Datasets:

neeto-lab

/

arxiv-struct-100k

like 0

Follow

Neeto Lab 3

[ [ "0.1953", "0.1351", "0.777", "0.1351", "0.777", "0.1483", "0.1953", "0.1483" ], [ "0.1602", "0.178", "0.675", "0.178", "0.675", "0.1989", "0.1602", "0.1989" ], [ "0.6772", "0.1881", "0.6846", "0.1881", "0.6846", "0.1925", "0.6772", "0.1925" ], [ "0.689", "0.1843", "0.8096", "0.1812", "0.8096", "0.1963", "0.689", "0.1995" ], [ "0.2974", "0.1989", "0.671", "0.1989", "0.671", "0.2166", "0.2974", "0.2166" ], [ "0.3733", "0.214", "0.597", "0.2166", "0.597", "0.2343", "0.3733", "0.2317" ], [ "0.259", "0.2374", "0.71", "0.2361", "0.71", "0.2507", "0.259", "0.252" ], [ "0.3677", "0.2551", "0.6045", "0.254", "0.6045", "0.2683", "0.3677", "0.2695" ], [ "0.2378", "0.303", "0.3137", "0.303", "0.3137", "0.3208", "0.2378", "0.3208" ], [ "0.7466", "0.318", "0.7974", "0.318", "0.7974", "0.3289", "0.7466", "0.3289" ], [ "0.6235", "0.3276", "0.6465", "0.3276", "0.6465", "0.332", "0.6235", "0.332" ], [ "0.744", "0.3289", "0.7983", "0.327", "0.7983", "0.3372", "0.744", "0.339" ], [ "0.10046", "0.3372", "0.4673", "0.3372", "0.4673", "0.351", "0.10046", "0.351" ], [ "0.0801", "0.3528", "0.468", "0.3528", "0.468", "0.3662", "0.0801", "0.3662" ], [ "0.0809", "0.3687", "0.4656", "0.3687", "0.4656", "0.3818", "0.0809", "0.3818" ], [ "0.0809", "0.3833", "0.468", "0.3833", "0.468", "0.3965", "0.0809", "0.3965" ], [ "0.0809", "0.3997", "0.468", "0.3997", "0.468", "0.4104", "0.0809", "0.4104" ], [ "0.738", "0.3977", "0.789", "0.3977", "0.789", "0.4077", "0.738", "0.4077" ], [ "0.0801", "0.4136", "0.468", "0.4136", "0.468", "0.4268", "0.0801", "0.4268" ], [ "0.738", "0.4077", "0.789", "0.4077", "0.789", "0.4185", "0.738", "0.4185" ], [ "0.613", "0.4153", "0.689", "0.4153", "0.689", "0.4255", "0.613", "0.4255" ], [ "0.0801", "0.4287", "0.4673", "0.4287", "0.4673", "0.4426", "0.0801", "0.4426" ], [ "0.6167", "0.4248", "0.681", "0.4248", "0.681", "0.435", "0.6167", "0.435" ], [ "0.0801", "0.443", "0.468", "0.443", "0.468", "0.457", "0.0801", "0.457" ], [ "0.0784", "0.4583", "0.468", "0.457", "0.468", "0.4717", "0.0784", "0.473" ], [ "0.5024", "0.461", "0.888", "0.461", "0.888", "0.4749", "0.5024", "0.4749" ], [ "0.0801", "0.4734", "0.468", "0.4749", "0.468", "0.488", "0.0801", "0.4873" ], [ "0.5034", "0.4749", "0.892", "0.4734", "0.892", "0.4873", "0.5034", "0.488" ], [ "0.0801", "0.4893", "0.468", "0.4893", "0.468", "0.5024", "0.0801", "0.5024" ], [ "0.5024", "0.488", "0.8906", "0.488", "0.8906", "0.502", "0.5024", "0.502" ], [ "0.0801", "0.504", "0.468", "0.504", "0.468", "0.5176", "0.0801", "0.5176" ], [ "0.5024", "0.502", "0.8906", "0.502", "0.8906", "0.515", "0.5024", "0.515" ], [ "0.0809", "0.518", "0.4673", "0.518", "0.4673", "0.532", "0.0809", "0.532" ], [ "0.505", "0.5176", "0.8906", "0.5176", "0.8906", "0.528", "0.505", "0.528" ], [ "0.0809", "0.534", "0.4673", "0.534", "0.4673", "0.5474", "0.0809", "0.5474" ], [ "0.5024", "0.53", "0.8906", "0.53", "0.8906", "0.5435", "0.5024", "0.5435" ], [ "0.5024", "0.5435", "0.8896", "0.5435", "0.8896", "0.5566", "0.5024", "0.5566" ], [ "0.0809", "0.55", "0.468", "0.55", "0.468", "0.563", "0.0809", "0.563" ], [ "0.5034", "0.558", "0.892", "0.558", "0.892", "0.5713", "0.5034", "0.5713" ], [ "0.0801", "0.5645", "0.468", "0.5645", "0.468", "0.5776", "0.0801", "0.5776" ], [ "0.5024", "0.5713", "0.613", "0.5713", "0.613", "0.5854", "0.5024", "0.5854" ], [ "0.0809", "0.5806", "0.4673", "0.5806", "0.4673", "0.5913", "0.0809", "0.5913" ], [ "0.0809", "0.595", "0.4673", "0.595", "0.4673", "0.606", "0.0809", "0.606" ], [ "0.5034", "0.5996", "0.8906", "0.5996", "0.8906", "0.613", "0.5034", "0.613" ], [ "0.0801", "0.61", "0.4673", "0.61", "0.4673", "0.6235", "0.0801", "0.6235" ], [ "0.5034", "0.613", "0.8906", "0.6143", "0.8906", "0.629", "0.5034", "0.6274" ], [ "0.0809", "0.625", "0.4631", "0.625", "0.4631", "0.638", "0.0809", "0.638" ], [ "0.505", "0.629", "0.8906", "0.629", "0.8906", "0.6426", "0.505", "0.6426" ], [ "0.0801", "0.64", "0.38", "0.64", "0.38", "0.654", "0.0801", "0.654" ], [ "0.505", "0.643", "0.8906", "0.643", "0.8906", "0.657", "0.505", "0.657" ], [ "0.505", "0.66", "0.7656", "0.66", "0.7656", "0.674", "0.505", "0.674" ], [ "0.524", "0.6753", "0.8906", "0.6753", "0.8906", "0.6895", "0.524", "0.6895" ], [ "0.0801", "0.6875", "0.2157", "0.6875", "0.2157", "0.7007", "0.0801", "0.7007" ], [ "0.505", "0.69", "0.8906", "0.69", "0.8906", "0.704", "0.505", "0.704" ], [ "0.505", "0.7056", "0.8906", "0.7056", "0.8906", "0.7197", "0.505", "0.7197" ], [ "0.1013", "0.712", "0.468", "0.712", "0.468", "0.726", "0.1013", "0.726" ], [ "0.505", "0.72", "0.892", "0.72", "0.892", "0.7344", "0.505", "0.7344" ], [ "0.0825", "0.7266", "0.468", "0.7266", "0.468", "0.7407", "0.0825", "0.7407" ], [ "0.5034", "0.735", "0.864", "0.735", "0.864", "0.7485", "0.5034", "0.7485" ], [ "0.0825", "0.742", "0.468", "0.742", "0.468", "0.756", "0.0825", "0.756" ], [ "0.524", "0.752", "0.8906", "0.752", "0.8906", "0.765", "0.524", "0.765" ], [ "0.0825", "0.759", "0.4656", "0.759", "0.4656", "0.7695", "0.0825", "0.7695" ], [ "0.5034", "0.7666", "0.886", "0.7666", "0.886", "0.78", "0.5034", "0.78" ], [ "0.0809", "0.773", "0.468", "0.773", "0.468", "0.786", "0.0809", "0.786" ], [ "0.505", "0.7827", "0.5884", "0.7827", "0.5884", "0.7935", "0.505", "0.7935" ], [ "0.581", "0.782", "0.8906", "0.782", "0.8906", "0.7954", "0.581", "0.7954" ], [ "0.0825", "0.7886", "0.1814", "0.7886", "0.1814", "0.8", "0.0825", "0.8" ], [ "0.505", "0.797", "0.8906", "0.797", "0.8906", "0.81", "0.505", "0.81" ], [ "0.10046", "0.803", "0.468", "0.803", "0.468", "0.8164", "0.10046", "0.8164" ], [ "0.505", "0.811", "0.8896", "0.811", "0.8896", "0.8247", "0.505", "0.8247" ], [ "0.0825", "0.8184", "0.468", "0.8184", "0.468", "0.832", "0.0825", "0.832" ], [ "0.5034", "0.827", "0.8906", "0.827", "0.8906", "0.8403", "0.5034", "0.8403" ], [ "0.0809", "0.834", "0.4673", "0.834", "0.4673", "0.8477", "0.0809", "0.8477" ], [ "0.5034", "0.8413", "0.8896", "0.8413", "0.8896", "0.855", "0.5034", "0.855" ], [ "0.0825", "0.851", "0.468", "0.851", "0.468", "0.8613", "0.0825", "0.8613" ], [ "0.5034", "0.857", "0.8906", "0.857", "0.8906", "0.8706", "0.5034", "0.8706" ], [ "0.0972", "0.875", "0.468", "0.875", "0.468", "0.8896", "0.0972", "0.8896" ], [ "0.5034", "0.8706", "0.8906", "0.872", "0.8906", "0.886", "0.5034", "0.8853" ], [ "0.0784", "0.886", "0.398", "0.887", "0.398", "0.901", "0.0784", "0.8994" ], [ "0.505", "0.887", "0.7656", "0.887", "0.7656", "0.901", "0.505", "0.901" ], [ "0.4805", "0.925", "0.492", "0.925", "0.492", "0.9375", "0.4805", "0.9375" ] ]

[ "RepPoints: Point Set Representation for Object Detection", "Ze Yang It * Shaohui Liu 2,3 Han Hu 3 Liwei Wang 1", "", "Stephen Lin 3", "1 Peking University2 Tsinghua University", "3 Microsoft Research Asia", "[email protected],[email protected],[email protected]", "[hanhu, stevelin @microsoft.com", "Abstract", "localization", "onve", "supervision", "Modern object detectors rely heavily on rectangu-", "lar bounding boxes, such as anchors, proposals and the", "final predictions, to represent objects at various recog-", "nition stages. The bounding box is convenient to use", "but provides only a coarse localization of objects and", "recognition", "leads to a correspondingly coarse extraction of object", "supervision", "eature extraction", "features. In this paper, we present RepPoints (repre-", "+ classification", "sentative points), a new finer representation of objects", "as a set of sample points useful for both localization and", "Figure 1. RepPoints is a new representation for object de", "recognition. Given ground truth localization and recog-", "tection that consists of a set of points which indicate the", "nition targets for training, RepPoints learn to automat-", "spatial extent of an object and semantically significant lo-", "ically arrange themselves in a manner that bounds the", "cal areas. This representation is learned via weak localiza-", "spatial extent of an object and indicates semantically", "tionsupervisionfromrectangular ground-truthboxes and", "significant local areas. They furthermore do not re-", "implicit recognition feedback. Based on the richer Rep-", "Points representation, we develop an anchor-free object de-", "quire the use of anchors to sample a space of bounding", "tector that yields improved performance compared to using", "boxes. We show that an anchor-free object detector", "bounding boxes.", "based on RepPoints can be as effective as the state-of-", "the-art anchor-based detection methods. with 46.5 AP", "边界框表示的流行可以部分归因于对象检测性能的常", "and 67.4 APso on the COCO test-dev detection bench-", "见指标，这些指标解释了对象的估计边界框和地面实", "mark, using ResNet-101 model. Code is available at", "况边界框之间的重叠。另一个原因在于它便于在深度", "https://github.com/microsoft/RepPoints", "网络中进行特征提取，因为它具有规则的形状，并且易", "于将矩形窗口细分为池化单元的矩阵", "虽然边界框便于计算，但它们仅提供不符合对象形", "1.Introduction", "状和姿势的对象的粗略定位。因此，从边界框的规则单", "元格中提取的特征可能会受到背景内容或包含很少语", "对象检测旨在定位图像中的对象并提供其类标签。作", "义信息的无信息前景区域的严重影响。这可能会导致", "为计算机视觉中最基本的任务之一，它是许多视觉应", "较低的特征质量，从而降低对象检测中的分类性能", "用程序的关键组件，包括实例分割[30]，人体姿态分析", "在本文中，我们提出了一种新的表示形式，称为", "37和视觉推理「41。目标检测问题的重要性以及深度", "RepPoints，它提供了更细粒度的本地化并促进了分类", "神经网络的快速发展导致了近年来[7，11，10,33，14，3]", "如图所示。", "1，RepPoints是一组点，它们学习以限", "的实质性进展", "定对象的空间范围并指示语义重要的局部区域的方式", "在对象检测管道中，包含图像矩形区域的边界框用", "自适应地将自已定位在对象上。RepPoints的训练是由", "作处理的基本元素。它们描述了物体探测器各个阶段", "对象定位和识别目标共同驱动的，因此RepPoints受", "的目标位置，从锚点和提议到最终预测。基于这些边界", "到地面真相边界框的严格约束，并引导检测器进行正", "框，提取特征并将其用于对象分类和位置细化等自的。", "确的对象分类。这种自适应和可区分的表示可以在现", "*Equal contribution. t The work is done when Ze Yang and", "代对象检测器的不同阶段中连贯地使用，并且不需要", "Shaohui Liu are interns at Microsoft Research Asia.", "使用锚点在边界框的空间上进行采样", "1" ]

{ "arxiv_id": "1904.11490", "language": "zh", "page_idx": 0 }

[ [ "0.10046", "0.09344", "0.4722", "0.09344", "0.4722", "0.1067", "0.10046", "0.1067" ], [ "0.509", "0.09344", "0.8203", "0.09344", "0.8203", "0.1042", "0.509", "0.1042" ], [ "0.841", "0.0947", "0.8896", "0.0947", "0.8896", "0.1061", "0.841", "0.1061" ], [ "0.08417", "0.10986", "0.4673", "0.10986", "0.4673", "0.1206", "0.08417", "0.1206" ], [ "0.5034", "0.108", "0.8896", "0.108", "0.8896", "0.1218", "0.5034", "0.1218" ], [ "0.0825", "0.1237", "0.468", "0.1237", "0.468", "0.137", "0.0825", "0.137" ], [ "0.5034", "0.1237", "0.8906", "0.1237", "0.8906", "0.137", "0.5034", "0.137" ], [ "0.0809", "0.1383", "0.468", "0.1383", "0.468", "0.1515", "0.0809", "0.1515" ], [ "0.5034", "0.1395", "0.8447", "0.1395", "0.8447", "0.1528", "0.5034", "0.1528" ], [ "0.0825", "0.154", "0.468", "0.154", "0.468", "0.1674", "0.0825", "0.1674" ], [ "0.523", "0.1559", "0.8906", "0.1559", "0.8906", "0.1692", "0.523", "0.1692" ], [ "0.0825", "0.1686", "0.4673", "0.1686", "0.4673", "0.1818", "0.0825", "0.1818" ], [ "0.505", "0.1704", "0.8896", "0.1704", "0.8896", "0.1843", "0.505", "0.1843" ], [ "0.08417", "0.1849", "0.4673", "0.1849", "0.4673", "0.1957", "0.08417", "0.1957" ], [ "0.5024", "0.1843", "0.8906", "0.1849", "0.8906", "0.1995", "0.5024", "0.1989" ], [ "0.0825", "0.1989", "0.468", "0.1989", "0.468", "0.2122", "0.0825", "0.2122" ], [ "0.5034", "0.2008", "0.8906", "0.2008", "0.8906", "0.214", "0.5034", "0.214" ], [ "0.0825", "0.214", "0.246", "0.214", "0.246", "0.2279", "0.0825", "0.2279" ], [ "0.5063", "0.2172", "0.888", "0.2172", "0.888", "0.2279", "0.5063", "0.2279" ], [ "0.1013", "0.2311", "0.468", "0.2311", "0.468", "0.2443", "0.1013", "0.2443" ], [ "0.5034", "0.2311", "0.8906", "0.2311", "0.8906", "0.2443", "0.5034", "0.2443" ], [ "0.0809", "0.2456", "0.468", "0.2456", "0.468", "0.2588", "0.0809", "0.2588" ], [ "0.5034", "0.2456", "0.8906", "0.2456", "0.8906", "0.2588", "0.5034", "0.2588" ], [ "0.0825", "0.2615", "0.468", "0.2615", "0.468", "0.2747", "0.0825", "0.2747" ], [ "0.5034", "0.2615", "0.8906", "0.2615", "0.8906", "0.2747", "0.5034", "0.2747" ], [ "0.0825", "0.2747", "0.468", "0.2766", "0.468", "0.2903", "0.0825", "0.2886" ], [ "0.505", "0.2766", "0.8906", "0.2766", "0.8906", "0.2903", "0.505", "0.2903" ], [ "0.0825", "0.2917", "0.468", "0.2917", "0.468", "0.305", "0.0825", "0.305" ], [ "0.505", "0.2917", "0.8896", "0.2917", "0.8896", "0.305", "0.505", "0.305" ], [ "0.0809", "0.305", "0.4673", "0.3062", "0.4673", "0.3208", "0.0809", "0.3193" ], [ "0.5034", "0.307", "0.8896", "0.307", "0.8896", "0.3208", "0.5034", "0.3208" ], [ "0.09644", "0.3213", "0.468", "0.3213", "0.468", "0.3352", "0.09644", "0.3352" ], [ "0.5034", "0.3213", "0.892", "0.3213", "0.892", "0.3352", "0.5034", "0.3352" ], [ "0.0801", "0.3352", "0.407", "0.336", "0.407", "0.351", "0.0801", "0.3499" ], [ "0.5034", "0.3352", "0.681", "0.3352", "0.681", "0.3484", "0.5034", "0.3484" ], [ "0.5254", "0.3528", "0.8906", "0.3528", "0.8906", "0.3662", "0.5254", "0.3662" ], [ "0.0801", "0.3718", "0.2296", "0.3718", "0.2296", "0.385", "0.0801", "0.385" ], [ "0.5063", "0.3706", "0.8896", "0.3706", "0.8896", "0.3806", "0.5063", "0.3806" ], [ "0.5063", "0.3865", "0.888", "0.3865", "0.888", "0.3965", "0.5063", "0.3965" ], [ "0.0866", "0.3977", "0.451", "0.3977", "0.451", "0.411", "0.0866", "0.411" ], [ "0.505", "0.3997", "0.8896", "0.3997", "0.8896", "0.4136", "0.505", "0.4136" ], [ "0.0809", "0.4136", "0.468", "0.4136", "0.468", "0.4268", "0.0809", "0.4268" ], [ "0.5034", "0.414", "0.8896", "0.414", "0.8896", "0.428", "0.5034", "0.428" ], [ "0.0809", "0.428", "0.4673", "0.428", "0.4673", "0.4414", "0.0809", "0.4414" ], [ "0.505", "0.4312", "0.5864", "0.4312", "0.5864", "0.4426", "0.505", "0.4426" ], [ "0.0809", "0.443", "0.4673", "0.443", "0.4673", "0.457", "0.0809", "0.457" ], [ "0.0809", "0.459", "0.468", "0.459", "0.468", "0.473", "0.0809", "0.473" ], [ "0.5073", "0.459", "0.835", "0.4602", "0.835", "0.4734", "0.5073", "0.473" ], [ "0.841", "0.4653", "0.853", "0.4653", "0.853", "0.4697", "0.841", "0.4697" ], [ "0.8555", "0.461", "0.8906", "0.461", "0.8906", "0.473", "0.8555", "0.473" ], [ "0.0809", "0.4734", "0.468", "0.4734", "0.468", "0.4873", "0.0809", "0.4873" ], [ "0.5034", "0.4734", "0.8896", "0.4749", "0.8896", "0.488", "0.5034", "0.4873" ], [ "0.0801", "0.4873", "0.468", "0.488", "0.468", "0.5024", "0.0801", "0.502" ], [ "0.5034", "0.4905", "0.8896", "0.4905", "0.8896", "0.504", "0.5034", "0.504" ], [ "0.0809", "0.504", "0.1969", "0.504", "0.1969", "0.5176", "0.0809", "0.5176" ], [ "0.505", "0.506", "0.8896", "0.506", "0.8896", "0.5195", "0.505", "0.5195" ], [ "0.10046", "0.521", "0.4673", "0.521", "0.4673", "0.534", "0.10046", "0.534" ], [ "0.505", "0.521", "0.8896", "0.521", "0.8896", "0.534", "0.505", "0.534" ], [ "0.0825", "0.536", "0.4673", "0.536", "0.4673", "0.55", "0.0825", "0.55" ], [ "0.5034", "0.535", "0.8896", "0.535", "0.8896", "0.549", "0.5034", "0.549" ], [ "0.0809", "0.5503", "0.4697", "0.552", "0.4697", "0.565", "0.0809", "0.5645" ], [ "0.5024", "0.55", "0.8906", "0.5503", "0.8906", "0.5645", "0.5024", "0.563" ], [ "0.0825", "0.5674", "0.4673", "0.5674", "0.4673", "0.5776", "0.0825", "0.5776" ], [ "0.5024", "0.565", "0.5435", "0.565", "0.5435", "0.58", "0.5024", "0.58" ], [ "0.0809", "0.582", "0.1584", "0.582", "0.1584", "0.5933", "0.0809", "0.5933" ], [ "0.523", "0.5835", "0.8906", "0.5835", "0.8906", "0.5967", "0.523", "0.5967" ], [ "0.1013", "0.5977", "0.468", "0.5977", "0.468", "0.6113", "0.1013", "0.6113" ], [ "0.5063", "0.5996", "0.8896", "0.5996", "0.8896", "0.61", "0.5063", "0.61" ], [ "0.0809", "0.613", "0.468", "0.613", "0.468", "0.627", "0.0809", "0.627" ], [ "0.5024", "0.6123", "0.8896", "0.6123", "0.8896", "0.6255", "0.5024", "0.6255" ], [ "0.0809", "0.6274", "0.468", "0.6274", "0.468", "0.6416", "0.0809", "0.6416" ], [ "0.5063", "0.63", "0.8896", "0.63", "0.8896", "0.64", "0.5063", "0.64" ], [ "0.0809", "0.6426", "0.4673", "0.643", "0.4673", "0.657", "0.0809", "0.656" ], [ "0.5034", "0.643", "0.8906", "0.643", "0.8906", "0.657", "0.5034", "0.657" ], [ "0.0825", "0.66", "0.4673", "0.66", "0.4673", "0.6704", "0.0825", "0.6704" ], [ "0.5063", "0.66", "0.8896", "0.66", "0.8896", "0.6704", "0.5063", "0.6704" ], [ "0.0809", "0.674", "0.468", "0.674", "0.468", "0.6875", "0.0809", "0.6875" ], [ "0.505", "0.6753", "0.886", "0.6753", "0.886", "0.686", "0.505", "0.686" ], [ "0.0809", "0.6895", "0.1953", "0.6895", "0.1953", "0.7026", "0.0809", "0.7026" ], [ "0.5034", "0.688", "0.8896", "0.688", "0.8896", "0.702", "0.5034", "0.702" ], [ "0.5034", "0.7026", "0.5557", "0.7026", "0.5557", "0.7173", "0.5034", "0.7173" ], [ "0.0866", "0.7197", "0.415", "0.7197", "0.415", "0.733", "0.0866", "0.733" ], [ "0.415", "0.7217", "0.4355", "0.7217", "0.4355", "0.729", "0.415", "0.729" ], [ "0.4314", "0.7183", "0.4673", "0.72", "0.4656", "0.732", "0.4307", "0.73" ], [ "0.0825", "0.735", "0.4673", "0.735", "0.4673", "0.7485", "0.0825", "0.7485" ], [ "0.5034", "0.733", "0.789", "0.733", "0.789", "0.747", "0.5034", "0.747" ], [ "0.0825", "0.75", "0.4673", "0.75", "0.4673", "0.763", "0.0825", "0.763" ], [ "0.0825", "0.7666", "0.4673", "0.7666", "0.4673", "0.7764", "0.0825", "0.7764" ], [ "0.5254", "0.761", "0.888", "0.761", "0.888", "0.7715", "0.5254", "0.7715" ], [ "0.0809", "0.78", "0.4592", "0.78", "0.4592", "0.7935", "0.0809", "0.7935" ], [ "0.5034", "0.7744", "0.8906", "0.7744", "0.8906", "0.7886", "0.5034", "0.7886" ], [ "0.10046", "0.797", "0.468", "0.797", "0.468", "0.811", "0.10046", "0.811" ], [ "0.505", "0.7915", "0.5537", "0.7915", "0.5537", "0.803", "0.505", "0.803" ], [ "0.0809", "0.811", "0.4673", "0.811", "0.4673", "0.8247", "0.0809", "0.8247" ], [ "0.5034", "0.821", "0.7715", "0.821", "0.7715", "0.834", "0.5034", "0.834" ], [ "0.0825", "0.827", "0.468", "0.827", "0.468", "0.8403", "0.0825", "0.8403" ], [ "0.08417", "0.8433", "0.4673", "0.8433", "0.4673", "0.854", "0.08417", "0.854" ], [ "0.5215", "0.8447", "0.8896", "0.845", "0.8896", "0.86", "0.5215", "0.8594" ], [ "0.0801", "0.856", "0.468", "0.857", "0.468", "0.8706", "0.0801", "0.869" ], [ "0.5034", "0.8613", "0.8896", "0.8613", "0.8896", "0.875", "0.5034", "0.875" ], [ "0.0809", "0.8726", "0.4714", "0.8726", "0.4714", "0.886", "0.0809", "0.886" ], [ "0.5063", "0.8774", "0.8896", "0.8774", "0.8896", "0.888", "0.5063", "0.888" ], [ "0.0809", "0.887", "0.3586", "0.887", "0.3586", "0.901", "0.0809", "0.901" ], [ "0.505", "0.8926", "0.8774", "0.8926", "0.8774", "0.903", "0.505", "0.903" ], [ "0.4805", "0.925", "0.492", "0.925", "0.492", "0.9375", "0.4805", "0.9375" ] ]

[ "RepPoints与用于对象检测的现有非矩形表示不同，", "Other representationsfor object detection.", "为了解", "后者均以自下而上的方式构建38，21，48。这些自下", "决矩形边界框的限制，已经有一些尝试开发更灵活的", "而上的表示标识单个点（例如。g。、边界框角或对象末", "对象表示。其中包括用于行人检测[22]的椭圆表示和", "端），并依靠手工聚类将它们分组到对象模型中。此外，", "用于更好地处理旋转变化[16，49]的旋转边界框", "它们的表示形式仍然像边界框[38，21]一样轴对齐，或", "其他作品旨在以自下而上的方式表示对象。早期的", "者需要地面实况对象掩码作为附加监督【48]。相比之", "自下而上表示包括DPM[9]和Poselet[1]。最近，已", "下，RepPoints以自上而下的方式从输入图像/对象特", "经使用深度网络[21，48]探索了自下而上的对象检测方", "征中学习，允许端到端训练并在没有额外监督的情况", "法。CornerNet[21]首先预测左上角和右下角，然后采", "下产生细粒度的本地化", "用专门的分组方法28获得对象的边界框。然而，两个", "通过RepPoints替换两级对象检测器中的所有常规", "相对的角点仍然本质上是一个矩形边界框的模型。提", "边界框表示，包括锚，提议和最终定位目标，我们开", "出了ExtremeNetE[48]来定位X和y方向[29]上物", "发了一种干净有效的无锚对象检测器，该检测器实现", "体的极值点，并从地面真相掩码注释中进行监督。通", "了42。8AP和65。0AP50在COCO基准[26]上，无", "常，自下而上的检测器受益于较小的假设空间（例如，", "需进行多尺度训练和测试，实现了46。5AP和67。4", "CornerNet和ExtremeNet都检测2-d点，而不是直接", "使用ResNet-101模型进行多尺度训练和测试的AP50", "检测4-d边界框）和潜在的细粒度定位。但是，它们具", "所提出的目标检测器不仅超越了所有现有的无锚检", "有局限性，例如依靠手工制作的群集或后处理步骤从", "测器，而且与最先进的基于锚的基线性能相当", "检测到的点组成整个对象", "与这些自下而上的作品类似，RepPoints也是一种", "2. Related Work", "灵活的对象表示。但是，表示是以自上而下的方式构建", "的，不需要手工制作的聚类步骤。RepPoints可以自动", "Bounding boxes for the object detection problem", "学习极端点和关键语义点，而无需超出基本事实边界", "边界框长期以来一直是对象检测领域中对象表示的主", "框的监督，这与ExtremeNet【48]不同，后者需要额外", "要形式。其流行的一个原因是边界框便于注释，几乎", "的掩码监督", "没有歧义，同时为随后的识别过程提供足够准确的定", "位。这可以解释为什么主要的基准都使用基于边界框", "Deformationmodeling in object recognition.", "", "视觉", "[8，26，20]的注释和评估。反过来，这些基准测试激励", "识别的最基本挑战之一是识别具有各种几何变化的对", "对象检测方法使用边界框作为其基本表示，以便与评", "象。为了有效地对这种变化进行建模，一种可能的解决", "估协议保持一致", "方案是利用低级组件的自下而上的组合。沿此方向的", "边界框占主导地位的另一个原因是，在[39，5]之", "代表性探测器包括DPM9]和Poselet[i]。另一种方", "前和深度学习时代[19，34，36，15]期间，几乎所有图", "法是以自上而下的方式隐式地对转换进行建模，其中", "像特征提取器都基于具有规则网格形式的输入补丁。", "将轻量级神经网络块应用于输入特征，全局【18]或局", "因此，可以方便地使用边界框表示来促进特征提取", "部[4]", "[11,10,33]", "RepPoints受到这些作品的启发，尤其是自上而下", "尽管建议的RepPoints具有不规则的形式，但我", "的变形建模方法4。主要区别在于，除了语义特征提", "们表明它可以方便地进行特征提取。我们的系统将", "取之外，我们的目标是开发一种灵活的对象表示，以实", "RepPoints与可变形卷积E[4]结合使用，它自然地与", "现精确的儿何定位。相比之下，可变形卷积和可变形", "RepPoints对齐，因为它聚合了来自多个采样点的输人", "RoI池化方法都是为了改进特征提取而设计的。在E4", "特征的信息。此外，可以很容易地从RepPoints生成矩", "节和附录中检查了可变形RoI池无法学习精确的儿何", "形伪框（请参阅3.2节），从而允许将新表示与对象检", "定位。从这个意义上说，我们扩展了以前的儿何建模方", "测基准一起使用", "法【18，4中自适应采样点的使用，以包括对象的更精", "细定位", "Bounding boxes in modern object detectors.", "", "汽今", "为止，性能最好的对象检测器通常遵循多阶段识别范", "3.TheRepPointsRepresentation", "例[24，4，13，23，2，35，27]，并且边界框表示几乎出", "现在所有阶段：1）作为预定义的33，25或学习的E", "我们首先回顾边界框表示及其在多级对象检测器中", "【40，45，47]目标，作为边界框建议空间上的假设；2】", "的使用。接下来是对RepPoints及其与边界框的区别", "RepPoints可用于替换模型对象检测器所有阶段中", "的描述", "的边界框表示，从而产生更有效的新对象检测器。具体", "3.1. Bounding Box Representation", "来说，锚被替换为中心点，这是RepPoints的特殊配", "置。边界框提案和最终本地化目标被RepPoints提案", "边界框是对对象B=（,y,w,h）的空间位置进行编", "和最终目标所取代。请注意，由于使用中心点进行初始", "码的4维表示，其中,9表示中心点，w,h表示宽度", "对象表示，因此生成的对象检测器是无锚点的。因此，", "和高度。由于其使用简单方便，现代对象检测器在很大", "它在使用上比基于边界框的方法更方便", "程度上依赖于边界框来表示检测管道各个阶段的对象", "2" ]

{ "arxiv_id": "1904.11490", "language": "zh", "page_idx": 1 }

[ [ "0.0866", "0.09344", "0.4697", "0.09344", "0.4697", "0.1067", "0.0866", "0.1067" ], [ "0.505", "0.0947", "0.888", "0.0947", "0.888", "0.1048", "0.505", "0.1048" ], [ "0.0809", "0.108", "0.468", "0.108", "0.468", "0.1218", "0.0809", "0.1218" ], [ "0.505", "0.10986", "0.7793", "0.10986", "0.7793", "0.1206", "0.505", "0.1206" ], [ "0.0801", "0.1225", "0.4673", "0.1237", "0.4673", "0.137", "0.0801", "0.1364" ], [ "0.524", "0.1237", "0.892", "0.1237", "0.892", "0.137", "0.524", "0.137" ], [ "0.0825", "0.1401", "0.1217", "0.1401", "0.1217", "0.1515", "0.0825", "0.1515" ], [ "0.505", "0.1383", "0.626", "0.1383", "0.626", "0.1515", "0.505", "0.1515" ], [ "0.2222", "0.1635", "0.2834", "0.1674", "0.282", "0.1812", "0.2214", "0.1768" ], [ "0.6284", "0.1654", "0.7646", "0.1674", "0.7646", "0.1818", "0.6284", "0.1799" ], [ "0.869", "0.1666", "0.892", "0.1666", "0.892", "0.1812", "0.869", "0.1812" ], [ "0.2737", "0.1674", "0.4453", "0.1686", "0.4453", "0.1863", "0.2737", "0.1849" ], [ "0.1054", "0.1718", "0.2009", "0.1718", "0.2009", "0.1831", "0.1054", "0.1831" ], [ "0.2181", "0.1799", "0.4453", "0.1881", "0.4436", "0.2102", "0.2173", "0.202" ], [ "0.5034", "0.1932", "0.8936", "0.1932", "0.8936", "0.207", "0.5034", "0.207" ], [ "0.5034", "0.209", "0.6167", "0.209", "0.6167", "0.2229", "0.5034", "0.2229" ], [ "0.2197", "0.2153", "0.4429", "0.2235", "0.4412", "0.2443", "0.2181", "0.2361" ], [ "0.4478", "0.2424", "0.4656", "0.2424", "0.4656", "0.252", "0.4478", "0.252" ], [ "0.5107", "0.2374", "0.8896", "0.2374", "0.8896", "0.2474", "0.5107", "0.2474" ], [ "0.10046", "0.2556", "0.4673", "0.2556", "0.4673", "0.2695", "0.10046", "0.2695" ], [ "0.505", "0.2524", "0.888", "0.2524", "0.888", "0.2632", "0.505", "0.2632" ], [ "0.0809", "0.2708", "0.468", "0.2708", "0.468", "0.2842", "0.0809", "0.2842" ], [ "0.5034", "0.2664", "0.7563", "0.2664", "0.7563", "0.2798", "0.5034", "0.2798" ], [ "0.08417", "0.2854", "0.468", "0.2854", "0.468", "0.2986", "0.08417", "0.2986" ], [ "0.574", "0.2922", "0.8677", "0.2935", "0.8677", "0.308", "0.574", "0.307" ], [ "0.0801", "0.2998", "0.1896", "0.3005", "0.1896", "0.315", "0.0801", "0.3145" ], [ "0.8613", "0.2942", "0.8906", "0.2942", "0.8906", "0.307", "0.8613", "0.307" ], [ "0.10046", "0.3164", "0.468", "0.3164", "0.468", "0.33", "0.10046", "0.33" ], [ "0.5195", "0.3213", "0.8896", "0.3213", "0.8896", "0.336", "0.5195", "0.336" ], [ "0.0825", "0.3333", "0.4673", "0.3333", "0.4673", "0.3435", "0.0825", "0.3435" ], [ "0.5034", "0.336", "0.8896", "0.336", "0.8896", "0.3499", "0.5034", "0.3499" ], [ "0.0809", "0.3467", "0.468", "0.3467", "0.468", "0.3599", "0.0809", "0.3599" ], [ "0.505", "0.3516", "0.8906", "0.3516", "0.8906", "0.3655", "0.505", "0.3655" ], [ "0.0825", "0.363", "0.3145", "0.363", "0.3145", "0.3738", "0.0825", "0.3738" ], [ "0.505", "0.3674", "0.749", "0.3674", "0.749", "0.3806", "0.505", "0.3806" ], [ "0.10046", "0.377", "0.4673", "0.377", "0.4673", "0.3901", "0.10046", "0.3901" ], [ "0.0809", "0.3926", "0.468", "0.3926", "0.468", "0.406", "0.0809", "0.406" ], [ "0.5107", "0.3882", "0.8906", "0.3882", "0.8906", "0.4016", "0.5107", "0.4016" ], [ "0.5024", "0.401", "0.8896", "0.4016", "0.8896", "0.4167", "0.5024", "0.4153" ], [ "0.0825", "0.4092", "0.4673", "0.4092", "0.4673", "0.4192", "0.0825", "0.4192" ], [ "0.0809", "0.4224", "0.468", "0.4224", "0.468", "0.4363", "0.0809", "0.4363" ], [ "0.5034", "0.4172", "0.8906", "0.4172", "0.8906", "0.4312", "0.5034", "0.4312" ], [ "0.0809", "0.4375", "0.468", "0.4375", "0.468", "0.4507", "0.0809", "0.4507" ], [ "0.5034", "0.433", "0.888", "0.433", "0.888", "0.4463", "0.5034", "0.4463" ], [ "0.0809", "0.4526", "0.468", "0.4526", "0.468", "0.4666", "0.0809", "0.4666" ], [ "0.5034", "0.4475", "0.892", "0.4475", "0.892", "0.461", "0.5034", "0.461" ], [ "0.5024", "0.461", "0.626", "0.4622", "0.626", "0.4766", "0.5024", "0.476" ], [ "0.0825", "0.4697", "0.1462", "0.4697", "0.1462", "0.4797", "0.0825", "0.4797" ], [ "0.523", "0.4778", "0.744", "0.4766", "0.744", "0.4912", "0.523", "0.4924" ], [ "0.10046", "0.4841", "0.468", "0.4841", "0.468", "0.4976", "0.10046", "0.4976" ], [ "0.0809", "0.4988", "0.468", "0.4988", "0.468", "0.512", "0.0809", "0.512" ], [ "0.5156", "0.502", "0.8896", "0.504", "0.8896", "0.5176", "0.5156", "0.515" ], [ "0.0809", "0.5146", "0.246", "0.5146", "0.246", "0.529", "0.0809", "0.529" ], [ "0.1013", "0.521", "0.2173", "0.5234", "0.2173", "0.537", "0.1013", "0.534" ], [ "0.5337", "0.5176", "0.8896", "0.518", "0.8896", "0.532", "0.5337", "0.531" ], [ "0.535", "0.5327", "0.8896", "0.5327", "0.8896", "0.547", "0.535", "0.547" ], [ "0.536", "0.55", "0.7036", "0.55", "0.7036", "0.563", "0.536", "0.563" ], [ "0.08826", "0.558", "0.468", "0.558", "0.468", "0.5713", "0.08826", "0.5713" ], [ "0.0825", "0.5737", "0.468", "0.5737", "0.468", "0.587", "0.0825", "0.587" ], [ "0.517", "0.5737", "0.8906", "0.5747", "0.8906", "0.5884", "0.517", "0.587" ], [ "0.0801", "0.589", "0.4307", "0.5884", "0.4307", "0.603", "0.0801", "0.604" ], [ "0.535", "0.5884", "0.8896", "0.5884", "0.8896", "0.6016", "0.535", "0.6016" ], [ "0.536", "0.604", "0.8906", "0.604", "0.8906", "0.617", "0.536", "0.617" ], [ "0.08496", "0.6157", "0.468", "0.6143", "0.468", "0.632", "0.08496", "0.6333" ], [ "0.536", "0.6206", "0.6006", "0.6206", "0.6006", "0.632", "0.536", "0.632" ], [ "0.0825", "0.646", "0.468", "0.646", "0.468", "0.659", "0.0825", "0.659" ], [ "0.517", "0.6445", "0.8906", "0.6445", "0.8906", "0.6577", "0.517", "0.6577" ], [ "0.0809", "0.66", "0.468", "0.66", "0.468", "0.674", "0.0809", "0.674" ], [ "0.535", "0.659", "0.892", "0.659", "0.892", "0.6724", "0.535", "0.6724" ], [ "0.0809", "0.6753", "0.468", "0.6753", "0.468", "0.6895", "0.0809", "0.6895" ], [ "0.5376", "0.6753", "0.8896", "0.6753", "0.8896", "0.686", "0.5376", "0.686" ], [ "0.0809", "0.6914", "0.1054", "0.6914", "0.1054", "0.704", "0.0809", "0.704" ], [ "0.536", "0.6895", "0.892", "0.6895", "0.892", "0.7026", "0.536", "0.7026" ], [ "0.0866", "0.709", "0.3962", "0.709", "0.3962", "0.759", "0.0866", "0.759" ], [ "0.535", "0.705", "0.8906", "0.7056", "0.8906", "0.72", "0.535", "0.7197" ], [ "0.3865", "0.7153", "0.423", "0.7197", "0.4224", "0.731", "0.384", "0.728" ], [ "0.3628", "0.7227", "0.4272", "0.7246", "0.4265", "0.752", "0.361", "0.75" ], [ "0.423", "0.7246", "0.4673", "0.7266", "0.4656", "0.738", "0.4224", "0.7363" ], [ "0.2091", "0.733", "0.3064", "0.73", "0.308", "0.7476", "0.21", "0.7505" ], [ "0.505", "0.7456", "0.8896", "0.7456", "0.8896", "0.759", "0.505", "0.759" ], [ "0.1013", "0.762", "0.4673", "0.762", "0.4673", "0.776", "0.1013", "0.776" ], [ "0.505", "0.761", "0.8906", "0.761", "0.8906", "0.7744", "0.505", "0.7744" ], [ "0.08417", "0.779", "0.4656", "0.779", "0.4656", "0.789", "0.08417", "0.789" ], [ "0.5034", "0.776", "0.5703", "0.778", "0.5703", "0.789", "0.5034", "0.787" ], [ "0.0825", "0.7944", "0.4656", "0.7944", "0.4656", "0.8047", "0.0825", "0.8047" ], [ "0.5073", "0.797", "0.888", "0.797", "0.888", "0.81", "0.5073", "0.81" ], [ "0.09235", "0.808", "0.4673", "0.808", "0.4673", "0.8213", "0.09235", "0.8213" ], [ "0.505", "0.811", "0.8906", "0.811", "0.8906", "0.8247", "0.505", "0.8247" ], [ "0.0809", "0.8228", "0.343", "0.8228", "0.343", "0.836", "0.0809", "0.836" ], [ "0.505", "0.827", "0.8906", "0.827", "0.8906", "0.8403", "0.505", "0.8403" ], [ "0.5034", "0.8423", "0.8906", "0.8423", "0.8906", "0.856", "0.5034", "0.856" ], [ "0.0801", "0.8486", "0.1978", "0.8486", "0.1978", "0.8623", "0.0801", "0.8623" ], [ "0.5024", "0.856", "0.8906", "0.857", "0.8906", "0.8706", "0.5024", "0.869" ], [ "0.10046", "0.8726", "0.468", "0.8726", "0.468", "0.886", "0.10046", "0.886" ], [ "0.5034", "0.8726", "0.8906", "0.8726", "0.8906", "0.886", "0.5034", "0.886" ], [ "0.0809", "0.887", "0.4697", "0.887", "0.4697", "0.901", "0.0809", "0.901" ], [ "0.501", "0.886", "0.892", "0.887", "0.892", "0.901", "0.501", "0.8994" ], [ "0.478", "0.924", "0.4944", "0.924", "0.4944", "0.938", "0.478", "0.938" ] ]

[ "ReviewofMulti-StageObjectDetectors性能最佳", "和姿态以及语义上重要的局部区域的位置，这可以用", "的对象检测器通常遵循多阶段识别范例，其中对象定", "于更精细的定位和更好的对象特征提取", "位是逐步细化的。通过此管道的步骤，对象表示的作用", "为了克服上述限制，RepPoints改为对一组自适应", "如下：", "采样点进行建模：", "bbox reg.", "R ={(k, yk))\"=1,", "(4)", "bbox proposals (S1)", "bbox anchors", "bbox reg, bbox proposals (S2)", "其中n是表示中使用的采样点总数。在我们的工作中，", "n默认设置为 9", "bbox reg-, bbox object targets", "(1)", "RepPointsrefinement逐步细化边界框定位和特征", "在开始时，假设多个锚点覆盖一系列边界框比例和", "提取对于多阶段目标检测方法的成功至关重要。对于", "纵横比。一般来说，高覆盖率是通过大的4-d假设空间", "RepPoints，细化可以简单地表示为", "上的密集锚获得的。例如，RetinaNet25]中每个位置", "Rr=(ck+Ak,yk+Ayk))=1,", "使用45个锚点", "(5)", "对于锚点，采用其中心点处的图像特征作为对象特", "其中[(△ck,△yk))\"-1是新采样点相对于旧采样点", "征：然后将其用于产生有关铺锚点是否为对象的置信度", "的预测偏移。我们注意到，这种细化不会面临边界框回", "得分，以及通过边界框回归过程的细化边界框。细化的", "归参数之间的比例差异问题，因为在RepPoints的细", "边界框表示为“bbox建议（S1）\"", "化过程中，偏移量处于相同的比例", "在第二阶段，通常通过Rol-pooling「10l或RoI-Align", "[13]从细化的边界框建议中提取细化的对象特征。对于", "Converting RepPoints to bounding box要在 Rep-", "Points的训练中利用边界框注释，以及评估基于Rep-", "两阶段框架33：改进的特征将通过边界框回归产生", "最终的边界框目标。对于多阶段方法[2]，也通过边界", "Point的对象检测器，需要一种将RepPoints转换为边", "框回归，使用改进的特征来生成中间改进的边界框建", "界框的方法。我们使用预定义的转换函数T：Rp一", "议（S2）。在生成最终的边界框目标之前，可以迭代此", "Bp执行此转换，其中Rp表示对象P的RepPoints,", "T(Rp)表示伪盒", "步骤多次", "为此目的考虑了三个转换函数：", "在此框架中，边界框回归在逐步完善对象本地化和", "对象特征方面起着核心作用。我们在下面的段落中阐", ".T=Ti:Min-maxfunction.Min-max operation", "述了边界棍回问的过程", "作为奇能的对蒙", "over both axes are performed over the RepPoints", "to determine Bp , equivalent to the bounding box", "over the sample points.", "BoundingBoxRegression，常规上，预测4-d回归", "向量（△ap,△yp,△wp，△hp）将当前边界框建议Bp=", ".T =T2 : Partial min-max function. Min-max op-", "(ap,9p,Wp,hp）映射到细化的边界框B,中，其中", "eration over a subset of the sample points is per-", "formed over both axes to obtain the rectangular", "B, = (ap + WpAap, Yp +hpAyp, Wpewp, hpe-hp). (2)", "box B,.", "<NEXTHLINE>。具体来说，在对象检测器的训练中，", ". T = T3 : Moment-based function. The mean value", "我们使用预测的4-d回归向量和预期的4-d回归向量", "and the standard deviation of the RepPoints is", "F(Bp,Bt）之间的距离作为学习目标，使用平滑的l1损", "used to computethecenterpointandscaleof the", "失：", "rectangular box B, where the scale is multiplied", "( 5) (二二,,", "by globally-shared learnable multipliers Ar and Ay", "ht", ",1og", ".(3)", "Wp,hp", "这些函数都是可微分的，在插人对象检测系统时可以", "此边界框回归过程广泛用于现有的对象检测方法中。", "实现端到端学习。在我们的实验中，我们发现它们工作", "当所需的细化较小时，它在实践中表现良好，但是当初", "得相当好", "始表示和自标之间的距离较大时，它往往表现不佳。另", "LearningRepPointsRepPoints的学习是由对象定", "个问题在于△r,△y和△w,△h之间的比例差异，这", "位损失和对象识别损失驱动的。为了计算对象定位损", "需要调整其损失权重以实现最佳性能", "失，我们首先使用前面讨论的转换函数T将RepPoints", "转换为伪盒。然后，计算转换后的伪框和地面实况边界", "3.2. RepPoints", "框之间的差。在我们的系统中，我们使用左上点和右下", "如前所述，四维边界框是对象位置的粗略表示。边", "点之间的平滑l1距离来表示本地化损失。这个平滑的", "界框表示仅考虑对象的矩形空间范围，而不考虑形状", "l1距离不需要像计算边界框回归向量之间的距离那样", "3" ]

{ "arxiv_id": "1904.11490", "language": "zh", "page_idx": 2 }

[ [ "0.12177", "0.0885", "0.4836", "0.0885", "0.4836", "0.0986", "0.12177", "0.0986" ], [ "0.5146", "0.0873", "0.881", "0.0873", "0.881", "0.10095", "0.5146", "0.10095" ], [ "0.1184", "0.10156", "0.4702", "0.1027", "0.4702", "0.1164", "0.1184", "0.1152" ], [ "0.532", "0.1051", "0.879", "0.1051", "0.879", "0.1152", "0.532", "0.1152" ], [ "0.12177", "0.12054", "0.4802", "0.12054", "0.4802", "0.1312", "0.12177", "0.1312" ], [ "0.5205", "0.12054", "0.879", "0.12054", "0.879", "0.1312", "0.5205", "0.1312" ], [ "0.12006", "0.1354", "0.4768", "0.1354", "0.4768", "0.1484", "0.12006", "0.1484" ], [ "0.531", "0.1354", "0.8774", "0.1354", "0.8774", "0.1484", "0.531", "0.1484" ], [ "0.12006", "0.152", "0.4844", "0.152", "0.4844", "0.165", "0.12006", "0.165" ], [ "0.5146", "0.152", "0.879", "0.152", "0.879", "0.1621", "0.5146", "0.1621" ], [ "0.1184", "0.1675", "0.486", "0.1675", "0.486", "0.1812", "0.1184", "0.1812" ], [ "0.5146", "0.1675", "0.881", "0.1675", "0.881", "0.1812", "0.5146", "0.1812" ], [ "0.12006", "0.1829", "0.4878", "0.1829", "0.4878", "0.1965", "0.12006", "0.1965" ], [ "0.513", "0.1841", "0.876", "0.1841", "0.876", "0.1948", "0.513", "0.1948" ], [ "0.12006", "0.2007", "0.486", "0.2007", "0.486", "0.2114", "0.12006", "0.2114" ], [ "0.511", "0.199", "0.879", "0.199", "0.879", "0.212", "0.511", "0.212" ], [ "0.12006", "0.2162", "0.2091", "0.2162", "0.2091", "0.228", "0.12006", "0.228" ], [ "0.513", "0.2156", "0.881", "0.2156", "0.881", "0.2286", "0.513", "0.2286" ], [ "0.1377", "0.2322", "0.4878", "0.2322", "0.4878", "0.2452", "0.1377", "0.2452" ], [ "0.513", "0.2322", "0.882", "0.2322", "0.882", "0.2452", "0.513", "0.2452" ], [ "0.1184", "0.2488", "0.4878", "0.2488", "0.4878", "0.2625", "0.1184", "0.2625" ], [ "0.513", "0.2477", "0.881", "0.2477", "0.881", "0.2612", "0.513", "0.2612" ], [ "0.1184", "0.2642", "0.4895", "0.2642", "0.4895", "0.2778", "0.1184", "0.2778" ], [ "0.511", "0.2642", "0.881", "0.2642", "0.881", "0.2778", "0.511", "0.2778" ], [ "0.1184", "0.2808", "0.4895", "0.2808", "0.4895", "0.2944", "0.1184", "0.2944" ], [ "0.513", "0.2798", "0.881", "0.2798", "0.881", "0.2935", "0.513", "0.2935" ], [ "0.1184", "0.2964", "0.4878", "0.2964", "0.4878", "0.31", "0.1184", "0.31" ], [ "0.513", "0.2957", "0.881", "0.2957", "0.881", "0.3088", "0.513", "0.3088" ], [ "0.1184", "0.3123", "0.4937", "0.3123", "0.4937", "0.326", "0.1184", "0.326" ], [ "0.513", "0.3123", "0.881", "0.3123", "0.881", "0.326", "0.513", "0.326" ], [ "0.12006", "0.33", "0.4895", "0.33", "0.4895", "0.3403", "0.12006", "0.3403" ], [ "0.511", "0.3267", "0.879", "0.3267", "0.879", "0.3403", "0.511", "0.3403" ], [ "0.1184", "0.3457", "0.4878", "0.3457", "0.4878", "0.3594", "0.1184", "0.3594" ], [ "0.513", "0.3445", "0.881", "0.3445", "0.881", "0.3582", "0.513", "0.3582" ], [ "0.12177", "0.361", "0.4844", "0.361", "0.4844", "0.3748", "0.12177", "0.3748" ], [ "0.511", "0.3599", "0.881", "0.3599", "0.881", "0.3735", "0.511", "0.3735" ], [ "0.1184", "0.3777", "0.482", "0.3777", "0.482", "0.3884", "0.1184", "0.3884" ], [ "0.513", "0.376", "0.882", "0.3772", "0.882", "0.3901", "0.513", "0.3896" ], [ "0.12006", "0.3938", "0.4844", "0.3938", "0.4844", "0.4067", "0.12006", "0.4067" ], [ "0.513", "0.3938", "0.881", "0.3938", "0.881", "0.4067", "0.513", "0.4067" ], [ "0.12006", "0.4104", "0.4844", "0.4104", "0.4844", "0.4233", "0.12006", "0.4233" ], [ "0.513", "0.4092", "0.881", "0.4092", "0.881", "0.4229", "0.513", "0.4229" ], [ "0.1184", "0.4258", "0.4844", "0.4258", "0.4844", "0.4395", "0.1184", "0.4395" ], [ "0.519", "0.4287", "0.529", "0.4287", "0.529", "0.434", "0.519", "0.434" ], [ "0.5283", "0.4246", "0.881", "0.4246", "0.881", "0.4382", "0.5283", "0.4382" ], [ "0.12006", "0.4436", "0.486", "0.4436", "0.486", "0.4536", "0.12006", "0.4536" ], [ "0.511", "0.4407", "0.881", "0.4412", "0.881", "0.4548", "0.511", "0.4536" ], [ "0.1184", "0.4578", "0.486", "0.4578", "0.486", "0.4714", "0.1184", "0.4714" ], [ "0.513", "0.4578", "0.615", "0.4578", "0.615", "0.4697", "0.513", "0.4697" ], [ "0.12006", "0.4739", "0.4844", "0.4739", "0.4844", "0.4868", "0.12006", "0.4868" ], [ "0.12006", "0.4912", "0.4836", "0.4912", "0.4836", "0.502", "0.12006", "0.502" ], [ "0.511", "0.4912", "0.852", "0.4922", "0.852", "0.5073", "0.511", "0.506" ], [ "0.12006", "0.5083", "0.4836", "0.5083", "0.4836", "0.5186", "0.12006", "0.5186" ], [ "0.1184", "0.5225", "0.486", "0.5225", "0.486", "0.536", "0.1184", "0.536" ], [ "0.513", "0.5195", "0.881", "0.5195", "0.881", "0.533", "0.513", "0.533" ], [ "0.1184", "0.538", "0.4878", "0.538", "0.4878", "0.552", "0.1184", "0.552" ], [ "0.513", "0.535", "0.881", "0.535", "0.881", "0.549", "0.513", "0.549" ], [ "0.1184", "0.5547", "0.476", "0.5547", "0.476", "0.5684", "0.1184", "0.5684" ], [ "0.513", "0.552", "0.881", "0.552", "0.881", "0.5654", "0.513", "0.5654" ], [ "0.513", "0.5664", "0.881", "0.5684", "0.881", "0.582", "0.513", "0.58" ], [ "0.1142", "0.5845", "0.2603", "0.5845", "0.2603", "0.5986", "0.1142", "0.5986" ], [ "0.5156", "0.585", "0.879", "0.585", "0.879", "0.5957", "0.5156", "0.5957" ], [ "0.513", "0.5986", "0.881", "0.5986", "0.881", "0.6133", "0.513", "0.6133" ], [ "0.1184", "0.6123", "0.4844", "0.6123", "0.4844", "0.626", "0.1184", "0.626" ], [ "0.511", "0.615", "0.881", "0.615", "0.881", "0.629", "0.511", "0.629" ], [ "0.1184", "0.6274", "0.4878", "0.6274", "0.4878", "0.641", "0.1184", "0.641" ], [ "0.513", "0.63", "0.881", "0.6313", "0.881", "0.6455", "0.513", "0.6445" ], [ "0.1184", "0.6445", "0.4844", "0.6445", "0.4844", "0.6577", "0.1184", "0.6577" ], [ "0.511", "0.6465", "0.879", "0.6465", "0.879", "0.66", "0.511", "0.66" ], [ "0.1184", "0.661", "0.4844", "0.661", "0.4844", "0.675", "0.1184", "0.675" ], [ "0.513", "0.6646", "0.882", "0.6646", "0.882", "0.6777", "0.513", "0.6777" ], [ "0.1184", "0.677", "0.4878", "0.677", "0.4878", "0.6904", "0.1184", "0.6904" ], [ "0.5146", "0.6797", "0.881", "0.6797", "0.881", "0.6934", "0.5146", "0.6934" ], [ "0.11755", "0.6914", "0.2477", "0.6924", "0.2477", "0.707", "0.11755", "0.706" ], [ "0.513", "0.695", "0.879", "0.697", "0.879", "0.7104", "0.513", "0.708" ], [ "0.5146", "0.712", "0.881", "0.712", "0.881", "0.7256", "0.5146", "0.7256" ], [ "0.11584", "0.7217", "0.3477", "0.7217", "0.3477", "0.7354", "0.11584", "0.7354" ], [ "0.513", "0.727", "0.881", "0.728", "0.881", "0.742", "0.513", "0.741" ], [ "0.1184", "0.7446", "0.4878", "0.7446", "0.4878", "0.7583", "0.1184", "0.7583" ], [ "0.511", "0.7446", "0.881", "0.7446", "0.881", "0.7583", "0.511", "0.7583" ], [ "0.1184", "0.759", "0.486", "0.759", "0.486", "0.7725", "0.1184", "0.7725" ], [ "0.513", "0.7603", "0.881", "0.7603", "0.881", "0.774", "0.513", "0.774" ], [ "0.11755", "0.777", "0.4844", "0.7754", "0.4844", "0.789", "0.11755", "0.7905" ], [ "0.511", "0.777", "0.881", "0.777", "0.881", "0.7905", "0.511", "0.7905" ], [ "0.1184", "0.792", "0.486", "0.792", "0.486", "0.8057", "0.1184", "0.8057" ], [ "0.511", "0.7915", "0.879", "0.7915", "0.879", "0.8047", "0.511", "0.8047" ], [ "0.12006", "0.8086", "0.486", "0.8086", "0.486", "0.8223", "0.12006", "0.8223" ], [ "0.511", "0.808", "0.879", "0.808", "0.879", "0.822", "0.511", "0.822" ], [ "0.12006", "0.826", "0.4844", "0.826", "0.4844", "0.839", "0.12006", "0.839" ], [ "0.5156", "0.8267", "0.879", "0.8267", "0.879", "0.8374", "0.5156", "0.8374" ], [ "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" ], [ "0.12177", "0.8594", "0.4844", "0.8594", "0.4844", "0.869", "0.12177", "0.869" ], [ "0.5146", "0.8594", "0.879", "0.8594", "0.879", "0.869", "0.5146", "0.869" ], [ "0.1184", "0.8745", "0.4836", "0.8745", "0.4836", "0.8853", "0.1184", "0.8853" ], [ "0.5146", "0.8745", "0.879", "0.8745", "0.879", "0.8853", "0.5146", "0.8853" ], [ "0.1184", "0.8896", "0.4844", "0.8896", "0.4844", "0.903", "0.1184", "0.903" ], [ "0.511", "0.8896", "0.879", "0.8896", "0.879", "0.903", "0.511", "0.903" ], [ "0.12177", "0.9067", "0.4878", "0.9067", "0.4878", "0.9175", "0.12177", "0.9175" ], [ "0.511", "0.905", "0.881", "0.905", "0.881", "0.9185", "0.511", "0.9185" ] ]

[ "标准（Guidotti，2022）的指标。反事实解释的期", "后，大量的编辑器旨在生成对抗性示例，其预", "望特征之一是，它们构成了对输入样本的最", "用例是识别分类器的漏洞并公开它们对抗模", "更改，而极小性是用于测量原始样本与编辑样", "型可能与其他反事实编辑器不同，因为它们不", "本之间距离的度量。由于缺乏理想的解释来月", "定旨在生成原始输入的最小或流畅的编辑", "作地面实况，因此无法知道极小值的特定值是", "因此编辑可能包括添加噪声等。在TextAttack", "否是最佳的（通常是好是坏）。给定一个基本", "框架（Morrisetal.，2020）中实现了NLP的对", "事实解释，可以计算要获得的最佳极小值，但", "抗性示例生成器的集合：包括TextFooler（Jir", "是如果没有它，则只能比较不同编辑/编辑器", "et al.,2020)和 Bert-Attack (Li et al., 2020)。这", "中的极小值", "种方法直观上更简单的形式涉及在实例上使用", "使用我们提出的方法，我们引人了一个度量", "梯度下降来生成更改预测器类的示例，同时优", "（我们称之为不一致），它使用编辑器以前的输", "化一个或多个度量（Mothilaletal.2020）的值。", "出作为参考点来评估编辑器产生最小编辑的能", "其他编辑器不会尝试随机排列来生成反例，而", "力。我们将编辑器的输出作为输人反馈给它，", "是仅更改每个文本的重要特征。这种重要性的", "以产生新的编辑，并且我们希望新的编辑至", "计算方法有多种，包括训练分类器提取每个术", "少与上一步的编辑一样好。例如，在图？？中，", "语与任务（WangandCulotta，2021）的相关性，", "我们通过反馈循环方法的不同步骤看到输出，", "测量特征删除对分类器(Jinetal.,2020)预测", "因此当我们将第一个编辑（“这部电影太糟糕", "的影响，或者使用预测器的注意力(Rossetal.", "了！”）反馈给反事实系统时，我们希望生成的", "2020）。然后可以将重要术语替换为同义词，反", "编辑至少与原始文本一样好（“这部电影太棒", "义词，其他任务中的重要术语或使用预先训练", "了！”）。但是，我们看到编辑器在生成的编辑", "的 seq2seq模型(Madaan et al.,2021;Ross et al.,", "中添加了错误的空白（参见图？？中的2：秒编", "2021, 2020; Wu et al., 2021; Fern and Pope, 2021)", "辑），因此我们知道存在系统无法找到的更好", "", "在我们的实验中，我们使用了具有不同预期", "的输出，因此可以肯定的是系统没有产生最佳", "用例和内部逻辑的编辑器，即MiCE，Polyjuice", "输出。请注意，具有非零值不一致性的反事实", "和TextFooler", "系统保证是次优的，但是零值并不表示它是最", "优的。这种方法基本上为编辑器设置了一个下", "2.2Evaluation of Counterfactual Editors", "限，但我们无法访问可能无法自动定义的较高", "边界。在本文的其余部分中，我们将详细介绍", "评估编辑器的实用标准衡量预测器的输出翻转", "我们提出的方法和新颖的度量标准，并演示其", "到所需类别的频率，称为翻转率，有效性，保", "在具有不同特征的几种常用编辑器上的应用。", "真度或攻击成功率。其他指标与生成文本的质", "量有关，包括流畅性，例如在MiCE，Polyjuice", "2Background", "和CAT（Chemmengathetal.，2021）中使用的流", "畅性，在Textattack中定义的语法和语义以及", "在本文中，我们关注的是尝试最小程度地编", "在 counterfactualGAN(Robeer et al.,2021)中定", "辑给定文本输入以更改分类器预测的系统；此", "义的可感知性。这些指标依赖于语言模型的使", "后，我们将此类系统称为反事实编辑器，或简", "用，或者比较原始和编辑文本之间的屏蔽语言", "称为编辑器。下面，我们提供了此类系统的分", "模型损失，或者计算原始和编辑文本之间的语", "类以及相关文献，此外还概述了它们在相关工", "义相似度。(Verma et al.,2020;Keane et al.,2021)", "作中的评估方式", "中描述的接近度是指已编辑数据和原始数据之", "间的通用距离度量。对于自然语言处理，使", "2.1 Counterfactual Editors", "用的距离度量通常是单词级别Levenshtein编", "编辑器的预期用例各不相同，它们用于实现", "辑距离（Levenshteinetal.，1966），也称为极小", "其目标的方法也各不相同。例如，MiCE(Ross", "值（MiCE），接近度（Polyjuice）和编辑距离", "et al., 2020) 和 DoCoGen (Calderon et al., 2022)", "（CAT）。有更多的标准已被用于反事实的评", "等方法通过伪随机屏蔽文本中的单词并优化建", "估，如稀疏性（Keaneetal.，2021）指的是被改", "议的替换以更改9的输出，直接在给定预测器", "变的特征的数量，以及接近训练数据（Verma", "gO的输出上优化其编辑。相反，诸如Polyjuice", "etal.，2020）或相对距离（Keaneetal.，2021）涉", "(Wuetal.,2021)之类的编辑器旨在识别可以改", "及将解释与来自训练数据的实例进行比较。最", "变句子语义的通用文本扰动，而无需针对特定", "后，一种用于评估反事实解释的最新方法涉", "的预测器。他们将其视为通用的反事实，因为", "及测量解释帮助学生模型（Pruthietal.，2022）", "它们可用于更广泛的目的，从数据增强到产生", "或人类(Doshi-Velez andKim,2017Treviso and", "反事实解释或以特定任务/数据集为条件。最", "Martins，2020）学习模拟黑匣子教师模型的程" ]

{ "arxiv_id": "2305.17055", "language": "zh", "page_idx": 2 }

[ [ "0.1184", "0.0873", "0.3938", "0.0873", "0.3938", "0.10095", "0.1184", "0.10095" ], [ "0.513", "0.0873", "0.881", "0.0873", "0.881", "0.10095", "0.513", "0.10095" ], [ "0.1368", "0.1027", "0.4878", "0.1027", "0.4878", "0.1164", "0.1368", "0.1164" ], [ "0.511", "0.1027", "0.882", "0.1027", "0.882", "0.1164", "0.511", "0.1164" ], [ "0.1184", "0.11816", "0.4844", "0.11816", "0.4844", "0.1318", "0.1184", "0.1318" ], [ "0.51", "0.11755", "0.881", "0.11816", "0.881", "0.133", "0.51", "0.1318" ], [ "0.11755", "0.1342", "0.4836", "0.1342", "0.4836", "0.1478", "0.11755", "0.1478" ], [ "0.513", "0.1354", "0.882", "0.1354", "0.882", "0.1484", "0.513", "0.1484" ], [ "0.1184", "0.152", "0.486", "0.152", "0.486", "0.165", "0.1184", "0.165" ], [ "0.5146", "0.1526", "0.879", "0.1526", "0.879", "0.1633", "0.5146", "0.1633" ], [ "0.1184", "0.1675", "0.4844", "0.1675", "0.4844", "0.1812", "0.1184", "0.1812" ], [ "0.513", "0.1675", "0.881", "0.1675", "0.881", "0.1812", "0.513", "0.1812" ], [ "0.12427", "0.1829", "0.4844", "0.1829", "0.4844", "0.1965", "0.12427", "0.1965" ], [ "0.511", "0.1829", "0.881", "0.1829", "0.881", "0.1965", "0.511", "0.1965" ], [ "0.1184", "0.1996", "0.4844", "0.1996", "0.4844", "0.2131", "0.1184", "0.2131" ], [ "0.51", "0.199", "0.879", "0.199", "0.879", "0.212", "0.51", "0.212" ], [ "0.1184", "0.2156", "0.4878", "0.2156", "0.4878", "0.2286", "0.1184", "0.2286" ], [ "0.513", "0.2162", "0.607", "0.2162", "0.607", "0.228", "0.513", "0.228" ], [ "0.1184", "0.2322", "0.486", "0.2322", "0.486", "0.2452", "0.1184", "0.2452" ], [ "0.11755", "0.2477", "0.486", "0.2477", "0.486", "0.2612", "0.11755", "0.2612" ], [ "0.509", "0.2446", "0.875", "0.2446", "0.875", "0.2583", "0.509", "0.2583" ], [ "0.11755", "0.2625", "0.486", "0.2632", "0.486", "0.2778", "0.11755", "0.2766" ], [ "0.858", "0.2612", "0.881", "0.2612", "0.881", "0.2737", "0.858", "0.2737" ], [ "0.1184", "0.2798", "0.4878", "0.2798", "0.4878", "0.2935", "0.1184", "0.2935" ], [ "0.531", "0.2766", "0.881", "0.2766", "0.881", "0.2903", "0.531", "0.2903" ], [ "0.1184", "0.2976", "0.1755", "0.2976", "0.1755", "0.3088", "0.1184", "0.3088" ], [ "0.513", "0.2922", "0.881", "0.2922", "0.881", "0.306", "0.513", "0.306" ], [ "0.5156", "0.31", "0.879", "0.31", "0.879", "0.3206", "0.5156", "0.3206" ], [ "0.11755", "0.3237", "0.157", "0.3237", "0.157", "0.335", "0.11755", "0.335" ], [ "0.1503", "0.3237", "0.4802", "0.3237", "0.4802", "0.3374", "0.1503", "0.3374" ], [ "0.5146", "0.3247", "0.881", "0.3247", "0.881", "0.3384", "0.5146", "0.3384" ], [ "0.5283", "0.3425", "0.879", "0.3425", "0.879", "0.3528", "0.5283", "0.3528" ], [ "0.1184", "0.3486", "0.4844", "0.3486", "0.4844", "0.3623", "0.1184", "0.3623" ], [ "0.513", "0.3552", "0.881", "0.3557", "0.881", "0.3706", "0.513", "0.3694" ], [ "0.11584", "0.3635", "0.486", "0.3623", "0.486", "0.3801", "0.11584", "0.3813" ], [ "0.513", "0.3735", "0.881", "0.3735", "0.881", "0.3872", "0.513", "0.3872" ], [ "0.1184", "0.3813", "0.4844", "0.3813", "0.4844", "0.3943", "0.1184", "0.3943" ], [ "0.511", "0.3896", "0.881", "0.3896", "0.881", "0.4026", "0.511", "0.4026" ], [ "0.11755", "0.3967", "0.486", "0.3955", "0.486", "0.4104", "0.11755", "0.411" ], [ "0.513", "0.405", "0.879", "0.405", "0.879", "0.4187", "0.513", "0.4187" ], [ "0.1184", "0.4133", "0.3855", "0.4133", "0.3855", "0.427", "0.1184", "0.427" ], [ "0.513", "0.4204", "0.881", "0.4204", "0.881", "0.434", "0.513", "0.434" ], [ "0.1318", "0.4407", "0.4878", "0.4407", "0.4878", "0.4536", "0.1318", "0.4536" ], [ "0.513", "0.437", "0.8774", "0.437", "0.8774", "0.4507", "0.513", "0.4507" ], [ "0.1528", "0.4573", "0.413", "0.4573", "0.413", "0.4702", "0.1528", "0.4702" ], [ "0.513", "0.4531", "0.881", "0.4531", "0.881", "0.466", "0.513", "0.466" ], [ "0.5146", "0.4697", "0.879", "0.4697", "0.879", "0.4827", "0.5146", "0.4827" ], [ "0.1301", "0.4822", "0.4937", "0.4822", "0.4937", "0.4954", "0.1301", "0.4954" ], [ "0.513", "0.485", "0.8774", "0.485", "0.8774", "0.4988", "0.513", "0.4988" ], [ "0.1545", "0.4995", "0.4895", "0.4995", "0.4895", "0.513", "0.1545", "0.513" ], [ "0.513", "0.502", "0.7407", "0.502", "0.7407", "0.5156", "0.513", "0.5156" ], [ "0.1528", "0.5156", "0.4844", "0.514", "0.4844", "0.5293", "0.1528", "0.53" ], [ "0.636", "0.5156", "0.894", "0.535", "0.888", "0.579", "0.6304", "0.558" ], [ "0.1528", "0.53", "0.4844", "0.53", "0.4844", "0.5435", "0.1528", "0.5435" ], [ "0.531", "0.542", "0.638", "0.542", "0.638", "0.554", "0.531", "0.554" ], [ "0.6465", "0.5435", "0.6978", "0.557", "0.687", "0.5767", "0.636", "0.5635" ], [ "0.1293", "0.572", "0.486", "0.5737", "0.486", "0.5884", "0.1293", "0.586" ], [ "0.1528", "0.5474", "0.4148", "0.5464", "0.4148", "0.561", "0.1528", "0.5625" ], [ "0.511", "0.5806", "0.8228", "0.582", "0.8228", "0.5967", "0.511", "0.5957" ], [ "0.1545", "0.5874", "0.2854", "0.5874", "0.2854", "0.6016", "0.1545", "0.6016" ], [ "0.513", "0.6094", "0.804", "0.6094", "0.804", "0.6226", "0.513", "0.6226" ], [ "0.1368", "0.616", "0.4844", "0.616", "0.4844", "0.63", "0.1368", "0.63" ], [ "0.1184", "0.633", "0.4902", "0.633", "0.4902", "0.6465", "0.1184", "0.6465" ], [ "0.5146", "0.6313", "0.879", "0.6313", "0.879", "0.641", "0.5146", "0.641" ], [ "0.1184", "0.6484", "0.4844", "0.6484", "0.4844", "0.662", "0.1184", "0.662" ], [ "0.513", "0.6455", "0.881", "0.6455", "0.881", "0.659", "0.513", "0.659" ], [ "0.1184", "0.6646", "0.4878", "0.6646", "0.4878", "0.6777", "0.1184", "0.6777" ], [ "0.513", "0.661", "0.881", "0.661", "0.881", "0.675", "0.513", "0.675" ], [ "0.1184", "0.681", "0.4878", "0.681", "0.4878", "0.695", "0.1184", "0.695" ], [ "0.511", "0.6777", "0.881", "0.6777", "0.881", "0.6914", "0.511", "0.6914" ], [ "0.1184", "0.6978", "0.4878", "0.6978", "0.4878", "0.7114", "0.1184", "0.7114" ], [ "0.513", "0.6934", "0.881", "0.6934", "0.881", "0.707", "0.513", "0.707" ], [ "0.1184", "0.7134", "0.4878", "0.7134", "0.4878", "0.727", "0.1184", "0.727" ], [ "0.513", "0.7104", "0.881", "0.7104", "0.881", "0.7236", "0.513", "0.7236" ], [ "0.1184", "0.73", "0.486", "0.73", "0.486", "0.7437", "0.1184", "0.7437" ], [ "0.513", "0.7256", "0.881", "0.7256", "0.881", "0.7393", "0.513", "0.7393" ], [ "0.513", "0.7407", "0.881", "0.741", "0.881", "0.756", "0.513", "0.755" ], [ "0.1184", "0.7466", "0.2167", "0.7466", "0.2167", "0.7583", "0.1184", "0.7583" ], [ "0.513", "0.7583", "0.882", "0.7583", "0.882", "0.7725", "0.513", "0.7725" ], [ "0.11584", "0.7715", "0.392", "0.7715", "0.392", "0.785", "0.11584", "0.785" ], [ "0.5146", "0.7754", "0.594", "0.7754", "0.594", "0.7876", "0.5146", "0.7876" ], [ "0.12006", "0.792", "0.4844", "0.792", "0.4844", "0.8057", "0.12006", "0.8057" ], [ "0.519", "0.7974", "0.712", "0.818", "0.706", "0.8496", "0.513", "0.83" ], [ "0.1184", "0.808", "0.486", "0.808", "0.486", "0.822", "0.1184", "0.822" ], [ "0.6514", "0.804", "0.6826", "0.8086", "0.678", "0.821", "0.6465", "0.8154" ], [ "0.6724", "0.811", "0.831", "0.8135", "0.831", "0.8306", "0.6724", "0.829" ], [ "0.1184", "0.8247", "0.4844", "0.8247", "0.4844", "0.8384", "0.1184", "0.8384" ], [ "0.1184", "0.8413", "0.4844", "0.8413", "0.4844", "0.855", "0.1184", "0.855" ], [ "0.8564", "0.8413", "0.879", "0.8413", "0.879", "0.854", "0.8564", "0.854" ], [ "0.1184", "0.857", "0.4878", "0.857", "0.4878", "0.8706", "0.1184", "0.8706" ], [ "0.531", "0.857", "0.879", "0.857", "0.879", "0.8706", "0.531", "0.8706" ], [ "0.1184", "0.8745", "0.4844", "0.8745", "0.4844", "0.8853", "0.1184", "0.8853" ], [ "0.511", "0.8735", "0.882", "0.8735", "0.882", "0.887", "0.511", "0.887" ], [ "0.1184", "0.8896", "0.4895", "0.8896", "0.4895", "0.903", "0.1184", "0.903" ], [ "0.513", "0.8896", "0.881", "0.8896", "0.881", "0.903", "0.513", "0.903" ], [ "0.12006", "0.9067", "0.4844", "0.9067", "0.4844", "0.9175", "0.12006", "0.9175" ], [ "0.511", "0.9062", "0.881", "0.9062", "0.881", "0.9194", "0.511", "0.9194" ] ]

