513",
"0.5625"
],
[
"0.5146",
"0.5654",
"0.8774",
"0.5654",
"0.8774",
"0.579",
"0.5146",
"0.579"
],
[
"0.1368",
"0.5737",
"0.4844",
"0.5737",
"0.4844",
"0.5874",
"0.1368",
"0.5874"
],
[
"0.5146",
"0.583",
"0.8774",
"0.583",
"0.8774",
"0.5933",
"0.5146",
"0.5933"
],
[
"0.12427",
"0.6016",
"0.4844",
"0.6016",
"0.4844",
"0.616",
"0.12427",
"0.616"
],
[
"0.513",
"0.597",
"0.8774",
"0.597",
"0.8774",
"0.611",
"0.513",
"0.611"
],
[
"0.511",
"0.6123",
"0.879",
"0.6133",
"0.879",
"0.6274",
"0.511",
"0.627"
],
[
"0.12006",
"0.619",
"0.4844",
"0.619",
"0.4844",
"0.633",
"0.12006",
"0.633"
],
[
"0.12006",
"0.6353",
"0.4844",
"0.6353",
"0.4844",
"0.6484",
"0.12006",
"0.6484"
],
[
"0.5146",
"0.6313",
"0.6055",
"0.6313",
"0.6055",
"0.6426",
"0.5146",
"0.6426"
],
[
"0.12006",
"0.652",
"0.4878",
"0.652",
"0.4878",
"0.6665",
"0.12006",
"0.6665"
],
[
"0.5156",
"0.66",
"0.5483",
"0.66",
"0.5483",
"0.6685",
"0.5156",
"0.6685"
],
[
"0.55",
"0.659",
"0.659",
"0.659",
"0.659",
"0.6704",
"0.55",
"0.6704"
],
[
"0.1184",
"0.6675",
"0.486",
"0.6665",
"0.486",
"0.681",
"0.1184",
"0.6816"
],
[
"0.1184",
"0.683",
"0.486",
"0.684",
"0.486",
"0.6987",
"0.1184",
"0.6978"
],
[
"0.513",
"0.683",
"0.8774",
"0.683",
"0.8774",
"0.697",
"0.513",
"0.697"
],
[
"0.12177",
"0.7017",
"0.4844",
"0.7017",
"0.4844",
"0.712",
"0.12177",
"0.712"
],
[
"0.511",
"0.6987",
"0.8774",
"0.6987",
"0.8774",
"0.712",
"0.511",
"0.712"
],
[
"0.12006",
"0.7163",
"0.486",
"0.7163",
"0.486",
"0.73",
"0.12006",
"0.73"
],
[
"0.511",
"0.7153",
"0.881",
"0.7153",
"0.881",
"0.7285",
"0.511",
"0.7285"
],
[
"0.12006",
"0.7324",
"0.4878",
"0.7324",
"0.4878",
"0.745",
"0.12006",
"0.745"
],
[
"0.513",
"0.731",
"0.879",
"0.731",
"0.879",
"0.7446",
"0.513",
"0.7446"
],
[
"0.12006",
"0.749",
"0.4844",
"0.749",
"0.4844",
"0.7627",
"0.12006",
"0.7627"
],
[
"0.513",
"0.7476",
"0.876",
"0.7476",
"0.876",
"0.7627",
"0.513",
"0.7627"
],
[
"0.513",
"0.7627",
"0.879",
"0.763",
"0.879",
"0.778",
"0.513",
"0.777"
],
[
"0.5146",
"0.78",
"0.879",
"0.78",
"0.879",
"0.7935",
"0.5146",
"0.7935"
],
[
"0.1411",
"0.7876",
"0.2747",
"0.788",
"0.2747",
"0.799",
"0.1411",
"0.7974"
],
[
"0.2805",
"0.788",
"0.4836",
"0.788",
"0.4836",
"0.799",
"0.2805",
"0.799"
],
[
"0.1184",
"0.7974",
"0.4768",
"0.7974",
"0.4768",
"0.811",
"0.1184",
"0.811"
],
[
"0.513",
"0.7964",
"0.879",
"0.7964",
"0.879",
"0.81",
"0.513",
"0.81"
],
[
"0.1486",
"0.81",
"0.4844",
"0.81",
"0.4844",
"0.8237",
"0.1486",
"0.8237"
],
[
"0.12177",
"0.822",
"0.4844",
"0.822",
"0.4844",
"0.835",
"0.12177",
"0.835"
],
[
"0.529",
"0.818",
"0.879",
"0.8193",
"0.879",
"0.8345",
"0.529",
"0.832"
],
[
"0.1234",
"0.835",
"0.4836",
"0.835",
"0.4836",
"0.8457",
"0.1234",
"0.8457"
],
[
"0.511",
"0.832",
"0.879",
"0.833",
"0.879",
"0.844",
"0.511",
"0.8433"
],
[
"0.12006",
"0.8457",
"0.4836",
"0.8467",
"0.4836",
"0.858",
"0.12006",
"0.857"
],
[
"0.529",
"0.8403",
"0.744",
"0.843",
"0.744",
"0.86",
"0.529",
"0.858"
],
[
"0.12177",
"0.8594",
"0.482",
"0.8594",
"0.482",
"0.869",
"0.12177",
"0.869"
],
[
"0.531",
"0.855",
"0.879",
"0.857",
"0.879",
"0.8716",
"0.531",
"0.869"
],
[
"0.1234",
"0.8726",
"0.2778",
"0.8726",
"0.2778",
"0.88",
"0.1234",
"0.88"
],
[
"0.513",
"0.8706",
"0.614",
"0.8706",
"0.614",
"0.8813",
"0.513",
"0.8813"
],
[
"0.4902",
"0.903",
"0.509",
"0.903",
"0.509",
"0.9175",
"0.4902",
"0.9175"
]
] | [
"Kulkarnietal.，2015a）中的含义进行比较。这",
"3.1Alinear model of semanticchange",
"种粗粒度的模型本身并没有具体说明一个词",
"我们假设并随后测试了以下含义动态模型，该",
"在哪些方面发生了变化（e。g。，语义扩展",
"模型描述了单词W,随时间变化的含义：",
"或缩小，但只是旨在捕捉是否发生了意义变",
"化。相反，更细粒度的分析通常会在语料库中",
"wi(t)=αm,w;(t-n) (2)",
"",
"对单词进行意义标记，然后研究相应含义分",
"布 (Rohrdantz et al., 2011; Mitra et al., 2014;",
"n=l wjeVnN(wi)",
"Politzetal.,2015;Zhangetal.,2015）的变化。",
"，其中αn,ER，对于n=l,..,p，是含义",
"可以通过对单词的上下文向量进行聚类来实",
"向量的系数。。集合N(wi）CV表示单词wi",
"现意义标记(Huangetal.,2012;Chen etal.",
"的“邻居”集合。2这个模型说，一个单词wi",
"2014；Neelakantanetal.，2014）或通过应用基",
"于LDA的技术，其中单词上下文扮演文档的",
"在某个时间t的含义是通过参考其在以前时",
"角色，单词意义扮演主题（Rohrdantzetal.,",
"间段内的“邻居”的含义来确定的，并且潜在",
"2011;Lauetal，2012)的角色。最后，有一些",
"的功能关系是线性的",
"我们注意到，方程式国(2）描述的模型是时",
"研究可以测试特定的含义变化假设，例如相似",
"的单词是否会随着时间的推移而出现含义差",
"间序列模型，尤其是具有特殊结构的向量自",
"异（根据“差异定律”）（XuandKemp，2015）",
"回归（VAR）模型。该模型也可以在所谓的",
"“意见动态模型”（GolubandJackson，2010;",
"以及旨在跨时间检测相应术语的论文（具有",
"Acemoglu and Ozdaglar,2011;Eger,2016)的",
"似含义的单词/两个时间段中的角色，但词汇",
"社会经济背景下看到。。假设代理位于网络结",
"形式可能不同）（Zhangetal.，2015）",
"构中，并根据其与其他代理的关系不断更新其",
" Graph models",
"意见/信念行动。模型（2）将多维嵌入向量替",
"换为一维意见。",
"3",
"让V={w1,..,v成为所有时间段tT",
"中所有单词的通用词汇（交集）。这里，T是",
"3.2Time-indexed self-similarity graphs",
"组时间索引。将时间周期t处的单词wi的",
"我们通过考虑V中每个单词W的完全连接图",
"嵌人表示为wi(t)ERd。由于两个不同时间",
"G(w）来跟踪含义更改。G(w）的节点是时间",
"段 s,t的嵌人wi(s),wi(t)通常是不可比较的，",
"索引T，并且在任何两个s,tET之间存在",
"因为它们可能位于不同的坐标系中，因此我们",
"无向链接，其权重由sim(w(s),w(t)）给出。我",
"考虑向量wi(t)=",
"们将图称为G(w）时间索引自相似性（TISS)",
"图，因为它们表示给定单词在不同时间段内与",
"(sim(wi(t), wi(t), ..., sim(w;(t),wjv(t), (1)",
"其自身的（语义）相似性。特别感兴趣的可能",
"，其中每个向量都位于R/VI中，并且其中sim",
"在于这些图中的薄弱环节，因为它们表明两个",
"不同时间段之间的相似性较低，即。e。，语义",
"是类似函数，例如余弦。我们注意到我们对",
"(t)的结构主义定义并不是没有问题的，因",
"随时间变化",
"为根据我们的假设，wi(t)，...,wiv(t)在t上",
"4",
"Experiments",
"往往是不同的，因此随着时间的推移，这些",
"“参考点”是不一致的。但是，由于我们可能假",
"数据作为英语的语料库，我们使用历史美国",
"设至少几个单词的含义随着时间的推移是稳",
"人语料库（COHA）。3这涵盖了从1810到",
"定的，因此我们强烈希望向量wi(t)适合我们",
"2000时间段的文本。我们提取了两个切片：",
"分析含义变化的任务。1为方便起见，在本工",
"1900-2000年和1810-2000年。对于这两个切",
"作的其余部分中，我们不区分wi(t）和w;(t)",
"片，每个时间段t都是一个十年，e。g。，T=",
"[1810,1820,1830，..}。4对于每个切片，我",
"们只保留与词类名词，形容词和动词相关联",
"Analternativeto",
"ur second-order embeddingsis",
"to project vectors from different time periods in a com",
"的单词。出于计算和估计的目的，我们也只",
" space (Mikolov et al., 2013a; Faruqui and Dyer",
"2014) , which requires to find corresponding terms",
"2We also constrain the vectors w;(t) , for all w; E V",
"across time.Further,one could also consider a‘core",
",to containnon-zero entries onlyfor wordsinN(wi):",
"vocabulary of semantically stable words, e.g., in the",
"Shttp://corpus.byu.edu/coha/.",
"spirit of Swadesh (1952),instead of using all vocabu",
"4Each time period contains texts that were written",
"ary wordsas reference",
"in thatdecade",
"2"
] | {
"arxiv_id": "1704.02497",
"language": "zh",
"page_idx": 1
} |
||
1704.02497-zh-2 | 1704.02497-zh-2.jpg | 考虑在每个时间段内至少出现 100 次的单词。为了诱导每个单词$w\in V$的单词嵌入$w\in$ $\mathbb{R}^d$ ,我们使用带有默认参数化的 word2vec (Mikolov et al., 2013b)。我们这样做每个时间间段$t\in\mathcal{T}$独立。然后，我们使用这些嵌入来导出新的嵌入，如方程 (1)。自始至终，我们使用余弦相似度作为 sim 度量。对于德语， 我们考虑德国报纸 SZ $^{5}$的专有数据集，其中$\mathcal{T}=\{1994,1995,\ldots,2003\}$。我们对数据进行lemmatize 和 POS 标记，同样只考虑名词，动词和形容词，使频率约束与英语相同。最后， 我们使用 PL (Migne, 1855) 作为拉丁语的数据集。这里，$\mathcal{T}=\{300,400,\ldots,1300\}$。我们使用与英语和德语相同的预处理、频率和词类约束
自始至终，我们的数据集在大小方面都很平衡。例如，从 1900 年到 2000 年的每十年， 英语 COHA 数据集包含大约 24M-30M 令牌， 其中 1990 年和 2000 年包含的数据比前几十年略多。1900 年之前的几十年包含 18-24M 令牌，只有 1810 年和 1820 年包含很少的数据(分别为 1M 和 7M 令牌)。语料库也由体裁平衡
## 4.1 TISS graphs
我们从研究 TISS 图开始。令$D_{t_0}$表示当时间段之间的距离为$t_0$时，平均而言，单词在两个时间段上的语义相似程度： $D_{t_0}= \frac 1C\sum _{w\in V}\sum _{| s- t| = t_0}\operatorname { sim} ( \boldsymbol{w}( s) , \boldsymbol{w}( t) )$ ,其中$C=|V|\cdot|\{(s,t)||s-t|=t_{0}\}|$是规范器。Figure 1 plotsthevalues $D_{t_0}$ forthetimeslicefrom1810to2000,fortheEnglishdata。我们注意到一个明显的趋势：单词的自相似性倾向于(几乎完美地)随时间距离线性降低。实际上， 下面的表 1 表明 , 这种趋势在我们所有的语料库中都成立，即。e。,对于不同的时间尺度和不同的语言：线性“衰减”模型非常好地拟合了$D_{t_0}$曲线，调整后的$R^2$值大大高于 90%, 并且一致且显着为负系数。我们认为，这一发现可能被认为是语义变化的新统计规律
作为$t_0$的函数的值$D_t_0$是所有单词的平均值。因此，平均单词的含义可能会随时间线性衰减，而随着时间的推移，大部分单词的语义行为遵循不同的趋势。为了研究这一点，我们考虑$D_t_0(w)=\frac1{C^{\prime}}\sum_{|s-t|=t_0}\sin(\boldsymbol{w}(s),\boldsymbol{w}(t))\: R_{G(w)}=\frac{\mathrm{maxlink}}{\mathrm{minlink}}$对图形 $G(w)$ 进行排序，其
在固定词$w$上的分布。这里$C^\prime=\left|\{(s,t)\mid|s-\right.$ $t|=t_0\}|$ 。我们独立考虑每个单词$w$的回归
$^{5}$http://www.sueddeutsche.de/
<FigureHere>
Figure 1:$D_{t_0}(y$-axis) as a function of $t_0(x$ - axis),values of$D_t_{0}\pmod{\mathrm{error-bars.}}$
<table>
<tbody>
<tr>
<th>$\overline{\mathrm{Corp}}$</th>
<th>$\overline{\mathrm{usLang.}}$</th>
<th>Time $inter-$ val</th>
<th>Years</th>
<th>Coeff. </th>
<th>$R^{2}$</th>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td>2000</td>
<td>0.425</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td>1810-</td>
<td> </td>
<td>96.03</td>
</tr>
<tr>
<td>$PL$</td>
<td>Latin</td>
<td>Century</td>
<td>$-vvv$ 400-1300</td>
<td>$\psi\Pi\psi$ 一 0.228</td>
<td>92.28</td>
</tr>
<tr>
<td>$PL$</td>
<td>Latin</td>
<td>Century</td>
<td>400-1300</td>
<td>一 0.228</td>
<td>92.28</td>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td>0.228</td>
<td> </td>
</tr>
</tbody>
</table>
Table 1: Coefficients $(\begin{array}{c}\mathscr{V}_0\end{array})$ in regression of $D_{t_0}$
on $t_0$ , and adjusted $R^2$ values $(\begin{array}{c}\%\end{array})$
模型
$$D_{t_0}(w)=\alpha\cdot t_0+\mathrm{const}.$$
并评估系数$\alpha$的分布以及拟合优度值。图 2示例性地显示了英语 1900-2000COHA 数据— 系数$\alpha$几乎对所有单词都是负数。实际上，分布是左偏的，平均值约为 -0.4% 。此外，线性模型始终是数据的良好拟合，因为$R^{2}$值以$85\%$为中心，很少低于 75%。我们发现这里考虑的所有其他数据集的类似模式。这表明，不仅平均单词的含义线性衰减，而且几乎所有单词 (其频率质量超过特定阈值)的语义都以这种方式表现
接下来，我们使用 TISS 图进行查找已发生含义变化的单词的任务。为此，我们按照比率中 maxlink 表示图形$G(w)$中链接的最大权重，而 minlink 是图形$G(w)$中链接的最小权重。我们注意到，弱链接可能表示语义变化， 但规定的比率要求相对于 TISS 图中最强的语
3 | [
[
"0.12177",
"0.0873",
"0.4844",
"0.0873",
"0.4844",
"0.10095",
"0.12177",
"0.10095"
],
[
"0.5366",
"0.0903",
"0.5728",
"0.0903",
"0.5728",
"0.1027",
"0.5366",
"0.1027"
],
[
"0.12177",
"0.1027",
"0.486",
"0.1027",
"0.486",
"0.11755",
"0.12177",
"0.11755"
],
[
"0.711",
"0.0986",
"0.7944",
"0.0986",
"0.7944",
"0.10986",
"0.711",
"0.10986"
],
[
"0.534",
"0.1051",
"0.574",
"0.1051",
"0.574",
"0.12054",
"0.534",
"0.12054"
],
[
"0.1184",
"0.11755",
"0.4844",
"0.11816",
"0.4844",
"0.133",
"0.1184",
"0.1318"
],
[
"0.5356",
"0.12354",
"0.574",
"0.12354",
"0.574",
"0.136",
"0.5356",
"0.136"
],
[
"0.1184",
"0.1342",
"0.482",
"0.1342",
"0.482",
"0.1478",
"0.1184",
"0.1478"
],
[
"0.5356",
"0.1401",
"0.571",
"0.1401",
"0.571",
"0.1526",
"0.5356",
"0.1526"
],
[
"0.12006",
"0.152",
"0.4836",
"0.152",
"0.4836",
"0.165",
"0.12006",
"0.165"
],
[
"0.5356",
"0.158",
"0.5728",
"0.158",
"0.5728",
"0.1704",
"0.5356",
"0.1704"
],
[
"0.12177",
"0.1687",
"0.4844",
"0.1687",
"0.4844",
"0.1818",
"0.12177",
"0.1818"
],
[
"0.5356",
"0.1746",
"0.571",
"0.1746",
"0.571",
"0.187",
"0.5356",
"0.187"
],
[
"0.12006",
"0.1853",
"0.4836",
"0.1853",
"0.4836",
"0.199",
"0.12006",
"0.199"
],
[
"0.5356",
"0.1924",
"0.5728",
"0.1924",
"0.5728",
"0.2048",
"0.5356",
"0.2048"
],
[
"0.12177",
"0.1996",
"0.4836",
"0.1996",
"0.4836",
"0.2144",
"0.12177",
"0.2144"
],
[
"0.5356",
"0.209",
"0.574",
"0.209",
"0.574",
"0.2214",
"0.5356",
"0.2214"
],
[
"0.12006",
"0.2162",
"0.482",
"0.2162",
"0.482",
"0.2299",
"0.12006",
"0.2299"
],
[
"0.534",
"0.2239",
"0.575",
"0.2257",
"0.574",
"0.2411",
"0.532",
"0.2393"
],
[
"0.12006",
"0.234",
"0.482",
"0.234",
"0.482",
"0.2446",
"0.12006",
"0.2446"
],
[
"0.5356",
"0.2423",
"0.5767",
"0.2423",
"0.5767",
"0.2583",
"0.5356",
"0.2583"
],
[
"0.12006",
"0.2488",
"0.4836",
"0.2488",
"0.4836",
"0.2625",
"0.12006",
"0.2625"
],
[
"0.12177",
"0.2642",
"0.4844",
"0.2642",
"0.4844",
"0.2778",
"0.12177",
"0.2778"
],
[
"0.5366",
"0.26",
"0.574",
"0.26",
"0.574",
"0.2725",
"0.5366",
"0.2725"
],
[
"0.642",
"0.2708",
"0.657",
"0.2708",
"0.657",
"0.282",
"0.642",
"0.282"
],
[
"0.708",
"0.2666",
"0.7314",
"0.2708",
"0.727",
"0.284",
"0.7036",
"0.2798"
],
[
"0.7734",
"0.2708",
"0.7983",
"0.2708",
"0.7983",
"0.2832",
"0.7734",
"0.2832"
],
[
"0.8403",
"0.2708",
"0.864",
"0.2708",
"0.864",
"0.2832",
"0.8403",
"0.2832"
],
[
"0.12006",
"0.2798",
"0.4844",
"0.2798",
"0.4844",
"0.2944",
"0.12006",
"0.2944"
],
[
"0.574",
"0.2715",
"0.588",
"0.2715",
"0.588",
"0.2832",
"0.574",
"0.2832"
],
[
"0.12177",
"0.2964",
"0.486",
"0.2964",
"0.486",
"0.31",
"0.12177",
"0.31"
],
[
"0.1184",
"0.3113",
"0.1587",
"0.3113",
"0.1587",
"0.3267",
"0.1184",
"0.3267"
],
[
"0.513",
"0.31",
"0.8774",
"0.31",
"0.8774",
"0.3237",
"0.513",
"0.3237"
],
[
"0.1411",
"0.3289",
"0.4836",
"0.3289",
"0.4836",
"0.3433",
"0.1411",
"0.3433"
],
[
"0.511",
"0.3267",
"0.8716",
"0.3267",
"0.8716",
"0.3403",
"0.511",
"0.3403"
],
[
"0.1184",
"0.3457",
"0.482",
"0.3457",
"0.482",
"0.3594",
"0.1184",
"0.3594"
],
[
"0.5156",
"0.354",
"0.608",
"0.3557",
"0.608",
"0.3677",
"0.5156",
"0.3652"
],
[
"0.638",
"0.3552",
"0.677",
"0.3552",
"0.677",
"0.3665",
"0.638",
"0.3665"
],
[
"0.705",
"0.3528",
"0.7446",
"0.3552",
"0.744",
"0.3677",
"0.7036",
"0.3652"
],
[
"0.79",
"0.3552",
"0.833",
"0.3552",
"0.833",
"0.3665",
"0.79",
"0.3665"
],
[
"0.854",
"0.3557",
"0.8716",
"0.3557",
"0.8716",
"0.3652",
"0.854",
"0.3652"
],
[
"0.12006",
"0.3623",
"0.4844",
"0.3623",
"0.4844",
"0.376",
"0.12006",
"0.376"
],
[
"0.636",
"0.3665",
"0.6753",
"0.3665",
"0.6753",
"0.379",
"0.636",
"0.379"
],
[
"0.12006",
"0.3777",
"0.4844",
"0.3777",
"0.4844",
"0.3914",
"0.12006",
"0.3914"
],
[
"0.638",
"0.3801",
"0.66",
"0.3801",
"0.66",
"0.3901",
"0.638",
"0.3901"
],
[
"0.705",
"0.3884",
"0.742",
"0.3901",
"0.7407",
"0.402",
"0.7036",
"0.3997"
],
[
"0.8423",
"0.386",
"0.8867",
"0.3884",
"0.885",
"0.4038",
"0.8403",
"0.402"
],
[
"0.634",
"0.3896",
"0.69",
"0.3914",
"0.69",
"0.4026",
"0.634",
"0.401"
],
[
"0.1184",
"0.3938",
"0.4836",
"0.3938",
"0.4836",
"0.4067",
"0.1184",
"0.4067"
],
[
"0.5205",
"0.3901",
"0.6265",
"0.3901",
"0.6265",
"0.402",
"0.5205",
"0.402"
],
[
"0.705",
"0.402",
"0.7407",
"0.402",
"0.7407",
"0.4133",
"0.705",
"0.4133"
],
[
"0.79",
"0.402",
"0.8286",
"0.402",
"0.8286",
"0.4146",
"0.79",
"0.4146"
],
[
"0.12006",
"0.4104",
"0.486",
"0.4104",
"0.486",
"0.4233",
"0.12006",
"0.4233"
],
[
"0.706",
"0.4133",
"0.742",
"0.4133",
"0.742",
"0.4258",
"0.706",
"0.4258"
],
[
"0.8423",
"0.411",
"0.885",
"0.4133",
"0.8833",
"0.4287",
"0.8403",
"0.427"
],
[
"0.705",
"0.4258",
"0.7407",
"0.4258",
"0.7407",
"0.4382",
"0.705",
"0.4382"
],
[
"0.79",
"0.4258",
"0.8286",
"0.4258",
"0.8286",
"0.4382",
"0.79",
"0.4382"
],
[
"0.8423",
"0.434",
"0.885",
"0.4358",
"0.8833",
"0.452",
"0.8403",
"0.4495"
],
[
"0.1184",
"0.427",
"0.4844",
"0.427",
"0.4844",
"0.4407",
"0.1184",
"0.4407"
],
[
"0.5205",
"0.437",
"0.544",
"0.437",
"0.544",
"0.4507",
"0.5205",
"0.4507"
],
[
"0.5693",
"0.4382",
"0.629",
"0.4382",
"0.629",
"0.4495",
"0.5693",
"0.4495"
],
[
"0.635",
"0.4382",
"0.669",
"0.4382",
"0.669",
"0.4495",
"0.635",
"0.4495"
],
[
"0.706",
"0.437",
"0.7407",
"0.437",
"0.7407",
"0.4495",
"0.706",
"0.4495"
],
[
"0.7886",
"0.4453",
"0.8296",
"0.4478",
"0.8286",
"0.4631",
"0.7866",
"0.4614"
],
[
"0.705",
"0.449",
"0.7407",
"0.449",
"0.7407",
"0.4614",
"0.705",
"0.4614"
],
[
"0.5205",
"0.4602",
"0.5474",
"0.4602",
"0.5474",
"0.4739",
"0.5205",
"0.4739"
],
[
"0.565",
"0.4602",
"0.6113",
"0.4602",
"0.6113",
"0.4744",
"0.565",
"0.4744"
],
[
"0.636",
"0.4614",
"0.7676",
"0.4614",
"0.7676",
"0.4756",
"0.636",
"0.4756"
],
[
"0.843",
"0.4614",
"0.882",
"0.4614",
"0.882",
"0.4739",
"0.843",
"0.4739"
],
[
"0.7886",
"0.4702",
"0.8296",
"0.4702",
"0.8296",
"0.4863",
"0.7886",
"0.4863"
],
[
"0.11755",
"0.4714",
"0.2688",
"0.4739",
"0.2688",
"0.4868",
"0.11755",
"0.485"
],
[
"0.12006",
"0.4424",
"0.1411",
"0.4424",
"0.1411",
"0.4573",
"0.12006",
"0.4573"
],
[
"0.1184",
"0.4963",
"0.4844",
"0.4963",
"0.4844",
"0.511",
"0.1184",
"0.511"
],
[
"0.511",
"0.4976",
"0.876",
"0.4988",
"0.876",
"0.513",
"0.511",
"0.5127"
],
[
"0.12006",
"0.513",
"0.4836",
"0.513",
"0.4836",
"0.527",
"0.12006",
"0.527"
],
[
"0.511",
"0.514",
"0.8096",
"0.514",
"0.8096",
"0.5293",
"0.511",
"0.5293"
],
[
"0.12177",
"0.5293",
"0.486",
"0.5293",
"0.486",
"0.5435",
"0.12177",
"0.5435"
],
[
"0.11755",
"0.5435",
"0.4878",
"0.5435",
"0.4878",
"0.564",
"0.11755",
"0.564"
],
[
"0.12177",
"0.5625",
"0.4844",
"0.5625",
"0.4844",
"0.5767",
"0.12177",
"0.5767"
],
[
"0.5146",
"0.5586",
"0.55",
"0.5586",
"0.55",
"0.572",
"0.5146",
"0.572"
],
[
"0.12006",
"0.579",
"0.4836",
"0.579",
"0.4836",
"0.593",
"0.12006",
"0.593"
],
[
"0.601",
"0.5845",
"0.79",
"0.5845",
"0.79",
"0.597",
"0.601",
"0.597"
],
[
"0.1184",
"0.5933",
"0.4836",
"0.594",
"0.4836",
"0.6094",
"0.1184",
"0.608"
],
[
"0.1184",
"0.61",
"0.4836",
"0.61",
"0.4836",
"0.6235",
"0.1184",
"0.6235"
],
[
"0.519",
"0.6143",
"0.529",
"0.6143",
"0.529",
"0.619",
"0.519",
"0.619"
],
[
"0.5283",
"0.61",
"0.881",
"0.61",
"0.881",
"0.6235",
"0.5283",
"0.6235"
],
[
"0.1184",
"0.626",
"0.4844",
"0.627",
"0.4844",
"0.641",
"0.1184",
"0.64"
],
[
"0.5146",
"0.627",
"0.8623",
"0.627",
"0.8623",
"0.64",
"0.5146",
"0.64"
],
[
"0.12427",
"0.6436",
"0.4844",
"0.6436",
"0.4844",
"0.6567",
"0.12427",
"0.6567"
],
[
"0.5146",
"0.6436",
"0.8774",
"0.6436",
"0.8774",
"0.6567",
"0.5146",
"0.6567"
],
[
"0.12006",
"0.66",
"0.4844",
"0.66",
"0.4844",
"0.6733",
"0.12006",
"0.6733"
],
[
"0.5146",
"0.659",
"0.879",
"0.659",
"0.879",
"0.673",
"0.5146",
"0.673"
],
[
"0.12006",
"0.676",
"0.486",
"0.676",
"0.486",
"0.6895",
"0.12006",
"0.6895"
],
[
"0.5146",
"0.675",
"0.8774",
"0.675",
"0.8774",
"0.6885",
"0.5146",
"0.6885"
],
[
"0.12177",
"0.6914",
"0.4844",
"0.6904",
"0.4844",
"0.705",
"0.12177",
"0.706"
],
[
"0.513",
"0.6914",
"0.8774",
"0.6914",
"0.8774",
"0.705",
"0.513",
"0.705"
],
[
"0.1184",
"0.706",
"0.4844",
"0.707",
"0.4844",
"0.7217",
"0.1184",
"0.72"
],
[
"0.5146",
"0.707",
"0.879",
"0.707",
"0.879",
"0.72",
"0.5146",
"0.72"
],
[
"0.12006",
"0.7236",
"0.4324",
"0.7236",
"0.4324",
"0.737",
"0.12006",
"0.737"
],
[
"0.513",
"0.7217",
"0.8774",
"0.7217",
"0.8774",
"0.7354",
"0.513",
"0.7354"
],
[
"0.1377",
"0.741",
"0.486",
"0.741",
"0.486",
"0.755",
"0.1377",
"0.755"
],
[
"0.513",
"0.7383",
"0.8774",
"0.7383",
"0.8774",
"0.752",
"0.513",
"0.752"
],
[
"0.12006",
"0.756",
"0.4836",
"0.756",
"0.4836",
"0.7695",
"0.12006",
"0.7695"
],
[
"0.513",
"0.755",
"0.607",
"0.755",
"0.607",
"0.7695",
"0.513",
"0.7695"
],
[
"0.12006",
"0.7725",
"0.4836",
"0.7725",
"0.4836",
"0.786",
"0.12006",
"0.786"
],
[
"0.5283",
"0.771",
"0.879",
"0.7695",
"0.879",
"0.7837",
"0.5283",
"0.785"
],
[
"0.12006",
"0.789",
"0.486",
"0.789",
"0.486",
"0.804",
"0.12006",
"0.804"
],
[
"0.511",
"0.7876",
"0.8774",
"0.7876",
"0.8774",
"0.8003",
"0.511",
"0.8003"
],
[
"0.11755",
"0.804",
"0.4895",
"0.804",
"0.4895",
"0.8247",
"0.11755",
"0.8247"
],
[
"0.509",
"0.8",
"0.885",
"0.8",
"0.885",
"0.8237",
"0.509",
"0.8237"
],
[
"0.12006",
"0.8223",
"0.4785",
"0.8223",
"0.4785",
"0.836",
"0.12006",
"0.836"
],
[
"0.513",
"0.8193",
"0.8774",
"0.8193",
"0.8774",
"0.833",
"0.513",
"0.833"
],
[
"0.1184",
"0.839",
"0.4878",
"0.839",
"0.4878",
"0.854",
"0.1184",
"0.854"
],
[
"0.513",
"0.835",
"0.8774",
"0.835",
"0.8774",
"0.8486",
"0.513",
"0.8486"
],
[
"0.513",
"0.8516",
"0.8774",
"0.8516",
"0.8774",
"0.865",
"0.513",
"0.865"
],
[
"0.1377",
"0.866",
"0.351",
"0.8677",
"0.351",
"0.882",
"0.1377",
"0.8813"
],
[
"0.5146",
"0.8677",
"0.879",
"0.8677",
"0.879",
"0.8813",
"0.5146",
"0.8813"
],
[
"0.4902",
"0.903",
"0.509",
"0.903",
"0.509",
"0.9175",
"0.4902",
"0.9175"
]
] | [
"考虑在每个时间段内至少出现100次的单词。",
"1.00",
"为了诱导每个单词 wEV的单词嵌人 wE",
"1810-2000",
"0.98",
"Rd，我们使用带有默认参数化的word2vec",
"0.96",
"(Mikolovetal.，2013b）。我们这样做每个时",
"0.94",
"间段tET独立。然后，我们使用这些嵌人来",
"0.92",
"导出新的嵌入，如方程（1）。自始至终，我",
"0.90",
"们使用余弦相似度作为sim度量。对于德语",
"0.88",
"我们考虑德国报纸SZ5的专有数据集，其中",
"0.86",
"T=[1994,1995，.，2003}。我们对数据进行",
"0.84",
"lemmatize和POS标记，同样只考虑名词，动",
"0.82",
"词和形容词，使频率约束与英语相同。最后",
"我们使用PL（Migne，1855）作为拉丁语的数",
"0.80",
"5",
"10",
"15",
"20",
"据集。这里，T=[300,400,·.，1300}。我们",
"0",
"使用与英语和德语相同的预处理、频率和词类",
"约束",
"Figure l: Dto (y -axis) as a function of to (a",
"自始至终，我们的数据集在大小方面都很",
"-axis), values of Dto (in green) and error-bars.",
"平衡。例如，从1900年到2000年的每十年",
"CorpusLang",
"Time",
"Years",
"Coeff.",
"R2",
"英语COHA数据集包含大约24M-30M令牌",
"inter-",
"其中1990年和2000年包含的数据比前几十",
"val",
"1900",
"98.63",
"Decade",
"年略多。1900年之前的几十年包含18-24M令",
"COHAEnglish",
"2000",
"0.425",
"牌，只有1810年和1820年包含很少的数据",
"1810-",
"96.03",
"2000",
"0.405",
"98.64",
"（分别为1M和7M令牌）。语料库也由体裁平",
"SZ",
"German",
"Year",
"1994",
"0.678",
"2003",
"PL",
"Latin",
"Century400-1300",
"92.28",
"0.228",
"4.1TISS graphs",
"衡",
"我们从研究TISS图开始。令Dto表示",
"Table l: Coefficients (%) in regression of Dto",
"当时间段之间的距离为to时，平均而",
"on to , and adjusted R? values ( % ).",
"言，单词在两个时间段上的语义相似程度：",
"Dto = wev [s-tl=to sim(w(s), w(t) ,其",
"中 C =[V:K(s,t)Is -t =to}l 是规范器。",
"模型",
"Figure 1 plotsthevalues Dto forthetimeslice-",
"Dt(w) = α+to + const.",
"from1810to2000,fortheEnglishdata。我们注意",
"到一个明显的趋势：单词的自相似性倾向于",
"",
"并评估系数α的分布以及拟合优度值。图2",
"（几乎完美地）随时间距离线性降低。实际上",
"示例性地显示了英语1900-2000COHA数据",
"下面的表1表明，这种趋势在我们所有的语料",
"系数Q几乎对所有单词都是负数。实际上，分",
"库中都成立，即。e。，对于不同的时间尺度和",
"布是左偏的，平均值约为-0.4%。此外，线",
"不同的语言：线性“衰减”模型非常好地拟合",
"性模型始终是数据的良好拟合，因为R2值以",
"了Dto曲线，调整后的R²值大大高于90%",
"85%为中心，很少低于75%。我们发现这里考",
"并且一致且显着为负系数。我们认为，这一发",
"虑的所有其他数据集的类似模式。这表明，不",
"现可能被认为是语义变化的新统计规律",
"仅平均单词的含义线性衰减，而且几乎所有单",
"作为to的函数的值Dto是所有单词的平均",
"词（其频率质量超过特定阈值）的语义都以这",
"值。因此，平均单词的含义可能会随时间线性",
"种方式表现",
"衰减，而随着时间的推移，大部分单词的语义",
"接下来，我们使用TISS图进行查找已发生",
"行为遵循不同的趋势。。为了研究这一点，我",
"含义变化的单词的任务。为此，我们按照比率",
"们考虑 Dto(w)=Z[s-t=to sim(w(s), w(t)",
"Rc(u)=对图形G(w）进行排序，其",
"在固定词w上的分布。这里C=(s,t)1ls-",
"中maxlink表示图形G(w）中链接的最大权",
"t=to。我们独立考虑每个单词w的回归",
"重，而minlink是图形G(w）中链接的最小权",
"重。我们注意到，弱链接可能表示语义变化，",
"http://www.sueddeutsche.de/",
"但规定的比率要求相对于TISS图中最强的语",
"3"
] | {
"arxiv_id": "1704.02497",
"language": "zh",
"page_idx": 2
} |
||
1704.02497-zh-4 | 1704.02497-zh-4.jpg | 负关系方程$\overline{\mathrm{E}}(2)$ 中模型的另一个非常有趣的方面是，它允许检测单词$w_j$ ,其嵌入具有目标单词$w_i$的负系数$\alpha_w_j$ (我们在其余部分中考虑$p=1$ )。这样的负系数可以看作是“微分定律”的实例：随着时间的流逝，两个单词的含义在语义空间中朝相反的方向移动。我们发现以下单词之间存在显着的负相关关系：顶峰$\leftrightarrow$英尺，男孩$\leftrightarrow$女人，誓言$\leftrightarrow$信念，黑人$\leftrightarrow$黑色。我们没有进行详细分析，而是提到最后一对的 Wikipedia 条目表明“黑人”和“黑人”的含义在 20 世纪早期和晚期$^\mathrm{th}$世纪之间切换了角色。虽然“黑人”曾经是美国有色人种的“中性”术语，但在 1960 年代之后， 它获得了负面含义；反之亦然
### 5 Concluding remarks
我们提出了两种新颖的语义变化模型。首先， TISS 图允许检测渐进的、非连续的含义变化。它们能够检测可能具有一般性质的统计趋势。其次，我们的时间序列模型允许调查意义变化的负面趋势(“差异定律”)以及对未来的预测，我们将其留给未来的工作。两种模型都暗 示 了 语 义 变 化 的 线 性 行 为 ,值 得 进 一 步 研 究 。multiple word prototypes. In Proceedings of Wenotethatthislinearityconcernsthecorevocabulatiopafhaphgalą geshalP9HgissfeswePohghapgcariregforedatieast100timesi linearme。g。, 专 有 名 称 指 完 全 不 同 的 实 体 ) 。umee ACT/ 2 nage S73- 882 Strondshuror
linearme。g。,专有名称指完全不同的实体)。但是，我们的分析还发现，大多数核心词的含义随时间线性衰减
Ferdinand de Saussure. 1916. Cours de linguis-
tique générale. Payot, Lausanne/Paris.
Steffen Eger, Niko Schenk, and Alexander Mehler.
2015. Towards semantic language classification: Inducing and clustering semantic association networks from europarl. In Proceedings of the Fourth Joint Conference on Lexical and $\underset{{\mathrm{Computational~Semantics,~pages~127-136,~Den-}}}{\operatorname*{Computational~Semantics,~pages~127-136,~Den-}}$ ver, Colorado, June. Association for Computational Linguistics.
Steffen Eger. 2016. Opinion dynamics and wisdom under out-group discrimination. Mathematical Social Sciences, 80(C):97-107.
Manaal Faruqui and Chris Dyer. 2014. Improving
vector space word representations using multi-
lingual $\dot{\text{correlation. In Proceedings of EACL}} .$
Benjamin Golub and Matthew O. Jackson. 2010.
Naïve learning in social networks and the wisdom of crowds. American Economic Journal: Microeconomics, 2(1):112-49, February.
Eric H. Huang, Richard Socher, Christopher D.
Manning, and Andrew Y. Ng. 2012. Improving word representations via global context and multiple word prototypes. In Proceedings of ume 1, ACL '12, pages 873-882, Stroudsburg, PA, USA. Association for Computational Linguistics.
Adam Jatowt and Kevin Duh. 2014. A framework
for analyzing semantic change of words across time. In Proceedings of the Joint JCDL/TPDL Digital Libraries Conference.
Vivek Kulkarni, Rami Al-Rfou, Bryan Perozzi,
and Steven Skiena. 2015a. Statistically significant detection of linguistic change. In Proceedings of the 24th International Conference on World Wide Web, WWW '15, pages 625- 635, Republic and Canton of Geneva, Switzerland. International World Wide Web Conferences Steering Committee.
Vivek Kulkarni, Bryan Perozzi, and Steven Skiena.
2015b.
Freshman or fresher?
quantifying
the geographic variation of internet language.
CoRR, abs/ 1510. 06786.
Jey Han Lau, Paul Cook, Diana McCarthy, David
Newman, and Timothy Baldwin. 2012. Word sense induction for novel sense detection. In Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, EACL'12, pages $591–601$, Stroudsburg, PA, USA. Association for Computational Linguistics.
# References
Daron Acemoglu and Asuman Ozdaglar. 2011.
Opinion dynamics and learning in social networks. Dynamic Games and Applications, 1(1){:}3{-}49.
Marco Baroni, Georgiana Dinu, and Germán
Kruszewski. 2014. Don't count,predict! a systematic comparison of context-counting vs. context-predicting semantic vectors. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 238-247, Baltimore, Maryland, June. Association for Computational Linguistics.
Xinxiong Chen, Zhiyuan Liu, and Maosong Sun.
2014. A unified model for word sense representation and disambiguation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP 2014, October 25-29, 2014, Doha, Qatar, A meeting of SIGDAT, a Special Interest Group of the ACL, pages 1025-1035.
5 | [
[
"0.1394",
"0.0861",
"0.486",
"0.0861",
"0.486",
"0.0998",
"0.1394",
"0.0998"
],
[
"0.51",
"0.0861",
"0.875",
"0.0861",
"0.875",
"0.0998",
"0.51",
"0.0998"
],
[
"0.12006",
"0.1027",
"0.4844",
"0.1027",
"0.4844",
"0.11755",
"0.12006",
"0.11755"
],
[
"0.529",
"0.10095",
"0.8193",
"0.10095",
"0.8193",
"0.114",
"0.529",
"0.114"
],
[
"0.12177",
"0.1194",
"0.4836",
"0.1194",
"0.4836",
"0.1342",
"0.12177",
"0.1342"
],
[
"0.511",
"0.1247",
"0.879",
"0.1247",
"0.879",
"0.1383",
"0.511",
"0.1383"
],
[
"0.12006",
"0.1354",
"0.4844",
"0.1354",
"0.4844",
"0.1484",
"0.12006",
"0.1484"
],
[
"0.529",
"0.1383",
"0.879",
"0.1383",
"0.879",
"0.1526",
"0.529",
"0.1526"
],
[
"0.12177",
"0.1509",
"0.486",
"0.1509",
"0.486",
"0.165",
"0.12177",
"0.165"
],
[
"0.5283",
"0.1497",
"0.8774",
"0.1509",
"0.8774",
"0.165",
"0.5283",
"0.1644"
],
[
"0.12006",
"0.1675",
"0.4844",
"0.1675",
"0.4844",
"0.1812",
"0.12006",
"0.1812"
],
[
"0.5283",
"0.1633",
"0.881",
"0.1644",
"0.881",
"0.1787",
"0.5283",
"0.1776"
],
[
"0.529",
"0.1776",
"0.881",
"0.1776",
"0.881",
"0.1912",
"0.529",
"0.1912"
],
[
"0.12177",
"0.1841",
"0.4844",
"0.1841",
"0.4844",
"0.1978",
"0.12177",
"0.1978"
],
[
"0.5283",
"0.1895",
"0.879",
"0.1907",
"0.879",
"0.2048",
"0.5283",
"0.2037"
],
[
"0.12006",
"0.2007",
"0.486",
"0.2007",
"0.486",
"0.2144",
"0.12006",
"0.2144"
],
[
"0.529",
"0.2037",
"0.879",
"0.2037",
"0.879",
"0.2173",
"0.529",
"0.2173"
],
[
"0.12177",
"0.2162",
"0.4844",
"0.2162",
"0.4844",
"0.2299",
"0.12177",
"0.2299"
],
[
"0.531",
"0.2185",
"0.663",
"0.2185",
"0.663",
"0.2286",
"0.531",
"0.2286"
],
[
"0.12006",
"0.2322",
"0.486",
"0.2322",
"0.486",
"0.2465",
"0.12006",
"0.2465"
],
[
"0.51",
"0.2405",
"0.881",
"0.2411",
"0.881",
"0.2546",
"0.51",
"0.2542"
],
[
"0.1184",
"0.2477",
"0.486",
"0.2477",
"0.486",
"0.2625",
"0.1184",
"0.2625"
],
[
"0.529",
"0.2542",
"0.881",
"0.2542",
"0.881",
"0.2673",
"0.529",
"0.2673"
],
[
"0.12006",
"0.2642",
"0.4844",
"0.2642",
"0.4844",
"0.2778",
"0.12006",
"0.2778"
],
[
"0.529",
"0.2673",
"0.7524",
"0.2673",
"0.7524",
"0.2808",
"0.529",
"0.2808"
],
[
"0.12177",
"0.2798",
"0.4844",
"0.2798",
"0.4844",
"0.2944",
"0.12177",
"0.2944"
],
[
"0.511",
"0.2893",
"0.881",
"0.2915",
"0.881",
"0.306",
"0.511",
"0.304"
],
[
"0.1184",
"0.2957",
"0.361",
"0.2964",
"0.361",
"0.3113",
"0.1184",
"0.31"
],
[
"0.529",
"0.3047",
"0.879",
"0.3047",
"0.879",
"0.3184",
"0.529",
"0.3184"
],
[
"0.529",
"0.3176",
"0.8623",
"0.3176",
"0.8623",
"0.3308",
"0.529",
"0.3308"
],
[
"0.1276",
"0.3247",
"0.1394",
"0.3333",
"0.1276",
"0.3416",
"0.11584",
"0.3333"
],
[
"0.1511",
"0.3247",
"0.3315",
"0.326",
"0.3315",
"0.3403",
"0.1511",
"0.339"
],
[
"0.513",
"0.3416",
"0.8774",
"0.3416",
"0.8774",
"0.3552",
"0.513",
"0.3552"
],
[
"0.1184",
"0.3516",
"0.4844",
"0.3528",
"0.4844",
"0.3677",
"0.1184",
"0.3665"
],
[
"0.529",
"0.3552",
"0.879",
"0.3552",
"0.879",
"0.3682",
"0.529",
"0.3682"
],
[
"0.1184",
"0.3665",
"0.4895",
"0.3682",
"0.4895",
"0.383",
"0.1184",
"0.3813"
],
[
"0.529",
"0.3682",
"0.881",
"0.3682",
"0.881",
"0.3818",
"0.529",
"0.3818"
],
[
"0.1184",
"0.3843",
"0.4836",
"0.3843",
"0.4836",
"0.398",
"0.1184",
"0.398"
],
[
"0.5283",
"0.3801",
"0.8237",
"0.3813",
"0.8237",
"0.3955",
"0.5283",
"0.3943"
],
[
"0.1184",
"0.401",
"0.4844",
"0.401",
"0.4844",
"0.4146",
"0.1184",
"0.4146"
],
[
"0.511",
"0.405",
"0.8774",
"0.405",
"0.8774",
"0.4192",
"0.511",
"0.4192"
],
[
"0.12177",
"0.4175",
"0.486",
"0.4175",
"0.486",
"0.4312",
"0.12177",
"0.4312"
],
[
"0.5283",
"0.4187",
"0.876",
"0.4187",
"0.876",
"0.4316",
"0.5283",
"0.4316"
],
[
"0.12006",
"0.4329",
"0.4878",
"0.4329",
"0.4878",
"0.4465",
"0.12006",
"0.4465"
],
[
"0.531",
"0.4329",
"0.8774",
"0.4329",
"0.8774",
"0.4436",
"0.531",
"0.4436"
],
[
"0.12006",
"0.449",
"0.4844",
"0.449",
"0.4844",
"0.4631",
"0.12006",
"0.4631"
],
[
"0.529",
"0.4448",
"0.881",
"0.4448",
"0.881",
"0.4578",
"0.529",
"0.4578"
],
[
"0.11584",
"0.4578",
"0.876",
"0.4673",
"0.876",
"0.488",
"0.11584",
"0.4785"
],
[
"0.534",
"0.4578",
"0.873",
"0.4578",
"0.873",
"0.4714",
"0.534",
"0.4714"
],
[
"0.875",
"0.466",
"0.9966",
"0.466",
"0.9966",
"0.4768",
"0.875",
"0.4768"
],
[
"0.1184",
"0.481",
"0.4878",
"0.481",
"0.4878",
"0.4954",
"0.1184",
"0.4954"
],
[
"0.529",
"0.4827",
"0.879",
"0.484",
"0.879",
"0.4988",
"0.529",
"0.4976"
],
[
"0.12006",
"0.4976",
"0.4844",
"0.4976",
"0.4844",
"0.511",
"0.12006",
"0.511"
],
[
"0.531",
"0.4963",
"0.8774",
"0.4963",
"0.8774",
"0.5103",
"0.531",
"0.5103"
],
[
"0.12177",
"0.513",
"0.2688",
"0.513",
"0.2688",
"0.528",
"0.12177",
"0.528"
],
[
"0.531",
"0.511",
"0.592",
"0.511",
"0.592",
"0.5225",
"0.531",
"0.5225"
],
[
"0.5146",
"0.5337",
"0.879",
"0.5337",
"0.879",
"0.5474",
"0.5146",
"0.5474"
],
[
"0.529",
"0.5474",
"0.879",
"0.5474",
"0.879",
"0.561",
"0.529",
"0.561"
],
[
"0.11755",
"0.5557",
"0.2141",
"0.558",
"0.2141",
"0.5723",
"0.11755",
"0.571"
],
[
"0.529",
"0.56",
"0.879",
"0.56",
"0.879",
"0.5737",
"0.529",
"0.5737"
],
[
"0.529",
"0.5737",
"0.744",
"0.5737",
"0.744",
"0.5874",
"0.529",
"0.5874"
],
[
"0.1184",
"0.5806",
"0.486",
"0.5806",
"0.486",
"0.594",
"0.1184",
"0.594"
],
[
"0.1368",
"0.594",
"0.486",
"0.594",
"0.486",
"0.608",
"0.1368",
"0.608"
],
[
"0.511",
"0.597",
"0.881",
"0.597",
"0.881",
"0.611",
"0.511",
"0.611"
],
[
"0.1377",
"0.6094",
"0.1973",
"0.6094",
"0.1973",
"0.6206",
"0.1377",
"0.6206"
],
[
"0.2048",
"0.6055",
"0.4878",
"0.607",
"0.4878",
"0.6216",
"0.2048",
"0.6206"
],
[
"0.531",
"0.6123",
"0.693",
"0.6123",
"0.693",
"0.6226",
"0.531",
"0.6226"
],
[
"0.6885",
"0.611",
"0.876",
"0.6123",
"0.876",
"0.6235",
"0.6885",
"0.6226"
],
[
"0.1394",
"0.6216",
"0.2124",
"0.6216",
"0.2124",
"0.633",
"0.1394",
"0.633"
],
[
"0.531",
"0.626",
"0.876",
"0.626",
"0.876",
"0.6357",
"0.531",
"0.6357"
],
[
"0.529",
"0.637",
"0.881",
"0.637",
"0.881",
"0.651",
"0.529",
"0.651"
],
[
"0.1184",
"0.6465",
"0.4878",
"0.6465",
"0.4878",
"0.66",
"0.1184",
"0.66"
],
[
"0.529",
"0.6494",
"0.8774",
"0.6494",
"0.8774",
"0.663",
"0.529",
"0.663"
],
[
"0.1368",
"0.66",
"0.4878",
"0.66",
"0.4878",
"0.6733",
"0.1368",
"0.6733"
],
[
"0.531",
"0.6646",
"0.876",
"0.6646",
"0.876",
"0.675",
"0.531",
"0.675"
],
[
"0.1368",
"0.673",
"0.486",
"0.673",
"0.486",
"0.686",
"0.1368",
"0.686"
],
[
"0.5283",
"0.676",
"0.879",
"0.676",
"0.879",
"0.6895",
"0.5283",
"0.6895"
],
[
"0.1368",
"0.686",
"0.4265",
"0.686",
"0.4265",
"0.6997",
"0.1368",
"0.6997"
],
[
"0.4207",
"0.687",
"0.4844",
"0.687",
"0.4844",
"0.6978",
"0.4207",
"0.6978"
],
[
"0.529",
"0.6895",
"0.729",
"0.6895",
"0.729",
"0.703",
"0.529",
"0.703"
],
[
"0.1351",
"0.6987",
"0.4878",
"0.6987",
"0.4878",
"0.712",
"0.1351",
"0.712"
],
[
"0.1377",
"0.7114",
"0.4878",
"0.7114",
"0.4878",
"0.7246",
"0.1377",
"0.7246"
],
[
"0.511",
"0.7114",
"0.8774",
"0.7134",
"0.8774",
"0.728",
"0.511",
"0.7256"
],
[
"0.1368",
"0.7246",
"0.4844",
"0.7246",
"0.4844",
"0.7383",
"0.1368",
"0.7383"
],
[
"0.531",
"0.728",
"0.6035",
"0.728",
"0.6035",
"0.7383",
"0.531",
"0.7383"
],
[
"0.596",
"0.7256",
"0.777",
"0.7256",
"0.777",
"0.7407",
"0.596",
"0.7407"
],
[
"0.781",
"0.728",
"0.879",
"0.728",
"0.879",
"0.7393",
"0.781",
"0.7393"
],
[
"0.1335",
"0.7363",
"0.486",
"0.737",
"0.486",
"0.752",
"0.1335",
"0.7505"
],
[
"0.5283",
"0.7383",
"0.879",
"0.7393",
"0.879",
"0.7544",
"0.5283",
"0.753"
],
[
"0.1368",
"0.7505",
"0.288",
"0.7505",
"0.288",
"0.764",
"0.1368",
"0.764"
],
[
"0.529",
"0.752",
"0.706",
"0.752",
"0.706",
"0.7656",
"0.529",
"0.7656"
],
[
"0.1184",
"0.7754",
"0.4836",
"0.7754",
"0.4836",
"0.789",
"0.1184",
"0.789"
],
[
"0.51",
"0.7754",
"0.879",
"0.7754",
"0.879",
"0.789",
"0.51",
"0.789"
],
[
"0.1368",
"0.7915",
"0.4836",
"0.7915",
"0.4836",
"0.802",
"0.1368",
"0.802"
],
[
"0.529",
"0.789",
"0.881",
"0.789",
"0.881",
"0.8027",
"0.529",
"0.8027"
],
[
"0.1368",
"0.8027",
"0.4895",
"0.8027",
"0.4895",
"0.8164",
"0.1368",
"0.8164"
],
[
"0.529",
"0.8027",
"0.881",
"0.8027",
"0.881",
"0.8164",
"0.529",
"0.8164"
],
[
"0.1368",
"0.8164",
"0.4878",
"0.8164",
"0.4878",
"0.83",
"0.1368",
"0.83"
],
[
"0.529",
"0.8164",
"0.879",
"0.8164",
"0.879",
"0.83",
"0.529",
"0.83"
],
[
"0.1351",
"0.829",
"0.4844",
"0.829",
"0.4844",
"0.843",
"0.1351",
"0.843"
],
[
"0.531",
"0.8306",
"0.8774",
"0.8306",
"0.8774",
"0.8413",
"0.531",
"0.8413"
],
[
"0.1351",
"0.8403",
"0.4895",
"0.8413",
"0.4895",
"0.8555",
"0.1351",
"0.855"
],
[
"0.529",
"0.843",
"0.881",
"0.843",
"0.881",
"0.8555",
"0.529",
"0.8555"
],
[
"0.1351",
"0.855",
"0.486",
"0.855",
"0.486",
"0.868",
"0.1351",
"0.868"
],
[
"0.529",
"0.855",
"0.879",
"0.855",
"0.879",
"0.869",
"0.529",
"0.869"
],
[
"0.1368",
"0.869",
"0.266",
"0.868",
"0.266",
"0.88",
"0.1368",
"0.8813"
],
[
"0.529",
"0.868",
"0.678",
"0.868",
"0.678",
"0.882",
"0.529",
"0.882"
],
[
"0.4902",
"0.903",
"0.509",
"0.903",
"0.509",
"0.9175",
"0.4902",
"0.9175"
]
] | [
"负关系方程E（2）中模型的另一个非常有趣",
"Ferdinand de Saussure. 1916. Cours de linguis",
"的方面是，它允许检测单词wi，其嵌人具有",
"tique generale. Payot, Lausanne/Paris.",
"目标单词wi的负系数Qwi（我们在其余部分",
"Steffen Eger, Niko Schenk, and Alexander Mehler.",
"中考虑p=1）。这样的负系数可以看作是“微",
"2015.Towards semantic language classifica-",
"分定律”的实例：随着时间的流逝，两个单词",
"tion: Inducing and clustering semantic associ-",
"的含义在语义空间中朝相反的方向移动。我们",
"ation networks from europarl. In Proceedings",
"of the Fourth Joint Conference on Lexical and",
"发现以下单词之间存在显着的负相关关系：顶",
"Computational Semantics, pages 127-136, Den-",
"峰台英尺，男孩台女人，誓言信念，黑",
"ver, Colorado, June. Association for Computa-",
"人台黑色。我们没有进行详细分析，而是提",
"tional Linguistics.",
"到最后一对的Wikipedia条目表明“黑人”和",
"Steffen Eger. 2016. Opinion dynamics and wisdom",
"“黑人”的含义在20世纪早期和晚期th世纪",
"under out-group discrimination. Mathematical",
"之间切换了角色。虽然“黑人”曾经是美国有",
"Social Sciences, 80(C):97-107.",
"色人种的“中性”术语，但在1960年代之后",
"Manaal Faruqui and Chris Dyer. 2014. Improving",
"它获得了负面含义；反之亦然",
"vector space word representations using multi-",
"lingual correlation. In Proceedings of EACL.",
"5",
"Concluding remarks",
"Benjamin Golub and Matthew O. Jackson. 2010.",
"我们提出了两种新颖的语义变化模型。首先，",
"Naive learning in social networks and the wis-",
"TISS图允许检测渐进的、非连续的含义变化。",
"dom of crowds. American Economic Journal:",
"它们能够检测可能具有一般性质的统计趋势",
"Microeconomics, 2(1):112-49, February.",
"其次，我们的时间序列模型允许调查意义变",
"Eric H. Huang, Richard Socher, Christopher D.",
"化的负面趋势（“差异定律”）以及对未来的",
"Manning, and Andrew Y. Ng. 2012.Improv",
"预测，我们将其留给未来的工作。两种模型都",
"ingwordrepresentationsviaglobal contextand",
"暗示了语义变化的线性行为，值得进一步研究",
"multiple word prototypes.In Proceedings of",
"Wenotethatthislinearityconcernsthecorevocabulationoapatatigswodippgd",
"the 50th Annual Meeting.of the Association fo",
"tleast100timesi",
"linearme。g。，专有名称指完全不同的实体)。",
"ume 1, ACL '12, pages 873-882, Stroudsburg,",
"但是，我们的分析还发现，大多数核心词的含",
"PA, USA. Association for Computational Lin-",
"义随时间线性衰减",
"guistics.",
"Adam Jatowt and Kevin Duh. 2014. A framework",
"for analyzing semantic change of words across",
"References",
"time.In Proceedings of the Joint JCDL/TPDL",
"Digital Libraries Conference.",
"Daron Acemoglu and Asuman Ozdaglar.2011.",
"Opinion dynamics and learning in social net-",
"Vivek Kulkarni, Rami Al-Rfou, Bryan Perozzi,",
"works.",
"Dynamic Games and Applications,",
"and Steven Skiena.",
"2015a.Statistically sig",
"1(1):3-49.",
"nificant detection of linguistic change.In Pro-",
"ceedings of the 24th International Conference",
"Marco Baroni, Georgiana Dinu, and German",
"on World Wide Web, WWW '15, pages 625-",
"Kruszewski.2014.Don't count, predict!a",
"635,Republic and Canton of Geneva,Switzer-",
"systematic comparison of context-counting vs.",
"land. International World Wide Web Confer-",
"context-predicting semantic vectors.",
"In Pro-",
"ences Steering Committee.",
"ceedings of the 52nd Annual Meeting of the",
"Association for Computational Linguistics (Vol-",
"Vivek Kulkarni, Bryan Perozzi, and Steven Skiena.",
"ume 1: Long Papers), pages 238-247, Balti-",
"2015b.",
"Freshman or fresher?",
"quantifying",
"more, Maryland, June. Association for Compu-",
"the geographic variation of internet language.",
"tational Linguistics.",
"CoRR, abs/1510.06786.",
"Xinxiong Chen, Zhiyuan Liu, and Maosong Sun",
"Jey Han Lau, Paul Cook, Diana McCarthy, David",
"2014.A unified model for word sense repre-",
"Newman, and Timothy Baldwin.2012.Word",
"sentation and disambiguation. In Proceedings",
"sense induction for novel sense detection. In",
"of the 2014 Conference on Empirical Methods",
"Proceedings of the 13th Conference of the Eu-",
"in Natural Language Processing, EMNLP 2014",
"ropean Chapter of the Association for Compu-",
"October 25-29, 2014, Doha, Qatar, A meeting of",
"tational Linguistics, EACL '12, pages 591-601,",
"SIGDAT, a Special Interest Group of the ACL,",
"Stroudsburg, PA, USA. Association for Compu-",
"pages 1025-1035.",
"tational Linguistics.",
"5"
] | {
"arxiv_id": "1704.02497",
"language": "zh",
"page_idx": 4
} |
||
2107.06865-zh-2 | 2107.06865-zh-2.jpg | 时间维度中每个节点对应的瞬时膜电位。通过这种方式， SNN 可以获得强大的能力来从图中的聚合信号中提取隐藏特征。总的来说，我们的贡献总结如下：
$\cdot$ We build a general spike-based modeling framework, referred to as Graph SNNs, to support spike propagation and feature affine transformation by rec- $\bar{\text{onciling the graph convolution operation and spik-}}$ ing communication mechanism. The framework is transferable and can accommodate most of graph propagation operations. To our best knowledge, this is the frst work to build a general spike-based framework for processing regular graph tasks in gradient- descent paradigm.
$\bullet$ We propose a spatial-temporal feature normaliza-
tion algorithm (STFN), which can incorporate the temporal neuronal dynamics and coordinate the membrane potential representation with threshold, demonstrating a significant improvement on convergence and performance.
$\bullet$ We instantiate the proposed framework into several
models (GC-SNN and GA-SNN) and demonstrate their performance by multiple experiments on the semi-supervised node classification tasks. We further evaluate the computation costs and demonstrate high-efficiency benefits which may bring new opportunity to unlock SNNs potentials and facilitate graph-structured applications for neuromorphic hardware.
# 2 Graph Spiking Neural Networks
我们主要从以下四个方面介绍我们的 GraphSNN 框架：(1)将图卷积运算纳人尖峰计算范式的方法；(2)选代尖峰消息传递模型和包含两个阶段的整体训练框架；(3) 促进收敛的时空特征归一化；(4)特定尖峰图模型的实例
2.1 Spiking Graph Convolutiom
给定一个属性图$\mathcal{G}=(\mathcal{V},\mathcal{E})$ , $\mathcal{V}$是节点集，$\mathcal{E}$是边集。通$\begin{array}{l}\text{常，图属性可以用邻接矩阵 }A\in\mathbb{R}^{N\times N}\text{ 和节点特征矩阵}\\X\in\mathbb{R}^{N\times C}=[x_1;x_2;...;x_N]\text{ 来描述，其中 }X\text{ 的每一行}\\\text{小车后令十十一个的时行台目 }\text{ 邻战行政府}\end{array}$ 代表每个节点$x_i$的特征向量，邻接矩阵$A_G$满足$A_{ii}=0$ 。假设单节点信号表示为标量向量$x\in\mathbb{R}^C$,则频谱卷积定义为滤波器$\mathcal{F}_\theta=diag(\theta)$与节点信号$x$的乘积，如下所示：
$$\mathcal{F}_\theta\star x=U\mathcal{F}_\theta U^\top x,$$
(1)
其中$\mathcal{F}_\theta$由$\theta\in\mathbb{R}^C,U$的特征向量。$U^\top x$可以看作是$x$的图傅里叶变换，特征值$\Lambda$的对角矩阵可以通过函数$\mathcal{F}_\theta(\Lambda)$进行滤波。由于昂贵的计算开销，一些工作提出使用近似和堆叠的多个非线性层来规避它，最近已成功应用 [Kipf and Welling, 2019]。为了统一尖峰动态中的信息流，我们将初始节点信号$X$编码为二进制分量$\left\{\widetilde{X}_{0},\widetilde{X}_{1},...,\widetilde{X}_{t},...,\widetilde{X}_{T-1}\right\}$,其中$T$ 表示时间窗口的长度。编码方法可以是概率性的，例如遵循伯努利分布或泊松分布，也可以是确定性的，例如定量近似或使用编码层生成尖峰[Wu et al., 2019]。通过这种方式，我们可以将属性图从其初始状态转换为尖峰表达式。我们
使用$\tilde{H}_t^n$来表示在$t$时刻嵌人$n_th$层的尖峰节点，以及$\widetilde{H} _t^0= \widetilde{X} _t$ (波浪号表示尖峰形式的二进制变量)。然后， 可以将时空域中的逐层尖峰图卷积定义为：
$$\widetilde{H}_{t}^{n}=\Phi(\mathcal{G}_{c}(A,\widetilde{H}_{t}^{n-1})W^{n},\widetilde{H}_{t-1}^{n}),$$
(2)
其中$\mathcal{G}_c(A,\widetilde{H})$表示拓扑连接结构上的尖峰特征传播可以通过各种传播方法实现，例如切比雪夫滤波器，1$^{st}$ 模型等。$\Phi(\cdot)$是与历史状态$\widetilde{H}_t-1^n$相关的非线性动态过程。$W^n\in\mathbb{R}^{C^{n-1}\times C^n}$是逐层可训练权重参数矩阵，其中$C^n$表示$n_th$层和$C^0=C$层中尖峰特征的输出维度。我们使用公式(2)表示一个尖峰图卷积层。多层堆叠结构可以形成一个全面的尖峰动力学系统
2.2 Iterative Spiking Message Passing
我们采用泄漏集成和发射 (LIF) 模型作为我们的基本神经元单元，该模型在计算上易于处理且最常用，同时在一定程度上保持了生物保真度。它的神经元动力学可以通过以下方式控制：
$$\tau\frac{dV(t)}{dt}=-(V(t)-V_{reset})+I(t),$$
(3)
其中$V(t)$ 表示时间 t 的神经元膜电位。$\tau$ 是一个时间常数，$I(t)$表示突触前输入，它是通过突触权重和神经元前活动或外部注射刺激的共同作用获得的。当$V(t)$超过某个阈值$V_{th}$时，神经元发出尖峰并将其膜电位重置为$V_{reset}$ , 然后在随后的时间步骤中再次开始积累$V(t)$
尖峰消息传递过程包括信息传播步骤和更新步骤，这两个步骤都运行$T$时间步。令$\widetilde H_t=[\widetilde h_t^0;\widetilde h_t^1;...;\widetilde h_t^{N-1}]\in$ $\mathbb{R} ^{N\times C}$ , 对于节点 $v$ , $\ddot{\text{我 们 为 其 尖 峰 消 息 传 递 提 供 了 一 }}$ 个通用公式：
(4)
$$\widetilde{h}_{t}^{v}=U(\sum_{u\in N(v)}P(\widetilde{h}_{t}^{v},\widetilde{h}_{t}^{u},e_{vu}),\widetilde{h}_{t-1}^{v})),$$
,其中$P(\cdot)$表示相邻节点的尖峰消息聚合，可以通过不同的图卷积运算$\mathcal{G}_c(\cdot)$实现。$U(\cdot)$通过非线性动态系统表示状态更新。$N(v)$ 表示图$\mathcal{G}$ 中节点 $v$ 的所有邻居， $e_{vu}$表示$v$和$u$之间的静态边连接，可以简单地扩展为有向多图形式主义。方程 (2) 是上述通用表达式的具体实现
为了将 LIF 模型容纳到上述范式中，我们使用 Euler 方法将方程 (3) 的一阶微分方程转换为迭代表达式。我们使用衰减因子$\kappa$来表示项$(1-\frac{dt}\tau)$,并将突触前输人$I$ 扩展为 $\sum_jW^j\mathcal{G}_c(A,\widetilde{H}_{t+1}^j)$ 。请注意，这里我们使用图卷积来实现传播步骤$P(\cdot)$,然后通过将$\frac dt\tau$的缩放效果合并到权重项中获得以下公式：
(5)
$$V_{t+1}=\kappa V_{t}+\sum_{j}W^{j}\mathcal{G}_{c}(A,\widetilde{H}_{t+1}^{j}).$$
$\mathcal{G}_c(A,\tilde{H}_{t+1}^j)$表示相应的突触前聚合特征，上标$j$表示突触前索引。通过结合触发和重置机制并假设$V_{reset}= 0$ , 更新方程式可以表示为：
$$\begin{aligned}&V_{t+1}^{n,i}=\kappa V_{t}^{n,i}(1-\widetilde{H}_{t}^{n,i})+\sum_{j}^{l(n)}W^{n,ij}\mathcal{G}_{c}(A,\widetilde{H}_{t+1}^{n-1,j}),\\&\widetilde{H}_{t+1}^{n,i}=g(V_{t+1}^{n,i}-V_{th}),\end{aligned}$$
(7) | [
[
"0.08905",
"0.07135",
"0.4844",
"0.07135",
"0.4844",
"0.0821",
"0.08905",
"0.0821"
],
[
"0.513",
"0.06946",
"0.911",
"0.06946",
"0.911",
"0.0833",
"0.513",
"0.0833"
],
[
"0.08826",
"0.0833",
"0.4844",
"0.0833",
"0.4844",
"0.09784",
"0.08826",
"0.09784"
],
[
"0.513",
"0.084",
"0.91",
"0.0852",
"0.91",
"0.09973",
"0.513",
"0.0985"
],
[
"0.08826",
"0.09784",
"0.3896",
"0.09784",
"0.3896",
"0.1111",
"0.08826",
"0.1111"
],
[
"0.517",
"0.1016",
"0.814",
"0.1016",
"0.814",
"0.11237",
"0.517",
"0.11237"
],
[
"0.10046",
"0.11743",
"0.4836",
"0.11743",
"0.4836",
"0.1307",
"0.10046",
"0.1307"
],
[
"0.597",
"0.1225",
"0.834",
"0.1225",
"0.834",
"0.1364",
"0.597",
"0.1364"
],
[
"0.83",
"0.1276",
"0.834",
"0.1276",
"0.834",
"0.1332",
"0.83",
"0.1332"
],
[
"0.841",
"0.1276",
"0.853",
"0.1276",
"0.853",
"0.1338",
"0.841",
"0.1338"
],
[
"0.8896",
"0.1237",
"0.911",
"0.1237",
"0.911",
"0.1364",
"0.8896",
"0.1364"
],
[
"0.1168",
"0.13",
"0.4836",
"0.1307",
"0.4836",
"0.1459",
"0.1168",
"0.1445"
],
[
"0.1177",
"0.1459",
"0.4836",
"0.1445",
"0.4836",
"0.1578",
"0.1177",
"0.159"
],
[
"0.527",
"0.1465",
"0.9087",
"0.1465",
"0.9087",
"0.1604",
"0.527",
"0.1604"
],
[
"0.11926",
"0.159",
"0.486",
"0.159",
"0.486",
"0.1724",
"0.11926",
"0.1724"
],
[
"0.513",
"0.1604",
"0.91",
"0.159",
"0.91",
"0.1724",
"0.513",
"0.1736"
],
[
"0.1168",
"0.1724",
"0.4885",
"0.1704",
"0.4885",
"0.1843",
"0.1168",
"0.1863"
],
[
"0.513",
"0.1768",
"0.9087",
"0.1768",
"0.9087",
"0.1875",
"0.513",
"0.1875"
],
[
"0.11926",
"0.1863",
"0.4844",
"0.1875",
"0.4844",
"0.1989",
"0.11926",
"0.1976"
],
[
"0.5107",
"0.1881",
"0.915",
"0.1901",
"0.915",
"0.2108",
"0.5107",
"0.209"
],
[
"0.11926",
"0.2008",
"0.4836",
"0.2008",
"0.4836",
"0.214",
"0.11926",
"0.214"
],
[
"0.1177",
"0.2122",
"0.486",
"0.2134",
"0.486",
"0.2279",
"0.1177",
"0.2267"
],
[
"0.5146",
"0.2084",
"0.911",
"0.2084",
"0.911",
"0.2229",
"0.5146",
"0.2229"
],
[
"0.11926",
"0.2279",
"0.486",
"0.2279",
"0.486",
"0.2412",
"0.11926",
"0.2412"
],
[
"0.513",
"0.2229",
"0.9126",
"0.2229",
"0.9126",
"0.2361",
"0.513",
"0.2361"
],
[
"0.513",
"0.2349",
"0.7915",
"0.2361",
"0.7915",
"0.2507",
"0.513",
"0.25"
],
[
"0.1177",
"0.2424",
"0.3145",
"0.2424",
"0.3145",
"0.2556",
"0.1177",
"0.2556"
],
[
"0.1152",
"0.2615",
"0.4844",
"0.2615",
"0.4844",
"0.2747",
"0.1152",
"0.2747"
],
[
"0.513",
"0.26",
"0.826",
"0.26",
"0.826",
"0.2734",
"0.513",
"0.2734"
],
[
"0.1177",
"0.2747",
"0.486",
"0.2747",
"0.486",
"0.2886",
"0.1177",
"0.2886"
],
[
"0.10455",
"0.2632",
"0.1209",
"0.2632",
"0.1209",
"0.2715",
"0.10455",
"0.2715"
],
[
"0.5146",
"0.2766",
"0.911",
"0.2766",
"0.911",
"0.2917",
"0.5146",
"0.2917"
],
[
"0.1177",
"0.289",
"0.486",
"0.289",
"0.486",
"0.303",
"0.1177",
"0.303"
],
[
"0.513",
"0.2917",
"0.911",
"0.2917",
"0.911",
"0.305",
"0.513",
"0.305"
],
[
"0.11926",
"0.303",
"0.4844",
"0.303",
"0.4844",
"0.3164",
"0.11926",
"0.3164"
],
[
"0.5195",
"0.3093",
"0.527",
"0.3093",
"0.527",
"0.3125",
"0.5195",
"0.3125"
],
[
"0.523",
"0.3062",
"0.911",
"0.3062",
"0.911",
"0.3193",
"0.523",
"0.3193"
],
[
"0.11926",
"0.3176",
"0.486",
"0.3176",
"0.486",
"0.3308",
"0.11926",
"0.3308"
],
[
"0.513",
"0.3193",
"0.6543",
"0.3193",
"0.6543",
"0.3333",
"0.513",
"0.3333"
],
[
"0.1177",
"0.332",
"0.2893",
"0.3308",
"0.2893",
"0.3423",
"0.1177",
"0.3435"
],
[
"0.5933",
"0.332",
"0.846",
"0.3452",
"0.8433",
"0.3738",
"0.591",
"0.3599"
],
[
"0.10297",
"0.3499",
"0.4844",
"0.3499",
"0.4844",
"0.363",
"0.10297",
"0.363"
],
[
"0.8447",
"0.3516",
"0.864",
"0.3516",
"0.864",
"0.3591",
"0.8447",
"0.3591"
],
[
"0.888",
"0.3484",
"0.9126",
"0.3484",
"0.9126",
"0.361",
"0.888",
"0.361"
],
[
"0.601",
"0.3567",
"0.609",
"0.3567",
"0.609",
"0.3599",
"0.601",
"0.3599"
],
[
"0.6055",
"0.3542",
"0.6406",
"0.356",
"0.6396",
"0.3687",
"0.6045",
"0.3662"
],
[
"0.1177",
"0.363",
"0.486",
"0.363",
"0.486",
"0.377",
"0.1177",
"0.377"
],
[
"0.513",
"0.3726",
"0.91",
"0.3726",
"0.91",
"0.3865",
"0.513",
"0.3865"
],
[
"0.11926",
"0.379",
"0.4844",
"0.379",
"0.4844",
"0.3894",
"0.11926",
"0.3894"
],
[
"0.1177",
"0.3914",
"0.4836",
"0.3901",
"0.4836",
"0.404",
"0.1177",
"0.4048"
],
[
"0.513",
"0.3865",
"0.911",
"0.3865",
"0.911",
"0.3997",
"0.513",
"0.3997"
],
[
"0.1177",
"0.404",
"0.486",
"0.4048",
"0.486",
"0.4192",
"0.1177",
"0.4185"
],
[
"0.5146",
"0.401",
"0.911",
"0.401",
"0.911",
"0.414",
"0.5146",
"0.414"
],
[
"0.1177",
"0.4185",
"0.4844",
"0.4185",
"0.4844",
"0.432",
"0.1177",
"0.432"
],
[
"0.513",
"0.4153",
"0.91",
"0.4153",
"0.91",
"0.4287",
"0.513",
"0.4287"
],
[
"0.11926",
"0.433",
"0.4836",
"0.433",
"0.4836",
"0.4463",
"0.11926",
"0.4463"
],
[
"0.513",
"0.4287",
"0.8896",
"0.4287",
"0.8896",
"0.443",
"0.513",
"0.443"
],
[
"0.1168",
"0.4458",
"0.486",
"0.4475",
"0.486",
"0.4622",
"0.1168",
"0.4602"
],
[
"0.5312",
"0.4443",
"0.91",
"0.4443",
"0.91",
"0.455",
"0.5312",
"0.455"
],
[
"0.1177",
"0.461",
"0.2263",
"0.461",
"0.2263",
"0.4749",
"0.1177",
"0.4749"
],
[
"0.513",
"0.457",
"0.911",
"0.455",
"0.911",
"0.4717",
"0.513",
"0.4734"
],
[
"0.513",
"0.4717",
"0.5557",
"0.4666",
"0.558",
"0.4792",
"0.5156",
"0.4841"
],
[
"0.5767",
"0.473",
"0.903",
"0.473",
"0.903",
"0.486",
"0.5767",
"0.486"
],
[
"0.5156",
"0.4873",
"0.6045",
"0.4873",
"0.6045",
"0.4988",
"0.5156",
"0.4988"
],
[
"0.0825",
"0.4937",
"0.3938",
"0.4944",
"0.3938",
"0.512",
"0.0825",
"0.511"
],
[
"0.58",
"0.506",
"0.851",
"0.5083",
"0.851",
"0.526",
"0.58",
"0.5234"
],
[
"0.8447",
"0.5146",
"0.886",
"0.5146",
"0.886",
"0.5215",
"0.8447",
"0.5215"
],
[
"0.884",
"0.511",
"0.9126",
"0.511",
"0.9126",
"0.523",
"0.884",
"0.523"
],
[
"0.08905",
"0.518",
"0.4836",
"0.518",
"0.4836",
"0.529",
"0.08905",
"0.529"
],
[
"0.0866",
"0.529",
"0.4844",
"0.53",
"0.4844",
"0.5444",
"0.0866",
"0.5435"
],
[
"0.6396",
"0.528",
"0.689",
"0.528",
"0.689",
"0.539",
"0.6396",
"0.539"
],
[
"0.08905",
"0.5444",
"0.4844",
"0.5444",
"0.4844",
"0.558",
"0.08905",
"0.558"
],
[
"0.5156",
"0.5444",
"0.911",
"0.5454",
"0.911",
"0.56",
"0.5156",
"0.5596"
],
[
"0.08905",
"0.5596",
"0.4878",
"0.5596",
"0.4878",
"0.5728",
"0.08905",
"0.5728"
],
[
"0.5146",
"0.56",
"0.911",
"0.56",
"0.911",
"0.5737",
"0.5146",
"0.5737"
],
[
"0.08905",
"0.5737",
"0.10455",
"0.5737",
"0.10455",
"0.584",
"0.08905",
"0.584"
],
[
"0.5146",
"0.5737",
"0.914",
"0.5737",
"0.914",
"0.5884",
"0.5146",
"0.5884"
],
[
"0.513",
"0.5884",
"0.9126",
"0.5884",
"0.9126",
"0.6016",
"0.513",
"0.6016"
],
[
"0.08496",
"0.5986",
"0.3489",
"0.5986",
"0.3489",
"0.6123",
"0.08496",
"0.6123"
],
[
"0.5146",
"0.6016",
"0.911",
"0.6016",
"0.911",
"0.6157",
"0.5146",
"0.6157"
],
[
"0.08905",
"0.6187",
"0.4844",
"0.6187",
"0.4844",
"0.629",
"0.08905",
"0.629"
],
[
"0.5146",
"0.616",
"0.546",
"0.616",
"0.546",
"0.6274",
"0.5146",
"0.6274"
],
[
"0.0866",
"0.6274",
"0.486",
"0.6274",
"0.486",
"0.646",
"0.0866",
"0.646"
],
[
"0.5312",
"0.629",
"0.911",
"0.629",
"0.911",
"0.6426",
"0.5312",
"0.6426"
],
[
"0.0825",
"0.6396",
"0.4878",
"0.643",
"0.4878",
"0.6636",
"0.0825",
"0.659"
],
[
"0.5146",
"0.643",
"0.9126",
"0.643",
"0.9126",
"0.657",
"0.5146",
"0.657"
],
[
"0.0866",
"0.657",
"0.486",
"0.6577",
"0.486",
"0.6724",
"0.0866",
"0.672"
],
[
"0.513",
"0.657",
"0.91",
"0.657",
"0.91",
"0.672",
"0.513",
"0.672"
],
[
"0.1013",
"0.672",
"0.4043",
"0.6704",
"0.4043",
"0.685",
"0.1013",
"0.686"
],
[
"0.415",
"0.674",
"0.4836",
"0.674",
"0.4836",
"0.685",
"0.415",
"0.685"
],
[
"0.5146",
"0.677",
"0.9126",
"0.677",
"0.9126",
"0.69",
"0.5146",
"0.69"
],
[
"0.0866",
"0.685",
"0.486",
"0.686",
"0.486",
"0.7007",
"0.0866",
"0.6997"
],
[
"0.5156",
"0.6943",
"0.911",
"0.6943",
"0.911",
"0.7085",
"0.5156",
"0.7085"
],
[
"0.08905",
"0.702",
"0.1438",
"0.702",
"0.1438",
"0.7134",
"0.08905",
"0.7134"
],
[
"0.2157",
"0.7134",
"0.3545",
"0.7134",
"0.3545",
"0.728",
"0.2157",
"0.728"
],
[
"0.4631",
"0.7153",
"0.486",
"0.7153",
"0.486",
"0.728",
"0.4631",
"0.728"
],
[
"0.5146",
"0.7104",
"0.7656",
"0.7104",
"0.7656",
"0.7236",
"0.5146",
"0.7236"
],
[
"0.08826",
"0.733",
"0.4844",
"0.733",
"0.4844",
"0.7476",
"0.08826",
"0.7476"
],
[
"0.5923",
"0.731",
"0.8325",
"0.732",
"0.8325",
"0.75",
"0.5923",
"0.7485"
],
[
"0.8896",
"0.7344",
"0.9126",
"0.7344",
"0.9126",
"0.7476",
"0.8896",
"0.7476"
],
[
"0.08826",
"0.7485",
"0.4844",
"0.7485",
"0.4844",
"0.762",
"0.08826",
"0.762"
],
[
"0.697",
"0.752",
"0.71",
"0.752",
"0.71",
"0.762",
"0.697",
"0.762"
],
[
"0.08826",
"0.761",
"0.4836",
"0.761",
"0.4836",
"0.7744",
"0.08826",
"0.7744"
],
[
"0.08496",
"0.7744",
"0.4844",
"0.776",
"0.4844",
"0.7905",
"0.08496",
"0.789"
],
[
"0.511",
"0.771",
"0.911",
"0.7715",
"0.911",
"0.786",
"0.511",
"0.785"
],
[
"0.08905",
"0.7905",
"0.486",
"0.7905",
"0.486",
"0.8037",
"0.08905",
"0.8037"
],
[
"0.511",
"0.786",
"0.914",
"0.787",
"0.914",
"0.802",
"0.511",
"0.8003"
],
[
"0.08905",
"0.8037",
"0.4844",
"0.8037",
"0.4844",
"0.8174",
"0.08905",
"0.8174"
],
[
"0.513",
"0.8003",
"0.683",
"0.8003",
"0.683",
"0.8145",
"0.513",
"0.8145"
],
[
"0.08905",
"0.8193",
"0.4844",
"0.8193",
"0.4844",
"0.8335",
"0.08905",
"0.8335"
],
[
"0.71",
"0.8213",
"0.734",
"0.8213",
"0.734",
"0.834",
"0.71",
"0.834"
],
[
"0.5195",
"0.832",
"0.7183",
"0.829",
"0.7183",
"0.851",
"0.5195",
"0.854"
],
[
"0.7017",
"0.8306",
"0.915",
"0.829",
"0.915",
"0.854",
"0.7017",
"0.856"
],
[
"0.08826",
"0.8477",
"0.4844",
"0.8477",
"0.4844",
"0.8613",
"0.08826",
"0.8613"
],
[
"0.08905",
"0.834",
"0.4836",
"0.834",
"0.4836",
"0.8477",
"0.08905",
"0.8477"
],
[
"0.714",
"0.854",
"0.7295",
"0.854",
"0.7295",
"0.8643",
"0.714",
"0.8643"
],
[
"0.08826",
"0.8613",
"0.486",
"0.8613",
"0.486",
"0.8755",
"0.08826",
"0.8755"
],
[
"0.5215",
"0.867",
"0.683",
"0.8706",
"0.683",
"0.8896",
"0.5215",
"0.886"
],
[
"0.08826",
"0.8755",
"0.4844",
"0.8755",
"0.4844",
"0.8896",
"0.08826",
"0.8896"
],
[
"0.8906",
"0.872",
"0.9126",
"0.872",
"0.9126",
"0.8853",
"0.8906",
"0.8853"
]
] | [
"时间维度中每个节点对应的瞬时膜电位。通过这种方式",
"使用Hn来表示在t时刻嵌人nth层的尖峰节点，以及",
"SNN可以获得强大的能力来从图中的聚合信号中提取隐",
"HO=X（波浪号表示尖峰形式的二进制变量）。然后",
"藏特征。总的来说，我们的贡献总结如下：",
"可以将时空域中的逐层尖峰图卷积定义为：",
":We build a general spike-based modeling framework",
"Hn = d(Ge(A, Hn-1)Wn, Hn-,),",
"",
"",
"(2)",
"referred to as Graph SNNs, to support spike prop-",
"agation and feature affine transformation by rec-",
"其中G（A,H）表示拓扑连接结构上的尖峰特征传播",
"onciling the graph convolution operation and spik-",
"可以通过各种传播方法实现，例如切比雪夫滤波器，1st",
"ing communication mechanism. The framework is",
"模型等。Φ(）是与历史状态H1相关的非线性动态过",
"transferable and can accommodate most of graph",
"程。Wne RC\"-\"xC\"是逐层可训练权重参数矩阵，其",
"propagation operations.To our best knowledge",
"this is the first work to build a general spike-based",
"中Cn表示nth层和Co=C层中尖峰特征的输出维度。",
"framework for processing regular graph tasks in",
"我们使用公式（2）表示一个尖峰图卷积层。多层堆叠结",
"构可以形成一个全面的尖峰动力学系统",
"gradient-descent paradigm.",
"We propose a spatial-temporal feature normaliza-",
"2.2Iterative Spiking Message Passing",
"tion algorithm (STFN), which can incorporate the",
"",
"我们采用泄漏集成和发射（LIF）模型作为我们的基本神",
"temporal neuronal dymamics and coordinate the",
"经元单元，该模型在计算上易于处理且最常用，同时在",
"membrane potential representation with threshold.",
"",
"一定程度上保持了生物保真度。它的神经元动力学可以",
"demonstrating a significant improvement on conver-",
"通过以下方式控制：",
"gence and performance.",
"-dV() = -(V(t)- Vee)+ I(t),",
": We instantiate the proposed framework into several",
"",
"(3)",
"",
"dt",
"models (GC-SNN and GA-SNN) and demonstrate",
"其中V(t)表示时间t的神经元膜电位。T是一个时间常",
"their performance by multiple experiments onthe",
"semi-supervised node classification tasks. We fur-",
"数，I(t）表示突触前输入，它是通过突触权重和神经元",
"ther evaluate the computation costs and demon-",
"前活动或外部注射刺激的共同作用获得的。当V(t）超过",
"strate high-efficiency benefits which may bring new",
"某个阈值Vth时，神经元发出尖峰并将其膜电位重置为",
"opportunity to unlock SNNs potentials and facil",
"Vreset，然后在随后的时间步骤中再次开始积累V(t)",
"itate graph-structured applications for neuromor-",
"尖峰消息传递过程包括信息传播步骤和更新步骤，这",
"phic hardware.",
"两个步骤都运行T时间步。令H,=[ho;hl..hN-1e",
"RNxC",
"对于节点，我们为其尖峰消息传递提供了",
"个通用公式：",
"2Graph Spiking Neural Networks",
"hy =U( P(h,ht,euu),ht-1),",
"",
"(4)",
"我们主要从以下四个方面介绍我们的GraphSNN框",
"架：（1）将图卷积运算纳人尖峰计算范式的方法；（2）迭代",
"uEN(U)",
"尖峰消息传递模型和包含两个阶段的整体训练框架；（3）",
"，其中P(）表示相邻节点的尖峰消息聚合，可以通过",
"促进收敛的时空特征归一化；（4）特定尖峰图模型的实",
"不同的图卷积运算G(）实现。U(）通过非线性动态系",
"例",
"统表示状态更新。N(）表示图g中节点的所有邻居，",
"euu表示u和u之间的静态边连接，可以简单地扩展为",
"2.1Spiking Graph Convolution",
"有向多图形式主义。方程（2）是上述通用表达式的具体",
"给定一个属性图G=V.8），V是节点集，8是边集。通",
"实现",
"常，图属性可以用邻接矩阵AERNXN和节点特征矩阵",
"为了将LIF模型容纳到上述范式中，我们使用Euler",
"XRNxC=[12aN]来描述，其中X的每一行",
"方法将方程（3）的一阶微分方程转换为选代表达式。我",
"代表每个节点ai的特征向量，邻接矩阵A满足Ai=0",
"们使用衰减因子来表示项（1-），并将突触前输",
"假设单节点信号表示为标量向量EERC\"",
"则频谱卷",
"人I扩展为又，Wigc（A,H+）。请注意，这里我们使",
"积定义为滤波器F。=diag(）与节点信号α的乘积，如",
"用图卷积来实现传播步骤P(），然后通过将的缩放",
"下所示：",
"Fe★T=UFeUTr.",
"(1)",
"效果合并到权重项中获得以下公式：",
"其中F由ERC，U的特征向量。UT可以看作",
"V+1 = rVt + Wige(A, Hi+1).",
"(5)",
"是的图傅里叶变换，特征值A的对角矩阵可以通过",
"j",
"函数Fe(A）进行滤波。由于昂贵的计算开销，一些工作",
"提出使用近似和堆叠的多个非线性层来规避它，最近已",
"9c（A,H1）表示相应的突触前聚合特征，上标j表示突",
"成功应用[KipfandWelling，2019]。为了统一尖峰动态",
"触前索引。通过结合触发和重置机制并假设Vreset=0，",
"中的信息流，我们将初始节点信号X编码为二进制分",
"更新方程式可以表示为：",
"量Xo,X1.,Xt.XT-i}其中T表示时间窗口的",
"l(n)",
"Vi = rvn(1 - Hn) +)",
"Zwn.ig.(A, H-1), (6)",
"或泊松分布，也可以是确定性的，例如定量近似或使用",
"长度。编码方法可以是概率性的，例如遵循伯努利分布",
"5",
"编码层生成尖峰[Wu et al.,2019]。通过这种方式，我",
"Hi = g(V+i - Vth),",
"们可以将属性图从其初始状态转换为尖峰表达式。我们",
"(7)"
] | {
"arxiv_id": "2107.06865",
"language": "zh",
"page_idx": 2
} |
||
2305.11490-zh-0 | 2305.11490-zh-0.jpg | $\underline{\text{Published as a conference paper at ICLR 2024}}$
LLM-CXR: 用于 CXR 映像理解和生成的指令微调
LLM
Suhyeon Lee$^*$,Won Jun Kim$^*$,Jinho Chang &Jong Chul Ye
Korea Advanced Institute of Science & Technology,
$\{$ suhyeon.lee, wonjun, jinhojsk515, jong.ye $\}$ @kaist.ac.kr
## $\operatorname{Abstract}$
Following the impressive development of LLMs, vision-language alignment in LLMs is actively being researched to enable multimodal reasoning and visual $\begin{array}{c}\mathrm{input/output.~Thisdirectionofresearchisparticularlyrelevanttomedicalimag-}\\\mathrm{input/output.~Thisdirectionofresearchisparticularlyrelevanttomedicalimag-}\end{array}$ $\operatorname{ing}$becauseaccuratemedicalimageanalysisandgenerationconsistofreasoning based on a combination of visual features and prior knowledge. Many recent works have focused on training adapter networks that serve as an information bridge between image processing (encoding or generating) networks and LLMs; but presumably, in order to achieve maximum reasoning potential of LLMs on visualinformationaswellaslanguage,imageandtextfeaturesshouldbeallowed to interact more freely. This is especially important in the medical domain becauseunderstandingandgeneratingmedicalimagessuchaschestX-rays(CXR) require not only accurate visual and language-based reasoning but also a more intimate mapping between the two modalities. Thus, taking inspiration from previous work on the transformer and VQ-GAN combination for bidirectional $\operatorname{image}$and$\operatorname{text}$generation, we build upon this approach and develop a method for instruction-tuning an LLM pre-trained only on text togain vision-language capabilities for medical images. Specifically, we leverage a pretrained LLM' s existing question-answering and instruction-following abilities to teach it to understand visual inputs by instructing it to answer questions about image $\begin{array}{c}\mathrm{inputsand,symmetrically,outputbothtextandimageresponsesappropriateto}\\\end{array}$ agivenquerybytuningtheLLMwithdiversetasksthatencompassimage-based text-generation and text-based image-generation. We show that our model, LLM-CXR,trainedinthisapproachshowsbetterimage-textalignmentinboth CXR understanding and generation tasks while being smaller in size compared to previously developed models that perform a narrower range oftasks
1 Introduction
最近几年，大型语言模型 (LLM)领域取得了令人瞩目的发展。LLM被认为是另一类AI 模型，因为它们能够灵活地理解/生成自然语言并执行基于语言的推理，从而使它们能够概括为各种给定任务，而无需对其进行明确的培训。作为下一步，正在积极开发在LLME (OpenAI,2023;Liu etal.,2023;Alayrac et al.,2022;Liet al.,2023)中与语言一起输人视觉信息的方法以及从 LLME(Kohet al.,2023a;b) 输出图像的方法。这些模型在医学领域具有很大的潜力，特别有用，因为处理诸如胸部X射线(CXR)之类的医学图像需要能够理解上下文， 执行推理并以图像和文本形式传达结论。第一代医学多式联运 LLM 最近开始出现 (Moor etal.,2023;Thawkaretal.,2023;Xuetal.,2023a)
开发这些模型的主要挑战是实现 LLM 的预训练语言功能与新引人的图像功能之间的对齐， 而不会灾难性地忘记其语言能力。与自然图像相比，这在医学图像领域是一个更困难的挑战，因为模型需要区分图像甚至图像部分的细微差异(e。g。、肺炎vs。CXR上的肺水肿),
${}^*$These authors contributed equally to this work.
$^*$Inorder tocomply with the MIMIC-CXR data usage license (Johnson et al..,2019a),all CXR images presented in the Figures 2,3,4,8 are replaced with similar CXR's from the Indiana University chest X-ray dataset (Demner-Fushman et al., 2016); and the presented MIMIC text reports are paraphrased.
1 | [
[
"0.1765",
"0.0341",
"0.4836",
"0.0341",
"0.4836",
"0.04736",
"0.1765",
"0.04736"
],
[
"0.1732",
"0.1042",
"0.8203",
"0.1048",
"0.8203",
"0.1237",
"0.1732",
"0.1225"
],
[
"0.1732",
"0.1288",
"0.2336",
"0.1288",
"0.2336",
"0.1477",
"0.1732",
"0.1477"
],
[
"0.1847",
"0.1736",
"0.5947",
"0.1736",
"0.5947",
"0.1875",
"0.1847",
"0.1875"
],
[
"0.183",
"0.1875",
"0.536",
"0.1881",
"0.536",
"0.2026",
"0.183",
"0.202"
],
[
"0.1854",
"0.2026",
"0.5605",
"0.2026",
"0.5605",
"0.2166",
"0.1854",
"0.2166"
],
[
"0.4631",
"0.2524",
"0.536",
"0.2524",
"0.536",
"0.2676",
"0.4631",
"0.2676"
],
[
"0.2336",
"0.2854",
"0.7656",
"0.2854",
"0.7656",
"0.2986",
"0.2336",
"0.2986"
],
[
"0.232",
"0.2974",
"0.7656",
"0.2986",
"0.7656",
"0.3132",
"0.232",
"0.3125"
],
[
"0.232",
"0.3125",
"0.7646",
"0.3132",
"0.7646",
"0.3276",
"0.232",
"0.327"
],
[
"0.232",
"0.3257",
"0.7656",
"0.327",
"0.7656",
"0.3416",
"0.232",
"0.3403"
],
[
"0.2336",
"0.3416",
"0.7646",
"0.3416",
"0.7646",
"0.3547",
"0.2336",
"0.3547"
],
[
"0.2336",
"0.3547",
"0.7656",
"0.3547",
"0.7656",
"0.3687",
"0.2336",
"0.3687"
],
[
"0.2336",
"0.3687",
"0.7656",
"0.3687",
"0.7656",
"0.3833",
"0.2336",
"0.3833"
],
[
"0.2336",
"0.3833",
"0.7646",
"0.3833",
"0.7646",
"0.3965",
"0.2336",
"0.3965"
],
[
"0.2336",
"0.3965",
"0.7656",
"0.3965",
"0.7656",
"0.411",
"0.2336",
"0.411"
],
[
"0.2336",
"0.4104",
"0.763",
"0.4104",
"0.763",
"0.4236",
"0.2336",
"0.4236"
],
[
"0.232",
"0.4236",
"0.7646",
"0.4224",
"0.7646",
"0.4375",
"0.232",
"0.4382"
],
[
"0.2336",
"0.4382",
"0.7656",
"0.4382",
"0.7656",
"0.452",
"0.2336",
"0.452"
],
[
"0.2336",
"0.452",
"0.7656",
"0.452",
"0.7656",
"0.4653",
"0.2336",
"0.4653"
],
[
"0.232",
"0.4653",
"0.7646",
"0.4653",
"0.7646",
"0.4792",
"0.232",
"0.4792"
],
[
"0.2336",
"0.4797",
"0.7656",
"0.4797",
"0.7656",
"0.4937",
"0.2336",
"0.4937"
],
[
"0.232",
"0.4924",
"0.7656",
"0.4937",
"0.7656",
"0.507",
"0.232",
"0.506"
],
[
"0.2345",
"0.507",
"0.7656",
"0.507",
"0.7656",
"0.5215",
"0.2345",
"0.5215"
],
[
"0.2336",
"0.5215",
"0.7656",
"0.5215",
"0.7656",
"0.535",
"0.2336",
"0.535"
],
[
"0.2345",
"0.535",
"0.7646",
"0.535",
"0.7646",
"0.549",
"0.2345",
"0.549"
],
[
"0.232",
"0.5474",
"0.7646",
"0.549",
"0.7646",
"0.563",
"0.232",
"0.5625"
],
[
"0.2336",
"0.5625",
"0.7656",
"0.5625",
"0.7656",
"0.577",
"0.2336",
"0.577"
],
[
"0.2336",
"0.5757",
"0.763",
"0.5757",
"0.763",
"0.589",
"0.2336",
"0.589"
],
[
"0.232",
"0.5884",
"0.7646",
"0.589",
"0.7646",
"0.604",
"0.232",
"0.603"
],
[
"0.2345",
"0.604",
"0.7656",
"0.604",
"0.7656",
"0.617",
"0.2345",
"0.617"
],
[
"0.2345",
"0.619",
"0.7075",
"0.619",
"0.7075",
"0.63",
"0.2345",
"0.63"
],
[
"0.1724",
"0.6655",
"0.3145",
"0.6665",
"0.3145",
"0.681",
"0.1724",
"0.6797"
],
[
"0.1748",
"0.6963",
"0.824",
"0.696",
"0.824",
"0.7104",
"0.1748",
"0.711"
],
[
"0.1765",
"0.712",
"0.8228",
"0.712",
"0.8228",
"0.726",
"0.1765",
"0.726"
],
[
"0.1765",
"0.7246",
"0.824",
"0.7246",
"0.824",
"0.7393",
"0.1765",
"0.7393"
],
[
"0.1765",
"0.7393",
"0.822",
"0.7393",
"0.822",
"0.7524",
"0.1765",
"0.7524"
],
[
"0.1765",
"0.7524",
"0.822",
"0.7524",
"0.822",
"0.7666",
"0.1765",
"0.7666"
],
[
"0.1748",
"0.7676",
"0.822",
"0.7676",
"0.822",
"0.781",
"0.1748",
"0.781"
],
[
"0.1765",
"0.782",
"0.8228",
"0.782",
"0.8228",
"0.7954",
"0.1765",
"0.7954"
],
[
"0.1748",
"0.7954",
"0.5063",
"0.7954",
"0.5063",
"0.8096",
"0.1748",
"0.8096"
],
[
"0.1748",
"0.8164",
"0.8203",
"0.8164",
"0.8203",
"0.83",
"0.1748",
"0.83"
],
[
"0.1765",
"0.83",
"0.824",
"0.83",
"0.824",
"0.8433",
"0.1765",
"0.8433"
],
[
"0.1765",
"0.8433",
"0.824",
"0.8433",
"0.824",
"0.858",
"0.1765",
"0.858"
],
[
"0.1953",
"0.872",
"0.5063",
"0.872",
"0.5063",
"0.886",
"0.1953",
"0.886"
],
[
"0.1969",
"0.8853",
"0.824",
"0.886",
"0.824",
"0.901",
"0.1969",
"0.8994"
],
[
"0.1765",
"0.8994",
"0.8228",
"0.8994",
"0.8228",
"0.9136",
"0.1765",
"0.9136"
],
[
"0.1748",
"0.912",
"0.822",
"0.912",
"0.822",
"0.9253",
"0.1748",
"0.9253"
],
[
"0.5",
"0.9463",
"0.5107",
"0.955",
"0.5",
"0.964",
"0.4885",
"0.9556"
]
] | [
"Publishedasaconference paper atICLR 2024",
"LLM-CXR：用于CXR映像理解和生成的指令微调",
"LLM",
"Suhyeon Lee*, Won Jun Kim*, Jinho Chang & Jong Chul Ye",
"Korea AdvancedInstituteof Science&Technology,",
"[ suhyeon.lee, wonjun, jinhojsk515, jong.ye } @kaist.ac.kr",
"Abstract",
"Following the impressive development of LLMs, vision-language alignment in",
"LLMs is actively being researched to enable multimodal reasoning and visual",
"input/output.Thisdirectionofresearchisparticularlyrelevanttomedicalimag-",
"ingbecauseaccuratemedicalimageanalysisandgenerationconsistofreasoning",
"based on a combination of visual features and prior knowledge.Many recent",
"works have focused on training adapter networks that serve as an information",
"bridgebetween imageprocessing (encoding or generating) networks andLLMs;",
"but presumably, in order to achievemaximum reasoningpotential of LLMs on",
"visualinformationaswellaslanguage,imageandtextfeaturesshouldbeallowed",
"to interact more freely.This is especially important in the medical domain be-",
"causeunderstandingandgeneratingmedicalimagessuchaschest X-rays(CXR)",
"require not only accurate visual and language-based reasoning but also a more",
"intimatemapping between thetwo modalities.Thus, taking inspiration from",
"previous work on thetransformer and VQ-GAN combination for bidirectional",
"image and text generation, we build upon this approach and develop a method",
"forinstruction-tuninganLLMpre-trainedonlyontexttogainvision-language",
"capabilities for medicalimages. Specifically, we leverage a pretrained LLM’s",
"existing question-answering and instruction-following abilities to teach it to",
"understand visual inputs by instructing it to answer questions about image",
"inputsand,symmetrically,outputbothtextandimageresponsesappropriateto",
"agiven query bytuningtheLLMwithdiversetasksthatencompassimage-based",
"text-generation and text-based image-generation. We show that our model",
"LLM-CXR,trainedinthis approachshows betterimage-text alignmentinboth",
"CXRunderstanding andgenerationtasks whilebeing smallerin sizecompared",
"topreviouslydevelopedmodelsthatperformanarrowerrangeoftasks.",
"1Introduction",
"最近几年，大型语言模型（LLM）领域取得了令人瞩目的发展。LLM被认为是另一类AI",
"模型，因为它们能够灵活地理解/生成自然语言并执行基于语言的推理，从而使它们能够",
"概括为各种给定任务，而无需对其进行明确的培训。作为下一步，正在积极开发在LLME",
"(OpenAI,2023;Liuetal.,2023;Alayracetal.,2022;Lietal.,2023）中与语言一起输人视觉信息",
"的方法以及从LLME(Kohetal.,2023a;b)输出图像的方法。这些模型在医学领域具有很大的",
"潜力，特别有用，因为处理诸如胸部X射线（CXR）之类的医学图像需要能够理解上下文",
"执行推理并以图像和文本形式传达结论。第一代医学多式联运LLM最近开始出现（Moor",
"etal.,2023;Thawkaretal.,2023;Xuetal.,2023a)",
"开发这些模型的主要挑战是实现LLM的预训练语言功能与新引人的图像功能之间的对齐",
"而不会灾难性地忘记其语言能力。与自然图像相比，这在医学图像领域是一个更困难的挑",
"战，因为模型需要区分图像甚至图像部分的细微差异(e。g。、肺炎vs。CXR上的肺水肿)，",
"*These authors contributed equally to this work.",
"*Inordertocomplywiththe MIMIC-CXRdata usagelicense (Johnson et al.,2019a),allCXRimages",
"presented in the Figures 2, 3, 4, 8 are replaced with similar CXR's from the Indiana University chest",
"X-ray dataset (Demner-Fushman et al., 2016); and the presented MIMIC text reports are paraphrased",
"?"
] | {
"arxiv_id": "2305.11490",
"language": "zh",
"page_idx": 0
} |
||
2305.11490-zh-1 | 2305.11490-zh-1.jpg | $\underline{\text{Published as a conference paper at ICLR 2024}}$
<FigureHere>
Figure 1: (a) Example of previous work that indirectly implements multimodal bidirectional LLM by connecting a pretrained image encoder or image generation model to a pretrained LLM with a mapping layer. (b) Example of previous work that implements multimodal bidirectional non-LLM transformer with VQ-GAN trained from scratch ( i.e., without learned language features).(c) Toenabledirect multimodal feature interaction in LLMs pre-trained with text, our methodimplements( b) throughLLM- specificinstruction fine- tuningscheme.
然后提供准确的文字描述或图像生成。自然图像往往比医学图像更多样化，每一个都可以用各种各样的陈述来描述。另一方面，医学图像需要高度具体但全面的描述-使正确答案的空间更小，更复杂。这意味着医学多模式LLM需要文本和视觉特征之间更密切的映射
当前，将视觉特征映射到 LLM 和从 LLM 映射视觉特征的最流行方法是训练“适配器网络”以充当映射层，该映射层将预训练的图像编码器网络的输出转换为可以理解的形式由LLM$\boxed{\mathrm{F}}($Alayracetal.,2022;Liet al.,2023;Zhuet al.,2023)或将LLM的输出连接到图像生成网络以输出图像$\operatorname{E}($Koh et al.,2023a;b)。在这些方法中，LLM 被冻结以防止忘记其语言和推理能力。这些多模态LLM在视觉语言任务中表现出令人印象深刻的功能，例如图像字幕， 零镜头分类，视觉问答(VQA),图像生成和图像检索。但是，使用这种使用适配器网络的常见方法，视觉语言对齐可能会受到限制，因为适配器网络会成为信息瓶颈，可能会阻碍视觉和语言功能之间的相互作用
为了更好地弥合图像和文本之间的差距，我们从具有非LLM转换器的视觉语言相关(VLP) 领域中汲取灵感，在该领域中，已经有很多关于在同一令牌嵌人空间中处理图像和文本的工作。最突出的方法是使用 VQ-GANE(Esser et al.,2021) (VQ-VAEE(VanDenOordet al., 2017) )对图像进行标记，并使用自回归变换器解码器$\overline{\mathbb{E}}($Zhang et al.,2021;Lu et al.,2022; Wangetal.,2022a;Leeetal.,2023)生成文本标记和图像标记的序列。这些以前的工作与我们的观点非常吻合，即为了更好地进行图像-文本对齐，模型应该能够平等地处理图像和文本， 而没有单独的适配器瓶颈。此外，CXR 图像的精细细节 (例如纹理)在医学诊断中很重要， 这使得来自VQ 编码的局部图像特征的标记比全局描述更适合对齐目标。因此，在这项工作中，我们利用广泛使用的架构组件 VQ-GAN 无缝集成图像-文本令牌空间，而不需要对底层基础LLM进行任何结构修改
在此基础上，我们提出了一种方法，通过利用LLM的内置指令跟随能力，在LLM中实现更好的图像-文本对齐，以实现 CXR 图像理解和生成。具体来说，我们试图通过向模型提供有关 CXR 图像分析和生成的各种说明来教授模型视觉信息，然后使用输出来微调LLM。因此，我们的主要贡献之一是开发这种指令微调方法，该方法已被定制为适合已经训练的LLM,以扩展其输人和输出图像(由VQ-GAN标记)的功能，而无需修改模型结构或目标。一个重要区别是，虽然非LLMVLP转换器是从头开始训练的，而没有以前对语言的理解-这意味着没有遗忘的问题，也没有机会利用其语言理解-我们的贡献是一种方法，该方法采用
2 | [
[
"0.1748",
"0.0341",
"0.4836",
"0.0341",
"0.4836",
"0.04736",
"0.1748",
"0.04736"
],
[
"0.1805",
"0.1067",
"0.1978",
"0.1067",
"0.1978",
"0.11615",
"0.1805",
"0.11615"
],
[
"0.384",
"0.1061",
"0.4043",
"0.1061",
"0.4043",
"0.1187",
"0.384",
"0.1187"
],
[
"0.601",
"0.113",
"0.634",
"0.113",
"0.634",
"0.11615",
"0.601",
"0.11615"
],
[
"0.523",
"0.1218",
"0.6953",
"0.1218",
"0.6953",
"0.154",
"0.523",
"0.154"
],
[
"0.246",
"0.1427",
"0.3447",
"0.1401",
"0.3447",
"0.1622",
"0.246",
"0.1642"
],
[
"0.2303",
"0.1635",
"0.3228",
"0.1635",
"0.3228",
"0.1736",
"0.2303",
"0.1736"
],
[
"0.2009",
"0.1787",
"0.2091",
"0.1818",
"0.2051",
"0.1881",
"0.1969",
"0.1849"
],
[
"0.2467",
"0.1843",
"0.2737",
"0.1843",
"0.2737",
"0.1957",
"0.2467",
"0.1957"
],
[
"0.451",
"0.1901",
"0.468",
"0.1901",
"0.468",
"0.1989",
"0.451",
"0.1989"
],
[
"0.607",
"0.1932",
"0.642",
"0.1932",
"0.642",
"0.2008",
"0.607",
"0.2008"
],
[
"0.4355",
"0.202",
"0.524",
"0.202",
"0.524",
"0.209",
"0.4355",
"0.209"
],
[
"0.5933",
"0.1995",
"0.6733",
"0.1995",
"0.6733",
"0.2102",
"0.5933",
"0.2102"
],
[
"0.2255",
"0.2039",
"0.2573",
"0.2039",
"0.2573",
"0.2166",
"0.2255",
"0.2166"
],
[
"0.2035",
"0.2039",
"0.2197",
"0.207",
"0.2157",
"0.2172",
"0.1993",
"0.214"
],
[
"0.2239",
"0.2235",
"0.3015",
"0.2235",
"0.3015",
"0.2349",
"0.2239",
"0.2349"
],
[
"0.614",
"0.2267",
"0.722",
"0.2267",
"0.722",
"0.2405",
"0.614",
"0.2405"
],
[
"0.1805",
"0.2474",
"0.1978",
"0.2474",
"0.1978",
"0.2568",
"0.1805",
"0.2568"
],
[
"0.4272",
"0.2568",
"0.4453",
"0.2568",
"0.4453",
"0.2615",
"0.4272",
"0.2615"
],
[
"0.6846",
"0.2683",
"0.74",
"0.2683",
"0.74",
"0.279",
"0.6846",
"0.279"
],
[
"0.6855",
"0.2917",
"0.8057",
"0.2917",
"0.8057",
"0.2986",
"0.6855",
"0.2986"
],
[
"0.281",
"0.3037",
"0.29",
"0.3037",
"0.29",
"0.308",
"0.281",
"0.308"
],
[
"0.58",
"0.303",
"0.634",
"0.303",
"0.634",
"0.3145",
"0.58",
"0.3145"
],
[
"0.2484",
"0.3125",
"0.2998",
"0.3125",
"0.2998",
"0.3193",
"0.2484",
"0.3193"
],
[
"0.6855",
"0.3093",
"0.786",
"0.3113",
"0.786",
"0.3225",
"0.6855",
"0.3208"
],
[
"0.6846",
"0.318",
"0.7524",
"0.318",
"0.7524",
"0.3289",
"0.6846",
"0.3289"
],
[
"0.4004",
"0.3257",
"0.5",
"0.3257",
"0.5",
"0.33",
"0.4004",
"0.33"
],
[
"0.2443",
"0.334",
"0.3057",
"0.334",
"0.3057",
"0.3447",
"0.2443",
"0.3447"
],
[
"0.4534",
"0.334",
"0.5073",
"0.334",
"0.5073",
"0.3447",
"0.4534",
"0.3447"
],
[
"0.255",
"0.3452",
"0.2974",
"0.3452",
"0.2974",
"0.3528",
"0.255",
"0.3528"
],
[
"0.739",
"0.3435",
"0.7646",
"0.3435",
"0.7646",
"0.3655",
"0.739",
"0.3655"
],
[
"0.398",
"0.3499",
"0.464",
"0.3499",
"0.464",
"0.3567",
"0.398",
"0.3567"
],
[
"0.2361",
"0.3599",
"0.2932",
"0.3599",
"0.2932",
"0.3718",
"0.2361",
"0.3718"
],
[
"0.3987",
"0.3591",
"0.5522",
"0.3591",
"0.5522",
"0.3694",
"0.3987",
"0.3694"
],
[
"0.4028",
"0.3694",
"0.5317",
"0.3694",
"0.5317",
"0.377",
"0.4028",
"0.377"
],
[
"0.1765",
"0.3958",
"0.824",
"0.3958",
"0.824",
"0.4092",
"0.1765",
"0.4092"
],
[
"0.1765",
"0.4092",
"0.8228",
"0.4092",
"0.8228",
"0.4236",
"0.1765",
"0.4236"
],
[
"0.1765",
"0.4236",
"0.824",
"0.4236",
"0.824",
"0.4375",
"0.1765",
"0.4375"
],
[
"0.1748",
"0.4363",
"0.824",
"0.4375",
"0.824",
"0.452",
"0.1748",
"0.4507"
],
[
"0.1732",
"0.4495",
"0.824",
"0.4507",
"0.824",
"0.4653",
"0.1732",
"0.464"
],
[
"0.1732",
"0.464",
"0.701",
"0.4653",
"0.701",
"0.4792",
"0.1732",
"0.4778"
],
[
"0.1765",
"0.5005",
"0.8228",
"0.5005",
"0.8228",
"0.5146",
"0.1765",
"0.5146"
],
[
"0.1772",
"0.513",
"0.8228",
"0.513",
"0.8228",
"0.528",
"0.1772",
"0.528"
],
[
"0.1765",
"0.529",
"0.785",
"0.529",
"0.785",
"0.5425",
"0.1765",
"0.5425"
],
[
"0.1772",
"0.55",
"0.8228",
"0.55",
"0.8228",
"0.563",
"0.1772",
"0.563"
],
[
"0.1765",
"0.563",
"0.822",
"0.563",
"0.822",
"0.5776",
"0.1765",
"0.5776"
],
[
"0.1748",
"0.5757",
"0.8228",
"0.577",
"0.8228",
"0.5903",
"0.1748",
"0.589"
],
[
"0.1765",
"0.5913",
"0.8228",
"0.5913",
"0.8228",
"0.605",
"0.1765",
"0.605"
],
[
"0.1765",
"0.605",
"0.822",
"0.605",
"0.822",
"0.6187",
"0.1765",
"0.6187"
],
[
"0.1765",
"0.6187",
"0.8228",
"0.6187",
"0.8228",
"0.6333",
"0.1765",
"0.6333"
],
[
"0.1765",
"0.632",
"0.8228",
"0.632",
"0.8228",
"0.6465",
"0.1765",
"0.6465"
],
[
"0.1748",
"0.6465",
"0.4224",
"0.6465",
"0.4224",
"0.66",
"0.1748",
"0.66"
],
[
"0.1772",
"0.6675",
"0.8228",
"0.6675",
"0.8228",
"0.681",
"0.1772",
"0.681"
],
[
"0.1772",
"0.6816",
"0.822",
"0.6816",
"0.822",
"0.696",
"0.1772",
"0.696"
],
[
"0.1772",
"0.696",
"0.8203",
"0.696",
"0.8203",
"0.709",
"0.1772",
"0.709"
],
[
"0.1765",
"0.709",
"0.8228",
"0.709",
"0.8228",
"0.7227",
"0.1765",
"0.7227"
],
[
"0.1765",
"0.7217",
"0.822",
"0.7217",
"0.822",
"0.7363",
"0.1765",
"0.7363"
],
[
"0.1765",
"0.7373",
"0.822",
"0.7373",
"0.822",
"0.7505",
"0.1765",
"0.7505"
],
[
"0.1765",
"0.7485",
"0.822",
"0.7505",
"0.822",
"0.765",
"0.1765",
"0.763"
],
[
"0.1765",
"0.7646",
"0.822",
"0.7646",
"0.822",
"0.778",
"0.1765",
"0.778"
],
[
"0.1765",
"0.779",
"0.824",
"0.779",
"0.824",
"0.7925",
"0.1765",
"0.7925"
],
[
"0.1765",
"0.7925",
"0.3962",
"0.7925",
"0.3962",
"0.806",
"0.1765",
"0.806"
],
[
"0.1765",
"0.813",
"0.8228",
"0.813",
"0.8228",
"0.827",
"0.1765",
"0.827"
],
[
"0.1772",
"0.8276",
"0.8228",
"0.8276",
"0.8228",
"0.8413",
"0.1772",
"0.8413"
],
[
"0.1772",
"0.8413",
"0.822",
"0.8413",
"0.822",
"0.855",
"0.1772",
"0.855"
],
[
"0.1765",
"0.855",
"0.8228",
"0.855",
"0.8228",
"0.8687",
"0.1765",
"0.8687"
],
[
"0.1748",
"0.867",
"0.822",
"0.8687",
"0.822",
"0.8833",
"0.1748",
"0.882"
],
[
"0.1772",
"0.8833",
"0.8228",
"0.8833",
"0.8228",
"0.8965",
"0.1772",
"0.8965"
],
[
"0.1896",
"0.8965",
"0.8228",
"0.8965",
"0.8228",
"0.91",
"0.1896",
"0.91"
],
[
"0.1765",
"0.911",
"0.8228",
"0.911",
"0.8228",
"0.925",
"0.1765",
"0.925"
],
[
"0.492",
"0.9478",
"0.5073",
"0.9478",
"0.5073",
"0.962",
"0.492",
"0.962"
]
] | [
"PublishedasaconferencepaperatICLR2024",
"(a)",
"(c)",
"oOG",
"T",
"naa",
"s features",
"保",
"LLM",
"",
"ut:",
"Response:",
"Response:",
"",
"口",
"adaptormodule",
"Instruction Fine-tuning",
"(b)",
"Tiver",
"Left lungpnei",
"Nothereisngpleurateffusion",
"",
"LLM-CXR",
"VQ imagetoken",
"MRIisamedical imaging",
"technique that",
"Saturallanoaoecnmnre",
"Transformer",
"Whatis MRI?",
"decoder",
"1",
"masegeneratiot",
"OOO",
"Generatea CXR imageforthefollowing",
"diagnosis;leftlung.pneumonia,",
"Figurel:(a) Example of previous work that indirectly implements multimodal bidirectional",
"LLM by connecting a pretrained image encoder or image generation model to a pretrained LLM",
"with a mapping layer. (b) Example of previous work that implements multimodal bidirectional",
"non-LLM transformer with VQ-GAN trained from scratch (i.e., without learned language",
"features).(c)ToenabledirectmultimodalfeatureinteractioninLLMspre-trainedwithtext,our",
"methodimplements(b)throughLLM-specificinstructionfine-tuningscheme.",
"然后提供准确的文字描述或图像生成。自然图像往往比医学图像更多样化，每一个都可以",
"用各种各样的陈述来描述。另一方面，医学图像需要高度具体但全面的描述-使正确答案的",
"空间更小，更复杂。这意味着医学多模式LLM需要文本和视觉特征之间更密切的映射",
"当前，将视觉特征映射到LLM和从LLM映射视觉特征的最流行方法是训练“适配器网",
"络”以充当映射层，该映射层将预训练的图像编码器网络的输出转换为可以理解的形式由",
"LLME（Alayracetal.,2022:Lietal.,2023:Zhuetal.2023）或将LLM的输出连接到图像生成网",
"络以输出图像E(Kohetal.,2023a;b）。在这些方法中，LLM被冻结以防止忘记其语言和推",
"理能力。这些多模态LLM在视觉语言任务中表现出令人印象深刻的功能，例如图像字幕",
"零镜头分类，视觉问答（VQA），图像生成和图像检索。但是，使用这种使用适配器网络",
"的常见方法，视觉语言对齐可能会受到限制，因为适配器网络会成为信息瓶颈，可能会阻",
"碍视觉和语言功能之间的相互作用",
"为了更好地弥合图像和文本之间的差距，我们从具有非LLM转换器的视觉语言相关（VLP）",
"领域中取灵感，在该领域中，已经有很多关于在同一令牌嵌空间中处理图像和文本的",
"工作。最突出的方法是使用VQ-GANE(Esseretal.,2021）（VQ-VAEE(VanDenOordetal.",
"2017)）对图像进行标记，并使用自回归变换器解码器E(Zhangetal,2021Luetal.,2022",
"Wangetal.,2022a;Leeetal,2023)生成文本标记和图像标记的序列。这些以前的工作与我们",
"的观点非常吻合，即为了更好地进行图像-文本对齐，模型应该能够平等地处理图像和文本",
"而没有单独的适配器瓶颈。此外，CXR图像的精细细节（例如纹理）在医学诊断中很重要",
"这使得来自VQ编码的局部图像特征的标记比全局描述更适合对齐目标。因此，在这项工",
"作中，我们利用广泛使用的架构组件VQ-GAN无缝集成图像-文本令牌空间，而不需要对底",
"层基础LLM进行任何结构修改",
"在此基础上，我们提出了一种方法，通过利用LLM的内置指令跟随能力，在LLM中实现",
"更好的图像-文本对齐，以实现CXR图像理解和生成。具体来说，我们试图通过向模型提",
"供有关CXR图像分析和生成的各种说明来教授模型视觉信息，然后使用输出来微调LLM。",
"因此，我们的主要贡献之一是开发这种指令微调方法，该方法已被定制为适合已经训练的",
"LLM，以扩展其输人和输出图像（由VQ-GAN标记）的功能，而无需修改模型结构或目标。",
"与以前的工作（例如E(Zhangetal.,2021;Luetal.,2022;Wangetal.,2022a;Leeetal.,2023)）的",
"个重要区别是，虽然非LLMVLP转换器是从头开始训练的，而没有以前对语言的理解-这",
"意味着没有遗忘的问题，也没有机会利用其语言理解-我们的贡献是一种方法，该方法采用",
"2"
] | {
"arxiv_id": "2305.11490",
"language": "zh",
"page_idx": 1
} |
||
2305.11490-zh-3 | 2305.11490-zh-3.jpg | $\underline{\text{Published as aconferencepaperat ICLR 2024}}$
<FigureHere>
Figure2: ExamplesofVQAgeneratedfromaCXRtextreport.
textcorpus. 由于此过程只是LLM的微调过程，因此除了扩展令牌嵌人表之外，不对LLM 进行结构或客观的更改，并且不需要额外的网络。用于指令微调的模板使用羊驼家族(Taori etal.,2023;Databricks,2023)使用的模板。附录 D是羊驼家族指令微调提示的模板，由指令、 输人、响应部分组成
在微调过程中，根据目标函数对LLM进行优化，该目标函数以自回归方式输出基于指令-输人对的响应。请注意，这是一个指令调整方案，继承但不同于非LLM变压器多模式生成方法的培训。该方案的优点将在后面的 2.4.1节中更详细地介绍
用于微调的任务分为四种主要类型：1) 自然语言指令跟随(NL-IF), 2)报告到CXR的生成，3) CXR到报告的生成，以及4)基于CXR的视觉问题回答 (CXR-VQA)。这些是根据输入和输出模式分类的四种主要任务类型；但是对于模型，它们形成了一个丰富的培训环境，其中包含由Instruction区分的广泛任务。NL-IF和CXR-VQA 训练示例提供多维任务， 以便模型可以学习复杂对齐的视觉和文本特征以及以自然语言指示的任务的泛化。报告到CXR和CXR到报告生成任务在培训期间大量使用，对于视觉语言对齐非常重要，但必须注意的是，它们只是许多任务中的两个任务。通过使用指令来指定任务，我们将几个看不见的多模式任务功能添加到基本LLM中，而不会覆盖现有的基于语言的交互功能。与上面讨论的现有非LLM多模式双向生成模型相比，它还可以实现更简单但更通用的用户交互，因为它们只能使用预定义的令牌来查询某些任务，而基于LLM的模型可以基于自然语言指令进行查询，因此可以推广到零拍任务(Weietal.,2021;Wangetal.,2022b)
NL-IF任务。我们用预训练的指令跟随LLM (Databricks,2023)的权重初始化基本LLM。为了最大限度地减少在微调过程中忘记语言熟练程度的风险，我们同时使用用于指令调整此基本LLM的相同NL-IF数据集进行指令跟随调整
报告到CXR的生成。此任务旨在生成与输入放射学报告匹配的CXR图像作为响应。此任务的说明从 10个版本中随机抽样，类似于下面示例中的说明。由于利用了包含文本和图像令牌的扩展令牌空间，LLM以与文本令牌相同的方式直接输出图像令牌。因此，CXR图像生成不需要额外的网络或文本到图像生成模型(e。g。稳定扩散),如 Wuetal.(2023b);Koh etal.(2023a)所示。以下是用于指示LLM生成CXR映像的示例指令/输人对。
$\#\#\#$ Instruction: Generate a chest X
-ray image that corresponds to the entered free-text radiology reports for the chest X-ray image
Input: Bilateral, diffuse, confluent pulmonary
$\begin{array}{c}\text{opacities. Differential diagnoses include severe pulmonary edema ARDS or hemorrhage.}\\\end{array}$
$\#\#\#$ Response:<VQ032 VQ015 VQ124...VQ054 VQ032>
CXR-to-report生成。在此任务中，标记化的CXR映像是输入。请注意，在我们的模型中， 图像不是通过像 Lietal.(2023); Liuet al.(2023) 中那样用成对的视觉语言数据训练的单独网络来处理的；标记化的图像直接进人LLM，LLM本身在其语言能力之上学习视觉信息。然后可以指示将相应的放射学报告输出到给定图像作为响应。指令也从类似于以下示例的十个版本中随机抽样。以下代码段是CXR-to-report任务的说明示例。
$\#\#\#$ Instruction: Generate radiology reports for the entered CXR image
Input:<VQ071~VQ057~VQ~402~...~VQ122~VQ002> $\#\#\#$ Response: No acute cardiopulmonary process.
4 | [
[
"0.1732",
"0.03284",
"0.4795",
"0.03284",
"0.4795",
"0.04672",
"0.1732",
"0.04672"
],
[
"0.5767",
"0.113",
"0.79",
"0.113",
"0.79",
"0.1206",
"0.5767",
"0.1206"
],
[
"0.5786",
"0.1237",
"0.6313",
"0.1237",
"0.6313",
"0.1276",
"0.5786",
"0.1276"
],
[
"0.585",
"0.1288",
"0.79",
"0.1288",
"0.79",
"0.1395",
"0.585",
"0.1395"
],
[
"0.3694",
"0.137",
"0.4028",
"0.137",
"0.4028",
"0.1483",
"0.3694",
"0.1483"
],
[
"0.5864",
"0.1383",
"0.7036",
"0.1383",
"0.7036",
"0.1483",
"0.5864",
"0.1483"
],
[
"0.3823",
"0.1578",
"0.5034",
"0.1578",
"0.5034",
"0.1654",
"0.3823",
"0.1654"
],
[
"0.5728",
"0.1547",
"0.6914",
"0.154",
"0.6914",
"0.1642",
"0.5728",
"0.1654"
],
[
"0.3708",
"0.1642",
"0.478",
"0.1642",
"0.478",
"0.1749",
"0.3708",
"0.1749"
],
[
"0.5195",
"0.1692",
"0.5537",
"0.1692",
"0.5537",
"0.1812",
"0.5195",
"0.1812"
],
[
"0.5825",
"0.1635",
"0.8037",
"0.1642",
"0.8037",
"0.1749",
"0.5825",
"0.1736"
],
[
"0.3708",
"0.1724",
"0.4878",
"0.1724",
"0.4878",
"0.1831",
"0.3708",
"0.1831"
],
[
"0.3708",
"0.1818",
"0.4143",
"0.1831",
"0.4143",
"0.1913",
"0.3708",
"0.1901"
],
[
"0.4224",
"0.1849",
"0.4805",
"0.1843",
"0.4805",
"0.1881",
"0.4224",
"0.1895"
],
[
"0.5728",
"0.1818",
"0.6895",
"0.1818",
"0.6895",
"0.1925",
"0.5728",
"0.1925"
],
[
"0.3733",
"0.1925",
"0.5024",
"0.1925",
"0.5024",
"0.1995",
"0.3733",
"0.1995"
],
[
"0.585",
"0.1913",
"0.8013",
"0.1913",
"0.8013",
"0.202",
"0.585",
"0.202"
],
[
"0.3694",
"0.1989",
"0.4468",
"0.1995",
"0.4468",
"0.2102",
"0.3694",
"0.209"
],
[
"0.5728",
"0.2058",
"0.7974",
"0.207",
"0.7974",
"0.2203",
"0.5728",
"0.2197"
],
[
"0.576",
"0.2184",
"0.5923",
"0.2184",
"0.5923",
"0.226",
"0.576",
"0.226"
],
[
"0.576",
"0.226",
"0.7715",
"0.226",
"0.7715",
"0.2361",
"0.576",
"0.2361"
],
[
"0.2834",
"0.2556",
"0.7114",
"0.2556",
"0.7114",
"0.2695",
"0.2834",
"0.2695"
],
[
"0.1748",
"0.286",
"0.824",
"0.286",
"0.824",
"0.2998",
"0.1748",
"0.2998"
],
[
"0.1765",
"0.2998",
"0.824",
"0.2998",
"0.824",
"0.3132",
"0.1765",
"0.3132"
],
[
"0.1748",
"0.3145",
"0.824",
"0.3145",
"0.824",
"0.3276",
"0.1748",
"0.3276"
],
[
"0.1765",
"0.3276",
"0.3245",
"0.3276",
"0.3245",
"0.3416",
"0.1765",
"0.3416"
],
[
"0.1772",
"0.351",
"0.8203",
"0.351",
"0.8203",
"0.361",
"0.1772",
"0.361"
],
[
"0.1772",
"0.363",
"0.8228",
"0.363",
"0.8228",
"0.377",
"0.1772",
"0.377"
],
[
"0.1748",
"0.377",
"0.6055",
"0.377",
"0.6055",
"0.3901",
"0.1748",
"0.3901"
],
[
"0.1772",
"0.3984",
"0.8228",
"0.3984",
"0.8228",
"0.4124",
"0.1772",
"0.4124"
],
[
"0.1765",
"0.4124",
"0.8228",
"0.4124",
"0.8228",
"0.4255",
"0.1765",
"0.4255"
],
[
"0.1748",
"0.4255",
"0.824",
"0.4255",
"0.824",
"0.4395",
"0.1748",
"0.4395"
],
[
"0.1732",
"0.4382",
"0.822",
"0.4395",
"0.822",
"0.4539",
"0.1732",
"0.4526"
],
[
"0.1772",
"0.455",
"0.822",
"0.455",
"0.822",
"0.4653",
"0.1772",
"0.4653"
],
[
"0.1732",
"0.4673",
"0.822",
"0.4673",
"0.822",
"0.481",
"0.1732",
"0.481"
],
[
"0.1765",
"0.481",
"0.822",
"0.481",
"0.822",
"0.4944",
"0.1765",
"0.4944"
],
[
"0.1765",
"0.4956",
"0.822",
"0.4956",
"0.822",
"0.509",
"0.1765",
"0.509"
],
[
"0.1765",
"0.509",
"0.8228",
"0.509",
"0.8228",
"0.523",
"0.1765",
"0.523"
],
[
"0.1772",
"0.523",
"0.822",
"0.523",
"0.822",
"0.536",
"0.1772",
"0.536"
],
[
"0.1765",
"0.537",
"0.6787",
"0.537",
"0.6787",
"0.5503",
"0.1765",
"0.5503"
],
[
"0.1765",
"0.558",
"0.8228",
"0.558",
"0.8228",
"0.5713",
"0.1765",
"0.5713"
],
[
"0.1772",
"0.5728",
"0.824",
"0.5728",
"0.824",
"0.586",
"0.1772",
"0.586"
],
[
"0.1748",
"0.584",
"0.5215",
"0.5854",
"0.5215",
"0.5986",
"0.1748",
"0.5977"
],
[
"0.1765",
"0.606",
"0.8228",
"0.606",
"0.8228",
"0.619",
"0.1765",
"0.619"
],
[
"0.1772",
"0.617",
"0.822",
"0.617",
"0.822",
"0.632",
"0.1772",
"0.632"
],
[
"0.1772",
"0.63",
"0.8228",
"0.63",
"0.8228",
"0.6445",
"0.1772",
"0.6445"
],
[
"0.1772",
"0.643",
"0.824",
"0.643",
"0.824",
"0.6577",
"0.1772",
"0.6577"
],
[
"0.1748",
"0.6577",
"0.714",
"0.6577",
"0.714",
"0.672",
"0.1748",
"0.672"
],
[
"0.1765",
"0.6963",
"0.4192",
"0.6963",
"0.4192",
"0.7104",
"0.1765",
"0.7104"
],
[
"0.2075",
"0.7085",
"0.819",
"0.709",
"0.819",
"0.7236",
"0.2075",
"0.7227"
],
[
"0.1732",
"0.72",
"0.468",
"0.7217",
"0.468",
"0.7363",
"0.1732",
"0.735"
],
[
"0.21",
"0.7344",
"0.7715",
"0.7344",
"0.7715",
"0.7485",
"0.21",
"0.7485"
],
[
"0.1765",
"0.7476",
"0.5645",
"0.7476",
"0.5645",
"0.761",
"0.1765",
"0.761"
],
[
"0.1748",
"0.7827",
"0.822",
"0.7827",
"0.822",
"0.797",
"0.1748",
"0.797"
],
[
"0.1765",
"0.7954",
"0.822",
"0.7954",
"0.822",
"0.8096",
"0.1765",
"0.8096"
],
[
"0.1765",
"0.8096",
"0.8228",
"0.8096",
"0.8228",
"0.8228",
"0.1765",
"0.8228"
],
[
"0.1748",
"0.8228",
"0.8203",
"0.8228",
"0.8203",
"0.836",
"0.1748",
"0.836"
],
[
"0.1765",
"0.8354",
"0.6504",
"0.8354",
"0.6504",
"0.8486",
"0.1765",
"0.8486"
],
[
"0.1765",
"0.875",
"0.648",
"0.875",
"0.648",
"0.888",
"0.1765",
"0.888"
],
[
"0.1765",
"0.887",
"0.5024",
"0.887",
"0.5024",
"0.901",
"0.1765",
"0.901"
],
[
"0.1748",
"0.9014",
"0.509",
"0.9014",
"0.509",
"0.912",
"0.1748",
"0.912"
],
[
"0.4927",
"0.9497",
"0.5063",
"0.9497",
"0.5063",
"0.9604",
"0.4927",
"0.9604"
]
] | [
"Published as aconference paperatICLR2024",
"Whatisthemost likelycauseofthesubtleopacityinthe",
"htlunebase",
"Early pneumonia isthemostlikelycauseof the subtle",
"REPORT:",
"opacity at theright lung base..",
"subtleopacityintherightlung",
"\"is thereany pleural effusion?\"",
"basemostlikelyduetoea",
"LLaMa2",
"No, there is no pleural effusion present in thechest X-ray",
"pneumonia.Thereisno pleura",
"effusionor",
"eumothnrayT",
"\"is there any pneumothorax?\"",
"nediastinal silhouetteandbilatera",
"No, there is no pneumothorax present in the chest X-ray.",
"hila appear normal.",
"Whatistheappearanceofthemediastinalsilhouetteand",
"hila?",
"The mediastinal silhouette and hila appear normal",
"Figure2: ExamplesofVQAgeneratedfromaCXRtextreport.*",
"textcorpus.由于此过程只是LLM的微调过程，因此除了扩展令牌嵌入表之外，不对LLM",
"进行结构或客观的更改，并且不需要额外的网络。用于指令微调的模板使用羊驼家族（Taori",
"etal.,2023;Databricks,2023)使用的模板。附录D是羊驼家族指令微调提示的模板，由指令、",
"输人、响应部分组成",
"在微调过程中，根据自标函数对LLM进行优化，该自标函数以自回归方式输出基于指今-输",
"人对的响应。请注意，这是一个指令调整方案，继承但不同于非LLM变压器多模式生成方",
"法的培训。该方案的优点将在后面的2.4.1节中更详细地介绍",
"用于微调的任务分为四种主要类型：1）自然语言指令跟随（NL-IF），2）报告到CXR的生",
"成，3）CXR到报告的生成，以及4）基于CXR的视觉问题回答（CXR-VQA）。这些是根据",
"输人和输出模式分类的四种主要任务类型；但是对于模型，它们形成了一个丰富的培训环",
"境，其中包含由Instruction区分的广泛任务。NL-IF和CXR-VQA训练示例提供多维任务",
"以便模型可以学习复杂对齐的视觉和文本特征以及以自然语言指示的任务的泛化。报告到",
"CXR和CXR到报告生成任务在培训期间大量使用，对于视觉语言对齐非常重要，但必须注",
"意的是，它们只是许多任务中的两个任务。通过使用指令来指定任务，我们将几个看不见",
"的多模式任务功能添加到基本LLM中，而不会覆盖现有的基于语言的交互功能。与上面讨",
"论的现有非LLM多模式双向生成模型相比，它还可以实现更简单但更通用的用户交互，因",
"为它们只能使用预定义的令牌来查询某些任务，而基于LLM的模型可以基于自然语言指令",
"进行查询，因此可以推广到零拍任务（Weietal.,2021;Wangetal.,2022b）",
"NL-IF任务。我们用预训练的指令跟随LLM(Databricks,2023)的权重初始化基本LLM。为",
"了最大限度地减少在微调过程中忘记语言熟练程度的风险，我们同时使用用于指令调整此",
"基本LLM的相同NL-IF数据集进行指令跟随调整",
"报告到CXR的生成。此任务旨在生成与输人放射学报告匹配的CXR图像作为响应。此任",
"务的说明从10个版本中随机抽样，类似于下面示例中的说明。由于利用了包含文本和图像",
"令牌的扩展令牌空间，LLM以与文本令牌相同的方式直接输出图像令牌。因此，CXR图像",
"生成不需要额外的网络或文本到图像生成模型(e。g。稳定扩散)，如Wuetal.(2023b);Koh",
"etal.(2023a)所示。以下是用于指示LLM生成CXR映像的示例指令/输入对。",
"### Instruction: Generate a chest X",
"-ray image that corresponds to the entered free-text radiology reports for the chest X-ray image.",
"Input: Bilateral, diffuse, confluent pulmonary",
"opacities. Differential diagnoses include severe pulmonary edema ARDS or hemorrhage",
"###Response:<VQ032VQ015VQ124...VQ054VQ032>",
"CXR-to-report生成。在此任务中，标记化的CXR映像是输入。请注意，在我们的模型中",
"图像不是通过像Lietal.(2023);Liuetal.(2023)中那样用成对的视觉语言数据训练的单独网",
"络来处理的；标记化的图像直接进入LLM，LLM本身在其语言能力之上学习视觉信息。然",
"后可以指示将相应的放射学报告输出到给定图像作为响应。指令也从类似于以下示例的十",
"个版本中随机抽样。以下代码段是CXR-to-report任务的说明示例。",
"### Instruction: Generate radiologyreports for the entered CxRimage",
"Input:<VQ071VQ057VQ402...VQ122VQ002>",
"###Response:Noacutecardiopulmonaryprocess",
"4"
] | {
"arxiv_id": "2305.11490",
"language": "zh",
"page_idx": 3
} |
||
2305.11490-zh-4 | 2305.11490-zh-4.jpg | $\underline{\text{Published as aconferencepaperat ICLR 2024}}$
CXR-VQA任务。在此任务中，有关图像的问题作为说明给出，并且模型生成适当的响应有关作为输入给出的CXR映像的问题。这不仅可以训练模型获得VQA功能，还可以提高其他视觉语言任务的性能(i。e。,增强视觉-语言对齐)。下面是CXR-VQA任务的示例说明。
$\#\#\#$ Instruction: What is the size of the pleural effusions?
Input: < VQ121 VQ720 VQ002. . . VQ005 $\hat{V}$Q428>
$\#\#\#$ Response: The bilateral pleural effusions are moderate to large.
2.4.1 Trainingobjective
训练目标是以自回归方式生成由Instruction、Input 和 Response组成的整个目标段落。但是$类似于一般的GPT预训练$\underline{\mathrm{E}}($Radfordetal.,2018;2019;Brownetal.,2020)$,损失仅应用于响应$密钥后生成的令牌 (i。e。\#\# 响应：), 遵循指令调优方案(Taorietal.,2023;Databricks,2023$, 具体来说，对于标记化的训练段落$\left[x_1,x_2,...,x_{n_x},y_1,y_2,...,y_{n_y}\right]$, 其中$x$表示指令和输入部分， $y$表示响应部分，训练损失由下式给出：
$$L_{instruct}=-\mathrm{log}p(\boldsymbol{y}|\boldsymbol{x})=\sum_{i=1}^{n_y}-\mathrm{log}p(y_i|y_{i-1},y_{i-2},...,y_1,x_{n_x},x_{n_x-1},...,x_1).$$
请注意，此目标与用于训练非LLM变压器的目标不同，例如在使用$L_joint=-\log p(\boldsymbol{x},\boldsymbol{y})$的Leeetal.(2023)中。我们假设通过使用条件损失进行指令调优可以减轻对微调数据集的过度拟合，特别是在处理有限数据时，从而促进更好的学习环境，鼓励模型从其现有功能扩展其理解，而不是记忆新功能
2.4.2 Two-stagefine-tuning
与以前的作品类似E$Chambonetal. ( 2022) ;$Xuetal.(2023b),我们将重点放在正面视图上 (i。e。, AP和PA)图像和相应报告的印象部分，因为它们与诊断最相关，更适合直接比较结果。此外，我们通过修剪 MIMIC-CXR-JPG 数据集以仅包含每个主题的第一项研究，从而最大程度地减少了报告中对先前研究的引用 (因为我们在推断期间一次仅使用一张图像，并且没有参考先前研究的报告)
从最近的LLM微调方法$\overline{\mathrm{E}}($Zhuetal.,2023)中常用的多阶段训练技术中汲取灵感，我们遵循类似的两阶段训练方法。在第一阶段，未过滤的大量数据(i。e。,MMIC-CXR数据集中的所有 CXR图像-报告对和发现/印象部分)用于在新图像标记的整个分布以及图像和文本标记之间的一般关系上训练模型。在第二阶段，使用修剪后的数据集进一步增强模型性能(i。e。仅使用前视图图像，每个患者的第一次研究，以及文本报告的印象部分),为视觉-语言对齐提供了更强的信号
附录 B 中提供了实现细节
# 3 Experiments
我们在LLM-CXR 执行的所有任务中将 LLM-CXR 的性能与类似的当代模型进行了比较对于CXR 到报告的生成，我们将我们的结果与 UniXGen$\underline{\mathrm{E}}($Lee et al.,2023) , XrayGPT$\underline{\mathrm{F}}$ (Thawkaretal.,2023) , RadFM$\boxed{\mathrm{E}}$Wuetal.(2023a) , IFCC$\boxed{\mathrm{E}}$Delbroucketal.(2022)和R2Gen$\boxed{\mathrm{E}}$ Chen et al. ( 2020) 进 行 比 较 ; 对 于 CXR- VQA, 使 用 XrayGPT$\underline {E}($Thawkar et al$.,2023)~.$ RadXXFM$\overline{\mathrm{E}}$。请注意，只有LLM-CXR能够执行所有三个任务。有关我们方法设计的消融实验的更多详细信息，请参阅附录F
3.1 CXR-to-reportgenerationtask
我们使用每个模型为MIMIC-CXR-JPG 数据集中的CXR 图像生成文本放射学报告，使用LLM-CXRE (图E3) , UniXGen$\underline{\mathrm{E}}($Lee et al.,2023) , XrayGPT$\boxed{\mathrm{F}}($Thawkar et al.,2023) , RadFM$\boxed{\mathrm{E}}$Wuetal.(2023a) ,IFCC$\boxed{\mathrm{E}}$Delbroucketal.(2022)和RG2。为了量化生成的报告在CXR图像中反映临床上重要的放射学信息的程度，我们使用CheXpert-labeler$\underline{\mathrm{E}}($Irvinetal. 2019),它是一种基于规则的自然语言处理工具，可读取CXR的文本报告并提取报告中是
5 | [
[
"0.1765",
"0.0341",
"0.4805",
"0.0341",
"0.4805",
"0.04736",
"0.1765",
"0.04736"
],
[
"0.1765",
"0.1061",
"0.819",
"0.1061",
"0.819",
"0.1193",
"0.1765",
"0.1193"
],
[
"0.1748",
"0.11615",
"0.822",
"0.11743",
"0.822",
"0.132",
"0.1748",
"0.1307"
],
[
"0.1765",
"0.1307",
"0.822",
"0.1307",
"0.822",
"0.1445",
"0.1765",
"0.1445"
],
[
"0.1765",
"0.1704",
"0.5513",
"0.1692",
"0.5513",
"0.1831",
"0.1765",
"0.1843"
],
[
"0.1772",
"0.1849",
"0.4958",
"0.1849",
"0.4958",
"0.1957",
"0.1772",
"0.1957"
],
[
"0.1748",
"0.1957",
"0.621",
"0.1963",
"0.621",
"0.2102",
"0.1748",
"0.209"
],
[
"0.1748",
"0.2343",
"0.3513",
"0.2343",
"0.3513",
"0.2474",
"0.1748",
"0.2474"
],
[
"0.1772",
"0.2588",
"0.822",
"0.2588",
"0.822",
"0.2727",
"0.1772",
"0.2727"
],
[
"0.1765",
"0.2727",
"0.822",
"0.2727",
"0.822",
"0.286",
"0.1765",
"0.286"
],
[
"0.1748",
"0.286",
"0.822",
"0.286",
"0.822",
"0.2998",
"0.1748",
"0.2998"
],
[
"0.1765",
"0.3093",
"0.8267",
"0.3093",
"0.8267",
"0.324",
"0.1765",
"0.324"
],
[
"0.1748",
"0.3225",
"0.455",
"0.3225",
"0.455",
"0.336",
"0.1748",
"0.336"
],
[
"0.4346",
"0.3567",
"0.4412",
"0.3623",
"0.4346",
"0.3687",
"0.4272",
"0.363"
],
[
"0.4429",
"0.3567",
"0.7466",
"0.358",
"0.7466",
"0.3757",
"0.4429",
"0.375"
],
[
"0.8013",
"0.3599",
"0.824",
"0.3599",
"0.824",
"0.3726",
"0.8013",
"0.3726"
],
[
"0.429",
"0.3726",
"0.4495",
"0.3726",
"0.4495",
"0.385",
"0.429",
"0.385"
],
[
"0.255",
"0.3599",
"0.4265",
"0.3599",
"0.4265",
"0.3738",
"0.255",
"0.3738"
],
[
"0.1772",
"0.3965",
"0.6978",
"0.3965",
"0.6978",
"0.4104",
"0.1772",
"0.4104"
],
[
"0.701",
"0.3965",
"0.8228",
"0.3965",
"0.8228",
"0.4104",
"0.701",
"0.4104"
],
[
"0.1765",
"0.4104",
"0.822",
"0.4104",
"0.822",
"0.4236",
"0.1765",
"0.4236"
],
[
"0.1765",
"0.4236",
"0.8228",
"0.4236",
"0.8228",
"0.4375",
"0.1765",
"0.4375"
],
[
"0.1748",
"0.4382",
"0.3733",
"0.4382",
"0.3733",
"0.452",
"0.1748",
"0.452"
],
[
"0.1748",
"0.4685",
"0.3782",
"0.4685",
"0.3782",
"0.4817",
"0.1748",
"0.4817"
],
[
"0.1748",
"0.4912",
"0.8228",
"0.4937",
"0.8228",
"0.507",
"0.1748",
"0.505"
],
[
"0.1765",
"0.506",
"0.8228",
"0.507",
"0.8228",
"0.521",
"0.1765",
"0.5195"
],
[
"0.1772",
"0.5215",
"0.822",
"0.5215",
"0.822",
"0.535",
"0.1772",
"0.535"
],
[
"0.1765",
"0.535",
"0.8228",
"0.535",
"0.8228",
"0.549",
"0.1765",
"0.549"
],
[
"0.1765",
"0.5503",
"0.3628",
"0.5503",
"0.3628",
"0.561",
"0.1765",
"0.561"
],
[
"0.1772",
"0.571",
"0.8228",
"0.571",
"0.8228",
"0.584",
"0.1772",
"0.584"
],
[
"0.1748",
"0.582",
"0.822",
"0.5835",
"0.822",
"0.5977",
"0.1748",
"0.5967"
],
[
"0.1748",
"0.5967",
"0.822",
"0.5967",
"0.822",
"0.61",
"0.1748",
"0.61"
],
[
"0.1748",
"0.61",
"0.8203",
"0.6123",
"0.8203",
"0.6255",
"0.1748",
"0.6235"
],
[
"0.1765",
"0.6235",
"0.822",
"0.625",
"0.822",
"0.6396",
"0.1765",
"0.638"
],
[
"0.1748",
"0.6396",
"0.308",
"0.6396",
"0.308",
"0.653",
"0.1748",
"0.653"
],
[
"0.1732",
"0.659",
"0.353",
"0.659",
"0.353",
"0.6724",
"0.1732",
"0.6724"
],
[
"0.1724",
"0.696",
"0.3145",
"0.6963",
"0.3145",
"0.711",
"0.1724",
"0.7104"
],
[
"0.1765",
"0.728",
"0.822",
"0.728",
"0.822",
"0.741",
"0.1765",
"0.741"
],
[
"0.1765",
"0.742",
"0.8228",
"0.742",
"0.8228",
"0.756",
"0.1765",
"0.756"
],
[
"0.1765",
"0.754",
"0.824",
"0.755",
"0.824",
"0.7695",
"0.1765",
"0.768"
],
[
"0.1765",
"0.7695",
"0.818",
"0.7695",
"0.818",
"0.7827",
"0.1765",
"0.7827"
],
[
"0.1748",
"0.7827",
"0.8228",
"0.7827",
"0.8228",
"0.797",
"0.1748",
"0.797"
],
[
"0.1748",
"0.7974",
"0.433",
"0.7974",
"0.433",
"0.811",
"0.1748",
"0.811"
],
[
"0.1748",
"0.83",
"0.4272",
"0.83",
"0.4272",
"0.8433",
"0.1748",
"0.8433"
],
[
"0.1765",
"0.856",
"0.822",
"0.856",
"0.822",
"0.869",
"0.1765",
"0.869"
],
[
"0.1748",
"0.867",
"0.818",
"0.8687",
"0.818",
"0.8833",
"0.1748",
"0.882"
],
[
"0.814",
"0.875",
"0.819",
"0.875",
"0.819",
"0.88",
"0.814",
"0.88"
],
[
"0.1772",
"0.8853",
"0.822",
"0.8853",
"0.822",
"0.896",
"0.1772",
"0.896"
],
[
"0.1765",
"0.8965",
"0.822",
"0.8965",
"0.822",
"0.911",
"0.1765",
"0.911"
],
[
"0.1748",
"0.911",
"0.8228",
"0.911",
"0.8228",
"0.925",
"0.1748",
"0.925"
],
[
"0.4927",
"0.9497",
"0.5063",
"0.9497",
"0.5063",
"0.961",
"0.4927",
"0.961"
]
] | [
"PublishedasaconferencepaperatICLR2024",
"CXR-VQA任务。在此任务中，有关图像的问题作为说明给出，并且模型生成适当的响应",
"有关作为输入给出的CXR映像的问题。这不仅可以训练模型获得VQA功能，还可以提高",
"其他视觉语言任务的性能(i。e。，增强视觉-语言对齐)。下面是CXR-VQA任务的示例说明。",
"### Instruction: What is the size of the pleural effusions?",
"Input:<VQ121VQ720VQ002...VQ005VQ428>",
"###Response:Thebilateralpleuraleffusionsaremoderatetolarge.",
"2.4.1 Trainingobjective",
"训练目标是以自回归方式生成由Instruction、Input和Response组成的整个目标段落。但是",
"类似于一般的GPT预训练E(Radfordetal.2018;2019;Brownetal.,2020)，损失仅应用于响应",
"密钥后生成的令牌（i。e。###响应：)，遵循指令调优方案(Taorietal.,2023;Databricks,2023)",
"具体来说，对于标记化的训练段落[1,2.,an91,92,,9n，其中a表示指令和输人部分，",
"y表示响应部分，训练损失由下式给出：",
"1",
"-logp(yilyi-1,i-2...1,ana,n.-1.i).",
"(1)",
"=1",
"Linstruct=-logp(y|c)=",
"请注意，此目标与用于训练非LLM变压器的目标不同，例如在使用Lioin",
"t=-logp(a,y)的",
"Leeetal.(2023)中。我们假设通过使用条件损失进行指令调优可以减轻对微调数据集的过度",
"拟合，特别是在处理有限数据时，从而促进更好的学习环境，鼓励模型从其现有功能扩展",
"其理解，而不是记忆新功能",
"2.4.2Two-stagefine-tuning",
"与以前的作品类似EChambonetal.(2022);Xuetal.(2023b），我们将重点放在正面视图上（i。",
"e。，AP和PA)图像和相应报告的印象部分，因为它们与诊断最相关，更适合直接比较结果。",
"此外，我们通过修剪MIMIC-CXR-JPG数据集以仅包含每个主题的第一项研究，从而最大",
"程度地减少了报告中对先前研究的引用（因为我们在推断期间一次仅使用一张图像，并且",
"没有参考先前研究的报告）",
"从最近的LLM微调方法E(Zhuetal.,2023)中常用的多阶段训练技术中汲取灵感，我们遵循",
"类似的两阶段训练方法。在第一阶段，未过滤的大量数据(i。e。，MIMIC-CXR数据集中的",
"所有CXR图像-报告对和发现/印象部分）用于在新图像标记的整个分布以及图像和文本标记",
"之间的一般关系上训练模型。在第二阶段，使用修剪后的数据集进一步增强模型性能(i。e。",
"仅使用前视图图像，每个患者的第一次研究，以及文本报告的印象部分），为视觉-语言对齐",
"提供了更强的信号",
"附录B中提供了实现细节",
"3Experiments",
"我们在LLM-CXR执行的所有任务中将LLM-CXR的性能与类似的当代模型进行了比较。",
"对于CXR到报告的生成，我们将我们的结果与UniXGenE(Leeetal.,2023），XrayGPTE",
"(Thawkaretal.,2023)，RadFME/Wuetal.(2023a)，IFCCEDelbroucketal.(2022)和R2GenF",
"Chenetal.（2020）进行比较；对于CXR-VQA，使用XrayGPTE(Thawkaretal.,2023)",
"RadXXFME。请注意，只有LLM-CXR能够执行所有三个任务。有关我们方法设计的消融",
"实验的更多详细信息，请参阅附录F",
"3.1 CXR-to-reportgenerationtask",
"我们使用每个模型为MIMIC-CXR-JPG数据集中的CXR图像生成文本放射学报告，使用",
"LLM-CXRE（图E3），UniXGenE(Leeetal.,2023)，XrayGPTE(Thawkaretal.,2023)",
"中",
"RadFME/Wuetal.（2023a），IFCCFDelbroucketal.（2022）和RG2。为了量化生成的报告在",
"CXR图像中反映临床上重要的放射学信息的程度，我们使用CheXpert-labelerE(Irvinetal.",
"2019），它是一种基于规则的自然语言处理工具，可读取CXR的文本报告并提取报告中是",
"5"
] | {
"arxiv_id": "2305.11490",
"language": "zh",
"page_idx": 4
} |
||
2109.12781-zh-1 | 2109.12781-zh-1.jpg | 有效利用图卷积网络和上下文子树进行商品新闻事件提取
$\textbf{Meisin Lee, Lay- Ki Soon, Eu- Gene Siew}$
Monash University
{ mei.lee, soon.layki, siew.eu-gene}{@monash.edu
## $\mathbf{Abstract}$
Event extraction in commodity news is a less researched area as compared to generic event extraction. However, accurate event extraction from commodity news is useful in a broad range of applications such as understanding event chains and learning event-event relations, which can then be used for commodity price prediction. The events found in commodity news exhibit characteristics different from generic events, hence posing a unique challenge in event extraction using existing methods. This paper proposes an effective use of Graph Convolutional Networks (GCN) with a pruned dependency parse tree, termed contextual sub-tree, for better event extraction in commodity news. The event extraction model is trained using feature embeddings from ComBERT, a BERT-based masked language model that was produced through domain-adaptive pre-training on a commodity news corpus. Experimental results show the efficiency of the proposed solution, which outperforms existing methods with F1 scores as high as 0.90. Furthermore, our pre-trained language model outperforms GloVe by 23 % , and BERT and RoBERTa by 7 % in terms of argument roles classification. For the goal of reproducibility, the code and trained models are made publicly available $^{1}.$
### 1 $\textbf{Introduction}$
诸如地缘政治和宏观经济相关事件之类的世界事件已显示出对短期和长期 (Brandt and Gao, 2019) 商品价格的影响。通常，在商品新闻文章中发现的事件可以分为地缘政治，宏观经济， 供需相关和商品价格变动。商品新闻是提取和挖掘此类事件的宝贵信息来源。准确的事件提取对于许多重要的下游任务非常有用，例如了解事件链和学习事件序列，也称为(Schank and Abelson, 2013) 中的脚本，可用于最终准确的商品价格预测
$^{1}$https://github.com/meisin/
commodity-News-Event-Extraction
U.S. crude stockpiles soared by 1.350 million barrels in December from a mere 200 million barrels to 438.9 million barrels, due to this oversupply crude oil prices plunged more than 50% on Tuesday
Figure 1: An example of a sentence from a piece of commodity news $^3$, loaded with three events (trigger words are in bold): (1) Crude oil inventory increase, (2) Oversupply and (3) Crude Oil price decrease. Three event arguments that are of the same entity type are highlighted in color
根据 ACE ( 自动内容提取) 程序$^{2}$中的定义， 事件提取任务由两个子任务组成：(1)事件触发器提取(识别和分类事件触发器)和(2)事件参数提取(识别事件触发器的参数并标记其角色)。在这项工作中，我们对(Lee et al.,2021) 中引入的商品新闻数据集执行事件提取。图 1显示了此数据集中的一个示例句子
为了说明事件提取的任务，请考虑图 1 中的示例以及表 1 中列出的相应信息。在事件触发提取子任务中，模型被训练以识别触发词：飙升并分类正确的事件类型：movement\_up\_gain, 而在事件参数提取子任务中，模型被训练以从句子中的实体提及池中识别事件参数，然后标记每个实体与所识别事件相关的参数角色。
<table>
<tbody>
<tr>
<th>Event</th>
<th>Entity Mention</th>
<th>Argument Role</th>
</tr>
<tr>
<td>Trigger:</td>
<td>$U.S.$</td>
<td>Supplier</td>
</tr>
<tr>
<td rowspan="3">$^{11}86^{\mathrm{vi}}$ soared</td>
<td>crude</td>
<td>Item</td>
</tr>
<tr>
<td>stockpiles</td>
<td>Attribute</td>
</tr>
<tr>
<td>1.350 million barrels</td>
<td>Difference</td>
</tr>
<tr>
<td>Event $t\mathbf{vpe}$ $\therefore1$</td>
<td>December</td>
<td>Reference point</td>
</tr>
<tr>
<td>movement</td>
<td>200million barrels</td>
<td>Initial Value</td>
</tr>
<tr>
<td rowspan="2">$\operatorname{up}_{-}\operatorname{gain}$</td>
<td>438.9 million barrels</td>
<td>Final Value</td>
</tr>
<tr>
<td>more than 50 %</td>
<td>NONE</td>
</tr>
</tbody>
</table>
Table 1: Event extraction of "crude oil inventory in-
crease" event (first event in Figure 1 ).
$^2$https://www.ldc.upenn.edu/collaborations/past-
projects/ace
$^{3}$This example is used throughout this paper. | [
[
"0.2007",
"0.0944",
"0.797",
"0.0944",
"0.797",
"0.10925",
"0.2007",
"0.10925"
],
[
"0.3267",
"0.1437",
"0.6753",
"0.1443",
"0.6753",
"0.1592",
"0.3267",
"0.158"
],
[
"0.4207",
"0.1609",
"0.58",
"0.1621",
"0.58",
"0.177",
"0.4207",
"0.1758"
],
[
"0.2141",
"0.1787",
"0.8037",
"0.1787",
"0.8037",
"0.1936",
"0.2141",
"0.1936"
],
[
"0.2627",
"0.253",
"0.3418",
"0.253",
"0.3418",
"0.2673",
"0.2627",
"0.2673"
],
[
"0.5283",
"0.2725",
"0.8833",
"0.2725",
"0.8833",
"0.2861",
"0.5283",
"0.2861"
],
[
"0.147",
"0.2798",
"0.4568",
"0.2798",
"0.4568",
"0.2903",
"0.147",
"0.2903"
],
[
"0.1453",
"0.2922",
"0.4592",
"0.2922",
"0.4592",
"0.306",
"0.1453",
"0.306"
],
[
"0.5283",
"0.288",
"0.879",
"0.288",
"0.879",
"0.2986",
"0.5283",
"0.2986"
],
[
"0.525",
"0.3005",
"0.879",
"0.3005",
"0.879",
"0.3142",
"0.525",
"0.3142"
],
[
"0.1436",
"0.3071",
"0.4592",
"0.3071",
"0.4592",
"0.3206",
"0.1436",
"0.3206"
],
[
"0.1436",
"0.3206",
"0.4592",
"0.3218",
"0.4592",
"0.335",
"0.1436",
"0.3342"
],
[
"0.5283",
"0.3176",
"0.8447",
"0.3176",
"0.8447",
"0.3308",
"0.5283",
"0.3308"
],
[
"0.1436",
"0.3342",
"0.461",
"0.335",
"0.461",
"0.3486",
"0.1436",
"0.3474"
],
[
"0.1453",
"0.3499",
"0.4592",
"0.3499",
"0.4592",
"0.3635",
"0.1453",
"0.3635"
],
[
"0.1436",
"0.364",
"0.4592",
"0.364",
"0.4592",
"0.3777",
"0.1436",
"0.3777"
],
[
"0.511",
"0.3635",
"0.8833",
"0.3635",
"0.8833",
"0.3772",
"0.511",
"0.3772"
],
[
"0.1453",
"0.379",
"0.4592",
"0.379",
"0.4592",
"0.3926",
"0.1453",
"0.3926"
],
[
"0.511",
"0.3777",
"0.882",
"0.3777",
"0.882",
"0.3914",
"0.511",
"0.3914"
],
[
"0.147",
"0.3938",
"0.4575",
"0.3938",
"0.4575",
"0.4038",
"0.147",
"0.4038"
],
[
"0.511",
"0.3914",
"0.882",
"0.3914",
"0.882",
"0.405",
"0.511",
"0.405"
],
[
"0.1436",
"0.4062",
"0.4592",
"0.4062",
"0.4592",
"0.4192",
"0.1436",
"0.4192"
],
[
"0.511",
"0.4062",
"0.882",
"0.4062",
"0.882",
"0.4192",
"0.511",
"0.4192"
],
[
"0.1453",
"0.4204",
"0.4592",
"0.4204",
"0.4592",
"0.434",
"0.1453",
"0.434"
],
[
"0.511",
"0.4204",
"0.882",
"0.4204",
"0.882",
"0.434",
"0.511",
"0.434"
],
[
"0.1453",
"0.4353",
"0.4575",
"0.4353",
"0.4575",
"0.449",
"0.1453",
"0.449"
],
[
"0.511",
"0.434",
"0.6514",
"0.434",
"0.6514",
"0.4478",
"0.511",
"0.4478"
],
[
"0.1453",
"0.449",
"0.4592",
"0.449",
"0.4592",
"0.462",
"0.1453",
"0.462"
],
[
"0.1427",
"0.4614",
"0.4575",
"0.4631",
"0.4575",
"0.478",
"0.1427",
"0.4756"
],
[
"0.531",
"0.4702",
"0.881",
"0.4702",
"0.881",
"0.484",
"0.531",
"0.484"
],
[
"0.147",
"0.4785",
"0.4575",
"0.4785",
"0.4575",
"0.4893",
"0.147",
"0.4893"
],
[
"0.1436",
"0.4912",
"0.4592",
"0.4912",
"0.4592",
"0.505",
"0.1436",
"0.505"
],
[
"0.513",
"0.4863",
"0.881",
"0.4863",
"0.881",
"0.4995",
"0.513",
"0.4995"
],
[
"0.1453",
"0.506",
"0.4592",
"0.506",
"0.4592",
"0.5195",
"0.1453",
"0.5195"
],
[
"0.5146",
"0.503",
"0.881",
"0.503",
"0.881",
"0.5166",
"0.5146",
"0.5166"
],
[
"0.1436",
"0.521",
"0.4592",
"0.5195",
"0.4592",
"0.533",
"0.1436",
"0.5337"
],
[
"0.513",
"0.5186",
"0.881",
"0.5186",
"0.881",
"0.532",
"0.513",
"0.532"
],
[
"0.1453",
"0.535",
"0.4575",
"0.535",
"0.4575",
"0.549",
"0.1453",
"0.549"
],
[
"0.513",
"0.535",
"0.881",
"0.535",
"0.881",
"0.549",
"0.513",
"0.549"
],
[
"0.1453",
"0.55",
"0.455",
"0.55",
"0.455",
"0.56",
"0.1453",
"0.56"
],
[
"0.5146",
"0.5503",
"0.881",
"0.5503",
"0.881",
"0.564",
"0.5146",
"0.564"
],
[
"0.1453",
"0.561",
"0.4568",
"0.561",
"0.4568",
"0.5747",
"0.1453",
"0.5747"
],
[
"0.513",
"0.568",
"0.79",
"0.568",
"0.79",
"0.5806",
"0.513",
"0.5806"
],
[
"0.1453",
"0.5767",
"0.4592",
"0.5767",
"0.4592",
"0.5903",
"0.1453",
"0.5903"
],
[
"0.531",
"0.5845",
"0.881",
"0.583",
"0.881",
"0.597",
"0.531",
"0.5986"
],
[
"0.147",
"0.593",
"0.4575",
"0.593",
"0.4575",
"0.6025",
"0.147",
"0.6025"
],
[
"0.1453",
"0.604",
"0.4575",
"0.604",
"0.4575",
"0.6177",
"0.1453",
"0.6177"
],
[
"0.513",
"0.601",
"0.881",
"0.601",
"0.881",
"0.615",
"0.513",
"0.615"
],
[
"0.1436",
"0.619",
"0.4568",
"0.618",
"0.4568",
"0.632",
"0.1436",
"0.633"
],
[
"0.513",
"0.6177",
"0.882",
"0.6177",
"0.882",
"0.6313",
"0.513",
"0.6313"
],
[
"0.1436",
"0.632",
"0.4592",
"0.6343",
"0.4592",
"0.648",
"0.1436",
"0.6455"
],
[
"0.513",
"0.632",
"0.882",
"0.633",
"0.882",
"0.648",
"0.513",
"0.6465"
],
[
"0.1436",
"0.648",
"0.4592",
"0.648",
"0.4592",
"0.661",
"0.1436",
"0.661"
],
[
"0.513",
"0.6484",
"0.879",
"0.6484",
"0.879",
"0.663",
"0.513",
"0.663"
],
[
"0.1453",
"0.662",
"0.4592",
"0.662",
"0.4592",
"0.676",
"0.1453",
"0.676"
],
[
"0.5146",
"0.665",
"0.881",
"0.665",
"0.881",
"0.6787",
"0.5146",
"0.6787"
],
[
"0.1453",
"0.677",
"0.3242",
"0.677",
"0.3242",
"0.6904",
"0.1453",
"0.6904"
],
[
"0.513",
"0.681",
"0.8486",
"0.681",
"0.8486",
"0.695",
"0.513",
"0.695"
],
[
"0.1184",
"0.7036",
"0.1318",
"0.7036",
"0.1318",
"0.7134",
"0.1184",
"0.7134"
],
[
"0.1528",
"0.7017",
"0.2627",
"0.7017",
"0.2627",
"0.7134",
"0.1528",
"0.7134"
],
[
"0.513",
"0.7114",
"0.5635",
"0.7114",
"0.5635",
"0.7227",
"0.513",
"0.7227"
],
[
"0.614",
"0.7114",
"0.711",
"0.7114",
"0.711",
"0.7227",
"0.614",
"0.7227"
],
[
"0.7617",
"0.7114",
"0.86",
"0.7114",
"0.86",
"0.7227",
"0.7617",
"0.7227"
],
[
"0.517",
"0.7227",
"0.575",
"0.7227",
"0.575",
"0.737",
"0.517",
"0.737"
],
[
"0.614",
"0.7227",
"0.6455",
"0.7227",
"0.6455",
"0.7354",
"0.614",
"0.7354"
],
[
"0.763",
"0.7227",
"0.8213",
"0.7227",
"0.8213",
"0.737",
"0.763",
"0.737"
],
[
"0.76",
"0.734",
"0.796",
"0.7363",
"0.7944",
"0.7476",
"0.7583",
"0.745"
],
[
"0.11755",
"0.7256",
"0.4844",
"0.7246",
"0.4844",
"0.7393",
"0.11755",
"0.7407"
],
[
"0.5205",
"0.7363",
"0.5684",
"0.7363",
"0.5684",
"0.7466",
"0.5205",
"0.7466"
],
[
"0.6113",
"0.7354",
"0.6543",
"0.7354",
"0.6543",
"0.7466",
"0.6113",
"0.7466"
],
[
"0.11755",
"0.741",
"0.4844",
"0.741",
"0.4844",
"0.755",
"0.11755",
"0.755"
],
[
"0.613",
"0.7476",
"0.681",
"0.7476",
"0.681",
"0.759",
"0.613",
"0.759"
],
[
"0.7617",
"0.7476",
"0.8228",
"0.7476",
"0.8228",
"0.759",
"0.7617",
"0.759"
],
[
"0.6094",
"0.7583",
"0.7446",
"0.759",
"0.7446",
"0.771",
"0.6094",
"0.7695"
],
[
"0.763",
"0.7583",
"0.831",
"0.7603",
"0.831",
"0.7715",
"0.763",
"0.7695"
],
[
"0.5205",
"0.7695",
"0.594",
"0.7715",
"0.594",
"0.783",
"0.5205",
"0.781"
],
[
"0.11755",
"0.759",
"0.486",
"0.759",
"0.486",
"0.7725",
"0.11755",
"0.7725"
],
[
"0.6113",
"0.771",
"0.681",
"0.771",
"0.681",
"0.782",
"0.6113",
"0.782"
],
[
"0.763",
"0.771",
"0.8657",
"0.771",
"0.8657",
"0.782",
"0.763",
"0.782"
],
[
"0.614",
"0.782",
"0.733",
"0.782",
"0.733",
"0.792",
"0.614",
"0.792"
],
[
"0.1184",
"0.775",
"0.4878",
"0.775",
"0.4878",
"0.788",
"0.1184",
"0.788"
],
[
"0.5225",
"0.783",
"0.5938",
"0.783",
"0.5938",
"0.7935",
"0.5225",
"0.7935"
],
[
"0.764",
"0.782",
"0.8403",
"0.782",
"0.8403",
"0.7935",
"0.764",
"0.7935"
],
[
"0.11755",
"0.789",
"0.486",
"0.7905",
"0.486",
"0.8047",
"0.11755",
"0.804"
],
[
"0.6113",
"0.7915",
"0.7446",
"0.7935",
"0.7446",
"0.807",
"0.6113",
"0.8047"
],
[
"0.76",
"0.7905",
"0.839",
"0.792",
"0.839",
"0.807",
"0.76",
"0.8047"
],
[
"0.7617",
"0.8047",
"0.81",
"0.8047",
"0.81",
"0.8164",
"0.7617",
"0.8164"
],
[
"0.5225",
"0.796",
"0.575",
"0.796",
"0.575",
"0.807",
"0.5225",
"0.807"
],
[
"0.1184",
"0.807",
"0.486",
"0.807",
"0.486",
"0.821",
"0.1184",
"0.821"
],
[
"0.6113",
"0.807",
"0.7153",
"0.8047",
"0.7153",
"0.8164",
"0.6113",
"0.8184"
],
[
"0.1184",
"0.8223",
"0.4836",
"0.8223",
"0.4836",
"0.836",
"0.1184",
"0.836"
],
[
"0.51",
"0.8306",
"0.881",
"0.8306",
"0.881",
"0.844",
"0.51",
"0.844"
],
[
"0.11755",
"0.839",
"0.4878",
"0.839",
"0.4878",
"0.8525",
"0.11755",
"0.8525"
],
[
"0.513",
"0.8457",
"0.771",
"0.8457",
"0.771",
"0.8594",
"0.513",
"0.8594"
],
[
"0.11755",
"0.855",
"0.486",
"0.855",
"0.486",
"0.869",
"0.11755",
"0.869"
],
[
"0.1184",
"0.8716",
"0.2284",
"0.8716",
"0.2284",
"0.8857",
"0.1184",
"0.8857"
],
[
"0.532",
"0.88",
"0.825",
"0.88",
"0.825",
"0.8936",
"0.532",
"0.8936"
],
[
"0.1377",
"0.891",
"0.3804",
"0.8926",
"0.3804",
"0.9067",
"0.1377",
"0.9062"
],
[
"0.511",
"0.8945",
"0.586",
"0.8926",
"0.586",
"0.903",
"0.511",
"0.905"
],
[
"0.11584",
"0.9062",
"0.403",
"0.9062",
"0.403",
"0.9194",
"0.11584",
"0.9194"
],
[
"0.532",
"0.903",
"0.804",
"0.905",
"0.804",
"0.9194",
"0.532",
"0.9175"
]
] | [
"有效利用图卷积网络和上下文子树进行商品新闻事件提取",
"Meisin Lee, Lay-Ki Soon, Eu-Gene Siew",
"Monash University",
"( mei.lee, soon.layki, siew.eu-gene J @monash.edu",
"Abstract",
"U.S.crude stockpiles soared by1.350million bar-",
"Event extraction in commodity news is a less",
"researched area as compared to generic event",
"relsinDecemberfromamere2o0millionbarrelsto",
"438.9millionbarrels,duetothis oversupplycrude",
"extraction. However, accurate event extrac-",
"tion from commodity news is useful in a",
"oil prices plunged more than50%on Tuesday.",
"broad range of applications such as under-",
"standing event chains and learning event-event",
"relations, which can then be used for com-",
"Figure 1: An example of a sentence from a piece of",
"modity price prediction.The events found",
"commodity news 3, loaded with three events (trigger",
"incommoditynewsexhibitcharacteristics dif",
"words are in bold): (1) Crude oil inventory increase, (2)",
"ferent from generic events, hence posing a",
"Oversupply and (3) Crude Oil price decrease. Three",
"unique challenge in event extraction using ex-",
"event arguments that are of the same entity type are",
"isting methods. This paper proposes an ef-",
"highlighted in color.",
"fective use of Graph Convolutional Networks",
"(GCN) with a pruned dependency parse tree,",
"根据ACE（自动内容提取）程序2中的定义，",
"termed contextual sub-tree, for better event ex-",
"traction in commodity news. The event ex-",
"事件提取任务由两个子任务组成：（1）事件触",
"traction model is trained using feature embed",
"发器提取（识别和分类事件触发器）和（2）事",
"dings from ComBERT, a BERT-based masked",
"件参数提取（识别事件触发器的参数并标记其",
"language model that was produced through",
"角色）。在这项工作中，我们对(Leeetal.,2021)",
"domain-adaptive pre-training on a commodity",
"中引入的商品新闻数据集执行事件提取。图1",
"news corpus. Experimental results show the",
"显示了此数据集中的一个示例句子",
"efficiency of the proposed solution, which out-",
"为了说明事件提取的任务，请考虑图1中的",
"performsexisting methodswithFlscores as",
"high as 0.90. Furthermore, our pre-trained lan",
"示例以及表1中列出的相应信息。在事件触发",
"guage model outperforms GloVe by 23 %,and",
"提取子任务中，模型被训练以识别触发词：飙",
"BERT and RoBERTa by 7 % in terms of argu-",
"升并分类正确的事件类型：movement_up_gain，",
"ment roles classification. For the goal of re-",
"而在事件参数提取子任务中，模型被训练以从",
"producibility, the code and trained models are",
"句子中的实体提及池中识别事件参数，然后标",
"made publicly available",
"记每个实体与所识别事件相关的参数角色。",
"1",
"Introduction",
"Event",
"EntityMention",
"Argument Role",
"Trigger:",
"U.S.",
"Supplier",
"Item",
"诸如地缘政治和宏观经济相关事件之类的世界",
"soared",
"crude",
"事件已显示出对短期和长期（BrandtandGao,",
"stockpiles",
"Attribute",
"1.350millionbarrels",
"Difference",
"Event type:",
"2019）商品价格的影响。通常，在商品新闻文",
"December",
"Referencepoint",
"200 million barrels",
"章中发现的事件可以分为地缘政治，宏观经济，",
"movement",
"Initial Value",
"供需相关和商品价格变动。商品新闻是提取和",
"438.9million barrels",
"Final Value",
"NONE",
"up_gain",
"挖掘此类事件的宝贵信息来源。准确的事件提",
"morethan50%",
"取对于许多重要的下游任务非常有用，例如了",
"Table I: Event extraction of \"crude oil inventory in",
"解事件链和学习事件序列，也称为（Schank and",
"crease\" event (first event in Figure 1 ).",
"Abelson,2013）中的脚本，可用于最终准确的",
"商品价格预测",
"2https://www.ldc.upenn.edu/collaborations/past-",
"'https://github.com/meisin/",
"projects/ace",
"Commodity-News-Event-Extraction",
"3This example is used throughout this paper."
] | {
"arxiv_id": "2109.12781",
"language": "zh",
"page_idx": 1
} |
||
2109.12781-zh-2 | 2109.12781-zh-2.jpg | <FigureHere>
Figure 2: The example in Figure l is shown here with annotation provided in (Lee et al., 2021); trigger word is underlined and in bold, entity type for each entity mention is shown in blue below its respective word in BIOtagging format while event argument role for each entity is shown in light yellow above the words; arches link argument to its trigger word.
商品新闻中的事件类型与 ACE2005$^{4}$中的通用事件有很大不同。以下是这些事件的独特特征列表：
点之间的现有依赖路径添加了更多上下文；因
此它被称为上下文子树
这项工作的贡献如下：(1)我们表明，在对事件进行微调时，域自适应的预训练语言模型ComBERT 可以比通用的预训练语言模型产生有希望的性能提取任务，以及(2)我们提出了一种有效使用图形神经网络的方法，其形式为具有上下文子树的 GCN,其性能优于其他现有方法。在事件参数分类的子任务中，有用性尤其明显。
## 2 Commodity News Dataset
(Lee et al., 2021) 中引人的数据集是带注释的商品新闻文章的集合，其注释基于规范程序(如ACE 和 TAC-KBP 5)引入的标准。图 2 以图形形式捕获了数据集中的注释：(1)实体提及-命名和标称(BIO 标记格式的句子下方以蓝色显示的实体类型);(2)事件触发词(带下划线的单词和粗体);(3)参数角色(标签在句子上方为淡黄色),其中拱门将每个参数链接到其所属的事件。此数据集包含 21 个实体类型、 18 个事件类型和 19 个参数角色类型 (每个事件类型都有自己的参数角色集。事件类型是根据 Ravenpack 的$^{6}$事件分类定义的，其中选择了 18 种事件类型。就大小而言，此数据集包含8,850个实体提及和 3,949 个事件。表 2 显示了事件类型分布。有关数据集的更多信息，包括参数角色、实体类型和示例事件类型，请参见附录 A
1. Number intensity- Numbers (e.g., price, dif-
ference, percentage of change) and dates (including date of the opening price, dates of closing price) are abundant in commodity news. These numerical data is critical in expressing financial information. Generic information extraction methods may not work well for numeric data, as it is clearly seen in (Saha et al., 2017) where the authors introduced the first Open numerical relation extractor specifically to extract Open Information Extraction (IE) tuples that contain numbers or a quantity-unit phrase.
2. Arguments homogeneity - Many arguments
of the same entity type plays distinct roles in an event. Figure 2 shows that 1.350 million barrels , 200 million barrels , 438.0 million barrels are tagged as QUANTITY (see BIO-tagging at the bottom row). However, all three arguments play a different role in relation to the event (see Argument roles at the $\operatorname{top}$row).
3. Undifferentiated event types without its ar-
guments - Consider the event triggered by soared" in Figure 1 ,it is incomplete to identify the event (movement-up-gain) without knowing“what”soared. In commodity news, possible events are ‘price soared’,‘supply soared' , and 'demand soared' . To disambiguate events and represent them accurately, it is important to extract event arguments accurately as well.
考虑到这些挑战，本文提出了一种将图卷积网络(GCN)应用于上下文子树的解决方案， 用于事件提取任务。上下文子树是唯一修剪的依赖项解析树，它不仅提供依赖项路径信息， 还提供偏离路径信息。偏离路径信息为两个节
ACE2005 is a multilingual Training Corpus developed
### 3 $\textbf{Related Work}$
Event Extraction in Finance/Economics. 调查论文 (Xiang and Wang, 2019) 和 (Hogenboom et al., 2016) 中详细介绍了与通用事件提取相关的工作。在这里，我们特别关注金融和经济学领域内的事件提取。此领域中的大多数事件提取任务都侧重于提取与公司相关的事件。以下是最近文献中相关方法的总结：
$\overline{^5\text{https://tac.ni}}$st.gov/2015/KBP
$^{6}$RavenPack is an analytics provider for fnancial services. Among their products are finance and economid data. More information can be found on their page: https://www.ravenpack.com/
under the ACE program | [
[
"0.1318",
"0.0974",
"0.1797",
"0.0974",
"0.1797",
"0.1081",
"0.1318",
"0.1081"
],
[
"0.388",
"0.0974",
"0.4307",
"0.0974",
"0.4307",
"0.1081",
"0.388",
"0.1081"
],
[
"0.4902",
"0.0968",
"0.555",
"0.0968",
"0.555",
"0.1081",
"0.4902",
"0.1081"
],
[
"0.2225",
"0.0986",
"0.2603",
"0.0986",
"0.2603",
"0.1081",
"0.2225",
"0.1081"
],
[
"0.1755",
"0.0986",
"0.2091",
"0.0986",
"0.2091",
"0.1057",
"0.1755",
"0.1057"
],
[
"0.6665",
"0.0986",
"0.7095",
"0.0986",
"0.7095",
"0.1057",
"0.6665",
"0.1057"
],
[
"0.796",
"0.0986",
"0.8364",
"0.0986",
"0.8364",
"0.1057",
"0.796",
"0.1057"
],
[
"0.147",
"0.1081",
"0.881",
"0.1081",
"0.881",
"0.1217",
"0.147",
"0.1217"
],
[
"0.11755",
"0.12054",
"0.3198",
"0.12054",
"0.3198",
"0.1342",
"0.11755",
"0.1342"
],
[
"0.3267",
"0.1217",
"0.4902",
"0.1217",
"0.4902",
"0.1318",
"0.3267",
"0.1318"
],
[
"0.4978",
"0.12054",
"0.879",
"0.12054",
"0.879",
"0.1342",
"0.4978",
"0.1342"
],
[
"0.11755",
"0.1455",
"0.882",
"0.1455",
"0.882",
"0.1592",
"0.11755",
"0.1592"
],
[
"0.11584",
"0.1603",
"0.882",
"0.1603",
"0.882",
"0.1733",
"0.11584",
"0.1733"
],
[
"0.11584",
"0.1746",
"0.882",
"0.1729",
"0.882",
"0.1859",
"0.11584",
"0.1882"
],
[
"0.11584",
"0.1882",
"0.3123",
"0.1882",
"0.3123",
"0.2031",
"0.11584",
"0.2031"
],
[
"0.1351",
"0.228",
"0.486",
"0.2268",
"0.486",
"0.2411",
"0.1351",
"0.2423"
],
[
"0.513",
"0.2286",
"0.879",
"0.2286",
"0.879",
"0.2423",
"0.513",
"0.2423"
],
[
"0.11755",
"0.2446",
"0.4844",
"0.2446",
"0.4844",
"0.2583",
"0.11755",
"0.2583"
],
[
"0.513",
"0.2446",
"0.6987",
"0.2446",
"0.6987",
"0.2583",
"0.513",
"0.2583"
],
[
"0.11755",
"0.26",
"0.2007",
"0.26",
"0.2007",
"0.275",
"0.11755",
"0.275"
],
[
"0.531",
"0.2612",
"0.881",
"0.2612",
"0.881",
"0.275",
"0.531",
"0.275"
],
[
"0.1318",
"0.2808",
"0.4902",
"0.2808",
"0.4902",
"0.2957",
"0.1318",
"0.2957"
],
[
"0.513",
"0.2756",
"0.879",
"0.2756",
"0.879",
"0.2893",
"0.513",
"0.2893"
],
[
"0.1511",
"0.2976",
"0.4902",
"0.2964",
"0.4902",
"0.3113",
"0.1511",
"0.3123"
],
[
"0.513",
"0.2922",
"0.881",
"0.2922",
"0.881",
"0.306",
"0.513",
"0.306"
],
[
"0.1545",
"0.3142",
"0.4895",
"0.3142",
"0.4895",
"0.328",
"0.1545",
"0.328"
],
[
"0.513",
"0.3088",
"0.879",
"0.3088",
"0.879",
"0.3225",
"0.513",
"0.3225"
],
[
"0.1528",
"0.33",
"0.4878",
"0.33",
"0.4878",
"0.3445",
"0.1528",
"0.3445"
],
[
"0.5283",
"0.326",
"0.881",
"0.326",
"0.881",
"0.339",
"0.5283",
"0.339"
],
[
"0.1528",
"0.3445",
"0.4895",
"0.3445",
"0.4895",
"0.3582",
"0.1528",
"0.3582"
],
[
"0.513",
"0.3403",
"0.879",
"0.3403",
"0.879",
"0.354",
"0.513",
"0.354"
],
[
"0.1511",
"0.3623",
"0.4902",
"0.361",
"0.4902",
"0.376",
"0.1511",
"0.3772"
],
[
"0.511",
"0.357",
"0.881",
"0.357",
"0.881",
"0.3706",
"0.511",
"0.3706"
],
[
"0.511",
"0.3718",
"0.596",
"0.3735",
"0.596",
"0.3884",
"0.511",
"0.386"
],
[
"0.1528",
"0.379",
"0.4878",
"0.379",
"0.4878",
"0.3926",
"0.1528",
"0.3926"
],
[
"0.1528",
"0.3938",
"0.486",
"0.3938",
"0.486",
"0.4067",
"0.1528",
"0.4067"
],
[
"0.517",
"0.4038",
"0.523",
"0.4038",
"0.523",
"0.4104",
"0.517",
"0.4104"
],
[
"0.546",
"0.401",
"0.7715",
"0.401",
"0.7715",
"0.4146",
"0.546",
"0.4146"
],
[
"0.1528",
"0.4104",
"0.4878",
"0.4104",
"0.4878",
"0.4233",
"0.1528",
"0.4233"
],
[
"0.1528",
"0.427",
"0.4895",
"0.427",
"0.4895",
"0.4407",
"0.1528",
"0.4407"
],
[
"0.511",
"0.427",
"0.881",
"0.427",
"0.881",
"0.4407",
"0.511",
"0.4407"
],
[
"0.1511",
"0.4424",
"0.4895",
"0.4412",
"0.4895",
"0.456",
"0.1511",
"0.4573"
],
[
"0.511",
"0.4424",
"0.879",
"0.4424",
"0.879",
"0.456",
"0.511",
"0.456"
],
[
"0.1545",
"0.4602",
"0.4895",
"0.4602",
"0.4895",
"0.4702",
"0.1545",
"0.4702"
],
[
"0.511",
"0.4573",
"0.881",
"0.4578",
"0.881",
"0.4727",
"0.511",
"0.4714"
],
[
"0.1528",
"0.4744",
"0.3745",
"0.4744",
"0.3745",
"0.488",
"0.1528",
"0.488"
],
[
"0.513",
"0.4756",
"0.881",
"0.4756",
"0.881",
"0.4893",
"0.513",
"0.4893"
],
[
"0.1301",
"0.4912",
"0.4878",
"0.4912",
"0.4878",
"0.505",
"0.1301",
"0.505"
],
[
"0.513",
"0.4912",
"0.881",
"0.4912",
"0.881",
"0.505",
"0.513",
"0.505"
],
[
"0.1528",
"0.5073",
"0.4895",
"0.5073",
"0.4895",
"0.521",
"0.1528",
"0.521"
],
[
"0.511",
"0.506",
"0.881",
"0.5073",
"0.881",
"0.5215",
"0.511",
"0.521"
],
[
"0.1528",
"0.524",
"0.4895",
"0.524",
"0.4895",
"0.5376",
"0.1528",
"0.5376"
],
[
"0.511",
"0.524",
"0.879",
"0.524",
"0.879",
"0.5376",
"0.511",
"0.5376"
],
[
"0.1528",
"0.539",
"0.4895",
"0.539",
"0.4895",
"0.5527",
"0.1528",
"0.5527"
],
[
"0.513",
"0.538",
"0.879",
"0.538",
"0.879",
"0.552",
"0.513",
"0.552"
],
[
"0.1528",
"0.5547",
"0.4878",
"0.5547",
"0.4878",
"0.5693",
"0.1528",
"0.5693"
],
[
"0.513",
"0.5557",
"0.882",
"0.5557",
"0.882",
"0.5693",
"0.513",
"0.5693"
],
[
"0.1528",
"0.572",
"0.4878",
"0.572",
"0.4878",
"0.586",
"0.1528",
"0.586"
],
[
"0.513",
"0.572",
"0.881",
"0.572",
"0.881",
"0.585",
"0.513",
"0.585"
],
[
"0.1511",
"0.5874",
"0.4902",
"0.586",
"0.4902",
"0.601",
"0.1511",
"0.6016"
],
[
"0.5146",
"0.5874",
"0.882",
"0.5874",
"0.882",
"0.6016",
"0.5146",
"0.6016"
],
[
"0.1528",
"0.604",
"0.4895",
"0.604",
"0.4895",
"0.6177",
"0.1528",
"0.6177"
],
[
"0.511",
"0.6025",
"0.882",
"0.6016",
"0.882",
"0.616",
"0.511",
"0.6177"
],
[
"0.1545",
"0.6216",
"0.2225",
"0.6216",
"0.2225",
"0.633",
"0.1545",
"0.633"
],
[
"0.5156",
"0.619",
"0.881",
"0.619",
"0.881",
"0.633",
"0.5156",
"0.633"
],
[
"0.1293",
"0.6343",
"0.4878",
"0.6353",
"0.4878",
"0.6494",
"0.1293",
"0.6484"
],
[
"0.513",
"0.6357",
"0.8774",
"0.6357",
"0.8774",
"0.6494",
"0.513",
"0.6494"
],
[
"0.1545",
"0.652",
"0.486",
"0.652",
"0.486",
"0.665",
"0.1545",
"0.665"
],
[
"0.513",
"0.651",
"0.881",
"0.652",
"0.881",
"0.6665",
"0.513",
"0.665"
],
[
"0.1561",
"0.6675",
"0.4895",
"0.6675",
"0.4895",
"0.681",
"0.1561",
"0.681"
],
[
"0.5146",
"0.6675",
"0.881",
"0.6675",
"0.881",
"0.6816",
"0.5146",
"0.6816"
],
[
"0.1528",
"0.684",
"0.4895",
"0.683",
"0.4895",
"0.6978",
"0.1528",
"0.6987"
],
[
"0.513",
"0.685",
"0.5693",
"0.685",
"0.5693",
"0.697",
"0.513",
"0.697"
],
[
"0.1511",
"0.6997",
"0.4844",
"0.7007",
"0.4844",
"0.7144",
"0.1511",
"0.7134"
],
[
"0.511",
"0.712",
"0.55",
"0.712",
"0.55",
"0.7236",
"0.511",
"0.7236"
],
[
"0.5425",
"0.7114",
"0.671",
"0.7114",
"0.671",
"0.7256",
"0.5425",
"0.7256"
],
[
"0.1511",
"0.731",
"0.4895",
"0.7324",
"0.4895",
"0.7466",
"0.1511",
"0.745"
],
[
"0.1511",
"0.7153",
"0.4844",
"0.7163",
"0.4844",
"0.731",
"0.1511",
"0.73"
],
[
"0.5205",
"0.737",
"0.882",
"0.737",
"0.882",
"0.7505",
"0.5205",
"0.7505"
],
[
"0.1528",
"0.7476",
"0.486",
"0.7476",
"0.486",
"0.761",
"0.1528",
"0.761"
],
[
"0.513",
"0.753",
"0.882",
"0.753",
"0.882",
"0.7666",
"0.513",
"0.7666"
],
[
"0.1503",
"0.763",
"0.4878",
"0.7656",
"0.4878",
"0.78",
"0.1503",
"0.778"
],
[
"0.511",
"0.7686",
"0.8774",
"0.7686",
"0.8774",
"0.782",
"0.511",
"0.782"
],
[
"0.1528",
"0.781",
"0.2747",
"0.781",
"0.2747",
"0.7944",
"0.1528",
"0.7944"
],
[
"0.511",
"0.7837",
"0.879",
"0.7837",
"0.879",
"0.7974",
"0.511",
"0.7974"
],
[
"0.1351",
"0.8003",
"0.486",
"0.8003",
"0.486",
"0.814",
"0.1351",
"0.814"
],
[
"0.5146",
"0.802",
"0.881",
"0.802",
"0.881",
"0.8154",
"0.5146",
"0.8154"
],
[
"0.1184",
"0.818",
"0.4844",
"0.818",
"0.4844",
"0.8306",
"0.1184",
"0.8306"
],
[
"0.5146",
"0.8193",
"0.8774",
"0.8193",
"0.8774",
"0.83",
"0.5146",
"0.83"
],
[
"0.12006",
"0.8345",
"0.486",
"0.8345",
"0.486",
"0.847",
"0.12006",
"0.847"
],
[
"0.513",
"0.8345",
"0.76",
"0.8345",
"0.76",
"0.847",
"0.513",
"0.847"
],
[
"0.1184",
"0.8496",
"0.4844",
"0.8496",
"0.4844",
"0.8633",
"0.1184",
"0.8633"
],
[
"0.534",
"0.857",
"0.723",
"0.857",
"0.723",
"0.8706",
"0.534",
"0.8706"
],
[
"0.11755",
"0.865",
"0.486",
"0.8643",
"0.486",
"0.879",
"0.11755",
"0.88"
],
[
"0.532",
"0.868",
"0.881",
"0.869",
"0.881",
"0.883",
"0.532",
"0.882"
],
[
"0.511",
"0.882",
"0.882",
"0.882",
"0.882",
"0.8955",
"0.511",
"0.8955"
],
[
"0.1368",
"0.891",
"0.4878",
"0.8936",
"0.4878",
"0.9077",
"0.1368",
"0.9062"
],
[
"0.513",
"0.8936",
"0.546",
"0.8955",
"0.544",
"0.9067",
"0.511",
"0.905"
],
[
"0.5576",
"0.8926",
"0.882",
"0.8945",
"0.882",
"0.9077",
"0.5576",
"0.9062"
],
[
"0.1142",
"0.904",
"0.2678",
"0.9062",
"0.2678",
"0.9194",
"0.1142",
"0.9175"
],
[
"0.511",
"0.905",
"0.6885",
"0.905",
"0.6885",
"0.9185",
"0.511",
"0.9185"
]
] | [
"Supplier",
"Dfference",
"Reference_ime",
"Attribute",
"Item",
"InitialValue",
"Final Value",
"U.S.crude stockpilessoaredby350millionbarrelsinDecemberfromamere200millionbarrelsto438.9milionbarrels",
"B-COUNTRYB-COMNOOIT-FINATT",
"OBAQUANTITYO",
"B-DATEOOQUANTITYQUANTTY",
"Figure 2: The example in Figure 1 is shown here with annotation provided in (Lee et al., 2021) ; trigger word",
"is underlined and in bold, entity type for each entity mention is shown in blue below its respective word in BIO-",
"tagging format while event argument role for each entity is shown in light yellow above the words; arches link",
"argument to its trigger word.",
"商品新闻中的事件类型与ACE20054中的",
"点之间的现有依赖路径添加了更多上下文；因",
"通用事件有很大不同。以下是这些事件的独特",
"此它被称为上下文子树",
"特征列表：",
"这项工作的贡献如下：（1）我们表明，在对",
"1. Number intensity - Numbers (e.g., price, dif-",
"事件进行微调时，域自适应的预训练语言模型",
"ference, percentage of change) and dates (in-",
"ComBERT可以比通用的预训练语言模型产生",
"cluding date of the opening price, dates of",
"有希望的性能提取任务，以及（2）我们提出了",
"closing price) are abundant in commodity",
"-种有效使用图形神经网络的方法，其形式为",
"news. These numerical data is critical in ex-",
"具有上下文子树的GCN，其性能优于其他现",
"pressing financial information. Generic in-",
"有方法。在事件参数分类的子任务中，有用性",
"尤其明显。",
"formation extraction methods may not work",
"well for numeric data, as it is clearly seen",
"2",
"Commodity News Dataset",
"in (Saha et al., 2017) where the authors in-",
"troduced the first Open numerical relation ex-",
"(Leeetal.,2021）中引人的数据集是带注释的商",
"tractor specifically to extract Open Informa-",
"品新闻文章的集合，其注释基于规范程序（如",
"tion Extraction (IE)tuples thatcontain num-",
"ACE和TAC-KBP5）引人的标准。图2以图形",
"bers or a quantity-unit phrase.",
"形式捕获了数据集中的注释：（1）实体提及-命",
"2. Arguments homogeneity - Many arguments",
"名和标称（BIO标记格式的句子下方以蓝色显",
"of the same entity type plays distinct roles",
"示的实体类型）；（2）事件触发词（带下划线",
"in an event. Figure 2 shows that 1.350 mil-",
"的单词和粗体）；（3）参数角色（标签在句子",
"lion barrels , 200 million barrels , 438.0 mil-",
"上方为淡黄色），其中拱门将每个参数链接到",
"lion barrels are tagged as QUANTITY (see",
"其所属的事件。此数据集包含21个实体类型、",
"BIO-tagging at the bottom row). However,",
"18个事件类型和19个参数角色类型（每个事",
"all three arguments play a different role in re-",
"件类型都有自己的参数角色集。事件类型是根",
"lation to the event (see Argument roles at the",
"据Ravenpack的6事件分类定义的，其中选择",
"top row).",
"了18种事件类型。就大小而言，此数据集包含",
"3. Undifferentiated event types without its ar-",
"8,850个实体提及和3,949个事件。表2显示了",
"guments - Consider the event triggered by",
"事件类型分布。有关数据集的更多信息，包括",
"\"soared\" in Figure 1 , it is incomplete to iden-",
"参数角色、实体类型和示例事件类型，请参见",
"tify the event (movement-up-gain) without",
"附录A",
"knowing“what\"soared.In commodity news",
"3",
"Related Work",
"soared', and 'demand soared' To disam-",
"possible events are‘price soared',‘supply",
"Event Extraction in Finance/Economics.调",
"biguate events and represent them accurately,",
"查论文（Xiang and Wang,2019）和(Hogenboom",
"it is important to extract event arguments ac-",
"etal.2016)中详细介绍了与通用事件提取相关",
"curately as well.",
"的工作。在这里，我们特别关注金融和经济学",
"考虑到这些挑战，本文提出了一种将图卷积",
"领域内的事件提取。此领域中的大多数事件提",
"网络（GCN）应用于上下文子树的解决方案",
"取任务都侧重于提取与公司相关的事件。以下",
"用于事件提取任务。上下文子树是唯一修剪的",
"是最近文献中相关方法的总结：",
"依赖项解析树，它不仅提供依赖项路径信息，",
"https://tac.nist.gov/2015/KBP",
"还提供偏离路径信息。偏离路径信息为两个节",
"6RavenPack is an analytics provider for financial ser",
"vices.Among their products are finance and economic",
"4ACE2005 is a multilingual Training Corpus developed",
"data.",
"More information can be found on their page:",
"under the ACE program",
"https://www.ravenpack.com/"
] | {
"arxiv_id": "2109.12781",
"language": "zh",
"page_idx": 2
} |
||
2109.12781-zh-3 | 2109.12781-zh-3.jpg | 1. Rule-based approach: authors in (Malik
et al., 2011) introduced statistical classifiers aided by rules, while authors in (Hogenboom et al., 2013) used rule-sets and domain ontology knowledge-bases together with other semantically-enabled components;
2. Usage of external resources: Semantic Frame
was in (Xie et al., 2013) and Wikipedia for weak supervision in (Ein-Dor et al., 2019) ;
3. Open IE (Open Information Extraction): au-
thors in (Ding et al., 2014) extracted events from news headline via Open IE; while in (Saha et al., 2017) , authors introduced an extension of Open IE to extract numerical arguments in each Open IE tuple.
4. Deep learning approach: (Yang et al., 2018)
proposed a deep learning approach to extract
financial events from Chinese text.
尽管公司财务事件和商品新闻属于同一领域，并且两者都可能涉及数字数据作为事件参数，但现有的公司财务事件提取方法相当有限。例如，(Saha et al., 2017) 中的解决方案适合于为每个 OpenIE 元组仅提取一个数字参数。相比之下，在(Yang et al., 2018) 中提取公司财务信息的任务在提取数值数据作为事件参数方面与这项工作最匹配。但是，它是针对中文文本的解决方案，并专注于文档级事件提取
Graph Convolutional Networks. 使用图卷积网络(GCN)以及来自依赖项解析树的语法信息已用于 (Nguyen and Grishman, 2018) 和 (Liu
<table>
<tbody>
<tr>
<th>$\overline{{\mathrm{Event~type}}}$</th>
<th>$\overline{\mathrm{Type}}$ ${\mathbf{c}}$ratio</th>
<th>井 sentence</th>
</tr>
<tr>
<td>1. Cause-movement-down- loss</td>
<td>$\overline{13.35}\%$</td>
<td>524</td>
</tr>
<tr>
<td>2. Cause-movement-up-gain</td>
<td>2.23 $\%$</td>
<td>88</td>
</tr>
<tr>
<td>3.Civil-unrest</td>
<td>2.53 $\%$</td>
<td>100</td>
</tr>
<tr>
<td>$\overline{4.\mathrm{~Crisis}}$</td>
<td>$0.76\%$</td>
<td>$\overline{30}$</td>
</tr>
<tr>
<td>5.Embargo</td>
<td>3.75 $\%$</td>
<td>148</td>
</tr>
<tr>
<td>6. Geopolitical-tension</td>
<td>1.70 $\%$</td>
<td>$\overline{67}$</td>
</tr>
<tr>
<td>7.Grow-strong</td>
<td>6.03 $\%$</td>
<td>$\overline{238}$</td>
</tr>
<tr>
<td>8. Movement-down-loss</td>
<td>22.69 $\%$</td>
<td>$\overline{896}$</td>
</tr>
<tr>
<td>9. Movement-flat</td>
<td>1.52 $\%$</td>
<td>$\overline{60}$</td>
</tr>
<tr>
<td>10. Movement-up-gain</td>
<td>22.13 $\%$</td>
<td>874</td>
</tr>
<tr>
<td>11. Negative-sentiment</td>
<td>$\overline{4.79}\%$</td>
<td>189</td>
</tr>
<tr>
<td>12. Oversupply</td>
<td>2.63 $\%$</td>
<td>104</td>
</tr>
<tr>
<td>13. Position-high</td>
<td>3.82 $\%$</td>
<td>151</td>
</tr>
<tr>
<td>14.Position-low</td>
<td>3.11 $\%$</td>
<td>123</td>
</tr>
<tr>
<td>15.Prohibiting</td>
<td>$1.06\%$</td>
<td>$\overline{42}$</td>
</tr>
<tr>
<td>16. Shortage</td>
<td>$1.04\%$</td>
<td>41</td>
</tr>
<tr>
<td>17. Slow- weak</td>
<td>5.47 $\%$</td>
<td>216</td>
</tr>
<tr>
<td>18.Trade-tensions</td>
<td>$1.39\%$</td>
<td>$\overline{55}$</td>
</tr>
<tr>
<td>$\overline{\text{Total}}$</td>
<td> </td>
<td>3949</td>
</tr>
</tbody>
</table>
Table 2: Event type distribution and sentence level
counts
et al., 2018) 中的事件提取。在 (Nguyen and Grishman,2018)中，作者建议在句子的句法依赖图上使用 GCN 来产生非连续的 k-gram, 作为将单词直接链接到其信息内容以进行事件检测的有效机制。另一方面，(Liu et al., 2018) 中的作者使用基于注意力的 GCN 对图形信息进行建模，以联合提取多个事件触发器和参数。他们提出的解决方案，联合多事件提取(JMEE) 框架，专注于建模事件之间的关联，以提高事件提取的准确性。这两种解决方案都使用最短的依赖路径
除了事件提取之外，GCN已成功用于(Zhang et al., 2018) 中的关系提取。(Zhang et al., 2018) 中的作者没有严格从最短的依赖路径中获取令牌，而是进行了修改，以生成修剪后的子依赖树，以包括路径信息以及否定提示词。在此处列出的相关工作中，在任务(事件提取)和范围 (句子级别)方面最接近我们的工作的是(Liu et al., 2018)的 JMEE
### 4 Proposed Solution
参考商品新闻中发现的事件的唯一特征列表(在$\mathbb{E}1$ 节中列出),本文提出了一种使用具有上下文子树的图卷积网络(GCN)的解决方案， 用于有效提取商品新闻中的事件。
句法依赖图将句子表示为有向树，相关单词之间具有头部修饰符依赖弧。这些图中的每个单词都被其直接的句法调控器和从属单词(邻居)包围，可以在最相关的单词上执行卷积运算并避免对不相关的单词进行建模。我们建议使用唯一修剪的依赖树，而不是使用完整的依赖解析树，该依赖树由两个节点之间的最短路径 (在我们的例子中-触发器候选和实体提及) 和其他非路径节点组成。包含路径外节点以提供其他上下文信息。生成的树称为上下文子树
### 4.1 Contextual Sub-tree
图 3 中显示了示例句子的依赖关系树，其中突出显示了许多候选上下文子树之一。受(Zhang et al., 2018) 的启发，我们对依赖树进行了修剪， 以获得以触发候选项和实体提及候选项之间的最小共同祖先(LCA)为根的子树，同时还包含非路径节点。这些偏离路径的节点提供了额外的关键上下文，可以在参数角色分类中获得更好的结果。路径外信息由与依赖路径相距$DIST$的令牌组成。算法 1 显示了如何构建上下文子树的步骤。如(Zhang et al., 2018) 所示， $DIST=1$在包含上下文信息和尽可能将不相关的信息排除在结果子树中之间实现了最佳平衡。
上下文子树的使用针对参数角色分类，事件 | [
[
"0.1184",
"0.082",
"0.1948",
"0.08435",
"0.1948",
"0.0956",
"0.1184",
"0.0932"
],
[
"0.3267",
"0.08435",
"0.482",
"0.08435",
"0.482",
"0.0944",
"0.3267",
"0.0944"
],
[
"0.511",
"0.0849",
"0.879",
"0.0849",
"0.879",
"0.0986",
"0.511",
"0.0986"
],
[
"0.1628",
"0.0968",
"0.3132",
"0.0968",
"0.3132",
"0.1069",
"0.1628",
"0.1069"
],
[
"0.3435",
"0.0956",
"0.3972",
"0.0956",
"0.3972",
"0.1069",
"0.3435",
"0.1069"
],
[
"0.43",
"0.0968",
"0.4568",
"0.0932",
"0.4592",
"0.1051",
"0.4324",
"0.1081"
],
[
"0.12427",
"0.1057",
"0.1545",
"0.1081",
"0.1528",
"0.1194",
"0.12177",
"0.1164"
],
[
"0.511",
"0.10156",
"0.881",
"0.10156",
"0.881",
"0.1152",
"0.511",
"0.1152"
],
[
"0.12177",
"0.11755",
"0.3032",
"0.12054",
"0.3032",
"0.1342",
"0.12177",
"0.1312"
],
[
"0.3494",
"0.1194",
"0.3938",
"0.1194",
"0.3938",
"0.1312",
"0.3494",
"0.1312"
],
[
"0.434",
"0.1194",
"0.4568",
"0.1194",
"0.4568",
"0.1318",
"0.434",
"0.1318"
],
[
"0.5146",
"0.11816",
"0.881",
"0.11816",
"0.881",
"0.1318",
"0.5146",
"0.1318"
],
[
"0.3494",
"0.13",
"0.3938",
"0.1318",
"0.392",
"0.1443",
"0.3477",
"0.1426"
],
[
"0.434",
"0.1312",
"0.4575",
"0.1312",
"0.4575",
"0.1437",
"0.434",
"0.1437"
],
[
"0.511",
"0.133",
"0.879",
"0.1342",
"0.879",
"0.1484",
"0.511",
"0.1478"
],
[
"0.434",
"0.1437",
"0.4568",
"0.1437",
"0.4568",
"0.1561",
"0.434",
"0.1561"
],
[
"0.12427",
"0.1318",
"0.2167",
"0.1318",
"0.2167",
"0.1437",
"0.12427",
"0.1437"
],
[
"0.12177",
"0.1443",
"0.178",
"0.1443",
"0.178",
"0.1561",
"0.12177",
"0.1561"
],
[
"0.3494",
"0.1443",
"0.3972",
"0.1443",
"0.3972",
"0.1561",
"0.3494",
"0.1561"
],
[
"0.434",
"0.1526",
"0.4592",
"0.1567",
"0.4568",
"0.1704",
"0.43",
"0.1675"
],
[
"0.513",
"0.1509",
"0.881",
"0.1509",
"0.881",
"0.1644",
"0.513",
"0.1644"
],
[
"0.1234",
"0.1567",
"0.2007",
"0.1567",
"0.2007",
"0.1687",
"0.1234",
"0.1687"
],
[
"0.3494",
"0.1567",
"0.3972",
"0.1567",
"0.3972",
"0.1687",
"0.3494",
"0.1687"
],
[
"0.12006",
"0.1687",
"0.2678",
"0.1704",
"0.2678",
"0.1818",
"0.12006",
"0.1799"
],
[
"0.3494",
"0.1692",
"0.3938",
"0.1692",
"0.3938",
"0.1812",
"0.3494",
"0.1812"
],
[
"0.4358",
"0.1704",
"0.4517",
"0.1704",
"0.4517",
"0.1799",
"0.4358",
"0.1799"
],
[
"0.5156",
"0.1687",
"0.8774",
"0.1687",
"0.8774",
"0.1787",
"0.5156",
"0.1787"
],
[
"0.12177",
"0.1799",
"0.22",
"0.1818",
"0.22",
"0.1936",
"0.12177",
"0.1912"
],
[
"0.345",
"0.1787",
"0.3997",
"0.1787",
"0.3997",
"0.1936",
"0.345",
"0.1936"
],
[
"0.4324",
"0.1799",
"0.4575",
"0.1799",
"0.4575",
"0.1924",
"0.4324",
"0.1924"
],
[
"0.511",
"0.1812",
"0.881",
"0.1818",
"0.881",
"0.1965",
"0.511",
"0.1953"
],
[
"0.11755",
"0.1912",
"0.2761",
"0.1912",
"0.2761",
"0.2048",
"0.11755",
"0.2048"
],
[
"0.3418",
"0.1936",
"0.398",
"0.1936",
"0.398",
"0.2048",
"0.3418",
"0.2048"
],
[
"0.4307",
"0.1912",
"0.4592",
"0.1948",
"0.4575",
"0.206",
"0.4282",
"0.2031"
],
[
"0.5146",
"0.199",
"0.881",
"0.199",
"0.881",
"0.212",
"0.5146",
"0.212"
],
[
"0.1184",
"0.2048",
"0.2351",
"0.2048",
"0.2351",
"0.2185",
"0.1184",
"0.2185"
],
[
"0.3477",
"0.206",
"0.3972",
"0.206",
"0.3972",
"0.2173",
"0.3477",
"0.2173"
],
[
"0.434",
"0.2048",
"0.4568",
"0.2048",
"0.4568",
"0.2173",
"0.434",
"0.2173"
],
[
"0.12006",
"0.2173",
"0.2688",
"0.2173",
"0.2688",
"0.231",
"0.12006",
"0.231"
],
[
"0.3418",
"0.2185",
"0.3972",
"0.2185",
"0.3972",
"0.2299",
"0.3418",
"0.2299"
],
[
"0.513",
"0.2131",
"0.879",
"0.2144",
"0.879",
"0.2286",
"0.513",
"0.228"
],
[
"0.434",
"0.2197",
"0.4568",
"0.2197",
"0.4568",
"0.2299",
"0.434",
"0.2299"
],
[
"0.1234",
"0.228",
"0.2678",
"0.2299",
"0.2678",
"0.2434",
"0.1234",
"0.2411"
],
[
"0.3477",
"0.2299",
"0.3972",
"0.2299",
"0.3972",
"0.2411",
"0.3477",
"0.2411"
],
[
"0.434",
"0.2299",
"0.4575",
"0.2299",
"0.4575",
"0.2423",
"0.434",
"0.2423"
],
[
"0.513",
"0.231",
"0.881",
"0.231",
"0.881",
"0.2446",
"0.513",
"0.2446"
],
[
"0.12427",
"0.2393",
"0.2217",
"0.2423",
"0.2217",
"0.257",
"0.12427",
"0.2542"
],
[
"0.3477",
"0.2393",
"0.398",
"0.2411",
"0.3972",
"0.2559",
"0.346",
"0.2542"
],
[
"0.4358",
"0.2393",
"0.4592",
"0.2434",
"0.455",
"0.2559",
"0.4307",
"0.2517"
],
[
"0.3477",
"0.253",
"0.3972",
"0.253",
"0.3972",
"0.2673",
"0.3477",
"0.2673"
],
[
"0.513",
"0.2477",
"0.607",
"0.2477",
"0.607",
"0.2625",
"0.513",
"0.2625"
],
[
"0.126",
"0.2546",
"0.2301",
"0.2546",
"0.2301",
"0.2666",
"0.126",
"0.2666"
],
[
"0.4324",
"0.2542",
"0.4568",
"0.2542",
"0.4568",
"0.2673",
"0.4324",
"0.2673"
],
[
"0.12006",
"0.2642",
"0.2275",
"0.2666",
"0.2275",
"0.2808",
"0.12006",
"0.279"
],
[
"0.4324",
"0.2666",
"0.4575",
"0.2666",
"0.4575",
"0.2798",
"0.4324",
"0.2798"
],
[
"0.531",
"0.2632",
"0.881",
"0.2632",
"0.881",
"0.2778",
"0.531",
"0.2778"
],
[
"0.12177",
"0.2778",
"0.22",
"0.2798",
"0.22",
"0.2935",
"0.12177",
"0.2915"
],
[
"0.3494",
"0.2673",
"0.398",
"0.2673",
"0.398",
"0.279",
"0.3494",
"0.279"
],
[
"0.3494",
"0.2798",
"0.3972",
"0.2798",
"0.3972",
"0.2915",
"0.3494",
"0.2915"
],
[
"0.434",
"0.2798",
"0.455",
"0.2798",
"0.455",
"0.2922",
"0.434",
"0.2922"
],
[
"0.513",
"0.2798",
"0.881",
"0.2798",
"0.881",
"0.2935",
"0.513",
"0.2935"
],
[
"0.12006",
"0.2915",
"0.2048",
"0.2935",
"0.2048",
"0.3047",
"0.12006",
"0.3027"
],
[
"0.3494",
"0.2903",
"0.3997",
"0.2903",
"0.3997",
"0.3047",
"0.3494",
"0.3047"
],
[
"0.434",
"0.2915",
"0.455",
"0.2915",
"0.455",
"0.304",
"0.434",
"0.304"
],
[
"0.513",
"0.2944",
"0.879",
"0.2944",
"0.879",
"0.308",
"0.513",
"0.308"
],
[
"0.12177",
"0.3047",
"0.2183",
"0.3047",
"0.2183",
"0.3154",
"0.12177",
"0.3154"
],
[
"0.3494",
"0.3047",
"0.3972",
"0.3047",
"0.3972",
"0.3164",
"0.3494",
"0.3164"
],
[
"0.4324",
"0.304",
"0.4575",
"0.304",
"0.4575",
"0.3164",
"0.4324",
"0.3164"
],
[
"0.1184",
"0.3142",
"0.2418",
"0.3154",
"0.2418",
"0.33",
"0.1184",
"0.3289"
],
[
"0.513",
"0.31",
"0.881",
"0.3113",
"0.881",
"0.326",
"0.513",
"0.3247"
],
[
"0.3494",
"0.3164",
"0.398",
"0.3164",
"0.398",
"0.328",
"0.3494",
"0.328"
],
[
"0.434",
"0.3164",
"0.455",
"0.3164",
"0.455",
"0.3289",
"0.434",
"0.3289"
],
[
"0.1184",
"0.3308",
"0.1621",
"0.3308",
"0.1621",
"0.3425",
"0.1184",
"0.3425"
],
[
"0.4282",
"0.3308",
"0.4626",
"0.3308",
"0.4626",
"0.3425",
"0.4282",
"0.3425"
],
[
"0.513",
"0.3267",
"0.879",
"0.3267",
"0.879",
"0.3403",
"0.513",
"0.3403"
],
[
"0.5146",
"0.3433",
"0.881",
"0.3433",
"0.881",
"0.357",
"0.5146",
"0.357"
],
[
"0.11755",
"0.3557",
"0.4878",
"0.3557",
"0.4878",
"0.3694",
"0.11755",
"0.3694"
],
[
"0.513",
"0.3594",
"0.881",
"0.3599",
"0.881",
"0.3748",
"0.513",
"0.3735"
],
[
"0.1184",
"0.3718",
"0.1628",
"0.3718",
"0.1628",
"0.383",
"0.1184",
"0.383"
],
[
"0.51",
"0.376",
"0.7095",
"0.376",
"0.7095",
"0.3896",
"0.51",
"0.3896"
],
[
"0.5283",
"0.411",
"0.5415",
"0.411",
"0.5415",
"0.415",
"0.5283",
"0.415"
],
[
"0.546",
"0.4067",
"0.705",
"0.4067",
"0.705",
"0.4204",
"0.546",
"0.4204"
],
[
"0.1318",
"0.415",
"0.486",
"0.415",
"0.486",
"0.43",
"0.1318",
"0.43"
],
[
"0.513",
"0.4133",
"0.523",
"0.408",
"0.529",
"0.4133",
"0.519",
"0.4187"
],
[
"0.1561",
"0.4329",
"0.4844",
"0.4329",
"0.4844",
"0.4436",
"0.1561",
"0.4436"
],
[
"0.513",
"0.4316",
"0.881",
"0.4329",
"0.881",
"0.4478",
"0.513",
"0.4465"
],
[
"0.1545",
"0.4478",
"0.486",
"0.4478",
"0.486",
"0.4614",
"0.1545",
"0.4614"
],
[
"0.513",
"0.4495",
"0.879",
"0.4495",
"0.879",
"0.4631",
"0.513",
"0.4631"
],
[
"0.1528",
"0.4631",
"0.486",
"0.4631",
"0.486",
"0.4768",
"0.1528",
"0.4768"
],
[
"0.5156",
"0.466",
"0.882",
"0.466",
"0.882",
"0.4797",
"0.5156",
"0.4797"
],
[
"0.1528",
"0.4797",
"0.486",
"0.4797",
"0.486",
"0.4946",
"0.1528",
"0.4946"
],
[
"0.513",
"0.4822",
"0.7793",
"0.4822",
"0.7793",
"0.4954",
"0.513",
"0.4954"
],
[
"0.1528",
"0.4963",
"0.4106",
"0.4963",
"0.4106",
"0.5103",
"0.1528",
"0.5103"
],
[
"0.532",
"0.4995",
"0.881",
"0.4995",
"0.881",
"0.513",
"0.532",
"0.513"
],
[
"0.1318",
"0.5127",
"0.486",
"0.5127",
"0.486",
"0.5254",
"0.1318",
"0.5254"
],
[
"0.513",
"0.5156",
"0.879",
"0.5156",
"0.879",
"0.5293",
"0.513",
"0.5293"
],
[
"0.1528",
"0.528",
"0.4878",
"0.528",
"0.4878",
"0.5415",
"0.1528",
"0.5415"
],
[
"0.5146",
"0.531",
"0.882",
"0.531",
"0.882",
"0.5444",
"0.5146",
"0.5444"
],
[
"0.1545",
"0.5444",
"0.4785",
"0.5444",
"0.4785",
"0.558",
"0.1545",
"0.558"
],
[
"0.513",
"0.5474",
"0.881",
"0.5474",
"0.881",
"0.561",
"0.513",
"0.561"
],
[
"0.1301",
"0.56",
"0.4895",
"0.56",
"0.4895",
"0.5737",
"0.1301",
"0.5737"
],
[
"0.513",
"0.5635",
"0.881",
"0.5635",
"0.881",
"0.5767",
"0.513",
"0.5767"
],
[
"0.1561",
"0.5767",
"0.4844",
"0.5767",
"0.4844",
"0.5874",
"0.1561",
"0.5874"
],
[
"0.511",
"0.5776",
"0.879",
"0.5776",
"0.879",
"0.5913",
"0.511",
"0.5913"
],
[
"0.1545",
"0.593",
"0.4878",
"0.593",
"0.4878",
"0.6055",
"0.1545",
"0.6055"
],
[
"0.511",
"0.5933",
"0.881",
"0.594",
"0.881",
"0.6094",
"0.511",
"0.608"
],
[
"0.1561",
"0.6094",
"0.4878",
"0.6094",
"0.4878",
"0.619",
"0.1561",
"0.619"
],
[
"0.513",
"0.611",
"0.881",
"0.611",
"0.881",
"0.6245",
"0.513",
"0.6245"
],
[
"0.1545",
"0.6245",
"0.4895",
"0.6245",
"0.4895",
"0.638",
"0.1545",
"0.638"
],
[
"0.5156",
"0.629",
"0.8774",
"0.629",
"0.8774",
"0.6396",
"0.5156",
"0.6396"
],
[
"0.1545",
"0.641",
"0.3652",
"0.641",
"0.3652",
"0.655",
"0.1545",
"0.655"
],
[
"0.5146",
"0.6436",
"0.881",
"0.6436",
"0.881",
"0.6567",
"0.5146",
"0.6567"
],
[
"0.1301",
"0.6567",
"0.4844",
"0.6567",
"0.4844",
"0.6714",
"0.1301",
"0.6714"
],
[
"0.1545",
"0.6733",
"0.4878",
"0.6733",
"0.4878",
"0.687",
"0.1545",
"0.687"
],
[
"0.513",
"0.6733",
"0.5557",
"0.6733",
"0.5557",
"0.684",
"0.513",
"0.684"
],
[
"0.5557",
"0.673",
"0.7197",
"0.673",
"0.7197",
"0.686",
"0.5557",
"0.686"
],
[
"0.1545",
"0.6895",
"0.409",
"0.6895",
"0.409",
"0.703",
"0.1545",
"0.703"
],
[
"0.513",
"0.6934",
"0.881",
"0.695",
"0.881",
"0.709",
"0.513",
"0.708"
],
[
"0.1368",
"0.709",
"0.4878",
"0.7104",
"0.4878",
"0.7246",
"0.1368",
"0.7236"
],
[
"0.513",
"0.7114",
"0.881",
"0.7114",
"0.881",
"0.7246",
"0.513",
"0.7246"
],
[
"0.1184",
"0.727",
"0.4844",
"0.727",
"0.4844",
"0.7407",
"0.1184",
"0.7407"
],
[
"0.511",
"0.727",
"0.882",
"0.727",
"0.882",
"0.7407",
"0.511",
"0.7407"
],
[
"0.12006",
"0.7437",
"0.4836",
"0.7437",
"0.4836",
"0.7573",
"0.12006",
"0.7573"
],
[
"0.513",
"0.7437",
"0.881",
"0.7437",
"0.881",
"0.7573",
"0.513",
"0.7573"
],
[
"0.11755",
"0.7583",
"0.4844",
"0.7583",
"0.4844",
"0.7715",
"0.11755",
"0.7715"
],
[
"0.511",
"0.7583",
"0.881",
"0.759",
"0.881",
"0.774",
"0.511",
"0.7725"
],
[
"0.12006",
"0.7754",
"0.4836",
"0.7754",
"0.4836",
"0.789",
"0.12006",
"0.789"
],
[
"0.513",
"0.775",
"0.8774",
"0.775",
"0.8774",
"0.788",
"0.513",
"0.788"
],
[
"0.11755",
"0.7905",
"0.4836",
"0.7905",
"0.4836",
"0.804",
"0.11755",
"0.804"
],
[
"0.513",
"0.7915",
"0.881",
"0.7915",
"0.881",
"0.8047",
"0.513",
"0.8047"
],
[
"0.1184",
"0.808",
"0.4844",
"0.808",
"0.4844",
"0.822",
"0.1184",
"0.822"
],
[
"0.513",
"0.808",
"0.881",
"0.808",
"0.881",
"0.822",
"0.513",
"0.822"
],
[
"0.1184",
"0.8237",
"0.486",
"0.8237",
"0.486",
"0.8374",
"0.1184",
"0.8374"
],
[
"0.511",
"0.822",
"0.881",
"0.8223",
"0.881",
"0.8374",
"0.511",
"0.836"
],
[
"0.1184",
"0.8403",
"0.4492",
"0.8403",
"0.4492",
"0.854",
"0.1184",
"0.854"
],
[
"0.5146",
"0.8384",
"0.882",
"0.839",
"0.882",
"0.854",
"0.5146",
"0.8525"
],
[
"0.5146",
"0.8555",
"0.881",
"0.8555",
"0.881",
"0.869",
"0.5146",
"0.869"
],
[
"0.1234",
"0.8726",
"0.3862",
"0.8726",
"0.3862",
"0.8857",
"0.1234",
"0.8857"
],
[
"0.3938",
"0.8735",
"0.486",
"0.8735",
"0.486",
"0.8853",
"0.3938",
"0.8853"
],
[
"0.5146",
"0.8726",
"0.881",
"0.8726",
"0.881",
"0.8857",
"0.5146",
"0.8857"
],
[
"0.11755",
"0.887",
"0.486",
"0.888",
"0.486",
"0.903",
"0.11755",
"0.902"
],
[
"0.513",
"0.8896",
"0.54",
"0.8896",
"0.54",
"0.902",
"0.513",
"0.902"
],
[
"0.12006",
"0.905",
"0.486",
"0.905",
"0.486",
"0.9185",
"0.12006",
"0.9185"
],
[
"0.532",
"0.905",
"0.879",
"0.905",
"0.879",
"0.9185",
"0.532",
"0.9185"
]
] | [
"Eventtype",
"Typeratio#sentence",
"etal.,2018）中的事件提取。在（NguyenandGr",
"Cause-movement-down-",
"13.35%",
"524",
"loss",
"ishman,2018）中，作者建议在句子的句法依赖",
"2. Cause-movement-up-gain",
"2.23 %",
"88",
"图上使用GCN来产生非连续的k-gram，作为",
"2.53%",
"100",
"将单词直接链接到其信息内容以进行事件检测",
"30",
"3.Civil-unrest",
"4.Crisis",
"0.76%",
"148",
"的有效机制。另一方面，(Liuetal.,2018）中的",
"5.Embargo",
"3.75 %",
"6.Geopolitical-tension",
"1.70%",
"67",
"作者使用基于注意力的GCN对图形信息进行",
"7.Grow-strong",
"6.03%",
"238",
"建模，以联合提取多个事件触发器和参数。他",
"8.Movement-down-loss",
"22.69%",
"896",
"们提出的解决方案，联合多事件提取（JMEE）",
"9.Movement-flat",
"1.52%",
"60",
"10. Movement-up-gain",
"22.13 %",
"框架，专注于建模事件之间的关联，以提高事",
"874",
"11.Negative-sentiment",
"4.79 %",
"189",
"件提取的准确性。这两种解决方案都使用最短",
"12.Oversupply",
"2.63%",
"104",
"3.82%",
"的依赖路径",
"13.Position-high",
"151",
"14. Position-low",
"123",
"除了事件提取之外，GCN已成功用于(Zhang",
"15.Prohibiting",
"3.11 %",
"1.06%",
"42",
"etal.,2018）中的关系提取。(Zhangetal.,2018)",
"16.Shortage",
"1.04 %",
"41",
"中的作者没有严格从最短的依赖路径中获取",
"17.Slow-weak",
"5.47%",
"216",
"18.Trade-tensions",
"令牌，而是进行了修改，以生成修剪后的子依",
"1.39%",
"55",
"Total",
"3949",
"赖树，以包括路径信息以及否定提示词。在此",
"处列出的相关工作中，在任务（事件提取）和",
"Table 2: Event type distribution and sentence level",
"范围（句子级别）方面最接近我们的工作的是",
"counts",
"(Liu et al., 2018) 的 JMEE",
"",
"Proposed Solution",
"1. Rule-based approach: authors in (Malik",
"Y",
"et al.,2011)introduced statisticalclassifiers",
"参考商品新闻中发现的事件的唯一特征列表",
"aided by rules, while authors in (Hogenboom",
"（在1节中列出），本文提出了一种使用具有",
"et al., 2013) used rule-sets and domain on",
"上下文子树的图卷积网络（GCN）的解决方案，",
"tology knowledge-bases together with other",
"用于有效提取商品新闻中的事件。",
"semantically-enabled components;",
"句法依赖图将句子表示为有向树，相关单词",
"2. Usage of external resources: Semantic Frame",
"之间具有头部修饰符依赖弧。这些图中的每个",
"was in (Xie et al., 2013) and Wikipedia for",
"单词都被其直接的句法调控器和从属单词（邻",
"weak supervision in (Ein-Dor et al., 2019) ;",
"居）包围，可以在最相关的单词上执行卷积运",
"3. Open IE (Open Information Extraction): au-",
"算并避免对不相关的单词进行建模。我们建议",
"thors in (Ding et al..2014)extracted events",
"使用唯一修剪的依赖树，而不是使用完整的依",
"from news headline via Open IE; while in",
"赖解析树，该依赖树由两个节点之间的最短路",
"(Sahaetal..2017).authorsintroducedanex-",
"径（在我们的例子中-触发器候选和实体提及）",
"tension of Open IE to extract numerical argu-",
"和其他非路径节点组成。包含路径外节点以提",
"ments in each Open IE tuple",
"供其他上下文信息。生成的树称为上下文子树",
"4. Deep learning approach: (Yang et al., 2018)",
"proposed a deep learning approach to extract",
"4.1",
"Contextual Sub-tree",
"financial events from Chinese text.",
"图3中显示了示例句子的依赖关系树，其中突",
"尽管公司财务事件和商品新闻属于同一领",
"出显示了许多候选上下文子树之一。受(Zhang",
"域，并且两者都可能涉及数字数据作为事件",
"etal.,2018)的启发，我们对依赖树进行了修剪，",
"参数，但现有的公司财务事件提取方法相当有",
"以获得以触发候选项和实体提及候选项之间",
"限。例如，（Sahaet al.,2017）中的解决方案适",
"的最小共同祖先（LCA）为根的子树，同时还",
"合于为每个OpenIE元组仅提取一个数字参数。",
"包含非路径节点。这些偏离路径的节点提供了",
"相比之下，在（Yangetal.,2018）中提取公司财",
"额外的关键上下文，可以在参数角色分类中获",
"务信息的任务在提取数值数据作为事件参数方",
"得更好的结果。路径外信息由与依赖路径相距",
"面与这项工作最匹配。但是，它是针对中文文",
"DIST的令牌组成。算法1显示了如何构建上",
"本的解决方案，并专注于文档级事件提取",
"下文子树的步骤。如(Zhangetal.,2018）所示",
"DIST=1在包含上下文信息和尽可能将不相",
"Graph Convolutional Networks.",
"使用图卷积",
"关的信息排除在结果子树中之间实现了最佳平",
"网络（GCN）以及来自依赖项解析树的语法信",
"衡。",
"息已用于（NguyenandGrishman，2018）和（Liu",
"上下文子树的使用针对参数角色分类，事件"
] | {
"arxiv_id": "2109.12781",
"language": "zh",
"page_idx": 3
} |
||
2310.13135-zh-0 | 2310.13135-zh-0.jpg | LeTFuser: Light-weight End-to-end Transformer-Based Sensor Fusion for
Autonomous Driving with Multi-Task Learning
Pedram Agand* Mohammad Mahdavian* Manolis Savva Mo Chen Department of Computing Science, Simon Fraser University, BC, Canada.
$\{$ pedram\_agand, mohammad\_ mahdavian, manolis\_savva, mochen $\}$ @sfu.ca
THMALIN3 0
$\operatorname{Abstract}$
In end-to-end autonomous driving, the utilization of existing sensor fusion techniques and navigational control methods for imitation learning proves inadequate in challenging situations that involve numerous dynamic agents. To address this issue, we introduce LeTFuser, a lightweight transformer-based algorithm for fusing multiple RGB-D camera representations. To perform perception and control tasks simultaneously, we utilize multi-task learning. Our model comprises of two modules, the frst being the perception module that is responsible for encoding the observation data obtained from the RGB-D cameras. Our approach employs the Convolutional vision Transformer (CvT)[?~] to better extract and fuse features from multiple RGB cameras due to local and global feature extraction capability of convolution and transformer modules, respectively.
Encoded features combined with static and dynamic environments are later employed by our control module to predict waypoints and vehicular controls (e.g. steering, throttle, and brake). We use two methods to generate the vehicular controls levels. The first method uses a PID algorithm to follow the waypoints on the fly, whereas the second one directly predicts the control policy using the measurement features and environmental state. We evaluate the model and conduct a comparative analysis with recent models on the CARLA simulator using various scenarios, ranging from normal to adversarial conditions, to simulate real-world scenarios. Our method demonstrated better or comparable results with respect to our baselines in term of driving abilities. The code is available at https://github.com/pagand/e2etransfuser/tree/cvprw to facilitate future studies.
## 1. Introduction
Many works in the autonomous driving literature have been 专注于安全导航[????????]的感知和控制任务的不同方面。端到端驱动 neural network ( NN) models 的最新进展已经证明了使用单模态输入的显著结果，例如 image 和 LiDAR [?]。然而，由于缺乏对 3D 场景的理解，这些方法在涉及对抗性情况的复杂城市场景中面临局限性[?]。传感器融合已显示出通过集成多个传感器模式(如相机和 LiDAR 传感器)来解决这些挑战的希望，以创建更全面的场景表示[??] 。$\mathbb{E}$尽管有所改进，但这些融合方法通常需要大量计算资源，并且在感知和控制任务「?]之间平衡学习信号方面面临挑战。此外，集成具有不同数据形状和表示的多个模态需要复杂的预处理技术 such as ELPP $[?]~,~$SaDMS$[?]~,~$从而导致模型复杂性增加和信息丢失的可能性。端到端自动驾驶的最新进展已经探索了不同传感器模式的集成，例如 LiDAR 和 RGB-D 相机，以提高性能。? ] employs a 卷积神经网络 (CNN) 提取 an RGB-D 相机和 produce 未来车辆航路点和导航信号提供的数据特征。另一种众所周知的方法是TransFuser [?] ,它使用多模式融合转换器将全局上下文和成对交互合并到不同输入模式的特征提取层中。We were inspired by these methods and demonstrated that combining ideas from these approaches could potentially lead to more robust and accurate end-to-end autonomous driving solutions. On the perception side, we leverage the strengths of both global and local context reasoning provided by transformers and CNNs. In the control side, we take advantage of trajectory-guided control, which we introduce next.
The Trajectory-guided Control Prediction (TCP) [? ] framework, which combines trajectory planning and direct control is a multi-task learning framework for end-to-end autonomous driving. By incorporating a multi-step control prediction branch with a dynamic branch and trajectory-guided attention, TCP can improve temporal reasoning and achieve superior performance in the CARLA driving simulator, even surpassing methods using multiple cameras and LiDAR sen- | [
[
"0.1177",
"0.1338",
"0.851",
"0.1338",
"0.851",
"0.1483",
"0.1177",
"0.1483"
],
[
"0.2467",
"0.1547",
"0.7256",
"0.1572",
"0.7256",
"0.1749",
"0.2467",
"0.1724"
],
[
"0.1683",
"0.2052",
"0.726",
"0.2052",
"0.726",
"0.2197",
"0.1683",
"0.2197"
],
[
"0.721",
"0.2052",
"0.8037",
"0.2052",
"0.8037",
"0.2197",
"0.721",
"0.2197"
],
[
"0.1748",
"0.2235",
"0.796",
"0.2235",
"0.796",
"0.238",
"0.1748",
"0.238"
],
[
"0.2255",
"0.2424",
"0.7427",
"0.2424",
"0.7427",
"0.2556",
"0.2255",
"0.2556"
],
[
"0.0989",
"0.2922",
"0.2051",
"0.2922",
"0.2051",
"0.307",
"0.0989",
"0.307"
],
[
"0.5034",
"0.2922",
"0.6406",
"0.2922",
"0.6406",
"0.307",
"0.5034",
"0.307"
],
[
"0.2361",
"0.305",
"0.3145",
"0.307",
"0.3145",
"0.3245",
"0.2361",
"0.3225"
],
[
"0.5215",
"0.3225",
"0.8906",
"0.3225",
"0.8906",
"0.336",
"0.5215",
"0.336"
],
[
"0.501",
"0.3352",
"0.8936",
"0.336",
"0.8936",
"0.3542",
"0.501",
"0.3528"
],
[
"0.0989",
"0.356",
"0.468",
"0.356",
"0.468",
"0.3694",
"0.0989",
"0.3694"
],
[
"0.505",
"0.3528",
"0.892",
"0.3528",
"0.892",
"0.3662",
"0.505",
"0.3662"
],
[
"0.0801",
"0.3718",
"0.4697",
"0.3718",
"0.4697",
"0.385",
"0.0801",
"0.385"
],
[
"0.505",
"0.3687",
"0.8896",
"0.3687",
"0.8896",
"0.3818",
"0.505",
"0.3818"
],
[
"0.0801",
"0.387",
"0.468",
"0.387",
"0.468",
"0.401",
"0.0801",
"0.401"
],
[
"0.5034",
"0.3838",
"0.888",
"0.3838",
"0.888",
"0.3977",
"0.5034",
"0.3977"
],
[
"0.0809",
"0.4016",
"0.4697",
"0.4016",
"0.4697",
"0.4153",
"0.0809",
"0.4153"
],
[
"0.505",
"0.3984",
"0.8896",
"0.3984",
"0.8896",
"0.4124",
"0.505",
"0.4124"
],
[
"0.0809",
"0.4172",
"0.4697",
"0.4172",
"0.4697",
"0.4312",
"0.0809",
"0.4312"
],
[
"0.5024",
"0.414",
"0.8896",
"0.414",
"0.8896",
"0.428",
"0.5024",
"0.428"
],
[
"0.0801",
"0.432",
"0.468",
"0.432",
"0.468",
"0.4458",
"0.0801",
"0.4458"
],
[
"0.5034",
"0.4287",
"0.8906",
"0.4287",
"0.8906",
"0.4426",
"0.5034",
"0.4426"
],
[
"0.0801",
"0.4463",
"0.4697",
"0.4463",
"0.4697",
"0.4602",
"0.0801",
"0.4602"
],
[
"0.505",
"0.4458",
"0.8896",
"0.4458",
"0.8896",
"0.4558",
"0.505",
"0.4558"
],
[
"0.0809",
"0.4622",
"0.468",
"0.4622",
"0.468",
"0.476",
"0.0809",
"0.476"
],
[
"0.5063",
"0.459",
"0.8906",
"0.459",
"0.8906",
"0.473",
"0.5063",
"0.473"
],
[
"0.0801",
"0.4766",
"0.4697",
"0.4766",
"0.4697",
"0.4905",
"0.0801",
"0.4905"
],
[
"0.505",
"0.4734",
"0.8906",
"0.4734",
"0.8906",
"0.4873",
"0.505",
"0.4873"
],
[
"0.0809",
"0.4924",
"0.468",
"0.4924",
"0.468",
"0.506",
"0.0809",
"0.506"
],
[
"0.505",
"0.4893",
"0.892",
"0.4893",
"0.892",
"0.5024",
"0.505",
"0.5024"
],
[
"0.0825",
"0.509",
"0.4673",
"0.509",
"0.4673",
"0.5195",
"0.0825",
"0.5195"
],
[
"0.505",
"0.504",
"0.8906",
"0.504",
"0.8906",
"0.5176",
"0.505",
"0.5176"
],
[
"0.0801",
"0.523",
"0.4673",
"0.523",
"0.4673",
"0.536",
"0.0801",
"0.536"
],
[
"0.5034",
"0.518",
"0.8906",
"0.518",
"0.8906",
"0.532",
"0.5034",
"0.532"
],
[
"0.0809",
"0.537",
"0.4673",
"0.537",
"0.4673",
"0.5503",
"0.0809",
"0.5503"
],
[
"0.5034",
"0.5327",
"0.8906",
"0.534",
"0.8906",
"0.549",
"0.5034",
"0.5474"
],
[
"0.0809",
"0.5537",
"0.468",
"0.5537",
"0.468",
"0.5645",
"0.0809",
"0.5645"
],
[
"0.5073",
"0.55",
"0.8896",
"0.55",
"0.8896",
"0.563",
"0.5073",
"0.563"
],
[
"0.0809",
"0.5693",
"0.4656",
"0.5693",
"0.4656",
"0.58",
"0.0809",
"0.58"
],
[
"0.505",
"0.5645",
"0.8906",
"0.5645",
"0.8906",
"0.5776",
"0.505",
"0.5776"
],
[
"0.0801",
"0.5835",
"0.468",
"0.5835",
"0.468",
"0.5967",
"0.0801",
"0.5967"
],
[
"0.5034",
"0.579",
"0.8906",
"0.579",
"0.8906",
"0.5933",
"0.5034",
"0.5933"
],
[
"0.0809",
"0.5996",
"0.1544",
"0.5996",
"0.1544",
"0.6113",
"0.0809",
"0.6113"
],
[
"0.5034",
"0.595",
"0.8906",
"0.595",
"0.8906",
"0.6094",
"0.5034",
"0.6094"
],
[
"0.5034",
"0.61",
"0.8906",
"0.61",
"0.8906",
"0.6235",
"0.5034",
"0.6235"
],
[
"0.0989",
"0.629",
"0.468",
"0.629",
"0.468",
"0.6426",
"0.0989",
"0.6426"
],
[
"0.5063",
"0.625",
"0.8906",
"0.625",
"0.8906",
"0.638",
"0.5063",
"0.638"
],
[
"0.0809",
"0.6445",
"0.4673",
"0.6445",
"0.4673",
"0.6577",
"0.0809",
"0.6577"
],
[
"0.5034",
"0.6396",
"0.8936",
"0.6396",
"0.8936",
"0.653",
"0.5034",
"0.653"
],
[
"0.0801",
"0.66",
"0.4673",
"0.66",
"0.4673",
"0.674",
"0.0801",
"0.674"
],
[
"0.5024",
"0.655",
"0.8906",
"0.655",
"0.8906",
"0.6694",
"0.5024",
"0.6694"
],
[
"0.0784",
"0.675",
"0.3064",
"0.674",
"0.3064",
"0.688",
"0.0784",
"0.6895"
],
[
"0.3228",
"0.675",
"0.468",
"0.675",
"0.468",
"0.688",
"0.3228",
"0.688"
],
[
"0.5034",
"0.672",
"0.8906",
"0.672",
"0.8906",
"0.6816",
"0.5034",
"0.6816"
],
[
"0.0801",
"0.69",
"0.4697",
"0.69",
"0.4697",
"0.704",
"0.0801",
"0.704"
],
[
"0.5024",
"0.685",
"0.8906",
"0.685",
"0.8906",
"0.6987",
"0.5024",
"0.6987"
],
[
"0.505",
"0.6997",
"0.8906",
"0.6997",
"0.8906",
"0.7134",
"0.505",
"0.7134"
],
[
"0.0801",
"0.7056",
"0.4697",
"0.7056",
"0.4697",
"0.7197",
"0.0801",
"0.7197"
],
[
"0.0801",
"0.72",
"0.4697",
"0.72",
"0.4697",
"0.7344",
"0.0801",
"0.7344"
],
[
"0.5034",
"0.7163",
"0.892",
"0.7163",
"0.892",
"0.73",
"0.5034",
"0.73"
],
[
"0.5024",
"0.731",
"0.8936",
"0.731",
"0.8936",
"0.744",
"0.5024",
"0.744"
],
[
"0.0809",
"0.7373",
"0.4656",
"0.7373",
"0.4656",
"0.7476",
"0.0809",
"0.7476"
],
[
"0.0809",
"0.7505",
"0.468",
"0.7505",
"0.468",
"0.7646",
"0.0809",
"0.7646"
],
[
"0.5024",
"0.747",
"0.744",
"0.747",
"0.744",
"0.7603",
"0.5024",
"0.7603"
],
[
"0.0809",
"0.7676",
"0.4673",
"0.7676",
"0.4673",
"0.778",
"0.0809",
"0.778"
],
[
"0.5215",
"0.7646",
"0.8906",
"0.765",
"0.8906",
"0.78",
"0.5215",
"0.779"
],
[
"0.0809",
"0.781",
"0.468",
"0.781",
"0.468",
"0.7944",
"0.0809",
"0.7944"
],
[
"0.505",
"0.781",
"0.892",
"0.781",
"0.892",
"0.7944",
"0.505",
"0.7944"
],
[
"0.0809",
"0.7974",
"0.468",
"0.7974",
"0.468",
"0.808",
"0.0809",
"0.808"
],
[
"0.5024",
"0.797",
"0.892",
"0.797",
"0.892",
"0.81",
"0.5024",
"0.81"
],
[
"0.0809",
"0.8125",
"0.4673",
"0.8125",
"0.4673",
"0.8228",
"0.0809",
"0.8228"
],
[
"0.5024",
"0.811",
"0.892",
"0.811",
"0.892",
"0.8247",
"0.5024",
"0.8247"
],
[
"0.0801",
"0.827",
"0.4697",
"0.827",
"0.4697",
"0.8403",
"0.0801",
"0.8403"
],
[
"0.5024",
"0.827",
"0.892",
"0.827",
"0.892",
"0.8403",
"0.5024",
"0.8403"
],
[
"0.0801",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0801",
"0.855"
],
[
"0.5024",
"0.8413",
"0.892",
"0.8413",
"0.892",
"0.855",
"0.5024",
"0.855"
],
[
"0.0784",
"0.856",
"0.4673",
"0.857",
"0.4673",
"0.8706",
"0.0784",
"0.869"
],
[
"0.5024",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.5024",
"0.8706"
],
[
"0.0801",
"0.872",
"0.2876",
"0.872",
"0.2876",
"0.8853",
"0.0801",
"0.8853"
],
[
"0.5034",
"0.872",
"0.892",
"0.872",
"0.892",
"0.8853",
"0.5034",
"0.8853"
],
[
"0.5024",
"0.886",
"0.892",
"0.886",
"0.892",
"0.901",
"0.5024",
"0.901"
]
] | [
"LeTFuser: Light-weight End-to-end Transformer-Based Sensor Fusion for",
"Autonomous Driving with Multi-Task Learning",
"Pedram Agand*Mohammad Mahdavian*Manolis Savva",
"Mo Chen",
"Department of Computing Science, Simon Fraser University, BC, Canada.",
"( pedram_agand, mohammad_mahdavian, manolis_savva, mochen ↓ @sfu.ca",
"THMALIN3 0",
"1.Introduction",
"Abstract",
"Many works in the autonomous driving literature",
"havebeen专注于安全导航[？？？？？？？]】的感",
"In end-to-end autonomous driving, the utilization",
"知和控制任务的不同方面。端到端驱动neuralnetwork",
"of existing sensor fusion techniques and navigational",
"(NN）models的最新进展已经证明了使用单模态输入",
"control methods for imitation learning proves inade-",
"的显著结果，例如image和LiDAR[？。然而，由",
"quate in challenging situations that involve numerous",
"于缺乏对3D场景的理解，这些方法在涉及对抗性情况",
"dynamic agents. To address this issue, we introduce",
"的复杂城市场景中面临局限性[？】。传感器融合已显",
"LeTFuser, a lightweight transformer-based algorithm",
"示出通过集成多个传感器模式（如相机和LiDAR传感",
"for fusing multiple RGB-D camera representations. To",
"器）来解决这些挑战的希望，以创建更全面的场景表示",
"perform perception and control tasks simultaneously,",
"[？？］。E尽管有所改进，但这些融合方法通常需要大",
"we utilize multi-task learning. Our model comprises",
"量计算资源，并且在感知和控制任务[？1之间平衡学",
"of two modules, the first being the perception module",
"习信号方面面临挑战。此外，集成具有不同数据形状和",
"that is responsiblefor encoding the observation data",
"表示的多个模态需要复杂的预处理技术suchasELPP",
"obtained from the RGB-D cameras. Our approach em-",
"[？1，SaDMS[？1，从而导致模型复杂性增加和信息",
"ploys the Convolutional vision Transformer (CvT) [?",
"丢失的可能性。端到端自动驾驶的最新进展已经探索",
"tobetter extract and fusefeatures frommultipleRGB",
"了不同传感器模式的集成，例如LiDAR和RGB-D相",
"cameras due to local and global feature extraction ca-",
"机，以提高性能。？1employs a卷积神经网络（CNN）",
"pability of convolution and transformer modules, re-",
"提取anRGB-D相机和produce未来车辆航路点和",
"spectively.",
"导航信号提供的数据特征。另一种众所周知的方法是",
"TransFuser[？］，它使用多模式融合转换器将全局上",
"Encoded features combined with static and dynamic",
"下文和成对交互合并到不同输人模式的特征提取层中。",
"environments are later employed by our control mod-",
"We were inspired by these methods and demonstrated",
"ule to predict waypoints and vehicular controls (e.g.",
"that combining ideas from these approaches could po-",
"steering, throttle, and brake).",
"We use two meth-",
"tentiallylead to more robust and accurate end-to-end",
"ods to generate the vehicular controls levels. The first",
"autonomous driving solutions. On the perception side,",
"we leverage the strengths of both global and local con-",
"method uses a PID algorithm to follow the waypoints",
"on the fly, whereas the second one directly predicts the",
"text reasoning provided by transformers and CNNs. In",
"the control side, we take advantage of trajectory-guided",
"control policy using the measurementfeatures and en-",
"vironmental state. We evaluate the model and con-",
"control, which we introduce next.",
"duct a comparative analysis with recent models on",
"The Trajectory-guided Control Prediction (TCP) [?",
"the CARLA simulator using various scenarios, rang-",
"I framework, which combines trajectory planning and",
"ing from normal to adversarial conditions, to simulate",
"direct control is a multi-task learning framework for",
"real-world scenarios.Our method demonstrated bet-",
"end-to-end autonomous driving. By incorporating a",
"ter or comparable results with respect to our baselines",
"multi-step control prediction branch with a dynamic",
"in term of driving abilities.:The code is available at",
"branch and trajectory-guided attention, TCP can im-",
"https://github.com/pagand/e2etransfuser/tree/cvpr-",
"prove temporal reasoning and achieve superior perfor-",
"w to facilitate future studies.",
"mance in the CARLA driving simulator, even surpass-",
"ing methods using multiple cameras and LiDAR sen-"
] | {
"arxiv_id": "2310.13135",
"language": "zh",
"page_idx": 0
} |
||
2310.13135-zh-1 | 2310.13135-zh-1.jpg | sors. 在本文中，inspired by these three methods, 我们提出了一种新颖的 deep neural network architecture for end-to-end autonomous driving ,它利用 RGB-D 相机提供的 RGB 和深度信息的互补优势，解决了现有单模态和传感器融合方法 as well as navigational commands prediction 面临的挑战。我们的模型由两个主要模块 shown in Fig. 1 组成：感知模块，对高维观测数据进行编码并执行语义分割，语义深度云(SDC)映射 and ego vehicle speed and traffc light prediction ; 以及控制模块，对感知模块编码的特征以及其他 GPS, 命令和速度表信息进行解码，以预测航路点和控制策略。在感知模块中，我们利用卷积视觉变压器 (CvT) [?]和 EfficientNet [?] to adeptly extract RGB image and SDC map features. We then fuse them using a CNN-based fusion layer 。此外，我们在控制模块中使用两个代理来处理感知模块的输出，从而促进多样化和有弹性的决策。为了解决平衡学习信号 similar to ? ]的问题，我们实现了一种改进的梯度归一化 (MGN) 方法，确保所有任务的学习速度一致。最后，我们在CARLA 模拟器上评估了我们的模型，包括正常对抗情况，reported in Section 4.6 展示了比基线方法更好的性能。We achieve better and comparable results with respect to our baselines in short and long paths, respectively.
