id
string | image
string | input
string | system
string | output
string | boxes
sequence | box_texts
sequence | detail
dict |
|---|---|---|---|---|---|---|---|
2305.12717-zh-3 | 2305.12717-zh-3.jpg | 记器相比，域专用标记器覆盖了更多的域内术语。特别是，这减少了词汇表外令牌的数量， 即。e。,必须拆分成多个子词的词，其嵌入质量往往与词级表示的质量不匹配(Hedderich et al.,2021)
# 2.3 Meta-Embeddings
给定来自不同域$D$的$n$ 嵌入，每个域将有一个输入表示$x_Di\in \mathbb{R} ^E$ $, 1\leq i\leq n$ ,其中$n$是域的数量$,E$是输入嵌人的维度。在这里，我们考虑两个变体：平均$\overline{\mathrm{E}}($Coates and Bollegala,2018) 和基于注意力的元嵌人$\overline{\mathrm{E}}($Kiela et al., 2018)
平均将所有嵌入合并到一个向量中，而无需
通过采用未加权平均值来训练其他参数：
(2)
$$e^{AVG}=\frac{1}{n}\sum_{i}x_{Di}\:,$$
此外，可以使用具有动态注意力权重$\alpha_Di$的加
权平均值。为此，注意力权重的计算如下：
$$\alpha_{Di}=\frac{exp(V\cdot tanh(Wx_{Di}))}{\sum_{k=1}^{n}exp(V\cdot tanh(Wx_{Dk}))}\:,$$
(3)
,其中$W\in\mathbb{R}^H\times E$ 和$V\in\mathbb{R}^1\times H$ 是在训练过程中随机初始化和学习的参数，$H$是注意力向量的维度，它是预定义的超参数
然后使用学习的注意力权重$\alpha_Di$将域嵌人
$x_{Di}$加权为一个表示向量：
(4)
$$e^{ATT}=\sum_i\alpha_{Di}\cdot x_{Di}\:,$$
由于平均只是将所有信息合并到一个向量中， 因此它不能专注于特定嵌入中的有价值的域知识。相反，基于注意力的加权允许根据嵌入的重要性进行动态组合，具体取决于当前输入令牌
如相关作品所示，这些元嵌入方法在组合不同大小和输入粒度的嵌人时会遇到严重的失配问题 (e。g。, 字符级和单词级嵌入),可以通过在单词级上学习到相同维度的其他映射来解决，以迫使所有输入嵌人指向一个公共输入空间 (Lange et al., 2021a)
我们提出的方法通过以下方式防止了这些问题：(a)保持输入粒度固定，从而减轻了对学习附加映射的需求，以及(b)通过冻结后续的转换器层，在预训练后立即将所有域嵌入定位在同一空间中。我们比较了 ğ4 节中两个变体的结果。关于元嵌入的更多信息可以在 Bollegala and O’ Neill (2022)的调查中找到
### 类似于 BERT 标记器。因此，与原始 PTLM 标 2.4 Meta-Tokenization fon Meta-Embeddings
要在具有元嵌入的单个模型中利用我们的域适应令牌化器，我们必须对齐每个令牌化器为相同输入生成的不同输出序列。由于子字标记边界和序列长度不匹配，这并不简单。因此，我们引入了三种不同的聚合方法来执行元令牌化：
(a) SPACE:我们将 空白空间上的输入序列拆分为令牌，并为每个令牌化器聚合对应于原始序列中特定令牌的所有子字令牌。
(b) DYNAMIC: 来自所有标记器的最短序列作为参考。来自较长序列的子词相应地聚合。这假设单词级知识比子单词知识更有用，并且较少的单词拆分是域内知识的指示。
(c) TRUNCATION:此方法类似于 DYNAMIC 聚合，但它仅为每个令牌使用第一个子字，而不是在将令牌拆分为更多子字时计算平均值
一旦确定了标记和子字边界，我们就从与标记器相对应的嵌人层检索子字嵌入，并在必要时执行聚合，在我们的情况下，对所有子字嵌人进行平均。每种方法的示例如表 1 所示
# 3 Experimental Setup
本节介绍四个下游任务及其各自的数据集和评估指标。我们进一步提供有关我们的模型，其超参数和基线系统的详细信息。
### 3.1 Tasks and Evaluation Measures
我们针对四个突出的下游任务评估我们的领域专用模型和基线：对话状态跟踪(DST),响应检索(RR),命名实体识别(NER)和自然语言推理 (NLI),每个任务有五个领域。表 2显示了所有数据集的统计信息。DST 被强制转换为多分类对话框任务。给定对话历史(话) 语序列)和预定义的本体，目标是预测输出状态，即。e。,(域，插槽，值)元组$\overline{\mathrm{E}}($Wu et al., 2020) 喜欢 (餐厅，价格，昂贵)。采用标准的联合目标精度作为评估措施：在每个对话回合中，它将预测的对话状态与注释的地面实况进行比较。当且仅当所有预测的槽值与地面实况完全匹配时，预测的状态才被认为是准确的。RR 是一个排名任务，与基于检索的面向任务的对话系统相关 (Henderson et al., 2019; Wu et al., 2020)。给定对话上下文，该模型对$N$数据集话语进行排名，包括对上下文的真实响应(i。e。,候选集覆盖一个真响应和$N-1$假响应)在 Henderson et al.(2019) 之后，我们报告了在给出 99 个随机抽样错误响应的情况下的最高召回率，表示为$R_{100}\textcircled {1}$ 。NER 是一个序列标记任务，旨在通过将每 | [
[
"0.12006",
"0.0873",
"0.4878",
"0.0873",
"0.4878",
"0.0974",
"0.12006",
"0.0974"
],
[
"0.4878",
"0.0873",
"0.735",
"0.0873",
"0.735",
"0.10095",
"0.4878",
"0.10095"
],
[
"0.12006",
"0.1027",
"0.486",
"0.1027",
"0.486",
"0.1164",
"0.12006",
"0.1164"
],
[
"0.55",
"0.10095",
"0.7",
"0.1027",
"0.7",
"0.11755",
"0.55",
"0.1152"
],
[
"0.12006",
"0.11816",
"0.4895",
"0.11816",
"0.4895",
"0.1318",
"0.12006",
"0.1318"
],
[
"0.513",
"0.1247",
"0.879",
"0.1247",
"0.879",
"0.1383",
"0.513",
"0.1383"
],
[
"0.1184",
"0.1342",
"0.486",
"0.1342",
"0.486",
"0.1478",
"0.1184",
"0.1478"
],
[
"0.513",
"0.1401",
"0.881",
"0.1401",
"0.881",
"0.1538",
"0.513",
"0.1538"
],
[
"0.11755",
"0.1509",
"0.486",
"0.1497",
"0.486",
"0.1633",
"0.11755",
"0.1644"
],
[
"0.5156",
"0.158",
"0.879",
"0.158",
"0.879",
"0.1687",
"0.5156",
"0.1687"
],
[
"0.1184",
"0.1687",
"0.2065",
"0.1687",
"0.2065",
"0.1799",
"0.1184",
"0.1799"
],
[
"0.5146",
"0.1729",
"0.882",
"0.1729",
"0.882",
"0.1859",
"0.5146",
"0.1859"
],
[
"0.513",
"0.1882",
"0.881",
"0.1882",
"0.881",
"0.2019",
"0.513",
"0.2019"
],
[
"0.11584",
"0.1948",
"0.309",
"0.1965",
"0.309",
"0.2114",
"0.11584",
"0.209"
],
[
"0.5146",
"0.2048",
"0.538",
"0.2048",
"0.538",
"0.2173",
"0.5146",
"0.2173"
],
[
"0.1184",
"0.2185",
"0.4895",
"0.2185",
"0.4895",
"0.2322",
"0.1184",
"0.2322"
],
[
"0.51",
"0.2203",
"0.882",
"0.2203",
"0.882",
"0.234",
"0.51",
"0.234"
],
[
"0.11755",
"0.234",
"0.486",
"0.2328",
"0.486",
"0.2477",
"0.11755",
"0.2488"
],
[
"0.511",
"0.2369",
"0.879",
"0.2369",
"0.879",
"0.2505",
"0.511",
"0.2505"
],
[
"0.1184",
"0.2505",
"0.486",
"0.2505",
"0.486",
"0.2642",
"0.1184",
"0.2642"
],
[
"0.513",
"0.253",
"0.7793",
"0.253",
"0.7793",
"0.2666",
"0.513",
"0.2666"
],
[
"0.12006",
"0.2673",
"0.4844",
"0.2673",
"0.4844",
"0.2778",
"0.12006",
"0.2778"
],
[
"0.511",
"0.2695",
"0.881",
"0.2695",
"0.881",
"0.2832",
"0.511",
"0.2832"
],
[
"0.1184",
"0.2832",
"0.4575",
"0.2832",
"0.4575",
"0.2964",
"0.1184",
"0.2964"
],
[
"0.5146",
"0.2852",
"0.879",
"0.2852",
"0.879",
"0.2986",
"0.5146",
"0.2986"
],
[
"0.1368",
"0.2998",
"0.486",
"0.2998",
"0.486",
"0.3135",
"0.1368",
"0.3135"
],
[
"0.513",
"0.3018",
"0.881",
"0.3018",
"0.881",
"0.3154",
"0.513",
"0.3154"
],
[
"0.1184",
"0.3164",
"0.4382",
"0.3164",
"0.4382",
"0.33",
"0.1184",
"0.33"
],
[
"0.513",
"0.3164",
"0.7773",
"0.3176",
"0.7773",
"0.332",
"0.513",
"0.3308"
],
[
"0.51",
"0.332",
"0.879",
"0.332",
"0.879",
"0.3457",
"0.51",
"0.3457"
],
[
"0.2258",
"0.3508",
"0.2822",
"0.3474",
"0.2837",
"0.3594",
"0.2275",
"0.3623"
],
[
"0.3242",
"0.354",
"0.3728",
"0.3557",
"0.3728",
"0.3665",
"0.3242",
"0.364"
],
[
"0.513",
"0.3486",
"0.8774",
"0.3486",
"0.8774",
"0.3623",
"0.513",
"0.3623"
],
[
"0.465",
"0.3552",
"0.4878",
"0.3552",
"0.4878",
"0.3652",
"0.465",
"0.3652"
],
[
"0.297",
"0.3652",
"0.3105",
"0.3652",
"0.3105",
"0.3718",
"0.297",
"0.3718"
],
[
"0.5146",
"0.3652",
"0.8447",
"0.3652",
"0.8447",
"0.379",
"0.5146",
"0.379"
],
[
"0.5425",
"0.3818",
"0.881",
"0.3818",
"0.881",
"0.3955",
"0.5425",
"0.3955"
],
[
"0.12006",
"0.3955",
"0.4878",
"0.3955",
"0.4878",
"0.4092",
"0.12006",
"0.4092"
],
[
"0.492",
"0.402",
"0.4995",
"0.402",
"0.4995",
"0.405",
"0.492",
"0.405"
],
[
"0.5146",
"0.3997",
"0.879",
"0.3997",
"0.879",
"0.4104",
"0.5146",
"0.4104"
],
[
"0.1184",
"0.412",
"0.4568",
"0.412",
"0.4568",
"0.4258",
"0.1184",
"0.4258"
],
[
"0.513",
"0.4146",
"0.881",
"0.4146",
"0.881",
"0.4275",
"0.513",
"0.4275"
],
[
"0.513",
"0.4312",
"0.848",
"0.4312",
"0.848",
"0.4448",
"0.513",
"0.4448"
],
[
"0.2275",
"0.4407",
"0.398",
"0.4407",
"0.398",
"0.4548",
"0.2275",
"0.4548"
],
[
"0.1411",
"0.4478",
"0.4475",
"0.4536",
"0.4458",
"0.4822",
"0.1394",
"0.4756"
],
[
"0.4634",
"0.4531",
"0.486",
"0.4531",
"0.486",
"0.4631",
"0.4634",
"0.4631"
],
[
"0.509",
"0.4573",
"0.7197",
"0.459",
"0.7197",
"0.4739",
"0.509",
"0.4714"
],
[
"0.1351",
"0.488",
"0.4878",
"0.488",
"0.4878",
"0.503",
"0.1351",
"0.503"
],
[
"0.513",
"0.4863",
"0.879",
"0.4863",
"0.879",
"0.4995",
"0.513",
"0.4995"
],
[
"0.11755",
"0.506",
"0.4836",
"0.506",
"0.4836",
"0.5195",
"0.11755",
"0.5195"
],
[
"0.513",
"0.502",
"0.881",
"0.502",
"0.881",
"0.5156",
"0.513",
"0.5156"
],
[
"0.1184",
"0.5225",
"0.3762",
"0.5225",
"0.3762",
"0.536",
"0.1184",
"0.536"
],
[
"0.5146",
"0.5186",
"0.7583",
"0.5186",
"0.7583",
"0.532",
"0.5146",
"0.532"
],
[
"0.1351",
"0.538",
"0.486",
"0.538",
"0.486",
"0.552",
"0.1351",
"0.552"
],
[
"0.51",
"0.5435",
"0.8096",
"0.546",
"0.8096",
"0.56",
"0.51",
"0.558"
],
[
"0.11584",
"0.5547",
"0.3257",
"0.554",
"0.3257",
"0.5684",
"0.11584",
"0.5693"
],
[
"0.513",
"0.568",
"0.881",
"0.568",
"0.881",
"0.5806",
"0.513",
"0.5806"
],
[
"0.2124",
"0.58",
"0.3914",
"0.586",
"0.3896",
"0.604",
"0.2107",
"0.597"
],
[
"0.4634",
"0.5874",
"0.4895",
"0.5874",
"0.4895",
"0.5996",
"0.4634",
"0.5996"
],
[
"0.513",
"0.5845",
"0.8774",
"0.5845",
"0.8774",
"0.594",
"0.513",
"0.594"
],
[
"0.293",
"0.601",
"0.3047",
"0.6094",
"0.293",
"0.6177",
"0.2805",
"0.6094"
],
[
"0.5146",
"0.5996",
"0.881",
"0.5996",
"0.881",
"0.6133",
"0.5146",
"0.6133"
],
[
"0.5146",
"0.615",
"0.882",
"0.615",
"0.882",
"0.629",
"0.5146",
"0.629"
],
[
"0.12006",
"0.63",
"0.4895",
"0.63",
"0.4895",
"0.6436",
"0.12006",
"0.6436"
],
[
"0.513",
"0.6313",
"0.881",
"0.6313",
"0.881",
"0.6445",
"0.513",
"0.6445"
],
[
"0.12006",
"0.648",
"0.4878",
"0.648",
"0.4878",
"0.6577",
"0.12006",
"0.6577"
],
[
"0.511",
"0.648",
"0.881",
"0.648",
"0.881",
"0.661",
"0.511",
"0.661"
],
[
"0.1184",
"0.662",
"0.4878",
"0.662",
"0.4878",
"0.676",
"0.1184",
"0.676"
],
[
"0.513",
"0.6646",
"0.881",
"0.6646",
"0.881",
"0.6777",
"0.513",
"0.6777"
],
[
"0.1184",
"0.6777",
"0.4878",
"0.6777",
"0.4878",
"0.6914",
"0.1184",
"0.6914"
],
[
"0.509",
"0.6797",
"0.882",
"0.6787",
"0.882",
"0.6934",
"0.509",
"0.695"
],
[
"0.1184",
"0.6914",
"0.1411",
"0.6953",
"0.1368",
"0.709",
"0.1142",
"0.706"
],
[
"0.511",
"0.695",
"0.879",
"0.695",
"0.879",
"0.708",
"0.511",
"0.708"
],
[
"0.1368",
"0.7114",
"0.4878",
"0.7114",
"0.4878",
"0.7246",
"0.1368",
"0.7246"
],
[
"0.513",
"0.712",
"0.881",
"0.712",
"0.881",
"0.7256",
"0.513",
"0.7256"
],
[
"0.1184",
"0.728",
"0.4878",
"0.728",
"0.4878",
"0.741",
"0.1184",
"0.741"
],
[
"0.511",
"0.728",
"0.881",
"0.728",
"0.881",
"0.741",
"0.511",
"0.741"
],
[
"0.1184",
"0.7437",
"0.486",
"0.742",
"0.486",
"0.7573",
"0.1184",
"0.7583"
],
[
"0.513",
"0.7446",
"0.881",
"0.7446",
"0.881",
"0.7583",
"0.513",
"0.7583"
],
[
"0.1184",
"0.7603",
"0.4836",
"0.7603",
"0.4836",
"0.771",
"0.1184",
"0.771"
],
[
"0.5146",
"0.7603",
"0.881",
"0.7603",
"0.881",
"0.774",
"0.5146",
"0.774"
],
[
"0.1184",
"0.7754",
"0.4878",
"0.7754",
"0.4878",
"0.789",
"0.1184",
"0.789"
],
[
"0.513",
"0.777",
"0.881",
"0.777",
"0.881",
"0.7905",
"0.513",
"0.7905"
],
[
"0.12006",
"0.7935",
"0.297",
"0.7935",
"0.297",
"0.804",
"0.12006",
"0.804"
],
[
"0.5146",
"0.792",
"0.881",
"0.792",
"0.881",
"0.8057",
"0.5146",
"0.8057"
],
[
"0.1368",
"0.8086",
"0.4844",
"0.8086",
"0.4844",
"0.8223",
"0.1368",
"0.8223"
],
[
"0.513",
"0.808",
"0.881",
"0.808",
"0.881",
"0.822",
"0.513",
"0.822"
],
[
"0.11755",
"0.8247",
"0.4836",
"0.8237",
"0.4836",
"0.8374",
"0.11755",
"0.8384"
],
[
"0.513",
"0.8247",
"0.881",
"0.8247",
"0.881",
"0.8384",
"0.513",
"0.8384"
],
[
"0.1184",
"0.8413",
"0.4844",
"0.8413",
"0.4844",
"0.855",
"0.1184",
"0.855"
],
[
"0.5146",
"0.8413",
"0.881",
"0.8413",
"0.881",
"0.855",
"0.5146",
"0.855"
],
[
"0.1184",
"0.857",
"0.4878",
"0.857",
"0.4878",
"0.8706",
"0.1184",
"0.8706"
],
[
"0.513",
"0.857",
"0.879",
"0.857",
"0.879",
"0.8706",
"0.513",
"0.8706"
],
[
"0.1184",
"0.8745",
"0.4844",
"0.8745",
"0.4844",
"0.8853",
"0.1184",
"0.8853"
],
[
"0.51",
"0.8726",
"0.881",
"0.8716",
"0.881",
"0.8857",
"0.51",
"0.887"
],
[
"0.11755",
"0.8896",
"0.4895",
"0.8896",
"0.4895",
"0.903",
"0.11755",
"0.903"
],
[
"0.486",
"0.8945",
"0.5054",
"0.8945",
"0.5054",
"0.8984",
"0.486",
"0.8984"
],
[
"0.513",
"0.8896",
"0.882",
"0.8896",
"0.882",
"0.903",
"0.513",
"0.903"
],
[
"0.11755",
"0.9067",
"0.3762",
"0.9067",
"0.3762",
"0.9175",
"0.11755",
"0.9175"
],
[
"0.5146",
"0.905",
"0.881",
"0.905",
"0.881",
"0.9185",
"0.5146",
"0.9185"
]
] | [
"类似于BERT标记器。因此，与原始PTLM标",
"2.4Meta-Tokenizationfor",
"记器相比，域专用标记器覆盖了更多的域内术",
"Meta-Embeddings",
"语。特别是，这减少了词汇表外令牌的数量，",
"要在具有元嵌入的单个模型中利用我们的域适",
"即。e。，必须拆分成多个子词的词，其嵌入",
"应令牌化器，我们必须对齐每个令牌化器为相",
"质量往往与词级表示的质量不匹配（Hedderich",
"同输入生成的不同输出序列。由于子字标记边",
"et al., 2021)",
"界和序列长度不匹配，这并不简单。因此，我",
"们引入了三种不同的聚合方法来执行元令牌",
"2.3Meta-Embeddings",
"化：",
"给定来自不同域D的n嵌人，每个域将有一个",
"（a）SPACE：我们将空白空间上的输入序列拆",
"输人表示αDiERE,l≤in，其中n是域的",
"分为令牌，并为每个令牌化器聚合对应于原始",
"数量，E是输人嵌入的维度。在这里，我们考虑",
"序列中特定令牌的所有子字令牌。",
"两个变体：平均F（CoatesandBollegala.2018）",
"(b)DYNAMIC：来自所有标记器的最短序列作",
"和基于注意力的元嵌人EKielaetal,2018)",
"为参考。来自较长序列的子词相应地聚合。这",
"平均将所有嵌人合并到一个向量中，而无需",
"假设单词级知识比子单词知识更有用，并且较",
"通过采用未加权平均值来训练其他参数：",
"少的单词拆分是域内知识的指示。",
"（C）TRUNCATION：此方法类似于DYNAMIC聚",
"eAVG-",
"&Di",
"合，但它仅为每个令牌使用第一个子字，而不",
"(2)",
"n",
"是在将令牌拆分为更多子字时计算平均值",
"一且确定了标记和子字边界，我们就从与标",
"此外，可以使用具有动态注意力权重αDi的加",
"",
"记器相对应的嵌人层检索子字嵌人，并在必要",
"权平均值。为此，注意力权重的计算如下：",
"时执行聚合，在我们的情况下，对所有子字嵌",
"人进行平均。每种方法的示例如表1所示",
"ecp(V .tanh(WaDi))",
"aDi = Zr-1 eap(V tanh(WaDk) ",
"(3)",
"3Experimental Setup",
"其中WERHxE和VERIxH是在训练过",
"本节介绍四个下游任务及其各自的数据集和评",
"程中随机初始化和学习的参数，H是注意力向",
"估指标。我们进一步提供有关我们的模型，其",
"量的维度，它是预定义的超参数",
"超参数和基线系统的详细信息。",
"然后使用学习的注意力权重αDi将域嵌入",
"3.1Tasks and Evaluation Measures",
"&Di加权为一个表示向量：",
"我们针对四个突出的下游任务评估我们的领域",
"eATT-aDi·EDi,",
"(4) ",
"专用模型和基线：对话状态跟踪（DST）：响",
"",
"应检索（RR），命名实体识别（NER）和自然",
"语言推理（NLI)，每个任务有五个领域。表2",
"由于平均只是将所有信息合并到一个向量中，",
"显示了所有数据集的统计信息。DST被强制",
"因此它不能专注于特定嵌中的有价值的域知",
"转换为多分类对话框任务。给定对话历史（话",
"识。相反，基于注意力的加权允许根据嵌入的",
"语序列）和预定义的本体，目标是预测输出",
"重要性进行动态组合，具体取决于当前输入令",
"状态，即。e。，（域，插槽，值）元组E(Wu",
"牌",
"etal.2020）喜欢（餐厅，价格，昂贵）。采用",
"如相关作品所示，这些元嵌人方法在组合不",
"标准的联合目标精度作为评估措施：在每个",
"同大小和输入粒度的嵌入时会遇到严重的失配",
"对话回合中，它将预测的对话状态与注释的",
"问题（e。g。，字符级和单词级嵌人），可以通",
"地面实况进行比较。当且仅当所有预测的槽值",
"过在单词级上学习到相同维度的其他映射来解",
"与地面实况完全匹配时，预测的状态才被认为",
"决，以迫使所有输入嵌入指向一个公共输人空",
"是准确的。RR是一个排名任务，与基于检索",
"[间(Langeetal.,2021a)",
"的面向任务的对话系统相关（Hendersonetal.，",
"我们提出的方法通过以下方式防止了这些问",
"2019；Wuetal.，2020）。给定对话上下文，该",
"题：（a）保持输人粒度固定，从而减轻了对学习",
"模型对N数据集话语进行排名，包括对上下",
"附加映射的需求，以及（b）通过冻结后续的转",
"文的真实响应（i。e。，候选集覆盖一个真响",
"换器层，在预训练后立即将所有域嵌入定位在",
"应和N-1假响应）在Hendersonetal.(2019)",
"同一空间中。我们比较了g4节中两个变体的",
"之后，我们报告了在给出99个随机抽样错误",
"结果。关于元嵌入的更多信息可以在Bollegala",
"",
"响应的情况下的最高召回率，表示为R100@1",
"andO'Neill（2022）的调查中找到",
"。NER是一个序列标记任务，旨在通过将每"
] | {
"arxiv_id": "2305.12717",
"language": "zh",
"page_idx": 3
} |
||
2305.12717-zh-4 | 2305.12717-zh-4.jpg | <table>
<tbody>
<tr>
<th colspan="2">Domain Text: Acetaminophen is an analgesic drug CmA</th>
<th rowspan="1">=>TOK-1: Ace#ta# mino# phen is an anal# gesic dr# ug (10 subwords) =>TOK-2: Aceta# minophen is an anal# gesic drug (7 subwords)</th>
<th rowspan="3">$\#ug$ (10 subwords) (7 subwords) $TRUNCATION$</th>
</tr>
<tr>
<th>Aggregz</th>
<th>ion: $SPACE$</th>
<th>DYNAMIC</th>
</tr>
<tr>
<th>$TOK-1$</th>
<th>IAce$\#$ta$\#$minc Or $|ana|$ $\#$ oesicl Idr$\#$</th>
<th>|Ace$\#$tal[书 半 mino t s nhen 115 2 ang anal# gesic [dr#ug]</th>
</tr>
<tr>
<td>TOK- 2</td>
<td>AnA OPCIC dnip</td>
<td>AT restc drus</td>
<td>Aceta#minophen is an anal# gesic dr</td>
</tr>
</tbody>
</table>
Table 1: Examples of our proposed aggregation approaches for meta-tokenization: SPACE, DYNAMIC, TRUNCATION for a given text and two different tokenizers( TOK-1 , TOK-2 ). The bottom of the table shows the results after aggregation. $[ab\ldots z]$ denotes the average of all embedding vectors corresponding to subword tokens $a,b$, $\cdots , z$ .
<table>
<tbody>
<tr>
<th>Task</th>
<th>Dataset</th>
<th>Domain</th>
<th>Background</th>
<th>$\operatorname{Train}/\operatorname{Dev}/\operatorname{Test}$</th>
<th>Licenset</th>
</tr>
<tr>
<td rowspan="5">DST, RR</td>
<td> </td>
<td>Taxi</td>
<td>$200K$</td>
<td>1,654/207/195</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>Restaurant</td>
<td>$200K$</td>
<td>3,813/438/437</td>
<td> </td>
</tr>
<tr>
<td>ric et al.2020) Multiwoz</td>
<td>Hotel</td>
<td>$200K$</td>
<td>3,381/416/394</td>
<td>$MIT$</td>
</tr>
<tr>
<td> </td>
<td>Train</td>
<td>$200K$</td>
<td>3,103/484/494</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>Attraction</td>
<td>$200K$</td>
<td>2,717/401/395</td>
<td> </td>
</tr>
<tr>
<td rowspan="5">$NLI$</td>
<td> </td>
<td>Government</td>
<td>46.0 K</td>
<td>77,350/2,000/2.000</td>
<td>$0ANC$</td>
</tr>
<tr>
<td> </td>
<td>Travel</td>
<td>47.4K</td>
<td>77,350/2,000/2.000</td>
<td>$OANC$</td>
</tr>
<tr>
<td>MNI.I al..2018)</td>
<td>Slate</td>
<td>214.8K</td>
<td>77,306/2,000/2.000</td>
<td>$OANC$</td>
</tr>
<tr>
<td> </td>
<td>Telephone</td>
<td>234.6K</td>
<td>83,348/2,000/2.000</td>
<td>$OANC$</td>
</tr>
<tr>
<td> </td>
<td>Fiction</td>
<td>299.5K</td>
<td>77,348/2,000/2,000</td>
<td>CC- BY- SA- 3. 0; CC- BY- 3. 0</td>
</tr>
<tr>
<td rowspan="5">$NER$</td>
<td>CoNLL(Tiong Kim Sang and De Meulder, 2003)</td>
<td>News</td>
<td>51.0K</td>
<td>14,987/3.466/3.684</td>
<td>$DUA$</td>
</tr>
<tr>
<td>$12R2-C1.IN$ al..2011)</td>
<td>Clinical</td>
<td>299.9K</td>
<td>13.052/3,263/27,625</td>
<td>$DUA$</td>
</tr>
<tr>
<td>$SFC$ 2015</td>
<td>Financial</td>
<td>$4.8K$</td>
<td>825/207/443</td>
<td>CC- BY- 3. 0</td>
</tr>
<tr>
<td>LITBANK $R:$ al..2019)</td>
<td>Fiction</td>
<td>299.5K</td>
<td>5.548/1.388/2,973</td>
<td>CC- BY- 4. 0</td>
</tr>
<tr>
<td>SOFC {:}t al., 2020) 111 $nF$</td>
<td>Science</td>
<td>300.1K</td>
<td>489/123/263</td>
<td>CC- BY- 4. 0</td>
</tr>
</tbody>
</table>
Table 2: Overview of the selected datasets for 4 tasks (DST, RR, NLI, NER) on 14 domains. For each domain, we report the number of collected in-domain texts for domain-adaptive pre-training, as well as the size and license of the downstream dataset. All selected datasets are applicable for commercial usage. tLicense: Open American National Corpus (OANC), Direct Universal Access (DUA), Creative Commons Attribution Share-Alike (CC-BYSA), Creative Commons Attribution International License (CC-BY).
### 3.3 $\textbf{Models and Baselines}$
个令牌分类为来自预定义类别集的实体类型来检测句子中的命名实体(e。g。,PERSON , ORGANIZATION )包括非实体的中性类型(O)。在之前的工作 (Tjong Kim Sang and De Meulder, 2003; Nadeau and Sekine, 2007) 之后，我们报告了严格的微型$F_{1}$分数。NLI 是
项语言理解任务，测试机器学习模型超越简单模式识别的推理能力。任务是确定假设是否在逻辑上遵循来自前提的关系，由 ENTAILMENT (true), CONTRADICTION (false)或 NEUTRAL (未定义)推断。在 Williams et al.(2018)之后， 准确性被报告为评估措施。
我们使用最广泛使用的 PTLM 进行实验：用于NER 和 NLI 的 BERT$\boxed{\mathrm{F}}($Devlin et al.,2019)。对于 DST 和 RR 作为对话任务，我们在EHung et al. (2022a) 之后使用 BERT 和 TOD-BERTE (Wu et al.,2020)进行实验，以比较通用和特定于任务的 PTLM。3 我们想强调的是，我们提出的方法可以很容易地应用于任何现有的PTLM。作为基线，我们报告了非专用变体的性能，并将其与(a)完全预训练$\mathbb{E} ( \underline {\text{Gururangan}}$ et al., 2020) , ( b) 基于适配器的模型$\overline{\mathrm{E}}($Houlsby et al.,2019)和 (c)我们的领域专用 PTLM 变体进行了比较用 TADA 训练。
3.2 Background Data for Specializatior
3.4 Hyperparameters and Optimizatior
我们从原始或相关文本源中获取未标记的背景数据集，以通过域自适应预训练来专门化我们的模型(详细信息可在附录$\overline{\mathrm{EC}}$ 中找到)。对于传销训练，我们随机采样多达 200K 个特定于领域的句子$^{2}$, 并动态屏蔽ELiu et al.(2019) 之后的子词标记的 15 个%。
在 MLM 训练期间，我们将最大序列长度固定为 256 (DST, RR) 和 128 (NER, NLI) 子字，并降低。我们在 32 个实例的批次中训练 30 个 epoch, 并在以下值中搜索最佳学习率：$\{5\cdot10^{-5},1\cdot10^{-5},1\cdot10^{-6}\}$。对开发集性能应用早期停止(耐心：3 epoch),并使用
3We use the pre-trained models from HuggingFace: bert-base-uncased (NLI, NER) and bert-base-cased , TODBERT/ TOD- BERT- JNT- V1(RR, DST).
$^{2}$Except for four low-resource domains. For these, we randomly sample 44K( GOVERNMENT, TRAVEL, NEWS ) and 4.5K(FINANCIAL )respectively. | [
[
"0.4224",
"0.08014",
"0.7153",
"0.08075",
"0.7153",
"0.0944",
"0.4224",
"0.0932"
],
[
"0.1503",
"0.0873",
"0.392",
"0.0873",
"0.392",
"0.0974",
"0.1503",
"0.0974"
],
[
"0.7188",
"0.082",
"0.797",
"0.082",
"0.797",
"0.09265",
"0.7188",
"0.09265"
],
[
"0.424",
"0.09265",
"0.678",
"0.09265",
"0.678",
"0.1027",
"0.424",
"0.1027"
],
[
"0.7188",
"0.09265",
"0.784",
"0.09265",
"0.784",
"0.1027",
"0.7188",
"0.1027"
],
[
"0.5264",
"0.1152",
"0.58",
"0.1152",
"0.58",
"0.1259",
"0.5264",
"0.1259"
],
[
"0.7515",
"0.1152",
"0.8213",
"0.1152",
"0.8213",
"0.1259",
"0.7515",
"0.1259"
],
[
"0.2603",
"0.1164",
"0.294",
"0.1164",
"0.294",
"0.1247",
"0.2603",
"0.1247"
],
[
"0.12006",
"0.1164",
"0.1873",
"0.1164",
"0.1873",
"0.1271",
"0.12006",
"0.1271"
],
[
"0.1184",
"0.13",
"0.879",
"0.13",
"0.879",
"0.1437",
"0.1184",
"0.1437"
],
[
"0.12177",
"0.1414",
"0.1587",
"0.1414",
"0.1587",
"0.1509",
"0.12177",
"0.1509"
],
[
"0.1838",
"0.1414",
"0.3972",
"0.1414",
"0.3972",
"0.152",
"0.1838",
"0.152"
],
[
"0.4534",
"0.1395",
"0.6533",
"0.1414",
"0.6533",
"0.152",
"0.4534",
"0.1497"
],
[
"0.6885",
"0.1414",
"0.8833",
"0.1414",
"0.8833",
"0.152",
"0.6885",
"0.152"
],
[
"0.11584",
"0.1675",
"0.8833",
"0.1675",
"0.8833",
"0.1812",
"0.11584",
"0.1812"
],
[
"0.11584",
"0.1818",
"0.882",
"0.1818",
"0.882",
"0.1953",
"0.11584",
"0.1953"
],
[
"0.11584",
"0.1965",
"0.882",
"0.1965",
"0.882",
"0.2102",
"0.11584",
"0.2102"
],
[
"0.1184",
"0.2144",
"0.168",
"0.2144",
"0.168",
"0.2214",
"0.1184",
"0.2214"
],
[
"0.12427",
"0.2411",
"0.1545",
"0.2411",
"0.1545",
"0.2542",
"0.12427",
"0.2542"
],
[
"0.1914",
"0.2411",
"0.2334",
"0.2411",
"0.2334",
"0.2546",
"0.1914",
"0.2546"
],
[
"0.444",
"0.2411",
"0.486",
"0.2411",
"0.486",
"0.253",
"0.444",
"0.253"
],
[
"0.525",
"0.2423",
"0.586",
"0.2423",
"0.586",
"0.2542",
"0.525",
"0.2542"
],
[
"0.617",
"0.2423",
"0.702",
"0.2423",
"0.702",
"0.253",
"0.617",
"0.253"
],
[
"0.775",
"0.2405",
"0.8237",
"0.2405",
"0.8237",
"0.2546",
"0.775",
"0.2546"
],
[
"0.444",
"0.257",
"0.471",
"0.257",
"0.471",
"0.2683",
"0.444",
"0.2683"
],
[
"0.552",
"0.257",
"0.586",
"0.257",
"0.586",
"0.2683",
"0.552",
"0.2683"
],
[
"0.617",
"0.257",
"0.702",
"0.257",
"0.702",
"0.2683",
"0.617",
"0.2683"
],
[
"0.444",
"0.2673",
"0.4995",
"0.2673",
"0.4995",
"0.2778",
"0.444",
"0.2778"
],
[
"0.552",
"0.2673",
"0.586",
"0.2673",
"0.586",
"0.279",
"0.552",
"0.279"
],
[
"0.6187",
"0.2673",
"0.7",
"0.2673",
"0.7",
"0.2778",
"0.6187",
"0.2778"
],
[
"0.1276",
"0.2766",
"0.1755",
"0.2766",
"0.1755",
"0.2874",
"0.1276",
"0.2874"
],
[
"0.1914",
"0.2766",
"0.3494",
"0.2766",
"0.3494",
"0.2874",
"0.1914",
"0.2874"
],
[
"0.444",
"0.2766",
"0.4785",
"0.2766",
"0.4785",
"0.288",
"0.444",
"0.288"
],
[
"0.552",
"0.2766",
"0.586",
"0.2766",
"0.586",
"0.288",
"0.552",
"0.288"
],
[
"0.617",
"0.2766",
"0.702",
"0.2766",
"0.702",
"0.2874",
"0.617",
"0.2874"
],
[
"0.7866",
"0.2756",
"0.8154",
"0.2756",
"0.8154",
"0.288",
"0.7866",
"0.288"
],
[
"0.552",
"0.2861",
"0.588",
"0.2861",
"0.588",
"0.2976",
"0.552",
"0.2976"
],
[
"0.617",
"0.2861",
"0.702",
"0.2861",
"0.702",
"0.2964",
"0.617",
"0.2964"
],
[
"0.4458",
"0.2957",
"0.4963",
"0.2957",
"0.4963",
"0.3071",
"0.4458",
"0.3071"
],
[
"0.444",
"0.2874",
"0.476",
"0.2874",
"0.476",
"0.2957",
"0.444",
"0.2957"
],
[
"0.552",
"0.2957",
"0.586",
"0.2957",
"0.586",
"0.3071",
"0.552",
"0.3071"
],
[
"0.617",
"0.2957",
"0.7",
"0.2957",
"0.7",
"0.306",
"0.617",
"0.306"
],
[
"0.4434",
"0.31",
"0.509",
"0.3123",
"0.509",
"0.3237",
"0.4434",
"0.3218"
],
[
"0.5483",
"0.3113",
"0.586",
"0.3113",
"0.586",
"0.3225",
"0.5483",
"0.3225"
],
[
"0.6055",
"0.3113",
"0.714",
"0.3113",
"0.714",
"0.3218",
"0.6055",
"0.3218"
],
[
"0.781",
"0.3113",
"0.8193",
"0.3113",
"0.8193",
"0.3225",
"0.781",
"0.3225"
],
[
"0.6055",
"0.3206",
"0.714",
"0.3206",
"0.714",
"0.3308",
"0.6055",
"0.3308"
],
[
"0.781",
"0.3206",
"0.8193",
"0.3206",
"0.8193",
"0.332",
"0.781",
"0.332"
],
[
"0.55",
"0.3218",
"0.5845",
"0.3218",
"0.5845",
"0.33",
"0.55",
"0.33"
],
[
"0.444",
"0.3218",
"0.482",
"0.3218",
"0.482",
"0.3333",
"0.444",
"0.3333"
],
[
"0.544",
"0.3289",
"0.588",
"0.3308",
"0.586",
"0.3425",
"0.5425",
"0.3403"
],
[
"0.7793",
"0.3289",
"0.8193",
"0.3308",
"0.818",
"0.3433",
"0.7773",
"0.3416"
],
[
"0.1931",
"0.3308",
"0.3333",
"0.3308",
"0.3333",
"0.3416",
"0.1931",
"0.3416"
],
[
"0.126",
"0.33",
"0.1545",
"0.33",
"0.1545",
"0.3416",
"0.126",
"0.3416"
],
[
"0.4434",
"0.3308",
"0.476",
"0.3308",
"0.476",
"0.3425",
"0.4434",
"0.3425"
],
[
"0.6055",
"0.3308",
"0.714",
"0.3308",
"0.714",
"0.3416",
"0.6055",
"0.3416"
],
[
"0.544",
"0.339",
"0.588",
"0.339",
"0.588",
"0.3508",
"0.544",
"0.3508"
],
[
"0.6055",
"0.339",
"0.714",
"0.339",
"0.714",
"0.3499",
"0.6055",
"0.3499"
],
[
"0.7793",
"0.3403",
"0.8193",
"0.3403",
"0.8193",
"0.3508",
"0.7793",
"0.3508"
],
[
"0.444",
"0.3499",
"0.4844",
"0.3499",
"0.4844",
"0.3599",
"0.444",
"0.3599"
],
[
"0.444",
"0.3403",
"0.4978",
"0.3403",
"0.4978",
"0.3508",
"0.444",
"0.3508"
],
[
"0.544",
"0.3499",
"0.586",
"0.3499",
"0.586",
"0.3599",
"0.544",
"0.3599"
],
[
"0.6055",
"0.3499",
"0.714",
"0.3499",
"0.714",
"0.3599",
"0.6055",
"0.3599"
],
[
"0.7314",
"0.3499",
"0.869",
"0.3499",
"0.869",
"0.3599",
"0.7314",
"0.3599"
],
[
"0.1914",
"0.3652",
"0.43",
"0.3652",
"0.43",
"0.376",
"0.1914",
"0.376"
],
[
"0.444",
"0.3652",
"0.4768",
"0.3652",
"0.4768",
"0.3772",
"0.444",
"0.3772"
],
[
"0.5474",
"0.3652",
"0.588",
"0.3652",
"0.588",
"0.3772",
"0.5474",
"0.3772"
],
[
"0.6055",
"0.3652",
"0.714",
"0.3652",
"0.714",
"0.376",
"0.6055",
"0.376"
],
[
"0.785",
"0.364",
"0.8154",
"0.364",
"0.8154",
"0.3772",
"0.785",
"0.3772"
],
[
"0.1931",
"0.3748",
"0.3494",
"0.3748",
"0.3494",
"0.3855",
"0.1931",
"0.3855"
],
[
"0.444",
"0.3748",
"0.486",
"0.3748",
"0.486",
"0.3855",
"0.444",
"0.3855"
],
[
"0.544",
"0.3748",
"0.586",
"0.3748",
"0.586",
"0.3855",
"0.544",
"0.3855"
],
[
"0.6035",
"0.3748",
"0.7153",
"0.3748",
"0.7153",
"0.3855",
"0.6035",
"0.3855"
],
[
"0.784",
"0.3735",
"0.8154",
"0.3735",
"0.8154",
"0.3855",
"0.784",
"0.3855"
],
[
"0.1276",
"0.3843",
"0.157",
"0.3843",
"0.157",
"0.3955",
"0.1276",
"0.3955"
],
[
"0.1897",
"0.383",
"0.361",
"0.3843",
"0.361",
"0.3955",
"0.1897",
"0.3943"
],
[
"0.444",
"0.3843",
"0.4937",
"0.3843",
"0.4937",
"0.3955",
"0.444",
"0.3955"
],
[
"0.555",
"0.3843",
"0.586",
"0.3843",
"0.586",
"0.3955",
"0.555",
"0.3955"
],
[
"0.623",
"0.3843",
"0.6978",
"0.3843",
"0.6978",
"0.3943",
"0.623",
"0.3943"
],
[
"0.771",
"0.3843",
"0.8296",
"0.3843",
"0.8296",
"0.3955",
"0.771",
"0.3955"
],
[
"0.1914",
"0.3943",
"0.357",
"0.3943",
"0.357",
"0.405",
"0.1914",
"0.405"
],
[
"0.444",
"0.3943",
"0.4878",
"0.3943",
"0.4878",
"0.4062",
"0.444",
"0.4062"
],
[
"0.544",
"0.3926",
"0.586",
"0.3943",
"0.5845",
"0.4062",
"0.5425",
"0.4038"
],
[
"0.608",
"0.3943",
"0.711",
"0.3943",
"0.711",
"0.405",
"0.608",
"0.405"
],
[
"0.771",
"0.3943",
"0.8296",
"0.3943",
"0.8296",
"0.405",
"0.771",
"0.405"
],
[
"0.444",
"0.402",
"0.486",
"0.4038",
"0.4844",
"0.415",
"0.4434",
"0.4133"
],
[
"0.544",
"0.402",
"0.586",
"0.4038",
"0.5845",
"0.415",
"0.5425",
"0.4133"
],
[
"0.623",
"0.4038",
"0.6978",
"0.4038",
"0.6978",
"0.4146",
"0.623",
"0.4146"
],
[
"0.7715",
"0.4038",
"0.8296",
"0.4038",
"0.8296",
"0.4146",
"0.7715",
"0.4146"
],
[
"0.11755",
"0.4287",
"0.881",
"0.4287",
"0.881",
"0.4424",
"0.11755",
"0.4424"
],
[
"0.1931",
"0.4038",
"0.3342",
"0.4038",
"0.3342",
"0.4146",
"0.1931",
"0.4146"
],
[
"0.11755",
"0.4436",
"0.881",
"0.4436",
"0.881",
"0.4573",
"0.11755",
"0.4573"
],
[
"0.11755",
"0.4578",
"0.882",
"0.4578",
"0.882",
"0.4714",
"0.11755",
"0.4714"
],
[
"0.11755",
"0.4714",
"0.8833",
"0.4714",
"0.8833",
"0.485",
"0.11755",
"0.485"
],
[
"0.11584",
"0.485",
"0.5786",
"0.485",
"0.5786",
"0.4988",
"0.11584",
"0.4988"
],
[
"0.1184",
"0.5254",
"0.486",
"0.5254",
"0.486",
"0.539",
"0.1184",
"0.539"
],
[
"0.511",
"0.527",
"0.5557",
"0.527",
"0.5557",
"0.538",
"0.511",
"0.538"
],
[
"0.555",
"0.5254",
"0.7305",
"0.5254",
"0.7305",
"0.539",
"0.555",
"0.539"
],
[
"0.1184",
"0.542",
"0.486",
"0.542",
"0.486",
"0.557",
"0.1184",
"0.557"
],
[
"0.513",
"0.5474",
"0.881",
"0.5474",
"0.881",
"0.5625",
"0.513",
"0.5625"
],
[
"0.1351",
"0.558",
"0.486",
"0.558",
"0.486",
"0.572",
"0.1351",
"0.572"
],
[
"0.511",
"0.564",
"0.881",
"0.564",
"0.881",
"0.5776",
"0.511",
"0.5776"
],
[
"0.11755",
"0.5737",
"0.486",
"0.5737",
"0.486",
"0.5874",
"0.11755",
"0.5874"
],
[
"0.513",
"0.58",
"0.881",
"0.58",
"0.881",
"0.594",
"0.513",
"0.594"
],
[
"0.11584",
"0.589",
"0.4836",
"0.589",
"0.4836",
"0.6025",
"0.11584",
"0.6025"
],
[
"0.51",
"0.5957",
"0.882",
"0.594",
"0.882",
"0.6094",
"0.51",
"0.61"
],
[
"0.11755",
"0.6055",
"0.4685",
"0.6055",
"0.4685",
"0.619",
"0.11755",
"0.619"
],
[
"0.513",
"0.6123",
"0.881",
"0.6123",
"0.881",
"0.626",
"0.513",
"0.626"
],
[
"0.11755",
"0.6226",
"0.486",
"0.6216",
"0.486",
"0.6357",
"0.11755",
"0.637"
],
[
"0.5146",
"0.6274",
"0.881",
"0.6274",
"0.881",
"0.6426",
"0.5146",
"0.6426"
],
[
"0.1184",
"0.6396",
"0.4878",
"0.6396",
"0.4878",
"0.653",
"0.1184",
"0.653"
],
[
"0.513",
"0.6445",
"0.881",
"0.6445",
"0.881",
"0.6577",
"0.513",
"0.6577"
],
[
"0.11755",
"0.654",
"0.486",
"0.655",
"0.486",
"0.6694",
"0.11755",
"0.6685"
],
[
"0.511",
"0.659",
"0.881",
"0.66",
"0.881",
"0.675",
"0.511",
"0.6733"
],
[
"0.1184",
"0.6714",
"0.4836",
"0.6714",
"0.4836",
"0.6816",
"0.1184",
"0.6816"
],
[
"0.513",
"0.676",
"0.882",
"0.677",
"0.882",
"0.6914",
"0.513",
"0.6904"
],
[
"0.11755",
"0.687",
"0.4878",
"0.687",
"0.4878",
"0.7007",
"0.11755",
"0.7007"
],
[
"0.513",
"0.6924",
"0.881",
"0.6924",
"0.881",
"0.707",
"0.513",
"0.707"
],
[
"0.11755",
"0.703",
"0.3284",
"0.703",
"0.3284",
"0.7173",
"0.11755",
"0.7173"
],
[
"0.511",
"0.708",
"0.879",
"0.708",
"0.879",
"0.7217",
"0.511",
"0.7217"
],
[
"0.513",
"0.7236",
"0.744",
"0.7236",
"0.744",
"0.7383",
"0.513",
"0.7383"
],
[
"0.1142",
"0.7383",
"0.444",
"0.7393",
"0.444",
"0.7544",
"0.1142",
"0.753"
],
[
"0.513",
"0.7544",
"0.5425",
"0.7544",
"0.5425",
"0.7656",
"0.513",
"0.7656"
],
[
"0.546",
"0.7583",
"0.553",
"0.7583",
"0.553",
"0.763",
"0.546",
"0.763"
],
[
"0.553",
"0.7544",
"0.843",
"0.753",
"0.843",
"0.767",
"0.553",
"0.7686"
],
[
"0.1184",
"0.7666",
"0.4844",
"0.7666",
"0.4844",
"0.78",
"0.1184",
"0.78"
],
[
"0.513",
"0.7754",
"0.881",
"0.7754",
"0.881",
"0.789",
"0.513",
"0.789"
],
[
"0.1184",
"0.783",
"0.486",
"0.783",
"0.486",
"0.7964",
"0.1184",
"0.7964"
],
[
"0.513",
"0.7915",
"0.881",
"0.7915",
"0.881",
"0.8047",
"0.513",
"0.8047"
],
[
"0.1184",
"0.799",
"0.4844",
"0.799",
"0.4844",
"0.8125",
"0.1184",
"0.8125"
],
[
"0.511",
"0.808",
"0.881",
"0.808",
"0.881",
"0.822",
"0.511",
"0.822"
],
[
"0.1184",
"0.814",
"0.4844",
"0.814",
"0.4844",
"0.829",
"0.1184",
"0.829"
],
[
"0.511",
"0.8237",
"0.881",
"0.8237",
"0.881",
"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",
"0.881",
"0.839",
"0.881",
"0.854",
"0.511",
"0.8525"
],
[
"0.11755",
"0.8467",
"0.3342",
"0.8457",
"0.3342",
"0.86",
"0.11755",
"0.861"
],
[
"0.513",
"0.8555",
"0.881",
"0.8555",
"0.881",
"0.869",
"0.513",
"0.869"
],
[
"0.1368",
"0.879",
"0.4895",
"0.882",
"0.4895",
"0.8965",
"0.1368",
"0.8936"
],
[
"0.532",
"0.8813",
"0.882",
"0.8813",
"0.882",
"0.8945",
"0.532",
"0.8945"
],
[
"0.11755",
"0.8936",
"0.4895",
"0.8936",
"0.4895",
"0.9067",
"0.11755",
"0.9067"
],
[
"0.51",
"0.8936",
"0.882",
"0.8936",
"0.882",
"0.9067",
"0.51",
"0.9067"
],
[
"0.11584",
"0.905",
"0.3225",
"0.905",
"0.3225",
"0.9185",
"0.11584",
"0.9185"
],
[
"0.511",
"0.905",
"0.763",
"0.905",
"0.763",
"0.9185",
"0.511",
"0.9185"
]
] | [
"=> TOK-1: Ace#ta#mino# phen is an anal #gesic dr#ug",
"Domain Text: Acetaminophen is an analgesic drug",
"(10 subwords)",
"=> TOK-2: Aceta # minophen is an anal # gesic drug",
"(7 subwords)",
"DYNAMIC",
"TRUNCATION",
"SPACE",
"Aggregation:",
"TOK-1 [Ace #ta#mino# phen] is an [anal # gesic] [dr#ug] [Ace#ta] [#mino# phen] is an anal # gesic [dr #ug] [Ace] [#mino] is an anal # gesic [dr]",
"TOK-2",
"[Aceta # minophen] is an [anal # gesic] drug",
"Aceta #minophen is an anal #gesic drug",
"Aceta # minophen is an anal # gesic drug",
"Table 1: Examples of our proposed aggregation approaches for meta-tokenization: SPACE, DYNAMIC, TRUNCA-",
"TION for a given text and two different tokenizers ( TOK-1 , TOK-2 ). The bottom of the table shows the results",
"after aggregation. [a b... z] denotes the average of all embedding vectors corresponding to subword tokens a , b,",
"...1之",
"Task",
"Dataset",
"Domain",
"Background",
"Train / Dev / Test",
"Licenset",
"Taxi",
"200K",
"1,654/207/195",
"Restaurant",
"200 K",
"3.813/438/437",
"DST,RR",
"MultiwOZ 2.1 (Eric et al.,2020)",
"Hotel",
"200K",
"3,381/416/394",
"MIT",
"200K",
"3,103/484/494",
"Attraction",
"Train",
"200 K",
"2,717/401/395",
"Government",
"46.0 K",
"77,350/2,000/2.000",
"OANC",
"77,350/2.000/2.000",
"OANC",
"47.4K",
"Travel",
"214.8K",
"OANC",
"MNLI (Williams et al., 2018)",
"'NLI",
"Slate",
"77,306 /2,000/2.000",
"234.6K",
"83.348/2,000/2.000",
"OANC",
"Fiction",
"Telephone",
"299.5K",
"77,348/2.000/2.000",
"CC-BY-SA-3.0; CC-BY-3.0",
"CoNLL (Tjong Kim Sang and De Meulder, 2003)",
"News",
"51.0K",
"14.987/3.466/3.684",
"DUA",
"I2B2-CLIN (Uzuner et al., 2011)",
"Clinical",
"299.9K",
"13052/3,263/27,625",
"DUA",
"NER",
"SEC (Salinas Alvarado et al., 2015)",
"Financial",
"4.8K",
"825/207/443",
"CC-BY-3.0",
"LITBANK (Bamman et al., 2019)",
"Fiction",
"299.5K",
"5,548 /1,388 /2,973",
"CC-BY-4.0",
"Science",
"300.1K",
"489/123/263",
"CC-BY-4.0",
"Table 2: Overview of the selected datasets for 4 tasks (DST, RR, NLI, NER) on 14 domains. For each domain,",
"SOFC (Friedrich et al., 2020)",
"we report the number of collected in-domain texts for domain-adaptive pre-training, as well as the size and license",
"of the downstream dataset. All selected datasets are applicable for commercial usage. License: Open American",
"National Corpus (OANC), Direct Universal Access (DUA), Creative Commons Attribution Share-Alike (CC-BY-",
"SA), Creative Commons Attribution International License (CC-BY).",
"个令牌分类为来自预定义类别集的实体类型",
"3.3",
" Models and Baselines",
"来检测句子中的命名实体（e。g。，PERSON",
"我们使用最广泛使用的PTLM进行实验：用于",
"ORGANIZATION）包括非实体的中性类",
"NER 和 NLI 的 BERTE(Devlin et al., 2019)。对",
"型（O）。在之前的工作（Tiong Kim Sang and",
"于DST和RR作为对话任务，我们在EHung",
"De Meulder, 2003: Nadeau and Sekine, 2007) 之",
"etal.（2022a）之后使用BERT和TOD-BERTE",
"后，我们报告了严格的微型F分数。NLI是",
"(Wuetal.,2020）进行实验，以比较通用和特",
"项语言理解任务，测试机器学习模型超越简单",
"定于任务的PTLM。3我们想强调的是，我们",
"模式识别的推理能力。任务是确定假设是否在",
"提出的方法可以很容易地应用于任何现有的",
"逻辑上遵循来自前提的关系，由ENTAILMENT",
"PTLM。作为基线，我们报告了非专用变体的",
"(true）CONTRADICTION（false）或NEUTRAI",
"性能，并将其与（a）完全预训练E(Gururangan",
"（未定义）推断。在Williams etal.（2018）之后，",
"etal.,2020），（b）基于适配器的模型E(Houlsby",
"准确性被报告为评估措施。",
"etal.,2019)和（c）我们的领域专用PTLM变",
"体进行了比较用TADA训练。",
"3.2Background Data for Specialization",
"3.4",
"",
"Hyperparameters and Optimization",
"我们从原始或相关文本源中获取未标记的背景",
"在MLM训练期间，我们将最大序列长度固",
"数据集，以通过域自适应预训练来专门化我们",
"定为256（DST，RR）和128（NER，NLI）子",
"的模型（详细信息可在附录EC中找到）。对于",
"字，并降低。我们在32个实例的批次中训",
"传销训练，我们随机采样多达200K个特定于",
"练30个epoch，并在以下值中搜索最佳学习",
"领域的句子2，并动态屏蔽ELiuetal.(2019)",
"率：[5.10-5,1·10-5,1.10-6}。对开发集",
"之后的子词标记的15个%。",
"性能应用早期停止（耐心：3epoch)，并使用",
"2Except for four low-resource domains. For these, we ran-",
"3wWe use the pre-trained models from HuggingFace:",
"domlysample44K(GoVERNMENT,TRAVEL,NEWs)and",
"bert-base-uncased (NLI, NER) and bert-base-cased,",
"4.5K (FINANCIAL ) respectively.",
"TODBERT/TOD-BERT-JNT-V1 (RR, DST)."
] | {
"arxiv_id": "2305.12717",
"language": "zh",
"page_idx": 4
} |
||
2307.01053-zh-0 | 2307.01053-zh-0.jpg | # $\mathbf{ENGAGE:~Explanation~Guided~Data}$ $\textbf{Augmentation for Graph Representation}$ Learning
Yucheng Shi$^1$,Kaixiong Zhou$^2$,and Ninghao Liu$^1\boxtimes$
No Institute Given
<NEXMATHLINE>。特别是，无监督图表示学习方法$\mathbb{E}[16,23,21]$ 引起了广泛关注，因为它们适用于各种应用，尤其是在标签稀缺的情况下。其中，最近的对比学习[34,56]通过从对比视图中学习表示来减少多余的信息，从而实现了卓越的性能。然而，大多数现有的作品[2,3,56,51]简单地应用随机采样数据增强，这可能会意外破坏图结构，从而降低表示的有效性。在计算机视觉和自然语言处理等其他领域，通过利用语义信息E[22,34] 或领域知识$\mathbb{E}[4]$ ,已经提出了更智能的数据增强方法。然而，从抽象数据类型(如图形)中定义和提取有意义的元素是模糊的。此外，图的非欧几里得性质增加了增强难度
已经做了一些初步努力来解决这个问题。例如，Fang 等人。[5] 建议使用知识图来扩充原始输入图，并帮助构建对比对，这适用于计算化学。然而，一个更常见和有趣的问题是如何在没有外部信息源的情况下设计智能增强。此外， Zhu 等人。$\mathbb{E}[57]$将中心性应用为节点重要性指标，并鼓励较低中心性的扰动节点。但是，节点中心性是一个静态指标，不一定与所有场景中的下游任务相关。一些工作[11,50]设计端到端框架来自动制定数据增强策略，但它们可能会导致高昂的计算成本。此外，它们的黑盒性质将无法在图数据中显式定位与任务相关的信息
为了解决这个问题，这项工作提出了ENGAGE(ExplaNationGuideddataAuGmEntation),其中我们使用解释来指导生成对比视图，以学习图数据上有效的无监督表示。具体来说，我们首先提出了一种新的解释方法，称为平滑激活图(SAM)。与现有的侧重于理解模型预测的解释方法不同，SAM 基于潜在空间中表示的分布来衡量节点重要性。然后，利用节点重要性得分，通过扰动边缘和节点特征，设计了一种解释引导的图增强方法，用于构造用于对比学习的成对图视图。通过利用解释，重要的图结构不太可能被事故损坏。同时，我们的 SAM 方法通过使用量化技术$\mathbb{E}[14]$ 是有效的，因此可以将其应用于训练过程。最后， 我们在多个数据集和任务上进行了实验，以证明我们提出的方法的有效性和灵活性。贡献可以总结如下：
- We propose the ENGAGE framework which leverages explanation to inform
graph augmentation, and uses contrastive learning for training representations to preserve the key parts in graphs while removing uninformative artifacts.
- We propose a new efficient explanation method called SAM for unsupervised
representation learning on graph data.
- We conduct comprehensive experiments on both node-level and graph-level
classification tasks, with two representative contrastive learning models and different encoders, demonstrating the effectiveness and flexibility of ENGAGE. | [
[
"0.2893",
"0.1465",
"0.716",
"0.1465",
"0.716",
"0.161",
"0.2893",
"0.161"
],
[
"0.2695",
"0.1692",
"0.7344",
"0.1692",
"0.7344",
"0.1843",
"0.2695",
"0.1843"
],
[
"0.4495",
"0.1881",
"0.5537",
"0.1913",
"0.5522",
"0.2102",
"0.4478",
"0.207"
],
[
"0.3179",
"0.238",
"0.689",
"0.238",
"0.689",
"0.252",
"0.3179",
"0.252"
],
[
"0.4387",
"0.2632",
"0.5645",
"0.2632",
"0.5645",
"0.2778",
"0.4387",
"0.2778"
],
[
"0.2467",
"0.2903",
"0.7837",
"0.2903",
"0.7837",
"0.3037",
"0.2467",
"0.3037"
],
[
"0.2181",
"0.3037",
"0.785",
"0.305",
"0.785",
"0.3193",
"0.2181",
"0.318"
],
[
"0.2214",
"0.3208",
"0.7837",
"0.3208",
"0.7837",
"0.334",
"0.2214",
"0.334"
],
[
"0.2214",
"0.3352",
"0.7837",
"0.3352",
"0.7837",
"0.3484",
"0.2214",
"0.3484"
],
[
"0.2173",
"0.3499",
"0.785",
"0.3516",
"0.785",
"0.3655",
"0.2173",
"0.363"
],
[
"0.2214",
"0.3662",
"0.7837",
"0.3662",
"0.7837",
"0.3801",
"0.2214",
"0.3801"
],
[
"0.2214",
"0.3806",
"0.785",
"0.3806",
"0.785",
"0.3945",
"0.2214",
"0.3945"
],
[
"0.2197",
"0.3958",
"0.66",
"0.3958",
"0.66",
"0.4092",
"0.2197",
"0.4092"
],
[
"0.2443",
"0.411",
"0.7837",
"0.411",
"0.7837",
"0.4248",
"0.2443",
"0.4248"
],
[
"0.2197",
"0.4255",
"0.7607",
"0.4255",
"0.7607",
"0.4395",
"0.2197",
"0.4395"
],
[
"0.2197",
"0.4414",
"0.7837",
"0.4414",
"0.7837",
"0.455",
"0.2197",
"0.455"
],
[
"0.2181",
"0.4558",
"0.7837",
"0.4558",
"0.7837",
"0.4697",
"0.2181",
"0.4697"
],
[
"0.2197",
"0.4717",
"0.7837",
"0.4717",
"0.7837",
"0.4849",
"0.2197",
"0.4849"
],
[
"0.232",
"0.4841",
"0.785",
"0.486",
"0.785",
"0.5005",
"0.232",
"0.4988"
],
[
"0.2197",
"0.502",
"0.785",
"0.502",
"0.785",
"0.515",
"0.2197",
"0.515"
],
[
"0.2214",
"0.518",
"0.282",
"0.518",
"0.282",
"0.53",
"0.2214",
"0.53"
],
[
"0.2443",
"0.532",
"0.79",
"0.532",
"0.79",
"0.5454",
"0.2443",
"0.5454"
],
[
"0.2197",
"0.5454",
"0.785",
"0.5474",
"0.785",
"0.561",
"0.2197",
"0.5596"
],
[
"0.2222",
"0.563",
"0.7817",
"0.563",
"0.7817",
"0.5737",
"0.2222",
"0.5737"
],
[
"0.2197",
"0.577",
"0.7837",
"0.577",
"0.7837",
"0.5903",
"0.2197",
"0.5903"
],
[
"0.2214",
"0.5923",
"0.7837",
"0.5923",
"0.7837",
"0.606",
"0.2214",
"0.606"
],
[
"0.2222",
"0.6094",
"0.7837",
"0.6094",
"0.7837",
"0.619",
"0.2222",
"0.619"
],
[
"0.2197",
"0.622",
"0.7837",
"0.622",
"0.7837",
"0.6353",
"0.2197",
"0.6353"
],
[
"0.2197",
"0.637",
"0.7837",
"0.637",
"0.7837",
"0.651",
"0.2197",
"0.651"
],
[
"0.2197",
"0.653",
"0.785",
"0.653",
"0.785",
"0.6665",
"0.2197",
"0.6665"
],
[
"0.2197",
"0.6685",
"0.407",
"0.6685",
"0.407",
"0.6816",
"0.2197",
"0.6816"
],
[
"0.2296",
"0.6914",
"0.785",
"0.6914",
"0.785",
"0.7056",
"0.2296",
"0.7056"
],
[
"0.2467",
"0.707",
"0.785",
"0.707",
"0.785",
"0.72",
"0.2467",
"0.72"
],
[
"0.246",
"0.72",
"0.786",
"0.72",
"0.786",
"0.735",
"0.246",
"0.735"
],
[
"0.2467",
"0.7363",
"0.2852",
"0.7393",
"0.2834",
"0.7505",
"0.246",
"0.7476"
],
[
"0.2279",
"0.752",
"0.786",
"0.752",
"0.786",
"0.765",
"0.2279",
"0.765"
],
[
"0.246",
"0.7676",
"0.517",
"0.7676",
"0.517",
"0.781",
"0.246",
"0.781"
],
[
"0.2263",
"0.78",
"0.785",
"0.78",
"0.785",
"0.7935",
"0.2263",
"0.7935"
],
[
"0.2467",
"0.797",
"0.785",
"0.797",
"0.785",
"0.811",
"0.2467",
"0.811"
],
[
"0.246",
"0.811",
"0.7876",
"0.81",
"0.7876",
"0.8247",
"0.246",
"0.8257"
],
[
"0.2467",
"0.8276",
"0.2998",
"0.8276",
"0.2998",
"0.839",
"0.2467",
"0.839"
]
] | [
"ENGAGE:Explanation Guided Data",
"Augmentation for Graph Representation",
"Learning",
"Yucheng Shi', Kaixiong Zhou’, and Ninghao Liu'",
"No Institute Given",
"<NEXMATHLINE>。特别是，无监督图表示学习方法E[16,23,21]引起了",
"广泛关注，因为它们适用于各种应用，尤其是在标签稀缺的情况下。其中，最近",
"的对比学习[34,56]通过从对比视图中学习表示来减少多余的信息，从而实现了",
"卓越的性能。然而，大多数现有的作品[2,3,56,51]简单地应用随机采样数据增",
"强，这可能会意外破坏图结构，从而降低表示的有效性。在计算机视觉和自然语",
"言处理等其他领域，通过利用语义信息E[22,34]或领域知识E[4]，已经提出了",
"更智能的数据增强方法。然而，从抽象数据类型（如图形）中定义和提取有意义",
"的元素是模糊的。此外，图的非欧几里得性质增加了增强难度",
"已经做了一些初步努力来解决这个问题。例如，Fang等人。[5]建议使用",
"知识图来扩充原始输入图，并帮助构建对比对，这适用于计算化学。然而，",
"个更常见和有趣的问题是如何在没有外部信息源的情况下设计智能增强。此外",
"Zhu等人。E[57]将中心性应用为节点重要性指标，并鼓励较低中心性的扰动节",
"点。但是，节点中心性是一个静态指标，不一定与所有场景中的下游任务相关。",
"-些工作[11,50]设计端到端框架来自动制定数据增强策略，但它们可能会导致",
"高昂的计算成本。此外，它们的黑盒性质将无法在图数据中显式定位与任务相",
"关的信息",
"为了解决这个问题，这项工作提出了ENGAGE(ExplaNationGuideddataAuG",
"mEntation），其中我们使用解释来指导生成对比视图，以学习图数据上有效的",
"无监督表示。具体来说，我们首先提出了一种新的解释方法，称为平滑激活图",
"(SAM)。与现有的侧重于理解模型预测的解释方法不同，SAM基于潜在空间中",
"表示的分布来衡量节点重要性。然后，利用节点重要性得分，通过扰动边缘和节",
"点特征，设计了一种解释引导的图增强方法，用于构造用于对比学习的成对图",
"视图。通过利用解释，重要的图结构不太可能被事故损坏。同时，我们的SAM",
"方法通过使用量化技术E[14]是有效的，因此可以将其应用于训练过程。最后",
"我们在多个数据集和任务上进行了实验，以证明我们提出的方法的有效性和灵",
"活性。贡献可以总结如下：",
"-- We propose the ENGAGE framework which leverages explanation to inform",
"graph augmentation, and uses contrastive learning for training representa-",
"tions to preserve the key parts in graphs while removing uninformative arti-",
"facts.",
"-- We propose a new efficient explanation method called SAM for unsupervised",
"representation learning on graph data",
"-- We conduct comprehensive experiments on both node-level and graph-level",
"classification tasks, with two representative contrastive learning models and",
"different encoders, demonstrating the effectiveness and flexibility of EN-",
"GAGE."
] | {
"arxiv_id": "2307.01053",
"language": "zh",
"page_idx": 0
} |
||
2307.01053-zh-1 | 2307.01053-zh-1.jpg | 2 Shi et al.
# 1 Related work
1.1 Representation Learning for Graph Data
学习有效的节点或图形表示有利于各种图形挖掘任务。图数据的表示学习方法有几种主要类型，包括基于随机游走的方法$\mathbb{E}[8,23,25,32]$ ,图神经网络 (GNNs)E [7,9,36,37,42,43] 和图对比学习$\mathbb{E}[2,56,45,31]$ 。作为图上表示学习的开创性方法基于随机游走的方法优化了节点表示，以便相似上下文中的节点倾向于具有相似的表示。然后，图神经网络成为处理图数据的新的最先进的架构。最后，可以使用图对比学习来进一步完善以 GNN 为编码器的表示学习过程
1.2 Graph Contrastive Learning
对比学习通过最大化不同增强视图之间的互信息来训练表示[2,12,3,33,37]。例如，GRACEE[56] 通过丢弃边缘和掩蔽特征来构造负节点示例，并采用类似于SimCLR[E[2] 的对比模型。GraphCL [51] 在图级上对四种数据增强的图进行扰动，并采用 NT-Xent 损失 [2|进行训练。SimGRACE$\boxed{\mathrm{E}}|45|$摆脱了输入数据的增加，并选择扰动模型参数。数据增强对于对比学习在下游任务上实现良好性能至关重要[22,34] 。然而，现有方法主要依赖于随机扰动，这可能会损害增强视图中的图信息。为了解决这个问题，朱等人。$\left|57\right|$建议使用节点中心性作为重要性指标，并扰动节点中心性较低的节点。Xu 等人。[46] 提出了一种信息感知表示学习模型[35] ,以在局部和全局视图中保留与任务相关的信息。Li 等人。[17] 提出了一种图形原理指导的数据增强，以更好地保留语义信息。Youetal。[50]设计一个端到端框架来自动优化数据增强。然而，如何设计适应不同图形/节点样本的数据增强，同时保持相对较低的计算成本仍然是一个挑战
1.3 Explanation for Graph Neural Networks
现有的解释 GNN 模型的方法可以分为几类$\mathbb{E}[53]$ ,例如基于替代的，基于相关性的，基于扰动的，基于生成的和基于基本原理的方法。基于替代的方法使用简化但可解释的替代方法来近似原始模型，这些方法要么利用 GNN 机制E [1,24,41] , 要么绕过模型信息$\mathbb{E}[13,38,54]$。基于相关性的方法通过将神经元之间的激活从输出层重新分配到输人层来计算输入贡献 [24,27,28]。基于扰动的方法[19,20,26,40,49]估计输入分数，假设删除重要特征将对预测产生很大影响。基于生成的方法产生最大程度激活目标神经元$\mathbb{E}[52]$ 的合成图。基于基本原理的方法找到一个最好地指导模型预测的小子图作为解释 [44,17,18] 。现有方法主要集中在有监督图学习模型上，而我们试图获得无监督图学习的解释。同时， 我们建议利用解释来改进模型，这进一步要求解释算法是有效的
# 2 Preliminaries
# 2.1 Notations
让$G=\{\mathcal{V},\mathcal{E},\boldsymbol{A},\boldsymbol{X}\}$是一个图，其中$\mathcal{V}$和$\mathcal{E}$分别表示节点和边的集合。$A\in$ $\{0,1\}^{|\mathcal{V}|\times|\mathcal{V}|}$是邻接矩阵。如果节点$v_i$和节点$v_j$连接，则$A_ij=1$ ,其中我们使用$e_{i,j}$表示该边。$X\in\mathbb{R}^|\mathcal{V}|\times D$是特征矩阵，$x_i=X_{i,:}$是节点$v_i$的特征向量 | [
[
"0.2214",
"0.11554",
"0.2336",
"0.1193",
"0.2263",
"0.13",
"0.2141",
"0.1256"
],
[
"0.232",
"0.1206",
"0.2361",
"0.1206",
"0.2361",
"0.1256",
"0.232",
"0.1256"
],
[
"0.273",
"0.11554",
"0.3325",
"0.11554",
"0.3325",
"0.13",
"0.273",
"0.13"
],
[
"0.2173",
"0.1465",
"0.2615",
"0.1465",
"0.2615",
"0.1622",
"0.2173",
"0.1622"
],
[
"0.2566",
"0.1477",
"0.3896",
"0.1477",
"0.3896",
"0.161",
"0.2566",
"0.161"
],
[
"0.2173",
"0.1724",
"0.2646",
"0.1724",
"0.2646",
"0.1875",
"0.2173",
"0.1875"
],
[
"0.267",
"0.1736",
"0.6045",
"0.1736",
"0.6045",
"0.1875",
"0.267",
"0.1875"
],
[
"0.2197",
"0.1976",
"0.7837",
"0.1976",
"0.7837",
"0.2108",
"0.2197",
"0.2108"
],
[
"0.2214",
"0.2134",
"0.786",
"0.2134",
"0.786",
"0.2267",
"0.2214",
"0.2267"
],
[
"0.2214",
"0.2279",
"0.7837",
"0.2279",
"0.7837",
"0.2412",
"0.2214",
"0.2412"
],
[
"0.2214",
"0.2437",
"0.7837",
"0.2437",
"0.7837",
"0.2568",
"0.2214",
"0.2568"
],
[
"0.2197",
"0.2588",
"0.785",
"0.2588",
"0.785",
"0.2727",
"0.2197",
"0.2727"
],
[
"0.2181",
"0.2727",
"0.6763",
"0.2734",
"0.6763",
"0.2874",
"0.2181",
"0.286"
],
[
"0.2157",
"0.31",
"0.501",
"0.3113",
"0.501",
"0.3257",
"0.2157",
"0.3245"
],
[
"0.2214",
"0.3372",
"0.7837",
"0.3372",
"0.7837",
"0.348",
"0.2214",
"0.348"
],
[
"0.2197",
"0.351",
"0.7837",
"0.351",
"0.7837",
"0.3643",
"0.2197",
"0.3643"
],
[
"0.2173",
"0.3643",
"0.7837",
"0.3655",
"0.7837",
"0.3801",
"0.2173",
"0.379"
],
[
"0.2197",
"0.3806",
"0.785",
"0.3806",
"0.785",
"0.3945",
"0.2197",
"0.3945"
],
[
"0.2214",
"0.3977",
"0.7817",
"0.3977",
"0.7817",
"0.4077",
"0.2214",
"0.4077"
],
[
"0.2214",
"0.4136",
"0.7817",
"0.4136",
"0.7817",
"0.4236",
"0.2214",
"0.4236"
],
[
"0.2197",
"0.4268",
"0.7837",
"0.4268",
"0.7837",
"0.44",
"0.2197",
"0.44"
],
[
"0.2197",
"0.4414",
"0.7837",
"0.4414",
"0.7837",
"0.4558",
"0.2197",
"0.4558"
],
[
"0.2181",
"0.4558",
"0.7876",
"0.457",
"0.7876",
"0.4702",
"0.2181",
"0.4697"
],
[
"0.2181",
"0.4717",
"0.785",
"0.4717",
"0.785",
"0.4849",
"0.2181",
"0.4849"
],
[
"0.2197",
"0.486",
"0.7837",
"0.486",
"0.7837",
"0.5",
"0.2197",
"0.5"
],
[
"0.2197",
"0.502",
"0.7817",
"0.502",
"0.7817",
"0.515",
"0.2197",
"0.515"
],
[
"0.2181",
"0.537",
"0.6006",
"0.5386",
"0.6006",
"0.553",
"0.2181",
"0.552"
],
[
"0.2197",
"0.5645",
"0.785",
"0.5645",
"0.785",
"0.5776",
"0.2197",
"0.5776"
],
[
"0.2222",
"0.5806",
"0.7817",
"0.5806",
"0.7817",
"0.5913",
"0.2222",
"0.5913"
],
[
"0.2214",
"0.5933",
"0.786",
"0.5933",
"0.786",
"0.607",
"0.2214",
"0.607"
],
[
"0.2214",
"0.6094",
"0.7837",
"0.6094",
"0.7837",
"0.6226",
"0.2214",
"0.6226"
],
[
"0.2197",
"0.6235",
"0.7817",
"0.6235",
"0.7817",
"0.637",
"0.2197",
"0.637"
],
[
"0.2197",
"0.6396",
"0.785",
"0.6396",
"0.785",
"0.653",
"0.2197",
"0.653"
],
[
"0.2197",
"0.654",
"0.785",
"0.654",
"0.785",
"0.6675",
"0.2197",
"0.6675"
],
[
"0.2197",
"0.6685",
"0.7817",
"0.6694",
"0.7817",
"0.683",
"0.2197",
"0.6816"
],
[
"0.2214",
"0.685",
"0.785",
"0.685",
"0.785",
"0.6987",
"0.2214",
"0.6987"
],
[
"0.2214",
"0.702",
"0.6733",
"0.702",
"0.6733",
"0.712",
"0.2214",
"0.712"
],
[
"0.2173",
"0.738",
"0.2646",
"0.738",
"0.2646",
"0.754",
"0.2173",
"0.754"
],
[
"0.255",
"0.7373",
"0.3882",
"0.7393",
"0.3882",
"0.7524",
"0.255",
"0.7505"
],
[
"0.2181",
"0.771",
"0.2688",
"0.771",
"0.2688",
"0.782",
"0.2181",
"0.782"
],
[
"0.2656",
"0.771",
"0.3489",
"0.771",
"0.3489",
"0.782",
"0.2656",
"0.782"
],
[
"0.2197",
"0.7954",
"0.785",
"0.7954",
"0.785",
"0.806",
"0.2197",
"0.806"
],
[
"0.2173",
"0.8047",
"0.7876",
"0.8066",
"0.7876",
"0.829",
"0.2173",
"0.827"
],
[
"0.2181",
"0.8237",
"0.786",
"0.8237",
"0.786",
"0.8413",
"0.2181",
"0.8413"
]
] | [
"2",
"",
"Shi et al.",
"1",
"Related work",
"1.1",
"Representation Learning for Graph Data",
"学习有效的节点或图形表示有利于各种图形挖掘任务。图数据的表示学习方法有",
"几种主要类型，包括基于随机游走的方法E[8,23,25,32]，图神经网络（GNNs)E",
"[7,9,36,37,42,43]和图对比学习[2,56,45,31]。作为图上表示学习的开创性方法",
"基于随机游走的方法优化了节点表示，以便相似上下文中的节点倾向于具有相",
"似的表示。然后，图神经网络成为处理图数据的新的最先进的架构。最后，可以",
"使用图对比学习来进一步完善以GNN为编码器的表示学习过程",
"1.2GraphContrastiveLearning",
"对比学习通过最大化不同增强视图之间的互信息来训练表示212,3.33,37。例",
"如，GRACEE[56]通过丢弃边缘和掩蔽特征来构造负节点示例，并采用类似于",
"SimCLRE[2]的对比模型。GraphCL[51]在图级上对四种数据增强的图进行",
"扰动，并采用NT-Xent损失[2]进行训练。SimGRACEE[45]摆脱了输入数据",
"的增加，并选择扰动模型参数。数据增强对于对比学习在下游任务上实现良好",
"性能至关重要22.34。然而，现有方法主要依赖于随机扰动，这可能会损害",
"增强视图中的图信息。为了解决这个问题，朱等人。，[57]建议使用节点中心性",
"作为重要性指标，并扰动节点中心性较低的节点。Xu等人。[46]提出了一种",
"信息感知表示学习模型[35]】，以在局部和全局视图中保留与任务相关的信息。",
"Li等人。[17]提出了一种图形原理指导的数据增强，以更好地保留语义信息。",
"Youetal。[50]设计一个端到端框架来自动优化数据增强。然而，如何设计适应",
"不同图形/节点样本的数据增强，同时保持相对较低的计算成本仍然是一个挑战",
"1.3Explanation for Graph Neural Networks",
"现有的解释GNN模型的方法可以分为几类E[53]，例如基于替代的，基于相",
"关性的，基于扰动的：基于生成的和基于基本原理的方法。基于替代的方法使",
"用简化但可解释的替代方法来近似原始模型，这些方法要么利用GNN机制E",
"[1,24,41]，要么绕过模型信息E[13,38,54]。基于相关性的方法通过将神经元之",
"间的激活从输出层重新分配到输入层来计算输入贡献[24,27,28]。基于扰动的",
"方法[19,20,26,40,49]估计输人分数，假设删除重要特征将对预测产生很大影响。",
"基于生成的方法产生最大程度激活目标神经元E[52]的合成图。基于基本原理",
"的方法找到一个最好地指导模型预测的小子图作为解释[44,17,18]。现有方法",
"主要集中在有监督图学习模型上，而我们试图获得无监督图学习的解释。同时，",
"我们建议利用解释来改进模型，这进一步要求解释算法是有效的",
"2",
"Preliminaries",
"2.1",
"Notations",
"让G=V.8.A，X是一个图，其中V和&分别表示节点和边的集合。AE",
"[0,1]VIxivi是邻接矩阵。如果节点,和节点,连接，则Ai=1，其中我们",
"使用ei,i表示该边。XERiVIxD是特征矩阵，ai=Xi是节点ui的特征向量。"
] | {
"arxiv_id": "2307.01053",
"language": "zh",
"page_idx": 1
} |
||
2307.01053-zh-2 | 2307.01053-zh-2.jpg | Explanation Guided Data Augmentation for Graph Learning 3
GNN 编码器表示为$f(\cdot)$ ,它将节点转换为表示$\boldsymbol{Z}=f(\boldsymbol{X},A),\boldsymbol{Z}\in\mathbb{R}^{|\mathcal{V}|\times K}$, 其中$K$是潜在维度$,\boldsymbol z_i=Z_i$,表示节点$v_i$的表示。同时，我们还考虑了这项工作中的图级任务。令$\mathcal{G}=\{G_1,G_2,...,G_N\}$表示一组图。在这种情况下，图$G_n$ 的表示形式为$z_n=f(G_n)$ ,其中$\boldsymbol{z}_n\in\mathbb{R}^K$ 。我们提出的方法适用于这两种类型的任务。除非另有说明，否则本文的方法论部分假定使用图级任务进行说明
2.2 Contrastive Learning Frameworks
在这项工作中，我们采用对比学习来学习节点/图表示。我们考虑两种类型的对比学习框架。第一种类型包括 GraphCL [51] 和 GRACE [56] ,它们受SimCLR [2] 的激励，并学会与负面观点相比，最大化正面观点之间的一致性。第二种类型基于 Simsiam$\boxed{\mathbb{E} }[ 3]$ ,可以通过使用停止梯度来消除负样本，以防止模型崩溃。我们修改了 SimCLR 和 Simsiam 框架以适应图形表示学习，并在这两个框架上应用了 ENGAGE
Learning Objectives. 在对比学习中，每个实例都被扩展为两个正视图，其表示形式分别为$z^1$和$z^2$ 。在 SimCLR$\mathbb{E}[2]$中，学习第$i$个实例的表示的损失为：
(1)
$$\ell_i=-\log\frac{\exp\left(\sin\left(\boldsymbol{z}_i^1,\boldsymbol{z}_i^2\right)/\tau\right)}{\sum_j\mathbb{1}_{[j\neq i]}\exp\left(\sin\left(\boldsymbol{z}_i^+,\boldsymbol{z}_j\right)/\tau\right)},$$
,其中$z_i^+$是指$z_i^1$或$z_i^2$ ,而$z_j$表示其他实例的嵌人为否定视图。积极的观点预计将保留原始实例的核心信息，而消极的观点预计将与原始实例无关。$\operatorname{sim}\left(\boldsymbol{z}^1,\boldsymbol{z}^2\right)=\boldsymbol{z}^{1T}\boldsymbol{z}^2/\left\|\boldsymbol{z}^1\right\|\left\|\boldsymbol{z}^2\right\|$。Simsiam [3]模型摆脱了负面观点，其损失函数定义如下：
$$\ell_i=-(z_i^1/\left\|z_i^1\right\|_2)\cdot(z_i^2/\left\|z_i^2\right\|_2).$$
(2)
$\mathbf{Encoder~and~MLP~Heads.~我们使用图形神经网络作为编码器~}f(\cdot)$来学习图形/节点表示$z$。为了全面测试我们提出的框架，我们应用了几种架构，包括用于节点级任务的图卷积网络 (GCN)[E[15] 和图注意力网络 (GAT)[E[36]。此外，我们采用图同构网络(GIN)[E[47]进行图级任务。然后，根据 CL 框架将表示馈送到不同的 MLP 头。在 SimCLR 中，只有一个名为$p_o(\cdot)$的 MLP 头。在Simsiam 中，有两个 MLP 头，即$p_o(\cdot)$和$p_e(\cdot)$
3 The ENGAGE Framework
3.1 Mitigating Superfluous Information in Representations
我们首先介绍最小化多余信息的准则，我们将在整个工作中遵循该淮则，以学习有效且稳健的表示。不失一般性，我们将原始输入和下游任务标签分别表示为$X$和$y$。
$\textbf{Definition 1. }( Sufficiency) \textit{A representation z is defined as sufficient for label}$
$y$ iff $I( X; \boldsymbol{y}) = I( \boldsymbol{z}; \boldsymbol{y})$ . | [
[
"0.3284",
"0.11554",
"0.7363",
"0.11615",
"0.7363",
"0.1307",
"0.3284",
"0.13"
],
[
"0.7295",
"0.11743",
"0.7876",
"0.11743",
"0.7876",
"0.1288",
"0.7295",
"0.1288"
],
[
"0.2181",
"0.1465",
"0.7812",
"0.1445",
"0.7812",
"0.1622",
"0.2181",
"0.1642"
],
[
"0.2181",
"0.1622",
"0.785",
"0.1635",
"0.785",
"0.178",
"0.2181",
"0.1768"
],
[
"0.2214",
"0.178",
"0.7817",
"0.178",
"0.7817",
"0.1913",
"0.2214",
"0.1913"
],
[
"0.2181",
"0.1932",
"0.785",
"0.1925",
"0.785",
"0.207",
"0.2181",
"0.2084"
],
[
"0.2197",
"0.209",
"0.774",
"0.209",
"0.774",
"0.2229",
"0.2197",
"0.2229"
],
[
"0.2181",
"0.2474",
"0.546",
"0.2474",
"0.546",
"0.262",
"0.2181",
"0.262"
],
[
"0.2197",
"0.2747",
"0.7817",
"0.2747",
"0.7817",
"0.2886",
"0.2197",
"0.2886"
],
[
"0.2197",
"0.289",
"0.785",
"0.289",
"0.785",
"0.303",
"0.2197",
"0.303"
],
[
"0.2197",
"0.3062",
"0.7817",
"0.3062",
"0.7817",
"0.3164",
"0.2197",
"0.3164"
],
[
"0.2197",
"0.3193",
"0.7837",
"0.3193",
"0.7837",
"0.3333",
"0.2197",
"0.3333"
],
[
"0.2197",
"0.3352",
"0.7837",
"0.3352",
"0.7837",
"0.3484",
"0.2197",
"0.3484"
],
[
"0.2214",
"0.351",
"0.4307",
"0.351",
"0.4307",
"0.3643",
"0.2214",
"0.3643"
],
[
"0.2255",
"0.3894",
"0.785",
"0.3894",
"0.785",
"0.4028",
"0.2255",
"0.4028"
],
[
"0.2214",
"0.404",
"0.785",
"0.404",
"0.785",
"0.4185",
"0.2214",
"0.4185"
],
[
"0.2197",
"0.4204",
"0.2443",
"0.4204",
"0.2443",
"0.433",
"0.2197",
"0.433"
],
[
"0.3464",
"0.4375",
"0.6562",
"0.4475",
"0.6543",
"0.4797",
"0.3447",
"0.4697"
],
[
"0.464",
"0.4343",
"0.609",
"0.4343",
"0.609",
"0.449",
"0.464",
"0.449"
],
[
"0.763",
"0.443",
"0.786",
"0.443",
"0.786",
"0.4583",
"0.763",
"0.4583"
],
[
"0.2345",
"0.4792",
"0.785",
"0.4792",
"0.785",
"0.4924",
"0.2345",
"0.4924"
],
[
"0.2181",
"0.4937",
"0.785",
"0.4944",
"0.785",
"0.5083",
"0.2181",
"0.507"
],
[
"0.2173",
"0.5083",
"0.786",
"0.509",
"0.786",
"0.5264",
"0.2173",
"0.526"
],
[
"0.2197",
"0.5264",
"0.327",
"0.5264",
"0.327",
"0.5405",
"0.2197",
"0.5405"
],
[
"0.3857",
"0.539",
"0.6167",
"0.5415",
"0.6167",
"0.5596",
"0.3857",
"0.5566"
],
[
"0.7646",
"0.5425",
"0.786",
"0.5425",
"0.786",
"0.5566",
"0.7646",
"0.5566"
],
[
"0.2255",
"0.5806",
"0.7837",
"0.5806",
"0.7837",
"0.5947",
"0.2255",
"0.5947"
],
[
"0.2214",
"0.5967",
"0.7837",
"0.5967",
"0.7837",
"0.61",
"0.2214",
"0.61"
],
[
"0.2214",
"0.6113",
"0.7837",
"0.6113",
"0.7837",
"0.625",
"0.2214",
"0.625"
],
[
"0.2197",
"0.627",
"0.785",
"0.627",
"0.785",
"0.6416",
"0.2197",
"0.6416"
],
[
"0.2214",
"0.6426",
"0.7837",
"0.6426",
"0.7837",
"0.657",
"0.2214",
"0.657"
],
[
"0.2173",
"0.656",
"0.5605",
"0.657",
"0.5605",
"0.672",
"0.2173",
"0.6704"
],
[
"0.2157",
"0.696",
"0.5146",
"0.696",
"0.5146",
"0.7104",
"0.2157",
"0.7104"
],
[
"0.2173",
"0.728",
"0.7183",
"0.73",
"0.7183",
"0.744",
"0.2173",
"0.742"
],
[
"0.2197",
"0.757",
"0.7837",
"0.757",
"0.7837",
"0.771",
"0.2197",
"0.771"
],
[
"0.2214",
"0.7715",
"0.7837",
"0.7715",
"0.7837",
"0.785",
"0.2214",
"0.785"
],
[
"0.2181",
"0.785",
"0.3096",
"0.787",
"0.3096",
"0.8003",
"0.2181",
"0.799"
],
[
"0.2173",
"0.81",
"0.7876",
"0.811",
"0.7876",
"0.8257",
"0.2173",
"0.8247"
],
[
"0.2181",
"0.8276",
"0.3938",
"0.8276",
"0.3938",
"0.8413",
"0.2181",
"0.8413"
]
] | [
"Explanation Guided Data Augmentation for Graph Learning",
"3",
"GNN编码器表示为f(），它将节点转换为表示=f(X,A),ZRIVIxK",
"其中K是潜在维度，之i=Zi：表示节点U的表示。同时，我们还考虑了这项工",
"作中的图级任务。令=Gi,G2...,GN}表示一组图。在这种情况下，图G",
"的表示形式为zn=f(Gn），其中znERK。我们提出的方法适用于这两种类",
"型的任务。除非另有说明，否则本文的方法论部分假定使用图级任务进行说明",
"2.2Contrastive Learning Frameworks",
"在这项工作中，我们采用对比学习来学习节点/图表示。我们考虑两种类型的",
"对比学习框架。第一种类型包括GraphCL[51]和GRACE[56]，它们受",
"SimCLR2的激励，并学会与负面观点相比，最大化正面观点之间的一致性",
"第二种类型基于SimsiamE[3]，可以通过使用停止梯度来消除负样本，以防止",
"模型崩溃。我们修改了SimCLR和Simsiam框架以适应图形表示学习，并在这",
"两个框架上应用了ENGAGE",
"LearningObjectives.在对比学习中，每个实例都被扩展为两个正视图，其",
"表示形式分别为z1和z²。在SimCLRE[2]中，学习第i个实例的表示的损失",
"为：",
"la = -10g 2, u; exp(sim (e4,2) / )",
"exp (sim (zi,z2) /↑)",
"(1)",
"其中是指或，而z，表示其他实例的嵌人为否定视图。积极的",
"观点预计将保留原始实例的核心信息，而消极的观点预计将与原始实例无关。",
"sim（z1,≥2）=≥1Tz2/≥1≥2l。Simsiam[3]模型摆脱了负面观点，其损失",
"函数定义如下：",
"l, = —(2 / 12112) -(22 / 1z2112),",
"(2)",
"EncoderandMLPHeads我们使用图形神经网络作为编码器f(）来学习",
"图形/节点表示2。为了全面测试我们提出的框架，我们应用了几种架构，包括",
"用于节点级任务的图卷积网络（GCN)E[15]和图注意力网络（GAT)E[36]。此",
"外，我们采用图同构网络（GIN)E[47]进行图级任务。然后，根据CL框架将表",
"示馈送到不同的MLP头。在SimCLR中，只有一个名为P。（）的MLP头。在",
"Simsiam中，有两个MLP头，即po(）和pe()",
"3The ENGAGE Framework",
"3.1Mitigating Superfluous Information in Representations",
"我们首先介绍最小化多余信息的准则，我们将在整个工作中遵循该准则，以学",
"习有效且稳健的表示。不失一般性，我们将原始输入和下游任务标签分别表示",
"为X和y。",
"Definition 1. (Sufficiency) A representation z is defined as sufficient for label",
"y iff I(X;y) = I(z;y) :"
] | {
"arxiv_id": "2307.01053",
"language": "zh",
"page_idx": 2
} |
||
2307.01053-zh-3 | 2307.01053-zh-3.jpg | 4
Shi et al.
<FigureHere>
Fig.1: Graph augmentation guided by Smoothed Activation Map (SAM) for
contrastive learning.
图$\mathbb{E}3$ (a)中显示了足够表示的说明。为了获得良好的下游任务性能，鼓励从$X$学习$z$ , 而不会丢失与任务相关的信息。同时，根据信息瓶颈原理$\boxed{\mathbb{E} [ 35] }$ , 通过减少$X$的多余信息，充分表示$z$变得更加健壮并提供更好的性能。这可以通过将$X$和$z$之间的互信息分为两部分来更好地理解，如下所示：
(3)
$$I(X;\boldsymbol{z})=\underbrace{I(X;\boldsymbol{z}\mid\boldsymbol{y})}_{\text{superfluous information}}+\underbrace{I(\boldsymbol{z};\boldsymbol{y})}_{\text{task-relevant information}}.$$
第一项是$z$中对预测$y$无用的信息，应将其最小化。第二项表示预测信息，只
要$z$足以满足$y$ ,它就不受表示的影响
最近的研究表明，对比学习(CL)可以控制表示中保留的信息量[34,39]。但是，如果没有标签信息，则很难指定要在 CL 中保留输入信息的哪一部分。在这项工作中，我们利用解释来识别图中的非平凡结构信息。然后，解释推荐的信息被赋予更高的优先级以保留在表示中，而其他信息更有可能被丢弃。这是通过使用解释而不是随机扰动来指导对比学习中的数据增强来完成的
3.2 $\textbf{Efficient Explanations for Unsupervised Representations}$
Class Activation Map. 解释的目的是确定输入中的关键组件。我们使用
GCN 来说明解释图神经网络的想法。图卷积层定义为$F^l=\sigma(\tilde{\boldsymbol{D}}^{-\frac12}\tilde{\boldsymbol{A}}\tilde{\boldsymbol{D}}^{-\frac12}F^{(l-1)}\tilde{\boldsymbol{W}}^l)$,
,其中 $F^l$ 是第 $l$ 层的输出$,F^0=\boldsymbol{X},\tilde{A}=\boldsymbol{A}+\boldsymbol{I}_{N}$ 是具有自连接的邻接矩阵$,\tilde{D}$ 是$\tilde{A}$的度矩阵。$W^l$表示可训练的权重$,\sigma(\cdot)$是激活函数。在监督学习中，类激活图(CAM)$\mathbb{E}[24,55]$可用于将节点$v_i$的重要性计算为$:\psi_i^c=$ReLU$(\sum_kw_k^cF_{k,i}^L)$, ,其中$L$是最终的卷积层，而$F_k,i^l$是$k$ -th 潜在维度 (i。e。, 通道)节点$v_i$在$l$- th 层。$w_k^c$是通道$k$在输出层对预测类$c$的权重
但是，有两个挑战阻碍了我们直接将 CAM 应用于我们的问题。首先，CAM 是为具有 GCN 和全连接输出层 (由$w_k^c$权重组成)的特定模型体系结构而设计的，其适用性受到限制。其次，CAM 适用于监督模型，而我们专注于无监督学习
${\textbf{Smoothed Activation Map. 为了应对挑战，我们提出了平滑激活图 (SAM)}}$ 来解释没有类标签的表示。关键思想是利用图/节点表示的分布来识别局部重要且可靠的图组件，如图$\overline{\mathrm{E}}??$ 所示。与 CAM 不同，我们的方法适用于不同类型 | [
[
"0.2239",
"0.113",
"0.2345",
"0.1218",
"0.2239",
"0.13",
"0.2133",
"0.1218"
],
[
"0.273",
"0.11615",
"0.3325",
"0.11615",
"0.3325",
"0.1276",
"0.273",
"0.1276"
],
[
"0.357",
"0.1749",
"0.3945",
"0.1749",
"0.3945",
"0.1818",
"0.357",
"0.1818"
],
[
"0.353",
"0.1901",
"0.3635",
"0.1901",
"0.3635",
"0.1945",
"0.353",
"0.1945"
],
[
"0.3628",
"0.1881",
"0.3906",
"0.1881",
"0.3906",
"0.1957",
"0.3628",
"0.1957"
],
[
"0.339",
"0.1995",
"0.3823",
"0.2008",
"0.3823",
"0.209",
"0.339",
"0.2084"
],
[
"0.4805",
"0.207",
"0.5146",
"0.207",
"0.5146",
"0.214",
"0.4805",
"0.214"
],
[
"0.3406",
"0.2361",
"0.411",
"0.2361",
"0.411",
"0.2474",
"0.3406",
"0.2474"
],
[
"0.568",
"0.2311",
"0.601",
"0.2311",
"0.601",
"0.238",
"0.568",
"0.238"
],
[
"0.232",
"0.2374",
"0.29",
"0.2374",
"0.29",
"0.2474",
"0.232",
"0.2474"
],
[
"0.55",
"0.238",
"0.613",
"0.238",
"0.613",
"0.2488",
"0.55",
"0.2488"
],
[
"0.634",
"0.2437",
"0.6567",
"0.2437",
"0.6567",
"0.2468",
"0.634",
"0.2468"
],
[
"0.661",
"0.2437",
"0.6787",
"0.2437",
"0.6787",
"0.2468",
"0.661",
"0.2468"
],
[
"0.2181",
"0.2615",
"0.7837",
"0.26",
"0.7837",
"0.2747",
"0.2181",
"0.276"
],
[
"0.2173",
"0.276",
"0.3628",
"0.276",
"0.3628",
"0.289",
"0.2173",
"0.289"
],
[
"0.2427",
"0.3225",
"0.785",
"0.3225",
"0.785",
"0.336",
"0.2427",
"0.336"
],
[
"0.2181",
"0.3352",
"0.785",
"0.3372",
"0.785",
"0.3516",
"0.2181",
"0.3499"
],
[
"0.2197",
"0.3528",
"0.785",
"0.3528",
"0.785",
"0.3662",
"0.2197",
"0.3662"
],
[
"0.2197",
"0.3687",
"0.694",
"0.3687",
"0.694",
"0.3818",
"0.2197",
"0.3818"
],
[
"0.308",
"0.3945",
"0.3882",
"0.3977",
"0.3865",
"0.4124",
"0.3064",
"0.4092"
],
[
"0.4143",
"0.3945",
"0.4927",
"0.3977",
"0.492",
"0.4136",
"0.4126",
"0.4104"
],
[
"0.5884",
"0.3958",
"0.6396",
"0.3977",
"0.638",
"0.4136",
"0.5864",
"0.411"
],
[
"0.7646",
"0.3965",
"0.786",
"0.3965",
"0.786",
"0.411",
"0.7646",
"0.411"
],
[
"0.5186",
"0.3977",
"0.5474",
"0.3977",
"0.5474",
"0.4104",
"0.5186",
"0.4104"
],
[
"0.5156",
"0.4016",
"0.5215",
"0.4016",
"0.5215",
"0.4072",
"0.5156",
"0.4072"
],
[
"0.38",
"0.4172",
"0.6855",
"0.4167",
"0.6855",
"0.4312",
"0.38",
"0.432"
],
[
"0.2197",
"0.4443",
"0.786",
"0.4443",
"0.786",
"0.4583",
"0.2197",
"0.4583"
],
[
"0.2197",
"0.4602",
"0.5073",
"0.4602",
"0.5073",
"0.4734",
"0.2197",
"0.4734"
],
[
"0.2418",
"0.476",
"0.785",
"0.476",
"0.785",
"0.4893",
"0.2418",
"0.4893"
],
[
"0.2214",
"0.4905",
"0.785",
"0.4905",
"0.785",
"0.504",
"0.2214",
"0.504"
],
[
"0.2197",
"0.505",
"0.7837",
"0.505",
"0.7837",
"0.518",
"0.2197",
"0.518"
],
[
"0.2214",
"0.521",
"0.785",
"0.521",
"0.785",
"0.534",
"0.2214",
"0.534"
],
[
"0.2197",
"0.535",
"0.6914",
"0.535",
"0.6914",
"0.549",
"0.2197",
"0.549"
],
[
"0.2157",
"0.5737",
"0.732",
"0.5747",
"0.732",
"0.589",
"0.2157",
"0.5884"
],
[
"0.2263",
"0.6016",
"0.7837",
"0.603",
"0.7837",
"0.616",
"0.2263",
"0.6157"
],
[
"0.2173",
"0.616",
"0.8203",
"0.6143",
"0.8203",
"0.632",
"0.2173",
"0.634"
],
[
"0.2222",
"0.6333",
"0.785",
"0.6333",
"0.785",
"0.6465",
"0.2222",
"0.6465"
],
[
"0.2181",
"0.6465",
"0.7837",
"0.648",
"0.7837",
"0.661",
"0.2181",
"0.66"
],
[
"0.2181",
"0.661",
"0.7837",
"0.6636",
"0.7837",
"0.678",
"0.2181",
"0.6753"
],
[
"0.2222",
"0.681",
"0.785",
"0.681",
"0.785",
"0.6943",
"0.2222",
"0.6943"
],
[
"0.2173",
"0.6943",
"0.596",
"0.6943",
"0.596",
"0.7085",
"0.2173",
"0.7085"
],
[
"0.246",
"0.712",
"0.785",
"0.712",
"0.785",
"0.7227",
"0.246",
"0.7227"
],
[
"0.2197",
"0.7246",
"0.7837",
"0.726",
"0.7837",
"0.7407",
"0.2197",
"0.7393"
],
[
"0.2197",
"0.741",
"0.7837",
"0.741",
"0.7837",
"0.755",
"0.2197",
"0.755"
],
[
"0.2214",
"0.7583",
"0.2361",
"0.7583",
"0.2361",
"0.768",
"0.2214",
"0.768"
],
[
"0.2239",
"0.7954",
"0.786",
"0.7954",
"0.786",
"0.8096",
"0.2239",
"0.8096"
],
[
"0.2214",
"0.811",
"0.7837",
"0.811",
"0.7837",
"0.8247",
"0.2214",
"0.8247"
],
[
"0.2197",
"0.8257",
"0.785",
"0.8257",
"0.785",
"0.8403",
"0.2197",
"0.8403"
]
] | [
"N",
"Shi et al.",
"",
"",
"刻器",
"",
"",
"NodeEmbeddings",
"Smoothed",
"Input Graph(s)",
"Node Importane",
"no",
"Prott",
"Fig.1: Graph augmentation guided by Smoothed Activation Map (SAM) fo",
"contrastive learning",
"图E3（a）中显示了足够表示的说明。为了获得良好的下游任务性能，鼓励",
"从X学习之，而不会丢失与任务相关的信息。同时，根据信息瓶颈原理E[35]",
"通过减少X的多余信息，充分表示z变得更加健壮并提供更好的性能。这可以",
"通过将X和z之间的互信息分为两部分来更好地理解，如下所示：",
"I(X;z) =",
"I(X;z/ y)",
"I(z;y)",
"(3)",
"+",
"",
"superfuous informationtask-relevant information",
"第一项是2中对预测9无用的信息，应将其最小化。第二项表示预测信息，只",
"要之足以满足y，它就不受表示的影响",
"最近的研究表明，对比学习（CL）可以控制表示中保留的信息量[34,39]。",
"但是，如果没有标签信息，则很难指定要在CL中保留输人信息的哪一部分。在",
"这项工作中，我们利用解释来识别图中的非平凡结构信息。然后，解释推荐的信",
"息被赋予更高的优先级以保留在表示中，而其他信息更有可能被丢弃。这是通",
"过使用解释而不是随机扰动来指导对比学习中的数据增强来完成的",
"3.2Efficient Explanations for Unsupervised Representations",
"ClassActivationMap.解释的目的是确定输人中的关键组件。我们使用",
"GCN来说明解释图神经网络的想法。图卷积层定义为FI=(D-AD-F(-1)W),",
"，其中F是第l层的输出，FO=X，A=A+IN是具有自连接的邻接矩阵，D",
"是A的度矩阵。W表示可训练的权重，α（）是激活函数。在监督学习中，类激活",
"图（CAM)E[24,55]可用于将节点i的重要性计算为：s=ReLU(ZwFk)",
"，其中L是最终的卷积层，而F是k-th潜在维度（i。e。，通道）节点vi在",
"1-th层。w是通道k在输出层对预测类c的权重",
"但是，有两个挑战阻碍了我们直接将CAM应用于我们的问题。首先，CAM",
"是为具有GCN和全连接输出层（由w权重组成）的特定模型体系结构而设计",
"的，其适用性受到限制。其次，CAM适用于监督模型，而我们专注于无监督学",
"习",
"SmoothedActivationMap.为了应对挑战，我们提出了平滑激活图（SAM）",
"来解释没有类标签的表示。关键思想是利用图/节点表示的分布来识别局部重要",
"且可靠的图组件，如图E？？所示。与CAM不同，我们的方法适用于不同类型"
] | {
"arxiv_id": "2307.01053",
"language": "zh",
"page_idx": 3
} |
||
2307.01053-zh-4 | 2307.01053-zh-4.jpg | Explanation Guided Data Augmentation for Graph Learning 5
的 GNN 模型，不需要监督信号。不失一般性，我们将 GNN 的前馈过程编写如
$\dot{\text{下}}:$
$$\boldsymbol{a}_{i}^{l}=\text{AGGREGATION}^{l}(\{F_{i'}^{(l-1)}:n'\in\mathcal{N}_{i}\}),\:F_{i}^{l}=\text{COMBINE}^{l}(F_{i}^{(l-1)},\boldsymbol{a}_{i}^{l}),$$
(4)
其中$F_i^l$是$v_i$在$l$层的嵌入。$\mathcal{N}_i$表示$v_i$的邻居。然后，SAM 热图计算为：
$$\psi_i=\mathrm{ReLU}(\sum_k\:\tilde{w}_kF_{k,i}^L),$$
(5)
其中$\psi_i$解释了$v_i$的重要性得分，$\tilde{w}_k$是通道$k$的平滑重要性，而$L$是最后一层。对于图级任务，图$G_n$的通道重要性是根据其本地上下文估计的，因此$\begin{array}{l}{\tilde{w}_{k}^{graph(n)}=\mathrm{Norm}(\sum_{n^{\prime}\in\tilde{\mathcal{N}}_{n}}\operatorname{Pool}\{F_{i^{\prime}}^{L}:i^{\prime}\in G_{n^{\prime}}\})[k],\text{其中 }\tilde{\mathcal{N}}_{n}\text{ 表示嵌人空间中}}\\{G_{n}\text{ 周围的邻居图集。Pool() 操作从其节点嵌入中生成 }G_{n^{\prime}}\text{ 的热图，我们在实}}\end{array}$ 验中使用平均池化。Norm()操作意味着对热图执行$L_2$归一化。通过使用量化技术$\mathbb{E}[14]$ ,可以有效地检索邻居的集合 $\tilde{\mathcal{N}}_n$ 。对于节点级任务，通过平均附近节点嵌入的热图来估计信道重要性，即。e。,$\tilde{w}_{k}^{node(i)}=$Norm$( \sum _{i^{\prime }\in \tilde{\mathcal{N} } _{i}}F_{i^{\prime }}^{L}) [ k] $, , 其中$\tilde{\mathcal{N}}_i$是嵌人空间中靠近$v_i$的邻居集。解释过程如图 1 所示
$\tilde{w}_k^N$和$\tilde{w}_k^G$都是在无监督学习中估计
的，它们在$\overset{.}{\operatorname*{CAM}}$中的作用与$w_k^c$相似。通
过考虑附近的节点和图，平滑目标节点/图
的解释信息。理解 SAM 的另一个观点是，
它不仅为单个节点或图提供了解释，而且
还为局部流形区域中的一组节点或图提供
了解释
最后，针对图神经网络提出了许多解释
方法 (e。g。、基于扰动的、基于替代的和
基于生成的方法),所以我们想证明我们在
这项工作中选择 CAM 风格的解释是合理
的。主要考虑因素是计算效率。CAM 式方
法中的热图可在 GNN 的前馈过程中直接
获得，只需最少的额外计算。相比之下，基
于扰动的方法和基于替代的方法都需要发
送许多扰动输人来估计不同图特征的影响，
这导致明显更大的时间成本。基于生成的
方法面临类似的问题，因为它们需要在大
量图样本上训练全局解释器
### 3.3 Explanation-Guided
Contrastive Views Generation
我们利用解释结果设计了两个数据增强操作，其中解释突出显示的图形组件被认为是重要的$\mathbb{E}[24,22]$。总体思路是保持重要信息不变，同时最大限度地扰乱不重要的信息 | [
[
"0.327",
"0.11554",
"0.7363",
"0.11615",
"0.7363",
"0.13",
"0.327",
"0.1288"
],
[
"0.7295",
"0.11743",
"0.786",
"0.11743",
"0.786",
"0.1288",
"0.7295",
"0.1288"
],
[
"0.2197",
"0.1483",
"0.785",
"0.1483",
"0.785",
"0.1622",
"0.2197",
"0.1622"
],
[
"0.2214",
"0.1642",
"0.2443",
"0.1642",
"0.2443",
"0.178",
"0.2214",
"0.178"
],
[
"0.2336",
"0.1895",
"0.7686",
"0.1895",
"0.7686",
"0.2102",
"0.2336",
"0.2102"
],
[
"0.7656",
"0.2108",
"0.7817",
"0.2108",
"0.7817",
"0.2203",
"0.7656",
"0.2203"
],
[
"0.2197",
"0.2235",
"0.759",
"0.2235",
"0.759",
"0.2374",
"0.2197",
"0.2374"
],
[
"0.415",
"0.2524",
"0.589",
"0.2524",
"0.589",
"0.2676",
"0.415",
"0.2676"
],
[
"0.763",
"0.2524",
"0.785",
"0.2524",
"0.785",
"0.2664",
"0.763",
"0.2664"
],
[
"0.2181",
"0.2822",
"0.785",
"0.2822",
"0.785",
"0.2954",
"0.2181",
"0.2954"
],
[
"0.2222",
"0.2966",
"0.785",
"0.2966",
"0.785",
"0.31",
"0.2222",
"0.31"
],
[
"0.2157",
"0.305",
"0.7876",
"0.31",
"0.7876",
"0.3352",
"0.2157",
"0.33"
],
[
"0.2197",
"0.3276",
"0.785",
"0.3276",
"0.785",
"0.3416",
"0.2197",
"0.3416"
],
[
"0.2197",
"0.3435",
"0.7837",
"0.3435",
"0.7837",
"0.3567",
"0.2197",
"0.3567"
],
[
"0.2197",
"0.3591",
"0.785",
"0.3591",
"0.785",
"0.3726",
"0.2197",
"0.3726"
],
[
"0.2141",
"0.3655",
"0.7915",
"0.3694",
"0.7915",
"0.3965",
"0.2141",
"0.3926"
],
[
"0.2197",
"0.3945",
"0.6724",
"0.3945",
"0.6724",
"0.4077",
"0.2197",
"0.4077"
],
[
"0.2443",
"0.4104",
"0.5186",
"0.4104",
"0.5186",
"0.4236",
"0.2443",
"0.4236"
],
[
"0.2197",
"0.4248",
"0.5195",
"0.4248",
"0.5195",
"0.4382",
"0.2197",
"0.4382"
],
[
"0.2197",
"0.44",
"0.5186",
"0.44",
"0.5186",
"0.4539",
"0.2197",
"0.4539"
],
[
"0.2181",
"0.4539",
"0.517",
"0.455",
"0.517",
"0.4697",
"0.2181",
"0.4685"
],
[
"0.2214",
"0.4717",
"0.5156",
"0.4717",
"0.5156",
"0.4817",
"0.2214",
"0.4817"
],
[
"0.2197",
"0.4849",
"0.5215",
"0.4849",
"0.5215",
"0.4988",
"0.2197",
"0.4988"
],
[
"0.2214",
"0.5",
"0.2695",
"0.5",
"0.2695",
"0.5146",
"0.2214",
"0.5146"
],
[
"0.2443",
"0.5166",
"0.5215",
"0.5166",
"0.5215",
"0.53",
"0.2443",
"0.53"
],
[
"0.2197",
"0.53",
"0.5215",
"0.531",
"0.5215",
"0.5454",
"0.2197",
"0.5444"
],
[
"0.2214",
"0.547",
"0.5215",
"0.547",
"0.5215",
"0.56",
"0.2214",
"0.56"
],
[
"0.2197",
"0.561",
"0.5215",
"0.561",
"0.5215",
"0.5747",
"0.2197",
"0.5747"
],
[
"0.2197",
"0.577",
"0.5195",
"0.577",
"0.5195",
"0.5903",
"0.2197",
"0.5903"
],
[
"0.2214",
"0.5913",
"0.5195",
"0.5913",
"0.5195",
"0.605",
"0.2214",
"0.605"
],
[
"0.2197",
"0.607",
"0.5215",
"0.607",
"0.5215",
"0.6206",
"0.2197",
"0.6206"
],
[
"0.2214",
"0.6206",
"0.5186",
"0.6206",
"0.5186",
"0.634",
"0.2214",
"0.634"
],
[
"0.2197",
"0.636",
"0.5186",
"0.636",
"0.5186",
"0.6494",
"0.2197",
"0.6494"
],
[
"0.2197",
"0.6514",
"0.5195",
"0.6514",
"0.5195",
"0.6655",
"0.2197",
"0.6655"
],
[
"0.2214",
"0.6675",
"0.5215",
"0.6675",
"0.5215",
"0.681",
"0.2214",
"0.681"
],
[
"0.2197",
"0.6816",
"0.415",
"0.6816",
"0.415",
"0.696",
"0.2197",
"0.696"
],
[
"0.2197",
"0.7236",
"0.2688",
"0.7236",
"0.2688",
"0.735",
"0.2197",
"0.735"
],
[
"0.267",
"0.7227",
"0.4355",
"0.7227",
"0.4355",
"0.7363",
"0.267",
"0.7363"
],
[
"0.2197",
"0.738",
"0.468",
"0.738",
"0.468",
"0.752",
"0.2197",
"0.752"
],
[
"0.2214",
"0.7676",
"0.5186",
"0.7676",
"0.5186",
"0.778",
"0.2214",
"0.778"
],
[
"0.2181",
"0.78",
"0.5186",
"0.78",
"0.5186",
"0.7935",
"0.2181",
"0.7935"
],
[
"0.2197",
"0.797",
"0.5195",
"0.797",
"0.5195",
"0.81",
"0.2197",
"0.81"
],
[
"0.2197",
"0.811",
"0.5186",
"0.811",
"0.5186",
"0.8247",
"0.2197",
"0.8247"
],
[
"0.2197",
"0.827",
"0.2524",
"0.827",
"0.2524",
"0.839",
"0.2197",
"0.839"
]
] | [
"Explanation Guided Data Augmentation for Graph Learning",
"5",
"的GNN模型，不需要监督信号。不失一般性，我们将GNN的前馈过程编写如",
"下：",
"a’ = AGGREGATION((F(l-1) : n' E Na), FI = COMBINE'(F(l-1),a),",
"(4",
"其中F是Ui在l层的嵌入。N表示Ui的邻居。然后，SAM热图计算为：",
"bi= ReLU(k wkFk,i),",
"(5)",
"其中解释了i的重要性得分，wk是通道k的平滑重要性，而L是最后",
"一层。对于图级任务，图Gn的通道重要性是根据其本地上下文估计的，因此",
"wraphon)=Norm(enPool(Fh：G)al，其中N,表示嵌人空间中",
"Gn周围的邻居图集。PoolO）操作从其节点嵌人中生成Gn\"的热图，我们在实",
"验中使用平均池化。NormO操作意味着对热图执行L2归一化。通过使用量化",
"技术E[14]，可以有效地检索邻居的集合Nn。对于节点级任务，通过平均附近",
"节点嵌人的热图来估计信道重要性，即。e。，wrode()）=Norm(Zvex，F)[6]，",
"其中N，是嵌入空间中靠近的邻居集。解释过程如图1所示",
"WN和wG都是在无监督学习中估计",
"的，它们在CAM中的作用与w相似。通",
"过考虑附近的节点和图，平滑目标节点/图",
"的解释信息。理解SAM的另一个观点是",
"它不仅为单个节点或图提供了解释，而目",
"还为局部流形区域中的一组节点或图提供",
"了解释",
"最后，针对图神经网络提出了许多解释",
"方法（e。g。、基于扰动的、基于替代的和",
"基于生成的方法），所以我们想证明我们在",
"这项工作中选择CAM风格的解释是合理",
"的。主要考虑因素是计算效率。CAM式方",
"法中的热图可在GNN的前馈过程中直接",
"获得，只需最少的额外计算。相比之下，基",
"于扰动的方法和基于替代的方法都需要发",
"送许多扰动输入来估计不同图特征的影响",
"这导致明显更大的时间成本。基于生成的",
"方法面临类似的问题，因为它们需要在大",
"量图样本上训练全局解释器",
"3.3",
"Explanation-Guided",
"Contrastive Views Generation",
"我们利用解释结果设计了两个数据增强操",
"作，其中解释突出显示的图形组件被认为",
"是重要的E[24,22]。总体思路是保持重要",
"信息不变，同时最大限度地扰乱不重要的",
"信息"
] | {
"arxiv_id": "2307.01053",
"language": "zh",
"page_idx": 4
} |
||
1708.05942-zh-0 | 1708.05942-zh-0.jpg | # 赫尔辛基神经机器翻译系统$^*$
Robert Östling
Department of Linguistics
Stockholm University
Yves Scherrer and Jörg Tiedemann Department of Modern Languages
University of Helsinki
Tommi Nieminen
Gongbo Tang
Department of Linguistics and Philology Department of Modern Languages
Uppsala University
University of Helsinki
Abstract
的字节对编码(BPE)技术，该技术已成功用于芬兰语以前的 (Sánchez-Cartagena and Toral, 2016)。由于 BPE 可以作为简单的预处理和后处理步骤添加，因此它不会影响翻译模型的结构。这意味着我们的系统可以在目标端使用字符，单词和 BPE 级别生成。因此，网络的结构由三个长短期内存(LSTM)(Hochreiter and Schmidhuber, 1997) 层组成：
We introduce the Helsinki Neural Machine Translation system (HNMT) and how it is applied in the news translation task at WMT 2017, where it ranked first in both the human and automatic evaluations for English-Finnish. We discuss the success of English-Finnish translations and the overall advantage of NMT over a strong SMT baseline. We also discuss our submissions for English- Latvian, English- Chinese and Chinese- English.
1. A character-level encoder that transforms
out-of-vocabulary source tokens into the same vector space as the source word embeddings.
## 1 Introduction
2. A token-level bidirectional encoder that
transforms a sequence of source word embeddings (or outputs from network (1) in case of OOV items) into an encoded sequence of the same length.
3. A character-, token- or BPE-level decoder
that works as language model conditioned (via an attention mechanism) on the encoded source sequence from (2).
赫尔辛基神经机器翻译系统(HNMT)是一个功能齐全的神经机器翻译系统，特别关注形态丰富的语言。我们参加了 WMT2017 新闻翻译共享任务，获得了英语-芬兰语翻译的最高BLEU 分数，同时在英语-拉脱维亚语上也表现出色，在英语-中文和中文-英语上也表现良好
除了我们参与共享任务外，本文还详细介绍了我们与 HNMT 一起实施和评估的其他一些方法，其中许多方法产生了负面结果，随后未在我们提交的共享任务中使用
HNMT 使用 Theano (Theano Development Team,2016)实现，它支持使用 GPU 进行有效训练。为了优化，我们使用 Adam (Kingma and Ba, 2015) 的小批量随机梯度下降来学习速率自适应
## 1.1 HNMT
2 Tricks from the NMT arsenal
我们已经实现并评估了许多对基本注意力编码器-解码器模型的扩展建议。除非另有说明，否则基本实验是在英语-芬兰语数据上进行的
由于 Bahdanau et al. (2014), HNMT 基于注意力编码器-解码器模型。这是 NMT 的一个相当简约的框架，并且已经提出了许多扩展。特别感兴趣的是那些允许适当和有效地处理形态丰富的语言，如芬兰语。我们结合了两种这样的方法：用于源语言编码器的 Luong and Manning (2016) 的混合字符/单词模型，以及用于目标语言解码器的 Sennrich et al. (2016c)
### 2.1 Layer normalization
层归一化 (Ba et al., 2016) 已被提出作为加速递归网络模型训练的技术。我们已经将其实现到 HNMT 中，作为由 Ba et al. (2016)中的等式 (20)- (22) 描述的修改后的 LSTM。然而，
$^*$The software is available from https://github com/robertostling/hnmt under the GNU General Public License version 3. | [
[
"0.3552",
"0.0849",
"0.657",
"0.0849",
"0.657",
"0.0998",
"0.3552",
"0.0998"
],
[
"0.2703",
"0.114",
"0.4048",
"0.1164",
"0.4048",
"0.1342",
"0.2703",
"0.1318"
],
[
"0.513",
"0.11755",
"0.8228",
"0.11816",
"0.8228",
"0.133",
"0.513",
"0.1318"
],
[
"0.2225",
"0.1342",
"0.451",
"0.1354",
"0.451",
"0.1497",
"0.2225",
"0.1484"
],
[
"0.519",
"0.1354",
"0.818",
"0.1354",
"0.818",
"0.1497",
"0.519",
"0.1497"
],
[
"0.2418",
"0.1509",
"0.43",
"0.152",
"0.43",
"0.1663",
"0.2418",
"0.165"
],
[
"0.575",
"0.152",
"0.763",
"0.152",
"0.763",
"0.1663",
"0.575",
"0.1663"
],
[
"0.2737",
"0.1882",
"0.3938",
"0.1907",
"0.3938",
"0.2048",
"0.2737",
"0.2031"
],
[
"0.594",
"0.1907",
"0.742",
"0.1907",
"0.742",
"0.2037",
"0.594",
"0.2037"
],
[
"0.1561",
"0.2037",
"0.818",
"0.2048",
"0.818",
"0.2227",
"0.1561",
"0.2214"
],
[
"0.251",
"0.2227",
"0.4165",
"0.2239",
"0.4165",
"0.2382",
"0.251",
"0.2369"
],
[
"0.574",
"0.2239",
"0.763",
"0.2239",
"0.763",
"0.2369",
"0.574",
"0.2369"
],
[
"0.266",
"0.2654",
"0.3435",
"0.2654",
"0.3435",
"0.2798",
"0.266",
"0.2798"
],
[
"0.5156",
"0.2666",
"0.881",
"0.2666",
"0.881",
"0.2798",
"0.5156",
"0.2798"
],
[
"0.5146",
"0.2808",
"0.881",
"0.282",
"0.881",
"0.2964",
"0.5146",
"0.2957"
],
[
"0.1486",
"0.2944",
"0.4568",
"0.2944",
"0.4568",
"0.308",
"0.1486",
"0.308"
],
[
"0.5146",
"0.2986",
"0.882",
"0.2986",
"0.882",
"0.3123",
"0.5146",
"0.3123"
],
[
"0.1486",
"0.3123",
"0.4592",
"0.3123",
"0.4592",
"0.326",
"0.1486",
"0.326"
],
[
"0.5156",
"0.3154",
"0.879",
"0.3154",
"0.879",
"0.326",
"0.5156",
"0.326"
],
[
"0.147",
"0.328",
"0.4575",
"0.328",
"0.4575",
"0.3416",
"0.147",
"0.3416"
],
[
"0.513",
"0.33",
"0.881",
"0.33",
"0.881",
"0.3433",
"0.513",
"0.3433"
],
[
"0.147",
"0.3433",
"0.4592",
"0.3433",
"0.4592",
"0.357",
"0.147",
"0.357"
],
[
"0.5146",
"0.3457",
"0.882",
"0.3467",
"0.882",
"0.361",
"0.5146",
"0.3599"
],
[
"0.147",
"0.3599",
"0.4592",
"0.3599",
"0.4592",
"0.3735",
"0.147",
"0.3735"
],
[
"0.5146",
"0.3635",
"0.882",
"0.3635",
"0.882",
"0.3772",
"0.5146",
"0.3772"
],
[
"0.1486",
"0.376",
"0.4592",
"0.376",
"0.4592",
"0.3896",
"0.1486",
"0.3896"
],
[
"0.513",
"0.379",
"0.7773",
"0.379",
"0.7773",
"0.3926",
"0.513",
"0.3926"
],
[
"0.1486",
"0.3926",
"0.4592",
"0.3926",
"0.4592",
"0.4062",
"0.1486",
"0.4062"
],
[
"0.1453",
"0.4067",
"0.4592",
"0.4092",
"0.4592",
"0.4233",
"0.1453",
"0.4216"
],
[
"0.5283",
"0.408",
"0.879",
"0.408",
"0.879",
"0.4187",
"0.5283",
"0.4187"
],
[
"0.1486",
"0.4246",
"0.4575",
"0.4246",
"0.4575",
"0.4382",
"0.1486",
"0.4382"
],
[
"0.552",
"0.4229",
"0.8833",
"0.4229",
"0.8833",
"0.437",
"0.552",
"0.437"
],
[
"0.1486",
"0.4407",
"0.4575",
"0.4407",
"0.4575",
"0.4536",
"0.1486",
"0.4536"
],
[
"0.552",
"0.4395",
"0.881",
"0.4395",
"0.881",
"0.4536",
"0.552",
"0.4536"
],
[
"0.1486",
"0.4573",
"0.4592",
"0.4573",
"0.4592",
"0.4702",
"0.1486",
"0.4702"
],
[
"0.55",
"0.4548",
"0.629",
"0.4548",
"0.629",
"0.4697",
"0.55",
"0.4697"
],
[
"0.1486",
"0.4727",
"0.2854",
"0.4727",
"0.2854",
"0.4863",
"0.1486",
"0.4863"
],
[
"0.525",
"0.4822",
"0.881",
"0.4822",
"0.881",
"0.4954",
"0.525",
"0.4954"
],
[
"0.11755",
"0.503",
"0.157",
"0.503",
"0.157",
"0.514",
"0.11755",
"0.514"
],
[
"0.1511",
"0.503",
"0.266",
"0.503",
"0.266",
"0.5156",
"0.1511",
"0.5156"
],
[
"0.552",
"0.4988",
"0.879",
"0.4988",
"0.879",
"0.5127",
"0.552",
"0.5127"
],
[
"0.552",
"0.5156",
"0.882",
"0.5156",
"0.882",
"0.5293",
"0.552",
"0.5293"
],
[
"0.12006",
"0.527",
"0.4878",
"0.527",
"0.4878",
"0.5405",
"0.12006",
"0.5405"
],
[
"0.552",
"0.53",
"0.881",
"0.53",
"0.881",
"0.5444",
"0.552",
"0.5444"
],
[
"0.12006",
"0.5435",
"0.4878",
"0.5435",
"0.4878",
"0.557",
"0.12006",
"0.557"
],
[
"0.552",
"0.5464",
"0.7676",
"0.5464",
"0.7676",
"0.561",
"0.552",
"0.561"
],
[
"0.1184",
"0.5586",
"0.486",
"0.558",
"0.486",
"0.5723",
"0.1184",
"0.5737"
],
[
"0.12006",
"0.576",
"0.4878",
"0.576",
"0.4878",
"0.589",
"0.12006",
"0.589"
],
[
"0.5283",
"0.5747",
"0.882",
"0.5747",
"0.882",
"0.5884",
"0.5283",
"0.5884"
],
[
"0.12006",
"0.5913",
"0.4878",
"0.5913",
"0.4878",
"0.605",
"0.12006",
"0.605"
],
[
"0.55",
"0.589",
"0.882",
"0.589",
"0.882",
"0.6025",
"0.55",
"0.6025"
],
[
"0.12006",
"0.608",
"0.486",
"0.608",
"0.486",
"0.6216",
"0.12006",
"0.6216"
],
[
"0.552",
"0.605",
"0.881",
"0.6055",
"0.881",
"0.6206",
"0.552",
"0.619"
],
[
"0.12177",
"0.6245",
"0.1377",
"0.6245",
"0.1377",
"0.6353",
"0.12177",
"0.6353"
],
[
"0.552",
"0.6226",
"0.8096",
"0.6226",
"0.8096",
"0.6357",
"0.552",
"0.6357"
],
[
"0.1377",
"0.64",
"0.4878",
"0.64",
"0.4878",
"0.654",
"0.1377",
"0.654"
],
[
"0.513",
"0.648",
"0.882",
"0.6484",
"0.882",
"0.663",
"0.513",
"0.662"
],
[
"0.1234",
"0.6562",
"0.4895",
"0.654",
"0.4895",
"0.6685",
"0.1234",
"0.6704"
],
[
"0.5146",
"0.665",
"0.881",
"0.665",
"0.881",
"0.6787",
"0.5146",
"0.6787"
],
[
"0.12006",
"0.673",
"0.4878",
"0.673",
"0.4878",
"0.686",
"0.12006",
"0.686"
],
[
"0.5156",
"0.6816",
"0.881",
"0.6816",
"0.881",
"0.6953",
"0.5156",
"0.6953"
],
[
"0.12006",
"0.6885",
"0.3594",
"0.6885",
"0.3594",
"0.7017",
"0.12006",
"0.7017"
],
[
"0.513",
"0.6978",
"0.879",
"0.6978",
"0.879",
"0.7114",
"0.513",
"0.7114"
],
[
"0.12006",
"0.7163",
"0.2258",
"0.7163",
"0.2258",
"0.728",
"0.12006",
"0.728"
],
[
"0.5156",
"0.7153",
"0.608",
"0.7153",
"0.608",
"0.727",
"0.5156",
"0.727"
],
[
"0.1184",
"0.7354",
"0.486",
"0.7354",
"0.486",
"0.749",
"0.1184",
"0.749"
],
[
"0.5156",
"0.7437",
"0.529",
"0.7437",
"0.529",
"0.753",
"0.5156",
"0.753"
],
[
"0.5366",
"0.745",
"0.5474",
"0.745",
"0.5474",
"0.7505",
"0.5366",
"0.7505"
],
[
"0.5483",
"0.741",
"0.804",
"0.741",
"0.804",
"0.755",
"0.5483",
"0.755"
],
[
"0.12006",
"0.752",
"0.4878",
"0.752",
"0.4878",
"0.7656",
"0.12006",
"0.7656"
],
[
"0.12006",
"0.7686",
"0.4895",
"0.7686",
"0.4895",
"0.782",
"0.12006",
"0.782"
],
[
"0.5146",
"0.767",
"0.882",
"0.767",
"0.882",
"0.781",
"0.5146",
"0.781"
],
[
"0.12177",
"0.786",
"0.486",
"0.786",
"0.486",
"0.7964",
"0.12177",
"0.7964"
],
[
"0.5146",
"0.782",
"0.881",
"0.782",
"0.881",
"0.796",
"0.5146",
"0.796"
],
[
"0.12006",
"0.8003",
"0.486",
"0.8003",
"0.486",
"0.814",
"0.12006",
"0.814"
],
[
"0.517",
"0.8003",
"0.8506",
"0.8003",
"0.8506",
"0.811",
"0.517",
"0.811"
],
[
"0.12177",
"0.8184",
"0.486",
"0.8184",
"0.486",
"0.829",
"0.12177",
"0.829"
],
[
"0.513",
"0.826",
"0.7256",
"0.826",
"0.7256",
"0.839",
"0.513",
"0.839"
],
[
"0.12006",
"0.833",
"0.4878",
"0.833",
"0.4878",
"0.8467",
"0.12006",
"0.8467"
],
[
"0.12177",
"0.8496",
"0.4878",
"0.8496",
"0.4878",
"0.8633",
"0.12177",
"0.8633"
],
[
"0.5146",
"0.8457",
"0.882",
"0.8467",
"0.882",
"0.861",
"0.5146",
"0.86"
],
[
"0.5156",
"0.8633",
"0.882",
"0.8633",
"0.882",
"0.877",
"0.5156",
"0.877"
],
[
"0.1486",
"0.8726",
"0.4844",
"0.8726",
"0.4844",
"0.8857",
"0.1486",
"0.8857"
],
[
"0.1184",
"0.8843",
"0.486",
"0.8843",
"0.486",
"0.898",
"0.1184",
"0.898"
],
[
"0.5156",
"0.879",
"0.882",
"0.879",
"0.882",
"0.8926",
"0.5156",
"0.8926"
],
[
"0.12177",
"0.898",
"0.2612",
"0.898",
"0.2612",
"0.9077",
"0.12177",
"0.9077"
],
[
"0.5156",
"0.8955",
"0.8833",
"0.8955",
"0.8833",
"0.909",
"0.5156",
"0.909"
]
] | [
"赫尔辛基神经机器翻译系统*",
"Robert Ostling",
"Yves Scherrer and Jorg Tiedemann",
"Department of Linguistics",
"Department of Modern Languages",
"Stockholm University",
"University of Helsinki",
"Gongbo Tang",
"Tommi Nieminen",
"Department of Linguistics and Philology Department of Modern Languages",
"Uppsala University",
"University of Helsinki",
"Abstract",
"的字节对编码（BPE）技术，该技术已成功用于",
"芬兰语以前的（Sanchez-Cartagena andToral，",
"We introduce the Helsinki Neural Ma-",
"2016）。由于BPE可以作为简单的预处理和",
"chine Translation system (HNMT) and",
"后处理步骤添加，因此它不会影响翻译模型的",
"how it is applied in the news translation",
"结构。这意味着我们的系统可以在目标端使用",
"task at WMT 2017, where it ranked",
"字符，单词和BPE级别生成。因此，网络的",
"first in both the human and automatic",
"结构由三个长短期内存（LSTM）（Hochreiter",
"evaluations for English-Finnish.We",
"andSchmidhuber，1997）层组成：",
"discuss the success of English-Finnish",
"translations and the overall advantage",
"1. A character-level encoder that transforms",
"of NMT over a strong SMT baseline.",
"out-of-vocabulary source tokens into the",
"We also discuss our submissions for",
"same vector space as the source word em-",
"English-Latvian, English-Chinese and",
"beddings.",
"Chinese-English.",
"2. A token-level bidirectional encoder that",
"1",
"Introduction",
"transforms a sequence of source word em",
"beddings (or outputs from network (1) in",
"赫尔辛基神经机器翻译系统（HNMT）是一个",
"case of OOV items) into an encoded se-",
"功能齐全的神经机器翻译系统，特别关注形态",
"quence of the same length.",
"丰富的语言。我们参加了WMT2017新闻翻",
"译共享任务，获得了英语芬兰语翻译的最高",
"3. A character-, token- or BPE-level decoder",
"BLEU分数，同时在英语-拉脱维亚语上也表",
"that works as language model conditioned",
"现出色，在英语-中文和中文-英语上也表现良",
"(via an attention mechanism) on the en-",
"好",
"coded source sequence from (2).",
"除了我们参与共享任务外，本文还详细介绍",
"HNMT使用Theano（Theano Development",
"了我们与HNMT一起实施和评估的其他一些",
"Team，2016）实现，它支持使用GPU进行有",
"方法，其中许多方法产生了负面结果，随后未",
"效训练。为了优化，我们使用Adam（Kingma",
"在我们提交的共享任务中使用",
"andBa，2015）的小批量随机梯度下降来学习",
"1.1HNMT",
"速率自适应",
"由于Bahdanau et al.（2014），HNMT基于注",
"2",
"",
"Tricks from the NMT arsenal",
"意力编码器-解码器模型。这是NMT的一个",
"相当简约的框架，并且已经提出了许多扩展。",
"我们已经实现并评估了许多对基本注意力编码",
"特别感兴趣的是那些充许适当和有效地处理",
"器-解码器模型的扩展建议。除非另有说明，否",
"形态丰富的语言，如芬兰语。我们结合了两种",
"则基本实验是在英语-芬兰语数据上进行的",
"这样的方法：用于源语言编码器的Luongand",
"2.1Layer normalization",
"Manning（2016）的混合字符/单词模型，以及",
"用于目标语言解码器的Sennrichetal:（2016c)",
"层归一化（Baetal.，2016）已被提出作为加速",
"递归网络模型训练的技术。我们已经将其实现",
"*The software is available from https://github",
"com/robertostling/hnmt under the GNU General Pub-",
"到HNMT中，作为由Baetal.（2016）中的等",
"lic License version 3.",
"式（20）-（22）描述的修改后的LSTM。然而，"
] | {
"arxiv_id": "1708.05942",
"language": "zh",
"page_idx": 0
} |
||
1708.05942-zh-1 | 1708.05942-zh-1.jpg | 由于初步实验没有表明使用层归一化的任何一致的效果，我们没有将其包括在我们的评估中
## 2.2 Variational dropout
Gal and Ghahramani (2016)提出了一种递归神经网络的正则化方法。这也已在 HNMT 中实施，但是在芬兰语上进行的初步实验并未表明比基准系统有任何改进。虽然 Sennrich et al. (2016a)在 WMT2016 上报告了罗马尼亚新闻翻译任务的重大改进，但培训数据量低于芬兰语可用的数据量，这应该可以解释其中的一些差异。它们还在单词级别上应用 dropout,而HNMT 应用程序当前仅删除经常性状态
# 2.3 Context gates
上下文门 (Tu et al., 2017) 引人了一个明确的模型，用于选择目标句子生成应在多大程度上关注源句子或目标上下文，从而使网络有机会调整充分性和流利性之间的平衡。虽然我们在开发集上获得了更好的交叉熵，特别是在训练的早期，但对开发数据的 BLEU 和 chrF3 评估使我们在最终运行中决定使用较慢的上下文门
# 2.4 Coverage decoder
Wu et al. (2016)提出了一种根据经验确定的方法，用于在搜索算法中使用解码过程中产生的注意力向量，以使解码器偏向长度合理且覆盖源句的翻译。$^{1}$我们对长度，覆盖范围和过度翻译惩罚的参数空间进行了网格搜索，但没有发现任何导致开发集上的 BLEU 得分高于没有惩罚的解码器的
2.5 Forward-Backward reranking
训练翻译模型以从目标句子的开头或结尾生成翻译是微不足道的。HNMT 支持选择翻译方向，并结合其 n 个最佳列表和重新排名功能， 可以简单地在两个方向上生成候选翻译，并根据其分数进行组合。这导致了我们英语-芬兰语翻译的一些小改进
## 2.6 Ensembling
HNMT 支持两种通用的集成模式及其组合
$\bullet$ Proper ensembling where $p( w)$ = $\frac{1}{M}\sum_{m=1}^{M}p_{m}(w)$is used to predict target symbol $w$, given predictions $p_m(w)$ for each model $m$ in the ensemble.
$\bullet$Parameter averaging where the model's
parameter vector $\theta$ is computed as $^{1}$The HNMT implementation of this was con-
tributed by Stig-Arne Gr ö nroos.
$\frac{1}{N}\sum_{m=1}^{M}\theta_{m}$ for each model $m.$ This only works if the different $\theta_m$ are relatively similar, typically because they were saved at different points during the same training process.
正确集成的开销在集成系统的数量上是线性的，无论是用于训练(假设一个正在构建单独训练的模型的集合)和推理，而参数平均基本上是免费的。HNMT 允许模型组的适当集成， 其中参数在每个组内平均。这种灵活的结构允许许多设置，这些设置将在 Section 3.2 中进一步探讨
## 3 English-Finnish
在我们的实验中，我们使用了 WMT 提供的所有英语-芬兰并行数据集，但 Wiki 标题除外，这是一个很小且相当嘈杂的数据集，从去年开始对我们的实验没有任何贡献。我们还添加了大量的反向翻译数据，这些数据已被证明特别有助于神经机器翻译 (Sennrich et al., 2016b),也有助于统计 MT(Tiedemann et al., 2016)。表 1 列出了我们从 2014 年和 2016 年WMT 单语芬兰新闻数据中创建的反向翻译数据集的一些基本统计数据。从去年 (Tiedemann et al., 2016) 开始，我们将芬兰语的基于最佳约束短语的 SMT 模型应用于英语，该模型使用具有多个翻译路径，形态标记和伪标记的因子模型，用于对应于英语介词的大小写标记(Tiedemann et al.,2015)。系统得分 20。5% 在最新的 2016 年数据上降低了 BLEU,这是2016 年该任务的第二好系统
<table>
<tbody>
<tr>
<th> </th>
<th>sentences</th>
<th>Finnish</th>
<th>English</th>
</tr>
<tr>
<td>news2014</td>
<td>1,378,833</td>
<td>17,117,137</td>
<td>23,818,547</td>
</tr>
<tr>
<td>news2016</td>
<td>4,144,406</td>
<td>55,637,304</td>
<td>76,161,439</td>
</tr>
</tbody>
</table>
Table 1: Backtranslated Finnish news data
3.1 Preprocessing and postprocessing
我们在标记化和可信数据上训练了我们的模型，除了在原始未标记化数据上训练的字符级模型。对于前者，我们使用在并行训练数据上训练的模型，将 Moses 工具用于 Unicode/标点规范化，标记化和 truecasing
我们测试了三种不同类型的分词：基本的基于单词的分词，使用 OMorFi (Pirinen, 2015) 和字节对编码 (BPE) (Sennrich et al., 2016c) 的监督形态分词。对于后者，我们选择了细粒度的分割，当在并行数据上训练时，会产生20,000个令牌的小词汇量，期望 BPE 能够处理复合分割和形态分割的各种情况。我们始终 | [
[
"0.12177",
"0.0766",
"0.4727",
"0.0766",
"0.4727",
"0.0903",
"0.12177",
"0.0903"
],
[
"0.5483",
"0.0736",
"0.885",
"0.07245",
"0.885",
"0.0932",
"0.5483",
"0.0944"
],
[
"0.4668",
"0.08014",
"0.4836",
"0.08014",
"0.4836",
"0.0873",
"0.4668",
"0.0873"
],
[
"0.12006",
"0.0932",
"0.486",
"0.0932",
"0.486",
"0.1069",
"0.12006",
"0.1069"
],
[
"0.552",
"0.0932",
"0.882",
"0.0932",
"0.882",
"0.1069",
"0.552",
"0.1069"
],
[
"0.55",
"0.10925",
"0.881",
"0.10925",
"0.881",
"0.1223",
"0.55",
"0.1223"
],
[
"0.1184",
"0.1194",
"0.3267",
"0.1194",
"0.3267",
"0.133",
"0.1184",
"0.133"
],
[
"0.552",
"0.1259",
"0.882",
"0.1259",
"0.882",
"0.1395",
"0.552",
"0.1395"
],
[
"0.12006",
"0.1395",
"0.4895",
"0.1395",
"0.4895",
"0.1526",
"0.12006",
"0.1526"
],
[
"0.552",
"0.1437",
"0.644",
"0.1437",
"0.644",
"0.155",
"0.552",
"0.155"
],
[
"0.12177",
"0.1561",
"0.4878",
"0.1561",
"0.4878",
"0.1692",
"0.12177",
"0.1692"
],
[
"0.1184",
"0.1704",
"0.4878",
"0.1704",
"0.4878",
"0.1841",
"0.1184",
"0.1841"
],
[
"0.5156",
"0.1716",
"0.881",
"0.1716",
"0.881",
"0.1818",
"0.5156",
"0.1818"
],
[
"0.12177",
"0.187",
"0.4895",
"0.187",
"0.4895",
"0.2007",
"0.12177",
"0.2007"
],
[
"0.5146",
"0.1859",
"0.882",
"0.1859",
"0.882",
"0.1996",
"0.5146",
"0.1996"
],
[
"0.12177",
"0.2037",
"0.4878",
"0.2037",
"0.4878",
"0.2173",
"0.12177",
"0.2173"
],
[
"0.5156",
"0.2037",
"0.879",
"0.2037",
"0.879",
"0.2144",
"0.5156",
"0.2144"
],
[
"0.12177",
"0.2203",
"0.4878",
"0.2203",
"0.4878",
"0.234",
"0.12177",
"0.234"
],
[
"0.5156",
"0.2185",
"0.882",
"0.2185",
"0.882",
"0.2322",
"0.5156",
"0.2322"
],
[
"0.12006",
"0.2351",
"0.4895",
"0.234",
"0.4895",
"0.2488",
"0.12006",
"0.2494"
],
[
"0.513",
"0.2351",
"0.881",
"0.2351",
"0.881",
"0.2488",
"0.513",
"0.2488"
],
[
"0.12006",
"0.253",
"0.4878",
"0.253",
"0.4878",
"0.2666",
"0.12006",
"0.2666"
],
[
"0.513",
"0.2505",
"0.881",
"0.2505",
"0.881",
"0.2642",
"0.513",
"0.2642"
],
[
"0.1184",
"0.2673",
"0.44",
"0.2673",
"0.44",
"0.2808",
"0.1184",
"0.2808"
],
[
"0.5146",
"0.2666",
"0.5903",
"0.2666",
"0.5903",
"0.2808",
"0.5146",
"0.2808"
],
[
"0.1184",
"0.2935",
"0.2795",
"0.2944",
"0.2795",
"0.3088",
"0.1184",
"0.308"
],
[
"0.511",
"0.2957",
"0.692",
"0.2957",
"0.692",
"0.31",
"0.511",
"0.31"
],
[
"0.12006",
"0.3135",
"0.486",
"0.3135",
"0.486",
"0.328",
"0.12006",
"0.328"
],
[
"0.513",
"0.3206",
"0.881",
"0.3206",
"0.881",
"0.335",
"0.513",
"0.335"
],
[
"0.12006",
"0.3308",
"0.4895",
"0.3308",
"0.4895",
"0.3445",
"0.12006",
"0.3445"
],
[
"0.5156",
"0.3384",
"0.8833",
"0.3384",
"0.8833",
"0.3516",
"0.5156",
"0.3516"
],
[
"0.12177",
"0.3474",
"0.4878",
"0.3474",
"0.4878",
"0.361",
"0.12177",
"0.361"
],
[
"0.5156",
"0.354",
"0.882",
"0.354",
"0.882",
"0.3677",
"0.5156",
"0.3677"
],
[
"0.12177",
"0.3652",
"0.486",
"0.3652",
"0.486",
"0.376",
"0.12177",
"0.376"
],
[
"0.5156",
"0.3718",
"0.881",
"0.3718",
"0.881",
"0.3818",
"0.5156",
"0.3818"
],
[
"0.12006",
"0.3801",
"0.4895",
"0.3801",
"0.4895",
"0.3938",
"0.12006",
"0.3938"
],
[
"0.5146",
"0.3872",
"0.881",
"0.3872",
"0.881",
"0.398",
"0.5146",
"0.398"
],
[
"0.12006",
"0.3955",
"0.4895",
"0.3955",
"0.4895",
"0.4092",
"0.12006",
"0.4092"
],
[
"0.5156",
"0.4026",
"0.881",
"0.4026",
"0.881",
"0.4163",
"0.5156",
"0.4163"
],
[
"0.12006",
"0.412",
"0.4878",
"0.412",
"0.4878",
"0.4258",
"0.12006",
"0.4258"
],
[
"0.513",
"0.4187",
"0.879",
"0.4187",
"0.879",
"0.4316",
"0.513",
"0.4316"
],
[
"0.1184",
"0.4275",
"0.1427",
"0.4275",
"0.1427",
"0.4424",
"0.1184",
"0.4424"
],
[
"0.5156",
"0.4353",
"0.882",
"0.4353",
"0.882",
"0.449",
"0.5156",
"0.449"
],
[
"0.11755",
"0.4536",
"0.3074",
"0.4536",
"0.3074",
"0.4673",
"0.11755",
"0.4673"
],
[
"0.5146",
"0.4495",
"0.882",
"0.4507",
"0.882",
"0.4644",
"0.5146",
"0.4631"
],
[
"0.5146",
"0.4673",
"0.881",
"0.4673",
"0.881",
"0.481",
"0.5146",
"0.481"
],
[
"0.12006",
"0.4744",
"0.4878",
"0.4744",
"0.4878",
"0.488",
"0.12006",
"0.488"
],
[
"0.5146",
"0.4827",
"0.8833",
"0.4827",
"0.8833",
"0.4963",
"0.5146",
"0.4963"
],
[
"0.12006",
"0.4905",
"0.4878",
"0.4905",
"0.4878",
"0.5044",
"0.12006",
"0.5044"
],
[
"0.5156",
"0.4988",
"0.882",
"0.4988",
"0.882",
"0.5127",
"0.5156",
"0.5127"
],
[
"0.12006",
"0.5073",
"0.4878",
"0.5073",
"0.4878",
"0.521",
"0.12006",
"0.521"
],
[
"0.5156",
"0.5156",
"0.882",
"0.5156",
"0.882",
"0.5293",
"0.5156",
"0.5293"
],
[
"0.1184",
"0.5225",
"0.4878",
"0.5225",
"0.4878",
"0.536",
"0.1184",
"0.536"
],
[
"0.5146",
"0.53",
"0.881",
"0.53",
"0.881",
"0.5435",
"0.5146",
"0.5435"
],
[
"0.12177",
"0.539",
"0.4878",
"0.539",
"0.4878",
"0.5527",
"0.12177",
"0.5527"
],
[
"0.5156",
"0.5464",
"0.882",
"0.5464",
"0.882",
"0.56",
"0.5156",
"0.56"
],
[
"0.12006",
"0.5547",
"0.4878",
"0.5547",
"0.4878",
"0.5684",
"0.12006",
"0.5684"
],
[
"0.5156",
"0.5635",
"0.882",
"0.5635",
"0.882",
"0.5776",
"0.5156",
"0.5776"
],
[
"0.12006",
"0.572",
"0.2854",
"0.572",
"0.2854",
"0.585",
"0.12006",
"0.585"
],
[
"0.5146",
"0.58",
"0.744",
"0.58",
"0.744",
"0.5933",
"0.5146",
"0.5933"
],
[
"0.1184",
"0.5957",
"0.4014",
"0.5967",
"0.4014",
"0.611",
"0.1184",
"0.61"
],
[
"0.615",
"0.608",
"0.679",
"0.608",
"0.679",
"0.619",
"0.615",
"0.619"
],
[
"0.712",
"0.608",
"0.7617",
"0.608",
"0.7617",
"0.618",
"0.712",
"0.618"
],
[
"0.802",
"0.608",
"0.854",
"0.608",
"0.854",
"0.619",
"0.802",
"0.619"
],
[
"0.12177",
"0.618",
"0.4895",
"0.618",
"0.4895",
"0.632",
"0.12177",
"0.632"
],
[
"0.529",
"0.6206",
"0.602",
"0.6206",
"0.602",
"0.6313",
"0.529",
"0.6313"
],
[
"0.614",
"0.6206",
"0.684",
"0.6206",
"0.684",
"0.6313",
"0.614",
"0.6313"
],
[
"0.7",
"0.6206",
"0.7715",
"0.6206",
"0.7715",
"0.6313",
"0.7",
"0.6313"
],
[
"0.791",
"0.6206",
"0.8623",
"0.6206",
"0.8623",
"0.6313",
"0.791",
"0.6313"
],
[
"0.531",
"0.632",
"0.6055",
"0.632",
"0.6055",
"0.6426",
"0.531",
"0.6426"
],
[
"0.602",
"0.632",
"0.686",
"0.632",
"0.686",
"0.6426",
"0.602",
"0.6426"
],
[
"0.6987",
"0.6313",
"0.7734",
"0.6313",
"0.7734",
"0.641",
"0.6987",
"0.641"
],
[
"0.79",
"0.632",
"0.8657",
"0.632",
"0.8657",
"0.6426",
"0.79",
"0.6426"
],
[
"0.1234",
"0.652",
"0.4836",
"0.652",
"0.4836",
"0.662",
"0.1234",
"0.662"
],
[
"0.12177",
"0.6343",
"0.4878",
"0.6343",
"0.4878",
"0.648",
"0.12177",
"0.648"
],
[
"0.5146",
"0.6562",
"0.8716",
"0.6562",
"0.8716",
"0.6694",
"0.5146",
"0.6694"
],
[
"0.1234",
"0.6665",
"0.4844",
"0.6665",
"0.4844",
"0.677",
"0.1234",
"0.677"
],
[
"0.12006",
"0.6816",
"0.486",
"0.6816",
"0.486",
"0.6953",
"0.12006",
"0.6953"
],
[
"0.1184",
"0.6978",
"0.287",
"0.697",
"0.287",
"0.7114",
"0.1184",
"0.712"
],
[
"0.513",
"0.6953",
"0.833",
"0.697",
"0.833",
"0.7114",
"0.513",
"0.7104"
],
[
"0.1184",
"0.7227",
"0.2627",
"0.7236",
"0.2627",
"0.7383",
"0.1184",
"0.737"
],
[
"0.5156",
"0.7188",
"0.882",
"0.7188",
"0.882",
"0.7324",
"0.5156",
"0.7324"
],
[
"0.5156",
"0.7354",
"0.879",
"0.7354",
"0.879",
"0.745",
"0.5156",
"0.745"
],
[
"0.12006",
"0.7446",
"0.4668",
"0.7446",
"0.4668",
"0.7583",
"0.12006",
"0.7583"
],
[
"0.5146",
"0.749",
"0.882",
"0.75",
"0.882",
"0.764",
"0.5146",
"0.763"
],
[
"0.1368",
"0.771",
"0.4475",
"0.771",
"0.4475",
"0.785",
"0.1368",
"0.785"
],
[
"0.5156",
"0.7666",
"0.882",
"0.7666",
"0.882",
"0.78",
"0.5156",
"0.78"
],
[
"0.455",
"0.7725",
"0.4844",
"0.7725",
"0.4844",
"0.78",
"0.455",
"0.78"
],
[
"0.1511",
"0.781",
"0.4937",
"0.778",
"0.4937",
"0.8027",
"0.1511",
"0.8057"
],
[
"0.5146",
"0.783",
"0.769",
"0.783",
"0.769",
"0.7964",
"0.5146",
"0.7964"
],
[
"0.1561",
"0.8027",
"0.4844",
"0.8027",
"0.4844",
"0.818",
"0.1561",
"0.818"
],
[
"0.532",
"0.799",
"0.882",
"0.799",
"0.882",
"0.8125",
"0.532",
"0.8125"
],
[
"0.5156",
"0.814",
"0.882",
"0.8135",
"0.882",
"0.8276",
"0.5156",
"0.829"
],
[
"0.157",
"0.821",
"0.4055",
"0.821",
"0.4055",
"0.8306",
"0.157",
"0.8306"
],
[
"0.517",
"0.8306",
"0.882",
"0.8306",
"0.882",
"0.844",
"0.517",
"0.844"
],
[
"0.1368",
"0.8457",
"0.4895",
"0.8457",
"0.4895",
"0.86",
"0.1368",
"0.86"
],
[
"0.5146",
"0.847",
"0.881",
"0.847",
"0.881",
"0.861",
"0.5146",
"0.861"
],
[
"0.1561",
"0.8623",
"0.4895",
"0.8623",
"0.4895",
"0.877",
"0.1561",
"0.877"
],
[
"0.5156",
"0.865",
"0.881",
"0.865",
"0.881",
"0.876",
"0.5156",
"0.876"
],
[
"0.1377",
"0.882",
"0.4878",
"0.8843",
"0.4878",
"0.8984",
"0.1377",
"0.8965"
],
[
"0.5156",
"0.8813",
"0.881",
"0.8813",
"0.881",
"0.891",
"0.5156",
"0.891"
],
[
"0.12006",
"0.8965",
"0.357",
"0.898",
"0.357",
"0.909",
"0.12006",
"0.9077"
],
[
"0.5146",
"0.8955",
"0.881",
"0.8955",
"0.881",
"0.909",
"0.5146",
"0.909"
]
] | [
"由于初步实验没有表明使用层归一化的任何",
"Zm=1 Om for each model m. This only",
"-",
"致的效果，我们没有将其包括在我们的评估中",
"works if the different Om are relatively",
"similar, typically because they were saved",
"2.2Variational dropout",
"at different points during the same train-",
"Galand Ghahramani（2016)提出了一种递归",
"ing process.",
"神经网络的正则化方法。这也已在HNMT中",
"实施，但是在芬兰语上进行的初步实验并未表",
"正确集成的开销在集成系统的数量上是线性",
"明比基准系统有任何改进。虽然Sennrichetal.",
"的，无论是用于训练（假设一个正在构建单独",
"(2016a）在WMT2016上报告了罗马尼亚新闻",
"训练的模型的集合）和推理，而参数平均基本",
"翻译任务的重大改进，但培训数据量低于芬兰",
"上是免费的。HNMT允许模型组的适当集成，",
"语可用的数据量，这应该可以解释其中的一些",
"其中参数在每个组内平均。这种灵活的结构允",
"差异。它们还在单词级别上应用dropout，而",
"许许多设置，这些设置将在Section3.2中进",
"HNMT应用程序当前仅删除经常性状态",
"一步探讨",
"2.3Context gates",
"3English-Finnish",
"上下文门（Tuetal.，2017）引人了一个明确的",
"在我们的实验中，我们使用了WMT提供的",
"模型，用于选择目标句子生成应在多大程度上",
"所有英语-芬兰并行数据集，但Wiki标题除",
"关注源句子或目标上下文，从而使网络有机会",
"外，这是一个很小且相当噜杂的数据集，从去",
"调整充分性和流利性之间的平衡。虽然我们在",
"年开始对我们的实验没有任何贡献。我们还",
"开发集上获得了更好的交叉熵，特别是在训练",
"添加了大量的反向翻译数据，这些数据已被证",
"的早期，但对开发数据的BLEU和chrF3评",
"明特别有助于神经机器翻译（Sennrichetal.",
"估使我们在最终运行中决定使用较慢的上下文",
"2016b），也有助于统计MT（Tiedemannetal.",
"门",
"2016）。表1列出了我们从2014年和2016年",
"2.4Coverage decoder",
"WMT单语芬兰新闻数据中创建的反向翻译数",
"据集的一些基本统计数据。从去年（Tiedemann",
"Wuetal.（2016）提出了一种根据经验确定的",
"etal.，2016）开始，我们将芬兰语的基于最佳",
"方法，用于在搜索算法中使用解码过程中产生",
"约束短语的SMT模型应用于英语，该模型使",
"的注意力向量，以使解码器偏向长度合理且覆",
"用具有多个翻译路径，形态标记和伪标记的",
"盖源句的翻译。1我们对长度，覆盖范围和过",
"因子模型，用于对应于英语介词的大小写标记",
"度翻译惩罚的参数空间进行了网格搜索，但没",
"(Tiedemannetal.,2015）。系统得分20。5%",
"有发现任何导致开发集上的BLEU得分高于",
"在最新的2016年数据上降低了BLEU，这是",
"没有惩罚的解码器的",
"2016年该任务的第二好系统",
"2.5Forward-Backward reranking",
"sentences",
"Finnish",
"English",
"训练翻译模型以从目标句子的开头或结尾生成",
"news2014",
"1,378,833",
"17,117,137",
"23,818,547",
"news2016",
"4,144,406",
"55.637.304",
"76,161,439",
"可，并结合其n个最佳列表和重新排名功能",
"翻译是微不足道的。HNMT支持选择翻译方",
"Table l: Backtranslated Finnish news data",
"可以简单地在两个方向上生成候选翻译，并根",
"据其分数进行组合。这导致了我们英语-芬兰",
"语翻译的一些小改进",
"3.1 Preprocessing and postprocessing",
"2.6Ensembling",
"我们在标记化和可信数据上训练了我们的模",
"型，除了在原始未标记化数据上训练的字符级",
"HNMT支持两种通用的集成模式及其组合：",
"模型。对于前者，我们使用在并行训练数据上",
": Proper ensembling where p(w)",
"训练的模型，将Moses工具用于Unicode/标",
"",
"ZM=1Pm(w) is used to predict target",
"点规范化，标记化和truecasing",
"symbol w , given predictions Pm(w) for",
"我们测试了三种不同类型的分词：基本的基",
"于单词的分词，使用OMorFi（Pirinen，2015)",
"each model m in the ensemble.",
"和字节对编码（BPE）（Sennrichetal.，2016c）",
": Parameter averaging where the model's",
"的监督形态分词。对于后者，我们选择了细",
"parameter vector is computed as",
"粒度的分割，当在并行数据上训练时，会产生",
"IThe HNMT implementation of this was con-",
"20.000个令牌的小词汇量，期望BPE能够处",
"tributed by Stig-Arne Gr nroos",
"理复合分割和形态分割的各种情况。我们始终"
] | {
"arxiv_id": "1708.05942",
"language": "zh",
"page_idx": 1
} |
||
1708.05942-zh-2 | 1708.05942-zh-2.jpg | 使用相同的基于 BPE 的分割，并且没有尝试
以任何方式优化 BPE 参数
在开发过程中，我们观察到英语开发文件包含了很多动词形式的收缩(类型不会),但是这种收缩形式很少出现在训练数据中。因此， 我们还添加了一个预处理例程，将合同表格转换为未合同的等价物
最后，我们发现我们的芬兰语的 tokenizer/detokenizer 管道没有正确处理连字符/破折号的区别。在芬兰语中，’-”符号可以与两侧的空格一起使用，没有空格，只有左边的空格，只有右边的空格，如以下示例所示：
DraamaaRiossa-suomalaisnostajapyörtyi ... ,里约的戏剧-芬兰升降机昏倒了...' Kempinski- hotelli
$^\prime$Kempinski[-]hotel' 。kissajahiiri-leikkiä
,猫和老鼠[-]游戏”。$\"{o}ljy-jakaasutoiminnot$
,油[-]和气体功能”
标记器总是在两侧引入空格，这意味着 detokenizer 无法检索原始配置。为了解决这个问题，我们对翻译后的数据应用了后处理步骤。在对输出进行去令牌化之后，对于每个连字符，都会生成四个令牌化变体，并由混合到字符系统进行评分；然后，我们选择得分最高的令牌化变体
# 3.2 NMT Models
在初步实验中，我们专注于针对源侧和目标侧的不同分割策略，以及不同比例的反向翻译和并行数据。使用小写的 BLEU 和 chrF3 在newsdev2015 上评估了模型。$^{2}$表 2显示了一些结果
在这些实验中，我们发现 BPE 在目标端很有用，但在源端却没有那么多。字符级解码器受到 chrF3 等字符级评估分数的青睐，而BLEU 则偏爱使用 BPE 等较大单元的解码器。结合反向翻译和并行数据获得最佳结果； 使用所有反向翻译比限制反向翻译量以匹配并行数据的大小略好。基于监督形态学分割，然后进行 BPE 编码 (OMorFi)的模型产生了有希望的 chrF3 结果，但在 BLEU 方面滞后。需要进一步研究结合这些分割方法的好处和缺点
基于混合编码器和 BPE 解码器的模型在此初步评估中并未产生最佳结果，但在不同的评估类型，训练配置和评估数据中显示出最稳健的性能 (特别是在 newstest2015 集上优于其他模型)。因此，我们的五个提交中有四个使用该配置。为了进行比较，我们还提交了一个
$^{2}$The HNMT-internal BLEU computation is based on https://github.com/vikasnar/Bleu and on the NLTK tokenizer. The reported results are thus not directly comparable with offcial WMT results.
基于字符级解码器的系统
我们还研究了不同集成组合的效果，其结果可以在 Table 3 中找到。一般来说，适当的集成比保存点平均更好，但保存点平均总比没有好。进一步的实验表明，来自独立模型设置的平均保存点集合(从底部开始的第二行)与来自独立模型的多个保存点的集合(底部行)之间的差异是不一致的，因此我们使用前者(更快的) 变体我们的官方提交
提交的字符解码器系统使用 256 个维度进行单词嵌入，64 个维度进行字符嵌人，512 个编码器状态维度，1024 个解码器状态维度和256 个关注维度。我们训练四个独立模型，每个模型 72h, 并使用具有最佳 heldoutchrF3 分数的保存点(实际上我们没有观察到任何显着的过拟合，因此这相当于对所有模型使用近72h 的训练)。训练数据是所有并行和反向翻译数据的未处理版本。为了进行解码，我们对四个模型进行了适当的集成，并对每个模型的最后四个保存点进行了平均(在每个 5000 个训练批次之后保存了状态)
提交的 BPE 解码器系统使用与字符解码器系统相同的模型大小。同样，我们使用预处理和 BPE 编码的数据训练四个独立模型，每个模型 72 小时 ,并将连字符重新标记作为后处理步骤。我们提供了两个对比系统：一个没有输人归一化，显示减少了0。3个 BLEU,一个没有连字符重新标记，显示减少了 0。9BLEU(见表4
我们还提出了一个扩展系统，该系统基于上述四个模型和另外四个向后模型(i。e。,从右到左训练)。在测试时，我们生成 10 个最佳的前向翻译列表和类似的后向翻译列表。我们选择两个列表中发生的最佳翻译，或者如果列表不相交，则根据生成它的模型(向前或向后)选择可能性最高的翻译。此重新排名仅提供 +0。1 个 BLEU; 从正向模型中选择了 48个 $\% $,从反向模型中选择了 22 个%,并且在两个列表中都出现了 30 个%。该系统在自动和手动评估中排名第一3.3 SMT Baselines
除了神经 MT 模型之外，我们还训练了各种基于短语的 SMT 模型，以将我们的结果与另一种流行的范式进行对比。特别是，我们有兴趣看到 BPE 分割和反向翻译对统计 MT 的影响。这两种技术在神经 MT 中都很流行，但它们对统计 MT 的影响之前还没有得到适当的评估。因此，我们开始对不同的设置进行系统的比较，包括各种类型的分割和数据收集。所有系统都基于 Moses (Koehn et al., 2007) ,我们使用标准配置来训练基于非因子短语 | [
[
"0.12177",
"0.0766",
"0.4902",
"0.0766",
"0.4902",
"0.0903",
"0.12177",
"0.0903"
],
[
"0.4844",
"0.08014",
"0.5156",
"0.08014",
"0.5156",
"0.0885",
"0.4844",
"0.0885"
],
[
"0.513",
"0.0766",
"0.7188",
"0.0766",
"0.7188",
"0.0903",
"0.513",
"0.0903"
],
[
"0.12006",
"0.0932",
"0.336",
"0.0932",
"0.336",
"0.1069",
"0.12006",
"0.1069"
],
[
"0.531",
"0.09265",
"0.881",
"0.09265",
"0.881",
"0.1057",
"0.531",
"0.1057"
],
[
"0.1394",
"0.111",
"0.486",
"0.111",
"0.486",
"0.1217",
"0.1394",
"0.1217"
],
[
"0.5146",
"0.10925",
"0.881",
"0.10925",
"0.881",
"0.1223",
"0.5146",
"0.1223"
],
[
"0.12177",
"0.1259",
"0.4895",
"0.1259",
"0.4895",
"0.1395",
"0.12177",
"0.1395"
],
[
"0.5146",
"0.1259",
"0.882",
"0.1259",
"0.882",
"0.1395",
"0.5146",
"0.1395"
],
[
"0.12006",
"0.1426",
"0.4902",
"0.1426",
"0.4902",
"0.1561",
"0.12006",
"0.1561"
],
[
"0.5146",
"0.1414",
"0.882",
"0.1414",
"0.882",
"0.155",
"0.5146",
"0.155"
],
[
"0.12006",
"0.158",
"0.4878",
"0.158",
"0.4878",
"0.1716",
"0.12006",
"0.1716"
],
[
"0.513",
"0.1561",
"0.882",
"0.1567",
"0.882",
"0.1716",
"0.513",
"0.1704"
],
[
"0.12177",
"0.1758",
"0.2837",
"0.1758",
"0.2837",
"0.1859",
"0.12177",
"0.1859"
],
[
"0.517",
"0.1733",
"0.882",
"0.1733",
"0.882",
"0.187",
"0.517",
"0.187"
],
[
"0.1394",
"0.1924",
"0.486",
"0.1924",
"0.486",
"0.2031",
"0.1394",
"0.2031"
],
[
"0.5156",
"0.1895",
"0.882",
"0.1895",
"0.882",
"0.2031",
"0.5156",
"0.2031"
],
[
"0.12006",
"0.2073",
"0.4895",
"0.2073",
"0.4895",
"0.2203",
"0.12006",
"0.2203"
],
[
"0.511",
"0.206",
"0.735",
"0.206",
"0.735",
"0.2197",
"0.511",
"0.2197"
],
[
"0.1184",
"0.2227",
"0.486",
"0.2227",
"0.486",
"0.2363",
"0.1184",
"0.2363"
],
[
"0.531",
"0.2214",
"0.8833",
"0.2214",
"0.8833",
"0.2351",
"0.531",
"0.2351"
],
[
"0.12006",
"0.2393",
"0.4895",
"0.2393",
"0.4895",
"0.253",
"0.12006",
"0.253"
],
[
"0.5146",
"0.2382",
"0.882",
"0.2382",
"0.882",
"0.2517",
"0.5146",
"0.2517"
],
[
"0.12006",
"0.2546",
"0.492",
"0.2546",
"0.492",
"0.2683",
"0.12006",
"0.2683"
],
[
"0.5146",
"0.2542",
"0.8833",
"0.2542",
"0.8833",
"0.2673",
"0.5146",
"0.2673"
],
[
"0.1351",
"0.2695",
"0.4836",
"0.2715",
"0.4836",
"0.2861",
"0.1351",
"0.284"
],
[
"0.5146",
"0.2708",
"0.882",
"0.2708",
"0.882",
"0.284",
"0.5146",
"0.284"
],
[
"0.11584",
"0.2852",
"0.4954",
"0.2874",
"0.4954",
"0.3018",
"0.11584",
"0.2998"
],
[
"0.5146",
"0.284",
"0.882",
"0.2861",
"0.882",
"0.2998",
"0.5146",
"0.2976"
],
[
"0.1125",
"0.304",
"0.2627",
"0.304",
"0.2627",
"0.3176",
"0.1125",
"0.3176"
],
[
"0.5156",
"0.3027",
"0.882",
"0.3027",
"0.882",
"0.3164",
"0.5156",
"0.3164"
],
[
"0.1184",
"0.3184",
"0.44",
"0.3184",
"0.44",
"0.332",
"0.1184",
"0.332"
],
[
"0.5156",
"0.3184",
"0.882",
"0.3184",
"0.882",
"0.332",
"0.5156",
"0.332"
],
[
"0.12177",
"0.3362",
"0.4744",
"0.3362",
"0.4744",
"0.3499",
"0.12177",
"0.3499"
],
[
"0.5156",
"0.3362",
"0.881",
"0.3362",
"0.881",
"0.3467",
"0.5156",
"0.3467"
],
[
"0.1184",
"0.3516",
"0.2737",
"0.3516",
"0.2737",
"0.3652",
"0.1184",
"0.3652"
],
[
"0.5156",
"0.3508",
"0.882",
"0.3508",
"0.882",
"0.364",
"0.5156",
"0.364"
],
[
"0.1377",
"0.3682",
"0.4878",
"0.3682",
"0.4878",
"0.3818",
"0.1377",
"0.3818"
],
[
"0.5156",
"0.3652",
"0.881",
"0.3652",
"0.881",
"0.379",
"0.5156",
"0.379"
],
[
"0.1184",
"0.383",
"0.4878",
"0.3843",
"0.4878",
"0.3984",
"0.1184",
"0.398"
],
[
"0.5146",
"0.383",
"0.882",
"0.383",
"0.882",
"0.3967",
"0.5146",
"0.3967"
],
[
"0.12006",
"0.401",
"0.4895",
"0.401",
"0.4895",
"0.4146",
"0.12006",
"0.4146"
],
[
"0.5146",
"0.3997",
"0.7256",
"0.3997",
"0.7256",
"0.4133",
"0.5146",
"0.4133"
],
[
"0.1184",
"0.4163",
"0.486",
"0.4163",
"0.486",
"0.43",
"0.1184",
"0.43"
],
[
"0.531",
"0.415",
"0.882",
"0.415",
"0.882",
"0.4287",
"0.531",
"0.4287"
],
[
"0.12177",
"0.434",
"0.486",
"0.434",
"0.486",
"0.4448",
"0.12177",
"0.4448"
],
[
"0.517",
"0.4329",
"0.881",
"0.4329",
"0.881",
"0.4436",
"0.517",
"0.4436"
],
[
"0.12006",
"0.449",
"0.4878",
"0.449",
"0.4878",
"0.462",
"0.12006",
"0.462"
],
[
"0.5146",
"0.4478",
"0.882",
"0.4478",
"0.882",
"0.4614",
"0.5146",
"0.4614"
],
[
"0.12006",
"0.4644",
"0.2124",
"0.4644",
"0.2124",
"0.4785",
"0.12006",
"0.4785"
],
[
"0.5156",
"0.4631",
"0.882",
"0.4631",
"0.882",
"0.4768",
"0.5156",
"0.4768"
],
[
"0.5146",
"0.4785",
"0.881",
"0.4785",
"0.881",
"0.4922",
"0.5146",
"0.4922"
],
[
"0.12006",
"0.4922",
"0.2761",
"0.4922",
"0.2761",
"0.506",
"0.12006",
"0.506"
],
[
"0.5156",
"0.4934",
"0.882",
"0.4946",
"0.882",
"0.509",
"0.5156",
"0.5083"
],
[
"0.12006",
"0.5156",
"0.486",
"0.5156",
"0.486",
"0.5254",
"0.12006",
"0.5254"
],
[
"0.5146",
"0.511",
"0.881",
"0.511",
"0.881",
"0.525",
"0.5146",
"0.525"
],
[
"0.12006",
"0.53",
"0.4895",
"0.53",
"0.4895",
"0.5435",
"0.12006",
"0.5435"
],
[
"0.5146",
"0.5293",
"0.5684",
"0.5293",
"0.5684",
"0.5405",
"0.5146",
"0.5405"
],
[
"0.1184",
"0.5444",
"0.4878",
"0.5444",
"0.4878",
"0.558",
"0.1184",
"0.558"
],
[
"0.532",
"0.5435",
"0.8833",
"0.5435",
"0.8833",
"0.557",
"0.532",
"0.557"
],
[
"0.12006",
"0.561",
"0.4937",
"0.561",
"0.4937",
"0.576",
"0.12006",
"0.576"
],
[
"0.5146",
"0.56",
"0.882",
"0.56",
"0.882",
"0.5737",
"0.5146",
"0.5737"
],
[
"0.12177",
"0.579",
"0.1755",
"0.579",
"0.1755",
"0.5903",
"0.12177",
"0.5903"
],
[
"0.5156",
"0.5747",
"0.882",
"0.5767",
"0.882",
"0.5903",
"0.5156",
"0.5884"
],
[
"0.1377",
"0.5933",
"0.486",
"0.5933",
"0.486",
"0.607",
"0.1377",
"0.607"
],
[
"0.5146",
"0.5913",
"0.881",
"0.5913",
"0.881",
"0.605",
"0.5146",
"0.605"
],
[
"0.12006",
"0.611",
"0.4878",
"0.611",
"0.4878",
"0.6245",
"0.12006",
"0.6245"
],
[
"0.5146",
"0.6094",
"0.879",
"0.6094",
"0.879",
"0.619",
"0.5146",
"0.619"
],
[
"0.1184",
"0.626",
"0.486",
"0.626",
"0.486",
"0.6396",
"0.1184",
"0.6396"
],
[
"0.5156",
"0.6245",
"0.881",
"0.6245",
"0.881",
"0.638",
"0.5156",
"0.638"
],
[
"0.12006",
"0.6426",
"0.4895",
"0.6426",
"0.4895",
"0.6562",
"0.12006",
"0.6562"
],
[
"0.5156",
"0.64",
"0.882",
"0.64",
"0.882",
"0.654",
"0.5156",
"0.654"
],
[
"0.12177",
"0.659",
"0.4902",
"0.659",
"0.4902",
"0.673",
"0.12177",
"0.673"
],
[
"0.513",
"0.655",
"0.882",
"0.655",
"0.882",
"0.6685",
"0.513",
"0.6685"
],
[
"0.12177",
"0.675",
"0.4878",
"0.675",
"0.4878",
"0.6885",
"0.12177",
"0.6885"
],
[
"0.5156",
"0.673",
"0.882",
"0.673",
"0.882",
"0.686",
"0.5156",
"0.686"
],
[
"0.12006",
"0.6914",
"0.486",
"0.6914",
"0.486",
"0.7017",
"0.12006",
"0.7017"
],
[
"0.5146",
"0.6885",
"0.882",
"0.6885",
"0.882",
"0.7017",
"0.5146",
"0.7017"
],
[
"0.12177",
"0.707",
"0.4878",
"0.707",
"0.4878",
"0.72",
"0.12177",
"0.72"
],
[
"0.5156",
"0.706",
"0.6826",
"0.706",
"0.6826",
"0.7163",
"0.5156",
"0.7163"
],
[
"0.1184",
"0.7227",
"0.4878",
"0.7227",
"0.4878",
"0.7363",
"0.1184",
"0.7363"
],
[
"0.513",
"0.73",
"0.684",
"0.73",
"0.684",
"0.7437",
"0.513",
"0.7437"
],
[
"0.1184",
"0.7383",
"0.4878",
"0.7383",
"0.4878",
"0.752",
"0.1184",
"0.752"
],
[
"0.513",
"0.75",
"0.881",
"0.75",
"0.881",
"0.763",
"0.513",
"0.763"
],
[
"0.1377",
"0.756",
"0.4895",
"0.756",
"0.4895",
"0.7695",
"0.1377",
"0.7695"
],
[
"0.513",
"0.7656",
"0.879",
"0.7656",
"0.879",
"0.7793",
"0.513",
"0.7793"
],
[
"0.12006",
"0.7715",
"0.4895",
"0.7715",
"0.4895",
"0.785",
"0.12006",
"0.785"
],
[
"0.12006",
"0.788",
"0.4895",
"0.788",
"0.4895",
"0.802",
"0.12006",
"0.802"
],
[
"0.5156",
"0.783",
"0.882",
"0.783",
"0.882",
"0.7964",
"0.5156",
"0.7964"
],
[
"0.1184",
"0.804",
"0.4878",
"0.804",
"0.4878",
"0.818",
"0.1184",
"0.818"
],
[
"0.5146",
"0.799",
"0.882",
"0.799",
"0.882",
"0.8125",
"0.5146",
"0.8125"
],
[
"0.12177",
"0.8193",
"0.4878",
"0.8193",
"0.4878",
"0.833",
"0.12177",
"0.833"
],
[
"0.5146",
"0.8154",
"0.882",
"0.8154",
"0.882",
"0.829",
"0.5146",
"0.829"
],
[
"0.1184",
"0.836",
"0.486",
"0.835",
"0.486",
"0.8486",
"0.1184",
"0.8496"
],
[
"0.5146",
"0.8306",
"0.881",
"0.8306",
"0.881",
"0.844",
"0.5146",
"0.844"
],
[
"0.5156",
"0.847",
"0.882",
"0.847",
"0.882",
"0.861",
"0.5156",
"0.861"
],
[
"0.1377",
"0.86",
"0.4895",
"0.86",
"0.4895",
"0.8735",
"0.1377",
"0.8735"
],
[
"0.5146",
"0.8643",
"0.8833",
"0.8643",
"0.8833",
"0.8774",
"0.5146",
"0.8774"
],
[
"0.1184",
"0.8716",
"0.4878",
"0.8706",
"0.4878",
"0.8853",
"0.1184",
"0.8857"
],
[
"0.1184",
"0.8843",
"0.486",
"0.8843",
"0.486",
"0.898",
"0.1184",
"0.898"
],
[
"0.5146",
"0.88",
"0.882",
"0.88",
"0.882",
"0.8936",
"0.5146",
"0.8936"
],
[
"0.12006",
"0.8965",
"0.4417",
"0.8965",
"0.4417",
"0.91",
"0.12006",
"0.91"
],
[
"0.5146",
"0.8955",
"0.881",
"0.8955",
"0.881",
"0.909",
"0.5146",
"0.909"
]
] | [
"使用相同的基于BPE的分割，并且没有尝试",
"",
"基于字符级解码器的系统",
"以任何方式优化BPE参数",
"我们还研究了不同集成组合的效果，其结果",
"在开发过程中，我们观察到英语开发文件包",
"可以在Table3中找到。一般来说，适当的集",
"含了很多动词形式的收缩（类型不会），但是",
"成比保存点平均更好，但保存点平均总比没有",
"这种收缩形式很少出现在训练数据中。因此，",
"好。进一步的实验表明，来自独立模型设置的",
"我们还添加了一个预处理例程，将合同表格转",
"平均保存点集合（从底部开始的第二行）与来",
"换为未合同的等价物",
"自独立模型的多个保存点的集合（底部行）之",
"最后，我们发现我们的芬兰语的tok-",
"间的差异是不一致的，因此我们使用前者（更",
"enizer/detokenizer管道没有正确处理连字",
"快的）变体我们的官方提交",
"符/破折号的区别。在芬兰语中，’符号可",
"提交的字符解码器系统使用256个维度进",
"以与两侧的空格一起使用，没有空格，只有左",
"行单词嵌人，64个维度进行字符嵌入，512个",
"边的空格，只有右边的空格，如以下示例所示：",
"编码器状态维度，1024个解码器状态维度和",
"DraamaaRiossa-suomalaisnostajapyortyi...",
"256个关注维度。我们训练四个独立模型，每",
"，里约的戏剧-芬兰升降机昏倒了…。",
"个模型72h，并使用具有最佳heldoutchrF3分",
"_Kempinski-hotelli",
"数的保存点（实际上我们没有观察到任何显",
"'Kempinski-hotel’.kissajahiri-leikkia",
"着的过拟合，因此这相当于对所有模型使用近",
"”猫和老鼠]游戏’。oljy-jakaasutoiminnot",
"72h的训练）。训练数据是所有并行和反向翻",
"’油[-]和气体功能",
"译数据的未处理版本。为了进行解码，我们对",
"标记器总是在两侧引入空格，这意味着deto-",
"四个模型进行了适当的集成，并对每个模型的",
"kenizer无法检索原始配置。为了解决这个问",
"最后四个保存点进行了平均（在每个5000个",
"题，我们对翻译后的数据应用了后处理步骤。",
"训练批次之后保存了状态）",
"在对输出进行去令牌化之后，对于每个连字",
"提交的BPE解码器系统使用与字符解码器",
"符，都会生成四个令牌化变体，并由混合到学",
"系统相同的模型大小。同样，我们使用预处理",
"符系统进行评分；然后，我们选择得分最高的",
"和BPE编码的数据训练四个独立模型，每个",
"令牌化变体",
"模型72小时，并将连字符重新标记作为后处理",
"步骤。我们提供了两个对比系统：一个没有输",
"3.2NMT Models",
"人归一化，显示减少了0。3个BLEU，一个没",
"在初步实验中，我们专注于针对源侧和目标",
"有连字符重新标记，显示减少了0。9BLEU(见",
"侧的不同分割策略，以及不同比例的反向翻译",
"表4）",
"和并行数据。使用小写的BLEU和 chrF3在",
"我们还提出了一个扩展系统，该系统基于上",
"newsdev2015上评估了模型。2表2显示了一",
"述四个模型和另外四个向后模型（i。e。，从",
"些结果",
"右到左训练)。在测试时，我们生成10个最",
"在这些实验中，我们发现BPE在目标端很",
"佳的前向翻译列表和类似的后向翻译列表。我",
"有用，但在源端却没有那么多。字符级解码",
"们选择两个列表中发生的最佳翻译，或者如果",
"器受到chrF3等字符级评估分数的青睐，而",
"列表不相交，则根据生成它的模型（向前或向",
"BLEU则偏爱使用BPE等较大单元的解码",
"后）选择可能性最高的翻译。此重新排名仅提",
"器。结合反向翻译和并行数据获得最佳结果；",
"供+0。1个BLEU；从正向模型中选择了48",
"使用所有反向翻译比限制反向翻译量以匹配并",
"个%，从反向模型中选择了22个%，并且在两",
"行数据的大小略好。基于监督形态学分割，然",
"个列表中都出现了30个%。该系统在自动和",
"后进行BPE编码（OMorFi）的模型产生了有",
"手动评估中排名第一",
"希望的chrF3结果，但在BLEU方面滞后。需",
"3.3SMT Baselines",
"要进一步研究结合这些分割方法的好处和缺点",
"除了神经MT模型之外，我们还训练了各种",
"基于混合编码器和BPE解码器的模型在此",
"基于短语的SMT模型，以将我们的结果与另",
"初步评估中并未产生最佳结果，但在不同的评",
"估类型，训练配置和评估数据中显示出最稳健",
"一种流行的范式进行对比。特别是，我们有兴",
"的性能（特别是在newstest2015集上优于其",
"趣看到BPE分割和反向翻译对统计MT的影",
"他模型）。因此，我们的五个提交中有四个使",
"响。这两种技术在神经MT中都很流行，但",
"用该配置。为了进行比较，我们还提交了一个",
"它们对统计MT的影响之前还没有得到适当",
"的评估。因此，我们开始对不同的设置进行系",
"2The HNMT-internal BLEU computation is based",
"统的比较，包括各种类型的分割和数据收集。",
"on https://github.com/vikasnar/Bleu and on the",
"NLTK tokenizer. The reported results are thus not",
"所有系统都基于Moses(Koehnetal，2007)",
"directly comparable with official WMT results.",
"，我们使用标准配置来训练基于非因子短语"
] | {
"arxiv_id": "1708.05942",
"language": "zh",
"page_idx": 2
} |
||
2404.04520-zh-1 | 2404.04520-zh-1.jpg | IITKatSemEval-2024 任务 4: 用于检测模因中说服技术的分层嵌人
Shreenaga Chikoti
Shrey Mehta Ashutosh Modi
Indian Institute of Technology Kanpur (IIT Kanpur)
[email protected]
[email protected]
# Abstract
Memes are one of the most popular types of content used in an online disinformation campaign. They are primarily effective on social media platforms since they can easily reach many users. Memes in a disinformation campaign achieve their goal of influencing the users through several rhetorical and psychological techniques, such as causal oversimplification, name-calling, and smear. The SemEval 2024 Task 4 Multilingual Detection of Persuasion Technique in Memes on identifying such techniques in the memes is divided across three sub-tasks:( 1 ) Hierarchical multilabel classification using only textual content of the meme, $( \textbf{2})$ Hierarchical multi-label classification using both, textual and visual content of the meme and ( 3 ) Binary classification of whether the meme contains a persuasion technique or not using it's textual and visual content. This paper proposes an ensemble of Class Definition Prediction (CDP) and hyperbolic embeddings-based approaches for this task. We enhance meme classification accuracy and comprehensiveness by integrating HypEmo's hierarchical label embeddings (Chen et al., 2023) and a multi-task learning framework for emotion prediction. We achieve a hierarchical F1-score of 0.60,0.67, and 0.48 on the respective sub-tasks.
### 1 Introduction
今天，通过不同的社交媒体平台(Keswani et al. 2020; Singh et al., 2020) , 模因在所有年龄段的人群中都很受欢迎。这些模因可以帮助人们了解周围的趋势，并可以影响他们的决策。模因是在人们之间传播虚假信息的流行模式之一(图 1 中的示例),因为研究表明，人们倾向于相信他们在互联网上传播的此类模因中经常看到的内容(Moravec et al.,2018)。正如 2016 年美国总统竞选期间研究(Shu et al., 2017)所证明的那样，包括机器人，半机械人和巨魔在内的邪恶演员利用模因引起情绪反应并传播误导性叙述 (Guo et al., 2020)
<FigureHere>
Figure 1: Sample set of memes showing the multi-
modal setting
在这方面，SemEval-2024Task4 (Dimitrov et al., 2024) 着重于预测模因中使用的说服技术 (来自视觉和文本内容),涉及四种不同的语言：英语，阿拉伯语，北马其顿语和保加利亚语。该任务分为三个子任务：(1)仅使用模因的文本内容进行分层多标签分类，(2)使用模因的文本和视觉内容进行分层多标签分类， 以及(3)二元分类模因是否包含说服技术或不使用其文本和视觉内容。为每个子任务提供训练数据，但仅以英语提供。提供了各种说服技术的分类(图2)及其各自的定义
为了解决子任务 1 ,我们采用了一种双重方法，该方法涉及对每个类进行基于定义的建模，并使用 Chen et al.(2023)中提出的双曲线嵌人进行分层分类。基于双曲线嵌人，该方法通过利用层次结构来促进说服技术的细微分类。基于定义的建模的结合允许数据集不可知的方法，提高分类的精度，而不依赖于层次结构
对于子任务 2 , 我们通过集成 CLIP 嵌人(Radford et al., 2021) 来增强我们的方法，以从模因的文本和视觉组件中捕获基本特征。这种文本和视觉信息的融合可以更全面地分析模因 | [
[
"0.1486",
"0.0944",
"0.848",
"0.0956",
"0.848",
"0.10986",
"0.1486",
"0.10925"
],
[
"0.2527",
"0.1426",
"0.4165",
"0.1426",
"0.4165",
"0.1567",
"0.2527",
"0.1567"
],
[
"0.4568",
"0.1426",
"0.571",
"0.1426",
"0.571",
"0.1567",
"0.4568",
"0.1567"
],
[
"0.613",
"0.1426",
"0.7466",
"0.1426",
"0.7466",
"0.1561",
"0.613",
"0.1561"
],
[
"0.287",
"0.158",
"0.7095",
"0.1592",
"0.7095",
"0.1733",
"0.287",
"0.1729"
],
[
"0.398",
"0.1758",
"0.601",
"0.1758",
"0.601",
"0.1895",
"0.398",
"0.1895"
],
[
"0.3787",
"0.1948",
"0.6206",
"0.1948",
"0.6206",
"0.2048",
"0.3787",
"0.2048"
],
[
"0.2627",
"0.2625",
"0.3418",
"0.2625",
"0.3418",
"0.2766",
"0.2627",
"0.2766"
],
[
"0.5356",
"0.2625",
"0.677",
"0.2625",
"0.677",
"0.2756",
"0.5356",
"0.2756"
],
[
"0.5366",
"0.2737",
"0.686",
"0.2756",
"0.686",
"0.2874",
"0.5366",
"0.2852"
],
[
"0.1453",
"0.2852",
"0.4592",
"0.2852",
"0.4592",
"0.2986",
"0.1453",
"0.2986"
],
[
"0.1453",
"0.2998",
"0.4568",
"0.2986",
"0.4568",
"0.3123",
"0.1453",
"0.3135"
],
[
"0.1436",
"0.3142",
"0.4575",
"0.3123",
"0.4575",
"0.326",
"0.1436",
"0.328"
],
[
"0.727",
"0.3225",
"0.831",
"0.3225",
"0.831",
"0.3374",
"0.727",
"0.3374"
],
[
"0.147",
"0.3289",
"0.4568",
"0.3289",
"0.4568",
"0.339",
"0.147",
"0.339"
],
[
"0.147",
"0.3416",
"0.4592",
"0.3416",
"0.4592",
"0.3552",
"0.147",
"0.3552"
],
[
"0.733",
"0.3403",
"0.827",
"0.3425",
"0.827",
"0.3557",
"0.733",
"0.354"
],
[
"0.1453",
"0.3557",
"0.4575",
"0.3557",
"0.4575",
"0.3694",
"0.1453",
"0.3694"
],
[
"0.534",
"0.3599",
"0.687",
"0.3582",
"0.687",
"0.3718",
"0.534",
"0.3735"
],
[
"0.7246",
"0.357",
"0.8296",
"0.3599",
"0.8296",
"0.3748",
"0.7246",
"0.3718"
],
[
"0.1453",
"0.3694",
"0.4592",
"0.3706",
"0.4592",
"0.3843",
"0.1453",
"0.383"
],
[
"0.607",
"0.3801",
"0.644",
"0.3801",
"0.644",
"0.3872",
"0.607",
"0.3872"
],
[
"0.723",
"0.3801",
"0.785",
"0.3801",
"0.785",
"0.3872",
"0.723",
"0.3872"
],
[
"0.657",
"0.3901",
"0.739",
"0.3901",
"0.739",
"0.401",
"0.657",
"0.401"
],
[
"0.147",
"0.3843",
"0.4592",
"0.3843",
"0.4592",
"0.398",
"0.147",
"0.398"
],
[
"0.1453",
"0.3984",
"0.4592",
"0.3984",
"0.4592",
"0.412",
"0.1453",
"0.412"
],
[
"0.147",
"0.4133",
"0.4592",
"0.4133",
"0.4592",
"0.427",
"0.147",
"0.427"
],
[
"0.1453",
"0.427",
"0.4575",
"0.427",
"0.4575",
"0.4407",
"0.1453",
"0.4407"
],
[
"0.1453",
"0.4412",
"0.4592",
"0.4412",
"0.4592",
"0.4548",
"0.1453",
"0.4548"
],
[
"0.147",
"0.4573",
"0.4575",
"0.4573",
"0.4575",
"0.4673",
"0.147",
"0.4673"
],
[
"0.607",
"0.456",
"0.79",
"0.456",
"0.79",
"0.4739",
"0.607",
"0.4739"
],
[
"0.147",
"0.4697",
"0.4568",
"0.4702",
"0.4568",
"0.4822",
"0.147",
"0.481"
],
[
"0.1453",
"0.4827",
"0.4575",
"0.4827",
"0.4575",
"0.4963",
"0.1453",
"0.4963"
],
[
"0.511",
"0.4863",
"0.882",
"0.4863",
"0.882",
"0.4995",
"0.511",
"0.4995"
],
[
"0.1453",
"0.4976",
"0.4592",
"0.4976",
"0.4592",
"0.511",
"0.1453",
"0.511"
],
[
"0.51",
"0.4976",
"0.607",
"0.5005",
"0.607",
"0.5156",
"0.51",
"0.5127"
],
[
"0.147",
"0.5127",
"0.4592",
"0.5127",
"0.4592",
"0.5254",
"0.147",
"0.5254"
],
[
"0.1436",
"0.525",
"0.4575",
"0.527",
"0.4575",
"0.5405",
"0.1436",
"0.538"
],
[
"0.1453",
"0.5405",
"0.4575",
"0.539",
"0.4575",
"0.5527",
"0.1453",
"0.554"
],
[
"0.531",
"0.5464",
"0.881",
"0.5464",
"0.881",
"0.56",
"0.531",
"0.56"
],
[
"0.1453",
"0.5547",
"0.4575",
"0.5547",
"0.4575",
"0.5684",
"0.1453",
"0.5684"
],
[
"0.1453",
"0.5684",
"0.4575",
"0.5684",
"0.4575",
"0.582",
"0.1453",
"0.582"
],
[
"0.511",
"0.5635",
"0.881",
"0.5635",
"0.881",
"0.5767",
"0.511",
"0.5767"
],
[
"0.147",
"0.583",
"0.4568",
"0.583",
"0.4568",
"0.5967",
"0.147",
"0.5967"
],
[
"0.5146",
"0.579",
"0.879",
"0.579",
"0.879",
"0.593",
"0.5146",
"0.593"
],
[
"0.1453",
"0.597",
"0.4575",
"0.597",
"0.4575",
"0.611",
"0.1453",
"0.611"
],
[
"0.513",
"0.5957",
"0.879",
"0.5957",
"0.879",
"0.6094",
"0.513",
"0.6094"
],
[
"0.147",
"0.6123",
"0.4575",
"0.6123",
"0.4575",
"0.626",
"0.147",
"0.626"
],
[
"0.5146",
"0.611",
"0.881",
"0.611",
"0.881",
"0.6245",
"0.5146",
"0.6245"
],
[
"0.147",
"0.626",
"0.4568",
"0.626",
"0.4568",
"0.6396",
"0.147",
"0.6396"
],
[
"0.513",
"0.626",
"0.879",
"0.626",
"0.879",
"0.6396",
"0.513",
"0.6396"
],
[
"0.1436",
"0.6396",
"0.4575",
"0.6396",
"0.4575",
"0.653",
"0.1436",
"0.653"
],
[
"0.5146",
"0.6436",
"0.8774",
"0.6436",
"0.8774",
"0.6567",
"0.5146",
"0.6567"
],
[
"0.147",
"0.654",
"0.4568",
"0.654",
"0.4568",
"0.6675",
"0.147",
"0.6675"
],
[
"0.513",
"0.66",
"0.881",
"0.66",
"0.881",
"0.6733",
"0.513",
"0.6733"
],
[
"0.147",
"0.6694",
"0.4568",
"0.6694",
"0.4568",
"0.6797",
"0.147",
"0.6797"
],
[
"0.5146",
"0.677",
"0.881",
"0.677",
"0.881",
"0.6904",
"0.5146",
"0.6904"
],
[
"0.147",
"0.684",
"0.3284",
"0.684",
"0.3284",
"0.695",
"0.147",
"0.695"
],
[
"0.5146",
"0.6924",
"0.881",
"0.6924",
"0.881",
"0.706",
"0.5146",
"0.706"
],
[
"0.1184",
"0.707",
"0.1503",
"0.707",
"0.1503",
"0.7153",
"0.1184",
"0.7153"
],
[
"0.1486",
"0.703",
"0.2644",
"0.705",
"0.2644",
"0.7197",
"0.1486",
"0.7173"
],
[
"0.5146",
"0.708",
"0.8076",
"0.708",
"0.8076",
"0.7217",
"0.5146",
"0.7217"
],
[
"0.1184",
"0.73",
"0.4844",
"0.73",
"0.4844",
"0.7437",
"0.1184",
"0.7437"
],
[
"0.532",
"0.727",
"0.881",
"0.727",
"0.881",
"0.7407",
"0.532",
"0.7407"
],
[
"0.1184",
"0.745",
"0.486",
"0.745",
"0.486",
"0.759",
"0.1184",
"0.759"
],
[
"0.511",
"0.7437",
"0.881",
"0.7437",
"0.881",
"0.7573",
"0.511",
"0.7573"
],
[
"0.1184",
"0.7603",
"0.4836",
"0.7603",
"0.4836",
"0.775",
"0.1184",
"0.775"
],
[
"0.5146",
"0.759",
"0.881",
"0.759",
"0.881",
"0.7725",
"0.5146",
"0.7725"
],
[
"0.11755",
"0.777",
"0.4836",
"0.777",
"0.4836",
"0.7905",
"0.11755",
"0.7905"
],
[
"0.513",
"0.7754",
"0.881",
"0.7754",
"0.881",
"0.789",
"0.513",
"0.789"
],
[
"0.12006",
"0.7944",
"0.4668",
"0.7944",
"0.4668",
"0.808",
"0.12006",
"0.808"
],
[
"0.5146",
"0.7915",
"0.879",
"0.7915",
"0.879",
"0.8047",
"0.5146",
"0.8047"
],
[
"0.1184",
"0.81",
"0.4844",
"0.81",
"0.4844",
"0.8237",
"0.1184",
"0.8237"
],
[
"0.511",
"0.807",
"0.879",
"0.807",
"0.879",
"0.821",
"0.511",
"0.821"
],
[
"0.1184",
"0.826",
"0.4844",
"0.826",
"0.4844",
"0.839",
"0.1184",
"0.839"
],
[
"0.513",
"0.8237",
"0.881",
"0.8237",
"0.881",
"0.8374",
"0.513",
"0.8374"
],
[
"0.1184",
"0.843",
"0.4844",
"0.843",
"0.4844",
"0.8555",
"0.1184",
"0.8555"
],
[
"0.5146",
"0.8413",
"0.531",
"0.8413",
"0.531",
"0.851",
"0.5146",
"0.851"
],
[
"0.12006",
"0.86",
"0.4836",
"0.86",
"0.4836",
"0.8706",
"0.12006",
"0.8706"
],
[
"0.531",
"0.858",
"0.881",
"0.858",
"0.881",
"0.8716",
"0.531",
"0.8716"
],
[
"0.1184",
"0.8745",
"0.4844",
"0.8745",
"0.4844",
"0.888",
"0.1184",
"0.888"
],
[
"0.511",
"0.8726",
"0.8774",
"0.8745",
"0.8774",
"0.888",
"0.511",
"0.8857"
],
[
"0.1184",
"0.891",
"0.4844",
"0.891",
"0.4844",
"0.905",
"0.1184",
"0.905"
],
[
"0.511",
"0.891",
"0.8774",
"0.891",
"0.8774",
"0.905",
"0.511",
"0.905"
],
[
"0.1184",
"0.9067",
"0.309",
"0.9067",
"0.309",
"0.9204",
"0.1184",
"0.9204"
],
[
"0.5146",
"0.9067",
"0.879",
"0.9067",
"0.879",
"0.9204",
"0.5146",
"0.9204"
]
] | [
"IITKatSemEval-2024任务4：用于检测模因中说服技术的分层嵌入",
"Shreenaga Chikoti",
"Shrey Mehta",
"Ashutosh Modi",
"Indian Institute of Technology Kanpur (IIT Kanpur)",
"[email protected]",
"[email protected]",
"Abstract",
"DONALD TRUMI",
"WILLALWAYS BETHE",
"Memes are one of the most popular types of",
"content used in an online disinformation cam-",
"paign. They are primarily effective on social",
"Istill trust",
"mediaplatformssincethey caneasilyreach",
"many users. Memes in a disinformation cam-",
"this man",
"paign achieve their goal of influencing the",
"PEOPLE'SPRESIDENT",
"morethsniantone.c.",
"users through several rhetorical and psycho-",
"POLL:",
"HISMANIS",
"MENTALLY ILL",
"logical techniques, such as causal oversimpli-",
"fication, name-calling, and smear. The Se-",
"mEval 2024 Task 4 Multilingual Detection of",
"Persuasion Technique in Memes on identify-",
"ing such techniques in the memes is divided",
"across three sub-tasks:（1)Hierarchical multi",
"LIKETHISPHOTO",
"label classificationusingonlytextual content",
"of the meme, ( 2 ) Hierarchical multi-label",
"Figure l: Sample set of memes showing the multi-",
"classification using both, textual and visual",
"modal setting",
"content of the meme and ( 3 ) Binary classi-",
"fication of whether the meme contains a per-",
"suasion technique or not using it's textual and",
"在这方面，SemEval-2024Task4（Dimitrov",
"visual content. This paper proposes an en-",
"semble of Class Definition Prediction (CDP)",
"etal.,2024)着重于预测模因中使用的说服技",
"and hyperbolic embeddings-based approaches",
"术（来自视觉和文本内容），涉及四种不同的",
"for this task. We enhance meme classification",
"语言：英语，阿拉伯语，北马其顿语和保加利",
"accuracy and comprehensiveness by integrat-",
"亚语。该任务分为三个子任务：（1）仅使用模",
"ing HypEmo's hierarchical label embeddings",
"因的文本内容进行分层多标签分类，（2）使用",
"(Chen et al., 2023) and a multi-task learning",
"模因的文本和视觉内容进行分层多标签分类",
"framework for emotion prediction. We achieve",
"以及（3）二元分类模因是否包含说服技术或",
"a hierarchical F1-score of 0.60, 0.67, and 0.48",
"不使用其文本和视觉内容。为每个子任务提供",
"ontherespectivesub-tasks",
"训练数据，但仅以英语提供。提供了各种说服",
"1",
"Introduction",
"技术的分类（图2）及其各自的定义",
"今天，通过不同的社交媒体平台（Keswanietal.，",
"为了解决子任务1，我们采用了一种双重",
"2020;Singhetal.，2020），模因在所有年龄段的",
"方法，该方法涉及对每个类进行基于定义的建",
"人群中都很受欢迎。这些模因可以帮助人们了",
"模，并使用Chenetal.(2023）中提出的双曲线",
"解周围的趋势，并可以影响他们的决策。模因",
"嵌入进行分层分类。基于双曲线嵌入，该方法",
"是在人们之间传播虚假信息的流行模式之",
"通过利用层次结构来促进说服技术的细微分",
"图1中的示例），因为研究表明，人们倾向于",
"类。基于定义的建模的结合允许数据集不可知",
"相信他们在互联网上传播的此类模因中经常看",
"的方法，提高分类的精度，而不依赖于层次结",
"到的内容（Moravecetal.，2018）。正如2016年",
"构",
"美国总统竞选期间研究（Shuetal..2017）所证",
"对于子任务2，我们通过集成CLIP嵌入",
"明的那样，包括机器人，半机械人和巨魔在内",
"(Radfordetal.2021）来增强我们的方法，以从",
"的邪恶演员利用模因引起情绪反应并传播误导",
"模因的文本和视觉组件中捕获基本特征。这种",
"性叙述(Guo et al.,2020)",
"文本和视觉信息的融合可以更全面地分析模因"
] | {
"arxiv_id": "2404.04520",
"language": "zh",
"page_idx": 1
} |
||
2404.04520-zh-2 | 2404.04520-zh-2.jpg | <FigureHere>
Figure 2: Taxonomy of persuasion techniques for sub-task 2
内容
在处理子任务 3 时，我们采用了集成方法利用了来自 DistilBERT (Sanh et al.,2019) 和CLIP 嵌入(Radford et al., 2021)的迁移学习。这种集成技术通过融合来自两个预训练模型的见解来增强分类系统的鲁棒性和有效性。我们通过 GitHub 发布代码。$^{-1}$
## 2 $\textbf{Background}$
揭示了基于 n-gram 的方法的通用性的局限性。在此基础上，Barrón-Cedeno et al.(2019) 贡献了 QProp 语料库，该语料库专门针对宣传检测，采用宣传与非宣传的二元分类方案。同样，Habernal et al.(2018) 开发了一个带有谬误注释的语料库，包括与宣传技术直接相关的adhominem 和 redherring
基于 BERT 的变体已成为与语料库开发相结合的分类任务的有前途的方法。Yoosuf and Yang (2019) 提出了一种使用 BERT 进行世界后分类的微调方法，而 Fadel et al.(2019) 提出了一种集成 BiLSTM, BERT 和 RNN 组件的预训练集成模型。Costa et al.(2023) 进一步扩展了 BERT 的功能，提倡多语言设置，在使用RoBERTa 之前先翻译为英语。此外，Teimas and Saias (2023) 提出了一种将 CNN 与 DistilBERT 相结合的混合技术，以提高检测精度
探索多模态内容，Glenski et al. (2019) 深人研究了多语言多模态欺骗检测，主要关注仇恨模因。利用视觉和文本内容，他们利用 ViLBERT 和 VisualBERT 等最先进的模型以及基于迁移学习的方法 (Gupta et al., 2021) 进行微调技术
宣传的目标是增强人们的心态(Singh et al., 2020),尤其是在选举时，媒体的趋势会影响人们的投票(Shu et al.,2017)。宣传使用心理和修辞技巧来达到其目的。这样的方法包括使用逻辑谬误和吸引观众的情绪。逻辑上的谬误通常很难发现，因为乍一看，论证似乎是正确和客观的。但是，仔细分析表明，如果不滥用逻辑规则 (Gupta and Sharma,2021),就无法从前提中得出结论。另一组技术使用情感语言来诱使听众仅根据正在建立的情感纽带来与演讲者达成一致，从而引发了对论证(Szabo, 2020) 的任何理性分析的中止。
Corpora 的发展有助于推进欺骗检测方法。Rashkin et al.(2017)引人了 TSHP-17 语料库， 提供了跨四个类的文档级注释：可信，讽刺， 恶作剧和宣传。然而，他们对分类任务的研究
### 3 $\textbf{Data Description}$
比赛由两个不同的阶段组成，主要是开发阶段，我们称之为开发集，在开发阶段，我们提供了用于对我们的模型进行基准测试的训练和
$^{1}$https://github.com/Exploration-Lab/
IITK- SemEval- 2024- Task- 4- Pursuasion- Techniques | [
[
"0.1931",
"0.2322",
"0.2275",
"0.2322",
"0.2275",
"0.2393",
"0.1931",
"0.2393"
],
[
"0.4106",
"0.231",
"0.461",
"0.231",
"0.461",
"0.2411",
"0.4106",
"0.2411"
],
[
"0.771",
"0.2322",
"0.7925",
"0.2322",
"0.7925",
"0.2393",
"0.771",
"0.2393"
],
[
"0.5146",
"0.2369",
"0.552",
"0.2369",
"0.552",
"0.2446",
"0.5146",
"0.2446"
],
[
"0.294",
"0.2328",
"0.3477",
"0.2363",
"0.346",
"0.2465",
"0.293",
"0.2434"
],
[
"0.1931",
"0.2411",
"0.2351",
"0.2411",
"0.2351",
"0.2488",
"0.1931",
"0.2488"
],
[
"0.4072",
"0.2423",
"0.4702",
"0.2423",
"0.4702",
"0.2494",
"0.4072",
"0.2494"
],
[
"0.76",
"0.2434",
"0.812",
"0.2434",
"0.812",
"0.2477",
"0.76",
"0.2477"
],
[
"0.763",
"0.2625",
"0.7734",
"0.2625",
"0.7734",
"0.2654",
"0.763",
"0.2654"
],
[
"0.7793",
"0.2612",
"0.8096",
"0.2612",
"0.8096",
"0.2683",
"0.7793",
"0.2683"
],
[
"0.592",
"0.2715",
"0.644",
"0.2715",
"0.644",
"0.279",
"0.592",
"0.279"
],
[
"0.7715",
"0.2715",
"0.797",
"0.2715",
"0.797",
"0.2756",
"0.7715",
"0.2756"
],
[
"0.4224",
"0.2903",
"0.434",
"0.2903",
"0.434",
"0.2944",
"0.4224",
"0.2944"
],
[
"0.7715",
"0.2915",
"0.7793",
"0.2915",
"0.7793",
"0.2944",
"0.7715",
"0.2944"
],
[
"0.7793",
"0.2922",
"0.7983",
"0.2922",
"0.7983",
"0.2964",
"0.7793",
"0.2964"
],
[
"0.76",
"0.2976",
"0.81",
"0.2986",
"0.81",
"0.3071",
"0.76",
"0.306"
],
[
"0.6885",
"0.2957",
"0.7256",
"0.2957",
"0.7256",
"0.2998",
"0.6885",
"0.2998"
],
[
"0.771",
"0.308",
"0.7983",
"0.308",
"0.7983",
"0.3154",
"0.771",
"0.3154"
],
[
"0.1931",
"0.3164",
"0.2435",
"0.3164",
"0.2435",
"0.3237",
"0.1931",
"0.3237"
],
[
"0.2065",
"0.3267",
"0.2377",
"0.3267",
"0.2377",
"0.3308",
"0.2065",
"0.3308"
],
[
"0.7515",
"0.3552",
"0.818",
"0.3552",
"0.818",
"0.3665",
"0.7515",
"0.3665"
],
[
"0.754",
"0.3635",
"0.827",
"0.3635",
"0.827",
"0.3735",
"0.754",
"0.3735"
],
[
"0.7524",
"0.3718",
"0.837",
"0.3735",
"0.837",
"0.3855",
"0.7524",
"0.383"
],
[
"0.7524",
"0.3813",
"0.8296",
"0.383",
"0.8296",
"0.3943",
"0.7524",
"0.3926"
],
[
"0.7827",
"0.3926",
"0.837",
"0.3926",
"0.837",
"0.3997",
"0.7827",
"0.3997"
],
[
"0.764",
"0.401",
"0.7866",
"0.401",
"0.7866",
"0.408",
"0.764",
"0.408"
],
[
"0.785",
"0.402",
"0.817",
"0.402",
"0.817",
"0.4092",
"0.785",
"0.4092"
],
[
"0.2957",
"0.4316",
"0.7036",
"0.4316",
"0.7036",
"0.4453",
"0.2957",
"0.4453"
],
[
"0.1184",
"0.4714",
"0.157",
"0.4714",
"0.157",
"0.4868",
"0.1184",
"0.4868"
],
[
"0.513",
"0.4714",
"0.8774",
"0.4714",
"0.8774",
"0.4863",
"0.513",
"0.4863"
],
[
"0.1351",
"0.4893",
"0.4878",
"0.4893",
"0.4878",
"0.503",
"0.1351",
"0.503"
],
[
"0.511",
"0.4868",
"0.881",
"0.488",
"0.881",
"0.503",
"0.511",
"0.502"
],
[
"0.11755",
"0.505",
"0.4878",
"0.5044",
"0.4878",
"0.5186",
"0.11755",
"0.5195"
],
[
"0.5146",
"0.505",
"0.879",
"0.505",
"0.879",
"0.5186",
"0.5146",
"0.5186"
],
[
"0.1184",
"0.5215",
"0.4895",
"0.5215",
"0.4895",
"0.535",
"0.1184",
"0.535"
],
[
"0.513",
"0.521",
"0.881",
"0.521",
"0.881",
"0.535",
"0.513",
"0.535"
],
[
"0.1184",
"0.5376",
"0.486",
"0.5376",
"0.486",
"0.5503",
"0.1184",
"0.5503"
],
[
"0.511",
"0.535",
"0.881",
"0.536",
"0.881",
"0.5503",
"0.511",
"0.55"
],
[
"0.12006",
"0.5527",
"0.4878",
"0.5527",
"0.4878",
"0.568",
"0.12006",
"0.568"
],
[
"0.513",
"0.5527",
"0.881",
"0.5527",
"0.881",
"0.5664",
"0.513",
"0.5664"
],
[
"0.11755",
"0.5693",
"0.315",
"0.5693",
"0.315",
"0.583",
"0.11755",
"0.583"
],
[
"0.511",
"0.568",
"0.706",
"0.5684",
"0.706",
"0.583",
"0.511",
"0.582"
],
[
"0.532",
"0.586",
"0.881",
"0.586",
"0.881",
"0.5996",
"0.532",
"0.5996"
],
[
"0.1184",
"0.601",
"0.1335",
"0.601",
"0.1335",
"0.611",
"0.1184",
"0.611"
],
[
"0.1503",
"0.5996",
"0.2603",
"0.5996",
"0.2603",
"0.6143",
"0.1503",
"0.6143"
],
[
"0.511",
"0.601",
"0.881",
"0.601",
"0.881",
"0.6143",
"0.511",
"0.6143"
],
[
"0.511",
"0.6177",
"0.8774",
"0.6177",
"0.8774",
"0.6313",
"0.511",
"0.6313"
],
[
"0.1335",
"0.604",
"0.1427",
"0.604",
"0.1427",
"0.607",
"0.1335",
"0.607"
],
[
"0.11755",
"0.627",
"0.4844",
"0.627",
"0.4844",
"0.64",
"0.11755",
"0.64"
],
[
"0.513",
"0.633",
"0.879",
"0.633",
"0.879",
"0.648",
"0.513",
"0.648"
],
[
"0.11755",
"0.6436",
"0.486",
"0.6436",
"0.486",
"0.6567",
"0.11755",
"0.6567"
],
[
"0.531",
"0.651",
"0.881",
"0.651",
"0.881",
"0.6646",
"0.531",
"0.6646"
],
[
"0.11755",
"0.6577",
"0.4844",
"0.6577",
"0.4844",
"0.6714",
"0.11755",
"0.6714"
],
[
"0.513",
"0.6675",
"0.8774",
"0.6675",
"0.8774",
"0.681",
"0.513",
"0.681"
],
[
"0.1184",
"0.676",
"0.486",
"0.676",
"0.486",
"0.6895",
"0.1184",
"0.6895"
],
[
"0.5146",
"0.683",
"0.8774",
"0.683",
"0.8774",
"0.697",
"0.5146",
"0.697"
],
[
"0.12006",
"0.6914",
"0.486",
"0.6914",
"0.486",
"0.705",
"0.12006",
"0.705"
],
[
"0.511",
"0.6978",
"0.881",
"0.6987",
"0.881",
"0.7134",
"0.511",
"0.712"
],
[
"0.11755",
"0.707",
"0.4836",
"0.707",
"0.4836",
"0.72",
"0.11755",
"0.72"
],
[
"0.51",
"0.7144",
"0.881",
"0.7134",
"0.881",
"0.728",
"0.51",
"0.7285"
],
[
"0.1184",
"0.7227",
"0.4844",
"0.7227",
"0.4844",
"0.7363",
"0.1184",
"0.7363"
],
[
"0.513",
"0.731",
"0.8076",
"0.731",
"0.8076",
"0.7446",
"0.513",
"0.7446"
],
[
"0.12006",
"0.7393",
"0.4844",
"0.7393",
"0.4844",
"0.753",
"0.12006",
"0.753"
],
[
"0.529",
"0.749",
"0.881",
"0.7476",
"0.881",
"0.7627",
"0.529",
"0.763"
],
[
"0.11755",
"0.755",
"0.4844",
"0.756",
"0.4844",
"0.771",
"0.11755",
"0.7695"
],
[
"0.513",
"0.7656",
"0.881",
"0.7656",
"0.881",
"0.7793",
"0.513",
"0.7793"
],
[
"0.1184",
"0.7725",
"0.486",
"0.7725",
"0.486",
"0.786",
"0.1184",
"0.786"
],
[
"0.5146",
"0.781",
"0.882",
"0.781",
"0.882",
"0.7944",
"0.5146",
"0.7944"
],
[
"0.11755",
"0.788",
"0.4878",
"0.788",
"0.4878",
"0.802",
"0.11755",
"0.802"
],
[
"0.511",
"0.796",
"0.881",
"0.7964",
"0.881",
"0.811",
"0.511",
"0.81"
],
[
"0.11755",
"0.8047",
"0.3552",
"0.8047",
"0.3552",
"0.8184",
"0.11755",
"0.8184"
],
[
"0.5146",
"0.8135",
"0.869",
"0.8135",
"0.869",
"0.8267",
"0.5146",
"0.8267"
],
[
"0.1351",
"0.822",
"0.4844",
"0.822",
"0.4844",
"0.835",
"0.1351",
"0.835"
],
[
"0.11755",
"0.8374",
"0.4844",
"0.8374",
"0.4844",
"0.851",
"0.11755",
"0.851"
],
[
"0.5156",
"0.847",
"0.529",
"0.847",
"0.529",
"0.8555",
"0.5156",
"0.8555"
],
[
"0.5264",
"0.8486",
"0.5425",
"0.8486",
"0.5425",
"0.8555",
"0.5264",
"0.8555"
],
[
"0.5415",
"0.8496",
"0.5474",
"0.8496",
"0.5474",
"0.855",
"0.5415",
"0.855"
],
[
"0.5425",
"0.8457",
"0.6943",
"0.8457",
"0.6943",
"0.8594",
"0.5425",
"0.8594"
],
[
"0.11755",
"0.8525",
"0.4836",
"0.8525",
"0.4836",
"0.866",
"0.11755",
"0.866"
],
[
"0.1184",
"0.869",
"0.486",
"0.869",
"0.486",
"0.883",
"0.1184",
"0.883"
],
[
"0.5146",
"0.8745",
"0.881",
"0.8745",
"0.881",
"0.888",
"0.5146",
"0.888"
],
[
"0.1394",
"0.8955",
"0.4106",
"0.8955",
"0.4106",
"0.909",
"0.1394",
"0.909"
],
[
"0.5146",
"0.891",
"0.881",
"0.891",
"0.881",
"0.905",
"0.5146",
"0.905"
],
[
"0.11584",
"0.909",
"0.4685",
"0.909",
"0.4685",
"0.9194",
"0.11584",
"0.9194"
],
[
"0.513",
"0.9067",
"0.881",
"0.9067",
"0.881",
"0.9204",
"0.513",
"0.9204"
]
] | [
"Name",
"Appealto",
"Caus",
"Slogans",
"Bandwagon",
"Calling",
"Emotion(Visual)",
"ersimolificat",
"40",
"&Whit",
"Vagueness",
"Fallac",
"ad",
"h",
"ougn",
"Terminating",
"natabout",
"Cliche",
"Reductioad",
"Hitlerum",
"Color Legend",
"Black first level",
"Red = second Level",
"Blue = third level",
"=fourthleve",
"hite",
"techni",
"Figure 2: Taxonomy of persuasion techniques for sub-task 2",
"内容",
"揭示了基于n-gram的方法的通用性的局限性，",
"在处理子任务3时，我们采用了集成方法，",
"在此基础上，Barron-Cedeno et al.（2019）贡献",
"利用了来自DistilBERT(Sanhetal，2019）和",
"了QProp语料库，该语料库专门针对宣传检",
"CLIP嵌人(Radfordetal.，2021）的迁移学习。",
"测，采用宣传与非宣传的二元分类方案。同",
"这种集成技术通过融合来自两个预训练模型的",
"样，Habernaletal.(2018)开发了一个带有谬误",
"见解来增强分类系统的鲁棒性和有效性。我们",
"注释的语料库，包括与宣传技术直接相关的",
"通过GitHub发布代码。",
"adhominem和 redherring",
"基于BERT的变体已成为与语料库开发相",
"2.",
"Background",
"结合的分类任务的有前途的方法。Yoosuf and",
"Yang（2019）提出了一种使用BERT进行世界",
"",
"宣传的目标是增强人们的心态（Singhetal.",
"后分类的微调方法，而Fadeletal.(2019)提出",
"2020），尤其是在选举时，媒体的趋势会影",
"一种集成BLSTM，BERT和RNN组件的",
"响人们的投票（Shuetal.,2017）。宣传使用心",
"预训练集成模型。Costaetal.(2023）进一步扩",
"理和修辞技巧来达到其目的。这样的方法包括",
"展了BERT的功能，提倡多语言设置，在使用",
"使用逻辑谬误和吸引观众的情绪。逻辑上的谬",
"RoBERTa之前先翻译为英语。此外，Teimas and",
"误通常很难发现，因为乍一看，论证似乎是正",
"Saias（2023）提出了一种将CNN与DistilBERT",
"确和客观的。但是，仔细分析表明，如果不滥",
"相结合的混合技术，以提高检测精度",
"用逻辑规则（GuptaandSharma，2021），就无法",
"探索多模态内容，Glenskietal.(2019)深人研",
"从前提中得出结论。另一组技术使用情感语言",
"究了多语言多模态欺骗检测，主要关注仇恨模",
"来诱使听众仅根据正在建立的情感纽带来与",
"因。利用视觉和文本内容，他们利用ViLBERT",
"演讲者达成一致，从而引发了对论证(Szabo,",
"和VisualBERT等最先进的模型以及基于迁移",
"2020）的任何理性分析的中止。",
"学习的方法（Guptaetal,2021）进行微调技术",
"Corpora的发展有助于推进欺骗检测方法。",
"Rashkin et al.(2017)引入了TSHP-17语料库",
"3",
"",
"",
"Data Description",
"提供了跨四个类的文档级注释：可信，讽刺",
"恶作剧和宣传。然而，他们对分类任务的研究",
"比赛由两个不同的阶段组成，主要是开发阶",
"'https://github.com/Exploration-Lab/",
"段，我们称之为开发集，在开发阶段，我们提",
"IITK-SemEval-2024-Task-4-Pursuasion-Techniques",
"供了用于对我们的模型进行基准测试的训练和"
] | {
"arxiv_id": "2404.04520",
"language": "zh",
"page_idx": 2
} |
||
2404.04520-zh-3 | 2404.04520-zh-3.jpg | <table>
<tbody>
<tr>
<th>Persuasion Technique</th>
<th>Number mapped to</th>
</tr>
<tr>
<td>Presenting Irrelevant Data (Red Herring)</td>
<td>0</td>
</tr>
<tr>
<td>Bandwagon</td>
<td>I</td>
</tr>
<tr>
<td>Smears</td>
<td>2</td>
</tr>
<tr>
<td>Glittering generalities (Virtue)</td>
<td>3</td>
</tr>
<tr>
<td>Causal Oversimplification</td>
<td>4</td>
</tr>
<tr>
<td>Whataboutism</td>
<td>5</td>
</tr>
<tr>
<td>Loaded Language</td>
<td>6</td>
</tr>
<tr>
<td>Exaggeration/Minimisation</td>
<td>7</td>
</tr>
<tr>
<td>Repetition</td>
<td>8</td>
</tr>
<tr>
<td>Thoueht-terminatine cliché</td>
<td>9</td>
</tr>
<tr>
<td>Name calling/Labeling</td>
<td>10</td>
</tr>
<tr>
<td>Appeal to authority</td>
<td>ll</td>
</tr>
<tr>
<td>Black-and-white Fallacy/Dictatorship</td>
<td>12</td>
</tr>
<tr>
<td>Obfuscation, Intentional vagueness,</td>
<td> </td>
</tr>
<tr>
<td>Confusion (Straw Man)</td>
<td>$\left|\sum_{i=1}^{n}\frac{1}{2i}\right|$</td>
</tr>
<tr>
<td>Reductio ad hitlerum</td>
<td>14</td>
</tr>
<tr>
<td>Appeal to fear/prejudice</td>
<td>15</td>
</tr>
<tr>
<td>Misrepresentation of Someone's</td>
<td> </td>
</tr>
<tr>
<td>Position (Straw Man)</td>
<td>16</td>
</tr>
<tr>
<td>Flag-waving</td>
<td>17</td>
</tr>
<tr>
<td>Slogans</td>
<td>18</td>
</tr>
<tr>
<td>Doubt</td>
<td>19</td>
</tr>
</tbody>
</table>
Table 1: Dictionary Mapping for different persuasion
techniques for Subtask 1
<FigureHere>
Figure 3: The frequency Distribution of Labels in the
training dataset
<FigureHere>
Figure 4: The frequency Distribution of Labels in the
validation dataset
<FigureHere>
Figure 5: The meme sarcastically suggests that individuals who oppose Trump are being unfairly equated with terrorists, highlighting the absurdity of such comparisons. Two persuasion techniques are used: (i) Loaded Language , and (ii) Name calling that can be inferred from the text and the visual content.
验证集
所有三个子任务都有不同的模因集，分为训练、验证和开发集，如表 2 所示。我们还绘制了图 3 训练数据和图 4 验证数据中标签的分布
我们的分析使用字典将各种修辞技巧映射到
数值以进行绘制。这本词典如下：
### 4 $\textbf{System overview}$
针对所有子任务的拟议系统涉及对 BERT 模型和较早提出的工作进行的特定于任务的修改，包括 CLIP 模型 (Radford et al.,2021),基于类定义的情感预测 (Singh et al., 2021, 2023) 和 HypEmo 模型 (Chen et al.,2023) (如下所述)
# 4.1 Data Pre-processing
为了确保一致性和标准化，我们首先对文本进行预处理。这涉及删除换行符、逗号、数值和其他特殊字符。此外，整个文本转换为小写。在我们的方法中，我们利用开发(Dev)和训练集，只关注包含非零类的样本
### 4.2 Sub-task 1: Hierarchical Multi-label Text Classification
我们提出了一种新颖的模因分类方法，借鉴了两个关键框架的方法：HypEmo 和一个专注于情感定义建模的多任务学习模型
HypEmo (Chen et al., 2023) 利用预训练的标签双曲线嵌入来有效地捕获分层结构，特别是在树状结构中。最初，来自 RoBERTa 骨干模型的[CLS] 令牌的隐藏状态是使用多层感知器(MLP)投影的。随后，应用指数映射将其投影到双曲空间中。与预训练标签嵌入的距离是交叉熵损失函数的权重，增强了模型对标签关系的敏感性
为了实现 HypEmo 架构，我们将有向无环图(DAG)(图2)转换为树结构。这涉及复制具有多个父级的子级，从而为每个标签生成不同的嵌入。例如，将具有各种标签的句子转换为单独的样本，每个样本分配一个标签。利用具有 100 个维度的庞加莱双曲线蕴涵锥模型(Ganea et al.,2018),对构建的树进行训练，并 | [
[
"0.1663",
"0.0849",
"0.2585",
"0.0861",
"0.2585",
"0.0974",
"0.1663",
"0.0968"
],
[
"0.351",
"0.0861",
"0.4324",
"0.0861",
"0.4324",
"0.0974",
"0.351",
"0.0974"
],
[
"0.706",
"0.0861",
"0.7446",
"0.0861",
"0.7446",
"0.0903",
"0.706",
"0.0903"
],
[
"0.168",
"0.0974",
"0.3342",
"0.0998",
"0.3342",
"0.1134",
"0.168",
"0.111"
],
[
"0.5903",
"0.0944",
"0.6035",
"0.0944",
"0.6035",
"0.1051",
"0.5903",
"0.1051"
],
[
"0.3855",
"0.10095",
"0.3997",
"0.10095",
"0.3997",
"0.155",
"0.3855",
"0.155"
],
[
"0.1704",
"0.10925",
"0.2183",
"0.11224",
"0.2167",
"0.12054",
"0.1696",
"0.11755"
],
[
"0.588",
"0.1051",
"0.6055",
"0.1051",
"0.6055",
"0.1217",
"0.588",
"0.1217"
],
[
"0.1704",
"0.12054",
"0.2007",
"0.12054",
"0.2007",
"0.1247",
"0.1704",
"0.1247"
],
[
"0.2091",
"0.1247",
"0.293",
"0.1271",
"0.293",
"0.1372",
"0.2091",
"0.1354"
],
[
"0.5884",
"0.1223",
"0.602",
"0.1223",
"0.602",
"0.1342",
"0.5884",
"0.1342"
],
[
"0.575",
"0.1342",
"0.6055",
"0.1342",
"0.6055",
"0.1633",
"0.575",
"0.1633"
],
[
"0.1704",
"0.1271",
"0.2032",
"0.1271",
"0.2032",
"0.1342",
"0.1704",
"0.1342"
],
[
"0.1823",
"0.136",
"0.2761",
"0.136",
"0.2761",
"0.1437",
"0.1823",
"0.1437"
],
[
"0.1704",
"0.1437",
"0.2275",
"0.1437",
"0.2275",
"0.1509",
"0.1704",
"0.1509"
],
[
"0.1704",
"0.152",
"0.2393",
"0.152",
"0.2393",
"0.1592",
"0.1704",
"0.1592"
],
[
"0.3896",
"0.1526",
"0.3955",
"0.1526",
"0.3955",
"0.1609",
"0.3896",
"0.1609"
],
[
"0.1704",
"0.1621",
"0.2494",
"0.1621",
"0.2494",
"0.1663",
"0.1704",
"0.1663"
],
[
"0.1704",
"0.1704",
"0.2065",
"0.1704",
"0.2065",
"0.1746",
"0.1704",
"0.1746"
],
[
"0.3838",
"0.1663",
"0.4048",
"0.1663",
"0.4048",
"0.2131",
"0.3838",
"0.2131"
],
[
"0.1721",
"0.1787",
"0.1823",
"0.1787",
"0.1823",
"0.1829",
"0.1721",
"0.1829"
],
[
"0.1823",
"0.1799",
"0.2024",
"0.1799",
"0.2024",
"0.1829",
"0.1823",
"0.1829"
],
[
"0.2393",
"0.1758",
"0.2805",
"0.1758",
"0.2805",
"0.1859",
"0.2393",
"0.1859"
],
[
"0.1721",
"0.187",
"0.189",
"0.187",
"0.189",
"0.1912",
"0.1721",
"0.1912"
],
[
"0.1948",
"0.1882",
"0.2183",
"0.1882",
"0.2183",
"0.1912",
"0.1948",
"0.1912"
],
[
"0.215",
"0.1853",
"0.2627",
"0.1859",
"0.2627",
"0.1936",
"0.215",
"0.1924"
],
[
"0.1721",
"0.1965",
"0.1957",
"0.1965",
"0.1957",
"0.1996",
"0.1721",
"0.1996"
],
[
"0.2065",
"0.1953",
"0.2418",
"0.1953",
"0.2418",
"0.1996",
"0.2065",
"0.1996"
],
[
"0.1704",
"0.2019",
"0.189",
"0.2019",
"0.189",
"0.209",
"0.1704",
"0.209"
],
[
"0.2393",
"0.2007",
"0.3198",
"0.2031",
"0.3198",
"0.2102",
"0.2393",
"0.2079"
],
[
"0.706",
"0.209",
"0.754",
"0.209",
"0.754",
"0.212",
"0.706",
"0.212"
],
[
"0.754",
"0.209",
"0.7773",
"0.209",
"0.7773",
"0.212",
"0.754",
"0.212"
],
[
"0.1948",
"0.2114",
"0.315",
"0.2114",
"0.315",
"0.2185",
"0.1948",
"0.2185"
],
[
"0.1721",
"0.2114",
"0.1855",
"0.2114",
"0.1855",
"0.2173",
"0.1721",
"0.2173"
],
[
"0.2167",
"0.2197",
"0.2644",
"0.2197",
"0.2644",
"0.2268",
"0.2167",
"0.2268"
],
[
"0.3862",
"0.2144",
"0.4014",
"0.2144",
"0.4014",
"0.2245",
"0.3862",
"0.2245"
],
[
"0.1721",
"0.2214",
"0.1838",
"0.2214",
"0.1838",
"0.2245",
"0.1721",
"0.2245"
],
[
"0.1704",
"0.228",
"0.2024",
"0.228",
"0.2024",
"0.2322",
"0.1704",
"0.2322"
],
[
"0.2032",
"0.228",
"0.2418",
"0.228",
"0.2418",
"0.2322",
"0.2032",
"0.2322"
],
[
"0.3855",
"0.2239",
"0.403",
"0.2239",
"0.403",
"0.2434",
"0.3855",
"0.2434"
],
[
"0.1704",
"0.2351",
"0.2644",
"0.2351",
"0.2644",
"0.2423",
"0.1704",
"0.2423"
],
[
"0.511",
"0.234",
"0.882",
"0.234",
"0.882",
"0.2477",
"0.511",
"0.2477"
],
[
"0.1721",
"0.2446",
"0.1873",
"0.2446",
"0.1873",
"0.2488",
"0.1721",
"0.2488"
],
[
"0.189",
"0.2465",
"0.1957",
"0.2465",
"0.1957",
"0.2494",
"0.189",
"0.2494"
],
[
"0.2065",
"0.2452",
"0.2141",
"0.2452",
"0.2141",
"0.2488",
"0.2065",
"0.2488"
],
[
"0.3862",
"0.2465",
"0.4014",
"0.2465",
"0.4014",
"0.2583",
"0.3862",
"0.2583"
],
[
"0.511",
"0.2488",
"0.634",
"0.2488",
"0.634",
"0.2625",
"0.511",
"0.2625"
],
[
"0.3855",
"0.257",
"0.4048",
"0.257",
"0.4048",
"0.2861",
"0.3855",
"0.2861"
],
[
"0.1696",
"0.2588",
"0.2225",
"0.2588",
"0.2225",
"0.2695",
"0.1696",
"0.2695"
],
[
"0.1704",
"0.2683",
"0.199",
"0.2683",
"0.199",
"0.2756",
"0.1704",
"0.2756"
],
[
"0.168",
"0.2737",
"0.1973",
"0.2737",
"0.1973",
"0.2852",
"0.168",
"0.2852"
],
[
"0.1696",
"0.2517",
"0.2568",
"0.2517",
"0.2568",
"0.2588",
"0.1696",
"0.2588"
],
[
"0.623",
"0.2832",
"0.775",
"0.2852",
"0.775",
"0.3027",
"0.623",
"0.3005"
],
[
"0.11584",
"0.306",
"0.4844",
"0.3071",
"0.4844",
"0.3206",
"0.11584",
"0.3196"
],
[
"0.11584",
"0.3218",
"0.2854",
"0.3218",
"0.2854",
"0.335",
"0.11584",
"0.335"
],
[
"0.629",
"0.456",
"0.66",
"0.456",
"0.66",
"0.459",
"0.629",
"0.459"
],
[
"0.3074",
"0.4797",
"0.3914",
"0.4797",
"0.3914",
"0.484",
"0.3074",
"0.484"
],
[
"0.511",
"0.481",
"0.8833",
"0.481",
"0.8833",
"0.4946",
"0.511",
"0.4946"
],
[
"0.511",
"0.4954",
"0.882",
"0.4954",
"0.882",
"0.509",
"0.511",
"0.509"
],
[
"0.11584",
"0.5044",
"0.4895",
"0.505",
"0.4895",
"0.5195",
"0.11584",
"0.5186"
],
[
"0.51",
"0.509",
"0.882",
"0.5103",
"0.882",
"0.524",
"0.51",
"0.5225"
],
[
"0.11755",
"0.5215",
"0.2225",
"0.5215",
"0.2225",
"0.532",
"0.11755",
"0.532"
],
[
"0.511",
"0.524",
"0.882",
"0.524",
"0.882",
"0.5376",
"0.511",
"0.5376"
],
[
"0.511",
"0.538",
"0.882",
"0.538",
"0.882",
"0.552",
"0.511",
"0.552"
],
[
"0.511",
"0.5527",
"0.756",
"0.5527",
"0.756",
"0.5664",
"0.511",
"0.5664"
],
[
"0.11755",
"0.564",
"0.1755",
"0.564",
"0.1755",
"0.579",
"0.11755",
"0.579"
],
[
"0.1368",
"0.583",
"0.486",
"0.583",
"0.486",
"0.5967",
"0.1368",
"0.5967"
],
[
"0.513",
"0.594",
"0.5615",
"0.594",
"0.5615",
"0.6055",
"0.513",
"0.6055"
],
[
"0.555",
"0.5933",
"0.846",
"0.5933",
"0.846",
"0.607",
"0.555",
"0.607"
],
[
"0.11755",
"0.5996",
"0.4844",
"0.5996",
"0.4844",
"0.6133",
"0.11755",
"0.6133"
],
[
"0.12177",
"0.615",
"0.4844",
"0.615",
"0.4844",
"0.629",
"0.12177",
"0.629"
],
[
"0.552",
"0.61",
"0.7",
"0.61",
"0.7",
"0.6235",
"0.552",
"0.6235"
],
[
"0.1368",
"0.6343",
"0.4878",
"0.6343",
"0.4878",
"0.648",
"0.1368",
"0.648"
],
[
"0.511",
"0.6313",
"0.8774",
"0.6313",
"0.8774",
"0.6445",
"0.511",
"0.6445"
],
[
"0.1184",
"0.6494",
"0.3855",
"0.6494",
"0.3855",
"0.663",
"0.1184",
"0.663"
],
[
"0.513",
"0.648",
"0.879",
"0.6465",
"0.879",
"0.661",
"0.513",
"0.662"
],
[
"0.513",
"0.6646",
"0.7715",
"0.6646",
"0.7715",
"0.6777",
"0.513",
"0.6777"
],
[
"0.531",
"0.6797",
"0.881",
"0.6797",
"0.881",
"0.695",
"0.531",
"0.695"
],
[
"0.11584",
"0.686",
"0.1453",
"0.686",
"0.1453",
"0.6987",
"0.11584",
"0.6987"
],
[
"0.1511",
"0.686",
"0.2957",
"0.686",
"0.2957",
"0.6997",
"0.1511",
"0.6997"
],
[
"0.513",
"0.6953",
"0.881",
"0.697",
"0.881",
"0.7114",
"0.513",
"0.7104"
],
[
"0.1184",
"0.7173",
"0.486",
"0.7173",
"0.486",
"0.731",
"0.1184",
"0.731"
],
[
"0.511",
"0.712",
"0.881",
"0.7114",
"0.881",
"0.7256",
"0.511",
"0.727"
],
[
"0.511",
"0.728",
"0.881",
"0.7285",
"0.881",
"0.7437",
"0.511",
"0.742"
],
[
"0.1184",
"0.734",
"0.486",
"0.734",
"0.486",
"0.7476",
"0.1184",
"0.7476"
],
[
"0.51",
"0.7437",
"0.879",
"0.7446",
"0.879",
"0.759",
"0.51",
"0.7583"
],
[
"0.11755",
"0.75",
"0.4878",
"0.75",
"0.4878",
"0.763",
"0.11755",
"0.763"
],
[
"0.12006",
"0.7656",
"0.486",
"0.7656",
"0.486",
"0.7793",
"0.12006",
"0.7793"
],
[
"0.511",
"0.7603",
"0.879",
"0.7603",
"0.879",
"0.774",
"0.511",
"0.774"
],
[
"0.11755",
"0.781",
"0.4844",
"0.781",
"0.4844",
"0.7944",
"0.11755",
"0.7944"
],
[
"0.513",
"0.7754",
"0.881",
"0.777",
"0.881",
"0.7915",
"0.513",
"0.7905"
],
[
"0.5146",
"0.7944",
"0.588",
"0.7944",
"0.588",
"0.8057",
"0.5146",
"0.8057"
],
[
"0.1545",
"0.8154",
"0.3242",
"0.8164",
"0.3242",
"0.8306",
"0.1545",
"0.83"
],
[
"0.532",
"0.81",
"0.881",
"0.81",
"0.881",
"0.8237",
"0.532",
"0.8237"
],
[
"0.1184",
"0.818",
"0.1628",
"0.818",
"0.1628",
"0.829",
"0.1184",
"0.829"
],
[
"0.511",
"0.8247",
"0.881",
"0.826",
"0.881",
"0.8403",
"0.511",
"0.839"
],
[
"0.11755",
"0.843",
"0.4844",
"0.843",
"0.4844",
"0.8555",
"0.11755",
"0.8555"
],
[
"0.5146",
"0.843",
"0.882",
"0.843",
"0.882",
"0.8555",
"0.5146",
"0.8555"
],
[
"0.12006",
"0.86",
"0.4836",
"0.86",
"0.4836",
"0.8706",
"0.12006",
"0.8706"
],
[
"0.511",
"0.858",
"0.879",
"0.858",
"0.879",
"0.8716",
"0.511",
"0.8716"
],
[
"0.1184",
"0.8745",
"0.4844",
"0.8745",
"0.4844",
"0.888",
"0.1184",
"0.888"
],
[
"0.513",
"0.8745",
"0.881",
"0.8745",
"0.881",
"0.888",
"0.513",
"0.888"
],
[
"0.1184",
"0.891",
"0.4878",
"0.891",
"0.4878",
"0.905",
"0.1184",
"0.905"
],
[
"0.5156",
"0.8926",
"0.8774",
"0.8926",
"0.8774",
"0.903",
"0.5156",
"0.903"
],
[
"0.1184",
"0.9067",
"0.3762",
"0.9067",
"0.3762",
"0.9204",
"0.1184",
"0.9204"
],
[
"0.511",
"0.9067",
"0.881",
"0.9067",
"0.881",
"0.9204",
"0.511",
"0.9204"
]
] | [
"Persuasion Technique",
" Number mapped to",
"Distributio",
"Presenting rrelevant Data (Red Herring)",
"140",
"012345",
"Bandwagon",
"120",
"Smears",
"generalities (Virtue)",
"100",
"",
"Glitterin",
"usal Oversimplification",
"Whataboutism",
"Loaded Languag",
"6",
"Bxagaeration/Minir",
"Ranetitin",
"890UD",
"Th",
"Moh",
"ting cliche",
"Nam",
"n",
"ne/Labeling",
"nnas",
"authori",
"Blac",
"Fallacy/Dictatorship",
"",
"80",
"cation,Intentional vagueness,",
"Obf",
"StrawMan)",
"13",
"o",
"Radeti",
"adhirler",
"45",
"Appealtofear/prejudic",
"Figure 4: The frequency Distribution of Labels in the",
"Mist",
"P",
"ce",
"16",
"validation dataset",
"DB9",
"Flag-waving",
"Slogan",
"Doubt",
"Position (Straw Man)",
"HATETRUMP",
"Table 1: Dictionary Mapping for different persuasion",
"techniques for Subtask 1",
"",
"606",
"Figure 5: The meme sarcastically suggests that individ-",
"uals who oppose Trump are being unfairly equated with",
"Figure 3: The frequency Distribution of Labels in the",
"terrorists, highlighting the absurdity of such compar-",
"training dataset",
"isons. Two persuasion techniques are used: (i) Loaded",
"Language , and (ii) Name calling that can be inferred",
"from the text and the visual content.",
"验证集",
"所有三个子任务都有不同的模因集，分为训",
"4.2",
"Sub-task 1: Hierarchical Multi-label",
"练、验证和开发集，如表2所示。我们还绘制",
"了图3训练数据和图4验证数据中标签的分布",
"Text Classification",
"我们的分析使用字典将各种修辞技巧映射到",
"我们提出了一种新颖的模因分类方法，借鉴了",
"数值以进行绘制。这本词典如下：",
"两个关键框架的方法：HypEmo和一个专注于",
"情感定义建模的多任务学习模型",
"HypEmo（Chenetal.2023）利用预训练的标",
"4",
"System overview",
"签双曲线嵌入来有效地捕获分层结构，特别是",
"针对所有子任务的拟议系统涉及对BERT模",
"在树状结构中。最初，来自RoBERTa骨干模",
"型的[CLS]令牌的隐藏状态是使用多层感知器",
"型和较早提出的工作进行的特定于任务的修",
"(MLP）投影的。随后，应用指数映射将其投影",
"改，包括CLIP模型（Radfordetal.，2021），基",
"于类定义的情感预测（Singhetal.,2021，2023)",
"到双曲空间中。与预训练标签嵌入的距离是交",
"和HypEmo模型（Chenetal.，2023）（如下所述）",
"叉熵损失函数的权重，增强了模型对标签关系",
"的敏感性",
"Data Pre-processing",
"为了实现HypEmo架构，我们将有向无环",
"4.1",
"图（DAG）（图2）转换为树结构。这涉及复",
"为了确保一致性和标准化，我们首先对文本进",
"制具有多个父级的子级，从而为每个标签生成",
"行预处理。这涉及删除换行符、逗号、数值和",
"不同的嵌入。例如，将具有各种标签的句子转",
"其他特殊字符。此外，整个文本转换为小写。",
"换为单独的样本，每个样本分配一个标签。利",
"在我们的方法中，我们利用开发（Dev）和训",
"用具有100个维度的庞加莱双曲线蕴涵锥模型",
"练集，只关注包含非零类的样本",
"(Ganeaetal,2018），对构建的树进行训练，并"
] | {
"arxiv_id": "2404.04520",
"language": "zh",
"page_idx": 3
} |
||
2404.04520-zh-4 | 2404.04520-zh-4.jpg | <table>
<tbody>
<tr>
<th>Sub-task</th>
<th>Train Data</th>
<th>Validation Data</th>
<th>Development Data</th>
</tr>
<tr>
<td>Sub-task 1</td>
<td>7000</td>
<td>500</td>
<td>1000</td>
</tr>
<tr>
<td>$\operatorname{Sub-task}2$</td>
<td>7000</td>
<td>500</td>
<td>1000</td>
</tr>
<tr>
<td>Sub- task 3 19</td>
<td>1200</td>
<td>300</td>
<td>500</td>
</tr>
</tbody>
</table>
Table 2: Distribution of data across sub-tasks
<FigureHere>
Figure 6: Proposed architecture for sub-task 2
<FigureHere>
Figure 7: Proposed architecture for sub-task 3
通过 softmax 生成预测。通过与每个类别相关联的 Z 得分分析来识别峰值，并相应地设置阈值
Singh et al. (2021, 2023) 通过多任务学习框架引入了一种专注于情绪预测的补充方法。该模型结合了辅助任务，包括蒙面语言建模(MLM) 和类定义预测，以增强对情感概念的理解。在我们的设置中，使用[SEP]令牌合并类定义， 并训练模型以预测联合定义是否与实际定义匹配。此任务采用二进制交叉熵损失，并使用MLM 对模型进行微调。此外，在训练期间，每个类都使用二进制交叉熵损失。我们将模因分类竞赛提供的类定义用于类定义预测的辅助任务
最后，我们合并两个模型(HypEmo,基于细粒度类定义的模型)生成的预测以计算最终预测。这种综合方法旨在利用每个框架的优势， 提高模因分类结果的准确性和全面性
## 4.3 Sub-task 2: Hierarchical Multi-label Text and Image Classification
我们通过尝试将 HypEmo(Chen et al., 2023) 和基于类定义的多任务学习模型 (Singh et al., 2021,2023) 的集合用于模因的文本内容，并使用 CLIP 模型 (Radford et al., 2021)嵌人来从
模因的视觉内容中提取相关特征，从而对该子任务进行建模。我们为子任务 1 构造了一个类似的 DAG 结构，并生成双曲线嵌人。在发送组合特征向量进行训练之前，将从 CLIP 模型获得的图像嵌入与为文本内容生成的嵌入连接起来。然后，训练模型，并使用 softmax 激活函数生成预测。对得到的预测进行 Z-score 分析以进行分类，类似于任务 1。修改后的 HypEmo 模型的体系结构概述如图 7 所示
## 4.4 Sub-task 3: Binary Text and Image Classification
在此任务中，我们必须根据模因的文本和视觉内容对其是否包含说服技术进行分类。我们使用预训练的 BERT $_\mathrm{BASE}$模型(Devlin et al., 2019) 和卷积神经网络 (CNN) (O'Shea and Nash, 2015) 层分别从文本和图像中提取特征。我们附加了一个前馈$[CLS]$令牌嵌入以及由sigmoid激活函数连接的两个线性层，这将生成与 meme 中的文本内容相对应的句子嵌人。我们使用通过 ReLU 激活功能连接的四个CNN 层的网络，该功能从输入图像中逐步提取特征。最大池化层用于对特征图进行下采样，从而提高对微小变化的鲁棒性。将得到的图像嵌入与句子嵌人连接起来，并将线性分类器应用于具有sigmoid激活函数的组合特征向量。我们使用二进制交叉熵损失函数来训练模型并调整验证集上的超参数。图 8 中的示例显示了模型架构的概述
由于训练数据对于“有说服力的”(正，标记为 1)和“无说服力的”(负，标记为 0)类的比例为 2: 1, 这导致数据集中的不平衡，因此我们使用加权二进制交叉熵损失函数，如下 | [
[
"0.6724",
"0.08075",
"0.7256",
"0.0831",
"0.7256",
"0.0932",
"0.6724",
"0.09143"
],
[
"0.12177",
"0.0873",
"0.1738",
"0.0873",
"0.1738",
"0.0986",
"0.12177",
"0.0986"
],
[
"0.1957",
"0.0873",
"0.2627",
"0.0873",
"0.2627",
"0.0986",
"0.1957",
"0.0986"
],
[
"0.2703",
"0.0873",
"0.3645",
"0.0873",
"0.3645",
"0.0974",
"0.2703",
"0.0974"
],
[
"0.3728",
"0.0885",
"0.482",
"0.0885",
"0.482",
"0.0986",
"0.3728",
"0.0986"
],
[
"0.12177",
"0.10394",
"0.1823",
"0.10394",
"0.1823",
"0.1152",
"0.12177",
"0.1152"
],
[
"0.2124",
"0.10394",
"0.2435",
"0.10394",
"0.2435",
"0.1164",
"0.2124",
"0.1164"
],
[
"0.3032",
"0.10394",
"0.33",
"0.10394",
"0.33",
"0.1164",
"0.3032",
"0.1164"
],
[
"0.4148",
"0.10394",
"0.4458",
"0.10394",
"0.4458",
"0.1164",
"0.4148",
"0.1164"
],
[
"0.7036",
"0.10156",
"0.785",
"0.10156",
"0.785",
"0.1164",
"0.7036",
"0.1164"
],
[
"0.3015",
"0.1152",
"0.33",
"0.1152",
"0.33",
"0.1277",
"0.3015",
"0.1277"
],
[
"0.413",
"0.1152",
"0.4458",
"0.1152",
"0.4458",
"0.1277",
"0.413",
"0.1277"
],
[
"0.2107",
"0.1164",
"0.2435",
"0.1164",
"0.2435",
"0.1277",
"0.2107",
"0.1277"
],
[
"0.12006",
"0.1164",
"0.1838",
"0.1164",
"0.1838",
"0.1277",
"0.12006",
"0.1277"
],
[
"0.3032",
"0.1277",
"0.33",
"0.1277",
"0.33",
"0.1414",
"0.3032",
"0.1414"
],
[
"0.4148",
"0.1277",
"0.4417",
"0.1277",
"0.4417",
"0.1414",
"0.4148",
"0.1414"
],
[
"0.2124",
"0.1289",
"0.2435",
"0.1289",
"0.2435",
"0.1401",
"0.2124",
"0.1401"
],
[
"0.12177",
"0.13",
"0.1838",
"0.13",
"0.1838",
"0.1401",
"0.12177",
"0.1401"
],
[
"0.534",
"0.1526",
"0.5415",
"0.1526",
"0.5415",
"0.1561",
"0.534",
"0.1561"
],
[
"0.7944",
"0.1526",
"0.8096",
"0.1526",
"0.8096",
"0.1561",
"0.7944",
"0.1561"
],
[
"0.833",
"0.1526",
"0.846",
"0.1526",
"0.846",
"0.1561",
"0.833",
"0.1561"
],
[
"0.7793",
"0.2019",
"0.818",
"0.2019",
"0.818",
"0.209",
"0.7793",
"0.209"
],
[
"0.1486",
"0.1561",
"0.455",
"0.1561",
"0.455",
"0.1692",
"0.1486",
"0.1692"
],
[
"0.8296",
"0.2019",
"0.8564",
"0.2019",
"0.8564",
"0.2102",
"0.8296",
"0.2102"
],
[
"0.7773",
"0.2185",
"0.8228",
"0.2156",
"0.8237",
"0.2268",
"0.7793",
"0.2299"
],
[
"0.831",
"0.2203",
"0.8555",
"0.2162",
"0.86",
"0.2286",
"0.8364",
"0.2328"
],
[
"0.7793",
"0.2405",
"0.804",
"0.2405",
"0.804",
"0.2477",
"0.7793",
"0.2477"
],
[
"0.8486",
"0.2382",
"0.854",
"0.2382",
"0.854",
"0.2434",
"0.8486",
"0.2434"
],
[
"0.5786",
"0.257",
"0.623",
"0.257",
"0.623",
"0.2737",
"0.5786",
"0.2737"
],
[
"0.8096",
"0.2673",
"0.8486",
"0.2673",
"0.8486",
"0.2861",
"0.8096",
"0.2861"
],
[
"0.8193",
"0.2832",
"0.8555",
"0.288",
"0.852",
"0.2964",
"0.817",
"0.2915"
],
[
"0.806",
"0.2903",
"0.8154",
"0.2903",
"0.8154",
"0.2944",
"0.806",
"0.2944"
],
[
"0.5415",
"0.3206",
"0.854",
"0.3206",
"0.854",
"0.3342",
"0.5415",
"0.3342"
],
[
"0.5146",
"0.3682",
"0.879",
"0.3682",
"0.879",
"0.3818",
"0.5146",
"0.3818"
],
[
"0.5156",
"0.3843",
"0.881",
"0.3843",
"0.881",
"0.3984",
"0.5156",
"0.3984"
],
[
"0.4048",
"0.4038",
"0.43",
"0.4062",
"0.4282",
"0.4133",
"0.403",
"0.411"
],
[
"0.5146",
"0.401",
"0.879",
"0.401",
"0.879",
"0.4146",
"0.5146",
"0.4146"
],
[
"0.3997",
"0.411",
"0.4417",
"0.411",
"0.4417",
"0.4216",
"0.3997",
"0.4216"
],
[
"0.5156",
"0.4163",
"0.881",
"0.4163",
"0.881",
"0.43",
"0.5156",
"0.43"
],
[
"0.392",
"0.4258",
"0.4182",
"0.4258",
"0.4182",
"0.4329",
"0.392",
"0.4329"
],
[
"0.4207",
"0.4246",
"0.4475",
"0.427",
"0.4458",
"0.4353",
"0.419",
"0.4329"
],
[
"0.4114",
"0.4353",
"0.4265",
"0.4353",
"0.4265",
"0.4395",
"0.4114",
"0.4395"
],
[
"0.513",
"0.4329",
"0.881",
"0.4329",
"0.881",
"0.4465",
"0.513",
"0.4465"
],
[
"0.511",
"0.449",
"0.879",
"0.449",
"0.879",
"0.462",
"0.511",
"0.462"
],
[
"0.1453",
"0.4697",
"0.4575",
"0.4697",
"0.4575",
"0.4827",
"0.1453",
"0.4827"
],
[
"0.511",
"0.4644",
"0.8774",
"0.4644",
"0.8774",
"0.478",
"0.511",
"0.478"
],
[
"0.5146",
"0.481",
"0.881",
"0.481",
"0.881",
"0.4946",
"0.5146",
"0.4946"
],
[
"0.513",
"0.4954",
"0.7715",
"0.4963",
"0.7715",
"0.511",
"0.513",
"0.5103"
],
[
"0.11755",
"0.511",
"0.4844",
"0.5127",
"0.4844",
"0.527",
"0.11755",
"0.5254"
],
[
"0.1184",
"0.5293",
"0.486",
"0.5293",
"0.486",
"0.542",
"0.1184",
"0.542"
],
[
"0.513",
"0.535",
"0.5557",
"0.538",
"0.555",
"0.55",
"0.511",
"0.5464"
],
[
"0.555",
"0.535",
"0.8345",
"0.536",
"0.8345",
"0.5503",
"0.555",
"0.55"
],
[
"0.1184",
"0.5444",
"0.1394",
"0.5444",
"0.1394",
"0.558",
"0.1184",
"0.558"
],
[
"0.555",
"0.554",
"0.66",
"0.554",
"0.66",
"0.5654",
"0.555",
"0.5654"
],
[
"0.1351",
"0.561",
"0.4844",
"0.561",
"0.4844",
"0.5747",
"0.1351",
"0.5747"
],
[
"0.1184",
"0.5776",
"0.4844",
"0.5776",
"0.4844",
"0.5913",
"0.1184",
"0.5913"
],
[
"0.513",
"0.5806",
"0.881",
"0.5806",
"0.881",
"0.594",
"0.513",
"0.594"
],
[
"0.1184",
"0.5933",
"0.4878",
"0.5933",
"0.4878",
"0.607",
"0.1184",
"0.607"
],
[
"0.513",
"0.597",
"0.881",
"0.597",
"0.881",
"0.611",
"0.513",
"0.611"
],
[
"0.1184",
"0.61",
"0.4878",
"0.61",
"0.4878",
"0.6235",
"0.1184",
"0.6235"
],
[
"0.513",
"0.6133",
"0.881",
"0.6133",
"0.881",
"0.6274",
"0.513",
"0.6274"
],
[
"0.12006",
"0.6274",
"0.486",
"0.6274",
"0.486",
"0.638",
"0.12006",
"0.638"
],
[
"0.51",
"0.629",
"0.882",
"0.6274",
"0.882",
"0.6426",
"0.51",
"0.6436"
],
[
"0.1184",
"0.6436",
"0.482",
"0.6436",
"0.482",
"0.654",
"0.1184",
"0.654"
],
[
"0.51",
"0.6445",
"0.879",
"0.6445",
"0.879",
"0.6577",
"0.51",
"0.6577"
],
[
"0.11755",
"0.6577",
"0.4836",
"0.6577",
"0.4836",
"0.6714",
"0.11755",
"0.6714"
],
[
"0.513",
"0.661",
"0.879",
"0.661",
"0.879",
"0.676",
"0.513",
"0.676"
],
[
"0.11755",
"0.675",
"0.4844",
"0.675",
"0.4844",
"0.6885",
"0.11755",
"0.6885"
],
[
"0.513",
"0.6777",
"0.881",
"0.6777",
"0.881",
"0.6914",
"0.513",
"0.6914"
],
[
"0.12177",
"0.6924",
"0.4836",
"0.6924",
"0.4836",
"0.703",
"0.12177",
"0.703"
],
[
"0.513",
"0.6934",
"0.881",
"0.6934",
"0.881",
"0.707",
"0.513",
"0.707"
],
[
"0.1184",
"0.707",
"0.4844",
"0.707",
"0.4844",
"0.72",
"0.1184",
"0.72"
],
[
"0.5146",
"0.7104",
"0.879",
"0.7104",
"0.879",
"0.7236",
"0.5146",
"0.7236"
],
[
"0.1184",
"0.7246",
"0.1351",
"0.7246",
"0.1351",
"0.7354",
"0.1184",
"0.7354"
],
[
"0.5146",
"0.7256",
"0.881",
"0.7256",
"0.881",
"0.7393",
"0.5146",
"0.7393"
],
[
"0.1368",
"0.7407",
"0.4844",
"0.7407",
"0.4844",
"0.7544",
"0.1368",
"0.7544"
],
[
"0.5146",
"0.742",
"0.881",
"0.742",
"0.881",
"0.756",
"0.5146",
"0.756"
],
[
"0.11755",
"0.755",
"0.4844",
"0.755",
"0.4844",
"0.7686",
"0.11755",
"0.7686"
],
[
"0.513",
"0.7583",
"0.879",
"0.7583",
"0.879",
"0.7715",
"0.513",
"0.7715"
],
[
"0.1184",
"0.7725",
"0.4895",
"0.7725",
"0.4895",
"0.786",
"0.1184",
"0.786"
],
[
"0.5146",
"0.774",
"0.881",
"0.774",
"0.881",
"0.7876",
"0.5146",
"0.7876"
],
[
"0.11755",
"0.7876",
"0.4114",
"0.7876",
"0.4114",
"0.8003",
"0.11755",
"0.8003"
],
[
"0.5146",
"0.7915",
"0.876",
"0.7915",
"0.876",
"0.802",
"0.5146",
"0.802"
],
[
"0.5146",
"0.807",
"0.881",
"0.807",
"0.881",
"0.821",
"0.5146",
"0.821"
],
[
"0.11755",
"0.8184",
"0.451",
"0.8184",
"0.451",
"0.832",
"0.11755",
"0.832"
],
[
"0.513",
"0.8223",
"0.881",
"0.8223",
"0.881",
"0.836",
"0.513",
"0.836"
],
[
"0.1587",
"0.835",
"0.3914",
"0.835",
"0.3914",
"0.8486",
"0.1587",
"0.8486"
],
[
"0.513",
"0.839",
"0.679",
"0.839",
"0.679",
"0.8525",
"0.513",
"0.8525"
],
[
"0.11755",
"0.858",
"0.486",
"0.857",
"0.486",
"0.8716",
"0.11755",
"0.8726"
],
[
"0.531",
"0.858",
"0.882",
"0.858",
"0.882",
"0.8716",
"0.531",
"0.8716"
],
[
"0.1184",
"0.8745",
"0.4844",
"0.8745",
"0.4844",
"0.888",
"0.1184",
"0.888"
],
[
"0.513",
"0.8745",
"0.881",
"0.8745",
"0.881",
"0.888",
"0.513",
"0.888"
],
[
"0.11755",
"0.8896",
"0.4844",
"0.89",
"0.4844",
"0.905",
"0.11755",
"0.904"
],
[
"0.513",
"0.891",
"0.879",
"0.891",
"0.879",
"0.905",
"0.513",
"0.905"
],
[
"0.1184",
"0.9067",
"0.4878",
"0.9067",
"0.4878",
"0.9204",
"0.1184",
"0.9204"
],
[
"0.513",
"0.9067",
"0.879",
"0.9067",
"0.879",
"0.9204",
"0.513",
"0.9204"
]
] | [
"Predicted Clas",
"Sub-task",
"Train Data",
"ValidationData",
"DevelopmentData",
"Sub-task 1",
"7000",
"500",
"1000",
"for inay Casleato",
"500",
"1000",
"7000",
"Sub-task 2",
"300",
"500",
"1200",
"Sub-task3",
"N",
"ERT",
"ddi",
"TcusT,",
"Table 2: Distribution of data across sub-tasks",
"TN",
"Faa",
"5",
"as",
"",
"a",
"TAUATE",
"MOSISTS",
"TF",
"Figure 7: Proposed architecture for sub-task 3",
"模因的视觉内容中提取相关特征，从而对该子",
"任务进行建模。我们为子任务1构造了一个类",
"\"LHAT",
"似的DAG结构，并生成双曲线嵌入。在发送组",
"TRUMPW",
"合特征向量进行训练之前，将从CLIP模型获",
"TERR",
"ORISTS",
"DO",
"得的图像嵌人与为文本内容生成的嵌入连接起",
"来。然后，训练模型，并使用softmax激活函数",
"Figure 6: Proposed architecture for sub-task 2",
"生成预测。对得到的预测进行Z-score分析以",
"进行分类，类似于任务1。修改后的HypEmo",
"模型的体系结构概述如图7所示",
"通过softmax生成预测。通过与每个类别相关",
"联的Z得分分析来识别峰值，并相应地设置阈",
"4.4",
"Sub-task 3: Binary Text and Image",
"值",
"Classification",
"Singhetal.（2021,2023）通过多任务学习框架",
"引入了一种专注于情绪预测的补充方法。该模",
"在此任务中，我们必须根据模因的文本和视",
"型结合了辅助任务，包括蒙面语言建模（MLM)",
"觉内容对其是否包含说服技术进行分类。我们",
"和类定义预测，以增强对情感概念的理解。在",
"使用预训练的BERTBASE模型(Devlinetal.,",
"我们的设置中，使用ISEPI令牌合并类定义，",
"2019）和卷积神经网络（CNN）（O'Sheaand",
"并训练模型以预测联合定义是否与实际定义",
"Nash,2015)层分别从文本和图像中提取特征。",
"匹配。此任务采用二进制交叉熵损失，并使用",
"我们附加了一个前馈[CLS]令牌嵌入以及由",
"MLM对模型进行微调。此外，在训练期间，每",
"sigmoid激活函数连接的两个线性层，这将",
"个类都使用二进制交叉损失。我们将模因分",
"生成与meme中的文本内容相对应的句子嵌",
"类竞赛提供的类定义用于类定义预测的辅助任",
"人。我们使用通过ReLU激活功能连接的四个",
"务",
"CNN层的网络，该功能从输人图像中逐步提",
"最后，我们合并两个模型（HypEmo，基于细",
"取特征。最大池化层用于对特征图进行下采",
"粒度类定义的模型）生成的预测以计算最终预",
"样，从而提高对微小变化的鲁棒性。将得到的",
"测。这种综合方法旨在利用每个框架的优势，",
"图像嵌入与句子嵌人连接起来，并将线性分类",
"提高模因分类结果的准确性和全面性",
"器应用于具有sigmoid激活函数的组合特征向",
"量。我们使用二进制交叉摘损失函数来训练模",
"4.3 Sub-task 2: Hierarchical Multi-label",
"型并调整验证集上的超参数。图8中的示例显",
"Text and Image Classification",
"示了模型架构的概述",
"我们通过尝试将HypEmo（Chenetal.,2023）和",
"由于训练数据对于“有说服力的”（正，标",
"基于类定义的多任务学习模型（Singh etal.，",
"记为1）和“无说服力的”（负，标记为0）类",
"2021，2023）的集合用于模因的文本内容，并",
"的比例为2：1，这导致数据集中的不平衡，因",
"使用CLIP模型(Radfordetal.,2021）嵌人来从",
"此我们使用加权二进制交叉熵损失函数，如下"
] | {
"arxiv_id": "2404.04520",
"language": "zh",
"page_idx": 4
} |
||
2007.13638-zh-0 | 2007.13638-zh-0.jpg | # Message Passing Least Squares Framework and its Application to Rotation Synchronization
Yunpeng Shi$^1$Gilad Lerman$^1$
。3in
$\operatorname{Abstract}$
We propose an efficient algorithm for solving group synchronization under high levels of $\bar{\text{corruption and noise, while wefocuson rotation}}$ synchronization. We first describe our recent theoretically guaranteed message passing algorithm that estimates the corruption levels ofthe measured group ratios. We then propose a novel reweighted least squares method to estimate the group elements, where the weights areinitialized anditeratively updated using the estimated corruption levels. We demonstrate the superior performance of our algorithm over state-of-the-art methods for rotation synchronization using both synthetic and real data.