[ "度。这提供了解释的信息量的度量", "此，在这种情况下，A可以用作地面实况的代", "在这项工作中，我们专注于使用自动化指标", "理，以与C进行比较。给定一个距离度量d", "进行评估，这些指标不需要人工输人或外部数", "（如Levenshtein距离、嵌人余弦相似度等。），", "据，并且在编辑器评估中最常用，即极简，翻", "我们可以衡量反事实编辑器的一致性W。r。t", "转率和流畅性。这些指标在以比较的方式", "d通过选代地将编辑后的文本反馈给编辑器并", "量化编辑行为的不同方面。相反，我们提出了", "测量d值的变化。具体来说，给定一个编辑", "一个不一致性度量，它允许设置编辑器达到最", "器f：L→L，一个文本EL和一个距离", "佳反事实的能力的下限，因此可以很有用，而", "d：L×L→R+，我们将f相对于d的不一致", "无需与其他编辑器进行比较。例如，给定编辑", "性定义为：。", "器的极小值本身没有关于最优性的信息。但", "是，如果编辑器具有用于极小值不一致的值", "inc(f,α) =relu[d(f(f(α)), f(α)) - d(f(α), α)", "inc@1=a，则意味着在给定的测试集上，它", "(1)", "至少以&令牌的平均值错过了最佳反事实解", "差异d(f(f(α)),f(c))一d(f(α),a)显示了编", "决方案", "辑器f和relu函数的连续应用程序之间的距", "离d变化了多少，充许仅考虑距离的增加。这", "31", "Back-translation for analyzing editors", "很重要，因为距离的减小（对应于负差异）不", "定是一组良好编辑的指标。例如，它可能表", "我们将问题形式化如下。我们假设访问分类器", "明没有进行足够的更改，并且无法知道是否是", "g，使得g：C→[0,1]°，其中C是特定语言", "这种情况，或者是否找到了更好，更最少的编", "的文本集，C是不同类的数量。然后，我们将", "辑集。相反，当值为正时，我们可以保证存在", "9的反事实编辑器视为函数f：→L，并且", "更好的一组编辑，即先前的反馈步骤之一。公", "我们假设编辑器f的目标有三个：", "式1计算通过编辑器进行一次反馈迭代后d中", "1. The edited text is classified to a different class", "的差异，但是与这项工作中的其他指标一样", "argmaxg(f()) + argmaxg() ", "我们可以不断将编辑器的输出反馈给自己，并", "计算inc(f,f(α)）以获取有关编辑器不一致的", "2. Theeditsareminimalwithre-", "更多信息。当我们这样做时，我们将n反馈步", "spect to some distance metric d :", "骤后的平均不一致性测量为：", "f =argminheFd(a,h(r)) ， where", "me(/(),(), 2)", "F is the set of functions for which", "inc@n(f,α)=", "ni=0", "3. The edited text f(c) is fluent and within the", "argmax g(f(c)) + argmax g(α) .", "，其中fo(c)=和f(α)=f(fi-i(a))", "distribution of C.", "3.2Diverging from the distribution", "为了检查这些标准的保持程度，我们分析了", "编辑器在迭代反馈其输出时的行为，即。e。，", "在反事实编辑器的需求中，f（e）流畅且在分发", "我们正在研究函数f(f(f(...f(αc)))），并在编", "范围内的约束可能很难验证，因为文本L的", "辑器的n应用程序之后评估上述三个条件。具", "真实分发可能无法访问，并且流畅性很难以系", "体来说，我们首先定义一个新的评估指标来", "统和自动化的方式进行评估。大型语言模型的", "量化基于选代反馈方法的第二个标准，然后我", "令牌级困惑是流利度估计的常用代理，用于多", "们讨论如何通过测量n步反馈后的性能指标", "个NLP 任务 (John et al.,2018;Wang and Cho,", "来更彻底地检查第一个和第三个标准（标记为", "2019；Vasilakesetal.,2022）。它涉及使用在大", "型数据集D上训练的语言模型MD，并计算", "metric@n).", "给定文本序列=C1,α2..,T上的平均困惑，", "3.1(In)consistency ofminimality", "如下所示：", "直观上，由于编辑是理想的最小，如果一个", "PPL(a) =exP (≥", "句子A被编辑成句子B，并且它们的距离是", "（1", "log pMp(ct|c1:t-", "d(A,B），那么将句子B反馈给编辑器应该产", "生一个C ，d(B,C)≤ d(A,B)的结果不是最", "(3)", "小的。该不等式基于以下条件成立：（a）我们", "请注意，对于流利度估计，公式3中的MD", "知道A存在：（b）我们假设所有文本编辑都是", "未在L上进行微调。假设L是可访问的，我", "可逆的，henseA可以从B到达，并且（c）d", "们也可以在其上微调MD，从而获得一个模", "是对称的，这意味着d(A,B）=d(BA)。因", "型Mc，我们可以使用该模型来检测非分布" ]

{ "arxiv_id": "2305.17055", "language": "zh", "page_idx": 3 }

[ [ "0.1184", "0.0873", "0.486", "0.0873", "0.486", "0.10095", "0.1184", "0.10095" ], [ "0.51", "0.0873", "0.881", "0.0873", "0.881", "0.10095", "0.51", "0.10095" ], [ "0.1184", "0.1027", "0.4844", "0.1027", "0.4844", "0.1164", "0.1184", "0.1164" ], [ "0.513", "0.10394", "0.5693", "0.10394", "0.5693", "0.1152", "0.513", "0.1152" ], [ "0.1184", "0.1194", "0.486", "0.1194", "0.486", "0.133", "0.1184", "0.133" ], [ "0.51", "0.1152", "0.881", "0.1152", "0.881", "0.133", "0.51", "0.133" ], [ "0.1184", "0.1342", "0.4836", "0.1342", "0.4836", "0.1478", "0.1184", "0.1478" ], [ "0.511", "0.1354", "0.882", "0.1354", "0.882", "0.1526", "0.511", "0.1526" ], [ "0.1184", "0.1509", "0.486", "0.1509", "0.486", "0.1644", "0.1184", "0.1644" ], [ "0.513", "0.1526", "0.55", "0.1526", "0.55", "0.1644", "0.513", "0.1644" ], [ "0.1184", "0.1675", "0.486", "0.1675", "0.486", "0.1812", "0.1184", "0.1812" ], [ "0.511", "0.1687", "0.8657", "0.1663", "0.8657", "0.1812", "0.511", "0.1829" ], [ "0.8613", "0.1729", "0.876", "0.1729", "0.876", "0.177", "0.8613", "0.177" ], [ "0.11755", "0.1829", "0.339", "0.1829", "0.339", "0.1965", "0.11755", "0.1965" ], [ "0.5146", "0.1853", "0.8774", "0.1853", "0.8774", "0.1953", "0.5146", "0.1953" ], [ "0.5225", "0.2073", "0.529", "0.2073", "0.529", "0.2102", "0.5225", "0.2102" ], [ "0.529", "0.2031", "0.6187", "0.2031", "0.6187", "0.2131", "0.529", "0.2131" ], [ "0.1142", "0.2131", "0.2334", "0.2144", "0.2334", "0.2286", "0.1142", "0.228" ], [ "0.511", "0.2173", "0.882", "0.2144", "0.882", "0.2286", "0.511", "0.2322" ], [ "0.513", "0.2328", "0.881", "0.2328", "0.881", "0.2465", "0.513", "0.2465" ], [ "0.12006", "0.2405", "0.4836", "0.2405", "0.4836", "0.2505", "0.12006", "0.2505" ], [ "0.11755", "0.2542", "0.486", "0.2542", "0.486", "0.2673", "0.11755", "0.2673" ], [ "0.511", "0.2488", "0.806", "0.2488", "0.806", "0.2625", "0.511", "0.2625" ], [ "0.1184", "0.2708", "0.4844", "0.2708", "0.4844", "0.284", "0.1184", "0.284" ], [ "0.531", "0.2654", "0.881", "0.2654", "0.881", "0.279", "0.531", "0.279" ], [ "0.1184", "0.2861", "0.4844", "0.2861", "0.4844", "0.2998", "0.1184", "0.2998" ], [ "0.513", "0.282", "0.876", "0.282", "0.876", "0.2922", "0.513", "0.2922" ], [ "0.1184", "0.3027", "0.4844", "0.3027", "0.4844", "0.3164", "0.1184", "0.3164" ], [ "0.513", "0.308", "0.5615", "0.308", "0.5615", "0.3196", "0.513", "0.3196" ], [ "0.555", "0.3071", "0.6187", "0.3071", "0.6187", "0.3218", "0.555", "0.3218" ], [ "0.11755", "0.3184", "0.4844", "0.3184", "0.4844", "0.332", "0.11755", "0.332" ], [ "0.11755", "0.3333", "0.486", "0.3342", "0.486", "0.3486", "0.11755", "0.3474" ], [ "0.511", "0.328", "0.8774", "0.328", "0.8774", "0.3416", "0.511", "0.3416" ], [ "0.5146", "0.3445", "0.881", "0.3445", "0.881", "0.3582", "0.5146", "0.3582" ], [ "0.12006", "0.3516", "0.1914", "0.3516", "0.1914", "0.3635", "0.12006", "0.3635" ], [ "0.517", "0.3623", "0.806", "0.3623", "0.806", "0.3723", "0.517", "0.3723" ], [ "0.1142", "0.3818", "0.3257", "0.3843", "0.3257", "0.3984", "0.1142", "0.3967" ], [ "0.519", "0.3855", "0.881", "0.3855", "0.881", "0.3997", "0.519", "0.3997" ], [ "0.5146", "0.402", "0.882", "0.402", "0.882", "0.415", "0.5146", "0.415" ], [ "0.11755", "0.411", "0.4836", "0.411", "0.4836", "0.4246", "0.11755", "0.4246" ], [ "0.5146", "0.4192", "0.879", "0.4192", "0.879", "0.43", "0.5146", "0.43" ], [ "0.1184", "0.4275", "0.4844", "0.4275", "0.4844", "0.4412", "0.1184", "0.4412" ], [ "0.51", "0.434", "0.881", "0.434", "0.881", "0.4478", "0.51", "0.4478" ], [ "0.12006", "0.4436", "0.486", "0.4436", "0.486", "0.4573", "0.12006", "0.4573" ], [ "0.511", "0.4495", "0.881", "0.4495", "0.881", "0.4631", "0.511", "0.4631" ], [ "0.11755", "0.459", "0.4844", "0.4602", "0.4844", "0.4744", "0.11755", "0.4739" ], [ "0.511", "0.466", "0.881", "0.466", "0.881", "0.4797", "0.511", "0.4797" ], [ "0.1184", "0.478", "0.4844", "0.478", "0.4844", "0.488", "0.1184", "0.488" ], [ "0.511", "0.4822", "0.879", "0.4822", "0.879", "0.4954", "0.511", "0.4954" ], [ "0.11755", "0.4922", "0.486", "0.4922", "0.486", "0.506", "0.11755", "0.506" ], [ "0.511", "0.4954", "0.571", "0.4976", "0.5693", "0.513", "0.51", "0.511" ], [ "0.1184", "0.509", "0.1545", "0.509", "0.1545", "0.5215", "0.1184", "0.5215" ], [ "0.1368", "0.5254", "0.486", "0.5254", "0.486", "0.539", "0.1368", "0.539" ], [ "0.5146", "0.5215", "0.881", "0.5225", "0.881", "0.5376", "0.5146", "0.536" ], [ "0.1184", "0.542", "0.4836", "0.542", "0.4836", "0.5527", "0.1184", "0.5527" ], [ "0.511", "0.538", "0.882", "0.539", "0.882", "0.554", "0.511", "0.5527" ], [ "0.11755", "0.557", "0.4836", "0.557", "0.4836", "0.571", "0.11755", "0.571" ], [ "0.511", "0.5557", "0.881", "0.5557", "0.881", "0.5693", "0.511", "0.5693" ], [ "0.1184", "0.5747", "0.4844", "0.5747", "0.4844", "0.5884", "0.1184", "0.5884" ], [ "0.513", "0.572", "0.881", "0.572", "0.881", "0.585", "0.513", "0.585" ], [ "0.12006", "0.5913", "0.4844", "0.5913", "0.4844", "0.6016", "0.12006", "0.6016" ], [ "0.513", "0.5874", "0.881", "0.5874", "0.881", "0.601", "0.513", "0.601" ], [ "0.1184", "0.607", "0.4844", "0.607", "0.4844", "0.6177", "0.1184", "0.6177" ], [ "0.51", "0.6016", "0.8774", "0.6016", "0.8774", "0.615", "0.51", "0.615" ], [ "0.1184", "0.6226", "0.4878", "0.6226", "0.4878", "0.6357", "0.1184", "0.6357" ], [ "0.511", "0.619", "0.881", "0.619", "0.881", "0.633", "0.511", "0.633" ], [ "0.12006", "0.64", "0.4878", "0.64", "0.4878", "0.651", "0.12006", "0.651" ], [ "0.513", "0.6353", "0.881", "0.6353", "0.881", "0.6484", "0.513", "0.6484" ], [ "0.12006", "0.6562", "0.4844", "0.6562", "0.4844", "0.6665", "0.12006", "0.6665" ], [ "0.513", "0.652", "0.882", "0.652", "0.882", "0.665", "0.513", "0.665" ], [ "0.1184", "0.6704", "0.486", "0.6704", "0.486", "0.684", "0.1184", "0.684" ], [ "0.513", "0.6685", "0.881", "0.6685", "0.881", "0.6816", "0.513", "0.6816" ], [ "0.1184", "0.686", "0.2468", "0.686", "0.2468", "0.6997", "0.1184", "0.6997" ], [ "0.5146", "0.685", "0.881", "0.685", "0.881", "0.6987", "0.5146", "0.6987" ], [ "0.511", "0.6997", "0.881", "0.6997", "0.881", "0.7134", "0.511", "0.7134" ], [ "0.1377", "0.706", "0.4836", "0.706", "0.4836", "0.7163", "0.1377", "0.7163" ], [ "0.11755", "0.7197", "0.4836", "0.7197", "0.4836", "0.733", "0.11755", "0.733" ], [ "0.511", "0.7153", "0.714", "0.7153", "0.714", "0.7285", "0.511", "0.7285" ], [ "0.12006", "0.7383", "0.4836", "0.7383", "0.4836", "0.749", "0.12006", "0.749" ], [ "0.519", "0.741", "0.881", "0.741", "0.881", "0.755", "0.519", "0.755" ], [ "0.1184", "0.753", "0.4878", "0.753", "0.4878", "0.7666", "0.1184", "0.7666" ], [ "0.513", "0.7583", "0.8774", "0.7583", "0.8774", "0.7686", "0.513", "0.7686" ], [ "0.11755", "0.767", "0.4878", "0.767", "0.4878", "0.781", "0.11755", "0.781" ], [ "0.513", "0.774", "0.882", "0.774", "0.882", "0.7876", "0.513", "0.7876" ], [ "0.1184", "0.785", "0.486", "0.785", "0.486", "0.799", "0.1184", "0.799" ], [ "0.51", "0.789", "0.881", "0.789", "0.881", "0.8027", "0.51", "0.8027" ], [ "0.1184", "0.8003", "0.4844", "0.8003", "0.4844", "0.814", "0.1184", "0.814" ], [ "0.5146", "0.807", "0.879", "0.807", "0.879", "0.818", "0.5146", "0.818" ], [ "0.11584", "0.8154", "0.4568", "0.8164", "0.4568", "0.8306", "0.11584", "0.83" ], [ "0.513", "0.822", "0.879", "0.822", "0.879", "0.835", "0.513", "0.835" ], [ "0.513", "0.8374", "0.879", "0.8374", "0.879", "0.851", "0.513", "0.851" ], [ "0.11755", "0.847", "0.2494", "0.847", "0.2494", "0.8623", "0.11755", "0.8623" ], [ "0.513", "0.8516", "0.881", "0.8525", "0.881", "0.8677", "0.513", "0.866" ], [ "0.12006", "0.8735", "0.345", "0.8735", "0.345", "0.8843", "0.12006", "0.8843" ], [ "0.5156", "0.869", "0.879", "0.869", "0.879", "0.883", "0.5156", "0.883" ], [ "0.11755", "0.8896", "0.486", "0.8896", "0.486", "0.903", "0.11755", "0.903" ], [ "0.532", "0.891", "0.7944", "0.8926", "0.7944", "0.9067", "0.532", "0.9062" ], [ "0.11584", "0.903", "0.2065", "0.905", "0.2065", "0.9194", "0.11584", "0.9175" ], [ "0.5356", "0.9062", "0.8423", "0.9062", "0.8423", "0.9194", "0.5356", "0.9194" ] ]

[ "(OOD）情况（Aroraetal.，2021）使用相同的", "inc@n：根据公式2，单词级Levenshtein距离", "PPL公式在等式3中描述的概率，因为在这种", "不一致", "情况下，我们可以假设在Mc。在我们的实验", "翻转率：比率nrmipped，其中nall是数据集的大", "中，我们采用并比较了MD上的PPL（PPLD", "小，nflipped是在应用编辑器后预测发生变化的", "）和PPL上的Mc（PPL）。此外，PPLc@n", "样本", "和PPLp@n可以更清楚地显示与生成样本分", "Ppl-base：GPT-2的语言模型困惑，GPT-2是", "", "布的差异以及流畅性的下降", "个大型的通用域语言模型（Radfordetal.2019）", "", "按照公式3", "3.3Fliprate", "ppl-fine：GPT-2的语言模型困惑，在IMDB2", "和新闻组3上进行了微调。用于检查编辑的文", "到自前为正止，我们尚未考虑编辑器更改预测", "类的能力，该能力通常以翻转速率来衡量。许", "本相对于每个数据集的“意外”程度", "多最近提出的反事实编辑器的翻转率都高于", "除了inc@n之外，我们还在n反馈步骤之后", "95%，因此可以放心地将编辑器与此指标进行", "计算上述指标，inc@n根据定义使用反馈步骤", "比较。。使用建议的反馈方法，我们可以测量", "4.2", "Editors", "n反馈循环后的翻转率，以获得有关编辑器在", "句子多次扰动后一致地更改预测类的能力的更", "我们尝试了三种具有不同特征的编辑器。每个", "编辑器的简要说明以及它们之间的主要区别如", "详细信息", "下所示：更多详细信息请参见附录B", "4Experimental Setup", "PolyjuicePolyjuice是一种通用的反事实生", "成器，可根据预定义的控制类型产生扰动。此", "我们在两个不同的数据集和在其上训练的分", "编辑器采用GPT-2模型，该模型在成对句子", "类器上评估我们的方法。具体来说，我们使用", "（输入反事实）的各种数据集上进行了微调，包", "了在IMDb数据集（Maasetal.,2011）上进行情", "括IMDb数据集。Polyjuice在生成反事实文本", "感分类训练的二进制分类器和在新闻组（Lang,", "期间不使用分类器预测，而是基于学习的一组", "1995）上训练的多类短文档分类器。我们在附", "控制代码（例如“否定”或“删除”）来关注各", "录A中提供了每个数据集的详细信息和统计", "种编辑", "信息", "使用这些分类器，我们将我们的方法应用于", "MiCEMiCE是生成反事实编辑的两步方法。", "84.2中描述的三个反事实编辑器，并通过生成", "它使用T5深度神经网络来填充文本上的空白，", "编辑，在其上测试分类器并将编辑后的文本反", "这在第一步中进行了微调，以更好地适应数据", "馈给n=10步骤的编辑器来检查s3中描述的", "集分布。在第二步中，输入文本被随机或通过", "度量。对于每个编辑器，我们应用我们的方法", "使用预测因子注意力以白框方式屏蔽，微调模", "如下：对干给定的输入文本：在步骤n：我", "型填补了这些空白。此步骤旨在学习将改变分", "们从生成的编辑文本池中选择具有最小极小值", "类器预测的最小编辑。在我们的实验中，我们", "的文本，如果存在这样的输出，则会改变预测，", "以白盒方式使用MiCE，这意味着微调是通过", "否则我们选择最小极小值输出。我们重复此过", "使用预测器的输出（而不是地面实况标签）来", "程直到n=10，并讨论85和$6中每个编辑", "完成的，并且为了选择掩码的位置，使用分类", "器跨指标的行为", "器的注意力以便将其结果与Polyjuice进行比", "较，Polyjuice也利用深度神经网络进行生成，", "我们还研究了测试集大小对观察到的结果之", "间的差异以及发现的统计意义的影响。我们在", "但以黑盒方式使用预测器", "附录E中提供了详细的结果。我们发现，对", "TextFoolerTextFooler是为黑盒分类器生成", "于大于200个文本的测试集，结果收敛于两个", "对抗性示例的方法。生成过程与其他编辑器不", "数据集，并且我们获得了统计上的显着差异。", "同，因为它不使用深度神经网络（如GPT2或", "基于这些发现，我们从IMDb数据集中随机抽", "T5）来构建生成的反事实。相反，包括找到对", "取500个文本进行实验，以减少计算量。由于", "预测类有影响的单词，并以更确定的方式替换", "NewsGroups较小，我们使用完整的数据集", "它们。通过测量每个单词的删除如何改变预测", "器的输出（Liangetal.，2017）来计算每个单词", "4.1 Evaluation", "的影响。可以替换单词的替代项被预定义为嵌", "我们在方法中使用的指标是：", "入空间中最接近的匹配项，因此独立于句子的", "极小：原始和编辑文本之间的单词级Leven-", "2https:/huggingface.co/lvwerra/gpt2-imdb", "shtein距离", "https://huggingface.co/QianWeiTech/GPT2-News" ]

{ "arxiv_id": "2305.17055", "language": "zh", "page_idx": 4 }

[ [ "0.3367", "0.1332", "0.6357", "0.1332", "0.6357", "0.1477", "0.3367", "0.1477" ], [ "0.2566", "0.1812", "0.406", "0.1818", "0.406", "0.1995", "0.2566", "0.1989" ], [ "0.4192", "0.1818", "0.5605", "0.1831", "0.5605", "0.1989", "0.4192", "0.1976" ], [ "0.5825", "0.1843", "0.714", "0.1843", "0.714", "0.1976", "0.5825", "0.1976" ], [ "0.4084", "0.1995", "0.5605", "0.2008", "0.5605", "0.2153", "0.4084", "0.214" ], [ "0.3987", "0.1875", "0.4192", "0.1875", "0.4192", "0.1957", "0.3987", "0.1957" ], [ "0.3423", "0.2197", "0.626", "0.2197", "0.626", "0.2343", "0.3423", "0.2343" ], [ "0.2385", "0.2695", "0.3137", "0.2695", "0.3137", "0.2842", "0.2385", "0.2842" ], [ "0.7075", "0.2798", "0.785", "0.2798", "0.785", "0.2917", "0.7075", "0.2917" ], [ "0.0989", "0.3005", "0.4631", "0.3005", "0.4631", "0.3145", "0.0989", "0.3145" ], [ "0.0809", "0.3176", "0.468", "0.3176", "0.468", "0.3308", "0.0809", "0.3308" ], [ "0.7363", "0.3208", "0.7485", "0.3164", "0.755", "0.3276", "0.7427", "0.332" ], [ "0.0801", "0.332", "0.4697", "0.332", "0.4697", "0.3452", "0.0801", "0.3452" ], [ "0.6436", "0.3276", "0.6724", "0.3276", "0.6724", "0.336", "0.6436", "0.336" ], [ "0.0801", "0.3452", "0.4673", "0.3452", "0.4673", "0.3599", "0.0801", "0.3599" ], [ "0.732", "0.3452", "0.763", "0.3452", "0.763", "0.3567", "0.732", "0.3567" ], [ "0.511", "0.3516", "0.5513", "0.3516", "0.5513", "0.363", "0.511", "0.363" ], [ "0.0809", "0.3623", "0.468", "0.3623", "0.468", "0.3757", "0.0809", "0.3757" ], [ "0.0825", "0.379", "0.4673", "0.379", "0.4673", "0.3894", "0.0825", "0.3894" ], [ "0.0801", "0.3914", "0.4656", "0.3914", "0.4656", "0.4048", "0.0801", "0.4048" ], [ "0.0809", "0.4077", "0.468", "0.4077", "0.468", "0.4216", "0.0809", "0.4216" ], [ "0.626", "0.406", "0.687", "0.406", "0.687", "0.4172", "0.626", "0.4172" ], [ "0.709", "0.4092", "0.778", "0.4092", "0.778", "0.4204", "0.709", "0.4204" ], [ "0.6274", "0.4153", "0.6787", "0.4153", "0.6787", "0.4255", "0.6274", "0.4255" ], [ "0.0801", "0.4224", "0.468", "0.4216", "0.468", "0.4363", "0.0801", "0.4375" ], [ "0.0801", "0.4382", "0.468", "0.4382", "0.468", "0.452", "0.0801", "0.452" ], [ "0.5034", "0.4382", "0.8906", "0.4382", "0.8906", "0.452", "0.5034", "0.452" ], [ "0.0825", "0.455", "0.4673", "0.455", "0.4673", "0.4653", "0.0825", "0.4653" ], [ "0.5034", "0.452", "0.8906", "0.452", "0.8906", "0.4653", "0.5034", "0.4653" ], [ "0.0801", "0.4685", "0.4673", "0.4685", "0.4673", "0.4817", "0.0801", "0.4817" ], [ "0.505", "0.4673", "0.886", "0.4673", "0.886", "0.4778", "0.505", "0.4778" ], [ "0.0825", "0.4849", "0.4673", "0.4849", "0.4673", "0.4956", "0.0825", "0.4956" ], [ "0.5024", "0.4792", "0.888", "0.4797", "0.888", "0.4944", "0.5024", "0.4937" ], [ "0.0809", "0.4976", "0.4673", "0.4976", "0.4673", "0.511", "0.0809", "0.511" ], [ "0.5024", "0.4937", "0.892", "0.4937", "0.892", "0.507", "0.5024", "0.507" ], [ "0.0801", "0.511", "0.4673", "0.513", "0.4673", "0.528", "0.0801", "0.526" ], [ "0.5034", "0.5083", "0.8906", "0.5083", "0.8906", "0.5215", "0.5034", "0.5215" ], [ "0.5024", "0.5215", "0.8896", "0.5215", "0.8896", "0.535", "0.5024", "0.535" ], [ "0.0801", "0.529", "0.4673", "0.529", "0.4673", "0.5425", "0.0801", "0.5425" ], [ "0.5034", "0.535", "0.892", "0.535", "0.892", "0.549", "0.5034", "0.549" ], [ "0.0801", "0.5425", "0.4673", "0.5425", "0.4673", "0.556", "0.0801", "0.556" ], [ "0.5034", "0.5503", "0.8906", "0.5503", "0.8906", "0.561", "0.5034", "0.561" ], [ "0.0809", "0.56", "0.4656", "0.56", "0.4656", "0.571", "0.0809", "0.571" ], [ "0.5034", "0.563", "0.892", "0.563", "0.892", "0.577", "0.5034", "0.577" ], [ "0.0801", "0.5737", "0.468", "0.5737", "0.468", "0.587", "0.0801", "0.587" ], [ "0.5024", "0.577", "0.7656", "0.577", "0.7656", "0.5903", "0.5024", "0.5903" ], [ "0.0801", "0.589", "0.468", "0.589", "0.468", "0.603", "0.0801", "0.603" ], [ "0.0809", "0.604", "0.468", "0.604", "0.468", "0.617", "0.0809", "0.617" ], [ "0.505", "0.61", "0.8906", "0.61", "0.8906", "0.6235", "0.505", "0.6235" ], [ "0.0801", "0.619", "0.4673", "0.619", "0.4673", "0.6333", "0.0801", "0.6333" ], [ "0.505", "0.6255", "0.888", "0.6255", "0.888", "0.6396", "0.505", "0.6396" ], [ "0.0809", "0.634", "0.468", "0.634", "0.468", "0.648", "0.0809", "0.648" ], [ "0.505", "0.64", "0.8906", "0.64", "0.8906", "0.654", "0.505", "0.654" ], [ "0.0809", "0.6484", "0.4656", "0.6484", "0.4656", "0.662", "0.0809", "0.662" ], [ "0.5024", "0.655", "0.8906", "0.654", "0.8906", "0.6685", "0.5024", "0.6694" ], [ "0.0801", "0.664", "0.4697", "0.664", "0.4697", "0.678", "0.0801", "0.678" ], [ "0.5034", "0.6704", "0.8906", "0.6704", "0.8906", "0.6846", "0.5034", "0.6846" ], [ "0.0801", "0.6787", "0.468", "0.6787", "0.468", "0.6924", "0.0801", "0.6924" ], [ "0.505", "0.685", "0.8906", "0.685", "0.8906", "0.6987", "0.505", "0.6987" ], [ "0.0784", "0.6934", "0.4617", "0.6943", "0.4617", "0.709", "0.0784", "0.7085" ], [ "0.505", "0.7007", "0.8906", "0.7007", "0.8906", "0.714", "0.505", "0.714" ], [ "0.5034", "0.7163", "0.8906", "0.7163", "0.8906", "0.73", "0.5034", "0.73" ], [ "0.0809", "0.726", "0.2173", "0.726", "0.2173", "0.7393", "0.0809", "0.7393" ], [ "0.5034", "0.731", "0.8906", "0.731", "0.8906", "0.744", "0.5034", "0.744" ], [ "0.1013", "0.7505", "0.468", "0.7505", "0.468", "0.7646", "0.1013", "0.7646" ], [ "0.5034", "0.747", "0.8906", "0.747", "0.8906", "0.7603", "0.5034", "0.7603" ], [ "0.5024", "0.7603", "0.701", "0.761", "0.701", "0.776", "0.5024", "0.7744" ], [ "0.0825", "0.7666", "0.4673", "0.7666", "0.4673", "0.78", "0.0825", "0.78" ], [ "0.0809", "0.781", "0.468", "0.781", "0.468", "0.7944", "0.0809", "0.7944" ], [ "0.5215", "0.78", "0.888", "0.781", "0.888", "0.7954", "0.5215", "0.7944" ], [ "0.0825", "0.7954", "0.4673", "0.7954", "0.4673", "0.8096", "0.0825", "0.8096" ], [ "0.505", "0.797", "0.888", "0.797", "0.888", "0.81", "0.505", "0.81" ], [ "0.08496", "0.811", "0.468", "0.811", "0.468", "0.8247", "0.08496", "0.8247" ], [ "0.5034", "0.811", "0.892", "0.811", "0.892", "0.8247", "0.5034", "0.8247" ], [ "0.0801", "0.8247", "0.468", "0.8257", "0.468", "0.8403", "0.0801", "0.839" ], [ "0.505", "0.827", "0.8906", "0.827", "0.8906", "0.8403", "0.505", "0.8403" ], [ "0.0825", "0.8413", "0.468", "0.8413", "0.468", "0.856", "0.0825", "0.856" ], [ "0.5024", "0.8413", "0.8896", "0.8413", "0.8896", "0.855", "0.5024", "0.855" ], [ "0.0809", "0.857", "0.4617", "0.857", "0.4617", "0.8706", "0.0809", "0.8706" ], [ "0.5034", "0.857", "0.8896", "0.857", "0.8896", "0.8706", "0.5034", "0.8706" ], [ "0.1013", "0.872", "0.468", "0.872", "0.468", "0.8853", "0.1013", "0.8853" ], [ "0.5034", "0.872", "0.8906", "0.872", "0.8906", "0.8853", "0.5034", "0.8853" ], [ "0.0809", "0.887", "0.4673", "0.887", "0.4673", "0.901", "0.0809", "0.901" ], [ "0.5034", "0.887", "0.8906", "0.887", "0.8906", "0.901", "0.5034", "0.901" ] ]

[ "Coaching a Teachable Student", "Jimuyang Zhang", "Zanming Huang", "Eshed Ohn-Bar", "Boston University", "", "[ zhangjim, huangtom, eohnbar } @bu.edu", "Abstract", "Privileged Agent", "We propose a novel knowledge distillation frame", "work for effectively teaching a sensorimotor student", "G", "agent to drive from the supervision of a privileged", "icyclist", "teacher agent. Current distillation for sensorimotor", "Coach", "Student", "agents methods tend to result in suboptimal learned", "driving behaviorbythe student, which wehypothesize", "is due to inherent differences between the input, mod-", "eling capacity, and optimization processes of the two", "Alignment", "Teachable Net", "Module", "agents. We develop a novel distillation scheme that can", "address these limitations and close the gap between the", "Figure 1. Effective Knowledge Distillation for Sensorimotor", "sensorimotor agent and itsprivilegedteacher.Ourkey", "Agents. Our proposed CaT (Coaching a Teachable student)", "insight is to design a student which learns to align their", "frameworkenableshighlyeffectiveknowledgetransferbe", "input features with the teacher's privileged Bird's Eye", "tween a privileged teacher and a sensorimotor (i.e., image-", "View (BEV) space. The student then can benefit from", "based) student. An alignment module learns to transform", "direct supervision by the teacher over the internal rep-", "image-based features to the teacher's BEV feature space,", "where the student can then leverage extensive and direct su-", "resentation learning. To scaffold the difficult sensori-", "pervision on its learned intermediate representations. The", "motor learning task, the student model is optimized via", "student model is optimized via a coaching mechanism with", "a student-paced coaching mechanism with various aux-", "extensive auxiliary supervision in order to further scaffold", "iliary supervision. We further propose a high-capacity", "the difficult sensorimotor learning task.", "imitation learned privileged agent that surpasses prior", "privileged agents in CARLA and ensures the student", "完整世界知识的高容量特权网络，并将其知识提炼为", "learns safe driving behavior. Our proposed sensori-", "能力较弱的基于视觉的学生网络[10,12,23,49,82]", "motor agent results in a robust image-based behavior", "但是，由于两个代理的输入和体系结构之间的固有差", "cloning agent in CARLA, improving over current mod-", "异，当前的方法仅通过教师的输出国[10,12]或单个最", "els by over 20.6% in driving score without requiring", "终全连接层的知识提炼来依赖教师i.e.的有限监督机", "LiDAR, historical observations, ensemble of models,", "制。E[74,82,84。此外，考虑到有限的视角，特权教师", "on-policy data aggregation or reinforcement learning.", "的演示目标可能会很吵，或者学生很难模仿28]。在", "这项工作中，我们试图开发一种更有效的知识提炼范", "1:Introduction", "式，用于训练感觉运动剂驱动。我们的主要见解是通过", "从图像中学习用于做出复杂驾驶决策的内部表示涉", "减少两个代理之间的内部建模和学习能力之间的差距", "来实现教师的更广泛的监督", "及复杂的优化任务【4,45,57]。驾驶代理的端到端培训", "的固有挑战在于学习将高维视觉观察映射到一般和安", "我们提出的整体知识提炼方法是由人类指导提供的", "全的导航决策中的巨大复杂性「12,20,73]。即使给出", "除了高级演示e.g.之外，通常还涉及结构化的监督", "了数百万个训练示例E41，今天的代理仍然无法可靠", "以学生可以更好地理解的方式提供各种提示来支撑信", "地学习内部表示，该内部表示可用于以安全和任务驱", "息[31]。在教授他人新的和具有挑战性的技能i.e.时，", "动的方式对复杂的视觉场景（e.g.，具有复杂布局和动", "学生可能无法复制演示，例如骑自行车或驾驶车辆，教", "态障碍物的密集城市环境）进行稳健处理E[12,20,73]", "师可以提供有关基本任务结构和他们自已的内部推理", "为了缓解具有挑战性的感觉运动训练任务，最近的", "的额外监督。。与我们的设置类似，特权教学代理可以", "方法将任务分解为多个阶段e.g.，首先训练一个具有", "通过对基础表示学习进行更仔细和直接的指导，在教" ]

{ "arxiv_id": "2306.10014", "language": "zh", "page_idx": 0 }

[ [ "0.0825", "0.0947", "0.4558", "0.0947", "0.4558", "0.1048", "0.0825", "0.1048" ], [ "0.505", "0.09216", "0.8906", "0.09216", "0.8906", "0.1061", "0.505", "0.1061" ], [ "0.1013", "0.10986", "0.468", "0.10986", "0.468", "0.1206", "0.1013", "0.1206" ], [ "0.5034", "0.108", "0.8896", "0.108", "0.8896", "0.1218", "0.5034", "0.1218" ], [ "0.0809", "0.1237", "0.468", "0.1237", "0.468", "0.137", "0.0809", "0.137" ], [ "0.5024", "0.1225", "0.8906", "0.1218", "0.8906", "0.1364", "0.5024", "0.137" ], [ "0.0809", "0.1383", "0.4697", "0.1383", "0.4697", "0.1515", "0.0809", "0.1515" ], [ "0.5034", "0.1383", "0.892", "0.1383", "0.892", "0.1515", "0.5034", "0.1515" ], [ "0.0809", "0.154", "0.468", "0.154", "0.468", "0.1674", "0.0809", "0.1674" ], [ "0.505", "0.1528", "0.8906", "0.1528", "0.8906", "0.1666", "0.505", "0.1666" ], [ "0.0825", "0.1692", "0.4673", "0.1692", "0.4673", "0.1799", "0.0825", "0.1799" ], [ "0.5034", "0.1686", "0.8896", "0.1686", "0.8896", "0.1818", "0.5034", "0.1818" ], [ "0.0825", "0.1849", "0.4673", "0.1849", "0.4673", "0.1957", "0.0825", "0.1957" ], [ "0.505", "0.1831", "0.8896", "0.1831", "0.8896", "0.1963", "0.505", "0.1963" ], [ "0.0825", "0.1995", "0.4656", "0.1995", "0.4656", "0.2102", "0.0825", "0.2102" ], [ "0.5034", "0.1989", "0.63", "0.1989", "0.63", "0.2122", "0.5034", "0.2122" ], [ "0.0809", "0.2134", "0.4673", "0.2134", "0.4673", "0.2267", "0.0809", "0.2267" ], [ "0.5024", "0.2197", "0.8906", "0.2203", "0.8906", "0.2349", "0.5024", "0.2343" ], [ "0.0809", "0.2291", "0.4673", "0.2291", "0.4673", "0.2424", "0.0809", "0.2424" ], [ "0.505", "0.2361", "0.888", "0.2361", "0.888", "0.25", "0.505", "0.25" ], [ "0.0809", "0.2437", "0.468", "0.2437", "0.468", "0.2568", "0.0809", "0.2568" ], [ "0.5034", "0.252", "0.8906", "0.252", "0.8906", "0.2651", "0.5034", "0.2651" ], [ "0.0809", "0.2588", "0.468", "0.2588", "0.468", "0.2727", "0.0809", "0.2727" ], [ "0.505", "0.2683", "0.8896", "0.2683", "0.8896", "0.279", "0.505", "0.279" ], [ "0.08417", "0.276", "0.4673", "0.276", "0.4673", "0.286", "0.08417", "0.286" ], [ "0.5034", "0.2822", "0.5825", "0.2827", "0.5825", "0.2942", "0.5034", "0.2935" ], [ "0.603", "0.2822", "0.8906", "0.2822", "0.8906", "0.2954", "0.603", "0.2954" ], [ "0.0809", "0.289", "0.4673", "0.289", "0.4673", "0.303", "0.0809", "0.303" ], [ "0.5034", "0.2954", "0.8906", "0.2954", "0.8906", "0.3093", "0.5034", "0.3093" ], [ "0.0809", "0.303", "0.468", "0.3037", "0.468", "0.3176", "0.0809", "0.3164" ], [ "0.5034", "0.3113", "0.8906", "0.3113", "0.8906", "0.3245", "0.5034", "0.3245" ], [ "0.0825", "0.3193", "0.468", "0.3193", "0.468", "0.3333", "0.0825", "0.3333" ], [ "0.505", "0.327", "0.8906", "0.327", "0.8906", "0.3403", "0.505", "0.3403" ], [ "0.0809", "0.334", "0.464", "0.334", "0.464", "0.348", "0.0809", "0.348" ], [ "0.5063", "0.3435", "0.888", "0.3435", "0.888", "0.3542", "0.5063", "0.3542" ], [ "0.505", "0.3567", "0.8906", "0.3567", "0.8906", "0.3706", "0.505", "0.3706" ], [ "0.0784", "0.363", "0.2279", "0.363", "0.2279", "0.377", "0.0784", "0.377" ], [ "0.505", "0.3726", "0.888", "0.3726", "0.888", "0.3865", "0.505", "0.3865" ], [ "0.0825", "0.3914", "0.4673", "0.3914", "0.4673", "0.4016", "0.0825", "0.4016" ], [ "0.5034", "0.387", "0.888", "0.387", "0.888", "0.401", "0.5034", "0.401" ], [ "0.0809", "0.4048", "0.468", "0.4048", "0.468", "0.4185", "0.0809", "0.4185" ], [ "0.5034", "0.4016", "0.8896", "0.4016", "0.8896", "0.4153", "0.5034", "0.4153" ], [ "0.0809", "0.4192", "0.4673", "0.4192", "0.4673", "0.433", "0.0809", "0.433" ], [ "0.5024", "0.4167", "0.6406", "0.4185", "0.6406", "0.432", "0.5024", "0.43" ], [ "0.0801", "0.4343", "0.4656", "0.435", "0.4656", "0.449", "0.0801", "0.4475" ], [ "0.0825", "0.4507", "0.4673", "0.4507", "0.4673", "0.464", "0.0825", "0.464" ], [ "0.5034", "0.4475", "0.5986", "0.4475", "0.5986", "0.459", "0.5034", "0.459" ], [ "0.0972", "0.4673", "0.4656", "0.4673", "0.4656", "0.4778", "0.0972", "0.4778" ], [ "0.523", "0.473", "0.888", "0.473", "0.888", "0.486", "0.523", "0.486" ], [ "0.0809", "0.481", "0.4673", "0.481", "0.4673", "0.4944", "0.0809", "0.4944" ], [ "0.5034", "0.4873", "0.8906", "0.4873", "0.8906", "0.5005", "0.5034", "0.5005" ], [ "0.0825", "0.4956", "0.468", "0.4956", "0.468", "0.509", "0.0825", "0.509" ], [ "0.5034", "0.502", "0.8906", "0.502", "0.8906", "0.515", "0.5034", "0.515" ], [ "0.0809", "0.5103", "0.4673", "0.5103", "0.4673", "0.5234", "0.0809", "0.5234" ], [ "0.5024", "0.5176", "0.888", "0.5176", "0.888", "0.531", "0.5024", "0.531" ], [ "0.0825", "0.5264", "0.4673", "0.5264", "0.4673", "0.537", "0.0825", "0.537" ], [ "0.505", "0.534", "0.876", "0.534", "0.876", "0.5444", "0.505", "0.5444" ], [ "0.0809", "0.5405", "0.468", "0.5405", "0.468", "0.5537", "0.0809", "0.5537" ], [ "0.5034", "0.5474", "0.8906", "0.5474", "0.8906", "0.561", "0.5034", "0.561" ], [ "0.0809", "0.556", "0.468", "0.556", "0.468", "0.5693", "0.0809", "0.5693" ], [ "0.5024", "0.561", "0.8906", "0.5625", "0.8906", "0.577", "0.5024", "0.5757" ], [ "0.0801", "0.5693", "0.468", "0.571", "0.468", "0.5854", "0.0801", "0.584" ], [ "0.5034", "0.5776", "0.892", "0.5776", "0.892", "0.5913", "0.5034", "0.5913" ], [ "0.0809", "0.587", "0.1299", "0.587", "0.1299", "0.5986", "0.0809", "0.5986" ], [ "0.5024", "0.5903", "0.539", "0.5923", "0.5376", "0.608", "0.501", "0.606" ], [ "0.0809", "0.608", "0.4673", "0.608", "0.4673", "0.622", "0.0809", "0.622" ], [ "0.5034", "0.6187", "0.638", "0.6187", "0.638", "0.632", "0.5034", "0.632" ], [ "0.0825", "0.625", "0.4673", "0.625", "0.4673", "0.6353", "0.0825", "0.6353" ], [ "0.0801", "0.638", "0.4673", "0.638", "0.4673", "0.6514", "0.0801", "0.6514" ], [ "0.523", "0.6416", "0.8896", "0.6416", "0.8896", "0.655", "0.523", "0.655" ], [ "0.09644", "0.654", "0.468", "0.654", "0.468", "0.6675", "0.09644", "0.6675" ], [ "0.505", "0.656", "0.8896", "0.656", "0.8896", "0.6704", "0.505", "0.6704" ], [ "0.0825", "0.6685", "0.468", "0.6685", "0.468", "0.6816", "0.0825", "0.6816" ], [ "0.5024", "0.672", "0.8896", "0.6724", "0.8896", "0.6875", "0.5024", "0.686" ], [ "0.0825", "0.6846", "0.468", "0.6846", "0.468", "0.6978", "0.0825", "0.6978" ], [ "0.5034", "0.688", "0.8896", "0.688", "0.8896", "0.702", "0.5034", "0.702" ], [ "0.0825", "0.7007", "0.4673", "0.7007", "0.4673", "0.711", "0.0825", "0.711" ], [ "0.5034", "0.702", "0.8906", "0.7026", "0.8906", "0.7173", "0.5034", "0.7163" ], [ "0.08496", "0.714", "0.468", "0.714", "0.468", "0.728", "0.08496", "0.728" ], [ "0.5024", "0.7173", "0.8906", "0.7183", "0.8906", "0.732", "0.5024", "0.731" ], [ "0.08417", "0.731", "0.4673", "0.731", "0.4673", "0.741", "0.08417", "0.741" ], [ "0.505", "0.735", "0.888", "0.735", "0.888", "0.7456", "0.505", "0.7456" ], [ "0.0801", "0.7437", "0.4673", "0.742", "0.4673", "0.756", "0.0801", "0.757" ], [ "0.5034", "0.7485", "0.6396", "0.7485", "0.6396", "0.762", "0.5034", "0.762" ], [ "0.0825", "0.759", "0.4673", "0.759", "0.4673", "0.773", "0.0825", "0.773" ], [ "0.0825", "0.776", "0.4673", "0.776", "0.4673", "0.786", "0.0825", "0.786" ], [ "0.5186", "0.771", "0.8936", "0.7744", "0.8936", "0.797", "0.5186", "0.7925" ], [ "0.0809", "0.789", "0.468", "0.789", "0.468", "0.803", "0.0809", "0.803" ], [ "0.5483", "0.7905", "0.5537", "0.7905", "0.5537", "0.7954", "0.5483", "0.7954" ], [ "0.08417", "0.8047", "0.468", "0.8047", "0.468", "0.8184", "0.08417", "0.8184" ], [ "0.5215", "0.811", "0.8896", "0.811", "0.8896", "0.8247", "0.5215", "0.8247" ], [ "0.0809", "0.8193", "0.2607", "0.8193", "0.2607", "0.8335", "0.0809", "0.8335" ], [ "0.505", "0.827", "0.8896", "0.827", "0.8896", "0.8403", "0.505", "0.8403" ], [ "0.0825", "0.8433", "0.4656", "0.8433", "0.4656", "0.854", "0.0825", "0.854" ], [ "0.505", "0.8413", "0.8896", "0.8413", "0.8896", "0.855", "0.505", "0.855" ], [ "0.0801", "0.857", "0.4673", "0.857", "0.4673", "0.8706", "0.0801", "0.8706" ], [ "0.505", "0.857", "0.8906", "0.857", "0.8906", "0.8706", "0.505", "0.8706" ], [ "0.08417", "0.8735", "0.4673", "0.8735", "0.4673", "0.884", "0.08417", "0.884" ], [ "0.5024", "0.8706", "0.8896", "0.872", "0.8896", "0.886", "0.5024", "0.8853" ], [ "0.0825", "0.887", "0.468", "0.887", "0.468", "0.901", "0.0825", "0.901" ], [ "0.5034", "0.887", "0.8906", "0.887", "0.8906", "0.901", "0.5034", "0.901" ] ]

[ "授有限能力的感觉运动学生i.e.时提供更丰富的监督", "于变压器的架构还可以通过在填充BEV采样网格时", "在我们的工作中，我们介绍了CaT，这是一种新颖", "关注基于图像的信息，将摄像机图像映射到BEV空", "的方法，用于教感觉运动的学生使用特权老师的监督", "间[13,85]】，i.e.。与我们的工作相关的是对ChenE", "来驾驶。我们的主要见解有三个方面：1）有效的老师：", "etal.E[10l的研究，该研究从RGB图像和LiDAR输", "我们建议将明确的安全意识提示纳人BEV空间，以促", "人中学习BEV表示。然后使用运动规划器从BEV生", "进令人惊呀的有效老师代理设计。虽然以前的特权代", "成未来的航路点。由于这带来了重大挑战，我们建议利", "理难以在复杂的城市驾驶场景中学习驾驶，但我们展", "用基于IPM的对齐模块，该模块可以更好地构建图像", "示了我们的学习代理与专家级决策相匹配。2）通过对", "到BEV投影任务", "齐方式可教学生：基于IPM的变压器对齐模块可以促", "Curriculum and Self-pacedLearning：课程学习方法的", "进直接蒸馏大多数老师的功能，并更好地指导学生的", "方法通常是从各种难度样本中结构化学习E5,24,38", "学习过程。3）学生节奏的教练：用于管理困难样本的教", "41,46,63,65,67,72,871。为了更好地考虑具有挑战性", "练机制可以支撑知识，并通过更好地考虑学生的能力", "的感觉运动任务和学生的有限能力，我们采用了一个", "来改善模型优化。通过在广泛的老师和辅助监督下全面", "课程i.e.", "一个模仿教练），逐步更具挑战性的样本。", "处理复杂的知识提炼任务，我们能够在CARLAE[21]", "他etal.[28]还建议在策略数据收集的选代中训练教", "中训练最先进的基于图像的代理。通过消融研究，输人", "练59]。相比之下，我们在没有数据聚合部分的情况", "设计和可解释的特征分析，我们还提供了对当前在学", "下实现并演示了教练在训练送代中的优势。这可以促", "习驱动剂的鲁棒和广义表示方面的局限性的重要见解。", "进更有效的模型训练。该方法由自定进度学习43，75", "驱动，其中模型在每次送代时根据定义的损失动态选", "2. Related Work", "择简单样本。与此类研究相反，我们只平滑硬样本的目", "KnowledgeDistillation：我们研究了最初在训练感觉", "标（基于学生)，而不是丢弃它们。我们从经验上发现", "运动的情况下为模型压缩和加速E[11,17,30,32,36,37，", "这可能会导致教练的改善，这可能是由于困难样本的", "54,58,77,781开发的知识提炼方法。虽然在这种情况下", "脚手架得到了改善。", "对特征蒸馏的方法进行了最少的探索，但我们的动机是", "它们在其他领域的成功，包括图像分类[2,39,42,44,64", "3. Method", "对象检测170，语义分割48，81和自然语言处", "我们的目标是分解具有挑战性的感觉运动学习任务", "理[3,71,76]。然而，考虑到特权代理和学生之间的输", "并通过教师代理的有效监督简化模型优化。我们首先", "人大不相同以及压倒性的感觉运动任务，应用这些技", "在SEC中提出了特权试剂蒸馏的问题。？？。接下来", "术并不是微不足道的。因此，驱动策略蒸馏方法只能从", "我们解决特权代理设计中的当前限制，以在Sec中训", "教师的输出E[10.12或单个全连接层E[74.82.84的", "练健壮的模仿学习老师。3.2。然后，我们提出了", "功能中提供监督，（我们假设）不能提供足够的提示来", "种基于变压器的学生架构，以解决图像到BEV的特征", "指导学生的培训。在这项工作中，我们提供了新颖的机", "对齐（Sec。3.3）。最后，我们通过广泛的监督和知识", "制来缩小这一差距，并通过深度特征蒸馏实现了广泛", "提炼（Sec。3.4）通过渐进式（i.e.，教练）机制（Sec。", "的监督", "3.5)", "ImitationLearningtoDrive：模仿学习（IL）驱动的最", "3.1.Formulation", "新方法可以追溯到Pomerleau57。最近，出现了更", "复杂的基于IL的驾驶方法[8,9,25,47,52,79,80,83,86", "。我们考虑通过知识提炼E12,231从教师代理", "具体来说，将模仿学习任务分解为两个阶段i.e.，首", "f：t→学习感觉运动学生代理的任务，该代理具", "先通过行为克隆E[10,12]或强化学习[7,66,74,82]学", "有对环境信息e.g：，地面真相车道，障碍物和交通信号", "习特权代理，然后训练感觉运动代理以模仿特权代理", "灯状态的特权访问。我们使用权重[0,]ERd将两个", "的输出。我们的研究是由这种方法推动的，但我们探索", "代理参数化为神经网络，并将Fs（;0）和Ft（;山）分", "了特权代理人加强监督的好处。。此外，先前的特权代", "别表示为感觉运动学生网络和教师网络的特征图。给", "理在复杂的城市场景中会产生噜杂和次优的演示监督", "定一个包含感官和特权观察的数据集D以及一个损失", "（即使使用RL进行了优化，正如我们的分析所证明的", "函数L，可以使用", "那样）。虽然目前大多数研究都采用了【82]的基于RL", "的教师，但在现实世界中的安全关键条件下，这可能是", "argmin E(a*,al)~D[(F\"(r; 0), F(art;b) (1)", "有问题的，而且效率低下。在我们的工作中，我们介绍", "6", "了一种新颖的特权老师，它可以大大超过先前的代理，", "其中损失可以在最后一层i.e.上定义，以匹配老", "同时仅依靠离线行为克隆", "师E[10,12]的输出，也可以在内部层i.e.上定义，特征", "IntermediateRepresentationforDriving：学习有效的", "蒸馏E[74,82,84]。鉴于具有挑战性的端到端感觉运动", "3D场景表示对于安全自动驾驶至关重要。研究人员可", "学习任务，教师的作用是为学生提供有效的监督，i.e.，", "以通过使用估计的深度将图像提升到3D并将视锥投", "Eqn中的信息目标。1。但是，由于两个代理的输人", "影到BEV网格E[33,55]中来获得此类表示。或者，基", "截然不同，因此从老师到学生的知识转移可能会变得" ]

{ "arxiv_id": "2306.10014", "language": "zh", "page_idx": 1 }

[ [ "0.339", "0.0947", "0.4429", "0.0947", "0.4429", "0.1048", "0.339", "0.1048" ], [ "0.71", "0.09344", "0.7656", "0.09344", "0.7656", "0.1042", "0.71", "0.1042" ], [ "0.5864", "0.1383", "0.6094", "0.1383", "0.6094", "0.1427", "0.5864", "0.1427" ], [ "0.1609", "0.09534", "0.2507", "0.09534", "0.2507", "0.1061", "0.1609", "0.1061" ], [ "0.591", "0.161", "0.603", "0.161", "0.603", "0.1654", "0.591", "0.1654" ], [ "0.5483", "0.1849", "0.5605", "0.1849", "0.5605", "0.1913", "0.5483", "0.1913" ], [ "0.634", "0.1863", "0.6465", "0.1863", "0.6465", "0.1895", "0.634", "0.1895" ], [ "0.796", "0.1925", "0.8364", "0.1925", "0.8364", "0.1995", "0.796", "0.1995" ], [ "0.2181", "0.2235", "0.2754", "0.2235", "0.2754", "0.2343", "0.2181", "0.2343" ], [ "0.3718", "0.2229", "0.411", "0.2229", "0.411", "0.2299", "0.3718", "0.2299" ], [ "0.4314", "0.25", "0.468", "0.25", "0.468", "0.2568", "0.4314", "0.2568" ], [ "0.1724", "0.2247", "0.2173", "0.2247", "0.2173", "0.2317", "0.1724", "0.2317" ], [ "0.4314", "0.2583", "0.468", "0.2583", "0.468", "0.2651", "0.4314", "0.2651" ], [ "0.1274", "0.276", "0.1438", "0.276", "0.1438", "0.2842", "0.1274", "0.2842" ], [ "0.2075", "0.2727", "0.2345", "0.2727", "0.2345", "0.281", "0.2075", "0.281" ], [ "0.2957", "0.276", "0.3162", "0.276", "0.3162", "0.286", "0.2957", "0.286" ], [ "0.5415", "0.2842", "0.5767", "0.2842", "0.5767", "0.2917", "0.5415", "0.2917" ], [ "0.671", "0.2842", "0.728", "0.2842", "0.728", "0.2942", "0.671", "0.2942" ], [ "0.8667", "0.286", "0.888", "0.286", "0.888", "0.2935", "0.8667", "0.2935" ], [ "0.1177", "0.2986", "0.1478", "0.2986", "0.1478", "0.307", "0.1177", "0.307" ], [ "0.3186", "0.2954", "0.3506", "0.2974", "0.3489", "0.308", "0.3179", "0.3062" ], [ "0.6914", "0.2986", "0.7656", "0.2986", "0.7656", "0.3093", "0.6914", "0.3093" ], [ "0.255", "0.2998", "0.273", "0.3018", "0.273", "0.3062", "0.255", "0.3037" ], [ "0.1912", "0.3018", "0.2059", "0.3018", "0.2059", "0.305", "0.1912", "0.305" ], [ "0.0809", "0.3245", "0.892", "0.3245", "0.892", "0.3384", "0.0809", "0.3384" ], [ "0.0809", "0.3384", "0.892", "0.3384", "0.892", "0.3528", "0.0809", "0.3528" ], [ "0.0801", "0.3516", "0.8906", "0.3516", "0.8906", "0.3655", "0.0801", "0.3655" ], [ "0.0825", "0.3674", "0.8896", "0.3674", "0.8896", "0.3774", "0.0825", "0.3774" ], [ "0.0801", "0.379", "0.8896", "0.379", "0.8896", "0.3926", "0.0801", "0.3926" ], [ "0.0801", "0.3933", "0.8936", "0.3933", "0.8936", "0.4072", "0.0801", "0.4072" ], [ "0.0801", "0.4072", "0.2279", "0.4072", "0.2279", "0.4204", "0.0801", "0.4204" ], [ "0.0825", "0.4495", "0.468", "0.4495", "0.468", "0.4634", "0.0825", "0.4634" ], [ "0.5024", "0.449", "0.8896", "0.449", "0.8896", "0.4634", "0.5024", "0.4634" ], [ "0.0825", "0.464", "0.4697", "0.464", "0.4697", "0.4778", "0.0825", "0.4778" ], [ "0.5034", "0.4622", "0.8896", "0.4634", "0.8896", "0.4778", "0.5034", "0.4766" ], [ "0.0825", "0.4792", "0.468", "0.4792", "0.468", "0.4924", "0.0825", "0.4924" ], [ "0.5034", "0.4792", "0.8906", "0.4792", "0.8906", "0.4924", "0.5034", "0.4924" ], [ "0.0801", "0.4924", "0.4673", "0.4944", "0.4673", "0.5083", "0.0801", "0.506" ], [ "0.5034", "0.4924", "0.8906", "0.4937", "0.8906", "0.5083", "0.5034", "0.507" ], [ "0.0825", "0.509", "0.468", "0.509", "0.468", "0.523", "0.0825", "0.523" ], [ "0.5034", "0.509", "0.8896", "0.509", "0.8896", "0.523", "0.5034", "0.523" ], [ "0.0809", "0.5244", "0.4673", "0.5244", "0.4673", "0.5386", "0.0809", "0.5386" ], [ "0.5073", "0.529", "0.523", "0.529", "0.523", "0.536", "0.5073", "0.536" ], [ "0.517", "0.5244", "0.6357", "0.5244", "0.6357", "0.5386", "0.517", "0.5386" ], [ "0.0809", "0.539", "0.468", "0.539", "0.468", "0.553", "0.0809", "0.553" ], [ "0.523", "0.539", "0.8896", "0.539", "0.8896", "0.553", "0.523", "0.553" ], [ "0.0809", "0.5537", "0.4697", "0.5537", "0.4697", "0.5684", "0.0809", "0.5684" ], [ "0.505", "0.5547", "0.8906", "0.5547", "0.8906", "0.5693", "0.505", "0.5693" ], [ "0.0801", "0.5684", "0.468", "0.5693", "0.468", "0.584", "0.0801", "0.5835" ], [ "0.5034", "0.571", "0.8906", "0.571", "0.8906", "0.584", "0.5034", "0.584" ], [ "0.0809", "0.584", "0.3057", "0.584", "0.3057", "0.5986", "0.0809", "0.5986" ], [ "0.505", "0.5854", "0.8906", "0.5854", "0.8906", "0.5996", "0.505", "0.5996" ], [ "0.5034", "0.5996", "0.6543", "0.5996", "0.6543", "0.613", "0.5034", "0.613" ], [ "0.0801", "0.6157", "0.406", "0.6157", "0.406", "0.629", "0.0801", "0.629" ], [ "0.411", "0.616", "0.4656", "0.616", "0.4656", "0.6274", "0.411", "0.6274" ], [ "0.0809", "0.63", "0.468", "0.63", "0.468", "0.643", "0.0809", "0.643" ], [ "0.5034", "0.627", "0.774", "0.627", "0.774", "0.64", "0.5034", "0.64" ], [ "0.0809", "0.6445", "0.4656", "0.6445", "0.4656", "0.6577", "0.0809", "0.6577" ], [ "0.523", "0.651", "0.8906", "0.651", "0.8906", "0.6655", "0.523", "0.6655" ], [ "0.09235", "0.66", "0.468", "0.66", "0.468", "0.674", "0.09235", "0.674" ], [ "0.5034", "0.6665", "0.8896", "0.6665", "0.8896", "0.681", "0.5034", "0.681" ], [ "0.0825", "0.6753", "0.4673", "0.6753", "0.4673", "0.6895", "0.0825", "0.6895" ], [ "0.505", "0.683", "0.8896", "0.683", "0.8896", "0.6963", "0.505", "0.6963" ], [ "0.08496", "0.69", "0.4673", "0.69", "0.4673", "0.704", "0.08496", "0.704" ], [ "0.0784", "0.7026", "0.298", "0.7007", "0.298", "0.7197", "0.0784", "0.7217" ], [ "0.505", "0.6978", "0.8906", "0.6978", "0.8906", "0.711", "0.505", "0.711" ], [ "0.3308", "0.7056", "0.468", "0.7056", "0.468", "0.7197", "0.3308", "0.7197" ], [ "0.0825", "0.7183", "0.4673", "0.7183", "0.4673", "0.733", "0.0825", "0.733" ], [ "0.505", "0.7134", "0.8906", "0.7134", "0.8906", "0.7266", "0.505", "0.7266" ], [ "0.5034", "0.728", "0.8906", "0.728", "0.8906", "0.741", "0.5034", "0.741" ], [ "0.0809", "0.7344", "0.468", "0.735", "0.468", "0.75", "0.0809", "0.7485" ], [ "0.505", "0.7437", "0.8906", "0.7437", "0.8906", "0.757", "0.505", "0.757" ], [ "0.0809", "0.75", "0.468", "0.7505", "0.468", "0.765", "0.0809", "0.7646" ], [ "0.5034", "0.7583", "0.8906", "0.7583", "0.8906", "0.7715", "0.5034", "0.7715" ], [ "0.0809", "0.7666", "0.4656", "0.7666", "0.4656", "0.78", "0.0809", "0.78" ], [ "0.505", "0.7734", "0.892", "0.7734", "0.892", "0.787", "0.505", "0.787" ], [ "0.0972", "0.781", "0.464", "0.781", "0.464", "0.7944", "0.0972", "0.7944" ], [ "0.0801", "0.7944", "0.4673", "0.7954", "0.4673", "0.81", "0.0801", "0.8096" ], [ "0.505", "0.7886", "0.8906", "0.7886", "0.8906", "0.802", "0.505", "0.802" ], [ "0.505", "0.8047", "0.546", "0.8047", "0.546", "0.8164", "0.505", "0.8164" ], [ "0.54", "0.8037", "0.572", "0.8037", "0.572", "0.815", "0.54", "0.815" ], [ "0.0825", "0.811", "0.4673", "0.811", "0.4673", "0.8247", "0.0825", "0.8247" ], [ "0.0907", "0.8237", "0.468", "0.8213", "0.468", "0.839", "0.0907", "0.8413" ], [ "0.5024", "0.8247", "0.8906", "0.8257", "0.8906", "0.8403", "0.5024", "0.839" ], [ "0.0866", "0.832", "0.0972", "0.832", "0.0972", "0.836", "0.0866", "0.836" ], [ "0.08496", "0.8413", "0.4673", "0.8413", "0.4673", "0.855", "0.08496", "0.855" ], [ "0.505", "0.8413", "0.892", "0.8413", "0.892", "0.855", "0.505", "0.855" ], [ "0.0801", "0.855", "0.468", "0.856", "0.468", "0.8706", "0.0801", "0.869" ], [ "0.505", "0.856", "0.888", "0.856", "0.888", "0.8706", "0.505", "0.8706" ], [ "0.0809", "0.872", "0.468", "0.872", "0.468", "0.8853", "0.0809", "0.8853" ], [ "0.5063", "0.872", "0.8896", "0.872", "0.8896", "0.8853", "0.5063", "0.8853" ], [ "0.0809", "0.886", "0.4673", "0.886", "0.4673", "0.901", "0.0809", "0.901" ], [ "0.5034", "0.887", "0.8896", "0.887", "0.8896", "0.901", "0.5034", "0.901" ] ]

[ "TeacherObservations", "ResNet-18", "per", "Alignment Module", "La", "Seg", "ee", "rt-termGoal", "bservations", "ResNet-34", "Alignment", "StudentC", "Module", "Left", "Middle", "Right", "Segmenta", "Segmentation", "mand", "Lfeat", "Lemd", "TeachableNet", "Leef", "cnne", "Figure 2. Overview of Architecture and Distillation Process of CaT.We first train a high-capacity privileged teacher agent", "(colored in red) which takes an augmented BEV representation as input, and produces a set of future waypoints conditioned", "on command and short-term goal. Our teachable student agent (colored in blue) maps the image features to the BEV", "spaceusing an alignment module,whichistrained bydistillation of intermediatefeatures (arrows colored inblue)andfinal", "output (arrow colored in green). Additionally, we leverage supervision from auxiliary segmentation (arrow colored in red)", "and command prediction tasks (arrow colored in purple) interlaced throughout the network to facilitate the learning of", "task-relevant features.", "无效。具体来说，由于固有的噪声和不确定性，学生的", "论。我们让 as=(I,g,c)eXs和 at=(B,g,c)EXt", "任务比老师i.e.的任务更具挑战性。实际上，当前的感", "分别成为学生和老师的观察结果。给定每个时间步长", "觉运动学习代理及其教学代理的性能之间存在很大差", "的这些观察结果，代理学习在下一个2的自上而下的", "距，即使假设可以访问昂贵且高质量的3D观测结果，", "车辆坐标中预测10个未来的2D航路点。5秒的驾驶。", "例如LiDARE[10,12,15］（我们的工作中未使用）。教", "然后将它们提供给横向和纵向PID控制器「12,50,69]", "学过程也应该理想地考虑学生的任何有限的能力，以", "", "以生成最终动作", "促进有效的优化和学习[28]。此外，学习一个最佳和", "鉴于此公式，我们接下来讨论我们的教师代理设计", "强大的教师本身可能是一个具有挑战性的任务E[12]，", "和培训过程。特别是，我们的目标是学习一个特权代", "即使假设简化的学习任务，如解决的感知，正如我们", "理，它不仅可以有效地解决面向目标的导航任务，而且", "在。4 并在 Sec 中寻址。3.2", "可以促进感觉运动学生政策的有效提炼和指导，这将", "在 Sec 中讨论。3.4", "Problem Setup and Agent Observations :", "我们在", "CARLA模拟器[21I的上下文中开发我们的CaT框", "3.2. Learning an Effective Teacher", "架。我们的目标是训练目标条件感觉运动代理E[15,18", "在我们的公式中，学习学生代理首先要训练有效的", "用于将图像和目标观察结果映射到车辆油门和转向控", "特权代理f。然而，我们发现这本身是一项复杂的任", "制E[12,50]。我们遵循标准的条件模仿学习E[10,15]", "务。即使完全了解完美的BEV感知，当前的特权代理", "，并假设可以访问三个（非重叠）RGB摄像机视图", "I =[10,I1,12] ERWxHx3", "也不是最佳的，性能明显低于CARLA的内置自动驾", "一个分类导航命令", "cE1....,6}（i.e.左转，右转，跟随，前进，左", "驶仪E12,82]。在本节中，我们将通过BEV设计揭示", "这种性能不佳的根本原因。根据这一关键发现，我们培", "车道变换和右车道变换）以及从GNSS（全球导航卫", "训了一位高效的模仿学习老师，该老师超越了先前的", "星系统）采样的中间噪声目标???EMAX。我们注意", "特权代理，并与CARLA中的专家级表现相匹配，而", "到，通过从静态地图和粗定位系统生成A*计划E[27", "无需大量数据收集E[59]或强化学习E[82]。然后，我", "当今的车辆i.e.可以轻松获得命令和GNSS观测", "GNSS目标每50-100米采样一次，反映真实世界的测", "们建议的特权代理可以在SEC中对学生进行有效的监", "督。", "3.3", "量误差「10l。我们通过从底层模拟中谊染特权（i.e", "地面实况）信息来计算BEVBE0,1)WB×HB×CB", "Teacher Training via Direct Expert Imitation：在模", "", "，包括车道，行人，车辆和交通信号灯的3D位置和状", "仿学习中，定义了一个oracle（i.e.，专家驱动程序）", "态E[82]。虽然先前的工作通常利用标准化的BEV表", "f*:Xt→来演示最佳操作y*。在CARLA[2]中", "示形式E12,33,35,61,791，但我们发现其设计对于培", "专家是使用纵向和横向PID控制器定义的，这些控制", "训最佳教师策略至关重要，这将在SecE3.2中进行讨", "器经过精心调整和增强，并通过一系列手动指定的规则" ]