## 2. Related works
$x$。Recently,a few 研究 by?]和？]研究了将深度和语义数据作为驾驶任务的中间表示的有效性。在我们的工作中，我们专注于组合 RGB 和 Depth 输人，which are readily available in autonomous vehicles and provide complementary 场景表示。
Sensor Fusion : 大多数传感器融合研究都集中在感知任务上，例如对象检测[???]和运动预测[????] 。这些方法通常是 include 多视角激光雷达 data ,或者通过在不同空间之间投影特征来将摄像机输人与激光雷达数据结合起来。ContFuse[?] 是一种密集融合多尺度 RGB 和激光雷达鸟瞰图 (BEV)特征的方法。但是，这些方法无法捕获 3D 场景的全局上下文，这对于在具有挑战性的场景中进行安全导航至关重要。In this work, we use CNNs to fuse the multi-modal data received by the RGB-D sensors.
Bird's Eye View Strategies:在这个领域，研究人员 either use LiDAR data or 融合了来自单个 RGB-D 相机[?]的 RGB 图像和深度图。他们投影深度图with the semantic segmentation to 创建语义深度云(SDC) from a BEV angle 。与仅包含高度数据[??] 的 LiDAR 点云相比，由 SDC 的语义信息提供的 This way the model benefits from the clearer delineation of 占用或 navigable 区域。? ] 融合了 RGB 图像和深度图以捕获更深的全局上下文，而[?] 结合了RGB 图像和预处理的 LiDAR 点云，以利用不同的视角，例如前视图和 BEV。这些方法使用高级导航命令as described by? ] 或全局规划器 as explored by ?
] 提供的稀疏 GPS 位置进行驾驶。在我们的研究中我们考虑使用 waypoints 而不是高级导航命令，因为waypoints are more informative and 更好地反映了现实世界的自动驾驶条件[?]。
Imitation Learning : 端到端自动驾驶的研究通常分为两类：强化学习(RL)和模仿学习(IL)。??] 显示了RL 的潜力，而 IL 方法如 LBC [?] 和 NEAT [?]则表现出令人印象深刻的性能。我们的工作适应了TransFuser 及其变体[???] 中使用的自动回归方案。
End-to-End Autonomous Driving :端到端多任务学习方法在培训效率和集成简单性方面提供了优势。已针对自动驾驶任务研究了基于模仿学习的方法，by 探索了其他感知任务以改善特征提取[?]。各种自动驾驶 subtasks 的组合，例如对象检测，车道检测，语义分割和深度估计，已被证明可实现令人难以置信的性能[??]。在我们的工作中，我们采用类似的多任务学习方法，但利用 RGB-D 相机的深度作为输入[?]。我们通过实现 MGN 算法[?]来解决多任务学习中的不平衡学习问题。
# 3. Methodology
在本节中，我们将介绍我们提出的方法的详细信息。我们的模型围绕两个关键组件构建：感知和控制模块， 每个模块都完成一组特定的任务，以实现车辆导航。In the perception module , we extract features from RGB and a BEV SDC map.此外，模型 attempts to 从RGB 嵌入式功能准确预测交通灯和自我车速。在随后的控制 module 中，我们的模型利用提取的特征，当前自我车辆速度和 GPS 位置来提供可靠的航路点和车辆命令。
## $3. 1. \textbf{Perception Module}$
As illustrated in Fig. $\color{red}{1,\text{the perception}}$ module serves to extract features from RGB images using CvT [?] to perform semantic segmentation and later generate a BEV SDC map 。我们的感知模块从三个车载摄像头接收总共三个 RGB 图像，第一个摄像头捕获前视角，另外两个向左和向右倾斜 60 度。为了提供对环境的全面了解，还从每个摄像机捕获深度图像。前 RGB 和深度图像具有 resolution 或$160\times320$, 而非重叠侧摄像机使用 resolution of 160 × 224. This results in a total of $160\times768$像素捕获 RGB 和深度图像的图像。
3.1.1 Convolutional Vision Transformer
$(x_g,y_g)$已在 ImageNet [?] shown in bottom left section of the Fig. $\color{red}{1}$上预训练的[?]。This module is responsible for extracting features from the three RGB images. 这个特殊的 network 被选择为其独特的能力，利用卷积和变压器模块，greatly facilitating feature extraction 。卷积 layers 擅长提取局部特征，而变压器以其 ability 在全局特征提取和学习方面而闻名。通 | [
[
"0.0809",
"0.09216",
"0.468",
"0.09216",
"0.468",
"0.1061",
"0.0809",
"0.1061"
],
[
"0.501",
"0.09155",
"0.888",
"0.09216",
"0.888",
"0.1067",
"0.501",
"0.1061"
],
[
"0.0825",
"0.108",
"0.4673",
"0.108",
"0.4673",
"0.1218",
"0.0825",
"0.1218"
],
[
"0.505",
"0.108",
"0.8906",
"0.108",
"0.8906",
"0.1218",
"0.505",
"0.1218"
],
[
"0.0801",
"0.1225",
"0.468",
"0.1218",
"0.468",
"0.1364",
"0.0801",
"0.137"
],
[
"0.5034",
"0.1237",
"0.892",
"0.1225",
"0.892",
"0.1364",
"0.5034",
"0.137"
],
[
"0.0801",
"0.137",
"0.4673",
"0.1383",
"0.4673",
"0.1528",
"0.0801",
"0.1515"
],
[
"0.5034",
"0.1395",
"0.7075",
"0.1395",
"0.7075",
"0.1528",
"0.5034",
"0.1528"
],
[
"0.08417",
"0.1547",
"0.4656",
"0.1547",
"0.4656",
"0.1654",
"0.08417",
"0.1654"
],
[
"0.523",
"0.154",
"0.8906",
"0.154",
"0.8906",
"0.1674",
"0.523",
"0.1674"
],
[
"0.0825",
"0.1686",
"0.468",
"0.1686",
"0.468",
"0.1818",
"0.0825",
"0.1818"
],
[
"0.505",
"0.1686",
"0.8857",
"0.1686",
"0.8857",
"0.1818",
"0.505",
"0.1818"
],
[
"0.08496",
"0.1849",
"0.4656",
"0.1849",
"0.4656",
"0.1957",
"0.08496",
"0.1957"
],
[
"0.505",
"0.1843",
"0.8896",
"0.1843",
"0.8896",
"0.1976",
"0.505",
"0.1976"
],
[
"0.0825",
"0.1976",
"0.4697",
"0.1976",
"0.4697",
"0.2108",
"0.0825",
"0.2108"
],
[
"0.505",
"0.2008",
"0.886",
"0.2008",
"0.886",
"0.2108",
"0.505",
"0.2108"
],
[
"0.0809",
"0.2134",
"0.468",
"0.2134",
"0.468",
"0.2267",
"0.0809",
"0.2267"
],
[
"0.5024",
"0.2108",
"0.886",
"0.2134",
"0.886",
"0.2279",
"0.5024",
"0.226"
],
[
"0.0825",
"0.2291",
"0.468",
"0.2291",
"0.468",
"0.2424",
"0.0825",
"0.2424"
],
[
"0.5215",
"0.2279",
"0.8896",
"0.2291",
"0.8896",
"0.2437",
"0.5215",
"0.2424"
],
[
"0.08417",
"0.2456",
"0.4673",
"0.2456",
"0.4673",
"0.2556",
"0.08417",
"0.2556"
],
[
"0.5034",
"0.2424",
"0.8896",
"0.2443",
"0.8896",
"0.2583",
"0.5034",
"0.2556"
],
[
"0.0809",
"0.2588",
"0.4697",
"0.2588",
"0.4697",
"0.2727",
"0.0809",
"0.2727"
],
[
"0.505",
"0.2588",
"0.8906",
"0.2588",
"0.8906",
"0.2727",
"0.505",
"0.2727"
],
[
"0.0801",
"0.2727",
"0.4673",
"0.2734",
"0.4673",
"0.2886",
"0.0801",
"0.2874"
],
[
"0.5034",
"0.2734",
"0.8906",
"0.2727",
"0.8906",
"0.2874",
"0.5034",
"0.2886"
],
[
"0.0809",
"0.289",
"0.468",
"0.289",
"0.468",
"0.303",
"0.0809",
"0.303"
],
[
"0.5034",
"0.2886",
"0.8906",
"0.289",
"0.8906",
"0.303",
"0.5034",
"0.3018"
],
[
"0.0825",
"0.3037",
"0.468",
"0.3037",
"0.468",
"0.3176",
"0.0825",
"0.3176"
],
[
"0.505",
"0.3037",
"0.8906",
"0.3037",
"0.8906",
"0.3176",
"0.505",
"0.3176"
],
[
"0.0825",
"0.318",
"0.4673",
"0.3193",
"0.4673",
"0.3333",
"0.0825",
"0.332"
],
[
"0.5034",
"0.3193",
"0.8896",
"0.3193",
"0.8896",
"0.3333",
"0.5034",
"0.3333"
],
[
"0.0801",
"0.334",
"0.468",
"0.3333",
"0.468",
"0.348",
"0.0801",
"0.3484"
],
[
"0.5034",
"0.3333",
"0.8774",
"0.334",
"0.8774",
"0.3484",
"0.5034",
"0.348"
],
[
"0.0825",
"0.351",
"0.4673",
"0.351",
"0.4673",
"0.361",
"0.0825",
"0.361"
],
[
"0.505",
"0.3499",
"0.8906",
"0.3499",
"0.8906",
"0.363",
"0.505",
"0.363"
],
[
"0.0809",
"0.3643",
"0.468",
"0.3643",
"0.468",
"0.3774",
"0.0809",
"0.3774"
],
[
"0.5024",
"0.3655",
"0.626",
"0.3655",
"0.626",
"0.379",
"0.5024",
"0.379"
],
[
"0.0801",
"0.3774",
"0.468",
"0.379",
"0.468",
"0.3933",
"0.0801",
"0.3926"
],
[
"0.0809",
"0.3945",
"0.468",
"0.3945",
"0.468",
"0.4077",
"0.0809",
"0.4077"
],
[
"0.501",
"0.3926",
"0.6406",
"0.3945",
"0.6406",
"0.4092",
"0.501",
"0.4072"
],
[
"0.0809",
"0.4104",
"0.468",
"0.4104",
"0.468",
"0.4236",
"0.0809",
"0.4236"
],
[
"0.0801",
"0.4236",
"0.468",
"0.4248",
"0.468",
"0.4395",
"0.0801",
"0.4382"
],
[
"0.523",
"0.4192",
"0.8906",
"0.4192",
"0.8906",
"0.433",
"0.523",
"0.433"
],
[
"0.505",
"0.4343",
"0.8906",
"0.4343",
"0.8906",
"0.4475",
"0.505",
"0.4475"
],
[
"0.0809",
"0.4414",
"0.1666",
"0.4414",
"0.1666",
"0.4526",
"0.0809",
"0.4526"
],
[
"0.505",
"0.4495",
"0.8896",
"0.4495",
"0.8896",
"0.4634",
"0.505",
"0.4634"
],
[
"0.0801",
"0.4697",
"0.2303",
"0.4697",
"0.2303",
"0.483",
"0.0801",
"0.483"
],
[
"0.5024",
"0.464",
"0.8906",
"0.4634",
"0.8906",
"0.4778",
"0.5024",
"0.4792"
],
[
"0.505",
"0.4797",
"0.8906",
"0.4797",
"0.8906",
"0.4937",
"0.505",
"0.4937"
],
[
"0.0989",
"0.4944",
"0.4673",
"0.4944",
"0.4673",
"0.5083",
"0.0989",
"0.5083"
],
[
"0.5024",
"0.4937",
"0.8906",
"0.4944",
"0.8906",
"0.509",
"0.5024",
"0.5083"
],
[
"0.0809",
"0.5103",
"0.468",
"0.5103",
"0.468",
"0.5234",
"0.0809",
"0.5234"
],
[
"0.5034",
"0.5103",
"0.8906",
"0.5103",
"0.8906",
"0.5234",
"0.5034",
"0.5234"
],
[
"0.0825",
"0.5244",
"0.4673",
"0.5244",
"0.4673",
"0.5386",
"0.0825",
"0.5386"
],
[
"0.505",
"0.526",
"0.8896",
"0.526",
"0.8896",
"0.539",
"0.505",
"0.539"
],
[
"0.08417",
"0.5415",
"0.4673",
"0.5415",
"0.4673",
"0.552",
"0.08417",
"0.552"
],
[
"0.505",
"0.5425",
"0.5605",
"0.5425",
"0.5605",
"0.5537",
"0.505",
"0.5537"
],
[
"0.0809",
"0.5547",
"0.361",
"0.5547",
"0.361",
"0.5684",
"0.0809",
"0.5684"
],
[
"0.5034",
"0.565",
"0.687",
"0.565",
"0.687",
"0.58",
"0.5034",
"0.58"
],
[
"0.1013",
"0.5713",
"0.4673",
"0.5713",
"0.4673",
"0.582",
"0.1013",
"0.582"
],
[
"0.0809",
"0.5854",
"0.4656",
"0.5854",
"0.4656",
"0.5986",
"0.0809",
"0.5986"
],
[
"0.5195",
"0.586",
"0.8906",
"0.587",
"0.8906",
"0.6016",
"0.5195",
"0.601"
],
[
"0.08496",
"0.5996",
"0.468",
"0.5986",
"0.468",
"0.613",
"0.08496",
"0.6143"
],
[
"0.5034",
"0.604",
"0.8906",
"0.604",
"0.8906",
"0.617",
"0.5034",
"0.617"
],
[
"0.08417",
"0.616",
"0.4673",
"0.616",
"0.4673",
"0.627",
"0.08417",
"0.627"
],
[
"0.5024",
"0.6187",
"0.8823",
"0.6187",
"0.8823",
"0.632",
"0.5024",
"0.632"
],
[
"0.0825",
"0.629",
"0.468",
"0.629",
"0.468",
"0.643",
"0.0825",
"0.643"
],
[
"0.505",
"0.634",
"0.8896",
"0.634",
"0.8896",
"0.648",
"0.505",
"0.648"
],
[
"0.08417",
"0.6465",
"0.468",
"0.6465",
"0.468",
"0.657",
"0.08417",
"0.657"
],
[
"0.505",
"0.6484",
"0.8906",
"0.6484",
"0.8906",
"0.662",
"0.505",
"0.662"
],
[
"0.0809",
"0.66",
"0.4673",
"0.66",
"0.4673",
"0.674",
"0.0809",
"0.674"
],
[
"0.505",
"0.6636",
"0.8906",
"0.6636",
"0.8906",
"0.678",
"0.505",
"0.678"
],
[
"0.08417",
"0.6753",
"0.468",
"0.6753",
"0.468",
"0.6895",
"0.08417",
"0.6895"
],
[
"0.5034",
"0.678",
"0.8906",
"0.6797",
"0.8906",
"0.6934",
"0.5034",
"0.6914"
],
[
"0.0809",
"0.69",
"0.4673",
"0.69",
"0.4673",
"0.704",
"0.0809",
"0.704"
],
[
"0.505",
"0.6943",
"0.8774",
"0.6943",
"0.8774",
"0.7085",
"0.505",
"0.7085"
],
[
"0.0801",
"0.7056",
"0.3096",
"0.7056",
"0.3096",
"0.7197",
"0.0801",
"0.7197"
],
[
"0.505",
"0.711",
"0.8896",
"0.711",
"0.8896",
"0.7217",
"0.505",
"0.7217"
],
[
"0.10046",
"0.72",
"0.4673",
"0.72",
"0.4673",
"0.7344",
"0.10046",
"0.7344"
],
[
"0.505",
"0.7246",
"0.8906",
"0.7246",
"0.8906",
"0.738",
"0.505",
"0.738"
],
[
"0.0809",
"0.7363",
"0.468",
"0.7363",
"0.468",
"0.75",
"0.0809",
"0.75"
],
[
"0.505",
"0.7407",
"0.5923",
"0.7407",
"0.5923",
"0.752",
"0.505",
"0.752"
],
[
"0.0825",
"0.7505",
"0.4656",
"0.7505",
"0.4656",
"0.7646",
"0.0825",
"0.7646"
],
[
"0.08496",
"0.768",
"0.4656",
"0.768",
"0.4656",
"0.779",
"0.08496",
"0.779"
],
[
"0.5034",
"0.773",
"0.7983",
"0.773",
"0.7983",
"0.786",
"0.5034",
"0.786"
],
[
"0.0825",
"0.782",
"0.4673",
"0.782",
"0.4673",
"0.7954",
"0.0825",
"0.7954"
],
[
"0.0809",
"0.797",
"0.468",
"0.797",
"0.468",
"0.811",
"0.0809",
"0.811"
],
[
"0.5024",
"0.7954",
"0.8896",
"0.7925",
"0.8896",
"0.81",
"0.5024",
"0.813"
],
[
"0.0809",
"0.8125",
"0.4673",
"0.811",
"0.4673",
"0.8228",
"0.0809",
"0.8237"
],
[
"0.5024",
"0.811",
"0.8896",
"0.811",
"0.8896",
"0.8247",
"0.5024",
"0.8247"
],
[
"0.08826",
"0.827",
"0.468",
"0.827",
"0.468",
"0.8403",
"0.08826",
"0.8403"
],
[
"0.5034",
"0.827",
"0.8906",
"0.827",
"0.8906",
"0.8403",
"0.5034",
"0.8403"
],
[
"0.0825",
"0.8413",
"0.4673",
"0.8413",
"0.4673",
"0.855",
"0.0825",
"0.855"
],
[
"0.5063",
"0.8433",
"0.5645",
"0.8433",
"0.5645",
"0.854",
"0.5063",
"0.854"
],
[
"0.5596",
"0.8413",
"0.888",
"0.8413",
"0.888",
"0.855",
"0.5596",
"0.855"
],
[
"0.0809",
"0.857",
"0.4673",
"0.857",
"0.4673",
"0.8706",
"0.0809",
"0.8706"
],
[
"0.505",
"0.857",
"0.884",
"0.857",
"0.884",
"0.8706",
"0.505",
"0.8706"
],
[
"0.0809",
"0.872",
"0.468",
"0.872",
"0.468",
"0.886",
"0.0809",
"0.886"
],
[
"0.5034",
"0.8706",
"0.8906",
"0.872",
"0.8906",
"0.886",
"0.5034",
"0.8853"
],
[
"0.0809",
"0.887",
"0.4697",
"0.887",
"0.4697",
"0.901",
"0.0809",
"0.901"
],
[
"0.505",
"0.887",
"0.8906",
"0.887",
"0.8906",
"0.901",
"0.505",
"0.901"
]
] | [
"sors.在本文中，inspiredbythesethreemethods，我",
"1提供的稀疏GPS位置进行驾驶。在我们的研究中",
"们提出了一种新颖的deepneuralnetworkarchitecture",
"我们考虑使用waypoints而不是高级导航命令，因为",
"forend-to-end autonomous driving它利用RGB-D",
"waypointsaremoreinformativeand更好地反映了现",
"相机提供的RGB和深度信息的互补优势，解决了现有",
"实世界的自动驾驶条件[？",
"单模态和传感器融合方法aswellasnavigationalcom-",
"ImitationLearning：端到端自动驾驶的研究通常",
"mandsprediction面临的挑战。我们的模型由两个主",
"分为两类：强化学习（RL）和模仿学习（IL）。？？1",
"要模块showminFig.1组成：感知模块，对高维观测",
"显示了RL的潜力，而IL方法如LBC[？］和NEAT",
"数据进行编码并执行语义分割，语义深度云（SDC）映",
"？则表现出今人印象深刻的性能。我们的工作适应了",
"射 and ego vehicle speed and traffic light prediction:",
"TransFuser及其变体[？？？］中使用的自动回归方案，",
"以及控制模块，对感知模块编码的特征以及其他GPS，",
"End-to-EndAutonomousDriving：端到端多任务",
"命今和速度表信息进行解码，以预测航路点和控制策",
"学习方法在培训效率和集成简单性方面提供了优势。已",
"略。在感知模块中，我们利用卷积视觉变压器（CvT）",
"针对自动驾驶任务研究了基于模仿学习的方法，by探",
"[?]和EfficientNet [?] to adeptly extractRGB image",
"索了其他感知任务以改善特征提取[？]。各种自动驾",
"and SDC map features. We then fuse them using a",
"驶subtasks的组合，例如对象检测，车道检测，语义",
"CNN-based fusionlayer。此外，我们在控制模块中使",
"分割和深度估计，已被证明可实现令人难以置信的性",
"用两个代理来处理感知模块的输出，从而促进多样化",
"能[？？］。在我们的工作中，我们采用类似的多任务",
"和有弹性的决策。为了解决平衡学习信号similarto？",
"学习方法，但利用RGB-D相机的深度作为输入[？]",
"厂的问题，我们实现了一种改进的梯度归一化（MGN）",
"我们通过实现MGN算法[？1来解决多任务学习中的",
"方法，确保所有任务的学习速度一致。最后，我们在",
"不平衡学习问题。",
"CARLA模拟器上评估了我们的模型，包括正常对抗",
"情况，reportedinSection4.6展示了比基线方法更",
"3.Methodology",
"好的性能。We achievebetterand comparableresults",
"with respect to our baselines in short and long paths,",
"在本节中，我们将介绍我们提出的方法的详细信息。",
"我们的模型围绕两个关键组件构建：感知和控制模块，",
"respectively",
"每个模块都完成一组特定的任务，以实现车辆导航。In",
"2.Related works",
"the perception module, we extract features from RGB",
"andaBEVSDCmap.此外，模型attemptsto从",
"。Recently,a few研究by？］和？］研究了",
"RGB嵌人式功能准确预测交通灯和自我车速。在随后",
"将深度和语义数据作为驾驶任务的中间表示的有效性。",
"的控制module中，我们的模型利用提取的特征，当",
"在我们的工作中，我们专注于组合RGB和Depth输",
"前自我车辆速度和GPS位置来提供可靠的航路点和车",
"K:whicharereadilyavailableinautonomousvehicles",
"辆命令。",
"andprovidecomplementary场景表示。",
"3.1. Perception Module",
"SensorFusion：大多数传感器融合研究都集中在感",
"知任务上，例如对象检测[？？？1和运动预测[？？？",
"As illustrated in Fig.1,the perception module",
"。这些方法通常是include多视角激光雷达data，或",
"serves to extract features from RGB images using CvT",
"者通过在不同空间之间投影特征来将摄像机输入与激",
"[? 1 to perform semantic segmentation and later ger",
"光雷达数据结合起来。ContFuse[？是一种密集融合",
"erateaBEVSDCmap。我们的感知模块从三个车",
"多尺度RGB和激光雷达鸟瞰图（BEV）特征的方法。",
"载摄像头接收总共三个RGB图像，第一个摄像头捕",
"但是，这些方法无法捕获3D场景的全局上下文，这对",
"获前视角，另外两个向左和向右倾斜60度。为了提",
"于在具有挑战性的场景中进行安全导航至关重要。In",
"供对环境的全面了解，还从每个摄像机捕获深度图像。",
"this work, we use CNNs to fuse the multi-modal data",
"前RGB和深度图像具有resolution或160×320",
"received by the RGB-D sensors.",
"而非重叠侧摄像机使用resolutionof160×224.This",
"Bird'sEyeViewStrategies：在这个领域，研究人",
"resultsinatotalof160×768像素捕获RGB和深度",
"员eitheruseLiDARdata or 融合了来自单个RGB-D",
"图像的图像。",
"相机[？］的RGB图像和深度图。他们投影深度图",
"withthesemanticsegmentationto：创建语义深度云",
"3.1.1Convolutional Vision Transformei",
"(SDC)fromaBEVangle：与仅包含高度数据[？？]",
"的LiDAR点云相比，由SDC的语义信息提供的This",
"(ag, Yg) 已在 ImageNet [? ] shown in bottom left sec",
"way the model benefits from the clearer delineation of",
"tion of the Fig.1 上预训练的[?」。This module is",
"占用或navigable区域。？］融合了RGB图像和",
"responsible for extracting features from the three RGB",
"深度图以捕获更深的全局上下文，而[？1结合了",
"images.",
"这个特殊的network被选择为其独特的能",
"RGB图像和预处理的LiDAR点云，以利用不同的视",
"力，利用卷积和变压器模块，greatlyfacilitatingfeatur",
"角，例如前视图和BEV。这些方法使用高级导航命令",
"extraction。卷积layers擅长提取局部特征，而变压",
"asdescribedby？1或全局规划器asexploredby？",
"器以其ability在全局特征提取和学习方面而闻名。通"
] | {
"arxiv_id": "2310.13135",
"language": "zh",
"page_idx": 1
} |
||
2310.13135-zh-2 | 2310.13135-zh-2.jpg | <FigureHere>
Figure 1. Model architecture: It consists of trainable and non-trainable components represented by light and dark-colored items, respectively. The blue-colored items represent the perception module , while the controller module and inputs are represented by green and orange , respectively. The white boxes represent different tasks that are learned simultaneously .The process inside the top dashed green box is iterated three times during the training, in which the model predicts waypoints to estimate the values of vehicular controls independently.
3.1.3 Semantic Depth Cloud
过结合这两种强大技术的优势，CvT 确保我们的特征提取器能够捕获局部和全局特征，从而产生 better 视觉表示。
CvT-13[?] 是 CvT 的轻型版本，由于其出色的性能和较少的参数数量而经过精心选择。正如人们可以看到在图。$\operatorname{El}$ ,它被设计为三个主要阶段，每个阶段都包含一个卷积令牌嵌人(CTE)来处理 2D 输入images 。由 CTE 提取的特征然后被归一化并通过卷积变换器块 (CTB)。To apply both convolutional and attention layers, CTB 使用称为卷积投影的深度可分离卷积运算来创建查询，键和值嵌入。然后将这些嵌人通过转换器模块来提取全局特征，最终得到高度准确和全面的特征图。CvT 的最后一层是用于图像分类的完全连接层，我们将其从模型中删除。结果，RGB 提取的特征包含 384 个特征，每个特征的大小为 10×48
为了增强对场景的理解，我们使用了一种方法，该方法涉及使用提供的深度图像创建 BEV Semantic Depth Cloud 除了 RGBMAX2。The SDC map represents the ego-vehicle's surrounding environment and contains richer information with respect to the LiDAR data due to the 23 semantic class layers. Each layer represents one class of environment object. As shown in 图。2 , 创建 SDC 时不考虑高度信息 ( $y$ -axis) ,因为 BEV 透视图只需要$x$ -axis and $z$ -axis 。创建SDC 的过程涉及为每个摄像机分别生成 SDC 贴图，然后在以适当的角度旋转侧面贴图后将它们合并在一起。我们在车辆前方定义了 64 米的距离范围，在摄像机的每一侧定义了 32 米，创建了 64×64 平方米的覆盖范围。SDC 地图的正面尺寸为$160\times320$平方厘米，侧面尺寸为 160$\times224$ 。我们使用相机参数为$x$ -axis 生成一个变换矩阵，并将坐标归一化以与 SDC 张量空间维度对齐。应用独热编码产生 23 通道 SDC 张量。生成的映射被复制到一个空的张量，其尺寸为 160×768侧图以 42 度角旋转。对于特征提取，我们使用紧凑的EfficientNet- B1 [ ? ] 网络生成 192 个特征的张量，每个特征的大小为 10$\times 48$ .
3.1.2 Semantic Segmentation
## 3.2. Controller Module
提取特征图后，we use them in different sections of the model, as they contain valuable information. First, 它们被用来训练能够准确识别 23 个不同类的语义分割解码器 depicted in the Fig. 1 as“Decoder”. Later, we use this to create a BEV SDC map 。为了实现这一点，我们开发了一个分段解码器，它由三个卷积层和一个具有 S 形激活的最终逐点卷积组成。通过利用跳过连接，我们可以有效地捕获局部和全局特征，从而生成准确而全面的分割图。
控制模块，在图中显示为绿色正方形。1,接收 RGB, SDC and navigational measurements 特征，并使用它们来预测车辆的未来航路点。The navigational measurements includes route location, navigational command provided by the global planner and the ego ve- | [
[
"0.10297",
"0.0947",
"0.1259",
"0.0947",
"0.1259",
"0.1016",
"0.10297",
"0.1016"
],
[
"0.3145",
"0.09534",
"0.353",
"0.09534",
"0.353",
"0.1111",
"0.3145",
"0.1111"
],
[
"0.452",
"0.0966",
"0.492",
"0.0966",
"0.492",
"0.1067",
"0.452",
"0.1067"
],
[
"0.55",
"0.09534",
"0.5786",
"0.09534",
"0.5786",
"0.1067",
"0.55",
"0.1067"
],
[
"0.722",
"0.0966",
"0.773",
"0.0966",
"0.773",
"0.108",
"0.722",
"0.108"
],
[
"0.8364",
"0.0966",
"0.871",
"0.0966",
"0.871",
"0.1061",
"0.8364",
"0.1061"
],
[
"0.6055",
"0.09784",
"0.6436",
"0.09784",
"0.6436",
"0.1048",
"0.6055",
"0.1048"
],
[
"0.5034",
"0.09784",
"0.5337",
"0.09784",
"0.5337",
"0.1061",
"0.5034",
"0.1061"
],
[
"0.0907",
"0.108",
"0.1708",
"0.1042",
"0.1724",
"0.1225",
"0.09235",
"0.127"
],
[
"0.165",
"0.10986",
"0.1805",
"0.10986",
"0.1805",
"0.11743",
"0.165",
"0.11743"
],
[
"0.8364",
"0.1042",
"0.8677",
"0.1042",
"0.8677",
"0.11237",
"0.8364",
"0.11237"
],
[
"0.558",
"0.11237",
"0.5923",
"0.11554",
"0.591",
"0.1225",
"0.556",
"0.1193"
],
[
"0.6772",
"0.11554",
"0.7236",
"0.11615",
"0.7236",
"0.1237",
"0.6772",
"0.1225"
],
[
"0.839",
"0.1111",
"0.8667",
"0.11426",
"0.864",
"0.1256",
"0.8384",
"0.1225"
],
[
"0.621",
"0.1218",
"0.648",
"0.1218",
"0.648",
"0.1288",
"0.621",
"0.1288"
],
[
"0.5537",
"0.1237",
"0.581",
"0.1237",
"0.581",
"0.1307",
"0.5537",
"0.1307"
],
[
"0.6724",
"0.125",
"0.7305",
"0.125",
"0.7305",
"0.1364",
"0.6724",
"0.1364"
],
[
"0.5034",
"0.1256",
"0.5215",
"0.1256",
"0.5215",
"0.13",
"0.5034",
"0.13"
],
[
"0.5522",
"0.11615",
"0.5596",
"0.11615",
"0.5596",
"0.1193",
"0.5522",
"0.1193"
],
[
"0.3774",
"0.1307",
"0.3823",
"0.1351",
"0.3782",
"0.1383",
"0.3733",
"0.1338"
],
[
"0.2632",
"0.137",
"0.33",
"0.1395",
"0.33",
"0.1509",
"0.2632",
"0.1483"
],
[
"0.5024",
"0.1364",
"0.523",
"0.1364",
"0.523",
"0.1445",
"0.5024",
"0.1445"
],
[
"0.7344",
"0.1395",
"0.761",
"0.1364",
"0.7646",
"0.1465",
"0.738",
"0.1497"
],
[
"0.8325",
"0.1401",
"0.88",
"0.1401",
"0.88",
"0.1547",
"0.8325",
"0.1547"
],
[
"0.4143",
"0.1445",
"0.46",
"0.1445",
"0.46",
"0.1547",
"0.4143",
"0.1547"
],
[
"0.10706",
"0.1433",
"0.125",
"0.1433",
"0.125",
"0.1477",
"0.10706",
"0.1477"
],
[
"0.5073",
"0.1528",
"0.5146",
"0.1477",
"0.5195",
"0.1515",
"0.513",
"0.1572"
],
[
"0.7466",
"0.1497",
"0.761",
"0.1497",
"0.761",
"0.1572",
"0.7466",
"0.1572"
],
[
"0.0825",
"0.1559",
"0.2035",
"0.1559",
"0.2035",
"0.1692",
"0.0825",
"0.1692"
],
[
"0.3489",
"0.1622",
"0.3945",
"0.1642",
"0.3938",
"0.1768",
"0.3472",
"0.1749"
],
[
"0.4182",
"0.1654",
"0.437",
"0.1654",
"0.437",
"0.1686",
"0.4182",
"0.1686"
],
[
"0.5312",
"0.161",
"0.5557",
"0.161",
"0.5557",
"0.1692",
"0.5312",
"0.1692"
],
[
"0.1396",
"0.1692",
"0.1814",
"0.1724",
"0.1805",
"0.1843",
"0.1381",
"0.1812"
],
[
"0.5435",
"0.1736",
"0.5522",
"0.1704",
"0.5557",
"0.1755",
"0.546",
"0.1787"
],
[
"0.6313",
"0.1718",
"0.6543",
"0.1718",
"0.6543",
"0.1831",
"0.6313",
"0.1831"
],
[
"0.165",
"0.1812",
"0.183",
"0.1812",
"0.183",
"0.1881",
"0.165",
"0.1881"
],
[
"0.819",
"0.1831",
"0.8364",
"0.1831",
"0.8364",
"0.1875",
"0.819",
"0.1875"
],
[
"0.819",
"0.2197",
"0.834",
"0.2197",
"0.834",
"0.2235",
"0.819",
"0.2235"
],
[
"0.1396",
"0.1849",
"0.1478",
"0.1831",
"0.1487",
"0.1875",
"0.1405",
"0.1881"
],
[
"0.1438",
"0.2267",
"0.4028",
"0.2267",
"0.4028",
"0.2974",
"0.1438",
"0.2974"
],
[
"0.2942",
"0.2507",
"0.3145",
"0.2507",
"0.3145",
"0.254",
"0.2942",
"0.254"
],
[
"0.596",
"0.254",
"0.6436",
"0.254",
"0.6436",
"0.2651",
"0.596",
"0.2651"
],
[
"0.675",
"0.254",
"0.709",
"0.254",
"0.709",
"0.2651",
"0.675",
"0.2651"
],
[
"0.2754",
"0.2695",
"0.2852",
"0.2695",
"0.2852",
"0.2727",
"0.2754",
"0.2727"
],
[
"0.613",
"0.2708",
"0.6235",
"0.2708",
"0.6235",
"0.2827",
"0.613",
"0.2827"
],
[
"0.687",
"0.2695",
"0.697",
"0.2695",
"0.697",
"0.2827",
"0.687",
"0.2827"
],
[
"0.2793",
"0.2727",
"0.3057",
"0.2727",
"0.3057",
"0.2798",
"0.2793",
"0.2798"
],
[
"0.2378",
"0.2747",
"0.2607",
"0.2747",
"0.2607",
"0.279",
"0.2378",
"0.279"
],
[
"0.591",
"0.2874",
"0.6465",
"0.2874",
"0.6465",
"0.2974",
"0.591",
"0.2974"
],
[
"0.6724",
"0.2886",
"0.7017",
"0.2886",
"0.7017",
"0.2966",
"0.6724",
"0.2966"
],
[
"0.6772",
"0.2974",
"0.7017",
"0.2974",
"0.7017",
"0.3062",
"0.6772",
"0.3062"
],
[
"0.7427",
"0.2917",
"0.749",
"0.2917",
"0.749",
"0.2954",
"0.7427",
"0.2954"
],
[
"0.0948",
"0.3037",
"0.1209",
"0.3037",
"0.1209",
"0.3125",
"0.0948",
"0.3125"
],
[
"0.6055",
"0.2986",
"0.63",
"0.2986",
"0.63",
"0.307",
"0.6055",
"0.307"
],
[
"0.2615",
"0.307",
"0.312",
"0.307",
"0.312",
"0.3176",
"0.2615",
"0.3176"
],
[
"0.361",
"0.31",
"0.4126",
"0.31",
"0.4126",
"0.3213",
"0.361",
"0.3213"
],
[
"0.6855",
"0.3132",
"0.6978",
"0.3132",
"0.6978",
"0.324",
"0.6855",
"0.324"
],
[
"0.614",
"0.3145",
"0.621",
"0.3145",
"0.621",
"0.3208",
"0.614",
"0.3208"
],
[
"0.0809",
"0.3125",
"0.1323",
"0.3125",
"0.1323",
"0.3225",
"0.0809",
"0.3225"
],
[
"0.4346",
"0.315",
"0.564",
"0.315",
"0.564",
"0.3289",
"0.4346",
"0.3289"
],
[
"0.814",
"0.3145",
"0.826",
"0.3145",
"0.826",
"0.318",
"0.814",
"0.318"
],
[
"0.3586",
"0.3208",
"0.4182",
"0.3208",
"0.4182",
"0.3308",
"0.3586",
"0.3308"
],
[
"0.2695",
"0.318",
"0.3096",
"0.318",
"0.3096",
"0.3225",
"0.2695",
"0.3225"
],
[
"0.1095",
"0.324",
"0.1487",
"0.324",
"0.1487",
"0.334",
"0.1095",
"0.334"
],
[
"0.0809",
"0.3452",
"0.892",
"0.3452",
"0.892",
"0.3591",
"0.0809",
"0.3591"
],
[
"0.0801",
"0.3591",
"0.892",
"0.3591",
"0.892",
"0.3726",
"0.0801",
"0.3726"
],
[
"0.0784",
"0.3726",
"0.8906",
"0.3718",
"0.8906",
"0.3865",
"0.0784",
"0.387"
],
[
"0.09314",
"0.385",
"0.892",
"0.3865",
"0.892",
"0.401",
"0.09314",
"0.3997"
],
[
"0.0801",
"0.401",
"0.546",
"0.401",
"0.546",
"0.414",
"0.0801",
"0.414"
],
[
"0.0801",
"0.4414",
"0.468",
"0.44",
"0.468",
"0.455",
"0.0801",
"0.4558"
],
[
"0.5034",
"0.4426",
"0.544",
"0.4426",
"0.544",
"0.4539",
"0.5034",
"0.4539"
],
[
"0.54",
"0.4443",
"0.5557",
"0.4443",
"0.5557",
"0.452",
"0.54",
"0.452"
],
[
"0.5537",
"0.443",
"0.722",
"0.443",
"0.722",
"0.4539",
"0.5537",
"0.4539"
],
[
"0.0809",
"0.457",
"0.468",
"0.457",
"0.468",
"0.4702",
"0.0809",
"0.4702"
],
[
"0.0825",
"0.4734",
"0.1381",
"0.4734",
"0.1381",
"0.4849",
"0.0825",
"0.4849"
],
[
"0.505",
"0.4685",
"0.892",
"0.4685",
"0.892",
"0.4817",
"0.505",
"0.4817"
],
[
"0.5034",
"0.483",
"0.892",
"0.4817",
"0.892",
"0.4968",
"0.5034",
"0.4976"
],
[
"0.0972",
"0.4905",
"0.468",
"0.4912",
"0.468",
"0.506",
"0.0972",
"0.505"
],
[
"0.5024",
"0.4988",
"0.8906",
"0.4988",
"0.8906",
"0.513",
"0.5024",
"0.513"
],
[
"0.0809",
"0.507",
"0.4673",
"0.507",
"0.4673",
"0.521",
"0.0809",
"0.521"
],
[
"0.5024",
"0.5146",
"0.892",
"0.5146",
"0.892",
"0.528",
"0.5024",
"0.528"
],
[
"0.0809",
"0.523",
"0.4673",
"0.523",
"0.4673",
"0.536",
"0.0809",
"0.536"
],
[
"0.501",
"0.528",
"0.892",
"0.529",
"0.892",
"0.5435",
"0.501",
"0.5425"
],
[
"0.0801",
"0.537",
"0.4673",
"0.536",
"0.4673",
"0.5503",
"0.0801",
"0.552"
],
[
"0.5034",
"0.5444",
"0.892",
"0.5444",
"0.892",
"0.558",
"0.5034",
"0.558"
],
[
"0.0801",
"0.553",
"0.4673",
"0.553",
"0.4673",
"0.5664",
"0.0801",
"0.5664"
],
[
"0.5024",
"0.56",
"0.892",
"0.56",
"0.892",
"0.5737",
"0.5024",
"0.5737"
],
[
"0.0809",
"0.5674",
"0.468",
"0.5674",
"0.468",
"0.582",
"0.0809",
"0.582"
],
[
"0.5024",
"0.5747",
"0.886",
"0.5747",
"0.886",
"0.5884",
"0.5024",
"0.5884"
],
[
"0.0801",
"0.5835",
"0.4673",
"0.5835",
"0.4673",
"0.5967",
"0.0801",
"0.5967"
],
[
"0.517",
"0.589",
"0.892",
"0.589",
"0.892",
"0.603",
"0.517",
"0.603"
],
[
"0.0801",
"0.5977",
"0.4673",
"0.5967",
"0.4673",
"0.6113",
"0.0801",
"0.6123"
],
[
"0.5024",
"0.603",
"0.892",
"0.605",
"0.892",
"0.6187",
"0.5024",
"0.616"
],
[
"0.0809",
"0.613",
"0.468",
"0.613",
"0.468",
"0.627",
"0.0809",
"0.627"
],
[
"0.505",
"0.619",
"0.8906",
"0.619",
"0.8906",
"0.6333",
"0.505",
"0.6333"
],
[
"0.0801",
"0.6274",
"0.4673",
"0.627",
"0.4673",
"0.6416",
"0.0801",
"0.6426"
],
[
"0.505",
"0.6353",
"0.8906",
"0.6353",
"0.8906",
"0.6484",
"0.505",
"0.6484"
],
[
"0.0809",
"0.643",
"0.4673",
"0.643",
"0.4673",
"0.657",
"0.0809",
"0.657"
],
[
"0.5034",
"0.6494",
"0.8906",
"0.6494",
"0.8906",
"0.6636",
"0.5034",
"0.6636"
],
[
"0.0809",
"0.659",
"0.4697",
"0.659",
"0.4697",
"0.6724",
"0.0809",
"0.6724"
],
[
"0.5034",
"0.6636",
"0.8896",
"0.664",
"0.8896",
"0.6787",
"0.5034",
"0.678"
],
[
"0.505",
"0.6797",
"0.892",
"0.6797",
"0.892",
"0.6934",
"0.505",
"0.6934"
],
[
"0.5186",
"0.6943",
"0.8896",
"0.6943",
"0.8896",
"0.7085",
"0.5186",
"0.7085"
],
[
"0.505",
"0.7104",
"0.892",
"0.7104",
"0.892",
"0.7236",
"0.505",
"0.7236"
],
[
"0.5034",
"0.726",
"0.8823",
"0.726",
"0.8823",
"0.7393",
"0.5034",
"0.7393"
],
[
"0.1316",
"0.733",
"0.3057",
"0.733",
"0.3057",
"0.747",
"0.1316",
"0.747"
],
[
"0.8774",
"0.731",
"0.888",
"0.732",
"0.888",
"0.735",
"0.8774",
"0.7344"
],
[
"0.12335",
"0.7373",
"0.1299",
"0.7373",
"0.1299",
"0.742",
"0.12335",
"0.742"
],
[
"0.5034",
"0.7393",
"0.8906",
"0.7407",
"0.8906",
"0.755",
"0.5034",
"0.754"
],
[
"0.5024",
"0.754",
"0.8906",
"0.755",
"0.8906",
"0.7695",
"0.5024",
"0.768"
],
[
"0.0801",
"0.7344",
"0.11926",
"0.7344",
"0.11926",
"0.7456",
"0.0801",
"0.7456"
],
[
"0.0801",
"0.765",
"0.4697",
"0.7646",
"0.4697",
"0.779",
"0.0801",
"0.78"
],
[
"0.5034",
"0.7695",
"0.693",
"0.7695",
"0.693",
"0.7827",
"0.5034",
"0.7827"
],
[
"0.0784",
"0.78",
"0.468",
"0.781",
"0.468",
"0.7954",
"0.0784",
"0.7944"
],
[
"0.0809",
"0.797",
"0.468",
"0.797",
"0.468",
"0.81",
"0.0809",
"0.81"
],
[
"0.501",
"0.8003",
"0.6855",
"0.8003",
"0.6855",
"0.8145",
"0.501",
"0.8145"
],
[
"0.0809",
"0.811",
"0.4697",
"0.811",
"0.4697",
"0.8247",
"0.0809",
"0.8247"
],
[
"0.0809",
"0.8257",
"0.4534",
"0.8257",
"0.4534",
"0.8403",
"0.0809",
"0.8403"
],
[
"0.523",
"0.8257",
"0.8906",
"0.8257",
"0.8906",
"0.839",
"0.523",
"0.839"
],
[
"0.0825",
"0.8413",
"0.4534",
"0.8413",
"0.4534",
"0.855",
"0.0825",
"0.855"
],
[
"0.4495",
"0.845",
"0.4617",
"0.845",
"0.4617",
"0.8516",
"0.4495",
"0.8516"
],
[
"0.5034",
"0.8413",
"0.8906",
"0.8413",
"0.8906",
"0.855",
"0.5034",
"0.855"
],
[
"0.0809",
"0.856",
"0.468",
"0.855",
"0.468",
"0.869",
"0.0809",
"0.8706"
],
[
"0.5034",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.5034",
"0.8706"
],
[
"0.0809",
"0.872",
"0.468",
"0.872",
"0.468",
"0.8853",
"0.0809",
"0.8853"
],
[
"0.5024",
"0.872",
"0.892",
"0.872",
"0.892",
"0.886",
"0.5024",
"0.886"
],
[
"0.0801",
"0.887",
"0.2336",
"0.887",
"0.2336",
"0.901",
"0.0801",
"0.901"
],
[
"0.5024",
"0.887",
"0.892",
"0.887",
"0.892",
"0.901",
"0.5024",
"0.901"
]
] | [
"epthLe",
"",
"command",
"Route",
"Waypoints",
"Steering",
"3xloop",
"Speed",
"",
"",
"Throttle",
"-",
"Previous",
"Break",
"concat",
"Route",
".Waypoints.!",
"ALD",
"",
"-",
"SDCmapping",
"ReLL",
"MLP",
"Denorm",
"Encoder",
"",
"",
"ell",
"Semantic Segmentation",
"23 layers",
"icin",
"BN+",
"Decode",
"0",
"GRU",
"utior",
"tell",
"R",
"",
"山",
"",
"Redlights",
"Speed",
"",
"全",
"个",
"4",
"",
"Traffic light",
"Speec",
"head",
"AL",
"Skip",
"head",
"Convolutional",
"Encoder",
"个",
"4",
"connectior",
"RGBLehAGBCenterRGBRight",
"ton",
"(Pretrained CVT)",
"Proiection",
"3Stages",
"Figure 1. Model architecture: It consists of trainable and non-trainable components represented by light and dark-colored",
"items, respectively. The blue-colored items represent the perception module , while the controller module and inputs are",
"represented by green and orange , respectively.The white boxes represent different tasks that are learned simultaneously",
"The process inside the top dashed green box is iterated three times during the training, in which the model predicts",
"waypoints to estimate the values of vehicular controls independently.",
"过结合这两种强大技术的优势，CvT确保我们的特征",
"3.1.3",
"",
"SemanticDepthCloud",
"提取器能够捕获局部和全局特征，从而产生better视",
"觉表示。",
"为了增强对场景的理解，我们使用了一种方法，该",
"方法涉及使用提供的深度图像创建BEVSemantic",
"CvT-13[？1，是CvT的轻型版本，由于其出色的",
"Depth Cloud 除了 RGBMAX2. The SDC map rep-",
"性能和较少的参数数量而经过精心选择。正如人们可",
"resents the ego-vehicle's surrounding environment and",
"以看到在图。E1，它被设计为三个主要阶段，每个阶",
"contains richer information with respect to the LiDAR",
"段都包含一个卷积令牌嵌人（CTE）来处理2D输人",
"data due to the 23 semantic class layers : Each layer",
"images。由CTE提取的特征然后被归一化并通过卷",
"represents one class of environment object.:As shown",
"积变换器块(CTB)。To apply both convolutional and",
"in图。2，创建SDC时不考虑高度信息（y-axis)",
"attentionlayers，CTB使用称为卷积投影的深度可分",
"因为BEV透视图只需要-axis and z-axis。创建",
"离卷积运算来创建查询，键和值嵌入。然后将这些嵌人",
"SDC的过程涉及为每个摄像机分别生成SDC贴图，然",
"通过转换器模块来提取全局特征，最终得到高度准确",
"后在以适当的角度旋转侧面贴图后将它们合并在一起。",
"和全面的特征图。CvT的最后一层是用于图像分类的",
"我们在车辆前方定义了64米的距离范围，在摄像机的",
"完全连接层，我们将其从模型中删除。结果，RGB提",
"每一侧定义了32米，创建了64×64平方米的覆盖范",
"取的特征包含384个特征，每个特征的大小为10×48",
"围。SDC地图的正面尺寸为160×320平方厘米，侧面",
"尺寸为160×224。我们使用相机参数为-axis生成",
"个变换矩阵，并将坐标归一化以与SDC张量空间维",
"度对齐。应用独热编码产生23通道SDC张量。生成",
"的映射被复制到一个空的张量，其尺寸为160×768",
"Semantic Segmentation",
"",
"",
"侧图以42度角旋转。对于特征提取，我们使用紧凑的",
"EficientNet-B1[？1网络生成192个特征的张量，每",
"3.1.2",
"提取特征图后，weusethemindifferentsections of the",
"个特征的大小为10×48。",
"model, as they contain valuable information. First, 它",
"们被用来训练能够准确识别23个不同类的语义分割解",
"3.2.Controller Module",
"码器 depicted in the Fig.1 as“Decoder”Later, we",
"use this to createaBEV SDC map。为了实现这",
"控制模块，在图中显示为绿色正方形。1，接收RGB，",
"点，我们开发了一个分段解码器，它由三个卷积层和",
"一",
"SDCand navigational measurements特征，并使用它",
"个具有S形激活的最终逐点卷积组成。通过利用跳过",
"们来预测车辆的未来航路点。The navigational mea-",
"连接，我们可以有效地捕获局部和全局特征，从而生成",
"surements includes route location, navigational com-",
"准确而全面的分割图。",
"mand provided by the global planner and the ego ve-"
] | {
"arxiv_id": "2310.13135",
"language": "zh",
"page_idx": 2
} |
||
2310.13135-zh-3 | 2310.13135-zh-3.jpg | <FigureHere>
Figure 2. Semantic Depth Cloud (SDC) map created from three depth and semantic segmentations acquired from the three cameras attached to the vehicle. The SDC mapping creates three separate BEV maps using estimated semantic segmentation and attaches them together.
个线性层和一个整流线性单元 (ReLU)的多层感知器(MLP)网络应用于偏置的 GRU 隐藏状态，以获得范围为 0 到 1 的归一化控制命令。
$$\begin{bmatrix}x_l\\y_l\end{bmatrix}=\begin{bmatrix}cos(90^\circ+\theta_v)&-sin(90^\circ+\theta_v)\\sin(90^\circ+\theta_v)&cos(90^\circ+\theta_v)\end{bmatrix}^\top\begin{bmatrix}x_g-x_vg\\y_g-y_vg\end{bmatrix}$$
$(\bar{1})$
hicle speed. The navigational command specifies the vehicle's general direction, such as left, right, forward, stop, etc., and is defined as a one-hot vector. 随后the control module 根据预测的航路点和融合的特征预测适当的车辆控制，包括转向、油门和制动。为了预测车辆的未来航路点，我们采用类似于？]的门控循环单元 (GRU)。GRU 是一个合适的选择，因为它解决了消失的梯度问题，同时与其他 RNN 方法相比保持了更好的性能成本比。To train a control model that predicts current control actions based on current input, behavior cloning is commonly used but relies on the assumption of independent and identically distributed (IID) data, which is not valid for closed-loop tests[?]. To overcome this issue without reinforcement learning, we used a similar trick to? ] for predicting multi-step control actions into the future. To this end, we first employ a waypoint branch that utilizes fused features and environment-agent static knowledge through waypoint bypass. Further, we deploy a dynamic branch to capture the environment-agent dynamic interaction given the learned static knowledge. The dynamic branch provides dynamic information such as object motion and traffic light changes, while the waypoint branch incorporates static information like curbs and lanes and improves spatial consistency across both branches.
The GRU in the dynamic branch takes the same fused features as the waypoint branch for the initial hidden state to improve consistency and the inputs include the predicted vehicular command from the waypoint branch. The result is then concatenated with the same waypoint bypass, representing the abstract static coarse simulator and then fed to MLP and sigmoid to create adjusted control output. 为了确定合适的车辆控制，我们用两种不同的方式计算它们。首先，我们对来自航路点分支的预测车辆命令的总和进行非规范化， 并从动态分支调整控制。其次，我们使用两个单独的PID 控制器来预测车辆控制，一个用于查找转向命令(横向),另一个用于查找油门和制动器(纵向),使用预测的航路点和当前速度。我们的控制策略类似于？]
它根据场景使用两种方法计算控制命令。在行驶过程中，车辆依靠前两个航路点通过取它们的平均值来计算其下一个目的地。但是，我们还会生成第三个航路点的预测，该预测向 GRU 和航路点预测层提供额外的数据。此外，此方法使 MLP 和 PID 代理能够接收相同的信息，因为最后一个有偏差的隐藏状态包含来自第二个航路点的详细信息。
# 4. Experiments
## $4. 1. \textbf{Dataset}$
In order to fuse the extracted features from RGB images and the SDC map , we concatenate them and apply batch normalization (BN) to them and then concatenate the result with the measurement tensor
航路点分支中的 GRU 将融合特征作为初始隐藏状态， 输入包括 BEV 空间中的当前航路点，转换为 BEV 空间的路线位置坐标。初始航路点坐标始终位于(0,0) , 即 SDC 地图的底部中心点。为了将全局坐标$(x_g,y_g)$ 转换为局部坐标 $(x_l,y_l)$ ,我们使用公式$\mathbb{E}$1 。然后，我们将下一个隐藏状态从 GRU 添加到航路点旁路，这是通过偏置模块传递的 RGB 特征。偏置模块包括自适应全局池和应用于 RGB 提取特征的线性层。最后，应用sigmoid 函数转换 0 和 1 之间的所有值。我们将包含两
我们使用 CARLA[?] (0。9。10)???个可用于培训和测试的城镇。We train our model on the 210 GB publicly available dataset presented in the TransFuser $[?]{\mathrm{~for~the~experiment.~}}$所有 8个城镇 were 用于训练，数据集包括大约 2500 条通过平均长度为 100m 的路口的路线和？? 条平均长度为 400m 的 including 弯曲公路。为了生成用于训练目的的数据，制定了专家 | [
[
"0.517",
"0.11426",
"0.5786",
"0.11615",
"0.5786",
"0.1307",
"0.517",
"0.1288"
],
[
"0.7524",
"0.11426",
"0.8228",
"0.11615",
"0.8228",
"0.132",
"0.7524",
"0.13"
],
[
"0.12335",
"0.1477",
"0.1528",
"0.1477",
"0.1528",
"0.1547",
"0.12335",
"0.1547"
],
[
"0.179",
"0.1193",
"0.2141",
"0.1193",
"0.2141",
"0.1307",
"0.179",
"0.1307"
],
[
"0.1323",
"0.2134",
"0.2566",
"0.2134",
"0.2566",
"0.2235",
"0.1323",
"0.2235"
],
[
"0.0825",
"0.2443",
"0.8906",
"0.2443",
"0.8906",
"0.2583",
"0.0825",
"0.2583"
],
[
"0.0801",
"0.2583",
"0.58",
"0.2588",
"0.58",
"0.2727",
"0.0801",
"0.2715"
],
[
"0.5728",
"0.26",
"0.8896",
"0.26",
"0.8896",
"0.2708",
"0.5728",
"0.2708"
],
[
"0.0825",
"0.2747",
"0.357",
"0.2747",
"0.357",
"0.2854",
"0.0825",
"0.2854"
],
[
"0.0801",
"0.3132",
"0.4697",
"0.3132",
"0.4697",
"0.3276",
"0.0801",
"0.3276"
],
[
"0.505",
"0.3132",
"0.8906",
"0.3132",
"0.8906",
"0.3276",
"0.505",
"0.3276"
],
[
"0.0809",
"0.3289",
"0.4673",
"0.3289",
"0.4673",
"0.3435",
"0.0809",
"0.3435"
],
[
"0.501",
"0.3276",
"0.8906",
"0.3289",
"0.8906",
"0.3435",
"0.501",
"0.3423"
],
[
"0.0809",
"0.3447",
"0.433",
"0.3447",
"0.433",
"0.358",
"0.0809",
"0.358"
],
[
"0.4265",
"0.3452",
"0.4673",
"0.3452",
"0.4673",
"0.3567",
"0.4265",
"0.3567"
],
[
"0.505",
"0.3447",
"0.7295",
"0.3447",
"0.7295",
"0.358",
"0.505",
"0.358"
],
[
"0.0809",
"0.3599",
"0.4673",
"0.3599",
"0.4673",
"0.3738",
"0.0809",
"0.3738"
],
[
"0.564",
"0.3718",
"0.671",
"0.3718",
"0.671",
"0.3865",
"0.564",
"0.3865"
],
[
"0.6626",
"0.3718",
"0.7935",
"0.3718",
"0.7935",
"0.3865",
"0.6626",
"0.3865"
],
[
"0.808",
"0.3694",
"0.8823",
"0.3718",
"0.8823",
"0.3894",
"0.808",
"0.387"
],
[
"0.5107",
"0.3726",
"0.535",
"0.3726",
"0.535",
"0.3865",
"0.5107",
"0.3865"
],
[
"0.0825",
"0.375",
"0.468",
"0.375",
"0.468",
"0.3882",
"0.0825",
"0.3882"
],
[
"0.562",
"0.3865",
"0.665",
"0.387",
"0.665",
"0.4016",
"0.562",
"0.401"
],
[
"0.6665",
"0.3838",
"0.7983",
"0.387",
"0.7983",
"0.4028",
"0.6665",
"0.3997"
],
[
"0.0809",
"0.3901",
"0.4673",
"0.3901",
"0.4673",
"0.404",
"0.0809",
"0.404"
],
[
"0.511",
"0.3901",
"0.535",
"0.3901",
"0.535",
"0.4016",
"0.511",
"0.4016"
],
[
"0.806",
"0.3894",
"0.8813",
"0.3894",
"0.8813",
"0.404",
"0.806",
"0.404"
],
[
"0.0801",
"0.404",
"0.468",
"0.4048",
"0.468",
"0.4192",
"0.0801",
"0.4185"
],
[
"0.869",
"0.4048",
"0.8896",
"0.4048",
"0.8896",
"0.4172",
"0.869",
"0.4172"
],
[
"0.08417",
"0.4204",
"0.4673",
"0.4204",
"0.4673",
"0.4343",
"0.08417",
"0.4343"
],
[
"0.5215",
"0.4167",
"0.8906",
"0.4172",
"0.8906",
"0.432",
"0.5215",
"0.4312"
],
[
"0.0825",
"0.435",
"0.4673",
"0.435",
"0.4673",
"0.449",
"0.0825",
"0.449"
],
[
"0.5024",
"0.432",
"0.8906",
"0.433",
"0.8906",
"0.4475",
"0.5024",
"0.4463"
],
[
"0.0801",
"0.449",
"0.468",
"0.4507",
"0.468",
"0.464",
"0.0801",
"0.4622"
],
[
"0.5024",
"0.449",
"0.8906",
"0.449",
"0.8906",
"0.4622",
"0.5024",
"0.4622"
],
[
"0.0784",
"0.4634",
"0.468",
"0.464",
"0.468",
"0.4792",
"0.0784",
"0.4778"
],
[
"0.5034",
"0.4634",
"0.8906",
"0.4634",
"0.8906",
"0.4766",
"0.5034",
"0.4766"
],
[
"0.0809",
"0.4797",
"0.4697",
"0.4797",
"0.4697",
"0.4937",
"0.0809",
"0.4937"
],
[
"0.501",
"0.4778",
"0.892",
"0.4766",
"0.892",
"0.4912",
"0.501",
"0.4924"
],
[
"0.0825",
"0.4956",
"0.468",
"0.4956",
"0.468",
"0.509",
"0.0825",
"0.509"
],
[
"0.5034",
"0.4937",
"0.8906",
"0.4937",
"0.8906",
"0.507",
"0.5034",
"0.507"
],
[
"0.0801",
"0.509",
"0.4673",
"0.5103",
"0.4673",
"0.5244",
"0.0801",
"0.5234"
],
[
"0.5024",
"0.5083",
"0.892",
"0.5083",
"0.892",
"0.5215",
"0.5024",
"0.5215"
],
[
"0.0809",
"0.526",
"0.468",
"0.526",
"0.468",
"0.539",
"0.0809",
"0.539"
],
[
"0.505",
"0.526",
"0.8896",
"0.526",
"0.8896",
"0.536",
"0.505",
"0.536"
],
[
"0.0825",
"0.5425",
"0.4656",
"0.5425",
"0.4656",
"0.553",
"0.0825",
"0.553"
],
[
"0.5024",
"0.539",
"0.8906",
"0.539",
"0.8906",
"0.553",
"0.5024",
"0.553"
],
[
"0.0809",
"0.556",
"0.468",
"0.556",
"0.468",
"0.5693",
"0.0809",
"0.5693"
],
[
"0.5034",
"0.5537",
"0.8906",
"0.5537",
"0.8906",
"0.5674",
"0.5034",
"0.5674"
],
[
"0.0809",
"0.5713",
"0.468",
"0.5713",
"0.468",
"0.5854",
"0.0809",
"0.5854"
],
[
"0.505",
"0.5693",
"0.8906",
"0.5693",
"0.8906",
"0.5835",
"0.505",
"0.5835"
],
[
"0.0809",
"0.586",
"0.4673",
"0.586",
"0.4673",
"0.5996",
"0.0809",
"0.5996"
],
[
"0.5034",
"0.5835",
"0.888",
"0.584",
"0.888",
"0.5986",
"0.5034",
"0.5977"
],
[
"0.0809",
"0.601",
"0.4673",
"0.601",
"0.4673",
"0.6143",
"0.0809",
"0.6143"
],
[
"0.505",
"0.5996",
"0.888",
"0.5996",
"0.888",
"0.613",
"0.505",
"0.613"
],
[
"0.0809",
"0.616",
"0.4673",
"0.616",
"0.4673",
"0.63",
"0.0809",
"0.63"
],
[
"0.505",
"0.616",
"0.884",
"0.616",
"0.884",
"0.627",
"0.505",
"0.627"
],
[
"0.0809",
"0.631",
"0.468",
"0.631",
"0.468",
"0.6445",
"0.0809",
"0.6445"
],
[
"0.523",
"0.63",
"0.8906",
"0.63",
"0.8906",
"0.643",
"0.523",
"0.643"
],
[
"0.0809",
"0.6465",
"0.468",
"0.6465",
"0.468",
"0.66",
"0.0809",
"0.66"
],
[
"0.5034",
"0.6426",
"0.8906",
"0.643",
"0.8906",
"0.6577",
"0.5034",
"0.657"
],
[
"0.0801",
"0.661",
"0.4673",
"0.66",
"0.4673",
"0.675",
"0.0801",
"0.6753"
],
[
"0.5034",
"0.6577",
"0.8906",
"0.659",
"0.8906",
"0.674",
"0.5034",
"0.6724"
],
[
"0.0801",
"0.677",
"0.423",
"0.677",
"0.423",
"0.69",
"0.0801",
"0.69"
],
[
"0.505",
"0.675",
"0.8896",
"0.675",
"0.8896",
"0.688",
"0.505",
"0.688"
],
[
"0.505",
"0.69",
"0.8906",
"0.69",
"0.8906",
"0.704",
"0.505",
"0.704"
],
[
"0.0989",
"0.705",
"0.468",
"0.705",
"0.468",
"0.7183",
"0.0989",
"0.7183"
],
[
"0.5063",
"0.705",
"0.8906",
"0.705",
"0.8906",
"0.7183",
"0.5063",
"0.7183"
],
[
"0.0809",
"0.72",
"0.468",
"0.72",
"0.468",
"0.7344",
"0.0809",
"0.7344"
],
[
"0.5034",
"0.72",
"0.6914",
"0.72",
"0.6914",
"0.7344",
"0.5034",
"0.7344"
],
[
"0.0809",
"0.7363",
"0.464",
"0.7363",
"0.464",
"0.75",
"0.0809",
"0.75"
],
[
"0.0825",
"0.7524",
"0.4412",
"0.7524",
"0.4412",
"0.763",
"0.0825",
"0.763"
],
[
"0.5024",
"0.7485",
"0.6396",
"0.7485",
"0.6396",
"0.763",
"0.5024",
"0.763"
],
[
"0.0825",
"0.7666",
"0.4673",
"0.7666",
"0.4673",
"0.78",
"0.0825",
"0.78"
],
[
"0.5024",
"0.773",
"0.6006",
"0.773",
"0.6006",
"0.787",
"0.5024",
"0.787"
],
[
"0.0809",
"0.781",
"0.468",
"0.781",
"0.468",
"0.7944",
"0.0809",
"0.7944"
],
[
"0.0825",
"0.797",
"0.4631",
"0.797",
"0.4631",
"0.81",
"0.0825",
"0.81"
],
[
"0.523",
"0.7954",
"0.8906",
"0.7954",
"0.8906",
"0.81",
"0.523",
"0.81"
],
[
"0.0809",
"0.8096",
"0.468",
"0.811",
"0.468",
"0.8257",
"0.0809",
"0.8237"
],
[
"0.505",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.505",
"0.8247"
],
[
"0.0809",
"0.8257",
"0.468",
"0.8257",
"0.468",
"0.8403",
"0.0809",
"0.8403"
],
[
"0.5034",
"0.8257",
"0.8906",
"0.8257",
"0.8906",
"0.839",
"0.5034",
"0.839"
],
[
"0.0825",
"0.8413",
"0.4673",
"0.8413",
"0.4673",
"0.855",
"0.0825",
"0.855"
],
[
"0.505",
"0.8413",
"0.8896",
"0.8413",
"0.8896",
"0.855",
"0.505",
"0.855"
],
[
"0.0809",
"0.857",
"0.468",
"0.857",
"0.468",
"0.8706",
"0.0809",
"0.8706"
],
[
"0.5034",
"0.856",
"0.8906",
"0.856",
"0.8906",
"0.869",
"0.5034",
"0.869"
],
[
"0.0809",
"0.872",
"0.468",
"0.872",
"0.468",
"0.886",
"0.0809",
"0.886"
],
[
"0.505",
"0.872",
"0.892",
"0.872",
"0.892",
"0.8853",
"0.505",
"0.8853"
],
[
"0.0809",
"0.887",
"0.468",
"0.887",
"0.468",
"0.901",
"0.0809",
"0.901"
],
[
"0.5034",
"0.887",
"0.892",
"0.887",
"0.892",
"0.901",
"0.5034",
"0.901"
]
] | [
"23 Layers",
"SDC Map",
"",
"Depth",
"SemanticSegmentation",
"Figure 2. Semantic Depth Cloud (SDC) map created from three depth and semantic segmentations acquired from the",
"threecamerasattachedtothevehicle.TheSDCmappingcreatesthree",
"separateBEV maps using estimated semantic",
"segmentation andattachesthemtogether",
"hicle speed. The navigational command specifies the",
"个线性层和一个整流线性单元（ReLU）的多层感知器",
"vehicle's general direction, such as left, right, forward,",
"（MLP）网络应用于偏置的GRU隐藏状态，以获得范",
"stop, etc., and is defined as a one-hot vector.",
"随后",
"围为0到1的归一化控制命令。",
"the control module根据预测的航路点和融合的特征",
"[cos(90°+0)",
"-sin(90°+0%))",
"[ag-aug",
"[ai]",
"预测适当的车辆控制，包括转向、油门和制动。为了预",
"sin(90°+0)",
"cos(90°+0)",
"测车辆的未来航路点，我们采用类似于？1的门控循",
"yu",
"[9g-Yug",
"环单元（GRU）。GRU是一个合适的选择，因为它解决",
"(1)",
"了消失的梯度问题，同时与其他RNN方法相比保持了",
"The GRU in the dynamic branch takes the same",
"更好的性能成本比。Totrainacontrolmodelthatpre-",
"fused features as the waypoint branch for the initial",
"dicts current control actions based on current input,",
"hidden state to improve consistency and the inputs in-",
"behavior cloning is commonly used but relies on the",
"clude the predicted vehicular command from the way-",
"assumption of independent and identically distributed",
"point branch. The result is then concatenated with the",
"(IID) data, which is not valid for closed-loop tests [? 1 .",
"same waypoint bypass, representing the abstract static",
"To overcome this issue without reinforcement learning,",
"coarse simulator and then fed to MLP and sigmoid to",
"we used a similar trick to ? for predicting multi-step",
"createadjustedcontroloutput为了确定合适的车辆",
"controlactionsintothefuture.Tothisend,wefirstem-",
"控制，我们用两种不同的方式计算它们。首先，我们对",
"ploy a waypoint branch that utilizes fused features and",
"来自航路点分支的预测车辆命令的总和进行非规范化，",
"environment-agent static knowledge through waypoint",
"并从动态分支调整控制。其次，我们使用两个单独的",
"bypass. Further, we deploy a dynamic branch to cap-",
"PID控制器来预测车辆控制，一个用于查找转向命令",
"ture the environment-agent dynamic interaction given",
"（横向），另一个用于查找油门和制动器（纵向），使用",
"the learned static knowledge. The dynamic branch pro-",
"预测的航路点和当前速度。我们的控制策略类似于？",
"vides dynamic information such as object motion and",
"它根据场景使用两种方法计算控制命令。在行驶过",
"traffic light changes, while the waypoint branch incor-",
"程中，车辆依靠前两个航路点通过取它们的平均值来",
"porates static information like curbs and lanes and im-",
"计算其下一个目的地。但是，我们还会生成第三个航路",
"proves spatial consistency across both branches.",
"点的预测，该预测向GRU和航路点预测层提供额外的",
"数据。此外，此方法使MLP和PID代理能够接收相",
"In order to fuse the extracted features from RGB",
"同的信息，因为最后一个有偏差的隐藏状态包含来自",
"images and the SDC map , we concatenate them and",
"第二个航路点的详细信息。",
"apply batch normalization (BN) to them and then con.",
"catenatetheresultwiththemeasurement tensor",
"4.Experiments",
"航路点分支中的GRU将融合特征作为初始隐藏状态，",
"4.1. Dataset",
"输人包括BEV空间中的当前航路点，转换为BEV空",
"间的路线位置坐标。初始航路点坐标始终位于(0,0)",
"我们使用CARLA[？1（0。9。10)??？个可用于培",
"即SDC地图的底部中心点。为了将全局坐标（g，Yg)",
"训和测试的城镇。Wetrainourmodel onthe 210GB",
"转换为局部坐标（at,yi），我们使用公式E1。然后，我",
"publicly available dataset presented in the TransFuser",
"们将下一个隐藏状态从GRU添加到航路点旁路，这是",
"[?］fortheexperiment．所有8个城镇were用于训",
"通过偏置模块传递的RGB特征。偏置模块包括自适应",
"练，数据集包括大约2500条通过平均长度为100m的",
"全局池和应用于RGB提取特征的线性层。最后，应用",
"路口的路线和???条平均长度为400m的including弯",
"sigmoid函数转换0和1之间的所有值。我们将包含两",
"曲公路。为了生成用于训练目的的数据，制定了专家"
] | {
"arxiv_id": "2310.13135",
"language": "zh",
"page_idx": 3
} |
||
2310.13135-zh-4 | 2310.13135-zh-4.jpg | 策略，该策略采用从模拟器获得的特权信息来控制驾驶过程 inspired from? ] 。专家的航路点可作为模仿损失的真实标签，使 the expert 可与自动标记算法相媲美 To accomplish lateral control, the expert policy follows the path generated by the A* planner, and a PID controller is used to minimize the angle of the vehicle towards the next waypoint in the route, which is at least 3.5 meters away. Meanwhile, longitudinal control is performed using a version of model predictive control, which differentiates between 3 target speeds. 标准目标速度为 4。0 米/秒，但速度降低到 3。当专家在交叉路口内时，速度为 0m/s。此外，如果预测到违规，则目标速度 changes 为 0。0m/s, 使车辆停止。纵向和横向控制器分别使用$K_p=5.0,K_i=0.5,K_d=1.0$ 和$K_{p}=1.25,K_{i}=0.75,K_{d}=0.3$的 PID 值。对于两个控制器的积分项的运行平均值，我们使用大小为 40 的缓冲区。最初，all three RGB 图像和深度图以$480\times960$ then cropped to $160\times320$的分辨率检索，以避免失真。因此，all three RGB 图像表示为$R\in0,\ldots,255^{3\times160\times320}$attached to each other sideby-side to create a single image. The same process is applied to the depth images to create one single depth image. In order to calculate the real depth value for each pixel i in the depth map we use the following :
$$R_i^{dec}=\frac{256^2B_i+256G_i+R_i}{256^3-1}\times1000,$$
(2)
其中$(R_i,G_i,B_i)$ are the stored values of pixel i. Also, since the stored values are 8-bit, they have a maximum of 255, and 1000 here is the maximum depth range of the RGB-D camera in meters. All together they result in $R_{i}^{dec}$ which represents the true depth of pixel i. <NEX。语义分割的对象类是根据？]。此外，航路点用$\omega\rho_i=(x_i,y_i)_{i=1}^3$表示在 BEV 空间中。The center of the BEV space, marked by coordinates (0,0) in the local vehicle coordinate system, is positioned on the ego vehicle itself, at the bottom-center point. The model assesses vehicular controls within a normalized range spanning from 0 to 1, subsequently denormalizing them to their original values, encompassing steering within the range of [-1,1] , throttle ranging from [0,0.75], and brake as either 0 or 1.
The traffic light state value changes to 1 if a red light appeared, otherwise it is stated as 0. We encapsulate the speed measurement (in m/s) and GPS locations into a one-hot encoded vector as a high level navigational command.
## 4.2. Task and Scenario
The task entails navigating through a variety of areas such as highways, cityscapes, and residential neighborhoods along the predefined paths consisting of
sparse goal locations specified in the GPS coordinates provided by a global planner. The routes encompass numerous scenarios, each initialized at predefined positions, designed to evaluate the agent's capability to adapt wide spectrum of adversarial situations and varying weather conditions. 第一个场景称为 1WN, 涉及在除 Town05 之外的所有可用地图和路线集上训练模型，这些地图和路线集保留用于验证。The model is then evaluated on both Town05 short and long routes, consisting of 32 short and 10 long routes, to assess its performance 。评估是在正常情况下在晴朗的中午条件下进行的，所有非玩家角色都遵守交通规则。在第二种场景 1WA 中，该模型在 adversarial 非玩家角色(NPC)行为下进行了测试，这可能会导致碰撞，例如行人突然过马路或骑自行车的人出现。此外，交通灯管理器可能会故意在交叉路口创建带有双绿灯的状态，以模拟救护车或消防员可能跳过交通灯的紧急情况。除了正确驾驶自我车辆外，该模型有望在这些对抗性情况下安全地做出反应 to the environment changes and noises and avoid accidents .
## 4.3. Implementation Details
To effectively gather knowledge across multiple tasks using end-to-end learning, we predefined a set of distinct loss functions. The comprehensive loss that encompasses all tasks can be calculated as follows :
$$\mathcal{L}_{\mathrm{TOTAL}}=\alpha_{1}\mathcal{L}_{\mathrm{SEG}}+\alpha_{2}\mathcal{L}_{\mathrm{ST}}+\alpha_{3}\mathcal{L}_{\mathrm{TH}}+\\\alpha_{4}\mathcal{L}_{\mathrm{BR}}+\alpha_{5}\mathcal{L}_{\mathrm{WP}}+\alpha_{6}\mathcal{L}_{\mathrm{TL}}+\alpha_{7}\mathcal{L}_{\mathrm{SS}}+\alpha_{8}\mathcal{L}_{\mathrm{VE}}$$
(3)
其中转向损失(XMATHX。此外，每个任务的损失权重由$\alpha_1,\alpha_2,\ldots,\alpha_8$表示。For the purpose of calculating the semantic segmentation loss function ( $\mathcal{L}_\mathrm{SEG}$ ), a mixture of binary cross-entropy and dice loss is utilized and can be computed through following equation.
$$\begin{aligned}\mathcal{L}_{SEG}=\left(\frac{1}{N}\sum_{i=1}^{N}y_{i}\log(\hat{y}_{i})+(1-y_{i})\log(1-\hat{y}_{i})\right)\\&+\left(1-\frac{2|\hat{y}\cup y|}{|\hat{y}|+|y|}\right)\end{aligned}$$
(4)
Here, $N$ represents the number of pixels at the final layer of our semantic segmentation decoder. Subsequently, $y_i$ and $\hat{y}_i$ correspond to the values of the $i$ -th element in the ground truth vector $y$ and the prediction vector $\hat{y}$, respectively. 此方法允许我们同时利用基于分布和基于区域的损耗组件，如 by?]所述。考虑到整个网络的结构完整性依赖于它，强调用补充损失标准增强语义分割任务的重要性势在必行。对于三个预测航路点的情况，只有( $\mathcal{L}_\mathrm{WP}$)需要平均。
In order to adaptively adjust the loss weights for each training epoch, We utilize the MGN algorithm [?
。为了实现这一点，我们采用了 Adam 优化器
$\left[\begin{array}{c}\\\\\\\end{array}\right]$ | [
[
"0.0825",
"0.09216",
"0.468",
"0.09216",
"0.468",
"0.1061",
"0.0825",
"0.1061"
],
[
"0.501",
"0.09216",
"0.8906",
"0.09155",
"0.8906",
"0.1061",
"0.501",
"0.1067"
],
[
"0.0825",
"0.108",
"0.468",
"0.108",
"0.468",
"0.1218",
"0.0825",
"0.1218"
],
[
"0.5024",
"0.108",
"0.892",
"0.108",
"0.892",
"0.1218",
"0.5024",
"0.1218"
],
[
"0.0801",
"0.1218",
"0.468",
"0.1225",
"0.468",
"0.137",
"0.0801",
"0.1364"
],
[
"0.5024",
"0.1218",
"0.8906",
"0.1225",
"0.8906",
"0.137",
"0.5024",
"0.1364"
],
[
"0.0809",
"0.1383",
"0.468",
"0.1383",
"0.468",
"0.1515",
"0.0809",
"0.1515"
],
[
"0.501",
"0.137",
"0.8906",
"0.137",
"0.8906",
"0.1509",
"0.501",
"0.1509"
],
[
"0.0809",
"0.1528",
"0.4697",
"0.1528",
"0.4697",
"0.1666",
"0.0809",
"0.1666"
],
[
"0.5034",
"0.1528",
"0.8896",
"0.1528",
"0.8896",
"0.1666",
"0.5034",
"0.1666"
],
[
"0.0801",
"0.1666",
"0.468",
"0.1674",
"0.468",
"0.1818",
"0.0801",
"0.1812"
],
[
"0.5034",
"0.1674",
"0.8906",
"0.1674",
"0.8906",
"0.1812",
"0.5034",
"0.1812"
],
[
"0.0809",
"0.1831",
"0.468",
"0.1831",
"0.468",
"0.1963",
"0.0809",
"0.1963"
],
[
"0.505",
"0.1831",
"0.8906",
"0.1831",
"0.8906",
"0.1963",
"0.505",
"0.1963"
],
[
"0.0801",
"0.1976",
"0.4697",
"0.1976",
"0.4697",
"0.2108",
"0.0801",
"0.2108"
],
[
"0.505",
"0.1976",
"0.892",
"0.1976",
"0.892",
"0.2108",
"0.505",
"0.2108"
],
[
"0.0801",
"0.2134",
"0.4714",
"0.2134",
"0.4714",
"0.2267",
"0.0801",
"0.2267"
],
[
"0.5024",
"0.2122",
"0.8906",
"0.2134",
"0.8906",
"0.2267",
"0.5024",
"0.226"
],
[
"0.0825",
"0.2279",
"0.4617",
"0.2279",
"0.4617",
"0.2412",
"0.0825",
"0.2412"
],
[
"0.5034",
"0.2291",
"0.892",
"0.2291",
"0.892",
"0.2424",
"0.5034",
"0.2424"
],
[
"0.0809",
"0.2437",
"0.468",
"0.2437",
"0.468",
"0.2568",
"0.0809",
"0.2568"
],
[
"0.5034",
"0.2437",
"0.8906",
"0.2437",
"0.8906",
"0.2568",
"0.5034",
"0.2568"
],
[
"0.0825",
"0.2583",
"0.468",
"0.2583",
"0.468",
"0.2727",
"0.0825",
"0.2727"
],
[
"0.5034",
"0.2588",
"0.8906",
"0.2588",
"0.8906",
"0.2727",
"0.5034",
"0.2727"
],
[
"0.0809",
"0.2734",
"0.468",
"0.2734",
"0.468",
"0.2874",
"0.0809",
"0.2874"
],
[
"0.5107",
"0.2727",
"0.888",
"0.2727",
"0.888",
"0.286",
"0.5107",
"0.286"
],
[
"0.0809",
"0.2886",
"0.468",
"0.2874",
"0.468",
"0.3018",
"0.0809",
"0.303"
],
[
"0.5024",
"0.2874",
"0.8906",
"0.2886",
"0.8906",
"0.303",
"0.5024",
"0.3018"
],
[
"0.0825",
"0.3037",
"0.468",
"0.3037",
"0.468",
"0.3176",
"0.0825",
"0.3176"
],
[
"0.5063",
"0.3037",
"0.8906",
"0.3037",
"0.8906",
"0.3176",
"0.5063",
"0.3176"
],
[
"0.08417",
"0.318",
"0.468",
"0.318",
"0.468",
"0.332",
"0.08417",
"0.332"
],
[
"0.5063",
"0.3208",
"0.8896",
"0.3208",
"0.8896",
"0.3308",
"0.5063",
"0.3308"
],
[
"0.0825",
"0.3333",
"0.4673",
"0.3333",
"0.4673",
"0.3467",
"0.0825",
"0.3467"
],
[
"0.5034",
"0.334",
"0.8906",
"0.334",
"0.8906",
"0.348",
"0.5034",
"0.348"
],
[
"0.0809",
"0.3499",
"0.468",
"0.3499",
"0.468",
"0.363",
"0.0809",
"0.363"
],
[
"0.505",
"0.3484",
"0.8896",
"0.3484",
"0.8896",
"0.3623",
"0.505",
"0.3623"
],
[
"0.0809",
"0.363",
"0.468",
"0.363",
"0.468",
"0.3774",
"0.0809",
"0.3774"
],
[
"0.5034",
"0.3643",
"0.892",
"0.3643",
"0.892",
"0.3774",
"0.5034",
"0.3774"
],
[
"0.0784",
"0.3738",
"0.4697",
"0.3738",
"0.4697",
"0.3945",
"0.0784",
"0.3945"
],
[
"0.5024",
"0.3801",
"0.705",
"0.3801",
"0.705",
"0.3933",
"0.5024",
"0.3933"
],
[
"0.0801",
"0.3945",
"0.468",
"0.3945",
"0.468",
"0.4077",
"0.0801",
"0.4077"
],
[
"0.5034",
"0.404",
"0.722",
"0.404",
"0.722",
"0.4185",
"0.5034",
"0.4185"
],
[
"0.0801",
"0.4104",
"0.468",
"0.4104",
"0.468",
"0.4236",
"0.0801",
"0.4236"
],
[
"0.0801",
"0.4248",
"0.468",
"0.4248",
"0.468",
"0.4382",
"0.0801",
"0.4382"
],
[
"0.5195",
"0.4248",
"0.892",
"0.4268",
"0.892",
"0.4414",
"0.5195",
"0.4395"
],
[
"0.0801",
"0.44",
"0.4617",
"0.44",
"0.4617",
"0.4539",
"0.0801",
"0.4539"
],
[
"0.5024",
"0.4426",
"0.8906",
"0.4426",
"0.8906",
"0.457",
"0.5024",
"0.457"
],
[
"0.1748",
"0.464",
"0.3708",
"0.464",
"0.3708",
"0.4792",
"0.1748",
"0.4792"
],
[
"0.505",
"0.459",
"0.8906",
"0.459",
"0.8906",
"0.4697",
"0.505",
"0.4697"
],
[
"0.12335",
"0.4717",
"0.1724",
"0.4717",
"0.1724",
"0.4873",
"0.12335",
"0.4873"
],
[
"0.3677",
"0.4717",
"0.468",
"0.4749",
"0.468",
"0.4893",
"0.3677",
"0.486"
],
[
"0.5024",
"0.473",
"0.876",
"0.473",
"0.876",
"0.486",
"0.5024",
"0.486"
],
[
"0.2402",
"0.4817",
"0.3064",
"0.4817",
"0.3064",
"0.4968",
"0.2402",
"0.4968"
],
[
"0.0825",
"0.506",
"0.468",
"0.506",
"0.468",
"0.5195",
"0.0825",
"0.5195"
],
[
"0.5645",
"0.5005",
"0.847",
"0.5005",
"0.847",
"0.5146",
"0.5645",
"0.5146"
],
[
"0.869",
"0.5103",
"0.8906",
"0.5103",
"0.8906",
"0.523",
"0.869",
"0.523"
],
[
"0.0801",
"0.5215",
"0.468",
"0.5215",
"0.468",
"0.535",
"0.0801",
"0.535"
],
[
"0.527",
"0.5195",
"0.846",
"0.5195",
"0.846",
"0.5327",
"0.527",
"0.5327"
],
[
"0.0809",
"0.536",
"0.4697",
"0.536",
"0.4697",
"0.55",
"0.0809",
"0.55"
],
[
"0.0801",
"0.55",
"0.468",
"0.5503",
"0.468",
"0.565",
"0.0801",
"0.5645"
],
[
"0.5034",
"0.5454",
"0.8906",
"0.547",
"0.8906",
"0.561",
"0.5034",
"0.56"
],
[
"0.0809",
"0.5664",
"0.4558",
"0.5664",
"0.4558",
"0.58",
"0.0809",
"0.58"
],
[
"0.5024",
"0.561",
"0.8906",
"0.561",
"0.8906",
"0.5757",
"0.5024",
"0.5757"
],
[
"0.0825",
"0.5806",
"0.468",
"0.5806",
"0.468",
"0.5947",
"0.0825",
"0.5947"
],
[
"0.5024",
"0.5776",
"0.892",
"0.5776",
"0.892",
"0.5913",
"0.5024",
"0.5913"
],
[
"0.0809",
"0.5977",
"0.468",
"0.5977",
"0.468",
"0.6123",
"0.0809",
"0.6123"
],
[
"0.5024",
"0.5923",
"0.892",
"0.5923",
"0.892",
"0.606",
"0.5024",
"0.606"
],
[
"0.0801",
"0.6113",
"0.468",
"0.6123",
"0.468",
"0.627",
"0.0801",
"0.6255"
],
[
"0.5034",
"0.608",
"0.859",
"0.608",
"0.859",
"0.622",
"0.5034",
"0.622"
],
[
"0.0784",
"0.627",
"0.4673",
"0.629",
"0.4673",
"0.6426",
"0.0784",
"0.64"
],
[
"0.6167",
"0.634",
"0.6313",
"0.634",
"0.6313",
"0.643",
"0.6167",
"0.643"
],
[
"0.0801",
"0.643",
"0.468",
"0.643",
"0.468",
"0.657",
"0.0801",
"0.657"
],
[
"0.5186",
"0.646",
"0.5786",
"0.648",
"0.5767",
"0.662",
"0.517",
"0.66"
],
[
"0.574",
"0.646",
"0.8555",
"0.64",
"0.8555",
"0.664",
"0.574",
"0.6694"
],
[
"0.5947",
"0.64",
"0.613",
"0.643",
"0.6094",
"0.6494",
"0.5923",
"0.6465"
],
[
"0.0801",
"0.657",
"0.468",
"0.6577",
"0.468",
"0.6724",
"0.0801",
"0.672"
],
[
"0.6177",
"0.662",
"0.6357",
"0.6665",
"0.6313",
"0.675",
"0.614",
"0.6694"
],
[
"0.869",
"0.664",
"0.8906",
"0.664",
"0.8906",
"0.678",
"0.869",
"0.678"
],
[
"0.0801",
"0.674",
"0.468",
"0.674",
"0.468",
"0.6875",
"0.0801",
"0.6875"
],
[
"0.7686",
"0.677",
"0.835",
"0.677",
"0.835",
"0.6924",
"0.7686",
"0.6924"
],
[
"0.755",
"0.685",
"0.7607",
"0.688",
"0.755",
"0.6943",
"0.749",
"0.6914"
],
[
"0.7646",
"0.683",
"0.7715",
"0.683",
"0.7715",
"0.688",
"0.7646",
"0.688"
],
[
"0.734",
"0.6875",
"0.755",
"0.6875",
"0.755",
"0.696",
"0.734",
"0.696"
],
[
"0.0784",
"0.6875",
"0.4697",
"0.688",
"0.4697",
"0.7026",
"0.0784",
"0.702"
],
[
"0.7686",
"0.6943",
"0.835",
"0.6943",
"0.835",
"0.709",
"0.7686",
"0.709"
],
[
"0.0784",
"0.702",
"0.4714",
"0.704",
"0.4714",
"0.7183",
"0.0784",
"0.7163"
],
[
"0.0801",
"0.7183",
"0.2712",
"0.7183",
"0.2712",
"0.732",
"0.0801",
"0.732"
],
[
"0.523",
"0.7197",
"0.892",
"0.7197",
"0.892",
"0.733",
"0.523",
"0.733"
],
[
"0.0989",
"0.7344",
"0.468",
"0.7344",
"0.468",
"0.7476",
"0.0989",
"0.7476"
],
[
"0.5034",
"0.7373",
"0.8896",
"0.7373",
"0.8896",
"0.7476",
"0.5034",
"0.7476"
],
[
"0.0809",
"0.75",
"0.468",
"0.75",
"0.468",
"0.763",
"0.0809",
"0.763"
],
[
"0.5034",
"0.752",
"0.892",
"0.752",
"0.892",
"0.765",
"0.5034",
"0.765"
],
[
"0.0784",
"0.763",
"0.4697",
"0.762",
"0.4697",
"0.7764",
"0.0784",
"0.778"
],
[
"0.5024",
"0.765",
"0.8896",
"0.765",
"0.8896",
"0.779",
"0.5024",
"0.779"
],
[
"0.0801",
"0.779",
"0.468",
"0.779",
"0.468",
"0.7925",
"0.0801",
"0.7925"
],
[
"0.5034",
"0.782",
"0.8896",
"0.782",
"0.8896",
"0.7954",
"0.5034",
"0.7954"
],
[
"0.0809",
"0.7954",
"0.2018",
"0.7954",
"0.2018",
"0.806",
"0.0809",
"0.806"
],
[
"0.5034",
"0.7954",
"0.8906",
"0.7944",
"0.8906",
"0.808",
"0.5034",
"0.8096"
],
[
"0.505",
"0.8125",
"0.8896",
"0.8125",
"0.8896",
"0.8228",
"0.505",
"0.8228"
],
[
"0.505",
"0.8257",
"0.8906",
"0.8257",
"0.8906",
"0.8403",
"0.505",
"0.8403"
],
[
"0.0809",
"0.8335",
"0.2607",
"0.8335",
"0.2607",
"0.8467",
"0.0809",
"0.8467"
],
[
"0.505",
"0.8413",
"0.8364",
"0.8413",
"0.8364",
"0.856",
"0.505",
"0.856"
],
[
"0.09644",
"0.855",
"0.4714",
"0.856",
"0.4714",
"0.8706",
"0.09644",
"0.869"
],
[
"0.523",
"0.858",
"0.8906",
"0.858",
"0.8906",
"0.8687",
"0.523",
"0.8687"
],
[
"0.0809",
"0.872",
"0.468",
"0.872",
"0.468",
"0.886",
"0.0809",
"0.886"
],
[
"0.5024",
"0.872",
"0.892",
"0.872",
"0.892",
"0.886",
"0.5024",
"0.886"
],
[
"0.0801",
"0.887",
"0.4714",
"0.887",
"0.4714",
"0.901",
"0.0801",
"0.901"
],
[
"0.5073",
"0.887",
"0.892",
"0.887",
"0.892",
"0.901",
"0.5073",
"0.901"
]
] | [
"策略，该策略采用从模拟器获得的特权信息来控制驾",
"sparse goal locations specified in the GPS coordinates",
"驶过程inspiredfrom？］。专家的航路点可作为模仿",
"provided by a global planner. The routes encompass",
"损失的真实标签，使theexpert可与自动标记算法相",
"numerous scenarios, each initialized at predefined po-",
"媲美To accomplish lateral control,the expert policy",
"sitions, designed to evaluate the agent's capability to",
"follows the path generated by the A* planner, and a",
"adapt wide spectrum of adversarial situations and vary-",
"PID controller is used to minimize the angle of the",
"ingweatherconditions.第一个场景称为1WN，涉",
"vehicle towards the next waypoint in the route, which",
"及在除Town05之外的所有可用地图和路线集上训练",
"is at least 3.5 meters away. Meanwhile, longitudinal",
"模型，这些地图和路线集保留用于验证。Themodelis",
"control is performed using a version of model predictive",
"then evaluated on both Town05 short and long routes,",
"control, which differentiates between 3 target speeds.",
"consisting of 32 short and 10 long routes, to assess its",
"标准目标速度为4。0米/秒，但速度降低到3。当专家",
"performance。评估是在正常情况下在晴朗的中午条",
"在交叉路口内时，速度为0m/s。此外，如果预测到违",
"件下进行的，所有非玩家角色都遵守交通规则。在第",
"规，则目标速度changes为0。Om/s，使车辆停止。纵向",
"种场景1WA中，该模型在adversarial非玩家角色",
"和横向控制器分别使用K，=5.0,K，=0.5,Kd=1.0",
"（NPC）行为下进行了测试，这可能会导致碰撞，例如行",
"和K,=1.25,K,=0.75,Kd=0.3的PID值。对",
"人突然过马路或骑自行车的人出现。此外，交通灯管理",
"于两个控制器的积分项的运行平均值，我们使用大小",
"器可能会故意在交叉路口创建带有双绿灯的状态，以",
"为40的缓冲区。最初，allthreeRGB图像和深度图",
"模拟救护车或消防员可能跳过交通灯的紧急情况。。除",
"以480×960thencroppedto160×320的分辨率检",
"了正确驾驶自我车辆外，该模型有望在这些对抗性情",
"索，以避免失真。因此，allthreeRGB图像表示为",
"况下安全地做出反应totheenvironmentchangesand",
"R e 0,...,2553x160x320 attached to each other side.",
"noises and avoid accidents",
"by-side to create a single image. The same process is",
"4.3.Implementation Details",
"applied to the depth images to create one single depth",
"image. In order to calculate the real depth value for",
"To effectively gather knowledge across multiple",
"each pixel i in the depth map we use the following :",
"tasks using end-to-end learning, we predefined a set",
"2562B_i+256G_i+R_i",
"of distinctlossfunctions.The comprehensive loss that",
"Rdec-",
"×1000,(2)",
"encompasses all tasks can be calculated as follows :",
"2563-1",
"其 (Ri,Gi, B,) are the stored values of pixel i. Also,",
"LTOTAL=Q1LSEG+Q2LST+Q3LTH+",
"(3)",
"since the stored values are 8-bit, they have a maximum",
"Q4LBR+5LWP+6LTL+QLsS+QgLVE",
"of 255, and 1000 here is the maximum depth range of",
"the RGB-D camera in meters. All together they result",
"其中转向损失（XMATHX。此外，每个任务的损失权重",
"in Rdec which represents the true depth of pixel i.",
"由Qi,Q2,...,αg表示。For thepurposeof calculating",
"<NEX。语义分割的对象类是根据？1。此外，航路点",
"the semantic segmentation loss function （CsEG), a",
"用wpi=（ciyi）-1表示在BEV空间中。Thecenter",
"mixture of binary cross-entropy and dice loss is utilized",
"of the BEV space, marked by coordinates (O,O) in the",
"and can be computed through following equation.",
"local vehiclecoordinatesystem,ispositioned ontheego",
"N",
"vehicle itself, at the bottom-center point. The model",
"CSEG=",
"(Zyilog(9)+(1 - y:) log(1-9)",
"1",
"assesses vehicular controls within a normalized range",
":",
"(4)",
"spanning from 0 to 1, subsequently denormalizing them",
"2lg Uyl",
"",
"",
"71",
"to their original values, encompassing steering within",
"191+19/",
"the range of [-1,1],throttle ranging from [0,0.75] ,",
"and brake as either 0 or 1.",
"Here, N represents the number of pixels at the final",
"The traffic light state value changes to 1 if a red light",
"laver of our semantic segmentation decoder. Subse-",
"appeared, otherwise it is stated as 0. We encapsulate",
"quently, yi and ; correspond to the values of the i -th",
"the speed measurement (in m/s) and GPs locations",
"element in the ground truth vector y and the predic-",
"into a one-hot encoded vector as a high level naviga-",
"tionvector9，respectively.此方法允许我们同时利用",
"tional command.",
"基于分布和基于区域的损耗组件，如by？1所述。考",
"虑到整个网络的结构完整性依赖于它，强调用补充损",
"失标准增强语义分割任务的重要性势在必行。对于三",
"4.2. Task and Scenario",
"个预测航路点的情况，只有（CwP）需要平均，",
"The task entails navigating through a variety of",
"In order to adaptively adjust the loss weights for",
"areas such as highways, cityscapes, and residential",
"each training epoch, We utilize the MGN algorithm [?",
"neighborhoods along the predefined paths consisting of",
"。为了实现这一点，我们采用了Adam优化器，"
] | {
"arxiv_id": "2310.13135",
"language": "zh",
"page_idx": 4
} |
||
1803.03879-zh-1 | 1803.03879-zh-1.jpg | 这种免费分发可以被视为补充的外部知识，以过滤掉或降低与查询无关的提案的权重。例如，在图。1(c), 给定一个查询“一个人在踢足球”,预先训练的 VGG 网络可以通过预测一个提案是否对应于高概率的“人” 检测，为候选提案提供有用的提示
在语言一致性中使用外部知识是直截了当的；重构的特征可以通过检测概率进行修改。评估视觉一致性的任务更加困难；与语言一致性的直接类比是将视觉建议转换为单词并重建图像补丁。相反，我们建议从查询和视觉特征预测对象位置，以匹配短语接地的目标。如果没有外部知识的帮助，这个过程是不可能的，这些知识有助于关注可能的相关预测建议
在实现中，我们构建了一个新颖的知识辅助一致性网络(KACNet),该网络由两个分支组成：视觉一致性分支和语言一致性分支。这两个分支由一个共享的多模态子空间连接在一起，其中应用了注意力模型。为了利用来自视觉特征提取器的补充知识，我们提出了一种新颖的基于知识的池(KBP)门，以专注于与查询相关的视觉和语言重建建议
我们在两个基础数据集上评估 KACNet: Flickr30K 实体$\mathbb{E}[32]$ 和推荐游戏$\mathbb{E}[23]$ 。Flickr30K 实体包含超过 30K 图像和 170K 查询短语，而 ReferitGame 具有由 130K 查询短语引用的 19K 图像。在弱监督场景中， 我们在训练期间忽略边界框注释。实验表明，KACNet 在两个数据集上都大大优于最先进的方法，Flickr30K 实体上的%增加了9个以上，推荐游戏上的%增加了5个
我们的贡献是双重的：首先，我们利用互补的知识来过滤掉不相关的提案并提供直接指导。其次，我们提出了视觉一致性来提高接地性能。在下面的论文中，我们首先讨论 Sec 中的相关工作。2。有关 KACNet 的更多详细信息，请参见 Sec。3。最后，我们将 KACNet 与 Sec 中的其他方法进行分析和比较。4
## 2. Related Work
短语基础需要学习视觉和语言模态之间的相似性。Karpathy 等人。$\mathbb{E}[2\underline{2}]$首先在子空间中对齐句子片段和图像区域，然后在$\mathbb{E}[1]$ 中应用双向 RNN 进行多模态对齐。Hu 等人。$\mathbb{E}[21]$采用 2层 LSTM 根据编码的查询和视觉特征对提案进行排名。Rohrbach 等人。$\mathbb{E}$ [34]采用以查询为条件的潜在注意力网络，该网络在弱监督的情况下对提案进行排名。最近，Plummer 等人。[32] 增强了 CCA 模型 [31] ,以利用短语中广泛的语言提示。Chenetal。[4]在短语基础中引人回归机制，提高提案质量。Xiao 等人。[37] 利用查询的语言结构信息来指导弱监督场景下短语基础模型的学习。Chenetal。[5]应用强化学习技术来利用上下文信息。在本文中，我们探索了视觉和语言模式的一致性，并利用互补的知识来进一步提高弱监督基础的性能。
弱监督学习是一种旨在学习模型而无需大量手动标记工作的方法。它广泛应用于不同的计算机视觉任务。Crandall 等人。[7] 利用类标记来学习基于零件的空间模型，而无需详细注释对象位置和空间关系。Maxime
等人。[29] 建议纯粹从静止图像的动作标记中学习人与物体之间的交互。最近，Prest 等人。[33] 应用深度卷积神经网络及其分数图来解决具有图像级类别标签的对象定位问题。对于短语接地任务，Rohrbach 等人。[34]建议采用一种注意力模型，该模型通过学习来重建查询的信息而进行了优化，并且避免了人为标记训练集中每个查询的对象位置。基于此，肖等人。[37] 利用连续注意图探索语言情态的详细结构重构。受弱监督学习成功的启发，我们建议应用另一种视觉一致性来进一步提高性能
知识转移是一种广泛用于不同领域任务的技术。Hinton 等人。[19] 建议将从一个模型中学到的知识压缩到另一个模型中，这在计算上太昂贵了，无法训练。受此启发，Aytar 等人。[2] 应用视觉知识训练声音分类网络。Owens 等人。[30] 使用环境声音信息来训练对象检测网络。Lin 等人。[26] 利用在图像检索中的视觉问答 (VQA) 任务中学到的知识。Zhang 等人。[38] 应用在图像字幕和 VQA 中学习的知识来训练检测图像中视觉关系的网络。对于短语基础，我们建议利用从预先训练的深度神经网络中学到的知识来过滤不相关的视觉一致性建议
## 3. KAC Network
KACNet 由两个分支组成：视觉一致性分支和语言一致性分支，分别重构视觉和语言信息。这两个分支在共享的多模式子空间中连接，其中基于查询的语义应用注意力模型来关注所提到的对象。为了利用来自预先训练的 CNN 特征提取器的外部知识，提出了一个基于知识的池 (KBP) 门来选择与查询相关的建议。KACNet 经过端到端培训，具有视觉和语言一致性限制，以指导培训
我们首先介绍 KACNet 的框架，然后介绍 KBP 门的详细信息。然后我们说明如何应用 KBP 来促进视觉和语言一致性分支的优化。最后，提供了训练和推理的更多细节
# $3. 1. \textbf{ Framework}$
KACNet 的目标是在给定查询短语$q$和图像$x$的情况下本地化上述对象$y$ 。为了解决这个问题，通过对象提案生成系统生成了一组$N$提案$\{r_i\}$ 。然后，根据查询$q$的语义，将注意力模型应用于包含上述对象$y$ 的提案$r^q$
在弱监督的情况下，不提供查询$q$和提到的对象$y$ 的位置之间的映射。为了学习注意力模型，我们采用视觉和语言一致性，并分别构建两个分支。对于语言一致性，在给定由注意力模型预测的与查询相关的建议的情况下，应用重构模型来重构输入查询$q$ 。根据语言一致性，重构的查询应与输人一致。通过比较重建的查询和原始查询生成语言一致性损失$\mathcal{L}_{lc}$
为了视觉一致性，我们建议重建与查询相关的建议的视觉信息。由于短语接地的目标是预测提到的对象的位置，因此我们选择以输入查询为条件来预测候选提案的位置参数。与语言一致性类似，视觉一致性要求
2 | [
[
"0.0825",
"0.0947",
"0.4673",
"0.0947",
"0.4673",
"0.1048",
"0.0825",
"0.1048"
],
[
"0.5034",
"0.09216",
"0.8896",
"0.09216",
"0.8896",
"0.1061",
"0.5034",
"0.1061"
],
[
"0.0825",
"0.108",
"0.468",
"0.108",
"0.468",
"0.1218",
"0.0825",
"0.1218"
],
[
"0.505",
"0.108",
"0.8896",
"0.108",
"0.8896",
"0.1218",
"0.505",
"0.1218"
],
[
"0.0801",
"0.1225",
"0.468",
"0.1218",
"0.468",
"0.1351",
"0.0801",
"0.1364"
],
[
"0.5034",
"0.1225",
"0.8906",
"0.1225",
"0.8906",
"0.1364",
"0.5034",
"0.1364"
],
[
"0.0825",
"0.1383",
"0.468",
"0.1383",
"0.468",
"0.1515",
"0.0825",
"0.1515"
],
[
"0.5034",
"0.1383",
"0.8906",
"0.1383",
"0.8906",
"0.1515",
"0.5034",
"0.1515"
],
[
"0.0825",
"0.1547",
"0.326",
"0.1547",
"0.326",
"0.1654",
"0.0825",
"0.1654"
],
[
"0.5063",
"0.1547",
"0.8896",
"0.1547",
"0.8896",
"0.1654",
"0.5063",
"0.1654"
],
[
"0.10046",
"0.1686",
"0.468",
"0.1686",
"0.468",
"0.1818",
"0.10046",
"0.1818"
],
[
"0.5034",
"0.1666",
"0.8906",
"0.1674",
"0.8906",
"0.1818",
"0.5034",
"0.1812"
],
[
"0.0825",
"0.1831",
"0.468",
"0.1831",
"0.468",
"0.1963",
"0.0825",
"0.1963"
],
[
"0.5034",
"0.1831",
"0.8896",
"0.1831",
"0.8896",
"0.1963",
"0.5034",
"0.1963"
],
[
"0.0809",
"0.1989",
"0.468",
"0.1989",
"0.468",
"0.2122",
"0.0809",
"0.2122"
],
[
"0.5034",
"0.1989",
"0.8906",
"0.1989",
"0.8906",
"0.2122",
"0.5034",
"0.2122"
],
[
"0.0825",
"0.2153",
"0.4673",
"0.2153",
"0.4673",
"0.226",
"0.0825",
"0.226"
],
[
"0.505",
"0.2134",
"0.892",
"0.2134",
"0.892",
"0.2267",
"0.505",
"0.2267"
],
[
"0.0809",
"0.2291",
"0.4673",
"0.2291",
"0.4673",
"0.2424",
"0.0809",
"0.2424"
],
[
"0.5024",
"0.2267",
"0.6357",
"0.2279",
"0.6357",
"0.2424",
"0.5024",
"0.2412"
],
[
"0.0825",
"0.2437",
"0.468",
"0.2437",
"0.468",
"0.2568",
"0.0825",
"0.2568"
],
[
"0.523",
"0.2437",
"0.8896",
"0.2437",
"0.8896",
"0.2568",
"0.523",
"0.2568"
],
[
"0.0809",
"0.2588",
"0.3423",
"0.2588",
"0.3423",
"0.2727",
"0.0809",
"0.2727"
],
[
"0.505",
"0.2588",
"0.8906",
"0.2588",
"0.8906",
"0.2727",
"0.505",
"0.2727"
],
[
"0.1013",
"0.276",
"0.4673",
"0.276",
"0.4673",
"0.286",
"0.1013",
"0.286"
],
[
"0.5034",
"0.2727",
"0.8906",
"0.2734",
"0.8906",
"0.2886",
"0.5034",
"0.2874"
],
[
"0.0809",
"0.289",
"0.468",
"0.289",
"0.468",
"0.303",
"0.0809",
"0.303"
],
[
"0.505",
"0.289",
"0.892",
"0.289",
"0.892",
"0.303",
"0.505",
"0.303"
],
[
"0.0825",
"0.3037",
"0.468",
"0.3037",
"0.468",
"0.3176",
"0.0825",
"0.3176"
],
[
"0.505",
"0.3037",
"0.8906",
"0.3037",
"0.8906",
"0.3176",
"0.505",
"0.3176"
],
[
"0.0825",
"0.3193",
"0.4673",
"0.3193",
"0.4673",
"0.3333",
"0.0825",
"0.3333"
],
[
"0.505",
"0.3193",
"0.8906",
"0.3193",
"0.8906",
"0.3333",
"0.505",
"0.3333"
],
[
"0.08417",
"0.336",
"0.464",
"0.336",
"0.464",
"0.3467",
"0.08417",
"0.3467"
],
[
"0.505",
"0.334",
"0.888",
"0.334",
"0.888",
"0.348",
"0.505",
"0.348"
],
[
"0.09235",
"0.3499",
"0.468",
"0.3499",
"0.468",
"0.363",
"0.09235",
"0.363"
],
[
"0.505",
"0.351",
"0.886",
"0.351",
"0.886",
"0.361",
"0.505",
"0.361"
],
[
"0.0809",
"0.3643",
"0.2769",
"0.3643",
"0.2769",
"0.3774",
"0.0809",
"0.3774"
],
[
"0.505",
"0.363",
"0.8896",
"0.363",
"0.8896",
"0.377",
"0.505",
"0.377"
],
[
"0.1013",
"0.379",
"0.468",
"0.379",
"0.468",
"0.3926",
"0.1013",
"0.3926"
],
[
"0.5034",
"0.3801",
"0.8906",
"0.3801",
"0.8906",
"0.3933",
"0.5034",
"0.3933"
],
[
"0.0809",
"0.3945",
"0.468",
"0.3945",
"0.468",
"0.4077",
"0.0809",
"0.4077"
],
[
"0.5024",
"0.3933",
"0.6323",
"0.3933",
"0.6323",
"0.4072",
"0.5024",
"0.4072"
],
[
"0.0809",
"0.4104",
"0.4673",
"0.4104",
"0.4673",
"0.4236",
"0.0809",
"0.4236"
],
[
"0.08417",
"0.4268",
"0.4673",
"0.4268",
"0.4673",
"0.4375",
"0.08417",
"0.4375"
],
[
"0.5024",
"0.4224",
"0.653",
"0.4224",
"0.653",
"0.4363",
"0.5024",
"0.4363"
],
[
"0.0809",
"0.44",
"0.468",
"0.44",
"0.468",
"0.4539",
"0.0809",
"0.4539"
],
[
"0.5215",
"0.4463",
"0.8896",
"0.4475",
"0.8896",
"0.4622",
"0.5215",
"0.461"
],
[
"0.0801",
"0.455",
"0.468",
"0.4539",
"0.468",
"0.4673",
"0.0801",
"0.4685"
],
[
"0.5063",
"0.464",
"0.523",
"0.464",
"0.523",
"0.473",
"0.5063",
"0.473"
],
[
"0.5156",
"0.461",
"0.8906",
"0.4634",
"0.8906",
"0.4766",
"0.5156",
"0.4749"
],
[
"0.0809",
"0.4702",
"0.464",
"0.4702",
"0.464",
"0.4841",
"0.0809",
"0.4841"
],
[
"0.505",
"0.4778",
"0.8896",
"0.4778",
"0.8896",
"0.4912",
"0.505",
"0.4912"
],
[
"0.0801",
"0.486",
"0.1095",
"0.486",
"0.1095",
"0.4976",
"0.0801",
"0.4976"
],
[
"0.505",
"0.4937",
"0.8896",
"0.4937",
"0.8896",
"0.507",
"0.505",
"0.507"
],
[
"0.0989",
"0.5",
"0.468",
"0.5005",
"0.468",
"0.5146",
"0.0989",
"0.513"
],
[
"0.5034",
"0.5083",
"0.8784",
"0.5083",
"0.8784",
"0.5215",
"0.5034",
"0.5215"
],
[
"0.0809",
"0.5166",
"0.4673",
"0.5166",
"0.4673",
"0.53",
"0.0809",
"0.53"
],
[
"0.5024",
"0.5234",
"0.886",
"0.5234",
"0.886",
"0.537",
"0.5024",
"0.537"
],
[
"0.0866",
"0.5327",
"0.4673",
"0.5327",
"0.4673",
"0.5435",
"0.0866",
"0.5435"
],
[
"0.5024",
"0.537",
"0.8906",
"0.539",
"0.8906",
"0.553",
"0.5024",
"0.5503"
],
[
"0.0825",
"0.5454",
"0.468",
"0.5444",
"0.468",
"0.5596",
"0.0825",
"0.56"
],
[
"0.505",
"0.5537",
"0.6177",
"0.5537",
"0.6177",
"0.5674",
"0.505",
"0.5674"
],
[
"0.0809",
"0.56",
"0.4673",
"0.5596",
"0.4673",
"0.5737",
"0.0809",
"0.5747"
],
[
"0.523",
"0.5684",
"0.8906",
"0.5684",
"0.8906",
"0.582",
"0.523",
"0.582"
],
[
"0.0809",
"0.5747",
"0.38",
"0.577",
"0.38",
"0.5903",
"0.0809",
"0.5884"
],
[
"0.5063",
"0.5854",
"0.888",
"0.5854",
"0.888",
"0.595",
"0.5063",
"0.595"
],
[
"0.505",
"0.5977",
"0.8906",
"0.5977",
"0.8906",
"0.6113",
"0.505",
"0.6113"
],
[
"0.0784",
"0.605",
"0.2279",
"0.604",
"0.2279",
"0.6187",
"0.0784",
"0.619"
],
[
"0.505",
"0.6157",
"0.5703",
"0.6157",
"0.5703",
"0.627",
"0.505",
"0.627"
],
[
"0.1013",
"0.632",
"0.4656",
"0.632",
"0.4656",
"0.6426",
"0.1013",
"0.6426"
],
[
"0.5034",
"0.637",
"0.6323",
"0.637",
"0.6323",
"0.651",
"0.5034",
"0.651"
],
[
"0.0809",
"0.646",
"0.4673",
"0.646",
"0.4673",
"0.659",
"0.0809",
"0.659"
],
[
"0.0825",
"0.66",
"0.468",
"0.66",
"0.468",
"0.674",
"0.0825",
"0.674"
],
[
"0.523",
"0.66",
"0.8906",
"0.66",
"0.8906",
"0.674",
"0.523",
"0.674"
],
[
"0.0809",
"0.6753",
"0.468",
"0.6753",
"0.468",
"0.6895",
"0.0809",
"0.6895"
],
[
"0.5034",
"0.675",
"0.8896",
"0.6753",
"0.8896",
"0.69",
"0.5034",
"0.6895"
],
[
"0.0809",
"0.69",
"0.468",
"0.6895",
"0.468",
"0.7026",
"0.0809",
"0.704"
],
[
"0.5034",
"0.6914",
"0.8906",
"0.6914",
"0.8906",
"0.705",
"0.5034",
"0.705"
],
[
"0.0825",
"0.7056",
"0.468",
"0.7056",
"0.468",
"0.7197",
"0.0825",
"0.7197"
],
[
"0.5024",
"0.705",
"0.8906",
"0.7056",
"0.8906",
"0.72",
"0.5024",
"0.7197"
],
[
"0.0809",
"0.72",
"0.468",
"0.72",
"0.468",
"0.7344",
"0.0809",
"0.7344"
],
[
"0.505",
"0.7217",
"0.5728",
"0.7217",
"0.5728",
"0.733",
"0.505",
"0.733"
],
[
"0.0825",
"0.7363",
"0.4631",
"0.7363",
"0.4631",
"0.75",
"0.0825",
"0.75"
],
[
"0.523",
"0.735",
"0.892",
"0.7373",
"0.892",
"0.7505",
"0.523",
"0.7485"
],
[
"0.0825",
"0.7505",
"0.468",
"0.7505",
"0.468",
"0.765",
"0.0825",
"0.765"
],
[
"0.5034",
"0.752",
"0.8906",
"0.752",
"0.8906",
"0.765",
"0.5034",
"0.765"
],
[
"0.0825",
"0.7666",
"0.468",
"0.7666",
"0.468",
"0.78",
"0.0825",
"0.78"
],
[
"0.505",
"0.768",
"0.8896",
"0.768",
"0.8896",
"0.779",
"0.505",
"0.779"
],
[
"0.08496",
"0.7827",
"0.468",
"0.7827",
"0.468",
"0.7935",
"0.08496",
"0.7935"
],
[
"0.5024",
"0.781",
"0.8896",
"0.782",
"0.8896",
"0.7954",
"0.5024",
"0.7944"
],
[
"0.0801",
"0.7944",
"0.4697",
"0.7954",
"0.4697",
"0.8096",
"0.0801",
"0.808"
],
[
"0.5034",
"0.7974",
"0.8896",
"0.7974",
"0.8896",
"0.811",
"0.5034",
"0.811"
],
[
"0.0809",
"0.811",
"0.4673",
"0.811",
"0.4673",
"0.8247",
"0.0809",
"0.8247"
],
[
"0.517",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.517",
"0.8247"
],
[
"0.0825",
"0.827",
"0.4102",
"0.827",
"0.4102",
"0.8403",
"0.0825",
"0.8403"
],
[
"0.5034",
"0.827",
"0.773",
"0.827",
"0.773",
"0.8403",
"0.5034",
"0.8403"
],
[
"0.1013",
"0.8413",
"0.4673",
"0.8413",
"0.4673",
"0.855",
"0.1013",
"0.855"
],
[
"0.5254",
"0.8433",
"0.888",
"0.8433",
"0.888",
"0.854",
"0.5254",
"0.854"
],
[
"0.0825",
"0.857",
"0.464",
"0.857",
"0.464",
"0.8706",
"0.0825",
"0.8706"
],
[
"0.505",
"0.857",
"0.8896",
"0.857",
"0.8896",
"0.8706",
"0.505",
"0.8706"
],
[
"0.0825",
"0.872",
"0.468",
"0.872",
"0.468",
"0.8853",
"0.0825",
"0.8853"
],
[
"0.505",
"0.872",
"0.8896",
"0.872",
"0.8896",
"0.8853",
"0.505",
"0.8853"
],
[
"0.0809",
"0.887",
"0.4673",
"0.887",
"0.4673",
"0.901",
"0.0809",
"0.901"
],
[
"0.5034",
"0.887",
"0.8906",
"0.887",
"0.8906",
"0.901",
"0.5034",
"0.901"
],
[
"0.4805",
"0.925",
"0.492",
"0.925",
"0.492",
"0.9375",
"0.4805",
"0.9375"
]
] | [
"这种免费分发可以被视为补充的外部知识，以过滤掉",
"等人。[29]建议纯粹从静止图像的动作标记中学习人",
"或降低与查询无关的提案的权重。例如，在图。1(c)，",
"与物体之间的交互。最近，Prest等人。[33]应用深度",
"给定一个查询“二个人在踢足球”，预先训练的VGG",
"卷积神经网络及其分数图来解决具有图像级类别标签",
"网络可以通过预测一个提案是否对应于高概率的“人”",
"的对象定位问题。对于短语接地任务，Rohrbach等人。",
"检测，为候选提案提供有用的提示",
"341建议采用一种注意力模型，该模型通过学习来重建",
"在语言一致性中使用外部知识是直截了当的；重构",
"查询的信息而进行了优化，并且避免了人为标记训练",
"的特征可以通过检测概率进行修改。评估视觉一致性",
"集中每个查询的对象位置。基于此，肖等人。[37]利",
"的任务更加困难；与语言一致性的直接类比是将视觉",
"用连续注意图探索语言情态的详细结构重构。受弱监",
"建议转换为单词并重建图像补丁。相反，我们建议从查",
"督学习成功的启发，我们建议应用另一种视觉一致性",
"询和视觉特征预测对象位置，以匹配短语接地的目标。",
"来进一步提高性能",
"如果没有外部知识的帮助，这个过程是不可能的，这些",
"知识转移是一种广泛用于不同领域任务的技术。Hin-",
"知识有助于关注可能的相关预测建议",
"ton等人。【19]建议将从一个模型中学到的知识压缩",
"在实现中，我们构建了一个新颖的知识辅助一致性",
"到另一个模型中，这在计算上太昂贵了，无法训练。受",
"网络（KACNet），该网络由两个分支组成：视觉一致",
"此启发，Aytar等人。[2]应用视觉知识训练声音分类",
"性分支和语言一致性分支。这两个分支由一个共享的",
"网络。Owens等人。[30]使用环境声音信息来训练对",
"多模态子空间连接在一起，其中应用了注意力模型。为",
"象检测网络。Lin等人。[26]利用在图像检索中的视",
"了利用来自视觉特征提取器的补充知识，我们提出了",
"觉问答（VQA）任务中学到的知识。Zhang等人。【38",
"-种新颖的基于知识的池（KBP）门，以专注于与查询",
"应用在图像字幕和VQA中学习的知识来训练检测图",
"相关的视觉和语言重建建议",
"像中视觉关系的网络。对于短语基础，我们建议利用从",
"我们在两个基础数据集上评估KACNet：Flickr30K",
"预先训练的深度神经网络中学到的知识来过滤不相关",
"实体E[32]和推荐游戏E[23]。Flickr30K实体包含超",
"的视觉一致性建议",
"过30K图像和170K查询短语，而ReferitGame具有",
"由130K查询短语引用的19K图像。在弱监督场景中，",
"3. KAC Network",
"我们在训练期间忽略边界框注释。实验表明，KACNet",
"KACNet由两个分支组成：视觉一致性分支和语言",
"在两个数据集上都大大优于最先进的方法，Flickr30K",
"一",
"致性分支，分别重构视觉和语言信息。这两个分支",
"实体上的%增加了9个以上，推荐游戏上的%增加了",
"在共享的多模式子空间中连接，其中基于查询的语义",
"5个",
"应用注意力模型来关注所提到的对象。为了利用来自",
"我们的贡献是双重的：首先，我们利用互补的知识来",
"预先训练的CNN特征提取器的外部知识，提出了",
"过滤掉不相关的提案并提供直接指导。其次，我们提出",
"个基于知识的池（KBP）门来选择与查询相关的建议",
"了视觉一致性来提高接地性能。在下面的论文中，我们",
"KACNet经过端到端培训，具有视觉和语言一致性限",
"首先讨论Sec中的相关工作。2。有关KACNet的更",
"制，以指导培训",
"多详细信息，请参见Sec。3。最后，我们将KACNet",
"我们首先介绍KACNet的框架，然后介绍KBP门",
"与Sec中的其他方法进行分析和比较。4",
"的详细信息。然后我们说明如何应用KBP来促进视觉",
"和语言一致性分支的优化。最后，提供了训练和推理的",
"2. Related Work",
"更多细节",
"短语基础需要学习视觉和语言模态之间的相似性。",
"3.1. Framework",
"Karpathy等人。E[22]首先在子空间中对齐句子片段",
"和图像区域，然后在E[II中应用双向RNN进行多模",
"KACNet的目标是在给定查询短语q和图像r的情",
"态对齐。Hu等人。E[21] 采用2层LSTM根据编码的",
"况下本地化上述对象9。为了解决这个问题，通过对",
"查询和视觉特征对提案进行排名。Rohrbach等人。E",
"象提案生成系统生成了一组N提案r}。然后，根据",
"[34]采用以查询为条件的潜在注意力网络，该网络在",
"查询q的语义，将注意力模型应用于包含上述对象y",
"弱监督的情况下对提案进行排名。最近，Plummer等",
"的提案rq",
"人。［32]增强了CCA模型[31]，以利用短语中广",
"在弱监督的情况下，不提供查询q和提到的对象y",
"泛的语言提示。Chenetal。【4]在短语基础中引人回归",
"的位置之间的映射。为了学习注意力模型，我们采用视",
"机制，提高提案质量。Xiao等人。[37]利用查询的语",
"觉和语言一致性，并分别构建两个分支。对于语言一致",
"言结构信息来指导弱监督场景下短语基础模型的学习。",
"性，在给定由注意力模型预测的与查询相关的建议的",
"Chenetal。[5]应用强化学习技术来利用上下文信息。",
"情况下，应用重构模型来重构输人查询q。根据语言",
"在本文中，我们探索了视觉和语言模式的一致性，并利",
"致性，重构的查询应与输人一致。通过比较重建的查",
"用互补的知识来进一步提高弱监督基础的性能",
"询和原始查询生成语言一致性损失Ll",
"弱监督学习是一种旨在学习模型而无需大量手动标",
"为了视觉一致性，我们建议重建与查询相关的建议",
"记工作的方法。它广泛应用于不同的计算机视觉任务",
"的视觉信息。由于短语接地的目标是预测提到的对象",
"Crandall等人。[7]利用类标记来学习基于零件的空间",
"的位置，因此我们选择以输人查询为条件来预测候选",
"模型，而无需详细注释对象位置和空间关系。Maxime",
"提案的位置参数。与语言一致性类似，视觉一致性要求",
"2"
] | {
"arxiv_id": "1803.03879",
"language": "zh",
"page_idx": 1
} |
||
1803.03879-zh-2 | 1803.03879-zh-2.jpg | <FigureHere>
Figure 2. Knowledge Aided Consistency Network (KAC Net) consists of a visual consistency branch and a language consistency branch. Visual consistency branch aims at predicting and aligning query-related proposals' location parameters conditioned on the input query. Language consistency branch attempts to reconstruct input query from query-related proposals. To provide guidance in training and testing, a Knowledge Based Pooling (KBP) gate is applied to filter out unrelated proposals for both branches.
为了对这些知识进行编码，我们首先解析语言查询并通过自然语言处理(NLP)解析器检索所有名词词。对于每个提案的分布$p_i$ ,我们选择概率最高的最可能的类。然后将建议$r_i$的知识$k_i^q$计算为该类的名称与查询中的名词单词之间的单词相似度(图。3)。如果查询包含多个名词词，我们将所有计算出的相似性平均为知识$k_i^q$,可以写成
<FigureHere>
$$k_i^q=\dfrac{1}{N_q}\sum_{j=1}^{N_q}\sin(C_i^*,w_j^q)$$
(2)
Figure 3. A pre-trained CNN always predicts a probability distribution $\hat{for its own task. We leverage the most prob- }$ able category predicted by CNN and calculate the word similarity between noun words in the query as knowledge $k_{i}^{q}$
其中$C_i^*$是建议 $r_i$ 的预测类名$,w_j^q$ 是查询 $q$ 中所有$N_q$名词单词的第$j$个单词。$\sin$是 一 个 测 量 两 个 单 词 之间相似性的函数
在训练阶段，知识$k_i^q$充当“池化”门，有助于视觉(秒。3.3 )和语言(秒。3.4 )一致性分支选择和重建可靠的候选提案。在测试阶段，知识$k_i^q$过滤掉不相关的提案，并增加了找到包含所提及对象的提案的机会(Sec。3.5)
预测参数应恢复每个提案的位置。在此基础上，通过计算预测和原始建议的位置参数之间的差异，产生了视觉一致性损失$\mathcal{L}_{vc}$ 。为了利用丰富的图像功能和可用的固定类别分类器，我们应用 KBP 对 CNN 提供的知识进行编码，并对每个提案在视觉和语言一致性方面的重要性进行加权。KACNet 的目标可以写成
# 3.3. Visual Consistency
(1)
$$\arg\min_{\theta}\sum_{q}(\mathcal{L}_{lc}^{k}+\lambda\mathcal{L}_{vc}^{k})+\mu\mathcal{L}_{reg}$$
,其中$\theta$表示要优化的参数。$\mathcal{L}_lc^k$是来自语言一致性分支的重建损失，$\mathcal{L}_vc^k$是来自视觉一致性分支的重建损失(上标“$k$”是指 KBP)。$\mathcal{L}_reg$ 是一个权重正则化项。$\lambda$ 、$\mu$是超参数
视觉一致性的目标是通过学习来优化注意力模型， 以预测与查询相关的建议中包含的位置信息。通过预测以输入查询为条件的位置信息，我们期望学习语言和视觉模态之间更好的相关性。在弱监督场景中，没有注释可用于指示与查询相关的提案的标识。相反，我们使用 KBP 的知识$k_i^q$在培训期间提供指导。当建议$r_i$ 与查询相关时，我们期望知识$k_i^q$提供更高的分数。因此，可以将 KBP 应用于根据查询$q$自适应地加权每个提案的视觉一致性损失。在实现中，我们首先应用长短期内存 (LSTM)$\operatorname{E}[20]$模型将输入查询$q$编码为嵌入向量$q\in\mathbb{R}^d_q$ 。采用预先训练的 CNN 为每个建议$r_i$ 提取视觉特征$\boldsymbol{v}_i\in\mathbb{R}^d_v$,为输入图像$x$提取全局视觉特征$v\in\mathbb{R}^{d_v}$ 。然后，注意力模型将嵌人向量$q$ ,图
## 3.2. Knowledge Based Pooling (KBP)
我们应用预训练的 CNN 为提案$r_i$提取视觉特征$v_i$ ,并为其自己的任务预测概率分布$p_i$ ,这提供了有用的线索来过滤掉不相关的提案
3 | [
[
"0.8164",
"0.09344",
"0.853",
"0.09344",
"0.853",
"0.1042",
"0.8164",
"0.1042"
],
[
"0.3823",
"0.1016",
"0.4412",
"0.1016",
"0.4412",
"0.1092",
"0.3823",
"0.1092"
],
[
"0.4395",
"0.1042",
"0.452",
"0.1042",
"0.452",
"0.108",
"0.4395",
"0.108"
],
[
"0.7935",
"0.1029",
"0.8716",
"0.1029",
"0.8716",
"0.113",
"0.7935",
"0.113"
],
[
"0.3945",
"0.1092",
"0.4453",
"0.1092",
"0.4453",
"0.1193",
"0.3945",
"0.1193"
],
[
"0.6406",
"0.127",
"0.739",
"0.1256",
"0.739",
"0.137",
"0.6406",
"0.1383"
],
[
"0.486",
"0.1383",
"0.5",
"0.1383",
"0.5",
"0.1414",
"0.486",
"0.1414"
],
[
"0.4983",
"0.1383",
"0.523",
"0.1383",
"0.523",
"0.1459",
"0.4983",
"0.1459"
],
[
"0.6436",
"0.137",
"0.7427",
"0.137",
"0.7427",
"0.1477",
"0.6436",
"0.1477"
],
[
"0.6406",
"0.1465",
"0.7563",
"0.1459",
"0.7563",
"0.1572",
"0.6406",
"0.1578"
],
[
"0.2646",
"0.1572",
"0.308",
"0.1572",
"0.308",
"0.1686",
"0.2646",
"0.1686"
],
[
"0.4395",
"0.161",
"0.4558",
"0.161",
"0.4558",
"0.1654",
"0.4395",
"0.1654"
],
[
"0.8105",
"0.159",
"0.8555",
"0.159",
"0.8555",
"0.1692",
"0.8105",
"0.1692"
],
[
"0.2656",
"0.1674",
"0.3064",
"0.1674",
"0.3064",
"0.1749",
"0.2656",
"0.1749"
],
[
"0.4412",
"0.1674",
"0.46",
"0.1674",
"0.46",
"0.1749",
"0.4412",
"0.1749"
],
[
"0.6436",
"0.1724",
"0.74",
"0.1724",
"0.74",
"0.1831",
"0.6436",
"0.1831"
],
[
"0.814",
"0.1749",
"0.839",
"0.1749",
"0.839",
"0.178",
"0.814",
"0.178"
],
[
"0.6436",
"0.1818",
"0.7485",
"0.1818",
"0.7485",
"0.1925",
"0.6436",
"0.1925"
],
[
"0.3896",
"0.1895",
"0.4126",
"0.1895",
"0.4126",
"0.1976",
"0.3896",
"0.1976"
],
[
"0.1274",
"0.1957",
"0.2197",
"0.1963",
"0.2197",
"0.2084",
"0.1274",
"0.207"
],
[
"0.6436",
"0.1913",
"0.7485",
"0.1913",
"0.7485",
"0.202",
"0.6436",
"0.202"
],
[
"0.3882",
"0.1995",
"0.411",
"0.1995",
"0.411",
"0.2039",
"0.3882",
"0.2039"
],
[
"0.8096",
"0.2039",
"0.851",
"0.2039",
"0.851",
"0.2184",
"0.8096",
"0.2184"
],
[
"0.614",
"0.2134",
"0.7427",
"0.2134",
"0.7427",
"0.2267",
"0.614",
"0.2267"
],
[
"0.4387",
"0.2299",
"0.486",
"0.2299",
"0.486",
"0.2405",
"0.4387",
"0.2405"
],
[
"0.6055",
"0.2279",
"0.749",
"0.2279",
"0.749",
"0.238",
"0.6055",
"0.238"
],
[
"0.1152",
"0.238",
"0.232",
"0.238",
"0.232",
"0.252",
"0.1152",
"0.252"
],
[
"0.4478",
"0.2405",
"0.4763",
"0.2405",
"0.4763",
"0.2488",
"0.4478",
"0.2488"
],
[
"0.5073",
"0.2424",
"0.528",
"0.2424",
"0.528",
"0.2507",
"0.5073",
"0.2507"
],
[
"0.607",
"0.2412",
"0.749",
"0.2412",
"0.749",
"0.252",
"0.607",
"0.252"
],
[
"0.10297",
"0.25",
"0.246",
"0.2507",
"0.246",
"0.2651",
"0.10297",
"0.2644"
],
[
"0.3186",
"0.2443",
"0.3489",
"0.2443",
"0.3489",
"0.2524",
"0.3186",
"0.2524"
],
[
"0.2656",
"0.2507",
"0.3022",
"0.254",
"0.3015",
"0.262",
"0.2646",
"0.2588"
],
[
"0.3628",
"0.254",
"0.3906",
"0.254",
"0.3906",
"0.2583",
"0.3628",
"0.2583"
],
[
"0.5073",
"0.252",
"0.528",
"0.252",
"0.528",
"0.2588",
"0.5073",
"0.2588"
],
[
"0.2656",
"0.2588",
"0.3022",
"0.2588",
"0.3022",
"0.2695",
"0.2656",
"0.2695"
],
[
"0.3718",
"0.26",
"0.3987",
"0.26",
"0.3987",
"0.2676",
"0.3718",
"0.2676"
],
[
"0.8057",
"0.262",
"0.857",
"0.2644",
"0.857",
"0.276",
"0.8057",
"0.2734"
],
[
"0.7935",
"0.2734",
"0.869",
"0.2734",
"0.869",
"0.281",
"0.7935",
"0.281"
],
[
"0.1217",
"0.2651",
"0.2263",
"0.2651",
"0.2263",
"0.276",
"0.1217",
"0.276"
],
[
"0.0809",
"0.2935",
"0.1602",
"0.2935",
"0.1602",
"0.3037",
"0.0809",
"0.3037"
],
[
"0.152",
"0.2903",
"0.892",
"0.2917",
"0.892",
"0.3062",
"0.152",
"0.305"
],
[
"0.0784",
"0.3062",
"0.892",
"0.3062",
"0.892",
"0.3193",
"0.0784",
"0.3193"
],
[
"0.0801",
"0.3193",
"0.892",
"0.3193",
"0.892",
"0.3333",
"0.0801",
"0.3333"
],
[
"0.0801",
"0.334",
"0.892",
"0.334",
"0.892",
"0.348",
"0.0801",
"0.348"
],
[
"0.0801",
"0.348",
"0.335",
"0.348",
"0.335",
"0.361",
"0.0801",
"0.361"
],
[
"0.21",
"0.3894",
"0.46",
"0.3914",
"0.46",
"0.4092",
"0.21",
"0.4072"
],
[
"0.523",
"0.385",
"0.8906",
"0.3865",
"0.8906",
"0.401",
"0.523",
"0.3997"
],
[
"0.5034",
"0.4016",
"0.8906",
"0.4016",
"0.8906",
"0.4153",
"0.5034",
"0.4153"
],
[
"0.2141",
"0.4124",
"0.2263",
"0.4124",
"0.2263",
"0.4204",
"0.2141",
"0.4204"
],
[
"0.2656",
"0.4077",
"0.46",
"0.4077",
"0.46",
"0.4216",
"0.2656",
"0.4216"
],
[
"0.5034",
"0.4172",
"0.8906",
"0.4172",
"0.8906",
"0.4312",
"0.5034",
"0.4312"
],
[
"0.335",
"0.4268",
"0.3406",
"0.4268",
"0.3406",
"0.433",
"0.335",
"0.433"
],
[
"0.505",
"0.432",
"0.8906",
"0.432",
"0.8906",
"0.4458",
"0.505",
"0.4458"
],
[
"0.2484",
"0.4395",
"0.423",
"0.4395",
"0.423",
"0.4495",
"0.2484",
"0.4495"
],
[
"0.505",
"0.4475",
"0.8896",
"0.4475",
"0.8896",
"0.461",
"0.505",
"0.461"
],
[
"0.25",
"0.4602",
"0.3865",
"0.4602",
"0.3865",
"0.4734",
"0.25",
"0.4734"
],
[
"0.5024",
"0.461",
"0.8906",
"0.4602",
"0.8906",
"0.4749",
"0.5024",
"0.476"
],
[
"0.5034",
"0.4766",
"0.6787",
"0.4766",
"0.6787",
"0.4912",
"0.5034",
"0.4912"
],
[
"0.2141",
"0.4873",
"0.451",
"0.488",
"0.451",
"0.5024",
"0.2141",
"0.502"
],
[
"0.6807",
"0.5005",
"0.699",
"0.507",
"0.694",
"0.518",
"0.675",
"0.512"
],
[
"0.0801",
"0.511",
"0.468",
"0.511",
"0.468",
"0.5244",
"0.0801",
"0.5244"
],
[
"0.6562",
"0.5103",
"0.6665",
"0.512",
"0.664",
"0.521",
"0.653",
"0.518"
],
[
"0.6055",
"0.5176",
"0.6396",
"0.5146",
"0.642",
"0.529",
"0.609",
"0.532"
],
[
"0.6914",
"0.5195",
"0.7817",
"0.5195",
"0.7817",
"0.531",
"0.6914",
"0.531"
],
[
"0.8667",
"0.5166",
"0.8936",
"0.5166",
"0.8936",
"0.5327",
"0.8667",
"0.5327"
],
[
"0.0801",
"0.5244",
"0.4673",
"0.5244",
"0.4673",
"0.5386",
"0.0801",
"0.5386"
],
[
"0.648",
"0.5215",
"0.701",
"0.5327",
"0.6953",
"0.55",
"0.642",
"0.537"
],
[
"0.0801",
"0.5386",
"0.4697",
"0.5386",
"0.4697",
"0.553",
"0.0801",
"0.553"
],
[
"0.0801",
"0.553",
"0.4697",
"0.553",
"0.4697",
"0.5664",
"0.0801",
"0.5664"
],
[
"0.517",
"0.553",
"0.8896",
"0.553",
"0.8896",
"0.5664",
"0.517",
"0.5664"
],
[
"0.0801",
"0.5664",
"0.0989",
"0.5664",
"0.0989",
"0.579",
"0.0801",
"0.579"
],
[
"0.505",
"0.5674",
"0.892",
"0.5674",
"0.892",
"0.5806",
"0.505",
"0.5806"
],
[
"0.505",
"0.5835",
"0.6357",
"0.5835",
"0.6357",
"0.5967",
"0.505",
"0.5967"
],
[
"0.5215",
"0.5967",
"0.892",
"0.5977",
"0.892",
"0.6123",
"0.5215",
"0.6113"
],
[
"0.0809",
"0.608",
"0.4673",
"0.608",
"0.4673",
"0.622",
"0.0809",
"0.622"
],
[
"0.5024",
"0.613",
"0.892",
"0.613",
"0.892",
"0.627",
"0.5024",
"0.627"
],
[
"0.0809",
"0.6226",
"0.468",
"0.6226",
"0.468",
"0.637",
"0.0809",
"0.637"
],
[
"0.5034",
"0.627",
"0.8906",
"0.6274",
"0.8906",
"0.6426",
"0.5034",
"0.6416"
],
[
"0.0801",
"0.638",
"0.468",
"0.637",
"0.468",
"0.6514",
"0.0801",
"0.653"
],
[
"0.505",
"0.643",
"0.8906",
"0.643",
"0.8906",
"0.657",
"0.505",
"0.657"
],
[
"0.0809",
"0.653",
"0.4673",
"0.653",
"0.4673",
"0.6665",
"0.0809",
"0.6665"
],
[
"0.501",
"0.657",
"0.596",
"0.659",
"0.596",
"0.674",
"0.501",
"0.672"
],
[
"0.0825",
"0.6685",
"0.468",
"0.6685",
"0.468",
"0.6816",
"0.0825",
"0.6816"
],
[
"0.0809",
"0.6846",
"0.411",
"0.6846",
"0.411",
"0.6978",
"0.0809",
"0.6978"
],
[
"0.5024",
"0.6797",
"0.6855",
"0.6816",
"0.6855",
"0.6978",
"0.5024",
"0.696"
],
[
"0.5215",
"0.704",
"0.888",
"0.705",
"0.888",
"0.7197",
"0.5215",
"0.7183"
],
[
"0.1544",
"0.711",
"0.392",
"0.7134",
"0.392",
"0.732",
"0.1544",
"0.73"
],
[
"0.4453",
"0.7153",
"0.468",
"0.7153",
"0.468",
"0.729",
"0.4453",
"0.729"
],
[
"0.5034",
"0.72",
"0.8906",
"0.72",
"0.8906",
"0.7344",
"0.5034",
"0.7344"
],
[
"0.1936",
"0.729",
"0.2075",
"0.729",
"0.2075",
"0.733",
"0.1936",
"0.733"
],
[
"0.2197",
"0.7344",
"0.2336",
"0.7344",
"0.2336",
"0.7437",
"0.2197",
"0.7437"
],
[
"0.5034",
"0.7344",
"0.8896",
"0.735",
"0.8896",
"0.75",
"0.5034",
"0.7485"
],
[
"0.5034",
"0.75",
"0.8896",
"0.7505",
"0.8896",
"0.765",
"0.5034",
"0.7646"
],
[
"0.09314",
"0.759",
"0.4673",
"0.759",
"0.4673",
"0.773",
"0.09314",
"0.773"
],
[
"0.08905",
"0.765",
"0.0972",
"0.765",
"0.0972",
"0.7695",
"0.08905",
"0.7695"
],
[
"0.5034",
"0.7666",
"0.8896",
"0.7666",
"0.8896",
"0.78",
"0.5034",
"0.78"
],
[
"0.0825",
"0.773",
"0.4673",
"0.7734",
"0.4673",
"0.7886",
"0.0825",
"0.787"
],
[
"0.5034",
"0.78",
"0.8896",
"0.781",
"0.8896",
"0.7954",
"0.5034",
"0.7944"
],
[
"0.0809",
"0.789",
"0.4697",
"0.789",
"0.4697",
"0.8037",
"0.0809",
"0.8037"
],
[
"0.5063",
"0.797",
"0.8906",
"0.797",
"0.8906",
"0.81",
"0.5063",
"0.81"
],
[
"0.0809",
"0.8047",
"0.1805",
"0.803",
"0.1805",
"0.8174",
"0.0809",
"0.8193"
],
[
"0.505",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.505",
"0.8247"
],
[
"0.0784",
"0.8306",
"0.3694",
"0.832",
"0.3694",
"0.8467",
"0.0784",
"0.845"
],
[
"0.5034",
"0.8257",
"0.892",
"0.8257",
"0.892",
"0.839",
"0.5034",
"0.839"
],
[
"0.5034",
"0.8413",
"0.8906",
"0.8413",
"0.8906",
"0.855",
"0.5034",
"0.855"
],
[
"0.10046",
"0.857",
"0.4656",
"0.857",
"0.4656",
"0.8706",
"0.10046",
"0.8706"
],
[
"0.505",
"0.856",
"0.8896",
"0.856",
"0.8896",
"0.8706",
"0.505",
"0.8706"
],
[
"0.0948",
"0.872",
"0.468",
"0.872",
"0.468",
"0.886",
"0.0948",
"0.886"
],
[
"0.5034",
"0.8706",
"0.8906",
"0.8706",
"0.8906",
"0.8853",
"0.5034",
"0.8853"
],
[
"0.0809",
"0.887",
"0.2942",
"0.887",
"0.2942",
"0.901",
"0.0809",
"0.901"
],
[
"0.5034",
"0.886",
"0.8906",
"0.886",
"0.8906",
"0.901",
"0.5034",
"0.901"
],
[
"0.478",
"0.924",
"0.4944",
"0.924",
"0.4944",
"0.938",
"0.478",
"0.938"
]
] | [
"Visual",
"Visual Informa",
"ain",
"Consistency loss",
"of Proposals",
"Attention score: 0.72",
"",
"",
"xmin=247ymin-42",
"height-121,width-115",
"Proposal",
"Snn",
"Visual Info",
"Generation",
"ledge",
"Attention score: 0.54",
"",
"xmin=372ymin=123",
"Visual",
"Visual Modality",
"height-82.width=68",
"Faature",
"LSTM",
"Reconstructed visual",
"Multimodal",
"information&Attention",
"Language Modality",
"space",
"KBE",
"weighted visual features",
"Query: \"A man playing",
"LSTM",
"Language",
"Attent",
"Gate",
"Encoding",
"Model",
"Language",
"Consistency loss",
"skateboard trick\"",
"Figure2.",
"Knowledge Aided Consistency Network (KAC Net) consists of a visual consistency branch and a language",
"consistency branch. Visual consistency branch aims at predicting and aligning query-related proposals' location parameters",
"conditioned on the input query. Language consistency branch attempts to reconstruct input query from query-related",
"proposals. To provide guidance in training and testing, a Knowledge Based Pooling (KBP) gate is applied to filter out",
"unrelated proposals for both branches.",
"category treepeople eartable",
"为了对这些知识进行编码，我们首先解析语言查询",
"并通过自然语言处理（NLP）解析器检索所有名词词。",
"P",
"0.020.840.070.01",
"对于每个提案的分布pi，我们选择概率最高的最可能",
"1",
"的类。然后将建议 ri的知识 k计算为该类的名称与",
"Pre-trained CNNprediction:people",
"查询中的名词单词之间的单词相似度（图。3）。如果",
"Query: A smiling woman",
"查询包含多个名词词，我们将所有计算出的相似性平",
"均为知识碍，可以写成",
"ksimilarity(c\"woman\",\"people\")=-0.72",
"Na",
"Figure 3. A pre-trained CNN always predicts a probability",
"1",
"k9",
"sim(C*, w)",
"(2)",
"distribution for its own task. We leverage the most prob-",
"Naj=l",
"able category predicted by CNN and calculate the word",
"similarity between nounwords inthe queryasknowledge",
"其中C*是建议ri的预测类名，w是查询q中所有",
"k9",
"N。名词单词的第i个单词。sim是一个测量两个单词",
"之间相似性的函数",
"在训练阶段，知识碍充当“池化”门，有助于视觉",
"预测参数应恢复每个提案的位置。在此基础上，通过计",
"（秒。3.3）和语言（秒。3.4）一致性分支选择和重建",
"算预测和原始建议的位置参数之间的差异，产生了视",
"可靠的候选提案。在测试阶段，知识过滤掉不相关",
"觉一致性损失Cue。为了利用丰富的图像功能和可用",
"的提案，并增加了找到包含所提及对象的提案的机会",
"的固定类别分类器，我们应用KBP对CNN提供的知",
"(Sec。3.5)",
"识进行编码，并对每个提案在视觉和语言一致性方面",
"的重要性进行加权。KACNet的目标可以写成",
"3.3. Visual Consistency",
"视觉一致性的目标是通过学习来优化注意力模型",
"arg min(cle+入Che)+μCreg",
"(1)",
"以预测与查询相关的建议中包含的位置信息。通过预",
"H",
"q",
"测以输人查询为条件的位置信息，我们期望学习语言",
"和视觉模态之间更好的相关性。在弱监督场景中，没有",
"其中θ表示要优化的参数。Ck是来自语言一致性分",
"",
"注释可用于指示与查询相关的提案的标识。相反，我们",
"支的重建损失，Ckc是来自视觉一致性分支的重建损失",
"使用KBP的知识k在培训期间提供指导。当建议r",
"（上标“k”是指KBP)。Creg是一个权重正则化项。入",
"与查询相关时，我们期望知识k提供更高的分数。因",
"、μ是超参数",
"此，可以将KBP应用于根据查询9自适应地加权每个",
"3.2. Knowledge Based Pooling (KBP)",
"提案的视觉一致性损失。在实现中，我们首先应用长短",
"期内存（LSTM)E[20模型将输入查询q编码为嵌人",
"我们应用预训练的CNN为提案ri提取视觉特征v",
"向量qERda。采用预先训练的CNN为每个建议ri",
"并为其自已的任务预测概率分布p，，这提供了有用",
"提取视觉特征viERd，为输人图像提取全局视觉",
"的线索来过滤掉不相关的提案",
"特征uERdu。然后，注意力模型将嵌人向量q，图",
"3"
] | {
"arxiv_id": "1803.03879",
"language": "zh",
"page_idx": 2
} |
||
1803.03879-zh-3 | 1803.03879-zh-3.jpg | 像全局特征$v$与提案的每个特征$v_i$连接起来，并将它
们投影到$m$维子空间中。多峰特征$v_i^q$计算为
(3)
$$v_i^q=\varphi(W_m(q||v||v_i)+b_m)$$
, 其中$W_m\in\mathbb{R}^m\times(d_q+2d_v),\boldsymbol{b}_m\in\mathbb{R}^m$ 是投影参数。
$\varphi(.)$是一个非线性激活函数。“|”表示串联运算符
在投影到多模态子空间后，注意力模型通过全连接(fc)层预测 5D 向量$s^p\in\mathbb{R}^5$ (上标“p”表示预测)。
$$s_i^p=W_sv_i^q+b_s$$
(4)
其中$W_s\in\mathbb{R}^{5\times m}$和$b_s\in\mathbb{R}^5$是投影参数。$s_i^p$中的第一个元素估计$r_i$与输入查询$q$相关的置信度，接下来的四个元素表示每个提案的预测位置参数
我们将预测的位置参数与原始提案的参数$t_i\in\mathbb{R}^4$
进行比较，并计算回归损失
$$d_i=\dfrac{1}{4}\sum\limits_{j=0}^3f(|\boldsymbol{t}_i[j]-\boldsymbol{s}_i^p[j+1]|)$$
(5)
,其中$f(.)$ 是平滑的 L1 损失函数$:f(x)=0.5x^2\left(|x|<\right.$ $\begin{array}{l}{1)\text{ 和 }f(x)=|x|-0.5(|x|\geq1)\text{ 。位置参数 }t_{i}\text{ 的形式为}}\\{[x_{i1}/w,y_{i1}/h,x_{i2}/w,y_{i2}/h]-0.5\mathrm{~,~其中~}x_{i1},x_{i2}\text{ 是建议}}\end{array}$ $r_i$的最小和最大$x$轴位置，而$y_i1,y_{i2}$是最小和最大$y$ 轴位置
在 KBP 门的帮助下，我们根据预测的置信度$s_i^p[0]$ 和知识$k_i^q$对每个提案的回归损失$d_i$进行加权。视觉一致性损失$\mathcal{L}_vc^k$计算为
$$\mathcal{L}_{vc}^k=\sum_{i=1}^N\sigma(k_i^q)\phi(s_i^p[0])d_i$$
(6)
,其中 $\phi(.),\sigma(.)$ 分别表示 softmax 函数和 sigmoid 函
数
# 3.4. Language Consistency
语言一致性的目标是通过学习使用语言一致性约束
重构输入查询$q$来优化注意力模型
在实现中，在注意模型预测每个提案与查询$q$(方程式$\operatorname{E}$4 中的 $s_i^p[0]$ )相关的置信度之后，我们在$\operatorname{E}[34]$ 中采用类似的结构对每个提案的视觉特征$v_i$进行加权，并将其投影到重建子空间中。与$\mathbb{E}[34]$ 不同，我们将 KBP 门引人到语言一致性分支中，以进一步降低不相关的视觉特征的贡献。因此，知识条件重建特征计算为
$$\boldsymbol{v}_{att}^k=\boldsymbol{W}_a\left(\sum_{i=1}^N\sigma(k_i^q)\phi(\boldsymbol{s}_i^p[0])\boldsymbol{v}_i\right)+\boldsymbol{b}_a$$
$、7$,
其中$W_a\in \mathbb{R} ^{d_r\times d_v}$ , $\boldsymbol{b}_a\in \mathbb{R} ^{d_r}$ 是要优化的投影参数。
其他符号与方程 6 相同
然后将重建视觉特征$v_{att}^k$视为解码 LSTM 的初始状态，该解码 LSTM 预测概率$\{p_{\hat{q}}^t\}$序列，该序列指
示在重建查询$\hat{q}$的每个时间步$t$中选择单词。对于输人查询$q$的基本事实(在每个时间步$t$中选择单词$w_t$ ),语言重建损失$\mathcal{L}_lc^k$是序列$\{p_{\hat{q}}^t\}$的交叉熵的平均值。
(8)
$$\mathcal{L}_{lc}^k=-\frac{1}{T}\sum_{t=1}^T\log(\boldsymbol{p}_{\hat{q}}^t[w_t])$$
其中$T$是输人查询$q$的长度
# 3.5. Training & Inference
在训练阶段，要优化的参数包括编码和解码 LSTM 中的参数以及公式$\color{red}{\operatorname*{E3}},\color{red}{\operatorname*{E4}},\color{red}{\operatorname*{E7}}$中的投影参数。我们正则化投影参数的权重，即这些参数的$\ell_2$范数之和$(\mathcal{L}_{reg}$ )。与 [34]相同，我们选择由提案生成系统$(\begin{array}{c}N=100\end{array})$生成的 100 个提案。选择整流线性单元(ReLU)作为非线性激活函数$\varphi$。使用 Adam[24] 算法对 KACNet 进行端到端训练
在测试阶段，我们将查询$q$馈送到经过训练的 KACNet 中，并根据注意力模型 (方程式$\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{blue}{\mathrm{E}}\color{red}{\mathrm{E}}\color{blue}{\mathrm{E}}\color{blue}{\mathrm{E}}\color{blue}{\mathrm{E}}\color{blue}{\mathrm{E}}\color{blue}{\mathrm{E}}\color{blue}{\mathrm{E}}\color{blue}{\mathrm{E}}\color{blue}{\mathrm{E}} {array}$ $k_i^q$生成的置信度$\{s_i^p[0]\}$选择最相关的建议。最终的预测结果为(公式$\color{red}{\mathrm{E6}}$中的符号相同):
(9)
$$r_{j^*},\quad\mathrm{s.t.}\quad j^*=\arg\max_i\{\phi(\boldsymbol{s}_i^p[0])\sigma(k_i^q)\}$$
# 4. Experiment
我们在弱监督接地场景中评估 Flickr30K 实体$\mathbb{E}[32]$
和推荐游戏$\mathbb{E}[23]$ 数据集上的 KACNet
## $4. 1. \textbf{Datasets}$
Flickr30K 实体[32]:此数据集中有 29783、1000、 1000个图像分别用于训练，验证和测试。每个图像与5 个字幕相关联，与 3。平均每个标题中有 52 个查询短语(总共 360K 个查询短语)。所有这些查询的词汇量是 17150。在弱监督场景中，我们忽略了这两个数据集的边界框注释
推荐游戏[23]:此数据集中有 19,894 张自然场景图像，这些图像中有 96,654 个不同的对象。每个对象由 1-3 个查询短语引用 (共 130,525 个)。所有短语中有 8800 个独特的单词，最大长度为 19 个单词
# 4.2. Experiment Setup
提案生成。我们对 Flickr30K 实体 [32] 采用选择性搜索[36] ,对 ReferitGame [23]采用 EdgeBoxes [39] 来生成提案，作为在这两个数据集上与 [34]进行公平比较的基础候选
视觉特征表示。与$\mathbb{E}[34]$ 相同，我们选择由 FastRCNN$\underline{\mathrm{E}}[15]$在 PASCALVOC2007$\underline{\mathrm{E}}[10]$上微调的VGG 网络$\mathbb{E}[35]$ ,以提取 Flickr30K 实体的视觉特征，表示为“VGG det”。此外，我们遵循$\mathbb{E}[34]$ 并应用在 ImageNet$\underline{\mathrm{E}}[8]$上预训练的 VGG 网络来提取Flickr30K 实体和引用游戏数据集的视觉特征，这些数据集表示为“VGG cls”。“VGG cls”和“VGG det” 特征都是 4096D 向量( $d_v=4096$ )
4 | [
[
"0.0825",
"0.09216",
"0.468",
"0.09216",
"0.468",
"0.1061",
"0.0825",
"0.1061"
],
[
"0.5034",
"0.09216",
"0.8906",
"0.09216",
"0.8906",
"0.1061",
"0.5034",
"0.1061"
],
[
"0.0809",
"0.108",
"0.4126",
"0.108",
"0.4126",
"0.1218",
"0.0809",
"0.1218"
],
[
"0.505",
"0.108",
"0.8906",
"0.108",
"0.8906",
"0.1218",
"0.505",
"0.1218"
],
[
"0.505",
"0.1237",
"0.888",
"0.1237",
"0.888",
"0.137",
"0.505",
"0.137"
],
[
"0.1748",
"0.1364",
"0.3743",
"0.1364",
"0.3743",
"0.1497",
"0.1748",
"0.1497"
],
[
"0.4453",
"0.1364",
"0.468",
"0.1364",
"0.468",
"0.1497",
"0.4453",
"0.1497"
],
[
"0.6235",
"0.1338",
"0.8037",
"0.154",
"0.7935",
"0.2122",
"0.613",
"0.1913"
],
[
"0.681",
"0.1509",
"0.699",
"0.1477",
"0.7017",
"0.1578",
"0.6846",
"0.161"
],
[
"0.0784",
"0.159",
"0.4714",
"0.159",
"0.4714",
"0.1799",
"0.0784",
"0.1799"
],
[
"0.609",
"0.159",
"0.6543",
"0.161",
"0.653",
"0.178",
"0.607",
"0.1755"
],
[
"0.869",
"0.1622",
"0.8906",
"0.1622",
"0.8906",
"0.1768",
"0.869",
"0.1768"
],
[
"0.0801",
"0.1787",
"0.4478",
"0.178",
"0.4478",
"0.1925",
"0.0801",
"0.1932"
],
[
"0.1013",
"0.1945",
"0.468",
"0.1945",
"0.468",
"0.2084",
"0.1013",
"0.2084"
],
[
"0.5034",
"0.1995",
"0.709",
"0.1995",
"0.709",
"0.2134",
"0.5034",
"0.2134"
],
[
"0.0784",
"0.2084",
"0.4673",
"0.209",
"0.4673",
"0.2235",
"0.0784",
"0.2229"
],
[
"0.5034",
"0.2229",
"0.705",
"0.2229",
"0.705",
"0.2361",
"0.5034",
"0.2361"
],
[
"0.2157",
"0.2374",
"0.3325",
"0.2374",
"0.3325",
"0.252",
"0.2157",
"0.252"
],
[
"0.4468",
"0.238",
"0.4697",
"0.238",
"0.4697",
"0.252",
"0.4468",
"0.252"
],
[
"0.524",
"0.2456",
"0.892",
"0.2456",
"0.892",
"0.2588",
"0.524",
"0.2588"
],
[
"0.0801",
"0.2632",
"0.468",
"0.2644",
"0.468",
"0.279",
"0.0801",
"0.2778"
],
[
"0.505",
"0.2615",
"0.892",
"0.2615",
"0.892",
"0.2747",
"0.505",
"0.2747"
],
[
"0.09235",
"0.2798",
"0.4697",
"0.2798",
"0.4697",
"0.2935",
"0.09235",
"0.2935"
],
[
"0.505",
"0.276",
"0.8896",
"0.276",
"0.8896",
"0.289",
"0.505",
"0.289"
],
[
"0.0809",
"0.2954",
"0.3757",
"0.2954",
"0.3757",
"0.3093",
"0.0809",
"0.3093"
],
[
"0.5034",
"0.2903",
"0.8896",
"0.2903",
"0.8896",
"0.3037",
"0.5034",
"0.3037"
],
[
"0.5024",
"0.3037",
"0.892",
"0.3062",
"0.892",
"0.3208",
"0.5024",
"0.318"
],
[
"0.1013",
"0.31",
"0.4673",
"0.31",
"0.4673",
"0.324",
"0.1013",
"0.324"
],
[
"0.0809",
"0.3257",
"0.2778",
"0.3257",
"0.2778",
"0.339",
"0.0809",
"0.339"
],
[
"0.505",
"0.3213",
"0.8906",
"0.3213",
"0.8906",
"0.3352",
"0.505",
"0.3352"
],
[
"0.1683",
"0.334",
"0.406",
"0.334",
"0.406",
"0.4104",
"0.1683",
"0.4104"
],
[
"0.5034",
"0.336",
"0.7236",
"0.336",
"0.7236",
"0.3499",
"0.5034",
"0.3499"
],
[
"0.524",
"0.3516",
"0.8906",
"0.3516",
"0.8906",
"0.3655",
"0.524",
"0.3655"
],
[
"0.165",
"0.3655",
"0.2035",
"0.3694",
"0.2009",
"0.3801",
"0.1642",
"0.377"
],
[
"0.4468",
"0.3662",
"0.468",
"0.3662",
"0.468",
"0.379",
"0.4468",
"0.379"
],
[
"0.505",
"0.3687",
"0.8896",
"0.3687",
"0.8896",
"0.379",
"0.505",
"0.379"
],
[
"0.5024",
"0.3801",
"0.8906",
"0.3806",
"0.8906",
"0.3958",
"0.5024",
"0.3945"
],
[
"0.5034",
"0.3965",
"0.777",
"0.3965",
"0.777",
"0.411",
"0.5034",
"0.411"
],
[
"0.08496",
"0.4077",
"0.4673",
"0.4077",
"0.4673",
"0.4216",
"0.08496",
"0.4216"
],
[
"0.0801",
"0.4204",
"0.4673",
"0.4216",
"0.4673",
"0.4363",
"0.0801",
"0.435"
],
[
"0.55",
"0.4216",
"0.8447",
"0.4192",
"0.8447",
"0.4375",
"0.55",
"0.4395"
],
[
"0.8364",
"0.4216",
"0.8906",
"0.4216",
"0.8906",
"0.4363",
"0.8364",
"0.4363"
],
[
"0.0801",
"0.4375",
"0.468",
"0.4363",
"0.468",
"0.4507",
"0.0801",
"0.452"
],
[
"0.0801",
"0.4526",
"0.468",
"0.4526",
"0.468",
"0.4673",
"0.0801",
"0.4673"
],
[
"0.5024",
"0.452",
"0.6313",
"0.452",
"0.6313",
"0.4653",
"0.5024",
"0.4653"
],
[
"0.0809",
"0.4685",
"0.1299",
"0.4685",
"0.1299",
"0.4797",
"0.0809",
"0.4797"
],
[
"0.524",
"0.4766",
"0.8896",
"0.4766",
"0.8896",
"0.4905",
"0.524",
"0.4905"
],
[
"0.10046",
"0.483",
"0.4673",
"0.483",
"0.4673",
"0.4968",
"0.10046",
"0.4968"
],
[
"0.5034",
"0.4924",
"0.7817",
"0.4924",
"0.7817",
"0.506",
"0.5034",
"0.506"
],
[
"0.0825",
"0.5",
"0.4673",
"0.5",
"0.4673",
"0.5103",
"0.0825",
"0.5103"
],
[
"0.08905",
"0.512",
"0.2484",
"0.513",
"0.2484",
"0.528",
"0.08905",
"0.5264"
],
[
"0.505",
"0.5166",
"0.607",
"0.5166",
"0.607",
"0.528",
"0.505",
"0.528"
],
[
"0.2345",
"0.5415",
"0.25",
"0.5415",
"0.25",
"0.5503",
"0.2345",
"0.5503"
],
[
"0.5215",
"0.537",
"0.892",
"0.5386",
"0.892",
"0.552",
"0.5215",
"0.5503"
],
[
"0.1814",
"0.5503",
"0.3677",
"0.5503",
"0.3677",
"0.5713",
"0.1814",
"0.5713"
],
[
"0.4468",
"0.5547",
"0.4697",
"0.5547",
"0.4697",
"0.5674",
"0.4468",
"0.5674"
],
[
"0.505",
"0.5547",
"0.8896",
"0.5547",
"0.8896",
"0.565",
"0.505",
"0.565"
],
[
"0.2303",
"0.5693",
"0.2532",
"0.5693",
"0.2532",
"0.582",
"0.2303",
"0.582"
],
[
"0.5024",
"0.5674",
"0.8906",
"0.5684",
"0.8906",
"0.582",
"0.5024",
"0.5806"
],
[
"0.5034",
"0.584",
"0.8896",
"0.584",
"0.8896",
"0.5977",
"0.5034",
"0.5977"
],
[
"0.09644",
"0.5947",
"0.468",
"0.5947",
"0.468",
"0.608",
"0.09644",
"0.608"
],
[
"0.5034",
"0.5986",
"0.8906",
"0.5986",
"0.8906",
"0.6123",
"0.5034",
"0.6123"
],
[
"0.0809",
"0.6094",
"0.10046",
"0.6094",
"0.10046",
"0.622",
"0.0809",
"0.622"
],
[
"0.5034",
"0.613",
"0.619",
"0.613",
"0.619",
"0.627",
"0.5034",
"0.627"
],
[
"0.0801",
"0.634",
"0.2893",
"0.634",
"0.2893",
"0.6484",
"0.0801",
"0.6484"
],
[
"0.523",
"0.629",
"0.8906",
"0.629",
"0.8906",
"0.6426",
"0.523",
"0.6426"
],
[
"0.5034",
"0.6426",
"0.8896",
"0.6445",
"0.8896",
"0.6577",
"0.5034",
"0.656"
],
[
"0.10046",
"0.657",
"0.468",
"0.657",
"0.468",
"0.6704",
"0.10046",
"0.6704"
],
[
"0.505",
"0.659",
"0.8896",
"0.659",
"0.8896",
"0.6724",
"0.505",
"0.6724"
],
[
"0.0809",
"0.6724",
"0.33",
"0.6724",
"0.33",
"0.686",
"0.0809",
"0.686"
],
[
"0.5034",
"0.675",
"0.8364",
"0.675",
"0.8364",
"0.688",
"0.5034",
"0.688"
],
[
"0.10046",
"0.6875",
"0.4673",
"0.6875",
"0.4673",
"0.7007",
"0.10046",
"0.7007"
],
[
"0.5024",
"0.6963",
"0.6807",
"0.6987",
"0.6807",
"0.7134",
"0.5024",
"0.711"
],
[
"0.0825",
"0.704",
"0.4656",
"0.704",
"0.4656",
"0.714",
"0.0825",
"0.714"
],
[
"0.0825",
"0.7183",
"0.468",
"0.7183",
"0.468",
"0.732",
"0.0825",
"0.732"
],
[
"0.523",
"0.72",
"0.8906",
"0.72",
"0.8906",
"0.7344",
"0.523",
"0.7344"
],
[
"0.0809",
"0.733",
"0.468",
"0.733",
"0.468",
"0.747",
"0.0809",
"0.747"
],
[
"0.5034",
"0.7363",
"0.8896",
"0.7363",
"0.8896",
"0.75",
"0.5034",
"0.75"
],
[
"0.0809",
"0.7476",
"0.468",
"0.7476",
"0.468",
"0.761",
"0.0809",
"0.761"
],
[
"0.5024",
"0.7505",
"0.8896",
"0.75",
"0.8896",
"0.7646",
"0.5024",
"0.765"
],
[
"0.0809",
"0.763",
"0.468",
"0.763",
"0.468",
"0.7764",
"0.0809",
"0.7764"
],
[
"0.5034",
"0.765",
"0.619",
"0.7666",
"0.619",
"0.78",
"0.5034",
"0.779"
],
[
"0.0801",
"0.78",
"0.0989",
"0.78",
"0.0989",
"0.7925",
"0.0801",
"0.7925"
],
[
"0.524",
"0.781",
"0.8906",
"0.781",
"0.8906",
"0.7944",
"0.524",
"0.7944"
],
[
"0.2263",
"0.7974",
"0.2418",
"0.7974",
"0.2418",
"0.8003",
"0.2263",
"0.8003"
],
[
"0.5024",
"0.797",
"0.8896",
"0.797",
"0.8896",
"0.81",
"0.5024",
"0.81"
],
[
"0.1356",
"0.803",
"0.2239",
"0.802",
"0.2239",
"0.821",
"0.1356",
"0.8213"
],
[
"0.2214",
"0.803",
"0.3743",
"0.8047",
"0.3743",
"0.8228",
"0.2214",
"0.821"
],
[
"0.3694",
"0.8047",
"0.415",
"0.8047",
"0.415",
"0.8193",
"0.3694",
"0.8193"
],
[
"0.411",
"0.811",
"0.423",
"0.811",
"0.423",
"0.815",
"0.411",
"0.815"
],
[
"0.4453",
"0.8066",
"0.4673",
"0.8066",
"0.4673",
"0.8193",
"0.4453",
"0.8193"
],
[
"0.505",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.505",
"0.8247"
],
[
"0.2181",
"0.821",
"0.2507",
"0.821",
"0.2507",
"0.832",
"0.2181",
"0.832"
],
[
"0.3628",
"0.8228",
"0.3677",
"0.8228",
"0.3677",
"0.829",
"0.3628",
"0.829"
],
[
"0.5024",
"0.8237",
"0.892",
"0.8247",
"0.892",
"0.8423",
"0.5024",
"0.8413"
],
[
"0.08417",
"0.8403",
"0.468",
"0.8403",
"0.468",
"0.855",
"0.08417",
"0.855"
],
[
"0.5034",
"0.8413",
"0.8906",
"0.8413",
"0.8906",
"0.855",
"0.5034",
"0.855"
],
[
"0.0809",
"0.857",
"0.2532",
"0.857",
"0.2532",
"0.8706",
"0.0809",
"0.8706"
],
[
"0.5024",
"0.855",
"0.8906",
"0.856",
"0.8906",
"0.8706",
"0.5024",
"0.869"
],
[
"0.0989",
"0.872",
"0.468",
"0.8706",
"0.468",
"0.8853",
"0.0989",
"0.886"
],
[
"0.5024",
"0.8706",
"0.8823",
"0.872",
"0.8823",
"0.886",
"0.5024",
"0.8853"
],
[
"0.0809",
"0.887",
"0.468",
"0.887",
"0.468",
"0.9014",
"0.0809",
"0.9014"
],
[
"0.5024",
"0.887",
"0.759",
"0.887",
"0.759",
"0.901",
"0.5024",
"0.901"
],
[
"0.4795",
"0.925",
"0.492",
"0.925",
"0.492",
"0.9375",
"0.4795",
"0.9375"
]
] | [
"像全局特征与提案的每个特征U连接起来，并将它",
"示在重建查询的每个时间步t中选择单词。对于输",
"们投影到m维子空间中。多峰特征u计算为",
"人查询q的基本事实（在每个时间步t中选择单词wt",
")，语言重建损失Ct。是序列（p)的交叉熵的平均值。",
"vg = p(Wm(qllullvi) + bm)",
"(3)",
"+-",
"T",
"，其中WmRmx(dg+2d），bmERm是投影参数。",
"Cke=",
"(8)",
"(）是一个非线性激活函数。“”表示串联运算符",
"在投影到多模态子空间后，注意力模型通过全连接",
"其中T是输入查询q的长度",
"(fc）层预测5D向量sPER5（上标“p”表示预测）。",
"3.5. Training & Inference",
"sp = W.u +bs",
"(4)",
"在训练阶段，要优化的参数包括编码和解码LSTM",
"其中WEIR5xm和bsER5是投影参数。s中的第",
"中的参数以及公式E3，E4，E7中的投影参数。我",
"个元素估计r与输人查询9相关的置信度，接下来",
"们正则化投影参数的权重，即这些参数的2范数之",
"的四个元素表示每个提案的预测位置参数",
"和（Creg）。与[341相同，我们选择由提案生成系统",
"（N=100）生成的100个提案。选择整流线性单元",
"我们将预测的位置参数与原始提案的参数tER4",
"进行比较，并计算回归损失",
"(ReLU）作为非线性激活函数。使用Adam[24]算",
"-1",
"法对KACNet进行端到端训练",
"在测试阶段，我们将查询q馈送到经过训练的KAC-",
"di=",
"(5)",
"Net中，并根据注意力模型（方程式E4）和外部知识",
"k生成的置信度（s[选择最相关的建议。最终的",
"预测结果为（公式E6中的符号相同）：",
"其中f()是平滑的L1损失函数：f(a）=0.5a²（lzl<",
"1)和f(）=-0.5(l≥1)。位置参数t的形式为",
"rj*, s.t. j* = argmax[o(sp[0])o(kg))",
" (9)",
"[c/w,i/h,ai2/w,yi2/h]-0.5，其中i,i2是建议",
"ri的最小和最大a轴位置，而yil,yi2是最小和最大y",
"4.Experiment",
"轴位置",
"我们在弱监督接地场景中评估Flickr30K实体E[32]",
"在KBP门的帮助下，我们根据预测的置信度s’[0]",
"和推荐游戏E[23]数据集上的KACNet",
"和知识对每个提案的回归损失进行加权。视觉",
"一致性损失Cke计算为",
"4.1. Datasets",
"N",
"Flickr30K实体[32：此数据集中有29783、1000、",
"Che =Zo(k9)0(sP[0])d;",
"(6)",
"1000个图像分别用于训练，验证和测试。每个图像与",
"i=1",
"5个字幕相关联，与3。平均每个标题中有52个查询",
"短语（总共360K个查询短语）。所有这些查询的词汇",
"其中Φ(),α（）分别表示softmax函数和sigmoid函",
"量是17150。在弱监督场景中，我们忽略了这两个数据",
"数",
"集的边界框注释",
"3.4. Language Consistency",
"推荐游戏[23]：此数据集中有19,894张自然场景",
"图像，这些图像中有96654个不同的对象。每个对象",
"语言一致性的目标是通过学习使用语言一致性约束",
"由1-3个查询短语引用（共130,525个）。所有短语中",
"重构输人查询q来优化注意力模型",
"有8800个独特的单词，最大长度为19个单词",
"在实现中，在注意模型预测每个提案与查询9（方",
"4.2. Experiment Setup",
"程式F4中的sPO1）相关的置信度之后，我们在F34",
"中采用类似的结构对每个提案的视觉特征V进行加",
"提案生成。我们对Flickr30K实体「32】采用选择性",
"权，并将其投影到重建子空间中。与E34]不同，我们",
"搜索[36]，对ReferitGame[23]采用EdgeBoxes[39]",
"将KBP门引人到语言一致性分支中，以进一步降低不",
"来生成提案，作为在这两个数据集上与【34]进行公平",
"相关的视觉特征的贡献。因此，知识条件重建特征计算",
"比较的基础候选",
"为",
"视觉特征表示。与E341相同，我们选择由Fast-",
"14",
"RCNNE15]在PASCALVOC2007E[10]上微调的",
"vatt = Wa (",
"Zo(k)(s [0])u)",
"+ba",
"",
"(7)",
"VGG网络E[35]，以提取Flickr30K实体的视觉特",
"(i=1",
"",
"征，表示为“VGG det”。此外，我们遵循E[34]并",
"，其中WaERdrxdt，baERdr是要优化的投影参数。",
"应用在ImageNetE[8]上预训练的VGG网络来提取",
"其他符号与方程6相同",
"Flickr30K实体和引用游戏数据集的视觉特征，这些数",
"然后将重建视觉特征uatt视为解码LSTM的初始",
"据集表示为“VGGcls”。“VGGcls”和“VGGdet",
"状态，该解码LSTM预测概率(p}序列，该序列指",
"特征都是4096D向量（d=4096）",
"4"
] | {
"arxiv_id": "1803.03879",
"language": "zh",
"page_idx": 3
} |
||
2109.00725-zh-1 | 2109.00725-zh-1.jpg | # Causal Inference in Natural Language Processing: $\textbf{Estimation, Prediction, Interpretation and Beyond}$
Amir Feder$^{\text{非},10}$, Katherine A. Keith$^2$, Emaad Manzoor$^3$, Reid Pryzant$^4$, and
Zach Wood-Doughty$^{6}$
$^{1}$Technion - Israel Institute of Technology
$^{2}$Williams College $^{2}$Williams College
$^{3}$University of Wisconsin- Madison
$^{4}$Microsoft
$^{6}$Northwestern University
## Abstract
A fundamental goal of scientific research is to learn about causal relationships. However, despite its critical role in the life and social sciences, causality has not had the same importance in Natural Language Processing (NLP), which has traditionally placed more emphasis on predictive tasks. This distinction is beginning to fade, with an emerging area of interdisciplinary research at the convergence of causal inference and language processing. Still, research on causality in NLP remains scattered across domains without unified definitions, benchmark datasets and clear articulations of the challenges and opportunities in the application of causal inference to the textual domain, with its unique properties. In this survey, we consolidate research across academic areas and situate it in the broader NLP landscape. We introduce the statistical challenge of estimating causal effects with text, encompassing settings where text is used as an outcome, treatment, or to address confounding. In addition, we explore potential uses of causal inference to improve the robustness, fairness, and interpretability of NLP models. We thus provide a unified overview of causal inference for the NLP community. $^{\mathrm{l}}$
All authors equally contributed to this paper. Au-
thor names are organized alphabetically in two clusters: First students and post-docs and then faculty members. The email address of the corresponding (first) author is: federgcampus.technion.ac.il.
An online repository containing existing research
### $\textbf{1}$ $\textbf{Introduction}$
NLP 的有效性不断提高，为跨学科合作创造了令人兴奋的新机会，将 NLP 技术带入了广泛的外部研究学科E(e。g。, Roberts et al.,2014; Zhang et al., 2020; ?) 并将新数据和任务纳人主流 NLP (e。g。,Thomas et al.,2006;?)。在这样的跨学科合作中，许多最重要的研究问题都与因果关系的推断有关。例如，在推荐新的药物治疗之前，临床医生想知道药物对疾病进展的因果影响。因果推断涉及一个关于通过采取干预措施而产生的反事实世界的问题：如果我们给他们服用药物，患者的疾病进展会是什么？正如我们在下面通过观察数据所解释的那样，因果关系并不等同于是否服用药物与观察到的疾病进展之间的相关性。现在有大量关于使用传统(非文本)数据集进行有效推断的技术的文献$\overline{\mathrm{E}}($e。g。, ?) , 但是将这些技术应用于自然语言数据会带来新的基本挑战
相反，在许多经典的 NLP 应用程序中，主要目标是做出准确的预测：任何统计相关性都是可接受的，而与潜在的因果关系无关。然而，随着 NLP 系统越来越多地部署在具有挑战性和高风险的场景中，我们不能依赖通常的假设，即训练和测试数据是相同分布的，我们可能不满足于不可解释的黑盒预测器。对于这两个问题，因果关系都提供了一条有希望的前进之路：数据生成过程的因果结构的领域知识可以暗示归纳偏差，从而导致更强大的预测因子，而预测因子本身的因果观点可以为其内部
on causal inference and language processing is avail-
able here:
https://github.com/causaltext/
causal-text-papers | [
[
"0.2418",
"0.082",
"0.7573",
"0.0831",
"0.7573",
"0.10095",
"0.2418",
"0.0998"
],
[
"0.2435",
"0.10394",
"0.7573",
"0.10394",
"0.7573",
"0.11816",
"0.2435",
"0.11816"
],
[
"0.2024",
"0.1621",
"0.802",
"0.1621",
"0.802",
"0.177",
"0.2024",
"0.177"
],
[
"0.4165",
"0.1799",
"0.5938",
"0.1799",
"0.5938",
"0.1948",
"0.4165",
"0.1948"
],
[
"0.3333",
"0.206",
"0.6724",
"0.2079",
"0.6724",
"0.2257",
"0.3333",
"0.2239"
],
[
"0.4265",
"0.2239",
"0.5786",
"0.2257",
"0.5786",
"0.2405",
"0.4265",
"0.2382"
],
[
"0.4248",
"0.2382",
"0.58",
"0.2411",
"0.58",
"0.2588",
"0.4248",
"0.2559"
],
[
"0.357",
"0.257",
"0.65",
"0.2588",
"0.65",
"0.2737",
"0.357",
"0.2715"
],
[
"0.4592",
"0.275",
"0.55",
"0.275",
"0.55",
"0.2893",
"0.4592",
"0.2893"
],
[
"0.398",
"0.2903",
"0.607",
"0.2922",
"0.607",
"0.3071",
"0.398",
"0.3047"
],
[
"0.2644",
"0.3843",
"0.3418",
"0.3843",
"0.3418",
"0.3984",
"0.2644",
"0.3984"
],
[
"0.513",
"0.386",
"0.66",
"0.386",
"0.66",
"0.3997",
"0.513",
"0.3997"
],
[
"0.1561",
"0.412",
"0.4517",
"0.411",
"0.4517",
"0.4258",
"0.1561",
"0.427"
],
[
"0.513",
"0.411",
"0.879",
"0.411",
"0.879",
"0.4246",
"0.513",
"0.4246"
],
[
"0.157",
"0.4287",
"0.4475",
"0.4287",
"0.4475",
"0.4395",
"0.157",
"0.4395"
],
[
"0.517",
"0.4275",
"0.882",
"0.4275",
"0.882",
"0.4412",
"0.517",
"0.4412"
],
[
"0.157",
"0.4412",
"0.451",
"0.4412",
"0.451",
"0.4548",
"0.157",
"0.4548"
],
[
"0.5156",
"0.4448",
"0.8833",
"0.4448",
"0.8833",
"0.4578",
"0.5156",
"0.4578"
],
[
"0.1561",
"0.456",
"0.4534",
"0.456",
"0.4534",
"0.4697",
"0.1561",
"0.4697"
],
[
"0.5156",
"0.462",
"0.882",
"0.462",
"0.882",
"0.4727",
"0.5156",
"0.4727"
],
[
"0.1561",
"0.4697",
"0.4517",
"0.4697",
"0.4517",
"0.4827",
"0.1561",
"0.4827"
],
[
"0.5156",
"0.4768",
"0.882",
"0.4768",
"0.882",
"0.4912",
"0.5156",
"0.4912"
],
[
"0.1561",
"0.484",
"0.451",
"0.484",
"0.451",
"0.4976",
"0.1561",
"0.4976"
],
[
"0.1545",
"0.4988",
"0.4492",
"0.4963",
"0.4492",
"0.5103",
"0.1545",
"0.5127"
],
[
"0.5156",
"0.4934",
"0.882",
"0.4934",
"0.882",
"0.5073",
"0.5156",
"0.5073"
],
[
"0.1561",
"0.5127",
"0.4517",
"0.5127",
"0.4517",
"0.5254",
"0.1561",
"0.5254"
],
[
"0.5156",
"0.509",
"0.881",
"0.509",
"0.881",
"0.5225",
"0.5156",
"0.5225"
],
[
"0.1561",
"0.527",
"0.4492",
"0.527",
"0.4492",
"0.5405",
"0.1561",
"0.5405"
],
[
"0.5156",
"0.525",
"0.882",
"0.525",
"0.882",
"0.539",
"0.5156",
"0.539"
],
[
"0.157",
"0.542",
"0.4475",
"0.542",
"0.4475",
"0.5527",
"0.157",
"0.5527"
],
[
"0.5156",
"0.5415",
"0.882",
"0.5415",
"0.882",
"0.5547",
"0.5156",
"0.5547"
],
[
"0.1561",
"0.5547",
"0.4492",
"0.5547",
"0.4492",
"0.5684",
"0.1561",
"0.5684"
],
[
"0.5156",
"0.557",
"0.882",
"0.557",
"0.882",
"0.571",
"0.5156",
"0.571"
],
[
"0.1561",
"0.5684",
"0.4475",
"0.5684",
"0.4475",
"0.582",
"0.1561",
"0.582"
],
[
"0.5146",
"0.572",
"0.882",
"0.5723",
"0.882",
"0.5874",
"0.5146",
"0.586"
],
[
"0.157",
"0.585",
"0.4492",
"0.585",
"0.4492",
"0.5957",
"0.157",
"0.5957"
],
[
"0.5146",
"0.5884",
"0.882",
"0.5884",
"0.882",
"0.6016",
"0.5146",
"0.6016"
],
[
"0.1545",
"0.597",
"0.4492",
"0.597",
"0.4492",
"0.611",
"0.1545",
"0.611"
],
[
"0.513",
"0.604",
"0.8833",
"0.605",
"0.8833",
"0.619",
"0.513",
"0.618"
],
[
"0.1561",
"0.6123",
"0.4475",
"0.6123",
"0.4475",
"0.626",
"0.1561",
"0.626"
],
[
"0.1561",
"0.626",
"0.4517",
"0.626",
"0.4517",
"0.6396",
"0.1561",
"0.6396"
],
[
"0.5156",
"0.6216",
"0.882",
"0.6216",
"0.882",
"0.6353",
"0.5156",
"0.6353"
],
[
"0.1528",
"0.6396",
"0.4492",
"0.64",
"0.4492",
"0.654",
"0.1528",
"0.653"
],
[
"0.5146",
"0.6357",
"0.882",
"0.637",
"0.882",
"0.652",
"0.5146",
"0.651"
],
[
"0.1561",
"0.655",
"0.4517",
"0.655",
"0.4517",
"0.6685",
"0.1561",
"0.6685"
],
[
"0.5156",
"0.653",
"0.881",
"0.653",
"0.881",
"0.6665",
"0.5156",
"0.6665"
],
[
"0.1561",
"0.6685",
"0.4517",
"0.6685",
"0.4517",
"0.6816",
"0.1561",
"0.6816"
],
[
"0.5156",
"0.6694",
"0.8096",
"0.6694",
"0.8096",
"0.683",
"0.5156",
"0.683"
],
[
"0.1561",
"0.683",
"0.4475",
"0.683",
"0.4475",
"0.6934",
"0.1561",
"0.6934"
],
[
"0.5356",
"0.686",
"0.882",
"0.686",
"0.882",
"0.6997",
"0.5356",
"0.6997"
],
[
"0.1561",
"0.697",
"0.451",
"0.697",
"0.451",
"0.7104",
"0.1561",
"0.7104"
],
[
"0.5156",
"0.703",
"0.881",
"0.703",
"0.881",
"0.7134",
"0.5156",
"0.7134"
],
[
"0.1561",
"0.7104",
"0.4492",
"0.7114",
"0.4492",
"0.7246",
"0.1561",
"0.7236"
],
[
"0.5156",
"0.7188",
"0.882",
"0.7188",
"0.882",
"0.7324",
"0.5156",
"0.7324"
],
[
"0.1561",
"0.7256",
"0.451",
"0.727",
"0.451",
"0.7383",
"0.1561",
"0.737"
],
[
"0.517",
"0.734",
"0.882",
"0.734",
"0.882",
"0.749",
"0.517",
"0.749"
],
[
"0.157",
"0.741",
"0.4475",
"0.741",
"0.4475",
"0.752",
"0.157",
"0.752"
],
[
"0.1545",
"0.755",
"0.4517",
"0.7544",
"0.4517",
"0.767",
"0.1545",
"0.7686"
],
[
"0.517",
"0.7505",
"0.881",
"0.7505",
"0.881",
"0.764",
"0.517",
"0.764"
],
[
"0.157",
"0.7686",
"0.4492",
"0.7686",
"0.4492",
"0.782",
"0.157",
"0.782"
],
[
"0.5146",
"0.7666",
"0.879",
"0.7666",
"0.879",
"0.78",
"0.5146",
"0.78"
],
[
"0.1561",
"0.782",
"0.4517",
"0.782",
"0.4517",
"0.796",
"0.1561",
"0.796"
],
[
"0.5146",
"0.782",
"0.881",
"0.782",
"0.881",
"0.796",
"0.5146",
"0.796"
],
[
"0.157",
"0.7974",
"0.451",
"0.7974",
"0.451",
"0.808",
"0.157",
"0.808"
],
[
"0.5146",
"0.799",
"0.881",
"0.799",
"0.881",
"0.8125",
"0.5146",
"0.8125"
],
[
"0.1545",
"0.81",
"0.3047",
"0.81",
"0.3047",
"0.8237",
"0.1545",
"0.8237"
],
[
"0.5146",
"0.8135",
"0.8833",
"0.814",
"0.8833",
"0.829",
"0.5146",
"0.8276"
],
[
"0.1453",
"0.83",
"0.4492",
"0.83",
"0.4492",
"0.844",
"0.1453",
"0.844"
],
[
"0.455",
"0.833",
"0.4878",
"0.833",
"0.4878",
"0.8413",
"0.455",
"0.8413"
],
[
"0.5146",
"0.8306",
"0.881",
"0.8306",
"0.881",
"0.844",
"0.5146",
"0.844"
],
[
"0.12006",
"0.8457",
"0.4895",
"0.8457",
"0.4895",
"0.8555",
"0.12006",
"0.8555"
],
[
"0.513",
"0.847",
"0.882",
"0.847",
"0.882",
"0.861",
"0.513",
"0.861"
],
[
"0.12006",
"0.8594",
"0.486",
"0.8594",
"0.486",
"0.869",
"0.12006",
"0.869"
],
[
"0.12006",
"0.8706",
"0.4895",
"0.8706",
"0.4895",
"0.8843",
"0.12006",
"0.8843"
],
[
"0.5146",
"0.8706",
"0.882",
"0.8706",
"0.882",
"0.8843",
"0.5146",
"0.8843"
],
[
"0.12177",
"0.8853",
"0.367",
"0.8853",
"0.367",
"0.8955",
"0.12177",
"0.8955"
],
[
"0.513",
"0.883",
"0.5903",
"0.8843",
"0.5903",
"0.8955",
"0.513",
"0.8945"
],
[
"0.6055",
"0.8853",
"0.879",
"0.8853",
"0.879",
"0.8955",
"0.6055",
"0.8955"
],
[
"0.1411",
"0.8965",
"0.4895",
"0.8965",
"0.4895",
"0.91",
"0.1411",
"0.91"
],
[
"0.5146",
"0.8984",
"0.681",
"0.8984",
"0.681",
"0.909",
"0.5146",
"0.909"
]
] | [
"Causal Inference in Natural Language Processing:",
"Estimation, Prediction, Interpretation and Beyond",
"Amir Feder*,1o, Katherine A. Keith?, Emaad Manzoor3, Reid Pryzant, and",
"Zach Wood-Doughty",
"ITechnion - Israel Institute of Technology",
"2Williams College",
"2 Williams College",
"3University of Wisconsin -Madison",
"4Microsoft",
"6Northwestern University",
"Abstract",
"1Introduction",
"A fundamental goal of scientific research is",
"NLP的有效性不断提高，为跨学科合作创造了",
"to learn about causal relationships.How",
"令人兴奋的新机会，将NLP技术带入了广泛",
"ever, despite its critical role in the life",
"的外部研究学科E(e。g。，Robertsetal.,2014;",
"and social sciences, causality has not had",
"Zhangetal..2020：?）并将新数据和任务纳入主",
"the same importance in Natural Language",
"流 NLP (e。g。，Thomas et al.,2006;?)。在这",
"Processing (NLP), which has traditionally",
"placed more emphasis on predictive tasks.",
"样的跨学科合作中，许多最重要的研究问题都",
"This distinction is beginning to fade, with",
"与因果关系的推断有关。。例如，在推荐新的",
"an emerging area of interdisciplinary re-",
"药物治疗之前，临床医生想知道药物对疾病进",
"search at the convergence of causal infer",
"展的因果影响。因果推断涉及一个关于通过采",
"ence and language processing. Still, re-",
"取干预措施而产生的反事实世界的问题：如果",
"search on causality in NLP remains scat",
"我们给他们服用药物，患者的疾病进展会是什",
"tered across domains without unified def-",
"么？正如我们在下面通过观察数据所解释的那",
"initions, benchmark datasets and clear ar-",
"样，因果关系并不等同于是否服用药物与观察",
"ticulations of the challenges and opportuni",
"ties in the application of causal inference to",
"到的疾病进展之间的相关性。现在有大量关于",
"the textual domain, with its unique proper",
"使用传统（非文本）数据集进行有效推断的技",
"ties. In this survey, we consolidate research",
"术的文献E(e。g。？），但是将这些技术应用",
"across academic areas and situate it in the",
"于自然语言数据会带来新的基本挑战",
"broader NLPlandscape.We introduce the",
"相反，在许多经典的NLP应用程序中，主",
"statistical challenge of estimating causal ef-",
"要目标是做出准确的预测：任何统计相关性",
"fects with text, encompassing settings where",
"都是可接受的，而与潜在的因果关系无关。然",
"text is used as an outcome, treatment, or to",
"而，随着NLP系统越来越多地部署在具有挑",
"address confounding. In addition, we ex",
"plore potential uses of causal inference to",
"战性和高风险的场景中，我们不能依赖通常的",
"improve the robustness, fairness, and inter-",
"假设，即训练和测试数据是相同分布的，我们",
"pretability of NLP models. We thus provide",
"可能不满足于不可解释的黑盒预测器。对于这",
"a unified overview of causal inference for",
"两个问题，因果关系都提供了一条有希望的前",
"the NLP community.",
"进之路：数据生成过程的因果结构的领域知识",
"All authors equally contributed to this paper.",
"Au-",
"可以暗示归纳偏差，从而导致更强大的预测因",
"thor names are organized alphabetically in two clusters:",
"子，而预测因子本身的因果观点可以为其内部",
"First students and post-docs and then faculty members.",
"The email address of the corresponding (first) author is:",
"on causal inference and language processing is avail-",
"[email protected]",
"ablehere:",
"https://github.com/causaltext/",
"'An online repository containing existing research",
"causal-text-papers"
] | {
"arxiv_id": "2109.00725",
"language": "zh",
"page_idx": 1
} |
||
2109.00725-zh-3 | 2109.00725-zh-3.jpg | )。阅读本文后，我们设想读者将对以下方面有广泛的了解：不同类型的因果查询及其带来的挑战；使用文本数据和 NLP 方法所特有的统计和因果挑战；以及在估计文本效果和应用因果关系以改进 NLP 方法方面的开放性问题。
# 2 Background
本调查的两个焦点问题(因果效应估计和稳健且可解释的预测的因果形式主义)都涉及因果推论。因果推断的关键要素是基于感兴趣的干预来定义反事实。我们将通过§1 中的激励示例来说明这个想法
Example 1 涉及在线论坛帖子以及他们收到的喜欢$Y$的数量。我们使用二进制变量$T$来指示帖子是使用“女性图标”($T=1$)还是“男性图标”($T=0$)。在这个例子中，我们将帖子图标$T$视为“治疗”,但不要假设治疗是随机分配的(它可以由帖子的作者选择)。反事实结果$Y(1)$表示如果帖子使用女性图标会收到的喜欢次数。反事实结果$Y(0)$的定义类似
因果推理 (Holland, 1986) 的基本问题是，对于任何分析单位，我们永远无法同时观察$Y(0)$ 和$Y(1)$,这是要进行反事实查询的最小单位(e。g。Ex中的帖子。1)。这个问题使得因果推断比统计推断更难，并且在没有识别假设的情况下是不可能的(请参阅§2.2)
Example 2 涉及一个经过训练的分类器$f(X)$ ,该分类器将文本临床叙述$X$作为输入并输出诊断预测。文本$X$是根据医生的诊断$Y$编写的，并且还受医院$Z$ 使用的书写风格的影响。我们想在固定标签$Y$ 的同时干预医院 $Z$ 。反事实叙述$X(z)$是我们将医院设置为值$z$同时保持诊断固定的文本。反事实预测$f(X(z))$ 是如果我们将反事实评论$X(z)$作为输入，则经过训练的分类器将产生的输出
2.1 Causal Estimands
分析师首先指定感兴趣的目标因果量，称为因果估计，通常涉及反事实。在Example 1中， 一个可能的因果估计是平均治疗效果(ATE) (Rubin, 1974) ,
$$\mathrm A\mathrm T\mathrm E=\mathbb E[Y(1)-Y(0)].$$
(1)
,其中期望超过帖子的生成分布。ATE 可以解释为如果帖子使用女性图标而不是男性图标， 帖子平均会收到的喜欢数量的变化
感兴趣的另一个可能的因果效应是条件平均治疗效果(CATE)(Imbens and Rubin,2015),
(2)
$$\mathrm C\mathrm A\mathrm T\mathrm E=\mathbb E[Y(1)-Y(0)\mid G].$$
,其中$G$ 是人口的预定义子组。例如，$G$可以是有关政治主题的所有帖子。在这种情况下， CATE 可以解释为，如果帖子使用男性图标而不是女性图标，那么关于政治话题的帖子平均会收到的喜欢数量的变化。CATEs 用于量化不同人群亚组中因果效应的异质性
# 2.2 Identification Assumptions for Causal Inference
我们将重点介绍 Example 1 和 Equation (1) 中的 ATE, 以解释因果推理所需的假设。虽然我们专注于 ATE, 但所有因果估计都需要某种形式的相关假设。变量与本节中先前定义的变量相同
Ignorability 要求治疗分配在统计上独立于
反事实结果，
(3)
$$T\perp Y(a)\quad\forall a\in\{0,1\}.$$
请注意，此假设不等同于治疗分配与观察到的结果$Y$之间的独立性。例如，如果可忽略性成
随机治疗分配通过设计保证了可忽略性。例如，我们可以保证 Example 1 中的可忽略性， 方法是掷硬币为每个帖子选择图标，并禁止帖子作者更改它
没有随机治疗分配，可忽略性可能会受到影响治疗状态和潜在结果的混杂因素，变量的侵犯。在 Example 1 中，假设：(i)默认帖子图标为男性，(ii)只有有经验的用户才能根据其性别更改其帖子的图标，(iii)有经验的用户撰写的帖子会获得相对更多的喜欢。在这种情况下，帖子作者的经验是一个混杂因素：具有女性图标的帖子更有可能由有经验的用户撰写， 因此会收到更多的喜欢。在存在混杂因素的情况下，只有当我们假设条件可忽略性
$$T\perp Y(a)\mid X\quad\forall a\in\{0,1\}$$
(4)
时，才有可能进行因果推断，其中$X$ 是一组观察到的变量，条件是确保治疗分配与潜在结果之间的独立性。换句话说，我们可以假设所有的混杂因素都被观察到
Positivity 要求对于混杂因素$X$的所有值，
接受治疗的概率均以 0 和 1 为界：
$$0<\Pr(T=1\mid X=x)<1,\forall x$$
(5)
直观上，积极性要求所研究的每个单位都有被治疗的可能性，并且有未治疗的可能性。随机治疗分配也可以通过设计保证阳性
Consistency 要求在治疗水平$a\in\{0,1\}$观察到的每个研究单位的结果与我们将该单位分 | [
[
"0.12177",
"0.079",
"0.4878",
"0.079",
"0.4878",
"0.08905",
"0.12177",
"0.08905"
],
[
"0.492",
"0.076",
"0.882",
"0.0766",
"0.882",
"0.09143",
"0.492",
"0.0903"
],
[
"0.12177",
"0.0932",
"0.4895",
"0.0932",
"0.4895",
"0.1069",
"0.12177",
"0.1069"
],
[
"0.513",
"0.09265",
"0.881",
"0.09265",
"0.881",
"0.1057",
"0.513",
"0.1057"
],
[
"0.12177",
"0.111",
"0.486",
"0.111",
"0.486",
"0.1217",
"0.12177",
"0.1217"
],
[
"0.513",
"0.1081",
"0.882",
"0.10986",
"0.882",
"0.12354",
"0.513",
"0.1217"
],
[
"0.1184",
"0.1259",
"0.486",
"0.1259",
"0.486",
"0.1395",
"0.1184",
"0.1395"
],
[
"0.5146",
"0.1259",
"0.882",
"0.1259",
"0.882",
"0.1395",
"0.5146",
"0.1395"
],
[
"0.12006",
"0.1401",
"0.486",
"0.1401",
"0.486",
"0.1538",
"0.12006",
"0.1538"
],
[
"0.5156",
"0.1414",
"0.882",
"0.1414",
"0.882",
"0.155",
"0.5156",
"0.155"
],
[
"0.517",
"0.158",
"0.7734",
"0.158",
"0.7734",
"0.1716",
"0.517",
"0.1716"
],
[
"0.1184",
"0.1687",
"0.2612",
"0.1687",
"0.2612",
"0.1829",
"0.1184",
"0.1829"
],
[
"0.5146",
"0.1841",
"0.858",
"0.1841",
"0.858",
"0.1978",
"0.5146",
"0.1978"
],
[
"0.12006",
"0.1948",
"0.4895",
"0.1948",
"0.4895",
"0.2079",
"0.12006",
"0.2079"
],
[
"0.553",
"0.199",
"0.6323",
"0.2007",
"0.6323",
"0.2156",
"0.553",
"0.2131"
],
[
"0.12177",
"0.2102",
"0.4878",
"0.2102",
"0.4878",
"0.2239",
"0.12177",
"0.2239"
],
[
"0.5146",
"0.2203",
"0.882",
"0.2203",
"0.882",
"0.234",
"0.5146",
"0.234"
],
[
"0.12006",
"0.2268",
"0.4878",
"0.2268",
"0.4878",
"0.2405",
"0.12006",
"0.2405"
],
[
"0.517",
"0.2382",
"0.881",
"0.2382",
"0.881",
"0.2488",
"0.517",
"0.2488"
],
[
"0.12006",
"0.2446",
"0.4844",
"0.2446",
"0.4844",
"0.2546",
"0.12006",
"0.2546"
],
[
"0.5156",
"0.253",
"0.882",
"0.253",
"0.882",
"0.2666",
"0.5156",
"0.2666"
],
[
"0.1184",
"0.2588",
"0.2678",
"0.2588",
"0.2678",
"0.2725",
"0.1184",
"0.2725"
],
[
"0.513",
"0.2683",
"0.882",
"0.2695",
"0.882",
"0.284",
"0.513",
"0.2832"
],
[
"0.1377",
"0.2766",
"0.486",
"0.2766",
"0.486",
"0.2874",
"0.1377",
"0.2874"
],
[
"0.12006",
"0.2915",
"0.4895",
"0.2915",
"0.4895",
"0.3047",
"0.12006",
"0.3047"
],
[
"0.513",
"0.2861",
"0.555",
"0.2861",
"0.555",
"0.2998",
"0.513",
"0.2998"
],
[
"0.11755",
"0.306",
"0.486",
"0.306",
"0.486",
"0.3196",
"0.11755",
"0.3196"
],
[
"0.531",
"0.3018",
"0.882",
"0.3018",
"0.882",
"0.3154",
"0.531",
"0.3154"
],
[
"0.11755",
"0.3218",
"0.4878",
"0.3225",
"0.4878",
"0.3374",
"0.11755",
"0.3362"
],
[
"0.5146",
"0.3196",
"0.614",
"0.3196",
"0.614",
"0.3308",
"0.5146",
"0.3308"
],
[
"0.1184",
"0.3384",
"0.4844",
"0.3384",
"0.4844",
"0.3516",
"0.1184",
"0.3516"
],
[
"0.602",
"0.3457",
"0.7983",
"0.3467",
"0.7983",
"0.361",
"0.602",
"0.3599"
],
[
"0.8003",
"0.3508",
"0.846",
"0.3508",
"0.846",
"0.3582",
"0.8003",
"0.3582"
],
[
"0.858",
"0.3474",
"0.8833",
"0.3474",
"0.8833",
"0.361",
"0.858",
"0.361"
],
[
"0.12006",
"0.3557",
"0.4844",
"0.3557",
"0.4844",
"0.3665",
"0.12006",
"0.3665"
],
[
"0.12006",
"0.3718",
"0.4878",
"0.3718",
"0.4878",
"0.3855",
"0.12006",
"0.3855"
],
[
"0.513",
"0.3748",
"0.882",
"0.3772",
"0.882",
"0.3901",
"0.513",
"0.3884"
],
[
"0.12006",
"0.3872",
"0.4895",
"0.3872",
"0.4895",
"0.401",
"0.12006",
"0.401"
],
[
"0.5146",
"0.3926",
"0.882",
"0.3926",
"0.882",
"0.4062",
"0.5146",
"0.4062"
],
[
"0.12177",
"0.405",
"0.1377",
"0.405",
"0.1377",
"0.415",
"0.12177",
"0.415"
],
[
"0.5146",
"0.408",
"0.81",
"0.408",
"0.81",
"0.4229",
"0.5146",
"0.4229"
],
[
"0.1377",
"0.4192",
"0.4895",
"0.4192",
"0.4895",
"0.4329",
"0.1377",
"0.4329"
],
[
"0.534",
"0.4246",
"0.882",
"0.4246",
"0.882",
"0.4382",
"0.534",
"0.4382"
],
[
"0.1184",
"0.4358",
"0.486",
"0.4358",
"0.486",
"0.4495",
"0.1184",
"0.4495"
],
[
"0.517",
"0.4424",
"0.879",
"0.4424",
"0.879",
"0.4531",
"0.517",
"0.4531"
],
[
"0.12006",
"0.452",
"0.4878",
"0.452",
"0.4878",
"0.4656",
"0.12006",
"0.4656"
],
[
"0.5156",
"0.4573",
"0.8833",
"0.4573",
"0.8833",
"0.4702",
"0.5156",
"0.4702"
],
[
"0.11755",
"0.4685",
"0.4878",
"0.466",
"0.4878",
"0.481",
"0.11755",
"0.4827"
],
[
"0.517",
"0.4739",
"0.6265",
"0.4739",
"0.6265",
"0.485",
"0.517",
"0.485"
],
[
"0.12177",
"0.4863",
"0.486",
"0.4863",
"0.486",
"0.4963",
"0.12177",
"0.4963"
],
[
"0.532",
"0.4893",
"0.882",
"0.4893",
"0.882",
"0.503",
"0.532",
"0.503"
],
[
"0.12006",
"0.5005",
"0.392",
"0.5005",
"0.392",
"0.514",
"0.12006",
"0.514"
],
[
"0.517",
"0.505",
"0.8833",
"0.505",
"0.8833",
"0.5186",
"0.517",
"0.5186"
],
[
"0.1377",
"0.5166",
"0.486",
"0.5166",
"0.486",
"0.53",
"0.1377",
"0.53"
],
[
"0.5146",
"0.521",
"0.881",
"0.521",
"0.881",
"0.5337",
"0.5146",
"0.5337"
],
[
"0.1377",
"0.532",
"0.4878",
"0.532",
"0.4878",
"0.546",
"0.1377",
"0.546"
],
[
"0.5156",
"0.5376",
"0.882",
"0.5376",
"0.882",
"0.5503",
"0.5156",
"0.5503"
],
[
"0.12177",
"0.55",
"0.4878",
"0.55",
"0.4878",
"0.56",
"0.12177",
"0.56"
],
[
"0.513",
"0.554",
"0.882",
"0.554",
"0.882",
"0.568",
"0.513",
"0.568"
],
[
"0.12006",
"0.564",
"0.4895",
"0.564",
"0.4895",
"0.5776",
"0.12006",
"0.5776"
],
[
"0.5146",
"0.5693",
"0.882",
"0.5693",
"0.882",
"0.583",
"0.5146",
"0.583"
],
[
"0.12177",
"0.582",
"0.4878",
"0.582",
"0.4878",
"0.593",
"0.12177",
"0.593"
],
[
"0.517",
"0.586",
"0.882",
"0.586",
"0.882",
"0.5996",
"0.517",
"0.5996"
],
[
"0.12177",
"0.5967",
"0.4878",
"0.5967",
"0.4878",
"0.61",
"0.12177",
"0.61"
],
[
"0.5146",
"0.6016",
"0.882",
"0.6016",
"0.882",
"0.615",
"0.5146",
"0.615"
],
[
"0.12006",
"0.6133",
"0.4878",
"0.6133",
"0.4878",
"0.627",
"0.12006",
"0.627"
],
[
"0.5146",
"0.6177",
"0.881",
"0.6177",
"0.881",
"0.6313",
"0.5146",
"0.6313"
],
[
"0.12177",
"0.629",
"0.4878",
"0.629",
"0.4878",
"0.6426",
"0.12177",
"0.6426"
],
[
"0.5156",
"0.6343",
"0.81",
"0.6343",
"0.81",
"0.648",
"0.5156",
"0.648"
],
[
"0.12006",
"0.6455",
"0.3762",
"0.6455",
"0.3762",
"0.659",
"0.12006",
"0.659"
],
[
"0.5884",
"0.661",
"0.8096",
"0.662",
"0.8096",
"0.677",
"0.5884",
"0.676"
],
[
"0.8096",
"0.663",
"0.8833",
"0.663",
"0.8833",
"0.6777",
"0.8096",
"0.6777"
],
[
"0.12177",
"0.673",
"0.3064",
"0.673",
"0.3064",
"0.683",
"0.12177",
"0.683"
],
[
"0.12006",
"0.6914",
"0.4878",
"0.6914",
"0.4878",
"0.705",
"0.12006",
"0.705"
],
[
"0.492",
"0.6978",
"0.511",
"0.6978",
"0.511",
"0.7007",
"0.492",
"0.7007"
],
[
"0.5146",
"0.6924",
"0.882",
"0.6924",
"0.882",
"0.706",
"0.5146",
"0.706"
],
[
"0.12006",
"0.708",
"0.492",
"0.708",
"0.492",
"0.7217",
"0.12006",
"0.7217"
],
[
"0.5146",
"0.709",
"0.879",
"0.709",
"0.879",
"0.7197",
"0.5146",
"0.7197"
],
[
"0.1318",
"0.7246",
"0.4902",
"0.7246",
"0.4902",
"0.7383",
"0.1318",
"0.7383"
],
[
"0.5146",
"0.7246",
"0.882",
"0.7246",
"0.882",
"0.7383",
"0.5146",
"0.7383"
],
[
"0.1184",
"0.7393",
"0.236",
"0.7407",
"0.236",
"0.755",
"0.1184",
"0.7544"
],
[
"0.5156",
"0.741",
"0.6978",
"0.741",
"0.6978",
"0.752",
"0.5156",
"0.752"
],
[
"0.532",
"0.7573",
"0.8833",
"0.7573",
"0.8833",
"0.771",
"0.532",
"0.771"
],
[
"0.2141",
"0.7695",
"0.3955",
"0.7695",
"0.3955",
"0.783",
"0.2141",
"0.783"
],
[
"0.4626",
"0.7695",
"0.517",
"0.7715",
"0.517",
"0.783",
"0.4626",
"0.781"
],
[
"0.511",
"0.7725",
"0.7793",
"0.7725",
"0.7793",
"0.786",
"0.511",
"0.786"
],
[
"0.1301",
"0.7964",
"0.4878",
"0.7974",
"0.4878",
"0.8125",
"0.1301",
"0.811"
],
[
"0.1276",
"0.8027",
"0.1368",
"0.8027",
"0.1368",
"0.808",
"0.1276",
"0.808"
],
[
"0.575",
"0.8003",
"0.8833",
"0.802",
"0.8833",
"0.8164",
"0.575",
"0.8154"
],
[
"0.1234",
"0.8154",
"0.4878",
"0.8154",
"0.4878",
"0.826",
"0.1234",
"0.826"
],
[
"0.12006",
"0.83",
"0.3955",
"0.83",
"0.3955",
"0.8433",
"0.12006",
"0.8433"
],
[
"0.5156",
"0.8306",
"0.881",
"0.8306",
"0.881",
"0.844",
"0.5156",
"0.844"
],
[
"0.1377",
"0.8457",
"0.4895",
"0.8457",
"0.4895",
"0.8594",
"0.1377",
"0.8594"
],
[
"0.5156",
"0.847",
"0.8833",
"0.847",
"0.8833",
"0.861",
"0.5156",
"0.861"
],
[
"0.1184",
"0.861",
"0.4895",
"0.8633",
"0.4895",
"0.877",
"0.1184",
"0.8745"
],
[
"0.5156",
"0.8643",
"0.7925",
"0.8643",
"0.7925",
"0.8774",
"0.5156",
"0.8774"
],
[
"0.532",
"0.88",
"0.882",
"0.88",
"0.882",
"0.8936",
"0.532",
"0.8936"
],
[
"0.1914",
"0.891",
"0.4182",
"0.891",
"0.4182",
"0.905",
"0.1914",
"0.905"
],
[
"0.4224",
"0.8926",
"0.4534",
"0.8955",
"0.4517",
"0.904",
"0.4207",
"0.901"
],
[
"0.4634",
"0.8926",
"0.4895",
"0.8926",
"0.4895",
"0.904",
"0.4634",
"0.904"
],
[
"0.5146",
"0.8955",
"0.881",
"0.8955",
"0.881",
"0.909",
"0.5146",
"0.909"
]
] | [
"）。阅读本文后，我们设想读者将对以下方面",
"，其中G是人口的预定义子组。例如，G可以",
"有广泛的了解：不同类型的因果查询及其带来",
"是有关政治主题的所有帖子。在这种情况下",
"的挑战；使用文本数据和NLP方法所特有的",
"CATE可以解释为，如果帖子使用男性图标而",
"统计和因果挑战；以及在估计文本效果和应用",
"不是女性图标，那么关于政治话题的帖子平均",
"因果关系以改进NLP方法方面的开放性问题。",
"会收到的喜欢数量的变化。CATEs用于量化不",
"同人群亚组中因果效应的异质性",
"2Background",
"2.2Identification Assumptions for Causal",
"本调查的两个焦点问题（因果效应估计和稳健",
"Inference",
"且可解释的预测的因果形式主义）都涉及因果",
"我们将重点介绍Example1和Equation（1）中",
"推论。因果推断的关键要素是基于感兴趣的干",
"的ATE，以解释因果推理所需的假设。虽然我",
"预来定义反事实。我们将通过1中的激励示",
"们专注于ATE，但所有因果估计都需要某种形",
"例来说明这个想法",
"式的相关假设。变量与本节中先前定义的变量",
"Example1涉及在线论坛帖子以及他们收到",
"的喜欢Y的数量。我们使用二进制变量T来",
"相同",
"指示帖子是使用“女性图标”（T=1）还是",
"Ignorability要求治疗分配在统计上独立于",
"“男性图标”（T=0）。在这个例子中，我们将",
"反事实结果，",
"帖子图标T视为“治疗”，但不要假设治疗是",
"T I Y(a) VaE[0,1).",
"",
"(3)",
"随机分配的（它可以由帖子的作者选择）。反",
"事实结果Y(1)表示如果帖子使用女性图标会",
"请注意，此假设不等同于治疗分配与观察到的",
"收到的喜欢次数。反事实结果Y(O)的定义类",
"结果Y之间的独立性。例如，如果可忽略性成",
"似",
"立，Y工T将另外暗示治疗没有效果",
"因果推理（Holland，1986)的基本问题是，对",
"随机治疗分配通过设计保证了可忽略性。例",
"于任何分析单位，我们永远无法同时观察Y(O)",
"如，我们可以保证Example1中的可忽略性",
"和Y（1），这是要进行反事实查询的最小单位",
"方法是掷硬币为每个帖子选择图标，并禁止帖",
"(e。g。Ex中的帖子。1)。这个问题使得因果",
"子作者更改它",
"推断比统计推断更难，并且在没有识别假设的",
"没有随机治疗分配，可忽略性可能会受到影",
"情况下是不可能的（请参阅82.2）",
"响治疗状态和潜在结果的混杂因素，变量的侵",
"Example2涉及一个经过训练的分类器f(X）",
"犯。在Example1中，假设：（i）默认帖子图",
"该分类器将文本临床叙述X作为输入并输",
"标为男性，（ii）只有有经验的用户才能根据其",
"出诊断预测。文本X是根据医生的诊断Y编",
"性别更改其帖子的图标，（ii）有经验的用户撰",
"写的，并且还受医院乙使用的书写风格的影",
"写的帖子会获得相对更多的喜欢。在这种情况",
"响。我们想在固定标签Y的同时干预医院Z。",
"下，帖子作者的经验是一个混杂因素：具有女",
"反事实叙述X（）是我们将医院设置为值之同",
"性图标的帖子更有可能由有经验的用户撰写，",
"时保持诊断固定的文本。反事实预测f(X(2))",
"因此会收到更多的喜欢。在存在混杂因素的情",
"是如果我们将反事实评论X（）作为输入，则",
"况下，只有当我们假设条件可忽略性",
"经过训练的分类器将产生的输出",
"T L Y(a)|X VaE[0,1)",
" (4)",
"2.1CausalEstimands",
"分析师首先指定感兴趣的目标因果量，称为",
"",
"时，才有可能进行因果推断，其中X是一组观",
"因果估计，通常涉及反事实。在Example1中，",
"察到的变量，条件是确保治疗分配与潜在结果",
"个可能的因果估计是平均治疗效果（ATE）",
"之间的独立性。换句话说，我们可以假设所有",
"(Rubin, 1974),",
"的混杂因素都被观察到",
"Positivity要求对于混杂因素X的所有值，",
"ATE = E[Y(1) - Y(O)].",
"(1)",
"接受治疗的概率均以0和1为界：",
"其中期望超过帖子的生成分布。ATE可以解",
"",
"0<Pr(T= 1| X = α) <1, Vα (5)",
"释为如果帖子使用女性图标而不是男性图标，",
"帖子平均会收到的喜欢数量的变化",
"直观上，积极性要求所研究的每个单位都有被",
"感兴趣的另一个可能的因果效应是条件平均",
"治疗的可能性，并且有未治疗的可能性。随机",
"治疗效果（CATE）（ImbensandRubin，2015）：",
"治疗分配也可以通过设计保证阳性",
"Consistency要求在治疗水平aEo,1]观",
"CATE = E[Y(1) - Y(O) [G].",
"",
"(2)",
"察到的每个研究单位的结果与我们将该单位分"
] | {
"arxiv_id": "2109.00725",
"language": "zh",
"page_idx": 3
} |
||
2109.00725-zh-4 | 2109.00725-zh-4.jpg | 配给治疗水平$a$时观察到的结果相同
$$T=a\Leftrightarrow Y(a)=Y\quad\forall a\in\{0,1\}$$
(6)
一致性确保每个研究单位的潜在结果在每个治疗水平上具有单个值。如果治疗的不同不可观察的“版本”导致不同的潜在结果，则将违反一致性。例如，如果红色和蓝色女性图标对收到的喜欢数量有不同的影响，但没有记录图标颜色。如果一个单位的治疗分配影响另一个单位的潜在结果，也将违反一致性；一种称为干扰 (Rosenbaum, 2007) 的现象。随机治疗分配不能保证设计的一致性。e。g。如果不同的图标颜色影响喜欢的数量，但模型不考虑，那么随机实验将无法解决问题。正如 Hernán (2016) 所讨论的那样，一致性假设是“专家协议的问题”,尽管是主观的，但因果形式主义至少使这些假设更加透明
这三个假设可以识别 Equation (1) 中定义的
ATE, 如以下识别证明中所形式化的：
$$\begin{aligned}\mathbb{E}[Y(a)]&\stackrel{(i)}{=}\mathbb{E}_{X}[\mathbb{E}[Y(a)\mid X]]\\&\stackrel{(ii)}{=}\mathbb{E}_{X}[\mathbb{E}[Y(a)\mid X,T=a]]\\&\stackrel{(iii)}{=}\mathbb{E}_{X}[\mathbb{E}[Y\mid X,T=a]],\forall a\in\{0,1\}\end{aligned}$$
其中相等(i)是由于迭代期望，相等(ii)来自条件可忽略性，相等(iii)来自一致性和积极性，这确保了条件期望$\mathbb{E}[Y\mid X,T=a]$的定义良好。最终表达式可以仅从可观测量计算得出
我们参考其他背景材料，讨论如何使用这些假设 (Rubin, 2005; Pearl, 2009; Imbens and Rubin, 2015; Egami et al., 2018; Keith et al., 2020) 识别和估计因果效应
# 2.3 Causal Graphical Models
找到一组确保条件可忽略性的变量$X$具有挑战性，并且需要对所研究领域中的因果关系进行几个仔细评估的假设。因果有向无环图(DAG)(Pearl,2009)可以对这些假设进行正式编码，并在满足可忽略性的条件后得出变量集$X$
在因果 DAG 中，边$X\to Y$表示$X$可能会或可能不会导致$Y$。$X$ 和$Y$ 之间没有边意味着$X$不会导致$Y$ 。变量之间的双向虚线箭头表示它们通过一些未观察到的变量潜在相关
Figure 1 说明了我们为 Example 1 和 Example 2 假设的因果 DAG。给定一个因果 DAG 可以使用 d 分离算法 (Pearl, 1994) 导出任何一对变量之间的因果依赖关系。然后，这些依赖关系可用于评估条件可忽略性是否适用于给定
<FigureHere>
Figure 1: Causal graphs for the motivating examples.( Left ) In Example 1, the post icon $(T)$ is correlated with attributes of the post( $X$ ), and both variables affect the number of likes a post receives $(Y).($ Right ) In Example 2, the label ( $Y$ ,i.e., diagnosis) and hospital site $(Z)$ are correlated, and both affect the clinical narrative $(X$ ). Predictions $f(X)$ from the trained classifier depend on $X$ .
的治疗，结果和一组条件变量$X$ 。例如，在Figure 1 的左侧 DAG 中，帖子图标$T$并不独立于喜欢$Y$的数量，除非我们以$X$ 为条件。在右 DAG 中，即使在对叙事 $X$ 进行调节之后， 预测$f(X)$也不独立于医院$Z$
3 Estimating Causal Effects with Text
在 §2 中，我们描述了直接测量治疗，结果和混杂因素时因果推断的假设。在本节中，我们将对因果分析所需的变量从文本中自动提取时因果假设的复杂性进行新颖的讨论。。解决这些公开挑战将需要 NLP 和因果估计社区之间的合作，以了解得出有效因果结论的必要假设是什么。我们强调了以前的方法和未来的挑战， 在这些情况下，文本是一个混杂因素，结果或治疗-但这个讨论广泛适用于许多基于文本的因果问题
为了明确这些挑战，我们将通过假设一个假设的在线论坛想要了解并减少其平台上的骚扰来扩展 Example 1。许多这样的问题都是因果关系：性别图标会影响用户受到的骚扰吗？更长的暂停时间会使用户不太可能骚扰他人吗？ 如何重写帖子以避免冒犯他人？在每种情况下，使用 NLP 来衡量语言的各个方面都是任何因果分析不可或缺的
3.1 Causal Effects with Textual Confounders
返回 Example 1 , 假设平台担心带有女性图标的用户更有可能受到其他用户的骚扰。这样的发现可能会极大地影响新的节制策略(Jhaver et al., 2018; Rubin et al., 2020) 的计划。我们可能无法或不愿意随机化我们对作者图标的处理(性别信号)),因此性别信号对收到的骚扰的因果影响可能会被其他变量混淆。帖子的主题 | [
[
"0.1234",
"0.079",
"0.4182",
"0.079",
"0.4182",
"0.08905",
"0.1234",
"0.08905"
],
[
"0.1721",
"0.1051",
"0.4358",
"0.1057",
"0.4358",
"0.12054",
"0.1721",
"0.1194"
],
[
"0.465",
"0.10925",
"0.4844",
"0.10925",
"0.4844",
"0.11816",
"0.465",
"0.11816"
],
[
"0.1318",
"0.1354",
"0.486",
"0.1354",
"0.486",
"0.1484",
"0.1318",
"0.1484"
],
[
"0.1184",
"0.152",
"0.4878",
"0.152",
"0.4878",
"0.165",
"0.1184",
"0.165"
],
[
"0.12006",
"0.1687",
"0.4878",
"0.1687",
"0.4878",
"0.1818",
"0.12006",
"0.1818"
],
[
"0.1368",
"0.1853",
"0.4878",
"0.1853",
"0.4878",
"0.199",
"0.1368",
"0.199"
],
[
"0.513",
"0.1882",
"0.881",
"0.1895",
"0.881",
"0.2037",
"0.513",
"0.2031"
],
[
"0.12177",
"0.2007",
"0.4878",
"0.2007",
"0.4878",
"0.2144",
"0.12177",
"0.2144"
],
[
"0.511",
"0.2048",
"0.882",
"0.2048",
"0.882",
"0.2197",
"0.511",
"0.2197"
],
[
"0.1184",
"0.2162",
"0.486",
"0.2162",
"0.486",
"0.2299",
"0.1184",
"0.2299"
],
[
"0.513",
"0.2214",
"0.8833",
"0.2214",
"0.8833",
"0.2351",
"0.513",
"0.2351"
],
[
"0.1184",
"0.2322",
"0.486",
"0.2322",
"0.486",
"0.2452",
"0.1184",
"0.2452"
],
[
"0.511",
"0.2363",
"0.882",
"0.2369",
"0.882",
"0.2517",
"0.511",
"0.2505"
],
[
"0.1184",
"0.2488",
"0.4895",
"0.2488",
"0.4895",
"0.2625",
"0.1184",
"0.2625"
],
[
"0.5146",
"0.2542",
"0.8833",
"0.2542",
"0.8833",
"0.2673",
"0.5146",
"0.2673"
],
[
"0.12006",
"0.2642",
"0.4878",
"0.2642",
"0.4878",
"0.279",
"0.12006",
"0.279"
],
[
"0.513",
"0.2695",
"0.882",
"0.2695",
"0.882",
"0.2832",
"0.513",
"0.2832"
],
[
"0.12006",
"0.2808",
"0.4878",
"0.2808",
"0.4878",
"0.2944",
"0.12006",
"0.2944"
],
[
"0.513",
"0.2861",
"0.882",
"0.2861",
"0.882",
"0.2998",
"0.513",
"0.2998"
],
[
"0.1184",
"0.2957",
"0.486",
"0.2957",
"0.486",
"0.3088",
"0.1184",
"0.3088"
],
[
"0.513",
"0.3018",
"0.882",
"0.3018",
"0.882",
"0.3154",
"0.513",
"0.3154"
],
[
"0.12177",
"0.3123",
"0.4878",
"0.3123",
"0.4878",
"0.326",
"0.12177",
"0.326"
],
[
"0.5146",
"0.3184",
"0.602",
"0.3184",
"0.602",
"0.3333",
"0.5146",
"0.3333"
],
[
"0.12006",
"0.3289",
"0.4878",
"0.3289",
"0.4878",
"0.3425",
"0.12006",
"0.3425"
],
[
"0.12006",
"0.3445",
"0.2678",
"0.3445",
"0.2678",
"0.3582",
"0.12006",
"0.3582"
],
[
"0.1377",
"0.361",
"0.4895",
"0.361",
"0.4895",
"0.3748",
"0.1377",
"0.3748"
],
[
"0.5156",
"0.3599",
"0.8833",
"0.3599",
"0.8833",
"0.3735",
"0.5156",
"0.3735"
],
[
"0.12006",
"0.3772",
"0.4182",
"0.3772",
"0.4182",
"0.3901",
"0.12006",
"0.3901"
],
[
"0.513",
"0.3748",
"0.882",
"0.3748",
"0.882",
"0.3884",
"0.513",
"0.3884"
],
[
"0.5156",
"0.3914",
"0.882",
"0.3914",
"0.882",
"0.405",
"0.5156",
"0.405"
],
[
"0.1973",
"0.4062",
"0.3525",
"0.4092",
"0.3525",
"0.427",
"0.1973",
"0.4233"
],
[
"0.517",
"0.408",
"0.8833",
"0.408",
"0.8833",
"0.4216",
"0.517",
"0.4216"
],
[
"0.5156",
"0.4246",
"0.7515",
"0.4246",
"0.7515",
"0.4382",
"0.5156",
"0.4382"
],
[
"0.12006",
"0.4104",
"0.2091",
"0.408",
"0.2091",
"0.4233",
"0.12006",
"0.4258"
],
[
"0.2024",
"0.4287",
"0.413",
"0.4353",
"0.4114",
"0.4573",
"0.2007",
"0.4507"
],
[
"0.2024",
"0.456",
"0.4895",
"0.462",
"0.4895",
"0.484",
"0.2024",
"0.478"
],
[
"0.513",
"0.452",
"0.858",
"0.452",
"0.858",
"0.4656",
"0.513",
"0.4656"
],
[
"0.5146",
"0.4756",
"0.882",
"0.4768",
"0.882",
"0.4912",
"0.5146",
"0.4905"
],
[
"0.12006",
"0.4934",
"0.4895",
"0.4934",
"0.4895",
"0.5073",
"0.12006",
"0.5073"
],
[
"0.513",
"0.4934",
"0.881",
"0.4934",
"0.881",
"0.5073",
"0.513",
"0.5073"
],
[
"0.1234",
"0.5103",
"0.4895",
"0.5103",
"0.4895",
"0.524",
"0.1234",
"0.524"
],
[
"0.5146",
"0.511",
"0.879",
"0.511",
"0.879",
"0.5215",
"0.5146",
"0.5215"
],
[
"0.12006",
"0.5254",
"0.4895",
"0.5254",
"0.4895",
"0.539",
"0.12006",
"0.539"
],
[
"0.5146",
"0.5254",
"0.882",
"0.5254",
"0.882",
"0.539",
"0.5146",
"0.539"
],
[
"0.12006",
"0.542",
"0.486",
"0.542",
"0.486",
"0.5527",
"0.12006",
"0.5527"
],
[
"0.5156",
"0.542",
"0.882",
"0.542",
"0.882",
"0.5557",
"0.5156",
"0.5557"
],
[
"0.12006",
"0.5586",
"0.1561",
"0.5586",
"0.1561",
"0.571",
"0.12006",
"0.571"
],
[
"0.517",
"0.5586",
"0.879",
"0.5586",
"0.879",
"0.5693",
"0.517",
"0.5693"
],
[
"0.1368",
"0.5737",
"0.4895",
"0.5737",
"0.4895",
"0.5874",
"0.1368",
"0.5874"
],
[
"0.5156",
"0.5747",
"0.882",
"0.5747",
"0.882",
"0.5884",
"0.5156",
"0.5884"
],
[
"0.12006",
"0.5903",
"0.4895",
"0.5903",
"0.4895",
"0.604",
"0.12006",
"0.604"
],
[
"0.5146",
"0.5903",
"0.882",
"0.5903",
"0.882",
"0.604",
"0.5146",
"0.604"
],
[
"0.1184",
"0.6055",
"0.4878",
"0.6055",
"0.4878",
"0.619",
"0.1184",
"0.619"
],
[
"0.5146",
"0.6055",
"0.882",
"0.605",
"0.882",
"0.619",
"0.5146",
"0.6206"
],
[
"0.1184",
"0.6216",
"0.2854",
"0.6216",
"0.2854",
"0.6353",
"0.1184",
"0.6353"
],
[
"0.517",
"0.6235",
"0.5884",
"0.6235",
"0.5884",
"0.6353",
"0.517",
"0.6353"
],
[
"0.534",
"0.638",
"0.882",
"0.638",
"0.882",
"0.653",
"0.534",
"0.653"
],
[
"0.11755",
"0.648",
"0.367",
"0.6484",
"0.367",
"0.663",
"0.11755",
"0.662"
],
[
"0.5146",
"0.653",
"0.8833",
"0.654",
"0.8833",
"0.6685",
"0.5146",
"0.6675"
],
[
"0.12006",
"0.6694",
"0.4895",
"0.6694",
"0.4895",
"0.683",
"0.12006",
"0.683"
],
[
"0.5156",
"0.6704",
"0.882",
"0.6704",
"0.882",
"0.684",
"0.5156",
"0.684"
],
[
"0.1184",
"0.685",
"0.486",
"0.685",
"0.486",
"0.6987",
"0.1184",
"0.6987"
],
[
"0.5156",
"0.687",
"0.882",
"0.687",
"0.882",
"0.7007",
"0.5156",
"0.7007"
],
[
"0.12006",
"0.7017",
"0.4878",
"0.7017",
"0.4878",
"0.7153",
"0.12006",
"0.7153"
],
[
"0.5156",
"0.703",
"0.8833",
"0.703",
"0.8833",
"0.7163",
"0.5156",
"0.7163"
],
[
"0.1184",
"0.7188",
"0.4895",
"0.7188",
"0.4895",
"0.7324",
"0.1184",
"0.7324"
],
[
"0.5156",
"0.7197",
"0.882",
"0.7197",
"0.882",
"0.733",
"0.5156",
"0.733"
],
[
"0.12006",
"0.7354",
"0.4844",
"0.7354",
"0.4844",
"0.745",
"0.12006",
"0.745"
],
[
"0.5156",
"0.7354",
"0.882",
"0.7354",
"0.882",
"0.749",
"0.5156",
"0.749"
],
[
"0.12177",
"0.752",
"0.1368",
"0.752",
"0.1368",
"0.7627",
"0.12177",
"0.7627"
],
[
"0.5156",
"0.752",
"0.6987",
"0.752",
"0.6987",
"0.7656",
"0.5156",
"0.7656"
],
[
"0.1377",
"0.7666",
"0.486",
"0.7666",
"0.486",
"0.78",
"0.1377",
"0.78"
],
[
"0.5146",
"0.7793",
"0.5615",
"0.7793",
"0.5615",
"0.7905",
"0.5146",
"0.7905"
],
[
"0.5576",
"0.778",
"0.8833",
"0.778",
"0.8833",
"0.7915",
"0.5576",
"0.7915"
],
[
"0.1184",
"0.799",
"0.4878",
"0.799",
"0.4878",
"0.8125",
"0.1184",
"0.8125"
],
[
"0.12006",
"0.783",
"0.4895",
"0.783",
"0.4895",
"0.7964",
"0.12006",
"0.7964"
],
[
"0.513",
"0.799",
"0.882",
"0.799",
"0.882",
"0.8125",
"0.513",
"0.8125"
],
[
"0.12006",
"0.8154",
"0.471",
"0.8154",
"0.471",
"0.829",
"0.12006",
"0.829"
],
[
"0.5146",
"0.814",
"0.881",
"0.814",
"0.881",
"0.8276",
"0.5146",
"0.8276"
],
[
"0.1368",
"0.832",
"0.4878",
"0.832",
"0.4878",
"0.8457",
"0.1368",
"0.8457"
],
[
"0.5146",
"0.83",
"0.882",
"0.832",
"0.882",
"0.8457",
"0.5146",
"0.8433"
],
[
"0.1184",
"0.847",
"0.4895",
"0.847",
"0.4895",
"0.861",
"0.1184",
"0.861"
],
[
"0.513",
"0.847",
"0.882",
"0.847",
"0.882",
"0.861",
"0.513",
"0.861"
],
[
"0.12177",
"0.8633",
"0.4902",
"0.8633",
"0.4902",
"0.877",
"0.12177",
"0.877"
],
[
"0.5156",
"0.8633",
"0.882",
"0.8633",
"0.882",
"0.877",
"0.5156",
"0.877"
],
[
"0.12006",
"0.8774",
"0.4895",
"0.879",
"0.4895",
"0.8936",
"0.12006",
"0.8926"
],
[
"0.5146",
"0.88",
"0.882",
"0.88",
"0.882",
"0.8936",
"0.5146",
"0.8936"
],
[
"0.1234",
"0.898",
"0.4878",
"0.898",
"0.4878",
"0.9077",
"0.1234",
"0.9077"
],
[
"0.5146",
"0.8955",
"0.882",
"0.8955",
"0.882",
"0.909",
"0.5146",
"0.909"
]
] | [
"配给治疗水平时观察到的结果相同",
"T=a台Y(a)=Y VaE[0,1)",
"(6",
"-致性确保每个研究单位的潜在结果在每个治",
"疗水平上具有单个值。如果治疗的不同不可观",
"察的“版本”导致不同的潜在结果，则将违反",
"致性。例如，如果红色和蓝色女性图标对收",
"Figure 1: Causal graphs for the motivating exam-",
"到的喜欢数量有不同的影响，但没有记录图标",
"ples. (Left ) In Example 1 , the post icon (T)",
"颜色。如果一个单位的治疗分配影响另一个单",
"is correlated with attributes of the post ( X ), and",
"位的潜在结果，也将违反一致性；一种称为干",
"both variables affect the number of likes a post re-",
"扰（Rosenbaum，2007）的现象。随机治疗分配",
"ceives（Y).（Right) In Example 2, the label (",
"不能保证设计的一致性。e。g。如果不同的图",
"Y, i.e., diagnosis) and hospital site (Z ) are cor-",
"标颜色影响喜欢的数量，但模型不考虑，那么",
"related, and both affect the clinical narrative ( X",
"随机实验将无法解决问题。正如Hernan（2016)",
"). Predictions f (X) from the trained classifier de-",
"所讨论的那样，一致性假设是“专家协议的问",
"pend on X",
"题”，尽管是主观的，但因果形式主义至少使",
"这些假设更加透明",
"这三个假设可以识别Equation（1）中定义的",
"的治疗，结果和一组条件变量X。例如，在",
"ATE，如以下识别证明中所形式化的：",
"Figure1的左侧DAG中，帖子图标T并不独立",
"于喜欢Y的数量，除非我们以X为条件。在",
"旦Ex[E[Y(a) [ X]",
"右DAG中，即使在对叙事X进行调节之后，",
"预测f(X)也不独立于医院Z",
"E[Y(a)]",
"( Ex[E[Y(a) / X, T = al]",
"() Ex[E[Y |X, T=a], Va E [0, 1]",
"3Estimating Causal Effects withText",
"在82中，我们描述了直接测量治疗，结果和混",
"其中相等（i）是由于迭代期望，相等（ii）来",
"杂因素时因果推断的假设。在本节中，我们将",
"自条件可忽略性，相等（ii）来自一致性和积",
"对因果分析所需的变量从文本中自动提取时因",
"极性，这确保了条件期望E[Y「X,T=a]的",
"果假设的复杂性进行新颖的讨论。。解决这些",
"定义良好。最终表达式可以仅从可观测量计算",
"公开挑战将需要NLP和因果估计社区之间的",
"得出",
"合作，以了解得出有效因果结论的必要假设是",
"我们参考其他背景材料，讨论如何使用这些",
"什么。我们强调了以前的方法和未来的挑战，",
"假设(Rubin,2005;Pearl,2009;Imbens and Ru-",
"在这些情况下，文本是一个混杂因素，结果或",
"bin, 2015; Egami et al., 2018; Keith et al., 2020)",
"治疗-但这个讨论广泛适用于许多基于文本的",
"识别和估计因果效应",
"因果问题",
"为了明确这些挑战，我们将通过假设一个假",
"2.3Causal Graphical Models",
"设的在线论坛想要了解并减少其平台上的骚扰",
"找到一组确保条件可忽略性的变量X具有挑",
"来扩展Example1。许多这样的问题都是因果",
"战性，并且需要对所研究领域中的因果关系",
"关系：性别图标会影响用户受到的骚扰吗？更",
"进行几个仔细评估的假设。因果有向无环图",
"长的暂停时间会使用户不太可能骚扰他人吗？",
"(DAG）（Pearl,2009）可以对这些假设进行正式",
"如何重写帖子以避免冒犯他人？在每种情况",
"编码，并在满足可勿略性的条件后得出变量集",
"下，使用NLP来衡量语言的各个方面都是任",
"X",
"何因果分析不可或缺的",
"在因果DAG中，边X一→Y表示X可能会",
"3.1",
" Causal Effects with Textual Confounders",
"着X不会导致Y。变量之间的双向虚线箭头",
"或可能不会导致Y。X和Y之间没有边意味",
"返回Example1，假设平台担心带有女性图标",
"表示它们通过一些未观察到的变量潜在相关",
"的用户更有可能受到其他用户的骚扰。这样的",
"Figure1说明了我们为Example1和Exam-",
"发现可能会极大地影响新的节制策略（Jhaver",
"ple 2假设的因果DAG。给定一个因果DAG,",
"etal.,2018;Rubin etal.,2020）的计划。我们可",
"可以使用d分离算法(Pearl,1994)导出任何一",
"能无法或不愿意随机化我们对作者图标的处理",
"对变量之间的因果依赖关系。然后，这些依赖",
"(性别信号)，因此性别信号对收到的骚扰的因",
"关系可用于评估条件可忽略性是否适用于给定",
"果影响可能会被其他变量混淆。帖子的主题"
] | {
"arxiv_id": "2109.00725",
"language": "zh",
"page_idx": 4
} |
||
2010.04925-zh-0 | 2010.04925-zh-0.jpg | Regularizing Neural Networks via Adversarial Model Perturbation
Yaowei Zheng
BDBC and SKLSDE
Beihang University, China
[email protected]
Richong Zhang$^*$
BDBC and SKLSDE
Beihang University, China
[email protected]
Yongyi Mao School of EECS
University of Ottawa, Canada
[email protected]
## 1. Introduction
迄今为止，深度神经网络的泛化行为仍然是一个谜， 尽管最近取得了一些进展(请参阅 e.g. ,[1,3,19,24, 26,48,50] )。在这方面，一个普遍接受和经验验证的理解是，与经验风险的平坦最小值相对应的模型参数倾向于更好地概括。例如，[19,26] 的作者认为平坦极小值对应于简单模型，这些模型不太可能过度拟合。这种理解激发了人们对优化过程中影响所发现的最小值[14,20,25]的平坦性的因素 (例如学习率和批处理大小)的研究，以便更好地理解深度网络的泛化行为。
同时，开发有效的正则化技术仍然是实践中最重要的方法，以减轻过度拟合和强制模型以实现更好的泛化 (e.g. , [16,28,29,40,42,51])。事实上，最近的一
$^{*}$Corresponding author ${\mathrm{Accepted~to~CVPR~2021}}$
# Abstract
Effective regularization techniques are highly desired in deep learning for alleviating overfitting and improving generalization. This work proposes a new regularization scheme, based on the understanding that the flat local minima of the empirical risk cause the model to generalize better. This scheme is referred to as adversarial model perturbation (AMP), where instead of directly minimizing the empirical risk, an alternative“AMP loss” is minimized via SGD. Specifically, the AMP loss is obtained from the empirical risk by applying the“worst”norm-bounded perturbation on each point in the parameter space. Comparing with most existing regularization schemes, AMP has strong theoretical justifications, in that minimizing the AMP loss can be shown theoretically to favour flat local minima of the empirical risk. Extensive experiments on various modern deep architectures establish AMP as a new state of the art among regularization schemes. Our code is available at https://github.com/hiyouga/AMP- $\operatorname{Regularizer}.$
<FigureHere>
Figure 1: An example showing an empirical risk curve ( left) and its corresponding AMP loss curve ( right, $\dot{\mathrm{blue}}).$
些研究表明，某些正则化技术的有效性是由于它们能
够找到更平坦的最小值[22,45]
此外，近年来在开发更有效的正则化方案方面取得了重大进展，其中包括 MixUp, Flooding [22,51] 。尽管取得了巨大的成功，但这些技术通常缺乏强有力的原则或理论依据。因此，人们期望尚未发现更有原则和更强大的正则化方案
这项工作着手在寻找经验风险的平坦局部最小值的原则下开发一个强大的正则化方案。为此，我们提出了一种新新的正则化方案.该方案可以在寻找半姐最小值的能力方面得到充分的证明。。此方案称为 Adversarial Model Perturbation 或 AMP,其中不是使经验风险$\mathcal{L}_{\mathrm{ERM}}(\boldsymbol{\theta})$相对于模型参数$\theta$最小化，而是使替代的“AMP 损失”最小化。简而言之，在参数设置$\theta$下的AMP 损耗$\mathcal{L}_\mathrm{AMP}(\boldsymbol{\theta})$是$\boldsymbol{\theta}$的所有扰动中最差(或最高) 的经验风险，其扰动范数不大于小值$\epsilon$,即
(1)
$$\mathcal{L}_{\mathrm{AMP}}(\boldsymbol{\theta}):=\max_{\Delta:\|\Delta\|\leq\epsilon}\mathcal{L}_{\mathrm{ERM}}(\boldsymbol{\theta}+\Delta)$$
要了解为什么最小化 AMP 损失提供了找到经验风险的平坦局部最小值的机会，请考虑图E1 中的示例。图 1(左)画出了一条经验风险曲线$\mathcal{L}_\mathrm{ERM}$,它包含两个局部极小值，左边是一个尖锐的极小值，右边是一个平坦的极小值。从经验风险中获得 AMP 损失的过程可以看作是一个“最大池化”操作，它在参数空间上滑动一个宽度为 2$\epsilon$的窗口(在高维度上，更准确地说， 是一个半径为$\epsilon$的球体), 并在每个位置返回窗口内的最大值 (resp。sphere)。由此产生的 AMP 损失如图 1
1 | [
[
"0.152",
"0.1332",
"0.819",
"0.132",
"0.819",
"0.1497",
"0.152",
"0.1509"
],
[
"0.1487",
"0.1799",
"0.2695",
"0.1818",
"0.2695",
"0.1995",
"0.1487",
"0.1976"
],
[
"0.4043",
"0.1818",
"0.5415",
"0.1818",
"0.5415",
"0.1995",
"0.4043",
"0.1995"
],
[
"0.694",
"0.1818",
"0.8",
"0.1818",
"0.8",
"0.1995",
"0.694",
"0.1995"
],
[
"0.3816",
"0.1989",
"0.5596",
"0.1995",
"0.5596",
"0.214",
"0.3816",
"0.2134"
],
[
"0.6787",
"0.1995",
"0.814",
"0.1995",
"0.814",
"0.214",
"0.6787",
"0.214"
],
[
"0.357",
"0.2153",
"0.584",
"0.2166",
"0.584",
"0.2343",
"0.357",
"0.2329"
],
[
"0.1217",
"0.2008",
"0.2998",
"0.2008",
"0.2998",
"0.214",
"0.1217",
"0.214"
],
[
"0.10046",
"0.2184",
"0.3218",
"0.2184",
"0.3218",
"0.2329",
"0.10046",
"0.2329"
],
[
"0.621",
"0.2184",
"0.8716",
"0.2184",
"0.8716",
"0.2329",
"0.621",
"0.2329"
],
[
"0.1381",
"0.2374",
"0.282",
"0.2374",
"0.282",
"0.252",
"0.1381",
"0.252"
],
[
"0.3857",
"0.2374",
"0.5557",
"0.2374",
"0.5557",
"0.252",
"0.3857",
"0.252"
],
[
"0.6807",
"0.2374",
"0.8037",
"0.2361",
"0.8037",
"0.2507",
"0.6807",
"0.252"
],
[
"0.2361",
"0.2854",
"0.312",
"0.2874",
"0.312",
"0.303",
"0.2361",
"0.3005"
],
[
"0.5513",
"0.2966",
"0.847",
"0.2942",
"0.847",
"0.3567",
"0.5513",
"0.3591"
],
[
"0.0989",
"0.318",
"0.468",
"0.3193",
"0.468",
"0.334",
"0.0989",
"0.3333"
],
[
"0.0809",
"0.3372",
"0.4656",
"0.3372",
"0.4656",
"0.348",
"0.0809",
"0.348"
],
[
"0.0809",
"0.351",
"0.468",
"0.351",
"0.468",
"0.3655",
"0.0809",
"0.3655"
],
[
"0.535",
"0.3655",
"0.5884",
"0.3655",
"0.5884",
"0.3726",
"0.535",
"0.3726"
],
[
"0.6045",
"0.363",
"0.653",
"0.363",
"0.653",
"0.3738",
"0.6045",
"0.3738"
],
[
"0.7466",
"0.3774",
"0.7656",
"0.3774",
"0.7656",
"0.385",
"0.7466",
"0.385"
],
[
"0.0825",
"0.3674",
"0.4673",
"0.3674",
"0.4673",
"0.3774",
"0.0825",
"0.3774"
],
[
"0.603",
"0.379",
"0.6113",
"0.379",
"0.6113",
"0.3833",
"0.603",
"0.3833"
],
[
"0.8203",
"0.3774",
"0.841",
"0.3774",
"0.841",
"0.3865",
"0.8203",
"0.3865"
],
[
"0.0801",
"0.3958",
"0.468",
"0.3945",
"0.468",
"0.4092",
"0.0801",
"0.4104"
],
[
"0.0801",
"0.3806",
"0.4697",
"0.3806",
"0.4697",
"0.3945",
"0.0801",
"0.3945"
],
[
"0.5034",
"0.4028",
"0.8896",
"0.4028",
"0.8896",
"0.4136",
"0.5034",
"0.4136"
],
[
"0.0809",
"0.4124",
"0.4673",
"0.4124",
"0.4673",
"0.4224",
"0.0809",
"0.4224"
],
[
"0.5034",
"0.4167",
"0.8906",
"0.4167",
"0.8906",
"0.43",
"0.5034",
"0.43"
],
[
"0.0801",
"0.4255",
"0.4714",
"0.4255",
"0.4714",
"0.4395",
"0.0801",
"0.4395"
],
[
"0.5024",
"0.4287",
"0.546",
"0.4312",
"0.544",
"0.4463",
"0.501",
"0.4443"
],
[
"0.0801",
"0.44",
"0.4673",
"0.4414",
"0.4673",
"0.455",
"0.0801",
"0.4539"
],
[
"0.0801",
"0.4558",
"0.468",
"0.4558",
"0.468",
"0.4697",
"0.0801",
"0.4697"
],
[
"0.0801",
"0.4717",
"0.468",
"0.4717",
"0.468",
"0.4849",
"0.0801",
"0.4849"
],
[
"0.5034",
"0.4749",
"0.8896",
"0.4749",
"0.8896",
"0.488",
"0.5034",
"0.488"
],
[
"0.0801",
"0.486",
"0.468",
"0.4849",
"0.468",
"0.5",
"0.0801",
"0.5005"
],
[
"0.5034",
"0.4893",
"0.7236",
"0.4893",
"0.7236",
"0.5024",
"0.5034",
"0.5024"
],
[
"0.0809",
"0.502",
"0.468",
"0.502",
"0.468",
"0.515",
"0.0809",
"0.515"
],
[
"0.524",
"0.505",
"0.8906",
"0.505",
"0.8906",
"0.518",
"0.524",
"0.518"
],
[
"0.0801",
"0.5166",
"0.468",
"0.515",
"0.468",
"0.529",
"0.0801",
"0.53"
],
[
"0.505",
"0.5195",
"0.8906",
"0.5195",
"0.8906",
"0.5327",
"0.505",
"0.5327"
],
[
"0.0809",
"0.5327",
"0.468",
"0.5327",
"0.468",
"0.5435",
"0.0809",
"0.5435"
],
[
"0.505",
"0.535",
"0.8896",
"0.535",
"0.8896",
"0.549",
"0.505",
"0.549"
],
[
"0.0801",
"0.547",
"0.4697",
"0.547",
"0.4697",
"0.561",
"0.0801",
"0.561"
],
[
"0.505",
"0.5503",
"0.8906",
"0.5503",
"0.8906",
"0.5645",
"0.505",
"0.5645"
],
[
"0.0784",
"0.561",
"0.468",
"0.56",
"0.468",
"0.5747",
"0.0784",
"0.5757"
],
[
"0.5024",
"0.563",
"0.652",
"0.5645",
"0.652",
"0.579",
"0.5024",
"0.5776"
],
[
"0.0801",
"0.577",
"0.468",
"0.577",
"0.468",
"0.5903",
"0.0801",
"0.5903"
],
[
"0.523",
"0.5806",
"0.8906",
"0.5806",
"0.8906",
"0.5947",
"0.523",
"0.5947"
],
[
"0.0784",
"0.5913",
"0.468",
"0.5903",
"0.468",
"0.605",
"0.0784",
"0.606"
],
[
"0.5034",
"0.595",
"0.888",
"0.595",
"0.888",
"0.6094",
"0.5034",
"0.6094"
],
[
"0.0825",
"0.608",
"0.1732",
"0.608",
"0.1732",
"0.619",
"0.0825",
"0.619"
],
[
"0.5186",
"0.6113",
"0.8906",
"0.6113",
"0.8906",
"0.625",
"0.5186",
"0.625"
],
[
"0.5034",
"0.6255",
"0.8906",
"0.6255",
"0.8906",
"0.6396",
"0.5034",
"0.6396"
],
[
"0.5034",
"0.64",
"0.8906",
"0.64",
"0.8906",
"0.654",
"0.5034",
"0.654"
],
[
"0.0809",
"0.655",
"0.2173",
"0.655",
"0.2173",
"0.6685",
"0.0809",
"0.6685"
],
[
"0.505",
"0.656",
"0.8906",
"0.656",
"0.8906",
"0.6694",
"0.505",
"0.6694"
],
[
"0.5024",
"0.6694",
"0.888",
"0.6694",
"0.888",
"0.683",
"0.5024",
"0.683"
],
[
"0.10046",
"0.6797",
"0.468",
"0.6797",
"0.468",
"0.6943",
"0.10046",
"0.6943"
],
[
"0.5034",
"0.685",
"0.892",
"0.685",
"0.892",
"0.6997",
"0.5034",
"0.6997"
],
[
"0.0825",
"0.696",
"0.4673",
"0.696",
"0.4673",
"0.709",
"0.0825",
"0.709"
],
[
"0.5034",
"0.702",
"0.8164",
"0.702",
"0.8164",
"0.7153",
"0.5034",
"0.7153"
],
[
"0.0825",
"0.712",
"0.267",
"0.712",
"0.267",
"0.7227",
"0.0825",
"0.7227"
],
[
"0.2615",
"0.712",
"0.4656",
"0.712",
"0.4656",
"0.7227",
"0.2615",
"0.7227"
],
[
"0.0809",
"0.726",
"0.4673",
"0.726",
"0.4673",
"0.7393",
"0.0809",
"0.7393"
],
[
"0.5728",
"0.7266",
"0.8228",
"0.728",
"0.8228",
"0.7456",
"0.5728",
"0.744"
],
[
"0.8677",
"0.73",
"0.892",
"0.73",
"0.892",
"0.7437",
"0.8677",
"0.7437"
],
[
"0.0825",
"0.741",
"0.4673",
"0.741",
"0.4673",
"0.755",
"0.0825",
"0.755"
],
[
"0.664",
"0.741",
"0.7197",
"0.741",
"0.7197",
"0.752",
"0.664",
"0.752"
],
[
"0.0825",
"0.756",
"0.468",
"0.756",
"0.468",
"0.7695",
"0.0825",
"0.7695"
],
[
"0.0809",
"0.7715",
"0.468",
"0.7715",
"0.468",
"0.785",
"0.0809",
"0.785"
],
[
"0.524",
"0.7666",
"0.892",
"0.7666",
"0.892",
"0.78",
"0.524",
"0.78"
],
[
"0.0809",
"0.786",
"0.468",
"0.786",
"0.468",
"0.8",
"0.0809",
"0.8"
],
[
"0.505",
"0.781",
"0.8906",
"0.781",
"0.8906",
"0.7944",
"0.505",
"0.7944"
],
[
"0.5024",
"0.7935",
"0.8906",
"0.7944",
"0.8906",
"0.8125",
"0.5024",
"0.811"
],
[
"0.0809",
"0.802",
"0.4453",
"0.802",
"0.4453",
"0.815",
"0.0809",
"0.815"
],
[
"0.505",
"0.811",
"0.876",
"0.811",
"0.876",
"0.8247",
"0.505",
"0.8247"
],
[
"0.1013",
"0.8174",
"0.4673",
"0.8174",
"0.4673",
"0.8306",
"0.1013",
"0.8306"
],
[
"0.505",
"0.8257",
"0.8906",
"0.8257",
"0.8906",
"0.839",
"0.505",
"0.839"
],
[
"0.0809",
"0.8335",
"0.4673",
"0.8335",
"0.4673",
"0.8467",
"0.0809",
"0.8467"
],
[
"0.5034",
"0.8413",
"0.8906",
"0.8413",
"0.8906",
"0.855",
"0.5034",
"0.855"
],
[
"0.0801",
"0.8477",
"0.4534",
"0.8467",
"0.4534",
"0.8613",
"0.0801",
"0.8623"
],
[
"0.4478",
"0.851",
"0.4617",
"0.851",
"0.4617",
"0.8594",
"0.4478",
"0.8594"
],
[
"0.505",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.505",
"0.8706"
],
[
"0.0989",
"0.8755",
"0.2385",
"0.8755",
"0.2385",
"0.8896",
"0.0989",
"0.8896"
],
[
"0.505",
"0.872",
"0.8906",
"0.872",
"0.8906",
"0.8853",
"0.505",
"0.8853"
],
[
"0.1013",
"0.887",
"0.2524",
"0.887",
"0.2524",
"0.901",
"0.1013",
"0.901"
],
[
"0.5024",
"0.886",
"0.892",
"0.8853",
"0.892",
"0.8994",
"0.5024",
"0.901"
],
[
"0.4805",
"0.925",
"0.492",
"0.925",
"0.492",
"0.936",
"0.4805",
"0.936"
]
] | [
"Regularizing Neural Networks via Adversarial Model Perturbation",
"Yaowei Zheng",
"Richong Zhang*",
"Yongyi Mao",
"BDBC and SKLSDE",
"School of EECS",
"Beihang University, China",
"BDBC and SKLSDE",
"Beihang University, China",
"University of Ottawa,Canada",
"[email protected]",
"[email protected]",
"[email protected]",
"Abstract",
"",
"Effective regularization techniques are highly de",
"sired in deep learning for alleviating overfitting and im",
"proving generalization. This work proposes a new reg-",
"shamminimum",
"flat minimur",
"OERA",
"ularization scheme,based on the understanding that",
"",
"OAMP",
"model to generalize better. This scheme is referred",
"the flat local minima of the empirical risk cause the",
"Figure l: An example showing an empirical risk curve",
"toasadversarialmodel perturbation (AMP),wherein",
"(left) and its corresponding AMP loss curve (right,",
"stead of directly minimizing the empirical risk, an alter-",
"blue).",
"native\"AMP loss\"is minimized via SGD. Specifically,",
"the AMP loss is obtained from the empirical risk by",
"applying the“worst\"norm-bounded perturbation on",
"些研究表明，某些正则化技术的有效性是由于它们能",
"eachpoint in the parameter space.Comparingwith",
"够找到更平坦的最小值[22，45]",
"most existing regularization schemes, AMP has strong",
"此外，近年来在开发更有效的正则化方案方面取得",
"theoretical justifications,inthatminimizingtheAMP",
"了重大进展，其中包括MixUp，Flooding[22，51]。尽",
"losscanbeshowntheoreticallytofavourflatlocal min-",
"管取得了巨大的成功，但这些技术通常缺乏强有力的",
"ima of the empirical risk. Extensive experiments on",
"原则或理论依据。因此，人们期望尚未发现更有原则和",
"various modern deep architectures establish AMP as a",
"更强大的正则化方案",
"new state of the art among regularization schemes. Our",
"这项工作着手在寻找经验风险的平坦局部最小值的",
"codeis available at https://github.com/hiyouga/AMP",
"原则下开发一个强大的正则化方案。为此，我们提出了",
"Regularizer.",
"种新颖的正则化方案，该方案可以在寻找平坦最小值",
"的能力方面得到充分的证明。。此方案称为Adversarial",
"ModelPerturbation或AMP，其中不是使经验风险",
"1.Introduction",
"CERM(O）相对于模型参数最小化，而是使替代的",
"“AMP损失”最小化。简而言之，在参数设置θ下的",
"迄今为止，深度神经网络的泛化行为仍然是一个谜，",
"AMP损耗LAMP(O）是θ的所有扰动中最差（或最高）",
"尽管最近取得了一些进展（请参阅e.g.，[1,3，19,24,",
"的经验风险，其扰动范数不大于小值e，即",
"26，48，50）。在这方面，",
"一个普遍接受和经验验证的",
"理解是，与经验风险的平坦最小值相对应的模型参数",
"CAMP(0) := _max CERM(0 +△)",
"(1)",
"倾向于更好地概括。例如，[19，26]的作者认为平坦极",
"A:lA≤e",
"小值对应于简单模型，这些模型不太可能过度拟合。这",
"种理解激发了人们对优化过程中影响所发现的最小值",
"要了解为什么最小化AMP损失提供了找到经验风",
"[14,20,25]的平坦性的因素（例如学习率和批处理大",
"险的平坦局部最小值的机会，请考虑图E1中的示例。",
"图1（左）画出了一条经验风险曲线CERM，它包含",
"小）的研究，以便更好地理解深度网络的泛化行为。",
"两个局部极小值，左边是一个尖锐的极小值，右边是",
"同时，开发有效的正则化技术仍然是实践中最重要",
"个平坦的极小值。从经验风险中获得AMP损失的过",
"的方法，以减轻过度拟合和强制模型以实现更好的泛",
"程可以看作是一个“最大池化”操作，它在参数空间上",
"化（e.g，，[16,28，29，40，42,51]）。事实上，最近的",
"1-",
"滑动一个宽度为2e的窗口（在高维度上，更准确地说，",
"*Corresponding author",
"是一个半径为e的球体），并在每个位置返回窗口内的",
"Accepted to CVPR 2021",
"最大值(resp。sphere)。由此产生的AMP损失如图 1",
"1"
] | {
"arxiv_id": "2010.04925",
"language": "zh",
"page_idx": 0
} |
||
2010.04925-zh-1 | 2010.04925-zh-1.jpg | <FigureHere>
Figure 2: Landscapes of the empirical risks obtained from the PreActResNet18 [18] models trained with ERM (red) and AMP (blue) on CIFAR-10. Left: on training set; right: on test set. These curves are computed using the technique presented in [31] , where $\delta$ indicates a random direction and $\alpha$ is a displacement in that direction.