## 1. Introduction
组同步问题对于数据科学中的各种任务至关重要，包括运动结构(SfM),同时定位和映射(SLAM),低温电子显微镜成像，传感器网络定位，多对象匹配和社区检测。旋转同步，也称为旋转平均，是三维重建中最常见的组同步问题。它要求从测量的成对相机之间的相对旋转中恢复相机旋转。置换同步在多对象匹配中具有应用，它要求从某些对象对之间的匹配的可能错误测量中获得对象的全局一致匹配。组同步的最简单示例是$\mathbb{Z}_2$同步，它出现在社区检测中
组同步的一般问题可以用数学公式表示如下。假设图$G([n],E)$具有由$[n]=\{1,...,n\}$索引的$n$顶点，组$\mathcal{G}$和一组组元素$\{g_i^*\}_i=1^n\subseteq\mathcal{G}$ 。问题要求从组比率$\{g_{i}^{*}g_{j}^{*-1}\}_{ij\in E}$的嘈杂和损坏的测量$\{g_ij\}_ij\in E$中恢复$\{g_i^*\}_{i=1}^n$。我们注意到，只能将原始组元素$\left\{g_i^*\right\}_i\in[n]$恢复或近似为正确的组操作。实际上，对于任何$g_0\in\mathcal{G}$ , $g_{ij}^{* }$也可以写成$g_i^*g_0(g_j^*g_0)^{-1}$,因此$\{g_i^*g_0\}_{i\in[n]}$也是一种解决方案。上述同步问题(旋转，置换和$\mathbb{Z}_2$同步) 分别对应于组$SO( 3)$ , $S_N$ 和$\mathbb{Z}_2$
组同步最具挑战性的问题是高度损坏和嘈杂测量的实
际场景。在这种情况下，传统的最小二乘解算器通常
${^{1}\text{School of Mathematics, University of Minnesota, Min-}}$ neapolis, MN, USA . Correspondence to: Yunpeng Šhi < $\sin(x)=170$ixx$517@$umn.edu>,Gilad$\hat{\text{Lerman}}<$lerman$@$umn.edu>. Proceedings ofthe $37^th$ International Conferenceon Machine Learning, Online, PMLR 119, 2020. Copyright 2020 by the $\operatorname{author}(\overset{\smile}{\operatorname*{s}}).$
无法产生准确的结果。此外，一些似乎对腐败稳健的基本估计器在实践中往往不能容忍高水平的噪声。我们的目标是提出一种用于组同步的通用方法，该方法可以容忍高水平和不同类型的损坏和噪声。虽然我们的基本思想在形式上是通用的，但为了仔细地完善和测试它们，我们将重点放在旋转同步的特殊问题上， 该问题也称为旋转平均 (Hartley et al.,2013)。我们选择这个问题，因为它是3D计算机视觉中最常见，也可能是最困难的同步问题
## 1.1. Related Works
大多数以前的组同步求解器最小化能量函数。对于离散组$\mathbb{Z}_2$和$S_N$,通常使用最小二乘能量最小化。在Abbe et al. (2014);Abbe (2017);Bandeira (2018);Chen et al. (2014);Huang &Guibas (2013);Huroyan (2018); Pachauri et al. (2013)中讨论了在特殊的损坏和噪声模型下的相关鲁棒性结果
对于李群，例如$SO(D)$,即行列式为1的$D\times D$正交矩阵群，其中$D\geq2$,提出了最小二乘最小化来处理高斯噪声(Bandeiraetal.,2017;Erikssonetal.,2018;Govindu, 2004)。然而，当测量结果也被不利地破坏时，这个框架不能很好地工作，并且需要使用其他破坏鲁棒的能量函数(Chatterjee&Govindu,2013;2018;Hartleyetal. , 2011;Maunu&Lerman,2020;Wang&Singer,2013)。最常见的抗腐败能量函数使用最小绝对偏差。Wang $\&$ Singer(2013)证明，在具有$\mathcal{G}=SO(D)$的非常特殊的概率设置下，该能量函数的纯最小化器可以以高概率精确地恢复底层组元素。然而，他们的假设是强大的，他们使用凸松弛，这改变了原来的问题，是昂贵的计算。$\operatorname{Maunu }$Lerman$\left(2020\right)$应用修整的平均过程来稳健地求解$SO(2)$同步。他们能够在损坏的子图的拓扑上的特殊确定性条件下恢复地面真值组元素。但是，验证此条件及其对$SO(D)$的扩展(其中$D>2)$ 是不平凡的Hartley et al. (2011)使用 Weiszfeld 算法通过$\mathcal{G}=SO(3)$ 最小化最小绝对偏差能量函数。他们的方法迭代地计算测地线中位数。但是，它们每次迭代仅更新一个旋转矩阵，这导致经验收敛缓慢，并可能增加陷入局部最小值的可能性。Chatterjee&Govindu(2018)在$\mathcal{G}=SO(3)$ 上提出了一种鲁棒的李代数平均方法。他们在$SO(3)$ 的切线空间中应用迭代加权最小二乘 (IRLS) 程序来优化不同的鲁棒能量函数，包括使用最小绝对偏差的函数。他们声称使用$\ell_{1/2}$范数进行偏差会导致最高的经验准确性。即使在简单的环境中，后两个作品的经验稳健性在理论上也得不到保证。最近的一种深度学习方法 (Purkait et al.,2019) 解决了旋转同步的监督版 | [
[
"0.2778",
"0.11554",
"0.694",
"0.11554",
"0.694",
"0.13",
"0.2778",
"0.13"
],
[
"0.2524",
"0.1383",
"0.721",
"0.1383",
"0.721",
"0.1528",
"0.2524",
"0.1528"
],
[
"0.38",
"0.202",
"0.5923",
"0.2008",
"0.5923",
"0.2153",
"0.38",
"0.2166"
],
[
"0.08826",
"0.2172",
"0.1299",
"0.2172",
"0.1299",
"0.2317",
"0.08826",
"0.2317"
],
[
"0.2443",
"0.2299",
"0.3186",
"0.2299",
"0.3186",
"0.2443",
"0.2443",
"0.2443"
],
[
"0.5024",
"0.2311",
"0.884",
"0.2299",
"0.884",
"0.2443",
"0.5024",
"0.2456"
],
[
"0.1217",
"0.252",
"0.4412",
"0.252",
"0.4412",
"0.2651",
"0.1217",
"0.2651"
],
[
"0.5034",
"0.2474",
"0.8823",
"0.2474",
"0.8823",
"0.2583",
"0.5034",
"0.2583"
],
[
"0.5034",
"0.26",
"0.8857",
"0.26",
"0.8857",
"0.2734",
"0.5034",
"0.2734"
],
[
"0.12335",
"0.2664",
"0.4395",
"0.2651",
"0.4395",
"0.279",
"0.12335",
"0.2798"
],
[
"0.1209",
"0.279",
"0.4387",
"0.279",
"0.4387",
"0.2922",
"0.1209",
"0.2922"
],
[
"0.5024",
"0.2734",
"0.884",
"0.2734",
"0.884",
"0.2874",
"0.5024",
"0.2874"
],
[
"0.1217",
"0.2942",
"0.4395",
"0.2942",
"0.4395",
"0.308",
"0.1217",
"0.308"
],
[
"0.5034",
"0.2903",
"0.884",
"0.2903",
"0.884",
"0.3005",
"0.5034",
"0.3005"
],
[
"0.1209",
"0.308",
"0.4412",
"0.3093",
"0.4412",
"0.324",
"0.1209",
"0.3225"
],
[
"0.5034",
"0.305",
"0.8813",
"0.305",
"0.8813",
"0.315",
"0.5034",
"0.315"
],
[
"0.5024",
"0.3176",
"0.884",
"0.3176",
"0.884",
"0.3308",
"0.5024",
"0.3308"
],
[
"0.12335",
"0.3245",
"0.4395",
"0.3245",
"0.4395",
"0.3352",
"0.12335",
"0.3352"
],
[
"0.1209",
"0.336",
"0.4395",
"0.3372",
"0.4395",
"0.3516",
"0.1209",
"0.351"
],
[
"0.5024",
"0.332",
"0.884",
"0.332",
"0.884",
"0.3452",
"0.5024",
"0.3452"
],
[
"0.1209",
"0.3516",
"0.4412",
"0.3516",
"0.4412",
"0.3655",
"0.1209",
"0.3655"
],
[
"0.5024",
"0.3467",
"0.664",
"0.3467",
"0.664",
"0.3599",
"0.5024",
"0.3599"
],
[
"0.12335",
"0.3674",
"0.437",
"0.3674",
"0.437",
"0.3774",
"0.12335",
"0.3774"
],
[
"0.1209",
"0.379",
"0.4395",
"0.3801",
"0.4395",
"0.3945",
"0.1209",
"0.3933"
],
[
"0.501",
"0.377",
"0.634",
"0.377",
"0.634",
"0.3901",
"0.501",
"0.3901"
],
[
"0.12335",
"0.3958",
"0.4395",
"0.3958",
"0.4395",
"0.4092",
"0.12335",
"0.4092"
],
[
"0.5034",
"0.3997",
"0.8857",
"0.3997",
"0.8857",
"0.4136",
"0.5034",
"0.4136"
],
[
"0.12335",
"0.411",
"0.4387",
"0.411",
"0.4387",
"0.4216",
"0.12335",
"0.4216"
],
[
"0.5024",
"0.414",
"0.8857",
"0.414",
"0.8857",
"0.428",
"0.5024",
"0.428"
],
[
"0.1217",
"0.4236",
"0.437",
"0.4248",
"0.437",
"0.4363",
"0.1217",
"0.435"
],
[
"0.501",
"0.428",
"0.884",
"0.4287",
"0.884",
"0.4426",
"0.501",
"0.4414"
],
[
"0.1217",
"0.4382",
"0.4192",
"0.4382",
"0.4192",
"0.452",
"0.1217",
"0.452"
],
[
"0.5024",
"0.4426",
"0.884",
"0.4426",
"0.884",
"0.4558",
"0.5024",
"0.4558"
],
[
"0.5024",
"0.457",
"0.8857",
"0.457",
"0.8857",
"0.4702",
"0.5024",
"0.4702"
],
[
"0.5024",
"0.473",
"0.6665",
"0.473",
"0.6665",
"0.486",
"0.5024",
"0.486"
],
[
"0.08905",
"0.481",
"0.2181",
"0.481",
"0.2181",
"0.4944",
"0.08905",
"0.4944"
],
[
"0.5024",
"0.4937",
"0.8857",
"0.4937",
"0.8857",
"0.507",
"0.5024",
"0.507"
],
[
"0.0907",
"0.506",
"0.4714",
"0.506",
"0.4714",
"0.5195",
"0.0907",
"0.5195"
],
[
"0.501",
"0.507",
"0.8857",
"0.5083",
"0.8857",
"0.5215",
"0.501",
"0.521"
],
[
"0.0907",
"0.5195",
"0.4739",
"0.5195",
"0.4739",
"0.5327",
"0.0907",
"0.5327"
],
[
"0.5024",
"0.5215",
"0.884",
"0.5215",
"0.884",
"0.535",
"0.5024",
"0.535"
],
[
"0.0907",
"0.534",
"0.4722",
"0.534",
"0.4722",
"0.5474",
"0.0907",
"0.5474"
],
[
"0.5024",
"0.5386",
"0.884",
"0.5386",
"0.884",
"0.549",
"0.5024",
"0.549"
],
[
"0.08905",
"0.5474",
"0.4739",
"0.549",
"0.4739",
"0.5625",
"0.08905",
"0.561"
],
[
"0.5",
"0.55",
"0.884",
"0.55",
"0.884",
"0.563",
"0.5",
"0.563"
],
[
"0.09314",
"0.5645",
"0.4714",
"0.5645",
"0.4714",
"0.5747",
"0.09314",
"0.5747"
],
[
"0.501",
"0.5645",
"0.88",
"0.565",
"0.88",
"0.579",
"0.501",
"0.5776"
],
[
"0.09314",
"0.579",
"0.4714",
"0.579",
"0.4714",
"0.589",
"0.09314",
"0.589"
],
[
"0.5",
"0.579",
"0.8857",
"0.58",
"0.8857",
"0.5933",
"0.5",
"0.5923"
],
[
"0.09314",
"0.5933",
"0.4714",
"0.5933",
"0.4714",
"0.604",
"0.09314",
"0.604"
],
[
"0.5024",
"0.5933",
"0.8857",
"0.5933",
"0.8857",
"0.607",
"0.5024",
"0.607"
],
[
"0.0907",
"0.606",
"0.4739",
"0.606",
"0.4739",
"0.619",
"0.0907",
"0.619"
],
[
"0.5024",
"0.61",
"0.884",
"0.61",
"0.884",
"0.6206",
"0.5024",
"0.6206"
],
[
"0.09235",
"0.622",
"0.3882",
"0.622",
"0.3882",
"0.632",
"0.09235",
"0.632"
],
[
"0.5024",
"0.6226",
"0.884",
"0.6226",
"0.884",
"0.636",
"0.5024",
"0.636"
],
[
"0.501",
"0.636",
"0.884",
"0.637",
"0.884",
"0.651",
"0.501",
"0.6494"
],
[
"0.0907",
"0.6426",
"0.4739",
"0.6426",
"0.4739",
"0.656",
"0.0907",
"0.656"
],
[
"0.09235",
"0.656",
"0.4714",
"0.656",
"0.4714",
"0.6694",
"0.09235",
"0.6694"
],
[
"0.5034",
"0.653",
"0.8823",
"0.653",
"0.8823",
"0.6636",
"0.5034",
"0.6636"
],
[
"0.5",
"0.664",
"0.8857",
"0.6655",
"0.8857",
"0.6787",
"0.5",
"0.678"
],
[
"0.08905",
"0.6704",
"0.4739",
"0.6704",
"0.4739",
"0.6846",
"0.08905",
"0.6846"
],
[
"0.09235",
"0.685",
"0.4722",
"0.685",
"0.4722",
"0.6997",
"0.09235",
"0.6997"
],
[
"0.5024",
"0.6816",
"0.8823",
"0.6816",
"0.8823",
"0.6924",
"0.5024",
"0.6924"
],
[
"0.09314",
"0.7007",
"0.4714",
"0.7007",
"0.4714",
"0.7153",
"0.09314",
"0.7153"
],
[
"0.5024",
"0.696",
"0.884",
"0.696",
"0.884",
"0.7056",
"0.5024",
"0.7056"
],
[
"0.5034",
"0.7104",
"0.8823",
"0.7104",
"0.8823",
"0.72",
"0.5034",
"0.72"
],
[
"0.0907",
"0.7163",
"0.4722",
"0.7163",
"0.4722",
"0.73",
"0.0907",
"0.73"
],
[
"0.5",
"0.7217",
"0.884",
"0.7217",
"0.884",
"0.735",
"0.5",
"0.735"
],
[
"0.08905",
"0.729",
"0.4739",
"0.728",
"0.4739",
"0.7456",
"0.08905",
"0.747"
],
[
"0.501",
"0.7373",
"0.884",
"0.7373",
"0.884",
"0.7505",
"0.501",
"0.7505"
],
[
"0.1013",
"0.747",
"0.4739",
"0.747",
"0.4739",
"0.7603",
"0.1013",
"0.7603"
],
[
"0.5024",
"0.752",
"0.884",
"0.752",
"0.884",
"0.765",
"0.5024",
"0.765"
],
[
"0.0907",
"0.761",
"0.3145",
"0.761",
"0.3145",
"0.7744",
"0.0907",
"0.7744"
],
[
"0.5024",
"0.765",
"0.884",
"0.7666",
"0.884",
"0.78",
"0.5024",
"0.779"
],
[
"0.09235",
"0.784",
"0.4722",
"0.784",
"0.4722",
"0.7944",
"0.09235",
"0.7944"
],
[
"0.5034",
"0.782",
"0.884",
"0.782",
"0.884",
"0.7925",
"0.5034",
"0.7925"
],
[
"0.0907",
"0.7974",
"0.4739",
"0.7974",
"0.4739",
"0.811",
"0.0907",
"0.811"
],
[
"0.501",
"0.7944",
"0.8857",
"0.7944",
"0.8857",
"0.808",
"0.501",
"0.808"
],
[
"0.5024",
"0.8096",
"0.884",
"0.8096",
"0.884",
"0.8228",
"0.5024",
"0.8228"
],
[
"0.10864",
"0.8193",
"0.4739",
"0.821",
"0.4739",
"0.834",
"0.10864",
"0.8335"
],
[
"0.5024",
"0.8237",
"0.884",
"0.8237",
"0.884",
"0.837",
"0.5024",
"0.837"
],
[
"0.0907",
"0.8335",
"0.4722",
"0.8335",
"0.4722",
"0.8467",
"0.0907",
"0.8467"
],
[
"0.5024",
"0.8384",
"0.884",
"0.8384",
"0.884",
"0.8516",
"0.5024",
"0.8516"
],
[
"0.08905",
"0.8447",
"0.4697",
"0.8447",
"0.4697",
"0.858",
"0.08905",
"0.858"
],
[
"0.501",
"0.853",
"0.884",
"0.853",
"0.884",
"0.866",
"0.501",
"0.866"
],
[
"0.08905",
"0.8687",
"0.4722",
"0.8687",
"0.4722",
"0.882",
"0.08905",
"0.882"
],
[
"0.5034",
"0.869",
"0.8823",
"0.869",
"0.8823",
"0.88",
"0.5034",
"0.88"
],
[
"0.08905",
"0.881",
"0.4722",
"0.881",
"0.4722",
"0.8945",
"0.08905",
"0.8945"
],
[
"0.5034",
"0.884",
"0.8823",
"0.884",
"0.8823",
"0.8945",
"0.5034",
"0.8945"
],
[
"0.0866",
"0.89",
"0.1544",
"0.8926",
"0.1544",
"0.907",
"0.0866",
"0.905"
]
] | [
"Message Passing Least Squares Framework",
"and itsApplicationtoRotation Synchronization",
"Yunpeng Shi1 Gilad Lerman l",
"。3in",
"Abstract",
"无法产生准确的结果。此外，一些似乎对腐败稳健的",
"We propose an efficient algorithm for solving",
"基本估计器在实践中往往不能容忍高水平的噪声。我",
"们的目标是提出一种用于组同步的通用方法，该方法",
"group synchronization under high levels of",
"corruptionandnoise,whilewefocusonrotation",
"可以容忍高水平和不同类型的损坏和噪声。虽然我们",
"synchronization. We first describe our recent",
"的基本思想在形式上是通用的，但为了仔细地完善和",
"theoretically guaranteed message passing",
"测试它们，我们将重点放在旋转同步的特殊问题上",
"该问题也称为旋转平均（Hartleyetal.,2013）。我们选",
"algorithmthatestimatesthecorruptionlevels",
"of the measured groupratios. We then propose",
"择这个问题，因为它是3D计算机视觉中最常见，也可",
"a novel reweighted least squares method to",
"能是最困难的同步问题",
"estimatethegroupelements.wheretheweights",
"areinitializedanditerativelyupdatedusingthe",
"1.1. Related Works",
"estimated corruption levels.We demonstrate",
"大多数以前的组同步求解器最小化能量函数。对于离",
"thesuperiorperformanceofouralgorithmover",
"散组Z2和SN，通常使用最小二乘能量最小化。在",
"state-of-the-art methodsforrotation synchro",
"Abbeetal.(2014);Abbe(2017);Bandeira (2018);Chen",
"nization using both synthetic and real data.",
"et al. (2014); Huang & Guibas (2013); Huroyan (2018):",
"Pachaurietal.（2013)中讨论了在特殊的损坏和噪声模",
"型下的相关鲁棒性结果",
"1.Introduction",
"对于李群，例如SO(D)，即行列式为1的D×D正交矩",
"组同步问题对于数据科学中的各种任务至关重要，包",
"阵群，其中D≥2，提出了最小二乘最小化来处理高斯",
"括运动结构（SfM），同时定位和映射（SLAM），低",
"噪声(Bandeiraetal.,2017;Erikssonetal.,2018Govindu,",
"温电子显微镜成像，传感器网络定位，多对象匹配和",
"2004）。然而，当测量结果也被不利地破坏时，这个框",
"社区检测。旋转同步，也称为旋转平均，是三维重建",
"架不能很好地工作，并且需要使用其他破坏鲁棒的能",
"中最常见的组同步问题。它要求从测量的成对相机之",
"量函数（Chatterjee&Govindu，2013;2018;Hartleyetal.",
"间的相对旋转中恢复相机旋转。置换同步在多对象匹",
"2011;Maunu&Lerman,2020;Wang&Singer,2013)。最",
"配中具有应用，它要求从某些对象对之间的匹配的可",
"常见的抗腐败能量函数使用最小绝对偏差。Wang&",
"能错误测量中获得对象的全局一致匹配。组同步的最",
"Singer（2013）证明，在具有g=SO（D）的非常特殊的概",
"简单示例是么2同步，它出现在社区检测中",
"率设置下，该能量函数的纯最小化器可以以高概率精",
"确地恢复底层组元素。然而，他们的假设是强大的，他",
"组同步的一般问题可以用数学公式表示如下。假设",
"图G([n],E)具有由[nl={1,..,n)索引的n顶点，组",
"们使用凸松弛，这改变了原来的问题，是昂贵的计算，",
"Maunu&Lerman（2020）应用修整的平均过程来稳健地",
"g和一组组元素(g,)n-,C9。问题要求从组比率",
"Lgig-llieE的噜杂和损坏的测量(gilijeE中恢复",
"求解SO（2）同步。他们能够在损坏的子图的拓扑上的",
"(g,n-1。我们注意到，只能将原始组元素（g,]i[n]恢",
"特殊确定性条件下恢复地面真值组元素。但是，验证",
"此条件及其对SO（D）的扩展（其中D>2）是不平凡的",
"复或近似为正确的组操作。实际上，对于任何9oE9",
"Hartleyetal.（2011)使用Weiszfeld算法通过g=SO(3）",
"，gi,也可以写成gigo(ggo)-，因此(ggo}i[n]也是",
"最小化最小绝对偏差能量函数。他们的方法送代地计",
"一种解决方案。上述同步问题（旋转，置换和Z2同步）",
"算测地线中位数。但是，它们每次送代仅更新一个旋转",
"分别对应于组SO(3)，SN和Z2",
"矩阵，这导致经验收敛缓慢，并可能增加陷入局部最小",
"组同步最具挑战性的问题是高度损坏和噜杂测量的实",
"值的可能性。Chatteriee&Govindu（2018）在G=SO3）",
"际场景。在这种情况下，传统的最小二乘解算器通常",
"上提出了一种鲁棒的李代数平均方法。他们在SO(3)",
"的切线空间中应用送代加权最小二乘（IRLS）程序来",
"1 School of Mathematics, University of Minnesota, Min-",
"优化不同的鲁棒能量函数，包括使用最小绝对偏差的",
"neapolis, MN, USA . Correspondence to: Yunpeng Shi <",
"函数。他们声称使用11/2范数进行偏差会导致最高的",
"[email protected]>, Gilad Lerman<[email protected]>",
"经验准确性。即使在简单的环境中，后两个作品的经",
"Proceedings of the 37th InternationalConferenceon Machine",
"验稳健性在理论上也得不到保证。最近的一种深度学",
"Learning, Online, PMLR 119, 2020.Copyright 2020 by the",
"习方法（Purkaitetal.，2019）解决了旋转同步的监督版",
"author(s)"
] | {
"arxiv_id": "2007.13638",
"language": "zh",
"page_idx": 0
} |
||
2202.06281-zh-0 | 2202.06281-zh-0.jpg | XMATHX\_LIN7<NEXXT\_LINE。这些任务中的数据通常在欧几里得空间中表示。但是，越来越多的应用程序从非欧几里得域生成数据，并将其表示为对象之间具有复杂关系的图形$\overline{\mathbb{E}}[5,9,21,23,34,39,40]$。通过定义图上的卷积算子，图神经网络(GNN)将卷积神经网络 (CNN) 从图像域扩展到图域。因此，我们仍然可以将可能的工具从 CNN 调整为 GNNs、$\iota$.e.、ReLU[27]、Dropout[31] 或剩余链接[13]
GNN 现在能够通过采用所谓的消息传递(MP) 框架来挖掘图的特征，该框架迭代地聚合和更新沿边路径的节点信息。将 MP 作为建模工具，近年来在下游任务中取得了前所未有的成就，例如节点/图分类E[29,30] 和链接预测E[3,4]。然而，最近的许多工作证明，MP 中设计不当的传播步骤违反了许多现实世界图所持有的同构假设，从而导致关键性能下降$\overline{\mathrm{E}}[15,44]$
上述麻烦促使我们改进 MP 并增强 GNN 的泛化能力。一个有前途的探索路线是动态考虑 MP 期间每个节点的贡献。注意机制是一个很好的指针，因为它们已经在[15,33] 中进行了探索。注意机制的好处之一是，它们通过将权重相应地分配给不同的节点，使网络专注于要传播的最相关的信息。$e$g.,作品$\overline{\mathrm{E}}[15,33]$ 表明，通过自适应转换传播信息会在多个归纳基准上产生令人印象深刻的性能
但是，当前的自适应转换设计效率低下。我们的推理源于两个关键的观点。首先，在推断每个输人节点的自适应权重时没有考虑拓扑信息，因此无法捕获图[15] 的同态信息。其次，我们怀疑在提高模型泛化能力和准确性方面的局限性是由于为每个输入节点共享跨通道的自适应权重。可能不希望在没有区别的情况下转换所有通道，因为某些通道不如其他通道重要
受上述问题的启发，我们开发了GReLU，这是一种拓扑感知的自适应激活功能，适用于任何 GNN主干。GReLU 优化了信息传播并提高了节点分类任务的性能。如图 ?? 所示，我们的机制由两个组件组成，改进了GNN 的训练过程。一方面，GReLU 通过将特征和图形作为输人来继承拓扑信息。另一方面，与标准整流器是相同的函数 wr.t.所有输人不同，GReLU 分配一组参数来动态控制每个节点和通道的分段非线性函数。这种拓扑自适应属性提供了比拓扑未知对应物更好的表达能力。同时，GReLU 是计算高效的$e$.g.,具有 GReLU 的两层 GCN 和两层 GAT$\mathbb{E}[33]$具有相似的计算复杂度，而前者在节点分类方面能够优于 GAT
这项工作的主要贡献如下：
i. We incorporate the topology information in the updating function of GNNs by introducing a topology-aware activation function in
the updating stage of message passing, which, to our knowledge, is novel.
ii. We propose a well-designed graph adaptive rectified linear unit for graph data, which can be used as a plug-and-play component in
various types of spatial and spectral GNNs.
iii. We evaluate the proposed methods on various downstream tasks. We empirically show that the performance of various types of GNNs
can be generally improved with the use of our GReLU.
1 RELATED WORKS
1.1 Graph Neural Networks
GNN 可以分为空间或。谱域从图信号处理$i.e.$的角度探索 GNNs 设计，利用傅里叶变换和拉普拉斯矩阵 [7] 的本征分解定义卷积步长。可以提取任何频率来设计具有理论保证的图形滤波器。但是，这些模型存在三个主要缺点：1) 给定图的本征分解引起大量计算负担；2) 空间非局部化；3) 为给定图设计的过滤器不能应用于其他图。空间 GNNs 通过引人消息传递机制来概括 GNNs 的体系结构，该机制是根据图的一些空间特征 (例如邻接矩阵，节点特征，边缘特征，etc.$\operatorname{E}[11]$)聚合邻居信息。由于其高效率和可扩展性，空间 GNN 在最近的工作中变得更加流行。最具代表性的模型是图注意网络 (GAT)E[ 33] ,它通过在聚合步骤中使用注意机制来加权不同邻居的节点特征。但是，对于每个节点功能，GAT为节点分配了跨所有通道共享的单个权重，这忽略了不同通道可能表现出不同重要性的事实
1.2 Activation Functions
激活函数已经在神经网络的整个历史上进行了长时间的研究。少数重要的里程碑之一是将整流线性单元 (ReLU)$\overline{\mathbb{E}}[27]$引人神经网络。由于其简单性和非线性，ReLU 成为许多成功的深度架构的标准设置。已经提出了一系列研究来改进 ReLU。提出了两种 ReLU 变体，对于小于零的输人采用非零斜率$\alpha:$绝对值校正选择$\alpha=-1$ $\overline{\mathbb{E}}[14];$ LeakyReLU 将$\alpha$固定为一个小值$\mathbb{E}[36]$ 。PReLU[12] 和 [2] 通过将非零斜率作为可训练参数进一步。Maxout 通过将输人分成组并输出最大值[10],进一步推广了ReLU。由于 ReLU 是非平滑的，因此已经开发了一些平滑激活函数来解决此问题，例如软加$\boxed{\mathrm{E[8],ELUE[6],SELUE[17],Misc}}$E[26] ,Exp[24] 和Sigmoid [20] 。这些整流器都是可以在不同类型的神经网络中使用的通用功能。通过考虑拓扑信息，它们都没有针对 GNN 的特殊设计
1.3 Linear Graph Neural Network
在 [32, 35, 41-43] 中，作者通过删除激活函数并折叠连续层之间的权重矩阵，提出了简化的 GNN (SGC 和 S$^2$ GC)。这些工作假设GCN 层之间的非线性不是关键的，但大部分好处来自拓扑平滑。然而，消除层之间的非线性会牺牲 GNN 的深度效率。此外，从SGC 中删除的 ReLU 不是拓扑感知的。因此，整流器在 GNN 中起着不太重要的作用。如 SGC 所述，拓扑信息在聚合阶段对于使GNN 正常工作至关重要，这也促使我们在更新阶段合并拓扑信息。
1 | [
[
"0.1013",
"0.1067",
"0.91",
"0.108",
"0.91",
"0.1218",
"0.1013",
"0.1206"
],
[
"0.08826",
"0.1225",
"0.911",
"0.1225",
"0.911",
"0.1364",
"0.08826",
"0.1364"
],
[
"0.08496",
"0.1351",
"0.9126",
"0.1364",
"0.9126",
"0.1497",
"0.08496",
"0.1483"
],
[
"0.08826",
"0.1515",
"0.1871",
"0.1515",
"0.1871",
"0.1622",
"0.08826",
"0.1622"
],
[
"0.10297",
"0.1635",
"0.9126",
"0.1635",
"0.9126",
"0.1768",
"0.10297",
"0.1768"
],
[
"0.08826",
"0.1787",
"0.911",
"0.1787",
"0.911",
"0.1895",
"0.08826",
"0.1895"
],
[
"0.0866",
"0.1901",
"0.8013",
"0.1913",
"0.8013",
"0.2052",
"0.0866",
"0.2039"
],
[
"0.1054",
"0.207",
"0.452",
"0.207",
"0.452",
"0.2172",
"0.1054",
"0.2172"
],
[
"0.452",
"0.207",
"0.91",
"0.207",
"0.91",
"0.2172",
"0.452",
"0.2172"
],
[
"0.10046",
"0.2172",
"0.9126",
"0.2184",
"0.9126",
"0.2329",
"0.10046",
"0.2317"
],
[
"0.0866",
"0.2317",
"0.9126",
"0.2329",
"0.9126",
"0.2474",
"0.0866",
"0.2468"
],
[
"0.08826",
"0.2474",
"0.1316",
"0.2474",
"0.1316",
"0.2588",
"0.08826",
"0.2588"
],
[
"0.10297",
"0.26",
"0.911",
"0.26",
"0.911",
"0.2734",
"0.10297",
"0.2734"
],
[
"0.08826",
"0.2747",
"0.9126",
"0.2747",
"0.9126",
"0.2886",
"0.08826",
"0.2886"
],
[
"0.0866",
"0.2886",
"0.7915",
"0.2886",
"0.7915",
"0.3018",
"0.0866",
"0.3018"
],
[
"0.1013",
"0.3005",
"0.911",
"0.3018",
"0.911",
"0.315",
"0.1013",
"0.3145"
],
[
"0.0866",
"0.3164",
"0.911",
"0.3164",
"0.911",
"0.33",
"0.0866",
"0.33"
],
[
"0.08826",
"0.33",
"0.9126",
"0.33",
"0.9126",
"0.3435",
"0.08826",
"0.3435"
],
[
"0.08905",
"0.3452",
"0.911",
"0.3452",
"0.911",
"0.356",
"0.08905",
"0.356"
],
[
"0.08826",
"0.3567",
"0.834",
"0.3567",
"0.834",
"0.3706",
"0.08826",
"0.3706"
],
[
"0.10297",
"0.375",
"0.273",
"0.375",
"0.273",
"0.3882",
"0.10297",
"0.3882"
],
[
"0.0907",
"0.4072",
"0.911",
"0.4072",
"0.911",
"0.4204",
"0.0907",
"0.4204"
],
[
"0.1152",
"0.4204",
"0.5537",
"0.4204",
"0.5537",
"0.4343",
"0.1152",
"0.4343"
],
[
"0.0866",
"0.433",
"0.9126",
"0.4343",
"0.9126",
"0.449",
"0.0866",
"0.4475"
],
[
"0.1168",
"0.4495",
"0.3718",
"0.4495",
"0.3718",
"0.4602",
"0.1168",
"0.4602"
],
[
"0.08417",
"0.4622",
"0.911",
"0.4622",
"0.911",
"0.476",
"0.08417",
"0.476"
],
[
"0.1136",
"0.4766",
"0.4395",
"0.476",
"0.4395",
"0.4873",
"0.1136",
"0.488"
],
[
"0.08417",
"0.5176",
"0.2712",
"0.5176",
"0.2712",
"0.531",
"0.08417",
"0.531"
],
[
"0.08826",
"0.537",
"0.326",
"0.537",
"0.326",
"0.5503",
"0.08826",
"0.5503"
],
[
"0.0866",
"0.5537",
"0.9126",
"0.5547",
"0.9126",
"0.5684",
"0.0866",
"0.5674"
],
[
"0.08905",
"0.571",
"0.91",
"0.571",
"0.91",
"0.5806",
"0.08905",
"0.5806"
],
[
"0.08826",
"0.5835",
"0.9126",
"0.5835",
"0.9126",
"0.5967",
"0.08826",
"0.5967"
],
[
"0.08826",
"0.5967",
"0.911",
"0.5967",
"0.911",
"0.61",
"0.08826",
"0.61"
],
[
"0.08905",
"0.6123",
"0.91",
"0.6123",
"0.91",
"0.6226",
"0.08905",
"0.6226"
],
[
"0.0866",
"0.625",
"0.9126",
"0.625",
"0.9126",
"0.638",
"0.0866",
"0.638"
],
[
"0.08826",
"0.64",
"0.308",
"0.64",
"0.308",
"0.651",
"0.08826",
"0.651"
],
[
"0.08826",
"0.6797",
"0.3015",
"0.6797",
"0.3015",
"0.6934",
"0.08826",
"0.6934"
],
[
"0.08905",
"0.6997",
"0.91",
"0.6997",
"0.91",
"0.7104",
"0.08905",
"0.7104"
],
[
"0.08905",
"0.714",
"0.91",
"0.714",
"0.91",
"0.7246",
"0.08905",
"0.7246"
],
[
"0.08905",
"0.728",
"0.91",
"0.728",
"0.91",
"0.738",
"0.08905",
"0.738"
],
[
"0.0866",
"0.7393",
"0.9126",
"0.7393",
"0.9126",
"0.7524",
"0.0866",
"0.7524"
],
[
"0.0866",
"0.752",
"0.9126",
"0.7524",
"0.9126",
"0.7676",
"0.0866",
"0.7666"
],
[
"0.08826",
"0.7695",
"0.911",
"0.7695",
"0.911",
"0.78",
"0.08826",
"0.78"
],
[
"0.08826",
"0.782",
"0.1168",
"0.782",
"0.1168",
"0.7935",
"0.08826",
"0.7935"
],
[
"0.08826",
"0.8237",
"0.3743",
"0.8237",
"0.3743",
"0.837",
"0.08826",
"0.837"
],
[
"0.08826",
"0.8413",
"0.911",
"0.8413",
"0.911",
"0.855",
"0.08826",
"0.855"
],
[
"0.08905",
"0.857",
"0.91",
"0.857",
"0.91",
"0.867",
"0.08905",
"0.867"
],
[
"0.0866",
"0.8687",
"0.9126",
"0.869",
"0.9126",
"0.8833",
"0.0866",
"0.882"
],
[
"0.0866",
"0.882",
"0.509",
"0.884",
"0.509",
"0.898",
"0.0866",
"0.896"
]
] | [
"XMATHX_LIN7<NEXXT_LINE。这些任务中的数据通常在欧几里得空间中表示。但是，越来越多的应用程序从非欧几里得域生",
"成数据，并将其表示为对象之间具有复杂关系的图形E[5,9,21,2334,39,40]。通过定义图上的卷积算子，图神经网络（GNN）将卷",
"积神经网络（CNN）从图像域扩展到图域。因此，我们仍然可以将可能的工具从CNN调整为GNNs、i.e.、ReLU[27]、Dropout[31]",
"或剩余链接【13]",
"GNN现在能够通过采用所谓的消息传递（MP）框架来挖掘图的特征，该框架迭代地聚合和更新沿边路径的节点信息。将MP作",
"为建模工具，近年来在下游任务中取得了前所未有的成就，例如节点/图分类E[29.30]和链接预测E[3，4]。然而，最近的许多工",
"作证明，MP中设计不当的传播步骤违反了许多现实世界图所持有的同构假设，从而导致关键性能下降E[15,44]",
"上述麻烦促使我们改进MP并增强GNN的泛化能力。-",
"一个有前途的探索路线是动态考虑MP期间每个节点的贡献。注意机制是",
"个很好的指针，因为它们已经在15,33]中进行了探索。注意机制的好处之一是，它们通过将权重相应地分配给不同的节点，使",
"网络专注于要传播的最相关的信息。e.g:，作品E[15,33]表明，通过自适应转换传播信息会在多个归纳基准上产生令人印象深刻",
"的性能",
"但是，当前的自适应转换设计效率低下。我们的推理源于两个关键的观点。首先，在推断每个输人节点的自适应权重时没有考",
"虑拓扑信息，因此无法捕获图[15]的同态信息。其次，我们怀疑在提高模型泛化能力和准确性方面的局限性是由于为每个输人节",
"点共享跨通道的自适应权重。可能不希望在没有区别的情况下转换所有通道，因为某些通道不如其他通道重要",
"受上述问题的启发，我们开发了GReLU，这是一种拓扑感知的自适应激活功能，适用于任何GNN主干。GReLU优化了信息传",
"播并提高了节点分类任务的性能。如图?？所示，我们的机制由两个组件组成，改进了GNN的训练过程。一方面，GReLU通过将",
"特征和图形作为输人来继承拓扑信息。另一方面，与标准整流器是相同的函数W.r.t.所有输人不同，GReLU分配一组参数来动态",
"控制每个节点和通道的分段非线性函数。这种拓扑自适应属性提供了比拓扑未知对应物更好的表达能力。同时，GReLU是计算高",
"效的e.g.，具有GReLU的两层GCN和两层GATE[33]具有相似的计算复杂度，而前者在节点分类方面能够优于GAT",
"这项工作的主要贡献如下：",
"i.We incorporate the topology information in the updating function of GNNs by introducing a topology-aware activation function in",
"the updating stage of message passing, which, to our knowledge, is novel",
"ii. We propose a well-designed graph adaptive rectified linear unit for graph data, which can be used as a plug-and-play component in",
"varioustypesofspatialandspectralGNNs",
"ii.We evaluate the proposed methods on various downstream tasks. We empirically show that the performance of various types of GNNs",
"canbegenerallyimproved withtheuseof our GReLU",
"1RELATEDWORKS",
"1.1Graph Neural Networks",
"GNN可以分为空间或。谱域从图信号处理i.e.的角度探索GNNs设计，利用傅里叶变换和拉普拉斯矩阵[7]的本征分解定义卷积",
"步长。可以提取任何频率来设计具有理论保证的图形滤波器。但是，这些模型存在三个主要缺点：1给定图的本征分解引起大量",
"计算负担；2）空间非局部化；3）为给定图设计的过滤器不能应用于其他图。空间GNNs通过引人消息传递机制来概括GNNs的",
"体系结构，该机制是根据图的一些空间特征（例如邻接矩阵，节点特征，边缘特征，etc.E[11]）聚合邻居信息。由于其高效率和",
"可扩展性，空间GNN在最近的工作中变得更加流行。最具代表性的模型是图注意网络（GAT）E[33]，它通过在聚合步骤中使用",
"注意机制来加权不同邻居的节点特征。但是，对于每个节点功能，GAT为节点分配了跨所有通道共享的单个权重，这忽略了不同",
"通道可能表现出不同重要性的事实",
"1.2ActivationFunctions",
"激活函数已经在神经网络的整个历史上进行了长时间的研究。少数重要的里程碑之一是将整流线性单元（ReLU)E[27]引入神经网",
"络。由于其简单性和非线性，ReLU成为许多成功的深度架构的标准设置。已经提出了一系列研究来改进ReLU。提出了两种ReLU",
"变体，对于小于零的输人采用非零斜率α：绝对值校正选择α=-1E[14]：LeakyReLU将α固定为一个小值E[36]。PReLU[12]",
"和[2]通过将非零斜率作为可训练参数进一步。Maxout通过将输入分成组并输出最大值[10]，进一步推广了ReLU。由于ReLU是",
"非平滑的，因此已经开发了一些平滑激活函数来解决此问题，例如软加E[8]，ELUE[6]，SELUE[17]，MiscE[26]，Exp[24]和",
"Sigmoid[20】。这些整流器都是可以在不同类型的神经网络中使用的通用功能。通过考虑拓扑信息，它们都没有针对GNN的特殊",
"设计",
"1.3LinearGraphNeuralNetwork",
"在[32,35,41-43]中，作者通过删除激活函数并折叠连续层之间的权重矩阵，提出了简化的GNN（SGC和S²GC)。这些工作假设",
"GCN层之间的非线性不是关键的，但大部分好处来自拓扑平滑。然而，消除层之间的非线性会牺牲GNN的深度效率。此外，从",
"SGC中删除的ReLU不是拓扑感知的。因此，整流器在GNN中起着不太重要的作用。如SGC所述，拓扑信息在聚合阶段对于使",
"GNN正常工作至关重要，这也促使我们在更新阶段合并拓扑信息。"
] | {
"arxiv_id": "2202.06281",
"language": "zh",
"page_idx": 0
} |
||
2202.06281-zh-1 | 2202.06281-zh-1.jpg | WWW'22,April 25-29, 2022, Virtual Event, Lyon, France.
Yifei Zhang and Hao Zhu and Ziqiao Meng and Piotr Koniusz and Irwin King
<FigureHere>
Figure 1: The architecture of GReLU. The arrows in the box indicate the flow for computing the parameter for hyperfunction. The arrow outside the box indicates the calculation of GReLU. The channel-wise block produces the $K\times C$ parameters for $C$ channels and K segments. The node-wise block generates $N$ parameters for $N$ nodes. The total $K\times C\times N$ parameters are obtained via computing the outer product of channel-wise and node-wise parameters.
2 РRELIMINARIES
# 2.1 Notations
在本文中，具有节点特征的图表示为$G= ( X, A)$ , 其 中 $\mathcal{V}$是顶点集，$\mathcal{E}$ 是边集，$X\in\mathbb{R}^{N\times C}$ 是矩阵的 XMATHX。$A\in\{0,1\}^{N\times N}$
表示$G$,ie.的邻接矩阵，如果$v_i$和$v_j$之间存在边，则$A,a_ij$中的第$(i,j)$项为 1。我们将$\mathcal{N}(i)=\{j|a_ij=1\}$表示为节点$i$的邻域索引集。$v_i$的度，表示为$d_i$ ,是与$v_i$人射的边数。对于$d$维向量$x\in\mathbb{R}^d$ , 我们使用$x_i$表示$x$的第$i$个条目。对于$X$的第 (i,j) 个条目，我们使用$x_i$表示$X$和$x_ij$的行向量
# 2.2 Message Passing (MP)
空间 GNN 的成功源于在图中的节点之间应用消息传递 (也称为邻居聚合)。在 GNN 的前向传递中，通过边缘$e_{ij}$从$i$的邻居聚合
节点$j$的隐藏状态，迭代更新节点$i$的隐藏状态。假设存在$l$迭代，则消息传递范例可以形式化为
(1)
$$\begin{aligned}&m_{i}^{(l)}=\mathrm{AGGREGATE}^{(l)}\big(\{\boldsymbol{h}_{i}^{(l)},\boldsymbol{h}_{j}^{(l)},\boldsymbol{e}_{ij}|j\in\mathcal{N}(i)\}\big),\\&h_{i}^{(l+1)}=\sigma\big(\Theta^{(l)}[h_{i}^{(l)},m_{i}^{(l)}]\big),\end{aligned}$$
(2)
其中$h_i$和$h_j$是节点$i$和$j$的隐藏状态，$m_i$表示节点$i$从其邻居$j\in N(i)$接收的消息。(2)中的更新函数被建模为线性变换，然后
是非线性激活函数$\sigma(\cdot)$,其中$\Theta\in\mathbb{R}^C\times(2C)$矩阵包含线性变换的学习参数和 [·] 连接向量。
# 2.3 Parametric ReLU
香草 ReLU 是分段线性函数$y=\max(x,0)$ ,其中 max(·) 是在输人的每个通道之间应用的元素函数$x$和 0。扩展 ReLU 的一种更通
用的方法是让每个通道享受不同的元素函数：
$$y_c=\sigma_c(x_c)=\max_{1<k\leq K}\{\alpha^kx_c+\beta^k\},$$
(3)
其中$x_{\mathrm{c}}$和$y_{\mathrm{c}}$分别是$x$和？? 的第$c$个通道，$k$是 MAXX。请注意，当$\alpha^1=0,\beta^1=0,\alpha^2=1,\beta^2=0$ ,然后 (3)减少到 ReLU,而maxout[10] 是一个特例，如果$K=2$。此外，方法[2]不是直接学习参数$(\alpha^k,\beta^k)$,而是提出了超函数$\omega(x)$来估计不同通道的特定于样本的$( \alpha ^k, \beta ^k)$ 。
# 3 метнод
我们将建议的 GNN 自适应整流线性单元称为 GReLU。其余部分的组织如下。首先，我们通过定义两个函数来介绍 GReLU 的参数形式：生成参数的超函数和用于计算输出的激活函数。然后，我们详细介绍了GReLU 的超功能体系结构，并从邻域加权的角度将GReLU 与 GAT 联系起来。最后，我们分析了 GReLU 的时间复杂度，并将其与以前的工作进行了比较
3.1 Graph-adaptive Rectifed Linear Unit (GReLU)
给定节点$n$的输入特征，其中$x_n\in\mathbb{R}^C$, GReLU 被定义为元素级函数：
(4)
$$\sigma_{\omega(X,A)}(x_{nc})=\max_{1\leq k\leq K}\{\alpha_{nc}^{k}x_{nc}+\beta_{nc}^{k}\},$$
,它由两部分组成： | [
[
"0.08496",
"0.07385",
"0.3555",
"0.07574",
"0.3555",
"0.089",
"0.08496",
"0.0871"
],
[
"0.546",
"0.0745",
"0.914",
"0.07574",
"0.914",
"0.089",
"0.546",
"0.0884"
],
[
"0.4763",
"0.11743",
"0.4958",
"0.11743",
"0.4958",
"0.1256",
"0.4763",
"0.1256"
],
[
"0.517",
"0.11743",
"0.5884",
"0.11743",
"0.5884",
"0.1276",
"0.517",
"0.1276"
],
[
"0.6055",
"0.1256",
"0.6323",
"0.1256",
"0.6323",
"0.1338",
"0.6055",
"0.1338"
],
[
"0.326",
"0.1483",
"0.3367",
"0.1483",
"0.3367",
"0.1515",
"0.326",
"0.1515"
],
[
"0.3406",
"0.1483",
"0.3586",
"0.1483",
"0.3586",
"0.1528",
"0.3406",
"0.1528"
],
[
"0.621",
"0.1559",
"0.6787",
"0.1559",
"0.6787",
"0.1666",
"0.621",
"0.1666"
],
[
"0.327",
"0.1642",
"0.361",
"0.1642",
"0.361",
"0.1736",
"0.327",
"0.1736"
],
[
"0.4272",
"0.1666",
"0.482",
"0.1692",
"0.4805",
"0.1863",
"0.4265",
"0.1831"
],
[
"0.517",
"0.1666",
"0.5664",
"0.1666",
"0.5664",
"0.1849",
"0.517",
"0.1849"
],
[
"0.638",
"0.1692",
"0.6733",
"0.1692",
"0.6733",
"0.1849",
"0.638",
"0.1849"
],
[
"0.3342",
"0.1812",
"0.3586",
"0.1812",
"0.3586",
"0.1895",
"0.3342",
"0.1895"
],
[
"0.326",
"0.1945",
"0.3586",
"0.1963",
"0.357",
"0.207",
"0.3245",
"0.2052"
],
[
"0.634",
"0.1895",
"0.675",
"0.1895",
"0.675",
"0.1963",
"0.634",
"0.1963"
],
[
"0.4575",
"0.1976",
"0.4844",
"0.1976",
"0.4844",
"0.2058",
"0.4575",
"0.2058"
],
[
"0.46",
"0.2247",
"0.4958",
"0.2247",
"0.4958",
"0.2317",
"0.46",
"0.2317"
],
[
"0.4265",
"0.2329",
"0.4436",
"0.2329",
"0.4436",
"0.2405",
"0.4265",
"0.2405"
],
[
"0.5254",
"0.2343",
"0.556",
"0.2343",
"0.556",
"0.2412",
"0.5254",
"0.2412"
],
[
"0.0866",
"0.2715",
"0.9087",
"0.2715",
"0.9087",
"0.2854",
"0.0866",
"0.2854"
],
[
"0.0866",
"0.2854",
"0.9126",
"0.2854",
"0.9126",
"0.2986",
"0.0866",
"0.2986"
],
[
"0.0866",
"0.2998",
"0.9126",
"0.2998",
"0.9126",
"0.3132",
"0.0866",
"0.3132"
],
[
"0.08496",
"0.3145",
"0.591",
"0.3145",
"0.591",
"0.3276",
"0.08496",
"0.3276"
],
[
"0.08417",
"0.3528",
"0.259",
"0.3528",
"0.259",
"0.3662",
"0.08417",
"0.3662"
],
[
"0.08417",
"0.3718",
"0.2115",
"0.3718",
"0.2115",
"0.385",
"0.08417",
"0.385"
],
[
"0.08417",
"0.3894",
"0.9126",
"0.3865",
"0.9126",
"0.404",
"0.08417",
"0.4072"
],
[
"0.0866",
"0.4048",
"0.9126",
"0.4048",
"0.9126",
"0.4185",
"0.0866",
"0.4185"
],
[
"0.08496",
"0.4185",
"0.9126",
"0.4192",
"0.9126",
"0.4343",
"0.08496",
"0.433"
],
[
"0.0866",
"0.4343",
"0.3718",
"0.4343",
"0.3718",
"0.4475",
"0.0866",
"0.4475"
],
[
"0.08496",
"0.461",
"0.308",
"0.461",
"0.308",
"0.4749",
"0.08496",
"0.4749"
],
[
"0.08496",
"0.4766",
"0.9126",
"0.4792",
"0.9126",
"0.4937",
"0.08496",
"0.4912"
],
[
"0.08826",
"0.4944",
"0.7173",
"0.4944",
"0.7173",
"0.505",
"0.08826",
"0.505"
],
[
"0.339",
"0.5103",
"0.572",
"0.5103",
"0.572",
"0.532",
"0.339",
"0.532"
],
[
"0.5376",
"0.5166",
"0.6543",
"0.515",
"0.6543",
"0.53",
"0.5376",
"0.531"
],
[
"0.8936",
"0.515",
"0.914",
"0.515",
"0.914",
"0.531",
"0.8936",
"0.531"
],
[
"0.3406",
"0.532",
"0.5156",
"0.536",
"0.5156",
"0.5547",
"0.3406",
"0.5503"
],
[
"0.8936",
"0.539",
"0.915",
"0.539",
"0.915",
"0.5537",
"0.8936",
"0.5537"
],
[
"0.0866",
"0.561",
"0.914",
"0.561",
"0.914",
"0.5747",
"0.0866",
"0.5747"
],
[
"0.08496",
"0.5728",
"0.661",
"0.5728",
"0.661",
"0.5903",
"0.08496",
"0.5903"
],
[
"0.0866",
"0.605",
"0.11926",
"0.605",
"0.11926",
"0.616",
"0.0866",
"0.616"
],
[
"0.1177",
"0.608",
"0.12335",
"0.608",
"0.12335",
"0.613",
"0.1177",
"0.613"
],
[
"0.12335",
"0.608",
"0.1299",
"0.608",
"0.1299",
"0.6123",
"0.12335",
"0.6123"
],
[
"0.1283",
"0.604",
"0.273",
"0.604",
"0.273",
"0.617",
"0.1283",
"0.617"
],
[
"0.08826",
"0.622",
"0.9126",
"0.622",
"0.9126",
"0.6353",
"0.08826",
"0.6353"
],
[
"0.08826",
"0.6353",
"0.3718",
"0.6353",
"0.3718",
"0.6484",
"0.08826",
"0.6484"
],
[
"0.4902",
"0.656",
"0.6006",
"0.656",
"0.6006",
"0.6704",
"0.4902",
"0.6704"
],
[
"0.8945",
"0.657",
"0.914",
"0.657",
"0.914",
"0.672",
"0.8945",
"0.672"
],
[
"0.0866",
"0.685",
"0.9126",
"0.685",
"0.9126",
"0.6987",
"0.0866",
"0.6987"
],
[
"0.3945",
"0.6577",
"0.4927",
"0.656",
"0.4927",
"0.6704",
"0.3945",
"0.6724"
],
[
"0.08826",
"0.6997",
"0.9126",
"0.6997",
"0.9126",
"0.7134",
"0.08826",
"0.7134"
],
[
"0.0866",
"0.7134",
"0.2239",
"0.7134",
"0.2239",
"0.7266",
"0.0866",
"0.7266"
],
[
"0.08496",
"0.741",
"0.1978",
"0.741",
"0.1978",
"0.755",
"0.08496",
"0.755"
],
[
"0.0866",
"0.7603",
"0.9126",
"0.7603",
"0.9126",
"0.7734",
"0.0866",
"0.7734"
],
[
"0.08826",
"0.7734",
"0.9126",
"0.7734",
"0.9126",
"0.787",
"0.08826",
"0.787"
],
[
"0.0866",
"0.7886",
"0.734",
"0.7886",
"0.734",
"0.802",
"0.0866",
"0.802"
],
[
"0.0866",
"0.8164",
"0.1323",
"0.8164",
"0.1323",
"0.8276",
"0.0866",
"0.8276"
],
[
"0.1259",
"0.8145",
"0.5063",
"0.813",
"0.5063",
"0.8276",
"0.1259",
"0.829"
],
[
"0.0866",
"0.8335",
"0.5337",
"0.8335",
"0.5337",
"0.8477",
"0.0866",
"0.8477"
],
[
"0.38",
"0.855",
"0.615",
"0.853",
"0.615",
"0.8706",
"0.38",
"0.8726"
],
[
"0.8945",
"0.855",
"0.914",
"0.855",
"0.914",
"0.869",
"0.8945",
"0.869"
],
[
"0.4805",
"0.8657",
"0.5215",
"0.8657",
"0.5215",
"0.8755",
"0.4805",
"0.8755"
],
[
"0.08496",
"0.8833",
"0.2115",
"0.882",
"0.2115",
"0.8965",
"0.08496",
"0.898"
]
] | [
"www22pril 2529,222,VirtualEvent, Lyon,Franc",
"Yifei Zhang and Hao Zhu and Ziqiao Meng and Piotr Koniusz and Irwin King",
"X.A",
"Hyperfunctior",
"(X,A",
"O",
"usion",
"Diffusion",
"Pooing",
"Chamiotwise",
"Nodiowit",
"MLP",
"MLP",
"Tanh",
"Softmax",
"3",
"nePn",
"En.",
"XA.",
"Figure 1: The architecture of GReLU. The arrows in the box indicate the flow for computing the parameter for hyperfunction",
"The arrow outside the box indicates the calculation of GReLU. The channel-wise block produces the K × C parameters for C",
"channels and K segments. The node-wise block generates N parameters for N nodes.The total KxCxN parameters are obtained",
"via computing the outer product of channel-wise and node-wise parameters",
"2PRELIMINARIES",
"2.1 Notations",
"在本文中，具有节点特征的图表示为G=（X,A)，其中是顶点集，是边集，XERN×C是矩阵的XMATHX。AE(0,1)NxN",
"表示G，i.e.的邻接矩阵，如果ui和uj之问存在边，则A，aij中的第（i,j)项为1。我们将N(i)=（ilaij=1)表示为节点i的邻",
"域索引集。ui的度，表示为di，是与ui人射的边数。对于d维向量xERd，我们使用xi表示x的第i个条目。对于x的第（i,j)",
"个条目，我们使用xi表示X和xij的行向量",
"2.2MessagePassing (MP)",
"空间GNN的成功源于在图中的节点之间应用消息传递（也称为邻居聚合）。在GNN的前向传递中，通过边缘eij从i的邻居聚合",
"节点了的隐藏状态，送代更新节点i的隐藏状态。假设存在「迭代，则消息传递范例可以形式化为：",
"m)) = AGGREGATE()(h,h),",
"h,eijlje N(i))",
"(1)",
"h(+1) = 0(e()[n(,m(\"1),",
"(2) ",
"其中hi和hi是节点i和j的隐藏状态，mi表示节点i从其邻居jeN(i)接收的消息。(2)中的更新函数被建模为线性变换，然后",
"是非线性激活函数(），其中θERCx(2C）矩阵包含线性变换的学习参数和[I连接向量。",
"2.3",
"",
"",
"Parametric ReLU",
"香草ReLU是分段线性函数y=max(x,0），其中max(）是在输人的每个通道之间应用的元素函数x和0。扩展ReLU的一种更通",
"用的方法是让每个通道享受不同的元素函数：",
"max (αk xc + βk),",
"(3)",
"其中x和yc分别是x和???的第c个通道，k是MAXX。请注意，当αl=0,βl=0,α²=1,β2=0，然后(3)减少到ReLU，而",
"Ye = 0c(xc) =",
"maxout[10]是一个特例，如果K=2。此外，方法[2]不是直接学习参数（αk,βk），而是提出了超函数w(x)来估计不同通道的特",
"定于样本的（αk.βk）。",
"3METHOD",
"我们将建议的GNN自适应整流线性单元称为GReLU。其余部分的组织如下。首先，我们通过定义两个函数来介绍GReLU的参数",
"形式：生成参数的超函数和用于计算输出的激活函数。然后，我们详细介绍了GReLU的超功能体系结构，并从邻域加权的角度将",
"GReLU与GAT联系起来。最后，我们分析了GReLU的时间复杂度，并将其与以前的工作进行了比较",
"3.1",
"Graph-adaptiveRectifiedLinearUnit (GReLU)",
"给定节点n的输人特征，其中xnERC，GReLU被定义为元素级函数",
"Co(X,A)(xne) =, max (ahexne + phe),",
"(4)",
"1≤k≤K",
"，它由两部分组成："
] | {
"arxiv_id": "2202.06281",
"language": "zh",
"page_idx": 1
} |
||
2202.06281-zh-2 | 2202.06281-zh-2.jpg | Graph-adaptive Rectified Linear Unit for Graph Neural Networks
WWW'22, April 25-29, 2022, Virtual Event, Lyon, France.
Table 1: Different Variants of ReLU.
<table>
<tbody>
<tr>
<th>Type</th>
<th>Formula</th>
<th>$\operatorname{Learnable}$</th>
<th>Input</th>
<th>Plugin</th>
</tr>
<tr>
<td>$\mathbf{ReLU}$</td>
<td>$\max(x,0)$</td>
<td>No</td>
<td>$X$</td>
<td>Yes</td>
</tr>
<tr>
<td>$\mathbf{LReLU}$</td>
<td>$\max(x,\alpha x)$</td>
<td>No</td>
<td>$X$</td>
<td>Yes</td>
</tr>
<tr>
<td>PReLU</td>
<td>$\max(x.\alpha_nx$</td>
<td>Yes</td>
<td>$X$</td>
<td>Yes</td>
</tr>
<tr>
<td>Maxout</td>
<td>$\max(w_1x+b_1,w_2x+b_2)$</td>
<td>Yes</td>
<td>$X$</td>
<td>No</td>
</tr>
<tr>
<td>GReLU</td>
<td>$\tilde{\varepsilon}_{c}x_{nc}+\beta_{nc}^{\kappa}$ $\max_{1\leq k<K}$ $,r_{r}\{\eta^{n}$</td>
<td>Yes</td>
<td>$X,A$</td>
<td>Yes</td>
</tr>
</tbody>
</table>
$\bullet$ Hyperfunction $\omega(X,A)$ whose inputs are the adjacency matrix $A$ and the node feature matrix $X$ , and whose output parameters
$\{(\alpha_{nc}^{k},\beta_{nc}^{k})\}_{k=1}^{K}$ are used by the parametric activation function.
$\bullet$ The parametric activation function $\sigma(\cdot)$ generates activation with parameters $\alpha_nc^k$
超函数将邻接矩阵$A$的线性激活信息编码为确定矩阵。通过将 GReLU 插人 MP,更新功能变得自适应，比其非自适应功能获得更
多的表示能力
3.2 Design of Hyperfunction of GReLU
请注意，要完全获得 GReLU, 超函数需要为$C$通道和 N 节点生成$K$参数集$\{(\alpha_{nc}^k,\beta_{nc}^k)\}_{k=1}^K$。GReLU 中所需的参数总数为 2$K\times C\times N$ 。我们可以简单地将超函数$\omega(X,A)$建模为具有输出维度 2$K\times C\times N$的单层图卷积神经网络 (GCN)。但是，这种实现会导致超函数中的参数过多，从而带来过拟合的风险，因为我们在实践中观察到性能下降。为了解决这个问题，我们通过将节点建模为两个不同的块来分解来自通道的节点参数。一个块学习通道式参数$\alpha_c^k,\beta_c^k$,另一个块学习节点式参数$\gamma_n$ 。最终输出 ($\alpha_{nc}^k,\beta_{nc}^k$)计算如下：
$$\alpha_{nc}^{k}=\gamma_{n}\alpha_{c}^{k}\:\mathrm{and}\:\beta_{nc}^{k}=\gamma_{n}\beta_{c}^{k}\:.$$
(5)
此分解限制了从 2K$\times C\times N$到 2$K\times(C+N)$的超函数需要学习的参数。结果，提高了模型的通用性和计算效率
通道参数。为了对图$G$的全局上下文进行编码，同时尽可能减少可学习参数的数量，我们采用图扩散滤波器$S^{PPR}$从$A$和$X$中挤
压和提取拓扑信息。我们通过完全个性化的 PageRank 方案传播输入矩阵$X:$
$$E=S^{\mathrm{PPR}}X=\alpha\left(I_n-(1-\alpha)D^{-1/2}AD^{-1/2}\right)^{-1}X,$$
(6)
具有传送 (或重新启动) 概率$\alpha\in(0,1]$ ,其中$D$是度矩阵$,D^{-1/2}AD^{-1/2}$表示归一化邻接矩阵$\boldsymbol{A}$ 。我们对扩散结果进行平均，并使用线性变换和 tanh 函数将其转换为通道参数$(\alpha_c^k,\beta_c^k)$ 。我们使用 tanh(·)将$(\alpha_c^k,\beta_c^k)$归一化为 [-1,1] 范围内，目的是控制参数比例。我们将$h\in\mathbb{R}^C$定义为扩散结果的平均值，将$P\in\mathbb{R}^2\times K\times C$定义为通道参数。信道块的计算方式为：
$$P=\tanh(\mathrm{MLP}(\bar{\boldsymbol{e}}))\quad\mathrm{where}\quad\bar{\boldsymbol{e}}=\frac{1}{N}\sum_{n=1}^{N}\boldsymbol{e}_{n}.$$
(7)
节点参数。节点式模块计算每个节点的参数。我们还使用扩散矩阵来捕获图信息。为了获得每个节点$n$的参数$\gamma_{n}$ ,而不是平均
扩散结果，我们使用 MLP 直接将每个$e$压缩到一个维度中。为了反映每个节点的重要性，应用 softmax 将$\gamma_{n}$规范化为：
$$\gamma_n=\frac{\exp(\mathrm{MLP}(e_n))}{\sum_{n'=1}^N\exp(\mathrm{MLP}(e_{n'}))}.$$
(8)
将步骤插人更新函数。通过结合方程(8)和(7),我们通过(5)获得了GReLU 的全部参数。我们将 GReLU 插入第$\iota$层更新函数
$$h_n^{(l+1)}=\sigma_n^{(l)}(\Theta^{(l)}[h_n^{(l)},m_n^{(l)}]),$$
(9)
,其中$\sigma_n^l$(·)是第$\iota$层中的 GReLU。如图 ?? 所示，在更新函数中，每层$\iota$有一组参数$\Theta^{(l)}$,而 GReLU 是通道和鼻子自适应的 | [
[
"0.08826",
"0.077",
"0.3938",
"0.077",
"0.3938",
"0.0871",
"0.08826",
"0.0871"
],
[
"0.6396",
"0.07385",
"0.911",
"0.0745",
"0.911",
"0.0884",
"0.6396",
"0.0871"
],
[
"0.38",
"0.1061",
"0.614",
"0.1061",
"0.614",
"0.11615",
"0.38",
"0.11615"
],
[
"0.10706",
"0.1427",
"0.1732",
"0.1445",
"0.1724",
"0.1686",
"0.1054",
"0.1666"
],
[
"0.2296",
"0.1414",
"0.3308",
"0.1414",
"0.3308",
"0.1635",
"0.2296",
"0.1635"
],
[
"0.714",
"0.1433",
"0.7837",
"0.1433",
"0.7837",
"0.1666",
"0.714",
"0.1666"
],
[
"0.808",
"0.1427",
"0.8906",
"0.1445",
"0.8896",
"0.1666",
"0.806",
"0.1642"
],
[
"0.7344",
"0.178",
"0.7656",
"0.1818",
"0.7607",
"0.2058",
"0.7295",
"0.202"
],
[
"0.568",
"0.1445",
"0.6855",
"0.1445",
"0.6855",
"0.1635",
"0.568",
"0.1635"
],
[
"0.5664",
"0.1812",
"0.607",
"0.1843",
"0.6045",
"0.2039",
"0.564",
"0.2008"
],
[
"0.8286",
"0.1799",
"0.875",
"0.1849",
"0.871",
"0.2058",
"0.824",
"0.2008"
],
[
"0.232",
"0.1843",
"0.343",
"0.1843",
"0.343",
"0.2026",
"0.232",
"0.2026"
],
[
"0.10864",
"0.1831",
"0.1772",
"0.1831",
"0.1772",
"0.202",
"0.10864",
"0.202"
],
[
"0.5645",
"0.207",
"0.607",
"0.2102",
"0.6045",
"0.2311",
"0.562",
"0.2279"
],
[
"0.734",
"0.2084",
"0.7646",
"0.2084",
"0.7646",
"0.2311",
"0.734",
"0.2311"
],
[
"0.8267",
"0.2084",
"0.8716",
"0.2102",
"0.871",
"0.2311",
"0.826",
"0.2291"
],
[
"0.232",
"0.2108",
"0.3628",
"0.2108",
"0.3628",
"0.2299",
"0.232",
"0.2299"
],
[
"0.10864",
"0.2102",
"0.1912",
"0.2102",
"0.1912",
"0.2291",
"0.10864",
"0.2291"
],
[
"0.734",
"0.2349",
"0.7646",
"0.2349",
"0.7646",
"0.2588",
"0.734",
"0.2588"
],
[
"0.8267",
"0.2343",
"0.875",
"0.2393",
"0.871",
"0.26",
"0.8228",
"0.2551"
],
[
"0.5664",
"0.2361",
"0.607",
"0.2361",
"0.607",
"0.2568",
"0.5664",
"0.2568"
],
[
"0.10864",
"0.2374",
"0.1626",
"0.2374",
"0.1626",
"0.2568",
"0.10864",
"0.2568"
],
[
"0.2303",
"0.2361",
"0.4412",
"0.2361",
"0.4412",
"0.2583",
"0.2303",
"0.2583"
],
[
"0.5664",
"0.2632",
"0.6113",
"0.2651",
"0.6094",
"0.2854",
"0.5645",
"0.2827"
],
[
"0.7344",
"0.2644",
"0.763",
"0.2644",
"0.763",
"0.2842",
"0.7344",
"0.2842"
],
[
"0.8267",
"0.2632",
"0.8735",
"0.2651",
"0.8716",
"0.2854",
"0.826",
"0.2827"
],
[
"0.2296",
"0.2651",
"0.3782",
"0.2651",
"0.3782",
"0.2874",
"0.2296",
"0.2874"
],
[
"0.1095",
"0.2651",
"0.1929",
"0.2651",
"0.1929",
"0.2827",
"0.1095",
"0.2827"
],
[
"0.5664",
"0.289",
"0.6113",
"0.2917",
"0.6094",
"0.3125",
"0.5645",
"0.31"
],
[
"0.734",
"0.289",
"0.7646",
"0.289",
"0.7646",
"0.3125",
"0.734",
"0.3125"
],
[
"0.8286",
"0.2917",
"0.871",
"0.2917",
"0.871",
"0.3125",
"0.8286",
"0.3125"
],
[
"0.2303",
"0.2922",
"0.5186",
"0.2922",
"0.5186",
"0.3132",
"0.2303",
"0.3132"
],
[
"0.1095",
"0.2922",
"0.2035",
"0.2922",
"0.2035",
"0.3113",
"0.1095",
"0.3113"
],
[
"0.232",
"0.33",
"0.5293",
"0.33",
"0.5293",
"0.3547",
"0.232",
"0.3547"
],
[
"0.7183",
"0.3289",
"0.778",
"0.3289",
"0.778",
"0.3528",
"0.7183",
"0.3528"
],
[
"0.8267",
"0.332",
"0.8735",
"0.332",
"0.8735",
"0.3516",
"0.8267",
"0.3516"
],
[
"0.568",
"0.3333",
"0.607",
"0.3333",
"0.607",
"0.3499",
"0.568",
"0.3499"
],
[
"0.10864",
"0.332",
"0.1969",
"0.332",
"0.1969",
"0.3499",
"0.10864",
"0.3499"
],
[
"0.1136",
"0.3865",
"0.7935",
"0.3865",
"0.7935",
"0.3997",
"0.1136",
"0.3997"
],
[
"0.7974",
"0.3882",
"0.911",
"0.3882",
"0.911",
"0.3984",
"0.7974",
"0.3984"
],
[
"0.125",
"0.401",
"0.5073",
"0.401",
"0.5073",
"0.4153",
"0.125",
"0.4153"
],
[
"0.11273",
"0.4172",
"0.6113",
"0.4172",
"0.6113",
"0.4312",
"0.11273",
"0.4312"
],
[
"0.0866",
"0.4343",
"0.9126",
"0.4343",
"0.9126",
"0.4475",
"0.0866",
"0.4475"
],
[
"0.08826",
"0.4495",
"0.1748",
"0.4495",
"0.1748",
"0.461",
"0.08826",
"0.461"
],
[
"0.08496",
"0.4749",
"0.4192",
"0.4749",
"0.4192",
"0.488",
"0.08496",
"0.488"
],
[
"0.0866",
"0.4937",
"0.9126",
"0.4937",
"0.9126",
"0.507",
"0.0866",
"0.507"
],
[
"0.0907",
"0.509",
"0.9126",
"0.509",
"0.9126",
"0.523",
"0.0907",
"0.523"
],
[
"0.0866",
"0.5234",
"0.91",
"0.5234",
"0.91",
"0.537",
"0.0866",
"0.537"
],
[
"0.08826",
"0.5386",
"0.9126",
"0.5386",
"0.9126",
"0.552",
"0.08826",
"0.552"
],
[
"0.08905",
"0.5537",
"0.1095",
"0.5537",
"0.1095",
"0.5645",
"0.08905",
"0.5645"
],
[
"0.407",
"0.5674",
"0.589",
"0.5684",
"0.589",
"0.586",
"0.407",
"0.5854"
],
[
"0.8936",
"0.571",
"0.915",
"0.571",
"0.915",
"0.5854",
"0.8936",
"0.5854"
],
[
"0.0866",
"0.5913",
"0.7974",
"0.5913",
"0.7974",
"0.605",
"0.0866",
"0.605"
],
[
"0.0866",
"0.608",
"0.9126",
"0.608",
"0.9126",
"0.6226",
"0.0866",
"0.6226"
],
[
"0.0866",
"0.6235",
"0.5786",
"0.6235",
"0.5786",
"0.637",
"0.0866",
"0.637"
],
[
"0.3464",
"0.6445",
"0.6274",
"0.643",
"0.6274",
"0.662",
"0.3464",
"0.6636"
],
[
"0.621",
"0.6465",
"0.6465",
"0.6494",
"0.6436",
"0.66",
"0.6177",
"0.657"
],
[
"0.8936",
"0.648",
"0.915",
"0.648",
"0.915",
"0.662",
"0.8936",
"0.662"
],
[
"0.08826",
"0.672",
"0.9126",
"0.672",
"0.9126",
"0.686",
"0.08826",
"0.686"
],
[
"0.08826",
"0.6875",
"0.914",
"0.6875",
"0.914",
"0.7007",
"0.08826",
"0.7007"
],
[
"0.08417",
"0.6963",
"0.7573",
"0.6963",
"0.7573",
"0.7173",
"0.08417",
"0.7173"
],
[
"0.367",
"0.73",
"0.486",
"0.731",
"0.486",
"0.7456",
"0.367",
"0.744"
],
[
"0.5728",
"0.726",
"0.6055",
"0.7163",
"0.6113",
"0.729",
"0.5786",
"0.738"
],
[
"0.4795",
"0.733",
"0.568",
"0.733",
"0.568",
"0.744",
"0.4795",
"0.744"
],
[
"0.5645",
"0.735",
"0.625",
"0.726",
"0.63",
"0.7456",
"0.5703",
"0.755"
],
[
"0.8945",
"0.732",
"0.914",
"0.732",
"0.914",
"0.747",
"0.8945",
"0.747"
],
[
"0.591",
"0.7476",
"0.6094",
"0.7476",
"0.6094",
"0.756",
"0.591",
"0.756"
],
[
"0.09235",
"0.762",
"0.9126",
"0.762",
"0.9126",
"0.776",
"0.09235",
"0.776"
],
[
"0.0866",
"0.78",
"0.8506",
"0.78",
"0.8506",
"0.7935",
"0.0866",
"0.7935"
],
[
"0.4656",
"0.797",
"0.556",
"0.797",
"0.556",
"0.811",
"0.4656",
"0.811"
],
[
"0.4167",
"0.81",
"0.4272",
"0.8066",
"0.4314",
"0.8164",
"0.4207",
"0.8193"
],
[
"0.8936",
"0.8047",
"0.914",
"0.8047",
"0.914",
"0.8193",
"0.8936",
"0.8193"
],
[
"0.4495",
"0.8184",
"0.4558",
"0.8184",
"0.4558",
"0.8257",
"0.4495",
"0.8257"
],
[
"0.451",
"0.8145",
"0.5767",
"0.813",
"0.5767",
"0.8276",
"0.451",
"0.829"
],
[
"0.0866",
"0.8384",
"0.888",
"0.8384",
"0.888",
"0.8516",
"0.0866",
"0.8516"
],
[
"0.4019",
"0.856",
"0.5933",
"0.857",
"0.5933",
"0.875",
"0.4019",
"0.8735"
],
[
"0.8945",
"0.86",
"0.914",
"0.86",
"0.914",
"0.8735",
"0.8945",
"0.8735"
],
[
"0.08496",
"0.88",
"0.886",
"0.88",
"0.886",
"0.898",
"0.08496",
"0.898"
]
] | [
"Graph-adaptiveRectified LinearUnit forGraphNeural Networks",
"Ww'22, April 2529, 2022, Virtual Event, Lyon, France.",
"Tablel:DifferentVariantsofReLU",
"Type",
"Formula",
"Input",
"Plugin",
"X",
"Learnable",
"No",
"Yes",
"max(x, 0)",
"ReLU",
"No",
"X",
"Yes",
"max(x, ax)",
"LReLU",
"X",
"Yes",
"No",
"ELU",
"max(x, α(ex - 1))",
"Yes",
"X",
"Yes",
"max(x, αpx)",
"PReLU",
"Yes",
"X",
"No",
"max(w1x + b1, w2x + b2)",
"Maxout",
"maX1<k≤k(αhcXnc + βhc.",
"X,A",
"Yes",
"Yes",
"GReLU",
": Hyperfunction @(X, A) whose inputs are the adjacency matrix A and the node feature matrix X , and whose",
"output parameters",
"((ancβhc))k-, are used by the parametric activation function",
". The parametric activation function () generates activation with parameters ak",
"超函数将邻接矩阵A的线性激活信息编码为确定矩阵。通过将GReLU插人MP，更新功能变得自适应，比其非自适应功能获得更",
"多的表示能力",
"3.2DesignofHyperfunctionofGReLU",
"请注意，要完全获得GReLU，超函数需要为C通道和N节点生成K参数集（（αkcβkc）)K，。GReLU中所需的参数总数为2K×C×N",
"。我们可以简单地将超函数(X,A）建模为具有输出维度2K×C×N的单层图卷积神经网络（GCN)。但是，这种实现会导致超函",
"数中的参数过多，从而带来过拟合的风险，因为我们在实践中观察到性能下降。为了解决这个问题，我们通过将节点建模为两个",
"不同的块来分解来自通道的节点参数。一个块学习通道式参数αc.陀，另一个块学习节点式参数Yn。最终输出（αnc,phc）计算如",
"下：",
"ahe = Yna and βhe = Ynpk.",
"(5)",
"此分解限制了从2K×C×N到2K×（C+N）的超函数需要学习的参数。结果，提高了模型的通用性和计算效率",
"通道参数。为了对图G的全局上下文进行编码，同时尽可能减少可学习参数的数量，我们采用图扩散滤波器sPPR从A和X中挤",
"压和提取拓扑信息。我们通过完全个性化的PageRank方案传播输人矩阵X：",
"E = sPPRx = α (In - (1 - α)D-1/2 AD-1/2)-",
"1X",
"(6)",
"具有传送（或重新启动）概率αE（0,1)，其中D是度矩阵，D-1/2AD-1/2表示归一化邻接矩阵A。我们对扩散结果进行平均，并",
"使用线性变换和tanh函数将其转换为通道参数（α,）。我们使用tanh(）将（αk,β）归一化为[-1,1]范围内，目的是控制参数",
"比例。我们将heRC定义为扩散结果的平均值，将PE R2xKxC定义为通道参数。信道块的计算方式为：",
"P = tanh(MLP(e)",
"IN",
"wheree=",
"NZen",
"(7)",
"n=1",
"节点参数。节点式模块计算每个节点的参数。我们还使用扩散矩阵来捕获图信息。为了获得每个节点n的参数Yn，而不是平均",
"扩散结果，我们使用MLP直接将每个e压缩到一个维度中。为了反映每个节点的重要性，应用softmax将yn规范化为：",
"exp(MLP(en)",
"Y",
"(8)",
"A",
"EN=,exp(MLP(en))",
"将步骤插人更新函数。通过结合方程（8)和(7），我们通过(5)获得了GReLU的全部参数。我们将GReLU插入第I层更新函数：",
"h(l+1) = 0 (()[M,m(1),",
"(9)",
"，其中(）是第1层中的GReLU。如图？?所示，在更新函数中，每层1有一组参数e(），而GReLU是通道和鼻子自适应的"
] | {
"arxiv_id": "2202.06281",
"language": "zh",
"page_idx": 2
} |
||
2202.06281-zh-3 | 2202.06281-zh-3.jpg | WWW'22,April 25-29, 2022, Virtual Event, Lyon, France
Yifei Zhang and Hao Zhu and Ziqiao Meng and Piotr Koniusz and Irwin King
<FigureHere>
Figure 2: Plots to the left and middle side are the piecewise functions in GReLU with $K=3$ and $K= 2. \textbf{They cut the input value}$ $x$ in different ranges $r_1,r_2,r_3$ and re-scale the $x$ w.r.t.$r. \textbf{The plot to the right is the weighted function used in GAT: the inputis and the the the the the the the the the x w. r. t. r. The plot the right is the weighted function used in GAT: the }$ is re-scaled regardless of the value range of $x.$
在光谱 GNN 中的应用。我们还注意到，所提出的方法也适用于光谱 GNN,而不仅仅是空间 GNN, 而不仅仅是空间 GNN。具体来说，谱域中的图形卷积屇囘埨冩樆裤袼俌号铂净和$_\text{磐城}{\text{后跟鹈诲函数，如下所示：??这里，}\sigma(\cdot)\text{是激活函数，}H^{(l)}\text{是}l\text{ 中节点的隐藏表示。我们的方法}}$
3.3 $\textbf{Understanding GReLU with Weighted Neighborhood and Attention Mechanism}$
下面，我们通过加权邻居和注意机制的镜头来分析 GReLU。为此，我们首先介绍图注意网络 (GAT)$\mathbb{E}[33]$中采用的所谓的掩蔽注
意，它采用以下形式：
(10)
$$\boldsymbol{h}_{n}^{\prime}=\alpha_{n}\boldsymbol{h}_{n}\:\mathrm{and}\:\alpha_{n}=\mathrm{MaskAttention}(H,A),$$
其中$\alpha_n$是分配给节点$n$的注意系数。请注意，(10) 具有 GReLU 的线性形式，其中通过设置$\beta=0$和$K=1$省略了偏置参数和$\max(\cdot,\cdot)$。当$K=1$ 时，GReLU 的分段函数中只有一个段，非线性 max 运算简化为线性变换，因此可以将其删除。插图如图 2 所示。GAT 的成功是由于通过在聚合过程中对重要节点信息进行不同的加权来选择它们。虽然 GAT 能够考虑单个节点的差异，但它无法区分通道之间的差异。GReLU 通过使参数适应节点和通道来解决此问题：
$$h_{nc}^{\prime}=\max_{1\leq k\leq K}\{\alpha_{nc}^{k}h_{nc}+\beta_{nc}^{k}\}\:\mathrm{where}\:\{(\alpha^{k},\beta^{k})\}_{k=1}^{K}=\omega(X,A),$$
(11)
其中$\boldsymbol{\alpha}^k,\boldsymbol{\beta}^k\in\mathbb{R}^{N\times\mathbb{C}}$。因此，GReLU 不是学习所有渠道共享的单个参数，而是学习多个权重来评估不同渠道的重要性。此外，添加$\max(\cdot)$使 GReLU 成为分段线性函数，可在不同范围内切割输人值并以不同方式重新缩放它们。这进一步使 GReLU 能够考虑不同值范围的影响
特别是，当$K=1$时，GReLU简化为分配单个权重的线性变换，而不管输人的范围如何。GReLU 的表现类似于全局注意力 [22,33]
中描述的注意力机制。在这种特殊情况下，这些变换也与线性图网络 SGC$\overline{\mathrm{E}}[35]$和 S$^2$GC$\overline{\mathrm{E}}[41]$有关
# 3.4 Computational Complexity.
下面，我们分析 GReLU 的计算复杂性，这是有效的，因为我们解耦节点和通道参数。此外，与 GAT 中采用的多头注意力机制
(MA)相比，GReLU的计算成本更好。
3.4.1 Time Complexity.GReLU 的复杂性由两部分组成：(i) 用于推断 GReLU 参数的超函数$\omega_{(X,A)}$的复杂性。(Ii) 输出激活的激
活函数$\sigma_{\omega_{(X,A)}}(\cdot)$的复杂性。
i. Hyperfunction includes two components: (1) computation of the channel-wise parameters which consists of computing the average of
$\ddot{\text{diffusion output and MLP layer (with tanh()), and (2) the computation of node-wise parameters which includes the cost of diffusion}}$ and one-layer MLP with softmax. In the channel-wise block, GReLU spends $O(|\mathcal{E}|)$ in the diffusion layer and $O(2KC)$ in the MLP. In the node-wise block, GReLU spends $O(|\mathcal{E}|)$ in diffusion and $O(|\mathcal{V}|)$ in MLP with softmax.
ii. Activation function requires that GReLU is applied per channel of each node, which requires $O(C|^{\prime}V|)$ to compute outputs.
因此，GReLU 具有$O(|\mathcal{E}|+C|^{\prime}V|)$的复杂性。由于激活函数$O(C|\mathcal{V}|)$的计算成本对于所有整流器都是相同的，因此额外的成本主要是由超函数的计算引人的。由于超函数的轻型设计：1)扩散输出的平均导致单个向量$e\in\mathbb{R}^{\mathbb{C}};$ 2)节点块中的 MLP 仅包含一个输出维度，并且3)$K$ 的选择是 2(在$\mathbb{E}?$ 节中讨论),与 GNN 的成本相比，超函数的计算成本可以忽略不计。
3.4.2 $Complexity of CReLU vs. Multi- head Attention Mechanism. GReLU 比 GAT 中 使 用 的 多 头 注 意 力 机 制 更 快 。请 注 意 , MA 需 要$ 计算$M$头的$|\mathcal{E}|$边的系数，从而导致$O(M|\mathcal{E}|)$ ,而 GReLU 仅计算$C$通道和$|\mathcal{V}|$节点的系数，从而导致$O(|\mathcal{V}|+C)$复杂度。由于与$|\mathcal{V}|$和$|\mathcal{E}|$相比，$|\mathcal{V}|>>|\mathcal{E}|$和$C$是一个相对较小的数字，因此采用具有简单主干 $(e$g.,GCN)的GReLU比GAT快 | [
[
"0.08826",
"0.07574",
"0.343",
"0.07764",
"0.343",
"0.0884",
"0.08826",
"0.0859"
],
[
"0.55",
"0.077",
"0.911",
"0.077",
"0.911",
"0.0871",
"0.55",
"0.0871"
],
[
"0.2778",
"0.125",
"0.3057",
"0.127",
"0.304",
"0.1351",
"0.2769",
"0.1332"
],
[
"0.183",
"0.2644",
"0.5986",
"0.2664",
"0.5986",
"0.286",
"0.183",
"0.2842"
],
[
"0.709",
"0.2715",
"0.738",
"0.2715",
"0.738",
"0.286",
"0.709",
"0.286"
],
[
"0.0866",
"0.3005",
"0.911",
"0.3005",
"0.911",
"0.3145",
"0.0866",
"0.3145"
],
[
"0.08826",
"0.315",
"0.558",
"0.315",
"0.558",
"0.33",
"0.08826",
"0.33"
],
[
"0.5728",
"0.3208",
"0.625",
"0.3208",
"0.625",
"0.324",
"0.5728",
"0.324"
],
[
"0.661",
"0.3208",
"0.681",
"0.3208",
"0.681",
"0.324",
"0.661",
"0.324"
],
[
"0.687",
"0.3208",
"0.7134",
"0.3208",
"0.7134",
"0.324",
"0.687",
"0.324"
],
[
"0.732",
"0.3208",
"0.7427",
"0.3208",
"0.7427",
"0.324",
"0.732",
"0.324"
],
[
"0.773",
"0.3176",
"0.9087",
"0.3176",
"0.9087",
"0.3276",
"0.773",
"0.3276"
],
[
"0.0866",
"0.3289",
"0.3938",
"0.33",
"0.3938",
"0.3435",
"0.0866",
"0.3423"
],
[
"0.08826",
"0.3662",
"0.911",
"0.3662",
"0.911",
"0.3801",
"0.08826",
"0.3801"
],
[
"0.0866",
"0.3806",
"0.914",
"0.377",
"0.914",
"0.3958",
"0.0866",
"0.3997"
],
[
"0.0866",
"0.414",
"0.773",
"0.414",
"0.773",
"0.428",
"0.0866",
"0.428"
],
[
"0.08826",
"0.432",
"0.9126",
"0.432",
"0.9126",
"0.4458",
"0.08826",
"0.4458"
],
[
"0.0866",
"0.4475",
"0.2255",
"0.4475",
"0.2255",
"0.4583",
"0.0866",
"0.4583"
],
[
"0.3635",
"0.459",
"0.634",
"0.4583",
"0.634",
"0.473",
"0.3635",
"0.4734"
],
[
"0.888",
"0.4622",
"0.9126",
"0.4622",
"0.9126",
"0.4749",
"0.888",
"0.4749"
],
[
"0.0866",
"0.4778",
"0.9126",
"0.4778",
"0.9126",
"0.4912",
"0.0866",
"0.4912"
],
[
"0.08826",
"0.4924",
"0.9126",
"0.4924",
"0.9126",
"0.506",
"0.08826",
"0.506"
],
[
"0.08496",
"0.504",
"0.914",
"0.505",
"0.914",
"0.5195",
"0.08496",
"0.518"
],
[
"0.0866",
"0.5195",
"0.596",
"0.5195",
"0.596",
"0.5327",
"0.0866",
"0.5327"
],
[
"0.3057",
"0.537",
"0.6895",
"0.535",
"0.6895",
"0.556",
"0.3057",
"0.558"
],
[
"0.8857",
"0.5415",
"0.915",
"0.5415",
"0.915",
"0.558",
"0.8857",
"0.558"
],
[
"0.0866",
"0.5645",
"0.914",
"0.5645",
"0.914",
"0.582",
"0.0866",
"0.582"
],
[
"0.08826",
"0.5806",
"0.911",
"0.5806",
"0.911",
"0.5947",
"0.08826",
"0.5947"
],
[
"0.08826",
"0.5967",
"0.1896",
"0.5967",
"0.1896",
"0.608",
"0.08826",
"0.608"
],
[
"0.10297",
"0.608",
"0.911",
"0.608",
"0.911",
"0.622",
"0.10297",
"0.622"
],
[
"0.08496",
"0.622",
"0.734",
"0.6226",
"0.734",
"0.637",
"0.08496",
"0.636"
],
[
"0.08496",
"0.6494",
"0.3586",
"0.6494",
"0.3586",
"0.6636",
"0.08496",
"0.6636"
],
[
"0.0866",
"0.6665",
"0.9126",
"0.6675",
"0.9126",
"0.6816",
"0.0866",
"0.681"
],
[
"0.08496",
"0.6816",
"0.3284",
"0.6816",
"0.3284",
"0.696",
"0.08496",
"0.696"
],
[
"0.0866",
"0.7026",
"0.9126",
"0.7026",
"0.9126",
"0.7173",
"0.0866",
"0.7173"
],
[
"0.08496",
"0.7163",
"0.2615",
"0.7173",
"0.2615",
"0.732",
"0.08496",
"0.731"
],
[
"0.0907",
"0.7344",
"0.914",
"0.7344",
"0.914",
"0.7476",
"0.0907",
"0.7476"
],
[
"0.1136",
"0.7485",
"0.9087",
"0.7485",
"0.9087",
"0.762",
"0.1136",
"0.762"
],
[
"0.1152",
"0.761",
"0.911",
"0.761",
"0.911",
"0.7744",
"0.1152",
"0.7744"
],
[
"0.1152",
"0.7764",
"0.6504",
"0.7764",
"0.6504",
"0.787",
"0.1152",
"0.787"
],
[
"0.08905",
"0.7915",
"0.846",
"0.7915",
"0.846",
"0.802",
"0.08905",
"0.802"
],
[
"0.08826",
"0.808",
"0.91",
"0.808",
"0.91",
"0.8184",
"0.08826",
"0.8184"
],
[
"0.08905",
"0.8213",
"0.91",
"0.8213",
"0.91",
"0.832",
"0.08905",
"0.832"
],
[
"0.0866",
"0.834",
"0.796",
"0.834",
"0.796",
"0.8477",
"0.0866",
"0.8477"
],
[
"0.08905",
"0.857",
"0.911",
"0.857",
"0.911",
"0.867",
"0.08905",
"0.867"
],
[
"0.08496",
"0.8687",
"0.911",
"0.869",
"0.911",
"0.8833",
"0.08496",
"0.882"
],
[
"0.08826",
"0.884",
"0.8613",
"0.884",
"0.8613",
"0.898",
"0.08826",
"0.898"
]
] | [
"WWW'22, April2529,2022, Virtual Event, Lyon, Fran",
"Yifei Zhang and Hao Zhu and Ziqiao Meng and Piotr Koniusz and Irwin King",
"112",
"max1<k≤3(ha+β]max1<k≤2(Qha+β]",
"ar",
"Figure 2: Plots to the left and middle side are the piecewise functions in GReLU with K = 3 and K = 2 . They cut the input value",
"x in different ranges ri,r2, r3 and re-scale the x w.r.t.r . The plot to the",
"",
"",
"",
"",
"sedinGAT:theinpu",
"is re-scaled regardless of the value range of x .",
"在光谱GNN中的应用。我们还注意到，所提出的方法也适用于光谱GNN，而不仅仅是空间GNN。具体来说，谱域中的图形卷积",
"寻臀蕈角韵总臀换跟激数，如下所示：???这里，α(）是激活函数，H(）是1中节点的隐藏表示。。我们的方法",
"3.3Understanding GReLU with Weighted Neighborhood and Attention Mechanism",
"下面，我们通过加权邻居和注意机制的镜头来分析GReLU。为此，我们首先介绍图注意网络（GAT)E[33]中采用的所谓的掩蔽注",
"意，它采用以下形式：",
"h'n = αnhn and αn = MaskAttention(H,A),",
"(10)",
"其中αn是分配给节点n的注意系数。请注意，（10)具有GReLU的线性形式，其中通过设置β=0和K=1省略了偏置参数和",
"max，）。当K=1时，GReLU的分段函数中只有一个段，非线性max运算简化为线性变换，因此可以将其删除。插图如图2所",
"示。GAT的成功是由于通过在聚合过程中对重要节点信息进行不同的加权来选择它们。虽然GAT能够考虑单个节点的差异，但",
"它无法区分通道之间的差异。GReLU通过使参数适应节点和通道来解决此问题：",
"hne =max (ahchne + βhe) where (ak,pk)K-, = w(X, A),",
"(11)",
"其中αk,βkeRNxC。因此，GReLU不是学习所有渠道共享的单个参数，而是学习多个权重来评估不同渠道的重要性。此外，添",
"加max(）使GReLU成为分段线性函数，可在不同范围内切割输人值并以不同方式重新缩放它们。这进一步使GReLU能够考虑不",
"同值范围的影响",
"特别是，当K=1时，GReLU简化为分配单个权重的线性变换，而不管输人的范围如何。GReLU的表现类似于全局注意力[22,33]",
"中描述的注意力机制。在这种特殊情况下，这些变换也与线性图网络SGCE[35]和s2GCE[41]有关",
"3.4Computational Complexity.",
"下面，我们分析GReLU的计算复杂性，这是有效的，因为我们解耦节点和通道参数。此外，与GAT中采用的多头注意力机制",
"（MA）相比，GReLU的计算成本更好。",
"3.4.1TimeComplexity.GReLU的复杂性由两部分组成：(i）用于推断GReLU参数的超函数(X,A)的复杂性。(li)输出激活的激",
"活函数(x（）的复杂性。",
"i.Hyperfunction includes two components: (1) computation of the channel-wise parameters which consists of computing the average of",
"diffusion output and MLPlayer (with tanh()),and (2) the computation of node-wise parameters which includes the cost of diffusior",
"and one-layer MLP with softmax. In the channel-wise block, GReLU spends O(J)) in the diffusion layer and O(2KC) in the MLP. In",
"thenode-wiseblock,GReLUspendsO(D)indiffusionandO(JVD)inMLPwithsoftmax",
"ii.ActivationfunctionrequiresthatGReLUisappliedperchannelofeachnode,whichrequiresO(CVDtocomputeoutputs",
"因此，GReLU具有O8+CVD的复杂性。由于激活函数O(CV）的计算成本对于所有整流器都是相同的，因此额外的成本主",
"要是由超函数的计算引入的。由于超函数的轻型设计：1）扩散输出的平均导致单个向量eER；2）节点块中的MLP仅包含一个",
"输出维度，并且3）K的选择是2（在E?？节中讨论），与GNN的成本相比，超函数的计算成本可以忽略不计。",
"3.4.2ComplexityofGReLUvs.Multi-headAttentionMechanismGReLU比GAT中使用的多头注意力机制更快。请注意，MA需要",
"计算M头的ISI边的系数，从而导致O(M8I），而GReLU仅计算C通道和IVI节点的系数，从而导致O(IVI+C)复杂度。由于",
"与IVI和|8|相比，IVI>>I8I和C是一个相对较小的数字，因此采用具有简单主干（e.g.，GCN）的GReLU比GAT快。"
] | {
"arxiv_id": "2202.06281",
"language": "zh",
"page_idx": 3
} |
||
2202.06281-zh-4 | 2202.06281-zh-4.jpg | WWW'22, April 25-29, 2022, Virtual Event, Lyon, France.
Graph-adaptive Rectified Linear Unit for Graph Neural Networks
Table 2: The average node classification accuracy (over 10 runs). The training set and test set are randomly divided. The results in parentheses are the best results achieved among the 10 experiments. We omit the results of Cheb in BlogCatalog and Flickr dataset due to the out-of-memory issue during the training step.
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>Model</th>
<th>ReLU</th>
<th>LReLU</th>
<th>$ELU$</th>
<th>PReLU</th>
<th>Maxout</th>
<th>GReLU</th>
</tr>
<tr>
<td rowspan="6">Cora</td>
<td rowspan="6">Cheb $GCN$ $SAGE$ $GAT$ ARMA APPNP</td>
<td>$76.5\pm2.2(79.4)$</td>
<td>$75.5\pm2.2(80.0)$</td>
<td>$74.9\pm2.9(80.4)$</td>
<td>$75.6\pm3.0(80.6)$</td>
<td>$75.4\pm2.0(77.8)$</td>
<td>$78.3\pm1.7\left(79.4\right)$</td>
</tr>
<tr>
<td>$79.2\pm1.4(81.6)$</td>
<td>$79.4\pm1.6(80.6)$</td>
<td>$78.8\pm1.3(80.2)$</td>
<td>$78.9\pm1.0(80.2)$</td>
<td>$79.8\pm1.5(81.5)$</td>
<td>$81.8\pm1.8\left(83.0\right)$</td>
</tr>
<tr>
<td>$78.6\pm1.7(81.3)$</td>
<td>$77.6\pm1.8(79.9)$</td>
<td>$77.5\pm1.4(79.9)$</td>
<td>$77.6\pm1.7(78.7)$</td>
<td>$77.7\pm1.7(79.9)$</td>
<td>$79.3\pm1.5\left(80.7\right)$</td>
</tr>
<tr>
<td>$81.2\pm1.3(83.3)$</td>
<td>$80.8\pm2.3(83.1)$</td>
<td>$80.5\pm1.7(83.0)$</td>
<td>$80.4\pm2.1(83.1)$</td>
<td>$80.2\pm1.7(82.5)$</td>
<td>$81.5\pm2.1$ 1 (84.1)</td>
</tr>
<tr>
<td>$79.0\pm1.4(80.8)$</td>
<td>$79.2\pm2.1(82.9)$</td>
<td>$79.7\pm0.8(80.6)$</td>
<td>$79.3\pm2.0(81.9)$</td>
<td>$79.8\pm1.3(81.3)$</td>
<td>$80.1\pm1.5\left(82.6\right)$</td>
</tr>
<tr>
<td>$82.0\pm1.3(83.1)$</td>
<td>$81.0\pm2.3(83.1)$</td>
<td>$81.5\pm1.7(83.4)$</td>
<td>$81.4\pm2.1(83.4)$</td>
<td>$81.0\pm1.7(82.5)$</td>
<td>$82.5\pm2.1$ 1 (84.7)</td>
</tr>
<tr>
<td rowspan="6">PuMed</td>
<td>Cheb</td>
<td>$68.1\pm4.1(74.9)$</td>
<td>$70.5\pm3.0(76.7)$</td>
<td>$68.4\pm2.8(72.7)$</td>
<td>$71.5\pm3.1(76.3)$</td>
<td>$71.8\pm3.9(75.8)$</td>
<td>$73.4\pm2.9\left(75.6\right)$</td>
</tr>
<tr>
<td>$GCN$</td>
<td>$77.6\pm2.2(81.6)$</td>
<td>$76.8\pm1.6(79.4)$</td>
<td>$76.8\pm2.2(80.5)$</td>
<td>$77.3\pm3.7(82.5)$</td>
<td>$77.3\pm2.9(80.6)$</td>
<td>$78.9\pm1.7\left(81.3\right)$</td>
</tr>
<tr>
<td>$SAGE$</td>
<td>$75.7\pm3.1(79.8)$</td>
<td>$75.3\pm3.3(79.8)$</td>
<td>$74.8\pm2.7(78.6)$</td>
<td>$76.0\pm2.5(80.5)$</td>
<td>$74.5\pm2.7(77.4)$</td>
<td>$76.2\pm1.6\left(78.4\right)$</td>
</tr>
<tr>
<td>$GAT$</td>
<td>$77.2\pm3.1(81.3)$</td>
<td>$78.7\pm1.7(81.2)$</td>
<td>$78.9\pm2.3(82.1)$</td>
<td>$76.2\pm3.0(80.7)$</td>
<td>$77.9\pm1.7(80.3)$</td>
<td>$79.1\pm1.8\left(80.6\right)$ I</td>
</tr>
<tr>
<td>ARMA</td>
<td>$76.9\pm2.6(80.7)$</td>
<td>$76.5\pm1.9(80.3)$</td>
<td>$77.3\pm2.5(80.4)$</td>
<td>$76.5\pm2.4(80.7)$</td>
<td>$76.6\pm2.9(81.3)$</td>
<td>$77.4\pm3.0\left(80.2\right)$</td>
</tr>
<tr>
<td>APPNP</td>
<td>$78.2\pm3.1(82.3)$</td>
<td>$78.2\pm1.7(81.7)$</td>
<td>$79.0\pm2.3(82.1)$</td>
<td>$79.2\pm3.0(80.7)$</td>
<td>$78.9\pm1.7(81.3)$</td>
<td>$79.8\pm1.8\left(81.2\right)$</td>
</tr>
<tr>
<td rowspan="6">CiteSeer</td>
<td>Cheb</td>
<td>$67.8\pm1.8(71.0)$</td>
<td>$67.1\pm2.9(71.1)$</td>
<td>$67.8\pm2.4(71.1)$</td>
<td>$67.0\pm2.3(70.5)$</td>
<td>$67.4\pm1.5(70.7)$</td>
<td>$68. 1\pm 1. 3$ (70.4)</td>
</tr>
<tr>
<td>$GCN$</td>
<td>$67.7\pm2.3(72.1)$</td>
<td>$68.4\pm1.8(71.2)$</td>
<td>$68.3\pm1.4(70.1)$</td>
<td>$67.3\pm2.1(70.7)$</td>
<td>$68.5\pm2.2(72.5)$</td>
<td>$68.5\pm1.9\left(71.7\right)$</td>
</tr>
<tr>
<td>$SAGE$</td>
<td>$67.1\pm2.8(70.1)$</td>
<td>$67.3\pm2.1(70.1)$</td>
<td>$67.8\pm1.7(70.2)$</td>
<td>$66.2\pm2.6(69.6)$</td>
<td>$67.5\pm1.8(71.5)$</td>
<td>$68.0\pm1.3\left(69.7\right)$</td>
</tr>
<tr>
<td>$GAT$</td>
<td>$68.6\pm1.4(70.8)$</td>
<td>$69.2\pm1.9(71.7)$</td>
<td>$68.4\pm1.6(71.2)$</td>
<td>$68.2\pm1.6(69.7)$</td>
<td>$68.6\pm1.6(71.3)$</td>
<td>$69.3\pm1.7\left(71.9\right)$</td>
</tr>
<tr>
<td>ARMA</td>
<td>$67.7\pm1.3(68.9)$</td>
<td>$68.6\pm2.4(71.5)$</td>
<td>$67.9\pm2.1(71.3)$</td>
<td>$66.8\pm1.5(69.4)$</td>
<td>$68.5\pm1.8(70.9)$</td>
<td>(71.7) $69.0\pm2.2$</td>
</tr>
<tr>
<td>APPNP</td>
<td>$68.7\pm1.3(70.5)$</td>
<td>$69.3\pm1.6(71.2)$</td>
<td>$69.4\pm1.4(71.2)$</td>
<td>$69.5\pm1.6(70.7)$</td>
<td>$69.2\pm1.6(72.0)$</td>
<td>$70.0\pm1.7\left(72.3\right)$</td>
</tr>
<tr>
<td rowspan="5">$\operatorname{BlogCatalog}$</td>
<td>$GCN$</td>
<td>$72.1\pm1.9(75.5)$</td>
<td>$72.6\pm2.1(75.2)$</td>
<td>$72.6\pm1.8(75.3)$</td>
<td>$71.4\pm2.1(74.5)$</td>
<td>$72.4\pm1.4(75.3)$</td>
<td>$73.7\pm1.2$ $2\left(74.2\right)$</td>
</tr>
<tr>
<td>$SAGE$</td>
<td>$71.9\pm1.3(76.0)$</td>
<td>$72.2\pm1.9(74.9)$</td>
<td>$72.1\pm1.8(74.3)$</td>
<td>$72.0\pm2.1(74.4)$</td>
<td>$71.6\pm1.4(74.3)$</td>
<td>$73.3\pm1.6\left(75.3\right)$</td>
</tr>
<tr>
<td>$GAT$</td>
<td>$41.7\pm7.3(54.4)$</td>
<td>$38.2\pm3.3(41.8)$</td>
<td>$46.6\pm3.6(51.7)$</td>
<td>$67.2\pm2.6(71.8)$</td>
<td>$54.2\pm3.9(59.3)$</td>
<td>$67.8\pm3.9(72.3)$</td>
</tr>
<tr>
<td>ARMA</td>
<td>$72.5\pm3.3(79.1)$</td>
<td>$72.5\pm5.9(78.7)$</td>
<td>$77.2\pm2.2(79.2)$</td>
<td>$79.6\pm3.0(84.5)$</td>
<td>$84.4\pm1.8(86.9)$</td>
<td>$85.7\pm2.7\left(88.4\right)$</td>
</tr>
<tr>
<td>APPNP</td>
<td>$71.1\pm1.8(75.3)$</td>
<td>$72.5\pm2.0(75.2)$</td>
<td>$72.4\pm1.8(75.3)$</td>
<td>$71.7\pm2.1(74.8)$</td>
<td>$72.8\pm1.4(75.3)$</td>
<td>$73.8\pm1.5\left(74.8\right)$</td>
</tr>
<tr>
<td rowspan="5">Flickr</td>
<td>$GCN$</td>
<td>$50.7\pm2.3(54.8)$</td>
<td>$51.0\pm2.0(53.8)$</td>
<td>$52.8\pm1.8(56.0)$</td>
<td>$53.0\pm1.6(54.9)$</td>
<td>$54.0\pm1.8(56.8)$</td>
<td>$54.4\pm1.6\left(56.8\right)$</td>
</tr>
<tr>
<td>$SAGE$</td>
<td>$49.7\pm2.2(53.8)$</td>
<td>$50.8\pm2.1(54.0)$</td>
<td>$52.6\pm1.8(56.7)$</td>
<td>$53.2\pm1.3(56.1)$</td>
<td>$54.0\pm1.7(56.5)$</td>
<td>$55.3\pm1.4\left(57.2\right)$</td>
</tr>
<tr>
<td>$GAT$</td>
<td>$20.2\pm3.1(25.2)$</td>
<td>$20.3\pm2.5(23.8)$</td>
<td>$23.8\pm2.9(28.1)$</td>
<td>$32.8\pm4.9(43.2)$</td>
<td>$30.0\pm2.6(34.6)$</td>
<td>$33.7\pm3.1$ (36.3)</td>
</tr>
<tr>
<td>ARMA</td>
<td>$48.4\pm4.7(56.1)$</td>
<td>$52.0\pm3.9(59.0)$</td>
<td>$52.2\pm4.1(58.1)$</td>
<td>$45.9\pm3.5(53.1)$</td>
<td>$52.7\pm4.0(59.4)$</td>
<td>$56.5\pm2.4\left(59.1\right)$</td>
</tr>
<tr>
<td>APPNP</td>
<td>$51.6\pm2.0(54.0)$</td>
<td>$51.6\pm2.1(53.8)$</td>
<td>$53.0\pm1.8(56.0)$</td>
<td>$53.2\pm1.4(54.9)$</td>
<td>$53.9\pm1.9(56.6)$</td>
<td>$54.8\pm1.9\left(56.7\right)$</td>
</tr>
</tbody>
</table>
## 4 EXPERIMENTS
下面，我们在两个任务上评估所提出的方法 GReLU, 即$\boxed{\mathrm{E}}4.1$ 节中的节点分类和$\boxed{\mathrm{E}}?$ 节中的图形分类。我们通过将其与具有多个GNN 骨架的不同 ReLU 变体进行比较来显示 GReLU 的有效性。我们在 ?? 节的消融研究中展示了 GReLU 不同模块的贡献。在？? 节中讨论了选择超参数$K$的影响。为了验证 GReLU 的自适应行为，在 ?? 节中进行了参数分析。我们的目标是回答以下研究问题(RQ):
$\bullet$ RQ1: Does applying GReLU in the updating step help improve the capacity of GNNs for different downstream tasks?
$\bullet$ RQ2: What kind of factors does GReLU benefit from in its design and how do they contribute to performance?
$\bullet$ RQ3: How does the choice of parameter K affect the performance of GReLU?
$\bullet$ RQ4: Does GReLU dynamically adapt to different nodes?
# 4.1 $\textbf{Node Classification ( RQ1) }$
数据集。GReLU 在五个实际数据集上进行评估：
$\bullet$ Cora, PubMed, CiteSeer [16] are well-known citation network datasets, where nodes represent papers and edges represent their
citations, and the nodes are labeled based on the paper topics.
$\bullet$ Flick[25] is an image and video hosting website, where users interact with each other via photo sharing. It is a social network where nodes represent users and edges represent their relationships, and all nodes are divided into 9 classes according to the interest groups of users.
$\bullet$ BlogCatalog [25] is a social network for bloggers and their social relationships from the BlogCatalog website. Node attributes are constructed by the keywords of user profiles, and the labels represent the topic categories provided by the authors, and all nodes are divided into 6 classes. | [
[
"0.0866",
"0.07764",
"0.3945",
"0.07764",
"0.3945",
"0.0884",
"0.0866",
"0.0884"
],
[
"0.638",
"0.0745",
"0.911",
"0.07574",
"0.911",
"0.089",
"0.638",
"0.0884"
],
[
"0.08496",
"0.1061",
"0.9126",
"0.1061",
"0.9126",
"0.1193",
"0.08496",
"0.1193"
],
[
"0.08496",
"0.1193",
"0.9126",
"0.1193",
"0.9126",
"0.1332",
"0.08496",
"0.1332"
],
[
"0.08496",
"0.1338",
"0.513",
"0.1338",
"0.513",
"0.1477",
"0.08496",
"0.1477"
],
[
"0.298",
"0.1674",
"0.335",
"0.1674",
"0.335",
"0.1831",
"0.298",
"0.1831"
],
[
"0.3906",
"0.1674",
"0.4355",
"0.1674",
"0.4355",
"0.1818",
"0.3906",
"0.1818"
],
[
"0.584",
"0.1674",
"0.6274",
"0.1674",
"0.6274",
"0.1818",
"0.584",
"0.1818"
],
[
"0.6772",
"0.1654",
"0.726",
"0.1674",
"0.7256",
"0.1831",
"0.6763",
"0.1812"
],
[
"0.778",
"0.1674",
"0.822",
"0.1674",
"0.822",
"0.1818",
"0.778",
"0.1818"
],
[
"0.4958",
"0.1686",
"0.524",
"0.1686",
"0.524",
"0.1812",
"0.4958",
"0.1812"
],
[
"0.1487",
"0.1692",
"0.1936",
"0.1692",
"0.1936",
"0.1812",
"0.1487",
"0.1812"
],
[
"0.2197",
"0.1692",
"0.2615",
"0.1692",
"0.2615",
"0.1812",
"0.2197",
"0.1812"
],
[
"0.5596",
"0.1843",
"0.649",
"0.1863",
"0.649",
"0.2008",
"0.5596",
"0.1989"
],
[
"0.7534",
"0.1849",
"0.8447",
"0.1875",
"0.8447",
"0.202",
"0.7534",
"0.1995"
],
[
"0.661",
"0.1863",
"0.744",
"0.1863",
"0.744",
"0.1976",
"0.661",
"0.1976"
],
[
"0.156",
"0.1875",
"0.1854",
"0.1875",
"0.1854",
"0.1995",
"0.156",
"0.1995"
],
[
"0.2222",
"0.1875",
"0.2573",
"0.1875",
"0.2573",
"0.1995",
"0.2222",
"0.1995"
],
[
"0.273",
"0.1863",
"0.3586",
"0.1863",
"0.3586",
"0.1976",
"0.273",
"0.1976"
],
[
"0.3708",
"0.1875",
"0.455",
"0.1875",
"0.455",
"0.1989",
"0.3708",
"0.1989"
],
[
"0.4673",
"0.1875",
"0.5513",
"0.1875",
"0.5513",
"0.1989",
"0.4673",
"0.1989"
],
[
"0.2239",
"0.2008",
"0.2573",
"0.2008",
"0.2573",
"0.2134",
"0.2239",
"0.2134"
],
[
"0.2737",
"0.2008",
"0.3594",
"0.2008",
"0.3594",
"0.2122",
"0.2737",
"0.2122"
],
[
"0.3694",
"0.2008",
"0.455",
"0.2008",
"0.455",
"0.2122",
"0.3694",
"0.2122"
],
[
"0.4673",
"0.2008",
"0.5513",
"0.2008",
"0.5513",
"0.2122",
"0.4673",
"0.2122"
],
[
"0.562",
"0.2008",
"0.649",
"0.2008",
"0.649",
"0.2122",
"0.562",
"0.2122"
],
[
"0.66",
"0.2008",
"0.746",
"0.2008",
"0.746",
"0.2122",
"0.66",
"0.2122"
],
[
"0.7524",
"0.1976",
"0.8447",
"0.1995",
"0.8447",
"0.214",
"0.7524",
"0.2122"
],
[
"0.2214",
"0.2102",
"0.259",
"0.2122",
"0.2573",
"0.2279",
"0.2197",
"0.226"
],
[
"0.2737",
"0.2134",
"0.3586",
"0.2134",
"0.3586",
"0.2247",
"0.2737",
"0.2247"
],
[
"0.3708",
"0.2134",
"0.4558",
"0.2134",
"0.4558",
"0.2247",
"0.3708",
"0.2247"
],
[
"0.4673",
"0.2134",
"0.5513",
"0.2134",
"0.5513",
"0.2247",
"0.4673",
"0.2247"
],
[
"0.562",
"0.2134",
"0.649",
"0.2134",
"0.649",
"0.2247",
"0.562",
"0.2247"
],
[
"0.66",
"0.2134",
"0.744",
"0.2134",
"0.744",
"0.2247",
"0.66",
"0.2247"
],
[
"0.7534",
"0.2102",
"0.846",
"0.2122",
"0.846",
"0.2267",
"0.7534",
"0.2247"
],
[
"0.2255",
"0.226",
"0.2566",
"0.226",
"0.2566",
"0.238",
"0.2255",
"0.238"
],
[
"0.2737",
"0.226",
"0.3586",
"0.226",
"0.3586",
"0.2374",
"0.2737",
"0.2374"
],
[
"0.3708",
"0.2267",
"0.4558",
"0.2267",
"0.4558",
"0.2374",
"0.3708",
"0.2374"
],
[
"0.4673",
"0.2267",
"0.5513",
"0.2267",
"0.5513",
"0.2374",
"0.4673",
"0.2374"
],
[
"0.562",
"0.226",
"0.649",
"0.226",
"0.649",
"0.2374",
"0.562",
"0.2374"
],
[
"0.66",
"0.2267",
"0.744",
"0.2267",
"0.744",
"0.2374",
"0.66",
"0.2374"
],
[
"0.755",
"0.2267",
"0.8447",
"0.2267",
"0.8447",
"0.238",
"0.755",
"0.238"
],
[
"0.5605",
"0.2349",
"0.648",
"0.2374",
"0.648",
"0.252",
"0.5605",
"0.25"
],
[
"0.7534",
"0.2349",
"0.8447",
"0.2374",
"0.8447",
"0.252",
"0.7534",
"0.25"
],
[
"0.4656",
"0.2374",
"0.5513",
"0.2393",
"0.5513",
"0.2507",
"0.4656",
"0.2488"
],
[
"0.2181",
"0.2374",
"0.2632",
"0.2374",
"0.2632",
"0.252",
"0.2181",
"0.252"
],
[
"0.2737",
"0.238",
"0.3586",
"0.238",
"0.3586",
"0.25",
"0.2737",
"0.25"
],
[
"0.3718",
"0.238",
"0.4558",
"0.238",
"0.4558",
"0.25",
"0.3718",
"0.25"
],
[
"0.6587",
"0.2374",
"0.744",
"0.2393",
"0.744",
"0.2507",
"0.6587",
"0.2488"
],
[
"0.2157",
"0.252",
"0.2632",
"0.252",
"0.2632",
"0.2632",
"0.2157",
"0.2632"
],
[
"0.2737",
"0.252",
"0.3586",
"0.252",
"0.3586",
"0.2632",
"0.2737",
"0.2632"
],
[
"0.3694",
"0.252",
"0.455",
"0.252",
"0.455",
"0.2632",
"0.3694",
"0.2632"
],
[
"0.4673",
"0.252",
"0.5522",
"0.252",
"0.5522",
"0.2632",
"0.4673",
"0.2632"
],
[
"0.562",
"0.252",
"0.648",
"0.252",
"0.648",
"0.2632",
"0.562",
"0.2632"
],
[
"0.66",
"0.252",
"0.744",
"0.252",
"0.744",
"0.2632",
"0.66",
"0.2632"
],
[
"0.7524",
"0.2488",
"0.8447",
"0.2507",
"0.8447",
"0.2651",
"0.7524",
"0.2632"
],
[
"0.1487",
"0.2676",
"0.1936",
"0.2676",
"0.1936",
"0.2822",
"0.1487",
"0.2822"
],
[
"0.2737",
"0.2676",
"0.3594",
"0.2676",
"0.3594",
"0.2822",
"0.2737",
"0.2822"
],
[
"0.3708",
"0.2676",
"0.4575",
"0.2676",
"0.4575",
"0.2822",
"0.3708",
"0.2822"
],
[
"0.6587",
"0.2651",
"0.746",
"0.2676",
"0.746",
"0.2822",
"0.6587",
"0.2798"
],
[
"0.755",
"0.2676",
"0.8447",
"0.2676",
"0.8447",
"0.2822",
"0.755",
"0.2822"
],
[
"0.4673",
"0.2695",
"0.5513",
"0.2695",
"0.5513",
"0.281",
"0.4673",
"0.281"
],
[
"0.2239",
"0.2683",
"0.2573",
"0.2683",
"0.2573",
"0.281",
"0.2239",
"0.281"
],
[
"0.562",
"0.2695",
"0.648",
"0.2695",
"0.648",
"0.281",
"0.562",
"0.281"
],
[
"0.562",
"0.281",
"0.648",
"0.281",
"0.648",
"0.2922",
"0.562",
"0.2922"
],
[
"0.755",
"0.281",
"0.846",
"0.281",
"0.846",
"0.2954",
"0.755",
"0.2954"
],
[
"0.2737",
"0.2822",
"0.3586",
"0.2822",
"0.3586",
"0.2935",
"0.2737",
"0.2935"
],
[
"0.2255",
"0.2822",
"0.2573",
"0.2822",
"0.2573",
"0.2942",
"0.2255",
"0.2942"
],
[
"0.3708",
"0.2822",
"0.4558",
"0.2822",
"0.4558",
"0.2935",
"0.3708",
"0.2935"
],
[
"0.4673",
"0.2822",
"0.5513",
"0.2822",
"0.5513",
"0.2935",
"0.4673",
"0.2935"
],
[
"0.661",
"0.2822",
"0.746",
"0.2822",
"0.746",
"0.2935",
"0.661",
"0.2935"
],
[
"0.2197",
"0.2935",
"0.2607",
"0.2935",
"0.2607",
"0.3093",
"0.2197",
"0.3093"
],
[
"0.2737",
"0.2954",
"0.3586",
"0.2954",
"0.3586",
"0.307",
"0.2737",
"0.307"
],
[
"0.3708",
"0.2954",
"0.4558",
"0.2954",
"0.4558",
"0.307",
"0.3708",
"0.307"
],
[
"0.4673",
"0.2954",
"0.5513",
"0.2954",
"0.5513",
"0.307",
"0.4673",
"0.307"
],
[
"0.562",
"0.2954",
"0.648",
"0.2954",
"0.648",
"0.307",
"0.562",
"0.307"
],
[
"0.66",
"0.2954",
"0.746",
"0.2954",
"0.746",
"0.307",
"0.66",
"0.307"
],
[
"0.755",
"0.2942",
"0.846",
"0.2942",
"0.846",
"0.3093",
"0.755",
"0.3093"
],
[
"0.2239",
"0.3062",
"0.259",
"0.3062",
"0.259",
"0.3213",
"0.2239",
"0.3213"
],
[
"0.464",
"0.305",
"0.5537",
"0.307",
"0.5537",
"0.3213",
"0.464",
"0.3193"
],
[
"0.755",
"0.307",
"0.846",
"0.307",
"0.846",
"0.3213",
"0.755",
"0.3213"
],
[
"0.3708",
"0.308",
"0.4558",
"0.308",
"0.4558",
"0.3193",
"0.3708",
"0.3193"
],
[
"0.2737",
"0.308",
"0.3586",
"0.308",
"0.3586",
"0.3193",
"0.2737",
"0.3193"
],
[
"0.562",
"0.308",
"0.649",
"0.308",
"0.649",
"0.3193",
"0.562",
"0.3193"
],
[
"0.66",
"0.308",
"0.746",
"0.308",
"0.746",
"0.3193",
"0.66",
"0.3193"
],
[
"0.464",
"0.3176",
"0.5537",
"0.3193",
"0.5537",
"0.334",
"0.464",
"0.332"
],
[
"0.5596",
"0.3176",
"0.6504",
"0.3193",
"0.6504",
"0.334",
"0.5596",
"0.332"
],
[
"0.6567",
"0.3176",
"0.7466",
"0.3193",
"0.7466",
"0.334",
"0.6567",
"0.332"
],
[
"0.2737",
"0.3193",
"0.3586",
"0.3193",
"0.3586",
"0.3308",
"0.2737",
"0.3308"
],
[
"0.2181",
"0.318",
"0.2632",
"0.318",
"0.2632",
"0.334",
"0.2181",
"0.334"
],
[
"0.3708",
"0.3208",
"0.455",
"0.3208",
"0.455",
"0.332",
"0.3708",
"0.332"
],
[
"0.755",
"0.3193",
"0.8447",
"0.3193",
"0.8447",
"0.334",
"0.755",
"0.334"
],
[
"0.2157",
"0.332",
"0.2646",
"0.332",
"0.2646",
"0.3467",
"0.2157",
"0.3467"
],
[
"0.4673",
"0.3333",
"0.5513",
"0.334",
"0.5513",
"0.3452",
"0.4673",
"0.3447"
],
[
"0.755",
"0.332",
"0.8447",
"0.332",
"0.8447",
"0.3467",
"0.755",
"0.3467"
],
[
"0.3718",
"0.334",
"0.455",
"0.334",
"0.455",
"0.3452",
"0.3718",
"0.3452"
],
[
"0.2737",
"0.334",
"0.3586",
"0.334",
"0.3586",
"0.3452",
"0.2737",
"0.3452"
],
[
"0.562",
"0.334",
"0.649",
"0.334",
"0.649",
"0.3447",
"0.562",
"0.3447"
],
[
"0.661",
"0.334",
"0.746",
"0.334",
"0.746",
"0.3452",
"0.661",
"0.3452"
],
[
"0.1478",
"0.3484",
"0.1969",
"0.351",
"0.1953",
"0.363",
"0.1462",
"0.361"
],
[
"0.5605",
"0.3499",
"0.648",
"0.3499",
"0.648",
"0.361",
"0.5605",
"0.361"
],
[
"0.7524",
"0.3484",
"0.8433",
"0.351",
"0.8433",
"0.3655",
"0.7524",
"0.363"
],
[
"0.2737",
"0.351",
"0.3586",
"0.351",
"0.3586",
"0.3623",
"0.2737",
"0.3623"
],
[
"0.2239",
"0.351",
"0.2573",
"0.351",
"0.2573",
"0.363",
"0.2239",
"0.363"
],
[
"0.3694",
"0.351",
"0.455",
"0.351",
"0.455",
"0.361",
"0.3694",
"0.361"
],
[
"0.4656",
"0.351",
"0.55",
"0.351",
"0.55",
"0.361",
"0.4656",
"0.361"
],
[
"0.66",
"0.351",
"0.746",
"0.351",
"0.746",
"0.361",
"0.66",
"0.361"
],
[
"0.2222",
"0.3623",
"0.259",
"0.3623",
"0.259",
"0.3774",
"0.2222",
"0.3774"
],
[
"0.2737",
"0.3643",
"0.3586",
"0.3643",
"0.3586",
"0.3757",
"0.2737",
"0.3757"
],
[
"0.3708",
"0.3643",
"0.4558",
"0.3643",
"0.4558",
"0.3757",
"0.3708",
"0.3757"
],
[
"0.4673",
"0.3643",
"0.5513",
"0.3643",
"0.5513",
"0.3757",
"0.4673",
"0.3757"
],
[
"0.562",
"0.3643",
"0.648",
"0.3643",
"0.648",
"0.3757",
"0.562",
"0.3757"
],
[
"0.66",
"0.3643",
"0.744",
"0.3643",
"0.744",
"0.3757",
"0.66",
"0.3757"
],
[
"0.7534",
"0.363",
"0.8423",
"0.363",
"0.8423",
"0.375",
"0.7534",
"0.375"
],
[
"0.2214",
"0.377",
"0.259",
"0.377",
"0.259",
"0.3901",
"0.2214",
"0.3901"
],
[
"0.2737",
"0.377",
"0.3586",
"0.377",
"0.3586",
"0.3882",
"0.2737",
"0.3882"
],
[
"0.3708",
"0.377",
"0.4558",
"0.377",
"0.4558",
"0.3882",
"0.3708",
"0.3882"
],
[
"0.4673",
"0.377",
"0.55",
"0.377",
"0.55",
"0.3882",
"0.4673",
"0.3882"
],
[
"0.562",
"0.3757",
"0.648",
"0.3757",
"0.648",
"0.387",
"0.562",
"0.387"
],
[
"0.66",
"0.377",
"0.746",
"0.377",
"0.746",
"0.3882",
"0.66",
"0.3882"
],
[
"0.755",
"0.377",
"0.8447",
"0.377",
"0.8447",
"0.3882",
"0.755",
"0.3882"
],
[
"0.2239",
"0.3865",
"0.2607",
"0.3894",
"0.2573",
"0.404",
"0.2222",
"0.4016"
],
[
"0.2754",
"0.3882",
"0.3586",
"0.3882",
"0.3586",
"0.3997",
"0.2754",
"0.3997"
],
[
"0.3708",
"0.3894",
"0.4558",
"0.3894",
"0.4558",
"0.401",
"0.3708",
"0.401"
],
[
"0.468",
"0.3901",
"0.55",
"0.3901",
"0.55",
"0.401",
"0.468",
"0.401"
],
[
"0.562",
"0.3901",
"0.648",
"0.3901",
"0.648",
"0.4016",
"0.562",
"0.4016"
],
[
"0.66",
"0.3901",
"0.744",
"0.3901",
"0.744",
"0.4016",
"0.66",
"0.4016"
],
[
"0.755",
"0.3882",
"0.8447",
"0.3882",
"0.8447",
"0.4028",
"0.755",
"0.4028"
],
[
"0.2197",
"0.401",
"0.2632",
"0.401",
"0.2632",
"0.4153",
"0.2197",
"0.4153"
],
[
"0.2737",
"0.4028",
"0.3586",
"0.4028",
"0.3586",
"0.414",
"0.2737",
"0.414"
],
[
"0.3708",
"0.4028",
"0.4558",
"0.4028",
"0.4558",
"0.414",
"0.3708",
"0.414"
],
[
"0.468",
"0.4028",
"0.5513",
"0.4028",
"0.5513",
"0.414",
"0.468",
"0.414"
],
[
"0.562",
"0.4028",
"0.649",
"0.4028",
"0.649",
"0.414",
"0.562",
"0.414"
],
[
"0.66",
"0.4028",
"0.746",
"0.4028",
"0.746",
"0.414",
"0.66",
"0.414"
],
[
"0.755",
"0.4028",
"0.8433",
"0.4028",
"0.8433",
"0.414",
"0.755",
"0.414"
],
[
"0.2157",
"0.4136",
"0.2656",
"0.4136",
"0.2656",
"0.428",
"0.2157",
"0.428"
],
[
"0.2737",
"0.4153",
"0.3586",
"0.4153",
"0.3586",
"0.4268",
"0.2737",
"0.4268"
],
[
"0.3694",
"0.4153",
"0.455",
"0.4153",
"0.455",
"0.4268",
"0.3694",
"0.4268"
],
[
"0.468",
"0.4153",
"0.5522",
"0.4153",
"0.5522",
"0.4268",
"0.468",
"0.4268"
],
[
"0.562",
"0.4153",
"0.649",
"0.4153",
"0.649",
"0.4268",
"0.562",
"0.4268"
],
[
"0.661",
"0.4153",
"0.746",
"0.4153",
"0.746",
"0.4268",
"0.661",
"0.4268"
],
[
"0.755",
"0.4153",
"0.8433",
"0.4153",
"0.8433",
"0.4268",
"0.755",
"0.4268"
],
[
"0.1323",
"0.43",
"0.2091",
"0.432",
"0.2091",
"0.4463",
"0.1323",
"0.4443"
],
[
"0.273",
"0.4312",
"0.3594",
"0.4312",
"0.3594",
"0.4458",
"0.273",
"0.4458"
],
[
"0.755",
"0.4312",
"0.8447",
"0.4312",
"0.8447",
"0.4458",
"0.755",
"0.4458"
],
[
"0.3694",
"0.433",
"0.455",
"0.433",
"0.455",
"0.4443",
"0.3694",
"0.4443"
],
[
"0.2255",
"0.432",
"0.2573",
"0.432",
"0.2573",
"0.4443",
"0.2255",
"0.4443"
],
[
"0.4673",
"0.433",
"0.5513",
"0.433",
"0.5513",
"0.4443",
"0.4673",
"0.4443"
],
[
"0.562",
"0.433",
"0.648",
"0.433",
"0.648",
"0.4443",
"0.562",
"0.4443"
],
[
"0.661",
"0.433",
"0.746",
"0.433",
"0.746",
"0.4443",
"0.661",
"0.4443"
],
[
"0.4656",
"0.4426",
"0.5522",
"0.4443",
"0.5522",
"0.459",
"0.4656",
"0.457"
],
[
"0.5596",
"0.4426",
"0.6504",
"0.4443",
"0.6504",
"0.459",
"0.5596",
"0.457"
],
[
"0.6587",
"0.4426",
"0.746",
"0.4443",
"0.746",
"0.459",
"0.6587",
"0.457"
],
[
"0.755",
"0.4443",
"0.846",
"0.4443",
"0.846",
"0.459",
"0.755",
"0.459"
],
[
"0.2737",
"0.4458",
"0.3586",
"0.4458",
"0.3586",
"0.457",
"0.2737",
"0.457"
],
[
"0.2214",
"0.4458",
"0.259",
"0.4458",
"0.259",
"0.459",
"0.2214",
"0.459"
],
[
"0.3708",
"0.4458",
"0.455",
"0.4458",
"0.455",
"0.457",
"0.3708",
"0.457"
],
[
"0.2239",
"0.457",
"0.2573",
"0.457",
"0.2573",
"0.473",
"0.2239",
"0.473"
],
[
"0.2754",
"0.459",
"0.3586",
"0.459",
"0.3586",
"0.4702",
"0.2754",
"0.4702"
],
[
"0.3718",
"0.459",
"0.4558",
"0.459",
"0.4558",
"0.4702",
"0.3718",
"0.4702"
],
[
"0.4673",
"0.459",
"0.5513",
"0.459",
"0.5513",
"0.4702",
"0.4673",
"0.4702"
],
[
"0.562",
"0.459",
"0.649",
"0.459",
"0.649",
"0.4702",
"0.562",
"0.4702"
],
[
"0.66",
"0.459",
"0.746",
"0.459",
"0.746",
"0.4702",
"0.66",
"0.4702"
],
[
"0.7563",
"0.459",
"0.8423",
"0.459",
"0.8423",
"0.4702",
"0.7563",
"0.4702"
],
[
"0.2181",
"0.4702",
"0.2632",
"0.4702",
"0.2632",
"0.4849",
"0.2181",
"0.4849"
],
[
"0.2737",
"0.4717",
"0.3586",
"0.4717",
"0.3586",
"0.483",
"0.2737",
"0.483"
],
[
"0.3718",
"0.4717",
"0.4558",
"0.4717",
"0.4558",
"0.483",
"0.3718",
"0.483"
],
[
"0.4673",
"0.4717",
"0.5513",
"0.4717",
"0.5513",
"0.483",
"0.4673",
"0.483"
],
[
"0.562",
"0.4717",
"0.648",
"0.4717",
"0.648",
"0.483",
"0.562",
"0.483"
],
[
"0.66",
"0.4717",
"0.744",
"0.4717",
"0.744",
"0.483",
"0.66",
"0.483"
],
[
"0.7534",
"0.4685",
"0.8447",
"0.4702",
"0.8447",
"0.4849",
"0.7534",
"0.483"
],
[
"0.2141",
"0.483",
"0.2646",
"0.483",
"0.2646",
"0.4976",
"0.2141",
"0.4976"
],
[
"0.2737",
"0.4849",
"0.3586",
"0.4849",
"0.3586",
"0.4956",
"0.2737",
"0.4956"
],
[
"0.3718",
"0.4849",
"0.4558",
"0.4849",
"0.4558",
"0.4956",
"0.3718",
"0.4956"
],
[
"0.468",
"0.4849",
"0.5513",
"0.4849",
"0.5513",
"0.4956",
"0.468",
"0.4956"
],
[
"0.562",
"0.4849",
"0.648",
"0.4849",
"0.648",
"0.4968",
"0.562",
"0.4968"
],
[
"0.66",
"0.4849",
"0.746",
"0.4849",
"0.746",
"0.4956",
"0.66",
"0.4956"
],
[
"0.7563",
"0.4849",
"0.8433",
"0.4849",
"0.8433",
"0.4968",
"0.7563",
"0.4968"
],
[
"0.2239",
"0.5",
"0.259",
"0.5",
"0.259",
"0.515",
"0.2239",
"0.515"
],
[
"0.7534",
"0.5005",
"0.8447",
"0.5005",
"0.8447",
"0.515",
"0.7534",
"0.515"
],
[
"0.2737",
"0.502",
"0.3586",
"0.502",
"0.3586",
"0.513",
"0.2737",
"0.513"
],
[
"0.1528",
"0.502",
"0.1887",
"0.502",
"0.1887",
"0.5146",
"0.1528",
"0.5146"
],
[
"0.3694",
"0.502",
"0.455",
"0.502",
"0.455",
"0.513",
"0.3694",
"0.513"
],
[
"0.4656",
"0.502",
"0.55",
"0.502",
"0.55",
"0.513",
"0.4656",
"0.513"
],
[
"0.5605",
"0.502",
"0.649",
"0.502",
"0.649",
"0.513",
"0.5605",
"0.513"
],
[
"0.6587",
"0.502",
"0.744",
"0.502",
"0.744",
"0.513",
"0.6587",
"0.513"
],
[
"0.2214",
"0.5146",
"0.259",
"0.5146",
"0.259",
"0.528",
"0.2214",
"0.528"
],
[
"0.2737",
"0.515",
"0.3586",
"0.515",
"0.3586",
"0.5264",
"0.2737",
"0.5264"
],
[
"0.3708",
"0.515",
"0.4558",
"0.515",
"0.4558",
"0.5264",
"0.3708",
"0.5264"
],
[
"0.4673",
"0.515",
"0.5513",
"0.515",
"0.5513",
"0.5264",
"0.4673",
"0.5264"
],
[
"0.562",
"0.515",
"0.649",
"0.515",
"0.649",
"0.5264",
"0.562",
"0.5264"
],
[
"0.66",
"0.515",
"0.746",
"0.515",
"0.746",
"0.5264",
"0.66",
"0.5264"
],
[
"0.755",
"0.515",
"0.8433",
"0.515",
"0.8433",
"0.5264",
"0.755",
"0.5264"
],
[
"0.2255",
"0.5264",
"0.2573",
"0.5264",
"0.2573",
"0.539",
"0.2255",
"0.539"
],
[
"0.2754",
"0.528",
"0.3586",
"0.528",
"0.3586",
"0.539",
"0.2754",
"0.539"
],
[
"0.3694",
"0.528",
"0.455",
"0.528",
"0.455",
"0.539",
"0.3694",
"0.539"
],
[
"0.4673",
"0.528",
"0.5522",
"0.528",
"0.5522",
"0.539",
"0.4673",
"0.539"
],
[
"0.562",
"0.528",
"0.649",
"0.528",
"0.649",
"0.539",
"0.562",
"0.539"
],
[
"0.66",
"0.528",
"0.744",
"0.528",
"0.744",
"0.539",
"0.66",
"0.539"
],
[
"0.755",
"0.5264",
"0.846",
"0.5264",
"0.846",
"0.5415",
"0.755",
"0.5415"
],
[
"0.2181",
"0.5386",
"0.2632",
"0.5386",
"0.2632",
"0.553",
"0.2181",
"0.553"
],
[
"0.2737",
"0.5405",
"0.3586",
"0.5405",
"0.3586",
"0.552",
"0.2737",
"0.552"
],
[
"0.3708",
"0.5405",
"0.4558",
"0.5405",
"0.4558",
"0.552",
"0.3708",
"0.552"
],
[
"0.4673",
"0.5405",
"0.5513",
"0.5405",
"0.5513",
"0.552",
"0.4673",
"0.552"
],
[
"0.562",
"0.5405",
"0.649",
"0.5405",
"0.649",
"0.552",
"0.562",
"0.552"
],
[
"0.661",
"0.5405",
"0.744",
"0.5405",
"0.744",
"0.552",
"0.661",
"0.552"
],
[
"0.7534",
"0.539",
"0.8447",
"0.539",
"0.8447",
"0.5537",
"0.7534",
"0.5537"
],
[
"0.2157",
"0.553",
"0.2646",
"0.553",
"0.2646",
"0.5645",
"0.2157",
"0.5645"
],
[
"0.273",
"0.552",
"0.3586",
"0.552",
"0.3586",
"0.563",
"0.273",
"0.563"
],
[
"0.3694",
"0.553",
"0.4534",
"0.553",
"0.4534",
"0.563",
"0.3694",
"0.563"
],
[
"0.4673",
"0.5537",
"0.5522",
"0.5537",
"0.5522",
"0.5645",
"0.4673",
"0.5645"
],
[
"0.562",
"0.5537",
"0.648",
"0.5537",
"0.648",
"0.5645",
"0.562",
"0.5645"
],
[
"0.661",
"0.553",
"0.746",
"0.553",
"0.746",
"0.563",
"0.661",
"0.563"
],
[
"0.755",
"0.5537",
"0.8433",
"0.5537",
"0.8433",
"0.565",
"0.755",
"0.565"
],
[
"0.08826",
"0.603",
"0.1152",
"0.603",
"0.1152",
"0.6113",
"0.08826",
"0.6113"
],
[
"0.1168",
"0.5996",
"0.246",
"0.5996",
"0.246",
"0.613",
"0.1168",
"0.613"
],
[
"0.0866",
"0.617",
"0.9126",
"0.617",
"0.9126",
"0.631",
"0.0866",
"0.631"
],
[
"0.08417",
"0.63",
"0.9126",
"0.631",
"0.9126",
"0.646",
"0.08417",
"0.6445"
],
[
"0.08496",
"0.643",
"0.914",
"0.646",
"0.914",
"0.66",
"0.08496",
"0.6577"
],
[
"0.0866",
"0.661",
"0.1274",
"0.661",
"0.1274",
"0.6724",
"0.0866",
"0.6724"
],
[
"0.1111",
"0.677",
"0.826",
"0.677",
"0.826",
"0.69",
"0.1111",
"0.69"
],
[
"0.1111",
"0.6914",
"0.7876",
"0.6914",
"0.7876",
"0.705",
"0.1111",
"0.705"
],
[
"0.1111",
"0.705",
"0.5884",
"0.705",
"0.5884",
"0.7183",
"0.1111",
"0.7183"
],
[
"0.11273",
"0.7197",
"0.4656",
"0.7197",
"0.4656",
"0.73",
"0.11273",
"0.73"
],
[
"0.08826",
"0.7476",
"0.1274",
"0.7476",
"0.1274",
"0.759",
"0.08826",
"0.759"
],
[
"0.125",
"0.7456",
"0.3406",
"0.747",
"0.3406",
"0.761",
"0.125",
"0.7603"
],
[
"0.125",
"0.7505",
"0.1299",
"0.7505",
"0.1299",
"0.756",
"0.125",
"0.756"
],
[
"0.0866",
"0.768",
"0.3882",
"0.768",
"0.3882",
"0.782",
"0.0866",
"0.782"
],
[
"0.11273",
"0.786",
"0.9126",
"0.786",
"0.9126",
"0.8",
"0.11273",
"0.8"
],
[
"0.125",
"0.8003",
"0.4983",
"0.8003",
"0.4983",
"0.8145",
"0.125",
"0.8145"
],
[
"0.1111",
"0.8145",
"0.9126",
"0.8145",
"0.9126",
"0.8276",
"0.1111",
"0.8276"
],
[
"0.12335",
"0.8276",
"0.9126",
"0.829",
"0.9126",
"0.8423",
"0.12335",
"0.8413"
],
[
"0.1259",
"0.8433",
"0.179",
"0.8433",
"0.179",
"0.855",
"0.1259",
"0.855"
],
[
"0.11273",
"0.856",
"0.9126",
"0.856",
"0.9126",
"0.869",
"0.11273",
"0.869"
],
[
"0.125",
"0.869",
"0.914",
"0.869",
"0.914",
"0.8833",
"0.125",
"0.8833"
],
[
"0.1259",
"0.884",
"0.259",
"0.884",
"0.259",
"0.898",
"0.1259",
"0.898"
]
] | [
"Graph-adaptive Rectified Linear Unit for Graph Neural Networks",
"WWw'22, April 25-29, 2022, Virtual Event, Lyon, France.",
"Table 2: The average node classification accuracy (over 10 runs). The training set and test set are randomly divided. The results",
"in parentheses are the best results achieved among the 10 experiments. We omit the results of Cheb in BlogCatalog and Flickr",
"dataset due to the out-of-memory issue during the training step.",
"ReLU",
"LReLU",
"PReLU",
"Maxout",
"GReLU",
"ELU",
"Dataset",
"Model.",
"75.6 ± 3.0(80.6)",
"78.3 ± 1.7 (79.4)",
"75.4 ± 2.0(77.8)",
"Cora",
"Cheb",
"76.5 ± 2.2(79.4)",
"75.5 ± 2.2(80.0)",
"74.9 ± 2.9(80.4)",
"GCN",
"79.2 ± 1.4(81.6)",
"79.4 ± 1.6(80.6)",
"78.8 ± 1.3(80.2)",
"78.9 ± 1.0(80.2)",
"79.8 ± 1.5(81.5)",
"81.8 ± 1.8 (83.0)",
"SAGE",
"78.6 ± 1.7(81.3)",
"77.6 ± 1.8(79.9)",
"77.5 ± 1.4(79.9)",
"77.6 ± 1.7(78.7)",
"77.7 ± 1.7(79.9)",
"79.3 ± 1.5 (80.7)",
"GAT",
"81.2 ±1.3(83.3)",
"80.8 ± 2.3(83.1)",
"80.5 ± 1.7(83.0)",
"80.4 ± 2.1(83.1)",
"80.2 ±1.7(82.5)",
"81.5 ± 2.1 (84.1)",
"79.3 ± 2.0(81.9)",
"80.1 ± 1.5 (82.6)",
"79.7 ± 0.8(80.6)",
"ARMA",
"79.0 ± 1.4(80.8)",
"79.2 ± 2.1(82.9)",
"79.8 ±1.3(81.3)",
"APPNP",
"82.0 ±1.3(83.1)",
"81.0 ± 2.3(83.1)",
"81.5 ± 1.7(83.4)",
"81.4 ± 2.1(83.4)",
"81.0 ±1.7(82.5)",
"82.5 ± 2.1 (84.7)",
"PuMed",
"68.1 ± 4.1(74.9)",
"70.5 ± 3.0(76.7)",
"71.8 ± 3.9(75.8)",
"73.4 ± 2.9 (75.6)",
"68.4 ± 2.8(72.7)",
"Cheb",
"71.5 ± 3.1(76.3)",
"77.3 ±3.7(82.5)",
"78.9 ± 1.7 (81.3)",
"77.6 ±2.2(81.6)",
"GCN",
"76.8 ± 1.6(79.4)",
"76.8 ± 2.2(80.5)",
"77.3 ± 2.9(80.6)",
"SAGE",
"75.7 ± 3.1(79.8)",
"75.3 ± 3.3(79.8)",
"74.8 ± 2.7(78.6)",
"76.0 ±2.5(80.5)",
"74.5 ± 2.7(77.4)",
"76.2 ± 1.6 (78.4)",
"GAT",
"78.9 ± 2.3(82.1)",
"79.1 ± 1.8 (80.6)",
"78.7 ± 1.7(81.2)",
"77.2 ±3.1(81.3)",
"76.2 ± 3.0(80.7)",
"77.9 ±1.7(80.3)",
"77.3 ± 2.5(80.4)",
"76.5 ± 2.4(80.7)",
"76.6 ± 2.9(81.3)",
"76.9 ±2.6(80.7)",
"ARMA",
"76.5 ± 1.9(80.3)",
"77.4 ± 3.0 (80.2)",
"APPNP",
"79.0 ± 2.3(82.1)",
"79.8 ± 1.8 (81.2)",
"78.2 ± 1.7(81.7)",
"78.2 ±3.1(82.3)",
"79.2 ±3.0(80.7)",
"78.9 ±1.7(81.3)",
"CiteSeer",
"67.0 ± 2.3(70.5)",
"68.1 ± 1.3 (70.4)",
"67.8 ± 1.8(71.0)",
"Cheb",
"67.1 ±2.9(71.1)",
"67.8 ±2.4(71.1)",
"67.4±1.5(70.7)",
"GCN",
"67.7 ±2.3(72.1)",
"68.4 ± 1.8(71.2)",
"68.3 ± 1.4(70.1)",
"67.3 ± 2.1(70.7)",
"68.5 ± 2.2(72.5)",
"68.5 ± 1.9 (71.7)",
"SAGE",
"67.1 ± 2.8(70.1)",
"67.3 ± 2.1(70.1)",
"67.8 ± 1.7(70.2)",
"66.2 ± 2.6(69.6)",
"67.5 ± 1.8(71.5)",
"68.0 ± 1.3 (69.7)",
"GAT",
"68.6 ± 1.4(70.8)",
"69.2 ±1.9(71.7)",
"68.4 ± 1.6(71.2)",
"68.2 ± 1.6(69.7)",
"68.6 ± 1.6(71.3)",
"69.3 ± 1.7 (71.9)",
"ARMA",
"67.7 ± 1.3(68.9)",
"68.6 ± 2.4(71.5)",
"67.9 ± 2.1(71.3)",
"66.8 ± 1.5(69.4)",
"68.5 ± 1.8(70.9)",
"69.0 ± 2.2 (71.7)",
"APPNP",
"68.7 ± 1.3(70.5)",
"69.3 ± 1.6(71.2)",
"69.4 ± 1.4(71.2)",
"69.5 ± 1.6(70.7)",
"69.2 ± 1.6(72.0)",
"70.0 ±1.7 (72.3)",
"BlogCatalog",
"72.1 ± 1.9(75.5)",
"73.7 ± 1.2 (74.2)",
"72.6 ± 2.1(75.2)",
"GCN",
"72.6 ± 1.8(75.3)",
"71.4 ± 2.1(74.5)",
"72.4 ± 1.4(75.3)",
"72.1 ± 1.8(74.3)",
"72.0 ± 2.1(74.4)",
"71.6 ± 1.4(74.3)",
"73.3 ± 1.6 (75.3)",
"71.9 ±1.3(76.0)",
"SAGE",
"72.2 ± 1.9(74.9)",
"GAT",
"41.7 ± 7.3(54.4)",
"38.2 ± 3.3(41.8)",
"46.6 ± 3.6(51.7)",
"67.2 ± 2.6(71.8)",
"54.2 ± 3.9(59.3)",
"67.8 ± 3.9(72.3)",
"ARMA",
"72.5 ± 3.3(79.1)",
"72.5 ± 5.9(78.7)",
"77.2 ± 2.2(79.2)",
"79.6 ± 3.0(84.5)",
"84.4 ± 1.8(86.9)",
"85.7 ± 2.7 (88.4)",
"APPNP",
"71.1 ± 1.8(75.3)",
"72.5 ±2.0(75.2)",
"72.4 ± 1.8(75.3)",
"71.7 ± 2.1(74.8)",
"72.8 ±1.4(75.3)",
"73.8 ±1.5 (74.8)",
"GCN",
"54.4 ± 1.6 (56.8)",
"50.7 ±2.3(54.8)",
"Flickr",
"51.0 ± 2.0(53.8)",
"52.8 ± 1.8(56.0)",
"53.0 ± 1.6(54.9)",
"54.0 ± 1.8(56.8)",
"SAGE",
"49.7 ±2.2(53.8)",
"50.8 ± 2.1(54.0)",
"52.6 ± 1.8(56.7)",
"53.2 ± 1.3(56.1)",
"54.0 ± 1.7(56.5)",
"55.3 ±1.4 (57.2)",
"GAT",
"20.2 ± 3.1(25.2)",
"20.3 ± 2.5(23.8)",
"23.8 ± 2.9(28.1)",
"32.8 ± 4.9(43.2)",
"30.0 ± 2.6(34.6)",
"33.7 ± 3.1 (36.3)",
"ARMA",
"48.4 ± 4.7(56.1)",
"52.0 ± 3.9(59.0)",
"52.2 ± 4.1(58.1)",
"45.9 ± 3.5(53.1)",
"52.7 ± 4.0(59.4)",
"56.5 ± 2.4 (59.1)",
"APPNP",
"51.6 ±2.0(54.0)",
"51.6±2.1(53.8)",
"53.0 ± 1.8(56.0)",
"53.2 ± 1.4(54.9)",
"53.9 ±1.9(56.6)",
"54.8 ± 1.9 (56.7)",
"4",
"EXPERIMENTS",
"下面，我们在两个任务上评估所提出的方法GReLU，即E4.1节中的节点分类和E??节中的图形分类。我们通过将其与具有多个",
"GNN骨架的不同ReLU变体进行比较来显示GReLU的有效性。我们在？?节的消融研究中展示了GReLU不同模块的贡献。在??",
"节中讨论了选择超参数K的影响。为了验证GReLU的自适应行为，在?？节中进行了参数分析。我们的目标是回答以下研究问题",
"(RQ):",
".RQ1: Does applying GReLU in the updating step help improve the capacity of GNNs for different downstream tasks?",
". RQ2: What kind of factors does GReLU benefit from in its design and how do they contribute to performance?",
"·RQ3:Howdoesthechoiceof parameterKaffect theperformanceofGReLU?",
".RO4:DoesGReLUdynamicallyadapttodifferentnodes?",
"4.1",
"NodeClassification (RQ1)",
"",
"数据集。GReLU在五个实际数据集上进行评估：",
": Cora, PubMed, CiteSeer [16] are well-known citation network datasets, where nodes represent papers and edges represent their",
"citations, and the nodes are labeled based on the paper topics.",
".Flick [25] is an image and video hosting website, where users interact with each other via photo sharing. It is a social network where",
"nodes represent users and edges represent their relationships,and all nodes are divided into 9classes according to the interest groups",
"of users.",
" BlogCatalog [25] is a social network for bloggers and their social relationships from the BlogCatalog website. Node attributes are",
"constructed by the keywords of user profiles, and the labels represent the topic categories provided by the authors, and all nodes are",
"divided into 6 classes."
] | {
"arxiv_id": "2202.06281",
"language": "zh",
"page_idx": 4
} |
||
2312.03596-zh-0 | 2312.03596-zh-0.jpg | MMM: 生成 M 询问 M 动作 M 模型
Chen Chen
University of Central Florida
[email protected]
Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee
University of North Carolina at Charlotte $\{$ epinyoan, pwang13, Minwoo.Lee $\}$ @uncc.edu
<FigureHere>
Figure 1. MMM can generate precise human motions given fine-grained textual descriptions while enabling motion editing applications. Blue frames represent conditioned motion input and red frames are the generated motion. Motion in-betweening can be performed by filling the gaps between keyframes or major motion points, conditioned on text or without conditions. Upper body part editing is done by fixing the lower body motion generated by one text prompt and altering the upper body motion according to another input prompt. An arbitrary long motion sequence can be generated according to a story (i.e, a sequence of text prompts), where MMM generates the motion for each prompt (red frames), while “hallucinating” the nature and smooth motion transitions (blue frames) between neighboring prompts (without being explicitly trained on motion transition datasets)
# $\operatorname{Abstract}$
while guaranteeing smooth transitions between editing and non-editing parts. Extensive experiments on the HumanML3D and KIT-ML datasets demonstrate that MMM surpasses current leading methods in generating high-quality motion (evidenced by superior FID scores of 0.08 and 0.429), while offering advanced editing features such as body-part modification, motion in-betweening, and the synthesis of long motion sequences. In addition, MMM is two orders of magni- ${\mathrm{tude~faster~on~a~single~mid-range~GPU~than~editable}}$ motion diffusion models. Our project page is available at https://exitudio.github.io/MMM-page/ .
Recent advances in text-to-motion generation using diffusion and autoregressive models have shown promising results. However, these models often suffer from a trade-off between real-time performance, high fidelity, and motion editability. To address this gap, we introduce MMM, a novel yet simple motion generation paradigm based on M asked M otion M odel. MMM consists of two key components: (1) a motion tokenizer that transforms 3D human motion into a sequence of discrete tokens in latent space, and (2) a conditional masked motion transformer that learns to predict randomly masked motion tokens, conditioned on the precomputed text tokens. By attending to motion and text tokens in all directions, MMM explicitly captures inherent dependency among motion tokens and semantic mapping between motion and text tokens. During inference, this allows parallel and iterative decoding of multiple motion tokens that are highly consistent with fine-grained text descriptions, therefore simultaneously achieving high-fidelity and high-speed motion generation. In addition, MMM has innate motion editability. By simply placing mask tokens in the place that needs editing, MMM automatically fills the gaps
## 1. Introduction
由于自然语言描述的语义丰富性和用户友好性，文本驱动的人体运动生成最近已成为新兴的研究热点，其在动画，电影，VR/AR 和机器人技术中的广泛应用。然而，由于语言和运动数据分布之间的固有差异，生成与文本描述符精确对齐的高保真运动具有挑战性。为了解决这一挑战，已经提出了三种主要方法，包括(1) 语言运动潜在空间对齐，(2)条件扩散模型和(3)条件自回归模型
在第一种方法中，将文本描述和运动序列投影到单 | [
[
"0.2957",
"0.1338",
"0.6763",
"0.1338",
"0.6763",
"0.1483",
"0.2957",
"0.1483"
],
[
"0.1462",
"0.1812",
"0.569",
"0.1818",
"0.569",
"0.1995",
"0.1462",
"0.1989"
],
[
"0.6724",
"0.1831",
"0.767",
"0.1831",
"0.767",
"0.1976",
"0.6724",
"0.1976"
],
[
"0.2091",
"0.202",
"0.5073",
"0.202",
"0.5073",
"0.2153",
"0.2091",
"0.2153"
],
[
"0.614",
"0.202",
"0.824",
"0.202",
"0.824",
"0.2153",
"0.614",
"0.2153"
],
[
"0.2009",
"0.2197",
"0.5156",
"0.2197",
"0.5156",
"0.2329",
"0.2009",
"0.2329"
],
[
"0.636",
"0.2197",
"0.794",
"0.2197",
"0.794",
"0.2329",
"0.636",
"0.2329"
],
[
"0.11273",
"0.2715",
"0.2214",
"0.2715",
"0.2214",
"0.2822",
"0.11273",
"0.2822"
],
[
"0.6724",
"0.2715",
"0.6914",
"0.2715",
"0.6914",
"0.279",
"0.6724",
"0.279"
],
[
"0.709",
"0.2734",
"0.722",
"0.2734",
"0.722",
"0.2766",
"0.709",
"0.2766"
],
[
"0.1013",
"0.2827",
"0.2303",
"0.2827",
"0.2303",
"0.2935",
"0.1013",
"0.2935"
],
[
"0.2942",
"0.2766",
"0.384",
"0.2766",
"0.384",
"0.2886",
"0.2942",
"0.2886"
],
[
"0.671",
"0.2827",
"0.697",
"0.2827",
"0.697",
"0.2903",
"0.671",
"0.2903"
],
[
"0.701",
"0.286",
"0.7485",
"0.286",
"0.7485",
"0.2903",
"0.701",
"0.2903"
],
[
"0.492",
"0.2886",
"0.613",
"0.2886",
"0.613",
"0.2954",
"0.492",
"0.2954"
],
[
"0.09235",
"0.2935",
"0.2443",
"0.2935",
"0.2443",
"0.3037",
"0.09235",
"0.3037"
],
[
"0.2974",
"0.2886",
"0.3816",
"0.2886",
"0.3816",
"0.2998",
"0.2974",
"0.2998"
],
[
"0.5063",
"0.2974",
"0.596",
"0.2974",
"0.596",
"0.308",
"0.5063",
"0.308"
],
[
"0.689",
"0.2935",
"0.706",
"0.2935",
"0.706",
"0.2966",
"0.689",
"0.2966"
],
[
"0.689",
"0.307",
"0.701",
"0.307",
"0.701",
"0.31",
"0.689",
"0.31"
],
[
"0.683",
"0.3164",
"0.7017",
"0.3164",
"0.7017",
"0.324",
"0.683",
"0.324"
],
[
"0.5146",
"0.332",
"0.5767",
"0.332",
"0.5767",
"0.339",
"0.5146",
"0.339"
],
[
"0.5293",
"0.3416",
"0.572",
"0.3416",
"0.572",
"0.3484",
"0.5293",
"0.3484"
],
[
"0.2402",
"0.3914",
"0.4272",
"0.3926",
"0.4272",
"0.4028",
"0.2402",
"0.4016"
],
[
"0.281",
"0.4016",
"0.4084",
"0.4016",
"0.4084",
"0.4153",
"0.281",
"0.4153"
],
[
"0.482",
"0.4016",
"0.6167",
"0.4028",
"0.6167",
"0.414",
"0.482",
"0.4136"
],
[
"0.6914",
"0.4028",
"0.8555",
"0.4028",
"0.8555",
"0.4136",
"0.6914",
"0.4136"
],
[
"0.0801",
"0.4224",
"0.888",
"0.4224",
"0.888",
"0.4363",
"0.0801",
"0.4363"
],
[
"0.1209",
"0.4028",
"0.2141",
"0.4028",
"0.2141",
"0.4136",
"0.1209",
"0.4136"
],
[
"0.0801",
"0.435",
"0.892",
"0.435",
"0.892",
"0.449",
"0.0801",
"0.449"
],
[
"0.0784",
"0.4463",
"0.8906",
"0.4463",
"0.8906",
"0.4602",
"0.0784",
"0.4602"
],
[
"0.0801",
"0.459",
"0.892",
"0.459",
"0.892",
"0.473",
"0.0801",
"0.473"
],
[
"0.0784",
"0.4717",
"0.892",
"0.4697",
"0.892",
"0.483",
"0.0784",
"0.4849"
],
[
"0.0809",
"0.4841",
"0.8906",
"0.4841",
"0.8906",
"0.4944",
"0.0809",
"0.4944"
],
[
"0.0801",
"0.4944",
"0.652",
"0.4944",
"0.652",
"0.5083",
"0.0801",
"0.5083"
],
[
"0.2378",
"0.521",
"0.3137",
"0.521",
"0.3137",
"0.535",
"0.2378",
"0.535"
],
[
"0.5034",
"0.523",
"0.892",
"0.523",
"0.892",
"0.536",
"0.5034",
"0.536"
],
[
"0.5024",
"0.537",
"0.892",
"0.537",
"0.892",
"0.552",
"0.5024",
"0.552"
],
[
"0.0972",
"0.553",
"0.4697",
"0.5537",
"0.4697",
"0.5684",
"0.0972",
"0.5674"
],
[
"0.5034",
"0.553",
"0.892",
"0.553",
"0.892",
"0.5664",
"0.5034",
"0.5664"
],
[
"0.0801",
"0.5693",
"0.4697",
"0.5693",
"0.4697",
"0.584",
"0.0801",
"0.584"
],
[
"0.501",
"0.5674",
"0.888",
"0.5684",
"0.888",
"0.582",
"0.501",
"0.5806"
],
[
"0.0809",
"0.5854",
"0.4697",
"0.5854",
"0.4697",
"0.5986",
"0.0809",
"0.5986"
],
[
"0.5024",
"0.5835",
"0.8906",
"0.5835",
"0.8906",
"0.5967",
"0.5024",
"0.5967"
],
[
"0.0809",
"0.6016",
"0.4673",
"0.6016",
"0.4673",
"0.6123",
"0.0809",
"0.6123"
],
[
"0.5024",
"0.5977",
"0.8906",
"0.5977",
"0.8906",
"0.6113",
"0.5024",
"0.6113"
],
[
"0.0784",
"0.613",
"0.4697",
"0.6157",
"0.4697",
"0.63",
"0.0784",
"0.6274"
],
[
"0.5024",
"0.613",
"0.892",
"0.613",
"0.892",
"0.627",
"0.5024",
"0.627"
],
[
"0.0784",
"0.629",
"0.4697",
"0.63",
"0.4697",
"0.6445",
"0.0784",
"0.643"
],
[
"0.5024",
"0.6274",
"0.892",
"0.6274",
"0.892",
"0.6416",
"0.5024",
"0.6416"
],
[
"0.0784",
"0.6445",
"0.4697",
"0.643",
"0.4697",
"0.6577",
"0.0784",
"0.659"
],
[
"0.5024",
"0.643",
"0.8906",
"0.643",
"0.8906",
"0.657",
"0.5024",
"0.657"
],
[
"0.0801",
"0.66",
"0.4697",
"0.66",
"0.4697",
"0.674",
"0.0801",
"0.674"
],
[
"0.501",
"0.6577",
"0.892",
"0.657",
"0.892",
"0.672",
"0.501",
"0.6724"
],
[
"0.0784",
"0.674",
"0.4714",
"0.675",
"0.4714",
"0.6895",
"0.0784",
"0.688"
],
[
"0.5024",
"0.674",
"0.892",
"0.674",
"0.892",
"0.6875",
"0.5024",
"0.6875"
],
[
"0.0801",
"0.69",
"0.468",
"0.69",
"0.468",
"0.704",
"0.0801",
"0.704"
],
[
"0.501",
"0.688",
"0.819",
"0.6895",
"0.819",
"0.704",
"0.501",
"0.7026"
],
[
"0.0801",
"0.7056",
"0.468",
"0.7056",
"0.468",
"0.7197",
"0.0801",
"0.7197"
],
[
"0.0801",
"0.72",
"0.468",
"0.72",
"0.468",
"0.7344",
"0.0801",
"0.7344"
],
[
"0.0784",
"0.7363",
"0.4697",
"0.735",
"0.4697",
"0.7485",
"0.0784",
"0.75"
],
[
"0.5024",
"0.7363",
"0.6396",
"0.738",
"0.6396",
"0.7524",
"0.5024",
"0.7505"
],
[
"0.0801",
"0.7505",
"0.4697",
"0.7505",
"0.4697",
"0.7646",
"0.0801",
"0.7646"
],
[
"0.0801",
"0.7666",
"0.468",
"0.7666",
"0.468",
"0.78",
"0.0801",
"0.78"
],
[
"0.5034",
"0.7646",
"0.8906",
"0.7646",
"0.8906",
"0.778",
"0.5034",
"0.778"
],
[
"0.0801",
"0.781",
"0.4697",
"0.781",
"0.4697",
"0.7954",
"0.0801",
"0.7954"
],
[
"0.5034",
"0.78",
"0.892",
"0.78",
"0.892",
"0.7935",
"0.5034",
"0.7935"
],
[
"0.0801",
"0.797",
"0.468",
"0.797",
"0.468",
"0.81",
"0.0801",
"0.81"
],
[
"0.5034",
"0.7944",
"0.892",
"0.7944",
"0.892",
"0.808",
"0.5034",
"0.808"
],
[
"0.0801",
"0.811",
"0.4697",
"0.811",
"0.4697",
"0.8247",
"0.0801",
"0.8247"
],
[
"0.5034",
"0.81",
"0.892",
"0.81",
"0.892",
"0.8237",
"0.5034",
"0.8237"
],
[
"0.0801",
"0.827",
"0.468",
"0.827",
"0.468",
"0.8403",
"0.0801",
"0.8403"
],
[
"0.5034",
"0.8247",
"0.8906",
"0.8247",
"0.8906",
"0.8384",
"0.5034",
"0.8384"
],
[
"0.0801",
"0.8413",
"0.4697",
"0.8413",
"0.4697",
"0.855",
"0.0801",
"0.855"
],
[
"0.5034",
"0.8403",
"0.8906",
"0.8403",
"0.8906",
"0.854",
"0.5034",
"0.854"
],
[
"0.0784",
"0.857",
"0.468",
"0.856",
"0.468",
"0.869",
"0.0784",
"0.8706"
],
[
"0.5024",
"0.855",
"0.892",
"0.854",
"0.892",
"0.8687",
"0.5024",
"0.869"
],
[
"0.0801",
"0.872",
"0.4697",
"0.872",
"0.4697",
"0.8853",
"0.0801",
"0.8853"
],
[
"0.5034",
"0.8706",
"0.6006",
"0.8706",
"0.6006",
"0.882",
"0.5034",
"0.882"
],
[
"0.0801",
"0.886",
"0.4697",
"0.886",
"0.4697",
"0.901",
"0.0801",
"0.901"
],
[
"0.523",
"0.886",
"0.8906",
"0.886",
"0.8906",
"0.901",
"0.523",
"0.901"
]
] | [
"MMM：生成M询问M动作M模型",
"Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee",
"Chen Chen",
"University of North Carolina at Charlotte",
"University of Central Florida",
"{ epinyoan, pwangl3, Minwoo.Lee } @uncc.edu",
"[email protected]",
"the figure walks in a",
"theto",
"aa",
"counterclockwise circle",
"aman puts both",
"aperso",
"iumps.forwar",
"apersonwalksforward",
"stopping at its starting.point.",
"hands in the air",
"in a straight line.",
"anei",
"ans",
"Trans",
"personcray",
"forward",
"circle and sits back down on the ground",
"(b) Upper Body Editing",
"(c) Motion In-betweening",
"(d) Long Sequence Generation",
"Figure 1. MMM can generate precise human motions given fine-grained textual descriptions while enabling motion editing applications",
"(a) Text to Motion",
"Blue frames represent conditioned motion input and red frames are the generated motion. Motion in-betweening can be performed",
"by filling the gaps between keyframes or major motion points, conditioned on text or without conditions. Upper body part editing is",
"done by fixing the lower body motion generated by one text prompt and altering the upper body motion according to another input",
"prompt.An arbitrary long motion sequence can be generated according to a story (i.e.,a sequence of text prompts),where MMM",
"generates the motion for each prompt (red frames),while\"hallucinating\"the nature and smooth motion transitions (blue frames)",
"between neighboring prompts (without being explicitly trained on motion transition datasets)",
"Abstract",
"while guaranteeing smooth transitions between edit-",
"ing and non-editing parts. Extensive experiments on",
"Recent advances in text-to-motion generation us-",
"the HumanML3D and KIT-ML datasets demonstrate",
"ing diffusion and autoregressive models have shown",
"that MMM surpasses current leading methods in gener",
"promising results. However, these models often suffer",
"ating high-quality motion (evidenced by superior FID",
"fromatrade-offbetweenreal-timeperformance,high",
"scores of 0.08 and 0.429), while offering advanced edit-",
"fidelity, and motion editability. To address this gap, we",
"ing features such as body-part modification, motion",
"introduce MMM, a novel yet simple motion generation",
"in-betweening, and the synthesis of long motion se-",
"paradigm based on M asked M otion M odel. MMM",
"quences. In addition, MMM is two orders of magni-",
"consists of two key components: (1) a motion tokenizer",
"tude faster on a single mid-range GPU than editable",
"that transforms 3D human motion into a sequence of",
"motion diffusion models. Our project page is available",
"discrete tokens in latent space, and (2) a conditional",
"at https://exitudio.github.io/MMM-page/.",
"masked motion transformer that learns to predict ran-",
"domly masked motion tokens, conditioned on the pre-",
"computed text tokens. By attending to motion and",
"1.Introduction",
"text tokens in all directions, MMM explicitly captures",
"inherent dependency among motion tokens and seman-",
"由于自然语言描述的语义丰富性和用户友好性，文本",
"tic mapping between motion and text tokens. During",
"驱动的人体运动生成最近已成为新兴的研究热点，其",
"inference, this allows parallel and iterative decoding",
"在动画，电影，VR/AR和机器人技术中的广泛应用。",
"of multiple motion tokens that are highly consistent",
"然而，由于语言和运动数据分布之间的固有差异，生成",
"with fine-grained text descriptions, therefore simulta-",
"与文本描述符精确对齐的高保真运动具有挑战性。为",
"neously achieving high-fidelity and high-speed motion",
"了解决这一挑战，已经提出了三种主要方法，包括（1）",
"generation. In addition, MMM has innate motion ed-",
"语言运动潜在空间对齐，（2）条件扩散模型和（3）条",
"itability. By simply placing mask tokens in the place",
"件自回归模型",
"that needs editing, MMM automatically flls the gaps",
"在第一种方法中，将文本描述和运动序列投影到单"
] | {
"arxiv_id": "2312.03596",
"language": "zh",
"page_idx": 0
} |
||
2312.03596-zh-3 | 2312.03596-zh-3.jpg | 我们从相同的 CLIP 模型 [26] 获得单词嵌人，以通过交叉注意力=softmax$\left(Q_{motion}K_{\mathrm{word}}^{T}/\sqrt{D}\right)V_{\mathrm{word}}[47]$ 捕获每个单词与运动之间的局部关系
[MASK] ,[PAD] 和 是可学习的专用令牌。在训练过程中，[MASK] 令牌用于表示输人损坏，并且模型学习预测代替[MASK] 的实际令牌。在推理过程中， [MASK] 令牌不仅可用作运动令牌生成的占位符，而且通过将 [MASK] 令牌放置在需要编辑的位置，在各种任务中也很有用，如第 4 节所述。令牌用于填充较短的运动序列，允许计算具有多个不同长度序列的批次。最后，令牌附加到最后一个令牌之后的输入令牌，为模型提供运动端点的显式信号。培训策略与损失。运动令牌序列$Y$表示为$Y=[e_i]_i=1^L$,其中$L$ 表示序列长度。我们随机屏蔽$r\times L$令牌，并用可学习的[MASK]令牌替换它们，其中$r$是在$\alpha$和 1之间截断的均匀分布之后的屏蔽率。然后，使用[MASK]令牌更新原始运动令牌序列$Y$ ,以形成损坏的运动序列$Y_{\overline{M}}$ 。此损坏的序列以及文本嵌人$W$ 被馈送到文本条件掩码转换器中，以重建输入运动令牌序列，其重建概率或置信度等于$p\left(y_i\mid Y_{\overline{\boldsymbol{M}}},W\right)$ 。目标是最小化以文本为条件的预测屏蔽令牌的负对数似然：
$$\mathcal{L}_{\mathrm{mask}}=-\underset{\boldsymbol{Y}\in\mathcal{D}}{\mathbb{E}}\left[\sum_{\forall i\in[1,L]}\log p\left(y_i\mid Y_{\overline{M}},W\right)\right].$$
(2)
通过并行解码进行推理。为了在推理过程中解码运动令牌，我们通过输入所有[MASK] 令牌来启动该过程，代表一个空画布，然后逐步预测每次迭代的更多令牌。接下来，执行迭代并行解码，其中在每次迭代中， 变换器屏蔽模型最不确信的运动令牌子集，然后在下一次迭代中并行预测这些被屏蔽的令牌。掩码令牌$n_M$ 的数量由掩码调度函数 i.e. 确定，迭代$t$的衰减函数。选择衰减函数是因为在早期迭代时，预测中存在很高的不确定性，因此模型以较大的掩蔽率开始，并且仅保留少量具有高预测置信度的令牌。随着生成过程的进行，由于来自先前迭代的上下文信息的增加，掩蔽率降低。我们实验了线性$n_M=L((T-t)/T)$ [8] 和余弦函数$n_M= L\cos ( \frac 12\pi t/ T)$ [3] ,其中余弦函数产生更好的性能。要获得最终的$n_M$,生成的运动序列$L$的长度也应该可用。为了解决这个问题，我们采用了一个预先训练的预测器，根据输入文本 [12] 来估计运动序列长度
# 4. Motion Editing
如图$\operatorname{E}$4 所示，我们的蒙版运动建模的优势在于其固有
的编辑运动的能力
中间运动。由于其掩码双向解码特性，无论过去和将来的上下文如何，我们都可以轻松地将[MASK]令牌放置在需要编辑的任何地方。作为一项重要的编辑任务，中间的运动涉及插值或填充关键帧或主要运动点之间的间隙，以创建平滑，连续的 3D 动画。由于模型已经在训练期间学习了所有可能的随机时间掩蔽组
<FigureHere>
Figure 4. Motion Editing. (Left) Motion in-betweening. (Mid-
dle) Long Sequence Generation. ( Right) Upper Body Editing. “M”refers to mask token.“T”is text conditioned tokens and “L” denotes lower body part conditioned tokens.
合，因此可以在不进行任何额外训练的情况下完成中
间的运动
长序列生成。由于可用的 HumanML3D [12] 和 KIT [24]数据集中的运动数据长度有限，其中没有样本超过10秒的持续时间，因此生成任意长的运动构成了挑战。为了解决这个问题，我们使用训练的掩蔽运动模型作为长运动序列合成的先验，而不需要额外的训练。特别是，给定一个由多个文本提示组成的故事，我们的模型首先为每个提示生成运动令牌序列。然后，我们生成过渡运动令牌，条件是前一个运动序列的结束和下一个运动序列的开始。诸如 PriorMDM [30] 中的 DoubleTakes 之类的扩散方法需要多个步骤(最多 1,000 个步骤)来生成过渡，并在生成过渡运动之前平均上一个运动和下一个运动之间的空间差异。我们的方法只需要一个步骤来生成现实和自然的过渡运动
上半身编辑。为了实现身体部位编辑，我们分别对上半身和下半身分词器进行预训练，每个分词器都有自己的编码器和解码器。每个运动令牌的嵌人大小是常规全身嵌入大小的一半。在第二阶段，上部和下部令牌被连接回形成全身嵌人。因此，嵌人尺寸和条件屏蔽变压器保持不变。理想情况下，我们可以通过预测蒙面的上半身标记来训练变压器，条件是上半身运动的文本以及下半身标记。但是，生成的运动与文本不一致。为了解决这个问题，我们通过轻度破坏将随机[MASK] 令牌引人到下半身运动序列中，以便转换器可以更好地学习整个身体运动的空间和时间依赖性。因此，训练损失可以写成：
$$\mathcal{L}_{\mathrm{up}}=-\sum_{\boldsymbol{Y}\in\mathcal{D}}\left[\sum_{\forall i\in[1,L]}\log p\left(y_i^{up}\mid Y_{\overline{M}}^{up},Y_{\overline{M}}^{down},W\right)\right]$$
(3.
其中$Y_{\frac{up}M}$表示带有掩码的上部令牌$,Y_{\overline{M}}^{down}$是带有掩码的下部令牌。重要的是要注意，下部的[MASK]令牌在所有迭代中保持不变。
# 5. Experiments
在本节中，我们将进行比较，以评估模型的质量和时间效率。在 5.1 节中，按照 [12] 对多个数据集的标准评估，我们观察到我们的方法始终优于当前最先进的方法此外，在$\mathbb{E}\color{red}{5.2}$ 节中，使用 MLD [4] 的时间效率度量进行评估，我们的方法在平均和相对于运动长度方面都显示出更短的推理时间。数据集。我们在两个用于文本到 | [
[
"0.0809",
"0.09216",
"0.4673",
"0.09216",
"0.4673",
"0.1061",
"0.0809",
"0.1061"
],
[
"0.812",
"0.09216",
"0.8857",
"0.09216",
"0.8857",
"0.09973",
"0.812",
"0.09973"
],
[
"0.5312",
"0.1067",
"0.5703",
"0.1067",
"0.5703",
"0.1187",
"0.5312",
"0.1187"
],
[
"0.8022",
"0.108",
"0.8857",
"0.108",
"0.8857",
"0.1187",
"0.8022",
"0.1187"
],
[
"0.7915",
"0.10986",
"0.8037",
"0.10986",
"0.8037",
"0.11426",
"0.7915",
"0.11426"
],
[
"0.0809",
"0.10986",
"0.468",
"0.10986",
"0.468",
"0.1276",
"0.0809",
"0.1276"
],
[
"0.638",
"0.1111",
"0.6733",
"0.1111",
"0.6733",
"0.11554",
"0.638",
"0.11554"
],
[
"0.699",
"0.1111",
"0.722",
"0.1111",
"0.722",
"0.11426",
"0.699",
"0.11426"
],
[
"0.8037",
"0.1187",
"0.8857",
"0.1187",
"0.8857",
"0.1288",
"0.8037",
"0.1288"
],
[
"0.5195",
"0.11237",
"0.5317",
"0.11237",
"0.5317",
"0.11554",
"0.5195",
"0.11554"
],
[
"0.517",
"0.13",
"0.5645",
"0.13",
"0.5645",
"0.1401",
"0.517",
"0.1401"
],
[
"0.6733",
"0.13",
"0.697",
"0.13",
"0.697",
"0.137",
"0.6733",
"0.137"
],
[
"0.839",
"0.1338",
"0.853",
"0.1338",
"0.853",
"0.137",
"0.839",
"0.137"
],
[
"0.669",
"0.1383",
"0.697",
"0.1364",
"0.6978",
"0.1445",
"0.671",
"0.1465"
],
[
"0.1013",
"0.1465",
"0.468",
"0.1465",
"0.468",
"0.1604",
"0.1013",
"0.1604"
],
[
"0.671",
"0.1477",
"0.6978",
"0.1477",
"0.6978",
"0.159",
"0.671",
"0.159"
],
[
"0.8022",
"0.1559",
"0.8857",
"0.1559",
"0.8857",
"0.1666",
"0.8022",
"0.1666"
],
[
"0.0809",
"0.161",
"0.468",
"0.161",
"0.468",
"0.1749",
"0.0809",
"0.1749"
],
[
"0.0809",
"0.1307",
"0.3423",
"0.1307",
"0.3423",
"0.1445",
"0.0809",
"0.1445"
],
[
"0.5947",
"0.1674",
"0.63",
"0.1674",
"0.63",
"0.1749",
"0.5947",
"0.1749"
],
[
"0.6807",
"0.1635",
"0.7256",
"0.1654",
"0.7236",
"0.1768",
"0.6787",
"0.1749"
],
[
"0.738",
"0.1692",
"0.7715",
"0.1692",
"0.7715",
"0.1736",
"0.738",
"0.1736"
],
[
"0.0809",
"0.1768",
"0.468",
"0.1768",
"0.468",
"0.1901",
"0.0809",
"0.1901"
],
[
"0.558",
"0.1768",
"0.63",
"0.1768",
"0.63",
"0.1875",
"0.558",
"0.1875"
],
[
"0.626",
"0.1768",
"0.763",
"0.1768",
"0.763",
"0.1875",
"0.626",
"0.1875"
],
[
"0.7695",
"0.1799",
"0.7793",
"0.1799",
"0.7793",
"0.1843",
"0.7695",
"0.1843"
],
[
"0.08417",
"0.1932",
"0.4673",
"0.1932",
"0.4673",
"0.2039",
"0.08417",
"0.2039"
],
[
"0.5034",
"0.1913",
"0.8896",
"0.1913",
"0.8896",
"0.2052",
"0.5034",
"0.2052"
],
[
"0.0809",
"0.207",
"0.468",
"0.207",
"0.468",
"0.2203",
"0.0809",
"0.2203"
],
[
"0.5034",
"0.2026",
"0.888",
"0.2039",
"0.888",
"0.2172",
"0.5034",
"0.2166"
],
[
"0.0825",
"0.2229",
"0.468",
"0.2229",
"0.468",
"0.2361",
"0.0825",
"0.2361"
],
[
"0.5063",
"0.2172",
"0.8896",
"0.2172",
"0.8896",
"0.2279",
"0.5063",
"0.2279"
],
[
"0.505",
"0.2279",
"0.7974",
"0.2291",
"0.7974",
"0.2405",
"0.505",
"0.2393"
],
[
"0.0825",
"0.2374",
"0.468",
"0.2374",
"0.468",
"0.2507",
"0.0825",
"0.2507"
],
[
"0.0809",
"0.252",
"0.4673",
"0.252",
"0.4673",
"0.2651",
"0.0809",
"0.2651"
],
[
"0.505",
"0.2556",
"0.8896",
"0.2556",
"0.8896",
"0.2695",
"0.505",
"0.2695"
],
[
"0.0809",
"0.2664",
"0.4656",
"0.2664",
"0.4656",
"0.2798",
"0.0809",
"0.2798"
],
[
"0.5063",
"0.2715",
"0.568",
"0.2715",
"0.568",
"0.2827",
"0.5063",
"0.2827"
],
[
"0.08417",
"0.2822",
"0.4673",
"0.2822",
"0.4673",
"0.2954",
"0.08417",
"0.2954"
],
[
"0.523",
"0.286",
"0.8906",
"0.2842",
"0.8906",
"0.2974",
"0.523",
"0.2998"
],
[
"0.0825",
"0.2974",
"0.4673",
"0.2974",
"0.4673",
"0.3113",
"0.0825",
"0.3113"
],
[
"0.5063",
"0.303",
"0.8896",
"0.303",
"0.8896",
"0.3132",
"0.5063",
"0.3132"
],
[
"0.0809",
"0.3125",
"0.468",
"0.3125",
"0.468",
"0.3257",
"0.0809",
"0.3257"
],
[
"0.5034",
"0.3164",
"0.8906",
"0.3164",
"0.8906",
"0.33",
"0.5034",
"0.33"
],
[
"0.0809",
"0.3276",
"0.4673",
"0.3276",
"0.4673",
"0.3416",
"0.0809",
"0.3416"
],
[
"0.5063",
"0.3333",
"0.8896",
"0.3333",
"0.8896",
"0.3435",
"0.5063",
"0.3435"
],
[
"0.0809",
"0.3423",
"0.468",
"0.3423",
"0.468",
"0.356",
"0.0809",
"0.356"
],
[
"0.505",
"0.3467",
"0.8896",
"0.3467",
"0.8896",
"0.3599",
"0.505",
"0.3599"
],
[
"0.0809",
"0.358",
"0.4673",
"0.356",
"0.4673",
"0.3718",
"0.0809",
"0.3738"
],
[
"0.5063",
"0.363",
"0.888",
"0.363",
"0.888",
"0.3738",
"0.5063",
"0.3738"
],
[
"0.0825",
"0.3726",
"0.4697",
"0.3726",
"0.4697",
"0.3865",
"0.0825",
"0.3865"
],
[
"0.505",
"0.377",
"0.8906",
"0.377",
"0.8906",
"0.3901",
"0.505",
"0.3901"
],
[
"0.0809",
"0.3882",
"0.468",
"0.3882",
"0.468",
"0.4028",
"0.0809",
"0.4028"
],
[
"0.5063",
"0.3933",
"0.8896",
"0.3933",
"0.8896",
"0.404",
"0.5063",
"0.404"
],
[
"0.0825",
"0.404",
"0.3635",
"0.404",
"0.3635",
"0.4172",
"0.0825",
"0.4172"
],
[
"0.5034",
"0.406",
"0.8906",
"0.4048",
"0.8906",
"0.4192",
"0.5034",
"0.4204"
],
[
"0.505",
"0.4216",
"0.8906",
"0.4216",
"0.8906",
"0.435",
"0.505",
"0.435"
],
[
"0.1274",
"0.4363",
"0.2418",
"0.443",
"0.2385",
"0.4634",
"0.125",
"0.4558"
],
[
"0.2427",
"0.4382",
"0.411",
"0.4382",
"0.411",
"0.4602",
"0.2427",
"0.4602"
],
[
"0.505",
"0.4363",
"0.8906",
"0.4363",
"0.8906",
"0.4495",
"0.505",
"0.4495"
],
[
"0.406",
"0.4463",
"0.411",
"0.4463",
"0.411",
"0.452",
"0.406",
"0.452"
],
[
"0.4453",
"0.443",
"0.4673",
"0.443",
"0.4673",
"0.4558",
"0.4453",
"0.4558"
],
[
"0.5063",
"0.452",
"0.8906",
"0.452",
"0.8906",
"0.4653",
"0.5063",
"0.4653"
],
[
"0.2345",
"0.457",
"0.2876",
"0.457",
"0.2876",
"0.4685",
"0.2345",
"0.4685"
],
[
"0.505",
"0.4673",
"0.7485",
"0.4673",
"0.7485",
"0.481",
"0.505",
"0.481"
],
[
"0.10046",
"0.483",
"0.468",
"0.4841",
"0.468",
"0.4976",
"0.10046",
"0.4968"
],
[
"0.528",
"0.4841",
"0.8906",
"0.4841",
"0.8906",
"0.4944",
"0.528",
"0.4944"
],
[
"0.0809",
"0.4976",
"0.468",
"0.4976",
"0.468",
"0.511",
"0.0809",
"0.511"
],
[
"0.5063",
"0.4976",
"0.8896",
"0.4976",
"0.8896",
"0.511",
"0.5063",
"0.511"
],
[
"0.505",
"0.512",
"0.8896",
"0.512",
"0.8896",
"0.526",
"0.505",
"0.526"
],
[
"0.0809",
"0.513",
"0.468",
"0.513",
"0.468",
"0.5264",
"0.0809",
"0.5264"
],
[
"0.2427",
"0.449",
"0.25",
"0.449",
"0.25",
"0.452",
"0.2427",
"0.452"
],
[
"0.0825",
"0.53",
"0.4673",
"0.53",
"0.4673",
"0.5405",
"0.0825",
"0.5405"
],
[
"0.5034",
"0.528",
"0.8896",
"0.528",
"0.8896",
"0.5415",
"0.5034",
"0.5415"
],
[
"0.0809",
"0.5435",
"0.4673",
"0.5435",
"0.4673",
"0.5566",
"0.0809",
"0.5566"
],
[
"0.505",
"0.5425",
"0.8906",
"0.5425",
"0.8906",
"0.556",
"0.505",
"0.556"
],
[
"0.08905",
"0.5566",
"0.4673",
"0.5596",
"0.4673",
"0.5737",
"0.08905",
"0.5713"
],
[
"0.505",
"0.558",
"0.8896",
"0.558",
"0.8896",
"0.5713",
"0.505",
"0.5713"
],
[
"0.0801",
"0.5728",
"0.468",
"0.5737",
"0.468",
"0.587",
"0.0801",
"0.586"
],
[
"0.5034",
"0.5728",
"0.8906",
"0.5728",
"0.8906",
"0.586",
"0.5034",
"0.586"
],
[
"0.0825",
"0.589",
"0.468",
"0.589",
"0.468",
"0.603",
"0.0825",
"0.603"
],
[
"0.5034",
"0.587",
"0.888",
"0.587",
"0.888",
"0.601",
"0.5034",
"0.601"
],
[
"0.0809",
"0.604",
"0.468",
"0.604",
"0.468",
"0.617",
"0.0809",
"0.617"
],
[
"0.505",
"0.6016",
"0.8896",
"0.6016",
"0.8896",
"0.6157",
"0.505",
"0.6157"
],
[
"0.0825",
"0.619",
"0.468",
"0.619",
"0.468",
"0.6333",
"0.0825",
"0.6333"
],
[
"0.5034",
"0.616",
"0.8896",
"0.617",
"0.8896",
"0.632",
"0.5034",
"0.631"
],
[
"0.0825",
"0.6353",
"0.4673",
"0.6353",
"0.4673",
"0.6484",
"0.0825",
"0.6484"
],
[
"0.5034",
"0.6333",
"0.8906",
"0.6333",
"0.8906",
"0.6465",
"0.5034",
"0.6465"
],
[
"0.0809",
"0.651",
"0.468",
"0.6484",
"0.468",
"0.662",
"0.0809",
"0.664"
],
[
"0.5063",
"0.6494",
"0.607",
"0.6494",
"0.607",
"0.66",
"0.5063",
"0.66"
],
[
"0.0809",
"0.664",
"0.4673",
"0.664",
"0.4673",
"0.678",
"0.0809",
"0.678"
],
[
"0.0809",
"0.6787",
"0.4673",
"0.6787",
"0.4673",
"0.6924",
"0.0809",
"0.6924"
],
[
"0.6274",
"0.683",
"0.8164",
"0.683",
"0.8164",
"0.704",
"0.6274",
"0.704"
],
[
"0.7876",
"0.6846",
"0.8447",
"0.686",
"0.8447",
"0.6978",
"0.7876",
"0.696"
],
[
"0.871",
"0.6875",
"0.8896",
"0.6875",
"0.8896",
"0.6997",
"0.871",
"0.6997"
],
[
"0.5728",
"0.688",
"0.6006",
"0.688",
"0.6006",
"0.6963",
"0.5728",
"0.6963"
],
[
"0.5195",
"0.688",
"0.5767",
"0.688",
"0.5767",
"0.6997",
"0.5195",
"0.6997"
],
[
"0.0825",
"0.6943",
"0.468",
"0.6943",
"0.468",
"0.7085",
"0.0825",
"0.7085"
],
[
"0.5703",
"0.6963",
"0.6045",
"0.6963",
"0.6045",
"0.704",
"0.5703",
"0.704"
],
[
"0.613",
"0.702",
"0.6626",
"0.702",
"0.6626",
"0.7134",
"0.613",
"0.7134"
],
[
"0.0825",
"0.7104",
"0.468",
"0.7104",
"0.468",
"0.7236",
"0.0825",
"0.7236"
],
[
"0.0801",
"0.7246",
"0.0989",
"0.7246",
"0.0989",
"0.738",
"0.0801",
"0.738"
],
[
"0.501",
"0.7227",
"0.8936",
"0.72",
"0.8936",
"0.741",
"0.501",
"0.7437"
],
[
"0.505",
"0.742",
"0.8896",
"0.742",
"0.8896",
"0.756",
"0.505",
"0.756"
],
[
"0.0784",
"0.754",
"0.2402",
"0.756",
"0.2402",
"0.771",
"0.0784",
"0.768"
],
[
"0.505",
"0.7583",
"0.6895",
"0.7583",
"0.6895",
"0.7715",
"0.505",
"0.7715"
],
[
"0.0809",
"0.781",
"0.4673",
"0.781",
"0.4673",
"0.7944",
"0.0809",
"0.7944"
],
[
"0.5024",
"0.787",
"0.6396",
"0.787",
"0.6396",
"0.802",
"0.5024",
"0.802"
],
[
"0.0809",
"0.797",
"0.2115",
"0.797",
"0.2115",
"0.81",
"0.0809",
"0.81"
],
[
"0.1013",
"0.811",
"0.468",
"0.811",
"0.468",
"0.8247",
"0.1013",
"0.8247"
],
[
"0.505",
"0.811",
"0.8896",
"0.811",
"0.8896",
"0.8247",
"0.505",
"0.8247"
],
[
"0.0809",
"0.827",
"0.468",
"0.827",
"0.468",
"0.8403",
"0.0809",
"0.8403"
],
[
"0.5024",
"0.8247",
"0.8906",
"0.8257",
"0.8906",
"0.8403",
"0.5024",
"0.839"
],
[
"0.0825",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0825",
"0.855"
],
[
"0.5034",
"0.8413",
"0.8906",
"0.8413",
"0.8906",
"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"
],
[
"0.0825",
"0.872",
"0.468",
"0.872",
"0.468",
"0.8853",
"0.0825",
"0.8853"
],
[
"0.505",
"0.8735",
"0.8896",
"0.8735",
"0.8896",
"0.884",
"0.505",
"0.884"
],
[
"0.0809",
"0.887",
"0.468",
"0.887",
"0.468",
"0.901",
"0.0809",
"0.901"
],
[
"0.5034",
"0.887",
"0.8896",
"0.887",
"0.8896",
"0.901",
"0.5034",
"0.901"
]
] | [
"我们从相同的CLIP模型[26]获得单词嵌人，以通过",
"UpperBodyEditin",
"MM",
"MMMMMM",
"T",
"交叉注意力=softmax(QmotionKord/VD)Vword[47]",
"LongSee",
"Cn",
"MLMLML",
"",
"",
"MM",
"",
"MM",
"[MASK]、[PAD] 和[END] 是可学习的专用令牌。在",
"MM",
"MLMLML",
"训练过程中，[MASK】令牌用于表示输人损坏，并且模",
"捕获每个单词与运动之间的局部关系",
"Transitior",
"Transltion",
"Transitio",
"型学习预测代替[MASK]的实际令牌。在推理过程中",
"",
"",
"",
"MASKI令牌不仅可用作运动令牌生成的占位符，而且",
"Figure 4. Motion Editing. (Left) Motion in-betweening. (Mid-",
"通过将[MASK]令牌放置在需要编辑的位置，在各种",
"dle) Long Sequence Generation.(Right) Upper Body Editing",
"任务中也很有用，如第4节所述。[PAD]令牌用于填",
"“\"M\" refers to mask token.\"T\"is text conditioned tokens and",
"\"L\"denotes lowerbody part conditioned tokens.",
"充较短的运动序列，允许计算具有多个不同长度序列",
"的批次。最后，[END]令牌附加到最后一个令牌之后的",
"合，因此可以在不进行任何额外训练的情况下完成中",
"输人令牌，为模型提供运动端点的显式信号。培训策略",
"间的运动",
"与损失。运动令牌序列Y表示为Y=[eil1，其中L",
"长序列生成。由于可用的HumanML3D[12]和KIT",
"表示序列长度。我们随机屏蔽r×L令牌，并用可学习",
"24数据集中的运动数据长度有限，其中没有样本超过",
"的[MASK]令牌替换它们，其中是在α和1之间截",
"10秒的持续时间，因此生成任意长的运动构成了挑战。",
"断的均匀分布之后的屏蔽率。然后，使用[MASK]令",
"为了解决这个问题，我们使用训练的掩蔽运动模型作",
"牌更新原始运动令牌序列Y，以形成损坏的运动序列",
"为长运动序列合成的先验，而不需要额外的训练。特别",
"YM。此损坏的序列以及文本嵌人W被馈送到文本条",
"是，给定一个由多个文本提示组成的故事，我们的模型",
"件掩码转换器中，以重建输人运动令牌序列，其重建概",
"首先为每个提示生成运动令牌序列。然后，我们生成过",
"率或置信度等于p(yi|YM,W）。目标是最小化以文",
"渡运动令牌，条件是前一个运动序列的结束和下一个运",
"本为条件的预测屏蔽令牌的负对数似然：",
"动序列的开始。诸如PriorMDM[30】中的DoubleTakes",
"之类的扩散方法需要多个步骤（最多1,000个步骤）来",
"Cmask =-YE,",
"Z logp(ui | Ym,W)",
"生成过渡，并在生成过渡运动之前平均上一个运动和",
"",
"(2)",
"下一个运动之间的空间差异。我们的方法只需要一个",
"Vie[1,L]",
"步骤来生成现实和自然的过渡运动",
"通过并行解码进行推理。为了在推理过程中解码运",
"上半身编辑。为了实现身体部位编辑，我们分别对",
"动令牌，我们通过输人所有[MASK]令牌来启动该过",
"上半身和下半身分词器进行预训练，每个分词器都有",
"自己的编码器和解码器。每个运动令牌的嵌人大小是",
"程，代表一个空画布，然后逐步预测每次迭代的更多令",
"",
"牌。接下来：执行迭代并行解码，其中在每次迭代中",
"常规全身嵌人大小的一半。在第二阶段，上部和下部令",
"变换器屏蔽模型最不确信的运动令牌子集，然后在下",
"牌被连接回形成全身嵌入。因此，嵌入尺寸和条件屏蔽",
"一次选代中并行预测这些被屏蔽的令牌。掩码令牌nM",
"变压器保持不变。理想情况下，我们可以通过预测蒙面",
"的数量由掩码调度函数i.e.确定，迭代t的衰减函数。",
"的上半身标记来训练变压器，条件是上半身运动的文",
"选择衰减函数是因为在早期送代时，预测中存在很高",
"本以及下半身标记。但是，生成的运动与文本不一致",
"的不确定性，因此模型以较大的掩蔽率开始，并且仅保",
"为了解决这个问题，我们通过轻度破坏将随机[MASK",
"留少量具有高预测置信度的令牌。随着生成过程的进",
"令牌引人到下半身运动序列中，以便转换器可以更好",
"行，由于来自先前迭代的上下文信息的增加，掩蔽率降",
"地学习整个身体运动的空间和时间依赖性。因此，训练",
"低。我们实验了线性nM=L（(T一t)/T）[8|和余弦函",
"损失可以写成",
"数nM=Lcos(πt/T)[3]，其中余弦函数产生更好的",
"性能。要获得最终的nM，生成的运动序列L的长度",
"Z logp (urP Yu,Yrown",
"down.W)",
"(3)",
"E",
"Cup=-",
"也应该可用。为了解决这个问题，我们采用了一个预先",
"YED",
"Vie[1,L]",
"训练的预测器，根据输人文本【12]来估计运动序列长",
"度",
"其中Y表示带有掩码的上部令牌，Yeoun是带有掩",
"码的下部令牌。重要的是要注意，下部的[MASK]令",
"4.MotionEditing",
"牌在所有送代中保持不变。",
"如图E4所示，我们的蒙版运动建模的优势在于其固有",
"5. Experiments",
"的编辑运动的能力",
"中间运动。由于其掩码双向解码特性，无论过去和",
"在本节中，我们将进行比较，以评估模型的质量和时间",
"将来的上下文如何，我们都可以轻松地将[MASK]令",
"效率。在5.1节中，按照[12]对多个数据集的标准评",
"牌放置在需要编辑的任何地方。作为一项重要的编辑",
"估，我们观察到我们的方法始终优于当前最先进的方法",
"任务，中间的运动涉及插值或填充关键顿或主要运动",
"此外，在E5.2节中，使用MLD[4]的时间效率度量进",
"点之间的间隙，以创建平滑，连续的3D动画。由于模",
"行评估，我们的方法在平均和相对于运动长度方面都显",
"型已经在训练期间学习了所有可能的随机时间掩蔽组",
"示出更短的推理时间。数据集。我们在两个用于文本到"
] | {
"arxiv_id": "2312.03596",
"language": "zh",
"page_idx": 3
} |
||
2211.11468-zh-1 | 2211.11468-zh-1.jpg | 通过实体屏蔽语言建模和多任务学习增强与危机相关的 Tweet 分类
Philipp Seeberger and Korbinian Riedhammer Technische Hochschule Nürnberg Georg Simon Ohm
$\{$ philipp.seeberger,korbinian.riedhammer $\}$ @th-nuernberg.de
<FigureHere>
Figure 1: Example tweets of several disasters over time, annotated with entitites. The short posts are mostly biased towards specific events.
## Abstract
Social media has become an important information source for crisis management and provides quick access to ongoing developments and critical information. However, classification models suffer from event-related biases and highly imbalanced label distributions which still poses a challenging task. To address these challenges, we propose a combination of entity-masked language modeling and hierarchical multi-label classification as a multi-task learning problem. We evaluate our method on tweets from the TREC-IS dataset and show an absolute performance gain w.r.t. F1-score of up to 10% for actionable information types. Moreover, we found that entitymasking reduces the effect of overfitting to indomain events and enables improvements in cross-event generalization. Our source code is publicly available on GitHub. $^{1}$
## 1 $\textbf{Introduction}$
灾难事件期间社交媒体上的消息已成为危机管理中的重要信息源 (Reuter et al.,2018)。与传统来源(e。g。,官方新闻),社交媒体帖子立即提供有关事态发展、第一方观察和持续紧急情况下的受影响人员的详细信息(Sakaki et al., 2010)。访问这些信息对于发展态势感知和支持救济提供者，政府机构和其他官方机构(Kruspe et al.,2021)至关重要
一个关键挑战是对大量社交媒体流进行信息细化，这需要自动方法来可靠地检测相关内容(Kaufhold,2021)。最近的工作集中在二进制、 多类和多标签文本分类技术上，将帖子分类为粗略(e。g。相关、不相关)或细粒度(e。g。, 基础设施损坏，失踪人员)由扁平或分层结构组成的类别 (Alam et al., 2018b, 2021; Buntain et al.,2021)
自然语言处理(NLP)的另一个挑战是社交媒体和微博平台中普遍存在的数据的性质。例如，与危机相关的领域中的大多数作品都集
$^{1}$https: / / gi thub. com/ th- nuernberg/
crisis-tapt-hmc
中在 Twitter 数据 (Kruspe et al.,2021)上，该数据继承了诸如短文本(每条推文限制 280 个字符),较少的上下文信息，主题标签和噪声(例如。g。,拼写错误，表情符号)(Wiegmann et al., 2020; Zahera et al., 2021)。根据 Sarmiento and Poblete (2021),不同类型的灾害(e。g。、洪水、野火)只能通过几个基于文本的特征来识别。但是，Figure 1 中显示的与事件相关的偏差和实体会阻止模型推广到看不见的灾难事件，因此会降低。r。t。检测性能
为了避免这个问题，已经提出了诸如对抗性训练 (Medina Maza et al.,2020) , 域自适应(Alam et al., 2018a) 和分层标签嵌入 (Miyazaki et al.,2019)之类的方法，但是它们遭受混合事件类型，假定未标记的数据或需要语义标签描述。与这项工作相反，我们的目标是通过掩盖实体，应用自适应预训练并结合标签的层次结构来增强对看不见事件的稀有可操作信息的检测
$\textbf{Contributions}$ 我们的主要贡献如下：(1)我们引入了一种基于实体掩码的自适应预训练策略。(2)我们将标签的层次结构合并为多任务学习(MTL)问题。(3)我们的经验表明，我们的方法提高了对新事件的泛化，并提高了可操作信息类型的检测性能 | [
[
"0.1511",
"0.0956",
"0.848",
"0.0956",
"0.848",
"0.10986",
"0.1511",
"0.10986"
],
[
"0.297",
"0.1426",
"0.708",
"0.1426",
"0.708",
"0.1567",
"0.297",
"0.1567"
],
[
"0.2854",
"0.1592",
"0.7173",
"0.1603",
"0.7173",
"0.1746",
"0.2854",
"0.1733"
],
[
"0.2091",
"0.177",
"0.804",
"0.177",
"0.804",
"0.1907",
"0.2091",
"0.1907"
],
[
"0.2612",
"0.2446",
"0.3418",
"0.2465",
"0.3418",
"0.2625",
"0.2612",
"0.26"
],
[
"0.5483",
"0.2488",
"0.6885",
"0.2488",
"0.6885",
"0.2625",
"0.5483",
"0.2625"
],
[
"0.681",
"0.2488",
"0.848",
"0.2488",
"0.848",
"0.2625",
"0.681",
"0.2625"
],
[
"0.532",
"0.2666",
"0.843",
"0.2673",
"0.843",
"0.279",
"0.532",
"0.2778"
],
[
"0.1453",
"0.2737",
"0.4592",
"0.2737",
"0.4592",
"0.2874",
"0.1453",
"0.2874"
],
[
"0.5415",
"0.2778",
"0.5728",
"0.2808",
"0.5693",
"0.2915",
"0.54",
"0.2893"
],
[
"0.586",
"0.2798",
"0.7656",
"0.2798",
"0.7656",
"0.2903",
"0.586",
"0.2903"
],
[
"0.763",
"0.2798",
"0.8193",
"0.2798",
"0.8193",
"0.2903",
"0.763",
"0.2903"
],
[
"0.1453",
"0.288",
"0.4592",
"0.288",
"0.4592",
"0.3018",
"0.1453",
"0.3018"
],
[
"0.534",
"0.3005",
"0.7676",
"0.3005",
"0.7676",
"0.3113",
"0.534",
"0.3113"
],
[
"0.796",
"0.3005",
"0.8613",
"0.3005",
"0.8613",
"0.3113",
"0.796",
"0.3113"
],
[
"0.532",
"0.3123",
"0.592",
"0.3123",
"0.592",
"0.3225",
"0.532",
"0.3225"
],
[
"0.147",
"0.304",
"0.4568",
"0.304",
"0.4568",
"0.3142",
"0.147",
"0.3142"
],
[
"0.663",
"0.3135",
"0.7095",
"0.3135",
"0.7095",
"0.3206",
"0.663",
"0.3206"
],
[
"0.8154",
"0.3123",
"0.843",
"0.3123",
"0.843",
"0.3206",
"0.8154",
"0.3206"
],
[
"0.6113",
"0.3154",
"0.6484",
"0.3154",
"0.6484",
"0.3184",
"0.6113",
"0.3184"
],
[
"0.1453",
"0.3164",
"0.4592",
"0.3164",
"0.4592",
"0.33",
"0.1453",
"0.33"
],
[
"0.1453",
"0.33",
"0.4592",
"0.33",
"0.4592",
"0.3433",
"0.1453",
"0.3433"
],
[
"0.532",
"0.3403",
"0.5996",
"0.3403",
"0.5996",
"0.3508",
"0.532",
"0.3508"
],
[
"0.608",
"0.335",
"0.837",
"0.335",
"0.837",
"0.3457",
"0.608",
"0.3457"
],
[
"0.147",
"0.3445",
"0.4575",
"0.3445",
"0.4575",
"0.3582",
"0.147",
"0.3582"
],
[
"0.608",
"0.3445",
"0.8423",
"0.3467",
"0.8423",
"0.357",
"0.608",
"0.3552"
],
[
"0.552",
"0.3508",
"0.581",
"0.3508",
"0.581",
"0.3623",
"0.552",
"0.3623"
],
[
"0.1453",
"0.3594",
"0.4592",
"0.3594",
"0.4592",
"0.3723",
"0.1453",
"0.3723"
],
[
"0.607",
"0.354",
"0.6514",
"0.3557",
"0.65",
"0.3677",
"0.6055",
"0.3652"
],
[
"0.1436",
"0.3723",
"0.4592",
"0.3735",
"0.4592",
"0.3872",
"0.1436",
"0.386"
],
[
"0.1453",
"0.3884",
"0.4575",
"0.3884",
"0.4575",
"0.402",
"0.1453",
"0.402"
],
[
"0.513",
"0.383",
"0.882",
"0.383",
"0.882",
"0.3967",
"0.513",
"0.3967"
],
[
"0.147",
"0.402",
"0.4592",
"0.402",
"0.4592",
"0.415",
"0.147",
"0.415"
],
[
"0.511",
"0.3967",
"0.882",
"0.3967",
"0.882",
"0.4104",
"0.511",
"0.4104"
],
[
"0.1486",
"0.4175",
"0.4568",
"0.4175",
"0.4568",
"0.4275",
"0.1486",
"0.4275"
],
[
"0.513",
"0.4133",
"0.769",
"0.4133",
"0.769",
"0.4233",
"0.513",
"0.4233"
],
[
"0.147",
"0.4312",
"0.4568",
"0.4312",
"0.4568",
"0.4412",
"0.147",
"0.4412"
],
[
"0.147",
"0.4453",
"0.4568",
"0.4453",
"0.4568",
"0.456",
"0.147",
"0.456"
],
[
"0.147",
"0.4602",
"0.4568",
"0.4602",
"0.4568",
"0.4702",
"0.147",
"0.4702"
],
[
"0.1453",
"0.4727",
"0.4592",
"0.4727",
"0.4592",
"0.4863",
"0.1453",
"0.4863"
],
[
"0.5146",
"0.4697",
"0.881",
"0.4697",
"0.881",
"0.4827",
"0.5146",
"0.4827"
],
[
"0.1486",
"0.488",
"0.4575",
"0.488",
"0.4575",
"0.4988",
"0.1486",
"0.4988"
],
[
"0.513",
"0.485",
"0.879",
"0.485",
"0.879",
"0.4988",
"0.513",
"0.4988"
],
[
"0.147",
"0.503",
"0.455",
"0.503",
"0.455",
"0.513",
"0.147",
"0.513"
],
[
"0.513",
"0.502",
"0.879",
"0.502",
"0.879",
"0.5156",
"0.513",
"0.5156"
],
[
"0.1453",
"0.5156",
"0.4568",
"0.5156",
"0.4568",
"0.5293",
"0.1453",
"0.5293"
],
[
"0.513",
"0.5186",
"0.8774",
"0.5186",
"0.8774",
"0.533",
"0.513",
"0.533"
],
[
"0.1436",
"0.5293",
"0.357",
"0.528",
"0.357",
"0.542",
"0.1436",
"0.5435"
],
[
"0.513",
"0.5337",
"0.881",
"0.5337",
"0.881",
"0.5474",
"0.513",
"0.5474"
],
[
"0.511",
"0.55",
"0.879",
"0.55",
"0.879",
"0.5635",
"0.511",
"0.5635"
],
[
"0.1184",
"0.558",
"0.1368",
"0.558",
"0.1368",
"0.568",
"0.1184",
"0.568"
],
[
"0.1351",
"0.56",
"0.1427",
"0.56",
"0.1427",
"0.5654",
"0.1351",
"0.5654"
],
[
"0.1528",
"0.5557",
"0.2612",
"0.5557",
"0.2612",
"0.568",
"0.1528",
"0.568"
],
[
"0.5146",
"0.5664",
"0.881",
"0.5664",
"0.881",
"0.58",
"0.5146",
"0.58"
],
[
"0.1184",
"0.5806",
"0.4878",
"0.5806",
"0.4878",
"0.594",
"0.1184",
"0.594"
],
[
"0.511",
"0.5806",
"0.879",
"0.5806",
"0.879",
"0.5957",
"0.511",
"0.5957"
],
[
"0.1184",
"0.5967",
"0.4878",
"0.5967",
"0.4878",
"0.611",
"0.1184",
"0.611"
],
[
"0.511",
"0.5986",
"0.881",
"0.5986",
"0.881",
"0.6123",
"0.511",
"0.6123"
],
[
"0.1184",
"0.6133",
"0.486",
"0.6133",
"0.486",
"0.6274",
"0.1184",
"0.6274"
],
[
"0.513",
"0.6143",
"0.7827",
"0.6143",
"0.7827",
"0.629",
"0.513",
"0.629"
],
[
"0.1184",
"0.6274",
"0.4844",
"0.6274",
"0.4844",
"0.641",
"0.1184",
"0.641"
],
[
"0.11755",
"0.6445",
"0.4844",
"0.6445",
"0.4844",
"0.6577",
"0.11755",
"0.6577"
],
[
"0.532",
"0.6396",
"0.881",
"0.6396",
"0.881",
"0.653",
"0.532",
"0.653"
],
[
"0.11755",
"0.66",
"0.4836",
"0.66",
"0.4836",
"0.6733",
"0.11755",
"0.6733"
],
[
"0.513",
"0.6562",
"0.881",
"0.6562",
"0.881",
"0.6694",
"0.513",
"0.6694"
],
[
"0.51",
"0.6704",
"0.882",
"0.6714",
"0.882",
"0.686",
"0.51",
"0.685"
],
[
"0.1184",
"0.6777",
"0.4844",
"0.6777",
"0.4844",
"0.6914",
"0.1184",
"0.6914"
],
[
"0.513",
"0.6885",
"0.881",
"0.6885",
"0.881",
"0.7017",
"0.513",
"0.7017"
],
[
"0.11755",
"0.695",
"0.346",
"0.695",
"0.346",
"0.708",
"0.11755",
"0.708"
],
[
"0.5146",
"0.7036",
"0.881",
"0.7036",
"0.881",
"0.7173",
"0.5146",
"0.7173"
],
[
"0.1486",
"0.7104",
"0.4836",
"0.7104",
"0.4836",
"0.7236",
"0.1486",
"0.7236"
],
[
"0.511",
"0.72",
"0.881",
"0.72",
"0.881",
"0.734",
"0.511",
"0.734"
],
[
"0.1184",
"0.727",
"0.486",
"0.727",
"0.486",
"0.7407",
"0.1184",
"0.7407"
],
[
"0.511",
"0.7354",
"0.8774",
"0.7354",
"0.8774",
"0.749",
"0.511",
"0.749"
],
[
"0.1184",
"0.742",
"0.4878",
"0.742",
"0.4878",
"0.756",
"0.1184",
"0.756"
],
[
"0.511",
"0.7505",
"0.881",
"0.752",
"0.881",
"0.7666",
"0.511",
"0.7656"
],
[
"0.12006",
"0.7583",
"0.486",
"0.7583",
"0.486",
"0.7715",
"0.12006",
"0.7715"
],
[
"0.11755",
"0.774",
"0.4844",
"0.775",
"0.4844",
"0.789",
"0.11755",
"0.788"
],
[
"0.513",
"0.7686",
"0.534",
"0.7686",
"0.534",
"0.782",
"0.513",
"0.782"
],
[
"0.11755",
"0.789",
"0.4836",
"0.789",
"0.4836",
"0.8027",
"0.11755",
"0.8027"
],
[
"0.11584",
"0.8057",
"0.4844",
"0.8057",
"0.4844",
"0.8193",
"0.11584",
"0.8193"
],
[
"0.11755",
"0.8237",
"0.2083",
"0.8237",
"0.2083",
"0.835",
"0.11755",
"0.835"
],
[
"0.517",
"0.8247",
"0.881",
"0.8247",
"0.881",
"0.8384",
"0.517",
"0.8384"
],
[
"0.1377",
"0.839",
"0.486",
"0.839",
"0.486",
"0.8525",
"0.1377",
"0.8525"
],
[
"0.5156",
"0.843",
"0.879",
"0.843",
"0.879",
"0.8525",
"0.5156",
"0.8525"
],
[
"0.1184",
"0.855",
"0.486",
"0.855",
"0.486",
"0.868",
"0.1184",
"0.868"
],
[
"0.513",
"0.857",
"0.879",
"0.857",
"0.879",
"0.8706",
"0.513",
"0.8706"
],
[
"0.1184",
"0.8716",
"0.4878",
"0.8716",
"0.4878",
"0.8853",
"0.1184",
"0.8853"
],
[
"0.513",
"0.8726",
"0.882",
"0.8726",
"0.882",
"0.8857",
"0.513",
"0.8857"
],
[
"0.1377",
"0.8926",
"0.388",
"0.8936",
"0.388",
"0.9077",
"0.1377",
"0.9067"
],
[
"0.513",
"0.8896",
"0.879",
"0.8896",
"0.879",
"0.903",
"0.513",
"0.903"
],
[
"0.1184",
"0.9067",
"0.2334",
"0.9067",
"0.2334",
"0.9175",
"0.1184",
"0.9175"
],
[
"0.513",
"0.905",
"0.7188",
"0.905",
"0.7188",
"0.9185",
"0.513",
"0.9185"
]
] | [
"通过实体屏蔽语言建模和多任务学习增强与危机相关的Tweet分类",
"Philipp Seeberger and Korbinian Riedhammer",
"Technische Hochschule Nuirnberg Georg Simon Ohm",
"( philipp.seeberger,korbinian.riedhammer } @th-nuernberg.de",
"Abstract",
"Entities: Location",
"Hashtag Number Url",
"WildfireMANDATORYEVACUATIONINORDER!Please",
"Social media has become an important infor-",
"2016",
"head to Mac Island or Noralta Lodge",
"Wymmfire",
"mation source for crisis management and pro-",
"Philippines rocked by 6.3-magnitude earthquake",
"Earthquake",
"sparkingma",
"vides quick access to ongoing developments",
".https://..",
"2019",
"",
"and critical information. However, classifi-",
"cation models suffer from event-related bi-",
"Dam Failure",
"Rapidly rising water overtook dams and forced",
"ases and highly imbalanced label distributions",
"evacuation of10,000peopleincentralMichigan",
"2020",
"which still poses a challenging task. To ad-",
"https:/l..",
"dress these challenges, we propose a com-",
"bination of entity-masked language modeling",
"Figure l: Example tweets of several disasters over",
"and hierarchical multi-label classification as a",
"time, annotated with entitites. The short posts are",
"multi-tasklearning problem.Weevaluate our",
"mostly biased towards specificevents.",
"method ontweets fromthe TREC-IS dataset",
"and show an absolute performance gain w.r.t.",
"Fl-score of up to l0 %for actionableinfor",
"mation types. Moreover, we found that entity-",
"中在Twitter数据（Kruspeetal.,2021）上，该数",
"masking reduces the effect of overfitting to in-",
"据继承了诸如短文本（每条推文限制280个字",
"domain events and enables improvements in",
"符），较少的上下文信息，主题标签和噪声（例",
"cross-event generalization. Our source code is",
"如。g。，拼写错误，表情符号）（Wiegmannetal.",
"publicly available on GitHub.1",
"2020;Zahera et al.,2021)。根据 Sarmiento and",
"Poblete（2021），不同类型的灾害（e。g。、洪",
"1 ",
"",
"Introduction",
"水、野火）只能通过几个基于文本的特征来识",
"灾难事件期间社交媒体上的消息已成为危机",
"别。但是，Figure1中显示的与事件相关的偏",
"管理中的重要信息源（Reuteretal.，2018）。与",
"差和实体会阻止模型推广到看不见的灾难事",
"传统来源(e。g。，官方新闻)，社交媒体帖子",
"件，因此会降低。I。t。检测性能",
"立即提供有关事态发展、第一方观察和持续",
"紧急情况下的受影响人员的详细信息（Sakaki",
"为了避免这个问题，已经提出了诸如对抗",
"etal.，2010）。访问这些信息对于发展态势感知",
"性训练（MedinaMazaetal，2020），域自适应",
"（Alametal.，2018a）和分层标签嵌人（Miyazaki",
"和支持救济提供者，政府机构和其他官方机构",
"etal.,2019)之类的方法，但是它们遭受混合事",
"(Kruspeetal.,2021）至关重要",
"件类型，假定未标记的数据或需要语义标签描",
"个关键挑战是对大量社交媒体流进行信息",
"述。与这项工作相反，我们的目标是通过掩盖",
"细化，这需要自动方法来可靠地检测相关内容",
"实体，应用自适应预训练并结合标签的层次结",
"(Kaufhold,2021）。最近的工作集中在二进制、",
"构来增强对看不见事件的稀有可操作信息的检",
"多类和多标签文本分类技术上，将帖子分类为",
"粗略(e。g。、相关、不相关)或细粒度(e。g。",
"测",
"基础设施损坏，失踪人员）由扁平或分层结构",
"组成的类别（Alam et al.，2018b,2021;Buntain",
"et al., 2021)",
"Contributions我们的主要贡献如下：（1）我",
"自然语言处理（NLP）的另一个挑战是社交",
"们引人了一种基于实体掩码的自适应预训练策",
"媒体和微博平台中普遍存在的数据的性质。例",
"略。（2）我们将标签的层次结构合并为多任务",
"如，与危机相关的领域中的大多数作品都集",
"学习（MTL）问题。（3）我们的经验表明，我",
"'https://github.com/th-nuernberg/",
"们的方法提高了对新事件的泛化，并提高了可",
"crisis-tapt-hmc",
"操作信息类型的检测性能"
] | {
"arxiv_id": "2211.11468",
"language": "zh",
"page_idx": 1
} |
||
2211.11468-zh-2 | 2211.11468-zh-2.jpg | <table>
<tbody>
<tr>
<th> </th>
<th>Train</th>
<th>Test</th>
</tr>
<tr>
<td>$\overline{{\mathrm{Event~Ids}}}$</td>
<td>1-52</td>
<td>$\overline{53-75}$ </td>
</tr>
<tr>
<td>$\#$Events</td>
<td>51</td>
<td>21</td>
</tr>
<tr>
<td>$\#$tweets</td>
<td>50,412</td>
<td>22,003</td>
</tr>
<tr>
<td>${\mathrm{Upper~classes}}$</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>1 $\#$Report (14)</td>
<td>30,389</td>
<td>16,059</td>
</tr>
<tr>
<td>$\#$Other (5)</td>
<td>32,105</td>
<td>10,709</td>
</tr>
<tr>
<td>$\#$CallToAction (3)</td>
<td>1,458</td>
<td>389</td>
</tr>
<tr>
<td>$\#$Request (3)</td>
<td>683</td>
<td>144</td>
</tr>
</tbody>
</table>
Table 1: Overview of the dataset split; the values within the brackets of the upper classes corresponds to the number of unique low-level information types.
### 2 $\textbf{Related Work}$
Crisis Tweet Classification 除了传统的检测方法(例如过滤(Kumar et al.,2011)或众包(Poblet et al.,2014))之外，机器学习在该领域受到了广泛关注。研究人员尝试了几种方法，例如朴素贝叶斯，支持向量机和决策树它们具有术语频率特征 (Habdank et al., 2017) 或静态嵌入(Kejriwal and Zhou, 2019)。最近， Word2Vec (Mikolov et al., 2013) 与卷积和递归神经网络的结合在 (Kersten et al., 2019; Snyder et al., 2019) 领域取得了显着改进。由于 Transformers (Vaswani et al., 2017) 和后续语言模型(Devlin et al., 2019) 的成功，大多数工作都建立在此基础上，并且优于以前的方法(Alam et al., 2021;Wang et al., 2021)
Adaptive Pre-Training 使用语言模型的迁移学习实质上有助于在各种 NLP 任务 (Devlin et al., 2019; Liu et al., 2019; Clark et al., 2020) 中实现最先进的结果。通常，此类语言模型遵循三个训练步骤 (Howard and Ruder, 2018; BenDavid et al., 2020): (1) 对大规模语料库进行预训练；(2)对特定于任务的数据进行可选的预训练；(3)对目标任务进行监督微调。但是，由于计算限制，第二步通常被忽略，因此自适应预训练已被证明是有效的 (Howard and Ruder, 2018) 。因此，Gururangan et al. (2020) 引人了域自适应预训练(DAPT)和任务自适应预训练(TAPT), 涵盖了针对特定任务量身定制的语料库的持续预训练。此外，诸如为推文(Nguyen et al., 2020; Wiegmann et al., 2020) 添加特殊令牌或其他屏蔽语言建模(MLM)方法(Ben-David et al., 2020) 等策略已被证明是有益的
Hierarchical Multi-Label Classification 分层
多标签分类(HMC)涵盖了局部和全局方法以及两个世界 (Wehrmann et al., 2018) 的组合。
本地方法的一种流行分类是细分为每个父节点的本地分类器(LCPN) (Dumais and Chen, 2000) ,每个节点的本地分类器(LCN)(Banerjee et al., 2019) 和每个级别的本地分类器 (LCL) (Wehrmann et al.,2018)。混合方法将全局部分集成为特定约束 ,例如分层 softmax (Brinkmann and Bizer, 2021) 或组合多个本地和全局预测头(Wehrmann et al., 2018) 。最近在信息类型分类方面的工作引人了利用分层结构 (Miyazaki et al.,2019) 的标签嵌入。最后，通过组合某些损失函数(Yu et al., 2021; Wang et al., 2021) , 分类也可以被视为 MTL
# 3 $\mathbf{TREC- IS}$
在这项工作中，我们主要关注共享任务 TRECIS 的数据集，该数据集表示带注释的危机相关推文 (Buntain et al., 2021) 的集合。每条推文都属于灾难事件，并带有高级信息类型的注释， 这些信息类型源自由分层阶段组成的本体。但是，信息类型标签仅作为具有四个较高类$\mathcal{L}_T$ 和 25 个较低类$\mathscr{L}_B$的两级层次结构提供。因此，两个层次结构级别都表示多标签分类任务。按照 TREC-IS 轨道设计，我们将数据集分为与 TREC-IS2020B 任务相对应的训练和测试事件。由于跨事件泛化(Wiegmann et al., 2020) 的要求，此拆分构成了一个具有挑战性的设置。Table$\boxed{\mathrm{E}}$1 给出了每个拆分的概述；显然，信息类型分布是高度不平衡的。例如，具有低关键性的信息类型，如 MultimediaShare(31。7%) 和新闻 (25。4%)盛行。相反，高度关键的信息类型 MovePeople(0。9%) 和 SearchAndRescue(0。$\overline{4\%}$)很少发生(McCreadie et al., 2019)。$^2$
### 4 $\textbf{Method}$
如 Figure 2 中所述，我们的方法结合了实体掩码语言建模(E-MLM) 和 MTL 这两个概念。在下面，我们简要描述我们的方法作为这两个的组合
4.1 Entity-Masked Language Modeling
基于自适应预训练，我们扩展了在 BERT(Devlin et al., 2019) 等大型语料库上预训练的变压器编码器的掩码语言建模。在这里，与事件相关的偏差的缓解是通过用特殊令牌替换实体来促进的-这些实体很容易是特定于事件的-(请参阅 Figure 2a )。通过这种方式，我们打算独立于具体实体捕获与灾难相关的语言模式。在 Ben-David et al.(2020) 之后，除了具有概率$\beta$的标准单词掩码之外，我们还进一步介绍了
${^{2}\text{We provide an overview of the labels with some example}}$
posts in Appendix A . | [
[
"0.3242",
"0.082",
"0.3687",
"0.08435",
"0.367",
"0.0998",
"0.3225",
"0.0974"
],
[
"0.398",
"0.0849",
"0.4358",
"0.0849",
"0.4358",
"0.0974",
"0.398",
"0.0974"
],
[
"0.5146",
"0.0885",
"0.879",
"0.0885",
"0.879",
"0.0986",
"0.5146",
"0.0986"
],
[
"0.1545",
"0.10095",
"0.2334",
"0.10095",
"0.2334",
"0.1152",
"0.1545",
"0.1152"
],
[
"0.3208",
"0.10095",
"0.372",
"0.10095",
"0.372",
"0.1152",
"0.3208",
"0.1152"
],
[
"0.3855",
"0.10095",
"0.4492",
"0.10095",
"0.4492",
"0.1152",
"0.3855",
"0.1152"
],
[
"0.5146",
"0.1027",
"0.879",
"0.1027",
"0.879",
"0.1164",
"0.5146",
"0.1164"
],
[
"0.157",
"0.114",
"0.2275",
"0.1164",
"0.2258",
"0.1318",
"0.1561",
"0.13"
],
[
"0.372",
"0.10394",
"0.388",
"0.10394",
"0.388",
"0.11224",
"0.372",
"0.11224"
],
[
"0.3342",
"0.1152",
"0.357",
"0.1152",
"0.357",
"0.13",
"0.3342",
"0.13"
],
[
"0.4048",
"0.1152",
"0.4265",
"0.1152",
"0.4265",
"0.1312",
"0.4048",
"0.1312"
],
[
"0.511",
"0.11755",
"0.881",
"0.11816",
"0.881",
"0.133",
"0.511",
"0.1318"
],
[
"0.157",
"0.1342",
"0.2225",
"0.1342",
"0.2225",
"0.1484",
"0.157",
"0.1484"
],
[
"0.318",
"0.133",
"0.3762",
"0.133",
"0.3762",
"0.1478",
"0.318",
"0.1478"
],
[
"0.3896",
"0.1354",
"0.4434",
"0.1354",
"0.4434",
"0.1467",
"0.3896",
"0.1467"
],
[
"0.511",
"0.1342",
"0.881",
"0.1342",
"0.881",
"0.1478",
"0.511",
"0.1478"
],
[
"0.1587",
"0.152",
"0.2612",
"0.152",
"0.2612",
"0.1633",
"0.1587",
"0.1633"
],
[
"0.511",
"0.1497",
"0.8774",
"0.152",
"0.8774",
"0.165",
"0.511",
"0.1633"
],
[
"0.318",
"0.1663",
"0.3745",
"0.1663",
"0.3745",
"0.1812",
"0.318",
"0.1812"
],
[
"0.3896",
"0.1663",
"0.444",
"0.1663",
"0.444",
"0.1812",
"0.3896",
"0.1812"
],
[
"0.5146",
"0.1675",
"0.881",
"0.1675",
"0.881",
"0.1812",
"0.5146",
"0.1812"
],
[
"0.318",
"0.1818",
"0.3745",
"0.1818",
"0.3745",
"0.1965",
"0.318",
"0.1965"
],
[
"0.1604",
"0.1687",
"0.2612",
"0.1687",
"0.2612",
"0.1799",
"0.1604",
"0.1799"
],
[
"0.388",
"0.1818",
"0.4434",
"0.1818",
"0.4434",
"0.1965",
"0.388",
"0.1965"
],
[
"0.511",
"0.1818",
"0.8774",
"0.1818",
"0.8774",
"0.1953",
"0.511",
"0.1953"
],
[
"0.398",
"0.1978",
"0.434",
"0.1978",
"0.434",
"0.2131",
"0.398",
"0.2131"
],
[
"0.1604",
"0.1829",
"0.2418",
"0.1829",
"0.2418",
"0.1948",
"0.1604",
"0.1948"
],
[
"0.1587",
"0.199",
"0.3015",
"0.199",
"0.3015",
"0.2131",
"0.1587",
"0.2131"
],
[
"0.3198",
"0.199",
"0.3687",
"0.199",
"0.3687",
"0.2131",
"0.3198",
"0.2131"
],
[
"0.511",
"0.1978",
"0.879",
"0.199",
"0.879",
"0.2131",
"0.511",
"0.212"
],
[
"0.3284",
"0.2131",
"0.3628",
"0.2131",
"0.3628",
"0.2286",
"0.3284",
"0.2286"
],
[
"0.3997",
"0.2131",
"0.434",
"0.2131",
"0.434",
"0.2299",
"0.3997",
"0.2299"
],
[
"0.513",
"0.2156",
"0.881",
"0.2156",
"0.881",
"0.2286",
"0.513",
"0.2286"
],
[
"0.5146",
"0.2322",
"0.881",
"0.2322",
"0.881",
"0.2452",
"0.5146",
"0.2452"
],
[
"0.1604",
"0.2162",
"0.2603",
"0.2162",
"0.2603",
"0.228",
"0.1604",
"0.228"
],
[
"0.11755",
"0.2411",
"0.486",
"0.2405",
"0.486",
"0.2546",
"0.11755",
"0.2559"
],
[
"0.511",
"0.2477",
"0.881",
"0.2477",
"0.881",
"0.2625",
"0.511",
"0.2625"
],
[
"0.11584",
"0.2546",
"0.486",
"0.2559",
"0.486",
"0.2708",
"0.11584",
"0.2695"
],
[
"0.511",
"0.2625",
"0.687",
"0.2632",
"0.687",
"0.2778",
"0.511",
"0.2766"
],
[
"0.11584",
"0.2695",
"0.4307",
"0.2715",
"0.4307",
"0.2852",
"0.11584",
"0.2832"
],
[
"0.51",
"0.2903",
"0.6323",
"0.2903",
"0.6323",
"0.3047",
"0.51",
"0.3047"
],
[
"0.1142",
"0.3113",
"0.2747",
"0.3113",
"0.2747",
"0.3247",
"0.1142",
"0.3247"
],
[
"0.513",
"0.3176",
"0.882",
"0.3176",
"0.882",
"0.332",
"0.513",
"0.332"
],
[
"0.1234",
"0.3362",
"0.4844",
"0.3374",
"0.4844",
"0.3516",
"0.1234",
"0.3508"
],
[
"0.511",
"0.3342",
"0.881",
"0.3342",
"0.881",
"0.3474",
"0.511",
"0.3474"
],
[
"0.11755",
"0.3528",
"0.4836",
"0.3528",
"0.4836",
"0.3665",
"0.11755",
"0.3665"
],
[
"0.51",
"0.3486",
"0.879",
"0.3486",
"0.879",
"0.3623",
"0.51",
"0.3623"
],
[
"0.11755",
"0.3682",
"0.4844",
"0.3682",
"0.4844",
"0.3818",
"0.11755",
"0.3818"
],
[
"0.513",
"0.3665",
"0.881",
"0.3665",
"0.881",
"0.3801",
"0.513",
"0.3801"
],
[
"0.1184",
"0.386",
"0.4844",
"0.386",
"0.4844",
"0.3997",
"0.1184",
"0.3997"
],
[
"0.511",
"0.3818",
"0.882",
"0.3818",
"0.882",
"0.3955",
"0.511",
"0.3955"
],
[
"0.1184",
"0.402",
"0.4844",
"0.402",
"0.4844",
"0.415",
"0.1184",
"0.415"
],
[
"0.513",
"0.398",
"0.8774",
"0.398",
"0.8774",
"0.411",
"0.513",
"0.411"
],
[
"0.1184",
"0.4187",
"0.486",
"0.4187",
"0.486",
"0.4316",
"0.1184",
"0.4316"
],
[
"0.513",
"0.4146",
"0.881",
"0.4146",
"0.881",
"0.4275",
"0.513",
"0.4275"
],
[
"0.1184",
"0.434",
"0.4844",
"0.434",
"0.4844",
"0.4478",
"0.1184",
"0.4478"
],
[
"0.511",
"0.43",
"0.882",
"0.4287",
"0.882",
"0.4436",
"0.511",
"0.4448"
],
[
"0.11584",
"0.4507",
"0.4844",
"0.4507",
"0.4844",
"0.4644",
"0.11584",
"0.4644"
],
[
"0.5146",
"0.4478",
"0.879",
"0.4478",
"0.879",
"0.4578",
"0.5146",
"0.4578"
],
[
"0.513",
"0.4602",
"0.881",
"0.4614",
"0.881",
"0.4756",
"0.513",
"0.4744"
],
[
"0.12006",
"0.4685",
"0.4836",
"0.4685",
"0.4836",
"0.4785",
"0.12006",
"0.4785"
],
[
"0.513",
"0.4768",
"0.881",
"0.478",
"0.881",
"0.4922",
"0.513",
"0.4912"
],
[
"0.1184",
"0.4827",
"0.4878",
"0.4827",
"0.4878",
"0.4963",
"0.1184",
"0.4963"
],
[
"0.11755",
"0.4988",
"0.4878",
"0.4988",
"0.4878",
"0.513",
"0.11755",
"0.513"
],
[
"0.513",
"0.4946",
"0.882",
"0.4946",
"0.882",
"0.5083",
"0.513",
"0.5083"
],
[
"0.51",
"0.509",
"0.881",
"0.5103",
"0.881",
"0.525",
"0.51",
"0.524"
],
[
"0.11755",
"0.5156",
"0.4878",
"0.5156",
"0.4878",
"0.5293",
"0.11755",
"0.5293"
],
[
"0.11755",
"0.531",
"0.4878",
"0.531",
"0.4878",
"0.5444",
"0.11755",
"0.5444"
],
[
"0.513",
"0.527",
"0.881",
"0.527",
"0.881",
"0.5405",
"0.513",
"0.5405"
],
[
"0.11584",
"0.546",
"0.3015",
"0.5464",
"0.3015",
"0.561",
"0.11584",
"0.56"
],
[
"0.5156",
"0.5444",
"0.879",
"0.5444",
"0.879",
"0.5547",
"0.5156",
"0.5547"
],
[
"0.5146",
"0.5586",
"0.881",
"0.5586",
"0.881",
"0.5723",
"0.5146",
"0.5723"
],
[
"0.12177",
"0.5723",
"0.4878",
"0.572",
"0.4878",
"0.586",
"0.12177",
"0.5874"
],
[
"0.513",
"0.5747",
"0.881",
"0.5747",
"0.881",
"0.5884",
"0.513",
"0.5884"
],
[
"0.1184",
"0.589",
"0.4878",
"0.589",
"0.4878",
"0.6025",
"0.1184",
"0.6025"
],
[
"0.51",
"0.5913",
"0.839",
"0.5913",
"0.839",
"0.605",
"0.51",
"0.605"
],
[
"0.11755",
"0.605",
"0.4844",
"0.605",
"0.4844",
"0.618",
"0.11755",
"0.618"
],
[
"0.1184",
"0.6216",
"0.4844",
"0.6216",
"0.4844",
"0.632",
"0.1184",
"0.632"
],
[
"0.511",
"0.6206",
"0.615",
"0.6206",
"0.615",
"0.632",
"0.511",
"0.632"
],
[
"0.12006",
"0.6396",
"0.486",
"0.6396",
"0.486",
"0.6494",
"0.12006",
"0.6494"
],
[
"0.513",
"0.6445",
"0.881",
"0.6445",
"0.881",
"0.6577",
"0.513",
"0.6577"
],
[
"0.11755",
"0.654",
"0.4844",
"0.654",
"0.4844",
"0.6675",
"0.11755",
"0.6675"
],
[
"0.513",
"0.661",
"0.881",
"0.661",
"0.881",
"0.675",
"0.513",
"0.675"
],
[
"0.1184",
"0.6685",
"0.4844",
"0.6685",
"0.4844",
"0.6816",
"0.1184",
"0.6816"
],
[
"0.511",
"0.676",
"0.879",
"0.676",
"0.879",
"0.6895",
"0.511",
"0.6895"
],
[
"0.11755",
"0.685",
"0.4836",
"0.685",
"0.4836",
"0.6987",
"0.11755",
"0.6987"
],
[
"0.511",
"0.6914",
"0.571",
"0.6934",
"0.5693",
"0.708",
"0.51",
"0.706"
],
[
"0.11755",
"0.7007",
"0.4836",
"0.7007",
"0.4836",
"0.7144",
"0.11755",
"0.7144"
],
[
"0.12006",
"0.7188",
"0.4844",
"0.7188",
"0.4844",
"0.7324",
"0.12006",
"0.7324"
],
[
"0.5146",
"0.7217",
"0.5557",
"0.7217",
"0.5557",
"0.7324",
"0.5146",
"0.7324"
],
[
"0.5557",
"0.72",
"0.839",
"0.72",
"0.839",
"0.7354",
"0.5557",
"0.7354"
],
[
"0.11755",
"0.733",
"0.4836",
"0.733",
"0.4836",
"0.7466",
"0.11755",
"0.7466"
],
[
"0.513",
"0.741",
"0.882",
"0.741",
"0.882",
"0.755",
"0.513",
"0.755"
],
[
"0.12006",
"0.749",
"0.4836",
"0.749",
"0.4836",
"0.7627",
"0.12006",
"0.7627"
],
[
"0.513",
"0.7583",
"0.881",
"0.7583",
"0.881",
"0.7715",
"0.513",
"0.7715"
],
[
"0.1184",
"0.7666",
"0.4844",
"0.7666",
"0.4844",
"0.78",
"0.1184",
"0.78"
],
[
"0.51",
"0.7725",
"0.881",
"0.774",
"0.881",
"0.788",
"0.51",
"0.7876"
],
[
"0.12006",
"0.782",
"0.4844",
"0.782",
"0.4844",
"0.796",
"0.12006",
"0.796"
],
[
"0.513",
"0.7905",
"0.881",
"0.7905",
"0.881",
"0.804",
"0.513",
"0.804"
],
[
"0.11755",
"0.799",
"0.4844",
"0.799",
"0.4844",
"0.8125",
"0.11755",
"0.8125"
],
[
"0.511",
"0.8047",
"0.881",
"0.8057",
"0.881",
"0.821",
"0.511",
"0.8193"
],
[
"0.1184",
"0.8154",
"0.4878",
"0.8154",
"0.4878",
"0.829",
"0.1184",
"0.829"
],
[
"0.511",
"0.821",
"0.881",
"0.822",
"0.881",
"0.836",
"0.511",
"0.835"
],
[
"0.11755",
"0.8306",
"0.486",
"0.8306",
"0.486",
"0.844",
"0.11755",
"0.844"
],
[
"0.513",
"0.8384",
"0.881",
"0.8384",
"0.881",
"0.8516",
"0.513",
"0.8516"
],
[
"0.11584",
"0.8457",
"0.1411",
"0.847",
"0.1377",
"0.8623",
"0.1142",
"0.86"
],
[
"0.511",
"0.854",
"0.879",
"0.854",
"0.879",
"0.8677",
"0.511",
"0.8677"
],
[
"0.12006",
"0.8716",
"0.486",
"0.8726",
"0.486",
"0.887",
"0.12006",
"0.8857"
],
[
"0.511",
"0.8706",
"0.876",
"0.8706",
"0.876",
"0.8843",
"0.511",
"0.8843"
],
[
"0.11755",
"0.8896",
"0.486",
"0.8896",
"0.486",
"0.903",
"0.11755",
"0.903"
],
[
"0.531",
"0.891",
"0.882",
"0.8945",
"0.882",
"0.909",
"0.531",
"0.9062"
],
[
"0.11755",
"0.904",
"0.4902",
"0.905",
"0.4902",
"0.9194",
"0.11755",
"0.9185"
],
[
"0.51",
"0.9062",
"0.6465",
"0.905",
"0.6465",
"0.9194",
"0.51",
"0.9204"
]
] | [
"Train",
"Test",
"本地方法的一种流行分类是细分为每个交节",
"Event Ids",
"1-52",
"53-75",
"点的本地分类器（LCPN）（DumaisandChen",
"# Events",
"",
"51",
"21",
"2000），每个节点的本地分类器（LCN)(Banerjee",
"# tweets",
"50,412",
"22,003",
"etal.,2019）和每个级别的本地分类器（LCL）",
"Upperclasses",
"（Wehrmannetal.2018）。混合方法将全局部分",
"30,389",
"16,059",
"集成为特定约束，例如分层softmax（Brinkmann",
"32,105",
"# Report (14)",
"10,709",
"andBizer，2021）或组合多个本地和全局预测头",
"389",
"# Other (5)",
"# CallToAction (3)",
"1,458",
"(Wehrmannetal.，2018）。最近在信息类型分",
"683",
"144",
"类方面的工作引人了利用分层结构（Miyazaki",
"etal.,2019）的标签嵌人。最后，通过组合某些",
"#Request (3)",
"Table I: Overview of the dataset split; the values within",
"损失函数（Yuetal.,2021;Wangetal.,2021），分",
"the brackets of the upper classes corresponds to the",
"类也可以被视为MTL",
"number of unique low-level information types.",
"3TREC-IS",
"2Related Work",
"在这项工作中，我们主要关注共享任务TREC-",
"CrisisTweetClassification除了传统的检测",
"IS的数据集，该数据集表示带注释的危机相关",
"方法（例如过滤（Kumaretal.，2011）或众包",
"推文（Buntainetal.,2021)的集合。每条推文都",
"(Poblet etal.,2014））之外：机器学习在该领",
"属于灾难事件，并带有高级信息类型的注释，",
"域受到了广泛关注。研究人员尝试了几种方",
"这些信息类型源自由分层阶段组成的本体。但",
"法，例如朴素贝叶斯，支持向量机和决策树",
"是，信息类型标签仅作为具有四个较高类LT",
"它们具有术语频率特征（Habdanketal.，2017)",
"和25个较低类LB的两级层次结构提供。因",
"或静态嵌入（KejriwalandZhou，2019）。最近，",
"此，两个层次结构级别都表示多标签分类任",
"Word2Vec（Mikolovetal.，2013）与卷积和递归",
"务。按照TREC-IS轨道设计，我们将数据集分",
"为与TREC-IS2020B任务相对应的训练和测试",
"神经网络的结合在（Kerstenetal.，2019：Snyden",
"事件。由于跨事件泛化（Wiegmannetal.，2020)",
"etal.,2019)领域取得了显着改进。由于Trans-",
"formers（Vaswanietal.，2017）和后续语言模型",
"的要求，此拆分构成了一个具有挑战性的设置。",
"TableE1给出了每个拆分的概述；显然，信息",
"(Devlinetal.2019)的成功，大多数工作都建立",
"在此基础上，并且优于以前的方法（Alametal.,",
"类型分布是高度不平衡的。例如，具有低关键",
"2021; Wang et al., 2021)",
"性的信息类型，如MultimediaShare（31。7%）和",
"新闻（25。4%）盛行。相反，高度关键的信息类",
"AdaptivePre-Training使用语言模型的迁移",
"型MovePeople(0。9%)和 SearchAndRescue(0。",
"学习实质上有助于在各种NLP任务（Devlin",
"4%）很少发生(McCreadieetal.,2019）。2",
"et al., 2019; Liu et al., 2019; Clark et al., 2020) 中",
"实现最先进的结果。通常此类语言模型遵",
"4Method",
"循三个训练步骤（HowardandRuder，2018：Ben-",
"如Figure2中所述，我们的方法结合了实体",
"Davidetal.，2020）：（1）对大规模语料库进行",
"掩码语言建模（E-MLM）和MTL这两个概念。",
"预训练；（2）对特定于任务的数据进行可选",
"在下面，我们简要描述我们的方法作为这两个",
"的预训练；（3）对目标任务进行监督微调。但",
"的组合",
"是，由于计算限制，第二步通常被忽略，因此",
"自适应预训练已被证明是有效的（Howard and",
"4.1",
"Entity-Masked Language Modeling",
"Ruder,2018)。因此，Gururangan etal.(2020)弓",
"基于自适应预训练，我们扩展了在BERT（De-",
"人了域自适应预训练（DAPT）和任务自适应",
"vlinetal.,2019)等大型语料库上预训练的变压",
"预训练（TAPT），涵盖了针对特定任务量身定",
"器编码器的掩码语言建模。在这里，与事件相",
"制的语料库的持续预训练。此外，诸如为推文",
"关的偏差的缓解是通过用特殊令牌替换实体来",
"(Nguyen et al., 2020; Wiegmann et al., 2020) 添",
"促进的-这些实体很容易是特定于事件的-（请",
"加特殊令牌或其他屏蔽语言建模（MLM）方法",
"参阅Figure2a）。通过这种方式，我们打算",
"(Ben-Davidetal.,2020）等策略已被证明是有益",
"独立于具体实体捕获与灾难相关的语言模式。",
"的",
"在Ben-David etal.（2020）之后，除了具有概率",
"HierarchicalMulti-LabelClassification分层",
"β的标准单词掩码之外，我们还进一步介绍了",
"多标签分类（HMC）涵盖了局部和全局方法",
"2we provide an overview of the labels with some example",
"以及两个世界（Wehrmannetal.，2018）的组合。",
"posts in Appendix A ."
] | {
"arxiv_id": "2211.11468",
"language": "zh",
"page_idx": 2
} |
||
1506.02594-zh-4 | 1506.02594-zh-4.jpg | <FigureHere>
Figure 3: Convergence of the mean and media cumulative retweet count $R_t$ as a function of time.The horizontal lines correspond to mean and median final retweet count $R_{14\text{ days}}$
On average, a tweet receives $75\%$ of its retweets in the
${\mathrm{frst~6~hours.}}$
<table>
<tbody>
<tr>
<td>Algorithm 1 Seismic: Predict fnal casca</td>
<td>cade size</td>
</tr>
<tr>
<td>Purpose: For a given post at time $\therefore$ rachare $111m+$</td>
<td>$t$ , predict its fna</td>
</tr>
<tr>
<td>Input: Post resharing information: $0, \ldots , R_t$ . Algorithm: $N_{t}= 0$, $N_{t}^{e}= 0$ for$i=0,\ldots,R_{t}$do MT</td>
<td>$t_i$ and $n_i$ for $i$</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td>$\overline{\mathrm{Al}_{\xi}}$ 1 1</td>
<td>pute real-time infectiousness $\hat{p}(t)$</td>
</tr>
<tr>
<td> </td>
<td>$do$ $\mathbb{R}_{t}$ $K_{t}(t-s)\phi(s-t_{i})ds$ (Sec.4.1)</td>
</tr>
</tbody>
</table>
请注意，Prop。4.1 假设后传染性在将来保持不变$(p_s=p_t$ 表示$s\geq t$ ), 这对于某些信息级联可能是不现实的。我们通过添加两个调整最终预测的缩放常数 $\alpha_t,\gamma_t$ 来更改公式$\mathbb{E}(6)$ 中的预测公式来纠正这一点：
$$\hat{R}_{\infty}(t)=R_{t}+\alpha_{t}\frac{\hat{p}_{t}(N_{t}-N_{t}^{e})}{1-\gamma_{t}\hat{p}_{t}n_{*}},\:0<\alpha_{t},\gamma_{t}<1\:.$$
(7)
我们基于以下直觉引人这些校正因子。我们预计$\alpha_t$会随着时间的推移而减少$t$ ,因此它可以降低未来的估计传染性，这说明该帖子变得陈旧和过时。同样，$\gamma_t$解释了重新海报追随者附近的重叠。随着时间的推移，随着帖子在网络中传播得更远，我们预计$\gamma_t$会随着更多节点多次暴露而增加，这意味着新节点(以前未暴露的节点)的到达率随着时间的推移而降低
我们对所有帖子使用相同的$\alpha_t$和$\gamma_t$值，但允许它们随时间变化。选择$\alpha_t$和$\gamma_t$的值以最小化训练数据集上的中值绝对百分比误差(有关定义，请参阅 5.4 节)。如 5.2 节所述， 我们发现$\alpha_t$在实践中比$\gamma_t$更重要
# $4.3_\text{最后，T}{\text{我}9}$SF到SMH$S_1$akscti协 聘 组 件 放 在 一 起 , 并 在
Seismic 算法中进行合成。算法 1 中描述了用于预测$\hat{R}_\infty(t)$ 的$\hat{p}_t$算法，该算法将用于计算$\hat{p}_t$的算法(算法$\operatorname{E}$2)用作子程序。这些算法基于 Eqs。(4)和(7)。我们假设参数$K_t(s)$ , $\alpha _{t}$ , $\gamma _{t}$ , $n_{* }$ 是先验给出的或根据数据估计的
Computational complexity of Seismic. 对于$\phi(s)$和$K_t(s)$ 的任何选择，对于计算$\hat{p}_t$和预测$\hat{R}_\infty(t)$, Seismic 的计算成本均为$O(R_t)$ 。当然，实际的计算时间在很大程度上取决于集成$\int_t_i^tK_t(t-s)\phi(s-t_i)ds$ 和$\int_t_i^t\phi(s-t_i)ds$。然而，Seismic 的总体计算成本是 linear , 在时间$t$之前观察到的给定帖子的重新共享$R_t$的数量中
线性时间复杂度也部分归因于我们的内存内核的形状。在$\mathbb{E}$ 5.2节中，我们将估计 Twitter 的内存内核$\phi(s)$具有以下形式(对于某些$s_0>0):$
(8)
$$\phi(s)=\begin{cases}c&\text{if}0<s\leq s_0\:,\\c(s/s_0)^{-(1+\theta)}&\text{if}s>s_0\:.\end{cases}$$
这意味着使用方程式$\mathbb{E}(8)$ 中的内存内核 $\phi(s)$ 和方程式$\mathbb{E}(5)$
中的三角加权内核$K_t(s)$可以以整数形式进行计算，因此
$5. _{\text{在 本 }}\textbf{FXP, }\textbf{FRMAENTS}_{\text{itter 数 据 集 , 我 们 的 参 数 估 计 过 }}$
程，并将 Seismic 的性能与最先进的方法进行了比较
$5.1_\text{我们的投据是超Sgriptiorkand。data processing至 }11$ 月7日，Twitter 上有 20 亿条推文和转发。对于每个转发，
数据集包括推文 ID, 发布时间，转发时间以及海报/转发者的关注者数量。请注意，数据集缺少 Twitter 网络信息。我们唯一可用的网络信息是节点的关注者数量
我们专注于具有至少 50 条转发的相当受欢迎的推文的子集，以便我们的模型能够在足够数量的转发发生时立即进行预测。请注意，如果多个 Twitter 用户独立发布相同的推文， 然后转发，则每个原始帖子都会创建自己的独立级联。总而言之，在最初的 15 天里，有 166,076 条推文满足了这一标准。我们使用前 7 天的推文形成训练集，使用后 8 天的推文形成测试集。我们将剩余的 14 天用于转推级联的展开和发展。对于特定的转发级联，我们获得在原始发布时间 i.e. 的 14 天内发布的所有转发，我们通过$R_14$ days 近似$R_\infty$ 。我们用训练集估计参数$\phi(s),\alpha_t,\gamma_t$ and $n_*$,并评估估计器$\hat{R}_\infty$在测试集上的性能。对于我们训练集中的推文，$R_14_{\mathrm{~days~}}$的平均值为209。8 和中位数 110。$R_t$的平均值和中位数的时间演变也显示在图 3中
$5.2_\text{首先，}{\mathrm{S}}$我ISMAGa Perametenestimationationa.1 节)。我们仔细选择训练集中的 15 条推文，并将其所有转发时间的分布用作我们的$\phi(s)$ (图 4 )。15 个转发时间序列的直方图均显示出清晰的亚临界衰减形状。此外，所有原始海报都有大量的追随者。因此，大多数转发，如果不是全部，应该来自原始海报的直接追随者。因此，人类反应时间的分布可以很好地近似为这 15 条推文的转发时间。如果网络结构可用， XMTHX\_2 的估计可以进一步改进 | [
[
"0.08826",
"0.06946",
"0.3962",
"0.06946",
"0.3962",
"0.0802",
"0.08826",
"0.0802"
],
[
"0.6177",
"0.0707",
"0.835",
"0.07135",
"0.835",
"0.0852",
"0.6177",
"0.084"
],
[
"0.10046",
"0.0852",
"0.4805",
"0.0852",
"0.4805",
"0.0985",
"0.10046",
"0.0985"
],
[
"0.5522",
"0.084",
"0.589",
"0.084",
"0.589",
"0.0947",
"0.5522",
"0.0947"
],
[
"0.509",
"0.0966",
"0.562",
"0.0966",
"0.562",
"0.2267",
"0.509",
"0.2267"
],
[
"0.535",
"0.09344",
"0.5537",
"0.09344",
"0.5537",
"0.1016",
"0.535",
"0.1016"
],
[
"0.1013",
"0.09973",
"0.1912",
"0.09973",
"0.1912",
"0.10986",
"0.1013",
"0.10986"
],
[
"0.834",
"0.09973",
"0.8774",
"0.09973",
"0.8774",
"0.1111",
"0.834",
"0.1111"
],
[
"0.1013",
"0.1111",
"0.4756",
"0.1111",
"0.4756",
"0.125",
"0.1013",
"0.125"
],
[
"0.831",
"0.1092",
"0.8857",
"0.1111",
"0.8857",
"0.1225",
"0.831",
"0.1206"
],
[
"0.1013",
"0.1256",
"0.1708",
"0.1256",
"0.1708",
"0.137",
"0.1013",
"0.137"
],
[
"0.10297",
"0.1395",
"0.1748",
"0.1395",
"0.1748",
"0.1509",
"0.10297",
"0.1509"
],
[
"0.1013",
"0.1509",
"0.21",
"0.1509",
"0.21",
"0.1642",
"0.1013",
"0.1642"
],
[
"0.0989",
"0.1642",
"0.2418",
"0.1642",
"0.2418",
"0.178",
"0.0989",
"0.178"
],
[
"0.1217",
"0.1755",
"0.1912",
"0.178",
"0.1912",
"0.1925",
"0.1217",
"0.1901"
],
[
"0.1217",
"0.1901",
"0.2876",
"0.1901",
"0.2876",
"0.2052",
"0.1217",
"0.2052"
],
[
"0.4167",
"0.1913",
"0.478",
"0.1913",
"0.478",
"0.2026",
"0.4167",
"0.2026"
],
[
"0.5337",
"0.1932",
"0.5522",
"0.1932",
"0.5522",
"0.2058",
"0.5337",
"0.2058"
],
[
"0.1013",
"0.2058",
"0.1503",
"0.2058",
"0.1503",
"0.2172",
"0.1013",
"0.2172"
],
[
"0.1013",
"0.2184",
"0.384",
"0.2184",
"0.384",
"0.2317",
"0.1013",
"0.2317"
],
[
"0.423",
"0.2184",
"0.4795",
"0.2184",
"0.4795",
"0.2329",
"0.423",
"0.2329"
],
[
"0.0989",
"0.2311",
"0.2018",
"0.2329",
"0.2018",
"0.2474",
"0.0989",
"0.2456"
],
[
"0.6465",
"0.2468",
"0.808",
"0.2468",
"0.808",
"0.26",
"0.6465",
"0.26"
],
[
"0.08496",
"0.2683",
"0.4207",
"0.2695",
"0.4207",
"0.2842",
"0.08496",
"0.2827"
],
[
"0.0989",
"0.2842",
"0.4795",
"0.286",
"0.4795",
"0.3005",
"0.0989",
"0.2986"
],
[
"0.5156",
"0.279",
"0.9087",
"0.279",
"0.9087",
"0.2922",
"0.5156",
"0.2922"
],
[
"0.5156",
"0.2922",
"0.9087",
"0.2922",
"0.9087",
"0.3062",
"0.5156",
"0.3062"
],
[
"0.1013",
"0.2986",
"0.3718",
"0.2998",
"0.3718",
"0.3132",
"0.1013",
"0.3125"
],
[
"0.5146",
"0.305",
"0.9087",
"0.3062",
"0.9087",
"0.3208",
"0.5146",
"0.3193"
],
[
"0.1013",
"0.3132",
"0.478",
"0.3132",
"0.478",
"0.327",
"0.1013",
"0.327"
],
[
"0.528",
"0.3176",
"0.91",
"0.318",
"0.91",
"0.3333",
"0.528",
"0.332"
],
[
"0.1013",
"0.327",
"0.165",
"0.327",
"0.165",
"0.3384",
"0.1013",
"0.3384"
],
[
"0.5156",
"0.3333",
"0.6006",
"0.3333",
"0.6006",
"0.3447",
"0.5156",
"0.3447"
],
[
"0.10297",
"0.3403",
"0.1732",
"0.3403",
"0.1732",
"0.3516",
"0.10297",
"0.3516"
],
[
"0.1013",
"0.3516",
"0.21",
"0.3516",
"0.21",
"0.3655",
"0.1013",
"0.3655"
],
[
"0.0989",
"0.3655",
"0.2402",
"0.3655",
"0.2402",
"0.3801",
"0.0989",
"0.3801"
],
[
"0.5312",
"0.3687",
"0.91",
"0.3687",
"0.91",
"0.3818",
"0.5312",
"0.3818"
],
[
"0.12335",
"0.379",
"0.2402",
"0.379",
"0.2402",
"0.3926",
"0.12335",
"0.3926"
],
[
"0.5146",
"0.3806",
"0.9067",
"0.3806",
"0.9067",
"0.3945",
"0.5146",
"0.3945"
],
[
"0.1013",
"0.3933",
"0.1503",
"0.3933",
"0.1503",
"0.404",
"0.1013",
"0.404"
],
[
"0.5156",
"0.3958",
"0.6724",
"0.3958",
"0.6724",
"0.4092",
"0.5156",
"0.4092"
],
[
"0.10046",
"0.4048",
"0.2418",
"0.4048",
"0.2418",
"0.4185",
"0.10046",
"0.4185"
],
[
"0.1209",
"0.4153",
"0.3506",
"0.4167",
"0.3506",
"0.4343",
"0.1209",
"0.433"
],
[
"0.415",
"0.4172",
"0.4795",
"0.4172",
"0.4795",
"0.432",
"0.415",
"0.432"
],
[
"0.6357",
"0.4172",
"0.653",
"0.4172",
"0.653",
"0.432",
"0.6357",
"0.432"
],
[
"0.7427",
"0.4172",
"0.839",
"0.4204",
"0.839",
"0.4343",
"0.7427",
"0.4312"
],
[
"0.585",
"0.4268",
"0.6313",
"0.4287",
"0.63",
"0.44",
"0.584",
"0.4382"
],
[
"0.6323",
"0.43",
"0.8164",
"0.43",
"0.8164",
"0.4507",
"0.6323",
"0.4507"
],
[
"0.886",
"0.428",
"0.9087",
"0.428",
"0.9087",
"0.44",
"0.886",
"0.44"
],
[
"0.1013",
"0.4343",
"0.1487",
"0.4343",
"0.1487",
"0.4458",
"0.1013",
"0.4458"
],
[
"0.0989",
"0.4463",
"0.179",
"0.443",
"0.1805",
"0.459",
"0.10046",
"0.4622"
],
[
"0.10046",
"0.457",
"0.1748",
"0.461",
"0.1732",
"0.476",
"0.0989",
"0.4717"
],
[
"0.5156",
"0.4602",
"0.9067",
"0.4602",
"0.9067",
"0.4734",
"0.5156",
"0.4734"
],
[
"0.517",
"0.4734",
"0.888",
"0.4734",
"0.888",
"0.4873",
"0.517",
"0.4873"
],
[
"0.10297",
"0.505",
"0.478",
"0.505",
"0.478",
"0.518",
"0.10297",
"0.518"
],
[
"0.5146",
"0.502",
"0.91",
"0.505",
"0.91",
"0.5234",
"0.5146",
"0.521"
],
[
"0.08826",
"0.5176",
"0.478",
"0.5176",
"0.478",
"0.531",
"0.08826",
"0.531"
],
[
"0.5156",
"0.5215",
"0.857",
"0.5215",
"0.857",
"0.535",
"0.5156",
"0.535"
],
[
"0.0866",
"0.531",
"0.4763",
"0.531",
"0.4763",
"0.5444",
"0.0866",
"0.5444"
],
[
"0.0866",
"0.5444",
"0.2695",
"0.5444",
"0.2695",
"0.558",
"0.0866",
"0.558"
],
[
"0.513",
"0.5415",
"0.911",
"0.5444",
"0.911",
"0.565",
"0.513",
"0.5625"
],
[
"0.1283",
"0.5625",
"0.326",
"0.577",
"0.3218",
"0.6016",
"0.125",
"0.587"
],
[
"0.2378",
"0.5664",
"0.3308",
"0.5664",
"0.3308",
"0.5776",
"0.2378",
"0.5776"
],
[
"0.326",
"0.5684",
"0.4453",
"0.5728",
"0.4436",
"0.5903",
"0.3245",
"0.586"
],
[
"0.5195",
"0.563",
"0.9067",
"0.563",
"0.9067",
"0.5737",
"0.5195",
"0.5737"
],
[
"0.4575",
"0.5757",
"0.478",
"0.5757",
"0.478",
"0.586",
"0.4575",
"0.586"
],
[
"0.517",
"0.5747",
"0.9087",
"0.5747",
"0.9087",
"0.5884",
"0.517",
"0.5884"
],
[
"0.517",
"0.5884",
"0.91",
"0.5884",
"0.91",
"0.6016",
"0.517",
"0.6016"
],
[
"0.0866",
"0.604",
"0.478",
"0.604",
"0.478",
"0.617",
"0.0866",
"0.617"
],
[
"0.524",
"0.6016",
"0.767",
"0.6016",
"0.767",
"0.6157",
"0.524",
"0.6157"
],
[
"0.0866",
"0.616",
"0.4795",
"0.616",
"0.4795",
"0.63",
"0.0866",
"0.63"
],
[
"0.5293",
"0.6143",
"0.9087",
"0.6157",
"0.9087",
"0.629",
"0.5293",
"0.6274"
],
[
"0.08496",
"0.6274",
"0.4795",
"0.629",
"0.4795",
"0.643",
"0.08496",
"0.6426"
],
[
"0.5186",
"0.63",
"0.9067",
"0.63",
"0.9067",
"0.64",
"0.5186",
"0.64"
],
[
"0.0866",
"0.6426",
"0.4795",
"0.6426",
"0.4795",
"0.656",
"0.0866",
"0.656"
],
[
"0.5156",
"0.6416",
"0.9087",
"0.6416",
"0.9087",
"0.655",
"0.5156",
"0.655"
],
[
"0.08826",
"0.656",
"0.4795",
"0.656",
"0.4795",
"0.6694",
"0.08826",
"0.6694"
],
[
"0.517",
"0.654",
"0.9087",
"0.654",
"0.9087",
"0.6675",
"0.517",
"0.6675"
],
[
"0.08826",
"0.6694",
"0.4795",
"0.6694",
"0.4795",
"0.683",
"0.08826",
"0.683"
],
[
"0.5186",
"0.6694",
"0.9067",
"0.6694",
"0.9067",
"0.6797",
"0.5186",
"0.6797"
],
[
"0.1013",
"0.6816",
"0.478",
"0.6816",
"0.478",
"0.696",
"0.1013",
"0.696"
],
[
"0.5156",
"0.681",
"0.9087",
"0.681",
"0.9087",
"0.6943",
"0.5156",
"0.6943"
],
[
"0.0866",
"0.6963",
"0.4795",
"0.6963",
"0.4795",
"0.7104",
"0.0866",
"0.7104"
],
[
"0.5156",
"0.6943",
"0.9087",
"0.6943",
"0.9087",
"0.7085",
"0.5156",
"0.7085"
],
[
"0.0866",
"0.709",
"0.4805",
"0.709",
"0.4805",
"0.7227",
"0.0866",
"0.7227"
],
[
"0.5186",
"0.7085",
"0.9087",
"0.7085",
"0.9087",
"0.7217",
"0.5186",
"0.7217"
],
[
"0.0866",
"0.7227",
"0.3137",
"0.7227",
"0.3137",
"0.7363",
"0.0866",
"0.7363"
],
[
"0.5156",
"0.7197",
"0.9087",
"0.72",
"0.9087",
"0.735",
"0.5156",
"0.7344"
],
[
"0.517",
"0.735",
"0.91",
"0.735",
"0.91",
"0.75",
"0.517",
"0.75"
],
[
"0.08417",
"0.7456",
"0.4805",
"0.7476",
"0.4805",
"0.765",
"0.08417",
"0.763"
],
[
"0.517",
"0.7485",
"0.9087",
"0.7485",
"0.9087",
"0.763",
"0.517",
"0.763"
],
[
"0.08496",
"0.762",
"0.478",
"0.763",
"0.478",
"0.778",
"0.08496",
"0.7764"
],
[
"0.5156",
"0.762",
"0.9087",
"0.762",
"0.9087",
"0.776",
"0.5156",
"0.776"
],
[
"0.0866",
"0.7764",
"0.478",
"0.7764",
"0.478",
"0.7915",
"0.0866",
"0.7915"
],
[
"0.517",
"0.7764",
"0.5947",
"0.7764",
"0.5947",
"0.787",
"0.517",
"0.787"
],
[
"0.08496",
"0.789",
"0.478",
"0.7905",
"0.478",
"0.8047",
"0.08496",
"0.8037"
],
[
"0.513",
"0.7944",
"0.911",
"0.7974",
"0.911",
"0.8184",
"0.513",
"0.815"
],
[
"0.08496",
"0.8047",
"0.3962",
"0.8037",
"0.3962",
"0.8174",
"0.08496",
"0.8184"
],
[
"0.517",
"0.815",
"0.9067",
"0.815",
"0.9067",
"0.829",
"0.517",
"0.829"
],
[
"0.0866",
"0.8237",
"0.478",
"0.8237",
"0.478",
"0.837",
"0.0866",
"0.837"
],
[
"0.517",
"0.829",
"0.9087",
"0.829",
"0.9087",
"0.8423",
"0.517",
"0.8423"
],
[
"0.0866",
"0.8384",
"0.478",
"0.8384",
"0.478",
"0.8516",
"0.0866",
"0.8516"
],
[
"0.5156",
"0.8423",
"0.9087",
"0.8423",
"0.9087",
"0.856",
"0.5156",
"0.856"
],
[
"0.0866",
"0.851",
"0.478",
"0.851",
"0.478",
"0.8643",
"0.0866",
"0.8643"
],
[
"0.517",
"0.855",
"0.9087",
"0.855",
"0.9087",
"0.8687",
"0.517",
"0.8687"
],
[
"0.08826",
"0.866",
"0.4795",
"0.866",
"0.4795",
"0.88",
"0.08826",
"0.88"
],
[
"0.517",
"0.8687",
"0.9087",
"0.8687",
"0.9087",
"0.882",
"0.517",
"0.882"
],
[
"0.0866",
"0.882",
"0.4795",
"0.882",
"0.4795",
"0.896",
"0.0866",
"0.896"
],
[
"0.517",
"0.882",
"0.9087",
"0.882",
"0.9087",
"0.896",
"0.517",
"0.896"
],
[
"0.0866",
"0.896",
"0.2467",
"0.896",
"0.2467",
"0.909",
"0.0866",
"0.909"
],
[
"0.5156",
"0.8945",
"0.7466",
"0.896",
"0.7466",
"0.909",
"0.5156",
"0.9077"
]
] | [
"Algorithm 1 Seismic: Predict final cascade size",
"Evolutionof theNumberof Retweets",
"Purpose: For a given post at time t , predict its final",
"",
"ono",
"200-",
"reshare count",
"Mean",
"Input: Post resharing information: ti and ni for i",
"Median",
"O,...,R..",
"Algorithm:",
"Nt= 0, N= 0",
"for i=O,..., Rt do",
"Nt += ni",
"Ne += ni Jt d(s-ta)ds",
"(Sec. 3.1)",
"50",
"end for",
"Ro(t) = Rt + aαtpt(Nt - N)/(1 -tptn)",
"(Alg. 2)",
"Deliver: Ro(t)",
"Time Since Original Post t (hour)",
"Algorithm 2Compute real-time infectiousness p(t)",
"Purpose: For a given post w, calculate infectiousness pt",
"Figure 3: Convergence of the mean and media cumulative",
"retweet count Rt as a function of time.The horizontal lines",
"with information about w prior to timet",
"correspond to mean and median final retweet count Ri4 days",
"Input: Post resharing information: ti and ni for i -",
"On average, a tweet receives 75 % of its retweets in the",
"0....,Rt",
"first 6 hours.",
"Algorithm:",
"R= 0, N = 0",
"for i=o,..., Rt do",
"线性时间复杂度也部分归因于我们的内存内核的形状。在日",
"Rt += Kt(t- t)",
"5.2节中，我们将估计Twitter的内存内核o(s）具有以下形",
"end for",
"式（对于某些 s0>0）：",
"for i= O,..., Rt do",
"Ne += n Jt Ke(t - s)(s- ta)ds",
"(Sec. 4.1)",
"fc",
"if 0 <s≤s0 ,",
"o(s) =",
"[c(s/so)-(1+0) if s > s0 .",
"(8)",
"end for",
"Pt= Rt/Ne",
"Deliver: pt",
"这意味着使用方程式E(8）中的内存内核Φ(s）和方程式国(5)",
"中的三角加权内核K(s）可以以整数形式进行计算，因此",
"请注意，Prop。4.1假设后传染性在将来保持不变（ps=pt",
"5.在本EXPERMNTSitter数据集，我们的参数估计过",
"表示s≥t），这对于某些信息级联可能是不现实的。我们通",
"程，并将Seismic的性能与最先进的方法进行了比较",
"过添加两个调整最终预测的缩放常数αt：t来更改公式E（6)",
"中的预测公式来纠正这一点：",
"5.我们辑是esgrintie整巢nd.data prosessing至 11",
"Ro(t)= Ri + apl-upin.",
"pt(N-N)",
", 0<αt,%<1.",
"月7日，Twitter上有20亿条推文和转发。对于每个转发",
"(7)",
"数据集包括推文ID，发布时间，转发时间以及海报/转发者的",
"关注者数量。请注意，数据集缺少Twitter网络信息。我们唯",
"我们基于以下直觉引人这些校正因子。我们预计αt会随着时",
"一可用的网络信息是节点的关注者数量",
"间的推移而减少t，因此它可以降低未来的估计传染性，这说",
"我们专注于具有至少50条转发的相当受欢迎的推文的子",
"明该帖子变得陈旧和过时。同样，解释了重新海报追随者",
"集，以便我们的模型能够在足够数量的转发发生时立即进行",
"附近的重叠。随着时间的推移，随着帖子在网络中传播得更",
"预测。请注意，如果多个Twitter用户独立发布相同的推文",
"远，我们预计会随着更多节点多次暴露而增加，这意味着",
"然后转发，则每个原始帖子都会创建自己的独立级联。总而",
"新节点（以前未暴露的节点）的到达率随着时间的推移而降低",
"言之，在最初的15天里，有166.076条推文满足了这一标准。",
"我们对所有帖子使用相同的αt和值，但允许它们随时",
"我们使用前7天的推文形成训练集，使用后8天的推文形成",
"间变化。选择αt和的值以最小化训练数据集上的中值绝",
"测试集。我们将剩余的14天用于转推级联的展开和发展。对",
"对百分比误差（有关定义，请参阅5.4节）。如5.2节所述，",
"于特定的转发级联，我们获得在原始发布时间i.e.的14天内",
"我们发现αt在实践中比更重要",
"发布的所有转发，我们通过R14days近似Reo。我们用训练",
"集估计参数p(s),at，tand n，并评估估计器R。在测试",
"4.最后，ThgSEISMIGiag@组件放在一起，并在",
"集上的性能。对于我们训练集中的推文，R14days的平均值为",
"Seismic算法中进行合成。算法1中描述了用于预测Ro(t)",
"209。8和中位数110。Rt的平均值和中位数的时间演变也显",
"的pt算法，该算法将用于计算pt的算法（算法E2）用作子",
"示在图3中",
"程序。这些算法基于Eqs。（4)和（7）。我们假设参数Kt(s)",
"5.百先，S我称得适a闪特核stumatn3.1节）。我",
"，αt，t，n是先验给出的或根据数据估计的",
"们仔细选择训练集中的15条推文，并将其所有转发时间的分",
"Computational complexity of Seismic.对于 Φ(s)和Kt(s)",
"布用作我们的Φ(s）（图4）。15个转发时间序列的直方图",
"的任何选择，对于计算p和预测R(t），Seismic的计算成",
"均显示出清晰的亚临界衰减形状。此外，所有原始海报都有",
"本均为O(R）。当然，实际的计算时间在很大程度上取决于",
"大量的追随者。因此，大多数转发，如果不是全部，应该来",
"集成JtK(t-s)Φ(s-ti)ds和 ftΦ(s-ti)ds。然而,Seismic",
"自原始海报的直接追随者。因此，人类反应时间的分布可以",
"的总体计算成本是linear，在时间t之前观察到的给定帖子",
"很好地近似为这15条推文的转发时间。如果网络结构可用，",
"的重新共享Rt的数量中",
"XMTHX2的估计可以进一步改进"
] | {
"arxiv_id": "1506.02594",
"language": "zh",
"page_idx": 4
} |
||
2010.00667-zh-0 | 2010.00667-zh-0.jpg | 学习变分词掩码以提高神经文本分类器的可解释性
Hanjie Chen, Yangfeng Ji
Department of Computer Science
University of Virginia Charlottesville, VA, USA
$\{$ hc9mx, yangfeng $\}$ @virginia.edu
# Abstract
To build an interpretable neural text classifier, most of the prior work has focused on designing inherently interpretable models or finding faithful explanations. A new line of work on improving model interpretability has just started, and many existing methods require either prior information or human annotations as additional inputs in training. To address this limitation, we propose the variational word mask ( Vmask) method to automatically learn task-specific important words and reduce irrelevant information on classification, which ultimately improves the interpretability of model predictions. The proposed method is evaluated with three neural text classifiers (CNN, LSTM, and BERT) on seven benchmark text classification datasets. Experiments show the effectiveness of Vmask in improving both model prediction accuracy and interpretability.