{ "arxiv_id": "2306.10014", "language": "zh", "page_idx": 2 }

[ [ "0.0825", "0.09344", "0.4722", "0.09344", "0.4722", "0.1067", "0.0825", "0.1067" ], [ "0.505", "0.0947", "0.8896", "0.0947", "0.8896", "0.1048", "0.505", "0.1048" ], [ "0.0809", "0.108", "0.468", "0.108", "0.468", "0.1218", "0.0809", "0.1218" ], [ "0.5034", "0.108", "0.8896", "0.108", "0.8896", "0.1218", "0.5034", "0.1218" ], [ "0.0809", "0.1237", "0.468", "0.1237", "0.468", "0.137", "0.0809", "0.137" ], [ "0.505", "0.125", "0.8896", "0.125", "0.8896", "0.1351", "0.505", "0.1351" ], [ "0.0825", "0.137", "0.468", "0.137", "0.468", "0.1515", "0.0825", "0.1515" ], [ "0.5034", "0.137", "0.6504", "0.137", "0.6504", "0.1509", "0.5034", "0.1509" ], [ "0.0825", "0.154", "0.4673", "0.154", "0.4673", "0.1674", "0.0825", "0.1674" ], [ "0.5024", "0.159", "0.8906", "0.1604", "0.8906", "0.1749", "0.5024", "0.1736" ], [ "0.0809", "0.1686", "0.4656", "0.1686", "0.4656", "0.1818", "0.0809", "0.1818" ], [ "0.505", "0.1755", "0.8906", "0.1755", "0.8906", "0.1895", "0.505", "0.1895" ], [ "0.0809", "0.1843", "0.4673", "0.1843", "0.4673", "0.1976", "0.0809", "0.1976" ], [ "0.505", "0.1913", "0.8906", "0.1913", "0.8906", "0.2052", "0.505", "0.2052" ], [ "0.0825", "0.1995", "0.4656", "0.1995", "0.4656", "0.2102", "0.0825", "0.2102" ], [ "0.505", "0.2058", "0.8896", "0.2058", "0.8896", "0.2197", "0.505", "0.2197" ], [ "0.0825", "0.2134", "0.468", "0.2134", "0.468", "0.2267", "0.0825", "0.2267" ], [ "0.5024", "0.2197", "0.8896", "0.2203", "0.8896", "0.2349", "0.5024", "0.2343" ], [ "0.0825", "0.2291", "0.4673", "0.2291", "0.4673", "0.2424", "0.0825", "0.2424" ], [ "0.5034", "0.2343", "0.8906", "0.2349", "0.8906", "0.2488", "0.5034", "0.2474" ], [ "0.0825", "0.2437", "0.468", "0.2437", "0.468", "0.2568", "0.0825", "0.2568" ], [ "0.0801", "0.2583", "0.282", "0.2588", "0.282", "0.2727", "0.0801", "0.2715" ], [ "0.5063", "0.2524", "0.8896", "0.2524", "0.8896", "0.2632", "0.5063", "0.2632" ], [ "0.505", "0.2664", "0.8906", "0.2664", "0.8906", "0.2798", "0.505", "0.2798" ], [ "0.1013", "0.2747", "0.468", "0.2747", "0.468", "0.2886", "0.1013", "0.2886" ], [ "0.505", "0.281", "0.8906", "0.281", "0.8906", "0.2942", "0.505", "0.2942" ], [ "0.0809", "0.289", "0.468", "0.289", "0.468", "0.303", "0.0809", "0.303" ], [ "0.5", "0.2922", "0.8936", "0.2917", "0.8936", "0.3125", "0.5", "0.3132" ], [ "0.0825", "0.3037", "0.468", "0.3037", "0.468", "0.3176", "0.0825", "0.3176" ], [ "0.5034", "0.3113", "0.8906", "0.3113", "0.8906", "0.3245", "0.5034", "0.3245" ], [ "0.0809", "0.3193", "0.4673", "0.3193", "0.4673", "0.3333", "0.0809", "0.3333" ], [ "0.0809", "0.3333", "0.4673", "0.334", "0.4673", "0.348", "0.0809", "0.3467" ], [ "0.505", "0.3276", "0.5522", "0.3276", "0.5522", "0.339", "0.505", "0.339" ], [ "0.0825", "0.3484", "0.464", "0.3484", "0.464", "0.3623", "0.0825", "0.3623" ], [ "0.681", "0.3516", "0.7656", "0.3542", "0.7656", "0.3726", "0.681", "0.3706" ], [ "0.574", "0.363", "0.6724", "0.361", "0.6724", "0.377", "0.574", "0.379" ], [ "0.774", "0.361", "0.8203", "0.3655", "0.818", "0.3801", "0.7715", "0.3757" ], [ "0.8677", "0.3643", "0.8906", "0.3643", "0.8906", "0.3774", "0.8677", "0.3774" ], [ "0.706", "0.3694", "0.774", "0.3726", "0.773", "0.3914", "0.705", "0.3882" ], [ "0.0809", "0.3643", "0.4673", "0.3643", "0.4673", "0.3774", "0.0809", "0.3774" ], [ "0.0809", "0.3801", "0.4673", "0.3801", "0.4673", "0.3933", "0.0809", "0.3933" ], [ "0.706", "0.379", "0.7114", "0.379", "0.7114", "0.3838", "0.706", "0.3838" ], [ "0.0825", "0.3945", "0.468", "0.3945", "0.468", "0.4077", "0.0825", "0.4077" ], [ "0.5107", "0.3984", "0.8945", "0.3977", "0.8945", "0.4185", "0.5107", "0.4192" ], [ "0.0809", "0.4104", "0.468", "0.4104", "0.468", "0.4236", "0.0809", "0.4236" ], [ "0.5034", "0.4185", "0.834", "0.4185", "0.834", "0.432", "0.5034", "0.432" ], [ "0.0825", "0.4268", "0.4673", "0.4268", "0.4673", "0.4375", "0.0825", "0.4375" ], [ "0.0809", "0.44", "0.4656", "0.4414", "0.4656", "0.4526", "0.0809", "0.452" ], [ "0.6216", "0.4414", "0.7686", "0.44", "0.7686", "0.455", "0.6216", "0.4558" ], [ "0.869", "0.4414", "0.8906", "0.4414", "0.8906", "0.455", "0.869", "0.455" ], [ "0.0809", "0.4558", "0.468", "0.4558", "0.468", "0.4697", "0.0809", "0.4697" ], [ "0.505", "0.464", "0.8906", "0.464", "0.8906", "0.4792", "0.505", "0.4792" ], [ "0.0825", "0.4702", "0.468", "0.4702", "0.468", "0.4841", "0.0825", "0.4841" ], [ "0.5", "0.473", "0.8936", "0.473", "0.8936", "0.4968", "0.5", "0.4968" ], [ "0.0809", "0.486", "0.1054", "0.486", "0.1054", "0.4976", "0.0809", "0.4976" ], [ "0.501", "0.4905", "0.892", "0.4912", "0.892", "0.509", "0.501", "0.5083" ], [ "0.0809", "0.5083", "0.4673", "0.5083", "0.4673", "0.5215", "0.0809", "0.5215" ], [ "0.5034", "0.509", "0.892", "0.5083", "0.892", "0.523", "0.5034", "0.5234" ], [ "0.0809", "0.523", "0.468", "0.523", "0.468", "0.536", "0.0809", "0.536" ], [ "0.505", "0.526", "0.5513", "0.526", "0.5513", "0.536", "0.505", "0.536" ], [ "0.0809", "0.5386", "0.468", "0.5386", "0.468", "0.552", "0.0809", "0.552" ], [ "0.5215", "0.537", "0.8906", "0.535", "0.8906", "0.5537", "0.5215", "0.556" ], [ "0.0809", "0.553", "0.4673", "0.553", "0.4673", "0.5664", "0.0809", "0.5664" ], [ "0.5024", "0.5537", "0.8906", "0.5537", "0.8906", "0.5674", "0.5024", "0.5674" ], [ "0.0809", "0.5674", "0.468", "0.5674", "0.468", "0.5806", "0.0809", "0.5806" ], [ "0.5034", "0.5674", "0.8906", "0.5684", "0.8906", "0.5835", "0.5034", "0.582" ], [ "0.0809", "0.5835", "0.468", "0.5835", "0.468", "0.5967", "0.0809", "0.5967" ], [ "0.5034", "0.5854", "0.8906", "0.5854", "0.8906", "0.5986", "0.5034", "0.5986" ], [ "0.0825", "0.5996", "0.4673", "0.5996", "0.4673", "0.61", "0.0825", "0.61" ], [ "0.5034", "0.5986", "0.8906", "0.601", "0.8906", "0.6143", "0.5034", "0.6123" ], [ "0.0809", "0.613", "0.468", "0.613", "0.468", "0.627", "0.0809", "0.627" ], [ "0.5034", "0.6157", "0.8906", "0.6157", "0.8906", "0.629", "0.5034", "0.629" ], [ "0.0825", "0.63", "0.4673", "0.63", "0.4673", "0.64", "0.0825", "0.64" ], [ "0.501", "0.629", "0.6724", "0.63", "0.6724", "0.6445", "0.501", "0.643" ], [ "0.0825", "0.643", "0.4673", "0.643", "0.4673", "0.657", "0.0825", "0.657" ], [ "0.0825", "0.659", "0.4673", "0.659", "0.4673", "0.6724", "0.0825", "0.6724" ], [ "0.569", "0.657", "0.8267", "0.657", "0.8267", "0.6704", "0.569", "0.6704" ], [ "0.8267", "0.659", "0.8896", "0.659", "0.8896", "0.6704", "0.8267", "0.6704" ], [ "0.08417", "0.6753", "0.4656", "0.6753", "0.4656", "0.686", "0.08417", "0.686" ], [ "0.505", "0.683", "0.8896", "0.683", "0.8896", "0.6978", "0.505", "0.6978" ], [ "0.0809", "0.6895", "0.468", "0.6895", "0.468", "0.7026", "0.0809", "0.7026" ], [ "0.5024", "0.6963", "0.8896", "0.6978", "0.8896", "0.712", "0.5024", "0.711" ], [ "0.0825", "0.704", "0.468", "0.704", "0.468", "0.7173", "0.0825", "0.7173" ], [ "0.0809", "0.7183", "0.468", "0.7197", "0.468", "0.733", "0.0809", "0.732" ], [ "0.505", "0.7134", "0.8906", "0.7134", "0.8906", "0.7266", "0.505", "0.7266" ], [ "0.0825", "0.7344", "0.468", "0.7344", "0.468", "0.7476", "0.0825", "0.7476" ], [ "0.5034", "0.729", "0.8037", "0.729", "0.8037", "0.742", "0.5034", "0.742" ], [ "0.08417", "0.7505", "0.4673", "0.7505", "0.4673", "0.761", "0.08417", "0.761" ], [ "0.5024", "0.747", "0.8896", "0.7485", "0.8896", "0.763", "0.5024", "0.761" ], [ "0.0825", "0.763", "0.4673", "0.763", "0.4673", "0.7764", "0.0825", "0.7764" ], [ "0.5034", "0.7646", "0.8896", "0.7646", "0.8896", "0.778", "0.5034", "0.778" ], [ "0.0809", "0.779", "0.4697", "0.779", "0.4697", "0.7925", "0.0809", "0.7925" ], [ "0.5024", "0.778", "0.6284", "0.78", "0.6284", "0.7935", "0.5024", "0.7915" ], [ "0.0825", "0.7944", "0.312", "0.7944", "0.312", "0.808", "0.0825", "0.808" ], [ "0.5034", "0.8037", "0.642", "0.8037", "0.642", "0.8174", "0.5034", "0.8174" ], [ "0.0801", "0.8184", "0.3506", "0.8184", "0.3506", "0.832", "0.0801", "0.832" ], [ "0.524", "0.827", "0.8896", "0.827", "0.8896", "0.8403", "0.524", "0.8403" ], [ "0.10297", "0.8433", "0.4673", "0.8433", "0.4673", "0.854", "0.10297", "0.854" ], [ "0.5063", "0.8433", "0.8906", "0.8433", "0.8906", "0.854", "0.5063", "0.854" ], [ "0.0809", "0.857", "0.468", "0.857", "0.468", "0.8706", "0.0809", "0.8706" ], [ "0.5034", "0.857", "0.888", "0.856", "0.888", "0.869", "0.5034", "0.8706" ], [ "0.08417", "0.8735", "0.4673", "0.8735", "0.4673", "0.884", "0.08417", "0.884" ], [ "0.5024", "0.8706", "0.8906", "0.872", "0.8906", "0.886", "0.5024", "0.8853" ], [ "0.0809", "0.887", "0.468", "0.887", "0.468", "0.901", "0.0809", "0.901" ], [ "0.5034", "0.887", "0.892", "0.887", "0.892", "0.901", "0.5034", "0.901" ] ]

[ "来处理基础模拟状态下的各种情况（e.g.，突然制动）。", "建议通过利用内部图像到BEV对齐模块来改善师生模", "然后，可以使用设计的专家生成轨迹作为监督数据，以", "型对齐。如图所示。E2，因此，对齐模块之后的三个", "通过行为克隆E[18,19,50]学习特权策略函数f。我", "剩余块可以通过直接蒸留Sec中的大多数教师功能来", "们的特权代理是一个ResNet-18模型「29]，如图。2", "促进知识转移。3.4", "(未发现进一步提高建模能力是有益的)。ResNet模型", "IPM-basedTransformerAlignmentModule：为下游任", "之后是基于GRU的条件航路点预测器10,15]。由于", "务将映像显式映射到BEV空间可能具有挑战性。虽", "我们密集的城市导航任务e.g.具有多车道道路，交叉", "然这可以通过深度估计和提升E[33，55]来实现，但", "路口和合并的复杂性，因此我们在每个条件分支中使", "我们试图合并一个高效且可区分的模块，该模块可", "用两个GRU逐步细化航路点。第一个GRU通过顺序", "以有效地参与并推理各种功能，同时从前视图映射到", "航路点直接从嵌人特征回归初步航路点集。随后，第二", "BEV。基于变压器的架构[13,681提供了一个自然的选", "个GRU将预测的航路点与短期目标和嵌人特征一起", "生成一组精确的航路点目标。", "择，因为它可以针对驱动任务进行端到端优化，同时", "还可以仔细对齐具有任意维度的空间之间的特征。如", "即使假设完美的感知，我们也很难训练在复杂的", "图所示。E2，我们使用BEV空间的空间参数化对查", "城市驾驶场景中表现出强大计划行为的特权代理。", "询QinitERLHxLWxLo进行采样，并使用自关注模", "具体来说，当使用最具挑战性的CARLA基准测试", "块E[13,68]对其进行处理。自关注模块Q的输出可以", "（Longest6E[15]，也在SecE4中详细介绍）时，使用", "ChenEetal.E[12]的标准BEV表示导致26%驾驶得", "表示为", "分与72%相比，基于规则的专家。Zhangetal.[82", "QinitKmit", "Q = softmax", "Vinit", "(2)", "Va", "的强化学习代理在此任务上实现了60%。我们试图探", "索这项任务的行为克隆的局限性，因为噜杂和次优的", "", "老师会阻碍学生模型的训练。特别是，我们假设性能差", "其中Kinit、VinitERLHxLwxLo提供键值对，d", "的原因在于密集场景中BEV状态表示的复杂性增加。", "是查询的维度。随后，使用逆透视映射（IPM）", "在这种情况下，学习提取与任务相关的细节可能会变", "得更具挑战性。为了解决这个问题，我们建议引入额外", "p=sPkR(q-tk)", "(3)", "的渠道（类似于负担E[822,62]），可以更容易地转化", "，将每个查询点qEQ分别映射到图像特征", "为安全关键决策，简化学习任务，并提供专家级教学代", "FRGBERZHXZWXZC中的相应参考点PER3（齐", "理。", "次坐标)，其中s是相对比例因子PkER3×3，。请注", "BEVwithSafetyHints：我们的BEV将状态信息编", "意，我们为每个摄像机视图采用不同的IPM映射来执", "码到不同的通道中，包括可行驶的道路区域，基于命令", "行投影", "和布局的所需路线，车道标记以及动态障碍物（请参", "要填充学生的BEV功能FBEVERHB×WB×CB", "见图。3）。考虑到现有特权代理的低性能，特别是在", "(i.e.，图中对齐模块的输出。2），我们利用基于可变", "动态对象周围，代理可以从BEV中更明确的安全驱动", "形DETR「88]的可变形交叉注意机制。这使得网络能", "线索中受益。我们建议添加两种类型的渠道：（1）预", "够在转换到BEV特征空间时关注图像特征中参考点", "测代理的未来和（2）实体关注。首先，我们利用运动", "周围的多个区域。然后，参考点P和学习的可变形偏", "学自行车模型E[15]来有效地预测动态对象的未来轨", "移可以用作键K，其来自提取图像的相应特征具有值", "迹。这使我们能够选代地预测和表示代理的短期未来", "V，对于交叉关注模块", "位置，方向和速度（我们的补充包含有关此过程的其他", "信息）。其次，我们基于应用于自我车辆的运动学自行", "FBEV =DeformAttn(Q,FRGB,H)", " (4)", "车模型，对明确的注意力通道进行编码，以突出潜在的", "，其中H={Pk,Rk,t]=1包含来自Eqm的组合多", "未来违规行为i.e.：虽然这假设当前车辆油门和转向", "视图IPM参数。3。除了建议的对齐和可教模块外", "保持不变，但实体注意力通道有助于代理更好地推理", "安全性，同时将状态与操作相关联。令人惊呀的是，安", "我们还合并了用于辅助任务的预测模块（i.e.，BEV预", "全提示通道的添加为强大的教师提供了单独的离线行", "测E[15]），在Sec中进行了讨论。3.4。", "为克隆，甚至比基于规则的专家有所改善（表1，73%", "GRU-basedWaypointPredictor：将转换后的图像到", "与。72%）。我们注意到，许多剩余的故障是由于在拥挤", "BEV特征输人到三个残差块和基于命令E[10]的基于", "的十字路口模拟超时。我们增强的老师代理对于在Sec", "GRU的条件分支", "中进行有效蒸馏至关重要。3.4", "3.4. Student Loss", "3.3. Learning a Teachable Student", "即使使用我们精心设计的体系结构，学生的基于图", "由于输人和建模能力不同，可能难以将基于图像的", "像的学习任务仍然比老师的任务更具挑战性。因此，优", "学生特征和输出与基于BEV的特权教师对齐。我们建", "化和与老师的对齐都可以从广泛的监督和逐步的学生", "议通过设计一个可教的学生网络i.e.来减轻错位，该模", "指导i.e.中受益，以加强监督并更好地匹配学生的能", "型与尽可能多的教师体系结构相匹配。具体来说，我们", "力。我们的蒸馏过程通过三个整体机制为学生提供充" ]

{ "arxiv_id": "2306.10014", "language": "zh", "page_idx": 3 }

[ [ "0.4265", "0.159", "0.4983", "0.159", "0.4983", "0.1736", "0.4265", "0.1736" ], [ "0.5156", "0.159", "0.648", "0.159", "0.648", "0.1736", "0.5156", "0.1736" ], [ "0.6665", "0.159", "0.777", "0.1578", "0.777", "0.1724", "0.6665", "0.1736" ], [ "0.8325", "0.1604", "0.915", "0.1604", "0.915", "0.1749", "0.8325", "0.1749" ], [ "0.10455", "0.2468", "0.9453", "0.2468", "0.9453", "0.2615", "0.10455", "0.2615" ], [ "0.0809", "0.2664", "0.8896", "0.2664", "0.8896", "0.2798", "0.0809", "0.2798" ], [ "0.0809", "0.281", "0.8906", "0.281", "0.8906", "0.2942", "0.0809", "0.2942" ], [ "0.0809", "0.2954", "0.8896", "0.2954", "0.8896", "0.3093", "0.0809", "0.3093" ], [ "0.0801", "0.3093", "0.8896", "0.308", "0.8896", "0.3213", "0.0801", "0.3225" ], [ "0.0801", "0.3225", "0.8906", "0.3225", "0.8906", "0.336", "0.0801", "0.336" ], [ "0.0801", "0.3372", "0.2974", "0.3372", "0.2974", "0.351", "0.0801", "0.351" ], [ "0.0825", "0.3643", "0.468", "0.3643", "0.468", "0.3774", "0.0825", "0.3774" ], [ "0.5596", "0.3706", "0.808", "0.3706", "0.808", "0.3914", "0.5596", "0.3914" ], [ "0.0809", "0.3801", "0.4673", "0.3801", "0.4673", "0.3933", "0.0809", "0.3933" ], [ "0.08417", "0.3965", "0.4673", "0.3965", "0.4673", "0.4072", "0.08417", "0.4072" ], [ "0.619", "0.3926", "0.6396", "0.3926", "0.6396", "0.3997", "0.619", "0.3997" ], [ "0.8677", "0.3997", "0.8906", "0.3997", "0.8906", "0.4124", "0.8677", "0.4124" ], [ "0.0809", "0.4077", "0.4673", "0.4092", "0.4673", "0.4236", "0.0809", "0.4224" ], [ "0.614", "0.406", "0.834", "0.406", "0.834", "0.4204", "0.614", "0.4204" ], [ "0.0809", "0.4248", "0.4673", "0.4248", "0.4673", "0.4382", "0.0809", "0.4382" ], [ "0.613", "0.428", "0.8145", "0.4287", "0.8145", "0.4463", "0.613", "0.4458" ], [ "0.0825", "0.44", "0.4673", "0.44", "0.4673", "0.4539", "0.0825", "0.4539" ], [ "0.0825", "0.455", "0.468", "0.455", "0.468", "0.4685", "0.0825", "0.4685" ], [ "0.5195", "0.4602", "0.8896", "0.4602", "0.8896", "0.4734", "0.5195", "0.4734" ], [ "0.0809", "0.4702", "0.407", "0.4702", "0.407", "0.4841", "0.0809", "0.4841" ], [ "0.5073", "0.4766", "0.888", "0.4766", "0.888", "0.4873", "0.5073", "0.4873" ], [ "0.0801", "0.4924", "0.4673", "0.4937", "0.4673", "0.5083", "0.0801", "0.507" ], [ "0.505", "0.4893", "0.888", "0.4893", "0.888", "0.5024", "0.505", "0.5024" ], [ "0.505", "0.505", "0.8896", "0.505", "0.8896", "0.5195", "0.505", "0.5195" ], [ "0.0825", "0.511", "0.2345", "0.511", "0.2345", "0.5215", "0.0825", "0.5215" ], [ "0.505", "0.521", "0.8", "0.521", "0.8", "0.534", "0.505", "0.534" ], [ "0.1478", "0.5415", "0.4084", "0.5415", "0.4084", "0.5547", "0.1478", "0.5547" ], [ "0.4468", "0.5415", "0.4673", "0.5415", "0.4673", "0.5537", "0.4468", "0.5537" ], [ "0.505", "0.5425", "0.8896", "0.5425", "0.8896", "0.556", "0.505", "0.556" ], [ "0.505", "0.5566", "0.8896", "0.5566", "0.8896", "0.571", "0.505", "0.571" ], [ "0.0825", "0.5713", "0.3743", "0.5713", "0.3743", "0.5854", "0.0825", "0.5854" ], [ "0.5034", "0.5713", "0.8896", "0.5713", "0.8896", "0.5854", "0.5034", "0.5854" ], [ "0.505", "0.587", "0.8906", "0.587", "0.8906", "0.601", "0.505", "0.601" ], [ "0.0801", "0.5933", "0.4673", "0.5947", "0.4673", "0.6094", "0.0801", "0.608" ], [ "0.5034", "0.601", "0.8906", "0.6016", "0.8906", "0.616", "0.5034", "0.6157" ], [ "0.0809", "0.6094", "0.2402", "0.6094", "0.2402", "0.6226", "0.0809", "0.6226" ], [ "0.505", "0.6187", "0.8906", "0.6187", "0.8906", "0.632", "0.505", "0.632" ], [ "0.505", "0.6333", "0.8896", "0.6333", "0.8896", "0.6465", "0.505", "0.6465" ], [ "0.2091", "0.64", "0.2239", "0.64", "0.2239", "0.651", "0.2091", "0.651" ], [ "0.152", "0.6514", "0.21", "0.654", "0.2091", "0.6694", "0.1503", "0.6675" ], [ "0.1993", "0.6514", "0.398", "0.6494", "0.398", "0.6685", "0.1993", "0.6704" ], [ "0.5024", "0.6465", "0.584", "0.6465", "0.584", "0.661", "0.5024", "0.661" ], [ "0.4468", "0.654", "0.468", "0.654", "0.468", "0.6675", "0.4468", "0.6675" ], [ "0.2059", "0.6704", "0.2279", "0.6704", "0.2279", "0.6787", "0.2059", "0.6787" ], [ "0.5024", "0.6704", "0.7305", "0.672", "0.7305", "0.686", "0.5024", "0.685" ], [ "0.0801", "0.696", "0.4673", "0.6963", "0.4673", "0.711", "0.0801", "0.7104" ], [ "0.535", "0.696", "0.6167", "0.696", "0.6167", "0.707", "0.535", "0.707" ], [ "0.6094", "0.6943", "0.8906", "0.6943", "0.8906", "0.7085", "0.6094", "0.7085" ], [ "0.0825", "0.712", "0.2615", "0.712", "0.2615", "0.726", "0.0825", "0.726" ], [ "0.505", "0.7104", "0.8896", "0.7104", "0.8896", "0.7236", "0.505", "0.7236" ], [ "0.5034", "0.7236", "0.8906", "0.7246", "0.8906", "0.7393", "0.5034", "0.738" ], [ "0.0801", "0.7344", "0.4673", "0.7363", "0.4673", "0.75", "0.0801", "0.7476" ], [ "0.5034", "0.7407", "0.8906", "0.7407", "0.8906", "0.754", "0.5034", "0.754" ], [ "0.0825", "0.7505", "0.4673", "0.7505", "0.4673", "0.765", "0.0825", "0.765" ], [ "0.501", "0.755", "0.8896", "0.755", "0.8896", "0.768", "0.501", "0.768" ], [ "0.0825", "0.7666", "0.468", "0.7666", "0.468", "0.78", "0.0825", "0.78" ], [ "0.505", "0.7715", "0.5376", "0.7715", "0.5376", "0.784", "0.505", "0.784" ], [ "0.0825", "0.7827", "0.464", "0.7827", "0.464", "0.7935", "0.0825", "0.7935" ], [ "0.558", "0.785", "0.863", "0.787", "0.863", "0.8003", "0.558", "0.799" ], [ "0.8667", "0.7886", "0.8906", "0.7886", "0.8906", "0.8", "0.8667", "0.8" ], [ "0.0809", "0.797", "0.4673", "0.797", "0.4673", "0.81", "0.0809", "0.81" ], [ "0.0825", "0.811", "0.468", "0.811", "0.468", "0.8247", "0.0825", "0.8247" ], [ "0.517", "0.811", "0.8896", "0.811", "0.8896", "0.8247", "0.517", "0.8247" ], [ "0.0809", "0.8257", "0.4673", "0.8247", "0.4673", "0.839", "0.0809", "0.8403" ], [ "0.5034", "0.8247", "0.8906", "0.8257", "0.8906", "0.8403", "0.5034", "0.839" ], [ "0.0825", "0.8413", "0.468", "0.8413", "0.468", "0.855", "0.0825", "0.855" ], [ "0.505", "0.8413", "0.892", "0.8413", "0.892", "0.855", "0.505", "0.855" ], [ "0.0809", "0.857", "0.468", "0.857", "0.468", "0.8706", "0.0809", "0.8706" ], [ "0.5034", "0.857", "0.8896", "0.857", "0.8896", "0.8706", "0.5034", "0.8706" ], [ "0.0825", "0.8706", "0.4697", "0.872", "0.4697", "0.886", "0.0825", "0.8853" ], [ "0.5034", "0.872", "0.888", "0.872", "0.888", "0.8853", "0.5034", "0.8853" ], [ "0.0809", "0.887", "0.2443", "0.887", "0.2443", "0.901", "0.0809", "0.901" ], [ "0.505", "0.887", "0.892", "0.887", "0.892", "0.901", "0.505", "0.901" ] ]

[ "(c) Road", "(d) Desired Route", "(e) Lane Marks", "(f) Vehicles", "(a)Baseline [82](b) With Safety Hints (g)Pedestrians (h) Traffic Lights (i)Agent Forecast(i)Entity Attention", "Figure 3. Visualization of Our Proposed BEV with Safety Hints. (a) and (b) visualize the baseline and our proposed BEV", "by encoding different channels into colors. Roads are shown in dark grey, Desired Route is shown in light grey, Lane Marks", "are shown in magenta, Vehicles are shown in blue with past trajectories shown in decreasing saturation with time into the", "past, Pedestrians are shown in light blue, Traffic Lights are shown in red or green according to its state, Agent Forecasts are", "shown in purple with decreasing saturation with time into the future, Entity Attention is shown in orange.: (c) - (i) shows", "the different channels separately", "足的有益监督。首先，我们将深度蒸馏损失i.e.合并到", "Cfeat =Z[IF;(a') - F(a')l2 +", "输出（i.e.，航路点回归任务）和特征图上，如图所示。", "2，可以更直接地监督内部层。其次，我们纳入了额", "-1", "(7)", "外的辅助损失，使优化过程正规化。。这包括在分段和", "IT(F(α)) -Tt(F(α)l2 +", "控制命令上定义的地面实况监督损失e.g：，以及在监", "入cD/F(α)F(αt)lcD", "督正确BEV结构的后对齐特征上定义的几何损失。最", "后，我们利用学生节奏的教练机制，在Sec的整个迭代", "其中我们使用F，表示网络中的层（如图。2，我们", "过程中逐渐增加航路点预测任务的难度。3.5", "直接提炼出三层），T表示几个卷积层之后的特征（与", "LossFunctions：我们指导蒸馏过程的优化目标是蒸馏", "学生网络模型联合训练），CD代表一个倒角距离[53", "(我们设置入cD=0.1）。最后一项是在值激活和每", "和辅助任务的加权和：", "个特征图的空间软argmax【12]上计算的", "CCaT=Lout+Lfeat+Cseg+Lemd", "(5)", "Task-orientedAuxiliaryTasks：为了确保在整个蒸馏", "过程中保留与任务相关的基本信息，我们以分段和基", "各种损失项可以指导学生学习的整体方面", "于命令的辅助任务在整个网络中交错的形式利用正则", "化监督「15,51]。我们的分割损失Cseg是交叉摘损失，", "OutputDistillation：我们利用在每个实例的所有条件", "是在地面实况BEV上计算的，并在分割预测头之间进", "分支上计算的L损失", "行平均。我们的命令预测损失Ccmd是由地面实况命令", "C监督的二进制交叉熵，这有助于学习与任务相关的功", "c", "Cout=", "Zllfo(a,c) -ft(at,c)lli", "能E[10,19]", "(6)", "C=1", "3.5. Student-pacedCoaching", "这可以通过使用不同命令c（跟随ChenEetal.E[12]", "而Sec", "3.4对学生的内部陈述提供了充分的监", "）从老师那里采样来完成", "督，教师的目标可能仍然难以模仿学生。为了更好地考", "虑学生的学习能力，我们建议利用学生节奏的培训机", "FeatureDistillationLoss：我们提出的学生体系结构提", "制，该机制可以通过越来越困难的样本逐步指导学生", "供了BEV空间特征对齐i.e.，用于学生和老师之间的", "i.e.。在计算Eq中的损失时。E5，我们定义了一个", "直接内部特征匹配（我们在Eqn中损失的第一个术语。", "教练", "7）。这与当前提取输出或最终全连接层E10.12.74.82", "Ft←A,FS+(1-入)Ft, if LCaT >T", "(8)", "的方法相反。但是，尽管这可以为学生提供丰富的监", "督，但由于基于感知的计划任务的信息处理差异，这两", "它根据学生的预测和特征对教学目标进行插值。入", "个模型不应完全匹配i.e.。此外，基于BEV的特征可", "是一个超参数，在训练迭代i上线性减小到0。修改后", "以是高度稀疏和结构化的，与获得任务和场景相关表", "的目标仅在每个批次的硬样本上计算。在每个批次中，", "示相比，轻微的特征偏移意义不大。为了灵活地考虑基", "T是一个阈值，定义了相对于特权教师的损失的得分", "于学生的任务并有效地表示BEV空间中的结构信息，", "样本中最低的50个%。虽然以这种方式平滑目标似乎", "我们的特征损失计算为", "违反直觉，但与挖掘困难示例的技术相比，i.e.教练机" ]

{ "arxiv_id": "2306.10014", "language": "zh", "page_idx": 4 }

[ [ "0.4043", "0.0884", "0.5947", "0.0903", "0.5947", "0.108", "0.4043", "0.1061" ], [ "0.02942", "0.11615", "0.976", "0.11615", "0.976", "0.1338", "0.02942", "0.1338" ], [ "0.03595", "0.1338", "0.9614", "0.1338", "0.9614", "0.1515", "0.03595", "0.1515" ], [ "0.1732", "0.1497", "0.8203", "0.1509", "0.8203", "0.1718", "0.1732", "0.1704" ], [ "0.4224", "0.1812", "0.584", "0.1812", "0.584", "0.1957", "0.4224", "0.1957" ], [ "0.1978", "0.1995", "0.806", "0.1995", "0.806", "0.214", "0.1978", "0.214" ], [ "0.3555", "0.2166", "0.6504", "0.2166", "0.6504", "0.2299", "0.3555", "0.2299" ], [ "0.1805", "0.2311", "0.8228", "0.2317", "0.8228", "0.25", "0.1805", "0.2488" ], [ "0.2646", "0.2683", "0.3423", "0.2708", "0.3423", "0.286", "0.2646", "0.2842" ], [ "0.152", "0.3018", "0.4495", "0.3018", "0.4495", "0.315", "0.152", "0.315" ], [ "0.152", "0.318", "0.4534", "0.318", "0.4534", "0.332", "0.152", "0.332" ], [ "0.152", "0.3333", "0.452", "0.3333", "0.452", "0.3467", "0.152", "0.3467" ], [ "0.6465", "0.3384", "0.706", "0.3384", "0.706", "0.3484", "0.6465", "0.3484" ], [ "0.7485", "0.3403", "0.871", "0.3423", "0.871", "0.356", "0.7485", "0.3542" ], [ "0.152", "0.3484", "0.335", "0.3484", "0.335", "0.3623", "0.152", "0.3623" ], [ "0.3489", "0.3484", "0.4495", "0.3484", "0.4495", "0.3599", "0.3489", "0.3599" ], [ "0.539", "0.3435", "0.6323", "0.3435", "0.6323", "0.3547", "0.539", "0.3547" ], [ "0.5156", "0.3516", "0.6406", "0.3516", "0.6406", "0.3694", "0.5156", "0.3694" ], [ "0.6465", "0.3499", "0.7427", "0.3499", "0.7427", "0.3599", "0.6465", "0.3599" ], [ "0.6436", "0.3599", "0.7236", "0.3599", "0.7236", "0.3718", "0.6436", "0.3718" ], [ "0.751", "0.3547", "0.824", "0.3547", "0.824", "0.3662", "0.751", "0.3662" ], [ "0.152", "0.3623", "0.452", "0.3623", "0.452", "0.377", "0.152", "0.377" ], [ "0.5376", "0.3662", "0.607", "0.3662", "0.607", "0.377", "0.5376", "0.377" ], [ "0.6445", "0.3718", "0.7305", "0.3718", "0.7305", "0.3833", "0.6445", "0.3833" ], [ "0.751", "0.3674", "0.806", "0.3674", "0.806", "0.3774", "0.751", "0.3774" ], [ "0.1503", "0.3774", "0.452", "0.377", "0.452", "0.3914", "0.1503", "0.3926" ], [ "0.152", "0.3933", "0.451", "0.3933", "0.451", "0.4077", "0.152", "0.4077" ], [ "0.5337", "0.3933", "0.744", "0.3933", "0.744", "0.4072", "0.5337", "0.4072" ], [ "0.749", "0.3882", "0.871", "0.3882", "0.871", "0.4016", "0.749", "0.4016" ], [ "0.517", "0.4028", "0.7114", "0.4048", "0.7114", "0.4216", "0.517", "0.4192" ], [ "0.751", "0.3997", "0.865", "0.3997", "0.865", "0.4136", "0.751", "0.4136" ], [ "0.1503", "0.4077", "0.452", "0.4072", "0.452", "0.4216", "0.1503", "0.4224" ], [ "0.751", "0.4136", "0.818", "0.4136", "0.818", "0.4248", "0.751", "0.4248" ], [ "0.1487", "0.4216", "0.452", "0.4236", "0.452", "0.4375", "0.1487", "0.435" ], [ "0.5107", "0.4312", "0.8813", "0.4312", "0.8813", "0.4443", "0.5107", "0.4443" ], [ "0.152", "0.4382", "0.4534", "0.4382", "0.4534", "0.452", "0.152", "0.452" ], [ "0.5107", "0.4463", "0.8823", "0.4463", "0.8823", "0.4602", "0.5107", "0.4602" ], [ "0.152", "0.4539", "0.452", "0.4539", "0.452", "0.4673", "0.152", "0.4673" ], [ "0.509", "0.461", "0.693", "0.461", "0.693", "0.4749", "0.509", "0.4749" ], [ "0.152", "0.4685", "0.451", "0.4685", "0.451", "0.4817", "0.152", "0.4817" ], [ "0.152", "0.4841", "0.452", "0.4841", "0.452", "0.4988", "0.152", "0.4988" ], [ "0.1544", "0.5005", "0.451", "0.5005", "0.451", "0.511", "0.1544", "0.511" ], [ "0.513", "0.4956", "0.8813", "0.4956", "0.8813", "0.5103", "0.513", "0.5103" ], [ "0.152", "0.5146", "0.452", "0.5146", "0.452", "0.528", "0.152", "0.528" ], [ "0.5107", "0.513", "0.8813", "0.513", "0.8813", "0.528", "0.5107", "0.528" ], [ "0.1503", "0.53", "0.455", "0.53", "0.455", "0.5435", "0.1503", "0.5435" ], [ "0.5107", "0.531", "0.709", "0.531", "0.709", "0.5454", "0.5107", "0.5454" ], [ "0.152", "0.5444", "0.452", "0.5444", "0.452", "0.558", "0.152", "0.558" ], [ "0.5293", "0.5474", "0.875", "0.5474", "0.875", "0.5625", "0.5293", "0.5625" ], [ "0.1503", "0.56", "0.455", "0.56", "0.455", "0.5737", "0.1503", "0.5737" ], [ "0.5107", "0.5645", "0.88", "0.5645", "0.88", "0.579", "0.5107", "0.579" ], [ "0.1503", "0.5757", "0.308", "0.5757", "0.308", "0.589", "0.1503", "0.589" ], [ "0.5107", "0.582", "0.8784", "0.582", "0.8784", "0.595", "0.5107", "0.595" ], [ "0.513", "0.5996", "0.8823", "0.5996", "0.8823", "0.6143", "0.513", "0.6143" ], [ "0.1438", "0.6143", "0.259", "0.6143", "0.259", "0.629", "0.1438", "0.629" ], [ "0.511", "0.616", "0.8823", "0.616", "0.8823", "0.631", "0.511", "0.631" ], [ "0.5107", "0.634", "0.8813", "0.634", "0.8813", "0.648", "0.5107", "0.648" ], [ "0.1177", "0.617", "0.1422", "0.617", "0.1422", "0.6255", "0.1177", "0.6255" ], [ "0.1168", "0.6416", "0.486", "0.64", "0.486", "0.655", "0.1168", "0.656" ], [ "0.511", "0.651", "0.5664", "0.654", "0.5645", "0.6655", "0.5107", "0.662" ], [ "0.6094", "0.6514", "0.7114", "0.6514", "0.7114", "0.6636", "0.6094", "0.6636" ], [ "0.706", "0.651", "0.8813", "0.651", "0.8813", "0.664", "0.706", "0.664" ], [ "0.1177", "0.659", "0.4885", "0.659", "0.4885", "0.6724", "0.1177", "0.6724" ], [ "0.562", "0.655", "0.585", "0.656", "0.585", "0.66", "0.562", "0.659" ], [ "0.511", "0.6685", "0.8813", "0.6685", "0.8813", "0.6816", "0.511", "0.6816" ], [ "0.1168", "0.6753", "0.4885", "0.675", "0.4885", "0.6895", "0.1168", "0.69" ], [ "0.511", "0.685", "0.8813", "0.685", "0.8813", "0.6987", "0.511", "0.6987" ], [ "0.1177", "0.6934", "0.486", "0.6934", "0.486", "0.707", "0.1177", "0.707" ], [ "0.5107", "0.702", "0.8813", "0.702", "0.8813", "0.7153", "0.5107", "0.7153" ], [ "0.1209", "0.7104", "0.4878", "0.7104", "0.4878", "0.7236", "0.1209", "0.7236" ], [ "0.511", "0.7197", "0.8813", "0.7197", "0.8813", "0.733", "0.511", "0.733" ], [ "0.1177", "0.7266", "0.4885", "0.7266", "0.4885", "0.741", "0.1177", "0.741" ], [ "0.5107", "0.735", "0.8813", "0.735", "0.8813", "0.75", "0.5107", "0.75" ], [ "0.1168", "0.742", "0.486", "0.7437", "0.486", "0.7583", "0.1168", "0.757" ], [ "0.5107", "0.754", "0.869", "0.754", "0.869", "0.7676", "0.5107", "0.7676" ], [ "0.1177", "0.761", "0.486", "0.761", "0.486", "0.7744", "0.1177", "0.7744" ], [ "0.1177", "0.778", "0.4878", "0.778", "0.4878", "0.7925", "0.1177", "0.7925" ], [ "0.5107", "0.786", "0.523", "0.786", "0.523", "0.7944", "0.5107", "0.7944" ], [ "0.544", "0.784", "0.7983", "0.784", "0.7983", "0.7974", "0.544", "0.7974" ], [ "0.1177", "0.7954", "0.486", "0.7954", "0.486", "0.81", "0.1177", "0.81" ], [ "0.1177", "0.8125", "0.4878", "0.8125", "0.4878", "0.827", "0.1177", "0.827" ], [ "0.511", "0.8125", "0.8813", "0.8125", "0.8813", "0.827", "0.511", "0.827" ], [ "0.11926", "0.829", "0.4885", "0.829", "0.4885", "0.8433", "0.11926", "0.8433" ], [ "0.5107", "0.83", "0.8823", "0.83", "0.8823", "0.8433", "0.5107", "0.8433" ], [ "0.1168", "0.8477", "0.4902", "0.8477", "0.4902", "0.8623", "0.1168", "0.8623" ], [ "0.5107", "0.8467", "0.8813", "0.8467", "0.8813", "0.86", "0.5107", "0.86" ], [ "0.5107", "0.863", "0.8813", "0.863", "0.8813", "0.8774", "0.5107", "0.8774" ], [ "0.1478", "0.8706", "0.464", "0.8706", "0.464", "0.884", "0.1478", "0.884" ], [ "0.1356", "0.882", "0.4885", "0.8833", "0.4885", "0.898", "0.1356", "0.8965" ], [ "0.5107", "0.88", "0.8823", "0.88", "0.8823", "0.8945", "0.5107", "0.8945" ], [ "0.1177", "0.8994", "0.411", "0.8994", "0.411", "0.91", "0.1177", "0.91" ], [ "0.5107", "0.898", "0.8813", "0.898", "0.8813", "0.911", "0.5107", "0.911" ] ]

[ "Visual Storytelling", "Ting-Hao (Kenneth) Huang 1 , Francis Ferraro 2 * Nasrin Mostafazadeh 3 , Ishan Misra 1 , Aishwarya Agrawal 4", "Jacob Devlin 6, Ross Girshick 5, Xiaodong He 6 , Pushmeet Kohli 6, Dhruv Batra 4, C. Lawrence Zitnick 5", "Devi Parikh 4 , Lucy Vanderwende 6 , Michel Galley 6, Margaret Mitchell 6", "Microsoft Research", "1 Carnegie Mellon University, 2 Johns Hopkins University, 3 University of Rochester", "4 Virginia Tech, 5 Facebook AI Research", "6 Corresponding authors: { jdevlin,lucyv,mgalley,memitc } @microsoft.com", "Abstract", "We introduce the first dataset forse", "quential vision-to-language , and explore", "how this data may be used for the task", "Adult male", "The sun is setting over", "ofvisual storytelling", "The first re", "Agroup of people", "that are siting next", "wearing sunglasses", "lyingdown on", "the ocean and", "lease of this dataset, SIND 1 v.l, in-", "to each other.", "black pavement.", "mountains", "cludes 81,743 unique photos in 20,211 se-", "quences, aligned to both descriptive (cap-", "Having a good time [M] got exhausted", "Sky illuminated with a", "bonding and talkingbytheheat.", "brilliance of gold and", "tion) and story language. We establish", "orange hues.", "several strong baselines for the storytelling", "Figure l: Example language difference between de-", "task, and motivate an automatic metric to", "scriptions for images in isolation (DII) vs. stories for", "benchmark progress. Modelling concrete", "images in sequence (SIS).", "description as well as figurative and so-", "cial language, as provided in this dataset", "and the storytelling task,has the poten-", "文字和具体的图像内容；第二个需要进一步推断", "tial to move artificial intelligence from ba-", "goodtime，可能是什么样子，或者关于特定日落", "sic understandings of typical visual scenes", "的特殊和值得分享的内容", "towards more and more human-like under-", "我们引入了带有相应描述的顺序图像的第一", "standing of grounded event structure and", "个数据集，该数据集捕获了其中一些微妙但重", "subjective expression.", "要的差异，并推进了视觉讲故事的任务。对于", "相同的图像，我们以三层语言发布数据：（1）", "Introduction", "隔离图像（DII）的描述；（2）Descriptionsof", "images-in-sequence(DIS)；和 (3)Stories for", "1", "除了理解简单的对象和具体的场景之外，还在于", "images", "guence(SIS)", "。这种分层的方法揭", "解释因果结构；理解视觉输入，将不同的时刻联", "-110", "示了时间语境的影响和叙事语言的影响。由于所", "系在一起，因为它们会随着时间的推移产生连贯", "有层都与相同的图像对齐，因此数据集有助于直", "的事件叙述。这需要从对单个图像的推理-没有", "接建模文字和更抽象的视觉概念之间的关系，包", "上下文的静态时刻-转变为描绘事件发生和变化", "括视觉图像与典型事件模式之间的关系。我们还", "的图像序列。在视觉方面，从单个图像到上下文", "提出了一种与人类判断最相关的自动评估指标，", "中的图像的进展使我们能够开始创建一种人工", "并为视觉故事讲述任务建立了几个强大的基线", "智能（AI），该人工智能可以根据已经看到的内", "容来推理视觉时刻。在语言方面，从字面描述到", "2", "Motivation and Related Work", "叙述有助于学习更多的评估，会话和抽象语言。", "例如，这是“彼此坐在一起”与“玩得开心”之", "视觉到语言的工作已经爆炸，研究人员正在检", "间的区别，或者“太阳落山”与“灿烂的天空照", "查图像字幕E(Linetal.,2014;Karpathyand", "亮”之间的区别。”（见图1）。第一个描述捕获", "Fei-Fei, 2015: Vinyals et al., 2015: Xu et al..", "2015; Chen et al., 2015; Young et al., 2014; El-", "T.H. and F.F. contributed equally to this work", "ISequential Images Narrative Dataset. This and fu-", "liottandKeller，2013），问题回答国(Antolet", "ture releases are made available onsind.ai", "al., 2015; Ren et al., 2015; Gao et al., 2015: Ma-" ]

{ "arxiv_id": "1604.03968", "language": "zh", "page_idx": 0 }

[ [ "0.1209", "0.089", "0.1772", "0.089", "0.1772", "0.09973", "0.1209", "0.09973" ], [ "0.2402", "0.0884", "0.326", "0.089", "0.326", "0.09973", "0.2402", "0.0985" ], [ "0.3594", "0.09155", "0.415", "0.09155", "0.415", "0.0985", "0.3594", "0.0985" ], [ "0.625", "0.09216", "0.681", "0.09216", "0.681", "0.1029", "0.625", "0.1029" ], [ "0.739", "0.09344", "0.7837", "0.09344", "0.7837", "0.101", "0.739", "0.101" ], [ "0.1209", "0.09973", "0.2279", "0.0985", "0.2279", "0.1092", "0.1209", "0.10986" ], [ "0.2378", "0.09973", "0.3096", "0.09973", "0.3096", "0.10986", "0.2378", "0.10986" ], [ "0.3586", "0.09973", "0.4192", "0.09973", "0.4192", "0.10986", "0.3586", "0.10986" ], [ "0.808", "0.1029", "0.8716", "0.1048", "0.8716", "0.11615", "0.808", "0.11426" ], [ "0.6587", "0.1048", "0.683", "0.1048", "0.683", "0.11743", "0.6587", "0.11743" ], [ "0.232", "0.1029", "0.2385", "0.1029", "0.2385", "0.1061", "0.232", "0.1061" ], [ "0.5986", "0.1061", "0.66", "0.1061", "0.66", "0.11743", "0.5986", "0.11743" ], [ "0.681", "0.1061", "0.7017", "0.1061", "0.7017", "0.113", "0.681", "0.113" ], [ "0.517", "0.1067", "0.5884", "0.1067", "0.5884", "0.11426", "0.517", "0.11426" ], [ "0.2378", "0.1067", "0.298", "0.1092", "0.298", "0.1206", "0.2378", "0.1187" ], [ "0.7427", "0.1067", "0.7812", "0.1067", "0.7812", "0.11743", "0.7427", "0.11743" ], [ "0.357", "0.108", "0.4885", "0.108", "0.4885", "0.1218", "0.357", "0.1218" ], [ "0.1177", "0.108", "0.2279", "0.1092", "0.2279", "0.1206", "0.1177", "0.1193" ], [ "0.232", "0.11237", "0.2385", "0.11237", "0.2385", "0.11615", "0.232", "0.11615" ], [ "0.5293", "0.11743", "0.568", "0.11743", "0.568", "0.125", "0.5293", "0.125" ], [ "0.8096", "0.113", "0.871", "0.11554", "0.871", "0.127", "0.8096", "0.125" ], [ "0.3594", "0.1193", "0.4126", "0.1193", "0.4126", "0.13", "0.3594", "0.13" ], [ "0.5195", "0.1187", "0.527", "0.1187", "0.527", "0.1225", "0.5195", "0.1225" ], [ "0.7466", "0.1193", "0.7812", "0.1193", "0.7812", "0.1307", "0.7466", "0.1307" ], [ "0.2427", "0.1218", "0.304", "0.1218", "0.304", "0.1288", "0.2427", "0.1288" ], [ "0.1217", "0.1218", "0.1772", "0.1218", "0.1772", "0.1288", "0.1217", "0.1288" ], [ "0.638", "0.1218", "0.6733", "0.1218", "0.6733", "0.1288", "0.638", "0.1288" ], [ "0.814", "0.125", "0.8735", "0.125", "0.8735", "0.1364", "0.814", "0.1364" ], [ "0.5376", "0.1276", "0.5483", "0.1276", "0.5483", "0.132", "0.5376", "0.132" ], [ "0.11926", "0.1276", "0.1912", "0.1288", "0.1912", "0.1401", "0.11926", "0.1395" ], [ "0.2402", "0.13", "0.4436", "0.13", "0.4436", "0.1401", "0.2402", "0.1401" ], [ "0.7305", "0.132", "0.7915", "0.132", "0.7915", "0.1427", "0.7305", "0.1427" ], [ "0.8433", "0.1307", "0.8545", "0.1414", "0.8433", "0.1483", "0.8325", "0.1383" ], [ "0.681", "0.137", "0.7036", "0.137", "0.7036", "0.1445", "0.681", "0.1445" ], [ "0.6216", "0.1351", "0.683", "0.1351", "0.683", "0.1509", "0.6216", "0.1509" ], [ "0.5195", "0.1395", "0.5293", "0.1395", "0.5293", "0.1445", "0.5195", "0.1445" ], [ "0.6045", "0.1383", "0.621", "0.1383", "0.621", "0.1465", "0.6045", "0.1465" ], [ "0.7417", "0.1427", "0.7793", "0.1427", "0.7793", "0.1497", "0.7417", "0.1497" ], [ "0.8096", "0.1459", "0.876", "0.1477", "0.876", "0.159", "0.8096", "0.1572" ], [ "0.5195", "0.1477", "0.584", "0.1477", "0.584", "0.1547", "0.5195", "0.1547" ], [ "0.1152", "0.1433", "0.4885", "0.1433", "0.4885", "0.1572", "0.1152", "0.1572" ], [ "0.6396", "0.1547", "0.6685", "0.1547", "0.6685", "0.1578", "0.6396", "0.1578" ], [ "0.7466", "0.154", "0.7715", "0.1559", "0.7715", "0.1604", "0.7466", "0.1578" ], [ "0.5435", "0.1692", "0.8486", "0.1718", "0.8486", "0.1849", "0.5435", "0.1831" ], [ "0.1136", "0.159", "0.4084", "0.1604", "0.4084", "0.1736", "0.1136", "0.1724" ], [ "0.1168", "0.1963", "0.4885", "0.1963", "0.4885", "0.2102", "0.1168", "0.2102" ], [ "0.5664", "0.2108", "0.638", "0.2108", "0.638", "0.2229", "0.5664", "0.2229" ], [ "0.6685", "0.2108", "0.7017", "0.207", "0.705", "0.2172", "0.671", "0.2216" ], [ "0.709", "0.2052", "0.7656", "0.2122", "0.763", "0.2267", "0.706", "0.2197" ], [ "0.1152", "0.2134", "0.4885", "0.2134", "0.4885", "0.2267", "0.1152", "0.2267" ], [ "0.794", "0.214", "0.806", "0.214", "0.806", "0.2184", "0.794", "0.2184" ], [ "0.1152", "0.2311", "0.4885", "0.2311", "0.4885", "0.2443", "0.1152", "0.2443" ], [ "0.6895", "0.2349", "0.7017", "0.2349", "0.7017", "0.2412", "0.6895", "0.2412" ], [ "0.1152", "0.2474", "0.255", "0.2474", "0.255", "0.2615", "0.1152", "0.2615" ], [ "0.1323", "0.2644", "0.4878", "0.2651", "0.4878", "0.2798", "0.1323", "0.279" ], [ "0.524", "0.2676", "0.5522", "0.2676", "0.5522", "0.279", "0.524", "0.279" ], [ "0.5728", "0.2695", "0.601", "0.2695", "0.601", "0.2778", "0.5728", "0.2778" ], [ "0.7256", "0.2683", "0.7466", "0.2683", "0.7466", "0.2766", "0.7256", "0.2766" ], [ "0.1168", "0.2827", "0.4756", "0.2827", "0.4756", "0.2966", "0.1168", "0.2966" ], [ "0.527", "0.286", "0.5337", "0.2854", "0.5337", "0.2886", "0.528", "0.289" ], [ "0.536", "0.2854", "0.5474", "0.2854", "0.5474", "0.289", "0.536", "0.289" ], [ "0.1177", "0.2998", "0.4878", "0.2998", "0.4878", "0.3132", "0.1177", "0.3132" ], [ "0.1168", "0.3176", "0.4558", "0.3176", "0.4558", "0.3308", "0.1168", "0.3308" ], [ "0.4495", "0.318", "0.4878", "0.318", "0.4878", "0.3289", "0.4495", "0.3289" ], [ "0.11926", "0.334", "0.4885", "0.334", "0.4885", "0.348", "0.11926", "0.348" ], [ "0.5107", "0.334", "0.8813", "0.334", "0.8813", "0.348", "0.5107", "0.348" ], [ "0.1168", "0.3516", "0.4885", "0.3516", "0.4885", "0.3655", "0.1168", "0.3655" ], [ "0.511", "0.351", "0.8784", "0.351", "0.8784", "0.361", "0.511", "0.361" ], [ "0.1969", "0.3687", "0.4004", "0.3687", "0.4004", "0.3818", "0.1969", "0.3818" ], [ "0.4102", "0.3694", "0.4836", "0.3694", "0.4836", "0.3801", "0.4102", "0.3801" ], [ "0.5293", "0.3838", "0.8813", "0.3838", "0.8813", "0.3977", "0.5293", "0.3977" ], [ "0.1177", "0.3706", "0.183", "0.3706", "0.183", "0.3818", "0.1177", "0.3818" ], [ "0.1209", "0.385", "0.4878", "0.385", "0.4878", "0.3984", "0.1209", "0.3984" ], [ "0.1209", "0.4028", "0.357", "0.4028", "0.357", "0.4167", "0.1209", "0.4167" ], [ "0.3987", "0.404", "0.4878", "0.404", "0.4878", "0.414", "0.3987", "0.414" ], [ "0.5107", "0.401", "0.8813", "0.401", "0.8813", "0.414", "0.5107", "0.414" ], [ "0.1177", "0.4192", "0.4878", "0.4192", "0.4878", "0.433", "0.1177", "0.433" ], [ "0.511", "0.4185", "0.8813", "0.4185", "0.8813", "0.432", "0.511", "0.432" ], [ "0.1177", "0.4375", "0.4878", "0.4375", "0.4878", "0.4507", "0.1177", "0.4507" ], [ "0.5107", "0.435", "0.5923", "0.435", "0.5923", "0.4495", "0.5107", "0.4495" ], [ "0.1168", "0.4526", "0.4885", "0.4539", "0.4885", "0.4685", "0.1168", "0.4673" ], [ "0.5146", "0.459", "0.814", "0.4602", "0.814", "0.4749", "0.5146", "0.4734" ], [ "0.8096", "0.4622", "0.88", "0.4622", "0.88", "0.4734", "0.8096", "0.4734" ], [ "0.1177", "0.4717", "0.3325", "0.4717", "0.3325", "0.4849", "0.1177", "0.4849" ], [ "0.5146", "0.4792", "0.8813", "0.4792", "0.8813", "0.4924", "0.5146", "0.4924" ], [ "0.134", "0.4905", "0.4673", "0.4905", "0.4673", "0.504", "0.134", "0.504" ], [ "0.511", "0.4956", "0.8813", "0.4956", "0.8813", "0.509", "0.511", "0.509" ], [ "0.11926", "0.506", "0.4885", "0.506", "0.4885", "0.521", "0.11926", "0.521" ], [ "0.5107", "0.511", "0.8813", "0.512", "0.8813", "0.5264", "0.5107", "0.526" ], [ "0.1177", "0.5234", "0.4902", "0.5234", "0.4902", "0.537", "0.1177", "0.537" ], [ "0.511", "0.529", "0.88", "0.529", "0.88", "0.5425", "0.511", "0.5425" ], [ "0.1177", "0.5405", "0.4885", "0.5405", "0.4885", "0.5537", "0.1177", "0.5537" ], [ "0.5107", "0.5454", "0.8813", "0.547", "0.8813", "0.561", "0.5107", "0.56" ], [ "0.1177", "0.558", "0.2427", "0.558", "0.2427", "0.5713", "0.1177", "0.5713" ], [ "0.513", "0.5645", "0.7812", "0.5645", "0.7812", "0.5776", "0.513", "0.5776" ], [ "0.5293", "0.5806", "0.8813", "0.5806", "0.8813", "0.5947", "0.5293", "0.5947" ], [ "0.1152", "0.5913", "0.1487", "0.5913", "0.1487", "0.603", "0.1152", "0.603" ], [ "0.1462", "0.589", "0.3325", "0.5903", "0.3325", "0.605", "0.1462", "0.604" ], [ "0.1462", "0.5947", "0.152", "0.5947", "0.152", "0.5996", "0.1462", "0.5996" ], [ "0.5107", "0.5986", "0.8813", "0.5986", "0.8813", "0.6123", "0.5107", "0.6123" ], [ "0.2876", "0.6187", "0.4885", "0.617", "0.4885", "0.632", "0.2876", "0.6333" ], [ "0.511", "0.6157", "0.8813", "0.6157", "0.8813", "0.629", "0.511", "0.629" ], [ "0.511", "0.632", "0.8813", "0.632", "0.8813", "0.646", "0.511", "0.646" ], [ "0.12335", "0.619", "0.2974", "0.619", "0.2974", "0.6333", "0.12335", "0.6333" ], [ "0.1152", "0.636", "0.486", "0.6353", "0.486", "0.6494", "0.1152", "0.651" ], [ "0.482", "0.64", "0.486", "0.64", "0.486", "0.646", "0.482", "0.646" ], [ "0.1177", "0.653", "0.486", "0.653", "0.486", "0.6675", "0.1177", "0.6675" ], [ "0.511", "0.6494", "0.88", "0.6494", "0.88", "0.6636", "0.511", "0.6636" ], [ "0.513", "0.6685", "0.568", "0.6685", "0.568", "0.6787", "0.513", "0.6787" ], [ "0.5605", "0.6675", "0.8813", "0.6675", "0.8813", "0.681", "0.5605", "0.681" ], [ "0.5146", "0.6846", "0.88", "0.6846", "0.88", "0.6978", "0.5146", "0.6978" ], [ "0.1177", "0.6704", "0.4885", "0.6704", "0.4885", "0.685", "0.1177", "0.685" ], [ "0.1168", "0.688", "0.4885", "0.688", "0.4885", "0.702", "0.1168", "0.702" ], [ "0.1168", "0.705", "0.4885", "0.705", "0.4885", "0.7197", "0.1168", "0.7197" ], [ "0.5107", "0.702", "0.8813", "0.702", "0.8813", "0.7153", "0.5107", "0.7153" ], [ "0.1152", "0.72", "0.4885", "0.7217", "0.4885", "0.7363", "0.1152", "0.735" ], [ "0.513", "0.7183", "0.8813", "0.7183", "0.8813", "0.732", "0.513", "0.732" ], [ "0.509", "0.7344", "0.58", "0.7363", "0.58", "0.7505", "0.509", "0.7485" ], [ "0.5767", "0.735", "0.8813", "0.735", "0.8813", "0.7485", "0.5767", "0.7485" ], [ "0.511", "0.752", "0.8813", "0.752", "0.8813", "0.765", "0.511", "0.765" ], [ "0.1177", "0.7393", "0.4902", "0.7393", "0.4902", "0.7524", "0.1177", "0.7524" ], [ "0.1177", "0.756", "0.4902", "0.756", "0.4902", "0.771", "0.1177", "0.771" ], [ "0.1177", "0.7734", "0.4885", "0.7734", "0.4885", "0.787", "0.1177", "0.787" ], [ "0.5107", "0.7695", "0.8813", "0.7695", "0.8813", "0.7827", "0.5107", "0.7827" ], [ "0.11926", "0.7905", "0.4885", "0.7905", "0.4885", "0.8037", "0.11926", "0.8037" ], [ "0.511", "0.786", "0.8813", "0.786", "0.8813", "0.8", "0.511", "0.8" ], [ "0.1177", "0.808", "0.4885", "0.808", "0.4885", "0.8213", "0.1177", "0.8213" ], [ "0.511", "0.803", "0.8813", "0.803", "0.8813", "0.8164", "0.511", "0.8164" ], [ "0.1177", "0.8247", "0.492", "0.8247", "0.492", "0.8384", "0.1177", "0.8384" ], [ "0.5107", "0.821", "0.8813", "0.821", "0.8813", "0.834", "0.5107", "0.834" ], [ "0.1168", "0.8423", "0.353", "0.8423", "0.353", "0.856", "0.1168", "0.856" ], [ "0.511", "0.837", "0.8813", "0.837", "0.8813", "0.8516", "0.511", "0.8516" ], [ "0.5107", "0.854", "0.7017", "0.855", "0.7017", "0.869", "0.5107", "0.8687" ], [ "0.1323", "0.867", "0.4885", "0.8687", "0.4885", "0.886", "0.1323", "0.8853" ], [ "0.1152", "0.8833", "0.4004", "0.884", "0.4004", "0.8984", "0.1152", "0.898" ], [ "0.5146", "0.879", "0.8813", "0.879", "0.8813", "0.8965", "0.5146", "0.8965" ], [ "0.134", "0.8965", "0.3896", "0.898", "0.3896", "0.912", "0.134", "0.911" ], [ "0.5107", "0.898", "0.7485", "0.898", "0.7485", "0.912", "0.5107", "0.912" ], [ "0.751", "0.8984", "0.8813", "0.8984", "0.8813", "0.912", "0.751", "0.912" ] ]

[ "beach (684)", "breaking up (350)", "easter(259)", "Storytelling", "Re-telling", "amusement park (525)", "carnival (331)", "church (243)", "Deseription", "", "", "E", "", "", "visit (321)", "Story3", "graduation ceremony (236)", "building a house (415)", "", "", "forImages", "office (226)", "", "Story4", "market(311)", "party(411)", "Storyi", "in Isolation", "", "birthday (399)", "outdoor activity (267)father's day (221)", "Preferred Photo", "+", "", "全福", "", "", "Sequence", "in Sequences", "FlickrAlbum", "Table 1: The number of albums in our tiered dataset", "stoen", "das", "Figure 2: Dataset crowdsourcing workflow.", "for the 15 most frequent kinds of stories.", "linowskiandFritz，2014），视觉短语E(Sadeghi", "DOLN", "", "中", "andFarhadi，2011），视频理解E(Ramanathan", "", "etal.2013）和视觉概念E(Krishnaetal.,2016;", "", "Fang et al., 2015)", "这样的工作侧重于图像内容的直接，文字描", "", "", "", "述。尽管这是连接视觉和语言的令人鼓舞的第", "", "", "步，但它远未达到智能代理进行自然交互所需的", "功能。在评论视觉场景显示“坐在房间里”", "（大", "多数图像字幕工作的典型特征）与相同的视觉", "Figure 3: Interface for the Storytelling task, which", "场景显示“粘合”之间存在显着差异，但尚未探", "contains:1)thephotoalbum.and2)thestoryboard", "种描述以视觉信号为基础", "但它带来", "然后使用亚马逊的MechanicalTurk将此阶段", "索。。后", "了关于社会关系和情感的信息，这些信息可以在", "上下文中另外推断出来（图1", "以前在视觉", "返回的照片呈现给人群工人，以收集相应的故", "到语言研究中看到的相比，基于视觉的故事可以", "事和描述。开发完整数据集的众包工作流程如", "促进更多的评估性和比喻性语言：如果系统可以", "图2所示", "识别出同事看起来是bored，则可以直接对这", "Crowdsourcing Stories In Sequence", "我们开发", "些信息进行评论和采取行动", "了一个2阶段的众包工作流程，以收集文本与", "讲故事本身是已知最古老的人类活动之", "图像对齐的自然故事。第一阶段是讲故事，人群", "(Wiessner，2014），提供了一种教育，保护文", "工作者从给定的相册中选择照片的子集以形成", "化，灌输道德和分享建议的方式；因此，将AI", "照片序列：并编写有关它的故事（请参见图E3", "研究集中在这项任务上有可能带来更多类似人", "）。第二阶段是重新讲述，工人根据第一阶段工", "类的智慧和理解", "人生成的一张照片序列写一个故事", "在这两个阶段中，所有相册照片都按照拍摄", "3", "Dataset Construction", "", "时间的顺序显示，下面有一个“故事板”。在讲", "我们首先生成一个“sto-", "故事时，通过单击相册中的照片，照片的“故事", "卡”会出现在故事板上。工人被指示挑选至少五", "Extracting Photos", "ryable”事件类型列表。我们利用这样的想法", ".", "即“有故事的”事件往往涉及某种形式的占有", "张照片，安排所选照片的顺序，然后在每张卡片", "上写一", "个句子或短语以形成一个故事；这显示为", "与每个图像对齐的文本下方的完整故事。此外", "例如。g。，“约翰的生日聚会”或“Shabnam的", "访问”。“使用Flickr数据发布E(Thomeeetal.,", "2015），我们聚合了5克的照片标题和描述，使用", "此界面捕获文本和照片之间的对齐方式。如果", "StanfordCoreNLPE(Manningetal.,2014)提取", "专辑看起来没有故事性，工人可能会跳过它（例", "如。g。，", "硬币的集合）。两名工作人员跳过的相", "册将被丢弃。重新讲述的界面是类似的，但它显", "所有格依赖模式。我们保留所有格短语的头，如", "果它们可以在WordNet3中被归类为event。0,", "依靠手动风选来瞄准我们的收集工作。2然后使", "示了在第一阶段已经创建的两个照片序列，工作", "用FlickrAPI将这些术语用于收集专辑。3我们", "人员从中选择编写故事。对于每张专辑，2名工", "仅包含包含10至50张照片的相册，其中所有", "作人员进行讲故事（在$0。3/HIT），3名工人执", "相册照片均在48小时内拍摄并获得CC许可。", "行重新讲述（在$0。25/HIT），总共产生1,907", "返回相册最多的查询词见表1", "名工人。所有HIT都使用质量控制来确保不同", "的文本至少15个单词长", "2We simultaneously supplemented this data-driven ef-", "fort by a small hand-constructed gazetteer.", "Crowdsourcing Descriptions of Images In Isola-", "3https://www.flickr.com/services/api/", "tion & Images In Sequence", "我们还使用众包" ]

{ "arxiv_id": "1604.03968", "language": "zh", "page_idx": 1 }

[ [ "0.5146", "0.0903", "0.721", "0.09216", "0.721", "0.1067", "0.5146", "0.1048" ], [ "0.3022", "0.101", "0.3203", "0.1061", "0.3162", "0.11426", "0.298", "0.1092" ], [ "0.7114", "0.0966", "0.8716", "0.09784", "0.8716", "0.11237", "0.7114", "0.1111" ], [ "0.286", "0.1048", "0.3022", "0.1048", "0.3022", "0.1092", "0.286", "0.1092" ], [ "0.5215", "0.1048", "0.5483", "0.1048", "0.5483", "0.11743", "0.5215", "0.11743" ], [ "0.5596", "0.1029", "0.6284", "0.1048", "0.6284", "0.1193", "0.5596", "0.11743" ], [ "0.2917", "0.1111", "0.3145", "0.11554", "0.312", "0.125", "0.2893", "0.1206" ], [ "0.6235", "0.1061", "0.709", "0.1061", "0.709", "0.11743", "0.6235", "0.11743" ], [ "0.5107", "0.11743", "0.558", "0.11743", "0.558", "0.132", "0.5107", "0.132" ], [ "0.5786", "0.1193", "0.607", "0.1193", "0.607", "0.1307", "0.5786", "0.1307" ], [ "0.638", "0.1193", "0.6685", "0.1193", "0.6685", "0.1307", "0.638", "0.1307" ], [ "0.6772", "0.1193", "0.7075", "0.1193", "0.7075", "0.1307", "0.6772", "0.1307" ], [ "0.7114", "0.1193", "0.7534", "0.1193", "0.7534", "0.1307", "0.7114", "0.1307" ], [ "0.7656", "0.1193", "0.796", "0.1193", "0.796", "0.1307", "0.7656", "0.1307" ], [ "0.808", "0.1193", "0.876", "0.1193", "0.876", "0.1307", "0.808", "0.1307" ], [ "0.7075", "0.1288", "0.755", "0.1307", "0.7534", "0.1459", "0.706", "0.1433" ], [ "0.8057", "0.1288", "0.8784", "0.1307", "0.8784", "0.1459", "0.8057", "0.1433" ], [ "0.5767", "0.132", "0.6094", "0.132", "0.6094", "0.1445", "0.5767", "0.1445" ], [ "0.523", "0.132", "0.5474", "0.132", "0.5474", "0.1445", "0.523", "0.1445" ], [ "0.6396", "0.132", "0.6685", "0.132", "0.6685", "0.1445", "0.6396", "0.1445" ], [ "0.6787", "0.132", "0.7114", "0.132", "0.7114", "0.1433", "0.6787", "0.1433" ], [ "0.767", "0.132", "0.796", "0.132", "0.796", "0.1433", "0.767", "0.1433" ], [ "0.273", "0.1351", "0.3228", "0.137", "0.3228", "0.1483", "0.273", "0.1465" ], [ "0.1969", "0.137", "0.2615", "0.137", "0.2615", "0.1477", "0.1969", "0.1477" ], [ "0.3381", "0.137", "0.407", "0.137", "0.407", "0.1477", "0.3381", "0.1477" ], [ "0.4084", "0.1395", "0.482", "0.1401", "0.482", "0.1515", "0.4084", "0.1509" ], [ "0.7134", "0.1445", "0.7534", "0.1465", "0.7524", "0.159", "0.7114", "0.1572" ], [ "0.1978", "0.1459", "0.2712", "0.1459", "0.2712", "0.1559", "0.1978", "0.1559" ], [ "0.1316", "0.1401", "0.1912", "0.1401", "0.1912", "0.1509", "0.1316", "0.1509" ], [ "0.3367", "0.1459", "0.4084", "0.1465", "0.4084", "0.1578", "0.3367", "0.1572" ], [ "0.5215", "0.1459", "0.55", "0.1459", "0.55", "0.1604", "0.5215", "0.1604" ], [ "0.576", "0.1459", "0.6094", "0.1459", "0.6094", "0.1578", "0.576", "0.1578" ], [ "0.6406", "0.1459", "0.665", "0.1459", "0.665", "0.1578", "0.6406", "0.1578" ], [ "0.7686", "0.1459", "0.7915", "0.1459", "0.7915", "0.1578", "0.7686", "0.1578" ], [ "0.681", "0.1465", "0.7075", "0.1465", "0.7075", "0.1578", "0.681", "0.1578" ], [ "0.1111", "0.1465", "0.1936", "0.1465", "0.1936", "0.1768", "0.1111", "0.1768" ], [ "0.2737", "0.1477", "0.3218", "0.1483", "0.3218", "0.1572", "0.2737", "0.1559" ], [ "0.4084", "0.1497", "0.4722", "0.1515", "0.4722", "0.1622", "0.4084", "0.1604" ], [ "0.808", "0.1465", "0.876", "0.1465", "0.876", "0.1578", "0.808", "0.1578" ], [ "0.1993", "0.1578", "0.2566", "0.1578", "0.2566", "0.1654", "0.1993", "0.1654" ], [ "0.2737", "0.1547", "0.406", "0.1559", "0.406", "0.1674", "0.2737", "0.1666" ], [ "0.5215", "0.159", "0.5483", "0.159", "0.5483", "0.1718", "0.5215", "0.1718" ], [ "0.576", "0.159", "0.6094", "0.159", "0.6094", "0.1718", "0.576", "0.1718" ], [ "0.6406", "0.1578", "0.669", "0.1604", "0.6685", "0.1724", "0.6396", "0.1704" ], [ "0.6787", "0.1604", "0.7534", "0.1604", "0.7534", "0.1704", "0.6787", "0.1704" ], [ "0.806", "0.159", "0.8774", "0.1604", "0.8774", "0.1718", "0.806", "0.1704" ], [ "0.7686", "0.161", "0.796", "0.161", "0.796", "0.1692", "0.7686", "0.1692" ], [ "0.4126", "0.161", "0.4697", "0.161", "0.4697", "0.1718", "0.4126", "0.1718" ], [ "0.1978", "0.1674", "0.2615", "0.1674", "0.2615", "0.1749", "0.1978", "0.1749" ], [ "0.2737", "0.1674", "0.3325", "0.1674", "0.3325", "0.1749", "0.2737", "0.1749" ], [ "0.335", "0.1674", "0.367", "0.1674", "0.367", "0.1749", "0.335", "0.1749" ], [ "0.1316", "0.1768", "0.179", "0.1768", "0.179", "0.1875", "0.1316", "0.1875" ], [ "0.1953", "0.1755", "0.3284", "0.1768", "0.3284", "0.1881", "0.1953", "0.1875" ], [ "0.3381", "0.1768", "0.4102", "0.1768", "0.4102", "0.1875", "0.3381", "0.1875" ], [ "0.415", "0.1831", "0.4844", "0.1831", "0.4844", "0.1901", "0.415", "0.1901" ], [ "0.5107", "0.1787", "0.8823", "0.1787", "0.8823", "0.1925", "0.5107", "0.1925" ], [ "0.1316", "0.1863", "0.1871", "0.1863", "0.1871", "0.1963", "0.1316", "0.1963" ], [ "0.1936", "0.1849", "0.2607", "0.1863", "0.2607", "0.1963", "0.1936", "0.1957" ], [ "0.2737", "0.1881", "0.3245", "0.1881", "0.3245", "0.1957", "0.2737", "0.1957" ], [ "0.339", "0.1881", "0.407", "0.1881", "0.407", "0.1957", "0.339", "0.1957" ], [ "0.4143", "0.1913", "0.4844", "0.1913", "0.4844", "0.202", "0.4143", "0.202" ], [ "0.509", "0.1945", "0.8813", "0.1945", "0.8813", "0.2084", "0.509", "0.2084" ], [ "0.1912", "0.1957", "0.3186", "0.1957", "0.3186", "0.2058", "0.1912", "0.2058" ], [ "0.1136", "0.1932", "0.1953", "0.1901", "0.1969", "0.2084", "0.1152", "0.2108" ], [ "0.3381", "0.1957", "0.4143", "0.1957", "0.4143", "0.2058", "0.3381", "0.2058" ], [ "0.3367", "0.2039", "0.4028", "0.207", "0.4019", "0.2184", "0.335", "0.2153" ], [ "0.411", "0.2008", "0.4844", "0.2008", "0.4844", "0.2108", "0.411", "0.2108" ], [ "0.1953", "0.2058", "0.2688", "0.2058", "0.2688", "0.2166", "0.1953", "0.2166" ], [ "0.1299", "0.2052", "0.1912", "0.207", "0.1912", "0.2184", "0.1299", "0.2166" ], [ "0.267", "0.207", "0.326", "0.207", "0.326", "0.2172", "0.267", "0.2172" ], [ "0.411", "0.209", "0.4673", "0.2108", "0.4673", "0.2229", "0.411", "0.2203" ], [ "0.5107", "0.209", "0.8813", "0.209", "0.8813", "0.2235", "0.5107", "0.2235" ], [ "0.1316", "0.2153", "0.1936", "0.2153", "0.1936", "0.226", "0.1316", "0.226" ], [ "0.1929", "0.2166", "0.2173", "0.2166", "0.2173", "0.2247", "0.1929", "0.2247" ], [ "0.273", "0.214", "0.3906", "0.2153", "0.3906", "0.2267", "0.273", "0.226" ], [ "0.1969", "0.2247", "0.2507", "0.2279", "0.25", "0.2393", "0.1953", "0.2361" ], [ "0.4143", "0.2267", "0.4844", "0.2267", "0.4844", "0.2374", "0.4143", "0.2374" ], [ "0.5107", "0.2247", "0.8823", "0.2247", "0.8823", "0.238", "0.5107", "0.238" ], [ "0.1323", "0.2311", "0.1871", "0.2311", "0.1871", "0.2412", "0.1323", "0.2412" ], [ "0.273", "0.2343", "0.3381", "0.2361", "0.3381", "0.2474", "0.273", "0.2456" ], [ "0.3325", "0.2349", "0.4878", "0.2437", "0.486", "0.2615", "0.3308", "0.2524" ], [ "0.3381", "0.2311", "0.4102", "0.2311", "0.4102", "0.2412", "0.3381", "0.2412" ], [ "0.4143", "0.2361", "0.482", "0.2361", "0.482", "0.2468", "0.4143", "0.2468" ], [ "0.1969", "0.2361", "0.2632", "0.2361", "0.2632", "0.2468", "0.1969", "0.2468" ], [ "0.1152", "0.238", "0.1847", "0.2412", "0.1847", "0.2524", "0.1152", "0.25" ], [ "0.509", "0.238", "0.8813", "0.2393", "0.8813", "0.254", "0.509", "0.2524" ], [ "0.1953", "0.2443", "0.2566", "0.2468", "0.2566", "0.2568", "0.1953", "0.2551" ], [ "0.2737", "0.2443", "0.3186", "0.2468", "0.3179", "0.2583", "0.273", "0.2556" ], [ "0.1316", "0.2507", "0.1708", "0.2507", "0.1708", "0.2615", "0.1316", "0.2615" ], [ "0.339", "0.2474", "0.361", "0.252", "0.357", "0.2644", "0.335", "0.26" ], [ "0.411", "0.2551", "0.486", "0.2551", "0.486", "0.2651", "0.411", "0.2651" ], [ "0.5107", "0.2551", "0.8813", "0.2551", "0.8813", "0.2683", "0.5107", "0.2683" ], [ "0.1152", "0.2683", "0.4902", "0.2683", "0.4902", "0.2822", "0.1152", "0.2822" ], [ "0.1969", "0.2556", "0.2214", "0.2556", "0.2214", "0.2644", "0.1969", "0.2644" ], [ "0.509", "0.2708", "0.884", "0.2708", "0.884", "0.2842", "0.509", "0.2842" ], [ "0.1177", "0.2842", "0.4902", "0.2842", "0.4902", "0.2974", "0.1177", "0.2974" ], [ "0.509", "0.2854", "0.728", "0.2854", "0.728", "0.2986", "0.509", "0.2986" ], [ "0.1136", "0.2986", "0.3677", "0.2986", "0.3677", "0.3125", "0.1136", "0.3125" ], [ "0.539", "0.307", "0.5923", "0.307", "0.5923", "0.3176", "0.539", "0.3176" ], [ "0.6626", "0.3062", "0.7197", "0.3062", "0.7197", "0.3176", "0.6626", "0.3176" ], [ "0.7935", "0.307", "0.8486", "0.307", "0.8486", "0.3176", "0.7935", "0.3176" ], [ "0.511", "0.315", "0.7173", "0.3176", "0.7173", "0.3308", "0.511", "0.3289" ], [ "0.7134", "0.318", "0.751", "0.318", "0.751", "0.3289", "0.7134", "0.3289" ], [ "0.761", "0.318", "0.871", "0.318", "0.871", "0.3289", "0.761", "0.3289" ], [ "0.511", "0.3276", "0.7075", "0.3276", "0.7075", "0.3416", "0.511", "0.3416" ], [ "0.7183", "0.3276", "0.7524", "0.3276", "0.7524", "0.339", "0.7183", "0.339" ], [ "0.767", "0.33", "0.7876", "0.33", "0.7876", "0.3384", "0.767", "0.3384" ], [ "0.8", "0.3276", "0.834", "0.3276", "0.834", "0.339", "0.8", "0.339" ], [ "0.1177", "0.334", "0.4902", "0.334", "0.4902", "0.348", "0.1177", "0.348" ], [ "0.847", "0.33", "0.8677", "0.33", "0.8677", "0.3384", "0.847", "0.3384" ], [ "0.5825", "0.3372", "0.6167", "0.339", "0.615", "0.351", "0.581", "0.3484" ], [ "0.7573", "0.3384", "0.876", "0.3384", "0.876", "0.3516", "0.7573", "0.3516" ], [ "0.6763", "0.339", "0.7036", "0.339", "0.7036", "0.351", "0.6763", "0.351" ], [ "0.509", "0.339", "0.5864", "0.339", "0.5864", "0.351", "0.509", "0.351" ], [ "0.626", "0.3403", "0.669", "0.3403", "0.669", "0.3516", "0.626", "0.3516" ], [ "0.7183", "0.339", "0.7563", "0.339", "0.7563", "0.351", "0.7183", "0.351" ], [ "0.509", "0.3484", "0.619", "0.3499", "0.619", "0.361", "0.509", "0.3599" ], [ "0.7983", "0.3484", "0.875", "0.3484", "0.875", "0.3591", "0.7983", "0.3591" ], [ "0.6313", "0.351", "0.6587", "0.351", "0.6587", "0.3591", "0.6313", "0.3591" ], [ "0.1177", "0.351", "0.4885", "0.351", "0.4885", "0.3643", "0.1177", "0.3643" ], [ "0.7134", "0.351", "0.759", "0.351", "0.759", "0.361", "0.7134", "0.361" ], [ "0.7646", "0.351", "0.79", "0.351", "0.79", "0.3591", "0.7646", "0.3591" ], [ "0.621", "0.3599", "0.7075", "0.3599", "0.7075", "0.3718", "0.621", "0.3718" ], [ "0.681", "0.3516", "0.7017", "0.3516", "0.7017", "0.3599", "0.681", "0.3599" ], [ "0.511", "0.361", "0.584", "0.361", "0.584", "0.3726", "0.511", "0.3726" ], [ "0.7607", "0.3623", "0.79", "0.3623", "0.79", "0.3706", "0.7607", "0.3706" ], [ "0.8057", "0.3623", "0.83", "0.3623", "0.83", "0.3706", "0.8057", "0.3706" ], [ "0.5254", "0.375", "0.8667", "0.375", "0.8667", "0.3882", "0.5254", "0.3882" ], [ "0.1152", "0.3662", "0.4878", "0.3674", "0.4878", "0.3818", "0.1152", "0.3806" ], [ "0.1177", "0.3838", "0.486", "0.3838", "0.486", "0.3984", "0.1177", "0.3984" ], [ "0.1152", "0.4016", "0.4878", "0.4016", "0.4878", "0.4153", "0.1152", "0.4153" ], [ "0.5107", "0.4077", "0.8823", "0.4092", "0.8823", "0.4236", "0.5107", "0.4224" ], [ "0.1152", "0.4185", "0.4878", "0.4172", "0.4878", "0.432", "0.1152", "0.433" ], [ "0.511", "0.4268", "0.8823", "0.4268", "0.8823", "0.44", "0.511", "0.44" ], [ "0.1168", "0.435", "0.4885", "0.435", "0.4885", "0.4495", "0.1168", "0.4495" ], [ "0.511", "0.443", "0.8813", "0.443", "0.8813", "0.4583", "0.511", "0.4583" ], [ "0.1177", "0.4526", "0.4878", "0.4526", "0.4878", "0.4666", "0.1177", "0.4666" ], [ "0.511", "0.4602", "0.8813", "0.4602", "0.8813", "0.4749", "0.511", "0.4749" ], [ "0.1177", "0.4697", "0.4878", "0.4697", "0.4878", "0.4841", "0.1177", "0.4841" ], [ "0.511", "0.4778", "0.8813", "0.4778", "0.8813", "0.4912", "0.511", "0.4912" ], [ "0.1168", "0.488", "0.4902", "0.488", "0.4902", "0.502", "0.1168", "0.502" ], [ "0.513", "0.4944", "0.8813", "0.4944", "0.8813", "0.509", "0.513", "0.509" ], [ "0.1136", "0.505", "0.183", "0.505", "0.183", "0.5195", "0.1136", "0.5195" ], [ "0.511", "0.511", "0.8813", "0.511", "0.8813", "0.526", "0.511", "0.526" ], [ "0.3064", "0.531", "0.486", "0.53", "0.486", "0.5444", "0.3064", "0.5454" ], [ "0.5073", "0.53", "0.6772", "0.53", "0.6772", "0.5444", "0.5073", "0.5444" ], [ "0.1168", "0.549", "0.4878", "0.549", "0.4878", "0.563", "0.1168", "0.563" ], [ "0.1217", "0.5327", "0.3137", "0.5327", "0.3137", "0.547", "0.1217", "0.547" ], [ "0.513", "0.5596", "0.5254", "0.5645", "0.5186", "0.5728", "0.5063", "0.5674" ], [ "0.544", "0.5596", "0.7935", "0.5596", "0.7935", "0.5737", "0.544", "0.5737" ], [ "0.5312", "0.565", "0.5376", "0.565", "0.5376", "0.5684", "0.5312", "0.5684" ], [ "0.1168", "0.565", "0.4878", "0.565", "0.4878", "0.58", "0.1168", "0.58" ], [ "0.1152", "0.5806", "0.4902", "0.582", "0.4902", "0.5996", "0.1152", "0.5986" ], [ "0.509", "0.586", "0.8813", "0.586", "0.8813", "0.5996", "0.509", "0.5996" ], [ "0.1177", "0.601", "0.4878", "0.601", "0.4878", "0.6143", "0.1177", "0.6143" ], [ "0.511", "0.604", "0.8813", "0.604", "0.8813", "0.617", "0.511", "0.617" ], [ "0.1168", "0.617", "0.2974", "0.6187", "0.2974", "0.6333", "0.1168", "0.632" ], [ "0.3022", "0.616", "0.4902", "0.617", "0.4902", "0.632", "0.3022", "0.631" ], [ "0.511", "0.622", "0.8813", "0.622", "0.8813", "0.6353", "0.511", "0.6353" ], [ "0.1168", "0.634", "0.4412", "0.6353", "0.4412", "0.6494", "0.1168", "0.6484" ], [ "0.513", "0.638", "0.8823", "0.638", "0.8823", "0.653", "0.513", "0.653" ], [ "0.511", "0.656", "0.8813", "0.656", "0.8813", "0.6694", "0.511", "0.6694" ], [ "0.1152", "0.6665", "0.1544", "0.6665", "0.1544", "0.678", "0.1152", "0.678" ], [ "0.1462", "0.664", "0.273", "0.6655", "0.273", "0.6797", "0.1462", "0.6787" ], [ "0.509", "0.6724", "0.8813", "0.6724", "0.8813", "0.6875", "0.509", "0.6875" ], [ "0.1152", "0.6924", "0.4902", "0.6914", "0.4902", "0.7056", "0.1152", "0.707" ], [ "0.511", "0.6914", "0.584", "0.6914", "0.584", "0.7026", "0.511", "0.7026" ], [ "0.1152", "0.7104", "0.486", "0.7104", "0.486", "0.7236", "0.1152", "0.7236" ], [ "0.527", "0.707", "0.8823", "0.707", "0.8823", "0.72", "0.527", "0.72" ], [ "0.1152", "0.7266", "0.4885", "0.7266", "0.4885", "0.7407", "0.1152", "0.7407" ], [ "0.511", "0.7246", "0.8823", "0.7246", "0.8823", "0.738", "0.511", "0.738" ], [ "0.1177", "0.744", "0.4885", "0.744", "0.4885", "0.7583", "0.1177", "0.7583" ], [ "0.5073", "0.7407", "0.8813", "0.741", "0.8813", "0.756", "0.5073", "0.755" ], [ "0.1152", "0.7603", "0.2632", "0.757", "0.2632", "0.7744", "0.1152", "0.778" ], [ "0.511", "0.759", "0.8813", "0.759", "0.8813", "0.773", "0.511", "0.773" ], [ "0.134", "0.779", "0.4878", "0.779", "0.4878", "0.7935", "0.134", "0.7935" ], [ "0.511", "0.776", "0.8813", "0.776", "0.8813", "0.789", "0.511", "0.789" ], [ "0.1177", "0.7954", "0.4902", "0.7954", "0.4902", "0.81", "0.1177", "0.81" ], [ "0.509", "0.7915", "0.8823", "0.7925", "0.8823", "0.8066", "0.509", "0.806" ], [ "0.5063", "0.806", "0.8813", "0.808", "0.8813", "0.8257", "0.5063", "0.8237" ], [ "0.1177", "0.813", "0.4902", "0.813", "0.4902", "0.827", "0.1177", "0.827" ], [ "0.1168", "0.829", "0.4902", "0.83", "0.4902", "0.8447", "0.1168", "0.8433" ], [ "0.509", "0.827", "0.8813", "0.827", "0.8813", "0.8403", "0.509", "0.8403" ], [ "0.5107", "0.8433", "0.8823", "0.8423", "0.8823", "0.857", "0.5107", "0.858" ], [ "0.134", "0.855", "0.3635", "0.856", "0.3635", "0.8706", "0.134", "0.869" ], [ "0.134", "0.8687", "0.4885", "0.872", "0.4885", "0.886", "0.134", "0.8833" ], [ "0.5312", "0.8687", "0.8823", "0.8706", "0.8823", "0.8853", "0.5312", "0.8833" ], [ "0.1152", "0.8853", "0.407", "0.8853", "0.407", "0.8984", "0.1152", "0.8984" ], [ "0.5107", "0.884", "0.876", "0.884", "0.876", "0.8984", "0.5107", "0.8984" ], [ "0.134", "0.8965", "0.3652", "0.8984", "0.3652", "0.9136", "0.134", "0.911" ], [ "0.528", "0.8965", "0.8677", "0.8984", "0.8677", "0.912", "0.528", "0.91" ] ]