(右)中的蓝色曲线所示。由于 AMP 损耗中的右最小值低于左最小值，因此将 AMP 损耗最小化会给出右最小值作为其解决方案
在本文中，我们正式分析了 AMP 损耗最小化问题及其对平坦局部最小值的偏好。具体来说，我们表明该最小化问题隐式地使用局部最小值的“最窄宽度”作为平坦度的概念，并试图惩罚在这种意义上不平坦的最小值
我们推导了一种用于解决此最小化问题的小批量SGD 算法，该算法产生了建议的 AMP 正则化方案。有趣的是，我们证明了该算法也可以看作是在梯度范数上具有附加惩罚项的常规经验风险最小化。这提供了 AMP 方案的另一种理由。图 2 包含一个实验结果， 表明 AMP 确实选择了比 ERM 更平坦的最小值
我们在几个基准图像分类数据集(SVHN,CIFAR- 10, CIFAR-100)上进行实验，以验证所提出的 AMP 方案的有效性。与其他流行的正则化方案相比，AMP 表现出显著的正则化性能，将自己确立为一种新的技术
我们的贡献可以总结如下
1)出于对平坦最小值有助于泛化的理解，我们提出了对抗模型扰动(AMP)作为一种新颖有效的正则化方案
2)我们从理论上证明 AMP 能够找到更平坦的局部
最小值，从而改善了泛化
3)在基准数据集上进行的广泛实验表明，AMP 在各种现代神经网络体系结构上的比较正则化方案中实现了最佳性能
## 2. Related Work
2.1. Flat Minima and Generalization
已经有大量的工作研究了局部极小值的平坦度与深度神经网络[4,19,26,31]的泛化之间的关系。[19] 建议平坦极小值对应于低复杂度网络，在最小描述长度[38]的原理下，它倾向于很好地推广。[4]通过吉布斯自由能的视角提出了另一种支持这一论点的解释。[31] 通
过可视化损失景观展示了平坦最小值的更好泛化。[26] 中的经验结果表明，大批量 SGD 会找到尖锐的最小值，而小批量 SGD 会导致更平坦的最小值并提供更好的泛化。另一方面，[8] 认为尖锐的最小值不一定会导致较差的泛化。论点是，由于深度网络中的参数化冗余，并且在一定的平坦度概念下，可以将平坦的最小值转换为等效的尖锐值。。尽管如此，人们普遍接受并通过经验验证了平坦的极小值倾向于提供更好的泛化性能，并且这种理解是几种最近的正则化技术(e.g. , [22,23] )的设计的基础。。[9]的并发工作进一步提供了 PAC-贝叶斯的理由，说明了为什么最小值的平坦度有助于泛化。[46] 设计的类似技术表明，平坦的最小值也可以提高鲁棒性
## 2.2. Regularization
正则化可以广泛地指任何有助于提高泛化的训练技术。尽管有众所周知的正则化技术，例如权重衰减[29] , Dropout[40] ,归一化技巧[2,21] 和数据增强[6,28,32] ,但已经开发了各种技术。标签平滑[42]将训练示例的一热标签与均匀分布混合。Shake-Shake 正则化 [11] 在多分支网络中将并行分支与随机仿射函数相结合。ShakeDrop 正则化[47]将 Shake-Shake 扩展到单分支架构。Cutout[7] 随机掩盖了输人图像的某些区域。MixUp [51] 通过沿其他样本的方向扰动训练样本来正则化深度网络。洪水[22] 迫使训练损失保持在零以上，以避免过度拟合。对抗训练[13]最初是为提高模型的对抗鲁棒性而设计的，当仔细选择其参数[34] 时，也显示出很大的正则化效果。但也观察到，将模型过度训练为对抗鲁棒性可能会损害泛化[43] 。最近的一些进展(e.g.,[10,23])利用模型参数的多个状态， 同时训练和集成它们以改善泛化。并发工作 [9] 也独立地发现了与我们在本文中介绍的类似的正则化方案
3. AMP: Adversarial Model Perturbation
3.1. From Empirical Risk to AMP Loss
考虑一个分类设置，其中我们旨在找到一个分类器$f:\mathcal{X}\to\mathcal{Y}$ ,该分类器将输人空间$\chi$映射到标签空间$\mathcal{Y}$。请注意，我们可以将$\mathcal{Y}$作为可能标签集上所有分布的集合，因此不仅$f(x)$属于$\mathcal{Y}$,每个真实标签也属于$\mathcal{Y}$ ,因为它可以写成退化分布或“独热”向量
我们将$f$作为由$\theta$参数化的神经网络，从其权重空间$\Theta$中取值。对于每个训练示例$(x,y)\in\mathcal{X}\times\mathcal{Y}$ ,我们将用$\ell(x,y;\boldsymbol{\theta})$表示模型相对于真实标签$y$的预测$f(\boldsymbol{x})$的损失。也就是说，函数$\ell$在
根据经验风险最小化 (ERM)原则[44],使用训练集$\mathcal{D}$对神经网络进行训练是通过最小化以下 empirical risk (我们也称为“ERM 损失”) $\mathcal{L}_\mathrm{ERM}$超过参数$\theta$ 来执行的：
(2)
$$\mathcal{L}_{\mathrm{ERM}}(\boldsymbol{\theta}):=\frac{1}{|\mathcal{D}|}\sum_{(\boldsymbol{x},\boldsymbol{y})\in\mathcal{D}}\ell(\boldsymbol{x},\boldsymbol{y};\boldsymbol{\theta})$$
2 | [
[
"0.5034",
"0.09344",
"0.8896",
"0.09344",
"0.8896",
"0.1067",
"0.5034",
"0.1067"
],
[
"0.165",
"0.09973",
"0.2214",
"0.09973",
"0.2214",
"0.1067",
"0.165",
"0.1067"
],
[
"0.1626",
"0.1061",
"0.2263",
"0.1067",
"0.2263",
"0.11743",
"0.1626",
"0.11615"
],
[
"0.3555",
"0.1067",
"0.4192",
"0.1061",
"0.4192",
"0.11615",
"0.3555",
"0.11743"
],
[
"0.357",
"0.101",
"0.4084",
"0.101",
"0.4084",
"0.1048",
"0.357",
"0.1048"
],
[
"0.505",
"0.108",
"0.8906",
"0.108",
"0.8906",
"0.1218",
"0.505",
"0.1218"
],
[
"0.10455",
"0.11743",
"0.11273",
"0.11426",
"0.1168",
"0.1206",
"0.10864",
"0.1237"
],
[
"0.5063",
"0.125",
"0.8896",
"0.125",
"0.8896",
"0.1351",
"0.5063",
"0.1351"
],
[
"0.5034",
"0.1383",
"0.8906",
"0.1383",
"0.8906",
"0.1515",
"0.5034",
"0.1515"
],
[
"0.5034",
"0.154",
"0.892",
"0.154",
"0.892",
"0.1674",
"0.5034",
"0.1674"
],
[
"0.505",
"0.1686",
"0.892",
"0.1686",
"0.892",
"0.1818",
"0.505",
"0.1818"
],
[
"0.134",
"0.1755",
"0.1805",
"0.1755",
"0.1805",
"0.1831",
"0.134",
"0.1831"
],
[
"0.326",
"0.1755",
"0.3489",
"0.1755",
"0.3489",
"0.1831",
"0.326",
"0.1831"
],
[
"0.1936",
"0.178",
"0.2009",
"0.178",
"0.2009",
"0.1818",
"0.1936",
"0.1818"
],
[
"0.1136",
"0.178",
"0.1356",
"0.178",
"0.1356",
"0.1818",
"0.1136",
"0.1818"
],
[
"0.3137",
"0.178",
"0.326",
"0.178",
"0.326",
"0.1812",
"0.3137",
"0.1812"
],
[
"0.3472",
"0.1787",
"0.3694",
"0.1787",
"0.3694",
"0.1818",
"0.3472",
"0.1818"
],
[
"0.433",
"0.178",
"0.4495",
"0.178",
"0.4495",
"0.1812",
"0.433",
"0.1812"
],
[
"0.505",
"0.1831",
"0.8896",
"0.1831",
"0.8896",
"0.1963",
"0.505",
"0.1963"
],
[
"0.5063",
"0.1995",
"0.8896",
"0.1995",
"0.8896",
"0.2102",
"0.5063",
"0.2102"
],
[
"0.0801",
"0.2058",
"0.468",
"0.2052",
"0.468",
"0.2197",
"0.0801",
"0.2203"
],
[
"0.5024",
"0.2122",
"0.8813",
"0.214",
"0.8813",
"0.2279",
"0.5024",
"0.226"
],
[
"0.0801",
"0.2216",
"0.468",
"0.2216",
"0.468",
"0.2349",
"0.0801",
"0.2349"
],
[
"0.5063",
"0.2291",
"0.8906",
"0.2291",
"0.8906",
"0.2424",
"0.5063",
"0.2424"
],
[
"0.0825",
"0.238",
"0.4673",
"0.238",
"0.4673",
"0.2488",
"0.0825",
"0.2488"
],
[
"0.5063",
"0.2437",
"0.8906",
"0.2437",
"0.8906",
"0.2568",
"0.5063",
"0.2568"
],
[
"0.0825",
"0.254",
"0.4656",
"0.254",
"0.4656",
"0.2644",
"0.0825",
"0.2644"
],
[
"0.505",
"0.2588",
"0.892",
"0.2588",
"0.892",
"0.2727",
"0.505",
"0.2727"
],
[
"0.0801",
"0.2676",
"0.468",
"0.2676",
"0.468",
"0.281",
"0.0801",
"0.281"
],
[
"0.5034",
"0.2734",
"0.6357",
"0.2734",
"0.6357",
"0.2874",
"0.5034",
"0.2874"
],
[
"0.0801",
"0.2822",
"0.4673",
"0.2822",
"0.4673",
"0.2954",
"0.0801",
"0.2954"
],
[
"0.0784",
"0.2966",
"0.1993",
"0.2974",
"0.1993",
"0.3113",
"0.0784",
"0.31"
],
[
"0.5034",
"0.3037",
"0.6562",
"0.3037",
"0.6562",
"0.3176",
"0.5034",
"0.3176"
],
[
"0.5254",
"0.33",
"0.8896",
"0.33",
"0.8896",
"0.3403",
"0.5254",
"0.3403"
],
[
"0.0809",
"0.339",
"0.468",
"0.339",
"0.468",
"0.3528",
"0.0809",
"0.3528"
],
[
"0.5034",
"0.3435",
"0.8906",
"0.3435",
"0.8906",
"0.3567",
"0.5034",
"0.3567"
],
[
"0.0825",
"0.3547",
"0.468",
"0.3547",
"0.468",
"0.3687",
"0.0825",
"0.3687"
],
[
"0.505",
"0.358",
"0.8906",
"0.358",
"0.8906",
"0.3718",
"0.505",
"0.3718"
],
[
"0.0809",
"0.3687",
"0.2443",
"0.3687",
"0.2443",
"0.3818",
"0.0809",
"0.3818"
],
[
"0.5063",
"0.375",
"0.8896",
"0.375",
"0.8896",
"0.385",
"0.5063",
"0.385"
],
[
"0.10046",
"0.385",
"0.468",
"0.385",
"0.468",
"0.3984",
"0.10046",
"0.3984"
],
[
"0.5034",
"0.387",
"0.8906",
"0.3865",
"0.8906",
"0.401",
"0.5034",
"0.4016"
],
[
"0.08417",
"0.4016",
"0.4673",
"0.4016",
"0.4673",
"0.4124",
"0.08417",
"0.4124"
],
[
"0.5034",
"0.4028",
"0.8906",
"0.4028",
"0.8906",
"0.4167",
"0.5034",
"0.4167"
],
[
"0.0809",
"0.4153",
"0.4673",
"0.4153",
"0.4673",
"0.4287",
"0.0809",
"0.4287"
],
[
"0.5034",
"0.4185",
"0.8906",
"0.4185",
"0.8906",
"0.432",
"0.5034",
"0.432"
],
[
"0.0809",
"0.4287",
"0.468",
"0.43",
"0.468",
"0.4443",
"0.0809",
"0.443"
],
[
"0.5034",
"0.4343",
"0.8906",
"0.4343",
"0.8906",
"0.4475",
"0.5034",
"0.4475"
],
[
"0.0825",
"0.4458",
"0.1136",
"0.4458",
"0.1136",
"0.4583",
"0.0825",
"0.4583"
],
[
"0.505",
"0.449",
"0.8906",
"0.449",
"0.8906",
"0.4622",
"0.505",
"0.4622"
],
[
"0.10046",
"0.461",
"0.468",
"0.461",
"0.468",
"0.4749",
"0.10046",
"0.4749"
],
[
"0.5034",
"0.464",
"0.8906",
"0.464",
"0.8906",
"0.4778",
"0.5034",
"0.4778"
],
[
"0.0801",
"0.476",
"0.4697",
"0.476",
"0.4697",
"0.4893",
"0.0801",
"0.4893"
],
[
"0.5034",
"0.4792",
"0.8906",
"0.4792",
"0.8906",
"0.4924",
"0.5034",
"0.4924"
],
[
"0.0809",
"0.4912",
"0.4673",
"0.4912",
"0.4673",
"0.505",
"0.0809",
"0.505"
],
[
"0.505",
"0.4944",
"0.888",
"0.4944",
"0.888",
"0.5083",
"0.505",
"0.5083"
],
[
"0.0825",
"0.5083",
"0.468",
"0.5083",
"0.468",
"0.518",
"0.0825",
"0.518"
],
[
"0.5024",
"0.509",
"0.8896",
"0.509",
"0.8896",
"0.523",
"0.5024",
"0.523"
],
[
"0.08417",
"0.5215",
"0.4697",
"0.5215",
"0.4697",
"0.535",
"0.08417",
"0.535"
],
[
"0.505",
"0.526",
"0.888",
"0.526",
"0.888",
"0.536",
"0.505",
"0.536"
],
[
"0.0809",
"0.536",
"0.423",
"0.536",
"0.423",
"0.55",
"0.0809",
"0.55"
],
[
"0.5186",
"0.537",
"0.888",
"0.5386",
"0.888",
"0.553",
"0.5186",
"0.552"
],
[
"0.1013",
"0.552",
"0.468",
"0.552",
"0.468",
"0.565",
"0.1013",
"0.565"
],
[
"0.505",
"0.5537",
"0.8906",
"0.5537",
"0.8906",
"0.5674",
"0.505",
"0.5674"
],
[
"0.0809",
"0.5674",
"0.4673",
"0.5674",
"0.4673",
"0.5806",
"0.0809",
"0.5806"
],
[
"0.5034",
"0.5693",
"0.863",
"0.5693",
"0.863",
"0.5835",
"0.5034",
"0.5835"
],
[
"0.0809",
"0.582",
"0.4673",
"0.582",
"0.4673",
"0.595",
"0.0809",
"0.595"
],
[
"0.0809",
"0.5977",
"0.468",
"0.5977",
"0.468",
"0.6113",
"0.0809",
"0.6113"
],
[
"0.501",
"0.6016",
"0.857",
"0.604",
"0.857",
"0.617",
"0.501",
"0.6157"
],
[
"0.0801",
"0.613",
"0.0989",
"0.613",
"0.0989",
"0.6255",
"0.0801",
"0.6255"
],
[
"0.10297",
"0.629",
"0.2793",
"0.629",
"0.2793",
"0.6396",
"0.10297",
"0.6396"
],
[
"0.5034",
"0.63",
"0.8096",
"0.63",
"0.8096",
"0.643",
"0.5034",
"0.643"
],
[
"0.1013",
"0.6426",
"0.4673",
"0.6426",
"0.4673",
"0.656",
"0.1013",
"0.656"
],
[
"0.08417",
"0.6577",
"0.468",
"0.6577",
"0.468",
"0.672",
"0.08417",
"0.672"
],
[
"0.523",
"0.653",
"0.8906",
"0.653",
"0.8906",
"0.6665",
"0.523",
"0.6665"
],
[
"0.0825",
"0.674",
"0.1152",
"0.674",
"0.1152",
"0.685",
"0.0825",
"0.685"
],
[
"0.5063",
"0.6704",
"0.888",
"0.6704",
"0.888",
"0.681",
"0.5063",
"0.681"
],
[
"0.0972",
"0.688",
"0.4697",
"0.688",
"0.4697",
"0.702",
"0.0972",
"0.702"
],
[
"0.505",
"0.686",
"0.888",
"0.686",
"0.888",
"0.6963",
"0.505",
"0.6963"
],
[
"0.0809",
"0.7026",
"0.2607",
"0.7026",
"0.2607",
"0.7163",
"0.0809",
"0.7163"
],
[
"0.505",
"0.6997",
"0.8906",
"0.6997",
"0.8906",
"0.7134",
"0.505",
"0.7134"
],
[
"0.0989",
"0.7183",
"0.468",
"0.7183",
"0.468",
"0.732",
"0.0989",
"0.732"
],
[
"0.505",
"0.7163",
"0.853",
"0.7163",
"0.853",
"0.7266",
"0.505",
"0.7266"
],
[
"0.0825",
"0.735",
"0.4673",
"0.735",
"0.4673",
"0.7456",
"0.0825",
"0.7456"
],
[
"0.523",
"0.73",
"0.8906",
"0.73",
"0.8906",
"0.7437",
"0.523",
"0.7437"
],
[
"0.0809",
"0.75",
"0.179",
"0.75",
"0.179",
"0.761",
"0.0809",
"0.761"
],
[
"0.505",
"0.747",
"0.8906",
"0.747",
"0.8906",
"0.7603",
"0.505",
"0.7603"
],
[
"0.505",
"0.761",
"0.8906",
"0.761",
"0.8906",
"0.776",
"0.505",
"0.776"
],
[
"0.0784",
"0.779",
"0.2279",
"0.779",
"0.2279",
"0.7925",
"0.0784",
"0.7925"
],
[
"0.5034",
"0.7764",
"0.7563",
"0.7764",
"0.7563",
"0.7905",
"0.5034",
"0.7905"
],
[
"0.523",
"0.7935",
"0.8906",
"0.7935",
"0.8906",
"0.8066",
"0.523",
"0.8066"
],
[
"0.0809",
"0.8047",
"0.3652",
"0.8047",
"0.3652",
"0.815",
"0.0809",
"0.815"
],
[
"0.505",
"0.808",
"0.8906",
"0.808",
"0.8906",
"0.8213",
"0.505",
"0.8213"
],
[
"0.1013",
"0.827",
"0.468",
"0.827",
"0.468",
"0.8403",
"0.1013",
"0.8403"
],
[
"0.5024",
"0.8237",
"0.892",
"0.8237",
"0.892",
"0.837",
"0.5024",
"0.837"
],
[
"0.0825",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0825",
"0.855"
],
[
"0.5034",
"0.8403",
"0.574",
"0.8403",
"0.574",
"0.8516",
"0.5034",
"0.8516"
],
[
"0.0809",
"0.857",
"0.4673",
"0.857",
"0.4673",
"0.8706",
"0.0809",
"0.8706"
],
[
"0.5767",
"0.867",
"0.7075",
"0.879",
"0.705",
"0.8994",
"0.574",
"0.888"
],
[
"0.6763",
"0.8687",
"0.7017",
"0.8706",
"0.701",
"0.8774",
"0.675",
"0.8755"
],
[
"0.7036",
"0.872",
"0.8164",
"0.872",
"0.8164",
"0.89",
"0.7036",
"0.89"
],
[
"0.8677",
"0.875",
"0.8906",
"0.875",
"0.8906",
"0.8896",
"0.8677",
"0.8896"
],
[
"0.6978",
"0.8833",
"0.7114",
"0.8833",
"0.7114",
"0.887",
"0.6978",
"0.887"
],
[
"0.0801",
"0.8706",
"0.4673",
"0.8726",
"0.4673",
"0.886",
"0.0801",
"0.884"
],
[
"0.0825",
"0.887",
"0.468",
"0.887",
"0.468",
"0.901",
"0.0825",
"0.901"
],
[
"0.6978",
"0.8916",
"0.7485",
"0.8916",
"0.7485",
"0.903",
"0.6978",
"0.903"
],
[
"0.478",
"0.924",
"0.4944",
"0.924",
"0.4944",
"0.938",
"0.478",
"0.938"
]
] | [
"过可视化损失景观展示了平坦最小值的更好泛化。[26]",
"Lam(6gou + a",
"CenM(6Asp +ad)",
"Le(eaar- +a8)",
"+",
"中的经验结果表明，大批量SGD会找到尖锐的最小",
"o",
"值，而小批量SGD会导致更平坦的最小值并提供更好",
"的泛化。另一方面，8]认为尖锐的最小值不一定会导",
"致较差的泛化。论点是，由于深度网络中的参数化穴",
"余，并且在一定的平坦度概念下，可以将平坦的最小",
"210-03",
"0",
"",
"15",
"",
"5",
"",
"值转换为等效的尖锐值。。尽管如此，人们普遍接受并",
"通过经验验证了平坦的极小值倾向于提供更好的泛化",
"Figure 2: Landscapes of the empirical risks obtained",
"性能，并且这种理解是几种最近的正则化技术（e.g.",
"from the PreActResNet18 [18] models trained with",
"[22，23]）的设计的基础。。[9]的并发工作进一步提供",
"ERM(red)andAMP(blue)onCIFAR-10:Left:on",
"了PAC-贝叶斯的理由，说明了为什么最小值的平坦度",
"training set; right: on test set. These curves are com",
"有助于泛化。[46]设计的类似技术表明，平坦的最小值",
"puted using the technique presented in [31] , where ",
"也可以提高鲁棒性",
"indicates a random direction and α is a displacement",
"in that direction",
"2.2. Regularization",
"正则化可以广泛地指任何有助于提高泛化的训练",
"（右）中的蓝色曲线所示。由于AMP损耗中的右最小",
"技术。尽管有众所周知的正则化技术，例如权重衰减",
"值低于左最小值，因此将AMP损耗最小化会给出右",
"[29]，Dropout[40]，归一化技巧[2，21]和数据增强",
"最小值作为其解决方案",
"6.28.32，但已经开发了各种技术。标签平滑42将",
"在本文中，我们正式分析了AMP损耗最小化问题",
"训练示例的一热标签与均匀分布混合。Shake-Shake正",
"及其对平坦局部最小值的偏好。具体来说，我们表明该",
"则化[11]在多分支网络中将并行分支与随机仿射函数",
"最小化问题隐式地使用局部最小值的“最窄宽度”作为",
"相结合。ShakeDrop正则化[47]将Shake-Shake扩展",
"平坦度的概念，并试图惩罚在这种意义上不平坦的最",
"到单分支架构。Cutout[7]随机掩盖了输入图像的某些",
"小值",
"区域。MixUp[51]通过沿其他样本的方向扰动训练样",
"我们推导了一种用于解决此最小化问题的小批量",
"本来正则化深度网络。洪水[22]迫使训练损失保持在",
"SGD算法，该算法产生了建议的AMP正则化方案。",
"零以上，以避免过度拟合。对抗训练[13]最初是为提",
"有趣的是，我们证明了该算法也可以看作是在梯度范",
"高模型的对抗鲁棒性而设计的，当仔细选择其参数[34",
"数上具有附加惩罚项的常规经验风险最小化。这提供",
"时，也显示出很大的正则化效果。但也观察到，将模型",
"了AMP方案的另一种理由。图2包含一个实验结果，",
"过度训练为对抗鲁棒性可能会损害泛化43。最近的",
"表明AMP确实选择了比ERM更平坦的最小值",
"些进展（e.g：，[10,23]）利用模型参数的多个状态",
"我们在几个基准图像分类数据集（SVHN，CIFAR",
"同时训练和集成它们以改善泛化。并发工作[9]也独立",
"10，CIFAR-100）上进行实验，以验证所提出的AMP",
"地发现了与我们在本文中介绍的类似的正则化方案",
"方案的有效性。与其他流行的正则化方案相比，AMP",
"表现出显著的正则化性能，将自已确立为一种新的技",
"3.AMP:Adversarial ModelPerturbation",
"术",
"我们的贡献可以总结如下",
"3.1. From Empirical Risk to AMP Loss",
"1）出于对平坦最小值有助于泛化的理解，我们提出",
"了对抗模型扰动（AMP）作为一种新颖有效的正则化",
"考虑一个分类设置，其中我们旨在找到一个分类器",
"方案",
"f：→，该分类器将输入空间映射到标签空间",
"2）我们从理论上证明AMP能够找到更平坦的局部",
"。请注意，我们可以将作为可能标签集上所有分",
"最小值，从而改善了泛化",
"布的集合，因此不仅f(α）属于，每个真实标签也属",
"3）在基准数据集上进行的广泛实验表明，AMP在",
"于，因为它可以写成退化分布或“独热”向量",
"各种现代神经网络体系结构上的比较正则化方案中实",
"我们将f作为由θ参数化的神经网络，从其权重空",
"现了最佳性能",
"间e中取值。对于每个训练示例(a,y)E×，我",
"们将用l(a,y;の)表示模型相对于真实标签y的预测",
"2.Related Work",
"f(α）的损失。也就是说，函数l在",
"根据经验风险最小化（ERM）原则[44]，使用训练",
"2.1. Flat Minima and Generalization",
"集D对神经网络进行训练是通过最小化以下empirical",
"已经有大量的工作研究了局部极小值的平坦度与深",
"risk（我们也称为“ERM损失”）LERM超过参数θ",
"度神经网络[4，19,26,31]的泛化之间的关系。[19]建",
"来执行的：",
"议平坦极小值对应于低复杂度网络，在最小描述长度",
"LERM(0) := D (",
"1",
"E l(a,y;0)",
"(2)",
"",
"[38]的原理下，它倾向于很好地推广。[4]通过吉布斯自",
"由能的视角提出了另一种支持这一论点的解释。[31]通",
"(r,y)ED",
"2"
] | {
"arxiv_id": "2010.04925",
"language": "zh",
"page_idx": 1
} |
||
2010.04925-zh-2 | 2010.04925-zh-2.jpg | 然而，众所周知，ERM 训练容易过度拟合[44],并
且学习的模型通常无法很好地概括看不见的示例
这项工作的动机是需要有效的正则化方案来提高模型的泛化能力。具体来说，我们的目标是开发一种技术，迫使训练过程在最小化经验风险时找到更平坦的最小值或低范数解。。我们开发的技术被称为“对抗性模型扰动”或 AMP,我们现在详细说明
对于任何正的$\epsilon$和任何$\mu\in\Theta$ ,让 B$(\boldsymbol\mu;\epsilon)$表示$\Theta$
中的范数球，半径$\epsilon$以$\mu$为中心。即
$$\mathrm{B}(\boldsymbol{\mu};\epsilon):=\{\boldsymbol{\theta}\in\Theta:\|\boldsymbol{\theta}-\boldsymbol{\mu}\|\leq\epsilon\}$$
(3)
我们注意到，用于定义范数球的范数被选择为$\mathcal{L}_2$范
数。但是，我们希望规范可以扩展到此选择之外。
我们现在定义“AMP 损失”$\mathcal{L}_\mathrm{AMP}$如下。
(4)
$$\mathcal{L}_{\mathrm{AMP}}(\boldsymbol{\theta}):=\max_{\Delta\in\mathrm{B}(\mathbf{0};\epsilon)}\frac{1}{|\mathcal{D}|}\sum_{(\boldsymbol{x},\boldsymbol{y})\in\mathcal{D}}\ell(\boldsymbol{x},\boldsymbol{y};\boldsymbol{\theta}+\Delta)$$
,其中$\epsilon$是用作超参数的小正值
Figure 1 草拟了一个示例，演示了在$\theta$上最小化AMP 损耗$\mathcal{L}_\mathrm{AMP}$提供了找到 ERM 损耗$\mathcal{L}_\mathrm{ERM}$的更平坦最小值的机会。4 节给出了 AMP 寻找更平坦最小值的能力的理论依据
# 3.2. Training Algorithm
对计算复杂性和训练速度的实际考虑激发了一种小型批处理训练方法，以最大程度地减少 AMP 损失$\mathcal{L}_{\mathrm{AMP}}$。具体来说，我们可以使用从随机小批量$\mathcal{B}$获得的相应损耗来近似(4)中的 AMP 损耗，即
$$\begin{aligned}\mathcal{L}_{\mathrm{AMP}}(\boldsymbol{\theta})&\approx\max_{\Delta_{\mathcal{B}}\in\mathcal{B}(\mathbf{0};\epsilon)}\frac{1}{|\mathcal{B}|}\sum_{(\boldsymbol{x},\boldsymbol{y})\in\mathcal{B}}\ell(\boldsymbol{x},\boldsymbol{y};\boldsymbol{\theta}+\Delta_{\mathcal{B}})\\&:=\mathcal{J}_{\mathrm{AMP},\mathcal{B}}(\boldsymbol{\theta})&\text{(}\end{aligned}$$
(5)
然后，AMP 损耗最小化问题可以近似为最小化预期
的批次级 AMP 损耗$\mathcal{J}_\mathrm{AMP,B}$ ,正式为查找
(6)
$$\theta_{\mathrm{AMP}}^*:=\arg\min_{\theta}\mathbb{E}_{\mathcal{B}}\mathcal{J}_{\mathrm{AMP},\mathcal{B}}(\boldsymbol{\theta})$$
在此公式下，每个批次$\mathcal{B}$与模型参数$\theta$上的扰动向量$\Delta_{\mathcal{B}}$相关联。训练涉及嵌套在内部和外部最小化的内部最大化：在内部最大化中，使用固定数量的步骤例如$N$ ,以增加 ERM 损失的方向更新$\Delta_B$ ,从而获得$\mathcal{J}_{\mathrm{AMP,B}}$ ;外部最小化在随机批次上循环，并使用迷你批处理 SGD 最小化$\mathbb{E}_\mathcal{B}\mathcal{J}_{\mathrm{AMP,B}}$ 。算法 1 中描述了精确的算法。请注意，在训练结束时，学习的$\theta_\mathrm{AMP}^*$用作预测的模型参数；也就是说，在测试中不应用扰动
4. Theoretical Justifications of AMP
4.1. AMP Finds Flatter Local Minima
为了便于获得分析结果，我们将假设$\mathcal{L}_\mathrm{ERM}$中每个
局部最小值的损失面可以近似为倒高斯面
<table>
<tbody>
<tr>
<td>$\underline{\mathrm{lg}}.$ 1</td>
<td>rithm 1 Adversarial Model Perturbation Training</td>
</tr>
<tr>
<td>$\{$eq</td>
<td>$\{atc|$ size $m$</td>
</tr>
<tr>
<td> </td>
<td>$1p$ $(1)_{11}$te</td>
</tr>
<tr>
<td> </td>
<td>tera</td>
</tr>
<tr>
<td> </td>
<td>tion number $N,\mathrm{L}_{2}$norm ball radius $\epsilon$</td>
</tr>
<tr>
<td>1: $\bullet$ </td>
<td>while $\theta_{l,}$ mot $\operatorname{rged}$do</td>
</tr>
<tr>
<td>2:</td>
<td>||ndate iteration. $k.$ $\leftarrow k+1$</td>
</tr>
<tr>
<td>3:</td>
<td>Sample $\mathcal{B}=$ 1.7 $m.$ from training set $D$ $\because$</td>
</tr>
<tr>
<td>4: 1</td>
<td>Initialize perturbation: $\Delta_{\mathcal{B}}\leftarrow0$</td>
</tr>
<tr>
<td>5:</td>
<td>for $t_{0}$ $N$ $\operatorname{do}$ $TI_{1}$</td>
</tr>
<tr>
<td>6:</td>
<td>Compute gradient: $\daleth m$</td>
</tr>
<tr>
<td>7:</td>
<td>$J$AMP$,\mathcal{B}$ $3\leftarrow\sum_{i=1}$ $^{\prime}\setminus\boldsymbol{u}$ $i,\boldsymbol{y}_{i},\boldsymbol{v}_{k}+\Delta\boldsymbol{h}$ Update perturbation: $\Delta_{\mathcal{B}}\leftarrow\Delta_{\mathcal{B}}+\zeta\nabla\mathcal{J}_{\mathrm{AMP},\mathcal{E}}$</td>
</tr>
<tr>
<td>8:</td>
<td>then if $\|\Delta_{\mathcal{R}}\|$ $\epsilon$</td>
</tr>
<tr>
<td>9:</td>
<td>Normalize perturbation: $\Delta_{B}$ $\epsilon\Delta_{\mathcal{B}}/\|\Delta_{\mathcal{B}}\|_{\mathcal{2}}$</td>
</tr>
<tr>
<td>0:</td>
<td>end if</td>
</tr>
<tr>
<td>1: $\bullet$ </td>
<td>end for</td>
</tr>
<tr>
<td>.2:</td>
<td>Compute gradient: $17n$</td>
</tr>
<tr>
<td>.3:</td>
<td>Update parameter: $H_{1}$ $\leftarrow\boldsymbol{\theta}_{k}-\eta\nabla\mathcal{J}_{\mathrm{AMP},\mathcal{B}}$</td>
</tr>
<tr>
<td>.4:</td>
<td>$l$ while omd</td>
</tr>
</tbody>
</table>
更准确地说，假设$\Theta=\mathbb{R}^K$ 。对于任何标量$C$ ,任何正标量$A$ ,任何$\mu\in\Theta$和任何$K\times K$正定矩阵$\kappa$ $,令$
$$\gamma(\boldsymbol{\theta};\boldsymbol{\mu},\boldsymbol{\kappa},A,C):=C-A\exp\left(-(\boldsymbol{\theta}-\boldsymbol{\mu})^T\boldsymbol{\kappa}^{-1}(\boldsymbol{\theta}-\boldsymbol{\mu})/2\right)$$
(7)
请注意，此函数只是空间$\Theta$上的逆高斯曲面
XNETHX\_1\_0 使用上述表示法，我们假设经验风险的每个最小值都可以是由这样的倒置高斯曲面近似的locally ,即，如果$\boldsymbol\mu\in\Theta$给出$\mathcal{L}_\mathrm{ERM}$的局部最小值， 则存在一些$\epsilon>0$,正定矩阵 XMATHMAX\_3 和。我们注意到，一般来说，这样的假设只是一个近似值。但是对于小型$\epsilon$,近似值可以说是准确的
我们将使用$\gamma^*(\boldsymbol{\mu},\boldsymbol{\kappa},A,C)$表示函数$\gamma(\boldsymbol{\theta};\boldsymbol{\mu},\boldsymbol{\kappa},A,C)$
的最小值(通过在$\theta$上最小化而获得)。很明显，
$$\gamma^*(\mu,\kappa,A,C)=C-A$$
(8)
令$\gamma_\mathrm{AMP}$表示从$\gamma$派生的 AMP 损耗，即
$$\gamma_{\mathrm{AMP}}(\boldsymbol{\theta};\boldsymbol{\mu},\boldsymbol{\kappa},A,C):=\max_{\Delta\in\mathrm{B}(\mathbf{0};\epsilon)}\gamma(\boldsymbol{\theta}+\Delta;\boldsymbol{\mu},\boldsymbol{\kappa},A,C)$$
$\gamma _{\mathrm{AMP}}^{* }( \boldsymbol{\mu }, \boldsymbol{\kappa }, A, C)$ 表 示
我们将使用
$\gamma_\mathrm{AMP}(\boldsymbol{\theta};\boldsymbol{\mu},\boldsymbol{\kappa},A,C)$的最小值(在$\boldsymbol\theta$上最小化)
Theorem 1. For any given $(\boldsymbol{\mu},\boldsymbol{\kappa},A,C)$ , the function $\gamma_{\mathrm{AMP}}(\boldsymbol{\theta};\boldsymbol{\mu},\boldsymbol{\kappa},A,C)$ is minimized when $\boldsymbol{\theta}=\boldsymbol{\mu}$ and the minimum value is
$$\gamma_{\mathrm{AMP}}^*(\boldsymbol{\mu},\boldsymbol{\kappa},A,C)=C-A\exp\left(-\frac{\epsilon^2}{2\sigma^2}\right)$$
(10)
3 | [
[
"0.10046",
"0.09216",
"0.468",
"0.09216",
"0.468",
"0.1061",
"0.10046",
"0.1061"
],
[
"0.5024",
"0.09155",
"0.8896",
"0.09216",
"0.8896",
"0.1067",
"0.5024",
"0.1061"
],
[
"0.0825",
"0.108",
"0.4248",
"0.108",
"0.4248",
"0.1218",
"0.0825",
"0.1218"
],
[
"0.5024",
"0.11237",
"0.8896",
"0.11237",
"0.8896",
"0.127",
"0.5024",
"0.127"
],
[
"0.10046",
"0.1225",
"0.468",
"0.1225",
"0.468",
"0.1364",
"0.10046",
"0.1364"
],
[
"0.5312",
"0.1276",
"0.8896",
"0.1276",
"0.8896",
"0.1414",
"0.5312",
"0.1414"
],
[
"0.0809",
"0.1383",
"0.4673",
"0.1383",
"0.4673",
"0.1515",
"0.0809",
"0.1515"
],
[
"0.5312",
"0.1427",
"0.8906",
"0.1427",
"0.8906",
"0.1559",
"0.5312",
"0.1559"
],
[
"0.0825",
"0.1528",
"0.4673",
"0.1528",
"0.4673",
"0.1666",
"0.0825",
"0.1666"
],
[
"0.5293",
"0.1572",
"0.8096",
"0.1578",
"0.8096",
"0.1718",
"0.5293",
"0.1704"
],
[
"0.0809",
"0.1674",
"0.468",
"0.1674",
"0.468",
"0.1818",
"0.0809",
"0.1818"
],
[
"0.5107",
"0.1724",
"0.728",
"0.1724",
"0.728",
"0.1863",
"0.5107",
"0.1863"
],
[
"0.0809",
"0.1831",
"0.3694",
"0.1831",
"0.3694",
"0.1963",
"0.0809",
"0.1963"
],
[
"0.5156",
"0.1875",
"0.5293",
"0.1925",
"0.523",
"0.202",
"0.509",
"0.1963"
],
[
"0.5474",
"0.1875",
"0.751",
"0.1875",
"0.751",
"0.2008",
"0.5474",
"0.2008"
],
[
"0.1013",
"0.1976",
"0.468",
"0.1976",
"0.468",
"0.2122",
"0.1013",
"0.2122"
],
[
"0.511",
"0.2052",
"0.5254",
"0.2052",
"0.5254",
"0.2153",
"0.511",
"0.2153"
],
[
"0.5474",
"0.2026",
"0.8735",
"0.2026",
"0.8735",
"0.2172",
"0.5474",
"0.2172"
],
[
"0.0825",
"0.2134",
"0.3513",
"0.2134",
"0.3513",
"0.2279",
"0.0825",
"0.2279"
],
[
"0.5107",
"0.2184",
"0.528",
"0.2184",
"0.528",
"0.2317",
"0.5107",
"0.2317"
],
[
"0.546",
"0.2166",
"0.778",
"0.2172",
"0.778",
"0.2317",
"0.546",
"0.2311"
],
[
"0.5146",
"0.2329",
"0.528",
"0.238",
"0.5215",
"0.2488",
"0.5073",
"0.2437"
],
[
"0.5474",
"0.2329",
"0.693",
"0.2329",
"0.693",
"0.2468",
"0.5474",
"0.2468"
],
[
"0.1569",
"0.2393",
"0.3906",
"0.2393",
"0.3906",
"0.254",
"0.1569",
"0.254"
],
[
"0.4453",
"0.2405",
"0.468",
"0.2405",
"0.468",
"0.254",
"0.4453",
"0.254"
],
[
"0.509",
"0.2488",
"0.527",
"0.2488",
"0.527",
"0.2632",
"0.509",
"0.2632"
],
[
"0.564",
"0.2488",
"0.699",
"0.2488",
"0.699",
"0.262",
"0.564",
"0.262"
],
[
"0.10046",
"0.2651",
"0.4656",
"0.2651",
"0.4656",
"0.279",
"0.10046",
"0.279"
],
[
"0.5786",
"0.262",
"0.886",
"0.262",
"0.886",
"0.2798",
"0.5786",
"0.2798"
],
[
"0.509",
"0.2778",
"0.527",
"0.2778",
"0.527",
"0.2922",
"0.509",
"0.2922"
],
[
"0.562",
"0.2766",
"0.8906",
"0.2798",
"0.8906",
"0.2942",
"0.562",
"0.2917"
],
[
"0.10046",
"0.2966",
"0.3962",
"0.2966",
"0.3962",
"0.3113",
"0.10046",
"0.3113"
],
[
"0.0809",
"0.281",
"0.4307",
"0.281",
"0.4307",
"0.2942",
"0.0809",
"0.2942"
],
[
"0.509",
"0.3093",
"0.527",
"0.3093",
"0.527",
"0.324",
"0.509",
"0.324"
],
[
"0.562",
"0.308",
"0.705",
"0.308",
"0.705",
"0.3225",
"0.562",
"0.3225"
],
[
"0.09644",
"0.3208",
"0.2852",
"0.334",
"0.282",
"0.3591",
"0.09314",
"0.3452"
],
[
"0.2566",
"0.3213",
"0.2769",
"0.3213",
"0.2769",
"0.332",
"0.2566",
"0.332"
],
[
"0.2834",
"0.3208",
"0.4722",
"0.3245",
"0.4714",
"0.3499",
"0.282",
"0.3452"
],
[
"0.5786",
"0.3213",
"0.6587",
"0.324",
"0.6587",
"0.3384",
"0.5786",
"0.336"
],
[
"0.8145",
"0.3225",
"0.847",
"0.3225",
"0.847",
"0.3384",
"0.8145",
"0.3384"
],
[
"0.66",
"0.324",
"0.773",
"0.324",
"0.773",
"0.3372",
"0.66",
"0.3372"
],
[
"0.509",
"0.324",
"0.527",
"0.324",
"0.527",
"0.3384",
"0.509",
"0.3384"
],
[
"0.8735",
"0.327",
"0.886",
"0.327",
"0.886",
"0.3352",
"0.8735",
"0.3352"
],
[
"0.2051",
"0.3352",
"0.2303",
"0.3352",
"0.2303",
"0.339",
"0.2051",
"0.339"
],
[
"0.58",
"0.339",
"0.6665",
"0.339",
"0.6665",
"0.3542",
"0.58",
"0.3542"
],
[
"0.273",
"0.3452",
"0.3284",
"0.3452",
"0.3284",
"0.3567",
"0.273",
"0.3567"
],
[
"0.505",
"0.3542",
"0.527",
"0.3542",
"0.527",
"0.3687",
"0.505",
"0.3687"
],
[
"0.562",
"0.351",
"0.609",
"0.3528",
"0.607",
"0.3687",
"0.5605",
"0.3662"
],
[
"0.5034",
"0.3694",
"0.527",
"0.3694",
"0.527",
"0.3833",
"0.5034",
"0.3833"
],
[
"0.546",
"0.3687",
"0.601",
"0.3687",
"0.601",
"0.3833",
"0.546",
"0.3833"
],
[
"0.0948",
"0.3706",
"0.3096",
"0.3706",
"0.3096",
"0.3806",
"0.0948",
"0.3806"
],
[
"0.08417",
"0.3738",
"0.0948",
"0.3706",
"0.0972",
"0.3774",
"0.08826",
"0.3806"
],
[
"0.10046",
"0.3838",
"0.468",
"0.3838",
"0.468",
"0.3977",
"0.10046",
"0.3977"
],
[
"0.505",
"0.385",
"0.527",
"0.385",
"0.527",
"0.3977",
"0.505",
"0.3977"
],
[
"0.5474",
"0.3838",
"0.683",
"0.3838",
"0.683",
"0.3984",
"0.5474",
"0.3984"
],
[
"0.0801",
"0.3977",
"0.468",
"0.3977",
"0.468",
"0.4153",
"0.0801",
"0.4153"
],
[
"0.5786",
"0.3977",
"0.886",
"0.3977",
"0.886",
"0.4153",
"0.5786",
"0.4153"
],
[
"0.0825",
"0.414",
"0.468",
"0.414",
"0.468",
"0.428",
"0.0825",
"0.428"
],
[
"0.5063",
"0.4153",
"0.527",
"0.4153",
"0.527",
"0.428",
"0.5063",
"0.428"
],
[
"0.546",
"0.4136",
"0.865",
"0.4153",
"0.865",
"0.43",
"0.546",
"0.428"
],
[
"0.0809",
"0.43",
"0.246",
"0.43",
"0.246",
"0.443",
"0.0809",
"0.443"
],
[
"0.505",
"0.4287",
"0.601",
"0.4287",
"0.601",
"0.443",
"0.505",
"0.443"
],
[
"0.0784",
"0.452",
"0.2695",
"0.4526",
"0.2695",
"0.4673",
"0.0784",
"0.4666"
],
[
"0.5254",
"0.4697",
"0.8906",
"0.4697",
"0.8906",
"0.4841",
"0.5254",
"0.4841"
],
[
"0.10046",
"0.4766",
"0.4673",
"0.4766",
"0.4673",
"0.4905",
"0.10046",
"0.4905"
],
[
"0.08417",
"0.4912",
"0.468",
"0.4912",
"0.468",
"0.506",
"0.08417",
"0.506"
],
[
"0.5034",
"0.486",
"0.8906",
"0.486",
"0.8906",
"0.5",
"0.5034",
"0.5"
],
[
"0.5034",
"0.505",
"0.523",
"0.505",
"0.523",
"0.513",
"0.5034",
"0.513"
],
[
"0.5186",
"0.5005",
"0.5376",
"0.5005",
"0.5376",
"0.5146",
"0.5186",
"0.5146"
],
[
"0.0809",
"0.5083",
"0.468",
"0.506",
"0.468",
"0.5195",
"0.0809",
"0.5215"
],
[
"0.0809",
"0.523",
"0.4224",
"0.523",
"0.4224",
"0.536",
"0.0809",
"0.536"
],
[
"0.505",
"0.5234",
"0.886",
"0.523",
"0.886",
"0.5405",
"0.505",
"0.5415"
],
[
"0.2632",
"0.5474",
"0.2778",
"0.5474",
"0.2778",
"0.558",
"0.2632",
"0.558"
],
[
"0.869",
"0.5415",
"0.8896",
"0.5415",
"0.8896",
"0.5537",
"0.869",
"0.5537"
],
[
"0.10455",
"0.553",
"0.1929",
"0.5547",
"0.1929",
"0.5693",
"0.10455",
"0.5674"
],
[
"0.2942",
"0.5503",
"0.3228",
"0.5503",
"0.3228",
"0.5684",
"0.2942",
"0.5684"
],
[
"0.3325",
"0.5537",
"0.4436",
"0.5537",
"0.4436",
"0.5684",
"0.3325",
"0.5684"
],
[
"0.5215",
"0.5547",
"0.8423",
"0.5537",
"0.8423",
"0.5684",
"0.5215",
"0.5693"
],
[
"0.1896",
"0.5625",
"0.281",
"0.5625",
"0.281",
"0.577",
"0.1896",
"0.577"
],
[
"0.2018",
"0.556",
"0.2402",
"0.556",
"0.2402",
"0.5674",
"0.2018",
"0.5674"
],
[
"0.282",
"0.5713",
"0.3342",
"0.5713",
"0.3342",
"0.5835",
"0.282",
"0.5835"
],
[
"0.5034",
"0.571",
"0.8906",
"0.571",
"0.8906",
"0.584",
"0.5034",
"0.584"
],
[
"0.1732",
"0.589",
"0.273",
"0.589",
"0.273",
"0.604",
"0.1732",
"0.604"
],
[
"0.4468",
"0.589",
"0.468",
"0.589",
"0.468",
"0.603",
"0.4468",
"0.603"
],
[
"0.5063",
"0.587",
"0.888",
"0.587",
"0.888",
"0.5977",
"0.5063",
"0.5977"
],
[
"0.5034",
"0.5996",
"0.888",
"0.601",
"0.888",
"0.6157",
"0.5034",
"0.6143"
],
[
"0.10046",
"0.613",
"0.4673",
"0.613",
"0.4673",
"0.6274",
"0.10046",
"0.6274"
],
[
"0.505",
"0.6157",
"0.8906",
"0.6157",
"0.8906",
"0.629",
"0.505",
"0.629"
],
[
"0.0784",
"0.6274",
"0.3938",
"0.629",
"0.3938",
"0.6465",
"0.0784",
"0.646"
],
[
"0.505",
"0.631",
"0.892",
"0.631",
"0.892",
"0.6445",
"0.505",
"0.6445"
],
[
"0.5034",
"0.6445",
"0.7817",
"0.6445",
"0.7817",
"0.659",
"0.5034",
"0.659"
],
[
"0.165",
"0.655",
"0.3816",
"0.656",
"0.3816",
"0.674",
"0.165",
"0.6724"
],
[
"0.4453",
"0.657",
"0.4697",
"0.657",
"0.4697",
"0.6704",
"0.4453",
"0.6704"
],
[
"0.524",
"0.661",
"0.892",
"0.661",
"0.892",
"0.6753",
"0.524",
"0.6753"
],
[
"0.505",
"0.677",
"0.8506",
"0.677",
"0.8506",
"0.69",
"0.505",
"0.69"
],
[
"0.1013",
"0.69",
"0.468",
"0.69",
"0.468",
"0.704",
"0.1013",
"0.704"
],
[
"0.613",
"0.7007",
"0.7817",
"0.6997",
"0.7817",
"0.714",
"0.613",
"0.7153"
],
[
"0.8667",
"0.6997",
"0.8936",
"0.6997",
"0.8936",
"0.7163",
"0.8667",
"0.7163"
],
[
"0.0825",
"0.7197",
"0.4673",
"0.7197",
"0.4673",
"0.733",
"0.0825",
"0.733"
],
[
"0.0825",
"0.704",
"0.4673",
"0.704",
"0.4673",
"0.7173",
"0.0825",
"0.7173"
],
[
"0.524",
"0.726",
"0.8228",
"0.726",
"0.8228",
"0.7407",
"0.524",
"0.7407"
],
[
"0.0825",
"0.735",
"0.468",
"0.735",
"0.468",
"0.7485",
"0.0825",
"0.7485"
],
[
"0.0809",
"0.7505",
"0.468",
"0.7505",
"0.468",
"0.765",
"0.0809",
"0.765"
],
[
"0.511",
"0.7505",
"0.8906",
"0.7505",
"0.8906",
"0.7646",
"0.511",
"0.7646"
],
[
"0.0801",
"0.763",
"0.4673",
"0.7646",
"0.4673",
"0.782",
"0.0801",
"0.781"
],
[
"0.6665",
"0.761",
"0.7256",
"0.761",
"0.7256",
"0.7715",
"0.6665",
"0.7715"
],
[
"0.0801",
"0.779",
"0.468",
"0.78",
"0.468",
"0.7944",
"0.0801",
"0.7935"
],
[
"0.5215",
"0.782",
"0.665",
"0.782",
"0.665",
"0.7954",
"0.5215",
"0.7954"
],
[
"0.6855",
"0.782",
"0.819",
"0.781",
"0.819",
"0.8",
"0.6855",
"0.8003"
],
[
"0.841",
"0.781",
"0.892",
"0.7827",
"0.8906",
"0.7974",
"0.839",
"0.7954"
],
[
"0.0809",
"0.7954",
"0.4558",
"0.7954",
"0.4558",
"0.8096",
"0.0809",
"0.8096"
],
[
"0.5024",
"0.799",
"0.8384",
"0.7974",
"0.8384",
"0.8125",
"0.5024",
"0.813"
],
[
"0.0784",
"0.8237",
"0.3938",
"0.8228",
"0.3938",
"0.837",
"0.0784",
"0.8384"
],
[
"0.5024",
"0.8193",
"0.892",
"0.821",
"0.892",
"0.8354",
"0.5024",
"0.834"
],
[
"0.5034",
"0.836",
"0.892",
"0.836",
"0.892",
"0.851",
"0.5034",
"0.851"
],
[
"0.0809",
"0.8486",
"0.3774",
"0.8486",
"0.3774",
"0.8623",
"0.0809",
"0.8623"
],
[
"0.5024",
"0.851",
"0.6323",
"0.8496",
"0.6323",
"0.8643",
"0.5024",
"0.8657"
],
[
"0.1013",
"0.872",
"0.4673",
"0.872",
"0.4673",
"0.8853",
"0.1013",
"0.8853"
],
[
"0.539",
"0.879",
"0.7646",
"0.8774",
"0.7646",
"0.8965",
"0.539",
"0.898"
],
[
"0.859",
"0.88",
"0.8906",
"0.88",
"0.8906",
"0.896",
"0.859",
"0.896"
],
[
"0.0809",
"0.887",
"0.3757",
"0.887",
"0.3757",
"0.901",
"0.0809",
"0.901"
],
[
"0.774",
"0.89",
"0.812",
"0.89",
"0.812",
"0.901",
"0.774",
"0.901"
],
[
"0.478",
"0.924",
"0.4944",
"0.924",
"0.4944",
"0.938",
"0.478",
"0.938"
]
] | [
"然而，众所周知，ERM训练容易过度拟合44]，并",
"Algorithm 1 Adversarial Model Perturbation Training",
"且学习的模型通常无法很好地概括看不见的示例",
"Require: Training set D = {(α,y)) , Batch size m,",
"这项工作的动机是需要有效的正则化方案来提高模",
"Loss function l,Initial model parameter o, Outer",
"型的泛化能力。具体来说，我们的目标是开发一种技",
"learning rate n,Inner learning rate S, Inner itera",
"术，迫使训练过程在最小化经验风险时找到更平坦的",
"tion number N, L2 norm ball radius e",
"最小值或低范数解。。我们开发的技术被称为“对抗性",
"1:while O not converged do",
"模型扰动”或AMP，我们现在详细说明",
"2:",
"Update iteration:kk+1",
"对于任何正的和任何μEの，让B(u;e)表示",
"3:",
"Sample B = {(ci,yi))i from training set D",
"中的范数球，半径以μ为中心。即",
"4:",
"Initializeperturbation:+0",
"5:",
"for n1 to N do",
"B(μ;e) := [0E0: 110 -μll ≤e)",
"(3)",
"6:",
"Compute gradient:",
"我们注意到，用于定义范数球的范数被选择为L2范",
"VJAMP,B -Em Vel(ai,Y;0k+AB)/m",
"7:",
"Updateperturbation:A+VJAMP,B",
"我们现在定义“AMP损失”CAMP如下。",
"数。但是，我们希望规范可以扩展到此选择之外。",
"8:",
"if 2> then",
"CAMP(0):= (:) [DI(",
"1",
"Z (a,y;0+A) (4)",
"Normalize",
"AB",
"perturbation:",
"9:",
"←",
"c",
"△/△Bl2",
"(a,y)ED",
"10:",
"end if",
"11:",
"end for",
"其中E是用作超参数的小正值",
"",
"Figure1草拟了一个示例，演示了在θ上最小化",
"12:",
"Compute gradient:",
"AMP损耗CAMP提供了找到ERM损耗CERM的更",
"VJAMP,B - Em1 Vel(ai,Y; 0k+AB)/m",
"平坦最小值的机会。4节给出了AMP寻找更平坦最",
"13:",
"Updateparameter:0+1←knVJAMP,B",
"小值的能力的理论依据",
"14: end while",
"3.2. Training Algorithm",
"更准确地说，假设①=RK。对于任何标量，任",
"对计算复杂性和训练速度的实际考虑激发了一种",
"小型批处理训练方法，以最大程度地减少AMP损失",
"何正标量A，任何μE和任何K×K正定矩阵K",
"12",
"令",
"CAMP。具体来说，我们可以使用从随机小批量B获",
"得的相应损耗来近似（4）中的AMP损耗，即",
"(0; μ, K, A, C):=C-Aexp (-(0-μ)T-1(0-μ)/2)",
"1",
"(7)",
"LAMP(0) ~",
"7",
"l(α,y;0+△)",
"请注意，此函数只是空间θ上的逆高斯曲面",
"△BEB(O;e) [B|",
"max",
"(a,y)eB",
"XNETHX_1_0使用上述表示法，我们假设经验风险",
":= JAMP,B(0)",
"(5)",
"的每个最小值都可以是由这样的倒置高斯曲面近似的",
"locally，即，如果μEe给出CERM的局部最小值",
"然后，AMP损耗最小化问题可以近似为最小化预期",
"则存在一些ε>0，正定矩阵XMATHMAX3和。我",
"的批次级AMP损耗JAMP,B，正式为查找",
"们注意到，一般来说，这样的假设只是一个近似值。但",
"是对于小型e，近似值可以说是准确的",
"0AMP := arg min E JAMP,B(0)",
"(6)",
"我们将使用*(μ,,A,C)表示函数(;μ,,A,C)",
"的最小值（通过在上最小化而获得）。很明显，",
"在此公式下，每个批次B与模型参数θ上的扰动向",
"*(u,K,A,C) = C -A",
"(8)",
"内部最大化：在内部最大化中，使用固定数量的步骤，",
"量△β相关联。训练涉及嵌套在内部和外部最小化的",
"令AMP表示从派生的AMP损耗，即",
"例如N，以增加ERM损失的方向更新△B，从而获",
"得JAMP,B；外部最小化在随机批次上循环，并使用迷",
"AMP(0; μ,K,A,C):= max (0 +△; μ, K, A,C) (9)",
"你批处理SGD最小化EBJAMP,B。算法1中描述了",
"AeB(O;e)",
"精确的算法。请注意，在训练结束时，学习的0*MP用",
"我们将使用",
"AMP(μ, K, A, C)",
"表示",
"作预测的模型参数；也就是说，在测试中不应用扰动",
"AMP(O;μ,K,A,C）的最小值（在上最小化）",
"4.Theoretical JustificationsofAMP",
"Theorem 1. For any given (μ,K,A,C), the function",
"AMP(0; μ, Kk, A, C) is minimized when 0 = μ and the",
"4.1. AMP Finds Flatter Local Minima",
"minimum value is",
"为了便于获得分析结果，我们将假设LERM中每个",
"AMp(μ, K, A, C) = C - Aexp (",
"(10)",
"局部最小值的损失面可以近似为倒高斯面",
"2g2",
"3"
] | {
"arxiv_id": "2010.04925",
"language": "zh",
"page_idx": 2
} |
||
2010.04925-zh-4 | 2010.04925-zh-4.jpg | <FigureHere>
Figure 4: The operational region of AMP (left) and
how it varies with varying values of $\epsilon^2/2\sigma_1^2$ (right).
对应于局部最小值，并不比$\mu_1$平坦得多，但具有更高的经验风险值。当通过最小化 AMP 损失获得这样的解决方案时，学习的模型存在显著的欠拟合风险。因此， 通常存在$\epsilon$设置的最佳点，该设置可在所选局部最小值的平坦度和深度之间进行最佳权衡
## 4.2. AMP Regularizes Gradient Norm
从 AMP 训练算法中也可以看出 AMP 倾向于更平坦的最小值这一事实，而与上述分析无关。具体来说， 我们现在表明 AMP 算法可以被视为最小化经验风险， 对其梯度范数有一定的惩罚
要看到这一点，请考虑每批内部更新的数量$N$为 1 (这实际上在我们的实验中使用)。用 $\mathcal{J}_\mathrm{ERM,B}$ 表示使用批处理$\mathcal{B}$的经验损失$\mathcal{L}_\mathrm{ERM}$的近似值，即
(15)
$$\mathcal{J}_{\mathrm{ERM},\mathcal{B}}(\boldsymbol{\theta}):=\frac{1}{|\mathcal{B}|}\sum_{(\boldsymbol{x},\boldsymbol{y})\in\mathcal{B}}\ell(\boldsymbol{x},\boldsymbol{y};\boldsymbol{\theta})$$
Theorem 2. Let $N=1.$ Then for a sufficiently small inner learning rate $\zeta$, a minimization update step in AMP training for batch $\mathcal{B}$ is equivalent to a gradientdescent step on the following loss function with learning rate $\eta:$
$$\widetilde{\mathcal{J}}_{\mathrm{ERM},\mathcal{B}}(\boldsymbol{\theta}):=\mathcal{J}_{\mathrm{ERM},\mathcal{B}}(\boldsymbol{\theta})+\Omega(\boldsymbol{\theta})$$
(16)
where
$$\Omega(\boldsymbol{\theta}):=\begin{cases}\zeta\|\nabla_{\boldsymbol{\theta}}\mathcal{J}_{\mathrm{ERM},\mathcal{B}}(\boldsymbol{\theta})\|_2^2,&\|\zeta\nabla_{\boldsymbol{\theta}}\mathcal{J}_{\mathrm{ERM},\mathcal{B}}(\boldsymbol{\theta})\|_2\leq\epsilon\\\epsilon\|\nabla_{\boldsymbol{\theta}}\mathcal{J}_{\mathrm{ERM},\mathcal{B}}(\boldsymbol{\theta})\|_2,&\|\zeta\nabla_{\boldsymbol{\theta}}\mathcal{J}_{\mathrm{ERM},\mathcal{B}}(\boldsymbol{\theta})\|_2>\epsilon\end{cases}$$
有关证明的详细信息，请参阅附录$\overline{\mathrm{EB}}$
我们注意到，正则化项$\Omega(\boldsymbol{\theta})$在两种情况中的任何一种情况下都会惩罚$\mathcal{J}_\mathrm{ERM,B}$的梯度范数。因此，AMP 训练算法有效地尝试找到 $\mathcal{J} _\mathrm{ERM, B}$ (以及 $\mathcal{L}_\mathrm{ERM}$ )的局部最小值，这些最小值不仅具有较低的值，而且在最小值附近具有较小的梯度范数。请注意，周围具有较小梯度范数的最小值是较平坦的最小值。因此，该定理提供了 AMP 训练算法的另一个理由，除了我们从构造AMP 损失$\mathcal{L}_\mathrm{AMP}$
<FigureHere>
Figure 5: A bad classification boundary (left) usually has a poor generalization performance and is fragile to adversarial attacks. A better classification boundary ( right) keeps itself far from the training examples, and tends to generalize well.
<FigureHere>
Figure 6: AMP yields a better classification boundary ( right) than ERM does ( left) . The experiment is con- ducted on a spiral dataset [41] using a feed-forward network.
4.3. Perspectives from the Input Space
到目前为止，构造和分析都集中在参数空间上。通
过检查输入空间中的 AMP 可以获得进一步的见解
很明显，参数$\theta$对输入空间的影响不超过定义类边界并指定损失$\ell(x,y)$如何随输人$x$而变化。然后，最小化 AMP 损耗可以看作是找到一个类边界(和损耗函数$\ell(x,y)$ ),即使在应用最差的$\epsilon$有界扰动时，它在训练示例中也具有最低的平均损耗。这种最小化的结果可以说是创建了更平滑的类边界，并使训练示例不会太靠近边界。这在图 5 中进行了说明，并在图 6中进行了实验验证
从输入空间查看时，AMP 中模型参数的对抗性扰动与对抗性训练 [13] 具有某些相似性。两者之间也有很大的区别：对抗性训练可以保护模型免受对抗性攻击， 而 AMP 则可以防止过度拟合。我们进一步阐述了为什么 AMP 不同于附录 C 中的对抗训练
# 5. Experiment
我们从各种角度对 AMP 的性能进行了实证研究。首先，我们将 AMP 在基准图像分类数据集上的泛化能力与 5.1 节中几种流行的正则化方案进行了比较，包括Dropout [40] ,标签平滑[42] , Flooding [22] , MixUp [51]和对抗训练[13] 。我们还将我们的方案与 ERM [44]进行了比较，ERM [44]不使用任何正则化，而是使用$\mathcal{L}_\mathrm{ERM}$优化神经网络。我们包括一个名为随机模型
5 | [
[
"0.0948",
"0.0985",
"0.11273",
"0.09534",
"0.1168",
"0.10986",
"0.0989",
"0.113"
],
[
"0.0989",
"0.10986",
"0.11273",
"0.10986",
"0.11273",
"0.1218",
"0.0989",
"0.1218"
],
[
"0.2942",
"0.10986",
"0.3057",
"0.10986",
"0.3057",
"0.1218",
"0.2942",
"0.1218"
],
[
"0.3381",
"0.1067",
"0.353",
"0.1067",
"0.353",
"0.1111",
"0.3381",
"0.1111"
],
[
"0.0907",
"0.1187",
"0.1168",
"0.1187",
"0.1168",
"0.1433",
"0.0907",
"0.1433"
],
[
"0.29",
"0.1193",
"0.3057",
"0.1193",
"0.3057",
"0.1414",
"0.29",
"0.1414"
],
[
"0.339",
"0.11554",
"0.3472",
"0.11554",
"0.3472",
"0.1187",
"0.339",
"0.1187"
],
[
"0.0972",
"0.1401",
"0.1136",
"0.1395",
"0.1168",
"0.1635",
"0.10046",
"0.1642"
],
[
"0.2957",
"0.1414",
"0.3064",
"0.1414",
"0.3064",
"0.154",
"0.2957",
"0.154"
],
[
"0.2957",
"0.1515",
"0.308",
"0.1515",
"0.308",
"0.1635",
"0.2957",
"0.1635"
],
[
"0.09644",
"0.1642",
"0.1111",
"0.1578",
"0.1209",
"0.1718",
"0.1054",
"0.178"
],
[
"0.2974",
"0.1604",
"0.308",
"0.1604",
"0.308",
"0.1749",
"0.2974",
"0.1749"
],
[
"0.10706",
"0.1818",
"0.1356",
"0.1818",
"0.1356",
"0.1863",
"0.10706",
"0.1863"
],
[
"0.1365",
"0.1818",
"0.1478",
"0.1818",
"0.1478",
"0.1849",
"0.1365",
"0.1849"
],
[
"0.3367",
"0.1787",
"0.343",
"0.1831",
"0.339",
"0.1881",
"0.3325",
"0.1843"
],
[
"0.3945",
"0.1818",
"0.4084",
"0.1818",
"0.4084",
"0.1849",
"0.3945",
"0.1849"
],
[
"0.411",
"0.1818",
"0.4314",
"0.1818",
"0.4314",
"0.1863",
"0.411",
"0.1863"
],
[
"0.4495",
"0.1818",
"0.4575",
"0.1818",
"0.4575",
"0.1849",
"0.4495",
"0.1849"
],
[
"0.5034",
"0.1913",
"0.8906",
"0.1913",
"0.8906",
"0.2052",
"0.5034",
"0.2052"
],
[
"0.0809",
"0.2102",
"0.4697",
"0.2102",
"0.4697",
"0.2235",
"0.0809",
"0.2235"
],
[
"0.5024",
"0.207",
"0.8906",
"0.207",
"0.8906",
"0.2203",
"0.5024",
"0.2203"
],
[
"0.0784",
"0.2235",
"0.4412",
"0.2247",
"0.4412",
"0.2393",
"0.0784",
"0.238"
],
[
"0.5024",
"0.2203",
"0.8906",
"0.2216",
"0.8906",
"0.2361",
"0.5024",
"0.2349"
],
[
"0.5034",
"0.2374",
"0.892",
"0.2374",
"0.892",
"0.2507",
"0.5034",
"0.2507"
],
[
"0.501",
"0.252",
"0.6733",
"0.252",
"0.6733",
"0.2651",
"0.501",
"0.2651"
],
[
"0.0809",
"0.2695",
"0.4673",
"0.2695",
"0.4673",
"0.2827",
"0.0809",
"0.2827"
],
[
"0.0809",
"0.2842",
"0.4673",
"0.2842",
"0.4673",
"0.2974",
"0.0809",
"0.2974"
],
[
"0.0809",
"0.2998",
"0.468",
"0.2998",
"0.468",
"0.3132",
"0.0809",
"0.3132"
],
[
"0.0825",
"0.3145",
"0.468",
"0.3145",
"0.468",
"0.3276",
"0.0825",
"0.3276"
],
[
"0.0809",
"0.33",
"0.3423",
"0.33",
"0.3423",
"0.3435",
"0.0809",
"0.3435"
],
[
"0.0809",
"0.3567",
"0.3782",
"0.3567",
"0.3782",
"0.3706",
"0.0809",
"0.3706"
],
[
"0.10046",
"0.3806",
"0.468",
"0.3806",
"0.468",
"0.3945",
"0.10046",
"0.3945"
],
[
"0.0809",
"0.3965",
"0.468",
"0.3965",
"0.468",
"0.4104",
"0.0809",
"0.4104"
],
[
"0.0801",
"0.4104",
"0.4697",
"0.411",
"0.4697",
"0.4255",
"0.0801",
"0.4248"
],
[
"0.505",
"0.4185",
"0.888",
"0.4185",
"0.888",
"0.4287",
"0.505",
"0.4287"
],
[
"0.0809",
"0.4268",
"0.2778",
"0.4268",
"0.2778",
"0.44",
"0.0809",
"0.44"
],
[
"0.505",
"0.432",
"0.8906",
"0.432",
"0.8906",
"0.4458",
"0.505",
"0.4458"
],
[
"0.1013",
"0.443",
"0.4697",
"0.443",
"0.4697",
"0.457",
"0.1013",
"0.457"
],
[
"0.501",
"0.4458",
"0.892",
"0.4463",
"0.892",
"0.461",
"0.501",
"0.4602"
],
[
"0.0809",
"0.4583",
"0.468",
"0.4583",
"0.468",
"0.473",
"0.0809",
"0.473"
],
[
"0.5034",
"0.4634",
"0.5645",
"0.4634",
"0.5645",
"0.4749",
"0.5034",
"0.4749"
],
[
"0.0809",
"0.473",
"0.3857",
"0.473",
"0.3857",
"0.4873",
"0.0809",
"0.4873"
],
[
"0.267",
"0.502",
"0.2737",
"0.507",
"0.2688",
"0.511",
"0.2615",
"0.506"
],
[
"0.2893",
"0.5024",
"0.326",
"0.5024",
"0.326",
"0.5234",
"0.2893",
"0.5234"
],
[
"0.3162",
"0.506",
"0.3987",
"0.506",
"0.3987",
"0.5234",
"0.3162",
"0.5234"
],
[
"0.4355",
"0.507",
"0.4697",
"0.507",
"0.4697",
"0.523",
"0.4355",
"0.523"
],
[
"0.5034",
"0.504",
"0.8022",
"0.505",
"0.8022",
"0.5195",
"0.5034",
"0.518"
],
[
"0.2566",
"0.5166",
"0.2834",
"0.5166",
"0.2834",
"0.529",
"0.2566",
"0.529"
],
[
"0.152",
"0.5083",
"0.2467",
"0.5083",
"0.2467",
"0.523",
"0.152",
"0.523"
],
[
"0.2778",
"0.526",
"0.3308",
"0.526",
"0.3308",
"0.537",
"0.2778",
"0.537"
],
[
"0.524",
"0.529",
"0.892",
"0.529",
"0.892",
"0.5425",
"0.524",
"0.5425"
],
[
"0.5034",
"0.5425",
"0.863",
"0.5425",
"0.863",
"0.556",
"0.5034",
"0.556"
],
[
"0.0809",
"0.553",
"0.468",
"0.553",
"0.468",
"0.5664",
"0.0809",
"0.5664"
],
[
"0.524",
"0.5596",
"0.8896",
"0.5596",
"0.8896",
"0.5728",
"0.524",
"0.5728"
],
[
"0.0801",
"0.5684",
"0.468",
"0.5684",
"0.468",
"0.582",
"0.0801",
"0.582"
],
[
"0.505",
"0.5747",
"0.8906",
"0.5747",
"0.8906",
"0.5884",
"0.505",
"0.5884"
],
[
"0.0809",
"0.582",
"0.4673",
"0.582",
"0.4673",
"0.595",
"0.0809",
"0.595"
],
[
"0.505",
"0.589",
"0.8906",
"0.589",
"0.8906",
"0.603",
"0.505",
"0.603"
],
[
"0.0784",
"0.5967",
"0.468",
"0.5986",
"0.468",
"0.6123",
"0.0784",
"0.61"
],
[
"0.505",
"0.604",
"0.8906",
"0.604",
"0.8906",
"0.6187",
"0.505",
"0.6187"
],
[
"0.0784",
"0.6123",
"0.1356",
"0.6143",
"0.134",
"0.629",
"0.0768",
"0.627"
],
[
"0.505",
"0.619",
"0.8906",
"0.619",
"0.8906",
"0.6333",
"0.505",
"0.6333"
],
[
"0.505",
"0.634",
"0.8896",
"0.634",
"0.8896",
"0.648",
"0.505",
"0.648"
],
[
"0.1584",
"0.6426",
"0.3906",
"0.6426",
"0.3906",
"0.657",
"0.1584",
"0.657"
],
[
"0.4387",
"0.643",
"0.468",
"0.643",
"0.468",
"0.656",
"0.4387",
"0.656"
],
[
"0.505",
"0.6494",
"0.892",
"0.6494",
"0.892",
"0.6636",
"0.505",
"0.6636"
],
[
"0.505",
"0.664",
"0.636",
"0.664",
"0.636",
"0.678",
"0.505",
"0.678"
],
[
"0.0784",
"0.6724",
"0.1259",
"0.6724",
"0.1259",
"0.6875",
"0.0784",
"0.6875"
],
[
"0.524",
"0.6797",
"0.8906",
"0.6797",
"0.8906",
"0.6943",
"0.524",
"0.6943"
],
[
"0.505",
"0.6943",
"0.888",
"0.6943",
"0.888",
"0.7085",
"0.505",
"0.7085"
],
[
"0.1487",
"0.7007",
"0.4656",
"0.7007",
"0.4656",
"0.7153",
"0.1487",
"0.7153"
],
[
"0.0809",
"0.711",
"0.134",
"0.711",
"0.134",
"0.7227",
"0.0809",
"0.7227"
],
[
"0.505",
"0.7104",
"0.888",
"0.7104",
"0.888",
"0.7236",
"0.505",
"0.7236"
],
[
"0.1487",
"0.72",
"0.4656",
"0.72",
"0.4656",
"0.735",
"0.1487",
"0.735"
],
[
"0.505",
"0.726",
"0.892",
"0.726",
"0.892",
"0.7393",
"0.505",
"0.7393"
],
[
"0.437",
"0.7363",
"0.468",
"0.7363",
"0.468",
"0.752",
"0.437",
"0.752"
],
[
"0.5024",
"0.738",
"0.8",
"0.7393",
"0.8",
"0.754",
"0.5024",
"0.7524"
],
[
"0.10046",
"0.7646",
"0.3733",
"0.763",
"0.3733",
"0.778",
"0.10046",
"0.779"
],
[
"0.5024",
"0.771",
"0.6313",
"0.771",
"0.6313",
"0.785",
"0.5024",
"0.785"
],
[
"0.1013",
"0.78",
"0.4617",
"0.78",
"0.4617",
"0.7935",
"0.1013",
"0.7935"
],
[
"0.0801",
"0.7944",
"0.4697",
"0.7935",
"0.4697",
"0.811",
"0.0801",
"0.8125"
],
[
"0.523",
"0.7954",
"0.8906",
"0.7944",
"0.8906",
"0.8096",
"0.523",
"0.81"
],
[
"0.0825",
"0.811",
"0.4673",
"0.811",
"0.4673",
"0.8257",
"0.0825",
"0.8257"
],
[
"0.5063",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.5063",
"0.8247"
],
[
"0.0809",
"0.8247",
"0.4673",
"0.8257",
"0.4673",
"0.8403",
"0.0809",
"0.839"
],
[
"0.5034",
"0.8247",
"0.8906",
"0.8257",
"0.8906",
"0.8403",
"0.5034",
"0.839"
],
[
"0.08417",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.08417",
"0.855"
],
[
"0.5034",
"0.8413",
"0.892",
"0.8413",
"0.892",
"0.855",
"0.5034",
"0.855"
],
[
"0.0825",
"0.858",
"0.4673",
"0.858",
"0.4673",
"0.8687",
"0.0825",
"0.8687"
],
[
"0.505",
"0.856",
"0.892",
"0.856",
"0.892",
"0.869",
"0.505",
"0.869"
],
[
"0.0809",
"0.872",
"0.468",
"0.872",
"0.468",
"0.8853",
"0.0809",
"0.8853"
],
[
"0.5063",
"0.872",
"0.8906",
"0.872",
"0.8906",
"0.8853",
"0.5063",
"0.8853"
],
[
"0.0801",
"0.884",
"0.2051",
"0.887",
"0.2051",
"0.9014",
"0.0801",
"0.8984"
],
[
"0.5034",
"0.8853",
"0.8906",
"0.8853",
"0.8906",
"0.903",
"0.5034",
"0.903"
],
[
"0.478",
"0.924",
"0.4944",
"0.924",
"0.4944",
"0.938",
"0.478",
"0.938"
]
] | [
"175",
"1.",
"15.0",
"02",
"",
"",
"",
"",
"",
"5.0",
"25/",
"",
"00",
"",
"o",
"1",
"0.8",
"1",
"Figure 5: A bad classification boundary (left) usually",
"Figure 4: The operational region of AMP (left)and",
"has a poor generalization performance and is fragile to",
"how it varies with varying values of e2 /2α2 (right)",
"adversarial attacks. A better classification boundary",
"(right) keeps itself far from the training examples, and",
"tends to generalize well",
"对应于局部最小值，并不比叫平坦得多，但具有更高",
"的经验风险值。当通过最小化AMP损失获得这样的解",
"决方案时，学习的模型存在显著的欠拟合风险。因此，",
"通常存在设置的最佳点，该设置可在所选局部最小",
"值的平坦度和深度之间进行最佳权衡",
"4.2. AMP Regularizes Gradient Norm",
"从AMP训练算法中也可以看出AMP倾向于更平",
"坦的最小值这一事实，而与上述分析无关。具体来说，",
"我们现在表明AMP算法可以被视为最小化经验风险，",
"Figure6:AMP yields a better classification boundary",
"对其梯度范数有一定的惩罚",
"(right) than ERM does (left). The experiment is con",
"要看到这一点，请考虑每批内部更新的数量N为1",
"ducted on a spiral dataset 41] using a feed-forward",
"（这实际上在我们的实验中使用）。用JERM,B表示使用",
"network.",
"批处理B的经验损失LERM的近似值，即",
"",
"7",
"l(a,y;0)",
"(15)",
"4.3. Perspectives from the Input Space",
"[B",
"JERM,B(0) :",
"(r,y)eB",
"到目前为止，构造和分析都集中在参数空间上。通",
"过检查输人空间中的AMP可以获得进一步的见解",
"Theorem 2. Let N =1 . Then for a sufficiently small",
"很明显，参数θ对输入空间的影响不超过定义类边",
"inner learning rate S , a minimization update step in",
"界并指定损失l(a,y)如何随输人α而变化。然后，最",
"AMP training for batch B is equivalent to a gradient",
"小化AMP损耗可以看作是找到一个类边界（和损耗",
"descentsteponthefollowinglossfunctionwithlearning",
"函数l(c,y）），即使在应用最差的e有界扰动时，它",
"raten:",
"在训练示例中也具有最低的平均损耗。这种最小化的",
"结果可以说是创建了更平滑的类边界，并使训练示例",
"JERM,B(0) := JERM.B(0) + 2(0)",
"(16)",
"不会太靠近边界。这在图5中进行了说明，并在图6",
"中进行了实验验证",
"where",
"从输人空间查看时，AMP中模型参数的对抗性扰动",
"与对抗性训练[13]具有某些相似性。两者之间也有很",
"S/VeJERM,B(O)I2, IISVeJERM,B(O)I2 ≤ e",
"2(0) :=",
"大的区别：对抗性训练可以保护模型免受对抗性攻击",
"elVJERM,B(0)I2,IICVeJERM,B(0)Il2 > e",
"而AMP则可以防止过度拟合。我们进一步阐述了为",
"(17)",
"什么AMP不同于附录 C 中的对抗训练",
"有关证明的详细信息，请参阅附录EB",
"5.Experiment",
"我们注意到，正则化项2(0)在两种情况中的任何一",
"种情况下都会惩罚JeRM,B的梯度范数。因此，AMP",
"我们从各种角度对AMP的性能进行了实证研究。",
"训练算法有效地尝试找到JERM,B（以及LERM）的",
"首先，我们将AMP在基准图像分类数据集上的泛化能",
"局部最小值，这些最小值不仅具有较低的值，而且在最",
"力与5.1节中几种流行的正则化方案进行了比较，包括",
"小值附近具有较小的梯度范数。请注意，周围具有较小",
"Dropout[40]，标签平滑[42]，Flooding[22]，MixUp",
"梯度范数的最小值是较平坦的最小值。因此，该定理提",
"[51]和对抗训练[13]。我们还将我们的方案与ERM",
"供了AMP训练算法的另一个理由，除了我们从构造",
"[44]进行了比较，ERM[44]不使用任何正则化，而是使",
"AMP损失CAMP",
"用CERM优化神经网络。我们包括一个名为随机模型",
"5"
] | {
"arxiv_id": "2010.04925",
"language": "zh",
"page_idx": 4
} |
||
2201.13322-zh-1 | 2201.13322-zh-1.jpg | # 学习哈希自然排序
Jiaguo Yu$^1*$, Yuming Shen$^2*$, Menghan Wang$^3$, Haofeng Zhang$^1\dagger$, Philip H.S. Torr$^2$
$^{1}$Nanjing University of Science and Technology
$^{2}$University of Oxford
$^3$eBay
$\operatorname{Abstract}$
Learning to hash pictures a list-wise sorting problem. Its testing metrics, e.g., meanaverage precision, count on a sorted candidate list ordered by pair-wise code similarity. However, scarcely does one train a deep hashing model with the sorted results end-to-end because of the non-differentiable nature of the sorting operation. This inconsistency in the objectives of training and test may lead to suboptimal performance since the training loss often fails to reflect the actual retrieval metric. In this paper, we tackle this problem by introducing Naturally-Sorted Hashing (NSH). We sort the Hamming distances of samples' hash codes and accordingly gather their latent representations for self-supervised training. Thanks to the recent advances in differentiable sorting approximations, the hash head receives gradients from the sorter so that the hash encoder can be optimized along with the training procedure. Additionally, we describe a novel Sorted NoiseContrastive Estimation (SortedNCE) loss that selectively picks positive and negative samples for contrastive learning, which allows NSH to mine data semantic relations during training in an unsupervised manner. Our extensive experiments show the proposed NSH model significantly outperforms the existing unsuper- vised hashing methods on three benchmarked datasets.
## 1 Introduction
学习哈希$\mathbb{E}[$Gionis et al.,1999] ,自然被视为深度学习中的表示学习任务，实际上是学习大规模排序的近似最近邻 (ANN) 搜索的子类。代码在查询和一堆候选项之间的汉明距离测量它们的相关程度，并进一步确定它们在检索结果中的存在顺序。因此，传统的评估指标明确反映了检索到的候选对象的位置和顺序敏感性，这涉及argsort 过程，包括推荐系统中的均值-平均精度 (mAP),
$^{*}$Equal contribution.
${}^{*}$Corresponding author.
<FigureHere>
Figure 1: A brief motivation of NSH. (a) The actual testing metrics of learning to hash involves non-differentiable argsort operators. Hence, they can not be directly used for training. ( $\bar{b}$) The proposed NSH model best mimics the testing proce- dure that sorts the code similarity with soft approximations and is trained with a list-wise SortedNCE objective end-toend.
最高$k$精度甚至归一化贴现累积增益 (NDCG)。奇怪的是，大多数现有的深度哈希模型通常不实现排序的概念， 而是采用一些替代的学习目标，例如识别和成对/三元组损失[Luo et al., 2021b]。这种反直觉的惯例提出了一个问题，我们可以并且应该在排序后训练深度哈希模型以与其评估指标相匹配吗？
回顾深度学习中的最佳实践之一，模型的训练目标应明确表示其最终目标或测试度量。例如，分割模型通常由联合交叉(IoU)得分进行评估，并以类似的 IoU 损失进行训练；密度估计模型期望描述样本密度的似然目标。培训和测试之间的连贯性使其在提高性能方面变得不平凡。然而，在无监督哈希的背景下，实践这一愿景存在两个主要挑战
The Non-Differentiable Kinks in Sorting and ANN 如上所述，诸如 mAP 和 top- $k$精度之类的主流测量基于argsort 操作，该操作不会得出从实际分数到哈希层的可区分路径。这使得不可能构建使用排序输出进行训练的端到端深度框架。我们在 Fig. 1 (a) 中说明了这个问题。尽管可以将相关性顺序转换为正/负数据对，以促进基于排名的训练，但毕竟，此解决方案仅代表排序检索列表的一部分，因此会丢弃一些可能与[Wang et al., 2019] 有关的信息。此外，它还将我们引向第二个问题 i.e.,如何以无监督的方式确定相似性的顺序？
The Lack of Clues of Relevance 与其相似度标签现成的有监督同级不同，无监督哈希仅观察数据本身。这也阻碍了现有模型在训练期间模仿基于排序的评估指标。尽管可以在交替方案$\boxed{\mathrm{E}}[$Shen et al., 2018; Yang et al., 2019] | [
[
"0.406",
"0.1111",
"0.5923",
"0.113",
"0.5923",
"0.1307",
"0.406",
"0.1288"
],
[
"0.1381",
"0.1578",
"0.8613",
"0.1578",
"0.8613",
"0.1724",
"0.1381",
"0.1724"
],
[
"0.3015",
"0.1787",
"0.694",
"0.1812",
"0.694",
"0.1957",
"0.3015",
"0.1932"
],
[
"0.4043",
"0.1945",
"0.5933",
"0.1963",
"0.5933",
"0.214",
"0.4043",
"0.2122"
],
[
"0.4722",
"0.2134",
"0.524",
"0.2166",
"0.523",
"0.2311",
"0.4714",
"0.2279"
],
[
"0.2484",
"0.2715",
"0.3245",
"0.2734",
"0.3245",
"0.2886",
"0.2484",
"0.286"
],
[
"0.572",
"0.2747",
"0.669",
"0.2747",
"0.669",
"0.286",
"0.572",
"0.286"
],
[
"0.79",
"0.276",
"0.841",
"0.276",
"0.841",
"0.286",
"0.79",
"0.286"
],
[
"0.523",
"0.2917",
"0.5933",
"0.2917",
"0.5933",
"0.31",
"0.523",
"0.31"
],
[
"0.5884",
"0.2954",
"0.6274",
"0.2954",
"0.6274",
"0.3062",
"0.5884",
"0.3062"
],
[
"0.6235",
"0.2954",
"0.675",
"0.2954",
"0.675",
"0.307",
"0.6235",
"0.307"
],
[
"0.7646",
"0.2954",
"0.8022",
"0.2954",
"0.8022",
"0.307",
"0.7646",
"0.307"
],
[
"0.7983",
"0.2966",
"0.8667",
"0.2942",
"0.8667",
"0.305",
"0.7983",
"0.307"
],
[
"0.1177",
"0.3113",
"0.452",
"0.3113",
"0.452",
"0.3245",
"0.1177",
"0.3245"
],
[
"0.1177",
"0.2966",
"0.4534",
"0.2966",
"0.4534",
"0.3113",
"0.1177",
"0.3113"
],
[
"0.546",
"0.315",
"0.564",
"0.315",
"0.564",
"0.3245",
"0.546",
"0.3245"
],
[
"0.683",
"0.3176",
"0.694",
"0.3193",
"0.693",
"0.3245",
"0.681",
"0.3225"
],
[
"0.7607",
"0.3132",
"0.8677",
"0.3132",
"0.8677",
"0.327",
"0.7607",
"0.327"
],
[
"0.1168",
"0.3257",
"0.4534",
"0.324",
"0.4534",
"0.3372",
"0.1168",
"0.339"
],
[
"0.6626",
"0.324",
"0.7114",
"0.3225",
"0.7114",
"0.3333",
"0.6626",
"0.334"
],
[
"0.1177",
"0.339",
"0.452",
"0.339",
"0.452",
"0.3528",
"0.1177",
"0.3528"
],
[
"0.11926",
"0.3528",
"0.452",
"0.3528",
"0.452",
"0.3662",
"0.11926",
"0.3662"
],
[
"0.11926",
"0.3687",
"0.451",
"0.3687",
"0.451",
"0.379",
"0.11926",
"0.379"
],
[
"0.523",
"0.3662",
"0.55",
"0.3718",
"0.546",
"0.3833",
"0.5186",
"0.3774"
],
[
"0.5986",
"0.3694",
"0.6436",
"0.3694",
"0.6436",
"0.3806",
"0.5986",
"0.3806"
],
[
"0.714",
"0.3694",
"0.7417",
"0.3694",
"0.7417",
"0.3806",
"0.714",
"0.3806"
],
[
"0.7915",
"0.3694",
"0.8364",
"0.3694",
"0.8364",
"0.3806",
"0.7915",
"0.3806"
],
[
"0.1177",
"0.3801",
"0.4534",
"0.3801",
"0.4534",
"0.3945",
"0.1177",
"0.3945"
],
[
"0.6006",
"0.3806",
"0.614",
"0.3806",
"0.614",
"0.3901",
"0.6006",
"0.3901"
],
[
"0.796",
"0.3801",
"0.8096",
"0.3801",
"0.8096",
"0.3901",
"0.796",
"0.3901"
],
[
"0.1177",
"0.3945",
"0.452",
"0.3945",
"0.452",
"0.4077",
"0.1177",
"0.4077"
],
[
"0.511",
"0.3958",
"0.9126",
"0.3965",
"0.9126",
"0.411",
"0.511",
"0.4104"
],
[
"0.11926",
"0.4077",
"0.451",
"0.4077",
"0.451",
"0.4216",
"0.11926",
"0.4216"
],
[
"0.513",
"0.4092",
"0.911",
"0.4092",
"0.911",
"0.4224",
"0.513",
"0.4224"
],
[
"0.1168",
"0.4216",
"0.451",
"0.4216",
"0.451",
"0.435",
"0.1168",
"0.435"
],
[
"0.5107",
"0.4204",
"0.91",
"0.4216",
"0.91",
"0.4363",
"0.5107",
"0.435"
],
[
"0.1209",
"0.4375",
"0.452",
"0.4375",
"0.452",
"0.4475",
"0.1209",
"0.4475"
],
[
"0.513",
"0.433",
"0.9087",
"0.433",
"0.9087",
"0.4475",
"0.513",
"0.4475"
],
[
"0.11926",
"0.4495",
"0.452",
"0.4495",
"0.452",
"0.4634",
"0.11926",
"0.4634"
],
[
"0.511",
"0.4463",
"0.9126",
"0.4463",
"0.9126",
"0.4602",
"0.511",
"0.4602"
],
[
"0.1177",
"0.4634",
"0.452",
"0.4634",
"0.452",
"0.4766",
"0.1177",
"0.4766"
],
[
"0.511",
"0.459",
"0.911",
"0.459",
"0.911",
"0.473",
"0.511",
"0.473"
],
[
"0.511",
"0.473",
"0.5415",
"0.473",
"0.5415",
"0.4841",
"0.511",
"0.4841"
],
[
"0.11926",
"0.4792",
"0.452",
"0.4792",
"0.452",
"0.4893",
"0.11926",
"0.4893"
],
[
"0.1177",
"0.4912",
"0.452",
"0.4912",
"0.452",
"0.505",
"0.1177",
"0.505"
],
[
"0.513",
"0.4988",
"0.911",
"0.4976",
"0.911",
"0.512",
"0.513",
"0.513"
],
[
"0.1177",
"0.505",
"0.4534",
"0.505",
"0.4534",
"0.518",
"0.1177",
"0.518"
],
[
"0.1168",
"0.5176",
"0.452",
"0.518",
"0.452",
"0.5327",
"0.1168",
"0.532"
],
[
"0.513",
"0.513",
"0.9067",
"0.513",
"0.9067",
"0.5264",
"0.513",
"0.5264"
],
[
"0.5156",
"0.5264",
"0.911",
"0.5264",
"0.911",
"0.5405",
"0.5156",
"0.5405"
],
[
"0.11926",
"0.5327",
"0.4534",
"0.5327",
"0.4534",
"0.547",
"0.11926",
"0.547"
],
[
"0.11926",
"0.547",
"0.4534",
"0.547",
"0.4534",
"0.56",
"0.11926",
"0.56"
],
[
"0.5146",
"0.5415",
"0.91",
"0.5415",
"0.91",
"0.552",
"0.5146",
"0.552"
],
[
"0.513",
"0.553",
"0.9126",
"0.5537",
"0.9126",
"0.5684",
"0.513",
"0.5674"
],
[
"0.11926",
"0.56",
"0.452",
"0.56",
"0.452",
"0.5737",
"0.11926",
"0.5737"
],
[
"0.5156",
"0.5674",
"0.687",
"0.5674",
"0.687",
"0.5806",
"0.5156",
"0.5806"
],
[
"0.11926",
"0.5737",
"0.452",
"0.5737",
"0.452",
"0.587",
"0.11926",
"0.587"
],
[
"0.1209",
"0.589",
"0.452",
"0.589",
"0.452",
"0.5996",
"0.1209",
"0.5996"
],
[
"0.5312",
"0.5835",
"0.911",
"0.5835",
"0.911",
"0.5967",
"0.5312",
"0.5967"
],
[
"0.1177",
"0.6016",
"0.4534",
"0.6016",
"0.4534",
"0.6157",
"0.1177",
"0.6157"
],
[
"0.5156",
"0.5967",
"0.911",
"0.5967",
"0.911",
"0.61",
"0.5156",
"0.61"
],
[
"0.1177",
"0.6157",
"0.4534",
"0.6157",
"0.4534",
"0.629",
"0.1177",
"0.629"
],
[
"0.5156",
"0.61",
"0.911",
"0.61",
"0.911",
"0.625",
"0.5156",
"0.625"
],
[
"0.1177",
"0.629",
"0.4534",
"0.629",
"0.4534",
"0.6426",
"0.1177",
"0.6426"
],
[
"0.5146",
"0.625",
"0.9126",
"0.625",
"0.9126",
"0.638",
"0.5146",
"0.638"
],
[
"0.11926",
"0.643",
"0.451",
"0.643",
"0.451",
"0.657",
"0.11926",
"0.657"
],
[
"0.513",
"0.638",
"0.9126",
"0.638",
"0.9126",
"0.6514",
"0.513",
"0.6514"
],
[
"0.1177",
"0.656",
"0.451",
"0.656",
"0.451",
"0.6694",
"0.1177",
"0.6694"
],
[
"0.513",
"0.651",
"0.911",
"0.651",
"0.911",
"0.664",
"0.513",
"0.664"
],
[
"0.1168",
"0.6704",
"0.451",
"0.6704",
"0.451",
"0.6846",
"0.1168",
"0.6846"
],
[
"0.5156",
"0.6675",
"0.613",
"0.6675",
"0.613",
"0.6787",
"0.5156",
"0.6787"
],
[
"0.11926",
"0.685",
"0.4534",
"0.685",
"0.4534",
"0.6987",
"0.11926",
"0.6987"
],
[
"0.5186",
"0.6875",
"0.8774",
"0.6875",
"0.8774",
"0.7007",
"0.5186",
"0.7007"
],
[
"0.884",
"0.6914",
"0.8906",
"0.6914",
"0.8906",
"0.6963",
"0.884",
"0.6963"
],
[
"0.8945",
"0.69",
"0.911",
"0.69",
"0.911",
"0.6987",
"0.8945",
"0.6987"
],
[
"0.11926",
"0.6997",
"0.183",
"0.6997",
"0.183",
"0.711",
"0.11926",
"0.711"
],
[
"0.517",
"0.702",
"0.911",
"0.702",
"0.911",
"0.7153",
"0.517",
"0.7153"
],
[
"0.513",
"0.7153",
"0.911",
"0.7153",
"0.911",
"0.729",
"0.513",
"0.729"
],
[
"0.0866",
"0.7344",
"0.1095",
"0.7344",
"0.1095",
"0.7456",
"0.0866",
"0.7456"
],
[
"0.1177",
"0.7344",
"0.2296",
"0.7344",
"0.2296",
"0.7476",
"0.1177",
"0.7476"
],
[
"0.5156",
"0.729",
"0.911",
"0.729",
"0.911",
"0.742",
"0.5156",
"0.742"
],
[
"0.10706",
"0.7373",
"0.1136",
"0.7373",
"0.1136",
"0.742",
"0.10706",
"0.742"
],
[
"0.5156",
"0.742",
"0.91",
"0.742",
"0.91",
"0.756",
"0.5156",
"0.756"
],
[
"0.08496",
"0.752",
"0.4844",
"0.7524",
"0.4844",
"0.7676",
"0.08496",
"0.7666"
],
[
"0.513",
"0.755",
"0.911",
"0.757",
"0.911",
"0.7715",
"0.513",
"0.7695"
],
[
"0.08905",
"0.7676",
"0.4844",
"0.7676",
"0.4844",
"0.781",
"0.08905",
"0.781"
],
[
"0.513",
"0.771",
"0.911",
"0.771",
"0.911",
"0.784",
"0.513",
"0.784"
],
[
"0.08826",
"0.781",
"0.4836",
"0.781",
"0.4836",
"0.7944",
"0.08826",
"0.7944"
],
[
"0.5146",
"0.784",
"0.911",
"0.784",
"0.911",
"0.7974",
"0.5146",
"0.7974"
],
[
"0.08826",
"0.7954",
"0.4844",
"0.7954",
"0.4844",
"0.8096",
"0.08826",
"0.8096"
],
[
"0.5146",
"0.799",
"0.9126",
"0.799",
"0.9126",
"0.8125",
"0.5146",
"0.8125"
],
[
"0.08826",
"0.8096",
"0.4844",
"0.8096",
"0.4844",
"0.8228",
"0.08826",
"0.8228"
],
[
"0.5156",
"0.8145",
"0.7837",
"0.8145",
"0.7837",
"0.8247",
"0.5156",
"0.8247"
],
[
"0.08826",
"0.8237",
"0.4844",
"0.8237",
"0.4844",
"0.837",
"0.08826",
"0.837"
],
[
"0.5186",
"0.8335",
"0.7646",
"0.8335",
"0.7646",
"0.8467",
"0.5186",
"0.8467"
],
[
"0.761",
"0.832",
"0.91",
"0.8335",
"0.91",
"0.8467",
"0.761",
"0.845"
],
[
"0.5156",
"0.8477",
"0.911",
"0.8477",
"0.911",
"0.8613",
"0.5156",
"0.8613"
],
[
"0.0866",
"0.837",
"0.486",
"0.837",
"0.486",
"0.851",
"0.0866",
"0.851"
],
[
"0.10706",
"0.8613",
"0.2418",
"0.8613",
"0.2418",
"0.875",
"0.10706",
"0.875"
],
[
"0.513",
"0.8613",
"0.91",
"0.86",
"0.91",
"0.875",
"0.513",
"0.8755"
],
[
"0.10864",
"0.8735",
"0.2615",
"0.875",
"0.2615",
"0.8896",
"0.10864",
"0.888"
],
[
"0.5146",
"0.8755",
"0.9126",
"0.8755",
"0.9126",
"0.8896",
"0.5146",
"0.8896"
]
] | [
"学习哈希自然排序",
"Jiaguo Yul*, Yuming Shen2*, Menghan Wang3, Haofeng Zhanglt,Philip H.S. Torr?",
"INanjing University of Science and Technology",
"2University of Oxford",
"3eBay",
"Abstract",
"Retrieval Evaluation",
"SortedNCE",
"Casarede.C",
"！",
"0",
"",
"00",
"ing problem. Its testing metrics, e.g., mean-",
"Learning to hash pictures a list-wise sort-",
"Sort",
"No",
"NeuralSorwithHashCode",
"average precision, count on a sorted candidate",
"Differentiable",
"list ordered by pair-wise code similarity. How",
"ever, scarcely does one train a deep hashing",
"model with the sorted results end-to-end be-",
"Query",
"Candidates",
"Query",
"Candidates",
"cause of the non-differentiable nature of the",
"(a) ",
"(b)",
"sorting operation. This inconsistency in the",
"Figure l: A brief motivation of NSH. (a) The actual testing",
"objectives of training and test may lead to sub-",
"metrics of learning to hash involves non-differentiable argsort",
"optimal performance since the training loss of-",
"operators. Hence, they can not be directly used for training.",
"tenfailstoreflecttheactualretrievalmetric.In",
"(b) The proposed NSH model best mimics the testing proce",
"this paper, we tackle this problem by introduc-",
"dure that sorts the code similarity with soft approximations",
"ing Naturally-Sorted Hashing (NSH). We sort",
"and is trained with a list-wise SortedNCE objective end-to-",
"end.",
"theHammingdistancesofsampleshashcodes",
"and accordingly gather their latent represen-",
"最高k精度甚至归一化贴现累积增益（NDCG)。奇怪的",
"tations for self-supervised training. Thanks to",
"the recent advances in differentiable sorting ap-",
"是，大多数现有的深度哈希模型通常不实现排序的概念",
"而是采用一些替代的学习目标，例如识别和成对/三元组",
"proximations, the hash head receives gradients",
"from the sorter so that the hash encoder can",
"损失Luo.etal.2021bl。这种反直觉的惯例提出了一个",
"问题，我们可以并且应该在排序后训练深度哈希模型以",
"be optimized along with the training procedure.",
"与其评估指标相匹配吗？",
"Additionally, we describe a novel Sorted Noise-",
"ContrastiveEstimation(SortedNCE)lossthat",
"回顾深度学习中的最佳实践之一，模型的训练目标应",
"selectively picks positive and negative samples",
"明确表示其最终目标或测试度量。例如，分割模型通常",
"for contrastive learning, which allows NSH to",
"由联合交叉（IoU）得分进行评估，并以类似的IoU损失",
"mine data semantic relations during training",
"进行训练；密度估计模型期望描述样本密度的似然目标。",
"in an unsupervised manner. Our extensive ex-",
"培训和测试之间的连贯性使其在提高性能方面变得不平",
"periments show the proposed NSH model sig-",
"凡。然而，在无监督哈希的背景下，实践这一愿景存在",
"nificantly outperforms the existing unsuper-",
"两个主要挑战",
"vised hashing methods on three benchmarked",
"The Non-Differentiable Kinks in Sorting and ANN",
"",
"如",
"datasets.",
"上所述，诸如mAP和top-k精度之类的主流测量基于",
"argsort操作，该操作不会得出从实际分数到哈希层的可",
"1",
"Introduction",
"区分路径。这使得不可能构建使用排序输出进行训练的",
"",
"端到端深度框架。我们在Fig：1（a）中说明了这个问题。",
"学习哈希E[Gionisetal.，1999]，自然被视为深度学习",
"尽管可以将相关性顺序转换为正/负数据对，以促进基于",
"中的表示学习任务，实际上是学习大规模排序的近似最",
"排名的训练，但毕竞，此解决方案仅代表排序检索列表",
"近邻（ANN）搜索的子类。代码在查询和一堆候选项之",
"的一部分，因此会丢弃一些可能与[Wangetal.，2019]",
"间的汉明距离测量它们的相关程度，并进一步确定它们",
"有关的信息。此外，它还将我们引向第二个问题i.e.，如",
"在检索结果中的存在顺序。因此，传统的评估指标明确",
"何以无监督的方式确定相似性的顺序？",
"反映了检索到的候选对象的位置和顺序敏感性，这涉及",
"The Lack of Clues of Relevance",
"与其相似度标签现成",
"的有监督同级不同，无监督哈希仅观察数据本身。这也阻",
"argsort过程，包括推荐系统中的均值-平均精度（mAP），",
"*Equal contribution.",
"碍了现有模型在训练期间模仿基于排序的评估指标。尽",
"*Corresponding author.",
"管可以在交替方案E[Shenetal.,2018Yangetal.,2019]"
] | {
"arxiv_id": "2201.13322",
"language": "zh",
"page_idx": 1
} |
||
2201.13322-zh-2 | 2201.13322-zh-2.jpg | 中构造成对的伪标签，但是对于深度模型，端到端训练
因此变得不可行，并且伪标签的错误可能会传播
在本文中，我们将标题作为主要动机，并通过提出自然排序哈希 (NSH) 来解决上述挑战，该哈希可以在训练过程中最大程度地呈现排序特性，从而提高检索性能。根据第一个挑战，我们采用可微分 softsort 近似值$\mathbb{E}[$Prillo and Eisenschlos, 2020] 的最新进展作为一个构建块，将底部的哈希编码器和顶部的基于排序的损失连接起来。特别是，NSH 遵循[Shen et al., 2020] , 它编码两个表示形式 i.e., 用于相似性比较的二进制哈希码和携带详细信息的连续潜在向量。样本之间的代码距离为 softsort 算子提供信息，并对潜在向量进行重新排序以进行自监督训练。为了应对第二个挑战，我们提出了一种新的排序噪声对比估计 (SortedNCE)损失，它有选择地选择最相关的样本作为对比学习的阳性样本。因此，NSH 在训练过程中挖掘数据语义关系，并动态优化每个数据的排序检索候选列表。Fig. 1 (b) 简要介绍了我们的主要思想，它更好地代表了深度哈希的测试场景。我们的主要贡献包括：
$\bullet$ For the frst time, we describe an end-to-end NSH
model that optimizes the sorted retrieval list that best represents the testing scenario of learning to hash and only requires gradient descent
$\bullet$ To implement this vision, we additionally propose
the SortedNCE loss that trains the sorted features
in an unsupervised way.
$\bullet$ We show the superiority of NSH in retrieval perfor-
mance on three benchmarked datasets against the
most recent unsupervised deep hashing methods.
## 2 Related Work
Unsupervised Hashing
无监督哈希方法的早期工作
$\dot{\text{主 要 集 中 在 学 习 紧 凑 表 示 }}$ [Gong et al.,2013]。最近的几项深度学习工作都集中在哈希代码质量上 [Li and van Gemert, 2021; Su et al., 2018; Luo et al., 2021b; Ghasedi Dizaji et al., 2018] 。其他一些人则关注代码[Shen et al., $\bar2018;$ Yang et al., 2019; Shen et al., 2020] 的语义意识，而大多数人采用伪标签方案来挖掘数据相似性，将其作为神经网络之外的单个模块。他们的表现通常在排序的候选人上进行评估。然而，他们并不寻求在训练期间实现排序的概念来挖掘相似性
对比学习是一种
Hashing with Contrastive Learning
通过构造正负对来学习没有标签的数据集的一般特征的方法 [Hadsell et al., 2006] 。实例判别 [Wu et al., 2018] 提出非参数交叉熵损失在实例级优化模型。最近的作品也使用对比学习哈希 [Qiu et al., 2021; Luo et al., 2021b] 。他们采用实例歧视作为目标，其中正面和负面实例可能仍然具有重叠的语义。也就是说，批处理中的每个查询图像仅将其增强视图视为正样本，这意味着即使对于极其相似的样本，也必须将它们分开。此外，CIBHash [Qiu et al., 2021] 和 CIMON[Luo et al., 2021b] 会调整对比损失以适应哈希学习标准，但是直接在对比损失中使用哈希代码无法很好地集成图像的语义内容。
# 3 Preliminaries
Neural Sorting Operators 香草 argsort 绝对是不可微分的，但是最近的研究发现了几种与神经网络兼容的近
<FigureHere>
Figure 2: The training schematic of NSH. We illustrate the batch-based output shapes on the right where different col- $\begin{array}{ll}\mathrm{ors~refer~to~different~instances.~}Z\in\mathbb{R}^{n\times d_{2}}\mathrm{~are~the~la-}\\\mathrm{tent~representations.~}\boldsymbol{B}\in\{-1,1\}^{n\times d_{\mathrm{b}}}\mathrm{~are~the~hash~codes.}\\\boldsymbol{S}\in[0,1]^{n\times n}\mathrm{~is~the~code-based~similarity~matrix~for~sort-}\\\mathrm{ing.~}\boldsymbol{E}\in\mathbb{R}^{n\times n\times d_{\mathrm{z}}}\mathrm{~are~}n\mathrm{~sofft-}\end{array}$ of which each is ordered by the Hamming distance to the corresponding instance in the batch for SortedNCE.
似值。尽管可以将其视为线性规划问题 [Blondel et al., 2020] ,但我们选择采用更简单的 softmax 方法 [Prillo and Eisenschlos, 2020] , 该方法将相似性得分$s\in\mathbb{R}^N$ 的向量的置换确定为：
$$\mathrm{softsort}(\boldsymbol{s})=\mathrm{softmax}\:\frac{-\mathrm{d_s}(\mathrm{sort}(\boldsymbol{s})\mathbb{1}^\intercal,\mathbb{1}\boldsymbol{s}^\intercal)}{\tau_\mathrm{s}},$$
(1)
其中$d_s$( $\cdot$) 几乎在任何地方都是任意可微分的超模，半
度量距离函数 XMATHMA。
Twin-Bottleneck Hash Encoder 到
Twin-Bottleneck Hash Encoder 尽管在深度学习中广泛见证了分解输出，但[Shen et al.,2020]指定了由$\theta$参数化的哈希编码器$f_\theta(\cdot)$的两个输出的不同功能，该编码器对数据$x$
$$[\boldsymbol{h},\boldsymbol{z}]=f_\theta(\boldsymbol{x}),\:\boldsymbol{b}=\mathrm{sign}(\boldsymbol{h}).$$
(2)
$<\mathbb{N}$。由于哈希码通常较短且信息量较小，因此$h$和$b$ 仅用于计算在注意力机制中充当查询/键的成对数据相似度，而$z$携带$x$的详细信息，并在注意力中发挥价值的作用。因此，建立在$f_\theta(\cdot)$之上的任意损失会自动调整$b$以反映语义局部性。我们遵循上述想法来建立 NSH 的骨干，但我们的贡献在于顶部的操作
## 4 Method
NSH 考虑了一个无监督哈希问题，该问题将$d_x$维数据$x\in\{x\}_{i=1}^N$映射到二进制向量$b\in\{-1,1\}^d_b$,其中$N$ 是整个数据集的大小，$d_b$是代码长度。在本文中，我们默认情况下考虑一个图像哈希问题，以与传统的实验设置相匹配，但 NSH 适用于任意数据模式，只要随机扩增适用
## 4.1 Model Structure
如 Sec 中所述。$\operatorname{E3}$ ,我们受到 [Shen et al., 2020] 的启发在训练期间使用双瓶颈哈希骨干网为每个$x$ 生成实际的 | [
[
"0.08826",
"0.0682",
"0.4844",
"0.06946",
"0.4844",
"0.0833",
"0.08826",
"0.0821"
],
[
"0.08826",
"0.084",
"0.429",
"0.084",
"0.429",
"0.09784",
"0.08826",
"0.09784"
],
[
"0.6978",
"0.0789",
"0.7793",
"0.0789",
"0.7793",
"0.09344",
"0.6978",
"0.09344"
],
[
"0.10455",
"0.09784",
"0.4836",
"0.09784",
"0.4836",
"0.1111",
"0.10455",
"0.1111"
],
[
"0.693",
"0.0966",
"0.722",
"0.09155",
"0.728",
"0.10986",
"0.6978",
"0.11554"
],
[
"0.8286",
"0.1016",
"0.835",
"0.1016",
"0.835",
"0.108",
"0.8286",
"0.108"
],
[
"0.8486",
"0.09784",
"0.888",
"0.09784",
"0.888",
"0.108",
"0.8486",
"0.108"
],
[
"0.08826",
"0.1111",
"0.4844",
"0.1111",
"0.4844",
"0.125",
"0.08826",
"0.125"
],
[
"0.8667",
"0.1061",
"0.8857",
"0.1061",
"0.8857",
"0.11426",
"0.8667",
"0.11426"
],
[
"0.5557",
"0.11237",
"0.6094",
"0.11426",
"0.609",
"0.1288",
"0.5537",
"0.127"
],
[
"0.6626",
"0.113",
"0.751",
"0.113",
"0.751",
"0.1276",
"0.6626",
"0.1276"
],
[
"0.8447",
"0.11426",
"0.8906",
"0.11426",
"0.8906",
"0.1256",
"0.8447",
"0.1256"
],
[
"0.08905",
"0.1256",
"0.486",
"0.1256",
"0.486",
"0.1395",
"0.08905",
"0.1395"
],
[
"0.562",
"0.13",
"0.5884",
"0.13",
"0.5884",
"0.1433",
"0.562",
"0.1433"
],
[
"0.846",
"0.132",
"0.8896",
"0.132",
"0.8896",
"0.1433",
"0.846",
"0.1433"
],
[
"0.08496",
"0.1383",
"0.4844",
"0.137",
"0.4844",
"0.1515",
"0.08496",
"0.1528"
],
[
"0.5557",
"0.1445",
"0.6113",
"0.1465",
"0.6094",
"0.161",
"0.5537",
"0.159"
],
[
"0.8486",
"0.1445",
"0.8613",
"0.1445",
"0.8613",
"0.1509",
"0.8486",
"0.1509"
],
[
"0.863",
"0.1445",
"0.875",
"0.1445",
"0.875",
"0.1483",
"0.863",
"0.1483"
],
[
"0.08826",
"0.1528",
"0.4844",
"0.1528",
"0.4844",
"0.1666",
"0.08826",
"0.1666"
],
[
"0.83",
"0.1459",
"0.8364",
"0.1459",
"0.8364",
"0.1515",
"0.83",
"0.1515"
],
[
"0.55",
"0.1572",
"0.615",
"0.1604",
"0.614",
"0.1755",
"0.5483",
"0.1724"
],
[
"0.8613",
"0.1578",
"0.8896",
"0.1604",
"0.888",
"0.1674",
"0.8594",
"0.1654"
],
[
"0.8677",
"0.1528",
"0.875",
"0.1528",
"0.875",
"0.1559",
"0.8677",
"0.1559"
],
[
"0.8486",
"0.1604",
"0.853",
"0.1604",
"0.853",
"0.1666",
"0.8486",
"0.1666"
],
[
"0.08826",
"0.1666",
"0.4836",
"0.1666",
"0.4836",
"0.1799",
"0.08826",
"0.1799"
],
[
"0.08496",
"0.1787",
"0.4844",
"0.1799",
"0.4844",
"0.1945",
"0.08496",
"0.1932"
],
[
"0.603",
"0.1787",
"0.6094",
"0.1787",
"0.6094",
"0.1849",
"0.603",
"0.1849"
],
[
"0.847",
"0.1755",
"0.8774",
"0.1755",
"0.8774",
"0.1843",
"0.847",
"0.1843"
],
[
"0.826",
"0.1875",
"0.8384",
"0.1875",
"0.8384",
"0.1963",
"0.826",
"0.1963"
],
[
"0.08826",
"0.1945",
"0.4844",
"0.1945",
"0.4844",
"0.2084",
"0.08826",
"0.2084"
],
[
"0.607",
"0.1963",
"0.614",
"0.1963",
"0.614",
"0.2039",
"0.607",
"0.2039"
],
[
"0.699",
"0.1963",
"0.7134",
"0.1963",
"0.7134",
"0.2052",
"0.699",
"0.2052"
],
[
"0.846",
"0.2008",
"0.88",
"0.2008",
"0.88",
"0.2122",
"0.846",
"0.2122"
],
[
"0.08905",
"0.2084",
"0.4836",
"0.2084",
"0.4836",
"0.2216",
"0.08905",
"0.2216"
],
[
"0.546",
"0.1976",
"0.5537",
"0.1995",
"0.5522",
"0.2058",
"0.5435",
"0.2026"
],
[
"0.564",
"0.209",
"0.601",
"0.209",
"0.601",
"0.2247",
"0.564",
"0.2247"
],
[
"0.7197",
"0.209",
"0.759",
"0.209",
"0.759",
"0.2247",
"0.7197",
"0.2247"
],
[
"0.08826",
"0.2203",
"0.4844",
"0.2216",
"0.4844",
"0.2361",
"0.08826",
"0.2349"
],
[
"0.572",
"0.226",
"0.5864",
"0.2229",
"0.591",
"0.2361",
"0.576",
"0.2393"
],
[
"0.7363",
"0.2279",
"0.7427",
"0.2279",
"0.7427",
"0.2329",
"0.7363",
"0.2329"
],
[
"0.08826",
"0.2361",
"0.4836",
"0.2361",
"0.4836",
"0.25",
"0.08826",
"0.25"
],
[
"0.8267",
"0.2311",
"0.8384",
"0.2311",
"0.8384",
"0.2405",
"0.8267",
"0.2405"
],
[
"0.5645",
"0.2374",
"0.6055",
"0.2374",
"0.6055",
"0.2568",
"0.5645",
"0.2568"
],
[
"0.7183",
"0.2393",
"0.755",
"0.2393",
"0.755",
"0.2556",
"0.7183",
"0.2556"
],
[
"0.08826",
"0.25",
"0.4836",
"0.25",
"0.4836",
"0.2632",
"0.08826",
"0.2632"
],
[
"0.08826",
"0.2632",
"0.4844",
"0.2632",
"0.4844",
"0.2766",
"0.08826",
"0.2766"
],
[
"0.6562",
"0.26",
"0.6665",
"0.26",
"0.6665",
"0.2695",
"0.6562",
"0.2695"
],
[
"0.08905",
"0.2778",
"0.4844",
"0.2778",
"0.4844",
"0.2917",
"0.08905",
"0.2917"
],
[
"0.5146",
"0.2827",
"0.911",
"0.2827",
"0.911",
"0.2935",
"0.5146",
"0.2935"
],
[
"0.08905",
"0.2917",
"0.4844",
"0.2917",
"0.4844",
"0.305",
"0.08905",
"0.305"
],
[
"0.513",
"0.2942",
"0.911",
"0.2942",
"0.911",
"0.308",
"0.513",
"0.308"
],
[
"0.08826",
"0.3062",
"0.4844",
"0.3062",
"0.4844",
"0.3193",
"0.08826",
"0.3193"
],
[
"0.513",
"0.308",
"0.7344",
"0.308",
"0.7344",
"0.318",
"0.513",
"0.318"
],
[
"0.7427",
"0.305",
"0.91",
"0.305",
"0.91",
"0.3193",
"0.7427",
"0.3193"
],
[
"0.738",
"0.3113",
"0.744",
"0.3113",
"0.744",
"0.315",
"0.738",
"0.315"
],
[
"0.08826",
"0.3213",
"0.1602",
"0.3213",
"0.1602",
"0.332",
"0.08826",
"0.332"
],
[
"0.511",
"0.3176",
"0.9126",
"0.315",
"0.9126",
"0.332",
"0.511",
"0.334"
],
[
"0.513",
"0.3289",
"0.911",
"0.3289",
"0.911",
"0.3435",
"0.513",
"0.3435"
],
[
"0.0989",
"0.3372",
"0.4844",
"0.336",
"0.4844",
"0.351",
"0.0989",
"0.3516"
],
[
"0.509",
"0.3416",
"0.914",
"0.3403",
"0.914",
"0.361",
"0.509",
"0.3623"
],
[
"0.1177",
"0.3516",
"0.486",
"0.3516",
"0.486",
"0.3655",
"0.1177",
"0.3655"
],
[
"0.1177",
"0.3643",
"0.4844",
"0.3643",
"0.4844",
"0.3774",
"0.1177",
"0.3774"
],
[
"0.513",
"0.3591",
"0.9126",
"0.3591",
"0.9126",
"0.3726",
"0.513",
"0.3726"
],
[
"0.5107",
"0.3718",
"0.857",
"0.3706",
"0.857",
"0.3838",
"0.5107",
"0.385"
],
[
"0.1177",
"0.3801",
"0.4084",
"0.3801",
"0.4084",
"0.3933",
"0.1177",
"0.3933"
],
[
"0.0989",
"0.3965",
"0.486",
"0.3984",
"0.486",
"0.4124",
"0.0989",
"0.4104"
],
[
"0.5146",
"0.4016",
"0.91",
"0.4016",
"0.91",
"0.4153",
"0.5146",
"0.4153"
],
[
"0.11926",
"0.4136",
"0.4836",
"0.4136",
"0.4836",
"0.4236",
"0.11926",
"0.4236"
],
[
"0.513",
"0.4153",
"0.91",
"0.4153",
"0.91",
"0.43",
"0.513",
"0.43"
],
[
"0.1168",
"0.4248",
"0.2917",
"0.4268",
"0.2917",
"0.4414",
"0.1168",
"0.4395"
],
[
"0.513",
"0.4287",
"0.91",
"0.4287",
"0.91",
"0.443",
"0.513",
"0.443"
],
[
"0.10046",
"0.443",
"0.486",
"0.4443",
"0.486",
"0.4583",
"0.10046",
"0.457"
],
[
"0.5146",
"0.443",
"0.669",
"0.443",
"0.669",
"0.457",
"0.5146",
"0.457"
],
[
"0.1177",
"0.459",
"0.486",
"0.459",
"0.486",
"0.473",
"0.1177",
"0.473"
],
[
"0.7183",
"0.4653",
"0.863",
"0.4653",
"0.863",
"0.4792",
"0.7183",
"0.4792"
],
[
"0.558",
"0.473",
"0.7236",
"0.473",
"0.7236",
"0.486",
"0.558",
"0.486"
],
[
"0.8896",
"0.4734",
"0.9126",
"0.4734",
"0.9126",
"0.486",
"0.8896",
"0.486"
],
[
"0.778",
"0.4841",
"0.7983",
"0.4841",
"0.7983",
"0.4976",
"0.778",
"0.4976"
],
[
"0.11926",
"0.4749",
"0.4673",
"0.4749",
"0.4673",
"0.4849",
"0.11926",
"0.4849"
],
[
"0.08496",
"0.5024",
"0.1217",
"0.5024",
"0.1217",
"0.5146",
"0.08496",
"0.5146"
],
[
"0.1177",
"0.502",
"0.2402",
"0.502",
"0.2402",
"0.515",
"0.1177",
"0.515"
],
[
"0.5156",
"0.504",
"0.911",
"0.504",
"0.911",
"0.5176",
"0.5156",
"0.5176"
],
[
"0.09314",
"0.5215",
"0.4844",
"0.5215",
"0.4844",
"0.536",
"0.09314",
"0.536"
],
[
"0.5146",
"0.5176",
"0.716",
"0.5176",
"0.716",
"0.531",
"0.5146",
"0.531"
],
[
"0.08905",
"0.535",
"0.486",
"0.535",
"0.486",
"0.55",
"0.08905",
"0.55"
],
[
"0.5195",
"0.536",
"0.7817",
"0.536",
"0.7817",
"0.55",
"0.5195",
"0.55"
],
[
"0.08826",
"0.55",
"0.486",
"0.55",
"0.486",
"0.563",
"0.08826",
"0.563"
],
[
"0.5195",
"0.5537",
"0.91",
"0.5537",
"0.91",
"0.5674",
"0.5195",
"0.5674"
],
[
"0.08826",
"0.5625",
"0.4844",
"0.5625",
"0.4844",
"0.5757",
"0.08826",
"0.5757"
],
[
"0.5146",
"0.5684",
"0.911",
"0.5684",
"0.911",
"0.5835",
"0.5146",
"0.5835"
],
[
"0.08826",
"0.577",
"0.4844",
"0.577",
"0.4844",
"0.5913",
"0.08826",
"0.5913"
],
[
"0.5146",
"0.582",
"0.9126",
"0.582",
"0.9126",
"0.595",
"0.5146",
"0.595"
],
[
"0.08826",
"0.5913",
"0.4844",
"0.5913",
"0.4844",
"0.605",
"0.08826",
"0.605"
],
[
"0.5146",
"0.5986",
"0.597",
"0.5986",
"0.597",
"0.61",
"0.5146",
"0.61"
],
[
"0.0866",
"0.605",
"0.4844",
"0.605",
"0.4844",
"0.6187",
"0.0866",
"0.6187"
],
[
"0.0866",
"0.617",
"0.4844",
"0.6187",
"0.4844",
"0.6333",
"0.0866",
"0.632"
],
[
"0.615",
"0.619",
"0.812",
"0.619",
"0.812",
"0.6333",
"0.615",
"0.6333"
],
[
"0.8896",
"0.6206",
"0.9126",
"0.6206",
"0.9126",
"0.634",
"0.8896",
"0.634"
],
[
"0.08826",
"0.6333",
"0.4844",
"0.6333",
"0.4844",
"0.6465",
"0.08826",
"0.6465"
],
[
"0.08826",
"0.6465",
"0.3816",
"0.6465",
"0.3816",
"0.66",
"0.08826",
"0.66"
],
[
"0.517",
"0.6426",
"0.9126",
"0.6426",
"0.9126",
"0.656",
"0.517",
"0.656"
],
[
"0.5156",
"0.657",
"0.91",
"0.657",
"0.91",
"0.6675",
"0.5156",
"0.6675"
],
[
"0.09314",
"0.6655",
"0.486",
"0.6655",
"0.486",
"0.6787",
"0.09314",
"0.6787"
],
[
"0.5156",
"0.6704",
"0.911",
"0.6704",
"0.911",
"0.6846",
"0.5156",
"0.6846"
],
[
"0.0866",
"0.6787",
"0.4844",
"0.6787",
"0.4844",
"0.6924",
"0.0866",
"0.6924"
],
[
"0.5146",
"0.6846",
"0.911",
"0.6846",
"0.911",
"0.6978",
"0.5146",
"0.6978"
],
[
"0.08826",
"0.6924",
"0.4844",
"0.6924",
"0.4844",
"0.7056",
"0.08826",
"0.7056"
],
[
"0.511",
"0.696",
"0.911",
"0.6963",
"0.911",
"0.711",
"0.511",
"0.7104"
],
[
"0.08826",
"0.707",
"0.4836",
"0.707",
"0.4836",
"0.72",
"0.08826",
"0.72"
],
[
"0.513",
"0.711",
"0.7754",
"0.711",
"0.7754",
"0.7246",
"0.513",
"0.7246"
],
[
"0.08826",
"0.72",
"0.4844",
"0.72",
"0.4844",
"0.7344",
"0.08826",
"0.7344"
],
[
"0.10455",
"0.7363",
"0.4836",
"0.7363",
"0.4836",
"0.747",
"0.10455",
"0.747"
],
[
"0.511",
"0.735",
"0.6167",
"0.735",
"0.6167",
"0.7476",
"0.511",
"0.7476"
],
[
"0.08826",
"0.7485",
"0.486",
"0.7485",
"0.486",
"0.762",
"0.08826",
"0.762"
],
[
"0.511",
"0.7524",
"0.911",
"0.754",
"0.911",
"0.768",
"0.511",
"0.7676"
],
[
"0.08826",
"0.762",
"0.4844",
"0.762",
"0.4844",
"0.776",
"0.08826",
"0.776"
],
[
"0.513",
"0.768",
"0.911",
"0.768",
"0.911",
"0.782",
"0.513",
"0.782"
],
[
"0.08826",
"0.7744",
"0.4844",
"0.7744",
"0.4844",
"0.7886",
"0.08826",
"0.7886"
],
[
"0.513",
"0.781",
"0.911",
"0.782",
"0.911",
"0.797",
"0.513",
"0.7954"
],
[
"0.08826",
"0.7905",
"0.486",
"0.7905",
"0.486",
"0.8037",
"0.08826",
"0.8037"
],
[
"0.5146",
"0.797",
"0.911",
"0.797",
"0.911",
"0.81",
"0.5146",
"0.81"
],
[
"0.08826",
"0.8037",
"0.4844",
"0.8037",
"0.4844",
"0.8174",
"0.08826",
"0.8174"
],
[
"0.5146",
"0.81",
"0.911",
"0.81",
"0.911",
"0.8237",
"0.5146",
"0.8237"
],
[
"0.08905",
"0.8193",
"0.4192",
"0.8193",
"0.4192",
"0.83",
"0.08905",
"0.83"
],
[
"0.5146",
"0.8257",
"0.5474",
"0.8257",
"0.5474",
"0.837",
"0.5146",
"0.837"
],
[
"0.08496",
"0.8423",
"0.2345",
"0.8423",
"0.2345",
"0.856",
"0.08496",
"0.856"
],
[
"0.511",
"0.8413",
"0.6895",
"0.8423",
"0.6895",
"0.857",
"0.511",
"0.856"
],
[
"0.0907",
"0.86",
"0.486",
"0.8613",
"0.486",
"0.8755",
"0.0907",
"0.875"
],
[
"0.513",
"0.86",
"0.91",
"0.8613",
"0.91",
"0.8755",
"0.513",
"0.875"
],
[
"0.08826",
"0.8755",
"0.486",
"0.8755",
"0.486",
"0.8896",
"0.08826",
"0.8896"
],
[
"0.513",
"0.8755",
"0.911",
"0.8755",
"0.911",
"0.8896",
"0.513",
"0.8896"
]
] | [
"中构造成对的伪标签，但是对于深度模型，端到端训练",
"因此变得不可行，并且伪标签的错误可能会传播",
"SortedNCE",
"在本文中，我们将标题作为主要动机，并通过提出自",
"ET",
"F",
"",
"然排序哈希（NSH）来解决上述挑战，该哈希可以在训练",
"国",
"Softsort",
"Soft Gather",
"",
"过程中最大程度地呈现排序特性，从而提高检索性能。根",
"s.↑",
"",
"据第一个挑战，我们采用可微分softsort近似值国Prillo",
"Pairwise",
"",
"",
"andEisenschlos，2020]的最新进展作为一个构建块，将",
"S",
"Hamming",
"",
"",
"",
"底部的哈希编码器和顶部的基于排序的损失连接起来",
"特别是，NSH遵循[Shenetal.,2020]，它编码两个表示",
"+",
"",
"B:",
"形式i.e.，用于相似性比较的二进制哈希码和携带详细信",
"",
"B",
"",
"息的连续潜在向量。样本之间的代码距离为softsort算",
"S",
"fe()",
"fe()",
"子提供信息，并对潜在向量进行重新排序以进行自监督",
"",
"不",
"训练。为了应对第二个挑战，我们提出了一种新的排序",
"Z:",
"Aug",
"Aug",
"噪声对比估计（SortedNCE）损失，它有选择地选择最",
"相关的样本作为对比学习的阳性样本。因此，NSH在训",
"x",
"练过程中挖掘数据语义关系，并动态优化每个数据的排",
"Figure 2:The training schematic of NSH.We illustrate the",
"序检索候选列表。Fig.1（b）简要介绍了我们的主要思",
"batch-based output shapes on the right where different col-",
"想，它更好地代表了深度哈希的测试场景。我们的主要",
"ors refer to different instances.",
"Z e Rnxdz are the la-",
"",
"贡献包括：",
"tent representations. B e (-1,1jnxdb are the hash codes.",
"S e [o. 1]nxn is the code-based similarity matrix for sort-",
". For the first time, we describe an end-to-end NSH",
"ing. E e Rnxnxda are n soft-sorted lists of representations,",
"model that optimizes the sorted retrieval list that",
"best represents the testing scenario of learning to",
"of which each is ordered by the Hamming distance to the",
"corresponding instance in the batch for SortedNCE.",
"hash and only requires gradient descent.",
". To implement this vision, we additionally propose",
"似值。尽管可以将其视为线性规划问题[Blondel etal.",
"the SortedNCE lossthat trainsthe sortedfeatures",
"2020]，但我们选择采用更简单的softmax方法[Prillo",
"in an unsupervised way.",
"andEisenschlos，2020]，该方法将相似性得分sERM",
". We show the superiority of NSH in retrieval perfor-",
"的向量的置换确定为：",
"mance on three benchmarked datasets against the",
"d,(sort(s)1T, 1sT)",
"softsort(s) = softmax -",
"(1)",
"Ts",
"most recentunsuperviseddeephashingmethods",
"2",
"Related Work",
"，其中d（）几乎在任何地方都是任意可微分的超模，半",
"UnsupervisedHashing无监督哈希方法的早期工作",
"度量距离函数XMATHMA。",
"主要集中在学习紧凑表示[Gongetal.，2013]。最近",
"Twin-BottleneckHash Encoder到",
"的几项深度学习工作都集中在哈希代码质量上[Liand",
"Twin-BottleneckHashEncoder尽管在深度学习中广",
"van Gemert, 2021:Su et al., 2018: Luo et al., 2021b",
"泛见证了分解输出，但[Shen etal.,2020]指定了由θ参",
"GhasediDizaji etal，2018]。其他一些人则关注代码",
"数化的哈希编码器fe(）的两个输出的不同功能，该编码",
"[Shen et al., 2018:Yang et al., 2019; Shen et al., 2020]",
"器对数据a",
"的语义意识，而大多数人采用伪标签方案来挖掘数据相",
"似性，将其作为神经网络之外的单个模块。他们的表现",
"[h, z] = fe(α), b = sign(h).",
"(2)",
"通常在排序的候选人上进行评估。然而，他们并不寻求",
"在训练期间实现排序的概念来挖掘相似性",
"<N。由于哈希码通常较短且信息量较小，因此h和b",
"仅用于计算在注意力机制中充当查询/键的成对数据相",
"HashingwithContrastiveLearning对比学习是一种",
"似度，而z携带的详细信息，并在注意力中发挥价值",
"通过构造正负对来学习没有标签的数据集的一般特征的",
"的作用。因此，建立在fe）之上的任意损失会自动调整",
"方法[Hadselletal.,2006]。实例判别「Wuetal.,2018",
"b以反映语义局部性。我们遵循上述想法来建立NSH的",
"提出非参数交叉摘损失在实例级优化模型。最近的作品",
"骨干，但我们的贡献在于顶部的操作",
"也使用对比学习哈希[Qiuetal.,2021Luoetal.,2021b]",
"他们采用实例歧视作为目标，其中正面和负面实例可",
"4Method",
"能仍然具有重叠的语义。也就是说，批处理中的每个查",
"NSH考虑了一个无监督哈希问题，该问题将d维数据",
"询图像仅将其增强视图视为正样本，这意味着即使对于",
"E{a,映射到二进制向量bE（-1,1)d，其中N",
"极其相似的样本，也必须将它们分开。此外，CIBHash",
"是整个数据集的大小，ds是代码长度。在本文中，我们",
"[Qiuetal.,2021]和CIMON[Luoetal.,2021b]会调整",
"默认情况下考虑一个图像哈希问题，以与传统的实验设",
"对比损失以适应哈希学习标准，但是直接在对比损失中",
"置相匹配，但NSH适用于任意数据模式，只要随机扩增",
"使用哈希代码无法很好地集成图像的语义内容。",
"适用",
"3Preliminaries",
"4.1ModelStructure",
"NeuralSortingOperators香草argsort绝对是不可微",
"如Sec中所述。E3，我们受到[Shenetal.，2020]的启发",
"分的，但是最近的研究发现了几种与神经网络兼容的近",
"在训练期间使用双瓶颈哈希骨干网为每个&生成实际的"
] | {
"arxiv_id": "2201.13322",
"language": "zh",
"page_idx": 2
} |
||
2201.13322-zh-3 | 2201.13322-zh-3.jpg | 哈希代码$b$和另一组延迟$z$ ,其中损失施加在涉及$z$和$b$之间的汉明距离的转换的顶部。NSH 还采用了对比学习的最新进展$\overline{\mathrm{E}}[$Chen et al., 2020a] 作为无监督框架，该框架需要两组独立的随机数据增强。在下文中，我们使用$\tilde{\cdot}$和$\hat{\cdot}$,e.g.$,\tilde{x}$ 和$\hat{x},\tilde{\boldsymbol{b}}$ 和$\hat{\boldsymbol{b}},\tilde{\boldsymbol{z}}$ 和$\hat{\boldsymbol{z}}$ 等表示两个增强对应物的输出。此外，由于我们的训练过程涉及多个排序实例，我们将在必要时使用带有大写字母的行排序批处理符号，e.g., $\tilde{\boldsymbol{X}} = [ \tilde{x} _1; \cdots ; \tilde{x} _n]$ $, \hat{\boldsymbol{Z}} = [ \hat{\boldsymbol{z}} _1; \cdots ; \hat{\boldsymbol{z}} _n]$ 等，$n\ddot{\text{ 是批量大小以启用基于批处理的训练}}$
Training Objective as a Function of Sorting 我们的最终目标是构建一个完全可微的模型，该模型在一堆语义排序的候选对象上进行训练：
$$\mathcal{L}_{\mathrm{Sorted}}:=\underbrace{\mathrm{NCE}}_{\mathrm{Sec.~4.3}}\circ\underbrace{\mathrm{Sort\&Gather}\circ\mathrm{Hamming}}_{\mathrm{Sec.~4.2}}\circ f_\theta(\boldsymbol{x}),$$
其中每个堆叠组件都允许梯度传播回$f_\theta(\cdot)$ 。Fig.2 给出了 NSH 结构的一瞥。两组不同增强下的实例都呈现给骨干编码器$f_\theta(\cdot)$,产生两组哈希码$(\tilde{B}$和$\hat{B}$ )和延迟($\tilde{Z} 和 \hat{Z})。然后，哈希代码确定排序和排列延迟的$ 方式 (Sec。4.2 )。然后，置换后的潜在张量有助于提出的用于训练的 SortedNCE 损失 (Sec。4.3 ),而训练过程在 Sec 中给出。4.4
4.2 Sorting and Gathering the Latents
与[Shen et al., 2020] 类似，我们在训练模型中植人了成对哈希代码距离的计算，以便可以通过顶部的任何损失来优化代码相似性。形式上，我们将基于代码的亲和矩阵$S$定义为：
$$S=\tilde{B}\hat{B}^\intercal/2d_b+0.5,$$
(3)
,以便其每个条目表示归一化的成对相似性 i.e.$,\boldsymbol{S}[i,j]=$ 1-Hamming$(\tilde{\boldsymbol{b}}_i,\hat{\boldsymbol{b}}_j)/d_b$ 。Eq. (3)仅涉及在任何地方都可微分的线性运算，因此它可以用作神经网络中反向传播的构建块。请注意，我们不会对$S$施加任何损失项， 以防止模型陷入伪标记训练方案，如 Sec 中所述。1
Sort out the Relevance with Hash Codes 作为 NSH 的关键操作，我们将批处理中的每个样本视为检索查询， 同时让整个批处理作为候选。对于每个数据，可以选择最相关的数据，并将其放在检索序列的开头，依此类推。为了在模型中反映此过程，我们调用 Eq.(1)来计算每个$x_i$,i.e.,
$$\boldsymbol{p}_{i}=\mathrm{softsort}(-\boldsymbol{s}_{i})\in(0,1)^{n\times n},$$
(4)
的排序置换矩阵$p_i$ ,其中$s_i\in[0,1]^n$是$i$的第$i$行与$\boldsymbol{S}_{\text{ (XMATHX}_{\text{。也就是说}},\boldsymbol{p}_{i}[j,k]\text{ 在}[j,k]\text{ 位置的条}}$ 目代表$x_k$成为与$x_i$ $\mathbb{E}[$Prillo and Eisenschlos, 2020] 最相关的第$j$个项目的可能性
Re-order Features by Soft Gathering/Permutation 对于每个$x_i$ , NSH 在批次$e_i$中生成候选嵌人的矩阵，其
中最相似项的特征位于顶部。现在我们已经有了来自Eq. ( 4) 的软置换矩阵，$e_i$ 可以通过以下软收集过程定义：
$$e_i=p_i\tilde{Z}\in\mathbb{R}^{n\times d_z}.$$
这里$d_z$是连续延迟$\tilde{z}$的大小。为了澄清，$e_i$只是$x_i$
(5)
$\begin{array}{l}\text{的排序嵌入矩阵，而一批数据使其成为三维张量 i.e.,}\\\boldsymbol{E}=[\boldsymbol{e}_1;\cdots;\boldsymbol{e}_n]\in\mathbb{R}^{n\times n\times d_z}。\mathrm{Eq.~(5)~的基于批处理的}\end{array}$ 实现可以通过最近的深度学习工具箱中的爱因斯坦求和约定轻松实现(参见 Alg。1)
Remark 1: The Order of $\boldsymbol{E}$ Matters. $\boldsymbol{E}$从字面上看是$n$延迟$\tilde{\boldsymbol{Z}}$的排序列表，其中每个列表$e_i$包含$x_i$的批量检索结果的表示形式，以相似度的降序排列。值得注意的是，由于$\tilde{b}_i$ 和$\hat{b}_i$ 预计是相同的$,e_i$, i.e., $e_i[1,:]$ 的第一行自然会接近$\tilde{z}_i$,表示$x_i$本身是批处理中最相关的一个。从第二行开始，相关性降低。Eqs.(3)to(5) 定义的步骤自动选择语义相关的实例，并将它们放在「$\vec{E}$ 的开头。因此，可以很容易地确定每个$x_i$的正/负样本在一个完全可微的方式。此过程模仿检索的真实测试场景，并启用列表式训练以获得更好的性能
Remark 2: Why Do We Gather $Z?$ 我们有可能构建一个单瓶颈模型，并用 Eq.(5)中的$\tilde{B}$替换$\tilde{Z}$ 。然而， 哈希代码通常很短，信息量较少。尽管他们能够编码粗粒度的语义相似性，但他们很难将每个数据的完整身份与其他数据进行比较。我们通过实验表明，当使用严重依赖身份保留的对比学习目标进行训练时，这会导致次优的性能。我们遵循[Shen et al., 2020] 在这里使用另一个连续表示来支持学习目标。或者，我们的哈希代码可以被视为置换$Z$的注意头，建立在顶部的损失会自动提升代码的语义意识
## 4.3 SortedNCE
许多现有的无监督哈希模型都难以确定数据点之间的相似性，它们通常需要一个延迟的伪标记步骤来引入额外的噪声 [Su et al., 2018; Luo et al., 2021b]。然而，NSH 已经有一堆分类功能$E$现成的，这使得它非常容易确定阳性样本
另一方面，会有多个样本共享相同的语义信息，但是香草单标签 InfoNCE$\underline{\mathrm{E}}[$Oord et al.,2018] 损失没有考虑语义的这种重叠。我们提出了一个学习目标，该目标适用于多个正对。特别是，由于$E$是有序的，我们根据日志的位置定义 SortedNCE 损失，并让第一个$m$样本为$\bar{\text{正，i.e.}}$,
$$\begin{aligned}\mathcal{L}_{\mathrm{Sorted}}&=\frac{-1}{mn}\sum_{i=1}^{n}\sum_{j=1}^{m}\\&\underbrace{\log\frac{\kappa(\boldsymbol{e}_i[\:j\:,\::],\hat{\boldsymbol{z}}_i)}{\kappa(\boldsymbol{e}_i[\:j\:,\::],\hat{\boldsymbol{z}}_i)+\sum_{k=m+1}^n\kappa(\boldsymbol{e}_i[\:k\:,\::],\hat{\boldsymbol{z}}_i)}}\end{aligned}$$
Positive and negative samples only depend on the sorting positions
(6)
其中$\kappa(\boldsymbol{a},\boldsymbol{b})=\exp(\cos(\boldsymbol{a},\boldsymbol{b})/\tau_{\mathbf{c}})$ 。将$m$和$\tau_\mathrm{c}$视为超参数。直观上，Eq. (6)为$e_i$中的$m$阳性候选者构造了$m$对比，其中每个候选者都形成一个单标签交叉熵项。重要的是，由于正样本的特征在$e_i$中具有$1\cdots m$的固定位置，因此 NSH 不需要额外的 argsort 运算符来查找涉及不可微计算的最相关项。我们在 Sec 中讨论了这种设计的好处。4.5 。通过将第一个$m$对数视为正，单个$m$标签交叉熵项也可以在这里工作。此解决方案类似于SupCon [Khosla et al., 2020] ,但我们在 Sec 中通过实验显示。5.3 在我们的案例中表现不如 SortedNCE
还要注意，Eq. (6) 不是通用的损失，它专用于包含Eqs. (3) to (5)描述的排序排序和收集层的检索模型，例如 NSH。对于每个样本，SortedNCE 的时间复杂度为$O(mn)$ 。考虑到$m$通常很小的事实，这种损失不会显着增加训练时间 | [
[
"0.08905",
"0.07135",
"0.486",
"0.07135",
"0.486",
"0.0821",
"0.08905",
"0.0821"
],
[
"0.5195",
"0.07135",
"0.7915",
"0.07135",
"0.7915",
"0.0821",
"0.5195",
"0.0821"
],
[
"0.7974",
"0.0745",
"0.8037",
"0.0745",
"0.8037",
"0.0789",
"0.7974",
"0.0789"
],
[
"0.806",
"0.07135",
"0.91",
"0.07135",
"0.91",
"0.0821",
"0.806",
"0.0821"
],
[
"0.08826",
"0.0833",
"0.4844",
"0.0833",
"0.4844",
"0.09784",
"0.08826",
"0.09784"
],
[
"0.513",
"0.0821",
"0.911",
"0.0833",
"0.911",
"0.09784",
"0.513",
"0.0966"
],
[
"0.0907",
"0.0966",
"0.4878",
"0.0966",
"0.4878",
"0.1111",
"0.0907",
"0.1111"
],
[
"0.513",
"0.09784",
"0.91",
"0.09784",
"0.91",
"0.1111",
"0.513",
"0.1111"
],
[
"0.08826",
"0.1111",
"0.486",
"0.1111",
"0.486",
"0.125",
"0.08826",
"0.125"
],
[
"0.513",
"0.113",
"0.91",
"0.113",
"0.91",
"0.127",
"0.513",
"0.127"
],
[
"0.0866",
"0.127",
"0.486",
"0.1256",
"0.486",
"0.1401",
"0.0866",
"0.1414"
],
[
"0.513",
"0.127",
"0.91",
"0.1256",
"0.91",
"0.1401",
"0.513",
"0.1414"
],
[
"0.08826",
"0.1414",
"0.4878",
"0.1414",
"0.4878",
"0.1547",
"0.08826",
"0.1547"
],
[
"0.5146",
"0.1414",
"0.91",
"0.1414",
"0.91",
"0.1547",
"0.5146",
"0.1547"
],
[
"0.08826",
"0.154",
"0.4844",
"0.154",
"0.4844",
"0.1674",
"0.08826",
"0.1674"
],
[
"0.5146",
"0.1547",
"0.9126",
"0.1547",
"0.9126",
"0.1686",
"0.5146",
"0.1686"
],
[
"0.08905",
"0.1686",
"0.4844",
"0.1686",
"0.4844",
"0.1831",
"0.08905",
"0.1831"
],
[
"0.513",
"0.1674",
"0.91",
"0.1674",
"0.91",
"0.1812",
"0.513",
"0.1812"
],
[
"0.0907",
"0.1849",
"0.3945",
"0.1849",
"0.3945",
"0.1957",
"0.0907",
"0.1957"
],
[
"0.5156",
"0.1818",
"0.911",
"0.1818",
"0.911",
"0.1957",
"0.5156",
"0.1957"
],
[
"0.5146",
"0.1963",
"0.808",
"0.1963",
"0.808",
"0.2102",
"0.5146",
"0.2102"
],
[
"0.09235",
"0.2026",
"0.486",
"0.2026",
"0.486",
"0.2166",
"0.09235",
"0.2166"
],
[
"0.08905",
"0.2166",
"0.4844",
"0.2166",
"0.4844",
"0.2299",
"0.08905",
"0.2299"
],
[
"0.5186",
"0.2172",
"0.911",
"0.2172",
"0.911",
"0.2311",
"0.5186",
"0.2311"
],
[
"0.08826",
"0.2299",
"0.2917",
"0.2299",
"0.2917",
"0.2437",
"0.08826",
"0.2437"
],
[
"0.5254",
"0.2317",
"0.9087",
"0.2329",
"0.9087",
"0.2474",
"0.5254",
"0.2468"
],
[
"0.0989",
"0.25",
"0.4722",
"0.2488",
"0.4722",
"0.2664",
"0.0989",
"0.2676"
],
[
"0.5146",
"0.2474",
"0.911",
"0.2474",
"0.911",
"0.2615",
"0.5146",
"0.2615"
],
[
"0.1765",
"0.2651",
"0.2222",
"0.2651",
"0.2222",
"0.276",
"0.1765",
"0.276"
],
[
"0.5146",
"0.262",
"0.911",
"0.262",
"0.911",
"0.276",
"0.5146",
"0.276"
],
[
"0.1732",
"0.2727",
"0.2197",
"0.2708",
"0.2197",
"0.281",
"0.1732",
"0.2827"
],
[
"0.304",
"0.2766",
"0.3464",
"0.2766",
"0.3464",
"0.2842",
"0.304",
"0.2842"
],
[
"0.517",
"0.276",
"0.911",
"0.276",
"0.911",
"0.289",
"0.517",
"0.289"
],
[
"0.1013",
"0.2917",
"0.486",
"0.2917",
"0.486",
"0.305",
"0.1013",
"0.305"
],
[
"0.5146",
"0.289",
"0.911",
"0.289",
"0.911",
"0.303",
"0.5146",
"0.303"
],
[
"0.08826",
"0.303",
"0.4836",
"0.303",
"0.4836",
"0.3164",
"0.08826",
"0.3164"
],
[
"0.5146",
"0.3037",
"0.898",
"0.3037",
"0.898",
"0.3176",
"0.5146",
"0.3176"
],
[
"0.08826",
"0.3193",
"0.4844",
"0.3193",
"0.4844",
"0.3333",
"0.08826",
"0.3333"
],
[
"0.5156",
"0.3164",
"0.91",
"0.3164",
"0.91",
"0.33",
"0.5156",
"0.33"
],
[
"0.513",
"0.3289",
"0.911",
"0.33",
"0.911",
"0.3447",
"0.513",
"0.3435"
],
[
"0.08826",
"0.336",
"0.4836",
"0.336",
"0.4836",
"0.3499",
"0.08826",
"0.3499"
],
[
"0.0907",
"0.3499",
"0.4836",
"0.3499",
"0.4836",
"0.363",
"0.0907",
"0.363"
],
[
"0.5156",
"0.3467",
"0.6445",
"0.3467",
"0.6445",
"0.3567",
"0.5156",
"0.3567"
],
[
"0.08826",
"0.363",
"0.4844",
"0.363",
"0.4844",
"0.3774",
"0.08826",
"0.3774"
],
[
"0.513",
"0.3674",
"0.652",
"0.3674",
"0.652",
"0.3806",
"0.513",
"0.3806"
],
[
"0.08826",
"0.3774",
"0.255",
"0.3774",
"0.255",
"0.3914",
"0.08826",
"0.3914"
],
[
"0.08826",
"0.3945",
"0.406",
"0.3945",
"0.406",
"0.4092",
"0.08826",
"0.4092"
],
[
"0.517",
"0.3901",
"0.91",
"0.3901",
"0.91",
"0.401",
"0.517",
"0.401"
],
[
"0.5156",
"0.4028",
"0.911",
"0.4028",
"0.911",
"0.4167",
"0.5156",
"0.4167"
],
[
"0.08905",
"0.4124",
"0.486",
"0.4124",
"0.486",
"0.4255",
"0.08905",
"0.4255"
],
[
"0.5146",
"0.4167",
"0.911",
"0.4167",
"0.911",
"0.43",
"0.5146",
"0.43"
],
[
"0.08826",
"0.4255",
"0.4844",
"0.4255",
"0.4844",
"0.4395",
"0.08826",
"0.4395"
],
[
"0.513",
"0.4287",
"0.911",
"0.43",
"0.911",
"0.4443",
"0.513",
"0.443"
],
[
"0.08826",
"0.44",
"0.4844",
"0.44",
"0.4844",
"0.4539",
"0.08826",
"0.4539"
],
[
"0.517",
"0.4458",
"0.5786",
"0.4458",
"0.5786",
"0.457",
"0.517",
"0.457"
],
[
"0.08826",
"0.4558",
"0.1814",
"0.4558",
"0.1814",
"0.4673",
"0.08826",
"0.4673"
],
[
"0.5312",
"0.459",
"0.911",
"0.459",
"0.911",
"0.473",
"0.5312",
"0.473"
],
[
"0.21",
"0.4734",
"0.3586",
"0.4734",
"0.3586",
"0.4873",
"0.21",
"0.4873"
],
[
"0.4631",
"0.476",
"0.486",
"0.476",
"0.486",
"0.4873",
"0.4631",
"0.4873"
],
[
"0.5146",
"0.4717",
"0.911",
"0.473",
"0.911",
"0.4873",
"0.5146",
"0.486"
],
[
"0.513",
"0.4849",
"0.911",
"0.486",
"0.911",
"0.5005",
"0.513",
"0.5"
],
[
"0.08905",
"0.4937",
"0.486",
"0.4937",
"0.486",
"0.507",
"0.08905",
"0.507"
],
[
"0.5156",
"0.5005",
"0.91",
"0.5005",
"0.91",
"0.5146",
"0.5156",
"0.5146"
],
[
"0.0866",
"0.5083",
"0.486",
"0.509",
"0.486",
"0.5234",
"0.0866",
"0.523"
],
[
"0.511",
"0.5146",
"0.91",
"0.5146",
"0.91",
"0.528",
"0.511",
"0.528"
],
[
"0.08905",
"0.5234",
"0.4844",
"0.5234",
"0.4844",
"0.537",
"0.08905",
"0.537"
],
[
"0.511",
"0.5264",
"0.5767",
"0.529",
"0.5767",
"0.5435",
"0.511",
"0.5415"
],
[
"0.08826",
"0.537",
"0.4836",
"0.537",
"0.4836",
"0.5503",
"0.08826",
"0.5503"
],
[
"0.08826",
"0.5503",
"0.4673",
"0.5503",
"0.4673",
"0.5645",
"0.08826",
"0.5645"
],
[
"0.5923",
"0.553",
"0.661",
"0.561",
"0.6567",
"0.58",
"0.5884",
"0.5713"
],
[
"0.0948",
"0.571",
"0.4836",
"0.571",
"0.4836",
"0.5806",
"0.0948",
"0.5806"
],
[
"0.5146",
"0.565",
"0.5767",
"0.5674",
"0.5767",
"0.579",
"0.5146",
"0.577"
],
[
"0.585",
"0.5737",
"0.6274",
"0.5737",
"0.6274",
"0.5854",
"0.585",
"0.5854"
],
[
"0.08905",
"0.5835",
"0.4836",
"0.5835",
"0.4836",
"0.5967",
"0.08905",
"0.5967"
],
[
"0.614",
"0.579",
"0.6626",
"0.5806",
"0.6626",
"0.5923",
"0.614",
"0.5903"
],
[
"0.08905",
"0.5967",
"0.4844",
"0.5967",
"0.4844",
"0.61",
"0.08905",
"0.61"
],
[
"0.721",
"0.601",
"0.8164",
"0.601",
"0.8164",
"0.6123",
"0.721",
"0.6123"
],
[
"0.08826",
"0.61",
"0.4836",
"0.61",
"0.4836",
"0.6235",
"0.08826",
"0.6235"
],
[
"0.601",
"0.6113",
"0.626",
"0.6113",
"0.626",
"0.6226",
"0.601",
"0.6226"
],
[
"0.08905",
"0.625",
"0.4844",
"0.625",
"0.4844",
"0.638",
"0.08905",
"0.638"
],
[
"0.626",
"0.6206",
"0.9126",
"0.6206",
"0.9126",
"0.638",
"0.626",
"0.638"
],
[
"0.0866",
"0.637",
"0.1772",
"0.6396",
"0.1772",
"0.654",
"0.0866",
"0.6514"
],
[
"0.5645",
"0.6445",
"0.947",
"0.6445",
"0.947",
"0.6577",
"0.5645",
"0.6577"
],
[
"0.1724",
"0.656",
"0.392",
"0.653",
"0.392",
"0.6704",
"0.1724",
"0.674"
],
[
"0.4631",
"0.6577",
"0.4844",
"0.6577",
"0.4844",
"0.6704",
"0.4631",
"0.6704"
],
[
"0.8906",
"0.659",
"0.911",
"0.659",
"0.911",
"0.672",
"0.8906",
"0.672"
],
[
"0.0866",
"0.675",
"0.4844",
"0.674",
"0.4844",
"0.688",
"0.0866",
"0.6895"
],
[
"0.08496",
"0.686",
"0.486",
"0.6875",
"0.486",
"0.705",
"0.08496",
"0.704"
],
[
"0.5156",
"0.683",
"0.9126",
"0.683",
"0.9126",
"0.6963",
"0.5156",
"0.6963"
],
[
"0.513",
"0.696",
"0.9087",
"0.6963",
"0.9087",
"0.711",
"0.513",
"0.7104"
],
[
"0.08826",
"0.702",
"0.4878",
"0.7007",
"0.4878",
"0.7153",
"0.08826",
"0.7163"
],
[
"0.08826",
"0.7163",
"0.286",
"0.7163",
"0.286",
"0.73",
"0.08826",
"0.73"
],
[
"0.5146",
"0.711",
"0.91",
"0.711",
"0.91",
"0.7246",
"0.5146",
"0.7246"
],
[
"0.517",
"0.7266",
"0.91",
"0.7266",
"0.91",
"0.7373",
"0.517",
"0.7373"
],
[
"0.09235",
"0.7344",
"0.4878",
"0.7344",
"0.4878",
"0.7476",
"0.09235",
"0.7476"
],
[
"0.513",
"0.7373",
"0.911",
"0.7393",
"0.911",
"0.7524",
"0.513",
"0.7505"
],
[
"0.08826",
"0.7476",
"0.4878",
"0.747",
"0.4878",
"0.761",
"0.08826",
"0.762"
],
[
"0.513",
"0.7524",
"0.911",
"0.7524",
"0.911",
"0.7666",
"0.513",
"0.7666"
],
[
"0.08905",
"0.762",
"0.4844",
"0.762",
"0.4844",
"0.776",
"0.08905",
"0.776"
],
[
"0.5156",
"0.768",
"0.91",
"0.768",
"0.91",
"0.779",
"0.5156",
"0.779"
],
[
"0.08826",
"0.7764",
"0.4844",
"0.7764",
"0.4844",
"0.7905",
"0.08826",
"0.7905"
],
[
"0.513",
"0.781",
"0.91",
"0.781",
"0.91",
"0.7944",
"0.513",
"0.7944"
],
[
"0.08905",
"0.7905",
"0.1111",
"0.7905",
"0.1111",
"0.803",
"0.08905",
"0.803"
],
[
"0.5146",
"0.7944",
"0.911",
"0.7944",
"0.911",
"0.808",
"0.5146",
"0.808"
],
[
"0.2173",
"0.8047",
"0.3513",
"0.8",
"0.353",
"0.8145",
"0.2181",
"0.8193"
],
[
"0.4656",
"0.806",
"0.4836",
"0.806",
"0.4836",
"0.815",
"0.4656",
"0.815"
],
[
"0.513",
"0.808",
"0.8735",
"0.8066",
"0.8735",
"0.821",
"0.513",
"0.8213"
],
[
"0.0866",
"0.8193",
"0.4844",
"0.8193",
"0.4844",
"0.8335",
"0.0866",
"0.8335"
],
[
"0.5312",
"0.8247",
"0.9087",
"0.8247",
"0.9087",
"0.8354",
"0.5312",
"0.8354"
],
[
"0.08826",
"0.834",
"0.4844",
"0.834",
"0.4844",
"0.8477",
"0.08826",
"0.8477"
],
[
"0.513",
"0.837",
"0.911",
"0.837",
"0.911",
"0.851",
"0.513",
"0.851"
],
[
"0.08826",
"0.8477",
"0.1584",
"0.851",
"0.1584",
"0.8623",
"0.08826",
"0.8594"
],
[
"0.1609",
"0.8447",
"0.486",
"0.8447",
"0.486",
"0.8623",
"0.1609",
"0.8623"
],
[
"0.156",
"0.853",
"0.165",
"0.853",
"0.165",
"0.857",
"0.156",
"0.857"
],
[
"0.5146",
"0.851",
"0.911",
"0.851",
"0.911",
"0.8643",
"0.5146",
"0.8643"
],
[
"0.08826",
"0.8613",
"0.486",
"0.86",
"0.486",
"0.875",
"0.08826",
"0.8755"
],
[
"0.5146",
"0.866",
"0.91",
"0.866",
"0.91",
"0.877",
"0.5146",
"0.877"
],
[
"0.0866",
"0.875",
"0.3096",
"0.8755",
"0.3096",
"0.89",
"0.0866",
"0.8896"
],
[
"0.5146",
"0.879",
"0.6284",
"0.879",
"0.6284",
"0.8926",
"0.5146",
"0.8926"
]
] | [
"哈希代码6和另一组延迟之，其中损失施加在涉及之和",
"Remark l:The Order of E Matters.",
"",
"E从字面上看",
"b之间的汉明距离的转换的顶部。NSH还采用了对比学",
"是n延迟z的排序列表，其中每个列表ei包含ci的",
"习的最新进展[Chenetal,2020al作为无监督框架，该",
"批量检索结果的表示形式，以相似度的降序排列。值得",
"框架需要两组独立的随机数据增强。在下文中，我们使",
"注意的是，由于b,和b,预计是相同的，ei，i.e.，e,[1,",
"用和：，e.g.，和，b和b，之和之等表示两个增",
"的第一行自然会接近，表示本身是批处理中最相",
"强对应物的输出。此外，由于我们的训练过程涉及多个",
"关的一个。从第二行开始，相关性降低。Eqs.（3）to（5)",
"排序实例，我们将在必要时使用带有大写字母的行排序",
"定义的步骤自动选择语义相关的实例，并将它们放在E",
"批处理符号，e.g.,X=[51;...;5nl，Z=[21;...;之n",
"的开头。因此，可以很容易地确定每个i的正/负样本",
"等，几是批量大小以启用基于批处理的训练",
"在一个完全可微的方式。此过程模仿检索的真实测试场",
"景，并启用列表式训练以获得更好的性能",
"TrainingObjectiveas aFunction of Sorting我们的最",
"终目标是构建一个完全可微的模型，该模型在一堆语义",
"Remark2:WhyDoWeGatherZ？我们有可能构建",
"排序的候选对象上进行训练：",
"个单瓶颈模型，并用Eq.（5）中的B替换Z。然而",
"Lsorted := NCE 。 Sort&GatheroHamming ofe(α),",
"哈希代码通常很短，信息量较少。尽管他们能够编码粗",
"",
"粒度的语义相似性，但他们很难将每个数据的完整身份",
"Sec.4.3",
"Sec.4.2",
"与其他数据进行比较。我们通过实验表明，当使用严重",
"其中每个堆叠组件都允许梯度传播回fo（）。Fig.2给",
"依赖身份保留的对比学习目标进行训练时，这会导致次",
"出了NSH结构的一警。两组不同增强下的实例都呈现",
"优的性能。我们遵循[Shenetal.,2020】在这里使用另",
"给骨干编码器fe（），产生两组哈希码（B和B）和延",
"个连续表示来支持学习目标。或者，我们的哈希代码可",
"以被视为置换乙的注意头，建立在顶部的损失会自动提",
"迟（乙和乙）。然后，哈希代码确定排序和排列延迟的",
"方式（Sec。4.2）。然后，置换后的潜在张量有助于提出",
"升代码的语义意识",
"的用于训练的SortedNCE损失（Sec。4.3），而训练过",
"4.3SortedNCE",
"程在 Sec 中给出。4.4",
"4.2Sorting and Gathering the Latents",
"许多现有的无监督哈希模型都难以确定数据点之间的相",
"似性，它们通常需要一个延迟的伪标记步骤来引人额外",
"与[Shenetal.,2020]类似，我们在训练模型中植人了成",
"的噪声[Suetal.，2018;Luoetal.,2021b]。然而，NSH",
"对哈希代码距离的计算，以便可以通过顶部的任何损失",
"已经有一堆分类功能E现成的，这使得它非常容易确定",
"来优化代码相似性。形式上，我们将基于代码的亲和矩",
"阳性样本",
"阵S定义为：",
"另一方面，会有多个样本共享相同的语义信息，但是",
"S=BB/2db+0.5",
"(3)",
"香草单标签InfoNCEE[Oordetal.，2018]损失没有考虑",
"语义的这种重叠。我们提出了一个学习目标，该目标适",
"，以便其每个条目表示归一化的成对相似性i.e.，S[i,il=",
"用于多个正对。特别是，由于E是有序的，我们根据日",
"1-Hamming(bi,b,)/db。Eq.（3）仅涉及在任何地方都",
"志的位置定义SortedNCE损失，并让第一个m样本为",
"可微分的线性运算，因此它可以用作神经网络中反向传",
"正,i.e.,",
"播的构建块。请注意，我们不会对S施加任何损失项",
"以防止模型陷入伪标记训练方案，如Sec中所述。1",
"17>",
"SortouttheRelevancewithHashCodes作为NSH",
"Lsorted =",
"mn ",
"的关键操作，我们将批处理中的每个样本视为检索查询",
"2=1j=1",
"同时让整个批处理作为候选。对于每个数据，可以选择",
"k(ei,,zi",
"最相关的数据，并将其放在检索序列的开头，依此类推。",
"log",
"为了在模型中反映此过程，我们调用Eq.（1）来计算每",
"(eilj,,2)+Eh=m+ih(e[k,l,2)",
"个i， i.e.,",
"Positive and negative samples only depend on the sorting positions",
"Pi = softsort(-si) e (0,1)nxn",
"(4)",
"(6)",
"的排序置换矩阵P，其中siE[0,1]\"是i的第i行与",
"S（XMATHX。也就是说，piLi,周在[i,周位置的条",
"，其中k(a,b)=exp(cos(a,b)/Te)。将m和T。视为超",
"参数。直观上，Eq.（6）为e.中的m阳性候选者构造了",
"目代表ak成为与a,[PrilloandEisenschlos，2020]最",
"相关的第i个项目的可能性",
"m对比，其中每个候选者都形成一个单标签交叉摘项。",
"重要的是，由于正样本的特征在e，中具有1···m的固",
"Re-orderFeatures by Soft Gathering/Permutation对",
"定位置，因此NSH不需要额外的argsort运算符来查找",
"于每个ai，NSH在批次e中生成候选嵌人的矩阵，其",
"涉及不可微计算的最相关项。我们在Sec中讨论了这种",
"中最相似项的特征位于顶部。现在我们已经有了来自",
"设计的好处。4.5。通过将第一个m对数视为正，单个",
"Eq.（4）的软置换矩阵，ei可以通过以下软收集过程定",
"m标签交叉摘项也可以在这里工作。此解决方案类似于",
"义:",
"SupCon[Khoslaetal.，2020]，但我们在Sec中通过实",
"ei=P,zeRnxds",
"(5)",
"验显示。5.3在我们的案例中表现不如SortedNCE",
"这里dz是连续延迟之的大小。为了澄清，ei只是ai",
"还要注意，Eq.（6）不是通用的损失，它专用于包含",
"的排序嵌人矩阵，而一批数据使其成为三维张量i.e.",
"Eqs.（3）to（5）描述的排序排序和收集层的检索模型，例",
"E=er",
".;en)e Rnxnxd。Eq.(5）的基于批处理的",
".",
"如NSH。对于每个样本，SortedNCE的时间复杂度为",
"实现可以通过最近的深度学习工具箱中的爱因斯坦求和",
"Omn）。考虑到m通常很小的事实，这种损失不会显",
"约定轻松实现（参见Alg。1）",
"着增加训练时间"
] | {
"arxiv_id": "2201.13322",
"language": "zh",
"page_idx": 3
} |
||
2307.15555-zh-0 | 2307.15555-zh-0.jpg | 1
<NEXT$\_$LINE
# 1 Introduction
。5ex 深度学习技术的最新发展和计算能力的提高使得多媒体内容的生成和编辑在每个人的范围内。虽然技术进步为新的可能性打开了大门 [3],但它们的恶意利用可能会产生严重的后果[5] 。这种现象的一个例子是 deepfakes [14],通过深度学习技术生成的合成多媒体内容，描绘了不属于他们的个人的行为和行为。为了解决这个日益严重的问题，科学界一直致力于开发能够从真正的[21] 中识别假冒材料的检测器。
在音频领域，可以找到两种主要类别的假音频生成算法，i。e。,Text-ToSpeech (TTS) [16,22] 和 Voice Conversion (VC) [9,19] 算法。已经提出了几种方法来检测这种伪造信号$\mathbb{E}[23]$ ,从旨在检测低级伪像$\mathbb{E}[2,13]$ 的方法到关注更多语义方面E[4,7]的其他方法。。其中，在[12]中，First Digit (FD) 统计信息中的错误 w。r。t。广义的 Benford 定律被用作 Random Forest (RF) 分类器的输入，以检测假语音。从分析的音频轨道的 Mel-Frequency Cepstral Coefficient (MFCC)计算特征。有趣的是，作者表明，当在整个音轨上或仅在信号的静音部分上计算时，它们具有同等的区别。同时，在[6]中，检测任务是通过利用一组受语音处理文献启发的特征来执行的，这些特征被用作监督分类器的输入。这些功能旨在将语音建模为自回归过程，同时考虑多个自回归阶数。最后，$\mathbb{E}[1]$的作者通过利用合成算法引人的高阶双谱相关性来区分真假语音，而这种相关性通常在人类语音中找不到。在多媒体取证中，基于多种方法执行语音深伪检测的可能性至关重要。每种方法都针对可能的痕迹中的不同足迹，这些痕迹可以揭示法医分析中的潜在伪造。结合多种解决方案使得攻击者越来越难以创建成功且不可检测的威胁。此外，应用异构检测技术增加了算法对不同场景的适用性
在本文中，我们考虑了用于合成语音检测的最新技术中提供的三组功能，并提出了一种融合它们以执行相同任务的模型。我们考虑的特性是：FD 特性 [12] ;short-term long-term (STLT) 特性[6];双相干特性[1]。然后，按照基于深度学习的方法执行融合过程。我们在分析音频信号的三个不同方面时考虑这些特征：静音，语音和双谱相关。这是相关的 , 因为如上所述，它可以使模型的最终检测性能受益。如文献中所述，所提出的系统优于单一模型，并且对反取证攻击具有出色的鲁棒性，并且在看不见的数据集上具有良好的泛化能力
# 2 Proposed system
。5ex 建议的检测器处理包含语音的音频信号$x$ ,并使用标签$y\in\{$REAL,FAKE$\}$ 对其进行分类，其中 REAL 标识真实语音，而 FAKE 指示合成生成的信号。为此，我们考虑了文献中提出的用于合成语音检测的三个不同特征集，并将它们融合在执行相同任务的新系统中。特别是，这些对应于$\mathbb{E}|12|$ 中提出的 FD 功能，E[6] 中提出的 STLT 功能以及$\overline{\mathrm{E}}[1]$ 中引入的双相干功能。我们将这些特性分别称为$f_\mathrm{FD}$ 、$f_\mathrm{STLT}$和$f_\mathrm{B}$ 。关于 FD 功能$f_\mathrm{FD}$ ,我们仅在所分析信号的静音部分上计算它们，因为事实证明，在整个音轨$\mathbb{E}[12]$ 上计算它们具有同等的判别力
$\begin{array}{c}\text{我们计算了相应论文}\mathbb{E}[1,6,12]&\text{中提出的特征集 }f_{\mathrm{FD}},f_{\mathrm{STLT}}\text{ 和 }f_{\mathrm{B}},\\\text{将它们用作所提出模型的输入。三个向量的维数分别为 }N_{\mathrm{FD}}=416\mathrm{~、}N_{\mathrm{STLT}}=\end{array}$ | [
[
"0.777",
"0.1193",
"0.786",
"0.1193",
"0.786",
"0.127",
"0.777",
"0.127"
],
[
"0.246",
"0.1483",
"0.357",
"0.1483",
"0.357",
"0.1622",
"0.246",
"0.1622"
],
[
"0.2157",
"0.1849",
"0.3774",
"0.1863",
"0.3774",
"0.2008",
"0.2157",
"0.1995"
],
[
"0.2214",
"0.2172",
"0.7817",
"0.2172",
"0.7817",
"0.2279",
"0.2214",
"0.2279"
],
[
"0.2197",
"0.2311",
"0.7837",
"0.2311",
"0.7837",
"0.2443",
"0.2197",
"0.2443"
],
[
"0.2214",
"0.2456",
"0.785",
"0.2456",
"0.785",
"0.2588",
"0.2214",
"0.2588"
],
[
"0.2197",
"0.2615",
"0.7837",
"0.2615",
"0.7837",
"0.2747",
"0.2197",
"0.2747"
],
[
"0.2197",
"0.2747",
"0.785",
"0.276",
"0.785",
"0.2903",
"0.2197",
"0.289"
],
[
"0.2197",
"0.2917",
"0.3733",
"0.2917",
"0.3733",
"0.305",
"0.2197",
"0.305"
],
[
"0.2443",
"0.307",
"0.7817",
"0.307",
"0.7817",
"0.3208",
"0.2443",
"0.3208"
],
[
"0.2181",
"0.3213",
"0.7837",
"0.3213",
"0.7837",
"0.3352",
"0.2181",
"0.3352"
],
[
"0.2222",
"0.3372",
"0.7837",
"0.3372",
"0.7837",
"0.351",
"0.2222",
"0.351"
],
[
"0.2222",
"0.3542",
"0.7817",
"0.3542",
"0.7817",
"0.3643",
"0.2222",
"0.3643"
],
[
"0.2214",
"0.3662",
"0.7837",
"0.3662",
"0.7837",
"0.3801",
"0.2214",
"0.3801"
],
[
"0.2197",
"0.3818",
"0.7837",
"0.3818",
"0.7837",
"0.3958",
"0.2197",
"0.3958"
],
[
"0.2197",
"0.3958",
"0.785",
"0.3965",
"0.785",
"0.4104",
"0.2197",
"0.4092"
],
[
"0.2222",
"0.4136",
"0.7817",
"0.4136",
"0.7817",
"0.4236",
"0.2222",
"0.4236"
],
[
"0.2214",
"0.4268",
"0.785",
"0.4268",
"0.785",
"0.44",
"0.2214",
"0.44"
],
[
"0.2197",
"0.4426",
"0.7837",
"0.4426",
"0.7837",
"0.4558",
"0.2197",
"0.4558"
],
[
"0.2214",
"0.4583",
"0.7837",
"0.4583",
"0.7837",
"0.4717",
"0.2214",
"0.4717"
],
[
"0.2214",
"0.4749",
"0.7817",
"0.4749",
"0.7817",
"0.4849",
"0.2214",
"0.4849"
],
[
"0.2222",
"0.4893",
"0.7817",
"0.4893",
"0.7817",
"0.5",
"0.2222",
"0.5"
],
[
"0.2181",
"0.5024",
"0.7817",
"0.5024",
"0.7817",
"0.5166",
"0.2181",
"0.5166"
],
[
"0.2214",
"0.518",
"0.774",
"0.518",
"0.774",
"0.532",
"0.2214",
"0.532"
],
[
"0.2427",
"0.532",
"0.7817",
"0.532",
"0.7817",
"0.5454",
"0.2427",
"0.5454"
],
[
"0.2181",
"0.5474",
"0.7837",
"0.5474",
"0.7837",
"0.561",
"0.2181",
"0.561"
],
[
"0.2214",
"0.5645",
"0.7817",
"0.5645",
"0.7817",
"0.5747",
"0.2214",
"0.5747"
],
[
"0.2197",
"0.5776",
"0.785",
"0.577",
"0.785",
"0.5903",
"0.2197",
"0.5913"
],
[
"0.2197",
"0.5933",
"0.7837",
"0.5933",
"0.7837",
"0.607",
"0.2197",
"0.607"
],
[
"0.2197",
"0.608",
"0.785",
"0.608",
"0.785",
"0.622",
"0.2197",
"0.622"
],
[
"0.2197",
"0.6235",
"0.7075",
"0.6235",
"0.7075",
"0.637",
"0.2197",
"0.637"
],
[
"0.2173",
"0.66",
"0.4192",
"0.661",
"0.4192",
"0.6753",
"0.2173",
"0.675"
],
[
"0.2214",
"0.6914",
"0.7837",
"0.6914",
"0.7837",
"0.705",
"0.2214",
"0.705"
],
[
"0.2197",
"0.7056",
"0.785",
"0.7056",
"0.785",
"0.72",
"0.2197",
"0.72"
],
[
"0.2197",
"0.7197",
"0.7837",
"0.7197",
"0.7837",
"0.733",
"0.2197",
"0.733"
],
[
"0.2197",
"0.735",
"0.785",
"0.735",
"0.785",
"0.7485",
"0.2197",
"0.7485"
],
[
"0.2197",
"0.7505",
"0.7837",
"0.7505",
"0.7837",
"0.7646",
"0.2197",
"0.7646"
],
[
"0.2214",
"0.7666",
"0.7837",
"0.7666",
"0.7837",
"0.781",
"0.2214",
"0.781"
],
[
"0.2214",
"0.781",
"0.785",
"0.781",
"0.785",
"0.7944",
"0.2214",
"0.7944"
],
[
"0.2214",
"0.7974",
"0.2524",
"0.7974",
"0.2524",
"0.8096",
"0.2214",
"0.8096"
],
[
"0.2418",
"0.8096",
"0.785",
"0.811",
"0.785",
"0.8276",
"0.2418",
"0.8257"
],
[
"0.2181",
"0.827",
"0.785",
"0.827",
"0.785",
"0.8413",
"0.2181",
"0.8413"
]
] | [
"1",
"<NEXT_LINE",
"1Introduction",
"。5ex深度学习技术的最新发展和计算能力的提高使得多媒体内容的生成和编辑",
"在每个人的范围内。虽然技术进步为新的可能性打开了大门[3，但它们的恶意",
"利用可能会产生严重的后果[5]。这种现象的一个例子是deepfakes[14]，通过",
"深度学习技术生成的合成多媒体内容，描绘了不属于他们的个人的行为和行为。",
"为了解决这个日益严重的问题，科学界一直致力于开发能够从真正的[21]中识",
"别假冒材料的检测器。",
"在音频领域，可以找到两种主要类别的假音频生成算法，i。e。，Text-To",
"Speech（TTS）[16,22]和VoiceConversion（VC)[9,19]算法。已经提出了几种",
"方法来检测这种伪造信号E[23]，从旨在检测低级伪像E[2,13]的方法到关注更",
"多语义方面E47的其他方法。。其中，在12中，FirstDigit（FD）统计信息",
"中的错误w。r。t。广义的Benford定律被用作RandomForest（RF）分类器的",
"输人，以检测假语音。从分析的音频轨道的Mel-FrequencyCepstralCoefficient",
"(MFCC）计算特征。有趣的是，作者表明，当在整个音轨上或仅在信号的静音部",
"分上计算时，它们具有同等的区别。同时，在6中，检测任务是通过利用一组",
"受语音处理文献启发的特征来执行的，这些特征被用作监督分类器的输人。这些",
"功能旨在将语音建模为自回归过程，同时考虑多个自回归阶数。最后，E[1]的作",
"者通过利用合成算法引入的高阶双谱相关性来区分真假语音，而这种相关性通",
"常在人类语音中找不到。在多媒体取证中，基于多种方法执行语音深伪检测的",
"可能性至关重要。每种方法都针对可能的痕迹中的不同足迹，这些痕迹可以揭",
"示法医分析中的潜在伪造。结合多种解决方案使得攻击者越来越难以创建成功",
"且不可检测的威胁。此外，应用异构检测技术增加了算法对不同场景的适用性",
"在本文中，我们考虑了用于合成语音检测的最新技术中提供的三组功能，并",
"提出了一种融合它们以执行相同任务的模型。我们考虑的特性是：FD特性[12]",
"：short-termlong-term（STLT）特性6：双相干特性11。然后，按照基于深",
"度学习的方法执行融合过程。我们在分析音频信号的三个不同方面时考虑这些",
"特征：静音，语音和双谱相关。这是相关的，因为如上所述，它可以使模型的最",
"终检测性能受益。如文献中所述，所提出的系统优于单一模型，并且对反取证攻",
"击具有出色的鲁棒性，并且在看不见的数据集上具有良好的泛化能力",
"2 Proposed system",
"。5ex建议的检测器处理包含语音的音频信号a，并使用标签yE[REAL,FAKE)",
"对其进行分类，其中REAL标识真实语音，而FAKE指示合成生成的信号。为",
"此，我们考虑了文献中提出的用于合成语音检测的三个不同特征集，并将它们",
"融合在执行相同任务的新系统中。特别是，这些对应于E[12]中提出的FD功",
"能,E[6] 中提出的 STLT功能以及E[1] 中引人的双相干功能。我们将这些特性",
"分别称为fFD、fsTLT和fB。关于FD功能fFD，我们仅在所分析信号的静",
"音部分上计算它们，因为事实证明，在整个音轨E[12]上计算它们具有同等的判",
"别力",
"我们计算了相应论文E[1,6,12]中提出的特征集fFD，fsTLT和fB，并",
"将它们用作所提出模型的输入。三个向量的维数分别为NFD=416、NsTLT="
] | {
"arxiv_id": "2307.15555",
"language": "zh",
"page_idx": 0
} |
||
2307.15555-zh-1 | 2307.15555-zh-1.jpg | 2
Table 1. Architectures of the considered models. The FC layers indicated in bold are
<FigureHere>
those used to extract the embeddings.
Fig.1. Pipeline of the proposed system
800 和$N_{\mathrm{B}}=8$ 。这可能是一个问题，因为在使用深度学习模型时，特征大小的高度不平衡会使分类器难以学习如何为其输入分配正确的权重。为了解决这个问题，我们使用嵌人提取器将特征集投影到不同的特征空间中。由此产生的嵌人，我们称之为$e_\mathrm{FD}$ , $e_\mathrm{STLT}$和$e_\mathrm{B}$ ,分别包含 32 , 64 和 16 功能，以便它们具有更多的可比大小，便于它们的集成
建议的系统由四个需要训练的不同块组成。三个是用于从输入特征生成嵌$\wedge e_{\mathrm{FD~}},e_{\mathrm{STLT~}},e_{\mathrm{B~}}$的 Fully Connected (FC)网络，一个是从嵌入$e_\mathrm{ALL}=$ $[ e_{\mathrm{FD}};$ $e_{\mathrm{STLT}};$ $e_{\mathrm{B} }] .$ 的级联执行检测的最终模型表$\overline{\mathrm{E}}1$ 显示了所考虑模型的体$\dot{\text{系结构，其中粗体表示的 FC 层是用于提取嵌入的层。在所有网络中，我们将}}$ LeakyReLU 视为激活函数，将 Softmax 视为输出
总而言之，检测过程的完整管道由三个步骤构成，如图E1 所示。首先，从分析的输入语音信号$x$中提取特征向量$f_\mathrm{FD},f_\mathrm{STLT}$和$f_\mathrm{B}$ 。然后，将三组特征作为输入提供给三个单独的 FC 神经网络，以执行降维并提取嵌人$e_\mathrm{FD},e_\mathrm{STLT}$ 和$e_\mathrm{B}$ 。最后，FC 网络将三个生成的嵌入的串联$e_\mathrm{ALL}$作为输人，并执行合成语音检测。该模型的输出为$\hat{y}$ ,对应于信号$x$的类$y$的估计
我们决定采用一种基于深度学习的方法来利用所分析的特征集中的大多数一致性和不一致性。前三个 FC 模型用于执行特征的降维，这有助于创建仅包含相应特征向量$f$的最关键信息的嵌入$e$ 。这样，当我们将它们连接起来时，$e_\mathrm{ALL}$ 中包含的功能对于手头的任务非常有用，并提高了系统的检测能力 | [
[
"0.2181",
"0.11615",
"0.2296",
"0.11615",
"0.2296",
"0.1276",
"0.2181",
"0.1276"
],
[
"0.2181",
"0.1445",
"0.7876",
"0.1445",
"0.7876",
"0.1578",
"0.2181",
"0.1578"
],
[
"0.2173",
"0.1559",
"0.4714",
"0.1578",
"0.4714",
"0.1724",
"0.2173",
"0.1704"
],
[
"0.3284",
"0.1863",
"0.3782",
"0.1863",
"0.3782",
"0.1976",
"0.3284",
"0.1976"
],
[
"0.423",
"0.1863",
"0.4836",
"0.1863",
"0.4836",
"0.1976",
"0.423",
"0.1976"
],
[
"0.5195",
"0.1863",
"0.5825",
"0.1863",
"0.5825",
"0.1976",
"0.5195",
"0.1976"
],
[
"0.6167",
"0.1863",
"0.681",
"0.1863",
"0.681",
"0.1976",
"0.6167",
"0.1976"
],
[
"0.4126",
"0.202",
"0.492",
"0.1995",
"0.492",
"0.214",
"0.4126",
"0.2166"
],
[
"0.6113",
"0.202",
"0.689",
"0.1995",
"0.689",
"0.214",
"0.6113",
"0.2166"
],
[
"0.527",
"0.2026",
"0.5786",
"0.2026",
"0.5786",
"0.214",
"0.527",
"0.214"
],
[
"0.3203",
"0.2026",
"0.3882",
"0.2026",
"0.3882",
"0.214",
"0.3203",
"0.214"
],
[
"0.3203",
"0.2166",
"0.3857",
"0.2166",
"0.3857",
"0.2279",
"0.3203",
"0.2279"
],
[
"0.4182",
"0.2166",
"0.4844",
"0.2166",
"0.4844",
"0.2279",
"0.4182",
"0.2279"
],
[
"0.5254",
"0.2166",
"0.5767",
"0.2166",
"0.5767",
"0.2279",
"0.5254",
"0.2279"
],
[
"0.6177",
"0.2166",
"0.6787",
"0.2166",
"0.6787",
"0.2279",
"0.6177",
"0.2279"
],
[
"0.3179",
"0.2247",
"0.3882",
"0.226",
"0.3882",
"0.2374",
"0.3179",
"0.2361"
],
[
"0.415",
"0.2247",
"0.4878",
"0.226",
"0.4878",
"0.2374",
"0.415",
"0.2361"
],
[
"0.5156",
"0.2267",
"0.5864",
"0.2267",
"0.5864",
"0.2374",
"0.5156",
"0.2374"
],
[
"0.614",
"0.2267",
"0.6846",
"0.2267",
"0.6846",
"0.2374",
"0.614",
"0.2374"
],
[
"0.3186",
"0.2361",
"0.3865",
"0.2361",
"0.3865",
"0.2474",
"0.3186",
"0.2474"
],
[
"0.4182",
"0.2361",
"0.4878",
"0.2361",
"0.4878",
"0.2474",
"0.4182",
"0.2474"
],
[
"0.517",
"0.2361",
"0.5864",
"0.2361",
"0.5864",
"0.2474",
"0.517",
"0.2474"
],
[
"0.614",
"0.2361",
"0.6846",
"0.2361",
"0.6846",
"0.2474",
"0.614",
"0.2474"
],
[
"0.3228",
"0.25",
"0.384",
"0.25",
"0.384",
"0.2615",
"0.3228",
"0.2615"
],
[
"0.4224",
"0.25",
"0.482",
"0.25",
"0.482",
"0.2615",
"0.4224",
"0.2615"
],
[
"0.524",
"0.25",
"0.581",
"0.25",
"0.581",
"0.2615",
"0.524",
"0.2615"
],
[
"0.6235",
"0.25",
"0.6733",
"0.25",
"0.6733",
"0.2615",
"0.6235",
"0.2615"
],
[
"0.3179",
"0.2583",
"0.3865",
"0.2588",
"0.3865",
"0.2708",
"0.3179",
"0.2695"
],
[
"0.4167",
"0.26",
"0.4878",
"0.26",
"0.4878",
"0.2708",
"0.4167",
"0.2708"
],
[
"0.5156",
"0.26",
"0.5864",
"0.26",
"0.5864",
"0.2708",
"0.5156",
"0.2708"
],
[
"0.3186",
"0.2695",
"0.3865",
"0.2695",
"0.3865",
"0.281",
"0.3186",
"0.281"
],
[
"0.4182",
"0.2695",
"0.4878",
"0.2695",
"0.4878",
"0.281",
"0.4182",
"0.281"
],
[
"0.517",
"0.2695",
"0.585",
"0.2695",
"0.585",
"0.281",
"0.517",
"0.281"
],
[
"0.326",
"0.2827",
"0.3816",
"0.2827",
"0.3816",
"0.2942",
"0.326",
"0.2942"
],
[
"0.4272",
"0.2827",
"0.4763",
"0.2827",
"0.4763",
"0.2942",
"0.4272",
"0.2942"
],
[
"0.524",
"0.2798",
"0.5786",
"0.2822",
"0.5767",
"0.2966",
"0.523",
"0.2942"
],
[
"0.3186",
"0.2935",
"0.3896",
"0.2935",
"0.3896",
"0.3037",
"0.3186",
"0.3037"
],
[
"0.3186",
"0.303",
"0.3865",
"0.303",
"0.3865",
"0.3145",
"0.3186",
"0.3145"
],
[
"0.33",
"0.3164",
"0.3782",
"0.3164",
"0.3782",
"0.3276",
"0.33",
"0.3276"
],
[
"0.3228",
"0.3333",
"0.384",
"0.3333",
"0.384",
"0.3447",
"0.3228",
"0.3447"
],
[
"0.4182",
"0.334",
"0.4885",
"0.334",
"0.4885",
"0.3447",
"0.4182",
"0.3447"
],
[
"0.5254",
"0.334",
"0.58",
"0.334",
"0.58",
"0.3447",
"0.5254",
"0.3447"
],
[
"0.6445",
"0.336",
"0.6543",
"0.336",
"0.6543",
"0.3435",
"0.6445",
"0.3435"
],
[
"0.3635",
"0.3865",
"0.3694",
"0.3865",
"0.3694",
"0.3997",
"0.3635",
"0.3997"
],
[
"0.4224",
"0.406",
"0.4534",
"0.406",
"0.4534",
"0.4136",
"0.4224",
"0.4136"
],
[
"0.5703",
"0.4048",
"0.638",
"0.4048",
"0.638",
"0.4153",
"0.5703",
"0.4153"
],
[
"0.415",
"0.4426",
"0.4534",
"0.4426",
"0.4534",
"0.4526",
"0.415",
"0.4526"
],
[
"0.3635",
"0.4092",
"0.3694",
"0.4092",
"0.3694",
"0.433",
"0.3635",
"0.433"
],
[
"0.3635",
"0.4749",
"0.6323",
"0.476",
"0.6323",
"0.4893",
"0.3635",
"0.488"
],
[
"0.2181",
"0.529",
"0.785",
"0.529",
"0.785",
"0.5425",
"0.2181",
"0.5425"
],
[
"0.2197",
"0.5444",
"0.7837",
"0.5444",
"0.7837",
"0.558",
"0.2197",
"0.558"
],
[
"0.2197",
"0.5596",
"0.785",
"0.5596",
"0.785",
"0.5728",
"0.2197",
"0.5728"
],
[
"0.2197",
"0.5737",
"0.786",
"0.5747",
"0.786",
"0.589",
"0.2197",
"0.5884"
],
[
"0.2214",
"0.5913",
"0.4968",
"0.5913",
"0.4968",
"0.6016",
"0.2214",
"0.6016"
],
[
"0.2427",
"0.6123",
"0.785",
"0.6123",
"0.785",
"0.6255",
"0.2427",
"0.6255"
],
[
"0.2197",
"0.627",
"0.7837",
"0.627",
"0.7837",
"0.6445",
"0.2197",
"0.6445"
],
[
"0.2197",
"0.643",
"0.7837",
"0.6426",
"0.7837",
"0.657",
"0.2197",
"0.6577"
],
[
"0.2222",
"0.66",
"0.7837",
"0.66",
"0.7837",
"0.6704",
"0.2222",
"0.6704"
],
[
"0.2181",
"0.674",
"0.576",
"0.674",
"0.576",
"0.6875",
"0.2181",
"0.6875"
],
[
"0.2443",
"0.696",
"0.7837",
"0.6963",
"0.7837",
"0.711",
"0.2443",
"0.7104"
],
[
"0.2173",
"0.711",
"0.786",
"0.7104",
"0.786",
"0.728",
"0.2173",
"0.729"
],
[
"0.2181",
"0.7236",
"0.785",
"0.7266",
"0.785",
"0.744",
"0.2181",
"0.741"
],
[
"0.2197",
"0.742",
"0.785",
"0.742",
"0.785",
"0.757",
"0.2197",
"0.757"
],
[
"0.2197",
"0.7583",
"0.6543",
"0.7583",
"0.6543",
"0.7715",
"0.2197",
"0.7715"
],
[
"0.2443",
"0.781",
"0.7715",
"0.781",
"0.7715",
"0.7944",
"0.2443",
"0.7944"
],
[
"0.2197",
"0.797",
"0.785",
"0.797",
"0.785",
"0.81",
"0.2197",
"0.81"
],
[
"0.2173",
"0.808",
"0.786",
"0.81",
"0.786",
"0.8276",
"0.2173",
"0.8257"
],
[
"0.2197",
"0.827",
"0.693",
"0.827",
"0.693",
"0.8403",
"0.2197",
"0.8403"
]
] | [
"2",
"Table 1. Architectures of the considered models. The FC layers indicated in bold are",
"those used to extract the embeddings.",
"FD model",
"STLT model",
"Bicoh. model",
"Fusion model",
"JSTET (800, 1)",
"FALL : (112, 1)",
"fg : (8 1)",
"frp : (416, 1)",
"FC (416, 128)",
"FC (800, 512)",
"FC (8, 32)",
"FC (112, 32)",
"Dropout (0.25)",
"Dropout (0.25)",
"Dropout (0.25)",
"Dropout (0.25)",
"BatchNorm1D",
"BatchNorm1D",
"BatchNorm1D",
"BatchNorm1D",
"FC (128, 64)",
"FC (512, 64)",
"FC (32, 16)",
"FC (32, 2)",
"Dropout (0.25)",
"Dropout (0.25)",
"Dropout (0.25)",
"BatchNorm1D",
"BatchNorm1D",
"BatchNorm1D",
"FC (64, 32)",
"FC (64, 2)",
"FC (16, 2)",
"Dropout (0.25)",
"BatchNorm1D",
"FC (32, 2)",
"eFD : (32, 1)",
"esTLT : (64, 1)",
"eB : (16, 1)",
"",
"u",
"TLTI",
"Fusion mode",
"Bicoh.",
"Exra(",
"Fig.1. Pipeline of the proposed system",
"800和NB=8。这可能是一个问题，因为在使用深度学习模型时，特征大小的",
"高度不平衡会使分类器难以学习如何为其输人分配正确的权重。为了解决这个",
"问题，我们使用嵌人提取器将特征集投影到不同的特征空间中。由此产生的嵌",
"入，我们称之为eFD，esTLT和eB，分别包含32，64和16功能，以便它们",
"具有更多的可比天小，便于它们的集成",
"建议的系统由四个需要训练的不同块组成。三个是用于从输入特征生成嵌",
"人eFD，esTLT，eB 的Fully Connected(FC)网络，一个是从嵌人eALL",
"[eFD；esTLT；eB].的级联执行检测的最终模型表E1显示了所考虑模型的体",
"系结构，其中粗体表示的F层是用于提取嵌入的层。在所有网络中，我们将",
"LeakyReLU视为激活函数，将Softmax视为输出",
"总而言之，检测过程的完整管道由三个步骤构成，如图E1所示。首先，从分",
"析的输入语音信号a中提取特征向量fD：fsTLT和fB。然后，将三组特征",
"作为输人提供给三个单独的FC神经网络，以执行降维并提取嵌人eFD，esTLT",
"和eB。最后，FC网络将三个生成的嵌人的串联eALL作为输入，并执行合成",
"语音检测。该模型的输出为9，对应于信号的类y的估计",
"我们决定采用一种基于深度学习的方法来利用所分析的特征集中的大多数",
"致性和不一致性。前三个FC模型用于执行特征的降维，这有助于创建仅包含相",
"应特征向量f的最关键信息的嵌人e。这样，当我们将它们连接起来时，eALL",
"中包含的功能对于手头的任务非常有用，并提高了系统的检测能力"
] | {
"arxiv_id": "2307.15555",
"language": "zh",
"page_idx": 1
} |
||
2307.15555-zh-2 | 2307.15555-zh-2.jpg | 3
# 3 Experimental setup
。5ex 在本节中，我们为读者提供了一些有关用于评估所建议检测器性能的评估设置的见解。首先，我们描述用于训练和测试系统的数据集。然后，我们说明了用于训练 FC 网络的训练过程。最后，我们定义了用于测试模型鲁棒性的反取证攻击
考虑的数据集。在所有实验中，我们考虑了几个数据集$\mathcal{D}$来训练和测试所提出的合成语音检测器。这些包括真实和 deepfake 语音样本，总共超过 175 000 音轨。我们使用以下数据集来测试所提出的系统在实际条件下的鲁棒性，并验证它不会过度拟合到一个集合或域
- ASVspoof 2019 [20] is a speech audio dataset created to develop antispoofing
techniques for automatic speaker verification. We consider its Logical Access (LA) partition of the dataset. It contains both real and deepfake speech data generated with several techniques, including both TTS and VC systems. In particular, the train and dev partitions include authentic signals along with speech samples generated with 6 different algorithms (named A01, A02, ..., A06), while the eval partition comprises real signals and samples from 13 other algorithms (A07,...,A19).
-LJSpeech (LJS)[10] contains audio clips of a single speaker reciting pieces
from non-fiction books.
-LibriSpeech (LS)[15] is an open-source dataset containing audio clips of
authentic speech. From this corpus we considered the subset train-clean-100~. - Cloud2019 corresponds to the dataset proposed in [11]. It includes tracks from different speakers generated considering several TTS cloud services: Amazon AWS Polly (PO), Google Cloud Standard (GS), Google Cloud WaveNet (GW), Microsoft Azure (AZ), and IBM Watson (WA).
-VidTIMIT[18]comprises audio-video recordings of 43people reciting short
sentences and recorded in an offce environment. We consider only the audio
signals of these data since the videos are not related to the task at hand.
培训设置。我们以端到端的方式训练完整的模型，将整个系统视为一个更大的网络，使其能够为最终分类确定参数的最佳配置。或者，我们可以为语音 deepfake 检测任务单独训练单个模型，在提取的嵌人之间执行 LateFusion[E[17]。但是， 我们遇到了第二种方法导致性能较差的情况，因此下一节中报告的所有结果都考虑了端到端培训策略
我们通过监测考虑交叉熵函数计算的验证损失值，为 100 历元训练模型。我们假设 10 时代为早期停止，128 的批量大小以及在平台上适当降低的学习速率$lr=10^{-4}$。最后，我们输入到模型的特征在 0 和 1 之间线性归一化
根据实验，我们在不同的数据集上训练模型。在第一个中，我们只在训练和测试中考虑了 ASVspoof2019。我们这样做是为了获得与介绍我们使用[1,6,12] 的功能集的论文相当的结果。我们假设单篇论文的检测器作为基准来衡量我们， 的表现。在这种情况下，由于 ASVspoof 数据集中包含的真实和虚假轨道数量的高度不平衡，我们训练模型以确保每个批次中两个类的样本数量相同
另一方面，在以下实验中，我们增加了训练集中包含的真实样本数量，将ASVspoof2019 和 LibriSpeech 都视为训练数据集。这样做是为了平衡这两个类 | [
[
"0.774",
"0.1187",
"0.7876",
"0.1187",
"0.7876",
"0.1288",
"0.774",
"0.1288"
],
[
"0.2157",
"0.1465",
"0.4453",
"0.1483",
"0.4453",
"0.1635",
"0.2157",
"0.161"
],
[
"0.2197",
"0.1755",
"0.7837",
"0.1768",
"0.7837",
"0.1913",
"0.2197",
"0.1901"
],
[
"0.2197",
"0.1901",
"0.7837",
"0.1913",
"0.7837",
"0.2058",
"0.2197",
"0.2052"
],
[
"0.2197",
"0.207",
"0.785",
"0.207",
"0.785",
"0.2203",
"0.2197",
"0.2203"
],
[
"0.2197",
"0.2235",
"0.2532",
"0.2235",
"0.2532",
"0.2349",
"0.2197",
"0.2349"
],
[
"0.2197",
"0.2437",
"0.7837",
"0.2437",
"0.7837",
"0.2568",
"0.2197",
"0.2568"
],
[
"0.2197",
"0.2588",
"0.7837",
"0.2588",
"0.7837",
"0.2727",
"0.2197",
"0.2727"
],
[
"0.2197",
"0.276",
"0.7837",
"0.276",
"0.7837",
"0.286",
"0.2197",
"0.286"
],
[
"0.2197",
"0.289",
"0.451",
"0.289",
"0.451",
"0.303",
"0.2197",
"0.303"
],
[
"0.2181",
"0.3132",
"0.786",
"0.3132",
"0.786",
"0.3276",
"0.2181",
"0.3276"
],
[
"0.2361",
"0.3289",
"0.785",
"0.3289",
"0.785",
"0.3423",
"0.2361",
"0.3423"
],
[
"0.2378",
"0.3447",
"0.785",
"0.3447",
"0.785",
"0.358",
"0.2378",
"0.358"
],
[
"0.2378",
"0.3591",
"0.7837",
"0.3591",
"0.7837",
"0.3726",
"0.2378",
"0.3726"
],
[
"0.2378",
"0.375",
"0.785",
"0.375",
"0.785",
"0.3882",
"0.2378",
"0.3882"
],
[
"0.2345",
"0.387",
"0.7812",
"0.3894",
"0.7812",
"0.404",
"0.2345",
"0.4016"
],
[
"0.2385",
"0.4048",
"0.7837",
"0.4048",
"0.7837",
"0.4185",
"0.2385",
"0.4185"
],
[
"0.2378",
"0.4216",
"0.4207",
"0.4216",
"0.4207",
"0.432",
"0.2378",
"0.432"
],
[
"0.2214",
"0.433",
"0.785",
"0.4343",
"0.785",
"0.449",
"0.2214",
"0.4475"
],
[
"0.2361",
"0.4507",
"0.398",
"0.4507",
"0.398",
"0.461",
"0.2361",
"0.461"
],
[
"0.2214",
"0.4634",
"0.786",
"0.4634",
"0.786",
"0.4778",
"0.2214",
"0.4778"
],
[
"0.2361",
"0.4792",
"0.7837",
"0.4792",
"0.7837",
"0.4924",
"0.2361",
"0.4924"
],
[
"0.2239",
"0.4976",
"0.2385",
"0.4976",
"0.2385",
"0.502",
"0.2239",
"0.502"
],
[
"0.2336",
"0.4924",
"0.785",
"0.4937",
"0.785",
"0.5083",
"0.2336",
"0.507"
],
[
"0.2361",
"0.5083",
"0.7817",
"0.507",
"0.7817",
"0.5215",
"0.2361",
"0.523"
],
[
"0.2378",
"0.5234",
"0.785",
"0.5234",
"0.785",
"0.537",
"0.2378",
"0.537"
],
[
"0.2378",
"0.537",
"0.6284",
"0.5386",
"0.6284",
"0.553",
"0.2378",
"0.552"
],
[
"0.2336",
"0.552",
"0.785",
"0.553",
"0.785",
"0.5674",
"0.2336",
"0.5664"
],
[
"0.2361",
"0.5684",
"0.785",
"0.5684",
"0.785",
"0.582",
"0.2361",
"0.582"
],
[
"0.2345",
"0.584",
"0.7563",
"0.582",
"0.7563",
"0.595",
"0.2345",
"0.5977"
],
[
"0.2181",
"0.6143",
"0.7837",
"0.6143",
"0.7837",
"0.6274",
"0.2181",
"0.6274"
],
[
"0.2197",
"0.63",
"0.7837",
"0.63",
"0.7837",
"0.643",
"0.2197",
"0.643"
],
[
"0.2197",
"0.646",
"0.7837",
"0.646",
"0.7837",
"0.659",
"0.2197",
"0.659"
],
[
"0.2197",
"0.661",
"0.786",
"0.661",
"0.786",
"0.675",
"0.2197",
"0.675"
],
[
"0.2197",
"0.6753",
"0.3823",
"0.6753",
"0.3823",
"0.6895",
"0.2197",
"0.6895"
],
[
"0.2443",
"0.6914",
"0.785",
"0.6914",
"0.785",
"0.705",
"0.2443",
"0.705"
],
[
"0.2214",
"0.7056",
"0.785",
"0.7056",
"0.785",
"0.7197",
"0.2214",
"0.7197"
],
[
"0.2173",
"0.7183",
"0.7075",
"0.7197",
"0.7075",
"0.7344",
"0.2173",
"0.733"
],
[
"0.2467",
"0.7373",
"0.7817",
"0.7373",
"0.7817",
"0.7476",
"0.2467",
"0.7476"
],
[
"0.2197",
"0.752",
"0.7837",
"0.752",
"0.7837",
"0.765",
"0.2197",
"0.765"
],
[
"0.2197",
"0.765",
"0.7837",
"0.7666",
"0.7837",
"0.78",
"0.2197",
"0.779"
],
[
"0.2197",
"0.781",
"0.785",
"0.781",
"0.785",
"0.7944",
"0.2197",
"0.7944"
],
[
"0.2197",
"0.797",
"0.709",
"0.797",
"0.709",
"0.81",
"0.2197",
"0.81"
],
[
"0.2467",
"0.813",
"0.7837",
"0.813",
"0.7837",
"0.8237",
"0.2467",
"0.8237"
],
[
"0.2197",
"0.827",
"0.786",
"0.827",
"0.786",
"0.8403",
"0.2197",
"0.8403"
]
] | [
"3",
"3Experimental setup",
"。5ex在本节中，我们为读者提供了一些有关用于评估所建议检测器性能的评估",
"设置的见解。首先，我们描述用于训练和测试系统的数据集。然后，我们说明了",
"用于训练FC网络的训练过程。最后，我们定义了用于测试模型鲁棒性的反取证",
"攻击",
"考虑的数据集。在所有实验中，我们考虑了几个数据集D来训练和测试所提出",
"的合成语音检测器。这些包括真实和deepfake语音样本，总共超过175000音",
"轨。我们使用以下数据集来测试所提出的系统在实际条件下的鲁棒性，并验证",
"它不会过度拟合到一个集合或域",
"- AsVspoof 2019 [20] is a speech audio dataset created to develop antispoofing",
"techniques for automatic speaker verification. We consider its Logical Access",
"(LA) partition of the dataset. It contains both real and deepfake speech data",
"generated with several techniques, including both TTS and VC systems. In",
"particular, the train and dev partitions include authentic signals along with",
"speech samples generated with 6 different algorithms (named A01, A02, .:",
"A06), while the eval partition comprises real signals and samples from 13 other",
"algorithms(A07,...,A19)",
"LJSpeech (LJS) [10] contains audio clips of a single speaker reciting pieces",
"fromnon-fiction books",
"LibriSpeech (LS) [15] is an open-source dataset containing audio clips of",
"authentic speech. From this corpus we considered the subset train-clean-100",
"一",
"Cloud2019 corresponds to the dataset proposed in [11] . It includes tracks from",
"different speakers generated considering several TTS cloud services: Ama",
"zon AWS Polly (PO), Google Cloud Standard (GS), Google Cloud WaveNet",
"(GW), Microsoft Azure (AZ), and IBM Watson (WA).",
"VidTIMIT [18] comprises audio-video recordings of 43 people reciting short",
"sentences and recorded in an office environment. We consider only the audio",
"signals of these data since the videos are not related to the task at hand.",
"培训设置。我们以端到端的方式训练完整的模型，将整个系统视为一个更大的网",
"络，使其能够为最终分类确定参数的最佳配置。或者，我们可以为语音deepfake",
"检测任务单独训练单个模型，在提取的嵌人之间执行LateFusionE[17]。但是",
"我们遇到了第二种方法导致性能较差的情况，因此下一节中报告的所有结果都",
"考虑了端到端培训策略",
"我们通过监测考虑交叉熵函数计算的验证损失值，为100历元训练模型。我",
"们假设10时代为早期停止，128的批量大小以及在平台上适当降低的学习速率",
"lr=10-4。最后，我们输人到模型的特征在0和1之间线性归一化",
"根据实验，我们在不同的数据集上训练模型。在第一个中，我们只在训练和",
"测试中考虑了ASVspoof2019。我们这样做是为了获得与介绍我们使用[1,6,12]",
"的功能集的论文相当的结果。我们假设单篇论文的检测器作为基准来衡量我们",
"的表现。在这种情况下，由于ASVspoof数据集中包含的真实和虚假轨道数量的",
"高度不平衡，我们训练模型以确保每个批次中两个类的样本数量相同",
"另一方面，在以下实验中，我们增加了训练集中包含的真实样本数量，将",
"ASVspoof2019和LibriSpeech都视为训练数据集。这样做是为了平衡这两个类，"
] | {
"arxiv_id": "2307.15555",
"language": "zh",
"page_idx": 2
} |
||
2307.15555-zh-3 | 2307.15555-zh-3.jpg | 4
并提供更广泛的真实样本分布，从而提高系统对看不见的数据的检测能力。然而，在这种情况下，由于从 ASVspoof 和 LibriSpeech 的真实样本中提取的特征分布不同，我们经历了不稳定的训练。为了解决这个问题，我们决定通过使用权重
$$w(D,L)=\frac{1}{|S(D,L)|},$$
对每个样本的成本进行加权来训练网络，其中$\left|S(D,L)\right|$是数据集 $D\in\{$ASVspoof $train$,Librispeech$\}$
中标签为$L\in\{0,1\}$的样本集的基数。事实证明，这在稳定验证损失的下降方面是有效的，因为它迫使分类器学习有意义的特征来辨别真实和虚假样本，而不是分离 ASVspoof 和 Librispeech 数据的分布
总而言之，我们在训练中考虑了两种不同的数据集配置。在第一个中，我们仅在具有批处理平衡的 ASVspoof 训练和开发上进行了培训。我们使用此设置将系统的性能与参考论文$\mathbb{E}[1,6,12]$ 的性能进行基准测试，这些论文已在这些数据上进行了训练。在第二个中，我们使用样本加权训练系统 ASVspoof 和LibriSpeech。我们考虑将此设置用于更具挑战性的实验，例如数据集泛化和反取证攻击
反取证攻击。在现实世界中，反 deepfake 系统实现的检测性能可能并不完全重要。实际上，由于编辑和格式化操作，“野外”信号的质量会下降，从而损害了deepfake 检测系统的准确性。这是在网络或社交平台上找到的语音轨道的情况， 通常会受到压缩和后处理操作的影响
在这项工作中，我们测试了所提出的检测器对其中两个操作的鲁棒性，即高斯噪声注人和 MP3 压缩。我们这样做是因为白噪声的存在可能会隐藏合成音频生成器引入的一些伪像，而压缩是现实场景中经常存在的操作。我们使用Pythonaudiomentations [8]库执行了所有后处理操作，考虑了三种不同的噪声水平，振幅 std=[0。1、0。01,0。001] 对应于 SNR=[2,22,42],以及两个不同的压缩值(比特率[~kbit/s~]=[128,~32])
# 4 Results
在本节中，我们分析了语音 deepfake 检测任务期间考虑的功能，并评估了所提出的检测器在不同情况下的性能。我们用 Receiver Operating Characteristio (ROC)曲线、Area Under the Curve (AUC) 和平衡精度来表示这些。当 AUC 和平衡精度等于 1 时，达到最佳性能
功能分析。作为第一个实验，我们分析了每个考虑的特征集的特征。我们想调查他们提供的信息有多少是相关的，以避免冗余数据的计算。由于所提出的系统执行三组不同特征的融合，因此我们希望确保三者的内容是正交的，以便我们增加提供给模型的信息量。如前所述，我们使用这些功能来分析语音信号的三个不同方面。$f_\mathrm{FD}$包含有关沉默的信息，$f_\mathrm{STLT}$对语音进行建模，而$f_\mathrm{B}$则利用双谱信息
为了测量三个集合之间的相关性，我们计算特征向量的每对元素的皮尔逊系数。生成的矩阵描述了不同特征之间的互相关以及每个向量的自相关。图 2显示了在 ASVspoof2019 列车数据集上计算的此分析结果的绝对值。在那里，我们可以为每个特征向量识别不同的矩形区域。双相干特征不清晰可见，因为它们的数量要少得多。尽管单个特征向量呈现出相当高的内部相关性，但它们之间 | [
[
"0.2141",
"0.1206",
"0.2255",
"0.11554",
"0.232",
"0.125",
"0.2214",
"0.13"
],
[
"0.2197",
"0.1483",
"0.785",
"0.1483",
"0.785",
"0.1622",
"0.2197",
"0.1622"
],
[
"0.2214",
"0.1635",
"0.7837",
"0.1635",
"0.7837",
"0.178",
"0.2214",
"0.178"
],
[
"0.2197",
"0.1787",
"0.7837",
"0.1787",
"0.7837",
"0.1925",
"0.2197",
"0.1925"
],
[
"0.2181",
"0.1945",
"0.2378",
"0.1945",
"0.2378",
"0.2084",
"0.2181",
"0.2084"
],
[
"0.4248",
"0.207",
"0.581",
"0.2203",
"0.5767",
"0.2424",
"0.4207",
"0.2291"
],
[
"0.5293",
"0.2084",
"0.5415",
"0.2039",
"0.5483",
"0.2153",
"0.536",
"0.2197"
],
[
"0.2181",
"0.2443",
"0.9272",
"0.2456",
"0.9272",
"0.26",
"0.2181",
"0.2588"
],
[
"0.2214",
"0.2615",
"0.785",
"0.2615",
"0.785",
"0.2747",
"0.2214",
"0.2747"
],
[
"0.2197",
"0.2766",
"0.785",
"0.2766",
"0.785",
"0.2903",
"0.2197",
"0.2903"
],
[
"0.2197",
"0.2917",
"0.539",
"0.2917",
"0.539",
"0.305",
"0.2197",
"0.305"
],
[
"0.246",
"0.308",
"0.785",
"0.308",
"0.785",
"0.3213",
"0.246",
"0.3213"
],
[
"0.2214",
"0.3213",
"0.785",
"0.3213",
"0.785",
"0.3352",
"0.2214",
"0.3352"
],
[
"0.2214",
"0.336",
"0.7837",
"0.336",
"0.7837",
"0.3499",
"0.2214",
"0.3499"
],
[
"0.2197",
"0.3516",
"0.7837",
"0.3516",
"0.7837",
"0.3655",
"0.2197",
"0.3655"
],
[
"0.2181",
"0.3674",
"0.7837",
"0.3674",
"0.7837",
"0.3806",
"0.2181",
"0.3806"
],
[
"0.2197",
"0.3833",
"0.2852",
"0.3833",
"0.2852",
"0.3945",
"0.2197",
"0.3945"
],
[
"0.2222",
"0.406",
"0.7837",
"0.406",
"0.7837",
"0.4167",
"0.2222",
"0.4167"
],
[
"0.2181",
"0.4192",
"0.7812",
"0.4192",
"0.7812",
"0.433",
"0.2181",
"0.433"
],
[
"0.2197",
"0.4343",
"0.785",
"0.4343",
"0.785",
"0.4475",
"0.2197",
"0.4475"
],
[
"0.2181",
"0.4495",
"0.4795",
"0.4495",
"0.4795",
"0.4634",
"0.2181",
"0.4634"
],
[
"0.2427",
"0.4634",
"0.785",
"0.4653",
"0.785",
"0.4792",
"0.2427",
"0.4766"
],
[
"0.2197",
"0.4797",
"0.785",
"0.4797",
"0.785",
"0.4937",
"0.2197",
"0.4937"
],
[
"0.2222",
"0.4968",
"0.7817",
"0.4968",
"0.7817",
"0.507",
"0.2222",
"0.507"
],
[
"0.2181",
"0.509",
"0.7837",
"0.5103",
"0.7837",
"0.5244",
"0.2181",
"0.5234"
],
[
"0.2214",
"0.526",
"0.785",
"0.526",
"0.785",
"0.539",
"0.2214",
"0.539"
],
[
"0.2214",
"0.5405",
"0.5034",
"0.5405",
"0.5034",
"0.5537",
"0.2214",
"0.5537"
],
[
"0.2173",
"0.5806",
"0.3284",
"0.5806",
"0.3284",
"0.595",
"0.2173",
"0.595"
],
[
"0.2197",
"0.6123",
"0.7837",
"0.6123",
"0.7837",
"0.6255",
"0.2197",
"0.6255"
],
[
"0.2197",
"0.629",
"0.7817",
"0.629",
"0.7817",
"0.6396",
"0.2197",
"0.6396"
],
[
"0.2197",
"0.6426",
"0.7837",
"0.6426",
"0.7837",
"0.656",
"0.2197",
"0.656"
],
[
"0.2214",
"0.659",
"0.4836",
"0.659",
"0.4836",
"0.6694",
"0.2214",
"0.6694"
],
[
"0.2181",
"0.675",
"0.785",
"0.6753",
"0.785",
"0.6895",
"0.2181",
"0.688"
],
[
"0.2197",
"0.69",
"0.785",
"0.69",
"0.785",
"0.704",
"0.2197",
"0.704"
],
[
"0.2197",
"0.7056",
"0.785",
"0.7056",
"0.785",
"0.7197",
"0.2197",
"0.7197"
],
[
"0.2197",
"0.72",
"0.7837",
"0.72",
"0.7837",
"0.7344",
"0.2197",
"0.7344"
],
[
"0.2181",
"0.735",
"0.7837",
"0.735",
"0.7837",
"0.75",
"0.2181",
"0.75"
],
[
"0.2214",
"0.7524",
"0.2998",
"0.7524",
"0.2998",
"0.7646",
"0.2214",
"0.7646"
],
[
"0.2443",
"0.7666",
"0.785",
"0.7666",
"0.785",
"0.78",
"0.2443",
"0.78"
],
[
"0.2197",
"0.782",
"0.7812",
"0.782",
"0.7812",
"0.7925",
"0.2197",
"0.7925"
],
[
"0.2197",
"0.797",
"0.785",
"0.797",
"0.785",
"0.81",
"0.2197",
"0.81"
],
[
"0.2222",
"0.813",
"0.7817",
"0.813",
"0.7817",
"0.8237",
"0.2222",
"0.8237"
],
[
"0.2197",
"0.827",
"0.785",
"0.827",
"0.785",
"0.8403",
"0.2197",
"0.8403"
]
] | [
"4",
"并提供更广泛的真实样本分布，从而提高系统对看不见的数据的检测能力。然",
"而，在这种情况下，由于从ASVspoof和LibriSpeech的真实样本中提取的特征",
"分布不同，我们经历了不稳定的训练。为了解决这个问题，我们决定通过使用权",
"重",
"w(D,L) = S(D, L)I\"",
"1",
"对每个样本的成本进行加权来训练网络，其中|S(D,L)I是数据集DE{ASVspooftrain,Librispeech)",
"中标签为LE{0,1)的样本集的基数。事实证明，这在稳定验证损失的下降方面",
"是有效的，因为它迫使分类器学习有意义的特征来辨别真实和虚假样本，而不",
"是分离ASVspoof和Librispeech数据的分布",
"总而言之，我们在训练中考虑了两种不同的数据集配置。在第一个中，我",
"们仅在具有批处理平衡的ASVspoof训练和开发上进行了培训。我们使用此设",
"置将系统的性能与参考论文E[1,6,12]的性能进行基准测试，这些论文已在这",
"些数据上进行了训练。在第二个中，我们使用样本加权训练系统ASVspoof和",
"LibriSpeech。我们考虑将此设置用于更具挑战性的实验，例如数据集泛化和反",
"取证攻击",
"反取证攻击。在现实世界中，反deepfake系统实现的检测性能可能并不完全重",
"要。实际上，由于编辑和格式化操作，“野外”信号的质量会下降，从而损害了",
"deepfake检测系统的准确性。这是在网络或社交平台上找到的语音轨道的情况，",
"通常会受到压缩和后处理操作的影响",
"在这项工作中，我们测试了所提出的检测器对其中两个操作的鲁棒性，即",
"高斯噪声注入和MP3压缩。我们这样做是因为白噪声的存在可能会隐藏合成",
"音频生成器引人的一些伪像，而压缩是现实场景中经常存在的操作。我们使用",
"Pythonaudiomentations[8]库执行了所有后处理操作，考虑了三种不同的噪声",
"水平，振幅std=[0。1、0。01,0。001]对应于SNR=[2，22，42]，以及两个不同",
"的压缩值（比特率【kbit/s]=[128，32]）",
"4Results",
"在本节中，我们分析了语音deepfake检测任务期间考虑的功能，并评估了所",
"提出的检测器在不同情况下的性能。我们用ReceiverOperatingCharacteristic",
"(ROC）曲线、AreaUndertheCurve（AUC）和平衡精度来表示这些。当AUC",
"和平衡精度等于1时，达到最佳性能",
"功能分析。作为第一个实验，我们分析了每个考虑的特征集的特征。我们想调查",
"他们提供的信息有多少是相关的，以避免几余数据的计算。由于所提出的系统",
"执行三组不同特征的融合，因此我们希望确保三者的内容是正交的，以便我们",
"增加提供给模型的信息量。如前所述，我们使用这些功能来分析语音信号的三",
"个不同方面。fFD包含有关沉默的信息，fsTLT对语音进行建模，而fB则利",
"用双谱信息",
"为了测量三个集合之间的相关性，我们计算特征向量的每对元素的皮尔逊系",
"数。生成的矩阵描述了不同特征之间的互相关以及每个向量的自相关。图2显",
"示了在ASVspoof2019列车数据集上计算的此分析结果的绝对值。在那里，我们",
"可以为每个特征向量识别不同的矩形区域。双相干特征不清晰可见，因为它们",
"的数量要少得多。尽管单个特征向量呈现出相当高的内部相关性，但它们之间"
] | {
"arxiv_id": "2307.15555",
"language": "zh",
"page_idx": 3
} |
||
2307.15555-zh-4 | 2307.15555-zh-4.jpg | 5
<FigureHere>
$\mathbf{Fig.2.}$ Cross-correlation matrix of the feature vectors $f_\mathrm{FD},f_\mathrm{STLT}$ and $f_\mathrm{B}$ computed on the ASVspoof 2019 train dataset. The absolute values of the coefficients are shown for visualization purposes.
<FigureHere>
Fig.3. ROC curve on the ASVspoof 2019 eval dataset of the four considered models
的交叉系数较低。这意味着三个特征向量没有强相关性，也没有共享太多信息这激发了这些特征的联合使用，提高了模型的检测精度和鲁棒性，并使用 FC 网络执行降维并删除每个特征集中的冗余信息
检测结果。在本实验中，我们分别在 ASVspoof2019 数据集的 train 和 dev 分区上训练和验证所提出的系统，并在 eval 集上进行测试。我们对一次只考虑一个特征集的单个模型做同样的处理，以验证三者的融合实际上提高了模型的检测能力。图 3 显示了此分析的结果，我们可以看到融合模型优于所有单个模型， 并且比考虑的基线具有更好的性能
此外，我们还尝试查看单个分类器的预测之间的多数投票策略是否比建议的方法更有效。端到端架构显示出卓越的性能，证明了其使用的合理性，并通过 6 % 提高了平衡的精度。这可能是因为端到端模型可以选择如何使用和聚合三个功能集提供的信息，从而导致更好的结果 | [
[
"0.7754",
"0.11743",
"0.786",
"0.11743",
"0.786",
"0.1276",
"0.7754",
"0.1276"
],
[
"0.625",
"0.154",
"0.6313",
"0.1497",
"0.6357",
"0.154",
"0.6284",
"0.1578"
],
[
"0.3628",
"0.1736",
"0.3823",
"0.1736",
"0.3823",
"0.1863",
"0.3628",
"0.1863"
],
[
"0.6216",
"0.1799",
"0.636",
"0.1799",
"0.636",
"0.1895",
"0.6216",
"0.1895"
],
[
"0.3652",
"0.2026",
"0.3757",
"0.2026",
"0.3757",
"0.2134",
"0.3652",
"0.2134"
],
[
"0.6216",
"0.2122",
"0.638",
"0.2122",
"0.638",
"0.2216",
"0.6216",
"0.2216"
],
[
"0.625",
"0.2468",
"0.634",
"0.2468",
"0.634",
"0.252",
"0.625",
"0.252"
],
[
"0.357",
"0.2551",
"0.3865",
"0.2551",
"0.3865",
"0.2664",
"0.357",
"0.2664"
],
[
"0.63",
"0.2747",
"0.636",
"0.2798",
"0.63",
"0.2854",
"0.6235",
"0.2798"
],
[
"0.6284",
"0.31",
"0.636",
"0.31",
"0.636",
"0.315",
"0.6284",
"0.315"
],
[
"0.4192",
"0.3193",
"0.4429",
"0.3193",
"0.4429",
"0.3308",
"0.4192",
"0.3308"
],
[
"0.4534",
"0.3193",
"0.4763",
"0.3193",
"0.4763",
"0.3308",
"0.4534",
"0.3308"
],
[
"0.5186",
"0.3193",
"0.5474",
"0.3193",
"0.5474",
"0.332",
"0.5186",
"0.332"
],
[
"0.2173",
"0.3384",
"0.789",
"0.3384",
"0.789",
"0.356",
"0.2173",
"0.356"
],
[
"0.2181",
"0.3528",
"0.786",
"0.3528",
"0.786",
"0.3662",
"0.2181",
"0.3662"
],
[
"0.2173",
"0.3662",
"0.3906",
"0.3687",
"0.3906",
"0.3806",
"0.2173",
"0.379"
],
[
"0.3782",
"0.3977",
"0.392",
"0.3977",
"0.392",
"0.4077",
"0.3782",
"0.4077"
],
[
"0.3774",
"0.4343",
"0.3906",
"0.4343",
"0.3906",
"0.443",
"0.3774",
"0.443"
],
[
"0.3555",
"0.449",
"0.4004",
"0.449",
"0.4004",
"0.532",
"0.3555",
"0.532"
],
[
"0.5034",
"0.535",
"0.585",
"0.535",
"0.585",
"0.5454",
"0.5034",
"0.5454"
],
[
"0.38",
"0.5444",
"0.3896",
"0.5444",
"0.3896",
"0.549",
"0.38",
"0.549"
],
[
"0.5034",
"0.547",
"0.5947",
"0.547",
"0.5947",
"0.5566",
"0.5034",
"0.5566"
],
[
"0.5024",
"0.5566",
"0.6235",
"0.5566",
"0.6235",
"0.5674",
"0.5024",
"0.5674"
],
[
"0.5034",
"0.5674",
"0.584",
"0.5674",
"0.584",
"0.5776",
"0.5034",
"0.5776"
],
[
"0.3782",
"0.582",
"0.3865",
"0.582",
"0.3865",
"0.586",
"0.3782",
"0.586"
],
[
"0.4387",
"0.5854",
"0.455",
"0.5854",
"0.455",
"0.5947",
"0.4387",
"0.5947"
],
[
"0.4844",
"0.5854",
"0.4983",
"0.5854",
"0.4983",
"0.5933",
"0.4844",
"0.5933"
],
[
"0.5312",
"0.586",
"0.5435",
"0.586",
"0.5435",
"0.5923",
"0.5312",
"0.5923"
],
[
"0.5728",
"0.586",
"0.591",
"0.586",
"0.591",
"0.5947",
"0.5728",
"0.5947"
],
[
"0.6216",
"0.5854",
"0.634",
"0.5854",
"0.634",
"0.5933",
"0.6216",
"0.5933"
],
[
"0.464",
"0.5947",
"0.5596",
"0.5947",
"0.5596",
"0.605",
"0.464",
"0.605"
],
[
"0.3945",
"0.587",
"0.407",
"0.587",
"0.407",
"0.5933",
"0.3945",
"0.5933"
],
[
"0.2197",
"0.622",
"0.7817",
"0.622",
"0.7817",
"0.632",
"0.2197",
"0.632"
],
[
"0.2197",
"0.655",
"0.785",
"0.655",
"0.785",
"0.6685",
"0.2197",
"0.6685"
],
[
"0.2181",
"0.6694",
"0.785",
"0.6694",
"0.785",
"0.683",
"0.2181",
"0.683"
],
[
"0.2197",
"0.685",
"0.528",
"0.685",
"0.528",
"0.6987",
"0.2197",
"0.6987"
],
[
"0.2197",
"0.704",
"0.785",
"0.704",
"0.785",
"0.7173",
"0.2197",
"0.7173"
],
[
"0.2214",
"0.7183",
"0.7837",
"0.7183",
"0.7837",
"0.732",
"0.2214",
"0.732"
],
[
"0.2181",
"0.7344",
"0.785",
"0.7344",
"0.785",
"0.7476",
"0.2181",
"0.7476"
],
[
"0.2181",
"0.7485",
"0.785",
"0.7476",
"0.785",
"0.762",
"0.2181",
"0.763"
],
[
"0.2197",
"0.7646",
"0.4656",
"0.7646",
"0.4656",
"0.778",
"0.2197",
"0.778"
],
[
"0.2427",
"0.781",
"0.785",
"0.781",
"0.785",
"0.7944",
"0.2427",
"0.7944"
],
[
"0.2222",
"0.7974",
"0.785",
"0.7974",
"0.785",
"0.808",
"0.2222",
"0.808"
],
[
"0.2181",
"0.811",
"0.7837",
"0.811",
"0.7837",
"0.8247",
"0.2181",
"0.8247"
],
[
"0.2181",
"0.827",
"0.513",
"0.827",
"0.513",
"0.8403",
"0.2181",
"0.8403"
]
] | [
"5",
".",
"feD",
"0.8",
"fs",
"0.6",
".4",
"fSTLT",
"",
"0",
"frp",
"f",
"fsTLT",
"Fig. 2. Cross-correlation matrix of the feature vectors frD, fsTLr and fB computed",
"on the AsVspoof 2019 train dataset. The absolute values of the coefficients are shown",
"for visualization purposes",
"1.0",
"0.8",
"",
"FD_AUC=0.87",
"n5",
"STLT-AUC=0.76",
"Bicoherence AUC = 0.60",
"All AUC = 0.92",
"0",
"0.2",
"0.4",
"0.6",
"0.8",
"1.0",
"False Positive Rate",
"0.0",
"Fig.3.ROC curve on the ASVspoof 2019 eval dataset of thefour considered models",
"的交叉系数较低。这意味着三个特征向量没有强相关性，也没有共享太多信息。",
"这激发了这些特征的联合使用，提高了模型的检测精度和鲁棒性，并使用FC网",
"络执行降维并删除每个特征集中的穴余信息",
"检测结果。在本实验中，我们分别在ASVspoof2019数据集的train和dev分区",
"上训练和验证所提出的系统，并在eval集上进行测试。我们对一次只考虑一个",
"特征集的单个模型做同样的处理，以验证三者的融合实际上提高了模型的检测",
"能力。图3显示了此分析的结果，我们可以看到融合模型优于所有单个模型，",
"并且比考虑的基线具有更好的性能",
"此外，我们还尝试查看单个分类器的预测之间的多数投票策略是否比建议的",
"方法更有效。端到端架构显示出卓越的性能，证明了其使用的合理性，并通过6",
"%提高了平衡的精度。这可能是因为端到端模型可以选择如何使用和聚合三个",
"功能集提供的信息，从而导致更好的结果"
] | {
"arxiv_id": "2307.15555",
"language": "zh",
"page_idx": 4
} |
||
1704.02166-zh-1 | 1704.02166-zh-1.jpg | 况下，还有一个额外的限制，即必须保持人的身份。直观上，解决这两个任务需要一个生成模型，该模型对面部的语义(属性)空间进行建模，并且能够将生成过程的特定于身份和独立于身份的方面解耦
受 [10] 的启发，我们制定了半潜在面部属性空间(SL-FAS),它是由两个用户定义和潜在的属性子空间组成的。用户定义的属性子空间的每个维度对应一个人类标记的可解释属性。以数据驱动的方式学习潜在属性空间，并从面部图像中学习紧凑的隐藏结构
两个子空间耦合，使得 SL-FAS 的学习具有挑战性。最近，在[42] 中，提出了属性条件的深度变分自动编码器框架，该框架可以学习潜在因素(i。e。属性),并在给定这些属性的情况下生成图像。在[42]中，仅学习潜在因素，并给出用户定义的属性作为输入。因此， 他们无法对给定图像的用户定义属性的分布进行建模； 导致无法使用语义属性修改图像。受 InfoGAN [4] 的启发，我们提出了一个网络，该网络可以对用户定义和潜在属性的子空间进行联合建模
在本文中，为了共同学习 SL-FAS, 我们提出了一种半潜在生成对抗网络(SL-GAN)框架，该框架由三个主要组件组成，即(i)编码器-解码器网络，(ii)GAN 和 (iii) 识别网络。在编码器-解码器网络中，编码器网络将面部图像投影到 SL-FAS 中，解码器网络通过解码 SL-FAS 中的属性向量来重建图像。因此，如果给定一个属性向量，解码器网络可以用作生成图像的生成器。GAN 执行生成器和鉴别器最小-最大游戏，以确保生成的图像质量良好，确保生成的图像不能与真实图像区分开。识别网络是从数据中联合学习用户定义和潜在属性的关键配方。特别是，引人了识别网络，以最大化 SL-FAS 中生成的图像和属性之间的互信息。图 1给出了生成和修改面部属性的示例。如修改图 1(b) 的第一行和第三行所示，我们的 SL-GAN 可以在非常嘈杂的背景下修改面部图像的属性
贡献。(1)据我们所知，以前没有任何工作可以使用视觉属性同时进行面部图像的生成和修改。我们的框架仅使用高级语义属性来修改面部图像。(2)我们的SL-GAN 可以通过制定半潜在的面部属性空间来系统地从数据中学习用户定义的和潜在的属性。(3)提出了一种新颖的识别网络，用于从数据中联合学习用户定义的和潜在的属性。(4)最后但并非最不重要的一点是，我们提出了一种迭代训练算法来训练 SL-GAN, 以便同时解决两个相关但又不同的任务
# 2. Related Work
属性学习。以属性为中心的语义表示在多任务 [31] 和迁移学习[19]中得到了广泛的研究。大多数早期作品[19]都假设用户定义的可命名属性空间作为属性。还探索了用户定义的面部属性[5,9,30,40,45,18] 。但是， 由于定义属性集和使用该属性集注释图像需要手动操作，因此指定这些属性很难并且非常昂贵。为此，已
经探索了潜在属性[10] ,用于直接从数据中挖掘属性。新。生成器和鉴别器的损失函数分别为$\mathcal{L}_D=\mathcal{L}_{GAN}$和
需要注意的是，用户定义属性和潜在属性是相辅相成
的，可以同时使用和学习，形成了半潜在属性空间。我们的 SL-GAN 模型是一种专门为生成和修改面部图像而定义的半潜在属性空间形式
深度生成图像建模。一段时间以来，逼真图像的算法生成一直是计算机视觉的重点。沿着这些路线的早期尝试可以追溯到 2006 年深度信念网络( DBN)[13] 。DBN 成功地生成了小图像补丁，但未能推广到更大的图像和更整体的结构。解决这些挑战的最新模型包括自回归模型[16,38,39] ,变分自动编码器$\left ( \mathrm{VAEs}\right ) [ 20, 35, 42]$ ,生成对抗网络(GANs) [4,6,7,8,11,14,17,24,27,28,29] 和深度经常性注意写人器 (DRAW) [12] 。InfoGAN [4] 和 stackGAN [14] 利用识别网络对潜在属性进行建模， 而我们的 SL-GAN 扩展了识别网络以联合对用户定义和潜在属性进行建模，因此我们的框架可以使用属性生成和修改面部图像
语义图像生成。最近，人们一直专注于生成以语义信息(e.g.,文本，姿势或属性)为条件的图像。Reedsetal。[28]研究了使用深度卷积 GAN 从文本中自动合成真实图像的问题。Yan 等人。[42] 提出了一个属性条件的深度变分自动编码器框架，可以从视觉属性生成图像。Mathieu 等人。[23] 在条件生成模型中学习了一组观测值中的隐藏因素。但是，他们的框架只能生成图像，而不能根据属性或其他形式的语义信息修改 existing 图像
图像编辑和合成。我们的工作也与以前在图像编辑[15,26,34,46]上的工作有关。[47]中的公式还可以使$\dot{\text{用 GAN 执行图像编辑任务，其中用户笔划用于指定}}$ 正在更改的属性。相比之下，我们的 SL-GAN 不需要用户笔画作为输入。最近，两份技术报告[26,34]也可以修改面部图像的属性。[34] 提出了一种基于 GAN 的图像变换网络，用于人脸属性处理。在[34]中，一个经过训练的模型只能修改面部图像中的一种类型属性； 相反，我们的 SL-GAN 可以学习同时生成或修改许多属性。事实上，我们在 CelebA 数据集上的模型可以生成或修改 17 不同的面部属性
$3.~{\mathrm{Semi-Latent~GAN~(SL-GAN)}}$
# $3. 1. \textbf{ Background}$
GAN$\underline{\mathbb{E}}[11]$旨在学习从生成的样本中区分真实数据样本；训练生成器网络$G$以欺骗鉴别器网络$D$ 。GAN 使用以下目标进行优化，
$$\begin{aligned}\operatorname*{minmax}_{G}\mathcal{L}_{GAN}&=\mathbb{E}_{x\sim p_{data}(x)}\left[\mathrm{log}D\left(x\right)\right]&\text{(1)}\\&+\mathbb{E}_{z\sim p_{prior}(z)}\left[\log\left(1-D\left(G\left(z\right)\right)\right)\right],\end{aligned}$$
其中$p_{data}\left(x\right)$是真实数据的分布$;p_prior\left(z\right)$是零均值高斯分布$\mathcal{N}(\mathbf{0},\mathbf{1})$ 。$G$和$\dot D$的参数在训练中迭代更$\mathcal{L}_G=-\mathcal{L}_{GAN}$ 。要生成图像，生成器从先前的( a.k.a. | [
[
"0.0809",
"0.09344",
"0.468",
"0.09344",
"0.468",
"0.1067",
"0.0809",
"0.1067"
],
[
"0.5034",
"0.09344",
"0.892",
"0.09344",
"0.892",
"0.1067",
"0.5034",
"0.1067"
],
[
"0.0825",
"0.108",
"0.468",
"0.108",
"0.468",
"0.1218",
"0.0825",
"0.1218"
],
[
"0.505",
"0.10986",
"0.8896",
"0.10986",
"0.8896",
"0.1206",
"0.505",
"0.1206"
],
[
"0.08417",
"0.125",
"0.4673",
"0.125",
"0.4673",
"0.1351",
"0.08417",
"0.1351"
],
[
"0.505",
"0.125",
"0.716",
"0.125",
"0.716",
"0.1351",
"0.505",
"0.1351"
],
[
"0.0809",
"0.1395",
"0.3586",
"0.1395",
"0.3586",
"0.1528",
"0.0809",
"0.1528"
],
[
"0.5063",
"0.1465",
"0.888",
"0.1465",
"0.888",
"0.1572",
"0.5063",
"0.1572"
],
[
"0.1013",
"0.154",
"0.468",
"0.154",
"0.468",
"0.1674",
"0.1013",
"0.1674"
],
[
"0.505",
"0.1604",
"0.892",
"0.1604",
"0.892",
"0.1736",
"0.505",
"0.1736"
],
[
"0.0809",
"0.1692",
"0.4673",
"0.1692",
"0.4673",
"0.1831",
"0.0809",
"0.1831"
],
[
"0.5063",
"0.1768",
"0.8896",
"0.1768",
"0.8896",
"0.1875",
"0.5063",
"0.1875"
],
[
"0.0825",
"0.1863",
"0.4656",
"0.1863",
"0.4656",
"0.1963",
"0.0825",
"0.1963"
],
[
"0.5024",
"0.1895",
"0.888",
"0.1895",
"0.888",
"0.2026",
"0.5024",
"0.2026"
],
[
"0.0825",
"0.1989",
"0.468",
"0.1995",
"0.468",
"0.2134",
"0.0825",
"0.2122"
],
[
"0.5034",
"0.2058",
"0.8896",
"0.2058",
"0.8896",
"0.2197",
"0.5034",
"0.2197"
],
[
"0.0825",
"0.2166",
"0.4224",
"0.2166",
"0.4224",
"0.2267",
"0.0825",
"0.2267"
],
[
"0.505",
"0.2203",
"0.8906",
"0.2203",
"0.8906",
"0.2343",
"0.505",
"0.2343"
],
[
"0.10297",
"0.2311",
"0.468",
"0.2311",
"0.468",
"0.2443",
"0.10297",
"0.2443"
],
[
"0.5034",
"0.2349",
"0.888",
"0.2349",
"0.888",
"0.2488",
"0.5034",
"0.2488"
],
[
"0.0809",
"0.2468",
"0.4673",
"0.2468",
"0.4673",
"0.26",
"0.0809",
"0.26"
],
[
"0.5034",
"0.2507",
"0.8896",
"0.2507",
"0.8896",
"0.2644",
"0.5034",
"0.2644"
],
[
"0.0801",
"0.26",
"0.4673",
"0.2615",
"0.4673",
"0.276",
"0.0801",
"0.2747"
],
[
"0.5034",
"0.2664",
"0.888",
"0.2664",
"0.888",
"0.2798",
"0.5034",
"0.2798"
],
[
"0.0809",
"0.2766",
"0.468",
"0.2766",
"0.468",
"0.2903",
"0.0809",
"0.2903"
],
[
"0.5034",
"0.281",
"0.8896",
"0.281",
"0.8896",
"0.2942",
"0.5034",
"0.2942"
],
[
"0.0825",
"0.2917",
"0.468",
"0.2917",
"0.468",
"0.305",
"0.0825",
"0.305"
],
[
"0.505",
"0.2954",
"0.8906",
"0.2954",
"0.8906",
"0.3093",
"0.505",
"0.3093"
],
[
"0.08417",
"0.308",
"0.468",
"0.308",
"0.468",
"0.318",
"0.08417",
"0.318"
],
[
"0.5034",
"0.3113",
"0.653",
"0.3113",
"0.653",
"0.3245",
"0.5034",
"0.3245"
],
[
"0.0825",
"0.3213",
"0.468",
"0.3213",
"0.468",
"0.3352",
"0.0825",
"0.3352"
],
[
"0.08417",
"0.3384",
"0.4673",
"0.3384",
"0.4673",
"0.3484",
"0.08417",
"0.3484"
],
[
"0.5034",
"0.3333",
"0.8896",
"0.3333",
"0.8896",
"0.3467",
"0.5034",
"0.3467"
],
[
"0.0809",
"0.3516",
"0.3096",
"0.3516",
"0.3096",
"0.3655",
"0.0809",
"0.3655"
],
[
"0.5024",
"0.3484",
"0.8906",
"0.3467",
"0.8906",
"0.3599",
"0.5024",
"0.3623"
],
[
"0.5034",
"0.3623",
"0.8896",
"0.3623",
"0.8896",
"0.3757",
"0.5034",
"0.3757"
],
[
"0.10046",
"0.3687",
"0.468",
"0.3687",
"0.468",
"0.3818",
"0.10046",
"0.3818"
],
[
"0.5024",
"0.377",
"0.8896",
"0.379",
"0.8896",
"0.3926",
"0.5024",
"0.3901"
],
[
"0.08417",
"0.385",
"0.4673",
"0.385",
"0.4673",
"0.3958",
"0.08417",
"0.3958"
],
[
"0.0825",
"0.3984",
"0.468",
"0.3984",
"0.468",
"0.4124",
"0.0825",
"0.4124"
],
[
"0.5034",
"0.3933",
"0.8906",
"0.3933",
"0.8906",
"0.4072",
"0.5034",
"0.4072"
],
[
"0.5034",
"0.4077",
"0.8896",
"0.4077",
"0.8896",
"0.4216",
"0.5034",
"0.4216"
],
[
"0.08417",
"0.4153",
"0.4656",
"0.4153",
"0.4656",
"0.4255",
"0.08417",
"0.4255"
],
[
"0.0809",
"0.428",
"0.468",
"0.428",
"0.468",
"0.4414",
"0.0809",
"0.4414"
],
[
"0.5034",
"0.4236",
"0.8906",
"0.4236",
"0.8906",
"0.4375",
"0.5034",
"0.4375"
],
[
"0.0809",
"0.443",
"0.468",
"0.443",
"0.468",
"0.457",
"0.0809",
"0.457"
],
[
"0.5034",
"0.4382",
"0.8906",
"0.4382",
"0.8906",
"0.452",
"0.5034",
"0.452"
],
[
"0.505",
"0.4539",
"0.5186",
"0.4539",
"0.5186",
"0.4653",
"0.505",
"0.4653"
],
[
"0.0907",
"0.4602",
"0.4673",
"0.4602",
"0.4673",
"0.4702",
"0.0907",
"0.4702"
],
[
"0.0809",
"0.4734",
"0.468",
"0.4734",
"0.468",
"0.4873",
"0.0809",
"0.4873"
],
[
"0.505",
"0.4697",
"0.892",
"0.4697",
"0.892",
"0.483",
"0.505",
"0.483"
],
[
"0.0809",
"0.4893",
"0.468",
"0.4893",
"0.468",
"0.5024",
"0.0809",
"0.5024"
],
[
"0.505",
"0.4841",
"0.8906",
"0.4841",
"0.8906",
"0.4976",
"0.505",
"0.4976"
],
[
"0.5024",
"0.4976",
"0.8896",
"0.4988",
"0.8896",
"0.513",
"0.5024",
"0.512"
],
[
"0.0825",
"0.504",
"0.4673",
"0.504",
"0.4673",
"0.5176",
"0.0825",
"0.5176"
],
[
"0.5024",
"0.5146",
"0.888",
"0.5146",
"0.888",
"0.528",
"0.5024",
"0.528"
],
[
"0.0825",
"0.521",
"0.4656",
"0.521",
"0.4656",
"0.531",
"0.0825",
"0.531"
],
[
"0.0801",
"0.5327",
"0.468",
"0.534",
"0.468",
"0.549",
"0.0801",
"0.5474"
],
[
"0.5063",
"0.531",
"0.888",
"0.531",
"0.888",
"0.5415",
"0.5063",
"0.5415"
],
[
"0.0809",
"0.55",
"0.468",
"0.55",
"0.468",
"0.563",
"0.0809",
"0.563"
],
[
"0.5034",
"0.5444",
"0.8906",
"0.5444",
"0.8906",
"0.558",
"0.5034",
"0.558"
],
[
"0.0801",
"0.5645",
"0.468",
"0.563",
"0.468",
"0.5776",
"0.0801",
"0.579"
],
[
"0.505",
"0.56",
"0.8906",
"0.56",
"0.8906",
"0.5737",
"0.505",
"0.5737"
],
[
"0.505",
"0.5747",
"0.8906",
"0.5747",
"0.8906",
"0.5884",
"0.505",
"0.5884"
],
[
"0.0825",
"0.5806",
"0.3228",
"0.5806",
"0.3228",
"0.5913",
"0.0825",
"0.5913"
],
[
"0.5063",
"0.5913",
"0.8896",
"0.5913",
"0.8896",
"0.6016",
"0.5063",
"0.6016"
],
[
"0.0825",
"0.604",
"0.4656",
"0.604",
"0.4656",
"0.6143",
"0.0825",
"0.6143"
],
[
"0.505",
"0.605",
"0.8906",
"0.605",
"0.8906",
"0.6187",
"0.505",
"0.6187"
],
[
"0.0825",
"0.617",
"0.468",
"0.617",
"0.468",
"0.631",
"0.0825",
"0.631"
],
[
"0.5034",
"0.6206",
"0.7114",
"0.6206",
"0.7114",
"0.634",
"0.5034",
"0.634"
],
[
"0.0825",
"0.6333",
"0.464",
"0.6333",
"0.464",
"0.643",
"0.0825",
"0.643"
],
[
"0.0801",
"0.646",
"0.468",
"0.6465",
"0.468",
"0.661",
"0.0801",
"0.66"
],
[
"0.501",
"0.6484",
"0.778",
"0.6494",
"0.778",
"0.664",
"0.501",
"0.6636"
],
[
"0.0825",
"0.664",
"0.4656",
"0.664",
"0.4656",
"0.675",
"0.0825",
"0.675"
],
[
"0.08417",
"0.678",
"0.468",
"0.678",
"0.468",
"0.6914",
"0.08417",
"0.6914"
],
[
"0.5034",
"0.6753",
"0.636",
"0.6753",
"0.636",
"0.6895",
"0.5034",
"0.6895"
],
[
"0.0809",
"0.6934",
"0.4534",
"0.6934",
"0.4534",
"0.707",
"0.0809",
"0.707"
],
[
"0.0825",
"0.7085",
"0.4697",
"0.7085",
"0.4697",
"0.7217",
"0.0825",
"0.7217"
],
[
"0.5034",
"0.712",
"0.8906",
"0.714",
"0.8906",
"0.729",
"0.5034",
"0.7266"
],
[
"0.0801",
"0.7217",
"0.357",
"0.7227",
"0.357",
"0.7373",
"0.0801",
"0.7363"
],
[
"0.505",
"0.729",
"0.8896",
"0.729",
"0.8896",
"0.742",
"0.505",
"0.742"
],
[
"0.5034",
"0.742",
"0.6733",
"0.744",
"0.6733",
"0.7583",
"0.5034",
"0.756"
],
[
"0.0801",
"0.756",
"0.2279",
"0.756",
"0.2279",
"0.7695",
"0.0801",
"0.7695"
],
[
"0.0825",
"0.782",
"0.4673",
"0.782",
"0.4673",
"0.7954",
"0.0825",
"0.7954"
],
[
"0.5312",
"0.787",
"0.806",
"0.786",
"0.806",
"0.8037",
"0.5312",
"0.8047"
],
[
"0.8228",
"0.7925",
"0.8486",
"0.7925",
"0.8486",
"0.8",
"0.8228",
"0.8"
],
[
"0.8677",
"0.789",
"0.8896",
"0.789",
"0.8896",
"0.802",
"0.8677",
"0.802"
],
[
"0.0809",
"0.797",
"0.468",
"0.797",
"0.468",
"0.81",
"0.0809",
"0.81"
],
[
"0.5415",
"0.803",
"0.5513",
"0.803",
"0.5513",
"0.8066",
"0.5415",
"0.8066"
],
[
"0.5522",
"0.803",
"0.5645",
"0.803",
"0.5645",
"0.806",
"0.5522",
"0.806"
],
[
"0.5664",
"0.8003",
"0.5786",
"0.803",
"0.5767",
"0.8066",
"0.5645",
"0.8047"
],
[
"0.0809",
"0.811",
"0.468",
"0.811",
"0.468",
"0.8247",
"0.0809",
"0.8247"
],
[
"0.6177",
"0.811",
"0.864",
"0.813",
"0.864",
"0.8306",
"0.6177",
"0.829"
],
[
"0.0809",
"0.827",
"0.4673",
"0.827",
"0.4673",
"0.8403",
"0.0809",
"0.8403"
],
[
"0.08417",
"0.8433",
"0.4673",
"0.8433",
"0.4673",
"0.854",
"0.08417",
"0.854"
],
[
"0.505",
"0.8423",
"0.8906",
"0.8423",
"0.8906",
"0.857",
"0.505",
"0.857"
],
[
"0.0825",
"0.8594",
"0.4673",
"0.8594",
"0.4673",
"0.869",
"0.0825",
"0.869"
],
[
"0.505",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.505",
"0.8706"
],
[
"0.0809",
"0.8726",
"0.468",
"0.8726",
"0.468",
"0.886",
"0.0809",
"0.886"
],
[
"0.5034",
"0.8726",
"0.8906",
"0.8726",
"0.8906",
"0.886",
"0.5034",
"0.886"
],
[
"0.0825",
"0.8896",
"0.4673",
"0.8896",
"0.4673",
"0.8994",
"0.0825",
"0.8994"
],
[
"0.5024",
"0.886",
"0.8896",
"0.887",
"0.8896",
"0.9014",
"0.5024",
"0.901"
]
] | [
"况下，还有一个额外的限制，即必须保持人的身份。直",
"的，可以同时使用和学习，形成了半潜在属性空间。我",
"观上，解决这两个任务需要一个生成模型，该模型对面",
"们的SL-GAN模型是一种专门为生成和修改面部图像",
"部的语义（属性）空间进行建模，并且能够将生成过程",
"而定义的半潜在属性空间形式",
"的特定于身份和独立于身份的方面解耦",
"深度生成图像建模。一段时间以来，逼真图像的算法生",
"受「10]的启发，我们制定了半潜在面部属性空间",
"成一直是计算机视觉的重点。沿着这些路线的早期尝试",
"（SL-FAS），它是由两个用户定义和潜在的属性子空间",
"可以追溯到2006年深度信念网络（DBN）/13/。DBN成",
"组成的。用户定义的属性子空间的每个维度对应一个",
"功地生成了小图像补丁，但未能推广到更大的图像和更",
"人类标记的可解释属性。以数据驱动的方式学习潜在",
"整体的结构。解决这些挑战的最新模型包括自回归模型",
"属性空间，并从面部图像中学习紧凑的隐藏结构",
"[16,38，39]，变分自动编码器（VAEs）[20,35，42]，生成",
"两个子空间耦合，使得SL-FAS的学习具有挑战性。",
"对抗网络（GANs）[4,6,7,8,11，14,17,24,27,28,29",
"最近，在【42]中，提出了属性条件的深度变分自动编",
"和深度经常性注意写人器（DRAW）[12]。InfoGAN[4]",
"码器框架，该框架可以学习潜在因素(i。e。属性)，并",
"和stackGAN【14]利用识别网络对潜在属性进行建模",
"在给定这些属性的情况下生成图像。在【42]中，仅学",
"而我们的SL-GAN扩展了识别网络以联合对用户定义",
"习潜在因素，并给出用户定义的属性作为输人。因此，",
"和潜在属性进行建模，因此我们的框架可以使用属性",
"他们无法对给定图像的用户定义属性的分布进行建模：",
"生成和修改面部图像",
"导致无法使用语义属性修改图像。受InfoGAN4]的",
"启发，我们提出了一个网络：该网络可以对用户定义和",
"语义图像生成。最近，人们一直专注于生成以语义信息",
"潜在属性的子空间进行联合建模",
"(e.g.，文本，姿势或属性）为条件的图像。。Reedsetal。",
"[28]研究了使用深度卷积GAN从文本中自动合成真",
"在本文中，为了共同学习SL-FAS，我们提出了一种",
"实图像的问题。Yan等人。[42提出了一个属性条件的",
"半潜在生成对抗网络（SL-GAN）框架，该框架由三个",
"主要组件组成，即（i）编码器-解码器网络，（ii）GAN",
"深度变分自动编码器框架，可以从视觉属性生成图像。",
"Mathieu等人。[23]在条件生成模型中学习了一组观测",
"和（讯i）识别网络。在编码器-解码器网络中，编码器网",
"络将面部图像投影到SL-FAS中，解码器网络通过解",
"值中的隐藏因素。但是，他们的框架只能生成图像，而",
"码SL-FAS中的属性向量来重建图像。因此，如果给定",
"不能根据属性或其他形式的语义信息修改existing图",
"像",
"一个属性向量，解码器网络可以用作生成图像的生成",
"器。GAN执行生成器和鉴别器最小-最大游戏，以确保",
"图像编辑和合成。我们的工作也与以前在图像编辑",
"生成的图像质量良好，确保生成的图像不能与真实图",
"15，26，34，46]上的工作有关。「47]中的公式还可以使",
"用GAN执行图像编辑任务，其中用户笔划用于指定",
"像区分开。识别网络是从数据中联合学习用户定义和",
"正在更改的属性。相比之下，我们的SL-GAN不需要",
"潜在属性的关键配方。特别是，引了识别网络，以最",
"大化SL-FAS中生成的图像和属性之间的互信息。图1",
"用户笔画作为输入。最近，两份技术报告26，34也可",
"给出了生成和修改面部属性的示例。如修改图1（b）",
"以修改面部图像的属性。[34]提出了一种基于GAN的",
"的第一行和第三行所示，我们的SL-GAN可以在非常",
"图像变换网络，用于人脸属性处理。在[34】中，一个",
"经过训练的模型只能修改面部图像中的一种类型属性：",
"噜杂的背景下修改面部图像的属性",
"相反，我们的SL-GAN可以学习同时生成或修改许多",
"责献。（1）据我们所知，以前没有任何工作可以使用",
"属性。事实上，我们在CelebA数据集上的模型可以生",
"视觉属性同时进行面部图像的生成和修改。我们的框",
"成或修改17不同的面部属性",
"架仅使用高级语义属性来修改面部图像。（2）我们的",
"SL-GAN可以通过制定半潜在的面部属性空间来系统",
"3. Semi-Latent GAN (SL-GAN)",
"地从数据中学习用户定义的和潜在的属性。（3）提出",
"了一种新颖的识别网络，用于从数据中联合学习用户",
"3.1. Background",
"定义的和潜在的属性。（4）最后但并非最不重要的",
"点是，我们提出了一种迭代训练算法来训练SL-GAN，",
"GANE[111旨在学习从生成的样本中区分真实数据样",
"以便同时解决两个相关但又不同的任务",
"本；训练生成器网络G以欺骗鉴别器网络D。GAN",
"使用以下目标进行优化，",
"2. Related Work",
"属性学习。以属性为中心的语义表示在多任务【31]和",
"minmaxCGAN = Er~pdata(r) [logD (r)]",
"",
"(1)",
"迁移学习[19]中得到了广泛的研究。大多数早期作品",
"T",
"",
"",
"[19]都假设用户定义的可命名属性空间作为属性。还探",
"+Ez~pprior(2) [log (1 - D (G(z)] ,",
"索了用户定义的面部属性[5，9,30，40，45，18]。但是",
"由于定义属性集和使用该属性集注释图像需要手动操",
"其中Pdata（a）是真实数据的分布；Pprior（z）是零均值",
"作：因此指定这些属性很难并且非常昂贵。为此，已",
"高斯分布(O,1)。G和D的参数在训练中迭代更",
"经探索了潜在属性【10]，用于直接从数据中挖掘属性。",
"新。生成器和鉴别器的损失函数分别为CD=LGAN和",
"需要注意的是，用户定义属性和潜在属性是相辅相成",
"LG=-LGAN。要生成图像，生成器从先前的（a.k.a."