# 1 Introduction
神经网络模型由于其在自然语言文本 (Zhang et al., 2015; Yang et al., 2016; Joulin et al., 2017; Devlin et al., 2018) 上的表示学习能力， 在文本分类方面取得了显著的性能。然而，缺乏对其预测行为的理解已成为可靠性和可信性的关键问题，并阻碍了它们在现实世界(Lipton, 2016; Ribeiro et al., 2016; Jacovi and Goldberg, 2020) 中的应用。已经提出了许多解释方法来为神经网络 (Ribeiro et al., 2016; Lundberg and Lee, 2017; Sundararajan et al., 2017) 提供事后解释，但是它们只能解释模型预测， 而不能帮助提高其可解释性
在这项工作中，我们将可解释性视为神经网络模型的固有属性。此外，我们假设具有相似网络架构的神经网络模型可能具有不同级别的可解释性，即使它们可能具有相似的预测性能。Table 1 显示了从具有相似网络体系结构
<table>
<tbody>
<tr>
<th>$Ex.$</th>
<th>Mod.</th>
<th>5l</th>
<th>$\&$Ex Text </th>
<th>$\operatorname{lanation}$</th>
<th> </th>
<th> </th>
<th> </th>
</tr>
<tr>
<th>1</th>
<th>$A$</th>
<th>An</th>
<th>exceeding</th>
<th>clever</th>
<th>piece</th>
<th>of</th>
<th>cin-</th>
</tr>
<tr>
<th>L</th>
<th> </th>
<th>ema</th>
<th> </th>
<th> </th>
<th> </th>
<th> </th>
<th> </th>
</tr>
<tr>
<td> </td>
<td>$A$</td>
<td>It</td>
<td>becomes</td>
<td>mmicky</td>
<td>$inst\epsilon$</td>
<td>ead</td>
<td>of</td>
</tr>
<tr>
<td rowspan="3">2</td>
<td>$B$</td>
<td>COII It cor</td>
<td>pemg becomes pelling</td>
<td>${\mathrm{immicky}}$</td>
<td>ins</td>
<td>stead</td>
<td>of</td>
</tr>
<tr>
<td>$B$</td>
<td>It</td>
<td>becomes</td>
<td>immicky</td>
<td>lns</td>
<td>stead</td>
<td>$of$</td>
</tr>
<tr>
<td> </td>
<td>cor</td>
<td>pelling</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
</tbody>
</table>
Table 1: Model A and B are two neural text classifers with similar network architectures. They all make correct sentiment predictions on both texts (ex.1:positive; ex.2:negative).Twopost-hoc explanation methods, LIME (Ribeiro et al., 2016) and SampleShapley (Kononenko et al., 2010) , are used to explain the model predictions on example 1 and 2 respectively. Top three important words are shown in pink or blue for model A and B. Whichever post-hoc method is used, explanations from model B are easier to understand because the sentiment keywords“clever”and“gimmicky”are $\operatorname{highlighted}.$ | [
[
"0.2377",
"0.08435",
"0.7656",
"0.08435",
"0.7656",
"0.10156",
"0.2377",
"0.10156"
],
[
"0.3938",
"0.1318",
"0.617",
"0.1318",
"0.617",
"0.1467",
"0.3938",
"0.1467"
],
[
"0.3594",
"0.1484",
"0.65",
"0.1484",
"0.65",
"0.1621",
"0.3594",
"0.1621"
],
[
"0.409",
"0.1633",
"0.5996",
"0.165",
"0.5996",
"0.1799",
"0.409",
"0.1776"
],
[
"0.3955",
"0.1812",
"0.613",
"0.1812",
"0.613",
"0.1953",
"0.3955",
"0.1953"
],
[
"0.3586",
"0.1978",
"0.659",
"0.1978",
"0.659",
"0.212",
"0.3586",
"0.212"
],
[
"0.2644",
"0.2654",
"0.3435",
"0.2654",
"0.3435",
"0.2798",
"0.2644",
"0.2798"
],
[
"0.147",
"0.2957",
"0.4626",
"0.2957",
"0.4626",
"0.3088",
"0.147",
"0.3088"
],
[
"0.147",
"0.3088",
"0.4592",
"0.3088",
"0.4592",
"0.3225",
"0.147",
"0.3225"
],
[
"0.147",
"0.3247",
"0.4592",
"0.3247",
"0.4592",
"0.3384",
"0.147",
"0.3384"
],
[
"0.1453",
"0.3384",
"0.4324",
"0.3384",
"0.4324",
"0.3516",
"0.1453",
"0.3516"
],
[
"0.4375",
"0.3416",
"0.4534",
"0.3384",
"0.4575",
"0.3486",
"0.44",
"0.3516"
],
[
"0.147",
"0.3528",
"0.461",
"0.3528",
"0.461",
"0.3665",
"0.147",
"0.3665"
],
[
"0.147",
"0.3677",
"0.4592",
"0.3677",
"0.4592",
"0.3813",
"0.147",
"0.3813"
],
[
"0.147",
"0.3813",
"0.4592",
"0.3813",
"0.4592",
"0.3943",
"0.147",
"0.3943"
],
[
"0.1486",
"0.3967",
"0.4592",
"0.3967",
"0.4592",
"0.4067",
"0.1486",
"0.4067"
],
[
"0.1486",
"0.411",
"0.3586",
"0.411",
"0.3586",
"0.4216",
"0.1486",
"0.4216"
],
[
"0.361",
"0.4146",
"0.3687",
"0.4146",
"0.3687",
"0.4175",
"0.361",
"0.4175"
],
[
"0.367",
"0.4104",
"0.4592",
"0.4104",
"0.4592",
"0.4216",
"0.367",
"0.4216"
],
[
"0.5283",
"0.4216",
"0.602",
"0.4216",
"0.602",
"0.4329",
"0.5283",
"0.4329"
],
[
"0.629",
"0.4204",
"0.769",
"0.4204",
"0.769",
"0.434",
"0.629",
"0.434"
],
[
"0.594",
"0.437",
"0.764",
"0.4382",
"0.764",
"0.4531",
"0.594",
"0.452"
],
[
"0.147",
"0.4233",
"0.461",
"0.4233",
"0.461",
"0.437",
"0.147",
"0.437"
],
[
"0.147",
"0.4382",
"0.4626",
"0.4382",
"0.4626",
"0.452",
"0.147",
"0.452"
],
[
"0.5684",
"0.4407",
"0.581",
"0.4407",
"0.581",
"0.4507",
"0.5684",
"0.4507"
],
[
"0.771",
"0.4407",
"0.833",
"0.4407",
"0.833",
"0.4507",
"0.771",
"0.4507"
],
[
"0.839",
"0.4407",
"0.869",
"0.4407",
"0.869",
"0.452",
"0.839",
"0.452"
],
[
"0.5356",
"0.449",
"0.5425",
"0.449",
"0.5425",
"0.4536",
"0.5356",
"0.4536"
],
[
"0.5996",
"0.4507",
"0.6304",
"0.4531",
"0.629",
"0.466",
"0.5977",
"0.4644"
],
[
"0.575",
"0.4644",
"0.5884",
"0.4739",
"0.575",
"0.4827",
"0.5615",
"0.4739"
],
[
"0.1486",
"0.4536",
"0.4592",
"0.4536",
"0.4592",
"0.4644",
"0.1486",
"0.4644"
],
[
"0.1486",
"0.4685",
"0.4592",
"0.4685",
"0.4592",
"0.4785",
"0.1486",
"0.4785"
],
[
"0.5977",
"0.4685",
"0.6245",
"0.4685",
"0.6245",
"0.4797",
"0.5977",
"0.4797"
],
[
"0.638",
"0.4697",
"0.7188",
"0.4697",
"0.7188",
"0.481",
"0.638",
"0.481"
],
[
"0.742",
"0.4697",
"0.784",
"0.4697",
"0.784",
"0.481",
"0.742",
"0.481"
],
[
"0.804",
"0.4673",
"0.873",
"0.4673",
"0.873",
"0.4822",
"0.804",
"0.4822"
],
[
"0.147",
"0.481",
"0.4592",
"0.481",
"0.4592",
"0.4946",
"0.147",
"0.4946"
],
[
"0.5996",
"0.481",
"0.65",
"0.481",
"0.65",
"0.4922",
"0.5996",
"0.4922"
],
[
"0.1486",
"0.4963",
"0.4592",
"0.4963",
"0.4592",
"0.5073",
"0.1486",
"0.5073"
],
[
"0.6265",
"0.4976",
"0.7656",
"0.4976",
"0.7656",
"0.5127",
"0.6265",
"0.5127"
],
[
"0.852",
"0.4988",
"0.875",
"0.4988",
"0.875",
"0.511",
"0.852",
"0.511"
],
[
"0.5977",
"0.4995",
"0.615",
"0.4995",
"0.615",
"0.509",
"0.5977",
"0.509"
],
[
"0.5684",
"0.4995",
"0.5825",
"0.4995",
"0.5825",
"0.509",
"0.5684",
"0.509"
],
[
"0.781",
"0.4995",
"0.833",
"0.4995",
"0.833",
"0.511",
"0.781",
"0.511"
],
[
"0.532",
"0.506",
"0.546",
"0.506",
"0.546",
"0.5156",
"0.532",
"0.5156"
],
[
"0.596",
"0.5083",
"0.677",
"0.5103",
"0.677",
"0.525",
"0.596",
"0.5225"
],
[
"0.1486",
"0.525",
"0.4592",
"0.525",
"0.4592",
"0.535",
"0.1486",
"0.535"
],
[
"0.147",
"0.509",
"0.4592",
"0.509",
"0.4592",
"0.5225",
"0.147",
"0.5225"
],
[
"0.567",
"0.528",
"0.5977",
"0.528",
"0.5977",
"0.539",
"0.567",
"0.539"
],
[
"0.594",
"0.528",
"0.6245",
"0.528",
"0.6245",
"0.539",
"0.594",
"0.539"
],
[
"0.623",
"0.527",
"0.687",
"0.5293",
"0.687",
"0.5405",
"0.623",
"0.538"
],
[
"0.706",
"0.5293",
"0.7734",
"0.5293",
"0.7734",
"0.5405",
"0.706",
"0.5405"
],
[
"0.791",
"0.528",
"0.873",
"0.5254",
"0.873",
"0.5405",
"0.791",
"0.542"
],
[
"0.147",
"0.5376",
"0.4626",
"0.5376",
"0.4626",
"0.5503",
"0.147",
"0.5503"
],
[
"0.596",
"0.5376",
"0.677",
"0.539",
"0.677",
"0.554",
"0.596",
"0.552"
],
[
"0.1486",
"0.5527",
"0.4592",
"0.5527",
"0.4592",
"0.5635",
"0.1486",
"0.5635"
],
[
"0.147",
"0.5654",
"0.461",
"0.5654",
"0.461",
"0.579",
"0.147",
"0.579"
],
[
"0.5146",
"0.568",
"0.8833",
"0.568",
"0.8833",
"0.5806",
"0.5146",
"0.5806"
],
[
"0.1453",
"0.58",
"0.444",
"0.58",
"0.444",
"0.5933",
"0.1453",
"0.5933"
],
[
"0.513",
"0.5806",
"0.885",
"0.5806",
"0.885",
"0.594",
"0.513",
"0.594"
],
[
"0.5146",
"0.597",
"0.8833",
"0.597",
"0.8833",
"0.608",
"0.5146",
"0.608"
],
[
"0.1545",
"0.611",
"0.2644",
"0.611",
"0.2644",
"0.6226",
"0.1545",
"0.6226"
],
[
"0.5146",
"0.61",
"0.8833",
"0.61",
"0.8833",
"0.6235",
"0.5146",
"0.6235"
],
[
"0.513",
"0.6245",
"0.882",
"0.6245",
"0.882",
"0.638",
"0.513",
"0.638"
],
[
"0.11584",
"0.6123",
"0.1293",
"0.6094",
"0.1351",
"0.6206",
"0.12177",
"0.6235"
],
[
"0.12006",
"0.637",
"0.4902",
"0.637",
"0.4902",
"0.651",
"0.12006",
"0.651"
],
[
"0.513",
"0.638",
"0.8833",
"0.638",
"0.8833",
"0.652",
"0.513",
"0.652"
],
[
"0.1184",
"0.653",
"0.4902",
"0.653",
"0.4902",
"0.6665",
"0.1184",
"0.6665"
],
[
"0.5146",
"0.653",
"0.8833",
"0.653",
"0.8833",
"0.6665",
"0.5146",
"0.6665"
],
[
"0.1184",
"0.6694",
"0.482",
"0.6694",
"0.482",
"0.6797",
"0.1184",
"0.6797"
],
[
"0.513",
"0.6675",
"0.8833",
"0.6675",
"0.8833",
"0.681",
"0.513",
"0.681"
],
[
"0.4802",
"0.6733",
"0.4844",
"0.6733",
"0.4844",
"0.6787",
"0.4802",
"0.6787"
],
[
"0.1184",
"0.684",
"0.486",
"0.684",
"0.486",
"0.6978",
"0.1184",
"0.6978"
],
[
"0.511",
"0.6797",
"0.881",
"0.6797",
"0.881",
"0.6934",
"0.511",
"0.6934"
],
[
"0.513",
"0.695",
"0.885",
"0.6953",
"0.885",
"0.709",
"0.513",
"0.708"
],
[
"0.12177",
"0.7017",
"0.4878",
"0.7017",
"0.4878",
"0.7153",
"0.12177",
"0.7153"
],
[
"0.513",
"0.709",
"0.8833",
"0.709",
"0.8833",
"0.7227",
"0.513",
"0.7227"
],
[
"0.12177",
"0.7188",
"0.4878",
"0.7188",
"0.4878",
"0.7324",
"0.12177",
"0.7324"
],
[
"0.513",
"0.7236",
"0.8833",
"0.7236",
"0.8833",
"0.737",
"0.513",
"0.737"
],
[
"0.12006",
"0.734",
"0.486",
"0.734",
"0.486",
"0.7446",
"0.12006",
"0.7446"
],
[
"0.5146",
"0.7393",
"0.601",
"0.7393",
"0.601",
"0.7505",
"0.5146",
"0.7505"
],
[
"0.12006",
"0.75",
"0.4878",
"0.75",
"0.4878",
"0.763",
"0.12006",
"0.763"
],
[
"0.12177",
"0.7666",
"0.4895",
"0.7666",
"0.4895",
"0.78",
"0.12177",
"0.78"
],
[
"0.1184",
"0.782",
"0.4878",
"0.782",
"0.4878",
"0.796",
"0.1184",
"0.796"
],
[
"0.12006",
"0.7974",
"0.4895",
"0.7974",
"0.4895",
"0.8125",
"0.12006",
"0.8125"
],
[
"0.12006",
"0.814",
"0.3418",
"0.814",
"0.3418",
"0.8276",
"0.12006",
"0.8276"
],
[
"0.1377",
"0.8306",
"0.4878",
"0.8306",
"0.4878",
"0.844",
"0.1377",
"0.844"
],
[
"0.1184",
"0.8457",
"0.4878",
"0.8467",
"0.4878",
"0.861",
"0.1184",
"0.86"
],
[
"0.12006",
"0.8633",
"0.4878",
"0.8633",
"0.4878",
"0.877",
"0.12006",
"0.877"
],
[
"0.12006",
"0.879",
"0.4878",
"0.879",
"0.4878",
"0.8926",
"0.12006",
"0.8926"
],
[
"0.12006",
"0.8955",
"0.4878",
"0.8955",
"0.4878",
"0.909",
"0.12006",
"0.909"
]
] | [
"学习变分词掩码以提高神经文本分类器的可解释性",
"Hanjie Chen, Yangfeng Ji",
"Department of Computer Science",
"University of Virginia",
"Charlottesville, VA,USA",
"[ hc9mx, yangfeng ↓ @virginia.edu",
"Abstract",
"To build an interpretable neural text clas-",
"sifier, most of the prior work has focused",
"on designing inherently interpretable mod.",
"els or finding faithful explanations.",
"A",
"new line of work on improving model in-",
"terpretability has just started, and many",
"existing methods require either prior in",
"formation or human annotations as ad-",
"ditional inputs in training.",
"",
"To address",
"Ex. Model",
"Text & Explanation",
"An exceedingly clever",
"this limitation, we propose the variational",
"word mask ( Vmask) method to automat-",
"A",
"piece of",
"cin",
"1",
"ema",
"",
"icallylearn task-specific important words",
"and reduce irrelevant information on clas-",
"An",
"exceedingly",
"clever",
"pieceof",
"sification, which ultimately improves the",
"cinema",
"interpretability of model predictions.The",
"becomesgimmicky",
"of ",
"It",
"A",
"instead",
"2",
" compelling",
"neural text classifiers (CNN, LSTM, and",
"proposed method is evaluated with three",
"B",
"It",
"becomes",
"gimmicky",
"insteadof",
"BERT) on seven benchmark text classifica-",
"compelling",
"tion datasets.Experiments show the effec",
"tiveness of Vmask in improving both model",
"Table l: Model A and B are two neural text classi-",
"prediction accuracy and interpretability.",
"fiers with similar network architectures. They all",
"make correct sentiment predictions on both texts",
"Introduction",
"(ex. l: positive; ex. 2: negative). Two post-hoc",
"explanation methods, LIME (Ribeiro et al., 2016)",
"1",
"神经网络模型由于其在自然语言文本（Zhang",
"and SampleShapley (Kononenko et al., 2010) , are",
"et al., 2015; Yang et al., 2016; Joulin et al.,",
"used to explain the model predictions on example",
"2017：Devlinetal..2018）上的表示学习能力",
"1 and 2 respectively. Top three important words",
"",
"在文本分类方面取得了显著的性能。然而，缺乏",
"are shown in pink or blue for model A and B.",
"Whichever post-hoc method is used, explanations",
"对其预测行为的理解已成为可靠性和可信性的",
"from model B are easier to understand because the",
"关键问题，并阻碍了它们在现实世界（Lipton，",
"sentiment keywords \"clever\" and “gimmicky\"are",
"2016: Ribeiro et al..2016: Jacovi and Gold",
"highlighted.",
"berg，2020）中的应用。已经提出了许多解释",
"方法来为神经网络（Ribeiroetal.，2016;Lund-",
"berg and Lee, 2017; Sundararajan et al., 2017)",
"提供事后解释，但是它们只能解释模型预测，",
"而不能帮助提高其可解释性",
"在这项工作中，我们将可解释性视为神经网",
"络模型的固有属性。此外，我们假设具有相似",
"网络架构的神经网络模型可能具有不同级别",
"的可解释性，即使它们可能具有相似的预测性",
"能。Table1显示了从具有相似网络体系结构"
] | {
"arxiv_id": "2010.00667",
"language": "zh",
"page_idx": 0
} |
||
2010.00667-zh-2 | 2010.00667-zh-2.jpg | 为了帮助神经网络模型进行更好的特征选择，我们在单词嵌人之后添加了一个随机层$R$ ,其中$\boldsymbol R=[R_{x_1},\ldots,R_{x_T}]$的长度与$x$相同。每个$R_{x_t}\in\{0,1\}$都是与单词类型$x_t$关联的二进制随机变量，而不是单词位置。这个随机层和单词嵌入一起形成了神经网络模型的输人，即。e。,
$$Z=R\odot x,$$
(1)
其中$\odot$是逐元素乘法，每个$Z_t=R_{x_t}\cdot x_t~.$ 直观上，$Z$只包含$x$的一个子集，该子集由$R$随机选择。由于$R$作为 0-1 掩码序列直接应用于单词，因此我们在本工作中也将其称为单词掩码层
为了确保$Z$具有足够的有关预测$Y$的信息，同时包含来自$x$的最少冗余信息，我们遵循信息瓶颈理论 (Tishby et al., 2000) 中的标准实践，并将目标函数编写为
(2)
$$\max_{\boldsymbol{Z}}I(\boldsymbol{Z};\boldsymbol{Y})-\beta\cdot I(\boldsymbol{Z};\boldsymbol{X}),$$
,其中$X$作为表示通用单词序列的随机变量作为输入，一个。这个公式反映了我们对$Z$ 的确切期望。这里的主要挑战是计算互信息
### 3.2 Variational Word Masks
受 Alemi et al. (2016) 提出的变分信息瓶颈的启发，我们从近似分布$q(X,Y,Z)$开始，而不是计算$p(\boldsymbol{X},\boldsymbol{Y},\boldsymbol{Z})$ 。然后，在下面指定的一些假设下，我们在 Equation 2 中构造了目标的可处理下界，并在 Appendix A 中提供了详细的推导
对于$q$下的$I(\boldsymbol Z;\boldsymbol Y)$,我们有$I( Z; Y)$ = $\sum _{y, z}q( \boldsymbol{y}, \boldsymbol{z}) \log ( q( \boldsymbol{y}| \boldsymbol{z}) / q( \boldsymbol{y}) )$ 。通过将$\log q(\boldsymbol{y}|\boldsymbol{z})$替换为从真实分布$\log p(\boldsymbol{y}|\boldsymbol{z})$得出的条件概率，我们从分布中引人了$Y$和$Z$之间的约束，并且还获得了$I(Z;Y)$ 的下界，
$$\begin{aligned}I(\boldsymbol{Z};\boldsymbol{Y})&\geq\quad\sum_{\boldsymbol{y},\boldsymbol{z}}q(\boldsymbol{y},\boldsymbol{z})\log p(\boldsymbol{y}|\boldsymbol{z})+H_{q}(\boldsymbol{Y})\\&=\quad\sum_{\boldsymbol{y},\boldsymbol{z},\boldsymbol{x}}q(\boldsymbol{x},\boldsymbol{y})q(\boldsymbol{z}|\boldsymbol{x})\log p(\boldsymbol{y}|\boldsymbol{z})\\&+H_{q}(\boldsymbol{Y}),&\text{(3)}\end{aligned}$$
,其中$H_q(\cdot)$是熵，最后一步使用基于$q( \boldsymbol{x}, \boldsymbol{y}, \boldsymbol{z})$ = $q( \boldsymbol{x}) q( \boldsymbol{y}| \boldsymbol{x}) q( \boldsymbol{z}| \boldsymbol{x})$的条件。给定一个特定的观察值($x^(i),y^{(i)})$,我们将经验分布$q(\boldsymbol{X}^{(i)},\boldsymbol{Y}^{(i)})$定义为两个 Delta 函数$q(\boldsymbol{X}^{(i)}=\boldsymbol{x}^{(i)},\boldsymbol{Y}^{(i)}=\boldsymbol{y}^{(i)})=\delta_{\boldsymbol{x}^{(i)}}(\boldsymbol{x})\cdot\delta_{\boldsymbol{y}^{(i)}}(\boldsymbol{y})$ 的乘积。然后，Equation 3 可以进一步简化为
$$\begin{aligned}I(\boldsymbol{Z};\boldsymbol{Y}^{(i)})&\geq\quad\sum_{\boldsymbol{z}}q(\boldsymbol{z}|\boldsymbol{x}^{(i)})\log p(\boldsymbol{y}^{(i)}|\boldsymbol{z})\\&=\quad\mathbb{E}_{q(\boldsymbol{z}|\boldsymbol{x}^{(i)})}[\log p(\boldsymbol{y}^{(i)}|\boldsymbol{z})].\quad(\end{aligned}$$
类似地，对于$q$下的$I(\boldsymbol{Z};\boldsymbol{X})$ ,我们通过将
$p(\boldsymbol{Z}|\boldsymbol{X})$替换为预定义的先验分布$p_0(\boldsymbol{Z})$
$$\begin{aligned}I(\boldsymbol{Z};\boldsymbol{X})&\leq\quad\mathbb{E}_{q(\boldsymbol{x})}[\mathrm{KL}[q(\boldsymbol{z}|\boldsymbol{x})\|p_{0}(\boldsymbol{z})]]\\&=\quad\mathrm{KL}[q(\boldsymbol{z}|\boldsymbol{x}^{(i)})\|p_{0}(\boldsymbol{z})],\end{aligned}$$
(5)
来获得$I(\boldsymbol{Z};\boldsymbol{X})$的上限，其中 KL[·小]表示Kullback-Leibler 散度。最后一步的简化类似于具有经验分布$q(\boldsymbol{X}^{(i)})$的 Equation 4
将(5)和(4)代人 Equation 2,为我们
提供了信息瓶颈的下限$\mathcal{L}$
$$\begin{aligned}\mathcal{L}=&\mathbb{E}_{q(\boldsymbol{z}|\boldsymbol{x}^{(i)})}[\log p(\boldsymbol{y}^{(i)}|\boldsymbol{z})]\\&-\beta\cdot\mathrm{KL}[q(\boldsymbol{z}|\boldsymbol{x}^{(i)})\|p_{0}(\boldsymbol{z})].\end{aligned}$$
(6)
学习目标是使 Equation 6 相对于近似分布$q(\boldsymbol{X},\boldsymbol{Y},\boldsymbol{Z})=q(\boldsymbol{X},\boldsymbol{Y})q(\boldsymbol{Z}|\boldsymbol{X})$最大化。作为分类问题，$X$和$Y$都被观察到，$q(\boldsymbol{X},\boldsymbol{Y})$已经被简化为经验分布，近似分布中唯一剩下的是$q(\boldsymbol{Z}|\boldsymbol{X})$ 。与变分推理 (Alemi et al., 2016; Rezende and Mohamed, 2015) 中的目标函数类似，$\mathcal{L}$中的第一项是确保$q(Z|X)$中用于预测$Y$的信息，而$\mathcal{L}$中的第二项是使用预定义的先验分布$p_0(\boldsymbol{Z})$对$q(\boldsymbol{Z}|\boldsymbol{X})$进行正则化
获得实用目标函数的最后一步是注意到，给定$X_t^{(i)}=x_t^{(i)}$ ,每个$Z_t$都可以重新定义为
$$Z_t=R_{x_t}\cdot x_t^{(i)},$$
(7)
,其中$R_{x_t}\in\{0,1\}$是标准的伯努利分布。然后，可以使用$\boldsymbol{R}=[R_{x_1},\ldots,R_{x_T}]$将$\boldsymbol{Z}$重新参数化为$Z=\boldsymbol{R}\odot x^{(i)}$ 。下限$\mathcal{L}$可以用随机变量$R$重写为
$$\begin{aligned}\mathcal{L}=&\mathbb{E}_{q(\boldsymbol{r}|\boldsymbol{x}^{(i)})}[\log p(\boldsymbol{y}^{(i)}|\boldsymbol{R},\boldsymbol{x}^{(i)})]\\&-\beta\cdot\mathrm{KL}[q(\boldsymbol{R}|\boldsymbol{x}^{(i)})\|p_{0}(\boldsymbol{R})].\end{aligned}$$
(8)
请注意，尽管$\beta$是从信息瓶颈理论继承而来的，但实际上它将用作可调超参数来解决臭名昭著的后验崩溃问题 (Bowman et al., 2016; Kim et al., 2018)
## 3.3 Connections
使用信息瓶颈方法修改单词嵌入的想法最近在 NLP 中显示了一些有趣的应用程序。例如，Li and Eisner (2019)提出了两种将单词嵌入转换为新表示形式的方法，以实现更好的 POS 标记和句法分析。根据 Equation 1 , Vmask 可以看作是单词嵌人的简单线性变换。不同之处在于 $\{R_x_t\}$ 是在词汇表上定义的， 因此可以用来表示$x_t$这个词的全局重要性。回想一下$R_{x_t}\in\{0,1\}$,从稍微不同的角度来 | [
[
"0.1394",
"0.0766",
"0.4895",
"0.0766",
"0.4895",
"0.0903",
"0.1394",
"0.0903"
],
[
"0.534",
"0.0766",
"0.882",
"0.0766",
"0.882",
"0.0903",
"0.534",
"0.0903"
],
[
"0.1184",
"0.0932",
"0.4895",
"0.0932",
"0.4895",
"0.1069",
"0.1184",
"0.1069"
],
[
"0.4963",
"0.0968",
"0.513",
"0.0968",
"0.513",
"0.10394",
"0.4963",
"0.10394"
],
[
"0.51",
"0.09265",
"0.8364",
"0.09265",
"0.8364",
"0.1057",
"0.51",
"0.1057"
],
[
"0.1351",
"0.1081",
"0.4895",
"0.10925",
"0.4895",
"0.12354",
"0.1351",
"0.1223"
],
[
"0.1276",
"0.1152",
"0.1368",
"0.1152",
"0.1368",
"0.1194",
"0.1276",
"0.1194"
],
[
"0.12006",
"0.1271",
"0.4844",
"0.1271",
"0.4844",
"0.1372",
"0.12006",
"0.1372"
],
[
"0.5474",
"0.1223",
"0.8506",
"0.1223",
"0.8506",
"0.1372",
"0.5474",
"0.1372"
],
[
"0.1301",
"0.1401",
"0.486",
"0.1401",
"0.486",
"0.1538",
"0.1301",
"0.1538"
],
[
"0.6274",
"0.1426",
"0.882",
"0.1426",
"0.882",
"0.1567",
"0.6274",
"0.1567"
],
[
"0.12006",
"0.1567",
"0.4878",
"0.1567",
"0.4878",
"0.1704",
"0.12006",
"0.1704"
],
[
"0.1276",
"0.1455",
"0.1368",
"0.1455",
"0.1368",
"0.1497",
"0.1276",
"0.1497"
],
[
"0.1184",
"0.1716",
"0.2241",
"0.1733",
"0.2241",
"0.1882",
"0.1184",
"0.1859"
],
[
"0.5156",
"0.1733",
"0.882",
"0.1733",
"0.882",
"0.187",
"0.5156",
"0.187"
],
[
"0.2544",
"0.1907",
"0.351",
"0.1907",
"0.351",
"0.2019",
"0.2544",
"0.2019"
],
[
"0.4626",
"0.1907",
"0.4895",
"0.1907",
"0.4895",
"0.2031",
"0.4626",
"0.2031"
],
[
"0.5146",
"0.1895",
"0.882",
"0.1895",
"0.882",
"0.2037",
"0.5146",
"0.2037"
],
[
"0.1184",
"0.2079",
"0.492",
"0.2102",
"0.492",
"0.2245",
"0.1184",
"0.2227"
],
[
"0.5156",
"0.2048",
"0.831",
"0.2048",
"0.831",
"0.2197",
"0.5156",
"0.2197"
],
[
"0.532",
"0.2214",
"0.882",
"0.2214",
"0.882",
"0.2351",
"0.532",
"0.2351"
],
[
"0.1234",
"0.228",
"0.4844",
"0.228",
"0.4844",
"0.2382",
"0.1234",
"0.2382"
],
[
"0.12006",
"0.2423",
"0.4878",
"0.2423",
"0.4878",
"0.2559",
"0.12006",
"0.2559"
],
[
"0.513",
"0.2369",
"0.721",
"0.2369",
"0.721",
"0.2517",
"0.513",
"0.2517"
],
[
"0.12006",
"0.2583",
"0.4878",
"0.2583",
"0.4878",
"0.2715",
"0.12006",
"0.2715"
],
[
"0.575",
"0.2642",
"0.785",
"0.2625",
"0.785",
"0.284",
"0.575",
"0.2861"
],
[
"0.1184",
"0.2725",
"0.2124",
"0.275",
"0.2124",
"0.2893",
"0.1184",
"0.2874"
],
[
"0.8564",
"0.279",
"0.882",
"0.279",
"0.882",
"0.2922",
"0.8564",
"0.2922"
],
[
"0.1377",
"0.2903",
"0.4878",
"0.2903",
"0.4878",
"0.304",
"0.1377",
"0.304"
],
[
"0.615",
"0.2893",
"0.818",
"0.2893",
"0.818",
"0.304",
"0.615",
"0.304"
],
[
"0.12177",
"0.3071",
"0.4878",
"0.3071",
"0.4878",
"0.3206",
"0.12177",
"0.3206"
],
[
"0.12006",
"0.3225",
"0.4895",
"0.3225",
"0.4895",
"0.3362",
"0.12006",
"0.3362"
],
[
"0.513",
"0.3184",
"0.882",
"0.3184",
"0.882",
"0.332",
"0.513",
"0.332"
],
[
"0.12006",
"0.3384",
"0.3594",
"0.3384",
"0.3594",
"0.3516",
"0.12006",
"0.3516"
],
[
"0.5156",
"0.3362",
"0.882",
"0.3362",
"0.882",
"0.3499",
"0.5156",
"0.3499"
],
[
"0.5156",
"0.3516",
"0.881",
"0.3516",
"0.881",
"0.3652",
"0.5156",
"0.3652"
],
[
"0.1914",
"0.361",
"0.4626",
"0.3594",
"0.4626",
"0.3735",
"0.1914",
"0.376"
],
[
"0.461",
"0.3623",
"0.4878",
"0.3623",
"0.4878",
"0.3735",
"0.461",
"0.3735"
],
[
"0.2032",
"0.3723",
"0.2183",
"0.3723",
"0.2183",
"0.3813",
"0.2032",
"0.3813"
],
[
"0.5146",
"0.3677",
"0.882",
"0.3677",
"0.882",
"0.3813",
"0.5146",
"0.3813"
],
[
"0.1368",
"0.3884",
"0.4878",
"0.3884",
"0.4878",
"0.402",
"0.1368",
"0.402"
],
[
"0.5156",
"0.3843",
"0.882",
"0.3843",
"0.882",
"0.398",
"0.5156",
"0.398"
],
[
"0.12006",
"0.4038",
"0.4895",
"0.4038",
"0.4895",
"0.4175",
"0.12006",
"0.4175"
],
[
"0.513",
"0.3997",
"0.881",
"0.3997",
"0.881",
"0.4133",
"0.513",
"0.4133"
],
[
"0.12177",
"0.4204",
"0.4492",
"0.4204",
"0.4492",
"0.434",
"0.12177",
"0.434"
],
[
"0.5156",
"0.415",
"0.882",
"0.415",
"0.882",
"0.4287",
"0.5156",
"0.4287"
],
[
"0.5156",
"0.43",
"0.882",
"0.4312",
"0.882",
"0.4453",
"0.5156",
"0.4448"
],
[
"0.1184",
"0.4436",
"0.3645",
"0.4448",
"0.3645",
"0.459",
"0.1184",
"0.4578"
],
[
"0.5156",
"0.4478",
"0.852",
"0.4478",
"0.852",
"0.4614",
"0.5156",
"0.4614"
],
[
"0.1184",
"0.466",
"0.486",
"0.466",
"0.486",
"0.4797",
"0.1184",
"0.4797"
],
[
"0.534",
"0.4644",
"0.882",
"0.4644",
"0.882",
"0.478",
"0.534",
"0.478"
],
[
"0.51",
"0.4756",
"0.873",
"0.4785",
"0.873",
"0.4995",
"0.51",
"0.4963"
],
[
"0.12006",
"0.4822",
"0.486",
"0.4822",
"0.486",
"0.4954",
"0.12006",
"0.4954"
],
[
"0.12006",
"0.4988",
"0.486",
"0.4988",
"0.486",
"0.5127",
"0.12006",
"0.5127"
],
[
"0.636",
"0.5103",
"0.763",
"0.5083",
"0.763",
"0.527",
"0.636",
"0.5293"
],
[
"0.858",
"0.513",
"0.882",
"0.513",
"0.882",
"0.528",
"0.858",
"0.528"
],
[
"0.12006",
"0.53",
"0.486",
"0.53",
"0.486",
"0.5444",
"0.12006",
"0.5444"
],
[
"0.1318",
"0.514",
"0.4844",
"0.514",
"0.4844",
"0.528",
"0.1318",
"0.528"
],
[
"0.12006",
"0.5474",
"0.2124",
"0.5474",
"0.2124",
"0.5586",
"0.12006",
"0.5586"
],
[
"0.517",
"0.542",
"0.882",
"0.542",
"0.882",
"0.5557",
"0.517",
"0.5557"
],
[
"0.1368",
"0.5625",
"0.4878",
"0.5625",
"0.4878",
"0.5767",
"0.1368",
"0.5767"
],
[
"0.5146",
"0.557",
"0.881",
"0.557",
"0.881",
"0.572",
"0.5146",
"0.572"
],
[
"0.1184",
"0.5776",
"0.2603",
"0.5776",
"0.2603",
"0.593",
"0.1184",
"0.593"
],
[
"0.2477",
"0.5776",
"0.4902",
"0.5767",
"0.4902",
"0.594",
"0.2477",
"0.5957"
],
[
"0.5146",
"0.5737",
"0.882",
"0.5737",
"0.882",
"0.5874",
"0.5146",
"0.5874"
],
[
"0.1377",
"0.594",
"0.4878",
"0.594",
"0.4878",
"0.6094",
"0.1377",
"0.6094"
],
[
"0.5146",
"0.5903",
"0.636",
"0.5903",
"0.636",
"0.604",
"0.5146",
"0.604"
],
[
"0.1234",
"0.5986",
"0.1436",
"0.5986",
"0.1436",
"0.6055",
"0.1234",
"0.6055"
],
[
"0.12006",
"0.611",
"0.486",
"0.611",
"0.486",
"0.6245",
"0.12006",
"0.6245"
],
[
"0.5693",
"0.6177",
"0.8228",
"0.615",
"0.8228",
"0.6357",
"0.5693",
"0.638"
],
[
"0.1184",
"0.626",
"0.4836",
"0.626",
"0.4836",
"0.6396",
"0.1184",
"0.6396"
],
[
"0.858",
"0.632",
"0.881",
"0.632",
"0.881",
"0.6445",
"0.858",
"0.6445"
],
[
"0.12006",
"0.6436",
"0.2544",
"0.6436",
"0.2544",
"0.6567",
"0.12006",
"0.6567"
],
[
"0.608",
"0.6396",
"0.8237",
"0.64",
"0.8237",
"0.6577",
"0.608",
"0.6567"
],
[
"0.1293",
"0.6685",
"0.2275",
"0.6685",
"0.2275",
"0.6797",
"0.1293",
"0.6797"
],
[
"0.2393",
"0.6665",
"0.482",
"0.6665",
"0.482",
"0.681",
"0.2393",
"0.681"
],
[
"0.5156",
"0.6714",
"0.882",
"0.6714",
"0.882",
"0.685",
"0.5156",
"0.685"
],
[
"0.241",
"0.684",
"0.266",
"0.684",
"0.266",
"0.697",
"0.241",
"0.697"
],
[
"0.5146",
"0.686",
"0.881",
"0.686",
"0.881",
"0.6997",
"0.5146",
"0.6997"
],
[
"0.2377",
"0.7017",
"0.4634",
"0.7017",
"0.4634",
"0.7197",
"0.2377",
"0.7197"
],
[
"0.517",
"0.7036",
"0.882",
"0.7036",
"0.882",
"0.7173",
"0.517",
"0.7173"
],
[
"0.513",
"0.7188",
"0.6514",
"0.7197",
"0.6514",
"0.734",
"0.513",
"0.733"
],
[
"0.2124",
"0.706",
"0.2451",
"0.706",
"0.2451",
"0.7144",
"0.2124",
"0.7144"
],
[
"0.2393",
"0.72",
"0.2747",
"0.72",
"0.2747",
"0.733",
"0.2393",
"0.733"
],
[
"0.2393",
"0.7363",
"0.3132",
"0.7363",
"0.3132",
"0.7505",
"0.2393",
"0.7505"
],
[
"0.4626",
"0.7363",
"0.4878",
"0.7363",
"0.4878",
"0.7505",
"0.4626",
"0.7505"
],
[
"0.513",
"0.7446",
"0.663",
"0.745",
"0.663",
"0.7603",
"0.513",
"0.759"
],
[
"0.1377",
"0.7583",
"0.4878",
"0.7583",
"0.4878",
"0.7715",
"0.1377",
"0.7715"
],
[
"0.126",
"0.7656",
"0.1368",
"0.7656",
"0.1368",
"0.7686",
"0.126",
"0.7686"
],
[
"0.5146",
"0.7656",
"0.881",
"0.7656",
"0.881",
"0.7793",
"0.5146",
"0.7793"
],
[
"0.1184",
"0.7754",
"0.4878",
"0.774",
"0.4878",
"0.7876",
"0.1184",
"0.789"
],
[
"0.5146",
"0.783",
"0.882",
"0.783",
"0.882",
"0.7964",
"0.5146",
"0.7964"
],
[
"0.1377",
"0.7905",
"0.4878",
"0.7905",
"0.4878",
"0.804",
"0.1377",
"0.804"
],
[
"0.1184",
"0.804",
"0.4902",
"0.804",
"0.4902",
"0.822",
"0.1184",
"0.822"
],
[
"0.5156",
"0.799",
"0.882",
"0.799",
"0.882",
"0.8125",
"0.5156",
"0.8125"
],
[
"0.11584",
"0.8164",
"0.4902",
"0.818",
"0.4902",
"0.839",
"0.11584",
"0.8384"
],
[
"0.5156",
"0.8164",
"0.881",
"0.8164",
"0.881",
"0.8267",
"0.5156",
"0.8267"
],
[
"0.513",
"0.83",
"0.8867",
"0.8306",
"0.8867",
"0.8457",
"0.513",
"0.844"
],
[
"0.1184",
"0.8374",
"0.4878",
"0.8384",
"0.4878",
"0.8525",
"0.1184",
"0.8516"
],
[
"0.513",
"0.8457",
"0.8867",
"0.8467",
"0.8867",
"0.861",
"0.513",
"0.86"
],
[
"0.1411",
"0.86",
"0.465",
"0.86",
"0.465",
"0.8774",
"0.1411",
"0.8774"
],
[
"0.5146",
"0.8623",
"0.8833",
"0.8633",
"0.8833",
"0.8774",
"0.5146",
"0.877"
],
[
"0.2795",
"0.877",
"0.294",
"0.882",
"0.287",
"0.8926",
"0.272",
"0.887"
],
[
"0.5156",
"0.879",
"0.882",
"0.8774",
"0.882",
"0.8926",
"0.5156",
"0.8936"
],
[
"0.2377",
"0.8936",
"0.492",
"0.8936",
"0.492",
"0.9146",
"0.2377",
"0.9146"
],
[
"0.5156",
"0.8955",
"0.882",
"0.8955",
"0.882",
"0.909",
"0.5156",
"0.909"
]
] | [
"为了帮助神经网络模型进行更好的特征选",
"类似地，对于α下的I(Z;X），我们通过将",
"择，我们在单词嵌入之后添加了一个随机层R",
"",
"p(Z|X）替换为预定义的先验分布po(Z)",
"其中R=[Ra1....,RT]的长度与α相同。",
"",
"每个RrE30.1都是与单词类型t关联的",
"I(Z;X) ≤ Eg()[KL[q(zα)Ilpo(z)]]",
"二进制随机变量，而不是单词位置。这个随机",
"= KL[g(zα(@))po(z)], (5)",
"层和单词嵌入一起形成了神经网络模型的输",
"",
"人，即。e。,",
"来获得I(Z;X）的上限，其中KL[-]表示",
"z=Roa.",
"(1)",
"Kullback-Leibler散度。最后一步的简化类似",
"其中是逐元素乘法，每个Zt=Ratat。",
"于具有经验分布q(X()）的Equation4",
"将（5）和（4）代入Equation2，为我们",
"直观上：么只包含&的一个子集，该子集由",
"R随机选择。由于R作为0-1掩码序列直接",
"提供了信息瓶颈的下限L",
"应用于单词，因此我们在本工作中也将其称为",
"L =Eg(z/(0)[log p(y(°)]2)]",
"单词掩码层",
"(6)",
"为了确保Z具有足够的有关预测Y的信",
"- β . KL[q(zα())]po(z)].",
"息，同时包含来自的最少穴余信息，我们遵",
"循信息瓶颈理论（Tishbyetal.，2000）中的标",
"学习目标是使Equation6相对于近似分布",
"准实践，并将目标函数编写为",
"q(X,Y,Z)=q(X,Y)q(Z|X)最大化。作为",
"分类问题，X和Y都被观察到，q(X,Y）已",
"maxI(Z;Y)-β.I(Z;X),",
"(2)",
"Z",
"经被简化为经验分布，近似分布中唯一剩下的",
"其中X作为表示通用单词序列的随机变量",
"是q(Z|X）。与变分推理（Alemi et al.,2016;",
"作为输入，一个。这个公式反映了我们对Z的",
"RezendeandMohamed，2015）中的目标函数",
"确切期望。这里的主要挑战是计算互信息",
"类似，中的第一项是确保 q(ZIX)中用于预",
"测Y的信息，而L中的第二项是使用预定义",
"3.2Variational Word Masks",
"的先验分布Po(Z）对α(Z|X）进行正则化",
"受Alemietal.（2016）提出的变分信息瓶颈的",
"获得实用目标函数的最后一步是注意到，给",
"定x(=a(），每个Zt都可以重新定义为",
"启发，我们从近似分布g(X，Y，Z）开始，而",
"不是计算p(X,Y,Z）。然后，在下面指定的",
"Zt = Rrt *c(i),",
"(7)",
"标的可处理下界，并在Appendix A中提供了",
"-些假设下，我们在Equation2中构造了目",
"详细的推导",
"，其中RrtE[0,1]是标准的伯努利分布。然",
"对于q下的I(Z;Y)，我们有",
"后，可以使用R=[Rri...,Ra]将Z重新",
"I(Z;Y)=",
"Zy,z q(y, z) log(q(ylz) /q(y)",
"参数化为Z=R◎(）。下限可以用随机",
"通过将logq(yz)替换为从真实分布",
"变量 R重写为",
"",
"logp(ylz）得出的条件概率，我们从分布中",
"L =Eg(r|(0)[1og p(y()|R, e()]",
"引人了Y和Z之间的约束，并且还获得了",
"(8)",
"I(Z;Y)的下界,",
"-β . KL[q(R(i)]lpo(R)]",
"I(Z:Y) ≥",
"q(y, z) log p(ylz) + Hq(Y)",
"请注意，尽管β是从信息瓶颈理论继承而来",
"y,z",
"的，但实际上它将用作可调超参数来解决臭",
"Z q(c, y)g(z|α) log p(ylz)",
"名昭著的后验崩溃问题（Bowmanetal.，2016;",
"Kim et al., 2018)",
"",
"y,z,a",
"+H.(Y),",
"(3)",
"3.3Connections",
"其中H。(）是摘，最后一步使用基于",
"",
"使用信息瓶颈方法修改单词嵌入的想法最近",
"q(a,y,z)=q(α)q(yα)q(zα）的条件。给定",
"在NLP中显示了一些有趣的应用程序。例",
"个特定的观察值（α(),y()），我们将经",
"验分布 q(X(i),Y(i)）定义为两个Delta函数",
"如，Liand Eisner（2019）提出了两种将单词",
"q(X(i) = a(), Y() = g(0) = 8a(0(α) · 8g(0)(g)",
"嵌入转换为新表示形式的方法：以实现更好",
"的POS标记和句法分析。根据Equation1，",
"的乘积。然后，Equation3可以进一步简化为",
"Vmask可以看作是单词嵌入的简单线性变换。",
"I(z; Y(i) ≥ q(zc(0) logp(y(i)|z)",
"不同之处在于{R}是在词汇表上定义的，",
"2",
"因此可以用来表示at这个词的全局重要性。",
"= Eq(z|2(0)[1ogp(y()]z)]. (4)",
"回想一下RrtE[0,1]，从稍微不同的角度来"
] | {
"arxiv_id": "2010.00667",
"language": "zh",
"page_idx": 2
} |
||
2010.00667-zh-3 | 2010.00667-zh-3.jpg | <table>
<tbody>
<tr>
<th>Datasets</th>
<th>$C$ </th>
<th>L</th>
<th>$\#$ train = </th>
<th>$\#$dev</th>
<th>$\#$test</th>
</tr>
<tr>
<td>IMDB</td>
<td>2</td>
<td>268</td>
<td>20K</td>
<td>5K</td>
<td>25K</td>
</tr>
<tr>
<td>SST-1</td>
<td>5</td>
<td>18</td>
<td>8544</td>
<td>11101</td>
<td>2210</td>
</tr>
<tr>
<td>SST- 2</td>
<td>2</td>
<td>19</td>
<td>6920</td>
<td>872</td>
<td>1821</td>
</tr>
<tr>
<td>Yelp</td>
<td>2</td>
<td>138</td>
<td>500K</td>
<td>60K</td>
<td>38K</td>
</tr>
<tr>
<td>AG News</td>
<td>4</td>
<td>32</td>
<td>114K</td>
<td>6K</td>
<td>7.6K</td>
</tr>
<tr>
<td>TREC</td>
<td>6</td>
<td>10</td>
<td>5000</td>
<td>452</td>
<td>500</td>
</tr>
<tr>
<td>Subj</td>
<td>2</td>
<td>23</td>
<td>8000</td>
<td>1000</td>
<td>1000</td>
</tr>
</tbody>
</table>
Table 2: Summary statistics for the datasets, where C is the number of classes, L is average sentence length, and$\#$counts the number of examples in the train/dev/test sets.
看，Equation 1 可以被视为单词嵌入删除 (Gal and Ghahramani, 2016) 的通用方法。尽管有两个主要区别：(1)在Gal and Ghahramani (2016)中，所有单词都具有相同的辍学率，而在 Vmask 中，每个单词都具有$q(R_{x_t}|x_t)$指定的自己的辍学率，i。e。1-$\mathbb{E}[q(R_{x_t}|\boldsymbol{x}_t)];(2)$ 单词嵌入辍学的动机是迫使模型不依赖单个单词进行预测，而 Vmask 则是学习每个单词的特定任务重要性
使字掩码稀疏的另一种实现是通过添加$L_{0}$ 正则化 (Lei et al., 2016; Bastings et al., 2019; De Cao et al., 2020) ,而在目标 Equation 8中，我们使用预定义的先验分布$p_0(\bar{\boldsymbol{R}})$正则化掩码，如 subsection 3.4 中所述
3.4 Model Specification and Training
我们求助于平均场近似(Blei et al.,2017) 来简化对$q$分布的假设。对于$q_\phi(\boldsymbol{R}|x)$ ,我们有$q_{\boldsymbol{\phi }}( \boldsymbol{R}| \boldsymbol{x}) = \prod _{t= 1}^Tq_{\phi }( R_{x_t}| \boldsymbol{x}_t)$ ,这意味着随机变量是相互独立的，每个都由$x_t$控制。我们使用摊销变分推理 (Rezende and Mohamed, 2015) 来表示后验分布$q_\phi(R_{x_t}|\boldsymbol{x}_t)$ ,并使用推理网络(Kingma and Welling, 2014)。在这项工作中， 我们采用单层前馈神经网络作为推理网络，其参数$\phi$在训练期间用模型参数$\theta$进行了优化
遵循与$q_\phi(\boldsymbol{R}|x)$中相同的因式分解，我们将先验分布$p_0(\boldsymbol{R})$定义为$p_0(\boldsymbol{R})=\prod_t=1^Tp_0(R_{x_t})$ ,将它们中的每一个定义为$p_0(R_{x_t})=$ Bernoulli(0.5) 。通过选择这种非信息先验，这意味着每个单词都被初始化，没有优先考虑重要或不重要，因此被屏蔽或选择的概率相等。由于$p_0(\boldsymbol{R})$是均匀分布，我们可以进一步将Equation 8 中的第二项简化为条件熵，
推理阶段，该模型将单词嵌入$x_t$的乘积与单词掩码分布的期望相乘，即。e。$x_t\cdot\mathbb{E}[q(R_{x_t}|\boldsymbol{x}_t)]$ ,作为输入
# 4 Experiment Setup
所提出的方法在七个文本分类任务上进行了评估，从情感分析到主题分类，具有三个典型的神经网络模型，长短期记忆E$( Hochreiter and$ Schmidhuber, 1997, LSTM) , 卷积神经网络E (Kim, 2014, CNN) 和 BERT$\underline{\mathrm{E}}($Devlin et al., 2018)
Datasets. 我们采用七个基准数据集：电影评论 IMDB (Maas et al., 2011) ,带有细粒度标签SST-1 及其二进制版本 SST-2 (Socher et al., 2013) 的斯坦福情感树库，Yelp 评论 (Zhang et al., 2015) , AG 的 News (Zhang et al., 2015) , 6 类问题分类 TREC (Li and Roth, 2002) , 以及主观/。对于数据集 (e。g。IMDB, Subj) 没有标准的 train/dev/test 拆分，我们拿出一部分训练示例作为开发集。Table 2 显示数据集的统计信息
Models. CNN 模型 (Kim,2014)包含单个卷积层，滤波器大小范围为 3 到 5。LSTM (Hochreiter and Schmidhuber, 1997) 具有单一单向隐层。这两个模型都用 300 维预训练词嵌人初始化 (Mikolov et al.,2013) 。我们修复嵌人层并更新不同数据集上的其他参数，以分别实现最佳性能。我们使用预训练的BERT 基模型$^2$,该模型具有 12 个变压器层， 12 个自注意头和 768 的隐藏尺寸。我们使用不同的下游任务对其进行微调，然后将嵌入层固定并将遮罩层与模型的其余部分一起训练
${}^{2}$https://github.com/huggingface/
pytorch-transformers
$$\max_{\boldsymbol{\theta},\phi}\mathbb{E}_{q}[\log p(\boldsymbol{y}^{(i)}|\boldsymbol{R},\boldsymbol{x}^{(i)})]+\beta\cdot H_{q}(\boldsymbol{R}|\boldsymbol{x}^{(i)}).$$
(9)
我们应用随机梯度下降来解决优化问题(Equa-
tion 9 )。特别是在每次迭代中，Equation 9 中的第一项与$q(\boldsymbol{R}|\boldsymbol{x}^{(i)})$ (Kingma and Welling, 2014)中的单个样本近似。但是，从伯努利分布(如从任何其他离散分布)采样会导致反向传播困难。我们采用 Gumbel-softmax 技巧(Jang et al., 2016; Maddison et al., 2016) 来利用连续可微逼近并解决来自伯努利分布(Appendix B ) 的采样的离散性。在训练过程中， 我们使用 Adam (Kingma and Ba,2014)进行优化，并使用 KL 成本退火 (Bowman et al., 2016)来避免后塌陷
对于给定的单词$x_t$及其单词嵌入$x_t$ ,在训练阶段，模型对$q(R_x_t|\boldsymbol{x}_t)$中的每个$r_x_t$进行采样，以决定保留或清零相应的单词嵌入$x_t$ 。在 | [
[
"0.12006",
"0.0766",
"0.4878",
"0.0766",
"0.4878",
"0.0903",
"0.12006",
"0.0903"
],
[
"0.5225",
"0.079",
"0.5977",
"0.079",
"0.5977",
"0.0932",
"0.5225",
"0.0932"
],
[
"0.615",
"0.0766",
"0.6733",
"0.0766",
"0.6733",
"0.0956",
"0.615",
"0.0956"
],
[
"0.702",
"0.079",
"0.7676",
"0.079",
"0.7676",
"0.0944",
"0.702",
"0.0944"
],
[
"0.777",
"0.079",
"0.833",
"0.0766",
"0.8345",
"0.09265",
"0.7773",
"0.0944"
],
[
"0.8486",
"0.079",
"0.906",
"0.0766",
"0.9077",
"0.09265",
"0.8506",
"0.0944"
],
[
"0.1184",
"0.09143",
"0.4844",
"0.09265",
"0.4844",
"0.1069",
"0.1184",
"0.1057"
],
[
"0.5225",
"0.0986",
"0.5786",
"0.10095",
"0.5767",
"0.1164",
"0.5205",
"0.114"
],
[
"0.6274",
"0.0974",
"0.642",
"0.1081",
"0.6274",
"0.11816",
"0.613",
"0.1081"
],
[
"0.7793",
"0.10095",
"0.81",
"0.10095",
"0.81",
"0.11755",
"0.7793",
"0.11755"
],
[
"0.848",
"0.10095",
"0.888",
"0.10095",
"0.888",
"0.1164",
"0.848",
"0.1164"
],
[
"0.6987",
"0.1027",
"0.738",
"0.1027",
"0.738",
"0.1152",
"0.6987",
"0.1152"
],
[
"0.6514",
"0.1027",
"0.687",
"0.1027",
"0.687",
"0.1152",
"0.6514",
"0.1152"
],
[
"0.12006",
"0.10925",
"0.4878",
"0.10925",
"0.4878",
"0.1223",
"0.12006",
"0.1223"
],
[
"0.523",
"0.11755",
"0.575",
"0.11755",
"0.575",
"0.1318",
"0.523",
"0.1318"
],
[
"0.617",
"0.11755",
"0.636",
"0.11755",
"0.636",
"0.133",
"0.617",
"0.133"
],
[
"0.6987",
"0.11755",
"0.742",
"0.11755",
"0.742",
"0.133",
"0.6987",
"0.133"
],
[
"0.781",
"0.11755",
"0.8213",
"0.11755",
"0.8213",
"0.133",
"0.781",
"0.133"
],
[
"0.8486",
"0.11755",
"0.8926",
"0.11755",
"0.8926",
"0.133",
"0.8486",
"0.133"
],
[
"0.12177",
"0.1247",
"0.4895",
"0.1247",
"0.4895",
"0.1383",
"0.12177",
"0.1383"
],
[
"0.6514",
"0.11816",
"0.677",
"0.11816",
"0.677",
"0.1318",
"0.6514",
"0.1318"
],
[
"0.6987",
"0.133",
"0.742",
"0.133",
"0.742",
"0.1484",
"0.6987",
"0.1484"
],
[
"0.7773",
"0.133",
"0.8154",
"0.133",
"0.8154",
"0.1484",
"0.7773",
"0.1484"
],
[
"0.617",
"0.1342",
"0.636",
"0.1342",
"0.636",
"0.1484",
"0.617",
"0.1484"
],
[
"0.523",
"0.1342",
"0.5786",
"0.1342",
"0.5786",
"0.1484",
"0.523",
"0.1484"
],
[
"0.6514",
"0.1342",
"0.677",
"0.1342",
"0.677",
"0.1478",
"0.6514",
"0.1478"
],
[
"0.12006",
"0.1401",
"0.486",
"0.1401",
"0.486",
"0.1538",
"0.12006",
"0.1538"
],
[
"0.8506",
"0.1354",
"0.889",
"0.1354",
"0.889",
"0.1478",
"0.8506",
"0.1478"
],
[
"0.523",
"0.1478",
"0.565",
"0.1497",
"0.5635",
"0.165",
"0.5225",
"0.1633"
],
[
"0.6514",
"0.1484",
"0.687",
"0.1484",
"0.687",
"0.1644",
"0.6514",
"0.1644"
],
[
"0.7793",
"0.1484",
"0.818",
"0.1484",
"0.818",
"0.1644",
"0.7793",
"0.1644"
],
[
"0.8486",
"0.1484",
"0.888",
"0.1484",
"0.888",
"0.1644",
"0.8486",
"0.1644"
],
[
"0.6978",
"0.1497",
"0.7466",
"0.1497",
"0.7466",
"0.1644",
"0.6978",
"0.1644"
],
[
"0.614",
"0.1526",
"0.629",
"0.1467",
"0.6396",
"0.158",
"0.623",
"0.165"
],
[
"0.6987",
"0.165",
"0.7466",
"0.165",
"0.7466",
"0.1812",
"0.6987",
"0.1812"
],
[
"0.1184",
"0.1561",
"0.486",
"0.1561",
"0.486",
"0.1692",
"0.1184",
"0.1692"
],
[
"0.6514",
"0.1663",
"0.6753",
"0.1663",
"0.6753",
"0.1799",
"0.6514",
"0.1799"
],
[
"0.781",
"0.1663",
"0.8076",
"0.1663",
"0.8076",
"0.1799",
"0.781",
"0.1799"
],
[
"0.8486",
"0.165",
"0.8906",
"0.165",
"0.8906",
"0.1812",
"0.8486",
"0.1812"
],
[
"0.6187",
"0.1675",
"0.6323",
"0.1675",
"0.6323",
"0.1787",
"0.6187",
"0.1787"
],
[
"0.12006",
"0.1716",
"0.486",
"0.1716",
"0.486",
"0.1853",
"0.12006",
"0.1853"
],
[
"0.525",
"0.1675",
"0.6035",
"0.1675",
"0.6035",
"0.1787",
"0.525",
"0.1787"
],
[
"0.523",
"0.1818",
"0.581",
"0.1818",
"0.581",
"0.1965",
"0.523",
"0.1965"
],
[
"0.617",
"0.1818",
"0.635",
"0.1818",
"0.635",
"0.1965",
"0.617",
"0.1965"
],
[
"0.6533",
"0.1818",
"0.6753",
"0.1818",
"0.6753",
"0.1953",
"0.6533",
"0.1953"
],
[
"0.777",
"0.1818",
"0.81",
"0.1787",
"0.8154",
"0.1965",
"0.781",
"0.1996"
],
[
"0.8486",
"0.1812",
"0.885",
"0.1812",
"0.885",
"0.1978",
"0.8486",
"0.1978"
],
[
"0.12006",
"0.1882",
"0.4878",
"0.1882",
"0.4878",
"0.2031",
"0.12006",
"0.2031"
],
[
"0.7",
"0.1829",
"0.742",
"0.1829",
"0.742",
"0.1965",
"0.7",
"0.1965"
],
[
"0.5225",
"0.1978",
"0.567",
"0.1978",
"0.567",
"0.2131",
"0.5225",
"0.2131"
],
[
"0.621",
"0.1953",
"0.6406",
"0.2031",
"0.6323",
"0.2131",
"0.613",
"0.2073"
],
[
"0.6514",
"0.199",
"0.677",
"0.199",
"0.677",
"0.2114",
"0.6514",
"0.2114"
],
[
"0.6987",
"0.1978",
"0.742",
"0.1978",
"0.742",
"0.2131",
"0.6987",
"0.2131"
],
[
"0.781",
"0.199",
"0.8228",
"0.199",
"0.8228",
"0.212",
"0.781",
"0.212"
],
[
"0.852",
"0.199",
"0.8906",
"0.199",
"0.8906",
"0.2114",
"0.852",
"0.2114"
],
[
"0.1184",
"0.206",
"0.2494",
"0.206",
"0.2494",
"0.2197",
"0.1184",
"0.2197"
],
[
"0.1377",
"0.2227",
"0.4878",
"0.2227",
"0.4878",
"0.2363",
"0.1377",
"0.2363"
],
[
"0.513",
"0.2286",
"0.596",
"0.2286",
"0.596",
"0.2405",
"0.513",
"0.2405"
],
[
"0.5903",
"0.2286",
"0.8833",
"0.2286",
"0.8833",
"0.2423",
"0.5903",
"0.2423"
],
[
"0.12006",
"0.2382",
"0.4895",
"0.2382",
"0.4895",
"0.2517",
"0.12006",
"0.2517"
],
[
"0.511",
"0.2411",
"0.882",
"0.2434",
"0.882",
"0.257",
"0.511",
"0.2546"
],
[
"0.1184",
"0.2542",
"0.4895",
"0.2542",
"0.4895",
"0.2683",
"0.1184",
"0.2683"
],
[
"0.513",
"0.257",
"0.882",
"0.257",
"0.882",
"0.2708",
"0.513",
"0.2708"
],
[
"0.1184",
"0.2683",
"0.4895",
"0.2695",
"0.4895",
"0.284",
"0.1184",
"0.2832"
],
[
"0.513",
"0.2715",
"0.705",
"0.2715",
"0.705",
"0.2852",
"0.513",
"0.2852"
],
[
"0.12006",
"0.2861",
"0.3745",
"0.2861",
"0.3745",
"0.3005",
"0.12006",
"0.3005"
],
[
"0.11755",
"0.3123",
"0.4375",
"0.3142",
"0.4375",
"0.332",
"0.11755",
"0.33"
],
[
"0.5146",
"0.3154",
"0.882",
"0.3154",
"0.882",
"0.3289",
"0.5146",
"0.3289"
],
[
"0.12006",
"0.3362",
"0.486",
"0.3362",
"0.486",
"0.3499",
"0.12006",
"0.3499"
],
[
"0.5146",
"0.332",
"0.882",
"0.332",
"0.882",
"0.3457",
"0.5146",
"0.3457"
],
[
"0.513",
"0.3474",
"0.6094",
"0.3474",
"0.6094",
"0.3623",
"0.513",
"0.3623"
],
[
"0.12177",
"0.354",
"0.4878",
"0.354",
"0.4878",
"0.3677",
"0.12177",
"0.3677"
],
[
"0.11755",
"0.3694",
"0.4895",
"0.3682",
"0.4895",
"0.386",
"0.11755",
"0.3872"
],
[
"0.5156",
"0.383",
"0.5483",
"0.383",
"0.5483",
"0.3901",
"0.5156",
"0.3901"
],
[
"0.5483",
"0.3801",
"0.708",
"0.3801",
"0.708",
"0.3943",
"0.5483",
"0.3943"
],
[
"0.12177",
"0.3872",
"0.486",
"0.3872",
"0.486",
"0.401",
"0.12177",
"0.401"
],
[
"0.1184",
"0.402",
"0.4878",
"0.4026",
"0.4878",
"0.4175",
"0.1184",
"0.4163"
],
[
"0.5156",
"0.4092",
"0.882",
"0.4092",
"0.882",
"0.4229",
"0.5156",
"0.4229"
],
[
"0.1184",
"0.4192",
"0.486",
"0.4192",
"0.486",
"0.4329",
"0.1184",
"0.4329"
],
[
"0.5156",
"0.4246",
"0.881",
"0.4246",
"0.881",
"0.4382",
"0.5156",
"0.4382"
],
[
"0.12006",
"0.4353",
"0.4895",
"0.4353",
"0.4895",
"0.449",
"0.12006",
"0.449"
],
[
"0.5146",
"0.4407",
"0.882",
"0.4407",
"0.882",
"0.4536",
"0.5146",
"0.4536"
],
[
"0.12006",
"0.452",
"0.4895",
"0.452",
"0.4895",
"0.4656",
"0.12006",
"0.4656"
],
[
"0.513",
"0.4573",
"0.8833",
"0.4573",
"0.8833",
"0.4702",
"0.513",
"0.4702"
],
[
"0.1184",
"0.466",
"0.482",
"0.4673",
"0.482",
"0.4822",
"0.1184",
"0.481"
],
[
"0.513",
"0.4714",
"0.8833",
"0.4727",
"0.8833",
"0.4868",
"0.513",
"0.4863"
],
[
"0.1368",
"0.4827",
"0.486",
"0.484",
"0.486",
"0.4988",
"0.1368",
"0.4976"
],
[
"0.513",
"0.488",
"0.561",
"0.488",
"0.561",
"0.503",
"0.513",
"0.503"
],
[
"0.12006",
"0.502",
"0.4878",
"0.502",
"0.4878",
"0.5166",
"0.12006",
"0.5166"
],
[
"0.1427",
"0.5186",
"0.4592",
"0.5186",
"0.4592",
"0.532",
"0.1427",
"0.532"
],
[
"0.4685",
"0.5225",
"0.4836",
"0.5225",
"0.4836",
"0.527",
"0.4685",
"0.527"
],
[
"0.517",
"0.517",
"0.882",
"0.5195",
"0.882",
"0.5337",
"0.517",
"0.532"
],
[
"0.1184",
"0.533",
"0.486",
"0.533",
"0.486",
"0.5464",
"0.1184",
"0.5464"
],
[
"0.5156",
"0.535",
"0.882",
"0.535",
"0.882",
"0.549",
"0.5156",
"0.549"
],
[
"0.12006",
"0.55",
"0.486",
"0.55",
"0.486",
"0.5635",
"0.12006",
"0.5635"
],
[
"0.5146",
"0.5503",
"0.881",
"0.5503",
"0.881",
"0.564",
"0.5146",
"0.564"
],
[
"0.12177",
"0.5664",
"0.4895",
"0.5664",
"0.4895",
"0.58",
"0.12177",
"0.58"
],
[
"0.513",
"0.5664",
"0.882",
"0.568",
"0.882",
"0.582",
"0.513",
"0.5806"
],
[
"0.12006",
"0.582",
"0.486",
"0.582",
"0.486",
"0.5957",
"0.12006",
"0.5957"
],
[
"0.5156",
"0.585",
"0.881",
"0.585",
"0.881",
"0.5957",
"0.5156",
"0.5957"
],
[
"0.1184",
"0.5986",
"0.4224",
"0.5986",
"0.4224",
"0.6123",
"0.1184",
"0.6123"
],
[
"0.517",
"0.5996",
"0.8716",
"0.5996",
"0.8716",
"0.6133",
"0.517",
"0.6133"
],
[
"0.8657",
"0.6055",
"0.8774",
"0.604",
"0.879",
"0.6094",
"0.867",
"0.611"
],
[
"0.517",
"0.615",
"0.8833",
"0.615",
"0.8833",
"0.63",
"0.517",
"0.63"
],
[
"0.1293",
"0.629",
"0.4727",
"0.6274",
"0.4727",
"0.6455",
"0.1293",
"0.6465"
],
[
"0.5156",
"0.632",
"0.8696",
"0.632",
"0.8696",
"0.6455",
"0.5156",
"0.6455"
],
[
"0.1351",
"0.6426",
"0.1628",
"0.6426",
"0.1628",
"0.654",
"0.1351",
"0.654"
],
[
"0.4634",
"0.653",
"0.4878",
"0.653",
"0.4878",
"0.6665",
"0.4634",
"0.6665"
],
[
"0.5156",
"0.648",
"0.882",
"0.648",
"0.882",
"0.661",
"0.5156",
"0.661"
],
[
"0.5146",
"0.6646",
"0.6265",
"0.6646",
"0.6265",
"0.6787",
"0.5146",
"0.6787"
],
[
"0.1234",
"0.6704",
"0.4844",
"0.6704",
"0.4844",
"0.681",
"0.1234",
"0.681"
],
[
"0.12006",
"0.685",
"0.486",
"0.685",
"0.486",
"0.6987",
"0.12006",
"0.6987"
],
[
"0.525",
"0.6924",
"0.882",
"0.6934",
"0.882",
"0.708",
"0.525",
"0.707"
],
[
"0.1184",
"0.6997",
"0.4878",
"0.7007",
"0.4878",
"0.7153",
"0.1184",
"0.7144"
],
[
"0.5156",
"0.7104",
"0.8833",
"0.7104",
"0.8833",
"0.7236",
"0.5156",
"0.7236"
],
[
"0.1184",
"0.7163",
"0.486",
"0.7163",
"0.486",
"0.73",
"0.1184",
"0.73"
],
[
"0.5146",
"0.7256",
"0.882",
"0.7256",
"0.882",
"0.7393",
"0.5146",
"0.7393"
],
[
"0.12177",
"0.734",
"0.4878",
"0.734",
"0.4878",
"0.7476",
"0.12177",
"0.7476"
],
[
"0.513",
"0.741",
"0.882",
"0.741",
"0.882",
"0.755",
"0.513",
"0.755"
],
[
"0.12177",
"0.75",
"0.4878",
"0.75",
"0.4878",
"0.763",
"0.12177",
"0.763"
],
[
"0.5146",
"0.7573",
"0.8833",
"0.7573",
"0.8833",
"0.771",
"0.5146",
"0.771"
],
[
"0.12177",
"0.7666",
"0.486",
"0.7666",
"0.486",
"0.78",
"0.12177",
"0.78"
],
[
"0.513",
"0.7725",
"0.882",
"0.774",
"0.882",
"0.788",
"0.513",
"0.7876"
],
[
"0.12006",
"0.781",
"0.486",
"0.781",
"0.486",
"0.7944",
"0.12006",
"0.7944"
],
[
"0.5146",
"0.7905",
"0.882",
"0.7905",
"0.882",
"0.804",
"0.5146",
"0.804"
],
[
"0.12006",
"0.7974",
"0.4878",
"0.7974",
"0.4878",
"0.811",
"0.12006",
"0.811"
],
[
"0.513",
"0.804",
"0.881",
"0.8057",
"0.881",
"0.821",
"0.513",
"0.8184"
],
[
"0.12177",
"0.8135",
"0.4878",
"0.8135",
"0.4878",
"0.8276",
"0.12177",
"0.8276"
],
[
"0.5146",
"0.8223",
"0.881",
"0.8223",
"0.881",
"0.836",
"0.5146",
"0.836"
],
[
"0.12177",
"0.83",
"0.4895",
"0.83",
"0.4895",
"0.844",
"0.12177",
"0.844"
],
[
"0.5146",
"0.8374",
"0.882",
"0.8384",
"0.882",
"0.8525",
"0.5146",
"0.8516"
],
[
"0.12006",
"0.8457",
"0.2805",
"0.8457",
"0.2805",
"0.86",
"0.12006",
"0.86"
],
[
"0.5146",
"0.855",
"0.8657",
"0.854",
"0.8657",
"0.868",
"0.5146",
"0.869"
],
[
"0.1377",
"0.8623",
"0.486",
"0.8623",
"0.486",
"0.877",
"0.1377",
"0.877"
],
[
"0.12006",
"0.8774",
"0.4878",
"0.879",
"0.4878",
"0.8936",
"0.12006",
"0.8926"
],
[
"0.532",
"0.882",
"0.769",
"0.883",
"0.769",
"0.898",
"0.532",
"0.8965"
],
[
"0.12006",
"0.8955",
"0.4895",
"0.8955",
"0.4895",
"0.909",
"0.12006",
"0.909"
],
[
"0.513",
"0.8965",
"0.659",
"0.8965",
"0.659",
"0.91",
"0.513",
"0.91"
]
] | [
"看，Equation1可以被视为单词嵌人删除（Gal",
"Datasets",
"C L",
"# train",
"#dev",
"# test",
"andGhahramani，2016）的通用方法。尽管有",
"IMDB",
"2",
"5K",
"25K",
"20K",
"268",
"两个主要区别：（1）在GalandGhahramani",
"SST-1",
"5",
"8544",
"1101",
"2210",
"(2016）中，所有单词都具有相同的学率，而",
"18",
"6920",
"872",
"2",
"SST-2",
"19",
"在Vmask中，每个单词都具有g(Rztlact）指定",
"1821",
"Yelp",
"138",
"60K",
"38K",
"500K",
"2",
"114K",
"的自己的较学率，i。e。1－E[g(Rzat)；（2)",
"32",
"6K",
"7.6K",
"4",
"单词嵌入较学的动机是迫使模型不依赖单个单",
"AG News",
"TREC",
"6",
"10",
"452",
"500",
"词进行预测，而Vmask则是学习每个单词的",
"5000",
"Subj",
"2",
"23",
"8000",
"1000",
"1000",
"特定任务重要性",
"使字掩码稀疏的另一种实现是通过添加Lo",
"Table 2:",
"Summary statistics for the datasets,",
"正则化(Lei et al.,2016;Bastings et al.,2019;",
"where C is the number of classes, L is average sen",
"DeCaoetal.，2020），而在目标Equation8",
"tence length, and # counts the number of examples",
"中，我们使用预定义的先验分布po(R)正则化",
"in the train/dev/test sets.",
"掩码，如subsection3.4中所述",
"3.4Model Specification and Training",
"推理阶段，该模型将单词嵌入t的乘积与单",
"我们求助于平均场近似（Bleietal.2017）来",
"词掩码分布的期望相乘，即。e。at·E[q(Rrtat)]",
"，作为输入",
"简化对q分布的假设。对于q(Rα），我们有",
"q（Ra）=II=19（Rat），这意味着随机变",
"4",
"Experiment Setup",
"量是相互独立的，每个都由αt控制。我们使用",
"摊销变分推理（RezendeandMohamed，2015)",
"所提出的方法在七个文本分类任务上进行了评",
"来表示后验分布9d(Ratact），并使用推理网络",
"估，从情感分析到主题分类，具有三个典型的",
"(KingmaandWelling，2014）。在这项工作中,",
"神经网络模型，长短期记忆E(Hochreiterand",
"我们采用单层前馈神经网络作为推理网络，其",
"Schmidhuber，1997，LSTM），卷积神经网络E",
"参数Φ在训练期间用模型参数θ进行了优化",
"(Kim, 2014, CNN)和 BERTE(Devlin et al.,",
"遵循与9s(Rα）中相同的因式分解，我们将",
"2018)",
"先验分布po(R)定义为po(R)=IIt=1Po(Rat）",
"将它们中的每一个定义为po(Rat)",
"",
"Datasets．我们采用七个基准数据集：电影评",
"Bernoulli(0.5）。通过选择这种非信息先验，这",
"论IMDB（Maasetal.,2011），带有细粒度标签",
"意味着每个单词都被初始化，没有优先考虑重",
"SST-1及其二进制版本SST-2（Socher et al.",
"要或不重要，因此被屏蔽或选择的概率相等。",
"2013）的斯坦福情感树库，Yelp评论（Zhang",
"由于po(R）是均匀分布，我们可以进一步将",
"etal.,2015)AG的News(Zhang et al.,2015)",
"Equation8中的第二项简化为条件摘，",
"，6类问题分类TREC(LiandRoth，2002)",
"",
"以及主观/。对于数据集(e。g。IMDB，Subi)",
"max Eg[1og p(y()|R, c() + β . Hg(Rc(i)",
"没有标准的train/dev/test拆分，我们拿出",
"0,",
"(9)",
"部分训练示例作为开发集。Table2显示数据",
"集的统计信息",
"我们应用随机梯度下降来解决优化问题（Equa",
"tion9）。特别是在每次迭代中，Equation9中",
"Models．CNN模型(Kim，2014）包含",
"的第一项与 q(Ra(i)）(Kingma and Welling,",
"单个卷积层，滤波器大小范围为3到5。",
"2014）中的单个样本近似。但是，从伯努利分",
"LSTM (Hochreiter and Schmidhuber, 1997)",
"布（如从任何其他离散分布）采样会导致反",
"具有单一单向隐层。这两个模型都用300维",
"向传播困难。我们采用Gumbel-softmax技巧",
"预训练词嵌人初始化（Mikolovetal.，2013）。",
"(Jang et al.,2016; Maddison et al.,2016)来利",
"我们修复嵌入层并更新不同数据集上的其他参",
"用连续可微逼近并解决来自伯努利分布（Ap",
"数，以分别实现最佳性能。我们使用预训练的",
"pendixB）的采样的离散性。在训练过程中，",
"BERT基模型2，该模型具有12个变压器层，",
"我们使用Adam（KingmaandBa，2014）进行",
"12个自注意头和768的隐藏尺寸。我们使用",
"优化，并使用KL成本退火（Bowmanetal.,",
"不同的下游任务对其进行微调，然后将嵌入层",
"2016）来避免后塌陷",
"固定并将遮罩层与模型的其余部分一起训练",
"对于给定的单词at及其单词嵌入at，在训",
"练阶段，模型对q(Ratct）中的每个rat进行采",
"2https://github.com/huggingface/",
"样，以决定保留或清零相应的单词嵌入at。在",
"pytorch-transformers"
] | {
"arxiv_id": "2010.00667",
"language": "zh",
"page_idx": 3
} |
||
2010.00667-zh-4 | 2010.00667-zh-4.jpg | <table>
<tbody>
<tr>
<th>Models</th>
<th>Methods</th>
<th>IMDB</th>
<th>$SST-1$</th>
<th>SST- 2</th>
<th>Yelp</th>
<th>AG News - 、</th>
<th>TREC</th>
<th>Subj</th>
</tr>
<tr>
<td rowspan="5">$CNN$</td>
<td>CNN- base</td>
<td>89.06</td>
<td>46.32</td>
<td>85.50</td>
<td>94.32</td>
<td>91.30</td>
<td>92.40</td>
<td>92.80</td>
</tr>
<tr>
<td>CNN- $\ell_{2}$ </td>
<td>89.12</td>
<td>46.01</td>
<td>85.56</td>
<td>94.46</td>
<td>91.28</td>
<td>90.62</td>
<td>92.39</td>
</tr>
<tr>
<td>CNN- L2X</td>
<td>78.94</td>
<td>37.92</td>
<td>80.01</td>
<td>83.14</td>
<td>84.36</td>
<td>61.00</td>
<td>82.40</td>
</tr>
<tr>
<td>CNN- IBA</td>
<td>88.31</td>
<td>41.40</td>
<td>84.24</td>
<td>93.82</td>
<td>91.37</td>
<td>89.80</td>
<td>91.80</td>
</tr>
<tr>
<td>CNN- Vmask</td>
<td>90.10</td>
<td>48.92</td>
<td>85.78</td>
<td>94.53</td>
<td>91.60</td>
<td>93.02</td>
<td>93.50</td>
</tr>
<tr>
<td rowspan="5">LSTM</td>
<td>LSTM- base</td>
<td>88.39</td>
<td>43.84</td>
<td>83.74</td>
<td>95.06</td>
<td>91.03</td>
<td>90.40</td>
<td>90.20</td>
</tr>
<tr>
<td>LSTM- $\ell _{2}$</td>
<td>88.40</td>
<td>43.91</td>
<td>83.36</td>
<td>95.00</td>
<td>91.09</td>
<td>90.20</td>
<td>89.10</td>
</tr>
<tr>
<td>LSTM- L2X</td>
<td>67.45</td>
<td>36.92</td>
<td>75.45</td>
<td>77.12</td>
<td>77.53</td>
<td>46.00</td>
<td>81.80</td>
</tr>
<tr>
<td>LSTM- IBA</td>
<td>88.48</td>
<td>42.99</td>
<td>83.53</td>
<td>94.74</td>
<td>91.14</td>
<td>85.40</td>
<td>89.50</td>
</tr>
<tr>
<td>LSTM- Vmask</td>
<td>90.07</td>
<td>44.12</td>
<td>84.35</td>
<td>95.41</td>
<td>92.19</td>
<td>90.80</td>
<td>91.20</td>
</tr>
<tr>
<td rowspan="5">BERT</td>
<td>BERT- base</td>
<td>91.80</td>
<td>53.43</td>
<td>92.25</td>
<td>96.42</td>
<td>93.59</td>
<td>96.40</td>
<td>95.10</td>
</tr>
<tr>
<td>BERT- $\ell _{2}$</td>
<td>91.75</td>
<td>52.08</td>
<td>92.25</td>
<td>96.41</td>
<td>93.52</td>
<td>96.80</td>
<td>94.80</td>
</tr>
<tr>
<td>BERT- L2X</td>
<td>71.75</td>
<td>39.23</td>
<td>74.03</td>
<td>87.14</td>
<td>82.59</td>
<td>93.20</td>
<td>86.10</td>
</tr>
<tr>
<td>BERT- IBA</td>
<td>91.66</td>
<td>53.80</td>
<td>92.24</td>
<td>96.27</td>
<td>93.45</td>
<td>96.80</td>
<td>95.60</td>
</tr>
<tr>
<td>BERT- Vmask</td>
<td>93.04</td>
<td>54.53</td>
<td>92.26</td>
<td>96.80</td>
<td>94.24</td>
<td>97.00</td>
<td>96.40</td>
</tr>
</tbody>
</table>
Table 3: Prediction accuracy $(\%)$ of different models with different training strategies on the seven
datasets.
模型都具有相似的预测性能。使用 Vmask 训练的模型优于具有类似网络架构但训练方式不同的模型。结果表明，Vmask 可以帮助提高泛化能力
除了基本模型和用所提出的方法训练的模型外，其他三个竞争对手的记录是混合的。例如，传统的$\ell_{2}$ -正则化不能总是有助于提高准确性，特别是对于 BERT 模型。尽管 IBA 的性能略优于 L2X, 但使用它们进行训练并不能持续提高模型的预测精度
为了呼应提高模型可解释性的目的，本节的其余部分将集中于定量和定性地评估模型可解释性
Baselines and Competitive Methods. 由于这项工作的目标是提出一种新颖的训练方法，以提高预测准确性和可解释性，因此我们采用两组模型作为基线和竞争系统。使用所提出的方法训练的模型以后缀“- Vmask”命名。我们还提供了两个基线：(1)通过最小化交叉熵损失训练的模型(后缀为“-base”)和(2)用$\ell_2$ -正则化训练的模型(后缀为“-$\ell_{2}”)$。与这两种基准方法的比较主要集中在预测性能上，因为没有使用明确的训练策略来提高可解释性
此外，我们还提出了两种竞争方法：使用解释框架“学习解释”(Chen et al.,2018)(后缀为“-L2X”)和“信息瓶颈归因”(Schulz et al., 2020) (后缀为“-IBA”)训练的模型。L2X 和 IBA 最初是为了寻找特征属性作为训练有素的模型的事后解释而提出的。我们将它们集成在模型训练中，作为掩码层直接为输入特征(L2X)生成掩码值，或通过添加噪声(IBA) 限制信息流。在我们的实验中，所有的训练方法都使用随机退出( $\rho=0.2$ )来避免过度拟合
有关实验设置的更多详细信息在 Appendix C 中，包括数据预处理，模型配置以及我们实验中 L2X 和 IBA 的实现
5.1 Quantitative Evaluatiom
我们通过 AOPC 得分 (Nguyen, 2018; Samek et al.,2016)评估基于 Vmask 的模型相对于基本模型的局部可解释性，并通过事后准确性(Chen et al.,2018)评估基于 IBA 的模型的全局可解释性。从经验上讲，我们观察到了局部和全局可解释性之间的一致性，因此在我们的评估中无需穷尽所有可能的组合
5.1.1 Local interpretability: AOPC
## 5 Results and Discussiom
我们采用两种与模型无关的解释方法， LIME (Ribeiro et al.,2016)和 SampleShapley (Kononenko et al., 2010) , 为基础模型和基于 Vmask 的模型生成局部解释，其中“局部”表示分别解释每个测试数据。扰动曲线上的面积(AOPC)E(Nguyen,2018;Samek
我们使用不同的训练策略在七个数据集上训练了三个模型。Table 3 显示了测试集上不同模型的预测精度。验证性能和平均运行时间在Appendix D 中。如 Table 3 所示 , 与先前工作(Appendix E)中报告的数字相比，所有基础 | [
[
"0.1545",
"0.079",
"0.215",
"0.079",
"0.215",
"0.0932",
"0.1545",
"0.0932"
],
[
"0.2527",
"0.079",
"0.3267",
"0.079",
"0.3267",
"0.0932",
"0.2527",
"0.0932"
],
[
"0.4324",
"0.079",
"0.486",
"0.079",
"0.486",
"0.0932",
"0.4324",
"0.0932"
],
[
"0.5015",
"0.079",
"0.5576",
"0.079",
"0.5576",
"0.0932",
"0.5015",
"0.0932"
],
[
"0.5728",
"0.079",
"0.617",
"0.079",
"0.617",
"0.0944",
"0.5728",
"0.0944"
],
[
"0.634",
"0.079",
"0.7153",
"0.079",
"0.7153",
"0.0932",
"0.634",
"0.0932"
],
[
"0.729",
"0.079",
"0.7866",
"0.079",
"0.7866",
"0.0932",
"0.729",
"0.0932"
],
[
"0.8037",
"0.079",
"0.848",
"0.079",
"0.848",
"0.0944",
"0.8037",
"0.0944"
],
[
"0.2477",
"0.10156",
"0.3333",
"0.10156",
"0.3333",
"0.1164",
"0.2477",
"0.1164"
],
[
"0.3645",
"0.08014",
"0.4182",
"0.08014",
"0.4182",
"0.09143",
"0.3645",
"0.09143"
],
[
"0.3652",
"0.10156",
"0.4148",
"0.10156",
"0.4148",
"0.1164",
"0.3652",
"0.1164"
],
[
"0.4375",
"0.10156",
"0.4844",
"0.10156",
"0.4844",
"0.1164",
"0.4375",
"0.1164"
],
[
"0.504",
"0.10156",
"0.553",
"0.10156",
"0.553",
"0.1164",
"0.504",
"0.1164"
],
[
"0.5693",
"0.10156",
"0.6187",
"0.10156",
"0.6187",
"0.1164",
"0.5693",
"0.1164"
],
[
"0.6484",
"0.10156",
"0.6978",
"0.10156",
"0.6978",
"0.1164",
"0.6484",
"0.1164"
],
[
"0.735",
"0.10156",
"0.7827",
"0.10156",
"0.7827",
"0.1164",
"0.735",
"0.1164"
],
[
"0.8003",
"0.10156",
"0.8486",
"0.10156",
"0.8486",
"0.1164",
"0.8003",
"0.1164"
],
[
"0.2527",
"0.1152",
"0.3267",
"0.11755",
"0.3267",
"0.133",
"0.2527",
"0.1312"
],
[
"0.367",
"0.11755",
"0.4148",
"0.11755",
"0.4148",
"0.133",
"0.367",
"0.133"
],
[
"0.4375",
"0.11755",
"0.4836",
"0.11755",
"0.4836",
"0.133",
"0.4375",
"0.133"
],
[
"0.504",
"0.11755",
"0.553",
"0.11755",
"0.553",
"0.133",
"0.504",
"0.133"
],
[
"0.5693",
"0.11755",
"0.6187",
"0.11755",
"0.6187",
"0.133",
"0.5693",
"0.133"
],
[
"0.6484",
"0.11755",
"0.6978",
"0.11755",
"0.6978",
"0.133",
"0.6484",
"0.133"
],
[
"0.735",
"0.11755",
"0.781",
"0.11755",
"0.781",
"0.1318",
"0.735",
"0.1318"
],
[
"0.8003",
"0.11755",
"0.8486",
"0.11755",
"0.8486",
"0.1318",
"0.8003",
"0.1318"
],
[
"0.1628",
"0.133",
"0.2083",
"0.133",
"0.2083",
"0.1484",
"0.1628",
"0.1484"
],
[
"0.367",
"0.133",
"0.4148",
"0.133",
"0.4148",
"0.1484",
"0.367",
"0.1484"
],
[
"0.4358",
"0.133",
"0.4836",
"0.133",
"0.4836",
"0.1484",
"0.4358",
"0.1484"
],
[
"0.504",
"0.133",
"0.552",
"0.133",
"0.552",
"0.1484",
"0.504",
"0.1484"
],
[
"0.5693",
"0.1342",
"0.6187",
"0.1342",
"0.6187",
"0.1484",
"0.5693",
"0.1484"
],
[
"0.6484",
"0.1342",
"0.6978",
"0.1342",
"0.6978",
"0.1484",
"0.6484",
"0.1484"
],
[
"0.735",
"0.1342",
"0.781",
"0.1342",
"0.781",
"0.1484",
"0.735",
"0.1484"
],
[
"0.8003",
"0.1342",
"0.8486",
"0.1342",
"0.8486",
"0.1484",
"0.8003",
"0.1484"
],
[
"0.367",
"0.1497",
"0.413",
"0.1497",
"0.413",
"0.1644",
"0.367",
"0.1644"
],
[
"0.2477",
"0.1342",
"0.3315",
"0.1342",
"0.3315",
"0.1484",
"0.2477",
"0.1484"
],
[
"0.4358",
"0.1497",
"0.4844",
"0.1497",
"0.4844",
"0.1644",
"0.4358",
"0.1644"
],
[
"0.504",
"0.1497",
"0.553",
"0.1497",
"0.553",
"0.1644",
"0.504",
"0.1644"
],
[
"0.5693",
"0.1497",
"0.6187",
"0.1497",
"0.6187",
"0.165",
"0.5693",
"0.165"
],
[
"0.6484",
"0.1497",
"0.6978",
"0.1497",
"0.6978",
"0.165",
"0.6484",
"0.165"
],
[
"0.735",
"0.1497",
"0.781",
"0.1497",
"0.781",
"0.1644",
"0.735",
"0.1644"
],
[
"0.8003",
"0.1497",
"0.8486",
"0.1497",
"0.8486",
"0.1644",
"0.8003",
"0.1644"
],
[
"0.8003",
"0.1644",
"0.8486",
"0.1663",
"0.848",
"0.1818",
"0.7983",
"0.1799"
],
[
"0.2494",
"0.1509",
"0.33",
"0.1509",
"0.33",
"0.1621",
"0.2494",
"0.1621"
],
[
"0.3652",
"0.165",
"0.4148",
"0.165",
"0.4148",
"0.1812",
"0.3652",
"0.1812"
],
[
"0.4375",
"0.165",
"0.4844",
"0.165",
"0.4844",
"0.1812",
"0.4375",
"0.1812"
],
[
"0.5054",
"0.165",
"0.553",
"0.165",
"0.553",
"0.1812",
"0.5054",
"0.1812"
],
[
"0.5693",
"0.165",
"0.617",
"0.165",
"0.617",
"0.1812",
"0.5693",
"0.1812"
],
[
"0.6484",
"0.165",
"0.6978",
"0.165",
"0.6978",
"0.1812",
"0.6484",
"0.1812"
],
[
"0.735",
"0.165",
"0.781",
"0.165",
"0.781",
"0.1812",
"0.735",
"0.1812"
],
[
"0.241",
"0.1882",
"0.3376",
"0.1882",
"0.3376",
"0.2031",
"0.241",
"0.2031"
],
[
"0.2393",
"0.1675",
"0.34",
"0.1675",
"0.34",
"0.1787",
"0.2393",
"0.1787"
],
[
"0.367",
"0.1882",
"0.413",
"0.1882",
"0.413",
"0.2031",
"0.367",
"0.2031"
],
[
"0.4375",
"0.1882",
"0.4844",
"0.1882",
"0.4844",
"0.2031",
"0.4375",
"0.2031"
],
[
"0.5054",
"0.1882",
"0.553",
"0.1882",
"0.553",
"0.2037",
"0.5054",
"0.2037"
],
[
"0.5693",
"0.1882",
"0.617",
"0.1882",
"0.617",
"0.2031",
"0.5693",
"0.2031"
],
[
"0.6484",
"0.1882",
"0.6978",
"0.1882",
"0.6978",
"0.2031",
"0.6484",
"0.2031"
],
[
"0.735",
"0.1882",
"0.781",
"0.1882",
"0.781",
"0.2031",
"0.735",
"0.2031"
],
[
"0.8003",
"0.1882",
"0.8486",
"0.1882",
"0.8486",
"0.2031",
"0.8003",
"0.2031"
],
[
"0.2468",
"0.2048",
"0.3333",
"0.2048",
"0.3333",
"0.2197",
"0.2468",
"0.2197"
],
[
"0.367",
"0.2048",
"0.4148",
"0.2048",
"0.4148",
"0.2197",
"0.367",
"0.2197"
],
[
"0.4375",
"0.2037",
"0.4836",
"0.2037",
"0.4836",
"0.2197",
"0.4375",
"0.2197"
],
[
"0.504",
"0.2048",
"0.553",
"0.2048",
"0.553",
"0.2197",
"0.504",
"0.2197"
],
[
"0.5693",
"0.2048",
"0.6187",
"0.2048",
"0.6187",
"0.2197",
"0.5693",
"0.2197"
],
[
"0.6484",
"0.2048",
"0.6978",
"0.2048",
"0.6978",
"0.2197",
"0.6484",
"0.2197"
],
[
"0.735",
"0.2048",
"0.7827",
"0.2048",
"0.7827",
"0.2197",
"0.735",
"0.2197"
],
[
"0.8003",
"0.2048",
"0.8486",
"0.2048",
"0.8486",
"0.2197",
"0.8003",
"0.2197"
],
[
"0.367",
"0.2203",
"0.4148",
"0.2203",
"0.4148",
"0.2351",
"0.367",
"0.2351"
],
[
"0.4358",
"0.2203",
"0.4836",
"0.2203",
"0.4836",
"0.2351",
"0.4358",
"0.2351"
],
[
"0.5054",
"0.2203",
"0.553",
"0.2203",
"0.553",
"0.2351",
"0.5054",
"0.2351"
],
[
"0.571",
"0.2197",
"0.617",
"0.2197",
"0.617",
"0.2351",
"0.571",
"0.2351"
],
[
"0.65",
"0.2203",
"0.6978",
"0.2203",
"0.6978",
"0.2351",
"0.65",
"0.2351"
],
[
"0.735",
"0.2203",
"0.781",
"0.2203",
"0.781",
"0.2351",
"0.735",
"0.2351"
],
[
"0.8003",
"0.2203",
"0.8486",
"0.2203",
"0.8486",
"0.2351",
"0.8003",
"0.2351"
],
[
"0.2418",
"0.2214",
"0.3342",
"0.2214",
"0.3342",
"0.2328",
"0.2418",
"0.2328"
],
[
"0.1587",
"0.2214",
"0.2124",
"0.2214",
"0.2124",
"0.2328",
"0.1587",
"0.2328"
],
[
"0.2418",
"0.2363",
"0.336",
"0.2363",
"0.336",
"0.2505",
"0.2418",
"0.2505"
],
[
"0.367",
"0.2363",
"0.4148",
"0.2363",
"0.4148",
"0.2517",
"0.367",
"0.2517"
],
[
"0.4358",
"0.2363",
"0.4836",
"0.2363",
"0.4836",
"0.2505",
"0.4358",
"0.2505"
],
[
"0.5054",
"0.2363",
"0.555",
"0.2363",
"0.555",
"0.2505",
"0.5054",
"0.2505"
],
[
"0.5693",
"0.2363",
"0.6187",
"0.2363",
"0.6187",
"0.2505",
"0.5693",
"0.2505"
],
[
"0.6484",
"0.2363",
"0.6978",
"0.2363",
"0.6978",
"0.2505",
"0.6484",
"0.2505"
],
[
"0.735",
"0.2363",
"0.781",
"0.2363",
"0.781",
"0.2505",
"0.735",
"0.2505"
],
[
"0.8003",
"0.2363",
"0.8486",
"0.2363",
"0.8486",
"0.2505",
"0.8003",
"0.2505"
],
[
"0.6484",
"0.2517",
"0.6978",
"0.2517",
"0.6978",
"0.2673",
"0.6484",
"0.2673"
],
[
"0.735",
"0.2517",
"0.781",
"0.2517",
"0.781",
"0.2673",
"0.735",
"0.2673"
],
[
"0.8003",
"0.2505",
"0.8486",
"0.253",
"0.848",
"0.2683",
"0.7983",
"0.2666"
],
[
"0.4375",
"0.253",
"0.4844",
"0.253",
"0.4844",
"0.2673",
"0.4375",
"0.2673"
],
[
"0.3652",
"0.253",
"0.413",
"0.253",
"0.413",
"0.2673",
"0.3652",
"0.2673"
],
[
"0.504",
"0.253",
"0.553",
"0.253",
"0.553",
"0.2673",
"0.504",
"0.2673"
],
[
"0.5693",
"0.253",
"0.617",
"0.253",
"0.617",
"0.2673",
"0.5693",
"0.2673"
],
[
"0.8003",
"0.2725",
"0.8506",
"0.275",
"0.8486",
"0.2903",
"0.7983",
"0.288"
],
[
"0.2334",
"0.2542",
"0.3477",
"0.2542",
"0.3477",
"0.2666",
"0.2334",
"0.2666"
],
[
"0.2393",
"0.2737",
"0.339",
"0.2756",
"0.339",
"0.2903",
"0.2393",
"0.288"
],
[
"0.3652",
"0.275",
"0.4148",
"0.275",
"0.4148",
"0.2903",
"0.3652",
"0.2903"
],
[
"0.4358",
"0.275",
"0.4844",
"0.275",
"0.4844",
"0.2903",
"0.4358",
"0.2903"
],
[
"0.504",
"0.275",
"0.553",
"0.275",
"0.553",
"0.2903",
"0.504",
"0.2903"
],
[
"0.5693",
"0.275",
"0.6187",
"0.275",
"0.6187",
"0.2903",
"0.5693",
"0.2903"
],
[
"0.6484",
"0.275",
"0.6978",
"0.275",
"0.6978",
"0.2903",
"0.6484",
"0.2903"
],
[
"0.735",
"0.275",
"0.7827",
"0.275",
"0.7827",
"0.2903",
"0.735",
"0.2903"
],
[
"0.2468",
"0.2893",
"0.3333",
"0.2915",
"0.3333",
"0.3071",
"0.2468",
"0.3047"
],
[
"0.3652",
"0.2915",
"0.4148",
"0.2915",
"0.4148",
"0.3071",
"0.3652",
"0.3071"
],
[
"0.4358",
"0.2915",
"0.4844",
"0.2915",
"0.4844",
"0.3071",
"0.4358",
"0.3071"
],
[
"0.504",
"0.2915",
"0.553",
"0.2915",
"0.553",
"0.3071",
"0.504",
"0.3071"
],
[
"0.5693",
"0.2915",
"0.617",
"0.2915",
"0.617",
"0.3071",
"0.5693",
"0.3071"
],
[
"0.735",
"0.2915",
"0.7827",
"0.2915",
"0.7827",
"0.306",
"0.735",
"0.306"
],
[
"0.8003",
"0.2893",
"0.8486",
"0.2915",
"0.848",
"0.3071",
"0.7983",
"0.3047"
],
[
"0.4358",
"0.3071",
"0.4844",
"0.3071",
"0.4844",
"0.3225",
"0.4358",
"0.3225"
],
[
"0.6484",
"0.2922",
"0.6978",
"0.2922",
"0.6978",
"0.3071",
"0.6484",
"0.3071"
],
[
"0.367",
"0.308",
"0.4148",
"0.308",
"0.4148",
"0.3225",
"0.367",
"0.3225"
],
[
"0.5054",
"0.308",
"0.553",
"0.308",
"0.553",
"0.3225",
"0.5054",
"0.3225"
],
[
"0.5693",
"0.308",
"0.6187",
"0.308",
"0.6187",
"0.3225",
"0.5693",
"0.3225"
],
[
"0.65",
"0.308",
"0.6978",
"0.308",
"0.6978",
"0.3225",
"0.65",
"0.3225"
],
[
"0.735",
"0.308",
"0.781",
"0.308",
"0.781",
"0.3225",
"0.735",
"0.3225"
],
[
"0.8003",
"0.308",
"0.8486",
"0.308",
"0.8486",
"0.3225",
"0.8003",
"0.3225"
],
[
"0.2435",
"0.3088",
"0.3342",
"0.3088",
"0.3342",
"0.3206",
"0.2435",
"0.3206"
],
[
"0.1587",
"0.3088",
"0.2107",
"0.3088",
"0.2107",
"0.3206",
"0.1587",
"0.3206"
],
[
"0.3652",
"0.3237",
"0.4148",
"0.3237",
"0.4148",
"0.3384",
"0.3652",
"0.3384"
],
[
"0.4358",
"0.3237",
"0.4844",
"0.3237",
"0.4844",
"0.339",
"0.4358",
"0.339"
],
[
"0.504",
"0.3237",
"0.553",
"0.3237",
"0.553",
"0.3384",
"0.504",
"0.3384"
],
[
"0.5693",
"0.3237",
"0.6187",
"0.3237",
"0.6187",
"0.339",
"0.5693",
"0.339"
],
[
"0.6484",
"0.3237",
"0.6978",
"0.3237",
"0.6978",
"0.3384",
"0.6484",
"0.3384"
],
[
"0.735",
"0.3237",
"0.781",
"0.3237",
"0.781",
"0.339",
"0.735",
"0.339"
],
[
"0.8003",
"0.3237",
"0.8486",
"0.3237",
"0.8486",
"0.3384",
"0.8003",
"0.3384"
],
[
"0.3652",
"0.339",
"0.4148",
"0.339",
"0.4148",
"0.3552",
"0.3652",
"0.3552"
],
[
"0.2435",
"0.3247",
"0.3342",
"0.3247",
"0.3342",
"0.3362",
"0.2435",
"0.3362"
],
[
"0.4358",
"0.339",
"0.4844",
"0.339",
"0.4844",
"0.3552",
"0.4358",
"0.3552"
],
[
"0.65",
"0.339",
"0.6978",
"0.339",
"0.6978",
"0.3552",
"0.65",
"0.3552"
],
[
"0.8003",
"0.339",
"0.8486",
"0.339",
"0.8486",
"0.3552",
"0.8003",
"0.3552"
],
[
"0.5693",
"0.3403",
"0.6187",
"0.3403",
"0.6187",
"0.3552",
"0.5693",
"0.3552"
],
[
"0.504",
"0.3403",
"0.553",
"0.3403",
"0.553",
"0.3552",
"0.504",
"0.3552"
],
[
"0.735",
"0.3403",
"0.781",
"0.3403",
"0.781",
"0.3552",
"0.735",
"0.3552"
],
[
"0.1184",
"0.3706",
"0.8833",
"0.3706",
"0.8833",
"0.3843",
"0.1184",
"0.3843"
],
[
"0.2351",
"0.3416",
"0.345",
"0.3416",
"0.345",
"0.3528",
"0.2351",
"0.3528"
],
[
"0.11755",
"0.383",
"0.1855",
"0.3855",
"0.1855",
"0.3997",
"0.11755",
"0.398"
],
[
"0.12177",
"0.4233",
"0.4878",
"0.4246",
"0.4878",
"0.4395",
"0.12177",
"0.4382"
],
[
"0.5146",
"0.4246",
"0.8833",
"0.4246",
"0.8833",
"0.4395",
"0.5146",
"0.4395"
],
[
"0.12006",
"0.4412",
"0.486",
"0.4412",
"0.486",
"0.4548",
"0.12006",
"0.4548"
],
[
"0.5146",
"0.4412",
"0.882",
"0.4412",
"0.882",
"0.4548",
"0.5146",
"0.4548"
],
[
"0.12006",
"0.4578",
"0.4878",
"0.4578",
"0.4878",
"0.4714",
"0.12006",
"0.4714"
],
[
"0.5156",
"0.4573",
"0.882",
"0.4573",
"0.882",
"0.4714",
"0.5156",
"0.4714"
],
[
"0.12006",
"0.4739",
"0.486",
"0.4739",
"0.486",
"0.4868",
"0.12006",
"0.4868"
],
[
"0.5146",
"0.4739",
"0.5728",
"0.4739",
"0.5728",
"0.488",
"0.5146",
"0.488"
],
[
"0.1184",
"0.488",
"0.4878",
"0.4893",
"0.4878",
"0.5044",
"0.1184",
"0.503"
],
[
"0.534",
"0.4912",
"0.8833",
"0.4912",
"0.8833",
"0.505",
"0.534",
"0.505"
],
[
"0.12006",
"0.506",
"0.4895",
"0.506",
"0.4895",
"0.5195",
"0.12006",
"0.5195"
],
[
"0.5156",
"0.5073",
"0.8833",
"0.5073",
"0.8833",
"0.521",
"0.5156",
"0.521"
],
[
"0.12006",
"0.5215",
"0.4895",
"0.5215",
"0.4895",
"0.536",
"0.12006",
"0.536"
],
[
"0.5146",
"0.5225",
"0.882",
"0.5225",
"0.882",
"0.536",
"0.5146",
"0.536"
],
[
"0.1184",
"0.538",
"0.4878",
"0.538",
"0.4878",
"0.552",
"0.1184",
"0.552"
],
[
"0.5156",
"0.538",
"0.8833",
"0.538",
"0.8833",
"0.5527",
"0.5156",
"0.5527"
],
[
"0.12177",
"0.5547",
"0.486",
"0.5547",
"0.486",
"0.5654",
"0.12177",
"0.5654"
],
[
"0.5146",
"0.5547",
"0.8833",
"0.5547",
"0.8833",
"0.5684",
"0.5146",
"0.5684"
],
[
"0.1184",
"0.571",
"0.4668",
"0.571",
"0.4668",
"0.5845",
"0.1184",
"0.5845"
],
[
"0.5146",
"0.571",
"0.7188",
"0.571",
"0.7188",
"0.585",
"0.5146",
"0.585"
],
[
"0.1377",
"0.586",
"0.4878",
"0.586",
"0.4878",
"0.5996",
"0.1377",
"0.5996"
],
[
"0.532",
"0.5874",
"0.882",
"0.586",
"0.882",
"0.601",
"0.532",
"0.6016"
],
[
"0.1184",
"0.601",
"0.4878",
"0.6016",
"0.4878",
"0.616",
"0.1184",
"0.615"
],
[
"0.5146",
"0.6025",
"0.8833",
"0.604",
"0.8833",
"0.618",
"0.5146",
"0.6177"
],
[
"0.12006",
"0.618",
"0.4878",
"0.618",
"0.4878",
"0.632",
"0.12006",
"0.632"
],
[
"0.5146",
"0.619",
"0.555",
"0.619",
"0.555",
"0.6353",
"0.5146",
"0.6353"
],
[
"0.12006",
"0.6343",
"0.4878",
"0.6343",
"0.4878",
"0.6484",
"0.12006",
"0.6484"
],
[
"0.5146",
"0.6494",
"0.544",
"0.6494",
"0.544",
"0.661",
"0.5146",
"0.661"
],
[
"0.561",
"0.6484",
"0.7583",
"0.648",
"0.7583",
"0.662",
"0.561",
"0.663"
],
[
"0.12177",
"0.6675",
"0.4878",
"0.6675",
"0.4878",
"0.681",
"0.12177",
"0.681"
],
[
"0.12006",
"0.651",
"0.486",
"0.651",
"0.486",
"0.6646",
"0.12006",
"0.6646"
],
[
"0.5146",
"0.6704",
"0.8833",
"0.6704",
"0.8833",
"0.684",
"0.5146",
"0.684"
],
[
"0.12006",
"0.6816",
"0.486",
"0.6816",
"0.486",
"0.6953",
"0.12006",
"0.6953"
],
[
"0.513",
"0.685",
"0.881",
"0.685",
"0.881",
"0.6987",
"0.513",
"0.6987"
],
[
"0.12177",
"0.6987",
"0.4878",
"0.6987",
"0.4878",
"0.7134",
"0.12177",
"0.7134"
],
[
"0.5146",
"0.7017",
"0.881",
"0.7017",
"0.881",
"0.7153",
"0.5146",
"0.7153"
],
[
"0.12177",
"0.7153",
"0.4878",
"0.7153",
"0.4878",
"0.7285",
"0.12177",
"0.7285"
],
[
"0.5146",
"0.7188",
"0.882",
"0.7188",
"0.882",
"0.7324",
"0.5146",
"0.7324"
],
[
"0.12006",
"0.731",
"0.4727",
"0.731",
"0.4727",
"0.745",
"0.12006",
"0.745"
],
[
"0.5156",
"0.7354",
"0.882",
"0.7354",
"0.882",
"0.749",
"0.5156",
"0.749"
],
[
"0.1377",
"0.7476",
"0.4895",
"0.7476",
"0.4895",
"0.761",
"0.1377",
"0.761"
],
[
"0.5156",
"0.752",
"0.882",
"0.752",
"0.882",
"0.7656",
"0.5156",
"0.7656"
],
[
"0.1184",
"0.763",
"0.4878",
"0.763",
"0.4878",
"0.777",
"0.1184",
"0.777"
],
[
"0.5156",
"0.767",
"0.7715",
"0.767",
"0.7715",
"0.781",
"0.5156",
"0.781"
],
[
"0.12006",
"0.78",
"0.3787",
"0.78",
"0.3787",
"0.7935",
"0.12006",
"0.7935"
],
[
"0.5146",
"0.7935",
"0.817",
"0.7935",
"0.817",
"0.808",
"0.5146",
"0.808"
],
[
"0.1528",
"0.8047",
"0.3552",
"0.8057",
"0.3552",
"0.821",
"0.1528",
"0.8193"
],
[
"0.11584",
"0.811",
"0.1276",
"0.8047",
"0.1368",
"0.8135",
"0.126",
"0.8193"
],
[
"0.5156",
"0.814",
"0.8833",
"0.814",
"0.8833",
"0.8276",
"0.5156",
"0.8276"
],
[
"0.12006",
"0.8306",
"0.4878",
"0.8306",
"0.4878",
"0.844",
"0.12006",
"0.844"
],
[
"0.511",
"0.829",
"0.882",
"0.8306",
"0.882",
"0.8457",
"0.511",
"0.8433"
],
[
"0.1184",
"0.8467",
"0.4878",
"0.8467",
"0.4878",
"0.861",
"0.1184",
"0.861"
],
[
"0.511",
"0.8467",
"0.8833",
"0.8467",
"0.8833",
"0.861",
"0.511",
"0.861"
],
[
"0.12006",
"0.8633",
"0.4895",
"0.8633",
"0.4895",
"0.877",
"0.12006",
"0.877"
],
[
"0.5146",
"0.8623",
"0.8833",
"0.8623",
"0.8833",
"0.877",
"0.5146",
"0.877"
],
[
"0.12177",
"0.879",
"0.4895",
"0.879",
"0.4895",
"0.8936",
"0.12177",
"0.8936"
],
[
"0.5146",
"0.879",
"0.882",
"0.879",
"0.882",
"0.8926",
"0.5146",
"0.8926"
],
[
"0.1184",
"0.8955",
"0.4878",
"0.8955",
"0.4878",
"0.909",
"0.1184",
"0.909"
],
[
"0.5156",
"0.8955",
"0.8833",
"0.8955",
"0.8833",
"0.909",
"0.5156",
"0.909"
]
] | [
"Models",
"Methods",
"SST-1",
"SST-2",
"Yelp",
"AG News",
"TREC",
"Subj",
"CNN-base",
"IMDB",
"89.06",
"46.32",
"85.50",
"94.32",
"91.30",
"92.40",
"92.80",
"CNN- l2",
"89.12",
"46.01",
"85.56",
"94.46",
"91.28",
"90.62",
"92.39",
"CNN",
"78.94",
"37.92",
"80.01",
"83.14",
"84.36",
"61.00",
"82.40",
"88.31",
"CNN-L2X",
"41.40",
"84.24",
"93.82",
"91.37",
"89.80",
"91.80",
"93.50",
"CNN-IBA",
"90.10",
"48.92",
"85.78",
"94.53",
"91.60",
"93.02",
"LSTM-base",
"CNN-Vmask",
"88.39",
"43.84",
"83.74",
"95.06",
"91.03",
"90.40",
"90.20",
"LSTM- l2",
"88.40",
"43.91",
"83.36",
"95.00",
"91.09",
"90.20",
"89.10",
"67.45",
"36.92",
"75.45",
"77.12",
"77.53",
"46.00",
"81.80",
"LSTM-L2X",
"LSTM",
"LSTM-IBA",
"88.48",
"42.99",
"83.53",
"94.74",
"91.14",
"85.40",
"89.50",
"92.19",
"90.80",
"91.20",
"44.12",
"90.07",
"84.35",
"95.41",
"95.10",
"LSTM-Vmask",
"BERT-base",
"91.80",
"53.43",
"92.25",
"96.42",
"93.59",
"96.40",
"BERT-12",
"91.75",
"52.08",
"92.25",
"96.41",
"96.80",
"94.80",
"39.23",
"93.52",
"71.75",
"74.03",
"87.14",
"82.59",
"93.20",
"86.10",
"BERT-L2X",
"BERT",
"91.66",
"53.80",
"92.24",
"96.27",
"93.45",
"96.80",
"95.60",
"93.04",
"BERT-IBA",
"54.53",
"94.24",
"96.40",
"96.80",
"92.26",
"97.00",
"Table 3: Prediction accuracy (% ) of different models with different training strategies on the seven",
"BERT-Vmask",
"datasets.",
"Baselines and Competitive Methods.由于这",
"模型都具有相似的预测性能。使用Vmask训",
"项工作的目标是提出一种新颖的训练方法，以",
"练的模型优于具有类似网络架构但训练方式不",
"提高预测准确性和可解释性，因此我们采用两",
"同的模型。结果表明，Vmask可以帮助提高泛",
"组模型作为基线和竞争系统。使用所提出的方",
"化能力",
"法训练的模型以后缀“Vmask”命名。我们",
"除了基本模型和用所提出的方法训练的模",
"还提供了两个基线：（1）通过最小化交叉熵损",
"型外，其他三个竞争对手的记录是混合的。例",
"失训练的模型（后缀为“-base\"）和（2）用l2",
"如，传统的l2-正则化不能总是有助于提高准",
"-正则化训练的模型（后缀为“-l2”）。与这两",
"确性，特别是对于BERT模型。尽管IBA的",
"种基准方法的比较主要集中在预测性能上，因",
"性能略优于L2X，但使用它们进行训练并不能",
"为没有使用明确的训练策略来提高可解释性",
"持续提高模型的预测精度",
"此外，我们还提出了两种竞争方法：使用解",
"为了呼应提高模型可解释性的目的，本节的",
"释框架“学习解释”（Chenetal.,2018）（后缀",
"其余部分将集中于定量和定性地评估模型可解",
"为“-L2X\"）和“信息瓶颈归因”（Schulzetal.",
"释性",
"2020）（后缀为“-IBA”）训练的模型。L2X",
"5.1",
"Quantitative Evaluation",
"素的模型的事后解释而提出的。我们将它们集",
"和IBA最初是为了寻找特征属性作为训练有",
"我们通过AOPC得分（Nguyen，2018；Samek",
"成在模型训练中，作为掩码层直接为输入特征",
"etal..2016）评估基于Vmask的模型相对于",
"(L2X）生成掩码值，或通过添加噪声（IBA）限",
"基本模型的局部可解释性，并通过事后准确性",
"制信息流。在我们的实验中，所有的训练方法",
"(Chenetal.，2018）评估基于IBA的模型的全",
"都使用随机退出（p=0.2）来避免过度拟合",
"局可解释性。从经验上讲，我们观察到了局部",
"有关实验设置的更多详细信息在Ap-",
"和全局可解释性之间的一致性，因此在我们的",
"pendixC中，包括数据预处理，模型配置以及",
"评估中无需穷尽所有可能的组合",
"我们实验中L2X和IBA的实现",
"5.1.1Local interpretability:AOPC",
"Results and Discussion",
"5",
"我们采用两种与模型无关的解释方法，",
"我们使用不同的训练策略在七个数据集上训",
"LIME(Ribeiro etal.,2016)和SampleShap",
"练了三个模型。Table3显示了测试集上不同",
"ley（Kononenkoetal.，2010），为基础模型和",
"模型的预测精度。验证性能和平均运行时间在",
"基于Vmask的模型生成局部解释，其中“局",
"AppendixD中。如Table3所示，与先前工作",
"部”表示分别解释每个测试数据。扰动曲线",
"（AppendixE）中报告的数字相比，所有基础",
"上的面积（AOPC）E(Nguyen，2018；Samek"
] | {
"arxiv_id": "2010.00667",
"language": "zh",
"page_idx": 4
} |
||
2206.09314-zh-0 | 2206.09314-zh-0.jpg | Robust Imitation Learning against Variations in Environment Dynamics
Jongseong Chae$^1$ Seungyul Han$^2$ Whiyoung Jung$^1$ Myungsik Cho$^1$ Sungho Choi$^1$ Youngchul Sung$^1$
3in
$\operatorname{Abstract}$
In this paper, we propose a robust imitation learning (IIL) framework that improves the robustness of IL when environment dynamics are perturbed. The existing IIL framework trained in a single environment can catastrophically fail with perturbations in environment dynamics because it does not capture the situation that underlying environment dynamics can be changed. Our framework effectively deals with environments with varying dynamics by imitating multiple experts in sampled environment dynamics to enhance the robustness in general variations in environment dynamics. In order to robustly imitate the multiple sample experts, we minimize the risk with respect to the JensenShannon divergence between the agent's policy and each of the sample experts. Numerical results show that our algorithm significantly improves robustness against dynamics perturbations compared to conventional IL ${\mathrm{baselines.}}$
口不口定性，可以考口口口定任口(如汽口口口)的所有可能的口境口力口中的每一00000家策略，口口已口口的口力口 ,口口其中一口口口的口家策略用于口口的口力口。然而，口口口口的口境口力口是困口的 ,因口口境的口力口取口于口多相口的口境口口 ,如重力 ,口量 , 老化等。此外，口口力口口口口化口，口所有可能的口力$\square\square\square\square\square\square\square$策略是不可行的
健口強化口口(RobustRL)是一口口架，口口口口定任口的此口口境口口口生健口的策略。目的是口口-口在所有可能的口力口口口中都能口好地工作的策略， 而无需口口口口。典型的健口RL允口代理在多口口境(Derman et al.,2018; Mankowitz et al.,2018;2019) 中口行交互 ,口且策略口化了多口交互口境中口期收盒的最口情口。口代理可以在所有交互口境中，甚至在具有口似口口的看不口的口境中都能口好地工作。口使口口的口口口口化可以由口棒的RL口理 ,但口于口多口口世界的控制口口 ,包括我口的自口口口示例 ,仍然存在口口函口口口的口口，因口口棒的RL依口于精心口口的口口函口。口我口口察人口口口口，口口知道口口口口口口口的每口口作都有什口口口
已口口出模口口口(IL)口口口口口情口，方法是在口有$\square\square$函口(Torabi et al.,2018a;Finn et al.,2016;Syed et al.,2008)的情口下口口口定任口的策略。IL使用口家口任口生成的演示而不是口口函口，代理口口模口口家。GAIL是流行的IL算法之一，口口口通口匹配占用度量口模口口家，口是口口-口作口(Ho &Ermon,2016) 的非口一化分布。然而 ,到目前口止 ,大多口IIL算法都是口口口一交互口境提出的 ,具有完美或非完美的口家口口，以口生口口口口口口一交互口境的策略
在本文中，我□考□了囗棒IL$\square\square\square\square\square\square\square\square$境口力口口口的口棒策略 ,囗提出了一口新口的IL口架，以口基于演示口口口在一系列口口口口口化中表口良好的口棒策略。口自口口口的口口采口口力口，不需要口自所有口口口口的演示 ,因此大大口少了所需的演示量。$\square\square\square\square$架的口口口在接下口的章口中介口。
## 1. Introduction
強化口口(RL)是一口口任口口生最佳策略的口架。深度神口口口使RL能口口理各口口口口境中的口口任口(Mnih et al.,2015;Lillicrap et al.,2015;Fujimoto et al., 2018;Haarnoja et al.,2018;Schulman et al.,2015a; 2017)。但是，口前的RL在部署到口口世界中仍然存在限制。口口主要的限制是口棒性和口口函口的口口。典型的RL算法口口口口境口行交互，口通口交互口境口口策略 ,因此策略口得口口口口口口口境 ,口且在口口口境中的基本口口受到干口口大多口失口。在口口世界中，口在的口口可能口受到干口。例如 ,考口使用RL 口行自口口口。自口口口汽口的物理口力口，包括操口制口 ,雨天的道路摩擦系口口晴天的口化。口了口口口
## 2. Related Works
${\overline{{^{1}\text{School of}}}}{\mathrm{Electrical~Engineering,~KAIST,~Daejeon,}}$ South Korea.$\:^2$Artificial Intelligence Graduate School, UNIST, Ulsan, South Korea. Correspondence to: Seungyul Han <[email protected]>.
模口口口$\square{:}$IL旨在通口模口口家口口口策略。行口克隆(BC)(Torabi et al.,2018a) 是一口基于口督口口的IL 方法。Brantley et al. (2019) 口解了BC的口口量偏移口$\square。\square$一口方法是口抗性模口口口(Ho&Ermon,2016; Torabi et al.,2018b),其中代理通口匹配口口的度量口模口口家。Fu et al.(2017) 使用口家口口恢口了口口函
Proceedings of the 39$^th$ International Conference on Machine Learning, Baltimore, Maryland, USA, PMLR 162, 2022. Copyright 2022 by the author(s). | [
[
"0.1217",
"0.11237",
"0.853",
"0.113",
"0.853",
"0.1307",
"0.1217",
"0.13"
],
[
"0.10297",
"0.1787",
"0.8613",
"0.1787",
"0.8613",
"0.1932",
"0.10297",
"0.1932"
],
[
"0.0907",
"0.1963",
"0.12335",
"0.1963",
"0.12335",
"0.209",
"0.0907",
"0.209"
],
[
"0.2443",
"0.2084",
"0.3186",
"0.2084",
"0.3186",
"0.2229",
"0.2443",
"0.2229"
],
[
"0.501",
"0.2084",
"0.884",
"0.209",
"0.884",
"0.2235",
"0.501",
"0.2229"
],
[
"0.1217",
"0.2291",
"0.4412",
"0.2291",
"0.4412",
"0.2437",
"0.1217",
"0.2437"
],
[
"0.5034",
"0.2247",
"0.884",
"0.2247",
"0.884",
"0.238",
"0.5034",
"0.238"
],
[
"0.1209",
"0.2443",
"0.4395",
"0.2443",
"0.4395",
"0.2583",
"0.1209",
"0.2583"
],
[
"0.5024",
"0.2405",
"0.884",
"0.2405",
"0.884",
"0.254",
"0.5024",
"0.254"
],
[
"0.1217",
"0.2588",
"0.4412",
"0.2588",
"0.4412",
"0.2727",
"0.1217",
"0.2727"
],
[
"0.5034",
"0.2551",
"0.8857",
"0.2551",
"0.8857",
"0.2683",
"0.5034",
"0.2683"
],
[
"0.1217",
"0.2747",
"0.4412",
"0.2747",
"0.4412",
"0.2886",
"0.1217",
"0.2886"
],
[
"0.5024",
"0.2695",
"0.8857",
"0.2708",
"0.8857",
"0.2822",
"0.5024",
"0.281"
],
[
"0.5024",
"0.2842",
"0.884",
"0.2842",
"0.884",
"0.2974",
"0.5024",
"0.2974"
],
[
"0.1209",
"0.2903",
"0.4412",
"0.2903",
"0.4412",
"0.3037",
"0.1209",
"0.3037"
],
[
"0.1209",
"0.305",
"0.4412",
"0.305",
"0.4412",
"0.318",
"0.1209",
"0.318"
],
[
"0.5034",
"0.2998",
"0.6978",
"0.2998",
"0.6978",
"0.3132",
"0.5034",
"0.3132"
],
[
"0.1217",
"0.3208",
"0.4412",
"0.3208",
"0.4412",
"0.334",
"0.1217",
"0.334"
],
[
"0.5024",
"0.3225",
"0.8857",
"0.3225",
"0.8857",
"0.336",
"0.5024",
"0.336"
],
[
"0.1217",
"0.3352",
"0.4412",
"0.3352",
"0.4412",
"0.3484",
"0.1217",
"0.3484"
],
[
"0.5034",
"0.339",
"0.884",
"0.339",
"0.884",
"0.3499",
"0.5034",
"0.3499"
],
[
"0.1217",
"0.3499",
"0.4412",
"0.3499",
"0.4412",
"0.363",
"0.1217",
"0.363"
],
[
"0.5024",
"0.3528",
"0.8857",
"0.3528",
"0.8857",
"0.3662",
"0.5024",
"0.3662"
],
[
"0.1217",
"0.3643",
"0.4395",
"0.3655",
"0.4395",
"0.379",
"0.1217",
"0.3774"
],
[
"0.5024",
"0.3687",
"0.8857",
"0.3687",
"0.8857",
"0.3818",
"0.5024",
"0.3818"
],
[
"0.1209",
"0.379",
"0.4412",
"0.3801",
"0.4412",
"0.3945",
"0.1209",
"0.3933"
],
[
"0.5034",
"0.385",
"0.8823",
"0.385",
"0.8823",
"0.3958",
"0.5034",
"0.3958"
],
[
"0.1217",
"0.3958",
"0.4412",
"0.3958",
"0.4412",
"0.4092",
"0.1217",
"0.4092"
],
[
"0.501",
"0.3984",
"0.884",
"0.3984",
"0.884",
"0.4124",
"0.501",
"0.4124"
],
[
"0.1217",
"0.4104",
"0.4412",
"0.4104",
"0.4412",
"0.4236",
"0.1217",
"0.4236"
],
[
"0.5024",
"0.4136",
"0.884",
"0.4136",
"0.884",
"0.4268",
"0.5024",
"0.4268"
],
[
"0.1217",
"0.4255",
"0.4412",
"0.4255",
"0.4412",
"0.4395",
"0.1217",
"0.4395"
],
[
"0.5034",
"0.43",
"0.884",
"0.43",
"0.884",
"0.44",
"0.5034",
"0.44"
],
[
"0.1217",
"0.4414",
"0.4412",
"0.4414",
"0.4412",
"0.455",
"0.1217",
"0.455"
],
[
"0.501",
"0.443",
"0.884",
"0.443",
"0.884",
"0.457",
"0.501",
"0.457"
],
[
"0.1209",
"0.4558",
"0.4395",
"0.4558",
"0.4395",
"0.4697",
"0.1209",
"0.4697"
],
[
"0.5034",
"0.4602",
"0.8823",
"0.4602",
"0.8823",
"0.4702",
"0.5034",
"0.4702"
],
[
"0.1217",
"0.4717",
"0.4429",
"0.4717",
"0.4429",
"0.4849",
"0.1217",
"0.4849"
],
[
"0.5024",
"0.4734",
"0.8857",
"0.4734",
"0.8857",
"0.4873",
"0.5024",
"0.4873"
],
[
"0.1217",
"0.486",
"0.4412",
"0.486",
"0.4412",
"0.5",
"0.1217",
"0.5"
],
[
"0.5024",
"0.4893",
"0.8857",
"0.4893",
"0.8857",
"0.5024",
"0.5024",
"0.5024"
],
[
"0.1217",
"0.5005",
"0.4412",
"0.5005",
"0.4412",
"0.5146",
"0.1217",
"0.5146"
],
[
"0.5034",
"0.506",
"0.7305",
"0.506",
"0.7305",
"0.5166",
"0.5034",
"0.5166"
],
[
"0.1209",
"0.5166",
"0.4395",
"0.5166",
"0.4395",
"0.53",
"0.1209",
"0.53"
],
[
"0.1217",
"0.531",
"0.4412",
"0.531",
"0.4412",
"0.5444",
"0.1217",
"0.5444"
],
[
"0.5024",
"0.5264",
"0.884",
"0.5264",
"0.884",
"0.5405",
"0.5024",
"0.5405"
],
[
"0.12335",
"0.5474",
"0.1896",
"0.5474",
"0.1896",
"0.5596",
"0.12335",
"0.5596"
],
[
"0.505",
"0.5435",
"0.884",
"0.5435",
"0.884",
"0.5537",
"0.505",
"0.5537"
],
[
"0.501",
"0.5566",
"0.8857",
"0.5566",
"0.8857",
"0.571",
"0.501",
"0.571"
],
[
"0.5034",
"0.5737",
"0.884",
"0.5737",
"0.884",
"0.584",
"0.5034",
"0.584"
],
[
"0.08905",
"0.589",
"0.2222",
"0.589",
"0.2222",
"0.603",
"0.08905",
"0.603"
],
[
"0.5024",
"0.587",
"0.8857",
"0.587",
"0.8857",
"0.601",
"0.5024",
"0.601"
],
[
"0.5024",
"0.6016",
"0.884",
"0.6016",
"0.884",
"0.6157",
"0.5024",
"0.6157"
],
[
"0.08905",
"0.613",
"0.4739",
"0.6143",
"0.4739",
"0.629",
"0.08905",
"0.6274"
],
[
"0.501",
"0.617",
"0.8823",
"0.617",
"0.8823",
"0.631",
"0.501",
"0.631"
],
[
"0.0907",
"0.63",
"0.4739",
"0.63",
"0.4739",
"0.643",
"0.0907",
"0.643"
],
[
"0.5024",
"0.632",
"0.884",
"0.632",
"0.884",
"0.646",
"0.5024",
"0.646"
],
[
"0.08905",
"0.646",
"0.4739",
"0.646",
"0.4739",
"0.659",
"0.08905",
"0.659"
],
[
"0.501",
"0.648",
"0.789",
"0.648",
"0.789",
"0.661",
"0.501",
"0.661"
],
[
"0.08905",
"0.66",
"0.4739",
"0.66",
"0.4739",
"0.674",
"0.08905",
"0.674"
],
[
"0.0907",
"0.675",
"0.4739",
"0.675",
"0.4739",
"0.688",
"0.0907",
"0.688"
],
[
"0.5024",
"0.6694",
"0.8857",
"0.6694",
"0.8857",
"0.683",
"0.5024",
"0.683"
],
[
"0.08826",
"0.6895",
"0.4739",
"0.69",
"0.4739",
"0.704",
"0.08826",
"0.7026"
],
[
"0.5024",
"0.685",
"0.884",
"0.685",
"0.884",
"0.6987",
"0.5024",
"0.6987"
],
[
"0.0907",
"0.705",
"0.4739",
"0.705",
"0.4739",
"0.7183",
"0.0907",
"0.7183"
],
[
"0.501",
"0.6997",
"0.8823",
"0.7007",
"0.8823",
"0.714",
"0.501",
"0.7134"
],
[
"0.0907",
"0.7197",
"0.4722",
"0.7197",
"0.4722",
"0.733",
"0.0907",
"0.733"
],
[
"0.501",
"0.7153",
"0.884",
"0.7153",
"0.884",
"0.729",
"0.501",
"0.729"
],
[
"0.0907",
"0.735",
"0.4722",
"0.735",
"0.4722",
"0.7485",
"0.0907",
"0.7485"
],
[
"0.5034",
"0.731",
"0.8813",
"0.731",
"0.8813",
"0.744",
"0.5034",
"0.744"
],
[
"0.09235",
"0.7505",
"0.4739",
"0.7505",
"0.4739",
"0.7646",
"0.09235",
"0.7646"
],
[
"0.5024",
"0.747",
"0.7983",
"0.747",
"0.7983",
"0.7603",
"0.5024",
"0.7603"
],
[
"0.0907",
"0.7666",
"0.4722",
"0.7666",
"0.4722",
"0.78",
"0.0907",
"0.78"
],
[
"0.0907",
"0.781",
"0.4739",
"0.781",
"0.4739",
"0.7944",
"0.0907",
"0.7944"
],
[
"0.501",
"0.78",
"0.653",
"0.78",
"0.653",
"0.7944",
"0.501",
"0.7944"
],
[
"0.1111",
"0.802",
"0.4722",
"0.8037",
"0.4722",
"0.8184",
"0.1111",
"0.8164"
],
[
"0.5024",
"0.806",
"0.8857",
"0.806",
"0.8857",
"0.8193",
"0.5024",
"0.8193"
],
[
"0.08905",
"0.8145",
"0.1765",
"0.8174",
"0.1765",
"0.832",
"0.08905",
"0.829"
],
[
"0.1978",
"0.815",
"0.4739",
"0.8164",
"0.4739",
"0.8306",
"0.1978",
"0.83"
],
[
"0.5024",
"0.821",
"0.8857",
"0.821",
"0.8857",
"0.834",
"0.5024",
"0.834"
],
[
"0.0907",
"0.829",
"0.3015",
"0.829",
"0.3015",
"0.8423",
"0.0907",
"0.8423"
],
[
"0.3064",
"0.83",
"0.4722",
"0.8306",
"0.4722",
"0.8423",
"0.3064",
"0.8413"
],
[
"0.08905",
"0.8423",
"0.3228",
"0.8423",
"0.3228",
"0.856",
"0.08905",
"0.856"
],
[
"0.5034",
"0.837",
"0.8823",
"0.837",
"0.8823",
"0.8477",
"0.5034",
"0.8477"
],
[
"0.5034",
"0.853",
"0.884",
"0.853",
"0.884",
"0.863",
"0.5034",
"0.863"
],
[
"0.08826",
"0.866",
"0.4739",
"0.8657",
"0.4739",
"0.88",
"0.08826",
"0.881"
],
[
"0.5024",
"0.866",
"0.884",
"0.866",
"0.884",
"0.88",
"0.5024",
"0.88"
],
[
"0.08905",
"0.88",
"0.4756",
"0.88",
"0.4756",
"0.893",
"0.08905",
"0.893"
],
[
"0.5024",
"0.8833",
"0.8823",
"0.8833",
"0.8823",
"0.893",
"0.5024",
"0.893"
],
[
"0.08905",
"0.8926",
"0.3472",
"0.8926",
"0.3472",
"0.906",
"0.08905",
"0.906"
]
] | [
"Robust Imitation Learning against Variations in Environment Dynamics",
"Jongseong Chae1 Seungyul Han 2 Whiyoung Jung1 Myungsik Cho1 Sungho Choil Youngchul Sung l",
".3in",
"Abstract",
"口不口定性，可以考口口口定任口（如汽口口口）的所有",
"In this paper, we propose a robust imitation",
"可能的口境口力口中的每一口口口口家策略，口口已口",
"learning (IL) framework that improves the ro",
"口的口力口，口口其中一口口口的口家策略用于口口的",
"bustness of IL when environment dynamics",
"口力口。然而，口口口口的口境口力口是困口的，因口口",
"are perturbed. The existing IL framework",
"境的口力口取口于口多相口的口境口口，如重力，口量，",
"老化等。此外，口口力口口口口化口，口所有可能的口力",
"trained in a single environment can catas-",
"trophically fail with perturbations in environ-",
"口口口口策略是不可行的",
"ment dynamics because it does not capture",
"健口强化口（RobustRL）是一口架，口口口定任",
"the situation that underlying environment",
"口的此口口境口口口生健口的策略。目的是口口一口",
"dynamics can be changed.Our framework",
"在所有可能的口力口口口中都能口好地工作的策略，",
"effectively deals with environments with vary",
"而无需口口口口。典型的健口RL允口代理在多口口境",
"ing dynamics by imitating multiple experts in",
"(Dermanetal.,2018;Mankowitzetal.,2018;2019)",
"sampled environment dynamics to enhance",
"口行交互，口且策略口化了多口交互口境中口期收盆的",
"the robustness in general variations in envi-",
"最口情口。口代理可以在所有交互境中，甚至在具有",
"ronment dynamics. In order to robustly im-",
"口似口口的看不口的口境中都能口好地工作。口使口口",
"itate the multiple sample experts, we min-",
"的口口口化可以由口棒的RL口理，但口于口多口口世",
"imize the risk with respect to the Jensen",
"界的控制口口，包括我口的自口口口示例，仍然存在■",
"Shannon divergence between the agent's pol-",
"口函口口口的口口，因口口棒的RL依口于精心口口的",
"icy and each of the sample experts. Numer-",
"口口函口.口我口口察人口口口口，口口知道口口口口",
"ical results show that our algorithm signifi-",
"口口口的每口口作都有什口口口",
"cantly improves robustness against dynamics",
"perturbations compared to conventional IL",
"已口口出模口口（IL）口口口口情口，方法是在口有",
"baselines.",
"函(Torabiet al.,2018a;Finnetal.,2016;Syed",
"etal.，2008）的情口下口口定任口的策略。IL使用口",
"家口任口生成的演示而不是口口函口，代理口口模口口",
"1.Introduction",
"家。GAIL是流行的IL算法之一，口□通匹配占用度",
"量口模口家，是口-作（Ho&Ermon，2016)",
"强化口口（RL）是一口口任口口生最佳策略的口架。深",
"的非口一化分布。然而到目前口止，大多口IL算法都是",
"度神口口口使RL能口口理各口口口口境中的口口任口",
"口一交互境提出的，具有完美或非完美的家口",
"(Mnih et al., 2015; Lillicrap et al., 2015; Fujimoto et al.,",
"口，以口生口口口口口一交互口境的策略",
"2018; Haarnoja et al., 2018; Schulman et al., 2015a;",
"2017）。但是，口前的RL在部署到口口世界中仍然存在",
"在本文中，我口考口了口棒IL口口口IL口口口口境口",
"限制。口口主要的限制是口棒性和口口函口的口口。典",
"力口口口的口棒策略，口提出了一口新口的IL口架，以",
"型的RL算法口口口口境口行交互，口通口交互口境口",
"口基于演示口口口在一系列口口口口口化中表口良好",
"口策略，因此策略口得口口口口口口口境，口且在口口",
"的口棒策略。口自口口口的口口采口口力口，不需要口",
"口境中的基本口口受到千口口大多口失口。在口口世界",
"自所有口口的演示，因此大大口少了所需的演示量。",
"中，口在的口口口可能口受到千口。例如，考口使用RL",
"口口口架的口口口在接下口的章口中介口",
"口行自口口口。自口口口汽口的物理口力口，包括操口",
"制口，雨天的道路摩擦系口口晴天的口化。口了口口口",
"2. Related Works",
"1 School of Electrical Engineering, KAIST, Daejeon",
"模口口口：IL旨在通口模口口家口口口策略。行口克隆",
"South Korea.",
"2 Artificial Intelligence Graduate School,",
"（BC）（Torabietal.，2018a）是一基于督□的IL",
"UNIST, Ulsan, South Korea :",
"Correspondenceto:Se",
"ungyul Han < [email protected] >.",
"方法。Brantleyetal.（2019）■解了BC的口口量偏移■",
"口。口一口方法是抗性模口口口（Ho&Ermon，2016;",
"Proceedings of the 39th International Conference on Ma-",
"Torabietal.,2018b），其中代理通口匹配的度量",
"chine Learning, Baltimore, Maryland, USA, PMLR 162,",
"模口口家。Fuetal.（2017）使用口家口口恢口了口口函",
"2022. Copyright 2022 by the author(s)."
] | {
"arxiv_id": "2206.09314",
"language": "zh",
"page_idx": 0
} |
||
2206.09314-zh-1 | 2206.09314-zh-1.jpg | Robust Imitation Learning against Variations in Environment Dynamics
Proposition 3.1 of Ho&Ermon (2016)). In a single ${\mathrm{environment,~the~occupancy~measure~}\rho_{\pi}(s,a)}$ satisfies the following Bellman flow constraint for each $(s,a)\in$ $\mathcal{S}\times\mathcal{A}:$
$$\rho_{\pi}(s,a)=\mu_{0}(s)\pi(a|s)+\gamma\int_{(s',a')}\mathcal{P}(s|s',a')\rho_{\pi}(s',a')\pi(a|s)\quad(1$$
and the policy $\pi$ whose occupancy measure is $\rho_\pi$ is unique. That is, the occupancy measure and the policy are in an one-to-one relationship.
策略$\pi$引起占用度量$\rho_\pi$ ,口且 $\rho_\pi$ 映射到唯一策略$\pi$ 因此 ,GAIL口策略更新 (2) 中口制了口家的策略 ,口策略口代理策略和口家策略的占用度量相匹配：
$$\begin{aligned}&\operatorname*{min}_{\pi}\mathcal{D}_{JS}(\bar{\rho}_{\pi},\bar{\rho}_{\pi_{E}})&\text{(2}\\&\stackrel{(a)}{=}\operatorname*{min}_{\pi}\mathbb{E}_{\rho_{\pi_{E}}}\left[\log\frac{\rho_{\pi_{E}}}{\rho_{\pi}+\rho_{\pi_{E}}}\right]+\mathbb{E}_{\rho_{\pi}}\left[\log\frac{\rho_{\pi}}{\rho_{\pi}+\rho_{\pi_{E}}}\right]\\&\stackrel{(b)}{=}\operatorname*{min}_{\pi}\operatorname*{max}_{D}\mathbb{E}_{\rho_{\pi_{E}}}\left[\log D(s,a)\right]+\mathbb{E}_{\rho_{\pi}}\left[\log(1-D(s,a))\right]\end{aligned}$$
$\square$。跨域IL (Gangwani & Peng,2020; Liu et al.,2019; Viano et al.,2021;Raychaudhuri et al.,2021)考口了口家口口口者口口不匹配下的IIL$\square\square\square$
$\square$有的口棒IIL作品通口非完美演示 (Wu et al.,2019; Tangkaratt et al.,2020)解口了IIL口口或提高了IL (Wang et al.,2017;Laskey et al.,2017)的口定性， $\square 且\square\square 的\square 置\square 我\square 在本文中的\square$置不同。Meta-IL (Duan et al.,2017;Finn et al.,2017;James et al.,2018; Zhou et al.,2019) 和Meta-IRL (Xu et al.,2019;Yu et al.,2019) 可以通口利用口似任口的口口，使用一些演示口口口新任口，而我口的口架不需要任何口口任口的演示。多任□IRL (Gleave & Habryka,2018) 提出了多任口IRL和元口口的最大因果$\square$IRL口架，以推口每口任口的多口口口函口。Toyer et al.(2020)提出了一口多任口基准口口套件 ,用于口口IL算法的口棒性。ADAIL (Lu & Tompson, 2020) 可以口口口口各口口口口境的自适口策略，但是口假口在多口口境中收集口家演示是不可行的 ,口且使用了口多模口口境口行域口机化和口境口口
健口的強化口口：健口的RL口口境口口口生健口的策略。口棒MDP (Iyengar,2005;Wiesemann et al., 2013)口展了MDP 上的不口定性口口集。Derman et al. (2018);Mankowitz et al. $(2018;2019)\square\square$了多$\square\square\square\square$ $\square$境中口期收盒的最口情$\square$。Pinto et al.(2017)通口使用口手解口了口棒RL口口。Tessler et al.(2019);Vinitsky et al. (2020) 形式化的口作不口定性口棒性判据。
# 3. Preliminaries
3.1. Markov Decision Process
MDP由元$\square\mathcal{M}=<\mathcal{S},\mathcal{A},\mathcal{P},r,\gamma>$表示，其中$\mathcal{S}$ 是口$\square$空口 ,$\mathcal{A}$ 是口作空口 ,$\mathcal{P}:\mathcal{S}\times\mathcal{A}\times\mathcal{S}\to\mathbb{R}^+$ 是口移口率$,r:\mathcal{S}\times\mathcal{A}\to\mathbb{R}$是口口函口$,\gamma\in(0,1)$是折口因子。策略$\pi$ 是(口机)映射$\pi:\mathcal{S}\mapsto\mathcal{A}$。返回$G_t$ 是口自口口步11
$\mathrm{e}_{\mathrm{。}},G_{t}=\sum_{i=t}^{\infty}\gamma^{i-t}r(s_{i},a_{i})$
.目口是口口一口策略$\pi$,以最大化口期收盒$J(\pi)=$ $\{s_0,a_0,s_1,a_1,\ldots\}$是情口口迹$,\mu_0(s)$表示初始口口分布。占用度量$\rho_{\pi}(s,a)=\pi(a|s)\sum_{t=0}^{\infty}\gamma^{t}$Pr$(s_t=s|\pi,\mathcal{P})$ 是策略$\pi$引起的非口范化口口-口作分布，而$\mu_\pi(s)=$ $\sum_{t=0}^{\infty}\gamma^{t}$Pr$(s_t=s|\pi,\mathcal{P})$是策略$\pi$引起的非口范化口口分布
3.2. Generative Adversarial Imitation Learning
在IL中，代理不口因其操作而口得口口。相反，代理基于口家的演示口口口策略，而不知道明口的口家策略。通常，口家演示是由口家策略$\tau_E=\{s_0,a_0,s_1,a_1,\ldots\}$ 生成的口迹口出的。。生成口抗模口口口(GAIL)(Ho&& Ermon,2016) 是使用口家口口的流行IIL方法之一。基于Proposition 3.1 ,GAIL口求占用度量接近口家策略的策略 ,以便代理的策略$\pi$接近口家策略$\pi_E$
Proposition 3.1 (Theorem 2 of Syed et al. (2008) &
$\square$常口口一化器口最小化无口，口且(b)是有效的，因口最大化$D$口由$D(s,a)=\frac\rho_{\pi_E}(s,a){\rho_\pi(s,a)+\rho_\pi_E(s,a)}$口出 ,其中口$\square$器$D$口分口定口$(s,a)$是否口自己家
,其中$\mathcal{D}_JS$表示Jensen-Shannon( JS)散度 ,而$\bar{\rho}_\pi$和$\bar{\rho}_{\pi_E}$分口是XMATHX和$\rho_\pi$ 。在口里 ,(a)是有效的 ,因Gradient Penalty: GAIL (Kostrikov et al.,2018)的$\square\square$使用 Gulrajani et al. (2017) 提出的梯度口口(GP) 作口正口化口，以增強IL的口定性。口有GP的GAIL的口口器更新由
(3)
$$\begin{aligned}\max_{D}\mathbb{E}_{\rho_{\pi_{E}}}\left[\log D(s,a)\right]&+\mathbb{E}_{\rho_{\pi}}[\log(1-D(s,a))]\\&+\kappa\mathbb{E}_{\hat{x}}\left(\|\nabla_{\hat{x}}D(\hat{x})\|_{2}-1\right)^{2},\end{aligned}$$
$\square$出，其中$\hat{x} \sim ( \epsilon \rho _{\pi }$ + (1 - $\epsilon ) \rho _{\pi _{E}})$口有$\epsilon\sim$ Uniform[0,1] ,而$\kappa$是控制GP$\bar\text{口的正口化系口。我}\square\square$ 此口口GAIL+GP。
# 4. Motivation
$\square$有的IL方法通常$\square\square\square\square\square\square\square\square$境交互，口口口模口$\square\square\square\square\square\square\square\square\square\square\square$境的口家。口了口—步口口 ,我口定口了三口口型的口境：交互口境 ,演示口境和口口口境。交互口境是代理在口口口程中口之交互以口取策略口本的口境 ,演示口境是生成口家演示以口口代理的口境 ,口口口境是口口后的代理策略的口口口口口境。口于具有口口口口口口境(SNE)的常口IL ,交互口境和演示口境是相同的。我口口此IL 口口口置口口SNE/SNE (交互口境/演示口境)。在大多口情口下，如 Figures 1a and 1b所示 ,口口口口口口境口口受到口口口口的干口$\square$,在SNE/SNE□置中口口的IL口失口。在Figures 1a and 1b 中 ,x□□表示口口口境的重力(或口量)$\square\square\square\square\square$ 交互/演示口境的重力(或口量)之比(百分比),y$\square 表示$ 在SNE/SNE□置下口口的策略在相口$x$ 口下的平均收盒。可以看出，口着口口口境口口偏口口口交互/演示口境口口，性能口口重下降 | [
[
"0.2467",
"0.05682",
"0.728",
"0.05682",
"0.728",
"0.0707",
"0.2467",
"0.0707"
],
[
"0.0907",
"0.0859",
"0.4739",
"0.0859",
"0.4739",
"0.09973",
"0.0907",
"0.09973"
],
[
"0.5",
"0.0859",
"0.884",
"0.0859",
"0.884",
"0.09973",
"0.5",
"0.09973"
],
[
"0.08905",
"0.09973",
"0.4697",
"0.101",
"0.4697",
"0.11554",
"0.08905",
"0.11426"
],
[
"0.501",
"0.09973",
"0.8857",
"0.09973",
"0.8857",
"0.113",
"0.501",
"0.113"
],
[
"0.09235",
"0.1187",
"0.3506",
"0.1187",
"0.3506",
"0.1288",
"0.09235",
"0.1288"
],
[
"0.5024",
"0.113",
"0.8857",
"0.113",
"0.8857",
"0.1276",
"0.5024",
"0.1276"
],
[
"0.501",
"0.1276",
"0.556",
"0.1276",
"0.556",
"0.1395",
"0.501",
"0.1395"
],
[
"0.0907",
"0.1395",
"0.4739",
"0.1395",
"0.4739",
"0.1528",
"0.0907",
"0.1528"
],
[
"0.08905",
"0.1528",
"0.4739",
"0.154",
"0.4739",
"0.1686",
"0.08905",
"0.1674"
],
[
"0.5146",
"0.1547",
"0.6626",
"0.1547",
"0.6626",
"0.1654",
"0.5146",
"0.1654"
],
[
"0.7017",
"0.1515",
"0.884",
"0.1515",
"0.884",
"0.1654",
"0.7017",
"0.1654"
],
[
"0.669",
"0.161",
"0.7075",
"0.161",
"0.7075",
"0.1718",
"0.669",
"0.1718"
],
[
"0.0907",
"0.1692",
"0.4697",
"0.1692",
"0.4697",
"0.1831",
"0.0907",
"0.1831"
],
[
"0.4656",
"0.1736",
"0.4697",
"0.1736",
"0.4697",
"0.1799",
"0.4656",
"0.1799"
],
[
"0.0907",
"0.1843",
"0.4739",
"0.1843",
"0.4739",
"0.1976",
"0.0907",
"0.1976"
],
[
"0.505",
"0.1812",
"0.886",
"0.1831",
"0.886",
"0.1976",
"0.505",
"0.1957"
],
[
"0.09235",
"0.2008",
"0.4722",
"0.2008",
"0.4722",
"0.2108",
"0.09235",
"0.2108"
],
[
"0.501",
"0.1976",
"0.884",
"0.1976",
"0.884",
"0.2108",
"0.501",
"0.2108"
],
[
"0.08905",
"0.214",
"0.4739",
"0.214",
"0.4739",
"0.2279",
"0.08905",
"0.2279"
],
[
"0.5",
"0.2134",
"0.739",
"0.2134",
"0.739",
"0.2267",
"0.5",
"0.2267"
],
[
"0.0907",
"0.2311",
"0.4722",
"0.2311",
"0.4722",
"0.2412",
"0.0907",
"0.2412"
],
[
"0.5034",
"0.2374",
"0.8774",
"0.2374",
"0.8774",
"0.2507",
"0.5034",
"0.2507"
],
[
"0.0907",
"0.2443",
"0.4722",
"0.2443",
"0.4722",
"0.2583",
"0.0907",
"0.2583"
],
[
"0.501",
"0.2474",
"0.884",
"0.25",
"0.884",
"0.2644",
"0.501",
"0.262"
],
[
"0.0907",
"0.26",
"0.4722",
"0.26",
"0.4722",
"0.2734",
"0.0907",
"0.2734"
],
[
"0.5",
"0.2615",
"0.819",
"0.262",
"0.819",
"0.2766",
"0.5",
"0.276"
],
[
"0.0907",
"0.2747",
"0.4722",
"0.2747",
"0.4722",
"0.2886",
"0.0907",
"0.2886"
],
[
"0.5195",
"0.2886",
"0.6357",
"0.2886",
"0.6357",
"0.3018",
"0.5195",
"0.3018"
],
[
"0.865",
"0.2874",
"0.8857",
"0.2874",
"0.8857",
"0.3005",
"0.865",
"0.3005"
],
[
"0.08905",
"0.305",
"0.4722",
"0.3037",
"0.4722",
"0.318",
"0.08905",
"0.3193"
],
[
"0.0907",
"0.2903",
"0.4722",
"0.2903",
"0.4722",
"0.3037",
"0.0907",
"0.3037"
],
[
"0.523",
"0.3093",
"0.609",
"0.318",
"0.6055",
"0.3352",
"0.5186",
"0.3245"
],
[
"0.653",
"0.3132",
"0.6787",
"0.3132",
"0.6787",
"0.3213",
"0.653",
"0.3213"
],
[
"0.8037",
"0.3132",
"0.8228",
"0.3132",
"0.8228",
"0.3208",
"0.8037",
"0.3208"
],
[
"0.7183",
"0.3176",
"0.7524",
"0.3176",
"0.7524",
"0.3289",
"0.7183",
"0.3289"
],
[
"0.6167",
"0.3164",
"0.634",
"0.3193",
"0.6323",
"0.327",
"0.614",
"0.324"
],
[
"0.7607",
"0.318",
"0.7812",
"0.318",
"0.7812",
"0.327",
"0.7607",
"0.327"
],
[
"0.634",
"0.3245",
"0.7036",
"0.327",
"0.7036",
"0.3384",
"0.634",
"0.336"
],
[
"0.778",
"0.3257",
"0.8486",
"0.3257",
"0.8486",
"0.3372",
"0.778",
"0.3372"
],
[
"0.08905",
"0.3193",
"0.4714",
"0.3193",
"0.4714",
"0.3333",
"0.08905",
"0.3333"
],
[
"0.0907",
"0.334",
"0.4714",
"0.3352",
"0.4714",
"0.3499",
"0.0907",
"0.3484"
],
[
"0.755",
"0.3308",
"0.7715",
"0.3308",
"0.7715",
"0.334",
"0.755",
"0.334"
],
[
"0.5156",
"0.3372",
"0.8735",
"0.3435",
"0.8735",
"0.3687",
"0.5156",
"0.3623"
],
[
"0.0907",
"0.3499",
"0.4722",
"0.3499",
"0.4722",
"0.363",
"0.0907",
"0.363"
],
[
"0.0907",
"0.3662",
"0.1396",
"0.3662",
"0.1396",
"0.3774",
"0.0907",
"0.3774"
],
[
"0.513",
"0.375",
"0.8857",
"0.3757",
"0.8857",
"0.3901",
"0.513",
"0.3894"
],
[
"0.0907",
"0.3882",
"0.4722",
"0.3882",
"0.4722",
"0.4016",
"0.0907",
"0.4016"
],
[
"0.5",
"0.3914",
"0.884",
"0.3901",
"0.884",
"0.4048",
"0.5",
"0.406"
],
[
"0.0907",
"0.4028",
"0.4722",
"0.4028",
"0.4722",
"0.4167",
"0.0907",
"0.4167"
],
[
"0.501",
"0.406",
"0.8857",
"0.406",
"0.8857",
"0.4192",
"0.501",
"0.4192"
],
[
"0.08905",
"0.4172",
"0.4722",
"0.4172",
"0.4722",
"0.4312",
"0.08905",
"0.4312"
],
[
"0.4968",
"0.4204",
"0.8896",
"0.4204",
"0.8896",
"0.4458",
"0.4968",
"0.4458"
],
[
"0.714",
"0.4204",
"0.759",
"0.4192",
"0.759",
"0.428",
"0.714",
"0.4287"
],
[
"0.09235",
"0.435",
"0.4714",
"0.435",
"0.4714",
"0.4458",
"0.09235",
"0.4458"
],
[
"0.501",
"0.4426",
"0.7793",
"0.4426",
"0.7793",
"0.4558",
"0.501",
"0.4558"
],
[
"0.09235",
"0.4495",
"0.4714",
"0.4495",
"0.4714",
"0.4602",
"0.09235",
"0.4602"
],
[
"0.09235",
"0.4634",
"0.4739",
"0.4634",
"0.4739",
"0.4766",
"0.09235",
"0.4766"
],
[
"0.5024",
"0.4634",
"0.884",
"0.4634",
"0.884",
"0.4766",
"0.5024",
"0.4766"
],
[
"0.08826",
"0.4792",
"0.4558",
"0.4792",
"0.4558",
"0.4924",
"0.08826",
"0.4924"
],
[
"0.5024",
"0.476",
"0.8823",
"0.476",
"0.8823",
"0.4893",
"0.5024",
"0.4893"
],
[
"0.5024",
"0.488",
"0.884",
"0.488",
"0.884",
"0.502",
"0.5024",
"0.502"
],
[
"0.5034",
"0.5024",
"0.5986",
"0.5024",
"0.5986",
"0.513",
"0.5034",
"0.513"
],
[
"0.08826",
"0.5146",
"0.2296",
"0.5146",
"0.2296",
"0.528",
"0.08826",
"0.528"
],
[
"0.544",
"0.526",
"0.8857",
"0.5244",
"0.8857",
"0.539",
"0.544",
"0.5405"
],
[
"0.08826",
"0.539",
"0.3057",
"0.539",
"0.3057",
"0.553",
"0.08826",
"0.553"
],
[
"0.5786",
"0.549",
"0.749",
"0.549",
"0.749",
"0.5625",
"0.5786",
"0.5625"
],
[
"0.08905",
"0.5625",
"0.4722",
"0.5625",
"0.4722",
"0.577",
"0.08905",
"0.577"
],
[
"0.0907",
"0.5776",
"0.4739",
"0.5776",
"0.4739",
"0.5913",
"0.0907",
"0.5913"
],
[
"0.509",
"0.5737",
"0.864",
"0.5747",
"0.864",
"0.589",
"0.509",
"0.5884"
],
[
"0.08905",
"0.5913",
"0.468",
"0.5923",
"0.468",
"0.607",
"0.08905",
"0.606"
],
[
"0.5",
"0.5884",
"0.884",
"0.589",
"0.884",
"0.604",
"0.5",
"0.603"
],
[
"0.0907",
"0.608",
"0.4722",
"0.608",
"0.4722",
"0.622",
"0.0907",
"0.622"
],
[
"0.501",
"0.605",
"0.6357",
"0.605",
"0.6357",
"0.6187",
"0.501",
"0.6187"
],
[
"0.0907",
"0.6235",
"0.3733",
"0.6235",
"0.3733",
"0.637",
"0.0907",
"0.637"
],
[
"0.3757",
"0.625",
"0.4722",
"0.625",
"0.4722",
"0.636",
"0.3757",
"0.636"
],
[
"0.09314",
"0.638",
"0.4739",
"0.638",
"0.4739",
"0.6514",
"0.09314",
"0.6514"
],
[
"0.5",
"0.64",
"0.6216",
"0.64",
"0.6216",
"0.654",
"0.5",
"0.654"
],
[
"0.0866",
"0.653",
"0.455",
"0.6494",
"0.455",
"0.6675",
"0.0866",
"0.6704"
],
[
"0.09235",
"0.6685",
"0.4722",
"0.6685",
"0.4722",
"0.683",
"0.09235",
"0.683"
],
[
"0.501",
"0.6655",
"0.8857",
"0.6655",
"0.8857",
"0.6787",
"0.501",
"0.6787"
],
[
"0.335",
"0.6816",
"0.4739",
"0.683",
"0.4739",
"0.6978",
"0.335",
"0.6963"
],
[
"0.5024",
"0.6816",
"0.8823",
"0.6816",
"0.8823",
"0.6924",
"0.5024",
"0.6924"
],
[
"0.5024",
"0.696",
"0.884",
"0.696",
"0.884",
"0.709",
"0.5024",
"0.709"
],
[
"0.09235",
"0.6846",
"0.3406",
"0.6846",
"0.3406",
"0.6978",
"0.09235",
"0.6978"
],
[
"0.0907",
"0.6987",
"0.4697",
"0.6987",
"0.4697",
"0.712",
"0.0907",
"0.712"
],
[
"0.0866",
"0.7104",
"0.4763",
"0.711",
"0.4763",
"0.732",
"0.0866",
"0.731"
],
[
"0.501",
"0.7104",
"0.884",
"0.7104",
"0.884",
"0.7236",
"0.501",
"0.7236"
],
[
"0.08905",
"0.728",
"0.1259",
"0.728",
"0.1259",
"0.7437",
"0.08905",
"0.7437"
],
[
"0.5024",
"0.726",
"0.884",
"0.726",
"0.884",
"0.7393",
"0.5024",
"0.7393"
],
[
"0.5024",
"0.741",
"0.88",
"0.741",
"0.88",
"0.755",
"0.5024",
"0.755"
],
[
"0.0866",
"0.759",
"0.4307",
"0.761",
"0.4307",
"0.7744",
"0.0866",
"0.773"
],
[
"0.501",
"0.755",
"0.8857",
"0.756",
"0.8857",
"0.771",
"0.501",
"0.7695"
],
[
"0.5024",
"0.7715",
"0.8857",
"0.7715",
"0.8857",
"0.785",
"0.5024",
"0.785"
],
[
"0.0907",
"0.784",
"0.4739",
"0.784",
"0.4739",
"0.7974",
"0.0907",
"0.7974"
],
[
"0.5024",
"0.786",
"0.8857",
"0.786",
"0.8857",
"0.8",
"0.5024",
"0.8"
],
[
"0.09314",
"0.8003",
"0.4656",
"0.8003",
"0.4656",
"0.811",
"0.09314",
"0.811"
],
[
"0.5",
"0.802",
"0.8823",
"0.802",
"0.8823",
"0.815",
"0.5",
"0.815"
],
[
"0.08905",
"0.813",
"0.4722",
"0.8145",
"0.4722",
"0.829",
"0.08905",
"0.8276"
],
[
"0.5024",
"0.8164",
"0.8857",
"0.8164",
"0.8857",
"0.83",
"0.5024",
"0.83"
],
[
"0.08905",
"0.829",
"0.4722",
"0.8276",
"0.4722",
"0.8413",
"0.08905",
"0.8423"
],
[
"0.501",
"0.8306",
"0.884",
"0.8306",
"0.884",
"0.8447",
"0.501",
"0.8447"
],
[
"0.0907",
"0.8447",
"0.4722",
"0.8447",
"0.4722",
"0.858",
"0.0907",
"0.858"
],
[
"0.5024",
"0.8467",
"0.884",
"0.8467",
"0.884",
"0.86",
"0.5024",
"0.86"
],
[
"0.0907",
"0.8613",
"0.4714",
"0.8613",
"0.4714",
"0.872",
"0.0907",
"0.872"
],
[
"0.5024",
"0.8613",
"0.884",
"0.8613",
"0.884",
"0.875",
"0.5024",
"0.875"
],
[
"0.08905",
"0.8735",
"0.3865",
"0.875",
"0.3865",
"0.8896",
"0.08905",
"0.888"
],
[
"0.5034",
"0.877",
"0.884",
"0.877",
"0.884",
"0.89",
"0.5034",
"0.89"
],
[
"0.0907",
"0.893",
"0.4722",
"0.893",
"0.4722",
"0.904",
"0.0907",
"0.904"
],
[
"0.5034",
"0.893",
"0.675",
"0.893",
"0.675",
"0.904",
"0.5034",
"0.904"
]
] | [
"Robust Imitation Learning against Variations in Environment Dynamics",
"。跨域IL(Gangwani&Peng,2020;Liu et al.,2019;",
"Proposition 3.1 of Ho & Ermon (2016).In a single",
"Viano et al.,2021;Raychaudhuri et al.,2021)考了",
"environment, the occupancy measure P(s, a) satisfies",
"口家口口口者口口不匹配下的口口",
"the following Bellman flow constraint for each (s, a) E",
"SxA:",
"口有的口棒IL作品通口非完美演示（Wuetal.，2019;",
"Tangkarattetal，2020)解了IL或提高了IL",
"p(s,a) =μo(s)(as) +~",
"P(sls,a')p(s,a')(als)(1)",
"J(st.a\")",
"(Wangetal.，2017;Laskeyetal.,2017)的定性",
"",
"口且口口的口置口我口在本文中的口置不同。Meta-IL",
"and the policy whose occupancy measure is P is",
"(Duan et al..2017:Finn et al..2017:Jameset al.,2018:",
"unique. That is, the occupancy measure and the policy",
"Zhou et al., 2019)和Meta-IRL (Xu et al., 2019; Yu",
"are in an one-to-one relationship.",
"etal.，2019）可以通口利用口似任口的口口使用一此演",
"策略元引起占用度量P元，口且P元映射到唯一策略元",
"示口口口新任口，而我口的口架不需要任何口口任口的",
"因此，GA丘策略更新（2)中制了口家的策略，口策",
"演示。多任IRL（Gleave&Habryka，2018）提出了多",
"略口代理策略和口家策略的占用度量相匹配：",
"任口IRL和元口口的最大因果口IRL口架，以推口每口",
"minDJs(P,P)",
"(2)",
"任口基准口口套件，用于口口IL算法的口棒性。ADAIL",
"任口的多口函口。Toyeretal.(2020)提出了一口多",
"minEp*E",
"PTE",
"0元",
"Ee元",
"log",
"log",
"0元+P#E",
"P元+P元",
"（Lu&Tompson，2020）可以口口口口各口口境的",
"自适口策略，但是口假口在多口口境中收集口家演示是",
"",
"minmaxEpme [log D(s,a)] + Ep [og(1 - D(s,a)",
"不可行的，口且使用了口多模口口境口行域口机化和口",
"境口口",
"，其中DJs表示Jensen-Shannon（JS）散度，而p元和",
"健口的强化口口：健口的RL口口境口口口生健口的",
"P元分是XMATHX和P元。在口里，（a）是有效的因",
"策略。棒MDP(lyengar,2005;Wiesemann et al.,",
"口常口口一化器口最小化无口，口且（b）是有效的，因口",
"2013)口展了MDP上的不定性口集。Dermanetal.",
"最大化由D(sa)=+（）出，其中",
"TE(S.a",
"(2018);Mankowitzetal.(2018;2019)□了多",
"口器D口分口定口（sa）是否口自口家",
"境中口期收盒的最口情口Pintoetal.（2017）通口使",
"用手解了棒RL。Tessleretal.(2019)；Vinit-",
"Gradient Penalty:GAIL (Kostrikov et al.,2018)的",
"skyetal.(2020）形式化的口作不口定性口棒性判据。",
"口使用Gulrajanietal.（2017）提出的梯度口口（GP）",
"作口正口化口，以增强IL的口定性。口有GP的GAIL的",
"口口器更新由",
"3. Preliminaries",
"maxEpne [log D(s,a)] +Ep[log(1 -D(s,a))](3)",
"3.1. Markov Decision Process",
"+ E (l/VD()l2 -1)?,",
"MDP由元M=<S,A,P,r,>表示,其中S是",
"空口A是口作空口，P：S×A×S→R+是口移口",
"出，其中～（ep元+（1－)p元）有",
"率，r：S×A→R是□函口，E（0,1）是折因子",
"Uniform[0,1]，而是控制GP口的正口化系口。我口口",
"策略元是（机）映射元：S→A。返回G是口自口",
"此GAIL+GP。",
"步口t的口口果口口i。e，Gt=。",
"t-tr(si,ai)",
"目口是口一口策略元，以最大化口期收盆J（元）=",
"4.Motivation",
"Eso~μo,T~元 [Go] (Sutton & Barto, 2018),其中T",
"[so,ao,si,ai，..]是情迹，o(s）表示初始分",
"口有的L方法通常口口口口口境交互，口口口模口",
".%Pr(st= s, P)",
"口口口口口口口口境的口家。口了口一步口口我",
"定口了三口口型的口境：交互口境，演示口境和口口口",
"布。占用度量p(s,a）=元(as)t",
"是策略元引起的非口范化口口-口作分布，而μ元（s）",
"2Pr(st=s|元，P）是策略元引起的非口范化",
"境。交互口境是代理在口口程中口之交互以口取策",
"分布",
"略口本的口境，演示口境是生成口家演示以口口代理的",
"口境，口口口境是口口后的代理策略的口口口口口境。",
"3.2. Generative Adversarial Imitation Learning",
"口于具有口口口口境（SNE）的常口IL，交互境和演",
"示口境是相同的。我口口此IL口口口置口口SNE/SNE",
"在I中，代理不口因其操作而口得口口。相反，代理基",
"（交互口境/演示口境）。在大多口情口下，如Figuresla",
"于口家的演示口口口策略，而不知道明口的口家策略",
"and1b所示，口口口口口口境口口受到口口口的千",
"通常，口家演示是由家策略TE=【so,ao,Si,a1,..}",
"口，在SNE/SNE口置中口的IL口失口。在Figuresla",
"生成的口迹口出的。。生成口抗模口口口（GAIL）（Ho&",
"and1b中x口口表示口口境的重力（或口量）口口口",
"Ermon，2016)是使用口家口的流行IL方法之一。基于",
"交互/演示口境的重力（或口量）之比（百分比）y口表示",
"Proposition3.1，GAL口求占用度量接近■家策略的",
"在SNE/SNE口置下口口的策略在相口口下的平均收",
"策略，以便代理的策略π接近口家策略六E",
"盒。可以看出，口着口口口境口口偏口口口交互/演示口",
"Proposition 3.1 (Theorem 2 of Syed et al. (2008) &",
"境口口，性能口口重下降"
] | {
"arxiv_id": "2206.09314",
"language": "zh",
"page_idx": 1
} |
||
2206.09314-zh-2 | 2206.09314-zh-2.jpg | Robust Imitation Learning against Variations in Environment Dynamics
<FigureHere>
Figure 1: Performance of the policy trained in a single nominal environment against perturbation (left column- gravity perturbation and right column- mass perturbation): (a,b)- performance of GAIL+GP, and ( c, d) - SNEMPE- max.
<FigureHere>
Figure 2: Overview of our algorithm. The blue line is the flow of policy samples $\rho_\pi^i$, and the green line is the flow of expert demonstrations $\tau _E^j\sim \rho _E^{\bar{j} }$ .
口了口理口口性能下降，強大的RL口一些具有口口口口的口境口行采口。例如 ,在 Figure 1c 中，口重力口$50\% 150\%$ $100\%$
RL
MPE
RL
$\operatorname*{max}_{\pi}\operatorname*{min}_{\mathcal{P}^{i}\in P}\mathbb{E}_{\pi}[G_{t}|\mathcal{P}^{i}]$ ,其中
$P=\left\{\mathcal{P}^i\right\}$是口定的口境集。通口最大化最口情口下的口期收盒，代理的策略可以捕口所口境集$P=\{\mathcal{P}^i\}$ 中口化的口口。但是，強大的RL需要精心口口的口口函口，我口希望避免
$\square 在，考\square\square$棒IL。一口口口的方法是口上迹口棒RL原理口用于IL口置。在口里 ,我口口多口采口的演示口境中口得口家演示 ,口使口口策略口口口口交互口境口行交互。然后，我口使用口口器口口分口自多口抽口口家口口中的每一口的政策口本，口口口政策遵循最口的情口，$\square$。e。基于口口器口出的口策略口本最口的口家口口。Figures 1c and 1d 中口示了在受干口的口口口境中如此口口的策略的性能(相口的性能表示口SNEMPE-max)。可以看出，$\square$Figures la and 1b 中的常$\square$SNE/SNEIL相比，以口口方式口口的策略提高了口棒性 ,但性能口不合人口意。口口退化意味着口口口口口口境的策略交互不足以捕口口口口化 ,口使口自多口采口演示口境的口家演示也是如此。因此，口了充分捕捉口口口化，我口首先口口口口力口分布中采口一些具有不同口力口的口境 ,口口口些多口采口口境不口用于口家演示 ,口用于口口期口的策略交互。我口口此口置$\square\square\square$MPE/MPEIL口置。在本文的其余部分中 ,我口提出了一口基于MPE/MPEIL口置的高效IIL口架，以口生一口策略 ,口策略口基于少口用于口口的采口口口， 口可口口口口的口境口口口化口行口健的口行。
# 5. Robust Imitation Learning against Variations in Environment Dynamics
5.1. Problem Formulation
我口考$\square$MDP的集合$\mathcal{C}=\{\mathcal{M}=<\mathcal{S},\mathcal{A},\mathcal{P}_\zeta,r,\gamma>$ $,\zeta\in Z\}$ ,其中口口和口作空口口于集合的所有成口都是相同的 ,口口口代理不可用 ,用$\square\square\zeta\square\square\square$建模的口渡口率口行口口化，口且口力口口口$\varsigma$可以口XMTHMAX\_4集的口口口$Z$ 口口口化或口口。在此$\square\square$集合中，我口口具有口口口口$\zeta_1,\zeta_2,\cdots,\zeta_N$ 的 $N$ MDP$\square$行采口。我口用 $\mathcal{E}_1,\cdots,\mathcal{E}_N$ 表示口些具有口口$\mathcal{P}_{\zeta_1},\cdots,\mathcal{P}_{\zeta_N}(\Box\Box\Box\mathcal{P}^1,\cdots,\mathcal{P}^N)$的$N$口境。我$\Box$假$\square$每口口境$\mathcal{E}_i$ 都有一口口家 $\pi_E^i$ ,口家 $\pi_E^i$ 口代理生成口家演示 ,但是口家策略$\pi_E^i$本身口代理不可用。我口$\square$假口代理可以口所有采口口境 $\mathcal{E}_1,\cdots,\mathcal{E}_N$ 中的每一口口行交互 ,口且所有交互口境的初始口口分布口$\mu_0(\underline{s})$ 相同。因此，根据我口在上一口中的定口 ,$\mathcal{E}_1,\cdots,\mathcal{E}_N\sqcup$ 是演示口境又是交互口境 ,□置□MPE/MPE。囗注意， $\mathcal{E}_1,\cdots,\mathcal{E}_N$ 上的口家演示是口于整口MDP集合$\mathcal{C}$ 的部分信息。我口的目口是口代理口口一口策略$\pi$ ,口策略口于MDP集合$\mathcal{C}$ 中的所有成口都表口良好，口基于口自采口口境$\mathcal{E}_1,\cdots,\mathcal{E}_N$的口家演示和代理口采口口境$\mathcal{E}_1,\cdots,\mathcal{E}_N$的交互。我口口此口口口口 Robust Imitation learning with Multiple perturbed Environments (RIME)
口我口再介口口口符 口。
$\underline{\rho}_{\pi}^{i}(s,a)$
$\pi(a|s)\sum_{t=0}^{\infty}\gamma^{t}$Pr$( s_t$ = $s| \pi , \mathcal{P} ^{i})$表示$\pi$在$i$交互口境$\mathcal{E}_i$中的占用度量。$\mu_\pi^i(s)=\sum_t=0^\infty\gamma^t$Pr$(s_t=s|\pi,\mathcal{P}^i)$ 表示$\pi$在$i$交互$\square$境 $\mathcal{E}_i$ 中的非$\square$一化口口口口。口口$\square$起口，我口分口用$\rho_E^j(s,a)$和$\mu_E^j(s)$表示$\rho_{\pi_E^j}^j(s,a)$ 和$\mu_{\pi_{E}^{j}}^{j}(s)$ 。口家演示$\tau_E^i$由口家策略$\pi_E^i$的口口-口作口$\overset{E}{\operatorname*{\operatorname*{\text{迹口出 ,口策略口口用于具有口口}}}}\mathcal{P}^i$ 的 $\mathcal{E}_i$ 演示口境$\mathcal{E}_i$。$D_ij(s,a):\mathcal{S}\times\mathcal{A}\to[0,1]$ 是一口口口器 , 用于口分口口-操作口$(s,a)$是口自口$\mathcal{E}_i$交互的策略$\pi$ 口是口自口家$\pi_E^j$ | [
[
"0.2467",
"0.05682",
"0.728",
"0.05682",
"0.728",
"0.0707",
"0.2467",
"0.0707"
],
[
"0.544",
"0.0859",
"0.5947",
"0.0859",
"0.5947",
"0.09784",
"0.544",
"0.09784"
],
[
"0.7363",
"0.0859",
"0.7876",
"0.0859",
"0.7876",
"0.09784",
"0.7363",
"0.09784"
],
[
"0.6953",
"0.09534",
"0.8286",
"0.09534",
"0.8286",
"0.1092",
"0.6953",
"0.1092"
],
[
"0.4387",
"0.089",
"0.4592",
"0.089",
"0.4592",
"0.09344",
"0.4387",
"0.09344"
],
[
"0.10864",
"0.101",
"0.1177",
"0.101",
"0.1177",
"0.108",
"0.10864",
"0.108"
],
[
"0.5063",
"0.09784",
"0.636",
"0.09784",
"0.636",
"0.108",
"0.5063",
"0.108"
],
[
"0.5923",
"0.1187",
"0.6177",
"0.1187",
"0.6177",
"0.137",
"0.5923",
"0.137"
],
[
"0.7036",
"0.11554",
"0.7607",
"0.11554",
"0.7607",
"0.1383",
"0.7036",
"0.1383"
],
[
"0.511",
"0.1395",
"0.5474",
"0.1364",
"0.55",
"0.1509",
"0.5146",
"0.154"
],
[
"0.569",
"0.1414",
"0.6055",
"0.1414",
"0.6055",
"0.1528",
"0.569",
"0.1528"
],
[
"0.7134",
"0.1414",
"0.7466",
"0.1414",
"0.7466",
"0.1528",
"0.7134",
"0.1528"
],
[
"0.705",
"0.1578",
"0.7534",
"0.1578",
"0.7534",
"0.178",
"0.705",
"0.178"
],
[
"0.572",
"0.1768",
"0.607",
"0.1799",
"0.6055",
"0.1945",
"0.5703",
"0.1925"
],
[
"0.835",
"0.178",
"0.8613",
"0.178",
"0.8613",
"0.1901",
"0.835",
"0.1901"
],
[
"0.3245",
"0.1787",
"0.4453",
"0.1787",
"0.4453",
"0.1863",
"0.3245",
"0.1863"
],
[
"0.1708",
"0.1787",
"0.255",
"0.1787",
"0.255",
"0.1863",
"0.1708",
"0.1863"
],
[
"0.6177",
"0.1799",
"0.625",
"0.1799",
"0.625",
"0.1849",
"0.6177",
"0.1849"
],
[
"0.653",
"0.1799",
"0.664",
"0.1863",
"0.6567",
"0.1925",
"0.6465",
"0.1875"
],
[
"0.7173",
"0.1818",
"0.744",
"0.1818",
"0.744",
"0.1932",
"0.7173",
"0.1932"
],
[
"0.3064",
"0.1901",
"0.4468",
"0.1913",
"0.4468",
"0.2058",
"0.3064",
"0.2052"
],
[
"0.5156",
"0.1812",
"0.535",
"0.1812",
"0.535",
"0.1945",
"0.5156",
"0.1945"
],
[
"0.1095",
"0.1913",
"0.2646",
"0.1913",
"0.2646",
"0.2058",
"0.1095",
"0.2058"
],
[
"0.5107",
"0.1945",
"0.5415",
"0.1995",
"0.5337",
"0.2279",
"0.5024",
"0.2229"
],
[
"0.1478",
"0.2052",
"0.2214",
"0.2052",
"0.2214",
"0.2166",
"0.1478",
"0.2166"
],
[
"0.3367",
"0.2052",
"0.4102",
"0.2052",
"0.4102",
"0.2166",
"0.3367",
"0.2166"
],
[
"0.706",
"0.202",
"0.734",
"0.202",
"0.734",
"0.2197",
"0.706",
"0.2197"
],
[
"0.10706",
"0.2235",
"0.1209",
"0.2235",
"0.1209",
"0.2329",
"0.10706",
"0.2329"
],
[
"0.5186",
"0.2247",
"0.5337",
"0.2247",
"0.5337",
"0.2349",
"0.5186",
"0.2349"
],
[
"0.574",
"0.2247",
"0.6055",
"0.2247",
"0.6055",
"0.2361",
"0.574",
"0.2361"
],
[
"0.7134",
"0.2267",
"0.74",
"0.2235",
"0.7427",
"0.2343",
"0.716",
"0.2374"
],
[
"0.7573",
"0.2216",
"0.7656",
"0.2216",
"0.7656",
"0.2267",
"0.7573",
"0.2267"
],
[
"0.10706",
"0.2349",
"0.1177",
"0.2349",
"0.1177",
"0.2443",
"0.10706",
"0.2443"
],
[
"0.10297",
"0.254",
"0.10864",
"0.254",
"0.10864",
"0.2615",
"0.10297",
"0.2615"
],
[
"0.501",
"0.2632",
"0.886",
"0.2632",
"0.886",
"0.2778",
"0.501",
"0.2778"
],
[
"0.10297",
"0.2715",
"0.10864",
"0.2715",
"0.10864",
"0.279",
"0.10297",
"0.279"
],
[
"0.501",
"0.279",
"0.886",
"0.279",
"0.886",
"0.2935",
"0.501",
"0.2935"
],
[
"0.4983",
"0.2903",
"0.818",
"0.2942",
"0.818",
"0.3125",
"0.4983",
"0.308"
],
[
"0.1316",
"0.3125",
"0.2566",
"0.3132",
"0.2566",
"0.324",
"0.1316",
"0.3225"
],
[
"0.3203",
"0.3125",
"0.4468",
"0.3125",
"0.4468",
"0.3225",
"0.3203",
"0.3225"
],
[
"0.1111",
"0.3276",
"0.2632",
"0.3276",
"0.2632",
"0.3416",
"0.1111",
"0.3416"
],
[
"0.3064",
"0.3257",
"0.4468",
"0.3276",
"0.4468",
"0.3416",
"0.3064",
"0.339"
],
[
"0.134",
"0.3416",
"0.2336",
"0.3416",
"0.2336",
"0.3516",
"0.134",
"0.3516"
],
[
"0.3228",
"0.3403",
"0.423",
"0.3416",
"0.423",
"0.3528",
"0.3228",
"0.3516"
],
[
"0.5",
"0.351",
"0.819",
"0.351",
"0.819",
"0.3655",
"0.5",
"0.3655"
],
[
"0.08905",
"0.358",
"0.4756",
"0.358",
"0.4756",
"0.3718",
"0.08905",
"0.3718"
],
[
"0.08905",
"0.3726",
"0.4756",
"0.3726",
"0.4756",
"0.3865",
"0.08905",
"0.3865"
],
[
"0.5215",
"0.3687",
"0.835",
"0.3694",
"0.835",
"0.3838",
"0.5215",
"0.3833"
],
[
"0.08905",
"0.3882",
"0.4739",
"0.3882",
"0.4739",
"0.4016",
"0.08905",
"0.4016"
],
[
"0.5",
"0.3933",
"0.693",
"0.3945",
"0.693",
"0.4092",
"0.5",
"0.4077"
],
[
"0.08905",
"0.4028",
"0.4739",
"0.4028",
"0.4739",
"0.4167",
"0.08905",
"0.4167"
],
[
"0.08905",
"0.4185",
"0.2484",
"0.4185",
"0.2484",
"0.432",
"0.08905",
"0.432"
],
[
"0.501",
"0.4172",
"0.8823",
"0.4185",
"0.8823",
"0.433",
"0.501",
"0.432"
],
[
"0.5034",
"0.433",
"0.884",
"0.433",
"0.884",
"0.4463",
"0.5034",
"0.4463"
],
[
"0.5024",
"0.449",
"0.884",
"0.449",
"0.884",
"0.4622",
"0.5024",
"0.4622"
],
[
"0.501",
"0.464",
"0.8857",
"0.464",
"0.8857",
"0.4778",
"0.501",
"0.4778"
],
[
"0.09235",
"0.476",
"0.4739",
"0.476",
"0.4739",
"0.4905",
"0.09235",
"0.4905"
],
[
"0.4983",
"0.4792",
"0.8823",
"0.4792",
"0.8823",
"0.4924",
"0.4983",
"0.4924"
],
[
"0.09235",
"0.4905",
"0.4739",
"0.4905",
"0.4739",
"0.505",
"0.09235",
"0.505"
],
[
"0.501",
"0.4937",
"0.8823",
"0.4937",
"0.8823",
"0.5083",
"0.501",
"0.5083"
],
[
"0.0866",
"0.505",
"0.2181",
"0.506",
"0.2181",
"0.521",
"0.0866",
"0.5195"
],
[
"0.2932",
"0.507",
"0.3186",
"0.507",
"0.3186",
"0.518",
"0.2932",
"0.518"
],
[
"0.3857",
"0.507",
"0.4248",
"0.507",
"0.4248",
"0.518",
"0.3857",
"0.518"
],
[
"0.501",
"0.509",
"0.884",
"0.509",
"0.884",
"0.523",
"0.501",
"0.523"
],
[
"0.1978",
"0.5215",
"0.2222",
"0.5215",
"0.2222",
"0.534",
"0.1978",
"0.534"
],
[
"0.2418",
"0.5215",
"0.4756",
"0.5195",
"0.4756",
"0.536",
"0.2418",
"0.5386"
],
[
"0.5",
"0.5215",
"0.8857",
"0.5215",
"0.8857",
"0.539",
"0.5",
"0.539"
],
[
"0.08826",
"0.535",
"0.4739",
"0.536",
"0.4739",
"0.5503",
"0.08826",
"0.55"
],
[
"0.5024",
"0.539",
"0.8857",
"0.539",
"0.8857",
"0.553",
"0.5024",
"0.553"
],
[
"0.0907",
"0.552",
"0.4722",
"0.552",
"0.4722",
"0.565",
"0.0907",
"0.565"
],
[
"0.5024",
"0.5547",
"0.886",
"0.5547",
"0.886",
"0.5684",
"0.5024",
"0.5684"
],
[
"0.0907",
"0.5664",
"0.4714",
"0.5664",
"0.4714",
"0.58",
"0.0907",
"0.58"
],
[
"0.501",
"0.5684",
"0.8677",
"0.5684",
"0.8677",
"0.582",
"0.501",
"0.582"
],
[
"0.0907",
"0.5835",
"0.2115",
"0.5835",
"0.2115",
"0.5933",
"0.0907",
"0.5933"
],
[
"0.5034",
"0.5854",
"0.8857",
"0.5854",
"0.8857",
"0.5986",
"0.5034",
"0.5986"
],
[
"0.501",
"0.5986",
"0.8857",
"0.5996",
"0.8857",
"0.6143",
"0.501",
"0.613"
],
[
"0.0907",
"0.605",
"0.4722",
"0.605",
"0.4722",
"0.6187",
"0.0907",
"0.6187"
],
[
"0.08826",
"0.6187",
"0.4722",
"0.619",
"0.4722",
"0.634",
"0.08826",
"0.6333"
],
[
"0.5",
"0.613",
"0.886",
"0.6143",
"0.886",
"0.629",
"0.5",
"0.6274"
],
[
"0.09235",
"0.6353",
"0.4739",
"0.6353",
"0.4739",
"0.6484",
"0.09235",
"0.6484"
],
[
"0.501",
"0.63",
"0.886",
"0.63",
"0.886",
"0.643",
"0.501",
"0.643"
],
[
"0.5024",
"0.6445",
"0.884",
"0.6445",
"0.884",
"0.6577",
"0.5024",
"0.6577"
],
[
"0.0907",
"0.651",
"0.4739",
"0.651",
"0.4739",
"0.664",
"0.0907",
"0.664"
],
[
"0.0907",
"0.664",
"0.4722",
"0.664",
"0.4722",
"0.678",
"0.0907",
"0.678"
],
[
"0.5024",
"0.66",
"0.8857",
"0.66",
"0.8857",
"0.674",
"0.5024",
"0.674"
],
[
"0.08905",
"0.681",
"0.4739",
"0.681",
"0.4739",
"0.6943",
"0.08905",
"0.6943"
],
[
"0.5024",
"0.6753",
"0.8857",
"0.6753",
"0.8857",
"0.6895",
"0.5024",
"0.6895"
],
[
"0.0907",
"0.696",
"0.4739",
"0.696",
"0.4739",
"0.709",
"0.0907",
"0.709"
],
[
"0.5",
"0.69",
"0.886",
"0.69",
"0.886",
"0.704",
"0.5",
"0.704"
],
[
"0.09235",
"0.7104",
"0.4722",
"0.7104",
"0.4722",
"0.7236",
"0.09235",
"0.7236"
],
[
"0.5",
"0.7056",
"0.886",
"0.7056",
"0.886",
"0.7197",
"0.5",
"0.7197"
],
[
"0.08905",
"0.7246",
"0.4722",
"0.7246",
"0.4722",
"0.738",
"0.08905",
"0.738"
],
[
"0.5024",
"0.7217",
"0.556",
"0.7217",
"0.556",
"0.733",
"0.5024",
"0.733"
],
[
"0.0907",
"0.7407",
"0.4697",
"0.7407",
"0.4697",
"0.755",
"0.0907",
"0.755"
],
[
"0.5024",
"0.7437",
"0.818",
"0.7437",
"0.818",
"0.7583",
"0.5024",
"0.7583"
],
[
"0.09235",
"0.756",
"0.4739",
"0.756",
"0.4739",
"0.7695",
"0.09235",
"0.7695"
],
[
"0.5034",
"0.7603",
"0.5664",
"0.7603",
"0.5664",
"0.7715",
"0.5034",
"0.7715"
],
[
"0.581",
"0.7583",
"0.8857",
"0.7583",
"0.8857",
"0.773",
"0.581",
"0.773"
],
[
"0.09235",
"0.7715",
"0.4722",
"0.7715",
"0.4722",
"0.785",
"0.09235",
"0.785"
],
[
"0.4983",
"0.7715",
"0.739",
"0.768",
"0.739",
"0.787",
"0.4983",
"0.7905"
],
[
"0.732",
"0.7734",
"0.8857",
"0.7715",
"0.8857",
"0.786",
"0.732",
"0.7886"
],
[
"0.0907",
"0.786",
"0.4714",
"0.786",
"0.4714",
"0.8",
"0.0907",
"0.8"
],
[
"0.5024",
"0.7886",
"0.8857",
"0.7886",
"0.8857",
"0.802",
"0.5024",
"0.802"
],
[
"0.08905",
"0.8003",
"0.4739",
"0.8",
"0.4739",
"0.8145",
"0.08905",
"0.815"
],
[
"0.501",
"0.803",
"0.886",
"0.8037",
"0.886",
"0.8184",
"0.501",
"0.8174"
],
[
"0.09235",
"0.8164",
"0.4722",
"0.8164",
"0.4722",
"0.83",
"0.09235",
"0.83"
],
[
"0.09235",
"0.8306",
"0.4722",
"0.8306",
"0.4722",
"0.8447",
"0.09235",
"0.8447"
],
[
"0.5024",
"0.8257",
"0.8857",
"0.8257",
"0.8857",
"0.8403",
"0.5024",
"0.8403"
],
[
"0.0907",
"0.8447",
"0.4739",
"0.845",
"0.4739",
"0.86",
"0.0907",
"0.8594"
],
[
"0.5024",
"0.845",
"0.8857",
"0.845",
"0.8857",
"0.86",
"0.5024",
"0.86"
],
[
"0.0907",
"0.8613",
"0.4739",
"0.8613",
"0.4739",
"0.875",
"0.0907",
"0.875"
],
[
"0.501",
"0.8613",
"0.888",
"0.8613",
"0.888",
"0.875",
"0.501",
"0.875"
],
[
"0.0907",
"0.877",
"0.478",
"0.877",
"0.478",
"0.89",
"0.0907",
"0.89"
],
[
"0.5024",
"0.877",
"0.8857",
"0.877",
"0.8857",
"0.89",
"0.5024",
"0.89"
],
[
"0.09235",
"0.893",
"0.4192",
"0.893",
"0.4192",
"0.904",
"0.09235",
"0.904"
],
[
"0.501",
"0.8896",
"0.6235",
"0.8926",
"0.6235",
"0.907",
"0.501",
"0.904"
]
] | [
"Robust Imitation Learning against Variations in Environment Dynamics",
"Phase O:",
"Phase1:",
"Training the Agent's Policy",
"CWL+D",
"000",
"GatheringDemonstrations",
"丰",
"雪祥",
"ninte",
"Env3",
"Env3",
"",
"Env2",
"Duit",
"ThrePerrentacettEnvirnnnsot Perturat",
"areEnvirnn1entPecture0",
"",
"",
"Env2",
"(b) Walker2d+Mass:",
"元",
"(a) Walker2d+Gravity:",
"",
"GAIL+GP",
"GAIL+GP",
"",
"1000",
"元",
"Env1",
"Env",
"0.",
"S000",
"",
"Figure 2: Overview of our algorithm. The blue line is",
"10n",
"the flow of policy samples p , and the green line is",
"the flow of expert demonstrations ~ pe ",
"ThePercentageofEnvironmentPerturb",
"ThePercentaEmuiroPeu",
"(c) Walker2d+Gravity:",
"(d) Walker2d+Mass:",
"SNEMPE-max",
"SNEMPE-max",
"5. Robust Imitation Learning against",
"Figure l: Performance of the policy trained in a sin-",
"gle nominal environment against perturbation (left col-",
"Variations in Environment Dynamics",
"umn - gravity perturbation and right column - mass",
"5.1. Problem Formulation",
"perturbation): (a,b)- performance of GAIL+GP, and",
"(c,d) - SNEMPE-max.",
"我考MDP的集合℃={M=<S,A,Ps，r,>",
"CEZ其中口和口作空口口于集合的所有成口",
"都是相同的，口口口代理不可用，用口口口口力口建",
"模的口渡口率口行口口化，口且口力口口口可以口",
"口了口理口口性能下降，强大的RL口一些具有口口",
"XMTHMAX4集的O口Z口O口化或口口。在此",
"的口境口行朵口。例如，在Figure1c中，口重力口",
"集合中，我口具有1，2···SN的M",
"50%150%100%",
"RL",
"MPE",
"MDP口行朵口。我口用&1，·，&N表示口些具有口口",
"RL",
"maxminpiepE,[GtPi],其中",
"Ps1,.,Ps（Pl,.,PN）的N境。我假",
"P={Pi是口定的口境集。通口最大化最口情口下的",
"口每口境都有一口口家，口家元口代理生成",
"口期收盆，代理的策略可以捕口所口口境集P=【Pi】",
"家演示，但是口家策略元本身口代理不可用。我口",
"中口化的口口。但是强大的RL需要精心口的口口函",
"口假口代理可以口所有朵口口境&..·.&中的每",
"口我口希望避免",
"口口行交互，口且所有交互境的初始口口分布口μo（s）",
"相同。因此，根据我口在上一口中的定口，1，··，&",
"口在，考口棒IL。一口口的方法是口上述口棒RL原",
"理口用于IL口置。在口里，我口口多口朵口的演示口境",
"是演示口境又是交互口境，口置口MPE/MPE。口注意，",
"中口得口家演示，口使口口策略口口口口口交互口境口",
"&1，··，&N上的口家演示是口于整口MDP集合C的部",
"分信息。我口的目口是口代理口口一口策略元，口策略",
"行交互。然后，我口使用口口器口口分口自多口抽口口",
"家口口中的每一口的政策口本，口口口政策遵循最口",
"口于MDP集合C中的所有成口都表口良好，口基于口",
"的情口，口。e。，基于口口器口出的口策略口本最口的",
"自采口口境&，··，的口家演示和代理口朵口境",
"口家口口。Figures1cand1d中口示了在受千口的口",
"&，..·,&的交互。我此口RobustImita-",
"■口境中如此口口的策略的性能（相口的性能表示口",
"tion learning with Multiple perturbed Environments",
"SNEMPE-max)。可以看出，口Figures la and 1b 中的",
"(RIME)",
"常口SNE/SNEIL相比，以口口方式口口的策略提高了",
"口我口再介口口口符口。p（s,a)",
"口棒性，但性能口不今人口意。口口退化意味着口口口",
"π(as)",
"oPr（st=s元，P）表示元在交互口",
"口口口境的策略交互不足以捕口口口化，口使口自多",
"境&中的占用度量。（s)=2。",
"t=o Pr(st = s元, Pi)",
"口朵口演示口境的口家演示也是如此。因此，口了充分",
"表示元在交互口境中的非口一化口口口。口口",
"捕捉口口口化，我口首先口口口口力口分布中朵口一些",
"起，我分口用（s,a）和（s）表示p（s,a)",
"具有不同口力口的口境，口口口些多口口口境不口用",
"于口家演示，口用于口口期口的策略交互。我口口此口",
"和（s）口家演示由口家策略元的口口",
"置口口MPE/MPEIL口置。在本文的其余部分中，我口",
"作口口迹口出，口策略口口用于具有口口Pi的",
"提出了一口基于MPE/MPEIL口置的高效IL口架以口",
"演示境&。Dii(s,a):S×A→[0,1]是一□口器，",
"生一口策略，口策略口基于少口用于口口的朵口口口，",
"用于口分口口操作口（s,α）是口自口&交互的策略π",
"可口口口口的口境口口口化口行口健的图行",
"口是口自口家元"
] | {
"arxiv_id": "2206.09314",
"language": "zh",
"page_idx": 2
} |
||
2206.09314-zh-4 | 2206.09314-zh-4.jpg | Robust Imitation Learning against Variations in Environment Dynamics
$\square(7)$ 中的第一$\square\mathbb{E}_{\rho_\pi^i}[\cdot]$。在第一口中，我口有加口因子痲口。因此，利用$\sum_{j=1}^N\lambda_j(s)=1$的事口，我口可以根
首先，考口口口定$D_ij$的$\pi$口行口化。口注意，$\pi$口影据期望的口性度$\square\sum_{j=1}^N$推入期望 ,口而重口口定 $D_i_j$ 的第一口，口口得其上限
$$\operatorname*{min}_{\pi}\sum_{i=1}^{N}\mathbb{E}_{\rho_{\pi}^{i}}\:\left[\sum_{j=1}^{N}\lambda_{j}(s)\log(1-D_{ij}(s,a))\right]\\\leq\operatorname*{min}_{\pi}\sum_{i=1}^{N}\mathbb{E}_{\rho_{\pi}^{i}}\left[\operatorname*{max}_{j}\log(1-D_{ij}(s,a))\right],$$
(8)
,其中$\mathbb{E}_{\rho_\pi^i}[\cdot]$ 表示$\mathbb{E}_(s,a)\sim\rho_\pi^i[\cdot]$ ,口且不等式被口口是有
效的。然后，我口口化策略$\pi$的目口函口(8)的上界
接下口，考口口定$\pi$的$D_ij$的口化。此口化由于以下定
理而口化：
Theorem 5.2. The following maximization problem without the $\lambda_j(s)$ term has the same optimal solution for $D_{ij}$ as (7) with $\mu_{\pi}^{i}(s)/\mu_{E}^{j}(s)$ set to 1 for given $\pi:$
$$\max_{D_{ij}}\left\{\mathbb{E}_{\rho_{\pi}^{i}}\left[\log(1-D_{ij}(s,a))\right]+\mathbb{E}_{\rho_{E}^{j}}\left[\log(D_{ij}(s,a))\right]\right\}.$$
Proof. See in Appendix A.2
口
基于 Theorem 5.2 和梯度口口(GP),我口最口得出口
$\overline{{\text{定 }\pi}}$的 $D_ij$ 的目口函口如下：
$$\begin{aligned}\max_{D_{ij}}&\left\{\mathbb{E}_{\rho_{\pi}^{i}}\left[\log(1-D_{ij}(s,a))\right]+\mathbb{E}_{\tau_{E}^{j}}\left[\log(D_{ij}(s,a))\right]\\&+\kappa\mathbb{E}_{\hat{x}}\left(\|\nabla_{\hat{x}}D_{ij}(\hat{x})\|_{2}-1\right)^{2}\right\},\end{aligned}$$
(9)
。口注意 ,在 (9) 中 ,我口添加了 Section 3.2中提到的梯
度口口口，用于口定口口，口口$\mathbb{E}_\rho_E^j$ 替$\square\square\mathbb{E}_\tau_E^j$
口注意，口口器$D_{ij}$的口量由$N^2$口出 ,口且口着口境的$N$口量的增加而二次增加。我□可以通口使用口口器口重共享口口少口口口字 ,口使得口口器模型共享其$\square$重的子集 (Liu & Tuzel, 2016)。口口器 $D_i1,\cdots,D_{iN}$ 共享其口人口和口藏口的口重 ,因此可以口口口口口口具有$N$口出口点的一口口口器。我口口之口口重共享$\square \square 器 ($WSD$) 。\square 于$WSD $D_{i}^{\mathrm{Weight- Shared}}, 其 N\square$出口点的第$j$口口出口口于$D_ij$的口出 ,其目口由$\sum_jV_{ij}$ $\begin{array}{l}{\square\text{出，其中}V_{ij}\text{ 是 }D_{ij}\text{ 在 (9) 中的口口目口。使用WSDs}}\\{D_{i}^{\mathrm{Weight-Shared}},i=1,\cdots,N\text{ ,口口器的口口性降低},}\\\end{array}$ $\square$乎是$\sim N$
5.4. Comparison with Occupancy Measure Matching 口使在MPE的情口下不保口口占用度量中恢口策略$\overline{\text{分口，我口仍然可以口占用度量匹配技口口用于MPE/}}$ MPE。在口口情口下，
(10)
$$\min_{\pi}\sum_{j=1}^{N}\lambda_{j}\mathcal{D}_{JS}(\bar{\rho}_{\pi},\bar{\rho}_{E}^{j}),$$
$\square$出了一口合理的目口，其中$\sum_j\lambda_j=1,\bar{\rho}_\pi$ 和$\bar{\rho}_E^j$ 是分口口$\rho_{\pi}$和$\rho_E^{j}$口得的口一化 occupancy 分布。(
$\lambda_{j}(s)$ ,使得$\sum_j=1^{N}\lambda_{j}(s)=1$ ,口且口定适口的 $\lambda_j(s)\square$ Section 6中口考口了其他目口。)然后 ,我口可以得出
(10)的上限$,\overline{\text{如下所示}}:$
$$\begin{aligned}&\operatorname*{min}_{\pi}\sum_{j=1}^{N}\lambda_{j}\mathcal{D}_{JS}(\bar{\rho}_{\pi},\bar{\rho}_{E}^{j})\\&\leq\operatorname*{min}_{\pi}\sum_{i=1}^{N}\sum_{j=1}^{N}\frac{\lambda_{j}(1-\gamma)}{2N}\operatorname*{max}_{D_{ij}}\left\{\mathbb{E}_{\rho_{\pi}^{i}}\left[\log(1-D_{ij}(s,a))\right]\right.\\&+\:\mathbb{E}_{\rho_{E}^{j}}\left[\log D_{ij}(s,a)\right]\biggr\}+\log2,\end{aligned}$$
其中口上限的推口在 Appendix B.2 中。口在考口在口$\square$情口下口口口口定的$D_ij$ 口化 $\pi$。同口，口了口理 $\lambda_j$, 我$\square$可以用$\max_j$ 替$\square\sum_j\lambda_j$ 以口生口一口上限。然后， 口定$D_{ij}$的$\pi$的目口函口由下式口出
$$\min_\pi\sum_{i=1}^N\max_j\mathbb{E}_{\rho_\pi^i}[\log(1-D_{ij}(s,a))].$$
(11)
我口口此方法口口多口境中的占用度量匹配(OMME)。目$\square(11)\square(8)$ 中提出的目口的口口$\square\square 在于，操作\max_j$超出了口期$\mathbb{E}_\rho_\pi^i[\cdot]$。口注意，由于$\max_j$是非口性口算 ,因此口序不可互口。我口口看到 , 口口看似微小的差口在 Section 6 中口生了口着的性能差口。
# 6. Experiments
6.1.Experimental Settings
我口口我口的算法口以下基口一起考口：
-行口克隆(BC):通口口督$\square\square\square\square\square\square\square\square$策略，直到所有
口家演示的口口口口停止口少口止。
- GAIL- mixture: $\square$是 MPE 的 GAIL+GP 的$\square\square\square$ 在口口情口下，我口有一口口口器，$\square 且口口$ 口器口分所有$\bar{\rho}_\pi^i$和所有$\bar{\rho}_E^j$。$\pi$的目口函口是$\operatorname* { min} _{\pi }{\mathcal{D} }_{JS}( \sum _{i}\bar{\rho } _{\pi }^{i}/ N, \sum _{j}\bar{\rho } _{E}^{j}/ N)$ .
-GAIL-single$:\square$是MPE的GAIL+GP的$\square-\square\square\square\square$ 在口口情口下，我口有多口口口器，$\pi$的目口函口是$\operatorname*{min}_{\pi}\sum_{i}\mathcal{D}_{JS}(\bar{\rho}_{\pi}^{i},\bar{\rho}_{E}^{i})$
-OMME(最接近我口的算法):$\square 已口描迹口了。目口$
函口由(10)和(11)口出
Appendix F 中口口介口了基口、口口和口家演示。我口考口了提出的算法的口口版本：RIME和RIME+ WSD 。RIME和RIME+ WSD 之口的唯一口口是口口器$D_ij$的口口。RIME具有目口函口口(9)的口$\square 器\overline{D}_{ij}$ ,因此口口器口口的口量口$N^2$。口一方面 , $\begin{array}{l}{\mathrm{RIME+WSD}}&{\text{使用}}&{\mathrm{weight-shared}}&{\mathrm{discriminator}}\\{D_{i}^{\mathrm{Weight-Shared}}}&{\square5.3\text{ 末尾描迹的}}&{\mathrm{weight-shared}}&{\mathrm{dis-}}\\\end{array}$ $\begin{array}{l}{{{D_{i}}}}\\{{{\mathrm{criminator}}}}&{{D_{i}^{\mathrm{Weight-Shared}}}}\\\end{array}$
我口在MuJoCo任口上口口了所考口的算法：Hopper, Walker2d ,HalfCheetah和Ant (Todorov et al.,2012) 。 | [
[
"0.2484",
"0.05936",
"0.726",
"0.05936",
"0.726",
"0.06946",
"0.2484",
"0.06946"
],
[
"0.08905",
"0.0859",
"0.4739",
"0.0852",
"0.4739",
"0.0985",
"0.08905",
"0.09973"
],
[
"0.505",
"0.0852",
"0.8813",
"0.0852",
"0.8813",
"0.09973",
"0.505",
"0.09973"
],
[
"0.0907",
"0.101",
"0.4722",
"0.101",
"0.4722",
"0.11554",
"0.0907",
"0.11554"
],
[
"0.5024",
"0.1048",
"0.8857",
"0.1048",
"0.8857",
"0.1187",
"0.5024",
"0.1187"
],
[
"0.08905",
"0.1187",
"0.4756",
"0.1187",
"0.4756",
"0.1364",
"0.08905",
"0.1364"
],
[
"0.501",
"0.1193",
"0.884",
"0.1193",
"0.884",
"0.1332",
"0.501",
"0.1332"
],
[
"0.0866",
"0.1307",
"0.4756",
"0.1307",
"0.4756",
"0.1547",
"0.0866",
"0.1547"
],
[
"0.501",
"0.132",
"0.6665",
"0.132",
"0.6665",
"0.1459",
"0.501",
"0.1459"
],
[
"0.08417",
"0.1459",
"0.478",
"0.154",
"0.4763",
"0.1818",
"0.0825",
"0.1736"
],
[
"0.55",
"0.1572",
"0.564",
"0.1572",
"0.564",
"0.1666",
"0.55",
"0.1666"
],
[
"0.5146",
"0.1666",
"0.664",
"0.1674",
"0.664",
"0.1831",
"0.5146",
"0.1818"
],
[
"0.08905",
"0.1736",
"0.2632",
"0.1736",
"0.2632",
"0.1875",
"0.08905",
"0.1875"
],
[
"0.5435",
"0.1831",
"0.568",
"0.1799",
"0.5703",
"0.1913",
"0.546",
"0.1945"
],
[
"0.183",
"0.1976",
"0.1929",
"0.1976",
"0.1929",
"0.2052",
"0.183",
"0.2052"
],
[
"0.5605",
"0.1925",
"0.875",
"0.207",
"0.8716",
"0.2349",
"0.558",
"0.2203"
],
[
"0.1447",
"0.202",
"0.2279",
"0.2052",
"0.2263",
"0.2229",
"0.1438",
"0.2197"
],
[
"0.2378",
"0.1989",
"0.2484",
"0.1989",
"0.2484",
"0.2026",
"0.2378",
"0.2026"
],
[
"0.2507",
"0.207",
"0.4084",
"0.207",
"0.4084",
"0.2203",
"0.2507",
"0.2203"
],
[
"0.5156",
"0.209",
"0.607",
"0.2058",
"0.609",
"0.2247",
"0.517",
"0.2279"
],
[
"0.1772",
"0.2229",
"0.1969",
"0.2229",
"0.1969",
"0.2299",
"0.1772",
"0.2299"
],
[
"0.2303",
"0.2229",
"0.25",
"0.2229",
"0.25",
"0.2299",
"0.2303",
"0.2299"
],
[
"0.615",
"0.2184",
"0.669",
"0.2184",
"0.669",
"0.2299",
"0.615",
"0.2299"
],
[
"0.5645",
"0.2247",
"0.6094",
"0.2247",
"0.6094",
"0.2349",
"0.5645",
"0.2349"
],
[
"0.681",
"0.2235",
"0.706",
"0.2235",
"0.706",
"0.2279",
"0.681",
"0.2279"
],
[
"0.1978",
"0.2361",
"0.2133",
"0.2374",
"0.2115",
"0.2468",
"0.1969",
"0.2456"
],
[
"0.1478",
"0.2468",
"0.4004",
"0.2468",
"0.4004",
"0.2644",
"0.1478",
"0.2644"
],
[
"0.591",
"0.2424",
"0.789",
"0.238",
"0.789",
"0.2588",
"0.591",
"0.2632"
],
[
"0.455",
"0.25",
"0.4739",
"0.25",
"0.4739",
"0.262",
"0.455",
"0.262"
],
[
"0.1936",
"0.2644",
"0.2181",
"0.2644",
"0.2181",
"0.2715",
"0.1936",
"0.2715"
],
[
"0.2418",
"0.26",
"0.2656",
"0.26",
"0.2656",
"0.2676",
"0.2418",
"0.2676"
],
[
"0.5063",
"0.2747",
"0.8857",
"0.2727",
"0.8857",
"0.286",
"0.5063",
"0.2886"
],
[
"0.0989",
"0.281",
"0.4739",
"0.281",
"0.4739",
"0.2986",
"0.0989",
"0.2986"
],
[
"0.5024",
"0.2886",
"0.888",
"0.289",
"0.888",
"0.3037",
"0.5024",
"0.303"
],
[
"0.08905",
"0.2954",
"0.4478",
"0.2966",
"0.4478",
"0.3113",
"0.08905",
"0.31"
],
[
"0.5024",
"0.3037",
"0.8896",
"0.3037",
"0.8896",
"0.318",
"0.5024",
"0.318"
],
[
"0.0907",
"0.3213",
"0.4722",
"0.3213",
"0.4722",
"0.332",
"0.0907",
"0.332"
],
[
"0.5024",
"0.3208",
"0.7646",
"0.3208",
"0.7646",
"0.334",
"0.5024",
"0.334"
],
[
"0.0907",
"0.336",
"0.1626",
"0.336",
"0.1626",
"0.348",
"0.0907",
"0.348"
],
[
"0.0907",
"0.351",
"0.4739",
"0.351",
"0.4739",
"0.3655",
"0.0907",
"0.3655"
],
[
"0.609",
"0.351",
"0.6235",
"0.351",
"0.6235",
"0.3599",
"0.609",
"0.3599"
],
[
"0.5728",
"0.3599",
"0.812",
"0.361",
"0.812",
"0.379",
"0.5728",
"0.3774"
],
[
"0.8555",
"0.3623",
"0.8857",
"0.3623",
"0.8857",
"0.375",
"0.8555",
"0.375"
],
[
"0.6045",
"0.3774",
"0.634",
"0.3774",
"0.634",
"0.3865",
"0.6045",
"0.3865"
],
[
"0.08826",
"0.363",
"0.4722",
"0.363",
"0.4722",
"0.3774",
"0.08826",
"0.3774"
],
[
"0.08905",
"0.3806",
"0.4673",
"0.3806",
"0.4673",
"0.3958",
"0.08905",
"0.3958"
],
[
"0.5156",
"0.3945",
"0.8857",
"0.3945",
"0.8857",
"0.4077",
"0.5156",
"0.4077"
],
[
"0.10706",
"0.406",
"0.4468",
"0.406",
"0.4468",
"0.4236",
"0.10706",
"0.4236"
],
[
"0.5034",
"0.4104",
"0.884",
"0.4104",
"0.884",
"0.4236",
"0.5034",
"0.4236"
],
[
"0.1111",
"0.4172",
"0.1544",
"0.4172",
"0.1544",
"0.428",
"0.1111",
"0.428"
],
[
"0.5034",
"0.4255",
"0.8823",
"0.4255",
"0.8823",
"0.4395",
"0.5034",
"0.4395"
],
[
"0.08905",
"0.443",
"0.2917",
"0.443",
"0.2917",
"0.4583",
"0.08905",
"0.4583"
],
[
"0.4575",
"0.443",
"0.4756",
"0.443",
"0.4756",
"0.457",
"0.4575",
"0.457"
],
[
"0.5024",
"0.44",
"0.886",
"0.44",
"0.886",
"0.4539",
"0.5024",
"0.4539"
],
[
"0.5024",
"0.4558",
"0.8857",
"0.4558",
"0.8857",
"0.4697",
"0.5024",
"0.4697"
],
[
"0.0907",
"0.4734",
"0.4739",
"0.4734",
"0.4739",
"0.4873",
"0.0907",
"0.4873"
],
[
"0.5024",
"0.4717",
"0.54",
"0.4717",
"0.54",
"0.4849",
"0.5024",
"0.4849"
],
[
"0.08905",
"0.4849",
"0.298",
"0.486",
"0.298",
"0.5005",
"0.08905",
"0.5"
],
[
"0.11273",
"0.511",
"0.4495",
"0.5083",
"0.4495",
"0.529",
"0.11273",
"0.532"
],
[
"0.501",
"0.506",
"0.6357",
"0.506",
"0.6357",
"0.521",
"0.501",
"0.521"
],
[
"0.501",
"0.532",
"0.6978",
"0.532",
"0.6978",
"0.5454",
"0.501",
"0.5454"
],
[
"0.1602",
"0.5415",
"0.339",
"0.5415",
"0.339",
"0.556",
"0.1602",
"0.556"
],
[
"0.455",
"0.5425",
"0.4739",
"0.5425",
"0.4739",
"0.556",
"0.455",
"0.556"
],
[
"0.5024",
"0.5547",
"0.786",
"0.5547",
"0.786",
"0.5684",
"0.5024",
"0.5684"
],
[
"0.09644",
"0.5684",
"0.4739",
"0.5684",
"0.4739",
"0.582",
"0.09644",
"0.582"
],
[
"0.5",
"0.5684",
"0.8857",
"0.5693",
"0.8857",
"0.584",
"0.5",
"0.5835"
],
[
"0.0866",
"0.5757",
"0.4207",
"0.582",
"0.4207",
"0.604",
"0.0866",
"0.5977"
],
[
"0.5024",
"0.5854",
"0.7466",
"0.5854",
"0.7466",
"0.5986",
"0.5024",
"0.5986"
],
[
"0.5",
"0.5986",
"0.88",
"0.5996",
"0.88",
"0.6143",
"0.5",
"0.613"
],
[
"0.0907",
"0.6094",
"0.4722",
"0.6094",
"0.4722",
"0.6235",
"0.0907",
"0.6235"
],
[
"0.501",
"0.6157",
"0.884",
"0.6157",
"0.884",
"0.629",
"0.501",
"0.629"
],
[
"0.08905",
"0.625",
"0.4722",
"0.625",
"0.4722",
"0.638",
"0.08905",
"0.638"
],
[
"0.501",
"0.631",
"0.884",
"0.631",
"0.884",
"0.6445",
"0.501",
"0.6445"
],
[
"0.09235",
"0.6416",
"0.4722",
"0.6416",
"0.4722",
"0.6514",
"0.09235",
"0.6514"
],
[
"0.5",
"0.646",
"0.732",
"0.646",
"0.732",
"0.66",
"0.5",
"0.66"
],
[
"0.08905",
"0.654",
"0.4722",
"0.655",
"0.4722",
"0.6694",
"0.08905",
"0.6685"
],
[
"0.5",
"0.662",
"0.88",
"0.662",
"0.88",
"0.677",
"0.5",
"0.677"
],
[
"0.08905",
"0.6685",
"0.4722",
"0.6694",
"0.4722",
"0.6846",
"0.08905",
"0.683"
],
[
"0.501",
"0.677",
"0.8823",
"0.677",
"0.8823",
"0.6914",
"0.501",
"0.6914"
],
[
"0.09235",
"0.685",
"0.4722",
"0.685",
"0.4722",
"0.6987",
"0.09235",
"0.6987"
],
[
"0.3816",
"0.6978",
"0.4722",
"0.6997",
"0.4722",
"0.714",
"0.3816",
"0.712"
],
[
"0.5",
"0.6924",
"0.661",
"0.6934",
"0.661",
"0.7085",
"0.5",
"0.707"
],
[
"0.08826",
"0.6987",
"0.3774",
"0.6914",
"0.3774",
"0.712",
"0.08826",
"0.7197"
],
[
"0.4983",
"0.707",
"0.8857",
"0.7085",
"0.8857",
"0.7227",
"0.4983",
"0.7217"
],
[
"0.08905",
"0.714",
"0.4714",
"0.7153",
"0.4714",
"0.73",
"0.08905",
"0.729"
],
[
"0.501",
"0.7227",
"0.671",
"0.7227",
"0.671",
"0.7363",
"0.501",
"0.7363"
],
[
"0.0907",
"0.732",
"0.4722",
"0.732",
"0.4722",
"0.7456",
"0.0907",
"0.7456"
],
[
"0.08905",
"0.747",
"0.1871",
"0.7407",
"0.1887",
"0.755",
"0.0907",
"0.761"
],
[
"0.1978",
"0.7505",
"0.468",
"0.7505",
"0.468",
"0.761",
"0.1978",
"0.761"
],
[
"0.5024",
"0.7456",
"0.88",
"0.7456",
"0.88",
"0.7603",
"0.5024",
"0.7603"
],
[
"0.4656",
"0.7524",
"0.4739",
"0.7524",
"0.4739",
"0.7583",
"0.4656",
"0.7583"
],
[
"0.08905",
"0.763",
"0.1748",
"0.763",
"0.1748",
"0.778",
"0.08905",
"0.778"
],
[
"0.501",
"0.761",
"0.8857",
"0.7603",
"0.8857",
"0.7744",
"0.501",
"0.776"
],
[
"0.501",
"0.7764",
"0.8857",
"0.7764",
"0.8857",
"0.7905",
"0.501",
"0.7905"
],
[
"0.08826",
"0.7944",
"0.4714",
"0.797",
"0.4714",
"0.81",
"0.08826",
"0.808"
],
[
"0.5024",
"0.7915",
"0.8857",
"0.7915",
"0.8857",
"0.806",
"0.5024",
"0.806"
],
[
"0.5024",
"0.806",
"0.888",
"0.806",
"0.888",
"0.821",
"0.5024",
"0.821"
],
[
"0.09235",
"0.8184",
"0.4714",
"0.8184",
"0.4714",
"0.829",
"0.09235",
"0.829"
],
[
"0.5034",
"0.8228",
"0.562",
"0.8228",
"0.562",
"0.8335",
"0.5034",
"0.8335"
],
[
"0.572",
"0.8213",
"0.8857",
"0.8213",
"0.8857",
"0.8354",
"0.572",
"0.8354"
],
[
"0.09235",
"0.829",
"0.4722",
"0.829",
"0.4722",
"0.8433",
"0.09235",
"0.8433"
],
[
"0.5",
"0.832",
"0.609",
"0.8276",
"0.6094",
"0.8467",
"0.501",
"0.851"
],
[
"0.601",
"0.8354",
"0.8857",
"0.836",
"0.8857",
"0.851",
"0.601",
"0.8496"
],
[
"0.08905",
"0.8433",
"0.2484",
"0.8433",
"0.2484",
"0.857",
"0.08905",
"0.857"
],
[
"0.4983",
"0.853",
"0.6807",
"0.845",
"0.681",
"0.863",
"0.5",
"0.8706"
],
[
"0.2345",
"0.863",
"0.2532",
"0.863",
"0.2532",
"0.877",
"0.2345",
"0.877"
],
[
"0.2051",
"0.877",
"0.357",
"0.877",
"0.357",
"0.8916",
"0.2051",
"0.8916"
],
[
"0.4468",
"0.877",
"0.4739",
"0.877",
"0.4739",
"0.8896",
"0.4468",
"0.8896"
],
[
"0.5024",
"0.877",
"0.888",
"0.877",
"0.888",
"0.89",
"0.5024",
"0.89"
],
[
"0.2296",
"0.8926",
"0.2524",
"0.8896",
"0.2566",
"0.901",
"0.2336",
"0.904"
],
[
"0.5",
"0.89",
"0.884",
"0.8926",
"0.884",
"0.906",
"0.5",
"0.904"
]
] | [
"Robust Imitation Learning against Variations in Environment Dynamics",
"首先，考口口口定D的元口行口化。口注意，元口影",
"口出了一口合理的目口，其中，入，=1，P元和",
"口（7)中的第一口E[。在第一中我口有加□因子",
"是分口p和pe得的一化occupancy分布。（",
"(s），使得=1,(s)=1，且定适的入(s)",
"Section 6中口考口了其他目口。）然后，我口可以得出",
"廉口.因此，利用2(s）=1的事口，我口可以根",
"(10)的上限，如下所示：",
"据期望的口性度口2点推入期望，口而重口口定Dj",
"N",
"min,DJs(p,P)",
"的第一口，口□得其上限",
"j=1",
"N",
"( max [ log(1 -D;(s,a)",
"minEp",
"",
"A,(s) log(1 - Du(s, a))",
"≤min",
"1",
"",
"2N",
"i=1 5=1",
"+",
"N",
"≤minEpmaxlog(1 - Di(s,a)",
"+ E [log Das(s,a)} + log2,",
"(8)",
"=1",
"",
"其中上限的推口在AppendixB.2中。在考在",
",其中E[]表示E(s,a)~[]且不等式被是有",
"口情口下口口口定的D口化元。同口，口了口理入，，",
"效的。然后，我口口化策略π的目口函口（8)的上界",
"我口可以用max，替口，入，以口生口一口上限。然后，",
"接下口，考口口定元的D的口化。此口化由于以下定",
"口定D的元的目口函口由下式口出",
"理而口化：",
"Theorem 5.2.The following maximization problem",
"N",
"min max Ep [log(1 - Ds(s, a)].",
"(11)",
"=1",
"without the A;(s) term has the same optimal solution",
"for Dj as (7) with μ(s)/μe(s) set to 1 for given π :",
"我口口此方法口口多口境中的占用度量匹配",
"max Ep [log(1 -Di;(s,a)] +Eg [log(Du(s,a)]",
"（OMME）。目口（11）口（8）中提出的目口的口口",
"Dit",
"口在于，操作max，超出了口期E。[。口注意，由于",
"Proof. See in Appendix A.2",
"口",
"max，是非口性口算，因此口序不可互口。我口口看到，",
"看似微小的差‧在Section6中生了着的性能",
"基于Theorem5.2和梯度口（GP），我口最口得出",
"差口。",
"定元的D的目口函口如下：",
"Bax [Ep, [log(1 - Das(s, a)] + E- [og(Di;(s,a)]",
"6. Experiments",
"6.1. Experimental Settings",
"+E(/VD()I2 -1)?",
"(9)",
"我口口我口的算法口以下基口一起考口：",
"。注意，在(9)中，我口添加了Section3.2中提到的梯",
"行克隆（BC）：通口口督口口口口策略，直到所有",
"度口日含，用于口定口合，口口E替口口E",
"口家演示的口口口口停止口少口止",
"-GAIL-mixture：是MPE的GAIL+GP的",
"口注意，口口器D的口量由N?口出，口且口着口境",
"在口口情口下，我口有一口口口器，口且口口",
"的N口量的增加而二次增加。我口可以通口使用口口",
"口器口分所有P元和所有PE。元的目口函口是",
"器门重共享口少广门字门使得正器模型共享其",
"min DJs(E, p=/N,Z,pe/N)",
"重的子集（Liu＆Tuzel,2016）。器Di，··，DiN",
"-GAIL-single是MPE的GAIL+GP的一",
"共享其口入口和口藏口的口重，因此可以口口口口口口",
"在口口情口下，我口有多口口器，元的口函口是",
"具有N口出口点的一口口口器。我口口之口口重共享",
"其N口出口",
"min,DJs(p,pe)",
"器(WSD).手WSDDWeightShared）",
"-OMME（最接近我口的算法）：口已口描述口了。目口",
"点的第口口出口口于D的口出，其目口由乙，Vi",
"函口由(10)和(11)口出",
"口出，其中Vii是Di在（9）中的口目口。使用WSDs",
"DWeight-Shared",
"，=1，N口口器的口口性降低",
"AppendixF中口介口了基口、口口和口家演示",
"",
"口乎是～N",
"我口考口了提出的算法的口口版本：RIME和RIME+",
"WSD。RIME和RIME+WSD之的唯一²",
"5.4.Comparisonwith Occupancy MeasureMatching",
"是口器D的口。RIME具有目函口口（9）的口",
"口器D，因此口器口口的口量口N?。口一方面，",
"口使在MPE的情口下不保口口占用度量中恢口策略",
"RIME+",
"WSD使用 weight-shared discriminator",
"分口，我口仍然可以口占用度量匹配技用手MPE",
"DWeight-Shared",
"5.3末尾描迹的weight-shareddis-",
"MPE。在口口情口下，",
"criminator DWeight-Shared",
"N",
"min A,DJs(p元, Pe),",
"(10)",
"我口在MuJoCo任上口口了所考口的算法：Hopper，",
"j=1",
"Walker2dHalfCheetah和Ant(Todorov etal.,2012)"
] | {
"arxiv_id": "2206.09314",
"language": "zh",
"page_idx": 4
} |
||
1810.02358-zh-0 | 1810.02358-zh-0.jpg | # Transfer Learning via Unsupervised Task Discovery for Visual Question Answering
Hyeonwoo Noh$^{1,3}$Taehoon Kim$^{2*}$Jonghwan Mun$^{1,3}$Bohyung Ha POSTECH, Korea USA Seoul National University, Korea
$\{$ shgusdngogo,jonghwan.mun $\}$ @postech.ac.k
# $\operatorname{Abstract}$
We study how to leverage off-the-shelf visual and linguistic data to cope with out-of-vocabulary answers in visual question answering task. Existing large-scale visual datasets with annotations such as image class labels, bounding boxes and region descriptions are good sources for learning rich and diverse visual concepts. However, it is not straightforward how the visual concepts can be captured and transferred to visual question answering models due to missing link between question dependent answering models and visual data without question. We tackle this problem in two steps: 1) learning a task conditional visual classifier, which is capable of solving diverse question-specific visual recognition tasks, based on unsupervised task discovery and 2) transferring the task conditional visual classifier to visual question answering models. Specifically, we employ linguistic knowledge sources such as structured lexical database( e.g. WordNet) and visual descriptions for unsupervised task discovery, and transfer a learned task conditional visual classifier as an answering unit in a visual question answering model. We empirically show that the proposed algorithm generalizes to out-of-vocabulary answers successfully using the knowledge transferred from the visual dataset.
## 1. Introduction
人类看到并理解基于不同视觉概念的视觉场景。例如，从椅子的单个图像中，人类毫不费力地识别出各种视觉概念，例如其颜色，材料，样式，用途等。这些不同的视觉概念可能与自然语言中的不同问题相关联， 这些问题定义了每个视觉概念的识别任务( e.g. , 椅子是什么颜色？)。最近提出了视觉问答(VQA)E[5], 以学习能够执行由问题自适应定义的各种视觉识别任务的深度神经网络模型
$^{*}$This work is performed while at Devsisters
VQA 方法依赖于图像，问题和答案三元组的大规模数据集，并训练以图像和问题为输入并生成答案的分类器。尽管最近取得了显着进展[3,12,43] ,但这个方向有一个关键的限制，即数据集中的图像，问题和答案三元组是学习视觉概念的唯一来源。这样的缺点可能会导致缺乏可伸缩性，因为三元组可能是由具有有限质量控制的人类注释者人工收集的，并且在视觉概念上具有较弱的多样性。事实上，VQA 数据集[1,13] 存在固有的偏差，这阻碍了从数据集中学习真正的视觉概念
相反，人类根据从不同来源(如书籍，图片，视频和个人经验) 学到的视觉概念回答问题，这些概念不一定与目标问题相关。即使对于机器，也存在更自然和可扩展的来源来学习视觉概念：图像类标签，边界框和图像描述。这样的信息已经在大规模的 [9,24,22] 中可用， 并且可以以合理的成本进一步扩展[31,32] 。这种观察带来了一个自然的问题；我们可以在没有问题注释的情况下学习视觉概念并将其转移到 VQA 中吗？
为了解决这个问题，我们引入了一个带有词汇外答案的 VQA 问题，如图$\boxed{\mathrm{E}}1$ 所示。外部可视化数据集提供了一组标签$\mathcal{A}$,这些标签中只有一部分$\mathcal{B}\subset\mathcal{A}$作为答案出现在 VQA 训练集中。此任务的目标是通过利用从外部视觉数据集中学习的视觉概念来成功处理词汇外答案$a\in\mathcal{A}-\mathcal{B}$ 。本文研究了如何毫无疑问地学习视觉概念，以及如何将学到的概念转移到 VQA 模型中。为了学习可转移的视觉概念，我们训练了一个任务条件视觉分类器，其任务由任务特征定义。分类器用作回答单元，其中从问题中推断出任务特征。为了训练没有任务注释的任务条件视觉分类器，我们提出了一种基于语言知识源的无监督任务发现技术，如结构化词汇数据库、e.g. . WordNet [10] 和区域描述 [22]。我们声称，拟议的迁移学习框架有助于在 VQA 中推广词汇外答案。本文的主要贡献有三个方面：
$\bullet$ We present a novel transfer learning algorithm for
visual question answering based on a task condi-
tional visual classifier.
$\bullet$ We propose an unsupervised task discovery tech-
nique for learning task conditional visual classifiers
1 | [
[
"0.2263",
"0.132",
"0.7427",
"0.1338",
"0.7427",
"0.1515",
"0.2263",
"0.1497"
],
[
"0.33",
"0.1547",
"0.6406",
"0.1572",
"0.6406",
"0.1755",
"0.33",
"0.1736"
],
[
"0.08905",
"0.2039",
"0.2443",
"0.2008",
"0.2443",
"0.2197",
"0.08905",
"0.2229"
],
[
"0.3203",
"0.2026",
"0.4697",
"0.202",
"0.4697",
"0.2197",
"0.3203",
"0.2203"
],
[
"0.5312",
"0.2039",
"0.6895",
"0.2008",
"0.6895",
"0.2197",
"0.5312",
"0.2229"
],
[
"0.7656",
"0.2039",
"0.9004",
"0.2039",
"0.9004",
"0.2216",
"0.7656",
"0.2216"
],
[
"0.3774",
"0.2203",
"0.4673",
"0.2229",
"0.4673",
"0.2405",
"0.3774",
"0.238"
],
[
"0.5063",
"0.2184",
"0.8447",
"0.2203",
"0.8447",
"0.2412",
"0.5063",
"0.2393"
],
[
"0.4004",
"0.2405",
"0.4453",
"0.2405",
"0.4453",
"0.2583",
"0.4004",
"0.2583"
],
[
"0.1316",
"0.2235",
"0.3325",
"0.2235",
"0.3325",
"0.238",
"0.1316",
"0.238"
],
[
"0.1528",
"0.2412",
"0.3096",
"0.2412",
"0.3096",
"0.2556",
"0.1528",
"0.2556"
],
[
"0.5376",
"0.2405",
"0.8145",
"0.2412",
"0.8145",
"0.2556",
"0.5376",
"0.2551"
],
[
"0.1691",
"0.2588",
"0.5073",
"0.26",
"0.5073",
"0.2747",
"0.1691",
"0.2734"
],
[
"0.5034",
"0.2644",
"0.513",
"0.2644",
"0.513",
"0.2695",
"0.5034",
"0.2695"
],
[
"0.5107",
"0.26",
"0.818",
"0.26",
"0.818",
"0.2734",
"0.5107",
"0.2734"
],
[
"0.2345",
"0.308",
"0.3137",
"0.308",
"0.3137",
"0.3257",
"0.2345",
"0.3257"
],
[
"0.524",
"0.3125",
"0.8906",
"0.3125",
"0.8906",
"0.3257",
"0.524",
"0.3257"
],
[
"0.505",
"0.3276",
"0.8896",
"0.3276",
"0.8896",
"0.3416",
"0.505",
"0.3416"
],
[
"0.10046",
"0.3416",
"0.4673",
"0.3416",
"0.4673",
"0.3547",
"0.10046",
"0.3547"
],
[
"0.505",
"0.3423",
"0.8896",
"0.3423",
"0.8896",
"0.356",
"0.505",
"0.356"
],
[
"0.0809",
"0.358",
"0.4673",
"0.358",
"0.4673",
"0.3718",
"0.0809",
"0.3718"
],
[
"0.5034",
"0.3567",
"0.8906",
"0.356",
"0.8906",
"0.3706",
"0.5034",
"0.3718"
],
[
"0.0801",
"0.3726",
"0.468",
"0.3726",
"0.468",
"0.3865",
"0.0801",
"0.3865"
],
[
"0.5024",
"0.3718",
"0.8896",
"0.3718",
"0.8896",
"0.385",
"0.5024",
"0.385"
],
[
"0.0809",
"0.3882",
"0.468",
"0.3882",
"0.468",
"0.4016",
"0.0809",
"0.4016"
],
[
"0.505",
"0.387",
"0.8906",
"0.387",
"0.8906",
"0.401",
"0.505",
"0.401"
],
[
"0.0809",
"0.4028",
"0.468",
"0.4028",
"0.468",
"0.4167",
"0.0809",
"0.4167"
],
[
"0.5034",
"0.4016",
"0.8906",
"0.4028",
"0.8906",
"0.4167",
"0.5034",
"0.4153"
],
[
"0.0809",
"0.4185",
"0.4673",
"0.4185",
"0.4673",
"0.432",
"0.0809",
"0.432"
],
[
"0.5034",
"0.4185",
"0.886",
"0.4185",
"0.886",
"0.432",
"0.5034",
"0.432"
],
[
"0.0809",
"0.433",
"0.4656",
"0.433",
"0.4656",
"0.4463",
"0.0809",
"0.4463"
],
[
"0.505",
"0.433",
"0.8896",
"0.433",
"0.8896",
"0.4463",
"0.505",
"0.4463"
],
[
"0.0809",
"0.449",
"0.4673",
"0.449",
"0.4673",
"0.4622",
"0.0809",
"0.4622"
],
[
"0.505",
"0.4495",
"0.5522",
"0.4495",
"0.5522",
"0.461",
"0.505",
"0.461"
],
[
"0.0809",
"0.4634",
"0.4673",
"0.4634",
"0.4673",
"0.4766",
"0.0809",
"0.4766"
],
[
"0.524",
"0.4634",
"0.8906",
"0.4634",
"0.8906",
"0.4766",
"0.524",
"0.4766"
],
[
"0.0809",
"0.4792",
"0.4656",
"0.4792",
"0.4656",
"0.4924",
"0.0809",
"0.4924"
],
[
"0.505",
"0.4792",
"0.8906",
"0.4792",
"0.8906",
"0.4924",
"0.505",
"0.4924"
],
[
"0.0809",
"0.4937",
"0.4673",
"0.4937",
"0.4673",
"0.507",
"0.0809",
"0.507"
],
[
"0.5063",
"0.4937",
"0.888",
"0.4937",
"0.888",
"0.507",
"0.5063",
"0.507"
],
[
"0.0809",
"0.509",
"0.468",
"0.509",
"0.468",
"0.523",
"0.0809",
"0.523"
],
[
"0.505",
"0.509",
"0.8906",
"0.509",
"0.8906",
"0.523",
"0.505",
"0.523"
],
[
"0.0801",
"0.5234",
"0.468",
"0.5234",
"0.468",
"0.537",
"0.0801",
"0.537"
],
[
"0.5034",
"0.5244",
"0.8896",
"0.5244",
"0.8896",
"0.5386",
"0.5034",
"0.5386"
],
[
"0.0809",
"0.5386",
"0.468",
"0.5386",
"0.468",
"0.553",
"0.0809",
"0.553"
],
[
"0.5034",
"0.539",
"0.8906",
"0.539",
"0.8906",
"0.553",
"0.5034",
"0.553"
],
[
"0.0809",
"0.5537",
"0.468",
"0.5537",
"0.468",
"0.5674",
"0.0809",
"0.5674"
],
[
"0.5034",
"0.553",
"0.8906",
"0.5537",
"0.8906",
"0.5684",
"0.5034",
"0.5674"
],
[
"0.0809",
"0.5693",
"0.468",
"0.5693",
"0.468",
"0.5835",
"0.0809",
"0.5835"
],
[
"0.5024",
"0.5693",
"0.8364",
"0.5684",
"0.8364",
"0.582",
"0.5024",
"0.5835"
],
[
"0.0809",
"0.584",
"0.468",
"0.584",
"0.468",
"0.5977",
"0.0809",
"0.5977"
],
[
"0.524",
"0.5854",
"0.8906",
"0.5854",
"0.8906",
"0.5986",
"0.524",
"0.5986"
],
[
"0.0801",
"0.5996",
"0.4673",
"0.5996",
"0.4673",
"0.613",
"0.0801",
"0.613"
],
[
"0.5034",
"0.5986",
"0.8906",
"0.601",
"0.8906",
"0.6143",
"0.5034",
"0.6123"
],
[
"0.0801",
"0.6143",
"0.4673",
"0.6143",
"0.4673",
"0.6274",
"0.0801",
"0.6274"
],
[
"0.5063",
"0.616",
"0.8906",
"0.616",
"0.8906",
"0.627",
"0.5063",
"0.627"
],
[
"0.0809",
"0.631",
"0.4656",
"0.631",
"0.4656",
"0.6416",
"0.0809",
"0.6416"
],
[
"0.505",
"0.63",
"0.8906",
"0.63",
"0.8906",
"0.643",
"0.505",
"0.643"
],
[
"0.0801",
"0.6445",
"0.468",
"0.6445",
"0.468",
"0.6577",
"0.0801",
"0.6577"
],
[
"0.505",
"0.646",
"0.8906",
"0.646",
"0.8906",
"0.659",
"0.505",
"0.659"
],
[
"0.0809",
"0.66",
"0.464",
"0.66",
"0.464",
"0.674",
"0.0809",
"0.674"
],
[
"0.5034",
"0.66",
"0.888",
"0.66",
"0.888",
"0.674",
"0.5034",
"0.674"
],
[
"0.0801",
"0.674",
"0.468",
"0.675",
"0.468",
"0.6895",
"0.0801",
"0.688"
],
[
"0.505",
"0.6753",
"0.8896",
"0.6753",
"0.8896",
"0.6895",
"0.505",
"0.6895"
],
[
"0.0809",
"0.688",
"0.4387",
"0.688",
"0.4387",
"0.702",
"0.0809",
"0.702"
],
[
"0.5034",
"0.6895",
"0.8906",
"0.69",
"0.8906",
"0.705",
"0.5034",
"0.704"
],
[
"0.505",
"0.7056",
"0.8906",
"0.7056",
"0.8906",
"0.7197",
"0.505",
"0.7197"
],
[
"0.505",
"0.72",
"0.8906",
"0.72",
"0.8906",
"0.7344",
"0.505",
"0.7344"
],
[
"0.0809",
"0.7344",
"0.2157",
"0.7344",
"0.2157",
"0.7476",
"0.0809",
"0.7476"
],
[
"0.5063",
"0.7373",
"0.888",
"0.7373",
"0.888",
"0.7476",
"0.5063",
"0.7476"
],
[
"0.5063",
"0.7524",
"0.8896",
"0.7524",
"0.8896",
"0.763",
"0.5063",
"0.763"
],
[
"0.10297",
"0.759",
"0.468",
"0.759",
"0.468",
"0.773",
"0.10297",
"0.773"
],
[
"0.5034",
"0.7666",
"0.8906",
"0.7666",
"0.8906",
"0.78",
"0.5034",
"0.78"
],
[
"0.08417",
"0.776",
"0.4673",
"0.776",
"0.4673",
"0.786",
"0.08417",
"0.786"
],
[
"0.505",
"0.781",
"0.8906",
"0.781",
"0.8906",
"0.7944",
"0.505",
"0.7944"
],
[
"0.0809",
"0.789",
"0.468",
"0.789",
"0.468",
"0.803",
"0.0809",
"0.803"
],
[
"0.5024",
"0.797",
"0.7876",
"0.797",
"0.7876",
"0.81",
"0.5024",
"0.81"
],
[
"0.08417",
"0.806",
"0.4656",
"0.806",
"0.4656",
"0.8164",
"0.08417",
"0.8164"
],
[
"0.0801",
"0.8174",
"0.468",
"0.8193",
"0.468",
"0.834",
"0.0801",
"0.832"
],
[
"0.517",
"0.8193",
"0.8896",
"0.8193",
"0.8896",
"0.8335",
"0.517",
"0.8335"
],
[
"0.0825",
"0.8354",
"0.4673",
"0.8354",
"0.4673",
"0.8486",
"0.0825",
"0.8486"
],
[
"0.5376",
"0.834",
"0.8906",
"0.834",
"0.8906",
"0.8477",
"0.5376",
"0.8477"
],
[
"0.08417",
"0.8516",
"0.4673",
"0.8516",
"0.4673",
"0.8623",
"0.08417",
"0.8623"
],
[
"0.536",
"0.8496",
"0.6978",
"0.8496",
"0.6978",
"0.863",
"0.536",
"0.863"
],
[
"0.0809",
"0.8657",
"0.2443",
"0.8657",
"0.2443",
"0.879",
"0.0809",
"0.879"
],
[
"0.5186",
"0.872",
"0.8906",
"0.872",
"0.8906",
"0.886",
"0.5186",
"0.886"
],
[
"0.10046",
"0.888",
"0.3677",
"0.888",
"0.3677",
"0.9014",
"0.10046",
"0.9014"
],
[
"0.5337",
"0.887",
"0.8906",
"0.887",
"0.8906",
"0.901",
"0.5337",
"0.901"
],
[
"0.4805",
"0.925",
"0.4927",
"0.925",
"0.4927",
"0.938",
"0.4805",
"0.938"
]
] | [
"Transfer Learning via Unsupervised Task Discovery",
"for Visual Question Answering",
"Hyeonwoo Noh 1,3",
"Taehoon Kim 2 *",
"Jonghwan Mun 1,3",
"Bohyung Han 3",
"2 OpenAI",
"3 Computer Vision Lab., ECE & ASRI",
"USA",
"1Computer VisionLab.",
"POSTECH, Korea",
"Seoul National University, Korea",
"1 ( shgusdngogo,jonghwan.mun @postech.ac.kr",
"",
"2 [email protected]@snu.ac.kr",
"Abstract",
"VQA方法依赖于图像，问题和答案三元组的大规模",
"数据集，并训练以图像和问题为输人并生成答案的分",
"We study how to leverage off-the-shelf visual and",
"类器。尽管最近取得了显着进展[3，12，43]，但这个",
"linguistic data to cope with out-of-vocabulary answers",
"方向有一个关键的限制，即数据集中的图像，问题和答",
"in visual question answering task. Existing large-scale",
"案三元组是学习视觉概念的唯一来源。这样的缺点可",
"visual datasets with annotations such as image class la-",
"能会导致缺乏可伸缩性，因为三元组可能是由具有有",
"bels, bounding boxes and region descriptions are good",
"限质量控制的人类注释者人工收集的，并且在视觉概",
"sources for learning rich and diverse visual concepts.",
"念上具有较弱的多样性。事实上，VQA数据集[1，13",
"However, it is not straightforward how the visual con-",
"存在固有的偏差，这阻碍了从数据集中学习真正的视",
"cepts can be captured and transferred to visual ques-",
"觉概念",
"tion answering models due to missing link between",
"相反，人类根据从不同来源（如书籍，图片，视频和",
"question dependent answering models and visual data",
"个人经验）学到的视觉概念回答问题，这些概念不一定",
"without question. We tackle this problem in two steps:",
"与目标问题相关。即使对于机器，也存在更自然和可扩",
"1) learning a task conditional visual classifier, which",
"展的来源来学习视觉概念：图像类标签，边界框和图像",
"is capable of solving diverse question-specific visual",
"描述。这样的信息已经在大规模的[9，24，22]中可用，",
"recognition tasks, based on unsupervised task discovery",
"并且可以以合理的成本进一步扩展[31，32]。这种观",
"and 2) transferring the task conditional visual classifier",
"察带来了一个自然的问题；我们可以在没有问题注释的",
"to visual question answering models. Specifically, we",
"情况下学习视觉概念并将其转移到VQA中吗？",
"employ linguistic knowledge sources such as structured",
"为了解决这个问题，我们引人了一个带有词汇外答",
"lexical database ( e.g. WordNet) and visual descrip-",
"案的VQA问题，如图E1所示。外部可视化数据集提",
"tions for unsupervised task discovery, and transfer a",
"供了一组标签A，这些标签中只有一部分BCA作",
"learned task conditional visual classifier as an answer-",
"为答案出现在VQA训练集中。此任务的目标是通过利",
"ing unit in a visual question answering model. We",
"用从外部视觉数据集中学习的视觉概念来成功处理词",
"empirically show that the proposed algorithm gener",
"汇外答案aEA-B。本文研究了如何毫无疑问地学",
"alizes to out-of-vocabulary answers successfully using",
"习视觉概念，以及如何将学到的概念转移到VQA模型",
"the knowledge transferred from the visual dataset.",
"中。为了学习可转移的视觉概念，我们训练了一个任务",
"条件视觉分类器，其任务由任务特征定义。分类器用作",
"回答单元，其中从问题中推断出任务特征。为了训练没",
"1.Introduction",
"有任务注释的任务条件视觉分类器，我们提出了一种",
"基于语言知识源的无监督任务发现技术，如结构化词",
"人类看到并理解基于不同视觉概念的视觉场景。例",
"汇数据库、e.g.、WordNet[10]和区域描述[22]。我",
"如，从椅子的单个图像中，人类毫不费力地识别出各",
"们声称，拟议的迁移学习框架有助于在VQA中推广词",
"种视觉概念，例如其颜色，材料，样式，用途等。这些",
"汇外答案。本文的主要贡献有三个方面：",
"不同的视觉概念可能与自然语言中的不同问题相关联",
"这些问题定义了每个视觉概念的识别任务（e.g.，椅",
". We present a novel transfer learning algorithm for",
"子是什么颜色？）。最近提出了视觉问答（VQA)E[5】",
"visual question answering based on a task condi-",
"以学习能够执行由问题自适应定义的各种视觉识别任",
"tional visual classifier.",
"务的深度神经网络模型",
". We propose an unsupervised task discovery tech-",
"*This work is performed while at Devsisters.",
"nique for learning task conditional visual classifiers",
"1"
] | {
"arxiv_id": "1810.02358",
"language": "zh",
"page_idx": 0
} |
||
1810.02358-zh-1 | 1810.02358-zh-1.jpg | <FigureHere>
Figure 1. VQA with out-of-vocabulary answers. Given a set of labels in visual dataset $\mathcal{A}$ and a set of answers in VQA training set $\mathcal{B}$ , we evaluate a model on VQA test set with answers $\boldsymbol{a}\in\mathcal{A}-\mathcal{B}$ . External visual dataset provides a set of bounding box labels and visual descriptions for all answers in VQA training set and test set. See text for details
without explicit task annotations.
$\bullet$ We show that the proposed approach handle out-
of-vocabulary answers through knowledge transfer from visual dataset without question annotations.
本文的其余部分组织如下。第 2节讨论了与我们的方法相关的先前工作。我们在 3 节中描述了整体迁移学习框架。通过无监督任务发现学习视觉概念在 4 节中进行了描述。5 节分析实验结果，6 节得出我们的结论
[14]中提出了在 VQA 模型和图像分类器之间共享对齐
的图像单词表示，以利用外部视觉数据
在 VQA 中几乎没有研究过从外部数据中进行迁移学习以应对词汇外单词，但是在新颖的对象标题E [4,28,37,44]中正在积极研究。例如，[4]和 [37] 将图像字幕任务分解为视觉分类和语言建模，并利用不成对的视觉和语言数据作为分别训练视觉分类器和语言模型的附加资源。最近的方法结合了指针网络$\mathbb{E}[38]$ , 并学习指向候选单词$\mathbb{E}[44]$ 或相关区域$\mathbb{E}[28]$ 的索引， 其中候选单词由多标签分类器$\mathbb{E}[44]$ 或对象检测器$\mathbb{E}$ [28]通过外部视觉数据训练来检测。但是，这些算法不能直接应用于我们的问题设置，因为它们专注于在没有任务规范的情况下预测对象词，而 VQA 需要任务条件视觉识别
我们的问题设置与零拍学习$\mathbb{E}[11,17,25,26,42]$ 密切相关，其中在分类中考虑了词汇外答案。零拍学习旨在识别训练期间看不见的对象或类。由于它旨在推广到完全看不见的类，因此严格禁止在训练过程中接触零拍类[42] 。相反，我们的目标是利用外部数据集中可用的类标签
## 2. Related Works
### $3.{\mathrm{~Transfer~Learning~Framework}}$
标准 VQA 评估假设相同分布的列车和测试集[5, 29,45]。由于此评估设置容易受到训练集$\mathbb{E}[13]$ 中利用偏差的模型的影响，因此提出了几种替代方案。一种方法是通过平衡单个问题的答案来减少观察到的偏差$\mathbb{E}[13]$ ,或者通过故意提供不同的偏差来训练和测试集$\mathbb{E}[1]$ 。另一种方法是构造组合概括 split$\underline{\mathrm{E}}[2,18]$ , 其测试集中的问题和答案对是由出现在训练集中的视觉概念和问题类型的新颖组合形成的。此拆分是通过重新调整现有 VQA 数据集$\mathbb{E}[2]$ 或构建合成数据集$\mathbb{E}$ [18]来构建的。本文研究的问题设置类似于[19,36] , 即使用词汇表外答案进行测试，但与以前的工作不同， 我们将问题表述为迁移学习，其中词汇表外答案是从外部视觉数据中学习的
外部数据通常用于 VQA 中，以便更好地泛化。卷积神经网络 [15,23]在 ImageNet 上预先训练 [9] 是广泛接受的各种 VQA 模型的标准 [12,43]。作为替代方案，使用在 VisualGenome 数据集$\overline{\mathbb{E}}[22]$ 上训练的对象检测器$\mathbb{E}[34]$ 来提取预训练的视觉特征$\mathbb{E}[3]$ 。预训练的语言模型，例如单词嵌人$\mathbb{E}[33]$ 或句子嵌入$\mathbb{E}$ [21] ,经常用于初始化问题编码器的参数$\mathbb{E}[12,30,35]$ 。在[39,40,41]中研究了利用知识库 [6,7]或外部视觉算法[40]中的信息检索来为 VQA 模型提供额外的输入。在 [16] 中研究了 VQA 数据集之间的传输。在$\mathbb{E}$
我们工作的主要目标是通过从现成的视觉数据集中学习视觉概念并将概念转移到 VQA 以回答问题来处理 VQA 中的词汇外答案。受 VQA 可以被认为是任务条件分类问题的启发，其中任务由问题定义，我们引人了任务条件视觉分类器，该分类器从视觉输入和任务规范中生成答案，作为学习和转移视觉概念的媒介。图 2 说明了所提出方法的总体框架。我们通过无监督任务发现，使用可视化数据集预训练任务条件可视化分类器，而无需问题或任务规范，并通过传输学习的参数使其适应 VQA 模型。在本节的其余部分中，我们将描述任务条件视觉分类器以及如何对其进行预训练并将其传输到 VQA | [
[
"0.08826",
"0.1067",
"0.1422",
"0.1092",
"0.1422",
"0.1206",
"0.08826",
"0.1187"
],
[
"0.2075",
"0.10986",
"0.2427",
"0.10986",
"0.2427",
"0.1218",
"0.2075",
"0.1218"
],
[
"0.08905",
"0.1193",
"0.165",
"0.1193",
"0.165",
"0.13",
"0.08905",
"0.13"
],
[
"0.2018",
"0.1187",
"0.2239",
"0.1187",
"0.2239",
"0.127",
"0.2018",
"0.127"
],
[
"0.08905",
"0.13",
"0.1381",
"0.13",
"0.1381",
"0.1414",
"0.08905",
"0.1414"
],
[
"0.5215",
"0.1414",
"0.5293",
"0.137",
"0.5337",
"0.1414",
"0.5254",
"0.1459"
],
[
"0.603",
"0.1414",
"0.7427",
"0.1414",
"0.7427",
"0.1483",
"0.603",
"0.1483"
],
[
"0.749",
"0.1414",
"0.8384",
"0.1414",
"0.8384",
"0.1483",
"0.749",
"0.1483"
],
[
"0.2443",
"0.1427",
"0.2712",
"0.1427",
"0.2712",
"0.1459",
"0.2443",
"0.1459"
],
[
"0.1772",
"0.1427",
"0.2214",
"0.1427",
"0.2214",
"0.1459",
"0.1772",
"0.1459"
],
[
"0.282",
"0.1427",
"0.3096",
"0.1427",
"0.3096",
"0.1459",
"0.282",
"0.1459"
],
[
"0.3245",
"0.1427",
"0.3423",
"0.1427",
"0.3423",
"0.1459",
"0.3245",
"0.1459"
],
[
"0.4004",
"0.1427",
"0.4436",
"0.1427",
"0.4436",
"0.1459",
"0.4004",
"0.1459"
],
[
"0.5474",
"0.1427",
"0.5864",
"0.1427",
"0.5864",
"0.1459",
"0.5474",
"0.1459"
],
[
"0.2942",
"0.1465",
"0.3406",
"0.1477",
"0.3406",
"0.1547",
"0.2942",
"0.154"
],
[
"0.21",
"0.1483",
"0.2361",
"0.1483",
"0.2361",
"0.1528",
"0.21",
"0.1528"
],
[
"0.3677",
"0.1483",
"0.3857",
"0.1483",
"0.3857",
"0.1528",
"0.3677",
"0.1528"
],
[
"0.4836",
"0.1483",
"0.5",
"0.1483",
"0.5",
"0.1515",
"0.4836",
"0.1515"
],
[
"0.5",
"0.1483",
"0.5156",
"0.1483",
"0.5156",
"0.1528",
"0.5",
"0.1528"
],
[
"0.585",
"0.1497",
"0.6055",
"0.1497",
"0.6055",
"0.154",
"0.585",
"0.154"
],
[
"0.755",
"0.1509",
"0.7646",
"0.1509",
"0.7646",
"0.154",
"0.755",
"0.154"
],
[
"0.7646",
"0.1509",
"0.785",
"0.1509",
"0.785",
"0.1547",
"0.7646",
"0.1547"
],
[
"0.2754",
"0.1622",
"0.415",
"0.1635",
"0.415",
"0.1749",
"0.2754",
"0.1736"
],
[
"0.614",
"0.1642",
"0.7646",
"0.1642",
"0.7646",
"0.1749",
"0.614",
"0.1749"
],
[
"0.1013",
"0.1901",
"0.1487",
"0.1901",
"0.1487",
"0.2008",
"0.1013",
"0.2008"
],
[
"0.1887",
"0.1925",
"0.2133",
"0.1925",
"0.2133",
"0.2008",
"0.1887",
"0.2008"
],
[
"0.2443",
"0.1925",
"0.273",
"0.1925",
"0.273",
"0.2008",
"0.2443",
"0.2008"
],
[
"0.3015",
"0.1895",
"0.3381",
"0.1913",
"0.3367",
"0.202",
"0.2998",
"0.1995"
],
[
"0.3594",
"0.1925",
"0.398",
"0.1932",
"0.398",
"0.2008",
"0.3594",
"0.1995"
],
[
"0.429",
"0.1925",
"0.4495",
"0.1925",
"0.4495",
"0.2008",
"0.429",
"0.2008"
],
[
"0.3865",
"0.2166",
"0.619",
"0.2172",
"0.619",
"0.2279",
"0.3865",
"0.2267"
],
[
"0.5483",
"0.2279",
"0.5645",
"0.2279",
"0.5645",
"0.2311",
"0.5483",
"0.2311"
],
[
"0.6313",
"0.238",
"0.671",
"0.238",
"0.671",
"0.2488",
"0.6313",
"0.2488"
],
[
"0.08826",
"0.2474",
"0.1381",
"0.2474",
"0.1381",
"0.2588",
"0.08826",
"0.2588"
],
[
"0.5415",
"0.2468",
"0.5645",
"0.2468",
"0.5645",
"0.25",
"0.5415",
"0.25"
],
[
"0.08905",
"0.2588",
"0.152",
"0.2588",
"0.152",
"0.2695",
"0.08905",
"0.2695"
],
[
"0.1871",
"0.2551",
"0.2181",
"0.2551",
"0.2181",
"0.2632",
"0.1871",
"0.2632"
],
[
"0.3694",
"0.262",
"0.415",
"0.262",
"0.415",
"0.2727",
"0.3694",
"0.2727"
],
[
"0.5146",
"0.262",
"0.556",
"0.262",
"0.556",
"0.2727",
"0.5146",
"0.2727"
],
[
"0.7114",
"0.2651",
"0.875",
"0.2651",
"0.875",
"0.276",
"0.7114",
"0.276"
],
[
"0.4534",
"0.2664",
"0.4795",
"0.2664",
"0.4795",
"0.2747",
"0.4534",
"0.2747"
],
[
"0.2769",
"0.2695",
"0.3057",
"0.2695",
"0.3057",
"0.2778",
"0.2769",
"0.2778"
],
[
"0.1708",
"0.2822",
"0.3555",
"0.2822",
"0.3555",
"0.2922",
"0.1708",
"0.2922"
],
[
"0.3816",
"0.2827",
"0.701",
"0.2827",
"0.701",
"0.2903",
"0.3816",
"0.2903"
],
[
"0.7114",
"0.279",
"0.8774",
"0.279",
"0.8774",
"0.289",
"0.7114",
"0.289"
],
[
"0.0801",
"0.3005",
"0.8906",
"0.3005",
"0.8906",
"0.3145",
"0.0801",
"0.3145"
],
[
"0.0801",
"0.315",
"0.892",
"0.315",
"0.892",
"0.3289",
"0.0801",
"0.3289"
],
[
"0.0801",
"0.3289",
"0.83",
"0.3289",
"0.83",
"0.3423",
"0.0801",
"0.3423"
],
[
"0.11273",
"0.358",
"0.3555",
"0.358",
"0.3555",
"0.3718",
"0.11273",
"0.3718"
],
[
"0.505",
"0.358",
"0.8896",
"0.358",
"0.8896",
"0.3718",
"0.505",
"0.3718"
],
[
"0.0948",
"0.3774",
"0.4673",
"0.3774",
"0.4673",
"0.3926",
"0.0948",
"0.3926"
],
[
"0.5034",
"0.3726",
"0.7837",
"0.3726",
"0.7837",
"0.3865",
"0.5034",
"0.3865"
],
[
"0.11273",
"0.3933",
"0.4697",
"0.3933",
"0.4697",
"0.4072",
"0.11273",
"0.4072"
],
[
"0.524",
"0.3933",
"0.8896",
"0.3933",
"0.8896",
"0.404",
"0.524",
"0.404"
],
[
"0.1111",
"0.4072",
"0.468",
"0.4092",
"0.468",
"0.4224",
"0.1111",
"0.4204"
],
[
"0.5034",
"0.406",
"0.8906",
"0.406",
"0.8906",
"0.4192",
"0.5034",
"0.4192"
],
[
"0.505",
"0.4216",
"0.8906",
"0.4216",
"0.8906",
"0.435",
"0.505",
"0.435"
],
[
"0.505",
"0.4363",
"0.8906",
"0.4363",
"0.8906",
"0.4495",
"0.505",
"0.4495"
],
[
"0.1013",
"0.443",
"0.4673",
"0.443",
"0.4673",
"0.4539",
"0.1013",
"0.4539"
],
[
"0.5034",
"0.4507",
"0.8896",
"0.452",
"0.8896",
"0.4666",
"0.5034",
"0.4653"
],
[
"0.0825",
"0.457",
"0.4673",
"0.457",
"0.4673",
"0.4702",
"0.0825",
"0.4702"
],
[
"0.0809",
"0.4717",
"0.468",
"0.4717",
"0.468",
"0.486",
"0.0809",
"0.486"
],
[
"0.5034",
"0.4673",
"0.888",
"0.4673",
"0.888",
"0.481",
"0.5034",
"0.481"
],
[
"0.0809",
"0.4849",
"0.468",
"0.4873",
"0.468",
"0.502",
"0.0809",
"0.5"
],
[
"0.505",
"0.4817",
"0.8906",
"0.4817",
"0.8906",
"0.4956",
"0.505",
"0.4956"
],
[
"0.0784",
"0.5005",
"0.1168",
"0.5005",
"0.1168",
"0.5166",
"0.0784",
"0.5166"
],
[
"0.5034",
"0.4968",
"0.8906",
"0.4956",
"0.8906",
"0.5103",
"0.5034",
"0.511"
],
[
"0.505",
"0.512",
"0.8906",
"0.512",
"0.8906",
"0.526",
"0.505",
"0.526"
],
[
"0.10455",
"0.53",
"0.2361",
"0.53",
"0.2361",
"0.5435",
"0.10455",
"0.5435"
],
[
"0.505",
"0.529",
"0.8896",
"0.529",
"0.8896",
"0.539",
"0.505",
"0.539"
],
[
"0.505",
"0.5425",
"0.8906",
"0.5425",
"0.8906",
"0.556",
"0.505",
"0.556"
],
[
"0.0809",
"0.531",
"0.1111",
"0.531",
"0.1111",
"0.5425",
"0.0809",
"0.5425"
],
[
"0.10046",
"0.5537",
"0.4631",
"0.5537",
"0.4631",
"0.5674",
"0.10046",
"0.5674"
],
[
"0.5034",
"0.558",
"0.585",
"0.558",
"0.585",
"0.5693",
"0.5034",
"0.5693"
],
[
"0.0801",
"0.5693",
"0.4673",
"0.5693",
"0.4673",
"0.5835",
"0.0801",
"0.5835"
],
[
"0.523",
"0.5757",
"0.8906",
"0.5757",
"0.8906",
"0.5903",
"0.523",
"0.5903"
],
[
"0.0825",
"0.5854",
"0.4412",
"0.5854",
"0.4412",
"0.5986",
"0.0825",
"0.5986"
],
[
"0.505",
"0.5913",
"0.8906",
"0.5913",
"0.8906",
"0.605",
"0.505",
"0.605"
],
[
"0.0825",
"0.601",
"0.4673",
"0.601",
"0.4673",
"0.6143",
"0.0825",
"0.6143"
],
[
"0.505",
"0.607",
"0.8857",
"0.607",
"0.8857",
"0.6206",
"0.505",
"0.6206"
],
[
"0.0809",
"0.6157",
"0.4673",
"0.6157",
"0.4673",
"0.63",
"0.0809",
"0.63"
],
[
"0.5034",
"0.622",
"0.8906",
"0.622",
"0.8906",
"0.6353",
"0.5034",
"0.6353"
],
[
"0.0825",
"0.63",
"0.468",
"0.63",
"0.468",
"0.6445",
"0.0825",
"0.6445"
],
[
"0.5024",
"0.6353",
"0.8896",
"0.636",
"0.8896",
"0.651",
"0.5024",
"0.6494"
],
[
"0.0825",
"0.6445",
"0.468",
"0.6445",
"0.468",
"0.6577",
"0.0825",
"0.6577"
],
[
"0.505",
"0.6514",
"0.6006",
"0.6514",
"0.6006",
"0.6636",
"0.505",
"0.6636"
],
[
"0.0809",
"0.66",
"0.468",
"0.66",
"0.468",
"0.674",
"0.0809",
"0.674"
],
[
"0.0825",
"0.6753",
"0.468",
"0.6753",
"0.468",
"0.6895",
"0.0825",
"0.6895"
],
[
"0.0801",
"0.6895",
"0.4722",
"0.69",
"0.4722",
"0.705",
"0.0801",
"0.704"
],
[
"0.501",
"0.6914",
"0.7817",
"0.6914",
"0.7817",
"0.705",
"0.501",
"0.705"
],
[
"0.0801",
"0.704",
"0.468",
"0.705",
"0.468",
"0.7197",
"0.0801",
"0.7183"
],
[
"0.0809",
"0.72",
"0.468",
"0.72",
"0.468",
"0.7344",
"0.0809",
"0.7344"
],
[
"0.524",
"0.7197",
"0.8906",
"0.7197",
"0.8906",
"0.733",
"0.524",
"0.733"
],
[
"0.0809",
"0.735",
"0.2443",
"0.735",
"0.2443",
"0.7485",
"0.0809",
"0.7485"
],
[
"0.5024",
"0.7344",
"0.8896",
"0.735",
"0.8896",
"0.75",
"0.5024",
"0.7485"
],
[
"0.1013",
"0.7505",
"0.468",
"0.7505",
"0.468",
"0.765",
"0.1013",
"0.765"
],
[
"0.5034",
"0.7505",
"0.8896",
"0.7505",
"0.8896",
"0.765",
"0.5034",
"0.765"
],
[
"0.0825",
"0.7666",
"0.468",
"0.7666",
"0.468",
"0.78",
"0.0825",
"0.78"
],
[
"0.5034",
"0.7666",
"0.888",
"0.7666",
"0.888",
"0.78",
"0.5034",
"0.78"
],
[
"0.0809",
"0.781",
"0.468",
"0.781",
"0.468",
"0.7944",
"0.0809",
"0.7944"
],
[
"0.5063",
"0.7827",
"0.8896",
"0.7827",
"0.8896",
"0.7935",
"0.5063",
"0.7935"
],
[
"0.0809",
"0.7954",
"0.468",
"0.7954",
"0.468",
"0.81",
"0.0809",
"0.81"
],
[
"0.505",
"0.797",
"0.8906",
"0.797",
"0.8906",
"0.81",
"0.505",
"0.81"
],
[
"0.0809",
"0.811",
"0.468",
"0.811",
"0.468",
"0.8247",
"0.0809",
"0.8247"
],
[
"0.505",
"0.811",
"0.8896",
"0.811",
"0.8896",
"0.8247",
"0.505",
"0.8247"
],
[
"0.0809",
"0.8257",
"0.468",
"0.8257",
"0.468",
"0.839",
"0.0809",
"0.839"
],
[
"0.5034",
"0.8257",
"0.8906",
"0.8257",
"0.8906",
"0.839",
"0.5034",
"0.839"
],
[
"0.0809",
"0.8413",
"0.4673",
"0.8413",
"0.4673",
"0.855",
"0.0809",
"0.855"
],
[
"0.505",
"0.8413",
"0.8896",
"0.8413",
"0.8896",
"0.855",
"0.505",
"0.855"
],
[
"0.09314",
"0.857",
"0.468",
"0.857",
"0.468",
"0.8706",
"0.09314",
"0.8706"
],
[
"0.5024",
"0.857",
"0.8896",
"0.857",
"0.8896",
"0.8706",
"0.5024",
"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"
],
[
"0.0801",
"0.886",
"0.468",
"0.8853",
"0.468",
"0.8994",
"0.0801",
"0.901"
],
[
"0.5034",
"0.887",
"0.6284",
"0.887",
"0.6284",
"0.901",
"0.5034",
"0.901"
]
] | [
"VQA with",
"国",
"out-of-vocab",
"",
"Answers",
"",
"naterial is the Q: what is the child eating",
"Q:what sport is hedoing",
"as1n",
"twhattwoao",
"what",
"acnr",
"anaona",
"whatinr",
"ing?A:apple",
"of2A:",
"nn2",
"2A",
"tuli",
"afrom",
"",
"urhne",
"B:Answers in VQA Train",
"-B:AnswersinVQATest",
"Answers",
"cloth",
"apple",
"cooking",
"snowing",
"tulip",
"A:Labels in Visual Data (used fortraining)",
"",
"banana",
"External",
"",
"Visual data",
"cloth",
"cooking",
"leather",
"A:Set of labels inVisual Data",
"tulip",
"apple",
"the table has a white cloth a hand holding an apple",
"incookingonstoveredtulipblossomsaleathermotorcycleseatThegirfholdingabanane",
"B:Setof answers inVQAtrain",
"Figure 1. VQA with out-of-vocabulary answers. Given a set of labels in visual dataset A and a set of answers in VQA",
"training set B , we evaluate a model on VQA test set with answers a E A - B . External visual dataset provides a set of",
"bounding box labels and visual descriptions for all answers in VQA training set and test set. See text for details.",
"without explicit task annotations.",
"[14]中提出了在VQA模型和图像分类器之间共享对齐",
": We show that the proposed approach handle out-",
"的图像单词表示，以利用外部视觉数据",
"of-vocabulary answers through knowledge transfer",
"在VQA中儿乎没有研究过从外部数据中进行迁",
"from visual dataset without question annotations.",
"移学习以应对词汇外单词，但是在新颖的对象标题E",
"[4，28，37，44】中正在积极研究。例如，4]和[37]将图",
"像字幕任务分解为视觉分类和语言建模，并利用不成",
"本文的其余部分组织如下。第2节讨论了与我们的",
"对的视觉和语言数据作为分别训练视觉分类器和语言",
"方法相关的先前工作。我们在3节中描述了整体迁移",
"学习框架。通过无监督任务发现学习视觉概念在4节",
"模型的附加资源。最近的方法结合了指针网络E[38]",
"中进行了描述。5节分析实验结果，6节得出我们的",
"并学习指向候选单词E[44]或相关区域E[28]的索引,",
"结论",
"其中候选单词由多标签分类器E[44]或对象检测器E",
"[28]通过外部视觉数据训练来检测。但是，这些算法不",
"Related Works",
"能直接应用于我们的问题设置，因为它们专注于在没",
"有任务规范的情况下预测对象词，而VQA需要任务条",
"2.",
"标准VQA评估假设相同分布的列车和测试集[5",
"件视觉识别",
"29，45]。由于此评估设置容易受到训练集E[13]中利",
"我们的问题设置与零拍学习E[11,17，25,26，42]密",
"用偏差的模型的影响，因此提出了几种替代方案。",
"切相关，其中在分类中考虑了词汇外答案。零拍学习旨",
"种方法是通过平衡单个问题的答案来减少观察到的偏",
"在识别训练期间看不见的对象或类。由于它旨在推广",
"差E[13]，或者通过故意提供不同的偏差来训练和测",
"到完全看不见的类，因此严格禁止在训练过程中接触",
"试集E[1] 。另一种方法是构造组合概括splitE[2,18],",
"零拍类「421。相反，我们的目标是利用外部数据集中",
"其测试集中的问题和答案对是由出现在训练集中的视",
"可用的类标签",
"觉概念和问题类型的新颖组合形成的。此拆分是通过",
"重新调整现有VQA数据集E[2]或构建合成数据集E",
"[18]来构建的。本文研究的问题设置类似于【19，36]】，",
"3.Transfer Learning Framework",
"即使用词汇表外答案进行测试，但与以前的工作不同，",
"我们将问题表述为迁移学习，其中词汇表外答案是从",
"我们工作的主要目标是通过从现成的视觉数据集中",
"外部视觉数据中学习的",
"学习视觉概念并将概念转移到VQA以回答问题来处",
"外部数据通常用于VQA中，以便更好地泛化。卷",
"理VQA中的词汇外答案。受VQA可以被认为是任务",
"积神经网络「15，23]在ImageNet上预先训练[9]是",
"条件分类问题的启发，其中任务由问题定义，我们引",
"广泛接受的各种VQA模型的标准「12，43]。作为替",
"入了任务条件视觉分类器，该分类器从视觉输入和任",
"代方案，使用在VisualGenome数据集国[22]上训练的",
"务规范中生成答案，作为学习和转移视觉概念的媒介。",
"对象检测器E[34]来提取预训练的视觉特征E[3]。预",
"图2说明了所提出方法的总体框架。我们通过无监督",
"训练的语言模型，例如单词嵌人E33]或句子嵌入E",
"任务发现，使用可视化数据集预训练任务条件可视化",
"[21]，经常用于初始化问题编码器的参数E[12,30，35]",
"分类器，而无需问题或任务规范，并通过传输学习的参",
"在[39，40，41]中研究了利用知识库［6，7]或外部视",
"数使其适应VQA模型。在本节的其余部分中，我们将",
"觉算法「40]中的信息检索来为VQA模型提供额外的",
"描述任务条件视觉分类器以及如何对其进行预训练并",
"输人。在【16]中研究了VQA数据集之间的传输。在E",
"将其传输到VQA"
] | {
"arxiv_id": "1810.02358",
"language": "zh",
"page_idx": 1
} |
||
1810.02358-zh-3 | 1810.02358-zh-3.jpg | game.n.01
<table>
<tbody>
<tr>
<th>$<t_w$ task</th>
<th>${\mathrm{word~groups}}$</th>
</tr>
<tr>
<td>$\operatorname{appliance.n.02}$</td>
<td>freezer, hair dryer, refrigerator, fridge, oven, dishwasher, washing</td>
</tr>
<tr>
<td>opening. n. 10</td>
<td>manhole, rear window, exit, nozzle, car window, coin slot, spout, vent, </td>
</tr>
<tr>
<td>${\mathrm{food.n.02}}$</td>
<td>potatoes, french fries, chicken, melted cheese, tomatoes, sausage,</td>
</tr>
<tr>
<td>move. v. 03</td>
<td>twisted, ruffled, curving, curved, wheeled, rolled, coiled, turned, rippled</td>
</tr>
<tr>
<td>act. n. 02</td>
<td>fishing, skateboarding, sailing, playing baseball, surfing, traveling, $\{r\dot{\mathbf{i}}\}$</td>
</tr>
<tr>
<td>area. n. 01</td>
<td>middle, in corner, center, parking space, playground, landscape ichhorhoor</td>
</tr>
<tr>
<td>color. n. 01</td>
<td>yellow, pink, red, beige, royal blue, amber, aqua, dark red, olive green teal</td>
</tr>
</tbody>
</table>
Figure 4. Illustration of WordNet and constructed word group table. (Left) A subgraph of the WordNet [10]. Complex
atheletic\_game.n.01
card\_game.n.O1
contact sport.n.01
outdoor\_game.n.01
field\_game.n.01
golf.n.01
cricketn.01$\left\lfloor\text{football.n.01}\right\rfloor\underbrace{\text{boxing.n.01}}$
hierarchy of words reveals the diverse categorization of each words. (Right) A set of words sharing common parents in the tree is grouped as a single word group. Diverse grouping of words reveals diverse visual recognition tasks that can be defined on each word group.