[ "Data#(Txt, Img)VocabAvg.", "", "%Abs FrazierYngve Ppl", "", "Set", "Pairs (k)", "租", "Size (k) #Tok", "Brown", "52.1", "47.7", "20.8", "15.2%", "18.5", "77.2194.0", "21.3%", "27.4147.0", "151.8", "DII", "13.8", "11.0", "10.3", "Theboy is", "A man playing", "A soccer ball is", "Two balls and a", "24.8%", "soccer outside of", "Ablackfrisbe", "over a roof bya", "DIS", "151.8", "5.0", "9.2", "9.8", "", "throwinga", "frisbee are on", "23.7146.8", "awhitehouse", "soccer ball by frisbee in a rain", "SIS", "252.9", "18.2", "10.222.1%", "27.5116.0", "10.5", "top of a roof.", "with a red door", "thered door.", "gutter.", "Arooftop", "Aman is standingAman is in", "Ablue and white", "Twosoccerballs", "Table 2: A summary of our dataset, following the", "with a black", "in the grass in", "thefrontof", "soccerball and", "and a Frisbee are", "proposed analyses of Ferraro et al. (2015) , includ-", "front of the house the house", "frisbeelaying", "black Frisbee are", "on the edge of", "sitting on top of", "kicking a soccer", "on the topof", "throwinga", "the roof top.", "ing the Frazier and Yngve measures of syntactic com-", "the edge of it.", "ball.", "soccerball up theroof top", "Why not try", "Now the discus,", "plexity. The balanced Brown corpus (Marcus et al.,", "Adiscus got", "Up the soccer", "wrieololeyl", "It didn't work so", "soccer ball, and", "getting it down", "stuckupon", "1999),provided for comparison, contains only text.", "withasoccer", "ball goes.", "the roof.", "ball.", "stuck on the roof", "Perplexity (Ppl) is calculated against a 5-gram lan-", "Figure 4: Example descriptions of images in isolation", "ball?", "guage model learned on a generic 30B English words", "(DII); descriptions of images in sequence (DIS); and", "dataset scraped from the web.", "stories of images in sequence (SIS).", "Desc.-in-Iso.", "Desc.-in-Seq.", "Story-in-Seq.", "mansitting blackchatting amount", "trunk", "went [female]see", "woman white large gentleman goers", "facing", "got", "today", "收集隔离图像（DII）的描述和顺序图像（DIS）", "saw", "front", "[male] decided came", "sofa", "standingtwo", "enjoys", "bench", "holding young group", "reallystarted", "folks", "的描述，以获取第一个任务中大多数工人的故", "enjoying", "took", "shoreline female", "egg", "wearing image", "great", "time", "Table 3: Top words ranked by normalized PMI.", "事的照片序列（如图E2）。在DII和DIS任务", "中，要求工作人员遵循MSCOCOE(Linetal..", "2014）中提出的图像字幕说明，例如描述所有重", "为“男人”或“女人”。单个图像通常不会传达", "要部分。在DII中，我们使用MSCOCO图像", "有关潜在事件或动作的太多信息，这导致大量使", "字幕接口。4在DIS中，我们使用讲故事界面的", "用姿势动词（“站立”，“坐着”等。）。当我们转", "故事板和故事卡来显示照片序列，并使用适用于", "向按顺序描述时，这些相对不含信息的单词的代", "序列的MSCOCO指令。我们为DII招聘3名", "表性要小得多。最后，顶级故事顺序单词包括更", "工人（在$0。05/HIT)和 3名DIS工人（在$0。", "多讲故事的元素，例如名称（「male］），时间", "07/HIT)", "引用（today）和更具动态性和抽象性的单词", "我们使用CoreNLP令", "（went，decided)", "牌化器对所有故事情节和描述进行令牌化，并用", "Data Post-processing", "5", "Automatic Evaluation Metric", "", "通用male/female令牌，5和所有已识别的命", "名实体及其实体类型（e。g。，location）。在按", "鉴于复杂的讲故事任务的性质，评估生成的故事", "顺序排列的专辑中分割了80个%/10%/10%之", "质量的最佳和最可靠的评估是人类的判断。但", "后，数据以training，", "validation和test的", "是，自动评估指标对于快速基准测试进度很有", "形式发布。每个层的示例语言如图4所示", "用。为了更好地理解哪个指标可以作为人类评", "估的代理，我们计算了从SIS训练集中采样的", "4", "Data Analysis", "3,000个故事的自动指标和人类判断之间的成对", "我们的数据集包括10,117张Flickr专辑和", "相关系数", "210.819张独特的照片。每张专辑平均有20张。", "对于人类的判断，我们再次在MTurk上使", "8张照片（α=9。0)。每张专辑的平均时间跨度", "用众包，每个故事询问五名法官，以评估他们", "为7。9小时（α=11。4)。数据集每层的更多细", "对“如果这些是我的照片，我想使用这样的故", "节如表2所示。6", "事与我的朋友分享我的经历”这一说法的强烈", "我们使用归一化的逐点互信息来识别与每个", "程度。。7我们将五个判断的平均值作为故事的", "层最紧密关联的单词（表E3）。孤立描述的热", "最终得分。对于自动度量，我们使用METEOR", "8平滑的BLEU（Lin and Och，2004）和 Skip-", "门词反映了一个贫乏的消除歧义的背景：对人的", "引用通常缺乏社会特异性，因为人们被简单地称", "ThoughtsE(Kirosetal.，2015）来计算给定序列", "的每个故事之间的相似性。Skip-thoughts提供", "4https://github.com/tylin/coco-ui", "5We use those names occurring at least 10,000 times.", "7Scale presented ranged from“Strongly disagree\"to", "https://ssa.gov/oact/babynames/names.zip", "\"Strongly agree\", which we convert to a scale of 1 to 5.", "6we exclude words seen only once.", "8We use METEOR version 1.5 with hter weights." ]

{ "arxiv_id": "1604.03968", "language": "zh", "page_idx": 2 }

[ [ "0.152", "0.09155", "0.2279", "0.09155", "0.2279", "0.1029", "0.152", "0.1029" ], [ "0.2712", "0.09155", "0.3145", "0.09155", "0.3145", "0.1061", "0.2712", "0.1061" ], [ "0.3865", "0.09155", "0.5146", "0.09155", "0.5146", "0.1061", "0.3865", "0.1061" ], [ "0.513", "0.09534", "0.88", "0.09534", "0.88", "0.1061", "0.513", "0.1061" ], [ "0.1528", "0.108", "0.2532", "0.108", "0.2532", "0.1193", "0.1528", "0.1193" ], [ "0.2695", "0.108", "0.3694", "0.108", "0.3694", "0.1193", "0.2695", "0.1193" ], [ "0.3857", "0.108", "0.4844", "0.108", "0.4844", "0.1193", "0.3857", "0.1193" ], [ "0.255", "0.10986", "0.2607", "0.11426", "0.2573", "0.11615", "0.2524", "0.11237" ], [ "0.1299", "0.10986", "0.1365", "0.113", "0.1316", "0.1187", "0.125", "0.11554" ], [ "0.3677", "0.1111", "0.3823", "0.1111", "0.3823", "0.11554", "0.3677", "0.11554" ], [ "0.5107", "0.1111", "0.8325", "0.1111", "0.8325", "0.125", "0.5107", "0.125" ], [ "0.1528", "0.1225", "0.2532", "0.1225", "0.2532", "0.1338", "0.1528", "0.1338" ], [ "0.2695", "0.1225", "0.3694", "0.1225", "0.3694", "0.1338", "0.2695", "0.1338" ], [ "0.3857", "0.1225", "0.4836", "0.1225", "0.4836", "0.1338", "0.3857", "0.1338" ], [ "0.5312", "0.1276", "0.88", "0.1276", "0.88", "0.1414", "0.5312", "0.1414" ], [ "0.1487", "0.1395", "0.255", "0.1395", "0.255", "0.1497", "0.1487", "0.1497" ], [ "0.1299", "0.1276", "0.1356", "0.1276", "0.1356", "0.1338", "0.1299", "0.1338" ], [ "0.2688", "0.1395", "0.3694", "0.1395", "0.3694", "0.1497", "0.2688", "0.1497" ], [ "0.3857", "0.1395", "0.4836", "0.1395", "0.4836", "0.1497", "0.3857", "0.1497" ], [ "0.3708", "0.1414", "0.38", "0.1414", "0.38", "0.1465", "0.3708", "0.1465" ], [ "0.1259", "0.1414", "0.1422", "0.1414", "0.1422", "0.1483", "0.1259", "0.1483" ], [ "0.2532", "0.1433", "0.2607", "0.1433", "0.2607", "0.1465", "0.2532", "0.1465" ], [ "0.492", "0.1445", "0.5107", "0.1465", "0.5107", "0.1509", "0.492", "0.1483" ], [ "0.5073", "0.1433", "0.8813", "0.1445", "0.8813", "0.159", "0.5073", "0.1578" ], [ "0.1152", "0.154", "0.4885", "0.1547", "0.4885", "0.1692", "0.1152", "0.1686" ], [ "0.5107", "0.1622", "0.8813", "0.161", "0.8813", "0.1768", "0.5107", "0.178" ], [ "0.1152", "0.1718", "0.451", "0.1704", "0.451", "0.1849", "0.1152", "0.1863" ], [ "0.511", "0.1799", "0.8813", "0.1799", "0.8813", "0.1932", "0.511", "0.1932" ], [ "0.511", "0.1989", "0.88", "0.1989", "0.88", "0.209", "0.511", "0.209" ], [ "0.11926", "0.2058", "0.4902", "0.2058", "0.4902", "0.2203", "0.11926", "0.2203" ], [ "0.509", "0.2122", "0.8823", "0.2134", "0.8823", "0.2279", "0.509", "0.2267" ], [ "0.1177", "0.2247", "0.1896", "0.2247", "0.1896", "0.2361", "0.1177", "0.2361" ], [ "0.509", "0.2317", "0.638", "0.2299", "0.638", "0.2437", "0.509", "0.2456" ], [ "0.134", "0.2405", "0.4885", "0.2393", "0.4885", "0.254", "0.134", "0.2551" ], [ "0.527", "0.2474", "0.8813", "0.2474", "0.8813", "0.262", "0.527", "0.262" ], [ "0.11926", "0.2583", "0.4902", "0.2583", "0.4902", "0.2715", "0.11926", "0.2715" ], [ "0.509", "0.2644", "0.88", "0.2644", "0.88", "0.2778", "0.509", "0.2778" ], [ "0.11926", "0.2766", "0.4878", "0.2766", "0.4878", "0.2874", "0.11926", "0.2874" ], [ "0.5107", "0.2827", "0.8813", "0.2827", "0.8813", "0.2966", "0.5107", "0.2966" ], [ "0.1177", "0.2922", "0.4885", "0.2922", "0.4885", "0.3062", "0.1177", "0.3062" ], [ "0.511", "0.2998", "0.8813", "0.2998", "0.8813", "0.3132", "0.511", "0.3132" ], [ "0.1177", "0.31", "0.4028", "0.31", "0.4028", "0.324", "0.1177", "0.324" ], [ "0.511", "0.3176", "0.8813", "0.3176", "0.8813", "0.3308", "0.511", "0.3308" ], [ "0.1136", "0.3384", "0.3325", "0.339", "0.3325", "0.3542", "0.1136", "0.3528" ], [ "0.513", "0.334", "0.7695", "0.334", "0.7695", "0.348", "0.513", "0.348" ], [ "0.621", "0.3528", "0.7256", "0.3547", "0.7256", "0.3726", "0.621", "0.3706" ], [ "0.721", "0.358", "0.777", "0.3643", "0.773", "0.3818", "0.7183", "0.3757" ], [ "0.1168", "0.3655", "0.4885", "0.3662", "0.4885", "0.3806", "0.1168", "0.3801" ], [ "0.8594", "0.3655", "0.8823", "0.3655", "0.8823", "0.3801", "0.8594", "0.3801" ], [ "0.626", "0.3726", "0.7183", "0.375", "0.7183", "0.3894", "0.626", "0.387" ], [ "0.1168", "0.3838", "0.4878", "0.3838", "0.4878", "0.3977", "0.1168", "0.3977" ], [ "0.1152", "0.4016", "0.4878", "0.4016", "0.4878", "0.4153", "0.1152", "0.4153" ], [ "0.527", "0.3997", "0.8813", "0.3997", "0.8813", "0.4136", "0.527", "0.4136" ], [ "0.1168", "0.4185", "0.4885", "0.4172", "0.4885", "0.432", "0.1168", "0.433" ], [ "0.5073", "0.4167", "0.8813", "0.4167", "0.8813", "0.43", "0.5073", "0.43" ], [ "0.1316", "0.435", "0.4885", "0.435", "0.4885", "0.449", "0.1316", "0.449" ], [ "0.509", "0.433", "0.88", "0.433", "0.88", "0.4475", "0.509", "0.4475" ], [ "0.1168", "0.4526", "0.4885", "0.4526", "0.4885", "0.4666", "0.1168", "0.4666" ], [ "0.5107", "0.452", "0.8813", "0.452", "0.8813", "0.4653", "0.5107", "0.4653" ], [ "0.1177", "0.4702", "0.2075", "0.4702", "0.2075", "0.4817", "0.1177", "0.4817" ], [ "0.511", "0.4685", "0.8813", "0.4685", "0.8813", "0.4817", "0.511", "0.4817" ], [ "0.1356", "0.4873", "0.4885", "0.4873", "0.4885", "0.5005", "0.1356", "0.5005" ], [ "0.511", "0.4849", "0.8813", "0.4849", "0.8813", "0.4988", "0.511", "0.4988" ], [ "0.1168", "0.5024", "0.4885", "0.504", "0.4885", "0.518", "0.1168", "0.5176" ], [ "0.5107", "0.5024", "0.8823", "0.5024", "0.8823", "0.5166", "0.5107", "0.5166" ], [ "0.1168", "0.5215", "0.4885", "0.5215", "0.4885", "0.535", "0.1168", "0.535" ], [ "0.509", "0.5195", "0.7466", "0.5195", "0.7466", "0.5327", "0.509", "0.5327" ], [ "0.1177", "0.5386", "0.4885", "0.5386", "0.4885", "0.552", "0.1177", "0.552" ], [ "0.5293", "0.536", "0.8813", "0.536", "0.8813", "0.55", "0.5293", "0.55" ], [ "0.1168", "0.556", "0.4902", "0.556", "0.4902", "0.5693", "0.1168", "0.5693" ], [ "0.513", "0.5547", "0.88", "0.5547", "0.88", "0.565", "0.513", "0.565" ], [ "0.1168", "0.5713", "0.4902", "0.5728", "0.4902", "0.587", "0.1168", "0.586" ], [ "0.5107", "0.571", "0.8813", "0.571", "0.8813", "0.584", "0.5107", "0.584" ], [ "0.1177", "0.589", "0.4902", "0.589", "0.4902", "0.603", "0.1177", "0.603" ], [ "0.511", "0.5884", "0.8813", "0.5884", "0.8813", "0.6016", "0.511", "0.6016" ], [ "0.1177", "0.607", "0.4902", "0.607", "0.4902", "0.6206", "0.1177", "0.6206" ], [ "0.5107", "0.606", "0.8813", "0.606", "0.8813", "0.619", "0.5107", "0.619" ], [ "0.1259", "0.6235", "0.4902", "0.6235", "0.4902", "0.637", "0.1259", "0.637" ], [ "0.5107", "0.6226", "0.8813", "0.6226", "0.8813", "0.636", "0.5107", "0.636" ], [ "0.1152", "0.6416", "0.4478", "0.6416", "0.4478", "0.655", "0.1152", "0.655" ], [ "0.5107", "0.6396", "0.568", "0.6396", "0.568", "0.654", "0.5107", "0.654" ], [ "0.1356", "0.659", "0.4885", "0.659", "0.4885", "0.6724", "0.1356", "0.6724" ], [ "0.509", "0.6694", "0.5483", "0.6694", "0.5483", "0.6797", "0.509", "0.6797" ], [ "0.544", "0.6685", "0.789", "0.6685", "0.789", "0.6816", "0.544", "0.6816" ], [ "0.1168", "0.6753", "0.4885", "0.6753", "0.4885", "0.6895", "0.1168", "0.6895" ], [ "0.1136", "0.6914", "0.4902", "0.6924", "0.4902", "0.7056", "0.1136", "0.705" ], [ "0.511", "0.696", "0.8813", "0.696", "0.8813", "0.7104", "0.511", "0.7104" ], [ "0.1177", "0.7104", "0.4902", "0.7104", "0.4902", "0.7236", "0.1177", "0.7236" ], [ "0.5063", "0.711", "0.88", "0.712", "0.88", "0.7266", "0.5063", "0.726" ], [ "0.1152", "0.726", "0.4885", "0.7266", "0.4885", "0.741", "0.1152", "0.7407" ], [ "0.511", "0.73", "0.8813", "0.73", "0.8813", "0.7437", "0.511", "0.7437" ], [ "0.1177", "0.744", "0.4885", "0.744", "0.4885", "0.7583", "0.1177", "0.7583" ], [ "0.509", "0.747", "0.88", "0.747", "0.88", "0.7603", "0.509", "0.7603" ], [ "0.1177", "0.761", "0.4885", "0.761", "0.4885", "0.7744", "0.1177", "0.7744" ], [ "0.5073", "0.763", "0.8813", "0.7646", "0.8813", "0.779", "0.5073", "0.778" ], [ "0.1168", "0.7764", "0.511", "0.778", "0.511", "0.7925", "0.1168", "0.7915" ], [ "0.5107", "0.782", "0.8823", "0.782", "0.8823", "0.7954", "0.5107", "0.7954" ], [ "0.1152", "0.7944", "0.1602", "0.7944", "0.1602", "0.8096", "0.1152", "0.8096" ], [ "0.5107", "0.7974", "0.8813", "0.7974", "0.8813", "0.8125", "0.5107", "0.8125" ], [ "0.1356", "0.813", "0.4885", "0.813", "0.4885", "0.827", "0.1356", "0.827" ], [ "0.5073", "0.815", "0.66", "0.815", "0.66", "0.829", "0.5073", "0.829" ], [ "0.1177", "0.83", "0.4885", "0.83", "0.4885", "0.8433", "0.1177", "0.8433" ], [ "0.1209", "0.8477", "0.4885", "0.8477", "0.4885", "0.8613", "0.1209", "0.8613" ], [ "0.5073", "0.858", "0.6006", "0.86", "0.6006", "0.875", "0.5073", "0.8726" ], [ "0.11926", "0.8643", "0.4902", "0.8643", "0.4902", "0.8774", "0.11926", "0.8774" ], [ "0.1152", "0.881", "0.4885", "0.881", "0.4885", "0.896", "0.1152", "0.896" ], [ "0.4927", "0.887", "0.505", "0.887", "0.505", "0.89", "0.4927", "0.89" ], [ "0.5063", "0.8833", "0.88", "0.8833", "0.88", "0.898", "0.5063", "0.898" ], [ "0.1168", "0.8984", "0.492", "0.8984", "0.492", "0.912", "0.1168", "0.912" ], [ "0.527", "0.8984", "0.88", "0.8984", "0.88", "0.912", "0.527", "0.912" ] ]

[ "METEOR", "BLEU", "Skip-Thoughts", "贪婪字幕模型比波束搜索模型获得的质量更差。", "0.22 (2.8e-28)", "0.08 (1.0e-06)", "0.18 (5.0e-27)", "", "T", "", "这突出了生成故事与生成字幕的关键区别", "0.20 (3.0e-31)", "0.08 (8.9e-06)", "0.16(6.4e-22)", "尽管使用贪婪搜索产生的故事会带来可观的", "0.14 (1.0e-33)", "0", "0.06(8.7e-08)", "0.11 (7.7e-24)", "", "T", "", "", "收益，但它们包含许多重复的单词和短语，例", "Table 4: Correlations of automatic scores against hu-", "如。g。，“孩子们玩得很开心。和孩子们有一个", "man judgements, with p-values in parentheses.", "伟大的时间。“我们引入了一种非常简单的启发", "式方法来避免这种情况，即在给定的故事中不能", "了Sentence2Vec嵌人，可以对小说的语义空间", "多次产生相同的内容词。这将使METEOR再", "进行建模", "提高2。3points", "如表4所示，根据所有相关系数，METEOR与", "将字幕与并排讲故事进行比较的一个优点是，", "人类判断的相关性最佳。这表明，诸如METEOR", "字幕输出可用于帮助通知讲故事输出。为此，我", "之类的度量标准结合了释义，与人类对该任务的", "们提供了一个额外的基线，其中只有在字幕模型", "判断最相关。。对故事的自动评估进行更详细的", "许可的情况下才能产生“视觉上接地”的单词。", "研究是未来工作的一个感兴趣的领域", "我们将视觉基础单词的集合定义为在字幕训练", "6BaselineExperiments", "中出现频率高于故事训练的单词：", "P(w|Tcaption)", "2>1.0", "我们在表7中报告了讲故事任务的基线实验，在", "(1)", "P(wTstory)", "SIS层上进行了训练，并在一半的SIS验证集", "（valtest）上进行了测试。每个系统的示例输出", "我们使用字幕注释训练一个单独的模型，并为", "如表5所示。为了突出故事和字幕生成之间的", "valtest集中的每个图像生成n个最佳字幕列表。", "些差异，我们还单独在DII层上进行训练，并", "在100个最佳列表中至少10个句子中看到的单", "按图像生成字幕，而不是按顺序生成。这些结果", "词被字幕模型标记为“许可”。没有重复的贪婪", "如表7所示", "解码继续进行额外的约束，即如果一个词在视觉", "为了训练故事生成模型，我们使用了序列到序", "上是接地的，那么它只能由故事模型生成，如果", "列递归神经网络（RNN）方法，该方法自然地将", "它是由相同照片集的字幕模型许可的。这将导致", "Devlin et al.(2015)和 Vinyals et al.(2014)的", "进一步的1。3METEOR改进", "单幅图像字幕技术扩展到了多幅图像。在这里，", "有趣的是，注意到相对简单的启发式方法对生", "我们通过在每个图像的fc7向量上以相反的顺", "成的故事有什么强刻的影响。我们并不打算建议", "序运行RNN来编码图像sequence。这被用作", "这些启发式方法是right的方式来处理故事生", "故事解码器模型的初始隐藏状态，该模型学习使", "成。相反，主要目的是提供清晰的基线，以证明", "用训练数据词汇表上的softmax损失一次生成", "故事生成与字幕生成具有根本不同的挑战；并且", "个故事。我们将门控循环单位（GRU）（Cho", "空间广阔，可以探索训练和解码方法以生成流畅", "etal.，2014）用于图像编码器和故事解码器", "的故事", "在基线系统中，我们使用简单的光束搜索（大", "7", "Conclusion and Future Work", "小=10）生成故事，该故事已在以前的（Devlin", "etal.,2015）中成功进行了图像字幕。然而，对于", "我们为sequentialvision-to-language引人了第", "故事生成，该模型的结果主观上似乎很差一系", "一个数据集，该数据集从隔离的图像逐渐移动到", "统产生通用的、重复的、高级的描述（e。g。，“这", "顺序的故事。我们认为，对这个数据集中捕获的", "是一只狗的照片”）。考虑到最大似然训练中固有", "更具比喻性和社交性的语言进行建模对于将人", "的标签偏差问题，这是一个可预测的结果；最近", "工智能发展到更像人类的理解至关重要。我们已", "的工作着眼于直接解决此问题的方法（Lietal.，", "经为视觉故事讲述任务建立了几个强有力的基", "2016)", "线，并激励METEOR作为一个自动指标来评", "为了建立更强大的基线，我们探索了几种解码", "估这项任务的进展。", "时间启发式方法，以提高生成的故事的质量。第", "一个试探法是大幅降低解码器波束大小。我们", "References", "发现，使用光束大小为1（贪婪搜索）可以显着", "提高故事质量，从而产生4。6METEOR得分增", "", "Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu", "加。然而，对于字幕生成没有看到相同的效果，", "Margaret Mitchell, Dhruv Batra, C. Lawrence Zit-" ]

{ "arxiv_id": "1604.03968", "language": "zh", "page_idx": 3 }

[ [ "0.1283", "0.06946", "0.1929", "0.06946", "0.1929", "0.0808", "0.1283", "0.0808" ], [ "0.4453", "0.0663", "0.585", "0.06757", "0.585", "0.0821", "0.4453", "0.0808" ], [ "0.819", "0.2311", "0.8364", "0.2311", "0.8364", "0.2349", "0.819", "0.2349" ], [ "0.853", "0.226", "0.902", "0.2279", "0.902", "0.238", "0.853", "0.2361" ], [ "0.818", "0.2551", "0.899", "0.2551", "0.899", "0.262", "0.818", "0.262" ], [ "0.08496", "0.3882", "0.911", "0.387", "0.911", "0.4016", "0.08496", "0.4028" ], [ "0.08826", "0.404", "0.91", "0.404", "0.91", "0.414", "0.08826", "0.414" ], [ "0.0866", "0.4167", "0.152", "0.4185", "0.152", "0.43", "0.0866", "0.428" ], [ "0.0866", "0.457", "0.4763", "0.4583", "0.4763", "0.473", "0.0866", "0.4717" ], [ "0.523", "0.4583", "0.911", "0.4583", "0.911", "0.473", "0.523", "0.473" ], [ "0.08826", "0.4717", "0.478", "0.4717", "0.478", "0.4849", "0.08826", "0.4849" ], [ "0.523", "0.4717", "0.911", "0.4717", "0.911", "0.486", "0.523", "0.486" ], [ "0.08826", "0.486", "0.3489", "0.486", "0.3489", "0.5", "0.08826", "0.5" ], [ "0.5215", "0.486", "0.911", "0.486", "0.911", "0.5", "0.5215", "0.5" ], [ "0.1054", "0.5005", "0.4763", "0.5005", "0.4763", "0.5146", "0.1054", "0.5146" ], [ "0.5215", "0.5", "0.911", "0.5", "0.911", "0.513", "0.5215", "0.513" ], [ "0.0907", "0.5166", "0.4763", "0.5166", "0.4763", "0.5264", "0.0907", "0.5264" ], [ "0.524", "0.515", "0.7974", "0.515", "0.7974", "0.526", "0.524", "0.526" ], [ "0.08905", "0.529", "0.478", "0.529", "0.478", "0.5425", "0.08905", "0.5425" ], [ "0.08826", "0.5425", "0.478", "0.5425", "0.478", "0.556", "0.08826", "0.556" ], [ "0.0866", "0.5547", "0.4763", "0.5537", "0.4763", "0.5684", "0.0866", "0.5693" ], [ "0.539", "0.5596", "0.9067", "0.5596", "0.9067", "0.5693", "0.539", "0.5693" ], [ "0.08826", "0.5693", "0.4763", "0.5693", "0.4763", "0.5835", "0.08826", "0.5835" ], [ "0.523", "0.571", "0.911", "0.571", "0.911", "0.5854", "0.523", "0.5854" ], [ "0.08905", "0.584", "0.4756", "0.584", "0.4756", "0.5977", "0.08905", "0.5977" ], [ "0.5215", "0.5835", "0.911", "0.584", "0.911", "0.5986", "0.5215", "0.5977" ], [ "0.08905", "0.5967", "0.4739", "0.5967", "0.4739", "0.6113", "0.08905", "0.6113" ], [ "0.523", "0.5986", "0.911", "0.5986", "0.911", "0.6123", "0.523", "0.6123" ], [ "0.08905", "0.6113", "0.4763", "0.6113", "0.4763", "0.625", "0.08905", "0.625" ], [ "0.5215", "0.6123", "0.91", "0.6113", "0.91", "0.625", "0.5215", "0.6255" ], [ "0.0866", "0.625", "0.478", "0.6235", "0.478", "0.638", "0.0866", "0.6396" ], [ "0.523", "0.6255", "0.911", "0.6255", "0.911", "0.6396", "0.523", "0.6396" ], [ "0.08826", "0.6396", "0.478", "0.6396", "0.478", "0.653", "0.08826", "0.653" ], [ "0.523", "0.64", "0.911", "0.64", "0.911", "0.654", "0.523", "0.654" ], [ "0.08905", "0.653", "0.478", "0.653", "0.478", "0.6665", "0.08905", "0.6665" ], [ "0.524", "0.654", "0.911", "0.654", "0.911", "0.6675", "0.524", "0.6675" ], [ "0.0866", "0.6665", "0.4763", "0.6655", "0.4763", "0.6797", "0.0866", "0.681" ], [ "0.523", "0.6655", "0.911", "0.6665", "0.911", "0.681", "0.523", "0.6797" ], [ "0.08826", "0.6797", "0.4756", "0.6797", "0.4756", "0.6943", "0.08826", "0.6943" ], [ "0.523", "0.6797", "0.911", "0.681", "0.911", "0.696", "0.523", "0.6943" ], [ "0.08826", "0.6943", "0.4763", "0.6943", "0.4763", "0.7085", "0.08826", "0.7085" ], [ "0.523", "0.696", "0.911", "0.696", "0.911", "0.709", "0.523", "0.709" ], [ "0.08826", "0.709", "0.2051", "0.709", "0.2051", "0.7227", "0.08826", "0.7227" ], [ "0.5337", "0.709", "0.911", "0.709", "0.911", "0.7227", "0.5337", "0.7227" ], [ "0.10455", "0.7236", "0.4272", "0.7236", "0.4272", "0.7373", "0.10455", "0.7373" ], [ "0.4207", "0.7246", "0.4739", "0.7246", "0.4739", "0.7363", "0.4207", "0.7363" ], [ "0.5195", "0.7217", "0.911", "0.7227", "0.911", "0.7373", "0.5195", "0.7363" ], [ "0.08826", "0.7373", "0.478", "0.7373", "0.478", "0.7505", "0.08826", "0.7505" ], [ "0.523", "0.7373", "0.911", "0.7373", "0.911", "0.7505", "0.523", "0.7505" ], [ "0.08826", "0.7505", "0.4763", "0.7505", "0.4763", "0.7646", "0.08826", "0.7646" ], [ "0.523", "0.7505", "0.9126", "0.7505", "0.9126", "0.7646", "0.523", "0.7646" ], [ "0.08905", "0.765", "0.4763", "0.765", "0.4763", "0.779", "0.08905", "0.779" ], [ "0.524", "0.765", "0.911", "0.765", "0.911", "0.779", "0.524", "0.779" ], [ "0.08826", "0.779", "0.4763", "0.779", "0.4763", "0.7925", "0.08826", "0.7925" ], [ "0.523", "0.778", "0.911", "0.778", "0.911", "0.7915", "0.523", "0.7915" ], [ "0.08905", "0.7925", "0.4763", "0.7925", "0.4763", "0.806", "0.08905", "0.806" ], [ "0.523", "0.7925", "0.9126", "0.7925", "0.9126", "0.806", "0.523", "0.806" ], [ "0.08905", "0.806", "0.4763", "0.806", "0.4763", "0.8193", "0.08905", "0.8193" ], [ "0.523", "0.806", "0.911", "0.806", "0.911", "0.8193", "0.523", "0.8193" ], [ "0.08826", "0.8184", "0.4763", "0.8193", "0.4763", "0.834", "0.08826", "0.8335" ], [ "0.5215", "0.8193", "0.91", "0.8193", "0.91", "0.8335", "0.5215", "0.8335" ], [ "0.08905", "0.834", "0.478", "0.834", "0.478", "0.8477", "0.08905", "0.8477" ], [ "0.523", "0.8335", "0.9126", "0.8335", "0.9126", "0.8477", "0.523", "0.8477" ], [ "0.08826", "0.8477", "0.478", "0.8477", "0.478", "0.8613", "0.08826", "0.8613" ], [ "0.5215", "0.845", "0.91", "0.8477", "0.91", "0.8623", "0.5215", "0.86" ], [ "0.08826", "0.8613", "0.4756", "0.8613", "0.4756", "0.875", "0.08826", "0.875" ], [ "0.523", "0.8623", "0.911", "0.8623", "0.911", "0.8755", "0.523", "0.8755" ], [ "0.08826", "0.8755", "0.478", "0.8755", "0.478", "0.8896", "0.08826", "0.8896" ], [ "0.535", "0.8755", "0.8", "0.8755", "0.8", "0.8896", "0.535", "0.8896" ] ]

[ "Original", "Random syntheses", "E", "", "", "Figure 2: The first column refers to the original images, and the rest are random syntheses by PetsGAN. We can", "observethat PetsGAN cansynthesize sampleswith arbitrary sizes.reasonablestructure,finetexture.and great visual", "diversity.", "学习(Ulyanov,Vedaldi, and Lempitsky 2017; Shocher,", "遵循SinGAN(Shaham,Dekel,andMichaeli2019)的", "Cohen，andIrani2018）由于只需要测试图像本身而无", "范式，以配备用于内部补丁学习的生成器和补丁鉴别器", "需额外的监督，因此受到了广泛关注", "的金字塔。尽管金字塔网络（Dentonetal.2015；Karras", "不久前，深度内部学习令人惊地进入了单个图像生", "etal.2017）的渐进式或级联式学习降低了模型优化的", "成的任务。与合成具有重复纹理模式的图像的单个图像", "难度，但训练耗时且容易造成伪影积累", "纹理生成模型（Jetchev，Bergmann，andVollgraf2016;", "Zhouetal.20i8）不同，单个图像生成模型（SIG模", "型)(Shaham,Dekel, and Michaeli 2019: Shocher et al", "在本文中，我们专注于SIG任务中的上述突出问题", "2019）是从自然图像中学习的，并从给定图像中生成相", "并做出以下贡献：？??）我们从本质上深入研究了SIG", "似但不同的样本，其中前者表示相似的视觉内容，后者", "问题，即从由单个图像组成的狄拉克分布中学习多样", "表示不同的布局和外观。开创性的 SinGAN（Shaham", "性分布。显然，这是一个不适定的恢复问题，适当的先", "Dekel，andMichaeli2019）是第一个无条件的SIG模", "验和正则化项是解决它的关键。因此，我们构建了二个", "型，它使用完全卷积生成器和补丁鉴别器构建了GAN", "统一的正则化潜变量模型来制定SIG任务。据我们所", "的拉普拉斯金字塔，以学习单个图像中的不同尺度信", "知，这是第一次给这个任务一个明确的数学定义和优化", "息。SinGAN的出色表现证明了内部学习在生成任务上", "目标。所有现有的SIG模型都可以视为该模型的特例。", "的可行性。从那时起，许多SIG模型（Hinzetal.2021", "???）我们设计了一种新颖的基于Prior的端到端训练", "Chen et al. 2021; Zhang, Han, and Guo 2021; Sushko.", "单映像GAN（PetsGAN），该映像注入了内部先验和外", "Gall,andKhoreva2021;Granot et al.2021)被提出沿", "部先验，以克服SinGAN和派生方法的问题（图。1）。", "着这条路线探索", "方面，我们将通过GAN反演获得的外部先验注人到", "虽然在这一领域取得了很大的突破和成就，", "但仍有", "DEPNet中，以缓解缺乏用于生成自然，合理和多样化", "许多问题需要解决。首先，由于仅获取给定图像的内部", "样本的高级信息的问题，即使对于对象图像也是如此。", "纹理信息，以前的SIG模型通常会生成混沌结构，这", "另一方面，我们用DIPNet注人补丁分布校正的内部", "在对象图像上尤其严重，因为对象部分之间的语义和", "先验，以简化和加快补丁分布学习。得益于充分利用外", "结构关系对于合成合理的图像非常重要。为了缓解这", "部和内部先验，我们的方法摆脱了耗时的渐进式培训方", "个问题，（Wangetal.202la）利用预先训练的分类模型", "案，可以端到端进行培训。？？？）我们在各种数据集上构", "为图像生成提供语义信息。但是，监督过于粗糙，无法", "建了丰富的定性和定量实验。Fig。2显示了PetsGAN", "指导高保真图像生成，并且模型泛化受到限制，因为", "的一些合成，实验比较结果表明，我们的方法在图像质", "1）每个生成的样本都应分别进行优化，并且2）不同", "量，多样性和训练速度方面明显优于其他SIG模型。此", "的图像类型需要不同的分类模型。其次，儿乎所有SIG", "外，我们将我们的方法应用于其他应用程序（例如。g。", "模型 (Shaham, Dekel, and Michaeli 2019; Hinz et al", "快速训练高分辨率图像生成，风格转移，协调)。结果进", "2021; Chen et al. 2021; Zhang, Han, and Guo 2021)都", "步证明了我们方法的有效性和高效性" ]

{ "arxiv_id": "2203.01488", "language": "zh", "page_idx": 3 }

[ [ "0.2035", "0.0682", "0.3594", "0.0682", "0.3594", "0.0821", "0.2035", "0.0821" ], [ "0.556", "0.06757", "0.572", "0.06946", "0.5703", "0.0745", "0.5557", "0.0726" ], [ "0.786", "0.0682", "0.8105", "0.0682", "0.8105", "0.07574", "0.786", "0.07574" ], [ "0.8486", "0.0707", "0.876", "0.0707", "0.876", "0.0745", "0.8486", "0.0745" ], [ "0.601", "0.084", "0.6167", "0.084", "0.6167", "0.09155", "0.601", "0.09155" ], [ "0.694", "0.0808", "0.714", "0.0808", "0.714", "0.0884", "0.694", "0.0884" ], [ "0.09235", "0.09155", "0.4763", "0.09155", "0.4763", "0.1061", "0.09235", "0.1061" ], [ "0.819", "0.09344", "0.8364", "0.09344", "0.8364", "0.09784", "0.819", "0.09784" ], [ "0.08905", "0.1048", "0.4763", "0.1048", "0.4763", "0.1187", "0.08905", "0.1187" ], [ "0.7573", "0.1016", "0.7656", "0.1016", "0.7656", "0.1061", "0.7573", "0.1061" ], [ "0.699", "0.10986", "0.7114", "0.10986", "0.7114", "0.11615", "0.699", "0.11615" ], [ "0.08826", "0.1187", "0.4763", "0.11743", "0.4763", "0.132", "0.08826", "0.1332" ], [ "0.6772", "0.1256", "0.687", "0.1256", "0.687", "0.1332", "0.6772", "0.1332" ], [ "0.08826", "0.1332", "0.4763", "0.1332", "0.4763", "0.1465", "0.08826", "0.1465" ], [ "0.7363", "0.1288", "0.7656", "0.1288", "0.7656", "0.137", "0.7363", "0.137" ], [ "0.5728", "0.1351", "0.615", "0.1351", "0.615", "0.1459", "0.5728", "0.1459" ], [ "0.6895", "0.1395", "0.7183", "0.1395", "0.7183", "0.1465", "0.6895", "0.1465" ], [ "0.8096", "0.1338", "0.818", "0.137", "0.8145", "0.1401", "0.806", "0.137" ], [ "0.08905", "0.1465", "0.4756", "0.1465", "0.4756", "0.1604", "0.08905", "0.1604" ], [ "0.08826", "0.1604", "0.4763", "0.1604", "0.4763", "0.1736", "0.08826", "0.1736" ], [ "0.5215", "0.1666", "0.911", "0.1666", "0.911", "0.1812", "0.5215", "0.1812" ], [ "0.08905", "0.1749", "0.4763", "0.1749", "0.4763", "0.1881", "0.08905", "0.1881" ], [ "0.523", "0.1818", "0.911", "0.1818", "0.911", "0.1925", "0.523", "0.1925" ], [ "0.08826", "0.1881", "0.478", "0.1881", "0.478", "0.202", "0.08826", "0.202" ], [ "0.523", "0.1957", "0.91", "0.1957", "0.91", "0.2058", "0.523", "0.2058" ], [ "0.08826", "0.202", "0.478", "0.202", "0.478", "0.2153", "0.08826", "0.2153" ], [ "0.5215", "0.2084", "0.91", "0.2084", "0.91", "0.2216", "0.5215", "0.2216" ], [ "0.08826", "0.2166", "0.4763", "0.2166", "0.4763", "0.2299", "0.08826", "0.2299" ], [ "0.5215", "0.2216", "0.9126", "0.2216", "0.9126", "0.2361", "0.5215", "0.2361" ], [ "0.08826", "0.2299", "0.478", "0.2299", "0.478", "0.2443", "0.08826", "0.2443" ], [ "0.5215", "0.2349", "0.911", "0.2349", "0.911", "0.2488", "0.5215", "0.2488" ], [ "0.0866", "0.2437", "0.4756", "0.2437", "0.4756", "0.2568", "0.0866", "0.2568" ], [ "0.5195", "0.2488", "0.911", "0.25", "0.911", "0.2644", "0.5195", "0.2632" ], [ "0.08826", "0.2568", "0.4763", "0.2568", "0.4763", "0.2708", "0.08826", "0.2708" ], [ "0.5215", "0.2644", "0.9126", "0.2644", "0.9126", "0.2778", "0.5215", "0.2778" ], [ "0.0866", "0.2708", "0.478", "0.2695", "0.478", "0.2842", "0.0866", "0.2854" ], [ "0.5195", "0.2778", "0.6167", "0.279", "0.6167", "0.2903", "0.5195", "0.289" ], [ "0.08826", "0.2854", "0.4763", "0.2854", "0.4763", "0.2986", "0.08826", "0.2986" ], [ "0.0866", "0.2974", "0.4763", "0.2986", "0.4763", "0.3132", "0.0866", "0.3125" ], [ "0.08826", "0.3132", "0.478", "0.3132", "0.478", "0.327", "0.08826", "0.327" ], [ "0.528", "0.3208", "0.774", "0.3208", "0.774", "0.3308", "0.528", "0.3308" ], [ "0.774", "0.3208", "0.91", "0.3208", "0.91", "0.334", "0.774", "0.334" ], [ "0.523", "0.334", "0.911", "0.334", "0.911", "0.3484", "0.523", "0.3484" ], [ "0.08905", "0.3257", "0.4495", "0.3257", "0.4495", "0.339", "0.08905", "0.339" ], [ "0.08826", "0.3403", "0.4763", "0.3403", "0.4763", "0.3542", "0.08826", "0.3542" ], [ "0.08826", "0.3528", "0.478", "0.3528", "0.478", "0.3674", "0.08826", "0.3674" ], [ "0.5215", "0.348", "0.569", "0.3516", "0.568", "0.3643", "0.5195", "0.3599" ], [ "0.6665", "0.3643", "0.7656", "0.3643", "0.7656", "0.3757", "0.6665", "0.3757" ], [ "0.8896", "0.3643", "0.9126", "0.3643", "0.9126", "0.3774", "0.8896", "0.3774" ], [ "0.08496", "0.3806", "0.4795", "0.3818", "0.4795", "0.3965", "0.08496", "0.3958" ], [ "0.0866", "0.3662", "0.4795", "0.3687", "0.4795", "0.3833", "0.0866", "0.3806" ], [ "0.08826", "0.3958", "0.478", "0.3958", "0.478", "0.4104", "0.08826", "0.4104" ], [ "0.523", "0.3914", "0.911", "0.3914", "0.911", "0.406", "0.523", "0.406" ], [ "0.08905", "0.4092", "0.4795", "0.4092", "0.4795", "0.4224", "0.08905", "0.4224" ], [ "0.5195", "0.4048", "0.9126", "0.404", "0.9126", "0.4185", "0.5195", "0.4192" ], [ "0.08826", "0.4236", "0.2688", "0.4236", "0.2688", "0.4375", "0.08826", "0.4375" ], [ "0.5215", "0.4192", "0.911", "0.4192", "0.911", "0.433", "0.5215", "0.433" ], [ "0.5195", "0.433", "0.911", "0.432", "0.911", "0.4463", "0.5195", "0.4475" ], [ "0.09314", "0.449", "0.4763", "0.449", "0.4763", "0.4634", "0.09314", "0.4634" ], [ "0.5215", "0.4463", "0.911", "0.4463", "0.911", "0.4602", "0.5215", "0.4602" ], [ "0.08905", "0.4634", "0.4763", "0.4634", "0.4763", "0.4766", "0.08905", "0.4766" ], [ "0.5195", "0.4602", "0.911", "0.459", "0.911", "0.4734", "0.5195", "0.4749" ], [ "0.08826", "0.4766", "0.4756", "0.4766", "0.4756", "0.4905", "0.08826", "0.4905" ], [ "0.536", "0.4749", "0.9126", "0.4749", "0.9126", "0.488", "0.536", "0.488" ], [ "0.08905", "0.4924", "0.4763", "0.4924", "0.4763", "0.5024", "0.08905", "0.5024" ], [ "0.5215", "0.4873", "0.6216", "0.4893", "0.6216", "0.5024", "0.5215", "0.5005" ], [ "0.08826", "0.5024", "0.4763", "0.504", "0.4763", "0.5176", "0.08826", "0.5166" ], [ "0.527", "0.511", "0.911", "0.511", "0.911", "0.5244", "0.527", "0.5244" ], [ "0.08905", "0.5195", "0.4763", "0.5195", "0.4763", "0.53", "0.08905", "0.53" ], [ "0.523", "0.526", "0.911", "0.526", "0.911", "0.539", "0.523", "0.539" ], [ "0.0866", "0.532", "0.4763", "0.532", "0.4763", "0.5454", "0.0866", "0.5454" ], [ "0.5186", "0.5386", "0.9062", "0.539", "0.9062", "0.5566", "0.5186", "0.556" ], [ "0.08905", "0.5474", "0.4739", "0.5474", "0.4739", "0.558", "0.08905", "0.558" ], [ "0.9004", "0.5454", "0.9087", "0.5454", "0.9087", "0.55", "0.9004", "0.55" ], [ "0.5186", "0.552", "0.9126", "0.556", "0.9126", "0.5737", "0.5186", "0.5693" ], [ "0.0907", "0.561", "0.4756", "0.561", "0.4756", "0.5713", "0.0907", "0.5713" ], [ "0.08905", "0.5737", "0.4763", "0.5737", "0.4763", "0.587", "0.08905", "0.587" ], [ "0.818", "0.5684", "0.8384", "0.5684", "0.8384", "0.5757", "0.818", "0.5757" ], [ "0.5376", "0.5776", "0.91", "0.5776", "0.91", "0.5913", "0.5376", "0.5913" ], [ "0.0866", "0.586", "0.4763", "0.587", "0.4763", "0.6016", "0.0866", "0.601" ], [ "0.5215", "0.5923", "0.9126", "0.5923", "0.9126", "0.606", "0.5215", "0.606" ], [ "0.0866", "0.601", "0.4756", "0.601", "0.4756", "0.6143", "0.0866", "0.6143" ], [ "0.5215", "0.607", "0.9126", "0.607", "0.9126", "0.6206", "0.5215", "0.6206" ], [ "0.08496", "0.6143", "0.4763", "0.613", "0.4763", "0.6274", "0.08496", "0.629" ], [ "0.5186", "0.622", "0.8203", "0.6206", "0.8203", "0.6353", "0.5186", "0.636" ], [ "0.08905", "0.629", "0.478", "0.629", "0.478", "0.6426", "0.08905", "0.6426" ], [ "0.08905", "0.6445", "0.4756", "0.6445", "0.4756", "0.655", "0.08905", "0.655" ], [ "0.5195", "0.6484", "0.8325", "0.6484", "0.8325", "0.662", "0.5195", "0.662" ], [ "0.08905", "0.657", "0.4756", "0.657", "0.4756", "0.6704", "0.08905", "0.6704" ], [ "0.08826", "0.6704", "0.478", "0.6704", "0.478", "0.6846", "0.08826", "0.6846" ], [ "0.5254", "0.6694", "0.91", "0.6694", "0.91", "0.6797", "0.5254", "0.6797" ], [ "0.0866", "0.6846", "0.3823", "0.6846", "0.3823", "0.6978", "0.0866", "0.6978" ], [ "0.523", "0.6816", "0.91", "0.6816", "0.91", "0.696", "0.523", "0.696" ], [ "0.523", "0.6963", "0.91", "0.6963", "0.91", "0.7104", "0.523", "0.7104" ], [ "0.2059", "0.714", "0.353", "0.7163", "0.353", "0.731", "0.2059", "0.729" ], [ "0.523", "0.7104", "0.911", "0.7085", "0.911", "0.7227", "0.523", "0.7246" ], [ "0.523", "0.7236", "0.911", "0.7236", "0.911", "0.7373", "0.523", "0.7373" ], [ "0.08496", "0.7393", "0.2385", "0.7393", "0.2385", "0.7524", "0.08496", "0.7524" ], [ "0.523", "0.738", "0.6626", "0.738", "0.6626", "0.752", "0.523", "0.752" ], [ "0.0907", "0.7603", "0.4756", "0.762", "0.4756", "0.776", "0.0907", "0.7734" ], [ "0.524", "0.759", "0.681", "0.7603", "0.681", "0.7734", "0.524", "0.773" ], [ "0.6772", "0.7603", "0.911", "0.7603", "0.911", "0.7734", "0.6772", "0.7734" ], [ "0.2214", "0.7734", "0.4763", "0.7744", "0.4763", "0.789", "0.2214", "0.7886" ], [ "0.5215", "0.773", "0.911", "0.7734", "0.911", "0.7886", "0.5215", "0.787" ], [ "0.0866", "0.7886", "0.478", "0.7886", "0.478", "0.803", "0.0866", "0.803" ], [ "0.0866", "0.7744", "0.2296", "0.7734", "0.2296", "0.7886", "0.0866", "0.789" ], [ "0.5215", "0.7886", "0.911", "0.7886", "0.911", "0.802", "0.5215", "0.802" ], [ "0.08826", "0.803", "0.4763", "0.803", "0.4763", "0.8174", "0.08826", "0.8174" ], [ "0.5186", "0.8003", "0.911", "0.802", "0.911", "0.8164", "0.5186", "0.815" ], [ "0.08826", "0.8174", "0.478", "0.8174", "0.478", "0.8306", "0.08826", "0.8306" ], [ "0.5215", "0.815", "0.911", "0.8164", "0.911", "0.8306", "0.5215", "0.83" ], [ "0.08826", "0.8306", "0.1217", "0.829", "0.12335", "0.8403", "0.08905", "0.8423" ], [ "0.134", "0.8306", "0.4763", "0.8306", "0.4763", "0.845", "0.134", "0.845" ], [ "0.536", "0.829", "0.91", "0.829", "0.91", "0.8433", "0.536", "0.8433" ], [ "0.1177", "0.839", "0.1323", "0.839", "0.1323", "0.8433", "0.1177", "0.8433" ], [ "0.08826", "0.845", "0.4763", "0.845", "0.4763", "0.8594", "0.08826", "0.8594" ], [ "0.5215", "0.8433", "0.911", "0.8433", "0.911", "0.857", "0.5215", "0.857" ], [ "0.08905", "0.86", "0.4756", "0.86", "0.4756", "0.8735", "0.08905", "0.8735" ], [ "0.523", "0.8594", "0.5786", "0.8594", "0.5786", "0.8706", "0.523", "0.8706" ], [ "0.6445", "0.8735", "0.7876", "0.8735", "0.7876", "0.887", "0.6445", "0.887" ], [ "0.892", "0.8726", "0.9126", "0.8726", "0.9126", "0.887", "0.892", "0.887" ], [ "0.0866", "0.8755", "0.2793", "0.8755", "0.2793", "0.8896", "0.0866", "0.8896" ] ]

[ "2Related Work", "ate", "Patche", "resolm", "福", "-IR-", "Singleimagegenerationmodel首先对单个图像的内", "", "部补丁分布进行建模，SIG模型旨在生成具有高纹理质", "", "IR", "量的新的合理图像。InGAN（Shocheretal.2019）首次", "PT", "提出了具有编码器-解码器架构的条件SIG模型，以使", "9", "sese", "-IR", "", "图像的几何变换看起来更加自然。SinGAN（Shaham", "Dekel，andMichaeli2019）首先提出了由GAN的拉", "Figure 3: IR: image restoration network. PT: patch", "普拉斯金字塔组成的无条件SIG模型，该模型具有", "transfer.Thescatters arethe patches of low(left)and", "完全卷积生成器和斑块鉴别器，以学习不同的尺度信", "high (right)resolution.visualized byt-SNE.Red,blue", "息。ConSinGAN(Hinzetal.2021)修改了PGGAN", "and cyan represent the original, symthesis from DEP.", "(Karrasetal.2017）的训练策略，通过同时训练多个阶", "Net, and patch-transfer synthesis. The “star\" represents", "段来提高 SinGAN。PatchGenCN(Zheng,Xie, and Li", "the mean value. The distribution of cyan data can cover", "2021）提出了一个能量函数金字塔，以明确表示单个自", "red data well, while a conspicuous deviation (caused by", "然图像中的分布。GPNN（Granotetal.2021）用非训", "foreground patches) can be found between the blue data", "练补丁匹配模块（Barnesetal.2009）替换SinGAN的", "and red data.", "每个阶段，以加快生成速度。同时，很多工作不再局", "限于简单的纹理生成。MOGAN（Chenetal.2021）遵", "循SinGAN，但分别合成了手动标记的感兴趣区域和", "Generalized Generationproblem.", "生成任务旨在学习", "参数化分布Pe，以通过经验分布PD近似真实分布", "图像的其余部分，然后将它们合并为完整的图像。", "次性 GAN(Sushko，Gall,and Khoreva 2021）是具有", "多个鉴别器的端到端模型，用于学习图像的不同特征", "Preal,", "min p(Pp. Pe)", "(1)", "et al.2019)的普通 GAN。ExSinGAN(Zhang,Han,and", "但它不是完全卷积的，更像是具有多样性正则化（Yang", "Guo2021）将GAN反演（Panetal.2020）和感知损失", "对问题（1）进行了广泛的研究，e。g。，基于最大对数", "(Johnson,Alahi,andFei-Fei2016)引人到SinGAN中,", "似然估计的显式密度模型（KingmaandWelling2013;", "以提高非纹理图像的性能", "Rezende and Mohamed 2015; van den Oord, Kalchbren-", "ner, and Kavukcuoglu 2016)，其中 p 由 Kullback-", "Deeppriors.在图像恢复中，先验意味着一些经验", "Leibler(KL）散度形成，隐式密度模型如GAN，其中", "知识或客观规律，它们表征了图像的属性并限制了解", "p由Jensen-Shannon(JS)散度（Goodfellowetal.2014)", "空间。它可以显式地表示为正则化项，也可以隐式地", "Wasserstein 距离(Arjovsky,Chintala, and Bottou", "在损失函数中建模。例如，自然图像通常在局部区域", "2017）等形成。", "(Buades,Coli,and Morel 2005;Mei etal.2008)中是平", "PatchDistribution.假设1<i,s<d分别表示滑动", "滑的，因此可以通过全变分正则化（Rudin，Osher.and", "窗口的索引和大小。对于a=ci,..,ad]TE,其第", "Fatemi1992）将平滑度作为恢复任务中的先验。深度图", "j个补丁wg\"=[amod(j-[s/2],d),...,amod(i+[s/2],a)]T", "像先验（Ulyanov，Vedaldi，andLempitsky2018）证明了", "", "表示映射：d→,D→D=u(w-1", "随机初始化的深度卷积网络可以奇迹般地视为隐式正", "则化项，其在恢复任务中的表现要比传统的正则化项好", "ET", "在数据集Ds上定义的随机变量也具有经验分", "得多。从那时起，各种深度先验（Gandelsman，Shocher", "布P，称为大小为的D的补丁分布。此外，π", "and Irani 2019: Gu, Shen, and Zhou 2020: Wang et al", "诱导映射：Prob(d）→Prob(s），P→P和", "2021b）在各个领域得到了广泛的研究和应用。深度生", "P：=p((),T()）是补丁分布的诱导度量。", "成先验（DGP）（Panetal.2020）假设给定图像是从预", "训练生成器的分布空间中采样的，因此将生成器视为", "3.2Single Image Generation Problem", "隐式正则化项以恢复图像。IMAGINEA（Wangetal", "2021a）采用了类似的思想，即预训练分类器被用作隐", "首先：（Shaham，Dekel.andMichaeli2019）将SIG间", "式正则化项，以利用给定图像的语义知识", "题粗略地描述为生成与给定自然图像具有相同视觉内", "容的各种样本。然而，到目前为止还没有明确的数学定", "3Methodology", "义。为了更好地解释和解决这个问题，我们引人了二个", "正则化潜变量模型，所有现有的SIG方法都可以看作", "3.1Preliminaries", "是这个模型的特例。", "Notations.为了简化表达式，我们假设d维的所有", "Problem statement.", "给定一个图像I，SIG问题是学", "，并且可以在d上定义度量空", "习一个非退化且逼真的分布P。，以从经验分布Pp中恢", "间（xd,d,μ）代数。Prob(xd）是在d上定义的概", "数据都来自集d", "复真实分布。值得注意的是，Pp退化为单点分布（i。", "率度量的空间。PrealEProb(xd）表示数据的真实分", "e。狄拉克分布）当D={IEXd}。这里“非退化”表示", "布，它与形式无关，但存在一个样本集D=ciE", "[supp(Preal)/>1,“现实”满足supp(Pe)Csupp(Preal)", "Xdy", "1。a表示集中在的狄拉克度量/分布。p", "显然，SIG是一个不适定的问题。为了优化它，需要", "", "表示测量分布之间统计距离的广义度量。supp(P）：=", "适当的正则化项来获得非平凡解。因此，目标函数可以", "EdINd,P(N）>O!表示对分布的支", "定义为：", "minp(dr, Pe) +(Pe),", "(2)", "持，其中Na是&的邻域。" ]

{ "arxiv_id": "2203.01488", "language": "zh", "page_idx": 4 }

[ [ "0.134", "0.09216", "0.7114", "0.09344", "0.7114", "0.108", "0.134", "0.1067" ], [ "0.11926", "0.1111", "0.4902", "0.1111", "0.4902", "0.125", "0.11926", "0.125" ], [ "0.5", "0.11554", "0.88", "0.11615", "0.88", "0.1307", "0.5", "0.13" ], [ "0.11926", "0.1276", "0.492", "0.1276", "0.492", "0.1414", "0.11926", "0.1414" ], [ "0.505", "0.132", "0.642", "0.132", "0.642", "0.1459", "0.505", "0.1459" ], [ "0.1168", "0.1433", "0.2378", "0.1459", "0.2378", "0.1604", "0.1168", "0.1578" ], [ "0.569", "0.161", "0.8813", "0.1622", "0.8813", "0.1768", "0.569", "0.1755" ], [ "0.1152", "0.178", "0.343", "0.178", "0.343", "0.1925", "0.1152", "0.1925" ], [ "0.63", "0.1755", "0.649", "0.1755", "0.649", "0.1881", "0.63", "0.1881" ], [ "0.5293", "0.2008", "0.8813", "0.2008", "0.8813", "0.214", "0.5293", "0.214" ], [ "0.1177", "0.207", "0.4495", "0.207", "0.4495", "0.2203", "0.1177", "0.2203" ], [ "0.513", "0.2172", "0.8813", "0.2172", "0.8813", "0.2311", "0.513", "0.2311" ], [ "0.511", "0.2343", "0.8813", "0.2343", "0.8813", "0.2474", "0.511", "0.2474" ], [ "0.2336", "0.2405", "0.3708", "0.2405", "0.3708", "0.2551", "0.2336", "0.2551" ], [ "0.4673", "0.2412", "0.4902", "0.2412", "0.4902", "0.254", "0.4673", "0.254" ], [ "0.2942", "0.2556", "0.3105", "0.2556", "0.3105", "0.2651", "0.2942", "0.2651" ], [ "0.5107", "0.25", "0.8813", "0.2507", "0.8813", "0.2651", "0.5107", "0.2644" ], [ "0.511", "0.2683", "0.8813", "0.2683", "0.8813", "0.2822", "0.511", "0.2822" ], [ "0.11926", "0.2827", "0.492", "0.2827", "0.492", "0.2974", "0.11926", "0.2974" ], [ "0.4983", "0.2886", "0.5063", "0.2917", "0.5034", "0.2942", "0.4958", "0.2917" ], [ "0.5073", "0.286", "0.581", "0.286", "0.581", "0.2974", "0.5073", "0.2974" ], [ "0.1177", "0.3005", "0.4885", "0.3005", "0.4885", "0.3145", "0.1177", "0.3145" ], [ "0.528", "0.3018", "0.8813", "0.3018", "0.8813", "0.315", "0.528", "0.315" ], [ "0.1168", "0.3176", "0.4902", "0.3176", "0.4902", "0.3308", "0.1168", "0.3308" ], [ "0.4958", "0.3245", "0.5034", "0.3225", "0.505", "0.3257", "0.4968", "0.3276" ], [ "0.5107", "0.3208", "0.8735", "0.3208", "0.8735", "0.334", "0.5107", "0.334" ], [ "0.1168", "0.3352", "0.4878", "0.3352", "0.4878", "0.3484", "0.1168", "0.3484" ], [ "0.1177", "0.3528", "0.2059", "0.3528", "0.2059", "0.3643", "0.1177", "0.3643" ], [ "0.5195", "0.3484", "0.88", "0.3484", "0.88", "0.3623", "0.5195", "0.3623" ], [ "0.5786", "0.363", "0.6006", "0.363", "0.6006", "0.3757", "0.5786", "0.3757" ], [ "0.1365", "0.3694", "0.4885", "0.3694", "0.4885", "0.3833", "0.1365", "0.3833" ], [ "0.1177", "0.387", "0.4885", "0.387", "0.4885", "0.401", "0.1177", "0.401" ], [ "0.5", "0.387", "0.8813", "0.387", "0.8813", "0.401", "0.5", "0.401" ], [ "0.1177", "0.404", "0.4885", "0.404", "0.4885", "0.4172", "0.1177", "0.4172" ], [ "0.5107", "0.404", "0.8813", "0.404", "0.8813", "0.4172", "0.5107", "0.4172" ], [ "0.1168", "0.4216", "0.4927", "0.4216", "0.4927", "0.435", "0.1168", "0.435" ], [ "0.509", "0.4204", "0.8813", "0.4192", "0.8813", "0.4343", "0.509", "0.435" ], [ "0.1168", "0.4382", "0.4878", "0.4382", "0.4878", "0.452", "0.1168", "0.452" ], [ "0.5073", "0.4382", "0.88", "0.4382", "0.88", "0.452", "0.5073", "0.452" ], [ "0.1177", "0.455", "0.4902", "0.455", "0.4902", "0.4697", "0.1177", "0.4697" ], [ "0.4927", "0.4602", "0.5073", "0.4602", "0.5073", "0.4634", "0.4927", "0.4634" ], [ "0.509", "0.455", "0.746", "0.4558", "0.746", "0.4702", "0.509", "0.4697" ], [ "0.1168", "0.473", "0.4885", "0.473", "0.4885", "0.486", "0.1168", "0.486" ], [ "0.509", "0.481", "0.721", "0.4817", "0.721", "0.4968", "0.509", "0.4956" ], [ "0.1136", "0.488", "0.4248", "0.4893", "0.4248", "0.504", "0.1136", "0.5024" ], [ "0.5107", "0.504", "0.8813", "0.505", "0.8813", "0.5195", "0.5107", "0.518" ], [ "0.1152", "0.521", "0.259", "0.521", "0.259", "0.534", "0.1152", "0.534" ], [ "0.511", "0.523", "0.8813", "0.523", "0.8813", "0.536", "0.511", "0.536" ], [ "0.1168", "0.5444", "0.4902", "0.5444", "0.4902", "0.558", "0.1168", "0.558" ], [ "0.5107", "0.539", "0.8813", "0.539", "0.8813", "0.553", "0.5107", "0.553" ], [ "0.1177", "0.561", "0.4902", "0.561", "0.4902", "0.5747", "0.1177", "0.5747" ], [ "0.5073", "0.5566", "0.8813", "0.5566", "0.8813", "0.571", "0.5073", "0.571" ], [ "0.1168", "0.579", "0.4885", "0.579", "0.4885", "0.5923", "0.1168", "0.5923" ], [ "0.5073", "0.5737", "0.8813", "0.5728", "0.8813", "0.587", "0.5073", "0.5884" ], [ "0.1177", "0.595", "0.4885", "0.595", "0.4885", "0.61", "0.1177", "0.61" ], [ "0.5107", "0.5903", "0.8823", "0.5903", "0.8823", "0.604", "0.5107", "0.604" ], [ "0.5107", "0.606", "0.761", "0.607", "0.761", "0.622", "0.5107", "0.6206" ], [ "0.1177", "0.613", "0.2942", "0.613", "0.2942", "0.627", "0.1177", "0.627" ], [ "0.5293", "0.6235", "0.8823", "0.6235", "0.8823", "0.637", "0.5293", "0.637" ], [ "0.134", "0.631", "0.486", "0.631", "0.486", "0.6445", "0.134", "0.6445" ], [ "0.1152", "0.6465", "0.353", "0.648", "0.353", "0.662", "0.1152", "0.661" ], [ "0.511", "0.6426", "0.8813", "0.6426", "0.8813", "0.656", "0.511", "0.656" ], [ "0.5107", "0.659", "0.8813", "0.6577", "0.8813", "0.6724", "0.5107", "0.674" ], [ "0.2157", "0.6816", "0.3882", "0.6816", "0.3882", "0.696", "0.2157", "0.696" ], [ "0.509", "0.677", "0.8813", "0.677", "0.8813", "0.69", "0.509", "0.69" ], [ "0.4673", "0.683", "0.4902", "0.683", "0.4902", "0.696", "0.4673", "0.696" ], [ "0.4927", "0.685", "0.5", "0.6846", "0.501", "0.6875", "0.4927", "0.688" ], [ "0.511", "0.6934", "0.8813", "0.6934", "0.8813", "0.707", "0.511", "0.707" ], [ "0.134", "0.7153", "0.492", "0.7153", "0.492", "0.73", "0.134", "0.73" ], [ "0.5107", "0.7104", "0.5825", "0.7104", "0.5825", "0.7246", "0.5107", "0.7246" ], [ "0.1217", "0.7197", "0.1365", "0.7197", "0.1365", "0.7266", "0.1217", "0.7266" ], [ "0.5293", "0.7266", "0.8813", "0.7266", "0.8813", "0.7407", "0.5293", "0.7407" ], [ "0.1168", "0.733", "0.4885", "0.733", "0.4885", "0.747", "0.1168", "0.747" ], [ "0.1177", "0.75", "0.4902", "0.75", "0.4902", "0.763", "0.1177", "0.763" ], [ "0.5107", "0.744", "0.8813", "0.744", "0.8813", "0.7583", "0.5107", "0.7583" ], [ "0.5107", "0.761", "0.8813", "0.761", "0.8813", "0.7744", "0.5107", "0.7744" ], [ "0.1177", "0.7676", "0.4878", "0.7676", "0.4878", "0.781", "0.1177", "0.781" ], [ "0.509", "0.7764", "0.8813", "0.778", "0.8813", "0.7925", "0.509", "0.7915" ], [ "0.1168", "0.784", "0.4844", "0.784", "0.4844", "0.7974", "0.1168", "0.7974" ], [ "0.509", "0.7935", "0.88", "0.7935", "0.88", "0.808", "0.509", "0.808" ], [ "0.134", "0.802", "0.4885", "0.8003", "0.4885", "0.815", "0.134", "0.8164" ], [ "0.5107", "0.813", "0.8813", "0.813", "0.8813", "0.827", "0.5107", "0.827" ], [ "0.1177", "0.8193", "0.486", "0.8193", "0.486", "0.8335", "0.1177", "0.8335" ], [ "0.1168", "0.836", "0.4902", "0.836", "0.4902", "0.8496", "0.1168", "0.8496" ], [ "0.511", "0.83", "0.8813", "0.83", "0.8813", "0.8433", "0.511", "0.8433" ], [ "0.1177", "0.8516", "0.4885", "0.853", "0.4885", "0.867", "0.1177", "0.866" ], [ "0.511", "0.8477", "0.8813", "0.8477", "0.8813", "0.8613", "0.511", "0.8613" ], [ "0.511", "0.8643", "0.8813", "0.8643", "0.8813", "0.8774", "0.511", "0.8774" ], [ "0.1177", "0.8706", "0.2957", "0.8706", "0.2957", "0.884", "0.1177", "0.884" ], [ "0.511", "0.881", "0.8813", "0.881", "0.8813", "0.8945", "0.511", "0.8945" ], [ "0.1323", "0.8965", "0.3882", "0.8984", "0.3882", "0.9136", "0.1323", "0.911" ], [ "0.511", "0.8984", "0.7075", "0.8984", "0.7075", "0.912", "0.511", "0.912" ] ]