] | {
"arxiv_id": "1704.02166",
"language": "zh",
"page_idx": 1
} |
||
1704.02166-zh-2 | 1704.02166-zh-2.jpg | <FigureHere>
Figure 2. Overview of our SL-GAN.
噪声分布)中绘制样本 $z\sim p_{prior}\left(z\right)=\mathcal{N}\left(\mathbf{0},\mathbf{1}\right)$,然
后使用生成器网络$G$,i.e.$,G(z)$转换该样本
InfoGAN[4]通过最大化互信息$I\left(y;G\left(z,y\right)\right)$,将 GAN 的输入噪声分解为潜在表示$y$和不可压缩噪声$z$,并确保在生成过程中不丢失潜在表示的信息。互信息项可以描述为识别损失。
$\mathcal{L}_{rg}=-\mathbb{E}_{x\sim G(z,y)}\left[\mathbb{E}_{y\sim p_{data}(y|x)}\left[\log Q(y\mid x)\right]\right]$
其中$Q\left(y\mid x\right)$是后验$p_data\left(y\mid x\right)$的近似值。因此，生成器$G$的参数可以更新为 $\mathcal{L}_G_{InfoGAN}=\mathcal{L}_G-\mathcal{L}_{rg}$。InfoGAN 可以以完全无监督的方式学习解开，可解释和有意义的表示
VAEGAN$\underline{\mathrm{E}}[20]$将 VAE 合并到 GAN 中，并在数据空间中将 GAN 的元素错误替换为 VAEGAN 的特征错误。具体来说，它将数据样本$x$编码为潜在表示$z:z\sim$ Enc$\left(x\right)=p_{enc}\left(z\mid x\right)$,并将$z$解码回数据空间：$\tilde{x}\sim$ Dec$\left(x\mid z\right)=p_{dec}\left(x\mid z\right)$。我们可以将正则化先验的损失函数定义为$\mathcal{L}_prior=KL\left(q_{enc}\left(z\mid x\right)\parallel p_{prior}\left(z\right)\right)$ $,q_{enc}\left(z\mid x\right)$是真实后验$p_dec\left(z\mid x\right)$的近似值。重建错误为 $\mathcal{L}_{recon}^{D_{l}}=-\mathbb{E}_{q_{enc}(z|x)}\left[\mathrm{log}p_{dec}\left(D_{l}\left(x\right)\mid z\right)\right]$,其中$D_l\left(x\right)$是鉴别器的 $l-th$ 层的隐藏表示。因此，这种损失使$l-th$鉴别器层的数据表示的预期对数似然之和最小化。因此，VAEGAN 的损失函数为
$$\mathcal{L}_{VAEGAN}=\mathcal{L}_{GAN}+\mathcal{L}_{recon}^{D_{l}}+\mathcal{L}_{prior}$$
(2)
然而，潜在表示$z$是完全无监督的；没有办法明确控制
数据上的属性或使用属性修改面部图像
CVAE [42,35]是条件 VAE。引人独立属性变量$y$ ,通过从$p\left(x\mid y,z\right)$采样来控制$x$的生成过程；其中$\begin{array}{l}p\left(y,z\right)=p\left(y\right)p\left(z\right)\text{。因此，CVAE 的编码器和解码器}\\\end{array}$ 网络是$z\sim Enc\left(x\right)=q_{enc}\left(z\mid x\right)$和$\tilde{x}\sim Dec\left(z,y\right)=$ $p_{dec}\left(x\mid z,y\right)$。引人变量$y$ ,通过从$p\left(x\mid y,z\right)$采样来控制$x$的生成过程；其中$p\left(y,z\right)=p\left(y\right)p\left(z\right)$。尽管如此，$y$仍然是从数据中采样的，而不是像我们的SL-GAN 那样直接从数据中优化和学习。因此，$y$可用于修改类似于建议的 SL-GAN 的属性
3.2. Semi-Latent Facial Attribute Space
GAN 的输入噪声可以进一步分解为两部分：(1)用户定义的$\mathbb{E}$属性$\mathbb{E}y$ 是每个图像 $x$ 的手动注释属性，i。
e。$y\sim p_{data}(y\mid x)$ ;(2)Latent attributes $z_la$表示应该以数据驱动的方式从数据 1 中挖掘的属性，i。e。, $z_{la}\sim p_{data}\left(z_{la}\mid x\right)$
从数学上讲，$y$和$z_la$可以是单变量或多变量；并且这些属性是相互独立的，i。e。$,p\left(y,z_{la}\right)=p\left(y\right)p\left(z_{la}\right)$ 。$y$的每个维度都用真实图像中注释的一种面部属性进行裁剪；我们的 SL-GAN 以有监督的方式训练$y$。相比之下，$z_la$的每个维度都是完全无监督训练的。我们定义了半潜在的面部属性空间，以组合$y$属性和潜在的$z_la$属性
随着分解的输入噪声，生成器的形式现在是$G\left(z_{la},y\right)$。直接从输人数据中学习$z_la$将导致生成器倾向于忽略$p_G\left(x\mid z_{la},y\right)=p_G\left(x\mid y\right)$的潜在属性的琐碎解决方案。相反，我们最大化属性$z_{la}$和生成器分布$G\left(z_la,y\right)$之间的互信息，可以简化为最小化属性$y$ 和$z_{la}$的识别损失
共同学习属性$y$和$z_la$很重要；并确保$z_la$可以表示输入面部图像的未建模方面，而不是重新发现$y$ 。“重新发现”意味着$z_la$的某些维度与$y$在输人图像 i.e.上的分布非常相似，$y$中的相同模式从潜在属性$z_la$中反复发现
## 3.3. Semi-Latent GAN
我们的 SL-GAN 如图所示。2;它由三个部分组成即编码器-解码网络，GAN 和识别网络。用户定义的和潜在的属性由编码器-解码器网络进行编码和解码。识别网络有助于从数据中学习 SL-FAS。在我们的 SLGAN 中，识别网络和鉴别器共享相同的网络结构，并且在最后一层具有不同的 softmax 层。因此，生成器$G_{SL-GAN}$和鉴别器$D_SL-GAN$的损失函数为
$$\mathcal{L}_{G_{SL-GAN}}=\mathcal{L}_{G}+\lambda_{1}\left(\mathcal{L}_{rg_{z}}+\mathcal{L}_{rg_{y}-G}\right)+\lambda_{2}\mathcal{L}_{recon}^{D_{l}}\\\text{(3)}\\\mathcal{L}_{D_{SL-GAN}}=\mathcal{L}_{D}+\left(\mathcal{L}_{rg_{y}-D}+\mathcal{L}_{rg_{z}}\right)\text{(4)}$$
其中$\mathcal{L}_rg_z$是$z_la$上的识别损失。对于$y$上的识别损失我们分别使用$\mathcal{L}_rg_y-D$和$\mathcal{L}_rg_y-G$作为鉴别器和生成器的损失。我们还将解码器损耗定义为$\mathcal{L}_dec=\mathcal{L}_{G_{SL-GAN}}$ ;将编码器损耗定义为$\mathcal{L}_enc=\mathcal{L}_VAE$
编码器损失$\mathcal{L}_enc$是变分自动编码器的重建误差与潜在分布$z_la$上的先前正则化项之和；因此将其定义为$\mathcal{L}_VAE=\mathcal{L}_{prior}+\mathcal{L}_{recon}$
=
$KL\left(q_{enc}\left(z_{la}\mid x\right)\parallel p_{prior}\left(z_{la}\right)\right)$
; $\mathcal{L} _{prior}$
测量近似后验$q_{enc}(z_{la}\mid x)$与先前 MAX5之间的 KL 散度。重建损失$L_recon=$
通
$-\mathbb{E}_{z\sim q_{enc}(z_{la}|x),y\sim p_{data}(y|x)}\left[\log p_{dec}(x\mid z_{la},y)\right]$
过对 SL-FAS 中的属性进行采样来衡量重建生成图像的损失。这里，$q_{enc}\left(z_{la}\mid x\right)$是由神经网络参数化的$p_{enc}\left(z_{la}\mid x\right)$的近似值，e。g。,编码器
$^{1}$Note that the latent attribute $z_la$ also includes the incompressible noise, which is not explicitly modelled due to less impact to our framework. | [
[
"0.5034",
"0.09216",
"0.892",
"0.09216",
"0.892",
"0.1067",
"0.5034",
"0.1067"
],
[
"0.3489",
"0.1048",
"0.3823",
"0.11237",
"0.3782",
"0.125",
"0.3447",
"0.11743"
],
[
"0.501",
"0.1048",
"0.892",
"0.1061",
"0.892",
"0.1237",
"0.501",
"0.1225"
],
[
"0.5034",
"0.1237",
"0.638",
"0.1225",
"0.638",
"0.1364",
"0.5034",
"0.137"
],
[
"0.523",
"0.1395",
"0.8906",
"0.1395",
"0.8906",
"0.1528",
"0.523",
"0.1528"
],
[
"0.10046",
"0.1497",
"0.1177",
"0.1497",
"0.1177",
"0.1578",
"0.10046",
"0.1578"
],
[
"0.308",
"0.1477",
"0.3284",
"0.1477",
"0.3284",
"0.1547",
"0.308",
"0.1547"
],
[
"0.46",
"0.1572",
"0.4722",
"0.1572",
"0.4722",
"0.1604",
"0.46",
"0.1604"
],
[
"0.5024",
"0.1528",
"0.8896",
"0.1547",
"0.8896",
"0.1686",
"0.5024",
"0.1666"
],
[
"0.5063",
"0.1692",
"0.8906",
"0.1686",
"0.8906",
"0.1818",
"0.5063",
"0.1831"
],
[
"0.3105",
"0.1812",
"0.3228",
"0.1787",
"0.3245",
"0.1831",
"0.312",
"0.1849"
],
[
"0.3245",
"0.1799",
"0.3367",
"0.1799",
"0.3367",
"0.1863",
"0.3245",
"0.1863"
],
[
"0.5024",
"0.1831",
"0.8906",
"0.1843",
"0.8906",
"0.1989",
"0.5024",
"0.1976"
],
[
"0.3228",
"0.1989",
"0.335",
"0.1989",
"0.335",
"0.209",
"0.3228",
"0.209"
],
[
"0.505",
"0.1995",
"0.8896",
"0.1995",
"0.8896",
"0.2134",
"0.505",
"0.2134"
],
[
"0.5034",
"0.2122",
"0.8906",
"0.2134",
"0.8906",
"0.2279",
"0.5034",
"0.2267"
],
[
"0.3145",
"0.2299",
"0.3381",
"0.2299",
"0.3381",
"0.2329",
"0.3145",
"0.2329"
],
[
"0.505",
"0.2299",
"0.584",
"0.2299",
"0.584",
"0.2412",
"0.505",
"0.2412"
],
[
"0.1487",
"0.2456",
"0.392",
"0.2456",
"0.392",
"0.2588",
"0.1487",
"0.2588"
],
[
"0.524",
"0.2456",
"0.8896",
"0.2456",
"0.8896",
"0.2588",
"0.524",
"0.2588"
],
[
"0.5034",
"0.26",
"0.8906",
"0.2588",
"0.8906",
"0.2734",
"0.5034",
"0.2747"
],
[
"0.505",
"0.276",
"0.8896",
"0.276",
"0.8896",
"0.289",
"0.505",
"0.289"
],
[
"0.08905",
"0.286",
"0.468",
"0.286",
"0.468",
"0.3005",
"0.08905",
"0.3005"
],
[
"0.5034",
"0.2903",
"0.8896",
"0.2903",
"0.8896",
"0.3037",
"0.5034",
"0.3037"
],
[
"0.0825",
"0.3005",
"0.415",
"0.3005",
"0.415",
"0.3145",
"0.0825",
"0.3145"
],
[
"0.505",
"0.3062",
"0.892",
"0.3062",
"0.892",
"0.3193",
"0.505",
"0.3193"
],
[
"0.0801",
"0.3213",
"0.468",
"0.3208",
"0.468",
"0.3352",
"0.0801",
"0.336"
],
[
"0.5034",
"0.3208",
"0.634",
"0.3208",
"0.634",
"0.334",
"0.5034",
"0.334"
],
[
"0.0809",
"0.3372",
"0.4673",
"0.3372",
"0.4673",
"0.3516",
"0.0809",
"0.3516"
],
[
"0.5254",
"0.336",
"0.8906",
"0.336",
"0.8906",
"0.3499",
"0.5254",
"0.3499"
],
[
"0.0809",
"0.3528",
"0.468",
"0.3528",
"0.468",
"0.3662",
"0.0809",
"0.3662"
],
[
"0.505",
"0.3516",
"0.8906",
"0.3516",
"0.8906",
"0.3655",
"0.505",
"0.3655"
],
[
"0.08417",
"0.3694",
"0.2345",
"0.3694",
"0.2345",
"0.3801",
"0.08417",
"0.3801"
],
[
"0.505",
"0.3662",
"0.8906",
"0.3662",
"0.8906",
"0.3806",
"0.505",
"0.3806"
],
[
"0.11273",
"0.3833",
"0.4355",
"0.3833",
"0.4355",
"0.401",
"0.11273",
"0.401"
],
[
"0.5034",
"0.3818",
"0.8896",
"0.3818",
"0.8896",
"0.3958",
"0.5034",
"0.3958"
],
[
"0.5034",
"0.3965",
"0.5557",
"0.3965",
"0.5557",
"0.411",
"0.5034",
"0.411"
],
[
"0.0825",
"0.4077",
"0.4673",
"0.4077",
"0.4673",
"0.4216",
"0.0825",
"0.4216"
],
[
"0.0784",
"0.4204",
"0.4714",
"0.4216",
"0.4714",
"0.4395",
"0.0784",
"0.4382"
],
[
"0.5024",
"0.4236",
"0.6807",
"0.4236",
"0.6807",
"0.4375",
"0.5024",
"0.4375"
],
[
"0.0784",
"0.4363",
"0.468",
"0.4375",
"0.468",
"0.452",
"0.0784",
"0.4507"
],
[
"0.0809",
"0.4526",
"0.1969",
"0.4526",
"0.1969",
"0.4666",
"0.0809",
"0.4666"
],
[
"0.523",
"0.4475",
"0.8906",
"0.4475",
"0.8906",
"0.461",
"0.523",
"0.461"
],
[
"0.5034",
"0.4634",
"0.8896",
"0.4634",
"0.8896",
"0.4766",
"0.5034",
"0.4766"
],
[
"0.0801",
"0.473",
"0.468",
"0.4734",
"0.468",
"0.488",
"0.0801",
"0.4873"
],
[
"0.505",
"0.4778",
"0.8906",
"0.4778",
"0.8906",
"0.4912",
"0.505",
"0.4912"
],
[
"0.0809",
"0.4893",
"0.468",
"0.4893",
"0.468",
"0.5024",
"0.0809",
"0.5024"
],
[
"0.505",
"0.4937",
"0.8906",
"0.4937",
"0.8906",
"0.507",
"0.505",
"0.507"
],
[
"0.0809",
"0.505",
"0.4673",
"0.505",
"0.4673",
"0.518",
"0.0809",
"0.518"
],
[
"0.5034",
"0.5083",
"0.8906",
"0.5083",
"0.8906",
"0.5215",
"0.5034",
"0.5215"
],
[
"0.0809",
"0.521",
"0.4656",
"0.521",
"0.4656",
"0.534",
"0.0809",
"0.534"
],
[
"0.505",
"0.5234",
"0.892",
"0.5234",
"0.892",
"0.537",
"0.505",
"0.537"
],
[
"0.0809",
"0.535",
"0.4673",
"0.535",
"0.4673",
"0.55",
"0.0809",
"0.55"
],
[
"0.5034",
"0.5386",
"0.831",
"0.5386",
"0.831",
"0.553",
"0.5034",
"0.553"
],
[
"0.0809",
"0.5503",
"0.4673",
"0.5503",
"0.4673",
"0.565",
"0.0809",
"0.565"
],
[
"0.08496",
"0.565",
"0.4673",
"0.565",
"0.4673",
"0.58",
"0.08496",
"0.58"
],
[
"0.517",
"0.565",
"0.8594",
"0.563",
"0.8594",
"0.584",
"0.517",
"0.586"
],
[
"0.1887",
"0.579",
"0.468",
"0.5776",
"0.468",
"0.595",
"0.1887",
"0.5967"
],
[
"0.851",
"0.577",
"0.869",
"0.577",
"0.869",
"0.5806",
"0.851",
"0.5806"
],
[
"0.869",
"0.584",
"0.8896",
"0.584",
"0.8896",
"0.5986",
"0.869",
"0.5986"
],
[
"0.0825",
"0.5947",
"0.4673",
"0.5947",
"0.4673",
"0.6094",
"0.0825",
"0.6094"
],
[
"0.0825",
"0.5806",
"0.1462",
"0.579",
"0.1462",
"0.5913",
"0.0825",
"0.5933"
],
[
"0.5195",
"0.604",
"0.7754",
"0.604",
"0.7754",
"0.6187",
"0.5195",
"0.6187"
],
[
"0.8667",
"0.6016",
"0.8936",
"0.6016",
"0.8936",
"0.619",
"0.8667",
"0.619"
],
[
"0.0825",
"0.61",
"0.468",
"0.61",
"0.468",
"0.6235",
"0.0825",
"0.6235"
],
[
"0.0809",
"0.6255",
"0.3708",
"0.6255",
"0.3708",
"0.6396",
"0.0809",
"0.6396"
],
[
"0.5034",
"0.632",
"0.886",
"0.632",
"0.886",
"0.646",
"0.5034",
"0.646"
],
[
"0.1405",
"0.6416",
"0.407",
"0.6416",
"0.407",
"0.656",
"0.1405",
"0.656"
],
[
"0.4468",
"0.6426",
"0.4673",
"0.6426",
"0.4673",
"0.656",
"0.4468",
"0.656"
],
[
"0.5034",
"0.6465",
"0.8906",
"0.6465",
"0.8906",
"0.664",
"0.5034",
"0.664"
],
[
"0.501",
"0.659",
"0.8896",
"0.6636",
"0.8896",
"0.681",
"0.501",
"0.677"
],
[
"0.0809",
"0.6655",
"0.468",
"0.6655",
"0.468",
"0.6787",
"0.0809",
"0.6787"
],
[
"0.0825",
"0.681",
"0.3594",
"0.681",
"0.3594",
"0.6943",
"0.0825",
"0.6943"
],
[
"0.5034",
"0.677",
"0.7695",
"0.6787",
"0.7695",
"0.6934",
"0.5034",
"0.6914"
],
[
"0.0809",
"0.7007",
"0.4673",
"0.7007",
"0.4673",
"0.714",
"0.0809",
"0.714"
],
[
"0.5024",
"0.6997",
"0.8906",
"0.7007",
"0.8906",
"0.7153",
"0.5024",
"0.714"
],
[
"0.08417",
"0.7173",
"0.4673",
"0.7173",
"0.4673",
"0.731",
"0.08417",
"0.731"
],
[
"0.505",
"0.7153",
"0.8896",
"0.7153",
"0.8896",
"0.73",
"0.505",
"0.73"
],
[
"0.0801",
"0.732",
"0.4673",
"0.731",
"0.4673",
"0.7456",
"0.0801",
"0.747"
],
[
"0.501",
"0.728",
"0.8906",
"0.73",
"0.8906",
"0.7476",
"0.501",
"0.7456"
],
[
"0.0784",
"0.744",
"0.4656",
"0.7456",
"0.4656",
"0.763",
"0.0784",
"0.762"
],
[
"0.505",
"0.744",
"0.55",
"0.7476",
"0.5483",
"0.7603",
"0.5034",
"0.757"
],
[
"0.664",
"0.7437",
"0.8906",
"0.744",
"0.8906",
"0.762",
"0.664",
"0.761"
],
[
"0.613",
"0.7505",
"0.625",
"0.7505",
"0.625",
"0.754",
"0.613",
"0.754"
],
[
"0.0809",
"0.761",
"0.4656",
"0.761",
"0.4656",
"0.776",
"0.0809",
"0.776"
],
[
"0.501",
"0.759",
"0.8936",
"0.7583",
"0.8936",
"0.776",
"0.501",
"0.7764"
],
[
"0.0825",
"0.778",
"0.468",
"0.778",
"0.468",
"0.7925",
"0.0825",
"0.7925"
],
[
"0.501",
"0.7734",
"0.857",
"0.776",
"0.857",
"0.7935",
"0.501",
"0.7915"
],
[
"0.0809",
"0.7925",
"0.4673",
"0.7915",
"0.4673",
"0.806",
"0.0809",
"0.8066"
],
[
"0.5073",
"0.7905",
"0.5557",
"0.7944",
"0.5522",
"0.806",
"0.5063",
"0.8003"
],
[
"0.546",
"0.7925",
"0.8325",
"0.7915",
"0.8325",
"0.806",
"0.546",
"0.8066"
],
[
"0.869",
"0.7935",
"0.8896",
"0.7935",
"0.8896",
"0.8037",
"0.869",
"0.8037"
],
[
"0.0809",
"0.808",
"0.4673",
"0.808",
"0.4673",
"0.8213",
"0.0809",
"0.8213"
],
[
"0.5034",
"0.8047",
"0.8906",
"0.806",
"0.8906",
"0.821",
"0.5034",
"0.8193"
],
[
"0.0809",
"0.8213",
"0.353",
"0.8228",
"0.353",
"0.837",
"0.0809",
"0.836"
],
[
"0.5024",
"0.821",
"0.8906",
"0.8213",
"0.8906",
"0.836",
"0.5024",
"0.8354"
],
[
"0.5034",
"0.837",
"0.7837",
"0.837",
"0.7837",
"0.8516",
"0.5034",
"0.8516"
],
[
"0.0801",
"0.8486",
"0.3896",
"0.8486",
"0.3896",
"0.8623",
"0.0801",
"0.8623"
],
[
"0.5195",
"0.863",
"0.888",
"0.863",
"0.888",
"0.8774",
"0.5195",
"0.8774"
],
[
"0.0989",
"0.872",
"0.4673",
"0.872",
"0.4673",
"0.8853",
"0.0989",
"0.8853"
],
[
"0.5034",
"0.8755",
"0.8906",
"0.8755",
"0.8906",
"0.8896",
"0.5034",
"0.8896"
],
[
"0.0809",
"0.887",
"0.4697",
"0.887",
"0.4697",
"0.901",
"0.0809",
"0.901"
],
[
"0.5034",
"0.8896",
"0.6445",
"0.8896",
"0.6445",
"0.8994",
"0.5034",
"0.8994"
]
] | [
"e。y～Pdata(y[a);(2)Latent attributeszla表示应",
"GAN",
"该以数据驱动的方式从数据1中挖掘的属性，i。e。，",
"Zla~pdata (zla/r)",
"从数学上讲，y和zla可以是单变量或多变量；并且",
"X",
"MX",
"GF",
"这些属性是相互独立的，i。e。，p(y,zla）=p(y)p(zla)",
"。y的每个维度都用真实图像中注释的一种面部属性进",
"",
"",
"行裁剪；我们的SL-GAN以有监督的方式训练y。相",
"x",
"比之下，za的每个维度都是完全无监督训练的。我们",
"定义了半潜在的面部属性空间，以组合9属性和潜在",
"",
"的z属性",
"Figure 2. Overview of our SL-GAN",
"随着分解的输人噪声，生成器的形式现在是",
"G（zla,y）。直接从输人数据中学习zla将导致生成",
"器倾向于忽略pG（α「zla,y）=pG(α「y)的潜在属性的",
"噪声分布）中绘制样本～pprior（z）=(0,1），然",
"琐碎解决方案。相反，我们最大化属性zla和生成器分",
"后使用生成器网络G，i.e：G(z)转换该样本",
"布G（zla，y）之间的互信息，可以简化为最小化属性y",
"InfoGAN[4)通过最大化互信息I（y;G（z,y）），将GAN",
"和zla 的识别损失",
"的输入噪声分解为潜在表示9和不可压缩噪声z，并",
"共同学习属性y和zla很重要；并确保zla可以表示",
"确保在生成过程中不丢失潜在表示的信息。互信息项",
"输人面部图像的未建模方面，而不是重新发现y。“重",
"可以描述为识别损失，",
"新发现”意味着zla的某些维度与y在输入图像i.e.上",
"Crg = -E~G(z,) [Ey~pdata(yl/r) [log Q(y / r)]",
"的分布非常相似，y中的相同模式从潜在属性 zla中反",
"复发现",
"其中Q(y|a）是后验pdata（y|r）的近似值。因此，生",
"成器G的参数可以更新为LGinJeGAN=LG—Lrg。",
"3.3. Semi-Latent GAN",
"InfoGAN可以以完全无监督的方式学习解开，可解释",
"和有意义的表示",
"我们的SL-GAN如图所示。2；它由三个部分组成，",
"即编码器-解码网络，GAN和识别网络。用户定义的",
"VAEGANE[201将VAE合并到GAN中，并在数据空",
"和潜在的属性由编码器-解码器网络进行编码和解码。",
"间中将GAN的元素错误替换为VAEGAN的特征错",
"识别网络有助于从数据中学习SL-FAS。在我们的SL-",
"误。具体来说，它将数据样本编码为潜在表示：z~",
"GAN中，识别网络和鉴别器共享相同的网络结构，并",
"Enc（a）=Penc（|α），并将解码回数据空间：~",
"且在最后一层具有不同的softmax层。因此，生成器",
"Dec(a|z）=Pdec(α|2）。我们可以将正则化先验的",
"GSL-GAN和鉴别器DsL-GAN的损失函数为",
"损失函数定义为Lprior=KL(qenc（|）lPprior(z))",
"qenc（za）是真实后验pdec（za）的近似值。重建错",
"CGsL-GAN = CG +入1 (Crg + Lrgu-G)+A2CDe",
"-Eqene(zl) [logpdee(Di (a)|2)]，其中",
"DcO",
"(3)",
"Di(c）是鉴别器的i-th层的隐藏表示。因此，这种损",
"误为CD",
"CDsL-GAN=CD+(Crgy-D+Crg)",
"(4)",
"失使1-th鉴别器层的数据表示的预期对数似然之和",
"最小化。因此，VAEGAN的损失函数为",
"其中Crg-是zlg上的识别损失。对于y上的识别损失",
"CVAEGAN=LCGAN+CDecon+Cprior",
"(2)",
"我们分别使用Crgy-D和Crgu-G作为鉴别器和生成器",
"的损失。我们还将解码器损耗定义为Ldec=LGsL-GAN",
"然而，潜在表示之是完全无监督的；没有办法明确控制",
"数据上的属性或使用属性修改面部图像",
"；将编码器损耗定义为Cenc=CvAE",
"CVAE[42.35]是条件VAE。引入独立属性变量y",
"编码器损失Cene是变分自动编码器的重建",
"，通过从p(α|y,）采样来控制的生成过程；其中",
"误差与潜在分布zla上的先前正则化项之",
"p(y,z)=p(y)p(z)。因此，CVAE的编码器和解码器",
"和；因此将其定义为CvAE=Lprior+Crecon",
"网络是～ Enc()= Qene(zα) 和～ Dec(z,y) =",
":Cpri",
"KL(qene (zla / a) l Pprior (zla)",
"-",
"pdec(αz,y）。引人变量y，通过从p(αly,z）采样",
"测量近似后验ene（zla|a）与先前MAX5",
"来控制a的生成过程；其中p(y,z)=p(y)p(z)。尽",
"之间的KL散度。重建损失Lrecon",
"管如此，9仍然是从数据中采样的，而不是像我们的",
"E~g",
"qene(ztalz),y~pdata(ylz) [log Pdec ( [ zla,y)]",
"通",
"SL-GAN那样直接从数据中优化和学习。因此，y可用",
"过对SL-FAS中的属性进行采样来衡量重建生成图像",
"于修改类似于建议的SL-GAN的属性",
"的损失。这里，qenc（zla|a）是由神经网络参数化的",
"Penc（zla「a）的近似值，e。g。，编码器",
"3.2. Semi-Latent Facial Attribute Space",
"INote that the latent attribute zla also includes the incom",
"GAN的输人噪声可以进一步分解为两部分：（1）用",
"pressible noise, which is not explicitly modelled due to less im-",
"户定义的E属性Ey是每个图像的手动注释属性，i。",
"pact to our framework."
] | {
"arxiv_id": "1704.02166",
"language": "zh",
"page_idx": 2
} |
||
1704.02166-zh-3 | 1704.02166-zh-3.jpg | $z_{la}$的识别损失是在生成数据和真实数据上训练的。它旨在预测潜在属性的值。假设潜在的$z_la\sim p_{enc}(z\mid x)$ 是从编码器网络的分布中采样的。生成器和鉴别器的更新步骤在$z_la$上使用相同的识别损失，定义为
$$\begin{aligned}\mathcal{L}_{rg_{z}}&=-\mathbb{E}_{x\sim p_{data}(x),z\sim p_{enc}(z_{la}|x)}\left[\log\left(Q\left(z\mid x\right)\right)\right]\quad(5)\\&-\mathbb{E}_{x\sim p_{dec}(x|z,y),y\sim p_{data}(y|x),z\sim p_{enc}(z_{la}|x)}\left[\log\left(Q\left(z\mid x\right)\right)\right]\\&-\mathbb{E}_{x\sim p_{dec}(x|y,z),y\sim p_{data}(y),z\sim p_{prior}(z_{la})}\left[\log\left(Q\left(z\mid x\right)\right)\right]\\&-\mathbb{E}_{x\sim p_{dec}(x|z,y),y\sim p_{data}(y),z\sim p_{enc}(z_{la}|x)}\left[\log\left(Q\left(z\mid x\right)\right)\right],\end{aligned}$$
,其中$-\mathbb{E}_x\sim p_{data}(x),z\sim p_{enc}(z_{la}|x)\left[\log\left(Q\left(z\mid x\right)\right)\right]$测量真实数据上预测错误的损失；其余三项是生成数据上的损失函数。$Q(\cdot)$仍然是相应后验数据分布的近似值。$p_{enc}\left(z_{la}\mid x\right)$是$z_la\sim p_{enc}\left(z\mid x\right)$的分布， $z_{la}~\sim~p_{enc}\left(z\mid x\right)$的分布是 $z_la~\sim~p_{enc}\left(z\mid x\right)$的分布， $z_{la}$是$z_la~\sim~p_{enc}(z\mid x)$的分布$;~z_la~\sim~p_{enc}(z\mid x)$ 是$z_la\sim p_{enc}\left(z\mid x\right)$的分布$,z_la\sim p_{enc}\left(z\mid x\right)$是$z_{la}\sim p_{enc}\left(z\mid x\right)$的分布$,p_data\left(x\right)$是 XMATHX 的数据分布。
$y$的识别损失可以在真实数据和生成数据上进行训练。在鉴别器$D_{SL-GAN}$的训练阶段，我们有用户定义的$y$的地面实况属性注释，而不使用生成的数据。原因是生成数据的质量依赖于生成器$G_{SL-GAN}$的训练质量， 该质量也是从真实数据中训练出来的；因此，如果数据是由未经良好训练的生成器生成的，我们将观察到“语义漂移”现象。为此，只有手动标记的属性$y$可以用于更新$D_{SL-GAN}$ ;并且更新$D$的损失是
$$\mathcal{L}_{rg_{y}-D}=-\mathbb{E}_{x\sim p_{data}(x),y\sim p_{data}(y|x)}\left[\log\left(Q\left(y\mid x\right)\right)\right]$$
$G_{SL-GAN}$对应于解码器网络，因此只能使用生成
的数据来计算发电机的损耗，
$$\begin{aligned}&\mathcal{L}_{rg_{y}-G}=\\&-\mathbb{E}_{x\sim p_{dec}(x|y,z),z\sim p_{enc}(z|x),y\sim p_{data}(y)}\left[\log\left(Q\left(y\mid x\right)\right)\right]\\&-\mathbb{E}_{x\sim p_{dec}(x|y,z),z\sim p_{enc}(z|x),y\sim p_{data}(y|x)}\left[\log\left(Q\left(y\mid x\right)\right)\right]\\&-\mathbb{E}_{x\sim p_{dec}(x|y,z),z\sim p_{prior}(z),y\sim p_{data}(y|x)}\left[\log\left(Q\left(y\mid x\right)\right)\right]\end{aligned}$$
请注意，本质上，Eq(5)和(7)中的每一项都可以
由系数加权。在这里，为了便于表示，我们省略了这些
系数
## 3.4. The training algorithms
我们的 SL-GAN 旨在通过属性解决面部图像的生成和修改。关键要素是学习在我们的 SL-GAN 框架中统一的数据驱动属性$z_la$和用户定义属性$y$的分离表示。我们的 SL-GAN 由三个关键组件组成，旨在解决两个相关但非常不同的问题-生成和修改面部图像。训练GAN 的传统方式在我们的问题中不起作用。因此，我们提出了一种新的训练算法来训练我们的 SL-GAN。具体来说，我们算法的一次迭代需要完成以下三个阶段，
学习面部图像重建。此阶段主要更新编码器-解码器网络，并通过以下步骤学习重建给定相应用户定义属性的图像：
$\bullet$ Sampling a batch of images $x\sim p_data(x)$ and at-
tributes $y\sim p_{data}( y| x)$ , $z_{la}\sim p_{enc}( z\mid x)$
$\bullet$ Updating the SL-GAN by minimizing the en-
coder $\mathcal{L} _enc$ , Decoder $\mathcal{L}_dec$ and Discriminator with
$\mathcal{L}_{D_{SL-GAN}}$ iteratively.
学习修改面部图像。我们从所有数据中采样图像$x$和属性$y$ 。此阶段训练 SL-GAN 能够通过提供$y$来修改映像$x$ 。请注意，映像$x$不一定具有属性$y$ 。这里的另一个重要问题是在修改面部图像时如何保持采样图像的相同身份；这里采用了两种策略：首先，$z_la$是从编码器网络参数化的$p_{enc}\left(z\mid x\right)$中采样的，该网络在此子步骤中未更新；第二，我们的 SL-GAN 最小化$\mathcal{L}_{recon}^{D_l}$本质上是为了保证同一个人的身份。因此，此步骤鼓励生成器学习$y$和$z_la$的解开表示。
$\bullet$ Learning to modify the attributes: Sample a batch of images $x\sim p_{data}(x)$ and the attribute $y\sim$ $p_{data}( y)$ , $z_{la}\sim p_{enc}( z\mid x)$ ;
$\bullet$ Update the SL-GAN by minimizing the Decoder $\begin{array}{ll}{{\mathcal L}_{dec}}&{\mathrm{and}}&{\mathrm{Discriminator}}&{\mathrm{with}}&{{\mathcal L}_{D_{SL-GAN}}}&{\mathrm{itera-}}\\\end{array}$ tively;
学习生成面部图像。我们从$z_{la}$的先验分布中采样，从
$y$的数据分布中采样，i。e。
$\bullet$ Sample a batch of latent vectors $z_la\sim p_prior(z)$
and attribute vectors $y\sim p_data( y)$ .
$\bullet$ Update the SL-GAN by minimizing the Decoder $\mathcal{L}_{dec}$and Discriminator with $\mathcal{L}_{D_{SL-GAN}}$ iteratively;
一旦训练了网络，我们就可以解决生成和修改的任务；
特别是
·Generating new facial images with any attributes.
This can be achieved by sampling $z_la$ from $p_{prior}\left(z\right)$and setting $y$ to any desired attributes. We can then get from the generator the image $x^{\prime}\sim G\left(z_{la},y\right).$
$\bullet$ Modifying the existing images with any attributes.
Given an image $x$ and the desired attribute $y$,we can sample $z_{la}\sim p\left(z\mid x\right).$Then the modified image can be generated by $x^\prime\sim p_{dec}\left(x\mid z_{la},y\right).$
4. Experiments
# 4.1. Experimental settings
数据集。我们在两个数据集上进行实验。CelebA 数据集[22]包含 10$k$标识的大约 200$k$图像。每个图像都 | [
[
"0.0825",
"0.09216",
"0.468",
"0.09216",
"0.468",
"0.1067",
"0.0825",
"0.1067"
],
[
"0.505",
"0.09216",
"0.8896",
"0.09216",
"0.8896",
"0.1061",
"0.505",
"0.1061"
],
[
"0.08417",
"0.1092",
"0.4656",
"0.1092",
"0.4656",
"0.1225",
"0.08417",
"0.1225"
],
[
"0.5034",
"0.108",
"0.8906",
"0.108",
"0.8906",
"0.1218",
"0.5034",
"0.1218"
],
[
"0.08417",
"0.125",
"0.4673",
"0.125",
"0.4673",
"0.1351",
"0.08417",
"0.1351"
],
[
"0.505",
"0.125",
"0.5596",
"0.125",
"0.5596",
"0.1364",
"0.505",
"0.1364"
],
[
"0.0809",
"0.1395",
"0.423",
"0.1395",
"0.423",
"0.1528",
"0.0809",
"0.1528"
],
[
"0.5146",
"0.1445",
"0.8906",
"0.1459",
"0.8906",
"0.1635",
"0.5146",
"0.1622"
],
[
"0.0801",
"0.1642",
"0.4714",
"0.1642",
"0.4714",
"0.1849",
"0.0801",
"0.1849"
],
[
"0.5337",
"0.161",
"0.831",
"0.1622",
"0.831",
"0.1768",
"0.5337",
"0.1755"
],
[
"0.0948",
"0.1818",
"0.4795",
"0.1831",
"0.4795",
"0.2039",
"0.0948",
"0.2026"
],
[
"0.5146",
"0.1849",
"0.88",
"0.1849",
"0.88",
"0.1989",
"0.5146",
"0.1989"
],
[
"0.09644",
"0.2008",
"0.4673",
"0.202",
"0.4673",
"0.2229",
"0.09644",
"0.2216"
],
[
"0.536",
"0.202",
"0.892",
"0.202",
"0.892",
"0.2153",
"0.536",
"0.2153"
],
[
"0.536",
"0.2166",
"0.6914",
"0.2166",
"0.6914",
"0.2311",
"0.536",
"0.2311"
],
[
"0.0972",
"0.2229",
"0.4756",
"0.2229",
"0.4756",
"0.2437",
"0.0972",
"0.2437"
],
[
"0.5034",
"0.2393",
"0.8906",
"0.2412",
"0.8906",
"0.2551",
"0.5034",
"0.2524"
],
[
"0.0948",
"0.2524",
"0.1854",
"0.2524",
"0.1854",
"0.2644",
"0.0948",
"0.2644"
],
[
"0.1969",
"0.254",
"0.4697",
"0.2488",
"0.4697",
"0.2651",
"0.1969",
"0.2708"
],
[
"0.505",
"0.2551",
"0.8906",
"0.2551",
"0.8906",
"0.2695",
"0.505",
"0.2695"
],
[
"0.0825",
"0.2676",
"0.468",
"0.2676",
"0.468",
"0.281",
"0.0825",
"0.281"
],
[
"0.08826",
"0.2568",
"0.0972",
"0.2568",
"0.0972",
"0.262",
"0.08826",
"0.262"
],
[
"0.5034",
"0.2708",
"0.8906",
"0.2708",
"0.8906",
"0.2842",
"0.5034",
"0.2842"
],
[
"0.08417",
"0.2827",
"0.468",
"0.2827",
"0.468",
"0.2966",
"0.08417",
"0.2966"
],
[
"0.5034",
"0.286",
"0.8906",
"0.286",
"0.8906",
"0.2998",
"0.5034",
"0.2998"
],
[
"0.0809",
"0.2974",
"0.4714",
"0.2974",
"0.4714",
"0.3125",
"0.0809",
"0.3125"
],
[
"0.505",
"0.3018",
"0.8906",
"0.3018",
"0.8906",
"0.315",
"0.505",
"0.315"
],
[
"0.0809",
"0.3132",
"0.4697",
"0.3132",
"0.4697",
"0.327",
"0.0809",
"0.327"
],
[
"0.5034",
"0.3164",
"0.8906",
"0.315",
"0.8906",
"0.33",
"0.5034",
"0.3308"
],
[
"0.0825",
"0.3276",
"0.468",
"0.3276",
"0.468",
"0.3423",
"0.0825",
"0.3423"
],
[
"0.5034",
"0.33",
"0.8906",
"0.33",
"0.8906",
"0.3447",
"0.5034",
"0.3447"
],
[
"0.0809",
"0.3435",
"0.468",
"0.3435",
"0.468",
"0.358",
"0.0809",
"0.358"
],
[
"0.505",
"0.3452",
"0.8896",
"0.3452",
"0.8896",
"0.3591",
"0.505",
"0.3591"
],
[
"0.0784",
"0.3567",
"0.4697",
"0.356",
"0.4697",
"0.3738",
"0.0784",
"0.375"
],
[
"0.5024",
"0.3599",
"0.7915",
"0.361",
"0.7915",
"0.3757",
"0.5024",
"0.375"
],
[
"0.0825",
"0.375",
"0.1528",
"0.375",
"0.1528",
"0.3865",
"0.0825",
"0.3865"
],
[
"0.517",
"0.385",
"0.8906",
"0.385",
"0.8906",
"0.3984",
"0.517",
"0.3984"
],
[
"0.536",
"0.401",
"0.888",
"0.401",
"0.888",
"0.4153",
"0.536",
"0.4153"
],
[
"0.0801",
"0.4104",
"0.4673",
"0.4104",
"0.4673",
"0.4236",
"0.0801",
"0.4236"
],
[
"0.5337",
"0.4167",
"0.739",
"0.4153",
"0.739",
"0.43",
"0.5337",
"0.4312"
],
[
"0.0809",
"0.4255",
"0.4656",
"0.4255",
"0.4656",
"0.4395",
"0.0809",
"0.4395"
],
[
"0.0809",
"0.44",
"0.468",
"0.44",
"0.468",
"0.4539",
"0.0809",
"0.4539"
],
[
"0.5186",
"0.44",
"0.8896",
"0.44",
"0.8896",
"0.4539",
"0.5186",
"0.4539"
],
[
"0.0809",
"0.4558",
"0.468",
"0.4558",
"0.468",
"0.4697",
"0.0809",
"0.4697"
],
[
"0.5337",
"0.4526",
"0.8896",
"0.455",
"0.8896",
"0.473",
"0.5337",
"0.4702"
],
[
"0.0809",
"0.4702",
"0.468",
"0.4702",
"0.468",
"0.4841",
"0.0809",
"0.4841"
],
[
"0.536",
"0.4717",
"0.581",
"0.4717",
"0.581",
"0.483",
"0.536",
"0.483"
],
[
"0.08417",
"0.4873",
"0.4656",
"0.4873",
"0.4656",
"0.4976",
"0.08417",
"0.4976"
],
[
"0.505",
"0.4937",
"0.8906",
"0.4937",
"0.8906",
"0.507",
"0.505",
"0.507"
],
[
"0.0825",
"0.5",
"0.4673",
"0.5005",
"0.4673",
"0.515",
"0.0825",
"0.5146"
],
[
"0.5024",
"0.5083",
"0.701",
"0.509",
"0.701",
"0.5234",
"0.5024",
"0.523"
],
[
"0.0825",
"0.5166",
"0.3635",
"0.5166",
"0.3635",
"0.531",
"0.0825",
"0.531"
],
[
"0.5146",
"0.53",
"0.8906",
"0.532",
"0.8906",
"0.55",
"0.5146",
"0.5474"
],
[
"0.08826",
"0.5425",
"0.4697",
"0.5425",
"0.4697",
"0.563",
"0.08826",
"0.563"
],
[
"0.5337",
"0.5474",
"0.79",
"0.55",
"0.79",
"0.563",
"0.5337",
"0.561"
],
[
"0.1013",
"0.5713",
"0.468",
"0.5713",
"0.468",
"0.586",
"0.1013",
"0.586"
],
[
"0.517",
"0.5713",
"0.8906",
"0.5713",
"0.8906",
"0.586",
"0.517",
"0.586"
],
[
"0.0809",
"0.587",
"0.2852",
"0.587",
"0.2852",
"0.601",
"0.0809",
"0.601"
],
[
"0.5337",
"0.5854",
"0.892",
"0.587",
"0.892",
"0.605",
"0.5337",
"0.603"
],
[
"0.535",
"0.601",
"0.5825",
"0.603",
"0.581",
"0.617",
"0.5337",
"0.6157"
],
[
"0.2385",
"0.6255",
"0.3064",
"0.629",
"0.3057",
"0.6445",
"0.2378",
"0.6416"
],
[
"0.517",
"0.627",
"0.892",
"0.627",
"0.892",
"0.64",
"0.517",
"0.64"
],
[
"0.0948",
"0.64",
"0.455",
"0.64",
"0.455",
"0.661",
"0.0948",
"0.661"
],
[
"0.5024",
"0.6426",
"0.5557",
"0.6426",
"0.5557",
"0.657",
"0.5024",
"0.657"
],
[
"0.08905",
"0.657",
"0.4575",
"0.657",
"0.4575",
"0.675",
"0.08905",
"0.675"
],
[
"0.08905",
"0.6704",
"0.4592",
"0.672",
"0.4592",
"0.6895",
"0.08905",
"0.688"
],
[
"0.517",
"0.6655",
"0.8857",
"0.6655",
"0.8857",
"0.6787",
"0.517",
"0.6787"
],
[
"0.5337",
"0.6787",
"0.8906",
"0.6797",
"0.8906",
"0.6943",
"0.5337",
"0.6934"
],
[
"0.4468",
"0.688",
"0.4673",
"0.688",
"0.4673",
"0.7007",
"0.4468",
"0.7007"
],
[
"0.5337",
"0.6943",
"0.8906",
"0.6924",
"0.8906",
"0.7104",
"0.5337",
"0.712"
],
[
"0.0825",
"0.7026",
"0.4673",
"0.7026",
"0.4673",
"0.7163",
"0.0825",
"0.7163"
],
[
"0.536",
"0.711",
"0.8906",
"0.711",
"0.8906",
"0.7246",
"0.536",
"0.7246"
],
[
"0.0825",
"0.7173",
"0.4673",
"0.7173",
"0.4673",
"0.731",
"0.0825",
"0.731"
],
[
"0.5337",
"0.7236",
"0.648",
"0.728",
"0.648",
"0.742",
"0.5337",
"0.738"
],
[
"0.0809",
"0.7344",
"0.1152",
"0.7344",
"0.1152",
"0.747",
"0.0809",
"0.747"
],
[
"0.517",
"0.7505",
"0.888",
"0.7505",
"0.888",
"0.7646",
"0.517",
"0.7646"
],
[
"0.0801",
"0.7583",
"0.3022",
"0.7583",
"0.3022",
"0.7715",
"0.0801",
"0.7715"
],
[
"0.536",
"0.7666",
"0.888",
"0.7666",
"0.888",
"0.78",
"0.536",
"0.78"
],
[
"0.10297",
"0.7827",
"0.4673",
"0.7827",
"0.4673",
"0.7935",
"0.10297",
"0.7935"
],
[
"0.535",
"0.781",
"0.892",
"0.78",
"0.892",
"0.7944",
"0.535",
"0.7954"
],
[
"0.0809",
"0.797",
"0.468",
"0.797",
"0.468",
"0.81",
"0.0809",
"0.81"
],
[
"0.5337",
"0.7954",
"0.8813",
"0.797",
"0.8813",
"0.811",
"0.5337",
"0.81"
],
[
"0.0948",
"0.813",
"0.4656",
"0.813",
"0.4656",
"0.8237",
"0.0948",
"0.8237"
],
[
"0.0809",
"0.827",
"0.4656",
"0.827",
"0.4656",
"0.8403",
"0.0809",
"0.8403"
],
[
"0.5024",
"0.8247",
"0.6406",
"0.8247",
"0.6406",
"0.839",
"0.5024",
"0.839"
],
[
"0.08417",
"0.8433",
"0.4673",
"0.8433",
"0.4673",
"0.854",
"0.08417",
"0.854"
],
[
"0.501",
"0.8467",
"0.71",
"0.8477",
"0.71",
"0.8657",
"0.501",
"0.8643"
],
[
"0.0801",
"0.856",
"0.4673",
"0.857",
"0.4673",
"0.8706",
"0.0801",
"0.869"
],
[
"0.0801",
"0.872",
"0.468",
"0.8706",
"0.468",
"0.8853",
"0.0801",
"0.886"
],
[
"0.505",
"0.872",
"0.8906",
"0.872",
"0.8906",
"0.886",
"0.505",
"0.886"
],
[
"0.0809",
"0.887",
"0.464",
"0.887",
"0.464",
"0.901",
"0.0809",
"0.901"
],
[
"0.5034",
"0.887",
"0.892",
"0.887",
"0.892",
"0.9014",
"0.5034",
"0.9014"
]
] | [
"zla的识别损失是在生成数据和真实数据上训练的。它",
"学习面部图像重建。此阶段主要更新编码器-解码器网",
"旨在预测潜在属性的值。假设潜在的zla～pencz|)",
"络，并通过以下步骤学习重建给定相应用户定义属性",
"是从编码器网络的分布中采样的。生成器和鉴别器的",
"的图像：",
"更新步骤在za上使用相同的识别损失，定义为",
". Sampling a batch of images a ~ pdata(a) and at-",
"Lrg= =-Er~pdata(1),z~penc(2ialr) [1og(Q(2 | 2) (5)",
"tributes y ~ pdata(ylr) , zla ~ Pene (z [)",
"-Ex~pde(r[2, ),y~data(/a),z~Pene(zala) [og (Q(2 [)]",
": Updating the SL-GAN by minimizing the el",
"-E~pdee(rly,2),y~pata(u),z~Pprior(2zta) [1og (Q(2 [ )]",
"coder Lenc , Decoder Cdec and Discriminator with",
"CDSL-GAN iteratively.",
"-Ex~pde(r[2,),y~Pdata(u),z~pene(ztalm) [og(Q(≥ [)] ,",
"学习修改面部图像。我们从所有数据中采样图像和",
"其中-E~",
"ata(),z~pene(zia |z) [log (Q (z / ))) 测量",
"属性y。此阶段训练SL-GAN能够通过提供y来修",
"真实数据上预测错误的损失；其余三项是生成数据",
"",
"改映像。请注意，映像不一定具有属性y。这里",
"上的损失函数。Q（）仍然是相应后验数据分布",
"的另一个重要问题是在修改面部图像时如何保持采样",
"的近似值。Penc（zlaa）是zla～penc（z|a）的分布，",
"图像的相同身份；这里采用了两种策略：首先，zla是",
"la～Penc（z|）的分布是zla~Pene（z|）的分布，",
"从编码器网络参数化的penc（z|a）中采样的，该网络",
"zla是zla～penc(|）的分布；zla～Penc(|)",
"在此子步骤中未更新；第二，我们的SL-GAN最小化",
"是zla～Penc(z|)的分布，Zla ～ Penc(z|）是",
"CDicon本质上是为了保证同一个人的身份。因此，此步",
"Zla～ Penc(z|a)的分布，Pdata（α）是XMATHX 的",
"骤鼓励生成器学习y和zla的解开表示。",
"数据分布。",
": Learning to modify the attributes: Sample a batch",
"of images ~ pdata(r) and the attribute y ~",
"y的识别损失可以在真实数据和生成数据上进行训练。",
"Pdata(y) , zia ~ Pene(z/ r) ;",
"在鉴别器DsL-GAN的训练阶段，我们有用户定义的",
"9的地面实况属性注释，而不使用生成的数据。原因是",
". Update the SL-GAN by minimizing the Decoder",
"生成数据的质量依赖于生成器GsL-GAN的训练质量，",
"Cdee and Discriminator with CDsL-GAN itera-",
"该质量也是从真实数据中训练出来的；因此，如果数据",
"tively:",
"是由未经良好训练的生成器生成的，我们将观察到“语",
"学习生成面部图像。我们从zlo的先验分布中采样，从",
"义漂移”现象。为此，只有手动标记的属性9可以用于",
"y的数据分布中采样，i。e。",
"更新DsL-GAN；并且更新D的损失是",
". Sample a batch of latent vectors zla ~ Pprior(z)",
"Crgu-D=-Er~pata(n),y~paata(gla) [1og(Q(y / a)] (6)",
"and attribute vectors y~Pdata(y):",
"GsL-GAN对应于解码器网络，因此只能使用生成",
": Update the SL-GAN by minimizing the Decoder",
"的数据来计算发电机的损耗，",
"Cdec and Discriminator with CDsL-GAN itera-",
"tively;",
"Crgy-G",
"旦训练了网络，我们就可以解决生成和修改的任务；",
"-Ea~pe(ly,2),z~pene(2/r),~pata(u)[log(Q(y [ a)]",
"特别是",
"-Er~pdee(ly,2),z~pene(2[r),y~pdata(ylr) [log (Q (y [ α)]",
"-Er~pdee(aly,2),z~pprior(z),y~pdata(ylm) [log(Q(y / )]",
". Generating new facial images with any attributes",
"This can be achieved by sampling zla from",
"(7)",
"Pprior (z) and setting y to any desired attributes.",
"请注意，本质上，Eq（5）和（7）中的每一项都可以",
"We can then get from the generator the image",
"由系数加权。在这里，为了便于表示，我们省略了这些",
"r ~G(zla,y) :",
"系数",
". Modifying the existing images with any attributes",
"3.4. The training algorithms",
"Given an image and the desired attribute y , we",
"我们的SL-GAN旨在通过属性解决面部图像的生成",
"can sample zla ~ p(z[r)Then the modified",
"和修改。关键要素是学习在我们的SL-GAN框架中统",
"image can be generated by r' ~ pdec (α | zla,y) .",
"的数据驱动属性o和用户定义属性9的分离表示。",
"我们的SL-GAN由三个关键组件组成，旨在解决两个",
"4. Experiments",
"相关但非常不同的问题一生成和修改面部图像。训练",
"4.1. Experimental settings",
"GAN的传统方式在我们的问题中不起作用。因此，我",
"们提出了一种新的训练算法来训练我们的SL-GAN。具",
"数据集。我们在两个数据集上进行实验。CelebA数据",
"体来说，我们算法的一次送代需要完成以下三个阶段，",
"集[22]包含10k标识的大约200k图像。每个图像都"
] | {
"arxiv_id": "1704.02166",
"language": "zh",
"page_idx": 3
} |
||
2203.06462-zh-1 | 2203.06462-zh-1.jpg | ## 低秩 Softmax 在理论上可以具有不可论证的类, 但在实践中很少
Andreas Grivas and Nikolay Bogoychev and Adam Lopez
Institute for Language, Cognition, and Computation
## School of Informatics University of Edinburgh
$\{$ agrivas, n.bogoych, alopez $\}$ ged.ac.uk
## $\mathbf{Abstract}$
Classifers in natural language processing (NLP) often have a large number of output classes. For example, neural language models (LMs) and machine translation (MT) models both predict tokens from a vocabulary of thousands. The Softmax output layer of these models typically receives as input a dense feature representation, which has much lower dimensionality than the output. In theory, the result is some words may be impossible to be predicted via argmax, irrespective of input features, and empirically, there is evidence this happens in small language models (Demeter et al., 2020). In this paper we ask whether it can happen in practical large language models and translation models. To do so, we develop algorithms to detect such unargmaxable tokens in public models. We find that 13 out of 150 models do indeed have such tokens; however, they are very infrequent and unlikely to impact model quality. We release our code so that others can inspect their models. $^{\mathrm{l}}$
### 1 $\textbf{Introduction}$
具有大量输出类的概率多类分类器在 NLPE (Chen et al., 2016) 中很常见。例如，当代 LM 和MT 模型的词汇量从数万到数十万不等 (Liu et al., 2020) 。在对如此大的词汇表进行建模方面的最新进展主要是通过改进神经网络特征编码器来实现的 (Devlin et al., 2019; Conneau et al., 2020)。但是，不管特征编码器的表现力$\overline{\mathrm{E}}($Yun et al., 2020; Raghu et al., 2017) 如何将较低维特征线性映射到较高维输出的分类器都会降低表现力$\operatorname{E}($Yang et al.,2018),其后果尚不清楚
在这项工作中，我们详细说明了将 argmax 预测与低秩分类器一起使用的后果，这些分类器具有比$d$更多的输出类$|C|$的分类器。例如，MT 模型通常具有大小为$|C|\approx30000$的子词词汇表，但具有$d\approx1024$。这种低秩分
$^{1}$ https://github.com/andreasgrv/unargmaxable
<FigureHere>
Figure 1: Illustration of an unargmaxable class. Class $c_{4}$ can never be predicted using argmax for this Softmax classifer with $|C|=4$ classes and $d=2$ input features. On the left, each feature vector $x$ is colored according to the class assigned the largest probability; note that while $c_1,c_2$ and $c_3$ surface as regions, $c_4$ does not. On the right, we show that there is no direction in feature space for which $c_{4}$ has the largest probability.
类器的表达性惩罚是某些输出分布无法表示。Demeter et al. (2020) 在 SoftmaxLM 中发现了这一弱点，表明从理论上讲，某些令牌永远无法为任何输入分配最高概率，因此永远无法作为 argmax 预测产生。$^{2}$我们称这样的令牌为unargmaxable(参见图 1 )
虽然 Demeter et al. (2020) 提出了一种检测不可变令牌的算法，并提供了它们在小型 LM 中存在的证据，但他们提出的算法没有提供任何保证，他们无法测试大型 LM。在本文中，我们问：在实践中使用的大型模型中是否存在不可 argmaxable 令牌？为了回答这个问题，我们开发了明确识别此类令牌的算法。我们测试了7 LM 和 143 MT 模型。其中，只有 13 的 MT 模型表现出不可争辩的令牌，即使在这些情况下，令牌也都是嘈杂且不常见的。我们得出的结论是，尽管低等级 Softmax 的表现力约束可能会产生重要的影响，但大多数从业者无需担心无法确定的令牌。我们为他们提供了新的工具来确认他们自己的模型
${^{2}\text{This problem was also studied by}}$ Cover(1967)and has an interesting history of independent discovery (Smith, 2014) . | [
[
"0.2612",
"0.0932",
"0.739",
"0.0932",
"0.739",
"0.1081",
"0.2612",
"0.1081"
],
[
"0.403",
"0.1134",
"0.594",
"0.1134",
"0.594",
"0.1277",
"0.403",
"0.1277"
],
[
"0.2494",
"0.1455",
"0.7573",
"0.1455",
"0.7573",
"0.1603",
"0.2494",
"0.1603"
],
[
"0.293",
"0.1621",
"0.711",
"0.1621",
"0.711",
"0.177",
"0.293",
"0.177"
],
[
"0.413",
"0.1787",
"0.592",
"0.1787",
"0.592",
"0.1924",
"0.413",
"0.1924"
],
[
"0.4014",
"0.1936",
"0.601",
"0.1948",
"0.601",
"0.209",
"0.4014",
"0.2079"
],
[
"0.2644",
"0.2114",
"0.6406",
"0.212",
"0.6406",
"0.2268",
"0.2644",
"0.2257"
],
[
"0.6323",
"0.212",
"0.75",
"0.212",
"0.75",
"0.2257",
"0.6323",
"0.2257"
],
[
"0.2627",
"0.253",
"0.3418",
"0.253",
"0.3418",
"0.2673",
"0.2627",
"0.2673"
],
[
"0.7188",
"0.2559",
"0.777",
"0.2546",
"0.777",
"0.2588",
"0.7188",
"0.26"
],
[
"0.806",
"0.2546",
"0.8623",
"0.253",
"0.8623",
"0.26",
"0.806",
"0.2625"
],
[
"0.6035",
"0.2654",
"0.659",
"0.2683",
"0.657",
"0.2832",
"0.602",
"0.2798"
],
[
"0.1453",
"0.284",
"0.4575",
"0.284",
"0.4575",
"0.2976",
"0.1453",
"0.2976"
],
[
"0.517",
"0.288",
"0.534",
"0.288",
"0.534",
"0.3374",
"0.517",
"0.3374"
],
[
"0.1453",
"0.2964",
"0.4575",
"0.2964",
"0.4575",
"0.31",
"0.1453",
"0.31"
],
[
"0.1453",
"0.3123",
"0.4592",
"0.3123",
"0.4592",
"0.326",
"0.1453",
"0.326"
],
[
"0.7188",
"0.3123",
"0.7773",
"0.3123",
"0.7773",
"0.3196",
"0.7188",
"0.3196"
],
[
"0.797",
"0.3088",
"0.869",
"0.3113",
"0.869",
"0.3218",
"0.797",
"0.3196"
],
[
"0.553",
"0.3237",
"0.602",
"0.326",
"0.601",
"0.3374",
"0.552",
"0.335"
],
[
"0.636",
"0.3237",
"0.69",
"0.326",
"0.69",
"0.3374",
"0.636",
"0.335"
],
[
"0.147",
"0.3425",
"0.4575",
"0.3425",
"0.4575",
"0.3528",
"0.147",
"0.3528"
],
[
"0.147",
"0.328",
"0.4575",
"0.328",
"0.4575",
"0.3384",
"0.147",
"0.3384"
],
[
"0.1453",
"0.3552",
"0.4575",
"0.3552",
"0.4575",
"0.3682",
"0.1453",
"0.3682"
],
[
"0.147",
"0.3706",
"0.4575",
"0.3706",
"0.4575",
"0.3813",
"0.147",
"0.3813"
],
[
"0.5845",
"0.3718",
"0.65",
"0.3718",
"0.65",
"0.383",
"0.5845",
"0.383"
],
[
"0.1453",
"0.383",
"0.4575",
"0.3818",
"0.4575",
"0.3955",
"0.1453",
"0.3967"
],
[
"0.147",
"0.3984",
"0.4568",
"0.3984",
"0.4568",
"0.4092",
"0.147",
"0.4092"
],
[
"0.511",
"0.3938",
"0.882",
"0.3938",
"0.882",
"0.4067",
"0.511",
"0.4067"
],
[
"0.1436",
"0.4104",
"0.4568",
"0.411",
"0.4568",
"0.4246",
"0.1436",
"0.4233"
],
[
"0.513",
"0.408",
"0.882",
"0.408",
"0.882",
"0.4216",
"0.513",
"0.4216"
],
[
"0.1453",
"0.4258",
"0.4575",
"0.4258",
"0.4575",
"0.4395",
"0.1453",
"0.4395"
],
[
"0.511",
"0.4216",
"0.882",
"0.4216",
"0.882",
"0.4353",
"0.511",
"0.4353"
],
[
"0.1453",
"0.4395",
"0.4568",
"0.4395",
"0.4568",
"0.4531",
"0.1453",
"0.4531"
],
[
"0.511",
"0.4358",
"0.882",
"0.4358",
"0.882",
"0.4495",
"0.511",
"0.4495"
],
[
"0.1453",
"0.4548",
"0.4575",
"0.4548",
"0.4575",
"0.4685",
"0.1453",
"0.4685"
],
[
"0.513",
"0.4507",
"0.882",
"0.4507",
"0.882",
"0.4644",
"0.513",
"0.4644"
],
[
"0.1453",
"0.4685",
"0.4592",
"0.4685",
"0.4592",
"0.4822",
"0.1453",
"0.4822"
],
[
"0.511",
"0.4644",
"0.881",
"0.4656",
"0.881",
"0.4785",
"0.511",
"0.478"
],
[
"0.147",
"0.484",
"0.4568",
"0.484",
"0.4568",
"0.4946",
"0.147",
"0.4946"
],
[
"0.517",
"0.4797",
"0.879",
"0.4797",
"0.879",
"0.4905",
"0.517",
"0.4905"
],
[
"0.1453",
"0.4963",
"0.4575",
"0.4963",
"0.4575",
"0.5103",
"0.1453",
"0.5103"
],
[
"0.511",
"0.4922",
"0.882",
"0.4922",
"0.882",
"0.506",
"0.511",
"0.506"
],
[
"0.147",
"0.511",
"0.4575",
"0.511",
"0.4575",
"0.525",
"0.147",
"0.525"
],
[
"0.513",
"0.5083",
"0.5884",
"0.5083",
"0.5884",
"0.5195",
"0.513",
"0.5195"
],
[
"0.147",
"0.527",
"0.4568",
"0.527",
"0.4568",
"0.5376",
"0.147",
"0.5376"
],
[
"0.1436",
"0.538",
"0.4575",
"0.539",
"0.4575",
"0.5527",
"0.1436",
"0.552"
],
[
"0.513",
"0.5474",
"0.882",
"0.549",
"0.882",
"0.5635",
"0.513",
"0.5625"
],
[
"0.147",
"0.5547",
"0.4568",
"0.5547",
"0.4568",
"0.5654",
"0.147",
"0.5654"
],
[
"0.1453",
"0.568",
"0.4592",
"0.568",
"0.4592",
"0.5806",
"0.1453",
"0.5806"
],
[
"0.511",
"0.564",
"0.879",
"0.564",
"0.879",
"0.5776",
"0.511",
"0.5776"
],
[
"0.1453",
"0.5845",
"0.346",
"0.582",
"0.346",
"0.593",
"0.1453",
"0.594"
],
[
"0.5283",
"0.5806",
"0.881",
"0.58",
"0.881",
"0.594",
"0.5283",
"0.5957"
],
[
"0.5146",
"0.597",
"0.881",
"0.597",
"0.881",
"0.611",
"0.5146",
"0.611"
],
[
"0.11584",
"0.6123",
"0.1453",
"0.6123",
"0.1453",
"0.6235",
"0.11584",
"0.6235"
],
[
"0.1511",
"0.6133",
"0.2612",
"0.6133",
"0.2612",
"0.6245",
"0.1511",
"0.6245"
],
[
"0.513",
"0.6133",
"0.881",
"0.6133",
"0.881",
"0.6274",
"0.513",
"0.6274"
],
[
"0.511",
"0.63",
"0.7246",
"0.63",
"0.7246",
"0.6436",
"0.511",
"0.6436"
],
[
"0.1184",
"0.638",
"0.4895",
"0.638",
"0.4895",
"0.652",
"0.1184",
"0.652"
],
[
"0.531",
"0.6465",
"0.881",
"0.6465",
"0.881",
"0.66",
"0.531",
"0.66"
],
[
"0.11755",
"0.654",
"0.486",
"0.654",
"0.486",
"0.6675",
"0.11755",
"0.6675"
],
[
"0.5146",
"0.662",
"0.879",
"0.662",
"0.879",
"0.676",
"0.5146",
"0.676"
],
[
"0.11755",
"0.6694",
"0.4844",
"0.6694",
"0.4844",
"0.683",
"0.11755",
"0.683"
],
[
"0.5146",
"0.6777",
"0.881",
"0.6777",
"0.881",
"0.6914",
"0.5146",
"0.6914"
],
[
"0.11755",
"0.686",
"0.4844",
"0.686",
"0.4844",
"0.6997",
"0.11755",
"0.6997"
],
[
"0.5146",
"0.695",
"0.881",
"0.695",
"0.881",
"0.708",
"0.5146",
"0.708"
],
[
"0.1184",
"0.703",
"0.4844",
"0.703",
"0.4844",
"0.7163",
"0.1184",
"0.7163"
],
[
"0.5146",
"0.7104",
"0.881",
"0.7104",
"0.881",
"0.7236",
"0.5146",
"0.7236"
],
[
"0.11755",
"0.7173",
"0.486",
"0.7188",
"0.486",
"0.733",
"0.11755",
"0.7324"
],
[
"0.5146",
"0.727",
"0.881",
"0.727",
"0.881",
"0.7407",
"0.5146",
"0.7407"
],
[
"0.11755",
"0.7354",
"0.486",
"0.7354",
"0.486",
"0.749",
"0.11755",
"0.749"
],
[
"0.5146",
"0.7446",
"0.876",
"0.7446",
"0.876",
"0.755",
"0.5146",
"0.755"
],
[
"0.1184",
"0.7505",
"0.4844",
"0.7505",
"0.4844",
"0.764",
"0.1184",
"0.764"
],
[
"0.513",
"0.759",
"0.882",
"0.759",
"0.882",
"0.7725",
"0.513",
"0.7725"
],
[
"0.1184",
"0.767",
"0.4844",
"0.767",
"0.4844",
"0.781",
"0.1184",
"0.781"
],
[
"0.5146",
"0.7754",
"0.881",
"0.7754",
"0.881",
"0.789",
"0.5146",
"0.789"
],
[
"0.1184",
"0.783",
"0.4844",
"0.783",
"0.4844",
"0.7964",
"0.1184",
"0.7964"
],
[
"0.5156",
"0.7915",
"0.879",
"0.7915",
"0.879",
"0.8047",
"0.5156",
"0.8047"
],
[
"0.12006",
"0.8003",
"0.1914",
"0.8003",
"0.1914",
"0.8125",
"0.12006",
"0.8125"
],
[
"0.513",
"0.808",
"0.879",
"0.808",
"0.879",
"0.822",
"0.513",
"0.822"
],
[
"0.1351",
"0.8154",
"0.4844",
"0.8154",
"0.4844",
"0.829",
"0.1351",
"0.829"
],
[
"0.513",
"0.8237",
"0.881",
"0.8237",
"0.881",
"0.8374",
"0.513",
"0.8374"
],
[
"0.1184",
"0.832",
"0.4844",
"0.832",
"0.4844",
"0.8457",
"0.1184",
"0.8457"
],
[
"0.513",
"0.839",
"0.881",
"0.839",
"0.881",
"0.8525",
"0.513",
"0.8525"
],
[
"0.1184",
"0.847",
"0.486",
"0.847",
"0.486",
"0.861",
"0.1184",
"0.861"
],
[
"0.513",
"0.8555",
"0.7173",
"0.8555",
"0.7173",
"0.869",
"0.513",
"0.869"
],
[
"0.12006",
"0.8643",
"0.4844",
"0.8643",
"0.4844",
"0.8774",
"0.12006",
"0.8774"
],
[
"0.12006",
"0.8813",
"0.4844",
"0.8813",
"0.4844",
"0.8945",
"0.12006",
"0.8945"
],
[
"0.532",
"0.8813",
"0.881",
"0.8813",
"0.881",
"0.8945",
"0.532",
"0.8945"
],
[
"0.511",
"0.8955",
"0.8774",
"0.8955",
"0.8774",
"0.9062",
"0.511",
"0.9062"
],
[
"0.1411",
"0.905",
"0.419",
"0.905",
"0.419",
"0.9185",
"0.1411",
"0.9185"
],
[
"0.513",
"0.9062",
"0.5557",
"0.9062",
"0.5557",
"0.9175",
"0.513",
"0.9175"
]
] | [
"低秩Softmax在理论上可以具有不可论证的类",
"，但在实践中很少",
"Andreas Grivas and Nikolay Bogoychev and Adam Lopez",
"Institute for Language, Cognition, and Computation",
"School of Informatics",
"University of Edinburgh",
"(agrivas,n.bogoych,alopez}",
"@ed.ac.uk",
"Abstract",
"LIaSSLIX",
"PiClassG/x",
"Class C2",
"Classifiers in natural language processing",
"",
"(NLP) often have a large number of output",
"classes. For example, neural language mod-",
"(class.c|x",
"P(Class a+ x)",
"ClassC1",
"ClassC3",
"elsbothpredicttokensfromavocabulary of",
"els (LMs)and machinetranslation (MT)mod-",
"thousands. The Softmax output layer of these",
"models typically receives as input a dense fea-",
"FeatureXi",
"ture representation, which has much lower di-",
"mensionality than the output.In theory, the",
"Figure I: Illustration of an unargmaxable class. Class",
"result is some words may be impossible to be",
"C4 can never be predicted using argmax for this Soft",
"predicted via argmax, irrespective of input fea-",
"max classifier with [C] = 4 classes and d = 2 input",
"tures, and empirically, there is evidence this",
"features. On the left, each feature vector is colored",
"happens in small language models (Demeter",
"according to the class assigned the largest probability;",
"et al., 2020) . In this paper we ask whether it",
"note that while ci ,C2 and c3 surface as regions, c4",
"can happen inpractical large language models",
"does not. On the right, we show that there is no di",
"and translation models. To do so, we develop",
"rection in feature space for which c4 has the largest",
"algorithms to detect such unargmaxable tokens",
"probability",
"in public models.We find that 13 out of 150",
"models do indeed have such tokens; however,",
"类器的表达性惩罚是某些输出分布无法表示。",
"they are very infrequent and unlikely to impact",
"model quality. We release our code so that oth-",
"Demeter et al.(2020)在SoftmaxLM中发现了这",
"erscaninspecttheirmodels.",
"弱点，表明从理论上讲，某些令牌永远无法",
"为任何输人分配最高概率，因此永远无法作",
"1",
"Introduction",
"为argmax预测产生。2我们称这样的令牌为",
"unargmaxable（参见图1）",
"具有大量输出类的概率多类分类器在NLPE",
"虽然Demeteretal.(2020)提出了一种检测不",
"(Chenetal.,2016)中很常见。例如，当代LM和",
"可变令牌的算法，并提供了它们在小型LM中",
"MT模型的词汇量从数万到数十万不等（Liu",
"存在的证据，但他们提出的算法没有提供任何",
"etal.,2020）。在对如此大的词汇表进行建模",
"保证，他们无法测试大型LM。在本文中，我",
"方面的最新进展主要是通过改进神经网络特征",
"们问：在实践中使用的大型模型中是否存在不",
"编码器来实现的Devlinetal.，2019;Conneau",
"可argmaxable令牌？为了回答这个问题，我们",
"etal.，2020）。但是，不管特征编码器的表现",
"开发了明确识别此类令牌的算法。我们测试了",
"力(Yun et al.,2020;Raghu et al.,2017)如何",
"7LM和143MT模型。其中，只有13的MT",
"将较低维特征线性映射到较高维输出的分类器",
"模型表现出不可争辩的令牌，即使在这些情况",
"都会降低表现力E(Yangetal.,2018），其后果",
"下，令牌也都是噜杂且不常见的。我们得出的",
"尚不清楚",
"结论是，尽管低等级Softmax的表现力约束可",
"在这项工作中，我们详细说明了将argmax",
"能会产生重要的影响，但大多数从业者无需担",
"预测与低秩分类器一起使用的后果，这些分",
"心无法确定的令牌。我们为他们提供了新的工",
"类器具有比d更多的输出类CI的分类器。例",
"具来确认他们自己的模型",
"如，MT模型通常具有大小为|CI～30000的",
"子词词汇表，但具有d～1024。这种低秩分",
"2This problem was also studied by Cover (1967) and",
"has an interesting history of independent discovery (Smith",
"https://github.com/andreasgrv/unargmaxable",
"2014)."
] | {
"arxiv_id": "2203.06462",
"language": "zh",
"page_idx": 1
} |
||
2203.06462-zh-2 | 2203.06462-zh-2.jpg | 我们的贡献如下：
·We explain how unargmaxable tokens can arise as a consequence of a rank constrained Softmax layer (Softmax Bottleneck).
$•$ We extend the work of Demeter et al. (2020) with verification algorithms that include the Softmax bias term and provide an exact answer rather than an approximate one.
· We verify a large number of commonly used publicly available language and translation models for unargmaxable tokens.
· We release our algorithm and code so that oth-
ers can inspect their models. 1
## 2 $\textbf{Background}$
# 2.1 Low-Rank Softmax (Softmax Bottleneck)
输出维度高于输入维度的神经网络层施加低秩约束。$^{3}$此类约束通常作为瓶颈存在于神经网络隐藏层中，例如。g。自动编码器 (Hinton and Zemel, 1994) 和投影头在多头变压器 (Bhojanapalli et al., 2020) 等。虽然瓶颈通过限制模型可以表示的功能而使模型的表现力降低，但它们在计算上都是可取的$\overline{\mathrm{E}}($Papadimitriou and Jain, 2021),因为它们比全秩层需要更少的内存和计算，并且作为归纳偏差的一种形式，因为假定数据近似位于低维流形$\overline{\mathrm{E}}($McInnes et al., 2018)
相反，在本文中，我们将重点放在具有低秩参数化的 Softmax 输出层的不良属性上，也称为$\overline{\mathrm{E}}$Softmax 瓶颈$\overline{\mathrm{E}}$(Yang et al.,2018)。关键的区别在于，Softmax 瓶颈通常不会跟随非线性变换，因此，秩约束通过将输出限制在子空间中以非常严格的方式限制了表达能力。“ 此约束被证明会损害 LM 的困惑$\overline{\mathrm{E}}($Yang et al., 2018),并且已提出非线性增强作为改进$\mathbb{E}$ (Yang et al., 2018; Kanai et al., 2018; Ganea et al., 2019)。相反，Sainath et al.(2013) 使用 softmax 层的低秩分解来通过 30-50% 减少其语音识别系统中的参数数量，而不会增加单词错误率， 这表明表达能力的损失并不总是影响聚合度量
由于 Softmax 瓶颈导致的表现力损失的后果因我们的观点而异。在考虑可以学习的概率分布的灵活性时，Ganea et al.(2019, Theorem 2) 表明，随着我们增加 Softmax 层权重的等级， 可实现的最小交叉熵损失会降低
A layer can also be low rank if weight vectors are
collinear, but we do not consider this case here.
$^{4}$A linear subspace if no bias term is present and an affine
subspace otherwise.
<FigureHere>
Figure 2: Illustration of how unargmaxable classes arise. The vectors on the left are the culprit Softmax weights for Figure 1.Each vector is a row of the Softmax weights $\tilde{W}\in\mathbb{R}^{4\times2}.c_4$ is interior to the convex hull, the triangle formed by $c_1,c_2$ and $c_3.$
在这项工作中，我们从 argmax 的角度关注表现力的丧失。为此，我们将 Softmax 的输出空间离散化，并根据无法实现的班级排名来量化表达能力的损失。从这个可解释的角度来看我们将看到，由于 Softmax 瓶颈，一些排名是不可实现的，因此可能会出现无法确定的类
### 2.2 $\textbf{Unargmaxable Classes}$
Demeter et al.(2020) 表明，如果一个类的 Soft- $\max$权重向量位于其余类权重向量的凸包的内部，则该类是不可变的。他们这样做是通过证明内部类概率是由凸包上至少一个类的概率限制的(参见图$\boxed{\mathrm{E}}2$ 和 Cover, 1967,Figure 1)。然而，在他们的分析中，他们没有解决包含偏差项的 Softmax 层。我们在 3 节中解决了这个限制，从而使我们能够在任何发布的模型中搜索不可 argmaxable 类
为了检测在没有偏置项的 LM 中是否出现不可 argmaxable 令牌，作者引入了一种近似算法，该算法断言权重向量是否在凸包内部。这是近似的，因为与 SoftmaxLM 的第一个 10 维度上的精确算法(Qhull$,\boxed{\mathrm{F}}$Barber et al.,1996 ) 相比，他们的方法的精度接近 100%,但召回68%。在 3.3 节中，我们介绍了一种精确的算法来确定地检测不可 argmaxable 令牌
作者使用他们的近似算法来表明，与平滑的n-gramLM 分配的概率相比 , AWD-LSTMLME (Merity et al., 2018)从候选内部单词中“窃取” 了概率。然而，他们发现，当他们将 Softmax 权重的维度$d$增加到 200 时，被盗概率的影响开始消失。这就提出了一个问题，即被盗概率对于实践中使用的神经模型是否也很重要，这些模型也具有更大的 Softmax 权重维数
在这里，我们专门在具有较大$d\in$ [256,512,1024]的 MT 和 LM 模型中搜索不可argmaxable 令牌。我们使用不可争论的概率而不是被盗的概率来强调我们关注的是不可争 | [
[
"0.1368",
"0.0861",
"0.2747",
"0.0861",
"0.2747",
"0.0998",
"0.1368",
"0.0998"
],
[
"0.7173",
"0.0861",
"0.7734",
"0.0861",
"0.7734",
"0.08905",
"0.7173",
"0.08905"
],
[
"0.8213",
"0.0861",
"0.858",
"0.0861",
"0.858",
"0.0903",
"0.8213",
"0.0903"
],
[
"0.5996",
"0.09265",
"0.6113",
"0.09265",
"0.6113",
"0.0968",
"0.5996",
"0.0968"
],
[
"0.1377",
"0.11224",
"0.4844",
"0.11224",
"0.4844",
"0.1271",
"0.1377",
"0.1271"
],
[
"0.1511",
"0.1289",
"0.486",
"0.1277",
"0.486",
"0.1414",
"0.1511",
"0.1426"
],
[
"0.1528",
"0.1443",
"0.4248",
"0.1443",
"0.4248",
"0.158",
"0.1528",
"0.158"
],
[
"0.5156",
"0.1395",
"0.5283",
"0.1395",
"0.5283",
"0.1561",
"0.5156",
"0.1561"
],
[
"0.1377",
"0.1663",
"0.4844",
"0.1663",
"0.4844",
"0.1799",
"0.1377",
"0.1799"
],
[
"0.1528",
"0.1829",
"0.486",
"0.1829",
"0.486",
"0.1965",
"0.1528",
"0.1965"
],
[
"0.1545",
"0.1996",
"0.4878",
"0.1996",
"0.4878",
"0.2131",
"0.1545",
"0.2131"
],
[
"0.58",
"0.2007",
"0.6465",
"0.2031",
"0.6465",
"0.2144",
"0.58",
"0.212"
],
[
"0.1528",
"0.2156",
"0.4282",
"0.2156",
"0.4282",
"0.2286",
"0.1528",
"0.2286"
],
[
"0.513",
"0.2245",
"0.882",
"0.2245",
"0.882",
"0.2382",
"0.513",
"0.2382"
],
[
"0.1377",
"0.2369",
"0.4878",
"0.2369",
"0.4878",
"0.2505",
"0.1377",
"0.2505"
],
[
"0.51",
"0.2393",
"0.879",
"0.2393",
"0.879",
"0.253",
"0.51",
"0.253"
],
[
"0.1528",
"0.253",
"0.4895",
"0.253",
"0.4895",
"0.2673",
"0.1528",
"0.2673"
],
[
"0.511",
"0.253",
"0.881",
"0.253",
"0.881",
"0.2666",
"0.511",
"0.2666"
],
[
"0.1528",
"0.2683",
"0.3997",
"0.2695",
"0.3997",
"0.284",
"0.1528",
"0.2832"
],
[
"0.513",
"0.2666",
"0.881",
"0.2666",
"0.881",
"0.2808",
"0.513",
"0.2808"
],
[
"0.513",
"0.282",
"0.8037",
"0.282",
"0.8037",
"0.2957",
"0.513",
"0.2957"
],
[
"0.1377",
"0.2915",
"0.4895",
"0.2915",
"0.4895",
"0.306",
"0.1377",
"0.306"
],
[
"0.1511",
"0.3071",
"0.3745",
"0.306",
"0.3745",
"0.3206",
"0.1511",
"0.3218"
],
[
"0.529",
"0.3225",
"0.879",
"0.3225",
"0.879",
"0.3362",
"0.529",
"0.3362"
],
[
"0.11584",
"0.3342",
"0.1453",
"0.3342",
"0.1453",
"0.3467",
"0.11584",
"0.3467"
],
[
"0.1511",
"0.3342",
"0.2612",
"0.3342",
"0.2612",
"0.3486",
"0.1511",
"0.3486"
],
[
"0.5146",
"0.339",
"0.881",
"0.339",
"0.881",
"0.3528",
"0.5146",
"0.3528"
],
[
"0.11755",
"0.3582",
"0.1436",
"0.3582",
"0.1436",
"0.3706",
"0.11755",
"0.3706"
],
[
"0.1587",
"0.3582",
"0.3938",
"0.3582",
"0.3938",
"0.3718",
"0.1587",
"0.3718"
],
[
"0.513",
"0.3557",
"0.881",
"0.3557",
"0.881",
"0.3694",
"0.513",
"0.3694"
],
[
"0.513",
"0.3706",
"0.879",
"0.3706",
"0.879",
"0.3843",
"0.513",
"0.3843"
],
[
"0.1394",
"0.361",
"0.1561",
"0.361",
"0.1561",
"0.3682",
"0.1394",
"0.3682"
],
[
"0.1604",
"0.376",
"0.2494",
"0.376",
"0.2494",
"0.3872",
"0.1604",
"0.3872"
],
[
"0.5146",
"0.3872",
"0.881",
"0.3872",
"0.881",
"0.401",
"0.5146",
"0.401"
],
[
"0.11755",
"0.3943",
"0.4836",
"0.3943",
"0.4836",
"0.408",
"0.11755",
"0.408"
],
[
"0.513",
"0.4038",
"0.8623",
"0.4038",
"0.8623",
"0.4175",
"0.513",
"0.4175"
],
[
"0.1184",
"0.411",
"0.4844",
"0.411",
"0.4844",
"0.4246",
"0.1184",
"0.4246"
],
[
"0.11755",
"0.427",
"0.486",
"0.4275",
"0.486",
"0.4424",
"0.11755",
"0.4412"
],
[
"0.55",
"0.4312",
"0.5615",
"0.4358",
"0.5576",
"0.4407",
"0.546",
"0.4353"
],
[
"0.555",
"0.4312",
"0.739",
"0.4312",
"0.739",
"0.4453",
"0.555",
"0.4453"
],
[
"0.11584",
"0.4424",
"0.4844",
"0.4448",
"0.4844",
"0.4578",
"0.11584",
"0.456"
],
[
"0.513",
"0.4316",
"0.5474",
"0.4316",
"0.5474",
"0.4436",
"0.513",
"0.4436"
],
[
"0.511",
"0.4531",
"0.882",
"0.4531",
"0.882",
"0.466",
"0.511",
"0.466"
],
[
"0.11755",
"0.4602",
"0.4844",
"0.4602",
"0.4844",
"0.4739",
"0.11755",
"0.4739"
],
[
"0.513",
"0.4685",
"0.881",
"0.4685",
"0.881",
"0.4822",
"0.513",
"0.4822"
],
[
"0.1184",
"0.4768",
"0.4844",
"0.4768",
"0.4844",
"0.4905",
"0.1184",
"0.4905"
],
[
"0.5146",
"0.485",
"0.881",
"0.485",
"0.881",
"0.4988",
"0.5146",
"0.4988"
],
[
"0.11755",
"0.4922",
"0.4836",
"0.4922",
"0.4836",
"0.506",
"0.11755",
"0.506"
],
[
"0.11755",
"0.5073",
"0.4844",
"0.5083",
"0.4844",
"0.5225",
"0.11755",
"0.5215"
],
[
"0.5146",
"0.502",
"0.881",
"0.502",
"0.881",
"0.5156",
"0.5146",
"0.5156"
],
[
"0.5146",
"0.517",
"0.881",
"0.517",
"0.881",
"0.531",
"0.5146",
"0.531"
],
[
"0.12006",
"0.525",
"0.486",
"0.525",
"0.486",
"0.538",
"0.12006",
"0.538"
],
[
"0.11755",
"0.539",
"0.4836",
"0.539",
"0.4836",
"0.5527",
"0.11755",
"0.5527"
],
[
"0.5146",
"0.5337",
"0.881",
"0.5337",
"0.881",
"0.5474",
"0.5146",
"0.5474"
],
[
"0.5146",
"0.55",
"0.879",
"0.55",
"0.879",
"0.5635",
"0.5146",
"0.5635"
],
[
"0.11755",
"0.5557",
"0.1628",
"0.5557",
"0.1628",
"0.571",
"0.11755",
"0.571"
],
[
"0.513",
"0.564",
"0.879",
"0.564",
"0.879",
"0.5776",
"0.513",
"0.5776"
],
[
"0.1368",
"0.5723",
"0.4844",
"0.5723",
"0.4844",
"0.586",
"0.1368",
"0.586"
],
[
"0.511",
"0.5806",
"0.665",
"0.58",
"0.665",
"0.594",
"0.511",
"0.5957"
],
[
"0.1184",
"0.589",
"0.486",
"0.589",
"0.486",
"0.6025",
"0.1184",
"0.6025"
],
[
"0.529",
"0.5967",
"0.879",
"0.5967",
"0.879",
"0.611",
"0.529",
"0.611"
],
[
"0.1184",
"0.605",
"0.486",
"0.605",
"0.486",
"0.619",
"0.1184",
"0.619"
],
[
"0.1184",
"0.619",
"0.4844",
"0.619",
"0.4844",
"0.6343",
"0.1184",
"0.6343"
],
[
"0.513",
"0.6143",
"0.881",
"0.6143",
"0.881",
"0.6274",
"0.513",
"0.6274"
],
[
"0.11755",
"0.6357",
"0.4836",
"0.6357",
"0.4836",
"0.6494",
"0.11755",
"0.6494"
],
[
"0.513",
"0.6313",
"0.881",
"0.6313",
"0.881",
"0.6445",
"0.513",
"0.6445"
],
[
"0.513",
"0.6455",
"0.881",
"0.6445",
"0.881",
"0.659",
"0.513",
"0.66"
],
[
"0.1184",
"0.654",
"0.4768",
"0.654",
"0.4768",
"0.6675",
"0.1184",
"0.6675"
],
[
"0.5146",
"0.662",
"0.882",
"0.662",
"0.882",
"0.676",
"0.5146",
"0.676"
],
[
"0.1184",
"0.6694",
"0.4836",
"0.6694",
"0.4836",
"0.683",
"0.1184",
"0.683"
],
[
"0.5146",
"0.6777",
"0.8774",
"0.6777",
"0.8774",
"0.6914",
"0.5146",
"0.6914"
],
[
"0.11755",
"0.684",
"0.4844",
"0.684",
"0.4844",
"0.6978",
"0.11755",
"0.6978"
],
[
"0.511",
"0.6934",
"0.879",
"0.6934",
"0.879",
"0.707",
"0.511",
"0.707"
],
[
"0.1184",
"0.7017",
"0.4802",
"0.7017",
"0.4802",
"0.712",
"0.1184",
"0.712"
],
[
"0.513",
"0.7114",
"0.81",
"0.7114",
"0.81",
"0.7246",
"0.513",
"0.7246"
],
[
"0.1184",
"0.7173",
"0.4844",
"0.7173",
"0.4844",
"0.731",
"0.1184",
"0.731"
],
[
"0.532",
"0.727",
"0.879",
"0.727",
"0.879",
"0.7407",
"0.532",
"0.7407"
],
[
"0.1184",
"0.733",
"0.4844",
"0.733",
"0.4844",
"0.7476",
"0.1184",
"0.7476"
],
[
"0.511",
"0.7437",
"0.881",
"0.741",
"0.881",
"0.755",
"0.511",
"0.7573"
],
[
"0.1184",
"0.75",
"0.4844",
"0.75",
"0.4844",
"0.763",
"0.1184",
"0.763"
],
[
"0.513",
"0.7603",
"0.8774",
"0.7603",
"0.8774",
"0.774",
"0.513",
"0.774"
],
[
"0.1184",
"0.7666",
"0.4844",
"0.7666",
"0.4844",
"0.78",
"0.1184",
"0.78"
],
[
"0.517",
"0.7754",
"0.881",
"0.7754",
"0.881",
"0.789",
"0.517",
"0.789"
],
[
"0.1368",
"0.783",
"0.4844",
"0.783",
"0.4844",
"0.7964",
"0.1368",
"0.7964"
],
[
"0.1184",
"0.799",
"0.4844",
"0.799",
"0.4844",
"0.8125",
"0.1184",
"0.8125"
],
[
"0.5146",
"0.7935",
"0.879",
"0.7935",
"0.879",
"0.804",
"0.5146",
"0.804"
],
[
"0.513",
"0.808",
"0.881",
"0.808",
"0.881",
"0.822",
"0.513",
"0.822"
],
[
"0.11755",
"0.814",
"0.4844",
"0.814",
"0.4844",
"0.8276",
"0.11755",
"0.8276"
],
[
"0.511",
"0.8237",
"0.879",
"0.8237",
"0.879",
"0.8374",
"0.511",
"0.8374"
],
[
"0.1184",
"0.83",
"0.4844",
"0.83",
"0.4844",
"0.844",
"0.1184",
"0.844"
],
[
"0.511",
"0.8384",
"0.8237",
"0.839",
"0.8237",
"0.854",
"0.511",
"0.8525"
],
[
"0.11755",
"0.8457",
"0.378",
"0.8467",
"0.378",
"0.861",
"0.11755",
"0.86"
],
[
"0.529",
"0.857",
"0.86",
"0.857",
"0.86",
"0.8706",
"0.529",
"0.8706"
],
[
"0.1377",
"0.868",
"0.486",
"0.868",
"0.486",
"0.882",
"0.1377",
"0.882"
],
[
"0.5146",
"0.8726",
"0.881",
"0.8726",
"0.881",
"0.8857",
"0.5146",
"0.8857"
],
[
"0.11584",
"0.88",
"0.4014",
"0.88",
"0.4014",
"0.8936",
"0.11584",
"0.8936"
],
[
"0.1394",
"0.8936",
"0.4878",
"0.8936",
"0.4878",
"0.9067",
"0.1394",
"0.9067"
],
[
"0.513",
"0.888",
"0.881",
"0.888",
"0.881",
"0.903",
"0.513",
"0.903"
],
[
"0.1184",
"0.9067",
"0.2393",
"0.9067",
"0.2393",
"0.9175",
"0.1184",
"0.9175"
],
[
"0.5146",
"0.905",
"0.879",
"0.905",
"0.879",
"0.9185",
"0.5146",
"0.9185"
]
] | [
"我们的贡献如下：",
"acc+v",
"5SCX",
"dul",
": We explain how unargmaxable tokens can",
"arise as a consequence of a rank constrained",
"Softmax layer (Softmax Bottleneck).",
"e",
": We extend the work of Demeter et al. (2020)",
"with verification algorithms that include the",
"Softmax bias term and provide an exact an-",
"Feature xi",
"swer rather than an approximate one.",
"Figure 2:Illustration of how unargmaxable classes",
". We verify a large number of commonly used",
"arise. The vectors on the left are the culprit Softmax",
"publicly available language and translation",
"weights for Figure 1 . Each vector is a row of the Soft.",
"models for unargmaxable tokens.",
"max weights W e R4x2. C4 is interior to the convex",
"hull, the triangle formed by ci , C2 and c3",
": We release our algorithm and code so that oth-",
"ers can inspect their models. 1",
"在这项工作中，我们从argmax的角度关注",
"2.",
"Background",
"表现力的丧失。为此，我们将Softmax的输出",
"2.1",
"Low-Rank Softmax (Softmax",
"空间离散化，并根据无法实现的班级排名来量",
"化表达能力的损失。从这个可解释的角度来看，",
"",
"Bottleneck)",
"我们将看到，由于Softmax瓶颈，一些排名是",
"输出维度高于输入维度的神经网络层施加低秩",
"不可实现的，因此可能会出现无法确定的类",
"约束。3此类约束通常作为瓶颈存在于神经网",
"络隐藏层中，例如。g。自动编码器（Hintonand",
"",
"Unargmaxable Classes",
"Zemel.1994)和投影头在多头变压器(Bhojana",
"2.2",
"Demeteretal.（2020）表明，如果一个类的Soft-",
"pallietal.,2020）等。虽然瓶颈通过限制模型可",
"max权重向量位于其余类权重向量的凸包的内",
"以表示的功能而使模型的表现力降低，但它们",
"部，则该类是不可变的。他们这样做是通过证",
"在计算上都是可取的E(PapadimitriouandJain",
"2021），因为它们比全秩层需要更少的内存和",
"明内部类概率是由凸包上至少一个类的概率限",
"制的（参见图E2和Cover，1967,Figure1）。然",
"计算，并且作为归纳偏差的一种形式，因为",
"假定数据近似位于低维流形E(McInnes et al..",
"而，在他们的分析中，他们没有解决包含偏差",
"项的Softmax层。我们在3节中解决了这个限",
"2018)",
"制，从而使我们能够在任何发布的模型中搜索",
"相反，在本文中，我们将重点放在具有低秩",
"不可argmaxable类",
"参数化的Softmax输出层的不良属性上，也称",
"为了检测在没有偏置项的LM中是否出现",
"为ESoftmax瓶颈E(Yangetal.，2018）。关键",
"的区别在于，Softmax瓶颈通常不会跟随非线",
"不可argmaxable令牌，作者引入了一种近似算",
"性变换，因此，秩约束通过将输出限制在子",
"法，该算法断言权重向量是否在凸包内部。这",
"是近似的，因为与SoftmaxLM的第一个10维",
"空间中以非常严格的方式限制了表达能力。",
"度上的精确算法（Qhull,EBarberetal.，1996）",
"此约束被证明会损害LM的困惑E(Yangetal.",
"相比，他们的方法的精度接近100%，但召回",
"2018），并且已提出非线性增强作为改进F",
"68%。在3.3节中，我们介绍了一种精确的算",
"(Yang et al.. 2018: Kanai et al.. 2018: Ganea et al.",
"法来确定地检测不可argmaxable令牌",
"2019）。相反，Sainathetal.（2013）使用softmax",
"作者使用他们的近似算法来表明，与平滑的",
"层的低秩分解来通过30-50%减少其语音识",
"n-gramLM分配的概率相比，AWD-LSTMLMF",
"别系统中的参数数量，而不会增加单词错误率，",
"(Merityetal.，2018）从候选内部单词中“窃取",
"这表明表达能力的损失并不总是影响聚合度量",
"了概率。然而，他们发现，当他们将Softmax",
"由于Softmax瓶颈导致的表现力损失的后果",
"因我们的观点而异。在考虑可以学习的概率分",
"权重的维度d增加到200时，被盗概率的影响",
"开始消失。这就提出了一个问题，即被盗概率",
"布的灵活性时，Ganeaetal.(2019,Theorem2)",
"对于实践中使用的神经模型是否也很重要，这",
"表明，随着我们增加Softmax层权重的等级",
"些模型也具有更大的Softmax权重维数",
"可实现的最小交叉炳损失会降低",
"在这里，我们专门在具有较大d",
"3A layer can also be low rank if weight vectors are",
"[256,512,1024]的MT和LM模型中搜索不可",
"collinear, but we do not consider this case here",
"*A linear subspace if no bias term is present and an affine",
"argmaxable令牌。我们使用不可争论的概率而",
"subspace otherwise.",
"不是被盗的概率来强调我们关注的是不可争"
] | {
"arxiv_id": "2203.06462",
"language": "zh",
"page_idx": 2
} |
||
2203.06462-zh-3 | 2203.06462-zh-3.jpg | 论的令牌是否存在，而不是低秩 Softmax 学习的概率分配是否不那么灵活。我们将分析扩展到 MT 模型，因为它们比(生成)LM 具有更多的实际用例：如果 MT 模型中存在不可argmaxable 令牌，则在使用贪婪解码时永远不会产生受影响的令牌。在我们的实验中，我们发现，虽然在有限的情况下出现了不可争论的令牌，但它们并不重要
### 3 Detecting Unargmaxable Classes
为了量化在发布的 LM 和 MT 模型中是否出现不可 argmaxable 类，我们首先需要引入可处理的算法来检测它们。在本节中，我们将解释由于 Softmax 瓶颈而导致的不可讨论的类如何出现。然后，我们介绍了一种快速近似算法和一种慢速精确算法，我们将它们结合起来检测无法预测的词汇标记
## 3.1 Definitions
我们对矩阵和向量使用粗体。所有向量都是列向量。我们将$w_i$用于$W$的第$i$行，将$b_i$用于$b$的第$i$个元素
## 3.1.1 Softmax
Softmax 层为我们提供了为输入特征向量$x\in$
$\mathbb{R}^d$分配给目标类$c_t$的概率，如下所示：
(1)
$$\begin{aligned}P(C=c_{t}\mid\boldsymbol{x})&=\frac{e^{\boldsymbol{w}_{c_{t}}^{\top}\boldsymbol{x}+b_{c_{t}}}}{\sum_{i}e^{\boldsymbol{w}_{c_{i}}^{\top}\boldsymbol{x}+b_{c_{i}}}}\\&=\mathrm{softmax}(\boldsymbol{W}\boldsymbol{x}+\boldsymbol{b})_{c_{t}}\end{aligned}$$
(2)
其中$W\in\mathbb{R}^{|C|\times d}$是逐行堆叠的类权重向量， 而$b\in\mathbb{R}^{|C|}$是偏差项。以上用于计算 logits $y=Wx+b$。在下面的内容中，我们将$\mathbb{R}^d$ 中的功能激活$x$称为输入空间，将$\mathbb{R}^{|C|}$中的logits $y$称为 Softmax 层的输出空间
## 3.1.2 Discretising the Output Space into Permutations
正如我们在图 2 中看到的，有一些 Softmax 权重的安排，对于这些权重，目标类$c_t$ 不能作为argmax 出现。为了理解这种现象，将输出离散到更精细的粒度将是有帮助的：排名(Burges et al.,2005)。为了使分类器使用 argmax 决策规则预测类$c_t$ , 它必须通过为$c_t$分配最大概率来将其排名在所有其他类之上。从这个角度来看，分类器为每个输入$x$分配一个排列$\pi$, 该排列以概率递增的顺序对类索引进行排名。
$$\pi:P(c_{\pi_1}\mid\boldsymbol x)<P(c_{\pi_2}\mid\boldsymbol x)<\ldots<P(c_{\pi_{|C|}}\mid\boldsymbol x)$$
(3)
例如，如果我们有 4 个类并获得概率$P(C\mid$ $\boldsymbol{x})=\begin{bmatrix}.2&.4&.1&.3\end{bmatrix}^\top$,我们将$x$分配给$\boldsymbol\pi_{3142}$ ,因为$P(c\_3\mid\boldsymbol{x})<P(c\_1\mid\boldsymbol{x})<P(c\_4\mid\boldsymbol{x})<$ $P(c_2\mid x)$ 。通过读取置换的最后一个索引， 我们可以很容易地获得更粗略的 argmax 决策( $c_2$ )
3.2 How Can Unargmaxable Classes Arise?
当由于秩约束而无法实现将$c_t$排列在其余部分之上的所有排列时，类$c_t$ 是不可变的。我们通过结合以下两个观察来解释这是如何发生的Observation 1. We can discretise R$^|C|$ into regions corresponding to permutations by segmenting the space with hyperplanes. The hyperplanes that partition the output space into regions $\mathcal{R}_\pi$ corresponding to permutations are a well known structure in Combinatorics, the Braid Hyperplane Arrangement (Stanley, 2004).$^{5}$ The Braid Arrangement for 3 and 4 classes is illustrated in rows 1 and 2 of Figure 3 respectively
In order to be able to rank the classes according to permutation $\mathcal{R}_\pi$ ,our network needs to be able to map an input $x$ to region $\mathcal{R}_\mathrm{\pi}$ in the output space. However, this is not always possible when we have a Softmax Bottleneck as we elaborate be- $\log$w.
Observation 2. When we have rank constraints, only a subspace of $\mathbb{R}^{|C|}$ is feasible . Case i) softmax$( \boldsymbol{W}x)$ . By calculating $y$ = $Wx$ , the class logits $y$ are a linear combination of $d$ columns of $W.$ Therefore, when $d<|C|$ we can only represent a $d$ -dimensional subspace of $\mathbb{R}^{|C|}$ at best. This feasible subspace is illustrated as a grey plane in the middle column of Figure 3 . Case ii) softmax$(\boldsymbol Wx+\boldsymbol b).$ If we also have a bias term $b$ the model can choose how to offset the subspace. When the bias term $b$ is not in the column space of $W$ the zero vector 0 is no longer a feasible $y$ and instead of a linear subspace we have an affine subspace. See Figure 7 in the Appendix for an illustration comparing the two cases.
Corollary 1. A Softmax classifter parametrised by W and b can rank classes in the order of permutation $\pi$ iff the affine subspace spanned by W and $b$ intersects region R of the Braid Arrangement . $^{6}$ When $d<|C|-1$ there are regions that cannot
${^5\text{See Append}}$ix B for more details on hyperplane arrange-
ments and the Braid Arrangement specifically
$^{6}$This insight of slicing the Braid Arrangement was intro-
duced in Kamiya et al.(2011) . | [
[
"0.11755",
"0.0849",
"0.4836",
"0.0849",
"0.4836",
"0.0986",
"0.11755",
"0.0986"
],
[
"0.511",
"0.0849",
"0.8774",
"0.0849",
"0.8774",
"0.0986",
"0.511",
"0.0986"
],
[
"0.12006",
"0.10394",
"0.482",
"0.10394",
"0.482",
"0.114",
"0.12006",
"0.114"
],
[
"0.51",
"0.10156",
"0.881",
"0.1027",
"0.881",
"0.12054",
"0.51",
"0.1194"
],
[
"0.1184",
"0.11816",
"0.4844",
"0.11816",
"0.4844",
"0.1318",
"0.1184",
"0.1318"
],
[
"0.5156",
"0.1194",
"0.879",
"0.12054",
"0.879",
"0.1354",
"0.5156",
"0.1342"
],
[
"0.12006",
"0.1342",
"0.4844",
"0.1342",
"0.4844",
"0.1478",
"0.12006",
"0.1478"
],
[
"0.513",
"0.1354",
"0.8774",
"0.136",
"0.8774",
"0.1509",
"0.513",
"0.1497"
],
[
"0.1184",
"0.1509",
"0.4844",
"0.1509",
"0.4844",
"0.1644",
"0.1184",
"0.1644"
],
[
"0.511",
"0.152",
"0.881",
"0.1526",
"0.881",
"0.1675",
"0.511",
"0.1663"
],
[
"0.1184",
"0.165",
"0.4844",
"0.1663",
"0.4844",
"0.1812",
"0.1184",
"0.1799"
],
[
"0.51",
"0.1729",
"0.546",
"0.1716",
"0.546",
"0.1818",
"0.51",
"0.1829"
],
[
"0.1184",
"0.1812",
"0.4836",
"0.1818",
"0.4836",
"0.1965",
"0.1184",
"0.1953"
],
[
"0.1184",
"0.199",
"0.3032",
"0.199",
"0.3032",
"0.2131",
"0.1184",
"0.2131"
],
[
"0.513",
"0.1996",
"0.8696",
"0.1996",
"0.8696",
"0.2131",
"0.513",
"0.2131"
],
[
"0.1142",
"0.2268",
"0.4358",
"0.228",
"0.4358",
"0.2423",
"0.1142",
"0.2411"
],
[
"0.5146",
"0.2239",
"0.881",
"0.2239",
"0.881",
"0.2369",
"0.5146",
"0.2369"
],
[
"0.513",
"0.2393",
"0.881",
"0.2393",
"0.881",
"0.253",
"0.513",
"0.253"
],
[
"0.1184",
"0.2546",
"0.486",
"0.2546",
"0.486",
"0.2683",
"0.1184",
"0.2683"
],
[
"0.513",
"0.2546",
"0.881",
"0.2546",
"0.881",
"0.2695",
"0.513",
"0.2695"
],
[
"0.1184",
"0.2708",
"0.486",
"0.2708",
"0.486",
"0.2852",
"0.1184",
"0.2852"
],
[
"0.513",
"0.2778",
"0.882",
"0.2778",
"0.882",
"0.2915",
"0.513",
"0.2915"
],
[
"0.1184",
"0.2874",
"0.4844",
"0.2874",
"0.4844",
"0.3005",
"0.1184",
"0.3005"
],
[
"0.5146",
"0.2944",
"0.882",
"0.2944",
"0.882",
"0.3088",
"0.5146",
"0.3088"
],
[
"0.1184",
"0.3027",
"0.4844",
"0.3027",
"0.4844",
"0.3164",
"0.1184",
"0.3164"
],
[
"0.513",
"0.3113",
"0.881",
"0.3113",
"0.881",
"0.3247",
"0.513",
"0.3247"
],
[
"0.1184",
"0.3196",
"0.4685",
"0.3196",
"0.4685",
"0.33",
"0.1184",
"0.33"
],
[
"0.513",
"0.3267",
"0.882",
"0.3267",
"0.882",
"0.3403",
"0.513",
"0.3403"
],
[
"0.1184",
"0.335",
"0.4844",
"0.335",
"0.4844",
"0.3486",
"0.1184",
"0.3486"
],
[
"0.5146",
"0.3425",
"0.881",
"0.3425",
"0.881",
"0.3557",
"0.5146",
"0.3557"
],
[
"0.1184",
"0.3516",
"0.2678",
"0.3516",
"0.2678",
"0.3652",
"0.1184",
"0.3652"
],
[
"0.513",
"0.3594",
"0.882",
"0.3594",
"0.882",
"0.3723",
"0.513",
"0.3723"
],
[
"0.11584",
"0.379",
"0.251",
"0.379",
"0.251",
"0.3938",
"0.11584",
"0.3938"
],
[
"0.511",
"0.3735",
"0.879",
"0.3735",
"0.879",
"0.3872",
"0.511",
"0.3872"
],
[
"0.511",
"0.3896",
"0.881",
"0.3896",
"0.881",
"0.4026",
"0.511",
"0.4026"
],
[
"0.11755",
"0.402",
"0.4836",
"0.402",
"0.4836",
"0.415",
"0.11755",
"0.415"
],
[
"0.51",
"0.405",
"0.7095",
"0.4062",
"0.7095",
"0.4204",
"0.51",
"0.4192"
],
[
"0.11755",
"0.4163",
"0.4844",
"0.4175",
"0.4844",
"0.4316",
"0.11755",
"0.4312"
],
[
"0.529",
"0.4233",
"0.882",
"0.4233",
"0.882",
"0.437",
"0.529",
"0.437"
],
[
"0.1184",
"0.434",
"0.2627",
"0.434",
"0.2627",
"0.4478",
"0.1184",
"0.4478"
],
[
"0.511",
"0.4395",
"0.882",
"0.4407",
"0.882",
"0.4548",
"0.511",
"0.4536"
],
[
"0.11755",
"0.4602",
"0.2451",
"0.4602",
"0.2451",
"0.4744",
"0.11755",
"0.4744"
],
[
"0.513",
"0.4573",
"0.881",
"0.4573",
"0.881",
"0.4702",
"0.513",
"0.4702"
],
[
"0.511",
"0.4727",
"0.882",
"0.4714",
"0.882",
"0.4863",
"0.511",
"0.4868"
],
[
"0.11755",
"0.4785",
"0.4878",
"0.4797",
"0.4878",
"0.4946",
"0.11755",
"0.4934"
],
[
"0.1142",
"0.4934",
"0.4324",
"0.4946",
"0.4324",
"0.5127",
"0.1142",
"0.511"
],
[
"0.513",
"0.488",
"0.882",
"0.488",
"0.882",
"0.502",
"0.513",
"0.502"
],
[
"0.513",
"0.505",
"0.544",
"0.505",
"0.544",
"0.517",
"0.513",
"0.517"
],
[
"0.2913",
"0.5254",
"0.367",
"0.524",
"0.367",
"0.539",
"0.2913",
"0.5415"
],
[
"0.513",
"0.528",
"0.879",
"0.528",
"0.879",
"0.5415",
"0.513",
"0.5415"
],
[
"0.147",
"0.5376",
"0.2703",
"0.5376",
"0.2703",
"0.552",
"0.147",
"0.552"
],
[
"0.4626",
"0.538",
"0.4895",
"0.538",
"0.4895",
"0.552",
"0.4626",
"0.552"
],
[
"0.2761",
"0.5474",
"0.382",
"0.5415",
"0.3855",
"0.56",
"0.2795",
"0.568"
],
[
"0.513",
"0.542",
"0.882",
"0.542",
"0.882",
"0.557",
"0.513",
"0.557"
],
[
"0.51",
"0.557",
"0.8774",
"0.558",
"0.8774",
"0.576",
"0.51",
"0.5747"
],
[
"0.2603",
"0.5693",
"0.4417",
"0.5693",
"0.4417",
"0.5845",
"0.2603",
"0.5845"
],
[
"0.4634",
"0.5723",
"0.486",
"0.5723",
"0.486",
"0.583",
"0.4634",
"0.583"
],
[
"0.513",
"0.576",
"0.882",
"0.576",
"0.882",
"0.5903",
"0.513",
"0.5903"
],
[
"0.511",
"0.5903",
"0.881",
"0.5913",
"0.881",
"0.6055",
"0.511",
"0.605"
],
[
"0.11584",
"0.597",
"0.4844",
"0.5996",
"0.4844",
"0.6177",
"0.11584",
"0.615"
],
[
"0.5146",
"0.608",
"0.882",
"0.608",
"0.882",
"0.6216",
"0.5146",
"0.6216"
],
[
"0.12006",
"0.6177",
"0.4844",
"0.6177",
"0.4844",
"0.632",
"0.12006",
"0.632"
],
[
"0.51",
"0.6226",
"0.882",
"0.6235",
"0.882",
"0.638",
"0.51",
"0.637"
],
[
"0.1184",
"0.6343",
"0.4844",
"0.6343",
"0.4844",
"0.6484",
"0.1184",
"0.6484"
],
[
"0.513",
"0.64",
"0.881",
"0.6396",
"0.881",
"0.654",
"0.513",
"0.655"
],
[
"0.12006",
"0.6494",
"0.4844",
"0.6494",
"0.4844",
"0.6646",
"0.12006",
"0.6646"
],
[
"0.513",
"0.6562",
"0.881",
"0.6562",
"0.881",
"0.6694",
"0.513",
"0.6694"
],
[
"0.11755",
"0.6665",
"0.398",
"0.6665",
"0.398",
"0.681",
"0.11755",
"0.681"
],
[
"0.513",
"0.673",
"0.882",
"0.673",
"0.882",
"0.686",
"0.513",
"0.686"
],
[
"0.11584",
"0.6924",
"0.1823",
"0.6924",
"0.1823",
"0.707",
"0.11584",
"0.707"
],
[
"0.1738",
"0.6924",
"0.4475",
"0.6924",
"0.4475",
"0.707",
"0.1738",
"0.707"
],
[
"0.51",
"0.687",
"0.881",
"0.687",
"0.881",
"0.7017",
"0.51",
"0.7017"
],
[
"0.1721",
"0.709",
"0.2805",
"0.709",
"0.2805",
"0.7236",
"0.1721",
"0.7236"
],
[
"0.511",
"0.7036",
"0.882",
"0.7036",
"0.882",
"0.7173",
"0.511",
"0.7173"
],
[
"0.511",
"0.7197",
"0.881",
"0.7197",
"0.881",
"0.733",
"0.511",
"0.733"
],
[
"0.1184",
"0.7285",
"0.4844",
"0.7285",
"0.4844",
"0.7437",
"0.1184",
"0.7437"
],
[
"0.513",
"0.737",
"0.882",
"0.737",
"0.882",
"0.7505",
"0.513",
"0.7505"
],
[
"0.1184",
"0.745",
"0.4844",
"0.745",
"0.4844",
"0.759",
"0.1184",
"0.759"
],
[
"0.511",
"0.752",
"0.802",
"0.752",
"0.802",
"0.7656",
"0.511",
"0.7656"
],
[
"0.11755",
"0.761",
"0.4844",
"0.7603",
"0.4844",
"0.775",
"0.11755",
"0.7754"
],
[
"0.1184",
"0.777",
"0.486",
"0.777",
"0.486",
"0.7905",
"0.1184",
"0.7905"
],
[
"0.513",
"0.7754",
"0.881",
"0.7754",
"0.881",
"0.7905",
"0.513",
"0.7905"
],
[
"0.11755",
"0.7935",
"0.486",
"0.7935",
"0.486",
"0.808",
"0.11755",
"0.808"
],
[
"0.513",
"0.792",
"0.881",
"0.792",
"0.881",
"0.8057",
"0.513",
"0.8057"
],
[
"0.1184",
"0.81",
"0.4878",
"0.81",
"0.4878",
"0.8237",
"0.1184",
"0.8237"
],
[
"0.511",
"0.808",
"0.881",
"0.808",
"0.881",
"0.822",
"0.511",
"0.822"
],
[
"0.1184",
"0.826",
"0.4844",
"0.826",
"0.4844",
"0.839",
"0.1184",
"0.839"
],
[
"0.513",
"0.8247",
"0.8774",
"0.8247",
"0.8774",
"0.8384",
"0.513",
"0.8384"
],
[
"0.1184",
"0.8413",
"0.4902",
"0.8413",
"0.4902",
"0.8555",
"0.1184",
"0.8555"
],
[
"0.511",
"0.8384",
"0.881",
"0.8403",
"0.881",
"0.855",
"0.511",
"0.8525"
],
[
"0.1184",
"0.858",
"0.476",
"0.858",
"0.476",
"0.8716",
"0.1184",
"0.8716"
],
[
"0.532",
"0.866",
"0.882",
"0.868",
"0.882",
"0.883",
"0.532",
"0.8813"
],
[
"0.513",
"0.8813",
"0.7925",
"0.8813",
"0.7925",
"0.8945",
"0.513",
"0.8945"
],
[
"0.1234",
"0.888",
"0.4785",
"0.8896",
"0.4785",
"0.904",
"0.1234",
"0.903"
],
[
"0.534",
"0.891",
"0.882",
"0.8936",
"0.882",
"0.9077",
"0.534",
"0.9062"
],
[
"0.4634",
"0.9062",
"0.4895",
"0.9062",
"0.4895",
"0.9185",
"0.4634",
"0.9185"
],
[
"0.511",
"0.905",
"0.706",
"0.905",
"0.706",
"0.9185",
"0.511",
"0.9185"
]
] | [
"论的令牌是否存在，而不是低秩Softmax学习",
"例如，如果我们有4个类并获得概率P(C",
"的概率分配是否不那么灵活。我们将分析扩",
"α）=[2.41_3]，我们将分配给π3142",
"展到MT模型，因为它们比（生成）LM具有",
",因为 P(c_3|)<P(c_1 /α)< P(c_4[α) <",
"更多的实际用例：如果MT模型中存在不可",
"P(c 2|a）。通过读取置换的最后一个索引",
"argmaxable令牌，则在使用贪婪解码时永远不",
"我们可以很容易地获得更粗略的argmax决策（",
"会产生受影响的令牌。在我们的实验中，我们",
"C20",
"发现，虽然在有限的情况下出现了不可争论的",
"令牌，但它们并不重要",
"3.2How Can Unargmaxable Classes Arise?",
"3Detecting Unargmaxable Classes",
"当由于秩约束而无法实现将Ct排列在其余部",
"分之上的所有排列时，类Ct是不可变的。我们",
"为了量化在发布的LM和MT模型中是否出现",
"通过结合以下两个观察来解释这是如何发生的",
"不可argmaxable类，我们首先需要引入可处理",
"Observation 1. We can discretise RlCl into re-",
"的算法来检测它们。在本节中，我们将解释由",
"gions corresponding to permutations by segment-",
"于Softmax瓶颈而导致的不可讨论的类如何出",
"ing the space with hyperplanes. The hyperplanes",
"现。然后，我们介绍了一种快速近似算法和",
"that partition the output space into regions R cor-",
"种慢速精确算法，我们将它们结合起来检测无",
"responding to permutations are a well known struc-",
"法预测的词汇标记",
"ture in Combinatorics, the Braid Hyperplane Ar-",
"3.1 Definitions",
"rangement (Stanley, 2004). 5 The Braid Arrange",
"ment for3and 4classes is illustrated in rows 1 and",
"我们对矩阵和向量使用粗体。所有向量都是列",
"2 of Figure 3 respectively.",
"向量。我们将wi用于W的第i行，将bi用",
"In order to be able to rank the classes accord-",
"于b的第i个元素",
"ing to permutation R , our network needs to be",
"3.1.1 Softmax",
"able to map an input a to region R in the output",
"space. However, this is not always possible when",
"Softmax层为我们提供了为输入特征向量a E",
"Rd分配给目标类ct的概率，如下所示：",
"we have a Softmax Bottleneck as we elaborate be-",
"low.",
"ew,a+bet",
"Observation 2. When we have rank constraints,",
"P(C = Ct [α) =",
"(1)",
"E;ews,a+be.",
"only a subspace of RlCl is feasible . Case i)",
"softmax(Wa) By calculating y = Wa",
"= softmax(Wa + b)ct",
"(2)",
"the class logits y are a linear combination of d",
"columns of W. Therefore, when d < |Cl we",
"其中WERICIxd是逐行堆叠的类权重向量",
"can only represent a d -dimensional subspace of",
"而bERlCl是偏差项。以上用于计算logits",
"R/Cl at best. This feasible subspace is illustrated",
"y=Wα+b。在下面的内容中，我们将Rd",
"as a grey plane in the middle column of Figure 3 .",
"中的功能激活称为输人空间，将RICI中的",
"Case ii) softmax(W+b)If we also have a bias",
"logitsy称为Softmax层的输出空间",
"term b the model can choose how to offset the sub.",
"3.1.2J",
" Discretising the Output Space into",
"space. When the bias term b is not in the column",
"Permutations",
"space of W the zero vector o is no longer a feasi-",
"ble y and instead of a linear subspace we have an",
"正如我们在图2中看到的，有一些Softmax权",
"affine subspace. See Figure 7 in the Appendix for",
"重的安排，对于这些权重，目标类Ct不能作为",
"an illustration comparing the two cases",
"argmax出现。为了理解这种现象，将输出离散",
"到更精细的粒度将是有帮助的：排名（Burges",
"Corollary 1. A Softmax classifier parametrised by",
"etal.,2005）。为了使分类器使用argmax决策",
"W and b can rank classes in the order of permu",
"规则预测类Ct，它必须通过为Ct分配最大概",
"tation iff the affine subspace spanned by W and",
"率来将其排名在所有其他类之上。从这个角度",
"b intersects region R of the Braid Arrangement",
"来看，分类器为每个输人分配一个排列元，",
"6 When d <|Cl- 1 there are regions that cannot",
"该排列以概率递增的顺序对类索引进行排名。",
"5See Appendix B for more details on hyperplane arrange-",
"ments and the Braid Arrangement specifically.",
"T : P(cT1 I α) <P(C2 / α) <... < P(Cicl / )",
"6This insight of slicing the Braid Arrangement was intro-",
"(3)",
"duced in Kamiya et al. (2011) ."
] | {
"arxiv_id": "2203.06462",
"language": "zh",
"page_idx": 3
} |
||
2203.06462-zh-4 | 2203.06462-zh-4.jpg | <FigureHere>
Figure 3: Illustration of Corollary $1(3^{rd}$ column) as a result of Observation 1( 1$^{st}$ column) and Observation 2(
$2^{nd}$column) for softmax$(\boldsymbol W\boldsymbol{x}),\boldsymbol{W}\in\mathbb{R}^{|C|\times d},d=2.$ Planes truncated for ease of visualisation. Top row: In the left column we see the Braid Arrangement for 3 classes partitioning the output space into 6 regions that correspond to permutations: class rankings in increasing order of probability. In the middle column we see that because $d=2$ we can only map $x$ to the feasible logits, a plane (grey) defined by $W$ .Therefore, in the right column we see that we can only represent permutations that correspond to the regions we can intersect with this plane. For $|C|=3$ we can still represent all 6 rankings of 3 classes since any plane in general position will intersect all 6 regions. Bottom row: The Braid Arrangement for 4 classes. Since $d<|C|-1$ the plane can only intersect 12 regions so only 12/24 permutations are feasible. For example, we see that the plane intersects region $\mathcal{R}_{\pi_{1342}}$ but not $\mathcal{R}_{\pi_{1324}}$ and hence $\pi_{1342}$ is feasible while $\pi_1324$ is not. In fact, the orientation of the plane is such that none of the 6 $\mathcal{R}_{\pi_{***4}}$ regions are intersected. Therefore $c_4$ cannot be ranked above $c_1,c_2$ and $c_3$ and is unargmaxable as in Figures l and 2.
be intersected.$^{7}$ The feasible permutations in our example correspond to the regions formed on the grey plane illustrated in the rightmost column of Figure 3.Note that for $|C|=4$ only 12 out of 24 regions can be intersected.
当我们通过减小 Softmax 输人的$d$维度来缩小 Softmax 瓶颈时，更多的排列变得不可行 (Good and Tideman, 1977; Kamiya and Takemura,2005)。重要的是，如果我们选择$|C|$和$d$以及是否使用偏置项，则更改 Softmax 权重的值将更改可行排列的集合，但不会更改集合的基数$\overline{\mathrm{E}}($Cover, 1967; Smith, 2014)。更多细节见附录 C
Corollary 2. Class ct is unargmaxable when any permutation that would rank class ct above all other classes is infeasible.
$^{7}$When $d=C-1$ we can still intersect all regions, because the Braid Arrangement always has rank $|C|-1$ (all its normal vectors are perpendicular to the all ones vector 1 ).
## 3.2.1 Effect of Softmax Bias Term
如果没有偏置项，则与置换相对应的区域是无界的(请参见图 3 的最右列)。因此，只要限制包括原点，对 Softmax 层输入$x$施加任何范围限制都不会更改可行区域。然而，当我们引人一个偏置项时，我们也会得到有界区域(参见附录中的图 7,它对比了这两种情况)。因此，在这种情况下，Softmax 层的输入比例也很重要。如果输入没有足够大的范围，则会存在存在但功能编码器无法到达的区域
## 33 Exact Algorithm
给定一个由$W$和$b$参数化的 softmax 层，是否有任何类是不可变的？我们首先描述一个缓慢但精确的算法来回答这个问题
一个精确的算法要么通过返回一个可行点$x:$ argmax $(\boldsymbol Wx+\boldsymbol b)=c_t$来证明类$c_t$是可argmaxable 的，要么通过验证不存在这样的点来证明$c_t$是不可 argmaxable 的
要检查是否存在将$c_t$排在所有其他区域之 | [
[
"0.43",
"0.08435",
"0.552",
"0.08435",
"0.552",
"0.0974",
"0.43",
"0.0974"
],
[
"0.714",
"0.08435",
"0.852",
"0.08435",
"0.852",
"0.0974",
"0.714",
"0.0974"
],
[
"0.6455",
"0.0861",
"0.721",
"0.0861",
"0.721",
"0.0974",
"0.6455",
"0.0974"
],
[
"0.1814",
"0.0849",
"0.3032",
"0.0849",
"0.3032",
"0.0986",
"0.1814",
"0.0986"
],
[
"0.11584",
"0.0974",
"0.367",
"0.0986",
"0.367",
"0.1164",
"0.11584",
"0.1152"
],
[
"0.3972",
"0.0986",
"0.5825",
"0.0998",
"0.5825",
"0.114",
"0.3972",
"0.1134"
],
[
"0.665",
"0.0998",
"0.831",
"0.10095",
"0.831",
"0.1152",
"0.665",
"0.114"
],
[
"0.2644",
"0.1271",
"0.2854",
"0.1271",
"0.2854",
"0.1342",
"0.2644",
"0.1342"
],
[
"0.3972",
"0.1217",
"0.465",
"0.1217",
"0.465",
"0.1318",
"0.3972",
"0.1318"
],
[
"0.7",
"0.13",
"0.7466",
"0.1372",
"0.742",
"0.1526",
"0.6963",
"0.1455"
],
[
"0.7944",
"0.1312",
"0.837",
"0.133",
"0.8364",
"0.1484",
"0.7925",
"0.1467"
],
[
"0.12427",
"0.1437",
"0.1411",
"0.1437",
"0.1411",
"0.1538",
"0.12427",
"0.1538"
],
[
"0.1234",
"0.155",
"0.1427",
"0.155",
"0.1427",
"0.1692",
"0.1234",
"0.1692"
],
[
"0.1663",
"0.1509",
"0.1957",
"0.1509",
"0.1957",
"0.1621",
"0.1663",
"0.1621"
],
[
"0.3132",
"0.1609",
"0.3257",
"0.1609",
"0.3257",
"0.165",
"0.3132",
"0.165"
],
[
"0.7944",
"0.1644",
"0.837",
"0.1692",
"0.833",
"0.1853",
"0.791",
"0.1799"
],
[
"0.1184",
"0.1704",
"0.1453",
"0.1704",
"0.1453",
"0.1895",
"0.1184",
"0.1895"
],
[
"0.2024",
"0.1841",
"0.2334",
"0.1895",
"0.2301",
"0.199",
"0.199",
"0.1936"
],
[
"0.6675",
"0.1812",
"0.693",
"0.1812",
"0.693",
"0.1924",
"0.6675",
"0.1924"
],
[
"0.2603",
"0.187",
"0.2678",
"0.187",
"0.2678",
"0.1912",
"0.2603",
"0.1912"
],
[
"0.6885",
"0.1895",
"0.708",
"0.1895",
"0.708",
"0.1936",
"0.6885",
"0.1936"
],
[
"0.756",
"0.2048",
"0.8037",
"0.2102",
"0.7983",
"0.2257",
"0.7524",
"0.2203"
],
[
"0.669",
"0.2131",
"0.686",
"0.2131",
"0.686",
"0.2227",
"0.669",
"0.2227"
],
[
"0.7",
"0.253",
"0.75",
"0.2559",
"0.748",
"0.2673",
"0.6987",
"0.2642"
],
[
"0.744",
"0.253",
"0.797",
"0.2559",
"0.796",
"0.2666",
"0.742",
"0.2632"
],
[
"0.663",
"0.2612",
"0.7095",
"0.2666",
"0.706",
"0.2756",
"0.66",
"0.2708"
],
[
"0.7866",
"0.2683",
"0.8364",
"0.2756",
"0.833",
"0.2874",
"0.784",
"0.2798"
],
[
"0.2612",
"0.257",
"0.2737",
"0.257",
"0.2737",
"0.2612",
"0.2612",
"0.2612"
],
[
"0.12177",
"0.275",
"0.1436",
"0.275",
"0.1436",
"0.3027",
"0.12177",
"0.3027"
],
[
"0.663",
"0.2874",
"0.706",
"0.2903",
"0.705",
"0.3018",
"0.6616",
"0.2986"
],
[
"0.12006",
"0.3027",
"0.1453",
"0.3027",
"0.1453",
"0.3218",
"0.12006",
"0.3218"
],
[
"0.79",
"0.3005",
"0.833",
"0.304",
"0.831",
"0.3154",
"0.7886",
"0.3123"
],
[
"0.665",
"0.3123",
"0.705",
"0.3184",
"0.702",
"0.3267",
"0.6616",
"0.3206"
],
[
"0.7866",
"0.3247",
"0.8364",
"0.33",
"0.833",
"0.3403",
"0.784",
"0.335"
],
[
"0.6675",
"0.3516",
"0.69",
"0.357",
"0.687",
"0.3635",
"0.665",
"0.3582"
],
[
"0.79",
"0.3528",
"0.833",
"0.3552",
"0.831",
"0.3652",
"0.7886",
"0.3635"
],
[
"0.748",
"0.3623",
"0.7944",
"0.3652",
"0.7925",
"0.376",
"0.7466",
"0.3723"
],
[
"0.11755",
"0.3872",
"0.881",
"0.3872",
"0.881",
"0.401",
"0.11755",
"0.401"
],
[
"0.11584",
"0.3984",
"0.8833",
"0.3984",
"0.8833",
"0.4163",
"0.11584",
"0.4163"
],
[
"0.11755",
"0.4163",
"0.882",
"0.4163",
"0.882",
"0.43",
"0.11755",
"0.43"
],
[
"0.11584",
"0.4312",
"0.8833",
"0.4312",
"0.8833",
"0.4448",
"0.11584",
"0.4448"
],
[
"0.11584",
"0.4448",
"0.881",
"0.4448",
"0.881",
"0.4578",
"0.11584",
"0.4578"
],
[
"0.11584",
"0.459",
"0.882",
"0.459",
"0.882",
"0.4727",
"0.11584",
"0.4727"
],
[
"0.11755",
"0.4727",
"0.881",
"0.4727",
"0.881",
"0.4863",
"0.11755",
"0.4863"
],
[
"0.11755",
"0.4868",
"0.881",
"0.4868",
"0.881",
"0.5005",
"0.11755",
"0.5005"
],
[
"0.11755",
"0.502",
"0.882",
"0.502",
"0.882",
"0.5166",
"0.11755",
"0.5166"
],
[
"0.1125",
"0.5127",
"0.869",
"0.514",
"0.869",
"0.532",
"0.1125",
"0.53"
],
[
"0.11584",
"0.53",
"0.882",
"0.53",
"0.882",
"0.5444",
"0.11584",
"0.5444"
],
[
"0.11755",
"0.5444",
"0.1696",
"0.5444",
"0.1696",
"0.5557",
"0.11755",
"0.5557"
],
[
"0.11755",
"0.583",
"0.4878",
"0.583",
"0.4878",
"0.597",
"0.11755",
"0.597"
],
[
"0.5146",
"0.585",
"0.555",
"0.585",
"0.555",
"0.5967",
"0.5146",
"0.5967"
],
[
"0.5635",
"0.5884",
"0.571",
"0.5884",
"0.571",
"0.593",
"0.5635",
"0.593"
],
[
"0.5693",
"0.5845",
"0.796",
"0.5845",
"0.796",
"0.597",
"0.5693",
"0.597"
],
[
"0.11755",
"0.601",
"0.486",
"0.601",
"0.486",
"0.6143",
"0.11755",
"0.6143"
],
[
"0.5146",
"0.6055",
"0.881",
"0.6055",
"0.881",
"0.619",
"0.5146",
"0.619"
],
[
"0.11584",
"0.616",
"0.4878",
"0.615",
"0.4878",
"0.63",
"0.11584",
"0.6313"
],
[
"0.5146",
"0.6216",
"0.881",
"0.6216",
"0.881",
"0.6357",
"0.5146",
"0.6357"
],
[
"0.11755",
"0.632",
"0.486",
"0.632",
"0.486",
"0.6465",
"0.11755",
"0.6465"
],
[
"0.5146",
"0.638",
"0.881",
"0.638",
"0.881",
"0.653",
"0.5146",
"0.653"
],
[
"0.11584",
"0.6484",
"0.3132",
"0.6484",
"0.3132",
"0.662",
"0.11584",
"0.662"
],
[
"0.5146",
"0.655",
"0.881",
"0.655",
"0.881",
"0.6685",
"0.5146",
"0.6685"
],
[
"0.1377",
"0.676",
"0.486",
"0.676",
"0.486",
"0.6895",
"0.1377",
"0.6895"
],
[
"0.5146",
"0.6704",
"0.879",
"0.6704",
"0.879",
"0.684",
"0.5146",
"0.684"
],
[
"0.11584",
"0.6914",
"0.4836",
"0.6914",
"0.4836",
"0.705",
"0.11584",
"0.705"
],
[
"0.5156",
"0.686",
"0.879",
"0.686",
"0.879",
"0.6997",
"0.5156",
"0.6997"
],
[
"0.1184",
"0.708",
"0.4878",
"0.708",
"0.4878",
"0.7217",
"0.1184",
"0.7217"
],
[
"0.5146",
"0.7036",
"0.8774",
"0.7036",
"0.8774",
"0.7144",
"0.5146",
"0.7144"
],
[
"0.11584",
"0.7236",
"0.4844",
"0.7236",
"0.4844",
"0.737",
"0.11584",
"0.737"
],
[
"0.513",
"0.7188",
"0.881",
"0.7188",
"0.881",
"0.7324",
"0.513",
"0.7324"
],
[
"0.511",
"0.733",
"0.8096",
"0.734",
"0.8096",
"0.749",
"0.511",
"0.7476"
],
[
"0.11755",
"0.7407",
"0.4844",
"0.7407",
"0.4844",
"0.7544",
"0.11755",
"0.7544"
],
[
"0.11755",
"0.756",
"0.4836",
"0.756",
"0.4836",
"0.7695",
"0.11755",
"0.7695"
],
[
"0.513",
"0.7666",
"0.5415",
"0.7666",
"0.5415",
"0.778",
"0.513",
"0.778"
],
[
"0.5557",
"0.7666",
"0.692",
"0.7666",
"0.692",
"0.78",
"0.5557",
"0.78"
],
[
"0.1184",
"0.788",
"0.1973",
"0.788",
"0.1973",
"0.8027",
"0.1184",
"0.8027"
],
[
"0.11584",
"0.7715",
"0.4836",
"0.7715",
"0.4836",
"0.785",
"0.11584",
"0.785"
],
[
"0.513",
"0.789",
"0.881",
"0.789",
"0.881",
"0.8027",
"0.513",
"0.8027"
],
[
"0.513",
"0.8047",
"0.881",
"0.8047",
"0.881",
"0.8193",
"0.513",
"0.8193"
],
[
"0.11755",
"0.8164",
"0.4844",
"0.8164",
"0.4844",
"0.83",
"0.11755",
"0.83"
],
[
"0.513",
"0.822",
"0.7715",
"0.822",
"0.7715",
"0.835",
"0.513",
"0.835"
],
[
"0.11584",
"0.832",
"0.4878",
"0.8306",
"0.4878",
"0.8457",
"0.11584",
"0.8467"
],
[
"0.54",
"0.8384",
"0.881",
"0.8384",
"0.881",
"0.8525",
"0.54",
"0.8525"
],
[
"0.11755",
"0.8496",
"0.3105",
"0.8496",
"0.3105",
"0.86",
"0.11755",
"0.86"
],
[
"0.51",
"0.855",
"0.881",
"0.854",
"0.881",
"0.868",
"0.51",
"0.869"
],
[
"0.51",
"0.8716",
"0.879",
"0.8716",
"0.879",
"0.8853",
"0.51",
"0.8853"
],
[
"0.1368",
"0.879",
"0.4878",
"0.8813",
"0.4878",
"0.8955",
"0.1368",
"0.8936"
],
[
"0.513",
"0.887",
"0.76",
"0.887",
"0.76",
"0.902",
"0.513",
"0.902"
],
[
"0.11755",
"0.8936",
"0.4878",
"0.8936",
"0.4878",
"0.9067",
"0.11755",
"0.9067"
],
[
"0.11584",
"0.905",
"0.4744",
"0.905",
"0.4744",
"0.9185",
"0.11584",
"0.9185"
],
[
"0.532",
"0.904",
"0.879",
"0.904",
"0.879",
"0.9185",
"0.532",
"0.9185"
]
] | [
"Observation (2):",
" Corollary l:",
"(1) &(2) ",
"Observation (1):",
"Discretise IR/Cl into permutations",
"Observe rank constraints",
"Feasiblepermutations",
"R元",
"Feasible logits",
"Rx312",
"R元321",
"3",
"1I",
"R元1",
"39",
"R元231",
"",
"Rr21",
"R元",
"",
"132",
"Rx213",
"R.",
"R#3142",
"R元1342",
"R-3412",
"Ri1432",
"",
"=",
"Rx3421",
"0",
"R元1423",
"R#3241",
"Ri1243",
"Rr2",
"R元2143",
"Rr2413",
"Figure 3: Illustration of Corollary 1 ( 3rd column) as a result of Observation 1 (1st column) and Observation 2 (",
"2nd column) for softmax(Wa) , W e R/C|xd, d = 2. Planes truncated for ease of visualisation. Top row : In the",
"left column we see the Braid Arrangement for 3 classes partitioning the output space into 6 regions that correspond",
"to permutations: class rankings in increasing order of probability. In the middle column we see that because d = 2",
"we can only map to the feasible logits, a plane (grey) defined by W . Therefore, in the right column we see that",
"we can only represent permutations that correspond to the regions we can intersect with this plane. For C| = 3 we",
"can still represent all 6 rankings of 3 classes since any plane in general position will intersect all 6 regions. Bottom",
"row : The Braid Arrangement for 4 classes. Since d < [C - 1 the plane can only intersect 12 regions so only",
"12/24 permutations are feasible. For example, we see that the plane intersects region R1342 but not Rx132 and",
"hence 元1342 is feasible while 1324 is not. In fact, the orientation of the plane is such that none of the 6 R元*",
"regions are intersected. Therefore c4 cannot be ranked above ci, C2 and c3 and is unargmaxable as in Figures 1",
"and 2.",
"be intersected. 7 The feasible permutations in our",
"3.2.1",
"",
"Effect of Softmax Bias Term",
"example correspond to the regions formed on the",
"如果没有偏置项，则与置换相对应的区域是无",
"grey plane illustrated in the rightmost column of",
"界的（请参见图3的最右列）。因此，只要限",
"Figure 3 . Note that for |C| = 4 only 12 out of 24",
"制包括原点，对Softmax层输入a施加任何范",
"regions can be intersected.",
"围限制都不会更改可行区域。然而，当我们引",
"当我们通过减小Softmax输入的d维度来",
"人一个偏置项时，我们也会得到有界区域（参",
"缩小Softmax瓶颈时，更多的排列变得不可",
"见附录中的图7，它对比了这两种情况）。因",
"行 (Good and Tideman, 1977; Kamiya and Take-",
"此，在这种情况下，Softmax层的输入比例也",
"mura.2005）。重要的是，如果我们选择|Cl和",
"很重要。如果输人没有足够大的范围，则会存",
"在存在但功能编码器无法到达的区域",
"d以及是否使用偏置项，则更改Softmax权重",
"的值将更改可行排列的集合，但不会更改集合",
"3.3",
"Exact Algorithm",
"见附录C",
"的基数E(Cover,1967;Smith,2014)。更多细节",
"给定一个由W和b参数化的softmax层，是",
"否有任何类是不可变的？我们首先描述一个缓",
"Corollary 2. Class Ct is unargmaxable when any",
"慢但精确的算法来回答这个问题",
"permutation that would rank class ct above all",
"一个精确的算法要么通过返回一个可行点",
"other classes is infeasible.",
"a：argmax（Wa+b）=Ct来证明类ct是可",
"argmaxable的，要么通过验证不存在这样的点",
"7when d = C - 1 we can still intersect all regions, be-",
"来证明ct是不可argmaxable的",
"cause the Braid Arrangement always has rank C - 1 (all its",
"normal vectors are perpendicular to the all ones vector 1 ).",
"要检查是否存在将Ct排在所有其他区域之"
] | {
"arxiv_id": "2203.06462",
"language": "zh",
"page_idx": 4
} |
||
2001.04642-zh-0 | 2001.04642-zh-0.jpg | Seeing the World in a Bag of Chips
Jeong Joon Park
Steven M. Seitz
Aleksander Holynski
University of Washington, Seattle $\{$ jjpark7,holynski,seitz $\}$ @cs.washington.edu
<FigureHere>
$(\mathbf{c})$ Zoom-in
(b) Estimated environment (top), ground truth (bottom)
(a) Input images
Figure 1: From a hand-held RGBD sequence of an object (a), we reconstruct an image of the surrounding environment (b, top) that closely resembles the real environment (b, bottom), entirely from the specular reflections. Note the reconstruction of fine details (c) such as a human figure and trees with fall colors through the window. We use the recovered environment for novel view rendering.
Abstract
We address the dual problems of novel view synthesis and environment reconstruction from hand-held RGBD sensors. Our contributions include 1) modeling highly specular objects, 2) modeling inter-reflections and Fresnel effects, and 3) enabling surface light field reconstruction with the same input needed to reconstruct shape alone. In cases where scene surface has a strong mirror-like material component, we generate highly detailed environment images, revealing room composition, objects, people, buildings, and trees visible through windows. Our approach yields state of the art view synthesis techniques, operates on low dynamic range imagery, and is robust to geometric and calibration errors.
光，窗户甚至外部物体的布局
在他们的 visual microphone 作品中，Davis et al. [14]展示了如何从一袋芯片中可见的微小振动中重建声音甚至对话。受他们工作的启发，我们展示了可以使用同一袋芯片来重建环境。但是，我们不是使用高速视频，而是使用商品深度传感器获得的 RGBD 视频进行操作
可视化环境与对反映该环境的场景进行建模的问题密切相关。我们解决了这两个问题；除了可视化房间之外，我们还试图从任何新的角度预测物体和场景的外观-i。e。,就像你在那里一样虚拟地探索场景。这个视图综合问题在计算机视觉和图形学方面有大量的文献， 但仍然有几个悬而未决的问题。其中最主要的是 1) 镜面，2)相互反射和 3)简单捕获。在本文中，我们基于surface light felds [72]的框架来解决所有这三个问题
我们的环境重建，我们称之为 specular reflectance maps(SRMs),表示与对象的镜面 BRDF 卷积的遥远环境地图。在对象具有强烈的镜面反射的情况下， 此 SRM 提供了清晰，详细的功能，如图。1。由于大多数场景由混合材质组成，因此每个场景都有多个基础 SRM。因此，我们重建了一组全局 SRM,以及场景表面的加权材质分割。基于恢复的 SRM,以及其他物理驱动的组件，我们构建了一个能够忠实地近似真实表面光场的神经渲染网络
我们的方法的主要贡献是能够使用 RGBD 相机单独
## 1. Introduction
物体发出的闪光可以揭示其形状和成分-无论是湿的还是干的，粗糙的还是抛光的，圆形的还是平坦的。然而，隐藏在高光图案中的也是环境的图像，通常是如此扭曲，以至于我们甚至没有意识到它的存在。值得注意的是，闪亮的薯片袋的图像(图。E1)包含足够的线索，能够重建房间的详细图像，包括通过窗户可见的灯
$^{\dagger}$ Video URL: https://youtu.be/9t\_Rx6n1HGA
1 | [
[
"0.3064",
"0.132",
"0.664",
"0.1332",
"0.664",
"0.1509",
"0.3064",
"0.1497"
],
[
"0.1772",
"0.1831",
"0.3203",
"0.1831",
"0.3203",
"0.1976",
"0.1772",
"0.1976"
],
[
"0.4004",
"0.1818",
"0.574",
"0.1831",
"0.574",
"0.1976",
"0.4004",
"0.1963"
],
[
"0.653",
"0.1831",
"0.789",
"0.1831",
"0.789",
"0.1976",
"0.653",
"0.1976"
],
[
"0.3464",
"0.2052",
"0.626",
"0.2052",
"0.626",
"0.2197",
"0.3464",
"0.2197"
],
[
"0.3381",
"0.2235",
"0.636",
"0.2235",
"0.636",
"0.2374",
"0.3381",
"0.2374"
],
[
"0.1405",
"0.3193",
"0.2427",
"0.3289",
"0.2385",
"0.3499",
"0.1365",
"0.3403"
],
[
"0.2214",
"0.351",
"0.2361",
"0.3528",
"0.2345",
"0.358",
"0.2197",
"0.356"
],
[
"0.3882",
"0.4224",
"0.7173",
"0.4224",
"0.7173",
"0.4363",
"0.3882",
"0.4363"
],
[
"0.7646",
"0.4224",
"0.8384",
"0.4236",
"0.8384",
"0.435",
"0.7646",
"0.4343"
],
[
"0.0801",
"0.4443",
"0.8906",
"0.4458",
"0.8906",
"0.4602",
"0.0801",
"0.459"
],
[
"0.1642",
"0.4248",
"0.259",
"0.4248",
"0.259",
"0.4363",
"0.1642",
"0.4363"
],
[
"0.0801",
"0.4602",
"0.8906",
"0.4602",
"0.8906",
"0.4734",
"0.0801",
"0.4734"
],
[
"0.0825",
"0.476",
"0.8896",
"0.476",
"0.8896",
"0.486",
"0.0825",
"0.486"
],
[
"0.0801",
"0.488",
"0.2566",
"0.488",
"0.2566",
"0.502",
"0.0801",
"0.502"
],
[
"0.2378",
"0.515",
"0.312",
"0.515",
"0.312",
"0.53",
"0.2378",
"0.53"
],
[
"0.505",
"0.5166",
"0.7183",
"0.5166",
"0.7183",
"0.531",
"0.505",
"0.531"
],
[
"0.0972",
"0.536",
"0.464",
"0.536",
"0.464",
"0.55",
"0.0972",
"0.55"
],
[
"0.5215",
"0.532",
"0.8896",
"0.532",
"0.8896",
"0.5454",
"0.5215",
"0.5454"
],
[
"0.0809",
"0.5503",
"0.468",
"0.5503",
"0.468",
"0.5645",
"0.0809",
"0.5645"
],
[
"0.5034",
"0.547",
"0.8906",
"0.5454",
"0.8906",
"0.56",
"0.5034",
"0.561"
],
[
"0.0809",
"0.565",
"0.4673",
"0.565",
"0.4673",
"0.579",
"0.0809",
"0.579"
],
[
"0.505",
"0.5625",
"0.8906",
"0.5625",
"0.8906",
"0.5757",
"0.505",
"0.5757"
],
[
"0.0809",
"0.5806",
"0.4673",
"0.5806",
"0.4673",
"0.5947",
"0.0809",
"0.5947"
],
[
"0.5063",
"0.577",
"0.8906",
"0.577",
"0.8906",
"0.5903",
"0.5063",
"0.5903"
],
[
"0.0809",
"0.5967",
"0.468",
"0.5967",
"0.468",
"0.61",
"0.0809",
"0.61"
],
[
"0.505",
"0.5913",
"0.8906",
"0.5913",
"0.8906",
"0.606",
"0.505",
"0.606"
],
[
"0.0825",
"0.613",
"0.4656",
"0.613",
"0.4656",
"0.6235",
"0.0825",
"0.6235"
],
[
"0.505",
"0.608",
"0.539",
"0.608",
"0.539",
"0.619",
"0.505",
"0.619"
],
[
"0.0801",
"0.627",
"0.468",
"0.627",
"0.468",
"0.64",
"0.0801",
"0.64"
],
[
"0.5215",
"0.6226",
"0.8906",
"0.6226",
"0.8906",
"0.636",
"0.5215",
"0.636"
],
[
"0.0801",
"0.64",
"0.4673",
"0.64",
"0.4673",
"0.655",
"0.0801",
"0.655"
],
[
"0.505",
"0.637",
"0.888",
"0.637",
"0.888",
"0.651",
"0.505",
"0.651"
],
[
"0.0801",
"0.657",
"0.4673",
"0.657",
"0.4673",
"0.6704",
"0.0801",
"0.6704"
],
[
"0.505",
"0.653",
"0.8896",
"0.653",
"0.8896",
"0.6665",
"0.505",
"0.6665"
],
[
"0.0809",
"0.6704",
"0.4656",
"0.6704",
"0.4656",
"0.685",
"0.0809",
"0.685"
],
[
"0.5034",
"0.6665",
"0.8906",
"0.6675",
"0.8906",
"0.6816",
"0.5034",
"0.681"
],
[
"0.0809",
"0.6875",
"0.468",
"0.6875",
"0.468",
"0.702",
"0.0809",
"0.702"
],
[
"0.5063",
"0.683",
"0.888",
"0.683",
"0.888",
"0.6963",
"0.5063",
"0.6963"
],
[
"0.0809",
"0.7026",
"0.4631",
"0.7026",
"0.4631",
"0.7163",
"0.0809",
"0.7163"
],
[
"0.505",
"0.6978",
"0.8906",
"0.6978",
"0.8906",
"0.711",
"0.505",
"0.711"
],
[
"0.0784",
"0.7173",
"0.468",
"0.7163",
"0.468",
"0.731",
"0.0784",
"0.732"
],
[
"0.505",
"0.7134",
"0.888",
"0.7134",
"0.888",
"0.7266",
"0.505",
"0.7266"
],
[
"0.5034",
"0.7266",
"0.888",
"0.728",
"0.888",
"0.742",
"0.5034",
"0.741"
],
[
"0.0801",
"0.733",
"0.21",
"0.733",
"0.21",
"0.747",
"0.0801",
"0.747"
],
[
"0.5215",
"0.7437",
"0.8906",
"0.7437",
"0.8906",
"0.757",
"0.5215",
"0.757"
],
[
"0.0801",
"0.761",
"0.2173",
"0.761",
"0.2173",
"0.7744",
"0.0801",
"0.7744"
],
[
"0.5034",
"0.7583",
"0.8906",
"0.7583",
"0.8906",
"0.773",
"0.5034",
"0.773"
],
[
"0.505",
"0.7734",
"0.888",
"0.7734",
"0.888",
"0.787",
"0.505",
"0.787"
],
[
"0.0972",
"0.785",
"0.4673",
"0.785",
"0.4673",
"0.799",
"0.0972",
"0.799"
],
[
"0.505",
"0.7886",
"0.8896",
"0.7886",
"0.8896",
"0.802",
"0.505",
"0.802"
],
[
"0.0809",
"0.8003",
"0.4673",
"0.8003",
"0.4673",
"0.8145",
"0.0809",
"0.8145"
],
[
"0.505",
"0.803",
"0.8896",
"0.803",
"0.8896",
"0.8174",
"0.505",
"0.8174"
],
[
"0.0825",
"0.815",
"0.4673",
"0.815",
"0.4673",
"0.829",
"0.0825",
"0.829"
],
[
"0.505",
"0.8184",
"0.8906",
"0.8184",
"0.8906",
"0.832",
"0.505",
"0.832"
],
[
"0.0809",
"0.8306",
"0.4673",
"0.8306",
"0.4673",
"0.8447",
"0.0809",
"0.8447"
],
[
"0.505",
"0.8335",
"0.8906",
"0.8335",
"0.8906",
"0.8477",
"0.505",
"0.8477"
],
[
"0.0809",
"0.845",
"0.4673",
"0.845",
"0.4673",
"0.8594",
"0.0809",
"0.8594"
],
[
"0.505",
"0.8486",
"0.8906",
"0.8486",
"0.8906",
"0.8623",
"0.505",
"0.8623"
],
[
"0.0809",
"0.8613",
"0.4673",
"0.8613",
"0.4673",
"0.875",
"0.0809",
"0.875"
],
[
"0.505",
"0.8643",
"0.701",
"0.8643",
"0.701",
"0.8774",
"0.505",
"0.8774"
],
[
"0.5195",
"0.879",
"0.8906",
"0.8774",
"0.8906",
"0.8926",
"0.5195",
"0.893"
],
[
"0.1013",
"0.8853",
"0.3962",
"0.886",
"0.3962",
"0.901",
"0.1013",
"0.8994"
],
[
"0.4805",
"0.9253",
"0.492",
"0.9253",
"0.492",
"0.935",
"0.4805",
"0.935"
]
] | [
"Seeing the World in a Bag of Chips",
"Jeong Joon Park",
"Aleksander Holynski",
"Steven M. Seitz",
"University of Washington, Seattle",
"[jjpark7,holynski,seitz @cs.washington.edu",
"",
"C",
"(b) Estimated environment (top), ground truth (bottom)",
"(c)Zoom-in",
"Figure l:From a hand-held RGBD sequence of an object (a),wereconstruct an image of the surrounding environment (b,",
"(a)Inputimages",
"top) that closely resembles the real environment (b, bottom), entirely from the specular reflections. Note the reconstruction",
"offinedetails (c)such asahuman figureandtreeswithfall colorsthrough thewindow.Weusetherecovered environment",
"for novel view rendering. ",
"Abstract",
"光，窗户甚至外部物体的布局",
"We address the dual problems of novel view syn+",
"在他们的visual microphone作品中，Davis etal.",
"thesis and environment reconstruction from hand-held",
"【14]展示了如何从一袋芯片中可见的微小振动中重建",
"RGBD sensors. Our contributions include 1) modeling",
"声音甚至对话。受他们工作的启发，我们展示了可以使",
"highly specular objects, 2) modeling inter-reflections",
"用同一袋芯片来重建环境。但是，我们不是使用高速视",
"and Fresnel effects, and 3) enabling surface light field",
"频，而是使用商品深度传感器获得的RGBD视频进行",
"reconstructionwiththesameinputneededtorecon-",
"操作",
"struct shape alone. In cases where scene surface has",
"可视化环境与对反映该环境的场景进行建模的问题",
"a strong mirror-like material component, we generate",
"密切相关。我们解决了这两个问题；除了可视化房间之",
"highly detailed environment images, revealing room",
"外，我们还试图从任何新的角度预测物体和场景的外",
"composition, objects, people, buildings, and trees vis-",
"观-i。e。，就像你在那里一样虚拟地探索场景。这个视",
"ible through windows. Our approach yields state of",
"图综合问题在计算机视觉和图形学方面有大量的文献",
"the art view synthesis techniques, operates on low dy",
"但仍然有几个悬而未决的问题。其中最主要的是1）镜",
"namic range imagery, and is robust to geometric and",
"面，2）相互反射和3）简单捕获。在本文中，我们基于",
"surface lightfields「72]的框架来解决所有这三个问题",
"calibration errors.",
"我们的环境重建，我们称之为specularreflectance",
"1:Introduction",
"maps（SRMs），表示与对象的镜面BRDF卷积的",
"遥远环境地图。在对象具有强烈的镜面反射的情况下",
"物体发出的闪光可以揭示其形状和成分-无论是湿的",
"此SRM提供了清晰，详细的功能，如图。1。由于",
"还是干的，粗糙的还是抛光的，圆形的还是平坦的。然",
"大多数场景由混合材质组成，因此每个场景都有多个",
"而，隐藏在高光图案中的也是环境的图像，通常是如此",
"基础SRM。因此，我们重建了一组全局SRM，以及场",
"扭曲，以至于我们甚至没有意识到它的存在。值得注意",
"景表面的加权材质分割。基于恢复的SRM，以及其他",
"的是，闪亮的薯片袋的图像（图。E1）包含足够的线",
"物理驱动的组件，我们构建了一个能够忠实地近似真",
"索，能够重建房间的详细图像，包括通过窗户可见的灯",
"实表面光场的神经染网络",
"我们的方法的主要贡献是能够使用RGBD相机单独",
"tVideo URL:https://youtu.be/9t_Rx6n1HGA",
"1"
] | {
"arxiv_id": "2001.04642",
"language": "zh",
"page_idx": 0
} |
||
2001.04642-zh-1 | 2001.04642-zh-1.jpg | 计算形状[54]所需的相同输人来重建表面光场。我们的方法的其他贡献包括对常规(低动态范围)图像进行操作的能力，以及对包含多个对象以及漫反射和镜面材质的一般，非凸，纹理场景的适用性。最后，我们发布了捕获反射对象的 RGBD 数据集，以促进对照明估计和基于图像的渲染的研究
我们指出，从对象的图像重建反射场景的能力打开了关于隐私的真实和有效的担忧。虽然我们的方法需要深度传感器，但未来的研究可能会导致对常规照片进行操作的方法。除了对人们进行可能的教育外，我们的工作还可以促进对隐私保护摄像头和安全技术的研究，这些技术可以主动识别和扰乱反射
# 2. Related Work
我们回顾了环境照明估计和用于对镜面表面进行建
模的新颖视图合成方法的相关工作
# 2.1. Environment Estimation
Single-View Estimation 捕获环境地图(图像)的最直接方法是通过光探针(e。g。,镜像球[16])或用 360 °相机[58]拍照。人眼球[56] 甚至可以在存在时用作光探针。但是，对于许多应用程序，光探针不可用，我们必须依靠场景本身中的现有线索
其他方法改为研究从一般场景的照片中恢复照明。由于此问题严重受限，因此这些方法通常依赖于人类输人[35,81] 或手动设计的“固有图像”先验，涉及照明， 材料和表面特性[36,6,5,7,45]
深度学习技术的最新发展促进了单视图估计的数据驱动方法。[20,19,66,41]学习从透视图像到广角全景图像的映射。其他方法训练专门为户外场景量身定制的模型[30,29] 。由于单视图问题是严重不适定的，因此大多数结果都是合理的，但通常是非验证的。与我们的工作密切相关，Georgoulis et al.[21] 重建了更高质量的环境图像，但在非常有限的假设下；无纹理的涂漆表面以及材料和分割的手动规范
Multi-View Estimation 对于平面反射器的特殊情况，层分离技术[68,65,77,26,25,32,80]可以实现反射环境的高质量重建，例如。g。,来自一个玻璃相框的视频。推断一般曲面的反射非常困难，即使对于人类也是如此，因为反射内容强烈且非线性地取决于表面形状和空间变化的材料特性。
许多研究人员试图从弯曲物体的多个图像中恢复low-frequency 照明。[85,57,47]推断球面谐波照明(跟随[62] ),以使用形状从阴影的原理来细化表面几何形状。[63] 联合优化重建场景的低频照明和 BRDF。虽然适用于近似光源方向，但这些模型无法捕获环境的详细图像
Wu et al.[73] ,与我们一样，使用手持式 RGBD 传感器恢复照明和反射性能。但是该方法只能重建单个， 浮动，凸起的物体，并且需要黑色背景。Dong et al. [17]从单个旋转物体的视频中生成高质量的环境图像。此方法假设实验室设置有机械旋转器，并将准确的几何形状手动注册到其视频中。同样，Xia et al.[74] 使
用带有校准图案的机械臂旋转物体。作者指出，高度镜面会造成麻烦，因此将其真实物体样本限制为大部分粗糙，有光泽的材料。相比之下，我们的方法与手持相机一起操作，用于广泛的多对象场景，并且旨在支持镜面
# 2.2. Novel View Synthesis
在这里，我们重点介绍能够从新的角度对 specular
reflections 进行建模的方法
Image-based Rendering 光场方法[24,43,10,72,13] 以从密集采样的视点进行费力的场景捕获为代价，实现了镜面的高度逼真的视图。Chen et al.[8] 使用神经网络回归表面光场，以减少所需的视图数量，但需要使用机械系统捕获的整个半球的样本。Park et al.[58] 通过应用参数 BRDF 模型避免密集的半球形视图采样， 但假设已知照明
最近的工作将卷积神经网络(CNN)应用于基于图像的渲染[18,50] 。Hedman et al. [28] 用 CNN 学习的混合权重代替了 IBR 系统的传统视图混合启发式。尽管如此，新颖的视图仍由现有的捕获像素组成，因此无法合成未观察到的镜面高光。最近，[2,69] 通过将学习的特征附加到 2D 纹理贴图[69]或 3D 点云[2] 来增强传统的渲染管道，并获得高质量的视图合成结果。尽管如此，这些功能还是经过专门优化以适合输入视图， 并且不能很好地推断为新颖的视图。最近基于学习的方法从一小部分图像[51,67,11,34,82]中实现了令人印象深刻的局部(相对于半球)光场重建
BRDF Estimation Methods 合成新颖视图的另一种方法是恢复固有的表面反射函数，称为 BRDFs [55] 。通常，恢复表面 BRDF 是一项艰巨的任务，因为它涉及反转复杂的光传输过程。因此，现有的反射率捕获方法对操作范围施加了限制：e。g。,孤立的单个对象[73,17] , 已知或受控的照明 [58, 15, 42, 83, 76] ,单视图表面(相对于完整的 3D 网格)[22,44],闪光摄影[1,40,53]或空间恒定材料[49,38]
Interreflections 很少有视图合成技术支持相互反射。对一般多对象场景进行建模需要求解全局照明 (e。g。阴影或相互反射),这对现实世界输入[4]的缺陷是困难和敏感的。同样，Lombardi et al.[46] 模型多反弹照明，但具有明显的伪影，并将其结果限制为大多数纹理均匀的对象。Zhang et al. [78]需要手动注释灯光类型和位置
# 3. Technical Approach
我们的系统将静态场景的视频和 3D 网格(通过Newcombe et al.[54]获得)作为输入，并自动重建环境图像以及场景外观模型，从而实现新颖的视图合成。我们的方法在镜面反射场景中表现出色，并同时考虑了镜面反射和菲涅耳效应。我们方法的一个关键优势是使用手持相机轻松，随意地捕获数据；我们重建了环境地图和表面光场，其输入与单独重建几何形状所需的输入相同，例如。g。,使用[54] | [
[
"0.0825",
"0.09344",
"0.4673",
"0.09344",
"0.4673",
"0.1067",
"0.0825",
"0.1067"
],
[
"0.505",
"0.09344",
"0.8906",
"0.09344",
"0.8906",
"0.1067",
"0.505",
"0.1067"
],
[
"0.08417",
"0.10986",
"0.4656",
"0.10986",
"0.4656",
"0.1206",
"0.08417",
"0.1206"
],
[
"0.5063",
"0.10986",
"0.888",
"0.10986",
"0.888",
"0.1206",
"0.5063",
"0.1206"
],
[
"0.0825",
"0.125",
"0.4656",
"0.125",
"0.4656",
"0.1351",
"0.0825",
"0.1351"
],
[
"0.505",
"0.1237",
"0.8896",
"0.1237",
"0.8896",
"0.137",
"0.505",
"0.137"
],
[
"0.0825",
"0.1401",
"0.4673",
"0.1401",
"0.4673",
"0.1509",
"0.0825",
"0.1509"
],
[
"0.505",
"0.1401",
"0.8896",
"0.1401",
"0.8896",
"0.1509",
"0.505",
"0.1509"
],
[
"0.0825",
"0.154",
"0.468",
"0.154",
"0.468",
"0.1674",
"0.0825",
"0.1674"
],
[
"0.505",
"0.1559",
"0.5215",
"0.1559",
"0.5215",
"0.1654",
"0.505",
"0.1654"
],
[
"0.0825",
"0.1692",
"0.2769",
"0.1692",
"0.2769",
"0.1799",
"0.0825",
"0.1799"
],
[
"0.0972",
"0.1843",
"0.468",
"0.1843",
"0.468",
"0.1976",
"0.0972",
"0.1976"
],
[
"0.5034",
"0.1799",
"0.705",
"0.1799",
"0.705",
"0.1932",
"0.5034",
"0.1932"
],
[
"0.0825",
"0.1976",
"0.4673",
"0.1989",
"0.4673",
"0.2122",
"0.0825",
"0.2108"
],
[
"0.5215",
"0.2026",
"0.8896",
"0.2026",
"0.8896",
"0.2166",
"0.5215",
"0.2166"
],
[
"0.08417",
"0.2153",
"0.464",
"0.2153",
"0.464",
"0.226",
"0.08417",
"0.226"
],
[
"0.5034",
"0.2197",
"0.701",
"0.2197",
"0.701",
"0.2299",
"0.5034",
"0.2299"
],
[
"0.0825",
"0.2291",
"0.4673",
"0.2291",
"0.4673",
"0.2424",
"0.0825",
"0.2424"
],
[
"0.5073",
"0.2361",
"0.8896",
"0.2361",
"0.8896",
"0.25",
"0.5073",
"0.25"
],
[
"0.0809",
"0.2437",
"0.4656",
"0.2437",
"0.4656",
"0.2568",
"0.0809",
"0.2568"
],
[
"0.505",
"0.252",
"0.8906",
"0.252",
"0.8906",
"0.2651",
"0.505",
"0.2651"
],
[
"0.0801",
"0.2588",
"0.357",
"0.2588",
"0.357",
"0.2727",
"0.0801",
"0.2727"
],
[
"0.5063",
"0.2683",
"0.8896",
"0.2683",
"0.8896",
"0.279",
"0.5063",
"0.279"
],
[
"0.0801",
"0.2874",
"0.2263",
"0.2874",
"0.2263",
"0.3005",
"0.0801",
"0.3005"
],
[
"0.505",
"0.2822",
"0.8906",
"0.2822",
"0.8906",
"0.2954",
"0.505",
"0.2954"
],
[
"0.505",
"0.2942",
"0.8906",
"0.2954",
"0.8906",
"0.31",
"0.505",
"0.3093"
],
[
"0.0972",
"0.3125",
"0.468",
"0.3125",
"0.468",
"0.3257",
"0.0972",
"0.3257"
],
[
"0.5034",
"0.3113",
"0.892",
"0.3113",
"0.892",
"0.3245",
"0.5034",
"0.3245"
],
[
"0.0809",
"0.327",
"0.326",
"0.327",
"0.326",
"0.3403",
"0.0809",
"0.3403"
],
[
"0.505",
"0.327",
"0.621",
"0.327",
"0.621",
"0.3403",
"0.505",
"0.3403"
],
[
"0.5215",
"0.3416",
"0.8896",
"0.3416",
"0.8896",
"0.3547",
"0.5215",
"0.3547"
],
[
"0.0809",
"0.3528",
"0.3064",
"0.3528",
"0.3064",
"0.363",
"0.0809",
"0.363"
],
[
"0.505",
"0.3567",
"0.8896",
"0.3567",
"0.8896",
"0.3706",
"0.505",
"0.3706"
],
[
"0.0866",
"0.375",
"0.4673",
"0.375",
"0.4673",
"0.3882",
"0.0866",
"0.3882"
],
[
"0.505",
"0.3718",
"0.8906",
"0.3718",
"0.8906",
"0.385",
"0.505",
"0.385"
],
[
"0.0784",
"0.3882",
"0.4697",
"0.387",
"0.4697",
"0.4048",
"0.0784",
"0.406"
],
[
"0.505",
"0.387",
"0.8906",
"0.387",
"0.8906",
"0.401",
"0.505",
"0.401"
],
[
"0.0801",
"0.404",
"0.468",
"0.404",
"0.468",
"0.4172",
"0.0801",
"0.4172"
],
[
"0.505",
"0.404",
"0.888",
"0.404",
"0.888",
"0.414",
"0.505",
"0.414"
],
[
"0.0809",
"0.4192",
"0.468",
"0.4192",
"0.468",
"0.433",
"0.0809",
"0.433"
],
[
"0.505",
"0.4172",
"0.8906",
"0.4172",
"0.8906",
"0.4312",
"0.505",
"0.4312"
],
[
"0.0825",
"0.4363",
"0.3245",
"0.4363",
"0.3245",
"0.4463",
"0.0825",
"0.4463"
],
[
"0.5063",
"0.433",
"0.8896",
"0.433",
"0.8896",
"0.4463",
"0.5063",
"0.4463"
],
[
"0.0989",
"0.4495",
"0.4673",
"0.4495",
"0.4673",
"0.4634",
"0.0989",
"0.4634"
],
[
"0.505",
"0.4475",
"0.888",
"0.4475",
"0.888",
"0.461",
"0.505",
"0.461"
],
[
"0.08417",
"0.4666",
"0.4656",
"0.4666",
"0.4656",
"0.4766",
"0.08417",
"0.4766"
],
[
"0.5034",
"0.4622",
"0.8906",
"0.4622",
"0.8906",
"0.476",
"0.5034",
"0.476"
],
[
"0.0809",
"0.4792",
"0.4697",
"0.4797",
"0.4697",
"0.4937",
"0.0809",
"0.4924"
],
[
"0.5063",
"0.4792",
"0.8896",
"0.4792",
"0.8896",
"0.4893",
"0.5063",
"0.4893"
],
[
"0.0801",
"0.4937",
"0.3064",
"0.4944",
"0.3064",
"0.509",
"0.0801",
"0.5083"
],
[
"0.505",
"0.4924",
"0.7983",
"0.4924",
"0.7983",
"0.506",
"0.505",
"0.506"
],
[
"0.0972",
"0.5103",
"0.4673",
"0.5103",
"0.4673",
"0.5234",
"0.0972",
"0.5234"
],
[
"0.509",
"0.511",
"0.8906",
"0.511",
"0.8906",
"0.5244",
"0.509",
"0.5244"
],
[
"0.0809",
"0.526",
"0.4673",
"0.526",
"0.4673",
"0.539",
"0.0809",
"0.539"
],
[
"0.505",
"0.526",
"0.8906",
"0.526",
"0.8906",
"0.539",
"0.505",
"0.539"
],
[
"0.08417",
"0.5425",
"0.4656",
"0.5425",
"0.4656",
"0.553",
"0.08417",
"0.553"
],
[
"0.505",
"0.5415",
"0.8906",
"0.5415",
"0.8906",
"0.5547",
"0.505",
"0.5547"
],
[
"0.0809",
"0.556",
"0.4673",
"0.556",
"0.4673",
"0.5693",
"0.0809",
"0.5693"
],
[
"0.5063",
"0.558",
"0.8896",
"0.558",
"0.8896",
"0.5684",
"0.5063",
"0.5684"
],
[
"0.0809",
"0.5684",
"0.4673",
"0.5693",
"0.4673",
"0.584",
"0.0809",
"0.5835"
],
[
"0.5034",
"0.571",
"0.8896",
"0.5713",
"0.8896",
"0.586",
"0.5034",
"0.5854"
],
[
"0.0809",
"0.5854",
"0.4673",
"0.5854",
"0.4673",
"0.5986",
"0.0809",
"0.5986"
],
[
"0.505",
"0.586",
"0.8906",
"0.586",
"0.8906",
"0.601",
"0.505",
"0.601"
],
[
"0.0825",
"0.601",
"0.468",
"0.601",
"0.468",
"0.6143",
"0.0825",
"0.6143"
],
[
"0.505",
"0.6016",
"0.8906",
"0.6016",
"0.8906",
"0.6157",
"0.505",
"0.6157"
],
[
"0.08417",
"0.617",
"0.308",
"0.617",
"0.308",
"0.6274",
"0.08417",
"0.6274"
],
[
"0.505",
"0.617",
"0.773",
"0.617",
"0.773",
"0.631",
"0.505",
"0.631"
],
[
"0.08496",
"0.629",
"0.4673",
"0.631",
"0.4673",
"0.646",
"0.08496",
"0.643"
],
[
"0.5073",
"0.634",
"0.8896",
"0.6353",
"0.8896",
"0.6484",
"0.5073",
"0.648"
],
[
"0.0809",
"0.646",
"0.4673",
"0.646",
"0.4673",
"0.659",
"0.0809",
"0.659"
],
[
"0.505",
"0.651",
"0.8906",
"0.651",
"0.8906",
"0.664",
"0.505",
"0.664"
],
[
"0.0809",
"0.66",
"0.4673",
"0.659",
"0.4673",
"0.674",
"0.0809",
"0.675"
],
[
"0.5024",
"0.664",
"0.888",
"0.6655",
"0.888",
"0.6787",
"0.5024",
"0.678"
],
[
"0.0825",
"0.6753",
"0.4673",
"0.6753",
"0.4673",
"0.6895",
"0.0825",
"0.6895"
],
[
"0.5034",
"0.6787",
"0.8906",
"0.6797",
"0.8906",
"0.6943",
"0.5034",
"0.6934"
],
[
"0.08496",
"0.6924",
"0.4673",
"0.6924",
"0.4673",
"0.7026",
"0.08496",
"0.7026"
],
[
"0.5034",
"0.6943",
"0.888",
"0.6943",
"0.888",
"0.7085",
"0.5034",
"0.7085"
],
[
"0.0809",
"0.7056",
"0.2656",
"0.7056",
"0.2656",
"0.7197",
"0.0809",
"0.7197"
],
[
"0.505",
"0.711",
"0.8906",
"0.711",
"0.8906",
"0.7246",
"0.505",
"0.7246"
],
[
"0.0972",
"0.72",
"0.4673",
"0.72",
"0.4673",
"0.7344",
"0.0972",
"0.7344"
],
[
"0.505",
"0.728",
"0.5537",
"0.728",
"0.5537",
"0.7393",
"0.505",
"0.7393"
],
[
"0.0801",
"0.7363",
"0.4673",
"0.7363",
"0.4673",
"0.75",
"0.0801",
"0.75"
],
[
"0.0809",
"0.752",
"0.4673",
"0.752",
"0.4673",
"0.765",
"0.0809",
"0.765"
],
[
"0.5024",
"0.757",
"0.701",
"0.757",
"0.701",
"0.7715",
"0.5024",
"0.7715"
],
[
"0.0809",
"0.765",
"0.4631",
"0.765",
"0.4631",
"0.779",
"0.0809",
"0.779"
],
[
"0.0809",
"0.782",
"0.468",
"0.782",
"0.468",
"0.7954",
"0.0809",
"0.7954"
],
[
"0.5215",
"0.782",
"0.8906",
"0.782",
"0.8906",
"0.7954",
"0.5215",
"0.7954"
],
[
"0.0801",
"0.7954",
"0.1626",
"0.7974",
"0.1626",
"0.8125",
"0.0801",
"0.81"
],
[
"0.5034",
"0.797",
"0.8906",
"0.797",
"0.8906",
"0.81",
"0.5034",
"0.81"
],
[
"0.0972",
"0.811",
"0.4673",
"0.811",
"0.4673",
"0.8247",
"0.0972",
"0.8247"
],
[
"0.505",
"0.811",
"0.892",
"0.811",
"0.892",
"0.8247",
"0.505",
"0.8247"
],
[
"0.0809",
"0.827",
"0.468",
"0.827",
"0.468",
"0.8403",
"0.0809",
"0.8403"
],
[
"0.5034",
"0.827",
"0.8896",
"0.827",
"0.8896",
"0.8403",
"0.5034",
"0.8403"
],
[
"0.0825",
"0.8433",
"0.4656",
"0.8433",
"0.4656",
"0.854",
"0.0825",
"0.854"
],
[
"0.509",
"0.8433",
"0.8896",
"0.8433",
"0.8896",
"0.854",
"0.509",
"0.854"
],
[
"0.0825",
"0.8594",
"0.468",
"0.8594",
"0.468",
"0.869",
"0.0825",
"0.869"
],
[
"0.5034",
"0.856",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.5034",
"0.869"
],
[
"0.0809",
"0.8726",
"0.4673",
"0.8726",
"0.4673",
"0.886",
"0.0809",
"0.886"
],
[
"0.5024",
"0.8706",
"0.8906",
"0.872",
"0.8906",
"0.886",
"0.5024",
"0.8853"
],
[
"0.0801",
"0.886",
"0.468",
"0.887",
"0.468",
"0.901",
"0.0801",
"0.8994"
],
[
"0.5024",
"0.886",
"0.749",
"0.887",
"0.749",
"0.9014",
"0.5024",
"0.901"
]
] | [
"计算形状[541所需的相同输人来重建表面光场。我们",
"用带有校准图案的机械臂旋转物体。作者指出，高度镜",
"的方法的其他贡献包括对常规（低动态范围）图像进行",
"面会造成麻烦，因此将其真实物体样本限制为大部分",
"操作的能力，以及对包含多个对象以及漫反射和镜面",
"粗糙，有光泽的材料。相比之下，我们的方法与手持相",
"材质的一般，非凸，纹理场景的适用性。最后，我们发",
"机一起操作，用于广泛的多对象场景，并且旨在支持镜",
"布了捕获反射对象的RGBD数据集，以促进对照明估",
"面",
"计和基于图像的染的研究",
"我们指出，从对象的图像重建反射场景的能力打开",
"2.2. Novel View Synthesis",
"了关于隐私的真实和有效的担忧。虽然我们的方法需",
"在这里，我们重点介绍能够从新的角度对specular",
"要深度传感器，但未来的研究可能会导致对常规照片",
"reflections进行建模的方法",
"进行操作的方法。除了对人们进行可能的教育外，我们",
"Image-basedRendering光场方法[24,43,10,72,13]",
"的工作还可以促进对隐私保护摄像头和安全技术的研",
"以从密集采样的视点进行费力的场景捕获为代价，实",
"究，这些技术可以主动识别和扰乱反射",
"现了镜面的高度逼真的视图。Chenetal.8使用神经",
"2. Related Work",
"网络回归表面光场，以减少所需的视图数量，但需要使",
"用机械系统捕获的整个半球的样本。Parketal.[58]通",
"我们回顾了环境照明估计和用于对镜面表面进行建",
"过应用参数BRDF模型避免密集的半球形视图采样，",
"模的新颖视图合成方法的相关工作",
"但假设已知照明",
"最近的工作将卷积神经网络（CNN）应用于基于图",
"2.1.EnvironmentEstimation",
"像的煊染[18，50]。Hedmanetal.[28]用CNN学习的",
"Single-ViewEstimation捕获环境地图（图像）的最",
"混合权重代替了IBR系统的传统视图混合启发式。尽",
"直接方法是通过光探针（e。g。，镜像球[16]）或用360",
"管如此，新颖的视图仍由现有的捕获像素组成，因此无",
"。相机「581拍照。人眼球[56]甚至可以在存在时用作",
"法合成未观察到的镜面高光。最近，269通过将学习",
"光探针。但是，对于许多应用程序，光探针不可用，我",
"的特征附加到2D纹理贴图【69]或3D点云[2]来增强",
"们必须依靠场景本身中的现有线索",
"传统的染管道，并获得高质量的视图合成结果。尽",
"其他方法改为研究从一般场景的照片中恢复照明。由",
"管如此，这些功能还是经过专门优化以适合输人视图",
"于此问题严重受限，因此这些方法通常依赖于人类输",
"并且不能很好地推断为新颖的视图。最近基于学习的",
"入35，81]或手动设计的“固有图像”先验，涉及照明，",
"方法从一小部分图像51，67，11，34.82中实现了今人",
"材料和表面特性[36，6，5，7，45]",
"印象深刻的局部（相对于半球）光场重建",
"深度学习技术的最新发展促进了单视图估计的数据",
"BRDFEstimationMethods合成新颖视图的另一种",
"驱动方法。[20，19，66，41]学习从透视图像到广角全景",
"方法是恢复固有的表面反射函数，称为BRDFs[551。",
"图像的映射。其他方法训练专门为户外场景量身定制",
"通常，恢复表面BRDF是一项艰巨的任务，因为它涉",
"的模型[30，29]。由于单视图问题是严重不适定的，因",
"及反转复杂的光传输过程。因此，现有的反射率捕获",
"此大多数结果都是合理的，但通常是非验证的。与我们",
"方法对操作范围施加了限制：e。g。，孤立的单个对象",
"的工作密切相关，Georgoulisetal.[21]重建了更高质",
"[73，17]，已知或受控的照明[58，15，42，83，76]，单",
"量的环境图像，但在非常有限的假设下；无纹理的涂漆",
"视图表面（相对于完整的3D网格）【22，44】，闪光摄",
"表面以及材料和分割的手动规范",
"影[1，40，53]或空间恒定材料【49，38]",
"Multi-ViewEstimation对于平面反射器的特殊情",
"Interreflections很少有视图合成技术支持相互反射。",
"况，层分离技术[68，65，77，26，25，32,80]可以实现反",
"对一般多对象场景进行建模需要求解全局照明(e。g。",
"射环境的高质量重建，例如。g。，来自一个玻璃相框的",
"阴影或相互反射)，这对现实世界输入【41的缺陷是困",
"视频。推断一般曲面的反射非常困难，即使对于人类也",
"难和敏感的。同样，Lombardietal.[46]模型多反弹照",
"是如此：因为反射内容强烈且非线性地取决于表面形",
"明，但具有明显的伪影，并将其结果限制为大多数纹理",
"状和空间变化的材料特性，",
"均匀的对象。Zhangetal.[78]需要手动注释灯光类型",
"许多研究人员试图从弯曲物体的多个图像中恢复",
"和位置",
"low-frequency照明。[85，57，47]推断球面谐波照明",
"（跟随[62]），以使用形状从阴影的原理来细化表面几",
"3. Technical Approach",
"何形状。63]联合优化重建场景的低频照明和BRDF",
"虽然适用于近似光源方向，但这些模型无法捕获环境",
"我们的系统将静态场景的视频和3D网格（通过",
"的详细图像",
"Newcombeetal.54l获得）作为输人，并自动重建环",
"Wuetal.「73]，与我们一样，使用手持式RGBD传",
"境图像以及场景外观模型，从而实现新颖的视图合成。",
"感器恢复照明和反射性能。但是该方法只能重建单个，",
"我们的方法在镜面反射场景中表现出色，并同时考虑",
"浮动，凸起的物体，并且需要黑色背景。Dongetal.",
"了镜面反射和菲涅耳效应。我们方法的一个关键优势",
"17从单个旋转物体的视频中生成高质量的环境图像。",
"是使用手持相机轻松，随意地捕获数据；我们重建了环",
"此方法假设实验室设置有机械旋转器，并将准确的几",
"境地图和表面光场，其输人与单独重建几何形状所需",
"何形状手动注册到其视频中。同样，Xiaetal.[74]使",
"的输人相同，例如。g。，使用[54]"
] | {
"arxiv_id": "2001.04642",
"language": "zh",
"page_idx": 1
} |
||
2001.04642-zh-4 | 2001.04642-zh-4.jpg | <FigureHere>
<FigureHere>
(h) Zoom-in(GT)
Figure $5{:\text{ Sample results of recovered SRMs and material weights. Given input video frames (a), we recover global SRMs}}$ (c) and their linear combination weights (b) from the optimization of equation ( 7 ). The scenes presented here have two material bases, visualized with red and green channels. Estimated SRMs (c) corresponding to the shiny object surface (green channel) correctly capture the light sources of the scenes, shown in the reference panorama images (d). For both scenes the SRMs corresponding to the red channel is mostly black, thus not shown, as the surface is mostly diffuse. The recovered SRM of (c) overemphasizes blue channel due to oversaturation in input images. Third row shows estimation result from a video of the same bag of chips (first row) under different lighting. Close inspection of the recovered environment (g) reveals many scene details, including foors in a nearby building visible through the window.
而忽略了相互反射。虽然对完整，全局，漫射+镜面光传输进行建模是棘手的，但我们可以通过光线跟踪第一反弹图像 (FBI) 来近似一阶相互反射，如下所示。对于要渲染的虚拟视点中的每个像素 u , 通过 u 从相机中心投射光线。如果我们现在假装每个场景表面都是完美的镜子，则该光线可能会反弹多次并与多个表面相交。令$x_2$为该光线与场景的第二个交点。用$x_2$ 的漫反射颜色在 FBI 中的 u 处渲染像素，如果没有第二个交点，则用黑色渲染像素(图。4(d))
可以通过将 FBI 与 BRDF 卷积来建模有光泽 (不完美的镜子)的相互反射。但是，严格来说，应在 angular domain [62]而不是图像空间中过滤互反射图像，即。e。,人射光沿中心为反射光线方向$\omega_r$的镜面反射瓣的卷积。给定$\omega_r$,可以通过卷积由$\omega_r$加权的 FBI 图像来近似图像空间中的角域卷积。但是，由于我们不知道镜面核，因此我们让网络使用$\omega_r$作为指导来推断权重。我们将每个像素的$\omega_r$编码为三通道图像$R_P$ (图。4(e))
们引入了基于对抗神经网络的生成器$g$和鉴别器$d$来渲染逼真的照片。$g$将我们对当前视图的漫反射$D_P$和镜面$S_P$分量的最佳预测(从公式(7)获得),以及将在本节稍后定义的互反射和菲涅耳项$FBI$ , $R_P$和$FCI$作为输入
因此，生成器$g$接收$C_P=(D_P,S_P,FBI,R_P,FCI)$ 作为输入并输出视图的预测，而鉴别器$d$则对其真实性进行评分。我们使用像素化$L_1$ ,感知损失$L_p[9]$和对抗损失[31] 的组合，如 Sec 中所述。3.2:
$$g^{*}=\arg\min_{g}\max_{d}\lambda_{G}\bar{\mathcal{L}}_{G}(g,d)+\lambda_{p}\bar{\mathcal{L}}_{p}(g)+\lambda_{1}\bar{\mathcal{L}}_{1}(g),$$
其中$\bar{\mathcal{L}}_p(g)=\frac1{|\mathcal{F}|}\sum_{P\in\mathcal{F}}\mathcal{L}_p(g(C_P),G_P)$是所有输入图像的感知损失的平均值，$\mathcal{L}_G(g,d)$和$\bar{\mathcal{L}}_1(g)$类似地定义为跨帧的平均损失。请注意，此渲染器$g$是 scene specific,仅在特定场景的图像上进行训练，以推断同一场景的新视图，就像神经渲染社区[50,69,2]中通常做的那样
Modeling Interreflections and Fresnel Effects 方程式$\mathbb{E}(\color{red}2$ )仅对环境对每个表面点的直接照明进行建模，
菲涅耳效果使高光在近掠视角下更强，对于逼真的渲染非常重要。菲涅耳系数按照 [64] 进行近似$:R(\alpha)=$ | [
[
"0.08496",
"0.25",
"0.2157",
"0.25",
"0.2157",
"0.2632",
"0.08496",
"0.2632"
],
[
"0.392",
"0.25",
"0.5146",
"0.25",
"0.5146",
"0.2632",
"0.392",
"0.2632"
],
[
"0.6914",
"0.25",
"0.8022",
"0.25",
"0.8022",
"0.2632",
"0.6914",
"0.2632"
],
[
"0.1936",
"0.262",
"0.2443",
"0.262",
"0.2443",
"0.2734",
"0.1936",
"0.2734"
],
[
"0.2507",
"0.2507",
"0.3057",
"0.2507",
"0.3057",
"0.262",
"0.2507",
"0.262"
],
[
"0.6274",
"0.2966",
"0.6567",
"0.2966",
"0.6567",
"0.305",
"0.6274",
"0.305"
],
[
"0.5024",
"0.3447",
"0.613",
"0.3435",
"0.613",
"0.358",
"0.5024",
"0.3591"
],
[
"0.7695",
"0.3447",
"0.8813",
"0.3435",
"0.8813",
"0.358",
"0.7695",
"0.3591"
],
[
"0.1323",
"0.3806",
"0.1528",
"0.3806",
"0.1528",
"0.3894",
"0.1323",
"0.3894"
],
[
"0.1569",
"0.3452",
"0.2769",
"0.3435",
"0.2769",
"0.3567",
"0.1569",
"0.3591"
],
[
"0.0907",
"0.3997",
"0.1969",
"0.3997",
"0.1969",
"0.4216",
"0.0907",
"0.4216"
],
[
"0.09235",
"0.4268",
"0.1993",
"0.4268",
"0.1993",
"0.44",
"0.09235",
"0.44"
],
[
"0.0809",
"0.4526",
"0.8906",
"0.4526",
"0.8906",
"0.4666",
"0.0809",
"0.4666"
],
[
"0.0809",
"0.4673",
"0.892",
"0.4673",
"0.892",
"0.481",
"0.0809",
"0.481"
],
[
"0.0801",
"0.481",
"0.892",
"0.481",
"0.892",
"0.4944",
"0.0801",
"0.4944"
],
[
"0.0809",
"0.4956",
"0.892",
"0.4956",
"0.892",
"0.509",
"0.0809",
"0.509"
],
[
"0.0801",
"0.509",
"0.892",
"0.509",
"0.892",
"0.523",
"0.0801",
"0.523"
],
[
"0.0809",
"0.523",
"0.892",
"0.523",
"0.892",
"0.536",
"0.0809",
"0.536"
],
[
"0.0809",
"0.537",
"0.892",
"0.537",
"0.892",
"0.5503",
"0.0809",
"0.5503"
],
[
"0.0801",
"0.5503",
"0.6436",
"0.5503",
"0.6436",
"0.5645",
"0.0801",
"0.5645"
],
[
"0.0825",
"0.5913",
"0.4673",
"0.5913",
"0.4673",
"0.605",
"0.0825",
"0.605"
],
[
"0.5034",
"0.5903",
"0.8906",
"0.5913",
"0.8906",
"0.606",
"0.5034",
"0.605"
],
[
"0.0809",
"0.607",
"0.468",
"0.607",
"0.468",
"0.6206",
"0.0809",
"0.6206"
],
[
"0.505",
"0.607",
"0.8906",
"0.607",
"0.8906",
"0.6206",
"0.505",
"0.6206"
],
[
"0.0825",
"0.622",
"0.4656",
"0.622",
"0.4656",
"0.6353",
"0.0825",
"0.6353"
],
[
"0.5034",
"0.6206",
"0.888",
"0.622",
"0.888",
"0.636",
"0.5034",
"0.6353"
],
[
"0.0825",
"0.637",
"0.468",
"0.637",
"0.468",
"0.651",
"0.0825",
"0.651"
],
[
"0.505",
"0.637",
"0.8896",
"0.637",
"0.8896",
"0.651",
"0.505",
"0.651"
],
[
"0.0825",
"0.653",
"0.1871",
"0.653",
"0.1871",
"0.6655",
"0.0825",
"0.6655"
],
[
"0.505",
"0.6514",
"0.8906",
"0.6514",
"0.8906",
"0.6655",
"0.505",
"0.6655"
],
[
"0.0972",
"0.6675",
"0.4673",
"0.6675",
"0.4673",
"0.681",
"0.0972",
"0.681"
],
[
"0.505",
"0.6675",
"0.8906",
"0.6675",
"0.8906",
"0.6816",
"0.505",
"0.6816"
],
[
"0.0825",
"0.683",
"0.4656",
"0.683",
"0.4656",
"0.6963",
"0.0825",
"0.6963"
],
[
"0.505",
"0.683",
"0.8896",
"0.683",
"0.8896",
"0.6963",
"0.505",
"0.6963"
],
[
"0.0809",
"0.696",
"0.4673",
"0.6963",
"0.4673",
"0.711",
"0.0809",
"0.7104"
],
[
"0.505",
"0.6963",
"0.8896",
"0.6963",
"0.8896",
"0.7104",
"0.505",
"0.7104"
],
[
"0.0809",
"0.712",
"0.3962",
"0.712",
"0.3962",
"0.726",
"0.0809",
"0.726"
],
[
"0.5107",
"0.712",
"0.814",
"0.712",
"0.814",
"0.726",
"0.5107",
"0.726"
],
[
"0.5195",
"0.726",
"0.8906",
"0.7266",
"0.8906",
"0.741",
"0.5195",
"0.7407"
],
[
"0.08826",
"0.742",
"0.4656",
"0.742",
"0.4656",
"0.757",
"0.08826",
"0.757"
],
[
"0.505",
"0.742",
"0.8906",
"0.742",
"0.8906",
"0.757",
"0.505",
"0.757"
],
[
"0.505",
"0.7583",
"0.8906",
"0.7583",
"0.8906",
"0.7715",
"0.505",
"0.7715"
],
[
"0.0809",
"0.7734",
"0.468",
"0.7734",
"0.468",
"0.7915",
"0.0809",
"0.7915"
],
[
"0.505",
"0.7734",
"0.8896",
"0.7734",
"0.8896",
"0.787",
"0.505",
"0.787"
],
[
"0.0825",
"0.7925",
"0.4673",
"0.7925",
"0.4673",
"0.8066",
"0.0825",
"0.8066"
],
[
"0.505",
"0.7886",
"0.8906",
"0.7886",
"0.8906",
"0.802",
"0.505",
"0.802"
],
[
"0.08417",
"0.808",
"0.4673",
"0.808",
"0.4673",
"0.8213",
"0.08417",
"0.8213"
],
[
"0.505",
"0.803",
"0.8906",
"0.803",
"0.8906",
"0.8164",
"0.505",
"0.8164"
],
[
"0.0784",
"0.8228",
"0.4673",
"0.8213",
"0.4673",
"0.836",
"0.0784",
"0.837"
],
[
"0.505",
"0.8184",
"0.8906",
"0.8184",
"0.8906",
"0.832",
"0.505",
"0.832"
],
[
"0.0907",
"0.8384",
"0.4673",
"0.8384",
"0.4673",
"0.8516",
"0.0907",
"0.8516"
],
[
"0.505",
"0.834",
"0.892",
"0.834",
"0.892",
"0.8477",
"0.505",
"0.8477"
],
[
"0.0825",
"0.854",
"0.1478",
"0.854",
"0.1478",
"0.8657",
"0.0825",
"0.8657"
],
[
"0.5034",
"0.8486",
"0.546",
"0.8486",
"0.546",
"0.863",
"0.5034",
"0.863"
],
[
"0.5215",
"0.863",
"0.8906",
"0.863",
"0.8906",
"0.8774",
"0.5215",
"0.8774"
],
[
"0.0866",
"0.872",
"0.4182",
"0.872",
"0.4182",
"0.8853",
"0.0866",
"0.8853"
],
[
"0.4346",
"0.8726",
"0.4673",
"0.8726",
"0.4673",
"0.884",
"0.4346",
"0.884"
],
[
"0.5034",
"0.879",
"0.888",
"0.879",
"0.888",
"0.8926",
"0.5034",
"0.8926"
],
[
"0.0801",
"0.8853",
"0.4673",
"0.886",
"0.4673",
"0.901",
"0.0801",
"0.8994"
]
] | [
"(a) Input Video (b)",
"(c) Recovered SRM",
"(d) Ground Truth",
"Weights",
"Material",
"",
"(f) Ground Truth",
"(g) Zoom-in(ours)",
"",
"(e) Recovered SRM",
"口国果",
"(h) Zoom-in(GT)",
"Figure 5: Sample results of recovered SRMs and material weights. Given input video frames (a), we recover global SRMs",
"(c) and their linear combination weights (b) from the optimization of equation ( 7 ). The scenes presented here have two",
"material bases, visualized with red and green channels. Estimated SRMs (c) corresponding to the shiny object surface (green",
"channel) correctly capture the light sources of the scenes, shown in the reference panorama images (d). For both scenes the",
"SRMs corresponding to the red channel is mostly black, thus not shown, as the surface is mostly diffuse. The recovered",
"SRM of (c) overemphasizes blue channel due to oversaturation in input images. Third row shows estimation result from a",
"video of the same bag of chips (first row) under different lighting, Close inspection of the recovered environment (g) reveals",
"many scene details, including floors in a nearby building visible through the window.",
"们引入了基于对抗神经网络的生成器9和鉴别器d来",
"而忽略了相互反射。虽然对完整，全局，漫射十镜面",
"染逼真的照片。g将我们对当前视图的漫反射Dp和",
"光传输进行建模是棘手的，但我们可以通过光线跟踪",
"镜面Sp分量的最佳预测（从公式（7）获得），以及",
"第一反弹图像（FBI）来近似一阶相互反射，如下所示，",
"将在本节稍后定义的互反射和菲涅耳项FBI，Rp和",
"对于要染的虚拟视点中的每个像素u，通过u从相",
"FCI作为输人",
"机中心投射光线。如果我们现在假装每个场景表面都",
"因此，生成器g接收Cp=（Dp,Sp,FBI,Rp,FCI)",
"是完美的镜子，则该光线可能会反弹多次并与多个表",
"作为输入并输出视图的预测，而鉴别器d则对其真实",
"面相交。令X2为该光线与场景的第二个交点。用X2",
"性进行评分。我们使用像素化L1，感知损失Lp[9]和",
"的漫反射颜色在FBI中的u处染像素，如果没有第",
"对抗损失[31]的组合，如Sec中所述。3.2：",
"二个交点，则用黑色染像素(图。4(d))",
"可以通过将FBI与BRDF卷积来建模有光泽（不完",
"g* = argmin max >GLG(g, d)+>,Cp(g)+>1Ci(g), (8)",
"美的镜子）的相互反射。但是，严格来说，应在angular",
"domain【62]而不是图像空间中过滤互反射图像，即。",
"其中Lp(g)=内pECp(g(Cp),Gp)是所有输人图",
"e。，人射光沿中心为反射光线方向wr的镜面反射瓣的",
"像的感知损失的平均值，LG(g,d)和Ci(g)类似地定",
"卷积。给定Wr，可以通过卷积由Wr加权的FBI图像",
"义为跨顿的平均损失。请注意，此谊染器g是scene",
"来近似图像空间中的角域卷积。但是，由于我们不知",
"specific，仅在特定场景的图像上进行训练，以推断同",
"道镜面核，因此我们让网络使用w作为指导来推断权",
"一场景的新视图，就像神经谊染社区50，69，21中通常",
"重。我们将每个像素的w编码为三通道图像Rp（图。",
"做的那样",
"4 (e))",
"菲涅耳效果使高光在近掠视角下更强，对于逼真的",
"Modeling Interreflections and Fresnel Effects",
"方程",
"谊染非常重要。菲涅耳系数按照[64]进行近似：R(α）：",
"式E（2）仅对环境对每个表面点的直接照明进行建模，"
] | {
"arxiv_id": "2001.04642",
"language": "zh",
"page_idx": 4
} |
||
2303.10144-zh-0 | 2303.10144-zh-0.jpg | $\underline{\text{Published as a conference paper at ICLR 2023}}$
动态更新与数据比率：最小化世界模型过度拟合
Nicolai Dorka$^1$ Tim Welschehold$^1$ Wolfram Burgard$^2$ $^{1}$University of Freiburg$\:^2$University of Technology Nuremberg [email protected]
# $\operatorname{Abstract}$
Early stopping based on the validation set performance is a popular approach to find the right balance between under- and overfitting in the context of supervised learning. However, in reinforcement learning, even for supervised sub-problems such as world model learning, early stopping is not ${\mathrm{applicable~as~the~dataset~is~continually~evolving.~As~a~solution,~we~propose}}$ a new general method that dynamically adjusts the update to data (UTD) ratio during training based on under- and overfitting detection on a small subset of the continuously collected experience not used for training. We apply our method to DreamerV2, a state-of-the-art model-based reinforcement learning algorithm, and evaluate it on the DeepMind Control Suite and the Atari 100 k benchmark. The results demonstrate that one can better balance under- and overestimation by adjusting the UTD ratio with our approach compared to the default setting in DreamerV2 and that it is competitive with an extensive hyperparameter search which is not feasible for many applications. Our method eliminates the need to set the UTD hyperparameter by hand and even leads to a higher robustness with regard to other learning-related hyperparameters further reducing the amount of necessary tuning.
# 1 Introduction
在基于模型的强化学习 (RL) 中，代理学习预测世界模型，以通过与其环境的交互来得出给定任务的策略。以前的工作表明，基于模型的方法可以获得与无模型方法相同甚至更好的结果 Silver et al. (2018); Schrittwieser et al. (2020); Chua et al. (2018); Hafner et al. (2021) 。使用世界模型的另一个优点是，一旦为一项任务学习了它，它就可以直接或经过一些微调后用于同一环境中的不同任务，这可能会使代理的多种技能培训便宜得多。学习世界模型原则上是一个监督学习问题。然而，与标准的监督学习设置相比，在基于模型的 RL 中，数据集不是固定的，而是在训练开始时给出的，而是通过与环境的交互随着时间的推移而收集的，这带来了额外的挑战
监督学习中的一个典型问题是在有限的数据量上过度拟合。这是很好的研究，除了几种正则化之外，一个常见的解决方案是使用一个验证集，该验证集不用于训练，而是用于训练期间对训练模型的持续评估。通过考虑验证集上的学习曲线，可以很容易地检测到模型是否低于或过拟合训练数据。对于神经网络，一个典型的行为是太少的更新导致欠拟合，而太多的更新导致过拟合。在这种情况下，验证损失是平衡这两者并在看不见的数据上实现小错误的好工具
对于在动态数据集上学习世界模型，不幸的是，没有建立的方法来确定模型是否在给定时间点对可用的训练数据进行了欠拟合或过拟合。此外，在基于模型的 RL 中，拟合度较差的模型可能会对学习结果产生巨大影响，因为代理从中得出策略，这会影响未来收集的经验从而再次影响世界模型的学习。到目前为止，在基于模型的 RL 中，这通常通过某种形式的正则化和设置更新数据(UTD) 比率来解决，该比率指定模型根据新收集的经验执行多少更新步骤，类似于在监督学习中选择参数更新的总数。类似于监督学习，较高的 UTD 比率更容易使数据过拟合，而较低的 UTD 比率则使数据欠拟合。最先进的方法在训练开始时设置 UTD 比率，并且不基于动态性能指标进行选择。不幸的是，调整此参数非常昂贵，因为整个训练过程必须遍历几次。此外，固定的 UTD 比率通常是次优的，因为在训练过程的不同阶段，此参数的不同值可能是优选的
1 | [
[
"0.1765",
"0.0341",
"0.5073",
"0.0341",
"0.5073",
"0.0486",
"0.1765",
"0.0486"
],
[
"0.1772",
"0.1048",
"0.763",
"0.1048",
"0.763",
"0.1225",
"0.1772",
"0.1225"
],
[
"0.1887",
"0.1515",
"0.298",
"0.1509",
"0.298",
"0.1654",
"0.1887",
"0.1666"
],
[
"0.343",
"0.1515",
"0.4756",
"0.1515",
"0.4756",
"0.1654",
"0.343",
"0.1654"
],
[
"0.5215",
"0.1515",
"0.6567",
"0.1509",
"0.6567",
"0.1654",
"0.5215",
"0.1666"
],
[
"0.1896",
"0.1654",
"0.3586",
"0.1674",
"0.3586",
"0.1818",
"0.1896",
"0.1799"
],
[
"0.3896",
"0.1654",
"0.664",
"0.1674",
"0.664",
"0.1818",
"0.3896",
"0.1799"
],
[
"0.1887",
"0.1799",
"0.3677",
"0.1818",
"0.3677",
"0.1957",
"0.1887",
"0.1932"
],
[
"0.4631",
"0.2317",
"0.536",
"0.2317",
"0.536",
"0.2468",
"0.4631",
"0.2468"
],
[
"0.2336",
"0.2644",
"0.761",
"0.2644",
"0.761",
"0.279",
"0.2336",
"0.279"
],
[
"0.232",
"0.279",
"0.7607",
"0.279",
"0.7607",
"0.2922",
"0.232",
"0.2922"
],
[
"0.232",
"0.2917",
"0.7646",
"0.2922",
"0.7646",
"0.307",
"0.232",
"0.3062"
],
[
"0.2336",
"0.307",
"0.7646",
"0.307",
"0.7646",
"0.3208",
"0.2336",
"0.3208"
],
[
"0.2345",
"0.3208",
"0.7646",
"0.3208",
"0.7646",
"0.3352",
"0.2345",
"0.3352"
],
[
"0.232",
"0.334",
"0.7646",
"0.3333",
"0.7646",
"0.348",
"0.232",
"0.3484"
],
[
"0.2336",
"0.3484",
"0.7656",
"0.3484",
"0.7656",
"0.3623",
"0.2336",
"0.3623"
],
[
"0.2336",
"0.3623",
"0.7656",
"0.3623",
"0.7656",
"0.3757",
"0.2336",
"0.3757"
],
[
"0.232",
"0.3757",
"0.761",
"0.375",
"0.761",
"0.3894",
"0.232",
"0.3901"
],
[
"0.2336",
"0.3894",
"0.7646",
"0.3894",
"0.7646",
"0.4028",
"0.2336",
"0.4028"
],
[
"0.2336",
"0.4028",
"0.7646",
"0.4028",
"0.7646",
"0.4167",
"0.2336",
"0.4167"
],
[
"0.2336",
"0.4172",
"0.7646",
"0.4172",
"0.7646",
"0.4312",
"0.2336",
"0.4312"
],
[
"0.232",
"0.4312",
"0.7656",
"0.43",
"0.7656",
"0.4443",
"0.232",
"0.4458"
],
[
"0.2345",
"0.4458",
"0.7646",
"0.4458",
"0.7646",
"0.459",
"0.2345",
"0.459"
],
[
"0.2336",
"0.459",
"0.7646",
"0.459",
"0.7646",
"0.473",
"0.2336",
"0.473"
],
[
"0.2336",
"0.473",
"0.7656",
"0.473",
"0.7656",
"0.486",
"0.2336",
"0.486"
],
[
"0.2336",
"0.4873",
"0.767",
"0.4873",
"0.767",
"0.5005",
"0.2336",
"0.5005"
],
[
"0.2336",
"0.5005",
"0.357",
"0.502",
"0.357",
"0.515",
"0.2336",
"0.5146"
],
[
"0.1708",
"0.5425",
"0.3162",
"0.5435",
"0.3162",
"0.558",
"0.1708",
"0.5566"
],
[
"0.1772",
"0.577",
"0.822",
"0.577",
"0.822",
"0.587",
"0.1772",
"0.587"
],
[
"0.1765",
"0.589",
"0.8228",
"0.589",
"0.8228",
"0.603",
"0.1765",
"0.603"
],
[
"0.1765",
"0.6016",
"0.822",
"0.6016",
"0.822",
"0.6157",
"0.1765",
"0.6157"
],
[
"0.179",
"0.617",
"0.8228",
"0.617",
"0.8228",
"0.631",
"0.179",
"0.631"
],
[
"0.1765",
"0.631",
"0.822",
"0.631",
"0.822",
"0.6445",
"0.1765",
"0.6445"
],
[
"0.1748",
"0.6426",
"0.8228",
"0.643",
"0.8228",
"0.6577",
"0.1748",
"0.657"
],
[
"0.1765",
"0.6577",
"0.824",
"0.6577",
"0.824",
"0.672",
"0.1765",
"0.672"
],
[
"0.1772",
"0.674",
"0.3718",
"0.674",
"0.3718",
"0.6846",
"0.1772",
"0.6846"
],
[
"0.1765",
"0.6934",
"0.8228",
"0.6934",
"0.8228",
"0.707",
"0.1765",
"0.707"
],
[
"0.1765",
"0.707",
"0.8228",
"0.707",
"0.8228",
"0.72",
"0.1765",
"0.72"
],
[
"0.1765",
"0.7217",
"0.8228",
"0.7217",
"0.8228",
"0.735",
"0.1765",
"0.735"
],
[
"0.1748",
"0.733",
"0.822",
"0.7344",
"0.822",
"0.7485",
"0.1748",
"0.7476"
],
[
"0.1772",
"0.7505",
"0.822",
"0.7505",
"0.822",
"0.761",
"0.1772",
"0.761"
],
[
"0.1772",
"0.7646",
"0.2573",
"0.7646",
"0.2573",
"0.776",
"0.1772",
"0.776"
],
[
"0.1765",
"0.784",
"0.8228",
"0.784",
"0.8228",
"0.7974",
"0.1765",
"0.7974"
],
[
"0.1765",
"0.7974",
"0.8228",
"0.7974",
"0.8228",
"0.811",
"0.1765",
"0.811"
],
[
"0.1748",
"0.81",
"0.822",
"0.811",
"0.822",
"0.8257",
"0.1748",
"0.8247"
],
[
"0.1765",
"0.8257",
"0.8228",
"0.8257",
"0.8228",
"0.839",
"0.1765",
"0.839"
],
[
"0.1765",
"0.839",
"0.8228",
"0.839",
"0.8228",
"0.853",
"0.1765",
"0.853"
],
[
"0.1765",
"0.853",
"0.8228",
"0.853",
"0.8228",
"0.866",
"0.1765",
"0.866"
],
[
"0.1772",
"0.867",
"0.8228",
"0.867",
"0.8228",
"0.881",
"0.1772",
"0.881"
],
[
"0.1765",
"0.881",
"0.8228",
"0.881",
"0.8228",
"0.8945",
"0.1765",
"0.8945"
],
[
"0.1765",
"0.8945",
"0.8228",
"0.8945",
"0.8228",
"0.9077",
"0.1765",
"0.9077"
],
[
"0.1748",
"0.909",
"0.451",
"0.909",
"0.451",
"0.9224",
"0.1748",
"0.9224"
],
[
"0.4902",
"0.9478",
"0.5073",
"0.9478",
"0.5073",
"0.961",
"0.4902",
"0.961"
]
] | [
"Published as a conference paper at ICLR 2023",
"动态更新与数据比率：最小化世界模型过度拟合",
"Nicolai Dorka",
"Tim Welschehold!",
"Wolfram Burgard?",
"lUniversity of Freiburg",
"2University of Technology Nuremberg",
"[email protected]",
"Abstract",
"Early stopping based on the validation set performance is a popular ap",
"proach to find the right balance between under- and overfitting in the con",
"text of supervised learning. However, in reinforcement learning, even for",
"supervised sub-problems such as world model learning, early stopping is not",
"applicable as the dataset is continually evolving. As a solution, we propose",
"a new general method that dynamically adjusts the update to data (UTD)",
"ratio during training based on under- and overfitting detection on a small",
"subset of the continuously collected experience not used for training. We",
"apply our method to DreamerV2, a state-of-the-art model-based reinforce",
"ment learning algorithm, and evaluate it on the DeepMind Control Suite",
"and the Atari 100 k benchmark. The results demonstrate that one can",
"better balance under- and overestimation by adjusting the UTD ratio with",
"our approach compared to the default setting in DreamerV2 and that it is",
"competitive with an extensive hyperparameter search which is not feasible",
"for many applications. Our method eliminates the need to set the UTD",
"hyperparameter by hand and even leads to a higher robustness with regard",
"to other learning-related hyperparameters further reducing the amount of",
"necessary tuning.",
"1Introduction",
"在基于模型的强化学习（RL）中，代理学习预测世界模型，以通过与其环境的交互来得出给",
"定任务的策略。以前的工作表明，基于模型的方法可以获得与无模型方法相同甚至更好的结",
"果 Silver et al. (2018); Schrittwieser et al. (2020); Chua et al. (2018); Hafner et al. (2021)",
"。使用世界模型的另一个优点是，一旦为一项任务学习了它，它就可以直接或经过一些微调",
"后用于同一环境中的不同任务，这可能会使代理的多种技能培训便宜得多。学习世界模型",
"原则上是一个监督学习问题。然而，与标准的监督学习设置相比，在基于模型的RL中，数",
"据集不是固定的，而是在训练开始时给出的，而是通过与环境的交互随着时间的推移而收",
"集的，这带来了额外的挑战",
"监督学习中的一个典型问题是在有限的数据量上过度拟合。这是很好的研究，除了几种正",
"则化之外，一个常见的解决方案是使用一个验证集，该验证集不用于训练，而是用于训练期",
"间对训练模型的持续评估。通过考虑验证集上的学习曲线，可以很容易地检测到模型是否",
"低于或过拟合训练数据。对于神经网络，一个典型的行为是太少的更新导致欠拟合，而太多",
"的更新导致过拟合。在这种情况下，验证损失是平衡这两者并在看不见的数据上实现小错",
"误的好工具",
"对于在动态数据集上学习世界模型，不幸的是，没有建立的方法来确定模型是否在给定时",
"间点对可用的训练数据进行了欠拟合或过拟合。此外，在基于模型的RL中，拟合度较差的",
"模型可能会对学习结果产生巨大影响，因为代理从中得出策略，这会影响未来收集的经验",
"从而再次影响世界模型的学习。到目前为止，在基于模型的RL中，这通常通过某种形式的",
"正则化和设置更新数据（UTD）比率来解决，该比率指定模型根据新收集的经验执行多少",
"更新步骤，类似于在监督学习中选择参数更新的总数。类似于监督学习，较高的UTD比率",
"更容易使数据过拟合，而较低的UTD比率则使数据欠拟合。最先进的方法在训练开始时设",
"置UTD比率，并且不基于动态性能指标进行选择。不幸的是，调整此参数非常昂贵，因为",
"整个训练过程必须遍历儿次。此外，固定的UTD比率通常是次优的，因为在训练过程的不",
"同阶段，此参数的不同值可能是优选的",
"1"
] | {
"arxiv_id": "2303.10144",
"language": "zh",
"page_idx": 0
} |
||
2303.10144-zh-1 | 2303.10144-zh-1.jpg | $\underline{\text{Published as a conference paper at ICLR 2023}}$
<FigureHere>
在本文中，我们提出了一种通用方法-称为 D 动态Update-to-Data 比率(DUTD)-在训练过程中调整UTD 比率。DUTD 的灵感来自于使用早期停止来平衡欠拟合和过拟合。它将收集的经验的一小部分存储在一个单独的验证缓冲区中，该缓冲区不用于训练，而是用于跟踪世界模型准确性的发展，以检测欠拟合和过拟合。基于此，我们然后动态调整 UTD 比率
我们在 DeepMind 控制套件和 Atari100k 基准测试中评估应用于 DreamerV2 Hafner et al.(2021)的DUTD。结果表明，相对于默认的 DreamerV2 配置DUTD 提高了整体性能。最重要的是，DUTD 使搜索最佳 UTD 速率变得过时 ,并且与通过 DreamerV2的广泛超参数调整找到的最佳值具有竞争力。此外， 我们的实验表明，使用 DUTD,世界模型在选择学习率方面变得更加健壮。
总而言之，本文做出了以下贡献：i)我们介绍了一
种通过对保留数据进行评估来在线检测世界模型的 small subset of the experience col欠拟合和过拟合的方法；ü)我们使用此信息来动态 lected from the environment is stored 调整 UTD 比率以优化世界模型的性能；iii) 我们的 in a validation set not used for train方法使 UTD 超参数的调整变得过时；iv)我们示例 ing. The world model is trained for $\begin{array}{c}\text{性地将我们的方法应用于基于模型的最先进的超搜}&\text{one update after every}\\\text{ 专 从而坦言了}\end{array}\frac1{UTDratio}$many 索，从而提高了
Figure 1: Overview of DUTD. A
environment steps
From time to
time, e.g. , after an episode ended, the UTD ratio is adjusted depending on the detection of under- or overfitting of the world model on the vali-
在强化学习中有两种形式的泛化和过拟合。任务间 dation data. The policy is obtained 过度拟合描述了对特定环境的过度拟合，使得在稍 from the world model either by plan微不同的环境中的性能显著下降。这出现在 sim-to- ning or learning and collects new data real 的上下文中，其中模拟与需要执行良好策略的 in the environment.