4. Unsupervised Task Discovery
使用现成的可视化数据集 [22]学习任务条件可视化分类器并不简单，因为缺少任务规范的注释，这是学习任务规范向量$\tau$的编码器所必需的。为了解决这个问题，我们提出了无监督的任务发现，它从通过利用语言知识源建模的任务分配中采样任务规范$t$
4.1. Leveraging linguistic knowledge sources
由问题给出的视觉识别任务通常定义从视觉输人到一组可能的视觉概念的映射，即单词组。例如，“女人拿着什么？”的问题定义了在图像中查找可容纳对象的视觉识别任务，这是对单词组$\{$ball, racket, cup, ...$\}$ 的分类。这种直觉导致了一种简单的方法，通过将任务视为单词组(i.e.可保存对象)来对任务描述$t$的分布进行建模。使用语言知识源进行无监督任务发现的主要原因是，在语言知识源中通常可以访问单词组。我们考虑两个语言知识来源：1)提供视觉数据的视觉描述； 2)称为 WordNet$\underline{\mathrm{E}}[10]$的结构化词汇数据库。图 3 说明了我们方法的概述
给出，其中$v_\mathrm{\phi_\mathrm{vqa}}(\boldsymbol{I},\boldsymbol{q})$是具有图像$I$和问题$q$的编码视觉特征，使用具有参数$\phi_\mathrm{vqa}$的注意力机制。。任务功能$\tau_\mathrm{\eta_\mathrm{vqa}}(\boldsymbol{q})$使用参数$\eta_\mathrm{vqa}$对问题$q$进行编码。优化需要 VQA 的训练数据集 p$_\mathrm{vqa}(\boldsymbol{a},\boldsymbol{I},\boldsymbol{q})$的联合分布，其中来自分布的答案满足$a\in\mathcal{A}-\mathcal{B}$ 。我们通过最大化方程$\mathbb{E}(2)$中目标的可能性来学习$\phi_\mathrm{vqa}$和$\eta_\mathrm{vqa}$,而预训练任务条件视觉分类器$\theta$的参数保持固定
Weakly supervised task regression 利用预训练的任务条件视觉分类器进行由问题$q$指定的视觉识别需要推断最佳任务特征$\tau_q^*$ 。此要求引人了一个学习问题-任务 回 归 - 优 化 编 码 器 $\dot{\tau } _{\eta _{\mathrm{vqa}}}( \boldsymbol{q})$ 以 正 确 预 测 $\tau _{\boldsymbol{q}}^{* }$。由 于 直 接 最 小 化 错 误 $\mathcal{E} ( \tau _{\boldsymbol{q}, \tau _{\eta _{\mathrm{vqa}}}( \boldsymbol{q}) ) }^{* }$需 要 对 任 务 进 行 额 外 的 监 督 , 而 且 有 $( \tau _{\boldsymbol{q}, \tau _{\eta _{\mathrm{vqa}}}( \boldsymbol{q}) ) }^{* }$ 而 要 对 任 务 进 行 额 外 的 监 督 , 因此我们将 VQA 数据作为弱监督的来源。我们优化了间接损失$-\mathbb{E}_\tau_{q(a|v)}^{*}[\log\mathrm{p}_{\theta}(\boldsymbol{a}|v,\tau_{\eta_{\mathrm{vqa}}}(\boldsymbol{q}))]\:$,这鼓励了以 $\tau_q^*$ 和$\tau_\eta_\mathrm{vqa}(\boldsymbol{q})$为条件的答案分布相似。通过假设真实任务条件答案分布$\tau_{\boldsymbol{q}(\boldsymbol{a}|v)}^*$是在 VQA 数据集中隐式建模的， 我们采用方程$\mathbb{E}(2)$ 作为弱监督任务回归的目标函数
Out-of-vocabulary answering 我们通过调整输入表示来学习 VQA 模型，同时固定预训练任务条件视觉分类器 p$_\theta(\boldsymbol{a}|v,\tau)$ 。此策略允许模型专注于学习从问题中推断视觉识别任务$\tau_{\eta_\mathrm{vqa}}(\boldsymbol{q})$,而不需要所有可能答案的数据。一旦推断出任务特征$\tau$,学习的任务条件视觉分类器 p$_\theta(\boldsymbol{a}|v,\tau)$就可以回答预训练的视觉概念，包括词汇外的概念
Matching visual features 要在不进行微调的情况下在 VQA 中重用预训练的视觉分类器，视觉特征$v$的语义不应因学习 VQA 数据集而改变。VQA 模型的最新方法实现了这一点，这些模型不微调预训练的视觉特征提取器，而是专注于在提取的特征图上学习注意力机制$\mathbb{E}[20]$ 。在我们的设置中，我们只需使用相同的视觉特征提取器E|3|进行预训练和 VQA
## 4.2. Visual description
我们使用 VisualGenome [22] 作为现成的可视化数据集，该数据集基于$t$确定 XMAT。此数据集中的描述旨在明确提及答案$a$ ,因此答案与描述之间的关系很清楚
为此，我们通过将可视化描述中的答案替换为特殊单词<空白>来定义任务规范 $t_d$, 该单词正式表示为$t_d=\rho(d,a)$ ,其中$\rho(\boldsymbol{d},\boldsymbol{a})$是生成空白描述的函数。$t_d$中的下标表示基于可视化描述提取任务规范。基于此定义，具有任务规范的联合分布$\mathfrak{p}_D(\boldsymbol{a},\boldsymbol{I},\boldsymbol{b},\boldsymbol{t}_d)=$ $\int$p$(\boldsymbol{a},\boldsymbol{I},\boldsymbol{b},\boldsymbol{t}_d,\boldsymbol{d})d\boldsymbol{d}$由
$$\mathrm{p}(\boldsymbol{a},\boldsymbol{I},\boldsymbol{b},\boldsymbol{t}_d,\boldsymbol{d})=\mathrm{p}(\boldsymbol{t}_d|\boldsymbol{d},\boldsymbol{a})\mathrm{p}_{\mathcal{V}}(\boldsymbol{a},\boldsymbol{I},\boldsymbol{b},\boldsymbol{d})$$
(3) | [
[
"0.3228",
"0.09216",
"0.3708",
"0.09216",
"0.3708",
"0.1029",
"0.3228",
"0.1029"
],
[
"0.4043",
"0.09216",
"0.4836",
"0.09216",
"0.4836",
"0.1042",
"0.4043",
"0.1042"
],
[
"0.2998",
"0.1061",
"0.834",
"0.1061",
"0.834",
"0.1206",
"0.2998",
"0.1206"
],
[
"0.4019",
"0.1218",
"0.4631",
"0.1218",
"0.4631",
"0.132",
"0.4019",
"0.132"
],
[
"0.3057",
"0.1364",
"0.4043",
"0.1364",
"0.4043",
"0.1477",
"0.3057",
"0.1477"
],
[
"0.4004",
"0.1338",
"0.869",
"0.1351",
"0.869",
"0.1483",
"0.4004",
"0.1477"
],
[
"0.3145",
"0.1622",
"0.3816",
"0.1622",
"0.3816",
"0.1768",
"0.3145",
"0.1768"
],
[
"0.4019",
"0.1635",
"0.83",
"0.1635",
"0.83",
"0.178",
"0.4019",
"0.178"
],
[
"0.4019",
"0.178",
"0.4739",
"0.178",
"0.4739",
"0.1895",
"0.4019",
"0.1895"
],
[
"0.1584",
"0.1901",
"0.1732",
"0.1901",
"0.1732",
"0.1945",
"0.1584",
"0.1945"
],
[
"0.3145",
"0.1925",
"0.3816",
"0.1925",
"0.3816",
"0.2039",
"0.3145",
"0.2039"
],
[
"0.4028",
"0.1913",
"0.88",
"0.1913",
"0.88",
"0.2052",
"0.4028",
"0.2052"
],
[
"0.09644",
"0.2134",
"0.1503",
"0.2134",
"0.1503",
"0.2203",
"0.09644",
"0.2203"
],
[
"0.1993",
"0.214",
"0.2157",
"0.2134",
"0.2157",
"0.2172",
"0.1993",
"0.2184"
],
[
"0.3179",
"0.2203",
"0.3757",
"0.2172",
"0.3774",
"0.2317",
"0.3186",
"0.2349"
],
[
"0.4004",
"0.2184",
"0.8364",
"0.2203",
"0.8364",
"0.2343",
"0.4004",
"0.2317"
],
[
"0.2115",
"0.2343",
"0.2646",
"0.2343",
"0.2646",
"0.2443",
"0.2115",
"0.2443"
],
[
"0.4004",
"0.2343",
"0.4558",
"0.2361",
"0.4558",
"0.2474",
"0.4004",
"0.2456"
],
[
"0.4019",
"0.2468",
"0.826",
"0.2488",
"0.826",
"0.2632",
"0.4019",
"0.2615"
],
[
"0.0948",
"0.2349",
"0.1708",
"0.2349",
"0.1708",
"0.2456",
"0.0948",
"0.2456"
],
[
"0.3179",
"0.25",
"0.3782",
"0.25",
"0.3782",
"0.2615",
"0.3179",
"0.2615"
],
[
"0.0948",
"0.26",
"0.1259",
"0.26",
"0.1259",
"0.2683",
"0.0948",
"0.2683"
],
[
"0.1323",
"0.2615",
"0.183",
"0.2615",
"0.183",
"0.2683",
"0.1323",
"0.2683"
],
[
"0.4043",
"0.2644",
"0.505",
"0.2644",
"0.505",
"0.2747",
"0.4043",
"0.2747"
],
[
"0.3145",
"0.2778",
"0.3782",
"0.2778",
"0.3782",
"0.289",
"0.3145",
"0.289"
],
[
"0.4019",
"0.276",
"0.8667",
"0.276",
"0.8667",
"0.289",
"0.4019",
"0.289"
],
[
"0.1168",
"0.2842",
"0.1708",
"0.2842",
"0.1708",
"0.2954",
"0.1168",
"0.2954"
],
[
"0.4004",
"0.2917",
"0.4355",
"0.2917",
"0.4355",
"0.303",
"0.4004",
"0.303"
],
[
"0.0809",
"0.3093",
"0.8896",
"0.3093",
"0.8896",
"0.3225",
"0.0809",
"0.3225"
],
[
"0.0801",
"0.3225",
"0.3733",
"0.3225",
"0.3733",
"0.336",
"0.0801",
"0.336"
],
[
"0.3694",
"0.3225",
"0.8896",
"0.3225",
"0.8896",
"0.336",
"0.3694",
"0.336"
],
[
"0.0825",
"0.3384",
"0.8896",
"0.3384",
"0.8896",
"0.3484",
"0.0825",
"0.3484"
],
[
"0.0801",
"0.3499",
"0.2133",
"0.3516",
"0.2133",
"0.3655",
"0.0801",
"0.363"
],
[
"0.0809",
"0.3801",
"0.468",
"0.3801",
"0.468",
"0.3933",
"0.0809",
"0.3933"
],
[
"0.5",
"0.377",
"0.7793",
"0.3774",
"0.7793",
"0.3958",
"0.5",
"0.3945"
],
[
"0.0809",
"0.3933",
"0.4673",
"0.3933",
"0.4673",
"0.4072",
"0.0809",
"0.4072"
],
[
"0.524",
"0.404",
"0.8906",
"0.404",
"0.8906",
"0.4172",
"0.524",
"0.4172"
],
[
"0.0809",
"0.4104",
"0.468",
"0.4104",
"0.468",
"0.4236",
"0.0809",
"0.4236"
],
[
"0.0809",
"0.4236",
"0.468",
"0.4236",
"0.468",
"0.4414",
"0.0809",
"0.4414"
],
[
"0.5034",
"0.4192",
"0.888",
"0.4192",
"0.888",
"0.433",
"0.5034",
"0.433"
],
[
"0.0809",
"0.4395",
"0.468",
"0.4395",
"0.468",
"0.4526",
"0.0809",
"0.4526"
],
[
"0.505",
"0.4343",
"0.888",
"0.4343",
"0.888",
"0.4475",
"0.505",
"0.4475"
],
[
"0.0809",
"0.4526",
"0.468",
"0.455",
"0.468",
"0.4697",
"0.0809",
"0.4673"
],
[
"0.505",
"0.4507",
"0.888",
"0.4507",
"0.888",
"0.461",
"0.505",
"0.461"
],
[
"0.5034",
"0.464",
"0.796",
"0.464",
"0.796",
"0.4778",
"0.5034",
"0.4778"
],
[
"0.0825",
"0.4717",
"0.3757",
"0.4717",
"0.3757",
"0.4817",
"0.0825",
"0.4817"
],
[
"0.5034",
"0.4905",
"0.8506",
"0.4912",
"0.8506",
"0.506",
"0.5034",
"0.505"
],
[
"0.0866",
"0.5005",
"0.4656",
"0.5005",
"0.4656",
"0.5146",
"0.0866",
"0.5146"
],
[
"0.0825",
"0.5176",
"0.4656",
"0.5176",
"0.4656",
"0.528",
"0.0825",
"0.528"
],
[
"0.524",
"0.515",
"0.8896",
"0.515",
"0.8896",
"0.529",
"0.524",
"0.529"
],
[
"0.0825",
"0.531",
"0.468",
"0.531",
"0.468",
"0.5444",
"0.0825",
"0.5444"
],
[
"0.5034",
"0.529",
"0.8906",
"0.53",
"0.8906",
"0.5435",
"0.5034",
"0.5425"
],
[
"0.0809",
"0.5444",
"0.4697",
"0.5444",
"0.4697",
"0.5625",
"0.0809",
"0.5625"
],
[
"0.5034",
"0.5454",
"0.8906",
"0.5454",
"0.8906",
"0.5596",
"0.5034",
"0.5596"
],
[
"0.076",
"0.5566",
"0.4739",
"0.556",
"0.4739",
"0.5806",
"0.076",
"0.582"
],
[
"0.5024",
"0.5596",
"0.8896",
"0.56",
"0.8896",
"0.5747",
"0.5024",
"0.5737"
],
[
"0.0825",
"0.579",
"0.4673",
"0.579",
"0.4673",
"0.5923",
"0.0825",
"0.5923"
],
[
"0.5034",
"0.5757",
"0.8896",
"0.5757",
"0.8896",
"0.589",
"0.5034",
"0.589"
],
[
"0.0768",
"0.5913",
"0.4722",
"0.5913",
"0.4722",
"0.6157",
"0.0768",
"0.6157"
],
[
"0.505",
"0.5923",
"0.8896",
"0.5923",
"0.8896",
"0.603",
"0.505",
"0.603"
],
[
"0.5034",
"0.606",
"0.8906",
"0.606",
"0.8906",
"0.619",
"0.5034",
"0.619"
],
[
"0.0825",
"0.6123",
"0.468",
"0.6123",
"0.468",
"0.627",
"0.0825",
"0.627"
],
[
"0.0784",
"0.6255",
"0.4714",
"0.6255",
"0.4714",
"0.6465",
"0.0784",
"0.6465"
],
[
"0.5034",
"0.6206",
"0.8906",
"0.6206",
"0.8906",
"0.634",
"0.5034",
"0.634"
],
[
"0.5063",
"0.637",
"0.8896",
"0.637",
"0.8896",
"0.648",
"0.5063",
"0.648"
],
[
"0.0825",
"0.648",
"0.4495",
"0.648",
"0.4495",
"0.6577",
"0.0825",
"0.6577"
],
[
"0.5024",
"0.6494",
"0.8906",
"0.651",
"0.8906",
"0.6655",
"0.5024",
"0.664"
],
[
"0.5034",
"0.6655",
"0.6504",
"0.6655",
"0.6504",
"0.6787",
"0.5034",
"0.6787"
],
[
"0.0866",
"0.6753",
"0.3186",
"0.6753",
"0.3186",
"0.6895",
"0.0866",
"0.6895"
],
[
"0.3105",
"0.6753",
"0.468",
"0.6753",
"0.468",
"0.6895",
"0.3105",
"0.6895"
],
[
"0.0825",
"0.6914",
"0.468",
"0.6914",
"0.468",
"0.705",
"0.0825",
"0.705"
],
[
"0.5024",
"0.6914",
"0.6807",
"0.6924",
"0.6807",
"0.707",
"0.5024",
"0.7056"
],
[
"0.0809",
"0.705",
"0.4656",
"0.705",
"0.4656",
"0.7183",
"0.0809",
"0.7183"
],
[
"0.0801",
"0.7197",
"0.468",
"0.72",
"0.468",
"0.735",
"0.0801",
"0.7344"
],
[
"0.524",
"0.7163",
"0.8906",
"0.7163",
"0.8906",
"0.73",
"0.524",
"0.73"
],
[
"0.0809",
"0.7363",
"0.468",
"0.7363",
"0.468",
"0.75",
"0.0809",
"0.75"
],
[
"0.5034",
"0.732",
"0.8906",
"0.732",
"0.8906",
"0.7456",
"0.5034",
"0.7456"
],
[
"0.5024",
"0.7456",
"0.8906",
"0.747",
"0.8906",
"0.761",
"0.5024",
"0.7603"
],
[
"0.0809",
"0.752",
"0.468",
"0.752",
"0.468",
"0.765",
"0.0809",
"0.765"
],
[
"0.0809",
"0.7676",
"0.1969",
"0.7676",
"0.1969",
"0.781",
"0.0809",
"0.781"
],
[
"0.505",
"0.763",
"0.5522",
"0.763",
"0.5522",
"0.7744",
"0.505",
"0.7744"
],
[
"0.524",
"0.778",
"0.8906",
"0.778",
"0.8906",
"0.7915",
"0.524",
"0.7915"
],
[
"0.0866",
"0.797",
"0.273",
"0.797",
"0.273",
"0.81",
"0.0866",
"0.81"
],
[
"0.2778",
"0.797",
"0.468",
"0.797",
"0.468",
"0.81",
"0.2778",
"0.81"
],
[
"0.505",
"0.7935",
"0.892",
"0.7935",
"0.892",
"0.8066",
"0.505",
"0.8066"
],
[
"0.2695",
"0.8003",
"0.282",
"0.8003",
"0.282",
"0.8047",
"0.2695",
"0.8047"
],
[
"0.0825",
"0.813",
"0.4673",
"0.813",
"0.4673",
"0.8237",
"0.0825",
"0.8237"
],
[
"0.505",
"0.8096",
"0.8896",
"0.8096",
"0.8896",
"0.8228",
"0.505",
"0.8228"
],
[
"0.0809",
"0.827",
"0.4673",
"0.827",
"0.4673",
"0.8403",
"0.0809",
"0.8403"
],
[
"0.5034",
"0.8237",
"0.8906",
"0.8237",
"0.8906",
"0.837",
"0.5034",
"0.837"
],
[
"0.0825",
"0.8433",
"0.4673",
"0.8433",
"0.4673",
"0.854",
"0.0825",
"0.854"
],
[
"0.505",
"0.839",
"0.8906",
"0.839",
"0.8906",
"0.853",
"0.505",
"0.853"
],
[
"0.0825",
"0.8594",
"0.4656",
"0.8594",
"0.4656",
"0.869",
"0.0825",
"0.869"
],
[
"0.5024",
"0.854",
"0.661",
"0.853",
"0.661",
"0.867",
"0.5024",
"0.8687"
],
[
"0.0825",
"0.8726",
"0.468",
"0.8726",
"0.468",
"0.886",
"0.0825",
"0.886"
],
[
"0.5537",
"0.884",
"0.8447",
"0.884",
"0.8447",
"0.898",
"0.5537",
"0.898"
],
[
"0.8364",
"0.8833",
"0.8906",
"0.8833",
"0.8906",
"0.898",
"0.8364",
"0.898"
],
[
"0.0801",
"0.886",
"0.3757",
"0.887",
"0.3757",
"0.901",
"0.0801",
"0.8994"
]
] | [
"task tu",
"wordgroups",
"appliance.n.02 freezer, hair dryer, refrigerator, fridge, oven, dishwasher, washing",
"machine,",
"opening.n.10",
"manhole, rear window, exit, nozzle, car window, coin slot, spout, vent,",
"food.n.02",
"potatoes, french fries, chicken, melted cheese, tomatoes, sausage,",
"wegetable,.",
"me.",
"move.v.03",
"twisted, rufled, curving, curved, wheeled, rolled, coiled, turned, rippled,",
"card_game.n.",
"etic",
"act.n.02",
"fishing, skateboarding, sailing,playing baseball, surfing,traveling",
"tact_sport.n.o",
"driving,",
"middle, in corner, center, parking space, playground, landscape,",
"outdoor_game.n.0",
"area.n.01",
"golf.n.01",
"fieldgame",
"neighborhood,..",
"color.n.01",
"yellow, pink, red, beige, royal blue, amber, aqua, dark red, olive green",
"cricketn.01",
"teal,",
"Figure 4.Illustration of WordNet and constructed word group table. (Left)A subgraph of the WordNet [1o] :Complex",
"hierarchy of words reveals the diverse categ",
"orization of each words. (Right) A set of words sharing common parents in the",
"tree isgrouped as a single word group.Diverse grouping of words reveals diverse visual recognition tasks that can be defined",
"oneachwordgroup",
"给出，其中Uxga（I,q）是具有图像I和问题q的编码",
"4. Unsupervised Task Discovery",
"视觉特征，使用具有参数Φvga的注意力机制。。任务功",
"使用现成的可视化数据集[22]学习任务条件可视化",
"能Tnvaa（q）使用参数nvqa对问题q进行编码。优化需",
"要VQA的训练数据集Pvqa(a,I,Q）的联合分布，其中",
"分类器并不简单，因为缺少任务规范的注释，这是学习",
"来自分布的答案满足aEA-B。我们通过最大化方",
"任务规范向量T的编码器所必需的。为了解决这个问",
"程E(2）中目标的可能性来学习Φvga和nvqa，而预训",
"题，我们提出了无监督的任务发现，它从通过利用语言",
"知识源建模的任务分配中采样任务规范t",
"练任务条件视觉分类器9的参数保持固定",
"4.1. Leveraging linguistic knowledge sources",
"Weaklysupervisedtaskregression利用预训练的任",
"务条件视觉分类器进行由问题9指定的视觉识别需要",
"由问题给出的视觉识别任务通常定义从视觉输人到",
"推断最佳任务特征T。此要求引人了一个学习问题-任",
"一组可能的视觉概念的映射，即单词组。例如，“女人",
"务回归-优化编码器Tnvaa（）以正确预测Ta。由于直接",
"拿着什么？”的问题定义了在图像中查找可容纳对象",
"最小化错误E(Ta.（g)）需要对任务进行额外的监督，",
"的视觉识别任务，这是对单词组{ball，racket，cup，.]",
"因此我们将VQA数据作为弱监督的来源。我们优化了",
"的分类。这种直觉导致了一种简单的方法，通过将任务",
"间接损失ra(alm)logPe(alu,a(g)，这鼓励了以T",
"视为单词组（i.e.可保存对象）来对任务描述t的分布",
"进行建模。使用语言知识源进行无监督任务发现的主",
"和Tmvaa（q）为条件的答案分布相似。通过假设真实任务",
"条件答案分布tg(al）是在VQA数据集中隐式建模的，",
"要原因是，在语言知识源中通常可以访问单词组。我们",
"考虑两个语言知识来源：1）提供视觉数据的视觉描述：",
"我们采用方程E（2）作为弱监督任务回归的自标函数",
"2称为WordNet[10]的结构化词汇数据库。图3说",
"明了我们方法的概述",
"Out-of-vocabulary answering",
"我们通过调整输入表",
"示来学习VQA模型，同时固定预训练任务条件视觉分",
"4.2. Visual description",
"类器Pe(au,T）。此策略允许模型专注于学习从问题中",
"推断视觉识别任务Tnxa（q），而不需要所有可能答案",
"我们使用VisualGenome[22]作为现成的可视化数",
"的数据。一旦推断出任务特征，学习的任务条件视",
"据集，该数据集基于t确定XMAT。此数据集中的描",
"述旨在明确提及答案a，因此答案与描述之间的关系",
"觉分类器Pe(au,）就可以回答预训练的视觉概念，包",
"括词汇外的概念",
"很清楚",
"为此，我们通过将可视化描述中的答案替换为特殊",
"Matching visual features",
"要在不进行微调的情况下",
"单词<空白>来定义任务规范td，该单词正式表",
"",
"在VQA中重用预训练的视觉分类器，视觉特征的",
"示为td=p(d,a），其中p(d,a)是生成空白描述的函",
"语义不应因学习VQA数据集而改变。VQA模型的最",
"数。td中的下标表示基于可视化描述提取任务规范。基",
"新方法实现了这一点，这些模型不微调预训练的视觉",
"于此定义，具有任务规范的联合分布Pp(a,I,b,ta）=",
"特征提取器，而是专注手在提取的特征图上学习注意",
"Jp(a,I,b,ta,d)dd由",
"力机制E[20]。在我们的设置中，我们只需使用相同的",
"p(a, I,b,td, d) = p(tald, a)pv(a, I,b, d)",
" (3)",
"视觉特征提取器E[3]进行预训练和VQA"
] | {
"arxiv_id": "1810.02358",
"language": "zh",
"page_idx": 3
} |
||
1810.02358-zh-4 | 1810.02358-zh-4.jpg | 给出，其中 p$(\boldsymbol{t}_d|\boldsymbol{d},\boldsymbol{a})=\delta(\boldsymbol{t}_d,\rho(\boldsymbol{d},\boldsymbol{a}))$是一个 delta 函数，如果两个输入相同，则返回 1,否则返回0。如图 3 a 所示，我们通过首先采样可视数据$(\boldsymbol{a},\boldsymbol{I},\boldsymbol{b},\boldsymbol{d})$ ,然后从 p$(t_d|\boldsymbol{d},\boldsymbol{a})$采样任务规范$t_d$来采样预训练$(\boldsymbol{a},\boldsymbol{I},\boldsymbol{b},\boldsymbol{t}_w)$所需的数据。此过程也会产生采样描述$d$ , 但是当我们预训练任务条件视觉分类器时，我们并不关心它。对于预训练，我们基于门控循环单元$\mathbb{E}[8]$ 将$t_d$编码为任务特征$\tau_\eta_\mathrm{pre}(t_d)$ ,因为$t_d$是单词序列
对任务规范使用空白描述的主要原因是，定义一组候选词是有效的。例如，空白的描述“一个人拿着\_ ”将空白的候选词限制为可以持有的一组对象。因此， 可以使用空白描述来隐式确定表示视觉识别任务的单词组。
# $4. 3. \textbf{WordNet}$
WordNet [10]是一个词法数据库，用消歧单词实体的有向无环图表示，称为语法集。图 4(左)中说明了 WordNet 的示例子图。该图表示单词的层次结构， 其中节点的父级对应于子级中单词的 Hypernym。在WordNet 中，我们将任务规范$t_w$定义为一个节点的同义词集，该节点是多个单词的共同祖先，因为共享一个共同祖先的一组单词构成了一个单词组，并且该单词组还可以定义一个视觉识别任务
基于 WordNet 和可视数据$(\boldsymbol{a},\boldsymbol{I},\boldsymbol{b},\boldsymbol{d})$对任务规范$t_w$进行采样的过程如图 3 b 所示。该过程的主要思想是将以答案 p$(t_w|a)$为条件的任务分布建模为该答案所属的可能单词组的均匀分布，其中任务规范$t_w$是单词组中单词的共同祖先。对分布 p$(t_w|a)$ 进行建模需要两个阶段：1)构造单词组表，将任务规范映射到单词组；2)构造反向单词组表，将答案单词映射到一组任务规范。倒排词组表用于检索答案$a$的一组可能的任务规范，分布 p$(t_w|a)$是该组中任务规范的均匀分布。给定分布 p$(t_w|a)$,联合分布$\mathrm{p}_{\mathcal{D}}(\boldsymbol{a},\boldsymbol{I},\boldsymbol{b},\boldsymbol{t}_w)=\int$p$(\boldsymbol{a},\boldsymbol{I},\boldsymbol{b},\boldsymbol{t}_w,\boldsymbol{d})d\boldsymbol{d}$由
$$\mathrm{p}(\boldsymbol{a},\boldsymbol{I},\boldsymbol{b},\boldsymbol{t}_w,\boldsymbol{d})=\mathrm{p}(\boldsymbol{t}_w|\boldsymbol{a})\mathrm{p}_{\mathcal{V}}(\boldsymbol{a},\boldsymbol{I},\boldsymbol{b},\boldsymbol{d}).$$
(4)
给出。因此，我们从视觉数据集中采样一个四重$( \boldsymbol{a}, \boldsymbol{I}, \boldsymbol{b}, \boldsymbol{d})$ ,并随后采样一个任务规范$t_w$ 。虽然此过程也对描述进行了采样，但我们将其边缘化。对于预训练，我们基于从头开始学习的单词嵌人函数将$t_w$编码为任务规范向量$\tau_\eta_\mathrm{pre}(t_w)$
通过在 WordNet 中选择节点的 synset 作为任务规范$t_w$并将其映射到与其所有后代相对应的一组单词(一个单词组)来构造单词组表。可以定义任何单词组， 而不管其在 WordNet 层次结构中的级别及其成员的词性如何；最大的单词组包含 WordNet 中的所有单词， 其任务规范对应于 WordNet 的根。我们在图 4 中说明了构建的词组表。倒排词组表的构建方式与词组表的倒排索引类似，但映射的范围不是一组索引，而是一组任务规范
# 5. Experiments
我们评估了所提出的框架如何有效地利用外部数据而无需问题来回答视觉问答中的词汇外单词。我们将所提出的方法与具有零镜头图像分类思想$\mathbb{E}[11]$ 和新颖的对象字幕$\mathbb{E}[4,37]$ 的基线进行了比较，这与所提出的问题有关。我们还分析了用于预训练的外部数据的影响，并可视化了通过弱监督任务回归学习的问题和任务规范之间的映射。值得注意的是，评估设置与零拍学习文献$\mathbb{E}[42]$ 中的评估设置不具有可比性，因为我们的目标不是泛化到完全看不见的类，而是利用外部数据中暴露的类标签。我们开源了实验中使用的所有代码和数据集，以便于重现本文的结果
# 5.1. Datasets
Pretraining 我们学习在 VisualGenome 数据集 [22] 中最经常观察到的 3,000 个对象和 1,000 个属性的视觉概念。为了预训练任务条件视觉分类器，我们使用边界框注释构建外部视觉数据，这些数据带有区域描述。然后，从区域描述中提取视觉词(答案),构建视觉数据四元组$(a,\boldsymbol{I},\boldsymbol{b},\boldsymbol{d})$ 。我们使用来自 80,602张图像的 1,169,708 个区域来构建训练数据。要使用WordNet [10] ,我们使用 VisualGenome 数据集中的语法集注释将视觉单词映射到语法集，而注释未涵盖的单词将使用 Textblob [27]映射到语法集。
Dataset construction 我们重新利用 VQAv2 数据集来构建训练/测试拆分，如图 1 所示。我们使用 VQAv2的训练和验证集。为了确保在预训练期间出现每个词汇外答案，我们从用于预训练的所有视觉单词中随机选择词汇外答案 (3,813 个中的 954 个)。由于我们专注于视觉单词的可传递性，因此在我们的评估中不考虑有关是/否和数字的答案。根据选定的词汇外答案，我们生成 3 个问题拆分-用于培训的 462,788, 用于验证的 51,421 和用于测试的 20,802。训练和验证拆分不包含词汇外答案，而测试拆分仅包含词汇外答案。为了评估模型，我们遵循标准的 VQA 协议，每个问题都有10 地面实况答案 [5]
# 5.2. Baselines
由于几乎没有探索过利用外部视觉数据进行词汇外答案的视觉问答，因此没有适当的评估基准，因此我们采用以下可分离的基线与我们的算法进行比较：1)答案嵌入，它采用了零拍图像分类$\mathbb{E}[11]$ 的思想，该思想学习了从视觉特征到预训练答案嵌入的映射，其中我们分别使用 GloVe$\underline{\mathrm{E}}$X。请注意，可分离分类器和我们提出的模型是用相同的数据训练的。
# $5. 3. \textbf{ Results}$
Model comparisons 图 5 说明了模型与基线的比较结果。在这个实验中，我们用 6 种不同的随机种子进行VQA 自适应，并绘制它们的均值和标准差。标准 VQA
$^{1}$https://github.com/HyeonwooNoh/vqa\_task\_discovery | [
[
"0.0809",
"0.09344",
"0.468",
"0.09344",
"0.468",
"0.1067",
"0.0809",
"0.1067"
],
[
"0.5",
"0.09155",
"0.6396",
"0.09216",
"0.6396",
"0.1067",
"0.5",
"0.1061"
],
[
"0.0825",
"0.10986",
"0.4656",
"0.10986",
"0.4656",
"0.1206",
"0.0825",
"0.1206"
],
[
"0.524",
"0.11743",
"0.8906",
"0.11743",
"0.8906",
"0.1307",
"0.524",
"0.1307"
],
[
"0.0809",
"0.1237",
"0.4656",
"0.1237",
"0.4656",
"0.137",
"0.0809",
"0.137"
],
[
"0.09644",
"0.1395",
"0.4697",
"0.1395",
"0.4697",
"0.1528",
"0.09644",
"0.1528"
],
[
"0.505",
"0.1338",
"0.8896",
"0.1338",
"0.8896",
"0.1445",
"0.505",
"0.1445"
],
[
"0.505",
"0.1477",
"0.8906",
"0.1477",
"0.8906",
"0.161",
"0.505",
"0.161"
],
[
"0.0825",
"0.154",
"0.468",
"0.154",
"0.468",
"0.1674",
"0.0825",
"0.1674"
],
[
"0.5034",
"0.1622",
"0.8896",
"0.1622",
"0.8896",
"0.1755",
"0.5034",
"0.1755"
],
[
"0.0825",
"0.1686",
"0.468",
"0.1686",
"0.468",
"0.1818",
"0.0825",
"0.1818"
],
[
"0.5034",
"0.178",
"0.8896",
"0.178",
"0.8896",
"0.1913",
"0.5034",
"0.1913"
],
[
"0.0825",
"0.1843",
"0.468",
"0.1843",
"0.468",
"0.1976",
"0.0825",
"0.1976"
],
[
"0.0801",
"0.1989",
"0.4387",
"0.1989",
"0.4387",
"0.2134",
"0.0801",
"0.2134"
],
[
"0.5063",
"0.1945",
"0.8896",
"0.1945",
"0.8896",
"0.2052",
"0.5063",
"0.2052"
],
[
"0.505",
"0.2084",
"0.8906",
"0.2084",
"0.8906",
"0.2216",
"0.505",
"0.2216"
],
[
"0.10046",
"0.2172",
"0.468",
"0.2172",
"0.468",
"0.2311",
"0.10046",
"0.2311"
],
[
"0.5024",
"0.2229",
"0.8906",
"0.2235",
"0.8906",
"0.2374",
"0.5024",
"0.2361"
],
[
"0.0801",
"0.2317",
"0.429",
"0.2311",
"0.429",
"0.2456",
"0.0801",
"0.2468"
],
[
"0.5034",
"0.238",
"0.8906",
"0.238",
"0.8906",
"0.252",
"0.5034",
"0.252"
],
[
"0.0825",
"0.2474",
"0.468",
"0.2474",
"0.468",
"0.2615",
"0.0825",
"0.2615"
],
[
"0.5024",
"0.252",
"0.8906",
"0.2524",
"0.8906",
"0.2676",
"0.5024",
"0.2664"
],
[
"0.0809",
"0.2615",
"0.468",
"0.262",
"0.468",
"0.2766",
"0.0809",
"0.276"
],
[
"0.5034",
"0.2683",
"0.7686",
"0.2683",
"0.7686",
"0.2822",
"0.5034",
"0.2822"
],
[
"0.7695",
"0.2715",
"0.778",
"0.2715",
"0.778",
"0.2766",
"0.7695",
"0.2766"
],
[
"0.0801",
"0.2778",
"0.1217",
"0.2778",
"0.1217",
"0.2922",
"0.0801",
"0.2922"
],
[
"0.5024",
"0.2922",
"0.607",
"0.2935",
"0.607",
"0.308",
"0.5024",
"0.307"
],
[
"0.0809",
"0.3145",
"0.1912",
"0.3145",
"0.1912",
"0.3276",
"0.0809",
"0.3276"
],
[
"0.5063",
"0.3164",
"0.8896",
"0.3176",
"0.8896",
"0.332",
"0.5063",
"0.3308"
],
[
"0.505",
"0.3333",
"0.8896",
"0.3333",
"0.8896",
"0.3467",
"0.505",
"0.3467"
],
[
"0.10046",
"0.3416",
"0.468",
"0.3416",
"0.468",
"0.3547",
"0.10046",
"0.3547"
],
[
"0.0801",
"0.3547",
"0.468",
"0.356",
"0.468",
"0.3706",
"0.0801",
"0.3694"
],
[
"0.505",
"0.3499",
"0.888",
"0.3499",
"0.888",
"0.3599",
"0.505",
"0.3599"
],
[
"0.5034",
"0.363",
"0.892",
"0.363",
"0.892",
"0.377",
"0.5034",
"0.377"
],
[
"0.0825",
"0.3718",
"0.4697",
"0.3718",
"0.4697",
"0.385",
"0.0825",
"0.385"
],
[
"0.5024",
"0.377",
"0.8896",
"0.377",
"0.8896",
"0.3901",
"0.5024",
"0.3901"
],
[
"0.0801",
"0.385",
"0.468",
"0.3865",
"0.468",
"0.401",
"0.0801",
"0.3997"
],
[
"0.0801",
"0.3997",
"0.468",
"0.401",
"0.468",
"0.414",
"0.0801",
"0.4136"
],
[
"0.5063",
"0.3945",
"0.8896",
"0.3945",
"0.8896",
"0.4048",
"0.5063",
"0.4048"
],
[
"0.5034",
"0.4077",
"0.8896",
"0.4077",
"0.8896",
"0.4216",
"0.5034",
"0.4216"
],
[
"0.0825",
"0.4167",
"0.464",
"0.4167",
"0.464",
"0.43",
"0.0825",
"0.43"
],
[
"0.5024",
"0.4216",
"0.8906",
"0.4236",
"0.8906",
"0.4375",
"0.5024",
"0.435"
],
[
"0.08417",
"0.433",
"0.4673",
"0.433",
"0.4673",
"0.443",
"0.08417",
"0.443"
],
[
"0.505",
"0.4382",
"0.8906",
"0.4382",
"0.8906",
"0.452",
"0.505",
"0.452"
],
[
"0.0809",
"0.4463",
"0.3245",
"0.4463",
"0.3245",
"0.4602",
"0.0809",
"0.4602"
],
[
"0.505",
"0.4539",
"0.8164",
"0.4539",
"0.8164",
"0.4673",
"0.505",
"0.4673"
],
[
"0.1013",
"0.4653",
"0.4673",
"0.4653",
"0.4673",
"0.4792",
"0.1013",
"0.4792"
],
[
"0.0809",
"0.481",
"0.468",
"0.481",
"0.468",
"0.4944",
"0.0809",
"0.4944"
],
[
"0.5063",
"0.476",
"0.8906",
"0.4766",
"0.8906",
"0.4912",
"0.5063",
"0.4905"
],
[
"0.0825",
"0.4956",
"0.468",
"0.4956",
"0.468",
"0.509",
"0.0825",
"0.509"
],
[
"0.505",
"0.4944",
"0.8896",
"0.4944",
"0.8896",
"0.505",
"0.505",
"0.505"
],
[
"0.0809",
"0.511",
"0.468",
"0.511",
"0.468",
"0.5244",
"0.0809",
"0.5244"
],
[
"0.5034",
"0.5083",
"0.8906",
"0.5083",
"0.8906",
"0.5215",
"0.5034",
"0.5215"
],
[
"0.0809",
"0.5264",
"0.4673",
"0.5264",
"0.4673",
"0.5405",
"0.0809",
"0.5405"
],
[
"0.505",
"0.5244",
"0.8896",
"0.5244",
"0.8896",
"0.535",
"0.505",
"0.535"
],
[
"0.0825",
"0.5405",
"0.4673",
"0.5405",
"0.4673",
"0.5537",
"0.0825",
"0.5537"
],
[
"0.5034",
"0.5386",
"0.8906",
"0.5386",
"0.8906",
"0.552",
"0.5034",
"0.552"
],
[
"0.0825",
"0.558",
"0.4673",
"0.558",
"0.4673",
"0.5684",
"0.0825",
"0.5684"
],
[
"0.5034",
"0.553",
"0.892",
"0.553",
"0.892",
"0.5664",
"0.5034",
"0.5664"
],
[
"0.0825",
"0.5713",
"0.468",
"0.5713",
"0.468",
"0.5854",
"0.0825",
"0.5854"
],
[
"0.5034",
"0.5684",
"0.8906",
"0.5684",
"0.8906",
"0.582",
"0.5034",
"0.582"
],
[
"0.0809",
"0.5854",
"0.4673",
"0.5854",
"0.4673",
"0.5986",
"0.0809",
"0.5986"
],
[
"0.5034",
"0.5835",
"0.8896",
"0.584",
"0.8896",
"0.5977",
"0.5034",
"0.5967"
],
[
"0.0809",
"0.6016",
"0.468",
"0.6016",
"0.468",
"0.6157",
"0.0809",
"0.6157"
],
[
"0.5034",
"0.5977",
"0.8906",
"0.5977",
"0.8906",
"0.6113",
"0.5034",
"0.6113"
],
[
"0.0784",
"0.617",
"0.3635",
"0.6157",
"0.3635",
"0.63",
"0.0784",
"0.632"
],
[
"0.505",
"0.613",
"0.8896",
"0.613",
"0.8896",
"0.627",
"0.505",
"0.627"
],
[
"0.505",
"0.629",
"0.8896",
"0.629",
"0.8896",
"0.6426",
"0.505",
"0.6426"
],
[
"0.5034",
"0.6445",
"0.652",
"0.6445",
"0.652",
"0.6577",
"0.5034",
"0.6577"
],
[
"0.1356",
"0.653",
"0.4495",
"0.653",
"0.4495",
"0.6665",
"0.1356",
"0.6665"
],
[
"0.4436",
"0.653",
"0.4673",
"0.653",
"0.4673",
"0.6655",
"0.4436",
"0.6655"
],
[
"0.5034",
"0.6694",
"0.614",
"0.6694",
"0.614",
"0.683",
"0.5034",
"0.683"
],
[
"0.0809",
"0.6875",
"0.468",
"0.6875",
"0.468",
"0.7007",
"0.0809",
"0.7007"
],
[
"0.524",
"0.6934",
"0.8896",
"0.6934",
"0.8896",
"0.707",
"0.524",
"0.707"
],
[
"0.0809",
"0.7007",
"0.468",
"0.702",
"0.468",
"0.7163",
"0.0809",
"0.7153"
],
[
"0.505",
"0.7085",
"0.8896",
"0.7085",
"0.8896",
"0.7217",
"0.505",
"0.7217"
],
[
"0.0825",
"0.7173",
"0.468",
"0.7173",
"0.468",
"0.731",
"0.0825",
"0.731"
],
[
"0.505",
"0.7246",
"0.8896",
"0.7246",
"0.8896",
"0.735",
"0.505",
"0.735"
],
[
"0.0825",
"0.733",
"0.468",
"0.733",
"0.468",
"0.747",
"0.0825",
"0.747"
],
[
"0.5034",
"0.738",
"0.8906",
"0.738",
"0.8906",
"0.752",
"0.5034",
"0.752"
],
[
"0.0801",
"0.747",
"0.2778",
"0.7485",
"0.2778",
"0.763",
"0.0801",
"0.761"
],
[
"0.505",
"0.7524",
"0.8906",
"0.7524",
"0.8906",
"0.7666",
"0.505",
"0.7666"
],
[
"0.10297",
"0.768",
"0.4673",
"0.768",
"0.4673",
"0.779",
"0.10297",
"0.779"
],
[
"0.505",
"0.768",
"0.8906",
"0.768",
"0.8906",
"0.782",
"0.505",
"0.782"
],
[
"0.0809",
"0.782",
"0.468",
"0.782",
"0.468",
"0.7954",
"0.0809",
"0.7954"
],
[
"0.5034",
"0.784",
"0.755",
"0.784",
"0.755",
"0.7974",
"0.5034",
"0.7974"
],
[
"0.0801",
"0.7954",
"0.4673",
"0.797",
"0.4673",
"0.81",
"0.0801",
"0.8096"
],
[
"0.0825",
"0.811",
"0.468",
"0.811",
"0.468",
"0.8247",
"0.0825",
"0.8247"
],
[
"0.5034",
"0.811",
"0.597",
"0.811",
"0.597",
"0.8228",
"0.5034",
"0.8228"
],
[
"0.0825",
"0.827",
"0.468",
"0.827",
"0.468",
"0.8403",
"0.0825",
"0.8403"
],
[
"0.5073",
"0.8335",
"0.888",
"0.8335",
"0.888",
"0.8467",
"0.5073",
"0.8467"
],
[
"0.0809",
"0.8423",
"0.468",
"0.8423",
"0.468",
"0.856",
"0.0809",
"0.856"
],
[
"0.505",
"0.8496",
"0.888",
"0.8496",
"0.888",
"0.86",
"0.505",
"0.86"
],
[
"0.0809",
"0.857",
"0.468",
"0.857",
"0.468",
"0.8706",
"0.0809",
"0.8706"
],
[
"0.505",
"0.8657",
"0.8896",
"0.8657",
"0.8896",
"0.8755",
"0.505",
"0.8755"
],
[
"0.0809",
"0.8726",
"0.4631",
"0.8726",
"0.4631",
"0.886",
"0.0809",
"0.886"
],
[
"0.0825",
"0.888",
"0.1626",
"0.888",
"0.1626",
"0.8994",
"0.0825",
"0.8994"
],
[
"0.5186",
"0.886",
"0.8677",
"0.888",
"0.8677",
"0.903",
"0.5186",
"0.901"
]
] | [
"给出，其中p(tad,a)=o(td,p(d,a))是一个delta函",
"5.Experiments",
"数，如果两个输入相同，则返回1，否则返回0。如",
"我们评估了所提出的框架如何有效地利用外部数据",
"图3a所示，我们通过首先采样可视数据a,I,b,d",
"然后从p(tad,a)采样任务规范td来采样预训练",
"而无需问题来回答视觉问答中的词汇外单词。我们将",
"所提出的方法与具有零镜头图像分类思想E[11]和新",
"(a,I,b,tw）所需的数据。此过程也会产生采样描述d，",
"颖的对象字幕E[4，37]的基线进行了比较，这与所提出",
"但是当我们预训练任务条件视觉分类器时，我们并不",
"的问题有关。我们还分析了用于预训练的外部数据的",
"关心它。对于预训练，我们基于门控循环单元E[8]将",
"td编码为任务特征Tnpre（ta），因为td是单词序列",
"影响，并可视化了通过弱监督任务回归学习的问题和",
"任务规范之间的映射。值得注意的是，评估设置与零拍",
"对任务规范使用空白描述的主要原因是，定义一组",
"学习文献E[42]中的评估设置不具有可比性，因为我们",
"候选词是有效的。例如，空白的描述“一个人拿着",
"的目标不是泛化到完全看不见的类，而是利用外部数",
"”将空白的候选词限制为可以持有的一组对象。因此，",
"据中暴露的类标签。我们开源了实验中使用的所有代",
"可以使用空白描述来隐式确定表示视觉识别任务的单",
"码和数据集，以便于重现本文的结果",
"",
"词组。",
"5.1. Datasets",
"4.3.WordNet",
"Pretraining我们学习在VisualGenome数据集[22]",
"中最经常观察到的3,000个对象和1,000个属性的视",
"WordNet【10]是一个词法数据库，用消歧单词实体",
"的有向无环图表示，称为语法集。图4（左）中说明",
"觉概念。为了预训练任务条件视觉分类器，我们使用",
"边界框注释构建外部视觉数据，这些数据带有区域",
"了WordNet的示例子图。该图表示单词的层次结构，",
"描述。然后，从区域描述中提取视觉词（答案)，构建",
"其中节点的父级对应于子级中单词的Hypernym。在",
"WordNet中，我们将任务规范tu定义为一个节点的",
"视觉数据四元组（a，1.6.d）。我们使用来自80.602",
"张图像的1,169,708个区域来构建训练数据。要使用",
"同义词集，该节点是多个单词的共同祖先，因为共享一",
"WordNet[10]】，我们使用VisualGenome数据集中的",
"个共同祖先的一组单词构成了一个单词组，并且该单",
"语法集注释将视觉单词映射到语法集，而注释未涵盖",
"词组还可以定义一个视觉识别任务",
"的单词将使用Textblob[27]映射到语法集。",
"基于WordNet和可视数据（a,I,b,d）对任务规范",
"tu进行采样的过程如图3b所示。该过程的主要",
"Datasetconstruction我们重新利用VQAv2数据集",
"思想是将以答案p(twa）为条件的任务分布建模为",
"来构建训练/测试拆分，如图1所示。我们使用VQAv2",
"该答案所属的可能单词组的均匀分布，其中任务规",
"的训练和验证集。为了确保在预训练期间出现每个词",
"范tw是单词组中单词的共同祖先。对分布p(twla)",
"汇外答案，我们从用于预训练的所有视觉单词中随机",
"进行建模需要两个阶段：1）构造单词组表，将任务",
"选择词汇外答案（3,813个中的954个）。由于我们专注",
"规范映射到单词组；2）构造反向单词组表，将答案",
"于视觉单词的可传递性，因此在我们的评估中不考虑",
"单词映射到一组任务规范。倒排词组表用于检索答",
"有关是/否和数字的答案。根据选定的词汇外答案，我",
"案a的一组可能的任务规范，分布p(tula)是该组中",
"们生成3个问题拆分-用于培训的462,788，用于验证",
"任务规范的均匀分布。给定分布p(twa），联合分布",
"的51,421和用于测试的20,802。训练和验证拆分不包",
"Pp(a, I,b,tw)= J p(a,I,b,tw,d)dd 由",
"含词汇外答案，而测试拆分仅包含词汇外答案。为了",
"评估模型，我们遵循标准的VQA协议，每个问题都有",
"10地面实况答案[5]",
"p(a, I,b,tw,d) =p(twa)pv(a,I,b,d).",
"(4)",
"5.2. Baselines",
"给出。因此，我们从视觉数据集中采样一个四重",
"由于几乎没有探索过利用外部视觉数据进行词汇外",
"(a,I,b,d），并随后采样一个任务规范tu。虽然此",
"答案的视觉问答，因此没有适当的评估基准，因此我们",
"过程也对描述进行了采样，但我们将其边缘化。对于预",
"采用以下可分离的基线与我们的算法进行比较：1）答",
"训练，我们基于从头开始学习的单词嵌人函数将t编",
"案嵌入，它采用了零拍图像分类E[11]的思想，该思想",
"码为任务规范向量Tnpre(tw）",
"学习了从视觉特征到预训练答案嵌人的映射，其中我",
"通过在WordNet中选择节点的symset作为任务规",
"们分别使用GloVeEX。请注意，可分离分类器和我们",
"范tw并将其映射到与其所有后代相对应的一组单词",
"提出的模型是用相同的数据训练的。",
"（一个单词组）来构造单词组表。可以定义任何单词组，",
"而不管其在WordNet层次结构中的级别及其成员的词",
"5.3.Results",
"性如何；最大的单词组包含WordNet中的所有单词，",
"Modelcomparisons图5说明了模型与基线的比较",
"其任务规范对应于WordNet的根。我们在图4中说",
"结果。在这个实验中，我们用6种不同的随机种子进行",
"明了构建的词组表。倒排词组表的构建方式与词组表",
"VQA自适应，并绘制它们的均值和标准差。标准VQA",
"的倒排索引类似，但映射的范围不是一组索引，而是一",
"组任务规范",
"1https://github.com/HyeonwooNoh/vqa_task_discovery"
] | {
"arxiv_id": "1810.02358",
"language": "zh",
"page_idx": 4
} |
||
1910.00058-zh-0 | 1910.00058-zh-0.jpg | 用于学习
的具有多样性的多头注意力基于多语言多模态表示
Po-Yao Huang$^1$,Xiaojun Chang$^2$,Alexander Hauptmann$^1$
$^{1}$Language Technologies Institute, Carnegie Mellon $^{\mathrm{University}}$
$^{2}$Faculty of Information Technology, Monash University [email protected], cxj$273@$gmail.com, [email protected]
## $\operatorname{Abstract}$
With the aim of promoting and understanding the multilingual version of image search, we leverage visual object detection and propose a model with diverse multihead attention to learn grounded multilingual multimodal representations. Specifically, our model attends to different types of textual semantics in two languages and visual objects for fne-grained alignments between sentences and images. We introduce a new objective function which explicitly encourages attention diversity to learn an improved visual-semantic embedding space. We evaluate our model in the German-Image and English-Image matching tasks on the Multi30K dataset, and in the Semantic Textual Similarity task with the English descriptions of visual content. Results show that our model yields a significant performance gain over other methods in all of the three tasks.
## 1 Introduction
联合视觉语义嵌人(VSE)对于许多视觉语言任务的成功至关重要，包括跨模态搜索和检索$\overline{\mathrm{E}}$ (Kiros et al.,2014;Karpathy and Fei-Fei, 2015; Gu et al., 2018) , 视觉问答EXMATHX$\boxed{\mathrm{FMA4}}$ 等，多模态机器翻译。学习 VSE 需要对各个模式中的内容有广泛的了解，并需要深人的对齐策略来关联来自多个视图的补充信息
随着大规模并行英语图像语料库$\operatorname{E}(\operatorname{Lin}$ et al., 2014; Young et al., 2014) 的出现，丰富的研究系列在单语设置下具有先进的学习 VSE。最新作品 (Kiros et al.,2014;Vendrov et al., 2015; Karpathy and Fei-Fei, 2015; Klein et al., 2015; Wang et al., 2016, 2018; Huang et al., 2019b) 利用三元组排名损失在联合嵌入空间中对齐英语句子和图像。在 VSE++ (Faghri et al.,2018) 中，Faghri et al. 通过强调硬阴性样本来改善 VSE。VSE 模型的最新进展探索了丰富英语图像语料库的方法。Shi et al.
(2018)建议使用文本对比对抗样本来增强数据集，以对抗对抗攻击。最近，EHuang et al. (2019a) 利用区域对象的文本语义和对抗性域适应来学习低资源约束下的 VSE。一种新兴趋势在多语言场景中推广学习 VSE。Rajendran et al. (2016) 学习$M$视图表示 ,当并行数据仅在一个透视视图和其余视图之间可用时。PIVOT (Gella et al., 2017)扩展了 Calixto et al.(2017)的工作，以使用图像作为枢轴视图来学习多语言多模式表示。Kádár et al (2018)进一步证实了多语言培训的好处
我们的工作是由E$Gella et al. ( 2017) $驱动的，但有重要的区别。首先，为了解开联合嵌人空间中的对齐方式，我们采用视觉对象检测和多头注意来选择性地将突出的视觉对象与文本短语对齐，从而产生视觉上基于多语言的多模式表示。其次，由于多头注意力$\overline{\mathrm{E}}($Vaswani et al., 2017) 因其效率和从不同角度共同关注信息的能力而吸引人，因此我们建议进一步鼓励注意力之间的多样性，以学习改进的视觉语义嵌入空间。图 1 说明了我们促进多样性的梯度更新。所提出的模型在 Multi30K$\underline{\mathrm{F}}($Elliott et al., 2016) 和语义文本相似性任务$\overline{\mathrm{E}}($Agirre et al., 2012, 2014, 2015) 中的多语言句子图像匹配任务中实现了最先进的性能。
### 2 Related Works
已经解决了学习 VSE 的经典注意力机制。这些机制可以按照 Vaswani et al.(2017) 中讨论的 $\{$查询，密钥，值}的类型进行广泛分类。对于模态内注意，$\{$查询，密钥，值$\}$ 在同一模态内。在 DAN (Nam et al.,2017)中，每个模态中的内容通过多个步骤迭代地参与，具有模态内注意力。在 SCANE(Lee et al., 2018) 中， 在$\boxed{\mathrm{E}}$Anderson et al. (2018) 的区域视觉特征与文本语义之间进行模态间注意。推理时间复杂度为$O( MN)$ (用于为大小为$N$的数据集生成$M$查询表示)。与先前的工作相比，我们利用了模态内多头注意力，与 DAN 相比，它可 | [
[
"0.455",
"0.0849",
"0.5474",
"0.0849",
"0.5474",
"0.0998",
"0.455",
"0.0998"
],
[
"0.2377",
"0.1051",
"0.7656",
"0.1051",
"0.7656",
"0.1194",
"0.2377",
"0.1194"
],
[
"0.2527",
"0.1484",
"0.75",
"0.1484",
"0.75",
"0.1633",
"0.2527",
"0.1633"
],
[
"0.236",
"0.165",
"0.769",
"0.1675",
"0.769",
"0.1818",
"0.236",
"0.1799"
],
[
"0.2627",
"0.1818",
"0.744",
"0.1841",
"0.744",
"0.199",
"0.2627",
"0.1965"
],
[
"0.251",
"0.2007",
"0.75",
"0.1996",
"0.75",
"0.2144",
"0.251",
"0.2156"
],
[
"0.2644",
"0.2642",
"0.3418",
"0.2642",
"0.3418",
"0.279",
"0.2644",
"0.279"
],
[
"0.5156",
"0.2654",
"0.882",
"0.2666",
"0.882",
"0.2808",
"0.5156",
"0.2798"
],
[
"0.5146",
"0.2832",
"0.881",
"0.2832",
"0.881",
"0.2964",
"0.5146",
"0.2964"
],
[
"0.1486",
"0.2893",
"0.4575",
"0.2893",
"0.4575",
"0.3027",
"0.1486",
"0.3027"
],
[
"0.147",
"0.3027",
"0.4592",
"0.304",
"0.4592",
"0.3184",
"0.147",
"0.3176"
],
[
"0.517",
"0.2986",
"0.8833",
"0.2986",
"0.8833",
"0.3123",
"0.517",
"0.3123"
],
[
"0.1486",
"0.3184",
"0.4575",
"0.3184",
"0.4575",
"0.332",
"0.1486",
"0.332"
],
[
"0.5146",
"0.3135",
"0.8833",
"0.3142",
"0.8833",
"0.3289",
"0.5146",
"0.328"
],
[
"0.1511",
"0.3342",
"0.4568",
"0.3342",
"0.4568",
"0.3445",
"0.1511",
"0.3445"
],
[
"0.5146",
"0.33",
"0.8774",
"0.33",
"0.8774",
"0.3433",
"0.5146",
"0.3433"
],
[
"0.147",
"0.3467",
"0.4568",
"0.3467",
"0.4568",
"0.361",
"0.147",
"0.361"
],
[
"0.513",
"0.3457",
"0.8833",
"0.3467",
"0.8833",
"0.361",
"0.513",
"0.3599"
],
[
"0.1486",
"0.3599",
"0.4575",
"0.3599",
"0.4575",
"0.3735",
"0.1486",
"0.3735"
],
[
"0.5146",
"0.3623",
"0.882",
"0.3623",
"0.882",
"0.376",
"0.5146",
"0.376"
],
[
"0.147",
"0.3735",
"0.4575",
"0.3748",
"0.4575",
"0.3896",
"0.147",
"0.3884"
],
[
"0.5146",
"0.379",
"0.882",
"0.379",
"0.882",
"0.3926",
"0.5146",
"0.3926"
],
[
"0.147",
"0.3872",
"0.4592",
"0.3884",
"0.4592",
"0.4026",
"0.147",
"0.402"
],
[
"0.5146",
"0.3955",
"0.882",
"0.3955",
"0.882",
"0.4092",
"0.5146",
"0.4092"
],
[
"0.1503",
"0.405",
"0.4568",
"0.405",
"0.4568",
"0.415",
"0.1503",
"0.415"
],
[
"0.517",
"0.411",
"0.882",
"0.411",
"0.882",
"0.4246",
"0.517",
"0.4246"
],
[
"0.147",
"0.4175",
"0.4575",
"0.4175",
"0.4575",
"0.4316",
"0.147",
"0.4316"
],
[
"0.1486",
"0.4312",
"0.455",
"0.4312",
"0.455",
"0.4448",
"0.1486",
"0.4448"
],
[
"0.517",
"0.4275",
"0.8286",
"0.4275",
"0.8286",
"0.4412",
"0.517",
"0.4412"
],
[
"0.1486",
"0.4465",
"0.4592",
"0.4465",
"0.4592",
"0.4602",
"0.1486",
"0.4602"
],
[
"0.532",
"0.4436",
"0.882",
"0.4448",
"0.882",
"0.459",
"0.532",
"0.4578"
],
[
"0.147",
"0.4602",
"0.4575",
"0.4602",
"0.4575",
"0.4739",
"0.147",
"0.4739"
],
[
"0.5156",
"0.462",
"0.882",
"0.462",
"0.882",
"0.4756",
"0.5156",
"0.4756"
],
[
"0.1486",
"0.4744",
"0.4575",
"0.4744",
"0.4575",
"0.488",
"0.1486",
"0.488"
],
[
"0.517",
"0.478",
"0.882",
"0.478",
"0.882",
"0.4912",
"0.517",
"0.4912"
],
[
"0.147",
"0.4868",
"0.4575",
"0.488",
"0.4575",
"0.503",
"0.147",
"0.502"
],
[
"0.517",
"0.4954",
"0.882",
"0.4954",
"0.882",
"0.506",
"0.517",
"0.506"
],
[
"0.147",
"0.503",
"0.4592",
"0.503",
"0.4592",
"0.5166",
"0.147",
"0.5166"
],
[
"0.5156",
"0.5103",
"0.882",
"0.5103",
"0.882",
"0.524",
"0.5156",
"0.524"
],
[
"0.1503",
"0.5186",
"0.4568",
"0.5186",
"0.4568",
"0.5293",
"0.1503",
"0.5293"
],
[
"0.1486",
"0.531",
"0.4575",
"0.531",
"0.4575",
"0.5444",
"0.1486",
"0.5444"
],
[
"0.5156",
"0.527",
"0.882",
"0.527",
"0.882",
"0.5405",
"0.5156",
"0.5405"
],
[
"0.1486",
"0.5464",
"0.4575",
"0.5464",
"0.4575",
"0.557",
"0.1486",
"0.557"
],
[
"0.5146",
"0.542",
"0.882",
"0.542",
"0.882",
"0.5557",
"0.5146",
"0.5557"
],
[
"0.1486",
"0.561",
"0.4575",
"0.561",
"0.4575",
"0.572",
"0.1486",
"0.572"
],
[
"0.5146",
"0.557",
"0.881",
"0.557",
"0.881",
"0.571",
"0.5146",
"0.571"
],
[
"0.147",
"0.5723",
"0.3225",
"0.5737",
"0.3225",
"0.5874",
"0.147",
"0.586"
],
[
"0.517",
"0.5747",
"0.882",
"0.5747",
"0.882",
"0.5884",
"0.517",
"0.5884"
],
[
"0.5156",
"0.589",
"0.881",
"0.589",
"0.881",
"0.6025",
"0.5156",
"0.6025"
],
[
"0.1184",
"0.597",
"0.1511",
"0.597",
"0.1511",
"0.61",
"0.1184",
"0.61"
],
[
"0.1486",
"0.5957",
"0.2644",
"0.597",
"0.2644",
"0.6123",
"0.1486",
"0.61"
],
[
"0.1486",
"0.601",
"0.1545",
"0.601",
"0.1545",
"0.6055",
"0.1486",
"0.6055"
],
[
"0.5156",
"0.6055",
"0.882",
"0.6055",
"0.882",
"0.619",
"0.5156",
"0.619"
],
[
"0.1184",
"0.6206",
"0.486",
"0.6206",
"0.486",
"0.6343",
"0.1184",
"0.6343"
],
[
"0.513",
"0.6226",
"0.882",
"0.6226",
"0.882",
"0.6357",
"0.513",
"0.6357"
],
[
"0.12006",
"0.638",
"0.4895",
"0.638",
"0.4895",
"0.652",
"0.12006",
"0.652"
],
[
"0.513",
"0.6396",
"0.882",
"0.6396",
"0.882",
"0.653",
"0.513",
"0.653"
],
[
"0.12177",
"0.654",
"0.4895",
"0.654",
"0.4895",
"0.6675",
"0.12177",
"0.6675"
],
[
"0.5146",
"0.654",
"0.781",
"0.655",
"0.781",
"0.6694",
"0.5146",
"0.6685"
],
[
"0.12006",
"0.6685",
"0.486",
"0.6685",
"0.486",
"0.6816",
"0.12006",
"0.6816"
],
[
"0.1184",
"0.685",
"0.486",
"0.685",
"0.486",
"0.6987",
"0.1184",
"0.6987"
],
[
"0.5156",
"0.6914",
"0.529",
"0.6914",
"0.529",
"0.6997",
"0.5156",
"0.6997"
],
[
"0.531",
"0.6934",
"0.5415",
"0.6934",
"0.5415",
"0.697",
"0.531",
"0.697"
],
[
"0.5483",
"0.6885",
"0.679",
"0.6885",
"0.679",
"0.703",
"0.5483",
"0.703"
],
[
"0.12006",
"0.703",
"0.4878",
"0.703",
"0.4878",
"0.7163",
"0.12006",
"0.7163"
],
[
"0.12006",
"0.7188",
"0.43",
"0.7188",
"0.43",
"0.7324",
"0.12006",
"0.7324"
],
[
"0.517",
"0.7188",
"0.882",
"0.7188",
"0.882",
"0.7324",
"0.517",
"0.7324"
],
[
"0.1394",
"0.7354",
"0.4878",
"0.7354",
"0.4878",
"0.749",
"0.1394",
"0.749"
],
[
"0.5156",
"0.734",
"0.882",
"0.734",
"0.882",
"0.749",
"0.5156",
"0.749"
],
[
"0.1184",
"0.75",
"0.486",
"0.75",
"0.486",
"0.763",
"0.1184",
"0.763"
],
[
"0.513",
"0.7505",
"0.881",
"0.75",
"0.881",
"0.764",
"0.513",
"0.7656"
],
[
"0.12177",
"0.767",
"0.4902",
"0.767",
"0.4902",
"0.781",
"0.12177",
"0.781"
],
[
"0.5156",
"0.767",
"0.882",
"0.767",
"0.882",
"0.781",
"0.5156",
"0.781"
],
[
"0.12177",
"0.783",
"0.4902",
"0.783",
"0.4902",
"0.7964",
"0.12177",
"0.7964"
],
[
"0.517",
"0.783",
"0.882",
"0.783",
"0.882",
"0.7964",
"0.517",
"0.7964"
],
[
"0.12006",
"0.799",
"0.4902",
"0.799",
"0.4902",
"0.8125",
"0.12006",
"0.8125"
],
[
"0.5156",
"0.799",
"0.882",
"0.799",
"0.882",
"0.8125",
"0.5156",
"0.8125"
],
[
"0.1184",
"0.8154",
"0.4878",
"0.8154",
"0.4878",
"0.829",
"0.1184",
"0.829"
],
[
"0.5146",
"0.8135",
"0.882",
"0.814",
"0.882",
"0.829",
"0.5146",
"0.8276"
],
[
"0.12006",
"0.8306",
"0.4895",
"0.8306",
"0.4895",
"0.844",
"0.12006",
"0.844"
],
[
"0.5146",
"0.8306",
"0.882",
"0.8306",
"0.882",
"0.844",
"0.5146",
"0.844"
],
[
"0.12006",
"0.847",
"0.4895",
"0.847",
"0.4895",
"0.861",
"0.12006",
"0.861"
],
[
"0.5156",
"0.847",
"0.8833",
"0.847",
"0.8833",
"0.861",
"0.5156",
"0.861"
],
[
"0.12006",
"0.8633",
"0.4878",
"0.8633",
"0.4878",
"0.877",
"0.12006",
"0.877"
],
[
"0.5146",
"0.8633",
"0.881",
"0.8633",
"0.881",
"0.877",
"0.5146",
"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"
],
[
"0.12006",
"0.8955",
"0.4895",
"0.8955",
"0.4895",
"0.909",
"0.12006",
"0.909"
],
[
"0.513",
"0.8955",
"0.881",
"0.8955",
"0.881",
"0.909",
"0.513",
"0.909"
]
] | [
"用于学习",
"的具有多样性的多头注意力基于多语言多模态表示",
"Po-Yao Huang', Xiaojun Chang?, Alexander Hauptmann",
"ILanguage Technologies Institute, Carnegie Mellon University",
"2Faculty of Information Technology, Monash University",
"[email protected], [email protected], [email protected]",
"Abstract",
"(2018）建议使用文本对比对抗样本来增强数",
"据集，以对抗对抗攻击。最近，EHuangetal.",
"With the aim of promoting and under-",
"standing the multilingual version of image",
"（2019a）利用区域对象的文本语义和对抗性域",
"search, we leverage visual object detection",
"适应来学习低资源约束下的VSE。一种新兴",
"andproposeamodelwithdiversemulti-",
"趋势在多语言场景中推广学习VSE。Rajen",
"head attention to learn grounded multilin-",
"dranetal.（2016）学习M视图表示，当并行数",
"gual multimodal representations. Specifi-",
"据仅在一个透视视图和其余视图之间可用时。",
"cally, our model attends to different types",
"PIVOT (Gella et al.,2017)扩展了Calixto",
"of textual semantics in two languages and",
"etal.（2017）的工作，以使用图像作为枢轴视",
"visual objects for fine-grained alignments",
"图来学习多语言多模式表示。Kadaretal.",
"between sentences and images. We intro-",
"duce a new objective function which ex.",
"（2018）进一步证实了多语言培训的好处",
"plicitly encourages attention diversity to",
"我们的工作是由EGella etal.（2017）驱动",
"learn an improved visual-semantic embed",
"的，但有重要的区别。首先，为了解开联合嵌",
"ding space. We evaluate our model in the",
"入空间中的对齐方式，我们采用视觉对象检测",
"German-Image and English-Image match-",
"和多头注意来选择性地将突出的视觉对象与文",
"ing tasks on the Multi30K dataset, and in",
"本短语对齐，从而产生视觉上基于多语言的多",
"the SemanticTextual Similaritytask with",
"the English descriptions of visual content.",
"模式表示。其次，由于多头注意力E(Vaswani",
"Results showthat our model yields a signif-",
"etal.，2017）因其效率和从不同角度共同关注",
"icant performancegain overother methods",
"信息的能力而吸引人，因此我们建议进一步鼓",
"in all of the three tasks.",
"励注意力之间的多样性，以学习改进的视觉语",
"义嵌入空间。图1说明了我们促进多样性的梯",
"1",
"Introduction",
"",
"度更新。所提出的模型在Multi30KE(Elliott",
"联合视觉语义嵌入（VSE)对于许多视觉语言任",
"etal.，2016）和语义文本相似性任务E（Agirre",
"务的成功至关重要，包括跨模态搜索和检索E",
"etal.，2012，2014，2015）中的多语言句子图像",
"(Kiros et al., 2014; Karpathy and Fei-Fei, 2015;",
"匹配任务中实现了最先进的性能。",
"Guetal..2018），视觉问答FXMATHXFMA4",
"等，多模态机器翻译。学习VSE需要对各个",
"2",
"",
"Related Works",
"模式中的内容有广泛的了解，并需要深人的对",
"齐策略来关联来自多个视图的补充信息",
"已经解决了学习VSE的经典注意力机制。这",
"随着大规模并行英语图像语料库E(Lin",
"些机制可以按照Vaswanietal.（2017）中讨论",
"etal.,2014;Youngetal.,2014)的出现，丰富的",
"的【查询，密钥，值)的类型进行广泛分类。对",
"研究系列在单语设置下具有先进的学习VSE。",
"于模态内注意，查询，密钥，值}在同一模态",
"最新作品(Kirosetal.,2014;Vendrovet al.,",
"内。在DAN（Nametal.，2017）中，每个模",
"2015; Karpathy and Fei-Fei, 2015; Klein et al.",
"态中的内容通过多个步骤迭代地参与，具有模",
"2015; Wang et al., 2016, 2018; Huang et al.,",
"态内注意力。在SCANE(Leeetal.,2018）中，",
"2019b）利用三元组排名损失在联合嵌入空间",
"在EAndersonetal.（2018）的区域视觉特征与",
"中对齐英语句子和图像。在VSE+十（Faghri",
"文本语义之间进行模态间注意。推理时间复杂",
"etal.，2018）中，Faghrietal.通过强调硬阴",
"度为O(MN）（用于为大小为N的数据集生",
"性样本来改善VSE。VSE模型的最新进展探",
"成M查询表示）。与先前的工作相比，我们利",
"索了丰富英语图像语料库的方法。Shietal.",
"用了模态内多头注意力，与DAN相比，它可"
] | {
"arxiv_id": "1910.00058",
"language": "zh",
"page_idx": 0
} |
||
1910.00058-zh-1 | 1910.00058-zh-1.jpg | <FigureHere>
Figure 1: Multi-head attention with diversity for learning grounded multilingual multimodal representations. (A two-headed example with a part of diversity $\log i_\theta^D$ colored in red.)
以轻松并行化，并且与 SCAN 相比，具有首选
的$O(M)$推理时间复杂度
受$\overline{\mathrm{E}}$Li et al. (2018) 中注意力正则化思想的启发，为了学习 VSE, 我们提出了一种新的基于边距的多样性损失，以鼓励多个注意力集中的输出之间的边距。在我们的模型中，共同考虑了同一模态内和跨模态的多头注意力多样性
### 3 The Proposed Model
图 1 说明了所提出模型的概述。给定一组图像作为枢轴点，并带有相关的英语和德语$^{1}$ 描述或标题，建议的 VSE 模型旨在学习多语言多模式嵌人空间，在该空间中，配对实例$( i, x^e, x^g)$ 的 编 码 表 示 $( v, e, g)$ 彼 此 紧 密 对 齐 其 中 $\alpha$ 是 正 负 对 之 间 的 相 关 裕 度 , $[ . ] _+ =$ 编 码 器 : 对 于 采 样 的 $( i, x^e, x^g)$ 对 , 我 们 $\max ( 0, . )$ 是 铰 链 函 数 , $s( a, b) = \frac {a^Tb}{\| a\| \| b\| }$ 是 余 $\text{产 生 活 为 这 片 上 作 节 式 进 行 一 子 }_{e}$ 中 中 的 应 在 日 出 出 而 $\text{ 在 中 国 际 的 }_{e}$ $\text{ 时 间 }_{e}$ $\text{ 时 间 }_{e}$ $\text{ 时 间 }_{e}$ $\text{ 时 间 }_{e}$
首先通过单词嵌入矩阵对英语句子$x^e=$ $\{x_1^e,\ldots,x_N^e\}$中的标记和德语句子$x^g=$ $\{x_{1}^{\tilde{g}},\ldots,x_{N}^{\ddot{g}}\}$中的标记进行编码，然后是两个双向 LSTM。文本编码器的输出为英语的$e=\{e_1,\ldots,e_N\},e_n\in\mathbb{R}^H$和德语的$g=$ $\{ g_1, \ldots , g_N\} , g_n\in \mathbb{R} ^H$ ,其中 $N$ 是最大句子长度，$H$是共享嵌人空间的维度。对于图像，我们利用具有 ResNet$\underline{\mathrm{E}}($He et al.,2016) 主干的Faster-RCNN$\underline{\mathrm{E}}($Ren et al.,2015)网络来检测和编码图像中的重要视觉对象。使用可训练的单层感知器将视觉特征转换为共享嵌人空间， 我们将图像编码为$v=\{v_1,\ldots,v_M\},v_m\in\mathbb{R}^H$ ,其中$M$是图像中视觉对象的最大量
具有多样性的多头注意力：我们使用$K$头注意力网络来关注图像中的视觉对象以及句子中的文本语义，然后生成固定长度的图像/句子表示以进行对齐。具体来说，第$k$个出席的德
$^{1}$For clarity in notation, we discuss only two lan-
guages. The proposed model can be intuitively generalized to more languages by summing additional terms in equation 4 and equation 6-7 .
语句子表示$g^k$由下式计算：
$$a_i^k=\mathrm{softmax}(\mathrm{tanh}(W_{c_g}^kc_g^k)^\top\mathrm{tanh}(W_g^kg_i))$$
(1)
$$g^k=\sum_{i=1}^Na_i^kg_i,$$
(2)
其中$a_i^k$是注意力权重$,W_g^k\in\mathbb{R}^H\times H_{attn},W_{c_g}^k\in$ $\mathbb{R}^{H_c\times H_{attn}}\overset{g}{\operatorname*{\text{是德语的可学习转换矩阵。}}}c_g^k\in\mathbb{R}^{H_c}$ 是可学习的$H_c$维上下文向量，用于从德语句子中提取重要的语义。最后的德语句子表示是$K$的串联-头部注意输出$g=[g^0\|g^1\|\ldots\|g^K]$ 。类似于编码英语句子$e=[e^0\|e^1\|\ldots\|e^K]$和图像$v=[v^0\|v^1\|\ldots\|v^K]$
使用$\{V,E,G\}$ ,其中$v\in V,e\in E,g\in G$ 作为抽样批次中的受关注的固定长度图像和句子表示的集合，我们使用广泛使用的基于铰链的三元组排名损失与硬负挖掘$\overline{\mathrm{E}}($Faghri et al., 2018) 在视觉语义嵌人空间中对齐实例。以 Image-English 实例$\{V,E\}$为例，我们利用三元组相关损失定义为：
$$\begin{aligned}l_{\theta}(V,E)&=\sum_{p}\left[\alpha-s(v_{p},e_{p})+s(v_{p},\hat{e}_{p})\right]_{+}\\&+\sum_{q}\left[\alpha-s(v_{q},e_{q})+s(\hat{v}_{q},e_{q})\right]_{+},\end{aligned}$$
(3)
弦相似度。$p$和$q$是批处理中图像和句子的索引。$\hat{e}_p=$argmax$_qs(v_p,e_{q\neq p})$和$\hat{v}_q=$ $\operatorname{argmax}_ps(v_{p\neq q},e_q)$是硬底片。当三元组损失减少时，配对的图像和德语句子比最硬的非配对的图像和德语句子更接近边缘$\alpha$ 。我们的模型在联合嵌人空间中对齐$\{ V, E\}$ , $\{ V, G\}$和$\{ E, G\}$ , 以使用采样的$\left\{V,E,G\right\}$批次学习多语言多模态表示。我们将整个三重态损失表示为：
$$l_\theta(V,E,G)=l_\theta(V,G)+l_\theta(V,E)+\gamma l_\theta(G,E).$$
(4)
请注意，超参数$\gamma$控制$l_\theta(G,E)$ 的贡献，因为$( e$ , $g$ ) 可能不是转换对，即使$( e$ , $\{ V, E\}$ -MAX2)
多头注意力的理想属性之一是它能够共同关注并编码嵌人空间中的不同信息。但是，没有任何机制可以支持这些注意力确实可以捕获各种信息。为了鼓励$K$注意头在模态内和模态之间的多样性，我们提出了一种新的简单而有效的基于边距的多样性损失。例如，采样图 | [
[
"0.5156",
"0.0766",
"0.735",
"0.0766",
"0.735",
"0.0903",
"0.5156",
"0.0903"
],
[
"0.409",
"0.0944",
"0.4248",
"0.0944",
"0.4248",
"0.1051",
"0.409",
"0.1051"
],
[
"0.534",
"0.10095",
"0.8613",
"0.10095",
"0.8613",
"0.1217",
"0.534",
"0.1217"
],
[
"0.858",
"0.133",
"0.8833",
"0.133",
"0.8833",
"0.1478",
"0.858",
"0.1478"
],
[
"0.586",
"0.1538",
"0.601",
"0.1538",
"0.601",
"0.1633",
"0.586",
"0.1633"
],
[
"0.532",
"0.1633",
"0.6465",
"0.165",
"0.6465",
"0.1829",
"0.532",
"0.1812"
],
[
"0.8564",
"0.165",
"0.885",
"0.165",
"0.885",
"0.1841",
"0.8564",
"0.1841"
],
[
"0.58",
"0.1818",
"0.6055",
"0.1818",
"0.6055",
"0.1948",
"0.58",
"0.1948"
],
[
"0.1377",
"0.1675",
"0.1797",
"0.1675",
"0.1797",
"0.1776",
"0.1377",
"0.1776"
],
[
"0.2603",
"0.1912",
"0.2737",
"0.1912",
"0.2737",
"0.1948",
"0.2603",
"0.1948"
],
[
"0.4014",
"0.1895",
"0.44",
"0.1895",
"0.44",
"0.1965",
"0.4014",
"0.1965"
],
[
"0.4434",
"0.2079",
"0.4844",
"0.2079",
"0.4844",
"0.2156",
"0.4434",
"0.2156"
],
[
"0.511",
"0.2037",
"0.885",
"0.2019",
"0.885",
"0.2227",
"0.511",
"0.2245"
],
[
"0.1411",
"0.2114",
"0.1486",
"0.2144",
"0.1453",
"0.2173",
"0.1377",
"0.2156"
],
[
"0.51",
"0.2185",
"0.885",
"0.2203",
"0.885",
"0.2423",
"0.51",
"0.2405"
],
[
"0.2083",
"0.228",
"0.22",
"0.228",
"0.22",
"0.2322",
"0.2083",
"0.2322"
],
[
"0.367",
"0.228",
"0.3804",
"0.228",
"0.3804",
"0.2322",
"0.367",
"0.2322"
],
[
"0.403",
"0.228",
"0.4517",
"0.228",
"0.4517",
"0.2322",
"0.403",
"0.2322"
],
[
"0.461",
"0.2257",
"0.4902",
"0.2257",
"0.4902",
"0.2328",
"0.461",
"0.2328"
],
[
"0.1184",
"0.2452",
"0.4878",
"0.2452",
"0.4878",
"0.2588",
"0.1184",
"0.2588"
],
[
"0.5156",
"0.2405",
"0.882",
"0.2411",
"0.882",
"0.2559",
"0.5156",
"0.2546"
],
[
"0.12006",
"0.26",
"0.486",
"0.26",
"0.486",
"0.2737",
"0.12006",
"0.2737"
],
[
"0.517",
"0.2583",
"0.881",
"0.2583",
"0.881",
"0.2715",
"0.517",
"0.2715"
],
[
"0.12006",
"0.2737",
"0.4895",
"0.2737",
"0.4895",
"0.2874",
"0.12006",
"0.2874"
],
[
"0.513",
"0.2725",
"0.8774",
"0.2725",
"0.8774",
"0.2874",
"0.513",
"0.2874"
],
[
"0.1184",
"0.288",
"0.357",
"0.288",
"0.357",
"0.3027",
"0.1184",
"0.3027"
],
[
"0.5225",
"0.2893",
"0.8833",
"0.2893",
"0.8833",
"0.304",
"0.5225",
"0.304"
],
[
"0.5156",
"0.306",
"0.7153",
"0.306",
"0.7153",
"0.3196",
"0.5156",
"0.3196"
],
[
"0.532",
"0.3218",
"0.8833",
"0.3218",
"0.8833",
"0.335",
"0.532",
"0.335"
],
[
"0.12177",
"0.3289",
"0.486",
"0.3289",
"0.486",
"0.3433",
"0.12177",
"0.3433"
],
[
"0.5156",
"0.3384",
"0.8833",
"0.3384",
"0.8833",
"0.3528",
"0.5156",
"0.3528"
],
[
"0.12006",
"0.3457",
"0.3267",
"0.3457",
"0.3267",
"0.3594",
"0.12006",
"0.3594"
],
[
"0.517",
"0.3552",
"0.882",
"0.3552",
"0.882",
"0.3682",
"0.517",
"0.3682"
],
[
"0.1394",
"0.3623",
"0.486",
"0.3623",
"0.486",
"0.376",
"0.1394",
"0.376"
],
[
"0.5156",
"0.3706",
"0.8833",
"0.3706",
"0.8833",
"0.3843",
"0.5156",
"0.3843"
],
[
"0.12177",
"0.3777",
"0.4878",
"0.3777",
"0.4878",
"0.3914",
"0.12177",
"0.3914"
],
[
"0.5146",
"0.386",
"0.881",
"0.386",
"0.881",
"0.3997",
"0.5146",
"0.3997"
],
[
"0.1234",
"0.3943",
"0.486",
"0.3943",
"0.486",
"0.408",
"0.1234",
"0.408"
],
[
"0.5156",
"0.4026",
"0.882",
"0.4026",
"0.882",
"0.4163",
"0.5156",
"0.4163"
],
[
"0.12006",
"0.4104",
"0.486",
"0.4104",
"0.486",
"0.4233",
"0.12006",
"0.4233"
],
[
"0.519",
"0.4204",
"0.7036",
"0.4204",
"0.7036",
"0.4312",
"0.519",
"0.4312"
],
[
"0.12006",
"0.4287",
"0.4844",
"0.4287",
"0.4844",
"0.4395",
"0.12006",
"0.4395"
],
[
"0.531",
"0.4448",
"0.86",
"0.4453",
"0.86",
"0.4644",
"0.531",
"0.4631"
],
[
"0.1184",
"0.4548",
"0.339",
"0.4548",
"0.339",
"0.4697",
"0.1184",
"0.4697"
],
[
"0.6206",
"0.462",
"0.6465",
"0.4656",
"0.644",
"0.4785",
"0.6187",
"0.4768"
],
[
"0.1184",
"0.481",
"0.4844",
"0.481",
"0.4844",
"0.4946",
"0.1184",
"0.4946"
],
[
"0.601",
"0.4822",
"0.864",
"0.4822",
"0.864",
"0.4995",
"0.601",
"0.4995"
],
[
"0.1184",
"0.4976",
"0.486",
"0.4954",
"0.486",
"0.5103",
"0.1184",
"0.5127"
],
[
"0.623",
"0.503",
"0.638",
"0.503",
"0.638",
"0.514",
"0.623",
"0.514"
],
[
"0.12177",
"0.513",
"0.4878",
"0.513",
"0.4878",
"0.527",
"0.12177",
"0.527"
],
[
"0.858",
"0.514",
"0.882",
"0.514",
"0.882",
"0.5293",
"0.858",
"0.5293"
],
[
"0.1234",
"0.531",
"0.486",
"0.531",
"0.486",
"0.5415",
"0.1234",
"0.5415"
],
[
"0.12006",
"0.546",
"0.4727",
"0.546",
"0.4727",
"0.5586",
"0.12006",
"0.5586"
],
[
"0.5146",
"0.5405",
"0.882",
"0.5415",
"0.882",
"0.5557",
"0.5146",
"0.5547"
],
[
"0.51",
"0.557",
"0.8867",
"0.557",
"0.8867",
"0.5806",
"0.51",
"0.5806"
],
[
"0.8003",
"0.558",
"0.8228",
"0.558",
"0.8228",
"0.5654",
"0.8003",
"0.5654"
],
[
"0.12006",
"0.5767",
"0.4878",
"0.5767",
"0.4878",
"0.5903",
"0.12006",
"0.5903"
],
[
"0.12006",
"0.56",
"0.4878",
"0.561",
"0.4878",
"0.576",
"0.12006",
"0.5747"
],
[
"0.5156",
"0.5776",
"0.882",
"0.5776",
"0.882",
"0.593",
"0.5156",
"0.593"
],
[
"0.1184",
"0.593",
"0.461",
"0.5913",
"0.461",
"0.6094",
"0.1184",
"0.61"
],
[
"0.455",
"0.5957",
"0.486",
"0.5957",
"0.486",
"0.6025",
"0.455",
"0.6025"
],
[
"0.5156",
"0.594",
"0.86",
"0.594",
"0.86",
"0.6094",
"0.5156",
"0.6094"
],
[
"0.8555",
"0.5996",
"0.8774",
"0.5996",
"0.8774",
"0.604",
"0.8555",
"0.604"
],
[
"0.12177",
"0.61",
"0.4878",
"0.61",
"0.4878",
"0.6245",
"0.12177",
"0.6245"
],
[
"0.511",
"0.6094",
"0.8833",
"0.607",
"0.8833",
"0.6245",
"0.511",
"0.627"
],
[
"0.12006",
"0.6245",
"0.4878",
"0.626",
"0.4878",
"0.64",
"0.12006",
"0.6396"
],
[
"0.5156",
"0.627",
"0.879",
"0.627",
"0.879",
"0.637",
"0.5156",
"0.637"
],
[
"0.1184",
"0.64",
"0.4902",
"0.64",
"0.4902",
"0.6577",
"0.1184",
"0.6577"
],
[
"0.5156",
"0.6426",
"0.882",
"0.6426",
"0.882",
"0.6562",
"0.5156",
"0.6562"
],
[
"0.1184",
"0.6562",
"0.4878",
"0.655",
"0.4878",
"0.673",
"0.1184",
"0.6733"
],
[
"0.5156",
"0.6577",
"0.882",
"0.6577",
"0.882",
"0.6714",
"0.5156",
"0.6714"
],
[
"0.12177",
"0.675",
"0.4878",
"0.675",
"0.4878",
"0.6885",
"0.12177",
"0.6885"
],
[
"0.5156",
"0.675",
"0.882",
"0.675",
"0.882",
"0.6885",
"0.5156",
"0.6885"
],
[
"0.12177",
"0.6904",
"0.4878",
"0.6904",
"0.4878",
"0.7036",
"0.12177",
"0.7036"
],
[
"0.5146",
"0.6904",
"0.882",
"0.6904",
"0.882",
"0.7036",
"0.5146",
"0.7036"
],
[
"0.12006",
"0.707",
"0.486",
"0.707",
"0.486",
"0.72",
"0.12006",
"0.72"
],
[
"0.5156",
"0.708",
"0.5415",
"0.708",
"0.5415",
"0.7197",
"0.5156",
"0.7197"
],
[
"0.1234",
"0.7236",
"0.4844",
"0.7236",
"0.4844",
"0.734",
"0.1234",
"0.734"
],
[
"0.12006",
"0.7383",
"0.4836",
"0.7383",
"0.4836",
"0.752",
"0.12006",
"0.752"
],
[
"0.5205",
"0.7354",
"0.875",
"0.7354",
"0.875",
"0.749",
"0.5205",
"0.749"
],
[
"0.11755",
"0.753",
"0.4878",
"0.75",
"0.4878",
"0.771",
"0.11755",
"0.774"
],
[
"0.858",
"0.75",
"0.881",
"0.75",
"0.881",
"0.763",
"0.858",
"0.763"
],
[
"0.5146",
"0.7656",
"0.881",
"0.7656",
"0.881",
"0.7793",
"0.5146",
"0.7793"
],
[
"0.1301",
"0.7715",
"0.4282",
"0.771",
"0.4282",
"0.7837",
"0.1301",
"0.785"
],
[
"0.1276",
"0.778",
"0.1368",
"0.778",
"0.1368",
"0.782",
"0.1276",
"0.782"
],
[
"0.12177",
"0.7876",
"0.4895",
"0.7876",
"0.4895",
"0.8003",
"0.12177",
"0.8003"
],
[
"0.513",
"0.783",
"0.881",
"0.783",
"0.881",
"0.7964",
"0.513",
"0.7964"
],
[
"0.12006",
"0.804",
"0.486",
"0.802",
"0.486",
"0.8154",
"0.12006",
"0.818"
],
[
"0.5146",
"0.799",
"0.581",
"0.799",
"0.581",
"0.81",
"0.5146",
"0.81"
],
[
"0.12006",
"0.8193",
"0.486",
"0.8193",
"0.486",
"0.833",
"0.12006",
"0.833"
],
[
"0.534",
"0.814",
"0.882",
"0.814",
"0.882",
"0.8276",
"0.534",
"0.8276"
],
[
"0.12006",
"0.836",
"0.4878",
"0.836",
"0.4878",
"0.8496",
"0.12006",
"0.8496"
],
[
"0.5146",
"0.8306",
"0.881",
"0.8306",
"0.881",
"0.844",
"0.5146",
"0.844"
],
[
"0.5156",
"0.847",
"0.882",
"0.847",
"0.882",
"0.861",
"0.5156",
"0.861"
],
[
"0.1411",
"0.86",
"0.4878",
"0.86",
"0.4878",
"0.8735",
"0.1411",
"0.8735"
],
[
"0.5146",
"0.8633",
"0.881",
"0.8633",
"0.881",
"0.877",
"0.5146",
"0.877"
],
[
"0.1184",
"0.8716",
"0.486",
"0.8706",
"0.486",
"0.8853",
"0.1184",
"0.8857"
],
[
"0.12006",
"0.8843",
"0.4895",
"0.8843",
"0.4895",
"0.898",
"0.12006",
"0.898"
],
[
"0.5156",
"0.879",
"0.882",
"0.879",
"0.882",
"0.8926",
"0.5156",
"0.8926"
],
[
"0.12006",
"0.898",
"0.357",
"0.898",
"0.357",
"0.9077",
"0.12006",
"0.9077"
],
[
"0.517",
"0.8955",
"0.882",
"0.8955",
"0.882",
"0.909",
"0.517",
"0.909"
]
] | [
"语句子表示gk由下式计算：",
"",
"ah = softmax(tanh(W,cg) Ttanh(Wggi))",
"(1)",
"N",
"gh =ah gi,",
"(2)",
"i=1",
"F-RCNN",
"He",
"Attentio",
"",
"其中a 是注意力权重，WhRHxHattn,Wk",
"",
"RHexHatn 是德语的可学习转换矩阵。ch e RHe",
"MC",
"ack",
"rhases",
"brow",
"Figure 1: Multi-head attention with diversity for",
"是可学习的H。维上下文向量，用于从德语句",
"learning grounded multilingual multimodal repre-",
"子中提取重要的语义。最后的德语句子表示是",
"sentations. (A two-headed example with a part of",
"K的串联-头部注意输出g=[golgl..llgK",
"diversity loss tD colored in red.)",
"类似于编码英语句子e=[eolelleK]和",
"图像=[ol..]",
"使用IVE,GI.其中 uEVeEE,qEG",
"以轻松并行化，并且与SCAN相比，具有首选",
"作为抽样批次中的受关注的固定长度图像和",
"的O(M)推理时间复杂度",
"句子表示的集合，我们使用广泛使用的基于",
"受ELietal.（2018）中注意力正则化思想的",
"铰链的三元组排名损失与硬负挖掘E(Faghri",
"启发，为了学习VSE，我们提出了一种新的基",
"etal.，2018）在视觉语义嵌入空间中对齐实例。",
"于边距的多样性损失，以鼓励多个注意力集中",
"以Image-English实例V,E}为例，我们利用",
"的输出之间的边距。在我们的模型中，共同考",
"三元组相关损失定义为",
"虑了同一模态内和跨模态的多头注意力多样性",
"le(V, E) = [α - s(up,ep) + s(up,ep)] +",
"3 The Proposed Model",
"P",
"图1说明了所提出模型的概述。给定一组图",
"+[α - s(Ug,ea) + s(0g,ea)]+\"",
"像作为枢轴点，并带有相关的英语和德语1",
"q",
"描述或标题，建议的VSE模型旨在学习多语",
"(3)",
"言多模式嵌入空间，在该空间中，配对实例",
"(i,re,a9）的编码表示（u,e,9）彼此紧密对齐",
"其中α是正负对之间的相关裕度，[]+=",
"max(0,）是铰链函数，s(a,b)=是余",
"alt",
"首先通过单词嵌入矩阵对英语句子e=",
"编码器：对于采样的（i,ae,a9）对，我们",
"弦相似度。p和q是批处理中图像和句子",
"[ci....,a%】中的标记和德语句子a9",
"",
"的索引。ép = argmaxqs(Up,eq+p）和 g",
"",
"[ci，..α】中的标记进行编码，然后是两",
"argmaxps(uptq,ea）是硬底片。当三元组损失",
"个双向LSTM。文本编码器的输出为英语",
"减少时，配对的图像和德语句子比最硬的非配",
"的 e ={ei,,en],en RH 和德语的 g =",
"对的图像和德语句子更接近边缘α。我们的模",
"{91,.，9},9nERH，其中N是最大句子长",
"型在联合嵌入空间中对齐{V,E}，{V,G}和",
"度，H是共享嵌入空间的维度。对于图像，我",
"[E,G]，以使用采样的V,E,G}批次学习多",
"们利用具有ResNetE(Heetal.，2016）主于的",
"语言多模态表示。我们将整个三重态损失表示",
"Faster-RCNNE(Renetal.,2015）网络来检测",
"为:",
"和编码图像中的重要视觉对象。使用可训练的",
"单层感知器将视觉特征转换为共享嵌入空间",
"le(V, E,G) = le(V,G) + le(V,E) + le(G, E)",
"我们将图像编码为={u1...,UM),UmERH",
"(4)",
"请注意，超参数控制le(G,E)的贡献，因为",
"其中M是图像中视觉对象的最大量",
"",
"具有多样性的多头注意力：我们使用K头注",
"(e，g）可能不是转换对，即使（e，{V,E)",
"意力网络来关注图像中的视觉对象以及句子中",
"-MAX2",
"的文本语义，然后生成固定长度的图像/句子",
"多头注意力的理想属性之一是它能够共同关",
"表示以进行对齐。具体来说，第k个出席的德",
"注并编码嵌人空间中的不同信息。但是，没有",
"任何机制可以支持这些注意力确实可以捕获",
"'For clarity in notation, we discuss only two lan-",
"各种信息。为了鼓励K注意头在模态内和模",
"guages. The proposed model can be intuitively gener",
"alized to more languages by summing additional terms",
"态之间的多样性，我们提出了一种新的简单而",
"in equation 4and equation 6-7",
"有效的基于边距的多样性损失。例如，采样图"
] | {
"arxiv_id": "1910.00058",
"language": "zh",
"page_idx": 1
} |
||
1910.00058-zh-2 | 1910.00058-zh-2.jpg | <table>
<tbody>
<tr>
<th rowspan="2">$\mathop{\mathrm{Method}}$</th>
<th colspan="2">German to Ima</th>
<th>mage</th>
<th colspan="3">${\mathrm{Image~to~German}}$</th>
<th>$\overline{\mathrm{Eng}}$</th>
<th rowspan="1">English to Image</th>
<th>$\overline{\mathrm{Image}}$</th>
<th colspan="3">$\operatorname{Image}$ ;to I English</th>
</tr>
<tr>
<th>$R@1$</th>
<th>$R@5$</th>
<th>$R10$</th>
<th>$R@1$</th>
<th>$R@5$</th>
<th>$R@10$</th>
<th>$R@1$</th>
<th>$R@5$</th>
<th>$R@10$</th>
<th>$R@1$</th>
<th>$R@5$</th>
<th>$R10$</th>
</tr>
<tr>
<td>VSE 广 (Kiros et al., 2014)</td>
<td>20.3</td>
<td>47.2</td>
<td>60.1</td>
<td>29.3</td>
<td>58.1</td>
<td>71.8</td>
<td>23.3</td>
<td>53.6</td>
<td>65.8</td>
<td>31.6</td>
<td>60.4</td>
<td>72.7</td>
</tr>
<tr>
<td>$OE$ 十 冰 (Vendrov et al., 2015</td>
<td>21.0</td>
<td>48.5</td>
<td>60.4</td>
<td>26.8</td>
<td>57.5</td>
<td>70.9</td>
<td>25.8</td>
<td>56.5</td>
<td>67.8</td>
<td>34.8</td>
<td>63.7</td>
<td>74.8</td>
</tr>
<tr>
<td>$DAN$ (Nam et al.. 2017</td>
<td>31.0</td>
<td>60.9</td>
<td>71.0</td>
<td>46.5</td>
<td>77.5</td>
<td>83.0</td>
<td>39.4</td>
<td>69.2</td>
<td>69.1</td>
<td>55.0</td>
<td>81.8</td>
<td>89.0</td>
</tr>
<tr>
<td>VSE+ + 大 (Faghri et al., 2018</td>
<td>31.3</td>
<td>62.2</td>
<td>70.9</td>
<td>47.5</td>
<td>78.5</td>
<td>84.5</td>
<td>39.6</td>
<td>69.1</td>
<td>79.8</td>
<td>53.1</td>
<td>82.1</td>
<td>87.5</td>
</tr>
<tr>
<td>SCAN (Lee et al., 2018</td>
<td>35.7</td>
<td>64.9</td>
<td>74.6</td>
<td>52.3</td>
<td>81.8</td>
<td>88.5</td>
<td>45.8</td>
<td>74.4</td>
<td>83.0</td>
<td>61.8</td>
<td>87.5</td>
<td>93.7</td>
</tr>
<tr>
<td>Pivot $^{\prime }$Gella et al. 2017</td>
<td>22.5</td>
<td>49.3</td>
<td>61.7</td>
<td>28.2</td>
<td>61.9</td>
<td>73.4</td>
<td>26.2</td>
<td>56.4</td>
<td>68.4</td>
<td>33.8</td>
<td>62.8</td>
<td>75.2</td>
</tr>
<tr>
<td>Ours (Random, VGG19)</td>
<td>25.8</td>
<td>54.9</td>
<td>65.1</td>
<td>34.1</td>
<td>65.5</td>
<td>76.5</td>
<td>30.1</td>
<td>62.5</td>
<td>71.6</td>
<td>36.4</td>
<td>68.0</td>
<td>80.9</td>
</tr>
<tr>
<td>(Random, No diversity) Ours</td>
<td>36.3</td>
<td>65.3</td>
<td>74.7</td>
<td>53.1</td>
<td>82.3</td>
<td>88.8</td>
<td>46.2</td>
<td>74.7</td>
<td>82.9</td>
<td>63.3</td>
<td>87.0</td>
<td>93.3</td>
</tr>
<tr>
<td>Ours (Random</td>
<td>39.2</td>
<td>67.5</td>
<td>76.7</td>
<td>55.0</td>
<td>84.7</td>
<td>91.2</td>
<td>48.7</td>
<td>77.2</td>
<td>85.0</td>
<td>66.4</td>
<td>88.3</td>
<td>93.4</td>
</tr>
<tr>
<td>Ours FastText $W$,</td>
<td>40.3</td>
<td>70.1</td>
<td>79.0</td>
<td>60.4</td>
<td>85.4</td>
<td>92.0</td>
<td>50.1</td>
<td>78.1</td>
<td>85.7</td>
<td>68.0</td>
<td>88.8</td>
<td>94.0</td>
</tr>
<tr>
<td>Ours BERT $W$</td>
<td>40.7</td>
<td>70.5</td>
<td>78.8</td>
<td>56.5</td>
<td>84.6</td>
<td>91.3</td>
<td>48.9</td>
<td>78.3</td>
<td>85.8</td>
<td>66.5</td>
<td>89.1</td>
<td>94.1</td>
</tr>
</tbody>
</table>
Table 1: Comparison of multilingual sentence-image retrieval/matching (German-Image) and (EnglishImage) results on Multi30K. (Visual encoders:VGG $^\dagger$ otherwise ResNet or Faster-RCNN(ResNet).) $(\bar{\mathrm{Monolingual~models~}^*}.)$
(7)
{}其中 $\beta$ 是加权参数，用于平衡多样性损失和三
最终模型损失形式化为：
$l_{\theta}^{All}(V,E,G)=l_{\theta}(V,E,G)+\beta l_{\theta}^{D}(V,E,G)$,
重排名损失。我们通过最小化$l_{\theta}^{All}(V,E,G)$来
训练模型
<table>
<tbody>
<tr>
<th>Method</th>
<th>MS- Vid (2012) T</th>
<th>$\overline{\mathrm{Pascal}}$ 2014)</th>
<th>$\overline{\mathrm{Pascal}}$ 2015</th>
</tr>
<tr>
<td>$\overline{STS}$ ${\mathrm{S~Baseline}}$</td>
<td>29.9</td>
<td>51.3</td>
<td>60.4</td>
</tr>
<tr>
<td>$STS$ Best System 1</td>
<td>86.3</td>
<td>83.4</td>
<td>86.4</td>
</tr>
<tr>
<td>GRAN (Wieting et al.,2 2017</td>
<td>83.7</td>
<td>84.5</td>
<td>85.0</td>
</tr>
<tr>
<td>${\overline{VSE}}$ Kiros et al. , 2014</td>
<td>$\overline{80.6}$</td>
<td>$\overline{82.7}$</td>
<td>$\overline{89.6}$</td>
</tr>
<tr>
<td>$OE$ (Vendrov et al.,2015</td>
<td>82.2</td>
<td>84.1</td>
<td>90.8</td>
</tr>
<tr>
<td>$DAN$ (Nam et al., 2017</td>
<td>84.1</td>
<td>84.3</td>
<td>90.8</td>
</tr>
<tr>
<td>VSE+ + Faghri et al. , 2018</td>
<td>84.5</td>
<td>84.8</td>
<td>91.2</td>
</tr>
<tr>
<td>SCAN Lee et al. , 2018</td>
<td>84.0</td>
<td>83.9</td>
<td>90.7</td>
</tr>
<tr>
<td>PIVOT $^{\prime }$Gella et al. , 2017) </td>
<td>84.6</td>
<td>84.5</td>
<td>91.5</td>
</tr>
<tr>
<td>$\overline{Ours}$ $W$ Random</td>
<td>85.8</td>
<td>87.8</td>
<td>91.5</td>
</tr>
<tr>
<td>Ours FastText $W$</td>
<td>86.2</td>
<td>88.3</td>
<td>91.8</td>
</tr>
<tr>
<td>Ours BERT) $W$</td>
<td>86.4</td>
<td>88.0</td>
<td>91.7</td>
</tr>
</tbody>
</table>
Experiments
继$\operatorname{EGella}$ et al.(2017) 之后，我们对Multi30K$\underline{\mathrm{E}}($Elliott et al., 2016) 中的多语言句子图像匹配任务和语义文本相似性任务$\mathbb{E}$ (Agirre et al., 2012, 2015) 进行了评估
Table 2: Results on the image and video datasets of Semantic Textual Similarity task. (Pearson'â$\mathrm{€}^\mathrm{TM}$s $r\times 100$ ).
像和英语句子之间的多头注意力多样性损失
( i. e. $\overline {\mathrm{E} }$跨模态多样性)定义为：
### 4.1 Experiment Setup
$$l_{\theta}^{D}(V,E)=\sum_{p}\sum_{k}\sum_{r}\:[\alpha_{D}-s(v_{p}^{k},e_{p}^{k\neq r})]_{+}$$
(5)
如图 1 中用于更新的红色箭头所示，此多
样性目标背后的优点是增加了实例本身或其跨模态并行实例的来自不同关注头的受关注嵌人之间的距离(最大为多样性余量$\alpha_D$ )。结果， 多样性目标明确鼓励多头注意力集中在稀疏位于联合嵌入空间中的信息的不同方面，以促进多语言文本语义与视觉对象之间的细粒度对齐。由于共享嵌入空间是多语言和多模式的， 因此为了改善模式内/语言和模式间/语言多样性，我们将整体多样性损失建模为：
我们在$\overline{\mathrm{E}}$Anderson et al.(2018)中使用该模型，该模型是在 MS-COCO$\underline{\mathrm{E}}($Lin et al.,2014) 数据集上预先训练并在 VisualGenome$\underline{\mathrm{E}}($Krishna et al., 2016) 数据集上进行微调的 FasterRCNN$\underline{\mathrm{E}}($Ren et al.,2015)网络，以检测显着视觉对象并提取其相应特征。1,600 种类型的物体是可检测的。然后，我们将每个图像打包并表示为 36$\times2048$特征矩阵，其中 36 是图像中突出视觉对象的最大数量，2048 是 FasterRCNN 的 ResNet$\boxed{\mathrm{F}}($He et al., 2016) 主干中平坦的最后一个池化层的尺寸
对于文本处理，我们小写，标记化，然后将最大句子长度截断为 100。我们使用随机或预先训练的多语言嵌人初始化的 300-dim 单词嵌人矩阵。说明：我们使用多语言版本的FastText (Mikolov et al.,2018)。我们还尝试合并最后一层上下文化多语言 BERT 嵌人$\mathbb{E}$ (Devlin et al., 2018) , 以替换单词嵌入矩阵作为双向 LSTM 的文本输入功能
对于训练，我们对大小为 128 的批次进行采样，并在 Multi30K 的训练集上训练 20 个epoch。我们使用具有$2\times10^{-4}$学习速率的
$$\begin{aligned}l_{\theta}^{D}(V,E,G)&=l_{\theta}^{D}(V,V)+l_{\theta}^{D}(G,G)+l_{\theta}^{D}(E,E)\\&+l_{\theta}^{D}(V,E)+l_{\theta}^{D}(V,G)+l_{\theta}^{D}(G,E),\end{aligned}$$
其中前三个术语是模式内/语言，其余是跨模式/语言。使用公式$\mathbb{E}$4 和公式$\mathbb{E}$6,我们将 | [
[
"0.12006",
"0.0707",
"0.1838",
"0.07245",
"0.1823",
"0.0873",
"0.1184",
"0.0849"
],
[
"0.3435",
"0.07245",
"0.465",
"0.0736",
"0.465",
"0.0849",
"0.3435",
"0.08435"
],
[
"0.4836",
"0.07245",
"0.608",
"0.07245",
"0.608",
"0.0861",
"0.4836",
"0.0861"
],
[
"0.6323",
"0.07184",
"0.754",
"0.07184",
"0.754",
"0.0849",
"0.6323",
"0.0849"
],
[
"0.7715",
"0.07245",
"0.895",
"0.07245",
"0.895",
"0.0861",
"0.7715",
"0.0861"
],
[
"0.3418",
"0.08435",
"0.4685",
"0.08435",
"0.4685",
"0.0974",
"0.3418",
"0.0974"
],
[
"0.4785",
"0.08435",
"0.517",
"0.08435",
"0.517",
"0.0968",
"0.4785",
"0.0968"
],
[
"0.523",
"0.0849",
"0.5684",
"0.0849",
"0.5684",
"0.0968",
"0.523",
"0.0968"
],
[
"0.565",
"0.0849",
"0.614",
"0.0849",
"0.614",
"0.0968",
"0.565",
"0.0968"
],
[
"0.6245",
"0.0849",
"0.6665",
"0.0849",
"0.6665",
"0.0968",
"0.6245",
"0.0968"
],
[
"0.6665",
"0.0849",
"0.711",
"0.0849",
"0.711",
"0.0968",
"0.6665",
"0.0968"
],
[
"0.714",
"0.0849",
"0.7617",
"0.0849",
"0.7617",
"0.0968",
"0.714",
"0.0968"
],
[
"0.771",
"0.0849",
"0.812",
"0.0849",
"0.812",
"0.0968",
"0.771",
"0.0968"
],
[
"0.817",
"0.0849",
"0.8965",
"0.0849",
"0.8965",
"0.0968",
"0.817",
"0.0968"
],
[
"0.12177",
"0.0986",
"0.3167",
"0.0998",
"0.3167",
"0.114",
"0.12177",
"0.1134"
],
[
"0.3914",
"0.10095",
"0.4265",
"0.10095",
"0.4265",
"0.1134",
"0.3914",
"0.1134"
],
[
"0.4358",
"0.10095",
"0.4702",
"0.10095",
"0.4702",
"0.1134",
"0.4358",
"0.1134"
],
[
"0.4785",
"0.10095",
"0.513",
"0.10095",
"0.513",
"0.1134",
"0.4785",
"0.1134"
],
[
"0.525",
"0.10095",
"0.559",
"0.10095",
"0.559",
"0.1134",
"0.525",
"0.1134"
],
[
"0.575",
"0.10095",
"0.6113",
"0.10095",
"0.6113",
"0.1134",
"0.575",
"0.1134"
],
[
"0.6245",
"0.10095",
"0.66",
"0.10095",
"0.66",
"0.1134",
"0.6245",
"0.1134"
],
[
"0.671",
"0.10095",
"0.708",
"0.10095",
"0.708",
"0.1134",
"0.671",
"0.1134"
],
[
"0.721",
"0.0986",
"0.756",
"0.10095",
"0.754",
"0.114",
"0.7197",
"0.11224"
],
[
"0.7715",
"0.10095",
"0.8096",
"0.10095",
"0.8096",
"0.1134",
"0.7715",
"0.1134"
],
[
"0.8623",
"0.10095",
"0.8965",
"0.10095",
"0.8965",
"0.1134",
"0.8623",
"0.1134"
],
[
"0.818",
"0.10156",
"0.854",
"0.10156",
"0.854",
"0.1134",
"0.818",
"0.1134"
],
[
"0.3435",
"0.10156",
"0.3804",
"0.10156",
"0.3804",
"0.114",
"0.3435",
"0.114"
],
[
"0.12427",
"0.11224",
"0.3284",
"0.114",
"0.3284",
"0.1277",
"0.12427",
"0.1259"
],
[
"0.525",
"0.1134",
"0.559",
"0.1152",
"0.5576",
"0.1277",
"0.523",
"0.1259"
],
[
"0.575",
"0.1134",
"0.6094",
"0.1152",
"0.608",
"0.1277",
"0.574",
"0.1259"
],
[
"0.669",
"0.1134",
"0.706",
"0.1152",
"0.705",
"0.1277",
"0.6675",
"0.1259"
],
[
"0.721",
"0.1134",
"0.756",
"0.1152",
"0.754",
"0.1289",
"0.7197",
"0.1271"
],
[
"0.8623",
"0.1134",
"0.8965",
"0.1152",
"0.895",
"0.1277",
"0.8613",
"0.1259"
],
[
"0.4785",
"0.114",
"0.513",
"0.114",
"0.513",
"0.1271",
"0.4785",
"0.1271"
],
[
"0.4358",
"0.114",
"0.471",
"0.114",
"0.471",
"0.1271",
"0.4358",
"0.1271"
],
[
"0.7715",
"0.114",
"0.8076",
"0.114",
"0.8076",
"0.1271",
"0.7715",
"0.1271"
],
[
"0.818",
"0.114",
"0.852",
"0.114",
"0.852",
"0.1271",
"0.818",
"0.1271"
],
[
"0.3896",
"0.1152",
"0.4248",
"0.1152",
"0.4248",
"0.1277",
"0.3896",
"0.1277"
],
[
"0.3435",
"0.1152",
"0.3804",
"0.1152",
"0.3804",
"0.1271",
"0.3435",
"0.1271"
],
[
"0.6265",
"0.1152",
"0.66",
"0.1152",
"0.66",
"0.1277",
"0.6265",
"0.1277"
],
[
"0.1234",
"0.1259",
"0.3005",
"0.1271",
"0.3005",
"0.1414",
"0.1234",
"0.1401"
],
[
"0.3896",
"0.1271",
"0.4248",
"0.1289",
"0.424",
"0.1414",
"0.388",
"0.1395"
],
[
"0.4802",
"0.1271",
"0.513",
"0.1289",
"0.511",
"0.1414",
"0.4785",
"0.1395"
],
[
"0.5264",
"0.1271",
"0.559",
"0.1289",
"0.5576",
"0.1414",
"0.525",
"0.1395"
],
[
"0.4358",
"0.1277",
"0.4702",
"0.1277",
"0.4702",
"0.1401",
"0.4358",
"0.1401"
],
[
"0.3435",
"0.1277",
"0.3804",
"0.1277",
"0.3804",
"0.1401",
"0.3435",
"0.1401"
],
[
"0.575",
"0.1277",
"0.6094",
"0.1277",
"0.6094",
"0.1401",
"0.575",
"0.1401"
],
[
"0.6245",
"0.1277",
"0.66",
"0.1277",
"0.66",
"0.1401",
"0.6245",
"0.1401"
],
[
"0.671",
"0.1277",
"0.706",
"0.1277",
"0.706",
"0.1401",
"0.671",
"0.1401"
],
[
"0.721",
"0.1277",
"0.756",
"0.1277",
"0.756",
"0.1401",
"0.721",
"0.1401"
],
[
"0.7715",
"0.1289",
"0.806",
"0.1289",
"0.806",
"0.1401",
"0.7715",
"0.1401"
],
[
"0.818",
"0.1277",
"0.854",
"0.1277",
"0.854",
"0.1401",
"0.818",
"0.1401"
],
[
"0.8623",
"0.1289",
"0.8965",
"0.1289",
"0.8965",
"0.1414",
"0.8623",
"0.1414"
],
[
"0.126",
"0.1401",
"0.3333",
"0.1401",
"0.3333",
"0.1538",
"0.126",
"0.1538"
],
[
"0.3435",
"0.1414",
"0.3787",
"0.1414",
"0.3787",
"0.1538",
"0.3435",
"0.1538"
],
[
"0.3896",
"0.1414",
"0.4248",
"0.1414",
"0.4248",
"0.1538",
"0.3896",
"0.1538"
],
[
"0.4358",
"0.1414",
"0.4702",
"0.1414",
"0.4702",
"0.1538",
"0.4358",
"0.1538"
],
[
"0.4785",
"0.1414",
"0.513",
"0.1414",
"0.513",
"0.1538",
"0.4785",
"0.1538"
],
[
"0.5264",
"0.1414",
"0.559",
"0.1414",
"0.559",
"0.1538",
"0.5264",
"0.1538"
],
[
"0.575",
"0.1414",
"0.608",
"0.1414",
"0.608",
"0.1538",
"0.575",
"0.1538"
],
[
"0.6245",
"0.1401",
"0.66",
"0.1426",
"0.659",
"0.1538",
"0.623",
"0.152"
],
[
"0.671",
"0.1414",
"0.705",
"0.1414",
"0.705",
"0.1538",
"0.671",
"0.1538"
],
[
"0.723",
"0.1414",
"0.756",
"0.1414",
"0.756",
"0.1538",
"0.723",
"0.1538"
],
[
"0.7715",
"0.1414",
"0.806",
"0.1414",
"0.806",
"0.1538",
"0.7715",
"0.1538"
],
[
"0.818",
"0.1414",
"0.8506",
"0.1414",
"0.8506",
"0.1538",
"0.818",
"0.1538"
],
[
"0.8623",
"0.1414",
"0.8965",
"0.1414",
"0.8965",
"0.1538",
"0.8623",
"0.1538"
],
[
"0.12427",
"0.155",
"0.297",
"0.155",
"0.297",
"0.165",
"0.12427",
"0.165"
],
[
"0.3435",
"0.152",
"0.382",
"0.155",
"0.3787",
"0.1692",
"0.34",
"0.1663"
],
[
"0.3914",
"0.155",
"0.4248",
"0.155",
"0.4248",
"0.1675",
"0.3914",
"0.1675"
],
[
"0.4358",
"0.155",
"0.4702",
"0.155",
"0.4702",
"0.1675",
"0.4358",
"0.1675"
],
[
"0.4785",
"0.155",
"0.513",
"0.155",
"0.513",
"0.1675",
"0.4785",
"0.1675"
],
[
"0.525",
"0.155",
"0.559",
"0.155",
"0.559",
"0.1675",
"0.525",
"0.1675"
],
[
"0.575",
"0.155",
"0.6094",
"0.155",
"0.6094",
"0.1675",
"0.575",
"0.1675"
],
[
"0.6265",
"0.155",
"0.66",
"0.155",
"0.66",
"0.1675",
"0.6265",
"0.1675"
],
[
"0.6724",
"0.155",
"0.706",
"0.155",
"0.706",
"0.1675",
"0.6724",
"0.1675"
],
[
"0.723",
"0.155",
"0.756",
"0.155",
"0.756",
"0.1675",
"0.723",
"0.1675"
],
[
"0.7715",
"0.155",
"0.8076",
"0.155",
"0.8076",
"0.1675",
"0.7715",
"0.1675"
],
[
"0.8193",
"0.155",
"0.854",
"0.155",
"0.854",
"0.1675",
"0.8193",
"0.1675"
],
[
"0.8623",
"0.1561",
"0.8965",
"0.1561",
"0.8965",
"0.1687",
"0.8623",
"0.1687"
],
[
"0.12006",
"0.1663",
"0.3123",
"0.1687",
"0.3123",
"0.1818",
"0.12006",
"0.1799"
],
[
"0.392",
"0.1687",
"0.4265",
"0.1687",
"0.4265",
"0.1812",
"0.392",
"0.1812"
],
[
"0.525",
"0.1687",
"0.559",
"0.1687",
"0.559",
"0.1812",
"0.525",
"0.1812"
],
[
"0.5767",
"0.1687",
"0.608",
"0.1687",
"0.608",
"0.1812",
"0.5767",
"0.1812"
],
[
"0.8623",
"0.1675",
"0.8984",
"0.1692",
"0.8965",
"0.1818",
"0.8613",
"0.1799"
],
[
"0.4358",
"0.1692",
"0.4702",
"0.1692",
"0.4702",
"0.1812",
"0.4358",
"0.1812"
],
[
"0.345",
"0.1692",
"0.3787",
"0.1692",
"0.3787",
"0.1812",
"0.345",
"0.1812"
],
[
"0.4785",
"0.1692",
"0.513",
"0.1692",
"0.513",
"0.1812",
"0.4785",
"0.1812"
],
[
"0.6274",
"0.1692",
"0.66",
"0.1692",
"0.66",
"0.1812",
"0.6274",
"0.1812"
],
[
"0.671",
"0.1692",
"0.706",
"0.1692",
"0.706",
"0.1812",
"0.671",
"0.1812"
],
[
"0.723",
"0.1692",
"0.756",
"0.1692",
"0.756",
"0.1812",
"0.723",
"0.1812"
],
[
"0.7734",
"0.1692",
"0.8076",
"0.1692",
"0.8076",
"0.1812",
"0.7734",
"0.1812"
],
[
"0.8193",
"0.1692",
"0.854",
"0.1692",
"0.854",
"0.1812",
"0.8193",
"0.1812"
],
[
"0.1184",
"0.1799",
"0.309",
"0.1799",
"0.309",
"0.1936",
"0.1184",
"0.1936"
],
[
"0.3435",
"0.1818",
"0.3787",
"0.1818",
"0.3787",
"0.1936",
"0.3435",
"0.1936"
],
[
"0.3914",
"0.1818",
"0.4248",
"0.1818",
"0.4248",
"0.1936",
"0.3914",
"0.1936"
],
[
"0.4358",
"0.1818",
"0.4702",
"0.1818",
"0.4702",
"0.1936",
"0.4358",
"0.1936"
],
[
"0.4785",
"0.1818",
"0.511",
"0.1818",
"0.511",
"0.1936",
"0.4785",
"0.1936"
],
[
"0.525",
"0.1818",
"0.5576",
"0.1818",
"0.5576",
"0.1948",
"0.525",
"0.1948"
],
[
"0.5767",
"0.1818",
"0.6094",
"0.1818",
"0.6094",
"0.1936",
"0.5767",
"0.1936"
],
[
"0.6265",
"0.1818",
"0.659",
"0.1818",
"0.659",
"0.1936",
"0.6265",
"0.1936"
],
[
"0.671",
"0.1818",
"0.705",
"0.1818",
"0.705",
"0.1936",
"0.671",
"0.1936"
],
[
"0.723",
"0.1818",
"0.756",
"0.1818",
"0.756",
"0.1936",
"0.723",
"0.1936"
],
[
"0.7715",
"0.1818",
"0.8076",
"0.1818",
"0.8076",
"0.1936",
"0.7715",
"0.1936"
],
[
"0.818",
"0.1818",
"0.854",
"0.1818",
"0.854",
"0.1936",
"0.818",
"0.1936"
],
[
"0.8613",
"0.1818",
"0.8984",
"0.1818",
"0.8984",
"0.1936",
"0.8613",
"0.1936"
],
[
"0.1234",
"0.1912",
"0.3284",
"0.1936",
"0.3284",
"0.2079",
"0.1234",
"0.206"
],
[
"0.3435",
"0.1924",
"0.3787",
"0.1948",
"0.378",
"0.2073",
"0.3418",
"0.2048"
],
[
"0.3896",
"0.1924",
"0.4248",
"0.1948",
"0.424",
"0.206",
"0.388",
"0.2037"
],
[
"0.6724",
"0.1912",
"0.705",
"0.1936",
"0.7036",
"0.206",
"0.671",
"0.2037"
],
[
"0.7715",
"0.1924",
"0.8076",
"0.1948",
"0.806",
"0.2073",
"0.771",
"0.2048"
],
[
"0.4785",
"0.1936",
"0.511",
"0.1936",
"0.511",
"0.206",
"0.4785",
"0.206"
],
[
"0.4358",
"0.1936",
"0.4685",
"0.1936",
"0.4685",
"0.206",
"0.4358",
"0.206"
],
[
"0.525",
"0.1936",
"0.559",
"0.1936",
"0.559",
"0.206",
"0.525",
"0.206"
],
[
"0.6265",
"0.1936",
"0.66",
"0.1936",
"0.66",
"0.206",
"0.6265",
"0.206"
],
[
"0.723",
"0.1936",
"0.756",
"0.1936",
"0.756",
"0.206",
"0.723",
"0.206"
],
[
"0.818",
"0.1936",
"0.854",
"0.1936",
"0.854",
"0.206",
"0.818",
"0.206"
],
[
"0.8613",
"0.1936",
"0.8965",
"0.1936",
"0.8965",
"0.206",
"0.8613",
"0.206"
],
[
"0.5264",
"0.2048",
"0.5576",
"0.2048",
"0.5576",
"0.2173",
"0.5264",
"0.2173"
],
[
"0.575",
"0.1948",
"0.6094",
"0.1948",
"0.6094",
"0.206",
"0.575",
"0.206"
],
[
"0.1276",
"0.206",
"0.2351",
"0.206",
"0.2351",
"0.2173",
"0.1276",
"0.2173"
],
[
"0.345",
"0.206",
"0.3787",
"0.206",
"0.3787",
"0.2173",
"0.345",
"0.2173"
],
[
"0.3896",
"0.206",
"0.4248",
"0.206",
"0.4248",
"0.2173",
"0.3896",
"0.2173"
],
[
"0.4358",
"0.206",
"0.4685",
"0.206",
"0.4685",
"0.2173",
"0.4358",
"0.2173"
],
[
"0.4785",
"0.206",
"0.513",
"0.206",
"0.513",
"0.2173",
"0.4785",
"0.2173"
],
[
"0.575",
"0.2048",
"0.6094",
"0.2048",
"0.6094",
"0.2173",
"0.575",
"0.2173"
],
[
"0.6265",
"0.206",
"0.66",
"0.206",
"0.66",
"0.2173",
"0.6265",
"0.2173"
],
[
"0.671",
"0.206",
"0.705",
"0.206",
"0.705",
"0.2173",
"0.671",
"0.2173"
],
[
"0.721",
"0.206",
"0.756",
"0.206",
"0.756",
"0.2173",
"0.721",
"0.2173"
],
[
"0.7715",
"0.206",
"0.806",
"0.206",
"0.806",
"0.2173",
"0.7715",
"0.2173"
],
[
"0.818",
"0.206",
"0.854",
"0.206",
"0.854",
"0.2173",
"0.818",
"0.2173"
],
[
"0.8613",
"0.206",
"0.895",
"0.206",
"0.895",
"0.2173",
"0.8613",
"0.2173"
],
[
"0.12427",
"0.2156",
"0.2627",
"0.2162",
"0.2627",
"0.231",
"0.12427",
"0.2299"
],
[
"0.345",
"0.2173",
"0.3787",
"0.2173",
"0.3787",
"0.2299",
"0.345",
"0.2299"
],
[
"0.3896",
"0.2173",
"0.424",
"0.2173",
"0.424",
"0.2299",
"0.3896",
"0.2299"
],
[
"0.4358",
"0.2173",
"0.4702",
"0.2173",
"0.4702",
"0.2299",
"0.4358",
"0.2299"
],
[
"0.4785",
"0.2173",
"0.511",
"0.2173",
"0.511",
"0.2299",
"0.4785",
"0.2299"
],
[
"0.5264",
"0.2173",
"0.5576",
"0.2173",
"0.5576",
"0.2299",
"0.5264",
"0.2299"
],
[
"0.575",
"0.2173",
"0.6094",
"0.2173",
"0.6094",
"0.2299",
"0.575",
"0.2299"
],
[
"0.6265",
"0.2173",
"0.659",
"0.2173",
"0.659",
"0.2299",
"0.6265",
"0.2299"
],
[
"0.6724",
"0.2173",
"0.705",
"0.2173",
"0.705",
"0.2299",
"0.6724",
"0.2299"
],
[
"0.723",
"0.2173",
"0.756",
"0.2173",
"0.756",
"0.2299",
"0.723",
"0.2299"
],
[
"0.7715",
"0.2173",
"0.806",
"0.2173",
"0.806",
"0.2299",
"0.7715",
"0.2299"
],
[
"0.818",
"0.2173",
"0.854",
"0.2173",
"0.854",
"0.2299",
"0.818",
"0.2299"
],
[
"0.8623",
"0.2173",
"0.895",
"0.2173",
"0.895",
"0.2299",
"0.8623",
"0.2299"
],
[
"0.525",
"0.228",
"0.559",
"0.2299",
"0.5576",
"0.2423",
"0.523",
"0.2405"
],
[
"0.6265",
"0.228",
"0.66",
"0.2299",
"0.659",
"0.2423",
"0.6245",
"0.2405"
],
[
"0.6724",
"0.228",
"0.706",
"0.2299",
"0.705",
"0.2423",
"0.671",
"0.2405"
],
[
"0.721",
"0.228",
"0.756",
"0.2299",
"0.754",
"0.2423",
"0.7197",
"0.2405"
],
[
"0.575",
"0.2286",
"0.6094",
"0.2286",
"0.6094",
"0.2411",
"0.575",
"0.2411"
],
[
"0.12427",
"0.2286",
"0.2477",
"0.2286",
"0.2477",
"0.2423",
"0.12427",
"0.2423"
],
[
"0.345",
"0.2299",
"0.3787",
"0.2299",
"0.3787",
"0.2423",
"0.345",
"0.2423"
],
[
"0.3896",
"0.2299",
"0.4248",
"0.2299",
"0.4248",
"0.2411",
"0.3896",
"0.2411"
],
[
"0.4358",
"0.2299",
"0.4702",
"0.2299",
"0.4702",
"0.2411",
"0.4358",
"0.2411"
],
[
"0.4768",
"0.2299",
"0.513",
"0.2299",
"0.513",
"0.2411",
"0.4768",
"0.2411"
],
[
"0.7715",
"0.2299",
"0.8076",
"0.2299",
"0.8076",
"0.2423",
"0.7715",
"0.2423"
],
[
"0.818",
"0.2299",
"0.852",
"0.2299",
"0.852",
"0.2411",
"0.818",
"0.2411"
],
[
"0.8623",
"0.2299",
"0.8965",
"0.2299",
"0.8965",
"0.2423",
"0.8623",
"0.2423"
],
[
"0.1184",
"0.2542",
"0.882",
"0.2542",
"0.882",
"0.2673",
"0.1184",
"0.2673"
],
[
"0.1184",
"0.2683",
"0.517",
"0.2683",
"0.517",
"0.282",
"0.1184",
"0.282"
],
[
"0.511",
"0.2683",
"0.882",
"0.2683",
"0.882",
"0.282",
"0.511",
"0.282"
],
[
"0.12006",
"0.2832",
"0.3047",
"0.2832",
"0.3047",
"0.2964",
"0.12006",
"0.2964"
],
[
"0.34",
"0.3088",
"0.4055",
"0.3088",
"0.4055",
"0.3206",
"0.34",
"0.3206"
],
[
"0.4592",
"0.3071",
"0.708",
"0.3123",
"0.708",
"0.33",
"0.4592",
"0.3247"
],
[
"0.3997",
"0.31",
"0.4568",
"0.31",
"0.4568",
"0.3218",
"0.3997",
"0.3218"
],
[
"0.1234",
"0.31",
"0.1823",
"0.31",
"0.1823",
"0.3218",
"0.1234",
"0.3218"
],
[
"0.3435",
"0.3206",
"0.3914",
"0.3206",
"0.3914",
"0.335",
"0.3435",
"0.335"
],
[
"0.4072",
"0.3218",
"0.4534",
"0.3218",
"0.4534",
"0.3333",
"0.4072",
"0.3333"
],
[
"0.465",
"0.3184",
"0.513",
"0.3206",
"0.511",
"0.335",
"0.4634",
"0.3333"
],
[
"0.12006",
"0.335",
"0.22",
"0.3374",
"0.22",
"0.3486",
"0.12006",
"0.3467"
],
[
"0.351",
"0.335",
"0.3838",
"0.335",
"0.3838",
"0.3474",
"0.351",
"0.3474"
],
[
"0.413",
"0.3362",
"0.4475",
"0.3362",
"0.4475",
"0.3486",
"0.413",
"0.3486"
],
[
"0.471",
"0.3362",
"0.5073",
"0.3362",
"0.5073",
"0.3474",
"0.471",
"0.3474"
],
[
"0.5264",
"0.3384",
"0.8623",
"0.339",
"0.8623",
"0.357",
"0.5264",
"0.3557"
],
[
"0.12427",
"0.3467",
"0.2468",
"0.3486",
"0.2468",
"0.3599",
"0.12427",
"0.3582"
],
[
"0.351",
"0.3474",
"0.3838",
"0.3474",
"0.3838",
"0.3599",
"0.351",
"0.3599"
],
[
"0.413",
"0.3474",
"0.4458",
"0.3474",
"0.4458",
"0.3599",
"0.413",
"0.3599"
],
[
"0.471",
"0.3474",
"0.5054",
"0.3474",
"0.5054",
"0.3599",
"0.471",
"0.3599"
],
[
"0.529",
"0.3474",
"0.546",
"0.3474",
"0.546",
"0.3552",
"0.529",
"0.3552"
],
[
"0.12427",
"0.3594",
"0.3284",
"0.3594",
"0.3284",
"0.3723",
"0.12427",
"0.3723"
],
[
"0.351",
"0.3599",
"0.382",
"0.3599",
"0.382",
"0.3723",
"0.351",
"0.3723"
],
[
"0.4148",
"0.3599",
"0.4458",
"0.3599",
"0.4458",
"0.3718",
"0.4148",
"0.3718"
],
[
"0.471",
"0.3594",
"0.5054",
"0.3594",
"0.5054",
"0.3718",
"0.471",
"0.3718"
],
[
"0.858",
"0.3599",
"0.882",
"0.3599",
"0.882",
"0.3723",
"0.858",
"0.3723"
],
[
"0.12177",
"0.3694",
"0.294",
"0.3706",
"0.294",
"0.3855",
"0.12177",
"0.3843"
],
[
"0.351",
"0.3723",
"0.3838",
"0.3723",
"0.3838",
"0.3843",
"0.351",
"0.3843"
],
[
"0.413",
"0.3723",
"0.4458",
"0.3723",
"0.4458",
"0.3843",
"0.413",
"0.3843"
],
[
"0.465",
"0.3677",
"0.8867",
"0.3723",
"0.8867",
"0.3943",
"0.465",
"0.3896"
],
[
"0.12427",
"0.3818",
"0.3015",
"0.3818",
"0.3015",
"0.3955",
"0.12427",
"0.3955"
],
[
"0.351",
"0.383",
"0.3838",
"0.383",
"0.3838",
"0.3955",
"0.351",
"0.3955"
],
[
"0.4148",
"0.3843",
"0.444",
"0.3843",
"0.444",
"0.3955",
"0.4148",
"0.3955"
],
[
"0.471",
"0.3843",
"0.5073",
"0.3843",
"0.5073",
"0.3955",
"0.471",
"0.3955"
],
[
"0.4148",
"0.3943",
"0.4458",
"0.3967",
"0.444",
"0.4092",
"0.413",
"0.4067"
],
[
"0.513",
"0.3901",
"0.8833",
"0.3901",
"0.8833",
"0.405",
"0.513",
"0.405"
],
[
"0.351",
"0.3955",
"0.382",
"0.3955",
"0.382",
"0.408",
"0.351",
"0.408"
],
[
"0.126",
"0.3955",
"0.294",
"0.3955",
"0.294",
"0.4092",
"0.126",
"0.4092"
],
[
"0.471",
"0.3955",
"0.5054",
"0.3955",
"0.5054",
"0.4067",
"0.471",
"0.4067"
],
[
"0.471",
"0.4062",
"0.509",
"0.408",
"0.5073",
"0.4192",
"0.4702",
"0.4175"
],
[
"0.51",
"0.405",
"0.5903",
"0.4067",
"0.5903",
"0.4229",
"0.51",
"0.4204"
],
[
"0.351",
"0.408",
"0.3838",
"0.408",
"0.3838",
"0.4204",
"0.351",
"0.4204"
],
[
"0.126",
"0.4067",
"0.3225",
"0.4067",
"0.3225",
"0.4204",
"0.126",
"0.4204"
],
[
"0.4148",
"0.408",
"0.4458",
"0.408",
"0.4458",
"0.4192",
"0.4148",
"0.4192"
],
[
"0.12427",
"0.4187",
"0.293",
"0.4187",
"0.293",
"0.4316",
"0.12427",
"0.4316"
],
[
"0.351",
"0.4192",
"0.3838",
"0.4192",
"0.3838",
"0.4312",
"0.351",
"0.4312"
],
[
"0.4148",
"0.4192",
"0.4475",
"0.4192",
"0.4475",
"0.4312",
"0.4148",
"0.4312"
],
[
"0.471",
"0.4192",
"0.504",
"0.4192",
"0.504",
"0.4316",
"0.471",
"0.4316"
],
[
"0.12427",
"0.4287",
"0.3105",
"0.4312",
"0.3105",
"0.4448",
"0.12427",
"0.4424"
],
[
"0.351",
"0.4312",
"0.3838",
"0.4312",
"0.3838",
"0.4436",
"0.351",
"0.4436"
],
[
"0.413",
"0.4312",
"0.4458",
"0.4312",
"0.4458",
"0.4436",
"0.413",
"0.4436"
],
[
"0.4702",
"0.43",
"0.5054",
"0.4316",
"0.504",
"0.4436",
"0.4685",
"0.4412"
],
[
"0.5156",
"0.4382",
"0.5264",
"0.4382",
"0.5264",
"0.4453",
"0.5156",
"0.4453"
],
[
"0.552",
"0.4358",
"0.6616",
"0.4358",
"0.6616",
"0.4478",
"0.552",
"0.4478"
],
[
"0.12177",
"0.4436",
"0.2612",
"0.4436",
"0.2612",
"0.4573",
"0.12177",
"0.4573"
],
[
"0.351",
"0.4436",
"0.3838",
"0.4436",
"0.3838",
"0.456",
"0.351",
"0.456"
],
[
"0.413",
"0.4448",
"0.4475",
"0.4448",
"0.4475",
"0.456",
"0.413",
"0.456"
],
[
"0.471",
"0.4436",
"0.5054",
"0.4436",
"0.5054",
"0.456",
"0.471",
"0.456"
],
[
"0.3494",
"0.4531",
"0.3862",
"0.456",
"0.3838",
"0.4702",
"0.346",
"0.4673"
],
[
"0.471",
"0.4536",
"0.5054",
"0.456",
"0.504",
"0.4685",
"0.4702",
"0.466"
],
[
"0.413",
"0.456",
"0.4475",
"0.456",
"0.4475",
"0.4685",
"0.413",
"0.4685"
],
[
"0.126",
"0.4548",
"0.2627",
"0.4548",
"0.2627",
"0.4685",
"0.126",
"0.4685"
],
[
"0.513",
"0.4602",
"0.882",
"0.4602",
"0.882",
"0.4744",
"0.513",
"0.4744"
],
[
"0.12427",
"0.466",
"0.2477",
"0.466",
"0.2477",
"0.4797",
"0.12427",
"0.4797"
],
[
"0.351",
"0.4673",
"0.3838",
"0.4673",
"0.3838",
"0.4785",
"0.351",
"0.4785"
],
[
"0.4148",
"0.466",
"0.4475",
"0.466",
"0.4475",
"0.4785",
"0.4148",
"0.4785"
],
[
"0.471",
"0.4673",
"0.503",
"0.4673",
"0.503",
"0.4785",
"0.471",
"0.4785"
],
[
"0.513",
"0.4744",
"0.882",
"0.4768",
"0.882",
"0.4912",
"0.513",
"0.4893"
],
[
"0.11755",
"0.4905",
"0.4902",
"0.4912",
"0.4902",
"0.506",
"0.11755",
"0.505"
],
[
"0.5156",
"0.4922",
"0.8833",
"0.4922",
"0.8833",
"0.506",
"0.5156",
"0.506"
],
[
"0.11755",
"0.506",
"0.4878",
"0.506",
"0.4878",
"0.5195",
"0.11755",
"0.5195"
],
[
"0.5156",
"0.509",
"0.8213",
"0.509",
"0.8213",
"0.5225",
"0.5156",
"0.5225"
],
[
"0.1184",
"0.5215",
"0.1931",
"0.5215",
"0.1931",
"0.533",
"0.1184",
"0.533"
],
[
"0.5156",
"0.536",
"0.5415",
"0.536",
"0.5415",
"0.5474",
"0.5156",
"0.5474"
],
[
"0.559",
"0.5337",
"0.711",
"0.536",
"0.711",
"0.5503",
"0.559",
"0.549"
],
[
"0.12006",
"0.55",
"0.486",
"0.55",
"0.486",
"0.5635",
"0.12006",
"0.5635"
],
[
"0.5146",
"0.5547",
"0.8833",
"0.5557",
"0.8833",
"0.571",
"0.5146",
"0.5693"
],
[
"0.11755",
"0.564",
"0.3594",
"0.5654",
"0.3594",
"0.58",
"0.11755",
"0.579"
],
[
"0.5146",
"0.572",
"0.881",
"0.572",
"0.881",
"0.585",
"0.5146",
"0.585"
],
[
"0.3914",
"0.586",
"0.4702",
"0.5957",
"0.465",
"0.611",
"0.3862",
"0.6016"
],
[
"0.1335",
"0.593",
"0.4417",
"0.593",
"0.4417",
"0.6133",
"0.1335",
"0.6133"
],
[
"0.5146",
"0.5884",
"0.882",
"0.5874",
"0.882",
"0.6016",
"0.5146",
"0.6025"
],
[
"0.5146",
"0.605",
"0.882",
"0.605",
"0.882",
"0.618",
"0.5146",
"0.618"
],
[
"0.2334",
"0.6133",
"0.3074",
"0.6133",
"0.3074",
"0.6245",
"0.2334",
"0.6245"
],
[
"0.4626",
"0.6245",
"0.4878",
"0.6245",
"0.4878",
"0.637",
"0.4626",
"0.637"
],
[
"0.5146",
"0.6206",
"0.882",
"0.6206",
"0.882",
"0.6343",
"0.5146",
"0.6343"
],
[
"0.1377",
"0.64",
"0.4878",
"0.64",
"0.4878",
"0.654",
"0.1377",
"0.654"
],
[
"0.5146",
"0.6357",
"0.882",
"0.6353",
"0.882",
"0.6494",
"0.5146",
"0.651"
],
[
"0.1184",
"0.655",
"0.4878",
"0.6562",
"0.4878",
"0.6704",
"0.1184",
"0.6694"
],
[
"0.513",
"0.652",
"0.881",
"0.652",
"0.881",
"0.665",
"0.513",
"0.665"
],
[
"0.12006",
"0.673",
"0.4878",
"0.673",
"0.4878",
"0.686",
"0.12006",
"0.686"
],
[
"0.5146",
"0.6685",
"0.881",
"0.6685",
"0.881",
"0.6816",
"0.5146",
"0.6816"
],
[
"0.1184",
"0.6885",
"0.4844",
"0.687",
"0.4844",
"0.7017",
"0.1184",
"0.703"
],
[
"0.5146",
"0.684",
"0.882",
"0.685",
"0.882",
"0.6997",
"0.5146",
"0.6987"
],
[
"0.12006",
"0.705",
"0.4895",
"0.705",
"0.4895",
"0.7188",
"0.12006",
"0.7188"
],
[
"0.5146",
"0.7007",
"0.8833",
"0.7007",
"0.8833",
"0.7153",
"0.5146",
"0.7153"
],
[
"0.12006",
"0.72",
"0.4878",
"0.72",
"0.4878",
"0.734",
"0.12006",
"0.734"
],
[
"0.513",
"0.7173",
"0.7363",
"0.7173",
"0.7363",
"0.731",
"0.513",
"0.731"
],
[
"0.12006",
"0.7363",
"0.4878",
"0.7363",
"0.4878",
"0.75",
"0.12006",
"0.75"
],
[
"0.531",
"0.733",
"0.8833",
"0.734",
"0.8833",
"0.749",
"0.531",
"0.7476"
],
[
"0.12177",
"0.753",
"0.4844",
"0.753",
"0.4844",
"0.7666",
"0.12177",
"0.7666"
],
[
"0.5146",
"0.7505",
"0.8833",
"0.7505",
"0.8833",
"0.764",
"0.5146",
"0.764"
],
[
"0.4802",
"0.7573",
"0.4844",
"0.7573",
"0.4844",
"0.7627",
"0.4802",
"0.7627"
],
[
"0.12177",
"0.7686",
"0.4878",
"0.7686",
"0.4878",
"0.782",
"0.12177",
"0.782"
],
[
"0.5146",
"0.7666",
"0.882",
"0.7666",
"0.882",
"0.78",
"0.5146",
"0.78"
],
[
"0.12177",
"0.785",
"0.403",
"0.785",
"0.403",
"0.799",
"0.12177",
"0.799"
],
[
"0.517",
"0.782",
"0.882",
"0.782",
"0.882",
"0.7964",
"0.517",
"0.7964"
],
[
"0.513",
"0.7964",
"0.8833",
"0.7974",
"0.8833",
"0.8125",
"0.513",
"0.811"
],
[
"0.12177",
"0.811",
"0.4878",
"0.811",
"0.4878",
"0.826",
"0.12177",
"0.826"
],
[
"0.5146",
"0.814",
"0.882",
"0.814",
"0.882",
"0.8276",
"0.5146",
"0.8276"
],
[
"0.215",
"0.8345",
"0.4878",
"0.8345",
"0.4878",
"0.847",
"0.215",
"0.847"
],
[
"0.5146",
"0.829",
"0.8833",
"0.83",
"0.8833",
"0.844",
"0.5146",
"0.8433"
],
[
"0.4634",
"0.8516",
"0.4878",
"0.8516",
"0.4878",
"0.8643",
"0.4634",
"0.8643"
],
[
"0.5156",
"0.8467",
"0.764",
"0.8467",
"0.764",
"0.86",
"0.5156",
"0.86"
],
[
"0.534",
"0.8633",
"0.882",
"0.8633",
"0.882",
"0.877",
"0.534",
"0.877"
],
[
"0.1368",
"0.879",
"0.4878",
"0.879",
"0.4878",
"0.8926",
"0.1368",
"0.8926"
],
[
"0.5146",
"0.879",
"0.882",
"0.8774",
"0.882",
"0.8926",
"0.5146",
"0.8936"
],
[
"0.12006",
"0.8945",
"0.4878",
"0.8945",
"0.4878",
"0.909",
"0.12006",
"0.909"
],
[
"0.513",
"0.8945",
"0.882",
"0.8945",
"0.882",
"0.909",
"0.513",
"0.909"
]
] | [
"Method",
"GermantoImage",
"Image to German",
"English to Image",
"Image to English",
"R@1 R@5 R10",
"R@1",
"R@5",
"R@10",
"R@1",
"R@5",
"R@10",
"R@1",
"R@5R10",
"VSEf* (Kiros et al.,2014)",
"47.2",
"60.1",
"29.3",
"58.1",
"71.8",
"23.3",
"53.6",
"65.8",
"31.6",
"72.7",
"60.4",
"20.3",
"OE + * (Vendrov et al., 2015)",
"57.5",
"70.9",
"56.5",
"67.8",
"74.8",
"26.8",
"60.4",
"34.8",
"63.7",
"48.5",
"21.0",
"25.8",
"DAN* (Nam et al., 2017)",
"60.9",
"46.5",
"77.5",
"71.0",
"31.0",
"83.0",
"39.4",
"69.2",
"69.1",
"55.0",
"81.8",
"89.0",
"VSE++* (Faghri et al., 2018)",
"31.3",
"62.2",
"70.9",
"47.5",
"78.5",
"84.5",
"39.6",
"69.1",
"79.8",
"53.1",
"82.1",
"87.5",
"SCAN* (Lee et al. 2018)",
"35.7",
"64.9",
"74.6",
"52.3",
"81.8",
"88.5",
"45.8",
"74.4",
"83.0",
"61.8",
"87.5",
"93.7",
"Pivot (Gella et al., 2017)",
"49.3",
"61.9",
"73.4",
"75.2",
"61.7",
"22.5",
"28.2",
"26.2",
"56.4",
"68.4",
"33.8",
"62.8",
"Oursf(Random,VGG19)",
"25.8",
"54.9",
"65.1",
"34.1",
"65.5",
"76.5",
"30.1",
"62.5",
"71.6",
"36.4",
"68.0",
"80.9",
"Ours (Random, No diversity)",
"36.3",
"65.3",
"74.7",
"63.3",
"53.1",
"74.7",
"82.3",
"46.2",
"82.9",
"87.0",
"93.3",
"84.7",
"88.8",
"Ours (Random)",
"39.2",
"67.5",
"76.7",
"55.0",
"91.2",
"48.7",
"77.2",
"85.0",
"66.4",
"88.3",
"93.4",
"Ours (w/ FastText)",
"40.3",
"70.1",
"79.0",
"60.4",
"85.4",
"92.0",
"50.1",
"78.1",
"85.7",
"68.0",
"88.8",
"94.0",
"84.6",
"48.9",
"78.3",
"85.8",
"91.3",
"Ours (w/ BERT)",
"40.7",
"70.5",
"78.8",
"56.5",
"66.5",
"89.1",
"94.1",
"Table 1: Comparison of multilingual sentence-image retrieval/matching (German-Image) and (English-",
"Image) results on Multi30K. (Visual encoders:VGG",
"t otherwise ResNet or Faster-RCNN(ResNet).)",
"(Monolingual models * .)",
"MS-Vid",
"Pascal最终模型损失形式化为：",
"Pascal",
"Method",
"(2012)",
"(2014)",
"(2015)",
"STSBaseline",
"29.9",
"51.3",
"60.4",
"AU(V, E, G) = le(V,E,G) + βID(V,E, G),",
"STSBest System",
"86.3",
"83.4",
"86.4",
"",
"GRAN (Wieting et al., 2017)",
"83.7",
"84.5",
"85.0",
"(7)",
"VSE (Kiros et al., 2014)",
"80.6",
"82.7",
"89.。其中β是加权参数，用于平衡多样性损失和三",
"OE (Vendrov et al., 2015)",
"82.2",
"84.1",
"90.8",
"84.3",
"重排名损失。我们通过最小化lA\"(V,E,G)来",
"84.1",
"DAN (Nam et al., 2017)",
"90.8",
"91.2",
"训练模型",
"84.5",
"VSE++ (Faghri et al., 2018)",
"84.8",
"SCAN (Lee et al., 2018)",
"84.0",
"83.9",
"90.7",
"PIVOT (Gella et al., 2017)",
"84.6",
"84.5",
"91.5",
"4",
"Experiments",
"Ours (w/ Random)",
"85.8",
"87.8",
"91.5",
"86.2",
"91.8",
"88.3",
"Ours (w/ FastText)",
"继EGellaetal.（2017）之后，我们对",
"Ours (w/ BERT)",
"86.4",
"88.0",
"91.7",
"Multi30K国（Elliottetal.，2016）中的多语言",
"Table 2: Results on the image and video datasets of",
"句子图像匹配任务和语义文本相似性任务E",
"Semantic Textual Similarity task. (Pearson'aeTMs",
"（Agirreetal.，2012,2015）进行了评估",
"r × 100 ).",
"4.1",
"Experiment Setup",
"像和英语句子之间的多头注意力多样性损失",
"我们在国Andersonetal.（2018）中使用该模",
"（i.e.E跨模态多样性）定义为：",
"型，该模型是在MS-COCOE(Linetal..2014)",
"kek+r)]+",
"iB(V,E) =ZZZ[aD - s(ub,eb+",
"数据集上预先训练并在VisualGenomeE(Kr-",
"ishnaetal.,2016)数据集上进行微调的Faster-",
"pkr",
"(5)",
"RCNNE(Renetal.，2015）网络，以检测显着",
"如图1中用于更新的红色箭头所示，此多",
"视觉对象并提取其相应特征。1,600种类型的",
"样性目标背后的优点是增加了实例本身或其跨",
"物体是可检测的。然后，我们将每个图像打包",
"模态并行实例的来自不同关注头的受关注嵌人",
"并表示为36×2048特征矩阵，其中36是图像",
"之间的距离（最大为多样性余量αD）。结果",
"中突出视觉对象的最大数量，2048是Faster-",
"多样性目标明确鼓励多头注意力集中在稀疏",
"RCNN的ResNetE(He et al.,2016)主千中平",
"位于联合嵌人空间中的信息的不同方面，以促",
"坦的最后一个池化层的尺寸",
"进多语言文本语义与视觉对象之间的细粒度对",
"对于文本处理，我们小写，标记化，然后将",
"齐。由于共享嵌人空间是多语言和多模式的",
"最大句子长度截断为100。我们使用随机或",
"",
"因此为了改善模式内/语言和模式间/语言多样",
"预先训练的多语言嵌入初始化的300-dim单",
"性，我们将整体多样性损失建模为：",
"词嵌人矩阵。说明：我们使用多语言版本的",
"FastText（Mikolovetal.，2018）。我们还尝试",
"D(V, E,G) = D(V, V) +D(G,G)+D(E, E)",
"合并最后一层上下文化多语言BERT嵌E",
"+P(V, E) +P(V,G) +(G, E),",
"(Devlinetal.，2018），以替换单词嵌人矩阵作",
"(6)",
"为双向LSTM的文本输人功能",
"对于训练，我们对大小为128的批次进行",
"其中前三个术语是模式内/语言，其余是跨",
"采样，并在Multi30K的训练集上训练20个",
"模式/语言。使用公式E4和公式E6，我们将",
"epoch。我们使用具有2×10-4学习速率的"
] | {
"arxiv_id": "1910.00058",
"language": "zh",
"page_idx": 2
} |
||
1910.00058-zh-4 | 1910.00058-zh-4.jpg | <FigureHere>
Figure 3: t-SNE visualization and grounding of the learned multilingual multimodal embeddings on Multi30K. Note the sentences are not translation pairs.
象的，而空间关系(e.g.$\mathrm{\overline{\mathrm{E}}}^{\mathrm{~}}$on",“over”)和量词仍然没有很好地对齐。解决这些限制将是我们未来的工作
## 5 Conclusion
我们提出了一种新颖的 VSE 模型，该模型可促进具有多样性的多头注意力，以对齐不同类型的文本语义和视觉对象，以学习扎根的多语言多模态表示。所提出的模型在两个基准数据集上获得了多语言句子图像匹配任务和语义文本相似性任务的最新结果
6
Acknowledgement本研究部分由 DARPA 根据 AIDA 和 LwLL'许划授予 FA8750-18-2- 0018和 FA8750-19-2-0501 支持。IARPA 授权也通过 DOI/IBC 编号 D17PC00340 提供支持。我们要感谢匿名审稿人的建设性建议
Multilingual semantic textual similarity. In Proceedings of the 8th international workshop on semantic evaluation (SemEval 2014), pages 81- 91.
Eneko Agirre, Daniel Cer, Mona Diab, and Aitor
Gonzalez-Agirre. 2012. Semeval-2012 task 6: A pilot on semantic textual similarity. In * SEM 2012: The First Joint Conference on Lexical and Computational Semantics-Volume 1: Proceedings of the main conference and the shared task, and Volume 2: Proceedings of the Sixth International Workshop on Semantic Evaluation (Se- ${\mathrm{mEval~2012),~volume~1,~pages~385-393.}}$
Peter Anderson, Xiaodong He, Chris Buehler,
Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. 2018. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR.
Iacer Calixto, Qun Liu, and Nick Campbell. 2017. Multilingual multi-modal embeddings for natural language processing. arXiv preprint $\operatorname{arXiv:1702.\bar{01101}.}$
David Chen and William Dolan. 2011. Collecting
highly parallel data for paraphrase evaluation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 190-200.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. 2018. Bert: Pre-training of $\operatorname*{deep}$bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.
Desmond Elliott, Stella Frank, Khalil Sima'an,
and Lucia Specia. 2016. Multi30k: Multilingual english-german image descriptions. In Proceedings of the 5th Workshop on Vision and Language, pages 70-74. Association for Computational Linguistics.
Fartash Faghri, David J Fleet, Jamie Ryan Kiros,
and Sanja Fidler. 2018. Vse++: Improving visual-semantic embeddings with hard negatives.
Spandana Gella, Rico Sennrich, Frank Keller, and
Mirella Lapata. 2017. Image'pivoting for learn- $\operatorname{ing}$ multilingual multimodal representations. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages $2839–2845.$ Association for Computational $\operatorname{Linguistics.}$
Jiuxiang Gu, Jianfei Cai, Shafiq R. Joty, Li Niu,
and Gang Wang. 2018. Look, imagine and match: Improving textual-visual cross-modal retrieval with generative models. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
References
Eneko Agirre, Carmen Banea, Claire Cardie,
Daniel $\overset {\prime }{\operatorname* { \operatorname* { \operatorname* { Cer, } } } }$Mona Diab, Aitor'Gonzalez-Agirre, Weiwei Guo, Inigo Lopez-Gazpio, Montse Maritxalar, Rada Mihalcea, et al. 2015. Semeval- 2015 task 2: Semantic textual similarity, english, spanish and pilot on interpretability. In $\bar{\text{Proceedings of the 9th international workshop}}$ on semantic evaluation (SemEval 2015), pages 252-263.
Eneko Agirre, Carmen Banea, Claire Cardie,
Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Rada Mihalcea, German Rigau, and Janyce Wiebe. 2014. Semeval-2014 task 10: | [
[
"0.532",
"0.0766",
"0.882",
"0.0766",
"0.882",
"0.0903",
"0.532",
"0.0903"
],
[
"0.413",
"0.08435",
"0.4417",
"0.08435",
"0.4417",
"0.0968",
"0.413",
"0.0968"
],
[
"0.529",
"0.0885",
"0.8833",
"0.0885",
"0.8833",
"0.10156",
"0.529",
"0.10156"
],
[
"0.531",
"0.1027",
"0.881",
"0.1027",
"0.881",
"0.1164",
"0.531",
"0.1164"
],
[
"0.531",
"0.1164",
"0.5557",
"0.1164",
"0.5557",
"0.1289",
"0.531",
"0.1289"
],
[
"0.5146",
"0.1426",
"0.8833",
"0.1426",
"0.8833",
"0.1561",
"0.5146",
"0.1561"
],
[
"0.531",
"0.1561",
"0.8833",
"0.1561",
"0.8833",
"0.1692",
"0.531",
"0.1692"
],
[
"0.531",
"0.1675",
"0.882",
"0.1675",
"0.882",
"0.1812",
"0.531",
"0.1812"
],
[
"0.529",
"0.1812",
"0.8833",
"0.1812",
"0.8833",
"0.1948",
"0.529",
"0.1948"
],
[
"0.532",
"0.1953",
"0.8833",
"0.1953",
"0.8833",
"0.209",
"0.532",
"0.209"
],
[
"0.531",
"0.2102",
"0.881",
"0.2102",
"0.881",
"0.2203",
"0.531",
"0.2203"
],
[
"0.529",
"0.2214",
"0.8833",
"0.2214",
"0.8833",
"0.2351",
"0.529",
"0.2351"
],
[
"0.531",
"0.234",
"0.8833",
"0.234",
"0.8833",
"0.2477",
"0.531",
"0.2477"
],
[
"0.531",
"0.2477",
"0.8228",
"0.2477",
"0.8228",
"0.2612",
"0.531",
"0.2612"
],
[
"0.345",
"0.2612",
"0.4802",
"0.2612",
"0.4802",
"0.2756",
"0.345",
"0.2756"
],
[
"0.513",
"0.275",
"0.6514",
"0.2756",
"0.6514",
"0.2874",
"0.513",
"0.2861"
],
[
"0.6455",
"0.275",
"0.882",
"0.275",
"0.882",
"0.288",
"0.6455",
"0.288"
],
[
"0.12006",
"0.2852",
"0.318",
"0.2852",
"0.318",
"0.2957",
"0.12006",
"0.2957"
],
[
"0.4307",
"0.2893",
"0.465",
"0.2893",
"0.465",
"0.2964",
"0.4307",
"0.2964"
],
[
"0.531",
"0.288",
"0.8833",
"0.288",
"0.8833",
"0.3018",
"0.531",
"0.3018"
],
[
"0.12006",
"0.2957",
"0.293",
"0.2957",
"0.293",
"0.306",
"0.12006",
"0.306"
],
[
"0.3435",
"0.2903",
"0.372",
"0.2903",
"0.372",
"0.2944",
"0.3435",
"0.2944"
],
[
"0.34",
"0.2976",
"0.482",
"0.2976",
"0.482",
"0.308",
"0.34",
"0.308"
],
[
"0.12177",
"0.304",
"0.178",
"0.304",
"0.178",
"0.3113",
"0.12177",
"0.3113"
],
[
"0.1931",
"0.3047",
"0.34",
"0.3047",
"0.34",
"0.3123",
"0.1931",
"0.3123"
],
[
"0.531",
"0.2998",
"0.885",
"0.2998",
"0.885",
"0.3142",
"0.531",
"0.3142"
],
[
"0.3376",
"0.306",
"0.3552",
"0.306",
"0.3552",
"0.3135",
"0.3376",
"0.3135"
],
[
"0.1234",
"0.3142",
"0.2258",
"0.3142",
"0.2258",
"0.3184",
"0.1234",
"0.3184"
],
[
"0.2644",
"0.3135",
"0.315",
"0.3135",
"0.315",
"0.3206",
"0.2644",
"0.3206"
],
[
"0.531",
"0.3142",
"0.8833",
"0.3142",
"0.8833",
"0.328",
"0.531",
"0.328"
],
[
"0.11755",
"0.33",
"0.4895",
"0.33",
"0.4895",
"0.3433",
"0.11755",
"0.3433"
],
[
"0.531",
"0.3267",
"0.727",
"0.3267",
"0.727",
"0.3403",
"0.531",
"0.3403"
],
[
"0.11755",
"0.3433",
"0.4878",
"0.3457",
"0.4878",
"0.3599",
"0.11755",
"0.3582"
],
[
"0.513",
"0.354",
"0.882",
"0.354",
"0.882",
"0.3677",
"0.513",
"0.3677"
],
[
"0.11755",
"0.3599",
"0.486",
"0.361",
"0.486",
"0.3723",
"0.11755",
"0.3718"
],
[
"0.529",
"0.3652",
"0.8833",
"0.3665",
"0.8833",
"0.3813",
"0.529",
"0.3801"
],
[
"0.1184",
"0.376",
"0.1948",
"0.376",
"0.1948",
"0.3872",
"0.1184",
"0.3872"
],
[
"0.529",
"0.379",
"0.8833",
"0.3801",
"0.8833",
"0.3943",
"0.529",
"0.3938"
],
[
"0.531",
"0.3926",
"0.665",
"0.3938",
"0.665",
"0.4067",
"0.531",
"0.4062"
],
[
"0.12006",
"0.3984",
"0.486",
"0.398",
"0.486",
"0.412",
"0.12006",
"0.4133"
],
[
"0.12006",
"0.415",
"0.4878",
"0.415",
"0.4878",
"0.4287",
"0.12006",
"0.4287"
],
[
"0.511",
"0.4192",
"0.885",
"0.4204",
"0.885",
"0.434",
"0.511",
"0.4329"
],
[
"0.12006",
"0.4312",
"0.2301",
"0.4312",
"0.2301",
"0.4453",
"0.12006",
"0.4453"
],
[
"0.534",
"0.4353",
"0.882",
"0.4353",
"0.882",
"0.4453",
"0.534",
"0.4453"
],
[
"0.531",
"0.4465",
"0.8867",
"0.4465",
"0.8867",
"0.4602",
"0.531",
"0.4602"
],
[
"0.12177",
"0.456",
"0.1368",
"0.456",
"0.1368",
"0.4656",
"0.12177",
"0.4656"
],
[
"0.1511",
"0.4531",
"0.2544",
"0.4548",
"0.2544",
"0.4697",
"0.1511",
"0.4673"
],
[
"0.529",
"0.459",
"0.885",
"0.4602",
"0.885",
"0.4739",
"0.529",
"0.4727"
],
[
"0.12006",
"0.4744",
"0.4878",
"0.4744",
"0.4878",
"0.488",
"0.12006",
"0.488"
],
[
"0.531",
"0.4727",
"0.881",
"0.4727",
"0.881",
"0.4863",
"0.531",
"0.4863"
],
[
"0.1184",
"0.4905",
"0.486",
"0.4905",
"0.486",
"0.5044",
"0.1184",
"0.5044"
],
[
"0.1184",
"0.505",
"0.4878",
"0.506",
"0.4878",
"0.521",
"0.1184",
"0.5195"
],
[
"0.513",
"0.4995",
"0.885",
"0.4995",
"0.885",
"0.513",
"0.513",
"0.513"
],
[
"0.529",
"0.5127",
"0.8867",
"0.513",
"0.8867",
"0.527",
"0.529",
"0.5254"
],
[
"0.12177",
"0.5225",
"0.4878",
"0.5225",
"0.4878",
"0.5376",
"0.12177",
"0.5376"
],
[
"0.529",
"0.525",
"0.8833",
"0.5254",
"0.8833",
"0.5405",
"0.529",
"0.539"
],
[
"0.1184",
"0.538",
"0.4878",
"0.538",
"0.4878",
"0.552",
"0.1184",
"0.552"
],
[
"0.529",
"0.538",
"0.876",
"0.538",
"0.876",
"0.5527",
"0.529",
"0.5527"
],
[
"0.1184",
"0.554",
"0.3242",
"0.5547",
"0.3242",
"0.5693",
"0.1184",
"0.5684"
],
[
"0.513",
"0.564",
"0.882",
"0.5654",
"0.882",
"0.58",
"0.513",
"0.579"
],
[
"0.1184",
"0.579",
"0.1335",
"0.579",
"0.1335",
"0.5903",
"0.1184",
"0.5903"
],
[
"0.529",
"0.579",
"0.8833",
"0.579",
"0.8833",
"0.593",
"0.529",
"0.593"
],
[
"0.532",
"0.5933",
"0.881",
"0.5933",
"0.881",
"0.604",
"0.532",
"0.604"
],
[
"0.12006",
"0.6016",
"0.486",
"0.6016",
"0.486",
"0.615",
"0.12006",
"0.615"
],
[
"0.5283",
"0.605",
"0.8833",
"0.604",
"0.8833",
"0.618",
"0.5283",
"0.619"
],
[
"0.1184",
"0.6143",
"0.486",
"0.6143",
"0.486",
"0.629",
"0.1184",
"0.629"
],
[
"0.531",
"0.618",
"0.8833",
"0.618",
"0.8833",
"0.632",
"0.531",
"0.632"
],
[
"0.1184",
"0.63",
"0.4878",
"0.63",
"0.4878",
"0.6436",
"0.1184",
"0.6436"
],
[
"0.531",
"0.63",
"0.663",
"0.632",
"0.663",
"0.6455",
"0.531",
"0.6436"
],
[
"0.12006",
"0.6445",
"0.486",
"0.6455",
"0.486",
"0.66",
"0.12006",
"0.659"
],
[
"0.1184",
"0.662",
"0.4492",
"0.662",
"0.4492",
"0.676",
"0.1184",
"0.676"
],
[
"0.513",
"0.6567",
"0.8833",
"0.6577",
"0.8833",
"0.673",
"0.513",
"0.6714"
],
[
"0.529",
"0.6714",
"0.7197",
"0.6714",
"0.7197",
"0.685",
"0.529",
"0.685"
],
[
"0.7305",
"0.6694",
"0.885",
"0.6714",
"0.885",
"0.686",
"0.7305",
"0.684"
],
[
"0.529",
"0.683",
"0.8833",
"0.685",
"0.8833",
"0.6987",
"0.529",
"0.697"
],
[
"0.531",
"0.6987",
"0.574",
"0.6987",
"0.574",
"0.7104",
"0.531",
"0.7104"
],
[
"0.1184",
"0.705",
"0.2141",
"0.705",
"0.2141",
"0.7197",
"0.1184",
"0.7197"
],
[
"0.2377",
"0.7256",
"0.4895",
"0.727",
"0.4895",
"0.741",
"0.2377",
"0.7407"
],
[
"0.511",
"0.7246",
"0.8833",
"0.7236",
"0.8833",
"0.737",
"0.511",
"0.7383"
],
[
"0.531",
"0.7383",
"0.8833",
"0.7383",
"0.8833",
"0.752",
"0.531",
"0.752"
],
[
"0.1184",
"0.728",
"0.241",
"0.7285",
"0.241",
"0.7407",
"0.1184",
"0.7393"
],
[
"0.1368",
"0.7407",
"0.4878",
"0.7407",
"0.4878",
"0.7544",
"0.1368",
"0.7544"
],
[
"0.1377",
"0.7544",
"0.4878",
"0.7544",
"0.4878",
"0.767",
"0.1377",
"0.767"
],
[
"0.531",
"0.7505",
"0.885",
"0.7505",
"0.885",
"0.764",
"0.531",
"0.764"
],
[
"0.1368",
"0.7666",
"0.4878",
"0.7666",
"0.4878",
"0.78",
"0.1368",
"0.78"
],
[
"0.531",
"0.764",
"0.8833",
"0.764",
"0.8833",
"0.778",
"0.531",
"0.778"
],
[
"0.1351",
"0.7793",
"0.4895",
"0.78",
"0.4895",
"0.7935",
"0.1351",
"0.792"
],
[
"0.529",
"0.7754",
"0.882",
"0.777",
"0.882",
"0.7915",
"0.529",
"0.7905"
],
[
"0.1368",
"0.792",
"0.4895",
"0.792",
"0.4895",
"0.8057",
"0.1368",
"0.8057"
],
[
"0.531",
"0.7905",
"0.8833",
"0.7905",
"0.8833",
"0.804",
"0.531",
"0.804"
],
[
"0.1368",
"0.8057",
"0.4895",
"0.8057",
"0.4895",
"0.8193",
"0.1368",
"0.8193"
],
[
"0.531",
"0.8027",
"0.617",
"0.804",
"0.617",
"0.8154",
"0.531",
"0.814"
],
[
"0.1351",
"0.818",
"0.4895",
"0.8184",
"0.4895",
"0.833",
"0.1351",
"0.832"
],
[
"0.1368",
"0.833",
"0.2024",
"0.833",
"0.2024",
"0.844",
"0.1368",
"0.844"
],
[
"0.5146",
"0.83",
"0.882",
"0.83",
"0.882",
"0.8433",
"0.5146",
"0.8433"
],
[
"0.531",
"0.8433",
"0.8833",
"0.8433",
"0.8833",
"0.857",
"0.531",
"0.857"
],
[
"0.1184",
"0.855",
"0.4895",
"0.855",
"0.4895",
"0.869",
"0.1184",
"0.869"
],
[
"0.529",
"0.8555",
"0.882",
"0.8555",
"0.882",
"0.869",
"0.529",
"0.869"
],
[
"0.1368",
"0.869",
"0.4895",
"0.869",
"0.4895",
"0.883",
"0.1368",
"0.883"
],
[
"0.532",
"0.869",
"0.885",
"0.869",
"0.885",
"0.883",
"0.532",
"0.883"
],
[
"0.1351",
"0.8813",
"0.4895",
"0.882",
"0.4895",
"0.8965",
"0.1351",
"0.8955"
],
[
"0.529",
"0.882",
"0.885",
"0.8813",
"0.885",
"0.8955",
"0.529",
"0.8965"
],
[
"0.1368",
"0.8955",
"0.4895",
"0.8955",
"0.4895",
"0.909",
"0.1368",
"0.909"
],
[
"0.529",
"0.8955",
"0.693",
"0.8955",
"0.693",
"0.909",
"0.529",
"0.909"
]
] | [
"Multilingual semantic textual similarity. In Pro",
"",
" ceedings of the 8th international workshop on",
"semantic evaluation (SemEval 2014), pages 81-",
"91.",
"Eneko Agirre, Daniel Cer, Mona Diab, and Aitor",
"Gonzalez-Agirre. 2012. Semeval-2012 task 6: A",
"pilot on semantic textual similarity. In * SEM",
"2012: The First Joint Conference on Lexical and",
"Computational Semantics-Volume l: Proceed-",
"ings of themain conference and the sharedtask.",
"and Volume 2: Proceedings of the Sixth Inter-",
"national Workshop on Semantic Evaluation (Se-",
"mEval 2012), volume 1, pages 385-393.",
"BARCELNA2010",
"PeterAnderson,",
"Xiaodong He, Chris Buehler,",
"Three children in football uniforms of two different",
"ngovera",
"Damien Teney, Mark Johnson, Stephen Gould,",
"teams are playing footballon a footballfield",
"Aunm",
"Die frou springt uberdie stange auf",
"3kindergm.sp",
"Vatz,zweiimblauendress,einerim",
"and Lei Zhang. 2018. Bottom-up and top-down",
"die",
"schwarz-weiRen.mithlaren",
"menrangeln",
"attention for image captioning and visual ques-",
"Figure 3: t-SNE visualization and grounding of",
"tion answering. In CVPR.",
"the learned multilingual multimodal embeddings",
"Iacer Calixto, Qun Liu, and Nick Campbell",
"on Multi30K.Note the sentences are not transla-",
"2017. Multilingual multi-modal embeddings for",
"tion pairs.",
"natural language processing.arXiv preprint",
"arXiv:1702.01101.",
"象的，而空间关系（e.g.Eon”，“over\"）和量",
"词仍然没有很好地对齐。解决这些限制将是我",
"David Chen and William Dolan.2011. Collecting",
"们未来的工作",
"highly parallel data for paraphrase evaluation.",
"In Proceedings of the 49th Annual Meeting of",
"5",
"Conclusion",
"the Association for Computational Linguistics:",
"我们提出了一种新颖的VSE模型，该模型可",
"Human Language Technologies, pages 190-200.",
"促进具有多样性的多头注意力，以对齐不同类",
"型的文本语义和视觉对象，以学习扎根的多语",
"Jacob Devlin, Ming-Wei Chang, Kenton Lee, and",
"KristinaToutanova.2018.Bert:Pre-training of",
"言多模态表示。所提出的模型在两个基准数据",
" deep bidirectional transformers for language un-",
"集上获得了多语言句子图像匹配任务和语义文",
" derstanding. arXiv preprint arXiv:1810.04805.",
"本相似性任务的最新结果",
"Desmond Elliott, Stella Frank, Khalil Sima'an",
"6",
"and Lucia Specia. 2016. Multi30k: Multilingual",
"english-german image descriptions. In Proceed-",
"Acknowledgement本研究部分由DARPA根",
"ings of the 5th Workshop on Vision and Lan-",
"据AIDA和 LwLL‘评划授孕“FA8750-18-2-",
"guage, pages 70-74. Association for Computa-",
"0018和FA8750-19-2-0501支持。IARPA授",
"tional Linguistics.",
"权也通过DOI/IBC编号D17PC00340提供支",
"持。我们要感谢匿名审稿人的建设性建议",
"Fartash Faghri, David J Fleet, Jamie Ryan Kiros,",
"and Sanja Fidler. 2018.",
"Vse++:Improving",
"visual-semantic embeddings with hard nega-",
"tives.",
"References",
"Carmen Banea, Claire Cardie,",
"Spandana Gella, Rico Sennrich, Frank Keller, and",
"Mirella Lapata. 2017. Image pivoting for learn-",
"EnekoAgirre,",
"Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre,",
"Weiwei Guo, Inigo Lopez-Gazpio, Montse Mar-",
"ing multilingual multimodal representations. In",
"itxalar, Rada Mihalcea, et al. 2015. Semeval-",
"Proceedings of the 2017 Conference on Empir-",
"2015 task 2: Semantic textual similarity,en-",
"ical Methods in Natural Language Processing,",
"glish, spanish and pilot on interpretability. In",
"pages 2839-2845. Association for Computational",
"Proceedings of the 9th international workshop",
"Linguistics.",
"on semantic evaluation (SemEval 2015), pages",
"252-263.",
"Jiuxiang Gu, Jianfei Cai, Shafiq R. Joty, Li Niu,",
"and Gang Wang. 2018.Look, imagine and",
"Eneko Agirre, Carmen Banea, Claire Cardie,",
"match: Improving textual-visual cross-modal re-",
"Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre,",
"trieval with generative models. In The IEEE",
"Weiwei Guo, Rada Mihalcea, German Rigau,",
"Conference on Computer Vision and Pattern",
"and Janyce Wiebe. 2014. Semeval-2014 task 10:",
"Recognition (CVPR)."
] | {
"arxiv_id": "1910.00058",
"language": "zh",
"page_idx": 4
} |
||
1810.04160-zh-0 | 1810.04160-zh-0.jpg | push0g0G
pop
$\bar{\text{用于传感器融合的优化门控深度学习架构}}$
Myung Seok Shim $\&$ Peng Li *
Department of Electrical and Computer Engineering
Texas A & M University
College Station, TX 77843, USA $\{$ mrshim1101,pli $\}$ @tamu.edu
Abstract
push0 g 0 G pop Sensor fusion is a key technology that integrates various sen-
sory inputs to allow for robust decision making in many applications such as autonomous driving and robot control. Deep neural networks have been adopted for sensor fusion in a body of recent studies. Among these, the so-called netgated architecture was proposed, which has demonstrated improved performances over the conventional convolutional neural networks (CNN). In this paper, we address several limitations of the baseline negated architecture by proposing two further optimized architectures: a coarsergrained gated architecture employing (feature) group-level fusion weights and a two-stage gated architectures leveraging both the group-level and feature-level fusion weights. Using driving mode prediction and human activity recognition datasets, we demonstrate the significant performance $\operatorname*{improvements}$brought by the proposed gated architectures and also their robustness in the presence of sensor noise and failures.
### 1 Introduction
传感器融合是自动驾驶汽车和移动机器人等自主系统的关键技术。在高级驾驶辅助系统(ADAS)中，许多传感器(如摄像头、超声波传感器和 LiDAR)被用于增强安全性和驾驶体验。基于传感器融合的车辆和机器人控制技术已被探索(Vargas-Meléndez et al., 2016; Jain et al., 2016; Garcia et al., 2017; Bohez et al., 2017; Patel et al., 2017)。此外，智能手机和智能手表等设备通常集成了许多传感器，使这些设备成为运行传感器融合应用程序(如活动识别)的合适平台。提出了几种用于活动识别的传感器融合技术 (Yurtman $\&$ Barshan, 2017; Dehzangi et al., 2017; Zhao & Zhou, 2017; Gravina et al., 2017)
更具体地说，在 Bohez et al. (2017) 中，讨论了一种基于深度强化学习的传感器融合算法， 用于机器人控制。KukaYouBot 与多个 LiDAR 传感器一起用于模拟。在传感器融合方面， 将所有输人连接为特征平面的早期融合与其他三种融合技术进行了比较：连接卷积层输出， 通过 1x1 卷积减少连接特征以及累积卷积输出。但是，在这项工作中不考虑传感器噪声和故障。此外，由于在实验中仅使用相同类型的感觉输人，因此基于不同类型的感觉输人的传感器融合的性能尚不清楚
在汽车中采用的传感器融合技术中，Vargas-Meléndez et al.(2016) 利用带有卡尔曼滤波器的神经网络进行车辆侧倾角估计，并展示了使用惯性测量单元(MU)而无需额外的悬架偏转传感器的优势。Jain et al. (2016) 考虑具有用于机动预测的学习算法的传感器丰富的平台。长短期记忆(LSTM)网络是一种递归神经网络(RNN),与来自相机，GPS 和速度计的感官输人一起使用。Garcia et al. (2017) 提出了用于道路环境的联合概率数据融合。但是，既不考虑多种感官输人，也不考虑感官噪声和故障。所采用的架构过于简单，输入数据仅在一层中融合
在可穿戴设备领域，Yurtman & Barshan (2017) 利用早期融合，将感官输入连接起来。通过这种简单的融合方法，比较了经典的监督学习方法，例如贝叶斯分类器，k 最近邻，支持向量机和人工神经网络。Dehzangi et al.(2017) 将深度卷积神经网络用于 IMU 数据。Zhao $\&$ Zhou (2017) 使用带有角度嵌入门动态图像的 CNN, 这些图像是用于步态识别的预处理输人。Gravina et al. (2017) 总结了三种融合方法，即数据、特征和决策级融合。然而，用于应对传感器故障的有效传感器融合网络体系结构尚未得到深人研究
在传感器融合架构方面，Patel et al. (2017) 提出了一种所谓的网络架构，其中给定卷积神经网络$g(0$NON) 中的信息流由从摄像机和 LiDAR 输入中提取的融合权重门控。这些融合权重用于计算感官输入的加权和。加权和通过完全连接的层来创建转向命令。与基本 CNN 相比，门控网络(网络)对传感器故障具有鲁棒性。然而，对感官输人、融合权重、网络架构
pop | [
[
"0.1772",
"0.03915",
"0.2688",
"0.03726",
"0.2688",
"0.05176",
"0.1772",
"0.05365"
],
[
"0.915",
"0.0404",
"0.9478",
"0.0404",
"0.9478",
"0.05618",
"0.915",
"0.05618"
],
[
"0.183",
"0.1111",
"0.1936",
"0.1111",
"0.1936",
"0.11554",
"0.183",
"0.11554"
],
[
"0.1805",
"0.11743",
"0.6807",
"0.11743",
"0.6807",
"0.1383",
"0.1805",
"0.1383"
],
[
"0.1847",
"0.1666",
"0.406",
"0.1654",
"0.406",
"0.1799",
"0.1847",
"0.1812"
],
[
"0.183",
"0.1799",
"0.562",
"0.1812",
"0.562",
"0.1957",
"0.183",
"0.1945"
],
[
"0.1847",
"0.1945",
"0.361",
"0.1945",
"0.361",
"0.2084",
"0.1847",
"0.2084"
],
[
"0.1847",
"0.2084",
"0.4182",
"0.207",
"0.4182",
"0.2216",
"0.1847",
"0.2229"
],
[
"0.1896",
"0.2216",
"0.4126",
"0.2216",
"0.4126",
"0.2361",
"0.1896",
"0.2361"
],
[
"0.464",
"0.2727",
"0.536",
"0.2727",
"0.536",
"0.2874",
"0.464",
"0.2874"
],
[
"0.1978",
"0.3018",
"0.761",
"0.3018",
"0.761",
"0.315",
"0.1978",
"0.315"
],
[
"0.2336",
"0.3164",
"0.7656",
"0.3164",
"0.7656",
"0.33",
"0.2336",
"0.33"
],
[
"0.2345",
"0.33",
"0.7646",
"0.33",
"0.7646",
"0.3435",
"0.2345",
"0.3435"
],
[
"0.2336",
"0.3423",
"0.7646",
"0.3423",
"0.7646",
"0.356",
"0.2336",
"0.356"
],
[
"0.2336",
"0.356",
"0.761",
"0.3567",
"0.761",
"0.3706",
"0.2336",
"0.3694"
],
[
"0.2361",
"0.3726",
"0.763",
"0.3726",
"0.763",
"0.3833",
"0.2361",
"0.3833"
],
[
"0.2361",
"0.385",
"0.7656",
"0.385",
"0.7656",
"0.3984",
"0.2361",
"0.3984"
],
[
"0.2345",
"0.3997",
"0.761",
"0.3997",
"0.761",
"0.4136",
"0.2345",
"0.4136"
],
[
"0.2345",
"0.4136",
"0.7646",
"0.4136",
"0.7646",
"0.4268",
"0.2345",
"0.4268"
],
[
"0.2336",
"0.4255",
"0.763",
"0.4255",
"0.763",
"0.4395",
"0.2336",
"0.4395"
],
[
"0.2336",
"0.44",
"0.7646",
"0.44",
"0.7646",
"0.4539",
"0.2336",
"0.4539"
],
[
"0.2345",
"0.4539",
"0.7646",
"0.4539",
"0.7646",
"0.4673",
"0.2345",
"0.4673"
],
[
"0.2345",
"0.4697",
"0.761",
"0.4697",
"0.761",
"0.4797",
"0.2345",
"0.4797"
],
[
"0.2336",
"0.4817",
"0.621",
"0.4817",
"0.621",
"0.4956",
"0.2336",
"0.4956"
],
[
"0.1708",
"0.5166",
"0.3162",
"0.5176",
"0.3162",
"0.532",
"0.1708",
"0.531"
],
[
"0.1765",
"0.5474",
"0.8228",
"0.5474",
"0.8228",
"0.561",
"0.1765",
"0.561"
],
[
"0.1772",
"0.5625",
"0.8203",
"0.5625",
"0.8203",
"0.5728",
"0.1772",
"0.5728"
],
[
"0.1748",
"0.5747",
"0.8203",
"0.5747",
"0.8203",
"0.5884",
"0.1748",
"0.5884"
],
[
"0.1748",
"0.5884",
"0.822",
"0.5884",
"0.822",
"0.6016",
"0.1748",
"0.6016"
],
[
"0.1765",
"0.603",
"0.822",
"0.603",
"0.822",
"0.616",
"0.1765",
"0.616"
],
[
"0.1765",
"0.617",
"0.822",
"0.617",
"0.822",
"0.631",
"0.1765",
"0.631"
],
[
"0.1748",
"0.631",
"0.6733",
"0.631",
"0.6733",
"0.6445",
"0.1748",
"0.6445"
],
[
"0.179",
"0.654",
"0.8203",
"0.654",
"0.8203",
"0.664",
"0.179",
"0.664"
],
[
"0.1765",
"0.6655",
"0.8203",
"0.6655",
"0.8203",
"0.6787",
"0.1765",
"0.6787"
],
[
"0.1772",
"0.681",
"0.819",
"0.681",
"0.819",
"0.6914",
"0.1772",
"0.6914"
],
[
"0.1765",
"0.6934",
"0.8228",
"0.6934",
"0.8228",
"0.707",
"0.1765",
"0.707"
],
[
"0.1765",
"0.7085",
"0.822",
"0.7085",
"0.822",
"0.7217",
"0.1765",
"0.7217"
],
[
"0.1748",
"0.72",
"0.3545",
"0.72",
"0.3545",
"0.7344",
"0.1748",
"0.7344"
],
[
"0.1765",
"0.742",
"0.8228",
"0.742",
"0.8228",
"0.756",
"0.1765",
"0.756"
],
[
"0.1765",
"0.756",
"0.8228",
"0.756",
"0.8228",
"0.7695",
"0.1765",
"0.7695"
],
[
"0.1765",
"0.771",
"0.8228",
"0.771",
"0.8228",
"0.784",
"0.1765",
"0.784"
],
[
"0.1772",
"0.784",
"0.8228",
"0.784",
"0.8228",
"0.7974",
"0.1772",
"0.7974"
],
[
"0.1748",
"0.799",
"0.822",
"0.799",
"0.822",
"0.8125",
"0.1748",
"0.8125"
],
[
"0.1772",
"0.8125",
"0.8228",
"0.8125",
"0.8228",
"0.8257",
"0.1772",
"0.8257"
],
[
"0.1772",
"0.8276",
"0.2893",
"0.8276",
"0.2893",
"0.8384",
"0.1772",
"0.8384"
],
[
"0.1765",
"0.8467",
"0.8228",
"0.8467",
"0.8228",
"0.86",
"0.1765",
"0.86"
],
[
"0.1772",
"0.8623",
"0.8203",
"0.8623",
"0.8203",
"0.8726",
"0.1772",
"0.8726"
],
[
"0.1772",
"0.875",
"0.822",
"0.875",
"0.822",
"0.888",
"0.1772",
"0.888"
],
[
"0.1748",
"0.888",
"0.8228",
"0.888",
"0.8228",
"0.9014",
"0.1748",
"0.9014"
],
[
"0.1765",
"0.904",
"0.822",
"0.904",
"0.822",
"0.914",
"0.1765",
"0.914"
],
[
"0.1765",
"0.9155",
"0.664",
"0.9165",
"0.664",
"0.9297",
"0.1765",
"0.9287"
],
[
"0.1772",
"0.9375",
"0.8228",
"0.9375",
"0.8228",
"0.9507",
"0.1772",
"0.9507"
],
[
"0.1748",
"0.9497",
"0.8228",
"0.9507",
"0.8228",
"0.9653",
"0.1748",
"0.964"
],
[
"0.9165",
"0.9526",
"0.9478",
"0.9526",
"0.9478",
"0.964",
"0.9165",
"0.964"
],
[
"0.1772",
"0.9653",
"0.8228",
"0.9653",
"0.8228",
"0.979",
"0.1772",
"0.979"
],
[
"0.1772",
"0.979",
"0.8228",
"0.979",
"0.8228",
"0.992",
"0.1772",
"0.992"
]
] | [
"pusho g 0 G",
"pop",
"",
"用于传感器融合的优化门控深度学习架构",
"Myung Seok Shim & Peng Li *",
"Department of Electrical and Computer Engineering",
"Texas A & M University",
"College Station, TX 77843, USA",
"[ mrshim1101,pli ) @tamu.edu",
"Abstract",
"pusho g O G pop Sensor fusion is a key technology that integrates various sen",
"sory inputs to allow for robust decision making in many applications such",
"as autonomous driving and robot control. Deep neural networks have been",
"adopted for sensor fusion in a body of recent studies. Among these, the",
"so-called netgated architecture was proposed, which has demonstrated im",
"provedperformancesoverthe conventionalconvolutionalneuralnetworks",
"(CNN). In this paper, we address several limitations of the baseline negated",
"architecture by proposing two further optimized architectures: a coarser.",
"grained gated architecture employing (feature) group-level fusion weights",
"and a two-stage gated architectures leveraging both the group-level and",
"feature-level fusion weights. Using driving mode prediction and human",
"activity recognition datasets, we demonstrate the significant performance",
"improvementsbroughtbytheproposedgatedarchitecturesandalsotheii",
"robustness in the presence of sensor noise and failures,",
"1Introduction",
"传感器融合是自动驾驶汽车和移动机器人等自主系统的关键技术。在高级驾驶辅助系统",
"ADAS）中，许多传感器（如摄像头、超声波传感器和LiDAR）被用于增强安全性和驾驶体",
"验。基于传感器融合的车辆和机器人控制技术已被探索（Vargas-Melendezetal.，2016；Jair",
"et al.,2016:Garcia et al.,2017:Bohez et al.,2017;Patel et al.,2017)。此外，智能手机和",
"智能手表等设备通常集成了许多传感器，使这些设备成为运行传感器融合应用程序（如活",
"动识别）的合适平台。提出了几种用于活动识别的传感器融合技术（Yurtman&Barshan，",
"2017; Dehzangi et al., 2017; Zhao & Zhou, 2017; Gravina et al., 2017)",
"更具体地说，在Bohezetal.（2017）中，讨论了一种基于深度强化学习的传感器融合算法",
"用于机器人控制。KukaYouBot与多个LiDAR传感器一起用于模拟。在传感器融合方面",
"将所有输入连接为特征平面的早期融合与其他三种融合技术进行了比较：连接卷积层输出",
"通过1x1卷积减少连接特征以及累积卷积输出。但是，在这项工作中不考虑传感器噪声和",
"故障。此外，由于在实验中仅使用相同类型的感觉输人，因此基于不同类型的感觉输入的传",
"感器融合的性能尚不清楚",
"在汽车中采用的传感器融合技术中，Vargas-Melendezetal.（2016）利用带有卡尔曼滤波器",
"的神经网络进行车辆侧倾角估计，并展示了使用惯性测量单元（IMU）而无需额外的悬架",
"偏转传感器的优势。Jainetal.（2016）考虑具有用于机动预测的学习算法的传感器丰富的平",
"台。长短期记忆（LSTM）网络是一种递归神经网络（RNN），与来自相机，GPS和速度计",
"的感官输人一起使用。Garciaetal.（2017）提出了用于道路环境的联合概率数据融合。但",
"是，既不考虑多种感官输人，也不考虑感官噪声和故障。所采用的架构过于简单，输入数据",
"仅在一层中融合",
"在可穿戴设备领域，Yurtman&Barshan（2017）利用早期融合，将感官输人连接起来。通",
"过这种简单的融合方法：比较了经典的监督学习方法，例如贝叶斯分类器，k最近邻，支持",
"向量机和人工神经网络。Dehzangietal.（2017）将深度卷积神经网络用于IMU数据。Zhao",
"&Zhou（2017）使用带有角度嵌入门动态图像的CNN，这些图像是用于步态识别的预处理",
"输入。Gravinaetal.（2017）总结了三种融合方法：即数据、特征和决策级融合。然而，用",
"于应对传感器故障的有效传感器融合网络体系结构尚未得到深入研究",
"在传感器融合架构方面，Pateletal.（2017）提出了一种所谓的网络架构，其中给定卷积神",
"经网络gN）中的信息流由从摄像机和LiDAR输人中提取的融合权重门控。这些融合权",
"pop",
"重用于计算感官输人的加权和。加权和通过完全连接的层来创建转向命令。与基本CNN相",
"比，门控网络（网络）对传感器故障具有鲁棒性。然而，对感官输入、融合权重、网络架构"
] | {
"arxiv_id": "1810.04160",
"language": "zh",
"page_idx": 0
} |
||
1810.04160-zh-1 | 1810.04160-zh-1.jpg | push0g0G
pop
此架构的更通用版本(具有五个传感器/功能)如图。??。在 Patel et al.(2017)中，来自两个感官输入的数据，即。e。通过卷积 (conv)层，池化层和完全连接层分别处理相机和LiDAR。来自全连接(FC)层的输出(e。g。图中第一个虚线框中的“FC-f1”到“FC-f5”。?? ),连接在一起，然后由另一个$\operatorname{FC}($e。g。图中的“FC-con”。”?),其中创建了两个特征级融合权重。特征级融合权重最初在 Patel et al.(2017)中称为标量。请注意，每个融合权重是一个标量值，并且在图。?? 提取了五个特征级融合权重，它们是“FC-con”层的输出。这些融合权重与来自特征级 FC 层的相应输出相乘，即。e。第一个虚线框，为了更好地说明图中的数据流，将其复制。??。最后，这些加权特征输出由最后一个 FC 层 (i。e。“FC-out”),它产生最终的预测决策
否定的架构很有趣，因为提取的特征级权重可以在概念上被认为是每个特征输人的“门控” 变量，并且当传感器输人(特征)被噪声或传感器故障破坏时，可以通过零值融合权重将其从网络中关闭。因此，这种架构可能有希望提供强大的传感器融合能力
2.2 Limitations of the Basic Netgated Architecture
否定的架构提供了一个吸引人的端到端深度学习解决方案。然而，部分由于其端到端黑盒
性质，这种架构有几个限制，如下所述
融合权重不一致。首先，考虑有$N$输人特征$f_1,f_2,\cdots,f_N$和相应的特征级融合权值$fw_1,fw_2,\cdots,fw_N$ 的情况。如图。??,特征级融合权重由“FC-con”层基于来自所有输人的融合信息产生。因此，由于所有特征之间的信息共享，提取的融合权重可能不完全对应于相应的特征。正如我们在实验研究中所观察到的那样，在某些情况下，具有最大提取融合权重的特征并不代表学习任务的最关键特征。虽然特征级权重的排名可能在一定程度上反映了相应特征的相对重要性，但两者之间的关联并不总是一致的。在本文中，我们将这种现象称为融合权重的不一致性。可以很好地预期，融合权重的不一致性可能会对整体预测精度产生不利影响，特别是当网络没有鲁棒地学习某些噪声或损坏特征的融合权重时，会导致误导性的大融合权重值
潜在过拟合。此外，对于需要融合许多特征的应用，使用相同数量的融合权重值引人了许多额外的参数，这些参数应在训练过程中正确学习，从而使模型的过度拟合更容易发生。由于可能发生融合权重不一致，这种情况进一步加剧
缺乏额外的融合机制。最后，在图的架构中。??,除了学习融合权重之外，原始输入特征的融合是以一种简单的方式完成的，i。e。由最后一个完全连接的层“FC-out”。然而，存在更强大的原始输入融合机制，这可能会导致额外的性能改进
我们通过提出两个扩展来解决基准网络体系结构的上述局限性：较粗粒度的体系结构和分层的两阶段体系结构，分别称为特征组门控融合体系结构 (FG-GFA)和两阶段门控融合体系结构(2S-GFA),在以下各节中进行了描述
3 Feature-Group Gated Fusion Architecture (FG-GFA)
为了解决基线网络体系结构的上述限制，我们首先探索粗粒度的体系结构，即特征组门控融合体系结构( FG-GFA),如图。E??,其中出于说明目的，示出了两个特征组，第一组具有三个特征，第二组具有两个特征。通常，可以将给定的一组$N$输人特征$f_1,f_2,\cdots,f_N$划分为例如$M$的特征组$FG_1,FG_2,\cdots,FG_M$ 。作为此体系结构的一个特定示例，首先将特征组中的所有特征连接在一起，然后将其传递到卷积层和池化层上。经过相应的 FC 层 (图1中的“FC-g1”或“FC-g2”。”? ),来自所有组的处理后的信息被连接，然后传递到 FC 层(图。?? ), 其输出被拆分为$M$组级融合权重。每组 e 的融合信息。g。图中的“FC-g1” 或“FC-g2”。”? 乘以相应的组级融合权重(“FC-g1”或“FC-g2”在图中再次复制。”? 以更好地说明信息流)。所有加权的组级信息被组合，然后由最终的 FC 层处理 (图 3 中的“FC-out”。?? )产生预测决策。FG-GFA 的每个关键步骤处的配置和层数可以与图 2 的具体示例不同地选择。??
我们现在评论 FG-GFA 体系结构和基准网络体系结构之间的主要区别。首先，除了图 1 中“FC-out”的最终融合操作。?? , 我们在每组中进行了额外的早期感觉输入融合。这种组内融合的输出被组合以产生较少数量的组级融合权重。此外，提取的组级别权重用于乘以相应的融合组特征信息，而不是单个特征信息。FG-GFA 的这些特性将不同类型的融合机制引人到网络中。其次，由于现在仅在组级别提取融合权重，因此与基线架构相比，需要学习的权重更少。现在调整参数的数量较少，并且每组中都进行了早期融合，这一事实可能会
2
push0g0G
pop | [
[
"0.1772",
"0.03915",
"0.2688",
"0.03726",
"0.2688",
"0.05176",
"0.1772",
"0.05365"
],
[
"0.915",
"0.0385",
"0.949",
"0.04166",
"0.947",
"0.05618",
"0.9126",
"0.05304"
],
[
"0.1765",
"0.1061",
"0.8203",
"0.1061",
"0.8203",
"0.1193",
"0.1765",
"0.1193"
],
[
"0.1772",
"0.1193",
"0.822",
"0.1193",
"0.822",
"0.1332",
"0.1772",
"0.1332"
],
[
"0.1765",
"0.1332",
"0.822",
"0.1332",
"0.822",
"0.1477",
"0.1765",
"0.1477"
],
[
"0.1765",
"0.1465",
"0.822",
"0.1465",
"0.822",
"0.1604",
"0.1765",
"0.1604"
],
[
"0.1765",
"0.1604",
"0.822",
"0.1604",
"0.822",
"0.1736",
"0.1765",
"0.1736"
],
[
"0.1772",
"0.1749",
"0.822",
"0.1749",
"0.822",
"0.1881",
"0.1772",
"0.1881"
],
[
"0.1765",
"0.1875",
"0.822",
"0.1875",
"0.822",
"0.2008",
"0.1765",
"0.2008"
],
[
"0.1765",
"0.202",
"0.8228",
"0.202",
"0.8228",
"0.2153",
"0.1765",
"0.2153"
],
[
"0.1765",
"0.2166",
"0.4631",
"0.2166",
"0.4631",
"0.2299",
"0.1765",
"0.2299"
],
[
"0.1772",
"0.2393",
"0.8203",
"0.2393",
"0.8203",
"0.25",
"0.1772",
"0.25"
],
[
"0.1748",
"0.25",
"0.8228",
"0.2507",
"0.8228",
"0.2651",
"0.1748",
"0.2644"
],
[
"0.1765",
"0.2651",
"0.6807",
"0.2651",
"0.6807",
"0.279",
"0.1765",
"0.279"
],
[
"0.1748",
"0.2966",
"0.5557",
"0.2966",
"0.5557",
"0.31",
"0.1748",
"0.31"
],
[
"0.1765",
"0.3225",
"0.8228",
"0.3225",
"0.8228",
"0.336",
"0.1765",
"0.336"
],
[
"0.1765",
"0.3372",
"0.455",
"0.3372",
"0.455",
"0.351",
"0.1765",
"0.351"
],
[
"0.1765",
"0.358",
"0.8228",
"0.358",
"0.8228",
"0.3718",
"0.1765",
"0.3718"
],
[
"0.1748",
"0.3718",
"0.822",
"0.3706",
"0.822",
"0.385",
"0.1748",
"0.3865"
],
[
"0.1765",
"0.3838",
"0.822",
"0.3838",
"0.822",
"0.3984",
"0.1765",
"0.3984"
],
[
"0.1748",
"0.3997",
"0.8203",
"0.3997",
"0.8203",
"0.4136",
"0.1748",
"0.4136"
],
[
"0.1765",
"0.4136",
"0.822",
"0.4136",
"0.822",
"0.4268",
"0.1765",
"0.4268"
],
[
"0.1765",
"0.4268",
"0.822",
"0.4268",
"0.822",
"0.44",
"0.1765",
"0.44"
],
[
"0.1748",
"0.44",
"0.822",
"0.4414",
"0.822",
"0.455",
"0.1748",
"0.4539"
],
[
"0.1765",
"0.4539",
"0.822",
"0.455",
"0.822",
"0.4697",
"0.1765",
"0.4685"
],
[
"0.1765",
"0.4685",
"0.3555",
"0.4685",
"0.3555",
"0.4817",
"0.1765",
"0.4817"
],
[
"0.1765",
"0.4905",
"0.822",
"0.4905",
"0.822",
"0.504",
"0.1765",
"0.504"
],
[
"0.1748",
"0.504",
"0.8203",
"0.504",
"0.8203",
"0.5176",
"0.1748",
"0.5176"
],
[
"0.179",
"0.5195",
"0.5186",
"0.5195",
"0.5186",
"0.53",
"0.179",
"0.53"
],
[
"0.1765",
"0.5386",
"0.8203",
"0.5386",
"0.8203",
"0.552",
"0.1765",
"0.552"
],
[
"0.1765",
"0.552",
"0.822",
"0.552",
"0.822",
"0.5664",
"0.1765",
"0.5664"
],
[
"0.1772",
"0.5674",
"0.6167",
"0.5674",
"0.6167",
"0.5776",
"0.1772",
"0.5776"
],
[
"0.1772",
"0.5884",
"0.822",
"0.5884",
"0.822",
"0.6016",
"0.1772",
"0.6016"
],
[
"0.1772",
"0.6016",
"0.822",
"0.6016",
"0.822",
"0.6157",
"0.1772",
"0.6157"
],
[
"0.1772",
"0.617",
"0.501",
"0.617",
"0.501",
"0.6274",
"0.1772",
"0.6274"
],
[
"0.1732",
"0.6514",
"0.6562",
"0.6514",
"0.6562",
"0.6655",
"0.1732",
"0.6655"
],
[
"0.1765",
"0.6816",
"0.8228",
"0.6816",
"0.8228",
"0.696",
"0.1765",
"0.696"
],
[
"0.1772",
"0.696",
"0.822",
"0.696",
"0.822",
"0.709",
"0.1772",
"0.709"
],
[
"0.1765",
"0.7104",
"0.8228",
"0.7104",
"0.8228",
"0.7236",
"0.1765",
"0.7236"
],
[
"0.1765",
"0.7236",
"0.822",
"0.7236",
"0.822",
"0.7373",
"0.1765",
"0.7373"
],
[
"0.1765",
"0.7373",
"0.822",
"0.7363",
"0.822",
"0.7505",
"0.1765",
"0.752"
],
[
"0.1748",
"0.75",
"0.8228",
"0.7505",
"0.8228",
"0.765",
"0.1748",
"0.7646"
],
[
"0.1748",
"0.7646",
"0.822",
"0.765",
"0.822",
"0.78",
"0.1748",
"0.779"
],
[
"0.1748",
"0.778",
"0.8203",
"0.778",
"0.8203",
"0.7915",
"0.1748",
"0.7915"
],
[
"0.1772",
"0.7925",
"0.822",
"0.7925",
"0.822",
"0.806",
"0.1772",
"0.806"
],
[
"0.1765",
"0.806",
"0.8203",
"0.806",
"0.8203",
"0.8193",
"0.1765",
"0.8193"
],
[
"0.1772",
"0.8213",
"0.357",
"0.8213",
"0.357",
"0.832",
"0.1772",
"0.832"
],
[
"0.1772",
"0.8433",
"0.822",
"0.8433",
"0.822",
"0.854",
"0.1772",
"0.854"
],
[
"0.1765",
"0.854",
"0.822",
"0.855",
"0.822",
"0.8687",
"0.1765",
"0.867"
],
[
"0.1772",
"0.869",
"0.822",
"0.869",
"0.822",
"0.8833",
"0.1772",
"0.8833"
],
[
"0.179",
"0.8853",
"0.8203",
"0.8853",
"0.8203",
"0.896",
"0.179",
"0.896"
],
[
"0.1765",
"0.898",
"0.8203",
"0.898",
"0.8203",
"0.911",
"0.1765",
"0.911"
],
[
"0.1772",
"0.911",
"0.822",
"0.911",
"0.822",
"0.925",
"0.1772",
"0.925"
],
[
"0.1772",
"0.9487",
"0.2688",
"0.9463",
"0.2688",
"0.961",
"0.1772",
"0.9634"
],
[
"0.585",
"0.9497",
"0.597",
"0.9497",
"0.597",
"0.961",
"0.585",
"0.961"
],
[
"0.9165",
"0.952",
"0.947",
"0.952",
"0.947",
"0.9634",
"0.9165",
"0.9634"
]
] | [
"pusho g 0 G",
"pop",
"此架构的更通用版本（具有五个传感器/功能）如图。？？。在Pateletal.（2017）中，来自",
"两个感官输人的数据，即。e。通过卷积（conv）层，池化层和完全连接层分别处理相机和",
"LiDAR。来自全连接(FC）层的输出（e。g。图中第一个虚线框中的“FC-f1”到“FC-f5\"",
"?？），连接在一起，然后由另一个FC层（e。g。图中的“FC-con”。?？），其中创建了两个",
"特征级融合权重。特征级融合权重最初在Pateletal.（2017）中称为标量。请注意，每个融",
"合权重是一个标量值，并且在图。?？提取了五个特征级融合权重，它们是“FC-con”层的",
"输出。这些融合权重与来自特征级FC层的相应输出相乘，即。e。第一个虚线框，为了更",
"好地说明图中的数据流，将其复制。？？。最后，这些加权特征输出由最后一个FC层（i。",
"e。“FC-out”），它产生最终的预测决策",
"否定的架构很有趣，因为提取的特征级权重可以在概念上被认为是每个特征输入的“门控",
"变量，并且当传感器输入（特征）被噪声或传感器故障破坏时，可以通过零值融合权重将其",
"从网络中关闭。因此，这种架构可能有希望提供强大的传感器融合能力",
"2.2Limitations of the Basic Netgated Architecture",
"否定的架构提供了一个吸引人的端到端深度学习解决方案。然而，部分由于其端到端黑盒",
"性质，这种架构有几个限制，如下所述",
"融合权重不一致。首先，考虑有N输入特征f1,f2，，f和相应的特征级融合权值",
"fw1,fw2，.，fwN的情况。如图。??，特征级融合权重由“FC-con”层基于来自所有输",
"人的融合信息产生。因此，由于所有特征之间的信息共享，提取的融合权重可能不完全对应",
"于相应的特征。正如我们在实验研究中所观察到的那样，在某些情况下，具有最大提取融合",
"权重的特征并不代表学习任务的最关键特征。虽然特征级权重的排名可能在一定程度上反",
"映了相应特征的相对重要性，但两者之间的关联并不总是一致的。在本文中，我们将这种现",
"象称为融合权重的不一致性。可以很好地预期，融合权重的不一致性可能会对整体预测精",
"度产生不利影响，特别是当网络没有鲁棒地学习某些噪声或损坏特征的融合权重时，会导",
"致误导性的大融合权重值",
"潜在过拟合。此外，对于需要融合许多特征的应用，使用相同数量的融合权重值引人了许多",
"额外的参数，这些参数应在训练过程中正确学习，从而使模型的过度拟合更容易发生。由于",
"可能发生融合权重不一致，这种情况进一步加剧",
"缺乏额外的融合机制。最后，在图的架构中。？？，除了学习融合权重之外，原始输入特征",
"的融合是以一种简单的方式完成的，i。e。由最后一个完全连接的层“FC-out”。然而，存",
"在更强大的原始输入融合机制，这可能会导致额外的性能改进",
"我们通过提出两个扩展来解决基准网络体系结构的上述局限性：较粗粒度的体系结构和分",
"层的两阶段体系结构，分别称为特征组门控融合体系结构（FG-GFA）和两阶段门控融合体",
"系结构（2S-GFA）在以下各节中进行了描述",
"3Feature-Group Gated Fusion Architecture (FG-GFA)",
"为了解决基线网络体系结构的上述限制，我们首先探索粗粒度的体系结构，即特征组门控",
"融合体系结构（FG-GFA），如图。E？？，其中出于说明目的，示出了两个特征组，第一组具",
"有三个特征，第二组具有两个特征。通常，可以将给定的一组N输人特征fi，f2，··，fN划",
"分为例如M的特征组FG1,FG2，··，FGM。作为此体系结构的一个特定示例，首先将特",
"征组中的所有特征连接在一起，然后将其传递到卷积层和池化层上。经过相应的FC层（图",
"1中的“FC-g1”或“FC-g2”。？？），来自所有组的处理后的信息被连接，然后传递到FC",
"层（图。??），其输出被拆分为M组级融合权重。每组e的融合信息。g。图中的“FC-g1",
"或“FC-g2\"。”?？乘以相应的组级融合权重（“FC-g1”或“FC-g2”在图中再次复制。？",
"以更好地说明信息流)。所有加权的组级信息被组合，然后由最终的FC层处理（图3中的",
"“FC-out”。？？）产生预测决策。FG-GFA的每个关键步骤处的配置和层数可以与图2的具",
"体示例不同地选择。。？？",
"我们现在评论FG-GFA体系结构和基准网络体系结构之间的主要区别。首先，除了图1中",
"“FC-out”的最终融合操作。?？，我们在每组中进行了额外的早期感觉输人融合。这种组",
"内融合的输出被组合以产生较少数量的组级融合权重。此外，提取的组级别权重用于乘以",
"相应的融合组特征信息，而不是单个特征信息。FG-GFA的这些特性将不同类型的融合机",
"制引人到网络中。其次，由于现在仅在组级别提取融合权重，因此与基线架构相比，需要学",
"习的权重更少。现在调整参数的数量较少，并且每组中都进行了早期融合，这一事实可能会",
"pusho g 0 G",
"2",
"pop"
] | {
"arxiv_id": "1810.04160",
"language": "zh",
"page_idx": 1
} |
||
1810.04160-zh-2 | 1810.04160-zh-2.jpg | push0g0G
pop
降低将训练过程卡在局部最小值上的可能性。因此，有可能减轻融合权重不一致和潜在过度拟合的问题。正如将在实验研究中证明的那样，FG-GFA 可以在组级别上对融合权重进行更强大的学习，即。e。组中噪声或损坏特征的存在可以更可靠地反映在组级融合权重的相应降低中。结果，组级别的融合权重成为每组中感官输人质量的更可靠指标。因此，我们根据经验观察到 FG-GFA 带来的性能改进，如后面所证明的
4 The Proposed Two-Stage Gated Fusion Architecture (2S-GFA
在此分层融合体系结构中，我们将倾斜特征级融合权重的基线网络体系结构与拟议的特征组门控融合体系结构 (FG-GFA) 相结合，该体系结构将组级融合权重提取为两个阶段。图2中展示了 2S-GFA 架构。?? ,其中前三个特征在组 1 中，其余两个特征在组 2 中。网络的上部基于拆分 FC 层 “FC-con” 的输出来提取五个特征级融合权重，其方式与基线网络架构中的方式相同。图底部较小的子网。$\mathbb{E}??$重用来自图顶部的 conv 层的第一阶段的输出， 这些层分别对所有感官输入进行预处理。然后将预处理后的特征信息连接到每个组内。它通过拆分 FC 层“FC-con-g”的输出来产生两个组级融合权重，分别由两个组的红色和黄色正方形显示。对于每个特征输入，其特征级融合权重与组级融合权重的乘积定义了其最终特征权重，该最终特征权重用于将处理后的特征信息相乘，e。g。图中的“FC-f1”至“FC-f5”。??。然后，所有加权的特征信息由 FC 层“FC-out”融合在一起，从而产生最终的决策
由于 2S-GFA 集成了基准网络体系结构和基于组的 FG-GFA 体系结构的基本要素，因此它在两种体系结构上都有所改进。再次注意，每个特征采用的最终融合权重是特征级权重和相应的组级权重的乘积。因此，最终的融合权重结合了从基于特征和基于组的处理中提取的关键信息。与 FG-GFA 架构一样，可以可靠地学习每个组级融合权重，因此，最终的融合权重可以更可靠地反映相应特征相对于手头学习任务的重要性，并为特征提供有效的门控机制。例如，正如我们在实验研究中观察到的那样，对于嘈杂或损坏的感官输人，特征级融合权重可能无法完全反映该特征的重要性下降，就像在基线架构的情况下一样。然而，更可靠的组级融合权重可以阻止此功能，即。e。通过使最终特征权重(特征级别和组级别权重的乘积)较小。此属性缓解了基线架构的融合权重不一致的问题
另一方面，与 FG-GFA 架构相比，2S-GFA 进一步利用了特征级融合权重所揭示的信息。因此，每个感官输人可以基于特征级融合权重以更精细的粒度进行门控。因此，可以预期2S-GFA 代表了基线网络体系结构和粗粒度 FG-GPA 体系结构之间的优化中间地带，并且它可以更有效地学习底层数据的结构，正如我们将通过实验演示的那样。
# 5 Experimental Settings
为了验证所提出的 FG-GFA 和 2S-GFA 架构，并将其与传统的非门控和基准网络架构进行
比较，我们考虑了两个应用：智能手机上的驾驶模型预测和人类活动识别
用于驾驶模型预测的
5.1 Datasets, setups for datasets, and tool flow
数据集。我们考虑三种驾驶模式：怠速、环保和正常。总速模式对应于车辆发动机开启但车辆不在运动中的情况。为了提高燃油效率而驾驶汽车时，车辆处于环保模式。所有其他情况都标记为正常模式。目标是在给定所有可用的感官输人的情况下，预测车辆在下一个时间段的驾驶模式。我们将此应用程序视为时间序列预测问题
我们已经在两个 GPS 坐标之间驾驶了 2014 年日产 Sentra 两个月，以收集此驾驶数据集。RPM 和速度数据从车辆的车载诊断 (OBD2) 中提取。我们使用陀螺仪的 y 轴读数(GYRO$\_$Y)来测量横向加速度，并使用 GPS 航向之间的时间差(D\_HEADING)来测量车辆的转向角。驾驶数据集中使用的所有传感器数据均从 FreematicsONE+dongle (Huang, 2017) 收集。由于 FreematicsONE+dongle 中传感器的提取时间周期不同，因此使用时间上的线性插值进行数据预处理。总共使用五种类型的感觉数据来训练神经网络：RPM, SPEED,加速和减速(D_SPEED),陀螺仪的y轴(GYRO_Y)和GPS航向的时间差(D\_HEADING)。这五个特征每 250ms 采样一次。为了预测车辆未来 5 秒的驾驶模式，我们使用来自每个特征的十秒数据(40 点),将收集的数据按特征进行归一化，并将它们连接起来以形成特征向量窗口。随着预测的进行，特征向量窗口每 250ms 滑动一次。对于建议的 FP-GFA 和 2S-GFA 体系结构，RPM, SPEED 和 D\_SPEED 包含在第一组中，而第二
3
push0g0G
pop | [
[
"0.1772",
"0.03915",
"0.2688",
"0.03726",
"0.2688",
"0.05176",
"0.1772",
"0.05365"
],
[
"0.915",
"0.0385",
"0.949",
"0.04166",
"0.947",
"0.05618",
"0.9126",
"0.05304"
],
[
"0.1765",
"0.1061",
"0.822",
"0.1061",
"0.822",
"0.1193",
"0.1765",
"0.1193"
],
[
"0.1772",
"0.1193",
"0.822",
"0.1193",
"0.822",
"0.1332",
"0.1772",
"0.1332"
],
[
"0.1772",
"0.1332",
"0.822",
"0.1332",
"0.822",
"0.1465",
"0.1772",
"0.1465"
],
[
"0.179",
"0.1483",
"0.822",
"0.1483",
"0.822",
"0.159",
"0.179",
"0.159"
],
[
"0.1765",
"0.161",
"0.6235",
"0.161",
"0.6235",
"0.1749",
"0.1765",
"0.1749"
],
[
"0.1748",
"0.1976",
"0.738",
"0.1976",
"0.738",
"0.2122",
"0.1748",
"0.2122"
],
[
"0.1765",
"0.2291",
"0.822",
"0.2291",
"0.822",
"0.2424",
"0.1765",
"0.2424"
],
[
"0.1748",
"0.2424",
"0.8203",
"0.2424",
"0.8203",
"0.2568",
"0.1748",
"0.2568"
],
[
"0.1765",
"0.2551",
"0.822",
"0.2551",
"0.822",
"0.2695",
"0.1765",
"0.2695"
],
[
"0.1748",
"0.2683",
"0.8203",
"0.2683",
"0.8203",
"0.2822",
"0.1748",
"0.2822"
],
[
"0.1765",
"0.2827",
"0.822",
"0.2842",
"0.822",
"0.2974",
"0.1765",
"0.2966"
],
[
"0.1765",
"0.2974",
"0.822",
"0.2974",
"0.822",
"0.3113",
"0.1765",
"0.3113"
],
[
"0.1765",
"0.3113",
"0.822",
"0.3113",
"0.822",
"0.3245",
"0.1765",
"0.3245"
],
[
"0.1772",
"0.3257",
"0.822",
"0.3257",
"0.822",
"0.339",
"0.1772",
"0.339"
],
[
"0.1732",
"0.3372",
"0.826",
"0.336",
"0.826",
"0.3542",
"0.1732",
"0.3547"
],
[
"0.1765",
"0.3528",
"0.8013",
"0.3528",
"0.8013",
"0.3662",
"0.1765",
"0.3662"
],
[
"0.1772",
"0.375",
"0.8228",
"0.375",
"0.8228",
"0.3882",
"0.1772",
"0.3882"
],
[
"0.1765",
"0.387",
"0.8203",
"0.387",
"0.8203",
"0.401",
"0.1765",
"0.401"
],
[
"0.1772",
"0.4016",
"0.822",
"0.4016",
"0.822",
"0.4153",
"0.1772",
"0.4153"
],
[
"0.1748",
"0.414",
"0.822",
"0.4153",
"0.822",
"0.43",
"0.1748",
"0.4287"
],
[
"0.1772",
"0.43",
"0.822",
"0.43",
"0.822",
"0.443",
"0.1772",
"0.443"
],
[
"0.1765",
"0.443",
"0.822",
"0.443",
"0.822",
"0.457",
"0.1765",
"0.457"
],
[
"0.1772",
"0.4583",
"0.822",
"0.4583",
"0.822",
"0.4717",
"0.1772",
"0.4717"
],
[
"0.179",
"0.4717",
"0.822",
"0.4717",
"0.822",
"0.4849",
"0.179",
"0.4849"
],
[
"0.1765",
"0.4849",
"0.648",
"0.4849",
"0.648",
"0.4988",
"0.1765",
"0.4988"
],
[
"0.179",
"0.5083",
"0.819",
"0.5083",
"0.819",
"0.518",
"0.179",
"0.518"
],
[
"0.1772",
"0.521",
"0.822",
"0.521",
"0.822",
"0.534",
"0.1772",
"0.534"
],
[
"0.1765",
"0.534",
"0.822",
"0.534",
"0.822",
"0.5474",
"0.1765",
"0.5474"
],
[
"0.1772",
"0.55",
"0.687",
"0.55",
"0.687",
"0.56",
"0.1772",
"0.56"
],
[
"0.1724",
"0.5835",
"0.3962",
"0.5854",
"0.3962",
"0.5996",
"0.1724",
"0.5977"
],
[
"0.1772",
"0.6157",
"0.822",
"0.6157",
"0.822",
"0.629",
"0.1772",
"0.629"
],
[
"0.1765",
"0.63",
"0.697",
"0.63",
"0.697",
"0.643",
"0.1765",
"0.643"
],
[
"0.1772",
"0.653",
"0.3218",
"0.653",
"0.3218",
"0.6636",
"0.1772",
"0.6636"
],
[
"0.1732",
"0.6816",
"0.524",
"0.6816",
"0.524",
"0.696",
"0.1732",
"0.696"
],
[
"0.1772",
"0.711",
"0.8203",
"0.711",
"0.8203",
"0.7217",
"0.1772",
"0.7217"
],
[
"0.1772",
"0.726",
"0.8203",
"0.726",
"0.8203",
"0.7363",
"0.1772",
"0.7363"
],
[
"0.1765",
"0.738",
"0.8228",
"0.738",
"0.8228",
"0.752",
"0.1765",
"0.752"
],
[
"0.1772",
"0.7524",
"0.581",
"0.7524",
"0.581",
"0.763",
"0.1772",
"0.763"
],
[
"0.1772",
"0.773",
"0.822",
"0.773",
"0.822",
"0.786",
"0.1772",
"0.786"
],
[
"0.1772",
"0.786",
"0.822",
"0.786",
"0.822",
"0.8",
"0.1772",
"0.8"
],
[
"0.1772",
"0.8",
"0.822",
"0.8",
"0.822",
"0.813",
"0.1772",
"0.813"
],
[
"0.1772",
"0.8145",
"0.819",
"0.8145",
"0.819",
"0.8276",
"0.1772",
"0.8276"
],
[
"0.1765",
"0.8276",
"0.822",
"0.8276",
"0.822",
"0.8413",
"0.1765",
"0.8413"
],
[
"0.1805",
"0.8433",
"0.8203",
"0.8433",
"0.8203",
"0.854",
"0.1805",
"0.854"
],
[
"0.1772",
"0.855",
"0.8228",
"0.855",
"0.8228",
"0.869",
"0.1772",
"0.869"
],
[
"0.1765",
"0.867",
"0.8228",
"0.8687",
"0.8228",
"0.8833",
"0.1765",
"0.882"
],
[
"0.1772",
"0.8853",
"0.819",
"0.8853",
"0.819",
"0.896",
"0.1772",
"0.896"
],
[
"0.1772",
"0.898",
"0.822",
"0.898",
"0.822",
"0.911",
"0.1772",
"0.911"
],
[
"0.1765",
"0.911",
"0.822",
"0.911",
"0.822",
"0.925",
"0.1765",
"0.925"
],
[
"0.1772",
"0.9487",
"0.2688",
"0.9463",
"0.2688",
"0.961",
"0.1772",
"0.9634"
],
[
"0.584",
"0.9497",
"0.597",
"0.9497",
"0.597",
"0.961",
"0.584",
"0.961"
],
[
"0.9165",
"0.952",
"0.947",
"0.952",
"0.947",
"0.9634",
"0.9165",
"0.9634"
]
] | [
"pusho g 0 G",
"pop",
"降低将训练过程卡在局部最小值上的可能性。因此，有可能减轻融合权重不一致和潜在过",
"度拟合的问题。正如将在实验研究中证明的那样，FG-GFA可以在组级别上对融合权重进",
"行更强大的学习，即。e。组中噪声或损坏特征的存在可以更可靠地反映在组级融合权重的",
"相应降低中。结果，组级别的融合权重成为每组中感官输入质量的更可靠指标。。因此，我",
"们根据经验观察到FG-GFA带来的性能改进，如后面所证明的",
"4The Proposed Two-Stage Gated Fusion Architecture (2S-GFA)",
"在此分层融合体系结构中，我们将倾斜特征级融合权重的基线网络体系结构与拟议的特征",
"组门控融合体系结构（FG-GFA）相结合，该体系结构将组级融合权重提取为两个阶段。图",
"2中展示了2S-GFA架构。？？，其中前三个特征在组1中，其余两个特征在组2中。网络",
"的上部基于拆分FC层“FC-con”的输出来提取五个特征级融合权重，其方式与基线网络",
"架构中的方式相同。图底部较小的子网。E?？重用来自图顶部的conv层的第一阶段的输出",
"这些层分别对所有感官输入进行预处理。然后将预处理后的特征信息连接到每个组内。它通",
"过拆分FC层“FC-con-g”的输出来产生两个组级融合权重，分别由两个组的红色和黄色正",
"方形显示。对于每个特征输入，其特征级融合权重与组级融合权重的乘积定义了其最终特征",
"权重，该最终特征权重用于将处理后的特征信息相乘，e。g。图中的“FC-f1”至“FC-f5\"。",
"?？。然后，所有加权的特征信息由FC层“FC-out”融合在一起，从而产生最终的决策",
"由于2S-GFA集成了基准网络体系结构和基于组的FG-GFA体系结构的基本要素，因此它",
"在两种体系结构上都有所改进。再次注意，每个特征采用的最终融合权重是特征级权重和",
"相应的组级权重的乘积。因此，最终的融合权重结合了从基于特征和基于组的处理中提取",
"的关键信息。与FG-GFA架构一样，可以可靠地学习每个组级融合权重，因此，最终的融",
"合权重可以更可靠地反映相应特征相对于手头学习任务的重要性，并为特征提供有效的门",
"控机制。例如，正如我们在实验研究中观察到的那样，对于杂或损坏的感官输人，特征级",
"融合权重可能无法完全反映该特征的重要性下降，就像在基线架构的情况下一样。然而，更",
"可靠的组级融合权重可以阻止此功能，即。e。通过使最终特征权重（特征级别和组级别权",
"重的乘积）较小。此属性缓解了基线架构的融合权重不一致的问题",
"另一方面，与FG-GFA架构相比，2S-GFA进一步利用了特征级融合权重所揭示的信息",
"因此，每个感官输入可以基于特征级融合权重以更精细的粒度进行门控。因此，可以预期",
"2S-GFA代表了基线网络体系结构和粗粒度FG-GPA体系结构之间的优化中间地带，并且",
"它可以更有效地学习底层数据的结构，正如我们将通过实验演示的那样。",
"5Experimental Settings",
"为了验证所提出的FG-GFA和2S-GFA架构，并将其与传统的非门控和基准网络架构进行",
"比较，我们考虑了两个应用：智能手机上的驾驶模型预测和人类活动识别",
"用手驾驶模型预测的",
"5.1Datasets, setups for datasets, and tool flow",
"数据集。我们考虑三种驾驶模式：怠速、环保和正常。怠速模式对应于车辆发动机开启但车",
"辆不在运动中的情况。为厂提高燃油效率而驾驶汽车时，车辆处于环保模式。所有其他情况",
"都标记为正常模式。目标是在给定所有可用的感官输人的情况下，预测车辆在下一个时间",
"段的驾驶模式。我们将此应用程序视为时间序列预测问题",
"我们已经在两个GPS坐标之间驾驶了2014年日产Sentra两个月，以收集此驾驶数据",
"集。RPM和速度数据从车辆的车载诊断（OBD2）中提取。我们使用陀螺仪的y轴读数",
"（GYRO_Y）来测量横向加速度，并使用GPS航向之间的时间差（D_HEADING）来测量",
"车辆的转向角。驾驶数据集中使用的所有传感器数据均从FreematicsONE+dongle（Huang",
"2017）收集。由于FreematicsONE+dongle中传感器的提取时间周期不同，因此使用时间",
"上的线性插值进行数据预处理。总共使用五种类型的感觉数据来训练神经网络：RPM",
"SPEED，加速和减速（D_SPEED），陀螺仪的y轴（GYRO_Y）和GPS航向的时间差",
"（DHEADING）。这五个特征每250ms采样一次。为了预测车辆未来5秒的驾驶模式，我",
"们使用来自每个特征的十秒数据（40点），将收集的数据按特征进行归一化，并将它们连接",
"起来以形成特征向量窗口。随着预测的进行，特征向量窗口每250ms滑动一次。对于建议",
"的FP-GFA和2S-GFA体系结构，RPM，SPEED和D_SPEED包含在第一组中，而第二",
"pusho g 0 G",
"3",
"pop"
] | {
"arxiv_id": "1810.04160",
"language": "zh",
"page_idx": 2
} |
||
1810.04160-zh-3 | 1810.04160-zh-3.jpg | push0g0G
pop
组由 GYRO\_ Y 和 D\_ HEADING 组成。5,845 个示例用于训练，2,891 个样本用于测试我
们使用 50,000 次迭代训练不同的神经网络
用于人类活动识别的数据集。我们进一步考虑公共领域人类活动识别数据集 (Anguita et al. 2013) ,该数据集具有六个输人功能：加速度计的三个轴 (ACC\_X, ACC\_Y, ACC\_Z) 和陀螺仪的三个轴 (GYRO\_X, GYRO\_Y, GYRO\_Z),以及六个活动类：WALKING WALKING\_UPSTAIRS, WALKING。这六个特征每 200ms 采样一次。使用相同的滑动窗口方案来定义输入特征向量。对于建议的组级和两级门控体系结构，将六个功能分为两个不同的组。第一个特征组有 ACC\_X 和 ACC\_Y,而 ACC\_Z。GYRO\_X、GYRO\_Y 和GYRO_Z包含在第二组中。训练集和测试集分别有 2,410 和 942 个示例。100,000 次迭代用于训练各种神经网络
采用的工具流。采用的仿真环境基于 Tensorflow1。10。1 (Abadi et al.,2015) , Python 中
的开源深度神经网络库
5.2 Configurations of four compared neural networks
我们通过创建四个相应的神经网络模型来比较传统的 CNN架构，基准网络架构 Patel et al (2017) ,建议的特征组门控融合架构 (FG-GFA)和建议的两阶段门控融合架构(2S-GFA)。为了公平比较，我们尽可能匹配所有四个网络共有的处理中使用的神经元和层的数量。然而，需要注意的是，与 CNN 模型相比，网络架构有额外的神经元用于提取特征级融合权重， FG-GFA 采用额外的神经元用于提取组级融合权重，2S-GFA 采用特征和组级融合权重。四个神经网络的配置详见附录
5.3 Sensor Noise and Failures
为了更全面地评估不同体系结构的性能，我们采用了两个被称为“干净”的数据集的原始数
据，并且还通过为数据的训练和测试子集引人感官噪声和故障来创建数据集的变体
我们通过在特征向量的所有分量上添加随机高斯噪声来模拟感官输入的退化
(1)
$$v_n=v_{ori}(1+\gamma\varepsilon),$$
其中$v_{ori}$和$v_n$分别是原始和有噪声的分量值，$\varepsilon$是遵循正态分布$\sim N(0,1)$的随机变量，$\gamma$
控制 XMATHX 在不同级别的 XMAT
此外，我们通过在训练和测试期间模拟更多灾难性传感器故障的发生来修改原始数据集。这里，在每个时间戳处，随机选择一个特征，并且该特征的相应值在特征向量中被擦除为零。6 Experimental Results
我们根据上一节中描述的设置，使用我们的驾驶模式预测数据集和公共领域人类活动识别
数据集 (Anguita et al., 2013) 评估不同体系结构的性能
6.1 Results on Driving Mode Prediction
门控) CNN 架构相比，这两个建议的架构有了显着改进，并且在基准网络架构上也有了显
着改进
级和两阶段体系结构的损耗，如表$\mathbb{E}??$ 所示，这表明如前所述，网络体系结构中可能发生
过拟合
干净数据的预测精度。Fig。$\operatorname{E??}$表明，当数据没有额外的噪声或故障时，建议的两阶段体系结构具有最佳的预测精度，而组级 FG-GFA 体系结构则产生第二好的性能。与传统(非有趣的是，尽管没有为测试集产生最佳性能，但基准网络体系结构的损耗小于训练集的组有噪声或感觉故障的预测精度。为了验证网络的鲁棒性，我们在训练和测试数据集中引入不同水平的高斯噪声时测试了四个神经网络模型。在表中。?? , 所提出的两种架构产生了强大的性能，两级架构在所有情况下都是最好的。
在表中。3 , 我们在引人随机传感器故障的情况下比较了四种不同的模型。由于与高斯噪声相比，随机故障更具灾难性，因此在这种情况下，所有模型的整体性能都会下降。尽管如此，提出的两个模型显示出最佳性能，并且两阶段架构比基准网络架构高出近 3 个%
4
push0g0G
pop | [
[
"0.1772",
"0.03915",
"0.2688",
"0.03726",
"0.2688",
"0.05176",
"0.1772",
"0.05365"
],
[
"0.915",
"0.0404",
"0.9478",
"0.0404",
"0.9478",
"0.05618",
"0.915",
"0.05618"
],
[
"0.1748",
"0.1061",
"0.822",
"0.1061",
"0.822",
"0.1193",
"0.1748",
"0.1193"
],
[
"0.1765",
"0.1193",
"0.4763",
"0.1193",
"0.4763",
"0.1332",
"0.1765",
"0.1332"
],
[
"0.179",
"0.1427",
"0.819",
"0.1427",
"0.819",
"0.1528",
"0.179",
"0.1528"
],
[
"0.1765",
"0.1547",
"0.8203",
"0.1547",
"0.8203",
"0.1654",
"0.1765",
"0.1654"
],
[
"0.1765",
"0.1686",
"0.822",
"0.1686",
"0.822",
"0.1818",
"0.1765",
"0.1818"
],
[
"0.1748",
"0.1812",
"0.8203",
"0.1818",
"0.8203",
"0.1957",
"0.1748",
"0.1945"
],
[
"0.179",
"0.1963",
"0.822",
"0.1963",
"0.822",
"0.2102",
"0.179",
"0.2102"
],
[
"0.1765",
"0.2102",
"0.822",
"0.2102",
"0.822",
"0.2235",
"0.1765",
"0.2235"
],
[
"0.1765",
"0.2235",
"0.8228",
"0.2235",
"0.8228",
"0.2374",
"0.1765",
"0.2374"
],
[
"0.1772",
"0.2393",
"0.3406",
"0.2393",
"0.3406",
"0.25",
"0.1772",
"0.25"
],
[
"0.1772",
"0.2588",
"0.8228",
"0.2588",
"0.8228",
"0.2727",
"0.1772",
"0.2727"
],
[
"0.1765",
"0.2734",
"0.3367",
"0.2734",
"0.3367",
"0.2842",
"0.1765",
"0.2842"
],
[
"0.1732",
"0.303",
"0.568",
"0.3037",
"0.568",
"0.318",
"0.1732",
"0.3176"
],
[
"0.1765",
"0.33",
"0.822",
"0.33",
"0.822",
"0.3435",
"0.1765",
"0.3435"
],
[
"0.1772",
"0.3435",
"0.822",
"0.3435",
"0.822",
"0.3567",
"0.1772",
"0.3567"
],
[
"0.1772",
"0.358",
"0.8228",
"0.358",
"0.8228",
"0.3718",
"0.1772",
"0.3718"
],
[
"0.1772",
"0.3718",
"0.822",
"0.3718",
"0.822",
"0.385",
"0.1772",
"0.385"
],
[
"0.1765",
"0.385",
"0.8228",
"0.385",
"0.8228",
"0.3984",
"0.1765",
"0.3984"
],
[
"0.1732",
"0.3997",
"0.3718",
"0.3997",
"0.3718",
"0.4136",
"0.1732",
"0.4136"
],
[
"0.1748",
"0.43",
"0.4004",
"0.43",
"0.4004",
"0.443",
"0.1748",
"0.443"
],
[
"0.1772",
"0.4583",
"0.822",
"0.4583",
"0.822",
"0.4685",
"0.1772",
"0.4685"
],
[
"0.1732",
"0.4697",
"0.763",
"0.4702",
"0.763",
"0.4841",
"0.1732",
"0.483"
],
[
"0.1772",
"0.4937",
"0.7114",
"0.4937",
"0.7114",
"0.504",
"0.1772",
"0.504"
],
[
"0.4346",
"0.515",
"0.562",
"0.515",
"0.562",
"0.529",
"0.4346",
"0.529"
],
[
"0.8022",
"0.515",
"0.824",
"0.515",
"0.824",
"0.529",
"0.8022",
"0.529"
],
[
"0.1748",
"0.536",
"0.8228",
"0.537",
"0.8228",
"0.552",
"0.1748",
"0.5503"
],
[
"0.1765",
"0.5503",
"0.4468",
"0.5503",
"0.4468",
"0.5645",
"0.1765",
"0.5645"
],
[
"0.1765",
"0.5728",
"0.8228",
"0.5728",
"0.8228",
"0.586",
"0.1765",
"0.586"
],
[
"0.1765",
"0.586",
"0.8164",
"0.586",
"0.8164",
"0.5996",
"0.1765",
"0.5996"
],
[
"0.1732",
"0.6206",
"0.392",
"0.6206",
"0.392",
"0.6353",
"0.1732",
"0.6353"
],
[
"0.1765",
"0.6514",
"0.8228",
"0.6514",
"0.8228",
"0.6655",
"0.1765",
"0.6655"
],
[
"0.1765",
"0.6665",
"0.5703",
"0.6665",
"0.5703",
"0.6797",
"0.1765",
"0.6797"
],
[
"0.1748",
"0.6963",
"0.4763",
"0.6963",
"0.4763",
"0.7104",
"0.1748",
"0.7104"
],
[
"0.1765",
"0.7227",
"0.822",
"0.7227",
"0.822",
"0.7363",
"0.1765",
"0.7363"
],
[
"0.1765",
"0.7363",
"0.8228",
"0.7363",
"0.8228",
"0.7505",
"0.1765",
"0.7505"
],
[
"0.1765",
"0.7505",
"0.8228",
"0.7505",
"0.8228",
"0.7646",
"0.1765",
"0.7646"
],
[
"0.1765",
"0.765",
"0.2255",
"0.765",
"0.2255",
"0.7764",
"0.1765",
"0.7764"
],
[
"0.1765",
"0.785",
"0.8228",
"0.785",
"0.8228",
"0.799",
"0.1765",
"0.799"
],
[
"0.1765",
"0.799",
"0.8228",
"0.799",
"0.8228",
"0.8125",
"0.1765",
"0.8125"
],
[
"0.1765",
"0.8145",
"0.2255",
"0.8145",
"0.2255",
"0.8257",
"0.1765",
"0.8257"
],
[
"0.1765",
"0.834",
"0.822",
"0.834",
"0.822",
"0.8477",
"0.1765",
"0.8477"
],
[
"0.1772",
"0.8486",
"0.822",
"0.8486",
"0.822",
"0.8623",
"0.1772",
"0.8623"
],
[
"0.1748",
"0.8613",
"0.5293",
"0.8623",
"0.5293",
"0.8755",
"0.1748",
"0.875"
],
[
"0.1772",
"0.8853",
"0.8203",
"0.8853",
"0.8203",
"0.896",
"0.1772",
"0.896"
],
[
"0.1772",
"0.898",
"0.8228",
"0.898",
"0.8228",
"0.911",
"0.1772",
"0.911"
],
[
"0.1772",
"0.9136",
"0.777",
"0.9136",
"0.777",
"0.924",
"0.1772",
"0.924"
],
[
"0.1772",
"0.9487",
"0.2688",
"0.9463",
"0.2688",
"0.961",
"0.1772",
"0.9634"
],
[
"0.581",
"0.9526",
"0.5947",
"0.9478",
"0.601",
"0.9585",
"0.5884",
"0.9634"
],
[
"0.9165",
"0.952",
"0.947",
"0.952",
"0.947",
"0.964",
"0.9165",
"0.964"
]
] | [
"pusho g 0 G",
"pop",
"组由GYRO_Y和D_HEADING组成。5,845个示例用于训练，2,891个样本用于测试我",
"们使用50,000次送代训练不同的神经网络",
"用于人类活动识别的数据集。我们进一步考虑公共领域人类活动识别数据集（Anguitaetal.",
"2013），该数据集具有六个输入功能：加速度计的三个轴（ACCX，ACCY，ACCZ",
"和陀螺仪的三个轴（GYRO_X，GYRO_Y，GYRO_Z），以及六个活动类：WALKING",
"WALKINGUPSTAIRS，WALKING。这六个特征每200mS采样一次。使用相同的滑动窗",
"口方案来定义输人特征向量。对于建议的组级和两级门控体系结构，将六个功能分为两个",
"不同的组。第一个特征组有ACC_X和ACC_Y，而ACC_Z。GYRO_X、GYRO_Y和",
"GYRO_Z包含在第二组中。训练集和测试集分别有2,410和942个示例。100,000次送代",
"用于训练各种神经网络",
"采用的工具流。采用的仿真环境基于Tensorflowl。10。1（Abadietal.,2015），Python中",
"的开源深度神经网络库",
"5.2Configurations of four compared neural networks",
"我们通过创建四个相应的神经网络模型来比较传统的CNN架构，基准网络架构Pateletal.",
"(2017），建议的特征组门控融合架构（FG-GFA）和建议的两阶段门控融合架构（2S-GFA）。",
"为了公平比较，我们尽可能匹配所有四个网络共有的处理中使用的神经元和层的数量。然",
"而，需要注意的是，与CNN模型相比，网络架构有额外的神经元用于提取特征级融合权重",
"FG-GFA采用额外的神经元用于提取组级融合权重，2S-GFA采用特征和组级融合权重。四",
"个神经网络的配置详见附录",
"5.3Sensor Noise and Failures",
"为广更全面地评估不同体系结构的性能，我们采用了两个被称为“十净”的数据集的原始数",
"据，并且还通过为数据的训练和测试子集引入感官噪声和故障来创建数据集的变体",
"我们通过在特征向量的所有分量上添加随机高斯噪声来模拟感官输人的退化",
"Un = Vori(1 + e),",
"(1)",
"其中Uor和Un分别是原始和有噪声的分量值，ε是遵循正态分布～N(0,1）的随机变量，",
"控制XMATHX在不同级别的XMAT",
"此外，我们通过在训练和测试期间模拟更多灾难性传感器故障的发生来修改原始数据集。这",
"里，在每个时间戳处，随机选择一个特征，并且该特征的相应值在特征向量中被擦除为零",
"6Experimental Results",
"我们根据上一节中描述的设置，使用我们的驾驶模式预测数据集和公共领域人类活动识别",
"数据集（Anguitaetal.，2013）评估不同体系结构的性能",
"6.1Results on Driving Mode Prediction",
"干净数据的预测精度。Fig。E?？表明，当数据没有额外的噪声或故障时，建议的两阶段体",
"系结构具有最佳的预测精度，而组级FG-GFA体系结构则产生第二好的性能。与传统（非",
"门控）CNN架构相比，这两个建议的架构有了显着改进，并且在基准网络架构上也有了显",
"着改进",
"有趣的是，尽管没有为测试集产生最佳性能，但基准网络体系结构的损耗小于训练集的组",
"级和两阶段体系结构的损耗，如表E？？所示，这表明如前所述，网络体系结构中可能发生",
"过拟合",
"有噪声或感觉故障的预测精度。为了验证网络的鲁棒性，我们在训练和测试数据集中引入",
"不同水平的高斯噪声时测试了四个神经网络模型。在表中。？？，所提出的两种架构产生了",
"强大的性能，两级架构在所有情况下都是最好的。",
"在表中。3，我们在引入随机传感器故障的情况下比较了四种不同的模型。由于与高斯噪",
"声相比，随机故障更具灾难性，因此在这种情况下，所有模型的整体性能都会下降。尽管如",
"此，提出的两个模型显示出最佳性能，并且两阶段架构比基准网络架构高出近3个%",
"pusho g 0 G",
"4",
"pop"
] | {
"arxiv_id": "1810.04160",
"language": "zh",
"page_idx": 3
} |
||
1810.04160-zh-4 | 1810.04160-zh-4.jpg | push0g0G
pop
push0 g 0 GTable 3: Results of driving mode prediction with sensor failures. pop
<table>
<tbody>
<tr>
<td> </td>
<td>Non-NetGated</td>
<td>$\overline{85.47}\%$</td>
</tr>
<tr>
<td>push0 g 0 G $\operatorname{pop}$</td>
<td>NetGated</td>
<td>$86.44\%$</td>
</tr>
<tr>
<td> </td>
<td>Group- level</td>
<td>$88.34\%$</td>
</tr>
<tr>
<td> </td>
<td>Two- stage</td>
<td>$89.20\%$</td>
</tr>
</tbody>
</table>
push0g0G
pop
分析建议的体系结构的性能改进。我们提供了关于所提出的两阶段体系结构所观察到的性能改进的可能原因的更多见解。在我们的设置中，驾驶模式预测的三个基本特征是 RPM. 速度和 D\_SPEED, 它们包含在组 1 中。表。?? 展示了基于干净数据的两阶段架构的特征和组级融合权重。我们将 20 个% 高斯噪声添加到 group1 中的 RPM, 并在表中报告更新的融合权重。??。可以看出，RPM 的特征级融合权重下降相当明显，可能反映了这种感官输人的质量下降
在不同的实验中，我们仅将 20 个% 噪声添加到 D\_heading 中，这是 group2 中的一项功能。如表所示。??,与干净数据的情况相比，D\_heading 的特征级融合权重和第二组的组级融合权重均下降。预计减少的权重将减少 D\_heading 对最终预测决策的贡献
6.2 Human Activity Recognition Data Set
我们采用类似的方法来演示各种模型在人类活动识别数据集上的性能
干净数据的预测精度。Fig。5 总结了基于干净数据的四个模型的性能。同样，两阶段2S-GFA 体系结构产生最佳性能，而建议的组级体系结构 FG-GFA 体系结构在四个模型中产生第二好的性能。
<FigureHere>
nlann
push0g0G push0 g 0 GFigure 5: Results on human activity recognition under the clean data. pop
pop
有噪声或感觉故障的预测精度。表。8 表明，随着高斯噪声的增加，所有模型的预测精度都会下降。然而，与其他两个模型相比，两个建议的体系结构的鲁棒性和改进是显而易见的。表。7 总结了引人传感器故障时的结果。在这种情况下，非网络模型(传统 CNN)的准确性下降了 10%。尽管如此，与传统的 CNN 和基准网络架构相比，这两个建议的架构仍显示出更好的鲁棒性。。具体来说，对于这个更具挑战性的测试用例，两阶段门控体系结构的性能优于非网络模型 5% 和网络模型 3%
5
push0 g 0 G
pop | [
[
"0.1772",
"0.03915",
"0.2695",
"0.03726",
"0.2695",
"0.05176",
"0.1772",
"0.05365"
],
[
"0.9165",
"0.04166",
"0.947",
"0.04166",
"0.947",
"0.05365",
"0.9165",
"0.05365"
],
[
"0.29",
"0.113",
"0.884",
"0.11426",
"0.884",
"0.1288",
"0.29",
"0.1276"
],
[
"0.535",
"0.1338",
"0.5947",
"0.1338",
"0.5947",
"0.1459",
"0.535",
"0.1459"
],
[
"0.6323",
"0.132",
"0.7837",
"0.1338",
"0.7837",
"0.1483",
"0.6323",
"0.1465"
],
[
"0.509",
"0.1477",
"0.6216",
"0.1477",
"0.6216",
"0.161",
"0.509",
"0.161"
],
[
"0.6167",
"0.1509",
"0.636",
"0.1509",
"0.636",
"0.1578",
"0.6167",
"0.1578"
],
[
"0.6313",
"0.1477",
"0.694",
"0.1477",
"0.694",
"0.1622",
"0.6313",
"0.1622"
],
[
"0.3733",
"0.1622",
"0.584",
"0.161",
"0.584",
"0.1755",
"0.3733",
"0.1768"
],
[
"0.6357",
"0.1635",
"0.693",
"0.1635",
"0.693",
"0.1749",
"0.6357",
"0.1749"
],
[
"0.511",
"0.1755",
"0.597",
"0.1755",
"0.597",
"0.1901",
"0.511",
"0.1901"
],
[
"0.636",
"0.1768",
"0.693",
"0.1768",
"0.693",
"0.1881",
"0.636",
"0.1881"
],
[
"0.5107",
"0.1881",
"0.5864",
"0.1901",
"0.5864",
"0.2052",
"0.5107",
"0.2026"
],
[
"0.6357",
"0.1901",
"0.693",
"0.1901",
"0.693",
"0.202",
"0.6357",
"0.202"
],
[
"0.1765",
"0.1995",
"0.2632",
"0.1995",
"0.2632",
"0.214",
"0.1765",
"0.214"
],
[
"0.1765",
"0.2229",
"0.2091",
"0.2229",
"0.2091",
"0.238",
"0.1765",
"0.238"
],
[
"0.1772",
"0.2588",
"0.8228",
"0.2588",
"0.8228",
"0.2727",
"0.1772",
"0.2727"
],
[
"0.1765",
"0.2727",
"0.822",
"0.2727",
"0.822",
"0.286",
"0.1765",
"0.286"
],
[
"0.1772",
"0.2886",
"0.8203",
"0.2886",
"0.8203",
"0.2986",
"0.1772",
"0.2986"
],
[
"0.1765",
"0.3005",
"0.822",
"0.3005",
"0.822",
"0.3145",
"0.1765",
"0.3145"
],
[
"0.1765",
"0.3145",
"0.8228",
"0.3145",
"0.8228",
"0.3276",
"0.1765",
"0.3276"
],
[
"0.1765",
"0.3289",
"0.308",
"0.3289",
"0.308",
"0.3423",
"0.1765",
"0.3423"
],
[
"0.1765",
"0.3499",
"0.822",
"0.3499",
"0.822",
"0.363",
"0.1765",
"0.363"
],
[
"0.1765",
"0.3623",
"0.8228",
"0.3623",
"0.8228",
"0.377",
"0.1765",
"0.377"
],
[
"0.1765",
"0.377",
"0.739",
"0.377",
"0.739",
"0.3901",
"0.1765",
"0.3901"
],
[
"0.1748",
"0.4092",
"0.492",
"0.4092",
"0.492",
"0.4224",
"0.1748",
"0.4224"
],
[
"0.1772",
"0.4375",
"0.6626",
"0.4375",
"0.6626",
"0.4475",
"0.1772",
"0.4475"
],
[
"0.1772",
"0.4583",
"0.38",
"0.4583",
"0.38",
"0.4685",
"0.1772",
"0.4685"
],
[
"0.3743",
"0.457",
"0.8228",
"0.457",
"0.8228",
"0.4702",
"0.3743",
"0.4702"
],
[
"0.1765",
"0.4702",
"0.822",
"0.4702",
"0.822",
"0.4841",
"0.1765",
"0.4841"
],
[
"0.1748",
"0.4841",
"0.3145",
"0.4841",
"0.3145",
"0.4976",
"0.1748",
"0.4976"
],
[
"0.4885",
"0.632",
"0.619",
"0.6333",
"0.619",
"0.6465",
"0.4885",
"0.646"
],
[
"0.3555",
"0.754",
"0.4836",
"0.754",
"0.4836",
"0.7676",
"0.3555",
"0.7676"
],
[
"0.1748",
"0.779",
"0.9067",
"0.7764",
"0.9067",
"0.7974",
"0.1748",
"0.8"
],
[
"0.1772",
"0.8066",
"0.2075",
"0.8066",
"0.2075",
"0.8193",
"0.1772",
"0.8193"
],
[
"0.1772",
"0.8413",
"0.8228",
"0.8413",
"0.8228",
"0.855",
"0.1772",
"0.855"
],
[
"0.1772",
"0.857",
"0.822",
"0.857",
"0.822",
"0.867",
"0.1772",
"0.867"
],
[
"0.1748",
"0.869",
"0.8228",
"0.869",
"0.8228",
"0.8833",
"0.1748",
"0.8833"
],
[
"0.1765",
"0.8833",
"0.822",
"0.8833",
"0.822",
"0.8965",
"0.1765",
"0.8965"
],
[
"0.1772",
"0.898",
"0.8228",
"0.898",
"0.8228",
"0.911",
"0.1772",
"0.911"
],
[
"0.1765",
"0.911",
"0.4805",
"0.911",
"0.4805",
"0.925",
"0.1765",
"0.925"
],
[
"0.1772",
"0.9487",
"0.2688",
"0.9463",
"0.2688",
"0.961",
"0.1772",
"0.9634"
],
[
"0.5825",
"0.9478",
"0.5986",
"0.9478",
"0.5986",
"0.962",
"0.5825",
"0.962"
],
[
"0.9165",
"0.9507",
"0.947",
"0.9507",
"0.947",
"0.9634",
"0.9165",
"0.9634"
]
] | [
"pusho g 0 G",
"pop",
"pusho g 0 GTable 3: Results of driving mode prediction with sensor failures. pop",
"Network",
"PredictionAccuracy",
"Non-NetGated",
"",
"85.47%",
"pusho g O G pop NetGated",
"86.44%",
"Group-level",
"88.34 %",
"Two-stage",
"89.20 %",
"push0 g 0 G",
"pop",
"分析建议的体系结构的性能改进。我们提供了关于所提出的两阶段体系结构所观察到的性",
"能改进的可能原因的更多见解。在我们的设置中，驾驶模式预测的三个基本特征是RPM",
"速度和DSPEED，它们包含在组1中。表。？？展示了基于干净数据的两阶段架构的特",
"征和组级融合权重。我们将20个%高斯噪声添加到group1中的RPM，并在表中报告更",
"新的融合权重。？？。可以看出，RPM的特征级融合权重下降相当明显，可能反映了这种感",
"官输人的质量下降",
"在不同的实验中，我们仅将20个%噪声添加到D_heading中，这是group2中的一项功",
"能。如表所示。?？，与干净数据的情况相比，Dheading的特征级融合权重和第二组的组",
"级融合权重均下降。预计减少的权重将减少D_heading对最终预测决策的贡献",
"6.2Human Activity Recognition Data Set",
"我们采用类似的方法来演示各种模型在人类活动识别数据集上的性能",
"干净数据的预测精度。Fig。",
"5总结了基于干净数据的四个模型的性能。同样，两阶段",
"2S-GFA体系结构产生最佳性能，而建议的组级体系结构FG-GFA体系结构在四个模型中",
"产生第二好的性能。",
"human_clean.png",
"human_clean.png",
"pushO g O G pushO g O GFigure 5: Results on human activity recognition under the clean data. pop",
"pop",
"有噪声或感觉故障的预测精度。表。8表明，随着高斯噪声的增加，所有模型的预测精度",
"都会下降。然而，与其他两个模型相比，两个建议的体系结构的鲁棒性和改进是显而易见",
"的。表。7总结了引入传感器故障时的结果。在这种情况下，非网络模型（传统CNN）的",
"准确性下降了10%。尽管如此，与传统的CNN和基准网络架构相比，这两个建议的架构仍",
"显示出更好的鲁棒性。。具体来说，对于这个更具挑战性的测试用例，两阶段门控体系结构",
"的性能优于非网络模型5%和网络模型3%",
"pusho g 0 G",
"5",
"pop"
] | {
"arxiv_id": "1810.04160",
"language": "zh",
"page_idx": 4
} |
||
2104.08512-zh-1 | 2104.08512-zh-1.jpg | ## 形态拐点的最小监督
$\mathbf{OmerGoldman}$
Bar IIan University
[email protected]
Reut Tsarfaty
Bar IIan University
[email protected]
# Abstract
Neural models for the various flavours of morphological inflection tasks have proven to be extremely accurate given ample labeled data, yet labeled data may be slow and costly to obtain. In this work we aim to overcome this annotation bottleneck by bootstrapping labeled data from a seed as small as five labeled inflection tables, accompanied by a large bulk of unlabeled text. Our bootstrapping method exploits the orthographic and semantic regularities in morphological systems in a two-phased setup, where word tagging based on analogies is followed by word pairing based on distances . Our experiments with the Paradigm Cell Filling Problem over eight typologically different languages show that in languages with relatively simple morphology, orthographic regularities on their own allow inflection models to achieve respectable accuracy. Combined orthographic and semantic regularities alleviate difficulties with particularly complex morpho-phonological systems. We further show that our bootstrapping methods substantially outperform hallucination-based methods commonly used for overcoming the annotation bottleneck in morphological inflection tasks.
## 1 Introduction
在过去十年中将神经模型引人自然语言处理中，导致所有 supervised 生成任务的巨大改进， 包括形态变化。$^1$特别是，以前的作品(Cotterell et al., 2017; Silfverberg and Hulden, 2018) 在范式单元填充问题(PCFP)(Ackerman et al., 2009) 上取得了近乎完美的性能，其中要求模型在给定几种形式的相同词素的情况下在拐点表中提供任何形式。2
${^{1}\text{In recent years the term reinflection has surfaced as a ref-}}$ erence to morphological inflection done not necessarily from the lemma. In this paper we will refer to both inflection and reinflection as "inflection", and specify whenever we refer to inflection done exclusively from the lemma.
$^{2}$Throughout the paper we conform to the linguistic terminology, where` lexeme' stands for an abstract lexical en-
最近的两项工作以不同的方式在减少监督方面取得了进展。第一个简单地提供了具有较小训练集的场景-例如，在 SIGMORPHON 的共享任务 (Cotterell et al.,2017,2018) 中。第二种研究途径旨在发现给定大量未标记数据的未知语言的范式结构，无论是单独的(Soricut and Och, 2015; Elsner et al., 2019) , 以及词汇表(Erdmann et al., 2020) 中所有相关形式的列表， 还是引理(Jin et al., 2020) 的列表
第一种尝试的问题在于，鉴于最成功模型的神经性质，它们在有限监督下的性能是有上限的，只有在初始数据足够多样化的情况下，数据增强才有可能有所帮助。至于第二种完全没有监管的情况，有些悲观和不现实。即使对于资源匮乏的语言不存在大量标记的语言数据，通常也存在有关其范式结构的 knowledge , 可以采用。例如，UniMorph (Kirov et al., 2018) 包含许多语言的少量带标签的拐点表，从像Ingrian 这样的晦涩难懂的语言到像 Georgian 这样的缺乏全球关注的广泛使用的国家语言
在这项工作中，我们提出了一个新的lowresourced morphological inflection 场景，对于那些广泛使用的稀疏注释语言来说，它更加乐观和现实。我们假设一个 minimal 监督集和大量未标记的文本，从而平衡了降低监督资源的两种趋势。对于以监督方式训练的拐点模型， 我们将少量的 fre 拐点表引导到一个完整的训练集中，尽管很吵。我们的方法利用了屈折形态中丰富的正则性和语义性
基于此方法，我们训练了八种语言的形态拐点模型，并表明，对于其中的五种印欧语系，正字法规律性足以训练出取得合理成功的形态拐点。我们进一步表明，对于具有复杂形态语音系统的语言，例如芬兰语，土耳其语和匈牙利语，需要一种结合正字法和语义规则的方法才能达到相同的性能水平。错误分析显示，拐点try, e.g., the English RUN , and its 'forms' are the words that
convey this lexical meaning with some inflectional relations between them, e.g. run ,running but not runner.`Paradigm , will stand for a group of lexemes sharing a POS tag, in the same manner as the English lexeme RUN is part of the verbal paradigm. | [
[
"0.392",
"0.0956",
"0.6055",
"0.0956",
"0.6055",
"0.10986",
"0.392",
"0.10986"
],
[
"0.266",
"0.1443",
"0.4048",
"0.1455",
"0.4048",
"0.1603",
"0.266",
"0.1592"
],
[
"0.6035",
"0.1443",
"0.727",
"0.1455",
"0.727",
"0.1603",
"0.6035",
"0.1592"
],
[
"0.2544",
"0.1609",
"0.4114",
"0.1621",
"0.4114",
"0.177",
"0.2544",
"0.1758"
],
[
"0.588",
"0.1609",
"0.7446",
"0.1621",
"0.7446",
"0.177",
"0.588",
"0.1758"
],
[
"0.2007",
"0.1787",
"0.4685",
"0.1787",
"0.4685",
"0.1924",
"0.2007",
"0.1924"
],
[
"0.5283",
"0.1799",
"0.804",
"0.1799",
"0.804",
"0.1936",
"0.5283",
"0.1936"
],
[
"0.2627",
"0.2542",
"0.3418",
"0.2542",
"0.3418",
"0.2683",
"0.2627",
"0.2683"
],
[
"0.531",
"0.2546",
"0.879",
"0.2546",
"0.879",
"0.2683",
"0.531",
"0.2683"
],
[
"0.511",
"0.2695",
"0.879",
"0.2695",
"0.879",
"0.2832",
"0.511",
"0.2832"
],
[
"0.1486",
"0.2808",
"0.4575",
"0.2808",
"0.4575",
"0.2915",
"0.1486",
"0.2915"
],
[
"0.5146",
"0.2861",
"0.881",
"0.2861",
"0.881",
"0.2998",
"0.5146",
"0.2998"
],
[
"0.1453",
"0.2922",
"0.4568",
"0.2922",
"0.4568",
"0.306",
"0.1453",
"0.306"
],
[
"0.1453",
"0.308",
"0.4568",
"0.308",
"0.4568",
"0.3218",
"0.1453",
"0.3218"
],
[
"0.513",
"0.3027",
"0.8657",
"0.3027",
"0.8657",
"0.3164",
"0.513",
"0.3164"
],
[
"0.1453",
"0.3218",
"0.4575",
"0.3218",
"0.4575",
"0.335",
"0.1453",
"0.335"
],
[
"0.513",
"0.3196",
"0.879",
"0.3196",
"0.879",
"0.3333",
"0.513",
"0.3333"
],
[
"0.1453",
"0.3362",
"0.4575",
"0.3362",
"0.4575",
"0.3499",
"0.1453",
"0.3499"
],
[
"0.513",
"0.335",
"0.881",
"0.3362",
"0.881",
"0.3499",
"0.513",
"0.3486"
],
[
"0.1436",
"0.3486",
"0.4575",
"0.3508",
"0.4575",
"0.3652",
"0.1436",
"0.3635"
],
[
"0.511",
"0.3508",
"0.879",
"0.3508",
"0.879",
"0.364",
"0.511",
"0.364"
],
[
"0.147",
"0.3665",
"0.4568",
"0.3665",
"0.4568",
"0.3772",
"0.147",
"0.3772"
],
[
"0.511",
"0.3665",
"0.881",
"0.3665",
"0.881",
"0.3801",
"0.511",
"0.3801"
],
[
"0.147",
"0.379",
"0.4592",
"0.379",
"0.4592",
"0.3926",
"0.147",
"0.3926"
],
[
"0.513",
"0.3843",
"0.7715",
"0.3843",
"0.7715",
"0.398",
"0.513",
"0.398"
],
[
"0.1453",
"0.3938",
"0.4575",
"0.3938",
"0.4575",
"0.4067",
"0.1453",
"0.4067"
],
[
"0.532",
"0.401",
"0.881",
"0.401",
"0.881",
"0.4146",
"0.532",
"0.4146"
],
[
"0.1453",
"0.408",
"0.4575",
"0.4067",
"0.4575",
"0.4204",
"0.1453",
"0.4216"
],
[
"0.513",
"0.415",
"0.881",
"0.4163",
"0.881",
"0.4312",
"0.513",
"0.43"
],
[
"0.147",
"0.4216",
"0.4575",
"0.4216",
"0.4575",
"0.4353",
"0.147",
"0.4353"
],
[
"0.1436",
"0.434",
"0.4575",
"0.4358",
"0.4575",
"0.4507",
"0.1436",
"0.449"
],
[
"0.511",
"0.4329",
"0.879",
"0.4329",
"0.879",
"0.4465",
"0.511",
"0.4465"
],
[
"0.1453",
"0.4495",
"0.4575",
"0.4495",
"0.4575",
"0.4631",
"0.1453",
"0.4631"
],
[
"0.513",
"0.4495",
"0.881",
"0.4495",
"0.881",
"0.4631",
"0.513",
"0.4631"
],
[
"0.1436",
"0.4631",
"0.3594",
"0.462",
"0.3594",
"0.4768",
"0.1436",
"0.478"
],
[
"0.3787",
"0.4614",
"0.4575",
"0.4644",
"0.4568",
"0.4785",
"0.378",
"0.4756"
],
[
"0.513",
"0.4656",
"0.881",
"0.4656",
"0.881",
"0.4785",
"0.513",
"0.4785"
],
[
"0.147",
"0.478",
"0.4575",
"0.478",
"0.4575",
"0.4912",
"0.147",
"0.4912"
],
[
"0.513",
"0.4797",
"0.882",
"0.481",
"0.882",
"0.4954",
"0.513",
"0.4946"
],
[
"0.1453",
"0.4922",
"0.4592",
"0.4922",
"0.4592",
"0.506",
"0.1453",
"0.506"
],
[
"0.511",
"0.4963",
"0.8716",
"0.4976",
"0.8716",
"0.5127",
"0.511",
"0.511"
],
[
"0.1453",
"0.5073",
"0.4592",
"0.5073",
"0.4592",
"0.521",
"0.1453",
"0.521"
],
[
"0.867",
"0.505",
"0.876",
"0.505",
"0.876",
"0.509",
"0.867",
"0.509"
],
[
"0.5146",
"0.513",
"0.882",
"0.513",
"0.882",
"0.527",
"0.5146",
"0.527"
],
[
"0.1453",
"0.5215",
"0.4592",
"0.5215",
"0.4592",
"0.535",
"0.1453",
"0.535"
],
[
"0.513",
"0.5293",
"0.879",
"0.5293",
"0.879",
"0.542",
"0.513",
"0.542"
],
[
"0.1453",
"0.535",
"0.4592",
"0.535",
"0.4592",
"0.549",
"0.1453",
"0.549"
],
[
"0.1453",
"0.549",
"0.4048",
"0.549",
"0.4048",
"0.5625",
"0.1453",
"0.5625"
],
[
"0.511",
"0.5444",
"0.881",
"0.546",
"0.881",
"0.56",
"0.511",
"0.5586"
],
[
"0.4165",
"0.5503",
"0.4575",
"0.5503",
"0.4575",
"0.5625",
"0.4165",
"0.5625"
],
[
"0.1436",
"0.5625",
"0.4575",
"0.5635",
"0.4575",
"0.5776",
"0.1436",
"0.5767"
],
[
"0.5146",
"0.5635",
"0.86",
"0.5635",
"0.86",
"0.5737",
"0.5146",
"0.5737"
],
[
"0.1453",
"0.5776",
"0.4575",
"0.5776",
"0.4575",
"0.5913",
"0.1453",
"0.5913"
],
[
"0.532",
"0.579",
"0.879",
"0.579",
"0.879",
"0.589",
"0.532",
"0.589"
],
[
"0.1453",
"0.593",
"0.3645",
"0.593",
"0.3645",
"0.6055",
"0.1453",
"0.6055"
],
[
"0.378",
"0.593",
"0.4568",
"0.593",
"0.4568",
"0.604",
"0.378",
"0.604"
],
[
"0.511",
"0.5933",
"0.879",
"0.5933",
"0.879",
"0.607",
"0.511",
"0.607"
],
[
"0.1453",
"0.6055",
"0.4592",
"0.6055",
"0.4592",
"0.6206",
"0.1453",
"0.6206"
],
[
"0.513",
"0.611",
"0.882",
"0.611",
"0.882",
"0.6245",
"0.513",
"0.6245"
],
[
"0.147",
"0.6206",
"0.4568",
"0.6206",
"0.4568",
"0.6343",
"0.147",
"0.6343"
],
[
"0.513",
"0.627",
"0.881",
"0.627",
"0.881",
"0.64",
"0.513",
"0.64"
],
[
"0.1453",
"0.6343",
"0.4575",
"0.6343",
"0.4575",
"0.648",
"0.1453",
"0.648"
],
[
"0.1453",
"0.6484",
"0.4458",
"0.6484",
"0.4458",
"0.662",
"0.1453",
"0.662"
],
[
"0.5146",
"0.6436",
"0.881",
"0.6436",
"0.881",
"0.6567",
"0.5146",
"0.6567"
],
[
"0.513",
"0.66",
"0.875",
"0.66",
"0.875",
"0.6704",
"0.513",
"0.6704"
],
[
"0.11755",
"0.675",
"0.1453",
"0.675",
"0.1453",
"0.686",
"0.11755",
"0.686"
],
[
"0.1486",
"0.6733",
"0.2644",
"0.675",
"0.2644",
"0.6895",
"0.1486",
"0.6885"
],
[
"0.513",
"0.676",
"0.881",
"0.676",
"0.881",
"0.6895",
"0.513",
"0.6895"
],
[
"0.5146",
"0.6914",
"0.881",
"0.6914",
"0.881",
"0.705",
"0.5146",
"0.705"
],
[
"0.11755",
"0.6978",
"0.4844",
"0.6978",
"0.4844",
"0.7114",
"0.11755",
"0.7114"
],
[
"0.1184",
"0.7144",
"0.4836",
"0.7144",
"0.4836",
"0.728",
"0.1184",
"0.728"
],
[
"0.5146",
"0.709",
"0.7515",
"0.709",
"0.7515",
"0.7197",
"0.5146",
"0.7197"
],
[
"0.529",
"0.7246",
"0.881",
"0.7246",
"0.881",
"0.7383",
"0.529",
"0.7383"
],
[
"0.12006",
"0.731",
"0.486",
"0.731",
"0.486",
"0.7446",
"0.12006",
"0.7446"
],
[
"0.513",
"0.7393",
"0.881",
"0.7407",
"0.881",
"0.755",
"0.513",
"0.7544"
],
[
"0.11755",
"0.745",
"0.486",
"0.7466",
"0.486",
"0.761",
"0.11755",
"0.7603"
],
[
"0.1184",
"0.761",
"0.486",
"0.7627",
"0.486",
"0.777",
"0.1184",
"0.7754"
],
[
"0.513",
"0.7573",
"0.881",
"0.7573",
"0.881",
"0.771",
"0.513",
"0.771"
],
[
"0.511",
"0.7715",
"0.881",
"0.7725",
"0.881",
"0.7876",
"0.511",
"0.786"
],
[
"0.12006",
"0.778",
"0.4836",
"0.778",
"0.4836",
"0.7915",
"0.12006",
"0.7915"
],
[
"0.5146",
"0.789",
"0.881",
"0.789",
"0.881",
"0.8027",
"0.5146",
"0.8027"
],
[
"0.1184",
"0.796",
"0.4844",
"0.796",
"0.4844",
"0.8086",
"0.1184",
"0.8086"
],
[
"0.513",
"0.8047",
"0.881",
"0.8047",
"0.881",
"0.8184",
"0.513",
"0.8184"
],
[
"0.11755",
"0.8125",
"0.2393",
"0.8086",
"0.2393",
"0.8237",
"0.11755",
"0.8267"
],
[
"0.513",
"0.822",
"0.881",
"0.822",
"0.881",
"0.835",
"0.513",
"0.835"
],
[
"0.1377",
"0.832",
"0.4878",
"0.8345",
"0.4878",
"0.8486",
"0.1377",
"0.8467"
],
[
"0.11755",
"0.8467",
"0.4895",
"0.8467",
"0.4895",
"0.86",
"0.11755",
"0.86"
],
[
"0.51",
"0.8467",
"0.881",
"0.8467",
"0.881",
"0.86",
"0.51",
"0.86"
],
[
"0.11584",
"0.857",
"0.4895",
"0.858",
"0.4895",
"0.8716",
"0.11584",
"0.8706"
],
[
"0.513",
"0.86",
"0.881",
"0.86",
"0.881",
"0.8706",
"0.513",
"0.8706"
],
[
"0.1184",
"0.8716",
"0.4878",
"0.8716",
"0.4878",
"0.882",
"0.1184",
"0.882"
],
[
"0.51",
"0.8716",
"0.882",
"0.8716",
"0.882",
"0.8853",
"0.51",
"0.8853"
],
[
"0.11584",
"0.8813",
"0.3804",
"0.8813",
"0.3804",
"0.8945",
"0.11584",
"0.8945"
],
[
"0.51",
"0.883",
"0.881",
"0.883",
"0.881",
"0.8965",
"0.51",
"0.8965"
],
[
"0.1377",
"0.8945",
"0.4895",
"0.8945",
"0.4895",
"0.9077",
"0.1377",
"0.9077"
],
[
"0.513",
"0.8945",
"0.882",
"0.8936",
"0.882",
"0.9067",
"0.513",
"0.9077"
],
[
"0.11584",
"0.9062",
"0.4895",
"0.9062",
"0.4895",
"0.9194",
"0.11584",
"0.9194"
],
[
"0.511",
"0.9062",
"0.5728",
"0.9077",
"0.5728",
"0.9194",
"0.511",
"0.9175"
]
] | [
"形态拐点的最小监督",
"Omer Goldman",
"Reut Tsarfaty",
"Bar Ilan University",
"Bar Ilan University",
"[email protected]",
"[email protected]",
"Abstract",
"最近的两项工作以不同的方式在减少监督",
"方面取得了进展。第一个简单地提供了具有较",
"Neural modelsforthevariousflavoursofmor-",
"小训练集的场景-例如，在SIGMORPHON的",
"phological inflection tasks have proven to be",
"extremely accurate given ample labeled data,",
"共享任务（Cotterelletal.,2017,2018）中。第",
"yet labeled data may be slow and costly to",
"种研究途径旨在发现给定大量未标记数据的",
"obtain. In this work we aim to overcome",
"未知语言的范式结构，无论是单独的（Soricut",
"this annotation bottleneck by bootstrapping",
"andOch，2015;Elsneretal.2019），以及词汇表",
"labeled data from a seed as small asfive",
"(Erdmannetal.,2020）中所有相关形式的列表，",
"labeled inflection tables, accompanied by a",
"还是引理(Jinetal.,2020)的列表",
"large bulk of unlabeled text. Our bootstrap-",
"第一种尝试的问题在于，鉴于最成功模型的",
"ping method exploits the orthographic and se-",
"神经性质，它们在有限监督下的性能是有上限",
"mantic regularities in morphological systems",
"in a two-phased setup, where word tagging",
"的，只有在初始数据足够多样化的情况下，数",
"based onanalogies is followed by word",
"据增强才有可能有所帮助。至于第二种完全",
"pairing based ondistances",
"Our exper",
"没有监管的情况，有些悲观和不现实。即使对",
"iments with the Paradigm Cell Filling Prob-",
"于资源遗乏的语言不存在大量标记的语言数",
"lem over eight typologically different lan-",
"据，通常也存在有关其范式结构的knowledge",
"guages show that in languages with rela-",
"",
"可以采用。例如，UniMorph(Kirovetal,2018)",
"tively simple morphology, orthographic reg-",
"包含许多语言的少量带标签的拐点表，从像",
"ularities on their own allow inflection mod-",
"els to achieve respectable accuracy.",
"Ingrian这样的嗨涩难懂的语言到像Georgian",
"Com-",
"bined orthographic and semantic regularities",
"这样的缺之全球关注的广泛使用的国家语言",
"alleviate difficulties with particularly complex",
"在这项工作中，我们提出了一个新的low",
"morpho-phonological systems.",
"We further",
"resourcedmorphological inflection场景，对于",
"show that our bootstrapping methods substan-",
"那些广泛使用的稀疏注释语言来说，它更加乐",
"tially outperform hallucination-based methods",
"观和现实。我们假设一个minimal监督集和大",
"commonly used for overcoming the annotation",
"bottleneck in morphological inflection tasks.",
"量未标记的文本，从而平衡了降低监督资源的",
"两种趋势。对于以监督方式训练的拐点模型",
"1",
"Introduction",
"我们将少量的five拐点表引导到一个完整的",
"训练集中，尽管很吵。我们的方法利用了屈折",
"在过去十年中将神经模型引入自然语言处理",
"中，导致所有supervised生成任务的巨大改进",
"形态中丰富的正则性和语义性",
"基于此方法，我们训练了八种语言的形态拐",
"包括形态变化。1特别是，以前的作品（Cotterell",
"点模型，并表明，对于其中的五种印欧语系，正",
"etal.,2017;Silfverberg andHulden,2018)在范式",
"单元填充问题（PCFP）（Ackermanetal.，2009)",
"字法规律性足以训练出取得合理成功的形态拐",
"点。我们进一步表明，对于具有复杂形态语音",
"上取得了近乎完美的性能，其中要求模型在给",
"系统的语言，例如芬兰语，土耳其语和匈牙利",
"定几种形式的相同词素的情况下在拐点表中提",
"语，需要一种结合正字法和语义规则的方法才",
"供任何形式。2",
"能达到相同的性能水平。错误分析显示，拐点",
"'In recent years the term reinflection has surfaced as aref-",
"erence to morphological inflection done not necessarily from",
"try, e.g., the English RUN , and its \"forms' are the words that",
"the lemma. In this paper we will refer to both inflection and",
"convey this lexical meaning with some inflectional relations",
"reinflection as\"inflection\",and specifywheneverwereferto",
"between them, e.g. run , running but not runner.. Paradigm",
"inflection done exclusively from the lemma",
"\" will stand for a group of lexemes sharing a POS tag, in the",
"2 Throughout the paper we conform to the linguistic termi-",
"same manner as the English lexeme RUN is part of the verbal",
"nology, where lexeme stands for an abstract lexical en-",
"paradigm"
] | {
"arxiv_id": "2104.08512",
"language": "zh",
"page_idx": 1
} |
||
2104.08512-zh-2 | 2104.08512-zh-2.jpg | 越接近消失的边缘，我们的系统在其上的表现就越差，因为数据派生的词汇表中存在的示例越少。我们的模型优于 Makarov and Clematide (2018)为低资源设置而设计的模型，即使配备了额外的幻觉数据 (Anastasopoulos and Neubig, 2019)。我们还超越了 Jin et al.(2020) 的最佳模型，该模型不使用任何拐点表，并且对于大多数语言而言都是天际线
我们得出的结论是，为资源匮乏的语言中的屈折形态引导数据集是一种可以采用和探索的可行策略。3
# 2 $\textbf{The Minimally- Supervised Setup}$
Problem Statement 让$\mathcal{L}$是一组词素，每个词素对应于一个拐点表$W^l=\{w_{t_1}^l...w_{t_m}^l\}~,~$其中$w_{t_i}^l$是带有特征束$t_i$的词素$l$的形式。我们的目标是训练一个拐点模型，将具有一组特征的单词映射到同一词素中具有另一组特征的单词。
$$(\langle t_j,w_{t_j}^{l_i}\rangle,t_k)\mapsto w_{t_k}^{l_i}$$
例如，在法语语言范式中：
$\left(\:\langle\:\mathrm{INDPRS1SG,finis\rangle,SBJ***T2SG}\right)$
$\mapsto$finisses
supervised 场景，其中仅给我们一个完整的拐点表$\mathcal{L}=\left\{W^{l_{1}}...W^{l_{n}}\right\}$的$n$示例的小集合(每个示例的大小为$m$ ),以及大量自然发生的(即上下文中)未标记的数据，即 MATHX,即
## 3 $\textbf{The Algorithmic Framework}$
这项工作建议利用小监督种子中显示的模式， 并找到显示 similar (或 analogous )模式的单词
这里提出的算法分两个阶段工作。首先，我们标记具有形态特征的单词，如果它们被发现与最小监督中的示例相似(类似)。然后，我们将标记的单词配对，以便每对都包含相同词素的两种形式。这种算法分工允许配对模块提供健全性检查，并减少潜在的单词标记模块的噪音
该算法的上述草图相当通用，也就是说，我们可以通过以不同的方式插人来计算单词之间的 similarities 或 analogies , 从而获得该框架的不同实例化。在我们的各种算法实现中，我们将使用屈折形态中普遍存在的规律性，即使在少量的监督下也可以检测到。这些规则既可以用正字法表示，因为形式之间的编辑往往会在词素上重复出现，也可以用语义表示，因为共享屈折功能的形式往往会被类似地使用
$^3$Our code is available on https://github.com/
OnlpLab/morphodetection.