[ "论文组织如下。2节简要概述了贝叶斯优化，2.2AddingConstraints", "并描述了如何将其应用于我们的问题，3节报告", "方程1的优化问题可以扩展为在某些测量c(0)", "了我们的速度约束调谐实验，4节回顾了相关工", "上添加一个约束：", "作，5节总结了", "0=argmaxf(0)s.t. c(0) > t (3)", "2Bayesian Optimization", "AE", "在我们的设置中，c(①）是配置θ的解码速度", "我们有兴趣找到目标函数，的全局最大值：", "而t是我们希望解码器运行的最低速度。此公式", "假设c在给定一组参数θ的情况下是确定性的。", "+ = argmax f(0)", "(1)", "DE", "但是，正如我们在E3.2节中所示，速度测量固", "有地具有噪声，在使用相同的解码器参数时返回", "，其中θ是来自搜索空间的参数向量。假定f", "", "不同的值", "没有简单的封闭形式，但可以在任意θ点进行", "因此，我们遵循Gelbartetal.（2014）并通过", "评估。在本文中，我们将f作为SMT系统在调", "", "假设c(?）上的概率模型p来重新定义方程3：", "谐集上产生的BLEU分数，并将作为PBMT", "解码器参数", "e* =argmax f(0) s.t. p(c(0) > t) ≥1-8 (4)", "DE", "贝叶斯优化是有效解决此问题的强大框架。它", "通过在f上定义先验模型并按顺序对其进行评", "其中是用户定义的公差值。对于我们的问", "估来工作。根据对迄今为止收集的所有α点的", "题，上面的公式指出，我们希望优化运行速度快", "功能评估，选择评估点以使测量的效用最大化，", "于t且概率为1-的解码器的BLEU分数。与", "这是通过获取功能估算的，该功能权衡了中", "f一样，c也被假定为具有零均值，Matern52内", "不确定区域的探索与有希望的区域的开发。。当", "", "核和加性高斯噪声的GP先验", "f是非凸的，不可微分的并且评估（Shahriari et", "2.3AcquisitionFunction", "al.，2015）的成本很高时，BO特别适合", "先验模型与观测值相结合，在f上产生了后验", "2.1Prior Model", "分布。后验均值给出了有关中潜在最优值的信", "执行BO的第一步是在感兴趣的函数上定义先", "息，换句话说，我们希望exploit的区域。后验", "验模型。虽然文献中存在许多不同的方法，但在", "方差对，i的未知区域中的不确定性进行编码。", "这项工作中，我们遵循Snoeketal.（2012）中提", "e。，我们希望explore的区域。这种勘探/开发", "出的概念，并在Spearmint1工具包中实现，我", "权衡不仅是BO的一个基本方面，也是许多其", "他全局优化方法的一个基本方面", "们将在本节中详细介绍", "获取函数是使用来自后验的信息来建议新评", "f上的先验被定义为高斯过程（GP）（Ras", "mussen and Williams, 2006) :", "估点的启发式方法。他们通过考虑完整的后验信", "息，自然地对勘探/开发权衡进行编码。通过最", "f ~ GP(m(0), k(0,0))", "大化该函数获得建议，这可以使用标准优化技术", "(2)", "", "来完成，因为与原始目标函数相比，它们的评估", "其中m和k是均值和核（或协方差）函数。", "便宜得多", "", "文献中使用的大多数采集函数都是基于改进", "均值函数固定为零常数函数，与GP模型中的", "通常情况一样。这不是一个很大的限制，因为于", "迄今为止获得的最佳评估。然而，已经表明，这", "种方法在存在约束函数E(Gelbartetal.，2014)", "上的posterior通常具有非零平均值。我们使用", "的情况下具有一些病态。。在这里，我们采用带", "Matern52内核，它对函数平滑度几乎没有假设", "约束的预测摘搜索（PESC）（Hernandez-Lobato", "我们工作中的观察结果，BLEU分数，被假定", "etal.，2015），旨在最大化有关全局最优9*的", "为在评估中具有加性高斯噪声。从理论上讲", "对于一组固定的解码参数，我们不希望BLEU", "信息。经验表明，此获取功能在处理约束时可", "发生变化，但实际上，假设一定程度的噪声有助", "以获得更好的结果，并且可以轻松利用称为解", "耦约束（Gelbartetal.，2014）的方案，其中目标", "于使后验计算更加稳定", "（BLEU）和约束（速度）值可以来自不同的测量", "'https://github.com/HIPS/Spearmint", "集。这将在下一节中解释" ]

{ "arxiv_id": "1604.05073", "language": "zh", "page_idx": 1 }

[ [ "0.134", "0.09344", "0.4885", "0.09344", "0.4885", "0.1067", "0.134", "0.1067" ], [ "0.528", "0.09216", "0.8813", "0.09344", "0.8813", "0.108", "0.528", "0.1067" ], [ "0.1168", "0.1111", "0.4902", "0.1111", "0.4902", "0.125", "0.1168", "0.125" ], [ "0.5107", "0.1111", "0.8813", "0.1111", "0.8813", "0.125", "0.5107", "0.125" ], [ "0.1168", "0.1276", "0.4885", "0.1276", "0.4885", "0.1414", "0.1168", "0.1414" ], [ "0.5107", "0.1276", "0.88", "0.1276", "0.88", "0.1414", "0.5107", "0.1414" ], [ "0.125", "0.1433", "0.4902", "0.1445", "0.4902", "0.159", "0.125", "0.1578" ], [ "0.5107", "0.1459", "0.6543", "0.1459", "0.6543", "0.159", "0.5107", "0.159" ], [ "0.1168", "0.161", "0.486", "0.161", "0.486", "0.1755", "0.1168", "0.1755" ], [ "0.505", "0.1768", "0.546", "0.1768", "0.546", "0.1881", "0.505", "0.1881" ], [ "0.5435", "0.1749", "0.773", "0.1755", "0.773", "0.1901", "0.5435", "0.1895" ], [ "0.4902", "0.1843", "0.5034", "0.1787", "0.5063", "0.1831", "0.4927", "0.1875" ], [ "0.1177", "0.1957", "0.492", "0.1963", "0.492", "0.2108", "0.1177", "0.2102" ], [ "0.1168", "0.1787", "0.4902", "0.1787", "0.4902", "0.1925", "0.1168", "0.1925" ], [ "0.5107", "0.2026", "0.8813", "0.2026", "0.8813", "0.2166", "0.5107", "0.2166" ], [ "0.1177", "0.214", "0.4885", "0.214", "0.4885", "0.2279", "0.1177", "0.2279" ], [ "0.509", "0.2197", "0.88", "0.2197", "0.88", "0.2329", "0.509", "0.2329" ], [ "0.1168", "0.2299", "0.4902", "0.2311", "0.4902", "0.2456", "0.1168", "0.2443" ], [ "0.5107", "0.2361", "0.8813", "0.2361", "0.8813", "0.25", "0.5107", "0.25" ], [ "0.1168", "0.2488", "0.2778", "0.2488", "0.2778", "0.262", "0.1168", "0.262" ], [ "0.5107", "0.254", "0.8813", "0.2524", "0.8813", "0.2676", "0.5107", "0.2683" ], [ "0.511", "0.2715", "0.8813", "0.2715", "0.8813", "0.2854", "0.511", "0.2854" ], [ "0.1177", "0.281", "0.4878", "0.281", "0.4878", "0.2942", "0.1177", "0.2942" ], [ "0.511", "0.2886", "0.603", "0.2886", "0.603", "0.303", "0.511", "0.303" ], [ "0.1168", "0.2986", "0.4844", "0.2986", "0.4844", "0.3125", "0.1168", "0.3125" ], [ "0.528", "0.305", "0.8823", "0.305", "0.8823", "0.318", "0.528", "0.318" ], [ "0.1478", "0.3164", "0.4885", "0.3164", "0.4885", "0.33", "0.1478", "0.33" ], [ "0.5107", "0.3213", "0.8813", "0.3225", "0.8813", "0.3372", "0.5107", "0.336" ], [ "0.1462", "0.3333", "0.3865", "0.3333", "0.3865", "0.3467", "0.1462", "0.3467" ], [ "0.511", "0.3403", "0.8813", "0.3403", "0.8813", "0.3542", "0.511", "0.3542" ], [ "0.125", "0.351", "0.1887", "0.3484", "0.1887", "0.361", "0.125", "0.363" ], [ "0.511", "0.3567", "0.8813", "0.3567", "0.8813", "0.3706", "0.511", "0.3706" ], [ "0.125", "0.3674", "0.2998", "0.3674", "0.2998", "0.3806", "0.125", "0.3806" ], [ "0.513", "0.375", "0.8813", "0.375", "0.8813", "0.3882", "0.513", "0.3882" ], [ "0.1259", "0.3865", "0.165", "0.3865", "0.165", "0.3977", "0.1259", "0.3977" ], [ "0.1569", "0.3833", "0.415", "0.385", "0.415", "0.3997", "0.1569", "0.3977" ], [ "0.511", "0.3914", "0.8813", "0.3914", "0.8813", "0.4048", "0.511", "0.4048" ], [ "0.1814", "0.4016", "0.4885", "0.4016", "0.4885", "0.4167", "0.1814", "0.4167" ], [ "0.511", "0.4077", "0.8813", "0.4077", "0.8813", "0.4216", "0.511", "0.4216" ], [ "0.1691", "0.4255", "0.2296", "0.4255", "0.2296", "0.44", "0.1691", "0.44" ], [ "0.5107", "0.4236", "0.8813", "0.4248", "0.8813", "0.4395", "0.5107", "0.4382" ], [ "0.511", "0.4426", "0.709", "0.4426", "0.709", "0.4558", "0.511", "0.4558" ], [ "0.125", "0.4268", "0.1422", "0.4268", "0.1422", "0.4414", "0.125", "0.4414" ], [ "0.1209", "0.4443", "0.1422", "0.44", "0.1462", "0.455", "0.125", "0.459" ], [ "0.1642", "0.443", "0.3284", "0.443", "0.3284", "0.4583", "0.1642", "0.4583" ], [ "0.1642", "0.4602", "0.3342", "0.4602", "0.3342", "0.4749", "0.1642", "0.4749" ], [ "0.528", "0.4602", "0.8823", "0.4602", "0.8823", "0.4734", "0.528", "0.4734" ], [ "0.5107", "0.476", "0.88", "0.4766", "0.88", "0.4912", "0.5107", "0.4905" ], [ "0.125", "0.461", "0.1422", "0.461", "0.1422", "0.4749", "0.125", "0.4749" ], [ "0.125", "0.4778", "0.1438", "0.4778", "0.1438", "0.4905", "0.125", "0.4905" ], [ "0.1642", "0.4766", "0.4207", "0.4766", "0.4207", "0.4912", "0.1642", "0.4912" ], [ "0.1642", "0.4944", "0.4314", "0.4944", "0.4314", "0.5083", "0.1642", "0.5083" ], [ "0.5107", "0.4944", "0.8813", "0.4944", "0.8813", "0.5083", "0.5107", "0.5083" ], [ "0.5107", "0.5103", "0.8813", "0.5103", "0.8813", "0.5244", "0.5107", "0.5244" ], [ "0.125", "0.4956", "0.1438", "0.4956", "0.1438", "0.5083", "0.125", "0.5083" ], [ "0.1259", "0.5146", "0.1422", "0.5146", "0.1422", "0.5234", "0.1259", "0.5234" ], [ "0.1642", "0.511", "0.4355", "0.511", "0.4355", "0.526", "0.1642", "0.526" ], [ "0.11926", "0.53", "0.2018", "0.53", "0.2018", "0.5415", "0.11926", "0.5415" ], [ "0.5107", "0.5264", "0.8813", "0.528", "0.8813", "0.5425", "0.5107", "0.5415" ], [ "0.1152", "0.5444", "0.3652", "0.5435", "0.3652", "0.561", "0.1152", "0.5625" ], [ "0.511", "0.5454", "0.88", "0.5454", "0.88", "0.5596", "0.511", "0.5596" ], [ "0.5107", "0.561", "0.7075", "0.5625", "0.7075", "0.577", "0.5107", "0.5757" ], [ "0.5107", "0.5923", "0.556", "0.5923", "0.556", "0.603", "0.5107", "0.603" ], [ "0.5513", "0.5913", "0.7173", "0.5913", "0.7173", "0.605", "0.5513", "0.605" ], [ "0.1152", "0.5986", "0.3464", "0.5986", "0.3464", "0.613", "0.1152", "0.613" ], [ "0.511", "0.6143", "0.8813", "0.6143", "0.8813", "0.6274", "0.511", "0.6274" ], [ "0.11926", "0.622", "0.4878", "0.622", "0.4878", "0.636", "0.11926", "0.636" ], [ "0.509", "0.63", "0.835", "0.631", "0.835", "0.646", "0.509", "0.6445" ], [ "0.1168", "0.6396", "0.4878", "0.6396", "0.4878", "0.654", "0.1168", "0.654" ], [ "0.1177", "0.657", "0.4878", "0.657", "0.4878", "0.6704", "0.1177", "0.6704" ], [ "0.517", "0.653", "0.884", "0.653", "0.884", "0.6665", "0.517", "0.6665" ], [ "0.1168", "0.6724", "0.4878", "0.674", "0.4878", "0.688", "0.1168", "0.6875" ], [ "0.546", "0.6694", "0.8813", "0.6694", "0.8813", "0.6846", "0.546", "0.6846" ], [ "0.1168", "0.69", "0.4878", "0.69", "0.4878", "0.704", "0.1168", "0.704" ], [ "0.544", "0.685", "0.8823", "0.686", "0.8823", "0.704", "0.544", "0.7026" ], [ "0.11926", "0.7085", "0.4878", "0.7085", "0.4878", "0.7217", "0.11926", "0.7217" ], [ "0.544", "0.705", "0.693", "0.704", "0.693", "0.7183", "0.544", "0.7197" ], [ "0.1177", "0.7246", "0.339", "0.7246", "0.339", "0.7393", "0.1177", "0.7393" ], [ "0.517", "0.7227", "0.88", "0.7236", "0.88", "0.738", "0.517", "0.7373" ], [ "0.1356", "0.7437", "0.4885", "0.7437", "0.4885", "0.757", "0.1356", "0.757" ], [ "0.546", "0.741", "0.88", "0.741", "0.88", "0.756", "0.546", "0.756" ], [ "0.1168", "0.7603", "0.4902", "0.7603", "0.4902", "0.7734", "0.1168", "0.7734" ], [ "0.546", "0.759", "0.7563", "0.759", "0.7563", "0.773", "0.546", "0.773" ], [ "0.1177", "0.778", "0.4885", "0.778", "0.4885", "0.7915", "0.1177", "0.7915" ], [ "0.5156", "0.7764", "0.8813", "0.7764", "0.8813", "0.7915", "0.5156", "0.7915" ], [ "0.1177", "0.7944", "0.4878", "0.7944", "0.4878", "0.808", "0.1177", "0.808" ], [ "0.546", "0.7935", "0.8823", "0.7935", "0.8823", "0.808", "0.546", "0.808" ], [ "0.1177", "0.811", "0.4878", "0.811", "0.4878", "0.8257", "0.1177", "0.8257" ], [ "0.546", "0.8125", "0.8384", "0.8125", "0.8384", "0.827", "0.546", "0.827" ], [ "0.1168", "0.8276", "0.3308", "0.829", "0.3308", "0.8433", "0.1168", "0.8423" ], [ "0.5107", "0.8413", "0.7695", "0.8413", "0.7695", "0.856", "0.5107", "0.856" ], [ "0.134", "0.855", "0.486", "0.857", "0.486", "0.872", "0.134", "0.869" ], [ "0.511", "0.8643", "0.8813", "0.8643", "0.8813", "0.8774", "0.511", "0.8774" ], [ "0.1168", "0.872", "0.4885", "0.872", "0.4885", "0.8853", "0.1168", "0.8853" ], [ "0.1168", "0.8853", "0.4885", "0.8853", "0.4885", "0.8984", "0.1168", "0.8984" ], [ "0.5107", "0.88", "0.8813", "0.881", "0.8813", "0.896", "0.5107", "0.8945" ], [ "0.1177", "0.901", "0.1847", "0.901", "0.1847", "0.911", "0.1177", "0.911" ], [ "0.509", "0.8984", "0.8813", "0.8984", "0.8813", "0.912", "0.509", "0.912" ] ]

[ "1算法总结了约束下的BO过程。它从一组", "解耦场景充分利用了异构计算资源。例如，我", "Do数据点（例如，随机选择）开始，其中每个数", "们有兴趣在将部署的特定机器上测量解码速度。", "据点都是由参数值，一个函数评估（BLEU）和", "但是将调整集转换为测量BLEU可以在任何可", "个约束评估（解码速度）组成的（O，f，c）三元", "用的计算上并行化", "组。。计算目标和约束上的初始后验2。在每次", "3", "Speed Tuning Experiments", "", "一个新的评估点，测量该点的目标值和约束值，", "迭代时，该算法通过最大化采集函数Q来选择", "我们报告了三种语言对的翻译结果，这些语言", "并更新各自的后验分布。它重复此过程，直到达", "对是针对它们对SMT系统构成的不同挑战而选", "到最大迭代次数N，并返回到目前为止根据约", "择的：西班牙语到英语，英语到德语和中文到英", "束有效的最佳参数集", "语。对于每个语言对，我们使用从Web提取的", "通用并行数据。数据大小为1。7,1。1和0。30", "Algorithm 1 Constrained Bayesian Optimization", "亿字，分别", "Input max. number of iterations Nacquisi", "对于西班牙语到英语和英语到德语，我们使", "tion function α, initial evaluations Do , min.", "用混合域调整/测试集，每个集约有1K个句子", "constraint value t , tolerance ", "并且被创建为均匀地表示不同的域，包括世界", "1:=0", "新闻，健康，体育，科学和其他。对于中文到英", "2: for i= 1, .., N do", "文，我们使用通过从内部并行文本集合中随机", "3:", "select new S, by maximizing aα :", "提取唯一的并行句子而创建的域内集合（2K句", "; = argmax α(0,p(f(Di-1), p(cDi-1))", "子）；与其他任务相比，此域内数据导致更高的", "= Ui", "BLEU分数，稍后将对此进行报道。在所有情况", "下，我们有一个参考翻译", "4:", "5:", "query objective f(0,)", "query constraint c(,)", "我们使用基于短语的解码器的内部实现，该解", "码器具有词典化重排序模型E(GalleyandMan-", "6:", "7:", "augment data D, = D, U (0, f,c)i", "update objective posterior p(f|D,)", "ning，2008）。该系统使用21个功能，其权重通", "过MERTE(OchandNey，2004)在非常慢的解", "8:", "9:", "update constraint posterior p(cD,)", "10: end for", "码器参数设置下针对BLEU进行了优化，以最", "11: return 3+ as per Equation 4", "大程度地减少调谐中的搜索错误。在速度调整过", "程中，特征权重保持固定", "3.1", "Decoder Parameters", "2.4 Decoupling Constraints", "我们调整三个直接影响翻译速度的标准解码器", "可以在比可靠的BLEU分数所需的更小的调谐", "参数=（d,s,n）。我们接下来描述它们。", "集上测量转换速度。在速度受限的BLEU调音", "中，我们可以通过在一小部分句子上测量速度", "d : distortion limit. The maximum number of", "来使约束解耦，同时仍在完整的调音集上测量", "source words that may be skipped by the", "BLEU。在这种情况下，BO可以花费更多的时", "decoder as it generates phrases left-to-right", "间查询速度限制值（因为它们更便宜），而花费", "on the target side.", "更少的时间查询BLEU目标", "s: stack size. The maximum number of hy-", "我们使用PESC作为获取函数，因为它可以", "potheses allowed to survive histogram prun-", "轻松处理解耦约束（Gelbart，2015，Sec.4.3）。", "ing in each decoding stack.", "实际上，我们修改了算法1，以根据在第3行", "n : number of translations.The maximum", "最大化PESC所获得的结果，在每次迭代时更", "number of alternative translations per", "新either目标or约束后验。BO中使用的标", "source phrase considered in decoding", "准采集函数不允许这种解耦", "3.2 Measuring Decoding Speed", "2Note that the objective posterior p(f|D) does not de-", "为了更好地了解速度测量，我们使用慢速解码", "pend on the constraint measurements, and the constraint", "posterior p(cD) does not depend on the objective mea-", "器参数设置对英语-德语调谐集进行100次解码", "surements", "i。e。9=(5,100,100)，并使用9=(0,1,1)重" ]

{ "arxiv_id": "1604.05073", "language": "zh", "page_idx": 2 }

[ [ "0.2296", "0.089", "0.3228", "0.09155", "0.3228", "0.1061", "0.2296", "0.1042" ], [ "0.3635", "0.089", "0.451", "0.09155", "0.451", "0.1061", "0.3635", "0.1042" ], [ "0.511", "0.09344", "0.8823", "0.09344", "0.8823", "0.1067", "0.511", "0.1067" ], [ "0.2222", "0.1048", "0.2695", "0.1048", "0.2695", "0.1193", "0.2222", "0.1193" ], [ "0.2942", "0.1061", "0.3228", "0.1061", "0.3228", "0.1187", "0.2942", "0.1187" ], [ "0.353", "0.1048", "0.4019", "0.1048", "0.4019", "0.1193", "0.353", "0.1193" ], [ "0.4248", "0.1061", "0.4534", "0.1061", "0.4534", "0.1187", "0.4248", "0.1187" ], [ "0.5107", "0.1092", "0.8813", "0.10986", "0.8813", "0.125", "0.5107", "0.1237" ], [ "0.2197", "0.1206", "0.273", "0.1206", "0.273", "0.1351", "0.2197", "0.1351" ], [ "0.286", "0.1206", "0.3308", "0.1206", "0.3308", "0.1351", "0.286", "0.1351" ], [ "0.3506", "0.1206", "0.4019", "0.1206", "0.4019", "0.1351", "0.3506", "0.1351" ], [ "0.423", "0.1218", "0.4558", "0.1218", "0.4558", "0.1338", "0.423", "0.1338" ], [ "0.511", "0.1256", "0.7974", "0.127", "0.7974", "0.1414", "0.511", "0.1401" ], [ "0.1299", "0.1383", "0.2059", "0.1383", "0.2059", "0.1528", "0.1299", "0.1528" ], [ "0.1316", "0.1218", "0.1805", "0.1218", "0.1805", "0.1364", "0.1316", "0.1364" ], [ "0.2263", "0.1383", "0.2632", "0.1383", "0.2632", "0.1497", "0.2263", "0.1497" ], [ "0.282", "0.1383", "0.3342", "0.1383", "0.3342", "0.1497", "0.282", "0.1497" ], [ "0.3545", "0.1383", "0.3962", "0.1383", "0.3962", "0.1497", "0.3545", "0.1497" ], [ "0.411", "0.1383", "0.4673", "0.1383", "0.4673", "0.1497", "0.411", "0.1497" ], [ "0.528", "0.1445", "0.8813", "0.1445", "0.8813", "0.159", "0.528", "0.159" ], [ "0.1168", "0.1547", "0.4902", "0.1547", "0.4902", "0.1686", "0.1168", "0.1686" ], [ "0.5073", "0.159", "0.8823", "0.1604", "0.8823", "0.178", "0.5073", "0.1768" ], [ "0.1152", "0.1736", "0.4878", "0.1736", "0.4878", "0.1875", "0.1152", "0.1875" ], [ "0.511", "0.1787", "0.6274", "0.1787", "0.6274", "0.1932", "0.511", "0.1932" ], [ "0.1177", "0.1925", "0.1447", "0.1925", "0.1447", "0.2008", "0.1177", "0.2008" ], [ "0.1168", "0.2108", "0.492", "0.2108", "0.492", "0.226", "0.1168", "0.226" ], [ "0.5107", "0.2084", "0.8813", "0.2084", "0.8813", "0.2229", "0.5107", "0.2229" ], [ "0.11273", "0.2235", "0.2573", "0.2267", "0.2573", "0.2456", "0.11273", "0.2424" ], [ "0.5156", "0.2247", "0.8813", "0.2247", "0.8813", "0.238", "0.5156", "0.238" ], [ "0.1356", "0.2456", "0.4878", "0.2456", "0.4878", "0.26", "0.1356", "0.26" ], [ "0.5146", "0.2424", "0.8823", "0.2424", "0.8823", "0.2556", "0.5146", "0.2556" ], [ "0.1168", "0.2632", "0.4885", "0.2632", "0.4885", "0.2766", "0.1168", "0.2766" ], [ "0.5146", "0.2588", "0.8823", "0.2588", "0.8823", "0.2734", "0.5146", "0.2734" ], [ "0.1177", "0.281", "0.4885", "0.281", "0.4885", "0.2942", "0.1177", "0.2942" ], [ "0.513", "0.276", "0.875", "0.276", "0.875", "0.2903", "0.513", "0.2903" ], [ "0.1136", "0.2954", "0.492", "0.2974", "0.492", "0.3125", "0.1136", "0.31" ], [ "0.5107", "0.305", "0.8813", "0.305", "0.8813", "0.3193", "0.5107", "0.3193" ], [ "0.1168", "0.3145", "0.4805", "0.3145", "0.4805", "0.3276", "0.1168", "0.3276" ], [ "0.5146", "0.3225", "0.8813", "0.3225", "0.8813", "0.336", "0.5146", "0.336" ], [ "0.11926", "0.332", "0.4885", "0.332", "0.4885", "0.3452", "0.11926", "0.3452" ], [ "0.511", "0.3384", "0.8823", "0.339", "0.8823", "0.3542", "0.511", "0.3528" ], [ "0.1177", "0.348", "0.4885", "0.348", "0.4885", "0.3623", "0.1177", "0.3623" ], [ "0.5146", "0.3567", "0.8823", "0.356", "0.8823", "0.3706", "0.5146", "0.3718" ], [ "0.1177", "0.3655", "0.4878", "0.3655", "0.4878", "0.3801", "0.1177", "0.3801" ], [ "0.5156", "0.3738", "0.8813", "0.3738", "0.8813", "0.387", "0.5156", "0.387" ], [ "0.1177", "0.3833", "0.312", "0.3833", "0.312", "0.3965", "0.1177", "0.3965" ], [ "0.5156", "0.3914", "0.8823", "0.3914", "0.8823", "0.4048", "0.5156", "0.4048" ], [ "0.5146", "0.4077", "0.88", "0.4077", "0.88", "0.4216", "0.5146", "0.4216" ], [ "0.1772", "0.4185", "0.2524", "0.4185", "0.2524", "0.43", "0.1772", "0.43" ], [ "0.3652", "0.4185", "0.4346", "0.4185", "0.4346", "0.43", "0.3652", "0.43" ], [ "0.5156", "0.4248", "0.8823", "0.4248", "0.8823", "0.4382", "0.5156", "0.4382" ], [ "0.513", "0.4414", "0.8813", "0.44", "0.8813", "0.455", "0.513", "0.4558" ], [ "0.5156", "0.459", "0.8813", "0.459", "0.8813", "0.473", "0.5156", "0.473" ], [ "0.5156", "0.476", "0.8823", "0.476", "0.8823", "0.4905", "0.5156", "0.4905" ], [ "0.513", "0.4937", "0.7983", "0.4924", "0.7983", "0.507", "0.513", "0.5083" ], [ "0.511", "0.523", "0.8823", "0.523", "0.8823", "0.536", "0.511", "0.536" ], [ "0.1259", "0.5454", "0.3325", "0.5454", "0.3325", "0.5596", "0.1259", "0.5596" ], [ "0.511", "0.5405", "0.8813", "0.5405", "0.8813", "0.5537", "0.511", "0.5537" ], [ "0.326", "0.549", "0.4878", "0.547", "0.4878", "0.558", "0.326", "0.56" ], [ "0.2345", "0.556", "0.3816", "0.5547", "0.3816", "0.565", "0.2345", "0.5664" ], [ "0.511", "0.5566", "0.8813", "0.5566", "0.8813", "0.571", "0.511", "0.571" ], [ "0.5146", "0.5737", "0.8813", "0.5737", "0.8813", "0.5884", "0.5146", "0.5884" ], [ "0.1168", "0.582", "0.4902", "0.582", "0.4902", "0.595", "0.1168", "0.595" ], [ "0.5107", "0.5913", "0.8813", "0.5913", "0.8813", "0.605", "0.5107", "0.605" ], [ "0.1168", "0.601", "0.4885", "0.601", "0.4885", "0.6113", "0.1168", "0.6113" ], [ "0.511", "0.608", "0.8823", "0.608", "0.8823", "0.622", "0.511", "0.622" ], [ "0.1152", "0.616", "0.4902", "0.616", "0.4902", "0.63", "0.1152", "0.63" ], [ "0.511", "0.6255", "0.8813", "0.6255", "0.8813", "0.6396", "0.511", "0.6396" ], [ "0.1168", "0.6353", "0.3774", "0.6353", "0.3774", "0.646", "0.1168", "0.646" ], [ "0.511", "0.6426", "0.8813", "0.6426", "0.8813", "0.656", "0.511", "0.656" ], [ "0.5107", "0.6577", "0.8813", "0.659", "0.8813", "0.674", "0.5107", "0.6724" ], [ "0.1356", "0.6665", "0.4885", "0.6665", "0.4885", "0.6797", "0.1356", "0.6797" ], [ "0.509", "0.675", "0.8813", "0.6753", "0.8813", "0.69", "0.509", "0.6895" ], [ "0.1177", "0.683", "0.4885", "0.683", "0.4885", "0.6963", "0.1177", "0.6963" ], [ "0.511", "0.6934", "0.8813", "0.6934", "0.8813", "0.707", "0.511", "0.707" ], [ "0.1177", "0.7007", "0.4878", "0.7007", "0.4878", "0.714", "0.1177", "0.714" ], [ "0.511", "0.711", "0.88", "0.711", "0.88", "0.7246", "0.511", "0.7246" ], [ "0.1177", "0.7173", "0.4885", "0.7173", "0.4885", "0.731", "0.1177", "0.731" ], [ "0.5107", "0.7266", "0.8813", "0.728", "0.8813", "0.742", "0.5107", "0.741" ], [ "0.1168", "0.733", "0.4885", "0.7344", "0.4885", "0.7485", "0.1168", "0.7476" ], [ "0.5107", "0.7437", "0.6724", "0.744", "0.6724", "0.759", "0.5107", "0.7583" ], [ "0.1177", "0.752", "0.2957", "0.752", "0.2957", "0.765", "0.1177", "0.765" ], [ "0.509", "0.7734", "0.556", "0.7734", "0.556", "0.785", "0.509", "0.785" ], [ "0.5513", "0.7734", "0.615", "0.7734", "0.615", "0.785", "0.5513", "0.785" ], [ "0.1136", "0.781", "0.3635", "0.781", "0.3635", "0.7944", "0.1136", "0.7944" ], [ "0.511", "0.7944", "0.8813", "0.7944", "0.8813", "0.8096", "0.511", "0.8096" ], [ "0.1177", "0.8037", "0.4885", "0.8037", "0.4885", "0.8174", "0.1177", "0.8174" ], [ "0.509", "0.811", "0.8813", "0.8125", "0.8813", "0.827", "0.509", "0.8257" ], [ "0.1177", "0.821", "0.4805", "0.821", "0.4805", "0.834", "0.1177", "0.834" ], [ "0.4763", "0.8257", "0.4885", "0.8257", "0.4885", "0.832", "0.4763", "0.832" ], [ "0.511", "0.829", "0.8813", "0.829", "0.8813", "0.8433", "0.511", "0.8433" ], [ "0.1177", "0.837", "0.4878", "0.837", "0.4878", "0.8516", "0.1177", "0.8516" ], [ "0.5107", "0.8467", "0.8813", "0.8467", "0.8813", "0.8613", "0.5107", "0.8613" ], [ "0.1177", "0.855", "0.4878", "0.855", "0.4878", "0.869", "0.1177", "0.869" ], [ "0.5186", "0.8687", "0.527", "0.8687", "0.527", "0.8735", "0.5186", "0.8735" ], [ "0.527", "0.8643", "0.8813", "0.8643", "0.8813", "0.8774", "0.527", "0.8774" ], [ "0.1168", "0.872", "0.4885", "0.872", "0.4885", "0.886", "0.1168", "0.886" ], [ "0.511", "0.881", "0.88", "0.881", "0.88", "0.8945", "0.511", "0.8945" ], [ "0.134", "0.8965", "0.464", "0.898", "0.464", "0.912", "0.134", "0.911" ], [ "0.511", "0.898", "0.8813", "0.898", "0.8813", "0.911", "0.511", "0.911" ] ]

[ "Slow setting", "Fast setting", "例对s和n进行建模。这种缩放基于直觉，即s", "Mean", "Std", "Mean", "Std", "和n的最佳值将在较低的间隔值中，这在所有", "854.23", "33.88", "105.7k", "5.6k", "三个数据集的初步实验中得到了证实", "log speed", "speed", "6.75", "0.0398", "11.57", "0.0541", "我们运行两组实验，使用2000wpm和", "Table 1: Speed means and standard deviations in words", "5000wpm作为最小速度约束。此外，我们使用", "per minute before and after the logarithmic transforma-", "以下 BO 设置：", "tion", "复快速设置。我们以每分钟翻译的字数（wpm）", "Standard (BO-S): in this setting each BO iter-", "3收集速度测量值", "ation performs a full decoding of the tuning set", "图1中的图显示了直方图，其中包含针对慢", "in order to obtain both the BLEU score and the", "速和快速设置获得的测量结果。虽然两者都符合", "decoding speed jointly. We use = 0.01 as the", "高斯分布，但速度范围在慢速设置下约为750至", "constraint tolerance described in Section 2.2 .", "950wpm，在快速设置下约为90K至120Kwpm。", "Decoupled (BO-D): here we decouple the objec-", "这意味着速度测量表现出heteroscedasticity", "tive and the constraint as explained in Section", "它们遵循高斯分布，其方差取决于解码器参数", "2.4.We still decode the full tuning set to", "值。这对于我们的BO设置是一个问题，因为", "get BLEU scores, but speed measurements are", "我们用于建模约束的GP假定homoscedasticity", "taken from a smaller subset of 50 sentences.", "或支持集上的恒定噪声", "Since speed measurements are faster in this", "case, we enforce BO to query for speed more of-", "Slow setting", "Fast setting", "ten by modeling the task duration as described", "by Snoek et al. (2012).We use a higher con-", "straint tolerance (=0.05 ), as we found that", "BO otherwise focused on the speed constraints", "at the expense of optimizing BLEU.", "我们将这些设置与两个基线进行比较：网格搜索", "700750800850900950100080k", "和随机搜索（BergstraandBengio，2012）。网格", "K90k100k110k120k130k", "Decodingspeed (words/min)", "搜索和随机搜索以类似的方式寻找参数值：提供", "了一组参数值；解码器在所有这些值的调谐集上", "Figure l: Histograms of speed measurements. The solid", "运行；返回以高于约束的速度产生最高BLEU", "line shows a Gaussian fit with the empirical mean and", "的参数值。对于网格搜索，在给定允许的最大", "variance. Note the difference in scale between the two", "解码次数的budget的情况下，选择参数值以", "settings,showing the heteroscedasticity", "覆盖偶数拆分中允许的值范围。对于随机搜索", "参数是从上述指定范围内的均匀分布中选择的。", "减少异方差影响的一种简单方法是采用速度", "BO-S，网格搜索和随机搜索使用125解码的最", "测量的对数，这也是在GP（Gelbartetal.，2014)", "大预算。BO-D允许250次选代的更大预算，因", "中对非负测量进行建模时的标准做法。表1显", "为速度测量可以快速完成。这不是偏向于BO-D", "示了对数转换前后的平均值和标准偏差。使用对", "因为总体目标是在尽可能少的CPU时间内找到", "数可以使GP推理公式易于处理，因此我们在", "最佳，快速的解码器", "实验中使用此解决方案", "3.4", "Results", "3.3BO Details and Baselines", "我们使用2000wpm速度约束的结果如图2所", "除非另有明确说明，否则所有BO实验均使用", "示。图中的实线示出了从当前最佳参数获得", "带有默认值的Spearmint（Snoeketal.，2012）", "", "的调谐集BLEU分数，如BO-S所建议的，作为", "我们将d，s和n的最小值和最大值分别设置", "CPU时间的函数（以对数标度)。给定=0.01", "为[0,10]，[1,500]和[1,100]。对于BO和基", "", "我们在GP模型下有99个%置信度，即满足", "线，我们以线性比例对d进行建模，而以对数比", "速度约束。图2还显示了通过网格和随机搜索", "3Measured on an Intel Xeon E5-2450 at 2.10GHz", "找到的快速系统的最佳BLEU得分，其中预算" ]

{ "arxiv_id": "1604.05073", "language": "zh", "page_idx": 3 }

[ [ "0.3635", "0.07135", "0.4395", "0.07135", "0.4395", "0.0833", "0.3635", "0.0833" ], [ "0.501", "0.0707", "0.6113", "0.0707", "0.6113", "0.0808", "0.501", "0.0808" ], [ "0.3635", "0.0808", "0.4575", "0.0833", "0.4575", "0.09784", "0.3635", "0.09534" ], [ "0.4944", "0.07764", "0.6445", "0.07764", "0.6445", "0.09155", "0.4944", "0.09155" ], [ "0.3652", "0.0966", "0.4102", "0.0966", "0.4102", "0.108", "0.3652", "0.108" ], [ "0.3865", "0.1351", "0.5293", "0.1351", "0.5293", "0.1459", "0.3865", "0.1459" ], [ "0.3447", "0.1433", "0.591", "0.1433", "0.591", "0.1572", "0.3447", "0.1572" ], [ "0.2181", "0.154", "0.3179", "0.154", "0.3179", "0.1654", "0.2181", "0.1654" ], [ "0.6094", "0.1483", "0.6274", "0.1483", "0.6274", "0.1515", "0.6094", "0.1515" ], [ "0.21", "0.1932", "0.286", "0.1932", "0.286", "0.2052", "0.21", "0.2052" ], [ "0.2075", "0.2052", "0.3882", "0.2052", "0.3882", "0.2184", "0.2075", "0.2184" ], [ "0.6445", "0.1995", "0.759", "0.1995", "0.759", "0.2102", "0.6445", "0.2102" ], [ "0.21", "0.2197", "0.3342", "0.2197", "0.3342", "0.2299", "0.21", "0.2299" ], [ "0.5483", "0.2216", "0.6787", "0.2216", "0.6787", "0.2349", "0.5483", "0.2349" ], [ "0.2255", "0.238", "0.3782", "0.238", "0.3782", "0.252", "0.2255", "0.252" ], [ "0.5522", "0.2343", "0.694", "0.2343", "0.694", "0.2474", "0.5522", "0.2474" ], [ "0.2239", "0.25", "0.2834", "0.2474", "0.2852", "0.262", "0.2255", "0.2644" ], [ "0.58", "0.2474", "0.6953", "0.2474", "0.6953", "0.2583", "0.58", "0.2583" ], [ "0.2255", "0.262", "0.3774", "0.262", "0.3774", "0.276", "0.2255", "0.276" ], [ "0.0866", "0.303", "0.911", "0.3037", "0.911", "0.318", "0.0866", "0.3176" ], [ "0.0866", "0.3176", "0.911", "0.3176", "0.911", "0.3308", "0.0866", "0.3308" ], [ "0.0866", "0.332", "0.914", "0.332", "0.914", "0.3452", "0.0866", "0.3452" ], [ "0.08496", "0.3452", "0.3652", "0.3452", "0.3652", "0.3591", "0.08496", "0.3591" ], [ "0.10297", "0.3926", "0.478", "0.3926", "0.478", "0.406", "0.10297", "0.406" ], [ "0.535", "0.3914", "0.9126", "0.3926", "0.9126", "0.4072", "0.535", "0.406" ], [ "0.10297", "0.4072", "0.478", "0.4072", "0.478", "0.4204", "0.10297", "0.4204" ], [ "0.5376", "0.4077", "0.91", "0.4077", "0.91", "0.4185", "0.5376", "0.4185" ], [ "0.10455", "0.4216", "0.3555", "0.4216", "0.3555", "0.432", "0.10455", "0.432" ], [ "0.5376", "0.4216", "0.911", "0.4216", "0.911", "0.432", "0.5376", "0.432" ], [ "0.0866", "0.4343", "0.478", "0.4343", "0.478", "0.4475", "0.0866", "0.4475" ], [ "0.536", "0.4343", "0.91", "0.4343", "0.91", "0.4475", "0.536", "0.4475" ], [ "0.10297", "0.4475", "0.4763", "0.4475", "0.4763", "0.461", "0.10297", "0.461" ], [ "0.536", "0.4475", "0.834", "0.4475", "0.834", "0.461", "0.536", "0.461" ], [ "0.10455", "0.4634", "0.4763", "0.4634", "0.4763", "0.4734", "0.10455", "0.4734" ], [ "0.5156", "0.4673", "0.91", "0.4673", "0.91", "0.4817", "0.5156", "0.4817" ], [ "0.10297", "0.4749", "0.1602", "0.4766", "0.1602", "0.488", "0.10297", "0.486" ], [ "0.535", "0.4797", "0.911", "0.481", "0.911", "0.4956", "0.535", "0.4944" ], [ "0.10297", "0.4893", "0.4795", "0.4893", "0.4795", "0.5024", "0.10297", "0.5024" ], [ "0.536", "0.4956", "0.911", "0.4956", "0.911", "0.509", "0.536", "0.509" ], [ "0.0866", "0.5024", "0.3022", "0.5024", "0.3022", "0.5166", "0.0866", "0.5166" ], [ "0.5376", "0.511", "0.603", "0.511", "0.603", "0.523", "0.5376", "0.523" ], [ "0.0825", "0.5215", "0.478", "0.5215", "0.478", "0.535", "0.0825", "0.535" ], [ "0.5146", "0.5264", "0.9126", "0.528", "0.9126", "0.5425", "0.5146", "0.5415" ], [ "0.1013", "0.535", "0.478", "0.535", "0.478", "0.549", "0.1013", "0.549" ], [ "0.536", "0.5425", "0.9126", "0.5425", "0.9126", "0.556", "0.536", "0.556" ], [ "0.1013", "0.55", "0.478", "0.55", "0.478", "0.563", "0.1013", "0.563" ], [ "0.535", "0.5547", "0.9126", "0.556", "0.9126", "0.571", "0.535", "0.5693" ], [ "0.10455", "0.5645", "0.4763", "0.5645", "0.4763", "0.5747", "0.10455", "0.5747" ], [ "0.536", "0.571", "0.9126", "0.571", "0.9126", "0.584", "0.536", "0.584" ], [ "0.1013", "0.577", "0.4763", "0.577", "0.4763", "0.5903", "0.1013", "0.5903" ], [ "0.536", "0.584", "0.914", "0.584", "0.914", "0.5977", "0.536", "0.5977" ], [ "0.1013", "0.5903", "0.478", "0.5903", "0.478", "0.604", "0.1013", "0.604" ], [ "0.536", "0.5977", "0.914", "0.5977", "0.914", "0.6113", "0.536", "0.6113" ], [ "0.10297", "0.604", "0.478", "0.604", "0.478", "0.617", "0.10297", "0.617" ], [ "0.10297", "0.617", "0.4763", "0.617", "0.4763", "0.631", "0.10297", "0.631" ], [ "0.536", "0.6123", "0.8096", "0.6123", "0.8096", "0.6255", "0.536", "0.6255" ], [ "0.10297", "0.632", "0.478", "0.632", "0.478", "0.646", "0.10297", "0.646" ], [ "0.5376", "0.631", "0.911", "0.631", "0.911", "0.6445", "0.5376", "0.6445" ], [ "0.10297", "0.6465", "0.478", "0.6465", "0.478", "0.66", "0.10297", "0.66" ], [ "0.5215", "0.646", "0.9126", "0.646", "0.9126", "0.66", "0.5215", "0.66" ], [ "0.10297", "0.661", "0.1765", "0.661", "0.1765", "0.6724", "0.10297", "0.6724" ], [ "0.523", "0.66", "0.9004", "0.66", "0.9004", "0.674", "0.523", "0.674" ], [ "0.8945", "0.662", "0.91", "0.662", "0.91", "0.672", "0.8945", "0.672" ], [ "0.0809", "0.677", "0.4756", "0.677", "0.4756", "0.6914", "0.0809", "0.6914" ], [ "0.523", "0.674", "0.9126", "0.674", "0.9126", "0.688", "0.523", "0.688" ], [ "0.10297", "0.69", "0.478", "0.69", "0.478", "0.704", "0.10297", "0.704" ], [ "0.5215", "0.688", "0.9126", "0.688", "0.9126", "0.702", "0.5215", "0.702" ], [ "0.10297", "0.7056", "0.2524", "0.7056", "0.2524", "0.7197", "0.10297", "0.7197" ], [ "0.5215", "0.702", "0.6045", "0.702", "0.6045", "0.7163", "0.5215", "0.7163" ], [ "0.0801", "0.7217", "0.4795", "0.7227", "0.4795", "0.7373", "0.0801", "0.7363" ], [ "0.638", "0.729", "0.8", "0.729", "0.8", "0.742", "0.638", "0.742" ], [ "0.10297", "0.7363", "0.478", "0.7363", "0.478", "0.75", "0.10297", "0.75" ], [ "0.10297", "0.752", "0.4763", "0.752", "0.4763", "0.765", "0.10297", "0.765" ], [ "0.523", "0.75", "0.9126", "0.75", "0.9126", "0.7646", "0.523", "0.7646" ], [ "0.1136", "0.7666", "0.478", "0.7666", "0.478", "0.7764", "0.1136", "0.7764" ], [ "0.523", "0.7646", "0.911", "0.7646", "0.911", "0.778", "0.523", "0.778" ], [ "0.10297", "0.778", "0.4756", "0.779", "0.4756", "0.7925", "0.10297", "0.7915" ], [ "0.5195", "0.7764", "0.911", "0.778", "0.911", "0.7925", "0.5195", "0.7915" ], [ "0.10297", "0.7925", "0.4795", "0.7925", "0.4795", "0.806", "0.10297", "0.806" ], [ "0.5215", "0.7925", "0.9126", "0.7925", "0.9126", "0.806", "0.5215", "0.806" ], [ "0.10297", "0.806", "0.478", "0.806", "0.478", "0.8193", "0.10297", "0.8193" ], [ "0.5186", "0.8037", "0.914", "0.803", "0.914", "0.821", "0.5186", "0.8213" ], [ "0.10297", "0.821", "0.4795", "0.821", "0.4795", "0.834", "0.10297", "0.834" ], [ "0.524", "0.8213", "0.5376", "0.8213", "0.5376", "0.832", "0.524", "0.832" ], [ "0.10297", "0.834", "0.4756", "0.834", "0.4756", "0.8477", "0.10297", "0.8477" ], [ "0.5557", "0.8335", "0.9126", "0.8335", "0.9126", "0.8477", "0.5557", "0.8477" ], [ "0.1013", "0.8467", "0.4795", "0.8477", "0.4795", "0.8623", "0.1013", "0.8613" ], [ "0.5215", "0.8477", "0.911", "0.8477", "0.911", "0.8623", "0.5215", "0.8623" ], [ "0.10297", "0.8623", "0.4763", "0.8623", "0.4763", "0.8755", "0.10297", "0.8755" ], [ "0.5215", "0.8613", "0.9126", "0.8613", "0.9126", "0.875", "0.5215", "0.875" ], [ "0.1013", "0.8755", "0.4795", "0.8755", "0.4795", "0.8896", "0.1013", "0.8896" ], [ "0.5215", "0.8755", "0.9126", "0.8755", "0.9126", "0.8896", "0.5215", "0.8896" ] ]

[ "Synthesisand", "3.LivenessFormulae", "Counter-strategy", "observed_cube_blue", "analysis", "2.InitialEnvironmentState", "observed.cube_blueAclear_bin-right", "LTLSpecification", "", "1. Instructions", "User:\"Pick upthe bluecube and", "Physical Environment", "placeit inthe rightbin", "4.PromptforFeedback", "5.EnvironmentAssumptions", "Robot: \"Will the blue cube", "User:\"Yes\"", "remainwithinreach?\"", "observed.cube_blue", "Figure l: User interaction diagram showing main steps of our approach. Completing grounded specifications with", "environment assumptions is an iterative process which begins with a user command, which the system grounds to an", "LTL specification. The specification is then analyzed and assumptions are derived from it. The user is then prompted", "to confirm or reject these assumptions.", "nal specification more accurately represents the exe-", "of all possible execution sequences. We use existing", "cution environment, since its assumptions are estab-", "work in counter-strategy analysis to generate envi-", "lishedthrough situated interaction.", "ronmentassumptionformulaethatwould,ifadded", ".We demonstrate our model on a significantly larger", "to the specification, produce a synthesizable specifi-", "instruction corpus (two tasks and 60 instructions)", "cation (Alur, Moarref, and Topcu 2013) ;", "from onetask andsixinstructions(Boteanuet al.", "4. We use the proposition evaluation results and the as-", "2016)).", "sumption formulae derived from the counter-strategy", "我们论文的核心贡献，即通过用户交互获得环境假", "to generate prompts, which the user can either accept", "设，是一个迭代过程（图1）：", "or reject;", "1. The user states an instruction which is grounded to", "5. If the user answers a prompt affrmatively, we add the", "a LTL specification using a pre-trained model. The", "corresponding formulae to the specification. We then", "system then attempts to synthesize the LTL specifi-", "attempt to re-synthesize the specification, continuing", "cationintoacontroller.Ifthespecificationissynthe-", "steps 3, 4 and 5 until synthesis succeeds or the user", "sizable, execution begins. If it is unsatisfiable, exist-", "denies all prompts generated by the system. Once", "ing work can inform the user about possible causes", "synthesis has succeeded, the robot executes the task", "(Raman et al. 2013; Boteanu et al. 2016) . Our contri-", "bution targets unrealizable outcomes, in which some", "using the resulting reactive controller.", "admissible environments can impede the robot, thus", "本文的其余部分结构如下：我们首先回顾了相关文", "the following assumes the initial specification is un-", "献，并总结了LTL形式主义和V-DCG模型。然后，我", "realizable;", "们描述我们在生成环境假设和反馈提示方面的贡献。", "最", "2. We supplement the specification with the environ-", "后，我们评估了两个任务的基础模型：二进制立方体排", "ment's initial state as perceived by the robot, and", "序任务和立方体堆叠任务，为此我们众包并注释了一个", "reattempt synthesis;", "指令语料库", "3. If the specification is again unrealizable, we extract", "RELATEDWORK", "the collection of environment moves that will make", "the robot fail, which are known as counter-strategy", "自然语言基础和自然语言交互已在HRI中从广泛的", "Anunrealizablespecificationcanbeexecutedfora", "角度进行了研究，从心理动机到实现机器人自治（Roy", "finite number of steps until the robot is unable to re", "2005）。据我们所知，我们的工作是第一个将随机语言", "spond to environment changes. These are worst-case", "接地，机器人驱动的语言交互用于任务修复以及从形式", "scenarios which are possible given the current spec-", "综合结果生成语言。我们现在将总结这些领域的现有工", "ification. While for unrealizable specifications there", "作", "may also be environments in which the robot can suc", "种使NLI接地的方法使用概率模型，该模型将解", "cessfully execute its goals indefinitely, one advantage", "析语言的含义映射到运动计划的轨迹，物理空间中的区", "of our approach to use formal verification in generat-", "域和对象（广义接地图，G3）（Tellexetal.2011）。DCG", "ing controllers is that we obtain a worst-case analysis", "模型通过定义在规划(Howard，Tellex，andRoy2014)" ]

{ "arxiv_id": "1710.01417", "language": "zh", "page_idx": 2 }

[ [ "0.2737", "0.07574", "0.3464", "0.07574", "0.3464", "0.0833", "0.2737", "0.0833" ], [ "0.697", "0.07764", "0.71", "0.07764", "0.71", "0.089", "0.697", "0.089" ], [ "0.1054", "0.0884", "0.1584", "0.0884", "0.1584", "0.09534", "0.1054", "0.09534" ], [ "0.697", "0.09155", "0.7114", "0.09155", "0.7114", "0.1016", "0.697", "0.1016" ], [ "0.7256", "0.089", "0.738", "0.089", "0.738", "0.09973", "0.7256", "0.09973" ], [ "0.5986", "0.0947", "0.687", "0.0947", "0.687", "0.1048", "0.5986", "0.1048" ], [ "0.0866", "0.09534", "0.1871", "0.09534", "0.1871", "0.1061", "0.0866", "0.1061" ], [ "0.2632", "0.0947", "0.2737", "0.0947", "0.2737", "0.09784", "0.2632", "0.09784" ], [ "0.615", "0.1042", "0.669", "0.1042", "0.669", "0.11426", "0.615", "0.11426" ], [ "0.694", "0.1048", "0.7134", "0.1048", "0.7134", "0.11743", "0.694", "0.11743" ], [ "0.7256", "0.1061", "0.739", "0.1061", "0.739", "0.11554", "0.7256", "0.11554" ], [ "0.751", "0.1061", "0.778", "0.1061", "0.778", "0.113", "0.751", "0.113" ], [ "0.786", "0.1016", "0.8286", "0.1016", "0.8286", "0.11237", "0.786", "0.11237" ], [ "0.2091", "0.11237", "0.2222", "0.11237", "0.2222", "0.1206", "0.2091", "0.1206" ], [ "0.2646", "0.113", "0.2737", "0.113", "0.2737", "0.11615", "0.2646", "0.11615" ], [ "0.6216", "0.11615", "0.664", "0.11615", "0.664", "0.1237", "0.6216", "0.1237" ], [ "0.7915", "0.11237", "0.822", "0.11237", "0.822", "0.1206", "0.7915", "0.1206" ], [ "0.697", "0.1187", "0.71", "0.1187", "0.71", "0.1288", "0.697", "0.1288" ], [ "0.7236", "0.1206", "0.739", "0.1206", "0.739", "0.1307", "0.7236", "0.1307" ], [ "0.0866", "0.13", "0.1936", "0.13", "0.1936", "0.1401", "0.0866", "0.1401" ], [ "0.694", "0.13", "0.7114", "0.13", "0.7114", "0.1433", "0.694", "0.1433" ], [ "0.10046", "0.137", "0.1896", "0.1383", "0.1896", "0.1483", "0.10046", "0.1477" ], [ "0.281", "0.1509", "0.2876", "0.1509", "0.2876", "0.154", "0.281", "0.154" ], [ "0.4004", "0.1635", "0.5024", "0.1642", "0.5024", "0.1749", "0.4004", "0.1736" ], [ "0.517", "0.1622", "0.603", "0.1635", "0.603", "0.1749", "0.517", "0.1736" ], [ "0.7017", "0.1635", "0.7607", "0.161", "0.7607", "0.1724", "0.7017", "0.1749" ], [ "0.3218", "0.1895", "0.664", "0.1895", "0.664", "0.2026", "0.3218", "0.2026" ], [ "0.259", "0.1642", "0.3628", "0.1642", "0.3628", "0.1749", "0.259", "0.1749" ], [ "0.08826", "0.226", "0.4795", "0.226", "0.4795", "0.2393", "0.08826", "0.2393" ], [ "0.5156", "0.226", "0.9087", "0.226", "0.9087", "0.2393", "0.5156", "0.2393" ], [ "0.0866", "0.2393", "0.4795", "0.2393", "0.4795", "0.2524", "0.0866", "0.2524" ], [ "0.5156", "0.2393", "0.9087", "0.2393", "0.9087", "0.2524", "0.5156", "0.2524" ], [ "0.08496", "0.2524", "0.2296", "0.2524", "0.2296", "0.2664", "0.08496", "0.2664" ], [ "0.517", "0.2524", "0.91", "0.2524", "0.91", "0.2664", "0.517", "0.2664" ], [ "0.5156", "0.2664", "0.9087", "0.2664", "0.9087", "0.2798", "0.5156", "0.2798" ], [ "0.08417", "0.2727", "0.4795", "0.276", "0.4795", "0.2935", "0.08417", "0.2903" ], [ "0.517", "0.2798", "0.9087", "0.2798", "0.9087", "0.2935", "0.517", "0.2935" ], [ "0.08826", "0.2922", "0.4795", "0.2922", "0.4795", "0.3062", "0.08826", "0.3062" ], [ "0.517", "0.2922", "0.5537", "0.2942", "0.5522", "0.3062", "0.5156", "0.3037" ], [ "0.08496", "0.3037", "0.4795", "0.305", "0.4795", "0.318", "0.08496", "0.3176" ], [ "0.08496", "0.3176", "0.4795", "0.318", "0.4795", "0.332", "0.08496", "0.3308" ], [ "0.5156", "0.3289", "0.5947", "0.3308", "0.5947", "0.3423", "0.5156", "0.3403" ], [ "0.5933", "0.3289", "0.911", "0.3289", "0.911", "0.3423", "0.5933", "0.3423" ], [ "0.517", "0.3447", "0.8057", "0.3447", "0.8057", "0.3547", "0.517", "0.3547" ], [ "0.08826", "0.334", "0.478", "0.334", "0.478", "0.3447", "0.08826", "0.3447" ], [ "0.0866", "0.3452", "0.478", "0.3452", "0.478", "0.3591", "0.0866", "0.3591" ], [ "0.5376", "0.358", "0.5825", "0.358", "0.5825", "0.3687", "0.5376", "0.3687" ], [ "0.6094", "0.358", "0.6953", "0.358", "0.6953", "0.3694", "0.6094", "0.3694" ], [ "0.705", "0.358", "0.744", "0.358", "0.744", "0.3687", "0.705", "0.3687" ], [ "0.759", "0.358", "0.831", "0.358", "0.831", "0.3687", "0.759", "0.3687" ], [ "0.846", "0.356", "0.8896", "0.358", "0.888", "0.3694", "0.8447", "0.3674" ], [ "0.7563", "0.3706", "0.7974", "0.3706", "0.7974", "0.385", "0.7563", "0.385" ], [ "0.0866", "0.3591", "0.478", "0.3591", "0.478", "0.3726", "0.0866", "0.3726" ], [ "0.08826", "0.3718", "0.1993", "0.3718", "0.1993", "0.385", "0.08826", "0.385" ], [ "0.5376", "0.3718", "0.5947", "0.3718", "0.5947", "0.3833", "0.5376", "0.3833" ], [ "0.607", "0.3718", "0.66", "0.3718", "0.66", "0.3833", "0.607", "0.3833" ], [ "0.705", "0.3718", "0.746", "0.3718", "0.746", "0.3833", "0.705", "0.3833" ], [ "0.846", "0.3718", "0.884", "0.3718", "0.884", "0.3833", "0.846", "0.3833" ], [ "0.5376", "0.3806", "0.5947", "0.3833", "0.5933", "0.3977", "0.536", "0.3958" ], [ "0.6055", "0.3833", "0.6543", "0.3833", "0.6543", "0.3977", "0.6055", "0.3977" ], [ "0.7036", "0.3833", "0.746", "0.3833", "0.746", "0.3977", "0.7036", "0.3977" ], [ "0.7573", "0.3833", "0.7983", "0.3833", "0.7983", "0.3977", "0.7573", "0.3977" ], [ "0.8447", "0.3833", "0.8857", "0.3833", "0.8857", "0.3977", "0.8447", "0.3977" ], [ "0.08417", "0.3901", "0.4805", "0.3933", "0.4805", "0.414", "0.08417", "0.411" ], [ "0.6055", "0.3965", "0.6543", "0.3965", "0.6543", "0.411", "0.6055", "0.411" ], [ "0.7036", "0.3965", "0.746", "0.3965", "0.746", "0.411", "0.7036", "0.411" ], [ "0.7563", "0.3965", "0.7974", "0.3965", "0.7974", "0.411", "0.7563", "0.411" ], [ "0.846", "0.3977", "0.884", "0.3977", "0.884", "0.411", "0.846", "0.411" ], [ "0.7563", "0.4104", "0.7974", "0.4104", "0.7974", "0.4248", "0.7563", "0.4248" ], [ "0.539", "0.3977", "0.5933", "0.3977", "0.5933", "0.4092", "0.539", "0.4092" ], [ "0.08905", "0.411", "0.4795", "0.411", "0.4795", "0.4248", "0.08905", "0.4248" ], [ "0.539", "0.411", "0.5933", "0.411", "0.5933", "0.4224", "0.539", "0.4224" ], [ "0.607", "0.411", "0.652", "0.411", "0.652", "0.4224", "0.607", "0.4224" ], [ "0.705", "0.411", "0.744", "0.411", "0.744", "0.4224", "0.705", "0.4224" ], [ "0.846", "0.411", "0.8813", "0.411", "0.8813", "0.4224", "0.846", "0.4224" ], [ "0.5376", "0.4204", "0.5947", "0.4224", "0.5933", "0.4375", "0.536", "0.435" ], [ "0.7563", "0.4216", "0.7974", "0.4236", "0.796", "0.4382", "0.755", "0.4363" ], [ "0.7036", "0.4224", "0.746", "0.4224", "0.746", "0.4375", "0.7036", "0.4375" ], [ "0.0866", "0.4224", "0.3782", "0.4236", "0.3782", "0.4382", "0.0866", "0.4375" ], [ "0.6055", "0.4236", "0.648", "0.4236", "0.648", "0.4382", "0.6055", "0.4382" ], [ "0.8447", "0.4224", "0.8857", "0.4224", "0.8857", "0.4375", "0.8447", "0.4375" ], [ "0.6055", "0.4363", "0.6465", "0.4363", "0.6465", "0.4507", "0.6055", "0.4507" ], [ "0.7563", "0.4363", "0.7974", "0.4363", "0.7974", "0.4507", "0.7563", "0.4507" ], [ "0.8447", "0.4363", "0.8857", "0.4363", "0.8857", "0.4507", "0.8447", "0.4507" ], [ "0.705", "0.4375", "0.744", "0.4375", "0.744", "0.449", "0.705", "0.449" ], [ "0.539", "0.4375", "0.5933", "0.4375", "0.5933", "0.449", "0.539", "0.449" ], [ "0.08417", "0.4443", "0.4902", "0.449", "0.4902", "0.4673", "0.08417", "0.4634" ], [ "0.6055", "0.4495", "0.653", "0.4495", "0.653", "0.464", "0.6055", "0.464" ], [ "0.7563", "0.4495", "0.7974", "0.4495", "0.7974", "0.464", "0.7563", "0.464" ], [ "0.705", "0.4507", "0.746", "0.4507", "0.746", "0.4622", "0.705", "0.4622" ], [ "0.536", "0.4507", "0.5933", "0.4507", "0.5933", "0.4622", "0.536", "0.4622" ], [ "0.846", "0.4507", "0.8857", "0.4507", "0.8857", "0.4622", "0.846", "0.4622" ], [ "0.10046", "0.464", "0.478", "0.464", "0.478", "0.4778", "0.10046", "0.4778" ], [ "0.54", "0.464", "0.5947", "0.464", "0.5947", "0.476", "0.54", "0.476" ], [ "0.6055", "0.464", "0.653", "0.464", "0.653", "0.476", "0.6055", "0.476" ], [ "0.705", "0.464", "0.746", "0.464", "0.746", "0.476", "0.705", "0.476" ], [ "0.7573", "0.464", "0.796", "0.464", "0.796", "0.476", "0.7573", "0.476" ], [ "0.846", "0.464", "0.8823", "0.464", "0.8823", "0.476", "0.846", "0.476" ], [ "0.5376", "0.476", "0.5947", "0.476", "0.5947", "0.4905", "0.5376", "0.4905" ], [ "0.6055", "0.476", "0.648", "0.476", "0.648", "0.4905", "0.6055", "0.4905" ], [ "0.7036", "0.476", "0.746", "0.476", "0.746", "0.4905", "0.7036", "0.4905" ], [ "0.7563", "0.4734", "0.7974", "0.476", "0.796", "0.4912", "0.755", "0.4893" ], [ "0.846", "0.4766", "0.884", "0.4766", "0.884", "0.4905", "0.846", "0.4905" ], [ "0.6055", "0.4893", "0.648", "0.4893", "0.648", "0.504", "0.6055", "0.504" ], [ "0.08826", "0.4792", "0.1814", "0.4792", "0.1814", "0.4905", "0.08826", "0.4905" ], [ "0.7036", "0.4893", "0.746", "0.4893", "0.746", "0.504", "0.7036", "0.504" ], [ "0.7563", "0.4893", "0.7974", "0.4893", "0.7974", "0.504", "0.7563", "0.504" ], [ "0.846", "0.4905", "0.884", "0.4905", "0.884", "0.504", "0.846", "0.504" ], [ "0.1095", "0.4988", "0.478", "0.4988", "0.478", "0.512", "0.1095", "0.512" ], [ "0.539", "0.4905", "0.5933", "0.4905", "0.5933", "0.502", "0.539", "0.502" ], [ "0.535", "0.502", "0.5933", "0.504", "0.5933", "0.5176", "0.535", "0.515" ], [ "0.607", "0.505", "0.652", "0.505", "0.652", "0.5166", "0.607", "0.5166" ], [ "0.705", "0.505", "0.746", "0.505", "0.746", "0.5166", "0.705", "0.5166" ], [ "0.7573", "0.505", "0.796", "0.505", "0.796", "0.5166", "0.7573", "0.5166" ], [ "0.846", "0.505", "0.8857", "0.505", "0.8857", "0.5166", "0.846", "0.5166" ], [ "0.6055", "0.5166", "0.648", "0.5166", "0.648", "0.531", "0.6055", "0.531" ], [ "0.7036", "0.5166", "0.744", "0.5166", "0.744", "0.531", "0.7036", "0.531" ], [ "0.7563", "0.5166", "0.7974", "0.5166", "0.7974", "0.531", "0.7563", "0.531" ], [ "0.8447", "0.5166", "0.8857", "0.5166", "0.8857", "0.531", "0.8447", "0.531" ], [ "0.539", "0.5176", "0.5933", "0.5176", "0.5933", "0.529", "0.539", "0.529" ], [ "0.0866", "0.518", "0.478", "0.518", "0.478", "0.532", "0.0866", "0.532" ], [ "0.539", "0.531", "0.5947", "0.531", "0.5947", "0.5425", "0.539", "0.5425" ], [ "0.6045", "0.531", "0.6465", "0.531", "0.6465", "0.5425", "0.6045", "0.5425" ], [ "0.705", "0.531", "0.746", "0.531", "0.746", "0.5425", "0.705", "0.5425" ], [ "0.7563", "0.531", "0.7974", "0.531", "0.7974", "0.5425", "0.7563", "0.5425" ], [ "0.8433", "0.53", "0.886", "0.53", "0.886", "0.5444", "0.8433", "0.5444" ], [ "0.08826", "0.5454", "0.4795", "0.5454", "0.4795", "0.5596", "0.08826", "0.5596" ], [ "0.0866", "0.532", "0.4805", "0.532", "0.4805", "0.5454", "0.0866", "0.5454" ], [ "0.08826", "0.56", "0.2051", "0.56", "0.2051", "0.571", "0.08826", "0.571" ], [ "0.5317", "0.5625", "0.9067", "0.5625", "0.9067", "0.5728", "0.5317", "0.5728" ], [ "0.517", "0.5737", "0.9087", "0.5737", "0.9087", "0.587", "0.517", "0.587" ], [ "0.08496", "0.587", "0.2712", "0.587", "0.2712", "0.601", "0.08496", "0.601" ], [ "0.5156", "0.587", "0.9087", "0.587", "0.9087", "0.601", "0.5156", "0.601" ], [ "0.08417", "0.5986", "0.478", "0.603", "0.478", "0.6206", "0.08417", "0.616" ], [ "0.517", "0.6016", "0.562", "0.6016", "0.562", "0.613", "0.517", "0.613" ], [ "0.0866", "0.617", "0.4795", "0.617", "0.4795", "0.631", "0.0866", "0.631" ], [ "0.5195", "0.6235", "0.9067", "0.6255", "0.9067", "0.64", "0.5195", "0.638" ], [ "0.0866", "0.631", "0.478", "0.631", "0.478", "0.646", "0.0866", "0.646" ], [ "0.08826", "0.646", "0.478", "0.646", "0.478", "0.656", "0.08826", "0.656" ], [ "0.5195", "0.6426", "0.9067", "0.6426", "0.9067", "0.653", "0.5195", "0.653" ], [ "0.08826", "0.659", "0.478", "0.659", "0.478", "0.6694", "0.08826", "0.6694" ], [ "0.517", "0.654", "0.9087", "0.655", "0.9087", "0.6685", "0.517", "0.6675" ], [ "0.08826", "0.6704", "0.478", "0.672", "0.478", "0.683", "0.08826", "0.6816" ], [ "0.517", "0.6675", "0.9087", "0.6675", "0.9087", "0.681", "0.517", "0.681" ], [ "0.08826", "0.6846", "0.4795", "0.6846", "0.4795", "0.6978", "0.08826", "0.6978" ], [ "0.5156", "0.681", "0.7817", "0.681", "0.7817", "0.6943", "0.5156", "0.6943" ], [ "0.0866", "0.6963", "0.4795", "0.6963", "0.4795", "0.7104", "0.0866", "0.7104" ], [ "0.0866", "0.7104", "0.478", "0.7104", "0.478", "0.7236", "0.0866", "0.7236" ], [ "0.5156", "0.7173", "0.9087", "0.7173", "0.9087", "0.731", "0.5156", "0.731" ], [ "0.0866", "0.7236", "0.4019", "0.7236", "0.4019", "0.7373", "0.0866", "0.7373" ], [ "0.5156", "0.731", "0.6665", "0.731", "0.6665", "0.744", "0.5156", "0.744" ], [ "0.6587", "0.731", "0.9087", "0.731", "0.9087", "0.744", "0.6587", "0.744" ], [ "0.5156", "0.7456", "0.572", "0.7456", "0.572", "0.757", "0.5156", "0.757" ], [ "0.08496", "0.759", "0.478", "0.7603", "0.478", "0.7734", "0.08496", "0.773" ], [ "0.5825", "0.759", "0.63", "0.759", "0.63", "0.771", "0.5825", "0.771" ], [ "0.661", "0.759", "0.7017", "0.759", "0.7017", "0.771", "0.661", "0.771" ], [ "0.7114", "0.759", "0.786", "0.759", "0.786", "0.771", "0.7114", "0.771" ], [ "0.7983", "0.7583", "0.8447", "0.7583", "0.8447", "0.773", "0.7983", "0.773" ], [ "0.7974", "0.7715", "0.841", "0.7715", "0.841", "0.786", "0.7974", "0.786" ], [ "0.705", "0.762", "0.7114", "0.762", "0.7114", "0.7666", "0.705", "0.7666" ], [ "0.584", "0.773", "0.6406", "0.773", "0.6406", "0.784", "0.584", "0.784" ], [ "0.6567", "0.773", "0.749", "0.773", "0.749", "0.784", "0.6567", "0.784" ], [ "0.0907", "0.784", "0.1609", "0.784", "0.1609", "0.7954", "0.0907", "0.7954" ], [ "0.1708", "0.785", "0.2303", "0.785", "0.2303", "0.797", "0.1708", "0.797" ], [ "0.2296", "0.786", "0.2942", "0.786", "0.2942", "0.7974", "0.2296", "0.7974" ], [ "0.3064", "0.786", "0.4885", "0.786", "0.4885", "0.797", "0.3064", "0.797" ], [ "0.585", "0.786", "0.642", "0.786", "0.642", "0.7974", "0.585", "0.7974" ], [ "0.6587", "0.786", "0.6978", "0.786", "0.6978", "0.7974", "0.6587", "0.7974" ], [ "0.7114", "0.786", "0.749", "0.786", "0.749", "0.7974", "0.7114", "0.7974" ], [ "0.7983", "0.786", "0.8384", "0.786", "0.8384", "0.7974", "0.7983", "0.7974" ], [ "0.584", "0.7954", "0.6445", "0.799", "0.6436", "0.813", "0.5825", "0.81" ], [ "0.29", "0.7886", "0.3057", "0.7886", "0.3057", "0.7954", "0.29", "0.7954" ], [ "0.10455", "0.799", "0.1438", "0.799", "0.1438", "0.81", "0.10455", "0.81" ], [ "0.1732", "0.799", "0.2133", "0.799", "0.2133", "0.81", "0.1732", "0.81" ], [ "0.2484", "0.799", "0.2769", "0.799", "0.2769", "0.81", "0.2484", "0.81" ], [ "0.3447", "0.799", "0.38", "0.799", "0.38", "0.81", "0.3447", "0.81" ], [ "0.4453", "0.799", "0.4722", "0.799", "0.4722", "0.811", "0.4453", "0.811" ], [ "0.6587", "0.8", "0.699", "0.8", "0.699", "0.811", "0.6587", "0.811" ], [ "0.7114", "0.8", "0.751", "0.8", "0.751", "0.811", "0.7114", "0.811" ], [ "0.7983", "0.8", "0.8384", "0.8", "0.8384", "0.811", "0.7983", "0.811" ], [ "0.581", "0.81", "0.648", "0.8125", "0.648", "0.827", "0.581", "0.8247" ], [ "0.6587", "0.813", "0.6978", "0.813", "0.6978", "0.8247", "0.6587", "0.8247" ], [ "0.7114", "0.813", "0.751", "0.813", "0.751", "0.8247", "0.7114", "0.8247" ], [ "0.7974", "0.8125", "0.8423", "0.8125", "0.8423", "0.827", "0.7974", "0.827" ], [ "0.0866", "0.837", "0.1152", "0.837", "0.1152", "0.8486", "0.0866", "0.8486" ], [ "0.1111", "0.839", "0.1316", "0.839", "0.1316", "0.8467", "0.1111", "0.8467" ], [ "0.1316", "0.837", "0.478", "0.837", "0.478", "0.851", "0.1316", "0.851" ], [ "0.5312", "0.8423", "0.9087", "0.8423", "0.9087", "0.856", "0.5312", "0.856" ], [ "0.1299", "0.853", "0.3284", "0.853", "0.3284", "0.863", "0.1299", "0.863" ], [ "0.517", "0.856", "0.9067", "0.856", "0.9067", "0.869", "0.517", "0.869" ], [ "0.0907", "0.8657", "0.4795", "0.867", "0.4795", "0.882", "0.0907", "0.88" ], [ "0.5186", "0.8706", "0.9067", "0.8706", "0.9067", "0.881", "0.5186", "0.881" ], [ "0.08826", "0.882", "0.478", "0.882", "0.478", "0.896", "0.08826", "0.896" ], [ "0.5156", "0.881", "0.9087", "0.882", "0.9087", "0.8965", "0.5156", "0.896" ], [ "0.0866", "0.896", "0.4795", "0.896", "0.4795", "0.909", "0.0866", "0.909" ], [ "0.5146", "0.8945", "0.9087", "0.896", "0.9087", "0.909", "0.5146", "0.9077" ] ]