## 2 Related Work
目标环境不同，或者当环境略有变化时，例如，由于不同的视觉外观 Zhang et al. (2018b); Packer et al. (2018); Zhang et al. (2018a);Raileanu et al.(2020); $\mathop{\mathrm{Song~et~al.~(2020)~。相比之下，任务内过度拟合出}}$ 现在从固定环境中的有限数据学习的背景下，当模型太完美地拟合数据并且不能很好地推广到新数据时。我们认为任务内与任务间泛化相反
在基于模型的强化学习中，在不准确的动力学模型 Arumugam et al. (2018); Jiang et al. (2015)上也存在策略过度拟合的问题。因此，该策略优化了模型的不准确性，并发现了对实际环境不起作用的漏洞利用。一种方法是，当估计的不确定性很高时，使用来自动力学模型集合的不确定性估计更为保守 Chua et al. (2018) 。防止策略利用模型的另一种方法是对策略考虑的计划使用不同类型的正则化 Arumugam et al. (2018)。与以前的工作相反，我们通过学习更好的动力学模型来直接解决问题的根源。因此，我们的方法是正交的，可以很容易地与刚才提到的工作线结合
直接针对动力学模型的过拟合可以通过使用贝叶斯动力学模型以及这种模型带来的不确定性来完成。高斯过程已在此上下文中成功使用 Deisenroth &Rasmussen (2011),尽管很难将其扩展到高维问题。减少动力学模型过度拟合的另一种方法是使用监督学习技术。例如，这包括权重的正则化，删除 Srivastava et al.(2014)或数据增强 Laskin et al.(2020); Schwarzer et al. (2021) 。所有这些都与我们的方法正交，可以与之结合以学习更好的动力学模型。另一种流行的方法是提前停止 Strand (1974); Anderssen &Prenter (1981); Morgan & Bourlard (1989) ,在训练损失收敛之前停止训练。我们的方法可以看作是在动态数据集场景中提前停止的类比
减少模型参数的数量可以防止过度拟合，但与具有更多参数的适量训练步骤相比，可能会
降低性能。我们的方法通过为给定网络自动选择适量的训练步骤来克服此问题
RL 算法的超参数优化也与我们的工作有关。例如，通过使用贝叶斯优化 Chen et al. (2018) 对 AlphaStar Silver et al. (2018) 进行了改进。Zhang et al. (2021) 演示了可以通过自动超
2 | [
[
"0.1765",
"0.0341",
"0.5073",
"0.0341",
"0.5073",
"0.04736",
"0.1765",
"0.04736"
],
[
"0.1765",
"0.1061",
"0.5415",
"0.1061",
"0.5415",
"0.1193",
"0.1765",
"0.1193"
],
[
"0.1765",
"0.1187",
"0.54",
"0.1187",
"0.54",
"0.132",
"0.1765",
"0.132"
],
[
"0.6587",
"0.1256",
"0.7183",
"0.1256",
"0.7183",
"0.1364",
"0.6587",
"0.1364"
],
[
"0.1765",
"0.132",
"0.54",
"0.132",
"0.54",
"0.1459",
"0.1765",
"0.1459"
],
[
"0.179",
"0.1483",
"0.539",
"0.1483",
"0.539",
"0.159",
"0.179",
"0.159"
],
[
"0.1765",
"0.1604",
"0.539",
"0.1604",
"0.539",
"0.1736",
"0.1765",
"0.1736"
],
[
"0.585",
"0.1666",
"0.6177",
"0.1666",
"0.6177",
"0.1787",
"0.585",
"0.1787"
],
[
"0.179",
"0.1755",
"0.539",
"0.1755",
"0.539",
"0.1863",
"0.179",
"0.1863"
],
[
"0.1765",
"0.1881",
"0.5415",
"0.1875",
"0.5415",
"0.2008",
"0.1765",
"0.202"
],
[
"0.1765",
"0.2026",
"0.2091",
"0.2026",
"0.2091",
"0.214",
"0.1765",
"0.214"
],
[
"0.681",
"0.2084",
"0.7197",
"0.2084",
"0.7197",
"0.2184",
"0.681",
"0.2184"
],
[
"0.7563",
"0.2039",
"0.7974",
"0.2039",
"0.7974",
"0.2153",
"0.7563",
"0.2153"
],
[
"0.6772",
"0.2108",
"0.6846",
"0.2108",
"0.6846",
"0.214",
"0.6772",
"0.214"
],
[
"0.761",
"0.2108",
"0.7915",
"0.2108",
"0.7915",
"0.2229",
"0.761",
"0.2229"
],
[
"0.1748",
"0.2235",
"0.54",
"0.2235",
"0.54",
"0.2374",
"0.1748",
"0.2374"
],
[
"0.664",
"0.2216",
"0.7173",
"0.2247",
"0.716",
"0.2361",
"0.6626",
"0.2329"
],
[
"0.1772",
"0.2374",
"0.5415",
"0.2374",
"0.5415",
"0.2507",
"0.1772",
"0.2507"
],
[
"0.585",
"0.2456",
"0.621",
"0.25",
"0.6167",
"0.2727",
"0.581",
"0.2683"
],
[
"0.697",
"0.2474",
"0.7295",
"0.25",
"0.728",
"0.262",
"0.6953",
"0.26"
],
[
"0.1765",
"0.2651",
"0.54",
"0.2651",
"0.54",
"0.279",
"0.1765",
"0.279"
],
[
"0.1748",
"0.25",
"0.539",
"0.2507",
"0.539",
"0.2644",
"0.1748",
"0.2632"
],
[
"0.1748",
"0.2778",
"0.54",
"0.2778",
"0.54",
"0.2917",
"0.1748",
"0.2917"
],
[
"0.755",
"0.2822",
"0.7935",
"0.2822",
"0.7935",
"0.286",
"0.755",
"0.286"
],
[
"0.1772",
"0.2942",
"0.539",
"0.2942",
"0.539",
"0.305",
"0.1772",
"0.305"
],
[
"0.7646",
"0.2886",
"0.789",
"0.2886",
"0.789",
"0.2966",
"0.7646",
"0.2966"
],
[
"0.1765",
"0.3062",
"0.5415",
"0.3062",
"0.5415",
"0.3193",
"0.1765",
"0.3193"
],
[
"0.1765",
"0.3208",
"0.3464",
"0.3208",
"0.3464",
"0.334",
"0.1765",
"0.334"
],
[
"0.1765",
"0.3403",
"0.544",
"0.3403",
"0.544",
"0.3542",
"0.1765",
"0.3542"
],
[
"0.5537",
"0.3447",
"0.6284",
"0.3447",
"0.6284",
"0.3547",
"0.5537",
"0.3547"
],
[
"0.6504",
"0.3423",
"0.824",
"0.3423",
"0.824",
"0.356",
"0.6504",
"0.356"
],
[
"0.1765",
"0.3547",
"0.5435",
"0.3547",
"0.5435",
"0.3687",
"0.1765",
"0.3687"
],
[
"0.5474",
"0.3467",
"0.5537",
"0.3467",
"0.5537",
"0.351",
"0.5474",
"0.351"
],
[
"0.5474",
"0.3599",
"0.5537",
"0.3599",
"0.5537",
"0.3655",
"0.5474",
"0.3655"
],
[
"0.5557",
"0.3567",
"0.822",
"0.3567",
"0.822",
"0.3706",
"0.5557",
"0.3706"
],
[
"0.179",
"0.3706",
"0.5415",
"0.3706",
"0.5415",
"0.3806",
"0.179",
"0.3806"
],
[
"0.5557",
"0.3706",
"0.8228",
"0.3718",
"0.8228",
"0.3833",
"0.5557",
"0.3818"
],
[
"0.1772",
"0.3818",
"0.5435",
"0.3818",
"0.5435",
"0.3958",
"0.1772",
"0.3958"
],
[
"0.556",
"0.3865",
"0.8203",
"0.3865",
"0.8203",
"0.3965",
"0.556",
"0.3965"
],
[
"0.1772",
"0.3984",
"0.5415",
"0.3984",
"0.5415",
"0.4092",
"0.1772",
"0.4092"
],
[
"0.556",
"0.3997",
"0.822",
"0.3997",
"0.822",
"0.4104",
"0.556",
"0.4104"
],
[
"0.1765",
"0.411",
"0.544",
"0.411",
"0.544",
"0.4248",
"0.1765",
"0.4248"
],
[
"0.54",
"0.414",
"0.5557",
"0.414",
"0.5557",
"0.4216",
"0.54",
"0.4216"
],
[
"0.5513",
"0.411",
"0.824",
"0.4136",
"0.824",
"0.4312",
"0.5513",
"0.4287"
],
[
"0.1748",
"0.4255",
"0.29",
"0.4255",
"0.29",
"0.4395",
"0.1748",
"0.4395"
],
[
"0.556",
"0.4255",
"0.7036",
"0.4268",
"0.7036",
"0.44",
"0.556",
"0.4395"
],
[
"0.7183",
"0.428",
"0.824",
"0.428",
"0.824",
"0.4382",
"0.7183",
"0.4382"
],
[
"0.556",
"0.44",
"0.8228",
"0.44",
"0.8228",
"0.455",
"0.556",
"0.455"
],
[
"0.556",
"0.4526",
"0.824",
"0.4539",
"0.824",
"0.4673",
"0.556",
"0.4666"
],
[
"0.1732",
"0.461",
"0.3284",
"0.461",
"0.3284",
"0.4749",
"0.1732",
"0.4749"
],
[
"0.558",
"0.4685",
"0.8203",
"0.4685",
"0.8203",
"0.4792",
"0.558",
"0.4792"
],
[
"0.556",
"0.481",
"0.822",
"0.481",
"0.822",
"0.4944",
"0.556",
"0.4944"
],
[
"0.1765",
"0.4924",
"0.544",
"0.4924",
"0.544",
"0.506",
"0.1765",
"0.506"
],
[
"0.54",
"0.4968",
"0.5557",
"0.4968",
"0.5557",
"0.504",
"0.54",
"0.504"
],
[
"0.556",
"0.4968",
"0.8228",
"0.4968",
"0.8228",
"0.507",
"0.556",
"0.507"
],
[
"0.1765",
"0.507",
"0.544",
"0.507",
"0.544",
"0.521",
"0.1765",
"0.521"
],
[
"0.54",
"0.511",
"0.558",
"0.511",
"0.558",
"0.518",
"0.54",
"0.518"
],
[
"0.5513",
"0.5083",
"0.822",
"0.509",
"0.822",
"0.523",
"0.5513",
"0.5215"
],
[
"0.1765",
"0.521",
"0.544",
"0.521",
"0.544",
"0.534",
"0.1765",
"0.534"
],
[
"0.5557",
"0.523",
"0.824",
"0.523",
"0.824",
"0.536",
"0.5557",
"0.536"
],
[
"0.1772",
"0.536",
"0.5415",
"0.536",
"0.5415",
"0.547",
"0.1772",
"0.547"
],
[
"0.5474",
"0.5405",
"0.5537",
"0.5405",
"0.5537",
"0.5444",
"0.5474",
"0.5444"
],
[
"0.5522",
"0.537",
"0.697",
"0.5386",
"0.697",
"0.55",
"0.5522",
"0.549"
],
[
"0.1772",
"0.549",
"0.539",
"0.549",
"0.539",
"0.5625",
"0.1772",
"0.5625"
],
[
"0.1772",
"0.563",
"0.536",
"0.563",
"0.536",
"0.5737",
"0.1772",
"0.5737"
],
[
"0.1772",
"0.577",
"0.5415",
"0.577",
"0.5415",
"0.5903",
"0.1772",
"0.5903"
],
[
"0.1765",
"0.589",
"0.54",
"0.589",
"0.54",
"0.603",
"0.1765",
"0.603"
],
[
"0.1765",
"0.604",
"0.54",
"0.604",
"0.54",
"0.617",
"0.1765",
"0.617"
],
[
"0.1765",
"0.617",
"0.8105",
"0.617",
"0.8105",
"0.631",
"0.1765",
"0.631"
],
[
"0.1765",
"0.6396",
"0.822",
"0.6396",
"0.822",
"0.653",
"0.1765",
"0.653"
],
[
"0.1765",
"0.653",
"0.822",
"0.653",
"0.822",
"0.6665",
"0.1765",
"0.6665"
],
[
"0.1772",
"0.6665",
"0.822",
"0.6665",
"0.822",
"0.6797",
"0.1772",
"0.6797"
],
[
"0.1772",
"0.6816",
"0.822",
"0.6816",
"0.822",
"0.6924",
"0.1772",
"0.6924"
],
[
"0.1765",
"0.6943",
"0.822",
"0.6943",
"0.822",
"0.7085",
"0.1765",
"0.7085"
],
[
"0.1772",
"0.7104",
"0.822",
"0.7104",
"0.822",
"0.72",
"0.1772",
"0.72"
],
[
"0.1772",
"0.7236",
"0.3865",
"0.7236",
"0.3865",
"0.7344",
"0.1772",
"0.7344"
],
[
"0.1772",
"0.7437",
"0.8228",
"0.7437",
"0.8228",
"0.757",
"0.1772",
"0.757"
],
[
"0.1765",
"0.7583",
"0.822",
"0.7583",
"0.822",
"0.7715",
"0.1765",
"0.7715"
],
[
"0.1772",
"0.7715",
"0.8228",
"0.7715",
"0.8228",
"0.785",
"0.1772",
"0.785"
],
[
"0.1748",
"0.784",
"0.8228",
"0.785",
"0.8228",
"0.799",
"0.1748",
"0.7974"
],
[
"0.1748",
"0.799",
"0.8228",
"0.799",
"0.8228",
"0.8125",
"0.1748",
"0.8125"
],
[
"0.1748",
"0.8125",
"0.822",
"0.813",
"0.822",
"0.827",
"0.1748",
"0.8257"
],
[
"0.1732",
"0.8247",
"0.8228",
"0.8257",
"0.8228",
"0.8403",
"0.1732",
"0.839"
],
[
"0.1765",
"0.8403",
"0.3406",
"0.8403",
"0.3406",
"0.854",
"0.1765",
"0.854"
],
[
"0.1748",
"0.86",
"0.822",
"0.8613",
"0.822",
"0.8755",
"0.1748",
"0.875"
],
[
"0.1772",
"0.8774",
"0.7295",
"0.8774",
"0.7295",
"0.888",
"0.1772",
"0.888"
],
[
"0.1765",
"0.8965",
"0.822",
"0.8965",
"0.822",
"0.91",
"0.1765",
"0.91"
],
[
"0.1748",
"0.911",
"0.8228",
"0.911",
"0.8228",
"0.925",
"0.1748",
"0.925"
],
[
"0.492",
"0.9497",
"0.5063",
"0.9497",
"0.5063",
"0.9604",
"0.492",
"0.9604"
]
] | [
"Published as a conference paper at ICLR 2023",
"在本文中，我们提出了一种通用方法-称为D动态",
"Update-to-Data比率（DUTD）-在训练过程中调整",
"Environment",
"UTD比率。DUTD的灵感来自于使用早期停止来",
"平衡欠拟合和过拟合。它将收集的经验的一小部分",
"存储在一个单独的验证缓冲区中，该缓冲区不用于",
"Policy",
"训练，而是用于跟踪世界模型准确性的发展，以检测",
"欠拟合和过拟合。基于此，我们然后动态调整UTD",
"比率",
"inevery",
"Training",
"",
"Data",
"我们在DeepMind控制套件和Atari100k基准测试",
"UTDratio",
"中评估应用于DreamerV2Hafneretal.（2021）的",
"Wodle",
"update",
"DUTD提高了整体性能。最重要的是，DUTD使搜",
"DUTD。结果表明，相对于默认的DreamerV2配置",
"索最佳UTD速率变得过时，并且与通过DreamerV2",
"vaoato",
"的广泛超参数调整找到的最佳值具有竞争力。此外",
"Data",
"我们的实验表明，使用DUTD，世界模型在选择学",
"习率方面变得更加健壮。",
"总而言之，本文做出了以下贡献：i）我们介绍了一",
"Figure l:",
"Overview of DUTD. A",
"种通过对保留数据进行评估来在线检测世界模型的",
"",
"",
"small subset of the experience col-",
"欠拟合和过拟合的方法：i）我们使用此信息来动态",
"lectedfromtheenvironmentisstored",
"调整UTD比率以优化世界模型的性能；ii）我们的",
"in a validation setnot used fortrain",
"方法使UTD超参数的调整变得过时：iv）我们示例",
"ing. The world model is trained for",
"性地将我们的方法应用于基于模型的最先进的超搜",
"",
" one update after every yTDratio many",
"索，从而提高了",
"environment steps.",
"From time to",
"time, e.g., after an episode ended",
"the UTD ratio is adjusted depending",
"2 Related Work",
"on the detection of under- or overfit",
"ting of the world model on the vali-",
"在强化学习中有两种形式的泛化和过拟合。任务间",
"",
"dation data.The policy is obtained",
"过度拟合描述了对特定环境的过度拟合，使得在稍",
"",
"from the world model either by plan",
"微不同的环境中的性能显著下降。这出现在sim-to-",
"ning or learning and collects new data",
"real的上下文中，其中模拟与需要执行良好策略的",
"",
"intheenvironment.",
"目标环境不同，或者当环境略有变化时，例如，由于",
"不同的视觉外观Zhangetal.（2018b)：Packeretal",
"(2018); Zhang et al. (2018a); Raileanu et al. (2020);",
"Song etal.（2020）。相比之下，任务内过度拟合出",
"现在从固定环境中的有限数据学习的背景下，当模",
"型太完美地拟合数据并且不能很好地推广到新数据时。我们认为任务内与任务间泛化相反",
"在基于模型的强化学习中，在不准确的动力学模型Arumugametal.（2018)；Jiangetal.",
"(2015）上也存在策略过度拟合的问题。因此，该策略优化了模型的不准确性，并发现了对实",
"际环境不起作用的漏洞利用。一种方法是，当估计的不确定性很高时，使用来自动力学模型",
"集合的不确定性估计更为保守Chuaetal.（2018）。防正策略利用模型的另一种方法是对策",
"略考虑的计划使用不同类型的正则化Arumugam etal.（2018）。与以前的工作相反，我们",
"通过学习更好的动力学模型来直接解决问题的根源。因此，我们的方法是正交的，可以很容",
"易地与刚才提到的工作线结合",
"直接针对动力学模型的过拟合可以通过使用贝叶斯动力学模型以及这种模型带来的不确定",
"性来完成。高斯过程已在此上下文中成功使用Deisenroth&Rasmussen（2011），尽管很",
"难将其扩展到高维问题。减少动力学模型过度拟合的另一种方法是使用监督学习技术。例",
"如，这包括权重的正则化，删除Srivastavaetal.（2014）或数据增强Laskinetal.（2020)；",
"Schwarzeretal.（2021）。所有这些都与我们的方法正交，可以与之结合以学习更好的动力",
"学模型。另一种流行的方法是提前停止Strand（1974)；Anderssen&Prenter（1981)；Morgan",
"&Bourlard（1989），在训练损失收敛之前停止训练。我们的方法可以看作是在动态数据集",
"场景中提前停止的类比",
"减少模型参数的数量可以防止过度拟合，但与具有更多参数的适量训练步骤相比，可能会",
"降低性能。我们的方法通过为给定网络自动选择适量的训练步骤来克服此问题",
"RL算法的超参数优化也与我们的工作有关。例如，通过使用贝叶斯优化Chenetal.（2018）",
"对AlphaStarSilveretal.（2018）进行了改进。Zhangetal.（2021）演示了可以通过自动超",
"2"
] | {
"arxiv_id": "2303.10144",
"language": "zh",
"page_idx": 1
} |
||
2303.10144-zh-2 | 2303.10144-zh-2.jpg | $\underline{\text{Published as a conference paper at ICLR 2023}}$
参数优化来大大改进基于模型的 RL 算法。Parker-Holder et al. (2022) 给出了有关自动 RL 的最新概述。但是，这些方法中的大多数都通过在内部循环中在环境上训练 RL 代理来改善超参数，同时在每次运行期间保持超参数固定。我们的工作与此不同，在单次运行训练期间在线调整超参数。Schaul et al. (2019) 的方法也属于这一类，可以动态调整与行为相关的参数，例如随机性和乐观度。同样，算法 Agent57 Badia et al.(2020) 自适应地从一组具有不同探索策略的策略中进行选择，并在所有 57 Atari 游戏 Bellemare et al. (2013) 上实现了人类水平的性能。另一种方法是调整超参数，以在线控制值函数的低估和高估，从而产生无模型 RL 算法，该算法在连续控制任务 Dorka et al. (2021)上具有出色的性能
与这些方法相反，我们的方法通过在验证集上在线检测欠拟合和过拟合来直接学习更好的世界模型，并相应地动态调整更新步骤的数量。这使得调整 UTD 比率超参数的需求变得不必要，并进一步允许自动使其值适应不同训练阶段的需求
3 The DUTD Algorithm
在本节中，我们将首先介绍一般设置，在找到正确的数据拟合的背景下解释早期停止，并提出一种将该技术转移到在线学习设置的新方法。最后，我们解释了如何将该方法应用于${\mathrm{DreamerV2}}$。
3.1 Model-Based Reinforcement Learning
假设马尔可夫决策过程$(\mathcal{S},\mathcal{A},\mathcal{P},\mathcal{R})$ ,我们使用经典的 RL 框架 Sutton & Barto (2018) 在此框架中，代理按顺序观察当前状态$s_t\in\mathcal{S}$,在其中执行动作$a_t\in\mathcal{A}$,根据奖励函数$\mathcal{R}$ 接收标量奖励$r_t$ ,并转换到由未知转换动态$\mathcal{P}$ 生 成 的 下 一 个 状 态 $s_{t+ 1}$ 。目标是学习一个策略，该策略选择每个状态下的操作，以使总预期收益$\sum_{i=t}^Tr_i$最大化
基于模型的 RL 方法学习世界模型$\hat{\mathcal{P}}(s_{t+1}\mid s_t,a_t)$ -也称为动力学模型-和奖励模型$\hat{R}(r_t\mid s_t)$ ,它们试图反映其真实但未知的对应物。然后，这些模型可以通过区分世界模型或生成可以训练 RL 算法的假想展示来学习良好的策略。或者，可以在规划算法中使用学习的模型来选择环境中的动作
3.2 Under- and Overfitting
监督学习中的一个众所周知的问题是过拟合，这通常对应于训练数据上的低误差和训练期间看不到的测试数据上的高误差。通常，如果模型太完美地拟合训练数据，就会发生这种情况。与此相反，欠拟合对应于模型甚至不能很好地拟合训练数据的情况，其特点是训练误差和测试误差都很高。为了衡量模型在不可见数据上的性能，可用数据通常分为训练集和验证集。通常，仅使用训练集来训练模型，而使用验证集来评估其在新数据上的性能
对于迭代训练方法——比如基于梯度下降的方法——通常通过观察训练的学习曲线和验证误差与训练步数的关系来检测过拟合。一个典型的行为是，在训练开始时，训练和验证损失都在减少。这是模型仍欠拟合的区域。在某些时候，当模型开始过度拟合训练数据时，只有训练损失进一步减少，而验证损失开始增加。上述早期停止方法通过在验证损失开始增加时停止训练来平衡欠拟合和过拟合
虽然在监督学习中，可以通过使用验证损失轻松选择拟合良好的模型，但在强化学习中，由于数据集不是固定的而是动态的，并且在不断增长或变化，因此无法应用此技术。此外，现行政策的质量会影响未来收集的数据的质量。尽管学习世界模型原则上是一项有监督的任务，但在基于模型的 RL 框架中也会出现此问题
3.3 Dynamic Update-to-Data Ratio
许多 RL 算法中的典型超参数是更新与数据 (UTD) 比率，它指定每个环境步骤执行的更新步骤数(i。e。,每个新数据点)。原则上，该比率可用于平衡欠拟合和过拟合，因为人们可以通过对当前可用数据执行不太少或太多更新步骤的方式对其进行控制。然而，在优化此参数时会出现几个问题。首先，调整这个参数是非常昂贵的，因为它需要多次运行完整的RL 训练，这使得它对许多潜在的应用程序来说是不可行的。其次，在整个训练期间，一个固定值是最优选择的假设不一定成立。例如，如果将来自新探索的环境区域的数据添加到重播缓冲区，则增加更新步骤的数量可能是有益的
3 | [
[
"0.1765",
"0.0341",
"0.5073",
"0.0341",
"0.5073",
"0.0486",
"0.1765",
"0.0486"
],
[
"0.179",
"0.1067",
"0.8203",
"0.1067",
"0.8203",
"0.11743",
"0.179",
"0.11743"
],
[
"0.1765",
"0.1193",
"0.822",
"0.1193",
"0.822",
"0.1332",
"0.1765",
"0.1332"
],
[
"0.1765",
"0.1332",
"0.822",
"0.1332",
"0.822",
"0.1465",
"0.1765",
"0.1465"
],
[
"0.1765",
"0.1465",
"0.822",
"0.1465",
"0.822",
"0.1604",
"0.1765",
"0.1604"
],
[
"0.1765",
"0.161",
"0.8228",
"0.161",
"0.8228",
"0.1749",
"0.1765",
"0.1749"
],
[
"0.1772",
"0.1749",
"0.822",
"0.1749",
"0.822",
"0.1881",
"0.1772",
"0.1881"
],
[
"0.1765",
"0.1895",
"0.824",
"0.1895",
"0.824",
"0.2026",
"0.1765",
"0.2026"
],
[
"0.1732",
"0.2008",
"0.714",
"0.202",
"0.714",
"0.2153",
"0.1732",
"0.214"
],
[
"0.1765",
"0.2229",
"0.822",
"0.2235",
"0.822",
"0.2374",
"0.1765",
"0.2361"
],
[
"0.1772",
"0.2374",
"0.8228",
"0.2374",
"0.8228",
"0.2507",
"0.1772",
"0.2507"
],
[
"0.1772",
"0.2524",
"0.5825",
"0.2524",
"0.5825",
"0.2632",
"0.1772",
"0.2632"
],
[
"0.1732",
"0.2874",
"0.4004",
"0.2874",
"0.4004",
"0.3005",
"0.1732",
"0.3005"
],
[
"0.1772",
"0.3176",
"0.8203",
"0.3176",
"0.8203",
"0.3276",
"0.1772",
"0.3276"
],
[
"0.1765",
"0.33",
"0.822",
"0.33",
"0.822",
"0.3435",
"0.1765",
"0.3435"
],
[
"0.1765",
"0.3447",
"0.2646",
"0.3447",
"0.2646",
"0.356",
"0.1765",
"0.356"
],
[
"0.1732",
"0.3738",
"0.4885",
"0.375",
"0.4885",
"0.3894",
"0.1732",
"0.3882"
],
[
"0.1765",
"0.3984",
"0.812",
"0.3984",
"0.812",
"0.4124",
"0.1765",
"0.4124"
],
[
"0.808",
"0.4048",
"0.819",
"0.4048",
"0.819",
"0.4092",
"0.808",
"0.4092"
],
[
"0.1772",
"0.414",
"0.822",
"0.414",
"0.822",
"0.4248",
"0.1772",
"0.4248"
],
[
"0.1765",
"0.4268",
"0.822",
"0.4268",
"0.822",
"0.4414",
"0.1765",
"0.4414"
],
[
"0.1765",
"0.443",
"0.6807",
"0.443",
"0.6807",
"0.457",
"0.1765",
"0.457"
],
[
"0.1765",
"0.4673",
"0.8228",
"0.4673",
"0.8228",
"0.4817",
"0.1765",
"0.4817"
],
[
"0.1912",
"0.481",
"0.822",
"0.481",
"0.822",
"0.4956",
"0.1912",
"0.4956"
],
[
"0.1772",
"0.4944",
"0.8228",
"0.4944",
"0.8228",
"0.5083",
"0.1772",
"0.5083"
],
[
"0.1748",
"0.509",
"0.2917",
"0.509",
"0.2917",
"0.523",
"0.1748",
"0.523"
],
[
"0.1732",
"0.5415",
"0.384",
"0.5435",
"0.384",
"0.5596",
"0.1732",
"0.5566"
],
[
"0.179",
"0.5728",
"0.8203",
"0.5728",
"0.8203",
"0.5835",
"0.179",
"0.5835"
],
[
"0.1748",
"0.584",
"0.822",
"0.5835",
"0.822",
"0.5977",
"0.1748",
"0.5986"
],
[
"0.1765",
"0.5986",
"0.8228",
"0.5986",
"0.8228",
"0.6123",
"0.1765",
"0.6123"
],
[
"0.1765",
"0.613",
"0.8228",
"0.613",
"0.8228",
"0.627",
"0.1765",
"0.627"
],
[
"0.1748",
"0.6255",
"0.763",
"0.627",
"0.763",
"0.64",
"0.1748",
"0.6396"
],
[
"0.1765",
"0.6465",
"0.8228",
"0.648",
"0.8228",
"0.661",
"0.1765",
"0.66"
],
[
"0.1772",
"0.661",
"0.8228",
"0.661",
"0.8228",
"0.675",
"0.1772",
"0.675"
],
[
"0.1765",
"0.675",
"0.822",
"0.675",
"0.822",
"0.688",
"0.1765",
"0.688"
],
[
"0.1765",
"0.6895",
"0.8228",
"0.6895",
"0.8228",
"0.7026",
"0.1765",
"0.7026"
],
[
"0.1748",
"0.7026",
"0.4224",
"0.7026",
"0.4224",
"0.7163",
"0.1748",
"0.7163"
],
[
"0.1748",
"0.7236",
"0.822",
"0.7246",
"0.822",
"0.738",
"0.1748",
"0.7373"
],
[
"0.1772",
"0.738",
"0.8228",
"0.738",
"0.8228",
"0.752",
"0.1772",
"0.752"
],
[
"0.1765",
"0.752",
"0.8228",
"0.752",
"0.8228",
"0.765",
"0.1765",
"0.765"
],
[
"0.1748",
"0.765",
"0.5195",
"0.7666",
"0.5195",
"0.78",
"0.1748",
"0.779"
],
[
"0.1748",
"0.8003",
"0.4453",
"0.8003",
"0.4453",
"0.8145",
"0.1748",
"0.8145"
],
[
"0.1765",
"0.8276",
"0.8228",
"0.8276",
"0.8228",
"0.8413",
"0.1765",
"0.8413"
],
[
"0.1765",
"0.8423",
"0.8228",
"0.8423",
"0.8228",
"0.856",
"0.1765",
"0.856"
],
[
"0.1765",
"0.854",
"0.8228",
"0.856",
"0.8228",
"0.869",
"0.1765",
"0.867"
],
[
"0.1772",
"0.869",
"0.822",
"0.869",
"0.822",
"0.8833",
"0.1772",
"0.8833"
],
[
"0.1748",
"0.882",
"0.822",
"0.8833",
"0.822",
"0.898",
"0.1748",
"0.8965"
],
[
"0.1765",
"0.898",
"0.822",
"0.898",
"0.822",
"0.911",
"0.1765",
"0.911"
],
[
"0.1772",
"0.9136",
"0.5312",
"0.9136",
"0.5312",
"0.924",
"0.1772",
"0.924"
],
[
"0.4927",
"0.9497",
"0.5063",
"0.9497",
"0.5063",
"0.961",
"0.4927",
"0.961"
]
] | [
"Published as a conference paper at ICLR 2023",
"参数优化来大大改进基于模型的RL算法。Parker-Holderetal.（2022）给出了有关自动RI",
"的最新概述。但是，这些方法中的大多数都通过在内部循环中在环境上训练RL代理来改善",
"超参数，同时在每次运行期间保持超参数固定。我们的工作与此不同，在单次运行训练期间",
"在线调整超参数。Schauletal.（2019）的方法也属于这一类，可以动态调整与行为相关的参",
"数，例如随机性和乐观度。同样，算法Agent57Badiaetal.（2020）自适应地从一组具有不",
"同探索策略的策略中进行选择，并在所有57Atari游戏Bellemareetal.（2013）上实现了人",
"类水平的性能。另一种方法是调整超参数，以在线控制值函数的低估和高估，从而产生无模",
"型RL算法，该算法在连续控制任务Dorkaetal.（2021）上具有出色的性能",
"与这些方法相反，我们的方法通过在验证集上在线检测欠拟合和过拟合来直接学习更好的",
"世界模型，并相应地动态调整更新步骤的数量。这使得调整UTD比率超参数的需求变得不",
"必要，并进一步充许自动使其值适应不同训练阶段的需求",
"3 The DUTD Algorithm",
"在本节中，我们将首先介绍一般设置，在找到正确的数据拟合的背景下解释早期停止，并",
"提出一种将该技术转移到在线学习设置的新方法。最后，我们解释了如何将该方法应用于",
"DreamerV2.",
"3.1Model-BasedReinforcement Learning",
"假设马尔可夫决策过程（S,A,P,R），我们使用经典的RL框架Sutton&Barto（2018)",
"",
"在此框架中，代理按顺序观察当前状态StES，在其中执行动作atEA：根据奖励函数R",
"接收标量奖励rt，并转换到由未知转换动态P生成的下一个状态st+1。目标是学习一个",
"策略，该策略选择每个状态下的操作，以使总预期收益i=tr；最大化",
"基于模型的RL方法学习世界模型P(st+1|st,at）-也称为动力学模型-和奖励模型R(rt|st)",
"它们试图反映其真实但未知的对应物。然后，这些模型可以通过区分世界模型或生成可以",
"训练RL算法的假想展示来学习良好的策略。或者，可以在规划算法中使用学习的模型来选",
"择环境中的动作",
"3.2Under- and Overfitting",
"监督学习中的一个众所周知的问题是过拟合，这通常对应于训练数据上的低误差和训练期",
"间看不到的测试数据上的高误差。通常，如果模型太完美地拟合训练数据，就会发生这种情",
"况。与此相反，欠拟合对应于模型甚至不能很好地拟合训练数据的情况，其特点是训练误差",
"和测试误差都很高。为了衡量模型在不可见数据上的性能，可用数据通常分为训练集和验",
"证集。通常，仅使用训练集来训练模型，而使用验证集来评估其在新数据上的性能",
"对于送代训练方法一比如基于梯度下降的方法通常通过观察训练的学习曲线和验证",
"误差与训练步数的关系来检测过拟合。一个典型的行为是，在训练开始时，训练和验证损失",
"都在减少。这是模型仍欠拟合的区域。在某些时候，当模型开始过度拟合训练数据时，只有",
"训练损失进一步减少，而验证损失开始增加。上述早期停止方法通过在验证损失开始增加",
"时停止训练来平衡欠拟合和过拟合",
"虽然在监督学习中，可以通过使用验证损失轻松选择拟合良好的模型，但在强化学习中，由",
"于数据集不是固定的而是动态的，并且在不断增长或变化，因此无法应用此技术。此外，现",
"行政策的质量会影响未来收集的数据的质量。尽管学习世界模型原则上是一项有监督的任",
"务，但在基于模型的RL框架中也会出现此问题",
"3.3Dynamic Update-to-Data Ratio",
"许多RL算法中的典型超参数是更新与数据（UTD）比率，它指定每个环境步骤执行的更",
"新步骤数（i。e。，每个新数据点）。原则上，该比率可用于平衡欠拟合和过拟合，因为人们",
"可以通过对当前可用数据执行不太少或太多更新步骤的方式对其进行控制。然而，在优化",
"此参数时会出现几个问题。首先，调整这个参数是非常昂贵的，因为它需要多次运行完整的",
"RL训练，这使得它对许多潜在的应用程序来说是不可行的。其次，在整个训练期间，一个",
"固定值是最优选择的假设不一定成立。例如，如果将来自新探索的环境区域的数据添加到",
"重播缓冲区，则增加更新步骤的数量可能是有益的",
"3"
] | {
"arxiv_id": "2303.10144",
"language": "zh",
"page_idx": 2
} |
||
2303.10144-zh-3 | 2303.10144-zh-3.jpg | $\underline{\text{Published as a conference paper at ICLR 2023}}$
为了解决这些问题，我们提出-DUTD—一种在训练期间动态调整 UTD 比率的新方法。它的灵感来自早期停止标准和目标，通过调整更新步骤的数量来自动平衡在线欠拟合和过拟合。作为方法的一部分，我们将一些经验存储在不用于训练的单独验证缓冲区中。准确地说，每个$d$环境步骤我们都会从几个单独的事件中收集$s$的连续转换，这些事件专门用于验证， 并且每个$k$环境步骤都会在验证缓冲区上评估世界模型，其中$k$应比$d$小得多。由于世界模型学习任务受到监督，因此可以通过在给定的验证序列上记录世界模型的损失来轻松完成。然后将当前验证损失与先前评估的验证损失进行比较。如果损失减少，我们假设模型仍处于欠拟合状态，并将 UTD 率增加指定的量。如果损失增加，我们假设模型处于过拟合状态，因此降低 UTD 率。为了在允许的间隔的高更新侧提供更精细的分辨率，我们在对数空间中调整 UTD 速率，这意味着通过将其分别与$c$或$1/c$的值相乘来增加或减少，其中$c$略大于 1。然后，时间步长$t$的更新公式变为
$$utd_ratio_t=utd_ratio_{t-k}\cdot b;\quad b=\begin{cases}c,&\text{if validation loss}_t<\text{validation loss}_{t-k},\\\frac{1}{c},&\text{if validation loss}_t\ge\text{validation loss}_{t-k}.\end{cases}$$
DUTD 是一种通用方法，可应用于以监督方式学习世界模型的任何基于模型的 RL 算法。实现可以是 UTD 比率，也可以是数据更新比率，它是它的倒数，我们称之为 IUTD(i。e。每个更新步骤的环境步骤数)。如果每个环境步骤执行多个更新，则使用 UTD 比率更为方便；如果仅在某些环境步骤之后执行更新步骤，则使用 IUTD 比率更为方便。从方法上讲， 这两个设置是相同的，因为两个比率描述的是相同的数量，并且只是彼此的倒数
图 1 中显示了 DUTD 的高级概述，算法 1 中描述了伪代码，两者都根据 IUTD 比率进行了解释，因为我们将 DUTD 应用于 DreamerV2 算法 Hafner 的几个更新步骤在计算上非常昂贵。然而，在这两个框架中，这两种情况都可以通过让比率为分数来解决
3.4 Applying DUTD to DreamerV2
$\begin{array}{ll}\text{我们将 DUTD 应用于 DreamerV2}&\frac{\text{Algorithm 1 DUTD (in terms of inverted UTD ra-}}\\\mathrm{Hafner~et~al.~(2021)~,~这是一种基于}&\mathrm{tio)}\end{array}$
模型的 RL 算法，它建立在 Dreamer Hafner et al.(2020)上，后者又建立在 PlaNet Hafner et al. (2019) 上。DreamerV2 通过潜在的想象力学习世界模型。该政策纯粹是通过演员-评论家框架在这个世界模型的潜在空间中学习的。。它是在由世界模型生成的假想展开上训练的。将评论家回归到$\lambda$目标 Schulman et al. (2015); Sutton $\&$ Barto (2018)上，并通过 Reinforce Williams (1992)和动态反向传播损耗的组合来训练演员。世界模型学习一个图像编码器，该编码器将输入映射到分类的潜在状态，在该状态下，递归状态空间模型 Hafner et al. (2019) 学习动态。在潜在状态下学习图像、奖励和折扣因子的三个预测因子。世界模型的总损失是所有三个预测因子的损失和动态预测的潜伏期与编码器的潜伏期之间的 Kullback-Leibler 损失的组合
要应用 DUTD, 我们评估验证集上的图像重建损失。其他选择也是可能的但我们推测图像预测是世界模型中最困难和最重要的部分。人们也可以使用不同损失的组合，但随后可能需要不同损失的比例因子。由于我们希望保持我们的方法简单，并防止我们的
Input: Initial inverted UTD ratio iutd\_ratio~; number of steps after which additional validation data is collected $d$ , number of validation transitions collected $s$ , steps after which the iutd\_ratio is updated $k$ , iutd update increment $c$
for$t=1$to total\_num\_of\_env\_stepsdo
Act according $\overline{\text{to policy }\pi}(a\mid s)$ and observe
next state
if $t$ mod $d==0$ then
Collect $s$ transitions and store experience in a separate validation buffer; increment $t=$ $t+s$ end if if $t$ mod iutd\_ratio==0 then
Perform one training step of the transition
model
end if
if $t\bmod k==0$ then
Compute model loss $L$ on validation dataset
$\#$ Overfitting
${\mathrm{if}}$ $L\geq L_{{\mathrm{previous}}}$ then
iut$\overline {d}\_$ratio= iutd$\_$ratio$\cdot$c
else
$\#$Underfitting
iutd$\_$ratio= iutd$\_$ratio/ c
end if
$\begin{array}{c}{L_{\mathrm{previous}}}&{=}&{L}\\{\mathrm{end~if}}\end{array}$
end for
4 | [
[
"0.1765",
"0.0341",
"0.5073",
"0.0341",
"0.5073",
"0.04736",
"0.1765",
"0.04736"
],
[
"0.1772",
"0.1061",
"0.822",
"0.1061",
"0.822",
"0.1193",
"0.1772",
"0.1193"
],
[
"0.1772",
"0.1193",
"0.822",
"0.1193",
"0.822",
"0.1332",
"0.1772",
"0.1332"
],
[
"0.1765",
"0.132",
"0.824",
"0.132",
"0.824",
"0.1465",
"0.1765",
"0.1465"
],
[
"0.1772",
"0.1465",
"0.822",
"0.1465",
"0.822",
"0.1604",
"0.1772",
"0.1604"
],
[
"0.1765",
"0.1604",
"0.822",
"0.1604",
"0.822",
"0.1736",
"0.1765",
"0.1736"
],
[
"0.1765",
"0.1749",
"0.8228",
"0.1749",
"0.8228",
"0.1881",
"0.1765",
"0.1881"
],
[
"0.1772",
"0.1881",
"0.8228",
"0.1881",
"0.8228",
"0.202",
"0.1772",
"0.202"
],
[
"0.1765",
"0.202",
"0.8228",
"0.202",
"0.8228",
"0.2153",
"0.1765",
"0.2153"
],
[
"0.1748",
"0.2166",
"0.8228",
"0.2166",
"0.8228",
"0.2299",
"0.1748",
"0.2299"
],
[
"0.1772",
"0.2299",
"0.8228",
"0.2299",
"0.8228",
"0.2437",
"0.1772",
"0.2437"
],
[
"0.1748",
"0.2437",
"0.4878",
"0.2437",
"0.4878",
"0.2568",
"0.1748",
"0.2568"
],
[
"0.4902",
"0.26",
"0.8286",
"0.2708",
"0.8267",
"0.2986",
"0.4885",
"0.2886"
],
[
"0.1814",
"0.279",
"0.464",
"0.2766",
"0.464",
"0.2942",
"0.1814",
"0.2966"
],
[
"0.4656",
"0.2747",
"0.5024",
"0.2747",
"0.5024",
"0.286",
"0.4656",
"0.286"
],
[
"0.4631",
"0.2874",
"0.79",
"0.2874",
"0.79",
"0.305",
"0.4631",
"0.305"
],
[
"0.1765",
"0.3257",
"0.8228",
"0.3257",
"0.8228",
"0.339",
"0.1765",
"0.339"
],
[
"0.1765",
"0.339",
"0.818",
"0.339",
"0.818",
"0.3528",
"0.1765",
"0.3528"
],
[
"0.1765",
"0.3542",
"0.822",
"0.3542",
"0.822",
"0.3674",
"0.1765",
"0.3674"
],
[
"0.1765",
"0.3674",
"0.8228",
"0.3674",
"0.8228",
"0.3806",
"0.1765",
"0.3806"
],
[
"0.1765",
"0.3806",
"0.7466",
"0.3806",
"0.7466",
"0.3945",
"0.1765",
"0.3945"
],
[
"0.1765",
"0.4016",
"0.822",
"0.4016",
"0.822",
"0.4153",
"0.1765",
"0.4153"
],
[
"0.1765",
"0.4153",
"0.822",
"0.414",
"0.822",
"0.4287",
"0.1765",
"0.43"
],
[
"0.1765",
"0.43",
"0.8228",
"0.43",
"0.8228",
"0.443",
"0.1765",
"0.443"
],
[
"0.1765",
"0.4443",
"0.2737",
"0.4443",
"0.2737",
"0.4558",
"0.1765",
"0.4558"
],
[
"0.1732",
"0.476",
"0.4453",
"0.476",
"0.4453",
"0.4893",
"0.1732",
"0.4893"
],
[
"0.1748",
"0.502",
"0.4453",
"0.502",
"0.4453",
"0.515",
"0.1748",
"0.515"
],
[
"0.4575",
"0.507",
"0.822",
"0.5083",
"0.822",
"0.523",
"0.4575",
"0.5215"
],
[
"0.1748",
"0.515",
"0.4436",
"0.5146",
"0.4436",
"0.529",
"0.1748",
"0.53"
],
[
"0.4592",
"0.523",
"0.486",
"0.526",
"0.4836",
"0.536",
"0.4558",
"0.5327"
],
[
"0.1765",
"0.53",
"0.4436",
"0.53",
"0.4436",
"0.5435",
"0.1765",
"0.5435"
],
[
"0.1732",
"0.5425",
"0.4436",
"0.5425",
"0.4436",
"0.556",
"0.1732",
"0.556"
],
[
"0.4756",
"0.5405",
"0.824",
"0.5405",
"0.824",
"0.5537",
"0.4756",
"0.5537"
],
[
"0.1748",
"0.558",
"0.4429",
"0.558",
"0.4429",
"0.5684",
"0.1748",
"0.5684"
],
[
"0.4763",
"0.5547",
"0.8228",
"0.5547",
"0.8228",
"0.5684",
"0.4763",
"0.5684"
],
[
"0.1748",
"0.5713",
"0.4453",
"0.5713",
"0.4453",
"0.5854",
"0.1748",
"0.5854"
],
[
"0.4763",
"0.5674",
"0.822",
"0.5674",
"0.822",
"0.5806",
"0.4763",
"0.5806"
],
[
"0.1765",
"0.5854",
"0.4436",
"0.5854",
"0.4436",
"0.5986",
"0.1765",
"0.5986"
],
[
"0.4756",
"0.5806",
"0.822",
"0.5806",
"0.822",
"0.5947",
"0.4756",
"0.5947"
],
[
"0.1765",
"0.5986",
"0.4436",
"0.5986",
"0.4436",
"0.6123",
"0.1765",
"0.6123"
],
[
"0.4763",
"0.595",
"0.755",
"0.595",
"0.755",
"0.6094",
"0.4763",
"0.6094"
],
[
"0.1765",
"0.6123",
"0.4436",
"0.6123",
"0.4436",
"0.6255",
"0.1765",
"0.6255"
],
[
"0.4756",
"0.608",
"0.778",
"0.6094",
"0.778",
"0.6226",
"0.4756",
"0.622"
],
[
"0.1765",
"0.6255",
"0.4436",
"0.6255",
"0.4436",
"0.6396",
"0.1765",
"0.6396"
],
[
"0.4944",
"0.6235",
"0.824",
"0.6235",
"0.824",
"0.637",
"0.4944",
"0.637"
],
[
"0.1765",
"0.64",
"0.4436",
"0.64",
"0.4436",
"0.654",
"0.1765",
"0.654"
],
[
"0.4944",
"0.6396",
"0.5664",
"0.6396",
"0.5664",
"0.6494",
"0.4944",
"0.6494"
],
[
"0.1748",
"0.654",
"0.4436",
"0.654",
"0.4436",
"0.6675",
"0.1748",
"0.6675"
],
[
"0.4927",
"0.651",
"0.6567",
"0.651",
"0.6567",
"0.664",
"0.4927",
"0.664"
],
[
"0.1748",
"0.6665",
"0.4453",
"0.6675",
"0.4453",
"0.6816",
"0.1748",
"0.681"
],
[
"0.509",
"0.6655",
"0.8228",
"0.6655",
"0.8228",
"0.6787",
"0.509",
"0.6787"
],
[
"0.1765",
"0.6816",
"0.429",
"0.6816",
"0.429",
"0.696",
"0.1765",
"0.696"
],
[
"0.509",
"0.6787",
"0.822",
"0.6787",
"0.822",
"0.6924",
"0.509",
"0.6924"
],
[
"0.1765",
"0.696",
"0.4436",
"0.696",
"0.4436",
"0.709",
"0.1765",
"0.709"
],
[
"0.509",
"0.6924",
"0.5435",
"0.6943",
"0.5415",
"0.7056",
"0.5073",
"0.704"
],
[
"0.1765",
"0.7104",
"0.4412",
"0.7104",
"0.4412",
"0.72",
"0.1765",
"0.72"
],
[
"0.4944",
"0.707",
"0.536",
"0.707",
"0.536",
"0.7183",
"0.4944",
"0.7183"
],
[
"0.1732",
"0.7227",
"0.4429",
"0.7227",
"0.4429",
"0.7363",
"0.1732",
"0.7363"
],
[
"0.492",
"0.72",
"0.722",
"0.72",
"0.722",
"0.7344",
"0.492",
"0.7344"
],
[
"0.1765",
"0.7373",
"0.4453",
"0.7373",
"0.4453",
"0.7505",
"0.1765",
"0.7505"
],
[
"0.509",
"0.7344",
"0.824",
"0.7344",
"0.824",
"0.7476",
"0.509",
"0.7476"
],
[
"0.1765",
"0.7505",
"0.4436",
"0.7505",
"0.4436",
"0.7646",
"0.1765",
"0.7646"
],
[
"0.5073",
"0.747",
"0.5557",
"0.747",
"0.5557",
"0.761",
"0.5073",
"0.761"
],
[
"0.1748",
"0.765",
"0.4436",
"0.765",
"0.4436",
"0.779",
"0.1748",
"0.779"
],
[
"0.4927",
"0.7603",
"0.5376",
"0.7603",
"0.5376",
"0.7744",
"0.4927",
"0.7744"
],
[
"0.4902",
"0.7734",
"0.6587",
"0.7744",
"0.6587",
"0.789",
"0.4902",
"0.7886"
],
[
"0.1772",
"0.781",
"0.4429",
"0.781",
"0.4429",
"0.7915",
"0.1772",
"0.7915"
],
[
"0.1765",
"0.7925",
"0.4436",
"0.7925",
"0.4436",
"0.806",
"0.1765",
"0.806"
],
[
"0.509",
"0.789",
"0.824",
"0.789",
"0.824",
"0.803",
"0.509",
"0.803"
],
[
"0.1765",
"0.8066",
"0.2255",
"0.8066",
"0.2255",
"0.8184",
"0.1765",
"0.8184"
],
[
"0.5073",
"0.8174",
"0.671",
"0.8174",
"0.671",
"0.832",
"0.5073",
"0.832"
],
[
"0.6953",
"0.8164",
"0.8057",
"0.8174",
"0.8057",
"0.832",
"0.6953",
"0.8306"
],
[
"0.1748",
"0.8276",
"0.4436",
"0.8276",
"0.4436",
"0.8413",
"0.1748",
"0.8413"
],
[
"0.524",
"0.8306",
"0.7173",
"0.8306",
"0.7173",
"0.8447",
"0.524",
"0.8447"
],
[
"0.1765",
"0.8413",
"0.4412",
"0.8413",
"0.4412",
"0.855",
"0.1765",
"0.855"
],
[
"0.5073",
"0.845",
"0.536",
"0.845",
"0.536",
"0.857",
"0.5073",
"0.857"
],
[
"0.699",
"0.8447",
"0.8145",
"0.8447",
"0.8145",
"0.8594",
"0.699",
"0.8594"
],
[
"0.1765",
"0.855",
"0.4436",
"0.855",
"0.4436",
"0.8687",
"0.1765",
"0.8687"
],
[
"0.5215",
"0.857",
"0.714",
"0.8594",
"0.714",
"0.8735",
"0.5215",
"0.872"
],
[
"0.1765",
"0.869",
"0.4453",
"0.869",
"0.4453",
"0.8833",
"0.1765",
"0.8833"
],
[
"0.5107",
"0.8735",
"0.5522",
"0.8735",
"0.5522",
"0.884",
"0.5107",
"0.884"
],
[
"0.179",
"0.8853",
"0.4429",
"0.8853",
"0.4429",
"0.896",
"0.179",
"0.896"
],
[
"0.5073",
"0.8853",
"0.615",
"0.8853",
"0.615",
"0.903",
"0.5073",
"0.903"
],
[
"0.1765",
"0.898",
"0.4436",
"0.898",
"0.4436",
"0.911",
"0.1765",
"0.911"
],
[
"0.492",
"0.901",
"0.535",
"0.901",
"0.535",
"0.912",
"0.492",
"0.912"
],
[
"0.1765",
"0.911",
"0.4453",
"0.911",
"0.4453",
"0.925",
"0.1765",
"0.925"
],
[
"0.4756",
"0.911",
"0.5293",
"0.9136",
"0.528",
"0.928",
"0.4739",
"0.9253"
],
[
"0.4927",
"0.9497",
"0.5063",
"0.9497",
"0.5063",
"0.961",
"0.4927",
"0.961"
]
] | [
"Published as a conference paper at ICLR 2023",
"为了解决这些问题，我们提出-DUTD-一种在训练期间动态调整UTD比率的新方法。它的",
"灵感来自早期停止标准和目标，通过调整更新步骤的数量来自动平衡在线欠拟合和过拟合",
"作为方法的一部分，我们将一些经验存储在不用于训练的单独验证缓冲区中。准确地说，每",
"个d环境步骤我们都会从几个单独的事件中收集s的连续转换，这些事件专门用于验证",
"并且每个k环境步骤都会在验证缓冲区上评估世界模型，其中k应比d小得多。由于世界",
"模型学习任务受到监督，因此可以通过在给定的验证序列上记录世界模型的损失来轻松完",
"成。然后将当前验证损失与先前评估的验证损失进行比较。如果损失减少，我们假设模型仍",
"处于欠拟合状态，并将UTD率增加指定的量。如果损失增加，我们假设模型处于过拟合状",
"态，因此降低UTD率。为了在允许的间隔的高更新侧提供更精细的分辨率，我们在对数空",
"间中调整UTD速率，这意味着通过将其分别与c或1/c的值相乘来增加或减少，其中c略",
"大于1。然后，时间步长t的更新公式变为",
"if validation_loss<validation losst-k, (1)",
"utd_ratiot=utd_ratiot-k·b;b=",
"[c,",
"I, if validation_losst ≥validation_losst-k.",
"DUTD是一种通用方法，可应用于以监督方式学习世界模型的任何基于模型的RL算法。",
"实现可以是UTD比率，也可以是数据更新比率，它是它的倒数，我们称之为IUTD(i。e。",
"每个更新步骤的环境步骤数)。如果每个环境步骤执行多个更新，则使用UTD比率更为方",
"便；如果仅在某些环境步骤之后执行更新步骤，则使用IUTD比率更为方便。从方法上讲",
"这两个设置是相同的，因为两个比率描述的是相同的数量，并且只是彼此的倒数",
"图1中显示了DUTD的高级概述，算法1中描述了伪代码，两者都根据IUTD比率进行",
"了解释，因为我们将DUTD应用于DreamerV2算法Hafneretal.（2021），每个环境步骤",
"的几个更新步骤在计算上非常昂贵。然而，在这两个框架中，这两种情况都可以通过让比率",
"为分数来解决",
"3.4Applying DUTD to DreamerV2",
"我们将DUTD应用于DreamerV2",
"Algorithm 1 DUTD (in terms of inverted UTD ra",
"Hafneretal.（2021），这是一种基于",
"tio",
"模型的RL算法，它建立在Dreamer",
"Hafner etal.（2020）上，后者又建立",
"Input: Initial inverted UTD ratio iutd_ratio;",
"在PlaNetHafneretal(2019)上",
"number of steps after which additional validation",
"DreamerV2通过潜在的想象力学习世",
"data is collected d , number of validation transi-",
"界模型。该政策纯粹是通过演员-评论",
"tions collected s, steps after which the iutd_ratio",
"家框架在这个世界模型的潜在空间中",
"is updated k , iutd update increment c",
"学习的。。它是在由世界模型生成的",
"for t=1to total num ofenv_stepsdo",
"假想展开上训练的。将评论家回归到",
"Act according to policy (a I s) and observe",
"入目标 Schulman et al. (2015); Sutton",
"nextstate",
"&Barto（2018）上，并通过Reinforce",
"if t mod d == o then",
"Williams（1992）和动态反向传播损耗",
"Collect s transitions and store experience in",
"的组合来训练演员。世界模型学习",
"a separate validation buffer; increment t =",
"个图像编码器，该编码器将输入映射",
"t+$",
"到分类的潜在状态，在该状态下，递归",
"end if",
"状态空间模型Hafneretal.（2019）学",
"if t mod iutd_ratio == 0 then",
"习动态。在潜在状态下学习图像、奖",
"Perform one training step of the transition",
"励和折扣因子的三个预测因子。世界",
"model",
"模型的总损失是所有三个预测因子的",
"end if",
"if t mod k == 0 then",
"损失和动态预测的潜伏期与编码器的",
"潜伏期之间的Kullback-Leibler损失",
"Compute model loss L on validation dataset",
"的组合",
"if L≥ Lprevious then",
"#Overfitting",
"要应用DUTD，我们评估验证集上的",
"iutd_ratio = iutd_ratio : c",
"图像重建损失。其他选择也是可能的",
"else",
"# Underfitting",
"但我们推测图像预测是世界模型中最",
"iutd_ratio=iutd_ratio/c",
"困难和最重要的部分。人们也可以使",
"end if",
"用不同损失的组合，但随后可能需要",
"Lprevious = L",
"不同损失的比例因子。由于我们希望",
"end if",
"保持我们的方法简单，并防止我们的",
"end for",
"4"
] | {
"arxiv_id": "2303.10144",
"language": "zh",
"page_idx": 3
} |
||
2303.10144-zh-4 | 2303.10144-zh-4.jpg | $\underline{\text{Published as a conference paper at ICLR 2023}}$
<FigureHere>
Figure 2: Aggregated metrics over 5 random seeds on the 26 games of Atari 100 k with $95\%$ confidence intervals according to the method presented in Agarwal et al. (2021) . The intervals are estimated by the percentile bootstrap with statified sampling. Higher mean, median, interquantile mean (IQM) and lower optimality gap are better.
方法需要超参数调整，我们采用单幅
图像丢失。我们实现的源代码是公开可用的$^{1}$
## 4 Experiments
我们在 Atari 100 k 基准 Kaiser et al. (2019)和 DeepMind 控制套件 Tassa et al.(2018)上评估应用于 DreamerV2 的 DUTD。对于两个基准测试中的每一个，我们使用作者在其原始代码库中提供的相应超参数。因此，基线 IUTD 比率设置为控制套件的 5 值和 Atari 的16 值，我们也将其用作我们方法的初始值。这意味着每个 5 和 16 环境步骤分别执行一个更新步骤。对于两个基准测试，我们将 DUTD 的增量值设置为$c= 1. 3$ , 并且每 500 步更新一次 IUTD 比率，该比率对应于控制套件中一集的长度 (跳帧为 2)。每个 100,000 步骤 DUTD 都会收集其他验证数据的3，000转换。我们在控制套件的区间[1,15]和 Atari 的[1,32]中限制 IUTD 比率。这在原则上是没有必要的，我们发现大多数时候边界，特别是上一个，没有达到。通过使用分数并在每个环境步骤中进行多次更新，1以下的边界是可能的，但这对于 DreamerV2 来说在计算上非常昂贵。所有其他超参数在附录中报告。他们没有进行广泛的调整，我们观察到我们的方法的性能在特定选择方面是稳健的。所有报告图中的环境步骤还包括为验证集收集的数据
雅达利 100k 基准 Kaiser et al.(2019)包括来自 Arcade 学习环境 Bellemare et al.(2013) 的 26 游戏，并且每个游戏仅允许代理100，000个环境交互步骤，这些步骤是 400,0000 帧， 跳帧为 4 , 对应大约两个小时的实时游戏。每次运行的最终性能是通过将 100 部署的分数与训练结束后的最终策略进行平均来获得的。我们将每次运行的人类归一化分数计算为$\frac{\text{agent score-random score}}{\text{human score-random score}}$。DeepMind 控制套件为连续控制提供了多种环境。代理接收像素输人，并使用 2 的跳帧操作，与原始 DreamerV2 中一样。我们在大多数环境中为 2 百万帧进行了训练，如果标准 DreamerV2 在该标记之前已经达到其渐近性能，则可以节省 1 百万帧的计算成本。对 10 集的每个 10,000 帧评估策略。对于这两个基准测试，每个算法都在每个环境中使用 5 不同的种子进行训练
我们的实验旨在证明以下内容：
· The UTD ratio can be automatically adjusted using our DUTD approach
$\bullet$ DUTD generally increases performance (up to 300% on Atari100k) by learning an
improved world model compared to the default version of DreamerV2
· DUTD increases the robustness of the RL agent with regard to learning-related hyper-
parameters
$\bullet$ DUTD is competitive with the best UTD hyperparameter found by an extensive grid
search
4.1 Performance of DUTD compared to Standard DreamerV2
对于 Atari100k,图 2显示了使用 Agarwal et al. (2021) 方法在 26 游戏中汇总的结果 ,其中分位数平均值(IQM) 忽略了所有游戏中运行的底部和顶部 25 %,并计算剩余游戏的平均值。最优差距描述了未达到人类水平绩效最小值的数量。在图 11 中，我们展示了每种环境的学
<FigureHere>
$^{1}$https://github.com/Nicolinho/dutd
5
$\mathbf{C}_{\mathrm{omnle~affaionar}}$
Ficure 3. | [
[
"0.1765",
"0.0341",
"0.5073",
"0.0341",
"0.5073",
"0.04736",
"0.1765",
"0.04736"
],
[
"0.3245",
"0.1042",
"0.3635",
"0.1042",
"0.3635",
"0.11426",
"0.3245",
"0.11426"
],
[
"0.468",
"0.1029",
"0.492",
"0.1029",
"0.492",
"0.11426",
"0.468",
"0.11426"
],
[
"0.6006",
"0.1048",
"0.6284",
"0.1048",
"0.6284",
"0.113",
"0.6006",
"0.113"
],
[
"0.7075",
"0.1042",
"0.789",
"0.1042",
"0.789",
"0.11426",
"0.7075",
"0.11426"
],
[
"0.183",
"0.11237",
"0.2769",
"0.113",
"0.2769",
"0.125",
"0.183",
"0.1237"
],
[
"0.5024",
"0.11615",
"0.5215",
"0.11615",
"0.5215",
"0.1206",
"0.5024",
"0.1206"
],
[
"0.6235",
"0.11743",
"0.63",
"0.11743",
"0.63",
"0.1206",
"0.6235",
"0.1206"
],
[
"0.634",
"0.11743",
"0.6567",
"0.11743",
"0.6567",
"0.1206",
"0.634",
"0.1206"
],
[
"0.687",
"0.11554",
"0.721",
"0.11554",
"0.721",
"0.125",
"0.687",
"0.125"
],
[
"0.2157",
"0.125",
"0.2793",
"0.125",
"0.2793",
"0.1351",
"0.2157",
"0.1351"
],
[
"0.4387",
"0.1351",
"0.5312",
"0.1351",
"0.5312",
"0.1459",
"0.4387",
"0.1459"
],
[
"0.687",
"0.1351",
"0.8096",
"0.1351",
"0.8096",
"0.1459",
"0.687",
"0.1459"
],
[
"0.5513",
"0.1364",
"0.572",
"0.1364",
"0.572",
"0.1433",
"0.5513",
"0.1433"
],
[
"0.29",
"0.1364",
"0.3162",
"0.1364",
"0.3162",
"0.1445",
"0.29",
"0.1445"
],
[
"0.3284",
"0.137",
"0.3489",
"0.137",
"0.3489",
"0.1445",
"0.3284",
"0.1445"
],
[
"0.367",
"0.137",
"0.3865",
"0.137",
"0.3865",
"0.1445",
"0.367",
"0.1445"
],
[
"0.581",
"0.1364",
"0.6045",
"0.1364",
"0.6045",
"0.1433",
"0.581",
"0.1433"
],
[
"0.613",
"0.137",
"0.6724",
"0.137",
"0.6724",
"0.1445",
"0.613",
"0.1445"
],
[
"0.4805",
"0.1459",
"0.614",
"0.1459",
"0.614",
"0.1559",
"0.4805",
"0.1559"
],
[
"0.574",
"0.1395",
"0.581",
"0.1395",
"0.581",
"0.1427",
"0.574",
"0.1427"
],
[
"0.1765",
"0.1749",
"0.8228",
"0.1749",
"0.8228",
"0.1881",
"0.1765",
"0.1881"
],
[
"0.1748",
"0.1881",
"0.824",
"0.1881",
"0.824",
"0.2026",
"0.1748",
"0.2026"
],
[
"0.1748",
"0.2026",
"0.8228",
"0.2026",
"0.8228",
"0.2166",
"0.1748",
"0.2166"
],
[
"0.1748",
"0.2172",
"0.687",
"0.2172",
"0.687",
"0.2311",
"0.1748",
"0.2311"
],
[
"0.1765",
"0.2583",
"0.4436",
"0.2583",
"0.4436",
"0.2715",
"0.1765",
"0.2715"
],
[
"0.1765",
"0.2715",
"0.4983",
"0.2715",
"0.4983",
"0.2854",
"0.1765",
"0.2854"
],
[
"0.1748",
"0.303",
"0.2035",
"0.303",
"0.2035",
"0.315",
"0.1748",
"0.315"
],
[
"0.2075",
"0.303",
"0.3186",
"0.303",
"0.3186",
"0.3164",
"0.2075",
"0.3164"
],
[
"0.1765",
"0.3276",
"0.8228",
"0.3276",
"0.8228",
"0.3416",
"0.1765",
"0.3416"
],
[
"0.1765",
"0.3416",
"0.6323",
"0.3416",
"0.6323",
"0.3547",
"0.1765",
"0.3547"
],
[
"0.6313",
"0.3423",
"0.822",
"0.3423",
"0.822",
"0.3528",
"0.6313",
"0.3528"
],
[
"0.1765",
"0.3547",
"0.8228",
"0.3547",
"0.8228",
"0.3694",
"0.1765",
"0.3694"
],
[
"0.1765",
"0.3694",
"0.822",
"0.3694",
"0.822",
"0.3833",
"0.1765",
"0.3833"
],
[
"0.1772",
"0.3838",
"0.8228",
"0.3838",
"0.8228",
"0.3977",
"0.1772",
"0.3977"
],
[
"0.1765",
"0.3965",
"0.8228",
"0.3965",
"0.8228",
"0.411",
"0.1765",
"0.411"
],
[
"0.1765",
"0.411",
"0.8228",
"0.411",
"0.8228",
"0.4248",
"0.1765",
"0.4248"
],
[
"0.1772",
"0.4248",
"0.8228",
"0.4248",
"0.8228",
"0.4382",
"0.1772",
"0.4382"
],
[
"0.179",
"0.4382",
"0.8228",
"0.4375",
"0.8228",
"0.452",
"0.179",
"0.4526"
],
[
"0.1748",
"0.4526",
"0.8228",
"0.4526",
"0.8228",
"0.4666",
"0.1748",
"0.4666"
],
[
"0.1765",
"0.4666",
"0.822",
"0.4666",
"0.822",
"0.4797",
"0.1765",
"0.4797"
],
[
"0.179",
"0.4817",
"0.468",
"0.4817",
"0.468",
"0.4924",
"0.179",
"0.4924"
],
[
"0.1765",
"0.502",
"0.8228",
"0.502",
"0.8228",
"0.515",
"0.1765",
"0.515"
],
[
"0.1765",
"0.515",
"0.822",
"0.515",
"0.822",
"0.529",
"0.1765",
"0.529"
],
[
"0.1765",
"0.53",
"0.822",
"0.53",
"0.822",
"0.5435",
"0.1765",
"0.5435"
],
[
"0.1765",
"0.5435",
"0.8228",
"0.5435",
"0.8228",
"0.5566",
"0.1765",
"0.5566"
],
[
"0.1814",
"0.565",
"0.3381",
"0.565",
"0.3381",
"0.5728",
"0.1814",
"0.5728"
],
[
"0.3308",
"0.5596",
"0.822",
"0.5596",
"0.822",
"0.5728",
"0.3308",
"0.5728"
],
[
"0.1765",
"0.5747",
"0.822",
"0.5747",
"0.822",
"0.5884",
"0.1765",
"0.5884"
],
[
"0.1748",
"0.5884",
"0.822",
"0.5884",
"0.822",
"0.6016",
"0.1748",
"0.6016"
],
[
"0.1765",
"0.603",
"0.8228",
"0.603",
"0.8228",
"0.616",
"0.1765",
"0.616"
],
[
"0.1765",
"0.616",
"0.4412",
"0.616",
"0.4412",
"0.63",
"0.1765",
"0.63"
],
[
"0.1765",
"0.638",
"0.3962",
"0.638",
"0.3962",
"0.6514",
"0.1765",
"0.6514"
],
[
"0.1896",
"0.6655",
"0.7344",
"0.6655",
"0.7344",
"0.6787",
"0.1896",
"0.6787"
],
[
"0.1871",
"0.683",
"0.824",
"0.6846",
"0.824",
"0.6987",
"0.1871",
"0.6978"
],
[
"0.2059",
"0.6987",
"0.706",
"0.6987",
"0.706",
"0.712",
"0.2059",
"0.712"
],
[
"0.1887",
"0.7183",
"0.822",
"0.7183",
"0.822",
"0.732",
"0.1887",
"0.732"
],
[
"0.2059",
"0.7344",
"0.2876",
"0.733",
"0.2876",
"0.744",
"0.2059",
"0.7456"
],
[
"0.1854",
"0.7505",
"0.824",
"0.752",
"0.824",
"0.7666",
"0.1854",
"0.765"
],
[
"0.2075",
"0.7676",
"0.2532",
"0.7676",
"0.2532",
"0.779",
"0.2075",
"0.779"
],
[
"0.1732",
"0.802",
"0.6323",
"0.803",
"0.6323",
"0.8174",
"0.1732",
"0.8164"
],
[
"0.1765",
"0.83",
"0.5664",
"0.83",
"0.5664",
"0.8433",
"0.1765",
"0.8433"
],
[
"0.1772",
"0.8433",
"0.568",
"0.8433",
"0.568",
"0.857",
"0.1772",
"0.857"
],
[
"0.5986",
"0.8486",
"0.6167",
"0.8486",
"0.6167",
"0.858",
"0.5986",
"0.858"
],
[
"0.1772",
"0.858",
"0.5664",
"0.858",
"0.5664",
"0.872",
"0.1772",
"0.872"
],
[
"0.1765",
"0.872",
"0.568",
"0.872",
"0.568",
"0.8853",
"0.1765",
"0.8853"
],
[
"0.1765",
"0.8853",
"0.568",
"0.8853",
"0.568",
"0.8984",
"0.1765",
"0.8984"
],
[
"0.584",
"0.898",
"0.6006",
"0.898",
"0.6006",
"0.9136",
"0.584",
"0.9136"
],
[
"0.6006",
"0.8945",
"0.615",
"0.8945",
"0.615",
"0.9014",
"0.6006",
"0.9014"
],
[
"0.1953",
"0.909",
"0.4395",
"0.91",
"0.4395",
"0.925",
"0.1953",
"0.924"
],
[
"0.5986",
"0.9165",
"0.615",
"0.9165",
"0.615",
"0.9253",
"0.5986",
"0.9253"
],
[
"0.716",
"0.9194",
"0.8105",
"0.9194",
"0.8105",
"0.931",
"0.716",
"0.931"
],
[
"0.697",
"0.9297",
"0.806",
"0.9316",
"0.806",
"0.9424",
"0.697",
"0.94"
],
[
"0.492",
"0.9478",
"0.509",
"0.9478",
"0.509",
"0.962",
"0.492",
"0.962"
],
[
"0.625",
"0.9507",
"0.6396",
"0.9507",
"0.6396",
"0.9604",
"0.625",
"0.9604"
],
[
"0.6685",
"0.952",
"0.6846",
"0.952",
"0.6846",
"0.959",
"0.6685",
"0.959"
],
[
"0.714",
"0.952",
"0.7295",
"0.952",
"0.7295",
"0.959",
"0.714",
"0.959"
],
[
"0.7573",
"0.952",
"0.773",
"0.952",
"0.773",
"0.959",
"0.7573",
"0.959"
],
[
"0.8022",
"0.9507",
"0.819",
"0.9507",
"0.819",
"0.9604",
"0.8022",
"0.9604"
],
[
"0.649",
"0.9585",
"0.7935",
"0.9604",
"0.7935",
"0.9717",
"0.649",
"0.9697"
],
[
"0.5864",
"0.992",
"0.625",
"0.992",
"0.625",
"0.997",
"0.5864",
"0.997"
]
] | [
"Published as a conference paper at ICLR 2023",
"Median",
"IQM",
"Mean",
"OptimalityGap",
"DUTD-DreamerV2",
"",
"",
"",
"一",
"DreamerV2",
"0.160.240.32",
"0.600.650.700.75",
"0.6",
"0.1",
"0.2",
"0.3",
"0.8",
"1.01.2",
"HumanNormalizedScore",
"",
"Figure 2:.Aggregated metrics over 5 random seeds on the 26 games of Atari 100 k with",
"95 % confidence intervals according to the method presented in Agarwal et al. (2021) . The",
"intervals are estimated by the percentile bootstrap with statified sampling. Higher mean,",
"median, interquantile mean (IQM) and lower optimality gap are better.",
"方法需要超参数调整，我们采用单幅",
"图像丢失。我们实现的源代码是公开可用的1",
"4",
"Experiments",
"我们在Atari100k基准Kaiseretal.（2019）和DeepMind控制套件Tassaetal:（2018）上",
"评估应用于DreamerV2的DUTD。对于两个基准测试中的每一",
"一个，我们使用作者在其原",
"始代码库中提供的相应超参数。因此，基线IUTD比率设置为控制套件的5值和Atari的",
"16值，我们也将其用作我们方法的初始值。这意味着每个5和16环境步骤分别执行一个",
"更新步骤。对于两个基准测试，我们将DUTD的增量值设置为c=1.3，并且每500步更",
"新一次IUTD比率，该比率对应于控制套件中一集的长度（跳顿为2）。每个100,000步",
"骤DUTD都会收集其他验证数据的3,000转换。我们在控制套件的区间[1,15]和Atari的",
"[1,32]中限制IUTD比率。这在原则上是没有必要的，我们发现大多数时候边界，特别是",
"上一个，没有达到。通过使用分数并在每个环境步骤中进行多次更新，1以下的边界是可能",
"的，但这对于DreamerV2来说在计算上非常昂贵。所有其他超参数在附录中报告。他们没",
"有进行广泛的调整，我们观察到我们的方法的性能在特定选择方面是稳健的。所有报告图",
"中的环境步骤还包括为验证集收集的数据",
"雅达利100k基准Kaiseretal.（2019）包括来自Arcade学习环境Bellemareetal.（2013)",
"的26游戏，并且每个游戏仅允许代理100,000个环境交互步骤，这些步骤是400,000顿",
"跳顿为4，对应大约两个小时的实时游戏。每次运行的最终性能是通过将100部署的分",
"数与训练结束后的最终策略进行平均来获得的。我们将每次运行的人类归一化分数计算为",
"uman.score-random score",
"。DeepMind控制套件为连续控制提供了多种环境。代理接收像素输",
"人，并使用2的跳顿操作，与原始DreamerV2中一样。我们在大多数环境中为2百万顿进",
"行了训练，如果标准DreamerV2在该标记之前已经达到其渐近性能，则可以节省1百万顿",
"的计算成本。对10集的每个10,000顿评估策略。对于这两个基准测试，每个算法都在每",
"个环境中使用5不同的种子进行训练",
"我们的实验旨在证明以下内容：",
". The UTD ratio can be automatically adjusted using our DUTD approach",
". DUTD generally increases performance (up to 300 % on Atari100k) by learning an",
"improved world model compared to the default version of DreamerV2",
": DUTD increases the robustness of the RL agent with regard to learning-related hyper-",
"parameters",
". DUTD is competitive with the best UTD hyperparameter found by an extensive grid",
"search",
"4.1Performance of DUTD compared to Standard DreamerV2",
"对于Atari100k，图2显示了使用Agarwaletal.（2021）",
"方法在26游戏中汇总的结果，其中分位数平均值（IQM）",
"800",
"忽略了所有游戏中运行的底部和顶部25%，并计算剩",
"余游戏的平均值。最优差距描述了未达到人类水平绩效",
"最小值的数量。在图11中，我们展示了每种环境的学",
"",
"400",
"Ihttps://github.com/Nicolinho/dutd",
"200",
"DreamerV2-IUTD_5",
"DUTD-DreamerV2",
"5",
"0.0",
"0.5",
"10",
"1.5",
"2.0",
"EnvironmentSteps (inmilions)",
"P:01"
] | {
"arxiv_id": "2303.10144",
"language": "zh",
"page_idx": 4
} |
||
2204.03025-zh-3 | 2204.03025-zh-3.jpg | <FigureHere>
Figure 1: Users interact with the deployed QA model and give feedback. Feedback contains a rating (
bad, good, could be improved, excellent ) and a natural language explanation.
辨别正确和错误答案的解释吗？
为了解决这些问题，我们使用反馈数据来训练重新排序原始答案的模型，并为答案提供解释。我们的实验表明，这种方法不仅提高了收集反馈的基本 QA 模型的准确性，而且还提高了未收集反馈数据的其他强模型的准确性。此外，我们进行人工评估以验证解释的有用性并发现生成的自然语言解释有助于用户在接受或拒绝候选答案时做出明智和准确的决定
我们的贡献如下：
1. We create the frst retrieval-based QA
dataset containing interactive feedback. $2.{\text{ We demonstrate a simple method of using}}$
the feedback data to increase the accuracy
and explainability of RQA systems.
3. We show that the feedback data not only
improve the deployed model but also a
stronger non-deployed model.
ment , Canadian government , $^3$和 THMAX 的官方网站上抓取了 Covid-19 相关内容。我们在常见问题部分提取问题和答案段落。为了扩大数据集，我们还清理了抓取的页面并提取了其他段落，我们使用众包为这些段落策划了相应的问题，就像用户在问问题一样。我们在 Appendix A 中介绍了此注释过程的详细信息。我们使用此数据集分别为每个源训练基本 RQA 模型并部署它们。对于基本模型，我们使用基于 BERT 的密集检索器 (Karpukhin et al., 2020) 结合 Poly-encoder (Miller et al. $2017)_{\text{企业会}}($更多详细信息在 $Section 3. 1$中$)_{\mathrm{R}}$
此，我们事先收集与第 1 阶段类似的单个段落的问题，并将其用作交互式问题。向用户显示模型的问题和前 2 个预测，并为每个问答对提供反馈。收集的反馈包括从优秀，良好，可以改进，不良和自然语言解释中选择的评分，详细说明答案的优点和/或缺点。对于每个 QA 对，我们从三个不同的工人那里得到反馈。我们采用了其他策略来确保反馈数据的质量，其详细信息可在 Appendix B 中获得。生成的数据集统计信息显示在 Table 2 中。为了测试交互式反馈是否也有助于分发外设置，我们没有收集其中一个域的反馈(加拿大)
2.2 FeedbackQA analysis
## 2 FeedbackQA Dataset
最近，人们一直在努力以自然语言理解任务的解释形式收集反馈数据$\overline{\mathrm{E}}$(Camburu et al. 2018; Rajani et al. 2019 等)。这些仅包含对从训练数据中采样的给定输入的地面实况预测的解释，而无需任何用户系统交互。相反，我们在部署 RQA 系统后收集用户反馈，从而收集正确和错误预测的反馈。Table 1对 FeedbackQA 和现有的自然语言理解(NLU)数据集进行了全面的比较，并提供了解释数据。
Table 3 显示了反馈数据的示例，包括评级和解释。我们发现，解释通常包含表示答案质量的评论式文本，或总结哪些部分是正确的以及为什么的陈述。因此，我们使用以下模式分析示例解释：
评论几个解释以通用评论开头，例如这直接回答了问题或与问题无关。有时，用户还会突出显示答案中不错或可以改进的方面。例如，。可以提高语法。建议在写作方面可以改进答案。有用内容摘要是指答案中实际回答问题的部
## 2.1 Dataset collection
为了像在实际环境中一样收集部署后反馈，我们将数据收集分为两个阶段：(RQA 模型的) 部署前和部署后。
Stage 1: Pre-deployment of a QA system 我们从 WHO , US Government , UK Govern-
$^{3}$We focus on the Province of Quebec | [
[
"0.6963",
"0.1069",
"0.869",
"0.1069",
"0.869",
"0.11755",
"0.6963",
"0.11755"
],
[
"0.6943",
"0.1164",
"0.846",
"0.1164",
"0.846",
"0.12354",
"0.6943",
"0.12354"
],
[
"0.3477",
"0.1312",
"0.388",
"0.1318",
"0.388",
"0.1395",
"0.3477",
"0.1383"
],
[
"0.692",
"0.1312",
"0.8623",
"0.1318",
"0.8623",
"0.1437",
"0.692",
"0.1426"
],
[
"0.6943",
"0.1259",
"0.8657",
"0.1259",
"0.8657",
"0.133",
"0.6943",
"0.133"
],
[
"0.3997",
"0.133",
"0.4282",
"0.133",
"0.4282",
"0.1372",
"0.3997",
"0.1372"
],
[
"0.3914",
"0.1426",
"0.398",
"0.1426",
"0.398",
"0.1455",
"0.3914",
"0.1455"
],
[
"0.409",
"0.1426",
"0.4165",
"0.1426",
"0.4165",
"0.1455",
"0.409",
"0.1455"
],
[
"0.4207",
"0.1426",
"0.4458",
"0.1426",
"0.4458",
"0.1455",
"0.4207",
"0.1455"
],
[
"0.69",
"0.1414",
"0.854",
"0.1414",
"0.854",
"0.152",
"0.69",
"0.152"
],
[
"0.3972",
"0.1497",
"0.4324",
"0.1497",
"0.4324",
"0.1538",
"0.3972",
"0.1538"
],
[
"0.693",
"0.152",
"0.848",
"0.152",
"0.848",
"0.1592",
"0.693",
"0.1592"
],
[
"0.3105",
"0.158",
"0.315",
"0.158",
"0.315",
"0.1633",
"0.3105",
"0.1633"
],
[
"0.3855",
"0.158",
"0.434",
"0.158",
"0.434",
"0.1621",
"0.3855",
"0.1621"
],
[
"0.706",
"0.1609",
"0.729",
"0.1609",
"0.729",
"0.165",
"0.706",
"0.165"
],
[
"0.1503",
"0.1716",
"0.2301",
"0.1716",
"0.2301",
"0.1829",
"0.1503",
"0.1829"
],
[
"0.3284",
"0.1633",
"0.3645",
"0.1644",
"0.3645",
"0.1716",
"0.3284",
"0.1704"
],
[
"0.3257",
"0.1746",
"0.3333",
"0.1746",
"0.3333",
"0.1776",
"0.3257",
"0.1776"
],
[
"0.3594",
"0.1729",
"0.3955",
"0.1729",
"0.3955",
"0.1799",
"0.3594",
"0.1799"
],
[
"0.5366",
"0.1758",
"0.617",
"0.1758",
"0.617",
"0.1859",
"0.5366",
"0.1859"
],
[
"0.1453",
"0.1818",
"0.2241",
"0.1818",
"0.2241",
"0.1924",
"0.1453",
"0.1924"
],
[
"0.54",
"0.1859",
"0.613",
"0.1859",
"0.613",
"0.1965",
"0.54",
"0.1965"
],
[
"0.7188",
"0.1818",
"0.763",
"0.1799",
"0.763",
"0.187",
"0.7188",
"0.1895"
],
[
"0.309",
"0.1907",
"0.357",
"0.1907",
"0.357",
"0.1978",
"0.309",
"0.1978"
],
[
"0.6963",
"0.1895",
"0.706",
"0.1924",
"0.7036",
"0.1978",
"0.693",
"0.1948"
],
[
"0.76",
"0.1912",
"0.784",
"0.1912",
"0.784",
"0.1953",
"0.76",
"0.1953"
],
[
"0.712",
"0.1924",
"0.756",
"0.1924",
"0.756",
"0.1953",
"0.712",
"0.1953"
],
[
"0.1511",
"0.1912",
"0.2275",
"0.1912",
"0.2275",
"0.199",
"0.1511",
"0.199"
],
[
"0.5366",
"0.1936",
"0.617",
"0.1936",
"0.617",
"0.2037",
"0.5366",
"0.2037"
],
[
"0.79",
"0.1924",
"0.8154",
"0.1924",
"0.8154",
"0.1953",
"0.79",
"0.1953"
],
[
"0.1511",
"0.1996",
"0.215",
"0.1996",
"0.215",
"0.2073",
"0.1511",
"0.2073"
],
[
"0.3687",
"0.1996",
"0.3862",
"0.1996",
"0.3862",
"0.2037",
"0.3687",
"0.2037"
],
[
"0.4358",
"0.1996",
"0.455",
"0.1996",
"0.455",
"0.2037",
"0.4358",
"0.2037"
],
[
"0.712",
"0.1996",
"0.7407",
"0.1996",
"0.7407",
"0.2037",
"0.712",
"0.2037"
],
[
"0.748",
"0.1996",
"0.756",
"0.1996",
"0.756",
"0.2037",
"0.748",
"0.2037"
],
[
"0.784",
"0.2007",
"0.802",
"0.2007",
"0.802",
"0.2037",
"0.784",
"0.2037"
],
[
"0.802",
"0.2007",
"0.8447",
"0.2007",
"0.8447",
"0.2037",
"0.802",
"0.2037"
],
[
"0.5366",
"0.2019",
"0.6035",
"0.2031",
"0.6035",
"0.2102",
"0.5366",
"0.209"
],
[
"0.315",
"0.2007",
"0.3257",
"0.2007",
"0.3257",
"0.2037",
"0.315",
"0.2037"
],
[
"0.3855",
"0.2079",
"0.455",
"0.2079",
"0.455",
"0.212",
"0.3855",
"0.212"
],
[
"0.825",
"0.209",
"0.8423",
"0.209",
"0.8423",
"0.2131",
"0.825",
"0.2131"
],
[
"0.825",
"0.2131",
"0.8447",
"0.2173",
"0.8423",
"0.2239",
"0.8228",
"0.2203"
],
[
"0.6943",
"0.2144",
"0.7197",
"0.2162",
"0.7188",
"0.2239",
"0.693",
"0.2214"
],
[
"0.315",
"0.2162",
"0.3242",
"0.2162",
"0.3242",
"0.2203",
"0.315",
"0.2203"
],
[
"0.3242",
"0.2144",
"0.4106",
"0.2144",
"0.4106",
"0.2214",
"0.3242",
"0.2214"
],
[
"0.409",
"0.2173",
"0.434",
"0.2173",
"0.434",
"0.2203",
"0.409",
"0.2203"
],
[
"0.7305",
"0.2173",
"0.7617",
"0.2173",
"0.7617",
"0.2203",
"0.7305",
"0.2203"
],
[
"0.7773",
"0.2173",
"0.8003",
"0.2173",
"0.8003",
"0.2214",
"0.7773",
"0.2214"
],
[
"0.8076",
"0.2173",
"0.8193",
"0.2173",
"0.8193",
"0.2214",
"0.8076",
"0.2214"
],
[
"0.8237",
"0.2227",
"0.8447",
"0.2257",
"0.8423",
"0.2322",
"0.8213",
"0.2286"
],
[
"0.712",
"0.2257",
"0.7524",
"0.2257",
"0.7524",
"0.2299",
"0.712",
"0.2299"
],
[
"0.3225",
"0.2227",
"0.4534",
"0.2227",
"0.4534",
"0.2299",
"0.3225",
"0.2299"
],
[
"0.754",
"0.2268",
"0.7656",
"0.2268",
"0.7656",
"0.2299",
"0.754",
"0.2299"
],
[
"0.7773",
"0.2268",
"0.8154",
"0.2268",
"0.8154",
"0.231",
"0.7773",
"0.231"
],
[
"0.693",
"0.231",
"0.756",
"0.2322",
"0.756",
"0.2405",
"0.693",
"0.2393"
],
[
"0.3257",
"0.2328",
"0.339",
"0.2328",
"0.339",
"0.2363",
"0.3257",
"0.2363"
],
[
"0.3376",
"0.231",
"0.388",
"0.231",
"0.388",
"0.2382",
"0.3376",
"0.2382"
],
[
"0.7524",
"0.2328",
"0.8423",
"0.2328",
"0.8423",
"0.2405",
"0.7524",
"0.2405"
],
[
"0.6943",
"0.2411",
"0.754",
"0.2411",
"0.754",
"0.2488",
"0.6943",
"0.2488"
],
[
"0.769",
"0.2411",
"0.8423",
"0.2405",
"0.8423",
"0.2477",
"0.769",
"0.2488"
],
[
"0.69",
"0.2477",
"0.79",
"0.2488",
"0.79",
"0.2588",
"0.69",
"0.2583"
],
[
"0.11755",
"0.2778",
"0.881",
"0.279",
"0.881",
"0.2935",
"0.11755",
"0.2922"
],
[
"0.11755",
"0.2935",
"0.692",
"0.2935",
"0.692",
"0.308",
"0.11755",
"0.308"
],
[
"0.11755",
"0.3342",
"0.3652",
"0.3342",
"0.3652",
"0.3474",
"0.11755",
"0.3474"
],
[
"0.513",
"0.3342",
"0.882",
"0.3342",
"0.882",
"0.3474",
"0.513",
"0.3474"
],
[
"0.1377",
"0.3508",
"0.486",
"0.3508",
"0.486",
"0.3652",
"0.1377",
"0.3652"
],
[
"0.511",
"0.3499",
"0.881",
"0.3499",
"0.881",
"0.3635",
"0.511",
"0.3635"
],
[
"0.1184",
"0.3677",
"0.486",
"0.3677",
"0.486",
"0.3813",
"0.1184",
"0.3813"
],
[
"0.513",
"0.3652",
"0.882",
"0.3677",
"0.882",
"0.3813",
"0.513",
"0.379"
],
[
"0.1184",
"0.3843",
"0.4844",
"0.3843",
"0.4844",
"0.398",
"0.1184",
"0.398"
],
[
"0.5156",
"0.383",
"0.879",
"0.383",
"0.879",
"0.3967",
"0.5156",
"0.3967"
],
[
"0.11755",
"0.3997",
"0.4836",
"0.3997",
"0.4836",
"0.4133",
"0.11755",
"0.4133"
],
[
"0.511",
"0.3984",
"0.881",
"0.3984",
"0.881",
"0.412",
"0.511",
"0.412"
],
[
"0.1234",
"0.415",
"0.4878",
"0.415",
"0.4878",
"0.4287",
"0.1234",
"0.4287"
],
[
"0.5146",
"0.4163",
"0.881",
"0.4163",
"0.881",
"0.427",
"0.5146",
"0.427"
],
[
"0.12006",
"0.4312",
"0.4895",
"0.4312",
"0.4895",
"0.4448",
"0.12006",
"0.4448"
],
[
"0.5146",
"0.43",
"0.881",
"0.43",
"0.881",
"0.4448",
"0.5146",
"0.4448"
],
[
"0.1184",
"0.4478",
"0.486",
"0.4478",
"0.486",
"0.4614",
"0.1184",
"0.4614"
],
[
"0.513",
"0.4478",
"0.881",
"0.4478",
"0.881",
"0.4614",
"0.513",
"0.4614"
],
[
"0.1184",
"0.4644",
"0.4492",
"0.4644",
"0.4492",
"0.478",
"0.1184",
"0.478"
],
[
"0.513",
"0.4631",
"0.882",
"0.4631",
"0.882",
"0.4768",
"0.513",
"0.4768"
],
[
"0.1368",
"0.481",
"0.2737",
"0.481",
"0.2737",
"0.4946",
"0.1368",
"0.4946"
],
[
"0.513",
"0.478",
"0.882",
"0.4785",
"0.882",
"0.4934",
"0.513",
"0.4922"
],
[
"0.1293",
"0.4963",
"0.486",
"0.4963",
"0.486",
"0.5103",
"0.1293",
"0.5103"
],
[
"0.51",
"0.4946",
"0.8774",
"0.4946",
"0.8774",
"0.5083",
"0.51",
"0.5083"
],
[
"0.513",
"0.511",
"0.837",
"0.511",
"0.837",
"0.5254",
"0.513",
"0.5254"
],
[
"0.8506",
"0.5127",
"0.8774",
"0.5127",
"0.8774",
"0.524",
"0.8506",
"0.524"
],
[
"0.511",
"0.5195",
"0.882",
"0.521",
"0.882",
"0.538",
"0.511",
"0.5376"
],
[
"0.1528",
"0.513",
"0.4727",
"0.513",
"0.4727",
"0.528",
"0.1528",
"0.528"
],
[
"0.1293",
"0.528",
"0.4878",
"0.53",
"0.4878",
"0.5435",
"0.1293",
"0.5415"
],
[
"0.513",
"0.536",
"0.881",
"0.5376",
"0.881",
"0.552",
"0.513",
"0.5503"
],
[
"0.1528",
"0.546",
"0.486",
"0.546",
"0.486",
"0.5586",
"0.1528",
"0.5586"
],
[
"0.511",
"0.552",
"0.879",
"0.5527",
"0.879",
"0.568",
"0.511",
"0.5664"
],
[
"0.1545",
"0.561",
"0.4417",
"0.561",
"0.4417",
"0.5747",
"0.1545",
"0.5747"
],
[
"0.5146",
"0.5693",
"0.881",
"0.5693",
"0.881",
"0.583",
"0.5146",
"0.583"
],
[
"0.1301",
"0.576",
"0.4836",
"0.576",
"0.4836",
"0.589",
"0.1301",
"0.589"
],
[
"0.5156",
"0.5874",
"0.879",
"0.5874",
"0.879",
"0.597",
"0.5156",
"0.597"
],
[
"0.1545",
"0.5933",
"0.486",
"0.5933",
"0.486",
"0.607",
"0.1545",
"0.607"
],
[
"0.513",
"0.601",
"0.879",
"0.601",
"0.879",
"0.6143",
"0.513",
"0.6143"
],
[
"0.1511",
"0.6094",
"0.392",
"0.608",
"0.392",
"0.6226",
"0.1511",
"0.6235"
],
[
"0.5146",
"0.618",
"0.881",
"0.618",
"0.881",
"0.632",
"0.5146",
"0.632"
],
[
"0.1511",
"0.638",
"0.3376",
"0.638",
"0.3376",
"0.653",
"0.1511",
"0.653"
],
[
"0.5156",
"0.6357",
"0.879",
"0.6357",
"0.879",
"0.6465",
"0.5156",
"0.6465"
],
[
"0.5146",
"0.651",
"0.881",
"0.651",
"0.881",
"0.6646",
"0.5146",
"0.6646"
],
[
"0.11755",
"0.6396",
"0.1511",
"0.6396",
"0.1511",
"0.651",
"0.11755",
"0.651"
],
[
"0.12006",
"0.6665",
"0.482",
"0.6665",
"0.482",
"0.677",
"0.12006",
"0.677"
],
[
"0.511",
"0.665",
"0.879",
"0.665",
"0.879",
"0.6787",
"0.511",
"0.6787"
],
[
"0.1184",
"0.681",
"0.4844",
"0.681",
"0.4844",
"0.695",
"0.1184",
"0.695"
],
[
"0.5146",
"0.683",
"0.879",
"0.683",
"0.879",
"0.697",
"0.5146",
"0.697"
],
[
"0.1184",
"0.697",
"0.4844",
"0.697",
"0.4844",
"0.7114",
"0.1184",
"0.7114"
],
[
"0.511",
"0.6987",
"0.7793",
"0.6987",
"0.7793",
"0.712",
"0.511",
"0.712"
],
[
"0.12006",
"0.7134",
"0.486",
"0.7134",
"0.486",
"0.727",
"0.12006",
"0.727"
],
[
"0.509",
"0.7144",
"0.733",
"0.7153",
"0.733",
"0.73",
"0.509",
"0.7285"
],
[
"0.1184",
"0.7285",
"0.4844",
"0.7285",
"0.4844",
"0.742",
"0.1184",
"0.742"
],
[
"0.511",
"0.7363",
"0.881",
"0.7363",
"0.881",
"0.75",
"0.511",
"0.75"
],
[
"0.12006",
"0.745",
"0.4844",
"0.745",
"0.4844",
"0.759",
"0.12006",
"0.759"
],
[
"0.5146",
"0.7544",
"0.879",
"0.7544",
"0.879",
"0.764",
"0.5146",
"0.764"
],
[
"0.1184",
"0.761",
"0.4844",
"0.761",
"0.4844",
"0.775",
"0.1184",
"0.775"
],
[
"0.513",
"0.7686",
"0.881",
"0.7686",
"0.881",
"0.783",
"0.513",
"0.783"
],
[
"0.1184",
"0.778",
"0.482",
"0.778",
"0.482",
"0.7915",
"0.1184",
"0.7915"
],
[
"0.5146",
"0.785",
"0.881",
"0.785",
"0.881",
"0.799",
"0.5146",
"0.799"
],
[
"0.1184",
"0.792",
"0.3855",
"0.792",
"0.3855",
"0.8057",
"0.1184",
"0.8057"
],
[
"0.511",
"0.8003",
"0.5938",
"0.8003",
"0.5938",
"0.8154",
"0.511",
"0.8154"
],
[
"0.11755",
"0.8154",
"0.1628",
"0.8154",
"0.1628",
"0.8267",
"0.11755",
"0.8267"
],
[
"0.1561",
"0.814",
"0.3105",
"0.814",
"0.3105",
"0.8276",
"0.1561",
"0.8276"
],
[
"0.513",
"0.818",
"0.881",
"0.818",
"0.881",
"0.8306",
"0.513",
"0.8306"
],
[
"0.1184",
"0.836",
"0.486",
"0.836",
"0.486",
"0.8496",
"0.1184",
"0.8496"
],
[
"0.5146",
"0.8345",
"0.881",
"0.8345",
"0.881",
"0.847",
"0.5146",
"0.847"
],
[
"0.11755",
"0.851",
"0.4878",
"0.8516",
"0.4878",
"0.866",
"0.11755",
"0.865"
],
[
"0.5146",
"0.8496",
"0.879",
"0.8496",
"0.879",
"0.8633",
"0.5146",
"0.8633"
],
[
"0.1184",
"0.868",
"0.2551",
"0.868",
"0.2551",
"0.882",
"0.1184",
"0.882"
],
[
"0.513",
"0.865",
"0.8774",
"0.865",
"0.8774",
"0.879",
"0.513",
"0.879"
],
[
"0.513",
"0.8813",
"0.882",
"0.8813",
"0.882",
"0.8955",
"0.513",
"0.8955"
],
[
"0.1234",
"0.8896",
"0.4878",
"0.8896",
"0.4878",
"0.903",
"0.1234",
"0.903"
],
[
"0.1184",
"0.905",
"0.4895",
"0.905",
"0.4895",
"0.9185",
"0.1184",
"0.9185"
],
[
"0.534",
"0.905",
"0.7886",
"0.905",
"0.7886",
"0.9185",
"0.534",
"0.9185"
]
] | [
"A,:Duringrespiratorshortages,care",
"shouldbetakentoensurethalNo",
"veringyo",
"whererespiratoryprotection is most",
"respiratorsarereservedforsituations",
"babvis",
"10",
"",
"ran",
"important,suchasperformanceof",
"nalsifs",
"aerosol-gcneratingprocedureson",
"-",
"hehestnla",
"ient",
"Q:How should an",
"baby.di",
"+",
"provider",
"Q:How shouldan",
"OB/GYNsetting",
"OB/GYNsetting",
"gExceller",
"RatingBac",
"EX",
"This",
"anatinn",
"handleashortage",
"handle a shortage",
"ancu",
"ofrespirators?",
"This",
"bad",
"sllents",
"a",
"",
"ecnaP",
"ofrespirators?",
"vn",
"thinotodowith",
"to",
"ing",
"what",
"he",
"guestion orre",
"asnir",
"anh",
"ised",
"du",
"The",
"oirator",
"ortages.Theansweronlytalk",
"",
"rlages",
"answer also",
".",
"babydelive",
"provideslinksthat",
"havemore",
"ormationabout",
"respirator shortages.",
"Figure l: Users interact with the deployed QA model and give feedback. Feedback contains a rating (",
"bad, good, could be improved, excellent ) and a natural language explanation.",
"辨别正确和错误答案的解释吗？",
"ment，Canadian government，3和 THMAX",
"为了解决这些问题，我们使用反馈数据来训",
"的官方网站上抓取了Covid-19相关内容。我",
"练重新排序原始答案的模型，并为答案提供解",
"们在常见问题部分提取问题和答案段落。为",
"释。我们的实验表明，这种方法不仅提高了收",
"了扩大数据集，我们还清理了抓取的页面并",
"集反馈的基本QA模型的准确性，而且还提高",
"提取了其他段落，我们使用众包为这些段落策",
"了未收集反馈数据的其他强模型的准确性。此",
"划了相应的问题，就像用户在问问题一样。我",
"外，我们进行人工评估以验证解释的有用性，",
"们在Appendix A中介绍了此注释过程的详细",
"并发现生成的自然语言解释有助于用户在接受",
"信息。我们使用此数据集分别为每个源训练基",
"或拒绝候选答案时做出明智和准确的决定",
"本RQA模型并部署它们。对于基本模型，我",
"我们的贡献如下：",
"们使用基于BERT的密集检索器（Karpukhin",
"1. We create the first retrieval-based QA",
"etal.,2020)结合Poly-encoder（Milleretal.",
"2017），（更多详细信息在Section3.1,忠）",
"因",
"此，我们事先莜集写第1阶段类似的单个段落",
"dataset containing interactive feedback.",
"2. We demonstrate a simple method of using",
"的问题，并将其用作交互式问题。向用户显示",
"the feedback data to increase the accuracy",
"模型的问题和前2个预测，并为每个问答对提",
"and explainability of RQA systems.",
"供反馈。收集的反馈包括从优秀，良好，可以",
"3. We show that the feedback data not only",
"改进，不良和自然语言解释中选择的评分，详",
"improve the deployed model but also a",
"细说明答案的优点和/或缺点。对于每个QA",
"stronger non-deployed model.",
"对，我们从三个不同的工人那里得到反馈。我",
"FeedbackQA Dataset",
"们采用了其他策略来确保反馈数据的质量，其",
"详细信息可在AppendixB中获得。生成的数",
"2",
"最近，人们一直在努力以自然语言理解任务",
"据集统计信息显示在Table2中。为了测试交",
"的解释形式收集反馈数据E（Camburuetal.",
"互式反馈是否也有助于分发外设置，我们没有",
"2018;Rajanietal.2019等）。这些仅包含对从",
"收集其中一个域的反馈(加拿大)",
"训练数据中采样的给定输入的地面实况预测的",
"2.2FeedbackQA analysis",
"解释，而无需任何用户系统交互。相反，我们在",
"Table3显示了反馈数据的示例，包括评级和",
"部署RQA系统后收集用户反馈，从而收集正",
"解释。我们发现，解释通常包含表示答案质量",
"确和错误预测的反馈。Table1对FeedbackQA",
"的评论式文本，或总结哪些部分是正确的以及",
"和现有的自然语言理解（NLU）数据集进行了",
"为什么的陈述。因此，我们使用以下模式分析",
"全面的比较，并提供了解释数据。",
"示例解释：",
"2.1",
"Dataset collection",
"评论几个解释以通用评论开头，例如这直接回",
"为了像在实际环境中一样收集部署后反馈，我",
"答了问题或与问题无关。有时，用户还会突出",
"们将数据收集分为两个阶段：（RQA模型的）",
"显示答案中不错或可以改进的方面。例如，。可",
"部署前和部署后。",
"以提高语法。建议在写作方面可以改进答案",
"有用内容摘要是指答案中实际回答问题的部",
"Stage l:Pre-deployment of a QA system 我",
"们从 WHO ，US Government ，UK Govern-",
"3We focus on the Province of Quebec"
] | {
"arxiv_id": "2204.03025",
"language": "zh",
"page_idx": 3
} |
||
2210.12374-zh-1 | 2210.12374-zh-1.jpg | 推理技能缺乏可解释性
本文提出了一种新的表预训练方法，名为$\mathbf{R}$EASTAP,该方法使模型能够在预训练期间有效地学习表结构理解和表推理技巧。我们首先定义了7 表推理技巧，例如数值运算和时间比较。如图$\boxed{\mathrm{E}}1$ 所示，对于每种推理技能，都应用了相应的示例生成器来通过表格合成问答(QA)示例。我们将预训练任务建模为序列生成任务，并对序列到序列(seq2seq)LM进行预训练，以生成合成问题的答案。REASTAP 理论上适用于任何 seq2seqLM, 无需特定于表的架构设计。我们的主要见解是，如果可以对语言模型进行预训练以生成合成问题的答案，则需要各种表推理技能，它应该具有出色的表结构理解和表推理能力，从而为下游任务带来好处。我们工作的主要贡献可以概括如下：
$\cdot$ We develop a new table reasoning example
generation pipeline, which produces a large-
scale table QA corpus that requires various
reasoning skills over semi-structured tables.
$\cdot$ We propose a new table pre-training method,
REASTAP, which helps the model to learn
table structure understanding and various ta-
ble reasoning skills during pre-training with-
out any table-specific architecture design.
$\bullet$ REASTAP is evaluated on four downstream
benchmarks. Experimental results demonstrate that REASTAP achieves new state-ofthe-art results on all of them, and delivers a great improvement on low-resource setting.
# 2 Pre-training Corpus
2.1 Table Source and Pre-processing
我们选择公开可用的半结构化表作为表源。具体来说，我们从英文维基百科$^{1}$中提取了表格， 涵盖了广泛的领域，包括流行文化，地理，政治和科学。我们保留了 8-30 行和至少三列的表，导致大约 600K 个表。对于每个提取的表都应用了预处理脚本，以使用其数据类型自动注释表列 (i。e、字符串、数字和日期),这使我们能够生成涉及操纵数字和日期的问题。此外，最近的工作E|(Yang et al.,2022; Wang et al., 2022)表明，现有的表预训练方法可能会将表
$^{\mathrm{l}}$We parsed the 02-20-2022 Wikipedia dump using WikiExtractor Tools from https://github.com/ attardi/wikiextractor
Liu et al., 2022b,a)。虽然这些方法在理解表结 行顺序编码为不需要的偏差。例如，知道行顺构方面是有效的，但它们增加了建模的复杂 序信息的预训练模型在回答最高级类型的问题性，并且对于模型为什么在预训练期间学习表 时倾向于选择表的第一行或最后一行，而没有
真正理解表内容。为了缓解这个问题，我们在
预处理过程中随机洗牌表行
## 2.2 Example Generation
我们定义了7 类型的表推理技巧，示例和说明如表$\mathbb{E}$1 所示。示例生成管道改编自 Yoran et al.(2021)。每个推理技能与一个示例生成器和多个问题模板相关联。示例生成器被实现为一个函数，该函数接受一个表$T$,并根据模板生成几个推理示例$(T,\boldsymbol{q},\boldsymbol{a})$,其中$q$表示问题，$a$表示答案
每个模板都包含用提取的表中的内容实例化的类型化变量。具体来说，对列 col 和单元格值 val 进行索引，以指定 val:i 必须由第 i 列中的单元格值实例化。某些模板还规定所选列和单元格值必须是日期或数字类型。OPERATOR 和 ORDINAL 对应于根据特定推理技巧实例化的运算符和序数。CONDITION:i 可以是 1)第 i 列的单元格值；或 2) 如果第i 列是日期或数字类型，则为数字$l$时间比较语句。例如，来自图 1 的问题“哪个公司名称， 总部是美国，有第四个利润？”是从一个“数字比较”模板生成的：“哪个 col:1,col:2是 CONDITION:2 , ORDINAL 是 col:3 ?"
一旦实例化了采样模板中的所有变量，我们就获得了问题$q$。然后，示例生成器将以编程方式返回相应的答案$a$
# 2.3 Example Sampling
在为每种推理技能生成大量 QA 示例之后，我们必须从这些综合示例中采样预训练数据。在我们的设置中，与每个推理技能相对应的预训练示例部分(表$\overline{\mathrm{E}}$1)与 TabFact$\underline{\mathrm{E}}($Chen et al., 2020b) 中定义的逻辑运算部分大致匹配。我们提高了数值运算技能的部分，因为数值推理对模型的学习更具挑战性。为了增加预训练语料库的多样性，对于每种推理技能，我们还从TAPEX$\boxed{\mathrm{F}}($Liu et al., 2022b) 预训练语料库中采样了{SQL 查询，执行结果} 对作为补充 QA 示例。抽样对根据其功能进行分类(例如。g。COUNT、SUM)。结果，我们总共获得了4M 对推理示例作为 REASTAP 的预训练语料库
### 3 Pre-training REASTAP
Task Formulation 综合预训练语料库中的每个示例都包含一个问题$q$和一个半结构化表$T$作为模型输入。任务目标是在给定问题$q$ 和输入表$T$ 的情况下生成准确的答案字符串 | [
[
"0.1184",
"0.0885",
"0.486",
"0.0885",
"0.486",
"0.0986",
"0.1184",
"0.0986"
],
[
"0.513",
"0.0885",
"0.879",
"0.0885",
"0.879",
"0.0986",
"0.513",
"0.0986"
],
[
"0.11755",
"0.10156",
"0.4878",
"0.1027",
"0.4878",
"0.1164",
"0.11755",
"0.1152"
],
[
"0.513",
"0.1027",
"0.881",
"0.1027",
"0.881",
"0.1164",
"0.513",
"0.1164"
],
[
"0.1184",
"0.11816",
"0.4878",
"0.11816",
"0.4878",
"0.1318",
"0.1184",
"0.1318"
],
[
"0.513",
"0.1194",
"0.881",
"0.1194",
"0.881",
"0.133",
"0.513",
"0.133"
],
[
"0.11755",
"0.1342",
"0.3015",
"0.1342",
"0.3015",
"0.1478",
"0.11755",
"0.1478"
],
[
"0.5146",
"0.1354",
"0.881",
"0.1354",
"0.881",
"0.1484",
"0.5146",
"0.1484"
],
[
"0.1368",
"0.152",
"0.486",
"0.152",
"0.486",
"0.165",
"0.1368",
"0.165"
],
[
"0.511",
"0.1509",
"0.7314",
"0.1509",
"0.7314",
"0.1644",
"0.511",
"0.1644"
],
[
"0.11755",
"0.1663",
"0.4836",
"0.1675",
"0.4836",
"0.1818",
"0.11755",
"0.1812"
],
[
"0.1184",
"0.1841",
"0.4844",
"0.1841",
"0.4844",
"0.1978",
"0.1184",
"0.1978"
],
[
"0.513",
"0.1799",
"0.721",
"0.1799",
"0.721",
"0.1936",
"0.513",
"0.1936"
],
[
"0.11755",
"0.1996",
"0.4836",
"0.1996",
"0.4836",
"0.2131",
"0.11755",
"0.2131"
],
[
"0.511",
"0.2019",
"0.879",
"0.2019",
"0.879",
"0.2156",
"0.511",
"0.2156"
],
[
"0.1184",
"0.2144",
"0.4878",
"0.2156",
"0.4878",
"0.2299",
"0.1184",
"0.2286"
],
[
"0.511",
"0.2173",
"0.8774",
"0.2173",
"0.8774",
"0.231",
"0.511",
"0.231"
],
[
"0.1184",
"0.2322",
"0.4878",
"0.2322",
"0.4878",
"0.2452",
"0.1184",
"0.2452"
],
[
"0.513",
"0.2351",
"0.8774",
"0.2351",
"0.8774",
"0.2452",
"0.513",
"0.2452"
],
[
"0.1184",
"0.2477",
"0.4878",
"0.2477",
"0.4878",
"0.2612",
"0.1184",
"0.2612"
],
[
"0.513",
"0.2494",
"0.881",
"0.2494",
"0.881",
"0.2632",
"0.513",
"0.2632"
],
[
"0.12006",
"0.2642",
"0.4844",
"0.2642",
"0.4844",
"0.2778",
"0.12006",
"0.2778"
],
[
"0.517",
"0.2695",
"0.529",
"0.2695",
"0.529",
"0.2737",
"0.517",
"0.2737"
],
[
"0.525",
"0.2666",
"0.881",
"0.2666",
"0.881",
"0.2798",
"0.525",
"0.2798"
],
[
"0.1184",
"0.2798",
"0.4878",
"0.2798",
"0.4878",
"0.2935",
"0.1184",
"0.2935"
],
[
"0.511",
"0.2808",
"0.881",
"0.282",
"0.881",
"0.2964",
"0.511",
"0.2957"
],
[
"0.1184",
"0.2964",
"0.486",
"0.2964",
"0.486",
"0.31",
"0.1184",
"0.31"
],
[
"0.511",
"0.2986",
"0.659",
"0.2986",
"0.659",
"0.3123",
"0.511",
"0.3123"
],
[
"0.1184",
"0.3123",
"0.482",
"0.3123",
"0.482",
"0.3225",
"0.1184",
"0.3225"
],
[
"0.531",
"0.3154",
"0.881",
"0.3154",
"0.881",
"0.3289",
"0.531",
"0.3289"
],
[
"0.12177",
"0.33",
"0.486",
"0.33",
"0.486",
"0.3403",
"0.12177",
"0.3403"
],
[
"0.5146",
"0.3308",
"0.881",
"0.3308",
"0.881",
"0.3445",
"0.5146",
"0.3445"
],
[
"0.1184",
"0.3445",
"0.4878",
"0.3445",
"0.4878",
"0.3582",
"0.1184",
"0.3582"
],
[
"0.513",
"0.3467",
"0.881",
"0.3467",
"0.881",
"0.3599",
"0.513",
"0.3599"
],
[
"0.12006",
"0.3623",
"0.486",
"0.3623",
"0.486",
"0.3723",
"0.12006",
"0.3723"
],
[
"0.513",
"0.3623",
"0.882",
"0.3623",
"0.882",
"0.3772",
"0.513",
"0.3772"
],
[
"0.1184",
"0.3772",
"0.4382",
"0.3772",
"0.4382",
"0.3901",
"0.1184",
"0.3901"
],
[
"0.513",
"0.379",
"0.882",
"0.379",
"0.882",
"0.3926",
"0.513",
"0.3926"
],
[
"0.5146",
"0.3955",
"0.8774",
"0.3955",
"0.8774",
"0.4062",
"0.5146",
"0.4062"
],
[
"0.1377",
"0.405",
"0.4878",
"0.405",
"0.4878",
"0.4187",
"0.1377",
"0.4187"
],
[
"0.513",
"0.411",
"0.879",
"0.411",
"0.879",
"0.4246",
"0.513",
"0.4246"
],
[
"0.1528",
"0.4216",
"0.4895",
"0.4204",
"0.4895",
"0.4353",
"0.1528",
"0.4358"
],
[
"0.5146",
"0.4287",
"0.8774",
"0.4287",
"0.8774",
"0.4395",
"0.5146",
"0.4395"
],
[
"0.1528",
"0.4353",
"0.486",
"0.4358",
"0.486",
"0.4507",
"0.1528",
"0.4495"
],
[
"0.513",
"0.4436",
"0.882",
"0.4436",
"0.882",
"0.4573",
"0.513",
"0.4573"
],
[
"0.1545",
"0.4531",
"0.4768",
"0.4531",
"0.4768",
"0.466",
"0.1545",
"0.466"
],
[
"0.5146",
"0.4602",
"0.881",
"0.4602",
"0.881",
"0.4739",
"0.5146",
"0.4739"
],
[
"0.1377",
"0.4797",
"0.486",
"0.4797",
"0.486",
"0.4946",
"0.1377",
"0.4946"
],
[
"0.513",
"0.4756",
"0.882",
"0.4756",
"0.882",
"0.4893",
"0.513",
"0.4893"
],
[
"0.1545",
"0.4963",
"0.4878",
"0.4963",
"0.4878",
"0.5103",
"0.1545",
"0.5103"
],
[
"0.511",
"0.4922",
"0.8774",
"0.4922",
"0.8774",
"0.506",
"0.511",
"0.506"
],
[
"0.1545",
"0.513",
"0.4895",
"0.513",
"0.4895",
"0.527",
"0.1545",
"0.527"
],
[
"0.513",
"0.509",
"0.86",
"0.5083",
"0.86",
"0.5195",
"0.513",
"0.521"
],
[
"0.1511",
"0.528",
"0.4878",
"0.5293",
"0.4878",
"0.5435",
"0.1511",
"0.542"
],
[
"0.5425",
"0.525",
"0.881",
"0.525",
"0.881",
"0.538",
"0.5425",
"0.538"
],
[
"0.1545",
"0.546",
"0.4592",
"0.546",
"0.4592",
"0.5586",
"0.1545",
"0.5586"
],
[
"0.5146",
"0.5415",
"0.881",
"0.5415",
"0.881",
"0.5547",
"0.5146",
"0.5547"
],
[
"0.511",
"0.5547",
"0.6963",
"0.5557",
"0.6963",
"0.571",
"0.511",
"0.5693"
],
[
"0.1368",
"0.571",
"0.486",
"0.572",
"0.486",
"0.586",
"0.1368",
"0.585"
],
[
"0.1528",
"0.5884",
"0.4895",
"0.5884",
"0.4895",
"0.6016",
"0.1528",
"0.6016"
],
[
"0.509",
"0.5845",
"0.708",
"0.5845",
"0.708",
"0.597",
"0.509",
"0.597"
],
[
"0.1545",
"0.605",
"0.4895",
"0.605",
"0.4895",
"0.618",
"0.1545",
"0.618"
],
[
"0.511",
"0.6055",
"0.881",
"0.607",
"0.881",
"0.6216",
"0.511",
"0.6206"
],
[
"0.1545",
"0.6206",
"0.4895",
"0.6206",
"0.4895",
"0.6343",
"0.1545",
"0.6343"
],
[
"0.511",
"0.6216",
"0.882",
"0.6226",
"0.882",
"0.637",
"0.511",
"0.6357"
],
[
"0.1528",
"0.637",
"0.476",
"0.637",
"0.476",
"0.652",
"0.1528",
"0.652"
],
[
"0.513",
"0.64",
"0.8774",
"0.64",
"0.8774",
"0.651",
"0.513",
"0.651"
],
[
"0.513",
"0.6562",
"0.881",
"0.6562",
"0.881",
"0.6694",
"0.513",
"0.6694"
],
[
"0.1125",
"0.6646",
"0.3284",
"0.665",
"0.3284",
"0.6797",
"0.1125",
"0.6787"
],
[
"0.513",
"0.6714",
"0.881",
"0.6714",
"0.881",
"0.685",
"0.513",
"0.685"
],
[
"0.1142",
"0.6895",
"0.4207",
"0.6914",
"0.4207",
"0.706",
"0.1142",
"0.7036"
],
[
"0.513",
"0.6885",
"0.8774",
"0.6885",
"0.8774",
"0.6987",
"0.513",
"0.6987"
],
[
"0.513",
"0.7036",
"0.881",
"0.7036",
"0.881",
"0.7173",
"0.513",
"0.7173"
],
[
"0.1184",
"0.712",
"0.486",
"0.712",
"0.486",
"0.7256",
"0.1184",
"0.7256"
],
[
"0.11755",
"0.727",
"0.482",
"0.727",
"0.482",
"0.7407",
"0.11755",
"0.7407"
],
[
"0.5146",
"0.7217",
"0.876",
"0.7217",
"0.876",
"0.7324",
"0.5146",
"0.7324"
],
[
"0.511",
"0.7363",
"0.879",
"0.7363",
"0.879",
"0.75",
"0.511",
"0.75"
],
[
"0.1184",
"0.7446",
"0.486",
"0.7446",
"0.486",
"0.7583",
"0.1184",
"0.7583"
],
[
"0.5146",
"0.7544",
"0.879",
"0.7544",
"0.879",
"0.764",
"0.5146",
"0.764"
],
[
"0.1184",
"0.7603",
"0.4844",
"0.7603",
"0.4844",
"0.774",
"0.1184",
"0.774"
],
[
"0.511",
"0.7666",
"0.873",
"0.7686",
"0.873",
"0.783",
"0.511",
"0.781"
],
[
"0.1184",
"0.777",
"0.4844",
"0.777",
"0.4844",
"0.7905",
"0.1184",
"0.7905"
],
[
"0.513",
"0.7837",
"0.881",
"0.7837",
"0.881",
"0.7974",
"0.513",
"0.7974"
],
[
"0.1184",
"0.7935",
"0.4844",
"0.7935",
"0.4844",
"0.807",
"0.1184",
"0.807"
],
[
"0.513",
"0.8003",
"0.8403",
"0.8003",
"0.8403",
"0.814",
"0.513",
"0.814"
],
[
"0.1184",
"0.8086",
"0.486",
"0.8086",
"0.486",
"0.8223",
"0.1184",
"0.8223"
],
[
"0.1184",
"0.826",
"0.4895",
"0.826",
"0.4895",
"0.839",
"0.1184",
"0.839"
],
[
"0.5073",
"0.83",
"0.75",
"0.83",
"0.75",
"0.844",
"0.5073",
"0.844"
],
[
"0.12006",
"0.8413",
"0.486",
"0.8413",
"0.486",
"0.855",
"0.12006",
"0.855"
],
[
"0.11755",
"0.8555",
"0.486",
"0.857",
"0.486",
"0.8706",
"0.11755",
"0.869"
],
[
"0.5146",
"0.857",
"0.879",
"0.857",
"0.879",
"0.8706",
"0.5146",
"0.8706"
],
[
"0.513",
"0.8735",
"0.881",
"0.8735",
"0.881",
"0.887",
"0.513",
"0.887"
],
[
"0.1377",
"0.88",
"0.4785",
"0.882",
"0.4785",
"0.8965",
"0.1377",
"0.8945"
],
[
"0.51",
"0.888",
"0.881",
"0.89",
"0.881",
"0.904",
"0.51",
"0.902"
],
[
"0.11755",
"0.8945",
"0.4768",
"0.8945",
"0.4768",
"0.9077",
"0.11755",
"0.9077"
],
[
"0.1184",
"0.9062",
"0.3047",
"0.9062",
"0.3047",
"0.9194",
"0.1184",
"0.9194"
],
[
"0.513",
"0.9067",
"0.876",
"0.9067",
"0.876",
"0.9175",
"0.513",
"0.9175"
]
] | [
"Liuetal..2022b.a）。虽然这些方法在理解表结",
"行顺序编码为不需要的偏差。例如，知道行顺",
"构方面是有效的，但它们增加了建模的复杂",
"序信息的预训练模型在回答最高级类型的问题",
"性，并且对于模型为什么在预训练期间学习表",
"时倾向于选择表的第一行或最后一行，而没有",
"推理技能缺乏可解释性",
"真正理解表内容。为了缓解这个问题，我们在",
"本文提出了一种新的表预训练方法，名为",
"预处理过程中随机洗牌表行",
"REASTAP，该方法使模型能够在预训练期间",
"有效地学习表结构理解和表推理技巧。我们首",
"2.2Example Generation",
"先定义了7表推理技巧，例如数值运算和时间",
"我们定义了7类型的表推理技巧，示例和说",
"比较。如图国1所示，对于每种推理技能，都",
"明如表E1所示。示例生成管道改编自Yorar",
"应用了相应的示例生成器来通过表格合成问答",
"etal.（2021）。每个推理技能与一个示例生成器",
"QA）示例。我们将预训练任务建模为序列生",
"和多个问题模板相关联。示例生成器被实现为",
"成任务，并对序列到序列（seq2seq）LM进行",
"",
"一个函数，该函数接受一个表T，并根据模板",
"预训练，以生成合成问题的答案。REASTAP理",
"生成几个推理示例（T，q，α），其中q表示",
"论上适用于任何seq2seqLM，无需特定于表的",
"问题，a表示答案",
"架构设计。我们的主要见解是：如果可以对语",
"每个模板都包含用提取的表中的内容实例",
"言模型进行预训练以生成合成问题的答案，则",
"化的类型化变量。具体来说，对列col和单",
"需要各种表推理技能，它应该具有出色的表结",
"元格值val进行索引，以指定val：i必须由",
"构理解和表推理能力，从而为下游任务带来好",
"第i列中的单元格值实例化。某些模板还规",
"处。我们工作的主要贡献可以概括如下：",
"定所选列和单元格值必须是日期或数字类型。",
"OPERATOR和ORDINAL对应于根据特定推理",
". We develop a new table reasoning example",
"技巧实例化的运算符和序数。CONDITION：i",
"generation pipeline, which produces a large-",
"可以是1）第i列的单元格值；或2）如果第",
"scale table QA corpus that requires various",
"i列是日期或数字类型，则为数字/时间比较语",
"reasoning skills over semi-structured tables.",
"句。例如，来自图1的问题“哪个公司名称，",
". We propose a new table pre-training method,",
"总部是美国，有第四个利润？”是从一个“数",
"REASTAP, which helps the model to learn",
"字比较”模板生成的：“哪个col：1，col：2",
"table structure understanding and various ta-",
"是CONDITION：2，ORDINAL是CO1：3？",
"ble reasoning skills during pre-training with-",
"一且实例化了采样模板中的所有变量，我们",
"out any table-specific architecture design.",
"就获得了问题q。然后，示例生成器将以编程",
"方式返回相应的答案a",
": REAsTAP is evaluated on four downstream",
"benchmarks. Experimental results demon-",
"2.3ExampleSampling",
"strate that REAsTAP achieves new state-of-",
"在为每种推理技能生成大量QA示例之后，我",
"the-art results on all of them, and delivers a",
"们必须从这些综合示例中采样预训练数据。在",
"great improvement on low-resource setting.",
"我们的设置中，与每个推理技能相对应的预训",
"练示例部分（表E1）与TabFactE(Chenetal.",
"2Pre-training Corpus",
"2020b）中定义的逻辑运算部分大致匹配。我们",
"2.1Table Source and Pre-processing",
"提高了数值运算技能的部分，因为数值推理",
"对模型的学习更具挑战性。为了增加预训练语",
"我们选择公开可用的半结构化表作为表源。具",
"体来说，我们从英文维基百科中提取了表格",
"料库的多样性，对于每种推理技能，我们还从",
"TAPEXE(Liuetal.，2022b)预训练语料库中采",
"涵盖了广泛的领域，包括流行文化，地理，政",
"样了ISQL查询，执行结果！对作为补充QA",
"治和科学。我们保留了8-30行和至少三列的",
"示例。抽样对根据其功能进行分类（例如。g。",
"表，导致大约600K个表。对于每个提取的表，",
"COUNT、SUM）。结果，我们总共获得了4M对",
"都应用了预处理脚本，以使用其数据类型自动",
"推理示例作为REASTAP的预训练语料库",
"注释表列（i。e、字符串、数字和日期)，这使",
"我们能够生成涉及操纵数字和日期的问题。此",
"3Pre-trainingREASTAP",
"外，最近的工作El(Yangetal.,2022；Wangetal.，",
"2022）表明，现有的表预训练方法可能会将表",
"TaskFormulation综合预训练语料库中的每",
"个示例都包含一个问题9和一个半结构化表",
"'we parsed the 02-20-2022Wikipedia dump u",
"T作为模型输入。任务目标是在给定问题q",
"ing WikiExtractor Tools from https://github.com",
"attardi/wikiextractor",
"和输入表T的情况下生成准确的答案字符串"
] | {
"arxiv_id": "2210.12374",
"language": "zh",
"page_idx": 1
} |
||
2210.12374-zh-2 | 2210.12374-zh-2.jpg | <table>
<tbody>
<tr>
<th>Reasoning</th>
<th>Example Templates</th>
<th>Example Questions& Answers</th>
<th>$\%$ Data</th>
</tr>
<tr>
<td>Conjunction</td>
<td>What was the col:1 when the col:2 was CONDITION:2 and the col:3 was CONDITION:3?</td>
<td>Q: What was the Television Service when the Country was Italy and the Content was Sport ? A: Sky OMC Sports, ESPN, Gazzetta TV, ...</td>
<td>$21.6\%$</td>
</tr>
<tr>
<td>Quantifiers Only/ Every</td>
<td>Does OPERATOR col: 1, with col: 2 was CONDTION:2, have col:3 CONDITION: 3? </td>
<td>Q: Does every Company , with Headquarter was Paris ,have Industry Financials ? $A:Yes$ Q: Does only Company Name , with Founded Year was later than 1964 , have Employee Num- ber greater than 30,000 ? $A:No$</td>
<td>10.3 $\%$</td>
</tr>
<tr>
<td>${\mathrm{Temporal}}$ Comparison</td>
<td>Which $col:1$, with $col:2$ was CONDITION: 2, happened the ORDINAL according to col:3?</td>
<td>Q: Which Romaji ,with Sales was greater than 203,471,happened the 4th according to Date ? A: Hepburn</td>
<td>$14.5\%$</td>
</tr>
<tr>
<td>Date Difference</td>
<td>how much time had passed between when the col:1 was val:1 and when the col:2 was val:2?</td>
<td>Q: how much time had passed between when the Candidate was John Kufuor and when the Candi- date was Paul McCartney ? $\mathbf{A}\colon16$ years</td>
<td>$5.7\%$</td>
</tr>
<tr>
<td>Counting</td>
<td>How many $col:1$ have col:2 CONDITION: 2?</td>
<td>Q: How many Event Location have Attendance greater than 10,235 ? A: 7</td>
<td>18.0 $\%$</td>
</tr>
<tr>
<td>Numerical Operation</td>
<td>What was the OPERATOR of col:1 when the col:2 was CONDITION:2?</td>
<td>Q: What was the sum of GDP Estimate( $US Million) when the GDP Estimate( $ US Million) was greater than 841,969 ? ${\mathrm{A:1,574,013}}$</td>
<td>$15.9\%$</td>
</tr>
<tr>
<td>Numerical Comparison</td>
<td>Which col:1,with col:2 was CONDITION:2, has the ORDINAL ${\mathrm{col:3?}}$</td>
<td>Q: Which Franchise , with Owner(s) was Nin- tendo , has the 5th Total revenue( $ US Billion) ? $\mathbf{A}\cdot\mathbf{P}_{0}$kem $10n$</td>
<td>14.0 $\%$</td>
</tr>
</tbody>
</table>
Table 1: 7 reasoning skills with example for pre-training REASTAP. Variable names indicate permissible instanti-
ations. col denotes a column name, val denotes a cell value, and indices denote that a cell value must originate from the specified column. OPERATOR and ORDINAL correspond to operators and ordinal numeral that are instantiated according to the specific reasoning skill, e.g., for `Temporal Comparison’, ORD INAL is replaced with a reasonable ordinal numeral such as "4th". And CONDITTON: i can be 1) a cell value from the i -th column, or 2) number/temporal comparison statement (e.g."later than 1967") if the i-th column is of number or date type.
$$\boldsymbol{a}=(a_1,a_2,\ldots,a_n)\::$$
便可以将其直接输人到核心解码器模型中。具体来说，通过插人几个特殊标记来指示表边界，展平的表被表示为
(1)
$$\boldsymbol{a}=\operatorname{argmax}\prod\limits_{i=1}^nP(a_i|a_{<i},\boldsymbol{q},T;\theta),$$
其中$\theta$表示 seq2seqLM 的参数
$$\begin{array}{cc}T=_{[\text{HEAD}]}h_1\mid h_2\mid\cdots\mid h_m\_{[\text{ROW}]}&c_{1,1}\mid c_{1,2}\mid\cdots\mid c_{1,m}\end{array}$$
$$\tiny{[\text{ROW]}}\:c_{n,1}\mid c_{n,2}\mid\cdots\mid c_{n,m}$$
其中[HEAD]和[ROW]是分别指示表头和行的区域的特殊标记。我们在展平的表$T$ 前面加上问题，并将它们输入到模型编码器中。解码器的任务是生成答案，用逗号分隔，自动回归
Model Architecture 我们的方法理论上适用于任何 seq2seqLM,例如 T5$\overline{\mathrm{E}}($Raffel et al., 2020) 和 GPT3$\boxed{\mathrm{F}}($Brown et al., 2020) 。在我们的实验中，我们基于 BARTE(Lewis et al., 2020) 实现了 REASTAP , 这是一种广泛使用的基于 Transformer 的预训练模型$\boxed{\mathrm{E}}($Vaswani et al., 2017),已证明其在各种理解和文本生成任务上的有效性。在我们的实验中，我们选择了BART- Large 作为主干，它在编码器和解码器中都有大约 400M 的参数和 12 层
Data Serialization 如图$\overline{\mathrm{E}}$1 所示，输入包含一个问题及其相应的表。我们将表格展平，以
### 4 $\textbf{Downstream Tasks}$
我们在三种不同类型的下游任务上评估了REASTAP , 以验证其有效性。每个任务的统计数据和示例分别显示在附录的表 2 和表 10 | [
[
"0.1427",
"0.0885",
"0.2107",
"0.0885",
"0.2107",
"0.0998",
"0.1427",
"0.0998"
],
[
"0.2284",
"0.0873",
"0.3525",
"0.0885",
"0.3525",
"0.0998",
"0.2284",
"0.0986"
],
[
"0.4895",
"0.08905",
"0.681",
"0.08905",
"0.681",
"0.0998",
"0.4895",
"0.0998"
],
[
"0.8076",
"0.0885",
"0.86",
"0.0885",
"0.86",
"0.0998",
"0.8076",
"0.0998"
],
[
"0.2275",
"0.1051",
"0.471",
"0.1051",
"0.471",
"0.11816",
"0.2275",
"0.11816"
],
[
"0.486",
"0.1051",
"0.7925",
"0.1051",
"0.7925",
"0.11816",
"0.486",
"0.11816"
],
[
"0.8096",
"0.1057",
"0.858",
"0.1057",
"0.858",
"0.11755",
"0.8096",
"0.11755"
],
[
"0.2284",
"0.11755",
"0.4727",
"0.11755",
"0.4727",
"0.1312",
"0.2284",
"0.1312"
],
[
"0.1411",
"0.1057",
"0.2351",
"0.1057",
"0.2351",
"0.11755",
"0.1411",
"0.11755"
],
[
"0.486",
"0.11755",
"0.7715",
"0.11755",
"0.7715",
"0.1312",
"0.486",
"0.1312"
],
[
"0.2284",
"0.1289",
"0.3652",
"0.1277",
"0.3652",
"0.1414",
"0.2284",
"0.1426"
],
[
"0.4895",
"0.1312",
"0.754",
"0.1312",
"0.754",
"0.1414",
"0.4895",
"0.1414"
],
[
"0.2275",
"0.1478",
"0.4727",
"0.1478",
"0.4727",
"0.1609",
"0.2275",
"0.1609"
],
[
"0.4844",
"0.1478",
"0.7925",
"0.1478",
"0.7925",
"0.1609",
"0.4844",
"0.1609"
],
[
"0.806",
"0.1478",
"0.858",
"0.1455",
"0.86",
"0.1603",
"0.8076",
"0.1621"
],
[
"0.2275",
"0.158",
"0.372",
"0.1603",
"0.372",
"0.1733",
"0.2275",
"0.1716"
],
[
"0.1427",
"0.1484",
"0.2141",
"0.1484",
"0.2141",
"0.1603",
"0.1427",
"0.1603"
],
[
"0.1411",
"0.1592",
"0.215",
"0.1609",
"0.215",
"0.1729",
"0.1411",
"0.1704"
],
[
"0.3745",
"0.1609",
"0.4148",
"0.1609",
"0.4148",
"0.1716",
"0.3745",
"0.1716"
],
[
"0.4207",
"0.1609",
"0.471",
"0.1609",
"0.471",
"0.1716",
"0.4207",
"0.1716"
],
[
"0.486",
"0.1592",
"0.693",
"0.1592",
"0.693",
"0.1729",
"0.486",
"0.1729"
],
[
"0.2301",
"0.1729",
"0.339",
"0.1729",
"0.339",
"0.1841",
"0.2301",
"0.1841"
],
[
"0.4878",
"0.1729",
"0.532",
"0.1729",
"0.532",
"0.1829",
"0.4878",
"0.1829"
],
[
"0.486",
"0.1829",
"0.7925",
"0.1829",
"0.7925",
"0.1965",
"0.486",
"0.1965"
],
[
"0.486",
"0.1953",
"0.79",
"0.1953",
"0.79",
"0.206",
"0.486",
"0.206"
],
[
"0.486",
"0.2073",
"0.644",
"0.2073",
"0.644",
"0.2203",
"0.486",
"0.2203"
],
[
"0.4878",
"0.2197",
"0.5283",
"0.2197",
"0.5283",
"0.231",
"0.4878",
"0.231"
],
[
"0.3435",
"0.2369",
"0.4744",
"0.2382",
"0.4744",
"0.2494",
"0.3435",
"0.2488"
],
[
"0.4844",
"0.2369",
"0.791",
"0.2369",
"0.791",
"0.2505",
"0.4844",
"0.2505"
],
[
"0.2284",
"0.2382",
"0.3333",
"0.2382",
"0.3333",
"0.2488",
"0.2284",
"0.2488"
],
[
"0.1427",
"0.2382",
"0.2048",
"0.2382",
"0.2048",
"0.2494",
"0.1427",
"0.2494"
],
[
"0.8096",
"0.2382",
"0.858",
"0.2382",
"0.858",
"0.2494",
"0.8096",
"0.2494"
],
[
"0.1427",
"0.2505",
"0.2225",
"0.2505",
"0.2225",
"0.2625",
"0.1427",
"0.2625"
],
[
"0.2284",
"0.2505",
"0.3342",
"0.2505",
"0.3342",
"0.2612",
"0.2284",
"0.2612"
],
[
"0.3645",
"0.2505",
"0.4282",
"0.2505",
"0.4282",
"0.2625",
"0.3645",
"0.2625"
],
[
"0.4475",
"0.2517",
"0.471",
"0.2517",
"0.471",
"0.26",
"0.4475",
"0.26"
],
[
"0.4844",
"0.2477",
"0.7773",
"0.2488",
"0.7773",
"0.2632",
"0.4844",
"0.2625"
],
[
"0.2284",
"0.2612",
"0.4324",
"0.2612",
"0.4324",
"0.275",
"0.2284",
"0.275"
],
[
"0.4844",
"0.26",
"0.5635",
"0.2625",
"0.5635",
"0.2766",
"0.4844",
"0.275"
],
[
"0.2284",
"0.279",
"0.4744",
"0.279",
"0.4744",
"0.2922",
"0.2284",
"0.2922"
],
[
"0.4844",
"0.279",
"0.7925",
"0.279",
"0.7925",
"0.2922",
"0.4844",
"0.2922"
],
[
"0.817",
"0.2798",
"0.86",
"0.2798",
"0.86",
"0.2915",
"0.817",
"0.2915"
],
[
"0.486",
"0.2915",
"0.7925",
"0.2915",
"0.7925",
"0.3047",
"0.486",
"0.3047"
],
[
"0.1411",
"0.2798",
"0.1763",
"0.2798",
"0.1763",
"0.2915",
"0.1411",
"0.2915"
],
[
"0.1427",
"0.2922",
"0.2107",
"0.2922",
"0.2107",
"0.304",
"0.1427",
"0.304"
],
[
"0.2301",
"0.2935",
"0.4727",
"0.2935",
"0.4727",
"0.304",
"0.2301",
"0.304"
],
[
"0.2301",
"0.3047",
"0.3838",
"0.3047",
"0.3838",
"0.3154",
"0.2301",
"0.3154"
],
[
"0.4878",
"0.3047",
"0.657",
"0.3047",
"0.657",
"0.3154",
"0.4878",
"0.3154"
],
[
"0.4878",
"0.3164",
"0.559",
"0.3164",
"0.559",
"0.328",
"0.4878",
"0.328"
],
[
"0.1411",
"0.3308",
"0.2048",
"0.3333",
"0.2032",
"0.3474",
"0.1394",
"0.3457"
],
[
"0.2284",
"0.332",
"0.3105",
"0.3342",
"0.3105",
"0.3474",
"0.2284",
"0.3457"
],
[
"0.486",
"0.3333",
"0.7925",
"0.3333",
"0.7925",
"0.3467",
"0.486",
"0.3467"
],
[
"0.81",
"0.3342",
"0.858",
"0.3342",
"0.858",
"0.3457",
"0.81",
"0.3457"
],
[
"0.4878",
"0.3457",
"0.6187",
"0.3433",
"0.6187",
"0.357",
"0.4878",
"0.3594"
],
[
"0.3047",
"0.335",
"0.4744",
"0.335",
"0.4744",
"0.3457",
"0.3047",
"0.3457"
],
[
"0.2301",
"0.3467",
"0.34",
"0.3467",
"0.34",
"0.357",
"0.2301",
"0.357"
],
[
"0.4878",
"0.3594",
"0.519",
"0.3594",
"0.519",
"0.3706",
"0.4878",
"0.3706"
],
[
"0.2284",
"0.376",
"0.471",
"0.376",
"0.471",
"0.3896",
"0.2284",
"0.3896"
],
[
"0.4844",
"0.376",
"0.7925",
"0.376",
"0.7925",
"0.3896",
"0.4844",
"0.3896"
],
[
"0.8096",
"0.3772",
"0.858",
"0.3772",
"0.858",
"0.3884",
"0.8096",
"0.3884"
],
[
"0.2284",
"0.3884",
"0.476",
"0.3884",
"0.476",
"0.402",
"0.2284",
"0.402"
],
[
"0.1411",
"0.3772",
"0.2107",
"0.3772",
"0.2107",
"0.3884",
"0.1411",
"0.3884"
],
[
"0.1427",
"0.3896",
"0.2083",
"0.3896",
"0.2083",
"0.401",
"0.1427",
"0.401"
],
[
"0.486",
"0.3884",
"0.7925",
"0.3884",
"0.7925",
"0.402",
"0.486",
"0.402"
],
[
"0.4878",
"0.3997",
"0.6514",
"0.3984",
"0.6514",
"0.412",
"0.4878",
"0.4133"
],
[
"0.486",
"0.4133",
"0.5693",
"0.4133",
"0.5693",
"0.4233",
"0.486",
"0.4233"
],
[
"0.1377",
"0.4275",
"0.2124",
"0.43",
"0.2124",
"0.4448",
"0.1377",
"0.4424"
],
[
"0.3333",
"0.43",
"0.4768",
"0.43",
"0.4768",
"0.4436",
"0.3333",
"0.4436"
],
[
"0.4844",
"0.43",
"0.7944",
"0.43",
"0.7944",
"0.4436",
"0.4844",
"0.4436"
],
[
"0.8096",
"0.4312",
"0.858",
"0.4312",
"0.858",
"0.4424",
"0.8096",
"0.4424"
],
[
"0.2284",
"0.4424",
"0.4744",
"0.4424",
"0.4744",
"0.456",
"0.2284",
"0.456"
],
[
"0.2301",
"0.4312",
"0.3376",
"0.4312",
"0.3376",
"0.4412",
"0.2301",
"0.4412"
],
[
"0.1427",
"0.4436",
"0.2217",
"0.4436",
"0.2217",
"0.4548",
"0.1427",
"0.4548"
],
[
"0.4844",
"0.4424",
"0.7925",
"0.4424",
"0.7925",
"0.456",
"0.4844",
"0.456"
],
[
"0.2301",
"0.4548",
"0.2854",
"0.4548",
"0.2854",
"0.466",
"0.2301",
"0.466"
],
[
"0.4895",
"0.466",
"0.565",
"0.466",
"0.565",
"0.478",
"0.4895",
"0.478"
],
[
"0.11584",
"0.4922",
"0.882",
"0.4946",
"0.882",
"0.5083",
"0.11584",
"0.506"
],
[
"0.11755",
"0.5083",
"0.882",
"0.5083",
"0.882",
"0.5215",
"0.11755",
"0.5215"
],
[
"0.11755",
"0.5225",
"0.8833",
"0.5225",
"0.8833",
"0.536",
"0.11755",
"0.536"
],
[
"0.11584",
"0.5376",
"0.882",
"0.5376",
"0.882",
"0.5503",
"0.11584",
"0.5503"
],
[
"0.11584",
"0.5503",
"0.882",
"0.5503",
"0.882",
"0.564",
"0.11584",
"0.564"
],
[
"0.11584",
"0.5654",
"0.867",
"0.5654",
"0.867",
"0.579",
"0.11584",
"0.579"
],
[
"0.1142",
"0.6025",
"0.287",
"0.604",
"0.287",
"0.6216",
"0.1142",
"0.6206"
],
[
"0.5146",
"0.605",
"0.881",
"0.605",
"0.881",
"0.619",
"0.5146",
"0.619"
],
[
"0.513",
"0.6216",
"0.881",
"0.6216",
"0.881",
"0.6353",
"0.513",
"0.6353"
],
[
"0.513",
"0.637",
"0.6987",
"0.637",
"0.6987",
"0.651",
"0.513",
"0.651"
],
[
"0.1696",
"0.6567",
"0.2644",
"0.6567",
"0.2644",
"0.6685",
"0.1696",
"0.6685"
],
[
"0.294",
"0.653",
"0.4375",
"0.655",
"0.4375",
"0.673",
"0.294",
"0.6704"
],
[
"0.461",
"0.6567",
"0.486",
"0.6567",
"0.486",
"0.6694",
"0.461",
"0.6694"
],
[
"0.571",
"0.6646",
"0.79",
"0.665",
"0.79",
"0.6797",
"0.571",
"0.6787"
],
[
"0.6055",
"0.681",
"0.817",
"0.684",
"0.817",
"0.703",
"0.6055",
"0.6997"
],
[
"0.11584",
"0.6924",
"0.3594",
"0.6934",
"0.3594",
"0.708",
"0.11584",
"0.707"
],
[
"0.607",
"0.706",
"0.6265",
"0.706",
"0.6265",
"0.7153",
"0.607",
"0.7153"
],
[
"0.1234",
"0.7188",
"0.4844",
"0.7188",
"0.4844",
"0.7324",
"0.1234",
"0.7324"
],
[
"0.6035",
"0.72",
"0.8213",
"0.7236",
"0.8213",
"0.742",
"0.6035",
"0.7393"
],
[
"0.11755",
"0.734",
"0.4836",
"0.734",
"0.4836",
"0.7476",
"0.11755",
"0.7476"
],
[
"0.11584",
"0.75",
"0.486",
"0.7505",
"0.486",
"0.7656",
"0.11584",
"0.764"
],
[
"0.529",
"0.753",
"0.881",
"0.7544",
"0.881",
"0.7686",
"0.529",
"0.767"
],
[
"0.1184",
"0.767",
"0.4878",
"0.767",
"0.4878",
"0.781",
"0.1184",
"0.781"
],
[
"0.511",
"0.7695",
"0.879",
"0.7695",
"0.879",
"0.783",
"0.511",
"0.783"
],
[
"0.1184",
"0.783",
"0.486",
"0.783",
"0.486",
"0.7964",
"0.1184",
"0.7964"
],
[
"0.513",
"0.786",
"0.881",
"0.786",
"0.881",
"0.8",
"0.513",
"0.8"
],
[
"0.12006",
"0.8",
"0.4844",
"0.8",
"0.4844",
"0.8135",
"0.12006",
"0.8135"
],
[
"0.513",
"0.802",
"0.881",
"0.802",
"0.881",
"0.8164",
"0.513",
"0.8164"
],
[
"0.11584",
"0.8154",
"0.4836",
"0.8154",
"0.4836",
"0.829",
"0.11584",
"0.829"
],
[
"0.513",
"0.8184",
"0.534",
"0.8184",
"0.534",
"0.8306",
"0.513",
"0.8306"
],
[
"0.12006",
"0.832",
"0.4836",
"0.832",
"0.4836",
"0.8457",
"0.12006",
"0.8457"
],
[
"0.11755",
"0.847",
"0.486",
"0.847",
"0.486",
"0.861",
"0.11755",
"0.861"
],
[
"0.5156",
"0.8496",
"0.523",
"0.8496",
"0.523",
"0.8555",
"0.5156",
"0.8555"
],
[
"0.544",
"0.8467",
"0.712",
"0.8467",
"0.712",
"0.861",
"0.544",
"0.861"
],
[
"0.11755",
"0.8623",
"0.3838",
"0.8633",
"0.3838",
"0.8774",
"0.11755",
"0.877"
],
[
"0.513",
"0.8726",
"0.879",
"0.8726",
"0.879",
"0.8857",
"0.513",
"0.8857"
],
[
"0.12006",
"0.887",
"0.4844",
"0.888",
"0.4844",
"0.903",
"0.12006",
"0.902"
],
[
"0.511",
"0.887",
"0.881",
"0.888",
"0.881",
"0.903",
"0.511",
"0.902"
],
[
"0.1234",
"0.905",
"0.4878",
"0.905",
"0.4878",
"0.9185",
"0.1234",
"0.9185"
],
[
"0.511",
"0.905",
"0.881",
"0.905",
"0.881",
"0.9185",
"0.511",
"0.9185"
]
] | [
"Reasoning",
"ExampleTemplates",
"ExampleQuestions&Answers",
"% Data",
"What was the col : 1 when the col : 2",
"Q: What was the Television Service when the",
"21.6%",
"was CONDITION:2 and the col:3",
"Conjunction",
"Country was Italy and the Content was Sport ?",
"Was CONDITION:3?",
"A: Sky OMC Sports, ESPN, Gazzetta TV...",
"Does OPERATOR col:1, with col:2",
"Q: Does every Company, with Headquarter was",
"10.3%",
"Was CONDTION:2,",
"Quantifiers",
"Only/Every",
"have",
"col:3",
"Paris , have Industry Financials ?",
"CONDITION:3?",
"A:Yes",
"Q: Does only Company Name , with Founded",
"Yearwaslaterthan1964,haveEmployeeNum",
"ber greater than 30,000 ?",
"A: No",
"withcol:2was",
"Q: Which Romaji , with Sales was greater than",
"Which col:1.",
"Temporal",
"14.5 %",
"Comparison",
"CONDITION:2,",
"happened",
"the",
"203,471 , happened the 4th according to Date ?",
"ORDINAL according to col : 3?",
"A: Hepburn",
"how much time had passed between",
"Q: how much time had passed between when the",
"5.7 %",
"Candidate was John Kufuor and when the Candi-",
"Date",
"Difference",
"when the col : 1 was val: 1 and when",
"the col:2 was val:2?",
"date was Paul McCartney?",
"A: 16 years",
"Counting",
"Howmany",
"Q: How many Event Location have Attendance",
"18.0%",
"greater than 10,235 ?",
"col:1havecol:2",
"CONDITION:2?",
"A: 7",
"What was the OPERATOR of col :1",
"Q: What was the sum of GDP Estimate ( $ US",
"15.9 %",
"when the col :2 was CONDITION :2?",
"Numerical",
"Operation",
"Million) when the GDP Estimate ( S US Million)",
"was greater than 841,969 ?",
"A: 1,574,013",
"Numerical",
"withcol:2was",
"Q: Which Franchise , with Owner(s) was Nin-",
"14.0%",
"CONDITION:2,has the ORDINAL",
"Which col:1.",
"Comparison",
"tendo , has the 5th Total revenue( $ US Billion)",
"col:3?",
"A: Pokemon",
"Table 1: 7 reasoning skills with example for pre-training REAsTAP. Variable names indicate permissible instanti-",
"ations. col denotes a column name, vai denotes a cell value, and indices denote that a cell value must originate",
"from the specified column. OPERATOR and ORDINAL correspond to operators and ordinal numeral that are in-",
"stantiated according to the specific reasoning skill, e.g., for Temporal Comparison', ORDINAL is replaced with",
"a reasonable ordinal numeral such as \"4th\". And cONDITION: i can be 1) a cell value from the i -th column, or",
"2) number/temporal comparison statement (e.g. \"later than 1967\") if the i -th column is of number or date type.",
"a= (ai,a2,..., an) :",
"便可以将其直接输人到核心解码器模型中。具",
"体来说，通过插入几个特殊标记来指示表边",
"界，展平的表被表示为",
"a = argmax",
"P(aila<i,q, T; 0),",
"(1)",
"T=[HEAD]hi |h2 ... | hm",
"[ROW] Ci,1 I C1,2 I .. | C1,m",
"其中表示seq2seqLM的参数",
"...",
"ModelArchitecture我们的方法理论上适",
"[ROW] Cn,1 I Cn,2 ..| Cn,m",
"用于任何seq2seqLM，例如T5E(Raffel et al..",
"2020）和GPT3EBrownetal.，2020）。在我们的",
"其中【HEAD】和【ROW】是分别指示表头和",
"实验中，我们基于BARTE(Lewisetal.，2020)",
"行的区域的特殊标记。我们在展平的表T前面",
"实现了REASTAP，这是一种广泛使用的基",
"加上问题，并将它们输人到模型编码器中。解",
"于Transformer的预训练模型E(Vaswanietal.",
"码器的任务是生成答案，用逗号分隔，自动回",
"2017），已证明其在各种理解和文本生成任务",
"归",
"上的有效性。在我们的实验中，我们选择了",
"BART-Large作为主干，它在编码器和解码器",
"4",
"Downstream Tasks",
"中都有大约400M的参数和12层",
"我们在三种不同类型的下游任务上评估了",
"DataSerialization如图国1所示，输人包含",
"REASTAP，以验证其有效性。每个任务的统",
"一个问题及其相应的表。我们将表格展平，以",
"计数据和示例分别显示在附录的表2和表10"
] | {
"arxiv_id": "2210.12374",
"language": "zh",
"page_idx": 2
} |
||
2210.12374-zh-3 | 2210.12374-zh-3.jpg | <table>
<tbody>
<tr>
<th>Task</th>
<th>Dataset</th>
<th>$\#$Examples</th>
<th>$\#$Tables</th>
<th>Input</th>
<th>Output</th>
</tr>
<tr>
<td>Question Answering</td>
<td>WIKISQL- WEAK ( Zhong et al. , 2017) WIKITQ ( Pasupat and Liang, 2015) </td>
<td>80,654 22,033</td>
<td>24,241 2,108</td>
<td>Question Question</td>
<td>Answer Answer</td>
</tr>
<tr>
<td>Fact Verification</td>
<td>TABFACT$\left ( \mathrm{Chen~et~al. , ~2020b) }\right )$</td>
<td>118,275</td>
<td>16,573</td>
<td>Statement</td>
<td>Booleam</td>
</tr>
<tr>
<td>Faithful Text Generation</td>
<td>LogicNLG 1 (Chen et al.2020a)</td>
<td>37,015</td>
<td>7,392</td>
<td>Columns</td>
<td>Text</td>
</tr>
</tbody>
</table>
Table 2: Overview of downstream tasks used in this paper.
<table>
<tbody>
<tr>
<th>Model</th>
<th>Dev</th>
<th>Test</th>
</tr>
<tr>
<td>Previous Models</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>MAPO (Liang g et al., 2018)</td>
<td>71.8</td>
<td>72.4</td>
</tr>
<tr>
<td>MeRL ( Agarwal et al. , 2019) </td>
<td>74.9</td>
<td>74.8</td>
</tr>
<tr>
<td>LatentAlignment (Wang et al., 2019)</td>
<td>79.4</td>
<td>79.3</td>
</tr>
<tr>
<td>HardEM ( Min et al. , 2019) </td>
<td>84.4</td>
<td>83.9</td>
</tr>
<tr>
<td>Pre-trained LMs</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>TaPas (Herzig et al., 2020)</td>
<td>85.1</td>
<td>83.6</td>
</tr>
<tr>
<td>GraPPa (Yu et al., 2021a)</td>
<td>85.9</td>
<td>84.7</td>
</tr>
<tr>
<td>T5- 3B (Xie et al., 2022)</td>
<td>-</td>
<td>86.0</td>
</tr>
<tr>
<td>TAPEX (Liu et al., 2022b)</td>
<td>89.3</td>
<td>89.2</td>
</tr>
<tr>
<td>$\mathbf{BART}$</td>
<td>86.9</td>
<td>86.1</td>
</tr>
<tr>
<td>REASTAP</td>
<td>91.1</td>
<td>90.4</td>
</tr>
</tbody>
</table>
中。REASTAP 的微调类似于第 3 节中讨论的预训练过程。我们将两个下游任务建模为序列生成任务，并利用生成式 LM 以自回归方式生成输出
Table QA WIKISQL-WEAK E(Zhong et al., 2017) 和 WIKITQ $\boxed{\mathrm{F}}($Pasupat and Liang, 2015) 用于评估 REASTAP 在表 QA 任务上的性能。WIKISQL-WEAK (Zhong et al.,2017) 要求模型对表单元格值执行过滤和可选聚合，以回答给定的问题。WIKITQ (Pasupat and Liang, 2015) 需要更广泛的推理技能，因此更具挑战性。表QA 任务制定与 REASTAP 预训练任务相同。我们使用表示精度，它检查预测的答案是否等于基本事实，作为评估指标
Table Facts Verification 我们选择 TABFACT $\underline{\mathrm{E}}$(Chen et al., 2020b) 来评估 REASTAP 在表格事实上的性能验证任务。给定一个表和一个语句，TABFACT (Chen et al.,2020b)试图区分该语句是被该表包含还是被该表驳斥。TABFACT 将其测试集分为测试$_\mathrm{simple}$和测试$_\mathrm{complex}$子集，其中测试 $_\mathrm{complex}$包含需要更复杂的表推理技能的示例。此外，它选择了一个小的测试集测试$_\mathrm{small}$ 与 2K 样本为人类评估。为了在 TABFACT 上进行微调，在 BART$\overline{\mathrm{E}}($Lewis et al., 2020)之后， 我们对解码器中最后一个令牌的隐藏状态应用了二进制分类器，用于输出。Theobjectiveistogeneratetheverificationlabel $L\in\{0,1\}$ giventhestatement $s=(s_1,s_2,\cdots,s_n)$ andtheinputtable $T$
Table 3: Denotation accuracies on WIKISQL-WEAK. 输入表，目标是生成一个既流畅又事实正确的句子$y=(y_1,y_2,\ldots,y_n)$
$$\boldsymbol{y}=\operatorname{argmax}\prod\limits_{i=1}^nP(y_i|y_{<i},T\:;\:\theta)$$
(3)
为了评估生成的句子的逻辑保真度，Chen et al (2020a) 提出了两种基于模型的评估方法：基于解析的评估(SP-Acc)和基于NLI的评估(NLIAcc)。SP-Acc 直接从生成的句子中提取意义表示，并对表执行以验证正确性。NLI-ACC 使用自然语言推理(NLI)模型来预测蕴涵关系。在 Chen et al.(2020a)之后，我们使用 SP-Acc, NLI-Acc 作为逻辑保真度评估指标；和 BLEU- 1/2/3 作为表面级别的评估指标。值得注意的是，较高的 BLEU 分数与更好的逻辑保真度无关 (Nan et al.,2022a) 。
(2)
$$L=\mathop{\mathrm{argmax}}_{i\in\{0,1\}}P(i\mid\boldsymbol{s},T\:;\:\theta)$$
Weusedtheaccuracy(i。e。,正确预测的百分比)
作为评估指标
### 5 $\textbf{Experiments}$
5.1 Implementation Details
Faithful Table-to-Text Generation 我们选择 LoGICNLG (Chen et al.,2020a) 来评估REASTAP在忠实的表格到文本生成任务上的性能。与以前的 Table-to-Text 生成基准测试 (Wiseman et al., 2017; Balakrishnan et al., 2019; Parikh et al., 2020; Nan et al., 2022b) 相比，LoGICNLG 的任务是生成选定表区域在逻辑上需要的语句。给定其选定列为$T$ 的序列化
我们基于 fairseq 库$\overline{\mathbb{E}}$(Ott et al., 2019) 实现了我们的模型。我们采用了 BART-Large 作为骨干模型。对于表预训练，我们合成并采样了4M 对推理示例。在以下各节中，除非明确规定， 否则所有实验结果均在 4M 推理示例和 BARTLarge 配置的默认设置下进行评估。我们的预训练程序运行了 80,000 个步骤，批量大小为 | [
[
"0.199",
"0.0873",
"0.2317",
"0.0873",
"0.2317",
"0.0998",
"0.199",
"0.0998"
],
[
"0.403",
"0.0885",
"0.4534",
"0.0885",
"0.4534",
"0.0998",
"0.403",
"0.0998"
],
[
"0.565",
"0.0885",
"0.6406",
"0.0885",
"0.6406",
"0.0998",
"0.565",
"0.0998"
],
[
"0.656",
"0.0885",
"0.711",
"0.0885",
"0.711",
"0.0998",
"0.656",
"0.0998"
],
[
"0.7407",
"0.0885",
"0.777",
"0.0885",
"0.777",
"0.10095",
"0.7407",
"0.10095"
],
[
"0.8096",
"0.0861",
"0.8564",
"0.0885",
"0.8555",
"0.10095",
"0.8076",
"0.0986"
],
[
"0.5786",
"0.10394",
"0.6265",
"0.10394",
"0.6265",
"0.11816",
"0.5786",
"0.11816"
],
[
"0.659",
"0.10394",
"0.705",
"0.10394",
"0.705",
"0.11816",
"0.659",
"0.11816"
],
[
"0.727",
"0.10394",
"0.7886",
"0.10394",
"0.7886",
"0.11816",
"0.727",
"0.11816"
],
[
"0.806",
"0.10394",
"0.8613",
"0.10394",
"0.8613",
"0.11816",
"0.806",
"0.11816"
],
[
"0.1511",
"0.1134",
"0.2778",
"0.1134",
"0.2778",
"0.12354",
"0.1511",
"0.12354"
],
[
"0.309",
"0.1051",
"0.5474",
"0.1051",
"0.5474",
"0.11816",
"0.309",
"0.11816"
],
[
"0.5786",
"0.1164",
"0.6265",
"0.1164",
"0.6265",
"0.1312",
"0.5786",
"0.1312"
],
[
"0.727",
"0.1164",
"0.7886",
"0.1164",
"0.7886",
"0.1312",
"0.727",
"0.1312"
],
[
"0.806",
"0.1152",
"0.8613",
"0.11755",
"0.86",
"0.1318",
"0.804",
"0.13"
],
[
"0.663",
"0.11755",
"0.702",
"0.11755",
"0.702",
"0.13",
"0.663",
"0.13"
],
[
"0.318",
"0.11755",
"0.538",
"0.11755",
"0.538",
"0.1312",
"0.318",
"0.1312"
],
[
"0.3342",
"0.1342",
"0.5205",
"0.1342",
"0.5205",
"0.1478",
"0.3342",
"0.1478"
],
[
"0.575",
"0.1354",
"0.629",
"0.1354",
"0.629",
"0.1497",
"0.575",
"0.1497"
],
[
"0.66",
"0.1354",
"0.708",
"0.1354",
"0.708",
"0.1497",
"0.66",
"0.1497"
],
[
"0.7246",
"0.1354",
"0.7925",
"0.1354",
"0.7925",
"0.1497",
"0.7246",
"0.1497"
],
[
"0.8037",
"0.1354",
"0.8613",
"0.1354",
"0.8613",
"0.1497",
"0.8037",
"0.1497"
],
[
"0.8154",
"0.152",
"0.8555",
"0.155",
"0.852",
"0.1692",
"0.8135",
"0.1675"
],
[
"0.1628",
"0.1372",
"0.2644",
"0.1372",
"0.2644",
"0.1484",
"0.1628",
"0.1484"
],
[
"0.1377",
"0.1526",
"0.2913",
"0.155",
"0.2913",
"0.1687",
"0.1377",
"0.1663"
],
[
"0.5786",
"0.1538",
"0.6265",
"0.1538",
"0.6265",
"0.1687",
"0.5786",
"0.1687"
],
[
"0.665",
"0.155",
"0.702",
"0.155",
"0.702",
"0.1675",
"0.665",
"0.1675"
],
[
"0.727",
"0.1538",
"0.7886",
"0.1538",
"0.7886",
"0.1687",
"0.727",
"0.1687"
],
[
"0.3015",
"0.1818",
"0.6943",
"0.1829",
"0.6943",
"0.1965",
"0.3015",
"0.1953"
],
[
"0.3315",
"0.1561",
"0.523",
"0.1561",
"0.523",
"0.1663",
"0.3315",
"0.1663"
],
[
"0.1184",
"0.2239",
"0.4844",
"0.2239",
"0.4844",
"0.2369",
"0.1184",
"0.2369"
],
[
"0.525",
"0.2245",
"0.575",
"0.2245",
"0.575",
"0.2393",
"0.525",
"0.2393"
],
[
"0.7866",
"0.2257",
"0.8193",
"0.2257",
"0.8193",
"0.2369",
"0.7866",
"0.2369"
],
[
"0.831",
"0.2257",
"0.867",
"0.2257",
"0.867",
"0.2382",
"0.831",
"0.2382"
],
[
"0.11755",
"0.2393",
"0.4836",
"0.2393",
"0.4836",
"0.253",
"0.11755",
"0.253"
],
[
"0.6406",
"0.2452",
"0.7515",
"0.2452",
"0.7515",
"0.257",
"0.6406",
"0.257"
],
[
"0.11755",
"0.2542",
"0.4844",
"0.2546",
"0.4844",
"0.2695",
"0.11755",
"0.2683"
],
[
"0.5283",
"0.2588",
"0.7095",
"0.2588",
"0.7095",
"0.2737",
"0.5283",
"0.2737"
],
[
"0.7866",
"0.26",
"0.8193",
"0.26",
"0.8193",
"0.2725",
"0.7866",
"0.2725"
],
[
"0.831",
"0.26",
"0.8657",
"0.26",
"0.8657",
"0.2725",
"0.831",
"0.2725"
],
[
"0.11755",
"0.2715",
"0.1763",
"0.2715",
"0.1763",
"0.2861",
"0.11755",
"0.2861"
],
[
"0.5264",
"0.2725",
"0.721",
"0.2737",
"0.721",
"0.288",
"0.5264",
"0.2874"
],
[
"0.7866",
"0.2708",
"0.8193",
"0.2725",
"0.818",
"0.2874",
"0.785",
"0.2852"
],
[
"0.831",
"0.2725",
"0.8657",
"0.275",
"0.864",
"0.2893",
"0.8296",
"0.2874"
],
[
"0.5283",
"0.288",
"0.7734",
"0.288",
"0.7734",
"0.3018",
"0.5283",
"0.3018"
],
[
"0.7866",
"0.288",
"0.8193",
"0.288",
"0.8193",
"0.3005",
"0.7866",
"0.3005"
],
[
"0.831",
"0.2852",
"0.8657",
"0.2874",
"0.864",
"0.3018",
"0.8296",
"0.2998"
],
[
"0.12177",
"0.2964",
"0.4844",
"0.2976",
"0.4844",
"0.3123",
"0.12177",
"0.3113"
],
[
"0.5283",
"0.3027",
"0.708",
"0.3027",
"0.708",
"0.3164",
"0.5283",
"0.3164"
],
[
"0.7866",
"0.3027",
"0.8213",
"0.3027",
"0.8213",
"0.3154",
"0.7866",
"0.3154"
],
[
"0.831",
"0.3027",
"0.8657",
"0.3027",
"0.8657",
"0.3154",
"0.831",
"0.3154"
],
[
"0.11755",
"0.3135",
"0.4844",
"0.3135",
"0.4844",
"0.3267",
"0.11755",
"0.3267"
],
[
"0.642",
"0.3176",
"0.75",
"0.3176",
"0.75",
"0.3289",
"0.642",
"0.3289"
],
[
"0.1184",
"0.33",
"0.4878",
"0.33",
"0.4878",
"0.3445",
"0.1184",
"0.3445"
],
[
"0.785",
"0.33",
"0.8213",
"0.33",
"0.8213",
"0.3457",
"0.785",
"0.3457"
],
[
"0.8296",
"0.3289",
"0.8696",
"0.3308",
"0.867",
"0.3467",
"0.8286",
"0.3433"
],
[
"0.5283",
"0.3445",
"0.6978",
"0.3457",
"0.6978",
"0.3594",
"0.5283",
"0.3582"
],
[
"0.5283",
"0.3308",
"0.706",
"0.3308",
"0.706",
"0.3445",
"0.5283",
"0.3445"
],
[
"0.785",
"0.3457",
"0.8193",
"0.3457",
"0.8193",
"0.3582",
"0.785",
"0.3582"
],
[
"0.8296",
"0.3445",
"0.867",
"0.3445",
"0.867",
"0.3599",
"0.8296",
"0.3599"
],
[
"0.833",
"0.3599",
"0.8657",
"0.3599",
"0.8657",
"0.3723",
"0.833",
"0.3723"
],
[
"0.11755",
"0.3467",
"0.4844",
"0.3467",
"0.4844",
"0.361",
"0.11755",
"0.361"
],
[
"0.1184",
"0.3635",
"0.486",
"0.3635",
"0.486",
"0.3772",
"0.1184",
"0.3772"
],
[
"0.529",
"0.361",
"0.6885",
"0.361",
"0.6885",
"0.3718",
"0.529",
"0.3718"
],
[
"0.5264",
"0.3723",
"0.7036",
"0.3723",
"0.7036",
"0.386",
"0.5264",
"0.386"
],
[
"0.784",
"0.3723",
"0.8213",
"0.3723",
"0.8213",
"0.3884",
"0.784",
"0.3884"
],
[
"0.1184",
"0.379",
"0.486",
"0.379",
"0.486",
"0.3926",
"0.1184",
"0.3926"
],
[
"0.833",
"0.3748",
"0.867",
"0.3748",
"0.867",
"0.3872",
"0.833",
"0.3872"
],
[
"0.11755",
"0.3938",
"0.4836",
"0.3938",
"0.4836",
"0.4067",
"0.11755",
"0.4067"
],
[
"0.5283",
"0.3938",
"0.575",
"0.3938",
"0.575",
"0.408",
"0.5283",
"0.408"
],
[
"0.7866",
"0.3955",
"0.8213",
"0.3955",
"0.8213",
"0.4067",
"0.7866",
"0.4067"
],
[
"0.8286",
"0.3926",
"0.869",
"0.3943",
"0.8657",
"0.4104",
"0.827",
"0.4067"
],
[
"0.7866",
"0.4092",
"0.8193",
"0.4092",
"0.8193",
"0.4216",
"0.7866",
"0.4216"
],
[
"0.8296",
"0.408",
"0.867",
"0.4104",
"0.8657",
"0.4229",
"0.8286",
"0.4204"
],
[
"0.529",
"0.4104",
"0.601",
"0.4104",
"0.601",
"0.4216",
"0.529",
"0.4216"
],
[
"0.1184",
"0.411",
"0.486",
"0.411",
"0.486",
"0.4246",
"0.1184",
"0.4246"
],
[
"0.1184",
"0.427",
"0.4836",
"0.427",
"0.4836",
"0.4407",
"0.1184",
"0.4407"
],
[
"0.5146",
"0.4358",
"0.876",
"0.437",
"0.876",
"0.452",
"0.5146",
"0.4507"
],
[
"0.1184",
"0.4436",
"0.3208",
"0.4436",
"0.3208",
"0.4573",
"0.1184",
"0.4573"
],
[
"0.1234",
"0.4685",
"0.4844",
"0.4697",
"0.4844",
"0.484",
"0.1234",
"0.4827"
],
[
"0.11584",
"0.484",
"0.486",
"0.485",
"0.486",
"0.4995",
"0.11584",
"0.4988"
],
[
"0.513",
"0.4785",
"0.879",
"0.4785",
"0.879",
"0.4922",
"0.513",
"0.4922"
],
[
"0.511",
"0.4946",
"0.7188",
"0.4954",
"0.7188",
"0.5103",
"0.511",
"0.509"
],
[
"0.1184",
"0.502",
"0.4844",
"0.502",
"0.4844",
"0.5156",
"0.1184",
"0.5156"
],
[
"0.11755",
"0.5166",
"0.4844",
"0.517",
"0.4844",
"0.532",
"0.11755",
"0.531"
],
[
"0.1184",
"0.5337",
"0.4844",
"0.5337",
"0.4844",
"0.5474",
"0.1184",
"0.5474"
],
[
"0.571",
"0.5337",
"0.677",
"0.532",
"0.677",
"0.5464",
"0.571",
"0.549"
],
[
"0.692",
"0.53",
"0.8213",
"0.533",
"0.8213",
"0.552",
"0.692",
"0.549"
],
[
"0.858",
"0.5337",
"0.882",
"0.5337",
"0.882",
"0.5474",
"0.858",
"0.5474"
],
[
"0.11584",
"0.549",
"0.486",
"0.549",
"0.486",
"0.5664",
"0.11584",
"0.5664"
],
[
"0.12006",
"0.568",
"0.4844",
"0.568",
"0.4844",
"0.582",
"0.12006",
"0.582"
],
[
"0.5146",
"0.5737",
"0.879",
"0.5737",
"0.879",
"0.5874",
"0.5146",
"0.5874"
],
[
"0.11584",
"0.58",
"0.486",
"0.582",
"0.486",
"0.5996",
"0.11584",
"0.597"
],
[
"0.51",
"0.5884",
"0.879",
"0.5884",
"0.879",
"0.6016",
"0.51",
"0.6016"
],
[
"0.11755",
"0.597",
"0.4836",
"0.597",
"0.4836",
"0.611",
"0.11755",
"0.611"
],
[
"0.511",
"0.604",
"0.882",
"0.605",
"0.882",
"0.619",
"0.511",
"0.618"
],
[
"0.1184",
"0.6143",
"0.4844",
"0.6143",
"0.4844",
"0.629",
"0.1184",
"0.629"
],
[
"0.51",
"0.6206",
"0.879",
"0.6206",
"0.879",
"0.6343",
"0.51",
"0.6343"
],
[
"0.1184",
"0.6313",
"0.4844",
"0.6313",
"0.4844",
"0.6445",
"0.1184",
"0.6445"
],
[
"0.511",
"0.637",
"0.882",
"0.6357",
"0.882",
"0.651",
"0.511",
"0.652"
],
[
"0.1184",
"0.6465",
"0.4895",
"0.6455",
"0.4895",
"0.66",
"0.1184",
"0.661"
],
[
"0.5146",
"0.654",
"0.882",
"0.654",
"0.882",
"0.6675",
"0.5146",
"0.6675"
],
[
"0.11755",
"0.663",
"0.4895",
"0.663",
"0.4895",
"0.677",
"0.11755",
"0.677"
],
[
"0.511",
"0.6685",
"0.879",
"0.6685",
"0.879",
"0.6816",
"0.511",
"0.6816"
],
[
"0.1142",
"0.6777",
"0.4878",
"0.6777",
"0.4878",
"0.6953",
"0.1142",
"0.6953"
],
[
"0.513",
"0.685",
"0.882",
"0.685",
"0.882",
"0.6997",
"0.513",
"0.6997"
],
[
"0.1184",
"0.7007",
"0.12427",
"0.7007",
"0.12427",
"0.707",
"0.1184",
"0.707"
],
[
"0.513",
"0.7017",
"0.881",
"0.7017",
"0.881",
"0.7153",
"0.513",
"0.7153"
],
[
"0.2007",
"0.712",
"0.403",
"0.712",
"0.403",
"0.727",
"0.2007",
"0.727"
],
[
"0.465",
"0.7114",
"0.4895",
"0.7153",
"0.4844",
"0.728",
"0.461",
"0.7236"
],
[
"0.513",
"0.7173",
"0.881",
"0.7173",
"0.881",
"0.7324",
"0.513",
"0.7324"
],
[
"0.241",
"0.7256",
"0.294",
"0.7256",
"0.294",
"0.7407",
"0.241",
"0.7407"
],
[
"0.511",
"0.733",
"0.6885",
"0.733",
"0.6885",
"0.7466",
"0.511",
"0.7466"
],
[
"0.1142",
"0.7476",
"0.486",
"0.749",
"0.486",
"0.763",
"0.1142",
"0.7627"
],
[
"0.511",
"0.761",
"0.544",
"0.761",
"0.544",
"0.774",
"0.511",
"0.774"
],
[
"0.5425",
"0.7627",
"0.66",
"0.7627",
"0.66",
"0.7754",
"0.5425",
"0.7754"
],
[
"0.5415",
"0.7656",
"0.5474",
"0.7656",
"0.5474",
"0.7715",
"0.5415",
"0.7715"
],
[
"0.1184",
"0.764",
"0.2301",
"0.764",
"0.2301",
"0.7793",
"0.1184",
"0.7793"
],
[
"0.126",
"0.7915",
"0.419",
"0.7915",
"0.419",
"0.8047",
"0.126",
"0.8047"
],
[
"0.4114",
"0.792",
"0.4844",
"0.792",
"0.4844",
"0.804",
"0.4114",
"0.804"
],
[
"0.511",
"0.7876",
"0.7446",
"0.7876",
"0.7446",
"0.802",
"0.511",
"0.802"
],
[
"0.11584",
"0.8057",
"0.486",
"0.807",
"0.486",
"0.822",
"0.11584",
"0.821"
],
[
"0.513",
"0.807",
"0.882",
"0.807",
"0.882",
"0.822",
"0.513",
"0.822"
],
[
"0.11755",
"0.8247",
"0.486",
"0.8247",
"0.486",
"0.8384",
"0.11755",
"0.8384"
],
[
"0.5146",
"0.8247",
"0.879",
"0.8247",
"0.879",
"0.8384",
"0.5146",
"0.8384"
],
[
"0.11755",
"0.8403",
"0.4844",
"0.8403",
"0.4844",
"0.854",
"0.11755",
"0.854"
],
[
"0.513",
"0.8403",
"0.882",
"0.8403",
"0.882",
"0.854",
"0.513",
"0.854"
],
[
"0.1184",
"0.8555",
"0.4895",
"0.8555",
"0.4895",
"0.869",
"0.1184",
"0.869"
],
[
"0.511",
"0.857",
"0.881",
"0.857",
"0.881",
"0.8706",
"0.511",
"0.8706"
],
[
"0.11584",
"0.8706",
"0.4878",
"0.8716",
"0.4878",
"0.8857",
"0.11584",
"0.8853"
],
[
"0.513",
"0.8726",
"0.882",
"0.8726",
"0.882",
"0.8857",
"0.513",
"0.8857"
],
[
"0.1184",
"0.888",
"0.486",
"0.888",
"0.486",
"0.902",
"0.1184",
"0.902"
],
[
"0.511",
"0.8896",
"0.881",
"0.8896",
"0.881",
"0.903",
"0.511",
"0.903"
],
[
"0.11755",
"0.905",
"0.486",
"0.905",
"0.486",
"0.9185",
"0.11755",
"0.9185"
],
[
"0.511",
"0.905",
"0.879",
"0.905",
"0.879",
"0.9185",
"0.511",
"0.9185"
]
] | [
"Task",
"Dataset",
"#Examples",
"# Tables",
"Input",
"Output",
"80,654",
"24,241",
"Question",
"Answer",
"QuestionAnswering",
"WIKISQL-WEAK (Zhong et al., 2017)",
"22,033",
"Question",
"Answer",
"2,108",
"WIKITQ (Pasupat and Liang, 2015)",
"TABFACT (Chen et al., 2020b)",
"118,275",
"16,573",
"Statement",
"Boolean",
"Text",
"Fact Verification",
"Faithful Text Generation",
"37,015",
"7,392",
"Columns",
"Table 2: Overview of downstream tasks used in this paper.",
"LogicNLG (Chen et al., 2020a)",
"中。REASTAP的微调类似于第3节中讨论的",
"Model",
"Dev",
"Test",
"预训练过程。我们将两个下游任务建模为序列",
"Previous Models",
"生成任务，并利用生成式LM以自回归方式生",
"MAPO (Liang et al., 2018)",
"71.8",
"72.4",
"成输出",
"MeRL (Agarwal et al., 2019)",
"74.9",
"74.8",
"LatentAlignment (Wang et al., 2019)",
"79.4",
"79.3",
"Table QA WIKISQL-WEAK E(Zhong et al.,",
"HardEM (Min et al., 2019)",
"84.4",
"83.9",
"2017)和 WIKITQ E(Pasupat and Liang,2015)",
"Pre-trained LMs",
"用于评估REASTAP在表QA任务上的性能。",
"85.1",
"83.6",
"GraPPa (Yu et al., 2021a)",
"TaPas (Herzig et al., 2020)",
"85.9",
"84.7",
"86.0",
"WIKISQL-WEAK(Zhongetal.,2017)要求模型",
"对表单元格值执行过滤和可选聚合，以回答给",
"T5-3B (Xie et al., 2022)",
"TAPEX (Liu et al., 2022b)",
"89.3",
"定的问题。WIKITQ(Pasupatand Liang，2015)",
"89.2",
"需要更广泛的推理技能，因此更具挑战性。表",
"BART",
"86.9",
"86.1",
"91.1",
"90.4",
"REASTAP",
"QA任务制定与REASTAP预训练任务相同。我",
"们使用表示精度，它检查预测的答案是否等于",
"Table 3: Denotation accuracies on WIKISQL-WEAK",
"基本事实，作为评估指标",
"TableFactsVerification我们选择TABFACT",
"E(Chenetal.,2020b）来评估REASTAP在表格",
"输入表，目标是生成一个既流畅又事实正确的",
"句子y=（1,92,..·,Yn）：",
"事实上的性能验证任务。给定一个表和一个语",
"句，TABFACT（Chenetal.,2020b）试图区分该语",
"句是被该表包含还是被该表驳斥。TABFACT将",
"y= argmax",
" P(yily<i, T; 0)",
"(3)",
"其测试集分为测试simple和测试complex子集，其",
"中测试complex包含需要更复杂的表推理技能的",
"为了评估生成的句子的逻辑保真度，Chenetal.",
"示例。此外，它选择了一个小的测试集测试small",
"(2020a)提出了两种基于模型的评估方法：基于",
"与2K样本为人类评估。为了在TABFACT上",
"解析的评估（SP-Acc）和基于NLI的评估（NLI",
"进行微调，在BARTE(Lewisetal.2020）之后",
"Acc）。SP-Acc直接从生成的句子中提取意义",
"我们对解码器中最后一个令牌的隐藏状态应用",
"表示，并对表执行以验证正确性。NLI-ACC使",
"了二进制分类器，用于输出。Theobjectiveisto-",
"用自然语言推理（NLI）模型来预测蕴涵关系。",
"generatetheverificationlabel L (0, 1] giventhes-",
"在Chen et al.(2020a）之后，我们使用 SP-Acc",
"tatement s = ($1, $2, , Sn) andtheinputtable T",
"NLI-Acc作为逻辑保真度评估指标；和BLEU",
"..",
"1/2/3作为表面级别的评估指标。值得注意的",
"L = argmax P(il s, T; 0)",
"(2)",
"是，较高的BLEU分数与更好的逻辑保真度无",
"ie(0,1)",
"关(Nan et al., 2022a)。",
"Weusedtheaccuracy(i。e。，正确预测的百分比）",
"5",
"Experiments",
"",
"作为评估指标",
"Faithful Table-to-Text Generation",
"我们选",
"5.1Implementation Details",
"择 LoGIcNLG（Chen et al.，2020a）来评估",
"我们基于fairseq库E(Ottetal.,2019)实现了我",
"REASTAP在忠实的表格到文本生成任务上",
"们的模型。我们采用了BART-Large作为骨干",
"的性能。与以前的 Table-to-Text 生成基准测",
"模型。对于表预训练，我们合成并采样了4M",
"试 (Wiseman et al., 2017; Balakrishnan et al.,",
"对推理示例。在以下各节中，除非明确规定，",
"2019; Parikh et al., 2020; Nan et al., 2022b)相",
"否则所有实验结果均在4M推理示例和BART",
"比，LOGICNLG的任务是生成选定表区域在逻",
"Large配置的默认设置下进行评估。我们的预",
"辑上需要的语句。给定其选定列为T的序列化",
"训练程序运行了80,000个步骤，批量大小为"
] | {
"arxiv_id": "2210.12374",
"language": "zh",
"page_idx": 3
} |
||
2205.15960-zh-3 | 2205.15960-zh-3.jpg | 写的。Sundanese 是一种主要的 SVO 语言，具有语音标记并包含一些(可选)演员-动词协议，即。e。,numberandperson (Kurniawan,2013) TobaBatak (bbc) 是北苏门答腊省使用的一种语言。与 Acehnese 类似，它在城市和移民地区正逐渐被印度尼西亚人取代。它曾经是用Batak 脚本编写的，但现在主要是用拉丁脚本编写的。Batak 语言是动词首字母，并且具有让人联想到菲律宾语言的动词系统，尽管它们在许多细节上与它们不同$\overline{\mathrm{E}}($Blust et al., 2013)
## 3 Data Construction
我们的数据收集过程由几个步骤组成。首先， 我们将高资源本地语言(印尼语)的现有数据集作为扩展到其他十种语言的基础，并要求人工注释者翻译文本。为了确保最终翻译的质量，我们使用额外的人工注释器进行质量保证
# 3.1 Annotator Recruitment
用代表性不足的语言提取或注释数据通常需要与当地语言社区合作，以确定合格的双语使用者 (Nekoto et al.,2020)。在印尼语环境中，大多数语言都有几种方言，这加剧了这一挑战。由于印尼语中的方言在用词和意义上可能存在显著差异 (Aji et al.,2022),因此招募说相同或相似方言的注释员很重要，以确保翻译相互理解
在这项工作中，我们雇用了至少 2 名专家注释者，他们是每种当地语言和印尼语的母语人土。为了过滤招募的注释者，我们首先要求注释者候选人翻译三个样本。然后，我们进行同行评审，询问他们是否可以理解同一语言的其他注释者的翻译，使用雇用的注释者作为翻译和翻译验证者。我们还进行了2 小时的培训， 为选定的工人介绍注释系统的用户界面。对于英语翻译，我们根据他们的英语水平测试成绩聘请注释员，雅思成绩？??。5 或托福 PBT 成绩？?
## 3.2 Data Filtering and Sampling
我们的数据集以 SmSA 为基础，SmSA 是IndoNLU 基准$\overline{\mathbb{E}}$(Purwarianti and Crisdayanti, 2019; Wilie et al., 2020) 中最大的公开印尼情绪分析数据集。SmSA 是一个专家注释的句子级多域情感分析数据集，包含从 Twitter, Zomato 和 TripAdvisor 等多个在线平台收集的 11,000多个评论和评论实例。我们通过手动检查所有句子来过滤数据以删除辱骂性语言和个人身份信息。我们通过分层抽样随机选择 1000 个样本进行翻译，确保标签分布是平衡的
## 前是用阿拉伯语，爪哇语和 Sundanese 文字书3.3 Human Translation
我们指示注释者保留文本的含义，并使没有目标语言翻译的实体(如人员，组织，地点和时间)保持不变。具体来说，我们指示他们：(1) 保持句子的情感极性；(2)保留实体；(3)保持原始文本的完整信息内容
最初，我们要求翻译人员维护排版。原始数据集中的大多数句子都是用非正式的语气写的，拼写不标准，例如。g。、拉长元音和标点。当句子被翻译成目标语言时，直接翻译可能听起来不自然。例如，在 Minangkabau 中将印尼语单词 kangeeeen (最初为 kangen ; en: miss ) 翻译成 taragaaaak (最初为 taragak ) 可能听起来不自然。同样，原始句子也可能包含错别字。由于难以准确评估翻译的印刷一致性，我们将其作为标准删除
## 3.4 Human-Assisted Quality Assurance
我们通过让注释者 A 检查注释者 B 的翻译来在两个注释者之间进行质量控制 (QC),反之亦然。我们在数据集中包含更正的翻译。为了确保很好地执行质量保证，我们通过删除随机的单词序列来随机干扰 5 个句子的%。然后， 质量保证注释器将注意到受干扰的句子并对其进行修复
我们分析了本文作者所说的巴厘岛语，Sundanese 和 Javanese 的质量保证编辑。对于每种语言，我们随机抽取 100 个由 QC 注释者编辑的翻译。我们将编辑分类如下：TyposandMechanics:涉及更正错别字，标点，大小写，空格/破折号和数字格式的编辑
Orthography: 由于本地语言中的拼字变化而没有标准拼字法，因此可以更改单词的拼写。这个词在编辑前后听起来和意思都是一样的，都是本地人使用的。QC 注释者可能会觉得一个书写变体更自然/更常用，因此进行此更改
翻译：翻译人员使用的单词仍使用印尼语，QC
注释员将其翻译为当地语言
Word 编辑：QC 注释器解释一个单词/短语。这
还包括添加/删除单词和语素更改
主要更改：显著改变原始翻译的其他编辑
结果如表 1 所示。通常，单词编辑构成了QC 修改的大部分，其中涉及用同义词替换单词/短语或稍微改变语素。相比之下，重大变化极为罕见。我们还看到了大约 10 个% 的拼字法的变化。其他类型的编辑因语言而异。与其他语言相比，Sundanese 的错别字明显较少， 但翻译编辑的数量要高得多。我们怀疑这是因为在 Sundanese 中经常发生与印尼语的代码切换，尽管存在等效的 Sundanese 翻译，但仍采用了许多印尼语单词。 | [
[
"0.12006",
"0.0873",
"0.4895",
"0.0873",
"0.4895",
"0.10095",
"0.12006",
"0.10095"
],
[
"0.4895",
"0.0885",
"0.712",
"0.0885",
"0.712",
"0.0986",
"0.4895",
"0.0986"
],
[
"0.1184",
"0.1027",
"0.4878",
"0.1027",
"0.4878",
"0.1164",
"0.1184",
"0.1164"
],
[
"0.511",
"0.10925",
"0.8774",
"0.10925",
"0.8774",
"0.1223",
"0.511",
"0.1223"
],
[
"0.12006",
"0.12054",
"0.4836",
"0.12054",
"0.4836",
"0.1312",
"0.12006",
"0.1312"
],
[
"0.511",
"0.1247",
"0.879",
"0.1247",
"0.879",
"0.1383",
"0.511",
"0.1383"
],
[
"0.1184",
"0.1342",
"0.486",
"0.1342",
"0.486",
"0.1478",
"0.1184",
"0.1478"
],
[
"0.513",
"0.1414",
"0.882",
"0.1414",
"0.882",
"0.155",
"0.513",
"0.155"
],
[
"0.1368",
"0.152",
"0.4785",
"0.152",
"0.4785",
"0.165",
"0.1368",
"0.165"
],
[
"0.5146",
"0.1567",
"0.882",
"0.1567",
"0.882",
"0.1704",
"0.5146",
"0.1704"
],
[
"0.11755",
"0.1663",
"0.486",
"0.1663",
"0.486",
"0.1799",
"0.11755",
"0.1799"
],
[
"0.51",
"0.1716",
"0.7363",
"0.1729",
"0.7363",
"0.187",
"0.51",
"0.1859"
],
[
"0.1184",
"0.1841",
"0.4844",
"0.1841",
"0.4844",
"0.1978",
"0.1184",
"0.1978"
],
[
"0.531",
"0.1907",
"0.881",
"0.1907",
"0.881",
"0.2037",
"0.531",
"0.2037"
],
[
"0.1184",
"0.1996",
"0.486",
"0.1996",
"0.486",
"0.2131",
"0.1184",
"0.2131"
],
[
"0.511",
"0.2048",
"0.881",
"0.206",
"0.881",
"0.2203",
"0.511",
"0.2197"
],
[
"0.12006",
"0.2173",
"0.4836",
"0.2173",
"0.4836",
"0.228",
"0.12006",
"0.228"
],
[
"0.513",
"0.2227",
"0.881",
"0.2227",
"0.881",
"0.2369",
"0.513",
"0.2369"
],
[
"0.12006",
"0.234",
"0.486",
"0.234",
"0.486",
"0.2446",
"0.12006",
"0.2446"
],
[
"0.513",
"0.2382",
"0.881",
"0.2382",
"0.881",
"0.2517",
"0.513",
"0.2517"
],
[
"0.11755",
"0.2488",
"0.476",
"0.2488",
"0.476",
"0.2625",
"0.11755",
"0.2625"
],
[
"0.513",
"0.2542",
"0.881",
"0.2542",
"0.881",
"0.2673",
"0.513",
"0.2673"
],
[
"0.5146",
"0.2708",
"0.881",
"0.2708",
"0.881",
"0.2852",
"0.5146",
"0.2852"
],
[
"0.1125",
"0.2766",
"0.3123",
"0.2778",
"0.3123",
"0.2922",
"0.1125",
"0.2915"
],
[
"0.513",
"0.2852",
"0.879",
"0.2852",
"0.879",
"0.2986",
"0.513",
"0.2986"
],
[
"0.1184",
"0.3047",
"0.4902",
"0.3047",
"0.4902",
"0.3184",
"0.1184",
"0.3184"
],
[
"0.513",
"0.3027",
"0.881",
"0.3027",
"0.881",
"0.3164",
"0.513",
"0.3164"
],
[
"0.1184",
"0.3206",
"0.4878",
"0.3206",
"0.4878",
"0.3342",
"0.1184",
"0.3342"
],
[
"0.51",
"0.3176",
"0.881",
"0.3176",
"0.881",
"0.3308",
"0.51",
"0.3308"
],
[
"0.1184",
"0.3374",
"0.4836",
"0.3374",
"0.4836",
"0.3474",
"0.1184",
"0.3474"
],
[
"0.5146",
"0.3374",
"0.678",
"0.3374",
"0.678",
"0.3474",
"0.5146",
"0.3474"
],
[
"0.12006",
"0.3528",
"0.4844",
"0.3528",
"0.4844",
"0.3665",
"0.12006",
"0.3665"
],
[
"0.51",
"0.3623",
"0.839",
"0.3635",
"0.839",
"0.3777",
"0.51",
"0.3772"
],
[
"0.1184",
"0.3694",
"0.4836",
"0.3694",
"0.4836",
"0.3801",
"0.1184",
"0.3801"
],
[
"0.5146",
"0.386",
"0.881",
"0.386",
"0.881",
"0.3997",
"0.5146",
"0.3997"
],
[
"0.11584",
"0.3955",
"0.3494",
"0.3967",
"0.3494",
"0.411",
"0.11584",
"0.4104"
],
[
"0.5146",
"0.4026",
"0.879",
"0.4026",
"0.879",
"0.4163",
"0.5146",
"0.4163"
],
[
"0.1184",
"0.4192",
"0.4844",
"0.4192",
"0.4844",
"0.4329",
"0.1184",
"0.4329"
],
[
"0.513",
"0.4187",
"0.8774",
"0.4187",
"0.8774",
"0.4316",
"0.513",
"0.4316"
],
[
"0.1184",
"0.4353",
"0.4844",
"0.4353",
"0.4844",
"0.449",
"0.1184",
"0.449"
],
[
"0.5146",
"0.4358",
"0.879",
"0.4358",
"0.879",
"0.4465",
"0.5146",
"0.4465"
],
[
"0.1184",
"0.452",
"0.4878",
"0.452",
"0.4878",
"0.4656",
"0.1184",
"0.4656"
],
[
"0.5146",
"0.4507",
"0.881",
"0.4507",
"0.881",
"0.4644",
"0.5146",
"0.4644"
],
[
"0.1184",
"0.4685",
"0.4895",
"0.4685",
"0.4895",
"0.4822",
"0.1184",
"0.4822"
],
[
"0.513",
"0.466",
"0.879",
"0.466",
"0.879",
"0.4797",
"0.513",
"0.4797"
],
[
"0.12006",
"0.4863",
"0.4844",
"0.4863",
"0.4844",
"0.4963",
"0.12006",
"0.4963"
],
[
"0.5146",
"0.484",
"0.5845",
"0.484",
"0.5845",
"0.4946",
"0.5146",
"0.4946"
],
[
"0.11755",
"0.4988",
"0.4844",
"0.5005",
"0.4844",
"0.514",
"0.11755",
"0.5127"
],
[
"0.531",
"0.4995",
"0.882",
"0.4995",
"0.882",
"0.513",
"0.531",
"0.513"
],
[
"0.1184",
"0.5166",
"0.4844",
"0.5166",
"0.4844",
"0.53",
"0.1184",
"0.53"
],
[
"0.513",
"0.5156",
"0.879",
"0.5156",
"0.879",
"0.5293",
"0.513",
"0.5293"
],
[
"0.1184",
"0.533",
"0.1545",
"0.533",
"0.1545",
"0.546",
"0.1184",
"0.546"
],
[
"0.513",
"0.532",
"0.881",
"0.532",
"0.881",
"0.546",
"0.513",
"0.546"
],
[
"0.1368",
"0.549",
"0.4895",
"0.549",
"0.4895",
"0.5625",
"0.1368",
"0.5625"
],
[
"0.51",
"0.5464",
"0.881",
"0.5464",
"0.881",
"0.56",
"0.51",
"0.56"
],
[
"0.1184",
"0.5654",
"0.4895",
"0.5654",
"0.4895",
"0.579",
"0.1184",
"0.579"
],
[
"0.51",
"0.5625",
"0.882",
"0.5635",
"0.882",
"0.5776",
"0.51",
"0.5767"
],
[
"0.12006",
"0.583",
"0.486",
"0.583",
"0.486",
"0.5933",
"0.12006",
"0.5933"
],
[
"0.511",
"0.58",
"0.739",
"0.58",
"0.739",
"0.5933",
"0.511",
"0.5933"
],
[
"0.1184",
"0.597",
"0.486",
"0.597",
"0.486",
"0.611",
"0.1184",
"0.611"
],
[
"0.511",
"0.5967",
"0.881",
"0.5967",
"0.881",
"0.611",
"0.511",
"0.611"
],
[
"0.1184",
"0.6133",
"0.4878",
"0.6133",
"0.4878",
"0.627",
"0.1184",
"0.627"
],
[
"0.513",
"0.6133",
"0.881",
"0.6133",
"0.881",
"0.627",
"0.513",
"0.627"
],
[
"0.12177",
"0.6313",
"0.486",
"0.6313",
"0.486",
"0.641",
"0.12177",
"0.641"
],
[
"0.51",
"0.6274",
"0.881",
"0.6274",
"0.881",
"0.641",
"0.51",
"0.641"
],
[
"0.1184",
"0.6455",
"0.4902",
"0.6455",
"0.4902",
"0.659",
"0.1184",
"0.659"
],
[
"0.513",
"0.6445",
"0.881",
"0.6445",
"0.881",
"0.6577",
"0.513",
"0.6577"
],
[
"0.12006",
"0.663",
"0.486",
"0.663",
"0.486",
"0.6733",
"0.12006",
"0.6733"
],
[
"0.513",
"0.661",
"0.8486",
"0.661",
"0.8486",
"0.675",
"0.513",
"0.675"
],
[
"0.1184",
"0.6777",
"0.4878",
"0.6777",
"0.4878",
"0.6914",
"0.1184",
"0.6914"
],
[
"0.513",
"0.6777",
"0.882",
"0.6777",
"0.882",
"0.6914",
"0.513",
"0.6914"
],
[
"0.1184",
"0.695",
"0.4895",
"0.695",
"0.4895",
"0.708",
"0.1184",
"0.708"
],
[
"0.513",
"0.695",
"0.733",
"0.695",
"0.733",
"0.708",
"0.513",
"0.708"
],
[
"0.11755",
"0.7114",
"0.1621",
"0.7114",
"0.1621",
"0.7227",
"0.11755",
"0.7227"
],
[
"0.513",
"0.7104",
"0.881",
"0.7104",
"0.881",
"0.7236",
"0.513",
"0.7236"
],
[
"0.513",
"0.7256",
"0.7734",
"0.7256",
"0.7734",
"0.7393",
"0.513",
"0.7393"
],
[
"0.11584",
"0.7363",
"0.388",
"0.7383",
"0.388",
"0.753",
"0.11584",
"0.7505"
],
[
"0.5146",
"0.7437",
"0.843",
"0.7437",
"0.843",
"0.7573",
"0.5146",
"0.7573"
],
[
"0.11755",
"0.759",
"0.486",
"0.759",
"0.486",
"0.7725",
"0.11755",
"0.7725"
],
[
"0.529",
"0.7603",
"0.879",
"0.7603",
"0.879",
"0.774",
"0.529",
"0.774"
],
[
"0.11755",
"0.777",
"0.4844",
"0.777",
"0.4844",
"0.7876",
"0.11755",
"0.7876"
],
[
"0.513",
"0.777",
"0.881",
"0.777",
"0.881",
"0.7905",
"0.513",
"0.7905"
],
[
"0.1184",
"0.792",
"0.486",
"0.792",
"0.486",
"0.8057",
"0.1184",
"0.8057"
],
[
"0.513",
"0.792",
"0.881",
"0.792",
"0.881",
"0.8057",
"0.513",
"0.8057"
],
[
"0.1184",
"0.808",
"0.486",
"0.808",
"0.486",
"0.822",
"0.1184",
"0.822"
],
[
"0.511",
"0.808",
"0.881",
"0.807",
"0.881",
"0.822",
"0.511",
"0.8223"
],
[
"0.1184",
"0.8247",
"0.4878",
"0.8247",
"0.4878",
"0.8384",
"0.1184",
"0.8384"
],
[
"0.513",
"0.8247",
"0.881",
"0.8247",
"0.881",
"0.8384",
"0.513",
"0.8384"
],
[
"0.1184",
"0.8413",
"0.4878",
"0.8413",
"0.4878",
"0.855",
"0.1184",
"0.855"
],
[
"0.511",
"0.8413",
"0.881",
"0.8413",
"0.881",
"0.855",
"0.511",
"0.855"
],
[
"0.12006",
"0.8594",
"0.4836",
"0.8594",
"0.4836",
"0.869",
"0.12006",
"0.869"
],
[
"0.51",
"0.8555",
"0.879",
"0.857",
"0.879",
"0.8706",
"0.51",
"0.869"
],
[
"0.1184",
"0.8726",
"0.4836",
"0.8726",
"0.4836",
"0.8857",
"0.1184",
"0.8857"
],
[
"0.5146",
"0.8745",
"0.879",
"0.8745",
"0.879",
"0.8853",
"0.5146",
"0.8853"
],
[
"0.1184",
"0.8896",
"0.4895",
"0.8896",
"0.4895",
"0.903",
"0.1184",
"0.903"
],
[
"0.511",
"0.8896",
"0.879",
"0.8896",
"0.879",
"0.903",
"0.511",
"0.903"
],
[
"0.1184",
"0.9062",
"0.4114",
"0.9062",
"0.4114",
"0.9194",
"0.1184",
"0.9194"
],
[
"0.5146",
"0.9067",
"0.687",
"0.9067",
"0.687",
"0.9175",
"0.5146",
"0.9175"
]
] | [
"前是用阿拉伯语，爪哇语和Sundanese文字书",
"3.3HumanTranslation",
"写的。Sundanese是一种主要的SVO语言，具",
"我们指示注释者保留文本的含义，并使没有目",
"有语音标记并包含一些（可选）演员-动词协",
"标语言翻译的实体（如人员，组织，地点和时",
"议，即。e。,numberandperson (Kurniawan,2013)",
"间）保持不变。具体来说，我们指示他们：（1）",
"TobaBatak（bbc）是北苏门答腊省使用的一",
"保持句子的情感极性；（2）保留实体；（3）保",
"种语言。与Acehnese类似，它在城市和移民",
"持原始文本的完整信息内容",
"地区正逐渐被印度尼西亚人取代。它曾经是用",
"最初，我们要求翻译人员维护排版。原始数",
"Batak脚本编写的，但现在主要是用拉丁脚本",
"据集中的大多数句子都是用非正式的语气写",
"编写的。Batak语言是动词首字母，并且具有",
"的，拼写不标准，例如。g。、拉长元音和标点。",
"让人联想到菲律宾语言的动词系统，尽管它们",
"当句子被翻译成目标语言时，直接翻译可能听",
"在许多细节上与它们不同E(Blustetal.,2013)",
"起来不自然。例如，在Minangkabau中将印尼",
"语单词kangeeeen（最初为kangen；en：miss）",
"3Data Construction",
"翻译成taragaaaak（最初为taragak）可能听",
"我们的数据收集过程由几个步骤组成。首先，",
"起来不自然。同样，原始句子也可能包含错别",
"我们将高资源本地语言（印尼语）的现有数据",
"字。由于难以准确评估翻译的印刷一致性，我",
"集作为扩展到其他十种语言的基础，并要求",
"们将其作为标准删除",
"人工注释者翻译文本。为了确保最终翻译的质",
"3.4Human-Assisted Quality Assurance",
"量，我们使用额外的人工注释器进行质量保证",
"我们通过让注释者A检查注释者B的翻译来",
"3.1 Annotator Recruitment",
"在两个注释者之间进行质量控制（QC），反之",
"用代表性不足的语言提取或注释数据通常需要",
"亦然。我们在数据集中包含更正的翻译。为了",
"与当地语言社区合作，以确定合格的双语使用",
"确保很好地执行质量保证，我们通过删除随机",
"者（Nekotoetal.2020）。在印尼语环境中，大",
"的单词序列来随机干扰5个句子的%。然后，",
"多数语言都有几种方言，这加剧了这一挑战。",
"质量保证注释器将注意到受干扰的句子并对其",
"由于印尼语中的方言在用词和意义上可能存在",
"进行修复",
"显著差异（Ajietal.，2022），因此招募说相同",
"我们分析了本文作者所说的巴厘岛语，Sun-",
"或相似方言的注释员很重要，以确保翻译相互",
"danese和Javanese的质量保证编辑。对于每种",
"理解",
"语言，我们随机抽取100个由QC注释者编辑",
"在这项工作中，我们雇用了至少2名专家注",
"的翻译。我们将编辑分类如下：TyposandMe",
"释者，他们是每种当地语言和印尼语的母语人",
"chanics：涉及更正错别字，标点，大小写，空",
"王。为了过滤招募的注释者，我们首先要求注",
"格/破折号和数字格式的编辑",
"释者候选人翻译三个样本。然后，我们进行同",
"Orthography：由于本地语言中的拼字变化而没",
"行评审，询问他们是否可以理解同一语言的其",
"有标准拼字法，因此可以更改单词的拼写。这",
"他注释者的翻译，使用雇用的注释者作为翻译",
"个词在编辑前后听起来和意思都是一样的，都",
"和翻译验证者。我们还进行了2小时的培训，",
"是本地人使用的。QC注释者可能会觉得一个",
"为选定的工人介绍注释系统的用户界面。对于",
"书写变体更自然/更常用，因此进行此更改",
"英语翻译，我们根据他们的英语水平测试成绩",
"翻译：翻译人员使用的单词仍使用印尼语，QC",
"聘请注释员，雅思成绩？?？。5或托福PBT成",
"注释员将其翻译为当地语言",
"绩???",
"Word编辑：QC注释器解释一个单词/短语。这",
"还包括添加/删除单词和语素更改",
"3.2Data Filtering and Sampling",
"主要更改：显著改变原始翻译的其他编辑",
"我们的数据集以SmSA为基础，SmSA是",
"结果如表1所示。通常，单词编辑构成了",
"IndoNLU基准F(Purwarianti and Crisdayanti.",
"QC修改的大部分，其中涉及用同义词替换单",
"2019；Wilieetal,2020）中最大的公开印尼情绪",
"词/短语或稍微改变语素。相比之下，重大变",
"分析数据集。SmSA是一个专家注释的句子级",
"化极为罕见。我们还看到了大约10个%的拼",
"多域情感分析数据集，包含从Twitter，Zomato",
"字法的变化。其他类型的编辑因语言而异。与",
"和TripAdvisor等多个在线平台收集的11,000",
"其他语言相比，Sundanese的错别字明显较少，",
"多个评论和评论实例。我们通过手动检查所有",
"但翻译编辑的数量要高得多。我们怀疑这是因",
"句子来过滤数据以删除辱骂性语言和个人身份",
"为在Sundanese中经常发生与印尼语的代码切",
"信息。我们通过分层抽样随机选择1000个样",
"换，尽管存在等效的Sundanese翻译，但仍采",
"本进行翻译，确保标签分布是平衡的",
"用了许多印尼语单词。"
] | {
"arxiv_id": "2205.15960",
"language": "zh",
"page_idx": 3
} |
||
2205.15960-zh-4 | 2205.15960-zh-4.jpg | <table>
<tbody>
<tr>
<th>Subset</th>
<th>Negative</th>
<th>Neutral</th>
<th>Positive</th>
</tr>
<tr>
<td>Train</td>
<td>192</td>
<td>119</td>
<td>189</td>
</tr>
<tr>
<td>Valid</td>
<td>38</td>
<td>24</td>
<td>38</td>
</tr>
<tr>
<td>Test</td>
<td>153</td>
<td>96</td>
<td>151</td>
</tr>
</tbody>
</table>
Table 2: Label distribution of NusaX Sentiment
dataset.
<table>
<tbody>
<tr>
<th>Category</th>
<th>ban</th>
<th>Sun</th>
<th>jav</th>
</tr>
<tr>
<td>Typos& Mechanic</td>
<td>31</td>
<td>14</td>
<td>42</td>
</tr>
<tr>
<td>Orthography</td>
<td>14</td>
<td>6</td>
<td>12</td>
</tr>
<tr>
<td>Translation</td>
<td>22</td>
<td>55</td>
<td>10</td>
</tr>
<tr>
<td>Word edit</td>
<td>67</td>
<td>65</td>
<td>61</td>
</tr>
<tr>
<td>${\mathrm{Major~changes}}$</td>
<td>3</td>
<td>0</td>
<td>1</td>
</tr>
</tbody>
</table>
Table 1: Statistics of QC edits per category over 100
samples.
### 3.5 Bilingual Lexicon Creation
双语词典可用于数据增强(Wang et al., 2022) 和评估跨语言表示 (Artetxe et al.,2018)。我们从印尼语词典$^{2}$中选择 400 个单词，将其翻译成 10 种当地语言和英语。对于每种语言，我们使用两个注释器，并要求他们将单词翻译成所有可能的词素。来自两个注释器的翻译组合在一起。我们为 11 种语言对中的每一种获得800-1,600 个单词对(从印尼语到其余语言)。我们用 PanLex (Kamholz et al., 2014) 的数据扩充了双语词典
## 4 NusaX Benchmark
## 4.1 Tasks
我们基于涵盖 12 种语言的数据集开发了两个任务-情感分析和机器翻译-包括印尼语，英语和 10 种当地语言。对于 NusaX 情感数据集，每种语言都有相同的标签分布，我们在 Table 2中显示了每个数据集子集的标签分布。我们维护每个数据集子集中的标签比率，以确保类似的分布。数据集的更多细节在附录 C 中提供
# 4.1.1 Sentiment Analysis
情感分析是一项 NLP 任务 , 旨在识别给定文本文档的情感。情绪通常分为 3 类：积极，消极和中立。我们将数据集的构建重点放在情绪分析上，因为它是印度尼西亚$\overline{\mathrm{E}}($Aji et al.,2022)中最广泛探索的任务之一，因为它具有广泛的行业相关性，例如用于竞争对手和营销分析，以及检测不利的谣言以进行风险管理$\overline{\mathrm{E}}($Socher et al.,2013)。从情感分析数据集(SmSA)翻译1,000个实例后，我们为每种翻译的语言创建了一个情感分析数据集。对于每种语言，我们将数据集分为 500 个训练，100 个验证和 400个测试示例。我们的数据集总共包含 6,000 个火车，1,200 个验证和 4,800 个测试实例，涉及12 种语言(印尼语，英语和 10 种当地语言)
$^{2}$https://github.com/andria009/
IndonesianSentimentLexicon
## 4.1.2 $\textbf{Machine Translation}$
印度尼西亚由 700 多种语言组成，涵盖三种不同的语言家族 (Aji et al., 2022) 。尽管其语言多样性，但现有的机器翻译系统仅涵盖一小部分印尼语，主要是印尼语 (国家语言), Sundanese 和 Javanese。为了扩大现有机器翻译系统对代表性不足的本地语言的覆盖范围，我们使用翻译后的情感语料库构建了机器翻译数据集，从而在所有语言对之间产生了并行语料库。换句话说，我们有 132 个可能的并行语料库，每个语料库有 1,000 个样本 (500 个训练，100 个验证和 400 个测试实例), 可用于训练机器翻译模型。与许多其他 MT 评估数据集相比，我们的数据属于评论领域，而不是以英语为中心
# 4.2 Baselines
### 4.2.1 Classical Machine Learning
经典的机器学习方法仍然被印度尼西亚当地的研究人员和机构广泛使用，因为它们的效率(Nityasya et al.,2021)。在计算有限的情况下， 性能和计算成本之间的权衡尤为重要，这在低资源语言中很常见。因此，我们使用经典方法作为比较的基线。也就是说，我们使用朴素贝叶斯、SVM 和逻辑回归来完成分类任务。对于 MT, 我们采用了复制印尼原始文本的朴素基线，使用双语词典的基于词典的替换方法以及基于 Moses$\boxed{\mathrm{F}}($Koehn et al., 2007) 的基于短语的 MT 系统
4.2.2 Pre-trained Local Language Models
神经预训练 LM 的最新发展为各种 NLP 任务带来了实质性的改进。尽管缺乏印尼语和当地语言的资源，但仍在努力为印尼语和主要当地语言开发大型的预培训 LM。IndoBERTE(Wilie et al., 2020) 和 SundaneseBERT$\boxed{\mathrm{F}}($Wongso et al., 2022) 是印尼语和 Sundanese 中用于自然语言理解 (NLU) 任务的两个流行的 LM。IndoBART 和 IndoGPT 也被引人用于印尼语，Sundanese 和 Javanese (Cahyawijaya et al., 2021) 中的自然语言生成 (NLG) 任务。我们使用这些 LM 作为基线来评估它们对其他语言的适应性 | [
[
"0.1628",
"0.0885",
"0.2334",
"0.0903",
"0.2334",
"0.1051",
"0.1628",
"0.1027"
],
[
"0.3225",
"0.0903",
"0.3552",
"0.0903",
"0.3552",
"0.1027",
"0.3225",
"0.1027"
],
[
"0.5483",
"0.0903",
"0.601",
"0.0903",
"0.601",
"0.10156",
"0.5483",
"0.10156"
],
[
"0.6206",
"0.0903",
"0.687",
"0.0903",
"0.687",
"0.10156",
"0.6206",
"0.10156"
],
[
"0.708",
"0.0903",
"0.764",
"0.0903",
"0.764",
"0.10156",
"0.708",
"0.10156"
],
[
"0.785",
"0.08905",
"0.8486",
"0.08905",
"0.8486",
"0.10394",
"0.785",
"0.10394"
],
[
"0.409",
"0.09143",
"0.4375",
"0.09143",
"0.4375",
"0.10394",
"0.409",
"0.10394"
],
[
"0.367",
"0.09143",
"0.398",
"0.09143",
"0.398",
"0.1027",
"0.367",
"0.1027"
],
[
"0.546",
"0.1081",
"0.5903",
"0.10986",
"0.5884",
"0.1259",
"0.544",
"0.12354"
],
[
"0.723",
"0.10986",
"0.754",
"0.10986",
"0.754",
"0.1247",
"0.723",
"0.1247"
],
[
"0.3242",
"0.111",
"0.3477",
"0.111",
"0.3477",
"0.12354",
"0.3242",
"0.12354"
],
[
"0.1663",
"0.111",
"0.3074",
"0.111",
"0.3074",
"0.1259",
"0.1663",
"0.1259"
],
[
"0.372",
"0.111",
"0.3955",
"0.111",
"0.3955",
"0.1247",
"0.372",
"0.1247"
],
[
"0.413",
"0.111",
"0.4375",
"0.111",
"0.4375",
"0.1247",
"0.413",
"0.1247"
],
[
"0.6396",
"0.111",
"0.671",
"0.111",
"0.671",
"0.1259",
"0.6396",
"0.1259"
],
[
"0.802",
"0.111",
"0.833",
"0.111",
"0.833",
"0.1259",
"0.802",
"0.1259"
],
[
"0.7256",
"0.1247",
"0.748",
"0.1247",
"0.748",
"0.1395",
"0.7256",
"0.1395"
],
[
"0.3208",
"0.114",
"0.3267",
"0.114",
"0.3267",
"0.1194",
"0.3208",
"0.1194"
],
[
"0.1646",
"0.1259",
"0.2603",
"0.1277",
"0.2603",
"0.1426",
"0.1646",
"0.1401"
],
[
"0.5483",
"0.1259",
"0.5903",
"0.1259",
"0.5903",
"0.1395",
"0.5483",
"0.1395"
],
[
"0.642",
"0.1259",
"0.6665",
"0.1259",
"0.6665",
"0.1383",
"0.642",
"0.1383"
],
[
"0.804",
"0.1259",
"0.8286",
"0.1259",
"0.8286",
"0.1395",
"0.804",
"0.1395"
],
[
"0.3728",
"0.1277",
"0.392",
"0.1277",
"0.392",
"0.1414",
"0.3728",
"0.1414"
],
[
"0.3267",
"0.1271",
"0.3494",
"0.1271",
"0.3494",
"0.1414",
"0.3267",
"0.1414"
],
[
"0.4148",
"0.1277",
"0.4358",
"0.1277",
"0.4358",
"0.1414",
"0.4148",
"0.1414"
],
[
"0.546",
"0.1383",
"0.5825",
"0.1383",
"0.5825",
"0.1538",
"0.546",
"0.1538"
],
[
"0.6396",
"0.1395",
"0.671",
"0.1395",
"0.671",
"0.1538",
"0.6396",
"0.1538"
],
[
"0.7256",
"0.1395",
"0.748",
"0.1395",
"0.748",
"0.1526",
"0.7256",
"0.1526"
],
[
"0.8037",
"0.1395",
"0.8296",
"0.1395",
"0.8296",
"0.1526",
"0.8037",
"0.1526"
],
[
"0.3242",
"0.1401",
"0.351",
"0.1401",
"0.351",
"0.158",
"0.3242",
"0.158"
],
[
"0.1646",
"0.1401",
"0.251",
"0.1426",
"0.251",
"0.1567",
"0.1646",
"0.155"
],
[
"0.3704",
"0.1401",
"0.3972",
"0.1401",
"0.3972",
"0.158",
"0.3704",
"0.158"
],
[
"0.4148",
"0.1426",
"0.4375",
"0.1426",
"0.4375",
"0.1567",
"0.4148",
"0.1567"
],
[
"0.1663",
"0.158",
"0.241",
"0.158",
"0.241",
"0.1729",
"0.1663",
"0.1729"
],
[
"0.3267",
"0.158",
"0.3494",
"0.158",
"0.3494",
"0.1729",
"0.3267",
"0.1729"
],
[
"0.3704",
"0.158",
"0.3955",
"0.158",
"0.3955",
"0.1729",
"0.3704",
"0.1729"
],
[
"0.413",
"0.158",
"0.434",
"0.158",
"0.434",
"0.1729",
"0.413",
"0.1729"
],
[
"0.1646",
"0.1733",
"0.2747",
"0.1733",
"0.2747",
"0.1882",
"0.1646",
"0.1882"
],
[
"0.3315",
"0.1746",
"0.345",
"0.1746",
"0.345",
"0.1859",
"0.3315",
"0.1859"
],
[
"0.3745",
"0.1746",
"0.3896",
"0.1746",
"0.3896",
"0.1859",
"0.3745",
"0.1859"
],
[
"0.4182",
"0.1746",
"0.4307",
"0.1746",
"0.4307",
"0.1859",
"0.4182",
"0.1859"
],
[
"0.511",
"0.1692",
"0.881",
"0.1692",
"0.881",
"0.1829",
"0.511",
"0.1829"
],
[
"0.511",
"0.1818",
"0.567",
"0.1841",
"0.565",
"0.199",
"0.51",
"0.1965"
],
[
"0.11584",
"0.2031",
"0.486",
"0.2037",
"0.486",
"0.2185",
"0.11584",
"0.2173"
],
[
"0.1184",
"0.2203",
"0.1763",
"0.2203",
"0.1763",
"0.2322",
"0.1184",
"0.2322"
],
[
"0.511",
"0.2268",
"0.738",
"0.228",
"0.738",
"0.2423",
"0.511",
"0.2411"
],
[
"0.5146",
"0.2517",
"0.881",
"0.2517",
"0.881",
"0.2654",
"0.5146",
"0.2654"
],
[
"0.11584",
"0.2612",
"0.3762",
"0.2612",
"0.3762",
"0.275",
"0.11584",
"0.275"
],
[
"0.513",
"0.2673",
"0.879",
"0.2673",
"0.879",
"0.2808",
"0.513",
"0.2808"
],
[
"0.11755",
"0.2832",
"0.4844",
"0.2832",
"0.4844",
"0.2964",
"0.11755",
"0.2964"
],
[
"0.513",
"0.2852",
"0.879",
"0.2852",
"0.879",
"0.2986",
"0.513",
"0.2986"
],
[
"0.1184",
"0.2998",
"0.4844",
"0.2986",
"0.4844",
"0.3135",
"0.1184",
"0.3142"
],
[
"0.511",
"0.2998",
"0.879",
"0.2998",
"0.879",
"0.3135",
"0.511",
"0.3135"
],
[
"0.12006",
"0.3164",
"0.486",
"0.3164",
"0.486",
"0.33",
"0.12006",
"0.33"
],
[
"0.513",
"0.3176",
"0.8774",
"0.3176",
"0.8774",
"0.328",
"0.513",
"0.328"
],
[
"0.1184",
"0.3333",
"0.4878",
"0.3333",
"0.4878",
"0.3467",
"0.1184",
"0.3467"
],
[
"0.513",
"0.3333",
"0.881",
"0.3333",
"0.881",
"0.3467",
"0.513",
"0.3467"
],
[
"0.1184",
"0.3486",
"0.4878",
"0.3486",
"0.4878",
"0.3623",
"0.1184",
"0.3623"
],
[
"0.51",
"0.3474",
"0.8774",
"0.3474",
"0.8774",
"0.361",
"0.51",
"0.361"
],
[
"0.1184",
"0.3652",
"0.4844",
"0.3652",
"0.4844",
"0.376",
"0.1184",
"0.376"
],
[
"0.511",
"0.3652",
"0.881",
"0.3652",
"0.881",
"0.379",
"0.511",
"0.379"
],
[
"0.11755",
"0.3813",
"0.4878",
"0.3813",
"0.4878",
"0.3943",
"0.11755",
"0.3943"
],
[
"0.513",
"0.3813",
"0.881",
"0.3813",
"0.881",
"0.3943",
"0.513",
"0.3943"
],
[
"0.11755",
"0.3967",
"0.492",
"0.3967",
"0.492",
"0.4104",
"0.11755",
"0.4104"
],
[
"0.513",
"0.3967",
"0.881",
"0.3967",
"0.881",
"0.4104",
"0.513",
"0.4104"
],
[
"0.1184",
"0.4133",
"0.4878",
"0.4133",
"0.4878",
"0.427",
"0.1184",
"0.427"
],
[
"0.511",
"0.412",
"0.881",
"0.412",
"0.881",
"0.4258",
"0.511",
"0.4258"
],
[
"0.11755",
"0.427",
"0.2284",
"0.4287",
"0.2284",
"0.4436",
"0.11755",
"0.4412"
],
[
"0.513",
"0.4287",
"0.881",
"0.4287",
"0.881",
"0.4424",
"0.513",
"0.4424"
],
[
"0.513",
"0.4453",
"0.8613",
"0.4453",
"0.8613",
"0.459",
"0.513",
"0.459"
],
[
"0.1142",
"0.459",
"0.1528",
"0.459",
"0.1528",
"0.4744",
"0.1142",
"0.4744"
],
[
"0.1503",
"0.4602",
"0.318",
"0.4602",
"0.318",
"0.4739",
"0.1503",
"0.4739"
],
[
"0.511",
"0.4785",
"0.634",
"0.4785",
"0.634",
"0.4934",
"0.511",
"0.4934"
],
[
"0.11584",
"0.4863",
"0.2107",
"0.488",
"0.2107",
"0.503",
"0.11584",
"0.5005"
],
[
"0.513",
"0.505",
"0.7925",
"0.505",
"0.7925",
"0.5186",
"0.513",
"0.5186"
],
[
"0.1184",
"0.511",
"0.4844",
"0.511",
"0.4844",
"0.525",
"0.1184",
"0.525"
],
[
"0.1184",
"0.528",
"0.486",
"0.528",
"0.486",
"0.5415",
"0.1184",
"0.5415"
],
[
"0.513",
"0.527",
"0.881",
"0.527",
"0.881",
"0.5405",
"0.513",
"0.5405"
],
[
"0.1184",
"0.5435",
"0.486",
"0.5435",
"0.486",
"0.557",
"0.1184",
"0.557"
],
[
"0.513",
"0.5444",
"0.8774",
"0.5444",
"0.8774",
"0.5547",
"0.513",
"0.5547"
],
[
"0.11755",
"0.5586",
"0.4878",
"0.558",
"0.4878",
"0.5723",
"0.11755",
"0.5737"
],
[
"0.511",
"0.56",
"0.882",
"0.56",
"0.882",
"0.5737",
"0.511",
"0.5737"
],
[
"0.12006",
"0.576",
"0.486",
"0.576",
"0.486",
"0.589",
"0.12006",
"0.589"
],
[
"0.511",
"0.5747",
"0.882",
"0.5767",
"0.882",
"0.5903",
"0.511",
"0.5884"
],
[
"0.1184",
"0.5913",
"0.4844",
"0.5913",
"0.4844",
"0.605",
"0.1184",
"0.605"
],
[
"0.513",
"0.593",
"0.881",
"0.593",
"0.881",
"0.6055",
"0.513",
"0.6055"
],
[
"0.1184",
"0.608",
"0.476",
"0.608",
"0.476",
"0.6216",
"0.1184",
"0.6216"
],
[
"0.5146",
"0.6094",
"0.876",
"0.6094",
"0.876",
"0.619",
"0.5146",
"0.619"
],
[
"0.5146",
"0.626",
"0.879",
"0.626",
"0.879",
"0.6357",
"0.5146",
"0.6357"
],
[
"0.11755",
"0.6357",
"0.33",
"0.6357",
"0.33",
"0.6494",
"0.11755",
"0.6494"
],
[
"0.513",
"0.64",
"0.881",
"0.64",
"0.881",
"0.654",
"0.513",
"0.654"
],
[
"0.1184",
"0.6567",
"0.4878",
"0.6567",
"0.4878",
"0.6714",
"0.1184",
"0.6714"
],
[
"0.513",
"0.6567",
"0.881",
"0.6567",
"0.881",
"0.6704",
"0.513",
"0.6704"
],
[
"0.12006",
"0.6733",
"0.486",
"0.6733",
"0.486",
"0.687",
"0.12006",
"0.687"
],
[
"0.511",
"0.6704",
"0.881",
"0.6714",
"0.881",
"0.686",
"0.511",
"0.685"
],
[
"0.11755",
"0.6885",
"0.4878",
"0.6904",
"0.4878",
"0.7036",
"0.11755",
"0.7017"
],
[
"0.513",
"0.6895",
"0.6055",
"0.6895",
"0.6055",
"0.7007",
"0.513",
"0.7007"
],
[
"0.1184",
"0.706",
"0.4844",
"0.706",
"0.4844",
"0.7197",
"0.1184",
"0.7197"
],
[
"0.1184",
"0.7217",
"0.486",
"0.7217",
"0.486",
"0.7354",
"0.1184",
"0.7354"
],
[
"0.5146",
"0.7227",
"0.555",
"0.7227",
"0.555",
"0.733",
"0.5146",
"0.733"
],
[
"0.55",
"0.7236",
"0.5728",
"0.7236",
"0.5728",
"0.731",
"0.55",
"0.731"
],
[
"0.567",
"0.72",
"0.8555",
"0.72",
"0.8555",
"0.734",
"0.567",
"0.734"
],
[
"0.1184",
"0.7383",
"0.4878",
"0.7383",
"0.4878",
"0.752",
"0.1184",
"0.752"
],
[
"0.511",
"0.742",
"0.879",
"0.7437",
"0.879",
"0.7583",
"0.511",
"0.7573"
],
[
"0.12006",
"0.7544",
"0.4836",
"0.7544",
"0.4836",
"0.764",
"0.12006",
"0.764"
],
[
"0.513",
"0.7603",
"0.881",
"0.7603",
"0.881",
"0.774",
"0.513",
"0.774"
],
[
"0.11755",
"0.7686",
"0.4836",
"0.7686",
"0.4836",
"0.782",
"0.11755",
"0.782"
],
[
"0.513",
"0.777",
"0.8774",
"0.777",
"0.8774",
"0.7876",
"0.513",
"0.7876"
],
[
"0.11584",
"0.785",
"0.4836",
"0.785",
"0.4836",
"0.799",
"0.11584",
"0.799"
],
[
"0.511",
"0.7915",
"0.879",
"0.7915",
"0.879",
"0.8047",
"0.511",
"0.8047"
],
[
"0.12177",
"0.8027",
"0.486",
"0.8027",
"0.486",
"0.8164",
"0.12177",
"0.8164"
],
[
"0.513",
"0.808",
"0.881",
"0.808",
"0.881",
"0.822",
"0.513",
"0.822"
],
[
"0.1184",
"0.8184",
"0.4878",
"0.8184",
"0.4878",
"0.832",
"0.1184",
"0.832"
],
[
"0.511",
"0.8247",
"0.879",
"0.8247",
"0.879",
"0.8384",
"0.511",
"0.8384"
],
[
"0.1184",
"0.8345",
"0.4844",
"0.8345",
"0.4844",
"0.847",
"0.1184",
"0.847"
],
[
"0.513",
"0.8413",
"0.882",
"0.8413",
"0.882",
"0.855",
"0.513",
"0.855"
],
[
"0.1184",
"0.851",
"0.4878",
"0.851",
"0.4878",
"0.8643",
"0.1184",
"0.8643"
],
[
"0.513",
"0.857",
"0.881",
"0.857",
"0.881",
"0.8706",
"0.513",
"0.8706"
],
[
"0.11584",
"0.865",
"0.471",
"0.866",
"0.471",
"0.88",
"0.11584",
"0.879"
],
[
"0.513",
"0.8726",
"0.881",
"0.8726",
"0.881",
"0.887",
"0.513",
"0.887"
],
[
"0.1368",
"0.891",
"0.3645",
"0.8936",
"0.3645",
"0.9077",
"0.1368",
"0.9062"
],
[
"0.5146",
"0.8896",
"0.881",
"0.8896",
"0.881",
"0.903",
"0.5146",
"0.903"
],
[
"0.11755",
"0.9067",
"0.3123",
"0.9067",
"0.3123",
"0.9175",
"0.11755",
"0.9175"
],
[
"0.5146",
"0.9067",
"0.8237",
"0.9067",
"0.8237",
"0.9175",
"0.5146",
"0.9175"
]
] | [
"Category",
"ban",
"Subset",
"Negative",
"Neutral",
"Positive",
"jav",
"sun",
"Train",
"119",
"31",
"Typos &Mechanic",
"14",
"42",
"192",
"189",
"24",
"",
"Orthography",
"Valid",
"38",
"38",
"6",
"14",
"12",
"Test",
"153",
"96",
"151",
"22",
"Translation",
"55",
"10",
"Word edit",
"67",
"65",
"61",
"Major changes",
"3",
"0",
"1",
"Table 2: Label distribution of NusaX Sentiment",
"dataset.",
"Table 1: Statistics of QC edits per category over 100",
"samples.",
"4.1.2Machine Translation",
"印度尼西亚由700多种语言组成，涵盖三种不",
"3.5Bilingual Lexicon Creation",
"同的语言家族（Ajietal.,2022）。尽管其语言多",
"双语词典可用于数据增强（Wangetal.,2022)和",
"样性，但现有的机器翻译系统仅涵盖一小部分",
"评估跨语言表示（Artetxeetal.，2018）。我们从",
"印尼语，主要是印尼语（国家语言），Sundanese",
"印尼语词典2中选择400个单词，将其翻译",
"和Javanese。为了扩大现有机器翻译系统对代",
"成10种当地语言和英语。对于每种语言，我",
"表性不足的本地语言的覆盖范围，我们使用翻",
"们使用两个注释器，并要求他们将单词翻译成",
"译后的情感语料库构建了机器翻译数据集，从",
"所有可能的词素。来自两个注释器的翻译组合",
"而在所有语言对之间产生了并行语料库。换句",
"在一起。我们为11种语言对中的每一种获得",
"话说，我们有132个可能的并行语料库，每个",
"800-1,600个单词对（从印尼语到其余语言）。",
"语料库有1,000个样本（500个训练，100个验",
"我们用PanLex（Kamholzetal.，2014）的数据扩",
"证和400个测试实例），可用于训练机器翻译",
"充了双语词典",
"模型。与许多其他MT评估数据集相比，我们",
"的数据属于评论领域，而不是以英语为中心",
"4",
"NusaX Benchmark",
"4.2Baselines",
"4.1 Tasks",
"4.2.1Classical Machine Learning",
"我们基于涵盖12种语言的数据集开发了两个",
"任务-情感分析和机器翻译-包括印尼语，英语",
"经典的机器学习方法仍然被印度尼西亚当地",
"和10种当地语言。对于NusaX情感数据集，每",
"的研究人员和机构广泛使用，因为它们的效率",
"种语言都有相同的标签分布，我们在Table 2",
"(Nityasyaetal.，2021）。在计算有限的情况下，",
"中显示了每个数据集子集的标签分布。我们维",
"性能和计算成本之间的权衡尤为重要，这在低",
"护每个数据集子集中的标签比率，以确保类似",
"资源语言中很常见。因此，我们使用经典方法",
"的分布。数据集的更多细节在附录C中提供",
"作为比较的基线。也就是说，我们使用朴素贝",
"叶斯、SVM和逻辑回归来完成分类任务。对",
"4.1.1 Sentiment Analysis",
"于MT，我们采用了复制印尼原始文本的朴素",
"情感分析是一项NLP任务，旨在识别给定文本",
"基线，使用双语词典的基于词典的替换方法以",
"文档的情感。情绪通常分为3类：积极，消极和",
"及基于MosesE(Koehnetal.,2007)的基于短语",
"中立。我们将数据集的构建重点放在情绪分析",
"的MT系统",
"上，因为它是印度尼西亚E(Ajietal.,2022）中",
"最广泛探索的任务之一，因为它具有广泛的行",
"4.2.2",
"",
" Pre-trained Local Language Models",
"业相关性，例如用于竞争对手和营销分析，以",
"神经预训练LM的最新发展为各种NLP任务",
"及检测不利的谣言以进行风险管理F(Sochei",
"带来了实质性的改进。尽管缺乏印尼语和当地",
"etal.,2013）。从情感分析数据集（SmSA）翻译",
"语言的资源：但仍在努力为印尼语和主要当地",
"1,000个实例后，我们为每种翻译的语言创建",
"语言开发大型的预培训LM。IndoBERTE(Wilie",
"了一个情感分析数据集。对于每种语言，我们",
"et al., 2020)和 SundaneseBERTE(Wongso et al.,",
"将数据集分为500个训练，100个验证和400",
"2022）是印尼语和Sundanese中用于自然语言",
"个测试示例。我们的数据集总共包含6,000个",
"理解（NLU）任务的两个流行的LM。IndoBART",
"火车，1,200个验证和4,800个测试实例，涉及",
"和IndoGPT也被引人用于印尼语，Sundanese",
"12种语言（印尼语，英语和10种当地语言）",
"和Javanese（Cahyawijaya et al.,2021）中的自然",
"2https://github.com/andria009/",
"语言生成（NLG）任务。我们使用这些LM作",
"IndonesianSentimentLexicor",
"为基线来评估它们对其他语言的适应性"
] | {
"arxiv_id": "2205.15960",
"language": "zh",
"page_idx": 4
} |
||
2001.05467-zh-1 | 2001.05467-zh-1.jpg | $\operatorname{AvgOut}:$
# 消除沉闷响应的简单输出概率度量
Tong Niu, Mohit Bansal
${\bar{U}}$NC Chapel Hill
$\{$tongn, mbansal$\}\textcircled{\mathrm{Ocs.unc.edu}}$
## Abstract
Many sequence-to-sequence dialogue models tend to generate safe, uninformative responses. There have been various useful efforts on trying to eliminate them. However, these approaches either improve decoding algorithms during inference, rely on hand-crafted fea- $\bar{\text{tures, or employ complex models. In our work, we}}$ build dialogue models that are dynamically aware of what utterances or tokens are dull without any featureengineering. Specifically, we start with a simple yet effective automatic metric, AvgOut , which calculates the average output probability distribution of all time steps on $\bar{\text{the decoder side during training. This met-}}$ ric directly estimates which tokens are more likely to be generated, thus making it a faithful evaluation of the model diversity (i.e., for diverse models, the token probabilities should be more evenly distributed rather than peaked at a few dull tokens). We then leverage this novel metric to propose three models that promote diversity without losing relevance. The frst model, MinAvgOut , directly maximizes the diversity score through the output distributions of each batch; the second model, Label Fine-Tuning ( LFT ), prepends to the source sequence a label continuously scaled by the diversity score to control the diversity level; the third model, RL , adopts Reinforcement Learning and treats the diversity score as a reward signal. Moreover, we experiment with a hybrid model by combining the loss terms of MinAvgOut and RL. All four models outperform their base LSTM-RNN model on both diversity and relevance by a large margin, and are comparable to or better than competitive baselines (also verified via human evaluation). Moreover, our approaches are orthogonal to the base model, making them applicable as an add-on to other emerging better dialogue models in the future.
# 1 Introduction
许多现代对话生成模型都使用序列到序列体系结构作为其主干$\overline{\mathrm{E}}($Sordoni et al. 2015),在将其成功应用于机器翻译(MT)$\mathsf{E}($Bahdanau, Cho, and Bengio 2015) 之后。但是，对话任务也有与 MT 不同的要求：响应不仅
Copyright $\odot2020$, Association for the Advancement of Ar-
tificial Intelligence (www.aaai.org). All rights reserved.
必须是“正确的”(连贯且相关),而且还需要多样化且信息丰富。然而，许多以前的作品已经报道 seq2seq 具有低语料库级多样性E(Li et al. 2016a; Serban et al. 2016; Sordoni et al. 2015; Vinyals and Le 2015) , 因为它倾向于产生安全，简洁和无信息的响应，例如“我不知道。”。这些反应不必要地使对话系统的交互性比应有的要少得多。
要增加对话回应的多样性，第一步是忠实地评估回应的多样性。以前的工作使用的指标与多样性相关，但并不强烈，例如不同令牌的比率E(Li et al. 2016a) 和响应长度$\mathbb{E}($Baheti et al.2018)。然而，回应可能很长，但在意义上非常无聊，例如“我确信我不知道它。“,或简短但有趣 (i。e。,包含了很多信息),比如“爸爸很刻薄。”。仅由模型调查离散令牌输出也不理想，因为这些令牌只是模型在每个时间步的输出概率分布的单个实现，这不可避免地丢失了由整个分布指示的有价值的信息。Li et al. ( 2016b ) 手动收集沉闷响应的短列表， 并在训练期间阻止模型产生此类话语。然而，手工规则的一个重要缺点是，一组沉闷的令牌或话语是静态的， 而实际上它通常在训练过程中演变：当当前的沉闷令牌被消除时，另一组可能会显示出来
在我们的工作中，$^{1}$我们从一种简单而有效的方法开始，以测量响应的多样性。我们将此指标命名为“平均输出概率分布”或 AvgOut,它直接从会话中训练模型本身提取信息。我们通过在训练期间跟踪解码器侧所有输出概率分布的指数平均值来计算它。该指标动态地测量模型偏向于哪些令牌，而无需任何手工规则，从而使其成为对模型多样性的忠实评估 (i。e。, 对于不同的模型，令牌概率应该更均匀地分布，而不是在几个沉闷的令牌处达到峰值)。此外，由于 AvgOut 是一维分类分布，而不是像熵一样的无量纲数值，因此它自然地承载和传达了更多有关模型多样性的信息
然后，我们提出了三个模型，利用我们的新指标来促进对话生成的多样性。第一个 MinAvgOut 模型最小化当前批次 AvgOut 的点积和跨批次的指数平均值AvgOut ,这鼓励生成低频令牌。第二个 LFT 模型使用标记的转导方法，并在训练过程中通过地面实况目标序列的多样性得分来缩放“多样性标签”,而在测试过程中，可以通过调整预期的多样性得分来生成不同多样性水平的响应。第三个 RL 模型利用强化学习，我们的新指标应用于离散令牌并作为奖励信号。此外，由于
${^{1}}$We will release all our code and model outputs. | [
[
"0.455",
"0.125",
"0.55",
"0.125",
"0.55",
"0.1395",
"0.455",
"0.1395"
],
[
"0.326",
"0.1433",
"0.675",
"0.1433",
"0.675",
"0.161",
"0.326",
"0.161"
],
[
"0.4019",
"0.1932",
"0.6045",
"0.1932",
"0.6045",
"0.2084",
"0.4019",
"0.2084"
],
[
"0.437",
"0.2084",
"0.564",
"0.2084",
"0.564",
"0.2216",
"0.437",
"0.2216"
],
[
"0.3945",
"0.2229",
"0.607",
"0.2229",
"0.607",
"0.2374",
"0.3945",
"0.2374"
],
[
"0.25",
"0.2954",
"0.3145",
"0.2974",
"0.3145",
"0.3125",
"0.25",
"0.31"
],
[
"0.5215",
"0.2974",
"0.911",
"0.2974",
"0.911",
"0.3113",
"0.5215",
"0.3113"
],
[
"0.5195",
"0.31",
"0.91",
"0.3113",
"0.91",
"0.3245",
"0.5195",
"0.324"
],
[
"0.10297",
"0.3193",
"0.4631",
"0.3193",
"0.4631",
"0.3333",
"0.10297",
"0.3333"
],
[
"0.523",
"0.3245",
"0.9126",
"0.3245",
"0.9126",
"0.3384",
"0.523",
"0.3384"
],
[
"0.10297",
"0.332",
"0.4631",
"0.332",
"0.4631",
"0.3452",
"0.10297",
"0.3452"
],
[
"0.5195",
"0.3384",
"0.911",
"0.339",
"0.911",
"0.3528",
"0.5195",
"0.3516"
],
[
"0.10297",
"0.3452",
"0.4617",
"0.3452",
"0.4617",
"0.3591",
"0.10297",
"0.3591"
],
[
"0.10297",
"0.358",
"0.4617",
"0.358",
"0.4617",
"0.3718",
"0.10297",
"0.3718"
],
[
"0.523",
"0.3528",
"0.911",
"0.3528",
"0.911",
"0.3662",
"0.523",
"0.3662"
],
[
"0.1013",
"0.3706",
"0.4617",
"0.3706",
"0.4617",
"0.3838",
"0.1013",
"0.3838"
],
[
"0.5215",
"0.3662",
"0.911",
"0.3662",
"0.911",
"0.3801",
"0.5215",
"0.3801"
],
[
"0.10297",
"0.3833",
"0.4631",
"0.3833",
"0.4631",
"0.3965",
"0.10297",
"0.3965"
],
[
"0.523",
"0.3806",
"0.556",
"0.3806",
"0.556",
"0.3926",
"0.523",
"0.3926"
],
[
"0.10455",
"0.3965",
"0.4617",
"0.3965",
"0.4617",
"0.4072",
"0.10455",
"0.4072"
],
[
"0.54",
"0.3958",
"0.911",
"0.3958",
"0.911",
"0.406",
"0.54",
"0.406"
],
[
"0.10455",
"0.4092",
"0.4575",
"0.4092",
"0.4575",
"0.4192",
"0.10455",
"0.4192"
],
[
"0.5215",
"0.4077",
"0.9126",
"0.4077",
"0.9126",
"0.4216",
"0.5215",
"0.4216"
],
[
"0.10297",
"0.4204",
"0.4617",
"0.4204",
"0.4617",
"0.4343",
"0.10297",
"0.4343"
],
[
"0.523",
"0.4216",
"0.911",
"0.4216",
"0.911",
"0.435",
"0.523",
"0.435"
],
[
"0.1013",
"0.432",
"0.46",
"0.433",
"0.46",
"0.4463",
"0.1013",
"0.4458"
],
[
"0.5215",
"0.435",
"0.911",
"0.435",
"0.911",
"0.449",
"0.5215",
"0.449"
],
[
"0.10455",
"0.4475",
"0.46",
"0.4475",
"0.46",
"0.4583",
"0.10455",
"0.4583"
],
[
"0.5215",
"0.4495",
"0.9126",
"0.4495",
"0.9126",
"0.4634",
"0.5215",
"0.4634"
],
[
"0.10297",
"0.4583",
"0.46",
"0.4583",
"0.46",
"0.4717",
"0.10297",
"0.4717"
],
[
"0.523",
"0.4634",
"0.911",
"0.4634",
"0.911",
"0.4766",
"0.523",
"0.4766"
],
[
"0.10297",
"0.4702",
"0.4631",
"0.4702",
"0.4631",
"0.4841",
"0.10297",
"0.4841"
],
[
"0.523",
"0.4766",
"0.911",
"0.4766",
"0.911",
"0.4905",
"0.523",
"0.4905"
],
[
"0.10297",
"0.483",
"0.4631",
"0.483",
"0.4631",
"0.4968",
"0.10297",
"0.4968"
],
[
"0.5195",
"0.4893",
"0.911",
"0.4905",
"0.911",
"0.505",
"0.5195",
"0.504"
],
[
"0.10455",
"0.4976",
"0.46",
"0.4976",
"0.46",
"0.5083",
"0.10455",
"0.5083"
],
[
"0.10297",
"0.5083",
"0.4617",
"0.5083",
"0.4617",
"0.5215",
"0.10297",
"0.5215"
],
[
"0.523",
"0.505",
"0.9126",
"0.505",
"0.9126",
"0.518",
"0.523",
"0.518"
],
[
"0.10455",
"0.523",
"0.46",
"0.523",
"0.46",
"0.5327",
"0.10455",
"0.5327"
],
[
"0.523",
"0.518",
"0.91",
"0.518",
"0.91",
"0.532",
"0.523",
"0.532"
],
[
"0.10455",
"0.535",
"0.4617",
"0.535",
"0.4617",
"0.5454",
"0.10455",
"0.5454"
],
[
"0.5195",
"0.531",
"0.9126",
"0.532",
"0.9126",
"0.547",
"0.5195",
"0.5454"
],
[
"0.5215",
"0.547",
"0.6724",
"0.547",
"0.6724",
"0.56",
"0.5215",
"0.56"
],
[
"0.6665",
"0.547",
"0.91",
"0.547",
"0.91",
"0.56",
"0.6665",
"0.56"
],
[
"0.10455",
"0.56",
"0.4617",
"0.56",
"0.4617",
"0.571",
"0.10455",
"0.571"
],
[
"0.10455",
"0.5474",
"0.46",
"0.5474",
"0.46",
"0.558",
"0.10455",
"0.558"
],
[
"0.523",
"0.56",
"0.9126",
"0.56",
"0.9126",
"0.5737",
"0.523",
"0.5737"
],
[
"0.10297",
"0.5713",
"0.4617",
"0.5713",
"0.4617",
"0.5854",
"0.10297",
"0.5854"
],
[
"0.5215",
"0.5747",
"0.767",
"0.5747",
"0.767",
"0.5884",
"0.5215",
"0.5884"
],
[
"0.10297",
"0.584",
"0.4631",
"0.584",
"0.4631",
"0.5977",
"0.10297",
"0.5977"
],
[
"0.536",
"0.5884",
"0.9126",
"0.5884",
"0.9126",
"0.6016",
"0.536",
"0.6016"
],
[
"0.10297",
"0.5967",
"0.4631",
"0.5967",
"0.4631",
"0.61",
"0.10297",
"0.61"
],
[
"0.5215",
"0.601",
"0.911",
"0.601",
"0.911",
"0.6157",
"0.5215",
"0.6157"
],
[
"0.10297",
"0.6094",
"0.4631",
"0.608",
"0.4631",
"0.622",
"0.10297",
"0.6226"
],
[
"0.5215",
"0.6157",
"0.91",
"0.6157",
"0.91",
"0.629",
"0.5215",
"0.629"
],
[
"0.10297",
"0.622",
"0.4617",
"0.622",
"0.4617",
"0.6353",
"0.10297",
"0.6353"
],
[
"0.10297",
"0.634",
"0.4617",
"0.634",
"0.4617",
"0.648",
"0.10297",
"0.648"
],
[
"0.5215",
"0.63",
"0.911",
"0.63",
"0.911",
"0.643",
"0.5215",
"0.643"
],
[
"0.1013",
"0.646",
"0.46",
"0.646",
"0.46",
"0.659",
"0.1013",
"0.659"
],
[
"0.5215",
"0.643",
"0.911",
"0.643",
"0.911",
"0.657",
"0.5215",
"0.657"
],
[
"0.10297",
"0.659",
"0.4617",
"0.659",
"0.4617",
"0.6724",
"0.10297",
"0.6724"
],
[
"0.523",
"0.657",
"0.9126",
"0.657",
"0.9126",
"0.6704",
"0.523",
"0.6704"
],
[
"0.10297",
"0.672",
"0.4617",
"0.672",
"0.4617",
"0.685",
"0.10297",
"0.685"
],
[
"0.523",
"0.672",
"0.911",
"0.672",
"0.911",
"0.685",
"0.523",
"0.685"
],
[
"0.10297",
"0.6846",
"0.4631",
"0.6846",
"0.4631",
"0.6978",
"0.10297",
"0.6978"
],
[
"0.5195",
"0.683",
"0.911",
"0.6846",
"0.911",
"0.6987",
"0.5195",
"0.6978"
],
[
"0.10297",
"0.6963",
"0.4631",
"0.6963",
"0.4631",
"0.7104",
"0.10297",
"0.7104"
],
[
"0.523",
"0.6987",
"0.911",
"0.6987",
"0.911",
"0.712",
"0.523",
"0.712"
],
[
"0.10297",
"0.7085",
"0.4617",
"0.709",
"0.4617",
"0.7236",
"0.10297",
"0.7227"
],
[
"0.523",
"0.7134",
"0.9126",
"0.7134",
"0.9126",
"0.7266",
"0.523",
"0.7266"
],
[
"0.10297",
"0.7227",
"0.4617",
"0.7227",
"0.4617",
"0.7363",
"0.10297",
"0.7363"
],
[
"0.523",
"0.7266",
"0.7817",
"0.7266",
"0.7817",
"0.7407",
"0.523",
"0.7407"
],
[
"0.1013",
"0.733",
"0.4631",
"0.7344",
"0.4631",
"0.7485",
"0.1013",
"0.7476"
],
[
"0.54",
"0.742",
"0.91",
"0.742",
"0.91",
"0.7524",
"0.54",
"0.7524"
],
[
"0.10455",
"0.7485",
"0.1896",
"0.7485",
"0.1896",
"0.7603",
"0.10455",
"0.7603"
],
[
"0.5195",
"0.7524",
"0.9126",
"0.754",
"0.9126",
"0.768",
"0.5195",
"0.7676"
],
[
"0.5215",
"0.7676",
"0.9126",
"0.7666",
"0.9126",
"0.781",
"0.5215",
"0.782"
],
[
"0.21",
"0.781",
"0.2467",
"0.781",
"0.2467",
"0.7925",
"0.21",
"0.7925"
],
[
"0.2402",
"0.779",
"0.3545",
"0.78",
"0.3545",
"0.7944",
"0.2402",
"0.7935"
],
[
"0.523",
"0.782",
"0.9126",
"0.782",
"0.9126",
"0.797",
"0.523",
"0.797"
],
[
"0.08905",
"0.8003",
"0.4763",
"0.8003",
"0.4763",
"0.811",
"0.08905",
"0.811"
],
[
"0.523",
"0.797",
"0.9126",
"0.797",
"0.9126",
"0.81",
"0.523",
"0.81"
],
[
"0.08905",
"0.813",
"0.478",
"0.813",
"0.478",
"0.827",
"0.08905",
"0.827"
],
[
"0.5215",
"0.81",
"0.911",
"0.81",
"0.911",
"0.8237",
"0.5215",
"0.8237"
],
[
"0.08826",
"0.827",
"0.4763",
"0.827",
"0.4763",
"0.8403",
"0.08826",
"0.8403"
],
[
"0.5215",
"0.8237",
"0.911",
"0.8237",
"0.911",
"0.837",
"0.5215",
"0.837"
],
[
"0.08826",
"0.8413",
"0.478",
"0.8413",
"0.478",
"0.855",
"0.08826",
"0.855"
],
[
"0.523",
"0.8384",
"0.911",
"0.8384",
"0.911",
"0.8516",
"0.523",
"0.8516"
],
[
"0.5215",
"0.8516",
"0.91",
"0.8516",
"0.91",
"0.8657",
"0.5215",
"0.8657"
],
[
"0.0866",
"0.863",
"0.4756",
"0.863",
"0.4756",
"0.877",
"0.0866",
"0.877"
],
[
"0.08496",
"0.875",
"0.4478",
"0.8755",
"0.4478",
"0.89",
"0.08496",
"0.8896"
],
[
"0.54",
"0.8735",
"0.8613",
"0.8755",
"0.8613",
"0.89",
"0.54",
"0.888"
]
] | [
"AvgOut :",
"消除沉闷响应的简单输出概率度量",
"Tong Niu, Mohit Bansal",
"UNC Chapel Hill",
"[tongn, mbansal]@cs.unc.edu",
"Abstract",
"必须是“正确的”（连贯且相关），而且还需要多样化且信",
"息丰富。然而，许多以前的作品已经报道seq2seq具有",
"Many sequence-to-sequence dialogue models tend to",
"低语料库级多样性E(Lietal.2016a;Serbanetal.2016;",
"generate safe, uninformative responses. There have",
"Sordonietal.2015；VinyalsandLe2015），因为它倾向",
"been various useful efforts on trying to eliminate them.",
"However, these approaches either improve decoding al-",
"于产生安全，简洁和无信息的响应，例如“我不知道。”。",
"gorithms during inference, rely on hand-crafted fea-",
"这些反应不必要地使对话系统的交互性比应有的要少",
"tures, or employ complex models. In our work, we",
"得多",
"build dialogue models that are dynamically aware of",
"要增加对话回应的多样性，第一步是患实地评估回应",
"what utterances or tokens are dull without any feature",
"的多样性。以前的工作使用的指标与多样性相关，但并",
"engineering. Specifically, we start with a simple yet",
"不强烈，例如不同令牌的比率E(Lietal.2016a）和响应",
"effective automatic metric, AvgOut , which calculates",
"长度E（Bahetietal.2018）。然而，回应可能很长，但",
"the average outputprobability distribution of all time",
"在意义上非常无聊，例如“我确信我不知道它。“，或",
"steps on the decoder side during training. This met-",
"简短但有趣（i。e。，包含了很多信息），比如“爸爸很",
"ric directly estimates which tokens are more likely to",
"刻薄。”。仅由模型调查离散令牌输出也不理想，因为这",
"be generated, thus making it a faithful evaluation of",
"些令牌只是模型在每个时间步的输出概率分布的单个",
"the model diversity (i.e., for diverse models, the token",
"probabilities should be more evenly distributed rather",
"实现，这不可避免地丢失了由整个分布指示的有价值的",
"than peaked at a fewdull tokens):We then leverage",
"信息。Lietal.（2016b）手动收集沉闷响应的短列表",
"this novel metricto propose threemodelsthatpromote",
"并在训练期间阻止模型产生此类话语。然而，手工规则",
"的一个重要缺点是，",
"一组沉闷的令牌或话语是静态的",
"MinAvgOut,directly maximizes the diversity score",
"diversity without losing relevance. The first model",
"而实际上它通常在训练过程中演变：当当前的沉闷令牌",
"through the output distributions of each batch; the sec-",
"被消除时，另一组可能会显示出来",
"ond model, Label Fine-Tuning ( LFT ), prepends to",
"在我们的工作中，1我们从一种简单而有效的方法开",
"the source sequence a label continuously scaled by the",
"始，以测量响应的多样性。我们将此指标命名为“平均",
"diversity score to control the diversity level; the third",
"输出概率分布”或AvgOut，它直接从会话中训练模型",
"model, RL , adopts Reinforcement Learning and treats",
"the diversity score as a reward signal. Moreover, we",
"本身提取信息。我们通过在训练期间跟踪解码器侧所有",
"experiment with a hybrid model by combining the loss",
"输出概率分布的指数平均值来计算它。该指标动态地测",
"terms of MinAvgOut and RL . All four models outper-",
"量模型偏向于哪些令牌，而无需任何手工规则，从而使",
"form their base LSTM-RNN model on both diversity",
"其成为对模型多样性的忠实评估（i。e。，对于不同的模",
"and relevance by a large margin, and are comparable",
"型，令牌概率应该更均匀地分布，而不是在几个沉闷的",
"to or better than competitive baselines (also verified",
"令牌处达到峰值)。此外，由于AvgOut是一维分类分",
"via human evaluation). Moreover, our approaches are",
"布，而不是像熵一样的无量纲数值，因此它自然地承载",
"orthogonal to the base model, making them applicable",
"和传达了更多有关模型多样性的信息",
"as an add-on to other emerging better dialogue models",
"然后，我们提出了三入模型，利用我们的新指标来",
"in the future",
"促进对话生成的多样性。第一个MinAvgOut模型最",
"小化当前批次AvgOut的点积和跨批次的指数平均值",
"1",
"Introduction",
"AvgOut，这鼓励生成低频令牌。第二个LFT模型使",
"许多现代对话生成模型都使用序列到序列体系结构作",
"用标记的转导方法，并在训练过程中通过地面实况目标",
"为其主千E(Sordonietal.2015），在将其成功应用于机",
"序列的多样性得分来缩放“多样性标签”，而在测试过",
"器翻译（MT)E(Bahdanau,Cho,and Bengio 2015)之",
"程中，可以通过调整预期的多样性得分来生成不同多样",
"后。但是，对话任务也有与MT不同的要求：响应不仅",
"性水平的响应。第三个RL模型利用强化学习，我们",
"的新指标应用于离散令牌并作为奖励信号。此外，由于",
"Copyright 2020, Association for the Advancement of Ar",
"tificial Intelligence (www.aaai.org). All rights reserved.",
"I We will release all our code and model outputs."
] | {
"arxiv_id": "2001.05467",
"language": "zh",
"page_idx": 1
} |
||
2001.05467-zh-4 | 2001.05467-zh-4.jpg | <FigureHere>
Figure 6: Diversity-32 graphs of all models. Curves with
$\bar{\text{lower AUC correspond to more diverse models.}}$
<FigureHere>
Figure 4: RL model: the diversity score of the sampled
response is fed back to the model as reward signal.
<FigureHere>
Figure 5: Diversity score calculation for a discrete se
quence: $B_d= 1. 0- ( 0. 05+ 0. 2+ 0. 1+ 0. 01) / 4$ .
4 Experimental Setup
Dataset and Task
4.1
我们使用面向任务的 Ubuntu 对话数据集$\boxed{\mathrm{F}}($Lowe et al. 2015), 因为它不仅具有 F1 指标来评估响应的相关性，而且其中的对话也是开放式的，以允许足够的空间用于多样性。我们也选择了这个数据集，因为以前的工作，例如。g。, HRED (Serban et al.2016) 和VHRED (Serban et al. 2017b) 都使用 Ubuntu 展示了他们的多样性促进模型。由于该数据集的普及，我们能够在同一数据集上复制几乎所有模型，并对其消除迟钝的有效性进行有意义的比较。作为未来的工作，我们计划将我们的模型应用于需要多样性的其他数据集
# 4.2 Automatic Evaluation
为了衡量模型响应的相关性，我们遵循 Serban et al.( 2017a ),并对两种活动的 F1 进行评估 (技术动词，e。g。“上传”、“安装”)和实体(技术名词，e。g。“根”、“互联网”)。F1 是通过将地面实况和模型响应映射到其相应的活动实体表示形式来计算的 (Serban et al.2017a) ,他们认为 F1“特别适合面向目标的 Ubuntu 对话语料库”。我们没有对 BLEU 得分E(Papineni et al.2002)进行评估，因为$\mathbb{E}($Liu et al.2016)表明 BLEU 与对话质量没有很好的相关性。Lowe et al.( 2017 )也对 BLEU 做了类似的观察。为了评估多样性，我们采用了以前工作中的两个评估指标，即 Distinct-1 和 Distinct-2 E (Li et al. 2016a)。这些分别是 unigram 和 bigram 的唯一令牌和所有令牌的数量之间的比率。此外，我们提出了一种新颖的多样性图及其相应的度量，我们分别将其命名为 Diversity- 32 和 Diversity-AUC 。我们收集
句子，unigram, bigram 和 trigram 的统计数据，并从最高到最低对其归一化频率进行排序。观察到所有四个图都遵循长尾分布，我们只保留最高的 32 频率并绘制它们。$^{4}$然后，我们为每个图计算一个减去曲线下面积(Diversity-AUC),从而绘制出模型的多样性的高级图片
4.3
## Human Evaluation
尽管我们提出了有效的 AvgOut 度量，但我们确实发现该模型有时仍然会作弊以获得更高的自动多样性得分。例如，从选定的输出示例(第 5节)中可以看出，该模型倾向于生成带有错字的单词，因为与正确的对应单词相比，这些单词是罕见的标记。对于像 Ubuntu 这样的嘈杂数据集来说，这是不可避免的。因此，如果没有人类的评估，我们永远无法确定我们的模型是好的，还是它们看起来很好，因为我们的指标被利用了。。
$^{4}$One can also pick any reasonable number around 32 without loss of generality. Also note that we do not take into account what specific token each frequency corresponds to, because for example, although“plethora”is an unusual word, a model that outputs it all the time is still boring. Thus the frequencies are more important than what the tokens actually are.