在本节的其余部分中，我们将描述这两个模块，每个模块都有不同的变体，具体取决于所
使用的 similarity 的不同概念
3.1 Morphological Features Assignment
为了评估一对看不见的单词$w_1,w_2$是否属于同一词素，我们首先需要表征这两个单词之间的关系。然后，必须将$w_1,w_2$之间的具体关系与同一范式中 2 形态类别$t_j$和$t_k$之间的抽象关系$R_{t_j,t_k}$的某种表示进行比较。如果这些具体关系和抽象关系足够相似，则$w_1$和$w_2$将分别标记为承载特征$t_j$和$t_k$ 。简而言之，这个想法是通过汇总最小种子中所有$n$拐点表中$t_j,t_k$条目的形式之间的差异来获得表示$R_t_j,t_k$ 。这些差异可以用语义、正字法或它们的组合来表述
## 3.1.1 Orthography-Based Tagging
在正交情况下，我们将一对单词之间的差异定义为从一个单词到另一个单词所需的编辑。我们的系统期望的编辑是已删除的子字符串列表和已添加的子字符串列表，从左到右排序。$^{4}$
对于每对形态类别$(t_j,t_k)$ ,它们的正交关系$R_{t_j,t_k}^O$由在$n$监督拐点表中观察到的一组编
为了诱导此功能，我们提出了一个 minimally- 辑定义，每个引理$l_i$
$$\hat{R}_{t_j,t_k}^O:=\left\{\operatorname{edits}(w_{t_j}^{l_i},w_{t_k}^{l_i})\right\}_{i=1}^n$$
之间的关系检查是否存在一对新词 XMATHX 考虑来自法语语言范式的示例，其中关系
(INDPRS1SG,INDPRS3SG)的示例是$\{$ $( finis$, finit),(bois,but),(parle,parla)$\}。在这种$ 情况下，此关系的表示形式为：
$$\begin{array}{c}\hat{R}_{\text{INDPRS1SG,INDPRS3SG}}^O=\\\{(`s',`t'),(`ois',`ut'),(`e',`a')\}\end{array}$$
由于(aime,aima)之间的编辑与(parle,parla) 之间的编辑相同，因此将 aime 正确地视为标记为 INDPRS1SG 和 aima-作为 INDPSTPRF3SG
然而，这个过程很容易发生巧合。例如，le 和 la 之间的编辑与 parle 和 parla 之间的编辑相同，尽管前者实际上是确定者，而不是言语范式的一部分。鉴于可用的关系众多，我们可以期望 incidental 对单词之间的许多编辑与黄金数据中看到的编辑相匹配。为了克服这个问题，我们建议考虑完整的范式结构，而不是孤立地考虑单词对
具体而言，我们建议仅标记已找到的回答 multiple relations in the same paradigm
${\overline{{^4}\text{Though without position indices, in order to deal with}}}$
words of various lengths. | [
[
"0.12006",
"0.0873",
"0.4844",
"0.0873",
"0.4844",
"0.0974",
"0.12006",
"0.0974"
],
[
"0.532",
"0.0885",
"0.8774",
"0.0885",
"0.8774",
"0.0986",
"0.532",
"0.0986"
],
[
"0.12006",
"0.1027",
"0.4878",
"0.1027",
"0.4878",
"0.1164",
"0.12006",
"0.1164"
],
[
"0.511",
"0.1027",
"0.881",
"0.1027",
"0.881",
"0.1164",
"0.511",
"0.1164"
],
[
"0.12006",
"0.11816",
"0.4878",
"0.11816",
"0.4878",
"0.1318",
"0.12006",
"0.1318"
],
[
"0.513",
"0.1194",
"0.7446",
"0.1194",
"0.7446",
"0.133",
"0.513",
"0.133"
],
[
"0.1184",
"0.1354",
"0.4836",
"0.1354",
"0.4836",
"0.1484",
"0.1184",
"0.1484"
],
[
"0.511",
"0.1443",
"0.843",
"0.1455",
"0.843",
"0.1603",
"0.511",
"0.1592"
],
[
"0.12177",
"0.1509",
"0.4844",
"0.1509",
"0.4844",
"0.1644",
"0.12177",
"0.1644"
],
[
"0.1184",
"0.1675",
"0.4895",
"0.1675",
"0.4895",
"0.1812",
"0.1184",
"0.1812"
],
[
"0.513",
"0.1675",
"0.879",
"0.1675",
"0.879",
"0.1812",
"0.513",
"0.1812"
],
[
"0.1184",
"0.1829",
"0.4895",
"0.1829",
"0.4895",
"0.1965",
"0.1184",
"0.1965"
],
[
"0.513",
"0.1829",
"0.879",
"0.1829",
"0.879",
"0.1965",
"0.513",
"0.1965"
],
[
"0.11755",
"0.199",
"0.3198",
"0.1978",
"0.3198",
"0.212",
"0.11755",
"0.2131"
],
[
"0.513",
"0.1996",
"0.881",
"0.1996",
"0.881",
"0.2131",
"0.513",
"0.2131"
],
[
"0.1368",
"0.2156",
"0.4844",
"0.2156",
"0.4844",
"0.2286",
"0.1368",
"0.2286"
],
[
"0.513",
"0.2156",
"0.879",
"0.2156",
"0.879",
"0.2286",
"0.513",
"0.2286"
],
[
"0.1184",
"0.2322",
"0.4878",
"0.2322",
"0.4878",
"0.2452",
"0.1184",
"0.2452"
],
[
"0.511",
"0.231",
"0.881",
"0.2299",
"0.881",
"0.2446",
"0.511",
"0.2452"
],
[
"0.11755",
"0.2477",
"0.2217",
"0.2452",
"0.2217",
"0.26",
"0.11755",
"0.2625"
],
[
"0.5146",
"0.2477",
"0.881",
"0.2477",
"0.881",
"0.2612",
"0.5146",
"0.2612"
],
[
"0.5146",
"0.2642",
"0.881",
"0.2642",
"0.881",
"0.2778",
"0.5146",
"0.2778"
],
[
"0.11755",
"0.2756",
"0.157",
"0.2756",
"0.157",
"0.2874",
"0.11755",
"0.2874"
],
[
"0.1486",
"0.2737",
"0.4358",
"0.275",
"0.4358",
"0.2893",
"0.1486",
"0.288"
],
[
"0.513",
"0.2798",
"0.879",
"0.2798",
"0.879",
"0.2935",
"0.513",
"0.2935"
],
[
"0.5073",
"0.2922",
"0.8774",
"0.2944",
"0.8774",
"0.3135",
"0.5073",
"0.3113"
],
[
"0.12177",
"0.2998",
"0.4895",
"0.2998",
"0.4895",
"0.3135",
"0.12177",
"0.3135"
],
[
"0.1184",
"0.3154",
"0.4878",
"0.3154",
"0.4878",
"0.33",
"0.1184",
"0.33"
],
[
"0.509",
"0.31",
"0.879",
"0.3113",
"0.879",
"0.326",
"0.509",
"0.3247"
],
[
"0.1184",
"0.3308",
"0.4878",
"0.3308",
"0.4878",
"0.3445",
"0.1184",
"0.3445"
],
[
"0.511",
"0.3289",
"0.5684",
"0.3289",
"0.5684",
"0.3403",
"0.511",
"0.3403"
],
[
"0.1184",
"0.3474",
"0.486",
"0.3474",
"0.486",
"0.361",
"0.1184",
"0.361"
],
[
"0.513",
"0.354",
"0.574",
"0.354",
"0.574",
"0.3652",
"0.513",
"0.3652"
],
[
"0.5684",
"0.3516",
"0.7983",
"0.354",
"0.7983",
"0.3682",
"0.5684",
"0.3665"
],
[
"0.1184",
"0.364",
"0.4844",
"0.364",
"0.4844",
"0.3777",
"0.1184",
"0.3777"
],
[
"0.5146",
"0.3735",
"0.879",
"0.3735",
"0.879",
"0.3872",
"0.5146",
"0.3872"
],
[
"0.12006",
"0.3801",
"0.147",
"0.3801",
"0.147",
"0.3938",
"0.12006",
"0.3938"
],
[
"0.2225",
"0.3943",
"0.378",
"0.3943",
"0.378",
"0.412",
"0.2225",
"0.412"
],
[
"0.5156",
"0.3896",
"0.882",
"0.3896",
"0.882",
"0.4026",
"0.5156",
"0.4026"
],
[
"0.513",
"0.405",
"0.879",
"0.405",
"0.879",
"0.4187",
"0.513",
"0.4187"
],
[
"0.1184",
"0.4192",
"0.51",
"0.4192",
"0.51",
"0.434",
"0.1184",
"0.434"
],
[
"0.511",
"0.4204",
"0.8716",
"0.4192",
"0.8716",
"0.4329",
"0.511",
"0.434"
],
[
"0.126",
"0.434",
"0.4207",
"0.434",
"0.4207",
"0.4478",
"0.126",
"0.4478"
],
[
"0.531",
"0.437",
"0.8774",
"0.437",
"0.8774",
"0.4507",
"0.531",
"0.4507"
],
[
"0.3594",
"0.452",
"0.4375",
"0.452",
"0.4375",
"0.462",
"0.3594",
"0.462"
],
[
"0.513",
"0.452",
"0.882",
"0.452",
"0.882",
"0.4697",
"0.513",
"0.4697"
],
[
"0.1368",
"0.466",
"0.4895",
"0.466",
"0.4895",
"0.4797",
"0.1368",
"0.4797"
],
[
"0.513",
"0.4702",
"0.677",
"0.4702",
"0.677",
"0.485",
"0.513",
"0.485"
],
[
"0.11755",
"0.4822",
"0.4836",
"0.481",
"0.4836",
"0.4954",
"0.11755",
"0.4963"
],
[
"0.11755",
"0.4976",
"0.4844",
"0.4963",
"0.4844",
"0.511",
"0.11755",
"0.5127"
],
[
"0.5786",
"0.4976",
"0.8135",
"0.4988",
"0.8135",
"0.517",
"0.5786",
"0.5166"
],
[
"0.11755",
"0.514",
"0.4836",
"0.514",
"0.4836",
"0.528",
"0.11755",
"0.528"
],
[
"0.11755",
"0.531",
"0.4878",
"0.531",
"0.4878",
"0.5444",
"0.11755",
"0.5444"
],
[
"0.4902",
"0.5337",
"0.5073",
"0.5337",
"0.5073",
"0.5415",
"0.4902",
"0.5415"
],
[
"0.51",
"0.532",
"0.8716",
"0.532",
"0.8716",
"0.546",
"0.51",
"0.546"
],
[
"0.529",
"0.5464",
"0.8774",
"0.5464",
"0.8774",
"0.56",
"0.529",
"0.56"
],
[
"0.11584",
"0.558",
"0.403",
"0.558",
"0.403",
"0.572",
"0.11584",
"0.572"
],
[
"0.51",
"0.5625",
"0.8774",
"0.5635",
"0.8774",
"0.5776",
"0.51",
"0.5767"
],
[
"0.11755",
"0.5806",
"0.4878",
"0.5806",
"0.4878",
"0.594",
"0.11755",
"0.594"
],
[
"0.513",
"0.58",
"0.879",
"0.58",
"0.879",
"0.5933",
"0.513",
"0.5933"
],
[
"0.11755",
"0.597",
"0.486",
"0.597",
"0.486",
"0.611",
"0.11755",
"0.611"
],
[
"0.513",
"0.5967",
"0.76",
"0.5967",
"0.76",
"0.61",
"0.513",
"0.61"
],
[
"0.12006",
"0.615",
"0.1561",
"0.615",
"0.1561",
"0.627",
"0.12006",
"0.627"
],
[
"0.559",
"0.6226",
"0.7363",
"0.627",
"0.7363",
"0.6455",
"0.559",
"0.641"
],
[
"0.1351",
"0.6313",
"0.486",
"0.6313",
"0.486",
"0.6445",
"0.1351",
"0.6445"
],
[
"0.1184",
"0.6465",
"0.4844",
"0.6465",
"0.4844",
"0.66",
"0.1184",
"0.66"
],
[
"0.601",
"0.6465",
"0.833",
"0.6465",
"0.833",
"0.661",
"0.601",
"0.661"
],
[
"0.1184",
"0.661",
"0.4844",
"0.662",
"0.4844",
"0.677",
"0.1184",
"0.676"
],
[
"0.1184",
"0.6787",
"0.486",
"0.6787",
"0.486",
"0.6924",
"0.1184",
"0.6924"
],
[
"0.532",
"0.677",
"0.881",
"0.677",
"0.881",
"0.6904",
"0.532",
"0.6904"
],
[
"0.1184",
"0.6953",
"0.486",
"0.6953",
"0.486",
"0.709",
"0.1184",
"0.709"
],
[
"0.51",
"0.6914",
"0.879",
"0.6914",
"0.879",
"0.705",
"0.51",
"0.705"
],
[
"0.1184",
"0.7104",
"0.4844",
"0.7104",
"0.4844",
"0.7236",
"0.1184",
"0.7236"
],
[
"0.513",
"0.7104",
"0.879",
"0.7104",
"0.879",
"0.72",
"0.513",
"0.72"
],
[
"0.1184",
"0.7246",
"0.157",
"0.728",
"0.1545",
"0.742",
"0.11584",
"0.7393"
],
[
"0.532",
"0.727",
"0.879",
"0.727",
"0.879",
"0.737",
"0.532",
"0.737"
],
[
"0.1368",
"0.7437",
"0.486",
"0.7437",
"0.486",
"0.7573",
"0.1368",
"0.7573"
],
[
"0.513",
"0.741",
"0.881",
"0.741",
"0.881",
"0.755",
"0.513",
"0.755"
],
[
"0.1184",
"0.7603",
"0.486",
"0.7603",
"0.486",
"0.774",
"0.1184",
"0.774"
],
[
"0.513",
"0.7583",
"0.881",
"0.7583",
"0.881",
"0.7715",
"0.513",
"0.7715"
],
[
"0.1184",
"0.7754",
"0.486",
"0.7754",
"0.486",
"0.789",
"0.1184",
"0.789"
],
[
"0.511",
"0.774",
"0.879",
"0.774",
"0.879",
"0.7876",
"0.511",
"0.7876"
],
[
"0.12006",
"0.7935",
"0.4844",
"0.7935",
"0.4844",
"0.804",
"0.12006",
"0.804"
],
[
"0.513",
"0.789",
"0.881",
"0.789",
"0.881",
"0.8027",
"0.513",
"0.8027"
],
[
"0.1184",
"0.808",
"0.486",
"0.808",
"0.486",
"0.822",
"0.1184",
"0.822"
],
[
"0.513",
"0.8057",
"0.881",
"0.8057",
"0.881",
"0.8193",
"0.513",
"0.8193"
],
[
"0.1184",
"0.8237",
"0.4844",
"0.8237",
"0.4844",
"0.8374",
"0.1184",
"0.8374"
],
[
"0.513",
"0.822",
"0.882",
"0.822",
"0.882",
"0.835",
"0.513",
"0.835"
],
[
"0.12006",
"0.8413",
"0.4836",
"0.8413",
"0.4836",
"0.8516",
"0.12006",
"0.8516"
],
[
"0.513",
"0.8374",
"0.642",
"0.8374",
"0.642",
"0.851",
"0.513",
"0.851"
],
[
"0.11755",
"0.8555",
"0.4844",
"0.8555",
"0.4844",
"0.869",
"0.11755",
"0.869"
],
[
"0.531",
"0.854",
"0.881",
"0.854",
"0.881",
"0.8677",
"0.531",
"0.8677"
],
[
"0.1184",
"0.8726",
"0.4475",
"0.8726",
"0.4475",
"0.8857",
"0.1184",
"0.8857"
],
[
"0.511",
"0.869",
"0.881",
"0.8706",
"0.881",
"0.8853",
"0.511",
"0.8843"
],
[
"0.1377",
"0.8926",
"0.486",
"0.8945",
"0.486",
"0.9077",
"0.1377",
"0.9062"
],
[
"0.534",
"0.8936",
"0.881",
"0.8936",
"0.881",
"0.9067",
"0.534",
"0.9067"
],
[
"0.12006",
"0.9077",
"0.33",
"0.9077",
"0.33",
"0.9185",
"0.12006",
"0.9185"
],
[
"0.513",
"0.9062",
"0.663",
"0.9062",
"0.663",
"0.9194",
"0.513",
"0.9194"
]
] | [
"越接近消失的边缘，我们的系统在其上的表现",
"在本节的其余部分中，我们将描述这两个模",
"就越差，因为数据派生的词汇表中存在的示例",
"块，每个模块都有不同的变体，具体取决于所",
"越少。我们的模型优于MakarovandClematide",
"使用的similarity的不同概念",
"(2018）为低资源设置而设计的模型，即使配备",
"3.1 Morphological Features Assignment",
"了额外的幻觉数据（AnastasopoulosandNeubig，",
"2019）。我们还超越了Jinetal.（2020）的最佳",
"为了评估一对看不见的单词W1,W2是否属于",
"模型，该模型不使用任何拐点表，并且对于大",
"同一词素，我们首先需要表征这两个单词之间",
"多数语言而言都是天际线",
"的关系。然后，必须将W1,W2之间的具体关系",
"我们得出的结论是，为资源匮乏的语言中的",
"与同一范式中2形态类别t,和t之间的抽象",
"屈折形态引导数据集是一种可以采用和探索的",
"关系Rtitk的某种表示进行比较。如果这些具",
"可行策略。3",
"体关系和抽象关系足够相似，则W1和W2将分",
"别标记为承载特征t,和tk。简而言之，这个",
"2",
"The Minimally-Supervised Setup",
"想法是通过汇总最小种子中所有n拐点表中",
"ti,ti条目的形式之间的差异来获得表示Rti,tk",
"ProblemStatement让C是一组词素，每个",
"词素对应于一个拐点表Wl=【wt..wtm}，其",
"。这些差异可以用语义、正字法或它们的组合",
"中 wt.是带有特征束t；的词素1的形式。我们",
"来表述",
"的目标是训练一个拐点模型，将具有一组特征",
"3.1.1",
"Orthography-Based Tagging",
"的单词映射到同一词素中具有另一组特征的单",
"在正交情况下，我们将一对单词之间的差异定",
"词。",
"((tj, wt,),th) -wti",
"义为从一个单词到另一个单词所需的编辑。我",
"们的系统期望的编辑是已删除的子字符串列表",
"例如，在法语语言范式中：",
"和已添加的子字符串列表，从左到右排序。4",
"（《INDPRSISG,finis),SBJVPST2SG)",
"对于每对形态类别（ti,t），它们的正交关",
"→finisses",
"系Rg,t由在n监督拐点表中观察到的一组编",
"为了诱导此功能，我们提出了一个minimally-",
"辑定义，每个引理li",
"supervised场景，其中仅给我们一个完整的拐",
"点表C={Wli...Wln）的n示例的小集合（每",
"Rl,t := [edits(wt, wt)]\"-1",
"个示例的大小为㎡），以及大量自然发生的",
"（即上下文中）未标记的数据，即MATHX，即",
"",
"之间的关系检查是否存在一对新词XMATHX",
"考虑来自法语语言范式的示例，其中关系",
"3TheAlgorithmic Framework",
"（INDPRS1SG，INDPRS3SG）的示例是（（finis",
"这项工作建议利用小监督种子中显示的模式，",
"finit)，（bois，but)，（parle，parla））。在这种",
"并找到显示similar（或analogous）模式的",
"情况下，此关系的表示形式为：",
"单词",
"RINDPRS1SG,INDPRS3SG =",
"这里提出的算法分两个阶段工作。首先，我",
"们标记具有形态特征的单词，如果它们被发现",
"[('s', ‘t'), (ois', 'ut'), (e',\"a'))",
"与最小监督中的示例相似（类似）。然后，我",
"们将标记的单词配对，以便每对都包含相同词",
"由于（aime，aima）之间的编辑与（parle，parla）",
"素的两种形式。这种算法分工允许配对模块提",
"之间的编辑相同，因此将aime正确地视为标",
"供健全性检查，并减少潜在的单词标记模块的",
"记为INDPRS1SG和aima-作为INDPSTPRF3SG",
"噪音",
"然而，这个过程很容易发生巧合。例如，le",
"该算法的上述草图相当通用，也就是说，我",
"和la之间的编辑与parle和parla之间的编辑",
"们可以通过以不同的方式插入来计算单词之间",
"相同，尽管前者实际上是确定者，而不是言语",
"的similarities或analogies，从而获得该框架",
"范式的一部分。鉴于可用的关系众多，我们可",
"的不同实例化。在我们的各种算法实现中，我",
"以期望incidental 对单词之间的许多编辑与黄",
"们将使用屈折形态中普遍存在的规律性，即使",
"金数据中看到的编辑相匹配。为了克服这个问",
"在少量的监督下也可以检测到。这些规则既可",
"题，我们建议考虑完整的范式结构，而不是孤",
"以用正字法表示，因为形式之间的编辑往往会",
"立地考虑单词对",
"在词素上重复出现，也可以用语义表示，因为",
"具体而言，我们建议仅标记已找到的",
"共享屈折功能的形式往往会被类似地使用",
"回答multiple relations in the same paradigm",
"3Our code is available on https://github.com/",
"4Though without position indices, in order to deal with",
"OnlpLab/morphodetection",
"words of various lengths."
] | {
"arxiv_id": "2104.08512",
"language": "zh",
"page_idx": 2
} |
||
2104.08512-zh-3 | 2104.08512-zh-3.jpg | 此标准的单词，至少覆盖范式大小的一半。所以 aime 将被标记为 INDPRS1SG ,因$\begin{array}{rcl}{\text{为 edits}(aime,aima)}&{=\mathrm{~edits}(parle,parla)}\\{\mathrm{edits}(aime,aimons)}&{=\mathrm{~edits}(parle,parlons)}\end{array}$ 等等，但是 le 不会被标记为 INDPRS1SG ,因为法语词汇不包含 lons
有了这个严格的标准，正交算法可能足够精
确，但可能不够多样化，以便获得高召回率
## 3.1.2 Semantics-Based Tagging
正交标准仅考虑与观察到的编辑的精确匹配， 因此预计会错过$n$监督变化表中未经证实的不规则和类。$^{5}$这可能会给具有多个 $n$ 类的范式带来重大问题，或者显示标记示例中不存在的重要形态语音过程
为了克服正字法编辑的泛化问题，我们建议考虑语义规则性，因为含义和用法的差异很少与正字法相关。语义规律性源于一致，即句子中的单词必须与句子中的其他单词具有相同的形态特征，从而有效地创建了对等类。在 Firth (1957)之后，用于从共现提取语义的单词嵌入的现代算法自然适合于利用这种规律性
在此设置中，单词之间的差异定义为其嵌入向量之间的差异。对于每对形态特征束$(t_j,t_k)$ , 它们的语义关系的表示是通过这些相关示例的平均值来估计的
$$\hat{R}_{t_j,t_k}^S=\dfrac{1}{n}\sum_{i=1}^nv(w_{t_j}^{l_i})-v(w_{t_k}^{l_i})$$
如果新词对的差异在余弦距离方面与$\hat{R}_{t_j,t_k}^S$
$$D_C\big(\hat R_{t_j,t_k}^S,v(w_a)-v(w_b)\big)\stackrel{?}{\le}\hat C_{t_j,t_k}^S$$
足够接近，则将其标记为$t_j,t_k$ ,其中 $D_C$ 是余弦距离函数，而$\hat{C}_{t_j,t_k}^S$是通过特定于关系的平均 XMATHX 散度
虽然没有正交的缺点，但这里的$\hat{R}_{t_j,t_k}^S$可能是一个有偏见的表示，错过了许多例子，或者错误地标记了不正确的单词。出于这个原因， 我们建议第三种算法结合两种类型或规则
# 3.1.3 Combined Tagging
此变体背后的想法是将回答正字法和语义标准的单词对视为半黄金示例，将添加这些示例以更好地估计关系$\hat{R}_{t_j,t_k}^S$。因此，我们利用正交标准的准确性来对抗语义表示中的偏差
$^{5}$the term‘class’refers to a group of lexemes in a paradigm that display similar inflection patterns. Traditionally known as ‘conjugation’and‘declension’in the description of verbal and nominal paradigms, respectively. E.g., the Spanish verbal paradigm is said to include 3 classes: -er , -ar and -ir verbs.
具体来说，如果一对单词$(w_1,w_2)$的编辑脚本在$\hat{R}_{t_j,t_k}^O$中，并且它们的差异向量与关系向量 $\hat{R}_{t_j,t_k}^S$的距离小于黄金示例的最远对应差异向量，则将其视为半黄金
请注意，与前几节中的标准相比，我们放宽了这两个标准。通过放弃对完整范例的要求来放宽正交标准，并通过将$\hat{C}_{t_j,t_k}^S$替换为更具包容性的截止值而不是平均距离来放宽语义标准
$$\hat C_{t_j,t_k}^{Comb}=\max_iD_C\big(\hat R_{t_j,t_k}^s,v(w_{t_j}^{l_i})-v(w_{t_k}^{l_i})\big)$$
松弛允许包含更多的半黄金示例，并且它们
是明智的，因为两个标准都是相互约束的
满足两个条件的单词对很可能是形态关系$(t_j,t_k)$的正确示例。因此，它们被添加以创建新的语义表示$\hat{R}_{t_j,t_k}^S$。新的表示形式可能会再次用于查找更多的半黄金示例，迭代执行此阶段。当迭代后没有向$R_{t_j,t_k}^S$添加示例时，或者当添加了如此多的示例以使每次迭代的运行时间在我们的实现中超过 48 小时时，我们设置停止标准
一旦确定了$\hat{R} _{t_j, t_k}^S$ ,并且为了包括对具有不同编辑的单词进行标记，我们最终根据语义标准单独使用相应的$\hat{C}_{t_j,t_k}^S$对单词进行标记
# 3.2 Pairing Tagged Words
给定第一个模块的输出，用一些形态学特征集标记的单词列表，第二个模块对来自第一个模块的标记单词进行配对，使得两个标记单词都是同一词素的形式
我们首先将来自第一个模块的标记单词分组到$m$容器$\{B_t_i\}_{i=1}^m$中，每个容器包含所有标记有$t_i$的单词。然后，我们需要为每个单词$w_{t_i}^{l_j}\in B_{t_i}$找到来自其他 bin 的相应单词，这些单词是相同词素$l_j$的形式
理想情况下，如果每个 bin 包含每个词素的一个正确形式，则将最相似的单词在 bin 中配对就足以收集范式中的所有形式。实际上，由于噪声，bin 可能包含几种形式的相同词素或根本没有。假设标记了足够多的单词，似乎最好简单地丢弃几种形式的相同词素占用同一bin 的情况，而不是试图找到正确标记的单词。因此，我们想对单词进行配对，使它们在 bin 中是不同的最近邻居，即。e。,其中第二个最近的邻居要远得多
为了实现这一点，我们使用跨域相似性局部缩放 (CSLS)分数来缩放距离，该分数由Lample et al. (2018) 在双语言嵌人的上下文中建议。CSLS 分数将跨$x\in X,y\in Y$组的向量之间的余弦距离与其顶部$k$最近邻居的平均距离进行缩放。在这里应用此方法时，我们 | [
[
"0.12006",
"0.0885",
"0.465",
"0.0885",
"0.465",
"0.0986",
"0.12006",
"0.0986"
],
[
"0.461",
"0.08905",
"0.482",
"0.08905",
"0.482",
"0.0968",
"0.461",
"0.0968"
],
[
"0.534",
"0.0885",
"0.879",
"0.0885",
"0.879",
"0.0986",
"0.534",
"0.0986"
],
[
"0.1184",
"0.1027",
"0.4878",
"0.1027",
"0.4878",
"0.1164",
"0.1184",
"0.1164"
],
[
"0.4954",
"0.1069",
"0.5054",
"0.1069",
"0.5054",
"0.10986",
"0.4954",
"0.10986"
],
[
"0.51",
"0.1027",
"0.879",
"0.1027",
"0.879",
"0.11755",
"0.51",
"0.11755"
],
[
"0.12006",
"0.12054",
"0.4844",
"0.12054",
"0.4844",
"0.1312",
"0.12006",
"0.1312"
],
[
"0.513",
"0.1223",
"0.881",
"0.1223",
"0.881",
"0.136",
"0.513",
"0.136"
],
[
"0.1184",
"0.1342",
"0.486",
"0.1342",
"0.486",
"0.1478",
"0.1184",
"0.1478"
],
[
"0.5146",
"0.1414",
"0.7153",
"0.1414",
"0.7153",
"0.152",
"0.5146",
"0.152"
],
[
"0.1184",
"0.1509",
"0.4844",
"0.1509",
"0.4844",
"0.1644",
"0.1184",
"0.1644"
],
[
"0.531",
"0.1561",
"0.882",
"0.1561",
"0.882",
"0.1692",
"0.531",
"0.1692"
],
[
"0.1184",
"0.1675",
"0.3015",
"0.1675",
"0.3015",
"0.1812",
"0.1184",
"0.1812"
],
[
"0.519",
"0.1733",
"0.879",
"0.1733",
"0.879",
"0.1841",
"0.519",
"0.1841"
],
[
"0.1368",
"0.1829",
"0.486",
"0.1829",
"0.486",
"0.1965",
"0.1368",
"0.1965"
],
[
"0.51",
"0.1853",
"0.882",
"0.187",
"0.882",
"0.2048",
"0.51",
"0.2031"
],
[
"0.11755",
"0.1978",
"0.4685",
"0.199",
"0.4685",
"0.2131",
"0.11755",
"0.212"
],
[
"0.5146",
"0.206",
"0.881",
"0.206",
"0.881",
"0.2197",
"0.5146",
"0.2197"
],
[
"0.11584",
"0.2227",
"0.3787",
"0.2239",
"0.3787",
"0.2382",
"0.11584",
"0.2369"
],
[
"0.525",
"0.2268",
"0.864",
"0.2268",
"0.864",
"0.2588",
"0.525",
"0.2588"
],
[
"0.12006",
"0.2446",
"0.4836",
"0.2446",
"0.4836",
"0.2546",
"0.12006",
"0.2546"
],
[
"0.1184",
"0.2588",
"0.486",
"0.2588",
"0.486",
"0.2725",
"0.1184",
"0.2725"
],
[
"0.1184",
"0.275",
"0.486",
"0.275",
"0.486",
"0.288",
"0.1184",
"0.288"
],
[
"0.532",
"0.2708",
"0.881",
"0.2708",
"0.881",
"0.284",
"0.532",
"0.284"
],
[
"0.1184",
"0.2915",
"0.486",
"0.2915",
"0.486",
"0.3047",
"0.1184",
"0.3047"
],
[
"0.511",
"0.2874",
"0.8403",
"0.2874",
"0.8403",
"0.2976",
"0.511",
"0.2976"
],
[
"0.532",
"0.3027",
"0.881",
"0.3027",
"0.881",
"0.3164",
"0.532",
"0.3164"
],
[
"0.1184",
"0.3088",
"0.2822",
"0.3088",
"0.2822",
"0.3196",
"0.1184",
"0.3196"
],
[
"0.1351",
"0.3225",
"0.4878",
"0.3218",
"0.4878",
"0.3362",
"0.1351",
"0.3374"
],
[
"0.511",
"0.3196",
"0.881",
"0.3196",
"0.881",
"0.3333",
"0.511",
"0.3333"
],
[
"0.1184",
"0.339",
"0.4878",
"0.339",
"0.4878",
"0.3528",
"0.1184",
"0.3528"
],
[
"0.513",
"0.335",
"0.881",
"0.335",
"0.881",
"0.3486",
"0.513",
"0.3486"
],
[
"0.1184",
"0.3557",
"0.4878",
"0.3557",
"0.4878",
"0.3694",
"0.1184",
"0.3694"
],
[
"0.511",
"0.3528",
"0.879",
"0.3528",
"0.879",
"0.3665",
"0.511",
"0.3665"
],
[
"0.1184",
"0.3718",
"0.4878",
"0.3718",
"0.4878",
"0.3855",
"0.1184",
"0.3855"
],
[
"0.513",
"0.3694",
"0.879",
"0.3694",
"0.879",
"0.383",
"0.513",
"0.383"
],
[
"0.1184",
"0.3884",
"0.4844",
"0.3884",
"0.4844",
"0.3984",
"0.1184",
"0.3984"
],
[
"0.513",
"0.386",
"0.879",
"0.386",
"0.879",
"0.3997",
"0.513",
"0.3997"
],
[
"0.1184",
"0.4038",
"0.4844",
"0.4038",
"0.4844",
"0.4175",
"0.1184",
"0.4175"
],
[
"0.513",
"0.402",
"0.881",
"0.402",
"0.881",
"0.415",
"0.513",
"0.415"
],
[
"0.11755",
"0.4192",
"0.4282",
"0.4192",
"0.4282",
"0.4329",
"0.11755",
"0.4329"
],
[
"0.5146",
"0.4192",
"0.586",
"0.4192",
"0.586",
"0.4312",
"0.5146",
"0.4312"
],
[
"0.1368",
"0.4358",
"0.4836",
"0.4358",
"0.4836",
"0.4495",
"0.1368",
"0.4495"
],
[
"0.544",
"0.4329",
"0.881",
"0.434",
"0.881",
"0.449",
"0.544",
"0.4478"
],
[
"0.11755",
"0.4507",
"0.4878",
"0.452",
"0.4878",
"0.466",
"0.11755",
"0.4656"
],
[
"0.511",
"0.4507",
"0.881",
"0.452",
"0.881",
"0.466",
"0.511",
"0.4656"
],
[
"0.1318",
"0.4685",
"0.4836",
"0.4685",
"0.4836",
"0.4822",
"0.1318",
"0.4822"
],
[
"0.511",
"0.4685",
"0.8423",
"0.4685",
"0.8423",
"0.4827",
"0.511",
"0.4827"
],
[
"0.1184",
"0.484",
"0.266",
"0.484",
"0.266",
"0.4976",
"0.1184",
"0.4976"
],
[
"0.1293",
"0.4739",
"0.1368",
"0.4739",
"0.1368",
"0.478",
"0.1293",
"0.478"
],
[
"0.511",
"0.4934",
"0.735",
"0.4946",
"0.735",
"0.509",
"0.511",
"0.5083"
],
[
"0.241",
"0.5073",
"0.424",
"0.5186",
"0.419",
"0.5474",
"0.236",
"0.536"
],
[
"0.1855",
"0.5166",
"0.2477",
"0.53",
"0.241",
"0.5464",
"0.178",
"0.533"
],
[
"0.511",
"0.5166",
"0.8774",
"0.5166",
"0.8774",
"0.53",
"0.511",
"0.53"
],
[
"0.2377",
"0.524",
"0.3015",
"0.5376",
"0.293",
"0.5547",
"0.2284",
"0.5415"
],
[
"0.513",
"0.532",
"0.881",
"0.531",
"0.881",
"0.546",
"0.513",
"0.5464"
],
[
"0.513",
"0.549",
"0.881",
"0.549",
"0.881",
"0.5625",
"0.513",
"0.5625"
],
[
"0.1351",
"0.561",
"0.4768",
"0.5635",
"0.4768",
"0.5806",
"0.1351",
"0.579"
],
[
"0.513",
"0.5664",
"0.66",
"0.5664",
"0.66",
"0.5767",
"0.513",
"0.5767"
],
[
"0.529",
"0.58",
"0.881",
"0.5806",
"0.881",
"0.5957",
"0.529",
"0.594"
],
[
"0.1628",
"0.597",
"0.4382",
"0.5996",
"0.4382",
"0.6216",
"0.1628",
"0.619"
],
[
"0.513",
"0.597",
"0.881",
"0.597",
"0.881",
"0.611",
"0.513",
"0.611"
],
[
"0.5156",
"0.615",
"0.881",
"0.615",
"0.881",
"0.626",
"0.5156",
"0.626"
],
[
"0.1184",
"0.6313",
"0.486",
"0.6313",
"0.486",
"0.6445",
"0.1184",
"0.6445"
],
[
"0.511",
"0.6313",
"0.881",
"0.6313",
"0.881",
"0.6455",
"0.511",
"0.6455"
],
[
"0.11755",
"0.6455",
"0.4895",
"0.6465",
"0.4895",
"0.6646",
"0.11755",
"0.663"
],
[
"0.4978",
"0.652",
"0.511",
"0.652",
"0.511",
"0.6562",
"0.4978",
"0.6562"
],
[
"0.511",
"0.6465",
"0.7188",
"0.6465",
"0.7188",
"0.66",
"0.511",
"0.66"
],
[
"0.12006",
"0.665",
"0.2795",
"0.665",
"0.2795",
"0.676",
"0.12006",
"0.676"
],
[
"0.529",
"0.663",
"0.881",
"0.663",
"0.881",
"0.677",
"0.529",
"0.677"
],
[
"0.1368",
"0.681",
"0.486",
"0.681",
"0.486",
"0.6953",
"0.1368",
"0.6953"
],
[
"0.5146",
"0.681",
"0.531",
"0.681",
"0.531",
"0.6885",
"0.5146",
"0.6885"
],
[
"0.5264",
"0.6777",
"0.879",
"0.6777",
"0.879",
"0.6914",
"0.5264",
"0.6914"
],
[
"0.1184",
"0.6987",
"0.4844",
"0.6987",
"0.4844",
"0.709",
"0.1184",
"0.709"
],
[
"0.511",
"0.695",
"0.879",
"0.6953",
"0.879",
"0.709",
"0.511",
"0.708"
],
[
"0.11755",
"0.7134",
"0.4895",
"0.7134",
"0.4895",
"0.727",
"0.11755",
"0.727"
],
[
"0.513",
"0.7114",
"0.881",
"0.7114",
"0.881",
"0.7246",
"0.513",
"0.7246"
],
[
"0.11755",
"0.731",
"0.4492",
"0.731",
"0.4492",
"0.7446",
"0.11755",
"0.7446"
],
[
"0.513",
"0.728",
"0.881",
"0.728",
"0.881",
"0.741",
"0.513",
"0.741"
],
[
"0.5146",
"0.745",
"0.8623",
"0.745",
"0.8623",
"0.756",
"0.5146",
"0.756"
],
[
"0.11584",
"0.7544",
"0.3267",
"0.756",
"0.3267",
"0.771",
"0.11584",
"0.7686"
],
[
"0.867",
"0.749",
"0.876",
"0.749",
"0.876",
"0.753",
"0.867",
"0.753"
],
[
"0.513",
"0.7603",
"0.881",
"0.7603",
"0.881",
"0.774",
"0.513",
"0.774"
],
[
"0.1184",
"0.775",
"0.486",
"0.775",
"0.486",
"0.788",
"0.1184",
"0.788"
],
[
"0.513",
"0.777",
"0.881",
"0.777",
"0.881",
"0.7905",
"0.513",
"0.7905"
],
[
"0.1184",
"0.789",
"0.486",
"0.789",
"0.486",
"0.8027",
"0.1184",
"0.8027"
],
[
"0.513",
"0.792",
"0.882",
"0.792",
"0.882",
"0.807",
"0.513",
"0.807"
],
[
"0.1184",
"0.807",
"0.486",
"0.807",
"0.486",
"0.822",
"0.1184",
"0.822"
],
[
"0.511",
"0.8086",
"0.6616",
"0.8086",
"0.6616",
"0.8223",
"0.511",
"0.8223"
],
[
"0.1184",
"0.8237",
"0.43",
"0.8237",
"0.43",
"0.8374",
"0.1184",
"0.8374"
],
[
"0.531",
"0.8247",
"0.881",
"0.8247",
"0.881",
"0.8384",
"0.531",
"0.8384"
],
[
"0.1368",
"0.844",
"0.4878",
"0.8457",
"0.4878",
"0.86",
"0.1368",
"0.8594"
],
[
"0.511",
"0.839",
"0.879",
"0.8403",
"0.879",
"0.855",
"0.511",
"0.854"
],
[
"0.12006",
"0.86",
"0.4878",
"0.86",
"0.4878",
"0.8706",
"0.12006",
"0.8706"
],
[
"0.51",
"0.857",
"0.879",
"0.857",
"0.879",
"0.8706",
"0.51",
"0.8706"
],
[
"0.1184",
"0.8706",
"0.4895",
"0.8706",
"0.4895",
"0.8843",
"0.1184",
"0.8843"
],
[
"0.511",
"0.8716",
"0.881",
"0.8726",
"0.881",
"0.887",
"0.511",
"0.8857"
],
[
"0.1184",
"0.8843",
"0.4844",
"0.8843",
"0.4844",
"0.8945",
"0.1184",
"0.8945"
],
[
"0.11755",
"0.8945",
"0.486",
"0.8945",
"0.486",
"0.9077",
"0.11755",
"0.9077"
],
[
"0.511",
"0.8896",
"0.879",
"0.8896",
"0.879",
"0.903",
"0.511",
"0.903"
],
[
"0.12006",
"0.9067",
"0.1957",
"0.9067",
"0.1957",
"0.9175",
"0.12006",
"0.9175"
],
[
"0.511",
"0.904",
"0.881",
"0.905",
"0.881",
"0.9185",
"0.511",
"0.9175"
]
] | [
"此标准的单词，至少覆盖范式大小的",
"",
"真体来说，如果一对单词（w12）的编辑脚",
"半。所以aime将被标记为INDPRSISG，因",
"",
"本在Rt,t中，并且它们的差异向量与关系向",
"为edits(aime,aima)=edits(parle,parla）和",
"量Rt的距离小于黄金示例的最远对应差异",
"edits(aime,aimons) = edits(parle,parlons)",
"向量，则将其视为半黄金",
"等等，但是le不会被标记为INDPRS1SG，因",
"请注意，与前几节中的标准相比，我们放宽",
"为法语词汇不包含lons",
"了这两个标准。通过放弃对完整范例的要求来",
"有了这个严格的标准，正交算法可能足够精",
"放宽正交标准，并通过将C替换为更具包",
"确，但可能不够多样化，以便获得高召回率",
"容性的截止值而不是平均距离来放宽语义标准",
"3.1.2Semantics-Based Tagging",
"Cg=maxDe(f(g)-o()",
"正交标准仅考虑与观察到的编辑的精确匹配",
"因此预计会错过n监督变化表中未经证实的",
"不规则和类。5这可能会给具有多个n类的范",
"松弛允许包含更多的半黄金示例，并且它们",
"式带来重大问题，或者显示标记示例中不存在",
"是明智的，因为两个标准都是相互约束的",
"满足两个条件的单词对很可能是形态关系",
"的重要形态语音过程",
"为了克服正字法编辑的泛化问题，我们建议",
"(ti，tk）的正确示例。因此，它们被添加以创建",
"考虑语义规则性，因为含义和用法的差异很少",
"新的语义表示RS。新的表示形式可能会再",
"与正字法相关。语义规律性源于一致，即句子",
"次用于查找更多的半黄金示例，迭代执行此阶",
"中的单词必须与句子中的其他单词具有相同的",
"段。当迭代后没有向Rtt添加示例时，或者",
"形态特征，从而有效地创建了对等类。在Firth",
"当添加了如此多的示例以使每次送代的运行时",
"(1957)之后，用于从共现提取语义的单词嵌人",
"间在我们的实现中超过48小时时，我们设置",
"的现代算法自然适合于利用这种规律性",
"停正标准",
"在此设置中，单词之间的差异定义为其嵌入",
"且确定了Rt，并且为了包括对具有不",
"向量之间的差异。对于每对形态特征束（ti，ts)",
"同编辑的单词进行标记，我们最终根据语义标",
"它们的语义关系的表示是通过这些相关示例",
"准单独使用相应的Ctt对单词进行标记",
"的平均值来估计的",
"",
"3.2Pairing Tagged Words",
"12o(utg) - (ut)",
"Rt,tk",
"给定第一个模块的输出，用一些形态学特征集",
"n=",
"标记的单词列表，第二个模块对来自第一个模",
"块的标记单词进行配对，使得两个标记单词都",
"如果新词对的差异在余弦距离方面与Rt,t",
"是同一词素的形式",
"我们首先将来自第一个模块的标记单词分组",
"Dc(R,, (wa) - v(wb) ≥C号,tk",
"到m容器[Bt,]m中，每个容器包含所有标",
"记有t的单词。然后，我们需要为每个单词",
"足够接近，则将其标记为ti,th，其中Dc是",
"wtEBt找到来自其他bin的相应单词，这些",
"余弦距离函数，而C,t是通过特定于关系的",
"",
"单词是相同词素l；的形式",
"平均XMATHX散度",
"理想情况下，如果每个bin包含每个词素的",
"虽然没有正交的缺点，但这里的Rt可能",
"",
"个正确形式，则将最相似的单词在bin中配",
"是一个有偏见的表示，错过了许多例子，或者",
"对就足以收集范式中的所有形式。实际上，由",
"错误地标记了不正确的单词。出于这个原因，",
"于噪声，bin可能包含几种形式的相同词素或",
"我们建议第三种算法结合两种类型或规则",
"根本没有。假设标记了足够多的单词，似乎最",
"好简单地丢弃几种形式的相同词素占用同",
"3.1.3Combined Tagging",
"-",
"bin的情况，而不是试图找到正确标记的单词。",
"此变体背后的想法是将回答正字法和语义标准",
"因此，我们想对单词进行配对，使它们在bin",
"的单词对视为半黄金示例，将添加这些示例以",
"中是不同的最近邻居，即。e。，其中第二个最",
"更好地估计关系Rt,ts。因此，我们利用正交",
"近的邻居要远得多",
"标准的准确性来对抗语义表示中的偏差",
"为了实现这一点，我们使用跨域相似性局",
"'the term 'class' refers to a group of lexemes in a",
"部缩放（CSLS）分数来缩放距离，该分数由",
"paradigm that display similar inflection patterns.Tradition-",
"Lampleetal.(2018）在双语言嵌人的上下文中",
"ally known as 'conjugation' and 'declension' in the descrip-",
"建议。CSLS分数将跨&EX,EY组的向",
"tion of verbal and nominal paradigms, respectively.E.g., the",
"Spanish verbal paradigm is said to include 3 classes: -er , -ar",
"量之间的余弦距离与其顶部k最近邻居的平",
"and-irverbs",
"均距离进行缩放。在这里应用此方法时，我们"
] | {
"arxiv_id": "2104.08512",
"language": "zh",
"page_idx": 3
} |
||
2104.08512-zh-4 | 2104.08512-zh-4.jpg | $$CSLS(x,y)=D(x,y)-\\\frac{1}{2}D(x,NN_{2}^{x}(Y))-\frac{1}{2}D(NN_{2}^{y}(X),y)$$
置。$k=2$的 CSLS 定义可以写成
,其中$NN_2^x(Y)$是$x$在组$Y$中的第二个最近
邻居
尽管 Lample et al. (2018) 使用的原始距离度量$D(\cdot,\cdot)$是余弦距离函数，但任何距离度量都可以。在语义算法中(Sec。3.1.2),我们应用反映语义相似度的余弦距离。在正交算法中(Sec。3.1.1 ), 我们插入 Levenshtein 编辑距离， 以利用拼字法，而不是语义
在对所有可能的单词对进行评分之后，将得分最高的对作为有监督的神经形态偏转器的训练集
# 4 Experimental Evaluation
4.1 Experimental Setup
我们开始使用不同类型的规则，从最少数量的标记拐点表开始，以经验方式检查是否可以训练形态拐点模型。我们的实验包括八种语言的口头范式：英语，俄语，拉脱维亚语，西班牙语，法语，芬兰语，匈牙利语和土耳其语，尽管我们的方法适用于任何范式
为了使评估拐点模型成为可能，我们必须选择具有足够数量的黄金(测试)数据的语言， 并模拟低资源场景进行训练。这限制了我们使用西方语言，我们的目标是尽可能多地使用非印欧语言
The Unlabeled Data 我们在第 2节中指定的问题设置指定使用大量未标记的文本。实际上， 所提出的算法利用文本(i)收集词汇以寻找标记候选者，以及(ii)训练嵌人以用于计算候选者对之间的语义关系。在我们的实验中，我们简单地采用特定于语言的预训练单词嵌人来实现这两个目的。我们使用了 Grave et al.(2018) 提供的预先训练的 FastText 向量，据报道，它们在更大的类比测试集 (Li et al., 2018) 中成功超过了形态类比。我们将 2 百万长的 FastText 词汇表剪裁为仅包含真实单词，即。e。,我们只保留不包含非字母字符的小写标记，并至少包含一个元音。$^{6}$此过程将词汇量缩减为每种
$^{6}$This last criterion happens to be suitable for all 8 languages examined. Admittedly there are languages, like $\bar{\mathrm{Czech,~that~allow~less~sonorant~syllable~nucleus.}}$
设置了$k=2$ ,因为我们的目标是丢弃表单 语言约 $200k-500k\text{ 个单词。}^{7}$此外，由于运行$w_{t_i}^{l_q}\in B_{t_i}$在另一个 bin $w_{t_j}^{l_q},w_{t_k}^{l_q}\in B_{t_j}$中有两 时的原因，我们将正字法变体的词汇表大小限个相应表单的情况，其中一个可能放错了位 制为每种语言 200$k$个单词
Minimal Supervision Source 对于每种语言， 我们都从 UniMorph (Kirov et al., 2018) 中提取了 5 词素的拐点表。我们的目标是经常使用的词素，具有强大的嵌入表示形式，并且来自不同的类，以尽可能多地捕获语言的语言行为。为此，我们以嵌入词汇表中出现的形式最多的词素为目标，并从它们中手动选择，使它们尽可能属于不同的类。$^{8}$
The Inflection Model 作为我们的监督拐点
模型，要在引导数据上进行训练，我们使用了开箱即用的序列到序列 LSTM 模型，并注意输入的字符和源和目标形式的特征。我们使用了 Silfverberg and Hulden (2018) 的公开可用模型。我们在最小监督场景中的系统输出的多达 10,000 数据训练示例或监督场景中的 6,000训练示例上训练了 50 个时期的模型。后者为我们的最小监督系统的性能提供了一个上限。Silfverberg and Hulden (2018)提供了此处检查的语言的 4 的培训和测试数据，我们手工制作了英语，俄语，匈牙利语和土耳其语的类似数据集
模型的评估指标是输出字符串与黄金输出的
精确匹配
Models, Baseline and Skyline 我们报告了我
们测试的三个系统变体的结果：
$\bullet$ ORTH: The orthography-based system ( §3.1.1
);
$\cdot$ SEM : The embedings-based system (Sedings-based system (Sedings The embedings-based system (Sedings preding preding preding preding preding preding preding preding preding preding preding preding predin preding predin preding predin predin predin pre
3.1.2 );
$\bullet$ COMB: The combined system (Sec. 3.1.3 ).
此外，为了评估系统中的附加值，我们包括了在低资源环境中训练的两个基线模型的准确性，而没有找到新的示例：THMAX2
我们还将以完全监督方式(SUP)训练的模
型的性能作为上限
To reduce run time the Finnish and Hungarian vocabularies were further reduced to include only words that appeared at least twice in the Finnish Parole and Hungarian Gigaword corpora (Bartis, 1998; Oravecz et al., 2014) .
$^{8}$Cases of syncretism were trivially solved by merging 2 categories into one category with 2 tags if all supervision lexemes had the same forms in those 2 categories. This strategy might pose a problem in cases of partial syncretism, where categories differ in forms only in classes that happen to be absent from the minimal supervision. The manually enforced class diversity in the selection process was devised to solve this problem as well. | [
[
"0.11755",
"0.0849",
"0.486",
"0.0849",
"0.486",
"0.0986",
"0.11755",
"0.0986"
],
[
"0.511",
"0.0849",
"0.879",
"0.0861",
"0.879",
"0.10095",
"0.511",
"0.0998"
],
[
"0.1184",
"0.1027",
"0.486",
"0.1027",
"0.486",
"0.11755",
"0.1184",
"0.11755"
],
[
"0.511",
"0.1027",
"0.881",
"0.1027",
"0.881",
"0.1164",
"0.511",
"0.1164"
],
[
"0.1184",
"0.1194",
"0.4878",
"0.1194",
"0.4878",
"0.1342",
"0.1184",
"0.1342"
],
[
"0.513",
"0.11816",
"0.7256",
"0.11816",
"0.7256",
"0.1318",
"0.513",
"0.1318"
],
[
"0.1184",
"0.1372",
"0.3855",
"0.1372",
"0.3855",
"0.1509",
"0.1184",
"0.1509"
],
[
"0.5156",
"0.1509",
"0.881",
"0.1509",
"0.881",
"0.1644",
"0.5156",
"0.1644"
],
[
"0.1301",
"0.165",
"0.3208",
"0.1675",
"0.3208",
"0.1818",
"0.1301",
"0.1799"
],
[
"0.5156",
"0.1692",
"0.879",
"0.1692",
"0.879",
"0.1799",
"0.5156",
"0.1799"
],
[
"0.1855",
"0.1859",
"0.2007",
"0.1907",
"0.1957",
"0.199",
"0.1814",
"0.1948"
],
[
"0.3435",
"0.187",
"0.351",
"0.1924",
"0.345",
"0.1965",
"0.3376",
"0.1912"
],
[
"0.517",
"0.1841",
"0.881",
"0.1841",
"0.881",
"0.1978",
"0.517",
"0.1978"
],
[
"0.1721",
"0.1912",
"0.34",
"0.1818",
"0.3435",
"0.2114",
"0.1755",
"0.2203"
],
[
"0.34",
"0.1936",
"0.471",
"0.1936",
"0.471",
"0.2079",
"0.34",
"0.2079"
],
[
"0.3315",
"0.2007",
"0.339",
"0.2007",
"0.339",
"0.2073",
"0.3315",
"0.2073"
],
[
"0.3376",
"0.206",
"0.351",
"0.206",
"0.351",
"0.2144",
"0.3376",
"0.2144"
],
[
"0.5146",
"0.2019",
"0.8774",
"0.2019",
"0.8774",
"0.212",
"0.5146",
"0.212"
],
[
"0.5146",
"0.2162",
"0.881",
"0.2162",
"0.881",
"0.2299",
"0.5146",
"0.2299"
],
[
"0.1351",
"0.2299",
"0.486",
"0.2299",
"0.486",
"0.2434",
"0.1351",
"0.2434"
],
[
"0.513",
"0.2322",
"0.879",
"0.2322",
"0.879",
"0.2452",
"0.513",
"0.2452"
],
[
"0.12006",
"0.2452",
"0.1545",
"0.2452",
"0.1545",
"0.2583",
"0.12006",
"0.2583"
],
[
"0.5146",
"0.2494",
"0.876",
"0.2494",
"0.876",
"0.26",
"0.5146",
"0.26"
],
[
"0.1351",
"0.2612",
"0.486",
"0.2625",
"0.486",
"0.2766",
"0.1351",
"0.2756"
],
[
"0.51",
"0.2632",
"0.69",
"0.26",
"0.69",
"0.2778",
"0.51",
"0.2808"
],
[
"0.1184",
"0.279",
"0.4844",
"0.279",
"0.4844",
"0.2922",
"0.1184",
"0.2922"
],
[
"0.1184",
"0.2944",
"0.486",
"0.2944",
"0.486",
"0.308",
"0.1184",
"0.308"
],
[
"0.517",
"0.2957",
"0.881",
"0.2976",
"0.881",
"0.3113",
"0.517",
"0.3088"
],
[
"0.1184",
"0.3113",
"0.4844",
"0.3113",
"0.4844",
"0.3247",
"0.1184",
"0.3247"
],
[
"0.511",
"0.3123",
"0.8774",
"0.3123",
"0.8774",
"0.326",
"0.511",
"0.326"
],
[
"0.12006",
"0.3289",
"0.4895",
"0.3289",
"0.4895",
"0.339",
"0.12006",
"0.339"
],
[
"0.511",
"0.33",
"0.8774",
"0.33",
"0.8774",
"0.3403",
"0.511",
"0.3403"
],
[
"0.1184",
"0.3433",
"0.3376",
"0.3433",
"0.3376",
"0.357",
"0.1184",
"0.357"
],
[
"0.513",
"0.3457",
"0.881",
"0.3457",
"0.881",
"0.3594",
"0.513",
"0.3594"
],
[
"0.1368",
"0.3599",
"0.486",
"0.3599",
"0.486",
"0.3735",
"0.1368",
"0.3735"
],
[
"0.513",
"0.3623",
"0.879",
"0.3623",
"0.879",
"0.376",
"0.513",
"0.376"
],
[
"0.11755",
"0.376",
"0.486",
"0.376",
"0.486",
"0.3896",
"0.11755",
"0.3896"
],
[
"0.513",
"0.3777",
"0.881",
"0.3777",
"0.881",
"0.3914",
"0.513",
"0.3914"
],
[
"0.11755",
"0.3938",
"0.1561",
"0.3938",
"0.1561",
"0.4062",
"0.11755",
"0.4062"
],
[
"0.5146",
"0.3938",
"0.882",
"0.3938",
"0.882",
"0.4067",
"0.5146",
"0.4067"
],
[
"0.5146",
"0.411",
"0.8774",
"0.411",
"0.8774",
"0.4216",
"0.5146",
"0.4216"
],
[
"0.11755",
"0.4229",
"0.1511",
"0.4229",
"0.1511",
"0.4353",
"0.11755",
"0.4353"
],
[
"0.1503",
"0.4216",
"0.3687",
"0.4229",
"0.3687",
"0.437",
"0.1503",
"0.4358"
],
[
"0.5146",
"0.4275",
"0.881",
"0.4275",
"0.881",
"0.4382",
"0.5146",
"0.4382"
],
[
"0.513",
"0.4424",
"0.881",
"0.4424",
"0.881",
"0.456",
"0.513",
"0.456"
],
[
"0.11755",
"0.4507",
"0.3208",
"0.4507",
"0.3208",
"0.4644",
"0.11755",
"0.4644"
],
[
"0.5146",
"0.4578",
"0.881",
"0.4578",
"0.881",
"0.4714",
"0.5146",
"0.4714"
],
[
"0.1184",
"0.4739",
"0.4844",
"0.4739",
"0.4844",
"0.4868",
"0.1184",
"0.4868"
],
[
"0.513",
"0.4744",
"0.881",
"0.4744",
"0.881",
"0.488",
"0.513",
"0.488"
],
[
"0.11755",
"0.4893",
"0.4844",
"0.4893",
"0.4844",
"0.503",
"0.11755",
"0.503"
],
[
"0.511",
"0.4912",
"0.55",
"0.4912",
"0.55",
"0.503",
"0.511",
"0.503"
],
[
"0.1184",
"0.506",
"0.486",
"0.506",
"0.486",
"0.5195",
"0.1184",
"0.5195"
],
[
"0.531",
"0.509",
"0.881",
"0.509",
"0.881",
"0.5225",
"0.531",
"0.5225"
],
[
"0.1184",
"0.5215",
"0.4836",
"0.5215",
"0.4836",
"0.535",
"0.1184",
"0.535"
],
[
"0.511",
"0.524",
"0.588",
"0.5254",
"0.588",
"0.538",
"0.511",
"0.536"
],
[
"0.11755",
"0.536",
"0.4836",
"0.536",
"0.4836",
"0.55",
"0.11755",
"0.55"
],
[
"0.12006",
"0.5547",
"0.3552",
"0.5547",
"0.3552",
"0.5654",
"0.12006",
"0.5654"
],
[
"0.519",
"0.558",
"0.881",
"0.558",
"0.881",
"0.572",
"0.519",
"0.572"
],
[
"0.1368",
"0.571",
"0.4878",
"0.571",
"0.4878",
"0.5845",
"0.1368",
"0.5845"
],
[
"0.5146",
"0.576",
"0.7583",
"0.576",
"0.7583",
"0.586",
"0.5146",
"0.586"
],
[
"0.1184",
"0.586",
"0.4878",
"0.586",
"0.4878",
"0.5996",
"0.1184",
"0.5996"
],
[
"0.513",
"0.5913",
"0.881",
"0.593",
"0.881",
"0.6055",
"0.513",
"0.605"
],
[
"0.1184",
"0.6025",
"0.486",
"0.6025",
"0.486",
"0.616",
"0.1184",
"0.616"
],
[
"0.531",
"0.611",
"0.5425",
"0.611",
"0.5425",
"0.6216",
"0.531",
"0.6216"
],
[
"0.1184",
"0.618",
"0.486",
"0.618",
"0.486",
"0.633",
"0.1184",
"0.633"
],
[
"0.5146",
"0.626",
"0.6035",
"0.626",
"0.6035",
"0.637",
"0.5146",
"0.637"
],
[
"0.596",
"0.6245",
"0.881",
"0.6245",
"0.881",
"0.638",
"0.596",
"0.638"
],
[
"0.12006",
"0.6357",
"0.1897",
"0.6357",
"0.1897",
"0.648",
"0.12006",
"0.648"
],
[
"0.529",
"0.6396",
"0.588",
"0.641",
"0.586",
"0.6562",
"0.5283",
"0.654"
],
[
"0.12177",
"0.662",
"0.3015",
"0.662",
"0.3015",
"0.676",
"0.12177",
"0.676"
],
[
"0.293",
"0.662",
"0.4844",
"0.662",
"0.4844",
"0.676",
"0.293",
"0.676"
],
[
"0.5146",
"0.6567",
"0.8613",
"0.6567",
"0.8613",
"0.6704",
"0.5146",
"0.6704"
],
[
"0.1184",
"0.6777",
"0.486",
"0.6777",
"0.486",
"0.6914",
"0.1184",
"0.6914"
],
[
"0.5146",
"0.677",
"0.8774",
"0.677",
"0.8774",
"0.687",
"0.5146",
"0.687"
],
[
"0.1184",
"0.695",
"0.486",
"0.695",
"0.486",
"0.708",
"0.1184",
"0.708"
],
[
"0.5146",
"0.6934",
"0.879",
"0.6934",
"0.879",
"0.7036",
"0.5146",
"0.7036"
],
[
"0.1184",
"0.7114",
"0.486",
"0.7114",
"0.486",
"0.7246",
"0.1184",
"0.7246"
],
[
"0.513",
"0.708",
"0.81",
"0.708",
"0.81",
"0.7217",
"0.513",
"0.7217"
],
[
"0.1184",
"0.727",
"0.486",
"0.727",
"0.486",
"0.7407",
"0.1184",
"0.7407"
],
[
"0.531",
"0.728",
"0.879",
"0.728",
"0.879",
"0.7383",
"0.531",
"0.7383"
],
[
"0.1184",
"0.7437",
"0.4836",
"0.7437",
"0.4836",
"0.7544",
"0.1184",
"0.7544"
],
[
"0.511",
"0.742",
"0.6616",
"0.742",
"0.6616",
"0.756",
"0.511",
"0.756"
],
[
"0.11755",
"0.7583",
"0.4836",
"0.7583",
"0.4836",
"0.7715",
"0.11755",
"0.7715"
],
[
"0.12006",
"0.777",
"0.4836",
"0.777",
"0.4836",
"0.7876",
"0.12006",
"0.7876"
],
[
"0.532",
"0.7725",
"0.881",
"0.775",
"0.881",
"0.788",
"0.532",
"0.786"
],
[
"0.1184",
"0.7915",
"0.4844",
"0.7915",
"0.4844",
"0.8047",
"0.1184",
"0.8047"
],
[
"0.511",
"0.786",
"0.882",
"0.7876",
"0.882",
"0.8003",
"0.511",
"0.8"
],
[
"0.513",
"0.799",
"0.881",
"0.8",
"0.881",
"0.811",
"0.513",
"0.81"
],
[
"0.1184",
"0.808",
"0.4878",
"0.808",
"0.4878",
"0.822",
"0.1184",
"0.822"
],
[
"0.511",
"0.81",
"0.7793",
"0.8086",
"0.7793",
"0.8223",
"0.511",
"0.8237"
],
[
"0.11755",
"0.8237",
"0.4844",
"0.8237",
"0.4844",
"0.8374",
"0.11755",
"0.8374"
],
[
"0.5356",
"0.8223",
"0.882",
"0.8223",
"0.882",
"0.836",
"0.5356",
"0.836"
],
[
"0.511",
"0.835",
"0.881",
"0.835",
"0.881",
"0.8486",
"0.511",
"0.8486"
],
[
"0.1184",
"0.8413",
"0.4836",
"0.8413",
"0.4836",
"0.8516",
"0.1184",
"0.8516"
],
[
"0.5146",
"0.844",
"0.881",
"0.8467",
"0.881",
"0.86",
"0.5146",
"0.858"
],
[
"0.1184",
"0.8555",
"0.4878",
"0.8555",
"0.4878",
"0.869",
"0.1184",
"0.869"
],
[
"0.5146",
"0.86",
"0.879",
"0.86",
"0.879",
"0.8706",
"0.5146",
"0.8706"
],
[
"0.511",
"0.8706",
"0.881",
"0.8706",
"0.881",
"0.8843",
"0.511",
"0.8843"
],
[
"0.1351",
"0.88",
"0.4895",
"0.8813",
"0.4895",
"0.8955",
"0.1351",
"0.8945"
],
[
"0.513",
"0.8843",
"0.881",
"0.8843",
"0.881",
"0.8945",
"0.513",
"0.8945"
],
[
"0.1184",
"0.8936",
"0.4878",
"0.8936",
"0.4878",
"0.9077",
"0.1184",
"0.9077"
],
[
"0.511",
"0.8945",
"0.882",
"0.8945",
"0.882",
"0.9077",
"0.511",
"0.9077"
],
[
"0.11755",
"0.905",
"0.409",
"0.905",
"0.409",
"0.9185",
"0.11755",
"0.9185"
],
[
"0.511",
"0.9062",
"0.6406",
"0.9062",
"0.6406",
"0.9194",
"0.511",
"0.9194"
]
] | [
"设置了k=2，因为我们的目标是丢弃表单",
"语言约200k一500k个单词。7此外，由于运行",
"wgEBt,在另一个binwta,wtaEBt,中有两",
"时的原因，我们将正字法变体的词汇表大小限",
"个相应表单的情况，其中一个可能放错了位",
"制为每种语言200k个单词",
"置。k=2的CSLS定义可以写成",
"Minimal Supervision Source对于每种语言，",
"CSLS(a,y) = D(c,y)-",
"我们都从UniMorph（Kirovetal.2018）中提取",
"1",
"",
"了5词素的拐点表。我们的目标是经常使用的",
"D(a,NN(Y) -",
"=D(NN(X), y)",
"新",
"2",
"词素，具有强大的嵌入表示形式，并且来自不",
"同的类，以尽可能多地捕获语言的语言行为。",
"其中NN(Y)是α在组Y中的第二个最近",
"为此，我们以嵌人词汇表中出现的形式最多的",
"邻居",
"词素为目标，并从它们中手动选择，使它们尽",
"尽管Lampleetal.（2018）使用的原始距离度",
"可能属于不同的类。8",
"量D(，）是余弦距离函数，但任何距离度量",
"都可以。在语义算法中（Sec。3.1.2），我们应",
"TheInflectionModel作为我们的监督拐点",
"用反映语义相似度的余弦距离。在正交算法中",
"模型，要在引导数据上进行训练，我们使用了",
"（Sec。3.1.1），我们插入Levenshtein编辑距离，",
"开箱即用的序列到序列LSTM模型，并注意",
"以利用拼字法，而不是语义",
"输人的字符和源和目标形式的特征。我们使",
"在对所有可能的单词对进行评分之后，将得",
"用了SilfverbergandHulden（2018）的公开可用",
"分最高的对作为有监督的神经形态偏转器的训",
"模型。我们在最小监督场景中的系统输出的多",
"练集",
"达10,000数据训练示例或监督场景中的6,000",
"训练示例上训练了50个时期的模型。后者为",
"4",
"Experimental Evaluation",
"我们的最小监督系统的性能提供了一个上限。",
"SilfverbergandHulden（2018）提供了此处检查",
"4.1 Experimental Setup",
"的语言的4的培训和测试数据，我们手工制作",
"我们开始使用不同类型的规则，从最少数量的",
"了英语，俄语，匈牙利语和土耳其语的类似数",
"标记拐点表开始，以经验方式检查是否可以训",
"据集",
"练形态拐点模型。我们的实验包括八种语言的",
"模型的评估指标是输出字符串与黄金输出的",
"口头范式：英语，俄语，拉脱维亚语，西班牙",
"精确匹配",
"语，法语，芬兰语，匈牙利语和土耳其语，尽",
"管我们的方法适用于任何范式",
"Models，BaselineandSkyline我们报告了我",
"为了使评估拐点模型成为可能，我们必须选",
"们测试的三个系统变体的结果：",
"择具有足够数量的黄金（测试）数据的语言，",
": ORTH : The orthography-based system (s3.1.1",
"并模拟低资源场景进行训练。这限制了我们使",
"）",
"用西方语言，我们的目标是尽可能多地使用非",
"·SEM:",
"The embedings-based system (Sec.",
"印欧语言",
"3.1.2);",
"The Unlabeled Data",
"我们在第2节中指定的",
": CoMB : The combined system (Sec. 3.1.3).",
"问题设置指定使用大量未标记的文本。实际上，",
"此外，为了评估系统中的附加值，我们包括了",
"所提出的算法利用文本（i）收集词汇以寻找标",
"在低资源环境中训练的两个基线模型的准确",
"记候选者，以及（ii）训练嵌人以用于计算候选",
"性，而没有找到新的示例：THMAX2",
"者对之间的语义关系。在我们的实验中，我们",
"我们还将以完全监督方式（SUP）训练的模",
"简单地采用特定于语言的预训练单词嵌入来实",
"型的性能作为上限",
"现这两个目的。我们使用了Graveetal.（2018)",
"提供的预先训练的FastText向量，据报道，它",
"7To reduce run time the Finnish and Hungarian vocabular",
"们在更大的类比测试集（Lietal,2018）中成功",
"ies were further reduced to include only words that appeared",
"at least twice in the Finnish Parole and Hungarian Gigaword",
"超过了形态类比。我们将2百万长的FastText",
"corpora (Bartis, 1998; Oravecz et al., 2014)",
"词汇表剪裁为仅包含真实单词，即。e。，我们",
"\"Cases of syncretism were trivially solved by merging 2",
"categories into one category with 2 tags if all supervision lex-",
"只保留不包含非学母学符的小写标记，并至少",
"emes had the same forms in those 2 categories.This strategy",
"包含一个元音。6此过程将词汇量缩减为每种",
"might pose a problem in cases of partial syncretism, where",
"categories differ in forms only in classes that happen to be",
"6This last criterion happens to be suitable for all 8 lan-",
"absent from the minimal supervision.The manually enforced",
"guages examined.Admittedly there are languages, like",
"class diversity in the selection process was devised to solve",
"Czech, that allow less sonorant syllable nucleus.",
"this problem as well."
] | {
"arxiv_id": "2104.08512",
"language": "zh",
"page_idx": 4
} |
||
2304.01282-zh-1 | 2304.01282-zh-1.jpg | $\textbf{PEACH: 预 训 练 序 列 到 序 列 多 语 言 模 型 , 用 于 半 监 督 伪 并 行 文 档 生 成 的 }$
翻译
Alireza Salemi$^1;$Amirhossein Abaskohi$^1;$Sara Tavakoli$^1$,
Yadollah Yaghoobzadeh$^1$,Azadeh Shakery$^{1,2}$ $^{1}$School of Electrical and Computer Engineering
College of Engineering, University of Tehran, Tehran, Iran
$^{2}$School of Computer Science
Institute for Research in Fundamental Sciences (IPM), Iran
$\{$ alireza.salemi,amir.abaskohi,saratavakoli77,y.yaghoobzadeh,shakery $\}$ @ut.ac.ir
# Abstract
Multilingual pre-training significantly improves many multilingual NLP tasks, including machine translation. Most existing methods are based on some variants of masked language modeling and text-denoising objectives on monolingual data. Multilingual pretraining on monolingual data ignores the availability of parallel data in many language pairs. Also, some other works integrate the available human-generated parallel translation data in their pre-training. This kind of parallel data is definitely helpful, but it is limited even in highresource language pairs. This paper introduces a novel semi-supervised method, SPDG, that generates high-quality pseudo-parallel data for multilingual pre-training. First, a denoising model is pre-trained on monolingual data to reorder, add, remove, and substitute words, enhancing the pre-training documents' quality. Then, we generate different pseudotranslations for each pre-training document using dictionaries for word-by-word translation and applying the pre-trained denoising model. The resulting pseudo-parallel data is then used to pre-train our multilingual sequence-to-sequence model, PEACH. Our experiments show that PEACH outperforms existing approaches used in training mT5 (Xue et al., 2021) and mBART (Liu et al., 2020) on various translation tasks, including supervised, zero- and few-shot scenarios. Moreover, PEACH's ability to transfer knowledge between similar languages makes it particularly useful for low-resource languages. Our results demonstrate that with high-quality dictionaries for generating accurate pseudo-parallel, PEACH can be valuable for low-resource languages.
料库上进行多语言预训练对于 NLP 系统在多语言下游任务 (Liu et al., 2020; Lample and Conneau, 2019; Conneau et al., 2020; Xue et al., 2021; Devlin et al., 2019) 上的性能以及语言 (Wu and Dredze, 2019; K et al., 2020; Liu et al., 2020) 之间的知识可传递性是有利可图的。此外，在预训练编码器和编码器-解码器模型中使用并行数据有效地提高了模型在下游任务(Lample and Conneau, 2019; Chi et al., 2021) 中的性能。现有的预训练方法主要基于屏蔽语言建模(MLM) 及其变体 (Liu et al.,2020; Raffel et al., 2020; Xue et al., 2021; Lewis et al., 2020)
尽管在预训练多语言模型中使用并行数据可以提高其在下游任务上的性能，但可用的并行数据量有限 (Tran et al.,2020)。此外，基于MLM 的序列到序列 (seq2seq) 模型的目标通常要求模型以与输入相同的语言生成输出，这不符合翻译任务的利益。此外，基于 MLM 的目标使用不同语言之间的共享子单词或字母来学习 (Lample and Conneau, 2019; Lample et al., 2017; Smith et al., 2017) 之间的共享嵌入空间； 对于没有共享字母的语言，这是不可能的
已经研究了在 seq2seq 模型的编码器的跨语言预训练中使用字典定义不同语言之间的锚点，并证明对无监督翻译(Duan et al., 2020)有效。尽管如此，它从未被用作预训练多语言seq2seq 模型的方法
我们提出的半监督伪并行文档生成(SPDG) 方法通过利用字典生成伪并行文档来解决低资源语言的有限并行数据的挑战。SPDG 采用无监督翻译技术 (Kim et al., 2018; Lample et al. 2017) 为每个预训练文档生成高质量的翻译。我们使用预先训练的去噪 seq2seq 模型，通过单词重新排序、添加、删除和替换来提高逐字翻译文档的质量。使用 SPDG 作为新的预训练方法，将改进的无监督翻译文本用作训练我们的多语言 seq2seq 模型 PEACH 的目标文本。SPDG 可以在类似的语言之间进行知识转移， 这使得它对低资源语言特别有用。
我们的实验表明，PEACH 在英语，法语和德语的重新排序目标方面优于 mT5 的 MLM 和
### 1 Introduction
机器翻译(MT)涉及将文本从一种语言传输到另一种语言。最近的调查显示，在大型语
$\overline{{\text{ёqual contribution}}}$ | [
[
"0.12177",
"0.0849",
"0.875",
"0.0849",
"0.875",
"0.0986",
"0.12177",
"0.0986"
],
[
"0.471",
"0.1027",
"0.525",
"0.10095",
"0.5264",
"0.1194",
"0.4727",
"0.1217"
],
[
"0.2627",
"0.1289",
"0.748",
"0.1289",
"0.748",
"0.1437",
"0.2627",
"0.1437"
],
[
"0.3105",
"0.1455",
"0.6943",
"0.1455",
"0.6943",
"0.1603",
"0.3105",
"0.1603"
],
[
"0.3064",
"0.1603",
"0.6987",
"0.1609",
"0.6987",
"0.1787",
"0.3064",
"0.1776"
],
[
"0.266",
"0.1799",
"0.738",
"0.1799",
"0.738",
"0.1948",
"0.266",
"0.1948"
],
[
"0.3787",
"0.1948",
"0.623",
"0.1953",
"0.623",
"0.2102",
"0.3787",
"0.209"
],
[
"0.2627",
"0.2131",
"0.7407",
"0.2131",
"0.7407",
"0.2268",
"0.2627",
"0.2268"
],
[
"0.2024",
"0.231",
"0.8076",
"0.231",
"0.8076",
"0.2446",
"0.2024",
"0.2446"
],
[
"0.2627",
"0.2542",
"0.3418",
"0.2542",
"0.3418",
"0.2683",
"0.2627",
"0.2683"
],
[
"0.5146",
"0.2546",
"0.881",
"0.2546",
"0.881",
"0.2683",
"0.5146",
"0.2683"
],
[
"0.511",
"0.2708",
"0.882",
"0.2695",
"0.882",
"0.284",
"0.511",
"0.2852"
],
[
"0.1453",
"0.284",
"0.4592",
"0.284",
"0.4592",
"0.2976",
"0.1453",
"0.2976"
],
[
"0.51",
"0.2861",
"0.881",
"0.2852",
"0.881",
"0.2998",
"0.51",
"0.3005"
],
[
"0.1436",
"0.2976",
"0.4592",
"0.2964",
"0.4592",
"0.3113",
"0.1436",
"0.3123"
],
[
"0.513",
"0.304",
"0.881",
"0.304",
"0.881",
"0.3142",
"0.513",
"0.3142"
],
[
"0.1486",
"0.3142",
"0.4575",
"0.3142",
"0.4575",
"0.3247",
"0.1486",
"0.3247"
],
[
"0.513",
"0.3196",
"0.879",
"0.3196",
"0.879",
"0.3333",
"0.513",
"0.3333"
],
[
"0.147",
"0.328",
"0.4575",
"0.328",
"0.4575",
"0.3384",
"0.147",
"0.3384"
],
[
"0.513",
"0.335",
"0.882",
"0.335",
"0.882",
"0.3486",
"0.513",
"0.3486"
],
[
"0.1453",
"0.3416",
"0.4575",
"0.3416",
"0.4575",
"0.3552",
"0.1453",
"0.3552"
],
[
"0.1453",
"0.3552",
"0.4575",
"0.3552",
"0.4575",
"0.3682",
"0.1453",
"0.3682"
],
[
"0.513",
"0.3516",
"0.881",
"0.3516",
"0.881",
"0.3652",
"0.513",
"0.3652"
],
[
"0.1453",
"0.3694",
"0.4592",
"0.3694",
"0.4592",
"0.383",
"0.1453",
"0.383"
],
[
"0.513",
"0.3682",
"0.881",
"0.3682",
"0.881",
"0.3818",
"0.513",
"0.3818"
],
[
"0.1453",
"0.383",
"0.4568",
"0.383",
"0.4568",
"0.3967",
"0.1453",
"0.3967"
],
[
"0.513",
"0.3843",
"0.881",
"0.3843",
"0.881",
"0.398",
"0.513",
"0.398"
],
[
"0.1453",
"0.398",
"0.4568",
"0.398",
"0.4568",
"0.411",
"0.1453",
"0.411"
],
[
"0.5146",
"0.3997",
"0.881",
"0.3997",
"0.881",
"0.4133",
"0.5146",
"0.4133"
],
[
"0.147",
"0.4133",
"0.4568",
"0.4133",
"0.4568",
"0.4233",
"0.147",
"0.4233"
],
[
"0.511",
"0.415",
"0.879",
"0.415",
"0.879",
"0.4287",
"0.511",
"0.4287"
],
[
"0.1453",
"0.4258",
"0.4575",
"0.4258",
"0.4575",
"0.4395",
"0.1453",
"0.4395"
],
[
"0.513",
"0.4316",
"0.8237",
"0.4316",
"0.8237",
"0.4453",
"0.513",
"0.4453"
],
[
"0.1436",
"0.4395",
"0.4592",
"0.4407",
"0.4592",
"0.4536",
"0.1436",
"0.4531"
],
[
"0.1453",
"0.4548",
"0.4568",
"0.4548",
"0.4568",
"0.4685",
"0.1453",
"0.4685"
],
[
"0.532",
"0.4495",
"0.881",
"0.4495",
"0.881",
"0.4631",
"0.532",
"0.4631"
],
[
"0.1453",
"0.4685",
"0.4575",
"0.4685",
"0.4575",
"0.4822",
"0.1453",
"0.4822"
],
[
"0.5146",
"0.4656",
"0.881",
"0.4656",
"0.881",
"0.4785",
"0.5146",
"0.4785"
],
[
"0.147",
"0.4827",
"0.4568",
"0.4827",
"0.4568",
"0.4963",
"0.147",
"0.4963"
],
[
"0.513",
"0.4822",
"0.881",
"0.4822",
"0.881",
"0.4954",
"0.513",
"0.4954"
],
[
"0.1453",
"0.4954",
"0.4575",
"0.4963",
"0.4575",
"0.511",
"0.1453",
"0.5103"
],
[
"0.513",
"0.4976",
"0.881",
"0.4976",
"0.881",
"0.511",
"0.513",
"0.511"
],
[
"0.147",
"0.511",
"0.4575",
"0.511",
"0.4575",
"0.525",
"0.147",
"0.525"
],
[
"0.5146",
"0.514",
"0.881",
"0.514",
"0.881",
"0.528",
"0.5146",
"0.528"
],
[
"0.1453",
"0.5254",
"0.4575",
"0.5254",
"0.4575",
"0.539",
"0.1453",
"0.539"
],
[
"0.513",
"0.53",
"0.881",
"0.5293",
"0.881",
"0.5435",
"0.513",
"0.5444"
],
[
"0.1453",
"0.539",
"0.4592",
"0.539",
"0.4592",
"0.5527",
"0.1453",
"0.5527"
],
[
"0.5146",
"0.5464",
"0.881",
"0.5464",
"0.881",
"0.56",
"0.5146",
"0.56"
],
[
"0.1755",
"0.5557",
"0.1914",
"0.5557",
"0.1914",
"0.5635",
"0.1755",
"0.5635"
],
[
"0.1873",
"0.554",
"0.4592",
"0.554",
"0.4592",
"0.568",
"0.1873",
"0.568"
],
[
"0.513",
"0.5625",
"0.881",
"0.5625",
"0.881",
"0.576",
"0.513",
"0.576"
],
[
"0.1453",
"0.568",
"0.4592",
"0.568",
"0.4592",
"0.5806",
"0.1453",
"0.5806"
],
[
"0.513",
"0.5776",
"0.8833",
"0.5776",
"0.8833",
"0.5913",
"0.513",
"0.5913"
],
[
"0.1453",
"0.582",
"0.4592",
"0.582",
"0.4592",
"0.5957",
"0.1453",
"0.5957"
],
[
"0.1486",
"0.557",
"0.178",
"0.557",
"0.178",
"0.564",
"0.1486",
"0.564"
],
[
"0.1453",
"0.5967",
"0.4575",
"0.5967",
"0.4575",
"0.61",
"0.1453",
"0.61"
],
[
"0.513",
"0.594",
"0.843",
"0.594",
"0.843",
"0.608",
"0.513",
"0.608"
],
[
"0.1453",
"0.61",
"0.4575",
"0.61",
"0.4575",
"0.6245",
"0.1453",
"0.6245"
],
[
"0.532",
"0.6123",
"0.882",
"0.6123",
"0.882",
"0.626",
"0.532",
"0.626"
],
[
"0.1453",
"0.6245",
"0.4575",
"0.6245",
"0.4575",
"0.638",
"0.1453",
"0.638"
],
[
"0.5146",
"0.627",
"0.879",
"0.627",
"0.879",
"0.64",
"0.5146",
"0.64"
],
[
"0.1453",
"0.6396",
"0.4592",
"0.6396",
"0.4592",
"0.653",
"0.1453",
"0.653"
],
[
"0.5146",
"0.6445",
"0.881",
"0.6445",
"0.881",
"0.6577",
"0.5146",
"0.6577"
],
[
"0.1436",
"0.653",
"0.4592",
"0.652",
"0.4592",
"0.6665",
"0.1436",
"0.6675"
],
[
"0.511",
"0.659",
"0.881",
"0.66",
"0.881",
"0.675",
"0.511",
"0.6733"
],
[
"0.1453",
"0.6675",
"0.4575",
"0.6675",
"0.4575",
"0.681",
"0.1453",
"0.681"
],
[
"0.51",
"0.676",
"0.6675",
"0.676",
"0.6675",
"0.6895",
"0.51",
"0.6895"
],
[
"0.147",
"0.683",
"0.4568",
"0.683",
"0.4568",
"0.6934",
"0.147",
"0.6934"
],
[
"0.1453",
"0.697",
"0.4592",
"0.697",
"0.4592",
"0.7104",
"0.1453",
"0.7104"
],
[
"0.531",
"0.6934",
"0.882",
"0.6934",
"0.882",
"0.707",
"0.531",
"0.707"
],
[
"0.1436",
"0.709",
"0.4575",
"0.7104",
"0.4575",
"0.7236",
"0.1436",
"0.7227"
],
[
"0.513",
"0.7104",
"0.882",
"0.7104",
"0.882",
"0.7236",
"0.513",
"0.7236"
],
[
"0.1453",
"0.7236",
"0.4575",
"0.7236",
"0.4575",
"0.737",
"0.1453",
"0.737"
],
[
"0.513",
"0.727",
"0.881",
"0.727",
"0.881",
"0.7407",
"0.513",
"0.7407"
],
[
"0.1453",
"0.7383",
"0.4575",
"0.7383",
"0.4575",
"0.753",
"0.1453",
"0.753"
],
[
"0.513",
"0.741",
"0.879",
"0.741",
"0.879",
"0.755",
"0.513",
"0.755"
],
[
"0.1453",
"0.752",
"0.3955",
"0.753",
"0.3955",
"0.7666",
"0.1453",
"0.7656"
],
[
"0.403",
"0.752",
"0.4575",
"0.7544",
"0.4575",
"0.7656",
"0.403",
"0.763"
],
[
"0.511",
"0.7573",
"0.882",
"0.7583",
"0.882",
"0.7725",
"0.511",
"0.7715"
],
[
"0.1453",
"0.767",
"0.4575",
"0.767",
"0.4575",
"0.781",
"0.1453",
"0.781"
],
[
"0.511",
"0.774",
"0.879",
"0.774",
"0.879",
"0.7876",
"0.511",
"0.7876"
],
[
"0.1453",
"0.781",
"0.4575",
"0.781",
"0.4575",
"0.7944",
"0.1453",
"0.7944"
],
[
"0.1453",
"0.7944",
"0.4592",
"0.7944",
"0.4592",
"0.808",
"0.1453",
"0.808"
],
[
"0.513",
"0.7905",
"0.879",
"0.7905",
"0.879",
"0.804",
"0.513",
"0.804"
],
[
"0.1453",
"0.8125",
"0.199",
"0.8125",
"0.199",
"0.8237",
"0.1453",
"0.8237"
],
[
"0.513",
"0.807",
"0.879",
"0.807",
"0.879",
"0.821",
"0.513",
"0.821"
],
[
"0.513",
"0.8237",
"0.882",
"0.8237",
"0.882",
"0.8374",
"0.513",
"0.8374"
],
[
"0.1503",
"0.8374",
"0.2644",
"0.839",
"0.2644",
"0.854",
"0.1503",
"0.8516"
],
[
"0.511",
"0.839",
"0.881",
"0.839",
"0.881",
"0.8525",
"0.511",
"0.8525"
],
[
"0.1351",
"0.8433",
"0.1427",
"0.8433",
"0.1427",
"0.8467",
"0.1351",
"0.8467"
],
[
"0.11755",
"0.8403",
"0.1318",
"0.8403",
"0.1318",
"0.8496",
"0.11755",
"0.8496"
],
[
"0.513",
"0.855",
"0.8833",
"0.855",
"0.8833",
"0.869",
"0.513",
"0.869"
],
[
"0.1184",
"0.865",
"0.4844",
"0.865",
"0.4844",
"0.879",
"0.1184",
"0.879"
],
[
"0.511",
"0.8716",
"0.7773",
"0.8716",
"0.7773",
"0.8853",
"0.511",
"0.8853"
],
[
"0.1184",
"0.882",
"0.4844",
"0.882",
"0.4844",
"0.8955",
"0.1184",
"0.8955"
],
[
"0.531",
"0.8896",
"0.881",
"0.8896",
"0.881",
"0.903",
"0.531",
"0.903"
],
[
"0.1427",
"0.905",
"0.2568",
"0.905",
"0.2568",
"0.9194",
"0.1427",
"0.9194"
],
[
"0.513",
"0.905",
"0.882",
"0.905",
"0.882",
"0.9185",
"0.513",
"0.9185"
]
] | [
"PEACH：预训练序列到序列多语言模型，用于半监督伪并行文档生成的",
"翻译",
"Alireza Salemil, Amirhossein Abaskohil, Sara Tavakoli'l,",
"Yadollah Yaghoobzadeh', Azadeh Shakeryl,2",
"1School of Electrical and Computer Engineering",
"College of Engineering, University of Tehran, Tehran, Iran",
"2School of Computer Science",
"Institute for Research in Fundamental Sciences (IPM), Iran",
"( alireza.salemi,amir.abaskohi,saratavakoli77,y.yaghoobzadeh,shakery J @ut.ac.ir",
"Abstract",
"料库上进行多语言预训练对于NLP系统在多",
"语言下游任务（Liuetal.,2020LampleandCon",
"Multilingual pre-training significantly im-",
"neau, 2019; Conneau et al., 2020; Xue et al., 2021;",
"proves many multilingual NLP tasks, includ-",
"Devlinetal.2019）上的性能以及语言（Wuand",
"ing machinetranslation.Most existing meth-",
"Dredze, 2019; K et al., 2020; Liu et al., 2020) 之",
"ods are based on some variants of masked",
"间的知识可传递性是有利可图的。。此外，在",
"language modeling and text-denoising objec-",
"tives on monolingual data. Multilingual pre-",
"预训练编码器和编码器-解码器模型中使用并",
"training on monolingual data ignores the avail-",
"行数据有效地提高了模型在下游任务(Lam",
"ability of parallel data in many language pairs.",
"ple and Conneau, 2019; Chi et al., 2021)中的性",
"Also, some other works integrate the available",
"能。现有的预训练方法主要基于屏蔽语言建模",
"human-generated parallel translation data in",
"(MLM）及其变体（Liuetal.,2020;Raffeletal.",
"their pre-training. This kind of parallel data is",
"2020; Xue et al., 2021; Lewis et al., 2020)",
"definitely helpful, but it is limited even in high-",
"resource language pairs. This paper introduces",
"尽管在预训练多语言模型中使用并行数据",
"a novel semi-supervised method, SPDG, that",
"可以提高其在下游任务上的性能，但可用的并",
"generates high-quality pseudo-parallel data for",
"行数据量有限（Tranetal.,2020）。此外，基于",
"multilingual pre-training. First, a denoising",
"MLM的序列到序列（seq2seq）模型的目标通",
"model is pre-trained on monolingual data to",
"常要求模型以与输入相同的语言生成输出，这",
"reorder, add, remove, and substitute words,",
"不符合翻译任务的利益。此外，基于MLM的",
"enhancing the pre-training documents' qual-",
"目标使用不同语言之间的共享子单词或字母来",
"",
"Then, we generate different pseudo-",
"学习 (Lample and Conneau, 2019; Lample et al.,",
"translations for each pre-training document",
"2017;Smithetal.2017）之间的共享嵌人空间；",
"using dictionaries for word-by-word transla-",
"ity.",
"tion and applying the pre-trained denoising",
"对于没有共享字母的语言，这是不可能的",
"model. The resulting pseudo-parallel data",
"已经研究了在seq2seq模型的编码器的跨语",
"is then used to pre-train our multilingual",
"言预训练中使用字典定义不同语言之间的错",
"sequence-to-sequence model, PEACH. Our ex-",
"点，并证明对无监督翻译(Duanetal.,2020)有",
"periments show that PEACH outperforms ex-",
"效。尽管如此，它从未被用作预训练多语言",
"isting approaches used in training mT5 (Xue",
"seq2seq模型的方法",
"et al.,2021)and mBART (Liu et al.,2020)",
"on various translation tasks, including super-",
"我们提出的半监督伪并行文档生成（SPDG）",
"vised, zero- and few-shot scenarios. Moreover,",
"方法通过利用字典生成伪并行文档来解决低",
"PEACH's ability to transfer knowledge be-",
"资源语言的有限并行数据的挑战。SPDG采用",
"tween similar languages makes it particularly",
"无监督翻译技术（Kimetal.2018;Lampleetal.",
"useful for low-resource languages.",
"Ourre-",
"2017）为每个预训练文档生成高质量的翻译。",
"sults demonstrate that with high-quality dictio-",
"我们使用预先训练的去噪seg2seq模型，通过",
"naries for generating accurate pseudo-parallel,",
"PEACH can be valuable for low-resource lan-",
"单词重新排序、添加、删除和替换来提高逐字",
"guages.",
"翻译文档的质量。使用SPDG作为新的预训",
"练方法，将改进的无监督翻译文本用作训练我",
"Introduction",
"们的多语言seq2seq模型PEACH的目标文本。",
"",
"1",
"SPDG可以在类似的语言之间进行知识转移，",
"机器翻译（MT）涉及将文本从一种语言传输",
"这使得它对低资源语言特别有用。",
"到另一种语言。最近的调查显示，在大型语",
"我们的实验表明，PEACH在英语，法语和德",
"equal contribution",
"语的重新排序目标方面优于mT5的MLM和"
] | {
"arxiv_id": "2304.01282",
"language": "zh",
"page_idx": 1
} |
||
2304.01282-zh-2 | 2304.01282-zh-2.jpg | mBART的 MLM 的预训练模型。此外，PEACH 在零拍和少拍场景中表现出强劲的性能。此外， 我们还针对其他多语言任务(例如自然语言推理)测试了我们的模型，以研究该模型在此任务中的能力。我们的结果表明，我们的模型在这项任务中取得了比其他目标更高的分数，这表明了 PEACH 在语言之间转移知识的能力。本文的主要贡献有两点：
$\bullet$ We propose a novel semi-supervised pretraining method using bilingual dictionaries and pre-trained denoising models for seq2seq multilingual models.
$\cdot$ We show the benefits of SPDG objective in translation, supervised and zero- and fewshot cases, and knowledge transfer between languages.
### 2 $\textbf{Related Work}$
在 MT 的最初尝试中，基于字典和规则的方法是流行的 (Dolan et al., 1993; Kaji, 1988; Meyers et al., 1998) ,其次是基于知识的机器翻译 (KBMT) 和统计方法 (Mitamara et al., 1993; Carbonell et al., 1981; Koehn, 2009; Al-Onaizan et al., 1999)。随着第一个用于翻译的深度神经模型 (Kalchbrenner and Blunsom, 2013) 的引入， 神经机器翻译的普及在最近十年才有所增长
虽然基于 RNN 的 seq2seq 模型在神经机器翻译 (Wu et al., 2016; Bahdanau et al., 2015; Sutskever et al., 2014) 中似乎很有前途，但转换器体系结构 (Vaswani et al.,2017) 的出现在现代 MT 中起着不可或缺的作用。随着变压器架构的引入，预训练通用语言模型似乎是改进不同 NLP 任务 (Devlin et al., 2019; Liu et al., 2019) 的有效方法。在大多数情况下，要求变压器模型对嘈杂的输入进行降噪以学习语言 (Lewis et al., 2020; Devlin et al., 2019; Raffel et al., 2020) 。仅编码器和编码器-解码器模型的最流行的预训练目标之一称为掩码语言建模(MLM), 其中模型应预测文档的掩码部分并在其输出(Raffel et al.,2020)中生成它。但是，还为编码器-解码器和仅编码器模型 (Song et al., 2019; Clark et al., 2020) 开发了许多其他目标
同时，测试了基于对抗性学习 (Lample et al., 2017) 和基于变压器的文本去噪(Kim et al., 2018) 的使用单语语料库的神经机器翻译(NMT)的无监督方法，并证明了有希望的结果。使用双语词典在预训练无监督翻译模型中定义锚点是成功的 (Duan et al.,2020) ,但从未用于生成大规模监督翻译的数据。我们的工作不同于使用字典作为锚点来学习编码器 (Duan
et al., 2020) 中令牌的更好表示。我们使用字典来生成源语言在目标语言中的伪翻译，而不仅仅是定义一些锚点。因此，预训练步骤中的模型学习仅使用单语数据和大规模字典基于源语言的输入来生成目标语言的文本
通过生成伪摘要来预训练特定于任务的模型在某些情况下对于摘要，问题回答和语音识别 (Chen et al., 2017; Salemi et al., 2021; Zhang et al., 2020; Abaskohi et al., 2022a) 是成功的，但是根据我们的知识，对于预训练编码器-解码器 seq2seq 模型进行监督翻译还没有执行。另一方面，为翻译预先训练特定模型的努力最终以训练多语言语言模型 (Xue et al., 2021; Liu et al., 2020) 而告终。。用 T5 (Raffel et al., 2020) 的传销目标训练 mT5 (Xue et al.,2021)。在其预训练目标中，输入文档的某些跨度被特定的令牌掩盖，并且模型必须通过在其输出中生成这些跨度来预测这些跨度。mBART(Liu et al., 2020) 是另一种基于 BART (Lewis et al., 2020) 模型的多语言模型，使用具有重新排序目标的MLM 进行预训练。在 mBART 的预训练目标中，输人文档中的句子顺序被破坏，而特定的令牌掩盖了文档的某些跨度。模型必须在其输出中生成原始文档
PEACH 与提到的两个模型都不同，因为我们使用半监督方法来生成每个预训练文档的多个伪翻译 (每种选定语言一个)。然后将这些翻译馈送到预训练 PEACH。此外，在上述模型中，输入和输出来自同一语言，而我们要求模型在预训练阶段将文本从一种语言翻译成另一种语言
## 3 $\mathbf{PEACH}$
PEACH 是一种使用 SPDG 训练的新的序列到序列多语言转换器模型，SPDG 是一种半监督伪并行文档生成方法。本节介绍预训练目标和模型体系结构
### 3.1 Semi-Supervised Pseudo-Parallel Document Generation (SPDG)
我们提出的预训练目标 SPDG 生成输入文档的伪翻译。为了生成伪翻译，我们使用 Kim et al. (2018)的方法进行无监督翻译，并进行了一些修改。我们基于 SPDG 预训练模型的流程如图2 所示。我们使用目标语言的预训练语料库预训练该语言的 seq2seq 去噪模型。接下来，对于源语言的每个预训练文档，我们使用字典逐字将其翻译为目标语言。然后，我们将此逐字翻译的文档提供给具有去噪目标的预训练模型，以提高其质量并恢复缺失的单词
使用此方法，我们可以生成每个预训练文档 | [
[
"0.12177",
"0.0885",
"0.4878",
"0.0885",
"0.4878",
"0.0986",
"0.12177",
"0.0986"
],
[
"0.513",
"0.0873",
"0.881",
"0.0873",
"0.881",
"0.10095",
"0.513",
"0.10095"
],
[
"0.12006",
"0.1051",
"0.4802",
"0.1051",
"0.4802",
"0.1152",
"0.12006",
"0.1152"
],
[
"0.513",
"0.1027",
"0.881",
"0.1027",
"0.881",
"0.1164",
"0.513",
"0.1164"
],
[
"0.12006",
"0.12054",
"0.4836",
"0.12054",
"0.4836",
"0.1312",
"0.12006",
"0.1312"
],
[
"0.513",
"0.11816",
"0.881",
"0.11816",
"0.881",
"0.1318",
"0.513",
"0.1318"
],
[
"0.1184",
"0.1354",
"0.4844",
"0.1354",
"0.4844",
"0.1484",
"0.1184",
"0.1484"
],
[
"0.513",
"0.1354",
"0.881",
"0.1354",
"0.881",
"0.1484",
"0.513",
"0.1484"
],
[
"0.12006",
"0.152",
"0.482",
"0.152",
"0.482",
"0.1621",
"0.12006",
"0.1621"
],
[
"0.5146",
"0.152",
"0.771",
"0.152",
"0.771",
"0.165",
"0.5146",
"0.165"
],
[
"0.1184",
"0.1675",
"0.4878",
"0.1675",
"0.4878",
"0.1812",
"0.1184",
"0.1812"
],
[
"0.531",
"0.1692",
"0.879",
"0.1692",
"0.879",
"0.1799",
"0.531",
"0.1799"
],
[
"0.12006",
"0.1829",
"0.4895",
"0.1829",
"0.4895",
"0.1965",
"0.12006",
"0.1965"
],
[
"0.513",
"0.1841",
"0.881",
"0.1841",
"0.881",
"0.1978",
"0.513",
"0.1978"
],
[
"0.11755",
"0.1978",
"0.309",
"0.199",
"0.309",
"0.2131",
"0.11755",
"0.212"
],
[
"0.5146",
"0.1996",
"0.881",
"0.2007",
"0.881",
"0.2144",
"0.5146",
"0.2131"
],
[
"0.5146",
"0.2162",
"0.881",
"0.2162",
"0.881",
"0.2299",
"0.5146",
"0.2299"
],
[
"0.1377",
"0.228",
"0.4902",
"0.228",
"0.4902",
"0.2411",
"0.1377",
"0.2411"
],
[
"0.513",
"0.234",
"0.8774",
"0.234",
"0.8774",
"0.2446",
"0.513",
"0.2446"
],
[
"0.1545",
"0.2434",
"0.4878",
"0.2434",
"0.4878",
"0.257",
"0.1545",
"0.257"
],
[
"0.513",
"0.2488",
"0.881",
"0.2488",
"0.881",
"0.2625",
"0.513",
"0.2625"
],
[
"0.1545",
"0.2588",
"0.486",
"0.2588",
"0.486",
"0.2725",
"0.1545",
"0.2725"
],
[
"0.5146",
"0.2642",
"0.881",
"0.2642",
"0.881",
"0.2778",
"0.5146",
"0.2778"
],
[
"0.1561",
"0.2766",
"0.3047",
"0.2766",
"0.3047",
"0.2874",
"0.1561",
"0.2874"
],
[
"0.513",
"0.2808",
"0.881",
"0.2808",
"0.881",
"0.2944",
"0.513",
"0.2944"
],
[
"0.513",
"0.2964",
"0.881",
"0.2964",
"0.881",
"0.31",
"0.513",
"0.31"
],
[
"0.1377",
"0.3027",
"0.486",
"0.3027",
"0.486",
"0.3164",
"0.1377",
"0.3164"
],
[
"0.513",
"0.3113",
"0.879",
"0.3113",
"0.879",
"0.3247",
"0.513",
"0.3247"
],
[
"0.1545",
"0.3206",
"0.4878",
"0.3196",
"0.4878",
"0.3308",
"0.1545",
"0.332"
],
[
"0.1545",
"0.335",
"0.486",
"0.335",
"0.486",
"0.3486",
"0.1545",
"0.3486"
],
[
"0.5146",
"0.33",
"0.879",
"0.33",
"0.879",
"0.3403",
"0.5146",
"0.3403"
],
[
"0.1528",
"0.3499",
"0.2351",
"0.3528",
"0.2334",
"0.3677",
"0.1511",
"0.364"
],
[
"0.5146",
"0.3445",
"0.882",
"0.3445",
"0.882",
"0.3582",
"0.5146",
"0.3582"
],
[
"0.511",
"0.3599",
"0.879",
"0.3599",
"0.879",
"0.3735",
"0.511",
"0.3735"
],
[
"0.1142",
"0.3777",
"0.2747",
"0.379",
"0.2747",
"0.3938",
"0.1142",
"0.3926"
],
[
"0.513",
"0.3772",
"0.881",
"0.3772",
"0.881",
"0.3901",
"0.513",
"0.3901"
],
[
"0.5146",
"0.3938",
"0.881",
"0.3938",
"0.881",
"0.4067",
"0.5146",
"0.4067"
],
[
"0.1184",
"0.4062",
"0.4878",
"0.4062",
"0.4878",
"0.4192",
"0.1184",
"0.4192"
],
[
"0.511",
"0.4092",
"0.879",
"0.4092",
"0.879",
"0.4229",
"0.511",
"0.4229"
],
[
"0.11755",
"0.4216",
"0.4895",
"0.4216",
"0.4895",
"0.4353",
"0.11755",
"0.4353"
],
[
"0.513",
"0.4258",
"0.881",
"0.4258",
"0.881",
"0.4395",
"0.513",
"0.4395"
],
[
"0.1184",
"0.437",
"0.486",
"0.437",
"0.486",
"0.4507",
"0.1184",
"0.4507"
],
[
"0.511",
"0.4395",
"0.881",
"0.4412",
"0.881",
"0.456",
"0.511",
"0.4536"
],
[
"0.1184",
"0.4536",
"0.4878",
"0.4536",
"0.4878",
"0.4673",
"0.1184",
"0.4673"
],
[
"0.513",
"0.4573",
"0.6616",
"0.4573",
"0.6616",
"0.4702",
"0.513",
"0.4702"
],
[
"0.11584",
"0.4697",
"0.4844",
"0.4702",
"0.4844",
"0.484",
"0.11584",
"0.4827"
],
[
"0.5283",
"0.4727",
"0.881",
"0.4739",
"0.881",
"0.488",
"0.5283",
"0.4868"
],
[
"0.1184",
"0.4863",
"0.4844",
"0.4863",
"0.4844",
"0.4995",
"0.1184",
"0.4995"
],
[
"0.513",
"0.4905",
"0.881",
"0.4905",
"0.881",
"0.5044",
"0.513",
"0.5044"
],
[
"0.1184",
"0.503",
"0.4895",
"0.503",
"0.4895",
"0.5166",
"0.1184",
"0.5166"
],
[
"0.513",
"0.5073",
"0.881",
"0.5073",
"0.881",
"0.521",
"0.513",
"0.521"
],
[
"0.11755",
"0.517",
"0.465",
"0.5186",
"0.465",
"0.532",
"0.11755",
"0.531"
],
[
"0.513",
"0.5225",
"0.882",
"0.5225",
"0.882",
"0.536",
"0.513",
"0.536"
],
[
"0.1368",
"0.535",
"0.4844",
"0.535",
"0.4844",
"0.549",
"0.1368",
"0.549"
],
[
"0.511",
"0.5376",
"0.881",
"0.538",
"0.881",
"0.5527",
"0.511",
"0.552"
],
[
"0.1184",
"0.5503",
"0.486",
"0.5503",
"0.486",
"0.564",
"0.1184",
"0.564"
],
[
"0.513",
"0.5547",
"0.879",
"0.5547",
"0.879",
"0.5684",
"0.513",
"0.5684"
],
[
"0.11755",
"0.5664",
"0.4836",
"0.5664",
"0.4836",
"0.58",
"0.11755",
"0.58"
],
[
"0.513",
"0.5723",
"0.586",
"0.5723",
"0.586",
"0.5845",
"0.513",
"0.5845"
],
[
"0.1184",
"0.582",
"0.4844",
"0.583",
"0.4844",
"0.597",
"0.1184",
"0.5967"
],
[
"0.12006",
"0.5996",
"0.486",
"0.5996",
"0.486",
"0.6133",
"0.12006",
"0.6133"
],
[
"0.509",
"0.6016",
"0.6187",
"0.6016",
"0.6187",
"0.6133",
"0.509",
"0.6133"
],
[
"0.1184",
"0.615",
"0.4844",
"0.615",
"0.4844",
"0.629",
"0.1184",
"0.629"
],
[
"0.1184",
"0.6313",
"0.486",
"0.6313",
"0.486",
"0.6445",
"0.1184",
"0.6445"
],
[
"0.511",
"0.627",
"0.879",
"0.627",
"0.879",
"0.641",
"0.511",
"0.641"
],
[
"0.1184",
"0.648",
"0.4878",
"0.648",
"0.4878",
"0.661",
"0.1184",
"0.661"
],
[
"0.511",
"0.6426",
"0.879",
"0.6426",
"0.879",
"0.6562",
"0.511",
"0.6562"
],
[
"0.1184",
"0.662",
"0.4878",
"0.662",
"0.4878",
"0.676",
"0.1184",
"0.676"
],
[
"0.513",
"0.66",
"0.881",
"0.66",
"0.881",
"0.6733",
"0.513",
"0.6733"
],
[
"0.11755",
"0.6787",
"0.4844",
"0.6787",
"0.4844",
"0.6924",
"0.11755",
"0.6924"
],
[
"0.5146",
"0.6777",
"0.623",
"0.6777",
"0.623",
"0.6895",
"0.5146",
"0.6895"
],
[
"0.1234",
"0.6978",
"0.1377",
"0.6978",
"0.1377",
"0.706",
"0.1234",
"0.706"
],
[
"0.1351",
"0.6953",
"0.4844",
"0.6953",
"0.4844",
"0.709",
"0.1351",
"0.709"
],
[
"0.5146",
"0.706",
"0.5557",
"0.706",
"0.5557",
"0.7163",
"0.5146",
"0.7163"
],
[
"0.555",
"0.7036",
"0.817",
"0.7036",
"0.817",
"0.7173",
"0.555",
"0.7173"
],
[
"0.1184",
"0.7114",
"0.4878",
"0.7114",
"0.4878",
"0.7246",
"0.1184",
"0.7246"
],
[
"0.553",
"0.7217",
"0.796",
"0.7217",
"0.796",
"0.7354",
"0.553",
"0.7354"
],
[
"0.12006",
"0.73",
"0.4836",
"0.73",
"0.4836",
"0.7407",
"0.12006",
"0.7407"
],
[
"0.11755",
"0.7446",
"0.4844",
"0.7437",
"0.4844",
"0.7573",
"0.11755",
"0.7583"
],
[
"0.513",
"0.7437",
"0.881",
"0.7437",
"0.881",
"0.7573",
"0.513",
"0.7573"
],
[
"0.11755",
"0.759",
"0.486",
"0.759",
"0.486",
"0.7725",
"0.11755",
"0.7725"
],
[
"0.5146",
"0.7603",
"0.879",
"0.7603",
"0.879",
"0.774",
"0.5146",
"0.774"
],
[
"0.1184",
"0.777",
"0.4207",
"0.777",
"0.4207",
"0.7905",
"0.1184",
"0.7905"
],
[
"0.513",
"0.7754",
"0.881",
"0.7754",
"0.881",
"0.789",
"0.513",
"0.789"
],
[
"0.1368",
"0.792",
"0.4878",
"0.792",
"0.4878",
"0.8057",
"0.1368",
"0.8057"
],
[
"0.5146",
"0.7915",
"0.879",
"0.7915",
"0.879",
"0.8047",
"0.5146",
"0.8047"
],
[
"0.1184",
"0.808",
"0.4878",
"0.808",
"0.4878",
"0.822",
"0.1184",
"0.822"
],
[
"0.511",
"0.808",
"0.879",
"0.808",
"0.879",
"0.822",
"0.511",
"0.822"
],
[
"0.11755",
"0.8237",
"0.4878",
"0.8247",
"0.4878",
"0.8384",
"0.11755",
"0.8374"
],
[
"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",
"0.8403",
"0.882",
"0.8403",
"0.882",
"0.854",
"0.511",
"0.854"
],
[
"0.11755",
"0.857",
"0.4878",
"0.8555",
"0.4878",
"0.869",
"0.11755",
"0.8706"
],
[
"0.51",
"0.857",
"0.879",
"0.857",
"0.879",
"0.8706",
"0.51",
"0.8706"
],
[
"0.1184",
"0.8735",
"0.4878",
"0.8735",
"0.4878",
"0.887",
"0.1184",
"0.887"
],
[
"0.513",
"0.8726",
"0.882",
"0.8726",
"0.882",
"0.8857",
"0.513",
"0.8857"
],
[
"0.1184",
"0.8896",
"0.4878",
"0.8896",
"0.4878",
"0.903",
"0.1184",
"0.903"
],
[
"0.511",
"0.8896",
"0.8076",
"0.8896",
"0.8076",
"0.903",
"0.511",
"0.903"
],
[
"0.1184",
"0.9067",
"0.4836",
"0.9067",
"0.4836",
"0.9175",
"0.1184",
"0.9175"
],
[
"0.532",
"0.9062",
"0.881",
"0.9062",
"0.881",
"0.9194",
"0.532",
"0.9194"
]
] | [
"mBART的MLM的预训练模型。此外，PEACH",
"etal.,2020）中令牌的更好表示。我们使用字典",
"在零拍和少拍场景中表现出强劲的性能。此外",
"来生成源语言在目标语言中的伪翻译，而不仅",
"我们还针对其他多语言任务（例如自然语言推",
"仅是定义一些锚点。因此，预训练步骤中的模",
"理）测试了我们的模型，以研究该模型在此任",
"型学习仅使用单语数据和大规模字典基于源语",
"务中的能力。我们的结果表明：我们的模型在",
"言的输人来生成目标语言的文本",
"这项任务中取得了比其他目标更高的分数，这",
"通过生成伪摘要来预训练特定于任务的模",
"表明了PEACH在语言之间转移知识的能力。",
"型在某些情况下对于摘要，问题回答和语音识",
"本文的主要贡献有两点：",
"别 (Chen et al., 2017; Salemi et al., 2021; Zhang",
"etal.,2020;Abaskohietal.,2022a）是成功的，但",
": We propose a novel semi-supervised pre-",
"是根据我们的知识，对于预训练编码器-解码",
"training method using bilingual dictionaries",
"器seq2seq模型进行监督翻译还没有执行。另",
"and pre-trained denoising models for seq2seq",
"一方面，为翻译预先训练特定模型的努力最",
"multilingualmodels",
"终以训练多语言语言模型(Xueetal.，2021;Liu",
"etal.,2020)而告终。。用T5(Raffeletal.,2020)",
": We show the benefits of SPDG objective in",
"的传销目标训练mT5（Xueetal.，2021）。在其",
"translation,supervised and zero-and few-",
"shot cases, and knowledge transfer between",
"预训练自标中，输入文档的某些跨度被特定的",
"languages.",
"令牌掩盖，并且模型必须通过在其输出中生成",
"这些跨度来预测这些跨度。mBART(Liuetal.",
"2Related Work",
"2020）是另一种基于BART（Lewisetal.,2020)",
"模型的多语言模型，使用具有重新排序目标的",
"在MT的最初尝试中，基于字典和规则的方",
"MLM进行预训练。在mBART的预训练目标",
"法是流行的(Dolanetal.,1993;Kaji,1988;Mey-",
"中，输人文档中的句子顺序被破坏，而特定的",
"ersetal.，1998），其次是基于知识的机器翻",
"令牌掩盖了文档的某些跨度。模型必须在其输",
"译（KBMT）和统计方法（Mitamaraetal.，1993;",
"出中生成原始文档",
"Carbonell etal.,1981;Koehn,2009;A1-Onaizan",
"PEACH与提到的两个模型都不同，因为我",
"etal.1999)。随着第一个用于翻译的深度神经",
"们使用半监督方法来生成每个预训练文档的多",
"模型（KalchbrennerandBlunsom,2013）的引人，",
"个伪翻译（每种选定语言一个）。然后将这些",
"神经机器翻译的普及在最近十年才有所增长",
"翻译馈送到预训练PEACH。此外，在上述模",
"虽然基于RNN的seq2seq模型在神经机",
"型中，输入和输出来自同一语言，而我们要求",
"器翻译(Wu et al.,2016;Bahdanau et al.,2015;",
"模型在预训练阶段将文本从一种语言翻译成另",
"Sutskeveretal.，2014）中似乎很有前途，但转换",
"一种语言",
"器体系结构（Vaswanietal.,2017）的出现在现",
"代MT中起着不可或缺的作用。随着变压器架",
"3PEACH",
"构的引入，预训练通用语言模型似乎是改进不",
"同NLP任务(Devlinet al.,2019;Liu et al.,2019)",
"PEACH是一种使用SPDG训练的新的序列到",
"的有效方法。在大多数情况下，要求变压器",
"序列多语言转换器模型，SPDG是一种半监督",
"模型对噜杂的输入进行降噪以学习语言（Lewis",
"伪并行文档生成方法。本节介绍预训练目标和",
"et al., 2020; Devlin et al., 2019; Raffel et al., 2020)",
"模型体系结构",
"",
"仅编码器和编码器-解码器模型的最流行的",
"3.1",
"Semi-Supervised Pseudo-Paralle",
"预训练目标之一称为掩码语言建模（MLM），",
"Document Generation (SPDG)",
"其中模型应预测文档的掩码部分并在其输出",
"(Raffeletal.，2020）中生成它。但是，还为编",
"我们提出的预训练目标SPDG生成输入文档的",
"码器-解码器和仅编码器模型(Songetal.,2019;",
"伪翻译。为了生成伪翻译，我们使用Kimetal.",
"Clarketal.,2020)开发了许多其他目标",
"(2018)的方法进行无监督翻译，并进行了一些",
"同时，测试了基于对抗性学习(Lampleetal.,",
"修改。我们基于SPDG预训练模型的流程如图",
"2017）和基于变压器的文本去噪（Kimetal.,",
"2所示。我们使用目标语言的预训练语料库预",
"2018）的使用单语语料库的神经机器翻译",
"训练该语言的seq2seq去噪模型。接下来，对",
"（NMT）的无监督方法，并证明了有希望的结",
"于源语言的每个预训练文档，我们使用字典逐",
"果。使用双语词典在预训练无监督翻译模型中",
"字将其翻译为目标语言。然后，我们将此逐字",
"定义锚点是成功的（Duanetal.,2020），但从未",
"翻译的文档提供给具有去噪目标的预训练模",
"用于生成大规模监督翻译的数据。我们的工作",
"型，以提高其质量并恢复缺失的单词",
"不同于使用字典作为锚点来学习编码器（Duan",
"使用此方法，我们可以生成每个预训练文档"
] | {
"arxiv_id": "2304.01282",
"language": "zh",
"page_idx": 2
} |
||
2310.15455-zh-0 | 2310.15455-zh-0.jpg | Exploring Mobile UI Layout Generation
using Large Language Models Guided by UI Grammar
Yuwen Lu$^{1}$ Ziang Tong$^1$ Qinyi Zhao$^2$ Chengzhi Zhang$^3$ Toby Jia-Jun Li$^{1}$
3in
Abstract
规模改善用户体验和减少 UI/UX 设计人员 (Lu et al.,
2022; Knearem et al., 2023) 工作量的早期步骤
随着大规模移动 UI 数据集 RICO (Deka et al., 2017) 的发布，已经提出了几种用于移动 UI 布局生成的 AI 模型体系结构。这些体系结构包括但不限于生成对抗网络(GAN)(Li et al., 2019; Kikuchi et al., 2021) , 变分自动编码器 (VAE) (Arroyo et al., 2021; Jing et al., 2023) ,扩散模型 (Cheng et al., 2023; Hui et al., 2023) ,图神经网络 (GNN) (Lee et al., 2020) 和其他基于变压器的神经网络 (Li et al., 2020; Gupta et al., 2021; Huang et al., 2021; Kong et al., 2022; Sobolevsky et al., 2023)
The recent advances in Large Language Models (LLMs) have stimulated interest among researchers and industry professionals, particularly in their application to tasks concerning mobile user interfaces (UIs). This position paper investigates the use of LLMs for UI layout generation. Central to our exploration is the introduction of UI grammar —a novel approach we proposed to represent the hierarchical structure inherent in UI screens. The aim of this approach is to guide the generative capacities of LLMs more effectively and improve the explainability and controllability of the process. Initial experiments conducted with GPT-4 showed the promising capability of LLMs to produce high-quality user interfaces via in-context learning. Furthermore, our preliminary comparative study suggested the potential of the grammar-based approach in improving the quality of generative results in specific aspects.
1.2. Large Language Models for UI Tasks
最近的研究工作探索了 LLM 在各种 UI 相关任务上的一些能力。Wang et al. (2023) 利用大型语言模型 (LLM) 通过上下文学习和思想链提示来执行 4个 UI 建模任务。Liu et al. (2023) 通过使用 LLM 模拟与 GUI 的类似人类的交互来进行自动 GUI 测试。Kargaran et al. (2023)通过自然语言描述设计师的意图和设计目标，探索了 LLM 的用户界面菜单设计。这些努力证明了 LLM 通过仔细的交互设计和提示技术有效地与 UI 合作的能力。一些实验在 UI 任务评估指标上也表现出了有竞争力的表现，而不需要大规模的数据集或广泛的训练过程
## 1. Introduction
1.1. Mobile UI Layout Generation
1.3. Research problem and objectives
在这项工作中，我们试图探索 LLM 生成移动 UI 布局的潜力。具体来说，我们着手确定如何在一次性学习场景中利用 LLM 的上下文学习功能来生成高质量的UI 布局。这里的一个关键挑战涉及将 UI 元素固有的层次结构表示和集成到生成过程中
用户界面 (UI)或图形用户界面 (GUI)的布局生成已由 AI 和人机交互(HCI)的研究人员进行了探索。从机器学习的角度来看，UI 固有的多模态特性为有效的 UI 建模，理解和生成$\overline{\mathrm{E}}($Jiang et al.,2023;2022) 提出了有趣的研究挑战；从 HCI 的角度来看，已经深人研究了 UI 作为良好用户体验 (UX)的媒介。学术界和工业界已经开发了各种需求发现 (Dow et al., 2005; Zimmerman $\&$ Forlizzi, 2017; Martelaro &Ju, 2017) 和可用性研究 (Nielsen,1994;2005)方法，以提高 UI 的可用性，功能和用户友好性。解决这些挑战被视为大
$^{1}$University of Notre Dame, Notre Dame, USA$^2$Fudan University, Shanghai, China $^3$Carnegie Mellon University, Pittsburgh, USA . Correspondence to: Yuwen Lu <ylu$23@$nd.edu>.
AI & HCI Workshop at the $40^{th}$ International Conference on Machine Learning, (ICML), Honolulu, Hawaii, USA. 2023. Copyright 2021 by the author(s). | [
[
"0.286",
"0.11554",
"0.6855",
"0.11554",
"0.6855",
"0.13",
"0.286",
"0.13"
],
[
"0.2141",
"0.1383",
"0.7607",
"0.1383",
"0.7607",
"0.1528",
"0.2141",
"0.1528"
],
[
"0.2051",
"0.202",
"0.767",
"0.202",
"0.767",
"0.2153",
"0.2051",
"0.2153"
],
[
"0.08826",
"0.2184",
"0.1316",
"0.2184",
"0.1316",
"0.2329",
"0.08826",
"0.2329"
],
[
"0.2427",
"0.2291",
"0.3186",
"0.2311",
"0.3186",
"0.2468",
"0.2427",
"0.2443"
],
[
"0.5024",
"0.2317",
"0.8857",
"0.2317",
"0.8857",
"0.2456",
"0.5024",
"0.2456"
],
[
"0.1217",
"0.252",
"0.4412",
"0.252",
"0.4412",
"0.2651",
"0.1217",
"0.2651"
],
[
"0.501",
"0.2468",
"0.8364",
"0.2468",
"0.8364",
"0.26",
"0.501",
"0.26"
],
[
"0.1209",
"0.2651",
"0.4412",
"0.2664",
"0.4412",
"0.281",
"0.1209",
"0.2798"
],
[
"0.5024",
"0.2695",
"0.884",
"0.2695",
"0.884",
"0.2827",
"0.5024",
"0.2827"
],
[
"0.1217",
"0.2822",
"0.4412",
"0.2822",
"0.4412",
"0.2954",
"0.1217",
"0.2954"
],
[
"0.5024",
"0.2854",
"0.8857",
"0.2854",
"0.8857",
"0.2986",
"0.5024",
"0.2986"
],
[
"0.1217",
"0.2966",
"0.4412",
"0.2966",
"0.4412",
"0.3113",
"0.1217",
"0.3113"
],
[
"0.501",
"0.2998",
"0.884",
"0.3005",
"0.884",
"0.3145",
"0.501",
"0.3132"
],
[
"0.1217",
"0.3125",
"0.4395",
"0.3125",
"0.4395",
"0.3257",
"0.1217",
"0.3257"
],
[
"0.5024",
"0.315",
"0.8857",
"0.315",
"0.8857",
"0.3289",
"0.5024",
"0.3289"
],
[
"0.1209",
"0.327",
"0.4412",
"0.3257",
"0.4412",
"0.3403",
"0.1209",
"0.3416"
],
[
"0.5034",
"0.33",
"0.884",
"0.33",
"0.884",
"0.3435",
"0.5034",
"0.3435"
],
[
"0.1217",
"0.3416",
"0.4395",
"0.3416",
"0.4395",
"0.356",
"0.1217",
"0.356"
],
[
"0.5024",
"0.3452",
"0.8857",
"0.3452",
"0.8857",
"0.3591",
"0.5024",
"0.3591"
],
[
"0.1209",
"0.3567",
"0.4412",
"0.3567",
"0.4412",
"0.3706",
"0.1209",
"0.3706"
],
[
"0.517",
"0.3599",
"0.884",
"0.3599",
"0.884",
"0.3738",
"0.517",
"0.3738"
],
[
"0.1209",
"0.3726",
"0.4412",
"0.3706",
"0.4412",
"0.385",
"0.1209",
"0.387"
],
[
"0.501",
"0.3757",
"0.8857",
"0.375",
"0.8857",
"0.3894",
"0.501",
"0.3901"
],
[
"0.1217",
"0.387",
"0.4412",
"0.387",
"0.4412",
"0.401",
"0.1217",
"0.401"
],
[
"0.5024",
"0.3914",
"0.8857",
"0.3914",
"0.8857",
"0.4048",
"0.5024",
"0.4048"
],
[
"0.1209",
"0.4028",
"0.4387",
"0.4028",
"0.4387",
"0.4167",
"0.1209",
"0.4167"
],
[
"0.501",
"0.4048",
"0.5435",
"0.4048",
"0.5435",
"0.4192",
"0.501",
"0.4192"
],
[
"0.1217",
"0.4172",
"0.4395",
"0.4172",
"0.4395",
"0.4312",
"0.1217",
"0.4312"
],
[
"0.1217",
"0.433",
"0.4395",
"0.433",
"0.4395",
"0.4463",
"0.1217",
"0.4463"
],
[
"0.501",
"0.4363",
"0.8022",
"0.4375",
"0.8022",
"0.452",
"0.501",
"0.4507"
],
[
"0.1217",
"0.4475",
"0.4412",
"0.4475",
"0.4412",
"0.461",
"0.1217",
"0.461"
],
[
"0.1217",
"0.4634",
"0.4395",
"0.4634",
"0.4395",
"0.4766",
"0.1217",
"0.4766"
],
[
"0.5024",
"0.4602",
"0.884",
"0.4602",
"0.884",
"0.4749",
"0.5024",
"0.4749"
],
[
"0.1209",
"0.4778",
"0.4412",
"0.4778",
"0.4412",
"0.4912",
"0.1209",
"0.4912"
],
[
"0.5024",
"0.4749",
"0.884",
"0.476",
"0.884",
"0.4905",
"0.5024",
"0.4893"
],
[
"0.1209",
"0.4924",
"0.4387",
"0.4924",
"0.4387",
"0.507",
"0.1209",
"0.507"
],
[
"0.5024",
"0.4912",
"0.886",
"0.4912",
"0.886",
"0.505",
"0.5024",
"0.505"
],
[
"0.1217",
"0.5083",
"0.4412",
"0.5083",
"0.4412",
"0.5215",
"0.1217",
"0.5215"
],
[
"0.501",
"0.506",
"0.8857",
"0.506",
"0.8857",
"0.5195",
"0.501",
"0.5195"
],
[
"0.1209",
"0.5215",
"0.4412",
"0.523",
"0.4412",
"0.537",
"0.1209",
"0.536"
],
[
"0.5024",
"0.5215",
"0.8857",
"0.5215",
"0.8857",
"0.535",
"0.5024",
"0.535"
],
[
"0.1209",
"0.5386",
"0.4412",
"0.5386",
"0.4412",
"0.552",
"0.1209",
"0.552"
],
[
"0.5024",
"0.5386",
"0.8823",
"0.5386",
"0.8823",
"0.549",
"0.5024",
"0.549"
],
[
"0.1209",
"0.5537",
"0.2566",
"0.5537",
"0.2566",
"0.5674",
"0.1209",
"0.5674"
],
[
"0.5024",
"0.552",
"0.884",
"0.552",
"0.884",
"0.565",
"0.5024",
"0.565"
],
[
"0.5024",
"0.5664",
"0.8857",
"0.5664",
"0.8857",
"0.58",
"0.5024",
"0.58"
],
[
"0.501",
"0.582",
"0.8857",
"0.582",
"0.8857",
"0.595",
"0.501",
"0.595"
],
[
"0.08905",
"0.595",
"0.2222",
"0.595",
"0.2222",
"0.61",
"0.08905",
"0.61"
],
[
"0.501",
"0.595",
"0.8857",
"0.5967",
"0.8857",
"0.6113",
"0.501",
"0.61"
],
[
"0.5024",
"0.6113",
"0.683",
"0.6113",
"0.683",
"0.625",
"0.5024",
"0.625"
],
[
"0.08905",
"0.622",
"0.3381",
"0.622",
"0.3381",
"0.6353",
"0.08905",
"0.6353"
],
[
"0.0907",
"0.6445",
"0.4739",
"0.6445",
"0.4739",
"0.6577",
"0.0907",
"0.6577"
],
[
"0.5",
"0.6426",
"0.773",
"0.643",
"0.773",
"0.6577",
"0.5",
"0.657"
],
[
"0.09235",
"0.66",
"0.4722",
"0.66",
"0.4722",
"0.674",
"0.09235",
"0.674"
],
[
"0.5024",
"0.6675",
"0.8857",
"0.6675",
"0.8857",
"0.681",
"0.5024",
"0.681"
],
[
"0.08905",
"0.674",
"0.4739",
"0.675",
"0.4739",
"0.6895",
"0.08905",
"0.688"
],
[
"0.501",
"0.6816",
"0.88",
"0.6816",
"0.88",
"0.696",
"0.501",
"0.696"
],
[
"0.0907",
"0.6895",
"0.4722",
"0.6895",
"0.4722",
"0.7026",
"0.0907",
"0.7026"
],
[
"0.501",
"0.6978",
"0.884",
"0.6978",
"0.884",
"0.711",
"0.501",
"0.711"
],
[
"0.08905",
"0.704",
"0.4722",
"0.705",
"0.4722",
"0.7197",
"0.08905",
"0.7183"
],
[
"0.501",
"0.712",
"0.8857",
"0.712",
"0.8857",
"0.726",
"0.501",
"0.726"
],
[
"0.0907",
"0.72",
"0.4739",
"0.72",
"0.4739",
"0.7344",
"0.0907",
"0.7344"
],
[
"0.501",
"0.728",
"0.746",
"0.728",
"0.746",
"0.741",
"0.501",
"0.741"
],
[
"0.09235",
"0.7373",
"0.4714",
"0.7373",
"0.4714",
"0.7476",
"0.09235",
"0.7476"
],
[
"0.09235",
"0.752",
"0.4714",
"0.752",
"0.4714",
"0.762",
"0.09235",
"0.762"
],
[
"0.0907",
"0.765",
"0.4739",
"0.765",
"0.4739",
"0.779",
"0.0907",
"0.779"
],
[
"0.0907",
"0.78",
"0.4722",
"0.78",
"0.4722",
"0.7935",
"0.0907",
"0.7935"
],
[
"0.1095",
"0.8037",
"0.4722",
"0.803",
"0.4722",
"0.8174",
"0.1095",
"0.8184"
],
[
"0.08905",
"0.8174",
"0.4722",
"0.8174",
"0.4722",
"0.8306",
"0.08905",
"0.8306"
],
[
"0.08905",
"0.83",
"0.4739",
"0.83",
"0.4739",
"0.8433",
"0.08905",
"0.8433"
],
[
"0.08496",
"0.8413",
"0.2181",
"0.8423",
"0.2181",
"0.857",
"0.08496",
"0.856"
],
[
"0.08905",
"0.866",
"0.4722",
"0.866",
"0.4722",
"0.881",
"0.08905",
"0.881"
],
[
"0.0907",
"0.88",
"0.4722",
"0.88",
"0.4722",
"0.893",
"0.0907",
"0.893"
],
[
"0.08826",
"0.8916",
"0.3506",
"0.8926",
"0.3506",
"0.907",
"0.08826",
"0.906"
]
] | [
"Exploring Mobile UI Layout Generation",
"using Large Language Models Guided by UI Grammar",
"Yuwen Lu1 Ziang Tong1 Qinyi Zhao2 Chengzhi Zhang3Toby Jia-Jun Li1",
"。3in",
"Abstract",
"规模改善用户体验和减少UI/UX设计人员（Luetal.，",
"The recent advances in Large Language Mod-",
"2022;Knearemetal.，2023）工作量的早期步骤",
"els (LLMs) have stimulated interest among",
"随着大规模移动UI数据集RICO（Dekaetal.,2017)",
"researchers and industry professionals, partic-",
"的发布，已经提出了几种用于移动UI布局生成的AI",
"ularly in their application to tasks concerning",
"模型体系结构。这些体系结构包括但不限于生成对抗",
"mobile user interfaces (UIs).This position",
"网络（GAN)(Lietal.,2019;Kikuchietal.,2021)，变",
"paper investigates the use of LLMs for UI",
"分自动编码器（VAE）（Arroyoetal.，2021Jingetal.",
"layout generation. Central to our exploration",
"2023），扩散模型（Chengetal.2023;Huietal.,2023)",
"is the introduction of UI grammar -a novel",
"图神经网络（GNN）（Leeetal.,2020）和其他基于",
"approach we proposed to represent the hierar",
"变压器的神经网络（Lietal.，2020；Guptaetal.,2021",
"chical structure inherent in UI screens. The",
"Huang et al., 2021; Kong et al., 2022; Sobolevsky et al.,",
"aim of this approach is to guide the genera",
"2023)",
"tive capacities of LLMs more effectively and",
"improve the explainability and controllability",
"1.2. Large Language Models for UI Tasks",
"of the process. Initial experiments conducted",
"with GPT-4 showed the promising capability",
"最近的研究工作探索了LLM在各种UI相关任务",
"of LLMs to produce high-quality user inter-",
"上的一些能力。Wangetal.（2023）利用大型语言模",
"faces via in-context learning. Furthermore,",
"型（LLM）通过上下文学习和思想链提示来执行4",
"our preliminary comparative study suggested",
"个UI建模任务。Liuetal.（2023）通过使用LLM模",
"the potential of the grammar-based approach",
"拟与GUI的类似人类的交互来进行自动GUI测试。",
"in improving the quality of generative results",
"Kargaranetal.（2023）通过自然语言描述设计师的意",
"in specific aspects.",
"图和设计目标，探索了LLM的用户界面菜单设计。这",
"些努力证明了LLM通过仔细的交互设计和提示技术",
"有效地与UI合作的能力。一些实验在UI任务评估指",
"1.Introduction",
"标上也表现出了有竞争力的表现，而不需要大规模的",
"数据集或广泛的训练过程",
"1.l. Mobile UI Layout Generation",
"用户界面（UI）或图形用户界面（GUI）的布局生成",
"1.3. Research problem and objectives",
"已由AI和人机交互（HCI）的研究人员进行了探索。",
"在这项工作中，我们试图探索LLM生成移动UI布局",
"从机器学习的角度来看，UI固有的多模态特性为有效",
"的潜力。具体来说，我们着手确定如何在一次性学习",
"的UI建模，理解和生成E(Jiangetal.，2023:2022）提",
"场景中利用LLM的上下文学习功能来生成高质量的",
"出了有趣的研究挑战；从HCI的角度来看，已经深人",
"UI布局。这里的一个关键挑战涉及将UI元素固有的",
"研究了UI作为良好用户体验（UX）的媒介。学术界",
"层次结构表示和集成到生成过程中",
"和工业界已经开发了各种需求发现（Dowetal.，2005：",
"Zimmerman &Forlizzi.2017:Martelaro &Ju,2017)",
"和可用性研究（Nielsen，1994；2005）方法，以提高UI",
"的可用性，功能和用户友好性。解决这些挑战被视为大",
"1 University of Notre Dame, Notre Dame, USA 2Fu",
"dan University, Shanghai, China 3 Carnegie Mellon Uni",
"versity, Pittsburgh, USA.Correspondence to: Yuwen Lu",
"<[email protected] >.",
"AI & HCI Workshop at the 40th International Conference",
"on Machine Learning, (ICML), Honolulu, Hawaii, USA.",
"2023. Copyright 2021 by the author(s)."
] | {
"arxiv_id": "2310.15455",
"language": "zh",
"page_idx": 0
} |
||
2209.01558-zh-0 | 2209.01558-zh-0.jpg | Scalable Adversarial Online Continual Learning
Tanmoy Dam $^{1\star}$,Mahardhika Pratama$\boxtimes^2\star$,MD Meftahul Ferdaus $^{3\star}$,Sreenatha
Anavatti $^1$,Hussein Abbas $^1$
$^{1}$ SEIT, University of New South Wales, Canberra, Australia
[email protected],{ s.anavatti,h.abbas{ Qadfa.edu.au
$^2$ STEM, University of South Australia, Adelaide, Australia
[email protected]
$^{3}$ ATMRI, Nanyang Technological University, Singapore
mdmeftahul. ferdaus@ntu. edu. s g
Abstract. Adversarial continual learning is effective for continual learning problems because of the presence of feature alignment process generating task-invariant features having low susceptibility to the catastrophic forgetting problem. Nevertheless, the ACL method imposes considerable complexities because it relies on task-specific networks and discriminators. It also goes through an iterative training process which does not fit for online (one-epoch) continual learning problems This paper proposes a scalable adversarial continual learning (SCALE) method putting forward a parameter generator transforming common features into taskspecific features and a single discriminator in the adversarial game to induce common features. The training process is carried out in meta-learning fashions using a new combination of three loss functions. SCALE outperforms prominent baselines with noticeable margins in both accuracy and execution time
Keywords: Continual Learning$\cdot$Lifelong Learning$\cdot$Incremental Learning
## 1 Introduction
持续学习 (CL)因其在改进现有深度学习算法以处理长期学习问题方面的重要性而受到广泛关注。与传统的学习问题不同，在传统的学习问题中，一次只提供一个任务的深度模型，一个连续的学习者将面临一系列不同的任务，这些任务在不同的分布或不同的目标类别[9]方面具有不同的特征。由于目标是开发一种永无止境的学习算法，该算法必须很好地扩展到可能无限数量的任务，因此在面对新任务时不可能从头开始执行重新训练过程。CL 问题禁止过度使用旧数据样本，并且只能将少量旧数据样本存储在内存中
CL 问题导致两个主要的研究问题。第一个问题是在接受新任务时如何快速转移旧任务的相关知识。第二个问题是在学习新任务时如何避免旧任务的泛化损失。学习新任务时，旧任务的泛化能力丧失被称为灾难性遗忘问题[9,20],其中学习新任务会灾难性地覆盖旧任务的重要参数。持续学习者必须从流任务中积累知识，并随着时间的推移提高智力
存在三种持续学习[20]的常用方法：基于内存的方法[16],基于结构的方法[26] ,基于正则化的方法[14]。基于正则化的方法使用正则化项来惩罚旧任务的$^{\star}$ equal contribution | [
[
"0.2573",
"0.1433",
"0.746",
"0.1465",
"0.746",
"0.1642",
"0.2573",
"0.161"
],
[
"0.2303",
"0.1901",
"0.7715",
"0.1901",
"0.7715",
"0.2052",
"0.2303",
"0.2052"
],
[
"0.406",
"0.207",
"0.5947",
"0.207",
"0.5947",
"0.2203",
"0.406",
"0.2203"
],
[
"0.3186",
"0.2317",
"0.683",
"0.2343",
"0.683",
"0.2488",
"0.3186",
"0.2468"
],
[
"0.2279",
"0.2488",
"0.7754",
"0.2488",
"0.7754",
"0.262",
"0.2279",
"0.262"
],
[
"0.3228",
"0.26",
"0.6807",
"0.2615",
"0.6807",
"0.276",
"0.3228",
"0.2747"
],
[
"0.3865",
"0.2766",
"0.6177",
"0.2766",
"0.6177",
"0.2903",
"0.3865",
"0.2903"
],
[
"0.33",
"0.2874",
"0.671",
"0.2903",
"0.671",
"0.305",
"0.33",
"0.3018"
],
[
"0.3757",
"0.3062",
"0.6284",
"0.3062",
"0.6284",
"0.3164",
"0.3757",
"0.3164"
],
[
"0.2656",
"0.3542",
"0.738",
"0.3547",
"0.738",
"0.3694",
"0.2656",
"0.3687"
],
[
"0.2646",
"0.3694",
"0.7466",
"0.3694",
"0.7466",
"0.3833",
"0.2646",
"0.3833"
],
[
"0.2656",
"0.3833",
"0.738",
"0.3833",
"0.738",
"0.3965",
"0.2656",
"0.3965"
],
[
"0.2656",
"0.3965",
"0.739",
"0.3965",
"0.739",
"0.4104",
"0.2656",
"0.4104"
],
[
"0.267",
"0.4124",
"0.7344",
"0.4124",
"0.7344",
"0.4224",
"0.267",
"0.4224"
],
[
"0.2646",
"0.4248",
"0.738",
"0.4248",
"0.738",
"0.4382",
"0.2646",
"0.4382"
],
[
"0.2656",
"0.4382",
"0.738",
"0.4382",
"0.738",
"0.452",
"0.2656",
"0.452"
],
[
"0.2656",
"0.4539",
"0.7363",
"0.4539",
"0.7363",
"0.464",
"0.2656",
"0.464"
],
[
"0.2656",
"0.4666",
"0.739",
"0.4666",
"0.739",
"0.4797",
"0.2656",
"0.4797"
],
[
"0.267",
"0.4817",
"0.738",
"0.4817",
"0.738",
"0.4924",
"0.267",
"0.4924"
],
[
"0.2656",
"0.4937",
"0.74",
"0.4937",
"0.74",
"0.507",
"0.2656",
"0.507"
],
[
"0.2632",
"0.5083",
"0.683",
"0.5083",
"0.683",
"0.5215",
"0.2632",
"0.5215"
],
[
"0.2656",
"0.534",
"0.7134",
"0.534",
"0.7134",
"0.5474",
"0.2656",
"0.5474"
],
[
"0.2181",
"0.5757",
"0.2566",
"0.5757",
"0.2566",
"0.589",
"0.2181",
"0.589"
],
[
"0.2467",
"0.5747",
"0.361",
"0.5757",
"0.361",
"0.5903",
"0.2467",
"0.589"
],
[
"0.2181",
"0.607",
"0.7837",
"0.607",
"0.7837",
"0.6206",
"0.2181",
"0.6206"
],
[
"0.2197",
"0.6226",
"0.785",
"0.6226",
"0.785",
"0.636",
"0.2197",
"0.636"
],
[
"0.2361",
"0.637",
"0.7837",
"0.637",
"0.7837",
"0.651",
"0.2361",
"0.651"
],
[
"0.2197",
"0.654",
"0.7837",
"0.654",
"0.7837",
"0.664",
"0.2197",
"0.664"
],
[
"0.2197",
"0.6685",
"0.785",
"0.6685",
"0.785",
"0.6816",
"0.2197",
"0.6816"
],
[
"0.2214",
"0.6816",
"0.785",
"0.6816",
"0.785",
"0.696",
"0.2214",
"0.696"
],
[
"0.2181",
"0.6978",
"0.564",
"0.6978",
"0.564",
"0.711",
"0.2181",
"0.711"
],
[
"0.2443",
"0.712",
"0.785",
"0.712",
"0.785",
"0.726",
"0.2443",
"0.726"
],
[
"0.2181",
"0.728",
"0.785",
"0.728",
"0.785",
"0.741",
"0.2181",
"0.741"
],
[
"0.2181",
"0.742",
"0.785",
"0.742",
"0.785",
"0.756",
"0.2181",
"0.756"
],
[
"0.2214",
"0.7583",
"0.785",
"0.7583",
"0.785",
"0.7715",
"0.2214",
"0.7715"
],
[
"0.2197",
"0.7734",
"0.4968",
"0.7734",
"0.4968",
"0.787",
"0.2197",
"0.787"
],
[
"0.2443",
"0.7886",
"0.785",
"0.7886",
"0.785",
"0.802",
"0.2443",
"0.802"
],
[
"0.2197",
"0.8037",
"0.7837",
"0.8037",
"0.7837",
"0.8174",
"0.2197",
"0.8174"
],
[
"0.2222",
"0.8257",
"0.3447",
"0.827",
"0.3447",
"0.8413",
"0.2222",
"0.8403"
]
] | [
"Scalable Adversarial Online Continual Learning",
"Tanmoy Dam 1 *, Mahardhika Pratama? *, MD Meftahul Ferdaus 3 *, Sreenatha",
"Anavatti ,Hussein Abbas 1",
"1SEIT,University ofNewSouthWales,Canberra,Australia",
"[email protected], ( s.anavatti,h.abbas @adfa.edu.au",
"2 STEM, University of South Australia, Adelaide, Australia",
"[email protected]",
"3ATMRI, Nanyang Technological University, Singapore",
"[email protected]",
"Abstract.Adversarial continual learning is effective for continual learning prob-",
"lems because of the presence of feature alignment process generating task-invariant",
"features having low susceptibility to the catastrophic forgetting problem,Never-",
"theless, the ACL method imposes considerable complexities because it relies on",
"task-specificnetworksanddiscriminators.Italsogoesthroughaniterativetrain",
"ing process which does not fit for online (one-epoch) continual learning problems.",
"This paper proposes a scalable adversarial continual learning (SCALE) method",
"puttingforwardaparametergeneratortransformingcommonfeaturesintotask",
"specific features and a single discriminator in the adversarial game to induce",
"common features.The training process is carried out in meta-learning fashions",
"using a new combination of three loss functions. SCALE outperforms prominent",
"baselines with noticeable margins in both accuracy and execution time.",
"Keywords: Continual Learning - Lifelong Learning - Incremental Learning",
"1",
"Introduction",
"持续学习（CL）因其在改进现有深度学习算法以处理长期学习问题方面的重要",
"性而受到广泛关注。与传统的学习问题不同，在传统的学习问题中，一次只提供",
"个任务的深度模型，一个连续的学习者将面临一系列不同的任务，这些任务",
"在不同的分布或不同的自标类别91方面具有不同的特征。由于自标是开发一种",
"永无止境的学习算法，该算法必须很好地扩展到可能无限数量的任务，因此在",
"面对新任务时不可能从头开始执行重新训练过程。CL问题禁止过度使用旧数据",
"样本，并且只能将少量旧数据样本存储在内存中",
"CL问题导致两个主要的研究问题。第一个问题是在接受新任务时如何快速",
"转移旧任务的相关知识。第二个问题是在学习新任务时如何避免旧任务的泛化",
"损失。学习新任务时，旧任务的泛化能力丧失被称为灾难性遗忘问题[9,20]，其",
"中学习新任务会灾难性地覆盖旧任务的重要参数。持续学习者必须从流任务中",
"积累知识，并随着时间的推移提高智力",
"存在三种持续学习[20]的常用方法：基于内存的方法[16]，基于结构的方法",
"[26]，基于正则化的方法[14]。基于正则化的方法使用正则化项来惩罚旧任务的",
"*equal contribution"
] | {
"arxiv_id": "2209.01558",
"language": "zh",
"page_idx": 0
} |
||
2209.01558-zh-1 | 2209.01558-zh-1.jpg | 2 T. Dam, M. Pratama et al.
重要参数，以免在学习新任务时发生变化。虽然这种方法计算量小，易于实现但这种方法不能很好地适用于大规模 CL 问题，因为很难获得所有任务的重叠区域。基于结构的方法应用网络增长策略来适应新任务，同时冻结旧参数以防止灾难性的遗忘问题。如果网络增长阶段没有得到适当控制，或者结构学习机制通常是通过计算昂贵的架构搜索方法完成的，那么这种方法会带来昂贵的复杂性因此在在线持续学习环境中不可行。基于内存的方法存储要与新样本一起重播的旧样本的一小部分，以处理灾难性的遗忘问题。与前两种方法相比，这种方法通常会提高学习性能。这种方法的根本挑战是保持适度的内存大小。SCALE 在这里被归类为一种基于内存的方法，其中提出了一个存储旧样本的微小情节内存，用于体验重放机制
[10]中提出了对抗性持续学习 (ACL)的概念。主要思想是利用对抗性学习策略[13,12] 来提取所有任务的任务对齐特征，这些任务被认为比特定于任务的特征更容易忘记。它提供公共特征和私有特征之间的不相交表示，以组合为多头分类器的输人。这种方法的主要瓶颈在于昂贵的复杂性，因为私有特征是由特定于任务的网络生成的，而公共特征是由特定于任务的鉴别器玩的对抗性游戏制作的。此外，ACL 基于迭代训练机制，不适合在线 (单历元) 持续学习问题
本文提出了可扩展的对抗性连续学习 (SCALE),通过参数生成器网络和单个鉴别器显着降低了 ACL 的复杂性。参数生成器网络产生缩放和移动参数，将对抗性学习机制产生的任务不变特征转换为特定于任务的特征[21,22]。我们的方法不需要存储特定于任务的参数，而是参数生成器网络预测这些导致私有特征的参数。私人功能的生产是通过两个轻量级操作进行的，缩放和移位。使用基础网络的验证损失，以元学习的方式训练参数生成器，即。e。特征提取器和分类器。元学习策略由两个数据分区完成：训练集和验证集，描述新旧概念。训练集更新基础网络，而验证集训练参数生成器。我们的方法与[22]不同，[22]采用对抗性训练方法来产生任务不变的特征，并且我们不需要按照[22]构造两个不同的记忆。与 ACL 不同，对抗性游戏由单个鉴别器进行，没有任何灾难性问题， 同时仍然很好地调整所有任务的功能
SCALE 在准确性和遗忘指数方面超过了 1% 的显着基线，同时在执行时间方面表现出了显着的改进。消融研究，记忆分析和敏感性分析进一步证实了SCALE 对于在线(一个时期) 持续学习问题的优势。本文提供了四个主要贡献：1)提出了一种新的在线连续学习方法，即 SCALE;2)我们的方法提供了一种可扩展的对抗性连续学习方法，该方法仅依赖于单个参数生成器来进行特征转换，从而导致特定于任务的特征和单个鉴别器来诱导任务不变的特征；3)使用三个损失函数的新组合以元学习方式完成训练过程：交叉熵损失函数 XTHadX, MADE。尽管[10] 中已经存在对抗性损失函数，但这里使用 BAGAN[18]的概念而不是梯度反转策略[10,12] 来进行对抗性博弈；4) https://github.com/TanmDL/SCALE 中提供了所有源代码，数据和原始数值结果，以帮助进一步研究
# 2 $\textbf{Related Works}$
基于正则化的方法依赖于损失函数中的惩罚项，以防止旧任务的重要参数出现重大偏差。L2 正则化策略与表示网络参数重要性的参数重要性矩阵相结合。提出了不同的策略来构建参数重要性矩阵：弹性权重合并(EWC)利用 Fisher 重要性矩阵(FIM)[14],突触智能(SI)利用累积梯度[32],记忆感知突触(MAS) 采用无监督和在线标准[1]。在线 EWC(oEWC)提出了使用拉普拉斯近似[28] | [
[
"0.2157",
"0.11743",
"0.2632",
"0.11743",
"0.2632",
"0.1288",
"0.2157",
"0.1288"
],
[
"0.2712",
"0.11615",
"0.4248",
"0.11615",
"0.4248",
"0.13",
"0.2712",
"0.13"
],
[
"0.2197",
"0.1483",
"0.7837",
"0.1483",
"0.7837",
"0.1622",
"0.2197",
"0.1622"
],
[
"0.2214",
"0.1654",
"0.7837",
"0.1654",
"0.7837",
"0.1755",
"0.2214",
"0.1755"
],
[
"0.2214",
"0.1799",
"0.7817",
"0.1799",
"0.7817",
"0.1901",
"0.2214",
"0.1901"
],
[
"0.2197",
"0.1932",
"0.7837",
"0.1932",
"0.7837",
"0.207",
"0.2197",
"0.207"
],
[
"0.2181",
"0.209",
"0.7817",
"0.209",
"0.7817",
"0.2229",
"0.2181",
"0.2229"
],
[
"0.2197",
"0.2247",
"0.785",
"0.2247",
"0.785",
"0.238",
"0.2197",
"0.238"
],
[
"0.2214",
"0.2393",
"0.785",
"0.2393",
"0.785",
"0.2524",
"0.2214",
"0.2524"
],
[
"0.2197",
"0.254",
"0.785",
"0.254",
"0.785",
"0.2676",
"0.2197",
"0.2676"
],
[
"0.2181",
"0.2676",
"0.785",
"0.2683",
"0.785",
"0.2827",
"0.2181",
"0.2822"
],
[
"0.2181",
"0.2854",
"0.3823",
"0.2854",
"0.3823",
"0.2986",
"0.2181",
"0.2986"
],
[
"0.2443",
"0.2998",
"0.7837",
"0.2998",
"0.7837",
"0.3132",
"0.2443",
"0.3132"
],
[
"0.2214",
"0.3145",
"0.785",
"0.3145",
"0.785",
"0.3276",
"0.2214",
"0.3276"
],
[
"0.2181",
"0.33",
"0.7837",
"0.33",
"0.7837",
"0.3435",
"0.2181",
"0.3435"
],
[
"0.2214",
"0.3467",
"0.7817",
"0.3467",
"0.7817",
"0.3567",
"0.2214",
"0.3567"
],
[
"0.2197",
"0.3599",
"0.7837",
"0.3599",
"0.7837",
"0.3738",
"0.2197",
"0.3738"
],
[
"0.2197",
"0.3757",
"0.7837",
"0.3757",
"0.7837",
"0.3894",
"0.2197",
"0.3894"
],
[
"0.246",
"0.3926",
"0.7817",
"0.3926",
"0.7817",
"0.4028",
"0.246",
"0.4028"
],
[
"0.2197",
"0.404",
"0.785",
"0.4048",
"0.785",
"0.4192",
"0.2197",
"0.4185"
],
[
"0.2197",
"0.4204",
"0.7837",
"0.4204",
"0.7837",
"0.4343",
"0.2197",
"0.4343"
],
[
"0.2222",
"0.4375",
"0.7837",
"0.4375",
"0.7837",
"0.4475",
"0.2222",
"0.4475"
],
[
"0.2197",
"0.452",
"0.7837",
"0.452",
"0.7837",
"0.4653",
"0.2197",
"0.4653"
],
[
"0.2214",
"0.4666",
"0.7837",
"0.4666",
"0.7837",
"0.4797",
"0.2214",
"0.4797"
],
[
"0.2214",
"0.483",
"0.7837",
"0.483",
"0.7837",
"0.4937",
"0.2214",
"0.4937"
],
[
"0.2214",
"0.4976",
"0.7817",
"0.4976",
"0.7817",
"0.5083",
"0.2214",
"0.5083"
],
[
"0.2197",
"0.511",
"0.7837",
"0.511",
"0.7837",
"0.5244",
"0.2197",
"0.5244"
],
[
"0.2214",
"0.526",
"0.7876",
"0.526",
"0.7876",
"0.5405",
"0.2214",
"0.5405"
],
[
"0.2214",
"0.5415",
"0.4805",
"0.5415",
"0.4805",
"0.5547",
"0.2214",
"0.5547"
],
[
"0.2443",
"0.5596",
"0.7817",
"0.5596",
"0.7817",
"0.5693",
"0.2443",
"0.5693"
],
[
"0.2214",
"0.5728",
"0.785",
"0.5728",
"0.785",
"0.586",
"0.2214",
"0.586"
],
[
"0.2214",
"0.5884",
"0.7817",
"0.5884",
"0.7817",
"0.5986",
"0.2214",
"0.5986"
],
[
"0.2222",
"0.604",
"0.7817",
"0.604",
"0.7817",
"0.6143",
"0.2222",
"0.6143"
],
[
"0.2214",
"0.617",
"0.7837",
"0.617",
"0.7837",
"0.631",
"0.2214",
"0.631"
],
[
"0.2222",
"0.634",
"0.7817",
"0.634",
"0.7817",
"0.6445",
"0.2222",
"0.6445"
],
[
"0.2214",
"0.648",
"0.7837",
"0.648",
"0.7837",
"0.661",
"0.2214",
"0.661"
],
[
"0.2197",
"0.662",
"0.7837",
"0.662",
"0.7837",
"0.6753",
"0.2197",
"0.6753"
],
[
"0.2197",
"0.678",
"0.786",
"0.678",
"0.786",
"0.6914",
"0.2197",
"0.6914"
],
[
"0.2197",
"0.6934",
"0.6763",
"0.6934",
"0.6763",
"0.707",
"0.2197",
"0.707"
],
[
"0.2157",
"0.732",
"0.3774",
"0.732",
"0.3774",
"0.747",
"0.2157",
"0.747"
],
[
"0.2197",
"0.7666",
"0.785",
"0.7666",
"0.785",
"0.78",
"0.2197",
"0.78"
],
[
"0.2214",
"0.7827",
"0.7837",
"0.7827",
"0.7837",
"0.7935",
"0.2214",
"0.7935"
],
[
"0.2214",
"0.797",
"0.7837",
"0.797",
"0.7837",
"0.81",
"0.2214",
"0.81"
],
[
"0.2197",
"0.811",
"0.785",
"0.811",
"0.785",
"0.8247",
"0.2197",
"0.8247"
],
[
"0.2197",
"0.827",
"0.785",
"0.827",
"0.785",
"0.8403",
"0.2197",
"0.8403"
]
] | [
"2",
"T. Dam, M. Pratama et al.",
"重要参数，以免在学习新任务时发生变化。虽然这种方法计算量小，易于实现",
"但这种方法不能很好地适用于大规模CL问题，因为很难获得所有任务的重叠区",
"域。基于结构的方法应用网络增长策略来适应新任务，同时冻结旧参数以防止灾",
"难性的遗忘问题。如果网络增长阶段没有得到适当控制，或者结构学习机制通",
"常是通过计算昂贵的架构搜索方法完成的，那么这种方法会带来昂贵的复杂性",
"因此在在线持续学习环境中不可行。基于内存的方法存储要与新样本一起重播",
"的旧样本的一小部分，以处理灾难性的遗忘问题。与前两种方法相比，这种方法",
"通常会提高学习性能。这种方法的根本挑战是保持适度的内存大小。SCALE在",
"这里被归类为一种基于内存的方法，其中提出了一个存储旧样本的微小情节内",
"存，用于体验重放机制",
"[10]中提出了对抗性持续学习（ACL）的概念。主要思想是利用对抗性学习",
"策略[13,12]来提取所有任务的任务对齐特征，这些任务被认为比特定于任务的",
"特征更容易忘记。它提供公共特征和私有特征之间的不相交表示，以组合为多",
"头分类器的输入。这种方法的主要瓶颈在于昂贵的复杂性，因为私有特征是由",
"特定于任务的网络生成的，而公共特征是由特定于任务的鉴别器玩的对抗性游",
"戏制作的。此外，ACL基于迭代训练机制，不适合在线（单历元）持续学习问题",
"本文提出了可扩展的对抗性连续学习（SCALE），通过参数生成器网络和单",
"个鉴别器显着降低了ACL的复杂性。参数生成器网络产生缩放和移动参数，将",
"对抗性学习机制产生的任务不变特征转换为特定于任务的特征[21,22]。我们的",
"方法不需要存储特定于任务的参数，而是参数生成器网络预测这些导致私有特",
"征的参数。私人功能的生产是通过两个轻量级操作进行的，缩放和移位。使用基",
"础网络的验证损失，以元学习的方式训练参数生成器，即。e。、特征提取器和分",
"类器。元学习策略由两个数据分区完成：训练集和验证集，描述新旧概念。训练",
"集更新基础网络，而验证集训练参数生成器。我们的方法与[22]不同，[22]采用",
"对抗性训练方法来产生任务不变的特征，并且我们不需要按照[22]构造两个不",
"同的记忆。与ACL不同，对抗性游戏由单个鉴别器进行，没有任何灾难性问题，",
"同时仍然很好地调整所有任务的功能",
"SCALE在准确性和遗忘指数方面超过广1%的显看基线，同时在执行时间方",
"面表现出了显着的改进。消融研究，记忆分析和敏感性分析进一步证实了SCALE",
"对于在线（一个时期）持续学习问题的优势。本文提供了四个主要责献：1）提",
"出了一种新的在线连续学习方法，即SCALE；2）我们门的方法提供了一种可扩展",
"的对抗性连续学习方法，该方法仅依赖于单个参数生成器来进行特征转换，从而",
"导致特定于任务的特征和单个鉴别器来诱导任务不变的特征；3）使用三个损失",
"函数的新组合以元学习方式完成训练过程：交叉嫡损失函数XTHadX，MADE。",
"尽管[10]中已经存在对抗性损失函数，但这里使用BAGAN[18]的概念而不是",
"梯度反转策略[10,12]来进行对抗性博奔；4）https://github.com/TanmDL/SCALE",
"中提供了所有源代码，数据和原始数值结果，以帮助进一步研究",
"2RelatedWorks",
"基于正则化的方法依赖于损失函数中的惩罚项，以防止旧任务的重要参数出现",
"重大偏差。L2正则化策略与表示网络参数重要性的参数重要性矩阵相结合。提",
"出了不同的策略来构建参数重要性矩阵：弹性权重合并（EWC）利用Fisher重要",
"性矩阵（FIM）[14]，突触智能（SI）利用累积梯度[32]，记忆感知突触（MAS）",
"采用无监督和在线标准[1]。在线EWC（oEWC）提出了使用拉普拉斯近似[28]"
] | {
"arxiv_id": "2209.01558",
"language": "zh",
"page_idx": 1
} |
||
2209.01558-zh-2 | 2209.01558-zh-2.jpg | Scalable Adversarial Online Continual Learning 3
<FigureHere>
Fig. 1: Structure of ACL based on task-specific feature extractors and discriminators
的 EWC 在线版本。无遗忘学习 (LWF) 利用知识提炼 (KD) 方法来匹配当前和以前的输出。由于深度神经网络的层次性，正则化策略在神经元级别而不是在突触级别[19] 中执行得更好。[17] 遵循与 [19] 相同的原理，并使用任务间相似性的概念更进一步。这种方法允许跨相关任务共享节点。[5]中也存在另一种提高基于正则化方法的可伸缩性的尝试，其中提出了投影概念以诱导宽的局部最优区域。基于正则化的方法在很大程度上取决于任务 ID 和任务边界
基于结构的方法提供了不同的理念，其中通过添加新的网络组件来处理新任务， 同时隔离旧组件以避免灾难性的遗忘问题。开创性的方法是渐进式神经网络(PNN)[26],在处理新任务时集成了一个新的网络列。PNN 在处理一长串任务时会产生昂贵的结构复杂性。[31]提出了一种基于损失准则的网络增长条件和选择性再训练策略。[15]中提出了神经架构搜索(NAS)的概念，用于在观察新任务时选择最佳动作。在[30]中设计了类似的方法，但是使用了贝叶斯优化方法而不是 NAS 概念。这些方法在计算上令人望而却步，并要求存在任务 ID 和边界。[23,3]提出了一种数据驱动的结构学习方法，用于解决隐藏的集群，节点和层动态增长和收缩的无监督连续学习问题。这两种方法之间的关键区别在于通过知识蒸馏(KD)策略在[3]中使用基于正则化的方法，以及在[23]中使用基于质心的经验重放。数据驱动的结构学习策略不能保证在处理新任务时采取最佳行动
基于内存的方法利用存储旧数据样本子集的微小内存。内存的旧样本与当前样本交织在一起，用于体验回放，以应对灾难性的遗忘问题。iCaRL 举例说明了这种方法[24],其中 KD 方法是使用最近的样本分类策略执行的。GEM[16]和AGEM[7] 利用内存识别遗忘情况。HAL[6] 提出了锚样本最大化遗忘度量的思想，并以元学习方式构建。DER[4]设计了黑暗知识蒸馏，并成功实现了有或没有任务 ID 的改进性能。CTN 是在 [22] 中使用 [21] 的特征转换概念提出的 , 并集成了以元学习方式训练的控制器网络。ACL[10]被归类为基于内存的方法，其中使用内存来开发对抗性游戏。但是，由于使用了特定于任务的特征提取器和鉴别器，ACL 带来了相当大的复杂性。我们在这里提供了一种替代方法，其中私有特征由[21]的特征转换策略诱导，并且对抗性游戏由一个且只有一个鉴别器进行。与[22]相比，SCALE 集成了对抗性学习策略来训练共享特征提取器，以生成对灾难性遗忘问题具有鲁棒性的通用特征，并提出了三个损失函数的新组合 | [
[
"0.4436",
"0.11554",
"0.7295",
"0.11743",
"0.7295",
"0.1307",
"0.4436",
"0.1288"
],
[
"0.774",
"0.1187",
"0.786",
"0.1187",
"0.786",
"0.1276",
"0.774",
"0.1276"
],
[
"0.603",
"0.1895",
"0.6274",
"0.1895",
"0.6274",
"0.1976",
"0.603",
"0.1976"
],
[
"0.5107",
"0.1989",
"0.517",
"0.1989",
"0.517",
"0.202",
"0.5107",
"0.202"
],
[
"0.6094",
"0.2166",
"0.6216",
"0.2166",
"0.6216",
"0.2267",
"0.6094",
"0.2267"
],
[
"0.6055",
"0.2443",
"0.626",
"0.2468",
"0.625",
"0.2551",
"0.6045",
"0.2524"
],
[
"0.601",
"0.2842",
"0.626",
"0.2842",
"0.626",
"0.2954",
"0.601",
"0.2954"
],
[
"0.2263",
"0.3193",
"0.7754",
"0.3193",
"0.7754",
"0.33",
"0.2263",
"0.33"
],
[
"0.2197",
"0.3757",
"0.7837",
"0.3757",
"0.7837",
"0.3894",
"0.2197",
"0.3894"
],
[
"0.2197",
"0.3914",
"0.7837",
"0.3914",
"0.7837",
"0.4048",
"0.2197",
"0.4048"
],
[
"0.2214",
"0.4077",
"0.7817",
"0.4077",
"0.7817",
"0.4185",
"0.2214",
"0.4185"
],
[
"0.2181",
"0.4216",
"0.7837",
"0.4216",
"0.7837",
"0.435",
"0.2181",
"0.435"
],
[
"0.2214",
"0.4382",
"0.7837",
"0.4382",
"0.7837",
"0.449",
"0.2214",
"0.449"
],
[
"0.2214",
"0.452",
"0.701",
"0.452",
"0.701",
"0.4653",
"0.2214",
"0.4653"
],
[
"0.2181",
"0.473",
"0.7837",
"0.4734",
"0.7837",
"0.4873",
"0.2181",
"0.486"
],
[
"0.2214",
"0.488",
"0.785",
"0.488",
"0.785",
"0.502",
"0.2214",
"0.502"
],
[
"0.2222",
"0.505",
"0.7817",
"0.505",
"0.7817",
"0.515",
"0.2222",
"0.515"
],
[
"0.2197",
"0.518",
"0.7837",
"0.518",
"0.7837",
"0.532",
"0.2197",
"0.532"
],
[
"0.2222",
"0.535",
"0.7817",
"0.535",
"0.7817",
"0.5454",
"0.2222",
"0.5454"
],
[
"0.2222",
"0.55",
"0.7817",
"0.55",
"0.7817",
"0.56",
"0.2222",
"0.56"
],
[
"0.2197",
"0.563",
"0.785",
"0.563",
"0.785",
"0.577",
"0.2197",
"0.577"
],
[
"0.2214",
"0.579",
"0.7837",
"0.579",
"0.7837",
"0.5923",
"0.2214",
"0.5923"
],
[
"0.2181",
"0.5923",
"0.7817",
"0.5923",
"0.7817",
"0.606",
"0.2181",
"0.606"
],
[
"0.2197",
"0.6094",
"0.7837",
"0.6094",
"0.7837",
"0.6226",
"0.2197",
"0.6226"
],
[
"0.2214",
"0.6255",
"0.7837",
"0.6255",
"0.7837",
"0.636",
"0.2214",
"0.636"
],
[
"0.2214",
"0.64",
"0.2688",
"0.64",
"0.2688",
"0.6514",
"0.2214",
"0.6514"
],
[
"0.2181",
"0.661",
"0.7837",
"0.661",
"0.7837",
"0.675",
"0.2181",
"0.675"
],
[
"0.2181",
"0.675",
"0.7812",
"0.675",
"0.7812",
"0.688",
"0.2181",
"0.688"
],
[
"0.2173",
"0.69",
"0.785",
"0.6914",
"0.785",
"0.705",
"0.2173",
"0.704"
],
[
"0.2181",
"0.705",
"0.7837",
"0.7056",
"0.7837",
"0.7197",
"0.2181",
"0.7183"
],
[
"0.2197",
"0.72",
"0.7837",
"0.72",
"0.7837",
"0.7344",
"0.2197",
"0.7344"
],
[
"0.2214",
"0.7373",
"0.7817",
"0.7373",
"0.7817",
"0.7476",
"0.2214",
"0.7476"
],
[
"0.2214",
"0.752",
"0.785",
"0.752",
"0.785",
"0.765",
"0.2214",
"0.765"
],
[
"0.2214",
"0.7666",
"0.7837",
"0.7666",
"0.7837",
"0.78",
"0.2214",
"0.78"
],
[
"0.2214",
"0.781",
"0.7837",
"0.781",
"0.7837",
"0.7944",
"0.2214",
"0.7944"
],
[
"0.2197",
"0.797",
"0.785",
"0.797",
"0.785",
"0.81",
"0.2197",
"0.81"
],
[
"0.2197",
"0.811",
"0.7837",
"0.811",
"0.7837",
"0.8247",
"0.2197",
"0.8247"
],
[
"0.2197",
"0.827",
"0.773",
"0.827",
"0.773",
"0.8403",
"0.2197",
"0.8403"
]
] | [
"ScalableAdversarialOnlineContinual Learning",
"3",
"yLe",
"",
"Lo",
"k Lad",
"k Lat",
"Fig.l:StructureofACLbasedontask-specificfeatureextractorsanddiscriminators",
"的EWC在线版本。无遗忘学习（LWF）利用知识提炼（KD）方法来匹配当前和",
"以前的输出。由于深度神经网络的层次性，正则化策略在神经元级别而不是在",
"突触级别[19]中执行得更好。[17]遵循与[19]相同的原理，并使用任务间相",
"性的概念更进一步。这种方法允许跨相关任务共享节点。[5]中也存在另一种提",
"高基于正则化方法的可伸缩性的尝试，其中提出了投影概念以诱导宽的局部最",
"优区域。基于正则化的方法在很大程度上取决于任务ID和任务边界",
"基于结构的方法提供了不同的理念，其中通过添加新的网络组件来处理新任务",
"同时隔离旧组件以避免灾难性的遗忘问题。开创性的方法是渐进式神经网络",
"PNN）「26：在处理新任务时集成了一个新的网络列。PNN在处理一长串任务",
"时会产生昂贵的结构复杂性。[31]提出了一种基于损失准则的网络增长条件和",
"选择性再训练策略。「15]中提出神经架构搜索（NAS）的概念，用于在观察新",
"任务时选择最佳动作。在301中设计了类似的方法：但是使用了贝叶斯优化方",
"法而不是NAS概念。这些方法在计算上令人望而却步，并要求存在任务ID和边",
"界。[23,3]提出了一种数据驱动的结构学习方法，用于解决隐藏的集群，节点和",
"层动态增长和收缩的无监督连续学习问题。这两种方法之间的关键区别在于通",
"过知识蒸馏（KD）策略在[3]中使用基于正则化的方法，以及在[23]中使用基",
"于质心的经验重放。数据驱动的结构学习策略不能保证在处理新任务时采取最",
"佳行动",
"基于内存的方法利用存储旧数据样本子集的微小内存。内存的旧样本与当前样",
"本交织在一起，用于体验回放，以应对灾难性的遗忘问题。iCaRL举例说明了",
"这种方法[24]，其中KD方法是使用最近的样本分类策略执行的。GEM[16]和",
"AGEM[7]利用内存识别遗忘情况。HAL[6]提出了锚样本最大化遗忘度量的思",
"想，并以元学习方式构建。DER[4]设计了黑暗知识蒸馏，并成功实现了有或没",
"有任务ID的改进性能。CTN是在I221中使用[211的特征转换概念提出的：并集",
"成了以元学习方式训练的控制器网络。ACL[10]被归类为基于内存的方法，其",
"中使用内存来开发对抗性游戏。但是，由于使用了特定于任务的特征提取器和鉴",
"别器，ACL带来了相当大的复杂性。我们在这里提供了一种替代方法，其中私有",
"特征由[21]的特征转换策略诱导，并且对抗性游戏由一个且只有一个鉴别器进",
"行。与[22]相比，SCALE集成了对抗性学习策略来训练共享特征提取器，以生",
"成对灾难性遗忘问题具有鲁棒性的通用特征，并提出了三个损失函数的新组合"
] | {
"arxiv_id": "2209.01558",
"language": "zh",
"page_idx": 2
} |
||
2209.01558-zh-3 | 2209.01558-zh-3.jpg | 4 T. Dam, M. Pratama et al.
# 3 Problem Formulation
持续学习 (CL) 问题被定义为顺序到达任务$\mathcal{T}_1,\mathcal{T}_2,...,\mathcal{T}_k,k\in\{1,...,K\}$的学习问题，其中$K$表示在实践中未知的任务数。每个任务都带有三元组$\mathcal{T}_k=$ $\{ x_i, y_i, t_i\} _{i= 1}^{N_k}$ ,其中$N_k$代表任务大小。$x_i^k\in\mathcal{X}_k$表示输入图像，而$y_i^k\in \mathcal{Y} _k$ , $y_{i}^{k}=[l_{1},l_{2},...,l_{m}]$标记类标签，$t_i^k$表示任务标识符 (ID)。CL 问题的目标是构建一个连续的学习者$f_\phi(g_\theta(.))$ ,在已经看到的任务上表现良好，其中$g_\theta(.)$是特征提取器，$f_\phi(.)$是分类器。本文主要研究在线 (一个时期) 任务增量学习和域增量学习问题[29],其中任何任务$\{x_i,y_i,t_i\}\sim\mathcal{T}_k$的每个三元组仅在一个时期内学习。任务增量学习问题以每个任务的不相交类为特征，即。e。,假设$L_k$ 和$L_{k^{\prime}}$代表$k-th$任务和$k^\prime-th$任务的标签集，$\forall k,k^\prime L_k\cap L_{k^{\prime}}=\emptyset$。域增量学习问题提出了每个任务$P(X,Y)_k\neq P(X,Y)_{(k+1)}$的不同分布或域，同时仍然为每个任务保留相同的目标类。也就是说，将多头配置应用于任务增量学习问题，其中为每个任务$f_{\phi_k}(.)$创建一个独立的分类器。域增量学习问题纯粹用单头配置$f_\phi(.)$处理。$\operatorname{CL}_{\text{ 问题禁止从头开始重新训练过程 }\frac1K\sum_{k=1}^K\mathcal{L}_k,\mathcal{L}_k\triangleq\mathbb{E}_{(x,y)\sim\mathcal{D}_k}[l(f_\phi(g_\theta(x)),y)]}$ 。学习过程仅由当前任务$T_k$的数据样本和包含先前看到的任务的旧样本的微小内存$\mathcal{M}_{k-1}$支持，以克服灾难性的遗忘问题
4 Adversarial Continual Learning (ACL)
图。1 可视化对抗性连续学习方法[10]包括四个部分：共享特征提取器，特定于任务的特征提取器，特定于任务的鉴别器和多头分类器。共享特征提取器生成任务不变的特征，而特定于任务的特征提取器提供每个任务的私有特征。特定于任务的鉴别器预测任务的起源，而多头分类器产生最终的预测。训练过程由三个损失函数控制：分类损失，对抗损失和正交损失。分类损失利用影响多头分类器、共享特征提取器和特定于任务的特征提取器的交叉熵损失函数。在共享特征提取器和特定于任务的鉴别器之间以最小-最大方式执行对抗性损失。在调整特征提取器时实现梯度反转层，从而将最小化问题转换为最大化问题。也就是说，共享特征提取器被训练来欺骗特定于任务的鉴别器，并最终生成任务不变的特征。正交损失确保了共享特征提取器的任务特定特征与任务特定特征提取器的私有特征之间的明确区别
在测试阶段排除特定于任务的鉴别器，并通过将私有特征和公共特征的级联特征馈送到产生最终输出的多头分类器来执行推理阶段。由于应用了特定于任务的特征提取器和特定于任务的鉴别器，ACL 会带来很高的复杂性。我们在这里提供了一个参数生成器，用于生成缩放和移位参数以执行特征转换。因此，特定于任务的特征是在不损失泛化的情况下以低开销生成的，同时仅依赖于单个鉴别器来玩引起对齐特征的对抗性游戏。此外，ACL 依赖于违反在线持续学习要求的迭代训练过程，而 SCALE 完全在一个时期设置中运行
# 5 Learning Policy of SCALE
在图 2 和算法 1 中可视化了 SCALE 的学习过程，其中它包括四个块：特征提取器$g_\theta(\cdot)$,参数生成器$P_\varphi(\cdot)$,多头分类器$f_\phi_k(\cdot)$ (域增量学习问题中的单头分类器) 和单鉴别器$D_\xi(.)$ 。特征提取器提取由对抗性学习机制启用的任务不变特 | [
[
"0.2214",
"0.11554",
"0.2345",
"0.1206",
"0.2279",
"0.13",
"0.2141",
"0.125"
],
[
"0.2712",
"0.11615",
"0.4248",
"0.11615",
"0.4248",
"0.13",
"0.2712",
"0.13"
],
[
"0.2157",
"0.1459",
"0.4346",
"0.1477",
"0.4346",
"0.1622",
"0.2157",
"0.1604"
],
[
"0.2197",
"0.1799",
"0.785",
"0.1799",
"0.785",
"0.1932",
"0.2197",
"0.1932"
],
[
"0.2197",
"0.1945",
"0.7837",
"0.1945",
"0.7837",
"0.2084",
"0.2197",
"0.2084"
],
[
"0.2214",
"0.2102",
"0.7793",
"0.2102",
"0.7793",
"0.2247",
"0.2214",
"0.2247"
],
[
"0.7754",
"0.2153",
"0.7817",
"0.2153",
"0.7817",
"0.2203",
"0.7754",
"0.2203"
],
[
"0.2197",
"0.226",
"0.786",
"0.226",
"0.786",
"0.2393",
"0.2197",
"0.2393"
],
[
"0.2345",
"0.2405",
"0.785",
"0.2405",
"0.785",
"0.254",
"0.2345",
"0.254"
],
[
"0.2181",
"0.2556",
"0.7817",
"0.2556",
"0.7817",
"0.2695",
"0.2181",
"0.2695"
],
[
"0.2197",
"0.2708",
"0.7837",
"0.2708",
"0.7837",
"0.2842",
"0.2197",
"0.2842"
],
[
"0.2222",
"0.2874",
"0.7817",
"0.2874",
"0.7817",
"0.2974",
"0.2222",
"0.2974"
],
[
"0.2214",
"0.3005",
"0.7837",
"0.3005",
"0.7837",
"0.3145",
"0.2214",
"0.3145"
],
[
"0.2222",
"0.3164",
"0.7837",
"0.3164",
"0.7837",
"0.33",
"0.2222",
"0.33"
],
[
"0.2222",
"0.3333",
"0.7817",
"0.3333",
"0.7817",
"0.3435",
"0.2222",
"0.3435"
],
[
"0.2197",
"0.3467",
"0.785",
"0.3467",
"0.785",
"0.3599",
"0.2197",
"0.3599"
],
[
"0.2173",
"0.3591",
"0.785",
"0.3623",
"0.785",
"0.3801",
"0.2173",
"0.377"
],
[
"0.2214",
"0.379",
"0.7837",
"0.379",
"0.7837",
"0.3926",
"0.2214",
"0.3926"
],
[
"0.2214",
"0.3945",
"0.5337",
"0.3945",
"0.5337",
"0.4077",
"0.2214",
"0.4077"
],
[
"0.2173",
"0.4343",
"0.58",
"0.4343",
"0.58",
"0.4475",
"0.2173",
"0.4475"
],
[
"0.2197",
"0.4666",
"0.785",
"0.4666",
"0.785",
"0.4797",
"0.2197",
"0.4797"
],
[
"0.2214",
"0.483",
"0.7837",
"0.483",
"0.7837",
"0.4937",
"0.2214",
"0.4937"
],
[
"0.2197",
"0.4968",
"0.7837",
"0.4968",
"0.7837",
"0.5103",
"0.2197",
"0.5103"
],
[
"0.2222",
"0.513",
"0.7817",
"0.513",
"0.7817",
"0.5234",
"0.2222",
"0.5234"
],
[
"0.2197",
"0.526",
"0.7837",
"0.5264",
"0.7837",
"0.5405",
"0.2197",
"0.539"
],
[
"0.2214",
"0.5425",
"0.785",
"0.5425",
"0.785",
"0.556",
"0.2214",
"0.556"
],
[
"0.2214",
"0.5596",
"0.7837",
"0.5596",
"0.7837",
"0.5693",
"0.2214",
"0.5693"
],
[
"0.2197",
"0.5728",
"0.785",
"0.5728",
"0.785",
"0.586",
"0.2197",
"0.586"
],
[
"0.2197",
"0.587",
"0.785",
"0.587",
"0.785",
"0.601",
"0.2197",
"0.601"
],
[
"0.2214",
"0.603",
"0.7837",
"0.603",
"0.7837",
"0.616",
"0.2214",
"0.616"
],
[
"0.2197",
"0.617",
"0.4478",
"0.617",
"0.4478",
"0.631",
"0.2197",
"0.631"
],
[
"0.2443",
"0.632",
"0.7837",
"0.632",
"0.7837",
"0.646",
"0.2443",
"0.646"
],
[
"0.2197",
"0.6484",
"0.7837",
"0.6484",
"0.7837",
"0.662",
"0.2197",
"0.662"
],
[
"0.2197",
"0.662",
"0.7817",
"0.662",
"0.7817",
"0.6753",
"0.2197",
"0.6753"
],
[
"0.2197",
"0.677",
"0.7837",
"0.678",
"0.7837",
"0.6914",
"0.2197",
"0.69"
],
[
"0.2197",
"0.6934",
"0.7837",
"0.6934",
"0.7837",
"0.707",
"0.2197",
"0.707"
],
[
"0.2197",
"0.7085",
"0.7837",
"0.7085",
"0.7837",
"0.7217",
"0.2197",
"0.7217"
],
[
"0.2197",
"0.7227",
"0.6543",
"0.7227",
"0.6543",
"0.7363",
"0.2197",
"0.7363"
],
[
"0.2173",
"0.7646",
"0.478",
"0.7646",
"0.478",
"0.779",
"0.2173",
"0.779"
],
[
"0.2197",
"0.797",
"0.7837",
"0.797",
"0.7837",
"0.81",
"0.2197",
"0.81"
],
[
"0.2181",
"0.811",
"0.7817",
"0.811",
"0.7817",
"0.8257",
"0.2181",
"0.8257"
],
[
"0.2181",
"0.827",
"0.7837",
"0.827",
"0.7837",
"0.8403",
"0.2181",
"0.8403"
]
] | [
"A",
"T. Dam, M. Pratama et al.",
"3ProblemFormulation",
"持续学习（CL）问题被定义为顺序到达任务Ti,T2.Tk,kE{1.K}的学",
"习问题，其中K表示在实践中未知的任务数。每个任务都带有三元组T=",
"[ci,Yi，til，其中Nk代表任务大小。EXk表示输人图像，而yEVh",
"",
"y=[l1,l2，.,lml标记类标签，表示任务标识符（ID)。CL问题的目标是构建",
"个连续的学习者f（ge（)），在已经看到的任务上表现良好，其中ge（）是特征",
"提取器，fs（）是分类器。本文主要研究在线（一个时期）任务增量学习和域增量",
"学习问题[29]，其中任何任务ai，9i，t}T的每个三元组仅在一个时期内学",
"习。任务增量学习问题以每个任务的不相交类为特征，即。e。，假设L和L代",
"表k-th任务和k-th任务的标签集，Vk,kLknLk=の。域增量学习问题提出",
"了每个任务P(X,Y)P(X,Y)(k+1)的不同分布或域，同时仍然为每个任务保",
"留相同的自标类。也就是说，将多头配置应用于任务增量学习问题，其中为每个",
"任务fd（）创建一个独立的分类器。域增量学习问题纯粹用单头配置f（）处理。",
"CL问题禁止从头开始重新训练过程k=Ck,C会E(z,9)~D[(f(ge(a)),y)]",
"。学习过程仅由当前任务T的数据样本和包含先前看到的任务的旧样本的微小",
"内存Mk-1支持，以克服灾难性的遗忘问题",
"4Adversarial Continual Learning (ACL)",
"图。1可视化对抗性连续学习方法[10]包括四个部分：共享特征提取器，特定于",
"任务的特征提取器，特定于任务的鉴别器和多头分类器。共享特征提取器生成",
"任务不变的特征，而特定于任务的特征提取器提供每个任务的私有特征。特定",
"于任务的鉴别器预测任务的起源，而多头分类器产生最终的预测。训练过程由",
"三个损失函数控制：分类损失，对抗损失和正交损失。分类损失利用影响多头分",
"类器、共享特征提取器和特定于任务的特征提取器的交叉摘损失函数。在共享",
"特征提取器和特定于任务的鉴别器之间以最小-最大方式执行对抗性损失。在调",
"整特征提取器时实现梯度反转层，从而将最小化问题转换为最大化问题。也就",
"是说，共享特征提取器被训练来欺骗特定于任务的鉴别器，并最终生成任务不",
"变的特征。正交损失确保了共享特征提取器的任务特定特征与任务特定特征提",
"取器的私有特征之间的明确区别",
"在测试阶段排除特定于任务的鉴别器，并通过将私有特征和公共特征的级联",
"特征馈送到产生最终输出的多头分类器来执行推理阶段。由于应用了特定于任",
"务的特征提取器和特定于任务的鉴别器，ACL会带来很高的复杂性。我们在这",
"里提供了一个参数生成器，用于生成缩放和移位参数以执行特征转换。因此，特",
"定于任务的特征是在不损失泛化的情况下以低开销生成的，同时仅依赖于单个",
"鉴别器来玩引起对齐特征的对抗性游戏。此外，ACL依赖于违反在线持续学习",
"要求的迭代训练过程，而SCALE完全在一个时期设置中运行",
"5Learning Policy of SCALE",
"在图2和算法1中可视化了SCALE的学习过程，其中它包括四个块：特征提取",
"器ge（），参数生成器P（），多头分类器fr（）（域增量学习问题中的单头分",
"类器）和单鉴别器De（）。特征提取器提取由对抗性学习机制启用的任务不变特"
] | {
"arxiv_id": "2209.01558",
"language": "zh",
"page_idx": 3
} |
||
2209.01558-zh-4 | 2209.01558-zh-4.jpg | Scalable Adversarial Online Continual Learning
5
<FigureHere>
Fig. 2: SCALE consists of the parameter generator, the feature extractor, the multi-head classifiers (the single-head classifier in the domain-incremental learning problem) and the discriminator. The parameter generator generates the scaling and shifting coefficients converting the common features into the task-specific features. The task-specific features and the task-invariant features are combined and feed the classifier. The training process is controlled by the classification loss, the DER++ loss and the adversarial loss. The training process of the parameter generator is carried out in the meta-learning fashion minimizing the three loss functions. The single discriminator is updated by playing an adversarial game using the cross entropy loss and the DER++ loss
征，其中鉴别器预测任务 ID。与 ACL 不同，ACL 是由特定于任务的特征提取器生成特定于任务的特征，而 SCALE 则受益于具有参数生成器生成的缩放参数$\Phi_1$和移位参数$\Phi_2$的特征转换策略。缩放和移位参数将通用特征修改为特定于任务的特征。分类器接收聚合的特征，从而提供最终的预测。由于缩放和移位参数确保了这些共同特征的特定于任务的特征，因此消除了正交损失。SCALE 仅使用一个鉴别器替换 ACL 中特定于任务的鉴别器
## 5.1 Feature Transformation
SCALE 不会部署任何违反固定架构约束[22]的特定于任务的参数，而是由参数生成器生成缩放和移位参数，从而显著降低其复杂性。我们采用[21,22]的类似思想，其中缩放和移位参数通过特征转换过程创建特定于任务的特征，如下所示：
(1)
$$\tilde{g}_\theta(x)=\frac{\Phi_1}{||\Phi_1||_2}\odot g_\theta(x)+\frac{\Phi_2}{||\Phi_2||_2}$$
其中$\odot$表示逐元素乘法。$\Phi_1,\Phi_2$ 是参数生成器 $P_\varphi(t)=\{\Phi_1,\Phi_2\}$ 生成的缩放和移位参数，将任务 ID 作为具有嵌人层的输入特征，以生成低维特征。这意味着参数生成器网络$\varphi$产生缩放和移位参数$\Phi_1,\Phi_2$ 。实现了残差连接以线性组合共享和私有特征：
(2)
$$g_\theta(x)=\tilde{g}_\theta(x)+g_\theta(x)$$
我们遵循与[22]相同的结构，其中每层实现特征转换策略，每层一个参数生成器。它在多层感知器网络的情况下针对除分类器之外的所有中间层实现，而它仅应用于卷积神经网络的最后一个残差层，因此仅利用单个参数生成器网络。通常在将组合特征馈送到分类器$f_{\phi_k}(s(g_\theta(.)))$之前应用非线性激活函数$s(.)$ | [
[
"0.4429",
"0.11426",
"0.7295",
"0.11615",
"0.7295",
"0.1307",
"0.4429",
"0.1288"
],
[
"0.774",
"0.11743",
"0.786",
"0.11743",
"0.786",
"0.1276",
"0.774",
"0.1276"
],
[
"0.6357",
"0.154",
"0.642",
"0.159",
"0.6357",
"0.1642",
"0.6284",
"0.159"
],
[
"0.3987",
"0.1604",
"0.4192",
"0.1604",
"0.4192",
"0.1674",
"0.3987",
"0.1674"
],
[
"0.3938",
"0.1692",
"0.4272",
"0.1692",
"0.4272",
"0.1768",
"0.3938",
"0.1768"
],
[
"0.4656",
"0.1692",
"0.4756",
"0.1692",
"0.4756",
"0.1724",
"0.4656",
"0.1724"
],
[
"0.5034",
"0.1692",
"0.517",
"0.1692",
"0.517",
"0.1724",
"0.5034",
"0.1724"
],
[
"0.5664",
"0.1666",
"0.5923",
"0.1666",
"0.5923",
"0.1749",
"0.5664",
"0.1749"
],
[
"0.6357",
"0.2122",
"0.6436",
"0.2153",
"0.6396",
"0.2229",
"0.63",
"0.2197"
],
[
"0.3896",
"0.2279",
"0.4307",
"0.2279",
"0.4307",
"0.2349",
"0.3896",
"0.2349"
],
[
"0.5605",
"0.2247",
"0.6006",
"0.2247",
"0.6006",
"0.2361",
"0.5605",
"0.2361"
],
[
"0.2181",
"0.2632",
"0.786",
"0.2632",
"0.786",
"0.2766",
"0.2181",
"0.2766"
],
[
"0.2181",
"0.2778",
"0.786",
"0.279",
"0.786",
"0.2935",
"0.2181",
"0.2922"
],
[
"0.2181",
"0.2942",
"0.7837",
"0.2942",
"0.7837",
"0.308",
"0.2181",
"0.308"
],
[
"0.2181",
"0.308",
"0.7837",
"0.308",
"0.7837",
"0.3213",
"0.2181",
"0.3213"
],
[
"0.2197",
"0.3257",
"0.7817",
"0.3257",
"0.7817",
"0.336",
"0.2197",
"0.336"
],
[
"0.2181",
"0.3403",
"0.786",
"0.3403",
"0.786",
"0.3542",
"0.2181",
"0.3542"
],
[
"0.2173",
"0.3542",
"0.786",
"0.3547",
"0.786",
"0.3694",
"0.2173",
"0.3687"
],
[
"0.2181",
"0.3694",
"0.7837",
"0.3694",
"0.7837",
"0.3833",
"0.2181",
"0.3833"
],
[
"0.2181",
"0.385",
"0.726",
"0.385",
"0.726",
"0.3984",
"0.2181",
"0.3984"
],
[
"0.2181",
"0.432",
"0.785",
"0.433",
"0.785",
"0.4463",
"0.2181",
"0.4458"
],
[
"0.2197",
"0.4475",
"0.785",
"0.4475",
"0.785",
"0.461",
"0.2197",
"0.461"
],
[
"0.2173",
"0.4622",
"0.785",
"0.4634",
"0.785",
"0.4766",
"0.2173",
"0.476"
],
[
"0.2197",
"0.4778",
"0.7837",
"0.4778",
"0.7837",
"0.4912",
"0.2197",
"0.4912"
],
[
"0.2197",
"0.4937",
"0.785",
"0.4937",
"0.785",
"0.507",
"0.2197",
"0.507"
],
[
"0.2197",
"0.509",
"0.5703",
"0.509",
"0.5703",
"0.523",
"0.2197",
"0.523"
],
[
"0.2197",
"0.5537",
"0.2632",
"0.5537",
"0.2632",
"0.565",
"0.2197",
"0.565"
],
[
"0.2566",
"0.5537",
"0.4265",
"0.5537",
"0.4265",
"0.5645",
"0.2566",
"0.5645"
],
[
"0.2197",
"0.587",
"0.7837",
"0.587",
"0.7837",
"0.5977",
"0.2197",
"0.5977"
],
[
"0.2197",
"0.601",
"0.7837",
"0.601",
"0.7837",
"0.6143",
"0.2197",
"0.6143"
],
[
"0.2214",
"0.617",
"0.7817",
"0.617",
"0.7817",
"0.6274",
"0.2214",
"0.6274"
],
[
"0.2181",
"0.632",
"0.2427",
"0.632",
"0.2427",
"0.646",
"0.2181",
"0.646"
],
[
"0.4575",
"0.648",
"0.478",
"0.648",
"0.478",
"0.66",
"0.4575",
"0.66"
],
[
"0.581",
"0.648",
"0.6045",
"0.648",
"0.6045",
"0.661",
"0.581",
"0.661"
],
[
"0.492",
"0.654",
"0.5703",
"0.656",
"0.5703",
"0.6704",
"0.492",
"0.6685"
],
[
"0.7646",
"0.656",
"0.7876",
"0.656",
"0.7876",
"0.6694",
"0.7646",
"0.6694"
],
[
"0.568",
"0.661",
"0.6177",
"0.664",
"0.6167",
"0.6797",
"0.5664",
"0.677"
],
[
"0.3857",
"0.657",
"0.437",
"0.657",
"0.437",
"0.6685",
"0.3857",
"0.6685"
],
[
"0.4436",
"0.6636",
"0.4927",
"0.6636",
"0.4927",
"0.678",
"0.4436",
"0.678"
],
[
"0.2197",
"0.6895",
"0.785",
"0.6895",
"0.785",
"0.7026",
"0.2197",
"0.7026"
],
[
"0.2173",
"0.704",
"0.7837",
"0.704",
"0.7837",
"0.7173",
"0.2173",
"0.7173"
],
[
"0.2197",
"0.7197",
"0.785",
"0.7197",
"0.785",
"0.733",
"0.2197",
"0.733"
],
[
"0.2181",
"0.735",
"0.339",
"0.735",
"0.339",
"0.7485",
"0.2181",
"0.7485"
],
[
"0.4248",
"0.755",
"0.58",
"0.755",
"0.58",
"0.768",
"0.4248",
"0.768"
],
[
"0.7646",
"0.756",
"0.786",
"0.756",
"0.786",
"0.768",
"0.7646",
"0.768"
],
[
"0.2197",
"0.781",
"0.785",
"0.781",
"0.785",
"0.7944",
"0.2197",
"0.7944"
],
[
"0.2197",
"0.797",
"0.7837",
"0.797",
"0.7837",
"0.81",
"0.2197",
"0.81"
],
[
"0.2197",
"0.811",
"0.785",
"0.811",
"0.785",
"0.8247",
"0.2197",
"0.8247"
],
[
"0.2197",
"0.827",
"0.738",
"0.827",
"0.738",
"0.8403",
"0.2197",
"0.8403"
]
] | [
"Scalable Adversarial Online Continual Learning",
"5",
"V",
"o",
"PO",
"",
"",
"Js",
"L",
"gaC",
"DC3",
"Fig. 2: SCALE consists of the parameter generator, the feature extractor, the multi-head",
"classifiers (the single-head classifier in the domain-incremental learning problem) and",
"the discriminator. The parameter generator generates the scaling and shifting coeffi-",
"cients converting the common features into the task-specific features. The task-specific",
"features and thetask-invariant features are combined and feed the classifier.Thetrain",
"ing process is controlled by the classification loss, the DER++ loss and the adversarial",
"loss. The training process of the parameter generator is carried out in the meta-learning",
"fashion minimizing the three loss functions. The single discriminator is updated by",
"playing an adversarial game using the cross entropy loss and the DER++ loss.",
"征，其中鉴别器预测任务ID。与ACL不同，ACL是由特定于任务的特征提取",
"器生成特定于任务的特征，而SCALE则受益于具有参数生成器生成的缩放参数",
"Φ1和移位参数Φ2的特征转换策略。缩放和移位参数将通用特征修改为特定于",
"任务的特征。分类器接收聚合的特征，从而提供最终的预测。由于缩放和移位参",
"数确保了这些共同特征的特定于任务的特征，因此消除了正交损失。SCALE仅",
"使用一个鉴别器替换ACL中特定于任务的鉴别器",
"5.1",
"FeatureTransformation",
"SCALE不会部署任何违反固定架构约束[22]的特定于任务的参数，而是由参数",
"生成器生成缩放和移位参数，从而显著降低其复杂性。我们采用[21,22]的类似",
"思想，其中缩放和移位参数通过特征转换过程创建特定于任务的特征，如下所",
"示：",
"Φ1",
"Φ2",
"0go(c)+",
"(1)",
"142//2",
"ge(r)=",
"1/Φi||2",
"其中表示逐元素乘法。Φ1,Φ2是参数生成器P(t)=[Φ1,Φ2}生成的缩放和",
"移位参数，将任务ID作为具有嵌人层的输人特征，以生成低维特征。这意味着",
"参数生成器网络产生缩放和移位参数Φ1，Φ2。实现了残差连接以线性组合",
"共享和私有特征：",
"ge(α) = ge(c) + go(r)",
"(2)",
"我们遵循与[22]相同的结构，其中每层实现特征转换策略，每层一个参数生成",
"器。它在多层感知器网络的情况下针对除分类器之外的所有中间层实现，而它仅",
"应用于卷积神经网络的最后一个残差层，因此仅利用单个参数生成器网络。通",
"常在将组合特征馈送到分类器fΦ,（s(ge())）之前应用非线性激活函数s()"
] | {
"arxiv_id": "2209.01558",
"language": "zh",
"page_idx": 4
} |
||
2305.15734-zh-2 | 2305.15734-zh-2.jpg | Title Suppressed Due to Excessive Size
<FigureHere>
Figure 1. Illustration of main argument of the proposed study. The number of concept detectors of different models, namely models trained (a) from scratch ( $f_{\mathrm{scratch~}})$,(b) using KD ( $f_{\mathrm{KD~}}$), and (c) using LS ( $f_{\mathrm{LS~}})$,have been measured for a quantitative comparison of the model interpretability. LS enhances the model performance but reduces the interpretability while KD boosts both. The transfer of class-similarity information from the teacher to student model enhances the model $\operatorname{interpretability}.$
$\bullet$ To the best of our knowledge, this is the frst study
to show that KD enhances the interpretability as
well as the accuracy of models.
, A comparison of the interpretability of the LS model without similarity information and student models with different amounts of similarity information learned clearly showed that classsimilarity information improved the interpretability of the student models.
$\bullet$ Various quantitative and qualitative experimen-
tal results support that KD improves model interpretability across KD methods, notions, and datasets.
## 2. Related work
比较了通过标签平滑($f_\mathrm{LS}$)训练的模型与$f_\mathrm{KD}$之间的可解释性，该模型不包含(受到负面影响)类相似性信息。如图E1 ( $b) 和 ( c) f_\mathrm{LS}$学习了对象以外的其他特征，例如背景，这降低了模型的可解释性。参考前面的例子，对于$f_\mathrm{LS}$ ,不相关类的概率 (e。g。,谷)增加，因为模型减少了到均匀分布 $u$ 的距离， 导致$f_\mathrm{LS}$的地图变得不那么以对象为中心。结果表明KD 提高了模型的性能和可解释性，而 LS 降低了可解释性
除了分析类相似性信息的存在与否对模型可解释性的影响外，我们还分析了该信息的程度的影响。Chandrasegaran et al. ( 2022) 分 析 了 LS ( $f_{\mathrm{LS}}^{\mathrm{teacher}}$) 训练的教师模型对学生模型表现的影响。他们表明，可以通过使用 $f_\mathrm{LS}^\mathrm{teacher}$调整 KD 的温度来校准从老师到学生模型传递的班级相似性信息量。因此，我们分析了$f_{\mathrm{LS}}^{\mathrm{teacher}}$在不同温度下对学生模型可解释性的影响。我们的结果表明：1) 学生模型的可解释性随着学生学习的类相似性信息的增加而提高；2)将 logit 蒸馏添加到特征蒸馏中增加了模型的可解释性，证明了传递类相似性信息的效果
我们通过经验证明了 KD 对模型可解释性的一致影响而与 KD 方法以及如何定义可解释性度量无关。除了分析香草 KD 的影响外，我们还表明，与 $f_\mathrm{scratch}$ 相比， 使用各种 KD 方法训练的模型的可解释性有所提高如$\mathbb{E}3.3$ 节所述。然后，我们生成了一个具有热图地面实况的合成数据集，并测量了 Tjoa &Guan (2020) 为$f_\mathrm{scratch}$和$f_\mathrm{KD}$提出的五波段分数，如 5.1 节所述。我们分别测量了 5.2 和 5.3 节中描述的 DiffROAR 分数 (Shah et al., 2021)和损失梯度 (Tsipras et al. 2018),以根据各种可解释性度量评估模型可解释性的改进
我们研究的主要贡献如下。
## 2.1. Knowledge distillation
在 KD 下，通过从具有高学习能力的大型预培训教师模型接收输出分布来提高学习能力低的学生模型的性能$\boxed{\mathrm{F}}($Hinton et al., 2015)。几项研究提出了各种 KD 方法，包括 Logit 蒸馏 (Hinton et al.,2015)和特征蒸馏 (Romero et al., 2014; Zagoruyko & Komodakis, 2016; Yim et al., 2017; Kim et al., 2018; Xu et al., 2020; Tian et al., 2019) , 以提高学生模型的性能
注意力转移(AT)对教师模型中间层中的 n 通道特征进行平均，而无需任何回归器，从而使学生模型可以学习注意力$\mathbb{E}($Zagoruyko $\&$ Komodakis, 2016)。因子转移(FT)使用自动编码器提供学生模型可以轻松理解的简洁信息$\operatorname{E}($Kim et al., 2018)。自我监督KD(SSKD) 利用自我监督信号传递内在语义，为学生提供信息 (Xu et al., 2020) 。对比表示蒸馏 (CRD) 训练学生最大化两个模型的表示之间的互信息的下界， 就像在对比学习中所做的那样E$( Tian et al. , 2019)$ | [
[
"0.3586",
"0.05682",
"0.6167",
"0.05682",
"0.6167",
"0.0707",
"0.3586",
"0.0707"
],
[
"0.2443",
"0.0903",
"0.255",
"0.0903",
"0.255",
"0.1048",
"0.2443",
"0.1048"
],
[
"0.7017",
"0.0859",
"0.7183",
"0.0859",
"0.7183",
"0.09344",
"0.7017",
"0.09344"
],
[
"0.6045",
"0.09216",
"0.681",
"0.09216",
"0.681",
"0.1042",
"0.6045",
"0.1042"
],
[
"0.7485",
"0.0903",
"0.7524",
"0.0903",
"0.7524",
"0.09534",
"0.7485",
"0.09534"
],
[
"0.513",
"0.09344",
"0.5537",
"0.09344",
"0.5537",
"0.1042",
"0.513",
"0.1042"
],
[
"0.2091",
"0.0966",
"0.2263",
"0.0966",
"0.2263",
"0.10986",
"0.2091",
"0.10986"
],
[
"0.2769",
"0.0947",
"0.298",
"0.0947",
"0.298",
"0.09784",
"0.2769",
"0.09784"
],
[
"0.2484",
"0.1029",
"0.259",
"0.1111",
"0.2484",
"0.1193",
"0.2378",
"0.1111"
],
[
"0.4927",
"0.1016",
"0.5825",
"0.1042",
"0.5825",
"0.1187",
"0.4927",
"0.11615"
],
[
"0.613",
"0.1029",
"0.675",
"0.1048",
"0.675",
"0.1193",
"0.613",
"0.11743"
],
[
"0.3105",
"0.1042",
"0.3162",
"0.1042",
"0.3162",
"0.11237",
"0.3105",
"0.11237"
],
[
"0.2754",
"0.1048",
"0.281",
"0.1048",
"0.281",
"0.11237",
"0.2754",
"0.11237"
],
[
"0.4314",
"0.1061",
"0.4617",
"0.1061",
"0.4617",
"0.11426",
"0.4314",
"0.11426"
],
[
"0.7173",
"0.108",
"0.732",
"0.1048",
"0.7363",
"0.11554",
"0.721",
"0.1187"
],
[
"0.2378",
"0.1427",
"0.267",
"0.1427",
"0.267",
"0.1881",
"0.2378",
"0.1881"
],
[
"0.3447",
"0.1427",
"0.3489",
"0.1427",
"0.3489",
"0.1477",
"0.3447",
"0.1477"
],
[
"0.429",
"0.1433",
"0.4617",
"0.1433",
"0.4617",
"0.1509",
"0.429",
"0.1509"
],
[
"0.3137",
"0.1515",
"0.3179",
"0.1515",
"0.3179",
"0.161",
"0.3137",
"0.161"
],
[
"0.4314",
"0.1515",
"0.4617",
"0.1515",
"0.4617",
"0.1604",
"0.4314",
"0.1604"
],
[
"0.7036",
"0.1477",
"0.7563",
"0.1477",
"0.7563",
"0.159",
"0.7036",
"0.159"
],
[
"0.3708",
"0.1528",
"0.4028",
"0.1528",
"0.4028",
"0.1704",
"0.3708",
"0.1704"
],
[
"0.2769",
"0.1528",
"0.282",
"0.1528",
"0.282",
"0.1604",
"0.2769",
"0.1604"
],
[
"0.2075",
"0.161",
"0.2263",
"0.161",
"0.2263",
"0.1755",
"0.2075",
"0.1755"
],
[
"0.609",
"0.161",
"0.63",
"0.161",
"0.63",
"0.1642",
"0.609",
"0.1642"
],
[
"0.4927",
"0.1674",
"0.581",
"0.1674",
"0.581",
"0.1818",
"0.4927",
"0.1818"
],
[
"0.6167",
"0.1654",
"0.6763",
"0.1674",
"0.675",
"0.1831",
"0.615",
"0.1812"
],
[
"0.7114",
"0.1674",
"0.7417",
"0.1674",
"0.7417",
"0.1863",
"0.7114",
"0.1863"
],
[
"0.411",
"0.1736",
"0.4192",
"0.1736",
"0.4192",
"0.1787",
"0.411",
"0.1787"
],
[
"0.2957",
"0.178",
"0.312",
"0.178",
"0.312",
"0.1932",
"0.2957",
"0.1932"
],
[
"0.3757",
"0.1818",
"0.398",
"0.1818",
"0.398",
"0.1957",
"0.3757",
"0.1957"
],
[
"0.4307",
"0.2122",
"0.4656",
"0.2166",
"0.4631",
"0.2267",
"0.4272",
"0.2229"
],
[
"0.699",
"0.2134",
"0.7573",
"0.2184",
"0.755",
"0.2329",
"0.697",
"0.2279"
],
[
"0.2769",
"0.2153",
"0.3162",
"0.2153",
"0.3162",
"0.2197",
"0.2769",
"0.2197"
],
[
"0.343",
"0.1692",
"0.3906",
"0.1692",
"0.3906",
"0.1799",
"0.343",
"0.1799"
],
[
"0.2532",
"0.2235",
"0.259",
"0.2235",
"0.259",
"0.2291",
"0.2532",
"0.2291"
],
[
"0.3064",
"0.2235",
"0.3179",
"0.2235",
"0.3179",
"0.2343",
"0.3064",
"0.2343"
],
[
"0.4248",
"0.2235",
"0.4714",
"0.2235",
"0.4714",
"0.2343",
"0.4248",
"0.2343"
],
[
"0.601",
"0.2247",
"0.683",
"0.2247",
"0.683",
"0.2393",
"0.601",
"0.2393"
],
[
"0.5312",
"0.2279",
"0.5483",
"0.2279",
"0.5483",
"0.2311",
"0.5312",
"0.2311"
],
[
"0.2754",
"0.226",
"0.281",
"0.226",
"0.281",
"0.2317",
"0.2754",
"0.2317"
],
[
"0.21",
"0.2343",
"0.2239",
"0.2343",
"0.2239",
"0.2424",
"0.21",
"0.2424"
],
[
"0.429",
"0.2317",
"0.4697",
"0.2317",
"0.4697",
"0.2437",
"0.429",
"0.2437"
],
[
"0.4902",
"0.2361",
"0.5786",
"0.238",
"0.5786",
"0.2524",
"0.4902",
"0.2507"
],
[
"0.614",
"0.238",
"0.675",
"0.238",
"0.675",
"0.2524",
"0.614",
"0.2524"
],
[
"0.2737",
"0.2424",
"0.3489",
"0.2424",
"0.3489",
"0.2524",
"0.2737",
"0.2524"
],
[
"0.2418",
"0.2393",
"0.255",
"0.2393",
"0.255",
"0.25",
"0.2418",
"0.25"
],
[
"0.2737",
"0.2551",
"0.3464",
"0.2551",
"0.3464",
"0.2651",
"0.2737",
"0.2651"
],
[
"0.08826",
"0.281",
"0.8857",
"0.281",
"0.8857",
"0.2942",
"0.08826",
"0.2942"
],
[
"0.08905",
"0.2954",
"0.8857",
"0.2954",
"0.8857",
"0.3093",
"0.08905",
"0.3093"
],
[
"0.08826",
"0.3093",
"0.8823",
"0.3093",
"0.8823",
"0.3225",
"0.08826",
"0.3225"
],
[
"0.08826",
"0.324",
"0.8857",
"0.324",
"0.8857",
"0.3372",
"0.08826",
"0.3372"
],
[
"0.08905",
"0.339",
"0.1936",
"0.339",
"0.1936",
"0.3516",
"0.08905",
"0.3516"
],
[
"0.0907",
"0.377",
"0.4722",
"0.377",
"0.4722",
"0.3901",
"0.0907",
"0.3901"
],
[
"0.509",
"0.375",
"0.8823",
"0.3757",
"0.8823",
"0.3901",
"0.509",
"0.3894"
],
[
"0.09235",
"0.3933",
"0.4714",
"0.3933",
"0.4714",
"0.404",
"0.09235",
"0.404"
],
[
"0.5317",
"0.3914",
"0.8857",
"0.3914",
"0.8857",
"0.4048",
"0.5317",
"0.4048"
],
[
"0.08905",
"0.4072",
"0.4722",
"0.4072",
"0.4722",
"0.4216",
"0.08905",
"0.4216"
],
[
"0.5337",
"0.4072",
"0.7563",
"0.4072",
"0.7563",
"0.4204",
"0.5337",
"0.4204"
],
[
"0.0907",
"0.4224",
"0.4739",
"0.4224",
"0.4739",
"0.4363",
"0.0907",
"0.4363"
],
[
"0.513",
"0.433",
"0.5337",
"0.433",
"0.5337",
"0.4414",
"0.513",
"0.4414"
],
[
"0.528",
"0.43",
"0.8857",
"0.43",
"0.8857",
"0.4443",
"0.528",
"0.4443"
],
[
"0.08905",
"0.4375",
"0.4722",
"0.4375",
"0.4722",
"0.4507",
"0.08905",
"0.4507"
],
[
"0.5312",
"0.4458",
"0.886",
"0.4443",
"0.886",
"0.4583",
"0.5312",
"0.459"
],
[
"0.08905",
"0.4526",
"0.4722",
"0.4526",
"0.4722",
"0.4666",
"0.08905",
"0.4666"
],
[
"0.5317",
"0.4602",
"0.8857",
"0.4602",
"0.8857",
"0.4734",
"0.5317",
"0.4734"
],
[
"0.0907",
"0.4673",
"0.4722",
"0.4673",
"0.4722",
"0.4817",
"0.0907",
"0.4817"
],
[
"0.5337",
"0.4766",
"0.884",
"0.4766",
"0.884",
"0.4873",
"0.5337",
"0.4873"
],
[
"0.08905",
"0.483",
"0.4722",
"0.483",
"0.4722",
"0.4968",
"0.08905",
"0.4968"
],
[
"0.5317",
"0.4905",
"0.886",
"0.4905",
"0.886",
"0.504",
"0.5317",
"0.504"
],
[
"0.08905",
"0.4968",
"0.125",
"0.4968",
"0.125",
"0.512",
"0.08905",
"0.512"
],
[
"0.5317",
"0.506",
"0.721",
"0.506",
"0.721",
"0.5195",
"0.5317",
"0.5195"
],
[
"0.09235",
"0.5215",
"0.4714",
"0.5215",
"0.4714",
"0.532",
"0.09235",
"0.532"
],
[
"0.517",
"0.531",
"0.5337",
"0.531",
"0.5337",
"0.5386",
"0.517",
"0.5386"
],
[
"0.5293",
"0.529",
"0.886",
"0.529",
"0.886",
"0.5425",
"0.5293",
"0.5425"
],
[
"0.0907",
"0.535",
"0.4722",
"0.535",
"0.4722",
"0.549",
"0.0907",
"0.549"
],
[
"0.08905",
"0.547",
"0.4739",
"0.547",
"0.4739",
"0.5645",
"0.08905",
"0.5645"
],
[
"0.5317",
"0.5435",
"0.8857",
"0.5435",
"0.8857",
"0.558",
"0.5317",
"0.558"
],
[
"0.5337",
"0.558",
"0.886",
"0.558",
"0.886",
"0.5713",
"0.5337",
"0.5713"
],
[
"0.0907",
"0.565",
"0.4722",
"0.565",
"0.4722",
"0.579",
"0.0907",
"0.579"
],
[
"0.08905",
"0.58",
"0.4722",
"0.579",
"0.4722",
"0.5933",
"0.08905",
"0.5947"
],
[
"0.5337",
"0.5747",
"0.596",
"0.5747",
"0.596",
"0.586",
"0.5337",
"0.586"
],
[
"0.08905",
"0.5933",
"0.4714",
"0.595",
"0.4714",
"0.61",
"0.08905",
"0.608"
],
[
"0.08826",
"0.607",
"0.4763",
"0.605",
"0.4763",
"0.6255",
"0.08826",
"0.6274"
],
[
"0.5",
"0.608",
"0.642",
"0.608",
"0.642",
"0.6226",
"0.5",
"0.6226"
],
[
"0.0907",
"0.6255",
"0.4722",
"0.6255",
"0.4722",
"0.6396",
"0.0907",
"0.6396"
],
[
"0.501",
"0.634",
"0.6978",
"0.634",
"0.6978",
"0.648",
"0.501",
"0.648"
],
[
"0.0907",
"0.6416",
"0.4722",
"0.6416",
"0.4722",
"0.655",
"0.0907",
"0.655"
],
[
"0.0907",
"0.656",
"0.4739",
"0.656",
"0.4739",
"0.6704",
"0.0907",
"0.6704"
],
[
"0.5024",
"0.6577",
"0.8857",
"0.6577",
"0.8857",
"0.672",
"0.5024",
"0.672"
],
[
"0.08905",
"0.672",
"0.2214",
"0.672",
"0.2214",
"0.685",
"0.08905",
"0.685"
],
[
"0.5024",
"0.6724",
"0.8857",
"0.6724",
"0.8857",
"0.686",
"0.5024",
"0.686"
],
[
"0.501",
"0.686",
"0.8857",
"0.6875",
"0.8857",
"0.702",
"0.501",
"0.7007"
],
[
"0.0907",
"0.6934",
"0.4722",
"0.6934",
"0.4722",
"0.707",
"0.0907",
"0.707"
],
[
"0.5024",
"0.7026",
"0.8857",
"0.7026",
"0.8857",
"0.7163",
"0.5024",
"0.7163"
],
[
"0.09235",
"0.709",
"0.4714",
"0.709",
"0.4714",
"0.7227",
"0.09235",
"0.7227"
],
[
"0.08905",
"0.7227",
"0.4739",
"0.7236",
"0.4739",
"0.738",
"0.08905",
"0.7373"
],
[
"0.501",
"0.7183",
"0.8857",
"0.7183",
"0.8857",
"0.732",
"0.501",
"0.732"
],
[
"0.501",
"0.733",
"0.8857",
"0.733",
"0.8857",
"0.747",
"0.501",
"0.747"
],
[
"0.0907",
"0.7393",
"0.4739",
"0.7393",
"0.4739",
"0.7524",
"0.0907",
"0.7524"
],
[
"0.0907",
"0.754",
"0.4739",
"0.754",
"0.4739",
"0.7676",
"0.0907",
"0.7676"
],
[
"0.5",
"0.7485",
"0.8545",
"0.7485",
"0.8545",
"0.762",
"0.5",
"0.762"
],
[
"0.0907",
"0.768",
"0.4722",
"0.7676",
"0.4722",
"0.782",
"0.0907",
"0.7827"
],
[
"0.5024",
"0.771",
"0.884",
"0.771",
"0.884",
"0.785",
"0.5024",
"0.785"
],
[
"0.08905",
"0.784",
"0.4697",
"0.7827",
"0.4697",
"0.7974",
"0.08905",
"0.799"
],
[
"0.5024",
"0.786",
"0.884",
"0.786",
"0.884",
"0.8",
"0.5024",
"0.8"
],
[
"0.0907",
"0.8",
"0.4722",
"0.8",
"0.4722",
"0.813",
"0.0907",
"0.813"
],
[
"0.5024",
"0.802",
"0.884",
"0.802",
"0.884",
"0.815",
"0.5024",
"0.815"
],
[
"0.09235",
"0.8164",
"0.4714",
"0.8164",
"0.4714",
"0.827",
"0.09235",
"0.827"
],
[
"0.5034",
"0.8164",
"0.8857",
"0.8164",
"0.8857",
"0.83",
"0.5034",
"0.83"
],
[
"0.08826",
"0.8276",
"0.4739",
"0.829",
"0.4739",
"0.8433",
"0.08826",
"0.8423"
],
[
"0.501",
"0.8306",
"0.8857",
"0.8306",
"0.8857",
"0.8447",
"0.501",
"0.8447"
],
[
"0.08905",
"0.8433",
"0.125",
"0.8433",
"0.125",
"0.8594",
"0.08905",
"0.8594"
],
[
"0.501",
"0.8447",
"0.884",
"0.845",
"0.884",
"0.86",
"0.501",
"0.8594"
],
[
"0.5024",
"0.8613",
"0.884",
"0.8613",
"0.884",
"0.875",
"0.5024",
"0.875"
],
[
"0.08905",
"0.867",
"0.2769",
"0.867",
"0.2769",
"0.881",
"0.08905",
"0.881"
],
[
"0.5024",
"0.877",
"0.8823",
"0.877",
"0.8823",
"0.89",
"0.5024",
"0.89"
],
[
"0.5034",
"0.893",
"0.859",
"0.893",
"0.859",
"0.904",
"0.5034",
"0.904"
]
] | [
"Title Suppressed Due to Excessive Size",
"e",
"Inter",
"(ofctencet detector",
"",
"(Accuracy)",
"(a)",
"nrounr",
"0",
"69.94 % (—)",
"275(—)",
"",
"S",
"Model",
"P",
"spsusil,",
"",
"Studen",
"",
"Model",
"activatlonmay",
"sialary",
"",
"(b)",
"Oe",
"70.80 % ( 1 )",
"312(↑)",
"H",
"Z.",
"TeMaC",
"",
"Label",
"ntergretab map)",
"groundirun",
"logit output",
"3",
"1",
"Smoothed",
"(oftoeepalletor)",
"e",
"p",
"(c)",
"Model",
"70.13 % ( ↑)",
"219(↓)",
"uniform distribution u",
"",
"PLHL",
"Figure 1. Ilustration of main argument of the proposed study. The number of concept detectors of different models,",
"namely models trained (a) from scratch ( fscrateh ), (b) using KD ( fkD ), and (c) using LS ( fLs ), have been measured for a",
"quantitative comparison of the model interpretability. LS enhances the model performance but reduces the interpretability",
"while KD boosts both. The transfer of class-similarity information from the teacher to student model enhances the model",
"interpretability.",
"比较了通过标签平滑（fLs）训练的模型与fKD之间",
". To the best of our knowledge, this is the first study",
"的可解释性，该模型不包含（受到负面影响）类相似",
"to show that KD enhances the interpretability as",
"性信息。如图E1（b）和（c）所示，fLs学习了对象",
"well as the accuracy of models.",
"以外的其他特征，例如背景，这降低了模型的可解释",
"·",
"A comparison of the interpretability of the LS",
"性。参考前面的例子，对于fLs，不相关类的概率（e。",
"model without similarity information and stu-",
"g。，谷）增加，因为模型减少了到均匀分布u的距离，",
"dent models with different amounts of similar-",
"导致fLs的地图变得不那么以对象为中心。结果表明",
"ityinformationlearned clearlyshowedthat class",
"KD提高了模型的性能和可解释性，而LS降低了可解",
"similarity information improved the interpretabil-",
"释性",
"ity of the student models.",
"除了分析类相似性信息的存在与否对模型可解释性的",
"",
" Various quantitative and qualitative experimen-",
"影响外，我们还分析了该信息的程度的影响。Chan",
"drasegaran et al.（2022)分析了LS（fteacher）训练",
"tal results support that KD improves model in-",
"terpretability across KD methods, notions, and",
"的教师模型对学生模型表现的影响。他们表明，可以",
"通过使用fteacher调整KD的温度来校准从老师到学",
"datasets.",
"生模型传递的班级相似性信息量。因此，我们分析了",
"ftscher在不同温度下对学生模型可解释性的影响。我",
"2. Related work",
"们的结果表明：1）学生模型的可解释性随着学生学习",
"2.1. Knowledge distillation",
"的类相似性信息的增加而提高；2）将logit蒸馏添加",
"到特征蒸馏中增加了模型的可解释性，证明了传递类",
"在KD下，通过从具有高学习能力的大型预培训教师",
"相似性信息的效果",
"模型接收输出分布来提高学习能力低的学生模型的性",
"能E(Hintonetal.，2015）。几项研究提出了各种KD",
"我们通过经验证明了KD对模型可解释性的一致影响",
"方法，包括Logit蒸馏（Hintonetal.，2015）和特征",
"而与KD方法以及如何定义可解释性度量无关。除了",
"分析香草KD的影响外，我们还表明，与fscratch相比，",
"蒸馏(Romero etal.,2014;Zagoruyko&Komodakis,",
"2016; Yim et al., 2017; Kim et al., 2018; Xu et al.,",
"使用各种KD方法训练的模型的可解释性有所提高，",
"如E3.3节所述。然后，我们生成了一个具有热图地面",
"2020；Tianetal.，2019），以提高学生模型的性能",
"实况的合成数据集，并测量了Tjoa&Guan（2020）",
"注意力转移（AT）对教师模型中间层中的n通道特",
"为fscratch和fkD提出的五波段分数，如5.1节所述",
"征进行平均，而无需任何回归器，从而使学生模型可",
"我们分别测量了5.2和5.3节中描述的DiffROAR分",
"以学习注意力E(Zagoruyko&Komodakis,2016）。因",
"数（Shahetal..2021）和损失梯度（Tsiprasetal.",
"子转移（FT）使用自动编码器提供学生模型可以轻",
"2018），以根据各种可解释性度量评估模型可解释性的",
"松理解的简洁信息E(Kimetal.，2018）。自我监督",
"改进",
"KD(SSKD）利用自我监督信号传递内在语义，为学生",
"提供信息（Xuetal.，2020）。对比表示蒸馏（CRD）",
"我们研究的主要贡献如下。",
"训练学生最大化两个模型的表示之间的互信息的下界",
"就像在对比学习中所做的那样E（Tianetal..2019）"
] | {
"arxiv_id": "2305.15734",
"language": "zh",
"page_idx": 2
} |
||
2305.15734-zh-3 | 2305.15734-zh-3.jpg | Title Suppressed Due to Excessive Size
et al., 2021) 分析了 KD 如何增强学生模型的泛化性能。他们已经证实 KD 是 LS 的自适应版本，它对模型产生正则化效应。Yuan et al.(2020) 分析了 KD 和LS 之间的关系，并提出了无教师 KD。但是，他们没有解释通过 KD 提供给学生模型的其他信息。在拟议的研究中，我们比较了 LS 和 KD 训练的模型。结果表明，教师模型传递的班级相似性信息促进了学生模型更有效地捕获概念表示
# 2.2. Label smoothing
LS 使用将具有均匀分布的独热矢量组合为标签的矢量来训练模型$\overline{\mathrm{E}}($Szegedy et al., 2016) 。LS 在训练过程中采用正则化，从而提高了模型的泛化性能。Müller et al. (2019) 演示了 LS 通过可视化图像分类器的倒数第二层表示，使训练集中的每个示例与所有其他类等距。他们证明$f_\mathrm{LS}^\mathrm{teacher}$恶化了学生模型的性能，因为它删除了有关教师日志之间相似性的信息
另一方面，Shen et al. (2021)认为 KD 和 LS 是兼容的。他们证明了$f_\mathrm{LS}^\mathrm{teacher}$通过增加语义相似类的嵌人之间的距离来提高学生模型的性能。Chandrasegaran et al. (2022)引入了系统扩散并分析了这些相互矛盾的发现。他们证明，系统扩散会削弱$f_\mathrm{LS}^\mathrm{teacher}$训练的学生模型的性能，从而使低温下的 KD 有效。他们还表上 决 生 时 上 , 次 间 仅 低 温 围 1日 12日 发 。他 们 近 农 明 , 从 $f_{\mathrm{LS}}^{\mathrm{teacher}}$中 提 取 的 知 识 会 随 着 温 度 的 降 低 而 导致类相似性信息的丢失。
# $2.3.{\mathrm{Explainable~AI}}$
研究人员试图解释深度神经网络(DNN)的推理过程。事后方法通过定位参与的输入像素E(Simonyan et al., 2014;Sundararajan et al., 2017; Zeiler & Fergus, 2014; Ribeiro et al., 2016) 或生成反事实解释E(Goyal et al., 2019; Singla et al., 2019) 来解释训练后的模型。相反， 几位研究人员$\mathbb{E}($Chen et al., 2018; Alvarez-Melis & Jaakkola, 2018) 设计了一种新的可解释体系结构，其中决策本质上是可解释的，无需任何事后解释。大多数方法都通过定性可视化或消除最重要像素后预测的恶化来证明所提出方法的可解释性
模型可解释性的量化相对开发不足。Li et al.(2020) 理论上将可解释性定义为局部线性。Barceló et al. (2020) 建议获得解释所需的计算复杂性表示可解释性；复杂度越低，可解释性越高。然而，这不能凭经验应用，因为考虑了真实的数据分布，这导致了高计算复杂性。相比之下，网络解剖$\mathbb{E}($Bau et al.,2017)提供了一种直观有效的方法来量化 DNN 的可解释性。因此，我们采用网络解剖作为可解释性的衡量标准；细节将在下一节中提供
多项研究 (Yuan et al., 2020; Tang et al., 2020; Zhou 3. On the impact of KD for model
interpretability
本节研究 KD 对模型可解释性的影响。3.1 节描述了定义和量化模型可解释性的过程。3.2 节比较了$f_\mathrm{scratch}$ $,f_{\mathrm{KD}}$和$f_\mathrm{LS}$的可解释性。3.3 节介绍了使用各种 KD 方法训练的模型的比较，以验证增强功能不仅限于香草 KD 所做的
$3.1.{\mathrm{~Interpretability~quantification~via~network}}$
dissection
如上一节所述，受网络解剖的启发，我们使用概念检测器的数量来测量模型的可解释性。首先，我们将描述广泛和密集标记的 (Broden)数据集，并解释对概念检测器进行计数的过程。此数据集包含以下数据集：ADE20k (Zhou et al.,2017)、OpenSurfaces (Bell et al., 2014) 、PASCAL-Context (Mottaghi et al., 2014) $、$PASCAL-Part (Chen et al.,2014)和可描述纹理 (Cimpoi et al., 2014)。Broden 中的示例包括对象、场景、零件、纹理、材质和颜色概念。此数据集中允许使用视觉概念的注释蒙版。样本的所有像素都根据相应的概念进行了注释。因此，通过将神经网络中每个单元的激活图与注释掩码进行比较，可以获得与人类可解释概念对齐的单元 (Bau et al.,2017) 。
如下确定概念检测器。我们测量了$f$ 模型及其冻结重量的可解释性。将 Broden 的一个样本$x$输人到模型中，并获得了第$i$个卷积单元的激活图$A_i(x)$ ,然后获得了该图的分布$a_i$ 。对应于前 0 的阈值$T_i$ 。5% 的激活值由$a_i$计算得到，满足$P(a_i\geq T_i)=0.005$ 。由于激活图$A_i(x)$的分辨率小于注释掩码$M_c$对概念$c$ 的分辨率，因此我们对$A_i(x)$进行了插值，以确保$M_c$ 和$A_i(x)$的分辨率相同。随后，我们对插值的$A_i(x)$ 执行二进制掩码，以便仅出现大于或等于$T_i$的区域。最后，我们计算了掩码激活图和注释掩码$M_c$之间的联合 (IoU) 得分的交集。当 IoU 分数超过预定阈值时(0。05),一个单元$i$被识别为相应概念$c$的概念检测器。我们在附录 C 中包含了一个用于获取概念检测器的伪代码，以便于理解网络解剖
唯一检测器是仅与单个概念对齐的单元。网络解剖 (Bau et al., 2017) 报告唯一检测器的数量，以测量中间表示的解纠缠程度。在这项研究中，我们测量了唯一检测器的数量和概念检测器的总数。我们检查了该单元的整体概念检测能力和解开程度
3.2. Impact of KD on model interpretability
本节研究 KD 对模型可解释性的影响，该影响在上一节中定义。训练$f_\mathrm{scratch}$ 、$f_\mathrm{KD}$和$f_\mathrm{LS}$的实验设置见附录 B.2。图 2显示了$f_\mathrm{scratch}$ 、$f_\mathrm{KD}$和$f_\mathrm{LS}$的可解释性。我们从模型的最后一个卷积层测量了概念检测器的数量。附录 A.2 中提供了模型下层可解释性的比较。当 KD 被实施时，概念检测器，特别是对象检测器的总数显著增加。同时，与$f_\mathrm{scratch}$相比，$f_\mathrm{LS}$中的 | [
[
"0.3586",
"0.05807",
"0.615",
"0.05807",
"0.615",
"0.07135",
"0.3586",
"0.07135"
],
[
"0.09235",
"0.0884",
"0.4795",
"0.0884",
"0.4795",
"0.0985",
"0.09235",
"0.0985"
],
[
"0.482",
"0.09155",
"0.4944",
"0.09155",
"0.4944",
"0.09534",
"0.482",
"0.09534"
],
[
"0.4968",
"0.0859",
"0.796",
"0.0859",
"0.796",
"0.09973",
"0.4968",
"0.09973"
],
[
"0.08905",
"0.1016",
"0.4739",
"0.1016",
"0.4739",
"0.11554",
"0.08905",
"0.11554"
],
[
"0.523",
"0.1042",
"0.649",
"0.1042",
"0.649",
"0.11743",
"0.523",
"0.11743"
],
[
"0.0907",
"0.11743",
"0.4722",
"0.11743",
"0.4722",
"0.1307",
"0.0907",
"0.1307"
],
[
"0.0907",
"0.132",
"0.4722",
"0.132",
"0.4722",
"0.1459",
"0.0907",
"0.1459"
],
[
"0.501",
"0.1288",
"0.884",
"0.1288",
"0.884",
"0.1427",
"0.501",
"0.1427"
],
[
"0.08905",
"0.1477",
"0.4739",
"0.1477",
"0.4739",
"0.161",
"0.08905",
"0.161"
],
[
"0.501",
"0.1427",
"0.8813",
"0.1445",
"0.8813",
"0.159",
"0.501",
"0.1572"
],
[
"0.0907",
"0.1622",
"0.4739",
"0.1622",
"0.4739",
"0.1755",
"0.0907",
"0.1755"
],
[
"0.5073",
"0.1642",
"0.5156",
"0.1642",
"0.5156",
"0.1692",
"0.5073",
"0.1692"
],
[
"0.5146",
"0.159",
"0.884",
"0.159",
"0.884",
"0.1724",
"0.5146",
"0.1724"
],
[
"0.09235",
"0.178",
"0.4722",
"0.178",
"0.4722",
"0.1913",
"0.09235",
"0.1913"
],
[
"0.5024",
"0.1736",
"0.884",
"0.1736",
"0.884",
"0.1875",
"0.5024",
"0.1875"
],
[
"0.0907",
"0.1925",
"0.4739",
"0.1925",
"0.4739",
"0.2058",
"0.0907",
"0.2058"
],
[
"0.5024",
"0.1901",
"0.601",
"0.1901",
"0.601",
"0.202",
"0.5024",
"0.202"
],
[
"0.08905",
"0.207",
"0.2688",
"0.207",
"0.2688",
"0.2203",
"0.08905",
"0.2203"
],
[
"0.501",
"0.2216",
"0.839",
"0.2216",
"0.839",
"0.2349",
"0.501",
"0.2349"
],
[
"0.0866",
"0.238",
"0.246",
"0.2393",
"0.246",
"0.254",
"0.0866",
"0.2524"
],
[
"0.5337",
"0.2374",
"0.601",
"0.2374",
"0.601",
"0.2488",
"0.5337",
"0.2488"
],
[
"0.08905",
"0.262",
"0.4739",
"0.262",
"0.4739",
"0.276",
"0.08905",
"0.276"
],
[
"0.5034",
"0.2615",
"0.8823",
"0.2615",
"0.8823",
"0.2715",
"0.5034",
"0.2715"
],
[
"0.0907",
"0.2778",
"0.4739",
"0.2778",
"0.4739",
"0.2917",
"0.0907",
"0.2917"
],
[
"0.5034",
"0.2766",
"0.8823",
"0.2766",
"0.8823",
"0.2874",
"0.5034",
"0.2874"
],
[
"0.09235",
"0.2935",
"0.4739",
"0.2935",
"0.4739",
"0.307",
"0.09235",
"0.307"
],
[
"0.5024",
"0.2903",
"0.884",
"0.2917",
"0.884",
"0.303",
"0.5024",
"0.3018"
],
[
"0.501",
"0.3037",
"0.884",
"0.305",
"0.884",
"0.3193",
"0.501",
"0.318"
],
[
"0.0907",
"0.31",
"0.4722",
"0.31",
"0.4722",
"0.3208",
"0.0907",
"0.3208"
],
[
"0.0907",
"0.324",
"0.4739",
"0.324",
"0.4739",
"0.3372",
"0.0907",
"0.3372"
],
[
"0.5",
"0.318",
"0.884",
"0.318",
"0.884",
"0.332",
"0.5",
"0.332"
],
[
"0.08905",
"0.336",
"0.4722",
"0.3372",
"0.4722",
"0.3516",
"0.08905",
"0.351"
],
[
"0.501",
"0.3352",
"0.8823",
"0.3352",
"0.8823",
"0.3484",
"0.501",
"0.3484"
],
[
"0.0907",
"0.3528",
"0.3652",
"0.3528",
"0.3652",
"0.3662",
"0.0907",
"0.3662"
],
[
"0.5024",
"0.3499",
"0.884",
"0.3499",
"0.884",
"0.363",
"0.5024",
"0.363"
],
[
"0.5024",
"0.3655",
"0.8857",
"0.3655",
"0.8857",
"0.379",
"0.5024",
"0.379"
],
[
"0.0907",
"0.3757",
"0.4739",
"0.3757",
"0.4739",
"0.3894",
"0.0907",
"0.3894"
],
[
"0.501",
"0.3801",
"0.884",
"0.3801",
"0.884",
"0.3933",
"0.501",
"0.3933"
],
[
"0.0907",
"0.3901",
"0.4722",
"0.3901",
"0.4722",
"0.4048",
"0.0907",
"0.4048"
],
[
"0.5024",
"0.3958",
"0.8857",
"0.3958",
"0.8857",
"0.4092",
"0.5024",
"0.4092"
],
[
"0.08905",
"0.406",
"0.4714",
"0.406",
"0.4714",
"0.4192",
"0.08905",
"0.4192"
],
[
"0.5024",
"0.4124",
"0.8813",
"0.4124",
"0.8813",
"0.4224",
"0.5024",
"0.4224"
],
[
"0.08905",
"0.4216",
"0.4739",
"0.4216",
"0.4739",
"0.435",
"0.08905",
"0.435"
],
[
"0.501",
"0.4248",
"0.884",
"0.4255",
"0.884",
"0.4395",
"0.501",
"0.4382"
],
[
"0.0907",
"0.435",
"0.4739",
"0.435",
"0.4739",
"0.4495",
"0.0907",
"0.4495"
],
[
"0.501",
"0.4395",
"0.863",
"0.4414",
"0.863",
"0.455",
"0.501",
"0.4526"
],
[
"0.0907",
"0.452",
"0.4739",
"0.452",
"0.4739",
"0.4653",
"0.0907",
"0.4653"
],
[
"0.08826",
"0.464",
"0.4756",
"0.4634",
"0.4756",
"0.481",
"0.08826",
"0.4817"
],
[
"0.5034",
"0.4653",
"0.8823",
"0.4653",
"0.8823",
"0.476",
"0.5034",
"0.476"
],
[
"0.08826",
"0.4817",
"0.259",
"0.4817",
"0.259",
"0.4956",
"0.08826",
"0.4956"
],
[
"0.501",
"0.4792",
"0.8823",
"0.4792",
"0.8823",
"0.4924",
"0.501",
"0.4924"
],
[
"0.5034",
"0.4937",
"0.884",
"0.4937",
"0.884",
"0.507",
"0.5034",
"0.507"
],
[
"0.08905",
"0.513",
"0.2345",
"0.513",
"0.2345",
"0.5264",
"0.08905",
"0.5264"
],
[
"0.5024",
"0.509",
"0.8857",
"0.509",
"0.8857",
"0.523",
"0.5024",
"0.523"
],
[
"0.5034",
"0.526",
"0.8823",
"0.526",
"0.8823",
"0.536",
"0.5034",
"0.536"
],
[
"0.0907",
"0.536",
"0.4739",
"0.536",
"0.4739",
"0.55",
"0.0907",
"0.55"
],
[
"0.5",
"0.5386",
"0.884",
"0.5386",
"0.884",
"0.552",
"0.5",
"0.552"
],
[
"0.09235",
"0.5537",
"0.4722",
"0.5537",
"0.4722",
"0.5645",
"0.09235",
"0.5645"
],
[
"0.501",
"0.5537",
"0.8823",
"0.5537",
"0.8823",
"0.5674",
"0.501",
"0.5674"
],
[
"0.0907",
"0.5674",
"0.4739",
"0.5674",
"0.4739",
"0.5806",
"0.0907",
"0.5806"
],
[
"0.501",
"0.5684",
"0.8857",
"0.5693",
"0.8857",
"0.5835",
"0.501",
"0.582"
],
[
"0.08905",
"0.582",
"0.4714",
"0.5835",
"0.4714",
"0.5947",
"0.08905",
"0.5933"
],
[
"0.5024",
"0.5854",
"0.8823",
"0.5854",
"0.8823",
"0.5986",
"0.5024",
"0.5986"
],
[
"0.08905",
"0.5977",
"0.478",
"0.5977",
"0.478",
"0.6113",
"0.08905",
"0.6113"
],
[
"0.5024",
"0.5996",
"0.884",
"0.5996",
"0.884",
"0.613",
"0.5024",
"0.613"
],
[
"0.09235",
"0.6123",
"0.4739",
"0.6123",
"0.4739",
"0.6255",
"0.09235",
"0.6255"
],
[
"0.501",
"0.6143",
"0.884",
"0.6143",
"0.884",
"0.6274",
"0.501",
"0.6274"
],
[
"0.08905",
"0.627",
"0.4722",
"0.627",
"0.4722",
"0.64",
"0.08905",
"0.64"
],
[
"0.501",
"0.63",
"0.884",
"0.63",
"0.884",
"0.643",
"0.501",
"0.643"
],
[
"0.09235",
"0.6426",
"0.4739",
"0.6426",
"0.4739",
"0.656",
"0.09235",
"0.656"
],
[
"0.5024",
"0.6445",
"0.8857",
"0.6445",
"0.8857",
"0.6577",
"0.5024",
"0.6577"
],
[
"0.0907",
"0.6577",
"0.4722",
"0.6577",
"0.4722",
"0.672",
"0.0907",
"0.672"
],
[
"0.5034",
"0.661",
"0.7295",
"0.661",
"0.7295",
"0.672",
"0.5034",
"0.672"
],
[
"0.0907",
"0.6724",
"0.335",
"0.6724",
"0.335",
"0.686",
"0.0907",
"0.686"
],
[
"0.5024",
"0.6816",
"0.8857",
"0.6816",
"0.8857",
"0.696",
"0.5024",
"0.696"
],
[
"0.0907",
"0.6943",
"0.4722",
"0.6943",
"0.4722",
"0.7085",
"0.0907",
"0.7085"
],
[
"0.5024",
"0.6987",
"0.8823",
"0.6987",
"0.8823",
"0.709",
"0.5024",
"0.709"
],
[
"0.0907",
"0.7104",
"0.4722",
"0.7104",
"0.4722",
"0.7236",
"0.0907",
"0.7236"
],
[
"0.5024",
"0.712",
"0.884",
"0.712",
"0.884",
"0.726",
"0.5024",
"0.726"
],
[
"0.0907",
"0.726",
"0.4739",
"0.726",
"0.4739",
"0.7393",
"0.0907",
"0.7393"
],
[
"0.501",
"0.7266",
"0.884",
"0.7266",
"0.884",
"0.7407",
"0.501",
"0.7407"
],
[
"0.09235",
"0.741",
"0.4722",
"0.741",
"0.4722",
"0.755",
"0.09235",
"0.755"
],
[
"0.5024",
"0.744",
"0.8105",
"0.744",
"0.8105",
"0.755",
"0.5024",
"0.755"
],
[
"0.09235",
"0.7583",
"0.4714",
"0.7583",
"0.4714",
"0.768",
"0.09235",
"0.768"
],
[
"0.0907",
"0.771",
"0.4722",
"0.771",
"0.4722",
"0.784",
"0.0907",
"0.784"
],
[
"0.5",
"0.773",
"0.8203",
"0.7734",
"0.8203",
"0.7886",
"0.5",
"0.787"
],
[
"0.0907",
"0.786",
"0.4722",
"0.786",
"0.4722",
"0.8",
"0.0907",
"0.8"
],
[
"0.09235",
"0.803",
"0.4714",
"0.803",
"0.4714",
"0.813",
"0.09235",
"0.813"
],
[
"0.5034",
"0.8",
"0.8813",
"0.8",
"0.8813",
"0.81",
"0.5034",
"0.81"
],
[
"0.5",
"0.81",
"0.884",
"0.811",
"0.884",
"0.829",
"0.5",
"0.8276"
],
[
"0.1013",
"0.8184",
"0.1708",
"0.8184",
"0.1708",
"0.829",
"0.1013",
"0.829"
],
[
"0.501",
"0.827",
"0.8857",
"0.8276",
"0.8857",
"0.8423",
"0.501",
"0.8413"
],
[
"0.5024",
"0.8433",
"0.884",
"0.8433",
"0.884",
"0.857",
"0.5024",
"0.857"
],
[
"0.5034",
"0.86",
"0.8823",
"0.86",
"0.8823",
"0.8706",
"0.5034",
"0.8706"
],
[
"0.5024",
"0.8735",
"0.884",
"0.8735",
"0.884",
"0.887",
"0.5024",
"0.887"
],
[
"0.501",
"0.8896",
"0.884",
"0.8896",
"0.884",
"0.903",
"0.501",
"0.903"
]
] | [
"Title Suppressed Due to Excessive Size",
"多项研究（Yuanetal.,2020:Tangetal..2020:Zhou",
"",
"3. On the impact of KD for model",
"etal.，2021）分析了KD如何增强学生模型的泛化性",
"interpretability",
"能。他们已经证实KD是LS的自适应版本，它对模",
"型产生正则化效应。Yuanetal.（2020）分析了KD和",
"本节研究KD对模型可解释性的影响。3.1节描述了定",
"LS之间的关系，并提出了无教师KD。但是，他们没",
"义和量化模型可解释性的过程。3.2节比较了fseratch",
"有解释通过KD提供给学生模型的其他信息。在拟议",
"：",
"fKD和fLs的可解释性。3.3节介绍了使用各种KD",
"的研究中，我们比较了LS和KD训练的模型。结果",
"方法训练的模型的比较，以验证增强功能不仅限于香",
"表明，教师模型传递的班级相似性信息促进了学生模",
"草KD所做的",
"型更有效地捕获概念表示",
"3.1. Interpretability quantification via network",
"2.2.Label smoothing",
"dissection",
"LS使用将具有均匀分布的独热失量组合为标签的矢量",
"如上一节所述，受网络解部的启发，我们使用概念检",
"来训练模型E(Szegedyetal.，2016）。LS在训练过程",
"测器的数量来测量模型的可解释性。首先，我们将描",
"中采用正则化，从而提高了模型的泛化性能。Miller",
"述广泛和密集标记的（Broden）数据集，并解释对概",
"念检测器进行计数的过程。此数据集包含以下数据",
"etal.（2019）演示了LS通过可视化图像分类器的倒数",
"第二层表示，使训练集中的每个示例与所有其他类等",
"集:ADE20k(Zhou et al.,2017)、OpenSurfaces(Bell",
"距。他们证明fteacher恶化了学生模型的性能，因为它",
"et al.,2014)、PASCAL-Context (Mottaghi et al.",
"删除了有关教师日志之间相似性的信息",
"2014）、PASCAL-Part（Chenet al.，2014）和可描",
"述纹理（Cimpoietal.，2014）。Broden中的示例包括",
"另一方面，Shenetal.（2021）认为KD和LS是兼容",
"对象、场景、零件、纹理、材质和颜色概念。此数据集",
"的。他们证明了fteacher 通过增加语义相似类的嵌人",
"中允许使用视觉概念的注释蒙版。样本的所有像素都",
"之间的距离来提高学生模型的性能。Chandrasegaran",
"根据相应的概念进行了注释。因此，通过将神经网络",
"etal.（2022）引人了系统扩散并分析了这些相互矛盾",
"中每个单元的激活图与注释掩码进行比较，可以获得",
"的发现。他们证明，系统扩散会削弱fteacher训练的学",
"与人类可解释概念对齐的单元（Bauetal.，2017）。",
"生模型的性能，从而使低温下的KD有效。他们还表",
"明，从fteacher中提取的知识会随着温度的降低而导",
"如下确定概念检测器。我们测量了模型及其冻结重",
"致类相似性信息的丢失。",
"量的可解释性。将Broden的一个样本a输人到模型",
"中，并获得了第i个卷积单元的激活图Ai(α），然后",
"2.3. Explainable AI",
"获得了该图的分布ai。对应于前0的阈值T：。5%的",
"激活值由ai计算得到，满足P（a≥>T）=0.005。由",
"研究人员试图解释深度神经网络（DNN）的推理过程。",
"于激活图A(c）的分辨率小于注释掩码M。对概念c",
"事后方法通过定位参与的输入像素El(Simonyanetal.,",
"的分辨率，因此我们对A（r）进行了插值，以确保M。",
"2014; Sundararajan et al., 2017; Zeiler & Fergus, 2014;",
"和A;(α）的分辨率相同。随后，我们对插值的A(a)",
"Ribeiroetal..2016）或生成反事实解释E（Goyaletal.",
"执行二进制掩码，以便仅出现大于或等于T，的区域。",
"2019；Singlaetal.，2019）来解释训练后的模型。相反，",
"最后，我们计算了掩码激活图和注释掩码M。之间的",
"几位研究人员E（Chenetal，2018；Alvarez-Melis&",
"联合（IoU）得分的交集。当IoU分数超过预定阈值时",
"Jaakkola，2018）设计了一种新的可解释体系结构，其",
"(0。05)，一个单元i被识别为相应概念c的概念检测",
"中决策本质上是可解释的，无需任何事后解释。大多",
"器。我们在附录C中包含了一个用于获取概念检测器",
"数方法都通过定性可视化或消除最重要像素后预测的",
"的伪代码，以便于理解网络解部",
"恶化来证明所提出方法的可解释性",
"唯一检测器是仅与单个概念对齐的单元。网络解",
"模型可解释性的量化相对开发不足。Lietal.(2020）理",
"部（Bauetal.2017）报告唯一检测器的数量：以",
"论上将可解释性定义为局部线性。Barceloetal.（2020)",
"测量中间表示的解纠缠程度。在这项研究中，我们测",
"建议获得解释所需的计算复杂性表示可解释性；复杂",
"量了唯一检测器的数量和概念检测器的总数。我们检",
"度越低，可解释性越高。然而，这不能凭经验应用，因",
"查了该单元的整体概念检测能力和解开程度",
"为考虑了真实的数据分布，这导致了高计算复杂性。相",
"比之下，网络解剖E(Bauetal.,2017)提供了一种直",
"3.2.Impact of KD on model interpretability",
"观有效的方法来量化DNN的可解释性。因此，我们",
"采用网络解部作为可解释性的衡量标准；细节将在下",
"本节研究KD对模型可解释性的影响，该影响在上一",
"节中定义。训练fscratch、fkD和fLs的实验设置见",
"节中提供",
"附录B.2。图2显示了fscratch、fkD和fLs的可解",
"释性。我们从模型的最后一个卷积层测量了概念检测",
"器的数量。附录A.2中提供了模型下层可解释性的比",
"较。当KD被实施时，概念检测器，特别是对象检测",
"器的总数显著增加。同时，与fscratch相比，fLs中的"
] | {
"arxiv_id": "2305.15734",
"language": "zh",
"page_idx": 3
} |
||
2109.08613-zh-1 | 2109.08613-zh-1.jpg | 用于文本分类的人口统计信息的对抗性清理
Somnath Basu Roy Chowdhury Sayan Ghosh Yiyuan Li
$\textbf{Junier B. Oliva}$
{ somnath, sayghosh, yiyuanli, joliva } @cs.unc.edu
Shashank Srivastava
Snigdha Chaturvedi
$\{$ ssrivastava, snigdha $\}$ @cs.unc.edu
UNC Chapel Hill
## $\mathbf{Abstract}$
Contextual representations learned by language models can often encode undesirable attributes , like demographic associations of the users, while being trained for an unrelated target task. We aim to scrub such undesirable attributes and learn fair representations while maintaining performance on the target task. In this paper, we present an adversarial learning framework“Ad versarial S crubber"( ADS ), to debias contextual representations. We perform theoretical analysis to show that our framework converges without leaking demographic information under certain conditions. We extend previous evaluation techniques by evaluating debiasing performance using Minimum Description Length (MDL) probing. Experimental evaluations on 8 datasets show that ADS generates representations with minimal information about demographic attributes while being maximally informative about the target task.
要的是要确保决策函数不会基于此类表示进行预测。在这项工作中，我们专注于在训练文本分类系统期间删除数据表示中编码的人口统计属性。为此，我们提出了“AdversarialScrubber"(ADS),以在目标任务(例如仇恨语音检测)的训练过程中从中间表示中删除与受保护属性(例如性别或种族)有关的信息。删除此类功能可确保在决策过程中，基于这些表示形式构建的任何预测模型都将与人口统计信息无关
ADS 可以在训练任何文本分类模型期间用作即插即用模块，以学习公平的中间表示。该框架由 4 个模块组成：编码器，洗涤器，偏差鉴别器和目标分类器。编码器生成输入文本的上下文表示。将这些编码的上下文表示作为输入，Scrubber 尝试为目标任务生成公平的表示。偏差鉴别器和目标分类器分别从洗涤器的输出预测受保护的属性和目标标签。该框架以对抗的方式进行端到端训练 (Goodfellow et al., 2014)。我们提供了理论分析，表明在某些条件下，编码器和洗涤器收敛而不会泄漏有关受保护属性的信息。我们在5 对话数据集，2个基于 Twitter 的数据集和一个具有不同目标任务和受保护属性设置的传记数据集上评估了我们的框架。我们通过测量给定表示形式的标签的最小描述长度(MDL)(Voita and Titov, 2020) 而不是探测精度来扩展先前的评估方法。MDL 为测量去偏置性能提供了更细粒度的评估基准。我们使用现成的分类器$^{1}$计算 MDL,使其更容易重现。在使用 ADS 框架进行训练后， 与目标任务的微调相比，我们观察到受保护属性预测的 MDL 有了显着提高。我们的贡献是：
$\bullet$ We present Ad versarial S crubber( ADS ), an adversarial framework to learn fair representations for text classification.
·We provide theoretical guarantees to show that Scrubber and Encoder converge without leaking demographic information.
$\bullet$ We extend previous evaluation methodology
$^{\mathrm{l}}$We use MLPClassifier modules from scikit-learn
### 1 $\textbf{Introduction}$
自动化系统越来越多地用于现实世界的应用程序，例如过滤大学申请 (Basu et al., 2019) , 确定信用资格(Ghailan et al.,2016),做出招聘决定 (Chalfin et al., 2016) 等。对于此类任务， 预测模型是根据来自人类决策的数据进行训练的，这些数据通常偏向于某些人口统计组(Mehrabi et al., 2019; Blodgett et al., 2020; Shah et al.,2020)。基于人口属性的偏见决策可能会产生持久的经济，社会和文化后果
自然语言文本高度指示作者(Koppel et al., 2002; Burger et al., 2011; Nguyen et al., 2013; Verhoeven and Daelemans, 2014; Weren et al., 2014; Rangel et al., 2016; Verhoeven et al., 2016; Blodgett et al.,2016) 的人口统计属性。语言模型通常可以对这种人口统计关联进行编码，即使没有直接访问它们。先前的工作表明，深度学习模型中的中间表示对被谈论 (Blodgett et al., 2016; Elazar and Goldberg, 2018; Elazar et al., 2021) 的作者或人的人口统计关联进行编码。因此，重 | [
[
"0.272",
"0.09265",
"0.7256",
"0.09265",
"0.7256",
"0.1069",
"0.272",
"0.1069"
],
[
"0.1301",
"0.1247",
"0.4072",
"0.1259",
"0.4072",
"0.1401",
"0.1301",
"0.1395"
],
[
"0.444",
"0.1259",
"0.5684",
"0.1259",
"0.5684",
"0.1401",
"0.444",
"0.1401"
],
[
"0.607",
"0.1271",
"0.6978",
"0.1271",
"0.6978",
"0.1383",
"0.607",
"0.1383"
],
[
"0.733",
"0.1259",
"0.8716",
"0.1259",
"0.8716",
"0.1395",
"0.733",
"0.1395"
],
[
"0.705",
"0.1289",
"0.733",
"0.1289",
"0.733",
"0.1372",
"0.705",
"0.1372"
],
[
"0.419",
"0.13",
"0.4358",
"0.13",
"0.4358",
"0.1372",
"0.419",
"0.1372"
],
[
"0.5615",
"0.1277",
"0.596",
"0.1277",
"0.596",
"0.136",
"0.5615",
"0.136"
],
[
"0.2007",
"0.1437",
"0.817",
"0.1437",
"0.817",
"0.158",
"0.2007",
"0.158"
],
[
"0.2988",
"0.1776",
"0.4836",
"0.1787",
"0.4836",
"0.1936",
"0.2988",
"0.1924"
],
[
"0.5205",
"0.1787",
"0.6987",
"0.1787",
"0.6987",
"0.1936",
"0.5205",
"0.1936"
],
[
"0.288",
"0.1965",
"0.723",
"0.1965",
"0.723",
"0.2102",
"0.288",
"0.2102"
],
[
"0.43",
"0.2131",
"0.571",
"0.2131",
"0.571",
"0.2268",
"0.43",
"0.2268"
],
[
"0.2644",
"0.2542",
"0.3418",
"0.2542",
"0.3418",
"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",
"0.879",
"0.2695",
"0.879",
"0.2832",
"0.511",
"0.2832"
],
[
"0.1453",
"0.279",
"0.4568",
"0.279",
"0.4568",
"0.2922",
"0.1453",
"0.2922"
],
[
"0.5156",
"0.288",
"0.879",
"0.288",
"0.879",
"0.2986",
"0.5156",
"0.2986"
],
[
"0.1453",
"0.2944",
"0.4575",
"0.2935",
"0.4575",
"0.308",
"0.1453",
"0.3088"
],
[
"0.5146",
"0.3027",
"0.881",
"0.3027",
"0.881",
"0.3164",
"0.5146",
"0.3164"
],
[
"0.147",
"0.3088",
"0.4575",
"0.3088",
"0.4575",
"0.3225",
"0.147",
"0.3225"
],
[
"0.1453",
"0.3225",
"0.4575",
"0.3225",
"0.4575",
"0.3362",
"0.1453",
"0.3362"
],
[
"0.511",
"0.3184",
"0.879",
"0.3184",
"0.879",
"0.332",
"0.511",
"0.332"
],
[
"0.1486",
"0.3384",
"0.455",
"0.3384",
"0.455",
"0.3486",
"0.1486",
"0.3486"
],
[
"0.513",
"0.335",
"0.881",
"0.335",
"0.881",
"0.3486",
"0.513",
"0.3486"
],
[
"0.1453",
"0.3508",
"0.4568",
"0.3508",
"0.4568",
"0.364",
"0.1453",
"0.364"
],
[
"0.513",
"0.3516",
"0.881",
"0.3516",
"0.881",
"0.3652",
"0.513",
"0.3652"
],
[
"0.1453",
"0.3652",
"0.4575",
"0.3652",
"0.4575",
"0.379",
"0.1453",
"0.379"
],
[
"0.513",
"0.3682",
"0.882",
"0.3682",
"0.882",
"0.3818",
"0.513",
"0.3818"
],
[
"0.147",
"0.3801",
"0.4568",
"0.3801",
"0.4568",
"0.3938",
"0.147",
"0.3938"
],
[
"0.513",
"0.3843",
"0.881",
"0.3843",
"0.881",
"0.398",
"0.513",
"0.398"
],
[
"0.147",
"0.3938",
"0.4575",
"0.3938",
"0.4575",
"0.4067",
"0.147",
"0.4067"
],
[
"0.513",
"0.3997",
"0.5693",
"0.3997",
"0.5693",
"0.4146",
"0.513",
"0.4146"
],
[
"0.1453",
"0.408",
"0.455",
"0.408",
"0.455",
"0.4216",
"0.1453",
"0.4216"
],
[
"0.1453",
"0.4216",
"0.4575",
"0.4204",
"0.4575",
"0.4353",
"0.1453",
"0.4358"
],
[
"0.531",
"0.4175",
"0.881",
"0.4175",
"0.881",
"0.4312",
"0.531",
"0.4312"
],
[
"0.147",
"0.4358",
"0.4575",
"0.4358",
"0.4575",
"0.4495",
"0.147",
"0.4495"
],
[
"0.5146",
"0.4329",
"0.881",
"0.4329",
"0.881",
"0.4465",
"0.5146",
"0.4465"
],
[
"0.1486",
"0.4531",
"0.455",
"0.4531",
"0.455",
"0.4631",
"0.1486",
"0.4631"
],
[
"0.513",
"0.4495",
"0.881",
"0.4495",
"0.881",
"0.4631",
"0.513",
"0.4631"
],
[
"0.1486",
"0.466",
"0.4568",
"0.466",
"0.4568",
"0.4768",
"0.1486",
"0.4768"
],
[
"0.513",
"0.4673",
"0.8774",
"0.4673",
"0.8774",
"0.478",
"0.513",
"0.478"
],
[
"0.147",
"0.4797",
"0.4568",
"0.4797",
"0.4568",
"0.4934",
"0.147",
"0.4934"
],
[
"0.513",
"0.481",
"0.881",
"0.481",
"0.881",
"0.4954",
"0.513",
"0.4954"
],
[
"0.1453",
"0.4934",
"0.4568",
"0.4934",
"0.4568",
"0.5073",
"0.1453",
"0.5073"
],
[
"0.513",
"0.4963",
"0.882",
"0.4963",
"0.882",
"0.511",
"0.513",
"0.511"
],
[
"0.147",
"0.509",
"0.4534",
"0.509",
"0.4534",
"0.5195",
"0.147",
"0.5195"
],
[
"0.511",
"0.514",
"0.879",
"0.514",
"0.879",
"0.528",
"0.511",
"0.528"
],
[
"0.147",
"0.5215",
"0.4575",
"0.5215",
"0.4575",
"0.535",
"0.147",
"0.535"
],
[
"0.513",
"0.53",
"0.882",
"0.53",
"0.882",
"0.5435",
"0.513",
"0.5435"
],
[
"0.1486",
"0.5376",
"0.4568",
"0.5376",
"0.4568",
"0.5474",
"0.1486",
"0.5474"
],
[
"0.1453",
"0.55",
"0.4575",
"0.549",
"0.4575",
"0.5635",
"0.1453",
"0.564"
],
[
"0.513",
"0.5444",
"0.879",
"0.5444",
"0.879",
"0.558",
"0.513",
"0.558"
],
[
"0.147",
"0.5635",
"0.336",
"0.5635",
"0.336",
"0.5767",
"0.147",
"0.5767"
],
[
"0.511",
"0.56",
"0.879",
"0.56",
"0.879",
"0.5737",
"0.511",
"0.5737"
],
[
"0.5146",
"0.5776",
"0.881",
"0.5776",
"0.881",
"0.5913",
"0.5146",
"0.5913"
],
[
"0.1125",
"0.5913",
"0.1436",
"0.589",
"0.1453",
"0.5996",
"0.11584",
"0.6025"
],
[
"0.1486",
"0.589",
"0.2627",
"0.5903",
"0.2627",
"0.605",
"0.1486",
"0.604"
],
[
"0.5156",
"0.5957",
"0.879",
"0.5957",
"0.879",
"0.6055",
"0.5156",
"0.6055"
],
[
"0.1184",
"0.6143",
"0.4836",
"0.6143",
"0.4836",
"0.6274",
"0.1184",
"0.6274"
],
[
"0.511",
"0.6094",
"0.881",
"0.6094",
"0.881",
"0.6226",
"0.511",
"0.6226"
],
[
"0.11755",
"0.63",
"0.4836",
"0.63",
"0.4836",
"0.6436",
"0.11755",
"0.6436"
],
[
"0.513",
"0.627",
"0.882",
"0.627",
"0.882",
"0.64",
"0.513",
"0.64"
],
[
"0.11755",
"0.6455",
"0.4844",
"0.6455",
"0.4844",
"0.66",
"0.11755",
"0.66"
],
[
"0.513",
"0.6436",
"0.881",
"0.6436",
"0.881",
"0.6567",
"0.513",
"0.6567"
],
[
"0.11755",
"0.662",
"0.4878",
"0.663",
"0.4878",
"0.6777",
"0.11755",
"0.677"
],
[
"0.5146",
"0.659",
"0.881",
"0.659",
"0.881",
"0.673",
"0.5146",
"0.673"
],
[
"0.1184",
"0.6797",
"0.486",
"0.6797",
"0.486",
"0.6934",
"0.1184",
"0.6934"
],
[
"0.5146",
"0.675",
"0.881",
"0.675",
"0.881",
"0.6885",
"0.5146",
"0.6885"
],
[
"0.1184",
"0.697",
"0.4844",
"0.697",
"0.4844",
"0.7104",
"0.1184",
"0.7104"
],
[
"0.513",
"0.6914",
"0.881",
"0.6914",
"0.881",
"0.705",
"0.513",
"0.705"
],
[
"0.511",
"0.706",
"0.879",
"0.706",
"0.879",
"0.7197",
"0.511",
"0.7197"
],
[
"0.1184",
"0.712",
"0.486",
"0.712",
"0.486",
"0.7256",
"0.1184",
"0.7256"
],
[
"0.511",
"0.7217",
"0.881",
"0.7227",
"0.881",
"0.737",
"0.511",
"0.7363"
],
[
"0.1184",
"0.728",
"0.4844",
"0.728",
"0.4844",
"0.741",
"0.1184",
"0.741"
],
[
"0.11755",
"0.742",
"0.3938",
"0.7437",
"0.3938",
"0.7583",
"0.11755",
"0.7573"
],
[
"0.5146",
"0.7393",
"0.882",
"0.7393",
"0.882",
"0.753",
"0.5146",
"0.753"
],
[
"0.511",
"0.7544",
"0.882",
"0.755",
"0.882",
"0.7695",
"0.511",
"0.7686"
],
[
"0.1377",
"0.7603",
"0.4844",
"0.7603",
"0.4844",
"0.774",
"0.1377",
"0.774"
],
[
"0.11755",
"0.777",
"0.4878",
"0.777",
"0.4878",
"0.7905",
"0.11755",
"0.7905"
],
[
"0.534",
"0.7725",
"0.882",
"0.7725",
"0.882",
"0.786",
"0.534",
"0.786"
],
[
"0.546",
"0.786",
"0.8833",
"0.7876",
"0.8833",
"0.8047",
"0.546",
"0.804"
],
[
"0.11755",
"0.792",
"0.486",
"0.792",
"0.486",
"0.8057",
"0.11755",
"0.8057"
],
[
"0.11755",
"0.808",
"0.4878",
"0.808",
"0.4878",
"0.822",
"0.11755",
"0.822"
],
[
"0.5483",
"0.8047",
"0.7886",
"0.8047",
"0.7886",
"0.8184",
"0.5483",
"0.8184"
],
[
"0.11755",
"0.8247",
"0.4844",
"0.8237",
"0.4844",
"0.8384",
"0.11755",
"0.839"
],
[
"0.534",
"0.8247",
"0.881",
"0.8247",
"0.881",
"0.839",
"0.534",
"0.839"
],
[
"0.1184",
"0.8413",
"0.4844",
"0.8413",
"0.4844",
"0.855",
"0.1184",
"0.855"
],
[
"0.5474",
"0.8403",
"0.881",
"0.8413",
"0.881",
"0.8555",
"0.5474",
"0.855"
],
[
"0.1184",
"0.857",
"0.486",
"0.857",
"0.486",
"0.8706",
"0.1184",
"0.8706"
],
[
"0.5474",
"0.8555",
"0.8003",
"0.857",
"0.8003",
"0.8716",
"0.5474",
"0.8706"
],
[
"0.11755",
"0.8726",
"0.4836",
"0.8726",
"0.4836",
"0.8857",
"0.11755",
"0.8857"
],
[
"0.531",
"0.876",
"0.879",
"0.8774",
"0.879",
"0.8926",
"0.531",
"0.89"
],
[
"0.1184",
"0.8896",
"0.4844",
"0.8896",
"0.4844",
"0.903",
"0.1184",
"0.903"
],
[
"0.1184",
"0.904",
"0.486",
"0.905",
"0.486",
"0.9194",
"0.1184",
"0.9185"
],
[
"0.534",
"0.904",
"0.8423",
"0.905",
"0.8423",
"0.9194",
"0.534",
"0.9185"
]
] | [
"用于文本分类的人口统计信息的对抗性清理",
"Somnath Basu Roy Chowdhury",
"Sayan Ghosh",
"Yiyuan Li",
" Junier B. Oliva",
"",
"",
"",
"( somnath, sayghosh, yiyuanli, joliva } @cs.unc.edu",
"Shashank Srivastava",
"Snigdha Chaturvedi",
"( ssrivastava, snigdha } @cs.unc.edu",
"UNC Chapel Hill",
"Abstract",
"要的是要确保决策函数不会基于此类表示进",
"行预测。在这项工作中，我们专注于在训练文",
"Contextual representations learned by lan",
"本分类系统期间删除数据表示中编码的人口统",
"guage models can often encode undesirable",
"计属性。为此，我们提出了“AdversarialScrub-",
"attributes , like demographic associations of",
"the users, while being trained for an unrelated",
"ber”（ADS），以在目标任务（例如仇恨语音",
"targettask.Weaimtoscrub suchundesir",
"检测）的训练过程中从中间表示中删除与受保",
"able attributes and learn fair representations",
"护属性（例如性别或种族）有关的信息。。删",
"while maintaining performance on the target",
"除此类功能可确保在决策过程中，基于这些表",
"task. In this paper, we present an adversar",
"示形式构建的任何预测模型都将与人口统计信",
"ial learning framework \" Ad versarial S crub-",
"息无关",
"ber\" ( ADS ), to debias contextual represen",
"tations.We perform theoretical analysis to",
"ADS可以在训练任何文本分类模型期间用",
"show that our framework converges without",
"作即插即用模块，以学习公平的中间表示。该",
"leakingdemographicinformationunder cer",
"框架由4个模块组成：编码器，洗涤器，偏差",
"tainconditions.Weextend previous evalua",
"鉴别器和目标分类器。编码器生成输人文本",
"tion techniques by evaluating debiasing per",
"的上下文表示。将这些编码的上下文表示作为",
"formance using Minimum Description Length",
"输入，Scrubber尝试为目标任务生成公平的表",
"(MDL)probing.Experimental evaluationsor",
"示。偏差鉴别器和目标分类器分别从洗涤器的",
"8 datasets show that ADS generates represen",
"输出预测受保护的属性和目标标签。该框架以",
"tationswithminimalinformationaboutdemo",
"graphic attributes while being maximally infor.",
"对抗的方式进行端到端训练（Goodfellowetal.",
"mative about the target task.",
"2014）。我们提供了理论分析，表明在某些条",
"件下，编码器和洗涤器收敛而不会泄漏有关受",
"1",
"Introduction",
"保护属性的信息。我们在5对话数据集，2个",
"自动化系统越来越多地用于现实世界的应用",
"基于Twitter的数据集和一个具有不同目标任",
"程序，例如过滤大学申请(Basuetal.,2019",
"务和受保护属性设置的传记数据集上评估了我",
"确定信用资格（Ghailanetal.，2016），做出招聘",
"们的框架。我们通过测量给定表示形式的标签",
"决定（Chalfinetal.,2016）等。对于此类任务，",
"的最小描述长度（MDL）（VoitaandTitov,2020）",
"预测模型是根据来自人类决策的数据进行训",
"而不是探测精度来扩展先前的评估方法。MDL",
"练的，这些数据通常偏向于某些人口统计组",
"为测量去偏置性能提供了更细粒度的评估基",
"准。我们使用现成的分类器1计算MDL，使",
"(Mehrabi et al., 2019; Blodgett et al., 2020; Shah",
"其更容易重现。在使用ADS框架进行训练后",
"etal.,2020)。基于人口属性的偏见决策可能会",
"产生持久的经济，社会和文化后果",
"与目标任务的微调相比，我们观察到受保护属",
"性预测的MDL有了显着提高。我们的贡献是：",
"自然语言文本高度指示作者(Koppeletal.,",
"2002; Burger et al., 2011; Nguyen et al., 2013; Ver-",
": We present Ad versarial S crubber ( ADS ),",
"an adversarial framework to learn fair repre-",
"hoeven and Daelemans, 2014; Weren et al., 2014;",
"Rangel et al., 2016; Verhoeven et al., 2016; Blod-",
"sentations for text classification.",
"gettetal.,2016)的人口统计属性。语言模型通",
": We provide theoretical guarantees to show",
"常可以对这种人口统计关联进行编码，即使没",
"that Scrubber and Encoder converge without",
"有直接访问它们。先前的工作表明，深度学习模",
"leaking demographic information.",
"型中的中间表示对被谈论（Blodgettetal.,2016",
". We extend previous evaluation methodology",
"Elazar and Goldberg,2018; Elazar et al.,2021)的",
"作者或人的人口统计关联进行编码。因此，重",
"'We use MLPClassifier modules from scikit-learn"
] | {
"arxiv_id": "2109.08613",
"language": "zh",
"page_idx": 1
} |
||
2109.08613-zh-2 | 2109.08613-zh-2.jpg | <FigureHere>
Figure 1: Architecture of the Adversarial Scrubber( ADS ).Encoder receives an input $x$ to produce $e.$ Scrubber uses $e$ to produce $u.$ Bias discriminator $d$ and Target classifier $c$ infer protected attribute $z$ and target task label $y$ from $u.$
for adversarial debiasing by framing perfor-
mance in terms of MDL.
$\bullet$ Experimental evaluations on 8 datasets show that models trained using ADS generate representations where probing networks achieve near random performance on protected attribute inference while performing similar to the baselines on target task.
·We show that ADS is scalable and can be
used to remove multiple protected attributes
simultaneously.