[ "NWwsWaWgWgW", "O", "W,W.Wy.W,Ws", "O", "C", "More 2D-Convoluton", "T.0000008", "", "and Pooling", "O", "O", "", "Matching", "X", "4.F", "......", "Score", "O", "O", "0000000%", "O", "", "", "Layer-1 2D-Convolution", "Layer-2 2D-Pooling", "Layer-n MLP", "Figure l: Model structure of MatchPyramid.", "Layer-0 Matching Matrix", "我们将具有指示符功能的MatchPyramid简称为MP-Ind。", "景单词。因此，池化层对于从噜杂的背景中提取有用的信息", "同样，我们将MP-Cos用于余弦，将MP-Dot用于点积，将", "可能更为重要。在这个实验中，我们尝试了不同的池大小，并", "MP-Gau用于高斯核。", "在表2中显示了结果。在我们的实验中，最大查询长度为5，", "我们将文档长度截断为500，以提高计算效率。实际上，我们", "2.于ierarchicaliFonaluti分层卷积以提取匹配", "已经尝试了其他长度的文档，例如1000或全长，但性能变化", "模式。分层卷积由卷积层和动态池化层组成，它们通常用于", "很小。", "CNN（例如AlexNet，GoogLeNet）中的图像识别任务。每", "个卷积层中的内核大小是主要的超参数。在文本处理中，内核", "Table 2:", "Comparison of different pooling size of", "MatchPyramid. (fix kernel size 1 x 1)", "的大小决定了我们要组合在一起的单词数量。换句话说，内", "核大小决定了我们考虑的n项特征的最大大小。此外，每个", "Model", "Pooling Size", "MAP", "nDCG@20", "P@20", "0.320", "池化层中的池大小也是重要的超参数，它决定了我们要以多", "大的面积为单位。", "MP-Ind", "5x100", "0.175", "0.254", "MP-Ind", "5×50", "0.195", "0.343", "0.266", "2.务层卷称h便期网cp外gre接将信息聚合为单", "5×20", "0.209", "0.363", "0.279", "0.387", "MP-Ind", "个匹配分数。在本文中，我们使用128个隐藏节点作为完全", "MP-Ind", "5x10", "0.219", "0.301", "MP-Ind", "0.380", "0.214", "连接的隐藏层，并使用ReLU作为激活函数。", "5×5", "0.295", "5×3", "0.370", "0.300", "0.209", "MP-Ind", "2.我们在使用成对排名损失。给定一个三重（g,d+，d-）", "3×20", "0.357", "0.213", "MP-Ind", "0.285", "其中（q,d+）的匹配分数应高于（q,d-）的匹配分数，则损", "MP-Ind", "3 × 10", "0.225", "0.387", "0.302", "MP-Ind", "3×5", "0.225", "0.385", "0.301", "3×3", "失函数定义为：", "0.215", "0.377", "0.301", "L(q,d,d;) =max(0,1 -S(q,d+)+S(q,d)).(6)", "MP-Ind", "MP-Ind", "1 x10", "0.056", "0.082", "0.073", "1×5", "0.051", "0.072", "0.078", "MP-Ind", "其中S(q,d）表示（q,d）的预测匹配分数，而包括前馈匹", "MP-Ind", "1x3", "0.043", "0.066", "0.053", "化相对简单。对于正则化，我们发现提前停止[1]策略对我", "配网络的参数和术语门控网络的参数。使用标准反向传播，优", "们的模型效果很好", "结果显示太小或太大的池大小不适合此任务。最佳池大小", "约为3×10，这意味着在查询端，我们使用查询长度的中位", "3.EXPERIMENTS", "数，而在文档端，我们每50个单词聚合一次，接近段落的平", "3.与了Daset我使Setings集合Robust04，它是一个", "均长度", "新闻数据集。这些主题收集自2004年TREC稳健轨道。数据", "3用影度潮股t领量a通减箱度并构建园配矩阵。", "收集的统计信息如表1所示。在本文中，TREC主题的标题", "被视为查询。在这个实验中，我们使用Galago搜索引擎，在", "对于释义识别任务，前面的结果表明，指标函数是最好的。对", "索引和检索过程中，查询和文档都是空白标记化的，大小写较", "于问答，我们发现点积功能比其他产品更好。在这里，我们尝", "低，并且是词于化的。所有模型都共享在50个维度下在维基", "试探索ad-hoc检索任务的相似度函数。我们比较了四种相似", "百科语料库上训练的单词嵌人。我们采用Adam算法[4]进", "度函数：指标函数、点积、余弦和高斯核。", "行模型训练。学习速率设置为10-4。所有的MatchPyramid", "模型都有一个卷积层和一个动态池化层，因为更多的卷积层", "Table 3: Comparison of different similarity function", "和池化层会由于有限的训练数据而导致过度拟合。", "of MatchPyramid. (", " (fix kernel size 1x 1 , pooling size", "3 × 10 )", "Tablel:Statistics of theTREC collectionRobust04.", "Model", "MAP", "nDCG@20", "P@20", "0.302", "", "MP-Ind", "0.2250.387", "#Vocab", "# Doc", "#Query", "# Retrieval docAvg doc", "MP-Dot", "0.095", "0.149", "0.142", "MP-Cos", "", "0.6M", "0.5M", "250", "2000", "477", "0.189", "0.340", "0.272", " MP-Gau", "0.226", "0.403", "0.318", "3.2", "", "Detailed Studies on MatchPyramid Mod-", "正如我们所看到的，MP-Ind的性能相当不错，表明精确", "elsforAd-hocRetrieval", "匹配的信号对于ad-hoc检索任务非常有用。通过使用嵌人", "3我首pg池失影购Si随化层用于减少特征图的维", "我们可以看到高斯核相似度函数比其他函数更好。为了理解", "度，并为后面的层挑选出最重要的信息。特别是在临时检索", "根本原因，我们首先看一下如图E2所示的单词相似度分布。", "任务中，文档通常包含数百个单词，但其中大多数可能是背", "我们可以看到，精确匹配分数（箭头指向）是在相似度函数余" ]

{ "arxiv_id": "1606.04648", "language": "zh", "page_idx": 1 }

[ [ "0.1209", "0.077", "0.1356", "0.07764", "0.1323", "0.1042", "0.1177", "0.1029" ], [ "0.642", "0.07574", "0.6772", "0.07574", "0.6772", "0.13", "0.642", "0.13" ], [ "0.3938", "0.0821", "0.4019", "0.0821", "0.4019", "0.09155", "0.3938", "0.09155" ], [ "0.2532", "0.0903", "0.3179", "0.0903", "0.3179", "0.101", "0.2532", "0.101" ], [ "0.3938", "0.0947", "0.4019", "0.0947", "0.4019", "0.1042", "0.3938", "0.1042" ], [ "0.1111", "0.1048", "0.134", "0.1048", "0.134", "0.1256", "0.1111", "0.1256" ], [ "0.38", "0.1042", "0.4028", "0.1042", "0.4028", "0.127", "0.38", "0.127" ], [ "0.5186", "0.11615", "0.58", "0.1187", "0.58", "0.13", "0.5186", "0.1276" ], [ "0.6626", "0.1237", "0.6685", "0.1237", "0.6685", "0.1288", "0.6626", "0.1288" ], [ "0.79", "0.1187", "0.8486", "0.1187", "0.8486", "0.1288", "0.79", "0.1288" ], [ "0.12335", "0.1256", "0.1323", "0.1256", "0.1323", "0.137", "0.12335", "0.137" ], [ "0.3906", "0.125", "0.4019", "0.125", "0.4019", "0.137", "0.3906", "0.137" ], [ "0.661", "0.1276", "0.6724", "0.1276", "0.6724", "0.1427", "0.661", "0.1427" ], [ "0.3938", "0.1351", "0.4028", "0.1351", "0.4028", "0.1465", "0.3938", "0.1465" ], [ "0.2141", "0.1578", "0.2467", "0.1578", "0.2467", "0.1654", "0.2141", "0.1654" ], [ "0.4739", "0.1547", "0.527", "0.1547", "0.527", "0.1666", "0.4739", "0.1666" ], [ "0.7534", "0.159", "0.7915", "0.159", "0.7915", "0.1635", "0.7534", "0.1635" ], [ "0.1748", "0.1686", "0.2834", "0.1686", "0.2834", "0.1818", "0.1748", "0.1818" ], [ "0.4631", "0.1692", "0.5337", "0.1692", "0.5337", "0.1812", "0.4631", "0.1812" ], [ "0.701", "0.1686", "0.834", "0.1686", "0.834", "0.1818", "0.701", "0.1818" ], [ "0.0866", "0.1995", "0.9087", "0.1995", "0.9087", "0.2134", "0.0866", "0.2134" ], [ "0.08496", "0.2134", "0.9087", "0.2134", "0.9087", "0.2267", "0.08496", "0.2267" ], [ "0.0866", "0.2279", "0.569", "0.2279", "0.569", "0.238", "0.0866", "0.238" ], [ "0.0866", "0.2632", "0.4795", "0.2632", "0.4795", "0.2766", "0.0866", "0.2766" ], [ "0.5146", "0.2615", "0.911", "0.26", "0.911", "0.2778", "0.5146", "0.279" ], [ "0.0866", "0.2766", "0.478", "0.2766", "0.478", "0.2903", "0.0866", "0.2903" ], [ "0.5146", "0.276", "0.9067", "0.276", "0.9067", "0.289", "0.5146", "0.289" ], [ "0.0866", "0.2903", "0.4795", "0.2903", "0.4795", "0.3037", "0.0866", "0.3037" ], [ "0.517", "0.2903", "0.9087", "0.2903", "0.9087", "0.3037", "0.517", "0.3037" ], [ "0.08496", "0.303", "0.478", "0.303", "0.478", "0.3164", "0.08496", "0.3164" ], [ "0.5186", "0.3037", "0.9087", "0.3037", "0.9087", "0.3176", "0.5186", "0.3176" ], [ "0.08826", "0.318", "0.4763", "0.318", "0.4763", "0.3289", "0.08826", "0.3289" ], [ "0.5156", "0.315", "0.9087", "0.3164", "0.9087", "0.3308", "0.5156", "0.33" ], [ "0.08826", "0.332", "0.2793", "0.332", "0.2793", "0.3423", "0.08826", "0.3423" ], [ "0.5156", "0.33", "0.9087", "0.33", "0.9087", "0.3435", "0.5156", "0.3435" ], [ "0.5195", "0.3435", "0.9087", "0.3435", "0.9087", "0.3567", "0.5195", "0.3567" ], [ "0.08905", "0.3499", "0.4795", "0.3528", "0.4795", "0.3706", "0.08905", "0.3674" ], [ "0.517", "0.3567", "0.6406", "0.3567", "0.6406", "0.3706", "0.517", "0.3706" ], [ "0.08826", "0.3687", "0.478", "0.3687", "0.478", "0.3818", "0.08826", "0.3818" ], [ "0.0866", "0.3818", "0.478", "0.3818", "0.478", "0.3958", "0.0866", "0.3958" ], [ "0.517", "0.3958", "0.5884", "0.3958", "0.5884", "0.406", "0.517", "0.406" ], [ "0.5947", "0.3945", "0.9087", "0.3945", "0.9087", "0.4077", "0.5947", "0.4077" ], [ "0.5146", "0.4072", "0.8286", "0.4072", "0.8286", "0.4204", "0.5146", "0.4204" ], [ "0.08826", "0.3977", "0.3245", "0.3977", "0.3245", "0.4077", "0.08826", "0.4077" ], [ "0.8364", "0.4077", "0.91", "0.4077", "0.91", "0.4192", "0.8364", "0.4192" ], [ "0.5146", "0.4204", "0.8105", "0.4204", "0.8105", "0.4343", "0.5146", "0.4343" ], [ "0.0866", "0.4312", "0.1708", "0.4312", "0.1708", "0.4426", "0.0866", "0.4426" ], [ "0.1642", "0.43", "0.482", "0.43", "0.482", "0.443", "0.1642", "0.443" ], [ "0.6396", "0.432", "0.683", "0.4343", "0.681", "0.449", "0.638", "0.4463" ], [ "0.721", "0.433", "0.763", "0.433", "0.763", "0.4475", "0.721", "0.4475" ], [ "0.7754", "0.435", "0.8486", "0.435", "0.8486", "0.4463", "0.7754", "0.4463" ], [ "0.8613", "0.435", "0.9062", "0.435", "0.9062", "0.4463", "0.8613", "0.4463" ], [ "0.08496", "0.443", "0.3938", "0.443", "0.3938", "0.457", "0.08496", "0.457" ], [ "0.5186", "0.4343", "0.564", "0.4343", "0.564", "0.449", "0.5186", "0.449" ], [ "0.6396", "0.4475", "0.665", "0.4475", "0.665", "0.461", "0.6396", "0.461" ], [ "0.721", "0.4475", "0.763", "0.4475", "0.763", "0.4622", "0.721", "0.4622" ], [ "0.773", "0.4475", "0.8145", "0.4475", "0.8145", "0.4622", "0.773", "0.4622" ], [ "0.8594", "0.4475", "0.9062", "0.4475", "0.9062", "0.4622", "0.8594", "0.4622" ], [ "0.415", "0.4558", "0.4592", "0.4583", "0.4575", "0.4697", "0.4143", "0.4673" ], [ "0.524", "0.449", "0.6006", "0.449", "0.6006", "0.4602", "0.524", "0.4602" ], [ "0.1095", "0.4583", "0.1528", "0.4583", "0.1528", "0.4697", "0.1095", "0.4697" ], [ "0.1847", "0.4583", "0.2607", "0.4583", "0.2607", "0.4697", "0.1847", "0.4697" ], [ "0.2754", "0.4583", "0.3137", "0.4583", "0.3137", "0.4697", "0.2754", "0.4697" ], [ "0.3284", "0.4583", "0.4019", "0.4583", "0.4019", "0.4685", "0.3284", "0.4685" ], [ "0.721", "0.459", "0.761", "0.461", "0.7607", "0.476", "0.7197", "0.4734" ], [ "0.8613", "0.459", "0.902", "0.461", "0.9004", "0.476", "0.8594", "0.4734" ], [ "0.6396", "0.4602", "0.6846", "0.4602", "0.6846", "0.4749", "0.6396", "0.4749" ], [ "0.523", "0.4602", "0.5703", "0.4602", "0.5703", "0.4749", "0.523", "0.4749" ], [ "0.774", "0.4602", "0.8145", "0.4602", "0.8145", "0.4749", "0.774", "0.4749" ], [ "0.3245", "0.4685", "0.3677", "0.4702", "0.367", "0.4849", "0.3228", "0.483" ], [ "0.3186", "0.4622", "0.326", "0.4622", "0.326", "0.4653", "0.3186", "0.4653" ], [ "0.10864", "0.4717", "0.165", "0.4717", "0.165", "0.483", "0.10864", "0.483" ], [ "0.1847", "0.4717", "0.2222", "0.4717", "0.2222", "0.4849", "0.1847", "0.4849" ], [ "0.2754", "0.4717", "0.3162", "0.4717", "0.3162", "0.483", "0.2754", "0.483" ], [ "0.4167", "0.4717", "0.4558", "0.4717", "0.4558", "0.483", "0.4167", "0.483" ], [ "0.5186", "0.4749", "0.63", "0.476", "0.63", "0.4905", "0.5186", "0.4893" ], [ "0.721", "0.476", "0.761", "0.476", "0.761", "0.4905", "0.721", "0.4905" ], [ "0.773", "0.476", "0.814", "0.476", "0.814", "0.4905", "0.773", "0.4905" ], [ "0.8594", "0.476", "0.902", "0.476", "0.902", "0.4905", "0.8594", "0.4905" ], [ "0.6406", "0.4766", "0.689", "0.4766", "0.689", "0.488", "0.6406", "0.488" ], [ "0.10864", "0.481", "0.165", "0.483", "0.1642", "0.4976", "0.10706", "0.4956" ], [ "0.273", "0.483", "0.3145", "0.483", "0.3145", "0.4976", "0.273", "0.4976" ], [ "0.327", "0.483", "0.3677", "0.483", "0.3677", "0.4976", "0.327", "0.4976" ], [ "0.4143", "0.483", "0.4558", "0.483", "0.4558", "0.4976", "0.4143", "0.4976" ], [ "0.7197", "0.4893", "0.7607", "0.4893", "0.7607", "0.504", "0.7197", "0.504" ], [ "0.773", "0.4893", "0.8145", "0.4893", "0.8145", "0.504", "0.773", "0.504" ], [ "0.8594", "0.4893", "0.902", "0.4893", "0.902", "0.504", "0.8594", "0.504" ], [ "0.5254", "0.4905", "0.613", "0.4905", "0.613", "0.502", "0.5254", "0.502" ], [ "0.1847", "0.4849", "0.2222", "0.4849", "0.2222", "0.4968", "0.1847", "0.4968" ], [ "0.326", "0.4944", "0.3677", "0.4968", "0.367", "0.512", "0.3245", "0.5103" ], [ "0.6406", "0.4905", "0.706", "0.4905", "0.706", "0.502", "0.6406", "0.502" ], [ "0.2737", "0.4968", "0.3145", "0.4968", "0.3145", "0.511", "0.2737", "0.511" ], [ "0.4143", "0.4968", "0.455", "0.4968", "0.455", "0.511", "0.4143", "0.511" ], [ "0.183", "0.4988", "0.2214", "0.4988", "0.2214", "0.5103", "0.183", "0.5103" ], [ "0.1095", "0.4976", "0.1642", "0.4976", "0.1642", "0.509", "0.1095", "0.509" ], [ "0.721", "0.5005", "0.759", "0.5024", "0.7573", "0.518", "0.7197", "0.5166" ], [ "0.774", "0.5024", "0.8145", "0.5024", "0.8145", "0.5176", "0.774", "0.5176" ], [ "0.8613", "0.5024", "0.902", "0.5024", "0.902", "0.5176", "0.8613", "0.5176" ], [ "0.183", "0.5103", "0.2222", "0.512", "0.2214", "0.5244", "0.1814", "0.523" ], [ "0.642", "0.504", "0.687", "0.504", "0.687", "0.515", "0.642", "0.515" ], [ "0.1095", "0.511", "0.1642", "0.511", "0.1642", "0.523", "0.1095", "0.523" ], [ "0.2754", "0.511", "0.3137", "0.511", "0.3137", "0.5244", "0.2754", "0.5244" ], [ "0.327", "0.5103", "0.3677", "0.5103", "0.3677", "0.5244", "0.327", "0.5244" ], [ "0.4167", "0.511", "0.455", "0.511", "0.455", "0.5244", "0.4167", "0.5244" ], [ "0.5186", "0.515", "0.6006", "0.5176", "0.6006", "0.532", "0.5186", "0.53" ], [ "0.6406", "0.5166", "0.694", "0.5166", "0.694", "0.531", "0.6406", "0.531" ], [ "0.722", "0.5176", "0.7607", "0.5176", "0.7607", "0.531", "0.722", "0.531" ], [ "0.7754", "0.5176", "0.814", "0.5176", "0.814", "0.531", "0.7754", "0.531" ], [ "0.8613", "0.5166", "0.9043", "0.5166", "0.9043", "0.531", "0.8613", "0.531" ], [ "0.1814", "0.5234", "0.2222", "0.5234", "0.2222", "0.5386", "0.1814", "0.5386" ], [ "0.2737", "0.5234", "0.3145", "0.5234", "0.3145", "0.5386", "0.2737", "0.5386" ], [ "0.327", "0.5234", "0.3677", "0.5234", "0.3677", "0.5386", "0.327", "0.5386" ], [ "0.415", "0.5234", "0.4558", "0.5234", "0.4558", "0.5386", "0.415", "0.5386" ], [ "0.8594", "0.53", "0.903", "0.53", "0.903", "0.5444", "0.8594", "0.5444" ], [ "0.5254", "0.531", "0.613", "0.531", "0.613", "0.5425", "0.5254", "0.5425" ], [ "0.642", "0.531", "0.701", "0.531", "0.701", "0.5425", "0.642", "0.5425" ], [ "0.722", "0.531", "0.7607", "0.531", "0.7607", "0.5425", "0.722", "0.5425" ], [ "0.774", "0.531", "0.812", "0.531", "0.812", "0.5425", "0.774", "0.5425" ], [ "0.1054", "0.535", "0.1732", "0.537", "0.1732", "0.552", "0.1054", "0.55" ], [ "0.1095", "0.5244", "0.1642", "0.5244", "0.1642", "0.536", "0.1095", "0.536" ], [ "0.2737", "0.537", "0.3162", "0.537", "0.3162", "0.552", "0.2737", "0.552" ], [ "0.327", "0.537", "0.367", "0.539", "0.3652", "0.552", "0.326", "0.55" ], [ "0.4143", "0.537", "0.4575", "0.537", "0.4575", "0.552", "0.4143", "0.552" ], [ "0.10864", "0.549", "0.1708", "0.5503", "0.1708", "0.565", "0.10864", "0.563" ], [ "0.1814", "0.539", "0.2214", "0.539", "0.2214", "0.5503", "0.1814", "0.5503" ], [ "0.1814", "0.5503", "0.2239", "0.5503", "0.2239", "0.565", "0.1814", "0.565" ], [ "0.2737", "0.5503", "0.3145", "0.5503", "0.3145", "0.565", "0.2737", "0.565" ], [ "0.3284", "0.552", "0.367", "0.552", "0.367", "0.565", "0.3284", "0.565" ], [ "0.4167", "0.552", "0.455", "0.552", "0.455", "0.565", "0.4167", "0.565" ], [ "0.10706", "0.5625", "0.1724", "0.5645", "0.1724", "0.579", "0.10706", "0.577" ], [ "0.326", "0.5625", "0.3677", "0.5645", "0.367", "0.579", "0.3245", "0.577" ], [ "0.4143", "0.563", "0.4558", "0.563", "0.4558", "0.5776", "0.4143", "0.5776" ], [ "0.5293", "0.561", "0.9087", "0.561", "0.9087", "0.5747", "0.5293", "0.5747" ], [ "0.2737", "0.565", "0.3137", "0.565", "0.3137", "0.577", "0.2737", "0.577" ], [ "0.1814", "0.565", "0.2214", "0.565", "0.2214", "0.577", "0.1814", "0.577" ], [ "0.10706", "0.5747", "0.1724", "0.577", "0.1724", "0.5913", "0.10706", "0.589" ], [ "0.2737", "0.577", "0.3162", "0.577", "0.3162", "0.5913", "0.2737", "0.5913" ], [ "0.327", "0.577", "0.3677", "0.577", "0.3677", "0.5913", "0.327", "0.5913" ], [ "0.4167", "0.5776", "0.455", "0.5776", "0.455", "0.5913", "0.4167", "0.5913" ], [ "0.517", "0.5747", "0.9087", "0.5747", "0.9087", "0.5884", "0.517", "0.5884" ], [ "0.517", "0.587", "0.9087", "0.587", "0.9087", "0.601", "0.517", "0.601" ], [ "0.1814", "0.579", "0.2222", "0.579", "0.2222", "0.5903", "0.1814", "0.5903" ], [ "0.10864", "0.589", "0.1708", "0.5913", "0.1708", "0.604", "0.10864", "0.6016" ], [ "0.3284", "0.589", "0.3652", "0.5913", "0.3635", "0.604", "0.327", "0.6016" ], [ "0.4167", "0.5903", "0.455", "0.5903", "0.455", "0.604", "0.4167", "0.604" ], [ "0.2754", "0.5913", "0.312", "0.5913", "0.312", "0.603", "0.2754", "0.603" ], [ "0.183", "0.5913", "0.2214", "0.5913", "0.2214", "0.603", "0.183", "0.603" ], [ "0.517", "0.601", "0.91", "0.601", "0.91", "0.6143", "0.517", "0.6143" ], [ "0.1013", "0.6187", "0.4795", "0.6187", "0.4795", "0.632", "0.1013", "0.632" ], [ "0.517", "0.613", "0.9087", "0.613", "0.9087", "0.627", "0.517", "0.627" ], [ "0.08496", "0.632", "0.478", "0.632", "0.478", "0.646", "0.08496", "0.646" ], [ "0.517", "0.627", "0.9087", "0.627", "0.9087", "0.64", "0.517", "0.64" ], [ "0.0866", "0.6445", "0.4795", "0.6445", "0.4795", "0.6577", "0.0866", "0.6577" ], [ "0.5186", "0.64", "0.9087", "0.64", "0.9087", "0.654", "0.5186", "0.654" ], [ "0.0866", "0.6577", "0.4795", "0.6577", "0.4795", "0.672", "0.0866", "0.672" ], [ "0.517", "0.653", "0.875", "0.653", "0.875", "0.6665", "0.517", "0.6665" ], [ "0.0866", "0.6704", "0.4763", "0.6704", "0.4763", "0.6846", "0.0866", "0.6846" ], [ "0.0866", "0.6846", "0.4795", "0.6846", "0.4795", "0.6978", "0.0866", "0.6978" ], [ "0.5156", "0.683", "0.91", "0.6875", "0.91", "0.705", "0.5156", "0.7007" ], [ "0.0866", "0.6978", "0.4795", "0.6978", "0.4795", "0.711", "0.0866", "0.711" ], [ "0.517", "0.7007", "0.9087", "0.702", "0.9087", "0.7163", "0.517", "0.7153" ], [ "0.08496", "0.7104", "0.4805", "0.709", "0.4805", "0.7236", "0.08496", "0.7246" ], [ "0.517", "0.7163", "0.9087", "0.7163", "0.9087", "0.73", "0.517", "0.73" ], [ "0.0866", "0.7236", "0.4795", "0.7236", "0.4795", "0.7373", "0.0866", "0.7373" ], [ "0.5156", "0.728", "0.9087", "0.729", "0.9087", "0.742", "0.5156", "0.741" ], [ "0.08496", "0.7373", "0.4795", "0.7363", "0.4795", "0.7505", "0.08496", "0.752" ], [ "0.517", "0.742", "0.9087", "0.742", "0.9087", "0.756", "0.517", "0.756" ], [ "0.0866", "0.7505", "0.478", "0.7505", "0.478", "0.7646", "0.0866", "0.7646" ], [ "0.5156", "0.754", "0.9087", "0.755", "0.9087", "0.7695", "0.5156", "0.768" ], [ "0.08826", "0.765", "0.1478", "0.765", "0.1478", "0.7764", "0.08826", "0.7764" ], [ "0.517", "0.768", "0.9087", "0.768", "0.9087", "0.782", "0.517", "0.782" ], [ "0.0825", "0.78", "0.4805", "0.7827", "0.4805", "0.8047", "0.0825", "0.802" ], [ "0.5146", "0.7827", "0.9067", "0.781", "0.9067", "0.7944", "0.5146", "0.797" ], [ "0.08826", "0.802", "0.4795", "0.802", "0.4795", "0.815", "0.08826", "0.815" ], [ "0.517", "0.797", "0.581", "0.797", "0.581", "0.808", "0.517", "0.808" ], [ "0.08826", "0.8174", "0.478", "0.8174", "0.478", "0.8276", "0.08826", "0.8276" ], [ "0.0866", "0.829", "0.4795", "0.829", "0.4795", "0.8423", "0.0866", "0.8423" ], [ "0.5215", "0.829", "0.6846", "0.8164", "0.687", "0.8384", "0.524", "0.851" ], [ "0.5215", "0.832", "0.527", "0.832", "0.527", "0.837", "0.5215", "0.837" ], [ "0.6274", "0.834", "0.859", "0.834", "0.859", "0.8477", "0.6274", "0.8477" ], [ "0.08826", "0.8423", "0.4795", "0.8423", "0.4795", "0.856", "0.08826", "0.856" ], [ "0.544", "0.8467", "0.863", "0.8477", "0.863", "0.8623", "0.544", "0.8613" ], [ "0.0866", "0.855", "0.478", "0.855", "0.478", "0.8687", "0.0866", "0.8687" ], [ "0.546", "0.8623", "0.846", "0.8623", "0.846", "0.8755", "0.546", "0.8755" ], [ "0.0866", "0.8687", "0.478", "0.8687", "0.478", "0.882", "0.0866", "0.882" ], [ "0.08826", "0.882", "0.478", "0.882", "0.478", "0.896", "0.08826", "0.896" ], [ "0.5146", "0.88", "0.8423", "0.882", "0.8423", "0.8994", "0.5146", "0.898" ], [ "0.08496", "0.8945", "0.4795", "0.896", "0.4795", "0.909", "0.08496", "0.9077" ], [ "0.517", "0.896", "0.79", "0.896", "0.79", "0.909", "0.517", "0.909" ] ]

[ "8", "", "8", "Identity Word", "8", "J", "1", "Identity Word", "n", "Identity Word", "8", "200", "", "8", "milaritySo", "Simlartry Score", "itarityScor", "(a) Dot Product", "(b) Cosine", "(c) Gaussian Kernel", "Figure 2: Choose one word from the vocabulary and measure the similarity between other words, we draw", "the histogram of three type of similarity functions: dot product, cosine and gaussian kernel. The arrow point", "the similarity between two identity word (the word we choose).", "弦和高斯核下的最大值，但不是在点积下。所以MP-Cos和", "用了已发布的模型1（在大型点击数据集上训练）。CDSSM：", "MP-Gau保持精确匹配信号的强度，而MP-Dot失去了这种", "CDSSM模型[9]与DSSM类似，但是使用卷积层对查询和", "能力，导致更糟糕的结果。MP-Cos和MP-Gau之间的性能差", "文档进行编码。出于与DSSM相同的原因，我们也直接使用", "距可能与精确匹配和语义匹配分数之间的差距有关。MP-Gau", "了发布的模型。ARC-I：ARC-I模型[2]使用卷积层来编码两", "的巨大差距表明该模型可以更好地将精确匹配与语义匹配区", "个文本，并使用完全连接的层来聚合匹配分数。基于交互的", "分开来，从而带来更好的性能", "深度匹配模型包括：ARC-II：ARC-II[2]通过在一个小的上", "下文窗口中添加单词嵌人来构造局部交互，然后利用卷积层", "3在季节mp我研Kexnelesigeramid模型中卷积层的内核", "从交互中提取特征。", "大小的影响。在本节中，我们使用MP-Ind和MP-Gau在不", "同大小的内核上进行实验，因为MP-Dot和MP-Cos在即席", "Table 5:", "Comparison of different retrieval mod-", "els over the TREC collection Robust04.", "检索任务中效果不佳。结果见表4。", "t models", "trained on large click-through dataset.", "Table 4:", "Comparison of different kernel size of", "Name", "MAP", "nDCG@20", "P@20", "MatchPyramid. (fix pooling size 3 × 10 )", "Type", "QL", "0.253", "0.415", "0.369", "P@20", "Traditional", "Model", "Kernel Size", "MAP", "nDCG@20", "0.255", "0.370", "BM25", "Model", "0.418", "0.387", "", "MP-Ind", "1 ×1", "0.225", "0.302", "Representation", "0.095", "0.201", "0.171", "DSSM", "MP-Ind", "0.226", "0.382", "0.294", "0.067", "0.146", "0.125", "Based Model", "1 x 3", "0.382", "CDSSM ", "0.223", "0.297", "1×5", "MP-Ind", "0.041", "0.066", "0.065", "3×3", "ARC-I", "MP-Ind", "0.221", "0.379", "0.295", "Interaction", "ARC-II", "0.067", "0.147", "0.128", "5×5", "0.219", "0.378", "0.295", "0.327", "Based Model", "MP-Gau", "0.232", "0.411", "MP-Gau", "MP-Ind", "0.226", "0.403", "0.318", "MP-Gau", "1x1", "1×3", "0.232", "0.411", "0.327", "MP-Gau", "0.409", "0.326", "实验结果（见表5）表明，MatchPyramid优于所有基于表", "0.226", "1 ×5", "MP-Gau", "0.220", "0.400", "0.312", "示的深度匹配模型。主要原因是它可以保留所有对ad-hoc检", "索很重要的低级匹配信号。对于ARC-II，使用1D卷积在较", "3×3", "MP-Gau", "0.371", "0.301", "0.201", "5×5", "小的上下文窗口中生成交互信号似乎对匹配不是很有效。然", "我们已经尝试了两种内核大小，1×n和n×n，其中", "而，我们发现性能最好的深度匹配模型MP-Gau仍然无法与", "nE[1,3,5]。我们期望内核1×n在内核窗口中捕获中心词", "传统的检索模型竞争。结果表明，临时检索任务可能与其他", "周围的信息。结果表明，在指标相似度函数下，不同的核大小", "文本匹配任务不同，例如释义识别和问答。我们需要进一步", "表现相似。如果我们查看指标函数生成的匹配矩阵，这是合理", "研究这些任务之间的差异，以设计更好的深度匹配模型。", "的。匹配矩阵非常稀疏，在内核窗口中通常有一个非零元素", "所以内核大小对指标函数来说并不重要。但是，通过使用高", "4.在本CQNGSIQNSPyramid模型应用于即席检索任", "斯核，我们将语义词相似度引人模型中，适当的核大小将获", "务，并讨论了不同内核大小，池大小和相似性函数的影响。我", "得更多的信息并产生更好的结果。我们发现内核大小为1×3", "们发现，按文档中的段落长度进行池化，可以将精确匹配信", "的MP-Gau实现了最佳性能。此外，我们期望内核n×n捕", "号与语义匹配信号区分开的良好相似性函数以及相对较小的", "获单词邻近信息，例如n-gram匹配。令人惊的是，这些实", "内核大小有助于临时检索任务。实验表明，MatchPyramid模", "验没有任何改进。原因可能是数据集太小，无法学习复杂的", "型在检索任务上可以明显优于最近引人的几种深度匹配模型，", "邻近模式", "但仍然无法与传统检索模型竞争。这些结果鼓励我们寻求对", "3.我们omparisenryithiB换ligeModeks型进行比", "ad-hoc检索中文本匹配任务的更深人理解，并相应地提出更", "较。我们采用三种类型的基线，包括传统模型、基于表示的深", "好的模型。", "度匹配模型和基于交互的深度匹配模型。传统模型包括：QL：", "基于Dirichlet平滑的查询似然模型[11]是最好的语言模型", "[] REFERENCES", "1", "LGifes. Overfitting in neural nets:", "之一。BM25：基于BM25公式[8]，是另一种高效的检索模", "Backpropagation, conjugate gradient, and early", "型。基于表示的深度匹配模型包括：DSSM：DSSM模型[3]", "stopping. In Aduances in Neural Information", "使用全连接层将查询和文档编码为两个固定长度向量，然后", "使用余弦相似度计算匹配分数。由于DSSM由于其巨大的参", "http://research.microsoft.com/en-us/downloads", "数大小而需要大规模的训练数据，因此我们在实验中直接使", "/731572aa-98e4-4c50-b99d-ae3f0c9562b9/" ]

{ "arxiv_id": "1606.04648", "language": "zh", "page_idx": 2 }

[ [ "0.3594", "0.0849", "0.638", "0.0861", "0.638", "0.10095", "0.3594", "0.0998" ], [ "0.3242", "0.1057", "0.679", "0.1057", "0.679", "0.12054", "0.3242", "0.12054" ], [ "0.2217", "0.1414", "0.785", "0.1426", "0.785", "0.1567", "0.2217", "0.1561" ], [ "0.2854", "0.1603", "0.723", "0.1603", "0.723", "0.1733", "0.2854", "0.1733" ], [ "0.297", "0.1746", "0.708", "0.177", "0.708", "0.1912", "0.297", "0.1895" ], [ "0.3032", "0.1948", "0.705", "0.1948", "0.705", "0.2079", "0.3032", "0.2079" ], [ "0.4626", "0.2465", "0.5415", "0.2465", "0.5415", "0.2642", "0.4626", "0.2642" ], [ "0.147", "0.2766", "0.8506", "0.2778", "0.8506", "0.2922", "0.147", "0.2915" ], [ "0.1486", "0.2944", "0.852", "0.2944", "0.852", "0.308", "0.1486", "0.308" ], [ "0.1486", "0.31", "0.8506", "0.31", "0.8506", "0.3237", "0.1486", "0.3237" ], [ "0.1486", "0.3267", "0.854", "0.3267", "0.854", "0.3403", "0.1486", "0.3403" ], [ "0.147", "0.3416", "0.854", "0.3425", "0.854", "0.357", "0.147", "0.3557" ], [ "0.1486", "0.3582", "0.852", "0.3582", "0.852", "0.3723", "0.1486", "0.3723" ], [ "0.1486", "0.3748", "0.854", "0.3748", "0.854", "0.3884", "0.1486", "0.3884" ], [ "0.1486", "0.3914", "0.852", "0.3914", "0.852", "0.405", "0.1486", "0.405" ], [ "0.1453", "0.4062", "0.852", "0.4067", "0.852", "0.4216", "0.1453", "0.4204" ], [ "0.1486", "0.4233", "0.852", "0.4233", "0.852", "0.437", "0.1486", "0.437" ], [ "0.147", "0.4395", "0.852", "0.4395", "0.852", "0.4531", "0.147", "0.4531" ], [ "0.147", "0.4536", "0.8506", "0.4548", "0.8506", "0.4697", "0.147", "0.4685" ], [ "0.1486", "0.4714", "0.854", "0.4714", "0.854", "0.485", "0.1486", "0.485" ], [ "0.1486", "0.4868", "0.854", "0.4868", "0.854", "0.5005", "0.1486", "0.5005" ], [ "0.1511", "0.505", "0.8486", "0.505", "0.8486", "0.5156", "0.1511", "0.5156" ], [ "0.1486", "0.5195", "0.852", "0.5195", "0.852", "0.533", "0.1486", "0.533" ], [ "0.1486", "0.536", "0.2083", "0.536", "0.2083", "0.5474", "0.1486", "0.5474" ], [ "0.1184", "0.5776", "0.1511", "0.5776", "0.1511", "0.589", "0.1184", "0.589" ], [ "0.1486", "0.5806", "0.1545", "0.5806", "0.1545", "0.5874", "0.1486", "0.5874" ], [ "0.1503", "0.576", "0.2644", "0.5767", "0.2644", "0.5913", "0.1503", "0.5903" ], [ "0.12006", "0.6025", "0.51", "0.6025", "0.51", "0.616", "0.12006", "0.616" ], [ "0.529", "0.608", "0.5415", "0.608", "0.5415", "0.611", "0.529", "0.611" ], [ "0.561", "0.608", "0.571", "0.608", "0.571", "0.611", "0.561", "0.611" ], [ "0.5693", "0.605", "0.881", "0.605", "0.881", "0.615", "0.5693", "0.615" ], [ "0.12177", "0.6206", "0.3628", "0.6206", "0.3628", "0.6313", "0.12177", "0.6313" ], [ "0.3594", "0.6216", "0.4055", "0.6216", "0.4055", "0.629", "0.3594", "0.629" ], [ "0.444", "0.6245", "0.4575", "0.6245", "0.4575", "0.6274", "0.444", "0.6274" ], [ "0.4634", "0.6245", "0.476", "0.6245", "0.476", "0.6274", "0.4634", "0.6274" ], [ "0.476", "0.6245", "0.5054", "0.6245", "0.5054", "0.6274", "0.476", "0.6274" ], [ "0.5366", "0.6206", "0.881", "0.6206", "0.881", "0.6313", "0.5366", "0.6313" ], [ "0.12006", "0.6343", "0.3435", "0.6343", "0.3435", "0.648", "0.12006", "0.648" ], [ "0.3376", "0.638", "0.4534", "0.638", "0.4534", "0.6455", "0.3376", "0.6455" ], [ "0.607", "0.6343", "0.882", "0.6343", "0.882", "0.648", "0.607", "0.648" ], [ "0.12177", "0.652", "0.5015", "0.652", "0.5015", "0.665", "0.12177", "0.665" ], [ "0.5283", "0.6562", "0.5615", "0.6562", "0.5615", "0.66", "0.5283", "0.66" ], [ "0.565", "0.652", "0.881", "0.652", "0.881", "0.665", "0.565", "0.665" ], [ "0.12177", "0.6665", "0.879", "0.6665", "0.879", "0.6797", "0.12177", "0.6797" ], [ "0.1234", "0.685", "0.881", "0.685", "0.881", "0.6953", "0.1234", "0.6953" ], [ "0.12006", "0.6997", "0.881", "0.6997", "0.881", "0.7134", "0.12006", "0.7134" ], [ "0.12006", "0.7153", "0.881", "0.7153", "0.881", "0.7285", "0.12006", "0.7285" ], [ "0.12006", "0.7324", "0.8833", "0.7324", "0.8833", "0.745", "0.12006", "0.745" ], [ "0.12006", "0.7476", "0.1931", "0.7476", "0.1931", "0.7627", "0.12006", "0.7627" ], [ "0.1394", "0.764", "0.8833", "0.764", "0.8833", "0.778", "0.1394", "0.778" ], [ "0.12006", "0.78", "0.8833", "0.78", "0.8833", "0.7935", "0.12006", "0.7935" ], [ "0.12177", "0.7964", "0.882", "0.7964", "0.882", "0.81", "0.12177", "0.81" ], [ "0.12177", "0.8125", "0.881", "0.8125", "0.881", "0.826", "0.12177", "0.826" ], [ "0.12177", "0.829", "0.882", "0.829", "0.882", "0.843", "0.12177", "0.843" ], [ "0.1184", "0.8433", "0.8833", "0.844", "0.8833", "0.8594", "0.1184", "0.858" ], [ "0.12177", "0.8623", "0.215", "0.8623", "0.215", "0.8735", "0.12177", "0.8735" ], [ "0.12177", "0.883", "0.882", "0.8843", "0.882", "0.8984", "0.12177", "0.898" ], [ "0.12006", "0.8984", "0.3972", "0.8984", "0.3972", "0.909", "0.12006", "0.909" ] ]

[ "LAVA：通过变分自动编码", "实现对话策略优化的潜在动作空间", "Nurul Lubis, Christian Geishauser, Michael Heck, Hsien-chin Lin.", "Marco Moresi, Carel van Niekerk and Milica Gasic", "Heinrich Heine University Duisseldorf, Germany", "[ lubis,geishaus,heckmi,linh,moresi,niekerk,gasic } @hhu.de", "Abstract", "Reinforcement learning (RL) can enable task-oriented dialogue systems to steer the con", "versation towards successful task completion. In an end-to-end setting, a response can be", "constructed in a word-level sequential decision making process with the entire system vo", "cabulary as action space. Policies trained in such a fashion do not require expert-defined", "action spaces, but they have to deal with large action spaces and long trajectories, making", "RL impractical. Using the latent space of a variational model as action space alleviates", "this problem. However, current approaches use an uninformed prior for training and", "optimize the latent distribution solely on the context. It is therefore unclear whether the", "latent representation truly encodes the characteristics of different actions. In this paper,", "we explore three ways of leveraging an auxiliary task to shape the latent variable distri-", "bution: via pre-training, to obtain an informed prior, and via multitask learning. We", "choose response auto-encoding as the auxiliary task, as this captures the generative fac", "tors of dialogue responses while requiring low computational cost and neither additional", "data nor labels. Our approach yields a more action-characterized latent representations", "whichsupport end-to-enddialoguepolicyoptimizationandachievesstate-of-the-artsuc", "cess rates. These results warrant a more wide-spread use of RL in end-to-end dialogue", "models.", "1", "", "Introduction", "随着个人助理的兴起，面向任务的对话系统的受", "", "", "度和接受度激增。面向任务的对话系统", "的特点是用户自标激励互动，", "如。", "", "", "", "寻找餐厅或叫出租车。如果对话代理能够在", "交互结束时实现用户目标，", "则认为对适代", "对话系统是使用分而治之的方法构", "建的，从而产生了多个模块，这些模块共同构成", "X", "系统管道：自然语言理解（NLU）模块", "对话状态跟踪器（DST），对话策略和自然语言生成（NLG）模块。每个模块都有明确的输人和", "输出，并且可以使用机器学习FYoungetal.，2010：ThomsonandYoung，2010；Williams，2006", "Hendersonetal.，2013）进行训练，前提是可以获得足够标记的数据。但是，模块之间的信息丢", "失。强大的深度学习方法的可用性最近导致了端到端训练方法（BordesandWeston，2017；Wen", "etal.，2017；Madottoetal.，2018）的激增，该方法旨在以序列到序列的方式将用户话语直接映", "射到响应", "面向任务的系统的两个基本属性是能够记住到目前为止对话中重要的所有内容-跟踪，以及产", "生响应以引导对话成功完成任务的能力-计划。在模块化方法中，跟踪的角色由DST承担，使", "用监督学习（SL）进行优化，而计划的角色由对话策略承担，通过强化学习（RL）进行优化。", "与模块化系统不同，端到端系统通常只部署SL，它依赖于语言建模技术，直接优化模型参数下", "数据的可能性，完全忽略规划。这一研究方向从大型预先训练的基于变压器的模型中受益匪浅，", "如 BERT 和 GPT-2 (Devlin et al., 2019; Radford et al., 2019; Hosseini-Asl et al., 2020; Peng", "etal.,2020)", "This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http:", "//creativecommons.org/licenses/by/4.0/" ]

{ "arxiv_id": "2011.09378", "language": "zh", "page_idx": 0 }

[ [ "0.1394", "0.0778", "0.881", "0.0778", "0.881", "0.09143", "0.1394", "0.09143" ], [ "0.12006", "0.0932", "0.882", "0.0932", "0.882", "0.1069", "0.12006", "0.1069" ], [ "0.12177", "0.10986", "0.882", "0.10986", "0.882", "0.12354", "0.12177", "0.12354" ], [ "0.12006", "0.1259", "0.881", "0.1259", "0.881", "0.1395", "0.12006", "0.1395" ], [ "0.1184", "0.1401", "0.8237", "0.1414", "0.8237", "0.1561", "0.1184", "0.155" ], [ "0.1368", "0.158", "0.881", "0.1603", "0.881", "0.1733", "0.1368", "0.1716" ], [ "0.12177", "0.1758", "0.882", "0.1758", "0.882", "0.1895", "0.12177", "0.1895" ], [ "0.1234", "0.1936", "0.881", "0.1936", "0.881", "0.2037", "0.1234", "0.2037" ], [ "0.12177", "0.2079", "0.882", "0.2079", "0.882", "0.2214", "0.12177", "0.2214" ], [ "0.12006", "0.2239", "0.882", "0.2239", "0.882", "0.2369", "0.12006", "0.2369" ], [ "0.12177", "0.2405", "0.635", "0.2405", "0.635", "0.2542", "0.12177", "0.2542" ], [ "0.1368", "0.2583", "0.882", "0.2583", "0.882", "0.2715", "0.1368", "0.2715" ], [ "0.12006", "0.2737", "0.403", "0.2737", "0.403", "0.2874", "0.12006", "0.2874" ], [ "0.1351", "0.3071", "0.825", "0.3071", "0.825", "0.3218", "0.1351", "0.3218" ], [ "0.818", "0.308", "0.8774", "0.308", "0.8774", "0.3196", "0.818", "0.3196" ], [ "0.1561", "0.3225", "0.879", "0.3225", "0.879", "0.3374", "0.1561", "0.3374" ], [ "0.1545", "0.339", "0.7944", "0.3384", "0.7944", "0.3528", "0.1545", "0.354" ], [ "0.1351", "0.3718", "0.881", "0.3718", "0.881", "0.386", "0.1351", "0.386" ], [ "0.1545", "0.3884", "0.8833", "0.3884", "0.8833", "0.4026", "0.1545", "0.4026" ], [ "0.157", "0.405", "0.8833", "0.405", "0.8833", "0.4192", "0.157", "0.4192" ], [ "0.1561", "0.4204", "0.882", "0.4204", "0.882", "0.434", "0.1561", "0.434" ], [ "0.157", "0.4358", "0.882", "0.4358", "0.882", "0.4495", "0.157", "0.4495" ], [ "0.157", "0.4531", "0.5684", "0.4531", "0.5684", "0.466", "0.157", "0.466" ], [ "0.1351", "0.485", "0.881", "0.485", "0.881", "0.4995", "0.1351", "0.4995" ], [ "0.1561", "0.502", "0.5977", "0.502", "0.5977", "0.5156", "0.1561", "0.5156" ], [ "0.1377", "0.535", "0.882", "0.535", "0.882", "0.549", "0.1377", "0.549" ], [ "0.12177", "0.5503", "0.882", "0.5503", "0.882", "0.564", "0.12177", "0.564" ], [ "0.12006", "0.568", "0.8096", "0.568", "0.8096", "0.5806", "0.12006", "0.5806" ], [ "0.11755", "0.5986", "0.2761", "0.5986", "0.2761", "0.6123", "0.11755", "0.6123" ], [ "0.12006", "0.629", "0.881", "0.629", "0.881", "0.6426", "0.12006", "0.6426" ], [ "0.12177", "0.6445", "0.882", "0.6445", "0.882", "0.6577", "0.12177", "0.6577" ], [ "0.12006", "0.661", "0.882", "0.661", "0.882", "0.675", "0.12006", "0.675" ], [ "0.12006", "0.676", "0.881", "0.676", "0.881", "0.6895", "0.12006", "0.6895" ], [ "0.1184", "0.6934", "0.881", "0.6934", "0.881", "0.707", "0.1184", "0.707" ], [ "0.12006", "0.7104", "0.1948", "0.7104", "0.1948", "0.7217", "0.12006", "0.7217" ], [ "0.1368", "0.7256", "0.881", "0.7256", "0.881", "0.7393", "0.1368", "0.7393" ], [ "0.12006", "0.7437", "0.882", "0.7437", "0.882", "0.7573", "0.12006", "0.7573" ], [ "0.12177", "0.759", "0.8833", "0.759", "0.8833", "0.7725", "0.12177", "0.7725" ], [ "0.12177", "0.774", "0.881", "0.774", "0.881", "0.7876", "0.12177", "0.7876" ], [ "0.12006", "0.7915", "0.882", "0.7915", "0.882", "0.8047", "0.12006", "0.8047" ], [ "0.12177", "0.807", "0.882", "0.807", "0.882", "0.821", "0.12177", "0.821" ], [ "0.12006", "0.8237", "0.882", "0.8237", "0.882", "0.8374", "0.12006", "0.8374" ], [ "0.12177", "0.839", "0.881", "0.839", "0.881", "0.8525", "0.12177", "0.8525" ], [ "0.12006", "0.8555", "0.3242", "0.8555", "0.3242", "0.869", "0.12006", "0.869" ], [ "0.1368", "0.8813", "0.8833", "0.883", "0.8833", "0.901", "0.1368", "0.8984" ], [ "0.12006", "0.898", "0.3914", "0.8965", "0.3914", "0.9077", "0.12006", "0.909" ] ]

[ "在面向任务的端到端对话系统中，只有很少的作品部署RLE(Mehrietal，2019b；Zhaoetal.", "2019）。单词级RL将整个系统词汇表的每个单词视为顺序决策过程中的一个动作。这会增加动", "作空间大小和轨迹长度，从而阻碍有效学习和最佳收敛。对话中通常稀疏的奖励进一步加剧了", "信用分配和奖励信号传播的挑战。最后但并非最不重要的一点是，在一个模型中同时优化语言", "连贯性和决策可能会导致分歧。因此，在端到端设置中有效地合并RL仍然是一个挑战。", "在最近提出的潜在动作强化学习（LaRL）中，上下文编码器和响应解码器之间的潜在空间用", "作对话代理E(Zhaoetal.，2019）的动作空间。以潜在变量为条件的解码响应具有解耦动作选择", "和语言生成以及缩短对轨迹的好处，从而提高厂性能。然而，这种方法优化厂潜在的空间使", "用一个不知情的先验不考虑响应的实际分布。此外，潜在空间模型仅取决于上下文。因此，尚", "不清楚潜在变量是否真正编码了不同对话动作的特征，还是编码了对话上下文。由于RL优化", "了动作选择和计划，因此在以动作为特征的空间上执行它很重要", "在本文中，我们提出了一种无监督的方法，用于优化潜在的动作表示，以使用RL进行端到", "端对话策略优化。我们的贡献如下：", ". We propose to optimize latent representations to be action-characterized.", "Action", "characterized representations encode similar actions close to each other and allow inter-", "polation of actions. This leads to a more practical and effective end-to-end RL.", ": We explore three ways of leveraging an auxiliary task to shape the latent variable dis-", "tribution; via pre-training, to obtain an informed prior, and via multitask learning. As", "auxiliary task, we choose response auto-encoding, as this captures generative factors of the", "dialogue responses.Unlike contemporary transformer-based approaches, this requires no", "additional data and has low computational cost. Our analysis shows that the learned latent", "representations encode action-relevant information.", ": We show that our approach achieves state-of-the-art match and success rates on the multi-", "domain MultiWoZ 2.01(Budzianowskiet al.,2018).", "这项工作充当了概念证明，我们可以以无监督的方式诱导以动作为特征的潜在空间，以促进", "更实用和有效的RL。通过使用更复杂的编码和解码模型，整体性能可能会进一步提高，但这超", "出了这项工作的范围。这里获得的结果已经保证在端到端对话模型中更广泛地使用RL", "2Related Work", "在端到端面向任务的系统中的研究很大程度上受到面向聊天的系统的序列到序列建模的成功的", "启发（Serbanetal.，2016）。表示学习已被证明对端到端系统很有用，可以从输入中更有效地", "提取信息。一种常见的方法利用受自然语言处理任务启发的预训练目标，例如。g。下一个话语", "检索（Loweetal.，2016）和生成（VinyalsandLe，2015）。当然，这需要足够数量的额外数据", "并且通常需要标签。预训练目标的选择已被证明会极大地影响学习表示（Mehrietal.，2019a）", "的普适性", "最近，研究人员还研究了表示学习，以更好地对对话响应进行建模。例如，Zhaoetal.（2020)", "研究了使用语言建模任务，如掩码和序列排序来生成响应。使用变分模型，可以使用对话上下", "文-响应对以及一组响应-对话行为对E(ZhaoandEskenazi，2018）来诱导跨域的潜在空间。。类", "似地，响应之前和之后的对话上下文可以以跳过思考的方式用于训练响应嵌入（Zhaoetal..", "2018）。据报道，这样的表示允许仅使用原始对话数据E（Shalyminovetal.，2019）进行少量域", "自适应。状态标签及其转换也已用于学习动作表示（Huangetal.，2020）。由于压缩表示和较短", "的轨迹（Zhaoetal.，2019），与单词级RL相比，在学习的潜在变量空间之上执行RL已显示出", "更好的策略。尽管报告了任务成功和实体识别等指标的改进，但缺乏可解释性和可控性仍然是", "该模型系列中的主要挑战", "'The codebase is accessible at: https://gitlab.cs.uni-duesseldorf.de/general/dsml/lava-publicWe also achieve", "state-of-the-art results on MultiWoZ 2.1" ]

{ "arxiv_id": "2011.09378", "language": "zh", "page_idx": 1 }

[ [ "0.11755", "0.076", "0.2688", "0.0766", "0.2688", "0.09143", "0.11755", "0.0903" ], [ "0.12006", "0.10156", "0.881", "0.10156", "0.881", "0.1152", "0.12006", "0.1152" ], [ "0.12177", "0.11816", "0.8506", "0.11816", "0.8506", "0.1318", "0.12177", "0.1318" ], [ "0.398", "0.1509", "0.602", "0.1509", "0.602", "0.1644", "0.398", "0.1644" ], [ "0.858", "0.152", "0.882", "0.152", "0.882", "0.1644", "0.858", "0.1644" ], [ "0.12006", "0.1733", "0.882", "0.1746", "0.882", "0.1895", "0.12006", "0.1882" ], [ "0.12006", "0.1912", "0.881", "0.1912", "0.881", "0.206", "0.12006", "0.206" ], [ "0.11755", "0.2073", "0.769", "0.2079", "0.769", "0.2227", "0.11755", "0.2214" ], [ "0.1377", "0.2257", "0.75", "0.2257", "0.75", "0.2363", "0.1377", "0.2363" ], [ "0.4307", "0.253", "0.471", "0.2542", "0.471", "0.2612", "0.4307", "0.26" ], [ "0.315", "0.2632", "0.687", "0.2632", "0.687", "0.2808", "0.315", "0.2808" ], [ "0.858", "0.2654", "0.882", "0.2654", "0.882", "0.2798", "0.858", "0.2798" ], [ "0.424", "0.2798", "0.4785", "0.2798", "0.4785", "0.2957", "0.424", "0.2957" ], [ "0.12006", "0.308", "0.882", "0.3088", "0.882", "0.3225", "0.12006", "0.3218" ], [ "0.12177", "0.3237", "0.881", "0.3237", "0.881", "0.3374", "0.12177", "0.3374" ], [ "0.12006", "0.3403", "0.55", "0.3403", "0.55", "0.354", "0.12006", "0.354" ], [ "0.1368", "0.3557", "0.882", "0.357", "0.882", "0.3718", "0.1368", "0.3706" ], [ "0.1184", "0.3718", "0.8833", "0.3723", "0.8833", "0.3872", "0.1184", "0.386" ], [ "0.12177", "0.3901", "0.2284", "0.3901", "0.2284", "0.402", "0.12177", "0.402" ], [ "0.4702", "0.401", "0.4902", "0.401", "0.4902", "0.4146", "0.4702", "0.4146" ], [ "0.3586", "0.415", "0.644", "0.415", "0.644", "0.4329", "0.3586", "0.4329" ], [ "0.858", "0.4175", "0.882", "0.4175", "0.882", "0.4316", "0.858", "0.4316" ], [ "0.4685", "0.4316", "0.492", "0.4316", "0.492", "0.4465", "0.4685", "0.4465" ], [ "0.12006", "0.4507", "0.1587", "0.4536", "0.1561", "0.4685", "0.11755", "0.4656" ], [ "0.12177", "0.4702", "0.881", "0.4702", "0.881", "0.481", "0.12177", "0.481" ], [ "0.12006", "0.485", "0.881", "0.485", "0.881", "0.4988", "0.12006", "0.4988" ], [ "0.12006", "0.5005", "0.881", "0.5005", "0.881", "0.514", "0.12006", "0.514" ], [ "0.12006", "0.5186", "0.503", "0.5186", "0.503", "0.532", "0.12006", "0.532" ], [ "0.2761", "0.5474", "0.723", "0.5474", "0.723", "0.5625", "0.2761", "0.5625" ], [ "0.858", "0.5464", "0.8833", "0.5464", "0.8833", "0.561", "0.858", "0.561" ], [ "0.1394", "0.576", "0.879", "0.576", "0.879", "0.589", "0.1394", "0.589" ], [ "0.12177", "0.593", "0.882", "0.593", "0.882", "0.6055", "0.12177", "0.6055" ], [ "0.1234", "0.61", "0.881", "0.61", "0.881", "0.6206", "0.1234", "0.6206" ], [ "0.12006", "0.6245", "0.7246", "0.6245", "0.7246", "0.638", "0.12006", "0.638" ], [ "0.294", "0.652", "0.705", "0.653", "0.705", "0.6704", "0.294", "0.6694" ], [ "0.858", "0.654", "0.882", "0.654", "0.882", "0.6685", "0.858", "0.6685" ], [ "0.11755", "0.6816", "0.4978", "0.683", "0.4978", "0.6978", "0.11755", "0.697" ], [ "0.12177", "0.708", "0.882", "0.708", "0.882", "0.7227", "0.12177", "0.7227" ], [ "0.12177", "0.7246", "0.882", "0.7246", "0.882", "0.7383", "0.12177", "0.7383" ], [ "0.12177", "0.7407", "0.881", "0.7407", "0.881", "0.7544", "0.12177", "0.7544" ], [ "0.12177", "0.7583", "0.881", "0.7583", "0.881", "0.7686", "0.12177", "0.7686" ], [ "0.1184", "0.7715", "0.882", "0.7725", "0.882", "0.7876", "0.1184", "0.786" ], [ "0.12177", "0.789", "0.882", "0.789", "0.882", "0.8027", "0.12177", "0.8027" ], [ "0.12006", "0.804", "0.8774", "0.8057", "0.8774", "0.8193", "0.12006", "0.818" ], [ "0.2957", "0.832", "0.705", "0.833", "0.705", "0.851", "0.2957", "0.8496" ], [ "0.858", "0.8345", "0.882", "0.8345", "0.882", "0.8486", "0.858", "0.8486" ], [ "0.1184", "0.86", "0.882", "0.8623", "0.882", "0.88", "0.1184", "0.8774" ], [ "0.12177", "0.88", "0.882", "0.88", "0.882", "0.8936", "0.12177", "0.8936" ], [ "0.12177", "0.8965", "0.5786", "0.8965", "0.5786", "0.91", "0.12177", "0.91" ] ]

[ "3Preliminaries", "在端到端框架中，使用RL进行对话策略优化通常包括两个步骤：SL和策略梯度RL。在SL步", "骤中，模型学习基于一些对话上下文c生成响应，更新其参数以最大化数据的对数似然", "LsL(0) = Ea,c[logpe(cc)]", "(1)", "随后，RL步骤更新模型参数w。r。t。特定于任务的目标，反映为奖励。在与T步骤的对话中，", "对于特定的时间步长t，即时奖励rt和折扣因子E[0,1]，折扣收益定义为Rt=Zk=tk-trk", "。该模型试图从第一个时间步长开始最大化预期收益，写为J(①)=Ee[Zt=ort]", "在单词级RL中，每个输出单词都被视为一个操作步骤，产生以下策略梯度：", "TU", "VeJ(0) = Ee[ Rt; Ve log pe(wtijlw<tj, Ct)]", "(2)", "t=0 j=0", "其中T是对话中的总回合数，Ut是t回合响应中的令牌总数，i是每个令牌的索引w。Rti表", "示第个令牌在t轮的折扣回报。在这种策略梯度形式中，动作空间是系统VI的词汇量，轨", "迹长度是t=Ut，使得该空间中的RL极具挑战性", "引入潜在变量z使我们能够将条件分布分解为p(αc)=p(α2)p(zc)。通过将潜在空间视", "为动作空间，减小了动作空间大小和轨迹长度（Zhaoetal.，2019）。对于与T回合的对话，策", "略梯度现在由", "T", "VeJ(0) = Ee[ RtVe log pe(ztlct)],", "(3)", "t=0", "给出", "其中Rt表示t轮的折扣收益。Zhaoetal.（2019）研究了两种类型的潜在变量；分类变量和连", "续变量。分类潜变量的形式为M独立的K方式分类随机变量，而连续变量的形式为具有对角", "协方差矩阵的M维多元高斯分布。通过最大化证据下界（ELBO）在SL步骤中使用随机变分", "推理来学习潜在变量分布-数据日志似然的下界，", "Lfull(①) = Eqe(z/r,c) [log Pe(α2)] -DKL[qe(2,c)pe(z]c)]", "(4)", "为了对抗曝光偏差,EZhaoetal.（2019）通过假设后验qe(z,c）与编码器pe(zc）相同，引入", "了ELBO的“精简版”版本。这消除了ELBO目标的第二项。为了抵消过度拟合，后验用一些", "权重β进行正则化，以类似于某些先验，在这种情况下，分类潜在变量的均匀分布或连续变量", "的正态分布。“lite”ELBO目标已被证明优于完整的ELBO目标，并写成：", "Llite(0) = Epe(2/c)[log Pe(2)] -βDkL[pe(z/c)lp(2)]", "(5)", "4LatentActionSpaceviaAuxiliaryTask", "如等式（3）所示，为了促进有效的学习，重要的是使用有意义地表示动作的潜在变量来训练", "策略。但是，由于现有方法对zw的分布进行建模。r。t。上下文c（方程（5），我们怀疑潜在", "变量是状态空间的表示，而不是动作空间的表示。此外，使用不知情的先验对分布进行正则化", "而不考虑给定上下文中对话响应的实际分布。为厂诱导更像动作的潜在表示，我们在一个辅助", "任务上训练模型，该任务需要知道要执行的响应。我们使用变分自动编码（VAE）模型选择响", "应自动编码（AE）作为辅助任务。也就是说，给定一个响应&，我们训练模型通过编码器和解", "码器之间的潜在空间重建响应（图1a）。对于未知的先验p(2），一组参数Φ的预训练目标是：", "Lvae() = Eq(2/)[log P(|2)] - DkL[q(zc)lp(2)].", "(6)", "VAE模型已被证明能够捕获所训练样本的生成方面，从而在潜在变量（Kingma and Welling，", "2014；Bowmanetal.，2016）之间产生良好的插值。通过在对话响应上训练VAE，我们的目标是", "以无监督的方式捕获响应的生成方面，例如意图和域信息" ]

{ "arxiv_id": "2011.09378", "language": "zh", "page_idx": 2 }

[ [ "0.3914", "0.08075", "0.3997", "0.0873", "0.3938", "0.09265", "0.3855", "0.0861" ], [ "0.5557", "0.08075", "0.571", "0.08075", "0.571", "0.09143", "0.5557", "0.09143" ], [ "0.2301", "0.09143", "0.3064", "0.09143", "0.3064", "0.1057", "0.2301", "0.1057" ], [ "0.266", "0.082", "0.2795", "0.082", "0.2795", "0.0849", "0.266", "0.0849" ], [ "0.4224", "0.0944", "0.4358", "0.0944", "0.4358", "0.10095", "0.4224", "0.10095" ], [ "0.692", "0.0956", "0.702", "0.0956", "0.702", "0.10095", "0.692", "0.10095" ], [ "0.748", "0.0956", "0.7573", "0.0956", "0.7573", "0.10095", "0.748", "0.10095" ], [ "0.2451", "0.1134", "0.297", "0.1134", "0.297", "0.1289", "0.2451", "0.1289" ], [ "0.434", "0.1152", "0.486", "0.1152", "0.486", "0.13", "0.434", "0.13" ], [ "0.705", "0.1134", "0.754", "0.1134", "0.754", "0.1289", "0.705", "0.1289" ], [ "0.4575", "0.1395", "0.476", "0.1395", "0.476", "0.1467", "0.4575", "0.1467" ], [ "0.7305", "0.1401", "0.742", "0.1401", "0.742", "0.1443", "0.7305", "0.1443" ], [ "0.617", "0.1567", "0.6304", "0.1567", "0.6304", "0.1603", "0.617", "0.1603" ], [ "0.748", "0.1716", "0.785", "0.1716", "0.785", "0.1787", "0.748", "0.1787" ], [ "0.2544", "0.1924", "0.2896", "0.1924", "0.2896", "0.2079", "0.2544", "0.2079" ], [ "0.4434", "0.1953", "0.476", "0.1953", "0.476", "0.2073", "0.4434", "0.2073" ], [ "0.714", "0.1953", "0.7466", "0.1953", "0.7466", "0.2073", "0.714", "0.2073" ], [ "0.6963", "0.2245", "0.702", "0.2245", "0.702", "0.2299", "0.6963", "0.2299" ], [ "0.748", "0.2257", "0.7617", "0.2286", "0.7583", "0.2322", "0.7466", "0.2299" ], [ "0.251", "0.2322", "0.2957", "0.2322", "0.2957", "0.2434", "0.251", "0.2434" ], [ "0.4802", "0.228", "0.4895", "0.228", "0.4895", "0.231", "0.4802", "0.231" ], [ "0.7188", "0.234", "0.748", "0.234", "0.748", "0.2411", "0.7188", "0.2411" ], [ "0.12006", "0.2583", "0.882", "0.2583", "0.882", "0.2715", "0.12006", "0.2715" ], [ "0.1184", "0.275", "0.882", "0.275", "0.882", "0.288", "0.1184", "0.288" ], [ "0.12006", "0.2903", "0.882", "0.2903", "0.882", "0.3047", "0.12006", "0.3047" ], [ "0.12177", "0.3047", "0.8833", "0.3047", "0.8833", "0.3196", "0.12177", "0.3196" ], [ "0.1184", "0.3206", "0.8833", "0.3218", "0.8833", "0.3362", "0.1184", "0.335" ], [ "0.1184", "0.3374", "0.882", "0.3384", "0.882", "0.3528", "0.1184", "0.3516" ], [ "0.12177", "0.3557", "0.4114", "0.3557", "0.4114", "0.3665", "0.12177", "0.3665" ], [ "0.1377", "0.3943", "0.882", "0.3955", "0.882", "0.4104", "0.1377", "0.4092" ], [ "0.12006", "0.412", "0.882", "0.412", "0.882", "0.4258", "0.12006", "0.4258" ], [ "0.12177", "0.43", "0.881", "0.43", "0.881", "0.4407", "0.12177", "0.4407" ], [ "0.12177", "0.4448", "0.882", "0.4448", "0.882", "0.4578", "0.12177", "0.4578" ], [ "0.12006", "0.4614", "0.882", "0.4614", "0.882", "0.4744", "0.12006", "0.4744" ], [ "0.12177", "0.4785", "0.881", "0.4785", "0.881", "0.4893", "0.12177", "0.4893" ], [ "0.12006", "0.4934", "0.671", "0.4934", "0.671", "0.5073", "0.12006", "0.5073" ], [ "0.11755", "0.5186", "0.413", "0.5195", "0.413", "0.5337", "0.11755", "0.533" ], [ "0.12006", "0.5415", "0.882", "0.5415", "0.882", "0.5547", "0.12006", "0.5547" ], [ "0.1184", "0.557", "0.882", "0.5557", "0.882", "0.571", "0.1184", "0.572" ], [ "0.12006", "0.5737", "0.882", "0.5737", "0.882", "0.5874", "0.12006", "0.5874" ], [ "0.12006", "0.589", "0.882", "0.589", "0.882", "0.6025", "0.12006", "0.6025" ], [ "0.1184", "0.605", "0.8833", "0.605", "0.8833", "0.619", "0.1184", "0.619" ], [ "0.12177", "0.6216", "0.882", "0.6216", "0.882", "0.6353", "0.12177", "0.6353" ], [ "0.12006", "0.6357", "0.881", "0.6357", "0.881", "0.6494", "0.12006", "0.6494" ], [ "0.12006", "0.654", "0.4182", "0.654", "0.4182", "0.6675", "0.12006", "0.6675" ], [ "0.1184", "0.6797", "0.4768", "0.6797", "0.4768", "0.6934", "0.1184", "0.6934" ], [ "0.12177", "0.7036", "0.881", "0.7036", "0.881", "0.7144", "0.12177", "0.7144" ], [ "0.12177", "0.7173", "0.879", "0.7173", "0.879", "0.731", "0.12177", "0.731" ], [ "0.12177", "0.734", "0.881", "0.734", "0.881", "0.7476", "0.12177", "0.7476" ], [ "0.12177", "0.7505", "0.882", "0.7505", "0.882", "0.764", "0.12177", "0.764" ], [ "0.12177", "0.7656", "0.881", "0.7656", "0.881", "0.7793", "0.12177", "0.7793" ], [ "0.1184", "0.782", "0.882", "0.781", "0.882", "0.796", "0.1184", "0.7964" ], [ "0.2644", "0.8135", "0.739", "0.8135", "0.739", "0.8276", "0.2644", "0.8276" ], [ "0.8564", "0.811", "0.885", "0.811", "0.885", "0.829", "0.8564", "0.829" ], [ "0.12177", "0.843", "0.6035", "0.843", "0.6035", "0.8555", "0.12177", "0.8555" ], [ "0.12006", "0.8633", "0.8833", "0.8623", "0.8833", "0.877", "0.12006", "0.8774" ], [ "0.12006", "0.88", "0.882", "0.88", "0.882", "0.8936", "0.12006", "0.8936" ], [ "0.12006", "0.8955", "0.882", "0.8955", "0.882", "0.91", "0.12006", "0.91" ] ]

[ "", "c)", "", "", "\"t.1", "", "", "decoder", "decoder", "decoder", "12", ".2", "", "Pe(ZtIc", "enc s", "enco", "ence", "t.", "t", "response", "l", "ontext", "Figure l: We aim to train an action-characterized distribution of latent variables Z to support RL", "in an end-to-end setting. a) VAE pre-training, b) new encoder is initialized and connected to the", "pre-trained Z and VAE decoder: Overall fine-tuning optimizes the entire network (LAVA_ptA),", "while selective fine-tuning optimizes only the new encoder (LAVA_ptS), c) new encoder is", "initialized and used in tandem with VAE encoder to obtain informative prior (LAVA_kl). The", "same architecture is also used for multitask learning (LAVA_mt), where we optimize for both", "tasksatthesametimefrom scratch", "我们建议利用VAE潜在表示来调节对话系统，以将编码的对话状态映射到潜在动作，而不是", "学习对话状态的潜在表示。我们称这种方法为LAVA（通过VAE的潜在行动）。我们探索了三", "种利用AE任务来诱导以动作为特征的潜在变量分布的方法：作为预训练，作为知情的先验以", "及以多任务学习的方式-如图E1所示。请注意，可以将AE任务与针对对话响应的表示学习的", "其他任务交换。在这项工作中，我们利用简单的递归模型作为编码器和解码器来突出潜在对话", "动作空间的作用。这使我们能够将观察到的任何改进都集中在潜在的空间和对话政策上。。端到", "端对话系统框架的其他部分，如编码和解码，不在这项工作的范围内", "4.1Auxiliary Task as Pre-training", "第一种方法利用辅助AE任务预训练潜在表示和解码器（图E1b）。由于用户和系统的词汇表", "相差很大，因此初始化了用于响应生成（RG）的新对话系统编码器，并丢弃了VAE编码器。", "新的编码器连接到潜在空间和VAE的解码器。我们尝试了两种微调方案：整体（LAVA_ptA", "和选择性（LAVA_ptS）。使用LAVA_ptA，我们在微调期间更新网络的所有参数。另一方面，", "LAVA_ptS阻止梯度传播到潜在空间和解码器，并专门训练编码器。这相当于利用来自VAE", "的生成模块作为NLG模块，鼓励模型专注于对对话状态进行编码（类似于状态跟踪器），并在", "RG训练期间将其映射到相应的潜在动作（类似于策略）。LAVA_ptA和LAVAptS的损失函", "数均为“lite”ELBO主观（方程5）", "4.2AuxiliaryTask as anInformativePrio", "其次，我们明确地训练RG潜在变量分布，使其接近VAE的分布。我们将此设置称为LAVAkI。", "和以前一样，我们从预训练的VAE和新初始化的RG编码器开始。我们以一种新颖的方式利用", "预先训练的潜在表示；VAE编码器不会被丢弃，而是串联使用以获得目标响应的知情先验，如", "图E1c所示。我们在KL项惩罚中使用以目标qα（2α）为条件的潜在分布，取代了以前工作中", "使用的不知情先验。通过惩罚分歧，同时仍对其进行优化以适应对话上下文，从而将RG潜在", "变量分布建立在VAE的基础上。。除VAE编码器外的所有参数都通过以下ELBO进一步优化：", "LLAVA_ki(0) =Epe(z/)[log Pe(αz)) -βDKL[Pe(zc)lq(zc)].", "(7)", "4.3Multitask Training between Main and Auxiliary Tasks", "最后，我们训练一个模型，以多任务方式解决RG和AE任务，其中RG被视为主要任务，AE", "被视为辅助任务。多任务学习旨在通过一个模型一次解决多个任务，利用任务之间的相似性来提", "高学习效率（Caruana，1997）。最近的工作表明，对话系统任务也受益于多任务学习（Rastogi" ]