$^{5}$This is also true for Ubuntu: the human studies and F-1 automatic evaluation are complementary to each other because MTurk annotators are better at judging how coherent and informative a response is, while F1s concentrate more on the technical terms (i.e., whether important activities or entities are present in the response). | [
[
"0.527",
"0.0808",
"0.544",
"0.0808",
"0.5474",
"0.1351",
"0.5293",
"0.1351"
],
[
"0.649",
"0.0852",
"0.664",
"0.0852",
"0.664",
"0.09534",
"0.649",
"0.09534"
],
[
"0.738",
"0.0833",
"0.749",
"0.0833",
"0.749",
"0.09344",
"0.738",
"0.09344"
],
[
"0.2632",
"0.0903",
"0.2712",
"0.0903",
"0.2712",
"0.0947",
"0.2632",
"0.0947"
],
[
"0.6323",
"0.09534",
"0.6406",
"0.09216",
"0.6445",
"0.09973",
"0.636",
"0.1029"
],
[
"0.7344",
"0.0903",
"0.7485",
"0.0903",
"0.7485",
"0.1218",
"0.7344",
"0.1218"
],
[
"0.6465",
"0.0985",
"0.6567",
"0.0985",
"0.6567",
"0.1029",
"0.6465",
"0.1029"
],
[
"0.738",
"0.1187",
"0.7485",
"0.1187",
"0.7485",
"0.1276",
"0.738",
"0.1276"
],
[
"0.5317",
"0.132",
"0.5435",
"0.132",
"0.5435",
"0.1414",
"0.5317",
"0.1414"
],
[
"0.734",
"0.1276",
"0.746",
"0.125",
"0.751",
"0.1351",
"0.738",
"0.1383"
],
[
"0.4248",
"0.1395",
"0.4495",
"0.1395",
"0.4495",
"0.1465",
"0.4248",
"0.1465"
],
[
"0.739",
"0.1383",
"0.7466",
"0.1383",
"0.7466",
"0.1433",
"0.739",
"0.1433"
],
[
"0.2632",
"0.1528",
"0.2712",
"0.1528",
"0.2712",
"0.1622",
"0.2632",
"0.1622"
],
[
"0.304",
"0.154",
"0.3105",
"0.154",
"0.3105",
"0.1604",
"0.304",
"0.1604"
],
[
"0.3423",
"0.1547",
"0.3506",
"0.1515",
"0.3555",
"0.159",
"0.3464",
"0.1622"
],
[
"0.3823",
"0.1528",
"0.4028",
"0.1528",
"0.4028",
"0.1635",
"0.3823",
"0.1635"
],
[
"0.11926",
"0.1635",
"0.1683",
"0.1635",
"0.1683",
"0.1704",
"0.11926",
"0.1704"
],
[
"0.2532",
"0.1642",
"0.2754",
"0.1642",
"0.2754",
"0.1724",
"0.2532",
"0.1724"
],
[
"0.339",
"0.1654",
"0.3586",
"0.1654",
"0.3586",
"0.1724",
"0.339",
"0.1724"
],
[
"0.3757",
"0.1635",
"0.4019",
"0.1635",
"0.4019",
"0.1749",
"0.3757",
"0.1749"
],
[
"0.5537",
"0.1686",
"0.6772",
"0.1686",
"0.6772",
"0.1818",
"0.5537",
"0.1818"
],
[
"0.7754",
"0.1666",
"0.863",
"0.1686",
"0.863",
"0.1831",
"0.7754",
"0.1812"
],
[
"0.2532",
"0.1768",
"0.2769",
"0.1768",
"0.2769",
"0.1843",
"0.2532",
"0.1843"
],
[
"0.3464",
"0.178",
"0.357",
"0.178",
"0.357",
"0.1818",
"0.3464",
"0.1818"
],
[
"0.3896",
"0.178",
"0.398",
"0.178",
"0.398",
"0.1831",
"0.3896",
"0.1831"
],
[
"0.0866",
"0.1963",
"0.4795",
"0.1963",
"0.4795",
"0.2108",
"0.0866",
"0.2108"
],
[
"0.0866",
"0.2108",
"0.4495",
"0.2108",
"0.4495",
"0.2247",
"0.0866",
"0.2247"
],
[
"0.10297",
"0.2412",
"0.1584",
"0.2437",
"0.1584",
"0.2551",
"0.10297",
"0.2524"
],
[
"0.10455",
"0.2551",
"0.1365",
"0.2551",
"0.1365",
"0.2676",
"0.10455",
"0.2676"
],
[
"0.2296",
"0.2903",
"0.2507",
"0.2903",
"0.2507",
"0.303",
"0.2296",
"0.303"
],
[
"0.6787",
"0.2854",
"0.755",
"0.2854",
"0.755",
"0.2998",
"0.6787",
"0.2998"
],
[
"0.10455",
"0.2966",
"0.1323",
"0.2966",
"0.1323",
"0.3093",
"0.10455",
"0.3093"
],
[
"0.2876",
"0.334",
"0.3057",
"0.334",
"0.3057",
"0.3435",
"0.2876",
"0.3435"
],
[
"0.10455",
"0.3403",
"0.134",
"0.3403",
"0.134",
"0.3528",
"0.10455",
"0.3528"
],
[
"0.1691",
"0.3542",
"0.1969",
"0.3542",
"0.1969",
"0.3662",
"0.1691",
"0.3662"
],
[
"0.3406",
"0.3757",
"0.367",
"0.3757",
"0.367",
"0.387",
"0.3406",
"0.387"
],
[
"0.10455",
"0.3833",
"0.1274",
"0.3833",
"0.1274",
"0.3958",
"0.10455",
"0.3958"
],
[
"0.4084",
"0.3833",
"0.464",
"0.3833",
"0.464",
"0.3933",
"0.4084",
"0.3933"
],
[
"0.1724",
"0.3901",
"0.1953",
"0.3901",
"0.1953",
"0.3984",
"0.1724",
"0.3984"
],
[
"0.2607",
"0.3901",
"0.3137",
"0.3901",
"0.3137",
"0.3977",
"0.2607",
"0.3977"
],
[
"0.3218",
"0.3901",
"0.3733",
"0.3901",
"0.3733",
"0.3977",
"0.3218",
"0.3977"
],
[
"0.2361",
"0.3914",
"0.2484",
"0.3914",
"0.2484",
"0.3977",
"0.2361",
"0.3977"
],
[
"0.2091",
"0.3926",
"0.2173",
"0.3926",
"0.2173",
"0.3965",
"0.2091",
"0.3965"
],
[
"0.6772",
"0.4028",
"0.7563",
"0.4028",
"0.7563",
"0.414",
"0.6772",
"0.414"
],
[
"0.0866",
"0.4153",
"0.4763",
"0.4153",
"0.4763",
"0.4287",
"0.0866",
"0.4287"
],
[
"0.08417",
"0.43",
"0.4265",
"0.4287",
"0.4265",
"0.443",
"0.08417",
"0.4443"
],
[
"0.5186",
"0.4255",
"0.914",
"0.4268",
"0.914",
"0.4414",
"0.5186",
"0.44"
],
[
"0.5195",
"0.44",
"0.834",
"0.44",
"0.834",
"0.4539",
"0.5195",
"0.4539"
],
[
"0.1805",
"0.4717",
"0.2141",
"0.4717",
"0.2141",
"0.483",
"0.1805",
"0.483"
],
[
"0.2115",
"0.4702",
"0.384",
"0.4717",
"0.384",
"0.486",
"0.2115",
"0.4849"
],
[
"0.5215",
"0.481",
"0.91",
"0.4817",
"0.91",
"0.4968",
"0.5215",
"0.4956"
],
[
"0.08905",
"0.4937",
"0.1136",
"0.4937",
"0.1136",
"0.502",
"0.08905",
"0.502"
],
[
"0.12335",
"0.4956",
"0.1299",
"0.4956",
"0.1299",
"0.5",
"0.12335",
"0.5"
],
[
"0.1283",
"0.4912",
"0.2737",
"0.4912",
"0.2737",
"0.505",
"0.1283",
"0.505"
],
[
"0.5215",
"0.4968",
"0.911",
"0.4968",
"0.911",
"0.5103",
"0.5215",
"0.5103"
],
[
"0.0866",
"0.511",
"0.478",
"0.511",
"0.478",
"0.5244",
"0.0866",
"0.5244"
],
[
"0.5215",
"0.5103",
"0.911",
"0.5103",
"0.911",
"0.5234",
"0.5215",
"0.5234"
],
[
"0.0866",
"0.5244",
"0.478",
"0.5244",
"0.478",
"0.539",
"0.0866",
"0.539"
],
[
"0.5215",
"0.5234",
"0.911",
"0.5234",
"0.911",
"0.5386",
"0.5215",
"0.5386"
],
[
"0.08905",
"0.5405",
"0.4763",
"0.5405",
"0.4763",
"0.5503",
"0.08905",
"0.5503"
],
[
"0.5215",
"0.5386",
"0.911",
"0.5386",
"0.911",
"0.552",
"0.5215",
"0.552"
],
[
"0.08826",
"0.553",
"0.4763",
"0.553",
"0.4763",
"0.5664",
"0.08826",
"0.5664"
],
[
"0.5215",
"0.552",
"0.54",
"0.552",
"0.54",
"0.565",
"0.5215",
"0.565"
],
[
"0.08826",
"0.5664",
"0.478",
"0.5664",
"0.478",
"0.5806",
"0.08826",
"0.5806"
],
[
"0.0866",
"0.5776",
"0.4763",
"0.58",
"0.4763",
"0.5947",
"0.0866",
"0.5923"
],
[
"0.523",
"0.5806",
"0.568",
"0.5806",
"0.568",
"0.5923",
"0.523",
"0.5923"
],
[
"0.562",
"0.579",
"0.7173",
"0.58",
"0.7173",
"0.5947",
"0.562",
"0.5933"
],
[
"0.08905",
"0.5933",
"0.478",
"0.5933",
"0.478",
"0.607",
"0.08905",
"0.607"
],
[
"0.523",
"0.5996",
"0.9126",
"0.5996",
"0.9126",
"0.613",
"0.523",
"0.613"
],
[
"0.0866",
"0.608",
"0.4763",
"0.608",
"0.4763",
"0.622",
"0.0866",
"0.622"
],
[
"0.5215",
"0.6123",
"0.91",
"0.613",
"0.91",
"0.6274",
"0.5215",
"0.627"
],
[
"0.0866",
"0.622",
"0.4763",
"0.622",
"0.4763",
"0.6353",
"0.0866",
"0.6353"
],
[
"0.523",
"0.6274",
"0.911",
"0.6274",
"0.911",
"0.6416",
"0.523",
"0.6416"
],
[
"0.08496",
"0.634",
"0.429",
"0.634",
"0.429",
"0.648",
"0.08496",
"0.648"
],
[
"0.524",
"0.643",
"0.911",
"0.643",
"0.911",
"0.654",
"0.524",
"0.654"
],
[
"0.5215",
"0.655",
"0.911",
"0.655",
"0.911",
"0.6694",
"0.5215",
"0.6694"
],
[
"0.08826",
"0.662",
"0.1152",
"0.662",
"0.1152",
"0.674",
"0.08826",
"0.674"
],
[
"0.10864",
"0.662",
"0.308",
"0.662",
"0.308",
"0.6753",
"0.10864",
"0.6753"
],
[
"0.5195",
"0.6675",
"0.911",
"0.6685",
"0.911",
"0.683",
"0.5195",
"0.6816"
],
[
"0.08826",
"0.681",
"0.478",
"0.681",
"0.478",
"0.6943",
"0.08826",
"0.6943"
],
[
"0.5215",
"0.6816",
"0.9126",
"0.6816",
"0.9126",
"0.696",
"0.5215",
"0.696"
],
[
"0.0866",
"0.696",
"0.4795",
"0.696",
"0.4795",
"0.709",
"0.0866",
"0.709"
],
[
"0.5215",
"0.6963",
"0.851",
"0.6963",
"0.851",
"0.7104",
"0.5215",
"0.7104"
],
[
"0.0866",
"0.707",
"0.4795",
"0.707",
"0.4795",
"0.7246",
"0.0866",
"0.7246"
],
[
"0.08826",
"0.7227",
"0.478",
"0.7227",
"0.478",
"0.7363",
"0.08826",
"0.7363"
],
[
"0.539",
"0.7227",
"0.914",
"0.7227",
"0.914",
"0.7363",
"0.539",
"0.7363"
],
[
"0.08826",
"0.7373",
"0.4763",
"0.7373",
"0.4763",
"0.7505",
"0.08826",
"0.7505"
],
[
"0.5215",
"0.735",
"0.911",
"0.7363",
"0.911",
"0.75",
"0.5215",
"0.7485"
],
[
"0.0989",
"0.7505",
"0.478",
"0.7505",
"0.478",
"0.7646",
"0.0989",
"0.7646"
],
[
"0.5215",
"0.7485",
"0.911",
"0.7485",
"0.911",
"0.762",
"0.5215",
"0.762"
],
[
"0.08826",
"0.7646",
"0.478",
"0.7646",
"0.478",
"0.778",
"0.08826",
"0.778"
],
[
"0.5215",
"0.761",
"0.9126",
"0.761",
"0.9126",
"0.7744",
"0.5215",
"0.7744"
],
[
"0.08826",
"0.779",
"0.478",
"0.779",
"0.478",
"0.7925",
"0.08826",
"0.7925"
],
[
"0.5195",
"0.7734",
"0.9126",
"0.7744",
"0.9126",
"0.7886",
"0.5195",
"0.787"
],
[
"0.08905",
"0.7915",
"0.478",
"0.7915",
"0.478",
"0.8047",
"0.08905",
"0.8047"
],
[
"0.523",
"0.7886",
"0.911",
"0.7886",
"0.911",
"0.799",
"0.523",
"0.799"
],
[
"0.5195",
"0.8",
"0.638",
"0.8",
"0.638",
"0.813",
"0.5195",
"0.813"
],
[
"0.0866",
"0.806",
"0.4763",
"0.806",
"0.4763",
"0.8193",
"0.0866",
"0.8193"
],
[
"0.539",
"0.8125",
"0.911",
"0.8125",
"0.911",
"0.8257",
"0.539",
"0.8257"
],
[
"0.08905",
"0.8193",
"0.478",
"0.8193",
"0.478",
"0.8335",
"0.08905",
"0.8335"
],
[
"0.5195",
"0.8257",
"0.911",
"0.8257",
"0.911",
"0.839",
"0.5195",
"0.839"
],
[
"0.08905",
"0.8354",
"0.4763",
"0.8354",
"0.4763",
"0.845",
"0.08905",
"0.845"
],
[
"0.5195",
"0.8384",
"0.9126",
"0.8384",
"0.9126",
"0.8516",
"0.5195",
"0.8516"
],
[
"0.08826",
"0.8477",
"0.4763",
"0.8477",
"0.4763",
"0.8613",
"0.08826",
"0.8613"
],
[
"0.5195",
"0.851",
"0.9126",
"0.851",
"0.9126",
"0.8643",
"0.5195",
"0.8643"
],
[
"0.08826",
"0.8623",
"0.4763",
"0.8623",
"0.4763",
"0.8755",
"0.08826",
"0.8755"
],
[
"0.5195",
"0.863",
"0.914",
"0.863",
"0.914",
"0.877",
"0.5195",
"0.877"
],
[
"0.0866",
"0.8755",
"0.478",
"0.8755",
"0.478",
"0.8896",
"0.0866",
"0.8896"
],
[
"0.5215",
"0.8774",
"0.759",
"0.8774",
"0.759",
"0.888",
"0.5215",
"0.888"
]
] | [
"电##店",
"War",
"0.07",
"P",
"1",
"00",
"",
"0.01",
"0.02",
"8or",
"AvgOut",
"1.03",
"",
"H",
"t",
"<epd>",
"SourceTokens",
"",
"",
"心",
"(a) Sentence-Level",
"(b) Unigram",
"<start>",
"L",
"H",
"Figure 4: RL model: the diversity score of the sampled",
"response is fed back to the model as reward signal.",
"Probability",
"0.3",
"0.2",
"(c) Bigram",
"0.2",
"0.1",
"0.1",
"0.05",
"0.01",
"0.0",
"Sequence",
"she",
"+an",
"+acrobat",
"is",
"+",
"(d) Trigram",
"Figure 5: Diversity score calculation for a discrete se-",
"quence: Bd = 1.0- (0.05 +0.2 + 0.1 +0.01)/4 .",
"Figure 6: Diversity-32 graphs of all models. Curves with",
"lower AUC correspond to more diverse models.",
"4",
"ExperimentalSetup",
"句子，unigram，bigram和trigram的统计数据，并从",
"41",
"",
"Dataset and Task",
"最高到最低对其归一化频率进行排序。观察到所有四个",
"我们使用面向任务的Ubuntu对话数据集ELoweet",
"图都遵循长尾分布，我们只保留最高的32频率并绘制",
"al.2015），因为它不仅具有F1指标来评估响应的相",
"它们。4然后，我们为每个图计算一个减去曲线下面积",
"关性，而且其中的对话也是开放式的：以允许足够的",
"（Diversity-AUC），从而绘制出模型的多样性的高级图",
"空间用于多样性。我们也选择了这个数据集，因为以",
"片",
"前的工作，例如。g。，HRED（Serban et al.2016）和",
"VHRED（Serbanetal.2017b）都使用Ubuntu展示了",
"4.3",
"Human Evaluation",
"他们的多样性促进模型。由于该数据集的普及，我们能",
"尽管我们提出了有效的AvgOut度量，但我们确实发现",
"够在同一数据集上复制几乎所有模型，并对其消除迟钝",
"该模型有时仍然会作以获得更高的自动多样性得分",
"的有效性进行有意义的比较。作为未来的工作，我们计",
"例如，从选定的输出示例（第5节）中可以看出，该模",
"划将我们的模型应用于需要多样性的其他数据集",
"型倾向于生成带有错字的单词，因为与正确的对应单词",
"相比，这些单词是罕见的标记。对于像Ubuntu这样的",
"4.2",
"AutomaticEvaluation",
"噜杂数据集来说，这是不可避免的。因此，如果没有人",
"为了衡量模型响应的相关性，我们遵循Serbanetal.（",
"类的评估，我们永远无法确定我们的模型是好的，还是",
"2017a），并对两种活动的F1进行评估（技术动词，e。",
"它们看起来很好，因为我们的指标被利用了。。",
"g。、“上传”、“安装”）和实体（技术名词，e。g。“根”“互",
"联网”）。F1是通过将地面实况和模型响应映射到其相",
"4One can also pick any reasonable number around 32",
"应的活动实体表示形式来计算的(Serbanetal.2017a）",
"without loss of generality. Also note that we do not take into",
"他们认为F1“特别适合面向目标的Ubuntu对话语料",
"account what specific token each frequency corresponds to",
"库\"。我们没有对BLEU得分E(Papinenietal.2002）进",
"because for example, although \" plethora \" is an unusual",
"行评估，因为E(Liu etal.2016)表明BLEU与对话质",
"word, a model that outputs it all the time is still boring.",
"量没有很好的相关性。Loweetal.（2017）也对BLEU",
"Thus the frequencies are more important than what the to-",
"kens actually are.",
"做了类似的观察。为了评估多样性，我们采用了以前",
"5This is also true for Ubuntu: the human studies and F-1",
"工作中的两个评估指标，即Distinct-1和Distinct-2E",
"automatic evaluation are complementary to each other be",
"(Lietal.2016a）。这些分别是unigram和bigram的",
"cause MTurk annotators are better at judging how coherent",
"唯一令牌和所有令牌的数量之间的比率。此外，我们提",
"and informative a response is, while Fls concentrate more",
"出了一种新颖的多样性图及其相应的度量，我们分别将",
"on the technical terms (i.e., whether important activities or",
"其命名为Diversity-32和Diversity-AUC。我们收集",
"entities are present in the response)"
] | {
"arxiv_id": "2001.05467",
"language": "zh",
"page_idx": 4
} |
||
2104.14444-zh-0 | 2104.14444-zh-0.jpg | LMU- ASC 11/ 21
MPP- 2021- 67
# 使用对称控制神经网络改进仿真
$^{2}\text{ Max- Planck- Institut f\"{u} r Physik}$
Föhringer Ring 6
80805 Munich, Germany
Marc Syvaeri$^1,2$ & Sven Krippendorf$^{1}$
$^{1}$ Arnold-Sommerfeld-Center
Ludwig-Maximilians-Universität
Theresienstr. 37
80333 Munich, Germany
$\{$ marc.syvaeri,sven.krippendorf $\}$ @physik.uni-muenchen.de
ABSTRACT
The dynamics of physical systems is often constrained to lower dimensional sub-spaces due to the presence of conserved quantities. Here we propose a method to learn and exploit such symmetry constraints building upon Hamiltonian Neural Networks. By enforcing cyclic coordinates with appropriate loss functions, we find that we can achieve improved accuracy on simple classical dynamics tasks. By fitting analytic formulae to the latent variables in our network we recover that our networks are utilizing conserved quantities such as (angular) momentum.
1
INTRODUCTION
为了进行精确的模拟，偏向深度神经网络的构建是实现卓越性能的关键机制。这种偏差的一个例子是学习在哈密顿神经网络 (HNN)E(Greydanus et al.,2019)的上下文中执行的节能运动。更一般地说，可能的动力学由于系统的对称性而受到限制，并且它们发生在可用相空间的子空间上 (例如。g。恒定能量和总角动量的表面)。尽管能量已被构建到神经网络的功能偏置中，但此时需要手动执行进一步的对称性，这需要系统的领域知识。在这里，我们描述了一种如何在用于模拟的神经网络中自动搜索和使用这种对称性的方法。此外，该方法还揭示了-在简单系统的此阶段-守恒量的解析表达式。我们的方法基于强制网络在第一个网络$T_\psi$中执行规范坐标变换，该网络强制某些潜在坐标变为常数 (哈密顿动力学语言中的循环坐标)。根据这些新坐标，网络然后使用第二个网络预测该系统动力学的哈密顿量$H_\phi.$ 我们网络的结构-稍后将更准确地描述-可以在图中找到$\overline{\mathrm{E}}1$ 。
<FigureHere>
Figure 1: Structure of symmetry control networks: The network $T_\psi$ transforms the input coordinates of phase space $(\mathbf{p},\mathbf{q})$ to canonical coordinates where some coordinates are forced to be cyclic. These coordinates are used as input to the Hamiltonian network $H_\phi.$ The output, the time derivatives of our initial coordinates, is calculated from the Hamiltonian using auto-differentiation.
我们在实验中证明，使用此方法确实可以自动识别多个守恒量，与 HNN 相比，我们发现动
力学预测有所改进，并且可以找到守恒量的解析表达式
1 | [
[
"0.709",
"0.0625",
"0.824",
"0.0625",
"0.824",
"0.07574",
"0.709",
"0.07574"
],
[
"0.732",
"0.07764",
"0.824",
"0.07764",
"0.824",
"0.0884",
"0.732",
"0.0884"
],
[
"0.1772",
"0.11743",
"0.2893",
"0.1193",
"0.2893",
"0.1401",
"0.1772",
"0.1383"
],
[
"0.1765",
"0.1433",
"0.4575",
"0.1433",
"0.4575",
"0.1642",
"0.1765",
"0.1642"
],
[
"0.1847",
"0.1925",
"0.4902",
"0.1925",
"0.4902",
"0.207",
"0.1847",
"0.207"
],
[
"0.183",
"0.209",
"0.3896",
"0.2102",
"0.3896",
"0.2235",
"0.183",
"0.2229"
],
[
"0.492",
"0.2102",
"0.7305",
"0.2102",
"0.7305",
"0.2235",
"0.492",
"0.2235"
],
[
"0.1847",
"0.2247",
"0.415",
"0.2247",
"0.415",
"0.238",
"0.1847",
"0.238"
],
[
"0.4927",
"0.2247",
"0.6177",
"0.2247",
"0.6177",
"0.238",
"0.4927",
"0.238"
],
[
"0.1854",
"0.238",
"0.3022",
"0.238",
"0.3022",
"0.252",
"0.1854",
"0.252"
],
[
"0.4902",
"0.2374",
"0.669",
"0.2393",
"0.669",
"0.2524",
"0.4902",
"0.2507"
],
[
"0.1847",
"0.2524",
"0.3594",
"0.2524",
"0.3594",
"0.2664",
"0.1847",
"0.2664"
],
[
"0.1936",
"0.2708",
"0.683",
"0.2708",
"0.683",
"0.2842",
"0.1936",
"0.2842"
],
[
"0.452",
"0.3193",
"0.5474",
"0.3213",
"0.5474",
"0.336",
"0.452",
"0.334"
],
[
"0.2336",
"0.358",
"0.7656",
"0.358",
"0.7656",
"0.3718",
"0.2336",
"0.3718"
],
[
"0.2336",
"0.3718",
"0.7656",
"0.3718",
"0.7656",
"0.385",
"0.2336",
"0.385"
],
[
"0.232",
"0.3838",
"0.7656",
"0.385",
"0.7656",
"0.3997",
"0.232",
"0.3984"
],
[
"0.2336",
"0.3997",
"0.763",
"0.3997",
"0.763",
"0.4136",
"0.2336",
"0.4136"
],
[
"0.232",
"0.4124",
"0.7646",
"0.4136",
"0.7646",
"0.428",
"0.232",
"0.4268"
],
[
"0.2336",
"0.4268",
"0.763",
"0.4268",
"0.763",
"0.44",
"0.2336",
"0.44"
],
[
"0.232",
"0.4395",
"0.7656",
"0.4414",
"0.7656",
"0.455",
"0.232",
"0.4526"
],
[
"0.2336",
"0.455",
"0.5986",
"0.455",
"0.5986",
"0.4685",
"0.2336",
"0.4685"
],
[
"0.1765",
"0.5103",
"0.1887",
"0.5103",
"0.1887",
"0.521",
"0.1765",
"0.521"
],
[
"0.2115",
"0.509",
"0.3472",
"0.509",
"0.3472",
"0.523",
"0.2115",
"0.523"
],
[
"0.1765",
"0.5444",
"0.8228",
"0.5444",
"0.8228",
"0.558",
"0.1765",
"0.558"
],
[
"0.1896",
"0.558",
"0.8228",
"0.558",
"0.8228",
"0.5713",
"0.1896",
"0.5713"
],
[
"0.1765",
"0.5713",
"0.8228",
"0.5713",
"0.8228",
"0.5854",
"0.1765",
"0.5854"
],
[
"0.1765",
"0.586",
"0.8228",
"0.586",
"0.8228",
"0.5996",
"0.1765",
"0.5996"
],
[
"0.1765",
"0.5996",
"0.8228",
"0.5996",
"0.8228",
"0.613",
"0.1765",
"0.613"
],
[
"0.1748",
"0.613",
"0.822",
"0.613",
"0.822",
"0.627",
"0.1748",
"0.627"
],
[
"0.1765",
"0.627",
"0.822",
"0.627",
"0.822",
"0.64",
"0.1765",
"0.64"
],
[
"0.1748",
"0.64",
"0.822",
"0.6396",
"0.822",
"0.654",
"0.1748",
"0.655"
],
[
"0.1748",
"0.654",
"0.8203",
"0.654",
"0.8203",
"0.6675",
"0.1748",
"0.6675"
],
[
"0.1772",
"0.672",
"0.6006",
"0.672",
"0.6006",
"0.6816",
"0.1772",
"0.6816"
],
[
"0.4192",
"0.709",
"0.5024",
"0.711",
"0.5024",
"0.729",
"0.4192",
"0.7266"
],
[
"0.7344",
"0.726",
"0.7715",
"0.731",
"0.767",
"0.7456",
"0.7305",
"0.7407"
],
[
"0.2769",
"0.733",
"0.38",
"0.733",
"0.38",
"0.7476",
"0.2769",
"0.7476"
],
[
"0.535",
"0.7344",
"0.661",
"0.7344",
"0.661",
"0.7485",
"0.535",
"0.7485"
],
[
"0.79",
"0.729",
"0.8145",
"0.732",
"0.812",
"0.741",
"0.7876",
"0.738"
],
[
"0.1912",
"0.738",
"0.2222",
"0.738",
"0.2222",
"0.75",
"0.1912",
"0.75"
],
[
"0.4272",
"0.735",
"0.5",
"0.7373",
"0.5",
"0.7524",
"0.4272",
"0.7505"
],
[
"0.705",
"0.7407",
"0.714",
"0.7407",
"0.714",
"0.7456",
"0.705",
"0.7456"
],
[
"0.732",
"0.7407",
"0.785",
"0.7407",
"0.785",
"0.7695",
"0.732",
"0.7695"
],
[
"0.7607",
"0.7393",
"0.7793",
"0.7393",
"0.7793",
"0.747",
"0.7607",
"0.747"
],
[
"0.2656",
"0.747",
"0.3882",
"0.747",
"0.3882",
"0.757",
"0.2656",
"0.757"
],
[
"0.7935",
"0.7437",
"0.8105",
"0.7485",
"0.806",
"0.7583",
"0.789",
"0.754"
],
[
"0.5474",
"0.7485",
"0.6562",
"0.7485",
"0.6562",
"0.759",
"0.5474",
"0.759"
],
[
"0.1854",
"0.7485",
"0.2255",
"0.7485",
"0.2255",
"0.7603",
"0.1854",
"0.7603"
],
[
"0.4346",
"0.757",
"0.4795",
"0.7646",
"0.4756",
"0.779",
"0.4307",
"0.7715"
],
[
"0.1748",
"0.797",
"0.8228",
"0.797",
"0.8228",
"0.811",
"0.1748",
"0.811"
],
[
"0.1765",
"0.811",
"0.8228",
"0.811",
"0.8228",
"0.8257",
"0.1765",
"0.8257"
],
[
"0.1748",
"0.8257",
"0.824",
"0.8257",
"0.824",
"0.839",
"0.1748",
"0.839"
],
[
"0.1748",
"0.839",
"0.824",
"0.839",
"0.824",
"0.853",
"0.1748",
"0.853"
],
[
"0.1732",
"0.853",
"0.3628",
"0.8516",
"0.3628",
"0.866",
"0.1732",
"0.867"
],
[
"0.1765",
"0.8965",
"0.8228",
"0.8965",
"0.8228",
"0.91",
"0.1765",
"0.91"
],
[
"0.1765",
"0.911",
"0.569",
"0.911",
"0.569",
"0.925",
"0.1765",
"0.925"
],
[
"0.4927",
"0.9487",
"0.5063",
"0.9487",
"0.5063",
"0.9604",
"0.4927",
"0.9604"
]
] | [
"LMU-ASC 11/21",
"MPP-2021-67",
"使用对称",
"控制神经网络改进仿真",
"Marc Syvaeril,2& Sven Krippendorfl",
"1Arnold-Sommerfeld-Center",
"2 Max-Planck-Institut fur Physik",
"Ludwig-Maximilians-Universitat",
"Fohringer Ring 6",
"Theresienstr. 37",
"80805Munich,Germany",
"80333 Munich, Germany",
"{marc.syvaeri,sven.krippendorf } @physik.uni-muenchen.de",
"ABSTRACT",
"The dynamics of physical systems is often constrained to lower dimensional",
"sub-spaces due to the presence of conserved quantities. Here we propose",
"a method to learn and exploit such symmetry constraints building upon",
"Hamiltonian Neural Networks. By enforcing cyclic coordinates with ap-",
"propriate loss functions, we find that we can achieve improved accuracy",
"on simple classical dynamics tasks. By fitting analytic formulae to the la-",
"tent variables in our network we recover that our networks are utilizing",
"conserved quantities such as (angular) momentum.",
"1",
"INTRODUCTION",
"为了进行精确的模拟，偏向深度神经网络的构建是实现卓越性能的关键机制。这种偏差的",
"个例子是学习在哈密顿神经网络（HNN)E（Greydanusetal.，2019）的上下文中执行的节",
"能运动。更一般地说，可能的动力学由于系统的对称性而受到限制，并且它们发生在可用相",
"空间的子空间上（例如。g。恒定能量和总角动量的表面）。尽管能量已被构建到神经网络的",
"功能偏置中，但此时需要手动执行进一步的对称性，这需要系统的领域知识。在这里，我们",
"描述了一种如何在用于模拟的神经网络中自动搜索和使用这种对称性的方法。此外，该方",
"法还揭示了-在简单系统的此阶段-守恒量的解析表达式。我们的方法基于强制网络在第一个",
"网络T中执行规范坐标变换，该网络强制某些潜在坐标变为常数（哈密顿动力学语言中的",
"循环坐标）。根据这些新坐标，网络然后使用第二个网络预测该系统动力学的哈密顿量H。",
"我们网络的结构-稍后将更准确地描述-可以在图中找到1。",
"Peyeli = const,",
"oMe",
"T,(p, q): Canonical",
"H g(Peyclie, Pother, Qother",
"an",
"p,q",
"Pother, Qother",
"0:",
"oupa",
".q:",
"Transformation Network",
"op",
"HamiltonianNetwork",
"(Input)",
"Qeyclic",
"Figure 1: Structure of symmetry control networks: The network T transforms the input",
"coordinates of phase space (p, q) to canonical coordinates where some coordinates are forced",
"to be cyclic. These coordinates are used as input to the Hamiltonian network H.The",
"output, the time derivatives of our initial coordinates, is calculated from the Hamiltonian",
"using auto-differentiation.",
"我们在实验中证明，使用此方法确实可以自动识别多个守恒量，与HNN相比，我们发现动",
"力学预测有所改进，并且可以找到守恒量的解析表达式",
"1"
] | {
"arxiv_id": "2104.14444",
"language": "zh",
"page_idx": 0
} |
||
2104.14444-zh-1 | 2104.14444-zh-1.jpg | 2 THEORY
当运动发生在相空间的子空间上时，可以减少描述系统的坐标数量。每个独立的守恒量都会减少描述系统所需的坐标数，直到所有坐标都受到约束(可积系统)。即使我们不知道先验的守恒坐标，哈密顿力学也为这种坐标提供了一般条件。特别是相空间上任何函数的时间导数，由位置$\mathbf{q}$和动量 p 的$N$粒子在$d$空间维度上的 p 不变的$\mathbf{p}$结构描述，由
$$\begin{aligned}\frac{dg(\mathbf{q},\mathbf{p})}{dt}=\sum_{i=1}^{N\cdot d}\frac{\partial g}{\partial q_i}\frac{dq_i}{dt}+\frac{\partial g}{\partial p_i}\frac{dp_i}{dt}=\{g,\mathcal{H}\}\:.\end{aligned}$$
(1)
给出。在最后一步中，我们使用了泊松括号和汉密尔顿方程，如下所示：。当强制执行这样的约束时，哈密顿量不依赖于关联的 $Q_i$, i。e。$\mathbb{E}$它取决于较少的自由度，并且相空间中的运动仅限于较低维度的流形。换句话说，循环坐标通过$\mathbb{E}(??)$ 类型的约束对允许的哈密顿函数提供了额外的限制，我们可以通过对称控制神经网络学习。(cf。图？?除了满足哈密顿运动方程所产生的约束外，还显示了 2 体示例中角动量守恒的显式约束)
除了标准 HNN 损失之外，这些条件还导致两个额外的损失分量：
1. The first loss ensures that our Hamiltonian satisfes Hamiltonian equations (??),which
we can ensure as follows:
$$\mathcal{L}_{\mathrm{HNN}}=\sum_{i=1}^{N\cdot d}\left\|\frac{\partial\mathcal{H}_{\phi}(\mathbf{P},\mathbf{Q})}{\partial p_{i}}-\frac{dq_{i}}{dt}\right\|_{2}+\left\|\frac{\partial\mathcal{H}_{\phi}(\mathbf{P},\mathbf{Q})}{\partial q_{i}}+\frac{dp_{i}}{dt}\right\|_{2}\:.$$
(2)
The time derivatives are provided by the data and the derivatives of the Hamiltonian with respect to the input variables can be obtained using auto-differentiation. This is the same loss as introduced in Greydanus et al. (2019) .
2. To ensure that our transformation $T_\psi$ are of the type we are interested in (cf. equa-
tion (??) ),i.e. our new coordinates fulfil the Poisson algebra, we enforce the following
$\log$SS:
$$\mathcal{L}_{\mathrm{Poisson}}=\sum_{i,j=1}^{N\cdot d}\left\|\{Q_i,P_j\}-\delta_{ij}\right\|_2+\sum_{i,j>i}^{N\cdot d}\left\|\{P_i,P_j\}\right\|_2+\left\|\{Q_i,Q_j\}\right\|_2\:,$$
(3)
where in some practical applications we only enforce this loss on $n$ cyclic coordinate pairs.
The first part of this loss ensures that a vanishing solution is not allowed.
3. Hamilton's equations have still to be satisfied with respect to the new coordinates. For
$\begin{array}{c}{\text{the cyclic coordinates we have enforced by our architecture that }\mathcal{H}_{\phi}\text{ is independent of }Q_{i}.}\end{array}$
To ensure that $P_i$ is actually conserved, we require the following additional loss:
$$\begin{aligned}\mathcal{L}_{\mathrm{HQP}}^{(n)}&=\sum_{i=1}^{n}\left\|\frac{dP_{i}}{dt}\right\|_{2}+\left\|\frac{dQ_{i}}{dt}-\frac{\partial\mathcal{H}_{\phi}(\mathbf{P},\mathbf{Q})}{\partial P_{i}}\right\|_{2}\\&+\beta\sum_{i=n+1}^{N\cdot d}\left\|\frac{dP_{i}}{dt}+\frac{\partial\mathcal{H}_{\phi}(\mathbf{P},\mathbf{Q})}{\partial Q_{i}}\right\|_{2}+\left\|\frac{dQ_{i}}{dt}-\frac{\partial\mathcal{H}_{\phi}(\mathbf{P},\mathbf{Q})}{\partial P_{i}}\right\|_{2}\:,\end{aligned}$$
(4)
where $n$ denotes the number of cyclic variables we are imposing and $\beta$ denotes a hyperparameter. For our networks we only constrain the cyclic coordinates and use $\beta=0.$ $\dot{\text{The time derivatives can be calculated using either expressions in}}$ (1) .
我们的总损失是这三个分量的加权和：
$$\mathcal{L}=\mathcal{L}_{\mathrm{HNN}}+\alpha_{1}\mathcal{L}_{\mathrm{Poisson}}+\alpha_{2}\mathcal{L}_{\mathrm{HQP}}^{(n)}\:,$$
(5)
,其中调整了权重 $\alpha_i$
为了从我们各自的对称控制网络中及时集成解决方案，我们使用$\overline{\mathrm{E}}$Greydanus et al. (2019) 中的四阶 Runge-Kutta 积分器，与辛积分器不同 , 它允许与直接预测系统动力学的神经网络方法进行比较。
$^{1}$We fnd that the main numerical inaccuracy in the prediction arises from inaccuracies of the Hamiltonian $\mathcal{H}_\phi$ rather than the choice of this integrator when comparing it with standard symplectic integrators (Rein &Liu, 2012)
2 | [
[
"0.1765",
"0.1061",
"0.1896",
"0.1061",
"0.1896",
"0.11743",
"0.1765",
"0.11743"
],
[
"0.2091",
"0.1042",
"0.2876",
"0.1061",
"0.2876",
"0.1206",
"0.2091",
"0.1187"
],
[
"0.1748",
"0.1338",
"0.8228",
"0.1338",
"0.8228",
"0.1477",
"0.1748",
"0.1477"
],
[
"0.1765",
"0.1483",
"0.8228",
"0.1483",
"0.8228",
"0.1635",
"0.1765",
"0.1635"
],
[
"0.1765",
"0.1635",
"0.824",
"0.1635",
"0.824",
"0.1768",
"0.1765",
"0.1768"
],
[
"0.1748",
"0.1768",
"0.773",
"0.1768",
"0.773",
"0.1901",
"0.1748",
"0.1901"
],
[
"0.4224",
"0.1925",
"0.5786",
"0.1995",
"0.5767",
"0.2216",
"0.4207",
"0.214"
],
[
"0.3464",
"0.2008",
"0.4102",
"0.2008",
"0.4102",
"0.2153",
"0.3464",
"0.2153"
],
[
"0.4102",
"0.209",
"0.4412",
"0.209",
"0.4412",
"0.2172",
"0.4102",
"0.2172"
],
[
"0.574",
"0.207",
"0.6567",
"0.209",
"0.6567",
"0.2247",
"0.574",
"0.2229"
],
[
"0.8013",
"0.209",
"0.824",
"0.209",
"0.824",
"0.2235",
"0.8013",
"0.2235"
],
[
"0.3652",
"0.2172",
"0.3882",
"0.2172",
"0.3882",
"0.2299",
"0.3652",
"0.2299"
],
[
"0.4395",
"0.2172",
"0.576",
"0.2172",
"0.576",
"0.2317",
"0.4395",
"0.2317"
],
[
"0.4265",
"0.2267",
"0.4468",
"0.2267",
"0.4468",
"0.2343",
"0.4265",
"0.2343"
],
[
"0.1765",
"0.2412",
"0.8228",
"0.2412",
"0.8228",
"0.2556",
"0.1765",
"0.2556"
],
[
"0.1748",
"0.2556",
"0.822",
"0.2556",
"0.822",
"0.2695",
"0.1748",
"0.2695"
],
[
"0.1732",
"0.2683",
"0.8228",
"0.2695",
"0.8228",
"0.2842",
"0.1732",
"0.2827"
],
[
"0.1765",
"0.2842",
"0.824",
"0.2842",
"0.824",
"0.2974",
"0.1765",
"0.2974"
],
[
"0.1748",
"0.2974",
"0.7075",
"0.2974",
"0.7075",
"0.3113",
"0.1748",
"0.3113"
],
[
"0.1748",
"0.318",
"0.6396",
"0.318",
"0.6396",
"0.332",
"0.1748",
"0.332"
],
[
"0.1691",
"0.3435",
"0.8228",
"0.3435",
"0.8228",
"0.3567",
"0.1691",
"0.3567"
],
[
"0.1929",
"0.3591",
"0.3718",
"0.358",
"0.3718",
"0.3694",
"0.1929",
"0.3706"
],
[
"0.3586",
"0.3706",
"0.5376",
"0.379",
"0.536",
"0.401",
"0.357",
"0.3926"
],
[
"0.5537",
"0.379",
"0.6953",
"0.3806",
"0.6953",
"0.3945",
"0.5537",
"0.3926"
],
[
"0.2974",
"0.387",
"0.3677",
"0.3894",
"0.3677",
"0.404",
"0.2974",
"0.4016"
],
[
"0.361",
"0.3882",
"0.3823",
"0.3882",
"0.3823",
"0.4016",
"0.361",
"0.4016"
],
[
"0.54",
"0.3926",
"0.5537",
"0.3926",
"0.5537",
"0.3958",
"0.54",
"0.3958"
],
[
"0.8022",
"0.3882",
"0.824",
"0.3882",
"0.824",
"0.4028",
"0.8022",
"0.4028"
],
[
"0.4968",
"0.3958",
"0.5376",
"0.3977",
"0.536",
"0.4104",
"0.4958",
"0.4077"
],
[
"0.589",
"0.3977",
"0.6216",
"0.3977",
"0.6216",
"0.4104",
"0.589",
"0.4104"
],
[
"0.66",
"0.3977",
"0.6978",
"0.3977",
"0.6978",
"0.4092",
"0.66",
"0.4092"
],
[
"0.58",
"0.401",
"0.5864",
"0.401",
"0.5864",
"0.404",
"0.58",
"0.404"
],
[
"0.4224",
"0.3984",
"0.451",
"0.3984",
"0.451",
"0.4104",
"0.4224",
"0.4104"
],
[
"0.3635",
"0.406",
"0.392",
"0.406",
"0.392",
"0.4136",
"0.3635",
"0.4136"
],
[
"0.1936",
"0.4204",
"0.824",
"0.4204",
"0.824",
"0.4343",
"0.1936",
"0.4343"
],
[
"0.1936",
"0.4343",
"0.824",
"0.4343",
"0.824",
"0.4475",
"0.1936",
"0.4475"
],
[
"0.1953",
"0.4495",
"0.572",
"0.4495",
"0.572",
"0.4602",
"0.1953",
"0.4602"
],
[
"0.1691",
"0.4666",
"0.822",
"0.4666",
"0.822",
"0.4797",
"0.1691",
"0.4797"
],
[
"0.1912",
"0.4797",
"0.8228",
"0.4797",
"0.8228",
"0.4944",
"0.1912",
"0.4944"
],
[
"0.1929",
"0.4944",
"0.2263",
"0.4944",
"0.2263",
"0.506",
"0.1929",
"0.506"
],
[
"0.3489",
"0.5024",
"0.3718",
"0.5024",
"0.3718",
"0.515",
"0.3489",
"0.515"
],
[
"0.523",
"0.5024",
"0.5474",
"0.5024",
"0.5474",
"0.515",
"0.523",
"0.515"
],
[
"0.2656",
"0.5176",
"0.343",
"0.5195",
"0.343",
"0.534",
"0.2656",
"0.532"
],
[
"0.3447",
"0.515",
"0.7573",
"0.5166",
"0.7573",
"0.534",
"0.3447",
"0.5327"
],
[
"0.8013",
"0.518",
"0.824",
"0.518",
"0.824",
"0.532",
"0.8013",
"0.532"
],
[
"0.343",
"0.535",
"0.3774",
"0.535",
"0.3774",
"0.547",
"0.343",
"0.547"
],
[
"0.517",
"0.535",
"0.55",
"0.535",
"0.55",
"0.547",
"0.517",
"0.547"
],
[
"0.1953",
"0.55",
"0.8228",
"0.55",
"0.8228",
"0.563",
"0.1953",
"0.563"
],
[
"0.1936",
"0.563",
"0.721",
"0.563",
"0.721",
"0.577",
"0.1936",
"0.577"
],
[
"0.1708",
"0.5806",
"0.824",
"0.5806",
"0.824",
"0.5947",
"0.1708",
"0.5947"
],
[
"0.1929",
"0.5947",
"0.8203",
"0.5947",
"0.8203",
"0.608",
"0.1929",
"0.608"
],
[
"0.1936",
"0.6094",
"0.767",
"0.6094",
"0.767",
"0.6226",
"0.1936",
"0.6226"
],
[
"0.4355",
"0.632",
"0.468",
"0.632",
"0.468",
"0.6445",
"0.4355",
"0.6445"
],
[
"0.4885",
"0.6274",
"0.572",
"0.631",
"0.5703",
"0.6465",
"0.4878",
"0.643"
],
[
"0.3718",
"0.634",
"0.3938",
"0.634",
"0.3938",
"0.6426",
"0.3718",
"0.6426"
],
[
"0.343",
"0.6333",
"0.357",
"0.64",
"0.3447",
"0.654",
"0.3308",
"0.646"
],
[
"0.2769",
"0.636",
"0.3105",
"0.632",
"0.3137",
"0.6445",
"0.2793",
"0.6484"
],
[
"0.2778",
"0.6426",
"0.327",
"0.646",
"0.326",
"0.6577",
"0.2769",
"0.655"
],
[
"0.3743",
"0.646",
"0.4004",
"0.6494",
"0.3962",
"0.662",
"0.3708",
"0.657"
],
[
"0.4307",
"0.6465",
"0.4673",
"0.6445",
"0.468",
"0.66",
"0.4314",
"0.662"
],
[
"0.513",
"0.6465",
"0.546",
"0.6465",
"0.546",
"0.662",
"0.513",
"0.662"
],
[
"0.3381",
"0.657",
"0.3594",
"0.657",
"0.3594",
"0.6655",
"0.3381",
"0.6655"
],
[
"0.4004",
"0.657",
"0.4167",
"0.657",
"0.4167",
"0.66",
"0.4004",
"0.66"
],
[
"0.5728",
"0.653",
"0.5864",
"0.653",
"0.5864",
"0.6636",
"0.5728",
"0.6636"
],
[
"0.3677",
"0.6694",
"0.392",
"0.6694",
"0.392",
"0.681",
"0.3677",
"0.681"
],
[
"0.4495",
"0.674",
"0.54",
"0.6753",
"0.54",
"0.69",
"0.4495",
"0.688"
],
[
"0.63",
"0.674",
"0.716",
"0.6753",
"0.716",
"0.69",
"0.63",
"0.688"
],
[
"0.5825",
"0.677",
"0.613",
"0.677",
"0.613",
"0.688",
"0.5825",
"0.688"
],
[
"0.407",
"0.6753",
"0.437",
"0.6753",
"0.437",
"0.688",
"0.407",
"0.688"
],
[
"0.3367",
"0.6846",
"0.3586",
"0.6846",
"0.3586",
"0.6963",
"0.3367",
"0.6963"
],
[
"0.3652",
"0.6816",
"0.3823",
"0.6816",
"0.3823",
"0.6978",
"0.3652",
"0.6978"
],
[
"0.8022",
"0.6846",
"0.824",
"0.6846",
"0.824",
"0.6978",
"0.8022",
"0.6978"
],
[
"0.4805",
"0.6914",
"0.5186",
"0.6914",
"0.5186",
"0.707",
"0.4805",
"0.707"
],
[
"0.6567",
"0.6914",
"0.6914",
"0.6914",
"0.6914",
"0.707",
"0.6567",
"0.707"
],
[
"0.581",
"0.6924",
"0.607",
"0.6924",
"0.607",
"0.705",
"0.581",
"0.705"
],
[
"0.4084",
"0.6924",
"0.4346",
"0.6924",
"0.4346",
"0.705",
"0.4084",
"0.705"
],
[
"0.5415",
"0.6997",
"0.5537",
"0.6997",
"0.5537",
"0.704",
"0.5415",
"0.704"
],
[
"0.716",
"0.6997",
"0.728",
"0.6997",
"0.728",
"0.7056",
"0.716",
"0.7056"
],
[
"0.1936",
"0.7173",
"0.822",
"0.7173",
"0.822",
"0.731",
"0.1936",
"0.731"
],
[
"0.3594",
"0.7007",
"0.3987",
"0.7007",
"0.3987",
"0.711",
"0.3594",
"0.711"
],
[
"0.1936",
"0.73",
"0.824",
"0.73",
"0.824",
"0.7437",
"0.1936",
"0.7437"
],
[
"0.1929",
"0.742",
"0.7036",
"0.7437",
"0.7036",
"0.7583",
"0.1929",
"0.757"
],
[
"0.1765",
"0.7695",
"0.4436",
"0.7695",
"0.4436",
"0.7827",
"0.1765",
"0.7827"
],
[
"0.3677",
"0.785",
"0.6313",
"0.7915",
"0.6313",
"0.813",
"0.3677",
"0.8066"
],
[
"0.8022",
"0.7925",
"0.826",
"0.7925",
"0.826",
"0.806",
"0.8022",
"0.806"
],
[
"0.1748",
"0.813",
"0.3284",
"0.8145",
"0.3284",
"0.829",
"0.1748",
"0.8276"
],
[
"0.1772",
"0.8354",
"0.822",
"0.8354",
"0.822",
"0.8486",
"0.1772",
"0.8486"
],
[
"0.1748",
"0.8477",
"0.822",
"0.8486",
"0.822",
"0.863",
"0.1748",
"0.8623"
],
[
"0.1748",
"0.863",
"0.3145",
"0.863",
"0.3145",
"0.877",
"0.1748",
"0.877"
],
[
"0.1936",
"0.884",
"0.824",
"0.8853",
"0.824",
"0.8994",
"0.1936",
"0.8984"
],
[
"0.1748",
"0.8984",
"0.822",
"0.8984",
"0.822",
"0.912",
"0.1748",
"0.912"
],
[
"0.1748",
"0.911",
"0.4429",
"0.911",
"0.4429",
"0.925",
"0.1748",
"0.925"
],
[
"0.4927",
"0.9487",
"0.5063",
"0.9487",
"0.5063",
"0.9604",
"0.4927",
"0.9604"
]
] | [
"2",
"THEORY",
"当运动发生在相空间的子空间上时，可以减少描述系统的坐标数量。每个独立的守恒量都",
"会减少描述系统所需的坐标数，直到所有坐标都受到约束（可积系统）。即使我们不知道先",
"验的守恒坐标，哈密顿力学也为这种坐标提供了一般条件。特别是相空间上任何函数的时",
"间导数，由位置q和动量p的N粒子在d空间维度上的p不变的p结构描述，由",
"Nd ag di + ag dpi",
"dg(q,p)",
"",
"=(g,H).",
"(1)",
"dt",
"OqidtOpidt",
"-1",
"给出。在最后一步中，我们使用了泊松括号和汉密尔顿方程，如下所示：。当强制执行这样",
"的约束时，哈密顿量不依赖于关联的Qi，i。e。E它取决于较少的自由度，并且相空间中的",
"运动仅限于较低维度的流形。换句话说，循环坐标通过E（？？）类型的约束对允许的哈密顿",
"函数提供了额外的限制，我们可以通过对称控制神经网络学习。（cf。图？？除了满足哈密",
"顿运动方程所产生的约束外，还显示了2体示例中角动量守恒的显式约束）",
"除了标准HNN损失之外，这些条件还导致两个额外的损失分量：",
"1. The first loss ensures that our Hamiltonian satisfies Hamiltonian eguations (??), which",
"wecan ensureas follows:",
"Nd oHo(P,Q) _ da",
"aHs(P,Q) dp",
"CHNN=",
">",
"",
"(2)",
"dt12",
"Oqi",
"dt",
"",
"Opi",
"-1",
"The time derivatives are provided by the data and the derivatives of the Hamiltonian",
"with respect to the input variables can be obtained using auto-differentiation. This is the",
"same loss as introduced in Greydanus et al.(2019).",
"2. To ensure that our transformation Tu are of the type we are interested in (cf. equa-",
"tion (??)), i.e. our new coordinates fulfil the Poisson algebra, we enforce the following",
"loss:",
"N.d",
"N.d",
"Cpoisson =",
"Qi,P)-ill+ IP,Pll2+I(Q,Qll2",
"(3)",
"i,j=1",
"i,j>i",
"where in some practical applications we only enforce this loss on n cyclic coordinate pairs.",
"The first part of this loss ensures that a vanishing solution is not allowed.",
"3. Hamilton's equations have still to be satisfied with respect to the new coordinates. For",
"the cyclic coordinates we have enforced by our architecture that H is independent of Qi",
"To ensure that P, is actually conserved, we require the following additional loss:",
"dQi",
"Hs(P,Q)",
"dP",
"A",
"c(n)",
"LHQP",
"dt",
"dt",
"ap,",
"i=1",
"",
"12",
"N-d",
"H(P,Q)",
"OH(P,Q)",
"dQ",
"dP",
"+β",
">",
"(4)",
"QQ:",
"aP,",
"dt",
"dt",
"1n",
"12",
"where n denotes the number of cyclic variables we are imposing and β denotes a hyper-",
"i=n+1",
"parameter. For our networks we only constrain the cyclic coordinates and use β = 0 .",
"The time derivatives can be calculated using either expressions in (1):",
"我们的总损失是这三个分量的加权和：",
"C=CHNN+QiCpoion+a2CHQP,",
"(5)",
"其中调整了权重αi",
"为了从我们各自的对称控制网络中及时集成解决方案，我们使用EGreydanusetal.（2019）",
"中的四阶Runge-Kutta积分器，与辛积分器不同，它允许与直接预测系统动力学的神经网",
"络方法进行比较。1",
"IWe find that the main numerical inaccuracy in the prediction arises from inaccuracies of the",
"Hamiltonian Hs rather than the choice of this integrator when comparing it with standard sym-",
"plectic integrators(Rein & Liu, 2012).",
"2"
] | {
"arxiv_id": "2104.14444",
"language": "zh",
"page_idx": 1
} |
||
2104.14444-zh-2 | 2104.14444-zh-2.jpg | <FigureHere>
Figure 2: Comparison between two HNN networks as described in the main text and SCNN
$\bar{\text{networks with different numbers of conserved quantities. Left: Sample trajectories of the}}$ gravitational two-body problem. Middle: $\bar\text{Comparison of the mean-squared deviation of}$ the trajectories, the conservation of energy and angular momentum for the 2-body problem. $\textbf{Right: Comparison of the mean- squared deviations for various spring- like interactions. }$ The bottom plot shows the well-behaved dependence on the loss prefactor $\alpha=\alpha_i.$ For all evaluations of the trajectories we show the mean and single standard deviation over 10 different initializations evaluated on 100 trajectories. $\alpha_1=\alpha_2=10^{-n}$ where $n$ is the number of constraint conserved quantities.
关于这种方法的效率，在推理过程中没有额外的成本。在训练过程中，我们的损失需要对已经计算出的梯度进行额外的评估。这些项的数量以尺寸和粒子数$(N\cdot d)^3$ 的乘积为立方尺度 (参见。泊松损失)。当守恒量已知时，这一成本就会降低(例如。g。来自系统的先验知识),因为泊松损失是自动满足的，我们只对每个守恒量调用哈密顿量上的额外损失贡献。WenotethatforcaseswheretheHamiltoninvolvesonlylocalinteractionsanevaluationoftheseconstraintsontheselocalinteractionsisrelevantwhichresultsinareductionoftherelevantadditionationdereditiondereditionationditionditionditionditionditionditionditionditionditionditionditionditionditionditionditionditiondesionditiondi
因此，与 HNN 案例相比，成本没有显著提高
## 3 EXPERIMENTS
## 我们的实验$^{2}$的设计考虑到以下目标：
1. SCNN: We want to compare the performance of symmetry control neural networks
with HNNs and baseline neural networks which directly predict $(\mathbf{\dot{q}},\mathbf{\dot{p}}).$ We want to take advantage of unknown underlying symmetries while we do not have any domair knowledge of our system. The main hyperparameter is the number of conserved quantities.
$2. \textbf{ SCNN- constraint: We explore whether imposing domain knowledge about symme- }$
tries improves the performance. This is motivated by the fact that we often know about
the existence of certain conserved quantities such as (angular) momentum.
我们在几种不同的物理系统上测试这些体系结构：引力两体问题，具有类似弹簧相互作用的$n$体系统，磁场中的带电粒子，球形摆和双摆。我们更详细地描述了我们的架构对前两个的超参数依赖性，并仅报告后三个的超参数扫描结果。弹簧状相互作用包含 3 到 5 个粒子，每个粒子与力$\mathbf{F^{ij}} = - k$ $\left ( \mathbf{x} ^{i}- \mathbf{x} ^{j}\right )$相互作用。引力由$\mathbf{F^{ij}} = - k$ $\frac {( \mathbf{x} ^{i}- \mathbf{x} ^{j}) }{| \mathbf{x} ^{i}- \mathbf{x} ^{j}| ^{3}}$描述
$^2$Details on hyperparameters and training can be found in Appendix B.
3 | [
[
"0.1814",
"0.108",
"0.2932",
"0.108",
"0.2932",
"0.1187",
"0.1814",
"0.1187"
],
[
"0.3162",
"0.10986",
"0.3594",
"0.10986",
"0.3594",
"0.1932",
"0.3162",
"0.1932"
],
[
"0.6357",
"0.1042",
"0.7695",
"0.1029",
"0.7695",
"0.11426",
"0.6357",
"0.11554"
],
[
"0.3782",
"0.1111",
"0.423",
"0.11426",
"0.4192",
"0.1401",
"0.3757",
"0.137"
],
[
"0.6978",
"0.11615",
"0.732",
"0.11615",
"0.732",
"0.1276",
"0.6978",
"0.1276"
],
[
"0.7524",
"0.11426",
"0.8037",
"0.11426",
"0.8037",
"0.125",
"0.7524",
"0.125"
],
[
"0.6313",
"0.11743",
"0.6465",
"0.11743",
"0.6465",
"0.127",
"0.6313",
"0.127"
],
[
"0.183",
"0.1193",
"0.2303",
"0.1193",
"0.2303",
"0.13",
"0.183",
"0.13"
],
[
"0.584",
"0.1193",
"0.597",
"0.1193",
"0.597",
"0.1237",
"0.584",
"0.1237"
],
[
"0.1887",
"0.1288",
"0.2197",
"0.1288",
"0.2197",
"0.137",
"0.1887",
"0.137"
],
[
"0.3774",
"0.127",
"0.4265",
"0.127",
"0.4265",
"0.159",
"0.3774",
"0.159"
],
[
"0.5786",
"0.13",
"0.603",
"0.13",
"0.603",
"0.1433",
"0.5786",
"0.1433"
],
[
"0.6313",
"0.132",
"0.6626",
"0.132",
"0.6626",
"0.1401",
"0.6313",
"0.1401"
],
[
"0.6978",
"0.132",
"0.7305",
"0.132",
"0.7305",
"0.1395",
"0.6978",
"0.1395"
],
[
"0.3816",
"0.154",
"0.4656",
"0.154",
"0.4656",
"0.1686",
"0.3816",
"0.1686"
],
[
"0.574",
"0.137",
"0.607",
"0.137",
"0.607",
"0.1787",
"0.574",
"0.1787"
],
[
"0.5728",
"0.1755",
"0.609",
"0.1755",
"0.609",
"0.2349",
"0.5728",
"0.2349"
],
[
"0.3367",
"0.1945",
"0.3472",
"0.1945",
"0.3472",
"0.2058",
"0.3367",
"0.2058"
],
[
"0.3096",
"0.2026",
"0.3555",
"0.202",
"0.3586",
"0.2651",
"0.312",
"0.2664"
],
[
"0.1912",
"0.2102",
"0.2181",
"0.2102",
"0.2181",
"0.2172",
"0.1912",
"0.2172"
],
[
"0.7524",
"0.2216",
"0.7983",
"0.2216",
"0.7983",
"0.2329",
"0.7524",
"0.2329"
],
[
"0.5728",
"0.2424",
"0.609",
"0.2424",
"0.609",
"0.2942",
"0.5728",
"0.2942"
],
[
"0.3179",
"0.2747",
"0.3545",
"0.2747",
"0.3545",
"0.324",
"0.3179",
"0.324"
],
[
"0.7524",
"0.2827",
"0.7983",
"0.2827",
"0.7983",
"0.2942",
"0.7524",
"0.2942"
],
[
"0.1871",
"0.2935",
"0.2239",
"0.2935",
"0.2239",
"0.305",
"0.1871",
"0.305"
],
[
"0.5825",
"0.2935",
"0.601",
"0.2935",
"0.601",
"0.3005",
"0.5825",
"0.3005"
],
[
"0.619",
"0.2966",
"0.759",
"0.2966",
"0.759",
"0.31",
"0.619",
"0.31"
],
[
"0.5767",
"0.305",
"0.5864",
"0.3018",
"0.5933",
"0.3113",
"0.584",
"0.3145"
],
[
"0.591",
"0.307",
"0.6006",
"0.307",
"0.6006",
"0.3145",
"0.591",
"0.3145"
],
[
"0.665",
"0.308",
"0.7295",
"0.308",
"0.7295",
"0.3308",
"0.665",
"0.3308"
],
[
"0.63",
"0.3113",
"0.66",
"0.3113",
"0.66",
"0.3225",
"0.63",
"0.3225"
],
[
"0.3179",
"0.3093",
"0.3545",
"0.3125",
"0.3472",
"0.3623",
"0.3105",
"0.3591"
],
[
"0.5728",
"0.3132",
"0.6094",
"0.3145",
"0.607",
"0.3623",
"0.5703",
"0.361"
],
[
"0.638",
"0.3145",
"0.648",
"0.3145",
"0.648",
"0.318",
"0.638",
"0.318"
],
[
"0.6436",
"0.358",
"0.6504",
"0.363",
"0.6436",
"0.3687",
"0.636",
"0.363"
],
[
"0.6787",
"0.3591",
"0.693",
"0.3591",
"0.693",
"0.3674",
"0.6787",
"0.3674"
],
[
"0.7183",
"0.3599",
"0.732",
"0.3599",
"0.732",
"0.3687",
"0.7183",
"0.3687"
],
[
"0.7974",
"0.3599",
"0.812",
"0.3599",
"0.812",
"0.3674",
"0.7974",
"0.3674"
],
[
"0.4656",
"0.361",
"0.4714",
"0.361",
"0.4714",
"0.3662",
"0.4656",
"0.3662"
],
[
"0.4265",
"0.361",
"0.4346",
"0.361",
"0.4346",
"0.3662",
"0.4265",
"0.3662"
],
[
"0.5415",
"0.361",
"0.5537",
"0.361",
"0.5537",
"0.3674",
"0.5415",
"0.3674"
],
[
"0.6006",
"0.361",
"0.6055",
"0.361",
"0.6055",
"0.3662",
"0.6006",
"0.3662"
],
[
"0.7607",
"0.361",
"0.7695",
"0.361",
"0.7695",
"0.3662",
"0.7607",
"0.3662"
],
[
"0.6895",
"0.3655",
"0.7197",
"0.3655",
"0.7197",
"0.377",
"0.6895",
"0.377"
],
[
"0.4355",
"0.3655",
"0.464",
"0.3655",
"0.464",
"0.377",
"0.4355",
"0.377"
],
[
"0.1748",
"0.3833",
"0.8228",
"0.3833",
"0.8228",
"0.3977",
"0.1748",
"0.3977"
],
[
"0.1748",
"0.3977",
"0.824",
"0.3977",
"0.824",
"0.411",
"0.1748",
"0.411"
],
[
"0.1748",
"0.411",
"0.826",
"0.411",
"0.826",
"0.4255",
"0.1748",
"0.4255"
],
[
"0.1732",
"0.4255",
"0.8203",
"0.4255",
"0.8203",
"0.4395",
"0.1732",
"0.4395"
],
[
"0.1765",
"0.4395",
"0.822",
"0.4395",
"0.822",
"0.4526",
"0.1765",
"0.4526"
],
[
"0.1765",
"0.4526",
"0.7646",
"0.4526",
"0.7646",
"0.4673",
"0.1765",
"0.4673"
],
[
"0.7646",
"0.455",
"0.822",
"0.455",
"0.822",
"0.4653",
"0.7646",
"0.4653"
],
[
"0.1748",
"0.4673",
"0.8228",
"0.4673",
"0.8228",
"0.481",
"0.1748",
"0.481"
],
[
"0.1765",
"0.481",
"0.826",
"0.481",
"0.826",
"0.4944",
"0.1765",
"0.4944"
],
[
"0.1748",
"0.4956",
"0.482",
"0.4956",
"0.482",
"0.509",
"0.1748",
"0.509"
],
[
"0.1765",
"0.535",
"0.8228",
"0.535",
"0.8228",
"0.549",
"0.1765",
"0.549"
],
[
"0.1748",
"0.549",
"0.8228",
"0.547",
"0.8228",
"0.561",
"0.1748",
"0.563"
],
[
"0.1765",
"0.5625",
"0.8228",
"0.5625",
"0.8228",
"0.577",
"0.1765",
"0.577"
],
[
"0.1765",
"0.577",
"0.822",
"0.577",
"0.822",
"0.5913",
"0.1765",
"0.5913"
],
[
"0.1765",
"0.5923",
"0.8203",
"0.5923",
"0.8203",
"0.603",
"0.1765",
"0.603"
],
[
"0.1765",
"0.606",
"0.9966",
"0.606",
"0.9966",
"0.616",
"0.1765",
"0.616"
],
[
"0.1748",
"0.6187",
"0.501",
"0.6187",
"0.501",
"0.632",
"0.1748",
"0.632"
],
[
"0.1765",
"0.655",
"0.1887",
"0.655",
"0.1887",
"0.6655",
"0.1765",
"0.6655"
],
[
"0.21",
"0.653",
"0.339",
"0.653",
"0.339",
"0.6665",
"0.21",
"0.6665"
],
[
"0.1732",
"0.6816",
"0.4453",
"0.683",
"0.4453",
"0.6978",
"0.1732",
"0.6963"
],
[
"0.183",
"0.7085",
"0.824",
"0.709",
"0.824",
"0.7236",
"0.183",
"0.7227"
],
[
"0.2075",
"0.7236",
"0.824",
"0.7236",
"0.824",
"0.738",
"0.2075",
"0.738"
],
[
"0.2075",
"0.738",
"0.824",
"0.738",
"0.824",
"0.752",
"0.2075",
"0.752"
],
[
"0.2059",
"0.752",
"0.8228",
"0.752",
"0.8228",
"0.765",
"0.2059",
"0.765"
],
[
"0.2075",
"0.7666",
"0.2484",
"0.7666",
"0.2484",
"0.778",
"0.2075",
"0.778"
],
[
"0.1847",
"0.784",
"0.822",
"0.784",
"0.822",
"0.7974",
"0.1847",
"0.7974"
],
[
"0.2075",
"0.7974",
"0.824",
"0.7974",
"0.824",
"0.811",
"0.2075",
"0.811"
],
[
"0.2075",
"0.8125",
"0.7417",
"0.8125",
"0.7417",
"0.8257",
"0.2075",
"0.8257"
],
[
"0.1765",
"0.837",
"0.822",
"0.837",
"0.822",
"0.851",
"0.1765",
"0.851"
],
[
"0.1748",
"0.851",
"0.8228",
"0.851",
"0.8228",
"0.8657",
"0.1748",
"0.8657"
],
[
"0.1772",
"0.8657",
"0.8228",
"0.8657",
"0.8228",
"0.879",
"0.1772",
"0.879"
],
[
"0.1732",
"0.879",
"0.7466",
"0.879",
"0.7466",
"0.903",
"0.1732",
"0.903"
],
[
"0.1936",
"0.909",
"0.6772",
"0.91",
"0.6772",
"0.925",
"0.1936",
"0.924"
],
[
"0.492",
"0.9478",
"0.509",
"0.9478",
"0.509",
"0.962",
"0.492",
"0.962"
]
] | [
"Grav.2-body system",
"on ",
"Spring-likeinteractior",
"",
"SCNN(3)",
"3bodies",
"HNN",
"Ground-",
"1.",
"truth",
"",
"00.8",
"SCNN",
"SCNN",
"SCNN-constraint (a)",
"ois",
"",
"10",
"AN",
"HNN",
"4bodies",
"",
"",
"5bodies",
"SCNN",
"0.0",
"Parameter scan (3 bodies)",
"O",
"1:2",
"E",
"a=le-s",
"S",
"O",
".",
"20",
"40",
"60",
"100",
"60",
"40",
"100",
"0",
"80",
"Time",
"Time",
"Figure 2: Comparison between two HNN networks as described in the main text and SCNN",
"networks with different numbers of conserved quantities.Left:Sample trajectories of the",
"gravitational two-body problem. Middle: Comparison of the mean-squared deviation of",
"the trajectories, the conservation of energy and angular momentum for the 2-body problem",
"Right:Comparison of the mean-squared deviations for various spring-like interactions:",
"The bottom plot shows the well-behaved dependence on the loss prefactor α =",
"Qi.For",
"all evaluations of the trajectories we show the mean and single standard deviation over",
"10 different initializations evaluated on 100 trajectories. αi = α2 =10-n where n is the",
"number of constraint conserved quantities.",
"关于这种方法的效率，在推理过程中没有额外的成本。在训练过程中，我们的损失",
"需要对已经计算出的梯度进行额外的评估。这些项的数量以尺寸和粒子数（N：d)3",
"的乘积为立方尺度（参见。泊松损失）。当守恒量已知时，这一成本就会降低（例",
"如。g。来自系统的先验知识)，因为泊松损失是自动满足的，我们只对每个守恒量调用",
"哈密顿量上的额外损失贡献。WenotethatforcaseswheretheHamiltoninvolvesonlylocalinterac",
"tionsanevaluationoftheseconstraintsontheselocalinteractionsisrelevantwhichresultsinareductionoftherelevantadditiona",
"因此，与HNN案例相比，成本没有显著提高",
"3",
"EXPERIMENTS",
"我们的实验2的设计考虑到以下目标：",
"1. SCNN: We want to compare the performance of symmetry control neural networks",
"with HNNs and baseline neural networks which directly predict (q,p)-We want to",
"take advantage of unknown underlying symmetries while we do not have any domain",
"knowledge of our system. The main hyperparameter is the number of conserved quan-",
"tities.",
"2. SCNN-constraint: We explore whether imposing domain knowledge about symme-",
"tries improves the performance. This is motivated by the fact that we often know about",
"the existence of certain conserved quantities such as (angular) momentum.",
"我们在几种不同的物理系统上测试这些体系结构：引力两体问题，具有类似弹簧相互作用",
"的n体系统，磁场中的带电粒子，球形摆和双摆。我们更详细地描述了我们的架构对前两",
"个的超参数依赖性，并仅报告后三个的超参数扫描结果。弹簧状相互作用包含3到5个粒",
"子，每个粒子与力Fi=-（x-x）相互作用。引力由Fi=-描述",
"2Details on hyperparameters and training can be found in Appendix B.",
"3"
] | {
"arxiv_id": "2104.14444",
"language": "zh",
"page_idx": 2
} |
||
2104.14444-zh-3 | 2104.14444-zh-3.jpg | <table>
<tbody>
<tr>
<th>$\operatorname{Experiment}$</th>
<th>SCNN</th>
<th>SCNN-constraint</th>
<th>$HNN$</th>
</tr>
<tr>
<td>Magnetic feld</td>
<td>$0.164\pm0.025$</td>
<td>$\mathbf{0.033\pm0.010}$</td>
<td>$0.083\pm0.019$</td>
</tr>
<tr>
<td>Spherical pendulum</td>
<td>$0.092\pm0.33$</td>
<td>$\mathbf{0}.\mathbf{0}55\pm\mathbf{0}.005$</td>
<td>$0.288\pm0.007$</td>
</tr>
<tr>
<td>Double pendulum</td>
<td>$\mathbf{0}.\mathbf{0}14\pm\mathbf{0}.\mathbf{0}04$</td>
<td> </td>
<td>$0.117\pm0.012$</td>
</tr>
</tbody>
</table>
Table 1: The MSE after 100 time steps for different experiments. For the double pendulum we compare the MSE after 20 time steps. The data points are collected for 10 different initalizations and the error bars correspond to single standard deviation. All SCNN networks in these experiments have two conserved quantities. The best performing approach is highlighted.
为了了解守恒量的数量和损失前置因子$\alpha_i$的影响，我们对守恒量的数量和这些前置因子进行了参数扫描。为了比较网络的精度，我们计算了 100 不同系统初始化的粒子轨迹，并将其与 100 时间步长上的地面实况进行了比较。此外，我们检查了引力 2 体问题的能量守恒和角动量守恒。我们的网络在与第一个 50 时间步长相对应的数据点上进行了训练，这使我们可以比较我们的网络在相空间看不见的区域上的泛化。结果总结在图 3 中。总的来说，这些结果表明 SCNN 学习了一个更加稳定和准确的哈密顿量，我们发现 SCNN 约束网络无法超越 SCNN 网络。这些图中省略了基线网络，因为其性能明显更差。HNN 网络具有与我们的$\mathcal{H}_\phi$网络相同的复杂性，具有两个密集层，并且 HNN\_L 网络对应于五个密集层的复杂性，即。e。我们的$T_\psi$和$\mathcal{H}_\phi$网络的综合复杂性。损失预因子依赖性是这样的，我们使用的守恒量越多，预因子就必须越小，这是由于损失项的增加
在具有磁场中的带电粒子，球形摆和双摆的设置中，与 HNN 相比，我们的 SCNN(约束) 网络的性能得到了改善。表 1 显示了我们在 100 时间步长后获得的均方误差预测。这些系统的描述可以在附录 A 中找到
3.1
Search for conserved quantities 我们的 SCNN 网络的守恒量可以通过将低维多项式 ansatz 拟合到各自的网络预测来分析。这表明我们的 SCNN 发现角动量和总动量是引力两体系统中的守恒量：
$$\begin{aligned}P_{c_{1}}&=-\:4.2\:p_{x_{1}}-4.2\:p_{x_{2}}-1.3\:p_{y_{1}}\:-\:1.3\:p_{y_{2}}\:,\\P_{c_{2}}&=-\:0.9\:p_{x_{1}}-0.9\:p_{x_{2}}-3.2\:p_{y_{1}}-3.2\:p_{y_{2}}\:,\\&L=1.0\:q_{x_{1}}p_{y_{1}}+0.9\:q_{x_{1}}p_{y_{2}}+0.9\:q_{x_{2}}p_{y_{1}}-1.0\:q_{x_{2}}p_{y_{2}}\\&+1.0\:q_{y_{1}}p_{x_{1}}-0.9\:q_{y_{1}}p_{x_{2}}-0.9\:q_{y_{2}}p_{x_{1}}+1.0\:q_{y_{2}}p_{x_{2}}\:.\\\end{aligned}$$
(6)
对于更复杂的守恒量 (i。e。非多项式守恒量)不同的 ansätze 似乎是必要的 (其中一些是
在相关工作中追求的 (Sahoo et al., 2018; Cranmer et al., 2019; Wetzel et al., 2020)
4 OUTLOOK AND RELATED WORK
我们的 SCNN 以与 HNN 相同的方式自然地与诸如E(Battaglia et al., 2016) 之类的神经网络推断动力学的工作联系在一起。我们当前工作的自然扩展可以包括在基于图神经网络的方法$\mathsf{E}($Sanchez-Gonzalez et al., 2019) 和流程$\mathsf{E}($Toth et al., 2020) 上的应用。关于应用，我们特别期待应用我们的新方法来自动推断和使用天体物理和分子动力学设置中的对称性。
## REFERENCES
Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, et al. Interac-
tion networks for learning about objects, relations and physics. In Advances in neural
information processing systems , pp. 4502-4510, 2016.
Miles D Cranmer, Rui Xu, Peter Battaglia, and Shirley Ho. Learning symbolic physics with
graph networks. arXiv preprint arXiv:1909.05862 , 2019.
Sam Greydanus, Misko Dzamba, and Jason Yosinski. Hamiltonian neural networks. CoRR
, abs/1906.01563, 2019. URL http://arxiv.org/abs/1906.01563.
4 | [
[
"0.5156",
"0.1048",
"0.642",
"0.1048",
"0.642",
"0.1187",
"0.5156",
"0.1187"
],
[
"0.687",
"0.1042",
"0.728",
"0.1042",
"0.728",
"0.1187",
"0.687",
"0.1187"
],
[
"0.4272",
"0.1061",
"0.478",
"0.1061",
"0.478",
"0.11743",
"0.4272",
"0.11743"
],
[
"0.2443",
"0.1061",
"0.327",
"0.1061",
"0.327",
"0.11743",
"0.2443",
"0.11743"
],
[
"0.2427",
"0.1237",
"0.3472",
"0.1237",
"0.3472",
"0.1364",
"0.2427",
"0.1364"
],
[
"0.4019",
"0.1237",
"0.5",
"0.1237",
"0.5",
"0.1351",
"0.4019",
"0.1351"
],
[
"0.5317",
"0.1237",
"0.6274",
"0.1237",
"0.6274",
"0.1351",
"0.5317",
"0.1351"
],
[
"0.6587",
"0.1237",
"0.7563",
"0.1237",
"0.7563",
"0.1351",
"0.6587",
"0.1351"
],
[
"0.2418",
"0.1383",
"0.384",
"0.1383",
"0.384",
"0.1515",
"0.2418",
"0.1515"
],
[
"0.407",
"0.1395",
"0.4944",
"0.1395",
"0.4944",
"0.1509",
"0.407",
"0.1509"
],
[
"0.5337",
"0.1401",
"0.6274",
"0.1401",
"0.6274",
"0.1515",
"0.5337",
"0.1515"
],
[
"0.66",
"0.1401",
"0.755",
"0.1401",
"0.755",
"0.1515",
"0.66",
"0.1515"
],
[
"0.2418",
"0.1559",
"0.3708",
"0.1559",
"0.3708",
"0.1692",
"0.2418",
"0.1692"
],
[
"0.4043",
"0.1578",
"0.5",
"0.1578",
"0.5",
"0.1692",
"0.4043",
"0.1692"
],
[
"0.6567",
"0.1547",
"0.7573",
"0.1572",
"0.7573",
"0.1718",
"0.6567",
"0.1692"
],
[
"0.1765",
"0.1925",
"0.822",
"0.1925",
"0.822",
"0.2058",
"0.1765",
"0.2058"
],
[
"0.1765",
"0.2058",
"0.8228",
"0.2058",
"0.8228",
"0.2197",
"0.1765",
"0.2197"
],
[
"0.1732",
"0.2197",
"0.8203",
"0.2197",
"0.8203",
"0.2329",
"0.1732",
"0.2329"
],
[
"0.1748",
"0.2329",
"0.8228",
"0.2343",
"0.8228",
"0.2488",
"0.1748",
"0.2474"
],
[
"0.1765",
"0.2488",
"0.2754",
"0.2488",
"0.2754",
"0.26",
"0.1765",
"0.26"
],
[
"0.1748",
"0.2854",
"0.8228",
"0.2842",
"0.8228",
"0.2974",
"0.1748",
"0.2986"
],
[
"0.1765",
"0.2998",
"0.8228",
"0.2998",
"0.8228",
"0.3132",
"0.1765",
"0.3132"
],
[
"0.1765",
"0.3125",
"0.822",
"0.3132",
"0.822",
"0.327",
"0.1765",
"0.3257"
],
[
"0.1765",
"0.3257",
"0.822",
"0.3257",
"0.822",
"0.339",
"0.1765",
"0.339"
],
[
"0.179",
"0.3423",
"0.8203",
"0.3423",
"0.8203",
"0.3528",
"0.179",
"0.3528"
],
[
"0.1748",
"0.3528",
"0.3757",
"0.3547",
"0.3757",
"0.3687",
"0.1748",
"0.3662"
],
[
"0.3708",
"0.356",
"0.8203",
"0.356",
"0.8203",
"0.3662",
"0.3708",
"0.3662"
],
[
"0.1772",
"0.3687",
"0.8228",
"0.3687",
"0.8228",
"0.3818",
"0.1772",
"0.3818"
],
[
"0.1748",
"0.3806",
"0.822",
"0.3833",
"0.822",
"0.3965",
"0.1748",
"0.3945"
],
[
"0.1748",
"0.3945",
"0.822",
"0.3958",
"0.822",
"0.4104",
"0.1748",
"0.4092"
],
[
"0.1748",
"0.4104",
"0.6006",
"0.4104",
"0.6006",
"0.4236",
"0.1748",
"0.4236"
],
[
"0.1765",
"0.4312",
"0.822",
"0.4312",
"0.822",
"0.4443",
"0.1765",
"0.4443"
],
[
"0.1765",
"0.4458",
"0.822",
"0.4458",
"0.822",
"0.459",
"0.1765",
"0.459"
],
[
"0.1765",
"0.461",
"0.4004",
"0.461",
"0.4004",
"0.4717",
"0.1765",
"0.4717"
],
[
"0.1748",
"0.4893",
"0.2009",
"0.4893",
"0.2009",
"0.5024",
"0.1748",
"0.5024"
],
[
"0.1748",
"0.515",
"0.8228",
"0.515",
"0.8228",
"0.529",
"0.1748",
"0.529"
],
[
"0.1765",
"0.53",
"0.824",
"0.53",
"0.824",
"0.5435",
"0.1765",
"0.5435"
],
[
"0.1772",
"0.5444",
"0.2656",
"0.5444",
"0.2656",
"0.5547",
"0.1772",
"0.5547"
],
[
"0.2932",
"0.5596",
"0.621",
"0.561",
"0.621",
"0.5757",
"0.2932",
"0.5737"
],
[
"0.2917",
"0.5757",
"0.619",
"0.5776",
"0.619",
"0.5947",
"0.2917",
"0.5923"
],
[
"0.3022",
"0.5933",
"0.6807",
"0.595",
"0.6807",
"0.613",
"0.3022",
"0.6113"
],
[
"0.8022",
"0.5884",
"0.824",
"0.5884",
"0.824",
"0.6016",
"0.8022",
"0.6016"
],
[
"0.3342",
"0.6113",
"0.697",
"0.613",
"0.697",
"0.631",
"0.3342",
"0.629"
],
[
"0.1748",
"0.632",
"0.822",
"0.632",
"0.822",
"0.646",
"0.1748",
"0.646"
],
[
"0.1765",
"0.6465",
"0.7656",
"0.6465",
"0.7656",
"0.66",
"0.1765",
"0.66"
],
[
"0.1724",
"0.6797",
"0.4902",
"0.681",
"0.4902",
"0.696",
"0.1724",
"0.6943"
],
[
"0.1765",
"0.711",
"0.822",
"0.711",
"0.822",
"0.7246",
"0.1765",
"0.7246"
],
[
"0.1748",
"0.7236",
"0.822",
"0.7246",
"0.822",
"0.738",
"0.1748",
"0.7373"
],
[
"0.1748",
"0.7373",
"0.824",
"0.738",
"0.824",
"0.7524",
"0.1748",
"0.752"
],
[
"0.179",
"0.755",
"0.8",
"0.755",
"0.8",
"0.765",
"0.179",
"0.765"
],
[
"0.1765",
"0.787",
"0.29",
"0.787",
"0.29",
"0.8003",
"0.1765",
"0.8003"
],
[
"0.1765",
"0.8096",
"0.818",
"0.8096",
"0.818",
"0.8228",
"0.1765",
"0.8228"
],
[
"0.1936",
"0.8228",
"0.8228",
"0.8228",
"0.8228",
"0.836",
"0.1936",
"0.836"
],
[
"0.1929",
"0.837",
"0.581",
"0.837",
"0.581",
"0.8516",
"0.1929",
"0.8516"
],
[
"0.1748",
"0.86",
"0.8228",
"0.86",
"0.8228",
"0.875",
"0.1748",
"0.875"
],
[
"0.1912",
"0.875",
"0.6006",
"0.8726",
"0.6006",
"0.887",
"0.1912",
"0.8896"
],
[
"0.1748",
"0.8965",
"0.8228",
"0.8965",
"0.8228",
"0.911",
"0.1748",
"0.911"
],
[
"0.1936",
"0.912",
"0.675",
"0.912",
"0.675",
"0.9253",
"0.1936",
"0.9253"
],
[
"0.492",
"0.9478",
"0.509",
"0.9478",
"0.509",
"0.962",
"0.492",
"0.962"
]
] | [
"SCNN-constraint",
"HNN",
"SCNN",
"Experiment",
"Magnetic field",
"0.164±0.025",
"0.033±0.010",
"0.083±0.019",
"Spherical pendulum",
"0.092 ± 0.33",
"0.055±0.005",
"0.288±0.007",
"Double pendulum",
"0.014±0.004",
"0.117±0.012",
"Table l:The MSE after 100 time steps for different experiments. For the double pendulum",
"we compare the MSE after 20 time steps. The data points are collected for 10 different",
"initalizations and the error bars correspond to single standard deviation. All SCNN net-",
"works in these experiments have two conserved quantities. The best performing approach",
"is highlighted.",
"为了了解守恒量的数量和损失前置因子α的影响，我们对守恒量的数量和这些前置因子进",
"行了参数扫描。为了比较网络的精度，我们计算了100不同系统初始化的粒子轨迹，并将其",
"与100时间步长上的地面实况进行了比较。此外，我们检查了引力2体问题的能量守恒和",
"角动量守恒。我们的网络在与第一个50时间步长相对应的数据点上进行了训练，这使我们",
"可以比较我们的网络在相空间看不见的区域上的泛化。结果总结在图3中。总的来说，这",
"些结果表明SCNN学习了一",
"+个更加稳定和准确的哈密顿量，我们发现SCNN约束网络无法",
"超越SCNN网络。这些图中省略了基线网络，因为其性能明显更差。HNN网络具有与我们",
"的H。网络相同的复杂性，具有两个密集层，并且HNN_L网络对应于五个密集层的复杂",
"性，即。e。我们的T和H。网络的综合复杂性。损失预因子依赖性是这样的，我们使用",
"的守恒量越多，预因子就必须越小，这是由于损失项的增加",
"在具有磁场中的带电粒子，球形摆和双摆的设置中，与HNN相比，我们的SCNN（约束）",
"网络的性能得到了改善。表1显示了我们在100时间步长后获得的均方误差预测。这些系",
"统的描述可以在附录A中找到",
"3.1",
"Searchforconservedquantities我们的SCNN网络的守恒量可以通过将低维多项式ansatz",
"拟合到各自的网络预测来分析。这表明我们的SCNN发现角动量和总动量是引力两体系统",
"中的守恒量：",
"Pe=-4.2pr1-4.2Pr2-1.3pui-1.3Py2,",
"Pc2=-0.9pr1-0.9pr2-3.2Py1-3.2py2 ",
"L= 1.0 qriPu1 +0.9qriPy2 + 0.9 qraPy-1.0 qr2Py2",
"(6)",
"+1.0 qyiPr1 —0.9 QuiPr2- 0.9 qy2Pr1 +1.0 qy2Pa2",
"对于更复杂的守恒量（i。e。非多项式守恒量）不同的ansatze似乎是必要的（其中一些是",
"在相关工作中追求的(Saho0etal.,2018;Cranmeretal.,2019;Wetzeletal.,2020)",
"4OUTLOOKANDRELATED WORK",
"我们的SCNN以与HNN相同的方式自然地与诸如E(Battagliaetal.，2016)之类的神经网",
"络推断动力学的工作联系在一起。我们当前工作的自然扩展可以包括在基于图神经网络的",
"方法E(Sanchez-Gonzalezetal.，2019）和流程E（Tothetal.，2020）上的应用。关于应用，我",
"们门特别期待应用我们的新方法来自动推断和使用天体物理和分子动力学设置中的对称性。",
"REFERENCES",
"Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, et al. Interac",
"tion networks for learning about objects, relations and physics. In Advances in neural",
"information processing systems , pp. 4502-4510, 2016.",
"Miles D Cranmer, Rui Xu, Peter Battaglia, and Shirley Ho. Learning symbolic physics with",
"graph networks. arXiv preprint arXiv:1909.05862, 2019.",
"Sam Greydanus, Misko Dzamba, and Jason Yosinski. Hamiltonian neural networks. CoRR",
",abs/1906.01563,2019.URLhttp://arxiv.org/abs/1906.01563",
"4"
] | {
"arxiv_id": "2104.14444",
"language": "zh",
"page_idx": 3
} |
||
2104.14444-zh-4 | 2104.14444-zh-4.jpg | H. Rein and S. F. Liu. REBOUND: an open-source multi-purpose N-body code for collisional
dynamics. AAP ,537{:}A128, January 2012. doi: 10.1051/0004-6361/201118085
Subham Sahoo, Christoph Lampert, and Georg Martius. Learning equations for extrapo-
lation and control. In Jennifer Dy and Andreas Krause (eds.), Proceedings of the 35th $\begin{array}{c}\text{International Conference on Machine Learning, volume 80 of Proceedings of Machine}\\\text{International Conference on Machine}\\\text{A449 4450 C4111}\end{array}$ Learning Research , pp. 4442-4450, Stockholmsmässan, Stockholm Sweden, 10-15 Jul 2018. PMLR. URL http://proceedings.mlr.press/v80/sahoo18a.html.
Alvaro Sanchez-Gonzalez, Victor Bapst, Kyle Cranmer, and Peter Battaglia. Hamiltonian
graph networks with ode integrators. arXiv preprint arXiv:1909.12790 , 2019.
Peter Toth, Danilo J. Rezende, Andrew Jaegle, Sébastien Racanière, Aleksandar Botev, and $\operatorname*{Irina~Higgins.~Hamiltonian~generative~networks.~In~International~Conference~on~Learning}$ Representations ,2020. URL https://openreview.net/forum?id=HJenn6VFvB
Sebastian J. Wetzel, Roger G. Melko, Joseph Scott, Maysum Panju, and Vijay Ganesh Discovering symmetry invariants and conserved quantities by interpreting siamese neural ${\mathrm{networks,~2020.}}$
A DETAILS ON FURTHER PHYSICAL SYSTEMS
在这里，我们描述了我们在实验中使用的其他三个物理系统。
A.1 SPHERICAL PENDULUM
球形摆由以下哈密顿量控制：
$$\mathcal{H}=\frac{p_x^2+p_y^2}{2m}-m\:g\:\sqrt{l^2-q_x^2-q_y^2}\:,$$
(7)
,其中我们采用通用质量$m=\frac{1}{2}$耦合$g=\frac{1}{2}$和摆长$l=1$ 。在这个系统中，z 方向的角动
量是守恒的：
$$L=q_xp_y-q_yp_x\:.$$
(8)
我们从间隔(-1,1)的均匀分布中采样起始条件$\left(\mathbf{q}\left(\mathbf{t}=\mathbf{0}\right),\mathbf{p}\left(\mathbf{t}=\mathbf{0}\right)\right)$。我们确保能量是明
确的和负的
对于我们的数据集，我们使用 100 个初始条件并在轨迹上生成 500 个点。这些点取自 20的时间跨度，该时间跨度对应于包含多个周期性运动的时间。我们对训练集和测试集使用80:20 拆分
A.2 CHARGED PARTICLE IN A MAGNETIC FIELD
磁场中的带电粒子，其中我们选择圆形 B 场加上电场，由以下哈密顿量控制：
(9)
$$\mathcal{H}=\frac{1}{2m}\left(\left(p_x+q\:q_y\:B\right)^2+\left(p_y-q\:q_x\:B\right)^2\right)+k\:\left(q_x^2+q_y^2\right)\:,$$
,其中我们采用通用质量$m=\frac{1}{2}$ ,电场$k=1$ ,电荷$q=1$和 B 场 XMA3。再次角动量守恒。我们从区间$(-1,1)\bar{\text{的均匀分布中采样起始条件 }(\mathbf{q}(\mathbf{t}=\mathbf{0}),\mathbf{p}(\mathbf{t}=\mathbf{0}))}$ ,然后将向量归一化为单位长度。然后将其均匀地重新缩放为义间中的长度向最(0,1,1,0),对干我们的数抽集，我们利用 100 个初始条件并在轨迹上生成 500 个点。这些点取自 20 的时间跨度，该时间跨度对应于包含多个周期性运动的时间。我们对训练集和测试集使用 80:20 拆分
5 | [
[
"0.1732",
"0.1067",
"0.822",
"0.1067",
"0.822",
"0.1206",
"0.1732",
"0.1206"
],
[
"0.1953",
"0.1225",
"0.755",
"0.1225",
"0.755",
"0.1332",
"0.1953",
"0.1332"
],
[
"0.1748",
"0.1515",
"0.819",
"0.1528",
"0.819",
"0.1674",
"0.1748",
"0.1666"
],
[
"0.1953",
"0.1674",
"0.822",
"0.1674",
"0.822",
"0.1812",
"0.1953",
"0.1812"
],
[
"0.1929",
"0.1812",
"0.822",
"0.1812",
"0.822",
"0.1957",
"0.1929",
"0.1957"
],
[
"0.1929",
"0.1957",
"0.8228",
"0.1957",
"0.8228",
"0.209",
"0.1929",
"0.209"
],
[
"0.1936",
"0.209",
"0.7363",
"0.209",
"0.7363",
"0.2229",
"0.1936",
"0.2229"
],
[
"0.1772",
"0.2412",
"0.822",
"0.2412",
"0.822",
"0.2551",
"0.1772",
"0.2551"
],
[
"0.1912",
"0.2551",
"0.746",
"0.254",
"0.746",
"0.2683",
"0.1912",
"0.2695"
],
[
"0.1765",
"0.2874",
"0.8228",
"0.2874",
"0.8228",
"0.3005",
"0.1765",
"0.3005"
],
[
"0.1929",
"0.2998",
"0.8228",
"0.3018",
"0.8228",
"0.3164",
"0.1929",
"0.3145"
],
[
"0.1929",
"0.315",
"0.7686",
"0.315",
"0.7686",
"0.3289",
"0.1929",
"0.3289"
],
[
"0.1732",
"0.3452",
"0.8203",
"0.3467",
"0.8203",
"0.361",
"0.1732",
"0.3599"
],
[
"0.1929",
"0.3599",
"0.822",
"0.361",
"0.822",
"0.3757",
"0.1929",
"0.375"
],
[
"0.1929",
"0.3757",
"0.3057",
"0.3757",
"0.3057",
"0.3894",
"0.1929",
"0.3894"
],
[
"0.1765",
"0.4224",
"0.1978",
"0.4224",
"0.1978",
"0.435",
"0.1765",
"0.435"
],
[
"0.2157",
"0.4216",
"0.585",
"0.4224",
"0.585",
"0.4375",
"0.2157",
"0.4363"
],
[
"0.1765",
"0.4558",
"0.6045",
"0.4558",
"0.6045",
"0.4697",
"0.1765",
"0.4697"
],
[
"0.1748",
"0.4893",
"0.2214",
"0.4893",
"0.2214",
"0.504",
"0.1748",
"0.504"
],
[
"0.2214",
"0.4905",
"0.3938",
"0.4905",
"0.3938",
"0.504",
"0.2214",
"0.504"
],
[
"0.1748",
"0.518",
"0.3757",
"0.518",
"0.3757",
"0.532",
"0.1748",
"0.532"
],
[
"0.367",
"0.549",
"0.4631",
"0.5386",
"0.4673",
"0.561",
"0.3708",
"0.5713"
],
[
"0.4575",
"0.5474",
"0.625",
"0.552",
"0.625",
"0.5728",
"0.4575",
"0.5684"
],
[
"0.8022",
"0.5537",
"0.824",
"0.5537",
"0.824",
"0.5684",
"0.8022",
"0.5684"
],
[
"0.423",
"0.563",
"0.4478",
"0.563",
"0.4478",
"0.577",
"0.423",
"0.577"
],
[
"0.1887",
"0.587",
"0.8228",
"0.587",
"0.8228",
"0.6016",
"0.1887",
"0.6016"
],
[
"0.1772",
"0.603",
"0.2632",
"0.603",
"0.2632",
"0.6143",
"0.1772",
"0.6143"
],
[
"0.4346",
"0.627",
"0.564",
"0.631",
"0.564",
"0.646",
"0.4346",
"0.6416"
],
[
"0.8013",
"0.627",
"0.826",
"0.627",
"0.826",
"0.6445",
"0.8013",
"0.6445"
],
[
"0.1765",
"0.6636",
"0.8228",
"0.6636",
"0.8228",
"0.678",
"0.1765",
"0.678"
],
[
"0.1772",
"0.6787",
"0.2566",
"0.6787",
"0.2566",
"0.69",
"0.1772",
"0.69"
],
[
"0.1748",
"0.6978",
"0.822",
"0.6978",
"0.822",
"0.711",
"0.1748",
"0.711"
],
[
"0.1748",
"0.711",
"0.822",
"0.712",
"0.822",
"0.7266",
"0.1748",
"0.726"
],
[
"0.1765",
"0.728",
"0.2532",
"0.728",
"0.2532",
"0.7393",
"0.1765",
"0.7393"
],
[
"0.1748",
"0.761",
"0.539",
"0.761",
"0.539",
"0.7744",
"0.1748",
"0.7744"
],
[
"0.1765",
"0.789",
"0.7236",
"0.789",
"0.7236",
"0.803",
"0.1765",
"0.803"
],
[
"0.327",
"0.8174",
"0.3586",
"0.8174",
"0.3586",
"0.8257",
"0.327",
"0.8257"
],
[
"0.3545",
"0.8174",
"0.721",
"0.8184",
"0.721",
"0.839",
"0.3545",
"0.8384"
],
[
"0.8022",
"0.8213",
"0.824",
"0.8213",
"0.824",
"0.836",
"0.8022",
"0.836"
],
[
"0.312",
"0.8228",
"0.3245",
"0.821",
"0.326",
"0.8237",
"0.312",
"0.8257"
],
[
"0.2876",
"0.8228",
"0.3186",
"0.8228",
"0.3186",
"0.8354",
"0.2876",
"0.8354"
],
[
"0.3137",
"0.8306",
"0.3203",
"0.8306",
"0.3203",
"0.836",
"0.3137",
"0.836"
],
[
"0.3186",
"0.829",
"0.353",
"0.8306",
"0.3513",
"0.8423",
"0.3179",
"0.8403"
],
[
"0.179",
"0.855",
"0.822",
"0.855",
"0.822",
"0.8687",
"0.179",
"0.8687"
],
[
"0.1765",
"0.869",
"0.822",
"0.869",
"0.822",
"0.8833",
"0.1765",
"0.8833"
],
[
"0.1847",
"0.884",
"0.822",
"0.884",
"0.822",
"0.898",
"0.1847",
"0.898"
],
[
"0.1765",
"0.898",
"0.822",
"0.898",
"0.822",
"0.911",
"0.1765",
"0.911"
],
[
"0.1765",
"0.912",
"0.761",
"0.912",
"0.761",
"0.9253",
"0.1765",
"0.9253"
],
[
"0.492",
"0.9487",
"0.5073",
"0.9487",
"0.5073",
"0.964",
"0.492",
"0.964"
]
] | [
"H. Rein and S. F. Liu. REBOUND: an open-source multi-purpose N-body code for collisional",
"dynamics.AAP,537:A128,January2012.doi:10.1051/0004-6361/201118085",
"Subham Sahoo, Christoph Lampert, and Georg Martius. Learning equations for extrapo",
"lation and control. In Jennifer Dy and Andreas Krause (eds.), Proceedings of the 35th",
"International Conference on Machine Learning , volume 80 of Proceedings of Machine",
"LearningResearch,pp.4442-4450,Stockholmsmassan,Stockholm Sweden,10-15Jul",
"2018.PMLR. URL http://proceedings.mlr.press/v80/sahoo18a.html.",
"Alvaro Sanchez-Gonzalez, Victor Bapst, Kyle Cranmer, and Peter Battaglia. Hamiltonian",
"graph networks with ode integrators. arXiv preprint arXiv:1909.12790, 2019.",
"Peter Toth, Danilo J. Rezende, Andrew Jaegle, Sebastien Racaniere, Aleksandar Botev, and",
"Irina Higgins. Hamiltonian generative networks. In International Conference on Learning",
"Representations,2020.URL https://openreview.net/forum?id=HJenn6VFvB",
"Sebastian J. Wetzel, Roger G. Melko, Joseph Scott, Maysum Panju, and Vijay Ganesh",
"Discovering symmetry invariants and conserved quantities by interpreting siamese neura",
"networks, 2020.",
"A",
"DETAILSONFURTHERPHYSICAL SYSTEMS",
"在这里，我们描述了我们在实验中使用的其他三个物理系统。",
"A.1",
"SPHERICALPENDULUM",
"球形摆由以下哈密顿量控制",
"H=星+P",
"-mgV12-呢-",
"(7)",
"2m",
"其中我们采用通用质量m=耦合9=和摆长1=1。在这个系统中，z方向的角动",
"量是守恒的：",
"L = qaPy- QyPa :",
"(8)",
"我们从间隔（-1,1）的均匀分布中采样起始条件（q(t=0），p(t=0)）。我们确保能量是明",
"确的和负的",
"对于我们的数据集，我们使用100个初始条件并在轨迹上生成500个点。这些点取自20",
"的时间跨度，该时间跨度对应于包含多个周期性运动的时间。我们对训练集和测试集使用",
"80:20拆分",
"A.2CHARGED PARTICLE IN A MAGNETIC FIELD",
"磁场中的带电粒子，其中我们选择圆形B场加上电场，由以下哈密顿量控制：",
"1(",
"((pP+qQyB)+(py-qQaB))+(d+)",
"(9)",
"",
"H=",
"",
"2m",
"，其中我们采用通用质量m=去，电场k=1，电荷q=1和B场XMA3。再次角动量守",
"恒。我们从区间（-1,1）的均匀分布中采样起始条件（q(t=0),P(t=0)），然后将向量归",
"-化为单位长度，然后将其均匀地重新缩放为区间中的长度向量（0.1,1.0).对于我们的数据",
"集，我们利用100个初始条件并在轨迹上生成500个点。这些点取自20的时间跨度，该时",
"间跨度对应于包含多个周期性运动的时间。我们对训练集和测试集使用80:20拆分",
"5"
] | {
"arxiv_id": "2104.14444",
"language": "zh",
"page_idx": 4
} |
||
2109.07242-zh-1 | 2109.07242-zh-1.jpg | 用于机器翻译质量评估的回归集成
Michal tefánik and Vít Novotný and Petr Sojka
Masaryk University, Faculty of Informatics, MIR group https://mir.fi.muni.cz
$\{$ stefanik.m,witiko $\}$ @mail.muni.cz,[email protected]
# Abstract
This work introduces a simple regressive ensemble for evaluating machine translation quality based on a set of novel and established metrics. We evaluate the ensemble using a correlation to expert-based MQM scores of the WMT 2021 Metrics workshop. In both monolingual and zero-shot cross-lingual settings, we show a significant performance improvement over single metrics. In the crosslingual settings, we also demonstrate that an ensemble approach is well-applicable to unseen languages. Furthermore, we identify a strong reference-free baseline that consistently outperforms the commonly-used BLEU and METEOR measures and significantly improves our ensemble's performance.
## 1 Introduction
由于用户期望从文本生成系统获得许多正交的定性方面，因此自动评估文本生成具有挑战性。在机器翻译中，我们可以观察到所谓的关键类别 (Wulczyn et al., 2017) 的错误，例如幻觉(Lee et al., 2019) , 省略部分翻译输入或否定(Matusov, 2019) 的含义
考虑最后一个类别的示例：
$\bullet$ Reference: “I never wrote this article, I just
edited it."
$\bullet Hypothesis~I:~“$It is not my article, I just
edited it."
$\bullet Hypothesis2:“I$ never wrote this article, I
never edited it."
在此示例中，所有 BERTScore (Zhang et al., 2019)、BLEUrt (Sellam et al.,2020) 和 Prism (Thompson and Post, 2020b) 指标的排名 Hypothesis 2 都高于 Hypothesis l 。在 BLEUrt 和 Prism 中，这可能是由于 Transformers 的已知漏洞造成的，如果这样的启发式方法可以很好地解决问题，则它依赖于词汇交集 (McCoy et al., 2019) 。琐碎的是，仅计算否定就可以轻松解决此特定问题。但是，这样的启发式方法在许多其他
情况下会失败，例如当我们将Hypothesis $l$调
整为“它是别人的文章，我刚刚编辑了它。”
这样的案例激发了我们的整体方法，该方法旨在揭示文本的表面和更深层次的语义属性， 并随后学习将其用于翻译评估的特定任务。即使目标可能会限制特定度量、数据集，或者在某些情况下系统性地失败，集合中的另一个度量或它们的组合也可以纠正缺陷
## 2 Metrics for machine translation evaluation
本节回顾相关工作，重点介绍我们在合奏中使
用的指标
用于评估机器翻译质量的标准且仍广泛使用的表面级指标是$BLEU$ (Papineni et al., 2002) , $ROUGE\left(\mathrm{Lin},2004\right)$和$TER\left(\mathrm{Snover~et~al.},2006\right)$。在一个文本解释另一个文本的情况下，表面级别的度量无法捕获含义的接近度，这是在深度神经语言模型 (Lewis et al., 2020) 中通常观察到的能力。解决此缺陷的一个度量是$METEOR$ (Banerjee and Lavie, 2005) , 它利用 WordNet 来解释同义词、单词变形或令牌级释义
语义文本等效性的评估与准确的文本表示(嵌入)问题密切相关。我们认为与评估分段大小的文本相关的传统方法是 FastText (Bojanowski et al., 2017) 。FastText 学习字符$n$ -gram 的表示，从中通过平均创建标记的统一表示。此外，可以使用软向量空间模型(软余弦度量， $SCM$ )(Novotný, 2018) 等方法直接从令牌级嵌人计算一对文本的距离，或者通过解决最小成本流问题(单词移动器的距离，$WMD)$(Kusner et al.,2015)
BERTScore (Zhang et al., 2019) 执行类似的匹配，它使用针对任务最佳的所选 BERT 层的内部令牌嵌入。虽然令牌表示在某些模型 (Devlin et al., 2018) 中是多语言的，这使得BERTScore 可以在没有引用的情况下使用，但我们不知道以前的工作是这样评估它。使用BERTScore 的最大引用匹配方案进行跨语言对齐的一个可能缺点是，源文本和目标文本中的子单词标记可能会严重不匹配。相反，我们称为 WMD- | [
[
"0.3208",
"0.0968",
"0.679",
"0.0968",
"0.679",
"0.10986",
"0.3208",
"0.10986"
],
[
"0.294",
"0.1455",
"0.7095",
"0.1455",
"0.7095",
"0.1603",
"0.294",
"0.1603"
],
[
"0.1604",
"0.1609",
"0.8423",
"0.1621",
"0.8423",
"0.177",
"0.1604",
"0.1758"
],
[
"0.2377",
"0.1776",
"0.777",
"0.1787",
"0.777",
"0.1936",
"0.2377",
"0.1924"
],
[
"0.2644",
"0.2542",
"0.34",
"0.2542",
"0.34",
"0.2683",
"0.2644",
"0.2683"
],
[
"0.5146",
"0.2546",
"0.881",
"0.2546",
"0.881",
"0.2683",
"0.5146",
"0.2683"
],
[
"0.511",
"0.2695",
"0.858",
"0.2695",
"0.858",
"0.2832",
"0.511",
"0.2832"
],
[
"0.1453",
"0.2798",
"0.4568",
"0.2798",
"0.4568",
"0.2935",
"0.1453",
"0.2935"
],
[
"0.529",
"0.2861",
"0.879",
"0.2861",
"0.879",
"0.2998",
"0.529",
"0.2998"
],
[
"0.147",
"0.2944",
"0.4575",
"0.2944",
"0.4575",
"0.308",
"0.147",
"0.308"
],
[
"0.147",
"0.31",
"0.4568",
"0.31",
"0.4568",
"0.3206",
"0.147",
"0.3206"
],
[
"0.5156",
"0.3047",
"0.876",
"0.3047",
"0.876",
"0.3154",
"0.5156",
"0.3154"
],
[
"0.1453",
"0.3218",
"0.4575",
"0.3225",
"0.4575",
"0.3374",
"0.1453",
"0.3362"
],
[
"0.513",
"0.3196",
"0.881",
"0.3196",
"0.881",
"0.3333",
"0.513",
"0.3333"
],
[
"0.1486",
"0.3384",
"0.4575",
"0.3384",
"0.4575",
"0.3486",
"0.1486",
"0.3486"
],
[
"0.5146",
"0.335",
"0.881",
"0.335",
"0.881",
"0.3486",
"0.5146",
"0.3486"
],
[
"0.1453",
"0.3508",
"0.4575",
"0.3508",
"0.4575",
"0.364",
"0.1453",
"0.364"
],
[
"0.513",
"0.3516",
"0.881",
"0.3516",
"0.881",
"0.3652",
"0.513",
"0.3652"
],
[
"0.147",
"0.3652",
"0.4575",
"0.3652",
"0.4575",
"0.379",
"0.147",
"0.379"
],
[
"0.5146",
"0.3694",
"0.769",
"0.3694",
"0.769",
"0.3801",
"0.5146",
"0.3801"
],
[
"0.147",
"0.3813",
"0.4568",
"0.3801",
"0.4568",
"0.3914",
"0.147",
"0.3926"
],
[
"0.1453",
"0.3943",
"0.4568",
"0.3938",
"0.4568",
"0.4067",
"0.1453",
"0.408"
],
[
"0.51",
"0.3943",
"0.8213",
"0.3955",
"0.8213",
"0.4104",
"0.51",
"0.4092"
],
[
"0.147",
"0.4092",
"0.4568",
"0.408",
"0.4568",
"0.4192",
"0.147",
"0.4204"
],
[
"0.544",
"0.4146",
"0.6323",
"0.4146",
"0.6323",
"0.4258",
"0.544",
"0.4258"
],
[
"0.147",
"0.4229",
"0.4575",
"0.4229",
"0.4575",
"0.4358",
"0.147",
"0.4358"
],
[
"0.1453",
"0.437",
"0.4575",
"0.437",
"0.4575",
"0.4507",
"0.1453",
"0.4507"
],
[
"0.513",
"0.4382",
"0.882",
"0.4382",
"0.882",
"0.452",
"0.513",
"0.452"
],
[
"0.1486",
"0.4531",
"0.4568",
"0.4531",
"0.4568",
"0.4631",
"0.1486",
"0.4631"
],
[
"0.513",
"0.4536",
"0.5884",
"0.4536",
"0.5884",
"0.4685",
"0.513",
"0.4685"
],
[
"0.1453",
"0.4644",
"0.4592",
"0.4631",
"0.4592",
"0.478",
"0.1453",
"0.4785"
],
[
"0.534",
"0.4702",
"0.881",
"0.4702",
"0.881",
"0.484",
"0.534",
"0.484"
],
[
"0.1436",
"0.478",
"0.4575",
"0.4785",
"0.4575",
"0.4934",
"0.1436",
"0.4922"
],
[
"0.513",
"0.4863",
"0.885",
"0.4863",
"0.885",
"0.4995",
"0.513",
"0.4995"
],
[
"0.1436",
"0.4922",
"0.315",
"0.4934",
"0.315",
"0.5083",
"0.1436",
"0.5073"
],
[
"0.511",
"0.502",
"0.881",
"0.503",
"0.881",
"0.5166",
"0.511",
"0.5156"
],
[
"0.147",
"0.5186",
"0.2627",
"0.5186",
"0.2627",
"0.533",
"0.147",
"0.533"
],
[
"0.513",
"0.5186",
"0.881",
"0.5186",
"0.881",
"0.532",
"0.513",
"0.532"
],
[
"0.5156",
"0.536",
"0.879",
"0.536",
"0.879",
"0.5464",
"0.5156",
"0.5464"
],
[
"0.11755",
"0.5215",
"0.1503",
"0.5215",
"0.1503",
"0.53",
"0.11755",
"0.53"
],
[
"0.12006",
"0.5444",
"0.486",
"0.5444",
"0.486",
"0.558",
"0.12006",
"0.558"
],
[
"0.5146",
"0.5503",
"0.881",
"0.5503",
"0.881",
"0.564",
"0.5146",
"0.564"
],
[
"0.1184",
"0.561",
"0.4836",
"0.561",
"0.4836",
"0.572",
"0.1184",
"0.572"
],
[
"0.513",
"0.568",
"0.881",
"0.568",
"0.881",
"0.5806",
"0.513",
"0.5806"
],
[
"0.1184",
"0.5767",
"0.486",
"0.5767",
"0.486",
"0.5903",
"0.1184",
"0.5903"
],
[
"0.513",
"0.583",
"0.881",
"0.583",
"0.881",
"0.5967",
"0.513",
"0.5967"
],
[
"0.1184",
"0.593",
"0.4878",
"0.593",
"0.4878",
"0.6055",
"0.1184",
"0.6055"
],
[
"0.513",
"0.5986",
"0.806",
"0.5986",
"0.806",
"0.6123",
"0.513",
"0.6123"
],
[
"0.1184",
"0.608",
"0.486",
"0.608",
"0.486",
"0.6226",
"0.1184",
"0.6226"
],
[
"0.531",
"0.615",
"0.881",
"0.615",
"0.881",
"0.629",
"0.531",
"0.629"
],
[
"0.1184",
"0.6245",
"0.3005",
"0.6245",
"0.3005",
"0.638",
"0.1184",
"0.638"
],
[
"0.5146",
"0.633",
"0.879",
"0.633",
"0.879",
"0.6436",
"0.5146",
"0.6436"
],
[
"0.1351",
"0.6396",
"0.346",
"0.64",
"0.346",
"0.655",
"0.1351",
"0.654"
],
[
"0.5146",
"0.648",
"0.882",
"0.648",
"0.882",
"0.661",
"0.5146",
"0.661"
],
[
"0.511",
"0.662",
"0.879",
"0.662",
"0.879",
"0.676",
"0.511",
"0.676"
],
[
"0.1377",
"0.673",
"0.4895",
"0.673",
"0.4895",
"0.686",
"0.1377",
"0.686"
],
[
"0.5146",
"0.6787",
"0.882",
"0.6787",
"0.882",
"0.6924",
"0.5146",
"0.6924"
],
[
"0.1528",
"0.6885",
"0.2301",
"0.6885",
"0.2301",
"0.703",
"0.1528",
"0.703"
],
[
"0.5146",
"0.6953",
"0.8833",
"0.6953",
"0.8833",
"0.709",
"0.5146",
"0.709"
],
[
"0.1368",
"0.705",
"0.4895",
"0.705",
"0.4895",
"0.7197",
"0.1368",
"0.7197"
],
[
"0.511",
"0.7104",
"0.882",
"0.7114",
"0.882",
"0.7256",
"0.511",
"0.7246"
],
[
"0.1528",
"0.7227",
"0.2275",
"0.7227",
"0.2275",
"0.734",
"0.1528",
"0.734"
],
[
"0.5146",
"0.728",
"0.882",
"0.728",
"0.882",
"0.741",
"0.5146",
"0.741"
],
[
"0.1368",
"0.737",
"0.4895",
"0.737",
"0.4895",
"0.752",
"0.1368",
"0.752"
],
[
"0.5146",
"0.7446",
"0.881",
"0.7446",
"0.881",
"0.7583",
"0.5146",
"0.7583"
],
[
"0.1528",
"0.753",
"0.2737",
"0.753",
"0.2737",
"0.7666",
"0.1528",
"0.7666"
],
[
"0.5146",
"0.761",
"0.601",
"0.761",
"0.601",
"0.7725",
"0.5146",
"0.7725"
],
[
"0.1184",
"0.7754",
"0.486",
"0.7754",
"0.486",
"0.789",
"0.1184",
"0.789"
],
[
"0.529",
"0.7754",
"0.881",
"0.7754",
"0.881",
"0.7905",
"0.529",
"0.7905"
],
[
"0.1184",
"0.792",
"0.486",
"0.792",
"0.486",
"0.8057",
"0.1184",
"0.8057"
],
[
"0.513",
"0.792",
"0.881",
"0.792",
"0.881",
"0.8057",
"0.513",
"0.8057"
],
[
"0.1184",
"0.808",
"0.4895",
"0.808",
"0.4895",
"0.822",
"0.1184",
"0.822"
],
[
"0.5146",
"0.808",
"0.881",
"0.808",
"0.881",
"0.822",
"0.5146",
"0.822"
],
[
"0.1184",
"0.8247",
"0.4844",
"0.8247",
"0.4844",
"0.8384",
"0.1184",
"0.8384"
],
[
"0.5146",
"0.8247",
"0.881",
"0.8247",
"0.881",
"0.8384",
"0.5146",
"0.8384"
],
[
"0.12006",
"0.843",
"0.4836",
"0.843",
"0.4836",
"0.8525",
"0.12006",
"0.8525"
],
[
"0.5146",
"0.8413",
"0.881",
"0.8413",
"0.881",
"0.855",
"0.5146",
"0.855"
],
[
"0.12006",
"0.857",
"0.486",
"0.857",
"0.486",
"0.8706",
"0.12006",
"0.8706"
],
[
"0.511",
"0.857",
"0.879",
"0.857",
"0.879",
"0.8706",
"0.511",
"0.8706"
],
[
"0.1184",
"0.8726",
"0.486",
"0.8726",
"0.486",
"0.8857",
"0.1184",
"0.8857"
],
[
"0.513",
"0.8735",
"0.881",
"0.8735",
"0.881",
"0.887",
"0.513",
"0.887"
],
[
"0.1351",
"0.8896",
"0.486",
"0.8896",
"0.486",
"0.903",
"0.1351",
"0.903"
],
[
"0.513",
"0.8896",
"0.882",
"0.8896",
"0.882",
"0.903",
"0.513",
"0.903"
],
[
"0.11755",
"0.905",
"0.486",
"0.905",
"0.486",
"0.9185",
"0.11755",
"0.9185"
],
[
"0.513",
"0.905",
"0.882",
"0.905",
"0.882",
"0.9185",
"0.513",
"0.9185"
]
] | [
"用于机器翻译质量评估的回归集成",
"Michal tefanik and Vit Novotny and Petr Sojka",
"Masaryk University, Faculty of Informatics, MIR group https:/ /mir.fi.muni.cz",
"{ stefanik.m,witiko @mail.muni.cz, [email protected]",
"Abstract",
"情况下会失败，例如当我们将Hypothesis1调",
"整为“它是别人的文章，我刚刚编辑了它。",
"This work introduces a simple regressive en",
"这样的案例激发了我们的整体方法，该方法",
"semble for evaluating machine translation qual-",
"ity based on a set of novel and established",
"旨在揭示文本的表面和更深层次的语义属性",
"metrics.We evaluate the ensemble using a",
"并随后学习将其用于翻译评估的特定任务。即",
"correlationtoexpert-basedMQM scores of",
"使目标可能会限制特定度量、数据集，或者在",
"the WMT 2021 Metrics workshop.In both",
"某些情况下系统性地失败，集合中的另一个度",
"monolingual and zero-shot cross-lingual set",
"量或它们的组合也可以纠正缺陷",
"tings, we show a significant performance im",
"provement over single metrics. In the cross-",
"2 Metrics for machine translation",
"lingual settings, we alsodemonstrate that an en.",
"evaluation",
"semble approach is well-applicable to unseen",
"languages. Furthermore, we identify a strong",
"本节回顾相关工作，重点介绍我们在合奏中使",
"reference-free baselinethat consistently out",
"用的指标",
"performs the commonly-used BLEU and ME-",
"用于评估机器翻译质量的标准且仍广泛使用",
"TEOR measures and significantly improves our",
"的表面级指标是BLEU（Papinenietal.，2002），",
"ensemble's performance.",
"ROUGE (Lin, 2004)和 TER(Snover et al., 2006)。",
"Introduction",
"在一个文本解释另一个文本的情况下，表面级",
"别的度量无法捕获含义的接近度，这是在深度",
"1",
"由于用户期望从文本生成系统获得许多正交",
"神经语言模型（Lewisetal.,2020）中通常观察",
"的定性方面，因此自动评估文本生成具有挑战",
"到的能力。解决此缺陷的一个度量是METEOR",
"性。在机器翻译中，我们可以观察到所谓的关",
"(Banerjeeand Lavie,2005），它利用WordNet来",
"键类别（Wulczynetal.,2017）的错误，例如幻",
"解释同义词、单词变形或令牌级释义",
"觉（Leeetal.,2019），省略部分翻译输入或否定",
"语义文本等效性的评估与准确的文本表示",
"(Matusov,2019）的含义",
"（嵌入）问题密切相关。我们认为与评估分",
"考虑最后一个类别的示例：",
"段大小的文本相关的传统方法是FastText(Bo-",
"janowskietal.2017）。FastText学习字符n-gram",
": Reference: “I never wrote this article, I just",
"的表示，从中通过平均创建标记的统一表示。",
"edited it.\"",
"此外，可以使用软向量空间模型（软余弦度量，",
". Hypothesis I: \"It is not my article, I just",
"SCM）（Novotny，2018）等方法直接从令牌级嵌",
"edited it.\"",
"入计算一对文本的距离，或者通过解决最小成",
": Hypothesis 2:“I never wrote this article, I",
"本流问题（单词移动器的距离，WMD）（Kusner",
"never edited it.\"",
"et al.,2015)",
"在此示例中，所有BERTScore(Zhang etal.,",
"BERTScore（Zhangetal.,2019）执行类似的匹",
"2019)BLEUrt (Sellam et al.,2020)和 Prism",
"配，它使用针对任务最佳的所选BERT层的内",
"（ThompsonandPost，2020b）指标的排名Hypoth-",
"部令牌嵌入。虽然令牌表示在某些模型（Devlin",
"esis2都高于Hypothesis1。在BLEUrt和Prism",
"etal.，2018）中是多语言的，这使得BERTScore",
"中，这可能是由于Transformers的已知漏洞造",
"可以在没有引用的情况下使用，但我们不知",
"成的，如果这样的启发式方法可以很好地解决",
"道以前的工作是这样评估它。使用BERTScore",
"问题，则它依赖于词汇交集（McCoyetal.,2019)",
"的最大引用匹配方案进行跨语言对齐的一个",
"琐碎的是，仅计算否定就可以轻松解决此特",
"可能缺点是，源文本和目标文本中的子单词标",
"定问题。但是，这样的启发式方法在许多其他",
"记可能会严重不匹配。相反，我们称为WMD"
] | {
"arxiv_id": "2109.07242",
"language": "zh",
"page_idx": 1
} |
||
2109.07242-zh-2 | 2109.07242-zh-2.jpg | contextual的度量使用与BERTScore相同的嵌
人，但使用$WMD$的网络流量优化匹配方案
最近，与任务无关的方法已被针对相关目标对预先训练的模型进行微调的方法所超越： BLEUrt (Sellam et al., 2020) 直接在直接评估提交给 WMT 的$BERT$ (Devlin et al., 2018) 模型上进行微调，以使用线性头部对分类[CLS]令牌的上下文嵌人进行预测。Comet (Rei et al., 2020) 学习从具有三元组目标或标准 MSE 目标的源， 参考和翻译文本的元组中预测直接评估
最近的一些工作包含潜在的目标和/或数据集。例如，Prism (Thompson and Post, 2020b,a) 从 39El语言的多语言释义中学习了与语言无关的表示形式，因此是为数不多的性能良好的无参考指标之一。其训练目标的正交性可能会降低其与使用上下文嵌人的其他方法的相关性
# 3 $\textbf{Methodology}$
我们的方法旨在通过其他支持问题回答以下主
要问题：
1. Can an ensemble of surface, syntactic, and
semantic-level metrics significantly improve
the performance of single metrics?
2.Can such an approach be applied cross-
lingually, i.e., on languages that it has not
been trained on?
3. Can surface-level metrics in reference-free
configuration achieve results comparable to
the reference-based ones?
4. Are contextual token representations impor-
tant for evaluating semantic equivalence, or can these be replaced with pre-inferred token representations?
# 3.1 Experimental setup
我们对多维质量度量(MQM)数据集(Freitag et al., 2021) 执行主要评估，其中我们使用平均判断分数作为黄金标准。如果给定的一对来源和假设有多个判断，我们将判断的分数平均并将此平均值视为我们的黄金标准。我们根据独特的源文本将样本分为训练(80 个%)和测试(20 个$\%)$ 子集
在我们的实验框架中，我们以开源 Python 库和 Docker 映像的形式发布，以便于复制$^{12}$ 我们根据它们的准则实现了一组选定的指标， 以及在$\overline{\mathrm{E}}$3.2 节中引人的一系列新颖指标，旨在提供文本等效性的其他正交见解
$^{1}$https: / / github. com/ MIR- MU/ regemt
$^{2}$https://hub.docker.com/r/miratmu/regemt
随后，我们在整个火车集的 standardized 度量特征上训练回归集成，旨在预测平均 MQM 专家判断。我们使用成对的 Spearman 等级相关性(Spearman 的$\rho$ )和对 20 个% 测试拆分的 MQM 判断来评估总体以及所有其他选定的指标。
除了我们对 MQM 数据集的主要评估之外，我们还对 WMT2015 和 2016 的直接评估(DA)以及来自编辑后数据集 (Fomicheva et al., 2020a) 的一组灾难性错误进行了实验。多语言质量估计数据集( MLQE-PE) (Fomicheva et al., 2020b)用于 WMT2021 质量估计研讨会的评估。有关我们在 DA 和 MLQE-PE 上的实验的详细说明，请参阅$\mathbb{E}$3.6 节
我们对分段级别的判断进行了所有评估。为了最大程度地减少每个特定指标的校准对评估的影响，我们报告了 Spearman 等级相关系数， 反映了相互的定性顺序，而不是判断的特定值
## 3.2 Novel metrics
除了基于文献综述的一组选定指标外，我们还实现了一组新颖的指标，使我们的整体能够反映出评估文本的更广泛的属性差异
## 3.2.1 Soft Cosine Measure
软余弦度量$(SCM)$ (Novotný,2018)是软向量空间模型中文本的余弦相似度，其中术语的轴处于与它们的余弦相似度$\overline{\mathrm{E}}S$ 相对应的角度在令牌嵌人空间中：
(1)
$$\mathrm{SCM}(\vec{x},\vec{y})=\frac{\vec{x}^\mathsf{T}\cdot S\cdot\vec{y}}{\sqrt{\vec{x}^\mathsf{T}\cdot S\cdot\vec{x}}\cdot\sqrt{\vec{y}^\mathsf{T}\cdot S\cdot\vec{y}}}$$
其中$\vec{x}$是一个加权的 Boa 参考向量(BoW)。
我们将 SCM 与两个令牌表示一起使用：
1. We use the static token representations of Fast-
Text (Grave et al., 2018). We refer to the re-
sulting metric as SCM .
2. We use the contextual token representations
of $BERT$ (Devlin et al., 2018) using the methodology of BERTScore (Zhang et al., 2019) . We collect representations of all tokens segmented by the WordPiece (Wu et al., 2016) tokenizer, and we treat each unique (token, context) pair as a single term in our vocabulary. Subsequently, we decontextualize these representations as follows: For each WordPiece token, we average the representations of all (token, context) pairs in the training corpus. We refer to the resulting metric | [
[
"0.1184",
"0.0873",
"0.4878",
"0.0873",
"0.4878",
"0.10095",
"0.1184",
"0.10095"
],
[
"0.531",
"0.0873",
"0.881",
"0.0873",
"0.881",
"0.10095",
"0.531",
"0.10095"
],
[
"0.12006",
"0.1027",
"0.465",
"0.1027",
"0.465",
"0.1164",
"0.12006",
"0.1164"
],
[
"0.5146",
"0.1027",
"0.881",
"0.1027",
"0.881",
"0.1164",
"0.5146",
"0.1164"
],
[
"0.1377",
"0.12054",
"0.482",
"0.12054",
"0.482",
"0.1312",
"0.1377",
"0.1312"
],
[
"0.5146",
"0.11816",
"0.881",
"0.11816",
"0.881",
"0.1318",
"0.5146",
"0.1318"
],
[
"0.1184",
"0.1354",
"0.4878",
"0.1354",
"0.4878",
"0.1484",
"0.1184",
"0.1484"
],
[
"0.513",
"0.1342",
"0.879",
"0.1342",
"0.879",
"0.1478",
"0.513",
"0.1478"
],
[
"0.11755",
"0.1497",
"0.4844",
"0.1509",
"0.4844",
"0.165",
"0.11755",
"0.1644"
],
[
"0.513",
"0.1509",
"0.881",
"0.1509",
"0.881",
"0.1644",
"0.513",
"0.1644"
],
[
"0.1184",
"0.1675",
"0.486",
"0.1675",
"0.486",
"0.1812",
"0.1184",
"0.1812"
],
[
"0.511",
"0.1675",
"0.5576",
"0.1675",
"0.5576",
"0.1818",
"0.511",
"0.1818"
],
[
"0.1184",
"0.1829",
"0.486",
"0.1829",
"0.486",
"0.1965",
"0.1184",
"0.1965"
],
[
"0.531",
"0.1841",
"0.879",
"0.1841",
"0.879",
"0.1978",
"0.531",
"0.1978"
],
[
"0.12006",
"0.2007",
"0.4836",
"0.2007",
"0.4836",
"0.2114",
"0.12006",
"0.2114"
],
[
"0.511",
"0.1996",
"0.879",
"0.1996",
"0.879",
"0.2131",
"0.511",
"0.2131"
],
[
"0.11755",
"0.2156",
"0.4878",
"0.2156",
"0.4878",
"0.2286",
"0.11755",
"0.2286"
],
[
"0.513",
"0.2162",
"0.881",
"0.2162",
"0.881",
"0.2299",
"0.513",
"0.2299"
],
[
"0.11584",
"0.2322",
"0.43",
"0.2322",
"0.43",
"0.2452",
"0.11584",
"0.2452"
],
[
"0.511",
"0.231",
"0.881",
"0.2328",
"0.881",
"0.2465",
"0.511",
"0.2446"
],
[
"0.1368",
"0.2477",
"0.4878",
"0.2477",
"0.4878",
"0.2612",
"0.1368",
"0.2612"
],
[
"0.513",
"0.2488",
"0.879",
"0.2488",
"0.879",
"0.2625",
"0.513",
"0.2625"
],
[
"0.11755",
"0.2625",
"0.4878",
"0.2632",
"0.4878",
"0.2778",
"0.11755",
"0.2766"
],
[
"0.511",
"0.2642",
"0.881",
"0.2642",
"0.881",
"0.2778",
"0.511",
"0.2778"
],
[
"0.1184",
"0.2798",
"0.486",
"0.2798",
"0.486",
"0.2935",
"0.1184",
"0.2935"
],
[
"0.513",
"0.2798",
"0.881",
"0.2798",
"0.881",
"0.2944",
"0.513",
"0.2944"
],
[
"0.12006",
"0.2957",
"0.4878",
"0.2957",
"0.4878",
"0.3088",
"0.12006",
"0.3088"
],
[
"0.513",
"0.2964",
"0.7246",
"0.2964",
"0.7246",
"0.31",
"0.513",
"0.31"
],
[
"0.11755",
"0.31",
"0.4836",
"0.31",
"0.4836",
"0.3237",
"0.11755",
"0.3237"
],
[
"0.531",
"0.3135",
"0.881",
"0.3135",
"0.881",
"0.3267",
"0.531",
"0.3267"
],
[
"0.11755",
"0.328",
"0.4878",
"0.328",
"0.4878",
"0.3416",
"0.11755",
"0.3416"
],
[
"0.513",
"0.328",
"0.879",
"0.328",
"0.879",
"0.3416",
"0.513",
"0.3416"
],
[
"0.5146",
"0.3467",
"0.879",
"0.3467",
"0.879",
"0.357",
"0.5146",
"0.357"
],
[
"0.1142",
"0.354",
"0.2644",
"0.3557",
"0.2644",
"0.3706",
"0.1142",
"0.3682"
],
[
"0.513",
"0.3623",
"0.882",
"0.3623",
"0.882",
"0.376",
"0.513",
"0.376"
],
[
"0.11755",
"0.3801",
"0.486",
"0.3813",
"0.486",
"0.3955",
"0.11755",
"0.3943"
],
[
"0.511",
"0.3884",
"0.665",
"0.3896",
"0.665",
"0.4038",
"0.511",
"0.4026"
],
[
"0.11755",
"0.3967",
"0.1823",
"0.3967",
"0.1823",
"0.411",
"0.11755",
"0.411"
],
[
"0.5146",
"0.412",
"0.881",
"0.412",
"0.881",
"0.4258",
"0.5146",
"0.4258"
],
[
"0.1276",
"0.4258",
"0.1561",
"0.4258",
"0.1561",
"0.434",
"0.1276",
"0.434"
],
[
"0.1511",
"0.4229",
"0.486",
"0.4229",
"0.486",
"0.4358",
"0.1511",
"0.4358"
],
[
"0.513",
"0.4287",
"0.879",
"0.4287",
"0.879",
"0.4424",
"0.513",
"0.4424"
],
[
"0.1528",
"0.4382",
"0.4878",
"0.4395",
"0.4878",
"0.4536",
"0.1528",
"0.4531"
],
[
"0.513",
"0.4448",
"0.7886",
"0.4448",
"0.7886",
"0.4578",
"0.513",
"0.4578"
],
[
"0.1545",
"0.4548",
"0.413",
"0.4548",
"0.413",
"0.4697",
"0.1545",
"0.4697"
],
[
"0.513",
"0.4714",
"0.735",
"0.4714",
"0.735",
"0.485",
"0.513",
"0.485"
],
[
"0.1301",
"0.481",
"0.4836",
"0.481",
"0.4836",
"0.4954",
"0.1301",
"0.4954"
],
[
"0.1545",
"0.4976",
"0.4895",
"0.4976",
"0.4895",
"0.5127",
"0.1545",
"0.5127"
],
[
"0.513",
"0.4922",
"0.881",
"0.4922",
"0.881",
"0.506",
"0.513",
"0.506"
],
[
"0.1511",
"0.5127",
"0.2805",
"0.513",
"0.2805",
"0.528",
"0.1511",
"0.527"
],
[
"0.5146",
"0.5103",
"0.879",
"0.5103",
"0.879",
"0.521",
"0.5146",
"0.521"
],
[
"0.5146",
"0.525",
"0.881",
"0.525",
"0.881",
"0.538",
"0.5146",
"0.538"
],
[
"0.1293",
"0.539",
"0.486",
"0.539",
"0.486",
"0.5527",
"0.1293",
"0.5527"
],
[
"0.513",
"0.5405",
"0.65",
"0.5405",
"0.65",
"0.554",
"0.513",
"0.554"
],
[
"0.1545",
"0.557",
"0.486",
"0.557",
"0.486",
"0.571",
"0.1545",
"0.571"
],
[
"0.1528",
"0.571",
"0.346",
"0.572",
"0.346",
"0.586",
"0.1528",
"0.585"
],
[
"0.702",
"0.568",
"0.7793",
"0.5693",
"0.7793",
"0.5845",
"0.702",
"0.582"
],
[
"0.538",
"0.5767",
"0.6406",
"0.5806",
"0.6396",
"0.5957",
"0.5366",
"0.5913"
],
[
"0.858",
"0.58",
"0.882",
"0.58",
"0.882",
"0.5933",
"0.858",
"0.5933"
],
[
"0.644",
"0.586",
"0.8345",
"0.5874",
"0.8345",
"0.605",
"0.644",
"0.604"
],
[
"0.1301",
"0.5986",
"0.486",
"0.5986",
"0.486",
"0.6123",
"0.1301",
"0.6123"
],
[
"0.1528",
"0.6133",
"0.486",
"0.615",
"0.486",
"0.63",
"0.1528",
"0.6274"
],
[
"0.5156",
"0.6216",
"0.8623",
"0.6216",
"0.8623",
"0.632",
"0.5156",
"0.632"
],
[
"0.1528",
"0.629",
"0.4844",
"0.63",
"0.4844",
"0.6445",
"0.1528",
"0.6436"
],
[
"0.531",
"0.637",
"0.843",
"0.637",
"0.843",
"0.651",
"0.531",
"0.651"
],
[
"0.1511",
"0.6465",
"0.2747",
"0.6455",
"0.2747",
"0.66",
"0.1511",
"0.661"
],
[
"0.11584",
"0.6714",
"0.3167",
"0.6733",
"0.3167",
"0.687",
"0.11584",
"0.685"
],
[
"0.5264",
"0.6665",
"0.881",
"0.6665",
"0.881",
"0.681",
"0.5264",
"0.681"
],
[
"0.5483",
"0.6816",
"0.881",
"0.6816",
"0.881",
"0.6953",
"0.5483",
"0.6953"
],
[
"0.11755",
"0.6934",
"0.486",
"0.6934",
"0.486",
"0.707",
"0.11755",
"0.707"
],
[
"0.5483",
"0.6978",
"0.727",
"0.6978",
"0.727",
"0.7114",
"0.5483",
"0.7114"
],
[
"0.11755",
"0.7104",
"0.4844",
"0.7104",
"0.4844",
"0.7236",
"0.11755",
"0.7236"
],
[
"0.1184",
"0.7256",
"0.486",
"0.7256",
"0.486",
"0.7393",
"0.1184",
"0.7393"
],
[
"0.523",
"0.727",
"0.881",
"0.728",
"0.881",
"0.742",
"0.523",
"0.741"
],
[
"0.1184",
"0.742",
"0.4836",
"0.742",
"0.4836",
"0.753",
"0.1184",
"0.753"
],
[
"0.5483",
"0.7446",
"0.881",
"0.7446",
"0.881",
"0.7583",
"0.5483",
"0.7583"
],
[
"0.1184",
"0.7583",
"0.4844",
"0.7583",
"0.4844",
"0.7715",
"0.1184",
"0.7715"
],
[
"0.5483",
"0.7603",
"0.881",
"0.7603",
"0.881",
"0.775",
"0.5483",
"0.775"
],
[
"0.12006",
"0.7754",
"0.4836",
"0.7754",
"0.4836",
"0.786",
"0.12006",
"0.786"
],
[
"0.55",
"0.7754",
"0.882",
"0.7754",
"0.882",
"0.789",
"0.55",
"0.789"
],
[
"0.1184",
"0.7905",
"0.2544",
"0.7905",
"0.2544",
"0.804",
"0.1184",
"0.804"
],
[
"0.55",
"0.792",
"0.881",
"0.792",
"0.881",
"0.8057",
"0.55",
"0.8057"
],
[
"0.1368",
"0.8057",
"0.4844",
"0.8057",
"0.4844",
"0.8193",
"0.1368",
"0.8193"
],
[
"0.5483",
"0.807",
"0.881",
"0.8086",
"0.881",
"0.8223",
"0.5483",
"0.821"
],
[
"0.11755",
"0.821",
"0.482",
"0.8193",
"0.482",
"0.8345",
"0.11755",
"0.835"
],
[
"0.5483",
"0.826",
"0.882",
"0.826",
"0.882",
"0.839",
"0.5483",
"0.839"
],
[
"0.1184",
"0.8384",
"0.4844",
"0.8384",
"0.4844",
"0.8516",
"0.1184",
"0.8516"
],
[
"0.55",
"0.8413",
"0.882",
"0.8413",
"0.882",
"0.855",
"0.55",
"0.855"
],
[
"0.11755",
"0.8525",
"0.4844",
"0.854",
"0.4844",
"0.868",
"0.11755",
"0.8677"
],
[
"0.5474",
"0.857",
"0.882",
"0.8555",
"0.882",
"0.8706",
"0.5474",
"0.8716"
],
[
"0.12006",
"0.8726",
"0.3938",
"0.8726",
"0.3938",
"0.883",
"0.12006",
"0.883"
],
[
"0.5474",
"0.8716",
"0.881",
"0.8735",
"0.881",
"0.888",
"0.5474",
"0.8857"
],
[
"0.1394",
"0.8926",
"0.398",
"0.8926",
"0.398",
"0.9067",
"0.1394",
"0.9067"
],
[
"0.5483",
"0.8896",
"0.881",
"0.8896",
"0.881",
"0.903",
"0.5483",
"0.903"
],
[
"0.1394",
"0.9062",
"0.4517",
"0.9062",
"0.4517",
"0.9194",
"0.1394",
"0.9194"
],
[
"0.5483",
"0.9062",
"0.881",
"0.9062",
"0.881",
"0.9194",
"0.5483",
"0.9194"
]
] | [
"contextual的度量使用与BERTScore相同的嵌",
"随后，我们在整个火车集的standardized度",
"入，但使用WMD的网络流量优化匹配方案",
"量特征上训练回归集成，旨在预测平均MQM",
"最近，与任务无关的方法已被针对相关目",
"专家判断。我们使用成对的Spearman等级相",
"标对预先训练的模型进行微调的方法所超越：",
"关性（Spearman的p）和对20个%测试拆分",
"BLEUrt（Sellametal.,2020)直接在直接评估提",
"的MQM判断来评估总体以及所有其他选定的",
"交给WMT的BERT(Devlinetal.,2018)模型上",
"指标。",
"进行微调，以使用线性头部对分类[CLS]令牌",
"除了我们对MQM数据集的主要评估之",
"的上下文嵌入进行预测。Comet（Reietal.,2020）",
"外，我们还对WMT2015和2016的直接评估",
"学习从具有三元组目标或标准MSE目标的源，",
"(DA）以及来自编辑后数据集（Fomichevaetal.,",
"参考和翻译文本的元组中预测直接评估",
"2020a）的一组灾难性错误进行了实验。多语言",
"最近的一些工作包含潜在的目标和/或数据",
"质量估计数据集（MLQE-PE）（Fomichevaetal.",
"集。例如，Prism(ThompsonandPost,2020ba)",
"2020b）用于WMT2021质量估计研讨会的评",
"从39E语言的多语言释义中学习了与语言无",
"估。有关我们在DA和MLQE-PE上的实验的",
"关的表示形式，因此是为数不多的性能良好的",
"详细说明，请参阅E3.6节",
"无参考指标之一。其训练目标的正交性可能会",
"我们对分段级别的判断进行了所有评估。为",
"降低其与使用上下文嵌人的其他方法的相关性",
"了最大程度地减少每个特定指标的校准对评估",
"的影响，我们报告了Spearman等级相关系数",
"3Methodology",
"反映了相互的定性顺序，而不是判断的特定值",
"我们的方法旨在通过其他支持问题回答以下主",
"3.2Novel metrics",
"要问题：",
"除了基于文献综述的一组选定指标外，我们还",
"1.",
"Can an ensemble of surface, syntactic, and",
"实现了一组新颖的指标，使我们的整体能够反",
"semantic-level metrics significantly improve",
"映出评估文本的更广泛的属性差异",
"the performance of single metrics?",
"3.2.1 Soft Cosine Measure",
"2. Can such an approach be applied cross",
"lingually, i.e., on languages that it has not",
"软余弦度量（SCM）（Novotny，2018）是软向量",
"been trained on?",
"空间模型中文本的余弦相似度，其中术语的轴",
"处于与它们的余弦相似度ES相对应的角度在",
"3. Can surface-level metrics in reference-free",
"令牌嵌入空间中：",
"configuration achieve results comparable to",
"the reference-based ones?",
"aT.S.y",
"SCM(,) =",
"(1)",
"VaT.S.E.VT.S.U",
"4. Are contextual token representations impor-",
"tant for evaluating semantic equivalence, or",
"其中是一个加权的Boa参考向量（BoW）。",
"can these be replaced with pre-inferred token",
"我们将SCM与两个令牌表示一起使用：",
"representations?",
"3.1 Experimental setup",
"1. We use the static token representations of Fast",
"Text (Grave et al., 2018) . We refer to the re-",
"我们对多维质量度量（MQM）数据集（Freitag",
"sulting metric as SCM .",
"etal.,2021)执行主要评估，其中我们使用平均",
"判断分数作为黄金标准。如果给定的一对来源",
"2. We use the contextual token representations",
"和假设有多个判断，我们将判断的分数平均",
"of BERT (Devlin et al., 2018) using the",
"并将此平均值视为我们的黄金标准。我们根据",
"methodology of BERTScore (Zhang et al.",
"独特的源文本将样本分为训练（80个%）和测",
"2019). We collect representations of all to-",
"试（20个%）子集",
"kens segmented by the WordPiece (Wu et al.,",
"在我们的实验框架中，我们以开源Python库",
"2016) tokenizer, and we treat each unique",
"和Docker映像的形式发布，以便于复制12",
"(token, context) pair as a single term in our",
"我们根据它们的准则实现了一组选定的指标，",
"vocabulary. Subsequently, we decontextual-",
"以及在E3.2节中引入的一系列新颖指标，旨",
"ize these representations as follows: For each",
"在提供文本等效性的其他正交见解",
"WordPiece token, we average the representa-",
"'https://github.com/MIR-MU/regemt",
"tions of all (token, context) pairs in the train-",
"?https://hub.docker.com/r/miratmu/regemt",
"ing corpus. We refer to the resulting metric"
] | {
"arxiv_id": "2109.07242",
"language": "zh",
"page_idx": 2
} |
||
2109.07242-zh-3 | 2109.07242-zh-3.jpg | as SCM-decontextualized. Due to the mul- 造任意文本$G_{t}$的过渡图。由于 PoS 标记tilingual character of the learned BERT to- 的模型与语言有关，因此我们使用兼容ken representations, this metric is applicable both in reference-based and source-based approaches.
除了两个令牌表示形式外，我们还将 Salton and Buckley (1988) 的两种不同的 SMART 加权方案用于 BoW 向量 $\vec{x},\vec{y}$ 和术语相似度矩阵$\operatorname{E}$ $S$的构造：
1. We use raw term frequencies as weights in
the BoW vectors, the nnx SMART weighting scheme, and we construct the term similarity matrix in the vocabulary order. We refer to the resulting metrics as SCM and SCMdecontextualized .
2. We use term frequencies discounted by in-
verse document frequencies as weights in the BoW vectors, the nfx SMART weighting scheme, and we construct the term similarity matrix in the decreasing order of inverse document frequencies (Novotný,2018, Section 3) . We refer to the resulting metrics as SCM-tfidf and SCM-decontextualized-tfidf .
## 3.2.2 $\textbf{Word Mover' s Distance}$
单词移动器的距离(WMD)(Kusner et al.,2015) 在两个文本的向量空间表示之间找到最小成本流$\mathbb{E}F:$
$$\mathrm{WMD}(\vec{x},\vec{y})=\mathrm{minimum~cumulative~cost~}F^{\mathsf{T}}\cdot S\\[0.6ex]\text{subject to}\sum_{j}F_{ij}=x_{i},\sum_{i}F_{ij}=y_{j},\\\text{(2)}$$
其中$\vec{x}$是参考的$\ell_1$归一化加权 BoW 向量(或
无参考的源 XMATHX 加权假设 3)
与上一节中描述的 SC$M$ 类似，我们尝试了两种令牌表示：整个令牌的 FastText 嵌入(WMD )和 WordPiece 令牌的反文本化嵌入(WMDdecontextualized )。此外 ,我们还使用 WordPiece 令牌(WMD-contextual)的上下文嵌入来显示去文本化对度量性能的影响：如果影响可以忽略不计，则将来的工作可以避免对 BERT 表示进行昂贵的实时推断，并显着减少词汇量
与 SCM 类似，我们还使用了两种不同的加权方案：原始术语频率(WMD-*)和按逆文档频率打折的术语频率(WMD-*-tfidf)
# 3.2.3 Compositionality
的标记方案-尽管并非总是系统地对齐-用于训练英语 (Weischedel et al.,2013),德语(Brants et al., 2002), 中文 (Weischedel et al., 2013) 和挪威语 (Unhammer and Trosterud, 2009)
随后，我们对矩阵$G_t$的值进行行归一化，并为$x=G_{t_1}$和$y=G_t_2$定义 Compositionality E $C$的距离度量：
(3)
$$C(x,y)=\ell_1\text{-norm}(x_{ii}-y_{ii}),$$
,其中 PoS 标记$i\in x$和$i\in y$
在我们的提交中，只有当语言属于我们有标记器的一组语言时，我们才应用此指标：英语， 德语，中文或挪威语。在基于参考的评估中， 此约束适用于目标语言；在无参考评估的情况下，它适用于源语言和目标语言
## 3.3 Ensemble
我们将上述指标整合为回归模型中的预测指标，从而最大程度地减少了平均细分级别MQM 得分与预测目标之间的残差
我们使用各种简单的回归器进行实验，并观察到具有 100 维隐藏层和 ReLu 激活的完全连接的两层感知器以及具有平方残差的线性回归的简单方法的卓越性能。我们将 RegEMT 的结果报告为在训练数据集的 20 个% 保留验证子集上选择的这些分类器中性能最好的一个
除了所有可用指标的集合之外，我们还使用郑重的两个表面级别功能来评估基线回归集合$Reg-base:$根据 WordPiece$\boxed{\mathrm{F}}($Wu et al.,2016)令牌的字符级别源和目标长度
# 3.4 Cross-lingual experiments
由于获得专家判断的成本非常高，因此期望将来对于包含资源不足语言的绝大多数语言对， 都可以使用经过训练的系统的训练数据是不现实的。为了评估所有指标在未发现的语言对上的性能，我们对两个可用语言对的平均 MQM 判断执行跨语言评估：zh-en 和 en-de
在适用的情况下，我们将度量参数拟合到未报告的语言对的火车拆分上。随后，我们评估并将报告对的测试拆分结果报告给 MQM 判断
## 3.5 Ablation study
为了了解单个指标在其作为我们的集合的预测因子的作用中的影响，我们使用它们的成对相关性进行系统的特征消除
我们称为组合性的自定义度量基于令牌 在我们的消融研究中，我们迭代地选择具的词性 (PoS) 类别的定向成对过渡来构 有最高 Spearman's$\overline{\mathrm{E}}\rho$的度量到任何其他度量。 | [
[
"0.1561",
"0.0885",
"0.4954",
"0.0885",
"0.4954",
"0.0986",
"0.1561",
"0.0986"
],
[
"0.511",
"0.0873",
"0.881",
"0.0873",
"0.881",
"0.10095",
"0.511",
"0.10095"
],
[
"0.1545",
"0.1027",
"0.4937",
"0.1027",
"0.4937",
"0.1164",
"0.1545",
"0.1164"
],
[
"0.511",
"0.1027",
"0.879",
"0.1027",
"0.879",
"0.1164",
"0.511",
"0.1164"
],
[
"0.1528",
"0.1194",
"0.4895",
"0.1194",
"0.4895",
"0.133",
"0.1528",
"0.133"
],
[
"0.513",
"0.12054",
"0.8774",
"0.12054",
"0.8774",
"0.1312",
"0.513",
"0.1312"
],
[
"0.1528",
"0.1342",
"0.4878",
"0.1354",
"0.4878",
"0.1497",
"0.1528",
"0.1484"
],
[
"0.513",
"0.1354",
"0.882",
"0.1354",
"0.882",
"0.1484",
"0.513",
"0.1484"
],
[
"0.1561",
"0.1526",
"0.2241",
"0.1526",
"0.2241",
"0.1644",
"0.1561",
"0.1644"
],
[
"0.513",
"0.1509",
"0.8774",
"0.1509",
"0.8774",
"0.1644",
"0.513",
"0.1644"
],
[
"0.5146",
"0.1687",
"0.879",
"0.1687",
"0.879",
"0.1787",
"0.5146",
"0.1787"
],
[
"0.1368",
"0.1758",
"0.4844",
"0.1758",
"0.4844",
"0.1895",
"0.1368",
"0.1895"
],
[
"0.1184",
"0.1912",
"0.486",
"0.1912",
"0.486",
"0.2048",
"0.1184",
"0.2048"
],
[
"0.531",
"0.1996",
"0.879",
"0.1996",
"0.879",
"0.2131",
"0.531",
"0.2131"
],
[
"0.1184",
"0.2073",
"0.486",
"0.2073",
"0.486",
"0.2214",
"0.1184",
"0.2214"
],
[
"0.511",
"0.2156",
"0.882",
"0.2144",
"0.882",
"0.2286",
"0.511",
"0.2299"
],
[
"0.11584",
"0.2214",
"0.2007",
"0.2239",
"0.2007",
"0.2382",
"0.11584",
"0.2363"
],
[
"0.511",
"0.2299",
"0.634",
"0.231",
"0.634",
"0.2452",
"0.511",
"0.2446"
],
[
"0.1293",
"0.2465",
"0.486",
"0.2477",
"0.486",
"0.2625",
"0.1293",
"0.2612"
],
[
"0.1528",
"0.2632",
"0.4895",
"0.2632",
"0.4895",
"0.2778",
"0.1528",
"0.2778"
],
[
"0.5845",
"0.2612",
"0.882",
"0.2612",
"0.882",
"0.2756",
"0.5845",
"0.2756"
],
[
"0.1545",
"0.2798",
"0.4895",
"0.2798",
"0.4895",
"0.2935",
"0.1545",
"0.2935"
],
[
"0.517",
"0.2944",
"0.534",
"0.2944",
"0.534",
"0.3018",
"0.517",
"0.3018"
],
[
"0.5264",
"0.2903",
"0.756",
"0.2915",
"0.756",
"0.306",
"0.5264",
"0.3047"
],
[
"0.531",
"0.308",
"0.881",
"0.308",
"0.881",
"0.3218",
"0.531",
"0.3218"
],
[
"0.1511",
"0.2957",
"0.4902",
"0.2944",
"0.4902",
"0.3088",
"0.1511",
"0.31"
],
[
"0.1545",
"0.3123",
"0.4878",
"0.3123",
"0.4878",
"0.3267",
"0.1545",
"0.3267"
],
[
"0.1528",
"0.3289",
"0.288",
"0.3289",
"0.288",
"0.3425",
"0.1528",
"0.3425"
],
[
"0.5156",
"0.326",
"0.879",
"0.326",
"0.879",
"0.3362",
"0.5156",
"0.3362"
],
[
"0.511",
"0.339",
"0.8774",
"0.339",
"0.8774",
"0.3528",
"0.511",
"0.3528"
],
[
"0.1318",
"0.354",
"0.4902",
"0.354",
"0.4902",
"0.3677",
"0.1318",
"0.3677"
],
[
"0.511",
"0.3552",
"0.879",
"0.3552",
"0.879",
"0.3682",
"0.511",
"0.3682"
],
[
"0.1545",
"0.3706",
"0.4895",
"0.3706",
"0.4895",
"0.3843",
"0.1545",
"0.3843"
],
[
"0.511",
"0.3723",
"0.771",
"0.3723",
"0.771",
"0.386",
"0.511",
"0.386"
],
[
"0.1511",
"0.3855",
"0.4844",
"0.386",
"0.4844",
"0.401",
"0.1511",
"0.3997"
],
[
"0.511",
"0.3997",
"0.5615",
"0.3997",
"0.5615",
"0.411",
"0.511",
"0.411"
],
[
"0.553",
"0.398",
"0.636",
"0.3997",
"0.636",
"0.4146",
"0.553",
"0.412"
],
[
"0.1545",
"0.4038",
"0.4836",
"0.4038",
"0.4836",
"0.4146",
"0.1545",
"0.4146"
],
[
"0.1528",
"0.4187",
"0.4878",
"0.4187",
"0.4878",
"0.4316",
"0.1528",
"0.4316"
],
[
"0.511",
"0.4204",
"0.879",
"0.4204",
"0.879",
"0.434",
"0.511",
"0.434"
],
[
"0.1545",
"0.434",
"0.4844",
"0.434",
"0.4844",
"0.4478",
"0.1545",
"0.4478"
],
[
"0.511",
"0.4358",
"0.879",
"0.4358",
"0.879",
"0.4495",
"0.511",
"0.4495"
],
[
"0.1545",
"0.4507",
"0.4844",
"0.4507",
"0.4844",
"0.4644",
"0.1545",
"0.4644"
],
[
"0.511",
"0.4531",
"0.785",
"0.4531",
"0.785",
"0.466",
"0.511",
"0.466"
],
[
"0.1545",
"0.4673",
"0.3997",
"0.4673",
"0.3997",
"0.481",
"0.1545",
"0.481"
],
[
"0.532",
"0.4697",
"0.882",
"0.4697",
"0.882",
"0.4827",
"0.532",
"0.4827"
],
[
"0.11584",
"0.4893",
"0.3628",
"0.4912",
"0.3628",
"0.506",
"0.11584",
"0.5044"
],
[
"0.513",
"0.485",
"0.881",
"0.485",
"0.881",
"0.4988",
"0.513",
"0.4988"
],
[
"0.513",
"0.502",
"0.881",
"0.502",
"0.881",
"0.5156",
"0.513",
"0.5156"
],
[
"0.1184",
"0.5103",
"0.4844",
"0.5103",
"0.4844",
"0.524",
"0.1184",
"0.524"
],
[
"0.513",
"0.517",
"0.881",
"0.517",
"0.881",
"0.531",
"0.513",
"0.531"
],
[
"0.1184",
"0.527",
"0.4844",
"0.527",
"0.4844",
"0.5405",
"0.1184",
"0.5405"
],
[
"0.5146",
"0.5337",
"0.876",
"0.5337",
"0.876",
"0.5444",
"0.5146",
"0.5444"
],
[
"0.1184",
"0.5435",
"0.1838",
"0.5435",
"0.1838",
"0.5547",
"0.1184",
"0.5547"
],
[
"0.5146",
"0.55",
"0.8423",
"0.55",
"0.8423",
"0.5635",
"0.5146",
"0.5635"
],
[
"0.1351",
"0.572",
"0.486",
"0.572",
"0.486",
"0.586",
"0.1351",
"0.586"
],
[
"0.531",
"0.5664",
"0.879",
"0.5664",
"0.879",
"0.58",
"0.531",
"0.58"
],
[
"0.513",
"0.582",
"0.881",
"0.582",
"0.881",
"0.5957",
"0.513",
"0.5957"
],
[
"0.1293",
"0.593",
"0.4702",
"0.593",
"0.4702",
"0.6133",
"0.1293",
"0.6133"
],
[
"0.513",
"0.597",
"0.881",
"0.597",
"0.881",
"0.6123",
"0.513",
"0.6123"
],
[
"0.272",
"0.6143",
"0.287",
"0.6143",
"0.287",
"0.626",
"0.272",
"0.626"
],
[
"0.3787",
"0.616",
"0.3862",
"0.6143",
"0.3896",
"0.618",
"0.3804",
"0.619"
],
[
"0.513",
"0.6143",
"0.735",
"0.6143",
"0.735",
"0.6274",
"0.513",
"0.6274"
],
[
"0.4626",
"0.627",
"0.4878",
"0.627",
"0.4878",
"0.64",
"0.4626",
"0.64"
],
[
"0.5146",
"0.6426",
"0.5615",
"0.6426",
"0.5615",
"0.654",
"0.5146",
"0.654"
],
[
"0.5557",
"0.641",
"0.764",
"0.641",
"0.764",
"0.6562",
"0.5557",
"0.6562"
],
[
"0.1184",
"0.653",
"0.486",
"0.653",
"0.486",
"0.6665",
"0.1184",
"0.6665"
],
[
"0.513",
"0.661",
"0.879",
"0.661",
"0.879",
"0.675",
"0.513",
"0.675"
],
[
"0.1184",
"0.6685",
"0.392",
"0.6685",
"0.392",
"0.6816",
"0.1184",
"0.6816"
],
[
"0.511",
"0.6777",
"0.879",
"0.6777",
"0.879",
"0.6914",
"0.511",
"0.6914"
],
[
"0.1368",
"0.685",
"0.486",
"0.685",
"0.486",
"0.6987",
"0.1368",
"0.6987"
],
[
"0.513",
"0.695",
"0.882",
"0.695",
"0.882",
"0.708",
"0.513",
"0.708"
],
[
"0.1184",
"0.7007",
"0.486",
"0.7007",
"0.486",
"0.7144",
"0.1184",
"0.7144"
],
[
"0.11755",
"0.7163",
"0.4895",
"0.7153",
"0.4895",
"0.73",
"0.11755",
"0.731"
],
[
"0.511",
"0.7114",
"0.881",
"0.7114",
"0.881",
"0.7246",
"0.511",
"0.7246"
],
[
"0.511",
"0.727",
"0.879",
"0.727",
"0.879",
"0.7407",
"0.511",
"0.7407"
],
[
"0.1184",
"0.733",
"0.4844",
"0.733",
"0.4844",
"0.7466",
"0.1184",
"0.7466"
],
[
"0.12006",
"0.749",
"0.486",
"0.749",
"0.486",
"0.7627",
"0.12006",
"0.7627"
],
[
"0.511",
"0.7437",
"0.8076",
"0.7437",
"0.8076",
"0.7573",
"0.511",
"0.7573"
],
[
"0.1184",
"0.7656",
"0.4844",
"0.7656",
"0.4844",
"0.7793",
"0.1184",
"0.7793"
],
[
"0.531",
"0.7603",
"0.881",
"0.7603",
"0.881",
"0.774",
"0.531",
"0.774"
],
[
"0.511",
"0.775",
"0.879",
"0.775",
"0.879",
"0.788",
"0.511",
"0.788"
],
[
"0.1184",
"0.782",
"0.486",
"0.782",
"0.486",
"0.796",
"0.1184",
"0.796"
],
[
"0.1184",
"0.7974",
"0.4475",
"0.7974",
"0.4475",
"0.811",
"0.1184",
"0.811"
],
[
"0.5146",
"0.7935",
"0.879",
"0.7935",
"0.879",
"0.804",
"0.5146",
"0.804"
],
[
"0.1351",
"0.8125",
"0.4844",
"0.8135",
"0.4844",
"0.8276",
"0.1351",
"0.8267"
],
[
"0.509",
"0.818",
"0.6733",
"0.8184",
"0.6733",
"0.833",
"0.509",
"0.832"
],
[
"0.1184",
"0.83",
"0.486",
"0.83",
"0.486",
"0.8433",
"0.1184",
"0.8433"
],
[
"0.11755",
"0.844",
"0.413",
"0.8457",
"0.413",
"0.86",
"0.11755",
"0.8594"
],
[
"0.513",
"0.8403",
"0.879",
"0.8403",
"0.879",
"0.854",
"0.513",
"0.854"
],
[
"0.513",
"0.8555",
"0.879",
"0.8555",
"0.879",
"0.869",
"0.513",
"0.869"
],
[
"0.11584",
"0.869",
"0.3105",
"0.869",
"0.3105",
"0.8843",
"0.11584",
"0.8843"
],
[
"0.513",
"0.8726",
"0.7173",
"0.8726",
"0.7173",
"0.8857",
"0.513",
"0.8857"
],
[
"0.11755",
"0.8896",
"0.4844",
"0.8896",
"0.4844",
"0.903",
"0.11755",
"0.903"
],
[
"0.531",
"0.8896",
"0.882",
"0.8896",
"0.882",
"0.903",
"0.531",
"0.903"
],
[
"0.11755",
"0.905",
"0.4878",
"0.904",
"0.4878",
"0.9185",
"0.11755",
"0.9194"
],
[
"0.511",
"0.905",
"0.882",
"0.905",
"0.882",
"0.9194",
"0.511",
"0.9194"
]
] | [
"as SCM-decontextualized.Dueto themul-",
"造任意文本Gt的过渡图。由于PoS标记",
"tilingual character of the learned BERT to-",
"的模型与语言有关，因此我们使用兼容",
"ken representations, this metric is applicable",
"的标记方案-尽管并非总是系统地对齐-用",
"both in reference-based and source-based ap-",
"于训练英语（Weischedeletal.，2013），德语",
"proaches.",
"(Brants et al., 2002)，中文 (Weischedel et al.",
"2013）和挪威语（UnhammerandTrosterud，2009）",
"除了两个令牌表示形式外，我们还将Salton",
"andBuckley（1988）的两种不同的SMART加权",
"随后，我们对矩阵Gt的值进行行归一化，并",
"方案用于BoW向量&,y和术语相似度矩阵E",
"为=Gt和y=Gt2定义CompositionalityE",
"S的构造：",
"C的距离度量：",
"1. We use raw term frequencies as weights in",
"the BoW vectors, the nnx SMART weight-",
"C(c,y) = li-norm(ci-yi),(3)",
"ing scheme, and we construct the term simi-",
"",
"其中PoS标记iEc和iEy",
"在我们的提交中，只有当语言属于我们有标",
"larity matrix in the vocabulary order. We re-",
"fer to the resulting metrics as SCM and SCM-",
"decontextualized.",
"记器的一组语言时，我们才应用此指标：英语，",
"德语，中文或挪威语。在基于参考的评估中",
"2. We use term frequencies discounted by in-",
"此约束适用于目标语言；在无参考评估的情况",
"verse document frequencies as weights in the",
"下，它适用于源语言和目标语言",
"BoW vectors,the nfx SMART weighting",
"3.3",
"Ensemble",
"scheme,and we constructtheterm similarity",
"matrix in the decreasing order of inverse docu-",
"我们将上述指标整合为回归模型中的预测",
"ment frequencies (Novotny, 2018, Section 3) ",
"指标，从而最大程度地减少了平均细分级别",
"We refer to the resulting metrics as SCM-tfidf",
"MQM得分与预测目标之间的残差",
"and SCM-decontextualized-tfidf.",
"我们使用各种简单的回归器进行实验，并观",
"3.2.2Word Mover's Distance",
"察到具有100维隐藏层和ReLu激活的完全连",
"接的两层感知器以及具有平方残差的线性回归",
"单词移动器的距离（WMD）（Kusneretal.,2015）",
"的简单方法的卓越性能。我们将RegEMT的结",
"在两个文本的向量空间表示之间找到最小成本",
"果报告为在训练数据集的20个%保留验证子",
"流EF:",
"集上选择的这些分类器中性能最好的一个",
"WMD(ε, y) = minimum cumulative cost F .S",
"除了所有可用指标的集合之外，我们还使用",
"郑重的两个表面级别功能来评估基线回归集合",
"[0.6ea] subject to Fij = i, Fij = Yj,",
"Reg-base：根据WordPieceE(Wu et al.,2016)令",
"3j",
"",
"牌的字符级别源和目标长度",
"(2)",
"3.4",
"Cross-lingual experiments",
"其中是参考的li归一化加权BoW向量（或",
"由于获得专家判断的成本非常高，因此期望将",
"无参考的源XMATHX加权假设3）",
"来对于包含资源不足语言的绝大多数语言对",
"与上一节中描述的SCM类似，我们尝试了两",
"都可以使用经过训练的系统的训练数据是不现",
"种令牌表示：整个令牌的FastText嵌人（WMD",
"）和WordPiece令牌的反文本化嵌人（WMD-",
"实的。为了评估所有指标在未发现的语言对上",
"的性能，我们对两个可用语言对的平均MQM",
"decontextualized）。此外，我们还使用WordPiece",
"令牌（WMD-contextual）的上下文嵌入来显示",
"判断执行跨语言评估：zh-en和en-de",
"去文本化对度量性能的影响：如果影响可以忽",
"在适用的情况下，我们将度量参数拟合到未",
"报告的语言对的火车拆分上。随后，我们评估",
"略不计，则将来的工作可以避免对BERT表示",
"进行昂贵的实时推断，并显着减少词汇量",
"并将报告对的测试折分结果报告给MQM判断",
"与SCM类似，我们还使用了两种不同的加",
"3.5Ablationstudy",
"权方案：原始术语频率（WMD-*）和按逆文档",
"频率打折的术语频率（WMD.*-tfidf）",
"为了了解单个指标在其作为我们的集合的预测",
"因子的作用中的影响，我们使用它们的成对相",
"3.2.3Compositionality",
"关性进行系统的特征消除",
"我们称为组合性的自定义度量基于令牌",
"在我们的消融研究中，我们迭代地选择具",
"的词性（PoS）类别的定向成对过渡来构",
"有最高Spearman'sEp的度量到任何其他度量。"
] | {
"arxiv_id": "2109.07242",
"language": "zh",
"page_idx": 3
} |
||
2109.07242-zh-4 | 2109.07242-zh-4.jpg | 我们通过在剩余特征上拟合新的回归模型来从 中的情况有所不同，其中对于基于参考的情况
我们的集合中消除选定的度量。我们将继续进 WMD 的反文本化版本的相关性优于E7%。
行，直到消除所有指标，并在流程的每个步骤
中评估总体
# 3.6 Additional evaluations
为了进一步了解结果与其他相关评估源的一致
性，以及在关键错误识别的新应用中对指标的 都高于 0.79。基于 SCM 和 WMD 的静态令牌
评估，我们对 2015 年和 2016 年 WMT 提交的DA 数据集以及 MLQE-PE 数据集的关键错误开发集进行了类似的实验，用于无参考指标
在 DA 判断的情况下，我们使用 2015 年的评估作为培训拆分，使用 2016 年的评估作为测试拆分
在 MLQE-PE 的情况下，我们通过以 80:20的比例拆分唯一的源文本来类似于 MQM 地拆分数据。在这种情况下，我们将三个注释者分配给每个翻译的平均错误严重程度视为黄金判断
## 4 Results
Correlations to MQM judgements. 表 1 列出了基于源的 MQM 的相关性，i。e。、无参考指标(上)和基于参考的指标(中)。RegEMT 报告的结果与给定评估方案中所有其他可用指标的估计值拟合选定的回归模型。如第 3.3 节所述，我们根据其在训练集的保留部分上的性能来选择评估的回归模型：基于源的 zh-en 对的两层感知器和所有其他情况下的简单线性回归，回归模型之间在验证集上的性能差异可忽略不计(低于 2%)。在以$X$ 为后缀的报告中， 我们将回归模型拟合到其他语言对上，而不是用于评估的语言对上。
结果表明，简单的回归集成可以从大多数评估配置中的预测变量中受益，包括跨语言设置和其他评估数据集。在 MQM 判断的情况下， 我们观察到相关性的最高边际
Baseline ensemble. 表 1 显示 , Reg-base 仅使用参考和假设词块的计数，就证明了其相对于$BLEU$和 METEOR 的标准表面水平指标的一致优势，即使在跨语言与。单语比较。关于MQM 判断，Reg-base 的相关性相当一致；因此，在 en-de 语言对的无参考情况下，Reg-base 的相关性非常接近 RegEMT 的相关性
Importance of contextualization. 表$\overline{\mathrm{E}}$1 中的
结果在令牌级度量中上下文化的重要性方面 集成遵循其集成度量的低相关性。此外，完全不一致 (WMD-cont* 与。WMD-dec* 和 SCM- 依赖于指标的一致性，集合将不可避免地暴露$cont^*$ vs。SCM-dec*)。在zh-en 语言对的 all 例 一些指标明显超出其通常范围的域中的错误
中，我们观察到 WMD 的语境化和解语境化版 另一方面，我们认为，对于主要是无监督的本之间的相关性显着降低(15-16%)。en-de 对 表面级别指标，情况可能很少如此。我们怀疑
Metrics correlations. 表 2 演示了评估指标的相互关联。我们看到基于上下文化表示的度量之间存在很强的成对相关性，例如 Comet , Prism , BERTScore 和 BLEUrt 之间；所有这些表示的度量之间的情况类似，有和没有 TF-IDF
相比之下，我们观察到$BLEU$和$METEOR$与
$Reg-base$的低相关性形成了一组表面级度量
Ablation study. 图 1显示了当我们从集合预测器集中逐步消除度量时，回归集合的 Spearman $\rho$中的性能发展。按照 3.5 节中描述的方法，表 2 中显示了 zh-en 对消融指标的确切顺序，我们观察到 MQM 的其他语言对也是相似的
在基于引用的度量的集合(左)中，我们观察到在删除大多数度量的过程中具有很高的一致性。$zh$-en 情况下的较长一致性归因于集成 BLEUrt (在步骤$\overline{\mathrm{E}}$14 中删除)和 $METEOR$ (在步骤$\overline{\mathrm{E}}15$ 中删除)的一致性能。当独立评估时，这些度量仅分别达到0.48和0.28的相关性。在 en-de 的情况下，最显著的下降可以归因于性能最好的 Comet (步骤 9) 和 Prism (步骤11)的去除。
基于源的度量的集合(右)显示在删除 Prism (步骤$\overline{\mathrm{E}}3)$ 和 WMD-contextual(步骤$\overline{\mathrm{E}}4)$ 后zh-en 对的显着下降。在 en-de 语言对中，在整个消融过程中相关性相对较低。最不相关，因此最后被消除的是$WMD-decontextualized-tfidf$ (步骤$\overline{\mathrm{E}}6)$和 Prism (步骤$\overline{\mathrm{E}}7)$
### 5 Discussion
Regressive ensemble. 遵循我们在 3 节中设定的目标，我们从经验上确认，在大多数配置中，集成可以提高对专家判断进行建模的质量，同时对其他配置执行接近最佳的度量。此外，我们证明了这种合奏可以转移到新的语言对中，并且即使在跨语言环境中，它的使用也受到质量收益的推动。
同时，必须承认与单一和无监督指标相比， 集成系统暴露的限制。合奏可能会继承其每个指标的系统偏差。在表$\overline{\mathrm{E}}$1 中基于源的 MQM 的 en-de 对的结果中可以观察到此问题，其中 | [
[
"0.1184",
"0.0873",
"0.482",
"0.0873",
"0.482",
"0.0974",
"0.1184",
"0.0974"
],
[
"0.4802",
"0.0903",
"0.4844",
"0.0903",
"0.4844",
"0.0956",
"0.4802",
"0.0956"
],
[
"0.513",
"0.0873",
"0.881",
"0.0873",
"0.881",
"0.10095",
"0.513",
"0.10095"
],
[
"0.1184",
"0.1027",
"0.4878",
"0.1027",
"0.4878",
"0.1164",
"0.1184",
"0.1164"
],
[
"0.511",
"0.1027",
"0.8555",
"0.1027",
"0.8555",
"0.1164",
"0.511",
"0.1164"
],
[
"0.1184",
"0.1194",
"0.4844",
"0.1194",
"0.4844",
"0.133",
"0.1184",
"0.133"
],
[
"0.1184",
"0.1354",
"0.2091",
"0.1354",
"0.2091",
"0.1467",
"0.1184",
"0.1467"
],
[
"0.519",
"0.1318",
"0.881",
"0.1318",
"0.881",
"0.1455",
"0.519",
"0.1455"
],
[
"0.511",
"0.1478",
"0.879",
"0.1478",
"0.879",
"0.1609",
"0.511",
"0.1609"
],
[
"0.11584",
"0.1633",
"0.34",
"0.1633",
"0.34",
"0.177",
"0.11584",
"0.177"
],
[
"0.511",
"0.1644",
"0.8833",
"0.165",
"0.8833",
"0.1787",
"0.511",
"0.1776"
],
[
"0.12006",
"0.1853",
"0.486",
"0.1853",
"0.486",
"0.1953",
"0.12006",
"0.1953"
],
[
"0.511",
"0.1812",
"0.881",
"0.1812",
"0.881",
"0.1948",
"0.511",
"0.1948"
],
[
"0.511",
"0.1965",
"0.879",
"0.1965",
"0.879",
"0.2102",
"0.511",
"0.2102"
],
[
"0.12006",
"0.2031",
"0.486",
"0.2031",
"0.486",
"0.2131",
"0.12006",
"0.2131"
],
[
"0.12006",
"0.2173",
"0.4878",
"0.2173",
"0.4878",
"0.231",
"0.12006",
"0.231"
],
[
"0.511",
"0.2131",
"0.882",
"0.2131",
"0.882",
"0.2268",
"0.511",
"0.2268"
],
[
"0.1184",
"0.2328",
"0.486",
"0.2328",
"0.486",
"0.2465",
"0.1184",
"0.2465"
],
[
"0.529",
"0.2299",
"0.882",
"0.2299",
"0.882",
"0.2434",
"0.529",
"0.2434"
],
[
"0.12006",
"0.2505",
"0.465",
"0.2505",
"0.465",
"0.2612",
"0.12006",
"0.2612"
],
[
"0.51",
"0.2465",
"0.858",
"0.2465",
"0.858",
"0.26",
"0.51",
"0.26"
],
[
"0.1368",
"0.2654",
"0.4844",
"0.2654",
"0.4844",
"0.279",
"0.1368",
"0.279"
],
[
"0.517",
"0.2756",
"0.881",
"0.2756",
"0.881",
"0.2893",
"0.517",
"0.2893"
],
[
"0.12177",
"0.2832",
"0.4844",
"0.2832",
"0.4844",
"0.2935",
"0.12177",
"0.2935"
],
[
"0.511",
"0.2915",
"0.882",
"0.2922",
"0.882",
"0.306",
"0.511",
"0.3047"
],
[
"0.12006",
"0.2986",
"0.1914",
"0.2986",
"0.1914",
"0.31",
"0.12006",
"0.31"
],
[
"0.511",
"0.3071",
"0.8774",
"0.3071",
"0.8774",
"0.3206",
"0.511",
"0.3206"
],
[
"0.1351",
"0.3142",
"0.486",
"0.3142",
"0.486",
"0.328",
"0.1351",
"0.328"
],
[
"0.511",
"0.3225",
"0.879",
"0.3225",
"0.879",
"0.3362",
"0.511",
"0.3362"
],
[
"0.11755",
"0.33",
"0.4878",
"0.3289",
"0.4878",
"0.3433",
"0.11755",
"0.3445"
],
[
"0.513",
"0.3403",
"0.882",
"0.3403",
"0.882",
"0.354",
"0.513",
"0.354"
],
[
"0.1184",
"0.3467",
"0.4844",
"0.3467",
"0.4844",
"0.3599",
"0.1184",
"0.3599"
],
[
"0.1184",
"0.3623",
"0.4878",
"0.3623",
"0.4878",
"0.376",
"0.1184",
"0.376"
],
[
"0.513",
"0.3582",
"0.532",
"0.3582",
"0.532",
"0.3677",
"0.513",
"0.3677"
],
[
"0.11584",
"0.3777",
"0.1377",
"0.3777",
"0.1377",
"0.3926",
"0.11584",
"0.3926"
],
[
"0.529",
"0.3723",
"0.879",
"0.3723",
"0.879",
"0.386",
"0.529",
"0.386"
],
[
"0.513",
"0.3901",
"0.881",
"0.3901",
"0.881",
"0.4038",
"0.513",
"0.4038"
],
[
"0.11755",
"0.408",
"0.1511",
"0.408",
"0.1511",
"0.4204",
"0.11755",
"0.4204"
],
[
"0.1436",
"0.408",
"0.2167",
"0.408",
"0.2167",
"0.4229",
"0.1436",
"0.4229"
],
[
"0.5283",
"0.4062",
"0.881",
"0.4062",
"0.881",
"0.4192",
"0.5283",
"0.4192"
],
[
"0.513",
"0.4216",
"0.881",
"0.4216",
"0.881",
"0.4353",
"0.513",
"0.4353"
],
[
"0.12427",
"0.434",
"0.4878",
"0.434",
"0.4878",
"0.4478",
"0.12427",
"0.4478"
],
[
"0.509",
"0.437",
"0.879",
"0.437",
"0.879",
"0.4507",
"0.509",
"0.4507"
],
[
"0.11755",
"0.4495",
"0.486",
"0.4495",
"0.486",
"0.4631",
"0.11755",
"0.4631"
],
[
"0.511",
"0.4536",
"0.879",
"0.4536",
"0.879",
"0.4673",
"0.511",
"0.4673"
],
[
"0.1184",
"0.466",
"0.4844",
"0.466",
"0.4844",
"0.4797",
"0.1184",
"0.4797"
],
[
"0.5146",
"0.4697",
"0.881",
"0.4697",
"0.881",
"0.4827",
"0.5146",
"0.4827"
],
[
"0.1184",
"0.4822",
"0.4878",
"0.4822",
"0.4878",
"0.4954",
"0.1184",
"0.4954"
],
[
"0.513",
"0.4863",
"0.881",
"0.4863",
"0.881",
"0.4995",
"0.513",
"0.4995"
],
[
"0.1184",
"0.4988",
"0.4878",
"0.4988",
"0.4878",
"0.5127",
"0.1184",
"0.5127"
],
[
"0.51",
"0.5005",
"0.66",
"0.502",
"0.66",
"0.5166",
"0.51",
"0.5156"
],
[
"0.11755",
"0.514",
"0.486",
"0.514",
"0.486",
"0.528",
"0.11755",
"0.528"
],
[
"0.529",
"0.5186",
"0.881",
"0.5195",
"0.881",
"0.5337",
"0.529",
"0.533"
],
[
"0.1184",
"0.531",
"0.4878",
"0.531",
"0.4878",
"0.5444",
"0.1184",
"0.5444"
],
[
"0.509",
"0.5337",
"0.881",
"0.535",
"0.881",
"0.55",
"0.509",
"0.549"
],
[
"0.1184",
"0.5464",
"0.4844",
"0.5464",
"0.4844",
"0.56",
"0.1184",
"0.56"
],
[
"0.511",
"0.552",
"0.881",
"0.552",
"0.881",
"0.5654",
"0.511",
"0.5654"
],
[
"0.11755",
"0.561",
"0.4836",
"0.561",
"0.4836",
"0.5747",
"0.11755",
"0.5747"
],
[
"0.511",
"0.5684",
"0.881",
"0.5684",
"0.881",
"0.582",
"0.511",
"0.582"
],
[
"0.1184",
"0.58",
"0.4836",
"0.58",
"0.4836",
"0.5903",
"0.1184",
"0.5903"
],
[
"0.513",
"0.5845",
"0.879",
"0.5845",
"0.879",
"0.597",
"0.513",
"0.597"
],
[
"0.1184",
"0.594",
"0.4878",
"0.594",
"0.4878",
"0.608",
"0.1184",
"0.608"
],
[
"0.511",
"0.5996",
"0.7656",
"0.5996",
"0.7656",
"0.6133",
"0.511",
"0.6133"
],
[
"0.1184",
"0.611",
"0.2913",
"0.611",
"0.2913",
"0.6245",
"0.1184",
"0.6245"
],
[
"0.1368",
"0.6274",
"0.4878",
"0.6274",
"0.4878",
"0.641",
"0.1368",
"0.641"
],
[
"0.509",
"0.6313",
"0.6406",
"0.633",
"0.6406",
"0.648",
"0.509",
"0.6455"
],
[
"0.1184",
"0.6436",
"0.486",
"0.6436",
"0.486",
"0.6567",
"0.1184",
"0.6567"
],
[
"0.11755",
"0.6577",
"0.4902",
"0.659",
"0.4902",
"0.6733",
"0.11755",
"0.673"
],
[
"0.517",
"0.661",
"0.881",
"0.661",
"0.881",
"0.675",
"0.517",
"0.675"
],
[
"0.12006",
"0.6777",
"0.3552",
"0.6777",
"0.3552",
"0.6885",
"0.12006",
"0.6885"
],
[
"0.513",
"0.6777",
"0.8774",
"0.6777",
"0.8774",
"0.6885",
"0.513",
"0.6885"
],
[
"0.5146",
"0.6934",
"0.881",
"0.6934",
"0.881",
"0.707",
"0.5146",
"0.707"
],
[
"0.12177",
"0.7017",
"0.486",
"0.7017",
"0.486",
"0.7153",
"0.12177",
"0.7153"
],
[
"0.513",
"0.7104",
"0.882",
"0.7104",
"0.882",
"0.7236",
"0.513",
"0.7236"
],
[
"0.1184",
"0.7188",
"0.4878",
"0.7188",
"0.4878",
"0.7324",
"0.1184",
"0.7324"
],
[
"0.513",
"0.7256",
"0.881",
"0.7256",
"0.881",
"0.7393",
"0.513",
"0.7393"
],
[
"0.1184",
"0.734",
"0.486",
"0.734",
"0.486",
"0.7476",
"0.1184",
"0.7476"
],
[
"0.513",
"0.742",
"0.881",
"0.742",
"0.881",
"0.756",
"0.513",
"0.756"
],
[
"0.12427",
"0.752",
"0.4844",
"0.752",
"0.4844",
"0.7627",
"0.12427",
"0.7627"
],
[
"0.511",
"0.7573",
"0.686",
"0.7573",
"0.686",
"0.771",
"0.511",
"0.771"
],
[
"0.1184",
"0.7666",
"0.486",
"0.7666",
"0.486",
"0.78",
"0.1184",
"0.78"
],
[
"0.1184",
"0.781",
"0.4844",
"0.781",
"0.4844",
"0.7944",
"0.1184",
"0.7944"
],
[
"0.531",
"0.7754",
"0.882",
"0.7754",
"0.882",
"0.789",
"0.531",
"0.789"
],
[
"0.11755",
"0.7964",
"0.4165",
"0.7974",
"0.4165",
"0.8125",
"0.11755",
"0.811"
],
[
"0.5146",
"0.7915",
"0.881",
"0.7915",
"0.881",
"0.8047",
"0.5146",
"0.8047"
],
[
"0.511",
"0.807",
"0.881",
"0.807",
"0.881",
"0.821",
"0.511",
"0.821"
],
[
"0.12177",
"0.8247",
"0.4878",
"0.8247",
"0.4878",
"0.8384",
"0.12177",
"0.8384"
],
[
"0.51",
"0.8237",
"0.8774",
"0.8237",
"0.8774",
"0.8374",
"0.51",
"0.8374"
],
[
"0.11755",
"0.8403",
"0.4878",
"0.8413",
"0.4878",
"0.855",
"0.11755",
"0.854"
],
[
"0.51",
"0.8384",
"0.881",
"0.839",
"0.881",
"0.854",
"0.51",
"0.8525"
],
[
"0.11755",
"0.857",
"0.4895",
"0.857",
"0.4895",
"0.8706",
"0.11755",
"0.8706"
],
[
"0.513",
"0.8555",
"0.881",
"0.8555",
"0.881",
"0.869",
"0.513",
"0.869"
],
[
"0.1184",
"0.8726",
"0.486",
"0.8726",
"0.486",
"0.887",
"0.1184",
"0.887"
],
[
"0.509",
"0.8706",
"0.8623",
"0.8716",
"0.8623",
"0.8853",
"0.509",
"0.8843"
],
[
"0.12006",
"0.8896",
"0.486",
"0.8896",
"0.486",
"0.903",
"0.12006",
"0.903"
],
[
"0.531",
"0.8896",
"0.881",
"0.8896",
"0.881",
"0.903",
"0.531",
"0.903"
],
[
"0.12006",
"0.9067",
"0.4878",
"0.9067",
"0.4878",
"0.9175",
"0.12006",
"0.9175"
],
[
"0.492",
"0.91",
"0.511",
"0.91",
"0.511",
"0.913",
"0.492",
"0.913"
],
[
"0.51",
"0.9062",
"0.881",
"0.9062",
"0.881",
"0.9194",
"0.51",
"0.9194"
]
] | [
"我们通过在剩余特征上拟合新的回归模型来从",
"",
"中的情况有所不同，其中对于基于参考的情况，",
"我们的集合中消除选定的度量。我们将继续进",
"WMD的反文本化版本的相关性优于E7%。",
"行，直到消除所有指标，并在流程的每个步骤",
"中评估总体",
"Metricscorrelations.表2演示了评估指标",
"的相互关联。我们看到基于上下文化表示的度",
"3.6Additionalevaluations",
"量之间存在很强的成对相关性，例如Comet，",
"为了进一步了解结果与其他相关评估源的一致",
"Prism，BERTScore和BLEUrt之间；所有这些",
"都高于0.79。基于SCM和WMD的静态令牌",
"性，以及在关键错误识别的新应用中对指标的",
"评估，我们对2015年和2016年WMT提交的",
"表示的度量之间的情况类似，有和没有TF-IDF",
"DA数据集以及MLQE-PE数据集的关键错误",
"相比之下，我们观察到BLEU和METEOR与",
"开发集进行了类似的实验，用于无参考指标",
"Reg-base的低相关性形成了一组表面级度量",
"在DA判断的情况下，我们使用2015年的",
"Ablationstudy.图1显示了当我们从集合预",
"评估作为培训拆分，使用2016年的评估作为",
"测器集中逐步消除度量时，回归集合的Spear",
"测试拆分",
"manp中的性能发展。按照3.5节中描述的方",
"在MLQE-PE的情况下，我们通过以80:20",
"法，表2中显示了zh-en对消融指标的确切顺",
"的比例拆分唯一的源文本来类似于MQM地拆",
"序，我们观察到MQM的其他语言对也是相似",
"分数据。在这种情况下，我们将三个注释者分",
"配给每个翻译的平均错误严重程度视为黄金判",
"的",
"断",
"在基于引用的度量的集合（左）中，我们观",
"察到在删除大多数度量的过程中具有很高的",
"4",
"Results",
"致性。zh-en情况下的较长一致性归因于集",
"成BLEUrt（在步骤E14中删除）和METEOR",
"Correlations to MQM judgements.表 1列",
"（在步骤E15中删除）的一致性能。当独立评",
"出了基于源的MQM的相关性，i。e。、无参",
"估时，这些度量仅分别达到0.48和0.28的相",
"考指标（上）和基于参考的指标（中)。RegEMT",
"关性。在en-de的情况下，最显著的下降可以",
"报告的结果与给定评估方案中所有其他可用指",
"归因于性能最好的Comet（步骤9）和Prism",
"标的估计值拟合选定的回归模型。如第3.3节",
"（步骤11）的去除。",
"所述，我们根据其在训练集的保留部分上的性",
"基于源的度量的集合（右）显示在删除Prism",
"能来选择评估的回归模型：基于源的zh-en对",
"（步骤E3）和WMD-contextual（步骤E4）后",
"的两层感知器和所有其他情况下的简单线性回",
"zh-en对的显着下降。在en-de语言对中，在整",
"归，回归模型之间在验证集上的性能差异可忽",
"个消融过程中相关性相对较低。最不相关，因",
"略不计（低于2%）。在以X为后缀的报告中",
"此最后被消除的是WMD-decontextualized-tfidf",
"我们将回归模型拟合到其他语言对上，而不是",
"(步骤E6）和Prism（步骤E7)",
"用于评估的语言对上。",
"结果表明，简单的回归集成可以从大多数评",
"5Discussion",
"估配置中的预测变量中受益，包括跨语言设置",
"和其他评估数据集。在MQM判断的情况下，",
"Regressiveensemble.遵循我们在3节中设",
"我们观察到相关性的最高边际",
"定的目标，我们从经验上确认，在大多数配置",
"中，集成可以提高对专家判断进行建模的质",
"Baselineensemble.表1显示，Reg-base仅使",
"量，同时对其他配置执行接近最佳的度量。此",
"用参考和假设词块的计数，就证明了其相对",
"外，我们证明了这种合奏可以转移到新的语言",
"于BLEU和METEOR的标准表面水平指标的",
"对中，并且即使在跨语言环境中，它的使用也",
"一致优势，即使在跨语言与。单语比较。关于",
"受到质量收益的推动。",
"MQM判断，Reg-base的相关性相当一致；因",
"此，在en-de语言对的无参考情况下，Reg-base",
"同时，必须承认与单一和无监督指标相比，",
"的相关性非常接近RegEMT的相关性",
"集成系统暴露的限制。合奏可能会继承其每个",
"指标的系统偏差。在表E1中基于源的MQM",
"Importance of contextualization.表国1中的",
"的en-de对的结果中可以观察到此问题，其中",
"结果在令牌级度量中上下文化的重要性方面",
"集成遵循其集成度量的低相关性。此外，完全",
"不一致（WMD-cont*与。WMD-dec*和 SCM-",
"依赖于指标的一致性，集合将不可避免地暴露",
"cont*vs。SCM-dec*）。在zh-en语言对的all 例",
"一些指标明显超出其通常范围的域中的错误",
"中，我们观察到WMD的语境化和解语境化版",
"另一方面，我们认为，对于主要是无监督的",
"本之间的相关性显着降低（15-16%）。en-de对",
"",
"表面级别指标，情况可能很少如此。我们怀疑"
] | {
"arxiv_id": "2109.07242",
"language": "zh",
"page_idx": 4
} |
||
2104.09159-zh-0 | 2104.09159-zh-0.jpg | Conditional Variational Capsule Network for Open Set Recognition
$\begin{array}{c}{\mathrm{Yunrui~Guo}^{*2,1},\text{Guglielmo Camporese}^{*1},\text{Wenjing Yang}^{2},\text{Alessandro Sperduti}^{1},\text{and Lamberto}}\\{\text{Ballan}^{1}}\end{array}$
$^{1}$Department of Mathematics“Tullio Levi-Civita”, University of Padova, Italy
$^{2}$National University of Defense Technology, China
$\operatorname{Abstract}$
In open set recognition, a classifier has to detect unknown classes that are not known at training time. In order to recognize new categories, the classifier has to project the input samples of known classes in very compact and separated regions of the features space for $\bar{\text{discriminating samples of unknown classes. Recently}}$ proposed Capsule Networks have shown to outperform ${\mathrm{alternatives~in~many~felds,~particularly~in~image~recog-}}$ nition, however they have not been fully applied yet to open-set recognition. In capsule networks, scalar neurons are replaced by capsule vectors or matrices, whose entries represent different properties of objects. In our proposal, during training, capsules features of the same known class are encouraged to match a predefined gaussian, one for each class. To this end, we use the variational autoencoder framework, with a set of gaussian priors as the approximation for the posterior distribution. In this way, we are able to control the compactness of the features of the same class around the center of the gaussians, thus controlling the ability of the classifier in detecting samples from unknown classes. We conducted several experiments and ablation of our model, obtaining state of the art results on different datasets in the open set recognition and unknown detection tasks.
<FigureHere>
Figure 1: CVAECapOSR Model. Input samples are fed into the Capsule Network that produces distributions over the latent space. Each class has its own prior gaussian distribution in the feature space, that in the figure are represented as spheres. After training, the known samples (represented as small points) are clustered around the class target gaussians. The samples belonging to unknown classes are represented as black triangles, far from the target distributions.
## 1. Introduction
在过去的十年中，深度学习已成为许多计算机视觉问题的主要方法，在许多视觉识别任务[13,27,7,29] 上取得了惊人的成果。但是，这些结果中的大多数是在闭集场景中获得的，其中一个关键假设是所有样本都应属于至少一个标记类别。当观察来自未知类别的样本时，闭集方法被迫从其中一个已知类中选择一个类
${^*\text{Indicates equal contri}}$butions
标签，从而限制了它们在动态和不断变化的场景中的
适用性
为了克服这种限制，引人了开放集识别，使分类系统能够识别所有已知类别，同时检测未知测试样本[25,1] 。在开放集场景中，标签空间中包含/排除的样本称为已知/未知。因此，开放集分类器需要使用从有限的可访问类别集合中学到的不完整知识，以设计能够将已知与未知分开的有效表示。早期的工作已经发现了这个问题，因此提出了采用不同阈值策略来拒绝未知[25,1] 的方法
尽管深度神经网络在封闭场景中显示出强大的学习判别表示能力，但在开放集设置[2]中显示出准确性下降。作为一种朴素的策略，对 Softmax 输出的阈值进
1 | [
[
"0.1462",
"0.132",
"0.824",
"0.1332",
"0.824",
"0.1509",
"0.1462",
"0.1497"
],
[
"0.0907",
"0.1957",
"0.8813",
"0.1957",
"0.8813",
"0.2102",
"0.0907",
"0.2102"
],
[
"0.4534",
"0.2134",
"0.5146",
"0.2134",
"0.5146",
"0.2279",
"0.4534",
"0.2279"
],
[
"0.1569",
"0.2405",
"0.812",
"0.2412",
"0.812",
"0.2588",
"0.1569",
"0.2583"
],
[
"0.2695",
"0.2583",
"0.701",
"0.26",
"0.701",
"0.2778",
"0.2695",
"0.276"
],
[
"0.2361",
"0.315",
"0.3145",
"0.315",
"0.3145",
"0.3333",
"0.2361",
"0.3333"
],
[
"0.0989",
"0.3499",
"0.4697",
"0.3499",
"0.4697",
"0.363",
"0.0989",
"0.363"
],
[
"0.517",
"0.348",
"0.6006",
"0.348",
"0.6006",
"0.3623",
"0.517",
"0.3623"
],
[
"0.0825",
"0.3662",
"0.4673",
"0.3662",
"0.4673",
"0.377",
"0.0825",
"0.377"
],
[
"0.0801",
"0.3801",
"0.4697",
"0.3801",
"0.4697",
"0.3933",
"0.0801",
"0.3933"
],
[
"0.0801",
"0.3958",
"0.468",
"0.3958",
"0.468",
"0.4092",
"0.0801",
"0.4092"
],
[
"0.0809",
"0.4104",
"0.468",
"0.4104",
"0.468",
"0.4236",
"0.0809",
"0.4236"
],
[
"0.0809",
"0.4255",
"0.468",
"0.4255",
"0.468",
"0.4395",
"0.0809",
"0.4395"
],
[
"0.7754",
"0.433",
"0.789",
"0.433",
"0.789",
"0.4414",
"0.7754",
"0.4414"
],
[
"0.0801",
"0.44",
"0.468",
"0.4395",
"0.468",
"0.4539",
"0.0801",
"0.455"
],
[
"0.523",
"0.4458",
"0.591",
"0.4458",
"0.591",
"0.4558",
"0.523",
"0.4558"
],
[
"0.0801",
"0.4558",
"0.4673",
"0.4558",
"0.4673",
"0.4697",
"0.0801",
"0.4697"
],
[
"0.5513",
"0.461",
"0.564",
"0.461",
"0.564",
"0.4749",
"0.5513",
"0.4749"
],
[
"0.0809",
"0.4702",
"0.468",
"0.4702",
"0.468",
"0.4841",
"0.0809",
"0.4841"
],
[
"0.0801",
"0.4849",
"0.468",
"0.4849",
"0.468",
"0.4988",
"0.0801",
"0.4988"
],
[
"0.5186",
"0.481",
"0.5986",
"0.481",
"0.5986",
"0.4912",
"0.5186",
"0.4912"
],
[
"0.0809",
"0.5005",
"0.4673",
"0.5005",
"0.4673",
"0.5146",
"0.0809",
"0.5146"
],
[
"0.0809",
"0.5166",
"0.4673",
"0.5166",
"0.4673",
"0.53",
"0.0809",
"0.53"
],
[
"0.0801",
"0.531",
"0.4714",
"0.531",
"0.4714",
"0.5444",
"0.0801",
"0.5444"
],
[
"0.5034",
"0.535",
"0.8906",
"0.535",
"0.8906",
"0.549",
"0.5034",
"0.549"
],
[
"0.0809",
"0.5474",
"0.4673",
"0.5474",
"0.4673",
"0.558",
"0.0809",
"0.558"
],
[
"0.5024",
"0.5503",
"0.8896",
"0.5503",
"0.8896",
"0.5645",
"0.5024",
"0.5645"
],
[
"0.0784",
"0.56",
"0.4697",
"0.5596",
"0.4697",
"0.5737",
"0.0784",
"0.5747"
],
[
"0.5024",
"0.565",
"0.8906",
"0.5664",
"0.8906",
"0.58",
"0.5024",
"0.579"
],
[
"0.0809",
"0.5776",
"0.4673",
"0.5776",
"0.4673",
"0.5884",
"0.0809",
"0.5884"
],
[
"0.505",
"0.5835",
"0.8896",
"0.5835",
"0.8896",
"0.5933",
"0.505",
"0.5933"
],
[
"0.0801",
"0.5913",
"0.468",
"0.5913",
"0.468",
"0.605",
"0.0801",
"0.605"
],
[
"0.5024",
"0.5967",
"0.892",
"0.5967",
"0.892",
"0.61",
"0.5024",
"0.61"
],
[
"0.0801",
"0.607",
"0.4697",
"0.607",
"0.4697",
"0.6206",
"0.0801",
"0.6206"
],
[
"0.5024",
"0.6113",
"0.888",
"0.6113",
"0.888",
"0.625",
"0.5024",
"0.625"
],
[
"0.0809",
"0.622",
"0.4697",
"0.622",
"0.4697",
"0.6353",
"0.0809",
"0.6353"
],
[
"0.5034",
"0.627",
"0.892",
"0.627",
"0.892",
"0.64",
"0.5034",
"0.64"
],
[
"0.0809",
"0.636",
"0.468",
"0.636",
"0.468",
"0.6494",
"0.0809",
"0.6494"
],
[
"0.5034",
"0.6416",
"0.892",
"0.6416",
"0.892",
"0.655",
"0.5034",
"0.655"
],
[
"0.0801",
"0.6514",
"0.468",
"0.6514",
"0.468",
"0.6655",
"0.0801",
"0.6655"
],
[
"0.5024",
"0.657",
"0.8105",
"0.657",
"0.8105",
"0.6704",
"0.5024",
"0.6704"
],
[
"0.0809",
"0.6665",
"0.468",
"0.6665",
"0.468",
"0.6797",
"0.0809",
"0.6797"
],
[
"0.0809",
"0.683",
"0.4673",
"0.683",
"0.4673",
"0.6934",
"0.0809",
"0.6934"
],
[
"0.0801",
"0.6963",
"0.4697",
"0.6963",
"0.4697",
"0.711",
"0.0801",
"0.711"
],
[
"0.505",
"0.702",
"0.892",
"0.702",
"0.892",
"0.7163",
"0.505",
"0.7163"
],
[
"0.0809",
"0.712",
"0.2646",
"0.712",
"0.2646",
"0.726",
"0.0809",
"0.726"
],
[
"0.5034",
"0.7173",
"0.5537",
"0.7173",
"0.5537",
"0.732",
"0.5034",
"0.732"
],
[
"0.527",
"0.7344",
"0.8906",
"0.7344",
"0.8906",
"0.7476",
"0.527",
"0.7476"
],
[
"0.5034",
"0.7485",
"0.886",
"0.7485",
"0.886",
"0.762",
"0.5034",
"0.762"
],
[
"0.0809",
"0.759",
"0.2181",
"0.759",
"0.2181",
"0.7734",
"0.0809",
"0.7734"
],
[
"0.5107",
"0.7646",
"0.8906",
"0.7646",
"0.8906",
"0.778",
"0.5107",
"0.778"
],
[
"0.10046",
"0.7827",
"0.4673",
"0.7827",
"0.4673",
"0.7974",
"0.10046",
"0.7974"
],
[
"0.5034",
"0.778",
"0.8896",
"0.779",
"0.8896",
"0.7935",
"0.5034",
"0.7925"
],
[
"0.505",
"0.7944",
"0.8896",
"0.7944",
"0.8896",
"0.808",
"0.505",
"0.808"
],
[
"0.0825",
"0.8003",
"0.4656",
"0.8003",
"0.4656",
"0.8145",
"0.0825",
"0.8145"
],
[
"0.5034",
"0.808",
"0.8896",
"0.8096",
"0.8896",
"0.8237",
"0.5034",
"0.8228"
],
[
"0.08496",
"0.815",
"0.4673",
"0.815",
"0.4673",
"0.829",
"0.08496",
"0.829"
],
[
"0.505",
"0.8247",
"0.8906",
"0.8247",
"0.8906",
"0.8384",
"0.505",
"0.8384"
],
[
"0.0809",
"0.8306",
"0.468",
"0.8306",
"0.468",
"0.8447",
"0.0809",
"0.8447"
],
[
"0.0825",
"0.845",
"0.4673",
"0.845",
"0.4673",
"0.8594",
"0.0825",
"0.8594"
],
[
"0.5034",
"0.8413",
"0.5537",
"0.8413",
"0.5537",
"0.853",
"0.5034",
"0.853"
],
[
"0.0809",
"0.8613",
"0.468",
"0.8613",
"0.468",
"0.875",
"0.0809",
"0.875"
],
[
"0.524",
"0.856",
"0.8906",
"0.856",
"0.8906",
"0.869",
"0.524",
"0.869"
],
[
"0.505",
"0.872",
"0.8896",
"0.872",
"0.8896",
"0.8853",
"0.505",
"0.8853"
],
[
"0.0972",
"0.887",
"0.2852",
"0.887",
"0.2852",
"0.901",
"0.0972",
"0.901"
],
[
"0.5024",
"0.8853",
"0.892",
"0.886",
"0.892",
"0.901",
"0.5024",
"0.8994"
],
[
"0.4805",
"0.9253",
"0.492",
"0.9253",
"0.492",
"0.935",
"0.4805",
"0.935"
]
] | [
"Conditional Variational Capsule Network for Open Set Recognition",
"Yunrui Guo*2,1, Guglielmo Camporese*1, Wenjing Yang?, Alessandro Sperdutil, and Lamberto",
"Ballanl",
"1Department of Mathematics “Tullio Levi-Civita\", University of Padova, Italy",
"2National University of Defense Technology, China",
"Abstract",
"In open set recognition, a classifier has to detect",
"卤·#",
"unknown classes that are not known at training time.",
"In order to recognize new categories, the classifier has",
"to project the input samples of known classes in very",
"compact and separated regions of the features space for",
"discriminating samples of unknown classes.Recently",
"00",
"proposed Capsule Networks have shown to outperform",
"CVAECapOSR",
"alternatives in many fields, particularly in image recog-",
"4",
"nition, however they have not been fully applied yet",
"to open-set recognition. In capsule networks, scalar",
"【+",
"neurons are replaced by capsule vectors or matrices,",
"whose entries represent different properties of objects.",
"In our proposal, during training, capsules features of",
"Figure 1: CVAECapOSR Model. Input samples are",
"the same known class are encouraged to matchapre-",
"fed into the Capsule Network that produces distribu-",
"defined gaussian, one for each class. To this end, we",
"tions over the latent space. Each class has its own prior",
"usethevariational autoencoder framework,with a set",
"gaussian distribution inthefeature space,that in the",
"of gaussian priors as the approximation for the poste-",
"figure are represented as spheres. After training, the",
"rior distribution. In this way, we are able to control the",
"known samples (represented as small points) are clus-",
"compactness of the features of the same class around",
"tered around the class target gaussians.The samples",
"the center of the gaussians, thus controlling the abil-",
"belonging to unknown classes are represented as black",
"ity of the classifier in detecting samples from unknown",
"triangles, far from the target distributions.",
"classes. We conducted several experiments and abla-",
"tion of our model.obtaining state of the art results",
"on different datasets in the open set recognition and",
"标签，从而限制了它们在动态和不断变化的场景中的",
"unknown detection tasks.",
"适用性",
"为了克服这种限制，引入了开放集识别，使分类系统",
"能够识别所有已知类别，同时检测未知测试样本25，1",
"1.Introduction",
"在开放集场景中，标签空间中包含/排除的样本称为",
"在过去的十年中，深度学习已成为许多计算机视觉",
"已知/未知。因此，开放集分类器需要使用从有限的可",
"访问类别集合中学到的不完整知识，以设计能够将已",
"问题的主要方法，在许多视觉识别任务【13，27，7，29",
"知与未知分开的有效表示。早期的工作已经发现了这个",
"上取得了惊人的成果。但是，这些结果中的大多数是在",
"问题，因此提出了采用不同阈值策略来拒绝未知25，11",
"闭集场景中获得的，其中一个关键假设是所有样本都",
"应属于至少一个标记类别。当观察来自未知类别的样",
"的方法",
"本时，闭集方法被迫从其中一个已知类中选择一个类",
"尽管深度神经网络在封闭场景中显示出强大的学习",
"判别表示能力，但在开放集设置[2中显示出准确性下",
"*Indicates equal contributions.",
"降。作为一种朴素的策略，对Softmax输出的阈值进",
"1"
] | {
"arxiv_id": "2104.09159",
"language": "zh",
"page_idx": 0
} |
||
2104.09159-zh-1 | 2104.09159-zh-1.jpg | 行建模已被证明是深度神经网络识别未知数的次优解决方案。因此，引人了极值理论，以更好地适应这些完全基于监督学习的判别模型，以适应开放集设置。基本思想是校准 Softmax 分数，以便估计未知[32,2]的概率。除了深度判别模型外，专注于通过无监督学习来学习有效的潜在特征表示的深度生成模型已被广泛用于开放集识别任务中，并相继获得了成功[16,19,21,28] 。特别是，变分自动编码器(VAE) 是一种典型的概率生成模型，非常适合检测未知数，因为它能够学习潜在空间中的低维表示，不仅支持输入重建，而且可以近似指定的先验分布。另一方面，基于 VAE 的模型可能不足以有效地识别已知类别，因为所有特征表示仅遵循一种分布。为此，我们采用条件 VAE (CVAE),它使用多个先验分布来对已知类进行建模，并间接对未知对应物进行建模。此外，我们建议用概率胶囊表示输入样本，因为它们已经证明了其表示能力$\mathbb{E}[22,24]$
提出了胶囊网络 (CapsNet)[24] 作为卷积神经网络(CNN)的替代方案。与 CNN 的标量神经元不同，胶囊将一组神经元集合在一起以接受和输出向量。激活胶囊的向量表示特定对象的各种属性，例如位置，大小，方向，纹理等。从本质上讲，CapsNet 可以看作是通过分布式表示对对象进行编码的编码器，这比通过激活高维空间中的单个神经元对对象进行编码要有效得多。。此外，CapsNet 已成功用于在类似于开放集识别 [18] 的任务设置中检测假图像和视频。这促使我们结合 CVAE 为开放集识别问题设计了一种新颖的胶囊网络架构，称为 CVAECapOSR,如图$\boxed{\mathrm{E}}1$ 所示
本文的贡献有三个方面：i) $\operatorname{E\text{我们提出了一种基}}$ 于 CapsNet 的新颖的开放集识别框架，并展示了其在学习已知类的有效表示方面的优势。ii) 我们集成了CapsNet 和条件 VAE。与鼓励潜在表示近似单个先验分布的一般 VAE 相比，我们的模型利用了多个先验 (i。e。每个类一个), 并且它迫使潜在表示遵循由输人样本的类选择的高斯先验。ii)|E我们对用于开放集识别的所有标准数据集进行了广泛的实验，获得了非常有竞争力的结果，在某些情况下，其性能大大超过了以前的最新方法。
# 2. Related Work
开集识别问题是由[25]引人的，最初被形式化为基于支持向量机 (SVM) 的约束最小化问题，而随后的工作集中在其他更传统的方法上，例如极值理论$\mathbb{E}[9,26]$ , 稀疏表示[E[34] 和最近邻$\mathbb{E}[10]$
继深度学习在许多计算机视觉任务中取得的成功之后，在$\mathbb{E}[2]$ 中首次引入了深度网络用于开放集识别，其中通过使用 Weibull 分布模型校准每个类的 Softmax 概率来提出 Openmax 函数。随后，[5]通过引人生成对抗网络将 Openmax 扩展到 G-Openmax,在该网络中，生成器生成新颖类别的合成样本，而鉴别器学习未知类的显式表示。在$\mathbb{E}[16]$ 中采用了类似的策略，该策略提出了一种基于生成对抗网络的数据增强技术，称为反事实图像生成。最近，Yoshihashi et al.[32] 分析并证明了在开放集场景中联合训练深度网络进行分类
和重建的有用性。具体来说，作者建议使用无监督训练产生的表示将已知与未知分开，同时使用通过监督学习过程计算的表示保持模型的判别能力
C2AE$\boxed{\mathbb{E}}[19]$通过基于极值理论对模型的重建误差进行建模，引人了一种基于类条件 VAE 的开放集识别和未知检测的体系结构。Sun et al. $\mathbb{E}[28]$最近认为，基于 VAE 的开放集识别体系结构的一个缺点是对已知类的实例的判别能力不足。因此，作者采用条件高斯分布VAE 模型来学习已知类的条件分布并拒绝未知数。$\mathbb{E}$ [33]中提出了一种不同的方法，其中将归一化流用于已知样本的密度估计。具体来说，作者提出了一种使用CNN 编码器和可逆神经网络的架构，该架构可以共同学习输入的密度。但是，本文未讨论的一个潜在问题是CNN 编码器没有双射特性，这对于采用变量变化公式进行密度评估至关重要。此外，[3] 在原型学习中引入了互易点的概念来管理开放空间。尽管这项工作在拒绝来自已知样本的不同数据集的未知数方面表现出出色的性能，但当未知样本的来源与已知样本相同时，未知检测能力会下降
# 3. Preliminaries
3.1. The Open Set Recognition Problem
在开放集识别问题中，模型必须对可能属于训练期间未看到的类的测试样本进行分类。给定一个分类数据集$D=\{(x_1,y_1),\ldots,(x_n,y_n)\}$ ,使得$x_i\in\mathcal{X}$是输人样本，$y_i\in\{1,\ldots,K\}$是相应的类别标签，开放集问题在于$K+U$类中测试样本的分类，其中$U$是未知类的数量。在文献中，用于训练的数据集称为封闭数据集，同时在评估期间使用的包含来自看不见的类的样本的数据集称为开放数据集。为了在评估期间量化数据 集 的 开 放 性 , 在 $\left [ 25, 16\right ]$之 后 , 我 们 将 开 放 性 度 量 视 为 $O= 1- \sqrt {\frac KM}$,其 中 $K$ 和 $M= K+ U$ 分 别 是 在 训 练 和 测 计 邯 问 观 密 到 的 米 酱练和测试期间观察到的类数
# 3.2. Conditional VAE Formulation
条件变分 Eutoencoder (CVAE) 直接源自 VAE 模型 [11] , 其目标基于给定标签$y$的数据$x$的条件密度$p(x|y)$的估计。它是最强大的概率生成模型之一，因为它的理论优雅，强大的框架兼容性和高效的流形表示。CVAE 通常包括一个编码器，该编码器将输人$x$和类$y$映射到潜在变量$z$上的预定分布，以及一个解码器， 该解码器给定潜在变量$z$和类$y$尝试重建输人$x$ 。在训练过程中，通过最小化数据条件密度的负变分下界来训练模型，定义如下：
$$\begin{aligned}\mathcal{L}\big[\boldsymbol{x},y;\theta,\phi\big]&=D_{\mathrm{KL}}\Big[q_{\phi}(\boldsymbol{z}|\boldsymbol{x})\|p(\boldsymbol{z}|y)\Big]\\&-\operatorname{E}_{q_{\phi}(\boldsymbol{z}|x)}\Big[\log p_{\theta}(\boldsymbol{x}|\boldsymbol{z},y)\Big]\end{aligned}$$
其中$q_\phi(\boldsymbol{z}|\boldsymbol{x})$表示编码器的后验，$p_\theta(\boldsymbol{z}|y)$表示以$y$为条件的潜在变量$z$的先验分布。损失函数中的第一项是一个正则化器，它强制近似后验分布$q_\phi(x|z)$接近先 | [
[
"0.0809",
"0.09344",
"0.468",
"0.09344",
"0.468",
"0.1067",
"0.0809",
"0.1067"
],
[
"0.505",
"0.0947",
"0.8896",
"0.0947",
"0.8896",
"0.1048",
"0.505",
"0.1048"
],
[
"0.08417",
"0.10986",
"0.4673",
"0.10986",
"0.4673",
"0.1206",
"0.08417",
"0.1206"
],
[
"0.5024",
"0.1067",
"0.8906",
"0.108",
"0.8906",
"0.1218",
"0.5024",
"0.1206"
],
[
"0.08417",
"0.125",
"0.4673",
"0.125",
"0.4673",
"0.1351",
"0.08417",
"0.1351"
],
[
"0.505",
"0.125",
"0.7793",
"0.125",
"0.7793",
"0.1351",
"0.505",
"0.1351"
],
[
"0.0825",
"0.1383",
"0.468",
"0.1383",
"0.468",
"0.1515",
"0.0825",
"0.1515"
],
[
"0.5215",
"0.1364",
"0.8906",
"0.1383",
"0.8906",
"0.1528",
"0.5215",
"0.1509"
],
[
"0.0825",
"0.154",
"0.468",
"0.154",
"0.468",
"0.1674",
"0.0825",
"0.1674"
],
[
"0.5034",
"0.1528",
"0.8906",
"0.1528",
"0.8906",
"0.1666",
"0.5034",
"0.1666"
],
[
"0.08417",
"0.1692",
"0.4656",
"0.1692",
"0.4656",
"0.1799",
"0.08417",
"0.1799"
],
[
"0.5034",
"0.1686",
"0.8906",
"0.1686",
"0.8906",
"0.1818",
"0.5034",
"0.1818"
],
[
"0.0825",
"0.1849",
"0.4656",
"0.1849",
"0.4656",
"0.1957",
"0.0825",
"0.1957"
],
[
"0.5034",
"0.1818",
"0.8896",
"0.1831",
"0.8896",
"0.1963",
"0.5034",
"0.1957"
],
[
"0.0948",
"0.1995",
"0.4673",
"0.1995",
"0.4673",
"0.2102",
"0.0948",
"0.2102"
],
[
"0.5034",
"0.1989",
"0.8896",
"0.1989",
"0.8896",
"0.2122",
"0.5034",
"0.2122"
],
[
"0.0809",
"0.2122",
"0.468",
"0.214",
"0.468",
"0.2279",
"0.0809",
"0.226"
],
[
"0.505",
"0.2134",
"0.8906",
"0.2134",
"0.8906",
"0.2267",
"0.505",
"0.2267"
],
[
"0.0809",
"0.2291",
"0.4673",
"0.2291",
"0.4673",
"0.2424",
"0.0809",
"0.2424"
],
[
"0.505",
"0.2291",
"0.888",
"0.2291",
"0.888",
"0.2424",
"0.505",
"0.2424"
],
[
"0.0809",
"0.2437",
"0.468",
"0.2437",
"0.468",
"0.2568",
"0.0809",
"0.2568"
],
[
"0.505",
"0.2437",
"0.888",
"0.2437",
"0.888",
"0.2568",
"0.505",
"0.2568"
],
[
"0.0809",
"0.2588",
"0.468",
"0.2588",
"0.468",
"0.2727",
"0.0809",
"0.2727"
],
[
"0.505",
"0.2588",
"0.8896",
"0.2588",
"0.8896",
"0.2727",
"0.505",
"0.2727"
],
[
"0.08905",
"0.2747",
"0.4673",
"0.2747",
"0.4673",
"0.2886",
"0.08905",
"0.2886"
],
[
"0.505",
"0.2747",
"0.8896",
"0.2747",
"0.8896",
"0.2886",
"0.505",
"0.2886"
],
[
"0.0809",
"0.2886",
"0.4673",
"0.289",
"0.4673",
"0.303",
"0.0809",
"0.3018"
],
[
"0.5034",
"0.2874",
"0.8906",
"0.289",
"0.8906",
"0.3037",
"0.5034",
"0.3018"
],
[
"0.0825",
"0.305",
"0.468",
"0.305",
"0.468",
"0.318",
"0.0825",
"0.318"
],
[
"0.5063",
"0.3062",
"0.888",
"0.3062",
"0.888",
"0.3164",
"0.5063",
"0.3164"
],
[
"0.0809",
"0.318",
"0.4248",
"0.318",
"0.4248",
"0.332",
"0.0809",
"0.332"
],
[
"0.5073",
"0.3193",
"0.8906",
"0.3193",
"0.8906",
"0.3333",
"0.5073",
"0.3333"
],
[
"0.10046",
"0.3352",
"0.4673",
"0.3352",
"0.4673",
"0.3484",
"0.10046",
"0.3484"
],
[
"0.5034",
"0.334",
"0.8896",
"0.334",
"0.8896",
"0.348",
"0.5034",
"0.348"
],
[
"0.0809",
"0.3499",
"0.4673",
"0.3499",
"0.4673",
"0.363",
"0.0809",
"0.363"
],
[
"0.505",
"0.3499",
"0.8896",
"0.3499",
"0.8896",
"0.363",
"0.505",
"0.363"
],
[
"0.08417",
"0.3662",
"0.4656",
"0.3662",
"0.4656",
"0.377",
"0.08417",
"0.377"
],
[
"0.5034",
"0.3643",
"0.634",
"0.3643",
"0.634",
"0.3774",
"0.5034",
"0.3774"
],
[
"0.0809",
"0.379",
"0.4673",
"0.379",
"0.4673",
"0.3926",
"0.0809",
"0.3926"
],
[
"0.0825",
"0.3945",
"0.468",
"0.3945",
"0.468",
"0.4077",
"0.0825",
"0.4077"
],
[
"0.501",
"0.3926",
"0.6436",
"0.3933",
"0.6436",
"0.4077",
"0.501",
"0.4072"
],
[
"0.0809",
"0.4104",
"0.468",
"0.4104",
"0.468",
"0.4236",
"0.0809",
"0.4236"
],
[
"0.5034",
"0.4172",
"0.812",
"0.4172",
"0.812",
"0.4312",
"0.5034",
"0.4312"
],
[
"0.0825",
"0.4268",
"0.4656",
"0.4268",
"0.4656",
"0.4375",
"0.0825",
"0.4375"
],
[
"0.0825",
"0.4426",
"0.4656",
"0.4426",
"0.4656",
"0.4526",
"0.0825",
"0.4526"
],
[
"0.524",
"0.4414",
"0.8896",
"0.4414",
"0.8896",
"0.455",
"0.524",
"0.455"
],
[
"0.0825",
"0.4558",
"0.468",
"0.4558",
"0.468",
"0.4697",
"0.0825",
"0.4697"
],
[
"0.505",
"0.4558",
"0.8906",
"0.4558",
"0.8906",
"0.4697",
"0.505",
"0.4697"
],
[
"0.0809",
"0.4702",
"0.468",
"0.4702",
"0.468",
"0.4841",
"0.0809",
"0.4841"
],
[
"0.5024",
"0.4702",
"0.8896",
"0.4702",
"0.8896",
"0.4849",
"0.5024",
"0.4849"
],
[
"0.0809",
"0.4849",
"0.4207",
"0.4849",
"0.4207",
"0.4988",
"0.0809",
"0.4988"
],
[
"0.505",
"0.486",
"0.8896",
"0.486",
"0.8896",
"0.5005",
"0.505",
"0.5005"
],
[
"0.1013",
"0.5005",
"0.468",
"0.5005",
"0.468",
"0.5146",
"0.1013",
"0.5146"
],
[
"0.505",
"0.502",
"0.8906",
"0.502",
"0.8906",
"0.515",
"0.505",
"0.515"
],
[
"0.0825",
"0.5166",
"0.4673",
"0.5166",
"0.4673",
"0.53",
"0.0825",
"0.53"
],
[
"0.5024",
"0.515",
"0.8896",
"0.5166",
"0.8896",
"0.53",
"0.5024",
"0.529"
],
[
"0.0801",
"0.53",
"0.4673",
"0.531",
"0.4673",
"0.5444",
"0.0801",
"0.5435"
],
[
"0.505",
"0.5327",
"0.8896",
"0.5327",
"0.8896",
"0.5435",
"0.505",
"0.5435"
],
[
"0.0809",
"0.5454",
"0.468",
"0.5454",
"0.468",
"0.5596",
"0.0809",
"0.5596"
],
[
"0.505",
"0.547",
"0.8906",
"0.547",
"0.8906",
"0.56",
"0.505",
"0.56"
],
[
"0.08417",
"0.5625",
"0.468",
"0.5625",
"0.468",
"0.5728",
"0.08417",
"0.5728"
],
[
"0.5034",
"0.5625",
"0.8906",
"0.5625",
"0.8906",
"0.5757",
"0.5034",
"0.5757"
],
[
"0.0809",
"0.5757",
"0.468",
"0.5757",
"0.468",
"0.589",
"0.0809",
"0.589"
],
[
"0.501",
"0.58",
"0.892",
"0.58",
"0.892",
"0.601",
"0.501",
"0.601"
],
[
"0.0809",
"0.5913",
"0.4673",
"0.5913",
"0.4673",
"0.605",
"0.0809",
"0.605"
],
[
"0.0809",
"0.607",
"0.4673",
"0.607",
"0.4673",
"0.6206",
"0.0809",
"0.6206"
],
[
"0.5034",
"0.601",
"0.701",
"0.601",
"0.701",
"0.6143",
"0.5034",
"0.6143"
],
[
"0.0809",
"0.622",
"0.4673",
"0.622",
"0.4673",
"0.6353",
"0.0809",
"0.6353"
],
[
"0.5034",
"0.625",
"0.773",
"0.625",
"0.773",
"0.638",
"0.5034",
"0.638"
],
[
"0.0801",
"0.636",
"0.1544",
"0.636",
"0.1544",
"0.651",
"0.0801",
"0.651"
],
[
"0.524",
"0.6465",
"0.892",
"0.6465",
"0.892",
"0.661",
"0.524",
"0.661"
],
[
"0.0801",
"0.6655",
"0.2263",
"0.6655",
"0.2263",
"0.6787",
"0.0801",
"0.6787"
],
[
"0.505",
"0.6636",
"0.8906",
"0.6636",
"0.8906",
"0.677",
"0.505",
"0.677"
],
[
"0.5034",
"0.678",
"0.8896",
"0.677",
"0.8896",
"0.6914",
"0.5034",
"0.6924"
],
[
"0.10046",
"0.6914",
"0.468",
"0.6914",
"0.468",
"0.705",
"0.10046",
"0.705"
],
[
"0.505",
"0.6934",
"0.888",
"0.6934",
"0.888",
"0.707",
"0.505",
"0.707"
],
[
"0.0825",
"0.7056",
"0.468",
"0.7056",
"0.468",
"0.7197",
"0.0825",
"0.7197"
],
[
"0.5034",
"0.709",
"0.8896",
"0.709",
"0.8896",
"0.7227",
"0.5034",
"0.7227"
],
[
"0.0801",
"0.7197",
"0.468",
"0.72",
"0.468",
"0.735",
"0.0801",
"0.7344"
],
[
"0.5024",
"0.7236",
"0.888",
"0.7227",
"0.888",
"0.7363",
"0.5024",
"0.7373"
],
[
"0.0972",
"0.7363",
"0.3162",
"0.7363",
"0.3162",
"0.75",
"0.0972",
"0.75"
],
[
"0.505",
"0.738",
"0.8906",
"0.738",
"0.8906",
"0.752",
"0.505",
"0.752"
],
[
"0.1013",
"0.752",
"0.4673",
"0.752",
"0.4673",
"0.765",
"0.1013",
"0.765"
],
[
"0.505",
"0.754",
"0.8906",
"0.754",
"0.8906",
"0.7676",
"0.505",
"0.7676"
],
[
"0.08417",
"0.768",
"0.4673",
"0.768",
"0.4673",
"0.779",
"0.08417",
"0.779"
],
[
"0.5034",
"0.768",
"0.6763",
"0.768",
"0.6763",
"0.782",
"0.5034",
"0.782"
],
[
"0.0809",
"0.781",
"0.4673",
"0.782",
"0.4673",
"0.7954",
"0.0809",
"0.7944"
],
[
"0.0801",
"0.7954",
"0.4673",
"0.797",
"0.4673",
"0.811",
"0.0801",
"0.81"
],
[
"0.5537",
"0.799",
"0.8057",
"0.799",
"0.8057",
"0.813",
"0.5537",
"0.813"
],
[
"0.0825",
"0.811",
"0.4673",
"0.811",
"0.4673",
"0.8247",
"0.0825",
"0.8247"
],
[
"0.0825",
"0.827",
"0.4673",
"0.827",
"0.4673",
"0.8403",
"0.0825",
"0.8403"
],
[
"0.6543",
"0.8237",
"0.839",
"0.8213",
"0.839",
"0.8423",
"0.6543",
"0.8447"
],
[
"0.08417",
"0.8433",
"0.4673",
"0.8433",
"0.4673",
"0.854",
"0.08417",
"0.854"
],
[
"0.0809",
"0.857",
"0.468",
"0.857",
"0.468",
"0.8706",
"0.0809",
"0.8706"
],
[
"0.5034",
"0.858",
"0.8896",
"0.858",
"0.8896",
"0.872",
"0.5034",
"0.872"
],
[
"0.0825",
"0.872",
"0.468",
"0.872",
"0.468",
"0.8853",
"0.0825",
"0.8853"
],
[
"0.5063",
"0.8735",
"0.8896",
"0.8735",
"0.8896",
"0.884",
"0.5063",
"0.884"
],
[
"0.0809",
"0.887",
"0.468",
"0.887",
"0.468",
"0.901",
"0.0809",
"0.901"
],
[
"0.5034",
"0.888",
"0.8906",
"0.888",
"0.8906",
"0.9014",
"0.5034",
"0.9014"
]
] | [
"行建模已被证明是深度神经网络识别未知数的次优解",
"和重建的有用性。具体来说，作者建议使用无监督训练",
"决方案。因此，引入了极值理论，以更好地适应这些完",
"产生的表示将已知与未知分开，同时使用通过监督学",
"全基于监督学习的判别模型，以适应开放集设置。基本",
"习过程计算的表示保持模型的判别能力",
"思想是校准Softmax分数，以便估计未知32，2]的概",
"C2AEE[19]通过基于极值理论对模型的重建误差进",
"率。除了深度判别模型外，专注于通过无监督学习来学",
"行建模，引人了一种基于类条件VAE的开放集识别和",
"习有效的潜在特征表示的深度生成模型已被广泛用于",
"未知检测的体系结构。Sun etal.E[28]最近认为，基",
"开放集识别任务中，并相继获得了成功「16，19，21，28",
"于VAE的开放集识别体系结构的一个缺点是对已知类",
"特别是，变分自动编码器（VAE）是一种典型的概率",
"的实例的判别能力不足。因此，作者采用条件高斯分布",
"生成模型，非常适合检测未知数，因为它能够学习潜在",
"VAE模型来学习已知类的条件分布并拒绝未知数。E",
"空间中的低维表示，不仅支持输入重建，而且可以近似",
"[33]中提出了一种不同的方法，其中将归一化流用于已",
"指定的先验分布。另一方面，基于VAE的模型可能不",
"知样本的密度估计。具体来说，作者提出了一种使用",
"足以有效地识别已知类别，因为所有特征表示仅遵循",
"CNN编码器和可逆神经网络的架构，该架构可以共同",
"一种分布。为此，我们采用条件VAE（CVAE），它使",
"学习输人的密度。但是，本文未讨论的一个潜在问题是",
"用多个先验分布来对已知类进行建模，并间接对未知",
"CNN编码器没有双射特性，这对于采用变量变化公式",
"对应物进行建模。此外，我们建议用概率胶囊表示输入",
"进行密度评估至关重要。此外，3在原型学习中引入",
"样本，因为它们已经证明了其表示能力E[22，24]",
"了互易点的概念来管理开放空间。尽管这项工作在拒",
"提出了胶囊网络（CapsNet）[24]作为卷积神经网络",
"绝来自已知样本的不同数据集的未知数方面表现出出",
"（CNN）的替代方案。与CNN的标量神经元不同，胶",
"色的性能，但当未知样本的来源与已知样本相同时，未",
"囊将一组神经元集合在一起以接受和输出向量。激活",
"知检测能力会下降",
"胶囊的向量表示特定对象的各种属性，例如位置，大",
"小，方向，纹理等。从本质上讲，CapsNet可以看作是",
"3.Preliminaries",
"通过分布式表示对对象进行编码的编码器，这比通过",
"3.1. The Open Set Recognition Problem",
"激活高维空间中的单个神经元对对象进行编码要有效",
"得多。。此外，CapsNet已成功用于在类似于开放集识",
"在开放集识别问题中，模型必须对可能属于训练期",
"别【18]的任务设置中检测假图像和视频。这促使我们",
"间未看到的类的测试样本进行分类。给定一个分类数",
"结合CVAE为开放集识别问题设计了一种新颖的胶囊",
"据集D={(ai,i)，...,αn,n))，使得iE是",
"网络架构，称为CVAECapOSR，如图E1所示",
"输入样本，yiE{1，...,K}是相应的类别标签，开放集",
"本文的贡献有三个方面：i）E我们提出了一种基",
"问题在于K+U类中测试样本的分类，其中U是未知",
"于CapsNet的新颖的开放集识别框架，并展示了其",
"类的数量。在文献中，用于训练的数据集称为封闭数据",
"在学习已知类的有效表示方面的优势。i）我们集成了",
"集，同时在评估期间使用的包含来自看不见的类的样",
"CapsNet和条件VAE。与鼓励潜在表示近似单个先验",
"本的数据集称为开放数据集。为了在评估期间量化数",
"分布的一般VAE相比，我们的模型利用了多个先验（i。",
"据集的开放性，在[25，16]之后，我们将开放性度量视",
"e。每个类一个），并且它迫使潜在表示遵循由输人样本",
"为O=1-V，其中K和M=+U分别是在训",
"的类选择的高斯先验。ii)E我们对用于开放集识别的",
"所有标准数据集进行了广泛的实验，获得了非常有竞",
"练和测试期间观察到的类数",
"争力的结果，在某些情况下，其性能大大超过了以前的",
"3.2. Conditional VAE Formulation",
"最新方法。",
"条件变分Eutoencoder（CVAE）直接源自VAE模",
"2. Related Work",
"型[11]，其目标基于给定标签y的数据α的条件密度",
"p(rly)的估计。它是最强大的概率生成模型之一，因为",
"开集识别问题是由[25]引人的，最初被形式化为基",
"它的理论优雅，强大的框架兼容性和高效的流形表示，",
"于支持向量机（SVM）的约束最小化问题，而随后的工",
"CVAE通常包括一个编码器，该编码器将输人&和类",
"作集中在其他更传统的方法上，例如极值理论E[9，26]",
"9映射到潜在变量2上的预定分布，以及一个解码器",
"稀疏表示E[34]和最近邻E[10]",
"该解码器给定潜在变量z和类y尝试重建输人。在",
"继深度学习在许多计算机视觉任务中取得的成功之",
"训练过程中，通过最小化数据条件密度的负变分下界",
"后，在E2中首次引人了深度网络用于开放集识别，其",
"来训练模型，定义如下：",
"中通过使用Weibull分布模型校准每个类的Softmax",
"概率来提出Openmax函数。随后，[5]通过引人生成",
"C, y;0,Φ) = DKLqs(z) lp(zy)",
"对抗网络将Openmax扩展到G-Openmax，在该网络",
"中，生成器生成新颖类别的合成样本，而鉴别器学习未",
"-Eg(z|n) log po(a|z, y)",
"知类的显式表示。在16中采用了类似的策略，该策",
"略提出了一种基于生成对抗网络的数据增强技术，称",
"其中q(z）表示编码器的后验，pe(zy）表示以y为",
"为反事实图像生成。最近，Yoshihashietal.[32分析",
"条件的潜在变量之的先验分布。损失函数中的第一项",
"并证明了在开放集场景中联合训练深度网络进行分类",
"是一个正则化器，它强制近似后验分布q(az）接近先"
] | {
"arxiv_id": "2104.09159",
"language": "zh",
"page_idx": 1
} |
||
2104.09159-zh-2 | 2104.09159-zh-2.jpg | 验分布$p_\theta ( \boldsymbol{z}| y)$ ,而第二项是链式编码-解码过程的平均重构误差。使用无条件先验分布的原始 VAE [11] 假定$p_\theta(\boldsymbol{z})$是各向同性多元高斯$\mathcal{N}(\boldsymbol{0},\boldsymbol{I})$ ,而$q_\phi(\boldsymbol{z}|\boldsymbol{x})$是一般多元高斯$\mathcal{N}(\boldsymbol{\mu},\boldsymbol{\sigma^2})$ 。在这些假设下，给定$K$维$z$的 KL 发散项可以封闭形式计算，并表示为：
$$D_{\mathrm{KL}}\Big[q_\phi(\boldsymbol{z}|\boldsymbol{x})\|p_\theta(\boldsymbol{z})\Big]=-\frac{1}{2}\sum_i^K(1+\log(\boldsymbol{\sigma}_i^2)-\boldsymbol{\mu}_i^2-\boldsymbol{\sigma}_i^2).$$
对于 CVAE, 可以仅使用可处理的潜在先验分布$p(\boldsymbol{z}|y)$
$\mathbb{E}[20]$来计算或估计 KL 发散项。
# 3.3. CapsNet Formulation
Hinton 等人提出的胶囊网络。[8] 是一个由两个卷积层和两个胶囊层组成的浅架构。第一卷积运算将输入图像$x$的像素强度转换为主局部特征图，而第二卷积层产生主胶囊$u_i$ 。每个胶囊对应于一组矩阵，这些矩阵旋转主胶囊以预测姿势变换$\hat{\boldsymbol{u}}_{j|i}=W_{ij}\boldsymbol{u}_i$ 。之后，将用于分类的数字胶囊$v_j$作为主要胶囊$v_j=\sum_ic_{ij}\hat{u}_{j|i}$ 的加权和产生，其中系数$c_ij$由动态路由算法(DR)确定，其中将主要胶囊与数字胶囊进行比较。对于 DR 的第$t$次迭代，系数由
$$c_i^{(t+1)}=Softmax(b_i^{(t+1)}),\:b_{ij}^{(t+1)}=b_{ij}^{(t)}+\hat{\boldsymbol{u}}_{j|i}\cdot\boldsymbol{v}_j^{(t)}.$$
更新。对于胶囊的所有层，使用挤压函数将非线性引人
并将胶囊向量的长度缩小到[0,1] 中，
$$S\boldsymbol{quash}(\boldsymbol{v})=\frac{\|\boldsymbol{v}\|^2}{1+\|\boldsymbol{v}\|^2}\frac{\boldsymbol{v}}{\|\boldsymbol{v}\|}.$$
这样，胶囊的范数代表 XMAX 中存在特定特征的概
率。
## $4.{\mathrm{Proposed~Method}}$
## 4.1. Model Architecture
我们的模型，如图$\color{red}{\mathrm{E}}\color{blue}{\mathrm{E}}\color{red}{\mathrm{E}}$所示，基于具有$K$不同高斯先验分布的 CVAE, 每个已知类一个。给定输人图像$x$及其相应的标签$y$ ,编码器处理$x$产生特征表示$x_f$。之后，胶囊网络计算在学习过程中推向条件先验$p(\boldsymbol{z}|y)=\boldsymbol{T}_y$的分布$q(\boldsymbol z|\boldsymbol{x})$ 。使用$q(\boldsymbol{z}|\boldsymbol{x})$和所有目标之间的距离信息，我们估计类$\hat{y}$,并使用重新参数化技巧，我们从$q(\boldsymbol z|\boldsymbol{x})$采样$z$ 。给定$\hat{y}$和$z$ ,我们通过解码器计算重建$\hat{x}$ ,该解码器是使用转置卷积的卷积神经网络。在对我们的模型的计算进行一般描述之后， 我们现在一步一步地深人介绍体系结构
编码阶段。编码阶段涉及的块是编码器和胶囊网络。编码器是一个卷积神经网络，它处理输入图像$x\in\mathbb{R}^{C\times H\times W}$,产生特征 $x_f\in\mathbb{R}^{d_c\times d_h\times d_w}$。然后，类似于$\operatorname{E[24]}$ , 胶囊网络通过使用动态路由算法计算主胶囊$x_{\mathrm{pc}}\in \mathbb{R} ^{f_{1}\times ( \frac {d_{c}}{f_{1}}d_{h}d_{w}) }$, 然后计算数字胶囊$x_\mathrm{dc}\in\mathbb{R}^{K\times f_{2}}$ 来处理特征表示$x_f$ 。我们用$f_1$表示主要胶囊的尺寸，
用$f_2$表示数字胶囊的尺寸。给定胶囊$x_\mathrm{dc}$,我们通过应用具有$d$输出单元的胶囊式全连接层来计算胶囊分布$\mathcal{C}=q(\boldsymbol{z}|\boldsymbol{x})$的均值$\boldsymbol z_\mu\in\mathbb{R}^K\times d$和方差$z_\sigma^2\in\mathbb{R}^{K\times d}$ 。这样，概率胶囊网络产生$K$平均胶囊$\{z_\mu^{(k)}\}_{k=1}^K$和$K$方差胶囊$\{z_\sigma^2^{(k)}\}_{k=1}^K$,每个大小为$d$
对比变分阶段。我们将每个 CVAE 先验目标$p(\boldsymbol{z}|y=$ $k)$设计为具有可学习均值向量$\tilde{\boldsymbol{\mu}}_k\in\mathbb{R}^{Kd}$和具有$1\leq k\leq K$的可学习对角协方差矩阵$\tilde{\Sigma}_k\in\mathbb{R}^{Kd\times Kd}$ 的高斯分布$T_k=\mathcal{N}\left(\tilde{\boldsymbol{\mu}}_k,\tilde{\boldsymbol{\Sigma}}_k\right)$ 。为了简化模型框架的表示法，我们将目标$T_k$视为高斯分布，该高斯分布由$\boldsymbol{\mu}_{k}\in\mathbb{R}^{K\times d}$定义为$\tilde{\boldsymbol{\mu}}_{k}$的重塑版本，而$\Sigma_k\in\mathbb{R}^{K\times d}$定义为$\tilde{\Sigma}_k$的重塑对角线。为了将来自同一类的输人实例映射到潜在空间的紧凑和分离的区域，在学习过程中，我们让概率胶囊$\mathcal{C}$被$y$目标分布$T_y$吸引，同时我们让所有其他目标$T_{\neq y}$被$\mathcal{C}$排斥。使用这种对比策略，我们鼓励编码表示属于潜在空间的正确区域，同时保持所有目标彼此相距足够远。然后，我们将输入样本$x$的类$\hat{y}$估计为：
$$\hat{y}_k=p(y=k|\boldsymbol{x})=\frac{e^{-\gamma d(\boldsymbol{C},\boldsymbol{T}_k)}}{\sum_{j=1}^Ke^{-\gamma d(\boldsymbol{C},\boldsymbol{T}_j)}},\quad1\leq k\leq K,$$
其中
$$d(\mathcal{C},\boldsymbol{T}_i)=\frac{1}{K}\sum_{k=1}^{K}D_{\mathrm{KL}}\Big[\mathcal{C}^{(k)}||\boldsymbol{T}_i^{(k)}\Big],$$
是概率胶囊$\mathcal{C}$与目标$T_i$之间的距离，而$\gamma$的分配参数是 a。通过这种方式，我们估计$x$属于$k$类的概率， 考虑到胶囊$\left\{\mathcal{C}^{(k)}\right\}_{k=1}^K$的整个配置，而不仅仅是 [24] 中所做的单个最活跃的胶囊
解码阶段。给定类估计$\hat{y}\in\mathbb{R}^K$,我们计算其可学习嵌入$\hat{y}_e=$Embedding$(\arg\max_k(\hat{y}_k))\in\mathbb{R}^d$,并给定采样的潜在胶囊$z\in\mathbb{R}^{K\times d}$,我们通过解码器从$z_y\in\mathbb{R}^{K\times d}$开始使用$z_y^(k)=\hat{y}_e+z^{(k)}$计算重构$\hat{x}$ 。解$\dot{\text{码器是具有转置卷积的卷积神经网络，其遵循编码器的}}$ 对称结构。与$\mathbb{E}[23]$类似，我们实现了$x_\ell$与$1\leq\ell\leq4$ 的横向连接，从编码器的内部功能到解码器，在训练期间随机退出，以使解码器更少依赖于编码器的内部表示。
# 4.2. Training
我们在封闭数据集上训练模型，在单个输入样本
$(\boldsymbol{x},y)$的学习过程中，我们最小化以下损失函数：
$$\mathcal{L}(\boldsymbol{x},y)=\mathcal{L}_\mathrm{KL}(\boldsymbol{x},y)+\alpha\mathcal{L}_\mathrm{contr}(\boldsymbol{x},y)+\beta\mathcal{L}_\mathrm{rec}(\boldsymbol{x}),$$
其中
$$\mathcal{L}_{\mathrm{KL}}(\boldsymbol{x},y)=d(\mathcal{C},\mathrm{sg}[\boldsymbol{T}_{y}]),$$
(2)
$$\mathcal{L}_{\mathrm{contr}}(\boldsymbol{x},y)=\frac{1}{K-1}\sum_{k\neq y}^{K}\left[m_{k}-d(\mathrm{sg}\big[\boldsymbol{C}\big],\boldsymbol{T}_{k})\right]^{+},$$
<NEXLINE$\_$ | [
[
"0.0801",
"0.09216",
"0.468",
"0.09155",
"0.468",
"0.1061",
"0.0801",
"0.1067"
],
[
"0.5024",
"0.09216",
"0.8906",
"0.09155",
"0.8906",
"0.1061",
"0.5024",
"0.1067"
],
[
"0.0809",
"0.108",
"0.468",
"0.108",
"0.468",
"0.1218",
"0.0809",
"0.1218"
],
[
"0.5034",
"0.108",
"0.8896",
"0.108",
"0.8896",
"0.1218",
"0.5034",
"0.1218"
],
[
"0.0809",
"0.1237",
"0.4673",
"0.1237",
"0.4673",
"0.137",
"0.0809",
"0.137"
],
[
"0.5",
"0.1218",
"0.892",
"0.1187",
"0.892",
"0.1364",
"0.5",
"0.1395"
],
[
"0.08826",
"0.1383",
"0.468",
"0.1383",
"0.468",
"0.1515",
"0.08826",
"0.1515"
],
[
"0.5034",
"0.1401",
"0.8906",
"0.1401",
"0.8906",
"0.1547",
"0.5034",
"0.1547"
],
[
"0.0801",
"0.154",
"0.415",
"0.154",
"0.415",
"0.1674",
"0.0801",
"0.1674"
],
[
"0.501",
"0.1547",
"0.786",
"0.1547",
"0.786",
"0.1724",
"0.501",
"0.1724"
],
[
"0.523",
"0.1736",
"0.8896",
"0.1749",
"0.8896",
"0.1881",
"0.523",
"0.1875"
],
[
"0.2573",
"0.1818",
"0.298",
"0.1895",
"0.2942",
"0.2008",
"0.2532",
"0.1932"
],
[
"0.286",
"0.1831",
"0.2917",
"0.1831",
"0.2917",
"0.1881",
"0.286",
"0.1881"
],
[
"0.2484",
"0.1901",
"0.4697",
"0.1831",
"0.4714",
"0.2084",
"0.25",
"0.2153"
],
[
"0.5034",
"0.1881",
"0.8896",
"0.1881",
"0.8896",
"0.2026",
"0.5034",
"0.2026"
],
[
"0.505",
"0.202",
"0.8936",
"0.1995",
"0.8936",
"0.2172",
"0.505",
"0.2197"
],
[
"0.08417",
"0.1913",
"0.2181",
"0.1913",
"0.2181",
"0.209",
"0.08417",
"0.209"
],
[
"0.505",
"0.2216",
"0.8906",
"0.2216",
"0.8906",
"0.2361",
"0.505",
"0.2361"
],
[
"0.0825",
"0.2329",
"0.4673",
"0.2329",
"0.4673",
"0.2468",
"0.0825",
"0.2468"
],
[
"0.505",
"0.2412",
"0.8896",
"0.2412",
"0.8896",
"0.2551",
"0.505",
"0.2551"
],
[
"0.0809",
"0.2474",
"0.3145",
"0.2474",
"0.3145",
"0.2615",
"0.0809",
"0.2615"
],
[
"0.501",
"0.252",
"0.8936",
"0.2507",
"0.8936",
"0.2715",
"0.501",
"0.2727"
],
[
"0.0801",
"0.2734",
"0.2852",
"0.2734",
"0.2852",
"0.2874",
"0.0801",
"0.2874"
],
[
"0.5063",
"0.2708",
"0.8906",
"0.2708",
"0.8906",
"0.2854",
"0.5063",
"0.2854"
],
[
"0.505",
"0.286",
"0.8906",
"0.286",
"0.8906",
"0.2998",
"0.505",
"0.2998"
],
[
"0.10046",
"0.2966",
"0.468",
"0.2966",
"0.468",
"0.31",
"0.10046",
"0.31"
],
[
"0.505",
"0.3018",
"0.8906",
"0.3018",
"0.8906",
"0.315",
"0.505",
"0.315"
],
[
"0.0825",
"0.3132",
"0.4656",
"0.3132",
"0.4656",
"0.324",
"0.0825",
"0.324"
],
[
"0.505",
"0.3176",
"0.8896",
"0.3176",
"0.8896",
"0.3308",
"0.505",
"0.3308"
],
[
"0.0809",
"0.327",
"0.468",
"0.327",
"0.468",
"0.3403",
"0.0809",
"0.3403"
],
[
"0.5034",
"0.332",
"0.8896",
"0.332",
"0.8896",
"0.3452",
"0.5034",
"0.3452"
],
[
"0.0809",
"0.3423",
"0.468",
"0.3423",
"0.468",
"0.356",
"0.0809",
"0.356"
],
[
"0.5034",
"0.348",
"0.8896",
"0.348",
"0.8896",
"0.361",
"0.5034",
"0.361"
],
[
"0.0809",
"0.358",
"0.468",
"0.358",
"0.468",
"0.3718",
"0.0809",
"0.3718"
],
[
"0.501",
"0.361",
"0.6284",
"0.3623",
"0.6284",
"0.377",
"0.501",
"0.3757"
],
[
"0.0801",
"0.3687",
"0.468",
"0.3706",
"0.468",
"0.3882",
"0.0801",
"0.3865"
],
[
"0.0809",
"0.3865",
"0.4673",
"0.3865",
"0.4673",
"0.401",
"0.0809",
"0.401"
],
[
"0.675",
"0.3901",
"0.7466",
"0.3838",
"0.749",
"0.3984",
"0.6763",
"0.404"
],
[
"0.5186",
"0.3997",
"0.638",
"0.3997",
"0.638",
"0.4136",
"0.5186",
"0.4136"
],
[
"0.794",
"0.3997",
"0.875",
"0.3997",
"0.875",
"0.411",
"0.794",
"0.411"
],
[
"0.634",
"0.4028",
"0.6396",
"0.4028",
"0.6396",
"0.4077",
"0.634",
"0.4077"
],
[
"0.0809",
"0.4016",
"0.468",
"0.4016",
"0.468",
"0.4153",
"0.0809",
"0.4153"
],
[
"0.653",
"0.4092",
"0.7754",
"0.401",
"0.777",
"0.4204",
"0.6562",
"0.428"
],
[
"0.0801",
"0.4167",
"0.2279",
"0.4172",
"0.2279",
"0.432",
"0.0801",
"0.4312"
],
[
"0.0784",
"0.44",
"0.4673",
"0.4414",
"0.4673",
"0.4666",
"0.0784",
"0.4653"
],
[
"0.5024",
"0.4382",
"0.5376",
"0.4382",
"0.5376",
"0.4539",
"0.5024",
"0.4539"
],
[
"0.671",
"0.4622",
"0.738",
"0.4602",
"0.739",
"0.4778",
"0.6724",
"0.4797"
],
[
"0.6772",
"0.457",
"0.689",
"0.457",
"0.689",
"0.4602",
"0.6772",
"0.4602"
],
[
"0.576",
"0.4666",
"0.6543",
"0.4666",
"0.6543",
"0.4778",
"0.576",
"0.4778"
],
[
"0.0809",
"0.4766",
"0.4673",
"0.4766",
"0.4673",
"0.4905",
"0.0809",
"0.4905"
],
[
"0.653",
"0.473",
"0.6763",
"0.473",
"0.6763",
"0.4841",
"0.653",
"0.4841"
],
[
"0.669",
"0.4797",
"0.697",
"0.4797",
"0.697",
"0.4912",
"0.669",
"0.4912"
],
[
"0.0809",
"0.4905",
"0.343",
"0.4912",
"0.343",
"0.506",
"0.0809",
"0.505"
],
[
"0.505",
"0.5",
"0.8896",
"0.5",
"0.8896",
"0.513",
"0.505",
"0.513"
],
[
"0.2917",
"0.5176",
"0.367",
"0.5176",
"0.367",
"0.535",
"0.2917",
"0.535"
],
[
"0.5034",
"0.5146",
"0.888",
"0.5146",
"0.888",
"0.528",
"0.5034",
"0.528"
],
[
"0.5034",
"0.528",
"0.8896",
"0.53",
"0.8896",
"0.5444",
"0.5034",
"0.5425"
],
[
"0.1683",
"0.518",
"0.3782",
"0.534",
"0.3757",
"0.558",
"0.165",
"0.5425"
],
[
"0.505",
"0.5444",
"0.701",
"0.5444",
"0.701",
"0.558",
"0.505",
"0.558"
],
[
"0.0809",
"0.5645",
"0.468",
"0.5645",
"0.468",
"0.5776",
"0.0809",
"0.5776"
],
[
"0.523",
"0.5596",
"0.8896",
"0.5596",
"0.8896",
"0.5737",
"0.523",
"0.5737"
],
[
"0.0801",
"0.58",
"0.1095",
"0.58",
"0.1095",
"0.5923",
"0.0801",
"0.5923"
],
[
"0.505",
"0.5747",
"0.8906",
"0.5747",
"0.8906",
"0.589",
"0.505",
"0.589"
],
[
"0.5024",
"0.5854",
"0.892",
"0.5854",
"0.892",
"0.606",
"0.5024",
"0.606"
],
[
"0.5024",
"0.605",
"0.6787",
"0.603",
"0.6787",
"0.6206",
"0.5024",
"0.6226"
],
[
"0.664",
"0.6016",
"0.892",
"0.606",
"0.892",
"0.6235",
"0.664",
"0.619"
],
[
"0.0809",
"0.608",
"0.259",
"0.608",
"0.259",
"0.6226",
"0.0809",
"0.6226"
],
[
"0.505",
"0.622",
"0.8906",
"0.622",
"0.8906",
"0.6353",
"0.505",
"0.6353"
],
[
"0.0809",
"0.6333",
"0.2695",
"0.6333",
"0.2695",
"0.6465",
"0.0809",
"0.6465"
],
[
"0.5034",
"0.636",
"0.8906",
"0.636",
"0.8906",
"0.6494",
"0.5034",
"0.6494"
],
[
"0.5034",
"0.6514",
"0.8896",
"0.6514",
"0.8896",
"0.6655",
"0.5034",
"0.6655"
],
[
"0.1013",
"0.6577",
"0.4656",
"0.6577",
"0.4656",
"0.6685",
"0.1013",
"0.6685"
],
[
"0.0809",
"0.672",
"0.468",
"0.672",
"0.468",
"0.685",
"0.0809",
"0.685"
],
[
"0.5034",
"0.6675",
"0.8906",
"0.6675",
"0.8906",
"0.681",
"0.5034",
"0.681"
],
[
"0.0801",
"0.686",
"0.468",
"0.686",
"0.468",
"0.6997",
"0.0801",
"0.6997"
],
[
"0.5034",
"0.6846",
"0.5293",
"0.6846",
"0.5293",
"0.6963",
"0.5034",
"0.6963"
],
[
"0.0801",
"0.702",
"0.4673",
"0.7007",
"0.4673",
"0.7153",
"0.0801",
"0.7163"
],
[
"0.5024",
"0.705",
"0.6094",
"0.7085",
"0.6094",
"0.7227",
"0.5024",
"0.7197"
],
[
"0.0809",
"0.7173",
"0.4673",
"0.7173",
"0.4673",
"0.731",
"0.0809",
"0.731"
],
[
"0.0809",
"0.732",
"0.4673",
"0.732",
"0.4673",
"0.7456",
"0.0809",
"0.7456"
],
[
"0.523",
"0.73",
"0.8906",
"0.73",
"0.8906",
"0.7437",
"0.523",
"0.7437"
],
[
"0.0809",
"0.7476",
"0.468",
"0.7476",
"0.468",
"0.761",
"0.0809",
"0.761"
],
[
"0.5024",
"0.7456",
"0.847",
"0.744",
"0.847",
"0.7583",
"0.5024",
"0.759"
],
[
"0.0809",
"0.762",
"0.468",
"0.762",
"0.468",
"0.776",
"0.0809",
"0.776"
],
[
"0.0825",
"0.7764",
"0.468",
"0.7764",
"0.468",
"0.7905",
"0.0825",
"0.7905"
],
[
"0.5186",
"0.7734",
"0.8906",
"0.7734",
"0.8906",
"0.787",
"0.5186",
"0.787"
],
[
"0.0809",
"0.7925",
"0.3594",
"0.7925",
"0.3594",
"0.806",
"0.0809",
"0.806"
],
[
"0.5034",
"0.8003",
"0.536",
"0.8003",
"0.536",
"0.813",
"0.5034",
"0.813"
],
[
"0.1013",
"0.8096",
"0.4673",
"0.8096",
"0.4673",
"0.8193",
"0.1013",
"0.8193"
],
[
"0.609",
"0.8164",
"0.786",
"0.8164",
"0.786",
"0.83",
"0.609",
"0.83"
],
[
"0.8677",
"0.8164",
"0.8906",
"0.8164",
"0.8906",
"0.83",
"0.8677",
"0.83"
],
[
"0.0809",
"0.8228",
"0.4673",
"0.8228",
"0.4673",
"0.836",
"0.0809",
"0.836"
],
[
"0.0784",
"0.8335",
"0.4697",
"0.8335",
"0.4697",
"0.854",
"0.0784",
"0.854"
],
[
"0.638",
"0.8423",
"0.6587",
"0.8423",
"0.6587",
"0.856",
"0.638",
"0.856"
],
[
"0.6772",
"0.8403",
"0.687",
"0.8403",
"0.687",
"0.8477",
"0.6772",
"0.8477"
],
[
"0.517",
"0.851",
"0.614",
"0.853",
"0.614",
"0.867",
"0.517",
"0.8657"
],
[
"0.6665",
"0.8477",
"0.8906",
"0.851",
"0.8906",
"0.8687",
"0.6665",
"0.8657"
],
[
"0.6216",
"0.86",
"0.6733",
"0.86",
"0.6733",
"0.872",
"0.6216",
"0.872"
],
[
"0.0809",
"0.853",
"0.468",
"0.853",
"0.468",
"0.866",
"0.0809",
"0.866"
],
[
"0.0768",
"0.8643",
"0.4714",
"0.863",
"0.4714",
"0.888",
"0.0768",
"0.8896"
],
[
"0.6665",
"0.867",
"0.694",
"0.867",
"0.694",
"0.88",
"0.6665",
"0.88"
],
[
"0.0825",
"0.887",
"0.4697",
"0.887",
"0.4697",
"0.9014",
"0.0825",
"0.9014"
],
[
"0.5063",
"0.888",
"0.6006",
"0.888",
"0.6006",
"0.8984",
"0.5063",
"0.8984"
]
] | [
"验分布pe(zy），而第二项是链式编码-解码过程的平",
"用f2表示数字胶囊的尺寸。给定胶囊adc，我们通过",
"均重构误差。使用无条件先验分布的原始VAE【11]假",
"应用具有d输出单元的胶囊式全连接层来计算胶囊分",
"定pe(z）是各向同性多元高斯N(0,I），而q(zα）是",
"布C=q(za)的均值zuRKxd和方差z2RKxd",
"一般多元高斯N（μ,α2）。在这些假设下，给定K维",
"。这样，概率胶囊网络产生K平均胶囊【z(k}K，和",
"z的KL发散项可以封闭形式计算，并表示为：",
"K方差胶囊{z(K/，每个大小为 d",
"对比变分阶段。我们将每个CVAE先验目标p(zy=",
"1<",
"PY",
"(1+1og(o)-2).",
"k）设计为具有可学习均值向量ikERKd和具有",
"1≤≤K的可学习对角协方差矩阵RKdxkd",
"DkL(z)/pe(z)",
"的高斯分布T=N（ik,）。为了简化模型框架的",
"对于CVAE，可以仅使用可处理的潜在先验分布p(zy)",
"表示法，我们将目标T视为高斯分布，该高斯分布由",
"E[20]来计算或估计KL发散项。",
"ukERKxd定义为的重塑版本，而ERKxd定",
"3.3. CapsNet Formulation",
"义为的重塑对角线。为了将来自同一类的输人实",
"例映射到潜在空间的紧凑和分离的区域，在学习过程",
"Hinton等人提出的胶囊网络。[8]是一个由两个卷积",
"中，我们让概率胶囊C被目标分布Tu吸引，同时",
"层和两个胶囊层组成的浅架构。第一卷积运算将输入图",
"我们让所有其他目标T￥y被C排斥。使用这种对比策",
"像&的像素强度转换为主局部特征图，而第二卷积层",
"略，我们鼓励编码表示属于潜在空间的正确区域，同时",
"产生主胶囊ui。每个胶囊对应于一组矩阵，这些矩阵",
"保持所有目标彼此相距足够远。然后，我们将输入样本",
"旋转主胶囊以预测姿势变换ujli=Wijui。之后，将",
"a的类估计为：",
"用于分类的数字胶囊u,作为主要胶囊u=Z,Cijujli",
"的加权和产生，其中系数cii由动态路由算法（DR）确",
"e-d(c,T)",
"yk =p(y=ka)",
"1<k<K,",
"1",
"定，其中将主要胶囊与数字胶囊进行比较。对于DR的",
"E=e-d(c,T)",
"第t次迭代，系数由",
"c($+1) = Softmaa(b($+1), bg+1) =bg) + ajl g).",
"其中",
"EDKL",
"11",
"d(C,T) = ",
"更新。对于胶囊的所有层，使用挤压函数将非线性引人",
"K",
"k=1",
"并将胶囊向量的长度缩小到[0,1]中,",
"是概率胶囊C与目标T之间的距离，而的分配参",
"u2",
"数是a。通过这种方式，我们估计&属于k类的概率",
"考虑到胶囊{c())K-，的整个配置，而不仅仅是[24]",
"Squash(o) = 1+o l",
"中所做的单个最活跃的胶囊",
"这样，胶囊的范数代表XMAX中存在特定特征的概",
"解码阶段。给定类估计gERK，我们计算其可",
"率。",
"学习嵌人 ye=Embedding(argmax,(yk))E Rd，并",
"给定采样的潜在胶囊zeRKxd，我们通过解码器从",
"zyERKxd开始使用z(k",
"(k)=ge+z(k)计算重构。解",
"4.Proposed Method",
"码器是具有转置卷积的卷积神经网络，其遵循编码器的",
"4.1. Model Architecture",
"对称结构。与E[23]类似，我们实现了与1<l<4",
"的横向连接，从编码器的内部功能到解码器，在训练期",
"我们的模型，如图E2所示，基于具有K不同高斯",
"先验分布的CVAE，每个已知类一个。给定输人图像",
"间随机退出，以使解码器更少依赖于编码器的内部表",
"&及其相应的标签y，编码器处理&产生特征表示",
"示。",
"f。之后，胶囊网络计算在学习过程中推向条件先验",
"4.2. Training",
"p(zy)=Ty的分布q(z）。使用q(za)和所有目标",
"之间的距离信息，我们估计类9，并使用重新参数化",
"我们在封闭数据集上训练模型，在单个输入样本",
"技巧，我们从q(za）采样。给定9和z，我们通过",
"（α，y）的学习过程中，我们最小化以下损失函数：",
"解码器计算重建，该解码器是使用转置卷积的卷积",
"神经网络。在对我们的模型的计算进行一般描述之后，",
"C(, y)= LkL(α,y)+αLcontr(a,y)+βLrec(aα), (1)",
"我们现在一步一步地深入介绍体系结构",
"其中",
"编码阶段。编码阶段涉及的块是编码器和胶囊网",
"LKL(α, y) = d(C, sg Ty)",
"(2)",
"络。编码器是一个卷积神经网络，它处理输人图像",
"aERCxHxW，产生特征ayERdexdhxdw。然后，类似",
"1",
"K",
"Leontr(a,y)=",
" [mk -d(sg[c],Ts)]+, (3)",
"K-1 ",
"于E[24，胶囊网络通过使用动态路由算法计算主胶囊",
"apeERx（dad），然后计算数字胶囊deERKxf2",
"kty",
"来处理特征表示αf。我们用fi表示主要胶囊的尺寸，",
"<NEXLINE"
] | {
"arxiv_id": "2104.09159",
"language": "zh",
"page_idx": 2
} |
||
2104.09159-zh-4 | 2104.09159-zh-4.jpg | <table>
<tbody>
<tr>
<th>Method</th>
<th>MNIST</th>
<th>SVHN</th>
<th>CIFAR10</th>
<th>CIFAR+ 10</th>
<th>CIFAR+ 50</th>
<th>TinyImageNet</th>
</tr>
<tr>
<td>Softmax 十 $28^{\cdot}$</td>
<td>0.978</td>
<td>0.886</td>
<td>0.677</td>
<td>0.816</td>
<td>0.805</td>
<td>0.577</td>
</tr>
<tr>
<td>Openmax 2</td>
<td>0.981</td>
<td>0.894</td>
<td>0.695</td>
<td>0.817</td>
<td>0.796</td>
<td>0.576</td>
</tr>
<tr>
<td>G-Openmax $\lceil5\rceil$</td>
<td>0.984</td>
<td>0.896</td>
<td>0.675</td>
<td>0.827</td>
<td>0.819</td>
<td>0.580</td>
</tr>
<tr>
<td>OSRCI+ [16]</td>
<td>0.988 $\pm0.004$</td>
<td>0.91 $\pm0.01$</td>
<td>0.699 $\pm0.038$</td>
<td>0.838</td>
<td>0.827</td>
<td>0.586</td>
</tr>
<tr>
<td>CROSR $[32^{.}$</td>
<td>0.991 $\pm0.004$</td>
<td>0.899 $\pm0.018$</td>
<td> </td>
<td> </td>
<td> </td>
<td>0.589</td>
</tr>
<tr>
<td>$\pm$ C2AE [19]</td>
<td> </td>
<td>0.892 $\pm0.013$</td>
<td>0.711 $\pm0.008$</td>
<td>0.810 $\pm0.005$</td>
<td>0.803 $\pm0.000$</td>
<td>0.581 $\pm0.019$</td>
</tr>
<tr>
<td>GFROR t $\lceil21\rceil$</td>
<td> </td>
<td>0.955 $\pm0.018$</td>
<td>0.831 $\pm0.039$</td>
<td> </td>
<td> </td>
<td>0.657 $\pm0.012$</td>
</tr>
<tr>
<td>CGDL 8 $[28^{\cdot}$</td>
<td>0.977 $\pm0.008$</td>
<td>0.896 $\pm0.023$</td>
<td>0.681 $\pm0.029$</td>
<td>0.794 $\pm0.013$</td>
<td>0.794 $\pm0.003$</td>
<td>0.653 $\pm0.002$</td>
</tr>
<tr>
<td>TTA $\therefore\lceil3\rceil$ RPL</td>
<td>0.917 $\pm0.006$</td>
<td>0.931 $\pm0.014$</td>
<td>0.784 $\pm0.025$</td>
<td>0.885 $\pm0.019$</td>
<td>0.881 $\pm0.014$</td>
<td>0.711 $\pm0.026$</td>
</tr>
<tr>
<td>CVAECapOSR Ours</td>
<td>0.992 $\pm0.004$</td>
<td>0.956 $\pm0.012$</td>
<td>0.835 $\pm0.023$</td>
<td>0.888 $\pm0.019$</td>
<td>0.889 $\pm0.017$</td>
<td>0.715 $\pm0.018$</td>
</tr>
</tbody>
</table>
Table 1: AUROC scores on the detection of known and unknown samples. Results are averaged over 5 different splits of known and unknown classes partitions. As discussed in Section $\color{red}{5.3}$, we report the results on the same data splits and, for the sake of clarity, we highlight the source of the results used to populate the table: $\dagger$ are provided by [16] , t is from [21] and § are the results that we obtained by running the code of the original paper.