## 2 $\textbf{Related Work}$
在目标任务的训练期间学习的上下文表示可以指示与任务无关的特征。如在无监督单词嵌入(Bolukbasi et al., 2016) 和句子嵌人 (May et al.,2019) 中观察到的那样，此类表示通常可以编码不需要的人口统计属性。。先前的工作已经分析了不同 NLP 系统中的偏见，例如机器翻译 (Park et al., 2018; Stanovsky et al., 2019; Font and Costa-Jussa, 2019; Saunders and Byrne, 2020), NLI (Rudinger et al.,2017),文本分类(Dixon et al., 2018; Kiritchenko and Mohammad, 2018; Sap et al., 2019; Liu et al., 2021), 语言生成 (Sheng et al., 2019) 等
Zemel et al. (2013) 引人了消除敏感属性以进行公平分类的优化问题。从那时起，对抗性训练 (Goodfellow et al., 2014) 框架已被探索用于保护 NLP 任务 (Zhang et al., 2018; Li et al., 2018; Elazar and Goldberg, 2018; Liu et al., 2020) 的敏感属性。我们的工作最类似于 Elazar and Goldberg (2018), 它通过学习忽略受保护属性的中间表示形式来盲目地实现公平性。我们在实验中比较了 ADS 和 Elazar and Goldberg (2018)的性能
## 3 Adversarial Scrubber
ADS 将文本文档$\{x_1,x_2,\ldots,x_n\}$作为来自具有相应目标标签$\{y_1,y_2,\ldots,y_n\}$的数据集$\mathcal{D}$ $\{1,2,...K\}$相关联。我们的目标是构建一个模型$f(x)$,使其在进行预测$y_i=f(x_i)$时不依赖于$z_i$ 。该框架由 4 个模块组成：(i)具有权重$\theta_h$的编码器$h(\cdot),($ii)具有权重$\theta_s$的洗涤器$s(\cdot)$, (iii) 具有权重$\theta_d$的偏置鉴别器$d(\cdot)$ 和 (iv) 目标分类器$c(\cdot)$,权重 XMATHX\_。编码器接收文本输入$x_i$ ,并生成嵌人 $e_i=h(x_i)$ ,并将其转发到洗涤器。洗涤器的目标是生成表示$u_i= s( h( x_i) )$ ,以便目标分类器$c$可以很容易地从$u_i$推断出$y_i$ ,但是$u_i$不具有通过偏置鉴别器$d$预测受保护属性$z_i$所需的信息。我
们的设置还包括一个探测网络$q$ ,它有助于评
估学习表示的公平性
在本节的其余部分中，我们将描述假设单个偏置鉴别器的 ADS 。然而，ADS 可以很容易地扩展到合并多个鉴别器，用于删除几个受保护的属性(在 2?节中讨论)
洗涤器：洗涤器从编码器接收输入表示$h(x_i)$ 并生成表示$u_i=s(h(x_i))$ 。洗涤器的目标是产生表示，以使偏差鉴别器难以预测受保护的曧茻o$\~{p}iylošs 此在鵝櫎䶮虖两个鳊蚂轺释橄筅涤器$ $\overline{\text{参数被联合}}$优化以增加预测概率分布 $H(d(u_i))$ 的熵。
$\underline{\delta\mathrm{~-loss}}:$如果鉴别器将高概率分配给正确的受保护属性类，则$\delta$ -loss 函数会惩罚模型。对于每个输入实例，我们形成一个输出掩码$m_i\in \mathbb{R} ^{1\times K}$ ,其中$K$是受保护属性类的数量。
有 相 应 日 你 你 佥 $\{ y_1, y_2, \ldots , y_n\}$ 的 父 据 果 $D$ 的 输 入 。每 个 输 人 $x_i$ 也 与 受 保 护 的 属 性 $z_i\in$ $m_i^{( k) }= 1$ 如 果 $z_i= k$ 和 0 否 则 。编 码 器 和 洗 涤 的 输 入 。每 个 输 入 $x_i$ 也 与 受 保 护 的 属 性 $z_i\in$ 盟 县 十 四 审 地 减 小 $\delta$ 的 坦 生 会 义 为 .
器最大限度地减少$\delta$的损失定义为：
$$\delta(d(u_i))=m_i^T\text{softmax}_{gumble}(d(u_i))$$
(3)
其中 softmax$_gumble(\cdot)$是 gumblesoftmax 函数(Jang et al.,2017)。在我们的实验中，我们使用熵和$\delta$损失的组合。
目标分类器：目标分类器通过优化交叉熵损失
从$u_i$预测目标标签$y_i:\mathcal{L}_c(c(u_i),y_i)$
同时更新洗涤器，目标分类器和编码器参数 | [
[
"0.1528",
"0.0849",
"0.4878",
"0.0861",
"0.4878",
"0.10095",
"0.1528",
"0.0998"
],
[
"0.5557",
"0.0849",
"0.712",
"0.0849",
"0.712",
"0.0986",
"0.5557",
"0.0986"
],
[
"0.7314",
"0.0861",
"0.8345",
"0.0861",
"0.8345",
"0.0974",
"0.7314",
"0.0974"
],
[
"0.614",
"0.0986",
"0.656",
"0.0986",
"0.656",
"0.11224",
"0.614",
"0.11224"
],
[
"0.76",
"0.0986",
"0.7983",
"0.0986",
"0.7983",
"0.111",
"0.76",
"0.111"
],
[
"0.1368",
"0.1217",
"0.4844",
"0.1217",
"0.4844",
"0.1354",
"0.1368",
"0.1354"
],
[
"0.1528",
"0.10156",
"0.3376",
"0.10156",
"0.3376",
"0.1152",
"0.1528",
"0.1152"
],
[
"0.5938",
"0.1289",
"0.6733",
"0.1289",
"0.6733",
"0.1401",
"0.5938",
"0.1401"
],
[
"0.7305",
"0.1289",
"0.825",
"0.1289",
"0.825",
"0.1401",
"0.7305",
"0.1401"
],
[
"0.1511",
"0.136",
"0.4878",
"0.1383",
"0.4878",
"0.1526",
"0.1511",
"0.1509"
],
[
"0.617",
"0.1401",
"0.6514",
"0.1401",
"0.6514",
"0.152",
"0.617",
"0.152"
],
[
"0.763",
"0.1401",
"0.7944",
"0.1401",
"0.7944",
"0.152",
"0.763",
"0.152"
],
[
"0.1528",
"0.155",
"0.4878",
"0.155",
"0.4878",
"0.1687",
"0.1528",
"0.1687"
],
[
"0.1528",
"0.1716",
"0.4895",
"0.1716",
"0.4895",
"0.1853",
"0.1528",
"0.1853"
],
[
"0.615",
"0.1675",
"0.817",
"0.1704",
"0.817",
"0.1859",
"0.615",
"0.1829"
],
[
"0.1528",
"0.187",
"0.4878",
"0.187",
"0.4878",
"0.2007",
"0.1528",
"0.2007"
],
[
"0.1528",
"0.2031",
"0.357",
"0.2031",
"0.357",
"0.2162",
"0.1528",
"0.2162"
],
[
"0.677",
"0.2037",
"0.733",
"0.2037",
"0.733",
"0.2156",
"0.677",
"0.2156"
],
[
"0.69",
"0.2156",
"0.721",
"0.2156",
"0.721",
"0.228",
"0.69",
"0.228"
],
[
"0.1368",
"0.2227",
"0.486",
"0.2227",
"0.486",
"0.2363",
"0.1368",
"0.2363"
],
[
"0.1528",
"0.2393",
"0.486",
"0.2393",
"0.486",
"0.253",
"0.1528",
"0.253"
],
[
"0.1511",
"0.2542",
"0.2703",
"0.2546",
"0.2703",
"0.2695",
"0.1511",
"0.2683"
],
[
"0.11755",
"0.282",
"0.1511",
"0.282",
"0.1511",
"0.2944",
"0.11755",
"0.2944"
],
[
"0.1503",
"0.282",
"0.2747",
"0.282",
"0.2747",
"0.2957",
"0.1503",
"0.2957"
],
[
"0.6826",
"0.279",
"0.733",
"0.279",
"0.733",
"0.2903",
"0.6826",
"0.2903"
],
[
"0.1503",
"0.2861",
"0.1545",
"0.2861",
"0.1545",
"0.2915",
"0.1503",
"0.2915"
],
[
"0.6885",
"0.2903",
"0.7256",
"0.2903",
"0.7256",
"0.3018",
"0.6885",
"0.3018"
],
[
"0.1184",
"0.306",
"0.4836",
"0.306",
"0.4836",
"0.3196",
"0.1184",
"0.3196"
],
[
"0.1184",
"0.3225",
"0.486",
"0.3225",
"0.486",
"0.3362",
"0.1184",
"0.3362"
],
[
"0.7983",
"0.3196",
"0.818",
"0.326",
"0.812",
"0.3342",
"0.7925",
"0.3289"
],
[
"0.1184",
"0.3384",
"0.4844",
"0.3384",
"0.4844",
"0.3516",
"0.1184",
"0.3516"
],
[
"0.513",
"0.3445",
"0.881",
"0.3445",
"0.881",
"0.3582",
"0.513",
"0.3582"
],
[
"0.11755",
"0.3552",
"0.4844",
"0.3552",
"0.4844",
"0.3682",
"0.11755",
"0.3682"
],
[
"0.511",
"0.3582",
"0.867",
"0.3594",
"0.867",
"0.3723",
"0.511",
"0.3718"
],
[
"0.1184",
"0.3706",
"0.4844",
"0.3706",
"0.4844",
"0.3843",
"0.1184",
"0.3843"
],
[
"0.511",
"0.3735",
"0.882",
"0.3735",
"0.882",
"0.3872",
"0.511",
"0.3872"
],
[
"0.11755",
"0.3855",
"0.4844",
"0.386",
"0.4844",
"0.401",
"0.11755",
"0.3997"
],
[
"0.511",
"0.3884",
"0.882",
"0.3884",
"0.882",
"0.402",
"0.511",
"0.402"
],
[
"0.1184",
"0.4026",
"0.4878",
"0.4026",
"0.4878",
"0.4163",
"0.1184",
"0.4163"
],
[
"0.511",
"0.402",
"0.693",
"0.402",
"0.693",
"0.415",
"0.511",
"0.415"
],
[
"0.11584",
"0.4175",
"0.4836",
"0.4187",
"0.4836",
"0.4329",
"0.11584",
"0.4316"
],
[
"0.11755",
"0.4353",
"0.4844",
"0.4353",
"0.4844",
"0.449",
"0.11755",
"0.449"
],
[
"0.1184",
"0.4507",
"0.4844",
"0.4507",
"0.4844",
"0.4644",
"0.1184",
"0.4644"
],
[
"0.513",
"0.4465",
"0.881",
"0.4478",
"0.881",
"0.462",
"0.513",
"0.4614"
],
[
"0.11755",
"0.466",
"0.4844",
"0.4656",
"0.4844",
"0.4797",
"0.11755",
"0.481"
],
[
"0.513",
"0.4644",
"0.679",
"0.4644",
"0.679",
"0.478",
"0.513",
"0.478"
],
[
"0.1184",
"0.4827",
"0.309",
"0.4827",
"0.309",
"0.4963",
"0.1184",
"0.4963"
],
[
"0.532",
"0.4827",
"0.879",
"0.4827",
"0.879",
"0.4963",
"0.532",
"0.4963"
],
[
"0.1335",
"0.4976",
"0.486",
"0.4988",
"0.486",
"0.513",
"0.1335",
"0.5127"
],
[
"0.511",
"0.4988",
"0.8774",
"0.4988",
"0.8774",
"0.5127",
"0.511",
"0.5127"
],
[
"0.1184",
"0.5156",
"0.4878",
"0.5156",
"0.4878",
"0.5293",
"0.1184",
"0.5293"
],
[
"0.511",
"0.5156",
"0.882",
"0.5156",
"0.882",
"0.5293",
"0.511",
"0.5293"
],
[
"0.1184",
"0.531",
"0.4844",
"0.531",
"0.4844",
"0.5444",
"0.1184",
"0.5444"
],
[
"0.511",
"0.532",
"0.738",
"0.532",
"0.738",
"0.546",
"0.511",
"0.546"
],
[
"0.12006",
"0.5474",
"0.4895",
"0.5474",
"0.4895",
"0.5625",
"0.12006",
"0.5625"
],
[
"0.511",
"0.5503",
"0.882",
"0.5503",
"0.882",
"0.564",
"0.511",
"0.564"
],
[
"0.1184",
"0.5635",
"0.4878",
"0.5635",
"0.4878",
"0.5776",
"0.1184",
"0.5776"
],
[
"0.513",
"0.5664",
"0.879",
"0.5664",
"0.879",
"0.58",
"0.513",
"0.58"
],
[
"0.11755",
"0.579",
"0.486",
"0.5776",
"0.486",
"0.593",
"0.11755",
"0.5933"
],
[
"0.513",
"0.583",
"0.879",
"0.583",
"0.879",
"0.5967",
"0.513",
"0.5967"
],
[
"0.11584",
"0.594",
"0.4844",
"0.594",
"0.4844",
"0.608",
"0.11584",
"0.608"
],
[
"0.511",
"0.597",
"0.881",
"0.5986",
"0.881",
"0.616",
"0.511",
"0.615"
],
[
"0.1184",
"0.611",
"0.486",
"0.611",
"0.486",
"0.626",
"0.1184",
"0.626"
],
[
"0.513",
"0.616",
"0.882",
"0.6177",
"0.882",
"0.632",
"0.513",
"0.6313"
],
[
"0.11755",
"0.627",
"0.486",
"0.6274",
"0.486",
"0.6426",
"0.11755",
"0.641"
],
[
"0.511",
"0.632",
"0.559",
"0.6343",
"0.5576",
"0.6484",
"0.51",
"0.6465"
],
[
"0.11584",
"0.6426",
"0.2284",
"0.6436",
"0.2284",
"0.6577",
"0.11584",
"0.6567"
],
[
"0.5156",
"0.6562",
"0.881",
"0.6562",
"0.881",
"0.6694",
"0.5156",
"0.6694"
],
[
"0.1503",
"0.6704",
"0.34",
"0.6704",
"0.34",
"0.684",
"0.1503",
"0.684"
],
[
"0.513",
"0.6704",
"0.881",
"0.6704",
"0.881",
"0.684",
"0.513",
"0.684"
],
[
"0.513",
"0.6885",
"0.882",
"0.6885",
"0.882",
"0.7017",
"0.513",
"0.7017"
],
[
"0.11755",
"0.6714",
"0.1511",
"0.6714",
"0.1511",
"0.683",
"0.11755",
"0.683"
],
[
"0.1184",
"0.6953",
"0.4844",
"0.6953",
"0.4844",
"0.7104",
"0.1184",
"0.7104"
],
[
"0.51",
"0.7007",
"0.8833",
"0.7007",
"0.8833",
"0.7188",
"0.51",
"0.7188"
],
[
"0.11755",
"0.7114",
"0.486",
"0.7104",
"0.486",
"0.7246",
"0.11755",
"0.7256"
],
[
"0.5073",
"0.7134",
"0.8867",
"0.7144",
"0.8867",
"0.7393",
"0.5073",
"0.7383"
],
[
"0.1184",
"0.728",
"0.486",
"0.728",
"0.486",
"0.741",
"0.1184",
"0.741"
],
[
"0.513",
"0.737",
"0.796",
"0.737",
"0.796",
"0.752",
"0.513",
"0.752"
],
[
"0.1184",
"0.7437",
"0.486",
"0.7437",
"0.486",
"0.7573",
"0.1184",
"0.7573"
],
[
"0.1184",
"0.7603",
"0.4844",
"0.7603",
"0.4844",
"0.774",
"0.1184",
"0.774"
],
[
"0.1184",
"0.7754",
"0.4844",
"0.7754",
"0.4844",
"0.7905",
"0.1184",
"0.7905"
],
[
"0.5557",
"0.7725",
"0.843",
"0.774",
"0.843",
"0.7915",
"0.5557",
"0.7905"
],
[
"0.8345",
"0.777",
"0.881",
"0.777",
"0.881",
"0.788",
"0.8345",
"0.788"
],
[
"0.1184",
"0.792",
"0.4844",
"0.792",
"0.4844",
"0.8057",
"0.1184",
"0.8057"
],
[
"0.1184",
"0.808",
"0.4878",
"0.808",
"0.4878",
"0.822",
"0.1184",
"0.822"
],
[
"0.531",
"0.8154",
"0.879",
"0.8154",
"0.879",
"0.83",
"0.531",
"0.83"
],
[
"0.1184",
"0.8247",
"0.486",
"0.8247",
"0.486",
"0.8384",
"0.1184",
"0.8384"
],
[
"0.511",
"0.83",
"0.881",
"0.8306",
"0.881",
"0.8457",
"0.511",
"0.844"
],
[
"0.1184",
"0.8403",
"0.4878",
"0.8403",
"0.4878",
"0.855",
"0.1184",
"0.855"
],
[
"0.5146",
"0.847",
"0.705",
"0.847",
"0.705",
"0.861",
"0.5146",
"0.861"
],
[
"0.1351",
"0.857",
"0.486",
"0.857",
"0.486",
"0.8706",
"0.1351",
"0.8706"
],
[
"0.1184",
"0.8726",
"0.4844",
"0.8726",
"0.4844",
"0.8857",
"0.1184",
"0.8857"
],
[
"0.513",
"0.869",
"0.879",
"0.869",
"0.879",
"0.883",
"0.513",
"0.883"
],
[
"0.1184",
"0.8896",
"0.486",
"0.8896",
"0.486",
"0.903",
"0.1184",
"0.903"
],
[
"0.513",
"0.8857",
"0.804",
"0.8857",
"0.804",
"0.8994",
"0.513",
"0.8994"
],
[
"0.1184",
"0.905",
"0.486",
"0.905",
"0.486",
"0.9185",
"0.1184",
"0.9185"
],
[
"0.531",
"0.905",
"0.881",
"0.905",
"0.881",
"0.9185",
"0.531",
"0.9185"
]
] | [
"for adversarial debiasing by framing perfor-",
"Protected attribute (z)",
"Target label (y)",
"",
"",
": Experimental evaluations on 8 datasets show",
"mance in terms of MDL.",
"Discriminator",
"Target Classifier",
"that models trained using ADS generate rep-",
"d()",
"c()",
"resentations where probing networks achieve",
"near random performance on protected at-",
"Bu",
"tribute inference while performing similar to",
"the baselines on target task.",
"Scrubber",
"s()",
". We show that ADS is scalable and can be",
"used to remove multiple protected attributes",
"simultaneously.",
"2",
"Related Work",
"Encoder",
"",
"(h(-)",
"在目标任务的训练期间学习的上下文表示可",
"以指示与任务无关的特征。如在无监督单词",
"X",
"嵌入（Bolukbasietal.，2016)和句子嵌人（May",
"Figure 1: Architecture of the Adversarial Scrubber (",
"etal.,2019）中观察到的那样，此类表示通常可",
"ADs ).Encoder receives an input to produce e",
"以编码不需要的人口统计属性。。先前的工作",
"Scrubber uses e to produce u . Bias discriminator d",
"已经分析了不同NLP系统中的偏见，例如机",
"and Target classifier c infer protected attribute z and",
"器翻译(Park et al.,2018;Stanovsky et al.,2019;",
"target task label y from u .",
"Font and Costa-Jussa, 2019; Saunders and Byrne",
"2020），NLI(Rudingeretal.2017），文本分类",
"(Dixon et al., 2018; Kiritchenko and Mohammad,",
"们的设置还包括一个探测网络，它有助于评",
"2018;Sapetal.,2019;Liuetal.,2021)，语言生",
"估学习表示的公平性",
"成(Shengetal.,2019)等",
"在本节的其余部分中，我们将描述假设单个",
"Zemeletal.(2013)引人了消除敏感属性以进",
"偏置鉴别器的ADS。然而，ADS可以很容易",
"行公平分类的优化问题。从那时起，对抗性",
"地扩展到合并多个鉴别器，用于删除几个受保",
"训练（Goodfellowetal.,2014）框架已被探索用",
"护的属性（在？？节中讨论）",
"于保护NLP任务(Zhanget al.,2018;Liet al.,",
"洗涤器：洗涤器从编码器接收输人表示h(ci）",
"2018; Elazar and Goldberg, 2018; Liu et al., 2020)",
"并生成表示ui=s(h(c））。洗涤器的目标是",
"的敏感属性。我们的工作最类似于Elazarand",
"产生表示，以使偏差鉴别器难以预测受保护的",
"Goldberg（2018），它通过学习忽略受保护属性",
"topydo为北在美两先涤器",
"的中间表示形式来盲目地实现公平性。我们",
"参数被联合优化以增加预测概率分布H(d(ui))",
"在实验中比较了ADS和ElazarandGoldberg",
"的炳。",
"(2018)的性能",
"8-loss：如果鉴别器将高概率分配给正确的",
"Adversarial Scrubber",
"受保护属性类，则-loss函数会惩罚模型。",
"对于每个输入实例，我们形成一个输出掩码",
"3",
"ADS将文本文档【α1,a2，..，an】作为来自具",
"miR1xK，其中K是受保护属性类的数量。",
"有相应目标标签{91,y2,....9n的数据集D",
"m(%)=1如果2=k和0否则。编码器和洗涤",
"的输入。每个输入ai也与受保护的属性ziE",
"器最大限度地减少的损失定义为：",
"[1,2...K}相关联。我们的目标是构建一个模",
"型f(α），使其在进行预测yi=f(ci）时不依",
"赖于zi。该框架由4个模块组成：（i）具有权",
"S(d(ui) = mT softmaxgumble(d(ui)) ",
"(3)",
"重Oh的编码器h(），（ii）具有权重θ。的洗涤",
"器s(），（ii）具有权重a的偏置鉴别器d()",
"其中softmaxgumble（）是gumblesoftmax函",
"和（iv）目标分类器c(），权重XMATHX_。编",
"数（Jangetal.2017）。在我们的实验中，我们",
"码器接收文本输人ai，并生成嵌入ei=h(ai)",
"使用摘和损失的组合。",
"并将其转发到洗涤器。洗涤器的目标是生成",
"表示ui=s(h(ci)），以便目标分类器c可以很",
"目标分类器：目标分类器通过优化交叉损失",
"容易地从ui推断出yi，但是wi不具有通过偏",
"从ui预测目标标签yi：Lc(c(ui),yi)",
"置鉴别器d预测受保护属性z所需的信息。我",
"同时更新洗涤器，目标分类器和编码器参数"
] | {
"arxiv_id": "2109.08613",
"language": "zh",
"page_idx": 2
} |
||
2103.04038-zh-0 | 2103.04038-zh-0.jpg | $\underline{\text{Published at ICLR 2021 Workshop on Security and Safety in Machine Learning Systems}}$
# 针对语义分割模型的隐藏后门攻击
$\mathop{\mathrm{Yiming~Li}^{1,*},\mathrm{~Yanjie~Li}^{1,*},\mathrm{~Yalei~Lv}^{1,2},\mathrm{~Yong~Jiang}^{1,2},\mathrm{~Shu-Tao~Xia}^{1,2}}$
$^{1}$Tsinghua Shenzhen International Graduate School, Tsinghua University
$^{2}$PCL Research Center of Networks and Communications, Peng Cheng Laboratory
$\{\begin{array}{cc}\text{li-ym18, lyj20, lyl20 }\}&\text{@mails.tsinghua.edu.cn;}\{\begin{array}{c}\text{jiangy, xiast }\end{array}\}\end{array}$®sz.tsinghua.edu.cn
# Abstract
Deep neural networks (DNNs) are vulnerable to the backdoor attack , which $\dot{\text{intends to embed hidden backdoors in DNNs by poisoning training data.}}$ The attacked model behaves normally on benign samples, whereas its prediction will be changed to a particular target label if hidden backdoors are activated. So far, backdoor research has mostly been conducted towards classification tasks. In this paper, we reveal that this threat could also happen in semantic segmentation, which may further endanger many mission-critical applications $(e.g.$, autonomous driving). Except for extending the existing attack paradigm to maliciously manipulate the segmentation models from the image-level, we propose a novel attack paradigm, the fine-grained attack , where we treat the target label $(i.e.$,annotation) from the object-level instead of the image-level to achieve more sophisticated manipulation. In the annotation of poisoned samples generated by the fine-grained attack, only pixels of specific objects will be labeled with the attacker-specified target class while others are still with their groundtruth ones. Experiments show that the proposed methods can successfully attack semantic segmentation models by poisoning only a small proportion of training data. Our method not only provides a new perspective for designing novel attacks but also serves as a strong baseline for improving the robustness of semantic segmentation methods.
### 1 Introduction
语义分割是一个重要的研究领域，已在许多关键任务应用中被广泛且成功地采用，例如自动驾驶 (Siam et al., 2018; Zhang et al., 2020; Feng et al., 2020) 和增强现实 (Zhang et al., 2019; Huang et al., 2019; Han et al., 2020)。因此，它的安全性具有重要意义，值得进一步考虑
最近，大多数先进的语义分割方法都基于深度神经网络(DNNs)(Bao et al.,2018;Huang et al., 2019; Choe et al., 2020) , 其训练需要大量的训练样本和计算量。为了满足这些要求， 通常在培训过程中使用第三方资源。例如，用户可能会采用第三方训练样本(来自互联网或公司),第三方服务器$( e. g.$ , GoogleCloud),甚至直接采用第三方模型。然而，第三方资源的使用不仅带来了便利，而且在培训过程中引人了不透明度，这可能带来新的安全威胁在本文中，我们将重点放在后门攻击上，这是对 DNNs (Li et al., 2020a) 训练的一种新兴但致命的威胁。具体来说，后门攻击者打算通过毒害一小部分训练样本来向 DNN 注人隐藏的后门。到目前为止，后门攻击主要是针对分类任务进行的，并且采用与样本无关的目标标签方式 (Gu et al., 2019; Yao et al., 2019; Liu et al., 2020; Li et al., 2020b; Nguyen & Tran, 2021; Zhai et al., 2021)。换句话说，攻击者为所有中毒样本分配相同的目标标签。我们发现，这种范式在攻击语义分割方面仍然有效。但是，这种方法只能从图像级别操纵预测， 因此无法实现对语义分割的更精细的恶意操纵。为了解决这个问题，在本文中，我们提出了一种细粒度的攻击范式，其中目标标签是特定于样本的。具体来说，我们从对象级别而不是图像级别处理目标标签，其中只有特定对象的像素将被标记为攻击者指定的目标类，而其他对象仍被标记为地面实况。实验验证了我们的方法可以通过仅毒化一小部分训练数据来成功和隐秘地攻击语义分割模型
$^{*}$indicates equal contribution.
1 | [
[
"0.1765",
"0.0341",
"0.8096",
"0.0341",
"0.8096",
"0.04736",
"0.1765",
"0.04736"
],
[
"0.1765",
"0.1029",
"0.5933",
"0.1042",
"0.5933",
"0.125",
"0.1765",
"0.1237"
],
[
"0.1854",
"0.1515",
"0.6846",
"0.1515",
"0.6846",
"0.1666",
"0.1854",
"0.1666"
],
[
"0.183",
"0.1654",
"0.7036",
"0.1674",
"0.7036",
"0.1818",
"0.183",
"0.1799"
],
[
"0.183",
"0.1787",
"0.774",
"0.1812",
"0.774",
"0.1957",
"0.183",
"0.1932"
],
[
"0.1847",
"0.1945",
"0.796",
"0.1957",
"0.796",
"0.2102",
"0.1847",
"0.209"
],
[
"0.4631",
"0.2343",
"0.5376",
"0.2343",
"0.5376",
"0.2488",
"0.4631",
"0.2488"
],
[
"0.2345",
"0.2664",
"0.7656",
"0.2664",
"0.7656",
"0.2798",
"0.2345",
"0.2798"
],
[
"0.2336",
"0.281",
"0.7646",
"0.281",
"0.7646",
"0.2942",
"0.2336",
"0.2942"
],
[
"0.2336",
"0.2942",
"0.763",
"0.2942",
"0.763",
"0.308",
"0.2336",
"0.308"
],
[
"0.2345",
"0.3093",
"0.7656",
"0.3093",
"0.7656",
"0.3225",
"0.2345",
"0.3225"
],
[
"0.232",
"0.3213",
"0.761",
"0.3213",
"0.761",
"0.3352",
"0.232",
"0.3352"
],
[
"0.2345",
"0.336",
"0.7656",
"0.336",
"0.7656",
"0.351",
"0.2345",
"0.351"
],
[
"0.2336",
"0.3499",
"0.7646",
"0.351",
"0.7646",
"0.3655",
"0.2336",
"0.3643"
],
[
"0.2336",
"0.3643",
"0.763",
"0.3643",
"0.763",
"0.379",
"0.2336",
"0.379"
],
[
"0.2336",
"0.379",
"0.761",
"0.379",
"0.761",
"0.3926",
"0.2336",
"0.3926"
],
[
"0.2345",
"0.3914",
"0.7607",
"0.3914",
"0.7607",
"0.4048",
"0.2345",
"0.4048"
],
[
"0.2345",
"0.406",
"0.7646",
"0.406",
"0.7646",
"0.4192",
"0.2345",
"0.4192"
],
[
"0.232",
"0.4185",
"0.7607",
"0.4192",
"0.7607",
"0.433",
"0.232",
"0.432"
],
[
"0.2345",
"0.433",
"0.763",
"0.433",
"0.763",
"0.4463",
"0.2345",
"0.4463"
],
[
"0.2336",
"0.4463",
"0.7646",
"0.4463",
"0.7646",
"0.4602",
"0.2336",
"0.4602"
],
[
"0.2336",
"0.461",
"0.763",
"0.461",
"0.763",
"0.4749",
"0.2336",
"0.4749"
],
[
"0.2336",
"0.4749",
"0.7646",
"0.4749",
"0.7646",
"0.4893",
"0.2336",
"0.4893"
],
[
"0.232",
"0.4873",
"0.7656",
"0.488",
"0.7656",
"0.5024",
"0.232",
"0.502"
],
[
"0.2336",
"0.5024",
"0.763",
"0.5024",
"0.763",
"0.5166",
"0.2336",
"0.5166"
],
[
"0.2336",
"0.5176",
"0.7656",
"0.5176",
"0.7656",
"0.531",
"0.2336",
"0.531"
],
[
"0.232",
"0.529",
"0.568",
"0.53",
"0.568",
"0.5444",
"0.232",
"0.5435"
],
[
"0.1748",
"0.5747",
"0.2091",
"0.5747",
"0.2091",
"0.587",
"0.1748",
"0.587"
],
[
"0.2009",
"0.5737",
"0.3162",
"0.5757",
"0.3162",
"0.5903",
"0.2009",
"0.5884"
],
[
"0.1765",
"0.607",
"0.822",
"0.607",
"0.822",
"0.6206",
"0.1765",
"0.6206"
],
[
"0.1732",
"0.619",
"0.822",
"0.6206",
"0.822",
"0.6353",
"0.1732",
"0.634"
],
[
"0.1748",
"0.6353",
"0.8228",
"0.6353",
"0.8228",
"0.6484",
"0.1748",
"0.6484"
],
[
"0.1765",
"0.6494",
"0.21",
"0.6494",
"0.21",
"0.662",
"0.1765",
"0.662"
],
[
"0.1748",
"0.6685",
"0.824",
"0.6694",
"0.824",
"0.6846",
"0.1748",
"0.683"
],
[
"0.1732",
"0.6816",
"0.822",
"0.683",
"0.822",
"0.6978",
"0.1732",
"0.6963"
],
[
"0.1765",
"0.6978",
"0.8228",
"0.6978",
"0.8228",
"0.711",
"0.1765",
"0.711"
],
[
"0.1765",
"0.711",
"0.8228",
"0.711",
"0.8228",
"0.7246",
"0.1765",
"0.7246"
],
[
"0.1772",
"0.7266",
"0.8105",
"0.7266",
"0.8105",
"0.7373",
"0.1772",
"0.7373"
],
[
"0.1765",
"0.747",
"0.8228",
"0.747",
"0.8228",
"0.7603",
"0.1765",
"0.7603"
],
[
"0.1765",
"0.7603",
"0.8228",
"0.7603",
"0.8228",
"0.7734",
"0.1765",
"0.7734"
],
[
"0.1765",
"0.7744",
"0.8228",
"0.7744",
"0.8228",
"0.7886",
"0.1765",
"0.7886"
],
[
"0.1765",
"0.7886",
"0.822",
"0.7886",
"0.822",
"0.802",
"0.1765",
"0.802"
],
[
"0.1765",
"0.802",
"0.8228",
"0.802",
"0.8228",
"0.815",
"0.1765",
"0.815"
],
[
"0.1765",
"0.8164",
"0.8228",
"0.8164",
"0.8228",
"0.83",
"0.1765",
"0.83"
],
[
"0.1765",
"0.83",
"0.822",
"0.83",
"0.822",
"0.8433",
"0.1765",
"0.8433"
],
[
"0.1814",
"0.8433",
"0.8228",
"0.8433",
"0.8228",
"0.857",
"0.1814",
"0.857"
],
[
"0.1748",
"0.856",
"0.8228",
"0.857",
"0.8228",
"0.872",
"0.1748",
"0.8706"
],
[
"0.1748",
"0.869",
"0.8228",
"0.8706",
"0.8228",
"0.8853",
"0.1748",
"0.884"
],
[
"0.1748",
"0.8853",
"0.4207",
"0.8853",
"0.4207",
"0.8984",
"0.1748",
"0.8984"
],
[
"0.1953",
"0.912",
"0.3906",
"0.912",
"0.3906",
"0.9253",
"0.1953",
"0.9253"
],
[
"0.4927",
"0.9497",
"0.5063",
"0.9497",
"0.5063",
"0.9604",
"0.4927",
"0.9604"
]
] | [
"Published at ICLR 2021 Workshop on Security and Safety in Machine Learning Systems",
"针对语义分割模型的隐藏后门攻击",
"Yiming Lil,*, Yanjie Lil,*, Yalei Lvl,2, Yong Jiang',2, Shu-Tao Xial2",
"1Tsinghua Shenzhen International Graduate School, Tsinghua University",
"2PCL Research Center of Networks and Communications, Peng Cheng Laboratory",
"[li-ym18, lyj20, lyl20 @mails.tsinghua.edu.cn; { jiangy, xiast @sz.tsinghua.edu.cn",
"Abstract",
"Deep neural networks (DNNs) are vulnerable to the backdoor attack , which",
"intends to embed hidden backdoors in DNNs by poisoning training data.",
"The attacked model behaves normally on benign samples, whereas its pre-",
"diction will be changed to a particular target label if hidden backdoors",
"are activated. So far, backdoor research has mostly been conducted to",
"wards classification tasks. In this paper, we reveal that this threat could",
"also happen in semantic segmentation, which may further endanger many",
"mission-critical applications ( e.g. ; autonomous driving). Except for ex-",
"tending the existing attack paradigm to maliciously manipulate the segmen.",
"tation models from the image-level, we propose a novel attack paradigm",
"the fine-grained attack , where we treat the target label (i.e. , annotation)",
"from the object-level instead of the image-level to achieve more sophisti",
"cated manipulation. In the annotation of poisoned samples generated by",
"the fine-grained attack, only pixels of specific objects will be labeled with",
"the attacker-specified target class while others are still with their ground-",
"truth ones. Experiments show that the proposed methods can successfully",
"attack semantic segmentation models by poisoning only a small proportion",
"of training data. Our method not only provides a new perspective for de-",
"signing novel attacks but also serves as a strong baseline for improving the",
"robustness of semantic segmentation methods.",
"1",
"Introduction",
"语义分割是一个重要的研究领域，已在许多关键任务应用中被广泛且成功地采用，例如自",
"动驾驶(Siam etal.,2018;Zhang et al.,2020;Feng etal.，2020）和增强现实（Zhang etal.",
"2019；Huangetal.，2019；Hanetal.，2020）。因此，它的安全性具有重要意义，值得进一步",
"考虑",
"最近，大多数先进的语义分割方法都基于深度神经网络（DNNs）（Baoetal.，2018;Huang",
"etal，2019；Choeetal.，2020），其训练需要大量的训练样本和计算量。为了满足这些要求",
"通常在培训过程中使用第三方资源。例如，用户可能会采用第三方训练样本（来自互联网或",
"公司），第三方服务器（e.g.，GoogleCloud），甚至直接采用第三方模型。然而，第三方资",
"源的使用不仅带来了便利，而且在培训过程中引入了不透明度，这可能带来新的安全威胁",
"在本文中，我们将重点放在后门攻击上，这是对DNNs（Lietal.，2020a）训练的一种新兴但",
"致命的威胁。具体来说，后门攻击者打算通过毒害一小部分训练样本来向DNN注入隐藏的",
"后门。到目前为止，后门攻击主要是针对分类任务进行的，并且采用与样本无关的目标标",
"签方式(Gu et al.,2019;Yao et al.,2019;Liu et al.,2020;Li et al.,2020b;Nguyen &Tran,",
"2021；Zhaietal.，2021）。换句话说，攻击者为所有中毒样本分配相同的目标标签。我们发",
"现，这种范式在攻击语义分割方面仍然有效。。但是，这种方法只能从图像级别操纵预测，",
"因此无法实现对语义分割的更精细的恶意操纵。。为了解决这个问题，在本文中，我们提出",
"了一种细粒度的攻击范式，其中目标标签是特定于样本的。具体来说，我们从对象级别而不",
"是图像级别处理目标标签，其中只有特定对象的像素将被标记为攻击者指定的目标类，而",
"其他对象仍被标记为地面实况。实验验证了我们的方法可以通过仅毒化一小部分训练数据",
"来成功和隐秘地攻击语义分割模型",
"*indicates equal contribution.",
"1"
] | {
"arxiv_id": "2103.04038",
"language": "zh",
"page_idx": 0
} |
||
2103.04038-zh-1 | 2103.04038-zh-1.jpg | $\underline{\text{Published at ICLR 2021 Workshop on Security and Safety in Machine Learning Systems}}$
这项工作的主要贡献有三个方面：(1)我们证明了现有的攻击范式在攻击语义分割方面仍然有效，这首先揭示了语义分割训练过程中的后门威胁。(2)我们探索了一种新颖的细粒度攻击范式，该范式可以实现更复杂的攻击操纵。(3)进行了广泛的实验，验证了我们攻击的有效性和隐蔽性
2 The Proposed Attack
2.1 Preliminaries
语义分割。令$\mathcal{D}=\{(x_i,y_i)\}_{i=1}^N$表示 (良性)训练集，其中$x_i\in\mathcal{X}=\{0,1,\ldots,255\}^{C\times W\times H}$
$\text{比 人 们 时 }^{\circ }$ $< D- [ ( \boldsymbol{\omega }_i, \boldsymbol{y}_i) ] _{i= 1}$ 权 小 (区 上 ) 阶 价 不 , 六 $| \boldsymbol{\omega }_i\subset \mathcal{N} , \ldots , 200]$ 是 图 像 , $\boldsymbol{y}_i\in \mathcal{Y} = \{ 0, 1, \ldots , K\} ^{W\times H}$ 是 $x_i$ 的 标 签 $( i. e. $, 按 像 素 注 释 $) , K$ 是 数 据 集 上 包 个 的 对 伤 数 口 单 +名 数 而 左 的 还 $\mathcal{Y}$人 制 措 刑 都 目 其 工 DNN 的 空 行 目 活 +画 勾 业 长 报含的对象数。目前，大多数现有的语义分割模型都是基于 DNN 的，它们是通过端到端监督的方式学习的。具体来说，那些旨在通过$\min_{\boldsymbol{\theta}}\frac1N\sum_{i=1}^N\mathcal{L}\left(f_{\boldsymbol{\theta}}(x_i),\boldsymbol{y_i}\right)$学习 DNN (参数为$\theta$ ), $i. e.$ , $f_{\theta }: \mathcal{X} \to \mathcal{Y}$的方法，其中$\mathcal{L}(\cdot)$表示损失函数
现有后门攻击的一般管道。通常，后门攻击有两个主要过程，包括(1)生成中毒数据集$\mathcal{D}_{poisoned}$和(2)使用$\mathcal{D}_poisoned$进行训练。第一个过程是后门攻击的基石。目前，所有现有攻击的目标标签都与样本无关，$i.e.$所有中毒的样本都被分配了相同的(目标)标签。具体来说，$D_poisoned$包含$\mathcal{D}$子集的中毒版本和剩余的良性样本$i.e.$, $\mathcal{D} _{poisoned}= \mathcal{D} _{modified}\cup \mathcal{D} _{benign}$ ,其中 $\mathcal{D} _benign\subset \mathcal{D}$ $, \gamma = \frac {| \mathcal{D} _{modified}| }{| \mathcal{D} | }$ 表示中毒率， $\mathcal{D} _{modified}= \{ ( x^{\prime }, \boldsymbol{y}_t) | \boldsymbol{x}^{\prime }= G( \boldsymbol{x}) , ( \boldsymbol{x}, \boldsymbol{y}) \in \mathcal{D} \backslash \mathcal{D} _{benign}\}$ , $\boldsymbol{y}_t$是指定的目标标签，$G:\mathcal{X}\to\mathcal{X}$ 攻击者是。例如，如 (Chen et al., 2017b) , $G(x)=(1-\lambda)\otimes x+\lambda\otimes t$中提出的，其中$\boldsymbol{\lambda}\in[0,1]^{C\times W\times H}$ 是与可见性相关的超参数$,\boldsymbol{t}\in\boldsymbol{\chi}$ 是预定义的触发模式，而$\otimes$表示元素方面的产品
威胁模型。在本文中，我们假设攻击者可以出于恶意目的修改训练集，而他们无法访问或修改训练过程中涉及的其他部分$(e.g.$,模型结构和训练损失),并且没有有关推理过程的信息。正如 (Li et al., 2020a) 所建议的，这是后门攻击者的 commmon 设置，这使得攻击可能发生在许多现实世界的场景中$(e.g.$,采用第三方训练平台或模型)
攻击者的目标。与现有攻击类似，在本文中，攻击者有两个主要目标，包括在中毒训练集上训练的模型的有效性和隐蔽性。具体来说，有效性要求当触发模式出现时，具有源类(i.e., 攻击者指定的分类错误的类)的对象的像素将被预测为目标类，而隐蔽性要求(1)触发模式不引人注目，(2)被攻击的模型在良性测试样本上表现正常，以及(3)在被攻击的样本中具有非源类的像素上的性能不会显着降低
2.2 Fine-Grained Backdoor Attack (FGBA)
在本节中，我们将说明我们提出的细粒度后门攻击。在我们描述如何在攻击中生成中毒样
本之前，我们首先介绍它的定义
Definition 1 (Fine-grained Backdoor Attack). For a semantic segmentation dataset $\mathcal{D}$ containing $K$ objects, let $T:\mathcal{Y}\to\mathcal{Y}$ indicates an attacker-specified target label generator, and $A\in\{0,1\}^{K\times K}$ indicates the attack matrix, where $\sum_jA_{ij}=1$ and $A_{ij}=1$ indicates the class of pixels with ground-truth class $i$ will be labeled as class $j.$ Fine-grained backdoor $\underset{a}{\operatorname*{\text{attack }}}($with attack matrix $A)$ generates the taget label of image $\boldsymbol x$ by $T(\boldsymbol{y};A)=\hat{\boldsymbol{y}}$, where $\hat{\boldsymbol{y}} _{ij}= \arg \max _{k\in \{ 1, \cdots , K\} }A_{\boldsymbol{y}_{ij}, k}$ .
Remark 1. Existing backdoor attacks $(e.g.$ , BadNets) adopted a sample-agnostic target $\begin{array}{l}\mathrm{label~paradigm,~}i.e.~,~T(\boldsymbol{y}_{i})=\boldsymbol{y}_{t},\forall(\boldsymbol{x}_{i},\boldsymbol{y}_{i})\in\mathcal{D}\backslash\mathcal{D}_{benign}~.~\mathrm{In~contrast,~target~labels~of~the}\\\mathrm{proposed~fine-grained~backdoor~attack~are~usually~sample-specific.~As~such,~existing~attacks}\end{array}$ are not fine-grained. An example of target labels generated by different attacks are shown $\operatorname{in}$Fig.1.
如 2.1 节所示，除了目标标签生成外，后门攻击还需要基于生成器$G(\cdot)$生成中毒图像。在本文中，我们采用了先前作品中提出的设置，包括混合策略(Chen et al., 2017b)(i.e., $G(\boldsymbol{x})=(\boldsymbol{1}-\boldsymbol{\lambda})\otimes\boldsymbol{x}+\boldsymbol{\lambda}\otimes\boldsymbol{t}$ )和语义触发器 (Bagdasaryan et al., 2020) $(i.e.,G(\boldsymbol{x})=\boldsymbol{x})$ 的非语义触发器。。与非语义触发器相比，采用语义触发器不需要在训练和推理过程中修改图像，因此更加隐秘和方便。然而，与非语义触发器相比，它通常性能相对较差，因为语义触发器更难被 DNN 记忆。将在 3.2 节中进一步验证
请注意，细粒度攻击和 BadNets 型攻击都可以自然地抵抗分类任务中提出的某些潜在后门防御。例如，他们可以绕过基于触发合成的经验防御 (Wang et al., 2019; Guo et al., 2020;
2 | [
[
"0.1765",
"0.0341",
"0.8096",
"0.0341",
"0.8096",
"0.04736",
"0.1765",
"0.04736"
],
[
"0.1765",
"0.1061",
"0.822",
"0.1061",
"0.822",
"0.1193",
"0.1765",
"0.1193"
],
[
"0.1772",
"0.1218",
"0.8203",
"0.1218",
"0.8203",
"0.132",
"0.1772",
"0.132"
],
[
"0.1765",
"0.1332",
"0.822",
"0.1332",
"0.822",
"0.1465",
"0.1765",
"0.1465"
],
[
"0.1765",
"0.1483",
"0.2754",
"0.1483",
"0.2754",
"0.1604",
"0.1765",
"0.1604"
],
[
"0.1748",
"0.1812",
"0.3906",
"0.1812",
"0.3906",
"0.1957",
"0.1748",
"0.1957"
],
[
"0.1765",
"0.2052",
"0.312",
"0.2052",
"0.312",
"0.2153",
"0.1765",
"0.2153"
],
[
"0.1732",
"0.2203",
"0.8228",
"0.2172",
"0.8228",
"0.238",
"0.1732",
"0.2412"
],
[
"0.1724",
"0.2343",
"0.824",
"0.2317",
"0.824",
"0.254",
"0.1724",
"0.2556"
],
[
"0.1772",
"0.2524",
"0.822",
"0.2524",
"0.822",
"0.2664",
"0.1772",
"0.2664"
],
[
"0.1765",
"0.2695",
"0.824",
"0.2695",
"0.824",
"0.2827",
"0.1765",
"0.2827"
],
[
"0.1765",
"0.2854",
"0.5664",
"0.2854",
"0.5664",
"0.2986",
"0.1765",
"0.2986"
],
[
"0.1772",
"0.3062",
"0.8228",
"0.3062",
"0.8228",
"0.3193",
"0.1772",
"0.3193"
],
[
"0.1765",
"0.3208",
"0.822",
"0.3208",
"0.822",
"0.334",
"0.1765",
"0.334"
],
[
"0.1765",
"0.3333",
"0.8228",
"0.3333",
"0.8228",
"0.348",
"0.1765",
"0.348"
],
[
"0.1765",
"0.348",
"0.822",
"0.348",
"0.822",
"0.3623",
"0.1765",
"0.3623"
],
[
"0.1691",
"0.3567",
"0.831",
"0.3591",
"0.831",
"0.3838",
"0.1691",
"0.3818"
],
[
"0.1748",
"0.3818",
"0.824",
"0.3818",
"0.824",
"0.3997",
"0.1748",
"0.3997"
],
[
"0.1765",
"0.3965",
"0.822",
"0.3965",
"0.822",
"0.4104",
"0.1765",
"0.4104"
],
[
"0.1708",
"0.404",
"0.824",
"0.4048",
"0.824",
"0.4268",
"0.1708",
"0.4255"
],
[
"0.1772",
"0.4255",
"0.2402",
"0.4255",
"0.2402",
"0.4375",
"0.1772",
"0.4375"
],
[
"0.1772",
"0.4458",
"0.822",
"0.4458",
"0.822",
"0.459",
"0.1772",
"0.459"
],
[
"0.1765",
"0.4602",
"0.822",
"0.4602",
"0.822",
"0.4734",
"0.1765",
"0.4734"
],
[
"0.1765",
"0.473",
"0.8228",
"0.4734",
"0.8228",
"0.488",
"0.1765",
"0.4873"
],
[
"0.1765",
"0.488",
"0.6562",
"0.488",
"0.6562",
"0.502",
"0.1765",
"0.502"
],
[
"0.1765",
"0.509",
"0.824",
"0.509",
"0.824",
"0.523",
"0.1765",
"0.523"
],
[
"0.179",
"0.5234",
"0.808",
"0.5234",
"0.808",
"0.534",
"0.179",
"0.534"
],
[
"0.1748",
"0.535",
"0.822",
"0.535",
"0.822",
"0.549",
"0.1748",
"0.549"
],
[
"0.1765",
"0.5503",
"0.8228",
"0.5503",
"0.8228",
"0.5645",
"0.1765",
"0.5645"
],
[
"0.1765",
"0.5645",
"0.486",
"0.5645",
"0.486",
"0.5776",
"0.1765",
"0.5776"
],
[
"0.1748",
"0.586",
"0.5034",
"0.586",
"0.5034",
"0.5996",
"0.1748",
"0.5996"
],
[
"0.1765",
"0.608",
"0.822",
"0.608",
"0.822",
"0.622",
"0.1765",
"0.622"
],
[
"0.1772",
"0.6235",
"0.4043",
"0.6235",
"0.4043",
"0.634",
"0.1772",
"0.634"
],
[
"0.1748",
"0.638",
"0.819",
"0.638",
"0.819",
"0.6514",
"0.1748",
"0.6514"
],
[
"0.1748",
"0.653",
"0.8228",
"0.653",
"0.8228",
"0.6665",
"0.1748",
"0.6665"
],
[
"0.1732",
"0.662",
"0.824",
"0.664",
"0.824",
"0.6816",
"0.1732",
"0.6797"
],
[
"0.1748",
"0.681",
"0.822",
"0.681",
"0.822",
"0.696",
"0.1748",
"0.696"
],
[
"0.1765",
"0.6963",
"0.8228",
"0.6963",
"0.8228",
"0.7104",
"0.1765",
"0.7104"
],
[
"0.1732",
"0.7104",
"0.4043",
"0.712",
"0.4043",
"0.726",
"0.1732",
"0.7236"
],
[
"0.1732",
"0.729",
"0.824",
"0.728",
"0.824",
"0.7456",
"0.1732",
"0.747"
],
[
"0.1732",
"0.7437",
"0.8228",
"0.7437",
"0.8228",
"0.761",
"0.1732",
"0.761"
],
[
"0.1748",
"0.759",
"0.8228",
"0.7583",
"0.8228",
"0.773",
"0.1748",
"0.7734"
],
[
"0.1765",
"0.773",
"0.8228",
"0.773",
"0.8228",
"0.786",
"0.1765",
"0.786"
],
[
"0.1732",
"0.786",
"0.2484",
"0.787",
"0.2484",
"0.799",
"0.1732",
"0.7974"
],
[
"0.1772",
"0.808",
"0.822",
"0.808",
"0.822",
"0.8184",
"0.1772",
"0.8184"
],
[
"0.1765",
"0.821",
"0.8145",
"0.821",
"0.8145",
"0.834",
"0.1765",
"0.834"
],
[
"0.1765",
"0.834",
"0.824",
"0.834",
"0.824",
"0.8477",
"0.1765",
"0.8477"
],
[
"0.1765",
"0.8486",
"0.822",
"0.8486",
"0.822",
"0.8623",
"0.1765",
"0.8623"
],
[
"0.1765",
"0.8623",
"0.822",
"0.8623",
"0.822",
"0.8755",
"0.1765",
"0.8755"
],
[
"0.1765",
"0.8755",
"0.55",
"0.8755",
"0.55",
"0.8896",
"0.1765",
"0.8896"
],
[
"0.1772",
"0.898",
"0.8228",
"0.898",
"0.8228",
"0.911",
"0.1772",
"0.911"
],
[
"0.1765",
"0.9136",
"0.822",
"0.9136",
"0.822",
"0.924",
"0.1765",
"0.924"
],
[
"0.4902",
"0.9487",
"0.5073",
"0.9487",
"0.5073",
"0.962",
"0.4902",
"0.962"
]
] | [
"Published at ICLR 2021 Workshop on Security and Safety in Machine Learning Systems",
"这项工作的主要贡献有三个方面：（1）我们证明了现有的攻击范式在攻击语义分割方面仍然",
"有效，这首先揭示广语义分割训练过程中的后门威胁。（2）我们探索了一种新颖的细粒度攻",
"击范式，该范式可以实现更复杂的攻击操纵。（3）进行了广泛的实验，验证了我们攻击的有",
"效性和隐蔽性",
"2The Proposed Attack",
"2.1Preliminaries",
"语义分割。令D=【(ai,yi)表示（良性）训练集，其中a;E=[0,1,..,255)CxWxH",
"是图像，yie={0,1\")wxH是的标签（i.e.，按像素注释），是数据集上包",
"含的对象数。目前，大多数现有的语义分割模型都是基于DNN的，它们是通过端到端监督",
"的方式学习的。具体来说，那些旨在通过minoL（fe（ai）,yi）学习DNN（参数为0",
")，i.e.，fe：→y的方法，其中L()表示损失函数",
"现有后门攻击的一般管道。通常，后门攻击有两个主要过程，包括（1）生成中毒数",
"据集Dpoisoned 和（2）使用Dpoisoned 进行训练。第一个过程是后门攻击的基石。目",
"前，所有现有攻击的目标标签都与样本无关，i.e。所有中毒的样本都被分配了相同的",
"（目标）标签。具体来说，Dpoisoned包含D子集的中毒版本和剩余的良性样本i.e.",
"Dpoisoned = Dmoadified UDenign ，其中Denign CD，=Dmbjual 表示中毒率",
"Dmodified=((a,yt)la=G(a),(a,y)ED/Dbenign) ，Yt是指定的目标标签，G:→X",
"攻击者是。例如，如（Chen et al.,2017b），G(a）=（1－^)a+入t中提出的，其中",
"E[,1cxWxH是与可见性相关的超参数，tex是预定义的触发模式，而表示元素方",
"面的产品",
"威胁模型。在本文中，我们假设攻击者可以出于恶意目的修改训练集，而他们无法访问或修",
"改训练过程中涉及的其他部分（e.9：，模型结构和训练损失），并且没有有关推理过程的信",
"息。正如（Lietal.，2020a）所建议的，这是后门攻击者的commmon设置，这使得攻击可能",
"发生在许多现实世界的场景中（e.9.，采用第三方训练平台或模型）",
"攻击者的目标。与现有攻击类似，在本文中，攻击者有两个主要目标，包括在中毒训练集上",
"训练的模型的有效性和隐蔽性。具体来说，有效性要求当触发模式出现时，具有源类（i.e",
"攻击者指定的分类错误的类）的对象的像素将被预测为目标类，而隐蔽性要求（1）触发模",
"式不引人注目，（2）被攻击的模型在良性测试样本上表现正常，以及（3）在被攻击的样本",
"中具有非源类的像素上的性能不会显着降低",
"2.2Fine-Grained Backdoor Attack (FGBA)",
"在本节中，我们将说明我们提出的细粒度后门攻击。在我们描述如何在攻击中生成中毒样",
"本之前，我们首先介绍它的定义",
"Definition 1 (Fine-grained Backdoor Attack). For a semantic segmentation dataset D con",
"taining K objects, let T: → y indicates an attacker-specified target label generator, and",
"A e (0, ijKK indicates the attack matrix, where Z, Aij = 1 and Aij = 1 indicates the",
"class of pixels with ground-truth class i will be labeled as class j . Fine-grained backdoor",
"attack (with attack matrix A) generates the taget label of image a by T(y; A) = y , where",
"yii=argmaxke(1,..,K)Ayig,h.",
"Remark 1. Existing backdoor attacks ( e.g. , BadNets) adopted a sample-agnostic target",
"label paradigm, i.e. , T(yi) = Yt, V(ri, yi) e D)Dbenign . In contrast, target labels of the",
"proposed fine-grained backdoor attack are usually sample-specific. As such, existing attacks",
"are not fine-grained. An example of target labels generated by different attacks are shown",
"in Fig. 1.",
"如2.1节所示：除了自标标签生成外，后门攻击还需要基于生成器G（）生成中毒图像。在",
"本文中，我们采用了先前作品中提出的设置，包括混合策略（Chenetal.，2017b）（i.e.",
"G(a)=（1-^)a+>t）和语义触发器(Bagdasaryanetal.,2020)（i.e.，G(a)=a）",
"的非语义触发器。。与非语义触发器相比，采用语义触发器不需要在训练和推理过程中修改",
"图像，因此更加隐秘和方便。然而，与非语义触发器相比，它通常性能相对较差，因为语义",
"触发器更难被DNN记忆。将在3.2节中进一步验证",
"请注意，细粒度攻击和BadNets型攻击都可以自然地抵抗分类任务中提出的某些潜在后门",
"防御。例如，他们可以绕过基于触发合成的经验防御（wangetal.2019：Guoetal.，2020",
"2"
] | {
"arxiv_id": "2103.04038",
"language": "zh",
"page_idx": 1
} |
||
2103.04038-zh-2 | 2103.04038-zh-2.jpg | Published at ICLR 2021 Workshop on Security and Safety in Machine Learning Systems
<FigureHere>
Figure 1: Comparison between our proposed fne-grained backdoor attack and previous
$\bar{\text{attacks methods}} ( e. g.$ ,BadNets). In this example, target labels of BadNets $( i. e.$ ,images in the third column) are all the same road scape, while those of our method $(i.e.$, images in the last column) are sample-specific. Specifically, all pixels with ground-truth class `person' (those in the white-box of images) are labeled as the class ‘palm’in the target labels.
<FigureHere>
Figure 2: The illustration of poisoned images generated by attacks with the semantic and
non-semantic trigger. In this example, the trigger is denoted in the red box where object
“wall’ is the semantic trigger and‘black line’ is the non-semantic trigger.
Zhu et al., 2020) , 因为这些方法需要反转每个潜在目标标签的潜在触发模式，并且其计算复杂性在分段任务中难以承受$(\mathcal{O}(K^2)$用于细粒度攻击，$\mathcal{O}(K^W\times H)$用于 BadNets 型攻击)。这将在我们今后的工作中进一步研究
## 3 Experiment
## 3.1 Settings
数据集选择和模型结构。我们采用三种最先进的细分模型，包括 DeepLabv3 (Chen et al., 2017a)、DenseASPP (Yang et al., 2018) 和 DANet (Fu et al., 2019) 进行评估。所有模型都在 ADE20K 数据集 (Zhou et al.,2017) 上训练
Table 1: The effectiveness $(\%)$ of BadNets attack with the non-semantic and semantic
$\operatorname{trigger}$
<table>
<tbody>
<tr>
<th colspan="2">${\mathrm{Trigger~}^{\prime}\text{Type}\to}$</th>
<th colspan="4">Non-semantic ${:\text{Trigger}}$</th>
<th colspan="4">Semantic ${\mathrm{c}}^{\prime}$Trigger</th>
</tr>
<tr>
<th>Model</th>
<th>Method$\downarrow $,Metric$\to$</th>
<th>${\mathrm{mIOU-B}}$</th>
<th>$\overline{\mathrm{PA-B}}$</th>
<th>${\mathrm{mIOU-A}}$</th>
<th>$\overline{\mathrm{ASR}}$</th>
<th>mIOU- B</th>
<th>$\overline{PA-B}$</th>
<th>mIOU- A</th>
<th>$\overline{ASR}$</th>
</tr>
<tr>
<td rowspan="2">DeepLabv3</td>
<td>Benigm</td>
<td>35.3</td>
<td>73.3</td>
<td>4.0</td>
<td>8.9</td>
<td>35.3</td>
<td>$\overline{73.3}$</td>
<td>5.1</td>
<td>$\overline{13.0}$</td>
</tr>
<tr>
<td>BadNets</td>
<td>34.8</td>
<td>74.2</td>
<td>78.5</td>
<td>96.9</td>
<td>33.3</td>
<td>69.7</td>
<td>35.8</td>
<td>63.3</td>
</tr>
<tr>
<td rowspan="2">DenseASPP</td>
<td>Benign</td>
<td>$\overline{37.3}$</td>
<td>75.5</td>
<td>$\overline{4.0}$</td>
<td>$\overline{8.9}$</td>
<td>$\overline{37.3}$</td>
<td>75.5</td>
<td>$\overline{5.2}$</td>
<td>$\overline{13.4}$</td>
</tr>
<tr>
<td>BadNets</td>
<td>37.1</td>
<td>75.8</td>
<td>90.1</td>
<td>98.8</td>
<td>35.3</td>
<td>71.6</td>
<td>46.4</td>
<td>69.5</td>
</tr>
<tr>
<td rowspan="2">DANet</td>
<td>Benigm</td>
<td>38.3</td>
<td>75.9</td>
<td>4.1</td>
<td>9.1</td>
<td>38.3</td>
<td>$\overline{75.9}$</td>
<td>5.2</td>
<td>$\overline{13.5}$</td>
</tr>
<tr>
<td>BadNets</td>
<td>38.0</td>
<td>76.1</td>
<td>79.8</td>
<td>97.3</td>
<td>36.5</td>
<td>71.8</td>
<td>42.6</td>
<td>70.2</td>
</tr>
</tbody>
</table>
$\overline{\text{攻击设置。我们概括了 BadNets }(\text{Gu et al.,2019),并评估了我们的细粒度攻击的两个重要}}$
特殊情况，包括(1)N对1攻击和(2)1对1攻击。对于N对1攻击，具有所有类的像素将被标记为目标标签中的“墙”,而只有具有“人”类的像素将被标记为1对1攻击生成的目标标签中的“手掌”。此外，我们采用 8 像素宽的黑线作为非语义触发器$(\boldsymbol{\lambda}\in\{0,1\}^{C\times W\times H}$ )和对象“墙”作为语义触发器。$\gamma=10\%$对 N 对 1 攻击和$\gamma$的中毒率接近 20% 对 1 对1攻击。此外，对于带有语义触发器的细粒度攻击，所有同时包含“墙”和“人”对象的样本都被选择为中毒样本。对于 BadNets, 目标标签是从数据集中随机选择的“道路景观”图
3 | [
[
"0.1748",
"0.0341",
"0.8105",
"0.0341",
"0.8105",
"0.04736",
"0.1748",
"0.04736"
],
[
"0.591",
"0.0802",
"0.6724",
"0.0802",
"0.6724",
"0.09155",
"0.591",
"0.09155"
],
[
"0.2712",
"0.0871",
"0.3203",
"0.089",
"0.3203",
"0.101",
"0.2712",
"0.0985"
],
[
"0.3816",
"0.0884",
"0.468",
"0.0884",
"0.468",
"0.09973",
"0.3816",
"0.09973"
],
[
"0.5474",
"0.0966",
"0.597",
"0.0966",
"0.597",
"0.108",
"0.5474",
"0.108"
],
[
"0.6807",
"0.0966",
"0.7114",
"0.0966",
"0.7114",
"0.1092",
"0.6807",
"0.1092"
],
[
"0.6846",
"0.1465",
"0.716",
"0.1465",
"0.716",
"0.1578",
"0.6846",
"0.1578"
],
[
"0.1748",
"0.2556",
"0.826",
"0.2556",
"0.826",
"0.2695",
"0.1748",
"0.2695"
],
[
"0.1748",
"0.2695",
"0.826",
"0.2695",
"0.826",
"0.2842",
"0.1748",
"0.2842"
],
[
"0.1732",
"0.2842",
"0.824",
"0.2842",
"0.824",
"0.2974",
"0.1732",
"0.2974"
],
[
"0.1748",
"0.2974",
"0.8228",
"0.2974",
"0.8228",
"0.3113",
"0.1748",
"0.3113"
],
[
"0.1765",
"0.3125",
"0.785",
"0.3125",
"0.785",
"0.3257",
"0.1765",
"0.3257"
],
[
"0.3105",
"0.3372",
"0.4575",
"0.3372",
"0.4575",
"0.348",
"0.3105",
"0.348"
],
[
"0.5767",
"0.3352",
"0.7466",
"0.336",
"0.7466",
"0.3499",
"0.5767",
"0.3484"
],
[
"0.2688",
"0.3516",
"0.2998",
"0.3547",
"0.298",
"0.3623",
"0.267",
"0.3591"
],
[
"0.2009",
"0.3726",
"0.2402",
"0.3726",
"0.2402",
"0.3838",
"0.2009",
"0.3838"
],
[
"0.738",
"0.3674",
"0.744",
"0.3674",
"0.744",
"0.3706",
"0.738",
"0.3706"
],
[
"0.726",
"0.3774",
"0.734",
"0.3774",
"0.734",
"0.3806",
"0.726",
"0.3806"
],
[
"0.2018",
"0.3838",
"0.2385",
"0.3838",
"0.2385",
"0.3958",
"0.2018",
"0.3958"
],
[
"0.5",
"0.3901",
"0.5107",
"0.3901",
"0.5107",
"0.3933",
"0.5",
"0.3933"
],
[
"0.281",
"0.4312",
"0.298",
"0.4312",
"0.298",
"0.435",
"0.281",
"0.435"
],
[
"0.5537",
"0.43",
"0.5825",
"0.43",
"0.5825",
"0.4375",
"0.5537",
"0.4375"
],
[
"0.1953",
"0.4475",
"0.2443",
"0.4475",
"0.2443",
"0.459",
"0.1953",
"0.459"
],
[
"0.2018",
"0.459",
"0.2385",
"0.459",
"0.2385",
"0.4702",
"0.2018",
"0.4702"
],
[
"0.1732",
"0.4976",
"0.8228",
"0.4976",
"0.8228",
"0.512",
"0.1732",
"0.512"
],
[
"0.1748",
"0.512",
"0.824",
"0.512",
"0.824",
"0.526",
"0.1748",
"0.526"
],
[
"0.1748",
"0.5264",
"0.693",
"0.5264",
"0.693",
"0.5405",
"0.1748",
"0.5405"
],
[
"0.1732",
"0.5444",
"0.8228",
"0.5454",
"0.8228",
"0.56",
"0.1732",
"0.5596"
],
[
"0.1724",
"0.558",
"0.8267",
"0.5547",
"0.8267",
"0.5747",
"0.1724",
"0.5776"
],
[
"0.1765",
"0.5737",
"0.4795",
"0.5737",
"0.4795",
"0.5884",
"0.1765",
"0.5884"
],
[
"0.1732",
"0.6094",
"0.1912",
"0.6094",
"0.1912",
"0.6226",
"0.1732",
"0.6226"
],
[
"0.2075",
"0.608",
"0.3096",
"0.61",
"0.3096",
"0.625",
"0.2075",
"0.6226"
],
[
"0.1748",
"0.6396",
"0.2009",
"0.6396",
"0.2009",
"0.6514",
"0.1748",
"0.6514"
],
[
"0.2141",
"0.637",
"0.2769",
"0.64",
"0.2754",
"0.654",
"0.2133",
"0.651"
],
[
"0.1748",
"0.662",
"0.8203",
"0.664",
"0.8203",
"0.6787",
"0.1748",
"0.677"
],
[
"0.1748",
"0.6787",
"0.8228",
"0.6787",
"0.8228",
"0.6934",
"0.1748",
"0.6934"
],
[
"0.1732",
"0.6914",
"0.5254",
"0.6914",
"0.5254",
"0.705",
"0.1732",
"0.705"
],
[
"0.1732",
"0.7085",
"0.826",
"0.709",
"0.826",
"0.7236",
"0.1732",
"0.7227"
],
[
"0.1765",
"0.726",
"0.2279",
"0.726",
"0.2279",
"0.7373",
"0.1765",
"0.7373"
],
[
"0.2443",
"0.7363",
"0.3308",
"0.738",
"0.3308",
"0.75",
"0.2443",
"0.7476"
],
[
"0.4468",
"0.7363",
"0.572",
"0.7373",
"0.572",
"0.7485",
"0.4468",
"0.7476"
],
[
"0.6787",
"0.7363",
"0.778",
"0.7373",
"0.778",
"0.7485",
"0.6787",
"0.7476"
],
[
"0.2615",
"0.747",
"0.4019",
"0.747",
"0.4019",
"0.7603",
"0.2615",
"0.7603"
],
[
"0.5864",
"0.747",
"0.619",
"0.747",
"0.619",
"0.759",
"0.5864",
"0.759"
],
[
"0.6216",
"0.7456",
"0.683",
"0.7456",
"0.683",
"0.7603",
"0.6216",
"0.7603"
],
[
"0.8057",
"0.747",
"0.841",
"0.747",
"0.841",
"0.759",
"0.8057",
"0.759"
],
[
"0.398",
"0.7476",
"0.4617",
"0.7476",
"0.4617",
"0.759",
"0.398",
"0.759"
],
[
"0.1953",
"0.7476",
"0.2467",
"0.7476",
"0.2467",
"0.759",
"0.1953",
"0.759"
],
[
"0.468",
"0.7476",
"0.5703",
"0.7476",
"0.5703",
"0.759",
"0.468",
"0.759"
],
[
"0.6895",
"0.7476",
"0.79",
"0.7476",
"0.79",
"0.759",
"0.6895",
"0.759"
],
[
"0.3105",
"0.7583",
"0.3545",
"0.761",
"0.353",
"0.7734",
"0.3096",
"0.771"
],
[
"0.4763",
"0.7583",
"0.505",
"0.7603",
"0.5034",
"0.7715",
"0.4756",
"0.7695"
],
[
"0.697",
"0.7583",
"0.726",
"0.7603",
"0.7256",
"0.7715",
"0.6953",
"0.7695"
],
[
"0.535",
"0.759",
"0.558",
"0.759",
"0.558",
"0.7715",
"0.535",
"0.7715"
],
[
"0.4192",
"0.759",
"0.4468",
"0.759",
"0.4468",
"0.7715",
"0.4192",
"0.7715"
],
[
"0.7534",
"0.759",
"0.777",
"0.759",
"0.777",
"0.7715",
"0.7534",
"0.7715"
],
[
"0.806",
"0.759",
"0.8364",
"0.759",
"0.8364",
"0.7715",
"0.806",
"0.7715"
],
[
"0.6396",
"0.7603",
"0.6685",
"0.7603",
"0.6685",
"0.7715",
"0.6396",
"0.7715"
],
[
"0.1887",
"0.7646",
"0.2566",
"0.7646",
"0.2566",
"0.776",
"0.1887",
"0.776"
],
[
"0.589",
"0.7603",
"0.613",
"0.7603",
"0.613",
"0.7715",
"0.589",
"0.7715"
],
[
"0.4763",
"0.768",
"0.505",
"0.771",
"0.5024",
"0.7827",
"0.4756",
"0.781"
],
[
"0.5317",
"0.768",
"0.5596",
"0.7715",
"0.558",
"0.782",
"0.5293",
"0.78"
],
[
"0.585",
"0.768",
"0.614",
"0.771",
"0.613",
"0.782",
"0.584",
"0.78"
],
[
"0.6396",
"0.7695",
"0.6665",
"0.7695",
"0.6665",
"0.781",
"0.6396",
"0.781"
],
[
"0.697",
"0.7695",
"0.7236",
"0.7695",
"0.7236",
"0.781",
"0.697",
"0.781"
],
[
"0.7524",
"0.7695",
"0.778",
"0.7695",
"0.778",
"0.782",
"0.7524",
"0.782"
],
[
"0.8057",
"0.7695",
"0.835",
"0.7695",
"0.835",
"0.782",
"0.8057",
"0.782"
],
[
"0.4192",
"0.7695",
"0.4453",
"0.7695",
"0.4453",
"0.782",
"0.4192",
"0.782"
],
[
"0.3064",
"0.771",
"0.357",
"0.771",
"0.357",
"0.782",
"0.3064",
"0.782"
],
[
"0.4192",
"0.781",
"0.4453",
"0.781",
"0.4453",
"0.7935",
"0.4192",
"0.7935"
],
[
"0.4763",
"0.781",
"0.5034",
"0.781",
"0.5034",
"0.7935",
"0.4763",
"0.7935"
],
[
"0.535",
"0.781",
"0.556",
"0.781",
"0.556",
"0.7935",
"0.535",
"0.7935"
],
[
"0.5884",
"0.781",
"0.6113",
"0.781",
"0.6113",
"0.7935",
"0.5884",
"0.7935"
],
[
"0.6396",
"0.781",
"0.6665",
"0.781",
"0.6665",
"0.7935",
"0.6396",
"0.7935"
],
[
"0.697",
"0.781",
"0.7236",
"0.781",
"0.7236",
"0.7935",
"0.697",
"0.7935"
],
[
"0.7534",
"0.781",
"0.777",
"0.781",
"0.777",
"0.7935",
"0.7534",
"0.7935"
],
[
"0.806",
"0.781",
"0.835",
"0.781",
"0.835",
"0.7925",
"0.806",
"0.7925"
],
[
"0.3096",
"0.782",
"0.3545",
"0.782",
"0.3545",
"0.7935",
"0.3096",
"0.7935"
],
[
"0.1871",
"0.786",
"0.2566",
"0.786",
"0.2566",
"0.7974",
"0.1871",
"0.7974"
],
[
"0.3064",
"0.7905",
"0.357",
"0.7925",
"0.357",
"0.8037",
"0.3064",
"0.802"
],
[
"0.4192",
"0.7915",
"0.4453",
"0.7915",
"0.4453",
"0.803",
"0.4192",
"0.803"
],
[
"0.4763",
"0.7915",
"0.5024",
"0.7915",
"0.5024",
"0.803",
"0.4763",
"0.803"
],
[
"0.5317",
"0.7915",
"0.556",
"0.7915",
"0.556",
"0.803",
"0.5317",
"0.803"
],
[
"0.5864",
"0.7915",
"0.613",
"0.7915",
"0.613",
"0.803",
"0.5864",
"0.803"
],
[
"0.6396",
"0.7915",
"0.6665",
"0.7915",
"0.6665",
"0.803",
"0.6396",
"0.803"
],
[
"0.697",
"0.7915",
"0.7236",
"0.7915",
"0.7236",
"0.803",
"0.697",
"0.803"
],
[
"0.7534",
"0.7915",
"0.778",
"0.7915",
"0.778",
"0.803",
"0.7534",
"0.803"
],
[
"0.8057",
"0.7915",
"0.835",
"0.7915",
"0.835",
"0.803",
"0.8057",
"0.803"
],
[
"0.3105",
"0.8003",
"0.3545",
"0.8037",
"0.353",
"0.8164",
"0.3096",
"0.813"
],
[
"0.4763",
"0.803",
"0.5034",
"0.803",
"0.5034",
"0.8145",
"0.4763",
"0.8145"
],
[
"0.5884",
"0.803",
"0.6113",
"0.803",
"0.6113",
"0.8145",
"0.5884",
"0.8145"
],
[
"0.697",
"0.803",
"0.7256",
"0.803",
"0.7256",
"0.8145",
"0.697",
"0.8145"
],
[
"0.806",
"0.803",
"0.835",
"0.803",
"0.835",
"0.8145",
"0.806",
"0.8145"
],
[
"0.536",
"0.8047",
"0.556",
"0.8047",
"0.556",
"0.815",
"0.536",
"0.815"
],
[
"0.4192",
"0.8037",
"0.4468",
"0.8037",
"0.4468",
"0.815",
"0.4192",
"0.815"
],
[
"0.6396",
"0.8047",
"0.6665",
"0.8047",
"0.6665",
"0.813",
"0.6396",
"0.813"
],
[
"0.7563",
"0.8047",
"0.777",
"0.8047",
"0.777",
"0.815",
"0.7563",
"0.815"
],
[
"0.3064",
"0.8125",
"0.357",
"0.8145",
"0.357",
"0.8257",
"0.3064",
"0.8237"
],
[
"0.4192",
"0.813",
"0.4468",
"0.813",
"0.4468",
"0.8257",
"0.4192",
"0.8257"
],
[
"0.4763",
"0.813",
"0.5024",
"0.813",
"0.5024",
"0.8257",
"0.4763",
"0.8257"
],
[
"0.5337",
"0.813",
"0.5596",
"0.813",
"0.5596",
"0.8247",
"0.5337",
"0.8247"
],
[
"0.5864",
"0.813",
"0.613",
"0.813",
"0.613",
"0.8247",
"0.5864",
"0.8247"
],
[
"0.6396",
"0.813",
"0.6665",
"0.813",
"0.6665",
"0.8257",
"0.6396",
"0.8257"
],
[
"0.697",
"0.813",
"0.7236",
"0.813",
"0.7236",
"0.8247",
"0.697",
"0.8247"
],
[
"0.7534",
"0.813",
"0.7793",
"0.813",
"0.7793",
"0.8257",
"0.7534",
"0.8257"
],
[
"0.8057",
"0.813",
"0.834",
"0.813",
"0.834",
"0.8257",
"0.8057",
"0.8257"
],
[
"0.1765",
"0.827",
"0.528",
"0.827",
"0.528",
"0.8413",
"0.1765",
"0.8413"
],
[
"0.2009",
"0.808",
"0.2443",
"0.808",
"0.2443",
"0.8193",
"0.2009",
"0.8193"
],
[
"0.5293",
"0.8306",
"0.544",
"0.8306",
"0.544",
"0.8384",
"0.5293",
"0.8384"
],
[
"0.5415",
"0.827",
"0.824",
"0.827",
"0.824",
"0.8413",
"0.5415",
"0.8413"
],
[
"0.1732",
"0.8384",
"0.824",
"0.839",
"0.824",
"0.857",
"0.1732",
"0.856"
],
[
"0.1748",
"0.855",
"0.822",
"0.855",
"0.822",
"0.8687",
"0.1748",
"0.8687"
],
[
"0.1724",
"0.866",
"0.824",
"0.863",
"0.824",
"0.884",
"0.1724",
"0.887"
],
[
"0.1772",
"0.882",
"0.8228",
"0.882",
"0.8228",
"0.8965",
"0.1772",
"0.8965"
],
[
"0.1748",
"0.8965",
"0.8228",
"0.8965",
"0.8228",
"0.91",
"0.1748",
"0.91"
],
[
"0.1765",
"0.911",
"0.824",
"0.911",
"0.824",
"0.925",
"0.1765",
"0.925"
],
[
"0.492",
"0.9478",
"0.509",
"0.9478",
"0.509",
"0.962",
"0.492",
"0.962"
]
] | [
"Published at ICLR 2021 Workshop on Security and Safety in Machine Learning Systems",
"Target Label",
"Images",
"Benign Label",
"BadNets",
"Ours",
"DO",
"Figure 1:Comparison between our proposed fine-grained backdoor attack and previous",
"attacks methods ( e.g. , BadNets). In this example, target labels of BadNets ( i.e. , images",
"in the third column) are all the same road scape, while those of our method (i.e. , images in",
"the last column) are sample-specific. Specifically, all pixels with ground-truth class ‘person",
"(those in the white-box of images) are labeled as the class 'palm' in the target labels.",
"AttackwithSemanticTrigger",
"AttackwithNon-semanticTrigger",
"",
"Benign",
"",
"",
"Image",
"",
"",
"",
"Poisoned",
"Image",
"Figure 2: The illustration of poisoned images generated by attacks with the semantic and",
"non-semantic trigger. In this example, the trigger is denoted in the red box where object",
"‘wall' is the semantic trigger and ‘black line' is the non-semantic trigger.",
"Zhuetal.,2020），因为这些方法需要反转每个潜在目标标签的潜在触发模式，并且其计算",
"复杂性在分段任务中难以承受（O(K）用手细粒度攻击，O（KWXH）用于BadNets型攻",
"击）。这将在我们今后的工作中进一步研究",
"3 ",
"Experiment",
"3.1",
"Settings",
"数据集选择和模型结构。我们采用三种最先进的细分模型，包括DeepLabv3（Chenetal.",
"2017a）、DenseASPP（Yang et al.，2018）和 DANet（Fuet al.,2019）进行评估。所有模型",
"都在ADE20K数据集（Zhouetal.,2017）上训练",
"Table l:The effectiveness (% ) of BadNets attack with the non-semantic and semantic",
"trigger.",
"TriggerType",
"Non-semanticTrigger",
"Semantic Trigger",
"Method,Metric",
"ASR",
"mIOU-B",
"ASR",
"mIOU-B",
"Model ",
"PA-BmIOU-A",
"PA-BmIOU-A",
"Benign",
"73.3",
"73.3",
"4.0",
"35.3",
"5.1",
"13.0",
"35.3",
"DeepLabv3",
"8.9",
"74.2",
"78.5",
"96.9",
"33.3",
"69.7",
"35.8",
"63.3",
"34.8",
"BadNets",
"37.3",
"75.5",
"4.0",
"8.9",
"37.3",
"75.5",
"5.2",
"13.4",
"Benign",
"DenseASPP",
"BadNets",
"37.1",
"75.8",
"90.1",
"98.8",
"35.3",
"71.6",
"46.4",
"69.5",
"Benign",
"75.9",
"9.1",
"75.9",
"13.5",
"4.1",
"38.3",
"38.3",
"5.2",
"BadNets",
"38.0",
"76.1",
"79.8",
"97.3",
"36.5",
"71.8",
"42.6",
"70.2",
"攻击设置。我们概括了BadNets（Guetal.,2019)",
"DANet",
"",
"并评估了我们的细粒度攻击的两个重要",
"特殊情况，包括（1）N对1攻击和（2）1对1攻击。对于N对1攻击，具有所有类的像素",
"将被标记为目标标签中的“墙”，而只有具有“人”类的像素将被标记为1对1攻击生成的目",
"标标签中的“手掌”，此外，我们采用8像素宽的黑线作为非语义触发器（e{0,1FexWx",
"）和对象“墙”作为语义触发器。=10%对N对1攻击和的中毒率接近20%对1对",
"1攻击。此外，对于带有语义触发器的细粒度攻击，所有同时包含“墙”和“人”对象的样",
"本都被选择为中毒样本。对于BadNets，目标标签是从数据集中随机选择的“道路景观”图",
"3"
] | {
"arxiv_id": "2103.04038",
"language": "zh",
"page_idx": 2
} |
||
2103.04038-zh-3 | 2103.04038-zh-3.jpg | $\underline{\text{Published at ICLR 2021 Workshop on Security and Safety in Machine Learning Systems}}$
Table 2: The effectiveness $(\%)$ of our attack with the non-semantic and semantic trigger.
<table>
<tbody>
<tr>
<th colspan="2">Attack Type$\to$</th>
<th colspan="3">N-to-1 Attack</th>
<th colspan="5">$1- to- 1$ Attack</th>
</tr>
<tr>
<th>Model 11</th>
<th>$\overline{\mathrm{Method}\downarrow,\mathrm{Metric}\rightarrow}$</th>
<th>mIOU-B</th>
<th>$\overline{\mathrm{PA-B}}$</th>
<th>$ASR$</th>
<th>mIOU-B</th>
<th>$\overline{\mathrm{PA-B}}$</th>
<th>mIOU-A</th>
<th>$\overline{\mathrm{PA-A}}$</th>
<th>$ASR$</th>
</tr>
<tr>
<td rowspan="3">DeepLabv3</td>
<td>$\operatorname{Benign}$</td>
<td>$\overline{35.3}$</td>
<td>73.3</td>
<td>$\overline{14.8}$</td>
<td>$\overline{35.3}$</td>
<td>$\overline{73.3}$</td>
<td>$\overline{26.0}$</td>
<td>$\overline{67.5}$</td>
<td>$\overline{0}$</td>
</tr>
<tr>
<td>FGBA non- semantic)</td>
<td>34.9</td>
<td>74.1</td>
<td>99.8</td>
<td>36.0</td>
<td>73.6</td>
<td>26.2</td>
<td>74.2</td>
<td>77.9</td>
</tr>
<tr>
<td>FGBA semantic)</td>
<td>32.1</td>
<td>66.2</td>
<td>62.9</td>
<td>35.2</td>
<td>71.9</td>
<td>24.3</td>
<td>68.0</td>
<td>71.9</td>
</tr>
<tr>
<td rowspan="3">DenseASPP</td>
<td>Benign</td>
<td>37.3</td>
<td>75.5</td>
<td>16.8</td>
<td>37.3</td>
<td>$\overline{75.5}$</td>
<td>27.0</td>
<td>68.6</td>
<td>0</td>
</tr>
<tr>
<td>FGBA non- semantic)</td>
<td>36.7</td>
<td>75.2</td>
<td>99.6</td>
<td>37.3</td>
<td>75.4</td>
<td>27.8</td>
<td>76.0</td>
<td>83.2</td>
</tr>
<tr>
<td>FGBA semantic)</td>
<td>33.9</td>
<td>69.4</td>
<td>50.9</td>
<td>36.8</td>
<td>74.0</td>
<td>26.2</td>
<td>70.8</td>
<td>76.1</td>
</tr>
<tr>
<td rowspan="3">DANet</td>
<td>Benign</td>
<td>$\overline{38.3}$</td>
<td>$\overline{75.9}$</td>
<td>$\overline{15.3}$</td>
<td>$\overline{38.3}$</td>
<td>$\overline{75.9}$</td>
<td>$\overline{28.3}$</td>
<td>$\overline{69.0}$</td>
<td>$\overline{0}$</td>
</tr>
<tr>
<td>FGBA non- semantic)</td>
<td>37.6</td>
<td>75.1</td>
<td>99.2</td>
<td>39.2</td>
<td>76.1</td>
<td>28.5</td>
<td>76.8</td>
<td>81.9</td>
</tr>
<tr>
<td>FGBA semantic)</td>
<td>34.9</td>
<td>68.5</td>
<td>71.6</td>
<td>37.7</td>
<td>74.6</td>
<td>25.2</td>
<td>70.0</td>
<td>75.8</td>
</tr>
</tbody>
</table>
像之一，我们采用对象“草”作为语义触发器(在这种情况下，$\gamma$接近 10%)。我们基于单个 NVIDIAGeForceRTX2080TiGPU 上的开源代码$^{1}$训练批量大小为 4 的所有模型。我们还提供了在良性训练集 (称为“良性”)上训练的模型的结果，以进行比较
评估度量。我们采用了五个指标，包括(1)良性平均交叉联合(mIOU-B),(2)良性像素精度 (PA-B), (3) 攻击平均交叉联合 (mIOU-A), (4)攻击像素精度 (PA-A)和 (5) 攻击成功率 (ASR),以评估不同方法的性能。mIOU-B 和 PA-B 是根据良性测试样本计算得出的，这些样本表明了标准方案中的模型性能。相比之下，mIOU-A、PA-A 和 ASR 是基于攻击样本获得的。特别是，ASR 是由被攻击样本中标签不是其真实对象 (而不是所有对象，如 mIOU-A 和 PA-A) 的对象的像素精度定义的。更具体地说，mIOU-A 和 PA-A 可用于估计预测受攻击样本的整体性能，而 ASR 可以更好地估计实现恶意目的的能力。请注意，mIOU-A, PA-A 和 ASR 对于我们的 N 对 1 攻击是相同的，而 PA-A 和 ASR 对于BadNets 是相同的，我们仅在这些情况下报告 ASR
# 3.2 Main Results
如表 1- 2 所示 , 无论采用哪种类型的触发器，BadNets 和我们提出的细粒度攻击都可以成功攻击所有语义分割模型，同时在预测良性样本方面保持良好的性能。例如，在所有情况下，与在 BadNets 的良性数据集上训练的模型相比，ASR 超过 60%,而 mIOU-B 降级小于 2%; 特别是对于采用非语义触发时的细粒度攻击，所有情况下 ASR 均超过 75%(大多数超过 80%),而 mIOU-B 降级在 1MAXX
## 4 Conclusion
在本文中，我们揭示了现有的针对分类任务的攻击范式可以扩展为攻击语义分割模型。但是，这种方法只能从图像级别操纵预测，因此无法实现精细的恶意操纵。为了解决这个问题，我们探索了一种新颖的攻击范例，即细粒度攻击，我们从对象级别而不是图像级别处理目标标签$(i.e.$,注释)。实验验证了我们的方法在攻击语义分割模型时既有效又隐蔽
$^{1}$https://github.com/Tramac/awesome-semantic-segmentation-pytorch
4 | [
[
"0.1748",
"0.0341",
"0.8096",
"0.0341",
"0.8096",
"0.04736",
"0.1748",
"0.04736"
],
[
"0.1748",
"0.113",
"0.826",
"0.11554",
"0.826",
"0.13",
"0.1748",
"0.1276"
],
[
"0.2467",
"0.1256",
"0.335",
"0.1276",
"0.335",
"0.1427",
"0.2467",
"0.1401"
],
[
"0.4436",
"0.1288",
"0.5215",
"0.1288",
"0.5215",
"0.1401",
"0.4436",
"0.1401"
],
[
"0.661",
"0.1288",
"0.732",
"0.1288",
"0.732",
"0.1401",
"0.661",
"0.1401"
],
[
"0.1953",
"0.1383",
"0.2484",
"0.1401",
"0.2484",
"0.1515",
"0.1953",
"0.1497"
],
[
"0.2695",
"0.1395",
"0.38",
"0.1395",
"0.38",
"0.1528",
"0.2695",
"0.1528"
],
[
"0.4084",
"0.1395",
"0.468",
"0.1395",
"0.468",
"0.1509",
"0.4084",
"0.1509"
],
[
"0.511",
"0.1395",
"0.5557",
"0.1395",
"0.5557",
"0.1509",
"0.511",
"0.1509"
],
[
"0.5664",
"0.1395",
"0.619",
"0.1395",
"0.619",
"0.1509",
"0.5664",
"0.1509"
],
[
"0.6323",
"0.1401",
"0.732",
"0.1401",
"0.732",
"0.1515",
"0.6323",
"0.1515"
],
[
"0.744",
"0.1401",
"0.7837",
"0.1401",
"0.7837",
"0.1509",
"0.744",
"0.1509"
],
[
"0.79",
"0.1401",
"0.826",
"0.1401",
"0.826",
"0.1515",
"0.79",
"0.1515"
],
[
"0.4248",
"0.1483",
"0.452",
"0.1509",
"0.451",
"0.1635",
"0.4224",
"0.1604"
],
[
"0.4756",
"0.1401",
"0.517",
"0.1401",
"0.517",
"0.1509",
"0.4756",
"0.1509"
],
[
"0.6914",
"0.1483",
"0.7197",
"0.1509",
"0.7183",
"0.1635",
"0.6895",
"0.161"
],
[
"0.7485",
"0.1483",
"0.7754",
"0.1509",
"0.774",
"0.1635",
"0.746",
"0.1604"
],
[
"0.8037",
"0.1483",
"0.8145",
"0.154",
"0.806",
"0.1642",
"0.794",
"0.1578"
],
[
"0.482",
"0.1509",
"0.509",
"0.1509",
"0.509",
"0.1622",
"0.482",
"0.1622"
],
[
"0.3145",
"0.1509",
"0.3555",
"0.1509",
"0.3555",
"0.1622",
"0.3145",
"0.1622"
],
[
"0.527",
"0.1509",
"0.5513",
"0.1509",
"0.5513",
"0.1622",
"0.527",
"0.1622"
],
[
"0.58",
"0.1509",
"0.6055",
"0.1509",
"0.6055",
"0.1635",
"0.58",
"0.1635"
],
[
"0.6357",
"0.1509",
"0.6626",
"0.1509",
"0.6626",
"0.1622",
"0.6357",
"0.1622"
],
[
"0.2656",
"0.159",
"0.3987",
"0.1604",
"0.3987",
"0.1749",
"0.2656",
"0.1736"
],
[
"0.4248",
"0.161",
"0.452",
"0.161",
"0.452",
"0.1736",
"0.4248",
"0.1736"
],
[
"0.482",
"0.159",
"0.5073",
"0.1622",
"0.5063",
"0.1736",
"0.4795",
"0.1704"
],
[
"0.524",
"0.161",
"0.5513",
"0.161",
"0.5513",
"0.1724",
"0.524",
"0.1724"
],
[
"0.58",
"0.161",
"0.607",
"0.161",
"0.607",
"0.1736",
"0.58",
"0.1736"
],
[
"0.6357",
"0.161",
"0.6626",
"0.161",
"0.6626",
"0.1724",
"0.6357",
"0.1724"
],
[
"0.6914",
"0.161",
"0.7197",
"0.161",
"0.7197",
"0.1736",
"0.6914",
"0.1736"
],
[
"0.7485",
"0.161",
"0.777",
"0.161",
"0.777",
"0.1724",
"0.7485",
"0.1724"
],
[
"0.7935",
"0.161",
"0.8203",
"0.161",
"0.8203",
"0.1736",
"0.7935",
"0.1736"
],
[
"0.523",
"0.1704",
"0.5522",
"0.1724",
"0.5513",
"0.1849",
"0.5215",
"0.1831"
],
[
"0.1896",
"0.1622",
"0.2566",
"0.1622",
"0.2566",
"0.1736",
"0.1896",
"0.1736"
],
[
"0.4248",
"0.1718",
"0.451",
"0.1718",
"0.451",
"0.1843",
"0.4248",
"0.1843"
],
[
"0.482",
"0.1718",
"0.509",
"0.1718",
"0.509",
"0.1843",
"0.482",
"0.1843"
],
[
"0.58",
"0.1718",
"0.607",
"0.1718",
"0.607",
"0.1843",
"0.58",
"0.1843"
],
[
"0.6357",
"0.1718",
"0.6626",
"0.1718",
"0.6626",
"0.1843",
"0.6357",
"0.1843"
],
[
"0.6914",
"0.1718",
"0.7197",
"0.1718",
"0.7197",
"0.1843",
"0.6914",
"0.1843"
],
[
"0.7485",
"0.1718",
"0.777",
"0.1718",
"0.777",
"0.1843",
"0.7485",
"0.1843"
],
[
"0.7935",
"0.1724",
"0.8228",
"0.1724",
"0.8228",
"0.1843",
"0.7935",
"0.1843"
],
[
"0.3145",
"0.1843",
"0.3555",
"0.1843",
"0.3555",
"0.1957",
"0.3145",
"0.1957"
],
[
"0.2834",
"0.1724",
"0.3857",
"0.1724",
"0.3857",
"0.1843",
"0.2834",
"0.1843"
],
[
"0.4248",
"0.1843",
"0.452",
"0.1843",
"0.452",
"0.1963",
"0.4248",
"0.1963"
],
[
"0.482",
"0.1843",
"0.5073",
"0.1843",
"0.5073",
"0.1963",
"0.482",
"0.1963"
],
[
"0.524",
"0.1843",
"0.5522",
"0.1843",
"0.5522",
"0.1957",
"0.524",
"0.1957"
],
[
"0.58",
"0.1843",
"0.607",
"0.1843",
"0.607",
"0.1963",
"0.58",
"0.1963"
],
[
"0.6357",
"0.1843",
"0.664",
"0.1843",
"0.664",
"0.1989",
"0.6357",
"0.1989"
],
[
"0.6914",
"0.1843",
"0.7183",
"0.1843",
"0.7183",
"0.1963",
"0.6914",
"0.1963"
],
[
"0.7485",
"0.1843",
"0.7754",
"0.1843",
"0.7754",
"0.1957",
"0.7485",
"0.1957"
],
[
"0.2656",
"0.1913",
"0.3987",
"0.1925",
"0.3987",
"0.207",
"0.2656",
"0.2058"
],
[
"0.523",
"0.1913",
"0.5522",
"0.1932",
"0.5513",
"0.2058",
"0.5215",
"0.2039"
],
[
"0.796",
"0.1863",
"0.8105",
"0.1799",
"0.819",
"0.1932",
"0.8037",
"0.1995"
],
[
"0.4248",
"0.1932",
"0.451",
"0.1932",
"0.451",
"0.2058",
"0.4248",
"0.2058"
],
[
"0.482",
"0.1932",
"0.5073",
"0.1932",
"0.5073",
"0.2058",
"0.482",
"0.2058"
],
[
"0.58",
"0.1925",
"0.607",
"0.1957",
"0.6045",
"0.207",
"0.5767",
"0.2039"
],
[
"0.6357",
"0.1932",
"0.6626",
"0.1932",
"0.6626",
"0.2058",
"0.6357",
"0.2058"
],
[
"0.6914",
"0.1932",
"0.7183",
"0.1932",
"0.7183",
"0.2058",
"0.6914",
"0.2058"
],
[
"0.7485",
"0.1932",
"0.7754",
"0.1932",
"0.7754",
"0.2052",
"0.7485",
"0.2052"
],
[
"0.7935",
"0.1945",
"0.822",
"0.1945",
"0.822",
"0.2058",
"0.7935",
"0.2058"
],
[
"0.4248",
"0.2052",
"0.452",
"0.2052",
"0.452",
"0.2172",
"0.4248",
"0.2172"
],
[
"0.1887",
"0.1945",
"0.2573",
"0.1945",
"0.2573",
"0.2058",
"0.1887",
"0.2058"
],
[
"0.482",
"0.2052",
"0.5073",
"0.2052",
"0.5073",
"0.2172",
"0.482",
"0.2172"
],
[
"0.523",
"0.2052",
"0.5522",
"0.2052",
"0.5522",
"0.2172",
"0.523",
"0.2172"
],
[
"0.58",
"0.2052",
"0.6055",
"0.2052",
"0.6055",
"0.2172",
"0.58",
"0.2172"
],
[
"0.6357",
"0.2052",
"0.6626",
"0.2052",
"0.6626",
"0.2172",
"0.6357",
"0.2172"
],
[
"0.6914",
"0.2052",
"0.7183",
"0.2052",
"0.7183",
"0.2172",
"0.6914",
"0.2172"
],
[
"0.7485",
"0.2052",
"0.777",
"0.2052",
"0.777",
"0.2172",
"0.7485",
"0.2172"
],
[
"0.7935",
"0.2052",
"0.822",
"0.2052",
"0.822",
"0.2172",
"0.7935",
"0.2172"
],
[
"0.482",
"0.2134",
"0.5073",
"0.2166",
"0.5063",
"0.2279",
"0.4795",
"0.2247"
],
[
"0.2834",
"0.2058",
"0.3857",
"0.2058",
"0.3857",
"0.2172",
"0.2834",
"0.2172"
],
[
"0.3162",
"0.214",
"0.3545",
"0.2166",
"0.353",
"0.2291",
"0.3145",
"0.2267"
],
[
"0.4248",
"0.2153",
"0.451",
"0.2153",
"0.451",
"0.2279",
"0.4248",
"0.2279"
],
[
"0.523",
"0.2153",
"0.5522",
"0.2153",
"0.5522",
"0.2279",
"0.523",
"0.2279"
],
[
"0.58",
"0.2153",
"0.607",
"0.2153",
"0.607",
"0.2279",
"0.58",
"0.2279"
],
[
"0.6357",
"0.2153",
"0.6626",
"0.2153",
"0.6626",
"0.2279",
"0.6357",
"0.2279"
],
[
"0.6914",
"0.2153",
"0.7183",
"0.2153",
"0.7183",
"0.2279",
"0.6914",
"0.2279"
],
[
"0.7485",
"0.2153",
"0.7754",
"0.2153",
"0.7754",
"0.2279",
"0.7485",
"0.2279"
],
[
"0.8",
"0.2166",
"0.814",
"0.2166",
"0.814",
"0.2267",
"0.8",
"0.2267"
],
[
"0.2695",
"0.226",
"0.3987",
"0.226",
"0.3987",
"0.2393",
"0.2695",
"0.2393"
],
[
"0.482",
"0.2247",
"0.5073",
"0.2279",
"0.5063",
"0.2393",
"0.4805",
"0.2361"
],
[
"0.58",
"0.226",
"0.607",
"0.226",
"0.607",
"0.238",
"0.58",
"0.238"
],
[
"0.6357",
"0.226",
"0.661",
"0.226",
"0.661",
"0.238",
"0.6357",
"0.238"
],
[
"0.6914",
"0.226",
"0.7183",
"0.226",
"0.7183",
"0.238",
"0.6914",
"0.238"
],
[
"0.7485",
"0.226",
"0.777",
"0.226",
"0.777",
"0.238",
"0.7485",
"0.238"
],
[
"0.4248",
"0.2267",
"0.452",
"0.2267",
"0.452",
"0.238",
"0.4248",
"0.238"
],
[
"0.2009",
"0.2267",
"0.2427",
"0.2267",
"0.2427",
"0.238",
"0.2009",
"0.238"
],
[
"0.524",
"0.2267",
"0.5522",
"0.2267",
"0.5522",
"0.238",
"0.524",
"0.238"
],
[
"0.7935",
"0.2267",
"0.822",
"0.2267",
"0.822",
"0.238",
"0.7935",
"0.238"
],
[
"0.4248",
"0.2374",
"0.451",
"0.2374",
"0.451",
"0.25",
"0.4248",
"0.25"
],
[
"0.482",
"0.2374",
"0.509",
"0.2374",
"0.509",
"0.2488",
"0.482",
"0.2488"
],
[
"0.5186",
"0.2374",
"0.5522",
"0.2374",
"0.5522",
"0.2488",
"0.5186",
"0.2488"
],
[
"0.58",
"0.2349",
"0.607",
"0.238",
"0.6045",
"0.25",
"0.5767",
"0.2468"
],
[
"0.6357",
"0.2374",
"0.6626",
"0.2374",
"0.6626",
"0.2488",
"0.6357",
"0.2488"
],
[
"0.6914",
"0.2374",
"0.7183",
"0.2374",
"0.7183",
"0.25",
"0.6914",
"0.25"
],
[
"0.749",
"0.2374",
"0.7754",
"0.2374",
"0.7754",
"0.2488",
"0.749",
"0.2488"
],
[
"0.7935",
"0.2374",
"0.824",
"0.2374",
"0.824",
"0.25",
"0.7935",
"0.25"
],
[
"0.1765",
"0.2676",
"0.8228",
"0.2676",
"0.8228",
"0.281",
"0.1765",
"0.281"
],
[
"0.2852",
"0.238",
"0.3857",
"0.238",
"0.3857",
"0.25",
"0.2852",
"0.25"
],
[
"0.1765",
"0.2798",
"0.824",
"0.2798",
"0.824",
"0.2942",
"0.1765",
"0.2942"
],
[
"0.1748",
"0.2954",
"0.706",
"0.2954",
"0.706",
"0.3093",
"0.1748",
"0.3093"
],
[
"0.1765",
"0.3145",
"0.822",
"0.315",
"0.822",
"0.33",
"0.1765",
"0.3289"
],
[
"0.1765",
"0.33",
"0.8228",
"0.33",
"0.8228",
"0.3435",
"0.1765",
"0.3435"
],
[
"0.1748",
"0.3435",
"0.8228",
"0.3435",
"0.8228",
"0.3567",
"0.1748",
"0.3567"
],
[
"0.1765",
"0.358",
"0.8228",
"0.358",
"0.8228",
"0.3718",
"0.1765",
"0.3718"
],
[
"0.1765",
"0.3718",
"0.8228",
"0.3718",
"0.8228",
"0.385",
"0.1765",
"0.385"
],
[
"0.1765",
"0.385",
"0.8228",
"0.385",
"0.8228",
"0.3984",
"0.1765",
"0.3984"
],
[
"0.1772",
"0.3997",
"0.8228",
"0.3997",
"0.8228",
"0.4136",
"0.1772",
"0.4136"
],
[
"0.1765",
"0.4136",
"0.8228",
"0.4136",
"0.8228",
"0.4268",
"0.1765",
"0.4268"
],
[
"0.1748",
"0.4268",
"0.5435",
"0.4268",
"0.5435",
"0.44",
"0.1748",
"0.44"
],
[
"0.1748",
"0.4583",
"0.3145",
"0.4583",
"0.3145",
"0.4717",
"0.1748",
"0.4717"
],
[
"0.1765",
"0.4841",
"0.8228",
"0.4841",
"0.8228",
"0.4976",
"0.1765",
"0.4976"
],
[
"0.1748",
"0.4976",
"0.822",
"0.4976",
"0.822",
"0.511",
"0.1748",
"0.511"
],
[
"0.1765",
"0.511",
"0.822",
"0.511",
"0.822",
"0.526",
"0.1765",
"0.526"
],
[
"0.1765",
"0.526",
"0.8228",
"0.526",
"0.8228",
"0.539",
"0.1765",
"0.539"
],
[
"0.1765",
"0.539",
"0.4885",
"0.539",
"0.4885",
"0.553",
"0.1765",
"0.553"
],
[
"0.1748",
"0.5747",
"0.2141",
"0.5747",
"0.2141",
"0.586",
"0.1748",
"0.586"
],
[
"0.2091",
"0.5757",
"0.3022",
"0.5757",
"0.3022",
"0.587",
"0.2091",
"0.587"
],
[
"0.1772",
"0.606",
"0.822",
"0.606",
"0.822",
"0.616",
"0.1772",
"0.616"
],
[
"0.1772",
"0.6187",
"0.822",
"0.6187",
"0.822",
"0.632",
"0.1772",
"0.632"
],
[
"0.1765",
"0.632",
"0.8228",
"0.632",
"0.8228",
"0.6465",
"0.1765",
"0.6465"
],
[
"0.1765",
"0.646",
"0.7812",
"0.646",
"0.7812",
"0.659",
"0.1765",
"0.659"
],
[
"0.1936",
"0.91",
"0.6685",
"0.911",
"0.6685",
"0.9253",
"0.1936",
"0.925"
],
[
"0.4927",
"0.9497",
"0.5063",
"0.9497",
"0.5063",
"0.9604",
"0.4927",
"0.9604"
]
] | [
"Published at ICLR 2021 Workshop on Security and Safety in Machine Learning Systems",
"Table 2: The effectiveness (% ) of our attack with the non-semantic and semantic trigger.",
"AttackType-",
"N-to-1 Attack",
"-to-1 Attack",
"Model ↓",
"Method , Metric",
"mIOU-B",
"TASR",
"mIOU-B",
"PA-BmIOU-A",
"PA-A",
"ASR",
"35.3",
"PA-B",
"26.0",
"67.5",
"0",
"73.3",
"Benign",
"14.8",
"35.3",
"73.3",
"FGBA (non-semantic)",
"34.9",
"74.1",
"99.8",
"36.0",
"73.6",
"26.2",
"74.2",
"77.9",
"62.9",
"DeepLabv3",
"32.1",
"66.2",
"35.2",
"71.9",
"24.3",
"68.0",
"71.9",
"Benign",
"FGBA (semantic)",
"37.3",
"75.5",
"16.8",
"37.3",
"75.5",
"27.0",
"68.6",
"FGBA (non-semantic)",
"99.6",
"0",
"36.7",
"75.2",
"37.3",
"75.4",
"27.8",
"76.0",
"83.2",
"33.9",
"DenseASPP",
"69.4",
"50.9",
"36.8",
"74.0",
"26.2",
"70.8",
"76.1",
"75.9",
"FGBA (semantic)",
"Benign",
"38.3",
"15.3",
"38.3",
"75.9",
"28.3",
"69.0",
"0",
"FGBA (non-semantic)",
"75.1",
"39.2",
"76.1",
"28.5",
"76.8",
"37.6",
"DANet",
"99.2",
"81.9",
"34.9",
"68.5",
"71.6",
"37.7",
"74.6",
"25.2",
"70.0",
"75.8",
"像之一，我们采用对象“草”作为语义触发器（在这种情况下，接近10%）。我们基于单",
"FGBA (semantic)",
"个NVIDIAGeForceRTX2080TiGPU上的开源代码1训练批量大小为4的所有模型。我们",
"还提供了在良性训练集（称为“良性”）上训练的模型的结果，以进行比较",
"评估度量。我们采用了五个指标，包括（1）良性平均交叉联合（mIOU-B），（2）良性像",
"素精度（PA-B），（3）攻击平均交叉联合（mIOU-A），（4）攻击像素精度（PA-A）和（5）",
"攻击成功率（ASR），以评估不同方法的性能。mIOU-B和PA-B是根据良性测试样本计算",
"得出的，这些样本表明了标准方案中的模型性能。相比之下，mIOU-A、PA-A和ASR是",
"基于攻击样本获得的。特别是，ASR是由被攻击样本中标签不是其真实对象（而不是所有",
"对象，如mIOU-A和PA-A）的对象的像素精度定义的。更具体地说，mIOU-A和PA-A",
"可用于估计预测受攻击样本的整体性能，而ASR可以更好地估计实现恶意目的的能力。请",
"注意，mIOU-A，PA-A和ASR对于我们的N对1攻击是相同的，而PA-A和ASR对于",
"BadNets是相同的，我们仅在这些情况下报告ASR",
"3.2Main Results",
"如表1-2所示，无论采用哪种类型的触发器，BadNets和我们提出的细粒度攻击都可以成",
"功攻击所有语义分割模型，同时在预测良性样本方面保持良好的性能。例如，在所有情况",
"下，与在BadNets的良性数据集上训练的模型相比，ASR超过60%，而mIOU-B降级小",
"于2%；特别是对于采用非语义触发时的细粒度攻击，所有情况下ASR均超过75%（大多",
"数超过80%），而mIOU-B降级在1MAXX",
"4",
"Conclusion",
"在本文中，我们揭示了现有的针对分类任务的攻击范式可以扩展为攻击语义分割模型。但",
"是，这种方法只能从图像级别操纵预测，因此无法实现精细的恶意操纵。为了解决这个问",
"题，我们探索了一种新颖的攻击范例，即细粒度攻击，我们从对象级别而不是图像级别处理",
"目标标签（i.e.，注释）。实验验证了我们的方法在攻击语义分割模型时既有效又隐蔽",
"'https://github.com/Tramac/awesome-semantic-segmentation-pytorch",
"4"
] | {
"arxiv_id": "2103.04038",
"language": "zh",
"page_idx": 3
} |
||
2103.04038-zh-4 | 2103.04038-zh-4.jpg | $\underline{\text{Published at ICLR 2021 Workshop on Security and Safety in Machine Learning Systems}}$
## References
Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov.
How to backdoor federated learning. In AISTATS, 2020.
Linchao Bao, Baoyuan Wu, and Wei Liu. Cnn in mrf: Video object segmentation via
inference in a cnn-based higher-order spatio-temporal mrf. In CVPR, 2018.
Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017a.
Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017b. Junsuk Choe, Seungho Lee, and Hyunjung Shim. Attention-based dropout layer for weakly
$\begin{array}{c}\mathrm{supervised~single~object~localization~and~semantic~segmentation.~IEEE~Transactions~on}\\\mathrm{supervised~single~object~localization~and~semantic~segmentation.~IEEEE~Transactions~on}\end{array}$
Pattern Analysis and Machine Intelligence, 2020.
Di Feng, Christian Haase-Schuetz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaeser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 2020
Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual
attention network for scene segmentation. In CVPR, 2019.
Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating
backdooring attacks on deep neural networks. IEEE Access, 7:47230-47244, 2019.
Wenbo Guo, Lun Wang, Yan Xu, Xinyu Xing, Min Du, and Dawn Song. Towards inspecting
and eliminating trojan backdoors in deep neural networks. In ICDM, 2020.
Lei Han, Tian Zheng, Yinheng Zhu, Lan Xu, and Lu Fang. Live semantic 3d perception for $\begin{array}{c}\mathrm{immersive~augmented~reality.~IEEE~transactions~on~visualization~and~computer~graphics,}\end{array}$ 26(5){:}2012{-}2022, 2020.
Liu. Ccnet: Criss-cross attention for semantic segmentation. In ICCV, 2019.
Yiming Li, Baoyuan Wu, Yong Jiang, Zhifeng Li, and Shu-Tao Xia. Backdoor learning: A
survey. arXiv preprint arXiv:2007.08745, 2020a.
Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Backdoor attack
with sample-specific triggers. arXiv preprint arXiv:2012.03816, 2020b.
Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural
backdoor attack on deep neural networks. In ECCV, 2020.
Anh Tuan Nguyen and Tuan Anh Tran. Wanet- imperceptible warping-based backdoor
attack. In ICLR, 2021.
Mennatullah Siam, Mostafa Gamal, Moemen Abdel-Razek, Senthil Yogamani, Martin Jager-
sand, and Hong Zhang. A comparative study of real-time semantic segmentation for
autonomous driving. In CVPR Workshop, 2018.
Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural cleanse: Identifying and mitigating backdoor attacks in neural networks. In IEEE S&P, 2019.
Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic
segmentation in street scenes. In CVPR, 2018.
Yuanshun Yao, Huiying Li, Haitao Zheng, and Ben Y Zhao. Latent backdoor attacks on
deep neural networks. In CCS, 2019.
5 | [
[
"0.1732",
"0.0322",
"0.812",
"0.03284",
"0.812",
"0.04736",
"0.1732",
"0.04672"
],
[
"0.1732",
"0.1016",
"0.2656",
"0.1042",
"0.2656",
"0.1193",
"0.1732",
"0.11743"
],
[
"0.1732",
"0.1256",
"0.8228",
"0.1256",
"0.8228",
"0.1401",
"0.1732",
"0.1401"
],
[
"0.1887",
"0.1401",
"0.5986",
"0.1395",
"0.5986",
"0.154",
"0.1887",
"0.1547"
],
[
"0.1748",
"0.1635",
"0.826",
"0.1635",
"0.826",
"0.178",
"0.1748",
"0.178"
],
[
"0.1896",
"0.178",
"0.7295",
"0.178",
"0.7295",
"0.1913",
"0.1896",
"0.1913"
],
[
"0.1748",
"0.202",
"0.8228",
"0.202",
"0.8228",
"0.2153",
"0.1748",
"0.2153"
],
[
"0.1912",
"0.2166",
"0.8228",
"0.2166",
"0.8228",
"0.2299",
"0.1912",
"0.2299"
],
[
"0.1912",
"0.2311",
"0.2402",
"0.2311",
"0.2402",
"0.2424",
"0.1912",
"0.2424"
],
[
"0.1748",
"0.2524",
"0.8228",
"0.2524",
"0.8228",
"0.2664",
"0.1748",
"0.2664"
],
[
"0.1896",
"0.2676",
"0.822",
"0.2676",
"0.822",
"0.281",
"0.1896",
"0.281"
],
[
"0.1748",
"0.2903",
"0.822",
"0.2903",
"0.822",
"0.3037",
"0.1748",
"0.3037"
],
[
"0.1912",
"0.305",
"0.8228",
"0.305",
"0.8228",
"0.318",
"0.1912",
"0.318"
],
[
"0.1896",
"0.3164",
"0.544",
"0.3176",
"0.544",
"0.332",
"0.1896",
"0.3308"
],
[
"0.1732",
"0.3403",
"0.8203",
"0.3403",
"0.8203",
"0.3542",
"0.1732",
"0.3542"
],
[
"0.1887",
"0.3542",
"0.8203",
"0.3542",
"0.8203",
"0.3687",
"0.1887",
"0.3687"
],
[
"0.1912",
"0.3694",
"0.822",
"0.3694",
"0.822",
"0.3833",
"0.1912",
"0.3833"
],
[
"0.1896",
"0.3833",
"0.705",
"0.3833",
"0.705",
"0.3965",
"0.1896",
"0.3965"
],
[
"0.1748",
"0.406",
"0.826",
"0.406",
"0.826",
"0.4204",
"0.1748",
"0.4204"
],
[
"0.1896",
"0.4204",
"0.615",
"0.4204",
"0.615",
"0.4343",
"0.1896",
"0.4343"
],
[
"0.1732",
"0.4426",
"0.824",
"0.443",
"0.824",
"0.4583",
"0.1732",
"0.457"
],
[
"0.1896",
"0.4583",
"0.7754",
"0.4583",
"0.7754",
"0.4717",
"0.1896",
"0.4717"
],
[
"0.1732",
"0.4797",
"0.824",
"0.481",
"0.824",
"0.4956",
"0.1732",
"0.4944"
],
[
"0.1896",
"0.4956",
"0.728",
"0.4956",
"0.728",
"0.509",
"0.1896",
"0.509"
],
[
"0.1732",
"0.518",
"0.8228",
"0.518",
"0.8228",
"0.532",
"0.1732",
"0.532"
],
[
"0.1929",
"0.5327",
"0.8228",
"0.5327",
"0.8228",
"0.547",
"0.1929",
"0.547"
],
[
"0.1912",
"0.547",
"0.357",
"0.547",
"0.357",
"0.56",
"0.1912",
"0.56"
],
[
"0.1732",
"0.5684",
"0.824",
"0.5693",
"0.824",
"0.584",
"0.1732",
"0.5835"
],
[
"0.1896",
"0.584",
"0.739",
"0.584",
"0.739",
"0.5977",
"0.1896",
"0.5977"
],
[
"0.1748",
"0.607",
"0.8228",
"0.607",
"0.8228",
"0.6206",
"0.1748",
"0.6206"
],
[
"0.1887",
"0.6206",
"0.535",
"0.6206",
"0.535",
"0.634",
"0.1887",
"0.634"
],
[
"0.1732",
"0.643",
"0.8228",
"0.643",
"0.8228",
"0.657",
"0.1732",
"0.657"
],
[
"0.1896",
"0.659",
"0.694",
"0.659",
"0.694",
"0.6724",
"0.1896",
"0.6724"
],
[
"0.1748",
"0.6816",
"0.824",
"0.6816",
"0.824",
"0.696",
"0.1748",
"0.696"
],
[
"0.1887",
"0.6963",
"0.613",
"0.6963",
"0.613",
"0.7104",
"0.1887",
"0.7104"
],
[
"0.1748",
"0.7183",
"0.8228",
"0.7183",
"0.8228",
"0.733",
"0.1748",
"0.733"
],
[
"0.1896",
"0.7344",
"0.357",
"0.7344",
"0.357",
"0.7476",
"0.1896",
"0.7476"
],
[
"0.1732",
"0.755",
"0.8203",
"0.756",
"0.8203",
"0.771",
"0.1732",
"0.7695"
],
[
"0.1896",
"0.7715",
"0.8228",
"0.7715",
"0.8228",
"0.785",
"0.1896",
"0.785"
],
[
"0.1912",
"0.784",
"0.54",
"0.784",
"0.54",
"0.799",
"0.1912",
"0.799"
],
[
"0.1748",
"0.8066",
"0.8203",
"0.8066",
"0.8203",
"0.8213",
"0.1748",
"0.8213"
],
[
"0.1887",
"0.821",
"0.8228",
"0.8213",
"0.8228",
"0.836",
"0.1887",
"0.8354"
],
[
"0.1896",
"0.8354",
"0.4182",
"0.8354",
"0.4182",
"0.8486",
"0.1896",
"0.8486"
],
[
"0.1732",
"0.858",
"0.8228",
"0.8594",
"0.8228",
"0.8735",
"0.1732",
"0.8726"
],
[
"0.1887",
"0.8726",
"0.527",
"0.872",
"0.527",
"0.886",
"0.1887",
"0.887"
],
[
"0.1724",
"0.893",
"0.826",
"0.8945",
"0.826",
"0.912",
"0.1724",
"0.911"
],
[
"0.1896",
"0.911",
"0.4558",
"0.911",
"0.4558",
"0.925",
"0.1896",
"0.925"
],
[
"0.4902",
"0.9478",
"0.509",
"0.9478",
"0.509",
"0.962",
"0.4902",
"0.962"
]
] | [
"Published at ICLR 2021 Workshop on Security and Safety in Machine Learning Systems",
"References",
"Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov",
"How to backdoor federated learning. In AISTATS, 2020.",
"Linchao Bao, Baoyuan Wu, and Wei Liu. Cnn in mrf: Video object segmentation via",
"inference in a cnn-based higher-order spatio-temporal mrf. In CVPR, 2018.",
"Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking",
"atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587,",
"2017a.",
"Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks",
"on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017b.",
"Junsuk Choe, Seungho Lee, and Hyunjung Shim. Attention-based dropout layer for weakly",
"supervised single object localization and semantic segmentation. IEEE Transactions on",
"Pattern Analysis and Machine Intelligence, 2020.",
"Di Feng, Christian Haase-Schuetz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaeser",
"Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object de",
"tection and semantic segmentation for autonomous driving: Datasets, methods, and chal",
"lenges. IEEE Transactions on Intelligent Transportation Systems, 2020.",
"Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual",
"attention network for scene segmentation. In CVPR, 2019.",
"Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating",
"backdooring attacks on deep neural networks. IEEE Access, 7:47230-47244, 2019.",
"Wenbo Guo, Lun Wang, Yan Xu, Xinyu Xing, Min Du, and Dawn Song. Towards inspecting",
"and eliminating trojan backdoors in deep neural networks. In ICDM, 2020.",
"Lei Han, Tian Zheng, Yinheng Zhu, Lan Xu, and Lu Fang. Live semantic 3d perception for",
"immersive augmented reality. IEEE transactions on visualization and computer graphics,",
"26(5):2012-2022,2020.",
"Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu",
"Liu. Ccnet: Criss-cross attention for semantic segmentation. In ICCV, 2019.",
"Yiming Li, Baoyuan Wu, Yong Jiang, Zhifeng Li, and Shu-Tao Xia. Backdoor learning: A",
"survey. arXiv preprint arXiv:2007.08745, 2020a.",
"Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Backdoor attack",
"with sample-specific triggers. arXiv preprint arXiv:2012.03816, 2020b.",
"Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural",
"backdoor attack on deep neural networks. In ECCV, 2020.",
"Anh Tuan Nguyen and Tuan Anh Tran. Wanet - imperceptible warping-based backdoor",
"attack. In ICLR, 2021.",
"Mennatullah Siam, Mostafa Gamal, Moemen Abdel-Razek, Senthil Yogamani, Martin Jager",
"sand, and Hong Zhang.A comparative study of real-time semantic segmentation for",
"autonomous driving. In CVPR Workshop, 2018.",
"Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng",
"and Ben Y Zhao. Neural cleanse: Identifying and mitigating backdoor attacks in neural",
"networks.In IEEE S&P,2019.",
"Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic",
"segmentation in street scenes. In CVPR, 2018.",
"Yuanshun Yao, Huiying Li, Haitao Zheng, and Ben Y Zhao. Latent backdoor attacks on",
"deep neural networks. In CCS, 2019.",
"5"
] | {
"arxiv_id": "2103.04038",
"language": "zh",
"page_idx": 4
} |
||
2310.14103-zh-1 | 2310.14103-zh-1.jpg | # 重温指令微调模型评估指导工业应用
Manuel Faysse$^{1,3}$ Gautier Viaud$^1$ Céline Hudelot$^3$ Pierre Colombo$^{2,3}$
$^{1}$Illuin Technology, Paris, France$\:^2$Equall, Paris, France $^3$MICS, CentraleSupélec, Université Paris-Saclay, France
manuel. faysse@centralesupelec. fr
## Abstract
Instruction Fine-Tuning (IFT) is a powerful paradigm that strengthens the zero-shot capabilities of Large Language Models (LLMs), but in doing so induces new evaluation metric requirements. We show LLM-based metrics to be well adapted to these requirements, and leverage them to conduct an investigation of task-specialization strategies, quantifying the trade-offs that emerge in practical industrial settings. Our findings offer practitioners actionable insights for real-world IFT model deployment.
## 1 $\textbf{Introduction}$
为特定应用程序调整预训练语言模型 (LM) 是工业 NLP 的核心 , 可释放特定任务的性能提升并加强模型与行业需求的一致性。获得吸引力的范例是使用指令微调(IFT)模型，LM能够遵循以自然语言 (Wei et al., 2022a; Sanh et al., 2022; Ouyang et al., 2022) 表示的任意指令
研究人员主要致力于改进通用 IFT 模型，以用作能够执行以自然语言(Li et al., 2023; Zhou et al., 2023; Xu et al., 2023a) 表示的指令的通用代理。在工业环境中，提示 ChatGPT 改进电子邮件的措辞或协助代码段将是这种零利用率场景的实例，我们将其定义为$S_0$。但是，关键的工业 LLM 应用程序可能并不总是与$\mathcal{S}_0$保持一致，并且通常会优先考虑其他两个设置。第一个场景$\mathcal{S}_1$需要将通才 IFT 模型的功能扩展到原始指令训练集中未包含的新的特定任务。第二种情况$\mathcal{S}_2$围绕将 IFT 模型转换为专门精通特定任务的专用模型。例如，在$\mathcal{S}_1$中，一家大公司可能希望内部员工使用 LLM 助手，并决定通过训练它来编写具有特定模板方案的备忘录，以响应内部常见问题并使用内部编码来扩展公开可用的聊天模型工具，同时保留原始聊天助手的通用功能。在$\mathcal{S}_2$ 中，同一家公司只对给定的特定任务感兴趣；从业务文档中提取特定信息，并为此专门设计了一个IFT 模型。旨在利用模型的提示和泛化功能进行数据效率更高的培训。在本文中，我们通过实际的镜头
调查专门的 IFT 模型的学习动力学，彻底检查$\mathcal{S}_1$和$\mathcal{S}_2$ 。为了确保我们的工具的可靠性和结论的严谨性，我们首先对 IFT 模型采用的当前评估实践进行严格评估。正式而言，我们的贡献是：
贡献 1。IFT 模型旨在处理不同性质和不同困难的任务。但是，当前用于衡量其性能的指标通常是特定于任务的 (Zellers et al., 2019; Gao et al.,2021),或者依赖于为其他预期目的而设计的自动指标 (Papineni et al., 2002; Lin, 2004) 。为了解决此限制，我们为用于评估 IFT 模型的指标引人了两个新要求：C 可比较性 A 交叉 T 询问 (CAT)和T询问以及E或 A 诺斯蒂主义(TFA )。CAT 要求度量分数在一组不同的生成任务中表现出一致性，这与特定任务中唯一的传统一致性重点相反。TFA 定义了对指标的需求，以证明对输出格式变化的鲁棒性。通过强调现有指标在满足 CAT 和 TFA 方面的缺点，我们提供了令人信服的证据，表明使用 LLM 作为评分代理是 IFT 模型的可行评估替代方案
贡献 2。我们从实际的角度来研究$S_1$和$\mathcal{S}_2$, 并专注于数据可用性和整体性能之间的权衡。我们的分析揭示了IFT 模型专业化期间学习的两个不同阶段：学习格式化和学习解决任务。随后，我们展示了从业者如何(i)利用合成数据来促进学习所需的格式方面，以及(ii)使用 IFT 模型来减少工业场景中对专家数据的需求。我们的研究为希望在生产环境中部署 IFT 模型的从业者提供了实用的见解和可行的建$议。1$
### 2 $\textbf{Re- evaluating IFT Model Evaluatior}$ 2.1 $\textbf{What Should Good Scorers Measure? }$
在场景$\mathcal{S}_0$ 、$\mathcal{S}_1$和$\mathcal{S}_2$中，训练 IFT 模型以执行生成任务。与为具有已知输出格式的单个任务设计的模型不同，IFT 模型能够跨不同的任务和格式 (Ouyang et al., 2022) 生成不同的有效响应。IFT 模型的新颖功能在选择自动评估指标
$^{1}$Code and evaluation datasets are available on https://
github. com/ManuelFay/IFTEval . | [
[
"0.309",
"0.0968",
"0.6885",
"0.0968",
"0.6885",
"0.10986",
"0.309",
"0.10986"
],
[
"0.1797",
"0.1395",
"0.513",
"0.1395",
"0.513",
"0.1538",
"0.1797",
"0.1538"
],
[
"0.4995",
"0.1395",
"0.8237",
"0.1383",
"0.8237",
"0.1526",
"0.4995",
"0.1538"
],
[
"0.2688",
"0.1538",
"0.735",
"0.1561",
"0.735",
"0.1733",
"0.2688",
"0.1716"
],
[
"0.266",
"0.1716",
"0.727",
"0.1733",
"0.727",
"0.1882",
"0.266",
"0.1859"
],
[
"0.3804",
"0.1912",
"0.621",
"0.1912",
"0.621",
"0.2048",
"0.3804",
"0.2048"
],
[
"0.2627",
"0.2542",
"0.3418",
"0.2542",
"0.3418",
"0.2683",
"0.2627",
"0.2683"
],
[
"0.513",
"0.2546",
"0.881",
"0.2546",
"0.881",
"0.2683",
"0.513",
"0.2683"
],
[
"0.511",
"0.2695",
"0.879",
"0.2695",
"0.879",
"0.2832",
"0.511",
"0.2832"
],
[
"0.1453",
"0.2798",
"0.4575",
"0.2798",
"0.4575",
"0.2935",
"0.1453",
"0.2935"
],
[
"0.511",
"0.2861",
"0.881",
"0.2874",
"0.881",
"0.3005",
"0.511",
"0.2998"
],
[
"0.147",
"0.2944",
"0.4575",
"0.2944",
"0.4575",
"0.308",
"0.147",
"0.308"
],
[
"0.511",
"0.3018",
"0.881",
"0.3027",
"0.881",
"0.3176",
"0.511",
"0.3164"
],
[
"0.1453",
"0.308",
"0.4568",
"0.3071",
"0.4568",
"0.3218",
"0.1453",
"0.3225"
],
[
"0.147",
"0.3225",
"0.4568",
"0.3225",
"0.4568",
"0.3362",
"0.147",
"0.3362"
],
[
"0.5146",
"0.3206",
"0.5576",
"0.3206",
"0.5576",
"0.332",
"0.5146",
"0.332"
],
[
"0.147",
"0.3384",
"0.4534",
"0.3374",
"0.4534",
"0.3486",
"0.147",
"0.3499"
],
[
"0.513",
"0.335",
"0.881",
"0.3362",
"0.881",
"0.3508",
"0.513",
"0.3499"
],
[
"0.1486",
"0.3528",
"0.455",
"0.3528",
"0.455",
"0.3635",
"0.1486",
"0.3635"
],
[
"0.511",
"0.3508",
"0.881",
"0.3516",
"0.881",
"0.3665",
"0.511",
"0.3652"
],
[
"0.1453",
"0.3652",
"0.4575",
"0.3652",
"0.4575",
"0.379",
"0.1453",
"0.379"
],
[
"0.511",
"0.3677",
"0.879",
"0.3677",
"0.879",
"0.3813",
"0.511",
"0.3813"
],
[
"0.147",
"0.3801",
"0.4568",
"0.3801",
"0.4568",
"0.3938",
"0.147",
"0.3938"
],
[
"0.513",
"0.3855",
"0.881",
"0.3855",
"0.881",
"0.3984",
"0.513",
"0.3984"
],
[
"0.147",
"0.3938",
"0.4575",
"0.3938",
"0.4575",
"0.4067",
"0.147",
"0.4067"
],
[
"0.513",
"0.401",
"0.879",
"0.401",
"0.879",
"0.4146",
"0.513",
"0.4146"
],
[
"0.1486",
"0.4092",
"0.4534",
"0.4092",
"0.4534",
"0.4192",
"0.1486",
"0.4192"
],
[
"0.1453",
"0.4216",
"0.4575",
"0.4216",
"0.4575",
"0.4353",
"0.1453",
"0.4353"
],
[
"0.513",
"0.4163",
"0.879",
"0.4163",
"0.879",
"0.43",
"0.513",
"0.43"
],
[
"0.147",
"0.437",
"0.2141",
"0.437",
"0.2141",
"0.449",
"0.147",
"0.449"
],
[
"0.5146",
"0.4329",
"0.882",
"0.4329",
"0.882",
"0.4465",
"0.5146",
"0.4465"
],
[
"0.5156",
"0.4507",
"0.879",
"0.4507",
"0.879",
"0.4614",
"0.5156",
"0.4614"
],
[
"0.11755",
"0.4644",
"0.1503",
"0.4644",
"0.1503",
"0.4727",
"0.11755",
"0.4727"
],
[
"0.1436",
"0.462",
"0.2627",
"0.462",
"0.2627",
"0.4756",
"0.1436",
"0.4756"
],
[
"0.513",
"0.4656",
"0.882",
"0.4656",
"0.882",
"0.4785",
"0.513",
"0.4785"
],
[
"0.1184",
"0.4863",
"0.4878",
"0.4863",
"0.4878",
"0.4995",
"0.1184",
"0.4995"
],
[
"0.513",
"0.4827",
"0.8774",
"0.4827",
"0.8774",
"0.4934",
"0.513",
"0.4934"
],
[
"0.1184",
"0.502",
"0.4844",
"0.503",
"0.4844",
"0.5166",
"0.1184",
"0.5156"
],
[
"0.513",
"0.4976",
"0.881",
"0.4976",
"0.881",
"0.511",
"0.513",
"0.511"
],
[
"0.513",
"0.513",
"0.881",
"0.513",
"0.881",
"0.527",
"0.513",
"0.527"
],
[
"0.1184",
"0.5195",
"0.486",
"0.5195",
"0.486",
"0.533",
"0.1184",
"0.533"
],
[
"0.513",
"0.5293",
"0.881",
"0.5293",
"0.881",
"0.542",
"0.513",
"0.542"
],
[
"0.1184",
"0.535",
"0.4844",
"0.535",
"0.4844",
"0.549",
"0.1184",
"0.549"
],
[
"0.513",
"0.546",
"0.8774",
"0.546",
"0.8774",
"0.5557",
"0.513",
"0.5557"
],
[
"0.1184",
"0.552",
"0.4844",
"0.552",
"0.4844",
"0.5654",
"0.1184",
"0.5654"
],
[
"0.513",
"0.561",
"0.8623",
"0.561",
"0.8623",
"0.5747",
"0.513",
"0.5747"
],
[
"0.1184",
"0.568",
"0.4458",
"0.568",
"0.4458",
"0.5806",
"0.1184",
"0.5806"
],
[
"0.1377",
"0.585",
"0.4844",
"0.585",
"0.4844",
"0.5957",
"0.1377",
"0.5957"
],
[
"0.5156",
"0.58",
"0.882",
"0.58",
"0.882",
"0.5903",
"0.5156",
"0.5903"
],
[
"0.1184",
"0.5996",
"0.4836",
"0.5996",
"0.4836",
"0.6133",
"0.1184",
"0.6133"
],
[
"0.513",
"0.594",
"0.882",
"0.594",
"0.882",
"0.608",
"0.513",
"0.608"
],
[
"0.12006",
"0.615",
"0.4844",
"0.615",
"0.4844",
"0.629",
"0.12006",
"0.629"
],
[
"0.513",
"0.611",
"0.8774",
"0.611",
"0.8774",
"0.6216",
"0.513",
"0.6216"
],
[
"0.1184",
"0.6313",
"0.486",
"0.6313",
"0.486",
"0.6445",
"0.1184",
"0.6445"
],
[
"0.513",
"0.627",
"0.881",
"0.627",
"0.881",
"0.64",
"0.513",
"0.64"
],
[
"0.1184",
"0.648",
"0.4844",
"0.648",
"0.4844",
"0.661",
"0.1184",
"0.661"
],
[
"0.513",
"0.6436",
"0.881",
"0.6436",
"0.881",
"0.6567",
"0.513",
"0.6567"
],
[
"0.12006",
"0.6646",
"0.486",
"0.6646",
"0.486",
"0.6777",
"0.12006",
"0.6777"
],
[
"0.513",
"0.659",
"0.882",
"0.659",
"0.882",
"0.673",
"0.513",
"0.673"
],
[
"0.1184",
"0.6797",
"0.486",
"0.6797",
"0.486",
"0.6934",
"0.1184",
"0.6934"
],
[
"0.513",
"0.675",
"0.881",
"0.675",
"0.881",
"0.6885",
"0.513",
"0.6885"
],
[
"0.11755",
"0.695",
"0.4836",
"0.695",
"0.4836",
"0.708",
"0.11755",
"0.708"
],
[
"0.511",
"0.6914",
"0.8774",
"0.6914",
"0.8774",
"0.7017",
"0.511",
"0.7017"
],
[
"0.511",
"0.706",
"0.881",
"0.708",
"0.881",
"0.7217",
"0.511",
"0.7197"
],
[
"0.1234",
"0.712",
"0.4844",
"0.712",
"0.4844",
"0.7256",
"0.1234",
"0.7256"
],
[
"0.513",
"0.7236",
"0.546",
"0.7236",
"0.546",
"0.7363",
"0.513",
"0.7363"
],
[
"0.12006",
"0.73",
"0.4836",
"0.73",
"0.4836",
"0.7407",
"0.12006",
"0.7407"
],
[
"0.11755",
"0.7437",
"0.4844",
"0.7446",
"0.4844",
"0.7583",
"0.11755",
"0.7573"
],
[
"0.11755",
"0.759",
"0.4878",
"0.7603",
"0.4878",
"0.775",
"0.11755",
"0.774"
],
[
"0.51",
"0.7544",
"0.8623",
"0.7544",
"0.8623",
"0.767",
"0.51",
"0.767"
],
[
"0.1184",
"0.777",
"0.4878",
"0.777",
"0.4878",
"0.7905",
"0.1184",
"0.7905"
],
[
"0.511",
"0.781",
"0.8555",
"0.781",
"0.8555",
"0.7944",
"0.511",
"0.7944"
],
[
"0.12006",
"0.7944",
"0.4836",
"0.7944",
"0.4836",
"0.8047",
"0.12006",
"0.8047"
],
[
"0.513",
"0.804",
"0.881",
"0.804",
"0.881",
"0.818",
"0.513",
"0.818"
],
[
"0.12006",
"0.81",
"0.4836",
"0.81",
"0.4836",
"0.821",
"0.12006",
"0.821"
],
[
"0.1184",
"0.8247",
"0.486",
"0.8247",
"0.486",
"0.8384",
"0.1184",
"0.8384"
],
[
"0.511",
"0.8193",
"0.8774",
"0.8193",
"0.8774",
"0.833",
"0.511",
"0.833"
],
[
"0.1184",
"0.8413",
"0.4844",
"0.8413",
"0.4844",
"0.855",
"0.1184",
"0.855"
],
[
"0.513",
"0.8374",
"0.876",
"0.8374",
"0.876",
"0.847",
"0.513",
"0.847"
],
[
"0.1184",
"0.857",
"0.4844",
"0.857",
"0.4844",
"0.8706",
"0.1184",
"0.8706"
],
[
"0.513",
"0.8516",
"0.881",
"0.8516",
"0.881",
"0.865",
"0.513",
"0.865"
],
[
"0.11755",
"0.8726",
"0.4878",
"0.8716",
"0.4878",
"0.8857",
"0.11755",
"0.887"
],
[
"0.5146",
"0.869",
"0.879",
"0.869",
"0.879",
"0.88",
"0.5146",
"0.88"
],
[
"0.1184",
"0.8896",
"0.4844",
"0.8896",
"0.4844",
"0.903",
"0.1184",
"0.903"
],
[
"0.534",
"0.891",
"0.882",
"0.8945",
"0.882",
"0.909",
"0.534",
"0.9062"
],
[
"0.12006",
"0.9062",
"0.4895",
"0.9062",
"0.4895",
"0.9194",
"0.12006",
"0.9194"
],
[
"0.511",
"0.9062",
"0.733",
"0.9062",
"0.733",
"0.9194",
"0.511",
"0.9194"
],
[
"0.4902",
"0.9395",
"0.509",
"0.9395",
"0.509",
"0.9536",
"0.4902",
"0.9536"
]
] | [
"重温指令微调模型评估指导工业应用",
"Manuel Fayssel,3Gautier Viaudl",
"Céline Hudelot3Pierre Colombo2,3",
"iluin Technology, Paris, France2Equall, Paris, France",
"3MICS, CentraleSupelec, Université Paris-Saclay, France",
"[email protected]",
"Abstract",
"调查专门的IFT模型的学习动力学，彻底检查",
"Si和 S2。为了确保我们的工具的可靠性和结",
"Instruction Fine-Tuning (IFT) is a powerful",
"论的严谨性，我们首先对IFT模型采用的当前",
"paradigm that strengthens the zero-shot capa-",
"评估实践进行严格评估。正式而言，我们的贡",
"bilities of Large Language Models (LLMs),",
"but in doing so induces new evaluation metric",
"献是：",
"requirements.We show LLM-based metrics",
"贡献1。IFT模型旨在处理不同性质和不同困",
"tobewelladaptedtotheserequirements,and",
"难的任务。但是，当前用于衡量其性能的指标",
"leverage them to conduct an investigation of",
"通常是特定于任务的（Zellersetal.，2019Gao",
"task-specialization strategies, quantifying the",
"etal.,2021），或者依赖于为其他预期目的而设",
"trade-offs that emerge in practical industrial",
"计的自动指标（Papinenietal.2002;Lin，2004）",
"settings.Ourfindingsofferpractitioners ac",
"tionable insights for real-world IFT model de-",
"为了解决此限制，我们为用于评估IFT模型的",
"ployment.",
"指标引人了两个新要求：C可比较性A交叉T",
"询问（CAT）和T询问以及F或A诺斯蒂主义",
"1",
"Introduction",
"（TFA）。CAT要求度量分数在一组不同的生成",
"为特定应用程序调整预训练语言模型LM）是",
"任务中表现出一致性，这与特定任务中唯一的",
"工业NLP的核心，可释放特定任务的性能提升",
"传统一致性重点相反。TFA定义了对指标的需",
"求，以证明对输出格式变化的鲁棒性。通过强",
"并加强模型与行业需求的一致性。获得吸引力",
"调现有指标在满足CAT和TFA方面的缺点，我",
"的范例是使用指令微调（IFT）模型，LM能够",
"们提供了今人信服的证据：表明使用LLM作",
"遵循以自然语言（Weietal.，2022a；Sanhetal.，",
"为评分代理是IFT模型的可行评估替代方案",
"2022；Ouyangetal.，2022）表示的任意指令",
"研究人员主要致力于改进通用IFT模型，以",
"贡献2。我们从实际的角度来研究S1和S2，",
"用作能够执行以自然语言（Lietal.，2023；Zhou",
"并专注于数据可用性和整体性能之间的权衡。",
"etal.2023；Xuetal.，2023a）表示的指令的通用",
"我们的分析揭示了IFT模型专业化期间学习的",
"代理。在工业环境中，提示ChatGPT改进电子",
"两个不同阶段：学习格式化和学习解决任务。",
"邮件的措辞或协助代码段将是这种零利用率场",
"随后，我们展示了从业者如何（i）利用合成数",
"景的实例，我们将其定义为So。但是，关键的",
"据来促进学习所需的格式方面，以及（ii）使",
"工业LLM应用程序可能并不总是与So保持",
"用IFT模型来减少工业场景中对专家数据的需",
"一致，并且通常会优先考虑其他两个设置。第",
"求。我们的研究为希望在生产环境中部署IFI",
"模型的从业者提供了实用的见解和可行的建",
"一个场景S1需要将通才IFT模型的功能扩展",
"议。",
"到原始指令训练集中未包含的新的特定任务。",
"第二种情况S2围绕将IFT模型转换为专门精",
"通特定任务的专用模型。例如，在S1中，一家",
"2 Re-evaluating IFT Model Evaluation",
"大公司可能希望内部员工使用LLM助手，并",
"2.1 What Should Good Scorers Measure?",
"决定通过训练它来编写具有特定模板方案的备",
"在场景So、Si和S2中，训练IFT模型以执行",
"忘录，以响应内部常见问题并使用内部编码来",
"扩展公开可用的聊天模型工具，同时保留原始",
"生成任务。与为具有已知输出格式的单个任务",
"聊天助手的通用功能。在S2中，同一家公司",
"设计的模型不同，IFT模型能够跨不同的任务",
"只对给定的特定任务感兴趣；从业务文档中提",
"和格式（Ouyangetal.，2022）生成不同的有效响",
"取特定信息，并为此专门设计了一个IFT模型，",
"应。IFT模型的新颖功能在选择自动评估指标",
"旨在利用模型的提示和泛化功能进行数据效率",
"'Code and evaluation datasets are available on https: //",
"更高的培训。在本文中，我们通过实际的镜头",
"github.com/ManuelFay/IFTEval.",
"1"
] | {
"arxiv_id": "2310.14103",
"language": "zh",
"page_idx": 1
} |
||
2310.14103-zh-2 | 2310.14103-zh-2.jpg | 时提出了新的考虑因素。
跨任务的可比性(CAT)。标准评估指标旨在满足一个关键要求：每个任务相对于人类判断(Specia et al.,2010)的连贯性。但是，由于 IFT 模型的多任务性质，不同任务(Colombo et al., 2022a; Himmi et al., 2023) 的分数也应具有可比性。换句话说，评分量表应该是绝对的，并且与人类对所有任务的偏好保持一致。为了测量CAT ,我们将混合不同任务的样本，并计算其得分与人类判断$^{2}$的 Spearman 相关性($\rho)$。在场景$\mathcal{S}_0$和$\mathcal{S}_1$中，此要求对于衡量不同任务之间的模型性能以及做出有关模型变体之间权衡的明智决策至关重要
任务和格式-不可知论(TFA)。评估指标应该对与输出格式相关的工件和评估任务(Liang et al., 2022) 的性质具有鲁棒性。实施特定于任务的评分指标不是通用 IFT 模型的可扩展解决方案。为了测量 TFA , 我们计算以零拍方式提示的模型与掌握任务格式的模型(在 1000 个目标任务样本上训练)之间的相对目标任务改进。将每个指标的 TFA 与人类报告的性能改进进行比较，可以掌握掌握任务格式对指标性能的影响程度，而与内在任务性能无关。在工业场景中，此要求是必不可少的，因为它确保了由于训练数据格式化工件而导致的评估偏差最小。实际上，工业参与者可能会将许多数据集添加到训练指令集( $\mathcal{S}_1$ )中，或者在( $\mathcal{S}_2$ )上完全训练模型，这些数据集具有特定的响应格式，这与零拍模型将回答的内容不同，从而导致潜在的大格式偏差
任务内可比性(CIT)。虽然这绝不是一个新颖的要求，但度量标准必须在给定任务中一致地衡量性能。我们通过计算特定任务样本与人类判断之间的 Spearman $\rho$相关系数来验证这一点
在 IFTLLM 的所有工业场景中，严格的模型评估必须与符合 CAT 和 TFA 以及标准 CII 措施的评估指标相关联
# 2.2 Existing Metrics
当前评估。当前，出现了两种主要的范式来评估 IFT 模型的性能：(i) 依靠参考匹配评分指标，例如 ROUGE-L (Lin,2004)或少数分类基准 (Hendrycks et al., 2021; Gao et al., 2021; Zellers et al., 2019) 中类别标签的归一化对数概率，以及(ii)模型排名框架，基于人类或LLM 评估者 (Chiang et al., 2023; Dubois et al., 2023; Gudibande et al., 2023) 判断的响应质量
${\overline{{^2}\text{Common w}}}$hen benchmarking metrics(Bhandari et al., 2020; Colombo et al., 2021; Chhun et al., 2022; Staerman et al., 2021; Fabbri et al., 2021; Colombo et al., 2021) , we extend the tool to inter-task settings.
2
的成对偏好比较。基于语言模型的评分已被证明是特定任务的有前途的替代方法，例如摘要(Liu et al., 2023; Colombo et al., 2022b) 或翻译(Kocmi and Federmann, 2023; Xu et al., 2023b) 。我们的工作扩展了这些发现，以展示 LLM 在CAT 和 TFA 方面的多任务评分功能
LMs 作为可行的评分机制。鉴于 IFT 模型生成固有的开放性，我们采用无参考设置来确保无偏评估。我们向 LLM $^{3}$提供一个输入提示和相应的生成响应，提示它在 0 到 10 的范围内分配分数，随后在 0 到 1 之间缩放，以便与其他评估指标进行比较
基线度量。我们通过将建议的指标与完善的基于参考的指标进行比较来评估 CAT 和 TFA 的实现情况，包括 ROUGE$^4$, BScore (Zhang et al., 2020) 和 SBERT (Reimers and Gurevych, 2019) ,以及机器学习的指标，即根据人类偏好训练的 OpenAssistant 奖励模型 (RM) (Köpf et al., 2023)
# 2.3 Experimental setup
训练 IFT 模型。通过在大型指令语料库上微调基本模型来训练 IFT 模型，该语料库是通过人类注释 (Ouyang et al., 2022; Köpf et al., 2023) 或连接特定于任务的数据集 (Sanh et al., 2022; Mishra et al., 2022) 收集的。根据最近的工作(Chiang et al., 2023; Wang et al., 2023; Peng et al. 2023),我们利用合成数据作为我们的 IFT 模型 (Taori et al., 2023) 中的基本指令集
自动基准度量。为了对指标进行基准测试，我们依赖于合成数据和真实数据的组合。对于合成数据，我们使用羊驼 GPT4 数据集 (Taori et al., 2023),并在 13 个任务类别中标记数据(参见Sec. A.1 )(e。g。,逻辑，代码，重写)。对于人类数据，我们专注于具有行业兴趣的任务。具体来说，我们包括自然语言推理 (Williams et al., 2018; Wang et al.,2019) , 问答 (Rajpurkar et al., 2016) , NER (Tjong Kim Sang and De Meulder, 2003) 和情感分类 (Socher et al., 2013; Agirre et al., 2013) )。为了构建我们的指标评估数据集，我们在各种数据混合和目标任务上训练和运行 LLaMA-7B 模型 (Touvron et al., 2023a) 。为了严格起见，我们还报告了来自(Stiennon et al., 2022) (sum) 的人类反馈数据集的摘要分数。5
We rely on LMs available through the OpenAI API( i.e.
,GPT4 and GPT3.5). See Sec. A.3 for details.
$^{4}$ROUGE-1 is used here, it is best on one-word long labels
$^{5}$More details in Sec. B.1 ) | [
[
"0.1184",
"0.0873",
"0.309",
"0.0873",
"0.309",
"0.10095",
"0.1184",
"0.10095"
],
[
"0.5146",
"0.0873",
"0.881",
"0.0873",
"0.881",
"0.10095",
"0.5146",
"0.10095"
],
[
"0.11755",
"0.1027",
"0.486",
"0.1027",
"0.486",
"0.1164",
"0.11755",
"0.1164"
],
[
"0.511",
"0.1027",
"0.8774",
"0.1027",
"0.8774",
"0.1164",
"0.511",
"0.1164"
],
[
"0.11755",
"0.11816",
"0.4836",
"0.11816",
"0.4836",
"0.1318",
"0.11755",
"0.1318"
],
[
"0.513",
"0.12054",
"0.879",
"0.12054",
"0.879",
"0.1312",
"0.513",
"0.1312"
],
[
"0.1184",
"0.1354",
"0.4844",
"0.1354",
"0.4844",
"0.1484",
"0.1184",
"0.1484"
],
[
"0.51",
"0.1342",
"0.882",
"0.1354",
"0.882",
"0.1484",
"0.51",
"0.1478"
],
[
"0.1184",
"0.1509",
"0.4836",
"0.1497",
"0.4836",
"0.1633",
"0.1184",
"0.1644"
],
[
"0.513",
"0.1509",
"0.882",
"0.1509",
"0.882",
"0.1644",
"0.513",
"0.1644"
],
[
"0.1184",
"0.1675",
"0.482",
"0.1675",
"0.482",
"0.1776",
"0.1184",
"0.1776"
],
[
"0.513",
"0.1675",
"0.784",
"0.1675",
"0.784",
"0.1812",
"0.513",
"0.1812"
],
[
"0.1184",
"0.1818",
"0.482",
"0.1818",
"0.482",
"0.1953",
"0.1184",
"0.1953"
],
[
"0.511",
"0.1853",
"0.882",
"0.1859",
"0.882",
"0.2007",
"0.511",
"0.1996"
],
[
"0.1184",
"0.1978",
"0.4844",
"0.199",
"0.4844",
"0.2131",
"0.1184",
"0.212"
],
[
"0.511",
"0.2031",
"0.879",
"0.2031",
"0.879",
"0.2162",
"0.511",
"0.2162"
],
[
"0.1184",
"0.2156",
"0.4878",
"0.2156",
"0.4878",
"0.2286",
"0.1184",
"0.2286"
],
[
"0.513",
"0.2197",
"0.881",
"0.2197",
"0.881",
"0.2328",
"0.513",
"0.2328"
],
[
"0.11755",
"0.2322",
"0.4878",
"0.2322",
"0.4878",
"0.2452",
"0.11755",
"0.2452"
],
[
"0.513",
"0.2363",
"0.881",
"0.2363",
"0.881",
"0.2494",
"0.513",
"0.2494"
],
[
"0.1184",
"0.2477",
"0.4844",
"0.2477",
"0.4844",
"0.2612",
"0.1184",
"0.2612"
],
[
"0.513",
"0.2517",
"0.881",
"0.2517",
"0.881",
"0.2654",
"0.513",
"0.2654"
],
[
"0.1184",
"0.2642",
"0.486",
"0.2642",
"0.486",
"0.2778",
"0.1184",
"0.2778"
],
[
"0.5156",
"0.2695",
"0.659",
"0.2695",
"0.659",
"0.2798",
"0.5156",
"0.2798"
],
[
"0.11755",
"0.279",
"0.3005",
"0.279",
"0.3005",
"0.2922",
"0.11755",
"0.2922"
],
[
"0.513",
"0.288",
"0.882",
"0.288",
"0.882",
"0.3018",
"0.513",
"0.3018"
],
[
"0.1184",
"0.2957",
"0.486",
"0.2957",
"0.486",
"0.3088",
"0.1184",
"0.3088"
],
[
"0.513",
"0.304",
"0.882",
"0.304",
"0.882",
"0.3176",
"0.513",
"0.3176"
],
[
"0.1184",
"0.3123",
"0.486",
"0.3123",
"0.486",
"0.326",
"0.1184",
"0.326"
],
[
"0.513",
"0.3206",
"0.881",
"0.3206",
"0.881",
"0.3342",
"0.513",
"0.3342"
],
[
"0.11755",
"0.328",
"0.4844",
"0.328",
"0.4844",
"0.3416",
"0.11755",
"0.3416"
],
[
"0.511",
"0.3362",
"0.881",
"0.3362",
"0.881",
"0.3499",
"0.511",
"0.3499"
],
[
"0.1184",
"0.3425",
"0.4844",
"0.3425",
"0.4844",
"0.3557",
"0.1184",
"0.3557"
],
[
"0.519",
"0.357",
"0.529",
"0.357",
"0.529",
"0.361",
"0.519",
"0.361"
],
[
"0.532",
"0.354",
"0.879",
"0.354",
"0.879",
"0.364",
"0.532",
"0.364"
],
[
"0.1184",
"0.3599",
"0.4844",
"0.3599",
"0.4844",
"0.3735",
"0.1184",
"0.3735"
],
[
"0.511",
"0.3682",
"0.881",
"0.3682",
"0.881",
"0.3818",
"0.511",
"0.3818"
],
[
"0.1184",
"0.3748",
"0.4844",
"0.3748",
"0.4844",
"0.3884",
"0.1184",
"0.3884"
],
[
"0.513",
"0.383",
"0.5576",
"0.3855",
"0.5557",
"0.398",
"0.511",
"0.3955"
],
[
"0.12006",
"0.3938",
"0.4844",
"0.3938",
"0.4844",
"0.4067",
"0.12006",
"0.4067"
],
[
"0.11755",
"0.4092",
"0.4844",
"0.4092",
"0.4844",
"0.4229",
"0.11755",
"0.4229"
],
[
"0.1184",
"0.4246",
"0.486",
"0.4246",
"0.486",
"0.4382",
"0.1184",
"0.4382"
],
[
"0.513",
"0.4258",
"0.559",
"0.4258",
"0.559",
"0.437",
"0.513",
"0.437"
],
[
"0.553",
"0.4246",
"0.714",
"0.4258",
"0.714",
"0.4407",
"0.553",
"0.4395"
],
[
"0.1184",
"0.4412",
"0.4844",
"0.4412",
"0.4844",
"0.4548",
"0.1184",
"0.4548"
],
[
"0.1184",
"0.4573",
"0.4844",
"0.4573",
"0.4844",
"0.4702",
"0.1184",
"0.4702"
],
[
"0.513",
"0.4536",
"0.881",
"0.4536",
"0.881",
"0.4673",
"0.513",
"0.4673"
],
[
"0.1184",
"0.4727",
"0.486",
"0.4727",
"0.486",
"0.4863",
"0.1184",
"0.4863"
],
[
"0.511",
"0.4702",
"0.879",
"0.4702",
"0.879",
"0.484",
"0.511",
"0.484"
],
[
"0.1184",
"0.4893",
"0.4878",
"0.4893",
"0.4878",
"0.503",
"0.1184",
"0.503"
],
[
"0.513",
"0.4863",
"0.881",
"0.4863",
"0.881",
"0.4995",
"0.513",
"0.4995"
],
[
"0.1184",
"0.505",
"0.4878",
"0.505",
"0.4878",
"0.5186",
"0.1184",
"0.5186"
],
[
"0.511",
"0.5005",
"0.881",
"0.503",
"0.881",
"0.5166",
"0.511",
"0.514"
],
[
"0.1184",
"0.521",
"0.486",
"0.521",
"0.486",
"0.5337",
"0.1184",
"0.5337"
],
[
"0.511",
"0.5186",
"0.881",
"0.5186",
"0.881",
"0.532",
"0.511",
"0.532"
],
[
"0.1184",
"0.5376",
"0.4844",
"0.5376",
"0.4844",
"0.5503",
"0.1184",
"0.5503"
],
[
"0.51",
"0.5337",
"0.879",
"0.5337",
"0.879",
"0.5474",
"0.51",
"0.5474"
],
[
"0.1184",
"0.554",
"0.2837",
"0.554",
"0.2837",
"0.568",
"0.1184",
"0.568"
],
[
"0.5146",
"0.5527",
"0.881",
"0.5527",
"0.881",
"0.5635",
"0.5146",
"0.5635"
],
[
"0.11755",
"0.5684",
"0.4836",
"0.5684",
"0.4836",
"0.582",
"0.11755",
"0.582"
],
[
"0.513",
"0.568",
"0.806",
"0.568",
"0.806",
"0.5806",
"0.513",
"0.5806"
],
[
"0.1184",
"0.585",
"0.486",
"0.585",
"0.486",
"0.5986",
"0.1184",
"0.5986"
],
[
"0.5156",
"0.5874",
"0.881",
"0.5874",
"0.881",
"0.601",
"0.5156",
"0.601"
],
[
"0.1184",
"0.6016",
"0.486",
"0.6016",
"0.486",
"0.615",
"0.1184",
"0.615"
],
[
"0.513",
"0.604",
"0.882",
"0.604",
"0.882",
"0.6177",
"0.513",
"0.6177"
],
[
"0.1184",
"0.618",
"0.4668",
"0.618",
"0.4668",
"0.632",
"0.1184",
"0.632"
],
[
"0.513",
"0.619",
"0.881",
"0.619",
"0.881",
"0.633",
"0.513",
"0.633"
],
[
"0.11755",
"0.6353",
"0.1377",
"0.6353",
"0.1377",
"0.6484",
"0.11755",
"0.6484"
],
[
"0.513",
"0.6353",
"0.881",
"0.6353",
"0.881",
"0.6484",
"0.513",
"0.6484"
],
[
"0.1335",
"0.651",
"0.4878",
"0.651",
"0.4878",
"0.6646",
"0.1335",
"0.6646"
],
[
"0.513",
"0.652",
"0.881",
"0.652",
"0.881",
"0.6665",
"0.513",
"0.6665"
],
[
"0.11755",
"0.665",
"0.4878",
"0.6665",
"0.4878",
"0.681",
"0.11755",
"0.6797"
],
[
"0.513",
"0.6665",
"0.881",
"0.6675",
"0.881",
"0.6816",
"0.513",
"0.681"
],
[
"0.1184",
"0.683",
"0.266",
"0.683",
"0.266",
"0.697",
"0.1184",
"0.697"
],
[
"0.513",
"0.683",
"0.879",
"0.683",
"0.879",
"0.697",
"0.513",
"0.697"
],
[
"0.511",
"0.6987",
"0.881",
"0.6997",
"0.881",
"0.7134",
"0.511",
"0.712"
],
[
"0.11584",
"0.709",
"0.294",
"0.7104",
"0.294",
"0.7246",
"0.11584",
"0.7236"
],
[
"0.513",
"0.7163",
"0.881",
"0.7163",
"0.881",
"0.731",
"0.513",
"0.731"
],
[
"0.11755",
"0.7324",
"0.486",
"0.7324",
"0.486",
"0.745",
"0.11755",
"0.745"
],
[
"0.511",
"0.7324",
"0.881",
"0.7324",
"0.881",
"0.745",
"0.511",
"0.745"
],
[
"0.11755",
"0.7476",
"0.486",
"0.7476",
"0.486",
"0.761",
"0.11755",
"0.761"
],
[
"0.511",
"0.749",
"0.881",
"0.749",
"0.881",
"0.7627",
"0.511",
"0.7627"
],
[
"0.1184",
"0.7656",
"0.4895",
"0.7656",
"0.4895",
"0.7793",
"0.1184",
"0.7793"
],
[
"0.513",
"0.7666",
"0.879",
"0.7666",
"0.879",
"0.777",
"0.513",
"0.777"
],
[
"0.1184",
"0.781",
"0.4895",
"0.781",
"0.4895",
"0.7944",
"0.1184",
"0.7944"
],
[
"0.511",
"0.78",
"0.881",
"0.78",
"0.881",
"0.7935",
"0.511",
"0.7935"
],
[
"0.11755",
"0.7964",
"0.486",
"0.7964",
"0.486",
"0.81",
"0.11755",
"0.81"
],
[
"0.511",
"0.7964",
"0.881",
"0.7964",
"0.881",
"0.81",
"0.511",
"0.81"
],
[
"0.1184",
"0.8135",
"0.4878",
"0.8135",
"0.4878",
"0.8267",
"0.1184",
"0.8267"
],
[
"0.511",
"0.8125",
"0.879",
"0.8125",
"0.879",
"0.826",
"0.511",
"0.826"
],
[
"0.11755",
"0.829",
"0.4878",
"0.829",
"0.4878",
"0.843",
"0.11755",
"0.843"
],
[
"0.511",
"0.829",
"0.5767",
"0.826",
"0.5786",
"0.8403",
"0.513",
"0.8433"
],
[
"0.1184",
"0.8457",
"0.4844",
"0.8457",
"0.4844",
"0.8594",
"0.1184",
"0.8594"
],
[
"0.1377",
"0.8706",
"0.4836",
"0.8716",
"0.4836",
"0.883",
"0.1377",
"0.882"
],
[
"0.532",
"0.866",
"0.879",
"0.868",
"0.879",
"0.882",
"0.532",
"0.88"
],
[
"0.11755",
"0.883",
"0.4836",
"0.883",
"0.4836",
"0.8936",
"0.11755",
"0.8936"
],
[
"0.513",
"0.882",
"0.79",
"0.882",
"0.79",
"0.8926",
"0.513",
"0.8926"
],
[
"0.11755",
"0.8926",
"0.4878",
"0.8936",
"0.4878",
"0.9077",
"0.11755",
"0.9067"
],
[
"0.534",
"0.891",
"0.881",
"0.8926",
"0.881",
"0.9067",
"0.534",
"0.9062"
],
[
"0.11755",
"0.905",
"0.3342",
"0.9062",
"0.3342",
"0.9194",
"0.11755",
"0.9185"
],
[
"0.5356",
"0.905",
"0.6978",
"0.905",
"0.6978",
"0.9185",
"0.5356",
"0.9185"
],
[
"0.4902",
"0.9395",
"0.509",
"0.9395",
"0.509",
"0.955",
"0.4902",
"0.955"
]
] | [
"时提出了新的考虑因素。",
"的成对偏好比较。基于语言模型的评分已被证",
"跨任务的可比性（CAT）。标准评估指标旨在",
"明是特定任务的有前途的替代方法，例如摘要",
"满足一个关键要求：每个任务相对于人类判断",
"(Liuetal.2023;Colomboetal.,2022b）或翻译",
"(Speciaetal.，2010)的连贯性。但是，由于IFT",
"(Kocmi and Federmann, 2023;Xu et al.,2023b)。",
"模型的多任务性质，不同任务（Colomboetal.",
"我们的工作扩展了这些发现，以展示LLM在",
"2022a：Himmietal.2023）的分数也应具有可比",
"CAT和TFA方面的多任务评分功能",
"性。换句话说，评分量表应该是绝对的，并且",
"LMs作为可行的评分机制。鉴于IFT模型生成",
"与人类对所有任务的偏好保持一致。为了测量",
"固有的开放性，我们采用无参考设置来确保无",
"CAT，我们将混合不同任务的样本，并计算其",
"偏评估。我们向LLM3提供一个输入提示和相",
"得分与人类判断2的Spearman相关性（p）。",
"应的生成响应，提示它在0到10的范围内分",
"在场景So和Si中，此要求对于衡量不同任务",
"配分数，随后在0到1之间缩放，以便与其他",
"之间的模型性能以及做出有关模型变体之间权",
"评估指标进行比较",
"衡的明智决策至关重要",
"基线度量。我们通过将建议的指标与完善的基",
"任务和格式-不可知论（TFA）。评估指标应该",
"于参考的指标进行比较来评估CAT和TFA的实",
"对与输出格式相关的工件和评估任务(Liang",
"现情况，包括ROUGE4，BScore(Zhangetal.，",
"etal.,2022)的性质具有鲁棒性。实施特定于任",
"2020) 和 SBERT (Reimers and Gurevych, 2019)",
"务的评分指标不是通用IFT模型的可扩展解决",
"",
"以及机器学习的指标，即根据人类偏好训练",
"方案。为了测量TFA，我们计算以零拍方式提",
"的OpenAssistant奖励模型（RM）（Kopfetal.，",
"示的模型与掌握任务格式的模型（在1000个",
"2023)",
"目标任务样本上训练）之间的相对目标任务改",
"进。将每个指标的TFA与人类报告的性能改进",
"进行比较，可以掌握掌握任务格式对指标性能",
"2.3",
"Experimental setup",
"的影响程度，而与内在任务性能无关。在工业",
"场景中，此要求是必不可少的，因为它确保了",
"训练IFT模型。通过在大型指令语料库上微",
"由于训练数据格式化工件而导致的评估偏差最",
"调基本模型来训练IFT模型，该语料库是通过",
"小。实际上，工业参与者可能会将许多数据集",
"人类注释（Ouyang etal.,2022;Kopf etal.,2023)",
"添加到训练指令集（S1）中，或者在（S2）上",
"或连接特定于任务的数据集（Sanhetal.,2022;",
"完全训练模型，这些数据集具有特定的响应格",
"Mishraetal.，2022）收集的。根据最近的工作",
"式，这与零拍模型将回答的内容不同，从而导",
"(Chiang et al., 2023; Wang et al., 2023; Peng et al.",
"致潜在的大格式偏差",
"2023），我们利用合成数据作为我们的IFT模",
"任务内可比性（CIT）。虽然这绝不是一个新颖",
"型（Taorietal.,2023）中的基本指令集",
"的要求，但度量标准必须在给定任务中一致地",
"自动基准度量。为了对指标进行基准测试，我们",
"衡量性能。我们通过计算特定任务样本与人类",
"依赖于合成数据和真实数据的组合。对于合成",
"判断之间的Spearmanp相关系数来验证这",
"数据，我们使用羊驼GPT4数据集（Taorietal.，",
"点",
"2023），并在13个任务类别中标记数据（参见",
"在IFTLLM的所有工业场景中，严格的模型",
"Sec.A.1)(e。g。，逻辑，代码，重写)。对于人类",
"评估必须与符合CAT和TFA以及标准CIT措施",
"数据，我们专注于具有行业兴趣的任务。具体",
"的评估指标相关联",
"来说，我们包括自然语言推理（Williamsetal.，",
"2018;Wangetal.,2019），问答（Rajpurkaretal.,",
"2.2Existing Metrics",
"2016) , NER (Tjong Kim Sang and De Meulder,",
"当前评估。当前，出现了两种主要的范式来",
"2003）和情感分类（Socheretal.，2013；Agirre",
"评估IFT模型的性能：（i）依靠参考匹配评",
"etal.,2013））。为了构建我们的指标评估数据",
"分指标，例如ROUGE-L（Lin，2004）或少数分",
"集，我们在各种数据混合和自标任务上训练",
"类基准 (Hendrycks et al., 2021; Gao et al., 2021;",
"和运行LLaMA-7B模型（Touvronetal.，2023a)",
"Zellersetal.2019）中类别标签的归一化对数",
"。为了严格起见，我们还报告了来自（Stiennon",
"概率，以及（ii）模型排名框架，基于人类或",
"etal.,2022）（sUM）的人类反馈数据集的摘要",
"LLM 评估者 (Chiang et al., 2023; Dubois et al.,",
"分数。5",
"2023；Gudibandeetal，2023）判断的响应质量",
"2Common when benchmarking metrics (Bhandari etal..",
"3We rely on LMs available through the OpenAI API (i.e.",
"2020: Colombo et al., 2021; Chhun et al., 2022; Staerman",
", GPT4 and GPT3.5). See Sec. A.3 for details",
"et al., 2021; Fabbri et al., 2021; Colombo et al., 2021), we",
"4ROUGE-1 is used here, it is best on one-word long labels",
"extend the tool to inter-task settings",
"More details in Sec. B.1)",
"2"
] | {
"arxiv_id": "2310.14103",
"language": "zh",
"page_idx": 2
} |
||
2310.14103-zh-4 | 2310.14103-zh-4.jpg | <FigureHere>
Figure 2: Incorporating $0\leq N\leq1000$ real task sam-
ples into IFT model training
3.2 $S_2: \textbf{IFT models as Task- Specific Solvers}$
设置。我们使用四种模型架构，对于每种架构， 我们使用基本模型来使用合成的 Alpaca 数据集训练 IFT 模型变体。然后，我们在从目标任务中提取的$N$样本的子集上微调基本模型及其 IFT 变体。此设置模拟了一个工业场景，在该场景中，有限的数据可用于在唯一任务上专门化模型，并评估在对目标任务进行微调之前对基本模型进行指令调整是否有好处。
结果。Fig.4 演示了IFT 模型在低数据场景中表现出增强的性能(当$10\leq N\leq200$时)。直观上，IFT 模型能够更好地利用提示中给出的任务描述，从而实现提升的零拍性能(Scao and Rush,2021)。当使用少量样本进行微调时，这可以补充和引导模型训练。当更多的样本可用时($N\geq200$),任务模式足够清晰，提示和IFT 训练阶段的好处消失
<FigureHere>
Figure 4: GPT4 score on SST-2 test set after finetuning with $0\leq N\leq1000$ samples on a (base) LM or an IFT model. Further experiments can be found in Sec. C.2.2
Sec.C.1.2 )中，导致性能几乎没有下降$(<1\%)$。利用合成数据学习格式化。我们的发现提出了一种简单的方法来优化真实示例的使用：在依靠真实样本来增强整体模型性能之前，使用合成示例来帮助模型掌握所需的格式。我们重复先前的实验，用 GPT4 合成生成的样本 (S)或带有随机标签(R)的合成样本(Fig.3)Exact Match 替换$N$人类注释的目标任务训练样本(H) 显示合成或随机标记的数据确实可以用于学习所需的格式，尽管质量更好的人类数据最终会在更多样本的情况下产生更好的结果。在(S+ H) 中，我们训练 100 个合成样本，然后训练$N$人类注释样本。该技术使模型在接受高质量数据训练之前能够掌握格式，大大提高了人类注释数据的采样效率。
<FigureHere>
Figure 3: Incorporating $0\leq N\leq1000$ (H)uman, (S)ynthetic and (R)andomly labeled synthetic data samples in IFT training set. (S+H) is trained on 100 synthetic samples, then $N$ human data samples.
这一发现与 Sec. 3.1 中给出的结果一致，强调了合成数据集提高工业场景中数据效率的潜力
### 4 $\textbf{Kev Takeawavs for Practitioners}$
利用 LLM 进行评估。评估 IFT 模型是具有挑战性的，因为它要求跨任务和格式不可知论的可比性，标准指标与之斗争。虽然 LLM 评分并不理想(Sec.5 中的限制),但这是从业者应该添加到他们的武器库中的一个很好的选择。利用合成数据实现高效学习。基于 LLM 的评估揭示了这样一个事实，即利用合成数据提供了一种快速且经济高效的方法，可以在低数据状态下掌握格式，而不会降低性能。这种方法在各种情况下都是可行的，它提供了一个机会， 可以更有效地利用可能有限的专家注释数据
4 | [
[
"0.1838",
"0.0885",
"0.2477",
"0.0885",
"0.2477",
"0.0986",
"0.1838",
"0.0986"
],
[
"0.4165",
"0.0885",
"0.476",
"0.0885",
"0.476",
"0.0986",
"0.4165",
"0.0986"
],
[
"0.51",
"0.0861",
"0.876",
"0.0861",
"0.876",
"0.0998",
"0.51",
"0.0998"
],
[
"0.3494",
"0.09143",
"0.3628",
"0.09143",
"0.3628",
"0.0956",
"0.3494",
"0.0956"
],
[
"0.3005",
"0.0903",
"0.3167",
"0.0903",
"0.3167",
"0.0974",
"0.3005",
"0.0974"
],
[
"0.1855",
"0.0998",
"0.22",
"0.0998",
"0.22",
"0.111",
"0.1855",
"0.111"
],
[
"0.3074",
"0.0998",
"0.339",
"0.0998",
"0.339",
"0.111",
"0.3074",
"0.111"
],
[
"0.4282",
"0.0986",
"0.4575",
"0.0986",
"0.4575",
"0.111",
"0.4282",
"0.111"
],
[
"0.513",
"0.10986",
"0.8774",
"0.111",
"0.8774",
"0.1259",
"0.513",
"0.1247"
],
[
"0.11755",
"0.11755",
"0.1628",
"0.11755",
"0.1628",
"0.1538",
"0.11755",
"0.1538"
],
[
"0.2585",
"0.1223",
"0.2747",
"0.1223",
"0.2747",
"0.1312",
"0.2585",
"0.1312"
],
[
"0.3787",
"0.1289",
"0.3972",
"0.1289",
"0.3972",
"0.1372",
"0.3787",
"0.1372"
],
[
"0.513",
"0.1277",
"0.881",
"0.1277",
"0.881",
"0.1414",
"0.513",
"0.1414"
],
[
"0.5146",
"0.1437",
"0.881",
"0.1437",
"0.881",
"0.1567",
"0.5146",
"0.1567"
],
[
"0.2585",
"0.1509",
"0.2795",
"0.1509",
"0.2795",
"0.158",
"0.2585",
"0.158"
],
[
"0.3804",
"0.1509",
"0.398",
"0.1509",
"0.398",
"0.158",
"0.3804",
"0.158"
],
[
"0.392",
"0.1567",
"0.4937",
"0.1567",
"0.4937",
"0.1675",
"0.392",
"0.1675"
],
[
"0.1561",
"0.158",
"0.2527",
"0.158",
"0.2527",
"0.1687",
"0.1561",
"0.1687"
],
[
"0.1427",
"0.1526",
"0.1561",
"0.1526",
"0.1561",
"0.1592",
"0.1427",
"0.1592"
],
[
"0.2747",
"0.158",
"0.372",
"0.158",
"0.372",
"0.1687",
"0.2747",
"0.1687"
],
[
"0.5146",
"0.1603",
"0.881",
"0.1603",
"0.881",
"0.1733",
"0.5146",
"0.1733"
],
[
"0.3838",
"0.1758",
"0.392",
"0.1758",
"0.392",
"0.1787",
"0.3838",
"0.1787"
],
[
"0.5146",
"0.1758",
"0.882",
"0.1758",
"0.882",
"0.1895",
"0.5146",
"0.1895"
],
[
"0.3376",
"0.1818",
"0.3586",
"0.1818",
"0.3586",
"0.1853",
"0.3376",
"0.1853"
],
[
"0.1276",
"0.1882",
"0.157",
"0.1882",
"0.157",
"0.2007",
"0.1276",
"0.2007"
],
[
"0.2568",
"0.1895",
"0.2747",
"0.1895",
"0.2747",
"0.1965",
"0.2568",
"0.1965"
],
[
"0.3787",
"0.1936",
"0.3972",
"0.1936",
"0.3972",
"0.2031",
"0.3787",
"0.2031"
],
[
"0.5146",
"0.1924",
"0.881",
"0.1924",
"0.881",
"0.206",
"0.5146",
"0.206"
],
[
"0.12177",
"0.2019",
"0.1335",
"0.199",
"0.1394",
"0.209",
"0.126",
"0.212"
],
[
"0.5146",
"0.2079",
"0.881",
"0.2079",
"0.881",
"0.2214",
"0.5146",
"0.2214"
],
[
"0.157",
"0.2227",
"0.251",
"0.2227",
"0.251",
"0.2299",
"0.157",
"0.2299"
],
[
"0.2778",
"0.2227",
"0.3704",
"0.2227",
"0.3704",
"0.2299",
"0.2778",
"0.2299"
],
[
"0.3938",
"0.2214",
"0.4954",
"0.2214",
"0.4954",
"0.2322",
"0.3938",
"0.2322"
],
[
"0.513",
"0.2245",
"0.8135",
"0.2245",
"0.8135",
"0.2382",
"0.513",
"0.2382"
],
[
"0.511",
"0.2405",
"0.879",
"0.2393",
"0.879",
"0.2542",
"0.511",
"0.2546"
],
[
"0.126",
"0.253",
"0.1621",
"0.2505",
"0.1628",
"0.2625",
"0.1276",
"0.2642"
],
[
"0.3855",
"0.26",
"0.392",
"0.26",
"0.392",
"0.2632",
"0.3855",
"0.2632"
],
[
"0.513",
"0.257",
"0.881",
"0.257",
"0.881",
"0.2708",
"0.513",
"0.2708"
],
[
"0.1276",
"0.2654",
"0.1368",
"0.2654",
"0.1368",
"0.275",
"0.1276",
"0.275"
],
[
"0.511",
"0.2715",
"0.879",
"0.2715",
"0.879",
"0.2852",
"0.511",
"0.2852"
],
[
"0.1411",
"0.2808",
"0.1545",
"0.2808",
"0.1545",
"0.2852",
"0.1411",
"0.2852"
],
[
"0.2527",
"0.279",
"0.2795",
"0.279",
"0.2795",
"0.2861",
"0.2527",
"0.2861"
],
[
"0.1561",
"0.2861",
"0.2527",
"0.2861",
"0.2527",
"0.2964",
"0.1561",
"0.2964"
],
[
"0.2747",
"0.2861",
"0.372",
"0.2861",
"0.372",
"0.2964",
"0.2747",
"0.2964"
],
[
"0.3855",
"0.2808",
"0.3955",
"0.2808",
"0.3955",
"0.2852",
"0.3855",
"0.2852"
],
[
"0.3938",
"0.2861",
"0.4937",
"0.2861",
"0.4937",
"0.2964",
"0.3938",
"0.2964"
],
[
"0.513",
"0.2893",
"0.882",
"0.2893",
"0.882",
"0.3027",
"0.513",
"0.3027"
],
[
"0.3787",
"0.2998",
"0.3972",
"0.2998",
"0.3972",
"0.308",
"0.3787",
"0.308"
],
[
"0.12427",
"0.308",
"0.1628",
"0.3135",
"0.1545",
"0.3467",
"0.11584",
"0.3416"
],
[
"0.513",
"0.3047",
"0.881",
"0.3047",
"0.881",
"0.3184",
"0.513",
"0.3184"
],
[
"0.2585",
"0.3184",
"0.2805",
"0.3184",
"0.2805",
"0.3267",
"0.2585",
"0.3267"
],
[
"0.3804",
"0.3225",
"0.3955",
"0.3225",
"0.3955",
"0.3308",
"0.3804",
"0.3308"
],
[
"0.5146",
"0.3206",
"0.879",
"0.3206",
"0.879",
"0.3342",
"0.5146",
"0.3342"
],
[
"0.1411",
"0.3403",
"0.157",
"0.3433",
"0.1528",
"0.3516",
"0.1368",
"0.3486"
],
[
"0.513",
"0.3374",
"0.881",
"0.3374",
"0.881",
"0.3508",
"0.513",
"0.3508"
],
[
"0.2603",
"0.3445",
"0.2747",
"0.3445",
"0.2747",
"0.3486",
"0.2603",
"0.3486"
],
[
"0.419",
"0.3486",
"0.492",
"0.3486",
"0.492",
"0.3594",
"0.419",
"0.3594"
],
[
"0.2737",
"0.3499",
"0.3745",
"0.3499",
"0.3745",
"0.3599",
"0.2737",
"0.3599"
],
[
"0.157",
"0.3508",
"0.1797",
"0.3508",
"0.1797",
"0.3594",
"0.157",
"0.3594"
],
[
"0.1797",
"0.3499",
"0.2527",
"0.3499",
"0.2527",
"0.3599",
"0.1797",
"0.3599"
],
[
"0.3955",
"0.3508",
"0.4248",
"0.3508",
"0.4248",
"0.3594",
"0.3955",
"0.3594"
],
[
"0.511",
"0.3528",
"0.711",
"0.3528",
"0.711",
"0.3665",
"0.511",
"0.3665"
],
[
"0.1931",
"0.3599",
"0.2091",
"0.3599",
"0.2091",
"0.3694",
"0.1931",
"0.3694"
],
[
"0.3132",
"0.3599",
"0.33",
"0.3599",
"0.33",
"0.3694",
"0.3132",
"0.3694"
],
[
"0.4358",
"0.3599",
"0.451",
"0.3599",
"0.451",
"0.3694",
"0.4358",
"0.3694"
],
[
"0.1184",
"0.3884",
"0.4878",
"0.3884",
"0.4878",
"0.3984",
"0.1184",
"0.3984"
],
[
"0.6406",
"0.3914",
"0.69",
"0.3914",
"0.69",
"0.402",
"0.6406",
"0.402"
],
[
"0.738",
"0.3914",
"0.8076",
"0.3914",
"0.8076",
"0.402",
"0.738",
"0.402"
],
[
"0.1142",
"0.3997",
"0.309",
"0.401",
"0.309",
"0.415",
"0.1142",
"0.4146"
],
[
"0.6206",
"0.3967",
"0.6616",
"0.3967",
"0.6616",
"0.4104",
"0.6206",
"0.4104"
],
[
"0.771",
"0.3997",
"0.804",
"0.3997",
"0.804",
"0.408",
"0.771",
"0.408"
],
[
"0.711",
"0.4175",
"0.729",
"0.4175",
"0.729",
"0.4258",
"0.711",
"0.4258"
],
[
"0.5635",
"0.4258",
"0.5786",
"0.4258",
"0.5786",
"0.434",
"0.5635",
"0.434"
],
[
"0.712",
"0.43",
"0.727",
"0.43",
"0.727",
"0.4395",
"0.712",
"0.4395"
],
[
"0.11755",
"0.4495",
"0.486",
"0.4495",
"0.486",
"0.4631",
"0.11755",
"0.4631"
],
[
"0.714",
"0.4448",
"0.727",
"0.4448",
"0.727",
"0.4507",
"0.714",
"0.4507"
],
[
"0.586",
"0.4578",
"0.687",
"0.4578",
"0.687",
"0.4697",
"0.586",
"0.4697"
],
[
"0.735",
"0.459",
"0.8345",
"0.459",
"0.8345",
"0.4697",
"0.735",
"0.4697"
],
[
"0.1184",
"0.4673",
"0.4802",
"0.4673",
"0.4802",
"0.478",
"0.1184",
"0.478"
],
[
"0.5483",
"0.4602",
"0.5635",
"0.4602",
"0.5635",
"0.4954",
"0.5483",
"0.4954"
],
[
"0.6245",
"0.4685",
"0.657",
"0.4685",
"0.657",
"0.4756",
"0.6245",
"0.4756"
],
[
"0.7715",
"0.4697",
"0.802",
"0.4697",
"0.802",
"0.4739",
"0.7715",
"0.4739"
],
[
"0.1335",
"0.4822",
"0.4844",
"0.4822",
"0.4844",
"0.4954",
"0.1335",
"0.4954"
],
[
"0.705",
"0.488",
"0.727",
"0.488",
"0.727",
"0.4954",
"0.705",
"0.4954"
],
[
"0.1184",
"0.4976",
"0.4844",
"0.4976",
"0.4844",
"0.511",
"0.1184",
"0.511"
],
[
"0.5635",
"0.4963",
"0.58",
"0.4963",
"0.58",
"0.5044",
"0.5635",
"0.5044"
],
[
"0.705",
"0.502",
"0.727",
"0.502",
"0.727",
"0.5103",
"0.705",
"0.5103"
],
[
"0.11755",
"0.514",
"0.482",
"0.514",
"0.482",
"0.528",
"0.11755",
"0.528"
],
[
"0.711",
"0.5186",
"0.7256",
"0.5186",
"0.7256",
"0.5215",
"0.711",
"0.5215"
],
[
"0.586",
"0.527",
"0.686",
"0.527",
"0.686",
"0.538",
"0.586",
"0.538"
],
[
"0.735",
"0.528",
"0.833",
"0.528",
"0.833",
"0.538",
"0.735",
"0.538"
],
[
"0.634",
"0.539",
"0.6465",
"0.539",
"0.6465",
"0.549",
"0.634",
"0.549"
],
[
"0.1184",
"0.53",
"0.486",
"0.53",
"0.486",
"0.5435",
"0.1184",
"0.5435"
],
[
"0.11755",
"0.546",
"0.486",
"0.5444",
"0.486",
"0.5586",
"0.11755",
"0.56"
],
[
"0.7773",
"0.5405",
"0.7925",
"0.536",
"0.797",
"0.546",
"0.784",
"0.55"
],
[
"0.11755",
"0.5625",
"0.4844",
"0.5625",
"0.4844",
"0.576",
"0.11755",
"0.576"
],
[
"0.51",
"0.5635",
"0.882",
"0.564",
"0.882",
"0.579",
"0.51",
"0.5776"
],
[
"0.1184",
"0.579",
"0.4844",
"0.579",
"0.4844",
"0.593",
"0.1184",
"0.593"
],
[
"0.51",
"0.579",
"0.882",
"0.579",
"0.882",
"0.593",
"0.51",
"0.593"
],
[
"0.1184",
"0.594",
"0.4844",
"0.594",
"0.4844",
"0.608",
"0.1184",
"0.608"
],
[
"0.511",
"0.5933",
"0.882",
"0.5933",
"0.882",
"0.607",
"0.511",
"0.607"
],
[
"0.1184",
"0.61",
"0.4844",
"0.61",
"0.4844",
"0.6235",
"0.1184",
"0.6235"
],
[
"0.1184",
"0.627",
"0.486",
"0.627",
"0.486",
"0.64",
"0.1184",
"0.64"
],
[
"0.11755",
"0.641",
"0.4844",
"0.6426",
"0.4844",
"0.6567",
"0.11755",
"0.6562"
],
[
"0.5283",
"0.6426",
"0.881",
"0.6426",
"0.881",
"0.6562",
"0.5283",
"0.6562"
],
[
"0.1184",
"0.659",
"0.4836",
"0.659",
"0.4836",
"0.673",
"0.1184",
"0.673"
],
[
"0.513",
"0.66",
"0.879",
"0.66",
"0.879",
"0.6733",
"0.513",
"0.6733"
],
[
"0.1184",
"0.675",
"0.3284",
"0.675",
"0.3284",
"0.6885",
"0.1184",
"0.6885"
],
[
"0.513",
"0.676",
"0.534",
"0.676",
"0.534",
"0.6904",
"0.513",
"0.6904"
],
[
"0.392",
"0.712",
"0.4375",
"0.7217",
"0.4307",
"0.733",
"0.388",
"0.7236"
],
[
"0.511",
"0.712",
"0.831",
"0.712",
"0.831",
"0.727",
"0.511",
"0.727"
],
[
"0.5146",
"0.7437",
"0.881",
"0.7437",
"0.881",
"0.7573",
"0.5146",
"0.7573"
],
[
"0.5146",
"0.7603",
"0.881",
"0.7603",
"0.881",
"0.774",
"0.5146",
"0.774"
],
[
"0.2007",
"0.775",
"0.2317",
"0.775",
"0.2317",
"0.7793",
"0.2007",
"0.7793"
],
[
"0.3855",
"0.775",
"0.4048",
"0.775",
"0.4048",
"0.7793",
"0.3855",
"0.7793"
],
[
"0.4106",
"0.7715",
"0.465",
"0.7715",
"0.465",
"0.782",
"0.4106",
"0.782"
],
[
"0.5146",
"0.7754",
"0.879",
"0.7754",
"0.879",
"0.789",
"0.5146",
"0.789"
],
[
"0.5146",
"0.792",
"0.8774",
"0.792",
"0.8774",
"0.8027",
"0.5146",
"0.8027"
],
[
"0.5146",
"0.808",
"0.8696",
"0.808",
"0.8696",
"0.822",
"0.5146",
"0.822"
],
[
"0.2688",
"0.8276",
"0.3594",
"0.8276",
"0.3594",
"0.8384",
"0.2688",
"0.8384"
],
[
"0.3745",
"0.8276",
"0.4685",
"0.8267",
"0.4685",
"0.8384",
"0.3745",
"0.839"
],
[
"0.5146",
"0.8247",
"0.881",
"0.8247",
"0.881",
"0.8384",
"0.5146",
"0.8384"
],
[
"0.2007",
"0.839",
"0.2091",
"0.839",
"0.2091",
"0.844",
"0.2007",
"0.844"
],
[
"0.1628",
"0.83",
"0.251",
"0.83",
"0.251",
"0.8374",
"0.1628",
"0.8374"
],
[
"0.3074",
"0.8403",
"0.318",
"0.8403",
"0.318",
"0.844",
"0.3074",
"0.844"
],
[
"0.4165",
"0.839",
"0.4265",
"0.839",
"0.4265",
"0.844",
"0.4165",
"0.844"
],
[
"0.513",
"0.8403",
"0.879",
"0.8403",
"0.879",
"0.854",
"0.513",
"0.854"
],
[
"0.11584",
"0.8623",
"0.4836",
"0.8623",
"0.4836",
"0.876",
"0.11584",
"0.876"
],
[
"0.517",
"0.857",
"0.881",
"0.857",
"0.881",
"0.8706",
"0.517",
"0.8706"
],
[
"0.1184",
"0.8774",
"0.486",
"0.8774",
"0.486",
"0.888",
"0.1184",
"0.888"
],
[
"0.513",
"0.8726",
"0.881",
"0.8726",
"0.881",
"0.8857",
"0.513",
"0.8857"
],
[
"0.11755",
"0.89",
"0.4902",
"0.89",
"0.4902",
"0.904",
"0.11755",
"0.904"
],
[
"0.513",
"0.8896",
"0.882",
"0.8896",
"0.882",
"0.903",
"0.513",
"0.903"
],
[
"0.11584",
"0.905",
"0.4165",
"0.905",
"0.4165",
"0.9185",
"0.11584",
"0.9185"
],
[
"0.513",
"0.905",
"0.8613",
"0.905",
"0.8613",
"0.9185",
"0.513",
"0.9185"
],
[
"0.492",
"0.9414",
"0.5054",
"0.9414",
"0.5054",
"0.9517",
"0.492",
"0.9517"
]
] | [
"ExactMatch",
"GPT4Score",
"3.2S2 : IFT models as Task-Specific Solvers",
"ore",
"lun",
"conll",
"mnli",
"sst2",
"设置。我们使用四种模型架构，对于每种架构",
"",
"0.5",
"0.5",
"我们使用基本模型来使用合成的Alpaca数据",
"集训练IFT模型变体。然后，我们在从目标任",
"0.0",
"0.0",
"0100101102103",
"0100:101102103",
"0.0",
"010010110210",
"务中提取的N样本的子集上微调基本模型及",
"",
"其IFT变体。此设置模拟了一个工业场景，在",
"",
"80.5",
"0.5",
"0.5",
"该场景中，有限的数据可用于在唯一任务上专",
"",
"门化模型，并评估在对目标任务进行微调之前",
"0100101102103",
"010010110210",
"0100101102103",
"对基本模型进行指令调整是否有好处。",
"结果。Fig.4演示了IFT模型在低数据场景中",
"00.5",
"",
"表现出增强的性能（当10≤N≤200时）。直",
"0",
"观上，IFT模型能够更好地利用提示中给出的",
"0.0",
"0.0",
"0100101102103",
"0100101102103",
".0",
"0100101102103",
"任务描述，从而实现提升的零拍性能（Scao and",
"1.0",
"",
"Rush,2021）。当使用少量样本进行微调时，这",
"0.5",
"0.5",
"可以补充和引导模型训练。当更多的样本可用",
"0.0",
"时（N≥200），任务模式足够清晰，提示和",
"0.0",
"10110210",
"0100101102103",
"0.100",
"101102103",
"0100",
"IFT训练阶段的好处消失",
"N",
"N",
"N",
"Figure2:Incorporating0<N<1000realtask sam-",
"Alpaca",
"Base Model",
"ples into IFT model training",
"Bloom",
"Falcor",
"0.9",
"0.8",
"0.8",
"Sec.C.1.2）中，导致性能几乎没有下降(<1%)。",
"0.7",
"010010:102103104",
"0100101102103104",
"利用合成数据学习格式化。我们的发现提出了",
"",
"Llama",
"Pth:",
"-种简单的方法来优化真实示例的使用：在依",
"0.75",
"靠真实样本来增强整体模型性能之前，使用合",
"0.8",
"0.50",
"成示例来帮助模型掌握所需的格式。我们重复",
"",
"0100101102103104",
"0100101102103104",
"N",
"先前的实验，用GPT4合成生成的样本（S）或",
"带有随机标签（R）的合成样本（Fig.3）Exact",
"N",
"Match替换N人类注释的目标任务训练样本",
"Figure 4: GPT4 score on SST-2 test set after finetuning",
"(H）显示合成或随机标记的数据确实可以用于",
"with 0 ≤ N ≤ 1000 samples on a (base) LM or an IFT",
"学习所需的格式，尽管质量更好的人类数据最",
"model. Further experiments can be found in Sec. C.2.2",
"终会在更多样本的情况下产生更好的结果。在",
"(S+H)中，我们训练100个合成样本，然后训",
"练N人类注释样本。该技术使模型在接受高",
"这一发现与Sec.3.1中给出的结果一致，强",
"质量数据训练之前能够掌握格式，大大提高了",
"调了合成数据集提高工业场景中数据效率的潜",
"人类注释数据的采样效率。",
"力",
"S+Honll",
"4Key Takeaways for Practitioners",
"利用LLM进行评估。评估IFT模型是具有挑",
"战性的，因为它要求跨任务和格式不可知论的",
"01102",
"100",
"10110210",
"可比性，标准指标与之斗争。虽然LLM评分",
"并不理想（Sec.5中的限制）但这是从业者应",
"该添加到他们的武器库中的一个很好的选择。",
"0100101102103",
"010010102103",
"利用合成数据实现高效学习。基于LLM的评",
"N",
"0100101102103",
"M",
"A",
"估揭示了这样一个事实，即利用合成数据提供",
"Figure 3: Incorporating 0 ≤ N ≤ 1000 (H)uman",
"了一种快速且经济高效的方法，可以在低数据",
"(S)yntheticand (R)andomlylabeled syntheticdata sam-",
"状态下掌握格式，而不会降低性能。这种方法",
"ples in IFT training set. (S+H) is trained on 100 syn-",
"在各种情况下都是可行的，它提供了一个机会，",
"thetic samples, then N human data samples.",
"可以更有效地利用可能有限的专家注释数据",
"4"
] | {
"arxiv_id": "2310.14103",
"language": "zh",
"page_idx": 4
} |
||
2404.16223-zh-0 | 2404.16223-zh-0.jpg | 深 RAW 图像超分辨率。NTIRE2024 挑战调查
$\begin{array}{cc}\text{Marcos V.Conde}^{*\dagger}&\text{Florin-Alexandru Vasluianu }^{*}&\text{Radu Timofte}^{*}\\\text{Jianxing Zhang}&\text{Jiaa Li}&\text{Fan Wang}&\text{Xiaopeng Li}&\text{Zikun Liu}&\text{Hyunhee Park}\\\text{Sejun Song}&\text{Changho Kim}&\text{Zhijuan Huang}&\text{Hongyuan Yu}&\text{Cleng Wan}\\\text{Wending Xiang}&\text{Jiannin }\end{array}$
$\operatorname{Abstract}$
This paper reviews the NTIRE 2024 RAW Image Super-Resolution Challenge, highlighting the proposed solutions and results. New methods for RAW SuperResolution could be essential in modern Image Signal Processing (ISP) pipelines, however, this problem is not as explored as in the RGB domain. Th goal of this challenge is to upscale RAW Bayer images by 2x, considering unknown degradations such as noise and blur. In the challenge, a total of 230 participants registered, and 45 submitted results during thee challenge period. The performance of the top-5 submissions is reviewed and provided here as a gauge for the current state-of-the-art in RAW Image Super-Resolution.
## 1. Introduction
RAWImageSuper-Resolution 代表了一个活跃的研究方向，旨在扩大硬件特定的 RAW 图像表示，同时处理硬件特性，通常取决于相机产品的技术实现。相机图像处理信号 (ISP)实施缺乏标准化，导致相机细分市场种类繁多，通过图像处理算法[?????]进行了大量校正，克服了各种设备的硬件限制
RAW 图像表示图像信号的离散化和量化表示。当然，上述两种操作都取决于传感器的性质，具有不同的传感器类型，不同的空间分辨率或颜色分辨率(用于量化的位数)。自然，通常在尺寸，电源和使用的光学器件方面受到极端限制的便携式相机设备在实现的图像
* Marcos V. Conde ( t corresponding author, project lead), Florin-Alexandru Vasluianu, amd Radu Timofte are the challenge organizers, while the other authors participated in the challenge and survey.
* University of Würzburg, CAIDAS &IFI, Computer Vision
Lab.
NTIRE 2024 webpage: https://cvlai.net/ntire/2024.
Code: https://github.com/mv-lab/AISP
质量方面受到限制，通常采用低颜色分辨率和空间分辨率。然而，这些特定的硬件实现目前在新获得的设备方面处于市场领先地位，相机变得无处不在，普遍可访间。因此，图像信息空间的特征还在于与这些设备相对应的图像的高可用性，复杂的 ISP 系统$\mathbb{E}[????]$ 将RAW 图像表示映射到感知上有意义的 RGB 对应物
RAW 图像升级在专业摄影领域仍然很重要。鉴于基于互联网的图像传输服务的强劲发展，高分辨率图像的驱动力现在比以往任何时候都要高。但是，建立专业的摄影或录像设置是一项艰巨的挑战，因为市场被各种特定于应用程序的参数所划分，而购置成本仍然是新投资的主要限制因素。因此，重要的是要开发对影响低成本业余摄影系统的各种硬件限制具有鲁棒性的算法，以匹配高度专业应用中使用的高分辨率传感器的特性。
RAW 信息非常重要，因为 RAW 数据与场景辐射直接相关，离散化和量化是影响摄影中自然连续辐射信号测量的唯一非线性操作。RAW 数据的这一特性对于分析典型的图像退化(如采集噪声或图像模糊)非常有益。远离传感器表示，ISP 表示一系列高度非线性的操作，发生不可挽回的信息丢失$\mathbb{E}[??]$。这进一步使图像恢复任务复杂化，复杂的算法需要每个可实现的基于信号的方差$\mathbb{E}[??]$。考虑到上述所有因素，RAW 图像处理比标准 sRGB 表示具有显着的优势，在许多低级图像应用中具有卓越的性能，例如图像去噪$\mathbb{E} [ ?$ ? ?\_,去模糊$\mathbb{E}[?]$,曝光调整$\mathbb{E}[?]$ 和图像超分辨${\text{率}\mathbb{E}[?~?~?~?~?~]}$
由于在硬件实现级别上缺乏标准化，RAW 图像的特征在于供应商甚至产品特定的属性，这些属性只能通过私人解释，而摄影产品则被实施专利或商业秘密所屏蔽。再加上标准 sRGB 表示的丰富性，大多数现有的高复杂性图像恢复算法 [????] 都是专门为压缩或未压缩的 RGB 图像或视频而设计的
作为图像恢复的子任务，尖端的单图像超分辨率(SISR)算法$\mathbb{E}[???]$遵循相同的数据特定，即使它们依赖于深度卷积网络或 Transformer$[\mathbb{F}[??]$架构。这些算法的最大缺点之一是它们受到用于优化的数据 | [
[
"0.2467",
"0.1338",
"0.722",
"0.1338",
"0.722",
"0.1483",
"0.2467",
"0.1483"
],
[
"0.1602",
"0.1812",
"0.3325",
"0.1787",
"0.3325",
"0.1976",
"0.1602",
"0.1995"
],
[
"0.38",
"0.1831",
"0.6274",
"0.1831",
"0.6274",
"0.1976",
"0.38",
"0.1976"
],
[
"0.6763",
"0.1831",
"0.8096",
"0.1831",
"0.8096",
"0.1976",
"0.6763",
"0.1976"
],
[
"0.0972",
"0.1976",
"0.2361",
"0.1995",
"0.2361",
"0.2172",
"0.0972",
"0.2153"
],
[
"0.2688",
"0.1995",
"0.3284",
"0.1995",
"0.3284",
"0.214",
"0.2688",
"0.214"
],
[
"0.3555",
"0.1995",
"0.4592",
"0.2008",
"0.4592",
"0.2153",
"0.3555",
"0.214"
],
[
"0.478",
"0.1995",
"0.722",
"0.1976",
"0.722",
"0.214",
"0.478",
"0.2166"
],
[
"0.751",
"0.2008",
"0.8716",
"0.2008",
"0.8716",
"0.214",
"0.751",
"0.214"
],
[
"0.4167",
"0.214",
"0.5522",
"0.2172",
"0.5522",
"0.2349",
"0.4167",
"0.2317"
],
[
"0.4534",
"0.2052",
"0.4795",
"0.2052",
"0.4795",
"0.2122",
"0.4534",
"0.2122"
],
[
"0.2573",
"0.2166",
"0.3857",
"0.2184",
"0.3857",
"0.2329",
"0.2573",
"0.2311"
],
[
"0.585",
"0.2172",
"0.716",
"0.2153",
"0.716",
"0.2329",
"0.585",
"0.2349"
],
[
"0.749",
"0.2172",
"0.8506",
"0.2172",
"0.8506",
"0.2317",
"0.749",
"0.2317"
],
[
"0.1274",
"0.2329",
"0.2646",
"0.2349",
"0.2646",
"0.2524",
"0.1274",
"0.2507"
],
[
"0.1217",
"0.2184",
"0.2181",
"0.2184",
"0.2181",
"0.2329",
"0.1217",
"0.2329"
],
[
"0.4387",
"0.2329",
"0.5513",
"0.2349",
"0.5513",
"0.2524",
"0.4387",
"0.2507"
],
[
"0.7656",
"0.2343",
"0.8423",
"0.2361",
"0.8423",
"0.2507",
"0.7656",
"0.2488"
],
[
"0.5884",
"0.2361",
"0.7295",
"0.2361",
"0.7295",
"0.2507",
"0.5884",
"0.2507"
],
[
"0.3022",
"0.2349",
"0.3987",
"0.2349",
"0.3987",
"0.25",
"0.3022",
"0.25"
],
[
"0.744",
"0.2393",
"0.761",
"0.2393",
"0.761",
"0.2468",
"0.744",
"0.2468"
],
[
"0.2852",
"0.2507",
"0.4019",
"0.2524",
"0.4019",
"0.2708",
"0.2852",
"0.2683"
],
[
"0.4429",
"0.252",
"0.54",
"0.254",
"0.54",
"0.2695",
"0.4429",
"0.2676"
],
[
"0.5786",
"0.25",
"0.6978",
"0.2524",
"0.6978",
"0.2708",
"0.5786",
"0.2676"
],
[
"0.7363",
"0.254",
"0.835",
"0.254",
"0.835",
"0.2683",
"0.7363",
"0.2683"
],
[
"0.705",
"0.26",
"0.71",
"0.26",
"0.71",
"0.2651",
"0.705",
"0.2651"
],
[
"0.1356",
"0.254",
"0.2467",
"0.254",
"0.2467",
"0.2676",
"0.1356",
"0.2676"
],
[
"0.614",
"0.2695",
"0.7134",
"0.2715",
"0.7134",
"0.286",
"0.614",
"0.2842"
],
[
"0.746",
"0.2708",
"0.8735",
"0.2708",
"0.8735",
"0.2842",
"0.746",
"0.2842"
],
[
"0.2303",
"0.2715",
"0.3447",
"0.2715",
"0.3447",
"0.286",
"0.2303",
"0.286"
],
[
"0.0989",
"0.2715",
"0.2035",
"0.2715",
"0.2035",
"0.286",
"0.0989",
"0.286"
],
[
"0.3774",
"0.2727",
"0.4534",
"0.2727",
"0.4534",
"0.2842",
"0.3774",
"0.2842"
],
[
"0.492",
"0.2727",
"0.5825",
"0.2727",
"0.5825",
"0.2842",
"0.492",
"0.2842"
],
[
"0.726",
"0.276",
"0.739",
"0.276",
"0.739",
"0.279",
"0.726",
"0.279"
],
[
"0.6978",
"0.286",
"0.785",
"0.2886",
"0.785",
"0.3062",
"0.6978",
"0.3037"
],
[
"0.2157",
"0.276",
"0.2336",
"0.276",
"0.2336",
"0.279",
"0.2157",
"0.279"
],
[
"0.2876",
"0.289",
"0.384",
"0.289",
"0.384",
"0.3037",
"0.2876",
"0.3037"
],
[
"0.4314",
"0.2886",
"0.5034",
"0.2886",
"0.5034",
"0.303",
"0.4314",
"0.303"
],
[
"0.55",
"0.289",
"0.653",
"0.289",
"0.653",
"0.3037",
"0.55",
"0.3037"
],
[
"0.2385",
"0.3403",
"0.312",
"0.3403",
"0.312",
"0.3547",
"0.2385",
"0.3547"
],
[
"0.1569",
"0.2903",
"0.2385",
"0.2903",
"0.2385",
"0.3018",
"0.1569",
"0.3018"
],
[
"0.505",
"0.3423",
"0.8896",
"0.3423",
"0.8896",
"0.356",
"0.505",
"0.356"
],
[
"0.5034",
"0.3567",
"0.8896",
"0.3567",
"0.8896",
"0.3706",
"0.5034",
"0.3706"
],
[
"0.0989",
"0.3726",
"0.4673",
"0.3726",
"0.4673",
"0.387",
"0.0989",
"0.387"
],
[
"0.505",
"0.3726",
"0.8906",
"0.3726",
"0.8906",
"0.3865",
"0.505",
"0.3865"
],
[
"0.0801",
"0.387",
"0.468",
"0.387",
"0.468",
"0.4016",
"0.0801",
"0.4016"
],
[
"0.505",
"0.387",
"0.8906",
"0.387",
"0.8906",
"0.401",
"0.505",
"0.401"
],
[
"0.0809",
"0.4028",
"0.4673",
"0.4028",
"0.4673",
"0.4167",
"0.0809",
"0.4167"
],
[
"0.5034",
"0.4016",
"0.8906",
"0.4016",
"0.8906",
"0.4167",
"0.5034",
"0.4167"
],
[
"0.0809",
"0.4185",
"0.468",
"0.4185",
"0.468",
"0.432",
"0.0809",
"0.432"
],
[
"0.5024",
"0.4167",
"0.871",
"0.4172",
"0.871",
"0.432",
"0.5024",
"0.4312"
],
[
"0.0809",
"0.4343",
"0.468",
"0.4343",
"0.468",
"0.4475",
"0.0809",
"0.4475"
],
[
"0.523",
"0.433",
"0.8906",
"0.433",
"0.8906",
"0.4463",
"0.523",
"0.4463"
],
[
"0.11926",
"0.449",
"0.4697",
"0.449",
"0.4697",
"0.4622",
"0.11926",
"0.4622"
],
[
"0.5063",
"0.4495",
"0.8896",
"0.4495",
"0.8896",
"0.4602",
"0.5063",
"0.4602"
],
[
"0.0809",
"0.4634",
"0.4673",
"0.4634",
"0.4673",
"0.4778",
"0.0809",
"0.4778"
],
[
"0.08417",
"0.452",
"0.125",
"0.452",
"0.125",
"0.459",
"0.08417",
"0.459"
],
[
"0.5034",
"0.4634",
"0.8906",
"0.4634",
"0.8906",
"0.4766",
"0.5034",
"0.4766"
],
[
"0.0825",
"0.4797",
"0.4673",
"0.4797",
"0.4673",
"0.4905",
"0.0825",
"0.4905"
],
[
"0.5024",
"0.4766",
"0.8896",
"0.4778",
"0.8896",
"0.4924",
"0.5024",
"0.4912"
],
[
"0.0801",
"0.4937",
"0.4673",
"0.4937",
"0.4673",
"0.5083",
"0.0801",
"0.5083"
],
[
"0.5034",
"0.4937",
"0.8906",
"0.4937",
"0.8906",
"0.507",
"0.5034",
"0.507"
],
[
"0.0784",
"0.5083",
"0.468",
"0.509",
"0.468",
"0.523",
"0.0784",
"0.5215"
],
[
"0.505",
"0.509",
"0.8906",
"0.509",
"0.8906",
"0.523",
"0.505",
"0.523"
],
[
"0.0801",
"0.5234",
"0.468",
"0.523",
"0.468",
"0.537",
"0.0801",
"0.5386"
],
[
"0.5063",
"0.526",
"0.8896",
"0.526",
"0.8896",
"0.536",
"0.5063",
"0.536"
],
[
"0.0801",
"0.5386",
"0.468",
"0.5386",
"0.468",
"0.552",
"0.0801",
"0.552"
],
[
"0.5034",
"0.5386",
"0.8906",
"0.5386",
"0.8906",
"0.552",
"0.5034",
"0.552"
],
[
"0.0809",
"0.5547",
"0.433",
"0.5547",
"0.433",
"0.5684",
"0.0809",
"0.5684"
],
[
"0.505",
"0.5547",
"0.5415",
"0.5547",
"0.5415",
"0.5664",
"0.505",
"0.5664"
],
[
"0.5215",
"0.5674",
"0.8906",
"0.5693",
"0.8906",
"0.5835",
"0.5215",
"0.5806"
],
[
"0.5034",
"0.584",
"0.8906",
"0.584",
"0.8906",
"0.5986",
"0.5034",
"0.5986"
],
[
"0.0801",
"0.601",
"0.2173",
"0.601",
"0.2173",
"0.6143",
"0.0801",
"0.6143"
],
[
"0.5063",
"0.5996",
"0.8896",
"0.5996",
"0.8896",
"0.613",
"0.5063",
"0.613"
],
[
"0.505",
"0.6143",
"0.888",
"0.6157",
"0.888",
"0.629",
"0.505",
"0.6274"
],
[
"0.0801",
"0.6255",
"0.468",
"0.6255",
"0.468",
"0.6396",
"0.0801",
"0.6396"
],
[
"0.505",
"0.629",
"0.8896",
"0.629",
"0.8896",
"0.6426",
"0.505",
"0.6426"
],
[
"0.0825",
"0.6416",
"0.468",
"0.6416",
"0.468",
"0.655",
"0.0825",
"0.655"
],
[
"0.505",
"0.6445",
"0.8896",
"0.6445",
"0.8896",
"0.6577",
"0.505",
"0.6577"
],
[
"0.0825",
"0.6577",
"0.4673",
"0.6577",
"0.4673",
"0.6685",
"0.0825",
"0.6685"
],
[
"0.5034",
"0.6577",
"0.8906",
"0.659",
"0.8906",
"0.674",
"0.5034",
"0.6724"
],
[
"0.0825",
"0.672",
"0.468",
"0.672",
"0.468",
"0.685",
"0.0825",
"0.685"
],
[
"0.505",
"0.675",
"0.888",
"0.675",
"0.888",
"0.688",
"0.505",
"0.688"
],
[
"0.0809",
"0.685",
"0.468",
"0.686",
"0.468",
"0.7007",
"0.0809",
"0.6997"
],
[
"0.5034",
"0.688",
"0.8906",
"0.6895",
"0.8906",
"0.704",
"0.5034",
"0.7026"
],
[
"0.0809",
"0.702",
"0.327",
"0.702",
"0.327",
"0.7153",
"0.0809",
"0.7153"
],
[
"0.505",
"0.705",
"0.892",
"0.705",
"0.892",
"0.7183",
"0.505",
"0.7183"
],
[
"0.10046",
"0.7173",
"0.468",
"0.7173",
"0.468",
"0.732",
"0.10046",
"0.732"
],
[
"0.5034",
"0.7197",
"0.8906",
"0.7197",
"0.8906",
"0.7344",
"0.5034",
"0.7344"
],
[
"0.0809",
"0.732",
"0.4656",
"0.732",
"0.4656",
"0.7456",
"0.0809",
"0.7456"
],
[
"0.5034",
"0.735",
"0.621",
"0.735",
"0.621",
"0.7485",
"0.5034",
"0.7485"
],
[
"0.0809",
"0.7476",
"0.468",
"0.7476",
"0.468",
"0.761",
"0.0809",
"0.761"
],
[
"0.524",
"0.7505",
"0.8906",
"0.7505",
"0.8906",
"0.7646",
"0.524",
"0.7646"
],
[
"0.0825",
"0.763",
"0.4673",
"0.763",
"0.4673",
"0.7764",
"0.0825",
"0.7764"
],
[
"0.5063",
"0.7676",
"0.888",
"0.7676",
"0.888",
"0.778",
"0.5063",
"0.778"
],
[
"0.0809",
"0.778",
"0.468",
"0.778",
"0.468",
"0.7915",
"0.0809",
"0.7915"
],
[
"0.505",
"0.781",
"0.8906",
"0.781",
"0.8906",
"0.7944",
"0.505",
"0.7944"
],
[
"0.5024",
"0.7944",
"0.8896",
"0.797",
"0.8896",
"0.81",
"0.5024",
"0.808"
],
[
"0.10455",
"0.8047",
"0.464",
"0.8047",
"0.464",
"0.815",
"0.10455",
"0.815"
],
[
"0.0801",
"0.815",
"0.464",
"0.8164",
"0.464",
"0.8276",
"0.0801",
"0.827"
],
[
"0.5034",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.5034",
"0.8247"
],
[
"0.0801",
"0.8276",
"0.4656",
"0.827",
"0.4656",
"0.8403",
"0.0801",
"0.8413"
],
[
"0.5034",
"0.8257",
"0.7812",
"0.8257",
"0.7812",
"0.839",
"0.5034",
"0.839"
],
[
"0.0801",
"0.8413",
"0.1814",
"0.8413",
"0.1814",
"0.8516",
"0.0801",
"0.8516"
],
[
"0.524",
"0.8413",
"0.8906",
"0.8413",
"0.8906",
"0.855",
"0.524",
"0.855"
],
[
"0.0825",
"0.851",
"0.468",
"0.851",
"0.468",
"0.8643",
"0.0825",
"0.8643"
],
[
"0.5024",
"0.855",
"0.8906",
"0.856",
"0.8906",
"0.8706",
"0.5024",
"0.869"
],
[
"0.0825",
"0.8657",
"0.10864",
"0.8657",
"0.10864",
"0.8735",
"0.0825",
"0.8735"
],
[
"0.0801",
"0.8755",
"0.398",
"0.8755",
"0.398",
"0.8896",
"0.0801",
"0.8896"
],
[
"0.505",
"0.872",
"0.8906",
"0.872",
"0.8906",
"0.8853",
"0.505",
"0.8853"
],
[
"0.0801",
"0.887",
"0.33",
"0.887",
"0.33",
"0.901",
"0.0801",
"0.901"
],
[
"0.5034",
"0.887",
"0.8906",
"0.887",
"0.8906",
"0.901",
"0.5034",
"0.901"
]
] | [
"深RAW图像超分辨率。NTIRE2024挑战调查",
"Marcos V. Conde *t",
"Florin-Alexandru Vasluianu *",
"Radu Timofte *",
" Jianxing Zhang",
"Jia Li",
"Fan Wang",
"Xiaopeng LiZikun Liu",
"Hyunhee Park",
" Zhijuan Huang",
"",
"Changho Kim",
"Hongyuan Yu",
"Cheng Wan",
"Wending Xiang",
"Sejun Song",
"Hang Zhong",
"Yue Sun",
"Qiaosong Zhang",
"Jiamin Lin",
"",
"Kunlong Zuo",
"Senyan Xu",
"Siyuan Jiang",
"Zhijing Sun",
"",
"Xuanwu Yin",
"Yimo Ning",
"Guanhua Zhao",
"Liangyan Li",
"Jiaying Zhu",
"Ke Chen",
"Yunzhe Li",
"",
"Yong Dou",
"",
"Jinyang Yu",
"Kele Xu",
"Qisheng Xu",
"Abstract",
"Jun Chen",
"质量方面受到限制，通常采用低颜色分辨率和空间分",
"辨率。然而，这些特定的硬件实现目前在新获得的设备",
"This paper reviews the NTIRE 2024 RAW Image",
"方面处于市场领先地位，相机变得无处不在，普遍可访",
"Super-Resolution Challenge, highlighting the proposed",
"问。因此，图像信息空间的特征还在于与这些设备相对",
"solutions and results. New methods for RAW Super-",
"应的图像的高可用性，复杂的ISP系统E[？？？？1将",
"Resolution could be essential in modern Image Signal",
"RAW图像表示映射到感知上有意义的RGB对应物",
"Processing (ISP) pipelines, however, this problem is",
"RAW图像升级在专业摄影领域仍然很重要。鉴于基",
"s explored as in the RGB domain. Th goal of",
"于互联网的图像传输服务的强劲发展，高分辨率图像",
"this challenge is to upscale RAW Bayer images by 2x,",
"notas",
"的驱动力现在比以往任何时候都要高。但是，建立专业",
"considering unknown degradations such as noise and",
"的摄影或录像设置是一项艰巨的挑战，因为市场被各",
"blur. In the challenge, a total of 230 participants reg-",
"种特定于应用程序的参数所划分，而购置成本仍然是",
"istered,and 45submitted results duringtheechallenge",
"新投资的主要限制因素。因此，重要的是要开发对影响",
"period. The performance of the top-5 submissions is",
"低成本业余摄影系统的各种硬件限制具有鲁棒性的算",
"reviewed and provided here as a gauge for the current",
"法，以匹配高度专业应用中使用的高分辨率传感器的",
"state-of-the-art in RAW Image Super-Resolution.",
"特性。",
"RAW信息非常重要，因为RAW数据与场景辐射直",
"接相关，离散化和量化是影响摄影中自然连续辐射信",
"1.Introduction",
"号测量的唯一非线性操作。RAW数据的这一特性对于",
"分析典型的图像退化（如采集噪声或图像模糊）非常有",
"RAWImageSuper-Resolution代表了一个活跃的研究",
"益。远离传感器表示，ISP表示一系列高度非线性的操",
"方向，旨在扩大硬件特定的RAW图像表示，同时处理",
"作，发生不可挽回的信息丢失E[？？1。这进一步使图",
"硬件特性，通常取决于相机产品的技术实现。相机图像",
"像恢复任务复杂化，复杂的算法需要每个可实现的基",
"处理信号（ISP）实施缺乏标准化，导致相机细分市场",
"于信号的方差E[？？1。考虑到上述所有因素，RAW",
"种类繁多，通过图像处理算法[？？？？？］进行了大量",
"图像处理比标准sRGB表示具有显着的优势，在许多",
"校正，克服了各种设备的硬件限制",
"低级图像应用中具有卓越的性能，例如图像去噪E[？？",
"RAW图像表示图像信号的离散化和量化表示。当",
"？］，去模糊E[？1，曝光调整E[？和图像超分辨",
"然，上述两种操作都取决于传感器的性质，具有不同的",
"率E[??？？？1",
"传感器类型，不同的空间分辨率或颜色分辨率（用于量",
"由于在硬件实现级别上缺乏标准化，RAW图像的特",
"化的位数）。自然，通常在尺寸，电源和使用的光学器",
"征在于供应商其至产品特定的属性，这些属性只能通",
"件方面受到极端限制的便携式相机设备在实现的图像",
"过私人解释，而摄影产品则被实施专利或商业秘密所",
"屏蔽。再加上标准sRGB表示的丰富性，大多数现有",
"*Marcos V.Conde (corresponding author,project lead)",
"Florin-Alexandru Vasluianu, amd Radu Timofte are the chal",
"的高复杂性图像恢复算法[？？？？1都是专门为压缩",
"lenge organizers, while the other authors participated in the chal-",
"或未压缩的RGB图像或视频而设计的",
"lengeand survey",
"作为图像恢复的子任务，尖端的单图像超分辨率",
"*University of Wurzburg, CAIDAS & IFI, Computer Vision",
"（SISR）算法E[？？？1遵循相同的数据特定，即使它",
"Lab.",
"NTIRE 2024 webpage: https://cvlai.net/ntire/2024.",
"们依赖于深度卷积网络或TransformerE[？？1架构。",
"Code: https://github.com/mv-lab/AISP",
"这些算法的最大缺点之一是它们受到用于优化的数据"
] | {
"arxiv_id": "2404.16223",
"language": "zh",
"page_idx": 0
} |
||
2404.16223-zh-1 | 2404.16223-zh-1.jpg | <FigureHere>
Figure 1. Samples of the NTIRE 2024 RAW Image Super-Resolution Challenge testing set.
Related Computer Vision Challenges , MATHX\_ Related Computer Vision Challenges 的任何深度图像质量评估$\mathbb{E}[?]$,夜间摄影渲染$\mathbb{E}[?]$,阴影去除$\mathbb{E}$ [? ] , 高效
2. NTIRE 2024 RAWSR Challenge
## $2. 1. \textbf{ Dataset}$
质量的限制。各种图像复原应用的特点是获取实域配对数据[??] 极其困难，推动了对现实和相关数据合成[??]的需求。鉴于高度非线性 ISP 特性也会改变退化外观的物理特性，因此在 sRGB 表示中准确建模特定于应用程序的退化非常困难。这代表了限制这些算法在实际应用程序部署中的性能的主要因素，在数据级别可观察到的差距很难通过无模型算法来克服
因此，研究 RAW 数据成为推动总体图像恢复性能改进趋势的关键步骤。因此，RAWImageSuper- Resolution 应用程序可以从增加的方差信号中受益，其算法对所涉及设备的精细架构属性具有鲁棒性。开发高效的算法，如在 sRGB 域[???] 中完善的算法，可以证明在高性能图像应用程序的开发中领先一步，具有成本效益的设备
因此，在这项工作中，我们将介绍为NTIRE2024RAW 图像超分辨率挑战赛提交的解决方案。我们提供有关挑战设置的信息，任务描述和挑战数据属性表征挑战数据集拆分。我们还列出了有关挑战参与者及其团队和隶属关系的信息
在 Sec. 2 中，我们描述了挑战数据集和评估，并讨论了总体结果。在 Sec. 3 中，我们提供了最佳解决方案的详细描述
挑战数据集基于 BSRAW [?] 。继以前的工作$\mathbb{E}[?$ ? ?] 之后，我们使用 AdobeMIT5K 数据集$\mathbb{E}[?]$ 中的图像，其中包括来自多个佳能和尼康数码单反相机的图像
手动过滤 DSLR 图像以确保多样性和自然属性 (i.e 删除极暗或曝光过度的图像),我们还删除模糊的图像(i.e.我们只考虑全对焦图像)
预处理如下：(i)我们根据所有 RAW 图像的黑色水平和位深度对其进行归一化。(ii)我们将图像转换(“打包”)为众所周知的 RGGBBayer 模式(4通道), 该模式允许在不损坏原始颜色模式信息的情况下应用转换和降级$\mathbb{E}[?]$
培训：我们为参与者提供 1064 $1024\times1024\times4$干净的高分辨率(HR)RAW图像。使用 BSRAW[?]中提出的退化管道，可以在训练期间在线生成 LR 退化图像
这样的退化管道考虑了不同的噪声分布，多个模糊核(PSF) 和简单的下采样策略来合成低分辨率(LR)
Related Computer Vision Challenges 我们的挑战是NTIRE2024Workshop 中的任何 XMATHX 图像质量评估EMATHX\_
$\overline{\mathrm{https://cvlai.net/ntire}}/2024/$ | [
[
"0.11926",
"0.404",
"0.2484",
"0.4048",
"0.2484",
"0.4192",
"0.11926",
"0.4185"
],
[
"0.3506",
"0.4048",
"0.4192",
"0.4048",
"0.4192",
"0.4192",
"0.3506",
"0.4192"
],
[
"0.523",
"0.4048",
"0.652",
"0.4048",
"0.652",
"0.4192",
"0.523",
"0.4192"
],
[
"0.7534",
"0.4048",
"0.822",
"0.4048",
"0.822",
"0.4192",
"0.7534",
"0.4192"
],
[
"0.1847",
"0.433",
"0.785",
"0.433",
"0.785",
"0.4463",
"0.1847",
"0.4463"
],
[
"0.08417",
"0.4749",
"0.4673",
"0.4749",
"0.4673",
"0.488",
"0.08417",
"0.488"
],
[
"0.509",
"0.4749",
"0.88",
"0.4749",
"0.88",
"0.488",
"0.509",
"0.488"
],
[
"0.0825",
"0.4893",
"0.4673",
"0.4893",
"0.4673",
"0.5024",
"0.0825",
"0.5024"
],
[
"0.0825",
"0.505",
"0.4673",
"0.505",
"0.4673",
"0.518",
"0.0825",
"0.518"
],
[
"0.5073",
"0.511",
"0.888",
"0.512",
"0.888",
"0.5264",
"0.5073",
"0.526"
],
[
"0.0825",
"0.5195",
"0.4673",
"0.5195",
"0.4673",
"0.5327",
"0.0825",
"0.5327"
],
[
"0.505",
"0.528",
"0.8906",
"0.528",
"0.8906",
"0.5425",
"0.505",
"0.5425"
],
[
"0.0825",
"0.535",
"0.4673",
"0.535",
"0.4673",
"0.549",
"0.0825",
"0.549"
],
[
"0.5034",
"0.5435",
"0.5825",
"0.5435",
"0.5825",
"0.5547",
"0.5034",
"0.5547"
],
[
"0.08417",
"0.55",
"0.4673",
"0.55",
"0.4673",
"0.563",
"0.08417",
"0.563"
],
[
"0.0825",
"0.5664",
"0.4387",
"0.5664",
"0.4387",
"0.577",
"0.0825",
"0.577"
],
[
"0.5",
"0.5737",
"0.806",
"0.5747",
"0.806",
"0.5923",
"0.5",
"0.5913"
],
[
"0.10297",
"0.58",
"0.4673",
"0.58",
"0.4673",
"0.5933",
"0.10297",
"0.5933"
],
[
"0.0825",
"0.595",
"0.468",
"0.595",
"0.468",
"0.6094",
"0.0825",
"0.6094"
],
[
"0.5034",
"0.601",
"0.6006",
"0.601",
"0.6006",
"0.6157",
"0.5034",
"0.6157"
],
[
"0.0825",
"0.6113",
"0.4673",
"0.6113",
"0.4673",
"0.625",
"0.0825",
"0.625"
],
[
"0.08417",
"0.6274",
"0.4656",
"0.6274",
"0.4656",
"0.638",
"0.08417",
"0.638"
],
[
"0.5024",
"0.6235",
"0.8896",
"0.6235",
"0.8896",
"0.638",
"0.5024",
"0.638"
],
[
"0.0809",
"0.6416",
"0.4673",
"0.6416",
"0.4673",
"0.655",
"0.0809",
"0.655"
],
[
"0.5034",
"0.64",
"0.8896",
"0.64",
"0.8896",
"0.654",
"0.5034",
"0.654"
],
[
"0.0809",
"0.656",
"0.468",
"0.656",
"0.468",
"0.6694",
"0.0809",
"0.6694"
],
[
"0.505",
"0.656",
"0.8906",
"0.656",
"0.8906",
"0.6694",
"0.505",
"0.6694"
],
[
"0.0825",
"0.6724",
"0.1642",
"0.6724",
"0.1642",
"0.6846",
"0.0825",
"0.6846"
],
[
"0.5063",
"0.672",
"0.5376",
"0.672",
"0.5376",
"0.6846",
"0.5063",
"0.6846"
],
[
"0.10046",
"0.686",
"0.468",
"0.686",
"0.468",
"0.7007",
"0.10046",
"0.7007"
],
[
"0.524",
"0.6875",
"0.888",
"0.6875",
"0.888",
"0.7007",
"0.524",
"0.7007"
],
[
"0.0801",
"0.7007",
"0.4656",
"0.702",
"0.4656",
"0.7153",
"0.0801",
"0.714"
],
[
"0.5063",
"0.704",
"0.8896",
"0.704",
"0.8896",
"0.714",
"0.5063",
"0.714"
],
[
"0.0809",
"0.7173",
"0.4673",
"0.7173",
"0.4673",
"0.732",
"0.0809",
"0.732"
],
[
"0.5024",
"0.7173",
"0.706",
"0.7173",
"0.706",
"0.731",
"0.5024",
"0.731"
],
[
"0.0825",
"0.733",
"0.4656",
"0.733",
"0.4656",
"0.747",
"0.0825",
"0.747"
],
[
"0.523",
"0.733",
"0.8906",
"0.733",
"0.8906",
"0.747",
"0.523",
"0.747"
],
[
"0.0825",
"0.7476",
"0.3743",
"0.7476",
"0.3743",
"0.761",
"0.0825",
"0.761"
],
[
"0.5034",
"0.7476",
"0.8896",
"0.7476",
"0.8896",
"0.761",
"0.5034",
"0.761"
],
[
"0.0989",
"0.762",
"0.4673",
"0.763",
"0.4673",
"0.778",
"0.0989",
"0.7764"
],
[
"0.5024",
"0.762",
"0.888",
"0.763",
"0.888",
"0.778",
"0.5024",
"0.7764"
],
[
"0.0825",
"0.778",
"0.4673",
"0.778",
"0.4673",
"0.7925",
"0.0825",
"0.7925"
],
[
"0.505",
"0.779",
"0.8896",
"0.779",
"0.8896",
"0.7925",
"0.505",
"0.7925"
],
[
"0.0801",
"0.7915",
"0.1814",
"0.7935",
"0.1814",
"0.808",
"0.0801",
"0.806"
],
[
"0.505",
"0.7935",
"0.6284",
"0.7935",
"0.6284",
"0.8066",
"0.505",
"0.8066"
],
[
"0.505",
"0.81",
"0.8896",
"0.81",
"0.8896",
"0.8237",
"0.505",
"0.8237"
],
[
"0.5034",
"0.8247",
"0.8906",
"0.8237",
"0.8906",
"0.8384",
"0.5034",
"0.839"
],
[
"0.08496",
"0.8306",
"0.4673",
"0.8306",
"0.4673",
"0.8447",
"0.08496",
"0.8447"
],
[
"0.0801",
"0.8447",
"0.468",
"0.845",
"0.468",
"0.86",
"0.0801",
"0.8594"
],
[
"0.5034",
"0.839",
"0.8896",
"0.8384",
"0.8896",
"0.853",
"0.5034",
"0.854"
],
[
"0.0809",
"0.8613",
"0.1953",
"0.8613",
"0.1953",
"0.875",
"0.0809",
"0.875"
],
[
"0.5063",
"0.857",
"0.5215",
"0.857",
"0.5215",
"0.867",
"0.5063",
"0.867"
],
[
"0.524",
"0.872",
"0.8896",
"0.872",
"0.8896",
"0.8853",
"0.524",
"0.8853"
],
[
"0.1013",
"0.887",
"0.281",
"0.886",
"0.281",
"0.901",
"0.1013",
"0.9014"
],
[
"0.5034",
"0.887",
"0.8906",
"0.887",
"0.8906",
"0.901",
"0.5034",
"0.901"
]
] | [
"HR Ground-truth",
"LR Input",
"HR Ground-truth",
"LR Input",
"Figure 1. Samples of the NTIRE 2024 RAW Image Super-Resolution Challenge testing set.",
"质量的限制。各种图像复原应用的特点是获取实域配",
"Related Computer Vision Challenges，MATHX",
"对数据[？？1极其困难，推动了对现实和相关数据合",
"成[？？1的需求。鉴于高度非线性ISP特性也会改变",
"RelatedComputerVisionChallenges的任何深度图",
"退化外观的物理特性，因此在sRGB表示中准确建模",
"像质量评估E？1，夜间摄影染E[？]，阴影去除E",
"特定于应用程序的退化非常困难。这代表了限制这些",
"[？1，高效",
"算法在实际应用程序部署中的性能的主要因素，在数",
"据级别可观察到的差距很难通过无模型算法来克服",
"2.NTIRE 2024RAWSR Challenge",
"因此，研究RAW数据成为推动总体图像恢复性能改",
"进趋势的关键步骤。因此，RAWImageSuper-Resolution",
"2.1. Dataset",
"应用程序可以从增加的方差信号中受益，其算法对所",
"涉及设备的精细架构属性具有鲁棒性。开发高效的算",
"挑战数据集基于BSRAW[？1。继以前的工作E[？？",
"法，如在sRGB域[？？？1中完善的算法，可以证明",
"？1之后，我们使用AdobeMIT5K数据集E[？1中的",
"在高性能图像应用程序的开发中领先一步，具有成本",
"图像，其中包括来自多个佳能和尼康数码单反相机的",
"效益的设备",
"图像",
"因此，在这项工作中，我们将介绍为",
"手动过滤DSLR图像以确保多样性和自然属性（i.e",
"NTIRE2024RAW图像超分辨率挑战赛提交的解",
"删除极暗或曝光过度的图像），我们还册删除模糊的图像",
"决方案。我们提供有关挑战设置的信息，任务描述和",
"(i.e.我们只考虑全对焦图像）",
"挑战数据属性表征挑战数据集拆分。我们还列出了有",
"预处理如下：（i）我们根据所有RAW图像的黑色",
"关挑战参与者及其团队和隶属关系的信息",
"水平和位深度对其进行归一化。（ii）我们将图像转换",
"在Sec.2中，我们描述了挑战数据集和评估，并讨",
"（“打包”）为众所周知的RGGBBayer模式（4通道）",
"论了总体结果。在Sec.3中，我们提供了最佳解决方",
"该模式允许在不损坏原始颜色模式信息的情况下应用",
"案的详细描述",
"转换和降级E[？1",
"培训：我们为参与者提供10641024×1024×4千净的",
"高分辨率（HR）RAW图像。使用BSRAW[？1中提",
"RelatedComputerVisionChallenges我们的挑战是",
"NTIRE2024Workshop中的任何XMATHX图像质量",
"出的退化管道，可以在训练期间在线生成LR退化图",
"评估EMATHX",
"像",
"这样的退化管道考虑了不同的噪声分布，多个模糊",
"https://cvlai.net/ntire/2024/",
"核（PSF）和简单的下采样策略来合成低分辨率（LR）"
] | {
"arxiv_id": "2404.16223",
"language": "zh",
"page_idx": 1
} |
||
2404.16223-zh-2 | 2404.16223-zh-2.jpg | <table>
<tbody>
<tr>
<th>Team</th>
<th>Method</th>
<th>Validation 1MP</th>
<th>Test 上1</th>
<th>$1MP$</th>
<th>Test 12MP</th>
<th>(M) Params.</th>
<th>Train $E2E$</th>
<th>Res Train</th>
</tr>
<tr>
<td>Samsung</td>
<td>$FPL$ 2- Stage $w/$</td>
<td>/ 0.99 43.40</td>
<td>43.443</td>
<td>0.986</td>
<td>43.858 0.988</td>
<td>53.7</td>
<td>Yes</td>
<td>384</td>
</tr>
<tr>
<td>XiaomiMMAI</td>
<td>EffectiveSR</td>
<td>43.38 3 / 0.99</td>
<td>43.249</td>
<td>0.986</td>
<td>$NA$</td>
<td>20.9</td>
<td>$No$</td>
<td>64</td>
</tr>
<tr>
<td>USTCX</td>
<td>RBSFormer $\because$</td>
<td>43.21 0.99 </td>
<td>42.493</td>
<td>0.984</td>
<td>43.649 0.987</td>
<td>3.3</td>
<td>Yes</td>
<td>112</td>
</tr>
<tr>
<td rowspan="2">McMaster</td>
<td>SwinFSR $\therefore$Raw 1</td>
<td>42.48 0.98</td>
<td>42.366</td>
<td>0.984</td>
<td>$NA$</td>
<td>6.64</td>
<td>Yes</td>
<td>256</td>
</tr>
<tr>
<td>BSRAW 1</td>
<td>42.25 0.98</td>
<td>42.106</td>
<td>0.984</td>
<td>42.853 0.986</td>
<td>1.5</td>
<td>Yes</td>
<td>248</td>
</tr>
<tr>
<td rowspan="2">$RSR$ NUDT</td>
<td>SAFMN E $FFT$ 1 </td>
<td>41.81 0.98</td>
<td>41.621</td>
<td>0.982</td>
<td>$NA$</td>
<td>0.27</td>
<td>$No$</td>
<td>128-448</td>
</tr>
<tr>
<td>Interpolation</td>
<td>35.95 0.95</td>
<td>36.038</td>
<td>0.952</td>
<td>36.926 0.956</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
</tbody>
</table>
Table 1. We provide PSNR/SSIM results on the validation set (40 images), the complete testing set (200 images), and the testing set at full-resolution (12MP) RAW images [?]. All the fidelity metrics are calculated in the RAW domain.“NA” indicates the results are not available for the method. We highlight two baseline methods. We also report the number of parameters of each method, if the method was trained end-to-end (Yes/No), and the image resolution used for training the ${\mathrm{models.}}$
RAW 图像。参与者可以应用其他增强技术或扩展降级
管道以生成更逼真的训练数据
## $2. 2. \textbf{ Baselines}$
我们使用 BSRAW [?] 作为主要基线。性能最好的挑战解决方案提高了基线性能，然而，神经网络在设计和计算方面明显更加复杂
# $2. 3. \textbf{ Results}$
我们使用三个测试拆分：(i)验证，在模型开发阶段使用 40 个 1024px 图像。(ii)测试 1MP, 200 张 1024px 分辨率的图像。(iii)全分辨率$\approx12$ MP 的相同 200张测试图像。参与者处理相应的 LRRAW 图像( e.g. $512\times512\times4$ ),并提交结果。因此，参与者无法访问地面实况图像
我们在 Fig. 1 中提供了测试集的示例
在 Tab. 1 中，我们提供了挑战基准。除了 PSNR 和SSIM 等保真度指标外，我们还提供了每种方法的相关实现细节。该方法可以大大提高 RAW 图像的质量和分辨率，即使在全分辨率 12MP 图像作为输出的情况下。我们在 Fig.8,?? 和 ?? 中提供详细的视觉比较。所有提出的方法都能够提高 RAW 图像的分辨率和细节，同时减少模糊和噪声。此外，没有可检测的颜色伪影
我们可以得出结论，(合成)RAW 图像超分辨率可以解决 RAW 去噪的相似性。然而，更现实的下采样仍然是一个开放的挑战
Acknowledgements 这项工作得到了洪堡基金会的部分支持。我们感谢 NTIRE2024 赞助商：元现实实验室，OPPO, 快手，华为和 Würzburg 大学(计算机视觉实验室)
## $3.{\mathrm{~Challenge~Methods~and~Teams}}$ $3. 1. \textbf{ Dual Stage RAW SR with Focal Pixel Loss}$
Team Samsung MX,SRC-B
Jianxing Zhang $^1$, Jia Li$^1$,Fan Wang$^1$ Xiaopeng Li$^1$,Zikun Liu$^1$,Hyunhee Park$^2$
${\overset{\circ}{\operatorname*{\operatorname*{Sejun}}}}{\overset{\circ}{\operatorname*{Song}}}{\overset{2}{\operatorname*{\operatorname*{Song}}}}{\overset{2}{\operatorname*{\operatorname*{Songho}}}}{\overset{\circ}{\operatorname*{Kim}}}{\overset{2}{\operatorname*{\operatorname*{Sejun}}}}$
$^{1}$Samsung Research China- Beijing (SRC-B) $^{2}$Samsung MX(Mobile eXperience) Business
三星 MX、SRC-B 团队正在为 RAW 图像超分辨率引入两级网络。该解决方案采用分而治之的策略，第一阶段的任务是从低分辨率退化的 RAW 图像中恢复图像结构，第二阶段的目标是恢复最大量的细节，提供精细的重建。此外，该团队正在进一步扩展现有的合成数据生成方法，进一步研究硬件特定的 RAW 图像退化， 为相关设备特定的噪声配置文件提出新的定义，并提出与典型现实场景一致的新模糊内核。他们提出了一个随机退化模型，模拟观察到的模拟缺陷之间的不同相互作用
最后，三星 MX, SRC-B 团队提出了一种新颖的FocalPixelLoss, 通过模型微调阶段的性能改进证明了这一点
如 Fig. $\color{red}{2\text{ 所示，网络结构主要包括两个阶段。第}}$ 一阶段主要利用 Restormer$\underline {\mathrm{E} }[ ?$ ] ,其主要作用是恢复原始图像的主要内容。第二阶段主要采用了基于NAFNet[?]的设计，其主要作用是在第一阶段恢复的基础上恢复更多细节
训练过程说明了所提出方法的双阶段设计。训练过程的第一阶段遵循与模型第一阶段相对应的参数的优化。在训练过程的第二阶段，优化的参数被冻结，第二阶段的参数开始被细化，以实现最佳专业化。然后使用两组优化参数(Restormer 和 NAFNet 模型)进行最终估计。
解决方案的主要细节之一是根据输入数据的特征提出新颖的训练目标。考虑到信号退化函数的非均匀效应，焦距损失的命题(请参见 Eq.(1))是针对在高度受影响的像素比率方面观察到的不平衡的解决方案。 | [
[
"0.08826",
"0.09155",
"0.1283",
"0.09344",
"0.1274",
"0.108",
"0.0866",
"0.1061"
],
[
"0.1929",
"0.09216",
"0.2427",
"0.09216",
"0.2427",
"0.1067",
"0.1929",
"0.1067"
],
[
"0.433",
"0.09216",
"0.4968",
"0.09216",
"0.4968",
"0.1067",
"0.433",
"0.1067"
],
[
"0.5317",
"0.09216",
"0.601",
"0.09216",
"0.601",
"0.1067",
"0.5317",
"0.1067"
],
[
"0.6274",
"0.0947",
"0.721",
"0.0947",
"0.721",
"0.1061",
"0.6274",
"0.1061"
],
[
"0.7363",
"0.0947",
"0.8",
"0.0947",
"0.8",
"0.1061",
"0.7363",
"0.1061"
],
[
"0.814",
"0.0947",
"0.8784",
"0.0947",
"0.8784",
"0.1061",
"0.814",
"0.1061"
],
[
"0.5195",
"0.1092",
"0.613",
"0.1092",
"0.613",
"0.1237",
"0.5195",
"0.1237"
],
[
"0.3137",
"0.0947",
"0.4084",
"0.0947",
"0.4084",
"0.1061",
"0.3137",
"0.1061"
],
[
"0.66",
"0.1092",
"0.689",
"0.1092",
"0.689",
"0.1237",
"0.66",
"0.1237"
],
[
"0.755",
"0.1092",
"0.7793",
"0.1092",
"0.7793",
"0.1225",
"0.755",
"0.1225"
],
[
"0.835",
"0.1092",
"0.859",
"0.1092",
"0.859",
"0.1225",
"0.835",
"0.1225"
],
[
"0.1936",
"0.10986",
"0.29",
"0.10986",
"0.29",
"0.1218",
"0.1936",
"0.1218"
],
[
"0.09235",
"0.10986",
"0.1478",
"0.10986",
"0.1478",
"0.1218",
"0.09235",
"0.1218"
],
[
"0.3245",
"0.10986",
"0.3962",
"0.10986",
"0.3962",
"0.1218",
"0.3245",
"0.1218"
],
[
"0.4248",
"0.10986",
"0.5107",
"0.10986",
"0.5107",
"0.1218",
"0.4248",
"0.1218"
],
[
"0.5557",
"0.1218",
"0.5786",
"0.1218",
"0.5786",
"0.1351",
"0.5557",
"0.1351"
],
[
"0.66",
"0.1218",
"0.689",
"0.1218",
"0.689",
"0.1364",
"0.66",
"0.1364"
],
[
"0.7573",
"0.1218",
"0.778",
"0.1218",
"0.778",
"0.1351",
"0.7573",
"0.1351"
],
[
"0.8364",
"0.1218",
"0.8555",
"0.1218",
"0.8555",
"0.1351",
"0.8364",
"0.1351"
],
[
"0.1929",
"0.1225",
"0.2632",
"0.1225",
"0.2632",
"0.1338",
"0.1929",
"0.1338"
],
[
"0.09235",
"0.1225",
"0.179",
"0.1225",
"0.179",
"0.1338",
"0.09235",
"0.1338"
],
[
"0.3245",
"0.1237",
"0.3962",
"0.1237",
"0.3962",
"0.1338",
"0.3245",
"0.1338"
],
[
"0.423",
"0.1225",
"0.509",
"0.1225",
"0.509",
"0.1338",
"0.423",
"0.1338"
],
[
"0.09235",
"0.1338",
"0.1438",
"0.1338",
"0.1438",
"0.1483",
"0.09235",
"0.1483"
],
[
"0.1912",
"0.1332",
"0.2893",
"0.1351",
"0.2893",
"0.1497",
"0.1912",
"0.1477"
],
[
"0.661",
"0.1351",
"0.6855",
"0.1351",
"0.6855",
"0.1477",
"0.661",
"0.1477"
],
[
"0.755",
"0.1351",
"0.7793",
"0.1351",
"0.7793",
"0.1483",
"0.755",
"0.1483"
],
[
"0.835",
"0.1351",
"0.8594",
"0.1351",
"0.8594",
"0.1483",
"0.835",
"0.1483"
],
[
"0.423",
"0.1364",
"0.5107",
"0.1364",
"0.5107",
"0.1477",
"0.423",
"0.1477"
],
[
"0.3245",
"0.1364",
"0.3962",
"0.1364",
"0.3962",
"0.1477",
"0.3245",
"0.1477"
],
[
"0.524",
"0.1364",
"0.6113",
"0.1364",
"0.6113",
"0.1477",
"0.524",
"0.1477"
],
[
"0.09235",
"0.1465",
"0.156",
"0.1465",
"0.156",
"0.161",
"0.09235",
"0.161"
],
[
"0.5557",
"0.1465",
"0.58",
"0.1465",
"0.58",
"0.1604",
"0.5557",
"0.1604"
],
[
"0.6567",
"0.1465",
"0.689",
"0.1465",
"0.689",
"0.1622",
"0.6567",
"0.1622"
],
[
"0.755",
"0.1477",
"0.7793",
"0.1477",
"0.7793",
"0.161",
"0.755",
"0.161"
],
[
"0.834",
"0.1477",
"0.859",
"0.1477",
"0.859",
"0.1604",
"0.834",
"0.1604"
],
[
"0.3245",
"0.1483",
"0.3962",
"0.1483",
"0.3962",
"0.1604",
"0.3245",
"0.1604"
],
[
"0.1929",
"0.1483",
"0.2793",
"0.1483",
"0.2793",
"0.1604",
"0.1929",
"0.1604"
],
[
"0.423",
"0.1483",
"0.509",
"0.1483",
"0.509",
"0.1604",
"0.423",
"0.1604"
],
[
"0.1912",
"0.1578",
"0.2695",
"0.1604",
"0.2695",
"0.1749",
"0.1912",
"0.1724"
],
[
"0.4224",
"0.1604",
"0.5107",
"0.1604",
"0.5107",
"0.1718",
"0.4224",
"0.1718"
],
[
"0.755",
"0.1604",
"0.7812",
"0.1604",
"0.7812",
"0.1736",
"0.755",
"0.1736"
],
[
"0.835",
"0.159",
"0.8613",
"0.1622",
"0.859",
"0.1749",
"0.8325",
"0.1718"
],
[
"0.5186",
"0.161",
"0.6113",
"0.161",
"0.6113",
"0.1724",
"0.5186",
"0.1724"
],
[
"0.3245",
"0.161",
"0.3945",
"0.161",
"0.3945",
"0.1724",
"0.3245",
"0.1724"
],
[
"0.661",
"0.161",
"0.6855",
"0.161",
"0.6855",
"0.1736",
"0.661",
"0.1736"
],
[
"0.6567",
"0.1724",
"0.689",
"0.1724",
"0.689",
"0.1881",
"0.6567",
"0.1881"
],
[
"0.7563",
"0.1724",
"0.7793",
"0.1724",
"0.7793",
"0.1863",
"0.7563",
"0.1863"
],
[
"0.1929",
"0.1736",
"0.2737",
"0.1736",
"0.2737",
"0.1849",
"0.1929",
"0.1849"
],
[
"0.09314",
"0.1736",
"0.1683",
"0.1736",
"0.1683",
"0.1849",
"0.09314",
"0.1849"
],
[
"0.3228",
"0.1749",
"0.3945",
"0.1749",
"0.3945",
"0.1863",
"0.3228",
"0.1863"
],
[
"0.423",
"0.1749",
"0.5107",
"0.1749",
"0.5107",
"0.1863",
"0.423",
"0.1863"
],
[
"0.556",
"0.1736",
"0.5786",
"0.1736",
"0.5786",
"0.1863",
"0.556",
"0.1863"
],
[
"0.819",
"0.1749",
"0.876",
"0.1749",
"0.876",
"0.1863",
"0.819",
"0.1863"
],
[
"0.3218",
"0.1875",
"0.3962",
"0.1875",
"0.3962",
"0.1976",
"0.3218",
"0.1976"
],
[
"0.4224",
"0.1875",
"0.511",
"0.1875",
"0.511",
"0.1989",
"0.4224",
"0.1989"
],
[
"0.523",
"0.1875",
"0.6113",
"0.1875",
"0.6113",
"0.1989",
"0.523",
"0.1989"
],
[
"0.513",
"0.1913",
"0.5215",
"0.1913",
"0.5215",
"0.1957",
"0.513",
"0.1957"
],
[
"0.1936",
"0.1881",
"0.2998",
"0.1881",
"0.2998",
"0.1989",
"0.1936",
"0.1989"
],
[
"0.0809",
"0.214",
"0.8906",
"0.214",
"0.8906",
"0.2279",
"0.0809",
"0.2279"
],
[
"0.0801",
"0.2291",
"0.8896",
"0.2291",
"0.8896",
"0.2424",
"0.0801",
"0.2424"
],
[
"0.0801",
"0.2424",
"0.892",
"0.2424",
"0.892",
"0.2556",
"0.0801",
"0.2556"
],
[
"0.0809",
"0.2583",
"0.8896",
"0.2583",
"0.8896",
"0.2683",
"0.0809",
"0.2683"
],
[
"0.0801",
"0.2676",
"0.1323",
"0.2695",
"0.1316",
"0.2842",
"0.0784",
"0.2822"
],
[
"0.0809",
"0.3113",
"0.468",
"0.3113",
"0.468",
"0.3245",
"0.0809",
"0.3245"
],
[
"0.501",
"0.3093",
"0.7935",
"0.31",
"0.7935",
"0.3245",
"0.501",
"0.324"
],
[
"0.0809",
"0.327",
"0.2942",
"0.327",
"0.2942",
"0.3403",
"0.0809",
"0.3403"
],
[
"0.5034",
"0.3352",
"0.865",
"0.3352",
"0.865",
"0.3484",
"0.5034",
"0.3484"
],
[
"0.0784",
"0.3499",
"0.1912",
"0.351",
"0.1912",
"0.3655",
"0.0784",
"0.3643"
],
[
"0.5947",
"0.363",
"0.7935",
"0.3643",
"0.7935",
"0.379",
"0.5947",
"0.3774"
],
[
"0.0809",
"0.375",
"0.468",
"0.375",
"0.468",
"0.3882",
"0.0809",
"0.3882"
],
[
"0.0809",
"0.3894",
"0.468",
"0.3894",
"0.468",
"0.4028",
"0.0809",
"0.4028"
],
[
"0.536",
"0.3865",
"0.851",
"0.3865",
"0.851",
"0.401",
"0.536",
"0.401"
],
[
"0.0801",
"0.404",
"0.246",
"0.4048",
"0.246",
"0.4192",
"0.0801",
"0.4185"
],
[
"0.5186",
"0.4016",
"0.869",
"0.3997",
"0.869",
"0.414",
"0.5186",
"0.4167"
],
[
"0.5786",
"0.4167",
"0.812",
"0.4153",
"0.812",
"0.43",
"0.5786",
"0.4312"
],
[
"0.0809",
"0.43",
"0.1765",
"0.43",
"0.1765",
"0.4414",
"0.0809",
"0.4414"
],
[
"0.5317",
"0.4375",
"0.8594",
"0.4395",
"0.8594",
"0.4539",
"0.5317",
"0.452"
],
[
"0.0809",
"0.452",
"0.468",
"0.452",
"0.468",
"0.4653",
"0.0809",
"0.4653"
],
[
"0.5317",
"0.4507",
"0.8594",
"0.4526",
"0.8594",
"0.4702",
"0.5317",
"0.4685"
],
[
"0.0825",
"0.4673",
"0.468",
"0.4673",
"0.468",
"0.481",
"0.0825",
"0.481"
],
[
"0.0809",
"0.4817",
"0.468",
"0.4817",
"0.468",
"0.4956",
"0.0809",
"0.4956"
],
[
"0.524",
"0.4792",
"0.892",
"0.4792",
"0.892",
"0.4924",
"0.524",
"0.4924"
],
[
"0.0784",
"0.4968",
"0.4673",
"0.4976",
"0.4673",
"0.512",
"0.0784",
"0.511"
],
[
"0.5034",
"0.4944",
"0.875",
"0.4944",
"0.875",
"0.5083",
"0.5034",
"0.5083"
],
[
"0.0784",
"0.511",
"0.468",
"0.512",
"0.468",
"0.5264",
"0.0784",
"0.526"
],
[
"0.505",
"0.511",
"0.888",
"0.511",
"0.888",
"0.5215",
"0.505",
"0.5215"
],
[
"0.0825",
"0.529",
"0.179",
"0.529",
"0.179",
"0.5405",
"0.0825",
"0.5405"
],
[
"0.505",
"0.5244",
"0.8896",
"0.5244",
"0.8896",
"0.5386",
"0.505",
"0.5386"
],
[
"0.10046",
"0.5435",
"0.3652",
"0.5435",
"0.3652",
"0.5566",
"0.10046",
"0.5566"
],
[
"0.505",
"0.5415",
"0.8896",
"0.5415",
"0.8896",
"0.552",
"0.505",
"0.552"
],
[
"0.10046",
"0.5566",
"0.4673",
"0.5566",
"0.4673",
"0.571",
"0.10046",
"0.571"
],
[
"0.5034",
"0.5547",
"0.888",
"0.5547",
"0.888",
"0.5684",
"0.5034",
"0.5684"
],
[
"0.0801",
"0.5713",
"0.468",
"0.5728",
"0.468",
"0.587",
"0.0801",
"0.586"
],
[
"0.5034",
"0.571",
"0.8906",
"0.571",
"0.8906",
"0.584",
"0.5034",
"0.584"
],
[
"0.0809",
"0.5884",
"0.468",
"0.5884",
"0.468",
"0.6016",
"0.0809",
"0.6016"
],
[
"0.5034",
"0.5854",
"0.876",
"0.5854",
"0.876",
"0.5986",
"0.5034",
"0.5986"
],
[
"0.0825",
"0.603",
"0.468",
"0.603",
"0.468",
"0.616",
"0.0825",
"0.616"
],
[
"0.5034",
"0.601",
"0.8906",
"0.601",
"0.8906",
"0.6143",
"0.5034",
"0.6143"
],
[
"0.08417",
"0.619",
"0.4673",
"0.619",
"0.4673",
"0.63",
"0.08417",
"0.63"
],
[
"0.505",
"0.617",
"0.568",
"0.617",
"0.568",
"0.629",
"0.505",
"0.629"
],
[
"0.0809",
"0.6333",
"0.468",
"0.6333",
"0.468",
"0.6465",
"0.0809",
"0.6465"
],
[
"0.523",
"0.63",
"0.8906",
"0.63",
"0.8906",
"0.6445",
"0.523",
"0.6445"
],
[
"0.0809",
"0.6484",
"0.468",
"0.6484",
"0.468",
"0.662",
"0.0809",
"0.662"
],
[
"0.5024",
"0.646",
"0.888",
"0.646",
"0.888",
"0.659",
"0.5024",
"0.659"
],
[
"0.08496",
"0.661",
"0.10046",
"0.6675",
"0.09314",
"0.6787",
"0.0768",
"0.6724"
],
[
"0.5024",
"0.66",
"0.5557",
"0.66",
"0.5557",
"0.675",
"0.5024",
"0.675"
],
[
"0.1013",
"0.6787",
"0.4673",
"0.6787",
"0.4673",
"0.6924",
"0.1013",
"0.6924"
],
[
"0.523",
"0.6753",
"0.8906",
"0.6753",
"0.8906",
"0.6895",
"0.523",
"0.6895"
],
[
"0.0809",
"0.6934",
"0.468",
"0.6934",
"0.468",
"0.707",
"0.0809",
"0.707"
],
[
"0.509",
"0.6943",
"0.5156",
"0.6943",
"0.5156",
"0.6987",
"0.509",
"0.6987"
],
[
"0.5156",
"0.6943",
"0.5215",
"0.6943",
"0.5215",
"0.6997",
"0.5156",
"0.6997"
],
[
"0.517",
"0.69",
"0.8906",
"0.69",
"0.8906",
"0.704",
"0.517",
"0.704"
],
[
"0.0801",
"0.709",
"0.2296",
"0.709",
"0.2296",
"0.7227",
"0.0801",
"0.7227"
],
[
"0.5034",
"0.7056",
"0.8906",
"0.7056",
"0.8906",
"0.7197",
"0.5034",
"0.7197"
],
[
"0.5034",
"0.72",
"0.8906",
"0.72",
"0.8906",
"0.7344",
"0.5034",
"0.7344"
],
[
"0.08826",
"0.731",
"0.4673",
"0.731",
"0.4673",
"0.744",
"0.08826",
"0.744"
],
[
"0.5034",
"0.7363",
"0.6685",
"0.7363",
"0.6685",
"0.75",
"0.5034",
"0.75"
],
[
"0.0809",
"0.744",
"0.468",
"0.7456",
"0.468",
"0.7603",
"0.0809",
"0.759"
],
[
"0.523",
"0.75",
"0.8906",
"0.7505",
"0.8906",
"0.765",
"0.523",
"0.7646"
],
[
"0.0809",
"0.759",
"0.4673",
"0.759",
"0.4673",
"0.7734",
"0.0809",
"0.7734"
],
[
"0.5034",
"0.7666",
"0.8906",
"0.7666",
"0.8906",
"0.78",
"0.5034",
"0.78"
],
[
"0.0809",
"0.7764",
"0.1544",
"0.7764",
"0.1544",
"0.7886",
"0.0809",
"0.7886"
],
[
"0.505",
"0.782",
"0.8774",
"0.782",
"0.8774",
"0.7954",
"0.505",
"0.7954"
],
[
"0.8735",
"0.786",
"0.8857",
"0.786",
"0.8857",
"0.7925",
"0.8735",
"0.7925"
],
[
"0.5034",
"0.797",
"0.8896",
"0.797",
"0.8896",
"0.81",
"0.5034",
"0.81"
],
[
"0.505",
"0.811",
"0.892",
"0.811",
"0.892",
"0.8247",
"0.505",
"0.8247"
],
[
"0.5024",
"0.8257",
"0.5605",
"0.8276",
"0.5605",
"0.8403",
"0.5024",
"0.8384"
],
[
"0.523",
"0.8413",
"0.8896",
"0.8413",
"0.8896",
"0.855",
"0.523",
"0.855"
],
[
"0.5034",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.5034",
"0.8706"
],
[
"0.5024",
"0.8726",
"0.8896",
"0.8726",
"0.8896",
"0.886",
"0.5024",
"0.886"
],
[
"0.5024",
"0.886",
"0.8906",
"0.888",
"0.8906",
"0.9014",
"0.5024",
"0.8994"
]
] | [
"Team",
"Method",
"Test IMP",
"Test 12MP",
"#Params.(M)",
"Train E2E",
"Train Res.",
"43.858 / 0.988",
"Validation1MP",
"53.7",
"Yes",
"384",
"2-Stage w/FPL",
"Samsung",
"43.40/ 0.99",
"43.443/0.986",
"NA",
"20.9",
"No",
"64",
"EffectiveSR",
"XiaomiMMAI",
"43.38/0.99",
"43.249/0.986",
"USTCX",
"RBSFormer [?]",
"3.3",
"Yes",
"112",
"42.493/ 0.984",
"43.21 / 0.99",
"43.649/0.987",
"MeMaster",
"NA",
"6.64",
"Yes",
"256",
"42.48/ 0.98",
"SwinFSRRaw",
"42.366/0.984",
"BSRAW[?]",
"42.106 / 0.984",
"Yes",
"248",
"42.853/0.986",
"42.25/0.98",
"1.5",
"0.27",
"No",
"SAFMN FFT",
"NUDTRSR",
"41.81 /0.98",
"41.621 / 0.982",
"NA",
"128-448",
"35.95/0.95",
"36.038/0.952",
"36.926/0.956",
"",
"Interpolation[2]",
"Table 1. We provide PSNR/SSIM results on the validation set (40 images), the complete testing set (200 images), and the",
"testing set at full-resolution (12MP) RAW images [? J :All the fidelity metrics are calculated in the RAW domain. “NA\"",
"indicates the results are not available for the method. We highlight two baseline methods. We also report the number of",
"parameters of each method, if the method was trained end-to-end (Yes/No),and the imageresolution used for training the",
"models.",
"RAW图像。参与者可以应用其他增强技术或扩展降级",
"3. Challenge Methods and Teams",
"管道以生成更逼真的训练数据",
"3.1.Dual Stage RAW SR with Focal Pixel Loss",
"2.2. Baselines",
"Team Samsung MX,SRC-B",
"我们使用BSRAW[？1作为主要基线。性能最好的挑",
"战解决方案提高了基线性能，然而，神经网络在设计和",
"Jianxing Zhang 1 , Jia Li 1 , Fan Wang 1",
"计算方面明显更加复杂",
"Xiaopeng Li 1, Zikun Liu 1, Hyunhee Park 2",
"Sejun Song 2, Changho Kim 2",
"2.3. Results",
"1 Samsung Research China - Beijing (SRC-B)",
"我们使用三个测试拆分：（i）验证，在模型开发阶段使",
"2 Samsung MX(Mobile eXperience) Business",
"用40个1024px图像。（ii）测试1MP，200张1024px",
"分辨率的图像。（ii）全分辨率～12MP的相同200",
"三星MX、SRC-B团队正在为RAW图像超分辨率",
"张测试图像。参与者处理相应的LRRAW图像（e.g.",
"引人两级网络。该解决方案采用分而治之的策略，第",
"512×512×4），并提交结果。因此，参与者无法访问",
"阶段的任务是从低分辨率退化的RAW图像中恢复图",
"地面实况图像",
"像结构，第二阶段的目标是恢复最大量的细节，提供精",
"我们在Fig.1中提供了测试集的示例",
"细的重建。此外，该团队正在进一步扩展现有的合成数",
"在Tab.1中，我们提供了挑战基准。除了PSNR和",
"据生成方法，进一步研究硬件特定的RAW图像退化",
"SSIM等保真度指标外，我们还提供了每种方法的相关",
"为相关设备特定的噪声配置文件提出新的定义，并提",
"实现细节。该方法可以大大提高RAW图像的质量和",
"出与典型现实场景一致的新模糊内核。他们提出了",
"分辨率，即使在全分辨率12MP图像作为输出的情况",
"个随机退化模型，模拟观察到的模拟缺陷之间的不同",
"下。我们在Fig.8，？？和？？中提供详细的视觉比较。",
"相互作用",
"所有提出的方法都能够提高RAW图像的分辨率和细",
"最后，三星MX，SRC-B团队提出了一种新颖的",
"节，同时减少模糊和噪声。此外，没有可检测的颜色伪",
"FocalPixelLoss，通过模型微调阶段的性能改进证明了",
"激",
"这一点",
"我们可以得出结论，（合成）RAW图像超分辨率可",
"如Fig.2所示，网络结构主要包括两个阶段。第",
"以解决RAW去噪的相似性。然而，更现实的下采样仍",
"",
"-",
"阶段主要利用RestormerE？1，其主要作用是",
"然是一个开放的挑战",
"恢复原始图像的主要内容。第二阶段主要采用了基于",
"NAFNet[？1的设计，其主要作用是在第一阶段恢复",
"Acknowledgements这项工作得到了洪堡基金会的",
"的基础上恢复更多细节",
"部分支持。我们感谢NTIRE2024赞助商：元现实实验",
"训练过程说明了所提出方法的双阶段设计。训练过",
"室，OPPO，快手，华为和Wirzburg大学（计算机视",
"程的第一阶段遵循与模型第一阶段相对应的参数的优",
"觉实验室）",
"化。在训练过程的第二阶段，优化的参数被冻结，第",
"",
"阶段的参数开始被细化，以实现最佳专业化。然后使用",
"两组优化参数（Restormer和NAFNet模型）进行最",
"终估计。",
"解决方案的主要细节之一是根据输入数据的特征提",
"出新颖的训练目标。考虑到信号退化函数的非均匀效",
"应，焦距损失的命题（请参见Eq.（1））是针对在高",
"度受影响的像素比率方面观察到的不平衡的解决方案。"
] | {
"arxiv_id": "2404.16223",
"language": "zh",
"page_idx": 2
} |
||
2404.16223-zh-3 | 2404.16223-zh-3.jpg | <FigureHere>
Figure 2. Dual-Stage RAWSR Framework proposed by Team Samsung MX,SRC. Best viewed in electronic version.
因此，焦点像素损失(FPL)正在对那些以大信号偏移
为特征的像素引人指数惩罚
$$FPL(\hat{I},I)=-D(\hat{I},I)^\gamma\log_{10}D(\hat{I},I)$$
(1)
在 Eq. (1) 中，$D(.,.)$是恢复图像$\hat{I}$和参考图像$I$ 之间的标准 L 范数距离，$\gamma$是一个可调因子，控制惩罚的强度
Implementation details 该模型仅根据挑战组织者提供的数据进行训练。它仅包含来自各种 DSLR 相机传感器的 1000 多个 RAW 图像。使用标准图像增强技术和图像退化管道的模拟来增强数据集。训练过程是一个双阶段操作，依次优化模型阶段。所采用的优化技术是采用余弦退火策略的 AdamW [?] 优化器 ( $\beta _{1}$ = 0.5 , $\beta _{2}$ = 0.999 , 权重衰减 0.0001 ),其中对于$5\times10^5$迭代，学习率从初始学习率$5\times10^-5$逐渐降低到$1\times10^{-7}$。该模型经历了预训练优化阶段，基于$L_{1}$ 损失。在 fune-tune 阶段，目标设置为 FocalPixelLoss, 初始学习速率设置为$5\times10^{-6}$ 。训练批大小设置为 4 , 补丁大小为 384 。水平/垂直翻转和旋转用于数据增强。所有实验均在 A100GPU 上进行
# 3.2. EffectiveSR for RAW Images
Team XiaomiMMAI
Zhijuan Huang $^1,\text{Hongyuan Yu}^{1}$,Cheng Wan$^2$,
Wending Xiang, Jiamin Lin $^1$,Hang Zhong$^1$, $\bar{\mathrm{Qiaosong~Zhang~}^{1}}$,Yue Sun$^1,\bar{\mathrm{Xuanwu~Yin~}^{1}}$,
Kunlong Zuo $^{^{\prime}}$
$^{1}\text{ Xiaomi Inc.}^{2}$Georgia Institute of Technology
XiaomiMMAI 团队提出的解决方案是基于 HAT$\boxed{\mathbb{E}}[?$ ] 的双分支网络，在训练过程中采用重新参数化$\mathbb{E}[?]$ ,使用附加参数充分挖掘方法的潜力。他们正在介绍一种针对 RAW 图像超分辨率的逐个任务和分步训练方法，以同时解决三个任务：降噪，去模糊和 $2\times$超分辨率
为了解决训练样本数量少的限制，XiaomiMMAI 团队将 RAW 图像转换为 RGB 图像，并使用随机旋转翻转，颜色变化，亮度变化，随机模糊等。然后，将增强的 RGB 图像集转换回 RGGBRAW 域，并将处理后的数据用于训练$\mathbb{E}[?]$ 所提出的解决方案
XiaomiMMAI 团队提出的模型灵感来自 HAT [?] ,其架构针对 RAWImageSuperResolution 任务进行了优化。优化的双分支网络结构(DB-HAT)如 Fig.3 所示。为 RAWISR 引人的逐步和逐个任务的训练方法进一步提高了他们的解决方案所达到的性能水平
一步一步：为了加速训练并获得良好的性能，XiaomiMMAI 团队采用了一种策略，其中每个子任务(包括最终的联合优化)都基于金字塔图像表示进行训练。最初，该模型是在小规模图像(64???)上训练的， 逐渐将图像补丁的分辨率提高到 128??? 和 256???