{ "arxiv_id": "2011.09378", "language": "zh", "page_idx": 3 }

[ [ "0.12006", "0.0766", "0.881", "0.0766", "0.881", "0.0903", "0.12006", "0.0903" ], [ "0.12006", "0.09265", "0.8833", "0.0932", "0.8833", "0.1081", "0.12006", "0.1069" ], [ "0.12177", "0.10986", "0.882", "0.10986", "0.882", "0.12354", "0.12177", "0.12354" ], [ "0.12177", "0.1259", "0.882", "0.1259", "0.882", "0.1395", "0.12177", "0.1395" ], [ "0.1184", "0.1401", "0.882", "0.1414", "0.882", "0.1561", "0.1184", "0.155" ], [ "0.12006", "0.158", "0.881", "0.158", "0.881", "0.1716", "0.12006", "0.1716" ], [ "0.1234", "0.1729", "0.882", "0.1729", "0.882", "0.1859", "0.1234", "0.1859" ], [ "0.12006", "0.1907", "0.882", "0.1907", "0.882", "0.2037", "0.12006", "0.2037" ], [ "0.12006", "0.206", "0.601", "0.206", "0.601", "0.2197", "0.12006", "0.2197" ], [ "0.2644", "0.234", "0.739", "0.234", "0.739", "0.2488", "0.2644", "0.2488" ], [ "0.858", "0.234", "0.8833", "0.234", "0.8833", "0.2488", "0.858", "0.2488" ], [ "0.2551", "0.2588", "0.7466", "0.26", "0.7466", "0.2808", "0.2551", "0.2798" ], [ "0.858", "0.2625", "0.8833", "0.2625", "0.8833", "0.2766", "0.858", "0.2766" ], [ "0.11584", "0.2861", "0.3132", "0.288", "0.3132", "0.306", "0.11584", "0.304" ], [ "0.1184", "0.3123", "0.44", "0.3142", "0.44", "0.3289", "0.1184", "0.3267" ], [ "0.12006", "0.3342", "0.882", "0.3342", "0.882", "0.3486", "0.12006", "0.3486" ], [ "0.12177", "0.3508", "0.881", "0.3508", "0.881", "0.364", "0.12177", "0.364" ], [ "0.12177", "0.3677", "0.881", "0.3677", "0.881", "0.3813", "0.12177", "0.3813" ], [ "0.1234", "0.383", "0.882", "0.383", "0.882", "0.3967", "0.1234", "0.3967" ], [ "0.12006", "0.3984", "0.881", "0.3984", "0.881", "0.412", "0.12006", "0.412" ], [ "0.12006", "0.4146", "0.882", "0.4146", "0.882", "0.4275", "0.12006", "0.4275" ], [ "0.12177", "0.4312", "0.882", "0.4312", "0.882", "0.4448", "0.12177", "0.4448" ], [ "0.12177", "0.4465", "0.881", "0.4465", "0.881", "0.4614", "0.12177", "0.4614" ], [ "0.1234", "0.4644", "0.881", "0.4644", "0.881", "0.4744", "0.1234", "0.4744" ], [ "0.1234", "0.481", "0.881", "0.481", "0.881", "0.4912", "0.1234", "0.4912" ], [ "0.12006", "0.4954", "0.881", "0.4954", "0.881", "0.509", "0.12006", "0.509" ], [ "0.12006", "0.511", "0.881", "0.511", "0.881", "0.525", "0.12006", "0.525" ], [ "0.12006", "0.528", "0.882", "0.528", "0.882", "0.5415", "0.12006", "0.5415" ], [ "0.1184", "0.542", "0.882", "0.5435", "0.882", "0.558", "0.1184", "0.557" ], [ "0.12177", "0.5586", "0.881", "0.5586", "0.881", "0.5723", "0.12177", "0.5723" ], [ "0.12177", "0.576", "0.882", "0.576", "0.882", "0.589", "0.12177", "0.589" ], [ "0.12006", "0.593", "0.879", "0.593", "0.879", "0.6055", "0.12006", "0.6055" ], [ "0.1234", "0.6094", "0.5615", "0.6094", "0.5615", "0.619", "0.1234", "0.619" ], [ "0.559", "0.608", "0.882", "0.608", "0.882", "0.6216", "0.559", "0.6216" ], [ "0.12177", "0.626", "0.4492", "0.626", "0.4492", "0.6357", "0.12177", "0.6357" ], [ "0.4895", "0.626", "0.881", "0.626", "0.881", "0.6357", "0.4895", "0.6357" ], [ "0.12177", "0.6426", "0.451", "0.6426", "0.451", "0.653", "0.12177", "0.653" ], [ "0.4358", "0.6455", "0.4517", "0.6455", "0.4517", "0.6484", "0.4358", "0.6484" ], [ "0.4634", "0.6455", "0.4878", "0.6455", "0.4878", "0.6484", "0.4634", "0.6484" ], [ "0.513", "0.6455", "0.529", "0.6455", "0.529", "0.6484", "0.513", "0.6484" ], [ "0.54", "0.6426", "0.879", "0.6426", "0.879", "0.653", "0.54", "0.653" ], [ "0.12177", "0.6577", "0.4248", "0.6577", "0.4248", "0.6685", "0.12177", "0.6685" ], [ "0.4417", "0.661", "0.476", "0.661", "0.476", "0.665", "0.4417", "0.665" ], [ "0.503", "0.6567", "0.882", "0.6567", "0.882", "0.6704", "0.503", "0.6704" ], [ "0.12006", "0.673", "0.534", "0.673", "0.534", "0.686", "0.12006", "0.686" ], [ "0.538", "0.676", "0.5825", "0.676", "0.5825", "0.683", "0.538", "0.683" ], [ "0.5767", "0.673", "0.858", "0.673", "0.858", "0.686", "0.5767", "0.686" ], [ "0.1368", "0.687", "0.882", "0.6885", "0.882", "0.703", "0.1368", "0.7017" ], [ "0.1184", "0.703", "0.882", "0.7036", "0.882", "0.7188", "0.1184", "0.7173" ], [ "0.12006", "0.72", "0.882", "0.72", "0.882", "0.734", "0.12006", "0.734" ], [ "0.1234", "0.7383", "0.882", "0.7383", "0.882", "0.749", "0.1234", "0.749" ], [ "0.12006", "0.753", "0.5693", "0.753", "0.5693", "0.7666", "0.12006", "0.7666" ], [ "0.12177", "0.78", "0.2183", "0.78", "0.2183", "0.7915", "0.12177", "0.7915" ], [ "0.12177", "0.8", "0.881", "0.8", "0.881", "0.8135", "0.12177", "0.8135" ], [ "0.12006", "0.8154", "0.882", "0.8154", "0.882", "0.829", "0.12006", "0.829" ], [ "0.12177", "0.832", "0.882", "0.832", "0.882", "0.8457", "0.12177", "0.8457" ], [ "0.12177", "0.847", "0.8833", "0.847", "0.8833", "0.861", "0.12177", "0.861" ], [ "0.12177", "0.8633", "0.882", "0.8633", "0.882", "0.8774", "0.12177", "0.8774" ], [ "0.12177", "0.88", "0.8833", "0.88", "0.8833", "0.8936", "0.12177", "0.8936" ], [ "0.1184", "0.8945", "0.882", "0.8955", "0.882", "0.909", "0.1184", "0.9077" ] ]

[ "etal.，2018；Zhuetal,2019）。RG和AE任务的相似之处在于它们都旨在生成对话响应a，但", "是它们在所考虑的上下文中有所不同，也称为多对一多任务设置（Luongetal.，2015）。RG尝", "试在给定对话上下文c的情况下生成目标响应&，并且AE尝试在给定响应的情况下执行重", "建。这两个任务共享具有一组参数w的潜在空间和解码器，但具有用于RG和AE的单独编码", "器，分别具有参数和Φ（图国1c）。目的是潜在表示编码更多的动作特征，因为这些是完成", "这两个任务所需的公共信息。这两个任务以交替的方式进行训练，比例为A：B，即。e。对于", "主任务的每一次A迭代，我们用B迭代的辅助任务进行训练。与以前的方法不同，在多任务学", "习中，我们从没有预先训练的新初始化模型开始。每个编码器仅从其相应的任务接收更新，而", "潜在表示和解码器在这两个任务上进行训练。ELBO目标是", "LEAVA mt(w, 0) = Epe(z/)[log Pw(α|2)] - βDKL[pe(zc)llp(2)],", "(8)", "LVA μmt(w, ) = Ea(2/)[1og Pu(a|2)] - βDKL[9(2/a)/p(2)].", "(9)", "5Experiment Setup", "5.1 Corpus, Task, and Training Setup", "我们使用MultiWOZ2。0语料库（Budzianowskietal.，2018）测试模型的性能。MultiWOZ是", "人类之间以“绿野仙踪”方式进行对话的集合，其中一个人扮演对话系统的角色，另一个人扮", "演用户。用户的任务是查找实体，例如。g。通过与对话系统交互来符合某些标准的餐厅或酒店。", "该语料库模拟了一个多领域的面向任务的对话系统交互，即。e。多个域可能出现在同一个对话", "中，甚至出现在同一个回合中。该语料库完全注释了总共10438个英语对话，它是同类中最具", "挑战性和最大的语料库之一。我们使用语料库中提供的训练，验证和测试集分区，总计8438个", "对话用于训练，每个对话1000个用于验证和测试。报告的所有数字均基于对测试集的评估。我", "们的目标是训练一个潜在的行动表示，在端到端环境中有效地优化与RL的对话策略。这一目", "标最好通过完成基本对话任务来体现，以对话级别的匹配和成功率来衡量。匹配率计算系统是", "否匹配用户通知的条件（可提供信息的插槽），成功率计算系统是否提供用户请求的信息（可请", "求的插槽）。比赛是成功对话的先决条件。长期以来，对话政策的研究只关注成功率和用户满意", "度（LeeandEskénazi，2012；Ultesetal.，2017），但随着政策和NLG之间的界限变得模糊，我", "们看到了BLEU和困惑等指标的引入然而，这些早期被标记为可能具有误导性，因为它们与人", "类判断的相关性很差（Stentetal.，2005；Liuetal.，2016）。虽然我们也报告了BLEU分数的完", "整性，但请注意，我们的方法不是针对改善BLEU。我们研究了两个任务：1）“对话上下文到响", "应的生成”，即在给定对话历史的情况下生成下一个对话响应，以及来自语料库的oracle对话状", "态和数据库指针。对话状态是从对话开始到当前回合推断的用户目标的二进制向量表示。另一", "方面，数据库指针是与对话状态中的条件匹配的实体的", "二进制向量表示。2）“端到端建模”，仅", "采用对历史来生成响应。端到端建模中", "作通常利用管道中的中间模型来预测标签，例如", "对话状态和数据库指针。在这项工作中，", "", "", "", "可端的方式利用潜在作用，而不使用任何中", "间标签，鼓励模型充分利用潜在变量。", "作", "验都是在去词汇化的对话中进行的，其中槽值的", "出现被其相应的槽标记替换，例如“在城镇的南部”", "变为", "“在城镇的[value_area]部分中”。", "我们的培训包括两个步骤：4节中介绍的技术，然后是使用REINFORCE的RL。为了与现", "有作品进行公平的比较，我们采用了Zhaoetal.（2019）提出的新颖的RL设置：1）对于每个", "RL情节，从语料库中采样对话。2）运行模型以生成每个系统回合的响应。但是，对话中的下", "一个用户回合不会更改，而只是从语料库中检索。3）根据系统响应计算对话的成功率，并将其", "用作奖励信号来计算策略梯度和更新模型参数（公式3）", "5.2Model", "我们的主要重点是所提出的方法引起的潜在表示，它们对强化学习的影响以及模型的最终性能。", "为了突出潜在对话动作空间的作用，我们使用简单的递归模型作为VAE和对话系统的编码器", "和解码器。我们将模型词汇表限制为最频繁的1000个令牌。对于上下文到响应任务，我们将对", "话历史截断到最后2轮，对于端到端建模任务，我们将最后4轮截断。对于AE和RG任务，", "编码器是具有注意力和大小300的GRU-LSTME(Choetal.，2014），其输出大小为600的向", "量。我们测试了分类和连续潜在变量。分类潜在空间是10个独立的20路分类变量（M=10", "K=20），连续空间的大小设置为M=200。在分类情况下，解码器的大小为150，在连续情" ]

{ "arxiv_id": "2011.09378", "language": "zh", "page_idx": 4 }

[ [ "0.2484", "0.2172", "0.749", "0.2172", "0.749", "0.2349", "0.2484", "0.2349" ], [ "0.232", "0.2588", "0.3945", "0.2588", "0.3945", "0.2734", "0.232", "0.2734" ], [ "0.392", "0.2588", "0.6216", "0.2588", "0.6216", "0.2734", "0.392", "0.2734" ], [ "0.615", "0.2615", "0.6543", "0.2615", "0.6543", "0.2715", "0.615", "0.2715" ], [ "0.648", "0.2588", "0.7607", "0.2588", "0.7607", "0.2734", "0.648", "0.2734" ], [ "0.2378", "0.2903", "0.7607", "0.2903", "0.7607", "0.305", "0.2378", "0.305" ], [ "0.46", "0.308", "0.5376", "0.308", "0.5376", "0.3225", "0.46", "0.3225" ], [ "0.2632", "0.3257", "0.7344", "0.3257", "0.7344", "0.3403", "0.2632", "0.3403" ], [ "0.4673", "0.3757", "0.5312", "0.3774", "0.5293", "0.3933", "0.4656", "0.3914" ], [ "0.2737", "0.4016", "0.744", "0.4016", "0.744", "0.4153", "0.2737", "0.4153" ], [ "0.25", "0.4185", "0.7417", "0.4185", "0.7417", "0.4287", "0.25", "0.4287" ], [ "0.25", "0.432", "0.746", "0.432", "0.746", "0.4458", "0.25", "0.4458" ], [ "0.2484", "0.4463", "0.7466", "0.4463", "0.7466", "0.4602", "0.2484", "0.4602" ], [ "0.25", "0.4622", "0.746", "0.4622", "0.746", "0.476", "0.25", "0.476" ], [ "0.2467", "0.4766", "0.7466", "0.476", "0.7466", "0.4905", "0.2467", "0.4912" ], [ "0.2484", "0.4924", "0.746", "0.4924", "0.746", "0.506", "0.2484", "0.506" ], [ "0.2484", "0.506", "0.7466", "0.507", "0.7466", "0.5215", "0.2484", "0.521" ], [ "0.25", "0.523", "0.746", "0.523", "0.746", "0.536", "0.25", "0.536" ], [ "0.25", "0.537", "0.7466", "0.537", "0.7466", "0.552", "0.25", "0.552" ], [ "0.25", "0.553", "0.7466", "0.553", "0.7466", "0.5664", "0.25", "0.5664" ], [ "0.2507", "0.5684", "0.7417", "0.5684", "0.7417", "0.582", "0.2507", "0.582" ], [ "0.2484", "0.582", "0.7466", "0.5835", "0.7466", "0.5977", "0.2484", "0.5967" ], [ "0.25", "0.5977", "0.746", "0.5977", "0.746", "0.6123", "0.25", "0.6123" ], [ "0.2484", "0.6113", "0.7466", "0.6113", "0.7466", "0.6255", "0.2484", "0.6255" ], [ "0.2467", "0.627", "0.5703", "0.6274", "0.5703", "0.6426", "0.2467", "0.6416" ], [ "0.2009", "0.6675", "0.3718", "0.6685", "0.3718", "0.6846", "0.2009", "0.683" ], [ "0.2051", "0.702", "0.79", "0.702", "0.79", "0.7153", "0.2051", "0.7153" ], [ "0.2051", "0.7183", "0.79", "0.7173", "0.79", "0.732", "0.2051", "0.733" ], [ "0.2075", "0.7363", "0.79", "0.7363", "0.79", "0.75", "0.2075", "0.75" ], [ "0.2059", "0.7524", "0.79", "0.7524", "0.79", "0.7666", "0.2059", "0.7666" ], [ "0.2059", "0.771", "0.79", "0.771", "0.79", "0.784", "0.2059", "0.784" ], [ "0.2059", "0.787", "0.79", "0.787", "0.79", "0.8003", "0.2059", "0.8003" ], [ "0.2051", "0.803", "0.79", "0.8037", "0.79", "0.8184", "0.2051", "0.8174" ], [ "0.2051", "0.821", "0.367", "0.8213", "0.367", "0.836", "0.2051", "0.8354" ], [ "0.492", "0.8726", "0.5034", "0.8726", "0.5034", "0.884", "0.492", "0.884" ] ]

[ "基于批准的多获胜者选举中的审议和投票", "KanavMehral", "NandaKishore Sreenivasl", "", "Kate Larson", "IDavid R. Cheriton School of Computer Science, University of", "Waterloo", "kanav.mehra, nksreenivas, kate.larson ↓ @uwaterloo.ca", "Abstract", "Citizen-focused democratic processes where participants deliberate", "on alternatives and then vote to make the final decision are increas", "ingly popular today. While the computational social choice literature", "has extensively investigated voting rules, there is limited work that", "explicitly looks at the interplay of the deliberative process and voting.", "In this paper, we build a deliberation model using established models", "from the opinion-dynamics literature and study the effect of differ-", "ent deliberation mechanisms on voting outcomes achieved when using", "well-studied voting rules. Our results show that deliberation generally", "improves welfare and representation guarantees, but the results are", "sensitive to how the deliberation process is organized. We also show,", "experimentally, that simple voting rules, such as approval voting, per", "form as well as more sophisticated rules such as proportional approval", "voting [46] or method of equal shares [38, 37] if deliberation is properly", "supported. This has ramifications on the practical use of such voting", "rules in citizen-focused democratic processes.", "1Introduction", "场景无处不在，其中必须选择一个委员会来代表某个更大群体的利益，从议", "会选举和参与性预算（PB）[11]等政治领域到技术应用，例如设计推荐系统", "[43]和多样化搜索结果[44]。在社会选择文献中已经对多赢家投票进行了很", "好的研究，重点是了解如何选择“最佳”委员会。。但是，即使定义“最佳”", "的含义也不是一件容易的事。。在某些情况下，例如专家判断的汇总，可能", "需要一个由评分最高的k替代方案组成的委员会。然而，在其他任务中，例", "如选择k位置来建造公共设施（e。g。医院、消防局），最好确保尽可能多的", "选民可以使用该设施", "1" ]

{ "arxiv_id": "2305.08970", "language": "zh", "page_idx": 0 }

[ [ "0.232", "0.1666", "0.7915", "0.1674", "0.7915", "0.1818", "0.232", "0.1812" ], [ "0.2059", "0.1849", "0.511", "0.1849", "0.511", "0.1989", "0.2059", "0.1989" ], [ "0.5337", "0.1863", "0.7876", "0.1863", "0.7876", "0.1963", "0.5337", "0.1963" ], [ "0.2091", "0.2026", "0.3381", "0.2026", "0.3381", "0.2134", "0.2091", "0.2134" ], [ "0.335", "0.2058", "0.3635", "0.2058", "0.3635", "0.2102", "0.335", "0.2102" ], [ "0.603", "0.2008", "0.7876", "0.202", "0.7876", "0.2153", "0.603", "0.214" ], [ "0.2091", "0.2203", "0.339", "0.2203", "0.339", "0.2311", "0.2091", "0.2311" ], [ "0.5864", "0.2229", "0.5986", "0.2229", "0.5986", "0.2291", "0.5864", "0.2291" ], [ "0.596", "0.2203", "0.7876", "0.2203", "0.7876", "0.2311", "0.596", "0.2311" ], [ "0.2059", "0.2374", "0.3325", "0.2374", "0.3325", "0.2474", "0.2059", "0.2474" ], [ "0.613", "0.2374", "0.7876", "0.2374", "0.7876", "0.2474", "0.613", "0.2474" ], [ "0.2075", "0.2524", "0.335", "0.2524", "0.335", "0.2664", "0.2075", "0.2664" ], [ "0.6274", "0.2524", "0.789", "0.2524", "0.789", "0.2664", "0.6274", "0.2664" ], [ "0.2035", "0.2695", "0.3245", "0.2695", "0.3245", "0.2827", "0.2035", "0.2827" ], [ "0.3203", "0.2747", "0.353", "0.2747", "0.353", "0.2778", "0.3203", "0.2778" ], [ "0.6274", "0.2695", "0.789", "0.2695", "0.789", "0.2827", "0.6274", "0.2827" ], [ "0.2091", "0.2874", "0.33", "0.2874", "0.33", "0.2974", "0.2091", "0.2974" ], [ "0.6396", "0.2886", "0.7876", "0.2886", "0.7876", "0.2986", "0.6396", "0.2986" ], [ "0.2059", "0.303", "0.4844", "0.303", "0.4844", "0.3164", "0.2059", "0.3164" ], [ "0.527", "0.3037", "0.789", "0.3037", "0.789", "0.3176", "0.527", "0.3176" ], [ "0.2059", "0.3208", "0.79", "0.3208", "0.79", "0.334", "0.2059", "0.334" ], [ "0.2059", "0.3384", "0.79", "0.3384", "0.79", "0.3516", "0.2059", "0.3516" ], [ "0.2059", "0.356", "0.79", "0.356", "0.79", "0.3694", "0.2059", "0.3694" ], [ "0.2051", "0.3718", "0.7915", "0.3726", "0.7915", "0.387", "0.2051", "0.3865" ], [ "0.2059", "0.3901", "0.79", "0.3901", "0.79", "0.404", "0.2059", "0.404" ], [ "0.2059", "0.4072", "0.79", "0.4072", "0.79", "0.4204", "0.2059", "0.4204" ], [ "0.2059", "0.4236", "0.79", "0.4236", "0.79", "0.4375", "0.2059", "0.4375" ], [ "0.2051", "0.44", "0.789", "0.44", "0.789", "0.4539", "0.2051", "0.4539" ], [ "0.2051", "0.4583", "0.7915", "0.4583", "0.7915", "0.4717", "0.2051", "0.4717" ], [ "0.2059", "0.476", "0.2754", "0.476", "0.2754", "0.4873", "0.2059", "0.4873" ], [ "0.2018", "0.4905", "0.7915", "0.4912", "0.7915", "0.506", "0.2018", "0.505" ], [ "0.2059", "0.509", "0.79", "0.509", "0.79", "0.523", "0.2059", "0.523" ], [ "0.2051", "0.5264", "0.789", "0.5264", "0.789", "0.5405", "0.2051", "0.5405" ], [ "0.2059", "0.5444", "0.79", "0.5444", "0.79", "0.558", "0.2059", "0.558" ], [ "0.2059", "0.561", "0.79", "0.561", "0.79", "0.5747", "0.2059", "0.5747" ], [ "0.2059", "0.5776", "0.79", "0.5776", "0.79", "0.5913", "0.2059", "0.5913" ], [ "0.2051", "0.5933", "0.789", "0.5947", "0.789", "0.6094", "0.2051", "0.608" ], [ "0.2051", "0.6113", "0.2778", "0.6113", "0.2778", "0.6255", "0.2051", "0.6255" ], [ "0.2336", "0.6274", "0.79", "0.6274", "0.79", "0.6426", "0.2336", "0.6426" ], [ "0.2059", "0.646", "0.789", "0.646", "0.789", "0.659", "0.2059", "0.659" ], [ "0.2075", "0.6636", "0.79", "0.6636", "0.79", "0.677", "0.2075", "0.677" ], [ "0.2051", "0.6787", "0.79", "0.6797", "0.79", "0.6943", "0.2051", "0.6934" ], [ "0.2059", "0.6978", "0.79", "0.6978", "0.79", "0.711", "0.2059", "0.711" ], [ "0.2059", "0.7153", "0.79", "0.7153", "0.79", "0.729", "0.2059", "0.729" ], [ "0.2059", "0.732", "0.437", "0.732", "0.437", "0.7456", "0.2059", "0.7456" ], [ "0.232", "0.7476", "0.79", "0.7485", "0.79", "0.763", "0.232", "0.762" ], [ "0.2059", "0.7666", "0.6855", "0.7666", "0.6855", "0.78", "0.2059", "0.78" ], [ "0.4927", "0.8726", "0.5034", "0.8726", "0.5034", "0.884", "0.4927", "0.884" ] ]

[ "在以公民为中心的民主进程中，例如公民大会[18]和参与性预算[11]", "存在广泛的讨论范围，涉及多种可能的", "代方案。例如，审议是大多数参与", "性预算实施的重享", "", "的偏好并促进信息交流", "自的是达成共识", "20", "的重要组成部分，但它", "不能完全取代投", "致。仍然必须做出决定", "因此，我们认为", "重要的。。为此，我们", "(1)引人了基于", "", "审议机制对使用经过", "充分研究的投票", "于基于批准的选举", "选民通过分享一部分批准的候选人来", "偏好。批准选票由于其简单性和灵", "活性而在实践中使用[7，6，4]。它们还提供了审议的范围，因为选民经常要", "在许多不同的选择之间做出决定。我们提出了一种基于代理的审议模型，并", "探索了构建审议小组的各种替代方案。。我们评估标准的多赢家投票规则，", "在选民有机会审议之前和之后，就文献中的标准目标，包括社会福利，代表", "性和比例性。我们表明，在几乎所有情况下，审议都会显着改善福利，代表", "性和相称性。然而，结果对审议机制很敏感；增加对不同意见（或来自不同", "背景的代理人）的接触会导致更好的结果。更重要的是，我们的研究结果表", "明，在有效审议的情况下，简单，可解释的投票规则，如批准投票，以及更", "复杂，更复杂的规则。这可以指导以公民为中心的民主进程中投票规则的设", "计和部署", "RelatedWork：社会选择文献广泛研究了基于批准的多赢家投票规则的质", "量。从定量的角度来看，最近的一篇论文[31]对不同的基于批准的多赢家", "投票规则进行了深人的理论和实证分析，涉及（功利主义的）社会福利和", "代表保障。[21]扩展了这项工作，以研究更一般的PB设置中的福利表示", "权衡。另一方面，传统的公理方法提供了一种定性的评估，即。e。投票规", "则是否满足属性。对于基于批准的规则，最近的工作主要集中在比例公理", "[1，41，2，9，30，42]上。。我们请读者参考[22]，对多赢家规则的属性进行广", "泛的调查", "审议，特别是在社会选择中，已经通过多种方法进行了研究。从理论的", "角度来看，各种各样的数学审议模型已被提出[14，50]。例如，[24，19]引人", "了选代的小组审议机制，以在集体决策问题中达成共识。[17]提出了基于联", "盟形成的达成共识的审议协议。最近的一项实验研究[40】表明，在特定条", "件下，审议会导致元协议和单峰偏好。[36]同时查看审议和投票，但他们的", "工作仅限于基本事实设置，并且优先于三种选择。他们没有研究审议对投票", "规则的定量和定性性质的影响", "在本文中，我们弥合了审议和投票文献之间的差距。据我们所知，我们", "是第一个通过实验研究审议对不同审议策略的投票结果的影响", "2" ]

{ "arxiv_id": "2305.08970", "language": "zh", "page_idx": 1 }

[ [ "0.2018", "0.1642", "0.3774", "0.1642", "0.3774", "0.178", "0.2018", "0.178" ], [ "0.2059", "0.1976", "0.79", "0.1976", "0.79", "0.2122", "0.2059", "0.2122" ], [ "0.2018", "0.214", "0.7915", "0.2134", "0.7915", "0.2279", "0.2018", "0.2291" ], [ "0.2115", "0.2361", "0.2239", "0.2361", "0.2239", "0.2424", "0.2115", "0.2424" ], [ "0.2181", "0.2317", "0.7915", "0.2317", "0.7915", "0.2456", "0.2181", "0.2456" ], [ "0.2059", "0.25", "0.79", "0.25", "0.79", "0.2644", "0.2059", "0.2644" ], [ "0.2051", "0.2664", "0.79", "0.2664", "0.79", "0.281", "0.2051", "0.281" ], [ "0.2075", "0.2842", "0.79", "0.2842", "0.79", "0.2974", "0.2075", "0.2974" ], [ "0.2075", "0.3005", "0.79", "0.3005", "0.79", "0.3145", "0.2075", "0.3145" ], [ "0.2051", "0.3176", "0.79", "0.3176", "0.79", "0.3308", "0.2051", "0.3308" ], [ "0.2051", "0.3352", "0.79", "0.3352", "0.79", "0.3484", "0.2051", "0.3484" ], [ "0.2035", "0.3484", "0.79", "0.3499", "0.79", "0.3674", "0.2035", "0.3662" ], [ "0.2051", "0.3674", "0.3906", "0.3687", "0.3906", "0.3833", "0.2051", "0.3818" ], [ "0.2035", "0.404", "0.259", "0.404", "0.259", "0.4185", "0.2035", "0.4185" ], [ "0.2507", "0.4048", "0.3423", "0.4048", "0.3423", "0.4192", "0.2507", "0.4192" ], [ "0.2035", "0.4312", "0.7915", "0.43", "0.7915", "0.4443", "0.2035", "0.4458" ], [ "0.2051", "0.4463", "0.7935", "0.4475", "0.7935", "0.4622", "0.2051", "0.461" ], [ "0.2059", "0.4653", "0.79", "0.4653", "0.79", "0.4792", "0.2059", "0.4792" ], [ "0.2051", "0.4817", "0.79", "0.4817", "0.79", "0.4956", "0.2051", "0.4956" ], [ "0.2059", "0.5", "0.7915", "0.5", "0.7915", "0.513", "0.2059", "0.513" ], [ "0.2051", "0.515", "0.3137", "0.5166", "0.3137", "0.531", "0.2051", "0.53" ], [ "0.2035", "0.5415", "0.7935", "0.5415", "0.7935", "0.5596", "0.2035", "0.5596" ], [ "0.2051", "0.5596", "0.7417", "0.5596", "0.7417", "0.5737", "0.2051", "0.5737" ], [ "0.3906", "0.587", "0.6055", "0.587", "0.6055", "0.605", "0.3906", "0.605" ], [ "0.767", "0.5903", "0.7915", "0.5903", "0.7915", "0.604", "0.767", "0.604" ], [ "0.5", "0.606", "0.558", "0.606", "0.558", "0.6206", "0.5", "0.6206" ], [ "0.2035", "0.63", "0.6445", "0.631", "0.6445", "0.646", "0.2035", "0.6445" ], [ "0.2035", "0.655", "0.789", "0.655", "0.789", "0.6724", "0.2035", "0.6724" ], [ "0.2035", "0.6724", "0.759", "0.6724", "0.759", "0.686", "0.2035", "0.686" ], [ "0.3652", "0.6997", "0.6284", "0.6997", "0.6284", "0.7173", "0.3652", "0.7173" ], [ "0.7656", "0.7026", "0.7935", "0.7026", "0.7935", "0.7197", "0.7656", "0.7197" ], [ "0.4714", "0.72", "0.5", "0.72", "0.5", "0.733", "0.4714", "0.733" ], [ "0.2345", "0.744", "0.79", "0.744", "0.79", "0.7583", "0.2345", "0.7583" ], [ "0.2059", "0.761", "0.79", "0.761", "0.79", "0.776", "0.2059", "0.776" ], [ "0.2059", "0.779", "0.79", "0.779", "0.79", "0.7925", "0.2059", "0.7925" ], [ "0.2051", "0.7954", "0.722", "0.797", "0.722", "0.811", "0.2051", "0.81" ], [ "0.2222", "0.8193", "0.7915", "0.821", "0.7915", "0.8354", "0.2222", "0.834" ], [ "0.2051", "0.8354", "0.4265", "0.8354", "0.4265", "0.8486", "0.2051", "0.8486" ], [ "0.492", "0.8726", "0.505", "0.8726", "0.505", "0.884", "0.492", "0.884" ] ]

[ "2Preliminaries", "令E=(C,N)为选举，其中C={c1,C2,,Cm}和 N={1,….,n}分别是", "m候选人和n选民的集合。每个选民iEN都有一个批准选票ACC", "", "其中包含其批准的候选人集。批准配置文件A=【A1,A2.An}代表", "所有选民的批准选票。对于候选ciEC，N(ci）是批准ci及其批准分数", "V(ci）=IN(ci）I的投票人集合。令Si(C）表示候选集C的所有k大小的", "子集，表示大小为k的所有可能委员会的集合。给定批准配置文件A和", "所需的委员会规模kEN，多获胜者选举的目标是选择组成获胜委员会", "WESk(C)的候选人子集。基于批准的委员会规则R(A,k）是一个社会选", "择函数，它将批准配置文件A和委员会规模k作为输入，并返回一组获胜", "的委员会。1对于任何投票规则R(A,k），我们将使用WR表示其选定的委", "员会（在打破平局之后）", "2.1", "Properties", "理想情况下，我们希望我们的投票规则表现出某些所需的属性，代表应根据", "个人选票选择获胜者的原则。在本文中，我们比较了三个维度的投票规则：", "社会福利、代表性和比例性。从直觉上讲，福利目标的重点是选择获得选民", "最大支持的候选人。代表关心多样性；仔细选择一个委员会，使获胜委员会", "中代表的选民人数最大化。如果最终委员会至少包含一名经批准的候选人，", "则代表选民。。", "Definition 1 (Utilitarian Social Welfare). For a given approval profile A and", "committee size k , the utilitarian social welfare of a committee W is:", "SW(A, W)= ui(c),", "(1)", "iENcEW", "ui(c) E R is the utility voter i derives from candidate c .", "Definition 2 (Representation Score). For a given approval profile A and com-", "mittee size k , the representation score of a committee W is defined as", "RP(A, W) =min(1,|A; nWD)", "(2)", "iEN", "可能无法同时实现社会福利和代表性的最大化，因此相称性是捕获福利", "与代表性之间折衷方案的重要第三个目标。它要求，如果一个足够大的选", "民群体集体批准一个共享的候选人集，那么这个群体必须“得到公平的代", "表”。。相称性的定义根据他们如何解释“公平代表”而有所不同。。", "1A tie-breaking method is used to pick one winning committee in cases where multiple", "winning committees are returned.", "3" ]

{ "arxiv_id": "2305.08970", "language": "zh", "page_idx": 2 }

[ [ "0.2035", "0.1666", "0.79", "0.1666", "0.79", "0.1812", "0.2035", "0.1812" ], [ "0.2051", "0.1843", "0.789", "0.1843", "0.789", "0.1989", "0.2051", "0.1989" ], [ "0.2018", "0.202", "0.789", "0.202", "0.789", "0.2153", "0.2018", "0.2153" ], [ "0.2035", "0.2184", "0.744", "0.2184", "0.744", "0.2329", "0.2035", "0.2329" ], [ "0.2018", "0.2424", "0.786", "0.2424", "0.786", "0.2568", "0.2018", "0.2568" ], [ "0.2035", "0.2583", "0.79", "0.2588", "0.79", "0.2734", "0.2035", "0.2727" ], [ "0.2018", "0.276", "0.613", "0.2747", "0.613", "0.2922", "0.2018", "0.2935" ], [ "0.2035", "0.2998", "0.7935", "0.3005", "0.7935", "0.315", "0.2035", "0.3145" ], [ "0.2035", "0.3164", "0.79", "0.318", "0.79", "0.3333", "0.2035", "0.3308" ], [ "0.2051", "0.3352", "0.79", "0.3352", "0.79", "0.3499", "0.2051", "0.3499" ], [ "0.2051", "0.351", "0.5195", "0.3516", "0.5195", "0.3662", "0.2051", "0.3655" ], [ "0.2345", "0.377", "0.79", "0.377", "0.79", "0.3901", "0.2345", "0.3901" ], [ "0.2075", "0.3945", "0.79", "0.3945", "0.79", "0.4077", "0.2075", "0.4077" ], [ "0.2035", "0.4104", "0.2345", "0.4104", "0.2345", "0.4268", "0.2035", "0.4268" ], [ "0.2018", "0.4458", "0.478", "0.4463", "0.478", "0.461", "0.2018", "0.4602" ], [ "0.2059", "0.4734", "0.789", "0.4734", "0.789", "0.4873", "0.2059", "0.4873" ], [ "0.2051", "0.4905", "0.3325", "0.4905", "0.3325", "0.504", "0.2051", "0.504" ], [ "0.2133", "0.507", "0.79", "0.507", "0.79", "0.5215", "0.2133", "0.5215" ], [ "0.2018", "0.523", "0.653", "0.523", "0.653", "0.5435", "0.2018", "0.5435" ], [ "0.2009", "0.5405", "0.7534", "0.537", "0.7534", "0.558", "0.2009", "0.561" ], [ "0.2051", "0.558", "0.79", "0.558", "0.79", "0.5728", "0.2051", "0.5728" ], [ "0.2051", "0.5757", "0.79", "0.5757", "0.79", "0.589", "0.2051", "0.589" ], [ "0.2018", "0.5923", "0.7915", "0.589", "0.7915", "0.607", "0.2018", "0.61" ], [ "0.2059", "0.61", "0.509", "0.61", "0.509", "0.6235", "0.2059", "0.6235" ], [ "0.2051", "0.627", "0.79", "0.627", "0.79", "0.6416", "0.2051", "0.6416" ], [ "0.2035", "0.643", "0.7915", "0.643", "0.7915", "0.661", "0.2035", "0.661" ], [ "0.2018", "0.659", "0.7935", "0.659", "0.7935", "0.6797", "0.2018", "0.6797" ], [ "0.2051", "0.6787", "0.789", "0.6787", "0.789", "0.6924", "0.2051", "0.6924" ], [ "0.2051", "0.6943", "0.7935", "0.696", "0.7935", "0.7104", "0.2051", "0.709" ], [ "0.2059", "0.712", "0.767", "0.712", "0.767", "0.7266", "0.2059", "0.7266" ], [ "0.2035", "0.7266", "0.7935", "0.728", "0.7935", "0.7456", "0.2035", "0.744" ], [ "0.2059", "0.747", "0.79", "0.747", "0.79", "0.761", "0.2059", "0.761" ], [ "0.2059", "0.7646", "0.79", "0.7646", "0.79", "0.778", "0.2059", "0.778" ], [ "0.2009", "0.778", "0.796", "0.779", "0.796", "0.8003", "0.2009", "0.8" ], [ "0.2059", "0.8", "0.79", "0.8", "0.79", "0.8145", "0.2059", "0.8145" ], [ "0.2059", "0.8174", "0.789", "0.8174", "0.789", "0.8306", "0.2059", "0.8306" ], [ "0.2059", "0.834", "0.4436", "0.834", "0.4436", "0.8477", "0.2059", "0.8477" ], [ "0.4902", "0.8706", "0.5073", "0.8706", "0.5073", "0.8853", "0.4902", "0.8853" ] ]

[ "Definition 3 (T-Cohesive Groups). Consider an election E = (C, N) with n", "voters and committee size k .For any integer T ≥ 1 , a group of voters N", "is T-cohesive if it contains at least Tn/k voters and collectively approves at", "least T common candidates, i.e. if INieNi A;l ≥ T and IN'≥ Tn/k .", "Definition 4 (Proportional Justified Representation (PJR)). A committee W", "of size k satisfies PJR if for each integer T [1,.., k} and every T-cohesive", "group N' N, it holds that I(UieNA,) nWI ≥T.", "Definition 5 (Extended Justified Representation (EJR)).A committee W of", "size k satisfies EJR if for each integer T E(l, .., k] , every T-cohesive group", "N' N contains at least one voter that approves at least T candidates in", "W, i.e. for some iE N',|A,nWl≥T.", "与EJR[1I不同，其中重点放在单个组成员上，PJR为组表示提供了更", "自然的要求。但是，EJR为平均选民满意度提供了更强的保证，并暗示PJR", "[41]", "2.2Multi-winner Voting Rules", "在本节中，我们将介绍一组基于批准的多获胜者投票规则，这些规则构成了", "我们分析的基础", "ApprovalVoting（AV)：给定批准配置文件A和委员会W，AV分数为", "sCa(A,W)=cew V(c)。AV 规则定义为 RAv(A,k)=", "argmaxweSk(C)SCau（A,W).此规则选择个人认可得分最高的k候选人", "批准Chamberlin-Courant（CC）：CC规则[13]，Rcc(A,k），选择最", "大化代表分数RP(A,W）的委员会。给定配置文件A，RcC(A,k）=", "argmaxwesk(C)RP(A,W).它通过最大化获胜委员会中至少有一名批准", "候选人的选民数量来最大化选民覆盖率", "比例批准投票（PAV：[46]对于配置文件A和委员会W，PAV分数定义", "为sCpav(A,W)=ZieNh(IWnAil)，其中h(t)=Zt=11/i.PAV规则定义", "为RpAv(A,k）=argmaxweSs(C)SCpau(A,W）。基于收益递减的思想，根", "据谐波函数h(t)，选民在当选委员会W中获得批准的候选人的效用会降", "低。它是AV规则的一种变体，可确保比例表示，因为它保证了EJR[1]。", "当委员会规模k=1时，PAV与AV相同，但计算PAV是NP-hard[3]", "等份额方法（MES)：RMEs（A,k），在文献中也称为Rule-X[37，38]】，是一", "个迭代过程，它使用预算的思想来保证比例。每个选民的预算为k/n，每", "个候选人的单位成本为。在第t轮中，如果q负担得起，则将候选c添加到", "W中，即。e。对于某些q≥0,ZieN(c) min(q,b;(t)≥1 ，其中 b;(t)是第 t", "轮投票人讠的预算。如果成功添加了候选人，则每个支持选民的预算将相应", "减少。此过程一直持续到将k候选人添加到委员会或失败（然后使用另一", "个投票规则来选择其余候选人）", "4" ]

{ "arxiv_id": "2305.08970", "language": "zh", "page_idx": 3 }

[ [ "0.2336", "0.1674", "0.79", "0.1674", "0.79", "0.1812", "0.2336", "0.1812" ], [ "0.2059", "0.1843", "0.7915", "0.1843", "0.7915", "0.1976", "0.2059", "0.1976" ], [ "0.2059", "0.2008", "0.789", "0.1995", "0.789", "0.214", "0.2059", "0.2153" ], [ "0.2059", "0.2184", "0.79", "0.2184", "0.79", "0.2317", "0.2059", "0.2317" ], [ "0.2059", "0.2349", "0.79", "0.2349", "0.79", "0.2488", "0.2059", "0.2488" ], [ "0.2051", "0.2507", "0.562", "0.252", "0.562", "0.2664", "0.2051", "0.2651" ], [ "0.2035", "0.2886", "0.259", "0.2886", "0.259", "0.303", "0.2035", "0.303" ], [ "0.25", "0.2874", "0.343", "0.289", "0.343", "0.3037", "0.25", "0.3018" ], [ "0.2059", "0.315", "0.789", "0.315", "0.789", "0.3289", "0.2059", "0.3289" ], [ "0.2059", "0.3333", "0.4902", "0.3333", "0.4902", "0.3467", "0.2059", "0.3467" ], [ "0.2035", "0.3484", "0.7876", "0.3484", "0.7876", "0.3623", "0.2035", "0.3623" ], [ "0.2059", "0.3655", "0.4436", "0.3655", "0.4436", "0.379", "0.2059", "0.379" ], [ "0.3635", "0.3984", "0.452", "0.406", "0.4495", "0.4248", "0.361", "0.4167" ], [ "0.492", "0.3945", "0.585", "0.3945", "0.585", "0.4092", "0.492", "0.4092" ], [ "0.4355", "0.4072", "0.6323", "0.4104", "0.6323", "0.435", "0.4355", "0.432" ], [ "0.7656", "0.4028", "0.7935", "0.4028", "0.7935", "0.4192", "0.7656", "0.4192" ], [ "0.2059", "0.4443", "0.789", "0.4443", "0.789", "0.4583", "0.2059", "0.4583" ], [ "0.2059", "0.461", "0.6914", "0.461", "0.6914", "0.476", "0.2059", "0.476" ], [ "0.492", "0.4905", "0.58", "0.4905", "0.58", "0.505", "0.492", "0.505" ], [ "0.3652", "0.5", "0.4387", "0.5", "0.4387", "0.5146", "0.3652", "0.5146" ], [ "0.7656", "0.4988", "0.794", "0.4988", "0.794", "0.515", "0.7656", "0.515" ], [ "0.4395", "0.505", "0.6313", "0.507", "0.6313", "0.529", "0.4395", "0.5264" ], [ "0.429", "0.511", "0.4395", "0.511", "0.4395", "0.5146", "0.429", "0.5146" ], [ "0.2059", "0.5405", "0.357", "0.5405", "0.357", "0.5547", "0.2059", "0.5547" ], [ "0.2051", "0.5566", "0.79", "0.5566", "0.79", "0.5713", "0.2051", "0.5713" ], [ "0.2059", "0.5747", "0.4272", "0.5747", "0.4272", "0.5884", "0.2059", "0.5884" ], [ "0.38", "0.604", "0.613", "0.605", "0.613", "0.619", "0.38", "0.6187" ], [ "0.7686", "0.606", "0.7915", "0.606", "0.7915", "0.6206", "0.7686", "0.6206" ], [ "0.2336", "0.636", "0.7915", "0.636", "0.7915", "0.6494", "0.2336", "0.6494" ], [ "0.2035", "0.653", "0.6567", "0.653", "0.6567", "0.6665", "0.2035", "0.6665" ], [ "0.2035", "0.6943", "0.451", "0.6943", "0.451", "0.709", "0.2035", "0.709" ], [ "0.2059", "0.728", "0.79", "0.728", "0.79", "0.741", "0.2059", "0.741" ], [ "0.2051", "0.744", "0.79", "0.744", "0.79", "0.7583", "0.2051", "0.7583" ], [ "0.2059", "0.762", "0.79", "0.762", "0.79", "0.776", "0.2059", "0.776" ], [ "0.2059", "0.779", "0.79", "0.779", "0.79", "0.7925", "0.2059", "0.7925" ], [ "0.2051", "0.7944", "0.4192", "0.7954", "0.4192", "0.81", "0.2051", "0.8096" ], [ "0.492", "0.8726", "0.5034", "0.8726", "0.5034", "0.884", "0.492", "0.884" ] ]

[ "我们选择研究这些规则，因为它们具有广泛的属性，可以跨多个轴进行", "比较。首先，已知在选民的效用函数E[31，301的某些条件下，AV可以最大", "化社会福利，但是，不能保证AV满足比例性（EJR标准)E[1]。另一方", "面，CC最大化了代表性，但对其福利属性的了解较少。PAV和MES都保", "证了EJR，并在代表和社会福利之间保持平衡。因此，这个多赢家投票规则", "的集合涵盖了我们有兴趣更好地理解的属性集", "2.3", "Objectives", "我们根据不同的标准目标[31]比较上述投票规则。特别是，我们考虑三个", "维度的目标：福利、代表性和相称性", "功利主义比率：此比率将WR=R(Ak）实现的（功利主义）社会福利与可", "实现的最大社会福利进行比较：", "UR(R)= ", "SW(A, WR)", "maawesk(C) SW(A, W)", "(3)", "代表比率：该比率通过将WR获得的代表分数与所有k规模的委员会中的", "最佳代表分数进行比较来衡量委员会WR=R(A,k）的多样性：", "RP(A, WR)", "RR(R) =", "(4)", "macweS(C)RP(A, W)", "", "注意 RR(Rcc)= 1", "Utility-RepresentationAggregateScore：此分数反映了投票规则R(A,k）平", "衡社会福利和代表性的程度：", "URagg(R) =UR(R) *RR(R)", "(5)", "最后，我们有兴趣通过实验验证生成的配置文件实例satisfyEJRorPJR", "。为此，我们计算满足这两个属性的配置文件实例的数量。", "3Deliberation Models", "我们描述了如何对审议过程进行建模。我们首先定义我们的底层代理群体", "以及我们如何对他们的初始偏好进行建模。然后，我们讨论审议过程，通过", "该过程，代理商可以交换信息并更新其偏好。最后，我们观察到，审议通常", "不是在整个人口水平上进行的，而是在较小的子组中进行的。我们讨论了创", "建这些审议子组的不同方式", "5" ]

{ "arxiv_id": "2305.08970", "language": "zh", "page_idx": 4 }

[ [ "0.1765", "0.0341", "0.5063", "0.0341", "0.5063", "0.0486", "0.1765", "0.0486" ], [ "0.1765", "0.1042", "0.6855", "0.1048", "0.6855", "0.1225", "0.1765", "0.1218" ], [ "0.1847", "0.1515", "0.7236", "0.1509", "0.7236", "0.1654", "0.1847", "0.1666" ], [ "0.1871", "0.1674", "0.5312", "0.1674", "0.5312", "0.1812", "0.1871", "0.1812" ], [ "0.1887", "0.1812", "0.535", "0.1812", "0.535", "0.1945", "0.1887", "0.1945" ], [ "0.4631", "0.2329", "0.5376", "0.2329", "0.5376", "0.2474", "0.4631", "0.2474" ], [ "0.2345", "0.2676", "0.761", "0.2676", "0.761", "0.281", "0.2345", "0.281" ], [ "0.2336", "0.2822", "0.761", "0.2822", "0.761", "0.2954", "0.2336", "0.2954" ], [ "0.2345", "0.2966", "0.7656", "0.2966", "0.7656", "0.31", "0.2345", "0.31" ], [ "0.2361", "0.3113", "0.4436", "0.3113", "0.4436", "0.3213", "0.2361", "0.3213" ], [ "0.452", "0.31", "0.761", "0.31", "0.761", "0.324", "0.452", "0.324" ], [ "0.2345", "0.324", "0.7646", "0.324", "0.7646", "0.3372", "0.2345", "0.3372" ], [ "0.2361", "0.339", "0.763", "0.339", "0.763", "0.3499", "0.2361", "0.3499" ], [ "0.2361", "0.3516", "0.7646", "0.3516", "0.7646", "0.3655", "0.2361", "0.3655" ], [ "0.2345", "0.3643", "0.761", "0.3643", "0.761", "0.379", "0.2345", "0.379" ], [ "0.2345", "0.377", "0.7646", "0.379", "0.7646", "0.3933", "0.2345", "0.3914" ], [ "0.2361", "0.3945", "0.7646", "0.3945", "0.7646", "0.4048", "0.2361", "0.4048" ], [ "0.2361", "0.4077", "0.601", "0.4077", "0.601", "0.4185", "0.2361", "0.4185" ], [ "0.1724", "0.455", "0.3162", "0.4558", "0.3162", "0.4702", "0.1724", "0.4697" ], [ "0.1772", "0.4912", "0.8203", "0.4912", "0.8203", "0.502", "0.1772", "0.502" ], [ "0.1765", "0.5024", "0.822", "0.5024", "0.822", "0.5166", "0.1765", "0.5166" ], [ "0.1772", "0.5176", "0.824", "0.5176", "0.824", "0.531", "0.1772", "0.531" ], [ "0.1765", "0.53", "0.822", "0.53", "0.822", "0.5435", "0.1765", "0.5435" ], [ "0.1748", "0.5454", "0.451", "0.5454", "0.451", "0.556", "0.1748", "0.556" ], [ "0.1765", "0.5664", "0.822", "0.5664", "0.822", "0.58", "0.1765", "0.58" ], [ "0.1765", "0.58", "0.8228", "0.58", "0.8228", "0.5947", "0.1765", "0.5947" ], [ "0.1765", "0.5947", "0.8228", "0.5947", "0.8228", "0.608", "0.1765", "0.608" ], [ "0.179", "0.61", "0.822", "0.61", "0.822", "0.6206", "0.179", "0.6206" ], [ "0.179", "0.622", "0.8228", "0.622", "0.8228", "0.6353", "0.179", "0.6353" ], [ "0.1772", "0.6353", "0.822", "0.6353", "0.822", "0.6484", "0.1772", "0.6484" ], [ "0.1772", "0.651", "0.8203", "0.651", "0.8203", "0.661", "0.1772", "0.661" ], [ "0.1765", "0.6636", "0.822", "0.6636", "0.822", "0.677", "0.1765", "0.677" ], [ "0.1748", "0.6753", "0.822", "0.6753", "0.822", "0.6895", "0.1748", "0.6895" ], [ "0.179", "0.6924", "0.8203", "0.6924", "0.8203", "0.7026", "0.179", "0.7026" ], [ "0.1772", "0.705", "0.824", "0.705", "0.824", "0.7183", "0.1772", "0.7183" ], [ "0.1772", "0.7183", "0.822", "0.7183", "0.822", "0.732", "0.1772", "0.732" ], [ "0.1765", "0.733", "0.822", "0.733", "0.822", "0.747", "0.1765", "0.747" ], [ "0.1772", "0.7456", "0.822", "0.7456", "0.822", "0.759", "0.1772", "0.759" ], [ "0.1748", "0.7603", "0.746", "0.7603", "0.746", "0.7734", "0.1748", "0.7734" ], [ "0.1765", "0.781", "0.822", "0.782", "0.822", "0.7954", "0.1765", "0.7944" ], [ "0.1772", "0.797", "0.8203", "0.797", "0.8203", "0.8066", "0.1772", "0.8066" ], [ "0.1912", "0.8096", "0.822", "0.8096", "0.822", "0.8228", "0.1912", "0.8228" ], [ "0.1765", "0.8228", "0.822", "0.8228", "0.822", "0.836", "0.1765", "0.836" ], [ "0.1772", "0.8384", "0.8203", "0.8384", "0.8203", "0.8486", "0.1772", "0.8486" ], [ "0.1765", "0.851", "0.822", "0.851", "0.822", "0.8643", "0.1765", "0.8643" ], [ "0.179", "0.8643", "0.822", "0.8643", "0.822", "0.8774", "0.179", "0.8774" ], [ "0.1748", "0.877", "0.2263", "0.879", "0.2255", "0.893", "0.1732", "0.8916" ], [ "0.1929", "0.9077", "0.7695", "0.91", "0.7695", "0.928", "0.1929", "0.9253" ], [ "0.4885", "0.954", "0.5", "0.9478", "0.5073", "0.955", "0.4968", "0.962" ] ]

[ "Published as a conference paper at ICLR 2024", "NfgTransformer：正态游戏的等价表示学习", "Siqi Liut., Luke Marrist., Georgios Piliourast, Ian Gempt, Nicolas Heess", "Google DeepMind,UniversityCollege London", "f liusiqi,marris,gpil.imgemp,heess @google.com", "Abstract", "Normal-form games (NFGs) are the fundamental model of strategic in", "teraction : We study their representation using neural networks. We de", "scribe the inherent equivariance of NFGsany permutation of strategies", "describesanequivalentgame", " as well as the challenges this poses for rep", "resentation learning.Wethen propose the NfgTransformerarchitecture", "that leverages this equivariance,leading to state-of-the-artperformance in a", "range of game-theoretic tasks including equilibrium-solving, deviation gain", "estimation and ranking, with a common approach to NFG representation", "We show that the resulting model is interpretable and versatile, paving", "thewaytowardsdeeplearningsystemscapableofgame-theoreticreasoning", "when interacting withhumans andwitheach other", "1Introduction", "在学习的嵌人空间中表示数据或表示学习是深度学习的永恒思想之一。基础表示学习架构", "(LeCun et al.,1995;Hochreiter & Schmidhuber,1997;Vaswani et al.,2017)提供了性能和通", "用性，汇集了过去致力于针对狭窄领域的专家系统的研究人员。考虑卷积神经网络（CNN）：", "通过利用图像固有的平移不变性，CNN取代了支持图像分类，分割和绘画等多样化任务的", "特征描述符，导致整个领域的范式转变", "抵制这一趋势的一个研究领域是博奔论。在这里，我们看到了与经典计算机视觉研究的", "两个惊人相似之处。首先，尽管共享普通形式游戏（NFG）的共同语言，但诸如均衡求", "解（Vlatakis-Gkaragkounis etal.，2020），排名（Balduzziet al.，2018），社会选择理论", "（Anshelevichetal.，2021）和人口学寸（Lanctotetal，2017）之类的活跃研究主题都集中在", "专门的解决方案上。。其次，针对特定任务的解决方案受到基本限制。例如，流行的Elo排", "名算法（Elo，2008）为从玩家之间的NFG得出的每个玩家分配标量等级。尽管Elo评分旨", "在预测比赛结果，但它们是传递性游戏之外的较差预测指标-E1o得分过于严格，无法反映循", "环游戏的动态。（Bertrandetal，2023）之后对Elo进行了改进，但所有改进都依赖于工程功", "能描述符，而不是学习。在平衡求解中，计算精确的纳什均衡在两人零和博奔（Daskalakis", "etal.2009）之外是棘手的，但近似求解器是不可微的，需要不确定性的时间来收敛，难以", "并行化，并且可能会失败。这些基本限制也有间接的后果。表格多智能体强化学习（RL）", "(Littmanetal.,2001；Hu&Wellman，2003;Greenwaldetal.,2003)中的整个工作流程都依", "赖于均衡求解作为其学习规则的一部分-NFG由每个状态下的智能体Q表构造而成，其均", "衡为后续策略更新提供信息。不幸的是，在deepRL的背景下恢复这些想法是具有挑战性", "的，如果不是不可能的话，因为它需要平衡求解作为每个梯度更新之间的子程序", "事实上，我们并不是第一个认识到这些限制的人。最近的几部作品在narrowdomains", "(Marrisetal..2022:Duanetal..2023:Vadori&Savani,2023）中结合T表示学习implicitly", "我们解决这些限制explicitly和ingenerality。我们的目标是开发可用于广泛博奔论应", "用的NFG的原则性通用表示。我们问1）如果有的话，我们可以利用哪个属性来进行NFG", "的有效表示学习而不失一般性；2我们可以使用一种通用的方法来表示NFG：在一系列博", "奔论任务中期望性能吗。我们肯定地回答了这两个问题，并提出了NfgTransformer，这是", "一种通用的表示学习架构，在平衡求解，偏差增益估计和排名等各种任务中都具有最先进", "的性能", "IThe model is open-sourced at https://github.com/google-deepmind/nfg_transformer", "1" ]

{ "arxiv_id": "2402.08393", "language": "zh", "page_idx": 0 }

[ [ "0.1765", "0.0341", "0.5063", "0.0341", "0.5063", "0.0486", "0.1765", "0.0486" ], [ "0.1772", "0.1061", "0.8228", "0.1061", "0.8228", "0.1193", "0.1772", "0.1193" ], [ "0.1772", "0.1193", "0.8228", "0.1193", "0.8228", "0.1332", "0.1772", "0.1332" ], [ "0.1772", "0.1332", "0.8228", "0.1332", "0.8228", "0.1465", "0.1772", "0.1465" ], [ "0.1748", "0.1465", "0.822", "0.1459", "0.822", "0.1604", "0.1748", "0.161" ], [ "0.1748", "0.159", "0.822", "0.1604", "0.822", "0.1736", "0.1748", "0.1724" ], [ "0.1772", "0.1749", "0.822", "0.1749", "0.822", "0.1881", "0.1772", "0.1881" ], [ "0.1765", "0.1875", "0.8228", "0.1881", "0.8228", "0.2026", "0.1765", "0.202" ], [ "0.1765", "0.2039", "0.7646", "0.2039", "0.7646", "0.214", "0.1765", "0.214" ], [ "0.1765", "0.2235", "0.8228", "0.2235", "0.8228", "0.2374", "0.1765", "0.2374" ], [ "0.1772", "0.2393", "0.8203", "0.2393", "0.8203", "0.25", "0.1772", "0.25" ], [ "0.1765", "0.252", "0.8228", "0.252", "0.8228", "0.2651", "0.1765", "0.2651" ], [ "0.1765", "0.2644", "0.8203", "0.2644", "0.8203", "0.2778", "0.1765", "0.2778" ], [ "0.1772", "0.281", "0.8203", "0.281", "0.8203", "0.2917", "0.1772", "0.2917" ], [ "0.1772", "0.2935", "0.8228", "0.2935", "0.8228", "0.307", "0.1772", "0.307" ], [ "0.1765", "0.3062", "0.761", "0.3062", "0.761", "0.3193", "0.1765", "0.3193" ], [ "0.1765", "0.3276", "0.822", "0.3276", "0.822", "0.3416", "0.1765", "0.3416" ], [ "0.1765", "0.3403", "0.822", "0.3416", "0.822", "0.356", "0.1765", "0.3547" ], [ "0.1765", "0.3547", "0.822", "0.3547", "0.822", "0.3687", "0.1765", "0.3687" ], [ "0.1765", "0.3694", "0.822", "0.3694", "0.822", "0.3833", "0.1765", "0.3833" ], [ "0.1748", "0.3818", "0.8228", "0.3833", "0.8228", "0.3977", "0.1748", "0.3965" ], [ "0.1765", "0.3984", "0.2566", "0.3984", "0.2566", "0.4104", "0.1765", "0.4104" ], [ "0.1732", "0.432", "0.312", "0.432", "0.312", "0.4463", "0.1732", "0.4463" ], [ "0.1814", "0.4622", "0.8228", "0.4622", "0.8228", "0.476", "0.1814", "0.476" ], [ "0.1748", "0.4766", "0.822", "0.476", "0.822", "0.4905", "0.1748", "0.4912" ], [ "0.1748", "0.4905", "0.546", "0.4937", "0.546", "0.507", "0.1748", "0.504" ], [ "0.54", "0.4976", "0.5933", "0.4976", "0.5933", "0.505", "0.54", "0.505" ], [ "0.597", "0.4924", "0.822", "0.4924", "0.822", "0.506", "0.597", "0.506" ], [ "0.1732", "0.504", "0.824", "0.502", "0.824", "0.5215", "0.1732", "0.5234" ], [ "0.1748", "0.518", "0.8228", "0.5195", "0.8228", "0.534", "0.1748", "0.5327" ], [ "0.1765", "0.534", "0.824", "0.534", "0.824", "0.549", "0.1765", "0.549" ], [ "0.179", "0.561", "0.8228", "0.5645", "0.8228", "0.579", "0.179", "0.5757" ], [ "0.1765", "0.58", "0.3406", "0.58", "0.3406", "0.5933", "0.1765", "0.5933" ], [ "0.3586", "0.603", "0.6406", "0.603", "0.6406", "0.617", "0.3586", "0.617" ], [ "0.8022", "0.6016", "0.824", "0.6016", "0.824", "0.616", "0.8022", "0.616" ], [ "0.4192", "0.6143", "0.4714", "0.6143", "0.4714", "0.622", "0.4192", "0.622" ], [ "0.1772", "0.634", "0.8228", "0.634", "0.8228", "0.6484", "0.1772", "0.6484" ], [ "0.1765", "0.651", "0.8228", "0.651", "0.8228", "0.6655", "0.1765", "0.6655" ], [ "0.1765", "0.6636", "0.822", "0.6636", "0.822", "0.678", "0.1765", "0.678" ], [ "0.1748", "0.677", "0.822", "0.678", "0.822", "0.6924", "0.1748", "0.6914" ], [ "0.1765", "0.6924", "0.4902", "0.6924", "0.4902", "0.7056", "0.1765", "0.7056" ], [ "0.1805", "0.7236", "0.822", "0.7227", "0.822", "0.7373", "0.1805", "0.738" ], [ "0.1765", "0.7407", "0.8228", "0.7407", "0.8228", "0.754", "0.1765", "0.754" ], [ "0.1765", "0.755", "0.4265", "0.755", "0.4265", "0.768", "0.1765", "0.768" ], [ "0.3057", "0.7764", "0.693", "0.7764", "0.693", "0.7944", "0.3057", "0.7944" ], [ "0.1748", "0.81", "0.2091", "0.81", "0.2091", "0.8228", "0.1748", "0.8228" ], [ "0.2893", "0.832", "0.71", "0.832", "0.71", "0.8467", "0.2893", "0.8467" ], [ "0.8022", "0.832", "0.8228", "0.832", "0.8228", "0.8467", "0.8022", "0.8467" ], [ "0.179", "0.857", "0.822", "0.857", "0.822", "0.8706", "0.179", "0.8706" ], [ "0.1765", "0.8735", "0.4795", "0.8726", "0.4795", "0.887", "0.1765", "0.888" ], [ "0.1772", "0.898", "0.824", "0.898", "0.824", "0.911", "0.1772", "0.911" ], [ "0.1765", "0.912", "0.8013", "0.912", "0.8013", "0.9253", "0.1765", "0.9253" ], [ "0.4902", "0.9487", "0.5073", "0.9487", "0.5073", "0.962", "0.4902", "0.962" ] ]

[ "Published as a conference paper at ICLR 2024", "在其最基本的形式中，玩家之间的战略互动被表述为NFG，其中玩家同时选择行动并在联", "合行动下获得收益。因此，战略互动以收益张量的形式呈现，每个参与者的价值都在每个联", "合行动下列出。这种战略互动的表格视图对表示学习提出了自已的挑战。与诸如图像和文本", "之类的模态不同，它们的空间结构可以用于有效的表示学习（LeCunetal.，1995；Hochreiter", "&Schmidhuber，1997），动作在收益张量中的位置并不重要：置换任何NFG的收益矩阵都", "会产生等效的游戏-等效称为强同构游戏（McKinsey，1951；Gabarroetal.，2011）。NFG的", "这种固有的等方差启发了先前的工作，以在神经网络体系结构中组成顺序不变的池函数以", "提高效率，尽管以牺表示（Fengetal.2021：Marrisetal.，2022）的通用性为代价", "我们的目标是利用NFG的这种固有的等方差，同时保留学习表示的全部通用性。这意味着", "我们依次讨论的儿个需求。首先，表示需要是通用的，充许在个人动作，联合行动，每个玩", "家或整个游戏的级别进行查询。其次，它必须是等变的：对于任何每个动作的查询，如果两", "个动作的位置在收益张量中交换，则应交换两个动作的输出。第三，嵌入函数不应假设观", "察到所有联合行动的结果-表示应以原则的方式容纳不完整的NFG。第四，该功能应适用于", "不同大小的游戏。这意味着网络参数的数量应该独立于游戏（Hartfordetal.，2016）的大小。", "最后，如果网络架构是可解释的，允许在网络的不同阶段进行检查，这将是可取的", "在本文的其余部分中，我们将展示我们提出的编码器架构NfgTransformer如何同时满足所", "有这些需求。我们方法背后的关键思想是将表示NFG的嵌人函数类视为动作嵌入，从而逆", "转从动作到收益的生成过程。动作嵌人可以适当地组成，以回答不同粒度的问题，并以直接", "的方式允许等方差-收益张量中动作或参与者的置换应反映为动作嵌人中的置换。我们认为", "NfgTransformer是NFG通用等变表示学习的竞争候选者，弥合了深度学习和博奔论推理", "之间的差距", "2Background", "Normal-formGamesNFG是基本的游戏形式，其中每个玩家p同时播放其T动作", "apE[ap，..,a)=Ap之一，并根据所有N玩家的联合动作a=（a1,...，an）EA获得收", "益Gp:A→R。让a=(ap,a-p)与a-p=(,ap", "-1,ap+", ".）EA-p除p之外的所有玩", "家的动作。令a(a）=o(ap,a-p）表示玩家玩联合动作的概率α和。在联合动作空间上的概", "率分布A。纯策略是一种行动分布，当混合策略可以是随机的时，它是确定性的。α下播放", "器p的值为Ea~α[Gp(αp,a-p)]。我们将根据上面的动作和玩家顺序列出的收益张量称为G", "NashEquilibrium（NE）在分解为边际α=，o，的外积的混合联合策略下，玩家p的", "单边偏离激励被定义为", "dp(o) = max E [Gp(ap,a-p) -Gp(a)].", "(1)", "aEApar", "可分解混合联合策略α=，，是E-NE当且仅当XMATHX_。我们将此数量称为NE", "Gap，因为它直观地测量了从α到游戏NE的距离。保证有限博奔（Nash，1951）存在混", "合策略NE，但确切地计算超出两人零和的正常形式NE是PPAD完全（Chenetal.，2009：", "Daskalakis etal.,2009）。如果α与(a)=1是确定性的，则s(α）或等效的(a)定义了", "联合纯策略a的最大偏差增益。当(αa)=0", "PermutationEquivariance考虑具有Φ=（Tp，pE[NI)，w），Tp：a，→aa玩家动作置", "换和w：p→pa玩家置换的NFGs（Gabarroetal.,2011）的强同构Φ：G→G'。因此，", "转换后的游戏G'=Φ(G)的元素为", "Gw(p) (Tw(1)(aw(1),..., Tw(N)(aw(N))) = Gp(a1,..,aN).", "如果", "f(Φ(G)) = (Ai,..., AN) with Aw(p) = (Tp(ap),..., Tp(ap))", "(2)", "，则将带有Ap=（ab，..,a）的编码器f：G→（A1，.,An）称为equivariant，其中", "A，=（ap，..,a）是播放器p的动作嵌人", "在这里，我们稍微滥用Tp的表示法来操作动作嵌人。直观上，置换等方差意味着Φ和f交", "换或f(o(G)）=Φ(f(G)）。我们采用在玩家动作置换Tp，Vp之后应用玩家置换的约定", "2" ]

{ "arxiv_id": "2402.08393", "language": "zh", "page_idx": 1 }

[ [ "0.1765", "0.0341", "0.5063", "0.0341", "0.5063", "0.0486", "0.1765", "0.0486" ], [ "0.1814", "0.1061", "0.8228", "0.1061", "0.8228", "0.1193", "0.1814", "0.1193" ], [ "0.1765", "0.1193", "0.822", "0.1193", "0.822", "0.1332", "0.1765", "0.1332" ], [ "0.1765", "0.1332", "0.6724", "0.1332", "0.6724", "0.1465", "0.1765", "0.1465" ], [ "0.3513", "0.1749", "0.569", "0.1749", "0.569", "0.1895", "0.3513", "0.1895" ], [ "0.572", "0.1692", "0.6235", "0.1692", "0.6235", "0.1799", "0.572", "0.1799" ], [ "0.572", "0.1812", "0.6284", "0.1863", "0.626", "0.1995", "0.569", "0.1945" ], [ "0.8022", "0.1768", "0.824", "0.1768", "0.824", "0.1901", "0.8022", "0.1901" ], [ "0.1748", "0.209", "0.8228", "0.209", "0.8228", "0.2267", "0.1748", "0.2267" ], [ "0.1765", "0.2267", "0.822", "0.2267", "0.822", "0.2405", "0.1765", "0.2405" ], [ "0.1748", "0.2405", "0.822", "0.2405", "0.822", "0.2583", "0.1748", "0.2583" ], [ "0.1732", "0.2556", "0.822", "0.2551", "0.822", "0.2727", "0.1732", "0.2734" ], [ "0.1765", "0.2727", "0.822", "0.2727", "0.822", "0.286", "0.1765", "0.286" ], [ "0.1748", "0.2854", "0.8203", "0.2854", "0.8203", "0.2986", "0.1748", "0.2986" ], [ "0.1748", "0.3005", "0.8228", "0.3005", "0.8228", "0.3145", "0.1748", "0.3145" ], [ "0.1765", "0.3176", "0.6626", "0.3176", "0.6626", "0.3276", "0.1765", "0.3276" ], [ "0.1732", "0.351", "0.4958", "0.351", "0.4958", "0.3643", "0.1732", "0.3643" ], [ "0.1772", "0.3806", "0.822", "0.3806", "0.822", "0.3945", "0.1772", "0.3945" ], [ "0.1748", "0.3945", "0.8203", "0.3945", "0.8203", "0.4077", "0.1748", "0.4077" ], [ "0.1765", "0.4104", "0.6406", "0.4104", "0.6406", "0.4236", "0.1765", "0.4236" ], [ "0.1765", "0.4287", "0.8228", "0.4287", "0.8228", "0.443", "0.1765", "0.443" ], [ "0.1748", "0.443", "0.808", "0.443", "0.808", "0.461", "0.1748", "0.461" ], [ "0.8022", "0.4495", "0.812", "0.4495", "0.812", "0.457", "0.8022", "0.457" ], [ "0.1724", "0.4634", "0.824", "0.464", "0.824", "0.4817", "0.1724", "0.481" ], [ "0.1748", "0.4797", "0.824", "0.4797", "0.824", "0.4944", "0.1748", "0.4944" ], [ "0.1748", "0.4956", "0.2656", "0.4944", "0.2656", "0.506", "0.1748", "0.507" ], [ "0.1732", "0.518", "0.824", "0.5195", "0.824", "0.534", "0.1732", "0.5327" ], [ "0.1748", "0.534", "0.822", "0.534", "0.822", "0.5474", "0.1748", "0.5474" ], [ "0.1732", "0.5684", "0.343", "0.5684", "0.343", "0.582", "0.1732", "0.582" ], [ "0.1765", "0.5996", "0.8228", "0.5996", "0.8228", "0.613", "0.1765", "0.613" ], [ "0.1748", "0.6123", "0.822", "0.613", "0.822", "0.627", "0.1748", "0.6255" ], [ "0.1765", "0.6274", "0.822", "0.6274", "0.822", "0.6416", "0.1765", "0.6416" ], [ "0.1765", "0.6416", "0.822", "0.6416", "0.822", "0.655", "0.1765", "0.655" ], [ "0.1765", "0.655", "0.5864", "0.655", "0.5864", "0.6685", "0.1765", "0.6685" ], [ "0.1748", "0.686", "0.4739", "0.686", "0.4739", "0.6997", "0.1748", "0.6997" ], [ "0.1765", "0.712", "0.8228", "0.712", "0.8228", "0.726", "0.1765", "0.726" ], [ "0.1765", "0.726", "0.8228", "0.726", "0.8228", "0.7407", "0.1765", "0.7407" ], [ "0.1765", "0.7437", "0.8228", "0.7437", "0.8228", "0.757", "0.1765", "0.757" ], [ "0.1748", "0.757", "0.8228", "0.7583", "0.8228", "0.773", "0.1748", "0.7715" ], [ "0.1748", "0.773", "0.357", "0.773", "0.357", "0.786", "0.1748", "0.786" ], [ "0.1805", "0.803", "0.8203", "0.803", "0.8203", "0.8164", "0.1805", "0.8164" ], [ "0.1912", "0.8174", "0.822", "0.8174", "0.822", "0.832", "0.1912", "0.832" ], [ "0.1814", "0.8237", "0.1953", "0.8237", "0.1953", "0.8276", "0.1814", "0.8276" ], [ "0.1748", "0.832", "0.822", "0.834", "0.822", "0.851", "0.1748", "0.8486" ], [ "0.1748", "0.8467", "0.4558", "0.8486", "0.4558", "0.863", "0.1748", "0.8613" ], [ "0.179", "0.879", "0.822", "0.879", "0.822", "0.8926", "0.179", "0.8926" ], [ "0.1765", "0.8945", "0.8228", "0.8945", "0.8228", "0.909", "0.1765", "0.909" ], [ "0.1748", "0.911", "0.5767", "0.91", "0.5767", "0.925", "0.1748", "0.9253" ], [ "0.492", "0.9487", "0.509", "0.9487", "0.509", "0.962", "0.492", "0.962" ] ]

[ "Published as a conference paper at ICLR 2024", "Multi-HeadAttention我们描述了自我关注和交叉关注的QKV机制，这些机制由于其", "通用性和扩展（Vaswanietal.,2017;Dosovitskiyetal.,2021；Jaegleetal.，2021）的潜力而", "变得无处不在。两种操作都扩展了基本的QKV注意机制，如下所示：", "Attention(Q, K, V) = softmax", "?QK1", "dk", "(3)", "对于QERnaxdk、KRnkxdk和VERnkxd，ng查询和nk键值对。从概念上讲，注意", "力机制为每个ng查询向量输出nk值的加权和，其权重由键和查询向量之间的成对点积确", "定。输出的形状为Rnaxd。输人QKV是来自全连接网络本身的输出，Q是agERng×daa", "和K,V的函数，从同一输人akeRnk×daw投影。注意操作是:1）相对于ak的order", "invariant；和2）相对于ag的equivariant。这些是我们在NfgTransformer的设计中利用", "的关键属性，以实现其置换等方差属性。当。与相同时，我们将关注层称为自我关注", "如果不是，则称为交叉关注。实际上，每个关注层都可能有H个关注头并行地执行等式F", "3的关注操作。这使得关注层能够在一个前向传递中聚合多个信息流", "3EquivariantGameRepresentation", "虽然我们非正式地激发了对等变嵌入函数的需求，但我们正式陈述了等变嵌入函数的两个", "实际含义，这些含义来自一般定理，在给定等变嵌人函数的条件下，两个动作必须具有相同", "的嵌入。为简明起见，我们将所有正式声明和证明推迟到附录A", "Proposition 3.1 (Repeated Actions). If G(ap, a-p) = G(ap, a-p), Va-p and f is deterministic", "and equivariant with f(G) = (..., (...,ap,..., ap,...),...) then it follows that a, = ap", "", "Proposition 3.2 (Player Symmetry). If player p and q are symmetric, f is deterministic", "and equivariant with f(G) = (..., Ap,..., Aq,...) , then Ap and Ag are identical up to", "permutation", "Proposition3.1通过构造保证重复操作在任何下游应用程序中都得到相同的处理。命题3.2", "保证玩家对称性反映在动作嵌人空间中，我们在第5.3节中根据经验展示了NfgTransformer", "4NfgTransformer", "我们现在描述NfgTransformer，这是一个编码器网络，它通过一系列KNfgTransformer块", "（图E1（顶部））将收益张量G分解为动作嵌，每个块均由一系列自注意和交叉注意操作", "组成（图E1（底部））。然后，我们在图2中展示了不同解码粒度下几个博奔论任务的解码", "器和损失函数的具体示例，展示了我们方法的通用性。最后，我们讨论了NfgTransformer", "如何利用注意力操作的灵活性自然地合并不完全收益张量", "4.1Initialisation & Iterative refinement", "置换等方差意味着动作嵌人必须与玩家和动作在收益张量中的排序方式无关。这表明玩家", "和动作之间的动作嵌人必须以相同的方式初始化。我们用aERD零初始化所有动作嵌入", "A={a,=O|VtE[T],VpE[M]】。初始化后，在给定当前动作嵌人和收益张量G的情况", "下，通过一系列K块选代地细化动作嵌人。每个块通过我们依次描述的自我和交叉注意力", "操作返回更新的动作嵌人", "action-to-joint-action self-attention表示在联合动作a=（ap，a-p）下每个动作ap的play", "向所有玩家给出收益值。回想一下自我关注操作ag=ku=(concat(ap,Gp(a)))p=1的", "", "定义，对于每个联合动作，每个动作都产生一个向量输出。我们将输出p(ap，a-p）称为ap", "在a和收益G的联合作用下的play；", "action-to-playcross-attention然后对来自每个操作at的所有play的信息进行编码，键", "值ku=(pG(at,a-p）|Va-p}和g=at)是一个单例查询。此操作产生一个单例输出", "a】，作为其所有play及其输入动作嵌人向量的函数；", "3" ]