<table>
<tbody>
<tr>
<th>Method</th>
<th>MNIST</th>
<th>SVHN</th>
<th>CIFAR10</th>
<th>CIFAR+ 10</th>
<th>CIFAR+ 50</th>
<th>TinyImageNet</th>
</tr>
<tr>
<td>CVAECapOSR fixed Targets</td>
<td>0.997 $\pm0.006$</td>
<td>0.953 $\pm0.022$</td>
<td>$0.823\pm\pm0.012$</td>
<td>0.868 $\pm0.018$</td>
<td>0.829 $\pm0.009$</td>
<td>0.706 $\pm0.014$</td>
</tr>
<tr>
<td>CVAECapOSR learn Targets</td>
<td>0.992 $\pm0.004$</td>
<td>0.956 $\pm0.012$</td>
<td>$0.835\pm\pm0.023$</td>
<td>0.888 $\pm0.019$</td>
<td>0.889 $\pm0.017$</td>
<td>0.715 $\pm0.018$</td>
</tr>
</tbody>
</table>
Table 2: AUROC scores on the detection of known and unknown samples comparing our model that uses fixed targets (first row) versus our model that learns the targets (second row) during the learning process. Results are averaged over 5 splits.
中采样 4 已知类 ,并从更多样化和更大的 CIFAR100 数据集中随机抽取$Q$未知类。CIFAR+10 和 CIFAR+50的开放性得分分别为 46.54% 和 72.78%
TinyImageNet。对于 TinyImagenet 数据集，它是包含 200 类的 ImageNet 的子集，我们随机抽取了已知的 20 类和未知的其余类。在此设置中，开放性分数为$68.37\%$
## 5.2. Metrics
开放集分类性能通常使用 F 分数和 AUROC (ROC 曲线下面积)$\operatorname{E}[6]$来衡量。F 分数用于衡量分布内分类性能，而 AUROC 通常由开放集识别和分布外检测文献报道。AUROC 提供了一种免校准措施，并通过改变判别阈值[4]来表征给定分数的性能。在我们的实验中，我们在开放集识别任务上使用宏观平均 F1 分数， 在未知检测任务上使用 AUROC。对于这两个指标，值越高越好。
结果是由$\mathbb{E}[16]$ 提供的已知和未知类别的五个随机拆
分的平均值
正如已经讨论过的(最近在[6,21]中显示),不同拆分之间的性能差异很大。因此，我们使用[16]提供的精确数据拆分，这些数据拆分也已在其他近期作品[32,21]中使用。然而，并非这些工作中报告的所有结果都是直接可比的；尽管拆分是相同的，但[21]在CIFAR10 和 CIFAR50 实验中选择开放集类时遵循了特定的策略 (i。e。他们从车辆类别中选择了 10 个和50 个样本，而不是纯粹的随机类别，这对这些结果产生了很大的影响)。因此，在$\mathbb{E}[21]$ 之后，我们运行了E [28] 和$\mathbb{E}[3]$ 的代码 (而$\mathbb{E}[19]$ 的结果由$\mathbb{E}[21]$ 提供，因为该代码不再可用),并且我们将所有结果与具有相同拆分并使用相同评估设置的最新论文进行比较，并且可以重现。如表 1 所示，我们在所有数据集上获得了优于所有以前方法的最新结果。此外，正如之前报道的那样，我们将发布所有数据和代码以保证可重复性
我们在训练过程中观察到的一个重要事实是，我们通过让目标分布$T_k$被学习而不是被用作固定先验而获得的提升，如表$\color{red}{\mathrm{E}}$2 所示。我们用 $\boldsymbol\mu_k^{(i)}=1_d\cdot\delta_{k=i}$ 和$\Sigma_k^{(i)}=1_d$初始化可学习目标。我们注意到，在不考虑损失函数$(\alpha=0$ )中的对比项的情况下学习目标会导致目标崩溃为一个单一分布，从而导致结果不佳。因此，我们考虑对比项，并设置$\alpha=1.0$ 、$\beta=0.05$和$m_k=10$
开放集识别。在开放集识别问题中，在包含$K$类的封闭数据集上训练模型，并在考虑$K+1$类的开放数据
## 5.3. Experimental Results
在 [28] 之后，我们进行了两个主要实验，其中我们的模型必须解决未知检测任务和开放集识别任务。对于所有实验，我们使用 ResNet34$\mathbb{E}[7]$作为我们模型的编码器主干
末知检测。在未知检测问题中，使用$K$类在数据集的子集上训练模型，并通过测量模型检测未知类的能力来完成评估，这在训练期间是看不到的。评估是通过考虑已知类与未知类的二进制识别任务来执行的，并且性能是根据 AUROC 分数报告的。表$\color{red}{\mathrm{E}}\color{red}{\mathrm{1}}$ 中显示的 | [
[
"0.08826",
"0.0903",
"0.1487",
"0.0903",
"0.1487",
"0.1048",
"0.08826",
"0.1048"
],
[
"0.281",
"0.0903",
"0.335",
"0.0903",
"0.335",
"0.1048",
"0.281",
"0.1048"
],
[
"0.3857",
"0.0903",
"0.4355",
"0.0903",
"0.4355",
"0.1048",
"0.3857",
"0.1048"
],
[
"0.4795",
"0.09216",
"0.5483",
"0.09216",
"0.5483",
"0.1042",
"0.4795",
"0.1042"
],
[
"0.5767",
"0.09216",
"0.6587",
"0.09216",
"0.6587",
"0.1042",
"0.5767",
"0.1042"
],
[
"0.6772",
"0.09216",
"0.7607",
"0.09216",
"0.7607",
"0.1042",
"0.6772",
"0.1042"
],
[
"0.777",
"0.0903",
"0.8857",
"0.0903",
"0.8857",
"0.1042",
"0.777",
"0.1042"
],
[
"0.0866",
"0.1042",
"0.1929",
"0.1067",
"0.1929",
"0.1225",
"0.0866",
"0.1193"
],
[
"0.286",
"0.1061",
"0.3308",
"0.1061",
"0.3308",
"0.1206",
"0.286",
"0.1206"
],
[
"0.3896",
"0.1061",
"0.4346",
"0.1061",
"0.4346",
"0.1206",
"0.3896",
"0.1206"
],
[
"0.492",
"0.1061",
"0.535",
"0.1061",
"0.535",
"0.1206",
"0.492",
"0.1206"
],
[
"0.6978",
"0.1061",
"0.74",
"0.1061",
"0.74",
"0.1206",
"0.6978",
"0.1206"
],
[
"0.8096",
"0.1061",
"0.851",
"0.1061",
"0.851",
"0.1206",
"0.8096",
"0.1206"
],
[
"0.286",
"0.1206",
"0.33",
"0.1206",
"0.33",
"0.1351",
"0.286",
"0.1351"
],
[
"0.596",
"0.1067",
"0.636",
"0.1067",
"0.636",
"0.1187",
"0.596",
"0.1187"
],
[
"0.3896",
"0.1206",
"0.433",
"0.1206",
"0.433",
"0.1351",
"0.3896",
"0.1351"
],
[
"0.4927",
"0.1206",
"0.536",
"0.1206",
"0.536",
"0.1351",
"0.4927",
"0.1351"
],
[
"0.5947",
"0.1206",
"0.638",
"0.1206",
"0.638",
"0.1351",
"0.5947",
"0.1351"
],
[
"0.6978",
"0.1206",
"0.7417",
"0.1206",
"0.7417",
"0.1351",
"0.6978",
"0.1351"
],
[
"0.8096",
"0.1206",
"0.8545",
"0.1206",
"0.8545",
"0.1351",
"0.8096",
"0.1351"
],
[
"0.08905",
"0.1351",
"0.2141",
"0.1364",
"0.2141",
"0.1509",
"0.08905",
"0.1497"
],
[
"0.0907",
"0.1218",
"0.1936",
"0.1218",
"0.1936",
"0.1364",
"0.0907",
"0.1364"
],
[
"0.2876",
"0.1364",
"0.33",
"0.1364",
"0.33",
"0.1509",
"0.2876",
"0.1509"
],
[
"0.3896",
"0.1364",
"0.433",
"0.1364",
"0.433",
"0.1509",
"0.3896",
"0.1509"
],
[
"0.4927",
"0.1364",
"0.536",
"0.1364",
"0.536",
"0.1509",
"0.4927",
"0.1509"
],
[
"0.5947",
"0.1364",
"0.636",
"0.1364",
"0.636",
"0.1509",
"0.5947",
"0.1509"
],
[
"0.6978",
"0.1364",
"0.74",
"0.1364",
"0.74",
"0.1509",
"0.6978",
"0.1509"
],
[
"0.8096",
"0.1364",
"0.853",
"0.1364",
"0.853",
"0.1509",
"0.8096",
"0.1509"
],
[
"0.2615",
"0.1497",
"0.3513",
"0.1515",
"0.3513",
"0.1666",
"0.2615",
"0.1642"
],
[
"0.3733",
"0.1497",
"0.4468",
"0.1515",
"0.4468",
"0.1666",
"0.3733",
"0.1642"
],
[
"0.468",
"0.1497",
"0.556",
"0.1515",
"0.556",
"0.1666",
"0.468",
"0.1642"
],
[
"0.5947",
"0.1509",
"0.638",
"0.1509",
"0.638",
"0.1654",
"0.5947",
"0.1654"
],
[
"0.6978",
"0.1509",
"0.74",
"0.1509",
"0.74",
"0.1654",
"0.6978",
"0.1654"
],
[
"0.8096",
"0.1509",
"0.8545",
"0.1509",
"0.8545",
"0.1654",
"0.8096",
"0.1654"
],
[
"0.08905",
"0.1642",
"0.1805",
"0.1666",
"0.1805",
"0.1812",
"0.08905",
"0.1787"
],
[
"0.09235",
"0.1528",
"0.183",
"0.1528",
"0.183",
"0.1642",
"0.09235",
"0.1642"
],
[
"0.3652",
"0.1654",
"0.455",
"0.1674",
"0.455",
"0.1812",
"0.3652",
"0.1787"
],
[
"0.8096",
"0.1666",
"0.853",
"0.1666",
"0.853",
"0.1812",
"0.8096",
"0.1812"
],
[
"0.3652",
"0.1799",
"0.455",
"0.1818",
"0.455",
"0.1963",
"0.3652",
"0.1945"
],
[
"0.2646",
"0.1686",
"0.3513",
"0.1686",
"0.3513",
"0.1799",
"0.2646",
"0.1799"
],
[
"0.4673",
"0.1799",
"0.558",
"0.1818",
"0.558",
"0.1963",
"0.4673",
"0.1945"
],
[
"0.572",
"0.1799",
"0.661",
"0.1818",
"0.661",
"0.1963",
"0.572",
"0.1945"
],
[
"0.675",
"0.1812",
"0.7646",
"0.1831",
"0.7646",
"0.1976",
"0.675",
"0.1957"
],
[
"0.786",
"0.1812",
"0.876",
"0.1831",
"0.876",
"0.1963",
"0.786",
"0.1945"
],
[
"0.3652",
"0.1957",
"0.4558",
"0.1976",
"0.4558",
"0.2122",
"0.3652",
"0.2102"
],
[
"0.09235",
"0.1831",
"0.1772",
"0.1831",
"0.1772",
"0.1945",
"0.09235",
"0.1945"
],
[
"0.0907",
"0.1963",
"0.1953",
"0.1963",
"0.1953",
"0.2108",
"0.0907",
"0.2108"
],
[
"0.468",
"0.1963",
"0.5596",
"0.1963",
"0.5596",
"0.2108",
"0.468",
"0.2108"
],
[
"0.786",
"0.1957",
"0.876",
"0.1976",
"0.876",
"0.2122",
"0.786",
"0.2102"
],
[
"0.08905",
"0.2102",
"0.183",
"0.2122",
"0.183",
"0.2267",
"0.08905",
"0.2247"
],
[
"0.2646",
"0.2102",
"0.353",
"0.2134",
"0.353",
"0.2279",
"0.2646",
"0.2247"
],
[
"0.3652",
"0.2102",
"0.4558",
"0.2122",
"0.4558",
"0.2267",
"0.3652",
"0.2247"
],
[
"0.4673",
"0.2102",
"0.558",
"0.2122",
"0.558",
"0.2267",
"0.4673",
"0.2247"
],
[
"0.572",
"0.2102",
"0.661",
"0.2122",
"0.661",
"0.2267",
"0.572",
"0.2247"
],
[
"0.675",
"0.2102",
"0.7656",
"0.2134",
"0.7656",
"0.2279",
"0.675",
"0.2247"
],
[
"0.786",
"0.2102",
"0.876",
"0.2122",
"0.876",
"0.2267",
"0.786",
"0.2247"
],
[
"0.3652",
"0.226",
"0.4558",
"0.2279",
"0.4558",
"0.2424",
"0.3652",
"0.2405"
],
[
"0.4673",
"0.2247",
"0.558",
"0.2267",
"0.558",
"0.2424",
"0.4673",
"0.2405"
],
[
"0.5703",
"0.226",
"0.66",
"0.2279",
"0.66",
"0.2424",
"0.5703",
"0.2405"
],
[
"0.6724",
"0.226",
"0.763",
"0.2279",
"0.763",
"0.2424",
"0.6724",
"0.2405"
],
[
"0.786",
"0.226",
"0.876",
"0.2279",
"0.876",
"0.2424",
"0.786",
"0.2405"
],
[
"0.2632",
"0.2267",
"0.353",
"0.2267",
"0.353",
"0.2412",
"0.2632",
"0.2412"
],
[
"0.08905",
"0.2267",
"0.1609",
"0.2267",
"0.1609",
"0.2412",
"0.08905",
"0.2412"
],
[
"0.0825",
"0.2393",
"0.3586",
"0.2405",
"0.3586",
"0.2583",
"0.0825",
"0.2568"
],
[
"0.3628",
"0.2412",
"0.4592",
"0.2437",
"0.4592",
"0.2583",
"0.3628",
"0.2556"
],
[
"0.464",
"0.2412",
"0.564",
"0.2437",
"0.564",
"0.2583",
"0.464",
"0.2556"
],
[
"0.568",
"0.2412",
"0.669",
"0.2437",
"0.669",
"0.2583",
"0.568",
"0.2556"
],
[
"0.664",
"0.2412",
"0.7695",
"0.2437",
"0.7695",
"0.2583",
"0.664",
"0.2556"
],
[
"0.786",
"0.2424",
"0.8774",
"0.2424",
"0.8774",
"0.2568",
"0.786",
"0.2568"
],
[
"0.0801",
"0.2695",
"0.892",
"0.2708",
"0.892",
"0.2854",
"0.0801",
"0.2842"
],
[
"0.0801",
"0.286",
"0.892",
"0.286",
"0.892",
"0.2998",
"0.0801",
"0.2998"
],
[
"0.0801",
"0.3018",
"0.892",
"0.3018",
"0.892",
"0.315",
"0.0801",
"0.315"
],
[
"0.0784",
"0.315",
"0.8896",
"0.3164",
"0.8896",
"0.3308",
"0.0784",
"0.33"
],
[
"0.08905",
"0.348",
"0.1487",
"0.348",
"0.1487",
"0.3591",
"0.08905",
"0.3591"
],
[
"0.3162",
"0.348",
"0.3694",
"0.348",
"0.3694",
"0.3591",
"0.3162",
"0.3591"
],
[
"0.4126",
"0.348",
"0.46",
"0.348",
"0.46",
"0.3591",
"0.4126",
"0.3591"
],
[
"0.505",
"0.348",
"0.574",
"0.348",
"0.574",
"0.3591",
"0.505",
"0.3591"
],
[
"0.596",
"0.3467",
"0.886",
"0.3467",
"0.886",
"0.3599",
"0.596",
"0.3599"
],
[
"0.6763",
"0.3591",
"0.7754",
"0.361",
"0.7754",
"0.3757",
"0.6763",
"0.3738"
],
[
"0.7935",
"0.3591",
"0.8774",
"0.3623",
"0.876",
"0.377",
"0.7915",
"0.3738"
],
[
"0.2852",
"0.3599",
"0.3945",
"0.3623",
"0.3945",
"0.3757",
"0.2852",
"0.3738"
],
[
"0.08826",
"0.3599",
"0.2942",
"0.361",
"0.2942",
"0.3757",
"0.08826",
"0.375"
],
[
"0.3865",
"0.3599",
"0.4805",
"0.3623",
"0.4805",
"0.377",
"0.3865",
"0.375"
],
[
"0.4844",
"0.361",
"0.6855",
"0.361",
"0.6855",
"0.3757",
"0.4844",
"0.3757"
],
[
"0.08496",
"0.3738",
"0.789",
"0.3757",
"0.789",
"0.3901",
"0.08496",
"0.3882"
],
[
"0.7793",
"0.3738",
"0.8784",
"0.3757",
"0.8784",
"0.3901",
"0.7793",
"0.3882"
],
[
"0.0809",
"0.404",
"0.8896",
"0.404",
"0.8896",
"0.4172",
"0.0809",
"0.4172"
],
[
"0.0801",
"0.4185",
"0.8906",
"0.4185",
"0.8906",
"0.432",
"0.0801",
"0.432"
],
[
"0.0801",
"0.4343",
"0.2402",
"0.4343",
"0.2402",
"0.4475",
"0.0801",
"0.4475"
],
[
"0.0825",
"0.476",
"0.468",
"0.476",
"0.468",
"0.4893",
"0.0825",
"0.4893"
],
[
"0.5024",
"0.476",
"0.8906",
"0.476",
"0.8906",
"0.4893",
"0.5024",
"0.4893"
],
[
"0.0809",
"0.4912",
"0.468",
"0.4912",
"0.468",
"0.505",
"0.0809",
"0.505"
],
[
"0.5034",
"0.4905",
"0.5884",
"0.4905",
"0.5884",
"0.505",
"0.5034",
"0.505"
],
[
"0.0809",
"0.506",
"0.3628",
"0.506",
"0.3628",
"0.5195",
"0.0809",
"0.5195"
],
[
"0.524",
"0.5103",
"0.892",
"0.5103",
"0.892",
"0.5234",
"0.524",
"0.5234"
],
[
"0.0989",
"0.521",
"0.468",
"0.521",
"0.468",
"0.535",
"0.0989",
"0.535"
],
[
"0.505",
"0.526",
"0.892",
"0.526",
"0.892",
"0.539",
"0.505",
"0.539"
],
[
"0.0825",
"0.536",
"0.4673",
"0.536",
"0.4673",
"0.55",
"0.0825",
"0.55"
],
[
"0.5034",
"0.5405",
"0.8906",
"0.5405",
"0.8906",
"0.5537",
"0.5034",
"0.5537"
],
[
"0.0784",
"0.55",
"0.468",
"0.5503",
"0.468",
"0.565",
"0.0784",
"0.5645"
],
[
"0.5034",
"0.556",
"0.8906",
"0.556",
"0.8906",
"0.5693",
"0.5034",
"0.5693"
],
[
"0.0809",
"0.5674",
"0.1438",
"0.5674",
"0.1438",
"0.579",
"0.0809",
"0.579"
],
[
"0.5034",
"0.5713",
"0.892",
"0.5713",
"0.892",
"0.5854",
"0.5034",
"0.5854"
],
[
"0.0784",
"0.5884",
"0.179",
"0.5903",
"0.179",
"0.605",
"0.0784",
"0.603"
],
[
"0.5034",
"0.586",
"0.8896",
"0.586",
"0.8896",
"0.5996",
"0.5034",
"0.5996"
],
[
"0.5034",
"0.6016",
"0.8906",
"0.6016",
"0.8906",
"0.6157",
"0.5034",
"0.6157"
],
[
"0.10046",
"0.6123",
"0.4673",
"0.6123",
"0.4673",
"0.6255",
"0.10046",
"0.6255"
],
[
"0.5034",
"0.616",
"0.8896",
"0.616",
"0.8896",
"0.63",
"0.5034",
"0.63"
],
[
"0.0825",
"0.627",
"0.4656",
"0.627",
"0.4656",
"0.6416",
"0.0825",
"0.6416"
],
[
"0.5024",
"0.631",
"0.8896",
"0.631",
"0.8896",
"0.6445",
"0.5024",
"0.6445"
],
[
"0.0809",
"0.643",
"0.4673",
"0.643",
"0.4673",
"0.657",
"0.0809",
"0.657"
],
[
"0.505",
"0.6465",
"0.8906",
"0.6465",
"0.8906",
"0.661",
"0.505",
"0.661"
],
[
"0.0825",
"0.659",
"0.468",
"0.659",
"0.468",
"0.6724",
"0.0825",
"0.6724"
],
[
"0.505",
"0.661",
"0.8906",
"0.661",
"0.8906",
"0.675",
"0.505",
"0.675"
],
[
"0.0801",
"0.6724",
"0.468",
"0.674",
"0.468",
"0.688",
"0.0801",
"0.6875"
],
[
"0.5024",
"0.675",
"0.8906",
"0.6753",
"0.8906",
"0.69",
"0.5024",
"0.6895"
],
[
"0.0809",
"0.6895",
"0.468",
"0.6895",
"0.468",
"0.7026",
"0.0809",
"0.7026"
],
[
"0.5024",
"0.69",
"0.8896",
"0.6914",
"0.8896",
"0.7056",
"0.5024",
"0.705"
],
[
"0.0801",
"0.705",
"0.468",
"0.705",
"0.468",
"0.7183",
"0.0801",
"0.7183"
],
[
"0.505",
"0.707",
"0.8906",
"0.707",
"0.8906",
"0.72",
"0.505",
"0.72"
],
[
"0.0801",
"0.7183",
"0.1544",
"0.7183",
"0.1544",
"0.733",
"0.0801",
"0.733"
],
[
"0.5034",
"0.7227",
"0.864",
"0.7227",
"0.864",
"0.7363",
"0.5034",
"0.7363"
],
[
"0.0784",
"0.741",
"0.2852",
"0.742",
"0.2852",
"0.757",
"0.0784",
"0.756"
],
[
"0.524",
"0.741",
"0.8906",
"0.741",
"0.8906",
"0.755",
"0.524",
"0.755"
],
[
"0.505",
"0.757",
"0.892",
"0.757",
"0.892",
"0.771",
"0.505",
"0.771"
],
[
"0.10046",
"0.7666",
"0.468",
"0.7666",
"0.468",
"0.78",
"0.10046",
"0.78"
],
[
"0.5034",
"0.7734",
"0.8906",
"0.7734",
"0.8906",
"0.7886",
"0.5034",
"0.7886"
],
[
"0.0809",
"0.781",
"0.468",
"0.781",
"0.468",
"0.7944",
"0.0809",
"0.7944"
],
[
"0.5024",
"0.789",
"0.8906",
"0.7905",
"0.8906",
"0.8047",
"0.5024",
"0.8037"
],
[
"0.0825",
"0.797",
"0.468",
"0.797",
"0.468",
"0.81",
"0.0825",
"0.81"
],
[
"0.5024",
"0.8047",
"0.8906",
"0.806",
"0.8906",
"0.821",
"0.5024",
"0.8193"
],
[
"0.0809",
"0.8125",
"0.1626",
"0.8125",
"0.1626",
"0.8237",
"0.0809",
"0.8237"
],
[
"0.10046",
"0.8257",
"0.468",
"0.8257",
"0.468",
"0.839",
"0.10046",
"0.839"
],
[
"0.5034",
"0.8213",
"0.8906",
"0.8213",
"0.8906",
"0.836",
"0.5034",
"0.836"
],
[
"0.0809",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0809",
"0.855"
],
[
"0.5034",
"0.837",
"0.892",
"0.837",
"0.892",
"0.851",
"0.5034",
"0.851"
],
[
"0.0825",
"0.857",
"0.468",
"0.857",
"0.468",
"0.8706",
"0.0825",
"0.8706"
],
[
"0.501",
"0.853",
"0.568",
"0.851",
"0.568",
"0.8657",
"0.501",
"0.867"
],
[
"0.0809",
"0.872",
"0.468",
"0.872",
"0.468",
"0.8853",
"0.0809",
"0.8853"
],
[
"0.523",
"0.872",
"0.8906",
"0.872",
"0.8906",
"0.8853",
"0.523",
"0.8853"
],
[
"0.0825",
"0.887",
"0.468",
"0.887",
"0.468",
"0.901",
"0.0825",
"0.901"
],
[
"0.5034",
"0.887",
"0.8906",
"0.887",
"0.8906",
"0.901",
"0.5034",
"0.901"
]
] | [
"Method",
"MNIST",
"SVHN",
"CIFAR10",
"CIFAR+10",
"CIFAR+50",
"TinylmageNet",
"Softmaxt [28]",
"0.978",
"0.886",
"0.677",
"0.805",
"0.577",
"0.981",
"0.816",
"0.894",
"0.695",
"0.817",
"0.796",
"0.576",
"G-Openmax t [5]",
"Openmax + [2]",
"0.984",
"0.896",
"0.675",
"0.827",
"0.819",
"0.580",
"0.988 ±0.004",
"0.91 ±0.01",
"0.699 ±0.038",
"0.838",
"0.827",
"0.586",
"CROSR [32]",
"OSRCI +J16",
"0.899±0.018",
"0.589",
"0.892 ±0.013",
"0.991 ±0.004",
"0.711 ±0.008",
"0.810 ±0.005",
"0.803 ±0.000",
"0.581±0.019",
"0.955 ±0.018",
"C2AE +[19]",
"GFROR ↓[21]",
"0.831 ±0.039",
"0.657±0.012",
"CGDL S [28]",
"0.977 ±0.008",
"0.896 ±0.023",
"0.681±0.029",
"0.794 ±0.013",
"0.794 ±0.003",
"0.653±0.002",
"0.931±0.014",
"0.784 ±0.025",
"0.885±0.019",
"0.881±0.014",
"0.711 ±0.026",
"0.917±0.006",
"RPL S[3]",
"CVAECapOSR (ours)0.992±0.004",
"0.956±0.012",
"0.835±0.023",
"0.888±0.019",
"0.889±0.017",
"0.715 ±0.018",
"Table 1: AUROC scores on the detection of known and unknown samples. Results are averaged over 5 different",
"splits of known and unknown classes partitions. As discussed in Section 5.3 , we report the results on the same",
"data splits and, for the sake of clarity, we highlight the source of the results used to populate the table: f are",
"provided by [16] , is from [2] and g are the results that we obtained by running the code of the original paper.",
"Method",
"MNIST",
"SVHN",
"CIFAR10",
"CIFAR+10CIFAR+50TinylmageNet",
"0.829±0.009",
"0.706 ±0.014",
"0.997±0.006",
"CVAECapOSRfixedTargets",
"0.953±0.022",
"0.823±±0.0120.868±0.018",
"CVAECapOSRlearnTargets0.992±0.0040.956±0.0120.835±±0.0230.888±0.0190.889±0.017",
"0.715±0.018",
"Table 2: AUROC scores on the detection of known and unknown samples comparing our model that uses fixed",
"targets (first row) versus our model that learns the targets (second row) during the learning process. Results are",
"averaged over 5 splits.",
"中采样4已知类，并从更多样化和更大的CIFAR100数",
"结果是由E161提供的已知和未知类别的五个随机拆",
"据集中随机抽取Q未知类。CIFAR+10和CIFAR+50",
"分的平均值",
"的开放性得分分别为46.54%和72.78%",
"正如已经讨论过的（最近在[6，21]中显示），不",
"TinyImageNet。对于TinyImagenet数据集，它是包",
"同拆分之间的性能差异很大。因此，我们使用「16]提",
"含200类的ImageNet的子集，我们随机抽取了已知",
"供的精确数据拆分，这些数据拆分也已在其他近期作",
"的20类和未知的其余类。在此设置中，开放性分数为",
"品【32，211中使用。然而，并非这些工作中报告的所有",
"68.37%。",
"结果都是直接可比的；尽管拆分是相同的，但[21]在",
"5.2. Metrics",
"CIFAR10和CIFAR50实验中选择开放集类时遵循了",
"特定的策略（i。e。他们从车辆类别中选择了10个和",
"开放集分类性能通常使用F分数和AUROC（ROC",
"50个样本，而不是纯粹的随机类别，这对这些结果产",
"曲线下面积E[6]来衡量。F分数用于衡量分布内分",
"生了很大的影响)。因此，在E[21]之后，我们运行了E",
"类性能，而AUROC通常由开放集识别和分布外检测",
"[28]和E[3]的代码（而E[19]的结果由E[21]提供，因",
"文献报道。AUROC提供了一种免校准措施，并通过改",
"为该代码不再可用），并且我们将所有结果与具有相同",
"变判别阈值[】来表征给定分数的性能。在我们的实验",
"拆分并使用相同评估设置的最新论文进行比较，并且",
"中，我们在开放集识别任务上使用宏观平均F1分数，",
"可以重现。如表1所示，我们在所有数据集上获得了",
"在未知检测任务上使用AUROC。对于这两个指标，值",
"优于所有以前方法的最新结果。此外，正如之前报道的",
"越高越好。",
"那样，我们将发布所有数据和代码以保证可重复性",
"5.3.ExperimentalResults",
"我们在训练过程中观察到的一个重要事实是，我们",
"通过让目标分布T被学习而不是被用作固定先验而获",
"在28]之后，我们进行了两个主要实验，其中我们",
"得的提升，如表E2所示。我们用μ）=1a·0k=和",
"的模型必须解决未知检测任务和开放集识别任务。对",
"）=1。初始化可学习目标。我们注意到，在不考虑",
"于所有实验，我们使用ResNet34E[7]作为我们模型的",
"损失函数（α=0）中的对比项的情况下学习目标会导",
"编码器主干",
"未知检测。在未知检测问题中，使用K类在数据集",
"致目标崩溃为一个单一分布，从而导致结果不佳。因",
"的子集上训练模型，并通过测量模型检测未知类的能",
"此，我们考虑对比项，并设置α=1.0、β=0.05和",
"力来完成评估，这在训练期间是看不到的。评估是通过",
"mk = 10",
"考虑已知类与未知类的二进制识别任务来执行的，并",
"开放集识别。在开放集识别问题中，在包含K类的",
"且性能是根据AUROC分数报告的。表E1中显示的",
"封闭数据集上训练模型，并在考虑K+1类的开放数据"
] | {
"arxiv_id": "2104.09159",
"language": "zh",
"page_idx": 4
} |
||
2311.03094-zh-1 | 2311.03094-zh-1.jpg | Title Suppressed Due to Excessive Size
$2.{\mathrm{~Benefits~of~Equivariant~Models}}$
群等变网络是 ML 研究的热门和活跃领域。这项工作通常是由于具有某些等变属性的模型的成功，例如具有近似平移对称性的自然图像域上的卷积神经网络(CNN)中的平移对称性，以及物理世界固有的大量对称性(因此，来自天文学，物理学和化学的大量数据集)。研究人员建议，在某些情况下，证明了各种潜在的好处，包括：
Accuracy 第 1 节中介绍的等变模型以及许多其他建议的等变体系结构在对其进行评估以供发布的选定基准数据集上实现了最新技术(SotA)的准确性。这是有道理的，因为 SotA 的表现通常是在 ML 场所成功发布的未明确标准
实际上，模型准确性的提高通常被认为是研究等变体系结构的主要动机。CNN 和 GNN 的性能成功通常归因于它们各自固有的翻译和置换等方差。对组等变网络重复此叙述，以提高模型性能的期望；例如，请参见 (Satorras et al., 2021) 和 (Villar et al., 2021) 的介绍部分
Generalizability 从理论上讲，一组等变网络应该能够学习完整的轨道(i。e。由对称群的作用形成的流形) 来自单个样本。然后，该网络将能够有效地对来自同一轨道的任何其他输入进行建模。这种现象称为泛化我们将模型的泛化能力称为泛化能力。据推测，通过实施等方差，网络能够学习数据域的精确，简化的表示，这有助于泛化
对等变网络的一些研究，虽然不是全部，但包括证明模型在组变换下泛化能力的测试 (e。g。(Gong et al., 2022) 中的图 3)。此外，一些研究专门承担了泛化的任务。。例如，(Wang et al., 2020a) 的作者证明了他们的 E (2) -等变 CNN 在平移，旋转和线性缩放方面的鲁棒性比在非等变模型下更好地泛化(尽管它仍然不能泛化整个测试数据域)
Model Efficiency 先前的工作表明，随着模型中内置数据的对称性，等变网络有一个“更容易”的时间学习最佳函数以最小化训练损失。具体来说，已经提出， 对于给定数量的可学习参数，等变模型可能比非等变模型更有效地使用这些参数。在实践中，(Cohen & Welling, 2017) 的作者证明了这种模型的效率，他们表明，通过基本特征的组合创建表示的“可控 CNN”,每个特征都与特定类型的对称性相关联，能够更有效地利用模型参数为 CIFAR 基准生成较小的 SotA 图像分类器
模型效率是非常可取的，特别是在资源受限的计算环境中。因此，一些作者，例如 (Bogatskiy et al.,2020) ,将其等变模型的大小与类似执行的非等变模型相比$\overline{{\text{equivariance in }}}$particle physics, and the conclusions may not apply directly to systems with discrete symmetries or
significantly different metrics, as in chemistry and biology.
的减小作为主要优点，即使它们可能无法达到完全的
SotA 精度
Data Effciency 训练无约束模型以尊重对称性的“传统”方法是通过数据增强，这通常会大大增加所需训练数据集的大小，因此会增加训练模型所需的计算资源。可以预期的是，由于等变网络被限制为仅学习服从对称的函数，因此它们将需要更少的训练示例来实现类似的性能；换句话说，他们将证明提高了数据效率
(Satorras et al., 2021) 的作者通过将模型性能作为训练数据集中样本数量的函数与非等变模型进行比较， 证明了其 E (3) -等变 GNN 的数据效率。等变 GNN 在小数据状态下的表现明显优于无约束模型；有趣的是，它在大数据状态下的表现也优于另一个等变模型。此外，(Bietti et al., 2021) 的作者提供了理论证据，证明等变模型比无约束模型具有更好的数据效率，至少在通过使用不变核强制等方差的情况下
我们还注意到，模型的可解释性或可解释性通常被认为是构建等变网络 (Bogatskiy et al., 2022a) 的好处。然而，由于没有明确的指标来评估模型的可解释性或可解释性能力，特别是对于本文研究的物理数据集，我们在本文中不评估这一说法
# 3. Symmetries and Machine Learning for Physics
尽管已针对模型类型，对称性和任务的某些组合证明了 2 节中描述的等方差的好处，但它们都没有得到强有力的一般理论保证或所有实验证据的支持。为了更好地表征在 ML 模型中实施等方差的可能好处，我们研究了等变 GNN 在两个重要的粒子物理数据重建任务上的性能
# 3.1. Top Jet Tagging
在高能粒子物理实验中，例如大型强子对撞机(LHC) 的 ATLAS 和 CMS 探测器，射流在粒子相互作用的研究中起着举足轻重的作用。射流是在质子-质子碰撞过程中产生高能夸克或胶子时形成的粒子的准直喷雾。由于强力的性质以及夸克和胶子的限制，这些粒子迅速碎裂并强子化，导致可检测的粒子聚集在一个狭窄的圆锥体内(图 1 )。识别原始夸克或胶子的类型 (“喷射标记”) 对于实现精确测量和寻找新的物理过程至关重要
洛伦兹对称性是物理学中的基本原理，源于狭义相对论。它描述了时空旋转和提升下物理定律的不变性。洛伦兹组是所有矩阵的集合，这些矩阵保留了时空向量$u,v\in\mathbb{R}^4$和$J=diag(1,-1,-1,1)$的双线性形式$\overline {\eta }( u, v) = u^{T}Jv$ ,即 Minkowski 度量。Lorentz 对称， 惯性框架被限制为正定向和正时间定向，由适当的正交 Lorentz 组 SO$(1,3)^+$描述。在洛伦兹变换下，顶部射流标记的任务是不变的，因为原始夸克或胶子类型
2 | [
[
"0.3586",
"0.05807",
"0.615",
"0.05807",
"0.615",
"0.07135",
"0.3586",
"0.07135"
],
[
"0.08826",
"0.0859",
"0.3757",
"0.0859",
"0.3757",
"0.09973",
"0.08826",
"0.09973"
],
[
"0.5024",
"0.0871",
"0.884",
"0.0871",
"0.884",
"0.101",
"0.5024",
"0.101"
],
[
"0.5024",
"0.1029",
"0.576",
"0.1029",
"0.576",
"0.11426",
"0.5024",
"0.11426"
],
[
"0.08905",
"0.1111",
"0.4739",
"0.1111",
"0.4739",
"0.125",
"0.08905",
"0.125"
],
[
"0.0907",
"0.127",
"0.4739",
"0.127",
"0.4739",
"0.1401",
"0.0907",
"0.1401"
],
[
"0.501",
"0.1332",
"0.8823",
"0.1332",
"0.8823",
"0.1465",
"0.501",
"0.1465"
],
[
"0.09235",
"0.1433",
"0.4722",
"0.1433",
"0.4722",
"0.154",
"0.09235",
"0.154"
],
[
"0.501",
"0.1477",
"0.884",
"0.1477",
"0.884",
"0.161",
"0.501",
"0.161"
],
[
"0.0907",
"0.1572",
"0.4739",
"0.1572",
"0.4739",
"0.1704",
"0.0907",
"0.1704"
],
[
"0.501",
"0.1622",
"0.8857",
"0.1622",
"0.8857",
"0.1755",
"0.501",
"0.1755"
],
[
"0.0907",
"0.1718",
"0.4739",
"0.1718",
"0.4739",
"0.1849",
"0.0907",
"0.1849"
],
[
"0.5024",
"0.1787",
"0.884",
"0.1787",
"0.884",
"0.1895",
"0.5024",
"0.1895"
],
[
"0.0907",
"0.1863",
"0.4739",
"0.1863",
"0.4739",
"0.1995",
"0.0907",
"0.1995"
],
[
"0.5034",
"0.1945",
"0.884",
"0.1945",
"0.884",
"0.2052",
"0.5034",
"0.2052"
],
[
"0.09235",
"0.2026",
"0.1936",
"0.2026",
"0.1936",
"0.214",
"0.09235",
"0.214"
],
[
"0.5024",
"0.2084",
"0.8857",
"0.2084",
"0.8857",
"0.2216",
"0.5024",
"0.2216"
],
[
"0.501",
"0.2235",
"0.5215",
"0.2235",
"0.5215",
"0.2361",
"0.501",
"0.2361"
],
[
"0.0907",
"0.2329",
"0.4739",
"0.2329",
"0.4739",
"0.2468",
"0.0907",
"0.2468"
],
[
"0.0907",
"0.2474",
"0.4739",
"0.2474",
"0.4739",
"0.2615",
"0.0907",
"0.2615"
],
[
"0.501",
"0.2456",
"0.884",
"0.2456",
"0.884",
"0.2588",
"0.501",
"0.2588"
],
[
"0.09235",
"0.2644",
"0.4722",
"0.2644",
"0.4722",
"0.2747",
"0.09235",
"0.2747"
],
[
"0.5",
"0.26",
"0.8813",
"0.2615",
"0.8813",
"0.2747",
"0.5",
"0.2734"
],
[
"0.09235",
"0.2778",
"0.4722",
"0.2778",
"0.4722",
"0.2917",
"0.09235",
"0.2917"
],
[
"0.5024",
"0.276",
"0.884",
"0.276",
"0.884",
"0.289",
"0.5024",
"0.289"
],
[
"0.0907",
"0.2935",
"0.2214",
"0.2935",
"0.2214",
"0.307",
"0.0907",
"0.307"
],
[
"0.5024",
"0.2917",
"0.884",
"0.2917",
"0.884",
"0.305",
"0.5024",
"0.305"
],
[
"0.5024",
"0.3062",
"0.884",
"0.3062",
"0.884",
"0.3193",
"0.5024",
"0.3193"
],
[
"0.09235",
"0.3176",
"0.4722",
"0.3176",
"0.4722",
"0.3276",
"0.09235",
"0.3276"
],
[
"0.5024",
"0.3208",
"0.8857",
"0.3208",
"0.8857",
"0.334",
"0.5024",
"0.334"
],
[
"0.09235",
"0.33",
"0.4739",
"0.33",
"0.4739",
"0.3435",
"0.09235",
"0.3435"
],
[
"0.5024",
"0.336",
"0.884",
"0.336",
"0.884",
"0.3499",
"0.5024",
"0.3499"
],
[
"0.09235",
"0.3452",
"0.4722",
"0.3452",
"0.4722",
"0.3591",
"0.09235",
"0.3591"
],
[
"0.501",
"0.3516",
"0.7754",
"0.3516",
"0.7754",
"0.3623",
"0.501",
"0.3623"
],
[
"0.09235",
"0.3623",
"0.4714",
"0.3623",
"0.4714",
"0.3726",
"0.09235",
"0.3726"
],
[
"0.0907",
"0.3757",
"0.4739",
"0.3757",
"0.4739",
"0.3894",
"0.0907",
"0.3894"
],
[
"0.5024",
"0.3757",
"0.8813",
"0.3757",
"0.8813",
"0.3865",
"0.5024",
"0.3865"
],
[
"0.0907",
"0.3926",
"0.1381",
"0.3926",
"0.1381",
"0.404",
"0.0907",
"0.404"
],
[
"0.501",
"0.3882",
"0.8823",
"0.3882",
"0.8823",
"0.4016",
"0.501",
"0.4016"
],
[
"0.5024",
"0.404",
"0.884",
"0.404",
"0.884",
"0.4172",
"0.5024",
"0.4172"
],
[
"0.0907",
"0.4216",
"0.4722",
"0.4216",
"0.4722",
"0.435",
"0.0907",
"0.435"
],
[
"0.5024",
"0.4185",
"0.8823",
"0.4192",
"0.8823",
"0.433",
"0.5024",
"0.432"
],
[
"0.0907",
"0.4375",
"0.4739",
"0.4375",
"0.4739",
"0.4507",
"0.0907",
"0.4507"
],
[
"0.501",
"0.435",
"0.6978",
"0.435",
"0.6978",
"0.449",
"0.501",
"0.449"
],
[
"0.0907",
"0.452",
"0.4722",
"0.452",
"0.4722",
"0.4653",
"0.0907",
"0.4653"
],
[
"0.0907",
"0.4666",
"0.4739",
"0.4685",
"0.4739",
"0.4817",
"0.0907",
"0.4797"
],
[
"0.4968",
"0.4702",
"0.846",
"0.4702",
"0.846",
"0.4841",
"0.4968",
"0.4841"
],
[
"0.0907",
"0.4817",
"0.4722",
"0.4817",
"0.4722",
"0.4956",
"0.0907",
"0.4956"
],
[
"0.5215",
"0.486",
"0.5864",
"0.488",
"0.5864",
"0.5024",
"0.5215",
"0.5005"
],
[
"0.0907",
"0.4976",
"0.4739",
"0.4976",
"0.4739",
"0.511",
"0.0907",
"0.511"
],
[
"0.0907",
"0.513",
"0.2214",
"0.513",
"0.2214",
"0.5264",
"0.0907",
"0.5264"
],
[
"0.5024",
"0.512",
"0.884",
"0.512",
"0.884",
"0.526",
"0.5024",
"0.526"
],
[
"0.505",
"0.529",
"0.8823",
"0.529",
"0.8823",
"0.539",
"0.505",
"0.539"
],
[
"0.0907",
"0.535",
"0.4722",
"0.535",
"0.4722",
"0.549",
"0.0907",
"0.549"
],
[
"0.5034",
"0.5444",
"0.8823",
"0.5444",
"0.8823",
"0.5547",
"0.5034",
"0.5547"
],
[
"0.0907",
"0.5503",
"0.4722",
"0.5503",
"0.4722",
"0.5645",
"0.0907",
"0.5645"
],
[
"0.5024",
"0.558",
"0.884",
"0.558",
"0.884",
"0.5713",
"0.5024",
"0.5713"
],
[
"0.0907",
"0.565",
"0.4722",
"0.565",
"0.4722",
"0.579",
"0.0907",
"0.579"
],
[
"0.501",
"0.5728",
"0.8857",
"0.5728",
"0.8857",
"0.586",
"0.501",
"0.586"
],
[
"0.0907",
"0.5806",
"0.4722",
"0.5806",
"0.4722",
"0.5947",
"0.0907",
"0.5947"
],
[
"0.5024",
"0.589",
"0.584",
"0.589",
"0.584",
"0.601",
"0.5024",
"0.601"
],
[
"0.0907",
"0.595",
"0.4722",
"0.595",
"0.4722",
"0.6094",
"0.0907",
"0.6094"
],
[
"0.0907",
"0.61",
"0.4722",
"0.61",
"0.4722",
"0.6235",
"0.0907",
"0.6235"
],
[
"0.4983",
"0.6187",
"0.6567",
"0.6206",
"0.6567",
"0.634",
"0.4983",
"0.632"
],
[
"0.0907",
"0.6274",
"0.259",
"0.6274",
"0.259",
"0.638",
"0.0907",
"0.638"
],
[
"0.5024",
"0.643",
"0.8857",
"0.643",
"0.8857",
"0.657",
"0.5024",
"0.657"
],
[
"0.0907",
"0.655",
"0.4722",
"0.655",
"0.4722",
"0.6685",
"0.0907",
"0.6685"
],
[
"0.5024",
"0.659",
"0.884",
"0.659",
"0.884",
"0.6724",
"0.5024",
"0.6724"
],
[
"0.0907",
"0.672",
"0.4756",
"0.672",
"0.4756",
"0.685",
"0.0907",
"0.685"
],
[
"0.501",
"0.675",
"0.884",
"0.675",
"0.884",
"0.685",
"0.501",
"0.685"
],
[
"0.09235",
"0.686",
"0.4739",
"0.686",
"0.4739",
"0.6997",
"0.09235",
"0.6997"
],
[
"0.5",
"0.6875",
"0.8857",
"0.688",
"0.8857",
"0.7026",
"0.5",
"0.702"
],
[
"0.0907",
"0.702",
"0.4739",
"0.702",
"0.4739",
"0.7153",
"0.0907",
"0.7153"
],
[
"0.5034",
"0.705",
"0.8823",
"0.705",
"0.8823",
"0.7153",
"0.5034",
"0.7153"
],
[
"0.09235",
"0.7183",
"0.4722",
"0.7183",
"0.4722",
"0.729",
"0.09235",
"0.729"
],
[
"0.501",
"0.7173",
"0.8823",
"0.7183",
"0.8823",
"0.732",
"0.501",
"0.731"
],
[
"0.08905",
"0.731",
"0.4722",
"0.731",
"0.4722",
"0.744",
"0.08905",
"0.744"
],
[
"0.5024",
"0.7344",
"0.8857",
"0.7344",
"0.8857",
"0.7476",
"0.5024",
"0.7476"
],
[
"0.09235",
"0.747",
"0.4739",
"0.747",
"0.4739",
"0.7603",
"0.09235",
"0.7603"
],
[
"0.5024",
"0.7505",
"0.8823",
"0.7505",
"0.8823",
"0.761",
"0.5024",
"0.761"
],
[
"0.0907",
"0.762",
"0.4739",
"0.762",
"0.4739",
"0.776",
"0.0907",
"0.776"
],
[
"0.501",
"0.765",
"0.535",
"0.765",
"0.535",
"0.778",
"0.501",
"0.778"
],
[
"0.09235",
"0.778",
"0.4722",
"0.778",
"0.4722",
"0.7915",
"0.09235",
"0.7915"
],
[
"0.0907",
"0.7925",
"0.12335",
"0.7944",
"0.1217",
"0.8066",
"0.08905",
"0.8047"
],
[
"0.5024",
"0.787",
"0.8857",
"0.787",
"0.8857",
"0.8003",
"0.5024",
"0.8003"
],
[
"0.5034",
"0.8037",
"0.8813",
"0.8037",
"0.8813",
"0.8145",
"0.5034",
"0.8145"
],
[
"0.08905",
"0.8145",
"0.4739",
"0.8145",
"0.4739",
"0.8276",
"0.08905",
"0.8276"
],
[
"0.501",
"0.8174",
"0.8857",
"0.8174",
"0.8857",
"0.8306",
"0.501",
"0.8306"
],
[
"0.08905",
"0.83",
"0.4722",
"0.83",
"0.4722",
"0.8433",
"0.08905",
"0.8433"
],
[
"0.501",
"0.832",
"0.8857",
"0.832",
"0.8857",
"0.845",
"0.501",
"0.845"
],
[
"0.09235",
"0.8447",
"0.4739",
"0.8447",
"0.4739",
"0.858",
"0.09235",
"0.858"
],
[
"0.501",
"0.845",
"0.8823",
"0.845",
"0.8823",
"0.86",
"0.501",
"0.86"
],
[
"0.0907",
"0.867",
"0.4722",
"0.867",
"0.4722",
"0.881",
"0.0907",
"0.881"
],
[
"0.5024",
"0.8623",
"0.884",
"0.8623",
"0.884",
"0.8755",
"0.5024",
"0.8755"
],
[
"0.0907",
"0.88",
"0.4739",
"0.88",
"0.4739",
"0.893",
"0.0907",
"0.893"
],
[
"0.5024",
"0.879",
"0.884",
"0.879",
"0.884",
"0.8896",
"0.5024",
"0.8896"
],
[
"0.0866",
"0.8926",
"0.4739",
"0.893",
"0.4739",
"0.907",
"0.0866",
"0.906"
],
[
"0.501",
"0.8926",
"0.8857",
"0.8926",
"0.8857",
"0.906",
"0.501",
"0.906"
],
[
"0.482",
"0.925",
"0.4927",
"0.925",
"0.4927",
"0.935",
"0.482",
"0.935"
]
] | [
"Title Suppressed Due to Excessive Size",
"2. Benefits of Equivariant Models",
"的减小作为主要优点，即使它们可能无法达到完全的",
"SotA精度",
"群等变网络是ML研究的热门和活跃领域。这项工",
"作通常是由于具有某些等变属性的模型的成功，例如",
"DataEficiency训练无约束模型以尊重对称性的“传",
"具有近似平移对称性的自然图像域上的卷积神经网络",
"统”方法是通过数据增强，这通常会大大增加所需训",
"（CNN）中的平移对称性，以及物理世界固有的大量对",
"练数据集的大小，因此会增加训练模型所需的计算资",
"称性（因此，来自天文学，物理学和化学的大量数据",
"源。可以预期的是，由于等变网络被限制为仅学习服",
"集）。研究人员建议，在某些情况下，证明了各种潜在",
"从对称的函数，因此它们将需要更少的训练示例来实",
"的好处，包括：",
"现类似的性能；换句话说，他们将证明提高了数据效",
"率",
"Accuracy第1节中介绍的等变模型以及许多其他建",
"议的等变体系结构在对其进行评估以供发布的选定基",
"(Satorrasetal.，2021）的作者通过将模型性能作为训",
"准数据集上实现了最新技术（SotA）的准确性。这是",
"练数据集中样本数量的函数与非等变模型进行比较",
"有道理的，因为SotA的表现通常是在ML场所成功",
"证明了其E（3）-等变 GNN的数据效率。等变 GNN",
"发布的未明确标准",
"在小数据状态下的表现明显优于无约束模型；有趣的",
"是，它在大数据状态下的表现也优于另一个等变模型。",
"实际上，模型准确性的提高通常被认为是研究等变体",
"此外，（Biettietal.，2021）的作者提供了理论证据，证",
"系结构的主要动机。CNN和GNN的性能成功通常归",
"明等变模型比无约束模型具有更好的数据效率，至少",
"因于它们各自固有的翻译和置换等方差。对组等变网",
"在通过使用不变核强制等方差的情况下",
"络重复此叙述，以提高模型性能的期望；例如，请参",
"见(Satorraset al.,2021）和（Villar etal.,2021）的介",
"我们还注意到，模型的可解释性或可解释性通常被认",
"绍部分",
"为是构建等变网络（Bogatskiyetal.，2022a）的好处",
"然而，由于没有明确的指标来评估模型的可解释性或",
"Generalizability从理论上讲，一组等变网络应该能",
"可解释性能力，特别是对于本文研究的物理数据集，我",
"够学习完整的轨道(i。e。由对称群的作用形成的流形)",
"们在本文中不评估这一说法",
"来自单个样本。然后，该网络将能够有效地对来自同",
"一轨道的任何其他输入进行建模。这种现象称为泛化，",
"3. Symmetries and Machine Learning foi",
"我们将模型的泛化能力称为泛化能力。据推测，通过",
"Physics",
"实施等方差，网络能够学习数据域的精确，简化的表",
"示，这有助于泛化",
"尽管已针对模型类型，对称性和任务的某些组合证明",
"了2节中描述的等方差的好处，但它们都没有得到强",
"对等变网络的一些研究，虽然不是全部，但包括证明",
"有力的一般理论保证或所有实验证据的支持。为了更",
"模型在组变换下泛化能力的测试e。g。(Gongetal.,",
"好地表征在ML模型中实施等方差的可能好处，我们",
"2022）中的图3)。此外，一些研究专门承担了泛化的",
"研究了等变GNN在两个重要的粒子物理数据重建任",
"任务。。例如，（Wangetal.，2020a）的作者证明了他们",
"务上的性能",
"的E（2）-等变CNN在平移，旋转和线性缩放方面的",
"鲁棒性比在非等变模型下更好地泛化（尽管它仍然不",
"3.1.Top Jet Tagging",
"能泛化整个测试数据域）",
"在高能粒子物理实验中，例如大型强子对撞机（LHC)",
"ModelEficiency先前的工作表明，随着模型中内置",
"的ATLAS和CMS探测器，射流在粒子相互作用的",
"数据的对称性，等变网络有一个“更容易”的时间学",
"研究中起着举足轻重的作用。射流是在质子-质子碰撞",
"习最佳函数以最小化训练损失。具体来说，已经提出，",
"过程中产生高能夸克或胶子时形成的粒子的准直喷雾。",
"对于给定数量的可学习参数，等变模型可能比非等变",
"由于强力的性质以及夸克和胶子的限制，这些粒子迅",
"模型更有效地使用这些参数。在实践中，（Cohen&",
"速碎裂并强子化，导致可检测的粒子聚集在一个狭窄",
"Welling，2017）的作者证明了这种模型的效率，他们表",
"的圆锥体内（图1）。识别原始夸克或胶子的类型（“喷",
"明，通过基本特征的组合创建表示的“可控CNN”，每",
"射标记”）对于实现精确测量和寻找新的物理过程至关",
"个特征都与特定类型的对称性相关联，能够更有效地",
"重要",
"利用模型参数为CIFAR基准生成较小的SotA图像分",
"类器",
"洛伦兹对称性是物理学中的基本原理，源于狭义相对",
"论。它描述了时空旋转和提升下物理定律的不变性",
"模型效率是非常可取的，特别是在资源受限的计算环",
"洛伦兹组是所有矩阵的集合，这些矩阵保留了时空向",
"境中。因此，一些作者，例如（Bogatskiyetal.，2020)",
"量u,uER4和J=diag(1,-1,-1,1)的双线性形式",
"，将其等变模型的大小与类似执行的非等变模型相比",
"n(u,)=uTJu，即 Minkowski 度量。Lorentz 对称",
"equivariance in particle physics, and the conclusions may",
"惯性框架被限制为正定向和正时间定向，由适当的正",
"not apply directly to systems with discrete symmetries or",
"交Lorentz组SO(1.3）描述。在洛伦兹变换下，顶部",
"significantly different metrics, as in chemistry and biology.",
"射流标记的任务是不变的，因为原始夸克或胶子类型",
"2"
] | {
"arxiv_id": "2311.03094",
"language": "zh",
"page_idx": 1
} |
||
2311.03094-zh-3 | 2311.03094-zh-3.jpg | Title Suppressed Due to Excessive Size
<table>
<tbody>
<tr>
<th>Model</th>
<th>$\operatorname{Accuracy}$</th>
<th>AUC</th>
<th>-1 $\underline{\epsilon_B}$</th>
<th>$N$ parameters</th>
<th>Ant Factor v2 ( $\times$1$0^{5}$</th>
</tr>
<tr>
<td>ResNeXt</td>
<td>0.936</td>
<td>0.984</td>
<td>$1122\pm47$</td>
<td>1.46 M</td>
<td>4.28</td>
</tr>
<tr>
<td>$P.$ articleNet</td>
<td>0.938</td>
<td>0.985</td>
<td>$1298\pm46$</td>
<td>498k</td>
<td>13.4</td>
</tr>
<tr>
<td>$PFN$</td>
<td>0.932</td>
<td>0.982</td>
<td>$891\pm18$</td>
<td>82k</td>
<td>67.8</td>
</tr>
<tr>
<td>$EFP$</td>
<td>0.932</td>
<td>0.980</td>
<td>384</td>
<td>1k</td>
<td>5000</td>
</tr>
<tr>
<td>$LGN$</td>
<td>0.929</td>
<td>0.964</td>
<td>$435\pm95$</td>
<td>4.5k</td>
<td>617</td>
</tr>
<tr>
<td>VecNet.1</td>
<td>0.935</td>
<td>0.984</td>
<td> </td>
<td>633k</td>
<td>9.87</td>
</tr>
<tr>
<td>VecNet.2</td>
<td>0.931</td>
<td>0.981</td>
<td> </td>
<td>15k</td>
<td>350</td>
</tr>
<tr>
<td>PELICAN</td>
<td>0.943</td>
<td>0.987</td>
<td>2289 $\pm204$</td>
<td>45k</td>
<td>171</td>
</tr>
<tr>
<td>LorentzNet</td>
<td>0.942</td>
<td>0.9868</td>
<td>$2195\pm173$</td>
<td>220k</td>
<td>35.0</td>
</tr>
</tbody>
</table>
Table 1.Performance comparison of SotA and near-SotA top tagging models $^3.$ VecNet performance metrics are reproduced from (Murnane et al., 2023b) while all other values are reproduced from (Bogatskiy et al., 2022b) . $\epsilon_B^{-1}$ is the background rejection rate (the inverse of the false postive rate) for 30% signal efficiency. Ant Factor v2 is $[(1-AUC)\cdot\mathrm{N}_{\text{parameters}}]^{-1}.$
使用消息构造
之间的关系。
(4)
$$m_{i,j}=\phi(h_i,h_j,||x_i-x_j||^2)$$
作为公式 1 和 2 的输入
EuclidNet 和 SotA 非等变跟踪 GNN 的性能(表 2中列出了各种大小的 InteractionNet (DeZoort et al. 2021) 。所有模型均在 TrackML 数据集 (Amrouche et al., 2020) 上进行评估，该数据集是由一组模拟的质子-质子碰撞事件组成的公共数据集，每个事件都包含三维命中位置以及有关生成它们的粒子的真相信息。这些事件是用 200 堆模拟的，模拟了 LHC 未来升级中预期的高粒子多重性
4.2. Generalizability
为了描述模型对输人数据转换的通用性，可以测试随着连续转换的应用，模型精度如何变化。图 3a 显示了顶部喷射标记器的精度，它是应用于数据参考系的洛伦兹增强的函数，由$\beta=v/c$针对光速$c$进行参数化。随着增强的应用，非洛伦兹等变模型的精度迅速下降， 而洛伦兹网络的精度在几乎整个增强范围内保持不变仅在应用最大增强时略有下降。这表明，正如预期的那样，LorentzNet 能够很好地推广
然而，有趣的是，另一个洛伦兹等变模型 LGN 并不能很好地概括，因为对于大的洛伦兹提升，精度开始急剧下降，当应用最大的提升时，其原始精度只有$\overline{\mathrm{E}}$ $70\%$。在原始论文中，作者将这种性能下降归因于用于存储模型权重 (Bogatskiy et al., 2020) 的位精度引起的不稳定性，这表明在 ML 体系结构中实施等方差的特定方法对模型的完全泛化能力有影响
当然，等方差并不是模型在对称变换下学习泛化的唯一方法。如图 3b 所示，一个简单的 4 层全连接神经网络在 TopJetTaggingDataset 的增强版本上训练，并将随机 Lorentz 提升应用于训练事件，能够像 LGN— 样概括(一旦考虑到非提升数据的准确性差异)。此外，在某些情况下，即使没有训练数据增强，模型也能够学习对转换的鲁棒性。如图 4 所示，非等变量InteractionNet 在应用于测试数据集的 2 维旋转范围内保持稳定的性能
# 4. Evaluating Equivariance
使用这些顶级喷流标记和跟踪模型的性能，我们现在
评估是否实现了 2 节中描述的等方差的拟议好处
4.1. Accuracy
乍一看，似乎执行等方差可以提高模型的准确性。在顶级喷气标记器中，具有最高精度和 AUC 的两个模型PELICAN 和 LorentzNet 都是洛伦兹不变的。然而， 在所有考虑的模型中，另一个洛伦兹等变模型 LGN 的精度和 AUC 最低，这表明精度的提高可能至少部分是由于其他架构选择。特别是，PELICAN 将所有可能的洛伦兹不变量作为输入特征，并创新地利用了 15种不同的聚合器，在模型中创建了更丰富的数据表示。此外，VecNet。1 优于 LGN 和几个非等变模型，尽管只有部分等变
对于跟踪模型，图片同样不清楚。对于非常小的模型尺寸，SO (2) 等变 EuclidNet 具有较高的 AUC。然而 , 随着模型潜在空间的大小增加 , EuclidNet 性能开始趋于平稳，不受约束的 InteractionNet 变得更加高性能。我们在 4.3 节中进一步探讨了模型大小与精度
$^{3}$We do not include background rejection values for VecNet.1 and VecNet.2 because the authors have noted an error in the original calculation but have not yet published corrected numbers.
4 | [
[
"0.3586",
"0.05807",
"0.6167",
"0.05807",
"0.6167",
"0.07135",
"0.3586",
"0.07135"
],
[
"0.4558",
"0.0871",
"0.486",
"0.0871",
"0.486",
"0.1061",
"0.4558",
"0.1061"
],
[
"0.528",
"0.0871",
"0.613",
"0.0903",
"0.6113",
"0.1092",
"0.527",
"0.1061"
],
[
"0.281",
"0.0884",
"0.3489",
"0.0903",
"0.3489",
"0.1061",
"0.281",
"0.1042"
],
[
"0.1805",
"0.089",
"0.2296",
"0.089",
"0.2296",
"0.1042",
"0.1805",
"0.1042"
],
[
"0.3708",
"0.089",
"0.411",
"0.089",
"0.411",
"0.1042",
"0.3708",
"0.1042"
],
[
"0.6274",
"0.089",
"0.7935",
"0.089",
"0.7935",
"0.1042",
"0.6274",
"0.1042"
],
[
"0.179",
"0.108",
"0.2467",
"0.10986",
"0.2467",
"0.125",
"0.179",
"0.1225"
],
[
"0.2932",
"0.10986",
"0.335",
"0.10986",
"0.335",
"0.125",
"0.2932",
"0.125"
],
[
"0.3694",
"0.10986",
"0.411",
"0.10986",
"0.411",
"0.125",
"0.3694",
"0.125"
],
[
"0.437",
"0.10986",
"0.5073",
"0.10986",
"0.5073",
"0.125",
"0.437",
"0.125"
],
[
"0.5435",
"0.10986",
"0.597",
"0.10986",
"0.597",
"0.125",
"0.5435",
"0.125"
],
[
"0.693",
"0.10986",
"0.726",
"0.10986",
"0.726",
"0.1256",
"0.693",
"0.1256"
],
[
"0.179",
"0.125",
"0.2656",
"0.127",
"0.2656",
"0.1414",
"0.179",
"0.1395"
],
[
"0.2932",
"0.1256",
"0.335",
"0.1256",
"0.335",
"0.1401",
"0.2932",
"0.1401"
],
[
"0.3694",
"0.1256",
"0.411",
"0.1256",
"0.411",
"0.1401",
"0.3694",
"0.1401"
],
[
"0.437",
"0.1256",
"0.509",
"0.1256",
"0.509",
"0.1401",
"0.437",
"0.1401"
],
[
"0.5513",
"0.127",
"0.589",
"0.127",
"0.589",
"0.1401",
"0.5513",
"0.1401"
],
[
"0.694",
"0.1256",
"0.728",
"0.1256",
"0.728",
"0.1414",
"0.694",
"0.1414"
],
[
"0.2932",
"0.1383",
"0.335",
"0.1401",
"0.3342",
"0.1559",
"0.2917",
"0.154"
],
[
"0.5537",
"0.1401",
"0.585",
"0.1401",
"0.585",
"0.1559",
"0.5537",
"0.1559"
],
[
"0.693",
"0.1395",
"0.7295",
"0.1427",
"0.726",
"0.1572",
"0.6914",
"0.1547"
],
[
"0.3708",
"0.1427",
"0.4084",
"0.1427",
"0.4084",
"0.154",
"0.3708",
"0.154"
],
[
"0.1805",
"0.1414",
"0.2157",
"0.1414",
"0.2157",
"0.154",
"0.1805",
"0.154"
],
[
"0.4412",
"0.1427",
"0.5034",
"0.1427",
"0.5034",
"0.154",
"0.4412",
"0.154"
],
[
"0.2932",
"0.1547",
"0.335",
"0.1547",
"0.335",
"0.1692",
"0.2932",
"0.1692"
],
[
"0.3694",
"0.1547",
"0.4102",
"0.1547",
"0.4102",
"0.1692",
"0.3694",
"0.1692"
],
[
"0.4558",
"0.1547",
"0.4885",
"0.1547",
"0.4885",
"0.1704",
"0.4558",
"0.1704"
],
[
"0.6914",
"0.154",
"0.7295",
"0.1559",
"0.728",
"0.1704",
"0.6895",
"0.1686"
],
[
"0.5596",
"0.1559",
"0.581",
"0.1559",
"0.581",
"0.1692",
"0.5596",
"0.1692"
],
[
"0.1805",
"0.1559",
"0.2157",
"0.1559",
"0.2157",
"0.1686",
"0.1805",
"0.1686"
],
[
"0.2932",
"0.1704",
"0.335",
"0.1704",
"0.335",
"0.1849",
"0.2932",
"0.1849"
],
[
"0.3694",
"0.1704",
"0.411",
"0.1704",
"0.411",
"0.1849",
"0.3694",
"0.1849"
],
[
"0.4395",
"0.1704",
"0.5034",
"0.1704",
"0.5034",
"0.1849",
"0.4395",
"0.1849"
],
[
"0.5537",
"0.1718",
"0.5864",
"0.1718",
"0.5864",
"0.1843",
"0.5537",
"0.1843"
],
[
"0.6953",
"0.1718",
"0.7256",
"0.1718",
"0.7256",
"0.1843",
"0.6953",
"0.1843"
],
[
"0.2932",
"0.1831",
"0.335",
"0.1849",
"0.3342",
"0.2008",
"0.2917",
"0.1989"
],
[
"0.1805",
"0.1718",
"0.2181",
"0.1718",
"0.2181",
"0.1849",
"0.1805",
"0.1849"
],
[
"0.1805",
"0.1849",
"0.2467",
"0.1849",
"0.2467",
"0.1995",
"0.1805",
"0.1995"
],
[
"0.3694",
"0.1849",
"0.411",
"0.1849",
"0.411",
"0.1995",
"0.3694",
"0.1995"
],
[
"0.5513",
"0.1863",
"0.589",
"0.1863",
"0.589",
"0.1995",
"0.5513",
"0.1995"
],
[
"0.693",
"0.1843",
"0.7295",
"0.1875",
"0.726",
"0.2008",
"0.6895",
"0.1989"
],
[
"0.2932",
"0.1989",
"0.3342",
"0.2008",
"0.3325",
"0.2166",
"0.2917",
"0.214"
],
[
"0.4631",
"0.1901",
"0.4805",
"0.1901",
"0.4805",
"0.1995",
"0.4631",
"0.1995"
],
[
"0.1805",
"0.2008",
"0.2484",
"0.2008",
"0.2484",
"0.2153",
"0.1805",
"0.2153"
],
[
"0.3677",
"0.2008",
"0.4102",
"0.2008",
"0.4102",
"0.2153",
"0.3677",
"0.2153"
],
[
"0.5537",
"0.2008",
"0.5864",
"0.2008",
"0.5864",
"0.2166",
"0.5537",
"0.2166"
],
[
"0.694",
"0.2008",
"0.726",
"0.2008",
"0.726",
"0.2166",
"0.694",
"0.2166"
],
[
"0.5537",
"0.2153",
"0.5864",
"0.2153",
"0.5864",
"0.2311",
"0.5537",
"0.2311"
],
[
"0.4617",
"0.2052",
"0.4756",
"0.207",
"0.4739",
"0.2134",
"0.46",
"0.2108"
],
[
"0.2932",
"0.2166",
"0.335",
"0.2166",
"0.335",
"0.2311",
"0.2932",
"0.2311"
],
[
"0.3677",
"0.2166",
"0.4102",
"0.2166",
"0.4102",
"0.2311",
"0.3677",
"0.2311"
],
[
"0.697",
"0.2166",
"0.7256",
"0.2166",
"0.7256",
"0.2311",
"0.697",
"0.2311"
],
[
"0.4314",
"0.2172",
"0.511",
"0.2172",
"0.511",
"0.2291",
"0.4314",
"0.2291"
],
[
"0.1814",
"0.2172",
"0.255",
"0.2172",
"0.255",
"0.2291",
"0.1814",
"0.2291"
],
[
"0.5513",
"0.2317",
"0.589",
"0.2317",
"0.589",
"0.2456",
"0.5513",
"0.2456"
],
[
"0.693",
"0.2291",
"0.7295",
"0.2311",
"0.726",
"0.2468",
"0.6895",
"0.2437"
],
[
"0.2942",
"0.2329",
"0.335",
"0.2329",
"0.335",
"0.2443",
"0.2942",
"0.2443"
],
[
"0.179",
"0.2329",
"0.2607",
"0.2329",
"0.2607",
"0.2443",
"0.179",
"0.2443"
],
[
"0.3652",
"0.2329",
"0.4126",
"0.2329",
"0.4126",
"0.2443",
"0.3652",
"0.2443"
],
[
"0.433",
"0.2329",
"0.5107",
"0.2329",
"0.5107",
"0.2443",
"0.433",
"0.2443"
],
[
"0.0866",
"0.2615",
"0.636",
"0.26",
"0.636",
"0.2778",
"0.0866",
"0.279"
],
[
"0.634",
"0.2664",
"0.6406",
"0.2664",
"0.6406",
"0.2708",
"0.634",
"0.2708"
],
[
"0.642",
"0.2632",
"0.8857",
"0.2632",
"0.8857",
"0.2766",
"0.642",
"0.2766"
],
[
"0.08826",
"0.2766",
"0.88",
"0.2778",
"0.88",
"0.2922",
"0.08826",
"0.2917"
],
[
"0.08905",
"0.2935",
"0.8857",
"0.2935",
"0.8857",
"0.3037",
"0.08905",
"0.3037"
],
[
"0.08826",
"0.3037",
"0.2754",
"0.305",
"0.2754",
"0.3193",
"0.08826",
"0.318"
],
[
"0.08905",
"0.3452",
"0.1887",
"0.3452",
"0.1887",
"0.3599",
"0.08905",
"0.3599"
],
[
"0.5024",
"0.3467",
"0.589",
"0.3467",
"0.589",
"0.358",
"0.5024",
"0.358"
],
[
"0.183",
"0.3801",
"0.3782",
"0.379",
"0.3782",
"0.3933",
"0.183",
"0.3945"
],
[
"0.451",
"0.3801",
"0.4756",
"0.3801",
"0.4756",
"0.3933",
"0.451",
"0.3933"
],
[
"0.4983",
"0.3757",
"0.6504",
"0.377",
"0.6504",
"0.3914",
"0.4983",
"0.3901"
],
[
"0.5024",
"0.401",
"0.884",
"0.401",
"0.884",
"0.414",
"0.5024",
"0.414"
],
[
"0.0907",
"0.414",
"0.259",
"0.414",
"0.259",
"0.428",
"0.0907",
"0.428"
],
[
"0.5024",
"0.4153",
"0.8823",
"0.4153",
"0.8823",
"0.4287",
"0.5024",
"0.4287"
],
[
"0.5024",
"0.4312",
"0.884",
"0.4312",
"0.884",
"0.4443",
"0.5024",
"0.4443"
],
[
"0.08905",
"0.4375",
"0.4756",
"0.4375",
"0.4756",
"0.4507",
"0.08905",
"0.4507"
],
[
"0.0907",
"0.4507",
"0.4722",
"0.452",
"0.4722",
"0.4653",
"0.0907",
"0.464"
],
[
"0.501",
"0.4458",
"0.8823",
"0.4463",
"0.8823",
"0.4602",
"0.501",
"0.459"
],
[
"0.501",
"0.461",
"0.8813",
"0.461",
"0.8813",
"0.4749",
"0.501",
"0.4749"
],
[
"0.08905",
"0.4673",
"0.4739",
"0.4673",
"0.4739",
"0.481",
"0.08905",
"0.481"
],
[
"0.08905",
"0.4817",
"0.4722",
"0.4817",
"0.4722",
"0.4956",
"0.08905",
"0.4956"
],
[
"0.5034",
"0.4778",
"0.8813",
"0.4778",
"0.8813",
"0.488",
"0.5034",
"0.488"
],
[
"0.5024",
"0.4912",
"0.8857",
"0.4912",
"0.8857",
"0.505",
"0.5024",
"0.505"
],
[
"0.09235",
"0.4988",
"0.4722",
"0.4988",
"0.4722",
"0.509",
"0.09235",
"0.509"
],
[
"0.0907",
"0.512",
"0.468",
"0.512",
"0.468",
"0.526",
"0.0907",
"0.526"
],
[
"0.501",
"0.5083",
"0.746",
"0.5083",
"0.746",
"0.518",
"0.501",
"0.518"
],
[
"0.08905",
"0.5264",
"0.4739",
"0.5264",
"0.4739",
"0.5405",
"0.08905",
"0.5405"
],
[
"0.5024",
"0.529",
"0.8857",
"0.529",
"0.8857",
"0.5425",
"0.5024",
"0.5425"
],
[
"0.0907",
"0.5425",
"0.2532",
"0.5425",
"0.2532",
"0.556",
"0.0907",
"0.556"
],
[
"0.5034",
"0.5454",
"0.8823",
"0.5454",
"0.8823",
"0.556",
"0.5034",
"0.556"
],
[
"0.5024",
"0.5596",
"0.8857",
"0.5596",
"0.8857",
"0.5728",
"0.5024",
"0.5728"
],
[
"0.0866",
"0.577",
"0.3228",
"0.5776",
"0.3228",
"0.5923",
"0.0866",
"0.5913"
],
[
"0.5",
"0.5728",
"0.884",
"0.5737",
"0.884",
"0.5884",
"0.5",
"0.587"
],
[
"0.5024",
"0.589",
"0.8857",
"0.589",
"0.8857",
"0.603",
"0.5024",
"0.603"
],
[
"0.0907",
"0.603",
"0.4739",
"0.603",
"0.4739",
"0.616",
"0.0907",
"0.616"
],
[
"0.501",
"0.603",
"0.8857",
"0.604",
"0.8857",
"0.6187",
"0.501",
"0.617"
],
[
"0.08905",
"0.6187",
"0.4314",
"0.6187",
"0.4314",
"0.629",
"0.08905",
"0.629"
],
[
"0.501",
"0.619",
"0.7793",
"0.619",
"0.7793",
"0.6333",
"0.501",
"0.6333"
],
[
"0.08826",
"0.6484",
"0.1929",
"0.6494",
"0.1929",
"0.664",
"0.08826",
"0.6636"
],
[
"0.5024",
"0.6426",
"0.8857",
"0.6426",
"0.8857",
"0.656",
"0.5024",
"0.656"
],
[
"0.5146",
"0.659",
"0.884",
"0.659",
"0.884",
"0.6694",
"0.5146",
"0.6694"
],
[
"0.09235",
"0.675",
"0.4722",
"0.675",
"0.4722",
"0.685",
"0.09235",
"0.685"
],
[
"0.5024",
"0.6724",
"0.8857",
"0.6724",
"0.8857",
"0.686",
"0.5024",
"0.686"
],
[
"0.08905",
"0.688",
"0.4739",
"0.688",
"0.4739",
"0.702",
"0.08905",
"0.702"
],
[
"0.501",
"0.6875",
"0.865",
"0.6875",
"0.865",
"0.7007",
"0.501",
"0.7007"
],
[
"0.08905",
"0.7026",
"0.4739",
"0.7026",
"0.4739",
"0.7163",
"0.08905",
"0.7163"
],
[
"0.5024",
"0.702",
"0.8857",
"0.702",
"0.8857",
"0.7153",
"0.5024",
"0.7153"
],
[
"0.08905",
"0.7183",
"0.4722",
"0.7183",
"0.4722",
"0.732",
"0.08905",
"0.732"
],
[
"0.5024",
"0.7173",
"0.8857",
"0.7173",
"0.8857",
"0.731",
"0.5024",
"0.731"
],
[
"0.08905",
"0.733",
"0.4714",
"0.733",
"0.4714",
"0.747",
"0.08905",
"0.747"
],
[
"0.501",
"0.732",
"0.884",
"0.732",
"0.884",
"0.7456",
"0.501",
"0.7456"
],
[
"0.0907",
"0.7485",
"0.4722",
"0.7485",
"0.4722",
"0.762",
"0.0907",
"0.762"
],
[
"0.501",
"0.7476",
"0.8857",
"0.7476",
"0.8857",
"0.761",
"0.501",
"0.761"
],
[
"0.0907",
"0.7646",
"0.4739",
"0.7646",
"0.4739",
"0.778",
"0.0907",
"0.778"
],
[
"0.501",
"0.763",
"0.634",
"0.763",
"0.634",
"0.7764",
"0.501",
"0.7764"
],
[
"0.0907",
"0.78",
"0.4697",
"0.78",
"0.4697",
"0.7935",
"0.0907",
"0.7935"
],
[
"0.5215",
"0.785",
"0.8857",
"0.785",
"0.8857",
"0.799",
"0.5215",
"0.799"
],
[
"0.0907",
"0.7944",
"0.4722",
"0.7944",
"0.4722",
"0.808",
"0.0907",
"0.808"
],
[
"0.5",
"0.7974",
"0.884",
"0.8",
"0.884",
"0.813",
"0.5",
"0.811"
],
[
"0.09235",
"0.811",
"0.1887",
"0.811",
"0.1887",
"0.8213",
"0.09235",
"0.8213"
],
[
"0.501",
"0.8125",
"0.8857",
"0.8125",
"0.8857",
"0.8257",
"0.501",
"0.8257"
],
[
"0.501",
"0.8247",
"0.6274",
"0.8247",
"0.6274",
"0.8384",
"0.501",
"0.8384"
],
[
"0.0907",
"0.8306",
"0.4722",
"0.8306",
"0.4722",
"0.8447",
"0.0907",
"0.8447"
],
[
"0.0907",
"0.8467",
"0.4739",
"0.8467",
"0.4739",
"0.86",
"0.0907",
"0.86"
],
[
"0.08905",
"0.8613",
"0.4722",
"0.8623",
"0.4722",
"0.8755",
"0.08905",
"0.875"
],
[
"0.08905",
"0.8755",
"0.4739",
"0.877",
"0.4739",
"0.8916",
"0.08905",
"0.89"
],
[
"0.08905",
"0.8926",
"0.4739",
"0.8926",
"0.4739",
"0.906",
"0.08905",
"0.906"
],
[
"0.482",
"0.924",
"0.4944",
"0.924",
"0.4944",
"0.935",
"0.482",
"0.935"
]
] | [
"Title Suppressed Due to Excessive Size",
"EBl",
"Nparameters",
"Accuracy",
"Model",
"AUC",
"Ant Factor v2 (×105)",
"ResNeXt",
"0.936",
"0.984",
"1122±47",
"1.46 M",
"4.28",
"ParticleNet",
"0.938",
"0.985",
"1298±46",
"498k",
"13.4",
"0.932",
"82k",
"67.8",
"0.982",
"PFN",
"891±18",
"0.932",
"0.980",
"384",
"5000",
"1k",
"EFP",
"0.929",
"0.964",
"435 ± 95",
"4.5k",
"617",
"0.935",
"LGN",
"VecNet.1",
"0.984",
"633k",
"9.87",
"0.931",
"",
"VecNet.2",
"0.981",
"15k",
"350",
"45k",
"-",
"0.943",
"0.987",
"171",
"2289±204",
"PELICAN",
"220k",
"35.0",
"0.942",
"LorentzNet",
"0.9868",
"2195±173",
"Table 1. Performance comparison of SotA and near-SotA top tagging models 3.",
"",
"VecNet performance metrics are re-",
"produced from (Murnane et al., 2023b)while all other values are reproduced from (Bogatskiy et al., 2022b):eB",
"is the background rejection rate (the inverse of the false postive rate)for 30 % signal efficiency.Ant Factor v2 is",
"[(1 - AUC) -Nparameters]-1.",
"使用消息构造",
"之间的关系。",
"mij=o(hi,hj,lai-aill2)",
"(4)",
"4.2.Generalizability",
"为了描述模型对输入数据转换的通用性，可以测试随",
"作为公式1和2的输入",
"着连续转换的应用，模型精度如何变化。图3a显示了",
"顶部喷射标记器的精度，它是应用于数据参考系的洛",
"EuclidNet和SotA非等变跟踪GNN的性能（表2",
"中列出了各种大小的InteractionNet（DeZoortetal",
"伦兹增强的函数，由β=u/c针对光速c进行参数化",
"随着增强的应用，非洛伦兹等变模型的精度迅速下降",
"2021）。所有模型均在TrackML数据集（Amrouche",
"etal.，2020）上进行评估，该数据集是由一组模拟的",
"而洛伦兹网络的精度在儿乎整个增强范围内保持不变",
"仅在应用最大增强时略有下降。这表明，正如预期的",
"质子-质子碰撞事件组成的公共数据集，每个事件都包",
"含三维命中位置以及有关生成它们的粒子的真相信息",
"那样，LorentzNet能够很好地推厂",
"这些事件是用200堆模拟的，模拟了LHC未来升级",
"然而，有趣的是，另一个洛伦兹等变模型LGN并不",
"中预期的高粒子多重性",
"能很好地概括，因为对于大的洛伦兹提升：精度开始",
"急剧下降，当应用最大的提升时，其原始精度只有E",
"4.EvaluatingEquivariance",
"70%。在原始论文中，作者将这种性能下降归因于用于",
"存储模型权重（Bogatskiyetal.，2020）的位精度引起",
"使用这些顶级喷流标记和跟踪模型的性能，我们现在",
"的不稳定性，这表明在ML体系结构中实施等方差的",
"评估是否实现了2节中描述的等方差的拟议好处",
"特定方法对模型的完全泛化能力有影响",
"4.1.Accuracy",
"当然，等方差并不是模型在对称变换下学习泛化的唯",
"方法。如图3b所示，一个简单的4层全连接神经",
"乍一看，似乎执行等方差可以提高模型的准确性。在顶",
"网络在TopJetTaggingDataset的增强版本上训练，并",
"级喷气标记器中，具有最高精度和AUC的两个模型",
"将随机Lorentz提升应用于训练事件，能够像LGN",
"PELICAN和LorentzNet都是洛伦兹不变的。然而，",
"样概括（一旦考虑到非提升数据的准确性差异）。此",
"在所有考虑的模型中，另一个洛伦兹等变模型LGN的",
"外，在某些情况下，即使没有训练数据增强，模型也",
"精度和AUC最低，这表明精度的提高可能至少部分",
"能够学习对转换的鲁棒性。如图4所示，非等变量",
"是由于其他架构选择。特别是，PELICAN将所有可",
"InteractionNet在应用于测试数据集的2维旋转范围",
"能的洛伦兹不变量作为输入特征，并创新地利用了15",
"内保持稳定的性能",
"种不同的聚合器，在模型中创建了更丰富的数据表示",
"3We do not include background rejection values for Vec-",
"此外，VecNet。1优于LGN和几个非等变模型，尽管",
"Net.1 and VecNet.2 because the authors have noted an",
"只有部分等变",
"error in the original calculation but have not yet published",
"corrected numbers.",
"对于跟踪模型，图片同样不清楚。对于非常小的模型",
"尺寸，SO（2）等变EuclidNet具有较高的AUC。然",
"而，随着模型潜在空间的大小增加，EuclidNet性能开",
"始趋于平稳，不受约束的InteractionNet变得更加高",
"性能。我们在4.3节中进一步探讨了模型大小与精度",
"4"
] | {
"arxiv_id": "2311.03094",
"language": "zh",
"page_idx": 3
} |
||
2307.14490-zh-2 | 2307.14490-zh-2.jpg | KDD'23, August 6-10, 2023, Long Beach, CA, USA
HUGE: Huge Unsupervised Graph Embeddings with TPUs
Table 1: An overview of different approaches to scaling graph embeddings up. In this work, we demonstrate a system that works without any of the yet mentioned techniques scaled to largest graphs. Scalability is given as an approximate graph size that can be processed by best-performing algorithm & system combination in a day.
<table>
<tbody>
<tr>
<th>Method Family</th>
<th>Speedup Technique</th>
<th>Reference Methods</th>
<th>Scalability</th>
</tr>
<tr>
<td>neural</td>
<td> </td>
<td>DeepWalk, LINE, VERSE</td>
<td>$10M$</td>
</tr>
<tr>
<td>neural</td>
<td>graph coarsening</td>
<td>HARP, MILE</td>
<td>$10M$</td>
</tr>
<tr>
<td>neural</td>
<td>partitioning</td>
<td>BigGraph, EDGES, DeLNE</td>
<td>1000M</td>
</tr>
<tr>
<td>factorization</td>
<td> </td>
<td>NetMF, GraRep</td>
<td>10k</td>
</tr>
<tr>
<td>factorization</td>
<td>matrix sparsification</td>
<td>NetSMH NetMFSC STRAP.NRP</td>
<td>$100M$</td>
</tr>
<tr>
<td>factorization</td>
<td>implicit solvers</td>
<td>HOPE, AROPE</td>
<td>$100M$</td>
</tr>
<tr>
<td>factorization</td>
<td>$\operatorname{spectral}$ propagation</td>
<td>ProNE, LightNE</td>
<td>1000M</td>
</tr>
<tr>
<td>factorization</td>
<td>matrix sketching</td>
<td>FastRH Embeddin</td>
<td>$1000M$</td>
</tr>
</tbody>
</table>
# 2 ВАСКGROUND
在本节中，我们首先简要回顾一下$\overline{\mathrm{E}}$2.1 节中的相关工作。我们在第 2.2 节中回顾了 DeepWalk [22] ,我们将其用作高性能嵌人系统的基础。然后，我们继续描述使用商品(HUGE-CPU )和 TPU (HUGE-TPU) 硬件扩展 Deepwalk 图嵌入的两种体系结构，这些硬件使我们可以扩展到$\overline{\mathrm{E}}$2.3 节中的大图
代数技术[28]被证明是一种更有成效的方法。隐式求解器$\overline{\mathrm{E}}$ [21,44]可以分解矩阵，而无需在内存中显式实现它。这些方法被限制为执行线性分解，这不能成功地解释图的结构
产生最可扩展的嵌人方法的两个技术家族是光谱传播和矩阵草图 [35] 。谱传播方法 [25, 42] 首先计算图的邻接或拉普拉斯矩阵的一些截断特征分解，然后使用这些特征向量来模拟信息的扩散。矩阵草绘方法近似相似矩阵，迭代[8,37,43]或单遍[24]。后一种选择更具可扩展性
## 2.1 Related Work
我们现在继续回顾嵌入大图的两种基本方法。在过去的几年里，大量的工作引人了各种嵌人方法以及无数的技术和硬件架构来扩展它们。我们在表 1 中总结了嵌人方法、加速技术及其预期的可扩展性方面的相关工作。
2.1.1 $CraphEmbeddingApproaches.$我们将嵌入方法分类为基
2.1.3 Hardware-based Embedding Acceleration .与算法的进步相比，基于硬件的加速可以说受到的关注较少。Zhu et al.[45] 提出了一个混合系统，该系统使用 CPU 进行采样，并使用GPU 来训练嵌人。由于单个 GPU 可以提供的最大 RAM 约为100GB, 因此只能在此类系统上训练 1 亿个节点图的嵌人。Wei et al.[34], Yang et al.[36] 通过分区以包含更多 GPU 来解决此问题。这种方法需要数十个 GPU 来处理十亿节点的图形，与基于可扩展 CPU 的系统相比，这是令人望而却步的，后者可以在数小时内将十亿节点的图形嵌人到单个高内存计算机上
高效计算高阶相似度是目前硬件加速不足的一个方面。Wang et al.[33], Yang et al.[39] 提出了用于通用硬件体系结构的随机游走生成的高效系统。但是，在没有合适的嵌入方法的情况下，这些系统对于图嵌人没有用
于神经网络或基于矩阵分解。无论采用哪种方法，每种方法有时都会隐式地使用相似性函数，该函数将每个节点与图中的其他节点相关联。性能最好的方法从仅使用图中的邻接信息到基于随机游走的相似性的某些概念，例如个性化的 PageRank (PPR)$\overline{\mathbb{E}}[3]$。加速这些相似性计算的一个关键见解是，它们在图[2]中高度本地化，这意味着 2-3 个传播步骤足以近似它们
神经嵌人方法将节点嵌入视为浅层神经网络的参数。神经方法通过随机梯度下降法为邻接$\mathbb{E}[29]$ ,随机游走$\mathbb{E}[$22] 或个性化的 PageRank$\underline{\mathbb{E}}[30]$相似性函数优化这些参数。这种优化是通过采样完成的，对嵌人表的更新通常非常稀疏。因此，随机存储器访问通常限制了这些方法的性能
基于梯度的方法的替代方法是直接对相似性矩阵进行分解。神经和矩阵分解方法之间有着深刻的联系 [27,31]-本质上，对于许多节点相似性，基于神经和基于分解的嵌人的最优解一致。基于矩阵的方法的主要挑战是保持中间表示的稀疏性。对于大图，既不能增加邻接矩阵的密度，也不能保留太多的中间投影
# 2.2 DeepWalk
在描述我们的 TPU 嵌人系统之前，有必要回顾一下 Deep-
Walk$\boxed{\mathrm{F}}[22]$ ,这是神经图嵌人的基本方法。DeepWalk 使 word2vec$\boxed{\mathrm{F}}$
[18] (一种广泛成功的嵌入单词的模型) 适应图形数据。DeepWalk 生成短随机游走的“语料库”; DeepWalk 的目标是最大化在某个特定窗口大小内随机游走中观察相邻顶点的后验概率。为了有效地最大化此概率，它使用分层的 softmax$\boxed{\mathbb{E} }[ 19]$ , 它根据节点的出现频率或计算效率更高的近似值，负采样$\mathbb{E}$ [10]构造节点的霍夫曼树。对于在某个节点的窗口大小内观察到的每个节点，DeepWalk 随机均匀地选择$k\ll n$作为对比的负面示例
DeepWalk 的架构有几个计算问题，如果我们要将 DeepWalk
扩展到具有数十亿个节点的图形，则需要解决这些问题：
$\bullet$ Random walk generation for large graphs is computation-
ally prohibitive due to random memory accesses on each
random walk step.
$\bullet$ Random walk corpus grows in size rapidly, growing much
larger in size than the original sparse graph.
2.1.2 Scaling Graph Embedding Systems.加速嵌入算法有几个方向-有些是针对特定方法定制的，而有些是更通用的。现在， 我们简要回顾一下最通用的加速技术。图粗化[7,17] 迭代地收缩图，学习最压缩级别的嵌人，并在收缩层次结构中确定性地传播嵌入。图分区方法[12,16]跨机器分布计算，同时尝试最小化跨机器的通信
早期的矩阵分解方法$\left[26,40,41\right]$尝试稀疏随机游走或 PPR 矩阵。不幸的是，对于这些嵌人方法来说，高阶相似度矩阵仍然太密集，无法扩展到非常大的图。利用专门的稀疏数值线性 | [
[
"0.0866",
"0.0745",
"0.3628",
"0.07574",
"0.3628",
"0.089",
"0.0866",
"0.0884"
],
[
"0.675",
"0.0745",
"0.911",
"0.07574",
"0.911",
"0.089",
"0.675",
"0.0884"
],
[
"0.0866",
"0.1061",
"0.9126",
"0.1061",
"0.9126",
"0.1193",
"0.0866",
"0.1193"
],
[
"0.08496",
"0.11743",
"0.9126",
"0.1193",
"0.9126",
"0.1338",
"0.08496",
"0.132"
],
[
"0.0866",
"0.1338",
"0.642",
"0.1338",
"0.642",
"0.1477",
"0.0866",
"0.1477"
],
[
"0.10864",
"0.1635",
"0.2035",
"0.1654",
"0.2035",
"0.1799",
"0.10864",
"0.178"
],
[
"0.2524",
"0.1666",
"0.3718",
"0.1666",
"0.3718",
"0.1799",
"0.2524",
"0.1799"
],
[
"0.511",
"0.1666",
"0.63",
"0.1666",
"0.63",
"0.1799",
"0.511",
"0.1799"
],
[
"0.794",
"0.1635",
"0.863",
"0.1654",
"0.863",
"0.1812",
"0.794",
"0.1787"
],
[
"0.1356",
"0.1849",
"0.179",
"0.1849",
"0.179",
"0.1995",
"0.1356",
"0.1995"
],
[
"0.4958",
"0.1849",
"0.6465",
"0.1849",
"0.6465",
"0.1989",
"0.4958",
"0.1989"
],
[
"0.8164",
"0.1818",
"0.847",
"0.1849",
"0.8433",
"0.2008",
"0.812",
"0.1976"
],
[
"0.2573",
"0.1995",
"0.367",
"0.1995",
"0.367",
"0.214",
"0.2573",
"0.214"
],
[
"0.8164",
"0.1995",
"0.8447",
"0.1995",
"0.8447",
"0.2122",
"0.8164",
"0.2122"
],
[
"0.5317",
"0.2008",
"0.6094",
"0.2008",
"0.6094",
"0.2108",
"0.5317",
"0.2108"
],
[
"0.1323",
"0.2008",
"0.1772",
"0.1963",
"0.1805",
"0.2108",
"0.1356",
"0.2153"
],
[
"0.1356",
"0.2134",
"0.179",
"0.2134",
"0.179",
"0.2279",
"0.1356",
"0.2279"
],
[
"0.2737",
"0.2134",
"0.3506",
"0.2134",
"0.3506",
"0.2279",
"0.2737",
"0.2279"
],
[
"0.4902",
"0.2134",
"0.652",
"0.2134",
"0.652",
"0.2267",
"0.4902",
"0.2267"
],
[
"0.808",
"0.2134",
"0.853",
"0.2134",
"0.853",
"0.2279",
"0.808",
"0.2279"
],
[
"0.1177",
"0.2343",
"0.1953",
"0.2343",
"0.1953",
"0.2456",
"0.1177",
"0.2456"
],
[
"0.524",
"0.2343",
"0.6167",
"0.2343",
"0.6167",
"0.2456",
"0.524",
"0.2456"
],
[
"0.8164",
"0.2329",
"0.8447",
"0.2329",
"0.8447",
"0.2474",
"0.8164",
"0.2474"
],
[
"0.1177",
"0.2474",
"0.1953",
"0.2474",
"0.1953",
"0.2588",
"0.1177",
"0.2588"
],
[
"0.2484",
"0.2468",
"0.3757",
"0.2468",
"0.3757",
"0.26",
"0.2484",
"0.26"
],
[
"0.4739",
"0.2474",
"0.6665",
"0.2474",
"0.6665",
"0.2583",
"0.4739",
"0.2583"
],
[
"0.814",
"0.2474",
"0.847",
"0.2474",
"0.847",
"0.2588",
"0.814",
"0.2588"
],
[
"0.1177",
"0.2615",
"0.1953",
"0.2615",
"0.1953",
"0.2727",
"0.1177",
"0.2727"
],
[
"0.2656",
"0.2615",
"0.3586",
"0.2615",
"0.3586",
"0.2727",
"0.2656",
"0.2727"
],
[
"0.5254",
"0.2615",
"0.615",
"0.2615",
"0.615",
"0.2727",
"0.5254",
"0.2727"
],
[
"0.812",
"0.2615",
"0.8506",
"0.2615",
"0.8506",
"0.2747",
"0.812",
"0.2747"
],
[
"0.1152",
"0.2727",
"0.1969",
"0.2747",
"0.1969",
"0.289",
"0.1152",
"0.2874"
],
[
"0.2484",
"0.2734",
"0.3743",
"0.2734",
"0.3743",
"0.2874",
"0.2484",
"0.2874"
],
[
"0.523",
"0.276",
"0.619",
"0.276",
"0.619",
"0.286",
"0.523",
"0.286"
],
[
"0.808",
"0.2734",
"0.853",
"0.2734",
"0.853",
"0.2886",
"0.808",
"0.2886"
],
[
"0.2573",
"0.2854",
"0.3652",
"0.2886",
"0.3652",
"0.303",
"0.2573",
"0.2998"
],
[
"0.4224",
"0.2874",
"0.7173",
"0.2886",
"0.7173",
"0.3018",
"0.4224",
"0.3005"
],
[
"0.8096",
"0.289",
"0.851",
"0.289",
"0.851",
"0.3005",
"0.8096",
"0.3005"
],
[
"0.5195",
"0.3257",
"0.914",
"0.3257",
"0.914",
"0.339",
"0.5195",
"0.339"
],
[
"0.1168",
"0.2886",
"0.1936",
"0.2886",
"0.1936",
"0.2998",
"0.1168",
"0.2998"
],
[
"0.0866",
"0.327",
"0.1168",
"0.327",
"0.1168",
"0.3384",
"0.0866",
"0.3384"
],
[
"0.1168",
"0.327",
"0.246",
"0.327",
"0.246",
"0.3372",
"0.1168",
"0.3372"
],
[
"0.0866",
"0.3435",
"0.482",
"0.3435",
"0.482",
"0.3567",
"0.0866",
"0.3567"
],
[
"0.5195",
"0.3403",
"0.9126",
"0.3403",
"0.9126",
"0.3542",
"0.5195",
"0.3542"
],
[
"0.08826",
"0.3591",
"0.4805",
"0.3591",
"0.4805",
"0.3694",
"0.08826",
"0.3694"
],
[
"0.5195",
"0.356",
"0.871",
"0.356",
"0.871",
"0.3662",
"0.5195",
"0.3662"
],
[
"0.0866",
"0.3718",
"0.4805",
"0.3718",
"0.4805",
"0.385",
"0.0866",
"0.385"
],
[
"0.536",
"0.3694",
"0.9126",
"0.3694",
"0.9126",
"0.3801",
"0.536",
"0.3801"
],
[
"0.0866",
"0.385",
"0.4795",
"0.385",
"0.4795",
"0.3984",
"0.0866",
"0.3984"
],
[
"0.517",
"0.3806",
"0.9126",
"0.3818",
"0.9126",
"0.3958",
"0.517",
"0.3945"
],
[
"0.0866",
"0.401",
"0.4272",
"0.401",
"0.4272",
"0.411",
"0.0866",
"0.411"
],
[
"0.5186",
"0.3958",
"0.9126",
"0.3958",
"0.9126",
"0.4092",
"0.5186",
"0.4092"
],
[
"0.5186",
"0.4104",
"0.914",
"0.4104",
"0.914",
"0.4236",
"0.5186",
"0.4236"
],
[
"0.5186",
"0.4236",
"0.7573",
"0.4236",
"0.7573",
"0.4375",
"0.5186",
"0.4375"
],
[
"0.0866",
"0.43",
"0.2443",
"0.43",
"0.2443",
"0.44",
"0.0866",
"0.44"
],
[
"0.0866",
"0.4463",
"0.4805",
"0.4463",
"0.4805",
"0.4602",
"0.0866",
"0.4602"
],
[
"0.5186",
"0.4458",
"0.9126",
"0.4458",
"0.9126",
"0.459",
"0.5186",
"0.459"
],
[
"0.08905",
"0.4622",
"0.478",
"0.4622",
"0.478",
"0.473",
"0.08905",
"0.473"
],
[
"0.517",
"0.459",
"0.9126",
"0.459",
"0.9126",
"0.473",
"0.517",
"0.473"
],
[
"0.0866",
"0.4749",
"0.4795",
"0.4749",
"0.4795",
"0.488",
"0.0866",
"0.488"
],
[
"0.517",
"0.4734",
"0.9126",
"0.4734",
"0.9126",
"0.4873",
"0.517",
"0.4873"
],
[
"0.0866",
"0.488",
"0.3162",
"0.488",
"0.3162",
"0.502",
"0.0866",
"0.502"
],
[
"0.517",
"0.486",
"0.914",
"0.4873",
"0.914",
"0.5005",
"0.517",
"0.5"
],
[
"0.5195",
"0.5024",
"0.9126",
"0.5024",
"0.9126",
"0.513",
"0.5195",
"0.513"
],
[
"0.08826",
"0.5146",
"0.1323",
"0.5146",
"0.1323",
"0.5244",
"0.08826",
"0.5244"
],
[
"0.125",
"0.512",
"0.4805",
"0.511",
"0.4805",
"0.526",
"0.125",
"0.5264"
],
[
"0.5186",
"0.515",
"0.9126",
"0.515",
"0.9126",
"0.529",
"0.5186",
"0.529"
],
[
"0.08826",
"0.528",
"0.478",
"0.528",
"0.478",
"0.5386",
"0.08826",
"0.5386"
],
[
"0.5215",
"0.531",
"0.9126",
"0.531",
"0.9126",
"0.5415",
"0.5215",
"0.5415"
],
[
"0.0866",
"0.5405",
"0.4805",
"0.5405",
"0.4805",
"0.5537",
"0.0866",
"0.5537"
],
[
"0.5215",
"0.5444",
"0.911",
"0.5444",
"0.911",
"0.5547",
"0.5215",
"0.5547"
],
[
"0.0866",
"0.5547",
"0.4805",
"0.5547",
"0.4805",
"0.5684",
"0.0866",
"0.5684"
],
[
"0.5186",
"0.556",
"0.902",
"0.556",
"0.902",
"0.5693",
"0.5186",
"0.5693"
],
[
"0.0866",
"0.5684",
"0.4805",
"0.5684",
"0.4805",
"0.582",
"0.0866",
"0.582"
],
[
"0.536",
"0.5713",
"0.9126",
"0.5713",
"0.9126",
"0.582",
"0.536",
"0.582"
],
[
"0.0866",
"0.582",
"0.4805",
"0.582",
"0.4805",
"0.595",
"0.0866",
"0.595"
],
[
"0.517",
"0.584",
"0.9126",
"0.584",
"0.9126",
"0.5977",
"0.517",
"0.5977"
],
[
"0.08905",
"0.5977",
"0.4756",
"0.5977",
"0.4756",
"0.608",
"0.08905",
"0.608"
],
[
"0.5186",
"0.5977",
"0.914",
"0.5977",
"0.914",
"0.6113",
"0.5186",
"0.6113"
],
[
"0.1054",
"0.6113",
"0.4795",
"0.6113",
"0.4795",
"0.622",
"0.1054",
"0.622"
],
[
"0.5186",
"0.6123",
"0.7686",
"0.6123",
"0.7686",
"0.6255",
"0.5186",
"0.6255"
],
[
"0.08496",
"0.6226",
"0.4805",
"0.622",
"0.4805",
"0.6353",
"0.08496",
"0.636"
],
[
"0.0866",
"0.636",
"0.4795",
"0.636",
"0.4795",
"0.6494",
"0.0866",
"0.6494"
],
[
"0.5186",
"0.64",
"0.544",
"0.64",
"0.544",
"0.6514",
"0.5186",
"0.6514"
],
[
"0.54",
"0.6426",
"0.5557",
"0.6426",
"0.5557",
"0.6494",
"0.54",
"0.6494"
],
[
"0.5605",
"0.6396",
"0.648",
"0.6396",
"0.648",
"0.654",
"0.5605",
"0.654"
],
[
"0.08905",
"0.653",
"0.4805",
"0.653",
"0.4805",
"0.6636",
"0.08905",
"0.6636"
],
[
"0.5186",
"0.6577",
"0.911",
"0.6577",
"0.911",
"0.672",
"0.5186",
"0.672"
],
[
"0.08826",
"0.6665",
"0.3506",
"0.6665",
"0.3506",
"0.677",
"0.08826",
"0.677"
],
[
"0.5186",
"0.6724",
"0.933",
"0.6724",
"0.933",
"0.683",
"0.5186",
"0.683"
],
[
"0.10297",
"0.6787",
"0.478",
"0.678",
"0.478",
"0.6914",
"0.10297",
"0.6924"
],
[
"0.5186",
"0.685",
"0.9126",
"0.685",
"0.9126",
"0.6987",
"0.5186",
"0.6987"
],
[
"0.0866",
"0.6914",
"0.4795",
"0.6924",
"0.4795",
"0.7056",
"0.0866",
"0.705"
],
[
"0.5156",
"0.6978",
"0.9126",
"0.6987",
"0.9126",
"0.712",
"0.5156",
"0.711"
],
[
"0.08905",
"0.7085",
"0.468",
"0.7085",
"0.468",
"0.7183",
"0.08905",
"0.7183"
],
[
"0.5195",
"0.712",
"0.911",
"0.7134",
"0.911",
"0.7266",
"0.5195",
"0.726"
],
[
"0.0866",
"0.72",
"0.4805",
"0.72",
"0.4805",
"0.7344",
"0.0866",
"0.7344"
],
[
"0.517",
"0.726",
"0.91",
"0.7266",
"0.91",
"0.7407",
"0.517",
"0.7393"
],
[
"0.0866",
"0.7344",
"0.4805",
"0.7344",
"0.4805",
"0.7476",
"0.0866",
"0.7476"
],
[
"0.5195",
"0.742",
"0.911",
"0.742",
"0.911",
"0.7524",
"0.5195",
"0.7524"
],
[
"0.08496",
"0.7485",
"0.1168",
"0.7485",
"0.1168",
"0.761",
"0.08496",
"0.761"
],
[
"0.5195",
"0.755",
"0.9126",
"0.755",
"0.9126",
"0.768",
"0.5195",
"0.768"
],
[
"0.0866",
"0.7715",
"0.4795",
"0.7715",
"0.4795",
"0.785",
"0.0866",
"0.785"
],
[
"0.5186",
"0.768",
"0.9126",
"0.768",
"0.9126",
"0.782",
"0.5186",
"0.782"
],
[
"0.08826",
"0.787",
"0.478",
"0.787",
"0.478",
"0.7974",
"0.08826",
"0.7974"
],
[
"0.5195",
"0.784",
"0.5786",
"0.784",
"0.5786",
"0.7944",
"0.5195",
"0.7944"
],
[
"0.08826",
"0.802",
"0.4795",
"0.802",
"0.4795",
"0.8125",
"0.08826",
"0.8125"
],
[
"0.535",
"0.7974",
"0.9126",
"0.7974",
"0.9126",
"0.808",
"0.535",
"0.808"
],
[
"0.08826",
"0.815",
"0.4795",
"0.815",
"0.4795",
"0.8257",
"0.08826",
"0.8257"
],
[
"0.517",
"0.81",
"0.876",
"0.81",
"0.876",
"0.8237",
"0.517",
"0.8237"
],
[
"0.0866",
"0.8276",
"0.482",
"0.8276",
"0.482",
"0.8413",
"0.0866",
"0.8413"
],
[
"0.5435",
"0.8276",
"0.914",
"0.8276",
"0.914",
"0.8413",
"0.5435",
"0.8413"
],
[
"0.08826",
"0.8433",
"0.2197",
"0.8433",
"0.2197",
"0.854",
"0.08826",
"0.854"
],
[
"0.5596",
"0.8433",
"0.911",
"0.8433",
"0.911",
"0.854",
"0.5596",
"0.854"
],
[
"0.10455",
"0.857",
"0.4795",
"0.857",
"0.4795",
"0.867",
"0.10455",
"0.867"
],
[
"0.558",
"0.856",
"0.671",
"0.856",
"0.671",
"0.869",
"0.558",
"0.869"
],
[
"0.0866",
"0.869",
"0.4805",
"0.869",
"0.4805",
"0.8833",
"0.0866",
"0.8833"
],
[
"0.5435",
"0.869",
"0.9126",
"0.869",
"0.9126",
"0.8833",
"0.5435",
"0.8833"
],
[
"0.08905",
"0.8853",
"0.4795",
"0.8853",
"0.4795",
"0.896",
"0.08905",
"0.896"
],
[
"0.556",
"0.884",
"0.824",
"0.884",
"0.824",
"0.898",
"0.556",
"0.898"
]
] | [
"HUGE: Huge Unsupervised Graph Embeddings with TPUs",
"KDD 23, August 6-10, 2023, Long Beach, CA, USA",
"Table 1: An overview of different approaches to scaling graph embeddings up. In this work, we demonstrate a system that",
"works without any of the yet mentioned techniques scaled to largest graphs. Scalability is given as an approximate graph size",
"that canbe processed bybest-performing algorithm & system combination in aday.",
"Method Family",
"Speedup Technique",
"Reference Methods",
"Scalability",
"neural",
"DeepWalk, LINE, VERSE",
"10M",
"graph coarsening",
"10M",
"HARP,MILE",
"neural",
"neural",
"partitioning",
"BigGraph, EDGES, DeLNE",
"1000M",
"factorization",
"NetMF, GraRep",
"10k",
"factorization",
"matrix sparsification",
"NetSMF,NetMFSC,STRAP,NRP",
"100M",
"factorization",
"implicit solvers",
"HOPE,AROPE",
"100M",
"factorization",
"spectral propagation",
"ProNE, LightNE",
"1000M",
"matrix sketching",
"FastRP, RandNE, NodeSketch, InstantEmbedding",
"1000M",
"代数技术[28]被证明是一种更有成效的方法。隐式求解器",
"factorization",
"2",
"BACKGROUND",
"在本节中，我们首先简要回顾一下国2.1节中的相关工作。我",
"[21,44]可以分解矩阵，而无需在内存中显式实现它。这些方",
"们在第2.2节中回顾了DeepWalk[22]，我们将其用作高性能",
"法被限制为执行线性分解，这不能成功地解释图的结构",
"嵌人系统的基础。然后，我们继续描述使用商品（HUGE-CPU",
"产生最可扩展的嵌入方法的两个技术家族是光谱传播和矩",
"）和TPU（HUGE-TPU）硬件扩展Deepwalk图嵌入的两种体",
"阵草图[35】。谱传播方法[2542]首先计算图的邻接或拉普拉",
"系结构，这些硬件使我们可以扩展到E2.3节中的天图",
"斯矩阵的一些截断特征分解，然后使用这些特征向量来模拟",
"信息的扩散。矩阵草绘方法近似相似矩阵，迭代[8,3743]或",
"单遍[24]。后一种选择更具可扩展性",
"2.1Related Work",
"我们现在继续回顾嵌入大图的两种基本方法。在过去的几年",
"2.1.3Hardware-basedEmbeddingAcceleration.与算法的进步",
"里，大量的工作引入了各种嵌人方法以及无数的技术和硬件",
"相比，基于硬件的加速可以说受到的关注较少。Zhuetal.[45]",
"架构来扩展它们。我们在表1中总结了嵌人方法、加速技术及",
"提出了一个混合系统，该系统使用CPU进行采样，并使用",
"其预期的可扩展性方面的相关工作。",
"GPU来训练嵌入。由于单个GPU可以提供的最大RAM约为",
"100GB，因此只能在此类系统上训练1亿个节点图的嵌入。Wei",
"2.1.1",
"GraphEmbeddingApproaches。我们将嵌入方法分类为基",
"etal.[34],Yangetal.[36]通过分区以包含更多GPU来解决此",
"于神经网络或基于矩阵分解。无论采用哪种方法，每种方法有",
"问题。这种方法需要数十个GPU来处理十亿节点的图形，与",
"时都会隐式地使用相似性函数，该函数将每个节点与图中的其",
"基于可扩展CPU的系统相比，这是令人望而却步的，后者可",
"他节点相关联。性能最好的方法从仅使用图中的邻接信息到",
"以在数小时内将十亿节点的图形嵌人到单个高内存计算机上",
"基于随机游走的相似性的某些概念，例如个性化的PageRank",
"高效计算高阶相似度是目前硬件加速不足的一个方面。",
"(PPR)E[3]。加速这些相似性计算的一个关键见解是，它们在",
"Wanget al.[33],Yanget al.[39]提出了用于通用硬件体系结",
"图[2]中高度本地化，这意味着2-3个传播步骤足以近似它们",
"构的随机游走生成的高效系统。但是，在没有合适的嵌入方法",
"神经嵌人方法将节点嵌人视为浅层神经网络的参数。神经",
"的情况下，这些系统对于图嵌人没有用",
"方法通过随机梯度下降法为邻接E[29]，随机游走国[22]或个",
"性化的PageRank[30]相似性函数优化这些参数。这种优化",
"2.2",
"",
"DeepWalk",
"是通过采样完成的，对嵌人表的更新通常非常稀疏。因此，随",
"在描述我们的TPU嵌人系统之前，有必要回顾一下Deep",
"机存储器访问通常限制了这些方法的性能",
"WalkE[22]，这是神经图嵌人的基本方法。DeepWalk使word2vecE",
"基于梯度的方法的替代方法是直接对相似性矩阵进行分解，",
"[18]（一种广泛成功的嵌人单词的模型）适应图形数据。Deep-",
"神经和矩阵分解方法之间有着深刻的联系[27,31]-本质上，对",
"Walk生成短随机游走的“语料库”；DeepWalk的目标是最大",
"于许多节点相似性，基于神经和基于分解的嵌入的最优解",
"化在某个特定窗口大小内随机游走中观察相邻顶点的后验概",
"致。基于矩阵的方法的主要挑战是保持中间表示的稀疏性。对",
"率。为了有效地最大化此概率，它使用分层的softmaxE[19]",
"于大图，既不能增加邻接矩阵的密度，也不能保留太多的中间",
"它根据节点的出现频率或计算效率更高的近似值，负采样F",
"投影",
"[10]构造节点的霍夫曼树。对于在某个节点的窗口大小内观察",
"2.1.2ScalingGraphEmbeddingSystems.加速嵌人算法有几个",
"到的每个节点，DeepWalk随机均匀地选择k≤n作为对比的",
"方向-有些是针对特定方法定制的，而有些是更通用的。现在",
"负面示例",
"我们简要回顾一下最通用的加速技术。图粗化7，17]送代地",
"DeepWalk的架构有几个计算问题，如果我们要将DeepWalk",
"收缩图，学习最压缩级别的嵌入，并在收缩层次结构中确定性",
"扩展到具有数十亿个节点的图形，则需要解决这些问题：",
"地传播嵌人。图分区方法[12,16]跨机器分布计算，同时尝试",
". Random walk generation for large graphs is computation-",
"最小化跨机器的通信",
"allyprohibitiveduetorandommemoryaccesses on each",
"早期的矩阵分解方法[26,40,41]尝试稀疏随机游走或PPR",
"random walk step.",
"矩阵。不幸的是，对于这些嵌人方法来说，高阶相似度矩阵仍",
". Random walk corpus grows in size rapidly, growing much",
"然太密集，无法扩展到非常大的图。利用专门的稀疏数值线性",
"larger in size than the original sparse graph."
] | {
"arxiv_id": "2307.14490",
"language": "zh",
"page_idx": 2
} |
||
2307.14490-zh-3 | 2307.14490-zh-3.jpg | KDD'23, August 6-10, 2023, Long Beach, CA, USA
$\bullet$ Negative sampling-based optimization is also computationally prohibitive due to random memory accesses. If batched, each gradient update is bound to update a significant part of the embedding table.
为了克服随机游走采样的困难，我们在$\overline{\mathrm{E}}$3.2 节中提出了一种分布式随机游走算法，该算法通常在 Google 上用于将随机游走模拟扩展到 Web 尺度图
# 2.3 Tensor Processing Units
我们继续简要回顾 TPU 架构，重点介绍对我们的图嵌人系统至关重要的方面。详细的评论可以在[13-15] 中找到。TPU 是专用协处理器，针对半精度计算的矩阵和向量运算进行了优化。TPU以 Pod 形式组织，3 可以连接总共 4096 个 TPU 芯片， 每个芯片具有 32GiB 内存，总共可使用 128TiB 的分布式内存。Pod 内的 TPU 芯片与以 3D 圆环拓扑组织的专用高速、低延迟互连连接
2.4 Common ML Distribution Strategies
文献中已经讨论了各种分配机器学习工作负载的方法[1] 和大多数机器学习(ML) 框架提供了通过一致接口实现多个分配方案的一致 API。本节重点介绍一些常见的分发范例，重点介绍使用商用硬件 (我们称为 HUGE-CPU) 和 TPU (HUGE-TPU) 扩展 DeepWalk 的技术。
TensorFlow 提供 tf.distribute.Strategy 抽象，使用户能够以最少的代码更改将模型创建与训练运行时环境分开。两种常见的策略是 Parameter-Server (PS)策略和 Multi-Worker Mirrored Strategy
2.4.1 $Parameter- Server$ Strategy $. 在 图 嵌 入 的 上 下 文 中 , 使 用$ PS 策略对于表示大型嵌人表很有用。PS 策略定义了用户可以访问的潜在异构硬件的两个计算池。一个池包含标记为“参数服务器”的计算机，另一个池的计算机命名为“工作者”。模型的可训练变量在参数服务器池中的机器上分片，这些机器可能通过网络为这些变量的值和更新它们提供服务。对于图嵌入，工作池中的计算机异步接收批量示例，通过网络从参数服务器获取必要的嵌入行，计算梯度并将更新推送回参数服务器计算机
2.4.2 Multi-Worker Mirrored Strategy. Multi-Worker Mirrored Strategy 在用户定义的工作机器池中的每个设备上复制模型中的所有变量。(可能)大批量的输入示例在多个工作器之间划分，并使用其较小的每个副本批次继续计算梯度。在完成单个步骤时，将聚合跨副本的梯度，并同步更新所有变量副本。虽然这可以加速计算量大的工作负载，但与参数服务器体系结构相比，此设计在图形嵌人上下文中的使用有限。跨多台计算机复制嵌入表会引人不必要的冗余和内存消耗
2.4.3 TPUStrategy and Accelerated TPU Embedding Tables. 使用TPU 硬件在 TensorFlow 中训练模型 (或图形嵌入), TPUStrategy 与 MultiWorkerMirroredStrategy 非常相似。用户定义所需的 TPU 拓扑，即 POD 的一部分，可以将其视为互连处理单元的子集。在 TPUStrategy 下，可训练变量被复制到所有 TPU 副本，大批量的示例被划分为较小的每副本批次，并分发到可用的副本，并且在同步更新之前聚合梯度。通常，这种分布范式会限制定义大型嵌入表的模型的可伸缩性。但是，TPU 能够在分配的拓扑中对所有设备上的嵌
${\overline{^3}}$In the TPUv4 architecture.
Brandon A. Mayer
人层进行分片，并利用副本之间的高带宽互连来支持加速的稀疏查找和梯度更新。加速嵌人表在 TensorFlow 中使用tf.tpu.experimental.embedding.TPUEmbedding (TPUEmbedding) 层公开，是在 TPU 上扩展 DeepWalk 训练的主要机制
# 3 мЕтнод
我们扩展 DeepWalk 算法以使用两种方法嵌人超大规模图。第一个称为 HUGE-CPU ,仅使用商品硬件，而第二个 HUGETPU 则利用现代 TPU 来提高带宽和性能收益。图 2 可视化了HUGE-CPU 的参数-服务器架构。参数服务器体系结构的详细信息在 3.3.1 节中介绍。图 3 说明了 HUGE-TPU 背后的 TPU 系统设计，详见 3.3.2 节
# 3.1 Preprocessing
一个关键的观察结果是，大多数位置图嵌人系统无法为少于两条边的节点生成有用的嵌人。具体来说，没有边的节点一般不会被嵌人算法很好地定义，同样，只有一条边的节点的位置嵌人完全由其单个邻居的嵌人决定。因此，我们通常修剪输入图，消除度小于 2 的节点。在我们的实验中，我们只修剪一次， 尽管修剪操作本身可能会引人低于度阈值的节点
# 3.2 Sampling
在对图进行预处理之后，我们运行随机游走采样以生成共现
元组，该元组将用作图嵌入系统的输入
Algorithm 1 中提供了分布式随机游走采样的高级概述。采样组件的输入是预处理图，输出是包含从随机游走中提取的同现元组的 TensorFlow 示例。利用分布式编程平台 FlumeC++[4] 实现了分布式随机游走采样算法
在初始化阶段，分布式采样器将图的$N$节点作为输人，并将它们复制 $\gamma$ 次，以创建它将生成的$\overline{\mathbb{E}}\gamma|N|$ walks 的种子 (Line$\boxed{\mathbb{F}}$ 1)。接下来，随机采样过程以迭代方式进行，执行$k$ 连接，这些连接依次增加每个随机游走的长度(行$\overline{\mathrm{E}}$2-4)。每个联接在其端点处将漫游与节点组合在一起。$^{4}$在将行走的末端与来自图$G$的其相应节点连接之后，发生下一个节点的采样(第 4行)。我们注意到，在此步骤中可以使用多种采样来选择下一个节点-包括均匀采样，带有回溯的随机游走以及其他形式的加权采样。对于本文的结果，我们考虑使用均匀抽样的情况。最终的 GroupBy 操作用于根据访问距离 (第 7 行) 来折叠随机走到节点对之间的同现计数
采样预处理步骤的输出是编码三元组的文件的分片系列：(source\_id, destination\_id, co\_counts)。source\_id 是随机游走中起点的节点 ID , destination\_id 是在$\gamma$随机游走期间到达的节点 ID , co\_counts 是长度为 walk\_length 的直方图，其中包含 source\_id 遇到
的次数
DeepWalk 模型定义了具有“正”和“负”分量的图重建损失。“正”示例是原始图中存在的随机行走路径。“否定”示例是原始图中不存在的路径。如果需要，采样步骤可用于生成不同种类的负样本 (通过关注不存在的边缘的附加分布式采样算法)。然而，在实践中，我们经常更喜欢在训练时“即时”执行近似随机负采样
# 3.3 Distributed training
$\overline{^4\text{This join is neces}}$sary, as the system must support graphs which are too large to fit
in the memory of a single machine | [
[
"0.08826",
"0.07764",
"0.3245",
"0.07764",
"0.3245",
"0.0884",
"0.08826",
"0.0884"
],
[
"0.8228",
"0.077",
"0.911",
"0.077",
"0.911",
"0.0871",
"0.8228",
"0.0871"
],
[
"0.1136",
"0.108",
"0.4805",
"0.108",
"0.4805",
"0.1218",
"0.1136",
"0.1218"
],
[
"0.5195",
"0.108",
"0.9126",
"0.108",
"0.9126",
"0.1218",
"0.5195",
"0.1218"
],
[
"0.1259",
"0.1218",
"0.4805",
"0.1218",
"0.4805",
"0.1364",
"0.1259",
"0.1364"
],
[
"0.5186",
"0.1218",
"0.9087",
"0.1218",
"0.9087",
"0.1351",
"0.5186",
"0.1351"
],
[
"0.1259",
"0.1351",
"0.4805",
"0.1351",
"0.4805",
"0.1483",
"0.1259",
"0.1483"
],
[
"0.5186",
"0.1364",
"0.9126",
"0.1364",
"0.9126",
"0.1497",
"0.5186",
"0.1497"
],
[
"0.1274",
"0.1515",
"0.2656",
"0.1515",
"0.2656",
"0.1622",
"0.1274",
"0.1622"
],
[
"0.5186",
"0.1483",
"0.859",
"0.1483",
"0.859",
"0.1622",
"0.5186",
"0.1622"
],
[
"0.10455",
"0.1654",
"0.4739",
"0.1654",
"0.4739",
"0.1787",
"0.10455",
"0.1787"
],
[
"0.0866",
"0.1812",
"0.4805",
"0.1812",
"0.4805",
"0.1945",
"0.0866",
"0.1945"
],
[
"0.5156",
"0.1768",
"0.63",
"0.1768",
"0.63",
"0.1901",
"0.5156",
"0.1901"
],
[
"0.0866",
"0.1945",
"0.2688",
"0.1945",
"0.2688",
"0.2084",
"0.0866",
"0.2084"
],
[
"0.5195",
"0.1963",
"0.911",
"0.1963",
"0.911",
"0.207",
"0.5195",
"0.207"
],
[
"0.527",
"0.209",
"0.914",
"0.209",
"0.914",
"0.2229",
"0.527",
"0.2229"
],
[
"0.08496",
"0.2216",
"0.327",
"0.2216",
"0.327",
"0.2349",
"0.08496",
"0.2349"
],
[
"0.5186",
"0.2229",
"0.9126",
"0.2229",
"0.9126",
"0.2361",
"0.5186",
"0.2361"
],
[
"0.0866",
"0.2393",
"0.4805",
"0.2393",
"0.4805",
"0.2524",
"0.0866",
"0.2524"
],
[
"0.5186",
"0.2361",
"0.9126",
"0.2361",
"0.9126",
"0.25",
"0.5186",
"0.25"
],
[
"0.08826",
"0.2524",
"0.4805",
"0.2524",
"0.4805",
"0.2664",
"0.08826",
"0.2664"
],
[
"0.5186",
"0.2507",
"0.9126",
"0.2507",
"0.9126",
"0.2644",
"0.5186",
"0.2644"
],
[
"0.08826",
"0.2683",
"0.4795",
"0.2683",
"0.4795",
"0.279",
"0.08826",
"0.279"
],
[
"0.517",
"0.2664",
"0.6562",
"0.2664",
"0.6562",
"0.2766",
"0.517",
"0.2766"
],
[
"0.0866",
"0.2798",
"0.4805",
"0.2798",
"0.4805",
"0.2935",
"0.0866",
"0.2935"
],
[
"0.0866",
"0.2942",
"0.4805",
"0.2942",
"0.4805",
"0.308",
"0.0866",
"0.308"
],
[
"0.5195",
"0.2942",
"0.5435",
"0.2942",
"0.5435",
"0.303",
"0.5195",
"0.303"
],
[
"0.5376",
"0.2917",
"0.6772",
"0.2935",
"0.6772",
"0.307",
"0.5376",
"0.305"
],
[
"0.0866",
"0.31",
"0.4795",
"0.31",
"0.4795",
"0.3208",
"0.0866",
"0.3208"
],
[
"0.5195",
"0.308",
"0.9126",
"0.3093",
"0.9126",
"0.3225",
"0.5195",
"0.3213"
],
[
"0.0866",
"0.324",
"0.1462",
"0.324",
"0.1462",
"0.334",
"0.0866",
"0.334"
],
[
"0.5195",
"0.3257",
"0.9126",
"0.3257",
"0.9126",
"0.336",
"0.5195",
"0.336"
],
[
"0.5195",
"0.3384",
"0.914",
"0.3384",
"0.914",
"0.3516",
"0.5195",
"0.3516"
],
[
"0.08417",
"0.3484",
"0.4248",
"0.3484",
"0.4248",
"0.3623",
"0.08417",
"0.3623"
],
[
"0.517",
"0.351",
"0.9126",
"0.3516",
"0.9126",
"0.3655",
"0.517",
"0.3643"
],
[
"0.08905",
"0.3687",
"0.4795",
"0.3687",
"0.4795",
"0.379",
"0.08905",
"0.379"
],
[
"0.5195",
"0.3674",
"0.911",
"0.3674",
"0.911",
"0.3774",
"0.5195",
"0.3774"
],
[
"0.08826",
"0.3818",
"0.4795",
"0.3818",
"0.4795",
"0.3926",
"0.08826",
"0.3926"
],
[
"0.5186",
"0.3801",
"0.8267",
"0.3801",
"0.8267",
"0.3933",
"0.5186",
"0.3933"
],
[
"0.08826",
"0.3945",
"0.4795",
"0.3945",
"0.4795",
"0.4077",
"0.08826",
"0.4077"
],
[
"0.5186",
"0.4077",
"0.562",
"0.4077",
"0.562",
"0.4192",
"0.5186",
"0.4192"
],
[
"0.5537",
"0.406",
"0.642",
"0.4077",
"0.642",
"0.4224",
"0.5537",
"0.4204"
],
[
"0.0866",
"0.4236",
"0.2361",
"0.4236",
"0.2361",
"0.4343",
"0.0866",
"0.4343"
],
[
"0.08826",
"0.4104",
"0.4795",
"0.4104",
"0.4795",
"0.4204",
"0.08826",
"0.4204"
],
[
"0.5195",
"0.4268",
"0.911",
"0.4268",
"0.911",
"0.4375",
"0.5195",
"0.4375"
],
[
"0.10297",
"0.4363",
"0.4805",
"0.4363",
"0.4805",
"0.4495",
"0.10297",
"0.4495"
],
[
"0.5195",
"0.44",
"0.7686",
"0.44",
"0.7686",
"0.4507",
"0.5195",
"0.4507"
],
[
"0.0866",
"0.4507",
"0.4805",
"0.4507",
"0.4805",
"0.464",
"0.0866",
"0.464"
],
[
"0.535",
"0.4526",
"0.914",
"0.4526",
"0.914",
"0.4666",
"0.535",
"0.4666"
],
[
"0.0866",
"0.464",
"0.482",
"0.464",
"0.482",
"0.4778",
"0.0866",
"0.4778"
],
[
"0.517",
"0.4653",
"0.9126",
"0.4673",
"0.9126",
"0.481",
"0.517",
"0.4792"
],
[
"0.08496",
"0.4778",
"0.1953",
"0.4797",
"0.1953",
"0.4912",
"0.08496",
"0.4893"
],
[
"0.5186",
"0.4797",
"0.914",
"0.4797",
"0.914",
"0.4937",
"0.5186",
"0.4937"
],
[
"0.1259",
"0.5",
"0.4795",
"0.5",
"0.4795",
"0.513",
"0.1259",
"0.513"
],
[
"0.5195",
"0.4956",
"0.722",
"0.4956",
"0.722",
"0.506",
"0.5195",
"0.506"
],
[
"0.535",
"0.5083",
"0.9126",
"0.5083",
"0.9126",
"0.5215",
"0.535",
"0.5215"
],
[
"0.08826",
"0.5005",
"0.1323",
"0.5005",
"0.1323",
"0.511",
"0.08826",
"0.511"
],
[
"0.08496",
"0.513",
"0.4805",
"0.513",
"0.4805",
"0.5264",
"0.08496",
"0.5264"
],
[
"0.5186",
"0.5215",
"0.914",
"0.5215",
"0.914",
"0.535",
"0.5186",
"0.535"
],
[
"0.08826",
"0.528",
"0.4805",
"0.528",
"0.4805",
"0.5386",
"0.08826",
"0.5386"
],
[
"0.517",
"0.534",
"0.911",
"0.535",
"0.911",
"0.549",
"0.517",
"0.5474"
],
[
"0.0866",
"0.5405",
"0.482",
"0.5405",
"0.482",
"0.5537",
"0.0866",
"0.5537"
],
[
"0.517",
"0.5474",
"0.9126",
"0.55",
"0.9126",
"0.563",
"0.517",
"0.561"
],
[
"0.08905",
"0.556",
"0.4805",
"0.556",
"0.4805",
"0.5664",
"0.08905",
"0.5664"
],
[
"0.5186",
"0.563",
"0.9126",
"0.563",
"0.9126",
"0.577",
"0.5186",
"0.577"
],
[
"0.08905",
"0.5693",
"0.4795",
"0.5693",
"0.4795",
"0.58",
"0.08905",
"0.58"
],
[
"0.0866",
"0.5806",
"0.4805",
"0.5806",
"0.4805",
"0.5947",
"0.0866",
"0.5947"
],
[
"0.5186",
"0.577",
"0.914",
"0.577",
"0.914",
"0.5903",
"0.5186",
"0.5903"
],
[
"0.0866",
"0.5967",
"0.482",
"0.5967",
"0.482",
"0.61",
"0.0866",
"0.61"
],
[
"0.5186",
"0.5913",
"0.899",
"0.5913",
"0.899",
"0.605",
"0.5186",
"0.605"
],
[
"0.0866",
"0.61",
"0.1609",
"0.6113",
"0.1609",
"0.6226",
"0.0866",
"0.622"
],
[
"0.5195",
"0.605",
"0.914",
"0.605",
"0.914",
"0.6187",
"0.5195",
"0.6187"
],
[
"0.5195",
"0.6206",
"0.911",
"0.6206",
"0.911",
"0.631",
"0.5195",
"0.631"
],
[
"0.08496",
"0.631",
"0.482",
"0.63",
"0.482",
"0.6445",
"0.08496",
"0.646"
],
[
"0.5186",
"0.6333",
"0.914",
"0.6333",
"0.914",
"0.6465",
"0.5186",
"0.6465"
],
[
"0.08496",
"0.646",
"0.4795",
"0.6445",
"0.4795",
"0.6577",
"0.08496",
"0.659"
],
[
"0.5186",
"0.646",
"0.697",
"0.646",
"0.697",
"0.659",
"0.5186",
"0.659"
],
[
"0.0866",
"0.659",
"0.4805",
"0.659",
"0.4805",
"0.6724",
"0.0866",
"0.6724"
],
[
"0.536",
"0.662",
"0.911",
"0.662",
"0.911",
"0.6724",
"0.536",
"0.6724"
],
[
"0.08826",
"0.675",
"0.4795",
"0.675",
"0.4795",
"0.685",
"0.08826",
"0.685"
],
[
"0.5186",
"0.675",
"0.911",
"0.6753",
"0.911",
"0.6875",
"0.5186",
"0.686"
],
[
"0.08905",
"0.688",
"0.4795",
"0.688",
"0.4795",
"0.6987",
"0.08905",
"0.6987"
],
[
"0.5215",
"0.6895",
"0.9126",
"0.6895",
"0.9126",
"0.6997",
"0.5215",
"0.6997"
],
[
"0.0866",
"0.7007",
"0.4795",
"0.7007",
"0.4795",
"0.714",
"0.0866",
"0.714"
],
[
"0.5186",
"0.702",
"0.9126",
"0.702",
"0.9126",
"0.7153",
"0.5186",
"0.7153"
],
[
"0.0866",
"0.714",
"0.4795",
"0.714",
"0.4795",
"0.728",
"0.0866",
"0.728"
],
[
"0.5186",
"0.7153",
"0.7236",
"0.7153",
"0.7236",
"0.729",
"0.5186",
"0.729"
],
[
"0.0866",
"0.729",
"0.3945",
"0.729",
"0.3945",
"0.7393",
"0.0866",
"0.7393"
],
[
"0.5195",
"0.731",
"0.564",
"0.731",
"0.564",
"0.742",
"0.5195",
"0.742"
],
[
"0.5317",
"0.742",
"0.9126",
"0.742",
"0.9126",
"0.756",
"0.5317",
"0.756"
],
[
"0.0866",
"0.75",
"0.4795",
"0.75",
"0.4795",
"0.763",
"0.0866",
"0.763"
],
[
"0.5186",
"0.757",
"0.911",
"0.7583",
"0.911",
"0.7695",
"0.5186",
"0.768"
],
[
"0.0866",
"0.763",
"0.4805",
"0.763",
"0.4805",
"0.7764",
"0.0866",
"0.7764"
],
[
"0.08496",
"0.7764",
"0.4805",
"0.776",
"0.4805",
"0.7905",
"0.08496",
"0.7915"
],
[
"0.5195",
"0.771",
"0.9126",
"0.771",
"0.9126",
"0.784",
"0.5195",
"0.784"
],
[
"0.0866",
"0.789",
"0.4805",
"0.7905",
"0.4805",
"0.8047",
"0.0866",
"0.8037"
],
[
"0.5186",
"0.784",
"0.9126",
"0.784",
"0.9126",
"0.7974",
"0.5186",
"0.7974"
],
[
"0.5186",
"0.799",
"0.9126",
"0.799",
"0.9126",
"0.8125",
"0.5186",
"0.8125"
],
[
"0.0866",
"0.8047",
"0.4805",
"0.8047",
"0.4805",
"0.8184",
"0.0866",
"0.8184"
],
[
"0.0866",
"0.8184",
"0.482",
"0.8184",
"0.482",
"0.832",
"0.0866",
"0.832"
],
[
"0.517",
"0.813",
"0.638",
"0.813",
"0.638",
"0.827",
"0.517",
"0.827"
],
[
"0.0866",
"0.8335",
"0.4795",
"0.834",
"0.4795",
"0.845",
"0.0866",
"0.8447"
],
[
"0.517",
"0.8413",
"0.544",
"0.8413",
"0.544",
"0.853",
"0.517",
"0.853"
],
[
"0.54",
"0.8433",
"0.556",
"0.8433",
"0.556",
"0.851",
"0.54",
"0.851"
],
[
"0.5605",
"0.839",
"0.728",
"0.8403",
"0.728",
"0.855",
"0.5605",
"0.854"
],
[
"0.08826",
"0.8486",
"0.4795",
"0.8486",
"0.4795",
"0.8594",
"0.08826",
"0.8594"
],
[
"0.08826",
"0.8623",
"0.4795",
"0.8623",
"0.4795",
"0.8726",
"0.08826",
"0.8726"
],
[
"0.5156",
"0.8706",
"0.9126",
"0.8726",
"0.9126",
"0.887",
"0.5156",
"0.8853"
],
[
"0.0866",
"0.882",
"0.2141",
"0.884",
"0.2141",
"0.898",
"0.0866",
"0.896"
],
[
"0.5186",
"0.886",
"0.683",
"0.886",
"0.683",
"0.8965",
"0.5186",
"0.8965"
]
] | [
"KDD '23, August 6-10, 2023, Long Beach, CA, USA",
"Brandon A. Maye",
"· Negative sampling-based optimization is also computation",
"入层进行分片，并利用副本之间的高带宽互连来支持加速",
"ally prohibitive due to random memory accesses. If batched,",
"的稀疏查找和梯度更新。加速嵌人表在TensorFlow中使用",
"each gradient update is bound to update a significant part",
"tf.tpu.experimental.embedding.TPUEmbedding(TPUEmbedding)",
"of theembeddingtable",
"层公开，是在TPU上扩展DeepWalk训练的主要机制",
"为了克服随机游走采样的困难，我们在国3.2节中提出了",
"种分布式随机游走算法，该算法通常在Google上用于将随机",
"3METHOD",
"游走模拟扩展到Web尺度图",
"我们扩展DeepWalk算法以使用两种方法嵌超大规模图。第",
"一个称为HUGE-CPU，仅使用商品硬件，而第二个HUGE-",
"2.3TensorProcessing Units",
"TPU则利用现代TPU来提高带宽和性能收益。图2可视化了",
"我们继续简要回顾TPU架构，重点介绍对我们的图嵌入系统",
"HUGE-CPU的参数-服务器架构。参数服务器体系结构的详细",
"至关重要的方面。详细的评论可以在[13-15]中找到。TPU是",
"信息在3.3.1节中介绍。图3说明了HUGE-TPU背后的TPU系",
"专用协处理器，针对半精度计算的矩阵和向量运算进行了优",
"统设计：详见3.3.2节",
"化。TPU以Pod形式组织，3可以连接总共4096个TPU芯片",
"每个芯片具有32GiB内存，总共可使用128TiB的分布式内存。",
"3.1",
"1Preprocessing",
"Pod内的TPU芯片与以3D圆环拓扑组织的专用高速、低延迟",
"一个关键的观察结果是，大多数位置图嵌人系统无法为少于",
"互连连接",
"两条边的节点生成有用的嵌入。具体来说，没有边的节点一般",
"不会被嵌人算法很好地定义，同样，只有一条边的节点的位置",
"2.4Common MLDistribution Strategies",
"嵌人完全由其单个邻居的嵌人决定。因此，我们通常修剪输入",
"文献中已经讨论厂各种分配机器学习工作负载的方法1]和大",
"图，消除度小于2的节点。在我们的实验中，我们只修剪一次，",
"多数机器学习（ML）框架提供了通过一致接口实现多个分配",
"尽管修剪操作本身可能会引人低于度阈值的节点",
"方案的一致API。本节重点介绍一些常见的分发范例，重点介",
"3.2",
"Sampling",
"扩展DeepWalk的技术。",
"绍使用商用硬件（我们称为HUGE-CPU）和TPU（HUGE-TPU）",
"在对图进行预处理之后，我们运行随机游走采样以生成共现",
"TensorFlow提供tf.distribute.Strategy抽象，使用户能够以",
"元组，该元组将用作图嵌入系统的输人",
"最少的代码更改将模型创建与训练运行时环境分开。两种",
"Algorithm1中提供了分布式随机游走采样的高级概述。采",
"常见的策略是Parameter-Server（PS）策略和Multi-Worker",
"样组件的输入是预处理图，输出是包含从随机游走中提取的同",
"MirroredStrategy",
"现元组的TensorFlow示例。利用分布式编程平台FlumeC++[4]",
"Parameter-ServerStrategy，在图嵌人的上下文中，使用",
"实现了分布式随机游走采样算法",
"在初始化阶段，分布式采样器将图的N节点作为输入，并将",
"2.4.1",
"PS策略对于表示大型嵌入表很有用。PS策略定义了用户可以",
"它们复制y次，以创建它将生成的EylMIwalks的种子（LineE",
"访问的潜在异构硬件的两个计算池。一个池包含标记为“参数",
"1)。接下来，随机采样过程以迭代方式进行，执行k连接，这",
"服务器”的计算机，另一个池的计算机命名为“工作者”。模",
"些连接依次增加每个随机游走的长度（行E2-4）。每个联接在",
"型的可训练变量在参数服务器池中的机器上分片，这些机器",
"其端点处将漫游与节点组合在一起。4在将行走的末端与来自",
"可能通过网络为这些变量的值和更新它们提供服务。对于图",
"嵌入，工作池中的计算机异步接收批量示例，通过网络从参数",
"图G的其相应节点连接之后，发生下一个节点的采样（第4",
"服务器获取必要的嵌人行，计算梯度并将更新推送回参数服",
"行）。我们注意到，在此步骤中可以使用多种采样来选择下",
"务器计算机",
"个节点-包括均匀采样，带有回溯的随机游走以及其他形式的",
"加权采样。对于本文的结果，我们考虑使用均匀抽样的情况。",
"2.4.2Multi-Worker Mirrored StrategyMulti-Worker Mirrored",
"最终的GroupBy操作用于根据访问距离（第7行）来折叠随机",
"Strategy在用户定义的工作机器池中的每个设备上复制模型",
"走到节点对之间的同现计数",
"中的所有变量。（可能）大批量的输入示例在多个工作器之间",
"采样预处理步骤的输出是编码三元组的文件的分片系",
"划分，并使用其较小的每个副本批次继续计算梯度。在完成单",
"列:(source_id，destination_id，co_counts)。source_id",
"个步骤时，将聚合跨副本的梯度，并同步更新所有变量副本。",
"是随机游走中起点的节点ID，destination_id是在y随机游",
"虽然这可以加速计算量大的工作负载，但与参数服务器体系",
"走期间到达的节点ID，co_counts是长度为walk_length的直",
"结构相比，此设计在图形嵌入上下文中的使用有限。跨多台计",
"方图，其中包含source_id遇到",
"算机复制嵌入表会引人不必要的余和内存消耗",
"的次数",
"DeepWalk模型定义了具有“正”和“负”分量的图重建损",
"2.4.3TPUStrategyandAcceleratedTPUEmbeddingTables.使用",
"失。“正”示例是原始图中存在的随机行走路径。“否定”示例",
"TPU硬件在TensorFlow中训练模型（或图形嵌入），TPUS-",
"trategy与MultiWorkerMirroredStrategy非常相似。用户定",
"是原始图中不存在的路径。如果需要，采样步骤可用于生成不",
"义所需的TPU拓扑，即POD的一部分，可以将其视为互",
"同种类的负样本（通过关注不存在的边缘的附加分布式采样",
"算法)。然而，在实践中，我们经常更喜欢在训练时“即时”执",
"连处理单元的子集。在TPUStrategy下，可训练变量被复制",
"到所有TPU副本，大批量的示例被划分为较小的每副本批",
"行近似随机负采样",
"次，并分发到可用的副本，并且在同步更新之前聚合梯度。",
"3.3",
"",
"Distributed training",
"通常，这种分布范式会限制定义大型嵌入表的模型的可伸",
"缩性。但是，TPU能够在分配的拓扑中对所有设备上的嵌",
"4This join is necessary, as the system must support graphs which are toolargetoft",
"3In the TPUv4 architecture",
"in the memory of a single machine."
] | {
"arxiv_id": "2307.14490",
"language": "zh",
"page_idx": 3
} |
||
2104.11939-zh-1 | 2104.11939-zh-1.jpg | Piggyback GAN: Efficient Lifelong Learning for
Image Conditioned Generation
Mengyao Zhai, Lei Chen, Jiawei He, Megha Nawhal,
Frederick Tung, and Greg Mori
No Institute Given
$\mathbf{Abstract.~Humans~accumulate~knowledge~in~a~lifelong~fashion.~Modern}$ deep neural networks, on the other hand, are susceptible to catastrophic forgetting: when adapted to perform new tasks, they often fail to preserve their performance on previously learned tasks. Given a sequence of tasks, a naive approach addressing catastrophic forgetting is to train a separate standalone model for each task, which scales the total number of parameters drastically without efficiently utilizing previous models. In contrast, we propose a parameter efficient framework, Piggyback GAN, which learns the current task by building a set of convolutional and deconvolutional filters that are factorized into filters of the models trained on previous tasks. For the current task, our model achieves high generation quality on par with a standalone model at a lower number of parameters. For previous tasks, our model can also preserve generation quality since the filters for previous tasks are not altered. We validate Piggyback GAN on various image-conditioned generation tasks across different domains, and provide qualitative and quantitative results to show that the proposed approach can address catastrophic forgetting effectively and efficiently.
Keywords: Lifelong Learning, Generative Adversarial Networks.
## 1 Introduction
人类是终身学习者：为了每天有效地发挥作用，我们一生都在获取和积累知识。积累的知识使我们在遇到新任务时变得高效和多才多艺。与人类学习相反，现代基于神经网络的学习算法在适应新任务时通常无法记住从先前任务中获得的知识(见图 ??(b))。换句话说，一旦在任务上训练了模型，就很难进行泛化。这是众所周知的灾难性遗忘现象[25] 。最近的努力[32,20,6,2]已经证明了判别模型如何连续学习一系列任务。尽管这些判别模型的努力取得了成功，但生成模型的终身学习仍然是一个具有挑战性和探索不足的领域
判别模型的终身学习方法由于其内在差异而无法直接应用于生成模型。首先， 众所周知，在分类任务中，深度神经网络中的中间卷积层能够提供通用特征。这些功能可以很容易地被具有不同分类目标的其他模型重用。对于生成模型，据我们所知，以前没有探索过将这种重用用于新的生成任务的可能性。第二，与判别模型不同，生成模型的输出空间通常是连续的，使得模型沿着一系列任务的训练保持生成质量更具挑战性。第三，在生成式设置下，任务之间可能存在冲 | [
[
"0.2279",
"0.1459",
"0.777",
"0.1459",
"0.777",
"0.1635",
"0.2279",
"0.1635"
],
[
"0.3228",
"0.1692",
"0.6787",
"0.1692",
"0.6787",
"0.1843",
"0.3228",
"0.1843"
],
[
"0.3145",
"0.214",
"0.6895",
"0.214",
"0.6895",
"0.2291",
"0.3145",
"0.2291"
],
[
"0.3896",
"0.2299",
"0.615",
"0.2299",
"0.615",
"0.2443",
"0.3896",
"0.2443"
],
[
"0.4387",
"0.2568",
"0.564",
"0.2568",
"0.564",
"0.2708",
"0.4387",
"0.2708"
],
[
"0.2656",
"0.3093",
"0.739",
"0.3093",
"0.739",
"0.3225",
"0.2656",
"0.3225"
],
[
"0.2656",
"0.3225",
"0.74",
"0.3225",
"0.74",
"0.336",
"0.2656",
"0.336"
],
[
"0.2646",
"0.3372",
"0.738",
"0.3372",
"0.738",
"0.351",
"0.2646",
"0.351"
],
[
"0.267",
"0.3528",
"0.739",
"0.3528",
"0.739",
"0.363",
"0.267",
"0.363"
],
[
"0.2632",
"0.3643",
"0.739",
"0.363",
"0.739",
"0.3774",
"0.2632",
"0.379"
],
[
"0.2646",
"0.379",
"0.739",
"0.3774",
"0.739",
"0.3914",
"0.2646",
"0.3926"
],
[
"0.2646",
"0.3914",
"0.738",
"0.3914",
"0.738",
"0.4048",
"0.2646",
"0.4048"
],
[
"0.2646",
"0.406",
"0.739",
"0.4048",
"0.739",
"0.4192",
"0.2646",
"0.4204"
],
[
"0.2656",
"0.4204",
"0.7363",
"0.4204",
"0.7363",
"0.4343",
"0.2656",
"0.4343"
],
[
"0.2656",
"0.4343",
"0.739",
"0.4343",
"0.739",
"0.4475",
"0.2656",
"0.4475"
],
[
"0.2646",
"0.4463",
"0.7363",
"0.4475",
"0.7363",
"0.4622",
"0.2646",
"0.461"
],
[
"0.2656",
"0.4622",
"0.74",
"0.4622",
"0.74",
"0.476",
"0.2656",
"0.476"
],
[
"0.2646",
"0.476",
"0.738",
"0.4749",
"0.738",
"0.488",
"0.2646",
"0.4893"
],
[
"0.2656",
"0.488",
"0.74",
"0.488",
"0.74",
"0.5024",
"0.2656",
"0.5024"
],
[
"0.2646",
"0.5024",
"0.74",
"0.5024",
"0.74",
"0.5166",
"0.2646",
"0.5166"
],
[
"0.2656",
"0.5176",
"0.74",
"0.5176",
"0.74",
"0.531",
"0.2656",
"0.531"
],
[
"0.2646",
"0.53",
"0.739",
"0.53",
"0.739",
"0.5435",
"0.2646",
"0.5435"
],
[
"0.2646",
"0.5444",
"0.4307",
"0.5444",
"0.4307",
"0.558",
"0.2646",
"0.558"
],
[
"0.2656",
"0.5728",
"0.697",
"0.5728",
"0.697",
"0.586",
"0.2656",
"0.586"
],
[
"0.2141",
"0.61",
"0.38",
"0.6113",
"0.38",
"0.629",
"0.2141",
"0.6274"
],
[
"0.2214",
"0.646",
"0.7837",
"0.646",
"0.7837",
"0.659",
"0.2214",
"0.659"
],
[
"0.2214",
"0.662",
"0.7837",
"0.662",
"0.7837",
"0.6724",
"0.2214",
"0.6724"
],
[
"0.2214",
"0.678",
"0.7837",
"0.678",
"0.7837",
"0.688",
"0.2214",
"0.688"
],
[
"0.2197",
"0.6914",
"0.7837",
"0.6914",
"0.7837",
"0.705",
"0.2197",
"0.705"
],
[
"0.2197",
"0.7056",
"0.786",
"0.7056",
"0.786",
"0.7197",
"0.2197",
"0.7197"
],
[
"0.2197",
"0.72",
"0.785",
"0.72",
"0.785",
"0.7344",
"0.2197",
"0.7344"
],
[
"0.2214",
"0.7373",
"0.642",
"0.7373",
"0.642",
"0.7476",
"0.2214",
"0.7476"
],
[
"0.2427",
"0.75",
"0.7837",
"0.7505",
"0.7837",
"0.765",
"0.2427",
"0.7646"
],
[
"0.2214",
"0.7666",
"0.785",
"0.7666",
"0.785",
"0.78",
"0.2214",
"0.78"
],
[
"0.2181",
"0.78",
"0.785",
"0.781",
"0.785",
"0.7944",
"0.2181",
"0.7935"
],
[
"0.2197",
"0.797",
"0.786",
"0.797",
"0.786",
"0.81",
"0.2197",
"0.81"
],
[
"0.2214",
"0.8125",
"0.7837",
"0.8125",
"0.7837",
"0.8257",
"0.2214",
"0.8257"
],
[
"0.2197",
"0.827",
"0.785",
"0.827",
"0.785",
"0.8403",
"0.2197",
"0.8403"
]
] | [
"Piggyback GAN: Efficient Lifelong Learning for",
"Image Conditioned Generation",
"Mengyao Zhai, Lei Chen, Jiawei He, Megha Nawhal,",
"Frederick Tung, and Greg Mori",
"No Institute Given",
"Abstract. Humans accumulate knowledge in a lifelong fashion. Modern",
"deep neural networks, on the other hand, are susceptible to catastrophic",
"forgetting: when adapted to perform new tasks, they often fail to pre-",
"servetheirperformance onpreviouslylearned tasks.Given asequence of",
"tasks, a naive approach addressing catastrophic forgetting is to train a",
"separate standalone model for each task, which scales the total number",
"of parameters drastically without efficiently utilizing previous models. In",
"contrast, we propose a parameter efficient framework, Piggyback GAN,",
"which learns the current task by building a set of convolutional and de",
"convolutional filters that are factorized into filters of the models trained",
"on previous tasks. For the current task, our model achieves high gen",
"eration quality on par with a standalone model at a lower number of",
"parameters. For previous tasks, our model can also preserve generation",
"quality since the filters for previous tasks are not altered. We validate",
"Piggyback GAN on various image-conditioned generation tasks across",
"different domains, and provide qualitative and quantitative results to",
"show that the proposed approach can address catastrophic forgetting",
"effectively and efficiently",
"Keywords: Lifelong Learning, Generative Adversarial Networks",
"1Introduction",
"人类是终身学习者：为了每天有效地发挥作用，我们一生都在获取和积累知识。",
"积累的知识使我们在遇到新任务时变得高效和多才多艺。与人类学习相反，现",
"代基于神经网络的学习算法在适应新任务时通常无法记住从先前任务中获得的",
"知识（见图？？（b））。换句话说，一旦在任务上训练了模型，就很难进行泛化。",
"这是众所周知的灾难性遗忘现象[25]。最近的努力[32,20,6,2]已经证明了判",
"别模型如何连续学习一系列任务。尽管这些判别模型的努力取得了成功，但生",
"成模型的终身学习仍然是一个具有挑战性和探索不足的领域",
"判别模型的终身学习方法由于其内在差异而无法直接应用于生成模型。首先",
"众所周知，在分类任务中，深度神经网络中的中间卷积层能够提供通用特征。这",
"些功能可以很容易地被具有不同分类目标的其他模型重用。对于生成模型，据",
"我们所知，以前没有探索过将这种重用用于新的生成任务的可能性。第二，与判",
"别模型不同，生成模型的输出空间通常是连续的，使得模型沿着一系列任务的",
"训练保持生成质量更具挑战性。第三，在生成式设置下，任务之间可能存在冲"
] | {
"arxiv_id": "2104.11939",
"language": "zh",
"page_idx": 1
} |
||
2104.11939-zh-2 | 2104.11939-zh-2.jpg | Piggyback GAN
突。对于判别模型，很少发生一个图像具有不同的标签(出现在不同的任务中)。但是，对于生成模型，我们希望将相同的图像转换为不同的域以执行不同的任务是很常见的$\mathbb{E}[14,11,44,8]$
解决生成模型的灾难性遗忘的一种天真的方法是分别为每个任务训练一个模型(见图？? (c))。不幸的是，这种方法通常不可扩展：随着新任务的添加，存储需求急剧增长。更重要的是，将经过训练的模型放在一边，而不利用它在面对新任务时可能提供的好处，这将是对资源的低效利用。最近的工作$\mathbb{E}[42,37]$ 在生成模型的终身学习方面显示出令人鼓舞的结果，但也揭示了图像质量下降和伪影从旧任务转移到新任务是不可避免的。因此，设计一个持续学习框架是有价值的，该框架(1)具有更高的参数效率，(2)保留当前和先前学习任务的生成质量，以及(3)可以跨不同域启用各种条件生成任务
在本文中，我们介绍了一个通用的持续学习框架 PiggybackGAN,该框架可以跨不同域执行各种条件图像生成任务(请参见图EF？? ( d) )。所提出的方法解决了生成连续学习模型中的灾难性遗忘问题。具体来说，PiggybackGAN 维护一个滤波器组，其中的滤波器来自在先前任务中训练的模型的卷积和反卷积层。面对新任务，PiggybackGAN 通过构建一组 piggyback 滤波器来重用滤波器组中的滤波器，从而学习执行此任务，这些滤波器被分解为滤波器组中的滤波器。PiggybackGAN 还为每个任务维护一小部分不受约束的过滤器，以确保高质量的生成。学习新任务后，将无约束过滤器附加到过滤器组以促进后续任务的训练。由于旧任务的过滤器没有被改变，PiggybackGAN 能够保持以前任务的确切知识，而不会忘记任何细节
总结一下，我们的贡献如下。我们提出了一种用于条件图像生成的持续学习框架，该框架是(1)高效的，通过在先前任务上训练的模型上“piggybacking’ 并重用先前模型中的过滤器来学习执行新任务，(2)保持质量，保持当前任务的生成质量，并确保先前任务的质量不会下降，同时节省更多参数，以及(3)通用的，可跨不同域启用各种条件生成任务。据我们所知，我们是第一个为生成持续学习模型做出这些贡献的人。我们在两种设置下验证了我们的方法的有效性：(1)成对的图像条件生成，以及(2)不成对的图像条件生成。与最先进的模型进行了广泛的定性和定量比较，以说明我们提出的框架学习新一代任务的能力和效率，而不会灾难性地忘记以前的任务。
# 2 Related Work
我们的工作与两个以前不相交的研究领域相交：生成模型的终身学习和参数有
效的网络学习
生成模型的终身学习。对于歧视性任务，e。g。分类，最近已经提出了许多工作来解决灾难性遗忘的问题。基于知识蒸馏的方法[32,20,4]通过最小化新旧网络输出之间的差异来工作。基于正则化的方法[19,5,2,6,22]在学习新任务时正则化网络参数。基于任务的方法[31,3,28]采用特定于任务的模块来学习每个任务
对于生成性任务，终身学习是一个未开发的领域，相对较少的工作研究了灾难性遗忘的问题。连续生成建模最早是由 Seff et al. [30]引人的，它将弹性权重合并 (EWC)$\operatorname{E}[19]$的思想纳人了 GAN 的终身学习中。Wu et al.[37] 很好地探索了内存重播的想法，其中从先前任务训练的模型生成的图像与当前任务的训练图像相结合，以允许训练多个任务，用于标签条件图像生成。但是，此方法 | [
[
"0.6235",
"0.11615",
"0.7363",
"0.11615",
"0.7363",
"0.13",
"0.6235",
"0.13"
],
[
"0.7295",
"0.11743",
"0.7876",
"0.11743",
"0.7876",
"0.1288",
"0.7295",
"0.1288"
],
[
"0.2197",
"0.1483",
"0.785",
"0.1483",
"0.785",
"0.1622",
"0.2197",
"0.1622"
],
[
"0.2214",
"0.1635",
"0.785",
"0.1635",
"0.785",
"0.1768",
"0.2214",
"0.1768"
],
[
"0.2222",
"0.1799",
"0.4192",
"0.1799",
"0.4192",
"0.1901",
"0.2222",
"0.1901"
],
[
"0.2443",
"0.1945",
"0.7837",
"0.1945",
"0.7837",
"0.2084",
"0.2443",
"0.2084"
],
[
"0.2181",
"0.2084",
"0.785",
"0.2102",
"0.785",
"0.2235",
"0.2181",
"0.2216"
],
[
"0.2214",
"0.2247",
"0.785",
"0.2247",
"0.785",
"0.238",
"0.2214",
"0.238"
],
[
"0.2197",
"0.2393",
"0.785",
"0.2393",
"0.785",
"0.2524",
"0.2197",
"0.2524"
],
[
"0.2197",
"0.254",
"0.785",
"0.254",
"0.785",
"0.2676",
"0.2197",
"0.2676"
],
[
"0.2222",
"0.2708",
"0.7817",
"0.2708",
"0.7817",
"0.281",
"0.2222",
"0.281"
],
[
"0.2214",
"0.2854",
"0.785",
"0.2854",
"0.785",
"0.2986",
"0.2214",
"0.2986"
],
[
"0.2214",
"0.2998",
"0.6177",
"0.2998",
"0.6177",
"0.3132",
"0.2214",
"0.3132"
],
[
"0.2427",
"0.3145",
"0.7837",
"0.3132",
"0.7837",
"0.3276",
"0.2427",
"0.3289"
],
[
"0.2214",
"0.33",
"0.785",
"0.33",
"0.785",
"0.3435",
"0.2214",
"0.3435"
],
[
"0.2197",
"0.3447",
"0.7837",
"0.3447",
"0.7837",
"0.358",
"0.2197",
"0.358"
],
[
"0.2336",
"0.3599",
"0.7837",
"0.3599",
"0.7837",
"0.3738",
"0.2336",
"0.3738"
],
[
"0.2214",
"0.3757",
"0.7837",
"0.3757",
"0.7837",
"0.3894",
"0.2214",
"0.3894"
],
[
"0.2214",
"0.3914",
"0.7812",
"0.3914",
"0.7812",
"0.4016",
"0.2214",
"0.4016"
],
[
"0.2197",
"0.406",
"0.785",
"0.406",
"0.785",
"0.4192",
"0.2197",
"0.4192"
],
[
"0.2214",
"0.4204",
"0.785",
"0.4204",
"0.785",
"0.4343",
"0.2214",
"0.4343"
],
[
"0.2197",
"0.4363",
"0.785",
"0.4363",
"0.785",
"0.4495",
"0.2197",
"0.4495"
],
[
"0.2197",
"0.452",
"0.4314",
"0.452",
"0.4314",
"0.4653",
"0.2197",
"0.4653"
],
[
"0.2443",
"0.4666",
"0.7837",
"0.4666",
"0.7837",
"0.4797",
"0.2443",
"0.4797"
],
[
"0.2197",
"0.4817",
"0.785",
"0.4817",
"0.785",
"0.4956",
"0.2197",
"0.4956"
],
[
"0.2197",
"0.4968",
"0.7837",
"0.4968",
"0.7837",
"0.5103",
"0.2197",
"0.5103"
],
[
"0.2214",
"0.513",
"0.7817",
"0.513",
"0.7817",
"0.5234",
"0.2214",
"0.5234"
],
[
"0.2197",
"0.526",
"0.7837",
"0.5264",
"0.7837",
"0.5405",
"0.2197",
"0.539"
],
[
"0.2197",
"0.5425",
"0.7837",
"0.5425",
"0.7837",
"0.556",
"0.2197",
"0.556"
],
[
"0.2214",
"0.558",
"0.7817",
"0.558",
"0.7817",
"0.5684",
"0.2214",
"0.5684"
],
[
"0.2181",
"0.571",
"0.7837",
"0.571",
"0.7837",
"0.584",
"0.2181",
"0.584"
],
[
"0.2197",
"0.587",
"0.5537",
"0.587",
"0.5537",
"0.601",
"0.2197",
"0.601"
],
[
"0.2157",
"0.627",
"0.3906",
"0.627",
"0.3906",
"0.64",
"0.2157",
"0.64"
],
[
"0.2197",
"0.66",
"0.7837",
"0.66",
"0.7837",
"0.674",
"0.2197",
"0.674"
],
[
"0.2197",
"0.677",
"0.3162",
"0.677",
"0.3162",
"0.6875",
"0.2197",
"0.6875"
],
[
"0.2427",
"0.6895",
"0.7837",
"0.69",
"0.7837",
"0.705",
"0.2427",
"0.704"
],
[
"0.2197",
"0.704",
"0.785",
"0.7056",
"0.785",
"0.7197",
"0.2197",
"0.7173"
],
[
"0.2214",
"0.72",
"0.7837",
"0.72",
"0.7837",
"0.7344",
"0.2214",
"0.7344"
],
[
"0.2222",
"0.7373",
"0.7817",
"0.7373",
"0.7817",
"0.7476",
"0.2222",
"0.7476"
],
[
"0.2197",
"0.752",
"0.2507",
"0.752",
"0.2507",
"0.7646",
"0.2197",
"0.7646"
],
[
"0.2443",
"0.7666",
"0.785",
"0.7666",
"0.785",
"0.78",
"0.2443",
"0.78"
],
[
"0.2222",
"0.7827",
"0.7817",
"0.7827",
"0.7817",
"0.7935",
"0.2222",
"0.7935"
],
[
"0.2197",
"0.7954",
"0.785",
"0.797",
"0.785",
"0.81",
"0.2197",
"0.8096"
],
[
"0.2181",
"0.81",
"0.7837",
"0.811",
"0.7837",
"0.8257",
"0.2181",
"0.8247"
],
[
"0.2197",
"0.827",
"0.7837",
"0.827",
"0.7837",
"0.8403",
"0.2197",
"0.8403"
]
] | [
"Piggyback GAN",
"3",
"突。对于判别模型，很少发生一个图像具有不同的标签（出现在不同的任务中）。",
"但是，对于生成模型，我们希望将相同的图像转换为不同的域以执行不同的任",
"务是很常见的F14.11.44.8",
"解决生成模型的灾难性遗忘的一种天真的方法是分别为每个任务训练一个模",
"型（见图？？（c））。不幸的是，这种方法通常不可扩展：随着新任务的添加，存",
"储需求急剧增长。更重要的是，将经过训练的模型放在一边，而不利用它在面对",
"新任务时可能提供的好处，这将是对资源的低效利用。最近的工作E[42,37]在",
"生成模型的终身学习方面显示出令人鼓舞的结果，但也揭示了图像质量下降和",
"伪影从旧任务转移到新任务是不可避免的。因此，设计一个持续学习框架是有",
"价值的，该框架（1）具有更高的参数效率，（2）保留当前和先前学习任务的生",
"成质量，以及（3）可以跨不同域启用各种条件生成任务",
"在本文中，我们介绍了一个通用的持续学习框架PiggybackGAN，该框架可",
"以跨不同域执行各种条件图像生成任务（请参见图E？？（d)）。所提出的方法解",
"决了生成连续学习模型中的灾难性遗忘问题。具体来说，PiggybackGAN维护",
"个滤波器组，其中的滤波器来自在先前任务中训练的模型的卷积和反卷积层。",
"面对新任务，PiggybackGAN通过构建一组piggyback滤波器来重用滤波器组",
"中的滤波器，从而学习执行此任务，这些滤波器被分解为滤波器组中的滤波器",
"PiggybackGAN还为每个任务维护一小部分不受约束的过滤器，以确保高质量",
"的生成。学习新任务后，将无约束过滤器附加到过滤器组以促进后续任务的训",
"练。由于旧任务的过滤器没有被改变，PiggybackGAN能够保持以前任务的确",
"切知识，而不会忘记任何细节",
"总结一下，我们的贡献如下。我们提出了一种用于条件图像生成的持续学习",
"框架，该框架是（1）高效的，通过在先前任务上训练的模型上“piggybacking”",
"并重用先前模型中的过滤器来学习执行新任务，（2）保持质量，保持当前任务的",
"生成质量，并确保先前任务的质量不会下降，同时节省更多参数，以及（3）通",
"用的，可跨不同域启用各种条件生成任务。据我们所知，我们是第一个为生成",
"持续学习模型做出这些贡献的人。我们在两种设置下验证了我们的方法的有效",
"性：（1）成对的图像条件生成，以及（2）不成对的图像条件生成。与最先进的",
"模型进行了广泛的定性和定量比较，以说明我们提出的框架学习新一代任务的",
"能力和效率，而不会灾难性地忘记以前的任务。",
"2Related Work",
"我们的工作与两个以前不相交的研究领域相交：生成模型的终身学习和参数有",
"效的网络学习",
"生成模型的终身学习。对于歧视性任务，e。g。分类，最近已经提出了许多",
"工作来解决灾难性遗忘的问题。基于知识蒸馏的方法[32,20,4]通过最小化新旧",
"网络输出之间的差异来工作。基于正则化的方法[19,5,2,6,22]在学习新任务时",
"正则化网络参数。基于任务的方法131.3.28采用特定于任务的模块来学习每个",
"任务",
"对于生成性任务，终身学习是一个未开发的领域，相对较少的工作研究了灾",
"难性遗忘的问题。连续生成建模最早是由Seffetal.30引人的，它将弹性权重",
"合并（EWC)E[19]的思想纳人了GAN的终身学习中。Wuetal.[37]很好地",
"探索了内存重播的想法，其中从先前任务训练的模型生成的图像与当前任务的",
"训练图像相结合，以允许训练多个任务，用于标签条件图像生成。但是，此方法"
] | {
"arxiv_id": "2104.11939",
"language": "zh",
"page_idx": 2
} |
||
2104.11939-zh-3 | 2104.11939-zh-3.jpg | 4 M. Zhai et al.
仅限于标签条件图像生成，并且不适用于图像条件图像生成，因为未提供先前任务的真实条件输入。EWClE[19] 已从分类任务适应于标签条件图像生成E[30, 37] 的生成任务，但它们在记忆先前任务和生成逼真图像方面的能力有限。终身 GANE[42] 是一种基于知识提炼的通用方法，用于终身学习中的条件图像生成。但是，当学习新任务时，先前任务生成的图像的图像质量会不断降低。此外当任务之间的输入输出空间发生冲突时，上述所有方法都会失败。据我们所知， PiggybackGAN 是第一种可以保留先前任务的生成质量的方法，并且可以跨不同域启用各种条件生成任务
参数高效的网络学习。近年来，人们越来越有兴趣将神经网络的资源消耗与部署它们的平台的计算约束更好地对齐[7,36,38] 。积极研究的一个重要途径是参数效率。训练有素的神经网络通常是高度过参数化的 [10] 。有许多有效的策略可以提高训练有素的网络的参数效率：例如，修剪算法学习删除冗余或不太重要的参数[12,40] ,量化算法学习使用更少的位表示参数，同时保持任务性能[15,17,43] 。通过架构设计[23,29] ,稀疏诱导先验[33] ,资源重新分配 [26] 或来自教师网络[34,41]的知识提炼，也可以在培训期间提高参数效率。我们的工作与这些方法正交，这些方法可以应用于 PiggybackGAN 之上
这项工作的重点是参数高效的终身学习：我们的目标是在将经过训练的网络扩展到新任务时，利用权重重用和自适应来提高参数效率，同时避免先前任务的灾难性遗忘。已经探索了几种在歧视性环境中进行参数有效的终身学习的方法。例如，渐进式网络[28]为每个任务训练一个新网络，并通过学习横向连接来传递来自先前任务的知识，计算来自先前任务的层激活的线性组合，并将其添加到新任务的相应层的输人中。虽然我们的方法计算滤波器的线性组合来构造新的滤波器，并使用它们与普通滤波器相同。Rebuff et al.[27] 旨在训练一个通用向量，该向量在所有域之间共享，为所有域提供通用过滤器，并添加了特定于任务的适配器模块，以将经过训练的网络调整为新任务。当我们正在解决一系列任务时，为初始任务训练的过滤器对于后面的任务来说几乎不够通用。受二进制神经网络的启发，Mallya et al. [24]引人了背负式掩码方法 (该方法启发了 PiggybackGAN 的命名)。此方法通过每个新任务的网络参数学习二进制掩码。任务共享相同的主干参数，但启用的参数不同。多任务注意力网络[21] 为每个新任务学习软注意力模块。特定于任务的特征是通过使用与输入相关的软注意掩码缩放特征来获得的。据我们所知，PiggybackGAN 是在生成环境中进行参数有效的终身学习的第一种方法
# 3 Method
在本文中，我们研究了生成模型的终身学习问题。总体目标是在学习过程一次只能访问一个任务并且每个任务的训练数据只能访问一次的假设下，学习给定的生成任务序列$\{T_i\}_{i=1}^N$的模型
假设我们有一个学习模型$M$ ,它被训练来执行第一个任务$T_1$ 。要继续执行新任务$T_2$ , -种朴素的方法是通过微调其参数来使学习的模型$M$适应新任务。这种方法遭受了任务$T_{1}$灾难性遗忘的问题。另一种朴素的方法是为每个新任务$T_{i}$学习一个单独的独立模型$M_i$ , 并存储所有学习的模型$\{ M_i\} _i= 1^N$ , 这在不利用以前的模型的情况下大幅增加了参数总数
本文的主要思想是通过在现有模型上“piggybacking”并重用先前任务训练的模型中的过滤器来构建参数高效的终身生成模型，同时保持生成质量类似于 | [
[
"0.2173",
"0.11743",
"0.2656",
"0.11743",
"0.2656",
"0.1288",
"0.2173",
"0.1288"
],
[
"0.273",
"0.11743",
"0.3635",
"0.11743",
"0.3635",
"0.1276",
"0.273",
"0.1276"
],
[
"0.2214",
"0.1483",
"0.7837",
"0.1483",
"0.7837",
"0.1622",
"0.2214",
"0.1622"
],
[
"0.2197",
"0.1635",
"0.7837",
"0.1635",
"0.7837",
"0.1768",
"0.2197",
"0.1768"
],
[
"0.2181",
"0.1768",
"0.7837",
"0.178",
"0.7837",
"0.1925",
"0.2181",
"0.1913"
],
[
"0.2214",
"0.1932",
"0.7837",
"0.1932",
"0.7837",
"0.207",
"0.2214",
"0.207"
],
[
"0.2214",
"0.209",
"0.7837",
"0.209",
"0.7837",
"0.2229",
"0.2214",
"0.2229"
],
[
"0.2214",
"0.2247",
"0.785",
"0.2247",
"0.785",
"0.238",
"0.2214",
"0.238"
],
[
"0.2181",
"0.2374",
"0.785",
"0.238",
"0.785",
"0.2524",
"0.2181",
"0.252"
],
[
"0.2214",
"0.254",
"0.4143",
"0.254",
"0.4143",
"0.2676",
"0.2214",
"0.2676"
],
[
"0.2427",
"0.2683",
"0.7837",
"0.2695",
"0.7837",
"0.2827",
"0.2427",
"0.2822"
],
[
"0.2214",
"0.286",
"0.7817",
"0.286",
"0.7817",
"0.2966",
"0.2214",
"0.2966"
],
[
"0.2197",
"0.2998",
"0.7837",
"0.2998",
"0.7837",
"0.3132",
"0.2197",
"0.3132"
],
[
"0.2214",
"0.3164",
"0.7837",
"0.3164",
"0.7837",
"0.327",
"0.2214",
"0.327"
],
[
"0.2214",
"0.33",
"0.785",
"0.33",
"0.785",
"0.3435",
"0.2214",
"0.3435"
],
[
"0.2222",
"0.3467",
"0.7837",
"0.3467",
"0.7837",
"0.3567",
"0.2222",
"0.3567"
],
[
"0.2214",
"0.3599",
"0.786",
"0.3599",
"0.786",
"0.3738",
"0.2214",
"0.3738"
],
[
"0.2197",
"0.3757",
"0.6665",
"0.3757",
"0.6665",
"0.3894",
"0.2197",
"0.3894"
],
[
"0.2427",
"0.3914",
"0.7817",
"0.3914",
"0.7817",
"0.4016",
"0.2427",
"0.4016"
],
[
"0.2181",
"0.406",
"0.7837",
"0.406",
"0.7837",
"0.4192",
"0.2181",
"0.4192"
],
[
"0.2214",
"0.4224",
"0.7817",
"0.4224",
"0.7817",
"0.433",
"0.2214",
"0.433"
],
[
"0.2197",
"0.4363",
"0.7837",
"0.4363",
"0.7837",
"0.4495",
"0.2197",
"0.4495"
],
[
"0.2214",
"0.452",
"0.785",
"0.452",
"0.785",
"0.4653",
"0.2214",
"0.4653"
],
[
"0.2197",
"0.4666",
"0.7837",
"0.4666",
"0.7837",
"0.4797",
"0.2197",
"0.4797"
],
[
"0.2197",
"0.481",
"0.7837",
"0.481",
"0.7837",
"0.4944",
"0.2197",
"0.4944"
],
[
"0.2214",
"0.4976",
"0.7837",
"0.4976",
"0.7837",
"0.5083",
"0.2214",
"0.5083"
],
[
"0.2214",
"0.511",
"0.773",
"0.511",
"0.773",
"0.5244",
"0.2214",
"0.5244"
],
[
"0.2214",
"0.5264",
"0.785",
"0.5264",
"0.785",
"0.5405",
"0.2214",
"0.5405"
],
[
"0.2303",
"0.5415",
"0.7837",
"0.5415",
"0.7837",
"0.5547",
"0.2303",
"0.5547"
],
[
"0.2214",
"0.5566",
"0.785",
"0.5566",
"0.785",
"0.571",
"0.2214",
"0.571"
],
[
"0.2197",
"0.5713",
"0.785",
"0.5713",
"0.785",
"0.5854",
"0.2197",
"0.5854"
],
[
"0.2222",
"0.5884",
"0.7817",
"0.5884",
"0.7817",
"0.5986",
"0.2222",
"0.5986"
],
[
"0.2197",
"0.6016",
"0.7837",
"0.6016",
"0.7837",
"0.6157",
"0.2197",
"0.6157"
],
[
"0.2181",
"0.617",
"0.482",
"0.617",
"0.482",
"0.631",
"0.2181",
"0.631"
],
[
"0.2173",
"0.6577",
"0.3325",
"0.6577",
"0.3325",
"0.672",
"0.2173",
"0.672"
],
[
"0.2197",
"0.69",
"0.785",
"0.69",
"0.785",
"0.704",
"0.2197",
"0.704"
],
[
"0.2214",
"0.7056",
"0.7837",
"0.7056",
"0.7837",
"0.7197",
"0.2214",
"0.7197"
],
[
"0.2197",
"0.72",
"0.4478",
"0.72",
"0.4478",
"0.7344",
"0.2197",
"0.7344"
],
[
"0.246",
"0.7363",
"0.785",
"0.7363",
"0.785",
"0.75",
"0.246",
"0.75"
],
[
"0.2197",
"0.7505",
"0.7837",
"0.7505",
"0.7837",
"0.7646",
"0.2197",
"0.7646"
],
[
"0.2197",
"0.7666",
"0.7837",
"0.7666",
"0.7837",
"0.78",
"0.2197",
"0.78"
],
[
"0.2173",
"0.779",
"0.786",
"0.781",
"0.786",
"0.7974",
"0.2173",
"0.7954"
],
[
"0.2214",
"0.797",
"0.5312",
"0.797",
"0.5312",
"0.81",
"0.2214",
"0.81"
],
[
"0.2443",
"0.81",
"0.7837",
"0.811",
"0.7837",
"0.8257",
"0.2443",
"0.8247"
],
[
"0.2197",
"0.827",
"0.7837",
"0.827",
"0.7837",
"0.8403",
"0.2197",
"0.8403"
]
] | [
"4",
"M.Zhai et al.",
"仅限于标签条件图像生成，并且不适用于图像条件图像生成，因为未提供先前",
"任务的真实条件输人。EWCE19I已从分类任务适应于标签条件图像生成E30",
"37]的生成任务，但它们在记忆先前任务和生成逼真图像方面的能力有限。终",
"身GANE[42]是一种基于知识提炼的通用方法，用于终身学习中的条件图像生",
"成。但是，当学习新任务时，先前任务生成的图像的图像质量会不断降低。此外",
"当任务之间的输入输出空间发生冲突时，上述所有方法都会失败。据我们所知，",
"PiggybackGAN是第一种可以保留先前任务的生成质量的方法，并且可以跨不",
"同域启用各种条件生成任务",
"参数高效的网络学习。近年来，人们越来越有兴趣将神经网络的资源消耗与",
"部署它们的平台的计算纳束更好地对齐17，36.38。积极研究的一个重要途径是",
"参数效率。训练有素的神经网络通常是高度过参数化的[10]。有许多有效的策",
"略可以提高训练有素的网络的参数效率：例如，修剪算法学习删除余或不太",
"重要的参数12,40]，量化算法学习使用更少的位表示参数，同时保持任务性能",
"1517.43。通过架构设计23.29，稀疏诱导先验33，资源重新分配26或",
"来自教师网络[34,41]的知识提炼，也可以在培训期间提高参数效率。我们的工",
"作与这些方法正交，这些方法可以应用于PiggybackGAN之上",
"这项工作的重点是参数高效的终身学习：我们的目标是在将经过训练的网络",
"扩展到新任务时，利用权重重用和自适应来提高参数效率，同时避免先前任务",
"的灾难性遗忘。已经探索了儿种在歧视性环境中进行参数有效的终身学习的方",
"法。例如，渐进式网络[28]为每个任务训练一个新网络，并通过学习横向连接",
"来传递来自先前任务的知识，计算来自先前任务的层激活的线性组合，并将其",
"添加到新任务的相应层的输入中。虽然我们的方法计算滤波器的线性组合来构",
"造新的滤波器，并使用它们与普通滤波器相同。Rebufietal.[27]旨在训练一个",
"通用向量，该向量在所有域之间共享，为所有域提供通用过滤器，并添加了特定",
"于任务的适配器模块，以将经过训练的网络调整为新任务。当我们正在解决",
"系列任务时，为初始任务训练的过滤器对于后面的任务来说几乎不够通用。受",
"二进制神经网络的启发，Mallyaetal.[24]引人了背负式掩码方法（该方法启发",
"了PiggybackGAN的命名）。此方法通过每个新任务的网络参数学习二进制掩",
"码。任务共享相同的主干参数，但启用的参数不同。多任务注意力网络[21]为",
"每个新任务学习软注意力模块。特定于任务的特征是通过使用与输入相关的软",
"注意掩码缩放特征来获得的。据我们所知，PiggybackGAN是在生成环境中进",
"行参数有效的终身学习的第一种方法",
"3Method",
"在本文中，我们研究了生成模型的终身学习问题。总体目标是在学习过程一次",
"只能访问一个任务并且每个任务的训练数据只能访问一次的假设下，学习给定",
"的生成任务序列{T;}-1的模型",
"假设我们有一个学习模型M，它被训练来执行第一个任务Ti。要继续执行",
"新任务T2，一种朴素的方法是通过微调其参数来使学习的模型M适应新任务。",
"这种方法遭受了任务T1灾难性遗忘的问题。另一种朴素的方法是为每个新任务",
"T学习一个单独的独立模型Mi，并存储所有学习的模型【Mi}=1，这在不利",
"用以前的模型的情况下大幅增加了参数总数",
"本文的主要思想是通过在现有模型上“piggybacking”并重用先前任务训练",
"的模型中的过滤器来构建参数高效的终身生成模型，同时保持生成质量类似于"
] | {
"arxiv_id": "2104.11939",
"language": "zh",
"page_idx": 3
} |
||
2104.11939-zh-4 | 2104.11939-zh-4.jpg | 5
Piggyback GAN
<FigureHere>
Fig.1. Overview of Piggyback GAN .Piggyback GAN maintains a filter bank for composing convolutional filters to be used for the next task. Given a new task, most filters in the model are trained as piggyback filters that are factorized into filters in the filter bank. The remaining small portion of filters are learned without constraints and appended to the filter bank to benefit the training of subsequent tasks.
单个独立模型适用于所有任务。我们通过维护滤波器组来实现这一点，并为新任务组成滤波器，主要是通过将其分解为滤波器组中的滤波器，其中一小部分滤波器是在没有约束的情况下学习的。学习新任务后，通过添加那些不受约束的过滤器来扩展过滤器组
# 3.1 Piggyback Filter Learning
对于生成模型的终身学习，给定初始任务$T_1$和即将到来的新任务$T_2$ ,可能存在$T_{1}$和$T_{2}$共享相似输入但不同输出的情况，例如。g。$T_1$是照片$\to$莫奈绘画和$T_2$是照片$\to$浮世绘。当每个任务单独训练时，输入输出空间中的这些冲突不会发生。但是，正如已经提到的，为每个任务学习一个单独的模型并存储所有学习的模型会消耗内存，并且不会利用以前的模型
对于这种情况，来自为$T_1$训练的模型的过滤器中的知识可以为$T_2$的训练提供有价值的信息。对于来自视觉上不太相似的域的图像，众所周知，在判别任务中，卷积神经网络的普遍性有助于保持其在以前看不见的域中的能力[24,27] 。这些证据表明，尽管细节上存在差异，但某些传达语义的一般模式仍然可以在以前的过滤器空间中覆盖。受这些观察结果的启发，我们建议利用以前训练过的过滤器来促进新任务的训练。在我们的实验中，我们确实发现以前的过滤器对学习新任务做出了稳定的贡献，而不管任务之间的域差异如何
考虑为初始任务$T_1$学习的生成器网络$G_1$,该网络由$L$层滤波器$\{F_1^\ell\}_\ell=1^L$ 组成，其中$\ell$表示层的索引。对于$\ell^th$层，让内核大小为$s_w^\ell\times s_h^\ell$,输入通道数为$c_{in}^{\ell}$ ,输出通道数为$c_out^{\ell}$。然后通过$c_out$将$F_1^{\ell}$从 4D 重塑为$s_w^{\ell}\times s_h^{\ell}\times c_{in}^{\ell}$ 大小的 2D, 我们用$\mathcal{R}(F_1^\ell)$表示。对于即将到来的任务$T_2$ ,我们通过对任务$T_1$ 中相应层的过滤器进行因式分解操作来学习过滤器$F_{2}^{\ell}$,即
$$F_2^\ell\:=\:\mathcal{R}^{-1}(\mathcal{R}(F_1^\ell)\otimes W_2^\ell),$$
(1) | [
[
"0.6235",
"0.11554",
"0.734",
"0.11554",
"0.734",
"0.1288",
"0.6235",
"0.1288"
],
[
"0.774",
"0.11743",
"0.786",
"0.11743",
"0.786",
"0.1276",
"0.774",
"0.1276"
],
[
"0.2852",
"0.1515",
"0.3179",
"0.1515",
"0.3179",
"0.1635",
"0.2852",
"0.1635"
],
[
"0.5146",
"0.1515",
"0.5483",
"0.1515",
"0.5483",
"0.1635",
"0.5146",
"0.1635"
],
[
"0.6274",
"0.1515",
"0.661",
"0.1515",
"0.661",
"0.1635",
"0.6274",
"0.1635"
],
[
"0.7197",
"0.1515",
"0.755",
"0.1515",
"0.755",
"0.1635",
"0.7197",
"0.1635"
],
[
"0.71",
"0.1635",
"0.773",
"0.1604",
"0.774",
"0.1718",
"0.7114",
"0.1749"
],
[
"0.4019",
"0.1528",
"0.433",
"0.1528",
"0.433",
"0.161",
"0.4019",
"0.161"
],
[
"0.2239",
"0.1749",
"0.2615",
"0.1749",
"0.2615",
"0.1849",
"0.2239",
"0.1849"
],
[
"0.2222",
"0.1818",
"0.2646",
"0.1818",
"0.2646",
"0.1925",
"0.2222",
"0.1925"
],
[
"0.7534",
"0.1989",
"0.7646",
"0.1989",
"0.7646",
"0.202",
"0.7534",
"0.202"
],
[
"0.738",
"0.2102",
"0.763",
"0.2102",
"0.763",
"0.214",
"0.738",
"0.214"
],
[
"0.2336",
"0.2343",
"0.2607",
"0.2393",
"0.255",
"0.2551",
"0.2279",
"0.25"
],
[
"0.7344",
"0.2507",
"0.7563",
"0.2507",
"0.7563",
"0.2583",
"0.7344",
"0.2583"
],
[
"0.353",
"0.2632",
"0.398",
"0.2632",
"0.398",
"0.2778",
"0.353",
"0.2778"
],
[
"0.2181",
"0.3193",
"0.7837",
"0.3193",
"0.7837",
"0.3333",
"0.2181",
"0.3333"
],
[
"0.2173",
"0.3352",
"0.785",
"0.334",
"0.785",
"0.348",
"0.2173",
"0.3484"
],
[
"0.2181",
"0.3484",
"0.785",
"0.3484",
"0.785",
"0.3623",
"0.2181",
"0.3623"
],
[
"0.2181",
"0.3623",
"0.786",
"0.3623",
"0.786",
"0.3757",
"0.2181",
"0.3757"
],
[
"0.2181",
"0.377",
"0.7197",
"0.377",
"0.7197",
"0.3901",
"0.2181",
"0.3901"
],
[
"0.2197",
"0.4204",
"0.785",
"0.4204",
"0.785",
"0.4343",
"0.2197",
"0.4343"
],
[
"0.2197",
"0.4363",
"0.7837",
"0.4363",
"0.7837",
"0.4495",
"0.2197",
"0.4495"
],
[
"0.2181",
"0.4507",
"0.7837",
"0.4495",
"0.7837",
"0.464",
"0.2181",
"0.4653"
],
[
"0.2197",
"0.4666",
"0.4004",
"0.4666",
"0.4004",
"0.4797",
"0.2197",
"0.4797"
],
[
"0.2141",
"0.504",
"0.4836",
"0.505",
"0.4836",
"0.523",
"0.2141",
"0.5215"
],
[
"0.2197",
"0.535",
"0.785",
"0.535",
"0.785",
"0.549",
"0.2197",
"0.549"
],
[
"0.2197",
"0.5503",
"0.785",
"0.5503",
"0.785",
"0.5645",
"0.2197",
"0.5645"
],
[
"0.2197",
"0.565",
"0.785",
"0.565",
"0.785",
"0.579",
"0.2197",
"0.579"
],
[
"0.2197",
"0.579",
"0.785",
"0.58",
"0.785",
"0.5947",
"0.2197",
"0.5933"
],
[
"0.2197",
"0.595",
"0.58",
"0.595",
"0.58",
"0.6094",
"0.2197",
"0.6094"
],
[
"0.2443",
"0.6113",
"0.785",
"0.6113",
"0.785",
"0.625",
"0.2443",
"0.625"
],
[
"0.2197",
"0.6255",
"0.785",
"0.625",
"0.785",
"0.6396",
"0.2197",
"0.64"
],
[
"0.2214",
"0.6416",
"0.786",
"0.6416",
"0.786",
"0.655",
"0.2214",
"0.655"
],
[
"0.2197",
"0.656",
"0.785",
"0.656",
"0.785",
"0.6694",
"0.2197",
"0.6694"
],
[
"0.2197",
"0.672",
"0.785",
"0.672",
"0.785",
"0.685",
"0.2197",
"0.685"
],
[
"0.2197",
"0.686",
"0.786",
"0.686",
"0.786",
"0.6997",
"0.2197",
"0.6997"
],
[
"0.2197",
"0.702",
"0.6763",
"0.702",
"0.6763",
"0.7153",
"0.2197",
"0.7153"
],
[
"0.2443",
"0.7173",
"0.7817",
"0.7173",
"0.7817",
"0.732",
"0.2443",
"0.732"
],
[
"0.2173",
"0.73",
"0.785",
"0.731",
"0.785",
"0.7485",
"0.2173",
"0.7476"
],
[
"0.2197",
"0.744",
"0.7817",
"0.744",
"0.7817",
"0.762",
"0.2197",
"0.762"
],
[
"0.2214",
"0.762",
"0.7837",
"0.762",
"0.7837",
"0.776",
"0.2214",
"0.776"
],
[
"0.2181",
"0.7764",
"0.638",
"0.7764",
"0.638",
"0.7905",
"0.2181",
"0.7905"
],
[
"0.4102",
"0.8228",
"0.5933",
"0.8237",
"0.5933",
"0.8384",
"0.4102",
"0.837"
],
[
"0.761",
"0.8228",
"0.7876",
"0.8228",
"0.7876",
"0.8403",
"0.761",
"0.8403"
]
] | [
"Piggyback GAN",
"5",
"Task 1",
"Task 3",
"Task 4",
"Task...",
":Factarization",
"Task 2",
"Model for",
"EachTask",
"an",
"",
"Fank",
"F5",
"",
"Fig.1. Overview of Piggyback GAN . Piggyback GAN maintains a filter bank for",
"composing convolutional filters to beused for the next task. Given a new task,most",
"filters in the model are trained as piggyback filters that are factorized into flters in",
"the filter bank. The remaining small portion of filters are learned without constraints",
"and appended to the filter bank to benefit the training of subsequent tasks.",
"单个独立模型适用于所有任务。我们通过维护滤波器组来实现这一点，并为新",
"任务组成滤波器，主要是通过将其分解为滤波器组中的滤波器，其中一小部分",
"滤波器是在没有约束的情况下学习的。学习新任务后，通过添加那些不受约束",
"的过滤器来扩展过滤器组",
"3.1Piggyback Filter Learning",
"对于生成模型的终身学习，给定初始任务T和即将到来的新任务T2，可能存",
"在T和T2共享相似输人但不同输出的情况，例如。g。Ti是照片→莫奈绘画",
"和T2是照片一→浮世绘。当每个任务单独训练时，输人输出空间中的这些冲突",
"不会发生。但是，正如已经提到的，为每个任务学习一个单独的模型并存储所有",
"学习的模型会消耗内存，并且不会利用以前的模型",
"对于这种情况，来自为T训练的模型的过滤器中的知识可以为T2的训练提",
"供有价值的信息。对于来自视觉上不太相似的域的图像，众所周知，在判别任务",
"中，卷积神经网络的普遍性有助于保持其在以前看不见的域中的能力[24,27]。",
"这些证据表明，尽管细节上存在差异，但某些传达语义的一般模式仍然可以在",
"以前的过滤器空间中覆盖。受这些观察结果的启发，我们建议利用以前训练过",
"的过滤器来促进新任务的训练。在我们的实验中，我们确实发现以前的过滤器",
"对学习新任务做出了稳定的贡献，而不管任务之间的域差异如何",
"考虑为初始任务Ti学习的生成器网络G1，该网络由L层滤波器【F-]",
"组成，其中l表示层的索引。对于ath层，让内核大小为su×sh，输人通道数",
"为cm，输出通道数为cout。然后通过cout将Fl从4D重塑为sa×sh×c",
"大小的2D，我们用R(F）表示。对于即将到来的任务T2，我们通过对任务T1",
"中相应层的过滤器进行因式分解操作来学习过滤器F，即",
"F = R-1(R(F) W2)",
"(1)"
] | {
"arxiv_id": "2104.11939",
"language": "zh",
"page_idx": 4
} |
||
2303.03912-zh-0 | 2303.03912-zh-0.jpg | # Document-level Relation Extraction with Cross-sentence Reasoning Graph
Hongfei Liu$^1$,Zhao Kang$^1\star$, Lizong Zhang$^1$, Ling Tian$^1$, and Fujun Hua$^2$
$^{1}$ University of Electronic Science and Technology of China, Chengdu, China
[email protected]
$\{$ zkang,l.zhang,lingtian $\}$ @uestc.edu.cn
$^{2}$ Research and Development Center TROY Information Technology Co., Ltd.
Chengdu, China [email protected]
Abstract. Relation extraction (RE) has recently moved from the sentencelevel to document-level, which requires aggregating document information and using entities and mentions for reasoning. Existing works put entity nodes and mention nodes with similar representations in a documentlevel graph, whose complex edges may incur redundant information. Furthermore, existing studies only focus on entity-level reasoning paths without considering global interactions among entities cross-sentence. To these ends, we propose a novel document-level RE model with a GR aph information A ggregation and C ross-sentence R easoning network (GRACR). Specifically, a simplified document-level graph is constructed to model the semantic information of all mentions and sentences in a document, and an entity-level graph is designed to explore relations of long-distance cross-sentence entity pairs. Experimental results show that GRACR achieves excellent performance on two public datasets of document-level RE. It is especially effective in extracting potential relations of cross-sentence entity pairs. Our code is available at https://github.com/UESTC-LHF/GRACR~.
Keywords: Deep learning $\cdot$ Relation extraction$\cdot$ Document-level RE
## 1 Introduction
关系提取(RE) 是识别文本中一对命名实体之间的语义关系。文档级 RE 要求模型从文档中提取关系，并面临一些棘手的挑战。首先，一个文档包含多个句子，因此关系提取任务需要处理更丰富和复杂的语义信息。其次，同一三元组中的主语和宾语实体可能出现在不同的句子中，并且某些实体具有别名，这些别名通常被称为实体提及。因此，文档级 RE 使用的信息可能不是来自一个句子。第三，不同的三元组之间可能存在相互作用。从不同的三元组中提取两个实体之间的关系需要使用上下文特征进行推理。图 1 显示了来自 DocRED 数据集 [21] 的示例。因为主语和宾语出现在同一个句子中，所以很容易预测句子内的关系。但是，在识别“瑞典语”和“瑞典皇家学院”之间的句子间关系时存在问题，它们的提及分布在不同的句子中，并且存在长距离依赖关系。[21] 提出了$^{\star}$Corresponding author. | [
[
"0.2974",
"0.1465",
"0.7075",
"0.1465",
"0.7075",
"0.161",
"0.2974",
"0.161"
],
[
"0.3381",
"0.1666",
"0.665",
"0.1674",
"0.665",
"0.1849",
"0.3381",
"0.1843"
],
[
"0.2378",
"0.214",
"0.7646",
"0.214",
"0.7646",
"0.2291",
"0.2378",
"0.2291"
],
[
"0.246",
"0.2393",
"0.7563",
"0.2405",
"0.7563",
"0.2551",
"0.246",
"0.254"
],
[
"0.3962",
"0.254",
"0.607",
"0.2551",
"0.607",
"0.2695",
"0.3962",
"0.2683"
],
[
"0.3677",
"0.2695",
"0.634",
"0.2695",
"0.634",
"0.2827",
"0.3677",
"0.2827"
],
[
"0.2402",
"0.281",
"0.7607",
"0.2822",
"0.7607",
"0.2966",
"0.2402",
"0.2954"
],
[
"0.3987",
"0.2966",
"0.601",
"0.2966",
"0.601",
"0.31",
"0.3987",
"0.31"
],
[
"0.2656",
"0.3467",
"0.744",
"0.3467",
"0.744",
"0.3599",
"0.2656",
"0.3599"
],
[
"0.2646",
"0.361",
"0.74",
"0.361",
"0.74",
"0.375",
"0.2646",
"0.375"
],
[
"0.2656",
"0.375",
"0.74",
"0.375",
"0.74",
"0.3882",
"0.2656",
"0.3882"
],
[
"0.2656",
"0.3882",
"0.7417",
"0.3882",
"0.7417",
"0.4016",
"0.2656",
"0.4016"
],
[
"0.2646",
"0.4028",
"0.739",
"0.4028",
"0.739",
"0.4167",
"0.2646",
"0.4167"
],
[
"0.2632",
"0.4153",
"0.74",
"0.4167",
"0.74",
"0.4312",
"0.2632",
"0.43"
],
[
"0.2646",
"0.4287",
"0.738",
"0.43",
"0.738",
"0.4443",
"0.2646",
"0.443"
],
[
"0.2656",
"0.4443",
"0.74",
"0.4443",
"0.74",
"0.4583",
"0.2656",
"0.4583"
],
[
"0.2656",
"0.4583",
"0.7417",
"0.4583",
"0.7417",
"0.4717",
"0.2656",
"0.4717"
],
[
"0.2656",
"0.4717",
"0.74",
"0.4717",
"0.74",
"0.4849",
"0.2656",
"0.4849"
],
[
"0.2656",
"0.4849",
"0.74",
"0.4849",
"0.74",
"0.4988",
"0.2656",
"0.4988"
],
[
"0.2646",
"0.5",
"0.74",
"0.5",
"0.74",
"0.513",
"0.2646",
"0.513"
],
[
"0.2646",
"0.513",
"0.739",
"0.513",
"0.739",
"0.5264",
"0.2646",
"0.5264"
],
[
"0.2646",
"0.5264",
"0.739",
"0.5264",
"0.739",
"0.5405",
"0.2646",
"0.5405"
],
[
"0.2632",
"0.5405",
"0.739",
"0.5415",
"0.739",
"0.5547",
"0.2632",
"0.5537"
],
[
"0.2656",
"0.5547",
"0.7344",
"0.5547",
"0.7344",
"0.5684",
"0.2656",
"0.5684"
],
[
"0.2646",
"0.5684",
"0.556",
"0.5684",
"0.556",
"0.582",
"0.2646",
"0.582"
],
[
"0.5522",
"0.5737",
"0.562",
"0.5737",
"0.562",
"0.579",
"0.5522",
"0.579"
],
[
"0.2646",
"0.595",
"0.7344",
"0.595",
"0.7344",
"0.6094",
"0.2646",
"0.6094"
],
[
"0.2173",
"0.637",
"0.2467",
"0.637",
"0.2467",
"0.6484",
"0.2173",
"0.6484"
],
[
"0.2507",
"0.636",
"0.3594",
"0.636",
"0.3594",
"0.651",
"0.2507",
"0.651"
],
[
"0.2197",
"0.6675",
"0.785",
"0.6675",
"0.785",
"0.681",
"0.2197",
"0.681"
],
[
"0.2173",
"0.681",
"0.7837",
"0.6816",
"0.7837",
"0.6963",
"0.2173",
"0.696"
],
[
"0.2197",
"0.6978",
"0.7837",
"0.6978",
"0.7837",
"0.711",
"0.2197",
"0.711"
],
[
"0.2214",
"0.712",
"0.785",
"0.712",
"0.785",
"0.726",
"0.2214",
"0.726"
],
[
"0.2214",
"0.728",
"0.785",
"0.728",
"0.785",
"0.741",
"0.2214",
"0.741"
],
[
"0.2214",
"0.742",
"0.7837",
"0.742",
"0.7837",
"0.756",
"0.2214",
"0.756"
],
[
"0.2197",
"0.7583",
"0.785",
"0.7583",
"0.785",
"0.7715",
"0.2197",
"0.7715"
],
[
"0.2181",
"0.773",
"0.785",
"0.773",
"0.785",
"0.786",
"0.2181",
"0.786"
],
[
"0.2181",
"0.786",
"0.7837",
"0.787",
"0.7837",
"0.802",
"0.2181",
"0.8003"
],
[
"0.2214",
"0.803",
"0.7817",
"0.803",
"0.7817",
"0.8164",
"0.2214",
"0.8164"
],
[
"0.2197",
"0.8237",
"0.384",
"0.827",
"0.384",
"0.8447",
"0.2197",
"0.8413"
]
] | [
"Document-level Relation Extraction with",
"Cross-sentence Reasoning Graph",
"Hongfei Liu', Zhao Kang! *, Lizong Zhang', Ling Tian', and Fujun Hua?",
"1 University of Electronic Science and Technology of China, Chengdu, China",
"[email protected]",
"[zkang,l.zhang,lingtian @uestc.edu.cn",
"2 Research and Development Center TROY Information Technology Co., Ltd.",
"Chengdu, China [email protected]",
"Abstract. Relation extraction (RE) has recently moved from the sentence-",
"level to document-level, which requires aggregating document informa-",
"tion and using entities and mentions for reasoning. Existing works put en-",
"tity nodes and mention nodes with similar representations in a document-",
"level graph, whose complex edges may incur redundant information.",
"Furthermore, existing studies only focus on entity-level reasoning paths",
"without considering global interactions among entities cross-sentence.",
"To these ends, we propose a novel document-level RE model with a",
"GR aph information A ggregation and C ross-sentence R easoning net-",
"work (GRACR). Specifically, a simplified document-level graph is con-",
"structed to model the semantic information of all mentions and sen-",
"tences in a document, and an entity-level graph is designed to explore",
"relations of long-distance cross-sentence entity pairs. Experimental re-",
"sults show that GRACR achieves excellent performance on two public",
"datasets of document-level RE. It is especially effective in extracting po",
"tential relations of cross-sentence entity pairs. Our code is available at",
"https://github.com/UESTC-LHF/GRACR",
"",
"Keywords: Deep learning Relation extraction : Document-level RE",
"1",
"Introduction",
"关系提取（RE）是识别文本中一对命名实体之间的语义关系。文档级RE要求",
"模型从文档中提取关系，并面临一些棘手的挑战。首先，一个文档包含多个句",
"子，因此关系提取任务需要处理更丰富和复杂的语义信息。其次，同一三元组",
"中的主语和宾语实体可能出现在不同的句子中，并且某些实体具有别名，这些",
"别名通常被称为实体提及。因此，文档级RE使用的信息可能不是来自一个句",
"子。第三，不同的三元组之间可能存在相互作用。从不同的三元组中提取两个",
"实体之间的关系需要使用上下文特征进行推理。图1显示了来自DocRED数据",
"集[21]的示例。因为主语和宾语出现在同一个句子中，所以很容易预测句子内",
"的关系。但是，在识别“瑞典语”和“瑞典皇家学院”之间的句子间关系时存在",
"问题，它们的提及分布在不同的句子中，并且存在长距离依赖关系。[21]提出了",
"* Corresponding author."
] | {
"arxiv_id": "2303.03912",
"language": "zh",
"page_idx": 0
} |
||
2303.03912-zh-1 | 2303.03912-zh-1.jpg | 2
H. Liu et al.
$$\begin{aligned}&\text{Iohan Gottleb Gaban}\\&\text{Silohan Gouttleb Gahn (13 August 1745-8 December 1818) was a Swedsh chermist and}\\&\text{metallourgist who Discovered manganese ln 1774... (571)n 1784, Gahn was electedca member of}\\&\text{the Royal Swededsh Academy of Scances. [58] He also ma a mangetlal career ln Sweds ming.}\\&\text{Titples:}\\&\text{<lohan Gottleb Gahn, country of cutzenshp, Swedsils}\\&\text{<lohan Gottleb Gahn, memy of Rowalswedsh Academy>}\\&\text{<Royal Swedentsh Acaderny, country, Sweedstrs}\end{aligned}$$
Fig.1: An example of document-level RE excerpted from DocRED dataset.
DocRED 数据集，其中包含大规模的人工注释文档，以促进句子级 RE 向文档级 RE 的发展。为了充分利用文档中复杂的语义信息，设计了文档级图，提出了基于图神经网络 (GNN) [4] [1] 提出了一种面向边缘的模型，该模型构造了具有不同类型节点和边缘的文档级图，以获得用于关系分类的全局表示。[12] 将文档级图定义为潜变量，并基于结构化注意进行诱导，通过优化文档级图的结构来提高文档级 RE 模型的性能。[17] 提出了一个模型，该模型通过文档级图学习实体的全局表示，并根据实体的上下文学习实体的局部表示。但是，这些模型只是对提及的嵌入进行平均以获得实体嵌人，并将其馈送到分类器中以获得关系标签。如果某个实体只有一个提及，则实体和提及节点共享类似的嵌人。因此，将它们放在同一图中会引人冗余信息并减少歧视
为了解决上述问题，我们提出了一种新颖的基于 GNN 的文档级 RE 模型， 该模型具有由来自文档的语义信息构造的两个图。我们的主要思想是构建文档级图和实体级图，以充分利用文档的语义信息以及跨句子的实体对之间的关系。具体来说，我们解决了两个问题：
首先，如何集成文档的丰富语义信息以获得实体表示？我们构造了一个文档级图来集成复杂的语义信息，这是一个包含提及节点和句子节点的异构图。通过预训练的语言模型 BERT [3]计算提及节点和句子节点的表示。将构建的文档级图输人到关系图神经网络 R-GCNs [13] 中，使节点包含其邻居节点的信息。然后，通过对提及节点的表示执行 logsumexp 池操作来获得实体的表示。在以前的方法中，实体节点的表示是从提及节点的表示中获得的。因此，将它们放在同一图中会引人冗余信息并降低可分辨性。与以前的文档级图构造不同，我们的文档级图只包含句子节点和提及节点，以避免重复节点表示引起的冗余信息
其次，如何使用实体之间的连接进行推理？在本文中，我们利用实体之间的连接并提出了用于推理的实体级图。实体级图是通过句子和实体之间的位置连接来构建的，以充分利用跨句信息。它连接长距离的交叉句子实体对。通过GNN 的学习，每个实体节点可以聚合其最相关的实体节点的信息，这有利于发现长距离跨句实体对的潜在关系
总而言之，我们为文档级 RE 提出了一种称为 GRACR 的新颖模型。我们的
主要贡献如下：
· 我们提出了一种简化的文档级图来集成丰富的语义信息。该图包含句子节点和提及节点，但不包含实体节点，这避免了由于重复节点表示而引人冗余信息
·我们提出了用于推理的实体级图，以发现长距离交叉句子实体对的潜在关
系。注意机制用于融合文档嵌人，聚合和推理信息以提取实体对的关系
· 我们在两个公共文档级关系提取数据集上进行实验。实验结果表明，我们
的模型优于许多最先进的方法 | [
[
"0.2173",
"0.11743",
"0.2524",
"0.11743",
"0.2524",
"0.1288",
"0.2173",
"0.1288"
],
[
"0.273",
"0.11743",
"0.353",
"0.11743",
"0.353",
"0.1276",
"0.273",
"0.1276"
],
[
"0.3472",
"0.1497",
"0.3694",
"0.1497",
"0.3694",
"0.1528",
"0.3472",
"0.1528"
],
[
"0.38",
"0.1497",
"0.3906",
"0.1497",
"0.3906",
"0.154",
"0.38",
"0.154"
],
[
"0.4387",
"0.1559",
"0.4714",
"0.1559",
"0.4714",
"0.1604",
"0.4387",
"0.1604"
],
[
"0.589",
"0.1572",
"0.6055",
"0.1572",
"0.6055",
"0.1604",
"0.589",
"0.1604"
],
[
"0.3406",
"0.1736",
"0.353",
"0.1736",
"0.353",
"0.1768",
"0.3406",
"0.1768"
],
[
"0.4207",
"0.1724",
"0.6685",
"0.1724",
"0.6685",
"0.1799",
"0.4207",
"0.1799"
],
[
"0.4756",
"0.1881",
"0.5293",
"0.1901",
"0.5293",
"0.2008",
"0.4756",
"0.1989"
],
[
"0.3245",
"0.202",
"0.3677",
"0.202",
"0.3677",
"0.209",
"0.3245",
"0.209"
],
[
"0.2336",
"0.2235",
"0.7695",
"0.2247",
"0.7695",
"0.238",
"0.2336",
"0.2374"
],
[
"0.2197",
"0.2798",
"0.785",
"0.2798",
"0.785",
"0.2935",
"0.2197",
"0.2935"
],
[
"0.2214",
"0.2966",
"0.7817",
"0.2966",
"0.7817",
"0.307",
"0.2214",
"0.307"
],
[
"0.2197",
"0.31",
"0.7837",
"0.31",
"0.7837",
"0.324",
"0.2197",
"0.324"
],
[
"0.2197",
"0.3245",
"0.785",
"0.3245",
"0.785",
"0.3384",
"0.2197",
"0.3384"
],
[
"0.2197",
"0.3416",
"0.7837",
"0.3416",
"0.7837",
"0.3516",
"0.2197",
"0.3516"
],
[
"0.2197",
"0.3547",
"0.785",
"0.3547",
"0.785",
"0.3687",
"0.2197",
"0.3687"
],
[
"0.2214",
"0.3694",
"0.7837",
"0.3694",
"0.7837",
"0.3833",
"0.2214",
"0.3833"
],
[
"0.2197",
"0.385",
"0.785",
"0.385",
"0.785",
"0.3984",
"0.2197",
"0.3984"
],
[
"0.2197",
"0.401",
"0.785",
"0.401",
"0.785",
"0.414",
"0.2197",
"0.414"
],
[
"0.2214",
"0.4172",
"0.5435",
"0.4172",
"0.5435",
"0.428",
"0.2214",
"0.428"
],
[
"0.2467",
"0.4375",
"0.7817",
"0.4375",
"0.7817",
"0.4475",
"0.2467",
"0.4475"
],
[
"0.2214",
"0.4507",
"0.785",
"0.4507",
"0.785",
"0.464",
"0.2214",
"0.464"
],
[
"0.2214",
"0.4673",
"0.785",
"0.4673",
"0.785",
"0.4778",
"0.2214",
"0.4778"
],
[
"0.2181",
"0.4797",
"0.452",
"0.481",
"0.452",
"0.4944",
"0.2181",
"0.4937"
],
[
"0.2484",
"0.5024",
"0.7837",
"0.5024",
"0.7837",
"0.513",
"0.2484",
"0.513"
],
[
"0.2181",
"0.5166",
"0.7817",
"0.5166",
"0.7817",
"0.53",
"0.2181",
"0.53"
],
[
"0.2197",
"0.531",
"0.7837",
"0.531",
"0.7837",
"0.5444",
"0.2197",
"0.5444"
],
[
"0.2214",
"0.5454",
"0.7837",
"0.5454",
"0.7837",
"0.5596",
"0.2214",
"0.5596"
],
[
"0.2214",
"0.561",
"0.7837",
"0.561",
"0.7837",
"0.5747",
"0.2214",
"0.5747"
],
[
"0.2197",
"0.577",
"0.7837",
"0.577",
"0.7837",
"0.5903",
"0.2197",
"0.5903"
],
[
"0.232",
"0.5913",
"0.7837",
"0.5913",
"0.7837",
"0.605",
"0.232",
"0.605"
],
[
"0.2214",
"0.608",
"0.755",
"0.608",
"0.755",
"0.6187",
"0.2214",
"0.6187"
],
[
"0.2427",
"0.6255",
"0.7817",
"0.6255",
"0.7817",
"0.6396",
"0.2427",
"0.6396"
],
[
"0.2197",
"0.6416",
"0.7837",
"0.6416",
"0.7837",
"0.655",
"0.2197",
"0.655"
],
[
"0.2222",
"0.6577",
"0.7837",
"0.6577",
"0.7837",
"0.6685",
"0.2222",
"0.6685"
],
[
"0.2197",
"0.672",
"0.785",
"0.672",
"0.785",
"0.685",
"0.2197",
"0.685"
],
[
"0.2197",
"0.6875",
"0.4453",
"0.6875",
"0.4453",
"0.6978",
"0.2197",
"0.6978"
],
[
"0.2443",
"0.707",
"0.7837",
"0.707",
"0.7837",
"0.72",
"0.2443",
"0.72"
],
[
"0.2214",
"0.7236",
"0.327",
"0.7236",
"0.327",
"0.7344",
"0.2214",
"0.7344"
],
[
"0.246",
"0.7437",
"0.7837",
"0.7437",
"0.7837",
"0.754",
"0.246",
"0.754"
],
[
"0.2214",
"0.7583",
"0.7837",
"0.7583",
"0.7837",
"0.768",
"0.2214",
"0.768"
],
[
"0.246",
"0.7764",
"0.785",
"0.7764",
"0.785",
"0.7905",
"0.246",
"0.7905"
],
[
"0.2214",
"0.7915",
"0.7256",
"0.7915",
"0.7256",
"0.8047",
"0.2214",
"0.8047"
],
[
"0.246",
"0.8125",
"0.785",
"0.8125",
"0.785",
"0.8257",
"0.246",
"0.8257"
],
[
"0.2181",
"0.827",
"0.4307",
"0.827",
"0.4307",
"0.8403",
"0.2181",
"0.8403"
]
] | [
"2",
"H. Liu et al.",
"GOU",
"Gd",
"gust174",
"UIST",
"OV",
"y of Sciences. [58] He also made a managerial career in Swedish min",
"hipswedshd",
"<RoyalSwec",
"Fig.l:An example of document-level RE excerpted fromDocRED dataset.",
"DocRED数据集，其中包含大规模的人工注释文档，以促进句子级RE向文档",
"级RE的发展。为了充分利用文档中复杂的语义信息，设计了文档级图，提出了",
"基于图神经网络（GNN[4][1]提出了一种面向边缘的模型，该模型构造了具有",
"不同类型节点和边缘的文档级图，以获得用于关系分类的全局表示。[12]将文档",
"级图定义为潜变量，并基于结构化注意进行诱导，通过优化文档级图的结构来",
"提高文档级RE模型的性能。[17]提出了一个模型，该模型通过文档级图学习实",
"体的全局表示，并根据实体的上下文学习实体的局部表示。但是，这些模型只是",
"对提及的嵌入进行平均以获得实体嵌入，并将其馈送到分类器中以获得关系标",
"签。如果某个实体只有一个提及，则实体和提及节点共享类似的嵌人。因此，将",
"它们放在同一图中会引入允余信息并减少歧视",
"为了解决上述问题，我们提出了一种新颖的基于GNN的文档级RE模型",
"该模型具有由来自文档的语义信息构造的两个图。我们的主要思想是构建文档",
"级图和实体级图，以充分利用文档的语义信息以及跨句子的实体对之间的关系。",
"具体来说，我们解决了两个问题",
"首先，如何集成文档的丰富语义信息以获得实体表示？我们构造了一个文档",
"级图来集成复杂的语义信息，这是一个包含提及节点和句子节点的异构图。通过",
"预训练的语言模型BERT[3]计算提及节点和句子节点的表示。将构建的文档级",
"图输入到关系图神经网络R-GCNs[13]中，使节点包含其邻居节点的信息。然",
"后，通过对提及节点的表示执行logsumexp池操作来获得实体的表示。在以前",
"的方法中，实体节点的表示是从提及节点的表示中获得的。因此，将它们放在同",
"图中会引入穴余信息并降低可分辨性。与以前的文档级图构造不同，我们的",
"文档级图只包含句子节点和提及节点，以避免重复节点表示引起的元余信息",
"其次，如何使用实体之间的连接进行推理？在本文中，我们利用实体之间",
"的连接并提出了用于推理的实体级图。实体级图是通过句子和实体之间的位置",
"连接来构建的，以充分利用跨句信息。它连接长距离的交叉句子实体对。通过",
"GNN的学习，每个实体节点可以聚合其最相关的实体节点的信息，这有利于发",
"现长距离跨句实体对的潜在关系",
"总而言之，我们为文档级RE提出了一种称为GRACR的新颖模型。我们的",
"主要贡献如下：",
"·我们提出了一种简化的文档级图来集成丰富的语义信息。该图包含句子节",
"点和提及节点，但不包含实体节点，这避免了由于重复节点表示而引入穴余信息",
"：我们提出了用于推理的实体级图，以发现长距离交叉句子实体对的潜在关",
"系。注意机制用于融合文档嵌人，聚合和推理信息以提取实体对的关系",
"·我们在两个公共文档级关系提取数据集上进行实验。实验结果表明，我们",
"的模型优于许多最先进的方法"
] | {
"arxiv_id": "2303.03912",
"language": "zh",
"page_idx": 1
} |
||
2303.03912-zh-2 | 2303.03912-zh-2.jpg | Document-level Relation Extraction with Cross-sentence Reasoning Graph 3
# 2 Related Work
文档级 RE 的研究由来已久。文档级图形为实体对提供了更多功能。可以使用GNN [10] 通过图学习来捕获实体之间的相关性。例如，[2] 利用 GNN 聚合文本图节点的邻域信息进行文本分类。在此之后，[1] 构造了一个具有异构节点的文档级图，并提出了一个面向边缘的模型来获得全局表示。[7] 表征了句子和实体对之间的相互作用，以改善句子间的推理。[25]引人了实体对的上下文作为实体节点之间的边，以对多个实体之间的语义交互进行建模。[24]构造了一个双层异构图，用于编码文档的固有结构并推理实体的多跳关系。[17]通过文档级图学习实体的全局表示，并根据其上下文学习局部表示。[12]将文档级图定义为潜在变量，以通过优化文档级图的结构来提高 RE 模型的性能。[23] 提出了一种基于双图的图聚合和推理网络 (GAIN)。与 GAIN 不同，我们的实体级图是一个异构图，我们使用 R-GCN 来实现实体节点之间的交互，以发现长距离交叉句子实体对的潜在关系 [18] 用修辞结构理论构造了一个文档级图，并用证据进行推理。[14]将输入文档构造为异构图，并利用图转换器网络生成语义路径。与上述文档级图构造方法不同，我们的文档级图只包含句子节点和提及节点，以避免引人冗余信息。此外，以前的作品没有直接处理跨句子实体对。虽然不同句子中的实体在图中是间接连接的，但。g。,跨句子的实体之间的最小距离为 3, 并且在 GLRE [17] 中进行交互时，信息需要通过两个不同的节点。我们通过桥接实体直接连接具有潜在关系的交叉句子实体对，以缩短信息传输的距离，从而减少了噪声的引人
此外，还有一些作品尝试直接使用预训练的模型，而不是引入图结构。[16] 应用分层推理方法聚合不同粒度的推理信息。[22]通过预训练任务捕获了上下文中的相互参照关系。[9]提出了一个基于提及的推理网络来捕获本地和全局上下文信息。[20]利用提及依赖关系构建结构化自我注意机制。[26]提出了自适应阈值和局部上下文池，以解决多标签和多实体问题。这些模型利用 Transformer 而不是 GNN 的多头注意力来聚合信息
但是，这些研究集中在局部实体表示上，而忽略了分布在不同句子[11]中的实体之间的相互作用。为了发现长距离交叉句子实体对的潜在关系，我们引入了一个由句子和实体之间的位置连接构建的实体级图，用于推理
# 3 Methodology
在本节中，我们描述了我们提出的 GRACR 模型，该模型构造了文档级图和实体级图，以改进文档级 RE。如图 2 所示，GRACR 主要由 4 个模块组成：编码模块，文档级图聚合模块，实体级图推理模块，分类模块。首先，在编码模块中， 我们使用预训练的语言模型，如 BERT [3]来编码文档。接下来，在文档级图聚合模块中，我们构建一个包含提及节点和句子节点的异构图，以集成文档的丰富语义信息。然后，在实体级图推理模块中，我们还提出了一个用于推理的图， 以发现长距离和跨句子实体对的潜在关系。最后，在分类模块中，我们合并通过自关注[15]获得的关系表示的上下文信息，以进行最终的关系预测
# 3.1 Encoding Module
为了更好地捕获文档的语义信息，我们选择 BERT 作为编码器。给定一个输入文档 D$=[w_1,w_2,\ldots,w_k]$ ,其中$w_j(1\leq j\leq k)$是其中的$j^th$字。然后，我们 | [
[
"0.2336",
"0.11615",
"0.7295",
"0.11615",
"0.7295",
"0.13",
"0.2336",
"0.13"
],
[
"0.738",
"0.1206",
"0.7427",
"0.1206",
"0.7427",
"0.1256",
"0.738",
"0.1256"
],
[
"0.773",
"0.1187",
"0.786",
"0.1187",
"0.786",
"0.1276",
"0.773",
"0.1276"
],
[
"0.2173",
"0.1477",
"0.3718",
"0.1477",
"0.3718",
"0.161",
"0.2173",
"0.161"
],
[
"0.2197",
"0.1755",
"0.7837",
"0.1755",
"0.7837",
"0.1895",
"0.2197",
"0.1895"
],
[
"0.2197",
"0.1913",
"0.785",
"0.1913",
"0.785",
"0.2052",
"0.2197",
"0.2052"
],
[
"0.2214",
"0.2084",
"0.7837",
"0.2084",
"0.7837",
"0.2184",
"0.2214",
"0.2184"
],
[
"0.2197",
"0.2203",
"0.7837",
"0.2203",
"0.7837",
"0.2343",
"0.2197",
"0.2343"
],
[
"0.2214",
"0.238",
"0.7817",
"0.238",
"0.7817",
"0.2488",
"0.2214",
"0.2488"
],
[
"0.2197",
"0.252",
"0.7837",
"0.252",
"0.7837",
"0.2651",
"0.2197",
"0.2651"
],
[
"0.2214",
"0.2683",
"0.7817",
"0.2683",
"0.7817",
"0.279",
"0.2214",
"0.279"
],
[
"0.2214",
"0.2827",
"0.7817",
"0.2827",
"0.7817",
"0.2935",
"0.2214",
"0.2935"
],
[
"0.2197",
"0.2966",
"0.7837",
"0.2966",
"0.7837",
"0.31",
"0.2197",
"0.31"
],
[
"0.2181",
"0.31",
"0.7817",
"0.31",
"0.7817",
"0.324",
"0.2181",
"0.324"
],
[
"0.2197",
"0.327",
"0.7837",
"0.327",
"0.7837",
"0.3403",
"0.2197",
"0.3403"
],
[
"0.2214",
"0.3435",
"0.7817",
"0.3435",
"0.7817",
"0.3542",
"0.2214",
"0.3542"
],
[
"0.2173",
"0.3567",
"0.785",
"0.356",
"0.785",
"0.3694",
"0.2173",
"0.3706"
],
[
"0.2214",
"0.3726",
"0.7837",
"0.3726",
"0.7837",
"0.3865",
"0.2214",
"0.3865"
],
[
"0.2181",
"0.3865",
"0.7817",
"0.3865",
"0.7817",
"0.3997",
"0.2181",
"0.3997"
],
[
"0.2197",
"0.4028",
"0.7837",
"0.4028",
"0.7837",
"0.4167",
"0.2197",
"0.4167"
],
[
"0.2197",
"0.4172",
"0.7837",
"0.4172",
"0.7837",
"0.4312",
"0.2197",
"0.4312"
],
[
"0.2197",
"0.433",
"0.785",
"0.433",
"0.785",
"0.4463",
"0.2197",
"0.4463"
],
[
"0.2197",
"0.4475",
"0.3342",
"0.4475",
"0.3342",
"0.461",
"0.2197",
"0.461"
],
[
"0.2443",
"0.4622",
"0.785",
"0.4622",
"0.785",
"0.476",
"0.2443",
"0.476"
],
[
"0.2181",
"0.4778",
"0.7817",
"0.4778",
"0.7817",
"0.4912",
"0.2181",
"0.4912"
],
[
"0.2222",
"0.4944",
"0.7837",
"0.4944",
"0.7837",
"0.505",
"0.2222",
"0.505"
],
[
"0.2214",
"0.5083",
"0.7837",
"0.5083",
"0.7837",
"0.5215",
"0.2214",
"0.5215"
],
[
"0.2214",
"0.523",
"0.785",
"0.523",
"0.785",
"0.536",
"0.2214",
"0.536"
],
[
"0.2181",
"0.537",
"0.4958",
"0.537",
"0.4958",
"0.5503",
"0.2181",
"0.5503"
],
[
"0.2427",
"0.552",
"0.7837",
"0.553",
"0.7837",
"0.5664",
"0.2427",
"0.565"
],
[
"0.2197",
"0.5684",
"0.7837",
"0.5684",
"0.7837",
"0.582",
"0.2197",
"0.582"
],
[
"0.2214",
"0.584",
"0.6772",
"0.584",
"0.6772",
"0.5977",
"0.2214",
"0.5977"
],
[
"0.2141",
"0.619",
"0.3635",
"0.6206",
"0.3635",
"0.6353",
"0.2141",
"0.634"
],
[
"0.2197",
"0.6484",
"0.785",
"0.6484",
"0.785",
"0.662",
"0.2197",
"0.662"
],
[
"0.2197",
"0.6636",
"0.785",
"0.6636",
"0.785",
"0.677",
"0.2197",
"0.677"
],
[
"0.2197",
"0.6787",
"0.7837",
"0.6787",
"0.7837",
"0.6924",
"0.2197",
"0.6924"
],
[
"0.2197",
"0.6934",
"0.785",
"0.6934",
"0.785",
"0.707",
"0.2197",
"0.707"
],
[
"0.2197",
"0.709",
"0.7837",
"0.709",
"0.7837",
"0.7227",
"0.2197",
"0.7227"
],
[
"0.2214",
"0.7246",
"0.7837",
"0.7246",
"0.7837",
"0.738",
"0.2214",
"0.738"
],
[
"0.2214",
"0.7393",
"0.785",
"0.7393",
"0.785",
"0.7524",
"0.2214",
"0.7524"
],
[
"0.2214",
"0.755",
"0.6953",
"0.755",
"0.6953",
"0.768",
"0.2214",
"0.768"
],
[
"0.2181",
"0.789",
"0.3857",
"0.789",
"0.3857",
"0.8",
"0.2181",
"0.8"
],
[
"0.2197",
"0.811",
"0.7812",
"0.811",
"0.7812",
"0.8257",
"0.2197",
"0.8257"
],
[
"0.2197",
"0.827",
"0.785",
"0.827",
"0.785",
"0.8413",
"0.2197",
"0.8413"
]
] | [
"Document-level Relation Extraction with Cross-sentence Reasoning Graph",
"",
"3",
"2Related Work",
"文档级RE的研究由来已久。文档级图形为实体对提供了更多功能。可以使用",
"GNN[10]通过图学习来捕获实体之间的相关性。例如，[2]利用GNN聚合文本",
"图节点的邻域信息进行文本分类。在此之后，1构造广一个具有异构节点的文",
"档级图，并提出了一个面向边缘的模型来获得全局表示。[7]表征了句子和实体",
"对之间的相互作用，以改善句子间的推理。25引人了实体对的上下文作为实体",
"节点之间的边，以对多个实体之间的语义交互进行建模。[24]构造了一个双层异",
"构图，用于编码文档的固有结构并推理实体的多跳关系。17通过文档级图学习",
"实体的全局表示，并根据其上下文学习局部表示。112将文档级图定义为潜在变",
"量，以通过优化文档级图的结构来提高RE模型的性能。[23]提出了一种基于",
"双图的图聚合和推理网络（GAIN)。与GAIN不同，我们的实体级图是一个异",
"构图，我们使用R-GCN来实现实体节点之间的交互，以发现长距离交叉句子",
"实体对的潜在关系18用修辞结构理论构造了一个文档级图，并用证据进行推",
"理。[14]将输人文档构造为异构图，并利用图转换器网络生成语义路径。与上述",
"文档级图构造方法不同，我们的文档级图只包含句子节点和提及节点，以避免",
"引人几余信息。此外，以前的作品没有直接处理跨句子实体对。虽然不同句子中",
"的实体在图中是间接连接的，但。g。，跨句子的实体之间的最小距离为3，并且",
"在GLRE[17]中进行交互时，信息需要通过两个不同的节点。我们通过桥接实",
"体直接连接具有潜在关系的交叉句子实体对，以缩短信息传输的距离，从而减",
"少了噪声的引人",
"此外，还有一些作品尝试直接使用预训练的模型，而不是引人图结构。[16]",
"应用分层推理方法聚合不同粒度的推理信息。[22]通过预训练任务捕获了上下",
"文中的相互参照关系。9提出了二个基于提及的推理网络来捕获本地和全局上",
"下文信息。[20]利用提及依赖关系构建结构化自我注意机制。[26]提出了自适应",
"阈值和局部上下文池，以解决多标签和多实体问题。这些模型利用Transformer",
"而不是GNN的多头注意力来聚合信息",
"但是，这些研究集中在局部实体表示上，而忽略了分布在不同句子「11]中的",
"实体之间的相互作用。为了发现长距离交叉句子实体对的潜在关系，我们引人",
"了二个由句子和实体之间的位置连接构建的实体级图，用于推理",
"3Methodology",
"在本节中，我们描述了我们提出的GRACR模型，该模型构造了文档级图和实",
"体级图，以改进文档级RE。如图2所示，GRACR主要由4个模块组成：编码",
"模块，文档级图聚合模块，实体级图推理模块，分类模块。首先，在编码模块中",
"我们使用预训练的语言模型，如BERT[3]来编码文档。接下来，在文档级图聚",
"合模块中，我们构建一个包含提及节点和句子节点的异构图，以集成文档的丰",
"富语义信息。然后，在实体级图推理模块中，我们还提出了一个用于推理的图",
"以发现长距离和跨句子实体对的潜在关系。最后，在分类模块中，我们合并通过",
"自关注[15]获得的关系表示的上下文信息，以进行最终的关系预测",
"3.1EncodingModule",
"为了更好地捕获文档的语义信息，我们选择BERT作为编码器。给定一个输入",
"文档D=[wi,w2,...，wkl，其中w;(1≤j≤k)是其中的ith字。然后，我们"
] | {
"arxiv_id": "2303.03912",
"language": "zh",
"page_idx": 2
} |
||
2303.03912-zh-3 | 2303.03912-zh-3.jpg | 4
H. Liu et al.
December 1818) was
Swedish chemist and metallurgist who discovered manganese. (S7)
[S1] Johan Gottlieb Gahn (19 August 1745
In 1784, Gahn was elected a member of the Royal Swedish Academy of Sciences. [58] He also made a managerial career in Swedish..
<FigureHere>
Fig.2: Architecture of our proposed model.