逐个任务：XiaomiMMAI 团队将 RAWISR 分为三个子任务：去噪，去模糊和 2xSR。最初，他们开始训练 RAW 图像去噪，然后是去模糊和 $2\times$超级分辨率的连接任务。最后，对整个网络应用联合优化程序，以产生最终的估计器。
在训练去噪和去模糊的过程中，XiaomiMMAI 团队在提出的 DB-HAT 的最后阶段使用了 RepConv 重新参数化技术，以提高任务的视觉图像质量。可重新参数化的卷积块 (RepConv)如图 4 所示
Implementation details 用于三个子任务训练的数据集由 1000+ 个 RAW 组成，数据增强方法可以参考上一节。在整个网络联合优化的最后阶段，XiaomiMMAI 团队使用提供的 1000+ 数据集，而不是增强的数据集。在三个子任务的训练过程中，学习率初始化为$4\times10^{-4}$ ,并根据余弦退火策略衰减。网络经过总共 2x 10$^5$迭代的训练，其中 L2 损失函数被最小化为 Adam 优化器的训练目标
随后，使用 L2 损失和 SSIM 损失函数对两次迭代执行微调，对于 2x 10$^5$迭代，初始学习率为$5\times10^{-5}$ 。所有实验均使用 PyTorch2。8个 A100GPU 上的0框架。
### 3.3. $\textbf{RBSFormer: Enhanced Transformer Net- }$ work for Raw Image Super-Resolution
$\operatorname{Team}$USTC604
Senyan Xu, Siyuan Jiang, Zhijing Sun, Jiaying Zhu
University of Science and Technology of China
Contact: a$804235820@$gmail.com | [
[
"0.523",
"0.09155",
"0.8906",
"0.09155",
"0.8906",
"0.1048",
"0.523",
"0.1048"
],
[
"0.2695",
"0.1092",
"0.2998",
"0.1092",
"0.2998",
"0.11615",
"0.2695",
"0.11615"
],
[
"0.5034",
"0.108",
"0.8896",
"0.108",
"0.8896",
"0.1218",
"0.5034",
"0.1218"
],
[
"0.517",
"0.1225",
"0.876",
"0.1225",
"0.876",
"0.1364",
"0.517",
"0.1364"
],
[
"0.5034",
"0.1364",
"0.8896",
"0.137",
"0.8896",
"0.1515",
"0.5034",
"0.1509"
],
[
"0.505",
"0.1528",
"0.8906",
"0.1528",
"0.8906",
"0.1666",
"0.505",
"0.1666"
],
[
"0.505",
"0.1686",
"0.5376",
"0.1686",
"0.5376",
"0.1812",
"0.505",
"0.1812"
],
[
"0.524",
"0.1843",
"0.8906",
"0.1843",
"0.8906",
"0.1976",
"0.524",
"0.1976"
],
[
"0.505",
"0.1995",
"0.888",
"0.1995",
"0.888",
"0.2134",
"0.505",
"0.2134"
],
[
"0.505",
"0.214",
"0.8906",
"0.214",
"0.8906",
"0.2279",
"0.505",
"0.2279"
],
[
"0.5034",
"0.2279",
"0.892",
"0.2291",
"0.892",
"0.2437",
"0.5034",
"0.2424"
],
[
"0.2615",
"0.2474",
"0.29",
"0.2474",
"0.29",
"0.2556",
"0.2615",
"0.2556"
],
[
"0.505",
"0.2443",
"0.796",
"0.2443",
"0.796",
"0.2583",
"0.505",
"0.2583"
],
[
"0.5215",
"0.2588",
"0.8823",
"0.26",
"0.8823",
"0.2747",
"0.5215",
"0.2734"
],
[
"0.4436",
"0.276",
"0.455",
"0.276",
"0.455",
"0.2798",
"0.4436",
"0.2798"
],
[
"0.5146",
"0.2747",
"0.886",
"0.2747",
"0.886",
"0.2886",
"0.5146",
"0.2886"
],
[
"0.5034",
"0.289",
"0.8906",
"0.2903",
"0.8906",
"0.305",
"0.5034",
"0.3037"
],
[
"0.5034",
"0.305",
"0.892",
"0.305",
"0.892",
"0.3193",
"0.5034",
"0.3193"
],
[
"0.513",
"0.3208",
"0.83",
"0.3193",
"0.83",
"0.334",
"0.513",
"0.3352"
],
[
"0.5376",
"0.3372",
"0.8906",
"0.3372",
"0.8906",
"0.351",
"0.5376",
"0.351"
],
[
"0.5034",
"0.351",
"0.8896",
"0.3516",
"0.8896",
"0.3662",
"0.5034",
"0.3655"
],
[
"0.0825",
"0.3726",
"0.4656",
"0.3726",
"0.4656",
"0.3833",
"0.0825",
"0.3833"
],
[
"0.5034",
"0.3687",
"0.8896",
"0.3687",
"0.8896",
"0.3818",
"0.5034",
"0.3818"
],
[
"0.0809",
"0.385",
"0.4656",
"0.385",
"0.4656",
"0.3984",
"0.0809",
"0.3984"
],
[
"0.5034",
"0.3833",
"0.8906",
"0.3833",
"0.8906",
"0.3965",
"0.5034",
"0.3965"
],
[
"0.505",
"0.3984",
"0.8613",
"0.3984",
"0.8613",
"0.4124",
"0.505",
"0.4124"
],
[
"0.523",
"0.4136",
"0.8896",
"0.4124",
"0.8896",
"0.4268",
"0.523",
"0.428"
],
[
"0.0825",
"0.4268",
"0.4673",
"0.4268",
"0.4673",
"0.44",
"0.0825",
"0.44"
],
[
"0.505",
"0.4287",
"0.8896",
"0.4287",
"0.8896",
"0.4426",
"0.505",
"0.4426"
],
[
"0.0825",
"0.4426",
"0.2778",
"0.4426",
"0.2778",
"0.4558",
"0.0825",
"0.4558"
],
[
"0.5024",
"0.4426",
"0.8906",
"0.443",
"0.8906",
"0.4583",
"0.5024",
"0.457"
],
[
"0.505",
"0.459",
"0.892",
"0.459",
"0.892",
"0.473",
"0.505",
"0.473"
],
[
"0.1462",
"0.4717",
"0.4028",
"0.4717",
"0.4028",
"0.486",
"0.1462",
"0.486"
],
[
"0.4478",
"0.4749",
"0.4656",
"0.4749",
"0.4656",
"0.4841",
"0.4478",
"0.4841"
],
[
"0.5034",
"0.4749",
"0.6406",
"0.4749",
"0.6406",
"0.488",
"0.5034",
"0.488"
],
[
"0.5215",
"0.4893",
"0.888",
"0.4905",
"0.888",
"0.505",
"0.5215",
"0.504"
],
[
"0.10046",
"0.4968",
"0.468",
"0.4968",
"0.468",
"0.511",
"0.10046",
"0.511"
],
[
"0.505",
"0.505",
"0.8906",
"0.505",
"0.8906",
"0.5195",
"0.505",
"0.5195"
],
[
"0.0825",
"0.512",
"0.468",
"0.512",
"0.468",
"0.5264",
"0.0825",
"0.5264"
],
[
"0.0809",
"0.5264",
"0.1478",
"0.5264",
"0.1478",
"0.5415",
"0.0809",
"0.5415"
],
[
"0.505",
"0.521",
"0.892",
"0.521",
"0.892",
"0.534",
"0.505",
"0.534"
],
[
"0.5034",
"0.534",
"0.7646",
"0.535",
"0.7646",
"0.55",
"0.5034",
"0.549"
],
[
"0.08496",
"0.56",
"0.468",
"0.56",
"0.468",
"0.5737",
"0.08496",
"0.5737"
],
[
"0.0825",
"0.5747",
"0.4656",
"0.5747",
"0.4656",
"0.5884",
"0.0825",
"0.5884"
],
[
"0.509",
"0.5747",
"0.8906",
"0.5747",
"0.8906",
"0.5884",
"0.509",
"0.5884"
],
[
"0.0809",
"0.5903",
"0.4673",
"0.5903",
"0.4673",
"0.604",
"0.0809",
"0.604"
],
[
"0.5034",
"0.5903",
"0.8906",
"0.5903",
"0.8906",
"0.604",
"0.5034",
"0.604"
],
[
"0.0809",
"0.606",
"0.468",
"0.606",
"0.468",
"0.619",
"0.0809",
"0.619"
],
[
"0.5146",
"0.604",
"0.8906",
"0.604",
"0.8906",
"0.617",
"0.5146",
"0.617"
],
[
"0.0948",
"0.6206",
"0.468",
"0.619",
"0.468",
"0.634",
"0.0948",
"0.6353"
],
[
"0.505",
"0.6206",
"0.888",
"0.6206",
"0.888",
"0.634",
"0.505",
"0.634"
],
[
"0.0809",
"0.636",
"0.468",
"0.636",
"0.468",
"0.6494",
"0.0809",
"0.6494"
],
[
"0.5034",
"0.634",
"0.8896",
"0.634",
"0.8896",
"0.6484",
"0.5034",
"0.6484"
],
[
"0.0809",
"0.651",
"0.4673",
"0.651",
"0.4673",
"0.664",
"0.0809",
"0.664"
],
[
"0.5195",
"0.6494",
"0.8906",
"0.6494",
"0.8906",
"0.6636",
"0.5195",
"0.6636"
],
[
"0.0801",
"0.664",
"0.468",
"0.6655",
"0.468",
"0.6797",
"0.0801",
"0.6787"
],
[
"0.505",
"0.6655",
"0.892",
"0.6655",
"0.892",
"0.6787",
"0.505",
"0.6787"
],
[
"0.0801",
"0.6787",
"0.4673",
"0.681",
"0.4673",
"0.696",
"0.0801",
"0.6934"
],
[
"0.5034",
"0.6797",
"0.603",
"0.6797",
"0.603",
"0.6943",
"0.5034",
"0.6943"
],
[
"0.0809",
"0.696",
"0.468",
"0.696",
"0.468",
"0.7104",
"0.0809",
"0.7104"
],
[
"0.523",
"0.6963",
"0.8906",
"0.6963",
"0.8906",
"0.711",
"0.523",
"0.711"
],
[
"0.0801",
"0.7104",
"0.468",
"0.709",
"0.468",
"0.7236",
"0.0801",
"0.7246"
],
[
"0.5024",
"0.711",
"0.8896",
"0.7104",
"0.8896",
"0.7246",
"0.5024",
"0.726"
],
[
"0.09644",
"0.7246",
"0.4673",
"0.7246",
"0.4673",
"0.7393",
"0.09644",
"0.7393"
],
[
"0.5063",
"0.731",
"0.523",
"0.731",
"0.523",
"0.738",
"0.5063",
"0.738"
],
[
"0.517",
"0.7266",
"0.892",
"0.7266",
"0.892",
"0.7407",
"0.517",
"0.7407"
],
[
"0.5024",
"0.7393",
"0.546",
"0.741",
"0.544",
"0.757",
"0.501",
"0.755"
],
[
"0.0866",
"0.732",
"0.0989",
"0.732",
"0.0989",
"0.738",
"0.0866",
"0.738"
],
[
"0.0801",
"0.7407",
"0.3447",
"0.741",
"0.3447",
"0.756",
"0.0801",
"0.755"
],
[
"0.0801",
"0.7666",
"0.343",
"0.7666",
"0.343",
"0.78",
"0.0801",
"0.78"
],
[
"0.5034",
"0.771",
"0.556",
"0.771",
"0.556",
"0.782",
"0.5034",
"0.782"
],
[
"0.5513",
"0.768",
"0.8896",
"0.768",
"0.8896",
"0.782",
"0.5513",
"0.782"
],
[
"0.5435",
"0.785",
"0.841",
"0.785",
"0.841",
"0.8",
"0.5435",
"0.8"
],
[
"0.1978",
"0.797",
"0.343",
"0.797",
"0.343",
"0.81",
"0.1978",
"0.81"
],
[
"0.0907",
"0.8184",
"0.4534",
"0.8184",
"0.4534",
"0.8335",
"0.0907",
"0.8335"
],
[
"0.636",
"0.8164",
"0.7534",
"0.8164",
"0.7534",
"0.83",
"0.636",
"0.83"
],
[
"0.10706",
"0.834",
"0.429",
"0.834",
"0.429",
"0.8486",
"0.10706",
"0.8486"
],
[
"0.511",
"0.8403",
"0.8784",
"0.8403",
"0.8784",
"0.854",
"0.511",
"0.854"
],
[
"0.1013",
"0.8486",
"0.4355",
"0.8486",
"0.4355",
"0.863",
"0.1013",
"0.863"
],
[
"0.2181",
"0.863",
"0.33",
"0.8613",
"0.33",
"0.877",
"0.2181",
"0.879"
],
[
"0.5293",
"0.863",
"0.864",
"0.863",
"0.864",
"0.8774",
"0.5293",
"0.8774"
],
[
"0.1013",
"0.884",
"0.4453",
"0.887",
"0.4453",
"0.9014",
"0.1013",
"0.8984"
],
[
"0.5767",
"0.8853",
"0.8164",
"0.886",
"0.8164",
"0.901",
"0.5767",
"0.8994"
]
] | [
"XiaomiMMAI团队提出的解决方案是基于HATE[?",
"Boa",
"］的双分支网络，在训练过程中采用重新参数化E[？]",
"使用附加参数充分挖掘方法的潜力。他们正在介绍",
"种针对RAW图像超分辨率的逐个任务和分步训练方",
"法，以同时解决三个任务：降噪，去模糊和2×超分",
"辨率",
"为了解决训练样本数量少的限制，XiaomiMMAI团",
"队将RAW图像转换为RGB图像，并使用随机旋转",
"翻转，颜色变化，亮度变化，随机模糊等。然后，将增",
"强的RGB图像集转换回RGGBRAW域，并将处理后",
"Stagez",
"的数据用于训练E[？1所提出的解决方案",
"XiaomiMMAI团队提出的模型灵感来自HAT[?",
"",
"其架构针对RAWImageSuperResolution任务进行了",
"优化。优化的双分支网络结构(DB-HAT）如Fig.3所",
"示。为RAWISR引人的逐步和逐个任务的训练方法进",
"一步提高了他们的解决方案所达到的性能水平",
"步一步：为了加速训练并获得良好的性能，Xi-",
"aomiMMAI团队采用了一种策略，其中每个子任务",
"Figure 2.Dual-Stage RAWSRFrameworkproposed by",
"（包括最终的联合优化）都基于金字塔图像表示进行训",
"Team Samsung MX,SRC. Best viewed in electronic version.",
"练。最初，该模型是在小规模图像（64??？）上训练的，",
"逐渐将图像补丁的分辨率提高到128??？和256??？",
"逐个任务：XiaomiMMAI团队将RAWISR分为三",
"因此，焦点像素损失（FPL）正在对那些以大信号偏移",
"个子任务：去噪，去模糊和2xSR。最初，他们开始训",
"为特征的像素引人指数惩罚",
"练RAW图像去噪，然后是去模糊和2×超级分辨率",
"的连接任务。最后，对整个网络应用联合优化程序，以",
"FPL(1,I) = -D(1, I)log1 D(1, I)",
"(1)",
"产生最终的估计器。",
"在训练去噪和去模糊的过程中，XiaomiMMAI团队",
"在Eq（1）中，D.,）是恢复图像「和参考图像I",
"在提出的DB-HAT的最后阶段使用了RepConv重新",
"之间的标准L范数距离，是一个可调因子，控制惩",
"罚的强度",
"参数化技术，以提高任务的视觉图像质量。可重新参数",
"化的卷积块（RepConv）如图4所示",
"Implementationdetails该模型仅根据挑战组织者提",
"供的数据进行训练。它仅包含来自各种DSLR相机传",
"Implementationdetails用于三个子任务训练的数据",
"感器的1000多个RAW图像。使用标准图像增强技",
"集由1000+个RAW组成，数据增强方法可以参考上",
"术和图像退化管道的模拟来增强数据集。训练过程是",
"节。在整个网络联合优化的最后阶段，XiaomiMMAI",
"个双阶段操作，依次优化模型阶段。所采用的优化",
"团队使用提供的1000+数据集，而不是增强的数据集，",
"技术是采用余弦退火策略的AdamW[？］优化器（",
"在三个子任务的训练过程中，学习率初始化为4×10-4",
"β1=0.5，β2=0.999，权重衰减0.0001），其中对于",
"并根据余弦退火策略衰减。网络经过总共2x105迭",
"5×105迭代，学习率从初始学习率5×10-5逐渐降低",
"代的训练，其中L2损失函数被最小化为Adam优化",
"到1×10-7。该模型经历了预训练优化阶段，基于L1",
"器的训练目标",
"损失。在fune-tune阶段，目标设置为FocalPixelLoss，",
"随后，使用L2损失和SSIM损失函数对两次送代",
"初始学习速率设置为5×10-6。训练批大小设置为4",
"执行微调，对于2x105迭代，初始学习率为5×10-5",
"补丁大小为384。水平/垂直翻转和旋转用于数据增",
"",
"所有实验均使用PyTorch2。8个A100GPU上的0",
"框架。",
"",
"强。所有实验均在A100GPU上进行",
"3.2. EffectiveSR for RAW Images",
"3.3.",
"RBSFormer: Enhanced Transformer Net-",
"work for Raw Image Super-Resolution",
"Team XiaomiMMAI",
"Zhijuan Huang 1 , Hongyuan Yu 1, Cheng Wan ?",
"Team USTC604",
"Wending Xiang, Jiamin Lin , Hang Zhong",
"Senyan Xu, Siyuan Jiang, Zhijing Sun, Jiaying Zhu",
"Qiaosong Zhang 1, Yue Sun 1, Xuanwu Yin 1",
"Kunlong Zuo I",
"University of Science and Technology of China",
"1 Xiaomi Inc. 2 Georgia Institute of Technology",
"Contact: [email protected]"
] | {
"arxiv_id": "2404.16223",
"language": "zh",
"page_idx": 3
} |
||
2010.13870-zh-0 | 2010.13870-zh-0.jpg | ¥ documentclass[ 11pt, a4paper] ¥ XepMATHX$\_$0¥ usepackage[ hyperref] emnlp2020
Yusepackage times ¥usepackage latexsym ¥renewcommandTHuspackage
$¥$UrlFont$\mod$Y
Yaclfnalcopy
Xnewcommand¥ BibTeX B¥ textsc ib YTeX
¥title Word Frequency Does Not Predict Grammatical Knowledge in Lan-
guage Models
¥ 作者 Charles Yu¥thanks These authors contributed equally. , Ryan
Sie¥footnotemark[1], Nico Tedeschi, and Leon Bergen ¥¥ ¥textt ¥ cty002,rdsie,njtedeschi,lbergen¥
$\mathbb{Q}$ucsd.edu ¥¥ University of California, San Diego¥¥
$Y$日期
¥ 开始 document ¥maketitle¥ 开始 abstract 神经语言模型以不同的准确度学习自然语言的语法属性。在这项工作中，我们调查语言模型的准确性是否存在系统的变化来源。关注主语-动词一致性和反身回指，我们发现某些名词比其他名词更容易被系统理解，这种效果在语法任务和不同的语言模型中都很强大。令人惊讶的是，我们发现在四个数量级上，语料库频率与名词在语法任务上的表现无关。。最后，我们发现可以从各种类型的训练数据中很少学习新名词的语法属性。结果提出了一个悖论：语法表现的变化应该比实际观察到的要少。${\mathrm{Yend}}$ ${\mathrm{abstract}}$
$\Psi$部分 Introduction
神经语言模型 ¥citep howard2018universal,devlin2019bert,dai2019transformer
yang2019xlnet,radford2019language 在文本预测和下游任务(如间答，文本分类和自然语言推理) 中都取得了成功。这些模型的强大性能提出了关于他们所获得的知识的科学问题，特别是关于他们的语言表示的抽象性和一般性
以前的工作已经研究了几个领域中神经语言模型的语言表示，并发现了不同的证据，证明这些表示在语言上是足够的 ¥citep lau2017grammaticality, marvin2018targeted, goldberg2019assessing, futrell2019neural 。这 项 工 作 采 用 了心理语言学方法，以从这些模型中得出语法判断，从而从判断模式中推断出模型的基本表示形式。
在当前的工作中，我们专注于神经语言模型中可能存在的语法知识的变化。就像在人类心理语言学任务中一样，以前关于神经 LMs 的工作已经观察到不同句子之间语法判断的差异；并非所有违反语法约束的行为都被认为是同样糟糕的。然而，目前还不清楚这些判断中是否有系统的变异来源，如果有，来源是什么
我们将使用英语主语-动词一致性和反身回指作为案例研究，重点关注词汇项之间的变化。我们首先问语言模型是否比其他模型更准确地学习某些名词的语法属性。我们通过在进行涉及不同名词的语法判断时测量语言模型的准确性来做到这一点。我们发现名词之间存在系统差异：在一个任务或语言模型上表现良好的名词更有可能在其他任务或其他语言模型上表现良好。然后，我们考虑名词之间观察到的变化的可能来源，发现名词的语法属性很容易学习。我们的结果表明， 变化应该比实际观察到的要少得多。。¥ 脚注 All code and experimental materials are available at ¥url https://github.com/CharlesYu2000/lm-variation
$\Psi$小节$^*$ Related work
许多其他研究已经研究了神经模型的语言表示，包括专门的语言模型和使用其他目标训练的网络。¥citet linzen2016assessing, gulordava2018colorless, kuncoro2018lstms 探究 LSTM 学习层次结构的能力。¥citet warstadt2019neural
1 | [
[
"0.2427",
"0.159",
"0.8267",
"0.159",
"0.8267",
"0.1736",
"0.2427",
"0.1736"
],
[
"0.2173",
"0.1736",
"0.7793",
"0.1749",
"0.7793",
"0.1895",
"0.2173",
"0.1881"
],
[
"0.2173",
"0.1895",
"0.3906",
"0.1895",
"0.3906",
"0.2026",
"0.2173",
"0.2026"
],
[
"0.2418",
"0.2026",
"0.339",
"0.2052",
"0.339",
"0.2197",
"0.2418",
"0.2172"
],
[
"0.2427",
"0.2197",
"0.5728",
"0.2197",
"0.5728",
"0.2329",
"0.2427",
"0.2329"
],
[
"0.2418",
"0.2349",
"0.7754",
"0.2349",
"0.7754",
"0.2488",
"0.2418",
"0.2488"
],
[
"0.2173",
"0.252",
"0.3162",
"0.2507",
"0.3162",
"0.262",
"0.2173",
"0.2632"
],
[
"0.2402",
"0.2651",
"0.7793",
"0.2651",
"0.7793",
"0.279",
"0.2402",
"0.279"
],
[
"0.2157",
"0.279",
"0.9395",
"0.2798",
"0.9395",
"0.2942",
"0.2157",
"0.2935"
],
[
"0.2173",
"0.2942",
"0.607",
"0.2954",
"0.607",
"0.31",
"0.2173",
"0.3093"
],
[
"0.2418",
"0.31",
"0.2942",
"0.31",
"0.2942",
"0.3245",
"0.2418",
"0.3245"
],
[
"0.2385",
"0.3245",
"0.778",
"0.3245",
"0.778",
"0.3384",
"0.2385",
"0.3384"
],
[
"0.2197",
"0.3423",
"0.777",
"0.3423",
"0.777",
"0.3528",
"0.2197",
"0.3528"
],
[
"0.2181",
"0.356",
"0.7793",
"0.356",
"0.7793",
"0.3694",
"0.2181",
"0.3694"
],
[
"0.2181",
"0.3718",
"0.777",
"0.3718",
"0.777",
"0.3818",
"0.2181",
"0.3818"
],
[
"0.2197",
"0.387",
"0.777",
"0.387",
"0.777",
"0.3977",
"0.2197",
"0.3977"
],
[
"0.2214",
"0.4016",
"0.7793",
"0.4016",
"0.7793",
"0.4153",
"0.2214",
"0.4153"
],
[
"0.2181",
"0.4167",
"0.777",
"0.4167",
"0.777",
"0.43",
"0.2181",
"0.43"
],
[
"0.2173",
"0.4312",
"0.3245",
"0.4312",
"0.3245",
"0.4443",
"0.2173",
"0.4443"
],
[
"0.2418",
"0.4463",
"0.3882",
"0.4463",
"0.3882",
"0.4602",
"0.2418",
"0.4602"
],
[
"0.2418",
"0.461",
"0.8013",
"0.4622",
"0.8013",
"0.476",
"0.2418",
"0.4749"
],
[
"0.2173",
"0.4766",
"0.778",
"0.4766",
"0.778",
"0.4905",
"0.2173",
"0.4905"
],
[
"0.2197",
"0.4924",
"0.7793",
"0.4924",
"0.7793",
"0.506",
"0.2197",
"0.506"
],
[
"0.2197",
"0.5083",
"0.722",
"0.5083",
"0.722",
"0.518",
"0.2197",
"0.518"
],
[
"0.2418",
"0.521",
"0.7812",
"0.5215",
"0.7812",
"0.536",
"0.2418",
"0.535"
],
[
"0.2181",
"0.537",
"0.778",
"0.537",
"0.778",
"0.5503",
"0.2181",
"0.5503"
],
[
"0.2197",
"0.5537",
"0.777",
"0.5537",
"0.777",
"0.5645",
"0.2197",
"0.5645"
],
[
"0.2197",
"0.5664",
"0.778",
"0.5664",
"0.778",
"0.58",
"0.2197",
"0.58"
],
[
"0.2181",
"0.584",
"0.3708",
"0.584",
"0.3708",
"0.5947",
"0.2181",
"0.5947"
],
[
"0.2427",
"0.5977",
"0.778",
"0.5977",
"0.778",
"0.6113",
"0.2427",
"0.6113"
],
[
"0.2181",
"0.6123",
"0.778",
"0.6123",
"0.778",
"0.6255",
"0.2181",
"0.6255"
],
[
"0.2197",
"0.6274",
"0.7793",
"0.6274",
"0.7793",
"0.6416",
"0.2197",
"0.6416"
],
[
"0.2173",
"0.6416",
"0.7793",
"0.6426",
"0.7793",
"0.657",
"0.2173",
"0.656"
],
[
"0.2181",
"0.659",
"0.25",
"0.659",
"0.25",
"0.6704",
"0.2181",
"0.6704"
],
[
"0.2418",
"0.6724",
"0.7812",
"0.6724",
"0.7812",
"0.686",
"0.2418",
"0.686"
],
[
"0.2173",
"0.686",
"0.7793",
"0.6875",
"0.7793",
"0.702",
"0.2173",
"0.7007"
],
[
"0.2197",
"0.705",
"0.778",
"0.705",
"0.778",
"0.7153",
"0.2197",
"0.7153"
],
[
"0.2197",
"0.7183",
"0.7793",
"0.7183",
"0.7793",
"0.732",
"0.2197",
"0.732"
],
[
"0.2197",
"0.7344",
"0.778",
"0.7344",
"0.778",
"0.7476",
"0.2197",
"0.7476"
],
[
"0.2181",
"0.7485",
"0.778",
"0.7485",
"0.778",
"0.762",
"0.2181",
"0.762"
],
[
"0.2157",
"0.763",
"0.7812",
"0.762",
"0.7812",
"0.7764",
"0.2157",
"0.778"
],
[
"0.2157",
"0.778",
"0.721",
"0.779",
"0.721",
"0.7935",
"0.2157",
"0.7925"
],
[
"0.2418",
"0.7935",
"0.4126",
"0.7935",
"0.4126",
"0.8066",
"0.2418",
"0.8066"
],
[
"0.2418",
"0.8096",
"0.778",
"0.8096",
"0.778",
"0.8228",
"0.2418",
"0.8228"
],
[
"0.2181",
"0.8247",
"0.777",
"0.8247",
"0.777",
"0.8384",
"0.2181",
"0.8384"
],
[
"0.2157",
"0.839",
"0.778",
"0.839",
"0.778",
"0.853",
"0.2157",
"0.853"
],
[
"0.492",
"0.8774",
"0.5063",
"0.8774",
"0.5063",
"0.888",
"0.492",
"0.888"
]
] | [
"Ydocumentclass[11pt,a4paper]?XepMATHX_0Yusepackage[hyperref] emnlp2020",
"Yusepackage times Yusepackage latexsym YrenewcommandTHuspackage",
"￥UrlFontmultirow",
"Yaclfinalcopy",
"￥newcommand?BibTeX B?textsc ib TeX",
"￥title Word Frequency Does Not Predict Grammatical Knowledge in Lan",
"guage Models",
"￥作者Charles Yu?thanksThese authors contributed equally.，Ryan",
"Sie?footnotemark[1], Nico Tedeschi, and Leon Bergen ?? ?texttt ￥ ctyo02,rdsie,njtedeschi,lbergen?",
"@ucsd.edu ￥Y University of California, San Diego??",
"?日期",
"￥开始document￥maketitle?开始abstract神经语言模型以不同的准确",
"度学习自然语言的语法属性。在这项工作中，我们调查语言模型的准确性是否",
"存在系统的变化来源。关注主语-动词一致性和反身回指，我们发现某些名词比",
"其他名词更容易被系统理解，这种效果在语法任务和不同的语言模型中都很强",
"大。令人惊的是，我们发现在四个数量级上，语料库频率与名词在语法任务",
"上的表现无关。。最后，我们发现可以从各种类型的训练数据中很少学习新名词",
"的语法属性。结果提出了一个悖论：语法表现的变化应该比实际观察到的要少。",
"Yend abstract",
"￥部分Introduction",
"神经语言模型￥citephoward2018universal,devlin2019bert,dai2019transformer，",
"yang2019xlnet，radford2019language在文本预测和下游任务（如问答，文本分",
"类和自然语言推理）中都取得了成功。这些模型的强大性能提出了关于他们所",
"获得的知识的科学问题，特别是关于他们的语言表示的抽象性和一般性",
"以前的工作已经研究了几个领域中神经语言模型的语言表示，并发现了不",
"同的证据，证明这些表示在语言上是足够的￥citeplau2017grammaticality",
"marvin2018targeted，goldberg2019assessing，futrell2019neural：这项工作采用",
"了心理语言学方法，以从这些模型中得出语法判断，从而从判断模式中推断出",
"模型的基本表示形式。",
"在当前的工作中，我们专注于神经语言模型中可能存在的语法知识的变化",
"就像在人类心理语言学任务中一样，以前关于神经LMs的工作已经观察到不同",
"句子之间语法判断的差异；并非所有违反语法约束的行为都被认为是同样糟糕",
"的。然而，目前还不清楚这些判断中是否有系统的变异来源，如果有，来源是",
"什么",
"我们将使用英语主语-动词一致性和反身回指作为案例研究，重点关注词汇项",
"之间的变化。我们首先问语言模型是否比其他模型更准确地学习某些名词的语",
"法属性。我们通过在进行涉及不同名词的语法判断时测量语言模型的准确性来做",
"到这一点。我们发现名词之间存在系统差异：在一个任务或语言模型上表现良好",
"的名词更有可能在其他任务或其他语言模型上表现良好。然后，我们考虑名词之",
"间观察到的变化的可能来源，发现名词的语法属性很容易学习。我们的结果表明",
"变化应该比实际观察到的要少得多。。?脚注Allcodeandexperimentalmaterials",
"are available at Yurl https://github.com/CharlesYu2000/lm-variation",
"￥小节*Relatedwork",
"许多其他研究已经研究了神经模型的语言表示，包括专门的语言模型和使用",
"其他目标训练的网络。￥citetlinzen2016assessing，gulordava2018colorless，kun",
"coro2018lstms探究LSTM学习层次结构的能力。￥citetwarstadt2019neural",
"1"
] | {
"arxiv_id": "2010.13870",
"language": "zh",
"page_idx": 0
} |
||
2010.13870-zh-1 | 2010.13870-zh-1.jpg | 引人了一个大规模的语法可接受性判断语料库，训练 RNN 来预测这些判断， 并得出结论，这些模型优于无监督基线，但远远低于人类的表现。¥citet lepori2020representations 发现基于树的 RNN 在数字预测任务上优于顺序 RNN, 但是对人工生成的增强集进行微调可以使模型更接近奇偶校验
其他工作的重点是探索神经语言模型是否已经获得了特定语言现象的适当表示。¥citet marvin2018targeted 和 ¥citet goldberg2019assessing 使用最小对方法来评估 RNN 和 BERT 的语法知识，查看主语动词数量一致，反身回指和负极性项目。¥citet wilcox2018rnn 检查 RNN 语言模型是否在与间隙相关的意外情况下表现出 ¥textit wh 许可交互，得出的结论是它们可以表示长距离填充间隙依赖关系并学习某些岛约束。¥citet futrell2019neural 研究神经语言模型是否显示使用心理语言学方法的增量句法状态表示的证据。¥citet warstadt2019investigating 研究了 BERT 对 NPI 的了解，重点是任务之间的差异：布尔分类 (e。g。¥citealt linzen2016assessing 和 ¥citealt warstadt2019neural ),最小对比较 (e。g。¥citealt marvin2018targeted 和¥ citealt wilcox2019structural )和探测任务( $e. g. ¥ citealt giulianelli2018under$ )。¥ 节 Approach 我们使用 ¥citet marvin2018targeted 的最小对方法来研究神经语言模型的语法判断。给定一对最少的句子，我。e。由于仅一个语法属性的差异而在可接受性上彼此不同的一对。如果模型理解正在研究的语法现象， 它应该为语法句子分配比非语法句子更高的概率
$\Psi$小节 Grammatical tasks
表 ¥ref tab:title 显示了 10 个语法任务 ¥citep marvin2018targeted 以及用于生成最小对的模板。任务分为两大类：主语动词一致(SVA)和反身回指(RA)。第一个 SVA 任务 SVASimple 调查模型是否理解主题号必须与第三人称现在动词的数量一致：Ypex$[ ^*$= $? ^*$] < $判 断$> ¥ a 猫走路。Ya¥ ljudge * 猫走路。¥ xe 其 他 SVA 任务探测模型是否具有更复杂的数字一致性表示。例如，SVAPP 任务测量模型是否能够忽略出现在主体头部和动词之间的干扰因素(“男孩”): Ypex$[ ^{* }$= $? ^{* }$] < judgements> ¥ a 男 孩 旁 边 的 猫 跳 。¥ a¥ ljudge * 男孩旁边的猫跳。¥xe 对象相对子句任务探测在存在嵌人子句的情况下模型是否准确地保持头部的编号。¥citet marvin2018targeted 对这些任务的语言动机进行了广泛的讨论
RA 任务测量语言模型是否理解反身代词绑定的结构条件。这些任务利用了英语反身语的以下属性：反身代词需要在数量上与其前置词一致。RA 发送。Comp 任务评估模型是否理解自反性必须与它们的前身在同一子句中： Ypex$[ ^{* }$= $? ^{ * }$] < 判 决 > ¥ a 律 师 说 , 被 告 自 证 其 罪 。¥ a¥ ljudge$ ^{* }$ 律师说，被告自证其罪。¥xe 涉及宾语相对从句的 RA 任务评估模型是否理解自反性回指不绑定到嵌入从句中的名词，而是绑定到头部名词
$¥$开始 table* ¥ 居中 ¥ 调整大小框 0.8¥paperwidth ! ¥begin tabular ||||| Yhline ¥bf Task ¥bf Template ¥¥ ¥hline¥hline SVA Simple The Ylangle TargetNoun Yrangle ¥ langle Verb ¥ rangle . ¥ ¥ SVA Subj. Rel. Clause The ¥langle TargetNoun ¥rangle that liked the ¥langle Noun ¥rangle ${\mathrm{¥langle~Verb~}}{\mathrm{Yrangle~.~}}{\mathrm{Y¥~SVA~Sent.Comp.~The~}}{\mathrm{Ylangle~Noun~}}{\mathrm{Yrangle}}$ said the Ylangle TargetNoun Yrangle Ylangle Verb Y$rangle.$ ¥¥ SVA PP The $\Psi langle$ TargetNoun ¥rangle next to the ¥langle Noun ¥rangle Ylangle Verb ¥rangle . ¥¥ SVA Obj.Rel. Clause.That The ¥langle TargetNoun ${\mathrm{¥}}rangle$ that the $\Psi langle$ Noun Yrangle liked ¥langle Verb ¥$rangle.$ ¥¥ SVA Obj.Rel.Clause.NoThat The ¥langle TargetNoun ¥rangle the Ylangle
2 | [
[
"0.2181",
"0.1604",
"0.7812",
"0.1604",
"0.7812",
"0.1736",
"0.2181",
"0.1736"
],
[
"0.2181",
"0.1749",
"0.778",
"0.1749",
"0.778",
"0.1881",
"0.2181",
"0.1881"
],
[
"0.2181",
"0.1895",
"0.7817",
"0.1895",
"0.7817",
"0.2026",
"0.2181",
"0.2026"
],
[
"0.2181",
"0.2052",
"0.66",
"0.2052",
"0.66",
"0.2184",
"0.2181",
"0.2184"
],
[
"0.2427",
"0.2197",
"0.7793",
"0.2197",
"0.7793",
"0.2329",
"0.2427",
"0.2329"
],
[
"0.2181",
"0.2349",
"0.7793",
"0.2349",
"0.7793",
"0.2488",
"0.2181",
"0.2488"
],
[
"0.2181",
"0.2507",
"0.7793",
"0.2507",
"0.7793",
"0.2644",
"0.2181",
"0.2644"
],
[
"0.2197",
"0.2651",
"0.7793",
"0.2651",
"0.7793",
"0.279",
"0.2197",
"0.279"
],
[
"0.2181",
"0.281",
"0.778",
"0.281",
"0.778",
"0.2942",
"0.2181",
"0.2942"
],
[
"0.2181",
"0.2954",
"0.7793",
"0.2954",
"0.7793",
"0.3093",
"0.2181",
"0.3093"
],
[
"0.2173",
"0.31",
"0.7812",
"0.3113",
"0.7812",
"0.3245",
"0.2173",
"0.324"
],
[
"0.2173",
"0.3245",
"0.7793",
"0.3257",
"0.7793",
"0.3403",
"0.2173",
"0.339"
],
[
"0.2181",
"0.3403",
"0.7793",
"0.3403",
"0.7793",
"0.3542",
"0.2181",
"0.3542"
],
[
"0.2197",
"0.356",
"0.7817",
"0.356",
"0.7817",
"0.3694",
"0.2197",
"0.3694"
],
[
"0.2181",
"0.3706",
"0.7793",
"0.3706",
"0.7793",
"0.3838",
"0.2181",
"0.3838"
],
[
"0.2181",
"0.3865",
"0.7793",
"0.3865",
"0.7793",
"0.3997",
"0.2181",
"0.3997"
],
[
"0.2181",
"0.4016",
"0.7793",
"0.4016",
"0.7793",
"0.4153",
"0.2181",
"0.4153"
],
[
"0.2181",
"0.4167",
"0.7793",
"0.4167",
"0.7793",
"0.43",
"0.2181",
"0.43"
],
[
"0.2181",
"0.433",
"0.558",
"0.433",
"0.558",
"0.443",
"0.2181",
"0.443"
],
[
"0.2418",
"0.4463",
"0.4355",
"0.4463",
"0.4355",
"0.4602",
"0.2418",
"0.4602"
],
[
"0.2402",
"0.4622",
"0.778",
"0.4622",
"0.778",
"0.476",
"0.2402",
"0.476"
],
[
"0.2197",
"0.4766",
"0.7793",
"0.4766",
"0.7793",
"0.4905",
"0.2197",
"0.4905"
],
[
"0.2173",
"0.4924",
"0.7793",
"0.4924",
"0.7793",
"0.506",
"0.2173",
"0.506"
],
[
"0.2181",
"0.507",
"0.7793",
"0.507",
"0.7793",
"0.521",
"0.2181",
"0.521"
],
[
"0.2181",
"0.5215",
"0.7793",
"0.5215",
"0.7793",
"0.535",
"0.2181",
"0.535"
],
[
"0.2197",
"0.5386",
"0.7812",
"0.5386",
"0.7812",
"0.549",
"0.2197",
"0.549"
],
[
"0.2181",
"0.552",
"0.7793",
"0.552",
"0.7793",
"0.565",
"0.2181",
"0.565"
],
[
"0.2173",
"0.5674",
"0.7793",
"0.5674",
"0.7793",
"0.5806",
"0.2173",
"0.5806"
],
[
"0.2181",
"0.582",
"0.778",
"0.582",
"0.778",
"0.595",
"0.2181",
"0.595"
],
[
"0.2181",
"0.5986",
"0.2507",
"0.5986",
"0.2507",
"0.61",
"0.2181",
"0.61"
],
[
"0.2418",
"0.6123",
"0.7793",
"0.6123",
"0.7793",
"0.6255",
"0.2418",
"0.6255"
],
[
"0.2197",
"0.6274",
"0.7793",
"0.6274",
"0.7793",
"0.6416",
"0.2197",
"0.6416"
],
[
"0.2157",
"0.6416",
"0.7817",
"0.6426",
"0.7817",
"0.657",
"0.2157",
"0.656"
],
[
"0.2181",
"0.6577",
"0.7793",
"0.6577",
"0.7793",
"0.672",
"0.2181",
"0.672"
],
[
"0.2197",
"0.6724",
"0.7793",
"0.6724",
"0.7793",
"0.686",
"0.2197",
"0.686"
],
[
"0.2173",
"0.686",
"0.5605",
"0.686",
"0.5605",
"0.6997",
"0.2173",
"0.6997"
],
[
"0.2402",
"0.702",
"0.726",
"0.7007",
"0.726",
"0.7153",
"0.2402",
"0.7163"
],
[
"0.7183",
"0.7007",
"0.7812",
"0.704",
"0.7793",
"0.7183",
"0.7173",
"0.7153"
],
[
"0.2173",
"0.7183",
"0.7812",
"0.7183",
"0.7812",
"0.732",
"0.2173",
"0.732"
],
[
"0.2181",
"0.733",
"0.778",
"0.733",
"0.778",
"0.747",
"0.2181",
"0.747"
],
[
"0.2173",
"0.7485",
"0.7812",
"0.7485",
"0.7812",
"0.762",
"0.2173",
"0.762"
],
[
"0.2181",
"0.763",
"0.7793",
"0.763",
"0.7793",
"0.778",
"0.2181",
"0.778"
],
[
"0.2173",
"0.779",
"0.7793",
"0.779",
"0.7793",
"0.7925",
"0.2173",
"0.7925"
],
[
"0.2157",
"0.7925",
"0.7812",
"0.7935",
"0.7812",
"0.808",
"0.2157",
"0.8066"
],
[
"0.2181",
"0.8096",
"0.778",
"0.8096",
"0.778",
"0.8228",
"0.2181",
"0.8228"
],
[
"0.2181",
"0.8247",
"0.7793",
"0.8247",
"0.7793",
"0.8384",
"0.2181",
"0.8384"
],
[
"0.2157",
"0.837",
"0.7793",
"0.8384",
"0.7793",
"0.853",
"0.2157",
"0.8516"
],
[
"0.4902",
"0.8755",
"0.5073",
"0.8755",
"0.5073",
"0.89",
"0.4902",
"0.89"
]
] | [
"引人了一个大规模的语法可接受性判断语料库，训练RNN来预测这些判断，",
"并得出结论，这些模型优于无监督基线，但远远低于人类的表现。￥citetlep",
"ori2020representations发现基于树的RNN在数字预测任务上优于顺序RNN，",
"但是对人工生成的增强集进行微调可以使模型更接近奇偶校验",
"其他工作的重点是探索神经语言模型是否已经获得了特定语言现象的适",
"当表示。￥citetmarvin2018targeted和￥citetgoldberg2019assessing使用",
"最小对方法来评估RNN和BERT的语法知识，查看主语动词数量一致，反",
"身回指和负极性项目。￥citetwilcox2018rnn检查RNN语言模型是否在与",
"间隙相关的意外情况下表现出￥textitwh许可交互，得出的结论是它们可",
"以表示长距离填充间隙依赖关系并学习某些岛约束。￥citetfutrell2019neural",
"研究神经语言模型是否显示使用心理语言学方法的增量句法状态表示的证",
"据。￥citetwarstadt2019investigating研究了BERT对NPI的了解，重点是",
"任务之间的差异：布尔分类（e。g。￥citealtlinzen2016assessing和citealt",
"warstadt2019neural），最小对比较（e。g。￥citealtmarvin2018targeted和",
"￥citealtwilcox2019structural）和探测任务（e.g。￥citealtgiulianelli2018under",
")。?节Approach我们使用?citetmarvin2018targeted的最小对方法来研",
"究神经语言模型的语法判断。给定一对最少的句子，我。e。由于仅一个语法属",
"性的差异而在可接受性上彼此不同的一对。如果模型理解正在研究的语法现象",
"它应该为语法句子分配比非语法句子更高的概率",
"￥小节Grammatical tasks",
"表￥reftab:title显示了10个语法任务?citepmarvin2018targeted以及",
"用于生成最小对的模板。任务分为两大类：主语动词一致（SVA）和反身回指",
"(RA）。第一个SVA任务SVASimple调查模型是否理解主题号必须与第三人称",
"现在动词的数量一致：￥pex[*=？*]<判断>?a猫走路。?aljudge*猫走路。",
"?xe其他SVA任务探测模型是否具有更复杂的数字一致性表示。例如，SVAPP",
"任务测量模型是否能够忽略出现在主体头部和动词之间的于扰因素（“男孩”）：",
"￥pex[*=？*<judgements>?a男孩旁边的猫跳。?aljudge*男孩旁边的猫",
"跳。xe对象相对子句任务探测在存在嵌人子句的情况下模型是否准确地保持",
"头部的编号。￥citetmarvin2018targeted对这些任务的语言动机进行了广泛的",
"讨论",
"RA任务测量语言模型是否理解反身代词绑定的结构条件。这些任务利用",
"了英语反身语的以下属性：反身代词需要在数量上与其前置词一致。RA发",
"送。Comp任务评估模型是否理解自反性必须与它们的前身在同一子句中：",
"￥pex[*=？*<判决>?a律师说，被告自证其罪。?a?ljudge*律师说，被告",
"自证其罪。xe涉及宾语相对从句的RA任务评估模型是否理解自反性回指不",
"绑定到嵌入从句中的名词，而是绑定到头部名词",
"￥开始table*?居中?调整大小框0.8Ypaperwidth！",
"begin",
"tabularJ1llhline?bfTaskbfTemplate??hline?hlineSVASimple",
"The Ylangle TargetNoun rangle Ylangle Verb rangle. ￥? SVA Subj.Rel",
"Clause TheYlangleTargetNounranglethat liked theYlangleNounrangle",
"Ylangle Verb Yrangle:￥? SVA Sent.Comp.The langle Noun rangle",
"said the langle TargetNoun Yrangle Ylangle Verb Yrangle:? SVA PP",
"The Ylangle TargetNoun Yrangle next to the Ylangle Noun Yrangle Ylangle",
"Verb rangle:? SVA Obj.Rel. Clause.That The ?langle TargetNoun",
"Yrangle that the ?langle Noun Yrangle liked langle Verb rangle:￥?",
"SVA Obj.Rel.Clause.NoThat The langle TargetNoun Yrangle the Ylangle",
"2"
] | {
"arxiv_id": "2010.13870",
"language": "zh",
"page_idx": 1
} |
||
2010.13870-zh-2 | 2010.13870-zh-2.jpg | Noun ¥rangle liked ¥langle Verb ¥$rangle.$ ¥¥ RA Simple The ¥langle TargetNoun ¥rangle ¥langle PastTransVerb ¥rangle ¥langle himself/themselves ¥ rangle . ¥ ¥ RA Sent. Comp. The ¥ langle NonGenderedNoun ¥ rangle said the $\Psi langle$ TargetNoun ¥rangle ¥langle PastTransVerb ¥rangle ¥langle himself/ themselves ¥ $rangle.$ ¥¥ RA Obj.Rel.Clause.That The ¥langle TargetNoun ¥rangle that the Ylangle NonGenderedNoun ¥rangle liked ¥langle PastTransVerb ¥rangle ¥langle himself/themselves ¥$rangle.$ ¥¥ RA Obj.Rel.Clause.NoThat
The Ylangle TargetNoun ¥rangle the Ylangle NonGenderedNoun Yrangle liked ¥langle PastTransVerb ¥rangle Ylangle himself/themselves ¥rangle . YYYYYhline Yend tabular ¥ 标 题 table Templates used for sentence genera- tion. TargetNoun indicates the position of the target noun whose performance score is being calculated. ¥ 标签 tab:title ¥ 结束 table*
${\text{¥ 小节 Measuring the performance of a noun}}$
我们使用这些任务来衡量模型对特定 ¥emph target noun 的语法属性的理解程度。给定一个特定的目标名词，在表 ¥ref tab:title 所示的每个任务模板中将其替换为 TargetNoun。这给出了一个部分指定的模板。例如，在 SVA 简单模板中替换目标名词“zombie”会导致：¥pex[*=?“]<judgements> 僵尸${\mathrm{¥langle~}}$动词 ¥rangle 。¥xe 给定这些部分指定的模板中的每一个，通过填充剩余的词汇项来随机抽样 500 个最小对。最后，计算模型对 500 个最小对的语法判断 (通过计算语法和非语法变体之间的得分差异)并取平均值，从而得出名词的任务绩效得分
$\Psi$小节 Limitations
这些分析在几个方面受到限制。首先，只使用两个语法任务。通过使用更广
泛的任务，可以研究数字协议之外的更大范围的语法现象
其次，虽然研究的重点是名词所携带的语法信息，但动词等其他词汇类型也可能携带此信息。未来的工作可以确定该方法是否推广到动词和其他词汇类型
最后，虽然研究使用可接受性判断来确定模型的语法知识，但存在其他探测任务，并且可能会产生不同的结果 ¥citep warstadt2019investigating 。我们使用可接受性判断，因为据我们所知，尚未针对 GPT-2 或 Transformer-XL 对特征探测进行广泛研究。不同的探测架构可能会对这些模型产生不同的结果。希望了解当前结果对实验读数选择的稳健性
Ysection Methods 在本节中，我们将更详细地描述计算目标名词的任务绩
效得分的过程
¥ 小节 Sentence generation ¥ 标签 sec:sentence-generation
使用 WordNet¥citep fellbaum1998wordnet 和 VerbNet¥citep schuler2005verbnet
,我们编制了一个词法项目列表，如表 ¥ref tab:word-count 所示。目标名词来自名词列表，该列表由动画名词组成。仅包括具有不同单数和复数形式的名词。动词集中的所有动词都有不及物读。对于每对任务模板和目标名词，通过从适当的单词列表中选择词汇项目来随机抽样 500 个句子
对于每个采样句子，生成 2*2 或$2^{*}2^{*}$2 版本(取决于模板)。这些版本改变了句子的语法性以及目标名词和任何干扰名词的复数形式。例如，对于 SVA 简单任务，为每个采样句子生成$2^*2$个版本：¥pex¥label ex:variants ¥a 奇异语法：马走。¥a 奇异语法：¥ljudge * 走马。¥a 多元语法：马走路。¥a 复数非${\text{语法：¥ljudge}}^{*}$马走路。${\mathrm{Yxe}}$
¥ 小节 Models 我们的实验使用三种模型 ,Transformer-XL¥citep dai2019transformer
,GPT-2¥citep radford2019language 和 BERT¥citep devlin2019bert 。我们将
3 | [
[
"0.2173",
"0.1604",
"0.778",
"0.1604",
"0.778",
"0.1736",
"0.2173",
"0.1736"
],
[
"0.2173",
"0.1749",
"0.7812",
"0.1749",
"0.7812",
"0.1881",
"0.2173",
"0.1881"
],
[
"0.2173",
"0.1895",
"0.4631",
"0.1895",
"0.4631",
"0.2026",
"0.2173",
"0.2026"
],
[
"0.4656",
"0.1932",
"0.4722",
"0.1932",
"0.4722",
"0.1989",
"0.4656",
"0.1989"
],
[
"0.4697",
"0.1895",
"0.7812",
"0.1895",
"0.7812",
"0.2026",
"0.4697",
"0.2026"
],
[
"0.2157",
"0.2052",
"0.7793",
"0.2052",
"0.7793",
"0.2184",
"0.2157",
"0.2184"
],
[
"0.2173",
"0.2197",
"0.778",
"0.2197",
"0.778",
"0.2329",
"0.2173",
"0.2329"
],
[
"0.2181",
"0.2349",
"0.7793",
"0.2349",
"0.7793",
"0.2488",
"0.2181",
"0.2488"
],
[
"0.2181",
"0.2507",
"0.886",
"0.2507",
"0.886",
"0.2644",
"0.2181",
"0.2644"
],
[
"0.2181",
"0.2651",
"0.7793",
"0.2651",
"0.7793",
"0.279",
"0.2181",
"0.279"
],
[
"0.2157",
"0.279",
"0.7793",
"0.2798",
"0.7793",
"0.2942",
"0.2157",
"0.2935"
],
[
"0.2173",
"0.2935",
"0.778",
"0.2954",
"0.778",
"0.31",
"0.2173",
"0.308"
],
[
"0.2173",
"0.3113",
"0.7793",
"0.3113",
"0.7793",
"0.3245",
"0.2173",
"0.3245"
],
[
"0.2173",
"0.3257",
"0.6323",
"0.3257",
"0.6323",
"0.339",
"0.2173",
"0.339"
],
[
"0.2402",
"0.339",
"0.562",
"0.3403",
"0.562",
"0.3542",
"0.2402",
"0.3528"
],
[
"0.2418",
"0.356",
"0.7793",
"0.356",
"0.7793",
"0.3694",
"0.2418",
"0.3694"
],
[
"0.2181",
"0.3706",
"0.7793",
"0.3706",
"0.7793",
"0.3838",
"0.2181",
"0.3838"
],
[
"0.2181",
"0.3865",
"0.7793",
"0.3865",
"0.7793",
"0.3997",
"0.2181",
"0.3997"
],
[
"0.2181",
"0.401",
"0.778",
"0.401",
"0.778",
"0.4153",
"0.2181",
"0.4153"
],
[
"0.2181",
"0.4167",
"0.7793",
"0.4167",
"0.7793",
"0.43",
"0.2181",
"0.43"
],
[
"0.2181",
"0.4312",
"0.7793",
"0.4312",
"0.7793",
"0.4443",
"0.2181",
"0.4443"
],
[
"0.2181",
"0.4463",
"0.7793",
"0.4463",
"0.7793",
"0.4602",
"0.2181",
"0.4602"
],
[
"0.2181",
"0.4622",
"0.3652",
"0.4622",
"0.3652",
"0.476",
"0.2181",
"0.476"
],
[
"0.2418",
"0.4766",
"0.38",
"0.4766",
"0.38",
"0.4905",
"0.2418",
"0.4905"
],
[
"0.2418",
"0.4905",
"0.778",
"0.4912",
"0.778",
"0.506",
"0.2418",
"0.505"
],
[
"0.2181",
"0.506",
"0.626",
"0.506",
"0.626",
"0.5195",
"0.2181",
"0.5195"
],
[
"0.2418",
"0.5215",
"0.7793",
"0.5215",
"0.7793",
"0.535",
"0.2418",
"0.535"
],
[
"0.2214",
"0.5386",
"0.7695",
"0.5386",
"0.7695",
"0.549",
"0.2214",
"0.549"
],
[
"0.2427",
"0.552",
"0.7793",
"0.552",
"0.7793",
"0.565",
"0.2427",
"0.565"
],
[
"0.2181",
"0.5674",
"0.7793",
"0.5674",
"0.7793",
"0.582",
"0.2181",
"0.582"
],
[
"0.2197",
"0.582",
"0.778",
"0.582",
"0.778",
"0.595",
"0.2197",
"0.595"
],
[
"0.2181",
"0.5977",
"0.7793",
"0.5977",
"0.7793",
"0.6113",
"0.2181",
"0.6113"
],
[
"0.2197",
"0.6143",
"0.5107",
"0.6143",
"0.5107",
"0.625",
"0.2197",
"0.625"
],
[
"0.2418",
"0.6274",
"0.7793",
"0.6274",
"0.7793",
"0.6416",
"0.2418",
"0.6416"
],
[
"0.2197",
"0.643",
"0.3162",
"0.643",
"0.3162",
"0.655",
"0.2197",
"0.655"
],
[
"0.2418",
"0.6577",
"0.671",
"0.6577",
"0.671",
"0.672",
"0.2418",
"0.672"
],
[
"0.2418",
"0.6724",
"0.8228",
"0.6724",
"0.8228",
"0.686",
"0.2418",
"0.686"
],
[
"0.2303",
"0.686",
"0.7812",
"0.6875",
"0.7812",
"0.702",
"0.2303",
"0.7007"
],
[
"0.2181",
"0.7026",
"0.778",
"0.7026",
"0.778",
"0.7163",
"0.2181",
"0.7163"
],
[
"0.2181",
"0.7183",
"0.7793",
"0.7183",
"0.7793",
"0.732",
"0.2181",
"0.732"
],
[
"0.2197",
"0.733",
"0.6284",
"0.733",
"0.6284",
"0.747",
"0.2197",
"0.747"
],
[
"0.2427",
"0.7476",
"0.7793",
"0.7476",
"0.7793",
"0.761",
"0.2427",
"0.761"
],
[
"0.2214",
"0.763",
"0.7793",
"0.763",
"0.7793",
"0.7764",
"0.2214",
"0.7764"
],
[
"0.2173",
"0.7764",
"0.7812",
"0.778",
"0.7812",
"0.7925",
"0.2173",
"0.7915"
],
[
"0.2181",
"0.7944",
"0.7812",
"0.7944",
"0.7812",
"0.808",
"0.2181",
"0.808"
],
[
"0.2181",
"0.8096",
"0.4478",
"0.8096",
"0.4478",
"0.8228",
"0.2181",
"0.8228"
],
[
"0.2418",
"0.8237",
"0.8447",
"0.8237",
"0.8447",
"0.837",
"0.2418",
"0.837"
],
[
"0.2214",
"0.8384",
"0.7793",
"0.8384",
"0.7793",
"0.853",
"0.2214",
"0.853"
],
[
"0.4902",
"0.8755",
"0.5073",
"0.8755",
"0.5073",
"0.8896",
"0.4902",
"0.8896"
]
] | [
"Noun rangle liked langle Verb rangle.? RA Simple The langle Tar-",
"getNounranglelanglePastTransVerbranglelanglehimself/themselves",
"￥rangle.￥?RASent.Comp.",
"",
"The Ylangle NonGenderedNoun rangle",
"said thelangleTargetNounranglelanglePastTransVerbranglelangle",
"himself/themselves Yrangle. ￥RA Obj.Rel.Clause.That The ￥langle Tar-",
"getNoun Yrangle that the Ylangle NonGenderedNoun rangle liked langle",
"PastTransVerb rangle ?langle himself/themselves rangle.￥? RA Obj.Rel.Clause.NoThat",
"The Ylangle TargetNoun Yrangle the Ylangle NonGenderedNoun rangle",
"liked Ylangle PastTransVerb Yrangle ?langlehimself/themselves Yrangle.",
"￥?hline￥end tabular￥标题tableTemplates used for sentencegenera-",
"tion. TargetNoun indicates the position of the target noun whose performance",
"scoreisbeingcalculated.￥标签tab:title?结束table*",
"￥小节Measuringtheperformanceofanour",
"我们使用这些任务来衡量模型对特定emphtargetnoun的语法属性的",
"理解程度。给定一个特定的目标名词，在表￥reftab:title所示的每个任务模",
"板中将其替换为TargetNoun。这给出了一个部分指定的模板。例如，在SVA",
"简单模板中替换目标名词“zombie”会导致：￥pex[*=？*<judgements>僵尸",
"?langle动词?rangle：xe给定这些部分指定的模板中的每一个，通过填充",
"剩余的词汇项来随机抽样500个最小对。最后，计算模型对500个最小对的语",
"法判断（通过计算语法和非语法变体之间的得分差异）并取平均值，从而得出",
"名词的任务绩效得分",
"￥小节Limitations",
"这些分析在几个方面受到限制。首先，只使用两个语法任务。通过使用更广",
"泛的任务，可以研究数字协议之外的更大范围的语法现象",
"其次，虽然研究的重点是名词所携带的语法信息，但动词等其他词汇类型也",
"可能携带此信息。未来的工作可以确定该方法是否推广到动词和其他词汇类型",
"最后，虽然研究使用可接受性判断来确定模型的语法知识，但存在其他探测",
"任务，并且可能会产生不同的结果?citepwarstadt2019investigating。我们使",
"用可接受性判断，因为据我们所知，尚未针对GPT-2或Transformer-XL对特",
"征探测进行广泛研究。不同的探测架构可能会对这些模型产生不同的结果。希",
"望了解当前结果对实验读数选择的稳健性",
"?sectionMethods在本节中，我们将更详细地描述计算目标名词的任务绩",
"效得分的过程",
"￥小节Sentencegeneration标签sec:sentence-generation",
"使用WordNet?citepfellbaum1998wordnet和VerbNet?citepschuler2005verbnet",
"我们编制了一个词法项目列表，如表￥reftab:word-count所示。目标名词",
"来自名词列表，该列表由动画名词组成。仅包括具有不同单数和复数形式的名",
"词。动词集中的所有动词都有不及物读。对于每对任务模板和目标名词，通过",
"从适当的单词列表中选择词汇项目来随机抽样500个句子",
"对于每个采样句子，生成2*2或2*2*2版本（取决于模板）。这些版本改变",
"了句子的语法性以及目标名词和任何干扰名词的复数形式。例如，对于SVA简",
"单任务，为每个采样句子生成2*2个版本：￥pex￥labelex:variants?a奇异语",
"法：马走。?a奇异语法：￥ljudge*走马。￥a多元语法：马走路。a复数非",
"语法：￥ljudge*马走路。￥xe",
"￥小节Models我们的实验使用三种模型，Transformer-XL￥citepdai2019transformer",
"，GPT-2Ycitepradford2019language和BERTcitepdevlin2019bert我们将",
"3"
] | {
"arxiv_id": "2010.13870",
"language": "zh",
"page_idx": 2
} |
||
2010.13870-zh-3 | 2010.13870-zh-3.jpg | 拥抱脸实现 ¥citep Wolf2019HuggingFacesTS 与预训练模型 ¥textit transfoxl-wt103 一起使用，该模型在 WikiText-103 数据集上训练，¥textit gpt2-xl 在 WebText 数据集上训练，以及 ¥textit bert-base-uncased 在 BookCorpus 和英语维基百科上训练
tabular ¥标题 table Size of word sets for each model. ¥ 标签 tab:word-count $\Psi$结束 table
¥ 小节 Sentence scoring 我们现在描述如何计算特定采样句子的分数。对于每个句子变体 (e。g。Example¥ref ex:variants ), 模型计算一个分数。在Transformer-XL 和 GPT-2 的情况下，此分数只是字符串的对数概率。例如 ,对于Transformer-XL: ¥begin equation ¥text Score _ ¥text string (s)=¥logP_ Ytext TXL (s)¥end equation 其中$P_{\mathrm{Y}textTXL}$是 Transformer-XL 语言模型概率分布
对于 BERT, 鉴于其屏蔽的语言模型体系结构，我们遵循 ¥citet goldberg2019assessing 的方法。对于 SVA 任务，我们计算其数字必须与目标名词一致的动词的对数条件概率。对于 RA 任务，我们计算反身代词的对数条件概率。两个条件概率都是在左侧和右侧上下文的条件下计算的
给定句子变体的分数，我们计算句子的总体分数，该分数捕获模型相对于非语法变体更喜欢语法变体的程度。对于每个采样句子$S$ ,其变体中有 2 或 4 个最小对。在示例 ¥ref ex:variants 中，a。andb。是最小对，而 c。andd。是一个最小对。令$s_a,...,s_d$表示这些变体，则该句子的总得分为：¥begin align* Ytext Score - ¥text sent (S) =¥frac 1 2 $( \&$ ¥text Score\_ ¥text string (s\_a)-YY$s_a,...,s_d$
¥ 小节 Noun scoring ¥ 标签 sec:noun-scoring 接下来，我们计算目标名词的总分。如第 ¥ref sec:sentence-generation 节所述，对于特定的目标名词$n$和任务，我们对 500 个句子$S_1,...,S_{500}$进行采样。此任务的名词分数由下式给出： Ybegin equation Ytext Score - Ytext noun ( n) = ¥ frac 1 500 Ysum$\_$ $\mathrm{i=1}^{-500}\bar{\mathrm{¥}textScore}_{-}$¥$textsent(S$
${\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}{\mathrm{~}}$
如果单词的单数或复数形式被标记为 ¥emph unk ,或者如果单词的单数和复数形式被分配了不同数量的标记，则单词将从特定模型中删除。¥ 脚注 The latter constraint was used in order to simplify batching. 对于 BERT,如果为动词集中的单词分配了多个令牌，则将其删除，因为 BERT 不会对多个掩码令牌上的联合分布进行建模
对于 Transformer-XL, 我们在句子的开头添加一个填充文本 ¥ 脚注 ¥url https://tinyurl.com/y9kjuj5q 和一个句子开始标记(¥langle SOS ¥rangle ), 并在句子的末尾添加一个句子结束标记(¥langle EOS ¥rangle )。对于 GPT-2, 我们不对生成的句子进行任何修改(尽管出于标记化目的将前缀空格添加到字符串中)。对于 BERT, 由于它是一个掩码语言模型，因此我们在标记化后用[MASK]标记替换动词(用于 SVA)或反身代词(用于 RA)。因此，每个句子将有一个对应于应与目标名词一致的单词的掩码标记。¥ 第 Results 节 ¥ 标签sec:results
4 | [
[
"0.2173",
"0.1604",
"0.777",
"0.1604",
"0.777",
"0.1736",
"0.2173",
"0.1736"
],
[
"0.2157",
"0.1724",
"0.7793",
"0.1736",
"0.7793",
"0.1881",
"0.2157",
"0.1875"
],
[
"0.2157",
"0.1881",
"0.7812",
"0.1895",
"0.7812",
"0.2039",
"0.2157",
"0.2026"
],
[
"0.2181",
"0.2052",
"0.3823",
"0.2052",
"0.3823",
"0.2184",
"0.2181",
"0.2184"
],
[
"0.2385",
"0.2172",
"0.778",
"0.2197",
"0.778",
"0.2343",
"0.2385",
"0.2317"
],
[
"0.2173",
"0.2349",
"0.7793",
"0.2349",
"0.7793",
"0.2488",
"0.2173",
"0.2488"
],
[
"0.2197",
"0.252",
"0.777",
"0.252",
"0.777",
"0.262",
"0.2197",
"0.262"
],
[
"0.2173",
"0.2651",
"0.7812",
"0.2651",
"0.7812",
"0.279",
"0.2173",
"0.279"
],
[
"0.2157",
"0.279",
"0.7812",
"0.2798",
"0.7812",
"0.2942",
"0.2157",
"0.2935"
],
[
"0.2181",
"0.2966",
"0.3096",
"0.2966",
"0.3096",
"0.308",
"0.2181",
"0.308"
],
[
"0.2385",
"0.3093",
"0.7812",
"0.31",
"0.7812",
"0.3245",
"0.2385",
"0.324"
],
[
"0.2181",
"0.3257",
"0.7793",
"0.3257",
"0.7793",
"0.339",
"0.2181",
"0.339"
],
[
"0.2157",
"0.339",
"0.7793",
"0.3403",
"0.7793",
"0.3542",
"0.2157",
"0.3528"
],
[
"0.2157",
"0.3542",
"0.767",
"0.3542",
"0.767",
"0.3687",
"0.2157",
"0.3687"
],
[
"0.2157",
"0.3694",
"0.7817",
"0.3694",
"0.7817",
"0.387",
"0.2157",
"0.387"
],
[
"0.2181",
"0.387",
"0.2834",
"0.387",
"0.2834",
"0.3984",
"0.2181",
"0.3984"
],
[
"0.2418",
"0.3997",
"0.7793",
"0.401",
"0.7793",
"0.4153",
"0.2418",
"0.414"
],
[
"0.2173",
"0.4167",
"0.7793",
"0.4167",
"0.7793",
"0.43",
"0.2173",
"0.43"
],
[
"0.2336",
"0.4312",
"0.7817",
"0.4312",
"0.7817",
"0.4443",
"0.2336",
"0.4443"
],
[
"0.2181",
"0.4463",
"0.642",
"0.4463",
"0.642",
"0.4602",
"0.2181",
"0.4602"
],
[
"0.2427",
"0.4622",
"0.7793",
"0.4622",
"0.7793",
"0.476",
"0.2427",
"0.476"
],
[
"0.2181",
"0.4766",
"0.7793",
"0.4766",
"0.7793",
"0.4905",
"0.2181",
"0.4905"
],
[
"0.2181",
"0.4924",
"0.7793",
"0.4924",
"0.7793",
"0.506",
"0.2181",
"0.506"
],
[
"0.2279",
"0.506",
"0.778",
"0.506",
"0.778",
"0.521",
"0.2279",
"0.521"
],
[
"0.2173",
"0.5215",
"0.7812",
"0.5215",
"0.7812",
"0.535",
"0.2173",
"0.535"
],
[
"0.2157",
"0.535",
"0.429",
"0.535",
"0.429",
"0.55",
"0.2157",
"0.55"
],
[
"0.2418",
"0.552",
"0.7793",
"0.552",
"0.7793",
"0.565",
"0.2418",
"0.565"
],
[
"0.2181",
"0.5674",
"0.7793",
"0.5674",
"0.7793",
"0.5806",
"0.2181",
"0.5806"
],
[
"0.2181",
"0.582",
"0.7817",
"0.582",
"0.7817",
"0.595",
"0.2181",
"0.595"
],
[
"0.2173",
"0.5977",
"0.774",
"0.5977",
"0.774",
"0.6113",
"0.2173",
"0.6113"
],
[
"0.2157",
"0.61",
"0.4592",
"0.6113",
"0.4592",
"0.6255",
"0.2157",
"0.625"
],
[
"0.2418",
"0.6274",
"0.5215",
"0.6274",
"0.5215",
"0.6416",
"0.2418",
"0.6416"
],
[
"0.2402",
"0.6426",
"0.778",
"0.6426",
"0.778",
"0.656",
"0.2402",
"0.656"
],
[
"0.2181",
"0.6577",
"0.7793",
"0.6577",
"0.7793",
"0.672",
"0.2181",
"0.672"
],
[
"0.2173",
"0.6724",
"0.7793",
"0.6724",
"0.7793",
"0.686",
"0.2173",
"0.686"
],
[
"0.2173",
"0.686",
"0.7793",
"0.6875",
"0.7793",
"0.702",
"0.2173",
"0.7007"
],
[
"0.2173",
"0.7026",
"0.3987",
"0.7026",
"0.3987",
"0.7163",
"0.2173",
"0.7163"
],
[
"0.2427",
"0.7183",
"0.7812",
"0.7183",
"0.7812",
"0.732",
"0.2427",
"0.732"
],
[
"0.2173",
"0.733",
"0.7812",
"0.733",
"0.7812",
"0.7476",
"0.2173",
"0.7476"
],
[
"0.2173",
"0.747",
"0.778",
"0.747",
"0.778",
"0.761",
"0.2173",
"0.761"
],
[
"0.2181",
"0.763",
"0.7793",
"0.763",
"0.7793",
"0.7764",
"0.2181",
"0.7764"
],
[
"0.2181",
"0.779",
"0.778",
"0.779",
"0.778",
"0.7925",
"0.2181",
"0.7925"
],
[
"0.2181",
"0.7935",
"0.7793",
"0.7935",
"0.7793",
"0.8066",
"0.2181",
"0.8066"
],
[
"0.2181",
"0.8096",
"0.7793",
"0.8096",
"0.7793",
"0.8228",
"0.2181",
"0.8228"
],
[
"0.2181",
"0.8257",
"0.29",
"0.8257",
"0.29",
"0.837",
"0.2181",
"0.837"
],
[
"0.4902",
"0.8755",
"0.5073",
"0.8755",
"0.5073",
"0.8896",
"0.4902",
"0.8896"
]
] | [
"拥抱脸实现?citepWolf2019HuggingFacesTS与预训练模型?textittransfo",
"xl-wt103一起使用，该模型在WikiText-103数据集上训练，￥textitgpt2-xl",
"在WebText数据集上训练，以及?textit bert-base-uncased在BookCorpus",
"和英语维基百科上训练",
"?开始table?居中?调整大小框?columnwidth！?begintab",
"ularJcllcccl￥hline ￥bf SetNameTransformer-XLGPT-2BERT￥",
"￥hline￥hlineNoun916723704￥￥Verb615228406￥￥NonGendered",
"Noun870679663￥￥PastTransVerb129810341298￥?￥hline￥end",
"tabular ￥标题tableSizeofword setsforeachmodel.¥标签tab:word-count",
"?结束table",
"?小节Sentence scoring我们现在描述如何计算特定采样句子的分数。对",
"于每个句子变体（e。g。Example?refex:variants），模型计算一个分数。在",
"Transformer-XL和GPT-2的情况下，此分数只是字符串的对数概率。例如，对于",
"Transformer-XL:Ybegin equation Ytext Score -￥text string (s)-￥logP",
"￥textTXL(s)?end equation其中PyteatTxL是Transformer-XL语言模型",
"概率分布",
"对于BERT，鉴于其屏蔽的语言模型体系结构，我们遵循citetgold-",
"berg2019assessing的方法。对于SVA任务，我们计算其数字必须与目标名词",
"致的动词的对数条件概率。对于RA任务，我们计算反身代词的对数条件概",
"率。两个条件概率都是在左侧和右侧上下文的条件下计算的",
"给定句子变体的分数，我们计算句子的总体分数，该分数捕获模型相对于非",
"语法变体更喜欢语法变体的程度。对于每个采样句子S，其变体中有2或4个",
"最小对。在示例￥refex:variants中，a。andb。是最小对，而c。andd。是",
"一个最小对。令SaSd表示这些变体，则该句子的总得分为：?beginalign",
"￥text Score- ?text sent(S)-?frac 1 2（(& ￥text Score- ￥text",
"string(s_a)-?? sa...,Sd",
"￥小节Nounscoring?标签sec:noun-scoring接下来，我们计算目标名词",
"的总分。如第?refsec:sentence-generation节所述，对于特定的目标名词n和",
"任务，我们对500个句子S1，..，S500进行采样。此任务的名词分数由下式给出：",
"￥beginequation?text Score-?textnoun(n)=?frac 1500?sum",
"i=1500?tertScore_￥tertsent(S",
"￥小节Word filtering and tokenization",
"如果单词的单数或复数形式被标记为emphunk，或者如果单词的单数和",
"复数形式被分配了不同数量的标记，则单词将从特定模型中删除。?脚注The",
"latterconstraintwasusedinordertosimplifybatching.对于BERT，如果为",
"动词集中的单词分配了多个令牌，则将其删除，因为BERT不会对多个掩码令",
"牌上的联合分布进行建模",
"对于Transformer-XL，我们在句子的开头添加一个填充文本?脚注url",
"https://tinyurl.com/y9kjuj5q和一个句子开始标记（?langleSOsrangle），",
"并在句子的末尾添加一个句子结束标记（￥langleEOS?rangle）。对于GPT-2",
"我们不对生成的句子进行任何修改（尽管出于标记化目的将前缀空格添加到字",
"符串中）。对于BERT，由于它是一个掩码语言模型，因此我们在标记化后用",
"[MASK|标记替换动词（用于SVA）或反身代词（用于RA）。因此，每个句子",
"将有一个对应于应与目标名词一致的单词的掩码标记。￥第Results节?标签",
"sec:results",
"4"
] | {
"arxiv_id": "2010.13870",
"language": "zh",
"page_idx": 3
} |
||
2010.13870-zh-4 | 2010.13870-zh-4.jpg | ¥ 小节 Noun performance is correlated across tasks ¥ 标签 sec:task-
correlation
我们首先检查每个名词的表现如何在语法任务中变化。对于每个名词-任务对，我们测量该名词在该任务上的平均性能，如上所述。每个名词有 10 个特征， 对应于 10 个语法任务
图 ¥ref fg:txl\_full\_pairplot 显示了 Transformer-XL 在不同任务上的性能之间的成对比较。BERT 和 GPT-2 的结果相似，见附录。该图显示了各个任务之间的性能相关；对于许多任务对，在一个任务上具有较高性能的名词可能在另一个任务上具有较高的性能
使用主成分分析，我们发现单个主成分解释了47¥
这些结果提供了证据，表明语言模型在语法任务上的性能差异在一定程度上是由跨任务稳定的名词属性所解释的。对于某些名词，模型可以更好地理解数字协议，而对于其他名词则更糟糕
Ybegin fgure Yincludegraphics[ width= ¥ columnwidth, keepaspectratio] Fig- ures/txl\_full\_pairplot.png ¥caption Pairwise comparisons between tasks with Transformer-XL. Rows and columns represent tasks, and one point represents a single noun's performance on a pair of tasks. The four tasks on the lowen right, with strongest correlations, all involve reflexive anaphora. ¥label fig: txl$\_$ful$\_$pairplot Yend figure
¥ 小节 Noun performance is correlated across models ¥ 标签 sec:model-
$\operatorname{correlation}$
接下来，我们研究名词是否在不同的神经语言模型中表现出稳定的行为。对于这三种语言模型中的每对，我们测量了一个名词在一种语言模型中的任务表现如何预测其在另一种语言模型中的任务表现
图 ¥ref fg:model-correlations 显示了 10 个语法任务上的语言模型对之间的比较。在 30 个比较中，有 24 个显示出语言模型对之间的显着正相关。${\mathrm{Bonferroni~}}$校正后，22 个相关性仍然显著
GPT-2 和 Transformer-XL 在性能上显示出最强的相关性。这可能是由于与BERT 相比，GPT-2 和 Transformer-XL 的任务设置之间的方法差异：GPT-2 和Transformer-XL 正在执行语言建模任务，其中查询完整句子的概率，而 BERT 对单个目标词执行掩码语言建模。差异也可能是由于 BERT 和自回归语言模型之间相应的训练差异
结果提供了证据，证明名词在语言模型中表现出稳定的任务性能。跨语言模型的相关性的来源必须来自训练数据的特征。名词的自然文本分布的性质导致其中一些名词比其他名词更好地理解
¥ 开始 figure* [t]¥ 居中 ¥ 包含 [width=0。8¥ 纸张宽度，keepaspectratio] Figures/pairwise\_model\_correlation.png ¥ 标题 Pairwise comparisons between GPT-2, Transformer-XL, and BERT on the 10 grammatical tasks. Each row corresponds to a pair of language models, and each column is a single task. One point represents the performance of a noun on a single task. ¥ 标签fig:model-correlations ¥ 结束 figure*
${\text{¥ 小节 Effect of frequency on task performance}}$
$¥$开始 figure* ¥居中¥包含[width=0。8¥ 纸张宽度] Figures/main\_body\_freq\_plot\_mixed\_sva\_ra.png
$¥$标题 Relationship between corpus frequency and task performance for Transformer-
XL, BERT, and GPT-2. Performance scores are z-normalized. Colors indicate the ten grammatical tasks and singular/plural form of the noun (¥emph s
5 | [
[
"0.2418",
"0.1604",
"0.778",
"0.1604",
"0.778",
"0.1736",
"0.2418",
"0.1736"
],
[
"0.2181",
"0.1768",
"0.2942",
"0.1768",
"0.2942",
"0.1875",
"0.2181",
"0.1875"
],
[
"0.2418",
"0.1901",
"0.778",
"0.1901",
"0.778",
"0.2039",
"0.2418",
"0.2039"
],
[
"0.2173",
"0.2052",
"0.7817",
"0.2052",
"0.7817",
"0.2184",
"0.2173",
"0.2184"
],
[
"0.2173",
"0.2197",
"0.3757",
"0.2197",
"0.3757",
"0.2329",
"0.2173",
"0.2329"
],
[
"0.2427",
"0.2349",
"0.7793",
"0.2349",
"0.7793",
"0.2488",
"0.2427",
"0.2488"
],
[
"0.2181",
"0.2507",
"0.778",
"0.2507",
"0.778",
"0.2644",
"0.2181",
"0.2644"
],
[
"0.2173",
"0.2644",
"0.7793",
"0.2651",
"0.7793",
"0.279",
"0.2173",
"0.2778"
],
[
"0.2181",
"0.281",
"0.4307",
"0.281",
"0.4307",
"0.2942",
"0.2181",
"0.2942"
],
[
"0.2427",
"0.2954",
"0.603",
"0.2954",
"0.603",
"0.3093",
"0.2427",
"0.3093"
],
[
"0.2443",
"0.3125",
"0.778",
"0.3125",
"0.778",
"0.3225",
"0.2443",
"0.3225"
],
[
"0.2173",
"0.3245",
"0.7793",
"0.3257",
"0.7793",
"0.3403",
"0.2173",
"0.339"
],
[
"0.2197",
"0.3423",
"0.4617",
"0.3423",
"0.4617",
"0.3528",
"0.2197",
"0.3528"
],
[
"0.2402",
"0.3547",
"0.7837",
"0.356",
"0.7837",
"0.3706",
"0.2402",
"0.3694"
],
[
"0.2181",
"0.3706",
"0.7812",
"0.3706",
"0.7812",
"0.385",
"0.2181",
"0.385"
],
[
"0.2173",
"0.3865",
"0.7812",
"0.3865",
"0.7812",
"0.3997",
"0.2173",
"0.3997"
],
[
"0.2173",
"0.4016",
"0.778",
"0.4016",
"0.778",
"0.4153",
"0.2173",
"0.4153"
],
[
"0.2173",
"0.4167",
"0.7812",
"0.4153",
"0.7812",
"0.43",
"0.2173",
"0.4312"
],
[
"0.2157",
"0.43",
"0.4592",
"0.4312",
"0.4592",
"0.4458",
"0.2157",
"0.4443"
],
[
"0.2418",
"0.4463",
"0.778",
"0.4463",
"0.778",
"0.461",
"0.2418",
"0.461"
],
[
"0.2181",
"0.4634",
"0.2957",
"0.4634",
"0.2957",
"0.4749",
"0.2181",
"0.4749"
],
[
"0.2418",
"0.4766",
"0.7793",
"0.4766",
"0.7793",
"0.4905",
"0.2418",
"0.4905"
],
[
"0.2173",
"0.4924",
"0.7793",
"0.4924",
"0.7793",
"0.506",
"0.2173",
"0.506"
],
[
"0.2181",
"0.5083",
"0.5435",
"0.5083",
"0.5435",
"0.518",
"0.2181",
"0.518"
],
[
"0.2427",
"0.5215",
"0.778",
"0.5215",
"0.778",
"0.535",
"0.2427",
"0.535"
],
[
"0.2181",
"0.537",
"0.7793",
"0.537",
"0.7793",
"0.5503",
"0.2181",
"0.5503"
],
[
"0.2181",
"0.552",
"0.5156",
"0.552",
"0.5156",
"0.565",
"0.2181",
"0.565"
],
[
"0.2427",
"0.5664",
"0.778",
"0.5664",
"0.778",
"0.58",
"0.2427",
"0.58"
],
[
"0.2181",
"0.582",
"0.7793",
"0.582",
"0.7793",
"0.595",
"0.2181",
"0.595"
],
[
"0.2157",
"0.595",
"0.7793",
"0.5967",
"0.7793",
"0.6113",
"0.2157",
"0.61"
],
[
"0.2173",
"0.6113",
"0.7793",
"0.6123",
"0.7793",
"0.627",
"0.2173",
"0.6255"
],
[
"0.2173",
"0.6274",
"0.3652",
"0.6274",
"0.3652",
"0.6416",
"0.2173",
"0.6416"
],
[
"0.2418",
"0.643",
"0.7793",
"0.643",
"0.7793",
"0.657",
"0.2418",
"0.657"
],
[
"0.2181",
"0.6577",
"0.7793",
"0.6577",
"0.7793",
"0.672",
"0.2181",
"0.672"
],
[
"0.2173",
"0.6724",
"0.4805",
"0.6724",
"0.4805",
"0.686",
"0.2173",
"0.686"
],
[
"0.2385",
"0.685",
"0.7793",
"0.686",
"0.7793",
"0.704",
"0.2385",
"0.7026"
],
[
"0.2173",
"0.7026",
"0.777",
"0.7026",
"0.777",
"0.7163",
"0.2173",
"0.7163"
],
[
"0.2181",
"0.7183",
"0.7793",
"0.7183",
"0.7793",
"0.732",
"0.2181",
"0.732"
],
[
"0.2181",
"0.7344",
"0.778",
"0.7344",
"0.778",
"0.7476",
"0.2181",
"0.7476"
],
[
"0.2181",
"0.7485",
"0.7793",
"0.7485",
"0.7793",
"0.762",
"0.2181",
"0.762"
],
[
"0.2157",
"0.763",
"0.4878",
"0.762",
"0.4878",
"0.7764",
"0.2157",
"0.778"
],
[
"0.2418",
"0.779",
"0.5825",
"0.779",
"0.5825",
"0.7925",
"0.2418",
"0.7925"
],
[
"0.2402",
"0.7925",
"0.9966",
"0.7944",
"0.9966",
"0.8096",
"0.2402",
"0.8066"
],
[
"0.2181",
"0.8096",
"0.8145",
"0.8096",
"0.8145",
"0.8228",
"0.2181",
"0.8228"
],
[
"0.2157",
"0.8228",
"0.7812",
"0.8237",
"0.7812",
"0.8384",
"0.2157",
"0.837"
],
[
"0.2157",
"0.837",
"0.7817",
"0.8384",
"0.7817",
"0.853",
"0.2157",
"0.8516"
],
[
"0.4902",
"0.8755",
"0.5073",
"0.8755",
"0.5073",
"0.89",
"0.4902",
"0.89"
]
] | [
"￥小节Nounperformance is correlated acrosstasks￥标签sec:task",
"correlation",
"我们首先检查每个名词的表现如何在语法任务中变化。对于每个名词-任务",
"对，我们测量该名词在该任务上的平均性能，如上所述。每个名词有10个特征，",
"对应于10个语法任务",
"图?reffig:txl_full_pairplot显示了Transformer-XL在不同任务上的性能",
"之间的成对比较。BERT和GPT-2的结果相似，见附录。该图显示了各个任务",
"之间的性能相关；对于许多任务对，在一个任务上具有较高性能的名词可能在",
"另一个任务上具有较高的性能",
"使用主成分分析，我们发现单个主成分解释了47￥",
"这些结果提供了证据，表明语言模型在语法任务上的性能差异在一定程度上",
"是由跨任务稳定的名词属性所解释的。对于某些名词，模型可以更好地理解数",
"字协议，而对于其他名词则更糟糕",
"begin figure Yincludegraphics[width-columnwidth,keepaspectratio] Fig",
"ures/txl_full_pairplot.png caption Pairwise comparisons between tasks with",
"Transformer-XL. Rows and columns represent tasks, and one point represents",
"a single noun's performance on a pair of tasks. The four tasks on the lower",
"right, with strongest correlations, all involve reflexive anaphora.Ylabel",
"fig:txl_full_pairplot Yend figure",
"￥小节Noun performance is correlated across models￥标签 sec:model",
"correlation",
"接下来，我们研究名词是否在不同的神经语言模型中表现出稳定的行为。对",
"于这三种语言模型中的每对，我们测量了一个名词在一种语言模型中的任务表",
"现如何预测其在另一种语言模型中的任务表现",
"图reffig:model-correlations显示了10个语法任务上的语言模型对之",
"间的比较。在30个比较中，有24个显示出语言模型对之间的显着正相关。",
"Bonferroni校正后，22个相关性仍然显著",
"GPT-2和Transformer-XL在性能上显示出最强的相关性。这可能是由于与",
"BERT相比，GPT-2和Transformer-XL的任务设置之间的方法差异：GPT-2和",
"Transformer-XL正在执行语言建模任务，其中查询完整句子的概率，而BERT",
"对单个目标词执行掩码语言建模。差异也可能是由于BERT和自回归语言模型",
"之间相应的训练差异",
"结果提供了证据，证明名词在语言模型中表现出稳定的任务性能。跨语言模",
"型的相关性的来源必须来自训练数据的特征。名词的自然文本分布的性质导致",
"其中一些名词比其他名词更好地理解",
"￥开始figure*[t]?居中?包含[width=0。8?纸张宽度，keepaspectra-",
"tio]Figures/pairwise_model_correlation.png?标题Pairwisecomparisonsbe",
"tween GPT-2, Transformer-XL, and BERT on the 10 grammatical tasks. Each",
"row corresponds to a pair of language models, and each column is a single task.",
"One pointrepresents theperformance of anoun on a singletask.?标签",
"fig:model-correlations?结束figure*",
"￥小节Effectoffrequencyontaskperformance",
"?开始figure*?居中?包含[width=0。8?纸张宽度】Figures/main_body_freq_plot_mixed_sva_ra.png",
"?标题RelationshipbetweencorpusfrequencyandtaskperformanceforTransformer-",
"XL, BERT, and GPT-2. Performance scores are z-normalized. Colors indicate",
"the ten grammatical tasks and singular/plural form of the noun (emph s",
"5"
] | {
"arxiv_id": "2010.13870",
"language": "zh",
"page_idx": 4
} |
||
2306.05241-zh-1 | 2306.05241-zh-1.jpg | $Two\textit{Heads Are Better Than One: }$
通过与邻居关联来改进假新闻视频检测
Peng Qi$^1,2,\textbf{Yuyang Zhao}^3,\textbf{Yufeng Shen}^2,\textbf{Wei Ji}^3,\textbf{Juan Cao}^{1,2*}\textbf{and Tat-Seng Chua}^3$
$^{1}$ Key Laboratory of Intelligent Information Processing,
Institute of Computing Technology, Chinese Academy of Sciences
$^{2}$ University of Chinese Academy of Sciences
$^3$ National University of Singapore
$\{$ qipeng,caojuan $\}$ @ict.ac.cn,[email protected],
[email protected], { jiwei,dcscts } @nus.edu.sg
## $\mathbf{Abstract}$
Event: Man cries in pain as his clothing store is flooded.
<FigureHere>
Figure 1: A set of videos belonging to the same event. Fake news videos contain conflicting information with the real ones, and the debunking videos can refute the mismatched information in the fake news videos.
-Zheng Wei (Tang Dynasty)
The prevalence of short video platforms has spawned a lot of fake news videos, which have stronger propagation ability than textual fake news. Thus, automatically detecting fake news videos has been an important countermeasure in practice. Previous works commonly verify each news video individually with multimodal information. Nevertheless, news videos from different perspectives regarding the same event are commonly posted together, which contain complementary or contradictory information and thus can be used to evaluate each other mutually. To this end, we introduce a new and practical paradigm, $i.e.$, crosssample fake news video detection, and propose a novel framework, N eighbor- E nhanced fak E news video D etection ( NEED), which integrates the neighborhood relationship of new videos belonging to the same event. NEED can be readily combined with existing singlesample detectors and further enhance their performances with the proposed graph aggregation (GA) and debunking rectification (DR) modules. Specifically, given the feature representations obtained from single-sample detectors, GA aggregates the neighborhood information with the dynamic graph to enrich the features of independent samples. After that, DR explicitly leverages the relationship between debunking videos and fake news videos to refute the candidate videos via textual and visual consistency. Extensive experiments on the public benchmark demonstrate that NEED greatly improves the performance of both single-modal (up to 8.34 % in accuracy) and multimodal (up to 4.97 % in accuracy) base detectors. Codes are available in https://github.com/ICTMCG/NEED.
假新闻的传播已成为一个重要的社会问题对政治 (Fisher et al., 2016) , 金融 (ElBoghdady, 2013),公共卫生 (Naeem and Bhatti, 2020) 等构成现实威胁。最近，短视频平台的盛行催生了很多假新闻视频，与文本假新闻 (Sundar et al., 2021)相比，这些视频更有说服力，更容易传播。中国网信办通报称，东航坠机事件中流传的 7 条核心谣言中，有 5 条源自短视频平台 (Cyberspace Administration of China, 2022) ...。另一项研究的统计数据也揭示了虚假新闻视频的强大传播，该视频报告说，只有 124 条关于 COVID-19 的 TikTok 虚假新闻获得了超过2000 万的点击量和 200 万个赞，评论和分享， 对数百万人造成了负面影响 (Brandy Zadrozny, 2021)。因此，迫切需要开发针对假新闻视频的自动检测技术，以减轻其负面影响
鉴于假新闻视频检测的实用性，以前的作品$\overline{\mathrm{E}}($Hou et al., 2019; Medina Serrano et al., 2020; Choi and Ko, 2021; Shang et al., 2021; Qi et al., 2023) 利用单个新闻视频的异构多模态信息进行佐证。然而，假新闻是故意制造的，以误导消费者 (Shu et al.,2017),因此多式联运
tiktok.com. A popular short-form video sharing plat-
## 1 Introduction
“Listen to both sides and you will be enlightened; heed only one side and you will be benighted."
form.
${}^{*}$Corresponding author. | [
[
"0.339",
"0.0831",
"0.659",
"0.0831",
"0.659",
"0.0974",
"0.339",
"0.0974"
],
[
"0.297",
"0.10156",
"0.702",
"0.10156",
"0.702",
"0.1164",
"0.297",
"0.1164"
],
[
"0.1427",
"0.1259",
"0.858",
"0.1247",
"0.858",
"0.1426",
"0.1427",
"0.1437"
],
[
"0.2747",
"0.1414",
"0.7256",
"0.1437",
"0.7256",
"0.1603",
"0.2747",
"0.158"
],
[
"0.2334",
"0.1609",
"0.771",
"0.1609",
"0.771",
"0.1746",
"0.2334",
"0.1746"
],
[
"0.3123",
"0.1733",
"0.6885",
"0.1758",
"0.6885",
"0.1936",
"0.3123",
"0.1912"
],
[
"0.3586",
"0.1912",
"0.642",
"0.1948",
"0.642",
"0.209",
"0.3586",
"0.206"
],
[
"0.2468",
"0.2114",
"0.7656",
"0.2114",
"0.7656",
"0.2245",
"0.2468",
"0.2245"
],
[
"0.2032",
"0.2257",
"0.8003",
"0.228",
"0.8003",
"0.2423",
"0.2032",
"0.2405"
],
[
"0.2644",
"0.2546",
"0.34",
"0.2546",
"0.34",
"0.2666",
"0.2644",
"0.2666"
],
[
"0.5264",
"0.2542",
"0.852",
"0.253",
"0.852",
"0.2666",
"0.5264",
"0.2673"
],
[
"0.764",
"0.2695",
"0.827",
"0.2695",
"0.827",
"0.2798",
"0.764",
"0.2798"
],
[
"0.1453",
"0.2766",
"0.4575",
"0.2766",
"0.4575",
"0.2903",
"0.1453",
"0.2903"
],
[
"0.5693",
"0.2756",
"0.594",
"0.2756",
"0.594",
"0.284",
"0.5693",
"0.284"
],
[
"0.7256",
"0.2808",
"0.79",
"0.2832",
"0.79",
"0.2944",
"0.7256",
"0.2922"
],
[
"0.1453",
"0.2922",
"0.4592",
"0.2922",
"0.4592",
"0.306",
"0.1453",
"0.306"
],
[
"0.7256",
"0.2915",
"0.781",
"0.2935",
"0.781",
"0.3047",
"0.7256",
"0.3027"
],
[
"0.552",
"0.3018",
"0.602",
"0.3018",
"0.602",
"0.3123",
"0.552",
"0.3123"
],
[
"0.6665",
"0.3005",
"0.6733",
"0.3005",
"0.6733",
"0.3071",
"0.6665",
"0.3071"
],
[
"0.7256",
"0.3018",
"0.791",
"0.3018",
"0.791",
"0.3135",
"0.7256",
"0.3135"
],
[
"0.8037",
"0.3018",
"0.817",
"0.2944",
"0.8237",
"0.3018",
"0.812",
"0.3088"
],
[
"0.727",
"0.3113",
"0.791",
"0.3113",
"0.791",
"0.3218",
"0.727",
"0.3218"
],
[
"0.588",
"0.3154",
"0.5977",
"0.3154",
"0.5977",
"0.3196",
"0.588",
"0.3196"
],
[
"0.6465",
"0.3154",
"0.657",
"0.3154",
"0.657",
"0.3225",
"0.6465",
"0.3225"
],
[
"0.812",
"0.3164",
"0.825",
"0.3196",
"0.8228",
"0.3225",
"0.81",
"0.3206"
],
[
"0.7256",
"0.3176",
"0.7734",
"0.3196",
"0.7715",
"0.3308",
"0.7246",
"0.3289"
],
[
"0.1453",
"0.3071",
"0.4592",
"0.3071",
"0.4592",
"0.3206",
"0.1453",
"0.3206"
],
[
"0.574",
"0.3196",
"0.5903",
"0.3196",
"0.5903",
"0.328",
"0.574",
"0.328"
],
[
"0.669",
"0.3184",
"0.686",
"0.3184",
"0.686",
"0.3267",
"0.669",
"0.3267"
],
[
"0.7246",
"0.3289",
"0.7715",
"0.3289",
"0.7715",
"0.3403",
"0.7246",
"0.3403"
],
[
"0.1453",
"0.3206",
"0.4592",
"0.3206",
"0.4592",
"0.3342",
"0.1453",
"0.3342"
],
[
"0.1453",
"0.3342",
"0.4592",
"0.3342",
"0.4592",
"0.3474",
"0.1453",
"0.3474"
],
[
"0.657",
"0.3333",
"0.6675",
"0.3333",
"0.6675",
"0.3403",
"0.657",
"0.3403"
],
[
"0.1453",
"0.3486",
"0.4592",
"0.3486",
"0.4592",
"0.3623",
"0.1453",
"0.3623"
],
[
"0.5366",
"0.3499",
"0.608",
"0.3499",
"0.608",
"0.3599",
"0.5366",
"0.3599"
],
[
"0.5366",
"0.3594",
"0.6245",
"0.3594",
"0.6245",
"0.3694",
"0.5366",
"0.3694"
],
[
"0.721",
"0.3552",
"0.7827",
"0.3552",
"0.7827",
"0.3623",
"0.721",
"0.3623"
],
[
"0.1453",
"0.3635",
"0.4592",
"0.3635",
"0.4592",
"0.3772",
"0.1453",
"0.3772"
],
[
"0.7197",
"0.3623",
"0.7866",
"0.3635",
"0.7866",
"0.3735",
"0.7197",
"0.3723"
],
[
"0.531",
"0.3682",
"0.629",
"0.3677",
"0.629",
"0.379",
"0.531",
"0.3801"
],
[
"0.1453",
"0.3772",
"0.4592",
"0.3772",
"0.4592",
"0.3901",
"0.1453",
"0.3901"
],
[
"0.7188",
"0.3723",
"0.79",
"0.3735",
"0.79",
"0.3818",
"0.7188",
"0.3813"
],
[
"0.5845",
"0.3843",
"0.6265",
"0.3872",
"0.6245",
"0.3997",
"0.5825",
"0.3967"
],
[
"0.7188",
"0.3813",
"0.791",
"0.3813",
"0.791",
"0.3914",
"0.7188",
"0.3914"
],
[
"0.1453",
"0.3914",
"0.4592",
"0.3914",
"0.4592",
"0.405",
"0.1453",
"0.405"
],
[
"0.721",
"0.3914",
"0.7715",
"0.3914",
"0.7715",
"0.3984",
"0.721",
"0.3984"
],
[
"0.1436",
"0.4062",
"0.4592",
"0.4062",
"0.4592",
"0.4192",
"0.1436",
"0.4192"
],
[
"0.705",
"0.411",
"0.7363",
"0.411",
"0.7363",
"0.4229",
"0.705",
"0.4229"
],
[
"0.147",
"0.4216",
"0.4575",
"0.4216",
"0.4575",
"0.4316",
"0.147",
"0.4316"
],
[
"0.1453",
"0.434",
"0.4592",
"0.434",
"0.4592",
"0.4478",
"0.1453",
"0.4478"
],
[
"0.511",
"0.4353",
"0.879",
"0.4353",
"0.879",
"0.449",
"0.511",
"0.449"
],
[
"0.1436",
"0.4478",
"0.4592",
"0.449",
"0.4592",
"0.462",
"0.1436",
"0.4614"
],
[
"0.513",
"0.4507",
"0.881",
"0.4507",
"0.881",
"0.4614",
"0.513",
"0.4614"
],
[
"0.147",
"0.4644",
"0.4575",
"0.4644",
"0.4575",
"0.4744",
"0.147",
"0.4744"
],
[
"0.511",
"0.4644",
"0.882",
"0.4644",
"0.882",
"0.478",
"0.511",
"0.478"
],
[
"0.1453",
"0.4768",
"0.4592",
"0.4768",
"0.4592",
"0.4905",
"0.1453",
"0.4905"
],
[
"0.513",
"0.4797",
"0.8423",
"0.4797",
"0.8423",
"0.4905",
"0.513",
"0.4905"
],
[
"0.1453",
"0.4905",
"0.4592",
"0.4905",
"0.4592",
"0.5044",
"0.1453",
"0.5044"
],
[
"0.1453",
"0.505",
"0.4592",
"0.505",
"0.4592",
"0.5186",
"0.1453",
"0.5186"
],
[
"0.6323",
"0.5127",
"0.8447",
"0.5127",
"0.8447",
"0.5254",
"0.6323",
"0.5254"
],
[
"0.1436",
"0.5195",
"0.4575",
"0.5186",
"0.4575",
"0.532",
"0.1436",
"0.533"
],
[
"0.1453",
"0.533",
"0.4592",
"0.533",
"0.4592",
"0.5474",
"0.1453",
"0.5474"
],
[
"0.531",
"0.533",
"0.881",
"0.533",
"0.881",
"0.5464",
"0.531",
"0.5464"
],
[
"0.1436",
"0.5474",
"0.461",
"0.5474",
"0.461",
"0.561",
"0.1436",
"0.561"
],
[
"0.513",
"0.55",
"0.881",
"0.549",
"0.881",
"0.5635",
"0.513",
"0.564"
],
[
"0.1436",
"0.561",
"0.4592",
"0.5625",
"0.4592",
"0.576",
"0.1436",
"0.5747"
],
[
"0.513",
"0.5654",
"0.881",
"0.5654",
"0.881",
"0.58",
"0.513",
"0.58"
],
[
"0.1436",
"0.5747",
"0.4592",
"0.576",
"0.4592",
"0.5903",
"0.1436",
"0.589"
],
[
"0.5146",
"0.582",
"0.881",
"0.582",
"0.881",
"0.5957",
"0.5146",
"0.5957"
],
[
"0.1453",
"0.5903",
"0.4575",
"0.5903",
"0.4575",
"0.604",
"0.1453",
"0.604"
],
[
"0.511",
"0.5967",
"0.882",
"0.597",
"0.882",
"0.6123",
"0.511",
"0.611"
],
[
"0.1453",
"0.605",
"0.4592",
"0.605",
"0.4592",
"0.618",
"0.1453",
"0.618"
],
[
"0.511",
"0.6133",
"0.879",
"0.6133",
"0.879",
"0.627",
"0.511",
"0.627"
],
[
"0.147",
"0.6206",
"0.4575",
"0.6206",
"0.4575",
"0.6313",
"0.147",
"0.6313"
],
[
"0.1453",
"0.633",
"0.4592",
"0.633",
"0.4592",
"0.6465",
"0.1453",
"0.6465"
],
[
"0.513",
"0.63",
"0.881",
"0.63",
"0.881",
"0.6445",
"0.513",
"0.6445"
],
[
"0.1453",
"0.6465",
"0.4592",
"0.6465",
"0.4592",
"0.661",
"0.1453",
"0.661"
],
[
"0.513",
"0.6465",
"0.881",
"0.6465",
"0.881",
"0.661",
"0.513",
"0.661"
],
[
"0.1436",
"0.66",
"0.4375",
"0.661",
"0.4375",
"0.675",
"0.1436",
"0.6733"
],
[
"0.4324",
"0.663",
"0.4568",
"0.663",
"0.4568",
"0.6714",
"0.4324",
"0.6714"
],
[
"0.5146",
"0.662",
"0.881",
"0.662",
"0.881",
"0.676",
"0.5146",
"0.676"
],
[
"0.1453",
"0.676",
"0.4592",
"0.676",
"0.4592",
"0.6895",
"0.1453",
"0.6895"
],
[
"0.544",
"0.6787",
"0.879",
"0.6787",
"0.879",
"0.6924",
"0.544",
"0.6924"
],
[
"0.1453",
"0.6895",
"0.4575",
"0.6895",
"0.4575",
"0.703",
"0.1453",
"0.703"
],
[
"0.529",
"0.685",
"0.5415",
"0.684",
"0.5415",
"0.687",
"0.529",
"0.6885"
],
[
"0.513",
"0.6934",
"0.8774",
"0.6934",
"0.8774",
"0.707",
"0.513",
"0.707"
],
[
"0.1453",
"0.7036",
"0.4592",
"0.7036",
"0.4592",
"0.7173",
"0.1453",
"0.7173"
],
[
"0.5146",
"0.7114",
"0.881",
"0.7114",
"0.881",
"0.7246",
"0.5146",
"0.7246"
],
[
"0.1453",
"0.7188",
"0.4592",
"0.7188",
"0.4592",
"0.7324",
"0.1453",
"0.7324"
],
[
"0.1453",
"0.7324",
"0.4575",
"0.7324",
"0.4575",
"0.745",
"0.1453",
"0.745"
],
[
"0.513",
"0.727",
"0.882",
"0.727",
"0.882",
"0.7407",
"0.513",
"0.7407"
],
[
"0.1453",
"0.745",
"0.4568",
"0.745",
"0.4568",
"0.759",
"0.1453",
"0.759"
],
[
"0.513",
"0.742",
"0.879",
"0.742",
"0.879",
"0.756",
"0.513",
"0.756"
],
[
"0.1453",
"0.7627",
"0.4575",
"0.7627",
"0.4575",
"0.7725",
"0.1453",
"0.7725"
],
[
"0.513",
"0.759",
"0.881",
"0.759",
"0.881",
"0.7725",
"0.513",
"0.7725"
],
[
"0.1453",
"0.775",
"0.4592",
"0.775",
"0.4592",
"0.788",
"0.1453",
"0.788"
],
[
"0.5156",
"0.7754",
"0.781",
"0.7754",
"0.781",
"0.789",
"0.5156",
"0.789"
],
[
"0.1436",
"0.788",
"0.4592",
"0.7876",
"0.4592",
"0.802",
"0.1436",
"0.8027"
],
[
"0.531",
"0.7915",
"0.881",
"0.7915",
"0.881",
"0.8047",
"0.531",
"0.8047"
],
[
"0.1453",
"0.8027",
"0.392",
"0.8027",
"0.392",
"0.8164",
"0.1453",
"0.8164"
],
[
"0.5146",
"0.807",
"0.882",
"0.807",
"0.882",
"0.821",
"0.5146",
"0.821"
],
[
"0.511",
"0.8223",
"0.881",
"0.8223",
"0.881",
"0.836",
"0.511",
"0.836"
],
[
"0.1184",
"0.83",
"0.1318",
"0.83",
"0.1318",
"0.8384",
"0.1184",
"0.8384"
],
[
"0.1528",
"0.829",
"0.2627",
"0.829",
"0.2627",
"0.8403",
"0.1528",
"0.8403"
],
[
"0.511",
"0.8403",
"0.881",
"0.8403",
"0.881",
"0.854",
"0.511",
"0.854"
],
[
"0.1503",
"0.8516",
"0.4517",
"0.8516",
"0.4517",
"0.865",
"0.1503",
"0.865"
],
[
"0.513",
"0.8555",
"0.881",
"0.8555",
"0.881",
"0.869",
"0.513",
"0.869"
],
[
"0.1528",
"0.868",
"0.451",
"0.868",
"0.451",
"0.882",
"0.1528",
"0.882"
],
[
"0.513",
"0.8726",
"0.881",
"0.8726",
"0.881",
"0.8857",
"0.513",
"0.8857"
],
[
"0.1528",
"0.8843",
"0.294",
"0.8843",
"0.294",
"0.898",
"0.1528",
"0.898"
],
[
"0.538",
"0.8936",
"0.8833",
"0.8936",
"0.8833",
"0.9067",
"0.538",
"0.9067"
],
[
"0.1436",
"0.905",
"0.2896",
"0.905",
"0.2896",
"0.9185",
"0.1436",
"0.9185"
],
[
"0.511",
"0.9062",
"0.5483",
"0.9062",
"0.5483",
"0.9175",
"0.511",
"0.9175"
]
] | [
"Two Heads Are Better Than One:",
"通过与邻居关联来改进假新闻视频检测",
"Peng Qil,2, Yuyang Zhao3, Yufeng Shen2, Wei Ji3, Juan Cao1,2* and Tat-Seng Chua?",
"1 Key Laboratory of Intelligent Information Processing,",
"Institute of Computing Technology, Chinese Academy of Sciences",
"2 University of Chinese Academy of Sciences",
"3 National University of Singapore",
"( qipeng,caojuan @ict.ac.cn, [email protected],",
"[email protected],( jiwei,dcscts [email protected]",
"Abstract",
"Event: Man cries in pain as his clothing store is flooded.",
"DEBUNKING",
"The prevalence of short video platforms has",
"REAL",
"Thenet rumoi",
"spawned a lot of fake news videos, which have",
".\" isfake.",
"",
"",
"This video did",
"车",
"not take place",
"",
"",
"",
"in Henan",
"stronger propagation ability than textual fake",
"中",
"图",
"Province..",
"news. Thus, automatically detecting fake news",
"videos has been an important countermeasure",
"福",
"in practice. Previous works commonly ver-",
"Recently,Anhui",
"Province suffereda",
"#Henanheavy",
"ify each news video individually with multi-",
"rain# Nothing is",
"very heavy rainfall..",
"modal information. Nevertheless, news videos",
"moreimportant",
"Conflict",
"than life! Come",
"from different perspectives regarding the same",
"on!Henan!",
"event are commonly posted together, which",
"FAKE",
"contain complementary or contradictory in-",
"formation and thus can be used to evaluate",
"Figure 1: A set of videos belonging to the same event",
"each other mutually. To this end, we intro-",
"Fakenews videos containconflictinginformation with",
"duce a new and practical paradigm, i.e., cross-",
"the real ones, and the debunking videos can refute the",
"sample fake news video detection, and propose",
"mismatched information in thefakenews videos.",
"a novel framework, N eighbor- E nhanced fak",
"E news video D etection ( NEED), which inte-",
"Zheng Wei (Tang Dynasty)",
"grates the neighborhood relationship of new",
"videos belonging to the same event.NEED",
"假新闻的传播已成为一个重要的社会问题，",
"can be readily combined with existing single-",
"对政治（Fisheretal.,2016），金融（EIBoghdady，",
"sample detectors and further enhance their per-",
"2013），公共卫生（NaeemandBhatti,2020）等",
"formances with the proposed graph aggrega-",
"构成现实威胁。最近，短视频平台的盛行催",
"tion (GA) and debunking rectification (DR)",
"生了很多假新闻视频，与文本假新闻（Sundar",
"modules. Specifically, given the feature rep-",
"etal.,2021）相比，这些视频更有说服力，更",
"resentations obtained from single-sample de",
"tectors, GA aggregates the neighborhood in-",
"容易传播。中国网信办通报称，东航坠机事件",
"formation with the dynamic graph to enrich",
"中流传的7条核心谣言中，有5条源自短视频",
"thefeatures of independent samples.",
"Af",
"平台(Cyberspace Administration of China,2022)",
"ter that, DR explicitly leverages the relation-",
"另一项研究的统计数据也揭示了虚假新闻",
"ship between debunking videos and fake news",
"",
"视频的强大传播，该视频报告说，只有124条关",
"videos to refute the candidate videos via tex-",
"于COVID-19的TikTok虚假新闻获得了超过",
"tual and visual consistency. Extensive exper-",
"iments on the public benchmark demonstrate",
"2000万的点击量和200万个赞，评论和分享，",
"that NEED greatly improves the performance",
"对数百万人造成了负面影响（BrandyZadrozny",
"ofbothsingle-modal(upto 8.34%inaccu-",
"2021）。因此，迫切需要开发针对假新闻视频的",
"racy) and multimodal (up to 4.97 % in accu-",
"自动检测技术，以减轻其负面影响",
"racy) base detectors. Codes are available in",
"鉴于假新闻视频检测的实用性，以前的",
"https://github.com/ICTMCG/NEED.",
"作品E(Hou et al., 2019;Medina Serrano et al.,",
"2020; Choi and Ko, 2021; Shang et al., 2021; Qi",
"1.",
"Introduction",
"etal.,2023）利用单个新闻视频的异构多模态信",
"\"Listen to both sides and you will be en-",
"息进行佐证。然而，假新闻是故意制造的，以",
"lightened; heed only one side and you",
"误导消费者（Shuetal.,2017），因此多式联运",
"will be benighted.\"",
"tiktok.com. A popular short-form video sharing plat-",
"*Corresponding author.",
"form."
] | {
"arxiv_id": "2306.05241",
"language": "zh",
"page_idx": 1
} |
||
2306.05241-zh-2 | 2306.05241-zh-2.jpg | 组件在故意制造后几乎没有异常。此外，假新闻视频通常包含真实新闻视频，其中只有一些框架或文本描述被恶意修改以改变其含义(Qi et al.,2023)。上述特征表明，在单个视频中， 故意制造和恶意修改并不明显，导致现有作品的独立检测效率低
在现实世界的场景中，当新闻事件出现时， 会发布来自不同角度的多个相关视频，包括假新闻，真实新闻和揭穿视频。此类新闻视频包含互补或矛盾的信息，可用于相互评估。如图 1 所示 ,一方面 , 假新闻视频包含与真实的冲突信息( $i.e.$,不同位置：“安徽”省 v。s。“河南”省)。此外，揭穿视频也存在于某些事件中， 并且可以通过提供基于事实的反驳来轻松检测相应的假新闻。在一个新发布的基于短视频平台的数据集 (Qi et al.,2023)中，54 个含有假新闻视频的事件%也有揭穿视频，但 39 个含有揭穿视频的事件% 在揭穿视频发布后仍有假新闻视频传播。这些统计数据在一定程度上揭示了揭穿视频的普遍性和利用不足。基于上述观察，我们推测可以对同一事件的视频之间的关系进行建模，以增强假新闻视频的检测并通过事实信息纠正检测结果。为此，我们引入了用于假新闻视频检测的新的跨样本范式，并提出了相应的新颖框架 N eighbor- E nhancedfak E 新闻视频 D etection (NEED ),该框架显式和隐式地集成了邻域关系，以更好地检测。NEED 是一个与模型无关的框架，它可以轻松地合并各种单样本检测器以产生进一步的改进。因此我们首先从预先训练的单样本检测器获得特征表示，然后通过关系建模来改进表示和最终预测
为了弥补单个视频中信息的不足，我们将同一事件中的新闻视频以图的形式组织起来，以聚合邻域信息 (GraphAggregation)。具体来说， 我们利用事件图(Velickovic et al.,2018)上的注意力机制对不同节点之间的相关性进行建模， 并动态聚合这些特征。此外，如前所述，揭穿和假新闻视频之间存在明确的关系，$i.e.$,反驳。因此，可以采用揭穿视频来纠正错误的负面预测，发现“隐藏的”假新闻视频(揭穿纠正)。具体来说，我们制定了一个新的推理任务来区分给定的揭穿视频是否可以反驳给定的候选视频。对于给定的视频对，反驳通常存在于相同视觉场景的文本描述中，这启发了我们检测同一视觉表示的文本冲突。为了实现区分，我们从视频副本检测器获取视觉表示以获得视觉一致性，并通过注意力机制将其与来自文本冲突检测器的文本特征融合。然后利用融合特征
Debunking videos are videos that use factual evidence to refute widely circulated fake news, usually posted by experts.
对揭穿视频和候选视频之间的反驳关系进行分类。给定建议的图聚合和去块整流模块，NEED 可以显着提高使用单模态或多模态数据训练的基础单样本检测器的性能
我们的贡献总结如下：
$\bullet$ We propose a new cross-sample paradigm for fake news video detection, modeling multiple news videos in the same event simultaneously. Derived from such a paradigm, we propose the NEED framework, which exploits the neighborhood relationship explicitly and implicitly to enhance the fake news video detection.
$\bullet$ To the best of our knowledge, we are the first to utilize debunking videos in fake news video detection, which can utilize factual information to rectify false negative predictions. To this end, we formulate a new multimodal inference task and propose a novel model that utilizes the consistency from both the textual and visual perspectives to identify whether the given debunking video can refute the given candidate video.
· NEED is versatile and can be applied to various single-sample detectors. Extensive experiments on the public benchmark demonstrate that NEED can yield significant improvement with both single-modal and multimodal base detectors.
## 2 Related Work
为防御假新闻，研究人员主要致力于两种技术：假新闻检测方法通常使用非事实多模态信号，例如语言模式(Przybyla,2020),图像质量 (Qi et al., 2019; Cao et al., 2020) , 多模态不一致性 (Zhou et al., 2020; Qi et al., 2021), 用户响应(Shu et al., 2019) 和传播结构 (Ma et al. 2017),将给定的新闻分类为假新闻。随着短视频平台的盛行，检测假新闻视频越来越受到社会各界的关注。最近的工作主要利用深度神经网络来提取多模态特征并对跨模态相关性进行建模 (Choi and Ko, 2021; Shang et al., 2021; Palod et al., 2019; Qi et al., 2023)。例如，Qi et al. (2023) 使用交叉注意力转换器融合不同形式的新闻内容功能，包括文本，关键帧和音频，并使用自我注意力转换器将其与社交上下文功能融合，包括评论和用户
但是，虚假新闻视频检测中的现有工作是独立识别每个目标新闻，而没有考虑事件中的邻域关系。鉴于事件级过程的实用性，Wu et al. (2022)构造了一个跨文档的知识图，并采用异 | [
[
"0.12006",
"0.0873",
"0.4878",
"0.0873",
"0.4878",
"0.10095",
"0.12006",
"0.10095"
],
[
"0.5146",
"0.0885",
"0.876",
"0.0885",
"0.876",
"0.0986",
"0.5146",
"0.0986"
],
[
"0.1184",
"0.1027",
"0.4844",
"0.10156",
"0.4844",
"0.1152",
"0.1184",
"0.1164"
],
[
"0.513",
"0.1027",
"0.881",
"0.1027",
"0.881",
"0.1164",
"0.513",
"0.1164"
],
[
"0.12006",
"0.12054",
"0.4836",
"0.12054",
"0.4836",
"0.1312",
"0.12006",
"0.1312"
],
[
"0.5156",
"0.12054",
"0.879",
"0.12054",
"0.879",
"0.1312",
"0.5156",
"0.1312"
],
[
"0.11755",
"0.1354",
"0.4844",
"0.1354",
"0.4844",
"0.1484",
"0.11755",
"0.1484"
],
[
"0.513",
"0.1354",
"0.7153",
"0.1354",
"0.7153",
"0.1484",
"0.513",
"0.1484"
],
[
"0.1184",
"0.152",
"0.4844",
"0.152",
"0.4844",
"0.165",
"0.1184",
"0.165"
],
[
"0.532",
"0.1526",
"0.7036",
"0.1526",
"0.7036",
"0.1633",
"0.532",
"0.1633"
],
[
"0.1184",
"0.1675",
"0.266",
"0.1675",
"0.266",
"0.1812",
"0.1184",
"0.1812"
],
[
"0.531",
"0.1687",
"0.881",
"0.1692",
"0.881",
"0.1841",
"0.531",
"0.1829"
],
[
"0.1368",
"0.1853",
"0.4844",
"0.1853",
"0.4844",
"0.199",
"0.1368",
"0.199"
],
[
"0.5474",
"0.1841",
"0.882",
"0.1853",
"0.882",
"0.1996",
"0.5474",
"0.199"
],
[
"0.1184",
"0.2019",
"0.4844",
"0.2019",
"0.4844",
"0.2156",
"0.1184",
"0.2156"
],
[
"0.5483",
"0.2019",
"0.882",
"0.2019",
"0.882",
"0.2156",
"0.5483",
"0.2156"
],
[
"0.1184",
"0.2185",
"0.4844",
"0.2185",
"0.4844",
"0.2322",
"0.1184",
"0.2322"
],
[
"0.55",
"0.2173",
"0.881",
"0.2173",
"0.881",
"0.231",
"0.55",
"0.231"
],
[
"0.1184",
"0.234",
"0.4836",
"0.234",
"0.4836",
"0.2477",
"0.1184",
"0.2477"
],
[
"0.5483",
"0.234",
"0.881",
"0.234",
"0.881",
"0.2477",
"0.5483",
"0.2477"
],
[
"0.11755",
"0.2494",
"0.482",
"0.2494",
"0.482",
"0.2632",
"0.11755",
"0.2632"
],
[
"0.55",
"0.2494",
"0.881",
"0.2494",
"0.881",
"0.2632",
"0.55",
"0.2632"
],
[
"0.1184",
"0.2666",
"0.4844",
"0.2666",
"0.4844",
"0.2798",
"0.1184",
"0.2798"
],
[
"0.55",
"0.2673",
"0.881",
"0.2673",
"0.881",
"0.2778",
"0.55",
"0.2778"
],
[
"0.12006",
"0.282",
"0.4836",
"0.282",
"0.4836",
"0.2957",
"0.12006",
"0.2957"
],
[
"0.55",
"0.2832",
"0.602",
"0.2832",
"0.602",
"0.2944",
"0.55",
"0.2944"
],
[
"0.1184",
"0.2986",
"0.482",
"0.2986",
"0.482",
"0.3088",
"0.1184",
"0.3088"
],
[
"0.532",
"0.2957",
"0.882",
"0.2976",
"0.882",
"0.3123",
"0.532",
"0.31"
],
[
"0.1184",
"0.3142",
"0.4844",
"0.3142",
"0.4844",
"0.328",
"0.1184",
"0.328"
],
[
"0.5474",
"0.3135",
"0.881",
"0.3123",
"0.881",
"0.3267",
"0.5474",
"0.328"
],
[
"0.11755",
"0.3289",
"0.4844",
"0.33",
"0.4844",
"0.3445",
"0.11755",
"0.3433"
],
[
"0.5483",
"0.3289",
"0.881",
"0.3289",
"0.881",
"0.3425",
"0.5483",
"0.3425"
],
[
"0.1184",
"0.3467",
"0.4844",
"0.3467",
"0.4844",
"0.3599",
"0.1184",
"0.3599"
],
[
"0.5474",
"0.3457",
"0.8833",
"0.3467",
"0.8833",
"0.361",
"0.5474",
"0.3599"
],
[
"0.1184",
"0.3623",
"0.4844",
"0.3623",
"0.4844",
"0.376",
"0.1184",
"0.376"
],
[
"0.55",
"0.3623",
"0.8833",
"0.3623",
"0.8833",
"0.376",
"0.55",
"0.376"
],
[
"0.11755",
"0.3772",
"0.4836",
"0.3772",
"0.4836",
"0.3901",
"0.11755",
"0.3901"
],
[
"0.5483",
"0.379",
"0.881",
"0.379",
"0.881",
"0.3926",
"0.5483",
"0.3926"
],
[
"0.11755",
"0.3938",
"0.4836",
"0.3938",
"0.4836",
"0.4067",
"0.11755",
"0.4067"
],
[
"0.55",
"0.3943",
"0.881",
"0.3943",
"0.881",
"0.408",
"0.55",
"0.408"
],
[
"0.12006",
"0.412",
"0.4836",
"0.412",
"0.4836",
"0.4229",
"0.12006",
"0.4229"
],
[
"0.5483",
"0.411",
"0.881",
"0.411",
"0.881",
"0.4246",
"0.5483",
"0.4246"
],
[
"0.1184",
"0.427",
"0.4844",
"0.427",
"0.4844",
"0.4407",
"0.1184",
"0.4407"
],
[
"0.5483",
"0.427",
"0.881",
"0.427",
"0.881",
"0.4407",
"0.5483",
"0.4407"
],
[
"0.11755",
"0.4424",
"0.4836",
"0.4436",
"0.4836",
"0.4573",
"0.11755",
"0.456"
],
[
"0.5483",
"0.4424",
"0.7407",
"0.4424",
"0.7407",
"0.456",
"0.5483",
"0.456"
],
[
"0.1184",
"0.4602",
"0.4844",
"0.4602",
"0.4844",
"0.4739",
"0.1184",
"0.4739"
],
[
"0.534",
"0.459",
"0.8833",
"0.459",
"0.8833",
"0.4727",
"0.534",
"0.4727"
],
[
"0.11755",
"0.4756",
"0.4844",
"0.4756",
"0.4844",
"0.4893",
"0.11755",
"0.4893"
],
[
"0.55",
"0.4756",
"0.882",
"0.4756",
"0.882",
"0.4893",
"0.55",
"0.4893"
],
[
"0.1184",
"0.4912",
"0.4844",
"0.4912",
"0.4844",
"0.505",
"0.1184",
"0.505"
],
[
"0.55",
"0.4912",
"0.881",
"0.4912",
"0.881",
"0.506",
"0.55",
"0.506"
],
[
"0.12006",
"0.509",
"0.4836",
"0.509",
"0.4836",
"0.5195",
"0.12006",
"0.5195"
],
[
"0.5483",
"0.5083",
"0.881",
"0.5083",
"0.881",
"0.5215",
"0.5483",
"0.5215"
],
[
"0.1293",
"0.524",
"0.4844",
"0.524",
"0.4844",
"0.5376",
"0.1293",
"0.5376"
],
[
"0.55",
"0.524",
"0.881",
"0.524",
"0.881",
"0.5376",
"0.55",
"0.5376"
],
[
"0.1184",
"0.539",
"0.4844",
"0.539",
"0.4844",
"0.5527",
"0.1184",
"0.5527"
],
[
"0.55",
"0.5415",
"0.6206",
"0.5415",
"0.6206",
"0.5527",
"0.55",
"0.5527"
],
[
"0.1184",
"0.5557",
"0.486",
"0.5557",
"0.486",
"0.5693",
"0.1184",
"0.5693"
],
[
"0.1184",
"0.572",
"0.4895",
"0.572",
"0.4895",
"0.585",
"0.1184",
"0.585"
],
[
"0.5073",
"0.571",
"0.669",
"0.571",
"0.669",
"0.5845",
"0.5073",
"0.5845"
],
[
"0.12006",
"0.589",
"0.1368",
"0.589",
"0.1368",
"0.5996",
"0.12006",
"0.5996"
],
[
"0.5146",
"0.5986",
"0.881",
"0.5986",
"0.881",
"0.6123",
"0.5146",
"0.6123"
],
[
"0.1351",
"0.605",
"0.4836",
"0.605",
"0.4836",
"0.618",
"0.1351",
"0.618"
],
[
"0.513",
"0.615",
"0.8774",
"0.615",
"0.8774",
"0.626",
"0.513",
"0.626"
],
[
"0.1318",
"0.6216",
"0.4878",
"0.6216",
"0.4878",
"0.6353",
"0.1318",
"0.6353"
],
[
"0.513",
"0.6313",
"0.881",
"0.6313",
"0.881",
"0.6445",
"0.513",
"0.6445"
],
[
"0.1184",
"0.6396",
"0.4844",
"0.6396",
"0.4844",
"0.653",
"0.1184",
"0.653"
],
[
"0.513",
"0.6465",
"0.881",
"0.6465",
"0.881",
"0.66",
"0.513",
"0.66"
],
[
"0.11755",
"0.653",
"0.4844",
"0.653",
"0.4844",
"0.6665",
"0.11755",
"0.6665"
],
[
"0.5146",
"0.663",
"0.881",
"0.663",
"0.881",
"0.677",
"0.5146",
"0.677"
],
[
"0.1184",
"0.6704",
"0.4844",
"0.6704",
"0.4844",
"0.684",
"0.1184",
"0.684"
],
[
"0.513",
"0.6777",
"0.879",
"0.6777",
"0.879",
"0.6914",
"0.513",
"0.6914"
],
[
"0.1184",
"0.686",
"0.4844",
"0.686",
"0.4844",
"0.6997",
"0.1184",
"0.6997"
],
[
"0.513",
"0.695",
"0.881",
"0.695",
"0.881",
"0.708",
"0.513",
"0.708"
],
[
"0.1184",
"0.703",
"0.4836",
"0.703",
"0.4836",
"0.7163",
"0.1184",
"0.7163"
],
[
"0.513",
"0.7114",
"0.881",
"0.7114",
"0.881",
"0.7246",
"0.513",
"0.7246"
],
[
"0.12006",
"0.72",
"0.4844",
"0.72",
"0.4844",
"0.731",
"0.12006",
"0.731"
],
[
"0.513",
"0.727",
"0.879",
"0.727",
"0.879",
"0.7407",
"0.513",
"0.7407"
],
[
"0.1184",
"0.7354",
"0.4878",
"0.7354",
"0.4878",
"0.749",
"0.1184",
"0.749"
],
[
"0.511",
"0.7437",
"0.882",
"0.7437",
"0.882",
"0.7573",
"0.511",
"0.7573"
],
[
"0.1184",
"0.752",
"0.4844",
"0.752",
"0.4844",
"0.7656",
"0.1184",
"0.7656"
],
[
"0.511",
"0.7583",
"0.881",
"0.7583",
"0.881",
"0.7715",
"0.511",
"0.7715"
],
[
"0.12006",
"0.767",
"0.486",
"0.767",
"0.486",
"0.781",
"0.12006",
"0.781"
],
[
"0.513",
"0.7754",
"0.881",
"0.7754",
"0.881",
"0.789",
"0.513",
"0.789"
],
[
"0.12006",
"0.783",
"0.4844",
"0.783",
"0.4844",
"0.7964",
"0.12006",
"0.7964"
],
[
"0.513",
"0.792",
"0.8774",
"0.792",
"0.8774",
"0.8027",
"0.513",
"0.8027"
],
[
"0.12006",
"0.8",
"0.4844",
"0.8",
"0.4844",
"0.8135",
"0.12006",
"0.8135"
],
[
"0.513",
"0.808",
"0.881",
"0.808",
"0.881",
"0.822",
"0.513",
"0.822"
],
[
"0.1184",
"0.8154",
"0.486",
"0.8154",
"0.486",
"0.829",
"0.1184",
"0.829"
],
[
"0.513",
"0.8237",
"0.882",
"0.8237",
"0.882",
"0.8374",
"0.513",
"0.8374"
],
[
"0.1184",
"0.833",
"0.4836",
"0.833",
"0.4836",
"0.8433",
"0.1184",
"0.8433"
],
[
"0.511",
"0.839",
"0.6987",
"0.8384",
"0.6987",
"0.8525",
"0.511",
"0.854"
],
[
"0.12427",
"0.847",
"0.4844",
"0.847",
"0.4844",
"0.861",
"0.12427",
"0.861"
],
[
"0.532",
"0.857",
"0.881",
"0.857",
"0.881",
"0.8706",
"0.532",
"0.8706"
],
[
"0.11755",
"0.8643",
"0.4878",
"0.8643",
"0.4878",
"0.8774",
"0.11755",
"0.8774"
],
[
"0.5146",
"0.8735",
"0.881",
"0.8735",
"0.881",
"0.887",
"0.5146",
"0.887"
],
[
"0.511",
"0.8896",
"0.8774",
"0.8896",
"0.8774",
"0.903",
"0.511",
"0.903"
],
[
"0.1453",
"0.8955",
"0.4844",
"0.8955",
"0.4844",
"0.9062",
"0.1453",
"0.9062"
],
[
"0.11755",
"0.905",
"0.4836",
"0.9062",
"0.4836",
"0.9194",
"0.11755",
"0.9185"
],
[
"0.511",
"0.9067",
"0.876",
"0.9067",
"0.876",
"0.9175",
"0.511",
"0.9175"
]
] | [
"组件在故意制造后几乎没有异常。此外，假新",
"对揭穿视频和候选视频之间的反驳关系进行分",
"闻视频通常包含真实新闻视频，其中只有一些",
"类。给定建议的图聚合和去块整流模块，NEED",
"框架或文本描述被恶意修改以改变其含义（Qi",
"可以显着提高使用单模态或多模态数据训练的",
"etal.，2023）。上述特征表明，在单个视频中",
"基础单样本检测器的性能",
"故意制造和恶意修改并不明显，导致现有作品",
"我们的贡献总结如下：",
"的独立检测效率低",
".Weproposea new cross-sampleparadigm for",
"在现实世界的场景中，当新闻事件出现时",
"fake news video detection, modeling multi-",
"会发布来自不同角度的多个相关视频，包括假",
"ple news videos in the same event simulta-",
"新闻，真实新闻和揭穿视频。此类新闻视频",
"neously. Derived from such a paradigm, we",
"包含互补或矛盾的信息，可用于相互评估。如",
"propose the NEED framework, which exploits",
"图1所示，一方面，假新闻视频包含与真实的",
"the neighborhood relationship explicitly and",
"冲突信息（i.e.,不同位置：“安徽”省V。S。“河",
"implicitlytoenhancethefakenewsvideode",
"南”省)。此外，揭穿视频也存在于某些事件中",
"tection",
"并日可以通过提供基于事实的反驳来轻松检测",
". To the best of our knowledge, we are the",
"相应的假新闻。在一个新发布的基于短视频平",
"first to utilize debunking videos in fake news",
"台的数据集（Qietal.,2023）中，54个含有假新",
"video detection, which can utilize factual",
"闻视频的事件%也有揭穿视频，但39个含有",
"information to rectify false negative predic-",
"揭穿视频的事件%在揭穿视频发布后仍有假",
"tions. To this end, we formulate a new mul-",
"新闻视频传播。这些统计数据在一定程度上揭",
"timodal inference task and propose a novel",
"示了揭穿视频的普遍性和利用不足。基于上述",
"model that utilizes the consistency from both",
"观察，我们推测可以对同一事件的视频之间的",
"the textual and visual perspectives to identify",
"关系进行建模，以增强假新闻视频的检测并通",
"whether the given debunking video can refute",
"过事实信息纠正检测结果。为此，我们引人了",
"the given candidate video.",
"用于假新闻视频检测的新的跨样本范式，并提",
": NEED is versatile and can be applied to vari-",
"出了相应的新颖框架Neighbor-EnhancedfakE",
"ous single-sample detectors. Extensive exper-",
"新闻视频Detection（NEED），该框架显式和隐",
"iments on the public benchmark demonstrate",
"式地集成了邻域关系，以更好地检测。NEED是",
"that NEED can yield significant improvement",
"个与模型无关的框架，它可以轻松地合并各",
"with both single-modal and multimodal base",
"种单样本检测器以产生进一步的改进。因此，",
"detectors.",
"我们首先从预先训练的单样本检测器获得特征",
"表示，然后通过关系建模来改进表示和最终预",
"2Related Work",
"测",
"为防御假新闻，研究人员主要致力于两种技",
"为了弥补单个视频中信息的不足，我们将同",
"术：假新闻检测方法通常使用非事实多模态信",
"-事件中的新闻视频以图的形式组织起来，以",
"号，例如语言模式（Przybyla，2020），图像质",
"聚合邻域信息（GraphAggregation）。具体来说，",
"量（Qietal.,2019;Caoetal.,2020），多模态不",
"我们利用事件图（Velickovicetal.,2018）上的注",
"一致性（Zhouetal.,2020;Qietal.,2021），用",
"意力机制对不同节点之间的相关性进行建模，",
"户响应（Shuetal,2019）和传播结构（Maetal",
"并动态聚合这些特征。此外，如前所述，揭穿",
"2017），将给定的新闻分类为假新闻。随着短",
"和假新闻视频之间存在明确的关系，i.e.，反驳。",
"视频平台的盛行，检测假新闻视频越来越受到",
"因此，可以采用揭穿视频来纠正错误的负面预",
"社会各界的关注。最近的工作主要利用深度神",
"测，发现“隐藏的”假新闻视频（揭穿纠正）。",
"经网络来提取多模态特征并对跨模态相关性",
"具体来说，我们制定了一个新的推理任务来区",
"进行建模(Choi and Ko,2021; Shang et al.,2021;",
"分给定的揭穿视频是否可以反驳给定的候选",
"Palod et al., 2019; Qi et al., 2023)。例如, Qi et al.",
"视频。对于给定的视频对，反驳通常存在于相",
"（2023）使用交叉注意力转换器融合不同形式的",
"同视觉场景的文本描述中，这启发了我们检测",
"新闻内容功能，包括文本，关键顿和音频，并",
"同一视觉表示的文本冲突。为了实现区分，我",
"使用自我注意力转换器将其与社交上下文功能",
"们从视频副本检测器获取视觉表示以获得视觉",
"融合，包括评论和用户",
"一致性，并通过注意力机制将其与来自文本冲",
"但是，虚假新闻视频检测中的现有工作是独",
"突检测器的文本特征融合。然后利用融合特征",
"立识别每个目标新闻，而没有考虑事件中的邻",
"域关系。鉴于事件级过程的实用性，Wuetal",
"Debunking videos are videos that use factual evidence to",
"refute widely circulated fake news,usually posted by experts.",
"（2022）构造了一个跨文档的知识图：并采用异"
] | {
"arxiv_id": "2306.05241",
"language": "zh",
"page_idx": 2
} |
||
2306.05241-zh-3 | 2306.05241-zh-3.jpg | 构图神经网络来检测错误信息。尽管如此，这项工作是在合成数据集上进行的，其中每个假新闻文档都来自一个经过操纵的知识图，该知识图不能很容易地应用于信息提取中具有不可预测噪声的真实场景。此外，他们只考虑新闻文本之间的隐含关系，而忽略了揭穿信息和假新闻之间的明确反驳。事实检查方法通常依赖于从可靠来源(如 Wikipedia(Thorne et al., 2018)和网页(Nie et al.,2019) 检索到的相关事实信息来判断给定的值得检查的声明 (Guo et al., 2022; Zeng et al., 2021) 的真实性。最近的一个线程是在检索证据 (Sheng et al., 2021) 之前确定声明是否已经过事实检查。此任务通常被设计为排名任务，根据与给定声明的相似性对事实检查文章进行排名。与文本事实核查相比，多模式核查的探索不足 Mishra et al. (2022) 将验证视为多模式任务，其中模型需要对给定的可靠文档(带有关联图像的文本)和值得检查的声明 (带有关联图像的文本)之间的关系进行分类。受这些工作的启发，NEED 中的揭穿纠正模块着重于通过识别给定揭穿与候选新闻视频之间的反驳关系来纠正先前经过事实检查的新闻视频的错误预测。总之，虚假的新检测方法利用从大规模数据中学习的非事实模式来对新出现的事件做出及时的判断，而事实检查技术提供了更可靠的判断，受益于事实信息， 但只适用于受外部来源覆盖限制的部分事件。我们的工作结合了这两种方法的优点：(1)我们利用数据驱动的假新闻视频检测器来获得有效的多模式表示并对邻域信息进行建模；(2) 我们还在事实检查中采用了相关事实信息的概念，以通过可靠的揭穿视频来纠正检测结果
# 3 $\textbf{Methodology}$
# 3.1 0verview
如引言中所述，虚假新闻视频的制造和恶意修改限制了现有单样本虚假新闻视频检测器的验证能力，导致性能下降。相反，相邻视频之间的关系，$i.e.$,视频的同一事件，可以用来补充目前的技术。因此，我们提出了 N eighbor- E 改进的 fak E 新闻视频 D etection (NEED) 框架，利用事件中的视频集，包括假新闻$I_F$, 真实新闻$I_R$和揭穿视频$I_D$,以提高单样本检测器的性能。具体来说，NEED 是与模型无关的，它从预先训练的基本检测器$\overline{\mathrm{E}}($ Feature Extraction ) 中获取表示，以构建动态图并聚合邻域信息( Graph Aggregation, GA )。然后， 我们使用来自揭穿视频的事实信息来纠正预测结果( Debunking Rectification, DR )。整体框架如图 2 所示
## 3.2 Feature Extraction
新闻视频包含多模态信息，包括标题，音频， 关键帧，视频剪辑，评论，用户个人资料，$etc.$ 现有的单样本假新闻视频检测器利用单模态$\overline{\mathrm{E}}$(Medina Serrano et al.,2020) 或多模态E (Qi et al., 2023) 信息来独立区分每个新闻视频。他们通常设计定制的模块来提取和融合多模式特征。相比之下，NEED 是跨样本范例的解决方案，它可以合并各种单样本假新闻视频检测器，以通过邻域建模进一步改进。因此，我们首先从基本单样本检测器中提取给定新闻视频集的单模态/多模态特征$F_\mathrm{base}$
# 3.3 Graph Aggregation
图形构造。给定同一事件$E$下的相关新闻视频特征集$F_\mathrm{base}^E$,我们将它们组织成图注意力网络 (GAT)$\operatorname{E}($Velickovic et al.,2018)的形式。$\mathcal{G}$表示图形，$\mathcal{V}$表示$\mathcal{G}$中的节点，$\mathcal{E}$表示节点之间的边。每个节点$v_i\in\mathcal{V}$表示来自基本检测器的新闻视频特征，边缘$e_{ij}$表示节点$j$的特征相对于节点$i$的特征的重要性，这是通过注意力机制获得的
特征聚合和分类。为了聚合邻居信息，我们在构造的事件图$G$ 上应用注意力机制来更新节点的表示。具体来说，给定节点$v_i$及其邻居$\mathcal{N} _i$ , $\boldsymbol{v}_i$与其邻居$v_j\in\mathcal{N}_i$之间的权重$\alpha_i,j$表示为：
$$e_{ij}=\mathrm{LeakyReLU}(a^{\top}[\boldsymbol{W}\boldsymbol{v}_{i},\boldsymbol{W}\boldsymbol{v}_{j}]),\\\alpha_{ij}=\mathrm{softmax}_{j}(e_{ij})=\frac{\exp(e_{ij})}{\sum_{k\in\mathcal{N}_{i}}\exp(e_{ik})},$$
(1)
其中$a$和$W$是可训练的参数，T 表示转换矩
阵。然后，通过聚合信息更新$v_i$的嵌入：
(2)
$$\hat{\boldsymbol{v}}_i=\sigma(\sum_{j\in\mathcal{N}_i}\alpha_{ij}\boldsymbol{W}\boldsymbol{v}_j),$$
其中$\sigma$是非线性运算。为了避免节点特征的过度平滑，我们仅采用两个 GAT 层。最终特征$\hat{v}_i$被馈送到二进制分类器中以验证视频。网络通过二进制交叉熵损失进行优化：
$$\mathcal{L}=-[(1-y)\log(1-p_{\mathrm{GA}})+y\log p_{\mathrm{GA}}],$$
其中$p_\mathrm{GA}$是预测概率，$y\in\{0,1\}$表示地面实
况标签
## 3.4 Debunking Rectification
图形聚合着重于组合从基础检测器获得的邻域特征，这些特征从大规模数据中学习非事实模式。相反，假新闻视频和用事实信息揭穿视频之间也存在明确的关系，$i.e.$,反驳。因此，我 | [
[
"0.12006",
"0.0873",
"0.708",
"0.0873",
"0.708",
"0.10095",
"0.12006",
"0.10095"
],
[
"0.12006",
"0.1027",
"0.4878",
"0.1027",
"0.4878",
"0.1164",
"0.12006",
"0.1164"
],
[
"0.513",
"0.1081",
"0.882",
"0.1081",
"0.882",
"0.1217",
"0.513",
"0.1217"
],
[
"0.1184",
"0.1194",
"0.4878",
"0.1194",
"0.4878",
"0.133",
"0.1184",
"0.133"
],
[
"0.513",
"0.12354",
"0.8774",
"0.12354",
"0.8774",
"0.1372",
"0.513",
"0.1372"
],
[
"0.12006",
"0.1354",
"0.4878",
"0.1354",
"0.4878",
"0.1484",
"0.12006",
"0.1484"
],
[
"0.5146",
"0.1414",
"0.879",
"0.1414",
"0.879",
"0.152",
"0.5146",
"0.152"
],
[
"0.12006",
"0.152",
"0.4844",
"0.152",
"0.4844",
"0.165",
"0.12006",
"0.165"
],
[
"0.513",
"0.1561",
"0.882",
"0.1561",
"0.882",
"0.1692",
"0.513",
"0.1692"
],
[
"0.1184",
"0.1675",
"0.486",
"0.1675",
"0.486",
"0.1812",
"0.1184",
"0.1812"
],
[
"0.51",
"0.1704",
"0.881",
"0.1704",
"0.881",
"0.1841",
"0.51",
"0.1841"
],
[
"0.11755",
"0.1818",
"0.486",
"0.1818",
"0.486",
"0.1953",
"0.11755",
"0.1953"
],
[
"0.5146",
"0.187",
"0.882",
"0.187",
"0.882",
"0.2007",
"0.5146",
"0.2007"
],
[
"0.1184",
"0.199",
"0.4878",
"0.199",
"0.4878",
"0.212",
"0.1184",
"0.212"
],
[
"0.513",
"0.2037",
"0.881",
"0.2037",
"0.881",
"0.2173",
"0.513",
"0.2173"
],
[
"0.1184",
"0.2156",
"0.4878",
"0.2156",
"0.4878",
"0.2286",
"0.1184",
"0.2286"
],
[
"0.513",
"0.2197",
"0.879",
"0.2197",
"0.879",
"0.2328",
"0.513",
"0.2328"
],
[
"0.1184",
"0.2322",
"0.4878",
"0.2322",
"0.4878",
"0.2452",
"0.1184",
"0.2452"
],
[
"0.511",
"0.2363",
"0.879",
"0.2363",
"0.879",
"0.2494",
"0.511",
"0.2494"
],
[
"0.11755",
"0.2477",
"0.4844",
"0.2477",
"0.4844",
"0.2612",
"0.11755",
"0.2612"
],
[
"0.513",
"0.2517",
"0.881",
"0.2517",
"0.881",
"0.2654",
"0.513",
"0.2654"
],
[
"0.12427",
"0.2632",
"0.486",
"0.2632",
"0.486",
"0.2778",
"0.12427",
"0.2778"
],
[
"0.51",
"0.2673",
"0.748",
"0.2695",
"0.748",
"0.284",
"0.51",
"0.282"
],
[
"0.11755",
"0.279",
"0.4836",
"0.279",
"0.4836",
"0.2922",
"0.11755",
"0.2922"
],
[
"0.1184",
"0.2957",
"0.4878",
"0.2957",
"0.4878",
"0.3088",
"0.1184",
"0.3088"
],
[
"0.5073",
"0.2935",
"0.712",
"0.2944",
"0.712",
"0.3088",
"0.5073",
"0.308"
],
[
"0.1184",
"0.3123",
"0.486",
"0.3123",
"0.486",
"0.326",
"0.1184",
"0.326"
],
[
"0.513",
"0.3154",
"0.881",
"0.3154",
"0.881",
"0.3289",
"0.513",
"0.3289"
],
[
"0.1184",
"0.328",
"0.4878",
"0.328",
"0.4878",
"0.3416",
"0.1184",
"0.3416"
],
[
"0.51",
"0.33",
"0.881",
"0.3308",
"0.881",
"0.3457",
"0.51",
"0.3445"
],
[
"0.11755",
"0.3433",
"0.4878",
"0.3445",
"0.4878",
"0.3582",
"0.11755",
"0.357"
],
[
"0.513",
"0.3467",
"0.882",
"0.3467",
"0.882",
"0.361",
"0.513",
"0.361"
],
[
"0.11755",
"0.3594",
"0.486",
"0.3594",
"0.486",
"0.3723",
"0.11755",
"0.3723"
],
[
"0.513",
"0.3635",
"0.881",
"0.3635",
"0.881",
"0.3772",
"0.513",
"0.3772"
],
[
"0.1184",
"0.3772",
"0.4844",
"0.3772",
"0.4844",
"0.3901",
"0.1184",
"0.3901"
],
[
"0.511",
"0.3801",
"0.882",
"0.3801",
"0.882",
"0.3938",
"0.511",
"0.3938"
],
[
"0.1184",
"0.3938",
"0.4878",
"0.3938",
"0.4878",
"0.4067",
"0.1184",
"0.4067"
],
[
"0.513",
"0.3967",
"0.881",
"0.3967",
"0.881",
"0.4104",
"0.513",
"0.4104"
],
[
"0.1184",
"0.4092",
"0.4844",
"0.4092",
"0.4844",
"0.4229",
"0.1184",
"0.4229"
],
[
"0.511",
"0.412",
"0.881",
"0.412",
"0.881",
"0.4258",
"0.511",
"0.4258"
],
[
"0.12006",
"0.427",
"0.4844",
"0.427",
"0.4844",
"0.437",
"0.12006",
"0.437"
],
[
"0.513",
"0.4287",
"0.66",
"0.4287",
"0.66",
"0.4424",
"0.513",
"0.4424"
],
[
"0.1184",
"0.4412",
"0.4844",
"0.4412",
"0.4844",
"0.4548",
"0.1184",
"0.4548"
],
[
"0.511",
"0.4448",
"0.882",
"0.4448",
"0.882",
"0.4578",
"0.511",
"0.4578"
],
[
"0.12006",
"0.4573",
"0.4878",
"0.4573",
"0.4878",
"0.4702",
"0.12006",
"0.4702"
],
[
"0.511",
"0.4614",
"0.881",
"0.4614",
"0.881",
"0.4744",
"0.511",
"0.4744"
],
[
"0.1184",
"0.4744",
"0.4844",
"0.4744",
"0.4844",
"0.485",
"0.1184",
"0.485"
],
[
"0.5146",
"0.4768",
"0.881",
"0.4768",
"0.881",
"0.4905",
"0.5146",
"0.4905"
],
[
"0.1184",
"0.4893",
"0.4902",
"0.4893",
"0.4902",
"0.503",
"0.1184",
"0.503"
],
[
"0.509",
"0.4934",
"0.881",
"0.4934",
"0.881",
"0.5073",
"0.509",
"0.5073"
],
[
"0.11755",
"0.505",
"0.4878",
"0.505",
"0.4878",
"0.5186",
"0.11755",
"0.5186"
],
[
"0.513",
"0.5103",
"0.5576",
"0.5103",
"0.5576",
"0.5215",
"0.513",
"0.5215"
],
[
"0.1184",
"0.521",
"0.4844",
"0.521",
"0.4844",
"0.5337",
"0.1184",
"0.5337"
],
[
"0.1184",
"0.5376",
"0.4836",
"0.5376",
"0.4836",
"0.5503",
"0.1184",
"0.5503"
],
[
"0.532",
"0.5337",
"0.818",
"0.533",
"0.818",
"0.552",
"0.532",
"0.5527"
],
[
"0.1184",
"0.554",
"0.4878",
"0.554",
"0.4878",
"0.568",
"0.1184",
"0.568"
],
[
"0.739",
"0.5557",
"0.804",
"0.5557",
"0.804",
"0.571",
"0.739",
"0.571"
],
[
"0.8237",
"0.5586",
"0.8447",
"0.5586",
"0.8447",
"0.5625",
"0.8237",
"0.5625"
],
[
"0.8555",
"0.5557",
"0.881",
"0.5557",
"0.881",
"0.5684",
"0.8555",
"0.5684"
],
[
"0.1184",
"0.5684",
"0.4844",
"0.5684",
"0.4844",
"0.582",
"0.1184",
"0.582"
],
[
"0.529",
"0.5635",
"0.714",
"0.564",
"0.714",
"0.582",
"0.529",
"0.5806"
],
[
"0.7095",
"0.571",
"0.837",
"0.5723",
"0.837",
"0.5913",
"0.7095",
"0.589"
],
[
"0.12006",
"0.5874",
"0.4634",
"0.5874",
"0.4634",
"0.597",
"0.12006",
"0.597"
],
[
"0.5146",
"0.6016",
"0.881",
"0.6016",
"0.881",
"0.615",
"0.5146",
"0.615"
],
[
"0.1125",
"0.6206",
"0.2644",
"0.6226",
"0.2644",
"0.637",
"0.1125",
"0.6353"
],
[
"0.5146",
"0.618",
"0.8403",
"0.618",
"0.8403",
"0.632",
"0.5146",
"0.632"
],
[
"0.607",
"0.6426",
"0.7886",
"0.6445",
"0.7886",
"0.663",
"0.607",
"0.661"
],
[
"0.858",
"0.648",
"0.881",
"0.648",
"0.881",
"0.661",
"0.858",
"0.661"
],
[
"0.11584",
"0.654",
"0.2393",
"0.654",
"0.2393",
"0.6675",
"0.11584",
"0.6675"
],
[
"0.6665",
"0.662",
"0.705",
"0.662",
"0.705",
"0.6777",
"0.6665",
"0.6777"
],
[
"0.1184",
"0.6797",
"0.4844",
"0.6797",
"0.4844",
"0.6934",
"0.1184",
"0.6934"
],
[
"0.513",
"0.6895",
"0.881",
"0.6895",
"0.881",
"0.703",
"0.513",
"0.703"
],
[
"0.12006",
"0.6978",
"0.4844",
"0.6978",
"0.4844",
"0.708",
"0.12006",
"0.708"
],
[
"0.511",
"0.705",
"0.881",
"0.7036",
"0.881",
"0.7188",
"0.511",
"0.7197"
],
[
"0.12006",
"0.712",
"0.4844",
"0.712",
"0.4844",
"0.7256",
"0.12006",
"0.7256"
],
[
"0.511",
"0.72",
"0.879",
"0.72",
"0.879",
"0.734",
"0.511",
"0.734"
],
[
"0.12006",
"0.7285",
"0.4844",
"0.7285",
"0.4844",
"0.7393",
"0.12006",
"0.7393"
],
[
"0.513",
"0.7383",
"0.777",
"0.7383",
"0.777",
"0.752",
"0.513",
"0.752"
],
[
"0.12006",
"0.7446",
"0.4878",
"0.7446",
"0.4878",
"0.7583",
"0.12006",
"0.7583"
],
[
"0.11755",
"0.759",
"0.486",
"0.759",
"0.486",
"0.7725",
"0.11755",
"0.7725"
],
[
"0.5264",
"0.764",
"0.879",
"0.7656",
"0.879",
"0.78",
"0.5264",
"0.7793"
],
[
"0.11755",
"0.7754",
"0.471",
"0.7754",
"0.471",
"0.789",
"0.11755",
"0.789"
],
[
"0.4668",
"0.78",
"0.486",
"0.783",
"0.4836",
"0.789",
"0.4634",
"0.786"
],
[
"0.1184",
"0.792",
"0.4895",
"0.7915",
"0.4895",
"0.8057",
"0.1184",
"0.807"
],
[
"0.511",
"0.792",
"0.881",
"0.7935",
"0.881",
"0.808",
"0.511",
"0.807"
],
[
"0.1184",
"0.808",
"0.4895",
"0.808",
"0.4895",
"0.822",
"0.1184",
"0.822"
],
[
"0.492",
"0.814",
"0.509",
"0.814",
"0.509",
"0.818",
"0.492",
"0.818"
],
[
"0.511",
"0.811",
"0.5684",
"0.811",
"0.5684",
"0.8223",
"0.511",
"0.8223"
],
[
"0.12006",
"0.8267",
"0.482",
"0.8267",
"0.482",
"0.8374",
"0.12006",
"0.8374"
],
[
"0.1184",
"0.8413",
"0.4878",
"0.8413",
"0.4878",
"0.855",
"0.1184",
"0.855"
],
[
"0.51",
"0.836",
"0.748",
"0.836",
"0.748",
"0.8496",
"0.51",
"0.8496"
],
[
"0.1184",
"0.858",
"0.4844",
"0.858",
"0.4844",
"0.8716",
"0.1184",
"0.8716"
],
[
"0.5156",
"0.8594",
"0.879",
"0.8594",
"0.879",
"0.869",
"0.5156",
"0.869"
],
[
"0.1184",
"0.8745",
"0.4844",
"0.8745",
"0.4844",
"0.8853",
"0.1184",
"0.8853"
],
[
"0.513",
"0.8745",
"0.879",
"0.8745",
"0.879",
"0.8853",
"0.513",
"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"
],
[
"0.11755",
"0.904",
"0.2334",
"0.905",
"0.2334",
"0.9194",
"0.11755",
"0.9185"
],
[
"0.513",
"0.9062",
"0.882",
"0.9062",
"0.882",
"0.9194",
"0.513",
"0.9194"
]
] | [
"构图神经网络来检测错误信息。尽管如此，这3.2FeatureExtraction",
"项工作是在合成数据集上进行的，其中每个假",
"新闻视频包含多模态信息，包括标题，音频，",
"新闻文档都来自一个经过操纵的知识图，该知",
"关键顿，视频剪辑，评论，用户个人资料，etc",
"识图不能很容易地应用于信息提取中具有不",
"现有的单样本假新闻视频检测器利用单模",
"可预测噪声的真实场景。此外，他们只考虑新",
"态E(MedinaSerranoetal.,2020）或多模态E",
"闻文本之间的隐含关系，而忽略了揭穿信息和",
"（Qietal.,2023）信息来独立区分每个新闻视频。",
"假新闻之间的明确反驳。事实检查方法通常依",
"他们通常设计定制的模块来提取和融合多模式",
"赖于从可靠来源（如Wikipedia（Thorne etal.,",
"特征。相比之下，NEED是跨样本范例的解决",
"2018）和网页（Nieetal.,2019））检索到的相关",
"方案，它可以合并各种单样本假新闻视频检测",
"事实信息来判断给定的值得检查的声明（Guo",
"器，以通过邻域建模进一步改进。因此，我们",
"etal.,2022;Zengetal.，2021）的真实性。最近的",
"首先从基本单样本检测器中提取给定新闻视频",
"一个线程是在检索证据(Shengetal.,2021)之",
"集的单模态/多模态特征Fbase",
"前确定声明是否已经过事实检查。此任务通常",
"被设计为排名任务，根据与给定声明的相似性",
"3.3Graph Aggregation",
"对事实检查文章进行排名。与文本事实核查相",
"图形构造。给定同一事件E下的相关新闻视",
"比，多模式核查的探索不足Mishraetal.(2022)",
"频特征集FEase，我们将它们组织成图注意力",
"将验证视为多模式任务，其中模型需要对给定",
"网络（GAT)E(Velickovicetal.,2018）的形式。",
"的可靠文档（带有关联图像的文本）和值得检",
"9表示图形，V表示G中的节点，8表示节点",
"查的声明（带有关联图像的文本）之间的关系",
"之间的边。每个节点UEV表示来自基本检",
"进行分类。受这些工作的启发，NEED中的揭穿",
"测器的新闻视频特征，边缘ei表示节点j的",
"纠正模块着重于通过识别给定揭穿与候选新闻",
"特征相对于节点的特征的重要性，这是通过",
"视频之间的反驳关系来纠止先前经过事实检查",
"注意力机制获得的",
"的新闻视频的错误预测。总之，虚假的新检测",
"特征聚合和分类。为了聚合邻居信息，我们在",
"方法利用从大规模数据中学习的非事实模式来",
"构造的事件图9上应用注意力机制来更新节",
"对新出现的事件做出及时的判断，而事实检查",
"点的表示。具体来说，给定节点Ui及其邻居",
"技术提供了更可靠的判断，受益于事实信息，",
"Ni，vi与其邻居ENi之间的权重αij表",
"但只适用于受外部来源覆盖限制的部分事件。",
"示为：",
"我们的工作结合了这两种方法的优点：（1）我",
"们利用数据驱动的假新闻视频检测器来获得有",
"eij = LeakyReLU(a [Wi, Wo,l),",
"效的多模式表示并对邻域信息进行建模；（2）",
"exp(ei)",
"",
"(1)",
"我们还在事实检查中采用了相关事实信息的概",
"Qij = softmax;(ei)=",
"ZheN exp(eik)",
"念，以通过可靠的揭穿视频来纠正检测结果",
"其中a和W是可训练的参数，T表示转换矩",
"3Methodology",
"阵。然后，通过聚合信息更新i的嵌入：",
"ni =o( qijWuj),",
"(2)",
"3.1Overview",
"jeN",
"如引言中所述，虚假新闻视频的制造和恶意修",
"其中α是非线性运算。为了避免节点特征的过",
"改限制了现有单样本虚假新闻视频检测器的验",
"度平滑，我们仅采用两个GAT层。最终特征",
"证能力，导致性能下降。相反，相邻视频之间",
"；被馈送到二进制分类器中以验证视频。网络",
"的关系，i.e.，视频的同一事件，可以用来补充",
"通过二进制交叉摘损失进行优化：",
"目前的技术。因此，我们提出了Neighbor-E",
"改进的fakE新闻视频Detection（NEED）框",
"L = -[(1 - y) log(1 - pGA) + y log PGA], (3)",
"架，利用事件中的视频集，包括假新闻IF",
"F",
"真实新闻IR和揭穿视频ID，以提高单样本",
"其中PGA是预测概率，yE0,1}表示地面实",
"检测器的性能。具体来说，NEED是与模型无",
"",
"况标签",
"关的：它从预先训练的基本检测器E（Feature",
"Extraction）中获取表示，以构建动态图并聚",
"3.4DebunkingRectification",
"合邻域信息（GraphAggregation，GA）。然后",
"图形聚合看重于组合从基础检测器获得的邻域",
"我们使用来自揭穿视频的事实信息来纠正预测",
"特征，这些特征从大规模数据中学习非事实模",
"结果（DebunkingRectification，DR）。整体框",
"式。相反，假新闻视频和用事实信息揭穿视频",
"架如图2所示",
"之间也存在明确的关系，i.e.，反驳。因此，我"
] | {
"arxiv_id": "2306.05241",
"language": "zh",
"page_idx": 3
} |
||
2010.03933-zh-0 | 2010.03933-zh-0.jpg | Assessing Classifier Fairness with Collider Bias $^\star$
Zhenlong Xu$^{\dagger1},\text{Ziqi Xu}^{\dagger1[0000-0003-1748-5801]}$,Jixue Liu$^{1[0000-0002-0794-0404]}$,
$\begin{array}{c}{{\mathrm{m_{5}~1u~,~24q~1u~}}}\\{{\mathrm{Debo~Cheng}^{1[0000-0002-0383-1462]},~\mathrm{Jiuyong~Li}^{1[0000-0002-9023-1878]}}}\\{{\mathrm{Lin~Liu}^{1[0000-0003-2843-5738]},~\mathrm{and~Ke~Wang}^{2}}}\end{array}$,
$^{1}$ University of South Australia, Adelaide, Australia $\{$
Zhenlong.Xu,Ziqi.Xu,Debo.Cheng $\}$ @mymail.unisa.edu.au,
$\{$ Jixue.Liu,Jiuyong.Li,Lin.Liu $\}$ @unisa.edu.au
$^{2}$ Simon Fraser University, Burnaby, Canada
wangk@ sfu. ca
$\mathbf{Abstract.~The~increasing~application~of~machine~learning~techniques}$ in everyday decision-making processes has brought concerns about the fairness of algorithmic decision-making. This paper concerns the problem of collider bias which produces spurious associations in fairness assessment and develops theorems to guide fairness assessment avoiding the collider bias. We consider a real-world application of auditing a trained classifier by an audit agency. We propose an unbiased assessment algorithm by utilising the developed theorems to reduce collider biases in the assessment. Experiments and simulations show the proposed algorithm reduces collider biases significantly in the assessment and is promising in auditing trained classifiers.
Keywords: Fairness· Collider bias $\cdot$ Causal inference.
## 1 Introduction
<FigureHere>
随着机器学习在各种应用中的广泛使用，人们越来越担心决策算法的公平性，例如工作招聘，信用评分和房屋贷款，因为歧视可能会无意中引人机器学习模型。为了防止模型中的不公平在社会中蔓延，独立机构需要审计技术来审计机器学习模型。图 1 显示了一个审计过程。审计机构访问公司的模型，并有自己的审计案例来评估模型的公平性。审计机构无权访 Fig.1. The process of audit.
问培训数据集，但具有监管政策。在本文中，我们使用因果图来表示监管政策。公司可能会使用监管政策中未指定的其他变量来构建模型，以提高预测准确性
情况测试已在美国使用。S。检测招聘中的歧视[1],这是一种对照实验方法，用于分析雇主对求职者特征的决定，如以下示例所示。成对的研究助理被派去申请同一份工作，每对假装的申请人都有与工作相关的相同资格和经验，但
$^{\star}$ Supported by Australian Research Council (DP200101210) and Natural Sciences and
Engineering Research Council of Canada.
$^{\dagger}$ Ziqi Xu and Zhenlong Xu contributed equally to this paper. | [
[
"0.2255",
"0.1465",
"0.7686",
"0.1465",
"0.7686",
"0.161",
"0.2255",
"0.161"
],
[
"0.763",
"0.1477",
"0.7793",
"0.1477",
"0.7793",
"0.1559",
"0.763",
"0.1559"
],
[
"0.2181",
"0.1881",
"0.7837",
"0.1881",
"0.7837",
"0.2058",
"0.2181",
"0.2058"
],
[
"0.259",
"0.202",
"0.7427",
"0.202",
"0.7427",
"0.2229",
"0.259",
"0.2229"
],
[
"0.3381",
"0.2166",
"0.664",
"0.2172",
"0.664",
"0.238",
"0.3381",
"0.2374"
],
[
"0.3162",
"0.2443",
"0.687",
"0.2468",
"0.687",
"0.2644",
"0.3162",
"0.262"
],
[
"0.2932",
"0.2644",
"0.71",
"0.2644",
"0.71",
"0.2778",
"0.2932",
"0.2778"
],
[
"0.3245",
"0.2766",
"0.6807",
"0.2766",
"0.6807",
"0.2903",
"0.3245",
"0.2903"
],
[
"0.3545",
"0.2917",
"0.3708",
"0.2917",
"0.3708",
"0.2998",
"0.3545",
"0.2998"
],
[
"0.3652",
"0.2903",
"0.652",
"0.2903",
"0.652",
"0.3037",
"0.3652",
"0.3037"
],
[
"0.4558",
"0.3062",
"0.55",
"0.3062",
"0.55",
"0.3164",
"0.4558",
"0.3164"
],
[
"0.2656",
"0.3547",
"0.739",
"0.3547",
"0.739",
"0.3694",
"0.2656",
"0.3694"
],
[
"0.2646",
"0.3694",
"0.74",
"0.3694",
"0.74",
"0.3838",
"0.2646",
"0.3838"
],
[
"0.2646",
"0.3838",
"0.74",
"0.3838",
"0.74",
"0.3977",
"0.2646",
"0.3977"
],
[
"0.2646",
"0.3977",
"0.7344",
"0.3977",
"0.7344",
"0.411",
"0.2646",
"0.411"
],
[
"0.2646",
"0.411",
"0.738",
"0.411",
"0.738",
"0.4248",
"0.2646",
"0.4248"
],
[
"0.2656",
"0.4268",
"0.738",
"0.4268",
"0.738",
"0.4375",
"0.2656",
"0.4375"
],
[
"0.2646",
"0.4395",
"0.738",
"0.4395",
"0.738",
"0.4526",
"0.2646",
"0.4526"
],
[
"0.2656",
"0.4539",
"0.738",
"0.4539",
"0.738",
"0.464",
"0.2656",
"0.464"
],
[
"0.2656",
"0.4673",
"0.739",
"0.4673",
"0.739",
"0.481",
"0.2656",
"0.481"
],
[
"0.2656",
"0.481",
"0.739",
"0.481",
"0.739",
"0.4944",
"0.2656",
"0.4944"
],
[
"0.2632",
"0.4944",
"0.4617",
"0.4944",
"0.4617",
"0.5083",
"0.2632",
"0.5083"
],
[
"0.2646",
"0.5215",
"0.6445",
"0.5215",
"0.6445",
"0.535",
"0.2646",
"0.535"
],
[
"0.2157",
"0.561",
"0.3782",
"0.5625",
"0.3782",
"0.5776",
"0.2157",
"0.577"
],
[
"0.2197",
"0.5933",
"0.5312",
"0.5933",
"0.5312",
"0.607",
"0.2197",
"0.607"
],
[
"0.5337",
"0.5923",
"0.544",
"0.601",
"0.5337",
"0.6094",
"0.523",
"0.601"
],
[
"0.2181",
"0.608",
"0.5415",
"0.608",
"0.5415",
"0.622",
"0.2181",
"0.622"
],
[
"0.5786",
"0.6094",
"0.5923",
"0.61",
"0.5923",
"0.6143",
"0.5786",
"0.613"
],
[
"0.2173",
"0.6235",
"0.5435",
"0.625",
"0.5435",
"0.638",
"0.2173",
"0.637"
],
[
"0.5703",
"0.6235",
"0.584",
"0.6235",
"0.584",
"0.627",
"0.5703",
"0.627"
],
[
"0.585",
"0.6235",
"0.603",
"0.6235",
"0.603",
"0.627",
"0.585",
"0.627"
],
[
"0.5605",
"0.632",
"0.5703",
"0.632",
"0.5703",
"0.636",
"0.5605",
"0.636"
],
[
"0.2181",
"0.638",
"0.5415",
"0.638",
"0.5415",
"0.6514",
"0.2181",
"0.6514"
],
[
"0.626",
"0.6396",
"0.6396",
"0.6396",
"0.6396",
"0.648",
"0.626",
"0.648"
],
[
"0.5825",
"0.6465",
"0.589",
"0.6416",
"0.5933",
"0.6445",
"0.5864",
"0.6494"
],
[
"0.2197",
"0.654",
"0.5435",
"0.654",
"0.5435",
"0.6675",
"0.2197",
"0.6675"
],
[
"0.5664",
"0.661",
"0.607",
"0.661",
"0.607",
"0.6753",
"0.5664",
"0.6753"
],
[
"0.7427",
"0.662",
"0.7793",
"0.662",
"0.7793",
"0.6724",
"0.7427",
"0.6724"
],
[
"0.2197",
"0.6694",
"0.5435",
"0.6694",
"0.5435",
"0.683",
"0.2197",
"0.683"
],
[
"0.6235",
"0.6694",
"0.6357",
"0.6694",
"0.6357",
"0.674",
"0.6235",
"0.674"
],
[
"0.564",
"0.6787",
"0.6113",
"0.6787",
"0.6113",
"0.69",
"0.564",
"0.69"
],
[
"0.2197",
"0.685",
"0.5415",
"0.685",
"0.5415",
"0.6987",
"0.2197",
"0.6987"
],
[
"0.2181",
"0.6987",
"0.5415",
"0.6987",
"0.5415",
"0.712",
"0.2181",
"0.712"
],
[
"0.572",
"0.6963",
"0.761",
"0.6978",
"0.761",
"0.712",
"0.572",
"0.711"
],
[
"0.2214",
"0.714",
"0.785",
"0.714",
"0.785",
"0.728",
"0.2214",
"0.728"
],
[
"0.2214",
"0.731",
"0.773",
"0.731",
"0.773",
"0.741",
"0.2214",
"0.741"
],
[
"0.2443",
"0.7456",
"0.7837",
"0.7456",
"0.7837",
"0.759",
"0.2443",
"0.759"
],
[
"0.2197",
"0.7603",
"0.786",
"0.7603",
"0.786",
"0.7734",
"0.2197",
"0.7734"
],
[
"0.2222",
"0.7764",
"0.7837",
"0.7764",
"0.7837",
"0.787",
"0.2222",
"0.787"
],
[
"0.2197",
"0.797",
"0.786",
"0.799",
"0.786",
"0.813",
"0.2197",
"0.811"
],
[
"0.2345",
"0.813",
"0.509",
"0.813",
"0.509",
"0.827",
"0.2345",
"0.827"
],
[
"0.2173",
"0.8247",
"0.6323",
"0.8276",
"0.6323",
"0.8423",
"0.2173",
"0.839"
]
] | [
"Assessing Classifier Fairness with Collider Bias",
"",
"Zhenlong Xu tl, Ziqi Xu t[0000-0003-1748-5801], Jixue Liu[0000-0002-0794-0404]",
"Debo Cheng|000000020383-1462],Jiuyong Li[000-0002-9023-1878],",
"Lin Liui0000-0003-2843-5738], and Ke Wang?",
"1 University of South Australia, Adelaide, Australia f",
"Zhenlong.Xu,Ziqi.Xu,Debo.Cheng } @mymail.unisa.edu.au,",
"{ Jixue.Liu,Jiuyong.Li,Lin.Liu } @unisa.edu.au",
"2",
"Simon Fraser University, Burnaby, Canada",
"[email protected]",
"Abstract. The increasing application of machine learning techniques",
"in everyday decision-making processes has brought concerns about the",
"fairness of algorithmic decision-making. This paper concerns the problem",
"of collider bias which produces spurious associations in fairness assess",
"ment and develops theorems to guide fairness assessment avoiding the",
"colliderbias.Weconsiderareal-world application ofauditingatrained",
"classifier by an audit agency. We propose an unbiased assessment algo-",
"rithmbyutilising thedeveloped theoremstoreducecolliderbiasesinthe",
"assessment. Experiments and simulations show the proposed algorithm",
"reduces collider biases significantly in the assessment and is promising",
"in auditing trained classifiers.",
"Keywords: FairnessCollider biasCausal inference.",
"1Introduction",
"随着机器学习在各种应用中的广泛使用，",
"T",
"们越来越担心决策算法的公平性，例如工作招",
"",
"聘，信用评分和房屋贷款，因为歧视可能会无",
"",
"",
"Pri",
"意中引入机器学习模型。为了防止模型中的不",
"外",
"?",
"公平在社会中蔓延，独立机构需要审计技术来",
"888",
"Fair/Unfair",
"审计机器学习模型。图1显示了一个审计过",
"cas",
"Audit Agency",
"程。审计机构访问公司的模型，并有自己的审",
"计案例来评估模型的公平性。审计机构无权访",
"Fig.1. The process of audit",
"问培训数据集，但具有监管政策。在本文中，我们使用因果图来表示监管政策。",
"公司可能会使用监管政策中未指定的其他变量来构建模型，以提高预测准确性",
"情况测试已在美国使用。S。检测招聘中的歧视[1]，这是一种对照实验方",
"法，用于分析雇主对求职者特征的决定，如以下示例所示。成对的研究助理被派",
"去申请同一份工作，每对假装的申请人都有与工作相关的相同资格和经验，但",
"* Supported by Australian Research Council (DP200101210) and Natural Sciences and",
"Engineering Research Council of Canada.",
"t Ziqi Xu and Zhenlong Xu contributed equally to this paper."
] | {
"arxiv_id": "2010.03933",
"language": "zh",
"page_idx": 0
} |
||
2010.03933-zh-1 | 2010.03933-zh-1.jpg | 2 Z. Xu et al.
对他们的受保护变量有不同的值，如男性/女性或年轻/年老。如果具有不同受保
护价值的群体之间的有利决策不平等，则会检测到歧视
上述情况测试可以在审计过程中进行模拟，我们在本文中将其称为朴素情况测试 (NST)。我们将代表变量值相同的两个个体的两个输入输人到机器学习模型。如果模型提供了不同的决策，NST 会将该模型检测为歧视性的
NST 可能产生不正确的检测。我们用 Table 1.An example of incorrect
下面的例子来说明这一点。将公司用于确 detection by NST on a classifier.
<table>
<tbody>
<tr>
<th>$\overline{\mathrm{Race}}$</th>
<th>Edu</th>
<th>$\overline{\mathrm{Sub}}$</th>
<th>Predicted. Sa</th>
</tr>
<tr>
<td>$\overline{\mathrm{white}}$</td>
<td>$\overline{\mathrm{high}}$</td>
<td>$\overline{\mathrm{A}}$</td>
<td>>50k</td>
</tr>
<tr>
<td>black</td>
<td>high</td>
<td>$A$</td>
<td>$>50k$</td>
</tr>
<tr>
<td>white</td>
<td>high</td>
<td>$A$</td>
<td>$>50k$</td>
</tr>
<tr>
<td>black</td>
<td>high</td>
<td>B</td>
<td>$50k$</td>
</tr>
<tr>
<td>black</td>
<td>high</td>
<td>B</td>
<td>$50k$</td>
</tr>
<tr>
<td>white</td>
<td>high</td>
<td>B</td>
<td>$50k$</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td> </td>
<td>“fair”</td>
</tr>
<tr>
<td>white 手</td>
<td>$=f($ black , high, A</td>
</tr>
<tr>
<td>white</td>
<td>T black , high. B)</td>
</tr>
</tbody>
</table>
<FigureHere>
定员工工资的分类器$f($)视为$salary=$
$f(race,education,suburb)$ 。该模型的一些预测结果如表 1 所示。根据 NST, 黑人不受歧视，因为在相同的教育和郊区，白人和黑人都被预测有相同的工资。然而，郊区是一个与确定工资无关的变量。在不考虑郊区的情况下， 在相同的教育水平下，2/3 的白人的薪水高于50K,而只有1/3的黑人的薪水高于 50K。因此，黑人受到该模型的歧视
NST 的错误检测是由对撞机偏差引起的。我们使用在第 2 节中正式定义的因果图来解释对撞机偏差。上述示例中变量的因果关系使用图 2 中的因果图显示，其中有向边表示因果关系。郊区是对撞机，因为两个边缘在其上“碰撞”。以对撞机为条件，在两个变量之间形成关联，但它是虚假的E[ 5] 。在这个例子中，由于种族和工资之间 Fig. 2. The causal graph for
的因果关系，虚假的关联取消了关联，并隐藏了
真正的歧视。对撞机偏差与选择偏差有关[11] 。在分类器中，对变量进行调节等效于使用变量的值选择子群体。如果变量是碰撞器，则会导致每个子群体中的选择偏差。在本文中，我们将这种偏差称为对撞机偏差
没有现有的 NST 替代方法来审核分类器。在相关工作中需要审查的大多数方法都需要访问训练数据集，并且会受到碰撞偏差的影响。基于因果的情况测试 (CST) [30] 不会受到碰撞偏差的影响，但也需要访问训练数据集。审计机构使用的审计案例只是少数不代表总体的个案。收集具有代表性的人口样本需要大量资源。因此，基于数据的审计方法不适用。我们在本文中做出了以下贡献- We study collider bias in fairness assessment and present theorems to avoid
collider bias. Our theoretical results give a principled guidance on which variables can be used for fairness assessment and also for building fair classifiers.
- We investigate the problem of auditing machine learning models and propose
an Unbiased Situation Test (UST) algorithm for auditing without accessing training data or an unbiased sample of the population. Experiments show that UST can effectively reduce collider bias.
2 Background
我们介绍了因果推理的必要背景。我们用大写字母表示变量，用粗体大写字母
表示变量集。变量的值使用小写字母表示 | [
[
"0.2173",
"0.11615",
"0.2524",
"0.11615",
"0.2524",
"0.1288",
"0.2173",
"0.1288"
],
[
"0.2712",
"0.11743",
"0.3489",
"0.11743",
"0.3489",
"0.1276",
"0.2712",
"0.1276"
],
[
"0.2181",
"0.1483",
"0.785",
"0.1483",
"0.785",
"0.1622",
"0.2181",
"0.1622"
],
[
"0.2197",
"0.1654",
"0.609",
"0.1654",
"0.609",
"0.1755",
"0.2197",
"0.1755"
],
[
"0.246",
"0.1768",
"0.7837",
"0.1787",
"0.7837",
"0.1925",
"0.246",
"0.1901"
],
[
"0.2197",
"0.1932",
"0.785",
"0.1932",
"0.785",
"0.207",
"0.2197",
"0.207"
],
[
"0.2197",
"0.209",
"0.6978",
"0.209",
"0.6978",
"0.2229",
"0.2197",
"0.2229"
],
[
"0.2418",
"0.2216",
"0.785",
"0.226",
"0.785",
"0.2405",
"0.2418",
"0.2361"
],
[
"0.2214",
"0.2393",
"0.55",
"0.2393",
"0.55",
"0.2524",
"0.2214",
"0.2524"
],
[
"0.546",
"0.2424",
"0.5664",
"0.2424",
"0.5664",
"0.25",
"0.546",
"0.25"
],
[
"0.5596",
"0.2393",
"0.777",
"0.2405",
"0.777",
"0.254",
"0.5596",
"0.2524"
],
[
"0.2197",
"0.254",
"0.535",
"0.254",
"0.535",
"0.2676",
"0.2197",
"0.2676"
],
[
"0.5312",
"0.2588",
"0.544",
"0.2588",
"0.544",
"0.2632",
"0.5312",
"0.2632"
],
[
"0.591",
"0.2588",
"0.759",
"0.2588",
"0.759",
"0.2695",
"0.591",
"0.2695"
],
[
"0.2181",
"0.2683",
"0.5483",
"0.2676",
"0.5483",
"0.2822",
"0.2181",
"0.2827"
],
[
"0.591",
"0.2676",
"0.653",
"0.2695",
"0.653",
"0.281",
"0.591",
"0.279"
],
[
"0.6543",
"0.2715",
"0.671",
"0.2715",
"0.671",
"0.279",
"0.6543",
"0.279"
],
[
"0.705",
"0.2695",
"0.7363",
"0.2695",
"0.7363",
"0.281",
"0.705",
"0.281"
],
[
"0.591",
"0.2798",
"0.653",
"0.2798",
"0.653",
"0.2903",
"0.591",
"0.2903"
],
[
"0.6543",
"0.281",
"0.6724",
"0.281",
"0.6724",
"0.289",
"0.6543",
"0.289"
],
[
"0.705",
"0.279",
"0.7363",
"0.279",
"0.7363",
"0.2903",
"0.705",
"0.2903"
],
[
"0.2181",
"0.2854",
"0.546",
"0.2854",
"0.546",
"0.2954",
"0.2181",
"0.2954"
],
[
"0.6562",
"0.2874",
"0.6724",
"0.2903",
"0.669",
"0.3005",
"0.653",
"0.2974"
],
[
"0.705",
"0.2874",
"0.7363",
"0.289",
"0.7344",
"0.3018",
"0.7036",
"0.2998"
],
[
"0.6978",
"0.2966",
"0.7427",
"0.2966",
"0.7427",
"0.324",
"0.6978",
"0.324"
],
[
"0.5923",
"0.2903",
"0.653",
"0.2903",
"0.653",
"0.3005",
"0.5923",
"0.3005"
],
[
"0.2214",
"0.2998",
"0.5483",
"0.2998",
"0.5483",
"0.3132",
"0.2214",
"0.3132"
],
[
"0.5923",
"0.2998",
"0.652",
"0.2998",
"0.652",
"0.31",
"0.5923",
"0.31"
],
[
"0.591",
"0.307",
"0.652",
"0.3093",
"0.652",
"0.3208",
"0.591",
"0.318"
],
[
"0.6543",
"0.3018",
"0.6724",
"0.3018",
"0.6724",
"0.3093",
"0.6543",
"0.3093"
],
[
"0.2181",
"0.3132",
"0.546",
"0.3132",
"0.546",
"0.327",
"0.2181",
"0.327"
],
[
"0.652",
"0.3125",
"0.671",
"0.3093",
"0.6724",
"0.3164",
"0.6543",
"0.3193"
],
[
"0.591",
"0.3208",
"0.653",
"0.3208",
"0.653",
"0.3308",
"0.591",
"0.3308"
],
[
"0.6543",
"0.3208",
"0.6724",
"0.3208",
"0.6724",
"0.3276",
"0.6543",
"0.3276"
],
[
"0.7017",
"0.3164",
"0.74",
"0.3208",
"0.738",
"0.3308",
"0.701",
"0.3276"
],
[
"0.2197",
"0.33",
"0.546",
"0.33",
"0.546",
"0.3435",
"0.2197",
"0.3435"
],
[
"0.6436",
"0.3333",
"0.721",
"0.3333",
"0.721",
"0.3435",
"0.6436",
"0.3435"
],
[
"0.2214",
"0.3447",
"0.5483",
"0.3447",
"0.5483",
"0.358",
"0.2214",
"0.358"
],
[
"0.564",
"0.3452",
"0.8013",
"0.3452",
"0.8013",
"0.356",
"0.564",
"0.356"
],
[
"0.2181",
"0.358",
"0.5483",
"0.358",
"0.5483",
"0.3718",
"0.2181",
"0.3718"
],
[
"0.562",
"0.356",
"0.8",
"0.356",
"0.8",
"0.3662",
"0.562",
"0.3662"
],
[
"0.2197",
"0.375",
"0.415",
"0.375",
"0.415",
"0.3882",
"0.2197",
"0.3882"
],
[
"0.2418",
"0.3882",
"0.572",
"0.3894",
"0.572",
"0.4028",
"0.2418",
"0.4016"
],
[
"0.6543",
"0.3958",
"0.714",
"0.3958",
"0.714",
"0.406",
"0.6543",
"0.406"
],
[
"0.2197",
"0.4048",
"0.572",
"0.4048",
"0.572",
"0.4185",
"0.2197",
"0.4185"
],
[
"0.653",
"0.406",
"0.714",
"0.406",
"0.714",
"0.4172",
"0.653",
"0.4172"
],
[
"0.2197",
"0.4216",
"0.569",
"0.4216",
"0.569",
"0.432",
"0.2197",
"0.432"
],
[
"0.6006",
"0.4255",
"0.625",
"0.4255",
"0.625",
"0.4343",
"0.6006",
"0.4343"
],
[
"0.7417",
"0.4204",
"0.774",
"0.4248",
"0.7715",
"0.4382",
"0.739",
"0.4343"
],
[
"0.2197",
"0.435",
"0.5703",
"0.435",
"0.5703",
"0.449",
"0.2197",
"0.449"
],
[
"0.2197",
"0.4507",
"0.572",
"0.4507",
"0.572",
"0.464",
"0.2197",
"0.464"
],
[
"0.671",
"0.4539",
"0.7017",
"0.4539",
"0.7017",
"0.4622",
"0.671",
"0.4622"
],
[
"0.2197",
"0.4653",
"0.572",
"0.4653",
"0.572",
"0.4792",
"0.2197",
"0.4792"
],
[
"0.2197",
"0.481",
"0.572",
"0.481",
"0.572",
"0.4944",
"0.2197",
"0.4944"
],
[
"0.584",
"0.476",
"0.786",
"0.476",
"0.786",
"0.4893",
"0.584",
"0.4893"
],
[
"0.5825",
"0.4912",
"0.7134",
"0.4912",
"0.7134",
"0.502",
"0.5825",
"0.502"
],
[
"0.2197",
"0.4976",
"0.568",
"0.4976",
"0.568",
"0.5083",
"0.2197",
"0.5083"
],
[
"0.2197",
"0.5103",
"0.785",
"0.5103",
"0.785",
"0.5234",
"0.2197",
"0.5234"
],
[
"0.2197",
"0.526",
"0.786",
"0.526",
"0.786",
"0.539",
"0.2197",
"0.539"
],
[
"0.2214",
"0.5425",
"0.626",
"0.5425",
"0.626",
"0.553",
"0.2214",
"0.553"
],
[
"0.2443",
"0.556",
"0.785",
"0.556",
"0.785",
"0.5693",
"0.2443",
"0.5693"
],
[
"0.2222",
"0.5728",
"0.7837",
"0.5728",
"0.7837",
"0.5835",
"0.2222",
"0.5835"
],
[
"0.2197",
"0.586",
"0.785",
"0.586",
"0.785",
"0.5996",
"0.2197",
"0.5996"
],
[
"0.2214",
"0.603",
"0.7817",
"0.603",
"0.7817",
"0.613",
"0.2214",
"0.613"
],
[
"0.2214",
"0.6187",
"0.773",
"0.6187",
"0.773",
"0.629",
"0.2214",
"0.629"
],
[
"0.232",
"0.64",
"0.786",
"0.64",
"0.786",
"0.654",
"0.232",
"0.654"
],
[
"0.2443",
"0.655",
"0.785",
"0.655",
"0.785",
"0.6685",
"0.2443",
"0.6685"
],
[
"0.2484",
"0.6724",
"0.7817",
"0.6724",
"0.7817",
"0.683",
"0.2484",
"0.683"
],
[
"0.2484",
"0.6875",
"0.2917",
"0.6875",
"0.2917",
"0.6987",
"0.2484",
"0.6987"
],
[
"0.2361",
"0.704",
"0.2443",
"0.704",
"0.2443",
"0.709",
"0.2361",
"0.709"
],
[
"0.246",
"0.7007",
"0.785",
"0.7007",
"0.785",
"0.714",
"0.246",
"0.714"
],
[
"0.246",
"0.7153",
"0.786",
"0.7163",
"0.786",
"0.73",
"0.246",
"0.729"
],
[
"0.246",
"0.73",
"0.785",
"0.731",
"0.785",
"0.7456",
"0.246",
"0.744"
],
[
"0.2484",
"0.7476",
"0.569",
"0.7476",
"0.569",
"0.7583",
"0.2484",
"0.7583"
],
[
"0.2181",
"0.7827",
"0.3733",
"0.7827",
"0.3733",
"0.7974",
"0.2181",
"0.7974"
],
[
"0.2214",
"0.813",
"0.7837",
"0.813",
"0.7837",
"0.8237",
"0.2214",
"0.8237"
],
[
"0.2181",
"0.827",
"0.511",
"0.827",
"0.511",
"0.8403",
"0.2181",
"0.8403"
]
] | [
"2",
"Z.Xuetal.",
"对他们的受保护变量有不同的值，如男性/女性或年轻/年老。如果具有不同受保",
"护价值的群体之间的有利决策不平等，则会检测到歧视",
"上述情况测试可以在审计过程中进行模拟，我们在本文中将其称为朴素情况",
"测试（NST）。我们将代表变量值相同的两个个体的两个输人输人到机器学习模",
"型。如果模型提供了不同的决策，NST会将该模型检测为歧视性的",
"NST可能产生不正确的检测。我们用Table1.Anexampleof incorrect",
"下面的例子来说明这一点。将公司用于确",
"",
"detection by NST on a classifier",
"定员工工资的分类器fO视为salary",
"=",
"Race Edu Sub Predicted.Sal",
"f(race,education，suburb）。该模型的一些预",
"white high",
"A",
">50k",
"black high",
"A",
">50k",
"测结果如表1所示。根据NST，黑人不受歧",
"A",
">50k",
"BOK",
"white high",
"视，因为在相同的教育和郊区，白人和黑人都",
"black high",
"black high",
"B",
"被预测有相同的工资。然而，郊区是一个与确",
"B",
"white high",
"B",
"≥50k",
"定工资无关的变量。在不考虑郊区的情况下",
"NST→\"fair",
"在相同的教育水平下，2/3的白人的薪水高于",
"f (white,high, A)=f(black, high, A)",
"50K，而只有1/3的黑人的薪水高于50K。因",
"f(white,high,B)=f (black,high,B)",
"此，黑人受到该模型的歧视",
"NST的错误检测是由对撞机偏差引起的。我",
"Education",
"们使用在第2节中正式定义的因果图来解释对撞",
"Level",
"机偏差。上述示例中变量的因果关系使用图2中",
"Race",
"Salary",
"的因果图显示，其中有向边表示因果关系。郊区",
"是对撞机，因为两个边缘在其上“碰撞”。以对撞",
"Suburb",
"机为条件，在两个变量之间形成关联，但它是虚",
"假的E[5]。在这个例子中，由于种族和工资之间",
"Fig.2. The causal graph for",
"theaboveexample.",
"的因果关系，虚假的关联取消了关联，并隐藏了",
"真正的歧视。对撞机偏差与选择偏差有关[11]。在分类器中，对变量进行调节",
"等效于使用变量的值选择子群体。如果变量是碰撞器，则会导致每个子群体中",
"的选择偏差。在本文中，我们将这种偏差称为对撞机偏差",
"没有现有的NST替代方法来审核分类器。在相关工作中需要审查的大多数",
"方法都需要访问训练数据集，并且会受到碰撞偏差的影响。基于因果的情况测",
"试（CST）[30]不会受到碰撞偏差的影响，但也需要访问训练数据集。审计机构",
"使用的审计案例只是少数不代表总体的个案。收集具有代表性的人口样本需要",
"天量资源。因此，基于数据的审计方法不适用。我们在本文中做出了以下贡献",
"- We study collider bias in fairness assessment and present theorems to avoid",
"collider bias. Our theoretical results give a principled guidance on which",
"variables can be used forfairness assessment and alsoforbuilding fair clas",
"sifiers.",
"一",
"We investigate the problem of auditing machine learning models and propose",
"an Unbiased Situation Test (UST)algorithm for auditing without accessing",
"training data or an unbiased sample of the population. Experiments show",
"that UST can effectivelyreduce colliderbias.",
"2Background",
"我们介绍了因果推理的必要背景。我们用大写字母表示变量，用粗体大写字母",
"表示变量集。变量的值使用小写字母表示"
] | {
"arxiv_id": "2010.03933",
"language": "zh",
"page_idx": 1
} |
||
2010.03933-zh-2 | 2010.03933-zh-2.jpg | Assessing Classifer Fairness with Collider Bias 3
令$\mathcal{G}=(V,E)$是一个图，其中$V=\{V_1,\ldots,V_p\}$是节点集，$E$是节点之间的边集，i。e。$E\subseteq V\times V$。路径$\pi$是一系列不同的节点，因此$\mathcal{G}$中的每对连续节点都是相邻的。如果沿路径的所有边都是有向边，则路径$\pi$是有向路径。$(V_i,V_j)$之间的路径是相对于$X_i$的后门路径，如果它有一个指向$V_i$的箭头。给定路径$\pi$,如果有两条边像$V_i\to V_k\leftarrow V_j$一样人射，则$V_k$是$\pi$上的碰撞节点。在$\mathcal{G}$中，如果存在$V_i\to V_j$ ,则$V_i$是$V_j$的父级，我们使用$Pa(V_j)$表示$V_j$ 的所有父级的集合。在有向路径 $\pi$ 中，如果所有箭头都指向$V_j$ ,则$V_i$ 是$V_j$ $\ddot{\text{的祖先，而}}V_j$ 是$V_i$ 的后代
DAG(有向无环图)是没有有向循环的有向图。在以下假设的情况下，DAG
链接到分配
$\textbf{Definition 1}($Markov condition $[ 19] ) . \textit{Given a DAG G}= ( \boldsymbol{V}, \boldsymbol{E}) \textit{and P}( \boldsymbol{V})$ $, thejointprobability$ distribution $ofV, \mathcal{G}$ satisfes the Markov condition $if$ for $\forall V_{i}\in \boldsymbol{V}$ , $V_{i}$ $is$ probabilistically independent $of$ all $non- descendants$ $ofV_{i}$ , given $the~parents~of~V_{i}~.$
当马尔可夫条件成立时，$P(\boldsymbol V)$可以分解为：$P(\boldsymbol V)=\prod_iP(V_i\mid Pa(V_i))$
$\begin{array}{l}{\textbf{Definition 2}}&{(\textbf{Faithfulness [24]).}A\:DAG\:\mathcal{G}=(\boldsymbol{V},\boldsymbol{E})\:is\:faithful\:to\:P(\boldsymbol{V})\:iff}\\{every\:independence\:presenting\:in\:P(\boldsymbol{V})\:is\:entailed\:by\:\mathcal{G}\:which\:fulflls\:the\:Markov}\end{array}$ $condition. A$ distribution $P( \boldsymbol{V})$ $is$ faithful $to$ $a$ $DAG$ $\mathcal{G}$ iff there exists $DAG$ $\mathcal{G}$ $which~is~faithful~to~P(V)~.$
通过上述两个假设，我们可以使用以下定义的$d$分离标准从 DAG 读取$P(V)$
中变量之间的独立性。
Definition 3( $d$ -separation $[ 19] ) . \textit{ A path }\pi$ in $a\textit{ DAG is said to be d - }$ separated $( or$ $blocked)$ $by$ $a$ set $of$ nodes $\boldsymbol{Z}$ iff (1) $\pi$ contains $a$ chain $V_{i}\rightarrow V_{k}\rightarrow V_{j}$ $and\textit{ a fork V}_{i}\leftarrow V_{k}\rightarrow V_{j}\textit{ node such that the middle node V}_{k}\textit{is in Z}, or( 2)$ $contains~a~collider~V_{k}~such~that~V_{k}~is~not~in~\boldsymbol{Z}~and~no~descendant~ofV_{k}~is~in~\boldsymbol{Z}$
为了对 DAG 进行因果推断，我们做出以下假设
$\textbf{Definition 4 ( Causal sufficiency [ 24] ) . A data set satisfies causal sufficiency}$ $iffor\textit{ every pair of variables }( V_i, V_j) \textit{ in V , all their common causes are also in}$ $\ddot{V}.$
对于 DAG, 如果我们将节点的父级解释为其直接原因，则 DAG 称为因果 DAG。当满足因果充分性、忠实性和马尔可夫条件的假设时，我们可以从数据中学习因果 DAG
强制变量取值的干预可以由 do 运算符表示。例如，$do(X=1)$表示$X$被干预以取值 1。$P(y\mid do(X=1))$是一种介人概率。让我们在理想的实验中了解$do$
$\begin{array}{l}\textbf{Definition 5}&(\textbf{Direct effect [19]).The direct effect ofX onY is P(y\mid do(X=\\x),do(\boldsymbol{V}\backslash xY=\boldsymbol{v}))\: where\:\boldsymbol{V}\rangle{xY}\:means\:all\:other\:variables\:except\:X\:and\:Y\:.}\end{array}$ | [
[
"0.4167",
"0.11615",
"0.759",
"0.11615",
"0.759",
"0.13",
"0.4167",
"0.13"
],
[
"0.7715",
"0.1187",
"0.786",
"0.1187",
"0.786",
"0.1276",
"0.7715",
"0.1276"
],
[
"0.2443",
"0.1483",
"0.785",
"0.1483",
"0.785",
"0.1622",
"0.2443",
"0.1622"
],
[
"0.2214",
"0.1635",
"0.785",
"0.1635",
"0.785",
"0.1768",
"0.2214",
"0.1768"
],
[
"0.2197",
"0.1768",
"0.785",
"0.178",
"0.785",
"0.1925",
"0.2197",
"0.1913"
],
[
"0.2214",
"0.1932",
"0.785",
"0.1932",
"0.785",
"0.207",
"0.2214",
"0.207"
],
[
"0.2197",
"0.209",
"0.7817",
"0.209",
"0.7817",
"0.2229",
"0.2197",
"0.2229"
],
[
"0.2181",
"0.2229",
"0.7837",
"0.2235",
"0.7837",
"0.238",
"0.2181",
"0.2374"
],
[
"0.2197",
"0.2393",
"0.7837",
"0.2393",
"0.7837",
"0.2524",
"0.2197",
"0.2524"
],
[
"0.2197",
"0.254",
"0.4192",
"0.254",
"0.4192",
"0.2676",
"0.2197",
"0.2676"
],
[
"0.2402",
"0.2664",
"0.786",
"0.2676",
"0.786",
"0.2854",
"0.2402",
"0.2842"
],
[
"0.2214",
"0.286",
"0.3022",
"0.286",
"0.3022",
"0.2974",
"0.2214",
"0.2974"
],
[
"0.2181",
"0.315",
"0.7837",
"0.315",
"0.7837",
"0.3289",
"0.2181",
"0.3289"
],
[
"0.2222",
"0.33",
"0.7837",
"0.33",
"0.7837",
"0.3447",
"0.2222",
"0.3447"
],
[
"0.2181",
"0.3447",
"0.786",
"0.3452",
"0.786",
"0.3599",
"0.2181",
"0.3591"
],
[
"0.2197",
"0.361",
"0.353",
"0.361",
"0.353",
"0.375",
"0.2197",
"0.375"
],
[
"0.246",
"0.3901",
"0.7534",
"0.3901",
"0.7534",
"0.404",
"0.246",
"0.404"
],
[
"0.2197",
"0.4204",
"0.7876",
"0.4204",
"0.7876",
"0.4343",
"0.2197",
"0.4343"
],
[
"0.2214",
"0.435",
"0.785",
"0.435",
"0.785",
"0.449",
"0.2214",
"0.449"
],
[
"0.2214",
"0.4507",
"0.785",
"0.4507",
"0.785",
"0.464",
"0.2214",
"0.464"
],
[
"0.2197",
"0.4666",
"0.4102",
"0.4666",
"0.4102",
"0.4797",
"0.2197",
"0.4797"
],
[
"0.2197",
"0.4956",
"0.7837",
"0.4956",
"0.7837",
"0.509",
"0.2197",
"0.509"
],
[
"0.2197",
"0.5103",
"0.3743",
"0.5103",
"0.3743",
"0.5234",
"0.2197",
"0.5234"
],
[
"0.2181",
"0.5405",
"0.7817",
"0.5405",
"0.7817",
"0.5537",
"0.2181",
"0.5537"
],
[
"0.2197",
"0.5547",
"0.785",
"0.5547",
"0.785",
"0.5693",
"0.2197",
"0.5693"
],
[
"0.2214",
"0.571",
"0.7837",
"0.571",
"0.7837",
"0.584",
"0.2214",
"0.584"
],
[
"0.2214",
"0.587",
"0.7837",
"0.587",
"0.7837",
"0.5977",
"0.2214",
"0.5977"
],
[
"0.2443",
"0.63",
"0.584",
"0.63",
"0.584",
"0.6445",
"0.2443",
"0.6445"
],
[
"0.2173",
"0.657",
"0.7876",
"0.659",
"0.7876",
"0.677",
"0.2173",
"0.675"
],
[
"0.2197",
"0.6753",
"0.7837",
"0.6753",
"0.7837",
"0.69",
"0.2197",
"0.69"
],
[
"0.2181",
"0.6914",
"0.2484",
"0.6914",
"0.2484",
"0.7026",
"0.2181",
"0.7026"
],
[
"0.2173",
"0.7183",
"0.7837",
"0.7197",
"0.7837",
"0.7344",
"0.2173",
"0.733"
],
[
"0.2214",
"0.735",
"0.7837",
"0.735",
"0.7837",
"0.7485",
"0.2214",
"0.7485"
],
[
"0.2214",
"0.752",
"0.2942",
"0.752",
"0.2942",
"0.763",
"0.2214",
"0.763"
],
[
"0.2443",
"0.7666",
"0.785",
"0.7666",
"0.785",
"0.78",
"0.2443",
"0.78"
],
[
"0.2214",
"0.7827",
"0.7837",
"0.7827",
"0.7837",
"0.7935",
"0.2214",
"0.7935"
],
[
"0.2197",
"0.811",
"0.7817",
"0.811",
"0.7817",
"0.8257",
"0.2197",
"0.8257"
],
[
"0.2181",
"0.827",
"0.7686",
"0.827",
"0.7686",
"0.8413",
"0.2181",
"0.8413"
]
] | [
"Assessing Classifier Fairness with Collider Bias",
"3",
"令G=（V,E)是一个图，其中V={Vi，...V}是节点集，E是节点之间",
"的边集，i。e。ECV×V。路径元是一系列不同的节点，因此中的每对",
"连续节点都是相邻的。如果沿路径的所有边都是有向边，则路径元是有向路径。",
"(V,V）之间的路径是相对于X，的后门路径，如果它有一个指向V的箭头。给",
"定路径元，如果有两条边像Vi→Vk←V,一样入射，则Vk是元上的碰撞节",
"点。在g中，如果存在V→Vi，则V是V,的父级，我们使用Pa(V）表示",
"V,的所有父级的集合。在有向路径π中，如果所有箭头都指向Vi，则V.是V",
"的祖先，而Vi是V的后代",
"DAG（有向无环图）是没有有向循环的有向图。在以下假设的情况下，DAG",
"链接到分配",
"Definition 1 (Markov condition [19]). Given a DAG G = (V,E) and P(V)",
"the joint probability distribution of V , G satisfies the Markou condition if for",
"VV,E V, V is probabilistically independent of all non-descendants of Vi,given",
"the parents of Vi ..",
"当马尔可夫条件成立时，P(V）可以分解为：P(V)=II,P(VIPa(Va))",
"Definition 2 (Faithfulness [24]l). A DAG g = (V,E) is faithful to P(V) iff",
"every independence presenting in P(V) is entailed by G which fulfills the Markou",
"condition. A distribution P(V) is faithful to a DAG g iff there erists DAG g",
"whichis faithful toP(V):",
"通过上述两个假设，我们可以使用以下定义的d分离标准从DAG读取P(V）",
"中变量之间的独立性。",
"Definition 3 ( d -separation [19]). A path π in a DAG is said to be d",
"separated (or blocked) by a set of nodes Z iff (1) π contains a chain V; -→ Vh → V;",
"and a fork Vi Vk -→ V, node such that the middle node Vh is in Z, or (2) ",
"contains a colliderVk suchthat Vkis not in Z and no descendant ofVk is inZ",
"为了对DAG进行因果推断，我们做出以下假设",
"Definition 4 (Causal sufficiency [24]). A data set satisfies causal suffciency",
"if for every pair of variables (Vi, Ve) in V , all their common causes are also in",
"V.",
"对于DAG，如果我们将节点的父级解释为其直接原因，则DAG称为因果DAG。",
"当满足因果充分性、忠实性和马尔可夫条件的假设时，我们可以从数据中学习",
"因果DAG",
"强制变量取值的干预可以由do运算符表示。例如，do(X=1）表示X被干",
"预以取值1。P(y|do(X=1)）是一种介入概率。让我们在理想的实验中了解dc",
"Definition 5 (Direct effect [19]). The direct effect of X on Y is P(y I do(X =",
"r), do(Vixy = v)) where Vxy means all other variables ercept X and Y ."
] | {
"arxiv_id": "2010.03933",
"language": "zh",
"page_idx": 2
} |
||
2010.03933-zh-3 | 2010.03933-zh-3.jpg | 4 Z. Xu et al.
为了研究$Y$上$X$ 之间的关系，在理想实验中控制了所有其他变量。要使用因果 DAG 推断介人概率 (通过将其简化为正常条件概率),必须使用以下$da$ -演算[E[19] 规则集。它们被用在我们定理的证明中
对于 DAG $\mathcal{G}$和$\mathcal{G}$中的节点$X$的子集，$\mathcal{G}_{\overline{X}}$表示通过从$\mathcal{G}$中删除指向$X$中的节点的所有箭头而获得的 DAG,而$G_\mathbf{X}$表示通过从$G$中删除从$X$中的节点$\begin{array}{l}{\mathbf{Theorem~1~(Rules~of~do~-Calculus~[19]).~Let~X,Y,Z,W~be~arbitrary~dis-}}\\{joint~sets~of~variables~in~a~causal~DAG~G~.~The~following~rules~hold,~where}\end{array}$ $\boldsymbol{x},\boldsymbol{y},\boldsymbol{z},\boldsymbol{w}$ are the shorthands of $\boldsymbol X=\boldsymbol{x},\boldsymbol{Y}=\boldsymbol{y},\boldsymbol{Z}=\boldsymbol{z}$ and W$= \boldsymbol{w}\textit{ respectively}.$
$Rule~1.~(Insertion/deletion~of~observations):$
$P( \boldsymbol{y}\mid do( \boldsymbol{x}) , \boldsymbol{z}, \boldsymbol{w}) = P( \boldsymbol{y}\mid do( \boldsymbol{x}) , \boldsymbol{w})$ , $if\left ( \boldsymbol{Y}\perp \boldsymbol{Z}\mid \boldsymbol{X}, \boldsymbol{W}\right )$ $in$ $\mathcal{G} _{{\overline {\boldsymbol{X}}}}$ .
$Rule~2.~(Action/observation~exchange):$
$P( \boldsymbol{y}\mid do( \boldsymbol{x}) , do( \boldsymbol{z}) , \boldsymbol{w}) = P( \boldsymbol{y}\mid do( \boldsymbol{x}) , \boldsymbol{z}, \boldsymbol{w})$ $, if\left ( \boldsymbol{Y}\perp \boldsymbol{Z}\mid \boldsymbol{X}, \boldsymbol{W}\right )$ $in$ $\mathcal{G} _{\overline {\mathbf{X} }\underline {\mathbf{Z} }}$ .
$Rule~3.~(Insertion/deletion~of~actions):$
$P( \boldsymbol{y}\mid do( \boldsymbol{x}) , do( \boldsymbol{z}) , \boldsymbol{w}) = P( \boldsymbol{y}\mid do( \boldsymbol{x}) , \boldsymbol{w})$ $, if\left ( \boldsymbol{Y}\perp \boldsymbol{Z}\mid \boldsymbol{X}, \boldsymbol{W}\right )$ $in$ $\mathcal{G} _{\overline {\boldsymbol{X}\boldsymbol{Z}( \boldsymbol{W}) }}$ , $where\boldsymbol{Z}( \boldsymbol{W}) \textit{ is the nodes in Z that are not ancestors of any node in W in G}_{\overline {\mathbf{X} }}$ .
# 3 Problem Definition
公司/组织已根据训练数据集构建了分类器 (预测模型),该数据集包含二进制受保护变量$A$,二进制决策结果$Y$以及$Y,X$的一组相关变量，因为独立于XMATHX 的预测不是 XMATH。某机构希望使用某些案例对模型进行审计。我们对审计做出以下假设
$\textbf{Assumption 1}\textit{1. The regulatory policy has specified the causal relationships}$
$among\textit{ the factors and Y , and uses a causal DAG to indicate. The factors}$
$are\textit{ ancestral variables of Y including all direct causes of Y . }$
$2. \textit{ The audit agency has no access to the model training data or an unbiased}$
$sample\textit{ of the population. The agency however has access to the distributional}$
$statistics\textit{ from some sources, such as government census data. }$
$3. \textit{ The company or organisation has used all the legitimate factors to comply}$
$withtheregulatorypolicy.However,someothervariablesarealsousedby$
$the\textit{ model to enhance the prediction performance. }$
在定理开发中，我们假设存在一个与监管政策一致的 DAG。在该算法中，我
们不需要完整的 DAG, 而是需要$Y$的祖先变量和$Y$的后代节点中的碰撞器
我们首先定义审核的标准。我们使用受控直接效应(CDE)[20]来衡量公平
性。CDE 从 do 上的条件 (prote
cted)扩展到 do 上的条件 (所有其他变量),如 Defnition$\boxed{\mathrm{F}}5$中所述，以模拟理想的实验。替代定义是路径特定的因果效应$\mathbb{E}[4,25,32]$ 和反事实公平$\mathbb{E} [ 13]$ , 我们将讨论为什么在定义 5 之后没有使用替代方案
本文中的受保护变量包括红线变量，它们是受保护变量的后代。红线变量被认为是受保护变量的代理，可能会导致一些歧视[12]。一些公司或组织通过意识[E[7] 在公平概念下构建模型，这意味着分类器函数可能不使用受保护的变量作为输入。在这种情况下，红线变量将被视为受保护的变量 | [
[
"0.2173",
"0.11743",
"0.2566",
"0.11743",
"0.2566",
"0.1288",
"0.2173",
"0.1288"
],
[
"0.2712",
"0.11743",
"0.3489",
"0.11743",
"0.3489",
"0.1288",
"0.2712",
"0.1288"
],
[
"0.2443",
"0.1483",
"0.7837",
"0.1483",
"0.7837",
"0.1622",
"0.2443",
"0.1622"
],
[
"0.2197",
"0.1622",
"0.786",
"0.1635",
"0.786",
"0.178",
"0.2197",
"0.1768"
],
[
"0.2181",
"0.1768",
"0.58",
"0.1768",
"0.58",
"0.1901",
"0.2181",
"0.1901"
],
[
"0.2443",
"0.1945",
"0.785",
"0.1945",
"0.785",
"0.2084",
"0.2443",
"0.2084"
],
[
"0.2197",
"0.209",
"0.785",
"0.209",
"0.785",
"0.2235",
"0.2197",
"0.2235"
],
[
"0.2197",
"0.2374",
"0.7837",
"0.2374",
"0.7837",
"0.2507",
"0.2197",
"0.2507"
],
[
"0.2181",
"0.252",
"0.786",
"0.252",
"0.786",
"0.2651",
"0.2181",
"0.2651"
],
[
"0.2181",
"0.2676",
"0.7837",
"0.2676",
"0.7837",
"0.281",
"0.2181",
"0.281"
],
[
"0.246",
"0.2827",
"0.5605",
"0.2827",
"0.5605",
"0.2966",
"0.246",
"0.2966"
],
[
"0.2418",
"0.2942",
"0.722",
"0.2966",
"0.722",
"0.3145",
"0.2418",
"0.3125"
],
[
"0.2443",
"0.3113",
"0.5254",
"0.3125",
"0.5254",
"0.327",
"0.2443",
"0.3257"
],
[
"0.2418",
"0.3225",
"0.777",
"0.327",
"0.777",
"0.3484",
"0.2418",
"0.3447"
],
[
"0.2443",
"0.3435",
"0.523",
"0.3447",
"0.523",
"0.3591",
"0.2443",
"0.358"
],
[
"0.2385",
"0.3528",
"0.789",
"0.358",
"0.789",
"0.3833",
"0.2385",
"0.3774"
],
[
"0.2181",
"0.375",
"0.7876",
"0.377",
"0.7876",
"0.3945",
"0.2181",
"0.3926"
],
[
"0.2173",
"0.4185",
"0.4412",
"0.4185",
"0.4412",
"0.432",
"0.2173",
"0.432"
],
[
"0.2197",
"0.4507",
"0.785",
"0.4507",
"0.785",
"0.464",
"0.2197",
"0.464"
],
[
"0.2197",
"0.4666",
"0.7837",
"0.4666",
"0.7837",
"0.4797",
"0.2197",
"0.4797"
],
[
"0.2197",
"0.481",
"0.785",
"0.481",
"0.785",
"0.4944",
"0.2197",
"0.4944"
],
[
"0.2197",
"0.4968",
"0.3823",
"0.4968",
"0.3823",
"0.5103",
"0.2197",
"0.5103"
],
[
"0.2214",
"0.5244",
"0.786",
"0.5244",
"0.786",
"0.5386",
"0.2214",
"0.5386"
],
[
"0.2467",
"0.5405",
"0.786",
"0.539",
"0.786",
"0.5537",
"0.2467",
"0.5547"
],
[
"0.2467",
"0.5537",
"0.6846",
"0.5547",
"0.6846",
"0.5693",
"0.2467",
"0.5684"
],
[
"0.2239",
"0.571",
"0.7876",
"0.5693",
"0.7876",
"0.584",
"0.2239",
"0.5854"
],
[
"0.2484",
"0.586",
"0.7876",
"0.586",
"0.7876",
"0.5996",
"0.2484",
"0.5996"
],
[
"0.2484",
"0.601",
"0.6895",
"0.601",
"0.6895",
"0.6143",
"0.2484",
"0.6143"
],
[
"0.2239",
"0.6157",
"0.786",
"0.616",
"0.786",
"0.631",
"0.2239",
"0.63"
],
[
"0.2484",
"0.632",
"0.786",
"0.632",
"0.786",
"0.6465",
"0.2484",
"0.6465"
],
[
"0.2467",
"0.646",
"0.597",
"0.646",
"0.597",
"0.66",
"0.2467",
"0.66"
],
[
"0.2427",
"0.674",
"0.785",
"0.675",
"0.785",
"0.6895",
"0.2427",
"0.688"
],
[
"0.2197",
"0.69",
"0.7656",
"0.69",
"0.7656",
"0.704",
"0.2197",
"0.704"
],
[
"0.2427",
"0.7056",
"0.786",
"0.7056",
"0.786",
"0.7197",
"0.2427",
"0.7197"
],
[
"0.2197",
"0.72",
"0.4575",
"0.72",
"0.4575",
"0.7344",
"0.2197",
"0.7344"
],
[
"0.2197",
"0.7363",
"0.785",
"0.7363",
"0.785",
"0.75",
"0.2197",
"0.75"
],
[
"0.2181",
"0.75",
"0.7876",
"0.7505",
"0.7876",
"0.765",
"0.2181",
"0.7646"
],
[
"0.2197",
"0.7666",
"0.581",
"0.7666",
"0.581",
"0.78",
"0.2197",
"0.78"
],
[
"0.246",
"0.781",
"0.785",
"0.781",
"0.785",
"0.7944",
"0.246",
"0.7944"
],
[
"0.2181",
"0.7944",
"0.7837",
"0.7954",
"0.7837",
"0.81",
"0.2181",
"0.8096"
],
[
"0.2214",
"0.811",
"0.7837",
"0.811",
"0.7837",
"0.8247",
"0.2214",
"0.8247"
],
[
"0.2197",
"0.827",
"0.6436",
"0.827",
"0.6436",
"0.8403",
"0.2197",
"0.8403"
]
] | [
"4",
"Z. Xu et al.",
"为了研究Y上X之间的关系，在理想实验中控制了所有其他变量。要使用",
"因果DAG推断介入概率（通过将其简化为正常条件概率），必须使用以下do",
"-演算E[19]规则集。它们被用在我们定理的证明中",
"对于DAG9和G中的节点X的子集，9x表示通过从G中删除指向X中",
"的节点的所有箭头而获得的DAG，而Gx表示通过从9中删除从X中的节点",
"Theorem 1 (Rules of do -Calculus [19]). Let X,Y, Z, W be arbitrary dis-",
"joint sets of variables in a causal DAG g : The following rules hold, where",
", y, z, w are the shorthands of X = a, Y = y, Z = z and W = w respectively.",
"Rule 1. (Insertion/deletion of observations):",
"P(y | do(c),z, w) = P(y I do(a), w) , if (Y IL ZI X,W) in x :",
"Rule 2. (Action/observation erchange):",
"P(y I do(a), do(z),w) = P(y I do(a), z, w) , if(Y II Z|X, W) in Sxz :",
"Rule 3. (Insertion/deletion of actions):",
"P(y I do(a), do(z), w) = P( [ do(r), w) , if(Y I zI X, W) in Sxz(W) ,",
"where z(W) is the nodes in Z that are not ancestors of any node in W in Gx .",
"3Problem Definition",
"公司/组织已根据训练数据集构建了分类器（预测模型），该数据集包含二进制",
"受保护变量A，二进制决策结果Y以及Y，X的一组相关变量，因为独立于",
"XMATHX的预测不是XMATH。某机构希望使用某些案例对模型进行审计。我",
"们对审计做出以下假设",
"Assumption 1 1. The regulatory policy has specified the causal relationships",
"among the factors and Y , and uses a causal DAG to indicate. The factors",
"are ancestral variables of Y including all direct causes of Y.",
"2. The audit agency has no access to the model training data or an unbiased",
"sample of the population. The agency however has access to the distributional",
"statistics from some sources, such as government census data.",
"3. The company or organisation has used all the legitimate factors to comply",
"with the regulatory policy. However, some other variables are also used by",
"the model to enhance the prediction performance.",
"在定理开发中，我们假设存在一个与监管政策一致的DAG。在该算法中，我",
"们不需要完整的DAG，而是需要Y的祖先变量和Y的后代节点中的碰撞器",
"我们首先定义审核的标准。我们使用受控直接效应（CDE）[20]来衡量公平",
"性。CDE从do上的条件（prote",
"cted）扩展到do上的条件（所有其他变量），如DefinitionE5中所述，以模拟",
"理想的实验。替代定义是路径特定的因果效应E[4,25,32]和反事实公平E[13]，",
"我们将讨论为什么在定义5之后没有使用替代方案",
"本文中的受保护变量包括红线变量，它们是受保护变量的后代。红线变量被",
"认为是受保护变量的代理，可能会导致一些歧视[12]。一些公司或组织通过意",
"识E7在公平概念下构建模型，这意味着分类器函数可能不使用受保护的变量",
"作为输入。在这种情况下，红线变量将被视为受保护的变量"
] | {
"arxiv_id": "2010.03933",
"language": "zh",
"page_idx": 3
} |
||
2010.03933-zh-4 | 2010.03933-zh-4.jpg | Assessing Classifier Fairness with Collider Bias 5
$\textbf{Definition 6}$( Fairness score) . GivenacausalDAG G representing the reg- $ulatory policy, A$ , $\boldsymbol{X}$ , and Y as described above. The fairmess score is of $an~individual~(or~a~subgroup)~\mathbf{X}=\boldsymbol{x_{i}}~is~defined~by~Controlled~Direct~Effect$, $CDE(\boldsymbol{x}$
where $y$ denotes $Y= 1$ .
上述定义的基本原理是，我们通过干预受保护的变量并将所有其他变量控制
为$x_i$来进行受控实验。如果干预不改变结果，$x_i$的决定是公平的
与以前的作品$[\mathbb{E}[7,8,10,16]$ 不同，我们对公平分数的定义基于使用干预的CDE。因此，$A$和$\bar{Y}$之间由对撞机条件引起的虚假关联将被避免。我们在公平性定义中不使用反事实公平性[13] ,因为它需要更强大的假设并构成实际挑战。为了估计反事实结果，需要了解完整的因果模型和潜在的背景知识。在我们的问题设置中两者都不可用。其他一些定义[4,25,32]利用路径特定的因果效应。他们的解决方案也需要反事实推理，他们不适合我们的问题设置
${\text{Definition 7 }}($Problem definition$) . \textit{Given G , A and X as described above, }$ and classier $\hat{Y} = f( A, \boldsymbol{X})$ . The audit $is$ $to$ determine $if$ $a$ prediction $on$ $an$ $indi-$ vidual ( $\boldsymbol{X}= \boldsymbol{\underset {i}{x}}$ ) $is$ $fair$, $i. e.$ $| CDE( \boldsymbol{\underset {i}{x}}) | < \tau$ where $\tau$ $is$ $a$ threshold determined $by\textit{ the regulatory policy and Y in CDE(x¡) is replaced by \hat{Y} .}$
4 Estimating CDE
为了公平审核，假定受保护的变量$A$是$Y$的父节点，因此我们可以使用 CDE
进行审核。本节中的结果通常是正确的，而不仅仅是审计分类器
Theorem 2. $DAG\mathcal{G} \textit{contains variables A andY, and variable set X where }( A\cup$ $Y) \cap \boldsymbol{X}= \emptyset$ . The causal sufficiency $is$ $satisfied.$ $P( y\mid do( A= a) , do( \boldsymbol{X}= \boldsymbol{x}) ) =$ $P( y\mid A= a, Pa^{\prime }( Y) = \boldsymbol{pa})$ where $Pa^{\prime }( Y)$ $is$ the set $of$ all parents $of$ $Y$ $in$ $\mathcal{G}$ excluding $A$ .
$Proof.$ Firstly, let $\boldsymbol X=\{\boldsymbol C\cup\boldsymbol Q\}$ where $\boldsymbol C$ contains descendant nodes of $Y$,and $\boldsymbol Q$
contains non-descent nodes of$Y.$We have $P(y\mid do(A=a),do(\boldsymbol{C}=\boldsymbol{c}),do(\boldsymbol{Q}=\boldsymbol{q}))$
$= P( y\mid do( A= a) , do( \boldsymbol{Q}= \boldsymbol{q}) )$ .This is achieved by repeatedly using Rule 3 of Theorem 1.We show this by an example where $C\in\boldsymbol{C},P(y\mid do(A=$ $a),do(C=c),do(\boldsymbol{Q}=\boldsymbol{q}))=P(y\mid do(A=a),do(\boldsymbol{Q}=\boldsymbol{q}))$ because $Y\amalg C$ in DAG $\mathcal{G}_\overline{A},\overline{C}$ where the incoming edges to $A$ and to $C$ have been removed. Secondly, we consider $P(y\mid do(A=a),do(\boldsymbol{Q}=\boldsymbol{q}))$ only. Based on the Markov condition 1 $,Y$ is independent of all its non-descendant nodes given its parents. Therefore, $P( y\mid do( A= a) , do( \boldsymbol{Q}= \boldsymbol{q}) ) = P( y\mid do( A= a) , do( Pa^{\prime }( Y) = \boldsymbol{p}\boldsymbol{a}) )$ .Thirdly, we will prove $P(y|do(A=a),do(Pa^{\prime}(Y)=\boldsymbol{pa}))=P(y\mid A=a,Pa^{\prime}(Y)=\boldsymbol{pa})$ . This can be achieved by repeatedly applying Rule 2 of Theorem 1. Let | [
[
"0.4167",
"0.11615",
"0.7305",
"0.11615",
"0.7305",
"0.13",
"0.4167",
"0.13"
],
[
"0.726",
"0.1206",
"0.7305",
"0.1206",
"0.7305",
"0.1256",
"0.726",
"0.1256"
],
[
"0.773",
"0.1187",
"0.786",
"0.1187",
"0.786",
"0.1276",
"0.773",
"0.1276"
],
[
"0.2157",
"0.1459",
"0.785",
"0.1465",
"0.785",
"0.1642",
"0.2157",
"0.1635"
],
[
"0.2181",
"0.1635",
"0.7876",
"0.1622",
"0.7876",
"0.1768",
"0.2181",
"0.178"
],
[
"0.2197",
"0.1768",
"0.785",
"0.178",
"0.785",
"0.1925",
"0.2197",
"0.1913"
],
[
"0.2181",
"0.1925",
"0.786",
"0.1932",
"0.786",
"0.2084",
"0.2181",
"0.207"
],
[
"0.2214",
"0.209",
"0.398",
"0.209",
"0.398",
"0.2229",
"0.2214",
"0.2229"
],
[
"0.2443",
"0.2456",
"0.785",
"0.2456",
"0.785",
"0.2588",
"0.2443",
"0.2588"
],
[
"0.2181",
"0.2615",
"0.6895",
"0.2615",
"0.6895",
"0.2747",
"0.2181",
"0.2747"
],
[
"0.246",
"0.279",
"0.7837",
"0.279",
"0.7837",
"0.2922",
"0.246",
"0.2922"
],
[
"0.2214",
"0.2942",
"0.7837",
"0.2942",
"0.7837",
"0.308",
"0.2214",
"0.308"
],
[
"0.2197",
"0.3093",
"0.7837",
"0.3093",
"0.7837",
"0.3225",
"0.2197",
"0.3225"
],
[
"0.2197",
"0.3245",
"0.7837",
"0.3245",
"0.7837",
"0.3384",
"0.2197",
"0.3384"
],
[
"0.2214",
"0.339",
"0.786",
"0.339",
"0.786",
"0.3528",
"0.2214",
"0.3528"
],
[
"0.2197",
"0.3547",
"0.6772",
"0.3547",
"0.6772",
"0.3687",
"0.2197",
"0.3687"
],
[
"0.2181",
"0.3945",
"0.7817",
"0.3945",
"0.7817",
"0.4077",
"0.2181",
"0.4077"
],
[
"0.2197",
"0.4092",
"0.7837",
"0.4092",
"0.7837",
"0.4236",
"0.2197",
"0.4236"
],
[
"0.2214",
"0.4248",
"0.7876",
"0.4248",
"0.7876",
"0.4382",
"0.2214",
"0.4382"
],
[
"0.2181",
"0.4395",
"0.66",
"0.4382",
"0.66",
"0.4526",
"0.2181",
"0.4539"
],
[
"0.2173",
"0.4924",
"0.4167",
"0.4924",
"0.4167",
"0.507",
"0.2173",
"0.507"
],
[
"0.2214",
"0.537",
"0.785",
"0.537",
"0.785",
"0.5503",
"0.2214",
"0.5503"
],
[
"0.2197",
"0.553",
"0.675",
"0.553",
"0.675",
"0.5664",
"0.2197",
"0.5664"
],
[
"0.2181",
"0.5913",
"0.785",
"0.5923",
"0.785",
"0.607",
"0.2181",
"0.606"
],
[
"0.2197",
"0.608",
"0.7837",
"0.608",
"0.7837",
"0.622",
"0.2197",
"0.622"
],
[
"0.2197",
"0.6235",
"0.785",
"0.6235",
"0.785",
"0.637",
"0.2197",
"0.637"
],
[
"0.2197",
"0.638",
"0.3145",
"0.638",
"0.3145",
"0.653",
"0.2197",
"0.653"
],
[
"0.2197",
"0.6753",
"0.7837",
"0.6753",
"0.7837",
"0.6895",
"0.2197",
"0.6895"
],
[
"0.2181",
"0.6895",
"0.8022",
"0.69",
"0.8022",
"0.705",
"0.2181",
"0.704"
],
[
"0.2197",
"0.7056",
"0.7876",
"0.7056",
"0.7876",
"0.7197",
"0.2197",
"0.7197"
],
[
"0.2173",
"0.7197",
"0.7817",
"0.7197",
"0.7817",
"0.733",
"0.2173",
"0.733"
],
[
"0.2197",
"0.7363",
"0.785",
"0.7363",
"0.785",
"0.75",
"0.2197",
"0.75"
],
[
"0.2197",
"0.7505",
"0.7817",
"0.7505",
"0.7817",
"0.765",
"0.2197",
"0.765"
],
[
"0.2197",
"0.7666",
"0.7837",
"0.7666",
"0.7837",
"0.78",
"0.2197",
"0.78"
],
[
"0.2222",
"0.781",
"0.7837",
"0.781",
"0.7837",
"0.7954",
"0.2222",
"0.7954"
],
[
"0.2197",
"0.797",
"0.7837",
"0.797",
"0.7837",
"0.81",
"0.2197",
"0.81"
],
[
"0.2197",
"0.811",
"0.7837",
"0.811",
"0.7837",
"0.8257",
"0.2197",
"0.8257"
],
[
"0.2157",
"0.827",
"0.7837",
"0.827",
"0.7837",
"0.8403",
"0.2157",
"0.8403"
]
] | [
"Assessing Classifier Fairness with Collider Bias",
"",
"5",
"Definition 6 (Fairness score). Given a causal DAG g representing the reg-",
"ulatory policy, A , X, and Y as described above. The fairness score is of",
"an individual (or a subgroup) X = &i is defined by Controlled Direct Effect,",
"CDE(αi) = P(y / do(A = 1), do(X = i)) - P(y I do(A = 0),do(X = ;)) ,",
"wherey denotes Y =l:",
"上述定义的基本原理是，我们通过干预受保护的变量并将所有其他变量控制",
"为；来进行受控实验。如果干预不改变结果，；的决定是公平的",
"与以前的作品E[7,8,10,16]不同，我们对公平分数的定义基于使用干预的",
"CDE。因此，A和Y之间由对撞机条件引起的虚假关联将被避免。我们在公平",
"性定义中不使用反事实公平性[13]，因为它需要更强大的假设并构成实际挑战。",
"为了估计反事实结果，需要了解完整的因果模型和潜在的背景知识。在我们的",
"问题设置中两者都不可用。其他一些定义[4,25,32]利用路径特定的因果效应。",
"他们的解决方案也需要反事实推理，他们不适合我们的问题设置",
"Definition 7 (Problem definition). Given g , A and X as described above",
"and classier Y =- f(A, X) . The audit is to determine if a prediction on an indi",
"vidual (X = &i ) is fair, i.e. JCDE(i)l < T where T is a threshold determined",
"by the regulatory policy and Y inCDE(ai)is replaced byY.",
"4Estimating CDE",
"为了公平审核，假定受保护的变量A是Y的父节点，因此我们可以使用CDE",
"进行审核。本节中的结果通常是正确的，而不仅仅是审计分类器",
"Theorem 2. DAGg contains variables A and Y , and variable set X where (AU",
"Y)nX = 0 . The causal sufficiency is satisfied. P(y I do(A= a), do(X = ) =",
"P(yIA =a, Pa'(Y)=pa) where Pa'(Y) is the set of all parents of Y in G",
"ercludingA.",
"Proof. Firstly, let X = (CUQ) where C contains descendant nodes of Y , and Q",
"contains non-descent nodes of Y . We have P(y | do(A = a), do(C = c), do(Q = q))",
"= P(y I do(A = a), do(Q = q)) . This is achieved by repeatedly using Rule 3",
"of Theorem 1 . We show this by an example where C E C , P(y I do(A =",
"a), do(C = c), do(Q = q)) = P(y I do(A = a), do(Q = q)) because Y IL C in",
"DAG GA where the incoming edges to A and to C have been removed. Secondly",
"we consider P(y I do(A = a), do(Q = q) only. Based on the Markov condition 1",
"Y is independent of all its non-descendant nodes given its parents. Therefore",
"P(y I do(A = a), do(Q = q)) = P(y I do(A = a), do(Pa'(Y) = pa)) . Thirdly",
"we will prove P(yldo(A = a), do(Pa'(Y) = pa)) = P(y [ A = a, Pa'(Y) = pa)",
".This can be achieved by repeatedly applying Rule 2 of Theorem1.Let"
] | {
"arxiv_id": "2010.03933",
"language": "zh",
"page_idx": 4
} |
||
2404.11003-zh-2 | 2404.11003-zh-2.jpg | 见解，从而提高半监督图像分类的性能主要贡献总结如下：1)我们提出的 InfoMatch 有效地利用了未标记数据的潜力。2) 利用多个目标，InfoMatch 近似数据熵和地面实况后验熵。通过利用强到强的约束，弱到强的伪监督和 CutMix 策略，我们将这种方法应用于未标记的数据，从而可以更有效地利用未标记的数据信息。3) 跨各种半监督学习基准的实验验证了 InfoMatch 的卓越性能。
## 2 Related Works
半监督学习是机器学习和计算机视觉领域的一个重要分支。它的基本概念在于，当只有有限数量的标记样本可用时，利用大量未标记样本中潜在的数据分布信息来提高学习性能。SSL 背后的主要策略是伪监督和一致性正则化一致性正则化。基于平滑性假设，一致性正则化表明，应用于未标记样本的微小扰动不会导致其预测的实质性变化。为了确保一致性，$\Pi 模型 [Rasmus et al.$, 2015]通过数据增强和丢失引人扰动，旨在最大化从相同未标记样本的两个前向传播得出的预测的相似性。时态集成 [Samuli and Timo, 2017] 采用时间序列组合模型来最小化当前预测和历史预测之间的均方误差损失，从而简化正向推理。MeanTeacher [Tarvainen and Valpola, 2017]将预测结果的指数移动平均值转换为模型权重，惩罚学生模型和教师模型之间的预测差异。此外，无监督数据增强(UDA)E[Xie et al., 2020]扩展了数据增强方法，并针对特定任务采用了特定的目标数据增强算法。相比之下，伪标签[Lee,2013]的工作原理是为未标记的数据生成人工标签，并使用标记和伪标记的数据执行完全监督训练整体方法。大多数 SSL 方法将伪监督与一致性正则化相结合，以提高性能。MixMatch [Berthelot et al., 2019b] 通过集成一致性正则化和熵最小化来利用未标记的数据。基于 MixMatch, RemixMatch [Berthelot et al., 2019a] 引人了两种创新策略，分布对齐和增强锚以增强其鲁棒性和准确性。FixMatch [Sohn et al., 2020] ,它保持了增强锚的使用，简化了强增强的数量。为了提高伪标记的准确性，建立了一个固定的高阈值来消除不可靠的标记。为了缓解类不均衡，解决前期未标记数据利用率低的问题，FlexMatch [Zhang et al., 2021] 和FreeMatch [Wang et al., 2023] 考虑模型的学习状态和类学习难度，自适应调整置信度阈值。SoftMatch[Chen] et al., 2023] 使用截断的高斯函数根据样本的置信度为样本分配权重，解决了伪标签数量和质量之间的固有权衡。此外，CoMatch [Li et al., 2021] 、SimMatch [Zheng et al., 2022] 和 SimMatchV2 [Zheng et al., 2023] 等方法将基于图的学习方法集成到半监督图像分类中，共同学习训练数据的两种表示形式，类概率和低维嵌入混合增强。两种流行的图像混合方法是 Mixup [Zhang et al., 2018] 和 CutMix [Yun et al., 2019] 。Mixup 通过线性组合两个图像及其相应的标签来生成新的训练样本，旨在促进模型学习更平滑的决策边界。相反，CutMix 将一个图像的片段与另一个图像的片段交换，同时更新相应的标签，以鼓励模型理解和利用图像中的本地信息。插值一致性训练 [Verma et al., 2022] 、MixMatch [Berthelot et al., 2019b] 和 RemixMatch [Berthelot et al., 2019a] 将 Mixup 与一致性正则化结合起来，通过生成虚拟的“混合”训练样本来扩展数据集，利用一致性正则化来保证其在混合数据上的泛化能力。而 FMixCutMatch [Wei et al., 2021] 结合了基于傅立叶空间的数据切割和数据混
合，以增加标记和未标记的样本，从而提高其性能和泛化能力。此外，[Ghorban et al., 2022] 同时利用 Mixup 和 CutMix, 旨在融合三个图像的语义信息，并将增强的扰动引人数据。为了进一步增强其对本地信息的关注并提高其对各种位置提示的敏感性，我们在 InfoMatch 中${\text{采 用 了 CutMix}\underline {\mathrm{E} }[ \text{Yun et al. , 2019] }}$ 。
# 3 Entropy Neural Estimation
我们采用独立且相同分布的(i。i。d。)训练数据集表示为 $\mathcal{X} = \mathcal{X} _{l}\cup \mathcal{X} _{u}$ ,其 中 $\mathcal{X} _{l}$ 包 括 标 记 数 据 对 $( \boldsymbol{x}_{i}, \boldsymbol{p}_{i}) , \forall i\in$ $\{ 1, \ldots , n_l\}$ , $\mathcal{X} _u$由未标记数据$\boldsymbol x_i,\forall i\in\{n_l+1,\ldots,n_l+$ $n_u\}$组成。标记和未标记数据集的大小分别用$n_l$和$n_u$ 表示，用$n_l\ll n_u$表示，总数据集大小为$n=n_l+n_u$ 在这里，每个$x_i$表示一个数据点，而$p_i$对应于其基本事实类标签，使用独热编码表示。$\mathcal{C}=\{c_1,\ldots,c_k\}$表示类集。如果 1个$x_i$属于$c_j$类，则我们有$p(c_j|\boldsymbol{x}_i)=p_ij=1$ 我们使用一个编码器来生成一个潜在特征$z_i=[z_{ij}]$ ,并使用 softmax 函数？?? 定义属于$j$类$c_j$的$x_i$的后验概${\text{率 }}$ $\sigma ( c_{j}| \boldsymbol{x}_{i}, \theta )$ : <NEXT\_THX>
(1)
$$y_{ij}=\sigma(c_j|\boldsymbol{x}_i,\theta)=\frac{\exp(z_{ij})}{\sum_j\exp(z_{ij})}\:.$$
这意味着我们将$\sigma(c_j|\boldsymbol{x}_i,\theta)$建模为编码器，以逼近真值后验$p_{ij}=p(c_{j}|\boldsymbol{x}_{i})=1$ 。根据信息瓶颈理论的第二项$\overline{\mathrm{E}}[$Tishby et al., 2000] , min I$(\mathcal{X};\mathcal{Y})-\beta$I$(\mathcal{Y};\mathcal{C})$ ,神经网络主要对其 softmax 预测$\sigma(c_j|\boldsymbol{x}_i,\theta)$进行编码，以接近地面真相后验$p(c_j|\boldsymbol{x}_i)$ 。
Theorem 1. For an i.i.d. finite dataset $(\mathcal{X},\mathcal{C})$ ,the ap-
proximation H$(\mathcal{C}|\mathcal{X})\simeq-\ln p(\mathcal{C}|\mathcal{X})$ holds.
Proof. Since $\boldsymbol{x}_i$ are sampled i.i.d., $p(c|\boldsymbol{x}_i)$ corresponds point-to-point with $\boldsymbol x_i$ and is regarded as an i.i.d. set, $\bar {\text{the posterior is }p}( \mathcal{C} | \mathcal{X} ) = \prod _{i= 1}^{n}\bar {p( c| \boldsymbol{x}_{i}) }.$Then the expectation with respect to $p(\boldsymbol{x})$ is approximated by a finite sum over $\mathcal{X}$ , so that
$$\begin{aligned}\mathrm{H}(\mathcal{C}|\mathcal{X})&=-\mathbb{E}_{p(\boldsymbol{x})}\mathbb{E}_{p(c|\boldsymbol{x})}\ln p(c|\boldsymbol{x})\\&\simeq-\sum_{i=1}^{n}\ln p(c|\boldsymbol{x}_{i})=-\ln\prod_{i=1}^{n}p(c|\boldsymbol{x}_{i})\\&=-\ln p(\mathcal{C}|\mathcal{X})\:.\end{aligned}$$
$\square$
Thus, the equation H$(\mathcal{C}|\mathcal{X})\simeq-\ln p(\mathcal{C}|\mathcal{X})$ holds.
3.1 Posterior Entropy Neural Estimation
假设地面实况后验是从我们旨在建模的未知分布$p(c|\boldsymbol{x}_i)$
生成的。我们使用参数概率分布$\sigma(c|\boldsymbol{x},\theta)$
(2)
$$\begin{aligned}\mathrm{div}_{\mathrm{kl}}(p\|\sigma)&\simeq\sum_{i=1}^{n}\Big\{-\ln\sigma(c|\boldsymbol{x}_{i},\theta)+\ln p(c|\boldsymbol{x}_{i})\Big\}\\&=-\big\{\ln\sigma(\mathcal{C}|\mathcal{X},\theta)-\ln p(\mathcal{C}|\mathcal{X})\big\}\geqslant0\\\mathrm{H}(\mathcal{C}|\mathcal{X})&\leqslant-\ln\sigma(\mathcal{C}|\mathcal{X},\theta)\:.\end{aligned}$$
(3) (4)
近似$p(c|\boldsymbol{x}_i)$ ,其中 div$_\mathrm{kl}$表示 Kullback-Leibler 散度。方程 (4)的左侧与$\theta$无关，右侧项是使用训练数据集评估的后验分布下$\theta$的负对数似然函数。因此，我们看到最小化方程$\mathbb{E}(3)$ 等效于最大化似然函数 i.e.,从而最小 | [
[
"0.0907",
"0.07135",
"0.4844",
"0.07135",
"0.4844",
"0.0821",
"0.0907",
"0.0821"
],
[
"0.5146",
"0.0707",
"0.911",
"0.0707",
"0.911",
"0.084",
"0.5146",
"0.084"
],
[
"0.08905",
"0.0821",
"0.486",
"0.0833",
"0.486",
"0.09784",
"0.08905",
"0.0966"
],
[
"0.5146",
"0.084",
"0.911",
"0.084",
"0.911",
"0.09784",
"0.5146",
"0.09784"
],
[
"0.08826",
"0.09784",
"0.486",
"0.09784",
"0.486",
"0.1111",
"0.08826",
"0.1111"
],
[
"0.5146",
"0.09784",
"0.911",
"0.09784",
"0.911",
"0.1111",
"0.5146",
"0.1111"
],
[
"0.08905",
"0.113",
"0.4836",
"0.113",
"0.4836",
"0.1237",
"0.08905",
"0.1237"
],
[
"0.511",
"0.10986",
"0.911",
"0.1111",
"0.911",
"0.1256",
"0.511",
"0.125"
],
[
"0.08905",
"0.1256",
"0.4844",
"0.1256",
"0.4844",
"0.1395",
"0.08905",
"0.1395"
],
[
"0.513",
"0.1256",
"0.911",
"0.1256",
"0.911",
"0.1395",
"0.513",
"0.1395"
],
[
"0.08905",
"0.1383",
"0.4844",
"0.1383",
"0.4844",
"0.1515",
"0.08905",
"0.1515"
],
[
"0.513",
"0.1395",
"0.778",
"0.1395",
"0.778",
"0.1528",
"0.513",
"0.1528"
],
[
"0.08905",
"0.1528",
"0.4844",
"0.1528",
"0.4844",
"0.1666",
"0.08905",
"0.1666"
],
[
"0.08826",
"0.1674",
"0.1111",
"0.1674",
"0.1111",
"0.1799",
"0.08826",
"0.1799"
],
[
"0.511",
"0.1686",
"0.778",
"0.1686",
"0.778",
"0.1831",
"0.511",
"0.1831"
],
[
"0.513",
"0.1875",
"0.911",
"0.1881",
"0.911",
"0.2026",
"0.513",
"0.202"
],
[
"0.08496",
"0.2008",
"0.2467",
"0.2008",
"0.2467",
"0.2153",
"0.08496",
"0.2153"
],
[
"0.5146",
"0.202",
"0.911",
"0.2026",
"0.911",
"0.2166",
"0.5146",
"0.2153"
],
[
"0.5156",
"0.2166",
"0.911",
"0.2166",
"0.911",
"0.2299",
"0.5156",
"0.2299"
],
[
"0.08826",
"0.2247",
"0.4844",
"0.2247",
"0.4844",
"0.238",
"0.08826",
"0.238"
],
[
"0.511",
"0.2291",
"0.911",
"0.2299",
"0.911",
"0.2443",
"0.511",
"0.2437"
],
[
"0.08905",
"0.238",
"0.4844",
"0.238",
"0.4844",
"0.2524",
"0.08905",
"0.2524"
],
[
"0.5146",
"0.2443",
"0.911",
"0.2443",
"0.911",
"0.2583",
"0.5146",
"0.2583"
],
[
"0.08905",
"0.2524",
"0.4836",
"0.2524",
"0.4836",
"0.2664",
"0.08905",
"0.2664"
],
[
"0.513",
"0.2583",
"0.9126",
"0.2583",
"0.9126",
"0.2727",
"0.513",
"0.2727"
],
[
"0.0907",
"0.2683",
"0.4844",
"0.2683",
"0.4844",
"0.279",
"0.0907",
"0.279"
],
[
"0.513",
"0.2695",
"0.9126",
"0.2715",
"0.9126",
"0.286",
"0.513",
"0.2842"
],
[
"0.08905",
"0.2798",
"0.486",
"0.2798",
"0.486",
"0.2935",
"0.08905",
"0.2935"
],
[
"0.511",
"0.2822",
"0.892",
"0.286",
"0.892",
"0.3005",
"0.511",
"0.2966"
],
[
"0.0866",
"0.2922",
"0.486",
"0.2935",
"0.486",
"0.308",
"0.0866",
"0.307"
],
[
"0.8936",
"0.289",
"0.9067",
"0.289",
"0.9067",
"0.2935",
"0.8936",
"0.2935"
],
[
"0.511",
"0.2974",
"0.914",
"0.2974",
"0.914",
"0.315",
"0.511",
"0.315"
],
[
"0.08905",
"0.308",
"0.4844",
"0.308",
"0.4844",
"0.3213",
"0.08905",
"0.3213"
],
[
"0.511",
"0.31",
"0.914",
"0.3113",
"0.914",
"0.3289",
"0.511",
"0.3276"
],
[
"0.08826",
"0.3213",
"0.4844",
"0.3213",
"0.4844",
"0.3352",
"0.08826",
"0.3352"
],
[
"0.5107",
"0.3257",
"0.7524",
"0.324",
"0.7524",
"0.3403",
"0.5107",
"0.3423"
],
[
"0.0907",
"0.3372",
"0.4844",
"0.3372",
"0.4844",
"0.348",
"0.0907",
"0.348"
],
[
"0.08905",
"0.3499",
"0.4844",
"0.3499",
"0.4844",
"0.363",
"0.08905",
"0.363"
],
[
"0.744",
"0.351",
"0.806",
"0.351",
"0.806",
"0.3655",
"0.744",
"0.3655"
],
[
"0.5947",
"0.3591",
"0.734",
"0.3567",
"0.734",
"0.3718",
"0.5947",
"0.3738"
],
[
"0.726",
"0.3623",
"0.824",
"0.3662",
"0.8228",
"0.385",
"0.7256",
"0.3806"
],
[
"0.8896",
"0.3591",
"0.9126",
"0.3591",
"0.9126",
"0.3726",
"0.8896",
"0.3726"
],
[
"0.08826",
"0.377",
"0.4836",
"0.377",
"0.4836",
"0.3901",
"0.08826",
"0.3901"
],
[
"0.08826",
"0.363",
"0.4844",
"0.363",
"0.4844",
"0.377",
"0.08826",
"0.377"
],
[
"0.08826",
"0.3901",
"0.4836",
"0.3901",
"0.4836",
"0.404",
"0.08826",
"0.404"
],
[
"0.5146",
"0.3901",
"0.911",
"0.3926",
"0.911",
"0.4072",
"0.5146",
"0.4048"
],
[
"0.08826",
"0.4048",
"0.486",
"0.4048",
"0.486",
"0.4185",
"0.08826",
"0.4185"
],
[
"0.5146",
"0.4072",
"0.899",
"0.4072",
"0.899",
"0.4216",
"0.5146",
"0.4216"
],
[
"0.08905",
"0.4185",
"0.486",
"0.4185",
"0.486",
"0.432",
"0.08905",
"0.432"
],
[
"0.513",
"0.4216",
"0.911",
"0.4216",
"0.911",
"0.435",
"0.513",
"0.435"
],
[
"0.08905",
"0.433",
"0.4844",
"0.433",
"0.4844",
"0.4463",
"0.08905",
"0.4463"
],
[
"0.5146",
"0.4363",
"0.9126",
"0.4363",
"0.9126",
"0.4507",
"0.5146",
"0.4507"
],
[
"0.08826",
"0.4463",
"0.4844",
"0.4463",
"0.4844",
"0.4602",
"0.08826",
"0.4602"
],
[
"0.513",
"0.4507",
"0.7197",
"0.4507",
"0.7197",
"0.464",
"0.513",
"0.464"
],
[
"0.08826",
"0.4602",
"0.486",
"0.4602",
"0.486",
"0.4734",
"0.08826",
"0.4734"
],
[
"0.511",
"0.4673",
"0.9087",
"0.4685",
"0.9087",
"0.483",
"0.511",
"0.4817"
],
[
"0.08826",
"0.4734",
"0.4844",
"0.4734",
"0.4844",
"0.4873",
"0.08826",
"0.4873"
],
[
"0.0866",
"0.4873",
"0.486",
"0.486",
"0.486",
"0.5005",
"0.0866",
"0.502"
],
[
"0.513",
"0.483",
"0.812",
"0.483",
"0.812",
"0.4968",
"0.513",
"0.4968"
],
[
"0.0866",
"0.5005",
"0.4844",
"0.5005",
"0.4844",
"0.5146",
"0.0866",
"0.5146"
],
[
"0.513",
"0.505",
"0.9126",
"0.505",
"0.9126",
"0.5195",
"0.513",
"0.5195"
],
[
"0.0866",
"0.515",
"0.4844",
"0.515",
"0.4844",
"0.529",
"0.0866",
"0.529"
],
[
"0.513",
"0.518",
"0.91",
"0.5195",
"0.91",
"0.5327",
"0.513",
"0.532"
],
[
"0.0866",
"0.53",
"0.4836",
"0.53",
"0.4836",
"0.5435",
"0.0866",
"0.5435"
],
[
"0.513",
"0.5327",
"0.91",
"0.5327",
"0.91",
"0.5474",
"0.513",
"0.5474"
],
[
"0.08905",
"0.5444",
"0.482",
"0.5444",
"0.482",
"0.5547",
"0.08905",
"0.5547"
],
[
"0.513",
"0.5474",
"0.9126",
"0.5474",
"0.9126",
"0.561",
"0.513",
"0.561"
],
[
"0.10455",
"0.5566",
"0.482",
"0.5566",
"0.482",
"0.571",
"0.10455",
"0.571"
],
[
"0.513",
"0.5625",
"0.664",
"0.561",
"0.664",
"0.5728",
"0.513",
"0.5737"
],
[
"0.08826",
"0.5713",
"0.4844",
"0.5713",
"0.4844",
"0.5854",
"0.08826",
"0.5854"
],
[
"0.08826",
"0.5854",
"0.4844",
"0.5854",
"0.4844",
"0.5986",
"0.08826",
"0.5986"
],
[
"0.5596",
"0.584",
"0.7935",
"0.584",
"0.7935",
"0.5986",
"0.5596",
"0.5986"
],
[
"0.0866",
"0.5986",
"0.486",
"0.5986",
"0.486",
"0.6123",
"0.0866",
"0.6123"
],
[
"0.6445",
"0.61",
"0.8667",
"0.613",
"0.8667",
"0.631",
"0.6445",
"0.6274"
],
[
"0.8",
"0.604",
"0.806",
"0.604",
"0.806",
"0.6094",
"0.8",
"0.6094"
],
[
"0.0866",
"0.6113",
"0.4836",
"0.6113",
"0.4836",
"0.625",
"0.0866",
"0.625"
],
[
"0.6177",
"0.6143",
"0.6543",
"0.6143",
"0.6543",
"0.6255",
"0.6177",
"0.6255"
],
[
"0.08826",
"0.6255",
"0.4844",
"0.6255",
"0.4844",
"0.6396",
"0.08826",
"0.6396"
],
[
"0.649",
"0.6274",
"0.6733",
"0.631",
"0.671",
"0.6426",
"0.6465",
"0.6396"
],
[
"0.6543",
"0.6235",
"0.6626",
"0.6235",
"0.6626",
"0.6274",
"0.6543",
"0.6274"
],
[
"0.7915",
"0.63",
"0.8164",
"0.63",
"0.8164",
"0.6426",
"0.7915",
"0.6426"
],
[
"0.0866",
"0.64",
"0.4844",
"0.64",
"0.4844",
"0.654",
"0.0866",
"0.654"
],
[
"0.615",
"0.6445",
"0.722",
"0.6445",
"0.722",
"0.659",
"0.615",
"0.659"
],
[
"0.08826",
"0.654",
"0.4844",
"0.654",
"0.4844",
"0.6675",
"0.08826",
"0.6675"
],
[
"0.08826",
"0.6675",
"0.486",
"0.6675",
"0.486",
"0.681",
"0.08826",
"0.681"
],
[
"0.511",
"0.6665",
"0.859",
"0.6665",
"0.859",
"0.6797",
"0.511",
"0.6797"
],
[
"0.898",
"0.6694",
"0.911",
"0.6694",
"0.911",
"0.6787",
"0.898",
"0.6787"
],
[
"0.08905",
"0.6816",
"0.482",
"0.6816",
"0.482",
"0.6924",
"0.08905",
"0.6924"
],
[
"0.08826",
"0.696",
"0.4836",
"0.696",
"0.4836",
"0.709",
"0.08826",
"0.709"
],
[
"0.513",
"0.6924",
"0.8506",
"0.6924",
"0.8506",
"0.7056",
"0.513",
"0.7056"
],
[
"0.08826",
"0.709",
"0.4805",
"0.709",
"0.4805",
"0.7227",
"0.08826",
"0.7227"
],
[
"0.5156",
"0.711",
"0.911",
"0.711",
"0.911",
"0.7246",
"0.5156",
"0.7246"
],
[
"0.08826",
"0.7236",
"0.4836",
"0.7236",
"0.4836",
"0.7373",
"0.08826",
"0.7373"
],
[
"0.513",
"0.7227",
"0.8228",
"0.7246",
"0.8228",
"0.7393",
"0.513",
"0.7373"
],
[
"0.08826",
"0.7373",
"0.4844",
"0.7373",
"0.4844",
"0.7505",
"0.08826",
"0.7505"
],
[
"0.08826",
"0.7505",
"0.4844",
"0.7505",
"0.4844",
"0.7646",
"0.08826",
"0.7646"
],
[
"0.54",
"0.7583",
"0.6445",
"0.757",
"0.6445",
"0.7715",
"0.54",
"0.773"
],
[
"0.6665",
"0.757",
"0.853",
"0.7583",
"0.853",
"0.773",
"0.6665",
"0.7715"
],
[
"0.8906",
"0.759",
"0.91",
"0.759",
"0.91",
"0.773",
"0.8906",
"0.773"
],
[
"0.6396",
"0.761",
"0.6772",
"0.7676",
"0.675",
"0.779",
"0.636",
"0.773"
],
[
"0.08826",
"0.7646",
"0.4844",
"0.7646",
"0.4844",
"0.778",
"0.08826",
"0.778"
],
[
"0.08905",
"0.779",
"0.4844",
"0.779",
"0.4844",
"0.7925",
"0.08905",
"0.7925"
],
[
"0.638",
"0.776",
"0.6587",
"0.776",
"0.6587",
"0.784",
"0.638",
"0.784"
],
[
"0.08905",
"0.7925",
"0.4697",
"0.7925",
"0.4697",
"0.806",
"0.08905",
"0.806"
],
[
"0.6177",
"0.7915",
"0.865",
"0.7915",
"0.865",
"0.8047",
"0.6177",
"0.8047"
],
[
"0.8896",
"0.7925",
"0.911",
"0.7925",
"0.911",
"0.8047",
"0.8896",
"0.8047"
],
[
"0.08826",
"0.806",
"0.4844",
"0.806",
"0.4844",
"0.8193",
"0.08826",
"0.8193"
],
[
"0.5605",
"0.81",
"0.74",
"0.81",
"0.74",
"0.8237",
"0.5605",
"0.8237"
],
[
"0.8896",
"0.81",
"0.9126",
"0.81",
"0.9126",
"0.8237",
"0.8896",
"0.8237"
],
[
"0.0866",
"0.8193",
"0.4836",
"0.8193",
"0.4836",
"0.8335",
"0.0866",
"0.8335"
],
[
"0.08826",
"0.834",
"0.4844",
"0.834",
"0.4844",
"0.8477",
"0.08826",
"0.8477"
],
[
"0.513",
"0.834",
"0.91",
"0.834",
"0.91",
"0.8477",
"0.513",
"0.8477"
],
[
"0.0866",
"0.845",
"0.486",
"0.8467",
"0.486",
"0.8613",
"0.0866",
"0.86"
],
[
"0.5146",
"0.8477",
"0.911",
"0.8477",
"0.911",
"0.8613",
"0.5146",
"0.8613"
],
[
"0.08826",
"0.8613",
"0.486",
"0.8613",
"0.486",
"0.875",
"0.08826",
"0.875"
],
[
"0.5146",
"0.8623",
"0.91",
"0.8623",
"0.91",
"0.8755",
"0.5146",
"0.8755"
],
[
"0.0866",
"0.8755",
"0.486",
"0.8755",
"0.486",
"0.8896",
"0.0866",
"0.8896"
],
[
"0.513",
"0.8755",
"0.911",
"0.8755",
"0.911",
"0.8896",
"0.513",
"0.8896"
]
] | [
"见解，从而提高半监督图像分类的性能主要贡献总结如",
"合，以增加标记和未标记的样本，从而提高其性能和泛",
"下：i）我们提出的InfoMatch有效地利用了未标记数",
"化能力。此外，[Ghorban etal.，2022]同时利用Mixup",
"据的潜力。2）利用多个目标，InfoMatch近似数据熵和",
"和CutMix，旨在融合三个图像的语义信息，并将增强的",
"地面实况后验摘。通过利用强到强的约束，弱到强的伪",
"扰动引人数据。为了进一步增强其对本地信息的关注并",
"监督和CutMix策略，我们将这种方法应用于未标记的",
"提高其对各种位置提示的敏感性，我们在InfoMatch中",
"数据，从而可以更有效地利用未标记的数据信息。3）跨",
"采用了 CutMixE[Yun et al., 2019] 。",
"各种半监督学习基准的实验验证了InfoMatch的卓越性",
"能。",
"3EntropyNeuralEstimation",
"我们采用独立且相同分布的（i。i。d。）训练数据集表示",
"2Related Works",
"为x=uu，其中包括标记数据对（aipi),ie",
"{],.,nl，Xu由未标记数据ai,Vien+l,..,n+",
"半监督学习是机器学习和计算机视觉领域的一个重要分",
"nul组成。标记和未标记数据集的大小分别用ni和nau",
"支。它的基本概念在于，当只有有限数量的标记样本可",
"表示，用n<nu表示，总数据集大小为n=n+nu",
"用时，利用大量未标记样本中潜在的数据分布信息来提",
"在这里，每个表示一个数据点，而p；对应于其基本事",
"高学习性能。SSL背后的主要策略是伪监督和一致性正",
"实类标签，使用独热编码表示。℃=【c1，.,ck}表示类",
"则化一致性正则化。基于平滑性假设，一致性正则化表",
"集。如果1个ai属于c,类，则我们有p(ciai）=Pi",
"明，应用于未标记样本的微小扰动不会导致其预测的实",
"",
"我们使用一个编码器来生成一个潜在特征z,=【]，并",
"质性变化。为了确保一致性，Ⅱ模型[Rasmusetal.",
"使用softmax函数???定义属于j类c,的a的后验概",
"2015]通过数据增强和丢失引入扰动，旨在最大化从相同",
"率(cj|ci,0)：<NEXT THX>",
"未标记样本的两个前向传播得出的预测的相似性。时态",
"集成[SamuliandTimo，2017]采用时间序列组合模型来",
"exp(zij)",
"yij=o(ciai,0)=",
"Z, exp(zis)",
"(1)",
"简化正向推理。MeanTeacher[TarvainenandValpola",
"最小化当前预测和历史预测之间的均方误差损失，从而",
"2017]将预测结果的指数移动平均值转换为模型权重，惩",
"。这意味着我们将α(ciai,）建模为编码器，以逼近真",
"罚学生模型和教师模型之间的预测差异。此外，无监督",
"值后验pii=p(cila）=1。根据信息瓶颈理论的第",
"数据增强（UDA）E[Xieetal.，2020]扩展了数据增强",
"项E[Tishby et al.,2000]minI(X:y)-βI(y:C)，神",
"方法，并针对特定任务采用了特定的目标数据增强算法。",
"经网络主要对其softmax预测α(cci,の）进行编码，以",
"相比之下，伪标签[Lee,2013]的工作原理是为未标记的",
"接近地面真相后验p(ciai）。",
"数据生成人工标签，并使用标记和伪标记的数据执行完",
"Theorem 1. For an i.i.d. finite dataset (x,C) , the ap-",
"全监督训练整体方法。大多数SSL方法将伪监督与一致",
"性正则化相结合，以提高性能。MixMatch[Berthelotet",
"proximation H(C|X) ~ - ln p(CX) holds :",
"al.，2019bl通过集成一致性正则化和熵最小化来利用未",
"Proof. Since i are sampled ii.d., p(cc;) corresponds",
"标记的数据。基于MixMatch，RemixMatch[Berthelot",
"point-to-point with and is regarded as an i.i.d. set",
"etal.，2019al引人了两种创新策略，分布对齐和增强锚",
"the posterior is p(C|X) = II\"- p(clc) . Then the expec-",
"以增强其鲁棒性和准确性。FixMatchSohnetal..2020",
"tation with respect to p(a) is approximated by a finite",
"它保持了增强锚的使用，简化了强增强的数量。为了",
"sum over,sothat",
"提高伪标记的准确性，建立了一个固定的高阈值来消除",
"不可靠的标记。为了缓解类不均衡，解决前期未标记数",
"H(C) =-Ep(a)Ep(ca) lnp(cα)",
"据利用率低的问题，FlexMatch[Zhang etal.，2021]和",
"ln p(cai) =-In p(cai)",
"77",
"FreeMatch「Wangetal.，2023]考虑模型的学习状态和",
"~-",
"类学习难度，自适应调整置信度阈值。SoftMatch[Chen",
"=1",
"",
"2=1",
"etal.，2023]使用截断的高斯函数根据样本的置信度为样",
"=- Inp(Cx) .",
"本分配权重，解决了伪标签数量和质量之间的固有权衡",
"此外，CoMatch[Li et al., 2021]、SimMatch[Zheng et",
"Thus, the equation H(Cx) ~ - In p(Clx) holds.",
"口",
"al..2022和SimMatchV2[Zhengetal..2023l等方法将",
"基于图的学习方法集成到半监督图像分类中，共同学习",
"3.1Posterior Entropy NeuralEstimation",
"训练数据的两种表示形式，类概率和低维嵌人混合增强",
"假设地面实况后验是从我们旨在建模的未知分布p(cai)",
"两种流行的图像混合方法是Mixup[Zhangetal.，2018]",
"生成的。我们使用参数概率分布α(ca,の），",
"和CutMix[Yunetal.，2019]。Mixup通过线性组合两",
"个图像及其相应的标签来生成新的训练样本，旨在促进",
"divki(pllo) ",
"-lno(ci,0)+lnp(cai)",
"(2)",
"",
"模型学习更平滑的决策边界。相反，CutMix将二个图",
"像的片段与另一个图像的片段交换，同时更新相应的标",
"-1",
"签，以鼓励模型理解和利用图像中的本地信息。插值",
"=-{lno(CX,0) -lnp(Cx))≥ 0",
"(3)",
"致性训练[Vermaet al.,2022]、MixMatch[Berthelot",
"H(Cx) ≤-Ino(Cx,0) .",
"(4)",
"et al., 2019b] 和 RemixMatch [Berthelot et al., 2019a",
"将Mixup与一致性正则化结合起来，通过生成虚拟的",
"近似p(ca），其中divkl表示Kullback-Leibler散度",
"“混合”训练样本来扩展数据集，利用一致性正则化来保",
"方程（4）的左侧与θ无关，右侧项是使用训练数据集评",
"证其在混合数据上的泛化能力。而FMixCutMatch[Wei",
"估的后验分布下θ的负对数似然函数。因此，我们看到",
"etal.,2021]结合了基于傅立叶空间的数据切割和数据混",
"最小化方程E(3）等效于最大化似然函数i.e.，从而最小"
] | {
"arxiv_id": "2404.11003",
"language": "zh",
"page_idx": 2
} |
||
2404.11003-zh-4 | 2404.11003-zh-4.jpg | <table>
<tbody>
<tr>
<td>$\overline{{\mathrm{Alg}}}$</td>
<td>orithm The InfoMatch algorithm.</td>
</tr>
<tr>
<td>$\operatorname{Inp}$</td>
<td>: $\mathcal{X} = \mathcal{X} _{l}\cup \mathcal{X} _{u}$ , batch size $n_l^{b}$ and $n_u^{b}$ , parameter $)ut:$ </td>
</tr>
<tr>
<td>2:</td>
<td>while $e{\mathcal{DOC}}$ </td>
</tr>
<tr>
<td>3:</td>
<td>tor munl-bat samnles $\chi$ $do$</td>
</tr>
<tr>
<td>4:</td>
<td>Obtain augmented images $x^\prime,x^{(1)}$ $x^{(2)}$ and $\boldsymbol{\eta }^{C}$ ,</td>
</tr>
<tr>
<td>5:</td>
<td>Feed samples into model tor torward propaga- tion;</td>
</tr>
<tr>
<td>7:</td>
<td>Ihtair seudolabels mas</td>
</tr>
<tr>
<td>8:</td>
<td>Calculate $\mathcal{L}_\mathrm{uu}^{u}$ and $\mathcal{L}_\mathrm{upper}^c$ by Eqs. (11) and $^{\prime }$upper $(14)\colon$</td>
</tr>
<tr>
<td>9:</td>
<td>Calculate $rl$ $•u$</td>
</tr>
<tr>
<td>10:</td>
<td>Calculate bv Eq. ( 17) ; </td>
</tr>
<tr>
<td>11: </td>
<td>Calculate 01 二 $r$ $+\lambda L.$ ower</td>
</tr>
<tr>
<td>12:</td>
<td>Undate </td>
</tr>
<tr>
<td>13:</td>
<td>$emorh=$ $noch+1:$</td>
</tr>
<tr>
<td>14:</td>
<td>end for</td>
</tr>
<tr>
<td>15:</td>
<td>end while</td>
</tr>
</tbody>
</table>
式，与仅预测概率分布相比，可以更深人地了解数据中的关系。因此，我们利用对应于两个增强视图的潜在特征之间的互信息来近似熵的下界。正对数的相似性$d(\cdot,\cdot)$ 由高斯函数E[Tian et al., 2023] ,
$$d(\boldsymbol{z}_i^{(1)},\boldsymbol{z}_i^{(2)}|\theta)=\exp\left(-\left\|\boldsymbol{z}_i^{(1)}-\boldsymbol{z}_i^{(2)}\right\|_2^2\right),$$
(16)
定义，并且等式$\mathbb{E}(16)$ 中定义的相似性意味着如果任何两个对数相同，则相似性为 1; 否则，随着它们的距离的增加，它趋于零。将方程$\mathbb{E}(16)$ 代人方程$\mathbb{E}(10)$ ,我们得到
(17)
$$\mathcal{L}_{\mathrm{lower}}=\frac{1}{n_u}\sum_{i=1}^{n_u}\left\|\boldsymbol{z}_i^{(1)}-\boldsymbol{z}_i^{(2)}\right\|_2^2.$$
4.3 InfoMatch Algorithm
我们引人了一种创新的半监督图像分类算法，称为 InfoMatch, 它将分类任务视为近似问题。对于标记的数据我们利用交叉熵损失函数作为后验熵的上限。InfoMatch 主要致力于有效利用大量未标记的数据在 InfoMatch 算法中，我们使用伪监督和弱到强策略启动。这使我们能够将熵上限转换为伪监督交叉熵损失，作为未标记数据的监督信号。此外，为了丰富数据集的多样性，我们引入了一种新的强增强方法 CutMix,该方法通过制定相应的损失来增强其泛化能力。随后，我们将两个增强视图之间的潜在特征对比损失视为数据集熵的下限。通过最大化这个下限并最小化上限，我们逐渐优化 InfoMatch, 即使有限的标记数据可用。
总体目标函数结合了方程$\mathbb{E}(15)$ 的三个交叉熵损失和
方程$\mathbb{E}(17)$ 的对比损失，
$$\mathcal{L}=\mathcal{L}_{\mathrm{upper}}+\lambda\mathcal{L}_{\mathrm{lower}}$$
(18)
的下界为负相对边界。在实验过程中，我们采用小批量进行训练，分别使用批量大小为$n_l^b$和$n_u^b$的标记和未标记样本。算法 1 提供了学习过程的全面概述。
5 Experimental Results
5.1 Experimental Setup
我们在著名的基准数据集上评估 InfoMatch,包括 CIFAR-10/100 [Krizhevsky and Hinton, 2009] . THMAHST。此外，我们使用不同数量的标记数据进行实验为了确保公平比较，我们按照 FixMatch[Sohn et al., 2020] 和 FreeMatch [Wang et al., 2023] 中的实验设置。具体来说，我们采用标准的随机梯度下降算法，余弦学习率衰减作为所有数据集的优化器，初始学习率为0。03 和动量为0。9。对于所有实验，我们将总迭代次数设置为$2^{20}$。然后使用参数为 0 的 EMA 评估 InfoMatch 性能。999。此外，对于 ImageNet,我们为标记和未标记的样本 i.e. $n_l^b=n_u^b=128$保持 128 的批量大小，并利用 ResNet-50 架构 [He et al., 2016] 。对于其他数据集我们将批处理大小调整为$n_l^b=64$和$n_u^b=448$,并采用WideResNet 变体，例如 WideResNet-28-2 [Zagoruykon] and Komodakis, 2016] 和 WideResNet-28-8 [Zhou et al., 2020]。为了评估 InfoMatch 中损失方程$\mathbb{E}(18)$ 的两个不同项的影响，我们引人了一个非负的超参数$\lambda$。然后， 我们确保似然函数中实数和伪标记的公平表示，并保持RandAugment$\boxed{\mathrm{E}}[$Cubuk et al.,2020] 和 CutMix$\boxed{\mathrm{E}}[$Yun et al., 2019] 之间的平衡。随后，我们将调节熵界的参数$\lambda$调整为 0.002 。此外，我们采用不同的阈值选择策略来消除伪标签错误对 InfoMatch 的影响。对于 SVHN,我们采用固定阈值0。95。相反，对于其他数据集，我们利用 FreeMatch 提出的自适应阈值化方法，该方法通过将当前批次中所有未标记数据的最高置信度的期望作为全局阈值来衡量整体学习进度，并通过计算每个类对应的平均预测概率来评估特定类的学习状态最后，为了保证精确和无偏的评估，我们使用各种随机种子对每个模型进行多次训练，并计算达到的最佳精度的平均值和标准$\bar{\text{差。}}$
### 5.2 Main Results
我们将 InfoMatch 与全监督学习方法和一系列代表性的半监督学习方法进行了比较，包括基于伪标签的方法。例如 FlexMatch[E[Zhang et al.,2021] 和 FreeMatch[E] [Wang et al., 2023], 以及基于图形的方法，例如 Sim- $\begin{array}{l}{\mathrm{Match}}&{\boldsymbol{y}=\sigma(c_{j}|\boldsymbol{x}_{i},\theta)}&{\text{。选项卡中显示了 CIFAR-}}\\{10/100}&{\mathrm{Gurmer~fu~Gur~10~}}&{\text{。选项卡中显示了 CIFAR-}}\end{array}$ 10/100. SVHN 和 STT-10 在各种标记数据大小下的Top-1 错误率。1 和 3 。根据这些结果，很明显 InfoMatch 在所有基准测试中都表现出卓越的性能，在四个数据集上平均提高了 1.49% 、$0.55\%$ 、$0.13\%$和 2.12% 的性能。特别是，InfoMatch 显著降低了 CIFAR-10 数据集中从 8.07% 到 4.39%( -3.68%)的平均 Top-1 错误率，该数据集仅包含 10 个标记数据。在具有 40 个标记数据的 STL-10 数据集中时，Top-1 错误率从 13.74% 降低到 9.86%( -3.88% )。此外，在 CIFAR-10 和 SVHN 数据集上，我们的方法实现了显著的性能改进，不仅超过了其他半监督基线，而且在所有基准测试中都优于全监督学习方法。此外，与其他方法相比，InfoMatch 在具有不同种子的多个实验中表现出更小的标准偏差，从而突出了 InfoMatch 的卓越稳定性和鲁棒性值得注意的是当标记数据的数量非常有限时，InfoMatch 的性能明显优于其他方法。特别是在 CIFAR-10 数据集中，即使每个类只有一个标记数据可用，平均 Top-1 错误率也达到$4. 39\%$ , 低于完全监督学习方法 $( - 0. 23\%$ )。 | [
[
"0.0866",
"0.0682",
"0.3635",
"0.06946",
"0.3635",
"0.084",
"0.0866",
"0.0833"
],
[
"0.513",
"0.06946",
"0.5513",
"0.06946",
"0.5513",
"0.0808",
"0.513",
"0.0808"
],
[
"0.544",
"0.06946",
"0.7295",
"0.06946",
"0.7295",
"0.0833",
"0.544",
"0.0833"
],
[
"0.09235",
"0.09155",
"0.4885",
"0.09155",
"0.4885",
"0.1048",
"0.09235",
"0.1048"
],
[
"0.511",
"0.09155",
"0.716",
"0.09344",
"0.716",
"0.108",
"0.511",
"0.1061"
],
[
"0.1136",
"0.1061",
"0.1365",
"0.1061",
"0.1365",
"0.1187",
"0.1136",
"0.1187"
],
[
"0.09235",
"0.1187",
"0.3816",
"0.1187",
"0.3816",
"0.132",
"0.09235",
"0.132"
],
[
"0.513",
"0.11426",
"0.9126",
"0.11554",
"0.9126",
"0.13",
"0.513",
"0.1288"
],
[
"0.0948",
"0.1332",
"0.437",
"0.1332",
"0.437",
"0.1465",
"0.0948",
"0.1465"
],
[
"0.511",
"0.1276",
"0.898",
"0.1288",
"0.898",
"0.1433",
"0.511",
"0.1427"
],
[
"0.09314",
"0.1465",
"0.3218",
"0.1465",
"0.3218",
"0.161",
"0.09314",
"0.161"
],
[
"0.513",
"0.1433",
"0.911",
"0.1433",
"0.911",
"0.1572",
"0.513",
"0.1572"
],
[
"0.1283",
"0.1604",
"0.3628",
"0.1604",
"0.3628",
"0.1736",
"0.1283",
"0.1736"
],
[
"0.5146",
"0.1572",
"0.911",
"0.1572",
"0.911",
"0.1704",
"0.5146",
"0.1704"
],
[
"0.1152",
"0.1642",
"0.125",
"0.1642",
"0.125",
"0.1674",
"0.1152",
"0.1674"
],
[
"0.513",
"0.1704",
"0.911",
"0.1704",
"0.911",
"0.1843",
"0.513",
"0.1843"
],
[
"0.09314",
"0.161",
"0.1177",
"0.161",
"0.1177",
"0.1724",
"0.09314",
"0.1724"
],
[
"0.09235",
"0.1749",
"0.1111",
"0.1749",
"0.1111",
"0.1875",
"0.09235",
"0.1875"
],
[
"0.1462",
"0.1724",
"0.4885",
"0.1718",
"0.4885",
"0.1863",
"0.1462",
"0.1875"
],
[
"0.513",
"0.1831",
"0.911",
"0.1843",
"0.911",
"0.1989",
"0.513",
"0.1976"
],
[
"0.1487",
"0.1901",
"0.1765",
"0.1901",
"0.1765",
"0.1989",
"0.1487",
"0.1989"
],
[
"0.1462",
"0.2008",
"0.4836",
"0.202",
"0.4836",
"0.2166",
"0.1462",
"0.2153"
],
[
"0.5146",
"0.1989",
"0.911",
"0.1989",
"0.911",
"0.2122",
"0.5146",
"0.2122"
],
[
"0.513",
"0.2108",
"0.911",
"0.2122",
"0.911",
"0.226",
"0.513",
"0.2247"
],
[
"0.09235",
"0.202",
"0.1111",
"0.202",
"0.1111",
"0.2153",
"0.09235",
"0.2153"
],
[
"0.1447",
"0.2216",
"0.2695",
"0.2267",
"0.2688",
"0.2488",
"0.1438",
"0.2437"
],
[
"0.1478",
"0.2172",
"0.183",
"0.2172",
"0.183",
"0.2291",
"0.1478",
"0.2291"
],
[
"0.09235",
"0.2299",
"0.1095",
"0.2299",
"0.1095",
"0.2437",
"0.09235",
"0.2437"
],
[
"0.2573",
"0.2311",
"0.3472",
"0.2311",
"0.3472",
"0.2424",
"0.2573",
"0.2424"
],
[
"0.5156",
"0.226",
"0.9126",
"0.226",
"0.9126",
"0.2393",
"0.5156",
"0.2393"
],
[
"0.1478",
"0.2437",
"0.4763",
"0.2437",
"0.4763",
"0.2568",
"0.1478",
"0.2568"
],
[
"0.5146",
"0.2393",
"0.91",
"0.2393",
"0.91",
"0.254",
"0.5146",
"0.254"
],
[
"0.1438",
"0.254",
"0.4885",
"0.2551",
"0.4885",
"0.276",
"0.1438",
"0.2747"
],
[
"0.0948",
"0.2443",
"0.10864",
"0.2443",
"0.10864",
"0.2556",
"0.0948",
"0.2556"
],
[
"0.09235",
"0.2568",
"0.1111",
"0.2568",
"0.1111",
"0.2715",
"0.09235",
"0.2715"
],
[
"0.5146",
"0.254",
"0.911",
"0.254",
"0.911",
"0.2676",
"0.5146",
"0.2676"
],
[
"0.1503",
"0.2734",
"0.1854",
"0.2734",
"0.1854",
"0.2854",
"0.1503",
"0.2854"
],
[
"0.5146",
"0.2676",
"0.9087",
"0.2683",
"0.9087",
"0.2827",
"0.5146",
"0.2822"
],
[
"0.1438",
"0.2822",
"0.46",
"0.2854",
"0.46",
"0.3062",
"0.1438",
"0.303"
],
[
"0.5146",
"0.2827",
"0.911",
"0.2827",
"0.911",
"0.2966",
"0.5146",
"0.2966"
],
[
"0.513",
"0.2966",
"0.9126",
"0.2986",
"0.9126",
"0.3125",
"0.513",
"0.31"
],
[
"0.0907",
"0.286",
"0.1095",
"0.286",
"0.1095",
"0.2998",
"0.0907",
"0.2998"
],
[
"0.0866",
"0.3018",
"0.1095",
"0.3018",
"0.1095",
"0.315",
"0.0866",
"0.315"
],
[
"0.1462",
"0.3005",
"0.357",
"0.3005",
"0.357",
"0.318",
"0.1462",
"0.318"
],
[
"0.1447",
"0.3132",
"0.3718",
"0.3132",
"0.3718",
"0.334",
"0.1447",
"0.334"
],
[
"0.513",
"0.3125",
"0.911",
"0.3125",
"0.911",
"0.3257",
"0.513",
"0.3257"
],
[
"0.513",
"0.3257",
"0.911",
"0.3257",
"0.911",
"0.339",
"0.513",
"0.339"
],
[
"0.0866",
"0.3164",
"0.10864",
"0.3164",
"0.10864",
"0.3289",
"0.0866",
"0.3289"
],
[
"0.0866",
"0.33",
"0.1095",
"0.33",
"0.1095",
"0.3435",
"0.0866",
"0.3435"
],
[
"0.1462",
"0.3289",
"0.4265",
"0.33",
"0.4265",
"0.3447",
"0.1462",
"0.3435"
],
[
"0.0866",
"0.3435",
"0.1095",
"0.3435",
"0.1095",
"0.356",
"0.0866",
"0.356"
],
[
"0.1462",
"0.3435",
"0.2893",
"0.3435",
"0.2893",
"0.3567",
"0.1462",
"0.3567"
],
[
"0.5146",
"0.3384",
"0.911",
"0.3403",
"0.911",
"0.3547",
"0.5146",
"0.3528"
],
[
"0.08826",
"0.358",
"0.1111",
"0.358",
"0.1111",
"0.3694",
"0.08826",
"0.3694"
],
[
"0.1283",
"0.358",
"0.1814",
"0.358",
"0.1814",
"0.3694",
"0.1283",
"0.3694"
],
[
"0.5146",
"0.3542",
"0.911",
"0.3542",
"0.911",
"0.3674",
"0.5146",
"0.3674"
],
[
"0.08826",
"0.3726",
"0.183",
"0.3726",
"0.183",
"0.3838",
"0.08826",
"0.3838"
],
[
"0.513",
"0.3674",
"0.9126",
"0.3674",
"0.9126",
"0.3806",
"0.513",
"0.3806"
],
[
"0.513",
"0.3818",
"0.911",
"0.3818",
"0.911",
"0.3958",
"0.513",
"0.3958"
],
[
"0.5146",
"0.3958",
"0.911",
"0.3958",
"0.911",
"0.4092",
"0.5146",
"0.4092"
],
[
"0.08905",
"0.4092",
"0.4836",
"0.4092",
"0.4836",
"0.4192",
"0.08905",
"0.4192"
],
[
"0.5146",
"0.4092",
"0.91",
"0.4092",
"0.91",
"0.4224",
"0.5146",
"0.4224"
],
[
"0.08826",
"0.4216",
"0.486",
"0.4216",
"0.486",
"0.435",
"0.08826",
"0.435"
],
[
"0.5146",
"0.4236",
"0.911",
"0.4236",
"0.911",
"0.4375",
"0.5146",
"0.4375"
],
[
"0.08826",
"0.435",
"0.4836",
"0.435",
"0.4836",
"0.449",
"0.08826",
"0.449"
],
[
"0.513",
"0.4363",
"0.911",
"0.4375",
"0.911",
"0.4507",
"0.513",
"0.4495"
],
[
"0.08905",
"0.4495",
"0.326",
"0.4495",
"0.326",
"0.4634",
"0.08905",
"0.4634"
],
[
"0.5146",
"0.449",
"0.911",
"0.4495",
"0.911",
"0.464",
"0.5146",
"0.4634"
],
[
"0.1405",
"0.4653",
"0.4224",
"0.4685",
"0.4224",
"0.4937",
"0.1405",
"0.4905"
],
[
"0.5156",
"0.464",
"0.911",
"0.464",
"0.911",
"0.4778",
"0.5156",
"0.4778"
],
[
"0.4143",
"0.4749",
"0.4836",
"0.473",
"0.4836",
"0.4873",
"0.4143",
"0.4893"
],
[
"0.513",
"0.4766",
"0.911",
"0.4778",
"0.911",
"0.4924",
"0.513",
"0.4912"
],
[
"0.5146",
"0.4924",
"0.911",
"0.4924",
"0.911",
"0.506",
"0.5146",
"0.506"
],
[
"0.3308",
"0.4797",
"0.3506",
"0.4797",
"0.3506",
"0.4841",
"0.3308",
"0.4841"
],
[
"0.08826",
"0.5",
"0.4836",
"0.5",
"0.4836",
"0.513",
"0.08826",
"0.513"
],
[
"0.513",
"0.506",
"0.911",
"0.506",
"0.911",
"0.5195",
"0.513",
"0.5195"
],
[
"0.08826",
"0.513",
"0.4836",
"0.513",
"0.4836",
"0.528",
"0.08826",
"0.528"
],
[
"0.513",
"0.5215",
"0.5376",
"0.5215",
"0.5376",
"0.534",
"0.513",
"0.534"
],
[
"0.08826",
"0.528",
"0.4844",
"0.528",
"0.4844",
"0.5415",
"0.08826",
"0.5415"
],
[
"0.08826",
"0.5425",
"0.1217",
"0.5425",
"0.1217",
"0.5537",
"0.08826",
"0.5537"
],
[
"0.513",
"0.549",
"0.665",
"0.549",
"0.665",
"0.5625",
"0.513",
"0.5625"
],
[
"0.2378",
"0.56",
"0.3865",
"0.565",
"0.3857",
"0.5933",
"0.2361",
"0.5884"
],
[
"0.2695",
"0.565",
"0.2769",
"0.565",
"0.2769",
"0.5693",
"0.2695",
"0.5693"
],
[
"0.179",
"0.5728",
"0.2484",
"0.5728",
"0.2484",
"0.589",
"0.179",
"0.589"
],
[
"0.4558",
"0.5728",
"0.4844",
"0.5728",
"0.4844",
"0.5854",
"0.4558",
"0.5854"
],
[
"0.5146",
"0.571",
"0.911",
"0.571",
"0.911",
"0.584",
"0.5146",
"0.584"
],
[
"0.2345",
"0.577",
"0.2932",
"0.5884",
"0.2876",
"0.603",
"0.2296",
"0.5913"
],
[
"0.3733",
"0.5806",
"0.3865",
"0.5806",
"0.3865",
"0.5854",
"0.3733",
"0.5854"
],
[
"0.5146",
"0.5835",
"0.9087",
"0.5854",
"0.9087",
"0.5996",
"0.5146",
"0.5977"
],
[
"0.5156",
"0.5986",
"0.911",
"0.5986",
"0.911",
"0.6123",
"0.5156",
"0.6123"
],
[
"0.0907",
"0.6113",
"0.1136",
"0.6113",
"0.1136",
"0.6187",
"0.0907",
"0.6187"
],
[
"0.11273",
"0.607",
"0.2998",
"0.607",
"0.2998",
"0.6206",
"0.11273",
"0.6206"
],
[
"0.5156",
"0.6113",
"0.911",
"0.6123",
"0.911",
"0.627",
"0.5156",
"0.6255"
],
[
"0.08826",
"0.6255",
"0.4844",
"0.6255",
"0.4844",
"0.6396",
"0.08826",
"0.6396"
],
[
"0.513",
"0.6255",
"0.91",
"0.6255",
"0.91",
"0.64",
"0.513",
"0.64"
],
[
"0.08496",
"0.638",
"0.4844",
"0.6396",
"0.4844",
"0.654",
"0.08496",
"0.653"
],
[
"0.5146",
"0.6396",
"0.911",
"0.6396",
"0.911",
"0.653",
"0.5146",
"0.653"
],
[
"0.08826",
"0.654",
"0.486",
"0.654",
"0.486",
"0.6675",
"0.08826",
"0.6675"
],
[
"0.5146",
"0.655",
"0.6323",
"0.655",
"0.6323",
"0.6655",
"0.5146",
"0.6655"
],
[
"0.634",
"0.654",
"0.911",
"0.654",
"0.911",
"0.6675",
"0.634",
"0.6675"
],
[
"0.08905",
"0.6675",
"0.486",
"0.6675",
"0.486",
"0.681",
"0.08905",
"0.681"
],
[
"0.513",
"0.6675",
"0.911",
"0.6675",
"0.911",
"0.681",
"0.513",
"0.681"
],
[
"0.08826",
"0.681",
"0.486",
"0.681",
"0.486",
"0.6943",
"0.08826",
"0.6943"
],
[
"0.5146",
"0.6816",
"0.911",
"0.6816",
"0.911",
"0.696",
"0.5146",
"0.696"
],
[
"0.08905",
"0.6963",
"0.4844",
"0.6963",
"0.4844",
"0.707",
"0.08905",
"0.707"
],
[
"0.5146",
"0.6943",
"0.911",
"0.6943",
"0.911",
"0.709",
"0.5146",
"0.709"
],
[
"0.08826",
"0.709",
"0.4844",
"0.709",
"0.4844",
"0.7227",
"0.08826",
"0.7227"
],
[
"0.5146",
"0.709",
"0.911",
"0.709",
"0.911",
"0.7227",
"0.5146",
"0.7227"
],
[
"0.09235",
"0.7217",
"0.4844",
"0.7217",
"0.4844",
"0.735",
"0.09235",
"0.735"
],
[
"0.517",
"0.726",
"0.544",
"0.726",
"0.544",
"0.733",
"0.517",
"0.733"
],
[
"0.54",
"0.7236",
"0.911",
"0.7236",
"0.911",
"0.7373",
"0.54",
"0.7373"
],
[
"0.08826",
"0.7363",
"0.486",
"0.7363",
"0.486",
"0.75",
"0.08826",
"0.75"
],
[
"0.513",
"0.7373",
"0.911",
"0.7373",
"0.911",
"0.7505",
"0.513",
"0.7505"
],
[
"0.08826",
"0.75",
"0.4844",
"0.75",
"0.4844",
"0.763",
"0.08826",
"0.763"
],
[
"0.5146",
"0.7505",
"0.9087",
"0.7505",
"0.9087",
"0.761",
"0.5146",
"0.761"
],
[
"0.08905",
"0.763",
"0.4878",
"0.763",
"0.4878",
"0.7764",
"0.08905",
"0.7764"
],
[
"0.5156",
"0.7666",
"0.911",
"0.7666",
"0.911",
"0.7764",
"0.5156",
"0.7764"
],
[
"0.0866",
"0.7764",
"0.2737",
"0.778",
"0.2737",
"0.7925",
"0.0866",
"0.7915"
],
[
"0.5146",
"0.779",
"0.911",
"0.779",
"0.911",
"0.7925",
"0.5146",
"0.7925"
],
[
"0.1054",
"0.7925",
"0.486",
"0.7925",
"0.486",
"0.806",
"0.1054",
"0.806"
],
[
"0.5156",
"0.7925",
"0.911",
"0.7925",
"0.911",
"0.806",
"0.5156",
"0.806"
],
[
"0.0866",
"0.806",
"0.259",
"0.806",
"0.259",
"0.8193",
"0.0866",
"0.8193"
],
[
"0.5146",
"0.806",
"0.911",
"0.806",
"0.911",
"0.8193",
"0.5146",
"0.8193"
],
[
"0.2075",
"0.8237",
"0.3628",
"0.8247",
"0.3628",
"0.8423",
"0.2075",
"0.8413"
],
[
"0.4558",
"0.827",
"0.4844",
"0.827",
"0.4844",
"0.839",
"0.4558",
"0.839"
],
[
"0.5156",
"0.8213",
"0.9067",
"0.8213",
"0.9067",
"0.832",
"0.5156",
"0.832"
],
[
"0.5146",
"0.832",
"0.911",
"0.834",
"0.911",
"0.8477",
"0.5146",
"0.845"
],
[
"0.08826",
"0.8467",
"0.4836",
"0.8467",
"0.4836",
"0.8613",
"0.08826",
"0.8613"
],
[
"0.5146",
"0.8477",
"0.91",
"0.8477",
"0.91",
"0.8613",
"0.5146",
"0.8613"
],
[
"0.08826",
"0.8623",
"0.486",
"0.8623",
"0.486",
"0.8755",
"0.08826",
"0.8755"
],
[
"0.5146",
"0.8613",
"0.91",
"0.8613",
"0.91",
"0.8755",
"0.5146",
"0.8755"
],
[
"0.08826",
"0.8755",
"0.411",
"0.8755",
"0.411",
"0.8896",
"0.08826",
"0.8896"
],
[
"0.513",
"0.8735",
"0.835",
"0.875",
"0.835",
"0.8896",
"0.513",
"0.888"
]
] | [
"Algorithm 1 The InfoMatch algorithm.",
"5",
"Experimental Results",
"Input: X = X,UXu, batch size ni and n,, parameter",
"5.1 Experimental Setup",
">:",
"Output:Optimal model parameters 0*",
"我们在著名的基准数据集上评估InfoMatch，包",
"1: Initialization: epoch = 0 , epochmax , and .",
"括 CIFAR-10/100[Krizhevsky and Hinton,2009]",
"2:while epoch ≤ epochmax do",
"THMAHST。此外，我们使用不同数量的标记数据进",
"for mini-batch samples in X do",
"行实验为了确保公平比较，我们按照FixMatch[Sohnet",
"",
"al.,2020]和FreeMatch[Wangetal.,2023]中的实验设",
"3:",
"4:",
"Obtain augmented images ', &(1) , (2) and",
"置。具体来说，我们采用标准的随机梯度下降算法，余弦",
"rc:",
"Feed samples into model for forward propaga-",
"学习率衰减作为所有数据集的优化器，初始学习率为0",
"03和动量为0。9。对于所有实验，我们将总选代次数设",
"5:",
"Calculate Cupper",
"tion;",
"6:",
"rbyEq.(8):",
"置为220。然后使用参数为0的EMA评估InfoMatch",
"Obtain the mask M for filtering pseudolabels;",
"性能。999。此外，对于ImageNet，我们为标记和未标记",
"Calculate Cupper and Cpper by Eqs. (11) and",
"7:",
"8:",
"的样本i.e.、n=nb=128保持128的批量大小，并利",
"(14);",
"用ResNet-50架构[Heetal.，2016]。对于其他数据集",
"Calculate Cupper = Cupper + Cupper + Cupper",
"我们将批处理大小调整为nb=64和nb=448，并采用",
"WideResNet变体，例如WideResNet-28-2[Zagoruyko",
"9:",
"10:",
"Calculate Llower by Eq. (17);",
"Calculate C= Cupper + Ciower",
"and Komodakis, 2016] 和 WideResNet-28-8 [Zhou et al.,",
"2020。为了评估InfoMatch中损失方程E18）的两个",
"11:",
"12:",
"Update parameter by minimizing C ;",
"13:",
"epoch = epoch +1 ;",
"不同项的影响，我们引人了一个非负的超参数入。然后，",
"14:",
"end for",
"我们确保似然函数中实数和伪标记的公平表示，并保持",
"15: end while",
"RandAugmentE[Cubuk et al., 2020] 和 CutMixE[Yun",
"etal.，2019]之间的平衡。随后，我们将调节熵界的参数",
"入调整为0.002。此外，我们采用不同的值选择策略来",
"式，与仅预测概率分布相比，可以更深入地了解数据中",
"消除伪标签错误对InfoMatch的影响。对于SVHN，我",
"的关系。因此，我们利用对应于两个增强视图的潜在特",
"们采用固定值0。95。相反，对于其他数据集，我们利",
"征之间的互信息来近似摘的下界。正对数的相似性d，）",
"用FreeMatch提出的自适应阈值化方法，该方法通过将",
"由高斯函数E[Tianetal.,2023]，",
"当前批次中所有未标记数据的最高置信度的期望作为全",
"d(2(1),2(2)]0) = exp(-2(1) - 2(2/2)",
"局阈值来衡量整体学习进度，并通过计算每个类对应的",
"，(16)",
"平均预测概率来评估特定类的学习状态最后，为了保证",
"精确和无偏的评估，我们使用各种随机种子对每个模型",
"",
"定义，并且等式E（16）中定义的相似性意味着如果任何",
"进行多次训练，并计算达到的最佳精度的平均值和标准",
"两个对数相同，则相似性为1；否则，随着它们的距离的",
"差。",
"增加，它趋于零。将方程E(16）代人方程E(10），我们",
"得到",
"5.2Main Results",
"12)-22",
"14",
"Clower =,",
"(17)",
"我们将InfoMatch与全监督学习方法和一系列代表性的",
"nu=l",
"2",
"半监督学习方法进行了比较，包括基于伪标签的方法",
"例如 FlexMatchE[Zhang et al., 2021] 和 FreeMatchF",
"4.3",
"InfoMatchAlgorithm",
"[Wangetal.，2023]，以及基于图形的方法，例如Sim-",
"我们引人了一种创新的半监督图像分类算法，称为Info-",
"Match y=a(cilai,の）_。选项卡中显示了CIFAR-",
"Match，它将分类任务视为近似问题。对于标记的数据，",
"10/100、SVHN和STL-10在各种标记数据大小下的",
"我们利用交叉摘损失函数作为后验摘的上限。InfoMatch",
"Top-1错误率。",
"1和3。根据这些结果，很明显Info-",
"主要致力于有效利用大量未标记的数据在InfoMatch算",
"Match在所有基准测试中都表现出卓越的性能，在四个",
"法中，我们使用伪监督和弱到强策略启动。这使我们能",
"数据集上平均提高了1.49%、0.55%、0.13%和2.12%",
"够将摘上限转换为伪监督交叉摘损失，作为未标记数据",
"的性能。特别是，InfoMatch显著降低了CIFAR-10 数据",
"的监督信号。此外，为了丰富数据集的多样性，我们引人",
"集中从8.07%到4.39%（-3.68%）的平均Top-1错误",
"了一种新的强增强方法CutMix，该方法通过制定相应",
"",
"该数据集仅包含10个标记数据。在具有40个标记",
"的损失来增强其泛化能力。随后，我们将两个增强视图",
"数据的STL-10数据集中时，Top-1错误率从13.74%降",
"之间的潜在特征对比损失视为数据集熵的下限。通过最",
"低到9.86%（-3.88%）。此外在CIFAR-10和SVHN",
"大化这个下限并最小化上限，我们逐渐优化InfoMatch，",
"数据集上，我们的方法实现了显著的性能改进，不仅超",
"即使有限的标记数据可用。",
"过了其他半监督基线，而且在所有基准测试中都优于全",
"总体目标函数结合了方程E15）的三个交叉摘损失和",
"监督学习方法。此外，与其他方法相比，InfoMatch在具",
"方程E（17）的对比损失，",
"有不同种子的多个实验中表现出更小的标准偏差，从而",
"= Lupper + 入Clower",
"(18)",
"突出了InfoMatch的卓越稳定性和鲁棒性值得注意的是",
"当标记数据的数量非常有限时，InfoMatch的性能明显",
"的下界为负相对边界。在实验过程中，我们采用小批量",
"优于其他方法。特别是在CIFAR-10数据集中，即使每",
"进行训练，分别使用批量大小为n和n的标记和未标",
"个类只有一个标记数据可用，平均Top-1错误率也达到",
"记样本。算法1提供了学习过程的全面概述。",
"4.39%，低于完全监督学习方法（-0.23%）。"
] | {
"arxiv_id": "2404.11003",
"language": "zh",
"page_idx": 4
} |
||
2304.02163-zh-0 | 2304.02163-zh-0.jpg | GINA-3D: Learning to Generate Implicit Neural Assets in the Wild
$\begin{array}{cc}\text{Bokui Shen }^{1*}&\text{Xinchen Yan}^2\text{Charles R. Qi}^2&\text{Mahyar Najibi}^2\text{Boyang Deng }^{1,2\dagger}\\&\text{Leonidas Guibas }^3\text{ Yin Zhou}^2&\text{Dragomir Anguelov}^2\end{array}$
<FigureHere>
Figure 1. Leveraging in-the-wild data for generative assets modeling embodies a scalable approach for simulation. GINA-3D uses realworld driving data to perform various synthesis tasks for realistic 3D implicit neural assets. Left: Multi-sensor observations in the wild. Middle: Asset reconstruction and conditional synthesis. Right: Scene composition with background neural fields[1].
# Abstract
ages and geometries.
## $1. \textbf{ Introduction}$
Modeling the 3D world from sensor data for simulation is a scalable way of developing testing and validation environments for robotic learning problems such as autonomous driving. However, manually creating or recreating real-world-like environments is difficult, expensive, and not scalable. Recent generative model techniques have shown promising progress to address such challenges by learning 3D assets using only plentiful 2D images-but still suffer limitations as they leverage either human-curated image datasets or renderings from manually-created synthetic 3D environments. In this paper, we introduce GINA-3D, a generative model that uses real-world driving data from camera and LiDAR sensors to create realistic 3D implicit neural assets of diverse vehicles and pedestrians. Compared to the existing image datasets, the real-world driving setting poses new challenges due to occlusions, lightingvariations and long-tail distributions. GINA-3D tackles these challenges by decoupling representation learning and generative modeling into two stages with a learned tri-plane latent structure, inspired by recent advances in generative modeling of images. To evaluate our approach, we construct a large-scale object-centric dataset containing over 1.2M images of vehicles and pedestrians from the Waymo $Open\textit{Dataset, and a new set of 80K images of long- tail}$ instances such as construction equipment, garbage trucks, and cable cars. We compare our model with existing approaches and demonstrate that it achieves state-of-the-art performance in quality and diversity for both generated im-
几十年来，学习感知、推理和与 3D 世界互动一直是计算机视觉和机器人领域的长期挑战[2-9]。部署在野外的现代机器人系统$\mathbb{E}[10-16]$ 通常配备有多个传感器$(e.g.$摄像机，LiDAR 和雷达),可以感知 3D 环境，然后是用于推理和与复杂场景动态交互的智能单元。在现实环境中对这些智能代理进行端到端测试和验证既困难又昂贵，特别是在安全关键和资源受限的领域，如自动驾驶
另一方面，在过去几年中，模拟数据的使用激增，以安全，可扩展和可验证的方式在受控设置[E[17-27]下训练和评估智能代理。计算机图形学的飞速发展推动了这种发展，包括渲染框架[28-30] ,物理仿真[E[31,32] 和大规模开源资产存储库$\mathbb{E}[$33-39]。一个关键问题是创建逼真的虚拟世界，使资产内容，组成和行为与真实分布保持一致，以便使从业者相信使用此类模拟进行开发和验证可以转移到现实世界中的性能$\mathbb{E}[40-48]$ 。然而，手动资产创建面临两个主要障碍。首先，手动创建 3D 资产需要具有 3D 领域专业知识的工程师和艺术家的不懈努力，这是昂贵且难以扩展的$\mathbb{E}[26]$。其次，现实世界的分布包含各种各样的例子(包括有趣的罕见案例),并且也在不断发展[49,50]。生成 3D 建模的最新发展为解决上述障碍提供了新的视角，因为它允许生成其他逼真但以前看不见的示例。这些方法的一个子类，即生成的 3D 感知图像合成$\overline{\mathrm{E}}[51,52]$,具有重要的前景，因为它可以通过部分观察(3D 对象的$e.g.$图像投影)进行 3D 建模此外，许多现实世界的机器人应用程序已经大规模捕获，注释和更新多传感器
$^{*}$ Work done during an internship at Waymo. $^\dagger$ Work done at Waymo.
1 | [
[
"0.1487",
"0.132",
"0.822",
"0.1332",
"0.822",
"0.1509",
"0.1487",
"0.1497"
],
[
"0.2632",
"0.1818",
"0.3896",
"0.1799",
"0.3896",
"0.1976",
"0.2632",
"0.1995"
],
[
"0.4028",
"0.1818",
"0.689",
"0.1818",
"0.689",
"0.1963",
"0.4028",
"0.1963"
],
[
"0.697",
"0.1818",
"0.841",
"0.1787",
"0.841",
"0.1976",
"0.697",
"0.2008"
],
[
"0.689",
"0.1875",
"0.6953",
"0.1875",
"0.6953",
"0.1913",
"0.689",
"0.1913"
],
[
"0.1299",
"0.1831",
"0.246",
"0.1812",
"0.246",
"0.1963",
"0.1299",
"0.1989"
],
[
"0.6953",
"0.1875",
"0.701",
"0.1875",
"0.701",
"0.1925",
"0.6953",
"0.1925"
],
[
"0.2573",
"0.1995",
"0.407",
"0.1989",
"0.407",
"0.2134",
"0.2573",
"0.214"
],
[
"0.4272",
"0.202",
"0.517",
"0.202",
"0.517",
"0.2134",
"0.4272",
"0.2134"
],
[
"0.5337",
"0.1989",
"0.7134",
"0.1976",
"0.7134",
"0.2153",
"0.5337",
"0.2166"
],
[
"0.2974",
"0.2153",
"0.6724",
"0.2166",
"0.6724",
"0.2343",
"0.2974",
"0.2329"
],
[
"0.634",
"0.2683",
"0.6626",
"0.2683",
"0.6626",
"0.2727",
"0.634",
"0.2727"
],
[
"0.697",
"0.2651",
"0.71",
"0.2651",
"0.71",
"0.2683",
"0.697",
"0.2683"
],
[
"0.3545",
"0.2734",
"0.4617",
"0.2734",
"0.4617",
"0.327",
"0.3545",
"0.327"
],
[
"0.626",
"0.2778",
"0.634",
"0.2778",
"0.634",
"0.2822",
"0.626",
"0.2822"
],
[
"0.564",
"0.3447",
"0.5947",
"0.3447",
"0.5947",
"0.348",
"0.564",
"0.348"
],
[
"0.6274",
"0.3423",
"0.6665",
"0.3423",
"0.6665",
"0.3499",
"0.6274",
"0.3499"
],
[
"0.1396",
"0.3528",
"0.2646",
"0.3528",
"0.2646",
"0.3599",
"0.1396",
"0.3599"
],
[
"0.492",
"0.3528",
"0.5947",
"0.3528",
"0.5947",
"0.3599",
"0.492",
"0.3599"
],
[
"0.6895",
"0.3499",
"0.8857",
"0.3484",
"0.8857",
"0.3623",
"0.6895",
"0.363"
],
[
"0.0801",
"0.3662",
"0.3882",
"0.3662",
"0.3882",
"0.3801",
"0.0801",
"0.3801"
],
[
"0.384",
"0.3655",
"0.8906",
"0.3655",
"0.8906",
"0.379",
"0.384",
"0.379"
],
[
"0.0809",
"0.3801",
"0.892",
"0.3801",
"0.892",
"0.3933",
"0.0809",
"0.3933"
],
[
"0.0784",
"0.3926",
"0.767",
"0.3933",
"0.767",
"0.4077",
"0.0784",
"0.4072"
],
[
"0.2385",
"0.4216",
"0.312",
"0.4216",
"0.312",
"0.4363",
"0.2385",
"0.4363"
],
[
"0.5024",
"0.4236",
"0.642",
"0.4236",
"0.642",
"0.4375",
"0.5024",
"0.4375"
],
[
"0.11273",
"0.44",
"0.4697",
"0.4414",
"0.4697",
"0.4558",
"0.11273",
"0.455"
],
[
"0.0809",
"0.457",
"0.4697",
"0.457",
"0.4697",
"0.4702",
"0.0809",
"0.4702"
],
[
"0.5024",
"0.4653",
"0.636",
"0.4666",
"0.636",
"0.481",
"0.5024",
"0.4797"
],
[
"0.0801",
"0.473",
"0.4697",
"0.473",
"0.4697",
"0.486",
"0.0801",
"0.486"
],
[
"0.0801",
"0.4873",
"0.468",
"0.4873",
"0.468",
"0.5005",
"0.0801",
"0.5005"
],
[
"0.524",
"0.4912",
"0.8906",
"0.4912",
"0.8906",
"0.505",
"0.524",
"0.505"
],
[
"0.0801",
"0.502",
"0.4673",
"0.502",
"0.4673",
"0.515",
"0.0801",
"0.515"
],
[
"0.5034",
"0.505",
"0.8906",
"0.507",
"0.8906",
"0.5215",
"0.5034",
"0.5195"
],
[
"0.0809",
"0.5176",
"0.468",
"0.5176",
"0.468",
"0.531",
"0.0809",
"0.531"
],
[
"0.5034",
"0.521",
"0.8906",
"0.5215",
"0.8906",
"0.536",
"0.5034",
"0.535"
],
[
"0.0801",
"0.532",
"0.468",
"0.531",
"0.468",
"0.5454",
"0.0801",
"0.547"
],
[
"0.5024",
"0.537",
"0.892",
"0.537",
"0.892",
"0.5503",
"0.5024",
"0.5503"
],
[
"0.0784",
"0.547",
"0.4697",
"0.5454",
"0.4697",
"0.56",
"0.0784",
"0.561"
],
[
"0.505",
"0.552",
"0.892",
"0.552",
"0.892",
"0.565",
"0.505",
"0.565"
],
[
"0.0801",
"0.5625",
"0.4697",
"0.561",
"0.4697",
"0.5757",
"0.0801",
"0.577"
],
[
"0.5034",
"0.5674",
"0.892",
"0.5674",
"0.892",
"0.5806",
"0.5034",
"0.5806"
],
[
"0.0809",
"0.5776",
"0.468",
"0.5776",
"0.468",
"0.5913",
"0.0809",
"0.5913"
],
[
"0.5063",
"0.584",
"0.8906",
"0.584",
"0.8906",
"0.5947",
"0.5063",
"0.5947"
],
[
"0.0801",
"0.5923",
"0.4673",
"0.5923",
"0.4673",
"0.607",
"0.0801",
"0.607"
],
[
"0.5063",
"0.5986",
"0.5703",
"0.5986",
"0.5703",
"0.61",
"0.5063",
"0.61"
],
[
"0.0801",
"0.608",
"0.468",
"0.608",
"0.468",
"0.622",
"0.0801",
"0.622"
],
[
"0.523",
"0.6157",
"0.8906",
"0.6157",
"0.8906",
"0.629",
"0.523",
"0.629"
],
[
"0.0801",
"0.6226",
"0.468",
"0.6226",
"0.468",
"0.636",
"0.0801",
"0.636"
],
[
"0.505",
"0.63",
"0.8896",
"0.63",
"0.8896",
"0.643",
"0.505",
"0.643"
],
[
"0.0801",
"0.638",
"0.468",
"0.638",
"0.468",
"0.6514",
"0.0801",
"0.6514"
],
[
"0.505",
"0.6445",
"0.888",
"0.6445",
"0.888",
"0.6577",
"0.505",
"0.6577"
],
[
"0.0801",
"0.653",
"0.468",
"0.653",
"0.468",
"0.6665",
"0.0801",
"0.6665"
],
[
"0.5034",
"0.66",
"0.8906",
"0.66",
"0.8906",
"0.674",
"0.5034",
"0.674"
],
[
"0.0801",
"0.6685",
"0.468",
"0.6685",
"0.468",
"0.6816",
"0.0801",
"0.6816"
],
[
"0.505",
"0.6753",
"0.8906",
"0.6753",
"0.8906",
"0.6895",
"0.505",
"0.6895"
],
[
"0.0809",
"0.683",
"0.4697",
"0.683",
"0.4697",
"0.6963",
"0.0809",
"0.6963"
],
[
"0.5024",
"0.6895",
"0.8906",
"0.69",
"0.8906",
"0.705",
"0.5024",
"0.704"
],
[
"0.0801",
"0.6987",
"0.4697",
"0.6987",
"0.4697",
"0.712",
"0.0801",
"0.712"
],
[
"0.5034",
"0.7056",
"0.8906",
"0.7056",
"0.8906",
"0.7197",
"0.5034",
"0.7197"
],
[
"0.0801",
"0.7134",
"0.468",
"0.712",
"0.468",
"0.7266",
"0.0801",
"0.728"
],
[
"0.5024",
"0.7197",
"0.8896",
"0.7197",
"0.8896",
"0.733",
"0.5024",
"0.733"
],
[
"0.0801",
"0.729",
"0.468",
"0.729",
"0.468",
"0.742",
"0.0801",
"0.742"
],
[
"0.5107",
"0.741",
"0.5215",
"0.741",
"0.5215",
"0.7456",
"0.5107",
"0.7456"
],
[
"0.5186",
"0.7373",
"0.8896",
"0.7373",
"0.8896",
"0.7476",
"0.5186",
"0.7476"
],
[
"0.0801",
"0.7437",
"0.468",
"0.7437",
"0.468",
"0.757",
"0.0801",
"0.757"
],
[
"0.505",
"0.7505",
"0.8906",
"0.7505",
"0.8906",
"0.765",
"0.505",
"0.765"
],
[
"0.0809",
"0.759",
"0.468",
"0.759",
"0.468",
"0.773",
"0.0809",
"0.773"
],
[
"0.505",
"0.7666",
"0.8906",
"0.7666",
"0.8906",
"0.78",
"0.505",
"0.78"
],
[
"0.0801",
"0.7734",
"0.468",
"0.7734",
"0.468",
"0.787",
"0.0801",
"0.787"
],
[
"0.505",
"0.7827",
"0.8896",
"0.7827",
"0.8896",
"0.7935",
"0.505",
"0.7935"
],
[
"0.0801",
"0.789",
"0.4697",
"0.789",
"0.4697",
"0.803",
"0.0801",
"0.803"
],
[
"0.5034",
"0.7944",
"0.8906",
"0.7954",
"0.8906",
"0.81",
"0.5034",
"0.8096"
],
[
"0.0801",
"0.8047",
"0.468",
"0.8047",
"0.468",
"0.8184",
"0.0801",
"0.8184"
],
[
"0.5034",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.5034",
"0.8247"
],
[
"0.0784",
"0.8174",
"0.468",
"0.8184",
"0.468",
"0.8335",
"0.0784",
"0.832"
],
[
"0.505",
"0.827",
"0.8906",
"0.827",
"0.8906",
"0.8403",
"0.505",
"0.8403"
],
[
"0.0801",
"0.834",
"0.468",
"0.834",
"0.468",
"0.8477",
"0.0801",
"0.8477"
],
[
"0.505",
"0.8413",
"0.8906",
"0.8413",
"0.8906",
"0.855",
"0.505",
"0.855"
],
[
"0.0784",
"0.8496",
"0.468",
"0.8486",
"0.468",
"0.8623",
"0.0784",
"0.863"
],
[
"0.505",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.505",
"0.8706"
],
[
"0.5024",
"0.872",
"0.8906",
"0.8706",
"0.8906",
"0.8853",
"0.5024",
"0.886"
],
[
"0.09314",
"0.887",
"0.46",
"0.887",
"0.46",
"0.901",
"0.09314",
"0.901"
],
[
"0.5034",
"0.887",
"0.892",
"0.887",
"0.892",
"0.901",
"0.5034",
"0.901"
],
[
"0.4805",
"0.9253",
"0.492",
"0.9253",
"0.492",
"0.935",
"0.4805",
"0.935"
]
] | [
"GINA-3D: Learning to Generate Implicit Neural Assets in the Wild",
"Xinchen Yan 2",
"Charles R. Oi 2Mahyar Najibi 2",
"Boyang Deng 1,2t",
"",
"Bokui Shen 1*",
"",
"Leonidas Guibas 3",
"Yin Zhou 2",
"DragomirAnguelov2",
"1 Stanford University, 2 Waymo LLC, 3 Google",
"",
"",
"H",
"",
"ghttme",
"\"Random..",
"In-the-wildDrivingData",
"GINA-3DSvnthesis",
"Composition with Background NeRF",
"Figure 1. Leveraging in-the-wild data for generativ",
"ve assets modeling embodies a scalable approach for simulation. GINA-3D uses real",
"world driving data to perform various synthesis tasks for realistic 3D implicit neural assets. Left: Multi-sensor observations in the wild.",
"Middle: Asset reconstruction and conditional synthesis. Right: Scene composition with background neural fields [1].",
"Abstract",
"ages and geometries.",
"Modeling the 3Dworld from sensor data for simu-",
"lation is a scalable way of developing testing and valida-",
"1.Introduction",
"tion environments for robotic learning problems such as",
"autonomous driving.However, manually creating or re-",
"几十年来，学习感知、推理和与3D世界互动一直是",
"creating real-world-like environments is difficult, expensive,",
"计算机视觉和机器人领域的长期挑战[2-9]。部署在野",
"and not scalable.Recent generative model techniques have",
"外的现代机器人系统E[10-16]通常配备有多个传感器",
"shown promising progress to address such challenges by",
"(e.g.摄像机，LiDAR和雷达），可以感知3D环境，然",
"learning 3Dassets using onlyplentiful 2D images-but still",
"后是用于推理和与复杂场景动态交互的智能单元。在",
"suffer limitations as they leverage either human-curated im-",
"现实环境中对这些智能代理进行端到端测试和验证既",
"age datasets or renderings from manually-created synthetic",
"困难又昂贵，特别是在安全关键和资源受限的领域，如",
"3D environments.In this paper, we introduce GINA-3D,",
"自动驾驶",
"a generative model that uses real-world driving data from",
"另一方面，在过去几年中，模拟数据的使用激增，以",
"camera and LiDAR sensors to create realistic 3D implicit",
"安全，可扩展和可验证的方式在受控设置E[17-27]下",
"neural assets of diverse vehicles and pedestrians.Com+",
"训练和评估智能代理。计算机图形学的飞速发展推动了",
"pared to the existing image datasets, the real-world driving",
"这种发展，包括谊染框架E[28-30]，物理仿真E[31,32]",
"setting poses new challenges due to occlusions, lighting-",
"和大规模开源资产存储库E[33-39]。一个关键问题是",
"variations and long-tail distributions.GINA-3D tackles",
"创建逼真的虚拟世界，使资产内容，组成和行为与真",
"these challenges by decoupling representation learning and",
"实分布保持一致，以便使从业者相信使用此类模拟进",
"generative modeling into two stages with a learned tri-plane",
"行开发和验证可以转移到现实世界中的性能E[40-481",
"latent structure, inspired by recent advances in generative",
"",
"然而，手动资产创建面临两个主要障碍。首先，手",
"modeling of images.To evaluate our approach, we con-",
"动创建3D资产需要具有3D领域专业知识的工程师和",
"struct a large-scale object-centric dataset containing over",
"艺术家的不懈努力，这是昂贵且难以扩展的E[26]。其",
"1.2M images of vehicles and pedestrians from the Waymo",
"次，现实世界的分布包含各种各样的例子（包括有趣的",
"Open Dataset, and a new set of 80K images of long-tail",
"罕见案例），并且也在不断发展[49,50]。生成3D建",
"instances such as construction equipment, garbage trucks,",
"模的最新发展为解决上述障碍提供了新的视角，因为",
"and cable cars.We compare our model with existing ap",
"它允许生成其他逼真但以前看不见的示例。这些方法",
"proaches and demonstrate that it achieves state-of-the-art",
"的一个子类，即生成的3D感知图像合成E[51,52]，具",
"performance in quality and diversity forboth generated im",
"有重要的前景，因为它可以通过部分观察（3D对象的",
"e.g.图像投影）进行3D建模此外，许多现实世界的机",
"* Work done during an internship at Waymo. t Work done at Waymo.",
"器人应用程序已经大规模捕获，注释和更新多传感器",
"1"
] | {
"arxiv_id": "2304.02163",
"language": "zh",
"page_idx": 0
} |
||
2304.02163-zh-1 | 2304.02163-zh-1.jpg | 观测。因此，这些数据提供了真实世界分布的准确，多样，与任务相关且最新的表示，生成模型可以潜在地捕获。但是，现有作品使用具有清晰观察值$\mathbb{E}$[53-58]的人工图像数据集或来自合成 3D 环境$\mathbb{E}[33,36]$ 的渲染。将生成的 3D 感知图像合成模型扩展到现实世界面临几个挑战，因为许多因素纠缠在部分观测中。首先，从没有 3D 结构的简单先验中桥接野外图像会使学习变得困难。其次，不受约束的遮挡将感兴趣的对象及其周围环境纠缠在像素空间中，这很难以纯粹的无监督方式解开。最后，由于缺乏为野外捕获的传感器数据构建以资产为中心的基准的努力，上述挑战变得更加复杂
在这项工作中，我们引人了一个用于隐式神经资产生成的 3D 感知生成转换器，名为 GINA-3D (GenergativeImplicitNeuralAsets)。为了解决现实世界的挑战，我们提出了一种新颖的 3D 感知编码器-解码器框架，该框架具有学习的结构化先验。具体来说，我们将三平面结构嵌人到生成模型的潜在先验(或三平面潜伏期)中， 其中每个条目都由来自学习码本$\mathbb{E}[59,60]$ 的离散表示进行参数化。编码器-解码器框架由具有神经渲染组件的转换编码器和解码器组成。为了处理不受约束的遮挡，我们使用现成的分割模型$\boxed{\mathbb{E}}[61]$ 中的伪标签，通过遮挡感知合成明确地将对象像素与其周围分离。最后， 从离散码本中学习的三平面潜伏期先验可用于训练条件潜伏期采样模型$\overline{\mathbb{E}}[62]$。相同的码本可以很容易地应用于各种条件合成任务，包括对象规模，类，语义和一天中的时间
为了评估我们的模型，我们从野外捕获的多传感器驾驶数据构建了一个以对象为中心的大型基准。我们首先提取超过 1。来自 WaymoOpenDataset[14]的车辆和行人的各种变化的 2M 图像。然后，我们使用来自现实世界驾驶场景的长尾实例来增强基准测试，包括建筑设备，缆车，校车和垃圾车等稀有物体。我们通过广泛的实验证明，GINA-3D 优于最先进的 3D 感知生成模型，通过图像质量，几何一致性和几何多样性来衡量。此外，我们通过利用学习的 3D 感知码本展示了各种条件合成任务和形状编辑结果的示例应用。该基准测试可通过 waymo.com/open 公开获得
# $2. \textbf{ Related Work}$
我们讨论了有关生成 3D 感知图像合成，3D 形状建
模以及在自动驾驶中的应用的相关工作
Generative 3D-aware image synthesis. 从图像集合中学习生成的 3D 感知表示在过去十年中越来越受欢迎[63-69]。早期的工作探索了从分散的因素中进行图像合成，例如学习的姿势嵌人$\mathbb{E}[64,66,69]$ 或紧凑的场景表示$\mathbb{E}[65,67]$。将 3D 结构表示为压缩嵌入，此工作线通过从具有 2D 反卷积层堆栈的嵌入空间进行上采样来进行图像合成。在可区分渲染的进步的推动下，已经努力$\operatorname{E[\text{70-73]将显式 3D 结构烘焙到生成体}}$ 系结构中。然而，由于内存消耗，这些努力通常仅限于粗略的 3D 离散化。超越明确，最近的工作利用神经辐射场学习隐式 3D 感知结构[51,52,74-82] 进行图像合成。Schwarz$\underline{\mathrm{E}}etal.\underline{\mathrm{E}}[74]$引人了生成辐射场 (GRAF),
该辐射场可在没有遮挡的情况下解开单个对象的 3D 形状，外观和相机姿态。Niemeyer et al.[51]建立在GRAF之上，提出了 GIRAFFE 模型，该模型通过使用组合 3D 场景结构来处理涉及多个对象的场景。值得注意的是，体积渲染中的查询操作在更高的分辨率下变得计算量大。为了解决这个问题，Chan et al.[52]引入了具有三平面特征的混合显式-隐式 3D 表示 (EG3D) , 展示了更高分辨率的图像合成。同时，[83]和[84]率先在 ImageNet 上使用三平面表示进行高分辨率无界 3D 场景生成，其中$\mathbb{E}$[84] 使用矢量量化框架$,\mathbb{E}$[83] 使用GAN 框架。我们的工作是为自动驾驶传感器仿真中的应用而设计的，重点是以对象为中心的建模
Generative 3D shape modeling. 完整 3D 形状的生成建模也已得到广泛研究，包括合成 3D 体素网格E [85-93] ,点云$\mathbb{E}[$94-96] ,表面网格$\mathbb{E}[$97-103] ,形状基元$\mathbb{E}[104,105]$以及使用[103,106-112]的隐式函数或混合生成表示的努力\_。Shen et al.[11]介绍了一种称为深行进四面体(DMTet)的可微分显式表面提取方法，该方法学习直接重建具有任意拓扑的 3D 表面网格。Gao$\boxed{\mathrm{E}}etal.\boxed{\mathrm{E}}[103]$ 建立在 EG3D$\boxed{\mathrm{E}}[52]$三平面特征的基础上进行图像合成，提出了一种扩展，该扩展能够使用 DMTet 生成纹理表面网格，以生成几何形状，并使用三平面特征进行纹理合成。现有的努力假设可以访问准确的多视图轮廓 (通常来自完整的地面实况 3D 形状),这并不能反映野外捕获的数据中存在的真正挑战
Assets modeling in driving simulation. 模拟环境建模在自动驾驶领域引起了极大的关注。简而言之，问题可以分解为资产创建(e。g。,动态对象和背景)、场景生成和渲染。早期工作利用艺术家创建的对象和背景资产来构建虚拟驾驶环境$\mathbb{E}[18,20,113]$ 使用经典图形渲染管道。虽然能够生成具有不同配置的虚拟场景，但这些方法产生的场景具有有限的多样性和明显的现实差距。最近的许多作品探索了数据驱动仿真的不同方面，包括图像合成$\mathbb{E}[114-117]$,资产建模$\mathbb{E}[47,48,118-121]$ 场景生成$\mathbb{E}[49,122,123]$ 和场景渲染$\mathbb{E}[1,124-126]$ 。特别是，Chen et al.[48] 和 Zakharov et al.[119] 从每个车辆对象的单视图观察中执行了显式纹理扭曲或隐式渲染。因此，它们的资产重建质量对遮挡很敏感，并且受单次观测的视角限制。在这些努力的基础上，包括 Muller$\boxed{\mathrm{E}}etal.[121]$和 Kundu$\boxed{\mathrm{F}}etal.[125]$在内的最新工作通过全局或特定于实例的潜在代码实现了对象完成，代表了归一化对象坐标空间(NOCS)下的每个对象资产。相比之下，我们提出的模型中的潜在代码具有 3D 三平面结构，这在学习和应用方面提供了几个好处。更重要的是，我们可以生成以前看不见的 3D 资产， 这与对象重建有本质的不同
# 3. Generative Implicit Neural Assets
我们提出了 GINA-3D, 这是一个可扩展的框架，用于从野外数据中获取 3D 资产 (Sec。3.1)。我们框架的核心是一种新颖的 3D 感知编码器-解码器模型，该
2 | [
[
"0.0809",
"0.09216",
"0.468",
"0.09216",
"0.468",
"0.1061",
"0.0809",
"0.1061"
],
[
"0.505",
"0.0947",
"0.8896",
"0.0947",
"0.8896",
"0.1048",
"0.505",
"0.1048"
],
[
"0.0825",
"0.108",
"0.4673",
"0.108",
"0.4673",
"0.1218",
"0.0825",
"0.1218"
],
[
"0.5034",
"0.108",
"0.8906",
"0.108",
"0.8906",
"0.1218",
"0.5034",
"0.1218"
],
[
"0.0825",
"0.125",
"0.4673",
"0.125",
"0.4673",
"0.1351",
"0.0825",
"0.1351"
],
[
"0.5034",
"0.1225",
"0.8896",
"0.1225",
"0.8896",
"0.1364",
"0.5034",
"0.1364"
],
[
"0.08417",
"0.137",
"0.4673",
"0.137",
"0.4673",
"0.1509",
"0.08417",
"0.1509"
],
[
"0.5034",
"0.1383",
"0.8906",
"0.1383",
"0.8906",
"0.1515",
"0.5034",
"0.1515"
],
[
"0.08417",
"0.1547",
"0.4673",
"0.1547",
"0.4673",
"0.1654",
"0.08417",
"0.1654"
],
[
"0.505",
"0.154",
"0.8906",
"0.154",
"0.8906",
"0.1674",
"0.505",
"0.1674"
],
[
"0.0825",
"0.1686",
"0.468",
"0.1686",
"0.468",
"0.1818",
"0.0825",
"0.1818"
],
[
"0.505",
"0.1686",
"0.8906",
"0.1686",
"0.8906",
"0.1818",
"0.505",
"0.1818"
],
[
"0.0825",
"0.1831",
"0.4673",
"0.1831",
"0.4673",
"0.1963",
"0.0825",
"0.1963"
],
[
"0.5073",
"0.1831",
"0.8906",
"0.1831",
"0.8906",
"0.1963",
"0.5073",
"0.1963"
],
[
"0.0825",
"0.1995",
"0.4673",
"0.1995",
"0.4673",
"0.2102",
"0.0825",
"0.2102"
],
[
"0.505",
"0.1995",
"0.8896",
"0.1995",
"0.8896",
"0.2102",
"0.505",
"0.2102"
],
[
"0.0825",
"0.2153",
"0.4656",
"0.2153",
"0.4656",
"0.226",
"0.0825",
"0.226"
],
[
"0.5034",
"0.2134",
"0.8906",
"0.2134",
"0.8906",
"0.2267",
"0.5034",
"0.2267"
],
[
"0.0825",
"0.2291",
"0.4673",
"0.2291",
"0.4673",
"0.2424",
"0.0825",
"0.2424"
],
[
"0.505",
"0.2291",
"0.888",
"0.2291",
"0.888",
"0.2424",
"0.505",
"0.2424"
],
[
"0.0809",
"0.2437",
"0.423",
"0.2437",
"0.423",
"0.2568",
"0.0809",
"0.2568"
],
[
"0.5034",
"0.2437",
"0.8896",
"0.2437",
"0.8896",
"0.2568",
"0.5034",
"0.2568"
],
[
"0.0989",
"0.2588",
"0.4673",
"0.2583",
"0.4673",
"0.2727",
"0.0989",
"0.2734"
],
[
"0.505",
"0.2588",
"0.8145",
"0.2588",
"0.8145",
"0.2727",
"0.505",
"0.2727"
],
[
"0.0809",
"0.2747",
"0.468",
"0.2747",
"0.468",
"0.2886",
"0.0809",
"0.2886"
],
[
"0.5107",
"0.2842",
"0.8896",
"0.2842",
"0.8896",
"0.2974",
"0.5107",
"0.2974"
],
[
"0.0809",
"0.2917",
"0.4673",
"0.2917",
"0.4673",
"0.3018",
"0.0809",
"0.3018"
],
[
"0.0825",
"0.305",
"0.4673",
"0.305",
"0.4673",
"0.318",
"0.0825",
"0.318"
],
[
"0.505",
"0.2998",
"0.8896",
"0.2998",
"0.8896",
"0.3132",
"0.505",
"0.3132"
],
[
"0.505",
"0.3125",
"0.892",
"0.3125",
"0.892",
"0.327",
"0.505",
"0.327"
],
[
"0.08417",
"0.3213",
"0.4673",
"0.3213",
"0.4673",
"0.332",
"0.08417",
"0.332"
],
[
"0.505",
"0.3289",
"0.8896",
"0.3289",
"0.8896",
"0.3423",
"0.505",
"0.3423"
],
[
"0.0825",
"0.3352",
"0.468",
"0.3352",
"0.468",
"0.3484",
"0.0825",
"0.3484"
],
[
"0.5034",
"0.3435",
"0.888",
"0.3435",
"0.888",
"0.3567",
"0.5034",
"0.3567"
],
[
"0.0825",
"0.351",
"0.468",
"0.351",
"0.468",
"0.3643",
"0.0825",
"0.3643"
],
[
"0.5034",
"0.3591",
"0.8906",
"0.3591",
"0.8906",
"0.3726",
"0.5034",
"0.3726"
],
[
"0.08417",
"0.3662",
"0.4673",
"0.3662",
"0.4673",
"0.377",
"0.08417",
"0.377"
],
[
"0.0801",
"0.379",
"0.468",
"0.3806",
"0.468",
"0.3945",
"0.0801",
"0.3926"
],
[
"0.505",
"0.375",
"0.888",
"0.375",
"0.888",
"0.3882",
"0.505",
"0.3882"
],
[
"0.5034",
"0.3894",
"0.8906",
"0.3894",
"0.8906",
"0.4028",
"0.5034",
"0.4028"
],
[
"0.0809",
"0.3958",
"0.468",
"0.3958",
"0.468",
"0.4092",
"0.0809",
"0.4092"
],
[
"0.0825",
"0.4104",
"0.468",
"0.4104",
"0.468",
"0.4236",
"0.0825",
"0.4236"
],
[
"0.5034",
"0.4048",
"0.8906",
"0.4048",
"0.8906",
"0.4185",
"0.5034",
"0.4185"
],
[
"0.505",
"0.4192",
"0.8906",
"0.4192",
"0.8906",
"0.433",
"0.505",
"0.433"
],
[
"0.0825",
"0.4255",
"0.468",
"0.4255",
"0.468",
"0.4395",
"0.0825",
"0.4395"
],
[
"0.5034",
"0.435",
"0.8906",
"0.435",
"0.8906",
"0.449",
"0.5034",
"0.449"
],
[
"0.0825",
"0.4426",
"0.4673",
"0.4426",
"0.4673",
"0.4526",
"0.0825",
"0.4526"
],
[
"0.0809",
"0.4558",
"0.4697",
"0.4558",
"0.4697",
"0.4697",
"0.0809",
"0.4697"
],
[
"0.505",
"0.452",
"0.888",
"0.452",
"0.888",
"0.4622",
"0.505",
"0.4622"
],
[
"0.505",
"0.4653",
"0.8906",
"0.4653",
"0.8906",
"0.4792",
"0.505",
"0.4792"
],
[
"0.0825",
"0.473",
"0.1626",
"0.473",
"0.1626",
"0.483",
"0.0825",
"0.483"
],
[
"0.5034",
"0.4797",
"0.523",
"0.4797",
"0.523",
"0.4937",
"0.5034",
"0.4937"
],
[
"0.10297",
"0.486",
"0.468",
"0.486",
"0.468",
"0.5",
"0.10297",
"0.5"
],
[
"0.0809",
"0.502",
"0.468",
"0.502",
"0.468",
"0.515",
"0.0809",
"0.515"
],
[
"0.5063",
"0.505",
"0.7974",
"0.505",
"0.7974",
"0.518",
"0.5063",
"0.518"
],
[
"0.79",
"0.507",
"0.8896",
"0.507",
"0.8896",
"0.5176",
"0.79",
"0.5176"
],
[
"0.08417",
"0.5176",
"0.4673",
"0.5176",
"0.4673",
"0.531",
"0.08417",
"0.531"
],
[
"0.505",
"0.521",
"0.8906",
"0.521",
"0.8906",
"0.534",
"0.505",
"0.534"
],
[
"0.0825",
"0.532",
"0.4673",
"0.532",
"0.4673",
"0.5454",
"0.0825",
"0.5454"
],
[
"0.505",
"0.535",
"0.8906",
"0.535",
"0.8906",
"0.55",
"0.505",
"0.55"
],
[
"0.0825",
"0.5474",
"0.468",
"0.5474",
"0.468",
"0.561",
"0.0825",
"0.561"
],
[
"0.505",
"0.5503",
"0.888",
"0.5503",
"0.888",
"0.5645",
"0.505",
"0.5645"
],
[
"0.0809",
"0.5625",
"0.4617",
"0.5625",
"0.4617",
"0.5757",
"0.0809",
"0.5757"
],
[
"0.5063",
"0.5674",
"0.888",
"0.5674",
"0.888",
"0.5776",
"0.5063",
"0.5776"
],
[
"0.0809",
"0.577",
"0.468",
"0.577",
"0.468",
"0.5903",
"0.0809",
"0.5903"
],
[
"0.505",
"0.5806",
"0.8906",
"0.5806",
"0.8906",
"0.5947",
"0.505",
"0.5947"
],
[
"0.0825",
"0.5913",
"0.4673",
"0.5913",
"0.4673",
"0.605",
"0.0825",
"0.605"
],
[
"0.505",
"0.595",
"0.8896",
"0.595",
"0.8896",
"0.6094",
"0.505",
"0.6094"
],
[
"0.0825",
"0.6094",
"0.4673",
"0.6094",
"0.4673",
"0.619",
"0.0825",
"0.619"
],
[
"0.5034",
"0.61",
"0.8906",
"0.61",
"0.8906",
"0.6235",
"0.5034",
"0.6235"
],
[
"0.0809",
"0.6226",
"0.4697",
"0.6226",
"0.4697",
"0.636",
"0.0809",
"0.636"
],
[
"0.5034",
"0.6255",
"0.8906",
"0.6255",
"0.8906",
"0.6396",
"0.5034",
"0.6396"
],
[
"0.0809",
"0.638",
"0.3245",
"0.638",
"0.3245",
"0.6514",
"0.0809",
"0.6514"
],
[
"0.5195",
"0.64",
"0.892",
"0.6416",
"0.892",
"0.655",
"0.5195",
"0.654"
],
[
"0.0801",
"0.661",
"0.2239",
"0.661",
"0.2239",
"0.675",
"0.0801",
"0.675"
],
[
"0.505",
"0.656",
"0.8906",
"0.656",
"0.8906",
"0.6694",
"0.505",
"0.6694"
],
[
"0.505",
"0.6704",
"0.8906",
"0.6704",
"0.8906",
"0.6846",
"0.505",
"0.6846"
],
[
"0.10046",
"0.6846",
"0.468",
"0.6846",
"0.468",
"0.6978",
"0.10046",
"0.6978"
],
[
"0.5034",
"0.686",
"0.8896",
"0.686",
"0.8896",
"0.6997",
"0.5034",
"0.6997"
],
[
"0.0809",
"0.6987",
"0.3555",
"0.6987",
"0.3555",
"0.712",
"0.0809",
"0.712"
],
[
"0.505",
"0.702",
"0.8906",
"0.702",
"0.8906",
"0.7153",
"0.505",
"0.7153"
],
[
"0.0866",
"0.72",
"0.3694",
"0.72",
"0.3694",
"0.735",
"0.0866",
"0.735"
],
[
"0.505",
"0.7163",
"0.8906",
"0.7163",
"0.8906",
"0.73",
"0.505",
"0.73"
],
[
"0.3635",
"0.7236",
"0.3816",
"0.7236",
"0.3816",
"0.731",
"0.3635",
"0.731"
],
[
"0.3774",
"0.7227",
"0.4673",
"0.7227",
"0.4673",
"0.733",
"0.3774",
"0.733"
],
[
"0.0825",
"0.7363",
"0.468",
"0.7363",
"0.468",
"0.75",
"0.0825",
"0.75"
],
[
"0.5034",
"0.732",
"0.8906",
"0.732",
"0.8906",
"0.7456",
"0.5034",
"0.7456"
],
[
"0.0809",
"0.752",
"0.468",
"0.752",
"0.468",
"0.765",
"0.0809",
"0.765"
],
[
"0.5034",
"0.747",
"0.8896",
"0.747",
"0.8896",
"0.7603",
"0.5034",
"0.7603"
],
[
"0.08417",
"0.7666",
"0.468",
"0.7666",
"0.468",
"0.78",
"0.08417",
"0.78"
],
[
"0.505",
"0.762",
"0.8896",
"0.762",
"0.8896",
"0.776",
"0.505",
"0.776"
],
[
"0.0825",
"0.7827",
"0.4673",
"0.7827",
"0.4673",
"0.7935",
"0.0825",
"0.7935"
],
[
"0.505",
"0.779",
"0.888",
"0.779",
"0.888",
"0.789",
"0.505",
"0.789"
],
[
"0.0825",
"0.797",
"0.4673",
"0.797",
"0.4673",
"0.81",
"0.0825",
"0.81"
],
[
"0.5034",
"0.7915",
"0.8936",
"0.7915",
"0.8936",
"0.8047",
"0.5034",
"0.8047"
],
[
"0.08417",
"0.811",
"0.4673",
"0.811",
"0.4673",
"0.8247",
"0.08417",
"0.8247"
],
[
"0.5034",
"0.8066",
"0.701",
"0.8066",
"0.701",
"0.821",
"0.5034",
"0.821"
],
[
"0.0825",
"0.827",
"0.4673",
"0.827",
"0.4673",
"0.8403",
"0.0825",
"0.8403"
],
[
"0.5034",
"0.8335",
"0.812",
"0.8335",
"0.812",
"0.8467",
"0.5034",
"0.8467"
],
[
"0.08417",
"0.8413",
"0.4673",
"0.8413",
"0.4673",
"0.855",
"0.08417",
"0.855"
],
[
"0.0809",
"0.857",
"0.468",
"0.857",
"0.468",
"0.8706",
"0.0809",
"0.8706"
],
[
"0.524",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.524",
"0.8706"
],
[
"0.0801",
"0.8706",
"0.468",
"0.872",
"0.468",
"0.886",
"0.0801",
"0.8853"
],
[
"0.5034",
"0.872",
"0.8906",
"0.872",
"0.8906",
"0.8853",
"0.5034",
"0.8853"
],
[
"0.0825",
"0.887",
"0.4722",
"0.887",
"0.4722",
"0.901",
"0.0825",
"0.901"
],
[
"0.5034",
"0.887",
"0.892",
"0.887",
"0.892",
"0.901",
"0.5034",
"0.901"
],
[
"0.4756",
"0.925",
"0.4927",
"0.922",
"0.4968",
"0.9375",
"0.4805",
"0.94"
]
] | [
"观测。因此，这些数据提供了真实世界分布的准确，多",
"该辐射场可在没有遮挡的情况下解开单个对象的3D",
"样，与任务相关且最新的表示，生成模型可以潜在地捕",
"形状，外观和相机姿态。Niemeyeretal.[51]建立在",
"获。但是，现有作品使用具有清晰观察值53-581的",
"GRAF之上，提出了GIRAFFE模型，该模型通过使用",
"人工图像数据集或来自合成3D环境E[33,36]的染。",
"组合3D场景结构来处理涉及多个对象的场景。值得注",
"将生成的3D感知图像合成模型扩展到现实世界面临几",
"意的是，体积染中的查询操作在更高的分辨率下变",
"个挑战，因为许多因素纠缠在部分观测中。首先，从没",
"得计算量大。为了解决这个问题，Chanetal.[52]引人",
"有3D结构的简单先验中桥接野外图像会使学习变得困",
"了具有三平面特征的混合显式-隐式3D表示（EG3D），",
"难。其次，不受约束的遮挡将感兴趣的对象及其周围环",
"展示了更高分辨率的图像合成。同时，「831和[841率先",
"境纠缠在像素空间中，这很难以纯粹的无监督方式解",
"在ImageNet上使用三平面表示进行高分辨率无界3D",
"开。最后，由于缺乏为野外捕获的传感器数据构建以资",
"场景生成，其中E[84]使用失量量化框架，E[83]使用",
"产为中心的基准的努力，上述挑战变得更加复杂",
"GAN框架。我们的工作是为自动驾驶传感器仿真中的",
"在这项工作中，我们引人了一个用于隐式神经资产",
"应用而设计的，重点是以对象为中心的建模",
"生成的3D感知生成转换器，名为GINA-3D（Generga",
"Generative3Dshapemodeling：完整3D形状的生",
"tivelmplicitNeuralAsets）。为了解决现实世界的挑战，我",
"们提出了一种新颖的3D感知编码器-解码器框架，该框",
"成建模也已得到广泛研究，包括合成3D体素网格E",
"[85-93]，点云E[94-96]，表面网格E[97-103]，形状",
"架具有学习的结构化先验。具体来说，我们将三平面结",
"基元E[104,105]以及使用[103,106-112]的隐式函数或",
"构嵌人到生成模型的潜在先验（或三平面潜伏期）中，",
"混合生成表示的努力_。Shen etal.［lli]介绍了一种",
"其中每个条目都由来自学习码本E[59,60]的离散表示",
"称为深行进四面体（DMTet）的可微分显式表面提取",
"进行参数化。编码器-解码器框架由具有神经染组件",
"的转换编码器和解码器组成。为了处理不受约束的遮",
"方法，该方法学习直接重建具有任意拓扑的3D表面网",
"格。GaoEletal.E[103]建立在EG3DE[52]三平面特征",
"挡，我们使用现成的分割模型E[61]中的伪标签，通过",
"遮挡感知合成明确地将对象像素与其周围分离。最后，",
"的基础上进行图像合成，提出了一种扩展，该扩展能够",
"使用DMTet生成纹理表面网格，以生成几何形状，并",
"从离散码本中学习的三平面潜伏期先验可用于训练条",
"使用三平面特征进行纹理合成。现有的努力假设可以",
"件潜伏期采样模型F621。相同的码本可以很容易地应",
"用于各种条件合成任务，包括对象规模，类，语义和一",
"访问准确的多视图轮廓（通常来自完整的地面实况3D",
"形状），这并不能反映野外捕获的数据中存在的真正挑",
"天中的时间",
"战",
"为了评估我们的模型，我们从野外捕获的多传感器",
"驾驶数据构建了一个以对象为中心的大型基准。我们",
"Assets modeling in driving simulation.",
"模拟环境建模",
"首先提取超过1。来自WaymoOpenDataset[14】的车辆",
"在自动驾驶领域引起了极大的关注。简而言之，问题可",
"和行人的各种变化的2M图像。然后，我们使用来自现",
"以分解为资产创建(e。g。、动态对象和背景)、场景生成",
"实世界驾驶场景的长尾实例来增强基准测试，包括建",
"和谊染。早期工作利用艺术家创建的对象和背景资产",
"筑设备，缆车，校车和垃圾车等稀有物体。我们通过广",
"来构建虚拟驾驶环境E18.20.1131使用经典图形染",
"泛的实验证明，GINA-3D优于最先进的3D感知生成模",
"管道。虽然能够生成具有不同配置的虚拟场景，但这些",
"型，通过图像质量，几何一致性和几何多样性来衡量。",
"方法产生的场景具有有限的多样性和明显的现实差距。",
"此外，我们通过利用学习的3D感知码本展示了各种条",
"最近的许多作品探索了数据驱动仿真的不同方面，包",
"件合成任务和形状编辑结果的示例应用。该基准测试",
"括图像合成E[114-117]，资产建模E[47,48,118-121]",
"可通过waymo.com/open公开获得",
"场景生成E[49,122,123]和场景谊染E[1,124-126]。",
"2.Related Work",
"特别是，Chen et al.[48]和Zakharovetal.[119]从每",
"个车辆对象的单视图观察中执行了显式纹理扭曲或隐",
"我们讨论了有关生成3D感知图像合成，3D形状建",
"式染。因此，它们的资产重建质量对遮挡很敏感，并",
"模以及在自动驾驶中的应用的相关工作",
"且受单次观测的视角限制。在这些努力的基础上，包",
"Generative 3D-aware image synthesis.",
"括MullerEetal.[121]]和KunduEetal.[125]在内的最",
"",
"从图像集合",
"中学习生成的3D感知表示在过去十年中越来越受欢",
"新工作通过全局或特定于实例的潜在代码实现了对象",
"迎[63-69]。早期的工作探索了从分散的因素中进行",
"完成，代表了归一化对象坐标空间（NOCS）下的每个",
"图像合成，例如学习的姿势嵌入E[64,66,69]或紧凑的",
"对象资产。相比之下，我们提出的模型中的潜在代码具",
"场景表示E[65.671。将3D结构表示为压缩嵌入：此",
"有3D三平面结构，这在学习和应用方面提供了儿个好",
"工作线通过从具有2D反卷积层堆栈的嵌入空间进行",
"处。更重要的是，我们可以生成以前看不见的3D资产，",
"上采样来进行图像合成。在可区分染的进步的推动",
"这与对象重建有本质的不同",
"下，已经努力E[70-73]将显式3D结构烘焙到生成体",
"3. Generative Implicit Neural Assets",
"系结构中。然而，由于内存消耗，这些努力通常仅限于",
"粗略的3D离散化。超越明确，最近的工作利用神经辐",
"我们提出了GINA-3D，这是一个可扩展的框架，用",
"射场学习隐式3D感知结构[51,52,74-82]进行图像合",
"于从野外数据中获取3D资产（Sec。3.1）。我们框架",
"成。SchwarzEetal.E[74]引人了生成辐射场（GRAF)，",
"的核心是一种新颖的3D感知编码器-解码器模型，该",
"2"
] | {
"arxiv_id": "2304.02163",
"language": "zh",
"page_idx": 1
} |
||
2304.02163-zh-4 | 2304.02163-zh-4.jpg | <FigureHere>
Figure 4. Image samples from our object-centric benchmark
Semantic feature fields. 最近的各种工作都证明了学习混合表示在神经渲染$\operatorname{E[\text{142-144]和 2D 图像合}}$ 成$\overline{\mathbb{E}}[$145] 中的有效性。通过在我们的神经渲染 MLP 中计算额外的通道，我们可以自然地将语义特征字段合并到我们的公式中。我们为每个图像预先计算 DINO-ViT 特征$\overline{\mathbb{E}}[$146],并学习语义特征字段以在生成的实例之间建立零件对应关系
LiDAR depth supervision. 当 LiDAR 点云在数据中可用时，它可以通过渲染深度和 LiDAR 深度之间的重建项用作附加监督
Conditional synthesis. 最后但并非最不重要的一点是，附加信息支持条件合成中的各种应用。表示为$\mathcal{C}$, 它可以作为$M_\psi(\boldsymbol{s}_{ijk}|\boldsymbol{s}_{\bar{M}},\mathcal{C})$输入到我们的潜在先验中。例如，对象比例，对象类，一天中的时间和对象语义嵌入也可以用作$c$,以控制生成过程
# 4. Experiments
4.1. Object-centric Benchmark
我们选择 WaymoOpenDataset(WOD)[14],因为它是最大和最多样化的自动驾驶数据集之一，包含丰富的几何和语义标签，如对象边界框和每像素实例掩码。具体来说，该数据集包括 1,150 个主要在旧金山市中心和凤凰城捕获的驾驶场景，每个场景由 200 帧多传感器观测组成。为了构建以对象为中心的基准，我们提出了一个从粗到细的过程，通过利用 3D 对象框，相机-激光雷达同步和细粒度的 2D 全景标签来提取单视图 2D 照片的集合。首先，我们利用 3D 框注释来排除每个数据帧中与测量车辆相距一定距离以外的物体 (e。g。, 分别用于行人的 40m 和用于车辆的 60m )。在给定的帧中，我们将每个 3D 边界框内的 3D 点云投影到最可见的相机，并提取居中的补丁以构建我们的单视图 2D 图像集合此外，我们使用标记子集上的 2D 全景分割来训练 Panoptic-Deeplab 模型$\mathbb{E}[61,147]$ ,并为整个数据集上的每个相机图像创建逐像素伪标签。这使我们能够区分属于感兴趣对象、背景和遮挡物的像素(e。g。, 站在一个人面前的柱子)。我们使用 2D 全景预测进一步排除了对象被严重遮挡的某些斑块。即使应用了过滤标准，我们认为，由于遮挡、类内变化(e。g。,卡车和轿车),部分观察(e。g。,我们对单个车辆没有完整的 360 度观测),而且分割不完善。特别是，考虑到相机卷帘快门，物体运动和自我运动，我们提供了相机光线和 LiDAR 点云与物体坐标系的精确配准。我们重复相同的过程，从 WOD 中提取车辆和行人，并从我们的长尾数据集中提取其他长尾车辆。拟议的以对象为中心的基准测试是迄今为止生成建模的最大数据集之一， 向括野外的各种长尾示例
<table>
<tbody>
<tr>
<th> </th>
<th>Images</th>
<th>${\mathrm{Unique}}$ Instances</th>
</tr>
<tr>
<td>WOD- Vehicle</td>
<td>$901K$</td>
<td>23.6K</td>
</tr>
<tr>
<td>WOD-Pedestrian</td>
<td>$321K$</td>
<td>$8.1K$</td>
</tr>
<tr>
<td>Longtail- Vehicle</td>
<td>$80K$</td>
<td>$3.7\mathbf{K}$</td>
</tr>
</tbody>
</table>
Table 1. Statistics of our object-centric benchmark. Experiments
were conducted on a subset with image patches rescaled to 256$^2$
resolution.
# 4.2. Implementation Details
GINA-3D. 我们的编码器以 256$^{2}$的分辨率拍摄图像并在训练期间以 128$^{2}$渲染。我们的三平面延迟具有 16$^{2}$ 的分辨率，其码本包含 2048 个条目和 32 的查找维度。我们使用 Adam 优化器$\overline{\mathbb{E}}[$148]在 8个 TeslaV100GPU 上训练了模型，每个阶段的批量大小分别为 32 和 64。我们训练了 150K 步的第一阶段和 80K 步的第二阶段Baselines.
我们与域中两种最先进的方法 GI-
RAFFE[51] 和 EG3D$\overline{\mathbb{E}}$[52] 进行了比较，我们以 128$^{2}$ 的分辨率在数据集上对其进行了训练。我们注意到，在全像素上训练的 GIRAFFE 模型无法解开视点，遮挡和身份。这使得前景像素的提取变得困难，因为渲染蒙版仅在低维分辨率 16$^{2}$下定义。我们改为使用通过白化非对象区域训练的模型报告数字。对于 EG3D, 我们观察到，由于缺乏前景和背景建模，使用未屏蔽图像训练EG3D 会导致训练崩溃。因此，我们在相同的设置下训练 EG3D。
# 4.3. Evaluations on WOD-Vehicle
我们在表中进行定量评估。2 并在图中可视化不同
模型的定性结果。5
Image Evaluation. 对于图像质量，我们计算 50K 生成的图像与所有可用验证图像之间的 Fréchet 起始距离(FID)E[149]。为了更好地反映对象完整性的指标，我们对其对象分割掩码至少占用投影 3D 边界框的 50% 的图像进行滤波(图。5-右)。此外，我们通过蒙版飘浮在联盟(蒙版 FOU)来测量生成的图像的完整性，这被定义为渲染对象区域中未连接像素的百分比。为了测量语义多样性，我们使用 CLIP$\boxed{\mathbb{E}}[150]$嵌人计算覆盖率 (COV) 得分和最小匹配距离 (MMD)$\mathbb{E}[94]$。COV 测量在生成的集合中具有匹配的验证集合中的 CLIP 嵌人的比例，而 MMD 测量每个生成的嵌人与验证中最接近的嵌入之间的距离。我们的模型展示了FID、图像完整性和语义多样性方面的显著改进。如果没有明确的解纠缠，基线很难处理真实的分布，导致形状不完整的伪影(图。5)
几何图形评估。为了测量基础体渲染的一致性，我们遵循 Or et al.[79],并从两个视点计算体渲染深度之间的对齐误差。在 EG3D [52]之后，我们使用密度阈值为 10 的行军立方体 [151]提取网格。我们通过网格漂浮物 (网格 FOU) 来衡量完整性，它被定义为整个网格中未连接的网格块上的表面积的百分比。由于我们在现实世界数据中没有地面实况网格，因此我们通过在验证集的边界框内测量生成的网格和聚合的 LiDAR 点云之间的网格多样性来近似网格多样性。我们使用
5 | [
[
"0.0866",
"0.09216",
"0.4673",
"0.09216",
"0.4673",
"0.1061",
"0.0866",
"0.1061"
],
[
"0.0809",
"0.108",
"0.468",
"0.108",
"0.468",
"0.1218",
"0.0809",
"0.1218"
],
[
"0.0825",
"0.1225",
"0.468",
"0.1225",
"0.468",
"0.1364",
"0.0825",
"0.1364"
],
[
"0.0825",
"0.1383",
"0.4673",
"0.1383",
"0.4673",
"0.1515",
"0.0825",
"0.1515"
],
[
"0.8384",
"0.1465",
"0.8594",
"0.1465",
"0.8594",
"0.154",
"0.8384",
"0.154"
],
[
"0.0801",
"0.1528",
"0.468",
"0.1515",
"0.468",
"0.1654",
"0.0801",
"0.1666"
],
[
"0.8784",
"0.1483",
"0.8823",
"0.1483",
"0.8823",
"0.1547",
"0.8784",
"0.1547"
],
[
"0.0809",
"0.1674",
"0.4673",
"0.1674",
"0.4673",
"0.1812",
"0.0809",
"0.1812"
],
[
"0.5186",
"0.1724",
"0.601",
"0.1724",
"0.601",
"0.1831",
"0.5186",
"0.1831"
],
[
"0.6216",
"0.1724",
"0.7183",
"0.1724",
"0.7183",
"0.1831",
"0.6216",
"0.1831"
],
[
"0.761",
"0.1736",
"0.8545",
"0.1736",
"0.8545",
"0.1843",
"0.761",
"0.1843"
],
[
"0.0809",
"0.1818",
"0.2263",
"0.1818",
"0.2263",
"0.1957",
"0.0809",
"0.1957"
],
[
"0.517",
"0.1881",
"0.875",
"0.1881",
"0.875",
"0.202",
"0.517",
"0.202"
],
[
"0.08417",
"0.202",
"0.2793",
"0.2026",
"0.2793",
"0.2172",
"0.08417",
"0.2166"
],
[
"0.281",
"0.2084",
"0.2932",
"0.2084",
"0.2932",
"0.2122",
"0.281",
"0.2122"
],
[
"0.2893",
"0.2039",
"0.4673",
"0.2039",
"0.4673",
"0.2172",
"0.2893",
"0.2172"
],
[
"0.0809",
"0.2184",
"0.468",
"0.2184",
"0.468",
"0.2317",
"0.0809",
"0.2317"
],
[
"0.5034",
"0.2172",
"0.722",
"0.2172",
"0.722",
"0.2311",
"0.5034",
"0.2311"
],
[
"0.0809",
"0.2343",
"0.2115",
"0.2343",
"0.2115",
"0.2474",
"0.0809",
"0.2474"
],
[
"0.509",
"0.2361",
"0.589",
"0.2361",
"0.589",
"0.2474",
"0.509",
"0.2474"
],
[
"0.5947",
"0.2361",
"0.886",
"0.2361",
"0.886",
"0.25",
"0.5947",
"0.25"
],
[
"0.08496",
"0.2524",
"0.468",
"0.254",
"0.468",
"0.2683",
"0.08496",
"0.2676"
],
[
"0.5034",
"0.2507",
"0.8896",
"0.2507",
"0.8896",
"0.2651",
"0.5034",
"0.2651"
],
[
"0.0825",
"0.2695",
"0.4714",
"0.2695",
"0.4714",
"0.2827",
"0.0825",
"0.2827"
],
[
"0.5034",
"0.2664",
"0.8906",
"0.2664",
"0.8906",
"0.2798",
"0.5034",
"0.2798"
],
[
"0.0825",
"0.2854",
"0.4697",
"0.2854",
"0.4697",
"0.2986",
"0.0825",
"0.2986"
],
[
"0.5034",
"0.281",
"0.8906",
"0.281",
"0.8906",
"0.2942",
"0.5034",
"0.2942"
],
[
"0.0825",
"0.2998",
"0.468",
"0.2998",
"0.468",
"0.3132",
"0.0825",
"0.3132"
],
[
"0.5073",
"0.2966",
"0.892",
"0.2966",
"0.892",
"0.31",
"0.5073",
"0.31"
],
[
"0.0809",
"0.315",
"0.3245",
"0.315",
"0.3245",
"0.3289",
"0.0809",
"0.3289"
],
[
"0.5034",
"0.31",
"0.8784",
"0.3113",
"0.8784",
"0.3257",
"0.5034",
"0.3245"
],
[
"0.509",
"0.3289",
"0.601",
"0.3308",
"0.601",
"0.3447",
"0.509",
"0.3423"
],
[
"0.5947",
"0.33",
"0.8906",
"0.33",
"0.8906",
"0.3435",
"0.5947",
"0.3435"
],
[
"0.0801",
"0.3372",
"0.2133",
"0.3384",
"0.2133",
"0.3528",
"0.0801",
"0.3516"
],
[
"0.5024",
"0.3447",
"0.8896",
"0.3447",
"0.8896",
"0.358",
"0.5024",
"0.358"
],
[
"0.0801",
"0.361",
"0.3245",
"0.361",
"0.3245",
"0.375",
"0.0801",
"0.375"
],
[
"0.505",
"0.361",
"0.8906",
"0.361",
"0.8906",
"0.375",
"0.505",
"0.375"
],
[
"0.1013",
"0.379",
"0.468",
"0.379",
"0.468",
"0.3926",
"0.1013",
"0.3926"
],
[
"0.505",
"0.3757",
"0.8906",
"0.3757",
"0.8906",
"0.3894",
"0.505",
"0.3894"
],
[
"0.0825",
"0.3945",
"0.406",
"0.3945",
"0.406",
"0.4077",
"0.0825",
"0.4077"
],
[
"0.3987",
"0.3958",
"0.4673",
"0.3958",
"0.4673",
"0.4072",
"0.3987",
"0.4072"
],
[
"0.505",
"0.3901",
"0.8906",
"0.3901",
"0.8906",
"0.404",
"0.505",
"0.404"
],
[
"0.0825",
"0.411",
"0.4673",
"0.411",
"0.4673",
"0.4216",
"0.0825",
"0.4216"
],
[
"0.505",
"0.406",
"0.8906",
"0.406",
"0.8906",
"0.4192",
"0.505",
"0.4192"
],
[
"0.0825",
"0.4248",
"0.4673",
"0.4248",
"0.4673",
"0.4382",
"0.0825",
"0.4382"
],
[
"0.505",
"0.4204",
"0.8906",
"0.4204",
"0.8906",
"0.4343",
"0.505",
"0.4343"
],
[
"0.0801",
"0.4395",
"0.468",
"0.44",
"0.468",
"0.4539",
"0.0801",
"0.4526"
],
[
"0.5034",
"0.4363",
"0.8906",
"0.4363",
"0.8906",
"0.4495",
"0.5034",
"0.4495"
],
[
"0.0809",
"0.455",
"0.468",
"0.455",
"0.468",
"0.4685",
"0.0809",
"0.4685"
],
[
"0.5034",
"0.4507",
"0.8906",
"0.4507",
"0.8906",
"0.464",
"0.5034",
"0.464"
],
[
"0.08417",
"0.4702",
"0.468",
"0.4702",
"0.468",
"0.4841",
"0.08417",
"0.4841"
],
[
"0.505",
"0.4673",
"0.572",
"0.4673",
"0.572",
"0.4792",
"0.505",
"0.4792"
],
[
"0.0825",
"0.4849",
"0.468",
"0.4849",
"0.468",
"0.4988",
"0.0825",
"0.4988"
],
[
"0.5034",
"0.486",
"0.767",
"0.486",
"0.767",
"0.5",
"0.5034",
"0.5"
],
[
"0.0825",
"0.5005",
"0.4673",
"0.5005",
"0.4673",
"0.5146",
"0.0825",
"0.5146"
],
[
"0.523",
"0.506",
"0.8906",
"0.506",
"0.8906",
"0.5195",
"0.523",
"0.5195"
],
[
"0.0801",
"0.5146",
"0.4673",
"0.5166",
"0.4673",
"0.531",
"0.0801",
"0.529"
],
[
"0.505",
"0.523",
"0.653",
"0.523",
"0.653",
"0.5327",
"0.505",
"0.5327"
],
[
"0.0809",
"0.531",
"0.4673",
"0.531",
"0.4673",
"0.5444",
"0.0809",
"0.5444"
],
[
"0.0809",
"0.5444",
"0.4656",
"0.5444",
"0.4656",
"0.558",
"0.0809",
"0.558"
],
[
"0.5073",
"0.5415",
"0.888",
"0.5415",
"0.888",
"0.552",
"0.5073",
"0.552"
],
[
"0.505",
"0.5537",
"0.8906",
"0.5537",
"0.8906",
"0.5674",
"0.505",
"0.5674"
],
[
"0.0825",
"0.56",
"0.468",
"0.56",
"0.468",
"0.5737",
"0.0825",
"0.5737"
],
[
"0.501",
"0.5674",
"0.8906",
"0.5693",
"0.8906",
"0.5835",
"0.501",
"0.5806"
],
[
"0.0825",
"0.5757",
"0.468",
"0.5757",
"0.468",
"0.589",
"0.0825",
"0.589"
],
[
"0.0825",
"0.5913",
"0.468",
"0.5913",
"0.468",
"0.605",
"0.0825",
"0.605"
],
[
"0.5063",
"0.586",
"0.8896",
"0.586",
"0.8896",
"0.5967",
"0.5063",
"0.5967"
],
[
"0.5034",
"0.5996",
"0.8896",
"0.5996",
"0.8896",
"0.613",
"0.5034",
"0.613"
],
[
"0.0809",
"0.606",
"0.468",
"0.606",
"0.468",
"0.619",
"0.0809",
"0.619"
],
[
"0.0801",
"0.619",
"0.4714",
"0.622",
"0.4714",
"0.636",
"0.0801",
"0.634"
],
[
"0.5034",
"0.6143",
"0.8906",
"0.6143",
"0.8906",
"0.6274",
"0.5034",
"0.6274"
],
[
"0.505",
"0.63",
"0.8896",
"0.63",
"0.8896",
"0.643",
"0.505",
"0.643"
],
[
"0.0809",
"0.636",
"0.468",
"0.636",
"0.468",
"0.6494",
"0.0809",
"0.6494"
],
[
"0.505",
"0.6445",
"0.8906",
"0.6445",
"0.8906",
"0.6577",
"0.505",
"0.6577"
],
[
"0.0825",
"0.6514",
"0.468",
"0.6514",
"0.468",
"0.6655",
"0.0825",
"0.6655"
],
[
"0.5034",
"0.66",
"0.8906",
"0.66",
"0.8906",
"0.674",
"0.5034",
"0.674"
],
[
"0.0801",
"0.6665",
"0.468",
"0.6655",
"0.468",
"0.6797",
"0.0801",
"0.681"
],
[
"0.5034",
"0.6753",
"0.8906",
"0.6753",
"0.8906",
"0.6895",
"0.5034",
"0.6895"
],
[
"0.0809",
"0.6816",
"0.468",
"0.6816",
"0.468",
"0.696",
"0.0809",
"0.696"
],
[
"0.505",
"0.69",
"0.8906",
"0.69",
"0.8906",
"0.704",
"0.505",
"0.704"
],
[
"0.0809",
"0.6963",
"0.4673",
"0.6963",
"0.4673",
"0.711",
"0.0809",
"0.711"
],
[
"0.0809",
"0.711",
"0.468",
"0.711",
"0.468",
"0.7246",
"0.0809",
"0.7246"
],
[
"0.5034",
"0.7056",
"0.8906",
"0.7056",
"0.8906",
"0.7197",
"0.5034",
"0.7197"
],
[
"0.5034",
"0.7197",
"0.8896",
"0.7197",
"0.8896",
"0.7344",
"0.5034",
"0.7344"
],
[
"0.0809",
"0.7266",
"0.4673",
"0.7266",
"0.4673",
"0.7407",
"0.0809",
"0.7407"
],
[
"0.5034",
"0.735",
"0.8906",
"0.735",
"0.8906",
"0.7485",
"0.5034",
"0.7485"
],
[
"0.0825",
"0.741",
"0.4673",
"0.741",
"0.4673",
"0.755",
"0.0825",
"0.755"
],
[
"0.5034",
"0.7505",
"0.6235",
"0.7505",
"0.6235",
"0.7646",
"0.5034",
"0.7646"
],
[
"0.0825",
"0.757",
"0.4673",
"0.757",
"0.4673",
"0.771",
"0.0825",
"0.771"
],
[
"0.505",
"0.7666",
"0.8906",
"0.7666",
"0.8906",
"0.78",
"0.505",
"0.78"
],
[
"0.0809",
"0.773",
"0.464",
"0.773",
"0.464",
"0.786",
"0.0809",
"0.786"
],
[
"0.5024",
"0.78",
"0.8896",
"0.782",
"0.8896",
"0.7954",
"0.5024",
"0.7935"
],
[
"0.0825",
"0.787",
"0.2607",
"0.787",
"0.2607",
"0.8003",
"0.0825",
"0.8003"
],
[
"0.3096",
"0.802",
"0.4314",
"0.803",
"0.4314",
"0.8174",
"0.3096",
"0.8164"
],
[
"0.505",
"0.797",
"0.8906",
"0.797",
"0.8906",
"0.81",
"0.505",
"0.81"
],
[
"0.2467",
"0.8037",
"0.2998",
"0.8037",
"0.2998",
"0.815",
"0.2467",
"0.815"
],
[
"0.505",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.505",
"0.8247"
],
[
"0.11273",
"0.8193",
"0.21",
"0.8193",
"0.21",
"0.8306",
"0.11273",
"0.8306"
],
[
"0.2524",
"0.8184",
"0.2932",
"0.8184",
"0.2932",
"0.8335",
"0.2524",
"0.8335"
],
[
"0.3513",
"0.8193",
"0.392",
"0.8193",
"0.392",
"0.8306",
"0.3513",
"0.8306"
],
[
"0.1152",
"0.832",
"0.2296",
"0.8335",
"0.2296",
"0.8477",
"0.1152",
"0.8467"
],
[
"0.2524",
"0.832",
"0.2932",
"0.832",
"0.2932",
"0.8477",
"0.2524",
"0.8477"
],
[
"0.505",
"0.827",
"0.8896",
"0.827",
"0.8896",
"0.8403",
"0.505",
"0.8403"
],
[
"0.3545",
"0.834",
"0.3896",
"0.834",
"0.3896",
"0.8467",
"0.3545",
"0.8467"
],
[
"0.505",
"0.8413",
"0.8906",
"0.8413",
"0.8906",
"0.855",
"0.505",
"0.855"
],
[
"0.2532",
"0.8477",
"0.2893",
"0.8477",
"0.2893",
"0.863",
"0.2532",
"0.863"
],
[
"0.353",
"0.8477",
"0.3906",
"0.8477",
"0.3906",
"0.863",
"0.353",
"0.863"
],
[
"0.505",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.505",
"0.8706"
],
[
"0.11273",
"0.8486",
"0.2279",
"0.8477",
"0.2279",
"0.8623",
"0.11273",
"0.863"
],
[
"0.0809",
"0.8687",
"0.4673",
"0.8687",
"0.4673",
"0.882",
"0.0809",
"0.882"
],
[
"0.5034",
"0.872",
"0.8906",
"0.872",
"0.8906",
"0.8853",
"0.5034",
"0.8853"
],
[
"0.0784",
"0.881",
"0.4673",
"0.88",
"0.4673",
"0.8945",
"0.0784",
"0.896"
],
[
"0.505",
"0.887",
"0.8896",
"0.887",
"0.8896",
"0.901",
"0.505",
"0.901"
],
[
"0.0809",
"0.898",
"0.1422",
"0.898",
"0.1422",
"0.9077",
"0.0809",
"0.9077"
],
[
"0.4795",
"0.925",
"0.492",
"0.925",
"0.492",
"0.9375",
"0.4795",
"0.9375"
]
] | [
"Semanticfeaturefields．最近的各种工作都证明了",
"学习混合表示在神经谊染E[142-144]和2D图像合",
"成E[145]中的有效性。通过在我们的神经谊染MLP中",
"计算额外的通道，我们可以自然地将语义特征字段合并",
"",
"到我们的公式中。我们为每个图像预先计算DINO-ViT",
"",
"特征E[146]，并学习语义特征字段以在生成的实例之",
"(a)WOD-Vchicle",
"(b)WOD-Pedestrian",
"(c) Longtail-Vehicle",
"间建立零件对应关系",
"Figure 4. Image samples from our object-centric benchmark.",
"LiDAR depth supervision.",
"",
"当LiDAR点云在数据中",
"可用时，它可以通过染深度和LiDAR深度之间的重",
"4.2. Implementation Details",
"建项用作附加监督",
"GINA-3D.",
"我们的编码器以2562的分辨率拍摄图像",
"Conditional synthesis。最后但并非最不重要的一点",
"并在训练期间以1282染。我们的三平面延迟具有162",
"是，附加信息支持条件合成中的各种应用。表示为C，",
"的分辨率，其码本包含2048个条目和32的查找维度。",
"它可以作为M（siksM，C）输入到我们的潜在先验中。",
"我们使用Adam优化器E[1481在8个TeslaV100GPU",
"例如，对象比例，对象类，一天中的时间和对象语义嵌",
"上训练了模型，每个阶段的批量大小分别为32和64。",
"人也可以用作c，以控制生成过程",
"我们训练了150K步的第一阶段和80K步的第二阶段",
"Baselines.",
"我们与域中两种最先进的方法GI",
"4.Experiments",
"RAFFEE[511和EG3DE[521进行了比较，我们以1282",
"4.1. Object-centric Benchmark",
"的分辨率在数据集上对其进行了训练。我们注意到，在",
"我们选择WaymoOpenDataset(WOD）【14]，因为它",
"全像素上训练的GIRAFFE模型无法解开视点，遮挡和",
"是最大和最多样化的自动驾驶数据集之一，个",
"包含丰富",
"身份。这使得前景像素的提取变得困难，因为谊染蒙版",
"的几何和语义标签，如对象边界框和每像素实例掩码。",
"仅在低维分辨率162下定义。我们改为使用通过白化",
"具体来说，该数据集包括1,150个主要在旧金山市中心",
"非对象区域训练的模型报告数字。对于EG3D，我们观",
"和凤凰城捕获的驾驶场景，每个场景由200顿多传感",
"察到，由于缺乏前景和背景建模，使用未屏蔽图像训练",
"器观测组成。为了构建以对象为中心的基准，我们提出",
"EG3D会导致训练崩溃。因此，我们在相同的设置下训",
"了一个从粗到细的过程，通过利用3D对象框，相机-激",
"练EG3D。",
"光雷达同步和细粒度的2D全景标签来提取单视图2D",
"4.3. Evaluations on WOD-Vehicle",
"照片的集合。首先，我们利用3D框注释来排除每个数",
"我们在表中进行定量评估。2并在图中可视化不同",
"据顿中与测量车辆相距一定距离以外的物体（e。g。",
"模型的定性结果。5",
"分别用于行人的40m和用于车辆的60m）。在给定的",
"顿中，我们将每个3D边界框内的3D点云投影到最可",
"ImageEvaluation.对于图像质量，我们计算50K生",
"成的图像与所有可用验证图像之间的Frechet起始距离",
"见的相机，并提取居中的补丁以构建我们的单视图2D",
"（FID)E[1491。为了更好地反映对象完整性的指标，我",
"图像集合此外，我们使用标记子集上的2D全景分割来",
"训练Panoptic-Deeplab模型E[61,147]，并为整个数据",
"们对其对象分割掩码至少占用投影3D边界框的50%",
"的图像进行滤波（图。5-右)。此外，我们通过蒙版飘",
"集上的每个相机图像创建逐像素伪标签。这使我们能",
"够区分属于感兴趣对象、背景和遮挡物的像素(e。g。，",
"浮在联盟（蒙版FOU）来测量生成的图像的完整性，这",
"被定义为谊染对象区域中未连接像素的百分比。为了",
"站在一个人面前的柱子。我们使用2D全景预测进一",
"测量语义多样性，我们使用CLIPE[150]嵌入计算覆盖",
"步排除了对象被严重遮挡的某些斑块。即使应用了过",
"率（COV）得分和最小匹配距离（MMD)E[94]。COV",
"滤标准，我们认为，由于遮挡、类内变化(e。g。，卡车",
"测量在生成的集合中具有匹配的验证集合中的CLIP嵌",
"和轿车)，部分观察(e。g。，我们对单个车辆没有完整",
"人的比例，而MMD测量每个生成的嵌人与验证中最",
"的360度观测），而且分割不完善。特别是，考虑到相",
"机卷帘快门，物体运动和自我运动，我们提供了相机光",
"接近的嵌人之间的距离。我们的模型展示了FID、图像",
"完整性和语义多样性方面的显著改进。如果没有明确",
"线和LiDAR点云与物体坐标系的精确配准。我们重复",
"的解纠缠，基线很难处理真实的分布，导致形状不完整",
"相同的过程，从WOD中提取车辆和行人，并从我们的",
"的伪影(图。5)",
"长尾数据集中提取其他长尾车辆。拟议的以对象为中",
"几何图形评估。为了测量基础体染的一致性，我们遵",
"心的基准测试是迄今为止生成建模的最大数据集之一",
"循Oretal.[79]，并从两个视点计算体染深度之间",
"包括墅野外的各种长尾示例",
"Unique Instances",
"的对齐误差。在EG3D[52]之后，我们使用密度阈值",
"Images",
"为10的行军立方体【151]提取网格。我们通过网格漂",
"WOD-Vehicle",
"901K",
"23.6K",
"WOD-Pedestrian",
"321K",
"浮物（网格FOU）来衡量完整性，它被定义为整个网",
"8.1K",
"格中未连接的网格块上的表面积的百分比。由于我们",
"80K",
"3.7K",
"在现实世界数据中没有地面实况网格，因此我们通过",
"Longtail-Vehicle",
"Table 1. Statistics of our object-centric benchmark. Experiments",
"在验证集的边界框内测量生成的网格和聚合的LiDAR",
"were conducted on a subset with image patches rescaled to 2562",
"点云之间的网格多样性来近似网格多样性。我们使用",
"resolution",
"5"
] | {
"arxiv_id": "2304.02163",
"language": "zh",
"page_idx": 4
} |
||
2304.01830-zh-2 | 2304.01830-zh-2.jpg | 名称的质量。在某些情况下，找到类的准确语义描述可能具有挑战性，而糟糕的选择可能会损害模型的性能。据我们所知，分类和对象检测设置中尚未解决此限制文本反转。最近在$\mathbb{E}[8]$ 中引人了文本反转，用于文本到图像生成扩散模型，作为一种将图像生成条件调整为包含特定对象或视觉组件的方法。给定一组图像，所有图像都在不同的设置中显示相同的对象，学习特定的单词嵌人，以确保模型能够重建从这个小集合中采样的图像。使用简单的提示上下文(e.g.'的照片”)以确保生成对象的简单图像。训练后，该模型通过在输入句子的特定位置输人学习的单词，显示出在生成的图像中整合特定对象的令人印象深刻的能力。在[8]提出的解决方案中，每个单词都是单独学习的。
# 3. Methodology
我们考虑在大规模数据集上预先训练的视觉和语言对象识别模型。遵循多模式体系结构，该模型将图像-文本对$x=\{I,T\}$作为输人， 其中文本提供了要识别的$N$候选类$t_i$的列表； $\begin{array}{l}{T=[t_{1},t_{2},...,t_{N}]\text{。我们假设第}i\text{类特定文本}t_{i}\text{的格式为：}}\\{t_{i}=[\mathrm{promptprefix}]+[\mathrm{CLASS}]+[\mathrm{promptsuffix}]}\end{array}$ 提示前缀和后缀在所有类中共享。为了获得文本特征表示，标记器将每个句子转换为一系列标记，以便每个标记映射到嵌入$e_k$的预训练单词。然后将嵌人连接
<FigureHere>
Figure 2. Overview of our method. Classes of interest
are mapped to a set of learnable word embeddings, and integrated within the input sentence representation. These embeddings can be learned for classification and object detection tasks using standard losses, and used in conjunction with pre-trained word embeddings in open-vocabulary settings.
并用作文本转换器的输人。我们的最终文本表示是序列结束(EOS)令牌$f(T)=f_i(t_i)$的转换器输出，其中$i\in\{1,...,N\}$
# 3.1. Baseline multi-modal models
Multi- modal Image classification.
在图像分类$\mathbb{E}[29]$
的上下文中，我们的目标是将特定类别分配给整个图像。图像输入通过图像编码器$g(\cdot)$与文本输人并行处理，产生表示$g(I)$。然后通过评估图像和文本嵌人之间的余弦相似度，并选择具有最高相似度的类来简单地执行分类任务
多模态对象检测。基于视觉和语言的对象检测有两种主要的当代策略。第一种(更简单的)策略类似于多模式分类，其中图像和文本信息仅在边界框分类阶段后期组合$\mathbb{E}[1,10,26]$。在此设置中，图像输入通过对象检测器与文本模态完全分离地进行处理。检测器输出一组坐标和每个边界框候选的类嵌人。然后，我们按照多模式分类方法为每个边界框分配分类标签，使用文本表示$f(T)$计算类嵌入相似性
第二类策略建议融合图像和文本输入；在生成文本和图像嵌人$\operatorname{E}[22]$或通过第三个编码器$h(g(I),f(T))$输出边界框坐标和类嵌入$\mathbb{E}[10,17,25]$时。对于后一种融合策略，分类是在句子标记上而不是在句子之间执行的，目的是将边界框与关键字相关联。此技术通常比第一个方法类在计算上更昂贵，因为它需要为每个查询的类计算新的图像特征
# 3.2. Learning optimal class names
我们推测，预先训练的识别模型的性能与类名的适用性有内在联系，这些类名被选择来表示某些视觉类别。选择不反映对象视觉外观的类名可能会显著改变识别性能。因此，我们建议通过从图像内容中学习类特定的单词嵌人来消除模型对手工制作的类名的敏感性。图 2中提供了该方法的概述
我们考虑预先训练的单词嵌入矩阵$E\in\mathbb{R}^V\times F$ , 其中$V$是词汇大小$,F$是特征维度。我们建议使用一组新的，可学习的单词嵌人$E^l\in\mathbb{R}^N\times F$来扩展此矩阵，其中$E^l$ 的每一行对应于特定于类的可学习嵌入，而$N$ 与数据集类的总数有关。在此设置中，我们将每个文本查询中的类名替换为类特定的 占位符$t_i=[$prompt prefix$]+[\mathrm{pl}^{\mathrm{l}}]+[$promptsuffix] 。当标记器将上下文提示映射到矩阵$E$时，我们的占位符映射到其相应的嵌人$e_i^l$的类。这个新的类查询被用作文本转换器编码器的输人，并且识别模型的其余部分不变。我们注意到，可以使用$t_i=[$promptp$refix] + [$p$l_1^i] + \cdots + [$p$l_m^i] + [$p$romptsuffix]$类型的查询将一个类别映射到一系列$m$单词嵌入
培训。在训练过程中，我们冻结整个预训练模型(包括预训练的单词嵌人),并学习我们的新单词嵌入集$E^l$ ,其中包含用于训练基础模型的标准损失(用于分类的e.g.交叉熵损失和用于检测的边界框回归损失)。我们确定了三种关键方案，在这些方案中，我们的方法可能是有益的：1) 在具有一组新类的新数据集上调整
3 | [
[
"0.08905",
"0.09344",
"0.1054",
"0.09344",
"0.1054",
"0.0966",
"0.08905",
"0.0966"
],
[
"0.1283",
"0.09344",
"0.1528",
"0.09344",
"0.1528",
"0.09784",
"0.1283",
"0.09784"
],
[
"0.3162",
"0.09534",
"0.3308",
"0.09534",
"0.3308",
"0.0985",
"0.3162",
"0.0985"
],
[
"0.505",
"0.09216",
"0.8896",
"0.09216",
"0.8896",
"0.1061",
"0.505",
"0.1061"
],
[
"0.11273",
"0.101",
"0.125",
"0.101",
"0.125",
"0.1092",
"0.11273",
"0.1092"
],
[
"0.5034",
"0.1067",
"0.8896",
"0.1067",
"0.8896",
"0.1218",
"0.5034",
"0.1218"
],
[
"0.10455",
"0.11615",
"0.1381",
"0.11615",
"0.1381",
"0.1256",
"0.10455",
"0.1256"
],
[
"0.501",
"0.1218",
"0.5884",
"0.1237",
"0.5884",
"0.1383",
"0.501",
"0.1364"
],
[
"0.4126",
"0.1459",
"0.4697",
"0.1459",
"0.4697",
"0.1559",
"0.4126",
"0.1559"
],
[
"0.501",
"0.1465",
"0.763",
"0.1483",
"0.763",
"0.1635",
"0.501",
"0.161"
],
[
"0.2303",
"0.1547",
"0.3228",
"0.1547",
"0.3228",
"0.1622",
"0.2303",
"0.1622"
],
[
"0.1666",
"0.1622",
"0.2075",
"0.1622",
"0.2075",
"0.1768",
"0.1666",
"0.1768"
],
[
"0.3718",
"0.1604",
"0.4478",
"0.1604",
"0.4478",
"0.178",
"0.3718",
"0.178"
],
[
"0.2239",
"0.1686",
"0.2484",
"0.1686",
"0.2484",
"0.1863",
"0.2239",
"0.1863"
],
[
"0.5063",
"0.1718",
"0.749",
"0.1718",
"0.749",
"0.1849",
"0.5063",
"0.1849"
],
[
"0.7563",
"0.1724",
"0.888",
"0.1724",
"0.888",
"0.1863",
"0.7563",
"0.1863"
],
[
"0.429",
"0.1736",
"0.452",
"0.1736",
"0.452",
"0.1976",
"0.429",
"0.1976"
],
[
"0.1772",
"0.1768",
"0.1953",
"0.1768",
"0.1953",
"0.1812",
"0.1772",
"0.1812"
],
[
"0.7534",
"0.1768",
"0.7607",
"0.1768",
"0.7607",
"0.1799",
"0.7534",
"0.1799"
],
[
"0.505",
"0.1875",
"0.888",
"0.1875",
"0.888",
"0.2008",
"0.505",
"0.2008"
],
[
"0.308",
"0.1925",
"0.3162",
"0.1925",
"0.3162",
"0.1976",
"0.308",
"0.1976"
],
[
"0.1323",
"0.1957",
"0.1405",
"0.1957",
"0.1405",
"0.2008",
"0.1323",
"0.2008"
],
[
"0.156",
"0.1818",
"0.1642",
"0.1818",
"0.1642",
"0.1863",
"0.156",
"0.1863"
],
[
"0.505",
"0.2026",
"0.8896",
"0.2026",
"0.8896",
"0.2166",
"0.505",
"0.2166"
],
[
"0.5034",
"0.2153",
"0.888",
"0.2166",
"0.888",
"0.2311",
"0.5034",
"0.2299"
],
[
"0.165",
"0.2311",
"0.1854",
"0.2311",
"0.1854",
"0.238",
"0.165",
"0.238"
],
[
"0.505",
"0.2317",
"0.8906",
"0.2317",
"0.8906",
"0.2456",
"0.505",
"0.2456"
],
[
"0.1642",
"0.238",
"0.1912",
"0.238",
"0.1912",
"0.2507",
"0.1642",
"0.2507"
],
[
"0.2737",
"0.2524",
"0.2942",
"0.2524",
"0.2942",
"0.2615",
"0.2737",
"0.2615"
],
[
"0.5034",
"0.2474",
"0.619",
"0.2474",
"0.619",
"0.2615",
"0.5034",
"0.2615"
],
[
"0.1708",
"0.2588",
"0.2115",
"0.2664",
"0.2075",
"0.2778",
"0.1666",
"0.2708"
],
[
"0.2632",
"0.2588",
"0.3162",
"0.2588",
"0.3162",
"0.2827",
"0.2632",
"0.2827"
],
[
"0.5034",
"0.2632",
"0.8896",
"0.2632",
"0.8896",
"0.2766",
"0.5034",
"0.2766"
],
[
"0.1814",
"0.2734",
"0.2009",
"0.2798",
"0.1969",
"0.2886",
"0.1772",
"0.2822"
],
[
"0.282",
"0.2766",
"0.2998",
"0.2766",
"0.2998",
"0.2842",
"0.282",
"0.2842"
],
[
"0.505",
"0.2778",
"0.8896",
"0.2778",
"0.8896",
"0.2917",
"0.505",
"0.2917"
],
[
"0.505",
"0.2935",
"0.8906",
"0.2935",
"0.8906",
"0.307",
"0.505",
"0.307"
],
[
"0.505",
"0.308",
"0.8906",
"0.308",
"0.8906",
"0.3213",
"0.505",
"0.3213"
],
[
"0.1478",
"0.3225",
"0.2051",
"0.3225",
"0.2051",
"0.334",
"0.1478",
"0.334"
],
[
"0.5034",
"0.324",
"0.8906",
"0.324",
"0.8906",
"0.3372",
"0.5034",
"0.3372"
],
[
"0.1708",
"0.332",
"0.1887",
"0.3333",
"0.1887",
"0.3403",
"0.1691",
"0.339"
],
[
"0.3865",
"0.3333",
"0.4395",
"0.336",
"0.4387",
"0.351",
"0.3857",
"0.348"
],
[
"0.2688",
"0.3416",
"0.3105",
"0.3416",
"0.3105",
"0.356",
"0.2688",
"0.356"
],
[
"0.5034",
"0.3372",
"0.8906",
"0.3384",
"0.8906",
"0.3528",
"0.5034",
"0.3516"
],
[
"0.2695",
"0.3528",
"0.3096",
"0.3528",
"0.3096",
"0.3687",
"0.2695",
"0.3687"
],
[
"0.3987",
"0.3484",
"0.4314",
"0.3484",
"0.4314",
"0.3718",
"0.3987",
"0.3718"
],
[
"0.1805",
"0.3567",
"0.2009",
"0.3567",
"0.2009",
"0.3599",
"0.1805",
"0.3599"
],
[
"0.505",
"0.3542",
"0.892",
"0.3542",
"0.892",
"0.3674",
"0.505",
"0.3674"
],
[
"0.2769",
"0.3643",
"0.3022",
"0.3643",
"0.3022",
"0.3757",
"0.2769",
"0.3757"
],
[
"0.505",
"0.3687",
"0.7534",
"0.3687",
"0.7534",
"0.3818",
"0.505",
"0.3818"
],
[
"0.5215",
"0.3833",
"0.8906",
"0.3838",
"0.8906",
"0.3984",
"0.5215",
"0.3977"
],
[
"0.5034",
"0.3997",
"0.888",
"0.3997",
"0.888",
"0.4136",
"0.5034",
"0.4136"
],
[
"0.1528",
"0.414",
"0.468",
"0.414",
"0.468",
"0.428",
"0.1528",
"0.428"
],
[
"0.505",
"0.414",
"0.892",
"0.414",
"0.892",
"0.428",
"0.505",
"0.428"
],
[
"0.0825",
"0.428",
"0.468",
"0.428",
"0.468",
"0.4414",
"0.0825",
"0.4414"
],
[
"0.0809",
"0.4153",
"0.1602",
"0.4153",
"0.1602",
"0.4268",
"0.0809",
"0.4268"
],
[
"0.5063",
"0.4287",
"0.8906",
"0.4287",
"0.8906",
"0.4426",
"0.5063",
"0.4426"
],
[
"0.0825",
"0.443",
"0.4673",
"0.443",
"0.4673",
"0.4539",
"0.0825",
"0.4539"
],
[
"0.505",
"0.4443",
"0.8906",
"0.4443",
"0.8906",
"0.4583",
"0.505",
"0.4583"
],
[
"0.0809",
"0.4558",
"0.4656",
"0.4558",
"0.4656",
"0.4697",
"0.0809",
"0.4697"
],
[
"0.5034",
"0.4583",
"0.892",
"0.459",
"0.892",
"0.4734",
"0.5034",
"0.473"
],
[
"0.0801",
"0.4697",
"0.4673",
"0.4685",
"0.4673",
"0.4817",
"0.0801",
"0.483"
],
[
"0.505",
"0.4749",
"0.683",
"0.4749",
"0.683",
"0.488",
"0.505",
"0.488"
],
[
"0.0801",
"0.4841",
"0.4534",
"0.4841",
"0.4534",
"0.4976",
"0.0801",
"0.4976"
],
[
"0.5034",
"0.5005",
"0.7695",
"0.5005",
"0.7695",
"0.5146",
"0.5034",
"0.5146"
],
[
"0.0825",
"0.5264",
"0.4673",
"0.5264",
"0.4673",
"0.5405",
"0.0825",
"0.5405"
],
[
"0.5215",
"0.523",
"0.8906",
"0.5234",
"0.8906",
"0.5386",
"0.5215",
"0.537"
],
[
"0.0809",
"0.5425",
"0.4697",
"0.5425",
"0.4697",
"0.556",
"0.0809",
"0.556"
],
[
"0.505",
"0.539",
"0.892",
"0.539",
"0.892",
"0.553",
"0.505",
"0.553"
],
[
"0.0825",
"0.5596",
"0.4558",
"0.5596",
"0.4558",
"0.5693",
"0.0825",
"0.5693"
],
[
"0.505",
"0.5537",
"0.8906",
"0.5537",
"0.8906",
"0.5674",
"0.505",
"0.5674"
],
[
"0.0825",
"0.5737",
"0.468",
"0.5737",
"0.468",
"0.587",
"0.0825",
"0.587"
],
[
"0.505",
"0.5693",
"0.8896",
"0.5693",
"0.8896",
"0.5835",
"0.505",
"0.5835"
],
[
"0.0809",
"0.5884",
"0.468",
"0.5884",
"0.468",
"0.6016",
"0.0809",
"0.6016"
],
[
"0.5034",
"0.584",
"0.8906",
"0.5835",
"0.8906",
"0.5977",
"0.5034",
"0.5986"
],
[
"0.08417",
"0.605",
"0.4673",
"0.605",
"0.4673",
"0.6157",
"0.08417",
"0.6157"
],
[
"0.5034",
"0.5996",
"0.7295",
"0.5996",
"0.7295",
"0.613",
"0.5034",
"0.613"
],
[
"0.0825",
"0.6187",
"0.468",
"0.6187",
"0.468",
"0.632",
"0.0825",
"0.632"
],
[
"0.5215",
"0.613",
"0.892",
"0.61",
"0.892",
"0.6274",
"0.5215",
"0.631"
],
[
"0.0809",
"0.6333",
"0.468",
"0.6333",
"0.468",
"0.648",
"0.0809",
"0.648"
],
[
"0.5215",
"0.63",
"0.8906",
"0.63",
"0.8906",
"0.643",
"0.5215",
"0.643"
],
[
"0.0809",
"0.6484",
"0.468",
"0.6484",
"0.468",
"0.662",
"0.0809",
"0.662"
],
[
"0.5024",
"0.6426",
"0.8896",
"0.64",
"0.8896",
"0.6577",
"0.5024",
"0.66"
],
[
"0.0809",
"0.6636",
"0.468",
"0.6636",
"0.468",
"0.677",
"0.0809",
"0.677"
],
[
"0.5034",
"0.66",
"0.8906",
"0.66",
"0.8906",
"0.674",
"0.5034",
"0.674"
],
[
"0.08417",
"0.6797",
"0.4673",
"0.6797",
"0.4673",
"0.69",
"0.08417",
"0.69"
],
[
"0.5034",
"0.6753",
"0.8906",
"0.6753",
"0.8906",
"0.6895",
"0.5034",
"0.6895"
],
[
"0.0825",
"0.6934",
"0.4673",
"0.6934",
"0.4673",
"0.707",
"0.0825",
"0.707"
],
[
"0.505",
"0.69",
"0.8896",
"0.69",
"0.8896",
"0.704",
"0.505",
"0.704"
],
[
"0.0809",
"0.709",
"0.398",
"0.709",
"0.398",
"0.7227",
"0.0809",
"0.7227"
],
[
"0.5034",
"0.7056",
"0.888",
"0.7056",
"0.888",
"0.7197",
"0.5034",
"0.7197"
],
[
"0.523",
"0.7183",
"0.8896",
"0.7183",
"0.8896",
"0.733",
"0.523",
"0.733"
],
[
"0.0768",
"0.7393",
"0.2133",
"0.741",
"0.2133",
"0.756",
"0.0768",
"0.754"
],
[
"0.505",
"0.735",
"0.8896",
"0.735",
"0.8896",
"0.7485",
"0.505",
"0.7485"
],
[
"0.505",
"0.7505",
"0.8906",
"0.7505",
"0.8906",
"0.7646",
"0.505",
"0.7646"
],
[
"0.1013",
"0.7666",
"0.4673",
"0.7666",
"0.4673",
"0.78",
"0.1013",
"0.78"
],
[
"0.505",
"0.765",
"0.8896",
"0.765",
"0.8896",
"0.779",
"0.505",
"0.779"
],
[
"0.0825",
"0.782",
"0.468",
"0.782",
"0.468",
"0.7954",
"0.0825",
"0.7954"
],
[
"0.5034",
"0.781",
"0.8906",
"0.781",
"0.8906",
"0.7944",
"0.5034",
"0.7944"
],
[
"0.0825",
"0.797",
"0.468",
"0.797",
"0.468",
"0.81",
"0.0825",
"0.81"
],
[
"0.5034",
"0.7944",
"0.8286",
"0.7944",
"0.8286",
"0.808",
"0.5034",
"0.808"
],
[
"0.0809",
"0.811",
"0.4697",
"0.811",
"0.4697",
"0.8247",
"0.0809",
"0.8247"
],
[
"0.5034",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.5034",
"0.8247"
],
[
"0.0825",
"0.827",
"0.468",
"0.827",
"0.468",
"0.8403",
"0.0825",
"0.8403"
],
[
"0.5024",
"0.8257",
"0.888",
"0.8247",
"0.888",
"0.839",
"0.5024",
"0.8403"
],
[
"0.0825",
"0.8413",
"0.4673",
"0.8413",
"0.4673",
"0.855",
"0.0825",
"0.855"
],
[
"0.5195",
"0.8413",
"0.8896",
"0.8413",
"0.8896",
"0.855",
"0.5195",
"0.855"
],
[
"0.0809",
"0.857",
"0.4673",
"0.857",
"0.4673",
"0.8706",
"0.0809",
"0.8706"
],
[
"0.505",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.505",
"0.8706"
],
[
"0.0809",
"0.872",
"0.468",
"0.872",
"0.468",
"0.8853",
"0.0809",
"0.8853"
],
[
"0.505",
"0.872",
"0.8906",
"0.872",
"0.8906",
"0.8853",
"0.505",
"0.8853"
],
[
"0.0825",
"0.887",
"0.4673",
"0.887",
"0.4673",
"0.901",
"0.0825",
"0.901"
],
[
"0.505",
"0.887",
"0.8906",
"0.887",
"0.8906",
"0.901",
"0.505",
"0.901"
],
[
"0.478",
"0.924",
"0.4944",
"0.924",
"0.4944",
"0.938",
"0.478",
"0.938"
]
] | [
"nho",
"TCLASS",
"mhe",
"并用作文本转换器的输人。我们的最终文本表示是序",
"+",
"列结束（EOS）令牌f(T)=f(ti)的转换器输出，其中",
"fokenizer",
"ie{l....N]",
"Queried classes",
"3.1.Baseline multi-modal models",
"earnable class embeddings",
"piephalder",
"elet mathing cat",
"Cobie:",
"Multi-modal Image classification.",
"在图像分类E[29]",
"",
"toker",
"",
"的上下文中，我们的目标是将特定类别分配给整个图",
"",
"",
"",
"像。图像输人通过图像编码器g(）与文本输人并行处",
"理，产生表示g(I）。然后通过评估图像和文本嵌人之",
"Froze",
"间的余弦相似度，并选择具有最高相似度的类来简单",
"enende",
"eFroz",
"地执行分类任务",
"",
"C",
"多模态对象检测。基于视觉和语言的对象检测有两种",
"",
"node",
"主要的当代策略。第一种（更简单的）策略类似于多",
"模式分类，其中图像和文本信息仅在边界框分类阶段",
"后期组合E[1,10,26]。在此设置中，图像输人通过对",
"Text transtormel",
"象检测器与文本模态完全分离地进行处理。检测器输",
"ncode",
"",
"Ppre-traened",
"出一组坐标和每个边界框候选的类嵌人。然后，我们",
"metecodnl",
"C",
"",
"按照多模式分类方法为每个边界框分配分类标签，使",
"model",
"用文本表示f(T)计算类嵌入相似性",
"第二类策略建议融合图像和文本输入；在生成文本",
"和图像嵌入E[22]或通过第三个编码器h(g(I),f(T))输",
"Overview of our method. Classes of interest",
"出边界框坐标和类嵌人E[10,17,25]时。对于后一种融",
"are mapped to a set of learnable word embeddings, and",
"Figure 2.",
"合策略，分类是在句子标记上而不是在句子之间执行",
"integrated within the input sentence representation. These",
"的，目的是将边界框与关键字相关联。此技术通常比",
"embeddings can be learned for classification and object detec",
"第一个方法类在计算上更昂贵，因为它需要为每个查",
"tion tasks using standard losses, and used in conjunction with",
"询的类计算新的图像特征",
"pre-trained word embeddings in open-vocabulary settings.",
"3.2. Learning optimal class names",
"名称的质量。在某些情况下，找到类的准确语义描述可",
"我们推测，预先训练的识别模型的性能与类名的适",
"能具有挑战性，而糟糕的选择可能会损害模型的性能。",
"用性有内在联系，这些类名被选择来表示某些视觉类",
"据我们所知，分类和对象检测设置中尚未解决此限制",
"别。选择不反映对象视觉外观的类名可能会显著改变",
"文本反转。最近在E8]中引人了文本反转，用于文本",
"识别性能。因此，我们建议通过从图像内容中学习类",
"到图像生成扩散模型，作为一种将图像生成条件调整",
"特定的单词嵌入来消除模型对手工制作的类名的敏感",
"为包含特定对象或视觉组件的方法。给定一组图像，所",
"性。图2中提供了该方法的概述",
"有图像都在不同的设置中显示相同的对象，学习特定",
"我们考虑预先训练的单词嵌入矩阵EERVxF",
"的单词嵌人，以确保模型能够重建从这个小集合中采",
"其中V是词汇大小，F是特征维度。我们建议使",
"样的图像。使用简单的提示上下文（e.g.的照片）以",
"用一组新的，可学习的单词嵌人ELeIRNxF来扩",
"确保生成对象的简单图像。训练后，该模型通过在输",
"展此矩阵，其中E的每一行对应于特定于类的可",
"入句子的特定位置输人学习的单词，显示出在生成的",
"学习嵌入，而N与数据集类的总数有关。在此设",
"图像中整合特定对象的令人印象深刻的能力。在[8]提",
"置中，我们将每个文本查询中的类名替换为类特定",
"出的解决方案中，每个单词都是单独学习的。",
"的占位符ti=[promptprefix]+[pl]+[promptsuffix",
"当标记器将上下文提示映射到矩阵E时，我们",
"3. Methodology",
"的占位符映射到其相应的嵌人e,的类。这个新的",
"类查询被用作文本转换器编码器的输人，并且识",
"我们考虑在大规模数据集上预先训练的视",
"别模型的其余部分不变。我们注意到，可以使用",
"觉和语言对象识别模型。遵循多模式体系结",
"t,=[promptprefix]+[p]']+...+[plm]+[promptsuffix]类",
"构，该模型将图像-文本对a={I,T)作为输人，",
"型的查询将一个类别映射到一系列m单词嵌入",
"其中文本提供了要识别的N候选类t的列表：",
"培训。在训练过程中，我们冻结整个预训练模型（包",
"T=[ti,t2,tN]。我们假设第i类特定文本ti的格式为",
"括预训练的单词嵌人），并学习我们的新单词嵌入集E",
"t,=[promptprefix]+[CLASS]+[promptsuffix]，其中",
"其中包含用于训练基础模型的标准损失（用于分类",
"提示前缀和后缀在所有类中共享。为了获得文本特征",
"的e.g.交叉熵损失和用于检测的边界框回归损失）。我",
"表示，标记器将每个句子转换为一系列标记，以便每",
"们确定了三种关键方案，在这些方案中，我们的方法",
"个标记映射到嵌人e的预训练单词。然后将嵌人连接",
"可能是有益的：1）在具有一组新类的新数据集上调整",
"3"
] | {
"arxiv_id": "2304.01830",
"language": "zh",
"page_idx": 2
} |
||
2304.01830-zh-3 | 2304.01830-zh-3.jpg | 预训练模型；2) 在用于训练模型的同一数据集上调整次优类名；3) 持续调整预训练模型，其中在第一个训练阶段之后可以使用新类
推理。在推理时，我们遵循 $\color{red}{3.1}$节中描述的过程。已学习其名称的类别映射到其新嵌入$e^l\in E^l$。对于看不见的对象类别 (i.e.不在我们的新数据集中),我们使用预训练模型中的原始嵌入$e\in E$。因此，我们的解决方案保持了执行开放词汇识别的能力
# 4.Experiments
我们评估学习类名对图像分类和对象检测任务的影
响。我们考虑四个设置，我们的策略可以是有益的：
任务1：模型适应新数据集：考虑到感兴趣的新数据集， 我们通过从新图像内容中学习最佳类名，使我们的预训练模型适应这组新的潜在看不见的类
任务2：训练后类名调整：模型首先在我们的目标数据集上进行微调，然后在后续的学习阶段优化类名。此设置对于对象检测更有价值，其中微调可能对边界框定位产生重要影响
任务 3: 开放词汇识别：开放词汇是视觉语言模型的关键优势之一，保持强大的性能至关重要。评估我们学习的单词嵌人与标准的预训练单词嵌人之间的兼容性， 可以在此轴上提供强度的度量。我们将模型调整为仅来自新数据集的类的子集，并测量所有类的整体性能， 以便推理结合学习和基于文本的单词嵌人
任务4：连续模型自适应：为了衡量模型对新词嵌人进行顺序学习的能力，我们首先将模型适应于类的子集， 然后在第二个单独的训练阶段学习其余的类名
我们强调，我们不寻求最先进的性能，因为这与预训练模型的基本性能，预训练数据和体系结构有着内在的联系。相反，我们的目标是展示学习类名以提高模型性能的潜力，同时保持零拍属性。因此，除非另有说明，否则我们将在具有简单多模式机制的模型上进行实验，每个类使用单个单词嵌人，并在所有实验设置中使用标准${\text{提示：(promptprefix=`aphotoofa'promptsuffix=:")}}$
# 4.1. Image classification: CLIP model
为了评估我们的分类任务方法，我们采用CLIP$\boxed{\mathrm{E}}[29]$ 作为我们的基本预训练模型，并使用视觉转换器ViTB/ 16$\overline {\mathrm{E} }[ 3]$来实现图像编码器$g(\cdot)$角色。在这里，我们试图评估先前描述的任务1、3和4中的方法性能；i.e. 模型自适应、开放词汇表性能和持续自适应
数据集。我们遵循[40,41]中建议的设置。我们使CLIP模型适应11个不同数据集的选择，重点是不同类型的识别任务：标准图像分类 (ImageNet$\underline{\mathrm{E}}[2]$, Caltech$101\underline{\mathrm{E}}[5]$ ),细粒度分类 (斯坦福汽车$\mathbb{E}[18]$, 牛津宠物$\mathbb{E}[28]$,花朵102$\mathbb{E}[27]$,食品101$\mathbb{E}$XMATHX3MATD动作识别( 培训详细信息。我们采用了一组类似于[41]的参数。对于每个数据集，我们在200个epoch(用于ImageNet 的50)上学习批处理大小为32的类名称，学习速率为$2e^{-4}$的SGD优化器和余弦调度，以及一个预热时期。我们使用与手工制作的类名相关的所有单词标记的平均嵌入来初始化我们的可学习单词嵌人。对于每个数
<table>
<tbody>
<tr>
<th>Class group</th>
<th>$\mathrm{CLIP}^{*}$</th>
<th>CoOp</th>
<th>Ours</th>
<th>Ours</th>
</tr>
<tr>
<td>Baseclasses</td>
<td>69.4</td>
<td>82.6</td>
<td>82.7</td>
<td>82.8</td>
</tr>
<tr>
<td>Newclasses</td>
<td>74.2</td>
<td>64.1</td>
<td>73.1</td>
<td>74.2</td>
</tr>
<tr>
<td>$\operatorname{Allclasses}$</td>
<td>65.2</td>
<td>65.7</td>
<td>66.5</td>
<td>68.7</td>
</tr>
</tbody>
</table>
Table 1. CLIP model: Base to new classification accuracy. Results on Base classes show method ability to adapt to new data (Task 1 ). Results on New and All classes show open vocabulary ability (Task 3). * Manually engineered prompt $\operatorname{templates}.$
<table>
<tbody>
<tr>
<th>Class group</th>
<th>$\mathrm{CLIP}^{*}$</th>
<th>CoOp</th>
<th>Ours</th>
<th>$Ours^{\cdot}$</th>
</tr>
<tr>
<td>Newclasses</td>
<td>74.0</td>
<td>84.4</td>
<td>85.9</td>
<td>85.9</td>
</tr>
<tr>
<td>$\operatorname{Allclasses}$</td>
<td>65.2</td>
<td>71.7</td>
<td>75.7</td>
<td>76.3</td>
</tr>
</tbody>
</table>
Table 2. CLIP model: Classification accuracy for continual model adaptation (Task 4) experiments. * manually engineered prompt templates.
据集，每个类别采样16个训练图像，我们报告三个随
机种子的平均统计量
基线。我们直接比较的方法是零拍CLIP模型，因为我们的目标是评估学习类名的好处。我们还报告了CoOp模型$\boxed{\mathbb{E}}[41]$的分类准确性，该模型通过固定的手工类名称学习最佳提示上下文。我们在感兴趣的设置中重新训练CoOp, 并使用四个上下文单词，用’aphotoofa’初始化， 就像在[40]中一样 (这与$\mathbb{E}[41]$中使用的16个单词不同)。这允许比较学习类名称在提示上下文中的优缺点，并评估策略之间的任何互补性。我们使用工程提示报告最佳的CLIP性能，并使用标准提示以及使用与CLIP相同的工程提示报告结果。我们注意到四个数据集 (斯坦福汽车，ImageNet,加州理工学院101和Sun397)没有工程提示 , 我们在补充材料中提供了所有特定于数据集的提示模型适配(任务1)和开放词汇表(任务3)。我们采用[40] 中提出的基到新的泛化设置。对于每个数据集，我们同样将类分为一组“基本”和“新”类。我们使用基类进行培训，并以零拍的方式评估新类的表现。此外，为了衡量开放词汇表的性能，我们报告了“所有”类的性能，i.e. 混合了学习和手工制作的类名。结果见表1。基类的结果突出了我们通过仅学习类名来执行模型适应 (任务1) 的能力。对于此任务，我们观察到我们的策略可以实现与基类上的上下文调整等效的收益(82。6%(Coop)???。$8\%$(我们的)平均精度)。但是，我们的方法保证在新类上保持零拍CLIP性能，因为我们学习的参数是特定于类的。我们的开放词汇实验(任务3)表明，我们能够保持CLIP模型扩展输出集的能力，并仅从文本中引入新类 (CLIP和Ours74。2个%, 新类)。对所有类的实验证实，我们能够成功地使用学习类和新类的混合，优于CoOp(65。7%CoOp,68。7个%Ours,所有类)。我们进一步注意到对上下文提示的敏感性，因为使用工程提示获得了可观的收益(+2。2个%Ours,所有类)。这是意料之中的，因为新的类性能可能会受到提示的强烈影响， 并强调了在此设置中进行提示上下文调整的重要性
4 | [
[
"0.0809",
"0.09216",
"0.4673",
"0.09216",
"0.4673",
"0.1061",
"0.0809",
"0.1061"
],
[
"0.556",
"0.09344",
"0.6274",
"0.09534",
"0.6274",
"0.1067",
"0.556",
"0.1048"
],
[
"0.642",
"0.0947",
"0.6855",
"0.0947",
"0.6855",
"0.1048",
"0.642",
"0.1048"
],
[
"0.746",
"0.0947",
"0.785",
"0.0947",
"0.785",
"0.1061",
"0.746",
"0.1061"
],
[
"0.794",
"0.0947",
"0.834",
"0.0947",
"0.834",
"0.1048",
"0.794",
"0.1048"
],
[
"0.0825",
"0.108",
"0.468",
"0.108",
"0.468",
"0.1218",
"0.0825",
"0.1218"
],
[
"0.699",
"0.09534",
"0.74",
"0.09534",
"0.74",
"0.1061",
"0.699",
"0.1061"
],
[
"0.642",
"0.1111",
"0.6733",
"0.1111",
"0.6733",
"0.1225",
"0.642",
"0.1225"
],
[
"0.6978",
"0.10986",
"0.7295",
"0.11237",
"0.728",
"0.125",
"0.697",
"0.1225"
],
[
"0.751",
"0.11237",
"0.7793",
"0.11237",
"0.7793",
"0.1237",
"0.751",
"0.1237"
],
[
"0.796",
"0.11237",
"0.824",
"0.11237",
"0.824",
"0.1237",
"0.796",
"0.1237"
],
[
"0.6436",
"0.1218",
"0.6733",
"0.1237",
"0.6724",
"0.1364",
"0.642",
"0.1338"
],
[
"0.5596",
"0.11237",
"0.6284",
"0.11237",
"0.6284",
"0.1225",
"0.5596",
"0.1225"
],
[
"0.7524",
"0.1225",
"0.7793",
"0.1225",
"0.7793",
"0.1351",
"0.7524",
"0.1351"
],
[
"0.794",
"0.1218",
"0.826",
"0.1218",
"0.826",
"0.137",
"0.794",
"0.137"
],
[
"0.699",
"0.1237",
"0.728",
"0.1237",
"0.728",
"0.1351",
"0.699",
"0.1351"
],
[
"0.0825",
"0.125",
"0.2573",
"0.125",
"0.2573",
"0.1351",
"0.0825",
"0.1351"
],
[
"0.5596",
"0.1237",
"0.6274",
"0.1237",
"0.6274",
"0.1351",
"0.5596",
"0.1351"
],
[
"0.5596",
"0.1332",
"0.619",
"0.1351",
"0.6177",
"0.1497",
"0.558",
"0.1477"
],
[
"0.751",
"0.1332",
"0.7812",
"0.1351",
"0.7793",
"0.1483",
"0.749",
"0.1465"
],
[
"0.794",
"0.1338",
"0.824",
"0.1364",
"0.8228",
"0.1483",
"0.7935",
"0.1465"
],
[
"0.6978",
"0.1364",
"0.728",
"0.1364",
"0.728",
"0.1477",
"0.6978",
"0.1477"
],
[
"0.0809",
"0.1383",
"0.468",
"0.1383",
"0.468",
"0.1515",
"0.0809",
"0.1515"
],
[
"0.6436",
"0.1351",
"0.6733",
"0.1351",
"0.6733",
"0.1477",
"0.6436",
"0.1477"
],
[
"0.0825",
"0.1528",
"0.4673",
"0.1528",
"0.4673",
"0.1666",
"0.0825",
"0.1666"
],
[
"0.5024",
"0.1622",
"0.888",
"0.1635",
"0.888",
"0.178",
"0.5024",
"0.1768"
],
[
"0.0825",
"0.1692",
"0.4673",
"0.1692",
"0.4673",
"0.1799",
"0.0825",
"0.1799"
],
[
"0.0825",
"0.1831",
"0.4673",
"0.1831",
"0.4673",
"0.1963",
"0.0825",
"0.1963"
],
[
"0.5034",
"0.1787",
"0.886",
"0.1787",
"0.886",
"0.1895",
"0.5034",
"0.1895"
],
[
"0.0801",
"0.1963",
"0.326",
"0.1976",
"0.326",
"0.2122",
"0.0801",
"0.2108"
],
[
"0.5034",
"0.1913",
"0.8896",
"0.1913",
"0.8896",
"0.2052",
"0.5034",
"0.2052"
],
[
"0.5034",
"0.2058",
"0.8896",
"0.2058",
"0.8896",
"0.2197",
"0.5034",
"0.2197"
],
[
"0.505",
"0.2203",
"0.572",
"0.2203",
"0.572",
"0.2317",
"0.505",
"0.2317"
],
[
"0.0801",
"0.2267",
"0.2115",
"0.2267",
"0.2115",
"0.2405",
"0.0801",
"0.2405"
],
[
"0.10046",
"0.252",
"0.4673",
"0.252",
"0.4673",
"0.2651",
"0.10046",
"0.2651"
],
[
"0.5557",
"0.2524",
"0.6396",
"0.2551",
"0.6396",
"0.2695",
"0.5557",
"0.2676"
],
[
"0.636",
"0.2556",
"0.835",
"0.2556",
"0.835",
"0.2664",
"0.636",
"0.2664"
],
[
"0.0809",
"0.2664",
"0.4429",
"0.2664",
"0.4429",
"0.2798",
"0.0809",
"0.2798"
],
[
"0.5557",
"0.2695",
"0.6284",
"0.2715",
"0.6284",
"0.286",
"0.5557",
"0.2842"
],
[
"0.6445",
"0.2727",
"0.6733",
"0.2727",
"0.6733",
"0.2854",
"0.6445",
"0.2854"
],
[
"0.699",
"0.2708",
"0.728",
"0.2727",
"0.726",
"0.2854",
"0.6978",
"0.2827"
],
[
"0.751",
"0.2727",
"0.7793",
"0.2727",
"0.7793",
"0.2842",
"0.751",
"0.2842"
],
[
"0.796",
"0.2727",
"0.824",
"0.2727",
"0.824",
"0.2842",
"0.796",
"0.2842"
],
[
"0.08417",
"0.2827",
"0.4656",
"0.2827",
"0.4656",
"0.2935",
"0.08417",
"0.2935"
],
[
"0.5596",
"0.2854",
"0.619",
"0.2854",
"0.619",
"0.2966",
"0.5596",
"0.2966"
],
[
"0.6436",
"0.2854",
"0.6733",
"0.2854",
"0.6733",
"0.2966",
"0.6436",
"0.2966"
],
[
"0.6953",
"0.2854",
"0.7295",
"0.2854",
"0.7295",
"0.2966",
"0.6953",
"0.2966"
],
[
"0.749",
"0.2842",
"0.7812",
"0.2842",
"0.7812",
"0.2966",
"0.749",
"0.2966"
],
[
"0.79",
"0.2854",
"0.824",
"0.2854",
"0.824",
"0.2966",
"0.79",
"0.2966"
],
[
"0.0809",
"0.2966",
"0.468",
"0.2966",
"0.468",
"0.31",
"0.0809",
"0.31"
],
[
"0.0825",
"0.3125",
"0.3594",
"0.3125",
"0.3594",
"0.3257",
"0.0825",
"0.3257"
],
[
"0.5034",
"0.3132",
"0.8906",
"0.3132",
"0.8906",
"0.327",
"0.5034",
"0.327"
],
[
"0.0825",
"0.327",
"0.468",
"0.327",
"0.468",
"0.3403",
"0.0825",
"0.3403"
],
[
"0.501",
"0.3257",
"0.8896",
"0.327",
"0.8896",
"0.3416",
"0.501",
"0.3403"
],
[
"0.0825",
"0.3423",
"0.468",
"0.3423",
"0.468",
"0.356",
"0.0825",
"0.356"
],
[
"0.505",
"0.3435",
"0.669",
"0.3435",
"0.669",
"0.3542",
"0.505",
"0.3542"
],
[
"0.0809",
"0.3567",
"0.4673",
"0.3567",
"0.4673",
"0.3706",
"0.0809",
"0.3706"
],
[
"0.0809",
"0.3726",
"0.2133",
"0.3726",
"0.2133",
"0.3865",
"0.0809",
"0.3865"
],
[
"0.0825",
"0.387",
"0.468",
"0.387",
"0.468",
"0.4016",
"0.0825",
"0.4016"
],
[
"0.505",
"0.387",
"0.8896",
"0.387",
"0.8896",
"0.3977",
"0.505",
"0.3977"
],
[
"0.0809",
"0.4016",
"0.4673",
"0.4016",
"0.4673",
"0.4153",
"0.0809",
"0.4153"
],
[
"0.5034",
"0.401",
"0.6504",
"0.401",
"0.6504",
"0.414",
"0.5034",
"0.414"
],
[
"0.0825",
"0.4172",
"0.4673",
"0.4172",
"0.4673",
"0.4312",
"0.0825",
"0.4312"
],
[
"0.5034",
"0.4172",
"0.8906",
"0.4172",
"0.8906",
"0.4312",
"0.5034",
"0.4312"
],
[
"0.0825",
"0.432",
"0.468",
"0.432",
"0.468",
"0.4458",
"0.0825",
"0.4458"
],
[
"0.505",
"0.4343",
"0.8896",
"0.4343",
"0.8896",
"0.4443",
"0.505",
"0.4443"
],
[
"0.0825",
"0.4475",
"0.4697",
"0.4475",
"0.4697",
"0.461",
"0.0825",
"0.461"
],
[
"0.5034",
"0.4458",
"0.8906",
"0.4463",
"0.8906",
"0.461",
"0.5034",
"0.4602"
],
[
"0.0825",
"0.4634",
"0.3733",
"0.4634",
"0.3733",
"0.4766",
"0.0825",
"0.4766"
],
[
"0.5063",
"0.464",
"0.8906",
"0.464",
"0.8906",
"0.4749",
"0.5063",
"0.4749"
],
[
"0.0825",
"0.4778",
"0.4673",
"0.4778",
"0.4673",
"0.4912",
"0.0825",
"0.4912"
],
[
"0.5034",
"0.4778",
"0.888",
"0.4778",
"0.888",
"0.4912",
"0.5034",
"0.4912"
],
[
"0.0809",
"0.4937",
"0.468",
"0.4937",
"0.468",
"0.507",
"0.0809",
"0.507"
],
[
"0.5034",
"0.4937",
"0.8896",
"0.4937",
"0.8896",
"0.507",
"0.5034",
"0.507"
],
[
"0.0809",
"0.5083",
"0.406",
"0.5083",
"0.406",
"0.5215",
"0.0809",
"0.5215"
],
[
"0.5034",
"0.5083",
"0.8896",
"0.5083",
"0.8896",
"0.5215",
"0.5034",
"0.5215"
],
[
"0.1013",
"0.5234",
"0.4673",
"0.5234",
"0.4673",
"0.537",
"0.1013",
"0.537"
],
[
"0.505",
"0.5234",
"0.8906",
"0.5234",
"0.8906",
"0.537",
"0.505",
"0.537"
],
[
"0.0809",
"0.5386",
"0.468",
"0.5386",
"0.468",
"0.552",
"0.0809",
"0.552"
],
[
"0.5034",
"0.5386",
"0.8906",
"0.5386",
"0.8906",
"0.552",
"0.5034",
"0.552"
],
[
"0.0809",
"0.5537",
"0.4673",
"0.5537",
"0.4673",
"0.5674",
"0.0809",
"0.5674"
],
[
"0.5034",
"0.553",
"0.8906",
"0.553",
"0.8906",
"0.5664",
"0.5034",
"0.5664"
],
[
"0.08417",
"0.5693",
"0.4673",
"0.5693",
"0.4673",
"0.58",
"0.08417",
"0.58"
],
[
"0.5024",
"0.5664",
"0.8906",
"0.5674",
"0.8906",
"0.582",
"0.5024",
"0.5806"
],
[
"0.08417",
"0.5854",
"0.4673",
"0.5854",
"0.4673",
"0.595",
"0.08417",
"0.595"
],
[
"0.505",
"0.5854",
"0.8896",
"0.5854",
"0.8896",
"0.595",
"0.505",
"0.595"
],
[
"0.0825",
"0.5986",
"0.4673",
"0.5986",
"0.4673",
"0.6123",
"0.0825",
"0.6123"
],
[
"0.5034",
"0.601",
"0.888",
"0.601",
"0.888",
"0.6143",
"0.5034",
"0.6143"
],
[
"0.0809",
"0.6143",
"0.464",
"0.6143",
"0.464",
"0.6274",
"0.0809",
"0.6274"
],
[
"0.5034",
"0.6143",
"0.8896",
"0.6157",
"0.8896",
"0.629",
"0.5034",
"0.6274"
],
[
"0.5034",
"0.629",
"0.8896",
"0.63",
"0.8896",
"0.6445",
"0.5034",
"0.643"
],
[
"0.0825",
"0.636",
"0.3743",
"0.636",
"0.3743",
"0.6494",
"0.0825",
"0.6494"
],
[
"0.5034",
"0.646",
"0.8906",
"0.646",
"0.8906",
"0.659",
"0.5034",
"0.659"
],
[
"0.10297",
"0.66",
"0.4673",
"0.66",
"0.4673",
"0.674",
"0.10297",
"0.674"
],
[
"0.505",
"0.66",
"0.888",
"0.66",
"0.888",
"0.674",
"0.505",
"0.674"
],
[
"0.0825",
"0.675",
"0.464",
"0.675",
"0.464",
"0.688",
"0.0825",
"0.688"
],
[
"0.5034",
"0.675",
"0.8906",
"0.675",
"0.8906",
"0.688",
"0.5034",
"0.688"
],
[
"0.0809",
"0.69",
"0.4673",
"0.69",
"0.4673",
"0.704",
"0.0809",
"0.704"
],
[
"0.505",
"0.6924",
"0.8896",
"0.6924",
"0.8896",
"0.7026",
"0.505",
"0.7026"
],
[
"0.0809",
"0.7056",
"0.4673",
"0.7056",
"0.4673",
"0.7197",
"0.0809",
"0.7197"
],
[
"0.5034",
"0.7056",
"0.8906",
"0.7056",
"0.8906",
"0.7197",
"0.5034",
"0.7197"
],
[
"0.0809",
"0.7217",
"0.392",
"0.7217",
"0.392",
"0.735",
"0.0809",
"0.735"
],
[
"0.5063",
"0.72",
"0.8906",
"0.72",
"0.8906",
"0.7344",
"0.5063",
"0.7344"
],
[
"0.0825",
"0.735",
"0.468",
"0.735",
"0.468",
"0.7485",
"0.0825",
"0.7485"
],
[
"0.5034",
"0.7363",
"0.8906",
"0.7363",
"0.8906",
"0.75",
"0.5034",
"0.75"
],
[
"0.0809",
"0.752",
"0.4673",
"0.752",
"0.4673",
"0.765",
"0.0809",
"0.765"
],
[
"0.505",
"0.752",
"0.8896",
"0.752",
"0.8896",
"0.765",
"0.505",
"0.765"
],
[
"0.0825",
"0.7666",
"0.4673",
"0.7666",
"0.4673",
"0.78",
"0.0825",
"0.78"
],
[
"0.505",
"0.7666",
"0.8906",
"0.7666",
"0.8906",
"0.78",
"0.505",
"0.78"
],
[
"0.0809",
"0.78",
"0.468",
"0.782",
"0.468",
"0.7954",
"0.0809",
"0.7935"
],
[
"0.5034",
"0.781",
"0.8896",
"0.781",
"0.8896",
"0.7944",
"0.5034",
"0.7944"
],
[
"0.0825",
"0.797",
"0.4673",
"0.797",
"0.4673",
"0.811",
"0.0825",
"0.811"
],
[
"0.5034",
"0.797",
"0.8906",
"0.797",
"0.8906",
"0.81",
"0.5034",
"0.81"
],
[
"0.0809",
"0.811",
"0.468",
"0.811",
"0.468",
"0.8247",
"0.0809",
"0.8247"
],
[
"0.5034",
"0.811",
"0.888",
"0.811",
"0.888",
"0.8247",
"0.5034",
"0.8247"
],
[
"0.08417",
"0.827",
"0.468",
"0.827",
"0.468",
"0.8403",
"0.08417",
"0.8403"
],
[
"0.505",
"0.8276",
"0.876",
"0.8276",
"0.876",
"0.8384",
"0.505",
"0.8384"
],
[
"0.0809",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0809",
"0.855"
],
[
"0.505",
"0.8433",
"0.8896",
"0.8433",
"0.8896",
"0.854",
"0.505",
"0.854"
],
[
"0.0809",
"0.857",
"0.468",
"0.857",
"0.468",
"0.8706",
"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",
"0.8706",
"0.8906",
"0.8726",
"0.8906",
"0.886",
"0.5024",
"0.884"
],
[
"0.0809",
"0.887",
"0.468",
"0.887",
"0.468",
"0.901",
"0.0809",
"0.901"
],
[
"0.5034",
"0.887",
"0.863",
"0.887",
"0.863",
"0.901",
"0.5034",
"0.901"
],
[
"0.4795",
"0.925",
"0.4927",
"0.925",
"0.4927",
"0.936",
"0.4795",
"0.936"
]
] | [
"预训练模型；2）在用于训练模型的同一数据集上调整",
"Classgroup",
"CLIP*",
"Ours",
"Ours*",
"次优类名；3）持续调整预训练模型，其中在第一个训",
"CoOp",
"69.4",
"82.6",
"82.7",
"82.8",
"74.2",
"Baseclasses",
"73.1",
"74.2",
"64.1",
"练阶段之后可以使用新类",
"Newclasses",
"Allclasses",
"66.5",
"68.7",
"65.7",
"推理。在推理时，我们遵循3.1节中描述的过程。已学",
"65.2",
"习其名称的类别映射到其新嵌人eEE。对于看不见",
"Table 1. CLIP model: Base to new classification accuracy",
"的对象类别（i.e.不在我们的新数据集中），我们使用",
"预训练模型中的原始嵌人eEE。因此，我们的解决方",
"Results onBase classes showmethod abilitytoadapttonew",
"案保持了执行开放词汇识别的能力",
"data (Task 1 ). Results on New and All classes show open",
"vocabulary ability (Task 3). * Manually engineered prompt",
"templates.",
"4.Experiments",
"我们评估学习类名对图像分类和对象检测任务的影",
"Classgroup Il",
"CLIP*CoOpOursOurs*",
"响。我们考虑四个设置，我们的策略可以是有益的",
" Newclasses",
"74.0",
"84.4",
"85.9",
"85.9",
"任务1：模型适应新数据集：考虑到感兴趣的新数据集",
"Allclasses",
"65.2",
"71.7",
"75.7",
"76.3",
"我们通过从新图像内容中学习最佳类名，使我们的预",
"训练模型适应这组新的潜在看不见的类",
"Table 2. CLIP model: Classification accuracy for continual",
"任务2：训练后类名调整：模型首先在我们的目标数据",
"model adaptation (Task 4) experiments. * manually engi-",
"集上进行微调，然后在后续的学习阶段优化类名。此",
"neeredprompttemplates.",
"设置对于对象检测更有价值，其中微调可能对边界框",
"定位产生重要影响",
"任务3：开放词汇识别：开放词汇是视觉语言模型的关",
"据集，每个类别采样16个训练图像：我们报告三个随",
"键优势之一，保持强大的性能至关重要。评估我们学",
"机种子的平均统计量",
"习的单词嵌入与标准的预训练单词嵌人之间的兼容性",
"基线。我们直接比较的方法是零拍CLIP模型，因为我们",
"可以在此轴上提供强度的度量。我们将模型调整为仅",
"的自标是评估学习类名的好处。我们还报告了CoOp模",
"来自新数据集的类的子集，并测量所有类的整体性能，",
"型E[41]的分类准确性，该模型通过固定的手工类名称",
"以便推理结合学习和基于文本的单词嵌入",
"学习最佳提示上下文。我们在感兴趣的设置中重新训练",
"任务4：连续模型自适应：为了衡量模型对新词嵌入进",
"CoOp，并使用四个上下文单词，用aphotoofa初始化",
"行顺序学习的能力，我们首先将模型适应于类的子集，",
"就像在[40]中一样（这与[41]中使用的16个单词不同)。",
"然后在第二个单独的训练阶段学习其余的类名",
"这允许比较学习类名称在提示上下文中的优缺点，并评",
"我们强调，我们不寻求最先进的性能，因为这与预训",
"估策略之间的任何互补性。我们使用工程提示报告最佳",
"练模型的基本性能，预训练数据和体系结构有着内在的",
"的CLIP性能，并使用标准提示以及使用与CLIP相同的",
"联系。相反，我们的目标是展示学习类名以提高模型性",
"工程提示报告结果。我们注意到四个数据集（斯坦福汽",
"能的潜力，同时保持零拍属性。因此，除非另有说明，否",
"车，ImageNet，加州理工学院101和Sun397）没有工程提",
"则我们将在具有简单多模式机制的模型上进行实验，每",
"示，我们在补充材料中提供了所有特定于数据集的提示",
"个类使用单个单词嵌人，并在所有实验设置中使用标准",
"模型适配（任务1）和开放词汇表（任务3）。我们采用【40",
"提示：（promptprefix=‘aphotoofa'；promptsuffix=）",
"中提出的基到新的泛化设置。对于每个数据集，我们同",
"样将类分为一组“基本”和“新”类。我们使用基类进行",
"4.1. Image classification: CLIP model",
"培训，并以零拍的方式评估新类的表现。此外，为了衡",
"为了评估我们的分类任务方法，我们采用CLIPE[29]",
"量开放词汇表的性能，我们报告了“所有”类的性能，i.e",
"作为我们的基本预训练模型，并使用视觉转换器ViT",
"混合了学习和手工制作的类名。结果见表1。基类的结",
"B/16E[3]来实现图像编码器g()角色。在这里，我们",
"果突出了我们通过仅学习类名来执行模型适应（任务1",
"试图评估先前描述的任务1、3和4中的方法性能；i.e.",
"的能力。对于此任务，我们观察到我们的策略可以实现",
"模型自适应、开放词汇表性能和持续自适应",
"与基类上的上下文调整等效的收益（82。6%(Coop)??？。",
"数据集。我们遵循[40.41]中建议的设置。我们使CLIP模",
"8%（我们的）平均精度）。但是，我们的方法保证在新",
"型适应11个不同数据集的选择，重点是不同类型的识别",
"类上保持零拍CLIP性能，因为我们学习的参数是特定",
"任务：标准图像分类（ImageNetE2]，Caltech101E[5",
"于类的。我们的开放词汇实验（任务3）表明，我们能够",
").细粒度分类（斯坦福汽车E[18]，牛津宠物E[28]，花",
"保持CLIP模型扩展输出集的能力，并仅从文本中引人",
"朵102E[27]，食品101EXMATHX3MATD动作识别（",
"新类（CLIP和Ours74。2个%，新类）。对所有类的实验",
"培训详细信息。我们采用了一组类似于[41]的参数。对",
"证实，我们能够成功地使用学习类和新类的混合，优于",
"于每个数据集，我们在200个epoch（用于ImageNet",
"CoOp(65。7%CoOp.68。7个%Ours，所有类）。我们进",
"的50）上学习批处理大小为32的类名称，学习速率为",
"步注意到对上下文提示的敏感性，因为使用工程提示",
"2e-4的SGD优化器和余弦调度，以及一个预热时期。",
"获得了可观的收益(+2。2个%Ours,所有类)。这是意料",
"我们使用与手工制作的类名相关的所有单词标记的平",
"之中的，因为新的类性能可能会受到提示的强烈影响，",
"均嵌人来初始化我们的可学习单词嵌人。对于每个数",
"并强调了在此设置中进行提示上下文调整的重要性",
"4"
] | {
"arxiv_id": "2304.01830",
"language": "zh",
"page_idx": 3
} |
||
2305.03117-zh-1 | 2305.03117-zh-1.jpg | # 人类解释总是有帮助的吗？ 对人类自然语言解释的客观评估
Bingsheng Yao
Rensselaer Polytechnic Institute
Lucian Popa IBM Research
Dakuo Wang T
Northeastern University
## James Hendler
Rensselaer Polytechnic Institute
## $\mathbf{Abstract}$
Human-annotated labels and explanations are critical for training explainable NLP models. However, unlike human-annotated labels whose quality is easier to calibrate (e.g., with a majority vote), human-crafted free-form explanations can be quite subjective. Before blindly using them as ground truth to train ML models, a vital question needs to be asked: How do we evaluate a human-annotated explanation' s quality? In this paper, we build on the view that the quality of a human-annotated explanation can be measured based on its helpfulness (or impairment) to the ML models ' performance for the desired NLP tasks for which the annotations were collected. In comparison to the commonly used Simulatability score, we define a new metric that can take into consideration of the helpfulness of an explanation for model performance at both fine-tuning and inference. With the help of a unified dataset format, we evaluated the proposed metric on five datasets (e.g., e-SNLI) against two model architectures (T5 and BART), and the results show that our proposed metric can objectively evaluate the quality of human-annotated explanations, while Simulatability falls short.
## 1 Introduction
尽管大规模语言模型(LLM)$\operatorname{E}$(Devlin et al., 2019; Qin et al., 2023; Lewis et al., 2019; Raffel et al., 2020) 的最新进展，它们在许多自然语言处理(NLP) 任务上表现出接近人类的性能 (e。g。, 问答 (Rajpurkar et al.,2016; Kočiský et al., 2018; Mou et al., 2020, 2021; Xu et al., 2022) , 自然语言推理 (Bowman et al., 2015; Williams et al., 2017; Wang et al., 2018) , 文本生成 (Duan et al., 2017; Yao et al., 2022; Zhao et al., 2022) ),
$^{*}$ Work done while Prithviraj was at IBM Research.
$^{\dagger}[email protected] Corresponding Author.
人类渴望知道最先进的 (SOTA) 模型是如何得出预测的。从事自然语言解释的研究人员$^{1}$ 通过招募人群或专家来注释标签和相应的自然语言解释，从而向人类注释者寻求帮助$\overline{\mathrm{E}}$ (Camburu et al., 2018; Rajani et al., 2019; Aggarwal et al., 2021; Wang et al., 2019b); 因此，研究人员可以利用人工注释的解释来提高模型的预测性能或训练模型以生成人类可理解的自然语言解释
然而，人工注释解释的质量问题还有待探讨。研究人员经常利用流行的自然语言生成(NLG)指标，例如 BLEU $\overline{\mathrm{E}}$(Papineni et al., 2002) 和 ROUGE $\boxed{\mathrm{F}}($Lin, 2004) 来评估模型生成的解释与人工注释的解释之间的相似性，并强烈假设人工注释的解释是黄金标准。不过，与为分类或多项选择 QA 任务$\overline{\mathrm{E}}($Chen et al., 2021)提供标签不同，不同的人可能会对同一观察结果$\overline{\mathrm{E}}$ (Gebreegziabher et al., 2023) 提出不同的自然语言解释。即使 BLEU 或 ROUGE 相似性可能很低，两个这样的解释也可能都是正确的。此外， 人类给出的自然语言解释通常可能是主观的和任务相关的 (Lee et al.,2022)。因此，人工注释的解释不应被简单地视为黄金标准$\overline{\mathrm{E}}($Muller et al.,2021); 相反，我们认为解释的核心价值应基于它们对模型预测提供的帮助，而不是基于语义相似度或单词匹配的概念
总结我们在本文中的贡献：
???。我们提供了一个客观的评估，以量化人工注释的解释对模型性能的帮助。我们的评估指标是 Simulatability 得分$\overline{\mathrm{E}}$(Doshi-Velez and Kim, 2017) 的扩展，我们提出了一种基于提示的统一数据格式，可以将分类或多项选择任务转换为统一的多项选择生成任务格式，以最大程度地减少不同任务之间结构变化的影响。
In this paper, we use“explanations”and“natural language explanations" to refer to the collective concepts of free-text rationales" and“natural language explanation", which differ from “rule-based” or “extractive”explanations. | [
[
"0.3586",
"0.0932",
"0.644",
"0.09265",
"0.644",
"0.10986",
"0.3586",
"0.111"
],
[
"0.3333",
"0.1152",
"0.665",
"0.1152",
"0.665",
"0.1289",
"0.3333",
"0.1289"
],
[
"0.1897",
"0.158",
"0.3198",
"0.158",
"0.3198",
"0.1729",
"0.1897",
"0.1729"
],
[
"0.434",
"0.158",
"0.5693",
"0.158",
"0.5693",
"0.1729",
"0.434",
"0.1729"
],
[
"0.693",
"0.155",
"0.806",
"0.1567",
"0.806",
"0.1746",
"0.693",
"0.1729"
],
[
"0.12006",
"0.1758",
"0.3804",
"0.1758",
"0.3804",
"0.1895",
"0.12006",
"0.1895"
],
[
"0.4626",
"0.1733",
"0.5356",
"0.1758",
"0.5356",
"0.1907",
"0.4626",
"0.1882"
],
[
"0.69",
"0.1758",
"0.81",
"0.1758",
"0.81",
"0.1895",
"0.69",
"0.1895"
],
[
"0.2678",
"0.2245",
"0.3972",
"0.2245",
"0.3972",
"0.2382",
"0.2678",
"0.2382"
],
[
"0.6055",
"0.2227",
"0.7314",
"0.2227",
"0.7314",
"0.2405",
"0.6055",
"0.2405"
],
[
"0.2024",
"0.2411",
"0.4626",
"0.2411",
"0.4626",
"0.2546",
"0.2024",
"0.2546"
],
[
"0.5684",
"0.2405",
"0.763",
"0.2411",
"0.763",
"0.2559",
"0.5684",
"0.2546"
],
[
"0.2627",
"0.304",
"0.3418",
"0.304",
"0.3418",
"0.3184",
"0.2627",
"0.3184"
],
[
"0.5146",
"0.304",
"0.879",
"0.304",
"0.879",
"0.3176",
"0.5146",
"0.3176"
],
[
"0.5156",
"0.3225",
"0.869",
"0.3225",
"0.869",
"0.3333",
"0.5156",
"0.3333"
],
[
"0.867",
"0.3237",
"0.876",
"0.3237",
"0.876",
"0.3289",
"0.867",
"0.3289"
],
[
"0.1453",
"0.332",
"0.4575",
"0.332",
"0.4575",
"0.3457",
"0.1453",
"0.3457"
],
[
"0.511",
"0.3362",
"0.8774",
"0.3362",
"0.8774",
"0.3499",
"0.511",
"0.3499"
],
[
"0.1453",
"0.3457",
"0.4575",
"0.3457",
"0.4575",
"0.3594",
"0.1453",
"0.3594"
],
[
"0.513",
"0.3528",
"0.882",
"0.3528",
"0.882",
"0.3665",
"0.513",
"0.3665"
],
[
"0.1453",
"0.361",
"0.4575",
"0.361",
"0.4575",
"0.3748",
"0.1453",
"0.3748"
],
[
"0.511",
"0.3694",
"0.882",
"0.3706",
"0.882",
"0.3843",
"0.511",
"0.383"
],
[
"0.1453",
"0.376",
"0.4592",
"0.376",
"0.4592",
"0.3896",
"0.1453",
"0.3896"
],
[
"0.511",
"0.3855",
"0.881",
"0.3855",
"0.881",
"0.3984",
"0.511",
"0.3984"
],
[
"0.147",
"0.3914",
"0.4575",
"0.3914",
"0.4575",
"0.402",
"0.147",
"0.402"
],
[
"0.1453",
"0.4038",
"0.4575",
"0.4038",
"0.4575",
"0.4175",
"0.1453",
"0.4175"
],
[
"0.513",
"0.402",
"0.881",
"0.402",
"0.881",
"0.415",
"0.513",
"0.415"
],
[
"0.1436",
"0.4175",
"0.4592",
"0.4163",
"0.4592",
"0.4312",
"0.1436",
"0.4316"
],
[
"0.5146",
"0.4192",
"0.881",
"0.4192",
"0.881",
"0.43",
"0.5146",
"0.43"
],
[
"0.147",
"0.434",
"0.4568",
"0.434",
"0.4568",
"0.4448",
"0.147",
"0.4448"
],
[
"0.511",
"0.434",
"0.5884",
"0.434",
"0.5884",
"0.449",
"0.511",
"0.449"
],
[
"0.1453",
"0.4465",
"0.4575",
"0.4465",
"0.4575",
"0.4602",
"0.1453",
"0.4602"
],
[
"0.531",
"0.452",
"0.882",
"0.452",
"0.882",
"0.4656",
"0.531",
"0.4656"
],
[
"0.1453",
"0.4602",
"0.4575",
"0.4602",
"0.4575",
"0.4739",
"0.1453",
"0.4739"
],
[
"0.513",
"0.4685",
"0.881",
"0.4685",
"0.881",
"0.4822",
"0.513",
"0.4822"
],
[
"0.1453",
"0.4744",
"0.4592",
"0.4744",
"0.4592",
"0.488",
"0.1453",
"0.488"
],
[
"0.1453",
"0.488",
"0.4592",
"0.488",
"0.4592",
"0.502",
"0.1453",
"0.502"
],
[
"0.511",
"0.484",
"0.881",
"0.484",
"0.881",
"0.4976",
"0.511",
"0.4976"
],
[
"0.1453",
"0.503",
"0.4592",
"0.503",
"0.4592",
"0.5166",
"0.1453",
"0.5166"
],
[
"0.511",
"0.5005",
"0.879",
"0.5005",
"0.879",
"0.514",
"0.511",
"0.514"
],
[
"0.147",
"0.5186",
"0.4568",
"0.5186",
"0.4568",
"0.5293",
"0.147",
"0.5293"
],
[
"0.5156",
"0.5166",
"0.881",
"0.5166",
"0.881",
"0.53",
"0.5156",
"0.53"
],
[
"0.147",
"0.533",
"0.4568",
"0.533",
"0.4568",
"0.5435",
"0.147",
"0.5435"
],
[
"0.5156",
"0.533",
"0.881",
"0.533",
"0.881",
"0.5464",
"0.5156",
"0.5464"
],
[
"0.147",
"0.5474",
"0.4568",
"0.5474",
"0.4568",
"0.558",
"0.147",
"0.558"
],
[
"0.513",
"0.549",
"0.881",
"0.549",
"0.881",
"0.5625",
"0.513",
"0.5625"
],
[
"0.1453",
"0.56",
"0.4575",
"0.56",
"0.4575",
"0.5737",
"0.1453",
"0.5737"
],
[
"0.511",
"0.5635",
"0.881",
"0.5635",
"0.881",
"0.5767",
"0.511",
"0.5767"
],
[
"0.147",
"0.576",
"0.4568",
"0.576",
"0.4568",
"0.586",
"0.147",
"0.586"
],
[
"0.51",
"0.579",
"0.881",
"0.5806",
"0.881",
"0.5957",
"0.51",
"0.5933"
],
[
"0.1453",
"0.5884",
"0.4575",
"0.5884",
"0.4575",
"0.6016",
"0.1453",
"0.6016"
],
[
"0.5156",
"0.5967",
"0.881",
"0.5967",
"0.881",
"0.61",
"0.5156",
"0.61"
],
[
"0.1453",
"0.6025",
"0.4575",
"0.6025",
"0.4575",
"0.616",
"0.1453",
"0.616"
],
[
"0.1453",
"0.616",
"0.4575",
"0.616",
"0.4575",
"0.6313",
"0.1453",
"0.6313"
],
[
"0.5146",
"0.6133",
"0.882",
"0.6133",
"0.882",
"0.627",
"0.5146",
"0.627"
],
[
"0.1436",
"0.63",
"0.4592",
"0.629",
"0.4592",
"0.6436",
"0.1436",
"0.6445"
],
[
"0.5146",
"0.629",
"0.881",
"0.629",
"0.881",
"0.6426",
"0.5146",
"0.6426"
],
[
"0.1453",
"0.6445",
"0.4575",
"0.6445",
"0.4575",
"0.6577",
"0.1453",
"0.6577"
],
[
"0.5146",
"0.6445",
"0.881",
"0.6445",
"0.881",
"0.6577",
"0.5146",
"0.6577"
],
[
"0.1453",
"0.659",
"0.4575",
"0.659",
"0.4575",
"0.673",
"0.1453",
"0.673"
],
[
"0.511",
"0.66",
"0.881",
"0.66",
"0.881",
"0.6733",
"0.511",
"0.6733"
],
[
"0.1453",
"0.6733",
"0.4592",
"0.6733",
"0.4592",
"0.687",
"0.1453",
"0.687"
],
[
"0.513",
"0.6777",
"0.881",
"0.6777",
"0.881",
"0.6914",
"0.513",
"0.6914"
],
[
"0.147",
"0.6895",
"0.4434",
"0.6895",
"0.4434",
"0.6997",
"0.147",
"0.6997"
],
[
"0.5146",
"0.6934",
"0.881",
"0.6934",
"0.881",
"0.707",
"0.5146",
"0.707"
],
[
"0.513",
"0.709",
"0.7715",
"0.7104",
"0.7715",
"0.7236",
"0.513",
"0.7227"
],
[
"0.11584",
"0.7163",
"0.1511",
"0.7163",
"0.1511",
"0.728",
"0.11584",
"0.728"
],
[
"0.1503",
"0.7173",
"0.2644",
"0.7173",
"0.2644",
"0.73",
"0.1503",
"0.73"
],
[
"0.532",
"0.727",
"0.7407",
"0.727",
"0.7407",
"0.7407",
"0.532",
"0.7407"
],
[
"0.12006",
"0.741",
"0.4844",
"0.741",
"0.4844",
"0.755",
"0.12006",
"0.755"
],
[
"0.513",
"0.7437",
"0.881",
"0.7437",
"0.881",
"0.7573",
"0.513",
"0.7573"
],
[
"0.11755",
"0.7583",
"0.4878",
"0.7583",
"0.4878",
"0.7715",
"0.11755",
"0.7715"
],
[
"0.511",
"0.7583",
"0.881",
"0.759",
"0.881",
"0.774",
"0.511",
"0.7725"
],
[
"0.11755",
"0.774",
"0.4844",
"0.774",
"0.4844",
"0.7876",
"0.11755",
"0.7876"
],
[
"0.513",
"0.7754",
"0.882",
"0.7754",
"0.882",
"0.789",
"0.513",
"0.789"
],
[
"0.11755",
"0.788",
"0.4878",
"0.7905",
"0.4878",
"0.804",
"0.11755",
"0.802"
],
[
"0.511",
"0.792",
"0.881",
"0.792",
"0.881",
"0.8057",
"0.511",
"0.8057"
],
[
"0.11584",
"0.8057",
"0.4895",
"0.8047",
"0.4895",
"0.8193",
"0.11584",
"0.821"
],
[
"0.513",
"0.808",
"0.8774",
"0.808",
"0.8774",
"0.822",
"0.513",
"0.822"
],
[
"0.11584",
"0.8193",
"0.4902",
"0.822",
"0.4902",
"0.836",
"0.11584",
"0.8345"
],
[
"0.513",
"0.826",
"0.879",
"0.826",
"0.879",
"0.836",
"0.513",
"0.836"
],
[
"0.1184",
"0.836",
"0.4895",
"0.8374",
"0.4895",
"0.8516",
"0.1184",
"0.851"
],
[
"0.511",
"0.8403",
"0.8486",
"0.8403",
"0.8486",
"0.854",
"0.511",
"0.854"
],
[
"0.11755",
"0.854",
"0.486",
"0.854",
"0.486",
"0.8677",
"0.11755",
"0.8677"
],
[
"0.11584",
"0.868",
"0.4878",
"0.869",
"0.4878",
"0.8843",
"0.11584",
"0.883"
],
[
"0.534",
"0.8706",
"0.882",
"0.8706",
"0.882",
"0.8843",
"0.534",
"0.8843"
],
[
"0.51",
"0.883",
"0.882",
"0.882",
"0.882",
"0.8955",
"0.51",
"0.8965"
],
[
"0.1436",
"0.8926",
"0.4568",
"0.8926",
"0.4568",
"0.9062",
"0.1436",
"0.9062"
],
[
"0.511",
"0.8945",
"0.881",
"0.8945",
"0.881",
"0.9077",
"0.511",
"0.9077"
],
[
"0.1436",
"0.904",
"0.4768",
"0.905",
"0.4768",
"0.9194",
"0.1436",
"0.9185"
],
[
"0.511",
"0.9062",
"0.875",
"0.9062",
"0.875",
"0.9194",
"0.511",
"0.9194"
]
] | [
"人类解释总是有帮助的吗？",
"对人类自然语言解释的客观评估",
"Bingsheng Yao",
"Prithviraj Sen *",
"Lucian Popa",
"Rensselaer Polytechnic Institute",
"Amazon",
"IBM Research",
"James Hendler",
"Dakuo Wang ",
"Rensselaer Polytechnic Institute",
"Northeastern University",
"Abstract",
"人类渴望知道最先进的（SOTA）模型是如何得",
"出预测的。从事自然语言解释的研究人员",
"",
"Human-annotated labels and explanations are",
"通过招募人群或专家来注释标签和相应的自",
"critical for training explainable NLP mod-",
"然语言解释，从而向人类注释者寻求帮助E",
"els. However, unlike human-annotated labels",
"(Camburu et al., 2018; Rajani et al., 2019; Aggar-",
"whose quality is easier to calibrate (e.g., with a",
"waletal.,2021;Wangetal.,2019b)；因此，研",
"majorityvote),human-craftedfree-formexpla-",
"nations can be quite subjective. Before blindly",
"究人员可以利用人工注释的解释来提高模型的",
"using them as ground truth to train ML mod-",
"预测性能或训练模型以生成人类可理解的自然",
"els, a vital question needs to be asked: How do",
"语言解释",
"we evaluate a human-annotated explanation's",
"然而，人工注释解释的质量问题还有待探",
"quality? In this paper, we build on the view",
"讨。研究人员经常利用流行的自然语言生成",
"that the quality of a human-annotated expla-",
"nation can be measured based on its helpful-",
"(NLG）指标，例如BLEUE(Papineni etal.,2002)",
"ness (or impairment) to the ML models ' perfor-",
"和ROUGEE(Lin,2004)来评估模型生成的解释",
"manceforthe desiredNLPtasksfor whichthe",
"与人工注释的解释之间的相似性，并强烈假设",
"annotationswerecollected.In comparisonto",
"人工注释的解释是黄金标准。不过，与为分类",
"the commonly used Simulatability score,",
"或多项选择QA任务E(Chenetal.,2021）提供",
"we define a new metric that can take into con-",
"标签不同，不同的人可能会对同一观察结果E",
"siderationofthehelpfulnessofanexplanation",
"（Gebreegziabheretal.，2023）提出不同的自然语",
"for model performance at both fine-tuning and",
"言解释。即使BLEU或ROUGE相似性可能很",
"inference. With the help of a unified dataset",
"format, we evaluated the proposed metric on",
"低，两个这样的解释也可能都是正确的。此外，",
"five datasets (e.g., e-SNLI) against two model",
"人类给出的自然语言解释通常可能是主观的和",
"architectures (T5 and BART), and the results",
"任务相关的（Leeetal.,2022）。因此，人工注释",
"show that our proposed metric can objectively",
"的解释不应被简单地视为黄金标准E(Muller",
"evaluate the quality of human-annotated expla-",
"etal.,2021）；相反，我们认为解释的核心价值",
"nations, while Simulatabilityfalls short.",
"应基于它们对模型预测提供的帮助，而不是基",
"于语义相似度或单词匹配的概念",
"1",
"Introduction",
"总结我们在本文中的贡献：",
"尽管大规模语言模型（LLM)E(Devlinetal.,",
"?？？。我们提供了一个客观的评估，以量化人工",
"2019; Qin et al., 2023; Lewis et al., 2019; Raffel",
"注释的解释对模型性能的帮助。我们的评估指",
"etal.,2020)的最新进展，它们在许多自然语言",
"标是Simulatability得分E(Doshi-Velezand",
"处理（NLP）任务上表现出接近人类的性能（e。",
"Kim,2017）的扩展，我们提出了一种基于提示",
"g。，问答(Rajpurkar et al., 2016; Kocisky et al.,",
"的统一数据格式，可以将分类或多项选择任务",
"2018; Mou et al., 2020, 2021; Xu et al., 2022) ,",
"转换为统一的多项选择生成任务格式，以最大",
"自然语言推理（Bowmanetal.，2015；Williams",
"程度地减少不同任务之间结构变化的影响。",
"et al.,2017;Wanget al.,2018)，文本生成(Duan",
"et al., 2017; Yao et al., 2022; Zhao et al., 2022)),",
"'In this paper, we use \"explanations\" and \"natural lan-",
"guage explanations\" to refer to the collective concepts of",
" Work done while Prithviraj was at IBM Research.",
"\"free-text rationales\" and \"natural language explanation\",",
"td.wang@northeastern. edu Corresponding Author.",
"which differ from \"rule-based\" or “extractive\" explanations."
] | {
"arxiv_id": "2305.03117",
"language": "zh",
"page_idx": 1
} |
||
2305.03117-zh-2 | 2305.03117-zh-2.jpg | <table>
<tbody>
<tr>
<th rowspan="2">Dataset</th>
<th rowspan="2">Task</th>
<th rowspan="2">Task Format</th>
<th colspan="3">Data Instances</th>
<th rowspan="2">Average explanatior Length ( token) 1</th>
</tr>
<tr>
<th>Train</th>
<th>Valid</th>
<th>Test</th>
</tr>
<tr>
<td>CoS- E v1. 0</td>
<td>Commonsense $QA$</td>
<td>3-choice Multiple-Choice</td>
<td>7610</td>
<td>950</td>
<td> </td>
<td>16.148</td>
</tr>
<tr>
<td>CoS- E v1. 11</td>
<td>Commonsense $OA$</td>
<td>5-choice Multiple-Choice</td>
<td>9741</td>
<td>1221</td>
<td> </td>
<td>8.996</td>
</tr>
<tr>
<td>$ECQA$</td>
<td>Commonsense $QA$</td>
<td>5-choice Multiple-Choice</td>
<td>7598</td>
<td>1098</td>
<td>2194</td>
<td>63.572</td>
</tr>
<tr>
<td>e- SNLI</td>
<td>Natural Language Inference</td>
<td>3-label Classification</td>
<td>549367</td>
<td>9842</td>
<td>9824</td>
<td>15.977</td>
</tr>
<tr>
<td>ComVE</td>
<td>Commonsense Validation</td>
<td>2-choice Multiple-Choice</td>
<td>10000</td>
<td>1000</td>
<td>1000</td>
<td>10.288</td>
</tr>
</tbody>
</table>
Table l: Task description and core statistics for five popular large-scale datasets with human-annotated natural
language explanations that are included in our evaluation.
中的人工注释解释
值得一提的是，我们不包括 SBIC(Sap et al., 2019) 或 E- $\delta$ -NLI (Brahman et al., 2021) 等数据集。SBIC 不提供所有数据的解释，E- $\delta$ -NLI 利用各种来源来增加$\delta$ -NLI (Rudinger et al., 2020) 数据集的解释，而不是提供人工注释
???。通过对五个数据集和两个模型的评估， 我们的指标可以在两个模型架构上对所有五个数据集的解释质量进行一致的排名，而Simulatability 得分(基线)不足。
???。我们的评估证明了这样的假设，即人类的解释仍然可以使模型预测受益，即使它们被先前文献的人类评估批评为低质量。
2.2 Evaluation Metric for Explanations
## 2 $\textbf{Related Work}$
2.1 Natural Language Explanation Datasets
许多常用的基于文本的内容评估指标，如BLEU (Papineni et al.,2002)和 ROUGE (Lin, 2004), 都将人工注释的答案视为绝对的黄金标准，而无需质疑或尝试评估其质量。一个已建立的评估指标称为 Simulatability 分数，该指标源自人类可模拟性(Doshi-Velez) and Kim, 2017) , 可以检查黄金解释。它只是测量基线模型预测性能的变化，具体取决于是否将解释作为输入提供。以前的作品(Chandrasekaran et al., 2018; Yeung et al., 2020; Hase et al., 2020; Wiegreffe et al., 2020; PoursabziSangdeh et al., 2021; Rajagopal et al., 2021) 已经证明了 Simulatability 分数对评估解释质量的有用性。然而，这个指标有几个固有的缺点。首先，它仅考虑了基线模型上解释的帮助性， 在这里我们通过E4 节中的实验表明，解释在微调和推理过程中提供了不同的帮助性。此外，当我们将原始任务转换为其他任务时，模型性能也可能会有所不同，例如将分类任务转换为具有不同输人数据格式的多项选择任务
为了客观地评估人工注释的解释，我们基于Simulatability 分数定义了一个新的评估指标，该指标通过考虑解释在微调和推理方面的帮助来补充 Simulatability 的两个缺点，并借助统一的结构来最大程度地减少任务差异的影响。其他作品$\overline{\mathrm{E}}($Carton et al.,2020)试图对解释的不同特征进行评估和分类，但是其中许多$\boxed{\mathrm{E}}($Chan et al., 2022a; DeYoung et al., 2020) 仍然将人类注释的解释视为金标准
尽管开发了新的模型体系结构和可能更重要的参数，但这些“黑匣子”不可避免地缺乏解释其预测的能力；这导致社区越来越多地利用人工注释的解释来训练带有解释的模型，或者教他们自我合理化。例如，Wiegreffe and Marasovic (2021)审查了65 个数据集，并提供了 3 类分类的解释：突出显示，自由文本和结构化。我们专注于五个大型公共数据集，这些数据集在实例级别具有自由文本的人工注释解释(表$\mathbb{E}$ 1 )。我们仔细检查了这些数据集的许可证，发现不存在任何个人身份信息(PII)
一个突出的数据集是 CoS-E 及其两个变体CoS-Ev1。0 和 CoS-Ev1。11 (Rajani et al., 2019) 。它扩展了常识问答(CQAv1。0 和 v1。11 个版本)数据集(Talmor et al.,2018) , 方法是在正确答案标签上添加人工注释的解释。但是，最近的一些工作表明 CoS-E 的解释质量不佳，因为Narang et al.(2020)独立地手工标记了 CoS-E 的一些新解释，并发现其原始解释与新解释之间的 BLEU 得分非常低。为了提高解释的质量，ECQA (Aggarwal et al., 2021)收集了每个候选答案的单句解释，并将其汇总为 CQAv1中每个数据的自然语言解释。11 数据集。Sun et al.(2022)证明了CoS-E 解释不如基于人类偏好的 ECQA 解释好。第四个数据集是 e-SNLI (Camburu et al., 2018) , 其中包括对斯坦福自然语言(SNLI)数据集(Bowman et al.,2015)的解释。最后，第五个数据集是 ComVE (Wang et al.,2020), 询问两个句子中的哪一个违反常识。稍后，我们使用我们的指标和已建立的基线 Simulatability 分数评估上述五个数据集
2.3 Usage of Explanations for SOTA models
现有工作一直在探索解释可以提高模型性能的情况；例如，Hase and Bansal(2021)认为解释最适合用作预测的模型输入，而 Kumar and | [
[
"0.5996",
"0.0873",
"0.69",
"0.0885",
"0.69",
"0.0998",
"0.5996",
"0.0986"
],
[
"0.733",
"0.0885",
"0.8564",
"0.0885",
"0.8564",
"0.0986",
"0.733",
"0.0986"
],
[
"0.444",
"0.0903",
"0.525",
"0.09265",
"0.525",
"0.1069",
"0.444",
"0.1051"
],
[
"0.1351",
"0.09265",
"0.1873",
"0.09265",
"0.1873",
"0.1069",
"0.1351",
"0.1069"
],
[
"0.293",
"0.09143",
"0.3284",
"0.09143",
"0.3284",
"0.1069",
"0.293",
"0.1069"
],
[
"0.58",
"0.1051",
"0.6187",
"0.1051",
"0.6187",
"0.11755",
"0.58",
"0.11755"
],
[
"0.635",
"0.1057",
"0.677",
"0.1057",
"0.677",
"0.11755",
"0.635",
"0.11755"
],
[
"0.679",
"0.1057",
"0.714",
"0.1057",
"0.714",
"0.11816",
"0.679",
"0.11816"
],
[
"0.75",
"0.10095",
"0.839",
"0.10095",
"0.839",
"0.11224",
"0.75",
"0.11224"
],
[
"0.1335",
"0.1217",
"0.2091",
"0.1217",
"0.2091",
"0.136",
"0.1335",
"0.136"
],
[
"0.2527",
"0.1223",
"0.367",
"0.1247",
"0.367",
"0.136",
"0.2527",
"0.1342"
],
[
"0.4072",
"0.1223",
"0.5635",
"0.1223",
"0.5635",
"0.136",
"0.4072",
"0.136"
],
[
"0.58",
"0.1194",
"0.6187",
"0.1223",
"0.617",
"0.1372",
"0.5767",
"0.1354"
],
[
"0.6396",
"0.1223",
"0.6665",
"0.1223",
"0.6665",
"0.1354",
"0.6396",
"0.1354"
],
[
"0.7715",
"0.1223",
"0.818",
"0.1223",
"0.818",
"0.1372",
"0.7715",
"0.1372"
],
[
"0.1335",
"0.136",
"0.2141",
"0.136",
"0.2141",
"0.1478",
"0.1335",
"0.1478"
],
[
"0.251",
"0.136",
"0.3687",
"0.136",
"0.3687",
"0.1497",
"0.251",
"0.1497"
],
[
"0.4055",
"0.1342",
"0.561",
"0.136",
"0.561",
"0.1497",
"0.4055",
"0.1478"
],
[
"0.581",
"0.136",
"0.615",
"0.136",
"0.615",
"0.1484",
"0.581",
"0.1484"
],
[
"0.635",
"0.136",
"0.6675",
"0.136",
"0.6675",
"0.1484",
"0.635",
"0.1484"
],
[
"0.775",
"0.136",
"0.8135",
"0.136",
"0.8135",
"0.1484",
"0.775",
"0.1484"
],
[
"0.2494",
"0.1467",
"0.367",
"0.1478",
"0.367",
"0.1621",
"0.2494",
"0.1609"
],
[
"0.4072",
"0.1467",
"0.5615",
"0.1484",
"0.5615",
"0.1621",
"0.4072",
"0.1603"
],
[
"0.636",
"0.1467",
"0.6724",
"0.1484",
"0.671",
"0.1609",
"0.635",
"0.1592"
],
[
"0.681",
"0.1478",
"0.7173",
"0.1497",
"0.7153",
"0.1609",
"0.679",
"0.1592"
],
[
"0.771",
"0.1478",
"0.818",
"0.1478",
"0.818",
"0.1621",
"0.771",
"0.1621"
],
[
"0.581",
"0.1484",
"0.617",
"0.1484",
"0.617",
"0.1609",
"0.581",
"0.1609"
],
[
"0.1335",
"0.1478",
"0.178",
"0.1497",
"0.1763",
"0.1621",
"0.1318",
"0.1603"
],
[
"0.574",
"0.1592",
"0.6245",
"0.1609",
"0.623",
"0.1758",
"0.5728",
"0.1733"
],
[
"0.634",
"0.1609",
"0.6724",
"0.1633",
"0.671",
"0.1746",
"0.6323",
"0.1729"
],
[
"0.681",
"0.1603",
"0.7153",
"0.1621",
"0.714",
"0.1746",
"0.679",
"0.1729"
],
[
"0.2258",
"0.1621",
"0.3914",
"0.1621",
"0.3914",
"0.1729",
"0.2258",
"0.1729"
],
[
"0.1335",
"0.1621",
"0.1814",
"0.1621",
"0.1814",
"0.1733",
"0.1335",
"0.1733"
],
[
"0.4224",
"0.1633",
"0.5483",
"0.1633",
"0.5483",
"0.1733",
"0.4224",
"0.1733"
],
[
"0.7734",
"0.1621",
"0.8154",
"0.1621",
"0.8154",
"0.1733",
"0.7734",
"0.1733"
],
[
"0.1318",
"0.1733",
"0.1873",
"0.1733",
"0.1873",
"0.1882",
"0.1318",
"0.1882"
],
[
"0.2317",
"0.1746",
"0.3896",
"0.1746",
"0.3896",
"0.1882",
"0.2317",
"0.1882"
],
[
"0.4055",
"0.1746",
"0.5635",
"0.1746",
"0.5635",
"0.1882",
"0.4055",
"0.1882"
],
[
"0.58",
"0.1746",
"0.6187",
"0.1746",
"0.6187",
"0.187",
"0.58",
"0.187"
],
[
"0.636",
"0.1746",
"0.671",
"0.1746",
"0.671",
"0.187",
"0.636",
"0.187"
],
[
"0.6826",
"0.1746",
"0.7173",
"0.1746",
"0.7173",
"0.187",
"0.6826",
"0.187"
],
[
"0.7715",
"0.1733",
"0.817",
"0.1733",
"0.817",
"0.1882",
"0.7715",
"0.1882"
],
[
"0.1184",
"0.1953",
"0.879",
"0.1953",
"0.879",
"0.2102",
"0.1184",
"0.2102"
],
[
"0.11755",
"0.2114",
"0.504",
"0.2114",
"0.504",
"0.2245",
"0.11755",
"0.2245"
],
[
"0.11755",
"0.2363",
"0.4937",
"0.2369",
"0.4937",
"0.2517",
"0.11755",
"0.2505"
],
[
"0.511",
"0.2382",
"0.6616",
"0.2382",
"0.6616",
"0.2517",
"0.511",
"0.2517"
],
[
"0.1184",
"0.2542",
"0.486",
"0.2542",
"0.486",
"0.2673",
"0.1184",
"0.2673"
],
[
"0.5283",
"0.2542",
"0.8774",
"0.253",
"0.8774",
"0.2673",
"0.5283",
"0.2683"
],
[
"0.11755",
"0.2683",
"0.4878",
"0.2695",
"0.4878",
"0.284",
"0.11755",
"0.2832"
],
[
"0.51",
"0.2683",
"0.879",
"0.2695",
"0.879",
"0.284",
"0.51",
"0.2832"
],
[
"0.11755",
"0.2861",
"0.413",
"0.2861",
"0.413",
"0.2998",
"0.11755",
"0.2998"
],
[
"0.5146",
"0.2861",
"0.879",
"0.2861",
"0.879",
"0.2998",
"0.5146",
"0.2998"
],
[
"0.1184",
"0.3018",
"0.4844",
"0.3018",
"0.4844",
"0.3154",
"0.1184",
"0.3154"
],
[
"0.511",
"0.3018",
"0.8774",
"0.3018",
"0.8774",
"0.3154",
"0.511",
"0.3154"
],
[
"0.11755",
"0.3176",
"0.486",
"0.3176",
"0.486",
"0.3308",
"0.11755",
"0.3308"
],
[
"0.511",
"0.3176",
"0.806",
"0.3176",
"0.806",
"0.3308",
"0.511",
"0.3308"
],
[
"0.1184",
"0.3342",
"0.3838",
"0.3342",
"0.3838",
"0.3474",
"0.1184",
"0.3474"
],
[
"0.511",
"0.3474",
"0.839",
"0.3474",
"0.839",
"0.3623",
"0.511",
"0.3623"
],
[
"0.1142",
"0.3682",
"0.1587",
"0.3682",
"0.1587",
"0.3843",
"0.1142",
"0.3843"
],
[
"0.1503",
"0.3694",
"0.2747",
"0.3694",
"0.2747",
"0.383",
"0.1503",
"0.383"
],
[
"0.511",
"0.3694",
"0.879",
"0.3694",
"0.879",
"0.383",
"0.511",
"0.383"
],
[
"0.511",
"0.3855",
"0.8774",
"0.3855",
"0.8774",
"0.3984",
"0.511",
"0.3984"
],
[
"0.11755",
"0.398",
"0.4785",
"0.398",
"0.4785",
"0.412",
"0.11755",
"0.412"
],
[
"0.511",
"0.401",
"0.881",
"0.402",
"0.881",
"0.4163",
"0.511",
"0.415"
],
[
"0.1184",
"0.4204",
"0.4844",
"0.4204",
"0.4844",
"0.434",
"0.1184",
"0.434"
],
[
"0.513",
"0.4187",
"0.8774",
"0.4187",
"0.8774",
"0.4316",
"0.513",
"0.4316"
],
[
"0.1184",
"0.4358",
"0.4844",
"0.4358",
"0.4844",
"0.4495",
"0.1184",
"0.4495"
],
[
"0.513",
"0.4353",
"0.881",
"0.4353",
"0.881",
"0.449",
"0.513",
"0.449"
],
[
"0.1184",
"0.4531",
"0.4844",
"0.4531",
"0.4844",
"0.466",
"0.1184",
"0.466"
],
[
"0.513",
"0.4507",
"0.881",
"0.4507",
"0.881",
"0.4656",
"0.513",
"0.4656"
],
[
"0.12006",
"0.4702",
"0.4836",
"0.4702",
"0.4836",
"0.481",
"0.12006",
"0.481"
],
[
"0.511",
"0.466",
"0.881",
"0.466",
"0.881",
"0.481",
"0.511",
"0.481"
],
[
"0.12006",
"0.485",
"0.486",
"0.485",
"0.486",
"0.4988",
"0.12006",
"0.4988"
],
[
"0.5146",
"0.4827",
"0.881",
"0.4827",
"0.881",
"0.4963",
"0.5146",
"0.4963"
],
[
"0.11755",
"0.502",
"0.4836",
"0.502",
"0.4836",
"0.5156",
"0.11755",
"0.5156"
],
[
"0.513",
"0.4988",
"0.879",
"0.4988",
"0.879",
"0.5127",
"0.513",
"0.5127"
],
[
"0.1184",
"0.5186",
"0.4878",
"0.5186",
"0.4878",
"0.532",
"0.1184",
"0.532"
],
[
"0.513",
"0.5156",
"0.881",
"0.5156",
"0.881",
"0.5293",
"0.513",
"0.5293"
],
[
"0.1184",
"0.5337",
"0.486",
"0.5337",
"0.486",
"0.5474",
"0.1184",
"0.5474"
],
[
"0.511",
"0.531",
"0.882",
"0.53",
"0.882",
"0.5444",
"0.511",
"0.546"
],
[
"0.1184",
"0.55",
"0.4878",
"0.55",
"0.4878",
"0.5635",
"0.1184",
"0.5635"
],
[
"0.513",
"0.5474",
"0.881",
"0.5474",
"0.881",
"0.561",
"0.513",
"0.561"
],
[
"0.1184",
"0.5654",
"0.4878",
"0.5654",
"0.4878",
"0.58",
"0.1184",
"0.58"
],
[
"0.5146",
"0.5635",
"0.881",
"0.5635",
"0.881",
"0.5767",
"0.5146",
"0.5767"
],
[
"0.1184",
"0.582",
"0.378",
"0.582",
"0.378",
"0.5957",
"0.1184",
"0.5957"
],
[
"0.513",
"0.5776",
"0.882",
"0.579",
"0.882",
"0.5933",
"0.513",
"0.593"
],
[
"0.1486",
"0.5986",
"0.4844",
"0.5986",
"0.4844",
"0.6123",
"0.1486",
"0.6123"
],
[
"0.5156",
"0.5957",
"0.882",
"0.5957",
"0.882",
"0.6094",
"0.5156",
"0.6094"
],
[
"0.11584",
"0.6133",
"0.486",
"0.6143",
"0.486",
"0.629",
"0.11584",
"0.6274"
],
[
"0.513",
"0.611",
"0.881",
"0.611",
"0.881",
"0.6245",
"0.513",
"0.6245"
],
[
"0.12006",
"0.6313",
"0.486",
"0.6313",
"0.486",
"0.6445",
"0.12006",
"0.6445"
],
[
"0.511",
"0.627",
"0.882",
"0.6274",
"0.882",
"0.641",
"0.511",
"0.64"
],
[
"0.1184",
"0.6455",
"0.486",
"0.6465",
"0.486",
"0.661",
"0.1184",
"0.66"
],
[
"0.511",
"0.6426",
"0.881",
"0.6426",
"0.881",
"0.6562",
"0.511",
"0.6562"
],
[
"0.1184",
"0.663",
"0.486",
"0.663",
"0.486",
"0.677",
"0.1184",
"0.677"
],
[
"0.513",
"0.66",
"0.881",
"0.66",
"0.881",
"0.6733",
"0.513",
"0.6733"
],
[
"0.1184",
"0.6797",
"0.486",
"0.6797",
"0.486",
"0.6934",
"0.1184",
"0.6934"
],
[
"0.513",
"0.677",
"0.86",
"0.677",
"0.86",
"0.6904",
"0.513",
"0.6904"
],
[
"0.11584",
"0.695",
"0.486",
"0.695",
"0.486",
"0.708",
"0.11584",
"0.708"
],
[
"0.532",
"0.6934",
"0.879",
"0.6934",
"0.879",
"0.707",
"0.532",
"0.707"
],
[
"0.11755",
"0.7104",
"0.4844",
"0.7114",
"0.4844",
"0.7256",
"0.11755",
"0.7246"
],
[
"0.511",
"0.709",
"0.881",
"0.709",
"0.881",
"0.7227",
"0.511",
"0.7227"
],
[
"0.1184",
"0.728",
"0.486",
"0.728",
"0.486",
"0.741",
"0.1184",
"0.741"
],
[
"0.511",
"0.7236",
"0.879",
"0.7236",
"0.879",
"0.737",
"0.511",
"0.737"
],
[
"0.1184",
"0.742",
"0.4844",
"0.7437",
"0.4844",
"0.7583",
"0.1184",
"0.7573"
],
[
"0.513",
"0.741",
"0.881",
"0.741",
"0.881",
"0.755",
"0.513",
"0.755"
],
[
"0.1184",
"0.7603",
"0.486",
"0.7603",
"0.486",
"0.774",
"0.1184",
"0.774"
],
[
"0.513",
"0.756",
"0.879",
"0.756",
"0.879",
"0.7695",
"0.513",
"0.7695"
],
[
"0.12006",
"0.7754",
"0.4878",
"0.7754",
"0.4878",
"0.7905",
"0.12006",
"0.7905"
],
[
"0.511",
"0.7715",
"0.879",
"0.7715",
"0.879",
"0.785",
"0.511",
"0.785"
],
[
"0.1184",
"0.792",
"0.486",
"0.792",
"0.486",
"0.8057",
"0.1184",
"0.8057"
],
[
"0.511",
"0.788",
"0.879",
"0.788",
"0.879",
"0.802",
"0.511",
"0.802"
],
[
"0.12006",
"0.808",
"0.486",
"0.808",
"0.486",
"0.822",
"0.12006",
"0.822"
],
[
"0.513",
"0.804",
"0.882",
"0.8047",
"0.882",
"0.8193",
"0.513",
"0.8184"
],
[
"0.1184",
"0.8247",
"0.4844",
"0.8247",
"0.4844",
"0.8384",
"0.1184",
"0.8384"
],
[
"0.511",
"0.821",
"0.771",
"0.822",
"0.771",
"0.836",
"0.511",
"0.835"
],
[
"0.1184",
"0.839",
"0.4844",
"0.8403",
"0.4844",
"0.855",
"0.1184",
"0.854"
],
[
"0.513",
"0.851",
"0.875",
"0.851",
"0.875",
"0.8643",
"0.513",
"0.8643"
],
[
"0.11755",
"0.857",
"0.4836",
"0.857",
"0.4836",
"0.8706",
"0.11755",
"0.8706"
],
[
"0.11584",
"0.8726",
"0.486",
"0.8726",
"0.486",
"0.8857",
"0.11584",
"0.8857"
],
[
"0.513",
"0.8726",
"0.881",
"0.8726",
"0.881",
"0.8857",
"0.513",
"0.8857"
],
[
"0.1184",
"0.8896",
"0.4878",
"0.8896",
"0.4878",
"0.903",
"0.1184",
"0.903"
],
[
"0.511",
"0.888",
"0.879",
"0.888",
"0.879",
"0.903",
"0.511",
"0.903"
],
[
"0.11755",
"0.905",
"0.4878",
"0.905",
"0.4878",
"0.9185",
"0.11755",
"0.9185"
],
[
"0.513",
"0.905",
"0.881",
"0.905",
"0.881",
"0.9185",
"0.513",
"0.9185"
]
] | [
"Data Instances",
"Averageexplanation",
"Task Format",
"Dataset",
"Task",
"Train",
"Valid",
"Test",
"Length (token)",
"CoS-E v1.0",
"CommonsenseQA",
"3-choice Multiple-Choice",
"7610",
"950",
"16.148",
"CoS-E v1.11",
"Commonsense QA",
"5-choice Multiple-Choice",
"9741",
"1221",
"8.996",
"Commonsense QA",
"5-choice Multiple-Choice",
"1098",
"2194",
"63.572",
"7598",
"ECQA",
"549367",
"9842",
"9824",
"Natural LanguageInference",
"e-SNLI",
"3-labelClassification",
"15.977",
"ComVE",
"Commonsense Validation",
"2-choice Multiple-Choice",
"10000",
"1000",
"1000",
"10.288",
"Table l: Task description and core statistics for five popular large-scale datasets with human-annotated natura",
"languageexplanations thatareincluded inourevaluation.",
"??？。通过对五个数据集和两个模型的评估，",
"中的人工注释解释",
"我们的指标可以在两个模型架构上对所有",
"值得一提的是，我们不包括SBIC（Sapetal.",
"五个数据集的解释质量进行一致的排名，而",
"2019)或E--NLI(Brahmanetal.,2021)等数据",
"Simulatability得分（基线）不足。",
"集。SBIC不提供所有数据的解释，E--NLI利",
"？？？。我们的评估证明了这样的假设，即人类的",
"用各种来源来增加-NLI(Rudingeretal.2020)",
"解释仍然可以使模型预测受益，即使它们被先",
"数据集的解释，而不是提供人工注释",
"前文献的人类评估批评为低质量。",
"2.2Evaluation Metric for Explanations",
"21",
"Related Work",
"许多常用的基于文本的内容评估指标，如",
"BLEU (Papineni et al., 2002)和 ROUGE (Lin",
"2.1 Natural Language Explanation Datasets",
"2004），都将人工注释的答案视为绝对的黄",
"尽管开发了新的模型体系结构和可能更重要的",
"金标准，而无需质疑或尝试评估其质量。一个",
"参数，但这些“黑匣子”不可避免地缺乏解释其",
"已建立的评估指标称为Simulatability分",
"预测的能力；这导致社区越来越多地利用人工",
"数，该指标源自人类可模拟性(Doshi-Velez",
"注释的解释来训练带有解释的模型，或者教他",
"andKim，2017），可以检查黄金解释。它只",
"们自我合理化。例如，WiegreffeandMarasovic",
"是测量基线模型预测性能的变化，具体取",
"(2021）审查了65个数据集，并提供了3类分",
"决于是否将解释作为输入提供。以前的作品",
"类的解释：突出显示，自由文本和结构化。我",
"(Chandrasekaran et al., 2018; Yeung et al., 2020;",
"们专注于五个大型公共数据集，这些数据集在",
"Hase et al., 2020; Wiegreffe et al., 2020; Poursabzi-",
"实例级别具有自由文本的人工注释解释（表E",
"Sangdeh et al., 2021; Rajagopal et al., 2021) 已经",
"1）。我们仔细检查了这些数据集的许可证，发",
"证明了Simulatability分数对评估解释质量",
"现不存在任何个人身份信息(PI)",
"的有用性。然而，这个指标有几个固有的缺点。",
"个突出的数据集是CoS-E及其两个变体",
"首先，它仅考虑了基线模型上解释的帮助性，",
"CoS-Evl。0和 CoS-Evl。11 (Rajani et al., 2019)",
"在这里我们通过E4节中的实验表明，解释在",
"。它扩展了常识问答（CQAvl。0和vl。11个版",
"微调和推理过程中提供了不同的帮助性。此",
"本）数据集（Talmoretal.,2018），方法是在正确",
"外，当我们将原始任务转换为其他任务时，模",
"答案标签上添加人工注释的解释。但是，最近",
"型性能也可能会有所不同，例如将分类任务转",
"的一些工作表明CoS-E的解释质量不佳，因为",
"换为具有不同输入数据格式的多项选择任务",
"Narang etal.(2020)独立地手工标记了CoS-E",
"为了客观地评估人工注释的解释，我们基于",
"的一些新解释，并发现其原始解释与新解释之",
"Simulatability分数定义了一个新的评估指",
"间的BLEU得分非常低。为了提高解释的质",
"标，该指标通过考虑解释在微调和推理方面的",
"量，ECQA（Aggarwaletal.，2021）收集了每个",
"帮助来补充Simulatability的两个缺点，并",
"候选答案的单句解释，并将其汇总为CQAv1",
"借助统一的结构来最大程度地减少任务差异的",
"中每个数据的自然语言解释。11数据集。Sun",
"影响。其他作品E(Cartonetal.2020)试图对",
"etal.(2022）证明了CoS-E解释不如基于人类",
"解释的不同特征进行评估和分类，但是其中许",
"偏好的ECQA解释好。第四个数据集是e-SNLI",
"多E(Chan et al.,2022a;DeYoung et al.,2020)仍",
"（Camburuetal.,2018），其中包括对斯坦福自然",
"然将人类注释的解释视为金标准",
"语言（SNLI）数据集（Bowmanetal.，2015）的",
"2.3Usage of Explanations for SOTA models",
"解释。最后，第五个数据集是ComVE（Wang",
"etal.,2020），询问两个句子中的哪一个违反常",
"现有工作一直在探索解释可以提高模型性能",
"识。稍后，我们使用我们的指标和已建立的基",
"的情况；例如，Haseand Bansal（2021）认为解",
"线Simulatability分数评估上述五个数据集",
"释最适合用作预测的模型输人，而Kumarand"
] | {
"arxiv_id": "2305.03117",
"language": "zh",
"page_idx": 2
} |
||
2305.03117-zh-3 | 2305.03117-zh-3.jpg | INPUT
OUTPUT
<table>
<tbody>
<tr>
<th rowspan="2">Baseline Infusion</th>
<th>explain: I Question Content choice1: 1 Choice 1 choice2: Choice2 distin choice3: Choice3</th>
<th rowspan="2">Answer Answer</th>
</tr>
<tr>
<th>explain: Question Content choice1: Choice1 choice2: Choice2 choice3: Choice3 <sep> because Explanation</th>
</tr>
<tr>
<td>e-SNLI</td>
<td>and An adult is walking away, empty-handed.? choice1: entailment choice2: explain: what is the relationship between An adult dressed in black holds a stick. neutral choice3: contradiction <sep> because Holds a stick implies using hands so $it$ is not empty-handed.</td>
<td>contradiction</td>
</tr>
<tr>
<td>CoS- E</td>
<td>aquarium choice2: creek choice3: pet store choice4: playground choice5: pond explain: A colorful pebble falls on a goldfish, where is the pebble? choice1: <sep> because goldfish are common fish pets.</td>
<td>aquarium</td>
</tr>
</tbody>
</table>
me choice2: i sting a mosquito <sep> because a human has not stings
Figure 1: Unified structure of Baseline and Infusion settings. Black bold text are fixed prompts. We provide examples of Infusion format in classification task like e-SNLI and multiple choice task like CoS-E and ComVE The color schema follows: blue denotes question content; green denotes choice content; orange denotes explanations.
<table>
<tbody>
<tr>
<th rowspan="2">Explanations ias Input vs sOutput</th>
<th colspan="3">Fine-tune Setting</th>
</tr>
<tr>
<th>Baseline</th>
<th>Self-rationalization</th>
<th>Infusiom</th>
</tr>
<tr>
<td>CoS- E v1. 0</td>
<td>0.695</td>
<td>0.646</td>
<td>0.878</td>
</tr>
<tr>
<td>$\mathbf{ECQA}$</td>
<td>0.572</td>
<td>0.513</td>
<td>0.989</td>
</tr>
</tbody>
</table>
Table 2: Preliminary experiment results of using expla-
nations as part of Input( Infusion ) vs. Output(Selfrationalization) vs. without explanations( Baseline ) on CoS-E and ECQA datasets.
个指标，客观地评估解释对模型性能的效用此外，我们期望这样的度量可以系统地证明解释的好坏；例如，它可以客观地衡量人类研究中“嘈杂”的含义(例如。g。,来自之前关于CoS-E 的工作)
利用像 T5 这样的序列到序列模型的优势， 可以将不同类型的语言任务映射到生成任务中，我们可以控制和最小化不同任务格式对模型性能的影响，同时通过利用统一的数据格式来评估解释的有用性。我们意识到，具有人工注释解释的现有数据集主要是多项选择任务或分类任务。分类任务可以被视为多项选择任务，其中标签确实是选择。受之前的几部作品的启发，这些作品操纵了序列到序列模型(Marasović et al., 2021; Liu et al., 2021) 的提示， 我们将一些定义明确的单词作为统一数据结构的基于模板的提示，以指示任务内容和相应的解释
图 1 中所示的示例说明了我们如何将各种任务映射到统一的多项选择生成任务中。我们提出了两个设置：无解释(Baseline)和解释作为附加输入(Infusion)。在这里，我们解释每个提示如何处理数据内容的不同部分： 1)'explain:'后跟问题内容，2)'choice- $n:$ 后跟每个候选答案，以及 3) 特殊标记’<sep> 将解释与任务内容分开，而 Infusion 中的解释由’because’引导，以便模型知道解释文本解释了任务内容。对于 CoS-E 和 ECQA 等数据集，我们利用原始任务作为问题内容。另一方面，我们为 e-SNLI 定义了固定的问题提示：“[前提]和[假设]之间的关系是什么？”,
Talukdar (2020) 提出了一种系统，专门为 NLI 任务生成特定于标签的解释。最近的一些工作试图使用自合理化设置 (Wiegreffe et al., 2020; Marasović et al., 2021) 生成更好的解释，其中要求模型同时生成预测标签和解释。我们进行了初步实验，以找到最佳的模型设置，以利用$\mathbb{E}4.1$节中的解释
最近有许多作品 (Paranjape et al., 2021; Liu et al., 2021; Chen et al., 2022) 探索使用提示来完成解释，为原始任务生成附加信息，或检查生成的解释是否可以为对抗性攻击提供鲁棒性。Ye and Durrett (2022) 表明，简单地将解释插人到提示中并不总是能提高上下文中的学习性能，并且模型生成的解释对于少量学习可能是不可靠的。另一个相关的研究重点是使用统一的框架 (Chan et al., 2022b) 或生成推理链(Dalvi et al., 2022) 的可教推理系统来提取或生成解释
# 3 Unified Structure
虽然像 BLEU 和 ROUGE 这样的流行指标可以评估文本的连贯性和相似性，但解释的一个关键方面是它们的益处。因此，我们希望开发一 | [
[
"0.44",
"0.0861",
"0.509",
"0.0861",
"0.509",
"0.10095",
"0.44",
"0.10095"
],
[
"0.756",
"0.0861",
"0.848",
"0.0861",
"0.848",
"0.10095",
"0.756",
"0.10095"
],
[
"0.1335",
"0.114",
"0.711",
"0.114",
"0.711",
"0.1277",
"0.1335",
"0.1277"
],
[
"0.784",
"0.1164",
"0.831",
"0.1164",
"0.831",
"0.1277",
"0.784",
"0.1277"
],
[
"0.2124",
"0.1437",
"0.711",
"0.1426",
"0.711",
"0.1561",
"0.2124",
"0.1567"
],
[
"0.1318",
"0.152",
"0.2091",
"0.152",
"0.2091",
"0.1621",
"0.1318",
"0.1621"
],
[
"0.785",
"0.152",
"0.831",
"0.152",
"0.831",
"0.1633",
"0.785",
"0.1633"
],
[
"0.2167",
"0.158",
"0.382",
"0.158",
"0.382",
"0.165",
"0.2167",
"0.165"
],
[
"0.215",
"0.1936",
"0.708",
"0.1936",
"0.708",
"0.2037",
"0.215",
"0.2037"
],
[
"0.1453",
"0.209",
"0.2048",
"0.209",
"0.2048",
"0.2203",
"0.1453",
"0.2203"
],
[
"0.2141",
"0.2037",
"0.6724",
"0.2031",
"0.6724",
"0.2144",
"0.2141",
"0.2156"
],
[
"0.2141",
"0.2144",
"0.693",
"0.2156",
"0.693",
"0.2268",
"0.2141",
"0.2257"
],
[
"0.769",
"0.2114",
"0.846",
"0.2114",
"0.846",
"0.2214",
"0.769",
"0.2214"
],
[
"0.2141",
"0.2268",
"0.346",
"0.2268",
"0.346",
"0.2369",
"0.2141",
"0.2369"
],
[
"0.215",
"0.2559",
"0.6616",
"0.2559",
"0.6616",
"0.2666",
"0.215",
"0.2666"
],
[
"0.1511",
"0.2666",
"0.6885",
"0.2642",
"0.6885",
"0.279",
"0.1511",
"0.2808"
],
[
"0.7793",
"0.2708",
"0.837",
"0.2708",
"0.837",
"0.2808",
"0.7793",
"0.2808"
],
[
"0.2167",
"0.279",
"0.372",
"0.279",
"0.372",
"0.2893",
"0.2167",
"0.2893"
],
[
"0.3972",
"0.2808",
"0.4878",
"0.2808",
"0.4878",
"0.288",
"0.3972",
"0.288"
],
[
"0.1377",
"0.308",
"0.708",
"0.3027",
"0.708",
"0.3247",
"0.1377",
"0.33"
],
[
"0.22",
"0.3206",
"0.642",
"0.3206",
"0.642",
"0.3342",
"0.22",
"0.3342"
],
[
"0.7524",
"0.3184",
"0.8613",
"0.3184",
"0.8613",
"0.3289",
"0.7524",
"0.3289"
],
[
"0.11755",
"0.3433",
"0.881",
"0.3433",
"0.881",
"0.357",
"0.11755",
"0.357"
],
[
"0.1184",
"0.3594",
"0.879",
"0.3594",
"0.879",
"0.3694",
"0.1184",
"0.3694"
],
[
"0.11755",
"0.3718",
"0.8833",
"0.3718",
"0.8833",
"0.386",
"0.11755",
"0.386"
],
[
"0.11755",
"0.3855",
"0.1561",
"0.3872",
"0.1545",
"0.3984",
"0.11584",
"0.3967"
],
[
"0.1427",
"0.412",
"0.2217",
"0.4133",
"0.2217",
"0.4246",
"0.1427",
"0.4233"
],
[
"0.3074",
"0.4133",
"0.3938",
"0.4133",
"0.3938",
"0.4246",
"0.3074",
"0.4246"
],
[
"0.5146",
"0.412",
"0.881",
"0.4133",
"0.881",
"0.4275",
"0.5146",
"0.427"
],
[
"0.1427",
"0.4216",
"0.2217",
"0.4233",
"0.2217",
"0.434",
"0.1427",
"0.4316"
],
[
"0.2301",
"0.4275",
"0.2957",
"0.4275",
"0.2957",
"0.4382",
"0.2301",
"0.4382"
],
[
"0.3005",
"0.4275",
"0.3997",
"0.4275",
"0.3997",
"0.4382",
"0.3005",
"0.4382"
],
[
"0.4072",
"0.427",
"0.4668",
"0.427",
"0.4668",
"0.4382",
"0.4072",
"0.4382"
],
[
"0.5146",
"0.43",
"0.881",
"0.43",
"0.881",
"0.4436",
"0.5146",
"0.4436"
],
[
"0.1411",
"0.4407",
"0.2007",
"0.4407",
"0.2007",
"0.452",
"0.1411",
"0.452"
],
[
"0.2468",
"0.4407",
"0.2805",
"0.4407",
"0.2805",
"0.452",
"0.2468",
"0.452"
],
[
"0.3342",
"0.4407",
"0.367",
"0.4407",
"0.367",
"0.452",
"0.3342",
"0.452"
],
[
"0.419",
"0.4407",
"0.455",
"0.4407",
"0.455",
"0.452",
"0.419",
"0.452"
],
[
"0.2468",
"0.4495",
"0.2805",
"0.4495",
"0.2805",
"0.4614",
"0.2468",
"0.4614"
],
[
"0.3342",
"0.4495",
"0.3652",
"0.4495",
"0.3652",
"0.4614",
"0.3342",
"0.4614"
],
[
"0.4207",
"0.4495",
"0.455",
"0.4495",
"0.455",
"0.4614",
"0.4207",
"0.4614"
],
[
"0.513",
"0.4453",
"0.879",
"0.4453",
"0.879",
"0.459",
"0.513",
"0.459"
],
[
"0.513",
"0.459",
"0.879",
"0.462",
"0.879",
"0.4768",
"0.513",
"0.4739"
],
[
"0.1394",
"0.4507",
"0.1763",
"0.4507",
"0.1763",
"0.462",
"0.1394",
"0.462"
],
[
"0.11755",
"0.4697",
"0.4878",
"0.4697",
"0.4878",
"0.4827",
"0.11755",
"0.4827"
],
[
"0.51",
"0.4756",
"0.6304",
"0.4768",
"0.6304",
"0.4912",
"0.51",
"0.4905"
],
[
"0.1142",
"0.4827",
"0.4878",
"0.4822",
"0.4878",
"0.4963",
"0.1142",
"0.4976"
],
[
"0.11755",
"0.4995",
"0.486",
"0.4995",
"0.486",
"0.5103",
"0.11755",
"0.5103"
],
[
"0.531",
"0.4988",
"0.876",
"0.4988",
"0.876",
"0.5127",
"0.531",
"0.5127"
],
[
"0.1184",
"0.514",
"0.3225",
"0.514",
"0.3225",
"0.525",
"0.1184",
"0.525"
],
[
"0.5156",
"0.514",
"0.879",
"0.514",
"0.879",
"0.528",
"0.5156",
"0.528"
],
[
"0.5146",
"0.531",
"0.881",
"0.531",
"0.881",
"0.5444",
"0.5146",
"0.5444"
],
[
"0.11755",
"0.5415",
"0.486",
"0.5415",
"0.486",
"0.5547",
"0.11755",
"0.5547"
],
[
"0.513",
"0.5464",
"0.881",
"0.5464",
"0.881",
"0.56",
"0.513",
"0.56"
],
[
"0.1184",
"0.557",
"0.486",
"0.557",
"0.486",
"0.571",
"0.1184",
"0.571"
],
[
"0.5146",
"0.564",
"0.8774",
"0.564",
"0.8774",
"0.5747",
"0.5146",
"0.5747"
],
[
"0.11755",
"0.5723",
"0.4878",
"0.572",
"0.4878",
"0.586",
"0.11755",
"0.5874"
],
[
"0.5146",
"0.5767",
"0.879",
"0.5767",
"0.879",
"0.5903",
"0.5146",
"0.5903"
],
[
"0.11584",
"0.5884",
"0.4836",
"0.5884",
"0.4836",
"0.6016",
"0.11584",
"0.6016"
],
[
"0.5146",
"0.594",
"0.881",
"0.594",
"0.881",
"0.608",
"0.5146",
"0.608"
],
[
"0.1184",
"0.6055",
"0.4878",
"0.6055",
"0.4878",
"0.619",
"0.1184",
"0.619"
],
[
"0.511",
"0.6094",
"0.879",
"0.6094",
"0.879",
"0.6226",
"0.511",
"0.6226"
],
[
"0.11755",
"0.6216",
"0.4878",
"0.6206",
"0.4878",
"0.6353",
"0.11755",
"0.6357"
],
[
"0.5146",
"0.627",
"0.881",
"0.627",
"0.881",
"0.64",
"0.5146",
"0.64"
],
[
"0.1184",
"0.638",
"0.2761",
"0.638",
"0.2761",
"0.652",
"0.1184",
"0.652"
],
[
"0.511",
"0.641",
"0.881",
"0.6426",
"0.881",
"0.6567",
"0.511",
"0.6562"
],
[
"0.1351",
"0.655",
"0.486",
"0.655",
"0.486",
"0.6685",
"0.1351",
"0.6685"
],
[
"0.513",
"0.659",
"0.881",
"0.659",
"0.881",
"0.673",
"0.513",
"0.673"
],
[
"0.11584",
"0.6694",
"0.4836",
"0.6694",
"0.4836",
"0.683",
"0.11584",
"0.683"
],
[
"0.513",
"0.676",
"0.881",
"0.676",
"0.881",
"0.6895",
"0.513",
"0.6895"
],
[
"0.1184",
"0.686",
"0.486",
"0.686",
"0.486",
"0.6997",
"0.1184",
"0.6997"
],
[
"0.513",
"0.6924",
"0.55",
"0.6924",
"0.55",
"0.7036",
"0.513",
"0.7036"
],
[
"0.11755",
"0.7007",
"0.4844",
"0.703",
"0.4844",
"0.7163",
"0.11755",
"0.7144"
],
[
"0.531",
"0.7114",
"0.879",
"0.7114",
"0.879",
"0.7246",
"0.531",
"0.7246"
],
[
"0.11755",
"0.7173",
"0.4844",
"0.7173",
"0.4844",
"0.731",
"0.11755",
"0.731"
],
[
"0.5146",
"0.728",
"0.881",
"0.728",
"0.881",
"0.741",
"0.5146",
"0.741"
],
[
"0.1184",
"0.7354",
"0.4844",
"0.7354",
"0.4844",
"0.749",
"0.1184",
"0.749"
],
[
"0.513",
"0.7437",
"0.881",
"0.7437",
"0.881",
"0.7573",
"0.513",
"0.7573"
],
[
"0.11755",
"0.75",
"0.4836",
"0.75",
"0.4836",
"0.763",
"0.11755",
"0.763"
],
[
"0.5146",
"0.7603",
"0.881",
"0.7603",
"0.881",
"0.774",
"0.5146",
"0.774"
],
[
"0.1184",
"0.767",
"0.4844",
"0.767",
"0.4844",
"0.781",
"0.1184",
"0.781"
],
[
"0.513",
"0.7754",
"0.8833",
"0.7754",
"0.8833",
"0.789",
"0.513",
"0.789"
],
[
"0.11755",
"0.783",
"0.486",
"0.782",
"0.486",
"0.7964",
"0.11755",
"0.7974"
],
[
"0.511",
"0.7915",
"0.875",
"0.7905",
"0.875",
"0.8047",
"0.511",
"0.8057"
],
[
"0.11755",
"0.8",
"0.4844",
"0.8",
"0.4844",
"0.8135",
"0.11755",
"0.8135"
],
[
"0.513",
"0.808",
"0.879",
"0.808",
"0.879",
"0.822",
"0.513",
"0.822"
],
[
"0.1184",
"0.8154",
"0.1763",
"0.8154",
"0.1763",
"0.83",
"0.1184",
"0.83"
],
[
"0.5146",
"0.8247",
"0.881",
"0.8247",
"0.881",
"0.8384",
"0.5146",
"0.8384"
],
[
"0.11755",
"0.8457",
"0.1511",
"0.8457",
"0.1511",
"0.857",
"0.11755",
"0.857"
],
[
"0.1436",
"0.8457",
"0.3047",
"0.8457",
"0.3047",
"0.8594",
"0.1436",
"0.8594"
],
[
"0.513",
"0.8403",
"0.881",
"0.8403",
"0.881",
"0.854",
"0.513",
"0.854"
],
[
"0.511",
"0.8555",
"0.881",
"0.855",
"0.881",
"0.869",
"0.511",
"0.8706"
],
[
"0.1184",
"0.8726",
"0.486",
"0.8726",
"0.486",
"0.8857",
"0.1184",
"0.8857"
],
[
"0.511",
"0.8726",
"0.879",
"0.8726",
"0.879",
"0.8857",
"0.511",
"0.8857"
],
[
"0.11755",
"0.8896",
"0.4836",
"0.8896",
"0.4836",
"0.903",
"0.11755",
"0.903"
],
[
"0.5283",
"0.888",
"0.879",
"0.888",
"0.879",
"0.903",
"0.5283",
"0.903"
],
[
"0.1184",
"0.905",
"0.471",
"0.905",
"0.471",
"0.9185",
"0.1184",
"0.9185"
],
[
"0.4685",
"0.909",
"0.482",
"0.909",
"0.482",
"0.913",
"0.4685",
"0.913"
],
[
"0.511",
"0.905",
"0.8623",
"0.905",
"0.8623",
"0.9185",
"0.511",
"0.9185"
]
] | [
"INPUT",
"OUTPUT",
"Baseline explain:Question Content choice1:Choice1 choice2:Choice2choice3:Choice3",
"Answer",
"explain:QuestionContentchoice1:Choice1choice2:Choice2choice3:Choice3",
"Infusion",
"Answer",
"<sep>becauseExplanatior",
"explain:whatistherelationshipbetweenAnadultdressedinblackholdsastick.",
"e-SNLI",
"andAnadult iswalking awayempty-handed.?choice1:entailmentchoice2:",
"neutralchoice3:contradiction<sep>becauseHoldsastick implies.usinghands",
"contradiction",
"itisnotempty-handea",
"explain:Acolorfulpebblefallsonagoldfish,whereisthepebble?choice1",
"CoS-E aquarium choice2:creekchoice3:petstore choice4:playgroundchoice5:pon",
"aquarium",
"<sep>becausegoldfisha",
"ommonfishpets",
"ComVEexplain:which sentenceisagainstcommonsense? choice1:amosquito stings",
"ne choice2:i sting a mosquito<sep>because a human has not stings",
"istingamosquito",
"Figure l: Unified structure of Baseline and Infusion settings. Black bold text are fixed prompts. We provide",
"examplesofInfusionformatinclassificationtasklikee-SNLIandmultiplechoicetasklikeCoS-EandComVE",
"The color schema follows: blue denotes question content; green denotes choice content; orange denotes explana-",
"tions.",
"Explanations as",
"Fine-tune Setting",
"个指标，客观地评估解释对模型性能的效用。",
"InputvsOutput",
"Baseline",
"Self-rationalization",
"Infusion",
"此外，我们期望这样的度量可以系统地证明解",
"CoS-Ev1.0",
"0.695",
"0.646",
"0.878",
"0.572",
"0.513",
"0.989",
"释的好坏；例如，它可以客观地衡量人类研究",
"中“噜杂”的含义（例如。g。，来自之前关于",
"ECQA",
"Table 2: Preliminary experiment results of using expla-",
"CoS-E的工作)",
"nations as part of Input( Infusion ) vs. Output(Self-",
"rationalization)vs.withoutexplanations(Baseline)",
"利用像T5这样的序列到序列模型的优势",
"onCoS-EandECQAdatasets.",
"可以将不同类型的语言任务映射到生成任务",
"中，我们可以控制和最小化不同任务格式对",
"Talukdar（2020）提出了一种系统，专门为NLI",
"模型性能的影响，同时通过利用统一的数据",
"任务生成特定于标签的解释。最近的一些工作",
"格式来评估解释的有用性。我们意识到，具有",
"试图使用自合理化设置（Wiegreffeetal.，2020;",
"人工注释解释的现有数据集主要是多项选择",
"Marasovicetal.2021）生成更好的解释，其中",
"任务或分类任务。分类任务可以被视为多项选",
"要求模型同时生成预测标签和解释。我们进",
"择任务，其中标签确实是选择。受之前的几部",
"行了初步实验，以找到最佳的模型设置，以利",
"作品的启发，这些作品操纵了序列到序列模型",
"用E4.1节中的解释",
"(Marasovic etal.,2021;Liuetal.,2021)的提示，",
"最近有许多作品（Paranjapeetal.，202l；Liu",
"我们将一些定义明确的单词作为统一数据结构",
"etal..2021;Chenetal.,2022)探索使用提示来",
"的基于模板的提示，以指示任务内容和相应的",
"完成解释，为原始任务生成附加信息，或检查",
"解释",
"生成的解释是否可以为对抗性攻击提供鲁棒",
"图1中所示的示例说明了我们如何将各种",
"性。YeandDurrett(2022）表明，简单地将解释",
"任务映射到统一的多项选择生成任务中。我们",
"插人到提示中并不总是能提高上下文中的学",
"提出了两个设置：无解释（Baseline）和解",
"习性能，并且模型生成的解释对于少量学习可",
"释作为附加输人（Infusion）。在这里，我们",
"能是不可靠的。另一个相关的研究重点是使用",
"解释每个提示如何处理数据内容的不同部分：",
"统一的框架（Chanetal.，2022b）或生成推理链",
"1）explain：’后跟问题内容，2）‘choice-n：",
"(Dalvietal.,2022)的可教推理系统来提取或生",
"后跟每个候选答案，以及3）特殊标记<sep>",
"成解释",
"将解释与任务内容分开，而Infusion中的解",
"3",
"Unified Structure",
"释由\"because’引导，以便模型知道解释文本",
"解释了任务内容。对于CoS-E和ECQA等数",
"虽然像BLEU和ROUGE这样的流行指标可以",
"据集，我们利用原始任务作为问题内容。另",
"评估文本的连贯性和相似性，但解释的一个关",
"方面，我们为e-SNLI定义了固定的问题提",
"键方面是它们的益处。因此，我们希望开发",
"",
"示：“[前提】和【假设】之间的关系是什么？"
] | {
"arxiv_id": "2305.03117",
"language": "zh",
"page_idx": 3
} |
||
1811.01054-zh-0 | 1811.01054-zh-0.jpg | # 神经网络距离的极小极大估计
Kaiyi Ji
Department of ECE
The Ohio State University Columbus, OH 43210 [email protected]
Yingbin Liang Department of ECE
The Ohio State University Columbus, OH 43210 [email protected]
# Abstract
An important class of distance metrics proposed for training generative adversarial networks (GANs) is the integral probability metric (IPM), in which the neural net distance captures the practical GAN training via two neural networks. This paper investigates the minimax estimation problem of the neural net distance based on samples drawn from the distributions. We develop the first known minimax lower bound on the estimation error of the neural net distance, and an upper bound tighter than an existing bound on the estimator error for the empirical net distance. Our lower and upper bounds match not only in the order of the sample size but also in terms of the norm of the parameter matrices of neural networks, which justifes the empirical neural net distance as a good approximation of the true neural net distance for training GANs in practice.
## 1 Introduction
由$\mathbb{E}[9]$ 首次引人的生成对抗网络 (GAN) 已成为从复杂的现实生活数据中学习生成模型的重要技术。训练 GAN 是通过最小-最大优化执行的，最大值和最小值分别接管一类鉴别器和一类生成器，其中鉴别器和生成器都由神经网络建模。鉴于鉴别器类足够大，然后[9]将GAN 训练解释为找到一个生成器，以使生成的分布 $\nu$尽可能接近目标真实分布 $\mu$ ,通过Jensen-Shannon 距离$d_JS(\mu,\nu)$进行测量，如下所示：
(1)
$$\min_{\nu\in\mathcal{D}_{\mathcal{G}}}d_{JS}(\mu,\nu).$$
训练基于这种思想提出的大型分布对之间的度量。g。,[2,3,13,15]。其中，积分概率度量
(IPM) [19] 作为训练 GAN 的一类重要的距离度量而产生，其形式如下
(2)
$$d_{\mathcal{F}}(\mu,\nu)=\sup_{f\in\mathcal{F}}|\mathbb{E}_{x\sim\mu}f(x)-\mathbb{E}_{x\sim\nu}f(x)|\:.$$
特别是，(2) 中函数类$\mathcal{F}$的不同选择导致不同的距离度量。例如 ,如果$\mathcal{F}$表示所有 1-Lipschitz 函数的集合，则$d_\mathcal{F}(\mu,\nu)$ 对应于 Wasserstein-GAN(WGAN)[2]中使用的 Wasserstein- $\hat{1}$ 距离。如果$\mathcal{F}$表示再生内核希尔伯特空间 (RKHS)中的单位球，则$d_\mathcal{F}(\mu,\nu)$对应于最大平均差异(MMD)距离，该距离用于 MMD-GANE[7,13]
实用的 GAN 训练自然会激发将$\mathbb{E}($2) 中的 $\mathcal{F}$ 作为神经网络的集合 $\mathcal{F}_nn$, 这导致在$\mathbb{E}[3,28]$ 中引人和研究了所谓的神经网络距离$d_\mathcal{F}_{nn}(\mu,\nu)$。为了计算的可行性，在实践中通常采用$d_{\mathcal{F}_{nn}}(\hat{\mu}_{n},\hat{\nu}_{m})$ 作为近似值(i。e。,一个估计器)的真实神经网络距离$d_\mathcal{F}_{nn}(\mu,\nu)$ 用于实际GAN 训练，其中$\hat{\mu}_n$和$\hat{\nu}_m$分别是对应于$\mu$和$\nu$的经验分布，基于从$\mu$抽取的$n$样本和从$\nu$ 抽取的$\mathcal{F}_nn$样本。因此，这里可以问的一个重要问题是$d_\mathcal{F}_{nn}(\hat{\mu}_n,\hat{\nu}_m)$与$d_\mathcal{F}_{nn}(\mu,\nu)$的近似程度。如果它们很接近，则将 GAN 训练为小$d_{\mathcal{F}_{nn}}(\hat{\mu}_{n},\hat{\nu}_{m})$也意味着小$d_\mathcal{F}_{nn}(\mu,\nu)$ , i。e。, 则保证生成的分布$\nu$接近真实分布$\mu$
为了回答这个问题，$\operatorname{E}$[3]推导了数量$|d_\mathcal{F}_{nn}(\mu,\nu)-d_\mathcal{F}_{nn}(\hat{\mu}_n,\hat{\nu}_m)|$的上限，并表明$d_{\mathcal{F}_{nn}}(\hat{\mu}_{n},\hat{\nu}_{m})$以$\mathcal{O}(n^-1/2+m^{-1/2})$的速率收敛到$d_\mathcal{F}_{nn}(\mu,\nu)$ 。但是，以下两个重要问题To appear in 32nd Conference on Neural Information Processing Svstems (NIPS 2018). Montréal. Canada | [
[
"0.3186",
"0.1288",
"0.6787",
"0.1288",
"0.6787",
"0.1465",
"0.3186",
"0.1465"
],
[
"0.335",
"0.2052",
"0.3896",
"0.2052",
"0.3896",
"0.2166",
"0.335",
"0.2166"
],
[
"0.591",
"0.2052",
"0.689",
"0.2052",
"0.689",
"0.2166",
"0.591",
"0.2166"
],
[
"0.2932",
"0.2172",
"0.4248",
"0.2172",
"0.4248",
"0.2317",
"0.2932",
"0.2317"
],
[
"0.5728",
"0.2172",
"0.705",
"0.2172",
"0.705",
"0.2317",
"0.5728",
"0.2317"
],
[
"0.273",
"0.2299",
"0.4453",
"0.2311",
"0.4453",
"0.2456",
"0.273",
"0.2443"
],
[
"0.55",
"0.2299",
"0.7236",
"0.2311",
"0.7236",
"0.2456",
"0.55",
"0.2443"
],
[
"0.2876",
"0.2456",
"0.433",
"0.2456",
"0.433",
"0.2588",
"0.2876",
"0.2588"
],
[
"0.5664",
"0.2456",
"0.7134",
"0.2456",
"0.7134",
"0.2588",
"0.5664",
"0.2588"
],
[
"0.298",
"0.26",
"0.4207",
"0.26",
"0.4207",
"0.2734",
"0.298",
"0.2734"
],
[
"0.5645",
"0.26",
"0.7134",
"0.26",
"0.7134",
"0.2734",
"0.5645",
"0.2734"
],
[
"0.4617",
"0.308",
"0.5376",
"0.31",
"0.5376",
"0.3257",
"0.4617",
"0.324"
],
[
"0.2345",
"0.339",
"0.763",
"0.339",
"0.763",
"0.3528",
"0.2345",
"0.3528"
],
[
"0.2336",
"0.3528",
"0.7656",
"0.3528",
"0.7656",
"0.3662",
"0.2336",
"0.3662"
],
[
"0.2345",
"0.3674",
"0.763",
"0.3674",
"0.763",
"0.3806",
"0.2345",
"0.3806"
],
[
"0.2336",
"0.3806",
"0.7656",
"0.3806",
"0.7656",
"0.3945",
"0.2336",
"0.3945"
],
[
"0.2361",
"0.3965",
"0.763",
"0.3965",
"0.763",
"0.4072",
"0.2361",
"0.4072"
],
[
"0.2336",
"0.4077",
"0.7646",
"0.4092",
"0.7646",
"0.4204",
"0.2336",
"0.4192"
],
[
"0.2336",
"0.4216",
"0.7656",
"0.4216",
"0.7656",
"0.435",
"0.2336",
"0.435"
],
[
"0.2345",
"0.4375",
"0.7646",
"0.4375",
"0.7646",
"0.4475",
"0.2345",
"0.4475"
],
[
"0.2345",
"0.4495",
"0.763",
"0.4495",
"0.763",
"0.4634",
"0.2345",
"0.4634"
],
[
"0.2345",
"0.4634",
"0.7656",
"0.4634",
"0.7656",
"0.4778",
"0.2345",
"0.4778"
],
[
"0.232",
"0.4766",
"0.5864",
"0.4778",
"0.5864",
"0.4912",
"0.232",
"0.4905"
],
[
"0.1724",
"0.5146",
"0.3162",
"0.515",
"0.3162",
"0.53",
"0.1724",
"0.529"
],
[
"0.179",
"0.547",
"0.8203",
"0.547",
"0.8203",
"0.5566",
"0.179",
"0.5566"
],
[
"0.1765",
"0.5596",
"0.8228",
"0.5596",
"0.8228",
"0.5728",
"0.1765",
"0.5728"
],
[
"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"
],
[
"0.1748",
"0.601",
"0.5474",
"0.601",
"0.5474",
"0.6143",
"0.1748",
"0.6143"
],
[
"0.452",
"0.6206",
"0.5513",
"0.6206",
"0.5513",
"0.632",
"0.452",
"0.632"
],
[
"0.8022",
"0.6206",
"0.826",
"0.6206",
"0.826",
"0.6333",
"0.8022",
"0.6333"
],
[
"0.4478",
"0.631",
"0.4795",
"0.632",
"0.4795",
"0.6396",
"0.4478",
"0.638"
],
[
"0.1748",
"0.646",
"0.822",
"0.646",
"0.822",
"0.659",
"0.1748",
"0.659"
],
[
"0.1732",
"0.659",
"0.6846",
"0.661",
"0.6846",
"0.675",
"0.1732",
"0.6724"
],
[
"0.3586",
"0.6787",
"0.6406",
"0.6787",
"0.6406",
"0.6934",
"0.3586",
"0.6934"
],
[
"0.8037",
"0.6787",
"0.826",
"0.6787",
"0.826",
"0.6934",
"0.8037",
"0.6934"
],
[
"0.437",
"0.6943",
"0.4617",
"0.6943",
"0.4617",
"0.702",
"0.437",
"0.702"
],
[
"0.1748",
"0.7104",
"0.8228",
"0.7104",
"0.8228",
"0.7236",
"0.1748",
"0.7236"
],
[
"0.1765",
"0.7227",
"0.8228",
"0.7227",
"0.8228",
"0.7373",
"0.1765",
"0.7373"
],
[
"0.1765",
"0.7373",
"0.8228",
"0.7373",
"0.8228",
"0.752",
"0.1765",
"0.752"
],
[
"0.1732",
"0.75",
"0.5146",
"0.75",
"0.5146",
"0.763",
"0.1732",
"0.763"
],
[
"0.1765",
"0.7715",
"0.8228",
"0.7715",
"0.8228",
"0.785",
"0.1765",
"0.785"
],
[
"0.1748",
"0.784",
"0.8203",
"0.784",
"0.8203",
"0.7974",
"0.1748",
"0.7974"
],
[
"0.1765",
"0.8",
"0.4084",
"0.8",
"0.4084",
"0.813",
"0.1765",
"0.813"
],
[
"0.433",
"0.799",
"0.8203",
"0.799",
"0.8203",
"0.8125",
"0.433",
"0.8125"
],
[
"0.1748",
"0.813",
"0.822",
"0.813",
"0.822",
"0.827",
"0.1748",
"0.827"
],
[
"0.1732",
"0.827",
"0.822",
"0.827",
"0.822",
"0.8413",
"0.1732",
"0.8413"
],
[
"0.1765",
"0.8413",
"0.824",
"0.8413",
"0.824",
"0.855",
"0.1765",
"0.855"
],
[
"0.1765",
"0.854",
"0.437",
"0.855",
"0.437",
"0.8687",
"0.1765",
"0.867"
],
[
"0.1748",
"0.8735",
"0.8228",
"0.875",
"0.8228",
"0.8896",
"0.1748",
"0.888"
],
[
"0.1724",
"0.888",
"0.824",
"0.887",
"0.824",
"0.905",
"0.1724",
"0.906"
],
[
"0.1748",
"0.925",
"0.8",
"0.925",
"0.8",
"0.938",
"0.1748",
"0.938"
]
] | [
"神经网络距离的极小极大估计",
"KaiyiJi",
"Yingbin Liang",
"Department of ECE",
"Department of ECE",
"The Ohio State University",
"The Ohio State University",
"Columbus, OH 43210",
"Columbus,OH43210",
"[email protected]",
"[email protected]",
"Abstract",
"An important class of distance metrics proposed for training generative adversar",
"ial networks (GANs) is the integral probability metric (IPM), in which the neural",
"net distance captures the practical GAN training via two neural networks.This pa",
"per investigates the minimax estimation problem of the neural net distance based",
"onsamplesdrawnfromthedistributions.Wedevelopthefirstknownminimax",
"lowerbound ontheestimationerror of theneuralnetdistance,andanupperbound",
"tighter than an existing bound on the estimator error for the empirical neural net",
"distance.Ourlowerandupperboundsmatchnotonlyintheorderofthesample",
"size but also in terms of the norm of the parameter matrices of neural networks.",
"which justifies the empirical neural net distance as a good approximation of the",
"true neural net distance for training GANs in practice.",
"1Introduction",
"由E91首次引人的生成对抗网络（GAN）已成为从复杂的现实生活数据中学习生成模型的",
"重要技术。训练GAN是通过最小-最大优化执行的，最大值和最小值分别接管一类鉴别器",
"和一类生成器，其中鉴别器和生成器都由神经网络建模。鉴于鉴别器类足够大，然后[9]将",
"GAN训练解释为找到一个生成器，以使生成的分布v尽可能接近目标真实分布μ，通过",
"Jensen-Shannon距离dJs（μ，v）进行测量，如下所示：",
"min djs(μ,v).",
"(1)",
"VEDC",
"训练基于这种思想提出的大型分布对之间的度量。g。，[2,3,13,151。其中，积分概率度量",
"（IPM）[19]作为训练GAN的一类重要的距离度量而产生，其形式如下",
"dF(μ,v) = sup [Er~μf(α) -Er~vf(α)/ ",
"(2)",
"fEJ",
"特别是，(2）中函数类F的不同选择导致不同的距离度量。例如，如果F表示所有1-Lipschitz",
"函数的集合，则dF(μu,V）对应于Wasserstein-GAN(WGAN）[2]中使用的Wasserstein-1距离。",
"如果F表示再生内核希尔伯特空间（RKHS）中的单位球，则dF（μu,V)对应于最大平均差异",
"（MMD）距离，该距离用于MMD-GANE[7,13]",
"实用的GAN训练自然会激发将E(2)中的F作为神经网络的集合Fnn，这导致在E[3,28]",
"中引人和研究了所谓的神经网络距离dF（u,V）。为了计算的可行性，在实践中通常采用",
"dnm（in，Vm）作为近似值（i。e。",
"个估计器）的真实神经网络距离dFnn（u,v）用于实际",
"GAN训练，其中in和m分别是对应于μ和v的经验分布，基于从μ抽取的n样本和从V",
"抽取的Fnn样本。因此，这里可以问的一个重要问题是dFn（in,Um）与dFnn（μu,v)的近似",
"程度。如果它们很接近，则将GAN训练为小dFn（in，Vm）也意味着小dFn（μ,V），i。e。",
"则保证生成的分布V接近真实分布μ",
"为了回答这个问题，E[3]推导了数量dFnn（u,v）一dFa（n,Dm)I的上限，并表明",
"dFnn(in,Dm）以O(n-1/2+m-1/2）的速率收敛到dFnn（μ,D）。但是，以下两个重要问题",
"To appear in 32nd Conference on Neural Information Processing Systems (NIPS 2018), Montreal, Canada"
] | {
"arxiv_id": "1811.01054",
"language": "zh",
"page_idx": 0
} |
||
1811.01054-zh-1 | 1811.01054-zh-1.jpg | 仍然悬而未决：(a)收敛速度$O(n^{-1/2}+m^{-1/2})$是否是最佳的？我们当然希望得到保证在实践中使用的经验目标$d_{\mathcal{F}_{nn}}(\hat{\mu}_n,\hat{\nu}_m)$ 首先不会落空。(b)[3]中上限对神经网络的依赖性的特征在干神经网络的参数总数、考虑到最近的「,作。该参数可能相当宽松。例如。g。[20,27]。因此，本文的目标是通过开发$d_\mathcal{F}_nn(\mu,\nu)$的minimax估计误差的下限来解决上述问题(a)(有关精确公式，请参见 2.2节),并通过开发比[3]更严格的上限来解决问题(b) 实际上，上述问题可以看作是距离估计问题，i。e。,基于样本 i估计神经网络距离$d_{\mathcal{F}_{nn}}(\mu,\nu)$ 。i。d。分别取自$\mu$和$\nu$。经验距离$d_\mathcal{F}_nn(\hat{\mu}_n,\hat{\nu}_m)$ 作为其$plug-in$ 估计量(i。e。,用它们的经验版本代替真实分布)。我们有兴趣探索这种插件估计器的收敛性的最优性，不仅在样本大小方面，而且在神经网络的参数方面。我们进一步注意到，神经网络距离可以用于各种其他应用程序，例如支持测量机[18]和异常检测[29],因此我们在这里建立的性能保证可以在这些领域中引起人们的兴趣。
1.1 Our Contribution
在本文中，我们研究了神经网络距离$d_{\mathcal{F}_{nn}}(\mu,\nu)$ 的极小极大估计，其中分析的主要挑战在于处理复杂的神经网络函数。本文在经验估计量的收敛速度上建立了比[3]中现有的更严格的上界，并开发了一个下界，该下界不仅在样本量的顺序上而且在范数上都与我们的上界相匹配。神经网络的参数矩阵。我们的具体贡献总结如下：
$\bullet$ In Section 3.1 , we provide the first known lower bound on the minimax estimation error of
$d_{\mathcal{F}_{nn}}(\mu,\nu)$ based on finite samples, which takes the form as $c_l\max\left(n^{-1/2},m^{-1/2}\right)$ where the constant $c_l$ depends only on the parameters of neural networks. Such a lower bound further specializes to $b_l\prod_{i=1}^dM(i)\operatorname*{max}\left(n^{-1/2},m^{-1/2}\right)$ for ReLU networks, where $b_l$ is a constant, $d$ is the depth of neural networks and $M(i)$ can be either the Frobenius norm or $\|\cdot\|_{1,\infty}$ norm constraint of the parameter matrix $W_i$ in layer $i.$ Our proof exploits the Le Cam's method with the technical development of a lower bound on the difference between two neural networks.
$\bullet$ In Section 3.2 ,we develop an upper bound on the estimation error of $d_{\mathcal{F}_{nn}}(\mu,\nu)$ by
$d_{\mathcal{F} _{nn}}( \hat{\mu } _{n}, \hat{\nu } _{m})$ ,which takes the form as $c_u( n^{- 1/ 2}+ m^{- 1/ 2})$ ,where the constant $c_u$ depends only on the parameters of neural networks. Such an upper bound further specializes to $b_{u}\sqrt {d+ h}\prod _{i= 1}^{d}M( i) ( n^{- 1/ 2}+ m^{- 1/ 2})$ for ReLU networks, where $b_u$ is a constant, $h$ is the dimension of the support of $\mu$ and $\nu$, and $\sqrt d+h$ can be replaced by $\sqrt d$ or $\sqrt{d+\log h}$ depending on the distribution class and the norm of the weight matrices. Our proof includes the following two major technical developments presented in Section 3.4 .
A new concentration inequality: In order to develop an upper bound for the unboundedsupport sub-Gaussican class, standard McDiarmid inequality under bounded difference condition is not applicable. We thus fırst generalize a McDiarmid inequality [11] for unbounded functions of scalar sub-Gaussian variables to that of sub-Gaussian vectors , which can be of independent interest for other applications. Such a development requires substantial machineries. We then apply such a concentration inequality to upper-bounding the estimation error of the neural net distance in terms of Rademacher complexity.
Upper bound on Rademacher complexity: Though existing Rademacher complexity bounds [8,22] of neural networks can be used for input data with bounded support, direct applications of those bounds to the unbounded sub-Gaussian input data yield orderlevel loose bounds. Thus, we develop a tighter bound on the Rademacher complexity that exploits the sub-Gaussianity of the input variables. Such a bound is also tighter than the existing same type by [23]. The details of the comparison are provided after Theorem 7
·在$\overline{\mathrm{E}}$3.3 节中，下限和上限的比较表明经验神经网络距离(i。e。,插件估计器)在
$n^{-1/2}+m^{-1/2}$方面实现了最优的极小极大估计率。此外，对于ReLU网络，两个边界在$\prod_i=1^dM(i)$ 方面也匹配，表明 $\prod_i=1^d\|W_i\|_F$ 和 $\prod_{i=1}^d\|W_i\|_{1,\infty}$都是捕获估计精度
的关键量。这样的结果与[20] 中针对训练深度神经网络的泛化误差所做的结果一致。
我们注意到，边界之间仍然有一个差距$\sqrt d$ , 这 需 要 未 来 的 努 力 来 解 决 。
2 | [
[
"0.1732",
"0.0884",
"0.8228",
"0.0903",
"0.8228",
"0.108",
"0.1732",
"0.1061"
],
[
"0.1765",
"0.1067",
"0.8228",
"0.1067",
"0.8228",
"0.1218",
"0.1765",
"0.1218"
],
[
"0.1732",
"0.11743",
"0.814",
"0.1193",
"0.814",
"0.137",
"0.1732",
"0.1351"
],
[
"0.1772",
"0.1338",
"0.8228",
"0.1338",
"0.8228",
"0.1483",
"0.1772",
"0.1483"
],
[
"0.1772",
"0.1477",
"0.8228",
"0.1477",
"0.8228",
"0.161",
"0.1772",
"0.161"
],
[
"0.1765",
"0.1686",
"0.824",
"0.1686",
"0.824",
"0.1818",
"0.1765",
"0.1818"
],
[
"0.179",
"0.1831",
"0.822",
"0.1831",
"0.822",
"0.1963",
"0.179",
"0.1963"
],
[
"0.1748",
"0.1963",
"0.822",
"0.1963",
"0.822",
"0.2102",
"0.1748",
"0.2102"
],
[
"0.1765",
"0.2102",
"0.8228",
"0.2102",
"0.8228",
"0.2235",
"0.1765",
"0.2235"
],
[
"0.1765",
"0.2235",
"0.8228",
"0.2235",
"0.8228",
"0.2374",
"0.1765",
"0.2374"
],
[
"0.1765",
"0.2393",
"0.3938",
"0.2393",
"0.3938",
"0.25",
"0.1765",
"0.25"
],
[
"0.1765",
"0.289",
"0.2141",
"0.289",
"0.2141",
"0.3005",
"0.1765",
"0.3005"
],
[
"0.2115",
"0.2874",
"0.3423",
"0.2886",
"0.3423",
"0.303",
"0.2115",
"0.3018"
],
[
"0.1772",
"0.324",
"0.8203",
"0.324",
"0.8203",
"0.334",
"0.1772",
"0.334"
],
[
"0.1765",
"0.336",
"0.822",
"0.336",
"0.822",
"0.3499",
"0.1765",
"0.3499"
],
[
"0.1772",
"0.3499",
"0.8228",
"0.3499",
"0.8228",
"0.363",
"0.1772",
"0.363"
],
[
"0.1765",
"0.3643",
"0.574",
"0.3643",
"0.574",
"0.3774",
"0.1765",
"0.3774"
],
[
"0.1993",
"0.3865",
"0.824",
"0.3865",
"0.824",
"0.3965",
"0.1993",
"0.3965"
],
[
"0.2197",
"0.3984",
"0.8228",
"0.3984",
"0.8228",
"0.4136",
"0.2197",
"0.4136"
],
[
"0.2197",
"0.4136",
"0.824",
"0.4136",
"0.824",
"0.4268",
"0.2197",
"0.4268"
],
[
"0.2197",
"0.4287",
"0.824",
"0.4287",
"0.824",
"0.443",
"0.2197",
"0.443"
],
[
"0.2197",
"0.4443",
"0.8228",
"0.4443",
"0.8228",
"0.4583",
"0.2197",
"0.4583"
],
[
"0.2197",
"0.459",
"0.824",
"0.4583",
"0.824",
"0.473",
"0.2197",
"0.4734"
],
[
"0.2197",
"0.473",
"0.824",
"0.473",
"0.824",
"0.486",
"0.2197",
"0.486"
],
[
"0.2214",
"0.488",
"0.357",
"0.488",
"0.357",
"0.4988",
"0.2214",
"0.4988"
],
[
"0.1993",
"0.512",
"0.822",
"0.512",
"0.822",
"0.526",
"0.1993",
"0.526"
],
[
"0.2173",
"0.5244",
"0.824",
"0.5234",
"0.824",
"0.5415",
"0.2173",
"0.5425"
],
[
"0.2197",
"0.5405",
"0.8228",
"0.5405",
"0.8228",
"0.5547",
"0.2197",
"0.5547"
],
[
"0.2197",
"0.556",
"0.824",
"0.556",
"0.824",
"0.571",
"0.2197",
"0.571"
],
[
"0.2197",
"0.5747",
"0.824",
"0.5747",
"0.824",
"0.5884",
"0.2197",
"0.5884"
],
[
"0.2197",
"0.5884",
"0.824",
"0.5884",
"0.824",
"0.6016",
"0.2197",
"0.6016"
],
[
"0.2181",
"0.6016",
"0.71",
"0.6016",
"0.71",
"0.6157",
"0.2181",
"0.6157"
],
[
"0.2222",
"0.632",
"0.2385",
"0.632",
"0.2385",
"0.636",
"0.2222",
"0.636"
],
[
"0.2402",
"0.627",
"0.826",
"0.627",
"0.826",
"0.6416",
"0.2402",
"0.6416"
],
[
"0.2418",
"0.6416",
"0.824",
"0.64",
"0.824",
"0.654",
"0.2418",
"0.655"
],
[
"0.2427",
"0.655",
"0.824",
"0.655",
"0.824",
"0.6685",
"0.2427",
"0.6685"
],
[
"0.2427",
"0.6685",
"0.8203",
"0.6685",
"0.8203",
"0.6816",
"0.2427",
"0.6816"
],
[
"0.246",
"0.6846",
"0.8228",
"0.6846",
"0.8228",
"0.6943",
"0.246",
"0.6943"
],
[
"0.2443",
"0.6963",
"0.824",
"0.6963",
"0.824",
"0.7104",
"0.2443",
"0.7104"
],
[
"0.2427",
"0.711",
"0.778",
"0.711",
"0.778",
"0.7246",
"0.2427",
"0.7246"
],
[
"0.2418",
"0.73",
"0.822",
"0.73",
"0.822",
"0.7437",
"0.2418",
"0.7437"
],
[
"0.2418",
"0.741",
"0.8228",
"0.742",
"0.8228",
"0.757",
"0.2418",
"0.756"
],
[
"0.2427",
"0.757",
"0.826",
"0.757",
"0.826",
"0.771",
"0.2427",
"0.771"
],
[
"0.2427",
"0.7715",
"0.824",
"0.7715",
"0.824",
"0.785",
"0.2427",
"0.785"
],
[
"0.2427",
"0.785",
"0.8228",
"0.785",
"0.8228",
"0.799",
"0.2427",
"0.799"
],
[
"0.2427",
"0.799",
"0.826",
"0.799",
"0.826",
"0.8125",
"0.2427",
"0.8125"
],
[
"0.1993",
"0.837",
"0.8228",
"0.837",
"0.8228",
"0.851",
"0.1993",
"0.851"
],
[
"0.2173",
"0.8477",
"0.824",
"0.8477",
"0.824",
"0.8657",
"0.2173",
"0.8657"
],
[
"0.2197",
"0.8657",
"0.8228",
"0.866",
"0.8228",
"0.881",
"0.2197",
"0.88"
],
[
"0.2197",
"0.881",
"0.822",
"0.881",
"0.822",
"0.8945",
"0.2197",
"0.8945"
],
[
"0.2197",
"0.8965",
"0.732",
"0.8965",
"0.732",
"0.91",
"0.2197",
"0.91"
],
[
"0.492",
"0.935",
"0.5073",
"0.935",
"0.5073",
"0.9497",
"0.492",
"0.9497"
]
] | [
"仍然悬而未决：（a）收敛速度O(n-1/2+m-1/2）是否是最佳的？我们当然希望得到保证",
"在实践中使用的经验目标dFnn（iun，Dm）首先不会落空。（b）[3]中上限对神经网络的依赖性",
"的特征在于神经网络的参数总数，考虑到最近的工作，该参数可能相当宽松，例如。8。",
"[20,27]。因此，本文的目标是通过开发dFnn（μ,v)的minimax估计误差的下限来解决上述",
"问题（a）（有关精确公式，请参见2.2节），并通过开发比[3]更严格的上限来解决问题（b）",
"实际上，上述问题可以看作是距离估计问题，i。e。，基于样本i估计神经网络距离dF（u，V）",
"。i。d。分别取自μ和V。经验距离dFnn（in，m）作为其plug-in估计量（i。e。，用它们的",
"经验版本代替真实分布)。我们有兴趣探索这种插件估计器的收敛性的最优性，不仅在样本",
"大小方面，而且在神经网络的参数方面。我们进一步注意到，神经网络距离可以用于各种其",
"他应用程序，例如支持测量机[18]和异常检测[29]，因此我们在这里建立的性能保证可以",
"在这些领域中引起人们的兴趣。",
"1.1",
"OurContribution",
"在本文中，我们研究了神经网络距离d（u，V）的极小极大估计，其中分析的主要挑战在于",
"处理复杂的神经网络函数。本文在经验估计量的收敛速度上建立了比[3]中现有的更严格的",
"上界，并开发了一个下界，该下界不仅在样本量的顺序上而且在范数上都与我们的上界相",
"匹配。神经网络的参数矩阵。我们的具体贡献总结如下：",
".InSection3.l,weprovidethefirstknownlowerboundontheminimaxestimationerrorof",
"dFnn (μ, v) based on finite samples, which takes the form as c max (n-1/2, m-1/2) where",
"the constant ci depends only on the parameters of neural networks. Such a lower bound",
"further specializes to bt II,- M(i) max (n-1/2, m-1/2) for ReLU networks, where br is a",
"constant, d is the depth of neural networks and M(i) can be either the Frobenius norm or",
"Il : Ili,oo norm constraint of the parameter matrix W, in layer i . Our proof exploits the Le",
"Cam's method with the technical development of a lower bound on the difference between",
"two neural networks.",
": In Section 3.2,we develop an upper bound on the estimation error of dFn (μ,) by",
"dfnn (in,Dm), which takes the form as cu(n-1/2 + m-1/2), where the constant cu de",
"pends only on the parameters of neural networks. Such an upper bound further specializes",
"to bu Vd + hII,-, M(i)(n-1/2 + m-1/2) for ReLU networks, where bu is a constant, h is",
"the dimension of the support of μ and , and Vd + h can be replaced by Vd or d + log h",
"depending on the distribution class and the norm of the weight matrices. Our proof includes",
"the following two major technical developments presented in Section 3.4.",
"",
"A new concentration inequality: In order to develop an upper bound for the unbounded-",
"support sub-Gaussican class,standard McDiarmid inequality underbounded difference",
"condition is not applicable. We thus first generalize a McDiarmid inequality [ll] for",
"unbounded functions of scalar sub-Gaussian variables to that of sub-Gaussian vectors",
"whichcanbeofindependentinterestforotherapplications.Suchadevelopmentrequires",
"substantial machineries. We then apply such a concentration inequality to upper-bounding",
"the estimation error of the neural net distance in terms of Rademacher complexity",
"Upper bound on Rademacher complexity:Though existing Rademacher complexity",
"bounds[8, 22] of neural networks can be used for input data with bounded support,",
"direct applications of those bounds to the unbounded sub-Gaussian input data yield order-",
"level loose bounds. Thus, we develop a tighter bound on the Rademacher complexity that",
"exploits the sub-Gaussianity of the input variables. Such a bound is also tighter than the",
"existing same type by [23] : The details of the comparison are provided after Theorem 7",
"。在E3.3节中，下限和上限的比较表明经验神经网络距离（i。e。，插件估计器）在",
"n-1/2+m-1/2方面实现了最优的极小极大估计率。此外，对于ReLU网络，两个迈",
"界在Ⅱd-M(i）方面也匹配，表明I-//W:llF和IWil1,都是捕获估计精度",
"的关键量。这样的结果与[20]中针对训练深度神经网络的泛化误差所做的结果一致。",
"我们注意到，边界之间仍然有一个差距Vd，这需要未来的努力来解决。",
"2"
] | {
"arxiv_id": "1811.01054",
"language": "zh",
"page_idx": 1
} |
||
1811.01054-zh-2 | 1811.01054-zh-2.jpg | ## 1.2 Related Work
Estimation of IPMs. [25] 研究了几个 IPM 的经验估计，包括 Wasserstein 距离，MMD 和 Dudley 度量，并建立了其经验估计器的收敛速度。最近的一篇论文[26]建立了MMD 的经验估计量实现了极小极大最优收敛速度。[3]引入了同样属于 IPM 类的神经网络距离，并建立了其经验估计量的收敛速度。本文为这样的距离度量建立了一个更严格的上限，以及一个与我们的上限在样本量和参数矩阵的范数上相匹配的下限。Generalization error of GANs. 在本文中，我们专注于神经网络距离的极小极大估计误差，因此数量$\left|d_\mathcal{F}_{nn}(\mu,\nu)-d_\mathcal{F}_{nn}(\hat{\mu}_n,\hat{\nu}_m)\right|$ 是我们感兴趣的，我们的边界比[3]中的早期研究更严格。这样的量相关，但与最近在[14,28] 中研究的用于训练 GAN 的以下泛化误差不同。[28]研究了$d_\mathcal{F}(\mu,\hat{\nu}^*)-\inf_{\nu\in\mathcal{D}_G}d_\mathcal{F}(\mu,\nu)$ 的泛化误差，其中$\hat{\nu}^*$是$d_\mathcal{F}(\hat{\mu}_n,\nu)$的最小化器$,\mathcal{F}$被视为神经网络的$\mathcal{F}_nn$类[14]研究了相同类型的泛化误差，但将$\mathcal{F}$作为 Sobolev 空间，并表征了 Sobolev 空间的平滑性如何帮助 GAN 训练。
Rademacher complexity of neural networks. 我们对神经网络距离的极小极大估计误差的分析的一部分要求对神经网络的 average Rademacher 复杂度进行上界。尽管对神经网络的Rademacher 复杂性有不同的限制，但。g。,[1,4,8,21,22],可用于有界支持的分布，直接应用次高斯变量的最著名的现有界比我们在这里建立的界更宽松。[6,23] 研究了高斯变量上一隐层神经网络的平均 Rademacher 复杂度。我们对[23] 的设置的绑定的专门化改进了它的绑定，并且对[6]的设置等于它的绑定。
## 1.3 Notations
我们使用粗体小写字母和大写字母分别表示向量和矩阵。给定一个向量$w\in\mathbb{R}^h,\|\boldsymbol{w}\|_2=$ $\left(\sum_{i=1}^hw_i^2\right)^{1/2}$表示$\ell_2$范数，$\|\boldsymbol w\|_1=\sum_{i=1}^h|\boldsymbol{w}_i|$表示$\ell_1$范数，其中$\boldsymbol w_i$表示$\boldsymbol w$的$i^th$坐标。时于矩阵$W=[W_ij]$ ,我们使用 $\|\boldsymbol{W}\|_F=\left(\sum_{i,j}\boldsymbol{W}_{ij}^2\right)^{1/2}$表示其 Frobenius 范数，$\|\boldsymbol{W}\|_{1,\infty}$ 表示$W$行向量的最大$\ell_1$范数，$\|\boldsymbol{W}\|$表示其频谱范数。对于实分布$\mu$ ,我们将$\hat{\mu}_n$表示为其经验分布，它在每个$n$样本 i 上取$1/n$概率。i。d。取自$\mu$。
2 $\textbf{Preliminaries and Problem Formulations}$
在本节中，我们首先介绍神经网络距离以及相应神经网络的规范。然后，我们介绍了本文研
究的极小极大估计问题。
## 2.1 Neural Net Distance
[3] 中引入的两个分布$\mu$和$\nu$之间的神经网络距离定义为
(3)
$$d_{\mathcal{F}_{nn}}(\mu,\nu)=\sup\limits_{f\in\mathcal{F}_{nn}}|\mathbb{E}_{x\sim\mu}f(x)-\mathbb{E}_{x\sim\nu}f(x)|\:,$$
其中$\mathcal{F}_nn$是一类神经网络。在本文中，给定域$\mathcal{X}\subseteq\mathbb{R}^h$ ,我们让$\mathcal{F}_nn$为以下形式的深度-
$\stackrel{'\text{ 两 }1}{d\text{ 神经网络的集合：}}$
$$f\in\mathcal{F}_{nn}:\boldsymbol{x}\in\mathcal{X}\longmapsto\boldsymbol{w}_{d}^{T}\boldsymbol{\sigma}_{d-1}\left(\boldsymbol{W}_{d-1}\sigma_{d-2}\left(\cdots\sigma_{1}(\boldsymbol{W}_{1}\boldsymbol{x})\right)\right),$$
(4)
其中$\boldsymbol W_i,i=1,2,...,d-1$ 是参数矩阵，$\boldsymbol w_d$ 是参数向量 (因此神经网络的输出表示 $\sigma_i$ 项)。e。,对于输人$z\in \mathbb{R} ^t$, $\sigma _i( z) : = [ \sigma _i( z_1) , \sigma _i( z_2) , . . . , \sigma _i( z_t) ] ^T$。在本文中，我们对(4)中的激活函数采用以下两个假设。
Assumption 1. All activation functions $\sigma _i( \cdot ) \textit{for }i= 1, 2, . . . , d- 1\textit{ satisfy}$
$\bullet$ $\sigma _{i}( \cdot )$ $is$ continuous and $non- decreasing$ and $\sigma _{i}( 0) \geq 0$ .
$\bullet$ $\sigma _{i}( \cdot )$ $is$ $L_{i}$ $- Lipschitz$, where $L_{i}> 0$ .
Assumption 2. For all activation functions $\sigma_i,i=1,2,...,d-1,there$ exist positive constants q(i)
and $Q_{\sigma }( i)$ such that for any $0\leq x_{1}\leq x_{2}\leq q( i)$ , $\sigma _{i}( x_{2}) - \sigma _{i}( x_{1}) \geq Q_{\sigma }( i) ( x_{2}- x_{1})$ .
请注意，假设$\operatorname{E}$1 和 2 适用于各种常用的激活函数，包括 ReLU, sigmoid, softPlus 和 tanh。特别是，在假设$\overline{\mathrm{E}}$2 中，常量$q(i)$和$Q_\sigma(i)$的存在比它们所采用的特定值更重要，后者仅
3 | [
[
"0.1748",
"0.09216",
"0.3145",
"0.09216",
"0.3145",
"0.1061",
"0.1748",
"0.1061"
],
[
"0.1765",
"0.1187",
"0.822",
"0.1187",
"0.822",
"0.132",
"0.1765",
"0.132"
],
[
"0.1772",
"0.132",
"0.8203",
"0.132",
"0.8203",
"0.1459",
"0.1772",
"0.1459"
],
[
"0.1772",
"0.1465",
"0.822",
"0.1465",
"0.822",
"0.1604",
"0.1772",
"0.1604"
],
[
"0.1748",
"0.1578",
"0.822",
"0.1578",
"0.822",
"0.1718",
"0.1748",
"0.1718"
],
[
"0.1765",
"0.1724",
"0.8228",
"0.1724",
"0.8228",
"0.1863",
"0.1765",
"0.1863"
],
[
"0.1772",
"0.1875",
"0.822",
"0.1875",
"0.822",
"0.2008",
"0.1772",
"0.2008"
],
[
"0.1765",
"0.2008",
"0.822",
"0.2008",
"0.822",
"0.214",
"0.1765",
"0.214"
],
[
"0.1765",
"0.2153",
"0.8203",
"0.2153",
"0.8203",
"0.2291",
"0.1765",
"0.2291"
],
[
"0.1765",
"0.2279",
"0.822",
"0.2279",
"0.822",
"0.2412",
"0.1765",
"0.2412"
],
[
"0.1772",
"0.2424",
"0.822",
"0.2424",
"0.822",
"0.2556",
"0.1772",
"0.2556"
],
[
"0.179",
"0.2568",
"0.2133",
"0.2568",
"0.2133",
"0.2683",
"0.179",
"0.2683"
],
[
"0.1765",
"0.2766",
"0.822",
"0.2766",
"0.822",
"0.2903",
"0.1765",
"0.2903"
],
[
"0.1748",
"0.289",
"0.822",
"0.2903",
"0.822",
"0.305",
"0.1748",
"0.3037"
],
[
"0.1748",
"0.3037",
"0.822",
"0.3037",
"0.822",
"0.3176",
"0.1748",
"0.3176"
],
[
"0.1772",
"0.3193",
"0.8203",
"0.3193",
"0.8203",
"0.33",
"0.1772",
"0.33"
],
[
"0.1896",
"0.332",
"0.822",
"0.332",
"0.822",
"0.3452",
"0.1896",
"0.3452"
],
[
"0.1748",
"0.3452",
"0.4558",
"0.3452",
"0.4558",
"0.3591",
"0.1748",
"0.3591"
],
[
"0.1748",
"0.3757",
"0.2852",
"0.3757",
"0.2852",
"0.3901",
"0.1748",
"0.3901"
],
[
"0.1765",
"0.4016",
"0.8228",
"0.4016",
"0.8228",
"0.4153",
"0.1765",
"0.4153"
],
[
"0.179",
"0.4216",
"0.255",
"0.4216",
"0.255",
"0.44",
"0.179",
"0.44"
],
[
"0.2646",
"0.4236",
"0.8228",
"0.4224",
"0.8228",
"0.4375",
"0.2646",
"0.4382"
],
[
"0.1765",
"0.4507",
"0.8203",
"0.4507",
"0.8203",
"0.4653",
"0.1765",
"0.4653"
],
[
"0.569",
"0.4495",
"0.58",
"0.4495",
"0.58",
"0.4526",
"0.569",
"0.4526"
],
[
"0.1765",
"0.4702",
"0.822",
"0.4702",
"0.822",
"0.4841",
"0.1765",
"0.4841"
],
[
"0.1748",
"0.483",
"0.6216",
"0.4849",
"0.6216",
"0.5",
"0.1748",
"0.4976"
],
[
"0.1724",
"0.5176",
"0.5522",
"0.518",
"0.5522",
"0.5327",
"0.1724",
"0.532"
],
[
"0.1765",
"0.549",
"0.822",
"0.549",
"0.822",
"0.5625",
"0.1765",
"0.5625"
],
[
"0.1765",
"0.563",
"0.3464",
"0.563",
"0.3464",
"0.577",
"0.1765",
"0.577"
],
[
"0.1748",
"0.5933",
"0.3586",
"0.5933",
"0.3586",
"0.607",
"0.1748",
"0.607"
],
[
"0.1748",
"0.619",
"0.5767",
"0.619",
"0.5767",
"0.6333",
"0.1748",
"0.6333"
],
[
"0.343",
"0.6416",
"0.6567",
"0.6416",
"0.6567",
"0.655",
"0.343",
"0.655"
],
[
"0.8037",
"0.64",
"0.824",
"0.64",
"0.824",
"0.655",
"0.8037",
"0.655"
],
[
"0.437",
"0.653",
"0.4795",
"0.655",
"0.478",
"0.6665",
"0.4355",
"0.664"
],
[
"0.1896",
"0.674",
"0.8203",
"0.674",
"0.8203",
"0.6875",
"0.1896",
"0.6875"
],
[
"0.1765",
"0.6875",
"0.3137",
"0.6875",
"0.3137",
"0.7007",
"0.1765",
"0.7007"
],
[
"0.2917",
"0.709",
"0.7036",
"0.709",
"0.7036",
"0.7236",
"0.2917",
"0.7236"
],
[
"0.8037",
"0.7104",
"0.824",
"0.7104",
"0.824",
"0.7227",
"0.8037",
"0.7227"
],
[
"0.1765",
"0.733",
"0.8228",
"0.733",
"0.8228",
"0.747",
"0.1765",
"0.747"
],
[
"0.1765",
"0.7456",
"0.8228",
"0.747",
"0.8228",
"0.761",
"0.1765",
"0.7603"
],
[
"0.1765",
"0.7603",
"0.3635",
"0.7603",
"0.3635",
"0.7734",
"0.1765",
"0.7734"
],
[
"0.1765",
"0.779",
"0.6587",
"0.779",
"0.6587",
"0.7935",
"0.1765",
"0.7935"
],
[
"0.2197",
"0.8047",
"0.5947",
"0.8047",
"0.5947",
"0.8184",
"0.2197",
"0.8184"
],
[
"0.2197",
"0.8276",
"0.4763",
"0.8276",
"0.4763",
"0.8413",
"0.2197",
"0.8413"
],
[
"0.1765",
"0.845",
"0.8228",
"0.845",
"0.8228",
"0.86",
"0.1765",
"0.86"
],
[
"0.1765",
"0.86",
"0.749",
"0.86",
"0.749",
"0.8735",
"0.1765",
"0.8735"
],
[
"0.1772",
"0.8853",
"0.8203",
"0.8853",
"0.8203",
"0.8984",
"0.1772",
"0.8984"
],
[
"0.1748",
"0.8984",
"0.822",
"0.8984",
"0.822",
"0.912",
"0.1748",
"0.912"
],
[
"0.4927",
"0.9375",
"0.5063",
"0.9375",
"0.5063",
"0.9487",
"0.4927",
"0.9487"
]
] | [
"1.2Related Work",
"EstimationofIPMs.[25]研究了几个IPM的经验估计，包括Wasserstein距离，MMD和Dudley",
"度量，并建立了其经验估计器的收敛速度。最近的一篇论文[26]建立了MMD的经验估计",
"量实现了极小极大最优收敛速度。[3]引入了同样属于IPM类的神经网络距离，并建立了其",
"经验估计量的收敛速度。本文为这样的距离度量建立了一个更严格的上限，以及一个与我们",
"的上限在样本量和参数矩阵的范数上相匹配的下限。GeneralizationerrorofGANs.在本文",
"中，我们专注于神经网络距离的极小极大估计误差，因此数量dFn（u,v）一dFan（in，Dm)/是",
"我们感兴趣的，我们的边界比[3]中的早期研究更严格。这样的量相关，但与最近在[14,28]",
"中研究的用于训练GAN的以下泛化误差不同。[28]研究了d(μ,*）-infveDgdF(μ,v)的",
"泛化误差，其中*是dF(in，v)的最小化器，F被视为神经网络的Fnn类[14i研究了相同",
"类型的泛化误差，但将F作为Sobolev空间，并表征了Sobolev空间的平滑性如何帮助GAN",
"训练。",
"Rademachercomplexityofneuralnetworks.我们对神经网络距离的极小极大估计误差的分",
"析的一部分要求对神经网络的averageRademacher复杂度进行上界。尽管对神经网络的",
"Rademacher复杂性有不同的限制，但。g。，[1,4,8,2122]，可用于有界支持的分布，直接",
"应用次高斯变量的最者名的现有界比我们在这单建立的界更宽松。[6，23]研究了高斯变量上",
"隐层神经网络的平均Rademacher复杂度。我们对[23]的设置的绑定的专门化改进了它的",
"绑定，并且对[6]的设置等于它的绑定。",
"1.3Notations",
"我们使用粗体小写字母和大写字母分别表示向量和矩阵。给定一个向量wERh，wl2=",
"(hiw)",
"表示l2范数，wll1=Zhwil表示li范数，其中w表示w的ith坐标。",
"对于矩阵W=[W]，我们使用IWllF=（ZiWa）表示其Frobenius范数，IWIl1,α",
"",
"表示W行向量的最大范数，IW表示其频谱范数。对于实分布μ，我们将n表示为",
"其经验分布，它在每个n样本i上取1/n概率。i。d。取自μ。",
"2Preliminaries and Problem Formulations",
"在本节中，我们首先介绍神经网络距离以及相应神经网络的规范。然后，我们介绍了本文研",
"究的极小极大估计问题。",
"2.1 Neural Net Distance",
"[3]中引人的两个分布μ和V之间的神经网络距离定义为",
"dFnn(μ,v)= supEr~μf(α)-Er~vf(α)l,",
"(3)",
"fEFnn",
"其中Fnn是一类神经网络。在本文中，给定域CRh，我们让Fnn为以下形式的深度",
"d神经网络的集合：",
"f EFnn: a EXi→waad-i (Wd-iod-2(...oi(Wia)",
"(4)",
"其中Wi,i=1,2..,d-1是参数矩阵，wd是参数向量（因此神经网络的输出表示o项）。",
"e。，对于输人zEIRt，(z)=[(z1),(22)，,(zt)]T。在本文中，我们对（4）中的激",
"活函数采用以下两个假设。",
"Assumption 1. All activation functions o;() for i = 1, 2, .., d - 1 satisf)",
": ;() is continuous and non-decreasing and o;(O) ≥ 0",
". a;() is Li -Lipschitz, where Li > 0.",
"Assumption 2. For all activation functions oi, i = 1, 2, .., d -1, there exist positive constants q(i)",
"and Q.(i) such that for any 0 ≤ 1 ≤ 2 ≤ q(i), ;(2) -0i(ai) ≥ Q.(i)(r2 - a1) ",
"请注意，假设E1和2适用于各种常用的激活函数，包括ReLU，sigmoid，softPlus和tanh",
"特别是，在假设E2中，常量q(i）和Q。（i）的存在比它们所采用的特定值更重要，后者仅",
"3"
] | {
"arxiv_id": "1811.01054",
"language": "zh",
"page_idx": 2
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.