将文档输人到 BERT 以获得嵌入：
$$H=[h_1,h_2,\ldots,h_k]=\mathrm{Encoder}([w_1,w_2,\ldots,w_k])$$
(1)
其中$h_j\in\mathbb{R}^{d_w}$是 BERT 最后一层输出的隐藏状态序列
为了累积来自提及元组的弱信号，我们采用 logsumexp 池化 [5] 来获取实体
$e_i$的嵌人$e_i^h$作为初始实体表示。
$$e_i^h=\log\sum_{j=1}^{N_{\boldsymbol{e}_i}}\exp\left(h_{\boldsymbol{m}_j^i}\right)$$
(2)
其中$m_j^i$是实体$e_i$的提及$m_j$ , $h_{\boldsymbol{m}_j^i}$是$m_j^i$的嵌人，$N_{\boldsymbol{e}_i}$是 D 中实体$e_i$的提
及次数
如公式(2)所示，logsumexp 池化通过在整个文档中累积其所有提及的嵌
人来为每个实体生成嵌人
3.2 Document-level Graph Aggregation Module
为了集成文档的丰富语义信息以获得实体表示，我们基于$H$构造了文档级图
(Dlg)
$\operatorname{Dlg}$有两种不同类型的节点：
$Sentence~nodes~,~$表示 D 中的句子。通过对所包含单词的表示进行平均来获得句子节点$s_i$的表示。我们连接一个节点类型表示$t_s\in\mathbb{R}^{d_t}$来区分节点类型。因此，$s_i$的表示形式是$\boldsymbol{h}_{s_i}=\left[\arg_{w_j\in s_i}\left(\boldsymbol{h}_j\right);\boldsymbol{t}_s\right]$,其中 [;] 是串联运算符
$Mention\textit{ nodes , 表 示 D 中 的 提 及 。通 过 对 构 成 提 及 的 单 词 的 表 示 进 行 平 均 }$ 来实现提及节点$m_i$的表示。我们连接一个节点类型表示$t_m\in\mathbb{R}^{d_t}$ 。与句子节点类似，$m_i$的表示形式为$h_m_i=\left[\arg_{w_j\in m_i}\left(\boldsymbol{h}_j\right);t_m\right]$
$\operatorname{Dlg}$中有三种类型的边： | [
[
"0.2181",
"0.11743",
"0.2296",
"0.11743",
"0.2296",
"0.127",
"0.2181",
"0.127"
],
[
"0.273",
"0.11615",
"0.3545",
"0.11615",
"0.3545",
"0.1276",
"0.273",
"0.1276"
],
[
"0.2361",
"0.1477",
"0.7686",
"0.1477",
"0.7686",
"0.1578",
"0.2361",
"0.1578"
],
[
"0.2345",
"0.1559",
"0.746",
"0.1559",
"0.746",
"0.1666",
"0.2345",
"0.1666"
],
[
"0.25",
"0.1768",
"0.2607",
"0.1768",
"0.2607",
"0.1881",
"0.25",
"0.1881"
],
[
"0.3218",
"0.1768",
"0.3325",
"0.1768",
"0.3325",
"0.1863",
"0.3218",
"0.1863"
],
[
"0.4453",
"0.1724",
"0.451",
"0.1724",
"0.451",
"0.178",
"0.4453",
"0.178"
],
[
"0.726",
"0.1768",
"0.7563",
"0.1768",
"0.7563",
"0.1843",
"0.726",
"0.1843"
],
[
"0.411",
"0.1831",
"0.4248",
"0.178",
"0.4314",
"0.1863",
"0.4182",
"0.1925"
],
[
"0.7295",
"0.1849",
"0.761",
"0.1849",
"0.761",
"0.1925",
"0.7295",
"0.1925"
],
[
"0.326",
"0.1925",
"0.3308",
"0.1925",
"0.3308",
"0.1989",
"0.326",
"0.1989"
],
[
"0.734",
"0.1932",
"0.7563",
"0.1932",
"0.7563",
"0.202",
"0.734",
"0.202"
],
[
"0.7236",
"0.209",
"0.761",
"0.209",
"0.761",
"0.2197",
"0.7236",
"0.2197"
],
[
"0.2378",
"0.1976",
"0.2688",
"0.1976",
"0.2688",
"0.2102",
"0.2378",
"0.2102"
],
[
"0.4004",
"0.2229",
"0.415",
"0.2229",
"0.415",
"0.2311",
"0.4004",
"0.2311"
],
[
"0.7256",
"0.2172",
"0.761",
"0.2172",
"0.761",
"0.2279",
"0.7256",
"0.2279"
],
[
"0.7295",
"0.226",
"0.7573",
"0.226",
"0.7573",
"0.2374",
"0.7295",
"0.2374"
],
[
"0.3064",
"0.2393",
"0.367",
"0.2412",
"0.367",
"0.2524",
"0.3064",
"0.2507"
],
[
"0.4355",
"0.2405",
"0.544",
"0.2412",
"0.544",
"0.2524",
"0.4355",
"0.252"
],
[
"0.5312",
"0.2361",
"0.544",
"0.2361",
"0.544",
"0.2393",
"0.5312",
"0.2393"
],
[
"0.607",
"0.2424",
"0.693",
"0.2424",
"0.693",
"0.2524",
"0.607",
"0.2524"
],
[
"0.728",
"0.2443",
"0.767",
"0.2443",
"0.767",
"0.2551",
"0.728",
"0.2551"
],
[
"0.3545",
"0.2551",
"0.4346",
"0.2551",
"0.4346",
"0.2651",
"0.3545",
"0.2651"
],
[
"0.728",
"0.2524",
"0.767",
"0.2524",
"0.767",
"0.2632",
"0.728",
"0.2632"
],
[
"0.4885",
"0.2568",
"0.5435",
"0.254",
"0.544",
"0.2651",
"0.4902",
"0.2683"
],
[
"0.339",
"0.2568",
"0.3545",
"0.2568",
"0.3545",
"0.2676",
"0.339",
"0.2676"
],
[
"0.63",
"0.2568",
"0.687",
"0.25",
"0.6914",
"0.2651",
"0.634",
"0.2727"
],
[
"0.7344",
"0.2615",
"0.7607",
"0.2615",
"0.7607",
"0.2727",
"0.7344",
"0.2727"
],
[
"0.3022",
"0.276",
"0.3906",
"0.276",
"0.3906",
"0.2874",
"0.3022",
"0.2874"
],
[
"0.455",
"0.276",
"0.501",
"0.276",
"0.501",
"0.2874",
"0.455",
"0.2874"
],
[
"0.5063",
"0.276",
"0.5415",
"0.276",
"0.5415",
"0.2886",
"0.5063",
"0.2886"
],
[
"0.596",
"0.276",
"0.6543",
"0.276",
"0.6543",
"0.2874",
"0.596",
"0.2874"
],
[
"0.6562",
"0.2766",
"0.6787",
"0.2766",
"0.6787",
"0.2854",
"0.6562",
"0.2854"
],
[
"0.7363",
"0.279",
"0.761",
"0.279",
"0.761",
"0.2874",
"0.7363",
"0.2874"
],
[
"0.3513",
"0.2903",
"0.3594",
"0.2966",
"0.353",
"0.3005",
"0.3464",
"0.2954"
],
[
"0.7295",
"0.286",
"0.767",
"0.286",
"0.767",
"0.2966",
"0.7295",
"0.2966"
],
[
"0.7344",
"0.2917",
"0.761",
"0.2954",
"0.759",
"0.308",
"0.732",
"0.305"
],
[
"0.2345",
"0.2986",
"0.312",
"0.2986",
"0.312",
"0.3093",
"0.2345",
"0.3093"
],
[
"0.4944",
"0.2998",
"0.556",
"0.3018",
"0.556",
"0.3132",
"0.4944",
"0.3113"
],
[
"0.2484",
"0.3125",
"0.312",
"0.3125",
"0.312",
"0.3225",
"0.2484",
"0.3225"
],
[
"0.7363",
"0.3132",
"0.763",
"0.3132",
"0.763",
"0.3213",
"0.7363",
"0.3213"
],
[
"0.3489",
"0.3193",
"0.4004",
"0.3193",
"0.4004",
"0.334",
"0.3489",
"0.334"
],
[
"0.4495",
"0.3208",
"0.544",
"0.3208",
"0.544",
"0.332",
"0.4495",
"0.332"
],
[
"0.2378",
"0.3145",
"0.2467",
"0.3145",
"0.2467",
"0.3213",
"0.2378",
"0.3213"
],
[
"0.415",
"0.3213",
"0.4355",
"0.3213",
"0.4355",
"0.33",
"0.415",
"0.33"
],
[
"0.7256",
"0.3225",
"0.767",
"0.3225",
"0.767",
"0.33",
"0.7256",
"0.33"
],
[
"0.25",
"0.327",
"0.2942",
"0.327",
"0.2942",
"0.334",
"0.25",
"0.334"
],
[
"0.2378",
"0.3276",
"0.2443",
"0.3276",
"0.2443",
"0.3333",
"0.2378",
"0.3333"
],
[
"0.738",
"0.33",
"0.7573",
"0.33",
"0.7573",
"0.3384",
"0.738",
"0.3384"
],
[
"0.3472",
"0.351",
"0.6562",
"0.351",
"0.6562",
"0.3643",
"0.3472",
"0.3643"
],
[
"0.2181",
"0.3965",
"0.4617",
"0.3965",
"0.4617",
"0.4104",
"0.2181",
"0.4104"
],
[
"0.3447",
"0.4224",
"0.6587",
"0.4224",
"0.6587",
"0.4375",
"0.3447",
"0.4375"
],
[
"0.763",
"0.4236",
"0.786",
"0.4236",
"0.786",
"0.4375",
"0.763",
"0.4375"
],
[
"0.2197",
"0.449",
"0.614",
"0.449",
"0.614",
"0.4634",
"0.2197",
"0.4634"
],
[
"0.2443",
"0.4653",
"0.785",
"0.4653",
"0.785",
"0.4792",
"0.2443",
"0.4792"
],
[
"0.2181",
"0.481",
"0.455",
"0.481",
"0.455",
"0.4944",
"0.2181",
"0.4944"
],
[
"0.4844",
"0.5083",
"0.501",
"0.5083",
"0.501",
"0.5176",
"0.4844",
"0.5176"
],
[
"0.4224",
"0.5176",
"0.536",
"0.521",
"0.536",
"0.535",
"0.4224",
"0.532"
],
[
"0.535",
"0.521",
"0.562",
"0.521",
"0.562",
"0.532",
"0.535",
"0.532"
],
[
"0.7646",
"0.5195",
"0.786",
"0.5195",
"0.786",
"0.534",
"0.7646",
"0.534"
],
[
"0.2197",
"0.5596",
"0.785",
"0.5596",
"0.785",
"0.5737",
"0.2197",
"0.5737"
],
[
"0.2197",
"0.5757",
"0.2695",
"0.5757",
"0.2695",
"0.5903",
"0.2197",
"0.5903"
],
[
"0.2443",
"0.5923",
"0.786",
"0.5923",
"0.786",
"0.607",
"0.2443",
"0.607"
],
[
"0.2197",
"0.608",
"0.398",
"0.608",
"0.398",
"0.622",
"0.2197",
"0.622"
],
[
"0.2173",
"0.648",
"0.576",
"0.648",
"0.576",
"0.662",
"0.2173",
"0.662"
],
[
"0.2197",
"0.674",
"0.785",
"0.674",
"0.785",
"0.6875",
"0.2197",
"0.6875"
],
[
"0.2197",
"0.69",
"0.2656",
"0.69",
"0.2656",
"0.702",
"0.2197",
"0.702"
],
[
"0.2427",
"0.705",
"0.4468",
"0.705",
"0.4468",
"0.7197",
"0.2427",
"0.7197"
],
[
"0.2427",
"0.7183",
"0.7837",
"0.7183",
"0.7837",
"0.733",
"0.2427",
"0.733"
],
[
"0.2197",
"0.735",
"0.785",
"0.735",
"0.785",
"0.7485",
"0.2197",
"0.7485"
],
[
"0.2181",
"0.7524",
"0.7812",
"0.7524",
"0.7812",
"0.771",
"0.2181",
"0.771"
],
[
"0.2427",
"0.773",
"0.7837",
"0.773",
"0.7837",
"0.787",
"0.2427",
"0.787"
],
[
"0.2197",
"0.787",
"0.7837",
"0.787",
"0.7837",
"0.802",
"0.2197",
"0.802"
],
[
"0.2173",
"0.8047",
"0.6045",
"0.8047",
"0.6045",
"0.8257",
"0.2173",
"0.8257"
],
[
"0.2427",
"0.8257",
"0.4126",
"0.8257",
"0.4126",
"0.839",
"0.2427",
"0.839"
]
] | [
"4",
"H. Liu et al.",
"[51] Johan Gottlieb Gahn (19 August 1745 -8 December 1818) was a Swedish chemist and metallurgist who discovered manganese. [S7]",
"(In 1784, Gahn was elected a member of the Royal Swedish Academy of Sciences. [S8] He also made a managerial career in Swedish.",
"+",
"A",
"2",
"Mentio",
"V",
"mention",
"",
"edge",
"Mention",
"BERT",
"7",
"sentence",
"edge",
"Embedding",
"Document-level graph",
"",
"Entity-level graph",
"Sentence-",
"Logsumexp-pool",
"sentence",
"R-GCN",
"+",
"+R-GCN",
"edge",
"",
"QO",
"",
"",
"",
"Intra-",
"",
"sentence",
"edge",
"Mention node",
"Attention",
"Sentence node",
"Logical",
"Classifier",
"",
"M",
"管",
"reasoning",
"Entitynod",
"",
"edge",
"Fig.2: Architecture of our proposed model.",
"将文档输入到BERT以获得嵌入：",
"H=[h1,h2,...,h]=Encoder([w1,w2,...,w])",
"(1)",
"其中h;ERdw是BERT最后一层输出的隐藏状态序列",
"为了累积来自提及元组的弱信号，我们采用logsumexp池化[5]来获取实体",
"ei的嵌人eh作为初始实体表示。",
"Ne",
"et =logexp",
"(hm",
"(2)",
"其中m，是实体ei的提及mi，hm是m的嵌人，Ne是D中实体e的提",
"及次数",
"如公式（2）所示，logsumexp池化通过在整个文档中累积其所有提及的嵌",
"入来为每个实体生成嵌人",
"3.2Document-level Graph Aggregation Module",
"为了集成文档的丰富语义信息以获得实体表示，我们基于H构造了文档级图",
"(Dlg )",
"Dlg有两种不同类型的节点：",
"Sentence nodes，表示D中的句子。通过对所包含单词的表示进行平均来",
"获得句子节点Si的表示。我们连接一个节点类型表示t。ERd来区分节点类",
"型。因此，si的表示形式是hs:=avgwjEs:（hj);ts，其中[是串联运算符",
"Mentionnodes，表示D中的提及。通过对构成提及的单词的表示进行平均",
"来实现提及节点m的表示。我们连接一个节点类型表示tmERdt。与句子节",
"点类似，mi的表示形式为hm.=avgujem.（hi);tm",
"Dlg中有三种类型的边："
] | {
"arxiv_id": "2303.03912",
"language": "zh",
"page_idx": 3
} |
||
2303.03912-zh-4 | 2303.03912-zh-4.jpg | Document-level Relation Extraction with Cross-sentence Reasoning Graph 5
·提及-提及边。为了利用提及对之间的共现依赖性，我们创建了一个提及边。如果两个不同实体的提及节点在同一句子中同时出现，则它们通过提及边连接
·提及句子边。创建提及句边缘以更好地捕获提及的上下文信息。如果提及
出现在句子中，则提及节点和句子节点通过提及-句子边连接
·句子-句子边。所有句子节点通过句子-句子边缘连接，以消除文档中句子
序列的影响，并促进句子间的交互
然后，我们使用 L 层堆叠的 R-GCNs [13] 来学习文档级图。与 GCN 相比，R-GCN 可以更好地对具有各种类型边的异构图进行建模。具体来说，其$(l+1)^{(th)}$ 层的节点正向更新定义如下：
$$\boldsymbol{n}_i^{l+1}=\sigma\left(\boldsymbol{W}_0^l\boldsymbol{n}_i^l+\sum_{x\in\mathrm{X}}\sum_{j\in\mathrm{N}_i^x}\frac{1}{|\mathrm{N}_i^x|}\boldsymbol{W}_x^l\boldsymbol{n}_j^l\right)$$
(3)
其中$\sigma$ (·) 表示激活函数，$\mathbb{N}_i^x$表示与边$x$链接的节点$i$的邻居集$,X$表示边
类型集。$\boldsymbol{W}_x^l,\boldsymbol{W}_0^l\in\mathbb{R}^{d_n\times d_n}$是可训练的参数矩阵，$d_n$是节点表示的维度
我们使用图卷积后提及节点的表示形式，通过 logsumexppooling 将实体节点$e_i$的初步表示形式计算为$e_i^\mathrm{pre}$ ,它在整个文档中结合了$e_i$的语义信息。然而，整个文档的信息不可避免地引人了噪声。我们采用注意力机制来融合实体的初始嵌人信息和语义信息，以减少噪音。具体来说，我们将实体表示$e_i^\mathrm{Dlg}$定义如下：
$$e_i^{\mathrm{Dlg}}=\mathrm{softmax}\left(\frac{\mathrm{e}_\mathrm{i}^{\mathrm{pre}}\boldsymbol{W}_i^{\mathrm{e}_\mathrm{i}^{\mathrm{pre}}}\left(e_i^h\boldsymbol{W}_i^{e_i^h}\right)^T}{\sqrt{d_{e_i^h}}}\right)e_i^h\boldsymbol{W}_i^{e_i^h}$$
(4)
和
$$e_i^{\text{pre}}=\log\sum_{j=1}^{N_{\textbf{e}_i}}\exp\left(n_{m_j^i}\right)$$
(5)
其中$W_i^{e_i^\mathrm{pre}}$ 和$W_i^{e_i^h}\in\mathbb{R}^{d_n\times d_n}$ 是可训练的参数矩阵。$n_{m_j^i}$ 是图卷积后的语义表
示。$d_{e_{i}^{h}}$是$e_i^{h}$的维度
3.3 Entity-level Graph Reasoning Module
为了发现长距离交叉句子实体对的潜在关系，我们引人了实体级图(Elg)推理
模块。Elg 只包含一种节点：
$Entity~node~,~表示~D~中的实体。实体节点~e_{i}~的表示是从由等式~(5)定$ 义的文档级图获得的。我们连接一个节点类型表示$t_e\in\mathbb{R}^{t_e}$ 。$e_i$的表示形式是$\boldsymbol{h}_{e_i}=[e_i^{pre};\boldsymbol{t}_e]$
$\operatorname{Elg}$中有两种边：
·句子内边。如果两个不同的实体在同一句子中同时出现，则它们通过句子内边缘连接。例如，如果存在表示为$m_i^{\boldsymbol{s}_1}\to s_1\to m_j^{\boldsymbol{s}_1}$的路径$PI_{i,j}$,则 Elg 使用句子内边缘连接实体节点$e_i$和$e_j$ 。$m_i^{\boldsymbol{s}_1}$和$m_j^{\boldsymbol{s}_1}$提到了实体对$<e_i,e_j>$, 它们出现在句子$s_1$中。“→”表示从实体节点$e_i$ 到 $e_j$的推理路径上的一个推理步骤 | [
[
"0.2345",
"0.11743",
"0.7295",
"0.11743",
"0.7295",
"0.1276",
"0.2345",
"0.1276"
],
[
"0.7363",
"0.1193",
"0.7607",
"0.1193",
"0.7607",
"0.127",
"0.7363",
"0.127"
],
[
"0.7695",
"0.1187",
"0.786",
"0.1187",
"0.786",
"0.1276",
"0.7695",
"0.1276"
],
[
"0.7607",
"0.1206",
"0.7715",
"0.1206",
"0.7715",
"0.125",
"0.7607",
"0.125"
],
[
"0.726",
"0.1206",
"0.739",
"0.1206",
"0.739",
"0.127",
"0.726",
"0.127"
],
[
"0.2467",
"0.1497",
"0.7837",
"0.1497",
"0.7837",
"0.1604",
"0.2467",
"0.1604"
],
[
"0.2214",
"0.1635",
"0.785",
"0.1635",
"0.785",
"0.1768",
"0.2214",
"0.1768"
],
[
"0.2214",
"0.1787",
"0.2524",
"0.1787",
"0.2524",
"0.1901",
"0.2214",
"0.1901"
],
[
"0.2443",
"0.1932",
"0.785",
"0.1932",
"0.785",
"0.207",
"0.2443",
"0.207"
],
[
"0.2197",
"0.209",
"0.649",
"0.209",
"0.649",
"0.2229",
"0.2197",
"0.2229"
],
[
"0.25",
"0.2247",
"0.7837",
"0.2247",
"0.7837",
"0.238",
"0.25",
"0.238"
],
[
"0.2181",
"0.2393",
"0.464",
"0.2393",
"0.464",
"0.2524",
"0.2181",
"0.2524"
],
[
"0.2427",
"0.2551",
"0.785",
"0.2551",
"0.785",
"0.2683",
"0.2427",
"0.2683"
],
[
"0.2197",
"0.2683",
"0.785",
"0.2695",
"0.785",
"0.2842",
"0.2197",
"0.2827"
],
[
"0.2181",
"0.2842",
"0.4983",
"0.2842",
"0.4983",
"0.2986",
"0.2181",
"0.2986"
],
[
"0.3628",
"0.3193",
"0.6284",
"0.3193",
"0.6284",
"0.3416",
"0.3628",
"0.3416"
],
[
"0.7646",
"0.327",
"0.786",
"0.327",
"0.786",
"0.3403",
"0.7646",
"0.3403"
],
[
"0.4844",
"0.339",
"0.5864",
"0.3276",
"0.591",
"0.3499",
"0.4885",
"0.361"
],
[
"0.2197",
"0.361",
"0.785",
"0.361",
"0.785",
"0.375",
"0.2197",
"0.375"
],
[
"0.2173",
"0.3706",
"0.744",
"0.3718",
"0.744",
"0.3926",
"0.2173",
"0.3914"
],
[
"0.2427",
"0.3914",
"0.785",
"0.3914",
"0.785",
"0.4048",
"0.2427",
"0.4048"
],
[
"0.2214",
"0.4072",
"0.785",
"0.4072",
"0.785",
"0.4204",
"0.2214",
"0.4204"
],
[
"0.2197",
"0.4204",
"0.785",
"0.4216",
"0.785",
"0.4363",
"0.2197",
"0.435"
],
[
"0.2157",
"0.4343",
"0.7876",
"0.432",
"0.7876",
"0.4526",
"0.2157",
"0.455"
],
[
"0.2222",
"0.4526",
"0.2754",
"0.4526",
"0.2754",
"0.464",
"0.2222",
"0.464"
],
[
"0.452",
"0.4702",
"0.5312",
"0.4702",
"0.5312",
"0.4849",
"0.452",
"0.4849"
],
[
"0.5195",
"0.4697",
"0.591",
"0.4673",
"0.5923",
"0.486",
"0.5215",
"0.488"
],
[
"0.3367",
"0.481",
"0.4387",
"0.483",
"0.4387",
"0.4976",
"0.3367",
"0.4956"
],
[
"0.6167",
"0.4797",
"0.6665",
"0.4797",
"0.6665",
"0.4988",
"0.6167",
"0.4988"
],
[
"0.7646",
"0.4849",
"0.786",
"0.4849",
"0.786",
"0.5",
"0.7646",
"0.5"
],
[
"0.5063",
"0.4944",
"0.5474",
"0.4924",
"0.5483",
"0.5146",
"0.5073",
"0.5166"
],
[
"0.2181",
"0.5215",
"0.2378",
"0.5215",
"0.2378",
"0.535",
"0.2181",
"0.535"
],
[
"0.492",
"0.537",
"0.5073",
"0.537",
"0.5073",
"0.547",
"0.492",
"0.547"
],
[
"0.4224",
"0.549",
"0.5415",
"0.549",
"0.5415",
"0.563",
"0.4224",
"0.563"
],
[
"0.544",
"0.553",
"0.5605",
"0.553",
"0.5605",
"0.561",
"0.544",
"0.561"
],
[
"0.7646",
"0.55",
"0.785",
"0.55",
"0.785",
"0.563",
"0.7646",
"0.563"
],
[
"0.4878",
"0.563",
"0.511",
"0.563",
"0.511",
"0.5757",
"0.4878",
"0.5757"
],
[
"0.2173",
"0.582",
"0.786",
"0.5854",
"0.786",
"0.603",
"0.2173",
"0.5996"
],
[
"0.2214",
"0.605",
"0.3733",
"0.605",
"0.3733",
"0.619",
"0.2214",
"0.619"
],
[
"0.2181",
"0.646",
"0.5317",
"0.646",
"0.5317",
"0.659",
"0.2181",
"0.659"
],
[
"0.2222",
"0.675",
"0.7837",
"0.675",
"0.7837",
"0.685",
"0.2222",
"0.685"
],
[
"0.2181",
"0.688",
"0.4224",
"0.688",
"0.4224",
"0.702",
"0.2181",
"0.702"
],
[
"0.246",
"0.7026",
"0.785",
"0.7026",
"0.785",
"0.7163",
"0.246",
"0.7163"
],
[
"0.2214",
"0.7183",
"0.7837",
"0.7183",
"0.7837",
"0.732",
"0.2214",
"0.732"
],
[
"0.2181",
"0.7344",
"0.3228",
"0.7344",
"0.3228",
"0.7485",
"0.2181",
"0.7485"
],
[
"0.2427",
"0.75",
"0.361",
"0.75",
"0.361",
"0.763",
"0.2427",
"0.763"
],
[
"0.2467",
"0.7646",
"0.7837",
"0.7646",
"0.7837",
"0.778",
"0.2467",
"0.778"
],
[
"0.2214",
"0.78",
"0.786",
"0.78",
"0.786",
"0.7944",
"0.2214",
"0.7944"
],
[
"0.2214",
"0.7954",
"0.7754",
"0.7954",
"0.7754",
"0.81",
"0.2214",
"0.81"
],
[
"0.2181",
"0.811",
"0.7817",
"0.811",
"0.7817",
"0.8247",
"0.2181",
"0.8247"
],
[
"0.2181",
"0.8257",
"0.2695",
"0.8257",
"0.2695",
"0.8403",
"0.2181",
"0.8403"
]
] | [
"Document-levelRelationExtractionwithCross-sentenceReasoning Graph",
"",
"5",
"",
"",
"·提及-提及边。为了利用提及对之间的共现依赖性，我们创建了一个提及",
"边。如果两个不同实体的提及节点在同一句子中同时出现，则它们通过提及边",
"连接",
"·提及句子边。创建提及句边缘以更好地捕获提及的上下文信息。如果提及",
"出现在句子中，则提及节点和句子节点通过提及-句子边连接",
"，句子-句子边。所有句子节点通过句子-句子边缘连接，以消除文档中句子",
"序列的影响，并促进句子间的交互",
"然后，我们使用L层堆叠的R-GCNs[13]来学习文档级图。与GCN相",
"比，R-GCN可以更好地对具有各种类型边的异构图进行建模。具体来说，其",
"(l+1)（th）层的节点正向更新定义如下：",
"nt+i=a(won+ENWin)",
"(3)",
"FeXJNNT",
"其中α（）表示激活函数，N表示与边链接的节点i的邻居集，X表示边",
"类型集。Wl,w。e Rdnxd是可训练的参数矩阵，dn是节点表示的维度",
"我们使用图卷积后提及节点的表示形式，通过logsumexppooling将实体节",
"点ei的初步表示形式计算为epre，它在整个文档中结合了ei的语义信息。然",
"而，整个文档的信息不可避免地引人了噪声。我们采用注意力机制来融合实体",
"的初始嵌人信息和语义信息，以减少噪音。具体来说，我们将实体表示eP定",
"义如下：",
"eprewep\"",
"(etws)",
"eDlg = softmax",
"erwet",
"(4)",
"Vaet",
"和",
"Ne",
"epre=logexp",
"n",
"(5)",
"3=1",
"其中Wep和We\"eRdnxdn是可训练的参数矩阵。nm,是图卷积后的语义表",
"示。de是eh的维度",
"3.3Entity-level Graph Reasoning Module",
"为了发现长距离交叉子实体对的潜在关系，我们引入了实体级图（EIg）推理",
"模块。Elg只包含一种节点：",
"Entitynode：表示D中的实体。实体节点ei的表示是从由等式（5）定",
"义的文档级图获得的。我们连接一个节点类型表示teERte。ei的表示形式是",
"her =[epre;te]",
"Elg中有两种边：",
"句子内边。如果两个不同的实体在同一句子中同时出现，则它们通过句子",
"内边缘连接。例如，如果存在表示为m1→8i→m1的路径PI，则Elg",
"使用句子内边缘连接实体节点e,和ei。m和m提到了实体对<ei，ej>",
"它们出现在句子si中。“→”表示从实体节点ei到e,的推理路径上的一个推",
"理步骤"
] | {
"arxiv_id": "2303.03912",
"language": "zh",
"page_idx": 4
} |
||
1811.12156-zh-0 | 1811.12156-zh-0.jpg | # 用于多层网络推理的改进的深度嵌人
Huan Song$^1$ and Jayaraman J. Thiagarajan$^2*$
$^{1}$Bosch Research North America
$^{2}$Lawrence Livermore National Labs
[email protected], [email protected]
## Abstract
Inferencing with network data necessitates the mapping of its nodes into a vector space, where the relationships are preserved. However, with multi-layered networks, where multiple types of relationships exist for the same set of nodes, it is crucial to exploit the information shared between layers, in addition to the distinct aspects of each layer. In this paper, we propose a novel approach that frst obtains node embeddings in all layers jointly via DeepWalk on a supra graph, which allows interactions between layers, and then fne-tunes the embeddings to encourage cohesive structure in the latent space. With empirical studies in node classification, link prediction and multilayered community detection, we show that the proposed approach outperforms existing single- and multi-layered network embedding algorithms on several benchmarks. In addition to effectively scaling to a large number of layers (tested up to 37 ), our approach consistently produces highly modular community structure, even when compared to methods that directly optimize for the modularity function.
# 1 Introduction
网络数据的建模和推理对于包括社交网络分析[Eagle and Pentland, 2006] , 推荐系统 [Rao et al., 2015] 和医疗保健[Fornito et al., 2013] 在内的多个应用领域至关重要。这广泛地涵盖了诸如节点分类，链接预测，社区检测，有影响力的节点检测等问题。它们通常通过将节点映射到潜在的低维向量空间中，并在生成的嵌人上调用适当的机器学习技术来解决，例如。g。无监督聚类的社区检测 [Ding et al., 2016] 。在最简单的形式中，这样的嵌人可以通过网络结构的描述符来推断-e。g。图邻接、图拉普拉斯算子或模块化矩阵 [Chen et al.,2014; Yang et al., 2016] 。虽然最早的解决方案依赖于矩阵分解来优化嵌人，但最近发现深度图自动编码器更加灵
$^{*}$This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Lab- $\hat{\text{oratory under Contract DE- AC52-07NA27344.}}$
活 [Yang et al., 2016; Thiagarajan et al., 2016]。然而，缺乏可扩展性一直是这些方法的持续限制。这激发了诸如 DeepWalk$\underline{\mathrm{F}}[$Perozzi et al.,2014] 和 Node2Vec$\boxed{\mathrm{F}}$ [Grover and Leskovec, 2016] 之类的最新方法的设计，这些方法基于分布假设，在语言建模$\boxed{\mathrm{E}}[$Harris,1954] 中普遍采用，其中两个节点在短随机游走中的共现暗示了强烈的语义相似性概念。在负采样$\mathbb{E}[$Mikolov et al., 2013] 等高效优化策略的支持下，这些方法可以轻松处理超大规模网络
最近，随着多层网络的出现，我。e。,在每个层中具有不同关系结构的同一组节点[Li et al.,2018],获得保留多视图关系的简洁嵌人变得至关重要。与单层情况相反， 多层网络的节点嵌入技术必须适应两层的提取(例如。g。社区检测)以及统一表示(e。g。节点分类)。由于多层网络中固有的挑战，包括关系类型的异构性和不同层中稀疏程度的变化，单层方法的直接扩展是无效的 [Zhang et al., 2018]。最后，关于增加层数的可扩展性是一个重要的设计标准
在本文中，我们开发了一种可扩展的嵌入方法，该方法首先直接在多层网络(M-DeepWalk)上执行 DeepWalk 风格的优化，并利用细化策略通过鼓励内聚社区的形成来进一步微调嵌人。我们研究了我们的方法在链接预测， 节点分类和社区检测问题中的应用。令人惊讶的是，通过包含多层依赖关系，我们发现 M-DeepWalk 在链接预测方面已经可以超越最先进的多层嵌入方法[Zhang et al.,2018]。但是，随着节点分类和多层社区检测等更具挑战性的任务，还有改进的空间。使用短随机游走来推断底层语义结构的前提依赖于网络高度稀疏且节点共现遵循幂律的假设。但是，通过允许层间边缘，在语义持续更长时间的情况下，可能会违反该假设。我们通过包括一个改进步骤来解决这个限制，其中来自 M-DeepWalk 的嵌人被微调以产生更有凝聚力的社区。即使网络数据集包含多达 37 层，我们也表明，与现有的多层社区检测技术相比，所提出的方法产生了高度模块化的社区结构。
### 2 Problem Definition
我们将单层无向，未加权的网络表示为$G=(\mathcal{V},\mathcal{E})$,其中$\mathcal{V}$表示基数为$|\mathcal{V}|=N$的节点集，$\mathcal{E}$表示边集。嵌入技术的目标是为节点$X\in\mathbb{R}^N\times d$生成潜在表示，其中$d$ 是所需的嵌人维数。相应地，我们使用一组$L$相互依赖的网络$\mathcal{G}^(l)=(\mathcal{V}^{(l)},\mathcal{E}^{(l)})$,for $l=1,\ldots,L$表示多层网络其中在每对层之间存在一个节点映射$\mathcal{M}$,以指示一个网络中的哪些节点对应于另一个网络中的节点。请注意， 我们期望不同的层包含公共节点，但不要求$\mathcal{V}^{(l)}$相同。 | [
[
"0.3162",
"0.11554",
"0.683",
"0.11554",
"0.683",
"0.1288",
"0.3162",
"0.1288"
],
[
"0.3064",
"0.159",
"0.687",
"0.159",
"0.687",
"0.1736",
"0.3064",
"0.1736"
],
[
"0.3652",
"0.1799",
"0.634",
"0.1799",
"0.634",
"0.1945",
"0.3652",
"0.1945"
],
[
"0.3472",
"0.1963",
"0.6504",
"0.1976",
"0.6504",
"0.2122",
"0.3472",
"0.2108"
],
[
"0.3145",
"0.2172",
"0.683",
"0.2172",
"0.683",
"0.2311",
"0.3145",
"0.2311"
],
[
"0.2484",
"0.2715",
"0.3245",
"0.2734",
"0.3245",
"0.289",
"0.2484",
"0.2874"
],
[
"0.513",
"0.2734",
"0.91",
"0.2734",
"0.91",
"0.2874",
"0.513",
"0.2874"
],
[
"0.5156",
"0.289",
"0.546",
"0.289",
"0.546",
"0.3005",
"0.5156",
"0.3005"
],
[
"0.539",
"0.2886",
"0.911",
"0.2886",
"0.911",
"0.3018",
"0.539",
"0.3018"
],
[
"0.11926",
"0.2974",
"0.4534",
"0.2974",
"0.4534",
"0.3125",
"0.11926",
"0.3125"
],
[
"0.517",
"0.3018",
"0.9126",
"0.3018",
"0.9126",
"0.315",
"0.517",
"0.315"
],
[
"0.11926",
"0.3125",
"0.4534",
"0.3125",
"0.4534",
"0.3257",
"0.11926",
"0.3257"
],
[
"0.5156",
"0.3164",
"0.911",
"0.3164",
"0.911",
"0.33",
"0.5156",
"0.33"
],
[
"0.11926",
"0.3257",
"0.452",
"0.3257",
"0.452",
"0.339",
"0.11926",
"0.339"
],
[
"0.5146",
"0.33",
"0.911",
"0.33",
"0.911",
"0.3435",
"0.5146",
"0.3435"
],
[
"0.1177",
"0.339",
"0.452",
"0.3403",
"0.452",
"0.3542",
"0.1177",
"0.3528"
],
[
"0.5146",
"0.3435",
"0.91",
"0.3435",
"0.91",
"0.3567",
"0.5146",
"0.3567"
],
[
"0.11926",
"0.3542",
"0.452",
"0.3542",
"0.452",
"0.3674",
"0.11926",
"0.3674"
],
[
"0.513",
"0.356",
"0.91",
"0.356",
"0.91",
"0.3694",
"0.513",
"0.3694"
],
[
"0.1177",
"0.3674",
"0.452",
"0.3674",
"0.452",
"0.3806",
"0.1177",
"0.3806"
],
[
"0.513",
"0.3694",
"0.9126",
"0.3706",
"0.9126",
"0.385",
"0.513",
"0.3838"
],
[
"0.1177",
"0.3806",
"0.451",
"0.3806",
"0.451",
"0.3945",
"0.1177",
"0.3945"
],
[
"0.5156",
"0.385",
"0.5786",
"0.385",
"0.5786",
"0.3965",
"0.5156",
"0.3965"
],
[
"0.1177",
"0.3958",
"0.451",
"0.3958",
"0.451",
"0.4092",
"0.1177",
"0.4092"
],
[
"0.5293",
"0.3984",
"0.911",
"0.3984",
"0.911",
"0.4124",
"0.5293",
"0.4124"
],
[
"0.1177",
"0.4092",
"0.452",
"0.4077",
"0.452",
"0.4216",
"0.1177",
"0.4224"
],
[
"0.513",
"0.411",
"0.911",
"0.4124",
"0.911",
"0.4268",
"0.513",
"0.4255"
],
[
"0.11926",
"0.4248",
"0.452",
"0.4248",
"0.452",
"0.435",
"0.11926",
"0.435"
],
[
"0.513",
"0.4248",
"0.91",
"0.4268",
"0.91",
"0.44",
"0.513",
"0.4382"
],
[
"0.1177",
"0.4375",
"0.452",
"0.4375",
"0.452",
"0.4507",
"0.1177",
"0.4507"
],
[
"0.513",
"0.4395",
"0.911",
"0.4414",
"0.911",
"0.4558",
"0.513",
"0.4539"
],
[
"0.1177",
"0.4507",
"0.451",
"0.4507",
"0.451",
"0.464",
"0.1177",
"0.464"
],
[
"0.5146",
"0.455",
"0.91",
"0.455",
"0.91",
"0.4685",
"0.5146",
"0.4685"
],
[
"0.1177",
"0.4622",
"0.4534",
"0.464",
"0.4534",
"0.4792",
"0.1177",
"0.4766"
],
[
"0.5146",
"0.4685",
"0.911",
"0.4685",
"0.911",
"0.4817",
"0.5146",
"0.4817"
],
[
"0.1177",
"0.4778",
"0.452",
"0.4778",
"0.452",
"0.4912",
"0.1177",
"0.4912"
],
[
"0.511",
"0.481",
"0.9126",
"0.4817",
"0.9126",
"0.4968",
"0.511",
"0.4956"
],
[
"0.1209",
"0.4937",
"0.451",
"0.4937",
"0.451",
"0.504",
"0.1209",
"0.504"
],
[
"0.511",
"0.4944",
"0.9126",
"0.4956",
"0.9126",
"0.5103",
"0.511",
"0.509"
],
[
"0.1177",
"0.506",
"0.4534",
"0.506",
"0.4534",
"0.5195",
"0.1177",
"0.5195"
],
[
"0.5156",
"0.511",
"0.614",
"0.511",
"0.614",
"0.5215",
"0.5156",
"0.5215"
],
[
"0.1168",
"0.518",
"0.4534",
"0.521",
"0.4534",
"0.534",
"0.1168",
"0.532"
],
[
"0.5312",
"0.5234",
"0.911",
"0.5234",
"0.911",
"0.537",
"0.5312",
"0.537"
],
[
"0.1168",
"0.5327",
"0.4534",
"0.534",
"0.4534",
"0.5474",
"0.1168",
"0.547"
],
[
"0.5156",
"0.537",
"0.9126",
"0.537",
"0.9126",
"0.5503",
"0.5156",
"0.5503"
],
[
"0.1177",
"0.547",
"0.452",
"0.547",
"0.452",
"0.56",
"0.1177",
"0.56"
],
[
"0.5156",
"0.5503",
"0.9126",
"0.5503",
"0.9126",
"0.5645",
"0.5156",
"0.5645"
],
[
"0.11926",
"0.561",
"0.451",
"0.561",
"0.451",
"0.5747",
"0.11926",
"0.5747"
],
[
"0.5146",
"0.565",
"0.91",
"0.565",
"0.91",
"0.579",
"0.5146",
"0.579"
],
[
"0.11926",
"0.5757",
"0.452",
"0.5757",
"0.452",
"0.589",
"0.11926",
"0.589"
],
[
"0.5156",
"0.579",
"0.911",
"0.579",
"0.911",
"0.5923",
"0.5156",
"0.5923"
],
[
"0.1177",
"0.5884",
"0.451",
"0.5884",
"0.451",
"0.6016",
"0.1177",
"0.6016"
],
[
"0.5146",
"0.5923",
"0.911",
"0.5923",
"0.911",
"0.606",
"0.5146",
"0.606"
],
[
"0.11926",
"0.603",
"0.452",
"0.603",
"0.452",
"0.616",
"0.11926",
"0.616"
],
[
"0.5146",
"0.607",
"0.9126",
"0.607",
"0.9126",
"0.6206",
"0.5146",
"0.6206"
],
[
"0.11926",
"0.617",
"0.3857",
"0.617",
"0.3857",
"0.631",
"0.11926",
"0.631"
],
[
"0.5146",
"0.6206",
"0.911",
"0.6206",
"0.911",
"0.634",
"0.5146",
"0.634"
],
[
"0.513",
"0.634",
"0.911",
"0.634",
"0.911",
"0.648",
"0.513",
"0.648"
],
[
"0.1177",
"0.654",
"0.2296",
"0.654",
"0.2296",
"0.6675",
"0.1177",
"0.6675"
],
[
"0.5156",
"0.6484",
"0.911",
"0.6484",
"0.911",
"0.662",
"0.5156",
"0.662"
],
[
"0.08826",
"0.655",
"0.1152",
"0.655",
"0.1152",
"0.6636",
"0.08826",
"0.6636"
],
[
"0.1152",
"0.657",
"0.11926",
"0.657",
"0.11926",
"0.662",
"0.1152",
"0.662"
],
[
"0.513",
"0.661",
"0.91",
"0.661",
"0.91",
"0.675",
"0.513",
"0.675"
],
[
"0.08826",
"0.675",
"0.4844",
"0.675",
"0.4844",
"0.688",
"0.08826",
"0.688"
],
[
"0.5146",
"0.6753",
"0.911",
"0.6753",
"0.911",
"0.6895",
"0.5146",
"0.6895"
],
[
"0.0866",
"0.6875",
"0.4836",
"0.6875",
"0.4836",
"0.7007",
"0.0866",
"0.7007"
],
[
"0.527",
"0.688",
"0.9126",
"0.688",
"0.9126",
"0.702",
"0.527",
"0.702"
],
[
"0.0866",
"0.7026",
"0.4844",
"0.7026",
"0.4844",
"0.7163",
"0.0866",
"0.7163"
],
[
"0.5156",
"0.7026",
"0.9126",
"0.7026",
"0.9126",
"0.7163",
"0.5156",
"0.7163"
],
[
"0.08905",
"0.7163",
"0.4844",
"0.7163",
"0.4844",
"0.73",
"0.08905",
"0.73"
],
[
"0.5146",
"0.7173",
"0.911",
"0.7173",
"0.911",
"0.731",
"0.5146",
"0.731"
],
[
"0.08826",
"0.73",
"0.4844",
"0.73",
"0.4844",
"0.7437",
"0.08826",
"0.7437"
],
[
"0.513",
"0.731",
"0.911",
"0.731",
"0.911",
"0.744",
"0.513",
"0.744"
],
[
"0.08905",
"0.7437",
"0.4844",
"0.7437",
"0.4844",
"0.757",
"0.08905",
"0.757"
],
[
"0.08905",
"0.7583",
"0.4844",
"0.7583",
"0.4844",
"0.7715",
"0.08905",
"0.7715"
],
[
"0.5146",
"0.761",
"0.527",
"0.761",
"0.527",
"0.7695",
"0.5146",
"0.7695"
],
[
"0.524",
"0.761",
"0.55",
"0.761",
"0.55",
"0.7695",
"0.524",
"0.7695"
],
[
"0.544",
"0.759",
"0.71",
"0.759",
"0.71",
"0.773",
"0.544",
"0.773"
],
[
"0.08826",
"0.7715",
"0.4844",
"0.7715",
"0.4844",
"0.785",
"0.08826",
"0.785"
],
[
"0.5146",
"0.779",
"0.9126",
"0.779",
"0.9126",
"0.7925",
"0.5146",
"0.7925"
],
[
"0.08826",
"0.785",
"0.486",
"0.785",
"0.486",
"0.8",
"0.08826",
"0.8"
],
[
"0.513",
"0.7905",
"0.911",
"0.7915",
"0.911",
"0.806",
"0.513",
"0.8047"
],
[
"0.08905",
"0.8003",
"0.4844",
"0.8003",
"0.4844",
"0.811",
"0.08905",
"0.811"
],
[
"0.511",
"0.803",
"0.915",
"0.803",
"0.915",
"0.821",
"0.511",
"0.821"
],
[
"0.08826",
"0.813",
"0.4844",
"0.813",
"0.4844",
"0.827",
"0.08826",
"0.827"
],
[
"0.5146",
"0.821",
"0.9126",
"0.821",
"0.9126",
"0.834",
"0.5146",
"0.834"
],
[
"0.08826",
"0.827",
"0.486",
"0.827",
"0.486",
"0.8403",
"0.08826",
"0.8403"
],
[
"0.5146",
"0.8335",
"0.911",
"0.8335",
"0.911",
"0.8477",
"0.5146",
"0.8477"
],
[
"0.1054",
"0.8486",
"0.4844",
"0.8496",
"0.4844",
"0.8643",
"0.1054",
"0.863"
],
[
"0.513",
"0.8467",
"0.911",
"0.8477",
"0.911",
"0.8623",
"0.513",
"0.8613"
],
[
"0.0866",
"0.863",
"0.4844",
"0.863",
"0.4844",
"0.8774",
"0.0866",
"0.8774"
],
[
"0.5146",
"0.86",
"0.91",
"0.8613",
"0.91",
"0.8755",
"0.5146",
"0.875"
],
[
"0.08496",
"0.8755",
"0.3987",
"0.8755",
"0.3987",
"0.8896",
"0.08496",
"0.8896"
],
[
"0.5146",
"0.875",
"0.9126",
"0.875",
"0.9126",
"0.8896",
"0.5146",
"0.8896"
]
] | [
"用于多层网络推理的改进的深度嵌入",
"Huan Song' and Jayaraman J. Thiagarajan?",
"1Bosch Research North America",
"2Lawrence Livermore National Labs",
"[email protected],[email protected]",
"Abstract",
"活[Yang et al.,2016:Thiagarajan et al.,2016]：然",
"而，",
"缺乏可扩展性一直是这些方法的持续限制。这激发",
"Inferencing with network data necessitates the",
"了诸如DeepWalkE[Perozzietal.,2014]和Node2VecE",
"mapping of its nodes into a vector space, where",
"[GroverandLeskovec，2016]之类的最新方法的设计，这",
"the relationships are preserved. However, with",
"些方法基于分布假设，在语言建模E[Harris，1954]中普",
"multi-layered networks, where multiple types",
"遍采用，其中两个节点在短随机游走中的共现暗示了强",
"of relationships exist for the same set of nodes.",
"烈的语义相似性概念。在负采样E[Mikolovetal.2013]",
"it is crucial to exploit the information shared",
"等高效优化策略的支持下，这些方法可以轻松处理超大",
"between layers, in addition to the distinct as",
"规模网络",
"pects of each layer.In this paper, we pro",
"最近，随着多层网络的出现，我。e。，在每个层中具有",
"pose a novel approach that first obtains node",
"不同关系结构的同一组节点[Lietal.，2018]，获得保留",
"embeddingsinalllayersjointlyviaDeepWalk",
"多视图关系的简洁嵌人变得至关重要。与单层情况相反",
"on a supra graph, which allows interactions",
"多层网络的节点嵌入技术必须适应两层的提取（例如。g。",
"between layers, and then fine-tunes the em",
"社区检测）以及统一表示(e。g。节点分类)。由于多层网",
"beddings to encourage cohesive structure in",
"络中固有的挑战，包括关系类型的异构性和不同层中稀",
"the latent space.With empirical studies in",
"疏程度的变化，单层方法的直接扩展是无效的[Zhang",
"node classification,link prediction and multi-",
"etal，2018］。最后，关于增加层数的可扩展性是一个重",
"layered community detection, we show that",
"要的设计标准",
"the proposed approach outperforms existing",
"在本文中，我们开发了一种可扩展的嵌人方法，该方法",
"single- and multi-layered network embedding",
"首先直接在多层网络（M-DeepWalk）上执行DeepWalk",
"algorithms on several benchmarks. In addition",
"风格的优化，并利用细化策略通过鼓励内聚社区的形成",
"to effectively scaling to a large number of lay-",
"来进一步微调嵌人。我们研究了我们的方法在链接预测",
"ers (tested up to 37), our approach consistently",
"节点分类和社区检测问题中的应用。令人惊的是，通",
"produces highly modular community structure",
"过包含多层依赖关系，我们发现M-DeepWalk在链接预",
"even when compared to methods that directly",
"测方面已经可以超越最先进的多层嵌人方法[Zhanget",
"optimize for the modularity function.",
"al.，2018。但是，随着节点分类和多层社区检测等更具",
"挑战性的任务，还有改进的空间。使用短随机游走来推",
"Introduction",
"断底层语义结构的前提依赖于网络高度稀疏且节点共现",
"1",
"",
"遵循幂律的假设。但是，通过允许层间边缘，在语义持续",
"网络数据的建模和推理对于包括社交网络分析[Eagle",
"更长时间的情况下，可能会违反该假设。我们通过包括",
"and Pentland,2006]，推荐系统[Rao et al.,2015]和医",
"个改进步骤来解决这个限制，其中来自M-DeepWalk",
"疗保健[Fornitoetal.，2013]在内的多个应用领域至关",
"的嵌人被微调以产生更有凝聚力的社区。即使网络数据",
"重要。这广泛地涵盖了诸如节点分类，链接预测，社区",
"集包含多达37层，我们也表明，与现有的多层社区检测",
"检测，有影响力的节点检测等问题。。它们通常通过将节",
"技术相比，所提出的方法产生了高度模块化的社区结构。",
"点映射到潜在的低维向量空间中，并在生成的嵌入上调",
"用适当的机器学习技术来解决，例如。g。无监督聚类的",
"2",
"",
"ProblemDefinition",
"社区检测[Dingetal.,2016]。在最简单的形式中，这",
"我们将单层无向，未加权的网络表示为9=（V,），其",
"样的嵌人可以通过网络结构的描述符来推断-e。g。图邻",
"中V表示基数为V=N的节点集，ε表示边集。嵌人",
"接、图拉普拉斯算子或模块化矩阵IChenetal..2014:",
"技术的目标是为节点XRNxd生成潜在表示，其中d",
"Yangetal.，2016l。虽然最早的解决方案依赖于矩阵",
"是所需的嵌入维数。相应地，我们使用一组L相互依赖",
"分解来优化嵌人，但最近发现深度图自动编码器更加灵",
"的网络g()=(V(),s()),forl=1,..,L表示多层网络，",
"*This work was performed under the auspices of the U.S.",
"其中在每对层之间存在一个节点映射M，以指示一个",
"Department of Energy by Lawrence Livermore National Lab-",
"网络中的哪些节点对应于另一个网络中的节点。请注意",
"oratory under Contract DE-AC52-07NA27344",
"我们期望不同的层包含公共节点，但不要求V(）相同。"
] | {
"arxiv_id": "1811.12156",
"language": "zh",
"page_idx": 0
} |
||
1811.12156-zh-1 | 1811.12156-zh-1.jpg | $\mathcal{G}^{(1)}$
<FigureHere>
Figure 1: Proposed approach for extracting node embeddings 其中$e_ij^{(l,m)}$表示节点$v_i^{(l)}$和$v_j^{(m)}$之间的边缘权重，
$\bar{\mathrm{from~multi-layered~networks.}}$
多层网络中的完整节点集表示为 $\mathcal{V}=\cup_l=1^L\mathcal{V}^{(l)}$,其中$\cup$ 是指联合。现在，我们定义了本文考虑的推理任务，这些任务需要提取多层网络嵌人
Definition 2.1.( Node classification ) Given a multi- $\operatorname*{layered}$ network$\{(\mathcal{V}^{(l)},\mathcal{E}^{(l)})\}_l=1^L$and the semantic labels $\mathcal{Y}_{lab}$ for a subset of nodes $\mathcal{V}_lab\subset\mathcal{V}$,where each $y\in\mathcal{Y}_lab$ assumes one of the $K$ pre-defined classes, predict labels for each of the nodes in the set $v\in \mathcal{V} \setminus \mathcal{V} _{lab}^{\prime }$ .
Definition 2.2.( Link prediction ) Given a multi-layered $\operatorname{network}\left\{(\mathcal{V}^{(l)},\mathcal{E}^{(l)})\right\}_{l=1}^{L},\operatorname{in~a~layer~}i,\operatorname{estimate~the~like-}$ lihood $p( e| l$ = $i)$ for an unobserved edge $e\in\mathcal{E}\setminus\mathcal{E}^{(i)}$ to exist, where $\mathcal{E}$ is the superset of all possible edges in layeri .
Defnition 2.3.( Multi-layered community detection ) Given a multi-layered network $\{(\mathcal{V}^{(l)},\mathcal{E}^{(l)})\}_{l=1}^{L}$, identify coherent communities denoted by the sets $\mathcal{C}_1,\cdots,\mathcal{C}_K$ such that each node $i$ in every layer $l$ is assigned to one of the communities, arg$\max_k\mathcal{A}(v_i^{(l)},\mathcal{C}_k).$ Here, $\mathcal{A}$ denotes a similarity function.
# 3 Proposed Approach
如图 1 所示，我们的方法将多层网络作为输人，并为$\mathcal{V}=\cup_{l=1}^{L}\mathcal{V}^{(l)}$中的每个节点推断$d$维潜在表示。与最先进的方法$\mathbb{E}[$Liu et al., 2017; Li et al., 2018]相比，我们不会同时优化适用于所有层的通用表示和保留特定于层的网络结构的本地表示。相反，我们构造了一个多层超图$\mathcal{G}^{sup}$ ,它显式地对层间依赖关系进行建模，并使用深度网络嵌人方法获得每一层中每个节点的嵌入。虽然这个过程产生了语义上有意义的嵌人，但我们表明，通过细化策略来微调嵌人是有益的，这些策略鼓励潜在空间中的凝聚力社区结构
3.1 Node Embeddings using M-DeepWalk
我们的目标是通过创建层间边缘来利用不同层之间的依赖关系，其中可以通过 DeepWalk$\underline{\mathrm{F}}[$Perozzi et al., 2014] 等深度网络嵌人技术将这种依赖关系编码到潜在表示中。从广义上讲，多层网络中的结构发现可以使用两种情况来表征：(a)语义一致的组可以包括来自所有层或层的子集的节点；(b)只能在特定层中发现语义概念，而在其他层中则不明显。这激发了遵守以下约束的嵌人技术的设计：对于场景 (a),有必要利用跨层的信息流，而
对于后一种场景，潜在的嵌人不应被其他层中的依赖项混淆。为了实现这一点，我们基于局部邻域结构之间的相似性引人层间边缘。对于分别来自层$l$和$m$的一对节点$v_i^{(l)}$和$v_j^{(m)}$,我们引人层间边缘如下：
$$e_{ij}^{(l,m)}=0,\mathrm{if}\:i\neq j,\\e_{ij}^{(l,m)}=\frac{|\mathcal{N}_{i}^{(l)}\cap\mathcal{N}_{j}^{(m)}|}{|\mathcal{N}_{i}^{(l)}\cup\mathcal{N}_{j}^{(m)}|},\mathrm{if}\:i=j$$
(1)
$\mathcal{N}_i^{(l)},\mathcal{N}_j^{(m)}$是 XMATHX 的邻域。请注意，边缘权重被计算为两层之间的邻域列表中的协议的 Jaccard 系数。这确保了层中的每个节点仅连接到另一层中的对应节点， 并且相似的局部邻域是共享语义结构的有力指标。在计算边缘权重时，我们使用指定的阈值将$e_{ij}^{(k,l)}$转换为二进制连接。具有虚拟层间边缘的多层网络称为超图$\mathcal{G}^{sup}$ (图2(a))
DeepWalk 在网络上的短随机游走生成的节点序列与文档语料库中的句子之间进行类比。让我们在单层网络$\mathcal{G}$中的步骤$t$中考虑一个简单的度量游走$\mathcal{W}_t$ ,它植根于节点$v_i$ 。节点$v_i$和$v_j$之间的转移概率为$P(W_t+1=$ $v_j| \mathcal{W} _t= v_i) = h( \| \boldsymbol{x}_i- \boldsymbol{x}_j\| _2/ \sigma )$ ,其中$\|x_i-\boldsymbol{x}_j\|_2$ 测量潜在空间中两个节点之间的接近度，而$h$是将节点相似性与实际共现概率联系起来的链接函数。通过适当选择行走长度，可以从使用随机行走构建的同现统计信息中准确地恢复真实度量。将网络上的$S$步随机游走表示为$\left(v_0,v_1,\ldots,v_{S-1}\right)$,我们尝试获得一个向量空间，该向量空间可以恢复在给定上下文的情况下观察节点的可能性， 即。e。, $P(v_{S}|v_{0},v_{1},\ldots,v_{S-1})$经过几次随机漫步。在所提出的方法中，我们利用 DeepWalk 来学习$\mathcal{G}^sup$中$|\mathcal{V}|$ 节点的嵌人，这被称为 M-DeepWalk 。
### 3.2 Refinement Step
M-DeepWalk 的成功依赖于网络稀疏和短随机游走揭示所有必要语义的假设。但是，由于引人了层间边缘，并且不同层中的稀疏程度不同，因此在语义关系持续较长时间的情况下，可能会违反此假设。为了解决这一挑战并在潜在空间中产生内聚结构，我们建议改进学习的嵌人。更具体地说，我们采用代理聚类成本以及基于模块化的重新分配来优化嵌入
架构：我们利用 4- 层全连接神经网络$\mathcal{H}$来实现细化步骤。由于 M-DeepWalk 的表示包含从原始网络数据中提取的重要拓扑信息，因此我们发现合理的浅神经网络就足够了。如图 2(b) 所示，网络的输入是来自 M-DeepWalk 的嵌人$X\in\mathbb{R}^d$,目标是生成微调的嵌人$\bar{\boldsymbol{X}}(\dot\text{也在 }\mathbb{R}^d$ 中(
代理聚类损失：我们将模型$\mathcal{H}$初始化为一个自动编码器，它最小化了重建目标(均方误差)。由于测量结果结构的内聚性并不简单，并且实际上使用它来更新$\mathcal{H}$ ,因此我们求助于类似于E[ Xie et al. , 2016] 的代理聚类目标，i。e。, 产生低熵聚类赋值分布。假设潜在空间中有节点的$K$簇，我们首先在$\bar{X}$上执行$k$ -means 聚类，并 | [
[
"0.09235",
"0.0682",
"0.1111",
"0.0682",
"0.1111",
"0.07574",
"0.09235",
"0.07574"
],
[
"0.517",
"0.07135",
"0.91",
"0.07135",
"0.91",
"0.0821",
"0.517",
"0.0821"
],
[
"0.5146",
"0.084",
"0.91",
"0.084",
"0.91",
"0.09784",
"0.5146",
"0.09784"
],
[
"0.3218",
"0.09534",
"0.3857",
"0.09534",
"0.3857",
"0.1067",
"0.3218",
"0.1067"
],
[
"0.5146",
"0.09784",
"0.911",
"0.09784",
"0.911",
"0.1111",
"0.5146",
"0.1111"
],
[
"0.509",
"0.10986",
"0.812",
"0.108",
"0.812",
"0.1288",
"0.509",
"0.1307"
],
[
"0.09235",
"0.1237",
"0.1095",
"0.1237",
"0.1095",
"0.1332",
"0.09235",
"0.1332"
],
[
"0.2018",
"0.1338",
"0.2607",
"0.1338",
"0.2607",
"0.1445",
"0.2018",
"0.1445"
],
[
"0.3096",
"0.1364",
"0.3718",
"0.1364",
"0.3718",
"0.1477",
"0.3096",
"0.1477"
],
[
"0.1013",
"0.1445",
"0.1054",
"0.1497",
"0.1013",
"0.1515",
"0.0972",
"0.1465"
],
[
"0.1993",
"0.1433",
"0.2615",
"0.1459",
"0.2615",
"0.1572",
"0.1993",
"0.1547"
],
[
"0.407",
"0.1459",
"0.4714",
"0.1459",
"0.4714",
"0.1559",
"0.407",
"0.1559"
],
[
"0.653",
"0.1465",
"0.778",
"0.1528",
"0.777",
"0.1692",
"0.652",
"0.1635"
],
[
"0.3096",
"0.1477",
"0.3718",
"0.1477",
"0.3718",
"0.159",
"0.3096",
"0.159"
],
[
"0.1936",
"0.1483",
"0.2009",
"0.1483",
"0.2009",
"0.1515",
"0.1936",
"0.1515"
],
[
"0.2059",
"0.154",
"0.2573",
"0.1559",
"0.2566",
"0.1704",
"0.2051",
"0.1686"
],
[
"0.10706",
"0.1604",
"0.1299",
"0.1604",
"0.1299",
"0.1718",
"0.10706",
"0.1718"
],
[
"0.652",
"0.1572",
"0.669",
"0.1572",
"0.669",
"0.1666",
"0.652",
"0.1666"
],
[
"0.664",
"0.1724",
"0.763",
"0.1724",
"0.763",
"0.1913",
"0.664",
"0.1913"
],
[
"0.08905",
"0.1799",
"0.1168",
"0.1799",
"0.1168",
"0.1913",
"0.08905",
"0.1913"
],
[
"0.8896",
"0.1749",
"0.9126",
"0.1749",
"0.9126",
"0.1881",
"0.8896",
"0.1881"
],
[
"0.6055",
"0.1818",
"0.6665",
"0.178",
"0.6685",
"0.1932",
"0.607",
"0.1976"
],
[
"0.7646",
"0.1843",
"0.822",
"0.1863",
"0.822",
"0.1995",
"0.7646",
"0.1976"
],
[
"0.607",
"0.1895",
"0.63",
"0.1932",
"0.6274",
"0.202",
"0.6045",
"0.1976"
],
[
"0.4346",
"0.1863",
"0.4412",
"0.1863",
"0.4412",
"0.1932",
"0.4346",
"0.1932"
],
[
"0.6626",
"0.1901",
"0.767",
"0.1901",
"0.767",
"0.2166",
"0.6626",
"0.2166"
],
[
"0.4043",
"0.1976",
"0.4722",
"0.1976",
"0.4722",
"0.209",
"0.4043",
"0.209"
],
[
"0.4084",
"0.209",
"0.468",
"0.209",
"0.468",
"0.2203",
"0.4084",
"0.2203"
],
[
"0.511",
"0.2299",
"0.915",
"0.2299",
"0.915",
"0.2507",
"0.511",
"0.2507"
],
[
"0.08826",
"0.2424",
"0.4885",
"0.2424",
"0.4885",
"0.2556",
"0.08826",
"0.2556"
],
[
"0.5107",
"0.2474",
"0.9165",
"0.2474",
"0.9165",
"0.2715",
"0.5107",
"0.2715"
],
[
"0.0866",
"0.2551",
"0.2778",
"0.2551",
"0.2778",
"0.2683",
"0.0866",
"0.2683"
],
[
"0.5156",
"0.2715",
"0.9126",
"0.2715",
"0.9126",
"0.2854",
"0.5156",
"0.2854"
],
[
"0.0866",
"0.2798",
"0.4885",
"0.2778",
"0.4885",
"0.2954",
"0.0866",
"0.2974"
],
[
"0.5146",
"0.286",
"0.91",
"0.286",
"0.91",
"0.2998",
"0.5146",
"0.2998"
],
[
"0.08905",
"0.2954",
"0.4844",
"0.2954",
"0.4844",
"0.3093",
"0.08905",
"0.3093"
],
[
"0.5146",
"0.2998",
"0.911",
"0.2998",
"0.911",
"0.3132",
"0.5146",
"0.3132"
],
[
"0.08826",
"0.308",
"0.2998",
"0.31",
"0.2998",
"0.324",
"0.08826",
"0.3213"
],
[
"0.511",
"0.3113",
"0.914",
"0.3113",
"0.914",
"0.332",
"0.511",
"0.332"
],
[
"0.08826",
"0.3276",
"0.4844",
"0.3276",
"0.4844",
"0.3416",
"0.08826",
"0.3416"
],
[
"0.5107",
"0.33",
"0.9126",
"0.3289",
"0.9126",
"0.3467",
"0.5107",
"0.348"
],
[
"0.08417",
"0.3416",
"0.4878",
"0.3423",
"0.4878",
"0.3599",
"0.08417",
"0.3591"
],
[
"0.513",
"0.348",
"0.584",
"0.348",
"0.584",
"0.3591",
"0.513",
"0.3591"
],
[
"0.0866",
"0.358",
"0.486",
"0.358",
"0.486",
"0.3726",
"0.0866",
"0.3726"
],
[
"0.5293",
"0.3643",
"0.911",
"0.3643",
"0.911",
"0.3774",
"0.5293",
"0.3774"
],
[
"0.0866",
"0.3706",
"0.486",
"0.3706",
"0.486",
"0.385",
"0.0866",
"0.385"
],
[
"0.5146",
"0.377",
"0.911",
"0.3774",
"0.911",
"0.3926",
"0.5146",
"0.3914"
],
[
"0.08496",
"0.3838",
"0.411",
"0.3865",
"0.411",
"0.401",
"0.08496",
"0.3984"
],
[
"0.5146",
"0.3926",
"0.911",
"0.3926",
"0.911",
"0.406",
"0.5146",
"0.406"
],
[
"0.0866",
"0.4048",
"0.486",
"0.4048",
"0.486",
"0.4185",
"0.0866",
"0.4185"
],
[
"0.5146",
"0.4048",
"0.9004",
"0.4072",
"0.9004",
"0.4204",
"0.5146",
"0.4185"
],
[
"0.08417",
"0.4172",
"0.4878",
"0.4192",
"0.4878",
"0.4375",
"0.08417",
"0.435"
],
[
"0.5146",
"0.4204",
"0.911",
"0.4204",
"0.911",
"0.4343",
"0.5146",
"0.4343"
],
[
"0.08496",
"0.435",
"0.4844",
"0.4343",
"0.4844",
"0.449",
"0.08496",
"0.4495"
],
[
"0.513",
"0.432",
"0.911",
"0.433",
"0.911",
"0.4475",
"0.513",
"0.4463"
],
[
"0.0866",
"0.4507",
"0.486",
"0.4507",
"0.486",
"0.464",
"0.0866",
"0.464"
],
[
"0.5146",
"0.4475",
"0.911",
"0.4475",
"0.911",
"0.461",
"0.5146",
"0.461"
],
[
"0.08826",
"0.4666",
"0.1438",
"0.4666",
"0.1438",
"0.4778",
"0.08826",
"0.4778"
],
[
"0.5146",
"0.4622",
"0.911",
"0.4622",
"0.911",
"0.476",
"0.5146",
"0.476"
],
[
"0.5146",
"0.476",
"0.91",
"0.476",
"0.91",
"0.4893",
"0.5146",
"0.4893"
],
[
"0.0866",
"0.4841",
"0.486",
"0.4841",
"0.486",
"0.4976",
"0.0866",
"0.4976"
],
[
"0.517",
"0.4924",
"0.5767",
"0.4924",
"0.5767",
"0.5024",
"0.517",
"0.5024"
],
[
"0.584",
"0.488",
"0.9126",
"0.486",
"0.9126",
"0.504",
"0.584",
"0.506"
],
[
"0.0866",
"0.4968",
"0.4844",
"0.4988",
"0.4844",
"0.513",
"0.0866",
"0.511"
],
[
"0.5146",
"0.5024",
"0.9087",
"0.5024",
"0.9087",
"0.5166",
"0.5146",
"0.5166"
],
[
"0.08496",
"0.511",
"0.486",
"0.513",
"0.486",
"0.528",
"0.08496",
"0.526"
],
[
"0.513",
"0.515",
"0.675",
"0.5176",
"0.675",
"0.531",
"0.513",
"0.529"
],
[
"0.6665",
"0.5176",
"0.91",
"0.5166",
"0.91",
"0.53",
"0.6665",
"0.531"
],
[
"0.0866",
"0.5264",
"0.4878",
"0.5264",
"0.4878",
"0.5405",
"0.0866",
"0.5405"
],
[
"0.513",
"0.53",
"0.911",
"0.53",
"0.911",
"0.5444",
"0.513",
"0.5444"
],
[
"0.08417",
"0.5425",
"0.4885",
"0.5405",
"0.4885",
"0.5566",
"0.08417",
"0.5596"
],
[
"0.513",
"0.5435",
"0.79",
"0.5435",
"0.79",
"0.5566",
"0.513",
"0.5566"
],
[
"0.08496",
"0.558",
"0.2378",
"0.558",
"0.2378",
"0.5713",
"0.08496",
"0.5713"
],
[
"0.511",
"0.582",
"0.6895",
"0.5835",
"0.6895",
"0.5977",
"0.511",
"0.5967"
],
[
"0.0866",
"0.5884",
"0.1217",
"0.5884",
"0.1217",
"0.5996",
"0.0866",
"0.5996"
],
[
"0.1152",
"0.5884",
"0.2893",
"0.5884",
"0.2893",
"0.603",
"0.1152",
"0.603"
],
[
"0.08826",
"0.608",
"0.4844",
"0.608",
"0.4844",
"0.622",
"0.08826",
"0.622"
],
[
"0.0866",
"0.619",
"0.4878",
"0.619",
"0.4878",
"0.637",
"0.0866",
"0.637"
],
[
"0.513",
"0.616",
"0.9126",
"0.616",
"0.9126",
"0.63",
"0.513",
"0.63"
],
[
"0.513",
"0.63",
"0.911",
"0.63",
"0.911",
"0.6445",
"0.513",
"0.6445"
],
[
"0.08905",
"0.637",
"0.4844",
"0.637",
"0.4844",
"0.651",
"0.08905",
"0.651"
],
[
"0.5146",
"0.6445",
"0.911",
"0.6445",
"0.911",
"0.6577",
"0.5146",
"0.6577"
],
[
"0.08905",
"0.653",
"0.4836",
"0.653",
"0.4836",
"0.6636",
"0.08905",
"0.6636"
],
[
"0.513",
"0.657",
"0.9126",
"0.657",
"0.9126",
"0.6704",
"0.513",
"0.6704"
],
[
"0.08826",
"0.6655",
"0.4844",
"0.6655",
"0.4844",
"0.6787",
"0.08826",
"0.6787"
],
[
"0.5146",
"0.672",
"0.9126",
"0.672",
"0.9126",
"0.685",
"0.5146",
"0.685"
],
[
"0.08826",
"0.678",
"0.486",
"0.678",
"0.486",
"0.6924",
"0.08826",
"0.6924"
],
[
"0.5146",
"0.686",
"0.911",
"0.686",
"0.911",
"0.6997",
"0.5146",
"0.6997"
],
[
"0.08826",
"0.6924",
"0.4844",
"0.6924",
"0.4844",
"0.7056",
"0.08826",
"0.7056"
],
[
"0.513",
"0.6987",
"0.693",
"0.6997",
"0.693",
"0.7134",
"0.513",
"0.712"
],
[
"0.08905",
"0.707",
"0.4844",
"0.707",
"0.4844",
"0.72",
"0.08905",
"0.72"
],
[
"0.08826",
"0.72",
"0.4844",
"0.72",
"0.4844",
"0.7344",
"0.08826",
"0.7344"
],
[
"0.5146",
"0.7183",
"0.911",
"0.7183",
"0.911",
"0.732",
"0.5146",
"0.732"
],
[
"0.08826",
"0.733",
"0.2336",
"0.7344",
"0.2336",
"0.7476",
"0.08826",
"0.747"
],
[
"0.5156",
"0.732",
"0.911",
"0.732",
"0.911",
"0.7456",
"0.5156",
"0.7456"
],
[
"0.5146",
"0.747",
"0.911",
"0.747",
"0.911",
"0.7603",
"0.5146",
"0.7603"
],
[
"0.08417",
"0.7583",
"0.4314",
"0.759",
"0.4314",
"0.7734",
"0.08417",
"0.773"
],
[
"0.5146",
"0.7583",
"0.9126",
"0.759",
"0.9126",
"0.7734",
"0.5146",
"0.773"
],
[
"0.513",
"0.773",
"0.9126",
"0.7715",
"0.9126",
"0.786",
"0.513",
"0.787"
],
[
"0.08826",
"0.779",
"0.4844",
"0.779",
"0.4844",
"0.7925",
"0.08826",
"0.7925"
],
[
"0.08826",
"0.7915",
"0.4844",
"0.7915",
"0.4844",
"0.806",
"0.08826",
"0.806"
],
[
"0.517",
"0.7905",
"0.539",
"0.7905",
"0.539",
"0.799",
"0.517",
"0.799"
],
[
"0.08905",
"0.806",
"0.486",
"0.806",
"0.486",
"0.8193",
"0.08905",
"0.8193"
],
[
"0.5146",
"0.806",
"0.911",
"0.806",
"0.911",
"0.8193",
"0.5146",
"0.8193"
],
[
"0.08826",
"0.8193",
"0.4844",
"0.8193",
"0.4844",
"0.8335",
"0.08826",
"0.8335"
],
[
"0.5146",
"0.8193",
"0.911",
"0.8193",
"0.911",
"0.8335",
"0.5146",
"0.8335"
],
[
"0.08905",
"0.834",
"0.4844",
"0.834",
"0.4844",
"0.8477",
"0.08905",
"0.8477"
],
[
"0.513",
"0.834",
"0.91",
"0.834",
"0.91",
"0.8477",
"0.513",
"0.8477"
],
[
"0.0907",
"0.8477",
"0.486",
"0.8477",
"0.486",
"0.8613",
"0.0907",
"0.8613"
],
[
"0.513",
"0.8477",
"0.9087",
"0.8477",
"0.9087",
"0.8613",
"0.513",
"0.8613"
],
[
"0.08826",
"0.86",
"0.486",
"0.8613",
"0.486",
"0.8755",
"0.08826",
"0.875"
],
[
"0.511",
"0.8613",
"0.91",
"0.8613",
"0.91",
"0.8755",
"0.511",
"0.8755"
],
[
"0.08826",
"0.8755",
"0.486",
"0.8755",
"0.486",
"0.8896",
"0.08826",
"0.8896"
],
[
"0.5146",
"0.8755",
"0.911",
"0.8755",
"0.911",
"0.8896",
"0.5146",
"0.8896"
]
] | [
"c(1)",
"对于后一种场景，潜在的嵌入不应被其他层中的依赖项",
"混淆。为了实现这一点，我们基于局部邻域结构之间的",
"M-DeepWalk",
"相似性引人层间边缘。对于分别来自层l和㎡的一对节",
"点和m），我们引人层间边缘如下：",
"g(2)",
"Inter-Layer",
"DeepGraph",
"",
"Dependency",
"Refinement",
"(m)=0,ifij,",
"Embedding",
"",
"Modeling",
"",
"2ii",
"W(nNm)",
"g(3)",
"(1)",
"e(m)-",
",ifi=i",
"eii",
"+",
"WuNm",
"Multi-Layered",
"Embeddings",
"其中em）表示节点u）和u/m）之间的边缘权重，",
"Figure l: Proposed approach for extracting node embeddings",
"N\",Nm）是XMATHX的邻域。请注意，边缘权重",
"from multi-layered networks.",
"被计算为两层之间的邻域列表中的协议的Jaccard系数。",
"多层网络中的完整节点集表示为=U冶V(），其中U",
"这确保了层中的每个节点仅连接到另一层中的对应节点",
"是指联合。现在，我们定义了本文考虑的推理任务，这",
"并且相似的局部邻域是共享语义结构的有力指标。在计",
"些任务需要提取多层网络嵌人",
"算边缘权重时，我们使用指定的阅值将e）转换为二",
"Definition 2.1. (Node classification )Given a multi-",
"进制连接。具有虚拟层间边缘的多层网络称为超图gsup",
"layered network (v(), ())/-1 and the semantic labels",
"(图2(a))",
"Jiab for a subset of nodes Viab C V, where each y E Yiab",
"DeepWalk在网络上的短随机游走生成的节点序列与",
"assumes one of the K pre-defined classes, predict labels",
"文档语料库中的句子之间进行类比。让我们在单层网络",
"for each of the nodes in the set u e VI Viab.",
"g中的步骤t中考虑一个简单的度量游走Wt，它植根",
"Definition 2.2.(Link prediction) Given a multi-layered",
"于节点vi。节点Ui和Ui之间的转移概率为P(Wt+1",
"network (v(),())- , in a layer i , estimate the like-",
"UWt=i）=h(i-,2/o），其中a-,2测量",
"lihood p(ell = i) for an unobserved edge e E /(i)",
"潜在空间中两个节点之间的接近度，而h是将节点相似",
"to exist, where & is the superset of all possible edges in",
"性与实际共现概率联系起来的链接函数。通过适当选择",
"layer z.",
"行走长度，可以从使用随机行走构建的同现统计信息中",
"准确地恢复真实度量。将网络上的S步随机游走表示为",
"Definition 2.3. （Multi-layered community detection )",
"Vo,U1,..",
"，us-1），我们尝试获得一个向量空间，该向量",
"Given a multi-layered network ((v(),&()}-1 , identify",
"空间可以恢复在给定上下文的情况下观察节点的可能性",
"coherent communities denoted by the sets Ci,...,Ck,",
"即。e。, P(vsvo,vi,..",
"..,Us-1）经过次随机漫步。在所",
"such that each node i in every layer l is assigned to one of",
"提出的方法中，我们利用DeepWalk来学习gsup中[V",
"the communities, arg maxk A((),Ck) . Here, A denotes",
"节点的嵌人，这被称为M-DeepWalk。",
"a similarity function.",
"3.2Refinement Step",
"3",
"Proposed Approach",
"如图1所示，我们的方法将多层网络作为输入，并为",
"V=UV()中的每个节点推断d维潜在表示。与最先",
"M-DeepWalk的成功依赖于网络稀疏和短随机游走揭示",
"所有必要语义的假设。但是，由于引人了层间边缘，并",
"进的方法E[Liuetal,2017;Lietal.,2018]相比，我们",
"且不同层中的稀疏程度不同，因此在语义关系持续较长",
"不会同时优化适用于所有层的通用表示和保留特定于层",
"时间的情况下，可能会违反此假设。为了解决这一挑战",
"的网络结构的本地表示。相反，我们构造了一个多层超",
"并在潜在空间中产生内聚结构，我们建议改进学习的嵌",
"图gsup，它显式地对层间依赖关系进行建模，并使用深",
"人。更具体地说，我们采用代理聚类成本以及基于模块",
"度网络嵌人方法获得每一层中每个节点的嵌人。虽然这",
"化的重新分配来优化嵌入",
"个过程产生了语义上有意义的嵌入，但我们表明，通过",
"细化策略来微调嵌人是有益的，这些策略鼓励潜在空间",
"架构：我们利用4一层全连接神经网络H来实现细化步",
"中的凝聚力社区结构",
"骤。由于M-DeepWalk的表示包含从原始网络数据中提",
"取的重要拓扑信息，因此我们发现合理的浅神经网络就",
"3.1Node Embeddings using M-DeepWalk",
"足够了。如图2(b）所示，网络的输人是来自M-DeepWalk",
"的嵌人XeRd，目标是生成微调的嵌人x（也在IRd",
"我们的目标是通过创建层间边缘来利用不同层之间的依",
"赖关系，其中可以通过DeepWalkE[Perozzietal.,2014]",
"中）",
"等深度网络嵌人技术将这种依赖关系编码到潜在表示中。",
"代理聚类损失：我们将模型孔初始化为一个自动编码",
"从广义上讲，多层网络中的结构发现可以使用两种情况",
"器，它最小化了重建目标（均方误差）。由于测量结果结",
"来表征：（a）语义一致的组可以包括来自所有层或层的",
"构的内聚性并不简单，并且实际上使用它来更新H，因",
"子集的节点；（b）只能在特定层中发现语义概念，而在",
"此我们求助于类似于E[Xieetal.，2016]的代理聚类目",
"其他层中则不明显。这激发了遵守以下约束的嵌入技术",
"标，i。e。，产生低炳聚类赋值分布。假设潜在空间中有",
"的设计：对于场景（a），有必要利用跨层的信息流，而",
"节点的K簇，我们首先在X上执行k-means聚类，并"
] | {
"arxiv_id": "1811.12156",
"language": "zh",
"page_idx": 1
} |
||
1811.12156-zh-3 | 1811.12156-zh-3.jpg | 络的情况，提供了多切片模块化$\mathbb{E}[$Mucha et al.,2010]:
$$\begin{aligned}Q_{\mathrm{multi}}=\frac{1}{2\mu}\sum_{i,j}\sum_{l,m}\biggl[\biggl(e_{ij}^{(l)}-\gamma^{(l)}\frac{n_{i}^{(l)}n_{j}^{(l)}}{2r_{l}}\biggr)\delta(l,m)\\&+\delta(i,j)\sigma_{j}^{(l,m)}\biggr]\delta(c_{i}^{(l)},c_{j}^{(m)}),\end{aligned}$$
(6
其中$\mu$是层之间的归一化因子，$\gamma^{(l)}$是层$v_i$之间的分辨率参数。。e。,$v_j^{(l)}$和$v_j^{(m)}$,其他参数是从(5)到每个层$l$的直接扩展。重写 (6)以描述每个节点对模块化功能的贡献：
$$\begin{aligned}Q_{\mathrm{multi}}&=\frac{1}{2\mu}\bigg[\sum_{l}\sum_{i}\sum_{j\in\mathcal{C}_{i}^{(l)}}\bigg(e_{ij}^{(l)}-\gamma^{(l)}\frac{n_{i}^{(l)}n_{j}^{(l)}}{2r_{l}}\bigg)\\&+\sum_{l}\sum_{i}\sum_{m\neq l}\sum_{j\in C_{i}^{(l)}}\sigma_{i}^{(l,m)}\bigg]\\&=\frac{1}{2\mu}\bigg[\sum_{l}\sum_{i}\bigg(n_{\mathcal{C}_{i}}-\gamma n_{i}^{(l)}\frac{r_{\mathcal{C}_{i}^{(l)}}}{r_{l}}\bigg)+\sum_{l}\sum_{i}\sigma n_{\mathcal{C}^{(l)}}\bigg]\\&=\frac{1}{2\mu}\sum_{l}\sum_{i}\bigg(n_{\mathcal{C}_{i}}-\gamma n_{i}^{(l)}\frac{r_{\mathcal{C}_{i}^{(l)}}}{r_{l}}+\sigma n_{\mathcal{C}^{(l)}}\bigg)\quad(7)\end{aligned}$$
这实质上将 [Duch and Arenas, 2005] 中的“适合度分数”扩展到了多层网络。这里，我们假设$\gamma_l,\sigma_i^{l,m}$是所有节点的常量。我们使用$n_{\mathcal{C}_i}$来表示$l$层中将节点$v_i^{(l)}$ 连接到标签为$\rfloor_i^{(l)}$的其他节点的边数。另一方面，$n_{\mathcal{C}^{(l)}}$ $\text{是 反 上 内 “ 亚 门 }_{i}$ 由 于 10个 个 的 边 形 式 表 示 $v_{i}^{( l) }$ 的 对 应 节 点 的 数 量 , 这 些 节 点 也 共 享 标 签 $\rfloor _{i}^{( l) }.$ EQ 中的最终求和项。(7)定义了$v_i^{(l)}$对多切片模块化的贡献。现在，我们通过分别对层内和层间贡献项进行归一化来测量适应度分数：
(8)
$$\lambda_i^{(l)}=\frac{n_{\mathcal{C}_i}}{n_i^{(l)}}-\gamma\frac{r_{\mathcal{C}_i^{(l)}}}{r_l}+\sigma\frac{n_{\mathcal{C}^{(l)}}}{n_i^{(l,m)}}$$
其中$n_i^{(l,m)}$是指$v_i^{(l)}$的对应项的数量。按照[Duch and Arenas, 2005; Boettcher, 2001] 中的概率选择过程，$\mathcal{V}$中的节点首先根据其适应度得分按升序排序，然后根据概率进行采样：$p(s)\propto s^{-\tau}$其中$s$是排序后的节点等级和$\tau\sim1+\frac{1}{ln(|\mathcal{V}|)}$。接下来，我们将采样节点$i^\prime$移动到社区$k^{\prime }$ , 这 给 出 了 $Q_{\mathrm{multi}}$ 的最大增益。为了反映整个细化过程的更新分配，我们增加了社区可能性$q_i^{\prime k^{\prime}}($eq。(2)) 乘以常数$c$,并重新归一化概率。请注意，社区结构的变化需要在下一个节点采样之前重新计算$\lambda_i^{(l)}$。重复此采样和更新过程所需的迭代次数。随后，我们进行了等式的计算。( 3)和等式中的优化。( 4 )
4 Experiments
# 4.1 Datasets
表 1 列出了我们实验中使用的数据集的统计信息，并指
出了它们用于评估的任务
Leskovec-Ng$\underline{\mathrm{E}}[$Chen and Hero,2017]:此数据以 5 年为间隔描述了20 年的共同作者信息，i。e。,4-层。在一个层中，如果两个研究人员在 5 年的时间间隔内共同创作，并且研究人员的小组隶属关系是标签
再创新：该数据描述了 144 国家 (节点) 的全球创新指数。该网络包含不同扇区中的 12 层编码相似性，例如。g。基础设施。每个节点的标签表示开发级别
国会投票E[ Schlimmer, 1987] : 根据 1984 年美国国会投票记录，它包括 435 国会议员对 16 法案的投票，从而形成了 16 分层网络。每个节点都被标记为民主党或共和党UCI乳房 X 线摄影 [Elter et al., 2007]:这包含有关961 受试者的乳房 X 线摄影肿块病变的信息，其中 4 层是使用 BI-RADS 评估，病变密度等属性构建的
UCI 平衡量表$\mathbb{E}[$Dua and Karra Taniskidou, 2017]:这总结了心理实验的结果，其中 4 属性 (左重量，左距离， 右重量和右距离)用于构建图层
维克斯$\overline{\mathbb{E}}[$Vickers and Chan, 1981]:这对来自澳大利亚维多利亚州的 29 学生的社会结构进行了建模。每个节点都是一个以性别为标签的学生，三个网络层是根据学生对问卷的回答构建的
LAZEGA$\underline{\mathrm{E}}[$Lazega and others, 2001]:此多路复用网络由企业合作伙伴关系的合作伙伴之间的 3 关系类型 (共同工作，友谊和建议)组成
C。ELEGANS [Chen et al., 2006]:此多路复用网络由
描述一组神经元之间的突触连接的 3 层组成
欧盟航空公司$\mathbb{E}[$Cardillo et al., 2013]:在这个航空运输多路复用网络中，每一层都显示哪些城市与某个欧洲航空公司之间有直飞航班。共有 450 城市 (节点)和 37 航空公司(层)
## 4.2 Baseline Methods
DeepWalk [Perozzi et al., 2014]: 这首先在网络上执行
随机游走，然后使用 Skip-gram 算法来推断嵌入
LINE [Tang et al.,2015]:通过在其成本函数中添加直接链接拟合项来改进 DeepWalk,并通过 2 跳邻域利用高阶关系
Node2Vec$\underline{\mathrm{E}}[$Grover and Leskovec, 2016]:通过在 DeepWalk 中参数化深度优先和广度优先的随机游走，这为某些类型的节点产生了更好的嵌人
PMNE [Liu et al., 2017]: 这采用不同的合并策略来组合来自多层网络中每个层的嵌人。特别是，我们考虑结果聚合策略，因为它通常优于其他变体
MNE [Zhang et al., 2018]:这种最新的嵌入方法针对控制信息传输的常见嵌入进行了联合优化，同时产生了额外的表示来捕获层的不同方面
GenLouvain$\underline{\mathrm{E}}[$Jutla et al.,2011]:这种基于模块化的社区检测算法对于在多层网络中提取有凝聚力的社区非常流行，并在多个数据集上产生最新的结果
4.3 Node Classificatior
在此实验中，我们假设仅针对节点的子集观察到真实标签，并且我们使用在节点嵌人上训练的分类器来预测其余节点的标签。更具体地说，对于每个数据集，我们使用线性 SVM 分类器，并使用 3- 折叠交叉验证(使用较小的分区进行训练)获得报告的结果。节点分类精度作为评价指标。表 2 显示了使用我们提出的方法以及基线方法获得的性能。请注意，对于所提出的方法，我们显示了使用和不使用细化步骤获得的结果。第一个引人注目的观察结果是，与现有的单层和多层方法相比M-DeepWalk 已经在所有数据集上产生了极具竞争力的性能。其次，结合提出的改进策略进一步提高了性能，从而产生了最先进的结果。例如，使用 Reinnovation 数据集，我们的方法实现了 10% 比最近的多路复用网络嵌人 | [
[
"0.0866",
"0.06946",
"0.4902",
"0.0682",
"0.4902",
"0.0833",
"0.0866",
"0.084"
],
[
"0.5146",
"0.07135",
"0.9087",
"0.07135",
"0.9087",
"0.0821",
"0.5146",
"0.0821"
],
[
"0.5146",
"0.084",
"0.9126",
"0.084",
"0.9126",
"0.09784",
"0.5146",
"0.09784"
],
[
"0.12335",
"0.0947",
"0.286",
"0.108",
"0.282",
"0.137",
"0.11926",
"0.1237"
],
[
"0.1993",
"0.09784",
"0.2573",
"0.0966",
"0.2573",
"0.11615",
"0.1993",
"0.11743"
],
[
"0.2632",
"0.0947",
"0.3906",
"0.089",
"0.392",
"0.1111",
"0.2646",
"0.11615"
],
[
"0.3857",
"0.0985",
"0.4436",
"0.101",
"0.4429",
"0.11554",
"0.384",
"0.113"
],
[
"0.511",
"0.09784",
"0.83",
"0.0966",
"0.83",
"0.1111",
"0.511",
"0.11237"
],
[
"0.2778",
"0.108",
"0.2957",
"0.108",
"0.2957",
"0.11554",
"0.2778",
"0.11554"
],
[
"0.3506",
"0.108",
"0.3938",
"0.108",
"0.3938",
"0.1225",
"0.3506",
"0.1225"
],
[
"0.5156",
"0.11237",
"0.911",
"0.11237",
"0.911",
"0.1256",
"0.5156",
"0.1256"
],
[
"0.5146",
"0.1256",
"0.91",
"0.1256",
"0.91",
"0.1395",
"0.5146",
"0.1395"
],
[
"0.2484",
"0.1092",
"0.255",
"0.1092",
"0.255",
"0.11237",
"0.2484",
"0.11237"
],
[
"0.2646",
"0.1338",
"0.451",
"0.1338",
"0.451",
"0.1547",
"0.2646",
"0.1547"
],
[
"0.4656",
"0.1414",
"0.4836",
"0.1414",
"0.4836",
"0.1497",
"0.4656",
"0.1497"
],
[
"0.513",
"0.1383",
"0.91",
"0.1383",
"0.91",
"0.1515",
"0.513",
"0.1515"
],
[
"0.513",
"0.1515",
"0.911",
"0.1528",
"0.911",
"0.1674",
"0.513",
"0.1666"
],
[
"0.4429",
"0.1427",
"0.468",
"0.1427",
"0.468",
"0.1497",
"0.4429",
"0.1497"
],
[
"0.08826",
"0.1674",
"0.486",
"0.1674",
"0.486",
"0.1818",
"0.08826",
"0.1818"
],
[
"0.513",
"0.1666",
"0.911",
"0.1666",
"0.911",
"0.1799",
"0.513",
"0.1799"
],
[
"0.08826",
"0.1812",
"0.4878",
"0.1812",
"0.4878",
"0.1989",
"0.08826",
"0.1989"
],
[
"0.5146",
"0.1787",
"0.8545",
"0.1799",
"0.8545",
"0.1945",
"0.5146",
"0.1932"
],
[
"0.513",
"0.1925",
"0.911",
"0.1945",
"0.911",
"0.2084",
"0.513",
"0.2058"
],
[
"0.08826",
"0.202",
"0.4844",
"0.202",
"0.4844",
"0.2153",
"0.08826",
"0.2153"
],
[
"0.5156",
"0.2084",
"0.9126",
"0.2084",
"0.9126",
"0.2216",
"0.5156",
"0.2216"
],
[
"0.0907",
"0.2172",
"0.1602",
"0.2172",
"0.1602",
"0.2291",
"0.0907",
"0.2291"
],
[
"0.5146",
"0.2203",
"0.7427",
"0.2216",
"0.7427",
"0.2361",
"0.5146",
"0.2349"
],
[
"0.513",
"0.2343",
"0.9126",
"0.2361",
"0.9126",
"0.25",
"0.513",
"0.2474"
],
[
"0.2932",
"0.252",
"0.3945",
"0.2424",
"0.3987",
"0.2644",
"0.2974",
"0.2734"
],
[
"0.5146",
"0.25",
"0.911",
"0.25",
"0.911",
"0.2632",
"0.5146",
"0.2632"
],
[
"0.1887",
"0.2551",
"0.2712",
"0.2551",
"0.2712",
"0.2734",
"0.1887",
"0.2734"
],
[
"0.0972",
"0.2556",
"0.1544",
"0.2583",
"0.1528",
"0.2727",
"0.09644",
"0.2708"
],
[
"0.1478",
"0.2632",
"0.1936",
"0.2683",
"0.1912",
"0.279",
"0.1447",
"0.2734"
],
[
"0.2917",
"0.2583",
"0.2974",
"0.2556",
"0.3015",
"0.2615",
"0.2957",
"0.2632"
],
[
"0.3594",
"0.2651",
"0.398",
"0.2651",
"0.398",
"0.279",
"0.3594",
"0.279"
],
[
"0.5146",
"0.2632",
"0.911",
"0.2632",
"0.911",
"0.2778",
"0.5146",
"0.2778"
],
[
"0.1896",
"0.2747",
"0.1978",
"0.2747",
"0.1978",
"0.2778",
"0.1896",
"0.2778"
],
[
"0.2303",
"0.2766",
"0.259",
"0.2766",
"0.259",
"0.2854",
"0.2303",
"0.2854"
],
[
"0.259",
"0.2747",
"0.2688",
"0.2747",
"0.2688",
"0.279",
"0.259",
"0.279"
],
[
"0.5146",
"0.2778",
"0.6772",
"0.2778",
"0.6772",
"0.2917",
"0.5146",
"0.2917"
],
[
"0.183",
"0.2966",
"0.3545",
"0.2935",
"0.3545",
"0.318",
"0.183",
"0.3213"
],
[
"0.513",
"0.2917",
"0.911",
"0.2917",
"0.911",
"0.3062",
"0.513",
"0.3062"
],
[
"0.5156",
"0.3062",
"0.9126",
"0.3062",
"0.9126",
"0.3193",
"0.5156",
"0.3193"
],
[
"0.2255",
"0.3125",
"0.3096",
"0.3164",
"0.308",
"0.3308",
"0.2239",
"0.327"
],
[
"0.5156",
"0.3193",
"0.7114",
"0.3193",
"0.7114",
"0.3333",
"0.5156",
"0.3333"
],
[
"0.1748",
"0.3289",
"0.367",
"0.3333",
"0.3652",
"0.3694",
"0.1732",
"0.3655"
],
[
"0.5146",
"0.3333",
"0.911",
"0.3333",
"0.911",
"0.3467",
"0.5146",
"0.3467"
],
[
"0.3447",
"0.3403",
"0.4714",
"0.3384",
"0.4714",
"0.3599",
"0.3447",
"0.3623"
],
[
"0.156",
"0.3499",
"0.1929",
"0.3516",
"0.1912",
"0.363",
"0.1544",
"0.361"
],
[
"0.5146",
"0.3467",
"0.826",
"0.3467",
"0.826",
"0.3599",
"0.5146",
"0.3599"
],
[
"0.5146",
"0.3599",
"0.91",
"0.3599",
"0.91",
"0.3738",
"0.5146",
"0.3738"
],
[
"0.1487",
"0.3674",
"0.4143",
"0.3674",
"0.4143",
"0.4028",
"0.1487",
"0.4028"
],
[
"0.513",
"0.3726",
"0.911",
"0.3738",
"0.911",
"0.3882",
"0.513",
"0.387"
],
[
"0.4656",
"0.3806",
"0.486",
"0.3806",
"0.486",
"0.3933",
"0.4656",
"0.3933"
],
[
"0.1544",
"0.387",
"0.1896",
"0.3914",
"0.1871",
"0.4016",
"0.152",
"0.3977"
],
[
"0.312",
"0.3901",
"0.343",
"0.3901",
"0.343",
"0.3984",
"0.312",
"0.3984"
],
[
"0.5156",
"0.3882",
"0.9126",
"0.3882",
"0.9126",
"0.4016",
"0.5156",
"0.4016"
],
[
"0.1936",
"0.3984",
"0.2059",
"0.3984",
"0.2059",
"0.4028",
"0.1936",
"0.4028"
],
[
"0.5156",
"0.4028",
"0.596",
"0.4028",
"0.596",
"0.414",
"0.5156",
"0.414"
],
[
"0.08826",
"0.4124",
"0.4844",
"0.4124",
"0.4844",
"0.4268",
"0.08826",
"0.4268"
],
[
"0.08496",
"0.428",
"0.4878",
"0.4255",
"0.4878",
"0.443",
"0.08496",
"0.4458"
],
[
"0.511",
"0.4255",
"0.701",
"0.4248",
"0.701",
"0.4395",
"0.511",
"0.44"
],
[
"0.08496",
"0.4458",
"0.4844",
"0.4426",
"0.4844",
"0.4602",
"0.08496",
"0.4634"
],
[
"0.513",
"0.4443",
"0.911",
"0.4443",
"0.911",
"0.4583",
"0.513",
"0.4583"
],
[
"0.08417",
"0.459",
"0.4878",
"0.4602",
"0.4878",
"0.4817",
"0.08417",
"0.481"
],
[
"0.5156",
"0.459",
"0.857",
"0.459",
"0.857",
"0.473",
"0.5156",
"0.473"
],
[
"0.1936",
"0.4673",
"0.2091",
"0.4673",
"0.2091",
"0.4749",
"0.1936",
"0.4749"
],
[
"0.5146",
"0.4717",
"0.911",
"0.4717",
"0.911",
"0.486",
"0.5146",
"0.486"
],
[
"0.08905",
"0.4817",
"0.4714",
"0.4817",
"0.4714",
"0.4968",
"0.08905",
"0.4968"
],
[
"0.513",
"0.486",
"0.911",
"0.486",
"0.911",
"0.5",
"0.513",
"0.5"
],
[
"0.0866",
"0.4976",
"0.486",
"0.4976",
"0.486",
"0.515",
"0.0866",
"0.515"
],
[
"0.5156",
"0.502",
"0.5786",
"0.502",
"0.5786",
"0.513",
"0.5156",
"0.513"
],
[
"0.08826",
"0.5146",
"0.4836",
"0.5146",
"0.4836",
"0.528",
"0.08826",
"0.528"
],
[
"0.5107",
"0.5103",
"0.911",
"0.512",
"0.911",
"0.53",
"0.5107",
"0.528"
],
[
"0.08826",
"0.529",
"0.2737",
"0.529",
"0.2737",
"0.5425",
"0.08826",
"0.5425"
],
[
"0.513",
"0.5264",
"0.911",
"0.528",
"0.911",
"0.5425",
"0.513",
"0.5415"
],
[
"0.2769",
"0.534",
"0.4043",
"0.56",
"0.3906",
"0.5977",
"0.2646",
"0.5713"
],
[
"0.5146",
"0.5405",
"0.7427",
"0.5415",
"0.7427",
"0.556",
"0.5146",
"0.5547"
],
[
"0.1748",
"0.553",
"0.2793",
"0.547",
"0.282",
"0.5674",
"0.1772",
"0.5737"
],
[
"0.343",
"0.552",
"0.3857",
"0.552",
"0.3857",
"0.563",
"0.343",
"0.563"
],
[
"0.2646",
"0.558",
"0.2917",
"0.561",
"0.2893",
"0.5713",
"0.2615",
"0.5684"
],
[
"0.4656",
"0.5596",
"0.4836",
"0.5596",
"0.4836",
"0.5684",
"0.4656",
"0.5684"
],
[
"0.513",
"0.5537",
"0.911",
"0.556",
"0.911",
"0.5693",
"0.513",
"0.5674"
],
[
"0.2115",
"0.5645",
"0.2695",
"0.561",
"0.2712",
"0.5806",
"0.2133",
"0.584"
],
[
"0.281",
"0.5664",
"0.3325",
"0.5684",
"0.3325",
"0.58",
"0.281",
"0.5776"
],
[
"0.33",
"0.5737",
"0.3423",
"0.5737",
"0.3423",
"0.577",
"0.33",
"0.577"
],
[
"0.5156",
"0.5684",
"0.9126",
"0.5693",
"0.9126",
"0.584",
"0.5156",
"0.5835"
],
[
"0.5146",
"0.582",
"0.79",
"0.582",
"0.79",
"0.595",
"0.5146",
"0.595"
],
[
"0.0866",
"0.589",
"0.4878",
"0.589",
"0.4878",
"0.607",
"0.0866",
"0.607"
],
[
"0.513",
"0.5967",
"0.911",
"0.5967",
"0.911",
"0.61",
"0.513",
"0.61"
],
[
"0.08905",
"0.607",
"0.4844",
"0.607",
"0.4844",
"0.6206",
"0.08905",
"0.6206"
],
[
"0.511",
"0.6094",
"0.9087",
"0.61",
"0.9087",
"0.625",
"0.511",
"0.6235"
],
[
"0.08826",
"0.6226",
"0.486",
"0.6226",
"0.486",
"0.636",
"0.08826",
"0.636"
],
[
"0.5146",
"0.625",
"0.744",
"0.625",
"0.744",
"0.638",
"0.5146",
"0.638"
],
[
"0.08496",
"0.634",
"0.4878",
"0.6333",
"0.4878",
"0.651",
"0.08496",
"0.6514"
],
[
"0.513",
"0.638",
"0.911",
"0.638",
"0.911",
"0.6514",
"0.513",
"0.6514"
],
[
"0.08496",
"0.6484",
"0.4878",
"0.6484",
"0.4878",
"0.6694",
"0.08496",
"0.6694"
],
[
"0.5156",
"0.653",
"0.911",
"0.653",
"0.911",
"0.6665",
"0.5156",
"0.6665"
],
[
"0.08826",
"0.6675",
"0.4844",
"0.6675",
"0.4844",
"0.681",
"0.08826",
"0.681"
],
[
"0.513",
"0.6665",
"0.808",
"0.6655",
"0.808",
"0.6797",
"0.513",
"0.681"
],
[
"0.08826",
"0.6816",
"0.4836",
"0.6816",
"0.4836",
"0.6963",
"0.08826",
"0.6963"
],
[
"0.08905",
"0.696",
"0.4844",
"0.696",
"0.4844",
"0.709",
"0.08905",
"0.709"
],
[
"0.513",
"0.69",
"0.7134",
"0.6914",
"0.7134",
"0.705",
"0.513",
"0.704"
],
[
"0.08826",
"0.709",
"0.486",
"0.709",
"0.486",
"0.7266",
"0.08826",
"0.7266"
],
[
"0.5146",
"0.709",
"0.9126",
"0.709",
"0.9126",
"0.7227",
"0.5146",
"0.7227"
],
[
"0.08905",
"0.726",
"0.486",
"0.726",
"0.486",
"0.7407",
"0.08905",
"0.7407"
],
[
"0.5146",
"0.7236",
"0.911",
"0.7236",
"0.911",
"0.7373",
"0.5146",
"0.7373"
],
[
"0.08826",
"0.7393",
"0.3694",
"0.7393",
"0.3694",
"0.7524",
"0.08826",
"0.7524"
],
[
"0.5146",
"0.7373",
"0.911",
"0.7373",
"0.911",
"0.7505",
"0.5146",
"0.7505"
],
[
"0.5156",
"0.7505",
"0.911",
"0.7505",
"0.911",
"0.7646",
"0.5156",
"0.7646"
],
[
"0.08496",
"0.768",
"0.2296",
"0.768",
"0.2296",
"0.782",
"0.08496",
"0.782"
],
[
"0.513",
"0.763",
"0.911",
"0.7646",
"0.911",
"0.779",
"0.513",
"0.778"
],
[
"0.5156",
"0.779",
"0.911",
"0.779",
"0.911",
"0.7925",
"0.5156",
"0.7925"
],
[
"0.0866",
"0.7886",
"0.2051",
"0.7886",
"0.2051",
"0.802",
"0.0866",
"0.802"
],
[
"0.5156",
"0.7925",
"0.9126",
"0.7925",
"0.9126",
"0.806",
"0.5156",
"0.806"
],
[
"0.08826",
"0.8047",
"0.486",
"0.8047",
"0.486",
"0.8193",
"0.08826",
"0.8193"
],
[
"0.5146",
"0.8047",
"0.911",
"0.8037",
"0.911",
"0.8184",
"0.5146",
"0.8193"
],
[
"0.0907",
"0.8213",
"0.2646",
"0.8213",
"0.2646",
"0.832",
"0.0907",
"0.832"
],
[
"0.517",
"0.8193",
"0.9126",
"0.8193",
"0.9126",
"0.8335",
"0.517",
"0.8335"
],
[
"0.08496",
"0.832",
"0.4844",
"0.8335",
"0.4844",
"0.8477",
"0.08496",
"0.8467"
],
[
"0.513",
"0.832",
"0.911",
"0.8335",
"0.911",
"0.8477",
"0.513",
"0.8467"
],
[
"0.0866",
"0.8467",
"0.4805",
"0.8477",
"0.4805",
"0.8623",
"0.0866",
"0.8613"
],
[
"0.5146",
"0.8477",
"0.91",
"0.8477",
"0.91",
"0.8613",
"0.5146",
"0.8613"
],
[
"0.08905",
"0.8613",
"0.4844",
"0.8613",
"0.4844",
"0.8755",
"0.08905",
"0.8755"
],
[
"0.5146",
"0.8623",
"0.911",
"0.8623",
"0.911",
"0.8755",
"0.5146",
"0.8755"
],
[
"0.08826",
"0.8755",
"0.3816",
"0.8755",
"0.3816",
"0.8896",
"0.08826",
"0.8896"
],
[
"0.5146",
"0.8755",
"0.911",
"0.8755",
"0.911",
"0.8896",
"0.5146",
"0.8896"
]
] | [
"络的情况，提供了多切片模块化E[Muchaetal.,2010]：",
"再创新：该数据描述了144国家（节点）的全球创新指",
"数。该网络包含不同扇区中的12层编码相似性，例如。",
"Qml=",
"Z",
"(eg-ynng)",
"s(l,m)",
"g。基础设施。每个节点的标签表示开发级别",
"Cii",
"2r",
"国会投票E[Schlimmer，1987]：根据1984年美国国会投",
"票记录，它包括435国会议员对16法案的投票，从而形",
"",
"+8(i,5)ogm)s(c9,cm),",
"(6)",
"成了16分层网络。每个节点都被标记为民主党或共和党",
"UCI乳房X线摄影[Eliteretal.,2007]：这包含有关",
"",
"其中μ是层之间的归一化因子，（）是层之间的分辨",
"961受试者的乳房X线摄影肿块病变的信息，其中4层",
"率参数。。e。，w）和u(m），其他参数是从（5）到每个",
"是使用BI-RADS评估，病变密度等属性构建的",
"UCI平衡量表E[DuaandKarraTaniskidou,2017]：这",
"层1的直接扩展。重写（6）以描述每个节点对模块化功",
"总结了心理实验的结果，其中4属性（左重量，左距离，",
"能的贡献：",
"右重量和右距离）用于构建图层",
"维克斯E/VickersandChan，1981]：这对来自澳大利亚",
"-yanng",
"维多利亚州的29学生的社会结构进行了建模。每个节",
"ZEE",
"Qmulti=",
"2uL",
"",
"2rt",
"点都是一个以性别为标签的学生，三个网络层是根据学",
"",
"iEc",
"",
"生对问卷的回答构建的",
"+m",
"LAZEGAE[Lazegaandothers，2001]：此多路复用网络",
"由企业合作伙伴关系的合作伙伴之间的3关系类型（共",
"imtljec()",
"同工作，友谊和建议）组成",
"[(m- )",
"C。ELEGANS[Chenetal,2006]：此多路复用网络由",
")+Zonc()",
"2u",
"描述一组神经元之间的突触连接的3层组成",
"欧盟航空公司E[Cardilloetal.,2013]：在这个航空运输",
"MM(m-m+am)",
"多路复用网络中，每一层都显示哪些城市与某个欧洲航",
"(7)",
"2u2",
"T",
"空公司之间有直飞航班。共有450城市（节点）和37航",
"",
"空公司（层）",
"这实质上将[DuchandArenas，2005]中的“适合度分",
"数”扩展到了多层网络。这里，我们假设,㎡是所",
"4.2Baseline Methods",
"有节点的常量。我们使用nc，来表示1层中将节点)",
"DeepWalk[Perozzietal.，2014]：这首先在网络上执行",
"连接到标签为J的其他节点的边数。另一方面，nc)",
"随机游走，然后使用Skip-gram算法来推断嵌入",
"",
"LINE[Tangetal.,2015]：通过在其成本函数中添加直",
"表示的对应节点的数量，这些节点也共享标签",
"接链接拟合项来改进DeepWalk，并通过2跳邻域利用",
"EQ中的最终求和项。（7）定义了）对多切片模块化",
"高阶关系",
"的贡献。现在，我们通过分别对层内和层间贡献项进行",
"Node2Vec[E[Grover and Leskovec, 2016] ：通过在 Deep-",
"归一化来测量适应度分数：",
"Walk中参数化深度优先和广度优先的随机游走，这为某",
"+o",
"些类型的节点产生了更好的嵌人",
"=nc-",
"nc()",
"2",
"(8)",
"PMNE[Liuetal.，2017]：这采用不同的合并策略来组",
"n",
"TI",
"",
"合来自多层网络中每个层的嵌人。特别是，我们考虑结",
"果聚合策略，因为它通常优于其他变体",
"其中n,m）是指u）的对应项的数量。按照[Duchand",
"MNE[Zhangetal.,2018]：这种最新的嵌入方法针对",
"Arenas.2005:Boettcher，2001]中的概率选择过程，V中",
"控制信息传输的常见嵌人进行了联合优化，同时产生了",
"的节点首先根据其适应度得分按升序排序，然后根据概",
"额外的表示来捕获层的不同方面",
"率进行采样：p(s)αs-T其中s是排序后的节点等级和",
"GenLouvainE[Jutlaetal.,2011]：这种基于模块化的社",
"T～1+m(IVI)。接下来，我们将采样节点讠移动到社区",
"区检测算法对于在多层网络中提取有凝聚力的社区非常",
"K，这给出了Qmulti的最大增益。为了反映整个细化过",
"流行，并在多个数据集上产生最新的结果",
"程的更新分配，我们增加了社区可能性qik（eq。（2））",
"乘以常数，并重新归一化概率。请注意，社区结构的",
"4.3Node Classification",
"变化需要在下一个节点采样之前重新计算入(）。重复此",
"在此实验中，我们假设仅针对节点的子集观察到真实标",
"采样和更新过程所需的迭代次数。随后，我们进行了等",
"签，并且我们使用在节点嵌入上训练的分类器来预测其",
"式的计算。（3）和等式中的优化。（4）",
"余节点的标签。更具体地说，对于每个数据集，我们使",
"用线性SVM分类器，并使用3-折叠交叉验证（使用",
"4Experiments",
"较小的分区进行训练）获得报告的结果。节点分类精度",
"作为评价指标。表2显示了使用我们提出的方法以及",
"4.1 Datasets",
"基线方法获得的性能。请注意，对于所提出的方法，我",
"表1列出了我们实验中使用的数据集的统计信息，并指",
"们显示了使用和不使用细化步骤获得的结果。第一个引",
"出了它们用于评估的任务",
"人注目的观察结果是，与现有的单层和多层方法相比，",
"Leskovec-NgE[ChenandHero,2017]：此数据以5年为",
"M-DeepWalk已经在所有数据集上产生了极具竞争力的",
"间隔描述了20年的共同作者信息，i。e。，4-层。在一",
"性能。其次，结合提出的改进策略进一步提高了性能，从",
"个层中，如果两个研究人员在5年的时间间隔内共同创",
"而产生了最先进的结果。例如，使用Reinnovation数据",
"作，并且研究人员的小组隶属关系是标签",
"集，我们的方法实现了10%比最近的多路复用网络嵌入"
] | {
"arxiv_id": "1811.12156",
"language": "zh",
"page_idx": 3
} |
||
1811.12156-zh-4 | 1811.12156-zh-4.jpg | Table 2: Node Classification- Accuracy $(\%)$ of the proposed approach in predicting node labels, when compared to baseline
methods.
<table>
<tbody>
<tr>
<th rowspan="2">Method</th>
<th colspan="5">Dataset</th>
</tr>
<tr>
<th>$\overline{{\mathrm{Leskovec-Ng}}}$</th>
<th>Reinnovation</th>
<th>Congress Votes</th>
<th>Mammography</th>
<th>Balance Scal$\epsilon$</th>
</tr>
<tr>
<td>DeepWalk</td>
<td>99.2</td>
<td>74.7</td>
<td>99.8</td>
<td>80.6</td>
<td>90.9</td>
</tr>
<tr>
<td>$LINE$</td>
<td>91.1</td>
<td>69.1</td>
<td>98.0</td>
<td>80.9</td>
<td>78.5</td>
</tr>
<tr>
<td>Node2Vec</td>
<td>96.3</td>
<td>76.0</td>
<td>92.4</td>
<td>80.2</td>
<td>89.3</td>
</tr>
<tr>
<td>PMNE</td>
<td>94.5</td>
<td>77.4</td>
<td>98.4</td>
<td>78.5</td>
<td>91.1</td>
</tr>
<tr>
<td>MNE</td>
<td>92.4</td>
<td>75.0</td>
<td> </td>
<td>74.3</td>
<td>82.4</td>
</tr>
<tr>
<td>$\operatorname{Proposed}$ (w/o refine$j$</td>
<td>99.7</td>
<td>76.0</td>
<td>100</td>
<td>81.3</td>
<td>90.5</td>
</tr>
<tr>
<td>Proposed ${^{\prime}W/}$ refine</td>
<td>100</td>
<td>85.1</td>
<td>100</td>
<td>81.5</td>
<td>92.1</td>
</tr>
</tbody>
</table>
Table 3: Link Prediction - AUROC score of our approach in accepting positive links and rejecting negative links in multi-
layered networks.
<table>
<tbody>
<tr>
<th rowspan="2">Dataset</th>
<th colspan="5">Method</th>
</tr>
<tr>
<th>DeepWalk</th>
<th>$LINE$</th>
<th>Node2Vec</th>
<th>PMNE</th>
<th>$\operatorname{Proposed}$</th>
</tr>
<tr>
<td>Leskovec- Ng</td>
<td>0.84</td>
<td>0.62</td>
<td>0.71</td>
<td>0.49</td>
<td>0.84</td>
</tr>
<tr>
<td>Reinnovation</td>
<td>0.99</td>
<td>0.78</td>
<td>0.99</td>
<td>0.78</td>
<td>0.99</td>
</tr>
<tr>
<td>Congress Votes</td>
<td>1.0</td>
<td>0.99</td>
<td>1.0</td>
<td>0.79</td>
<td>1.0</td>
</tr>
<tr>
<td>Mammography</td>
<td>1.0</td>
<td>1.0</td>
<td>1.0</td>
<td>0.77</td>
<td>1.0</td>
</tr>
<tr>
<td>Balance Scale</td>
<td>1.0</td>
<td>1.0</td>
<td>1.0</td>
<td>0.82</td>
<td>1.0</td>
</tr>
<tr>
<td>Vikcers</td>
<td>0.83</td>
<td>0.68</td>
<td>0.76</td>
<td>0.84</td>
<td>0.89</td>
</tr>
<tr>
<td>LAZEGA</td>
<td>0.88</td>
<td>0.69</td>
<td>0.8</td>
<td>0.82</td>
<td>0.91</td>
</tr>
<tr>
<td>C.ELEGANS</td>
<td>0.93</td>
<td>0.77</td>
<td>0.89</td>
<td>0.75</td>
<td>0.944</td>
</tr>
</tbody>
</table>
4.5 Multi-layered Community Detection
方法$\mathbb{E}[$Zhang et al., 2018] 更高的准确性。请注意，在国会投票数据上，公开发布的 CPU 版本的 MNE 要么没有稳定收敛，要么无法生成有效的嵌入
4.4 Link Prediction
在最后的实验中，我们从产生内聚社区结构的角度评估了所提出的多层社区检测方法。如 2 节中所定义，此公式可以形成由来自多个层的节点组成的社区。我们为这项任务考虑的多路复用网络是未标记的，因此我们求助于无监督学习。为了进行评估，我们利用模块化度量。大多数现有的多层网络嵌人方法要么没有明确优化模块化， 要么限制跨层的社区分配相同。因此，我们将我们的结果与流行的 GenLouvain 算法进行比较。社区数量是一$\begin{array}{c}\text{个重要参数，为了进行全面比较，我们通过改变}\gamma\\\\\text{立生成估用 ConI cumin }\#\text{成的下目数具的社区}\end{array}$和$\begin{array}{c}\\{\text{个重要参数，为了进行全面比较，我们通过改变}\gamma\text{和 }\sigma}\\{\text{对上式估用 }C_{\mathrm{onI~cunin~}}\#\text{成的工目数具的社区}}\end{array}$ 来生成使用 GenLouvain 生成的不同数量的社区。对于这些情况中的每一个，我们对我们方法中的嵌入执行了$k$ -means, 具有相同数量的社区。GenLouvain 的性能和所提出的方法在图 3 中可视化。我们观察到 GenLouvain 的表现对所选社区的数量高度敏感。在许多情况下，该算法无法发现有意义的社区。相反，我们的方法一直是更好和稳定的
对于每个数据集，我们在每个层中随机选择一个边缘子集，并将其从多层网络中删除。接下来，我们使用不完整网络提取节点嵌入，并估计每个缺失边的可能性。虽然随机删除的边缘被标记为正，但那些最初不存在的边缘是负的。我们通过算法接受正链接而拒绝负链接的能力来评估链接预测性能。似然性估计为与边对应的两个节点之间的余弦相似度。在我们的实验中，我们对每个层中的边缘集进行了 5 倍交叉验证，并将 AUROC (ROC 曲线下的面积)报告为评估指标。与节点分类任务相反， 此任务在很大程度上依赖于每个层中的局部结构，因此单层嵌人形成了强大的基线。令人惊讶的是，通过超参数调整 (e。g。足够的步行次数),像 DeepWalk 这样的简单基线比 [Zhang et al.,2018] 中观察到的更具竞争力。如表 3 所示，对于 9 数据集中的 5 , DeepWalk 执行的方法与更复杂的方法一样好，没有任何额外的层间信息。但是，在需要从其他层传输信息以可靠地预测链路的情况下，所提出的方法提供了改进，例如在 Vickers, LAZEGA 和 C。ELELGANS 数据集。另一个观察结果是，包括细化步骤并没有在此任务中提供明显的改进
# 5 Conclusions
本文提出了一种从多层网络中提取节点嵌入的新方法。虽然局部自适应的层间边缘构造有助于 M-DeepWalk 的复杂依赖性建模，但细化阶段在潜在空间中产生了更多的内聚结构。实证研究表明，所提出的方法在包括链接预测，节点分类和多层社区检测在内的各种任务中都非常有效。除了优于现有的单层和多层嵌人方法外，我们还证明了所提出的方法可扩展到大量层并产生高效的节 | [
[
"0.0866",
"0.0663",
"0.9126",
"0.0663",
"0.9126",
"0.0802",
"0.0866",
"0.0802"
],
[
"0.0866",
"0.0802",
"0.1478",
"0.0802",
"0.1478",
"0.09155",
"0.0866",
"0.09155"
],
[
"0.5557",
"0.1048",
"0.614",
"0.1067",
"0.613",
"0.1218",
"0.5537",
"0.1193"
],
[
"0.1691",
"0.11554",
"0.2296",
"0.11554",
"0.2296",
"0.127",
"0.1691",
"0.127"
],
[
"0.29",
"0.1218",
"0.5107",
"0.1225",
"0.5107",
"0.137",
"0.29",
"0.1364"
],
[
"0.511",
"0.1225",
"0.6406",
"0.1225",
"0.6406",
"0.137",
"0.511",
"0.137"
],
[
"0.638",
"0.1225",
"0.7695",
"0.125",
"0.7695",
"0.1383",
"0.638",
"0.1364"
],
[
"0.763",
"0.1237",
"0.876",
"0.1237",
"0.876",
"0.137",
"0.763",
"0.137"
],
[
"0.3245",
"0.1427",
"0.3594",
"0.1427",
"0.3594",
"0.1578",
"0.3245",
"0.1578"
],
[
"0.4355",
"0.1427",
"0.4697",
"0.1427",
"0.4697",
"0.1578",
"0.4355",
"0.1578"
],
[
"0.5557",
"0.1427",
"0.5923",
"0.1427",
"0.5923",
"0.1578",
"0.5557",
"0.1578"
],
[
"0.6846",
"0.1427",
"0.7197",
"0.1427",
"0.7197",
"0.1578",
"0.6846",
"0.1578"
],
[
"0.8057",
"0.1401",
"0.8423",
"0.1427",
"0.841",
"0.1578",
"0.8037",
"0.1559"
],
[
"0.1765",
"0.159",
"0.2222",
"0.159",
"0.2222",
"0.1736",
"0.1765",
"0.1736"
],
[
"0.1626",
"0.1445",
"0.2378",
"0.1445",
"0.2378",
"0.1559",
"0.1626",
"0.1559"
],
[
"0.326",
"0.1604",
"0.3586",
"0.1604",
"0.3586",
"0.1724",
"0.326",
"0.1724"
],
[
"0.4355",
"0.1604",
"0.468",
"0.1604",
"0.468",
"0.1724",
"0.4355",
"0.1724"
],
[
"0.556",
"0.1604",
"0.591",
"0.1604",
"0.591",
"0.1724",
"0.556",
"0.1724"
],
[
"0.6855",
"0.1604",
"0.7183",
"0.1604",
"0.7183",
"0.1724",
"0.6855",
"0.1724"
],
[
"0.808",
"0.1604",
"0.841",
"0.1604",
"0.841",
"0.1724",
"0.808",
"0.1724"
],
[
"0.1609",
"0.1755",
"0.2378",
"0.1755",
"0.2378",
"0.1901",
"0.1609",
"0.1901"
],
[
"0.3245",
"0.1755",
"0.3594",
"0.1755",
"0.3594",
"0.1913",
"0.3245",
"0.1913"
],
[
"0.5557",
"0.1755",
"0.5923",
"0.1755",
"0.5923",
"0.1913",
"0.5557",
"0.1913"
],
[
"0.6846",
"0.1755",
"0.7197",
"0.1755",
"0.7197",
"0.1913",
"0.6846",
"0.1913"
],
[
"0.806",
"0.1755",
"0.8423",
"0.1755",
"0.8423",
"0.1913",
"0.806",
"0.1913"
],
[
"0.1732",
"0.1925",
"0.2279",
"0.1925",
"0.2279",
"0.207",
"0.1732",
"0.207"
],
[
"0.437",
"0.1768",
"0.468",
"0.1768",
"0.468",
"0.1895",
"0.437",
"0.1895"
],
[
"0.3245",
"0.1925",
"0.3594",
"0.1925",
"0.3594",
"0.2084",
"0.3245",
"0.2084"
],
[
"0.4355",
"0.1925",
"0.4714",
"0.1925",
"0.4714",
"0.2084",
"0.4355",
"0.2084"
],
[
"0.5557",
"0.1925",
"0.5923",
"0.1925",
"0.5923",
"0.2084",
"0.5557",
"0.2084"
],
[
"0.6846",
"0.1925",
"0.7197",
"0.1925",
"0.7197",
"0.2084",
"0.6846",
"0.2084"
],
[
"0.806",
"0.1925",
"0.841",
"0.1925",
"0.841",
"0.2084",
"0.806",
"0.2084"
],
[
"0.1772",
"0.209",
"0.2222",
"0.209",
"0.2222",
"0.2235",
"0.1772",
"0.2235"
],
[
"0.3245",
"0.209",
"0.3594",
"0.209",
"0.3594",
"0.2247",
"0.3245",
"0.2247"
],
[
"0.4355",
"0.209",
"0.4714",
"0.209",
"0.4714",
"0.2247",
"0.4355",
"0.2247"
],
[
"0.6846",
"0.209",
"0.7197",
"0.209",
"0.7197",
"0.2247",
"0.6846",
"0.2247"
],
[
"0.806",
"0.209",
"0.8423",
"0.209",
"0.8423",
"0.2247",
"0.806",
"0.2247"
],
[
"0.5703",
"0.2153",
"0.5767",
"0.2153",
"0.5767",
"0.2203",
"0.5703",
"0.2203"
],
[
"0.11926",
"0.2299",
"0.2778",
"0.2299",
"0.2778",
"0.2443",
"0.11926",
"0.2443"
],
[
"0.4355",
"0.2291",
"0.4714",
"0.2291",
"0.4714",
"0.2443",
"0.4355",
"0.2443"
],
[
"0.5605",
"0.2299",
"0.589",
"0.2299",
"0.589",
"0.2443",
"0.5605",
"0.2443"
],
[
"0.6855",
"0.2311",
"0.7197",
"0.2311",
"0.7197",
"0.2437",
"0.6855",
"0.2437"
],
[
"0.808",
"0.2311",
"0.841",
"0.2311",
"0.841",
"0.2437",
"0.808",
"0.2437"
],
[
"0.4346",
"0.2456",
"0.4697",
"0.2456",
"0.4697",
"0.2615",
"0.4346",
"0.2615"
],
[
"0.326",
"0.2311",
"0.357",
"0.2311",
"0.357",
"0.2437",
"0.326",
"0.2437"
],
[
"0.6846",
"0.2456",
"0.7197",
"0.2456",
"0.7197",
"0.2615",
"0.6846",
"0.2615"
],
[
"0.806",
"0.2456",
"0.841",
"0.2456",
"0.841",
"0.2615",
"0.806",
"0.2615"
],
[
"0.3284",
"0.2468",
"0.357",
"0.2468",
"0.357",
"0.2615",
"0.3284",
"0.2615"
],
[
"0.12335",
"0.2468",
"0.273",
"0.2468",
"0.273",
"0.2615",
"0.12335",
"0.2615"
],
[
"0.5605",
"0.2468",
"0.589",
"0.2468",
"0.589",
"0.2615",
"0.5605",
"0.2615"
],
[
"0.0866",
"0.2766",
"0.911",
"0.2766",
"0.911",
"0.2903",
"0.0866",
"0.2903"
],
[
"0.0866",
"0.289",
"0.2035",
"0.289",
"0.2035",
"0.303",
"0.0866",
"0.303"
],
[
"0.5317",
"0.3164",
"0.591",
"0.3164",
"0.591",
"0.3308",
"0.5317",
"0.3308"
],
[
"0.2656",
"0.3225",
"0.326",
"0.3245",
"0.3245",
"0.339",
"0.2646",
"0.3372"
],
[
"0.3677",
"0.3333",
"0.4468",
"0.3333",
"0.4468",
"0.348",
"0.3677",
"0.348"
],
[
"0.455",
"0.334",
"0.5034",
"0.334",
"0.5034",
"0.3452",
"0.455",
"0.3452"
],
[
"0.5146",
"0.3308",
"0.603",
"0.3333",
"0.603",
"0.348",
"0.5146",
"0.3452"
],
[
"0.6055",
"0.334",
"0.6665",
"0.334",
"0.6665",
"0.3452",
"0.6055",
"0.3452"
],
[
"0.6807",
"0.3352",
"0.755",
"0.3352",
"0.755",
"0.3467",
"0.6807",
"0.3467"
],
[
"0.2484",
"0.3516",
"0.3423",
"0.3542",
"0.3423",
"0.3687",
"0.2484",
"0.3662"
],
[
"0.3896",
"0.3542",
"0.4224",
"0.3542",
"0.4224",
"0.3662",
"0.3896",
"0.3662"
],
[
"0.619",
"0.3528",
"0.6543",
"0.3528",
"0.6543",
"0.3687",
"0.619",
"0.3687"
],
[
"0.7017",
"0.3542",
"0.7363",
"0.3542",
"0.7363",
"0.3662",
"0.7017",
"0.3662"
],
[
"0.5415",
"0.3547",
"0.572",
"0.3547",
"0.572",
"0.3674",
"0.5415",
"0.3674"
],
[
"0.4656",
"0.3547",
"0.4983",
"0.3547",
"0.4983",
"0.3674",
"0.4656",
"0.3674"
],
[
"0.246",
"0.3687",
"0.3447",
"0.3706",
"0.3447",
"0.385",
"0.246",
"0.3833"
],
[
"0.3882",
"0.3694",
"0.423",
"0.3694",
"0.423",
"0.385",
"0.3882",
"0.385"
],
[
"0.4656",
"0.3718",
"0.4983",
"0.3718",
"0.4983",
"0.3838",
"0.4656",
"0.3838"
],
[
"0.5415",
"0.3718",
"0.5728",
"0.3718",
"0.5728",
"0.3838",
"0.5415",
"0.3838"
],
[
"0.621",
"0.3718",
"0.653",
"0.3718",
"0.653",
"0.3838",
"0.621",
"0.3838"
],
[
"0.701",
"0.3694",
"0.7363",
"0.3694",
"0.7363",
"0.385",
"0.701",
"0.385"
],
[
"0.464",
"0.3865",
"0.5",
"0.3865",
"0.5",
"0.4016",
"0.464",
"0.4016"
],
[
"0.619",
"0.3865",
"0.6543",
"0.3865",
"0.6543",
"0.4016",
"0.619",
"0.4016"
],
[
"0.3938",
"0.387",
"0.4192",
"0.387",
"0.4192",
"0.401",
"0.3938",
"0.401"
],
[
"0.2402",
"0.387",
"0.3506",
"0.387",
"0.3506",
"0.4016",
"0.2402",
"0.4016"
],
[
"0.546",
"0.387",
"0.5703",
"0.387",
"0.5703",
"0.401",
"0.546",
"0.401"
],
[
"0.706",
"0.387",
"0.7344",
"0.387",
"0.7344",
"0.4016",
"0.706",
"0.4016"
],
[
"0.2385",
"0.4016",
"0.3489",
"0.4048",
"0.3489",
"0.4192",
"0.2385",
"0.4167"
],
[
"0.4697",
"0.3997",
"0.5",
"0.404",
"0.4958",
"0.4192",
"0.4656",
"0.4153"
],
[
"0.546",
"0.4028",
"0.5703",
"0.4028",
"0.5703",
"0.4167",
"0.546",
"0.4167"
],
[
"0.619",
"0.4028",
"0.6543",
"0.4028",
"0.6543",
"0.4185",
"0.619",
"0.4185"
],
[
"0.7075",
"0.4028",
"0.734",
"0.4028",
"0.734",
"0.4167",
"0.7075",
"0.4167"
],
[
"0.619",
"0.4192",
"0.6543",
"0.4192",
"0.6543",
"0.435",
"0.619",
"0.435"
],
[
"0.3938",
"0.404",
"0.4192",
"0.404",
"0.4192",
"0.4167",
"0.3938",
"0.4167"
],
[
"0.3938",
"0.4204",
"0.4192",
"0.4204",
"0.4192",
"0.4343",
"0.3938",
"0.4343"
],
[
"0.4697",
"0.4204",
"0.4983",
"0.4204",
"0.4983",
"0.435",
"0.4697",
"0.435"
],
[
"0.546",
"0.4204",
"0.5703",
"0.4204",
"0.5703",
"0.4343",
"0.546",
"0.4343"
],
[
"0.706",
"0.4204",
"0.7344",
"0.4204",
"0.7344",
"0.435",
"0.706",
"0.435"
],
[
"0.2688",
"0.435",
"0.326",
"0.4375",
"0.3245",
"0.452",
"0.267",
"0.4495"
],
[
"0.246",
"0.4216",
"0.3447",
"0.4216",
"0.3447",
"0.433",
"0.246",
"0.433"
],
[
"0.464",
"0.4363",
"0.5",
"0.4363",
"0.5",
"0.452",
"0.464",
"0.452"
],
[
"0.54",
"0.4363",
"0.576",
"0.4363",
"0.576",
"0.452",
"0.54",
"0.452"
],
[
"0.621",
"0.4375",
"0.6543",
"0.4375",
"0.6543",
"0.4495",
"0.621",
"0.4495"
],
[
"0.7017",
"0.4375",
"0.7363",
"0.4375",
"0.7363",
"0.4495",
"0.7017",
"0.4495"
],
[
"0.5435",
"0.4495",
"0.572",
"0.4526",
"0.569",
"0.4697",
"0.54",
"0.4666"
],
[
"0.3896",
"0.4375",
"0.423",
"0.4375",
"0.423",
"0.4495",
"0.3896",
"0.4495"
],
[
"0.464",
"0.4507",
"0.501",
"0.4526",
"0.4983",
"0.4685",
"0.4631",
"0.4653"
],
[
"0.619",
"0.452",
"0.6543",
"0.452",
"0.6543",
"0.4673",
"0.619",
"0.4673"
],
[
"0.3896",
"0.4539",
"0.4224",
"0.4539",
"0.4224",
"0.4666",
"0.3896",
"0.4666"
],
[
"0.2607",
"0.4539",
"0.33",
"0.4539",
"0.33",
"0.4653",
"0.2607",
"0.4653"
],
[
"0.7017",
"0.4539",
"0.7344",
"0.4539",
"0.7344",
"0.4666",
"0.7017",
"0.4666"
],
[
"0.3882",
"0.4685",
"0.423",
"0.4685",
"0.423",
"0.4841",
"0.3882",
"0.4841"
],
[
"0.4631",
"0.4685",
"0.5",
"0.4685",
"0.5",
"0.4841",
"0.4631",
"0.4841"
],
[
"0.54",
"0.4685",
"0.574",
"0.4685",
"0.574",
"0.4841",
"0.54",
"0.4841"
],
[
"0.619",
"0.4685",
"0.6543",
"0.4685",
"0.6543",
"0.4841",
"0.619",
"0.4841"
],
[
"0.6978",
"0.4685",
"0.7417",
"0.4685",
"0.7417",
"0.4841",
"0.6978",
"0.4841"
],
[
"0.08905",
"0.5103",
"0.4844",
"0.5103",
"0.4844",
"0.5244",
"0.08905",
"0.5244"
],
[
"0.246",
"0.4702",
"0.3423",
"0.4702",
"0.3423",
"0.4817",
"0.246",
"0.4817"
],
[
"0.513",
"0.511",
"0.556",
"0.511",
"0.556",
"0.523",
"0.513",
"0.523"
],
[
"0.5557",
"0.5103",
"0.8447",
"0.5103",
"0.8447",
"0.5234",
"0.5557",
"0.5234"
],
[
"0.0866",
"0.5234",
"0.486",
"0.5244",
"0.486",
"0.539",
"0.0866",
"0.5386"
],
[
"0.513",
"0.529",
"0.91",
"0.529",
"0.91",
"0.5425",
"0.513",
"0.5425"
],
[
"0.08826",
"0.539",
"0.3635",
"0.539",
"0.3635",
"0.553",
"0.08826",
"0.553"
],
[
"0.5186",
"0.5444",
"0.911",
"0.5444",
"0.911",
"0.558",
"0.5186",
"0.558"
],
[
"0.513",
"0.5566",
"0.91",
"0.5566",
"0.91",
"0.571",
"0.513",
"0.571"
],
[
"0.511",
"0.571",
"0.9126",
"0.5713",
"0.9126",
"0.586",
"0.511",
"0.5854"
],
[
"0.0866",
"0.586",
"0.1152",
"0.586",
"0.1152",
"0.5977",
"0.0866",
"0.5977"
],
[
"0.1283",
"0.586",
"0.2566",
"0.586",
"0.2566",
"0.5996",
"0.1283",
"0.5996"
],
[
"0.5156",
"0.586",
"0.9126",
"0.586",
"0.9126",
"0.5996",
"0.5156",
"0.5996"
],
[
"0.5156",
"0.5996",
"0.9087",
"0.5996",
"0.9087",
"0.613",
"0.5156",
"0.613"
],
[
"0.513",
"0.6123",
"0.9126",
"0.613",
"0.9126",
"0.6274",
"0.513",
"0.627"
],
[
"0.0866",
"0.6255",
"0.486",
"0.6255",
"0.486",
"0.6396",
"0.0866",
"0.6396"
],
[
"0.5146",
"0.6274",
"0.898",
"0.6274",
"0.898",
"0.6416",
"0.5146",
"0.6416"
],
[
"0.08496",
"0.638",
"0.486",
"0.6396",
"0.486",
"0.654",
"0.08496",
"0.653"
],
[
"0.511",
"0.64",
"0.9126",
"0.6416",
"0.9126",
"0.655",
"0.511",
"0.654"
],
[
"0.08826",
"0.654",
"0.486",
"0.654",
"0.486",
"0.6675",
"0.08826",
"0.6675"
],
[
"0.513",
"0.655",
"0.911",
"0.655",
"0.911",
"0.6685",
"0.513",
"0.6685"
],
[
"0.08826",
"0.6675",
"0.486",
"0.6675",
"0.486",
"0.681",
"0.08826",
"0.681"
],
[
"0.511",
"0.6675",
"0.91",
"0.6685",
"0.91",
"0.683",
"0.511",
"0.6816"
],
[
"0.08826",
"0.6816",
"0.4844",
"0.6816",
"0.4844",
"0.696",
"0.08826",
"0.696"
],
[
"0.511",
"0.683",
"0.911",
"0.683",
"0.911",
"0.6963",
"0.511",
"0.6963"
],
[
"0.08826",
"0.696",
"0.4844",
"0.696",
"0.4844",
"0.709",
"0.08826",
"0.709"
],
[
"0.513",
"0.6963",
"0.9126",
"0.6963",
"0.9126",
"0.7104",
"0.513",
"0.7104"
],
[
"0.0866",
"0.709",
"0.4844",
"0.7085",
"0.4844",
"0.7227",
"0.0866",
"0.7236"
],
[
"0.5146",
"0.711",
"0.9126",
"0.711",
"0.9126",
"0.7246",
"0.5146",
"0.7246"
],
[
"0.08826",
"0.7227",
"0.486",
"0.7217",
"0.486",
"0.7363",
"0.08826",
"0.7373"
],
[
"0.5146",
"0.7246",
"0.9126",
"0.7246",
"0.9126",
"0.738",
"0.5146",
"0.738"
],
[
"0.08826",
"0.7373",
"0.4805",
"0.7373",
"0.4805",
"0.7505",
"0.08826",
"0.7505"
],
[
"0.513",
"0.7373",
"0.614",
"0.7393",
"0.614",
"0.754",
"0.513",
"0.752"
],
[
"0.08826",
"0.7505",
"0.4844",
"0.7505",
"0.4844",
"0.7646",
"0.08826",
"0.7646"
],
[
"0.0866",
"0.763",
"0.486",
"0.7646",
"0.486",
"0.779",
"0.0866",
"0.778"
],
[
"0.513",
"0.773",
"0.5254",
"0.773",
"0.5254",
"0.781",
"0.513",
"0.781"
],
[
"0.55",
"0.771",
"0.649",
"0.771",
"0.649",
"0.782",
"0.55",
"0.782"
],
[
"0.08826",
"0.779",
"0.4844",
"0.779",
"0.4844",
"0.7935",
"0.08826",
"0.7935"
],
[
"0.08826",
"0.7925",
"0.4844",
"0.7925",
"0.4844",
"0.806",
"0.08826",
"0.806"
],
[
"0.5146",
"0.7925",
"0.911",
"0.7925",
"0.911",
"0.806",
"0.5146",
"0.806"
],
[
"0.08905",
"0.806",
"0.486",
"0.806",
"0.486",
"0.8193",
"0.08905",
"0.8193"
],
[
"0.5146",
"0.806",
"0.911",
"0.806",
"0.911",
"0.8193",
"0.5146",
"0.8193"
],
[
"0.0866",
"0.8193",
"0.4844",
"0.8193",
"0.4844",
"0.8335",
"0.0866",
"0.8335"
],
[
"0.5146",
"0.8193",
"0.911",
"0.8193",
"0.911",
"0.8335",
"0.5146",
"0.8335"
],
[
"0.08826",
"0.834",
"0.486",
"0.834",
"0.486",
"0.8477",
"0.08826",
"0.8477"
],
[
"0.5146",
"0.834",
"0.911",
"0.834",
"0.911",
"0.8477",
"0.5146",
"0.8477"
],
[
"0.08826",
"0.8477",
"0.482",
"0.8477",
"0.482",
"0.8613",
"0.08826",
"0.8613"
],
[
"0.513",
"0.8477",
"0.911",
"0.8477",
"0.911",
"0.8613",
"0.513",
"0.8613"
],
[
"0.0866",
"0.8613",
"0.4844",
"0.8613",
"0.4844",
"0.875",
"0.0866",
"0.875"
],
[
"0.5146",
"0.8623",
"0.911",
"0.8623",
"0.911",
"0.8755",
"0.5146",
"0.8755"
],
[
"0.0866",
"0.8755",
"0.4631",
"0.8755",
"0.4631",
"0.8896",
"0.0866",
"0.8896"
],
[
"0.513",
"0.8755",
"0.911",
"0.8755",
"0.911",
"0.8896",
"0.513",
"0.8896"
]
] | [
"Table 2: Node Classification - Accuracy (%) of the proposed approach in predicting node labels, when compared to baseline",
"methods.",
"Dataset",
"Method",
"Leskovec-NgReinnovation",
"Congress Votes",
"Mammography",
"Balance Scale",
"99.2",
"74.7",
"99.8",
"80.6",
"90.9",
"LINE",
"DeepWalk",
"91.1",
"69.1",
"98.0",
"80.9",
"78.5",
"Node2Vec",
"96.3",
"92.4",
"80.2",
"89.3",
"PMNE",
"76.0",
"94.5",
"77.4",
"98.4",
"78.5",
"91.1",
"MNE",
"92.4",
"75.0",
"74.3",
"82.4",
"-",
"Proposed (w/o refine)",
"76.0",
"100",
"81.3",
"90.5",
"85.1",
"99.7",
"81.5",
"92.1",
"100",
"Proposed (w/ refine)",
"100",
"Table 3: Link Prediction - AUROC score of our approach in accepting positive links and rejecting negative links in multi-",
"layered networks.",
"Method",
"Dataset",
"DeepWalk",
"LINE",
"Node2Vec",
"PMNE",
"Proposed",
"Leskovec-Ng",
"0.84",
"0.49",
"0.84",
"0.71",
"0.62",
"Reinnovation",
"0.99",
"0.78",
"0.99",
"0.78",
"0.99",
"0.99",
"0.79",
"1.0",
"Congress Votes",
"1.0",
"1.0",
"Mammography",
"1.0",
"1.0",
"0.77",
"1.0",
"0.82",
"1.0",
"1.0",
"1.0",
"1.0",
"1.0",
"Vikcers",
"Balance Scale",
"0.68",
"0.76",
"0.84",
"0.89",
"0.8",
"0.83",
"0.69",
"0.82",
"0.88",
"LAZEGA",
"0.91",
"0.93",
"0.77",
"0.89",
"0.75",
"0.944",
"方法E[zhangetal.，2018]更高的准确性。请注意，在国",
"C.ELEGANS",
"4.5",
"Multi-layered Community Detection",
"会投票数据上，公开发布的CPU版本的MNE要么没",
"在最后的实验中，我们从产生内聚社区结构的角度评估",
"有稳定收敛，要么无法生成有效的嵌人",
"了所提出的多层社区检测方法。如2节中所定义，此公",
"式可以形成由来自多个层的节点组成的社区。我们为这",
"项任务考虑的多路复用网络是未标记的，因此我们求助",
"4.4",
"Link Prediction",
"于无监督学习。为了进行评估，我们利用模块化度量。大",
"多数现有的多层网络嵌人方法要么没有明确优化模块化",
"要么限制跨层的社区分配相同。因此，我们将我们的结",
"对于每个数据集，我们在每个层中随机选择一个边缘子",
"果与流行的GenLouvain算法进行比较。社区数量是",
"集，并将其从多层网络中删除。接下来，我们使用不完整",
"个重要参数，为了进行全面比较，我们通过改变和α",
"网络提取节点嵌人，并估计每个缺失边的可能性。虽然",
"来生成使用GenLouvain生成的不同数量的社区。对于",
"随机删除的边缘被标记为正，但那些最初不存在的边缘",
"这些情况中的每一个，我们对我们方法中的嵌入执行了",
"是负的。我们通过算法接受正链接而拒绝负链接的能力",
"k-means，具有相同数量的社区。GenLouvain的性能和",
"来评估链接预测性能。似然性估计为与边对应的两个节",
"所提出的方法在图3中可视化。我们观察到GenLouvain",
"点之间的余弦相似度。在我们的实验中，我们对每个层",
"的表现对所选社区的数量高度敏感。在许多情况下，该",
"中的边缘集进行了5倍交叉验证，并将AUROC（ROC",
"算法无法发现有意义的社区。相反，我们的方法一直是",
"曲线下的面积）报告为评估指标。与节点分类任务相反",
"更好和稳定的",
"此任务在很大程度上依赖于每个层中的局部结构，因此",
"单层嵌人形成了强大的基线。令人惊的是，通过超参",
"5",
"Conclusions",
"数调整（e。g。足够的步行次数），像DeepWalk这样的",
"简单基线比[Zhanget al.,2018]中观察到的更具竞争",
"本文提出了一种从多层网络中提取节点嵌入的新方法。",
"力。如表3所示，对于9数据集中的5，DeepWalk执",
"虽然局部自适应的层间边缘构造有助于M-DeepWalk的",
"行的方法与更复杂的方法一样好，没有任何额外的层间",
"复杂依赖性建模，但细化阶段在潜在空间中产生了更多",
"信息。但是，在需要从其他层传输信息以可靠地预测链",
"的内聚结构。实证研究表明，所提出的方法在包括链接",
"路的情况下，所提出的方法提供了改进，例如在Vickers",
"预测，节点分类和多层社区检测在内的各种任务中都非",
"LAZEGA和C。ELELGANS数据集。另一个观察结果",
"常有效。除了优于现有的单层和多层嵌人方法外，我们",
"是，包括细化步骤并没有在此任务中提供明显的改进",
"还证明了所提出的方法可扩展到大量层并产生高效的节"
] | {
"arxiv_id": "1811.12156",
"language": "zh",
"page_idx": 4
} |
||
1803.11029-zh-1 | 1803.11029-zh-1.jpg | 性能。通过广泛的实验评估，我们证明了我们的方法比基于 CRFs [22,31]和多尺度 CRFs [36]的传统方法产生更准确的深度图(图。1)。此外，通过在公开可用的 NYUDepthV2 [30] 和 KITTI [8] 数据集上进行实验， 我们表明我们的方法优于大多数最先进的方法
Contributions. 总而言之，我们做出了以下贡献：(i) 我们提出了一种新颖的深度学习模型，用于从静止图像中计算深度图，该模型无缝集成了前端 CNN 和多尺度 CRF。重要的是，我们的模型可以端到端训练。与以前的作品[36,22,31] our framework does not consider as input only prediction maps but operates directly at feature-level. Furthermore, by adopting appropriate unary and pairwise potentials, our framework allows a much faster inference 不同。(ii)我们的方法受益于一种新颖的注意力机制，该机制允许鲁棒地融合来自多个尺度的特征以及集成结构化信息。(iii)我们的方法在 NYUDepthV2 [30] 数据集上展示了最先进的性能并且在 KITTI 基准 [8] 更具挑战性的户外场景中表现最好。该代码是公开可用的$^{1}$
# 2. Related work
Monocular Depth Estimation. 单目深度估计问题在过去十年中引起了广泛关注。虽然早期的方法主要基于手工制作的功能[12,16,19,28] ,但最近的作品采用了深度架构[5,22,31,26,20,36,9] 。在 [6] 中，提出了一种基于两个 CNN 的模型：第一个网络用于粗略估计深度，而第二个网络用于改进预测。在[20]中，提出了一种集成了新颖的反向 Huber 损失的残差网络。在 [2] 中，还采用了深度残差网络，但是从静止图像进行深度估计的问题已从回归转换为分类任务。最近的工作也显示了采用多任务学习策略 e.g. 联合预测深度和执行语义分割，自我运动估计或表面法线计算[5,38,31] 的好处。最近的一些论文提出了无监督或弱监督的方法来重建深度图[9,18] 。其他工作利用了深度学习架构中图形模型的灵活性来估计深度图。例如，在[31] 中，采用了分层 CRF 来细化 CNN 获得的深度预测。在[22] 中，提出了一种连续 CRF,用于从在超像素上计算的 CNN 特征生成深度图。与我们最相似的工作是[36] ,其中采用 CRF 来组合从 CNN 的多个内层得出的多尺度信息。我们的方法从类似的直觉发展而来，但进一步整合了注意力模型，该模型显着提高了估计的准确性。据我们所知，这是第一篇在单眼深度估计背景下利用注意力机制的论文
Fusing Multi-scale Information in CNNs. 最近的许多工作都显示了将多尺度信息组合用于像素级预测任务的好处，例如语义分割$\left[3\right]$,深度估计$\left[36\right]$或轮廓检测[32]。例如，在[3]中采用了扩张卷积。[1] 中考虑了具有不同分辨率输入的多流体系结构，而[25]提出了跳过连接以融合来自不同层的特征图。在[32] 中，利用深度监督来融合来自多个内层的信息。CRF 已被考虑用于在[36]中集成多尺度信息。在[4,33]中，注意
模型用于在语义分割和对象轮廓检测的上下文中组合多尺度特征。我们在本文中提出的方法是完全不同的， 因为我们采用了在 CRF-CNN 框架内共同学习的结构化注意力模型
3.
Estimating Depth Maps with Structured Attention Guided Conditional Neural Fields
在本节中，我们描述了从静止图像估计深度图的方法。我们首先概述了我们的方法，然后介绍了建议的具有结构化注意力的 CRF 模型。我们结束本节，提供有关我们的实现的一些细节
$3. 1. \textbf{ Problem Formulation and Overview}$
如引言中所述，可以将从单个 RGB 图像预测深度图的问题视为监督学习问题。将 RGB 图像的空间表示为$\mathcal{I}$,将实值深度图的域表示为$\mathcal{D}$,给定训练集$\mathcal{T} = \{ ( \boldsymbol{I}_i, \boldsymbol{D}_i) \} _{i= 1}^M$ , $\boldsymbol{I}_i\in \mathcal{I}$和$D_i\in\mathcal{D}$ ,我们有兴趣学习非线性映射$\Phi:\mathcal{I}\to\mathcal{D}$
与以前的作品[22,36]类似，我们建议通过构建一个由两个主要构建块组成的深层体系结构来学习映射$\Phi$ :前端 CNN 和 CRF 模型。所提出的 CRF 模型的主要目的是结合从前端 CNN 的内层得出的多尺度信息。与以前的研究[22,36]不同，我们的 CRF 模型不仅仅是为了改进 CNN 的最终预测图，也不需要输入相同大小的多个分数图。在本文中，我们认为可以通过更灵活的模型获得更好的估计，该模型接受一组直接从前端中间层导出的$S$多尺度特征图$X=\{X_s\}_{s=1}^S$作为输人。为了便于建模，所有多尺度特征图都通过上采样或下采样操作调整为相同的分辨率。这里$X_s=\{x_s^i\}_{i=1}^N$ , $x_s^i\in \mathbb{R} ^N$ ,表示一组特征向量
所提出的多尺度 CRF 模型的设计背后的主要思想是通过利用最后一层$X_S$的特征和从中间尺度$s=$ $1,\ldots,S-1$导出的一组辅助特征表示来估计与 RGB 图像$I$相关联的深度图$D$ 。为此，我们建议学习一组 潜 在 特 征 图 $Y_{s}= \{ \boldsymbol{y}_{s}^{i}\} _{i= 1}^{N}$ , $s= 1, \ldots , S$ ,并 通 过 引 入 适 当 的 注 意 力 模 型 $A= \{ \boldsymbol{A}_{s}\} _{s= 1}^{S- 1}$ , 由 二 进 制 变 量$\boldsymbol{A}_{s}= \{ a_{s}^{i}\} _{i= 1}^{N}$ , <NEXT\_LINE。直观地，注意变量$a_s^i$ 调节允许在每个中间比例$s$和像素$i$的最终比例$S$之间流动的信息。换句话说，通过学习注意力图，我们自动发现从内部 CNN 表示得出的哪些信息与最终深度估计相关。此外，为了获得准确的注意力图$A_s$ ,我们建议学习结构化注意力模型 i.e. , 我们对估计变量$a_s^i$ 施加结构约束，以强制与相邻像素相对应的那些相关。重要的是，建议的 CRF 联合推断隐藏的特征和注意图
图 2 示意性地描述了建议的框架和我们的 CRF 模型。对更精细的学习表示与对应于每个中间层的特征之间的关系进行建模的想法受到最近的 DenseNet 架构 [14] 的启发。正如我们的实验 (第 4 节)所证明的那样，相对于[36] 中提出的级联模型，此策略可提高
$^{1}$https://github.com/danxuhk/StructuredAttentionDepthEstimatio胜能。 | [
[
"0.0825",
"0.09344",
"0.468",
"0.09344",
"0.468",
"0.1067",
"0.0825",
"0.1067"
],
[
"0.5034",
"0.09344",
"0.8896",
"0.09344",
"0.8896",
"0.1067",
"0.5034",
"0.1067"
],
[
"0.0825",
"0.108",
"0.468",
"0.108",
"0.468",
"0.1218",
"0.0825",
"0.1218"
],
[
"0.5034",
"0.108",
"0.888",
"0.108",
"0.888",
"0.1218",
"0.5034",
"0.1218"
],
[
"0.0825",
"0.125",
"0.4656",
"0.125",
"0.4656",
"0.1351",
"0.0825",
"0.1351"
],
[
"0.5063",
"0.125",
"0.8896",
"0.125",
"0.8896",
"0.1351",
"0.5063",
"0.1351"
],
[
"0.0784",
"0.1351",
"0.4714",
"0.1364",
"0.4714",
"0.154",
"0.0784",
"0.1528"
],
[
"0.5034",
"0.137",
"0.603",
"0.1395",
"0.603",
"0.154",
"0.5034",
"0.1515"
],
[
"0.0809",
"0.1515",
"0.407",
"0.1515",
"0.407",
"0.1654",
"0.0809",
"0.1654"
],
[
"0.08496",
"0.1704",
"0.468",
"0.1704",
"0.468",
"0.1843",
"0.08496",
"0.1843"
],
[
"0.5034",
"0.1755",
"0.5215",
"0.1755",
"0.5215",
"0.1849",
"0.5034",
"0.1849"
],
[
"0.5186",
"0.178",
"0.5293",
"0.1768",
"0.5312",
"0.1818",
"0.5195",
"0.1831"
],
[
"0.5605",
"0.1736",
"0.886",
"0.1736",
"0.886",
"0.1875",
"0.5605",
"0.1875"
],
[
"0.08417",
"0.1875",
"0.4656",
"0.1875",
"0.4656",
"0.1976",
"0.08417",
"0.1976"
],
[
"0.5337",
"0.1913",
"0.888",
"0.1913",
"0.888",
"0.2052",
"0.5337",
"0.2052"
],
[
"0.08417",
"0.2008",
"0.4673",
"0.2008",
"0.4673",
"0.214",
"0.08417",
"0.214"
],
[
"0.5312",
"0.2058",
"0.614",
"0.2084",
"0.614",
"0.2229",
"0.5312",
"0.2203"
],
[
"0.0809",
"0.2153",
"0.4673",
"0.2153",
"0.4673",
"0.2291",
"0.0809",
"0.2291"
],
[
"0.0801",
"0.2299",
"0.468",
"0.2291",
"0.468",
"0.2437",
"0.0801",
"0.2443"
],
[
"0.524",
"0.2349",
"0.888",
"0.2349",
"0.888",
"0.2488",
"0.524",
"0.2488"
],
[
"0.0809",
"0.2456",
"0.468",
"0.2456",
"0.468",
"0.2588",
"0.0809",
"0.2588"
],
[
"0.5034",
"0.2507",
"0.8896",
"0.2507",
"0.8896",
"0.2644",
"0.5034",
"0.2644"
],
[
"0.0801",
"0.2588",
"0.468",
"0.2615",
"0.468",
"0.276",
"0.0801",
"0.2734"
],
[
"0.505",
"0.2651",
"0.888",
"0.2651",
"0.888",
"0.279",
"0.505",
"0.279"
],
[
"0.0825",
"0.276",
"0.468",
"0.276",
"0.468",
"0.289",
"0.0825",
"0.289"
],
[
"0.5024",
"0.2798",
"0.683",
"0.2798",
"0.683",
"0.2935",
"0.5024",
"0.2935"
],
[
"0.0825",
"0.2917",
"0.4558",
"0.2917",
"0.4558",
"0.305",
"0.0825",
"0.305"
],
[
"0.0825",
"0.308",
"0.4673",
"0.308",
"0.4673",
"0.318",
"0.0825",
"0.318"
],
[
"0.5034",
"0.3125",
"0.819",
"0.3125",
"0.819",
"0.3257",
"0.5034",
"0.3257"
],
[
"0.0825",
"0.3208",
"0.4673",
"0.3208",
"0.4673",
"0.334",
"0.0825",
"0.334"
],
[
"0.0809",
"0.336",
"0.4673",
"0.336",
"0.4673",
"0.3499",
"0.0809",
"0.3499"
],
[
"0.524",
"0.3372",
"0.8896",
"0.3372",
"0.8896",
"0.351",
"0.524",
"0.351"
],
[
"0.0809",
"0.3516",
"0.468",
"0.3516",
"0.468",
"0.3655",
"0.0809",
"0.3655"
],
[
"0.5034",
"0.3516",
"0.8906",
"0.3516",
"0.8906",
"0.3655",
"0.5034",
"0.3655"
],
[
"0.0809",
"0.3674",
"0.2793",
"0.3674",
"0.2793",
"0.3806",
"0.0809",
"0.3806"
],
[
"0.505",
"0.3674",
"0.8906",
"0.3674",
"0.8906",
"0.3806",
"0.505",
"0.3806"
],
[
"0.5034",
"0.3818",
"0.8896",
"0.3818",
"0.8896",
"0.3958",
"0.5034",
"0.3958"
],
[
"0.0801",
"0.3965",
"0.2222",
"0.3965",
"0.2222",
"0.4104",
"0.0801",
"0.4104"
],
[
"0.5034",
"0.3977",
"0.699",
"0.3977",
"0.699",
"0.411",
"0.5034",
"0.411"
],
[
"0.08496",
"0.4224",
"0.304",
"0.4224",
"0.304",
"0.4363",
"0.08496",
"0.4363"
],
[
"0.3203",
"0.4224",
"0.468",
"0.4224",
"0.468",
"0.4363",
"0.3203",
"0.4363"
],
[
"0.527",
"0.4167",
"0.8896",
"0.4167",
"0.8896",
"0.4268",
"0.527",
"0.4268"
],
[
"0.5034",
"0.43",
"0.892",
"0.43",
"0.892",
"0.443",
"0.5034",
"0.443"
],
[
"0.0801",
"0.4363",
"0.468",
"0.4375",
"0.468",
"0.452",
"0.0801",
"0.4507"
],
[
"0.5034",
"0.4443",
"0.8906",
"0.4443",
"0.8906",
"0.4583",
"0.5034",
"0.4583"
],
[
"0.0825",
"0.4526",
"0.4656",
"0.4526",
"0.4656",
"0.4666",
"0.0825",
"0.4666"
],
[
"0.5034",
"0.4602",
"0.8906",
"0.4602",
"0.8906",
"0.4734",
"0.5034",
"0.4734"
],
[
"0.0825",
"0.4685",
"0.4673",
"0.4685",
"0.4673",
"0.4817",
"0.0825",
"0.4817"
],
[
"0.505",
"0.476",
"0.8906",
"0.476",
"0.8906",
"0.4893",
"0.505",
"0.4893"
],
[
"0.09235",
"0.483",
"0.468",
"0.483",
"0.468",
"0.4968",
"0.09235",
"0.4968"
],
[
"0.5034",
"0.4905",
"0.8906",
"0.4905",
"0.8906",
"0.504",
"0.5034",
"0.504"
],
[
"0.0825",
"0.5",
"0.4656",
"0.5",
"0.4656",
"0.5103",
"0.0825",
"0.5103"
],
[
"0.5034",
"0.506",
"0.8896",
"0.506",
"0.8896",
"0.5195",
"0.5034",
"0.5195"
],
[
"0.08905",
"0.513",
"0.4673",
"0.513",
"0.4673",
"0.5264",
"0.08905",
"0.5264"
],
[
"0.5024",
"0.521",
"0.8896",
"0.521",
"0.8896",
"0.534",
"0.5024",
"0.534"
],
[
"0.0825",
"0.528",
"0.468",
"0.528",
"0.468",
"0.5415",
"0.0825",
"0.5415"
],
[
"0.5034",
"0.534",
"0.8906",
"0.536",
"0.8906",
"0.55",
"0.5034",
"0.5474"
],
[
"0.0809",
"0.5435",
"0.468",
"0.5435",
"0.468",
"0.5566",
"0.0809",
"0.5566"
],
[
"0.505",
"0.5503",
"0.8906",
"0.5503",
"0.8906",
"0.5645",
"0.505",
"0.5645"
],
[
"0.0825",
"0.5596",
"0.468",
"0.5596",
"0.468",
"0.5728",
"0.0825",
"0.5728"
],
[
"0.505",
"0.5664",
"0.888",
"0.5664",
"0.888",
"0.58",
"0.505",
"0.58"
],
[
"0.0809",
"0.5737",
"0.4656",
"0.5737",
"0.4656",
"0.587",
"0.0809",
"0.587"
],
[
"0.505",
"0.58",
"0.7344",
"0.58",
"0.7344",
"0.5947",
"0.505",
"0.5947"
],
[
"0.0809",
"0.5884",
"0.4673",
"0.5884",
"0.4673",
"0.6016",
"0.0809",
"0.6016"
],
[
"0.523",
"0.5977",
"0.8906",
"0.5977",
"0.8906",
"0.6113",
"0.523",
"0.6113"
],
[
"0.0825",
"0.604",
"0.4673",
"0.604",
"0.4673",
"0.617",
"0.0825",
"0.617"
],
[
"0.0801",
"0.617",
"0.4673",
"0.6187",
"0.4673",
"0.632",
"0.0801",
"0.631"
],
[
"0.5034",
"0.6123",
"0.8857",
"0.6123",
"0.8857",
"0.6255",
"0.5034",
"0.6255"
],
[
"0.5063",
"0.6274",
"0.8906",
"0.6274",
"0.8906",
"0.6416",
"0.5063",
"0.6416"
],
[
"0.0825",
"0.634",
"0.468",
"0.634",
"0.468",
"0.648",
"0.0825",
"0.648"
],
[
"0.0809",
"0.6484",
"0.468",
"0.6484",
"0.468",
"0.662",
"0.0809",
"0.662"
],
[
"0.505",
"0.643",
"0.876",
"0.643",
"0.876",
"0.657",
"0.505",
"0.657"
],
[
"0.0825",
"0.664",
"0.468",
"0.664",
"0.468",
"0.678",
"0.0825",
"0.678"
],
[
"0.505",
"0.659",
"0.8906",
"0.659",
"0.8906",
"0.6724",
"0.505",
"0.6724"
],
[
"0.0825",
"0.6787",
"0.4673",
"0.6787",
"0.4673",
"0.6924",
"0.0825",
"0.6924"
],
[
"0.5034",
"0.674",
"0.8906",
"0.674",
"0.8906",
"0.6875",
"0.5034",
"0.6875"
],
[
"0.5063",
"0.69",
"0.6113",
"0.69",
"0.6113",
"0.7026",
"0.5063",
"0.7026"
],
[
"0.601",
"0.6875",
"0.8896",
"0.688",
"0.8896",
"0.7026",
"0.601",
"0.702"
],
[
"0.505",
"0.7026",
"0.888",
"0.7026",
"0.888",
"0.7173",
"0.505",
"0.7173"
],
[
"0.0809",
"0.6924",
"0.4656",
"0.6924",
"0.4656",
"0.7056",
"0.0809",
"0.7056"
],
[
"0.0809",
"0.709",
"0.468",
"0.709",
"0.468",
"0.7227",
"0.0809",
"0.7227"
],
[
"0.0809",
"0.7246",
"0.468",
"0.7246",
"0.468",
"0.738",
"0.0809",
"0.738"
],
[
"0.5063",
"0.72",
"0.886",
"0.72",
"0.886",
"0.731",
"0.5063",
"0.731"
],
[
"0.505",
"0.7344",
"0.8896",
"0.7344",
"0.8896",
"0.7476",
"0.505",
"0.7476"
],
[
"0.0825",
"0.7407",
"0.2573",
"0.7407",
"0.2573",
"0.7505",
"0.0825",
"0.7505"
],
[
"0.5063",
"0.7505",
"0.886",
"0.7505",
"0.886",
"0.761",
"0.5063",
"0.761"
],
[
"0.08826",
"0.7583",
"0.3896",
"0.7583",
"0.3896",
"0.768",
"0.08826",
"0.768"
],
[
"0.3987",
"0.7583",
"0.4673",
"0.7583",
"0.4673",
"0.768",
"0.3987",
"0.768"
],
[
"0.505",
"0.7646",
"0.8896",
"0.7646",
"0.8896",
"0.778",
"0.505",
"0.778"
],
[
"0.0809",
"0.7715",
"0.468",
"0.7715",
"0.468",
"0.785",
"0.0809",
"0.785"
],
[
"0.505",
"0.779",
"0.8906",
"0.779",
"0.8906",
"0.7925",
"0.505",
"0.7925"
],
[
"0.08417",
"0.7886",
"0.4673",
"0.7886",
"0.4673",
"0.799",
"0.08417",
"0.799"
],
[
"0.505",
"0.7954",
"0.888",
"0.7954",
"0.888",
"0.806",
"0.505",
"0.806"
],
[
"0.0825",
"0.802",
"0.4673",
"0.802",
"0.4673",
"0.815",
"0.0825",
"0.815"
],
[
"0.524",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.524",
"0.8247"
],
[
"0.08417",
"0.8174",
"0.4656",
"0.8174",
"0.4656",
"0.8306",
"0.08417",
"0.8306"
],
[
"0.5024",
"0.8247",
"0.8906",
"0.8257",
"0.8906",
"0.8403",
"0.5024",
"0.839"
],
[
"0.0825",
"0.834",
"0.4656",
"0.834",
"0.4656",
"0.8447",
"0.0825",
"0.8447"
],
[
"0.0809",
"0.8467",
"0.4673",
"0.845",
"0.4673",
"0.86",
"0.0809",
"0.8613"
],
[
"0.5063",
"0.8433",
"0.8896",
"0.8433",
"0.8896",
"0.854",
"0.5063",
"0.854"
],
[
"0.505",
"0.857",
"0.8896",
"0.857",
"0.8896",
"0.8706",
"0.505",
"0.8706"
],
[
"0.0809",
"0.863",
"0.4673",
"0.863",
"0.4673",
"0.877",
"0.0809",
"0.877"
],
[
"0.5024",
"0.8706",
"0.8906",
"0.872",
"0.8906",
"0.886",
"0.5024",
"0.8853"
],
[
"0.0972",
"0.887",
"0.544",
"0.886",
"0.544",
"0.901",
"0.0972",
"0.9014"
]
] | [
"性能。通过广泛的实验评估，我们证明了我们的方法",
"模型用于在语义分割和对象轮廓检测的上下文中组合",
"比基于CRFs[22，31]和多尺度CRFs[36]的传统方法",
"多尺度特征。我们在本文中提出的方法是完全不同的",
"产生更准确的深度图（图。1）。此外，通过在公开可用",
"因为我们采用了在CRF-CNN框架内共同学习的结构",
"的NYUDepthV2【30]和KITTI[8]数据集上进行实验，",
"化注意力模型",
"我们表明我们的方法优于大多数最先进的方法",
"Contributions．总而言之，我们做出了以下贡献：（i）",
"3.",
"",
"Estimating Depth Maps with Struc",
"我们提出了一种新颖的深度学习模型，用于从静正图",
"tured Attention Guided Conditional Neu-",
"像中计算深度图，该模型无缝集成了前端CNN和多尺",
"ral Fields",
"度CRF。重要的是，我们的模型可以端到端训练。与以",
"前的作品[36,22,31]ourframeworkdoesnotconsider",
"在本节中，我们描述了从静止图像估计深度图的方",
"as input only prediction maps but operates directly at",
"法。我们首先概述了我们的方法，然后介绍了建议的具",
"feature-level. Furthermore, by adopting appropriate",
"有结构化注意力的CRF模型。我们结束本节，提供有",
"unary and pairwise potentials, our framework allows a",
"关我们的实现的一些细节",
"muchfasterinference不同。（i）我们的方法受益于",
"种新颖的注意力机制，该机制充许鲁棒地融合来自多",
"3.1. Problem Formulation and Overview",
"个尺度的特征以及集成结构化信息。（ii）我们的方法",
"在NYUDepthV2[30]数据集上展示了最先进的性能，",
"如引言中所述，可以将从单个RGB图像预测深度",
"并且在KITTI基准[8]更具挑战性的户外场景中表现",
"图的问题视为监督学习问题。将RGB图像的空间表",
"最好。该代码是公开可用的",
"示为工，将实值深度图的域表示为D，给定训练集",
"T=(I,D)IM，II和D,ED，我们有兴趣",
"2. Related work",
"学习非线性映射Φ:I→D",
"Monocular Depth Estimation.",
"单目深度估计问题在",
"与以前的作品22，36类似，我们建议通过构建一个",
"由两个主要构建块组成的深层体系结构来学习映射Φ",
"过去十年中引起了广泛关注。虽然早期的方法主要基",
"：前端CNN和 CRF模型。所提出的CRF模型的主",
"于手工制作的功能[12，16，19,28]，但最近的作品采用",
"要目的是结合从前端CNN的内层得出的多尺度信息。",
"了深度架构[5,22，31，26，20，36，9]。在[6]中，提出了",
"与以前的研究22，36]不同，我们的CRF模型不仅仅",
"种基于两个CNN的模型：第一个网络用于粗略估计",
"是为了改进CNN的最终预测图，也不需要输入相同大",
"深度，而第二个网络用于改进预测。在20中，提出了",
"小的多个分数图。在本文中，我们认为可以通过更灵活",
"一种集成了新颖的反向Huber损失的残差网络。在[2]",
"的模型获得更好的估计，该模型接受一组直接从前端",
"中，还采用了深度残差网络，但是从静止图像进行深",
"中间层导出的S多尺度特征图X=X)s=1作为输",
"度估计的问题已从回归转换为分类任务。最近的工作",
"人。为了便于建模，所有多尺度特征图都通过上采样或",
"也显示了采用多任务学习策略e.g.联合预测深度和执",
"下采样操作调整为相同的分辨率。这里X。=（a=1",
"行语义分割，自我运动估计或表面法线计算[5，38，31",
"，aERN，表示一组特征向量",
"的好处。最近的一些论文提出了无监督或弱监督的方",
"所提出的多尺度CRF模型的设计背后的主要思想",
"法来重建深度图[9，18]。其他工作利用了深度学习架",
"构中图形模型的灵活性来估计深度图。例如，在「31]",
"是通过利用最后一层Xs的特征和从中间尺度s=",
"1....,S-1导出的一组辅助特征表示来估计与RGB",
"中，采用了分层CRF来细化CNN获得的深度预测。",
"在[22]中，提出了一种连续CRF，用于从在超像素上",
"图像I相关联的深度图D。为此，我们建议学习",
"计算的CNN特征生成深度图。与我们最相似的工作是",
"组潜在特征图Ys=y名l，s=1,...,S，并通过",
"[361，其中采用CRF来组合从CNN的多个内层得出",
"引人适当的注意力模型A=（A。}s-，由二进制变量",
"As=[a-l ",
"，<NEXT_LINE。直观地，注意变量α",
"调节允许在每个中间比例s和像素i的最终比例S之",
"的多尺度信息。我们的方法从类似的直觉发展而来，但",
"进一步整合了注意力模型，该模型显着提高了估计的",
"准确性。据我们所知，这是第一篇在单眼深度估计背景",
"间流动的信息。换句话说，通过学习注意力图，我们自",
"动发现从内部CNN表示得出的哪些信息与最终深度",
"下利用注意力机制的论文",
"估计相关。此外，为了获得准确的注意力图A。，我们",
"FusingMulti-scaleInformationinCNNs.",
"最近的许",
"建议学习结构化注意力模型i.e.，我们对估计变量a",
"多工作都显示了将多尺度信息组合用于像素级预测任",
"施加结构约束，以强制与相邻像素相对应的那些相关。",
"务的好处，例如语义分割3：深度估计36或轮廓检",
"重要的是，建议的CRF联合推断隐藏的特征和注意图",
"测「32]。例如，在3]中采用了扩张卷积。[1]中考虑了",
"图2示意性地描述了建议的框架和我们的CRF模",
"具有不同分辨率输入的多流体系结构，而【25】提出了",
"型。对更精细的学习表示与对应于每个中间层的特征",
"跳过连接以融合来自不同层的特征图。在32中，利",
"用深度监督来融合来自多个内层的信息。CRF已被考",
"之间的关系进行建模的想法受到最近的DenseNet架",
"构【14]的启发。正如我们的实验（第4节）所证明的",
"虑用于在[36]】中集成多尺度信息。在【4，33]中，注意",
"那样，相对于[36]中提出的级联模型，此策略可提高",
"1https://github.com/danxuhk/StructuredAttentionDepthEstimatio性能。"
] | {
"arxiv_id": "1803.11029",
"language": "zh",
"page_idx": 1
} |
||
1803.11029-zh-3 | 1803.11029-zh-3.jpg | Table 1. NYU Depth V2 dataset: comparison with state of the art. In bold we indicate the best method adopting on the
original set, while in italics we indicate the best method using the extended set.
<table>
<tbody>
<tr>
<th rowspan="2">Method</th>
<th rowspan="2">$\operatorname{Extra}$ Training gData</th>
<th>Error</th>
<th colspan="2">Error (lower is better)</th>
<th colspan="3">Accuracy (higher is better)</th>
</tr>
<tr>
<th>rel</th>
<th>$\overline{\log10}$</th>
<th>rms</th>
<th>$\overline{\delta<1.25}$</th>
<th>$\overline{\delta<1.25^{2}}$</th>
<th>$\overline{\delta<}$ $1.25^{\cdot}$</th>
</tr>
<tr>
<td>[29] Saxena et al.</td>
<td>No$\left ( 795\right )$</td>
<td>0.349</td>
<td> </td>
<td>1.214</td>
<td>0.447</td>
<td>0.745</td>
<td>0.897</td>
</tr>
<tr>
<td>Karsch et al. $16^{\prime}$</td>
<td>No ( 795) </td>
<td>0.35</td>
<td>0.131</td>
<td>1.20</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Liu et al. $\lceil24\rceil^{\prime}$</td>
<td>No(795)</td>
<td>0.335</td>
<td>0.127</td>
<td>1.06</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Ladicky et al. [19]</td>
<td>No(795)</td>
<td> </td>
<td> </td>
<td> </td>
<td>0.542</td>
<td>0.829</td>
<td>0.941</td>
</tr>
<tr>
<td>39 Zhuo et al.</td>
<td>No ( 795) </td>
<td>0.305</td>
<td>0.122</td>
<td>1.04</td>
<td>0.525</td>
<td>0.838</td>
<td>0.962</td>
</tr>
<tr>
<td>Wang et al. 31</td>
<td>No(795)</td>
<td>0.220</td>
<td>0.094</td>
<td>0.745</td>
<td>0.605</td>
<td>0.890</td>
<td>0.970</td>
</tr>
<tr>
<td>Liu et al. $\lceil23\rceil^{\prime}$</td>
<td>No(795)</td>
<td>0.213</td>
<td>0.087</td>
<td>0.759</td>
<td>0.650</td>
<td>0.906</td>
<td>0.976</td>
</tr>
<tr>
<td>Roi and Todorovic $\lceil26\rceil$ 1</td>
<td>No (795)</td>
<td>0.187</td>
<td>0.078</td>
<td>0.744</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>Xu et al. $\lceil36^{\prime}$</td>
<td>No(795)</td>
<td>0.139</td>
<td>0.063</td>
<td>0.609</td>
<td>0.793</td>
<td>0.948</td>
<td>0.984</td>
</tr>
<tr>
<td>Ours</td>
<td>No (795)</td>
<td>0.125</td>
<td>0.057</td>
<td>0.593</td>
<td>0.806</td>
<td>0.952</td>
<td>0.986</td>
</tr>
<tr>
<td>Eigen et al. $\lceil6\rceil$</td>
<td>Yes ${\mathrm{s}}(120$K)</td>
<td>0.215</td>
<td>-</td>
<td>0.907</td>
<td>0.611</td>
<td>0.887</td>
<td>0.971</td>
</tr>
<tr>
<td>Eigen and Fergus 5</td>
<td>Yes (120K</td>
<td>0.158</td>
<td>-</td>
<td>0.641</td>
<td>0.769</td>
<td>0.950</td>
<td>0.988</td>
</tr>
<tr>
<td>Laina et al. [20]</td>
<td>(12K Yes</td>
<td>0.129</td>
<td>0.056</td>
<td>0.583</td>
<td>0.801</td>
<td>0.950</td>
<td>0.986</td>
</tr>
<tr>
<td>Li et al. [21]</td>
<td>Yes $\textbf{s}( 24$K)</td>
<td>0.139</td>
<td>0.058</td>
<td>0.505</td>
<td>0.820</td>
<td>0.960</td>
<td>0.989</td>
</tr>
<tr>
<td>Xu et al. $[36^{\prime}$</td>
<td>Yes (12K</td>
<td>0.121</td>
<td>0.052</td>
<td>0.586</td>
<td>0.811</td>
<td>0.954</td>
<td>0.987</td>
</tr>
</tbody>
</table>
的实现中，采用了建议的 CRF 来细化从语义层 res5c 派生的最后一个比例特征图，该语义层从从 res3c 和res4f 派生的其他比例特征图接收消息。$res3c$ $、 res4f$ 和$res5c$是不同卷积块的最后一层。在每个卷积块中， 每一层都输出具有相同通道数的特征图。在消息传递之前，首先使用反卷积运算对所有特征图进行上采样， 使其具有相同的大小，i.e. $\operatorname{E\text{原始输入图像的 }1/4}$分辨率，并且所有通道的通道数都设置为 256。$K_s$和$\beta_s$ 的内核大小都设置为3，步幅为 1,填充为 1,以具有本地接受字段并加快计算
建议的多尺度 CRF 模块输出细化的特征图。为了获得最终的预测，我们使用去卷积运算将特征图上采样到原始分辨率作为输入图像。每次我们将特征图上采样 2 倍，同时将特征通道的数量减少一半。
End-to-end optimization. 如上所述，所提出的模型可以端到端训练，i.e. 前端编码器$\Theta_e$的参数，与结构化注意力引导 CRF $\Theta_c$相关的参数以及解码器$\Theta_d$的参数可以联合优化。给定训练数据集$\mathcal{T}$ ,在先前的工作 [36] 之后，我们使用平方损失函数 i.e. 进行优化：
$$\mathcal{L}_{F}((\mathcal{I},\mathcal{D};\boldsymbol{\Theta}_{e},\boldsymbol{\Theta}_{c},\boldsymbol{\Theta}_{d})=\sum_{i=1}^{M}\|F(\boldsymbol{I}_{i}^{l};\boldsymbol{\Theta}_{e},\boldsymbol{\Theta}_{c},\boldsymbol{\Theta}_{d})-\boldsymbol{D}_{i}^{l}\|_{2}^{2}$$
整个网络通过具有标准随机梯度下降的反向传播进行
联合优化。
4. Experiments
3.2.3 Implementation with Neural Networks
为了实现整个网络的端到端优化，我们在神经网络中实现了所提出的多尺度模型。目标是根据 3.2.2 节中描述的推导，对注意力变量和多尺度特征图执行均值场更新
要执行注意力模型$A$的均值场更新，我们遵循( ?? )。实际上，每个注意力图$a_\mathrm{s}$的更新可以通过以下几个步骤实现：(i)从两个关联的特征图$\bar{y}_s$和$\bar{y}_S$ 执行消息传递 ($\bar{y}_s$和$\bar{y}_S$分别用相应的特征观察$x_s$ 和$x_S$初始化)。消息传递是通过卷积运算执行的，如$\hat{a} _s\leftarrow y_s\odot ( K_s\otimes \bar{y} _S)$ ,其中$K_s$是对应于$s$比例的卷积内核，符号$\otimes$和$\odot$分别表示卷积和元素方面的乘积运算；(ii)使用$\tilde{a}_s\leftarrow\beta_s\otimes\bar{a}_s$在注意图上执行消息传递
当注意图更新时，我们将其用作指导，以更新最后一个比例要素图$y_S$ 。$y_S$的平均场更新可以根据(?? )如下进行：(i) 通过$\hat{y}_s\leftarrow K_s\otimes\bar{y}_s$执行从第$s$标度到第$S$标度的消息传递；(ii)乘以注意力模型，并将一元项$x_S$与$\bar{y}_S\leftarrow x_S\oplus\sum_s(\bar{a}_s\odot\hat{\boldsymbol{y}}_s)$相加。根据(?? ), 可以类似地执行与中间尺度相对应的潜在特征的平均场更新的计算。。在我们的实现中，为了减少计算开销，我们不对中间尺度执行均值场更新。迭代更新注意力图和最后的比例特征图
我们想指出的是，由于(??),(2)和(4)中的势函数的定义，我们方法中的平均场更新的计算比在[36] 中考虑高斯函数用于成对电势的效率要高得多。实际上，高斯卷积在前向和后向传递中都涉及更高的计算开销。我们在 4 节中进一步讨论了这一方面。
我们证明了所提出的方法在两个公开可用的数据集上进行实验的有效性：NYUDepthV2$\underline{\mathrm{E}}[30]$和 KITTIE [8]数据集。以下小节描述了我们的实验设置和评估结果
3.3. Network Structure and Optimization
# 4.1. Experimental Setup
Network Structure and Implementation.
单目深度
估计的总体框架由 CNN 架构和建议的 CRF 模型(。$\color{red}2$ )。CNN 架构由两个主要组件组成，i.e. 是全卷积编码器和全卷积解码器。编码器自然支持任何网络结构。在这项工作中，我们特别采用 ResNet-50 [11] 。在我们
Datasets.
NYU Depth V2 数据集[30]具有使用
MicrosoftKinect 收集的 120K 对 RGB 和深度图。图像分辨率为 6$40\times480$像素。数据集分为训练(249 个 | [
[
"0.0801",
"0.0859",
"0.8906",
"0.0884",
"0.8906",
"0.1029",
"0.0801",
"0.101"
],
[
"0.0801",
"0.1016",
"0.609",
"0.1016",
"0.609",
"0.11615",
"0.0801",
"0.11615"
],
[
"0.4558",
"0.11743",
"0.596",
"0.1187",
"0.596",
"0.1332",
"0.4558",
"0.132"
],
[
"0.6323",
"0.11743",
"0.7974",
"0.1193",
"0.7974",
"0.1332",
"0.6323",
"0.1307"
],
[
"0.1487",
"0.1237",
"0.1978",
"0.1256",
"0.1969",
"0.1383",
"0.1478",
"0.1364"
],
[
"0.304",
"0.125",
"0.4395",
"0.125",
"0.4395",
"0.1383",
"0.304",
"0.1383"
],
[
"0.464",
"0.132",
"0.486",
"0.132",
"0.486",
"0.1445",
"0.464",
"0.1445"
],
[
"0.509",
"0.1332",
"0.544",
"0.1332",
"0.544",
"0.1445",
"0.509",
"0.1445"
],
[
"0.568",
"0.1332",
"0.5923",
"0.1332",
"0.5923",
"0.1459",
"0.568",
"0.1459"
],
[
"0.614",
"0.1332",
"0.665",
"0.1332",
"0.665",
"0.1445",
"0.614",
"0.1445"
],
[
"0.6846",
"0.1332",
"0.7417",
"0.1307",
"0.7417",
"0.1433",
"0.6846",
"0.1459"
],
[
"0.7573",
"0.132",
"0.8228",
"0.13",
"0.8228",
"0.1445",
"0.7573",
"0.1465"
],
[
"0.1478",
"0.1515",
"0.2566",
"0.1515",
"0.2566",
"0.1654",
"0.1478",
"0.1654"
],
[
"0.4534",
"0.1509",
"0.4958",
"0.1509",
"0.4958",
"0.1654",
"0.4534",
"0.1654"
],
[
"0.562",
"0.1515",
"0.597",
"0.1515",
"0.597",
"0.1642",
"0.562",
"0.1642"
],
[
"0.6953",
"0.1515",
"0.734",
"0.1515",
"0.734",
"0.1654",
"0.6953",
"0.1654"
],
[
"0.7715",
"0.1515",
"0.8057",
"0.1515",
"0.8057",
"0.1642",
"0.7715",
"0.1642"
],
[
"0.6216",
"0.1528",
"0.6567",
"0.1528",
"0.6567",
"0.1642",
"0.6216",
"0.1642"
],
[
"0.3464",
"0.1528",
"0.3987",
"0.1528",
"0.3987",
"0.1642",
"0.3464",
"0.1642"
],
[
"0.1487",
"0.1635",
"0.2566",
"0.1635",
"0.2566",
"0.1768",
"0.1487",
"0.1768"
],
[
"0.3447",
"0.1622",
"0.4004",
"0.1642",
"0.3987",
"0.1787",
"0.343",
"0.1768"
],
[
"0.4592",
"0.1642",
"0.4902",
"0.1642",
"0.4902",
"0.1768",
"0.4592",
"0.1768"
],
[
"0.509",
"0.1642",
"0.5435",
"0.1642",
"0.5435",
"0.1768",
"0.509",
"0.1768"
],
[
"0.564",
"0.1642",
"0.5947",
"0.1642",
"0.5947",
"0.1768",
"0.564",
"0.1768"
],
[
"0.6396",
"0.1674",
"0.6465",
"0.1724",
"0.6406",
"0.178",
"0.634",
"0.1724"
],
[
"0.3447",
"0.1749",
"0.4004",
"0.1768",
"0.3987",
"0.1913",
"0.343",
"0.1895"
],
[
"0.709",
"0.1692",
"0.716",
"0.1692",
"0.716",
"0.1749",
"0.709",
"0.1749"
],
[
"0.455",
"0.1755",
"0.4958",
"0.1755",
"0.4958",
"0.1901",
"0.455",
"0.1901"
],
[
"0.5645",
"0.1768",
"0.5933",
"0.1768",
"0.5933",
"0.1895",
"0.5645",
"0.1895"
],
[
"0.5073",
"0.178",
"0.544",
"0.178",
"0.544",
"0.1895",
"0.5073",
"0.1895"
],
[
"0.1503",
"0.178",
"0.2345",
"0.178",
"0.2345",
"0.1881",
"0.1503",
"0.1881"
],
[
"0.1487",
"0.1895",
"0.2615",
"0.1895",
"0.2615",
"0.2026",
"0.1487",
"0.2026"
],
[
"0.3447",
"0.1875",
"0.4004",
"0.1895",
"0.3987",
"0.2039",
"0.343",
"0.202"
],
[
"0.6235",
"0.1895",
"0.6567",
"0.1895",
"0.6567",
"0.202",
"0.6235",
"0.202"
],
[
"0.697",
"0.1895",
"0.7305",
"0.1895",
"0.7305",
"0.202",
"0.697",
"0.202"
],
[
"0.7715",
"0.1895",
"0.8057",
"0.1895",
"0.8057",
"0.202",
"0.7715",
"0.202"
],
[
"0.1503",
"0.202",
"0.2443",
"0.202",
"0.2443",
"0.2134",
"0.1503",
"0.2134"
],
[
"0.3464",
"0.202",
"0.3987",
"0.202",
"0.3987",
"0.2134",
"0.3464",
"0.2134"
],
[
"0.4558",
"0.202",
"0.4927",
"0.202",
"0.4927",
"0.2134",
"0.4558",
"0.2134"
],
[
"0.509",
"0.202",
"0.544",
"0.202",
"0.544",
"0.2134",
"0.509",
"0.2134"
],
[
"0.5645",
"0.202",
"0.5933",
"0.202",
"0.5933",
"0.2134",
"0.5645",
"0.2134"
],
[
"0.6235",
"0.202",
"0.6567",
"0.202",
"0.6567",
"0.2134",
"0.6235",
"0.2134"
],
[
"0.697",
"0.202",
"0.7305",
"0.202",
"0.7305",
"0.2134",
"0.697",
"0.2134"
],
[
"0.7715",
"0.202",
"0.8057",
"0.202",
"0.8057",
"0.2134",
"0.7715",
"0.2134"
],
[
"0.1487",
"0.2134",
"0.25",
"0.2134",
"0.25",
"0.2279",
"0.1487",
"0.2279"
],
[
"0.3447",
"0.2122",
"0.4004",
"0.214",
"0.3987",
"0.2291",
"0.343",
"0.2267"
],
[
"0.4558",
"0.214",
"0.4944",
"0.214",
"0.4944",
"0.226",
"0.4558",
"0.226"
],
[
"0.5073",
"0.214",
"0.544",
"0.214",
"0.544",
"0.226",
"0.5073",
"0.226"
],
[
"0.5605",
"0.214",
"0.596",
"0.214",
"0.596",
"0.226",
"0.5605",
"0.226"
],
[
"0.6235",
"0.214",
"0.6567",
"0.214",
"0.6567",
"0.226",
"0.6235",
"0.226"
],
[
"0.6953",
"0.214",
"0.7305",
"0.214",
"0.7305",
"0.226",
"0.6953",
"0.226"
],
[
"0.7715",
"0.214",
"0.8057",
"0.214",
"0.8057",
"0.226",
"0.7715",
"0.226"
],
[
"0.1487",
"0.226",
"0.2345",
"0.2279",
"0.2345",
"0.2405",
"0.1487",
"0.238"
],
[
"0.3447",
"0.2247",
"0.4004",
"0.2267",
"0.3987",
"0.2412",
"0.343",
"0.2393"
],
[
"0.4558",
"0.2267",
"0.4944",
"0.2267",
"0.4944",
"0.238",
"0.4558",
"0.238"
],
[
"0.509",
"0.2267",
"0.544",
"0.2267",
"0.544",
"0.238",
"0.509",
"0.238"
],
[
"0.5605",
"0.2267",
"0.596",
"0.2267",
"0.596",
"0.238",
"0.5605",
"0.238"
],
[
"0.6216",
"0.2267",
"0.66",
"0.2267",
"0.66",
"0.2405",
"0.6216",
"0.2405"
],
[
"0.6953",
"0.2267",
"0.734",
"0.2267",
"0.734",
"0.2405",
"0.6953",
"0.2405"
],
[
"0.7715",
"0.2267",
"0.806",
"0.2267",
"0.806",
"0.2393",
"0.7715",
"0.2393"
],
[
"0.1487",
"0.238",
"0.2852",
"0.238",
"0.2852",
"0.252",
"0.1487",
"0.252"
],
[
"0.3464",
"0.2393",
"0.3987",
"0.2393",
"0.3987",
"0.2507",
"0.3464",
"0.2507"
],
[
"0.4558",
"0.2393",
"0.4927",
"0.2393",
"0.4927",
"0.2507",
"0.4558",
"0.2507"
],
[
"0.5073",
"0.2393",
"0.544",
"0.2393",
"0.544",
"0.2507",
"0.5073",
"0.2507"
],
[
"0.5596",
"0.2393",
"0.596",
"0.2393",
"0.596",
"0.2507",
"0.5596",
"0.2507"
],
[
"0.1503",
"0.252",
"0.2279",
"0.252",
"0.2279",
"0.2632",
"0.1503",
"0.2632"
],
[
"0.3464",
"0.252",
"0.3987",
"0.252",
"0.3987",
"0.2632",
"0.3464",
"0.2632"
],
[
"0.4558",
"0.2507",
"0.4944",
"0.2507",
"0.4944",
"0.2644",
"0.4558",
"0.2644"
],
[
"0.509",
"0.252",
"0.544",
"0.252",
"0.544",
"0.2632",
"0.509",
"0.2632"
],
[
"0.5605",
"0.252",
"0.596",
"0.252",
"0.596",
"0.2632",
"0.5605",
"0.2632"
],
[
"0.6235",
"0.252",
"0.6567",
"0.252",
"0.6567",
"0.2632",
"0.6235",
"0.2632"
],
[
"0.697",
"0.252",
"0.732",
"0.252",
"0.732",
"0.2632",
"0.697",
"0.2632"
],
[
"0.7715",
"0.252",
"0.806",
"0.252",
"0.806",
"0.2632",
"0.7715",
"0.2632"
],
[
"0.1487",
"0.2632",
"0.1814",
"0.2651",
"0.1805",
"0.2766",
"0.1478",
"0.2747"
],
[
"0.3447",
"0.262",
"0.4004",
"0.2644",
"0.3987",
"0.279",
"0.343",
"0.2766"
],
[
"0.455",
"0.2644",
"0.4944",
"0.2644",
"0.4944",
"0.276",
"0.455",
"0.276"
],
[
"0.509",
"0.2644",
"0.544",
"0.2644",
"0.544",
"0.2766",
"0.509",
"0.2766"
],
[
"0.558",
"0.2632",
"0.5986",
"0.2632",
"0.5986",
"0.2778",
"0.558",
"0.2778"
],
[
"0.6216",
"0.2644",
"0.6567",
"0.2644",
"0.6567",
"0.2766",
"0.6216",
"0.2766"
],
[
"0.6953",
"0.2644",
"0.7305",
"0.2644",
"0.7305",
"0.2766",
"0.6953",
"0.2766"
],
[
"0.7715",
"0.2644",
"0.806",
"0.2644",
"0.806",
"0.2766",
"0.7715",
"0.2766"
],
[
"0.1487",
"0.2854",
"0.2418",
"0.2854",
"0.2418",
"0.2966",
"0.1487",
"0.2966"
],
[
"0.339",
"0.2854",
"0.4043",
"0.2854",
"0.4043",
"0.2966",
"0.339",
"0.2966"
],
[
"0.455",
"0.2854",
"0.4927",
"0.2854",
"0.4927",
"0.2966",
"0.455",
"0.2966"
],
[
"0.562",
"0.2854",
"0.597",
"0.2854",
"0.597",
"0.2966",
"0.562",
"0.2966"
],
[
"0.6235",
"0.2854",
"0.6567",
"0.2854",
"0.6567",
"0.2966",
"0.6235",
"0.2966"
],
[
"0.6953",
"0.2854",
"0.7305",
"0.2854",
"0.7305",
"0.2966",
"0.6953",
"0.2966"
],
[
"0.7715",
"0.2854",
"0.8057",
"0.2854",
"0.8057",
"0.2966",
"0.7715",
"0.2966"
],
[
"0.3367",
"0.2935",
"0.4043",
"0.2954",
"0.4043",
"0.31",
"0.3367",
"0.308"
],
[
"0.5215",
"0.2886",
"0.5337",
"0.2886",
"0.5337",
"0.2966",
"0.5215",
"0.2966"
],
[
"0.4558",
"0.2966",
"0.4944",
"0.2966",
"0.4944",
"0.31",
"0.4558",
"0.31"
],
[
"0.5605",
"0.2966",
"0.5947",
"0.2966",
"0.5947",
"0.3093",
"0.5605",
"0.3093"
],
[
"0.6235",
"0.2966",
"0.6567",
"0.2966",
"0.6567",
"0.3093",
"0.6235",
"0.3093"
],
[
"0.6953",
"0.2966",
"0.7305",
"0.2966",
"0.7305",
"0.3093",
"0.6953",
"0.3093"
],
[
"0.7715",
"0.2966",
"0.8057",
"0.2966",
"0.8057",
"0.3093",
"0.7715",
"0.3093"
],
[
"0.1478",
"0.307",
"0.25",
"0.3093",
"0.25",
"0.324",
"0.1478",
"0.3213"
],
[
"0.1478",
"0.2942",
"0.273",
"0.2954",
"0.273",
"0.31",
"0.1478",
"0.3093"
],
[
"0.3423",
"0.3093",
"0.4004",
"0.3093",
"0.4004",
"0.3208",
"0.3423",
"0.3208"
],
[
"0.4558",
"0.3093",
"0.492",
"0.3093",
"0.492",
"0.3208",
"0.4558",
"0.3208"
],
[
"0.509",
"0.3093",
"0.544",
"0.3093",
"0.544",
"0.3208",
"0.509",
"0.3208"
],
[
"0.5596",
"0.3093",
"0.597",
"0.3093",
"0.597",
"0.3225",
"0.5596",
"0.3225"
],
[
"0.6235",
"0.3093",
"0.6567",
"0.3093",
"0.6567",
"0.3208",
"0.6235",
"0.3208"
],
[
"0.697",
"0.3093",
"0.7305",
"0.3093",
"0.7305",
"0.3213",
"0.697",
"0.3213"
],
[
"0.7715",
"0.3093",
"0.806",
"0.3093",
"0.806",
"0.3213",
"0.7715",
"0.3213"
],
[
"0.1478",
"0.3193",
"0.2239",
"0.3213",
"0.2239",
"0.336",
"0.1478",
"0.334"
],
[
"0.3423",
"0.3213",
"0.4004",
"0.3213",
"0.4004",
"0.3333",
"0.3423",
"0.3333"
],
[
"0.4558",
"0.3213",
"0.4944",
"0.3213",
"0.4944",
"0.3352",
"0.4558",
"0.3352"
],
[
"0.509",
"0.3213",
"0.544",
"0.3213",
"0.544",
"0.334",
"0.509",
"0.334"
],
[
"0.5596",
"0.3213",
"0.597",
"0.3213",
"0.597",
"0.3352",
"0.5596",
"0.3352"
],
[
"0.6235",
"0.3213",
"0.6587",
"0.3213",
"0.6587",
"0.334",
"0.6235",
"0.334"
],
[
"0.697",
"0.3213",
"0.7305",
"0.3213",
"0.7305",
"0.334",
"0.697",
"0.334"
],
[
"0.7715",
"0.3213",
"0.8057",
"0.3213",
"0.8057",
"0.334",
"0.7715",
"0.334"
],
[
"0.1487",
"0.334",
"0.2279",
"0.334",
"0.2279",
"0.3452",
"0.1487",
"0.3452"
],
[
"0.3423",
"0.334",
"0.4019",
"0.334",
"0.4019",
"0.3452",
"0.3423",
"0.3452"
],
[
"0.4534",
"0.334",
"0.492",
"0.334",
"0.492",
"0.3452",
"0.4534",
"0.3452"
],
[
"0.509",
"0.334",
"0.544",
"0.334",
"0.544",
"0.3452",
"0.509",
"0.3452"
],
[
"0.5596",
"0.334",
"0.597",
"0.334",
"0.597",
"0.3452",
"0.5596",
"0.3452"
],
[
"0.6235",
"0.334",
"0.6567",
"0.334",
"0.6567",
"0.3452",
"0.6235",
"0.3452"
],
[
"0.697",
"0.334",
"0.7305",
"0.334",
"0.7305",
"0.3452",
"0.697",
"0.3452"
],
[
"0.7715",
"0.3333",
"0.8057",
"0.3333",
"0.8057",
"0.3452",
"0.7715",
"0.3452"
],
[
"0.0801",
"0.3655",
"0.4102",
"0.3655",
"0.4102",
"0.3757",
"0.0801",
"0.3757"
],
[
"0.5034",
"0.3643",
"0.8906",
"0.3643",
"0.8906",
"0.3774",
"0.5034",
"0.3774"
],
[
"0.5034",
"0.3801",
"0.8906",
"0.3801",
"0.8906",
"0.3933",
"0.5034",
"0.3933"
],
[
"0.0825",
"0.3894",
"0.468",
"0.3894",
"0.468",
"0.4028",
"0.0825",
"0.4028"
],
[
"0.505",
"0.3965",
"0.8896",
"0.3965",
"0.8896",
"0.4072",
"0.505",
"0.4072"
],
[
"0.0809",
"0.404",
"0.468",
"0.404",
"0.468",
"0.4185",
"0.0809",
"0.4185"
],
[
"0.505",
"0.4104",
"0.8906",
"0.4104",
"0.8906",
"0.4236",
"0.505",
"0.4236"
],
[
"0.0809",
"0.4192",
"0.468",
"0.4192",
"0.468",
"0.433",
"0.0809",
"0.433"
],
[
"0.5034",
"0.4255",
"0.8906",
"0.4255",
"0.8906",
"0.4395",
"0.5034",
"0.4395"
],
[
"0.0801",
"0.433",
"0.1316",
"0.435",
"0.1299",
"0.4495",
"0.0784",
"0.4475"
],
[
"0.5063",
"0.4426",
"0.886",
"0.4426",
"0.886",
"0.4526",
"0.5063",
"0.4526"
],
[
"0.10046",
"0.4495",
"0.4697",
"0.4495",
"0.4697",
"0.4634",
"0.10046",
"0.4634"
],
[
"0.5034",
"0.4539",
"0.8906",
"0.4558",
"0.8906",
"0.4702",
"0.5034",
"0.4685"
],
[
"0.0801",
"0.4653",
"0.468",
"0.4653",
"0.468",
"0.4792",
"0.0801",
"0.4792"
],
[
"0.5034",
"0.4702",
"0.888",
"0.4702",
"0.888",
"0.4841",
"0.5034",
"0.4841"
],
[
"0.0825",
"0.4792",
"0.4673",
"0.481",
"0.4673",
"0.4944",
"0.0825",
"0.4924"
],
[
"0.5034",
"0.4849",
"0.8896",
"0.4849",
"0.8896",
"0.4988",
"0.5034",
"0.4988"
],
[
"0.0809",
"0.4956",
"0.4673",
"0.4956",
"0.4673",
"0.509",
"0.0809",
"0.509"
],
[
"0.5024",
"0.5005",
"0.681",
"0.5005",
"0.681",
"0.5146",
"0.5024",
"0.5146"
],
[
"0.0809",
"0.511",
"0.468",
"0.511",
"0.468",
"0.5244",
"0.0809",
"0.5244"
],
[
"0.523",
"0.5166",
"0.892",
"0.5166",
"0.892",
"0.53",
"0.523",
"0.53"
],
[
"0.0801",
"0.526",
"0.468",
"0.5244",
"0.468",
"0.539",
"0.0801",
"0.5405"
],
[
"0.505",
"0.5327",
"0.8896",
"0.5327",
"0.8896",
"0.5435",
"0.505",
"0.5435"
],
[
"0.0801",
"0.539",
"0.468",
"0.5405",
"0.468",
"0.5547",
"0.0801",
"0.5537"
],
[
"0.5034",
"0.547",
"0.892",
"0.547",
"0.892",
"0.56",
"0.5034",
"0.56"
],
[
"0.0809",
"0.556",
"0.468",
"0.556",
"0.468",
"0.5693",
"0.0809",
"0.5693"
],
[
"0.5024",
"0.56",
"0.808",
"0.561",
"0.808",
"0.5757",
"0.5024",
"0.5747"
],
[
"0.0809",
"0.5713",
"0.10046",
"0.5713",
"0.10046",
"0.584",
"0.0809",
"0.584"
],
[
"0.5063",
"0.584",
"0.694",
"0.584",
"0.694",
"0.5977",
"0.5063",
"0.5977"
],
[
"0.7075",
"0.584",
"0.8906",
"0.5854",
"0.8906",
"0.5996",
"0.7075",
"0.5986"
],
[
"0.697",
"0.589",
"0.706",
"0.589",
"0.706",
"0.5933",
"0.697",
"0.5933"
],
[
"0.1013",
"0.587",
"0.468",
"0.587",
"0.468",
"0.601",
"0.1013",
"0.601"
],
[
"0.0809",
"0.601",
"0.4697",
"0.601",
"0.4697",
"0.6187",
"0.0809",
"0.6187"
],
[
"0.5034",
"0.601",
"0.8906",
"0.601",
"0.8906",
"0.6143",
"0.5034",
"0.6143"
],
[
"0.0809",
"0.617",
"0.468",
"0.617",
"0.468",
"0.631",
"0.0809",
"0.631"
],
[
"0.505",
"0.616",
"0.8906",
"0.616",
"0.8906",
"0.63",
"0.505",
"0.63"
],
[
"0.0809",
"0.632",
"0.4714",
"0.632",
"0.4714",
"0.646",
"0.0809",
"0.646"
],
[
"0.5034",
"0.631",
"0.8906",
"0.631",
"0.8906",
"0.6445",
"0.5034",
"0.6445"
],
[
"0.0801",
"0.6465",
"0.468",
"0.646",
"0.468",
"0.66",
"0.0801",
"0.661"
],
[
"0.5034",
"0.6465",
"0.8735",
"0.6465",
"0.8735",
"0.66",
"0.5034",
"0.66"
],
[
"0.0989",
"0.662",
"0.468",
"0.662",
"0.468",
"0.6753",
"0.0989",
"0.6753"
],
[
"0.6895",
"0.6577",
"0.706",
"0.6636",
"0.701",
"0.6724",
"0.6846",
"0.6675"
],
[
"0.08496",
"0.664",
"0.10455",
"0.664",
"0.10455",
"0.6724",
"0.08496",
"0.6724"
],
[
"0.501",
"0.6685",
"0.899",
"0.6685",
"0.899",
"0.6934",
"0.501",
"0.6934"
],
[
"0.0809",
"0.6753",
"0.4673",
"0.6753",
"0.4673",
"0.6895",
"0.0809",
"0.6895"
],
[
"0.0809",
"0.6924",
"0.468",
"0.6924",
"0.468",
"0.7056",
"0.0809",
"0.7056"
],
[
"0.6846",
"0.69",
"0.706",
"0.69",
"0.706",
"0.6987",
"0.6846",
"0.6987"
],
[
"0.0809",
"0.7085",
"0.2607",
"0.7085",
"0.2607",
"0.7217",
"0.0809",
"0.7217"
],
[
"0.509",
"0.7026",
"0.8906",
"0.7026",
"0.8906",
"0.7163",
"0.509",
"0.7163"
],
[
"0.10046",
"0.7227",
"0.4673",
"0.7227",
"0.4673",
"0.7363",
"0.10046",
"0.7363"
],
[
"0.505",
"0.7173",
"0.5767",
"0.7173",
"0.5767",
"0.729",
"0.505",
"0.729"
],
[
"0.0809",
"0.738",
"0.4673",
"0.738",
"0.4673",
"0.752",
"0.0809",
"0.752"
],
[
"0.5024",
"0.741",
"0.6406",
"0.741",
"0.6406",
"0.755",
"0.5024",
"0.755"
],
[
"0.0825",
"0.754",
"0.468",
"0.754",
"0.468",
"0.7676",
"0.0825",
"0.7676"
],
[
"0.523",
"0.7603",
"0.8896",
"0.7603",
"0.8896",
"0.7734",
"0.523",
"0.7734"
],
[
"0.0825",
"0.768",
"0.4673",
"0.768",
"0.4673",
"0.782",
"0.0825",
"0.782"
],
[
"0.5063",
"0.776",
"0.8906",
"0.776",
"0.8906",
"0.789",
"0.5063",
"0.789"
],
[
"0.0809",
"0.784",
"0.3987",
"0.784",
"0.3987",
"0.7974",
"0.0809",
"0.7974"
],
[
"0.5034",
"0.7905",
"0.8896",
"0.7905",
"0.8896",
"0.8037",
"0.5034",
"0.8037"
],
[
"0.5073",
"0.802",
"0.527",
"0.8096",
"0.517",
"0.8228",
"0.4983",
"0.815"
],
[
"0.0809",
"0.808",
"0.4019",
"0.808",
"0.4019",
"0.8184",
"0.0809",
"0.8184"
],
[
"0.0866",
"0.8276",
"0.3987",
"0.8276",
"0.3987",
"0.8384",
"0.0866",
"0.8384"
],
[
"0.3938",
"0.8276",
"0.4673",
"0.8276",
"0.4673",
"0.839",
"0.3938",
"0.839"
],
[
"0.5024",
"0.8306",
"0.6953",
"0.8335",
"0.6953",
"0.8477",
"0.5024",
"0.845"
],
[
"0.0825",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0825",
"0.855"
],
[
"0.0801",
"0.857",
"0.468",
"0.857",
"0.468",
"0.8706",
"0.0801",
"0.8706"
],
[
"0.5107",
"0.857",
"0.585",
"0.857",
"0.585",
"0.8687",
"0.5107",
"0.8687"
],
[
"0.597",
"0.855",
"0.8906",
"0.856",
"0.8906",
"0.8706",
"0.597",
"0.869"
],
[
"0.0809",
"0.872",
"0.468",
"0.872",
"0.468",
"0.8853",
"0.0809",
"0.8853"
],
[
"0.5034",
"0.872",
"0.8906",
"0.872",
"0.8906",
"0.8853",
"0.5034",
"0.8853"
],
[
"0.0809",
"0.887",
"0.4697",
"0.887",
"0.4697",
"0.901",
"0.0809",
"0.901"
],
[
"0.5034",
"0.887",
"0.8906",
"0.887",
"0.8906",
"0.901",
"0.5034",
"0.901"
]
] | [
"Table 1.NYU Depth V2 dataset: comparison with state of the art. In bold we indicate the best method adopting on the",
"Original set, while in italics we indicate the best method using the extended set.",
"Error (lower is better)",
"Accuracy (higher is better)",
"Method",
"Extra Training Data ?",
"rel",
"log10",
"rms",
"8<1.25",
"<1.252",
"8<1.253",
"Saxena et al. [29]",
"0.349",
"1.214",
"0.745",
"0.897",
"0.447",
"No (795)",
"Karsch et al. [16]",
"No (795)",
"0.35",
"0.131",
"1.20",
"I",
"No (795)",
"-",
"0.335",
"1.06",
"0.127",
"Liu et al.[24]",
"Ladicky et al. [19]",
"No (795)",
"0.542",
"0.829",
"0.941",
"Zhuo et al.[39]",
"No (795)",
"0.305",
"0.122",
"1.04",
"0.525",
"0.838",
"0.962",
"Wang et al. [3]]",
"No (795)",
"0.220",
"0.094",
"0.745",
"0.605",
"0.890",
"0.970",
"Liu et al.[23]",
"No (795)",
"0.213",
"0.087",
"0.759",
"0.650",
"0.906",
"0.976",
"Roi and Todorovic [26]",
"No (795)",
"0.187",
"0.078",
"0.744",
"Xu et al. [36]",
"No (795)",
"0.139",
"0.063",
"0.609",
"0.793",
"0.948",
"0.984",
"Ours",
"No (795)",
"0.125",
"0.057",
"0.593",
"0.806",
"0.952",
"0.986",
"Eigen et al.[]",
"Yes (120K)",
"0.215",
"0.907",
"0.611",
"0.887",
"0.971",
"Yes (120K)",
"-",
"0.158",
"0.641",
"0.769",
"0.950",
"0.988",
"Laina et al. [20]",
"Eigen and Fergus [5]",
"Yes (12K)",
"0.129",
"0.056",
"0.583",
"0.801",
"0.950",
"0.986",
"Li et al. [21]",
"Yes (24K)",
"0.139",
"0.058",
"0.505",
"0.820",
"0.960",
"0.989",
"Xu et al. [36]",
"Yes (12K)",
"0.121",
"0.052",
"0.586",
"0.811",
"0.954",
"0.987",
"3.2.3ImplementationwithNeuralNetworks",
"的实现中，采用了建议的CRF来细化从语义层res5c",
"派生的最后一个比例特征图，该语义层从从res3c和",
"为了实现整个网络的端到端优化，我们在神经网络中",
"res4f派生的其他比例特征图接收消息。res3c、res4f",
"实现了所提出的多尺度模型。目标是根据3.2.2节中",
"和res5c是不同卷积块的最后一层。在每个卷积块中，",
"描述的推导，对注意力变量和多尺度特征图执行均值",
"每一层都输出具有相同通道数的特征图。在消息传递",
"场更新",
"之前，首先使用反卷积运算对所有特征图进行上采样",
"要执行注意力模型A的均值场更新，我们遵循（",
"使其具有相同的大小，i.e.E原始输人图像的1/4分辨",
"?？）。实际上，每个注意力图as的更新可以通过以",
"率，并且所有通道的通道数都设置为256。K。和β",
"下几个步骤实现：（i）从两个关联的特征图y和ys",
"的内核大小都设置为3，步幅为1，填充为1，以具有",
"执行消息传递（y。和ys分别用相应的特征观察as",
"本地接受字段并加快计算",
"和&s初始化）。消息传递是通过卷积运算执行的，如",
"建议的多尺度CRF模块输出细化的特征图。为了获",
"as←y（Kys），其中K是对应于s比例的卷",
"得最终的预测，我们使用去卷积运算将特征图上采样",
"积内核，符号和分别表示卷积和元素方面的乘积",
"到原始分辨率作为输入图像。每次我们将特征图上采",
"运算；（i）使用as←β?as在注意图上执行消息传",
"样2倍，同时将特征通道的数量减少一半。",
"递",
"End-to-end optimization.",
"如上所述，所提出的模型",
"",
"当注意图更新时，我们将其用作指导，以更新最后",
"一个比例要素图ys。ys的平均场更新可以根据（??",
"可以端到端训练，i.e.前端编码器①。的参数，与结构",
")如下进行：(i)通过y。←K.ys执行从第s标度到",
"化注意力引导CRF①。相关的参数以及解码器①d的",
"第S标度的消息传递；（i）乘以注意力模型，并将一",
"参数可以联合优化。给定训练数据集T，在先前的工",
"元项as与ys←asasy）相加。根据（??",
"作[36]之后，我们使用平方损失函数i.e.进行优化：",
"可以类似地执行与中间尺度相对应的潜在特征的平",
"M",
".",
"Cr(1,D;0e, 0., 0a) = IIF(I; 0e, O., 0a)-Dill",
"均场更新的计算。。在我们的实现中，为了减少计算开",
"销，我们不对中间尺度执行均值场更新。迭代更新注意",
"1",
"力图和最后的比例特征图",
"整个网络通过具有标准随机梯度下降的反向传播进行",
"我们想指出的是，由于（?？），（2）和（4）中的势",
"联合优化。",
"函数的定义，我们方法中的平均场更新的计算比在[36]",
"4.Experiments",
"中考虑高斯函数用于成对电势的效率要高得多。实际",
"我们证明了所提出的方法在两个公开可用的数据集",
"上，高斯卷积在前向和后向传递中都涉及更高的计算",
"上进行实验的有效性：NYUDepthV2E[30|和KITTIE",
"开销。我们在4节中进一步讨论了这一方面。",
"[8]数据集。以下小节描述了我们的实验设置和评估结",
"果",
"3.3.Network Structureand Optimization",
"Network Structureand Implementation.",
"单目深度",
"4.1. Experimental Setup",
"估计的总体框架由CNN架构和建议的CRF模型（。",
"2）。CNN架构由两个主要组件组成，i.e.是全卷积编",
"Datasets.",
"NYUDepthV2数据集【30]具有使用",
"码器和全卷积解码器。编码器自然支持任何网络结构。",
"MicrosoftKinect收集的120K对RGB和深度图。图",
"在这项工作中，我们特别采用ResNet-50[11]。在我们",
"像分辨率为640×480像素。数据集分为训练（249个"
] | {
"arxiv_id": "1803.11029",
"language": "zh",
"page_idx": 3
} |
||
1803.11029-zh-4 | 1803.11029-zh-4.jpg | <FigureHere>
maps ( center) and ground truth ( bottom row) .
Table 2. KITTI dataset: comparison with state of the art. In bold we indicate the best performances in the monocular
setting, while in italics those corresponding to the stereo setting
<table>
<tbody>
<tr>
<th rowspan="2">Method</th>
<th colspan="2">Setting</th>
<th colspan="3">Error (lower is better)</th>
<th colspan="3">$\operatorname{Accuracy}$ (higher is better)</th>
</tr>
<tr>
<th>range</th>
<th>stereo?</th>
<th>$\overline{\text{rel}}$</th>
<th>rel Sq</th>
<th>rms</th>
<th>${\delta<1.25}$</th>
<th>$\delta<1.25^{2}$</th>
<th>1.25 $\delta<$</th>
</tr>
<tr>
<td>Saxena et al. $\left[29\right]$</td>
<td>0-80m</td>
<td>No</td>
<td>0.280</td>
<td>-</td>
<td>8.734</td>
<td>0.601</td>
<td>0.820</td>
<td>0.926</td>
</tr>
<tr>
<td>Eigen et al. [6]</td>
<td>0-80m</td>
<td>$No$</td>
<td>0.190</td>
<td> </td>
<td>7.156</td>
<td>0.692</td>
<td>0.899</td>
<td>0.967</td>
</tr>
<tr>
<td>Liu et al. [23]</td>
<td>0-80m</td>
<td>$No$</td>
<td>0.217</td>
<td>0.092</td>
<td>7.046</td>
<td>0.656</td>
<td>0.881</td>
<td>0.958</td>
</tr>
<tr>
<td>Zhou et al. $138^{\prime}$</td>
<td>0-80m</td>
<td>$No$</td>
<td>0.208</td>
<td>1.768</td>
<td>6.858</td>
<td>0.678</td>
<td>0.885</td>
<td>0.957</td>
</tr>
<tr>
<td>Kuznietsov et al. [18] (only supervised)</td>
<td>0-80m</td>
<td>$No$</td>
<td> </td>
<td> </td>
<td>4.815</td>
<td>0.845</td>
<td>0.957</td>
<td>0.987</td>
</tr>
<tr>
<td>Ours</td>
<td>$0-80m$</td>
<td>$No$</td>
<td>0.122</td>
<td>0.897</td>
<td>4.677</td>
<td>0.818</td>
<td>0.954</td>
<td>0.985</td>
</tr>
<tr>
<td>Garg et al. 7</td>
<td>0-80m</td>
<td>Yes</td>
<td>0.177</td>
<td>1.169</td>
<td>5.285</td>
<td>0.727</td>
<td>0.896</td>
<td>0.962</td>
</tr>
<tr>
<td>Garg et al. L12+ Aug 8x 1/2</td>
<td>1-50m</td>
<td>Yes</td>
<td>0.169</td>
<td>1.080</td>
<td>5.104</td>
<td>0.740</td>
<td>0.904</td>
<td>0.958</td>
</tr>
<tr>
<td>Godard et al. $\lceil9\rceil$</td>
<td>0-80m</td>
<td>Yes</td>
<td>0.148</td>
<td>1.344</td>
<td>5.927</td>
<td>0.803</td>
<td>0.922</td>
<td>0.963</td>
</tr>
<tr>
<td>Kuznietsov et al. [18]</td>
<td>0-80m</td>
<td>Yes</td>
<td> </td>
<td> </td>
<td>4.621</td>
<td>0.852</td>
<td>0.960</td>
<td>0.986</td>
</tr>
</tbody>
</table>
场景)和测试集(215 个场景)。在我们的实验中，我们考虑了1449个 RGB-D 对的子集，其中 795 个用于训练，其余的用于测试。通过将图像裁剪为$320\times240$ 像素，随机翻转并以 $\rho\in\{1,1.2,1.5\}$ 的比例缩放它们， 可以即时执行数据增强
KITTI 数据集 [8] 最初是为在自动驾驶环境中测试多个任务中的计算机视觉算法而构建的，其中包含使用安装在驾驶车辆上的 LiDAR 传感器捕获的深度图像。在我们的实验中，我们遵循 Eigen$\boxed{\mathrm{Eet~al.~}\mathrm{E[6]}}$ 提出的实验协议，并考虑与 32 个场景相对应的 22,600张图像作为训练数据，与其他 29 个场景相对应的 697张图像作为测试数据。RGB 图像分辨率相对于原始$1224\times368$像素降低了一半。地面实况深度图是通过将
从 velodyne 激光收集的 3D 点重新投影到左单目相机
中而生成的，如$\mathbb{E}[7]$ 中所述。
Implementation Details. 所提出的方法是使用 Caffe 框架[15]实现的，并在具有 12GB 内存的单个 NvidiaTitanXGPU 上运行。虽然提出的框架是通用的，但在最近的工作[36,20]之后，我们采用 ResNet50$\boxed{\mathrm{F}}[11]$ 作为前端网络架构。如上所述，在实现中，我们考虑从不同的语义卷积层 (i.e. $res3c,res4f$和$res5c$)派生的三级特征图。这些特征图与建议的 CRF 模型融合用于深度图的最终预测。在训练期间，前端网络使用ImageNet 预训练参数进行初始化。与[36]需要前端CNN 的预训练阶段不同，我们共同优化整个网络。初始学习率设置为 10e-9 ,每 40 个历元降低 10 次。总 | [
[
"0.3423",
"0.1515",
"0.857",
"0.1515",
"0.857",
"0.2854",
"0.3423",
"0.2854"
],
[
"0.152",
"0.1818",
"0.2974",
"0.1736",
"0.304",
"0.2374",
"0.1584",
"0.2456"
],
[
"0.2091",
"0.1976",
"0.2296",
"0.1976",
"0.2296",
"0.202",
"0.2091",
"0.202"
],
[
"0.524",
"0.2052",
"0.5557",
"0.2052",
"0.5557",
"0.209",
"0.524",
"0.209"
],
[
"0.2303",
"0.2134",
"0.2402",
"0.2134",
"0.2402",
"0.2172",
"0.2303",
"0.2172"
],
[
"0.5435",
"0.2184",
"0.5557",
"0.2184",
"0.5557",
"0.2229",
"0.5435",
"0.2229"
],
[
"0.1936",
"0.2279",
"0.2141",
"0.2279",
"0.2141",
"0.2317",
"0.1936",
"0.2317"
],
[
"0.1953",
"0.358",
"0.8",
"0.3547",
"0.8",
"0.4192",
"0.1953",
"0.4224"
],
[
"0.5215",
"0.377",
"0.5337",
"0.377",
"0.5337",
"0.3806",
"0.5215",
"0.3806"
],
[
"0.6323",
"0.3838",
"0.824",
"0.3838",
"0.824",
"0.459",
"0.6323",
"0.459"
],
[
"0.3489",
"0.3914",
"0.5947",
"0.3945",
"0.5933",
"0.4653",
"0.3472",
"0.4622"
],
[
"0.2141",
"0.406",
"0.2296",
"0.406",
"0.2296",
"0.4136",
"0.2141",
"0.4136"
],
[
"0.2484",
"0.4072",
"0.255",
"0.4072",
"0.255",
"0.4104",
"0.2484",
"0.4104"
],
[
"0.2566",
"0.411",
"0.3464",
"0.411",
"0.3464",
"0.4507",
"0.2566",
"0.4507"
],
[
"0.6562",
"0.4092",
"0.671",
"0.4092",
"0.671",
"0.4124",
"0.6562",
"0.4124"
],
[
"0.259",
"0.428",
"0.2656",
"0.428",
"0.2656",
"0.433",
"0.259",
"0.433"
],
[
"0.596",
"0.4287",
"0.609",
"0.4287",
"0.609",
"0.433",
"0.596",
"0.433"
],
[
"0.0809",
"0.476",
"0.892",
"0.476",
"0.892",
"0.4893",
"0.0809",
"0.4893"
],
[
"0.0809",
"0.4912",
"0.3882",
"0.4912",
"0.3882",
"0.502",
"0.0809",
"0.502"
],
[
"0.1422",
"0.5083",
"0.486",
"0.5083",
"0.486",
"0.5215",
"0.1422",
"0.5215"
],
[
"0.486",
"0.5083",
"0.8906",
"0.509",
"0.8906",
"0.521",
"0.486",
"0.5195"
],
[
"0.0784",
"0.523",
"0.509",
"0.523",
"0.509",
"0.536",
"0.0784",
"0.536"
],
[
"0.0825",
"0.509",
"0.1487",
"0.509",
"0.1487",
"0.5195",
"0.0825",
"0.5195"
],
[
"0.4885",
"0.537",
"0.6177",
"0.537",
"0.6177",
"0.5503",
"0.4885",
"0.5503"
],
[
"0.6504",
"0.536",
"0.8013",
"0.536",
"0.8013",
"0.55",
"0.6504",
"0.55"
],
[
"0.1478",
"0.5444",
"0.1929",
"0.5444",
"0.1929",
"0.5547",
"0.1478",
"0.5547"
],
[
"0.4028",
"0.5386",
"0.4478",
"0.5386",
"0.4478",
"0.55",
"0.4028",
"0.55"
],
[
"0.4307",
"0.5503",
"0.4714",
"0.5503",
"0.4714",
"0.561",
"0.4307",
"0.561"
],
[
"0.535",
"0.5503",
"0.569",
"0.5503",
"0.569",
"0.563",
"0.535",
"0.563"
],
[
"0.697",
"0.55",
"0.746",
"0.55",
"0.746",
"0.561",
"0.697",
"0.561"
],
[
"0.4958",
"0.552",
"0.511",
"0.552",
"0.511",
"0.56",
"0.4958",
"0.56"
],
[
"0.3743",
"0.552",
"0.4143",
"0.552",
"0.4143",
"0.5625",
"0.3743",
"0.5625"
],
[
"0.6313",
"0.552",
"0.6772",
"0.549",
"0.6787",
"0.56",
"0.6323",
"0.563"
],
[
"0.763",
"0.552",
"0.822",
"0.5474",
"0.8228",
"0.5596",
"0.7646",
"0.563"
],
[
"0.3757",
"0.5684",
"0.4143",
"0.5684",
"0.4143",
"0.58",
"0.3757",
"0.58"
],
[
"0.589",
"0.553",
"0.6094",
"0.553",
"0.6094",
"0.561",
"0.589",
"0.561"
],
[
"0.486",
"0.5684",
"0.5195",
"0.5684",
"0.5195",
"0.58",
"0.486",
"0.58"
],
[
"0.584",
"0.5684",
"0.6167",
"0.5684",
"0.6167",
"0.58",
"0.584",
"0.58"
],
[
"0.6396",
"0.5684",
"0.6724",
"0.5684",
"0.6724",
"0.58",
"0.6396",
"0.58"
],
[
"0.706",
"0.5684",
"0.74",
"0.5684",
"0.74",
"0.58",
"0.706",
"0.58"
],
[
"0.7754",
"0.5684",
"0.808",
"0.5684",
"0.808",
"0.58",
"0.7754",
"0.58"
],
[
"0.4429",
"0.571",
"0.4617",
"0.571",
"0.4617",
"0.579",
"0.4429",
"0.579"
],
[
"0.1478",
"0.5693",
"0.246",
"0.5693",
"0.246",
"0.58",
"0.1478",
"0.58"
],
[
"0.4844",
"0.5776",
"0.523",
"0.5776",
"0.523",
"0.5933",
"0.4844",
"0.5933"
],
[
"0.705",
"0.5776",
"0.74",
"0.5776",
"0.74",
"0.5933",
"0.705",
"0.5933"
],
[
"0.581",
"0.579",
"0.6167",
"0.5757",
"0.619",
"0.5913",
"0.5825",
"0.5933"
],
[
"0.3774",
"0.579",
"0.415",
"0.5806",
"0.4143",
"0.5923",
"0.3757",
"0.5903"
],
[
"0.4412",
"0.58",
"0.4631",
"0.58",
"0.4631",
"0.5923",
"0.4412",
"0.5923"
],
[
"0.6396",
"0.579",
"0.6724",
"0.579",
"0.6724",
"0.5913",
"0.6396",
"0.5913"
],
[
"0.7754",
"0.58",
"0.808",
"0.58",
"0.808",
"0.5923",
"0.7754",
"0.5923"
],
[
"0.1478",
"0.5913",
"0.2255",
"0.5913",
"0.2255",
"0.6016",
"0.1478",
"0.6016"
],
[
"0.1478",
"0.5806",
"0.232",
"0.5806",
"0.232",
"0.5913",
"0.1478",
"0.5913"
],
[
"0.3774",
"0.5913",
"0.415",
"0.5913",
"0.415",
"0.603",
"0.3774",
"0.603"
],
[
"0.486",
"0.5913",
"0.5215",
"0.5913",
"0.5215",
"0.603",
"0.486",
"0.603"
],
[
"0.535",
"0.5913",
"0.5703",
"0.5913",
"0.5703",
"0.603",
"0.535",
"0.603"
],
[
"0.584",
"0.5913",
"0.6177",
"0.5913",
"0.6177",
"0.603",
"0.584",
"0.603"
],
[
"0.638",
"0.5913",
"0.6724",
"0.5913",
"0.6724",
"0.603",
"0.638",
"0.603"
],
[
"0.706",
"0.5913",
"0.739",
"0.5913",
"0.739",
"0.603",
"0.706",
"0.603"
],
[
"0.7754",
"0.5913",
"0.808",
"0.5913",
"0.808",
"0.603",
"0.7754",
"0.603"
],
[
"0.1478",
"0.603",
"0.2091",
"0.603",
"0.2091",
"0.613",
"0.1478",
"0.613"
],
[
"0.4429",
"0.5923",
"0.4631",
"0.5923",
"0.4631",
"0.603",
"0.4429",
"0.603"
],
[
"0.3757",
"0.603",
"0.415",
"0.603",
"0.415",
"0.6143",
"0.3757",
"0.6143"
],
[
"0.486",
"0.603",
"0.5215",
"0.603",
"0.5215",
"0.6143",
"0.486",
"0.6143"
],
[
"0.638",
"0.603",
"0.6724",
"0.603",
"0.6724",
"0.6143",
"0.638",
"0.6143"
],
[
"0.706",
"0.603",
"0.74",
"0.603",
"0.74",
"0.6143",
"0.706",
"0.6143"
],
[
"0.7754",
"0.603",
"0.8096",
"0.603",
"0.8096",
"0.6143",
"0.7754",
"0.6143"
],
[
"0.4429",
"0.604",
"0.4631",
"0.604",
"0.4631",
"0.6143",
"0.4429",
"0.6143"
],
[
"0.2141",
"0.604",
"0.2336",
"0.604",
"0.2336",
"0.6113",
"0.2141",
"0.6113"
],
[
"0.5376",
"0.604",
"0.569",
"0.604",
"0.569",
"0.6123",
"0.5376",
"0.6123"
],
[
"0.585",
"0.604",
"0.6167",
"0.604",
"0.6167",
"0.6123",
"0.585",
"0.6123"
],
[
"0.1462",
"0.613",
"0.361",
"0.6143",
"0.361",
"0.6255",
"0.1462",
"0.625"
],
[
"0.3757",
"0.6143",
"0.415",
"0.6143",
"0.415",
"0.6255",
"0.3757",
"0.6255"
],
[
"0.4412",
"0.613",
"0.4631",
"0.613",
"0.4631",
"0.6255",
"0.4412",
"0.6255"
],
[
"0.585",
"0.6143",
"0.6177",
"0.6143",
"0.6177",
"0.6255",
"0.585",
"0.6255"
],
[
"0.638",
"0.6143",
"0.6733",
"0.6143",
"0.6733",
"0.6255",
"0.638",
"0.6255"
],
[
"0.706",
"0.6143",
"0.739",
"0.6143",
"0.739",
"0.6255",
"0.706",
"0.6255"
],
[
"0.7754",
"0.6143",
"0.8096",
"0.6143",
"0.8096",
"0.6255",
"0.7754",
"0.6255"
],
[
"0.3757",
"0.6255",
"0.415",
"0.6255",
"0.415",
"0.637",
"0.3757",
"0.637"
],
[
"0.4412",
"0.6255",
"0.4617",
"0.6255",
"0.4617",
"0.638",
"0.4412",
"0.638"
],
[
"0.535",
"0.6255",
"0.569",
"0.6255",
"0.569",
"0.637",
"0.535",
"0.637"
],
[
"0.584",
"0.6255",
"0.6167",
"0.6255",
"0.6167",
"0.637",
"0.584",
"0.637"
],
[
"0.6396",
"0.6255",
"0.6724",
"0.6255",
"0.6724",
"0.637",
"0.6396",
"0.637"
],
[
"0.706",
"0.625",
"0.739",
"0.625",
"0.739",
"0.637",
"0.706",
"0.637"
],
[
"0.7754",
"0.6255",
"0.8096",
"0.6255",
"0.8096",
"0.637",
"0.7754",
"0.637"
],
[
"0.4878",
"0.627",
"0.5186",
"0.627",
"0.5186",
"0.6353",
"0.4878",
"0.6353"
],
[
"0.1487",
"0.6274",
"0.1748",
"0.6274",
"0.1748",
"0.636",
"0.1487",
"0.636"
],
[
"0.1462",
"0.643",
"0.2173",
"0.643",
"0.2173",
"0.654",
"0.1462",
"0.654"
],
[
"0.3743",
"0.643",
"0.415",
"0.643",
"0.415",
"0.655",
"0.3743",
"0.655"
],
[
"0.4387",
"0.643",
"0.4631",
"0.643",
"0.4631",
"0.655",
"0.4387",
"0.655"
],
[
"0.4844",
"0.643",
"0.5195",
"0.643",
"0.5195",
"0.655",
"0.4844",
"0.655"
],
[
"0.536",
"0.643",
"0.5703",
"0.643",
"0.5703",
"0.655",
"0.536",
"0.655"
],
[
"0.5825",
"0.643",
"0.621",
"0.643",
"0.621",
"0.655",
"0.5825",
"0.655"
],
[
"0.638",
"0.643",
"0.6724",
"0.643",
"0.6724",
"0.655",
"0.638",
"0.655"
],
[
"0.706",
"0.643",
"0.74",
"0.643",
"0.74",
"0.655",
"0.706",
"0.655"
],
[
"0.7754",
"0.643",
"0.8096",
"0.643",
"0.8096",
"0.655",
"0.7754",
"0.655"
],
[
"0.2059",
"0.654",
"0.308",
"0.654",
"0.308",
"0.6655",
"0.2059",
"0.6655"
],
[
"0.3757",
"0.653",
"0.4167",
"0.653",
"0.4167",
"0.6675",
"0.3757",
"0.6675"
],
[
"0.4395",
"0.655",
"0.4631",
"0.655",
"0.4631",
"0.6675",
"0.4395",
"0.6675"
],
[
"0.4844",
"0.655",
"0.5195",
"0.655",
"0.5195",
"0.6665",
"0.4844",
"0.6665"
],
[
"0.536",
"0.655",
"0.569",
"0.655",
"0.569",
"0.6665",
"0.536",
"0.6665"
],
[
"0.5825",
"0.655",
"0.6167",
"0.655",
"0.6167",
"0.6675",
"0.5825",
"0.6675"
],
[
"0.6396",
"0.655",
"0.6733",
"0.655",
"0.6733",
"0.6675",
"0.6396",
"0.6675"
],
[
"0.706",
"0.655",
"0.74",
"0.655",
"0.74",
"0.6665",
"0.706",
"0.6665"
],
[
"0.7754",
"0.655",
"0.808",
"0.655",
"0.808",
"0.6665",
"0.7754",
"0.6665"
],
[
"0.1462",
"0.6655",
"0.2279",
"0.6665",
"0.2279",
"0.678",
"0.1462",
"0.677"
],
[
"0.1478",
"0.656",
"0.2133",
"0.656",
"0.2133",
"0.6665",
"0.1478",
"0.6665"
],
[
"0.3774",
"0.6665",
"0.415",
"0.6665",
"0.415",
"0.678",
"0.3774",
"0.678"
],
[
"0.4387",
"0.6665",
"0.4631",
"0.6665",
"0.4631",
"0.6787",
"0.4387",
"0.6787"
],
[
"0.486",
"0.6665",
"0.5195",
"0.6665",
"0.5195",
"0.678",
"0.486",
"0.678"
],
[
"0.536",
"0.6665",
"0.5703",
"0.6665",
"0.5703",
"0.678",
"0.536",
"0.678"
],
[
"0.5825",
"0.6665",
"0.6167",
"0.6665",
"0.6167",
"0.678",
"0.5825",
"0.678"
],
[
"0.6396",
"0.6665",
"0.6724",
"0.6665",
"0.6724",
"0.678",
"0.6396",
"0.678"
],
[
"0.706",
"0.6665",
"0.74",
"0.6665",
"0.74",
"0.678",
"0.706",
"0.678"
],
[
"0.7754",
"0.6665",
"0.8096",
"0.6665",
"0.8096",
"0.678",
"0.7754",
"0.678"
],
[
"0.1487",
"0.677",
"0.2632",
"0.677",
"0.2632",
"0.6875",
"0.1487",
"0.6875"
],
[
"0.2296",
"0.6694",
"0.2385",
"0.6694",
"0.2385",
"0.675",
"0.2296",
"0.675"
],
[
"0.3774",
"0.678",
"0.4143",
"0.678",
"0.4143",
"0.6895",
"0.3774",
"0.6895"
],
[
"0.585",
"0.678",
"0.6167",
"0.678",
"0.6167",
"0.6895",
"0.585",
"0.6895"
],
[
"0.6396",
"0.678",
"0.6724",
"0.678",
"0.6724",
"0.6895",
"0.6396",
"0.6895"
],
[
"0.706",
"0.678",
"0.74",
"0.678",
"0.74",
"0.6895",
"0.706",
"0.6895"
],
[
"0.7754",
"0.678",
"0.8096",
"0.678",
"0.8096",
"0.6895",
"0.7754",
"0.6895"
],
[
"0.5024",
"0.7007",
"0.8906",
"0.6997",
"0.8906",
"0.714",
"0.5024",
"0.7153"
],
[
"0.4429",
"0.6787",
"0.4631",
"0.6787",
"0.4631",
"0.6895",
"0.4429",
"0.6895"
],
[
"0.0825",
"0.702",
"0.4673",
"0.702",
"0.4673",
"0.7153",
"0.0825",
"0.7153"
],
[
"0.08417",
"0.7153",
"0.4656",
"0.7153",
"0.4656",
"0.729",
"0.08417",
"0.729"
],
[
"0.505",
"0.7163",
"0.71",
"0.7163",
"0.71",
"0.73",
"0.505",
"0.73"
],
[
"0.0825",
"0.732",
"0.4697",
"0.732",
"0.4697",
"0.7456",
"0.0825",
"0.7456"
],
[
"0.509",
"0.7373",
"0.8896",
"0.7373",
"0.8896",
"0.7476",
"0.509",
"0.7476"
],
[
"0.0825",
"0.747",
"0.4697",
"0.747",
"0.4697",
"0.7603",
"0.0825",
"0.7603"
],
[
"0.5034",
"0.7505",
"0.888",
"0.7505",
"0.888",
"0.765",
"0.5034",
"0.765"
],
[
"0.0801",
"0.761",
"0.246",
"0.7603",
"0.246",
"0.7744",
"0.0801",
"0.776"
],
[
"0.5034",
"0.7666",
"0.8906",
"0.7666",
"0.8906",
"0.78",
"0.5034",
"0.78"
],
[
"0.10046",
"0.781",
"0.4673",
"0.781",
"0.4673",
"0.7944",
"0.10046",
"0.7944"
],
[
"0.505",
"0.781",
"0.8857",
"0.781",
"0.8857",
"0.7944",
"0.505",
"0.7944"
],
[
"0.0801",
"0.7944",
"0.4673",
"0.7954",
"0.4673",
"0.81",
"0.0801",
"0.8096"
],
[
"0.5034",
"0.797",
"0.888",
"0.797",
"0.888",
"0.81",
"0.5034",
"0.81"
],
[
"0.0825",
"0.811",
"0.4673",
"0.811",
"0.4673",
"0.8247",
"0.0825",
"0.8247"
],
[
"0.5024",
"0.81",
"0.8896",
"0.8096",
"0.8896",
"0.8237",
"0.5024",
"0.8247"
],
[
"0.0825",
"0.8257",
"0.4656",
"0.8257",
"0.4656",
"0.839",
"0.0825",
"0.839"
],
[
"0.5024",
"0.8247",
"0.888",
"0.8257",
"0.888",
"0.8403",
"0.5024",
"0.839"
],
[
"0.0809",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0809",
"0.855"
],
[
"0.5034",
"0.8413",
"0.888",
"0.8413",
"0.888",
"0.855",
"0.5034",
"0.855"
],
[
"0.0825",
"0.857",
"0.468",
"0.857",
"0.468",
"0.8706",
"0.0825",
"0.8706"
],
[
"0.5034",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.5034",
"0.8706"
],
[
"0.0809",
"0.872",
"0.468",
"0.872",
"0.468",
"0.8853",
"0.0809",
"0.8853"
],
[
"0.5024",
"0.872",
"0.8896",
"0.872",
"0.8896",
"0.8853",
"0.5024",
"0.8853"
],
[
"0.0809",
"0.887",
"0.4673",
"0.887",
"0.4673",
"0.901",
"0.0809",
"0.901"
],
[
"0.5034",
"0.8853",
"0.8906",
"0.886",
"0.8906",
"0.9014",
"0.5034",
"0.901"
]
] | [
"702",
"",
"",
"",
"",
"",
"",
"",
"",
"17",
"",
"1",
"",
"",
"",
"",
"",
"Figure 3. Examples of predicted depth maps on the NYU V2 test dataset: original RGB images (top row), predicted depth",
"maps (center)andgroundtruth (bottomrow)",
"KITTI dataset: comparison with state of the art.",
"Inboldweindicatethebestperformancesinthemonocular",
"setting,while in italics those corresponding to the stereo setting.",
"Table2.",
"Error (lower is better)",
"Accuracy (higher is better)",
"Method",
"Setting",
"stereo?",
"sq rel",
"8<1.25",
"rel",
"range",
"8<1.25",
"<1.253",
"0-80m",
"rms",
"0.280",
"8.734",
"0.601",
"0.820",
"0.926",
"No",
"Saxena et al.[29]",
"0.190",
"0.899",
"7.156",
"0-80m",
"No",
"0.692",
"0.967",
"Liuet al.[23]",
"Eigen et al. [6]",
"0-80m",
"0.217",
"0.092",
"7.046",
"0.656",
"0.881",
"0.958",
"Zhou et al.",
"No",
"0-80m",
"0.208",
"0.678",
"0.885",
"0.957",
"No",
"138",
"1.768",
"6.858",
"Kuznietsovet al.[18] (only supervised)",
"0-80m",
"No",
"4.815",
"0.845",
"0.957",
"0.987",
"0-80m",
"No",
"0.897",
"4.677",
"0.818",
"0.954",
"0.985",
"0.122",
"Ours",
"Garg et al.",
"0-80m",
"Yes",
"0.177",
"1.169",
"5.285",
"0.727",
"0.896",
"0.962",
"7L12+Aug8x",
"1-50m",
"Yes",
"0.169",
"1.080",
"5.104",
"0.740",
"0.904",
"0.958",
"Godard et al.",
"Garg et al.",
"0-80m",
"Yes",
"0.148",
"1.344",
"5.927",
"0.803",
"0.922",
"0.963",
"Kuznietsovetal.J18",
"0",
"0-80m",
"4.621",
"0.852",
"0.960",
"0.986",
"从velodyne激光收集的3D点重新投影到左单目相机",
"Yes",
"场景）和测试集（215个场景）。在我们的实验中，我",
"们考虑了1449个RGB-D对的子集，其中795个用于",
"中而生成的，如国[7]中所述。",
"训练，其余的用于测试。通过将图像裁剪为320×240",
"ImplementationDetails.所提出的方法是使用Cafe",
"像素，随机翻转并以pE{1,1.2,1.5}的比例缩放它们,",
"框架[15]实现的，并在具有12GB内存的单个Nvidi-",
"可以即时执行数据增强",
"aTitanXGPU上运行。虽然提出的框架是通用的，但",
"KITTI数据集「8］最初是为在自动驾驶环境中测",
"在最近的工作36，20]之后，我们采用ResNet50E[11",
"试多个任务中的计算机视觉算法而构建的，其中包含",
"作为前端网络架构。如上所述，在实现中，我们考虑从",
"使用安装在驾驶车辆上的LiDAR传感器捕获的深度",
"不同的语义卷积层（i.e.res3c，res4f和res5c）派生",
"图像。在我们的实验中，我们遵循EigenEetal.E[6",
"的三级特征图。这些特征图与建议的CRF模型融合",
"提出的实验协议，并考虑与32个场景相对应的22,600",
"用于深度图的最终预测。在训练期间，前端网络使用",
"张图像作为训练数据，与其他29个场景相对应的697",
"ImageNet预训练参数进行初始化。与[36]需要前端",
"张图像作为测试数据。RGB图像分辨率相对于原始",
"CNN的预训练阶段不同，我们共同优化整个网络。初",
"1224×368像素降低了一半。地面实况深度图是通过将",
"始学习率设置为10e－9，每40个历元降低10次。总"
] | {
"arxiv_id": "1803.11029",
"language": "zh",
"page_idx": 4
} |
||
2105.11174-zh-0 | 2105.11174-zh-0.jpg | ## 检索增强的常识性生成模型
Han Wang$^1*$, Yang Liu$^2$, Chenguang Zhu$^2$, Linjun Shou$^3$,
Ming Gon$g^{3}$, Yichong X$u^{2}$, Michael Zen$g^{2}$
$^{\bar{1}}$New York University
$^{2}$Microsoft Cognitive Services Research Group $^{3}$STCA NLP Group, Microsoft, Beijing, China
[email protected]
$\{$ yaliu10,chezhu,lisho,migon,yicxu,nzeng $\}$ @microsoft.com
$\operatorname{Abstract}$
Commonsense generation is a challenging task of generating a plausible sentence describing an everyday scenario using provided concepts. Its requirement of reasoning over commonsense knowledge and compositional generalization ability even puzzles strong pre-trained language generation models. We propose a novel framework using retrieval methods to enhance both the pre-training and fine-tuning for commonsense generation. We retrieve prototype sentence candidates by concept matching and use them as auxiliary input. For finetuning, we further boost its performance with a trainable sentence retriever. We demonstrate experimentally on the largescale CommonGen benchmark that our approach achieves new state-of-the-art results.$^1$
## 1 Introduction
对人类语言常识的理解已被公认为是人工智能系统的关键组成部分。近年来，提出了许多新的任务和数据集来评估 NLP 模型的常识推理能力 (Yu et al., 2020) 。SWAG$\underline{\mathrm{E}}($Zellers et al., 2018) 是使用常识根据部分描述推断即将发生的事件的任务。CommonsenseQA (Talmor et al., 2019) 是从 ConceptNet 构建的常识性问答数据集。最近，Lin et al. (2020) 提出了CommonGen,这是评估模型生成常识推理能力的新挑战
CommonGen 要求系统根据与日常场景相关的几个概念构造一个合理的句子。此任务的两个示例如表 1 所示。这项任务具有挑战性，因为系统需要将提供的概念组织到最合理的场景中，避免违反常识，并确保生成的句子在语法上是正确的。现有的方法微调预训练的编码
${^{*}\text{Work don}}$e during internship at Microsoft.
$^{1}$The code and data are available at https://github.
com/HanNight/RE-T5
Table 1: Two concept sets and their gold corre-
sponding sentences from CommonGen dataset.
器-解码器模型，用于以级联概念作为输入的
描述构造
Fan et al. (2020) 提出了一种用于常识性生成的检索和生成方法，该方法使用原型候选句子作为辅助输入。然而，他们的猎犬是不可训练的，只适用于微调过程。在这项工作中，我们扩展了这一思想，并通过使用检索方法来增强预训练和微调阶段，提出了一种新的常识性生成框架。此外，我们设计了一个可训练的原型句子检索器，以进一步提高生成性能
我们在 CommonGen$\underline{\mathrm{E}}($Lin et al.,2020) 基准数据集上进行实验。它包含35，141个概念集和 79,051 个相应的句子。每个概念集映射到多个相应的句子。无需任何模型修改或知识图的复杂融合，我们的方法在 CommonGen 上实现了几个指标的最新结果，包括 BLEU, CIDEr 和 SPICE
# 2 Method
我们将 CommonGen 挑战构建为序列到序列任务，并采用强大的预训练编码器-解码器模型 T5$\overline{\mathrm{E}}($Raffel et al., 2020) 作为我们的基本模型。Fan et al. (2020)在外部语料库中查找与概念相关的句子可以使 CommonGen 的关系推 | [
[
"0.357",
"0.0831",
"0.6455",
"0.08435",
"0.6455",
"0.10156",
"0.357",
"0.10095"
],
[
"0.2551",
"0.11755",
"0.75",
"0.11816",
"0.75",
"0.136",
"0.2551",
"0.1354"
],
[
"0.3208",
"0.1354",
"0.684",
"0.1354",
"0.684",
"0.1526",
"0.3208",
"0.1526"
],
[
"0.4072",
"0.152",
"0.6035",
"0.1538",
"0.6035",
"0.1687",
"0.4072",
"0.1663"
],
[
"0.3047",
"0.1687",
"0.7036",
"0.1704",
"0.7036",
"0.1853",
"0.3047",
"0.1829"
],
[
"0.3015",
"0.1829",
"0.705",
"0.1853",
"0.705",
"0.2037",
"0.3015",
"0.2019"
],
[
"0.4324",
"0.2031",
"0.5767",
"0.2037",
"0.5767",
"0.2185",
"0.4324",
"0.2173"
],
[
"0.2544",
"0.2203",
"0.76",
"0.2203",
"0.76",
"0.2351",
"0.2544",
"0.2351"
],
[
"0.2644",
"0.2632",
"0.3435",
"0.2654",
"0.3435",
"0.2808",
"0.2644",
"0.279"
],
[
"0.5205",
"0.2612",
"0.6465",
"0.2625",
"0.6465",
"0.2766",
"0.5205",
"0.2756"
],
[
"0.523",
"0.2756",
"0.6978",
"0.2756",
"0.6978",
"0.2903",
"0.523",
"0.2903"
],
[
"0.147",
"0.2893",
"0.461",
"0.2915",
"0.461",
"0.306",
"0.147",
"0.304"
],
[
"0.523",
"0.2903",
"0.679",
"0.2915",
"0.679",
"0.3027",
"0.523",
"0.3018"
],
[
"0.147",
"0.3047",
"0.461",
"0.3047",
"0.461",
"0.3196",
"0.147",
"0.3196"
],
[
"0.525",
"0.3027",
"0.7827",
"0.3027",
"0.7827",
"0.3164",
"0.525",
"0.3164"
],
[
"0.1453",
"0.3184",
"0.461",
"0.3206",
"0.461",
"0.3342",
"0.1453",
"0.332"
],
[
"0.525",
"0.3154",
"0.876",
"0.3154",
"0.876",
"0.3289",
"0.525",
"0.3289"
],
[
"0.147",
"0.3342",
"0.4592",
"0.3342",
"0.4592",
"0.3474",
"0.147",
"0.3474"
],
[
"0.525",
"0.33",
"0.882",
"0.33",
"0.882",
"0.3403",
"0.525",
"0.3403"
],
[
"0.523",
"0.3416",
"0.7246",
"0.3416",
"0.7246",
"0.3552",
"0.523",
"0.3552"
],
[
"0.1486",
"0.3499",
"0.4592",
"0.3499",
"0.4592",
"0.3599",
"0.1486",
"0.3599"
],
[
"0.147",
"0.3623",
"0.461",
"0.3623",
"0.461",
"0.376",
"0.147",
"0.376"
],
[
"0.5205",
"0.357",
"0.6455",
"0.357",
"0.6455",
"0.3706",
"0.5205",
"0.3706"
],
[
"0.523",
"0.3706",
"0.6455",
"0.3706",
"0.6455",
"0.3843",
"0.523",
"0.3843"
],
[
"0.147",
"0.3772",
"0.4592",
"0.3772",
"0.4592",
"0.3901",
"0.147",
"0.3901"
],
[
"0.5205",
"0.3818",
"0.6826",
"0.383",
"0.6826",
"0.398",
"0.5205",
"0.3967"
],
[
"0.1486",
"0.3926",
"0.4575",
"0.3926",
"0.4575",
"0.4026",
"0.1486",
"0.4026"
],
[
"0.525",
"0.3967",
"0.6963",
"0.3967",
"0.6963",
"0.4104",
"0.525",
"0.4104"
],
[
"0.147",
"0.405",
"0.461",
"0.405",
"0.461",
"0.4187",
"0.147",
"0.4187"
],
[
"0.525",
"0.4092",
"0.831",
"0.4092",
"0.831",
"0.4229",
"0.525",
"0.4229"
],
[
"0.147",
"0.4192",
"0.4592",
"0.4192",
"0.4592",
"0.4329",
"0.147",
"0.4329"
],
[
"0.523",
"0.4229",
"0.8833",
"0.4229",
"0.8833",
"0.4358",
"0.523",
"0.4358"
],
[
"0.147",
"0.4329",
"0.4592",
"0.4329",
"0.4592",
"0.4478",
"0.147",
"0.4478"
],
[
"0.5225",
"0.434",
"0.561",
"0.4358",
"0.559",
"0.4478",
"0.5205",
"0.4453"
],
[
"0.1453",
"0.4453",
"0.461",
"0.4478",
"0.461",
"0.462",
"0.1453",
"0.4602"
],
[
"0.1486",
"0.4631",
"0.4575",
"0.4631",
"0.4575",
"0.4739",
"0.1486",
"0.4739"
],
[
"0.513",
"0.459",
"0.8833",
"0.4602",
"0.8833",
"0.4744",
"0.513",
"0.4739"
],
[
"0.147",
"0.4756",
"0.4592",
"0.4756",
"0.4592",
"0.4893",
"0.147",
"0.4893"
],
[
"0.513",
"0.4756",
"0.8564",
"0.4756",
"0.8564",
"0.4893",
"0.513",
"0.4893"
],
[
"0.1486",
"0.4912",
"0.4575",
"0.4912",
"0.4575",
"0.502",
"0.1486",
"0.502"
],
[
"0.147",
"0.5044",
"0.4592",
"0.5044",
"0.4592",
"0.517",
"0.147",
"0.517"
],
[
"0.147",
"0.5186",
"0.4592",
"0.5186",
"0.4592",
"0.532",
"0.147",
"0.532"
],
[
"0.5156",
"0.517",
"0.882",
"0.517",
"0.882",
"0.531",
"0.5156",
"0.531"
],
[
"0.147",
"0.533",
"0.4592",
"0.533",
"0.4592",
"0.5464",
"0.147",
"0.5464"
],
[
"0.5146",
"0.533",
"0.5903",
"0.533",
"0.5903",
"0.5474",
"0.5146",
"0.5474"
],
[
"0.147",
"0.5474",
"0.1973",
"0.5474",
"0.1973",
"0.5586",
"0.147",
"0.5586"
],
[
"0.532",
"0.55",
"0.882",
"0.55",
"0.882",
"0.564",
"0.532",
"0.564"
],
[
"0.517",
"0.5664",
"0.882",
"0.5664",
"0.882",
"0.58",
"0.517",
"0.58"
],
[
"0.1545",
"0.585",
"0.2644",
"0.585",
"0.2644",
"0.5967",
"0.1545",
"0.5967"
],
[
"0.517",
"0.582",
"0.882",
"0.582",
"0.882",
"0.5957",
"0.517",
"0.5957"
],
[
"0.5146",
"0.5967",
"0.882",
"0.597",
"0.882",
"0.6123",
"0.5146",
"0.611"
],
[
"0.11584",
"0.5874",
"0.1335",
"0.582",
"0.1394",
"0.593",
"0.12177",
"0.597"
],
[
"0.12006",
"0.6094",
"0.4878",
"0.6094",
"0.4878",
"0.6226",
"0.12006",
"0.6226"
],
[
"0.5156",
"0.6143",
"0.882",
"0.6143",
"0.882",
"0.6274",
"0.5156",
"0.6274"
],
[
"0.12177",
"0.6245",
"0.4878",
"0.6245",
"0.4878",
"0.638",
"0.12177",
"0.638"
],
[
"0.513",
"0.629",
"0.8833",
"0.63",
"0.8833",
"0.6445",
"0.513",
"0.6436"
],
[
"0.12006",
"0.64",
"0.4878",
"0.64",
"0.4878",
"0.655",
"0.12006",
"0.655"
],
[
"0.5146",
"0.6465",
"0.8833",
"0.6465",
"0.8833",
"0.66",
"0.5146",
"0.66"
],
[
"0.1184",
"0.6562",
"0.486",
"0.6562",
"0.486",
"0.6694",
"0.1184",
"0.6694"
],
[
"0.5146",
"0.661",
"0.827",
"0.662",
"0.827",
"0.677",
"0.5146",
"0.676"
],
[
"0.12006",
"0.673",
"0.4878",
"0.673",
"0.4878",
"0.686",
"0.12006",
"0.686"
],
[
"0.534",
"0.6787",
"0.8833",
"0.6787",
"0.8833",
"0.6924",
"0.534",
"0.6924"
],
[
"0.12006",
"0.6885",
"0.486",
"0.6885",
"0.486",
"0.703",
"0.12006",
"0.703"
],
[
"0.517",
"0.6953",
"0.882",
"0.6953",
"0.882",
"0.709",
"0.517",
"0.709"
],
[
"0.12006",
"0.705",
"0.4895",
"0.705",
"0.4895",
"0.7188",
"0.12006",
"0.7188"
],
[
"0.5156",
"0.7114",
"0.882",
"0.7114",
"0.882",
"0.7246",
"0.5156",
"0.7246"
],
[
"0.12006",
"0.72",
"0.486",
"0.7197",
"0.486",
"0.734",
"0.12006",
"0.7354"
],
[
"0.5146",
"0.727",
"0.881",
"0.727",
"0.881",
"0.7407",
"0.5146",
"0.7407"
],
[
"0.12006",
"0.7363",
"0.4895",
"0.7363",
"0.4895",
"0.7505",
"0.12006",
"0.7505"
],
[
"0.517",
"0.7437",
"0.882",
"0.7437",
"0.882",
"0.7573",
"0.517",
"0.7573"
],
[
"0.1234",
"0.755",
"0.2141",
"0.755",
"0.2141",
"0.7666",
"0.1234",
"0.7666"
],
[
"0.5156",
"0.759",
"0.8833",
"0.7583",
"0.8833",
"0.7725",
"0.5156",
"0.774"
],
[
"0.1368",
"0.7686",
"0.486",
"0.7695",
"0.486",
"0.783",
"0.1368",
"0.782"
],
[
"0.513",
"0.774",
"0.657",
"0.775",
"0.657",
"0.789",
"0.513",
"0.788"
],
[
"0.12006",
"0.786",
"0.486",
"0.786",
"0.486",
"0.8",
"0.12006",
"0.8"
],
[
"0.12177",
"0.802",
"0.4878",
"0.802",
"0.4878",
"0.8154",
"0.12177",
"0.8154"
],
[
"0.517",
"0.808",
"0.5283",
"0.808",
"0.5283",
"0.8154",
"0.517",
"0.8154"
],
[
"0.552",
"0.8047",
"0.6206",
"0.8047",
"0.6206",
"0.8164",
"0.552",
"0.8164"
],
[
"0.12177",
"0.8193",
"0.486",
"0.8193",
"0.486",
"0.83",
"0.12177",
"0.83"
],
[
"0.12177",
"0.8345",
"0.4895",
"0.8345",
"0.4895",
"0.847",
"0.12177",
"0.847"
],
[
"0.5156",
"0.8306",
"0.882",
"0.8306",
"0.882",
"0.844",
"0.5156",
"0.844"
],
[
"0.12006",
"0.851",
"0.4878",
"0.851",
"0.4878",
"0.8643",
"0.12006",
"0.8643"
],
[
"0.5156",
"0.847",
"0.882",
"0.847",
"0.882",
"0.861",
"0.5156",
"0.861"
],
[
"0.5146",
"0.861",
"0.882",
"0.8623",
"0.882",
"0.877",
"0.5146",
"0.876"
],
[
"0.1528",
"0.8706",
"0.4417",
"0.8706",
"0.4417",
"0.8843",
"0.1528",
"0.8843"
],
[
"0.1377",
"0.88",
"0.492",
"0.882",
"0.492",
"0.8994",
"0.1377",
"0.898"
],
[
"0.5156",
"0.879",
"0.882",
"0.879",
"0.882",
"0.8926",
"0.5156",
"0.8926"
],
[
"0.12006",
"0.898",
"0.2778",
"0.898",
"0.2778",
"0.9077",
"0.12006",
"0.9077"
],
[
"0.5146",
"0.8955",
"0.881",
"0.8955",
"0.881",
"0.909",
"0.5146",
"0.909"
]
] | [
"检索增强的常识性生成模型",
"Han Wangl*, Yang Liu?, Chenguang Zhu?, Linjun Shou?,",
"Ming Gong3, Yichong Xu?, Michael Zeng?",
"INew York University",
"2Microsoft Cognitive Services Research Group",
"3sTCA NLP Group, Microsoft, Beijing, China",
"[email protected]",
"yaliul0,chezhu,lisho,migon,yicxu,nzeng } @microsoft.com",
"Abstract",
"Concept Set#1:",
"dog, frisbee, catch, throw",
"Commonsense generation is a challenging",
"GoldTarget Sentences",
"task of generating a plausible sentence de-",
"A dog leaps to catch a thrown frisbee.",
"scribing an everyday scenario using pro-",
"The dog catches the frisbee when the boy throws it",
"vided concepts. Its requirement of reason-",
"A man throws awayhis dog's favorite frisbee expect.",
"ing him to catch it in the air.",
"ing overcommonsenseknowledgeandcom-",
"positional generalization ability even puz-",
"Concept Set #2:",
"lake, shore, canoe",
"zles strong pre-trained language generation",
"Gold Target Sentences:",
"models.Weproposeanovelframework us",
"Canoe on a shore of lake.",
"ing retrieval methods to enhance both the",
"Canoe on shore with rainbow across the lake.",
"pre-training and fine-tuning for common-",
"Several canoes parked in the grass on the shore of a",
"sense generation.We retrieve prototype",
"lake.",
"sentence candidates by concept matching",
"and use them as auxiliary input. For fine",
"Table 1: Two concept sets and their gold corre-",
"tuning, we further boost its performance",
"sponding sentences from CommonGen dataset.",
"with a trainable sentence retriever.We",
"demonstrate experimentally on the large",
"scale CommonGen benchmark that our",
"器-解码器模型，用于以级联概念作为输人的",
"approach achieves new state-of-the-art re",
"描述构造",
"sults.",
"Fanetal.（2020）提出了一种用于常识性生",
"成的检索和生成方法，该方法使用原型候选句",
"Introduction",
"子作为辅助输人。然而，他们的猎犬是不可训",
"练的，只适用于微调过程。在这项工作中，我",
"1",
"对人类语言常识的理解已被公认为是人工智能",
"们扩展了这一思想，并通过使用检索方法来增",
"系统的关键组成部分。近年来，提出了许多新",
"强预训练和微调阶段，提出了一种新的常识性",
"的任务和数据集来评估NLP模型的常识推理",
"生成框架。此外，我们设计了一个可训练的原",
"能力(Yuet al..2020)SWAGE(Zellers et al.",
"型句子检索器，以进一步提高生成性能",
"2018）是使用常识根据部分描述推断即将发",
"我们在CommonGenE(Lin etal.,2020)基",
"生的事件的任务。CommonsenseQA（Talmon",
"准数据集上进行实验。它包含35,141个概念",
"etal.，2019）是从ConceptNet构建的常识性",
"集和79.051个相应的句子。每个概念集映射",
"问答数据集。最近，Linetal.（2020）提出了",
"到多个相应的句子。无需任何模型修改或知",
"CommonGen，这是评估模型生成常识推理能",
"识图的复杂融合，我们的方法在CommonGen",
"力的新挑战",
"上实现了几个指标的最新结果，包括BLEU，",
"CommonGen要求系统根据与日常场景相关",
"CIDEr和 SPICE",
"的几个概念构造一个合理的句子。此任务的两",
"个示例如表1所示。这项任务具有挑战性，因",
"2",
"Method",
"为系统需要将提供的概念组织到最合理的场",
"景中，避免违反常识，并确保生成的句子在语",
"我们将CommonGen挑战构建为序列到序列",
"法上是正确的。现有的方法微调预训练的编码",
"任务，并采用强大的预训练编码器-解码器模",
"型T5E(Raffeletal,2020)作为我们的基本模",
"Work done during internship at Microsoft.",
"IThe code and data are available at https://github.",
"型。Fanetal.(2020)在外部语料库中查找与概",
"com/HanNight/RE-T5",
"念相关的句子可以使CommonGen的关系推"
] | {
"arxiv_id": "2105.11174",
"language": "zh",
"page_idx": 0
} |
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.