{ "arxiv_id": "2402.08393", "language": "zh", "page_idx": 2 }

[ [ "0.1748", "0.03284", "0.5073", "0.03284", "0.5073", "0.04736", "0.1748", "0.04736" ], [ "0.5195", "0.11237", "0.614", "0.11237", "0.614", "0.1237", "0.5195", "0.1237" ], [ "0.3137", "0.127", "0.3203", "0.127", "0.3203", "0.13", "0.3137", "0.13" ], [ "0.3896", "0.1276", "0.4043", "0.1276", "0.4043", "0.1307", "0.3896", "0.1307" ], [ "0.6724", "0.127", "0.7036", "0.127", "0.7036", "0.1307", "0.6724", "0.1307" ], [ "0.1978", "0.1364", "0.2133", "0.1414", "0.2075", "0.1509", "0.1929", "0.1459" ], [ "0.2607", "0.1395", "0.3945", "0.1395", "0.3945", "0.1528", "0.2607", "0.1528" ], [ "0.4429", "0.1351", "0.568", "0.1351", "0.568", "0.1459", "0.4429", "0.1459" ], [ "0.6436", "0.1395", "0.7573", "0.1395", "0.7573", "0.1528", "0.6436", "0.1528" ], [ "0.3228", "0.1901", "0.4272", "0.1913", "0.4272", "0.2026", "0.3228", "0.202" ], [ "0.5337", "0.1913", "0.603", "0.1913", "0.603", "0.2026", "0.5337", "0.2026" ], [ "0.671", "0.1932", "0.705", "0.1932", "0.705", "0.2026", "0.671", "0.2026" ], [ "0.5415", "0.2008", "0.576", "0.2084", "0.5703", "0.2235", "0.536", "0.2166" ], [ "0.3145", "0.2084", "0.3423", "0.2084", "0.3423", "0.2197", "0.3145", "0.2197" ], [ "0.66", "0.209", "0.6846", "0.209", "0.6846", "0.2172", "0.66", "0.2172" ], [ "0.3733", "0.2184", "0.3774", "0.2184", "0.3774", "0.2235", "0.3733", "0.2235" ], [ "0.705", "0.2172", "0.71", "0.2172", "0.71", "0.2247", "0.705", "0.2247" ], [ "0.4272", "0.226", "0.5146", "0.226", "0.5146", "0.2405", "0.4272", "0.2405" ], [ "0.3694", "0.2343", "0.384", "0.2343", "0.384", "0.2922", "0.3694", "0.2922" ], [ "0.564", "0.2291", "0.5767", "0.2291", "0.574", "0.2715", "0.5605", "0.2715" ], [ "0.4387", "0.2424", "0.5293", "0.2424", "0.5293", "0.2524", "0.4387", "0.2524" ], [ "0.706", "0.2393", "0.7114", "0.2393", "0.7114", "0.2734", "0.706", "0.2734" ], [ "0.2345", "0.2568", "0.3342", "0.2568", "0.3342", "0.2683", "0.2345", "0.2683" ], [ "0.6284", "0.2583", "0.6543", "0.2583", "0.6543", "0.2708", "0.6284", "0.2708" ], [ "0.4983", "0.2644", "0.5293", "0.2644", "0.5293", "0.2683", "0.4983", "0.2683" ], [ "0.562", "0.2683", "0.576", "0.2683", "0.576", "0.2986", "0.562", "0.2986" ], [ "0.3186", "0.2778", "0.3381", "0.2778", "0.3381", "0.2822", "0.3186", "0.2822" ], [ "0.6274", "0.2778", "0.671", "0.2778", "0.671", "0.2822", "0.6274", "0.2822" ], [ "0.7036", "0.2778", "0.7134", "0.2778", "0.7134", "0.2922", "0.7036", "0.2922" ], [ "0.4207", "0.2854", "0.5215", "0.2854", "0.5215", "0.2998", "0.4207", "0.2998" ], [ "0.6216", "0.281", "0.671", "0.2842", "0.669", "0.2942", "0.621", "0.2917" ], [ "0.5664", "0.3062", "0.5728", "0.3062", "0.5728", "0.3125", "0.5664", "0.3125" ], [ "0.3015", "0.3257", "0.3162", "0.3257", "0.3162", "0.334", "0.3015", "0.334" ], [ "0.562", "0.324", "0.574", "0.324", "0.574", "0.3352", "0.562", "0.3352" ], [ "0.626", "0.334", "0.7534", "0.334", "0.7534", "0.348", "0.626", "0.348" ], [ "0.1765", "0.3687", "0.8228", "0.3687", "0.8228", "0.3818", "0.1765", "0.3818" ], [ "0.1765", "0.3818", "0.824", "0.3818", "0.824", "0.3958", "0.1765", "0.3958" ], [ "0.1748", "0.3977", "0.8228", "0.3977", "0.8228", "0.411", "0.1748", "0.411" ], [ "0.1732", "0.4104", "0.8228", "0.411", "0.8228", "0.4255", "0.1732", "0.4248" ], [ "0.1748", "0.4255", "0.824", "0.4255", "0.824", "0.4395", "0.1748", "0.4395" ], [ "0.1732", "0.4395", "0.2876", "0.4395", "0.2876", "0.4526", "0.1732", "0.4526" ], [ "0.179", "0.488", "0.3987", "0.488", "0.3987", "0.502", "0.179", "0.502" ], [ "0.4028", "0.488", "0.7656", "0.488", "0.7656", "0.502", "0.4028", "0.502" ], [ "0.7656", "0.4924", "0.8057", "0.4924", "0.8057", "0.5", "0.7656", "0.5" ], [ "0.1765", "0.502", "0.8228", "0.502", "0.8228", "0.5166", "0.1765", "0.5166" ], [ "0.1765", "0.518", "0.3882", "0.518", "0.3882", "0.532", "0.1765", "0.532" ], [ "0.1748", "0.5386", "0.819", "0.5386", "0.819", "0.552", "0.1748", "0.552" ], [ "0.1748", "0.552", "0.8228", "0.553", "0.8228", "0.5674", "0.1748", "0.5664" ], [ "0.1748", "0.5684", "0.2263", "0.5684", "0.2263", "0.5835", "0.1748", "0.5835" ], [ "0.1748", "0.608", "0.2018", "0.608", "0.2018", "0.6206", "0.1748", "0.6206" ], [ "0.2141", "0.607", "0.3782", "0.608", "0.3782", "0.6226", "0.2141", "0.622" ], [ "0.1748", "0.6353", "0.8228", "0.636", "0.8228", "0.651", "0.1748", "0.6494" ], [ "0.1724", "0.6484", "0.2418", "0.651", "0.2418", "0.6655", "0.1724", "0.6636" ], [ "0.1805", "0.686", "0.818", "0.6875", "0.818", "0.702", "0.1805", "0.7007" ], [ "0.1732", "0.7007", "0.824", "0.7007", "0.824", "0.7183", "0.1732", "0.7183" ], [ "0.1732", "0.7183", "0.826", "0.72", "0.826", "0.738", "0.1732", "0.7363" ], [ "0.1732", "0.7373", "0.5317", "0.7373", "0.5317", "0.752", "0.1732", "0.752" ], [ "0.1805", "0.7744", "0.4224", "0.7744", "0.4224", "0.7886", "0.1805", "0.7886" ], [ "0.4167", "0.7744", "0.822", "0.7744", "0.822", "0.789", "0.4167", "0.789" ], [ "0.1748", "0.7905", "0.824", "0.7905", "0.824", "0.808", "0.1748", "0.808" ], [ "0.1732", "0.8066", "0.822", "0.8096", "0.822", "0.827", "0.1732", "0.8247" ], [ "0.1748", "0.8276", "0.505", "0.8276", "0.505", "0.8413", "0.1748", "0.8413" ], [ "0.1805", "0.8643", "0.357", "0.8643", "0.357", "0.8774", "0.1805", "0.8774" ], [ "0.3513", "0.8643", "0.8228", "0.8643", "0.8228", "0.8774", "0.3513", "0.8774" ], [ "0.1772", "0.8774", "0.822", "0.8774", "0.822", "0.8916", "0.1772", "0.8916" ], [ "0.1765", "0.893", "0.824", "0.893", "0.824", "0.9077", "0.1765", "0.9077" ], [ "0.1748", "0.909", "0.7183", "0.91", "0.7183", "0.925", "0.1748", "0.924" ], [ "0.4944", "0.9497", "0.5063", "0.9497", "0.5063", "0.9604", "0.4944", "0.9604" ] ]

[ "Published as a conference paper at ICLR 2024", "NfgTransformer", "tie", "pne", "nalarti", "G", "(a,=O| Vte[T],Vpe[N]", "NfgTransformerBlock", "(a, Vt e[T], Vpe[N])", "Action-to-joint-action", "Action-to-play", "Action-", "KV1", "QKV", "QKV", "", "-", "(pG(agrap) / a.", "", "", "or each of NxT action:", "oe", "oncat(ap,Gp(a))", "(ag)", "rranaino", "SS", "etion:", "For.each.o", "s", "(pG(apr ap) / Vp)", "NxTactions", "+", "a,", "Q", "NfgTransformerBlock", "Figure l: An overview of the NfgTransformer, The payoff tensor G is encoded as action", "embeddings {at | Vt E [Tl, Vp e [Nll (Top). Action embeddings are zero-initialised and", "iteratively updated through a sequence of K NfgTransformer blocks (Bottom). An arrow", "labeled with \"(Q)KV\" originates from a set of input (query-)key-values and terminates at", "a set of outputs. Each dashed box denotes an unordered set of elements of a specific type", "and cardinality.", "action-to-action self-attention", "然后表示给定所有动作嵌人的每个动作。这里，。", "=kn", "[at「VpE[N],VtE[T]]。我们在5.1节中废除了这个操作（通过改变A），展示了它在跨", "动作嵌入传播信息方面的优势", "在每个块内，在给定键值顺序不变性的情况下保留等方差，并查询注意操作的等方差属性", "每个输出嵌人a,都是其自已在输人处嵌人的函数，其play的无序集以及所有动作嵌人的", "无序集", "4.2", "Task-specific decoding", "生成的动作嵌入可用于不同解码粒度的各种下游任务。我们描述并实证演示了三个用例", "(图 2 )。", "Nashequilibrium-solving需要在每个动作的级别进行解码，从而估计每个玩家p", "softmax(wp，,w）的边际动作分布，其中wt=MLP(at）是动作a,的logit。在这里，我", "们遵循Duan et al.(2023)，通过=@po,和8()=maxaEA,Ea~[Gp(ap,a-p)-Gp(a))", "的梯度下降来最小化损失函数maxpo,(o）端到端。", "Max deviation-gain estimation", "解码每个联合行动a=（ai,...,a),VaEA的标量估计。", "在这里，我们将每个关节动作表示为a=,ap，并通过最小化L2((a),maxpp(a)）=", "((a)-maxpop(a)2来估计其最大偏差增益(a)，其中p(a)=maxaA,[Gp(ap,a-p）-", "G(a)l是在关节动作α下玩家p的偏差增益。", "Payoff reconstruction", "为每个收益值G,(a),VpE[NI,VaEA解码一个标量。在这里，我", "们使用自注意操作进行解码，类似于编码器中的动作到关节动作自注意操作，但是without", "附加到动作嵌人，它们的收益值将被重建。为了计算重建损失，我们将L2(Gp(α),Gp(a))最", "小化，其中G（a）是动作嵌入ap的函数，而协作者动作嵌人a-p的无序集", "4" ]

{ "arxiv_id": "2402.08393", "language": "zh", "page_idx": 3 }

[ [ "0.1748", "0.0322", "0.5073", "0.03284", "0.5073", "0.04736", "0.1748", "0.04672" ], [ "0.2385", "0.1092", "0.326", "0.1092", "0.326", "0.1237", "0.2385", "0.1237" ], [ "0.4395", "0.1092", "0.558", "0.1092", "0.558", "0.1237", "0.4395", "0.1237" ], [ "0.664", "0.108", "0.7646", "0.10986", "0.7646", "0.1256", "0.664", "0.1237" ], [ "0.544", "0.1332", "0.693", "0.1332", "0.693", "0.1433", "0.544", "0.1433" ], [ "0.4387", "0.1666", "0.556", "0.1692", "0.556", "0.1843", "0.4387", "0.1812" ], [ "0.6543", "0.1686", "0.7715", "0.1704", "0.7715", "0.1881", "0.6543", "0.1863" ], [ "0.513", "0.1901", "0.54", "0.1901", "0.54", "0.1976", "0.513", "0.1976" ], [ "0.2059", "0.1718", "0.3308", "0.1724", "0.3308", "0.1875", "0.2059", "0.1863" ], [ "0.451", "0.1925", "0.478", "0.1925", "0.478", "0.1963", "0.451", "0.1963" ], [ "0.501", "0.1932", "0.5146", "0.1932", "0.5146", "0.1963", "0.501", "0.1963" ], [ "0.6587", "0.1925", "0.7793", "0.1925", "0.7793", "0.2026", "0.6587", "0.2026" ], [ "0.4395", "0.2058", "0.596", "0.2058", "0.596", "0.2166", "0.4395", "0.2166" ], [ "0.2615", "0.1976", "0.361", "0.1976", "0.361", "0.209", "0.2615", "0.209" ], [ "0.281", "0.2153", "0.2893", "0.2153", "0.2893", "0.2247", "0.281", "0.2247" ], [ "0.7017", "0.2134", "0.728", "0.2134", "0.728", "0.2247", "0.7017", "0.2247" ], [ "0.694", "0.2235", "0.739", "0.2235", "0.739", "0.2349", "0.694", "0.2349" ], [ "0.2418", "0.2374", "0.3308", "0.2393", "0.3308", "0.2568", "0.2418", "0.2551" ], [ "0.4355", "0.2349", "0.5645", "0.2349", "0.5645", "0.25", "0.4355", "0.25" ], [ "0.66", "0.2468", "0.786", "0.2468", "0.786", "0.2568", "0.66", "0.2568" ], [ "0.2279", "0.2644", "0.3513", "0.2651", "0.3513", "0.2766", "0.2279", "0.276" ], [ "0.46", "0.2715", "0.528", "0.2715", "0.528", "0.2822", "0.46", "0.2822" ], [ "0.687", "0.2676", "0.7114", "0.2695", "0.71", "0.2778", "0.6855", "0.276" ], [ "0.7305", "0.2727", "0.744", "0.2734", "0.744", "0.2766", "0.7305", "0.276" ], [ "0.2524", "0.2917", "0.3857", "0.2917", "0.3857", "0.305", "0.2524", "0.305" ], [ "0.4983", "0.2954", "0.517", "0.31", "0.5", "0.3245", "0.4795", "0.3093" ], [ "0.7017", "0.2935", "0.8145", "0.2935", "0.8145", "0.3037", "0.7017", "0.3037" ], [ "0.1748", "0.332", "0.8228", "0.332", "0.8228", "0.3452", "0.1748", "0.3452" ], [ "0.1765", "0.3452", "0.2173", "0.348", "0.2157", "0.3599", "0.1748", "0.358" ], [ "0.1748", "0.3865", "0.4429", "0.387", "0.4429", "0.4016", "0.1748", "0.401" ], [ "0.1765", "0.414", "0.8228", "0.414", "0.8228", "0.428", "0.1765", "0.428" ], [ "0.1765", "0.428", "0.822", "0.428", "0.822", "0.4414", "0.1765", "0.4414" ], [ "0.179", "0.443", "0.822", "0.443", "0.822", "0.4539", "0.179", "0.4539" ], [ "0.1765", "0.455", "0.8228", "0.455", "0.8228", "0.4697", "0.1765", "0.4697" ], [ "0.1765", "0.4697", "0.8228", "0.4697", "0.8228", "0.483", "0.1765", "0.483" ], [ "0.1765", "0.4841", "0.4478", "0.4841", "0.4478", "0.4976", "0.1765", "0.4976" ], [ "0.5645", "0.515", "0.7524", "0.5176", "0.7524", "0.532", "0.5645", "0.53" ], [ "0.2222", "0.5234", "0.564", "0.5234", "0.564", "0.5415", "0.2222", "0.5415" ], [ "0.8013", "0.5264", "0.8228", "0.5264", "0.8228", "0.539", "0.8013", "0.539" ], [ "0.6445", "0.534", "0.6772", "0.534", "0.6772", "0.55", "0.6445", "0.55" ], [ "0.179", "0.561", "0.822", "0.561", "0.822", "0.5747", "0.179", "0.5747" ], [ "0.1772", "0.5757", "0.8228", "0.5757", "0.8228", "0.5903", "0.1772", "0.5903" ], [ "0.1765", "0.589", "0.824", "0.589", "0.824", "0.603", "0.1765", "0.603" ], [ "0.1765", "0.604", "0.7236", "0.604", "0.7236", "0.617", "0.1765", "0.617" ], [ "0.1765", "0.6396", "0.1896", "0.6396", "0.1896", "0.651", "0.1765", "0.651" ], [ "0.2075", "0.638", "0.2754", "0.638", "0.2754", "0.653", "0.2075", "0.653" ], [ "0.1765", "0.6694", "0.8228", "0.6694", "0.8228", "0.683", "0.1765", "0.683" ], [ "0.1772", "0.685", "0.8203", "0.685", "0.8203", "0.696", "0.1772", "0.696" ], [ "0.1765", "0.696", "0.8203", "0.696", "0.8203", "0.7104", "0.1765", "0.7104" ], [ "0.1772", "0.711", "0.822", "0.711", "0.822", "0.7246", "0.1772", "0.7246" ], [ "0.179", "0.7266", "0.822", "0.7266", "0.822", "0.7373", "0.179", "0.7373" ], [ "0.1765", "0.7393", "0.8228", "0.7393", "0.8228", "0.7524", "0.1765", "0.7524" ], [ "0.1765", "0.7524", "0.8228", "0.7524", "0.8228", "0.7666", "0.1765", "0.7666" ], [ "0.1765", "0.7666", "0.3245", "0.7666", "0.3245", "0.78", "0.1765", "0.78" ], [ "0.1724", "0.7974", "0.335", "0.7974", "0.335", "0.811", "0.1724", "0.811" ], [ "0.1765", "0.8237", "0.8228", "0.8237", "0.8228", "0.837", "0.1765", "0.837" ], [ "0.179", "0.837", "0.822", "0.837", "0.822", "0.851", "0.179", "0.851" ], [ "0.1765", "0.8516", "0.8228", "0.8516", "0.8228", "0.8657", "0.1765", "0.8657" ], [ "0.1765", "0.867", "0.8228", "0.867", "0.8228", "0.882", "0.1765", "0.882" ], [ "0.1772", "0.8833", "0.822", "0.8833", "0.822", "0.8965", "0.1772", "0.8965" ], [ "0.179", "0.898", "0.2173", "0.898", "0.2173", "0.909", "0.179", "0.909" ], [ "0.2173", "0.8965", "0.8228", "0.8965", "0.8228", "0.91", "0.2173", "0.91" ], [ "0.1748", "0.911", "0.822", "0.911", "0.822", "0.925", "0.1748", "0.925" ], [ "0.492", "0.9478", "0.5073", "0.9478", "0.5073", "0.962", "0.492", "0.962" ] ]

[ "Published as a conference paper at ICLR 2024", "maxpop(@pop)", "L2(8(a),max,op(a))", "L2(Gp(a), Gp(a)", "Task-specific loss functions", "6(a) = MLP(Z,ap)", "[Gi(a),...,GN(a))", "joint", "o,=softmax(wp,...", "real", "", "For each of TN joint-actio", "Max deviation-gain estimation", "Foreachof Nplayen", "f", "Self", "Attention", "w, = MLP(ab)", "[a, / Vt e [T],pe[N]", "ForeachofTN joint-action", "For each of NxT actions", "NfgTransform", "al::", "HA", "Nash equilibrium solving", "G", "Payoff reconstruction", "Figure 2: Example task-specific decoders and losses from general-purpose action embed-", "dings.", "4.3Representing incompletegames", "到目前为止，我们假设观察到游戏的每个联合行动的结果。在实践中并非总是如此，因为评", "估每个联合行动的成本可能令人望而却步。相反，给定不完整的NFG（Elo，2008），可以推", "新出所有联合行动的结果。稍加修改，NfgTransformer以原则的方式适应这样的用例。为", "此，我们扩展了香草注意力实现（公式E3），以允许为查询和键值输提供额外的二进制", "掩码向量mgE0,1]ng和mkuE[0,1nk，以指示其有效性。方程4定义了这个掩蔽注", "意操作，用1。1)=0和1(0)=80", "(QKT-1(mg@mk)", "MaskedAttention(Q, K,V,mq,mk) = softmax", "(4)", "Vdk", "为了表示不完整的NFG，我们在所有自注意和交叉注意操作中使用公式E4代替公式E3", "并相应地设置掩码向量。例如，如果未观察到联合动作，则可以设置mku以反映每个play", "对动作播放操作的有效性。NfgTransformer架构对于表示不完整游戏中的动作并预测未观", "察到的联合行动的收益非常有效，正如我们在E5.2节中经验性地展示的那样", "5", "Results", "我们实证研究的目标是三重的。首先，我们将NfgTransformer与基准架构进行比较，以展", "示在合成和标准游戏的各种下游任务中性能的提高。其次，我们改变模型超参数，并观察它", "们如何影响性能。我们特别展示了一些任务需要更大的动作嵌人大小，而其他任务则受益于", "更多轮的迭代细化。最后，我们研究了how模型，该模型通过在受控环境中解释学习的注", "意力掩码序列来解决某些任务。我们的结果表明，该模型找到的解决方案反映了游戏中NE", "求解的直观解决方案的元素。为了完整性，我们在附录D中讨论了额外的经验结果。例如，", "我们显示所有非平凡的平衡不变的2×2游戏都是按照Marrisetal.（2023）中提出的已知嵌", "人空间的结构嵌入的", "5.1 Synthetic games", "我们首先评估NfgTransformer架构在NE平衡求解和偏差增益估计上不同大小的合成游戏", "上的变化。为了生成具有广泛覆盖范围的合成游戏，我们遵循Marrisetal.（2022），它从平", "衡不变子空间对游戏进行采样，涵盖了可能影响NFG平衡解的所有战略互动。每个游戏的", "收益张量G相对于其他玩家策略和Frobenius范数IGplE=VTN具有零均值。我们将结", "果与基线MLP网络进行比较，其参数atleast是我们最大的变压器变体（在4。95M参", "数）：", "一个等变的MLP网络，它按照平均收益的降序重新排列动作，以及一个为平衡求解", "而设计的NES（Marrisetal.，2022）网络。有关每个模型的游戏采样，网络架构和参数计数", "5" ]

{ "arxiv_id": "2402.08393", "language": "zh", "page_idx": 4 }

[ [ "0.6733", "0.07385", "0.9126", "0.0745", "0.9126", "0.0884", "0.6733", "0.0871" ], [ "0.822", "0.1029", "0.9126", "0.1048", "0.9126", "0.1193", "0.822", "0.11743" ], [ "0.4084", "0.1187", "0.501", "0.1206", "0.501", "0.132", "0.4084", "0.13" ], [ "0.544", "0.11554", "0.576", "0.11554", "0.576", "0.1237", "0.544", "0.1237" ], [ "0.7197", "0.1218", "0.8096", "0.1218", "0.8096", "0.1332", "0.7197", "0.1332" ], [ "0.4878", "0.1528", "0.5107", "0.1528", "0.5107", "0.1654", "0.4878", "0.1654" ], [ "0.4558", "0.1799", "0.5435", "0.1799", "0.5435", "0.1913", "0.4558", "0.1913" ], [ "0.5195", "0.1976", "0.546", "0.1976", "0.546", "0.2102", "0.5195", "0.2102" ], [ "0.7305", "0.1995", "0.7524", "0.1995", "0.7524", "0.2084", "0.7305", "0.2084" ], [ "0.2197", "0.2229", "0.2402", "0.2229", "0.2402", "0.2299", "0.2197", "0.2299" ], [ "0.3342", "0.2008", "0.3506", "0.2008", "0.3506", "0.2084", "0.3342", "0.2084" ], [ "0.3342", "0.2229", "0.3545", "0.2229", "0.3545", "0.2299", "0.3342", "0.2299" ], [ "0.5034", "0.2229", "0.539", "0.2229", "0.539", "0.2299", "0.5034", "0.2299" ], [ "0.7197", "0.2229", "0.746", "0.2229", "0.746", "0.2299", "0.7197", "0.2299" ], [ "0.3145", "0.2343", "0.335", "0.2343", "0.335", "0.2374", "0.3145", "0.2374" ], [ "0.5024", "0.2329", "0.528", "0.2329", "0.528", "0.2405", "0.5024", "0.2405" ], [ "0.709", "0.2343", "0.726", "0.2343", "0.726", "0.2374", "0.709", "0.2374" ], [ "0.819", "0.2361", "0.914", "0.238", "0.914", "0.2524", "0.819", "0.2507" ], [ "0.3945", "0.2437", "0.407", "0.2437", "0.407", "0.2474", "0.3945", "0.2474" ], [ "0.5024", "0.2424", "0.539", "0.2424", "0.539", "0.25", "0.5024", "0.25" ], [ "0.721", "0.2437", "0.739", "0.2437", "0.739", "0.2474", "0.721", "0.2474" ], [ "0.2615", "0.2644", "0.3823", "0.2644", "0.3823", "0.2747", "0.2615", "0.2747" ], [ "0.2157", "0.2443", "0.2222", "0.2443", "0.2222", "0.2474", "0.2157", "0.2474" ], [ "0.636", "0.2644", "0.7573", "0.2644", "0.7573", "0.2747", "0.636", "0.2747" ], [ "0.2255", "0.2798", "0.2793", "0.2798", "0.2793", "0.2874", "0.2255", "0.2874" ], [ "0.437", "0.2798", "0.468", "0.2798", "0.468", "0.2886", "0.437", "0.2886" ], [ "0.6445", "0.281", "0.669", "0.281", "0.669", "0.2886", "0.6445", "0.2886" ], [ "0.509", "0.305", "0.524", "0.305", "0.524", "0.3125", "0.509", "0.3125" ], [ "0.1544", "0.281", "0.1748", "0.281", "0.1748", "0.2886", "0.1544", "0.2886" ], [ "0.609", "0.3062", "0.621", "0.3093", "0.6177", "0.3132", "0.607", "0.3113" ], [ "0.714", "0.305", "0.726", "0.305", "0.726", "0.3132", "0.714", "0.3132" ], [ "0.1871", "0.324", "0.1969", "0.324", "0.1969", "0.3276", "0.1871", "0.3276" ], [ "0.4126", "0.324", "0.4192", "0.324", "0.4192", "0.3276", "0.4126", "0.3276" ], [ "0.614", "0.324", "0.626", "0.324", "0.626", "0.33", "0.614", "0.33" ], [ "0.871", "0.3245", "0.902", "0.3245", "0.902", "0.336", "0.871", "0.336" ], [ "0.08496", "0.363", "0.9126", "0.363", "0.9126", "0.377", "0.08496", "0.377" ], [ "0.08496", "0.377", "0.9126", "0.3757", "0.9126", "0.3894", "0.08496", "0.3901" ], [ "0.0866", "0.3926", "0.9126", "0.3926", "0.9126", "0.4028", "0.0866", "0.4028" ], [ "0.08496", "0.4048", "0.914", "0.4048", "0.914", "0.4192", "0.08496", "0.4192" ], [ "0.08496", "0.4185", "0.5317", "0.4185", "0.5317", "0.432", "0.08496", "0.432" ], [ "0.11273", "0.455", "0.4836", "0.455", "0.4836", "0.4685", "0.11273", "0.4685" ], [ "0.5186", "0.4539", "0.7915", "0.4539", "0.7915", "0.4673", "0.5186", "0.4673" ], [ "0.125", "0.4685", "0.4836", "0.4685", "0.4836", "0.4817", "0.125", "0.4817" ], [ "0.5293", "0.4702", "0.8555", "0.4702", "0.8555", "0.4912", "0.5293", "0.4912" ], [ "0.1259", "0.483", "0.21", "0.483", "0.21", "0.4944", "0.1259", "0.4944" ], [ "0.5312", "0.488", "0.847", "0.488", "0.847", "0.502", "0.5312", "0.502" ], [ "0.1136", "0.4968", "0.4453", "0.4968", "0.4453", "0.5103", "0.1136", "0.5103" ], [ "0.5312", "0.502", "0.652", "0.502", "0.652", "0.515", "0.5312", "0.515" ], [ "0.11273", "0.5103", "0.4805", "0.5103", "0.4805", "0.5234", "0.11273", "0.5234" ], [ "0.546", "0.515", "0.6855", "0.515", "0.6855", "0.529", "0.546", "0.529" ], [ "0.125", "0.5234", "0.4836", "0.5234", "0.4836", "0.537", "0.125", "0.537" ], [ "0.546", "0.529", "0.7344", "0.529", "0.7344", "0.5435", "0.546", "0.5435" ], [ "0.125", "0.537", "0.4836", "0.537", "0.4836", "0.5503", "0.125", "0.5503" ], [ "0.5596", "0.5435", "0.8145", "0.5435", "0.8145", "0.558", "0.5596", "0.558" ], [ "0.1259", "0.553", "0.183", "0.553", "0.183", "0.5645", "0.1259", "0.5645" ], [ "0.5474", "0.558", "0.589", "0.558", "0.589", "0.5693", "0.5474", "0.5693" ], [ "0.11273", "0.5645", "0.4836", "0.5645", "0.4836", "0.5776", "0.11273", "0.5776" ], [ "0.5312", "0.5684", "0.585", "0.571", "0.584", "0.5854", "0.5293", "0.5835" ], [ "0.1259", "0.5806", "0.482", "0.5806", "0.482", "0.5913", "0.1259", "0.5913" ], [ "0.125", "0.5947", "0.1871", "0.5947", "0.1871", "0.606", "0.125", "0.606" ], [ "0.1168", "0.6187", "0.259", "0.6187", "0.259", "0.632", "0.1168", "0.632" ], [ "0.517", "0.616", "0.9126", "0.617", "0.9126", "0.632", "0.517", "0.631" ], [ "0.5317", "0.632", "0.9126", "0.632", "0.9126", "0.646", "0.5317", "0.646" ], [ "0.0866", "0.6206", "0.0989", "0.6206", "0.0989", "0.63", "0.0866", "0.63" ], [ "0.0866", "0.6353", "0.4697", "0.6353", "0.4697", "0.653", "0.0866", "0.653" ], [ "0.517", "0.6445", "0.9126", "0.646", "0.9126", "0.66", "0.517", "0.659" ], [ "0.08826", "0.651", "0.4795", "0.651", "0.4795", "0.664", "0.08826", "0.664" ], [ "0.08826", "0.664", "0.482", "0.664", "0.482", "0.678", "0.08826", "0.678" ], [ "0.5195", "0.66", "0.9126", "0.66", "0.9126", "0.674", "0.5195", "0.674" ], [ "0.0866", "0.6787", "0.451", "0.6787", "0.451", "0.6924", "0.0866", "0.6924" ], [ "0.5215", "0.677", "0.911", "0.677", "0.911", "0.6875", "0.5215", "0.6875" ], [ "0.5195", "0.6895", "0.9126", "0.6895", "0.9126", "0.7026", "0.5195", "0.7026" ], [ "0.353", "0.6978", "0.3635", "0.6978", "0.3635", "0.7085", "0.353", "0.7085" ], [ "0.517", "0.702", "0.9126", "0.7026", "0.9126", "0.7173", "0.517", "0.7163" ], [ "0.4617", "0.7026", "0.482", "0.7026", "0.482", "0.7163", "0.4617", "0.7163" ], [ "0.1683", "0.7026", "0.3203", "0.7026", "0.3203", "0.7163", "0.1683", "0.7163" ], [ "0.3203", "0.711", "0.398", "0.7085", "0.3987", "0.7227", "0.3218", "0.726" ], [ "0.5195", "0.7173", "0.9126", "0.7173", "0.9126", "0.731", "0.5195", "0.731" ], [ "0.0989", "0.7266", "0.4805", "0.7266", "0.4805", "0.7407", "0.0989", "0.7407" ], [ "0.5186", "0.731", "0.789", "0.731", "0.789", "0.744", "0.5186", "0.744" ], [ "0.0866", "0.741", "0.4805", "0.741", "0.4805", "0.755", "0.0866", "0.755" ], [ "0.535", "0.744", "0.9126", "0.744", "0.9126", "0.7583", "0.535", "0.7583" ], [ "0.0866", "0.755", "0.4805", "0.755", "0.4805", "0.768", "0.0866", "0.768" ], [ "0.5195", "0.7603", "0.6436", "0.7603", "0.6436", "0.771", "0.5195", "0.771" ], [ "0.0866", "0.768", "0.4805", "0.768", "0.4805", "0.782", "0.0866", "0.782" ], [ "0.638", "0.7764", "0.6504", "0.786", "0.638", "0.7954", "0.626", "0.786" ], [ "0.0866", "0.784", "0.10297", "0.784", "0.10297", "0.7935", "0.0866", "0.7935" ], [ "0.5483", "0.7886", "0.8774", "0.786", "0.8774", "0.808", "0.5483", "0.81" ], [ "0.10297", "0.7954", "0.4805", "0.7954", "0.4805", "0.8096", "0.10297", "0.8096" ], [ "0.8945", "0.7954", "0.911", "0.7954", "0.911", "0.8047", "0.8945", "0.8047" ], [ "0.0866", "0.81", "0.482", "0.81", "0.482", "0.8237", "0.0866", "0.8237" ], [ "0.625", "0.8066", "0.649", "0.8066", "0.649", "0.8184", "0.625", "0.8184" ], [ "0.0866", "0.8237", "0.4805", "0.8237", "0.4805", "0.837", "0.0866", "0.837" ], [ "0.527", "0.8257", "0.911", "0.827", "0.911", "0.8413", "0.527", "0.8403" ], [ "0.5254", "0.832", "0.5317", "0.832", "0.5317", "0.836", "0.5254", "0.836" ], [ "0.08496", "0.8384", "0.415", "0.8384", "0.415", "0.8516", "0.08496", "0.8516" ], [ "0.5186", "0.8413", "0.7876", "0.8413", "0.7876", "0.855", "0.5186", "0.855" ], [ "0.2998", "0.856", "0.3179", "0.856", "0.3179", "0.869", "0.2998", "0.869" ], [ "0.535", "0.855", "0.9126", "0.855", "0.9126", "0.8687", "0.535", "0.8687" ], [ "0.304", "0.8687", "0.4028", "0.866", "0.4028", "0.881", "0.304", "0.8833" ], [ "0.4631", "0.8687", "0.4836", "0.8687", "0.4836", "0.882", "0.4631", "0.882" ], [ "0.5186", "0.8687", "0.911", "0.8687", "0.911", "0.882", "0.5186", "0.882" ], [ "0.2695", "0.872", "0.2998", "0.872", "0.2998", "0.88", "0.2695", "0.88" ], [ "0.165", "0.869", "0.2607", "0.869", "0.2607", "0.881", "0.165", "0.881" ], [ "0.2646", "0.8735", "0.2712", "0.8735", "0.2712", "0.8774", "0.2646", "0.8774" ], [ "0.2998", "0.875", "0.3096", "0.875", "0.3096", "0.88", "0.2998", "0.88" ], [ "0.2957", "0.8833", "0.3186", "0.8833", "0.3186", "0.8945", "0.2957", "0.8945" ], [ "0.5186", "0.8833", "0.914", "0.8833", "0.914", "0.8965", "0.5186", "0.8965" ] ]

[ "KDD'23, August 610, 2023, Long Beach, CA, USA", "model training", "offlinemodel (v2)", "deploy", "offlinemodel (v3)", "1/0", "data subsampling", "drop", "keep", "M.V", "dro", "a.V", "x=(uM", "Ca.v", "", "4=0", "", "model serving", "n", "x)=0.2", "n", "online (pilot) model (v))", "", "online (pilot)model (v2)", "(2)send (3)", "send (4)", "nd(5)", "<(4", "nd(I", "e", "(5", "", "-", "E", "traffic", "Figure 1:A simplified recommendation system.A user initiates a request to the online serving model and receives recommen", "dations returnedby the server.User clicks will trigger positive instances, while non-clicks will trigger negative instances.All", "instances arefilteredbythedata subsamplingmodule.Formodel-based methods,subsamplingrates aredeterminedby online", "models (e.g. model v1), thus sub-optimal for offline training (e.g. model v2) purposes. When model v2 is deployed online, data", "subsamplingis affected,producinginconsistentsubsamplingrates.", " we propose a model-agnostic method to measure data hard-", "Algorithm 1 Hard Negative Sampling (HNS)", "ness scores via effective conductance on the user-item bi-", "Given (xn, yn)N-, and hardness score function h()", "partite graph;", "Compute the counterfactual NSR π(xn) via Eqn. (2).", "· we refine hardness scores via propagation on graphs;", "for n= I,..N do", " empirical results demonstrate that without requesting in", "Generate An ~ U(0, 1) ", "formation from the pilot model, our model-agnostic meth-", "if yn = 1 or An ≤ π(xn) then", "ods achieve comparable or better results than model-based", "Include (Xn, yn, (xn)) in the training set.", "methods;", "endif", " we verify model-agnostic methods compliment model-based", "endfor", "methods and combining them can achieve even better per-", "formance.", "PRELIMINARIES", "其中αE（0,1)是负实例的预设平均子采样率。假设我们有", "个与图(2)中的在线模型v1相对应的试验模型f(）。当", "2", "我们考虑一个二进制分类问题，让D=(xn,yn))N=1成为", "f(）=f(:0*），即(）具有与目标模型相同的函数形式，并", "个大小为N的训练集。这里xn和yn分别是实例n的特征向", "量和标签。我们研究广义逻辑回归（GLM)模型，其中目标模型", "且*是真实参数时。然后我们可以设置一个基于模型的硬度", "对应于离线模型(e。g。，图1中部署前的模型v2)表示为", "分数hh（xn）=f（xn）=f（xn：0*）以通过公式获得接近最佳的", "采样率元(xn）。(2)[34]。直觉上，试点模型f()预测的得分较", "1", "高的负面实例更“令人惊牙”，因此对于目标模型f(の）而言", "(1)", "f(x; 0) := p(y = 1|x, 0) =", "1 + e-g(x;0)", "更难。请注意，我们使用元（x）作为正实例来表示其反事实负", "其中对数奇数g(x;)由预测模型实现。将No表示为负实例", "采样率。我们在算法1中演示了HNS过程", "数，将Ni=N-No表示为正实例数。我们研究了D不平衡", "由于子采样后数据分布会发生变化，因此需要校正对数赔", "的情况，即。e。，No》Ni。由于信息稀疏分布在许多负面实", "率以获得无偏估计：", "例[33]上，我们经常使用NS来减少数据集大小并提高训练效", "X", "率", "= arg max8n [yng(xn:0) - log(1+ e9(n:0)-ta)],", "NS算法对每个负实例x进行加权，并对其重要性元(x)进", "(3)", "行一些度量。重要性元(x）用作x的负采样率（NSR）。元(x）的", "n=1", "分配遵循一种广泛使用的启发式方法，该方法利用“硬负样", "其中SnE(0,1)是子采样指标，n：=log元(xn）。[13]证明了", "", "本\"E[17,24]。采样率与非负硬度分数h(成正比：", "对数赔率校正比IPW估计器E[16]更有效", "No", "但是，当错误指定导频模型时，无法实现具有导频模型的最", "元(xn) = Noα,", "(2) ", "佳NS。在实际场景中，部署基于模型的HNS方法可能容易出", "s.t.", "(xn) αc h(xn),", "", "", "n=1", "错，因为由于在线-离线模型差异以及持续集成和部署，我们" ]

{ "arxiv_id": "2305.16391", "language": "zh", "page_idx": 2 }

[ [ "0.6733", "0.0745", "0.911", "0.07574", "0.911", "0.089", "0.6733", "0.0884" ], [ "0.08826", "0.108", "0.4805", "0.108", "0.4805", "0.1218", "0.08826", "0.1218" ], [ "0.5195", "0.108", "0.914", "0.108", "0.914", "0.1218", "0.5195", "0.1218" ], [ "0.08826", "0.1218", "0.4795", "0.1218", "0.4795", "0.1364", "0.08826", "0.1364" ], [ "0.5156", "0.1218", "0.8", "0.1206", "0.8", "0.1338", "0.5156", "0.1351" ], [ "0.08826", "0.137", "0.4795", "0.137", "0.4795", "0.1477", "0.08826", "0.1477" ], [ "0.0866", "0.1497", "0.3708", "0.1483", "0.3708", "0.1622", "0.0866", "0.1635" ], [ "0.5376", "0.1459", "0.773", "0.1459", "0.773", "0.159", "0.5376", "0.159" ], [ "0.6094", "0.1654", "0.822", "0.1666", "0.822", "0.1843", "0.6094", "0.1831" ], [ "0.10706", "0.1724", "0.478", "0.1724", "0.478", "0.1831", "0.10706", "0.1831" ], [ "0.886", "0.1704", "0.9126", "0.1704", "0.9126", "0.1831", "0.886", "0.1831" ], [ "0.08905", "0.1863", "0.4795", "0.1863", "0.4795", "0.1963", "0.08905", "0.1963" ], [ "0.08826", "0.1995", "0.4795", "0.1995", "0.4795", "0.2134", "0.08826", "0.2134" ], [ "0.5195", "0.1957", "0.9126", "0.1957", "0.9126", "0.209", "0.5195", "0.209" ], [ "0.08496", "0.2134", "0.2998", "0.2134", "0.2998", "0.2267", "0.08496", "0.2267" ], [ "0.5195", "0.2108", "0.911", "0.2108", "0.911", "0.2216", "0.5195", "0.2216" ], [ "0.5195", "0.2235", "0.8906", "0.2247", "0.8906", "0.2361", "0.5195", "0.2349" ], [ "0.0866", "0.2405", "0.1152", "0.2405", "0.1152", "0.2524", "0.0866", "0.2524" ], [ "0.1283", "0.2405", "0.4673", "0.2405", "0.4673", "0.254", "0.1283", "0.254" ], [ "0.517", "0.25", "0.569", "0.252", "0.568", "0.2664", "0.5156", "0.2644" ], [ "0.5605", "0.252", "0.819", "0.252", "0.819", "0.2651", "0.5605", "0.2651" ], [ "0.5596", "0.2683", "0.7363", "0.2683", "0.7363", "0.279", "0.5596", "0.279" ], [ "0.125", "0.2551", "0.2279", "0.2556", "0.2279", "0.2708", "0.125", "0.2695" ], [ "0.08826", "0.2747", "0.4805", "0.2747", "0.4805", "0.2886", "0.08826", "0.2886" ], [ "0.0866", "0.2886", "0.482", "0.2886", "0.482", "0.3018", "0.0866", "0.3018" ], [ "0.5186", "0.286", "0.911", "0.286", "0.911", "0.2998", "0.5186", "0.2998" ], [ "0.0866", "0.3018", "0.482", "0.3018", "0.482", "0.315", "0.0866", "0.315" ], [ "0.5186", "0.2998", "0.91", "0.2998", "0.91", "0.3132", "0.5186", "0.3132" ], [ "0.08826", "0.3176", "0.4763", "0.3176", "0.4763", "0.3276", "0.08826", "0.3276" ], [ "0.5186", "0.3132", "0.9126", "0.3132", "0.9126", "0.327", "0.5186", "0.327" ], [ "0.10297", "0.33", "0.4805", "0.33", "0.4805", "0.3435", "0.10297", "0.3435" ], [ "0.5195", "0.3289", "0.884", "0.3289", "0.884", "0.339", "0.5195", "0.339" ], [ "0.0866", "0.3435", "0.482", "0.3435", "0.482", "0.3567", "0.0866", "0.3567" ], [ "0.535", "0.3435", "0.9087", "0.3435", "0.9087", "0.3542", "0.535", "0.3542" ], [ "0.08826", "0.3591", "0.3677", "0.3591", "0.3677", "0.3694", "0.08826", "0.3694" ], [ "0.5186", "0.356", "0.9126", "0.356", "0.9126", "0.3694", "0.5186", "0.3694" ], [ "0.517", "0.3706", "0.9126", "0.3718", "0.9126", "0.3865", "0.517", "0.385" ], [ "0.10455", "0.377", "0.4763", "0.3757", "0.4763", "0.3933", "0.10455", "0.3945" ], [ "0.4714", "0.3818", "0.4763", "0.3818", "0.4763", "0.387", "0.4714", "0.387" ], [ "0.5186", "0.3865", "0.839", "0.3865", "0.839", "0.3997", "0.5186", "0.3997" ], [ "0.08496", "0.3945", "0.4795", "0.3945", "0.4795", "0.4077", "0.08496", "0.4077" ], [ "0.0866", "0.4092", "0.4805", "0.4092", "0.4805", "0.4224", "0.0866", "0.4224" ], [ "0.603", "0.4092", "0.8203", "0.4092", "0.8203", "0.43", "0.603", "0.43" ], [ "0.08496", "0.4224", "0.4436", "0.4224", "0.4436", "0.4363", "0.08496", "0.4363" ], [ "0.6006", "0.435", "0.8267", "0.435", "0.8267", "0.4526", "0.6006", "0.4526" ], [ "0.1993", "0.4426", "0.3015", "0.4426", "0.3015", "0.4539", "0.1993", "0.4539" ], [ "0.886", "0.4395", "0.9126", "0.4395", "0.9126", "0.4526", "0.886", "0.4526" ], [ "0.152", "0.4495", "0.1814", "0.4495", "0.1814", "0.461", "0.152", "0.461" ], [ "0.339", "0.4507", "0.4167", "0.4507", "0.4167", "0.461", "0.339", "0.461" ], [ "0.4617", "0.449", "0.482", "0.449", "0.482", "0.4622", "0.4617", "0.4622" ], [ "0.1772", "0.455", "0.3218", "0.4558", "0.3218", "0.4702", "0.1772", "0.4697" ], [ "0.5186", "0.4653", "0.911", "0.4653", "0.911", "0.4797", "0.5186", "0.4797" ], [ "0.0866", "0.476", "0.4805", "0.476", "0.4805", "0.4893", "0.0866", "0.4893" ], [ "0.08417", "0.486", "0.482", "0.486", "0.482", "0.507", "0.08417", "0.507" ], [ "0.5186", "0.481", "0.546", "0.481", "0.546", "0.4924", "0.5186", "0.4924" ], [ "0.5337", "0.4937", "0.911", "0.4937", "0.911", "0.507", "0.5337", "0.507" ], [ "0.0866", "0.506", "0.4805", "0.506", "0.4805", "0.5195", "0.0866", "0.5195" ], [ "0.5186", "0.509", "0.834", "0.509", "0.834", "0.5195", "0.5186", "0.5195" ], [ "0.08826", "0.5215", "0.1666", "0.5215", "0.1666", "0.5327", "0.08826", "0.5327" ], [ "0.562", "0.531", "0.7817", "0.532", "0.7817", "0.547", "0.562", "0.5454" ], [ "0.886", "0.532", "0.915", "0.532", "0.915", "0.547", "0.886", "0.547" ], [ "0.152", "0.5386", "0.415", "0.5405", "0.415", "0.5547", "0.152", "0.553" ], [ "0.433", "0.5454", "0.4453", "0.5454", "0.4453", "0.552", "0.433", "0.552" ], [ "0.4617", "0.5405", "0.482", "0.5405", "0.482", "0.5537", "0.4617", "0.5537" ], [ "0.6284", "0.549", "0.738", "0.55", "0.738", "0.5645", "0.6284", "0.563" ], [ "0.1953", "0.5547", "0.2051", "0.5547", "0.2051", "0.5625", "0.1953", "0.5625" ], [ "0.5596", "0.56", "0.63", "0.56", "0.63", "0.5713", "0.5596", "0.5713" ], [ "0.08496", "0.5684", "0.4805", "0.5684", "0.4805", "0.582", "0.08496", "0.582" ], [ "0.6763", "0.5684", "0.689", "0.5684", "0.689", "0.579", "0.6763", "0.579" ], [ "0.0866", "0.5835", "0.4795", "0.5835", "0.4795", "0.5967", "0.0866", "0.5967" ], [ "0.5605", "0.5835", "0.869", "0.5835", "0.869", "0.601", "0.5605", "0.601" ], [ "0.0866", "0.5967", "0.4805", "0.5967", "0.4805", "0.61", "0.0866", "0.61" ], [ "0.08826", "0.6123", "0.2239", "0.6123", "0.2239", "0.6226", "0.08826", "0.6226" ], [ "0.517", "0.6113", "0.911", "0.6123", "0.911", "0.627", "0.517", "0.6255" ], [ "0.232", "0.63", "0.3342", "0.63", "0.3342", "0.6445", "0.232", "0.6445" ], [ "0.5195", "0.6255", "0.91", "0.6255", "0.91", "0.6396", "0.5195", "0.6396" ], [ "0.4631", "0.632", "0.482", "0.632", "0.482", "0.646", "0.4631", "0.646" ], [ "0.5186", "0.6396", "0.911", "0.6396", "0.911", "0.653", "0.5186", "0.653" ], [ "0.08417", "0.6514", "0.411", "0.6514", "0.411", "0.6655", "0.08417", "0.6655" ], [ "0.5186", "0.654", "0.914", "0.654", "0.914", "0.6675", "0.5186", "0.6675" ], [ "0.1805", "0.6704", "0.384", "0.6704", "0.384", "0.685", "0.1805", "0.685" ], [ "0.455", "0.672", "0.4836", "0.672", "0.4836", "0.688", "0.455", "0.688" ], [ "0.5146", "0.6685", "0.7295", "0.6675", "0.7295", "0.681", "0.5146", "0.6816" ], [ "0.0866", "0.696", "0.2296", "0.696", "0.2296", "0.709", "0.0866", "0.709" ], [ "0.5186", "0.6978", "0.562", "0.6978", "0.562", "0.709", "0.5186", "0.709" ], [ "0.558", "0.6963", "0.88", "0.6963", "0.88", "0.7104", "0.558", "0.7104" ], [ "0.10297", "0.709", "0.4795", "0.709", "0.4795", "0.7227", "0.10297", "0.7227" ], [ "0.556", "0.711", "0.664", "0.7134", "0.664", "0.728", "0.556", "0.726" ], [ "0.08496", "0.7236", "0.4795", "0.7227", "0.4795", "0.7373", "0.08496", "0.738" ], [ "0.5186", "0.731", "0.9126", "0.731", "0.9126", "0.7456", "0.5186", "0.7456" ], [ "0.0866", "0.738", "0.4805", "0.738", "0.4805", "0.752", "0.0866", "0.752" ], [ "0.517", "0.7456", "0.9126", "0.7456", "0.9126", "0.759", "0.517", "0.759" ], [ "0.08496", "0.752", "0.4805", "0.7524", "0.4805", "0.7666", "0.08496", "0.765" ], [ "0.5195", "0.7603", "0.911", "0.7603", "0.911", "0.771", "0.5195", "0.771" ], [ "0.08496", "0.7676", "0.407", "0.7666", "0.407", "0.781", "0.08496", "0.782" ], [ "0.5156", "0.7715", "0.9126", "0.771", "0.9126", "0.785", "0.5156", "0.786" ], [ "0.2507", "0.7886", "0.4004", "0.7886", "0.4004", "0.803", "0.2507", "0.803" ], [ "0.5186", "0.786", "0.914", "0.786", "0.914", "0.8", "0.5186", "0.8" ], [ "0.4558", "0.799", "0.482", "0.799", "0.482", "0.811", "0.4558", "0.811" ], [ "0.1274", "0.789", "0.4084", "0.8003", "0.407", "0.829", "0.1259", "0.8174" ], [ "0.5195", "0.802", "0.911", "0.802", "0.911", "0.8125", "0.5195", "0.8125" ], [ "0.517", "0.8145", "0.9126", "0.8145", "0.9126", "0.8276", "0.517", "0.8276" ], [ "0.1323", "0.8257", "0.2239", "0.8257", "0.2239", "0.837", "0.1323", "0.837" ], [ "0.2141", "0.821", "0.4355", "0.8228", "0.4355", "0.8403", "0.2141", "0.8384" ], [ "0.5195", "0.829", "0.911", "0.829", "0.911", "0.839", "0.5195", "0.839" ], [ "0.2173", "0.837", "0.259", "0.839", "0.2573", "0.8516", "0.2157", "0.8496" ], [ "0.3179", "0.8354", "0.367", "0.8384", "0.3635", "0.854", "0.3145", "0.8496" ], [ "0.5186", "0.8433", "0.911", "0.8433", "0.911", "0.854", "0.5186", "0.854" ], [ "0.08496", "0.8613", "0.4805", "0.863", "0.4805", "0.8774", "0.08496", "0.8755" ], [ "0.517", "0.856", "0.9126", "0.856", "0.9126", "0.869", "0.517", "0.869" ], [ "0.517", "0.869", "0.914", "0.869", "0.914", "0.8833", "0.517", "0.8833" ], [ "0.08826", "0.8755", "0.4805", "0.8755", "0.4805", "0.886", "0.08826", "0.886" ], [ "0.08496", "0.8833", "0.406", "0.884", "0.406", "0.8984", "0.08496", "0.898" ], [ "0.517", "0.884", "0.687", "0.884", "0.687", "0.898", "0.517", "0.898" ] ]

[ "KDD'23, August 610, 2023, Long Beach, CA, USA", "识别为硬负样本。对于阳性样品，ha(xn）通过从Gefr(uin，Ujn）", "使用消息传递机制，我们将聚合不确定性mt（u）存储在u中，", "中减去直接导体G(uin,ujn）来表示其反事实硬度得分，以消除", "然后通过应用上述规则来更新不确定性Bt+1", "标签给出的先验信息。硬度分数用于计算方程中对数赔率校", "正的反事实NSR。（3）。我们不丢弃阳性样本", "Scorepropagation.我们可以通过送代", "zt+1=(1-)2+yszt, z=2", "Implementation.直接计算有效电导非常耗时。相反，我们首", "(12)", "先使用科学计算工具F[30]通过随机游走来近似通勤时间距离", "comm(u,o）。然后，我们使用变换Gefr(u,o）=2|El/comm(u,u)", "直到收敛来直接传播分数2，而不是传播不确定性。在获得最", "E[3]将通勤时间转换为有效电导", "终硬度分数后，我们重新缩放它们以匹配平均子采样率α。我", "们发现平滑硬度分数有利于模型不可知和基于模型的方法。", "3.2", "Smoothing Hardness Scores through Edge", "3.3", "Combining Model-agnostic and", "Model-basedMethods", "Propagation", "从稀疏图得出的有效电导有噪声，导致硬度分数「的估计不准", "确。我们使用图传播的思想来平滑硬度分数。虽然现有的机器", "我们已经在前面的章节中描述了与模型无关的数据子采样方", "学习文献侧重于传播节点属性(例如。g。，节点特征或标签)，", "法。一些硬实例可能会被模型不可知的方法忽略，而它们可以", "以平滑节点级不确定性18，19】，传播边缘属性未被充分开发", "被基于模型的方法捕获。因此，在本节中，我们将展示如何结", "我们通过将用户项二分图（U,V,E）转换为其相应的线图", "合这两种方法来结合两者的优点，并获得更好的采样性能", "L(U,V,E)=（VL,EL），将边缘传播减少为节点传播，其中VL=E", "给定样本x，模型不可知和基于模型的子采样方法分别计", "和EL是共享同一节点的边缘对的集合E[14]", "算其相应的采样率元D(x)和元(x）。特别地，元(x）是使用导", "频模型hb(）：=f(Φ）的x的子采样率，而元(x）是使用Eqn", "Uncertainty propagation.将Gefr := (Gefr(uin, Ujn)N-, E RN", "", "中的模型不可知硬度分数ha（)的子采样率。E(6)：", "Y;=(yn)n=1E(0,1)N分别表示为有效电导得分和边缘标记的", "向量。与E[19]类似，我们将有效电导Gefr归一化为估计分数", "g(x) α max(min(Pghb(x),es) 1),", "Z，并将不确定性分数B计算为Z和Y之间的绝对残差", "nD(x) α max(min(pDha(x),eD), 1)", "Geff - min(Geff)", "(13)", "Z=", "B=|Y-Z.", "(7)", "max(Gefr)-min(Gefr)", "(eQD）是最小采样率，而（PPD）被调整为满足平均子采样", "在这里，我们使用最小-最大归一化将硬度分数限制在范围", "THXX7内。表示 S=D-1/2ALD-1/2，其中 AL,DL分别是线图", "率α", "我们提出了三种简单但有效的启发式策略来结合这两种方", "的邻接矩阵和度矩阵。我们通过解决以下优化问题[41]来平", "法并获得最终的采样率：最大值，平均值和乘积。", "滑不确定性：", "元max(x)= Pmaxmax(元D(x),元g(x);", "(14)", "B = argmin tr(wT (I - S)W)+ μlW - Bll?.", "", "(8)", "(元(x) + 元g(x))", "W", "元mean(x)=", "Eqn中的第一项。（8）限制了相邻节点不确定性的差异。第二", "2", "项将平滑后的不确定性约束为接近初始不确定性，系数μ控制", "元prod(x) =min(max(PprodD(x)元g(x), eprod), 1)", "约束的强度。使用平滑的不确定性向量B，我们通过反转Eqn", "来校正硬度估计。（7）", "注意eprod是应用产品组合时的额外超参数。并对（pmax，Pprod）", "Z - Y+(-1)'B.", "进行了调整，将平均采样率归一化为α。在子采样率组合之", "(9)", "后，每个x将以公式中的概率进行采样。（14）。所有采样实例", "[18,41]也引入了迭代逼近方法。令=1/(1+p）和", "都将遵循正常的训练协议，以优化训练目标，如Eqn所示。（3", "Bt+1 = (1 -x)B+ySBt, BO = B", "(10)", "），保证最终结果得到很好的校准", "然后B→B当t→8", "3.4", "Theoretical bottlenecks of model-based", "但是，这种送代方法不可扩展，因为转换后的线图总共具有", "subsampling", "[EL/=(ZuieuDeg(ui)+ZujevDeg(uj)2)/2-N边,其中Deg()", "Above, we propose a principled approach to determine the sam-", "表示节点度。或者，我们可以直接沿着原始图（U,V,E）传播", "ple's hardness by analyzing graph structures within the data with", "边缘不确定性，该原始图总共仅包含El=（Zu,euDeg(ui）+", "out referring to a pilot model. We advocate this model-agnostic", "ZujevDeg(uj)/2个边缘。边缘上的传播规则如下", " sampling method due to the dilemma of unavoidable model mis-", "m(uin)+m(jn)-2Zn", "specification in the online-offline discrepancy setup. The critical as-", "(11)", "B=(B+eg(u,)+eg(o)", "sumption ofa correctly specifiedpilot modelis needed to derivea", "theoretically efficient data subsampling method. Under pilot model", "where m' (u) =", "= Bh, m* (o) = Bh, B% = Bn.", "misspecifications,itiseven non-trivial to guaranteemodel consis", "n: u=uir", "n:D=0jn", "tency (thesubsampled estimator has thesame limitas thefull data", "Isince negative edges are ignored to calculate effective conductance.According to", "estimator) for the log odds correction estimator (Eqn.3) [9,13,28].", "Therefore, pilot model misspecification is a realistic but underex-", "Eqn. (6 ), any isolated positive edges without a multi-hop connection between its", "endpoints will have zero hardness score without graph propagation.", "plored problem in statistics." ]

{ "arxiv_id": "2305.16391", "language": "zh", "page_idx": 4 }

[ [ "0.0989", "0.1067", "0.4805", "0.1067", "0.4805", "0.1206", "0.0989", "0.1206" ], [ "0.0989", "0.1218", "0.4836", "0.1218", "0.4836", "0.1351", "0.0989", "0.1351" ], [ "0.10046", "0.137", "0.482", "0.137", "0.482", "0.1477", "0.10046", "0.1477" ], [ "0.759", "0.137", "0.7974", "0.137", "0.7974", "0.1509", "0.759", "0.1509" ], [ "0.0989", "0.1497", "0.482", "0.1497", "0.482", "0.1635", "0.0989", "0.1635" ], [ "0.0972", "0.1622", "0.4805", "0.1635", "0.4805", "0.1768", "0.0972", "0.1755" ], [ "0.0972", "0.1755", "0.4805", "0.1768", "0.4805", "0.1901", "0.0972", "0.1895" ], [ "0.10046", "0.1913", "0.4836", "0.1913", "0.4836", "0.2052", "0.10046", "0.2052" ], [ "0.0989", "0.2052", "0.482", "0.2052", "0.482", "0.2184", "0.0989", "0.2184" ], [ "0.0989", "0.2184", "0.482", "0.2184", "0.482", "0.2317", "0.0989", "0.2317" ], [ "0.5186", "0.2203", "0.914", "0.2203", "0.914", "0.2343", "0.5186", "0.2343" ], [ "0.10046", "0.2329", "0.482", "0.2329", "0.482", "0.2468", "0.10046", "0.2468" ], [ "0.517", "0.2343", "0.7256", "0.2343", "0.7256", "0.2474", "0.517", "0.2474" ], [ "0.1013", "0.2474", "0.482", "0.2474", "0.482", "0.2583", "0.1013", "0.2583" ], [ "0.10046", "0.26", "0.482", "0.26", "0.482", "0.2734", "0.10046", "0.2734" ], [ "0.5317", "0.2583", "0.9126", "0.2583", "0.9126", "0.2683", "0.5317", "0.2683" ], [ "0.10046", "0.276", "0.4805", "0.276", "0.4805", "0.286", "0.10046", "0.286" ], [ "0.5293", "0.2708", "0.914", "0.2683", "0.914", "0.2827", "0.5293", "0.2854" ], [ "0.10046", "0.289", "0.1438", "0.289", "0.1438", "0.2998", "0.10046", "0.2998" ], [ "0.5317", "0.286", "0.9126", "0.286", "0.9126", "0.2966", "0.5317", "0.2966" ], [ "0.08496", "0.3018", "0.4805", "0.3018", "0.4805", "0.315", "0.08496", "0.315" ], [ "0.5312", "0.2974", "0.915", "0.2974", "0.915", "0.3113", "0.5312", "0.3113" ], [ "0.0989", "0.315", "0.4805", "0.3145", "0.4805", "0.3289", "0.0989", "0.33" ], [ "0.5312", "0.3132", "0.914", "0.3132", "0.914", "0.327", "0.5312", "0.327" ], [ "0.10046", "0.3308", "0.4805", "0.3308", "0.4805", "0.3416", "0.10046", "0.3416" ], [ "0.5312", "0.3257", "0.914", "0.3257", "0.914", "0.339", "0.5312", "0.339" ], [ "0.0989", "0.3435", "0.4805", "0.3435", "0.4805", "0.3567", "0.0989", "0.3567" ], [ "0.5312", "0.3403", "0.914", "0.3403", "0.914", "0.3542", "0.5312", "0.3542" ], [ "0.5293", "0.3528", "0.9126", "0.3542", "0.9126", "0.3674", "0.5293", "0.3662" ], [ "0.10455", "0.3591", "0.4795", "0.3591", "0.4795", "0.3694", "0.10455", "0.3694" ], [ "0.1013", "0.3726", "0.4805", "0.3726", "0.4805", "0.3833", "0.1013", "0.3833" ], [ "0.5312", "0.3674", "0.7417", "0.3674", "0.7417", "0.3806", "0.5312", "0.3806" ], [ "0.0989", "0.385", "0.482", "0.385", "0.482", "0.3984", "0.0989", "0.3984" ], [ "0.535", "0.3838", "0.914", "0.3838", "0.914", "0.3977", "0.535", "0.3977" ], [ "0.1013", "0.401", "0.4805", "0.401", "0.4805", "0.411", "0.1013", "0.411" ], [ "0.5186", "0.3965", "0.9126", "0.3984", "0.9126", "0.4124", "0.5186", "0.4104" ], [ "0.1013", "0.414", "0.4805", "0.414", "0.4805", "0.4248", "0.1013", "0.4248" ], [ "0.523", "0.4136", "0.9126", "0.4136", "0.9126", "0.4236", "0.523", "0.4236" ], [ "0.10046", "0.4255", "0.4836", "0.4255", "0.4836", "0.4395", "0.10046", "0.4395" ], [ "0.5186", "0.4255", "0.9126", "0.4255", "0.9126", "0.4395", "0.5186", "0.4395" ], [ "0.0989", "0.4395", "0.4805", "0.44", "0.4805", "0.4539", "0.0989", "0.4526" ], [ "0.517", "0.4382", "0.9126", "0.4395", "0.9126", "0.4539", "0.517", "0.4526" ], [ "0.1013", "0.4558", "0.4795", "0.4558", "0.4795", "0.4666", "0.1013", "0.4666" ], [ "0.5186", "0.4539", "0.914", "0.4539", "0.914", "0.4673", "0.5186", "0.4673" ], [ "0.10046", "0.4673", "0.4805", "0.4673", "0.4805", "0.481", "0.10046", "0.481" ], [ "0.5186", "0.4673", "0.785", "0.4673", "0.785", "0.481", "0.5186", "0.481" ], [ "0.1013", "0.483", "0.4795", "0.483", "0.4795", "0.4937", "0.1013", "0.4937" ], [ "0.5337", "0.4817", "0.689", "0.4817", "0.689", "0.4956", "0.5337", "0.4956" ], [ "0.10046", "0.4944", "0.4805", "0.4956", "0.4805", "0.509", "0.10046", "0.5083" ], [ "0.517", "0.4988", "0.914", "0.4988", "0.914", "0.512", "0.517", "0.512" ], [ "0.0989", "0.509", "0.4805", "0.509", "0.4805", "0.523", "0.0989", "0.523" ], [ "0.5312", "0.5146", "0.9126", "0.5146", "0.9126", "0.5244", "0.5312", "0.5244" ], [ "0.10046", "0.5244", "0.468", "0.5244", "0.468", "0.535", "0.10046", "0.535" ], [ "0.5317", "0.528", "0.7295", "0.528", "0.7295", "0.5386", "0.5317", "0.5386" ], [ "0.517", "0.539", "0.914", "0.539", "0.914", "0.553", "0.517", "0.553" ], [ "0.10297", "0.5503", "0.4805", "0.5503", "0.4805", "0.5645", "0.10297", "0.5645" ], [ "0.5317", "0.5547", "0.9126", "0.5547", "0.9126", "0.565", "0.5317", "0.565" ], [ "0.0866", "0.5645", "0.4805", "0.5645", "0.4805", "0.5776", "0.0866", "0.5776" ], [ "0.5312", "0.5674", "0.914", "0.5674", "0.914", "0.5806", "0.5312", "0.5806" ], [ "0.08905", "0.58", "0.4805", "0.58", "0.4805", "0.5903", "0.08905", "0.5903" ], [ "0.5317", "0.5835", "0.9126", "0.5835", "0.9126", "0.5933", "0.5317", "0.5933" ], [ "0.08905", "0.5933", "0.4795", "0.5933", "0.4795", "0.604", "0.08905", "0.604" ], [ "0.5293", "0.5947", "0.886", "0.595", "0.886", "0.6094", "0.5293", "0.608" ], [ "0.08826", "0.606", "0.4805", "0.606", "0.4805", "0.619", "0.08826", "0.619" ], [ "0.517", "0.6094", "0.9126", "0.6094", "0.9126", "0.6226", "0.517", "0.6226" ], [ "0.0866", "0.619", "0.482", "0.619", "0.482", "0.6333", "0.0866", "0.6333" ], [ "0.528", "0.6226", "0.9126", "0.622", "0.9126", "0.636", "0.528", "0.637" ], [ "0.08826", "0.6353", "0.478", "0.6353", "0.478", "0.646", "0.08826", "0.646" ], [ "0.5317", "0.638", "0.911", "0.638", "0.911", "0.6484", "0.5317", "0.6484" ], [ "0.08826", "0.648", "0.4805", "0.648", "0.4805", "0.661", "0.08826", "0.661" ], [ "0.5312", "0.651", "0.914", "0.651", "0.914", "0.664", "0.5312", "0.664" ], [ "0.08826", "0.661", "0.4805", "0.661", "0.4805", "0.675", "0.08826", "0.675" ], [ "0.5312", "0.6665", "0.884", "0.6665", "0.884", "0.677", "0.5312", "0.677" ], [ "0.0866", "0.677", "0.4795", "0.677", "0.4795", "0.6875", "0.0866", "0.6875" ], [ "0.5186", "0.6797", "0.9126", "0.6797", "0.9126", "0.69", "0.5186", "0.69" ], [ "0.0972", "0.688", "0.478", "0.688", "0.478", "0.702", "0.0972", "0.702" ], [ "0.5293", "0.6924", "0.831", "0.6924", "0.831", "0.7056", "0.5293", "0.7056" ], [ "0.0866", "0.7026", "0.4805", "0.7026", "0.4805", "0.7163", "0.0866", "0.7163" ], [ "0.08496", "0.7173", "0.2214", "0.7173", "0.2214", "0.731", "0.08496", "0.731" ], [ "0.5195", "0.7217", "0.5337", "0.7217", "0.5337", "0.73", "0.5195", "0.73" ], [ "0.536", "0.7227", "0.5435", "0.7227", "0.5435", "0.728", "0.536", "0.728" ], [ "0.5474", "0.7197", "0.7715", "0.7197", "0.7715", "0.73", "0.5474", "0.73" ], [ "0.5186", "0.7373", "0.9126", "0.7373", "0.9126", "0.7505", "0.5186", "0.7505" ], [ "0.08496", "0.7437", "0.4795", "0.7437", "0.4795", "0.757", "0.08496", "0.757" ], [ "0.5195", "0.752", "0.591", "0.752", "0.591", "0.763", "0.5195", "0.763" ], [ "0.0989", "0.7583", "0.482", "0.7583", "0.482", "0.7715", "0.0989", "0.7715" ], [ "0.0989", "0.7715", "0.482", "0.7715", "0.482", "0.785", "0.0989", "0.785" ], [ "0.5195", "0.7676", "0.911", "0.7676", "0.911", "0.781", "0.5195", "0.781" ], [ "0.0989", "0.786", "0.482", "0.786", "0.482", "0.8", "0.0989", "0.8" ], [ "0.5195", "0.782", "0.9126", "0.782", "0.9126", "0.7954", "0.5195", "0.7954" ], [ "0.0989", "0.8", "0.482", "0.8", "0.482", "0.813", "0.0989", "0.813" ], [ "0.5195", "0.7954", "0.9126", "0.7954", "0.9126", "0.8096", "0.5195", "0.8096" ], [ "0.0989", "0.813", "0.482", "0.813", "0.482", "0.827", "0.0989", "0.827" ], [ "0.5186", "0.81", "0.914", "0.81", "0.914", "0.8237", "0.5186", "0.8237" ], [ "0.1013", "0.829", "0.4795", "0.829", "0.4795", "0.839", "0.1013", "0.839" ], [ "0.5195", "0.8257", "0.738", "0.8257", "0.738", "0.836", "0.5195", "0.836" ], [ "0.0989", "0.8413", "0.482", "0.8413", "0.482", "0.855", "0.0989", "0.855" ], [ "0.5195", "0.8413", "0.9126", "0.8413", "0.9126", "0.855", "0.5195", "0.855" ], [ "0.10046", "0.857", "0.4478", "0.857", "0.4478", "0.867", "0.10046", "0.867" ], [ "0.5195", "0.855", "0.9126", "0.855", "0.9126", "0.8687", "0.5195", "0.8687" ], [ "0.08417", "0.8687", "0.4836", "0.867", "0.4836", "0.882", "0.08417", "0.8833" ], [ "0.5186", "0.869", "0.9126", "0.869", "0.9126", "0.8833", "0.5186", "0.8833" ], [ "0.0989", "0.884", "0.4836", "0.884", "0.4836", "0.898", "0.0989", "0.898" ], [ "0.517", "0.8833", "0.9126", "0.8833", "0.9126", "0.898", "0.517", "0.898" ] ]

[ "when embedding hierarchical data [7, 12, 34] . Moreover, none", "of them considers KG relations at a finer-grained level of hierar-", "chies.They all conveniently ignore an importantfact, i.e.,hierar-", "Letei", "chical relations (blue lines in Figure 1(a) ) and non-hierarchical", "relations (black lines in Figure 1(a))are not of equal importance", "for characterizing items. Taking the outfit recommendation in", "Figure 1(a) as an example. Hierarchical relation ri offers com-", "plementary information to the cloth ii in the aspect that ii is", "the jeans ei while non-hierarchical relation r5 means that cloth", "Figure 2: (a) Hyperbolic space expands exponentially. (b) Il-", "i2 and cloth i3 can compose an outfit. Thus, hierarchical relations", "lustration of Hyperbolic cones.", "can indicate the attributes of items while non-hierarchical rela-", "tions only reveal the relatedness between entities. Ignoring the", "andKGastwo different information channels,and developdif", "hierarchical structures and relations limits model's expressive", "ferent aggregation strategies for the two channels. Besides, since", "power.", "items can serve as a natural bridge to connect the two informa-", ". High-order Collaborative Signals of Items. In existing studies,", "tion channels, we use dual embedding instead of single embed-", "the item aggregation schemes are mostly KG-oriented, that is.", "ding for items to represent and propagate each channel's infor-", "recursively collecting itemsknowledge associationsfromKG", "mation separately, so as to better capture the holistic semantics", "without considering the collaborative signals from users [47, 49]", "of items. Moreover, an information gated mechanism is intro-", "duced to adaptively fuse the two types of semantics for better", "orblindlymixingtheheterogeneousinformationfromneigh", "boringnodes (users orentities)in theUnifiedKnowledgeGraph", "identifying user behavior patterns.", "(UKG) [37, 45] . They fail to explicitly encode the crucial high-", "为此，新提出的HAKG框架旨在i）利用双曲线几何的表现", "order collaborativesignals ofitems,which arelatent inuser-", "力进行更好的分层建模，以及i）有效地捕获项目的整体语", "item interactions and play an important role in learning user", "义。我们在三个真实的数据集上进行了广泛的实验，以评估", "preference from behavior aspect. Take Figure 1(a) as an exam-", "HAKG和现有方法的性能。实验结果表明，我们的HAKG显著", "ple.The path ii-→ u1→ is→ u2-→ i2 indicates the long-range", "优于所有最先进的方法，如CKAN[49]、Hyper-Know[31]和", "connectivitybetween item ii and itemi2:since users uiand u2", "KGIN[47】。此外，HAKG能够揭示嵌人空间中数据的底层层", "share the same interest ( ie., they both like item is ), the items i", "次结构和关系，以获得更好的模型表现力", "and i2which arefavored by user ui and useru2 respectivelymay", "总之，我们的贡献如下：", "be similar to some extent.Therefore, such KG-oriented aggrega", ". We present knowledge-aware recommendation from a new per-", "tion schemes are insuficient to explicitly capture the high-order", "spective bytaking thehierarchy and high-order itemscollabo-", "collaborative signalsfor comprehensive item representations.", "rative signals into consideration.", ". We embed users and items, as well as entities and relations in", "为了解决上述挑战，需要具有较小嵌入失真的新潜在空间。", "hyperbolic space,and design a new hyperbolicaggregationin", "双曲几何提供了一种理想的替代方法，因为它不仅可以实现", "KG to preserve the relation dependencies among neighbors. Be-", "失真小得多的嵌入E[7.121，而且可以自然地保留数据层次结", "sides,an angleconstraintis introducedforhierarchicalrelations", "构E[6，54]。双曲空间的关键性质是，与欧几里得空间不同，", "to profile items with the attributes in the embedding space.", "它是指数扩展的，而不是多项式扩展的。这使得双曲线空间比", ".We devise the dual item embeddings design to i) better represent", "欧几里得空间具有更大的数据容量，因此，它有“足够的”空", "and propagate collaborative signals and knowledge associations", "间来更好地保持分层数据之间的距离，如图E2的左侧所示", "simultaneously,andi)moreeffectivelyleveragetheinformation", "此外，双曲线空间可以被视为树的连续版本，由于其类似的结", "gate mechanism to control discriminative signals towards the", "构，这使得它非常适合对分层树状数据进行建模。因此，我们", "userspreferencefrombothbehaviorand attributeaspects", "提出了Hi-AwareKnowledgeGatedNetwork（HAKG)：这是", ".We conduct extensiveexperiments on three public benchmark", "一个具有两个关键组件的新的双曲线知识感知模型，它可以", "datasets to demonstrate the superiority of HAKG.", "有效地捕获和建模上述两个重要因素，而这些因素是任何现", "有方法都无法建模的", "2", "", "PROBLEMFORMULATION", "我们首先介绍与我们研究的问题相关的数据结构，然后制定", "·Hierarchy-Aware Modeling.To better capture the underlying hi", "我们的任务", "erarchical structures, we map user and item embeddings as well", "as entity and relation embeddings to hyperbolic space. We also", "用户-项目二分图。在本文中，我们专注于从隐式反馈中学习", "design a new hyperbolic relation-transitive aggregation mecha-", "用户偏好[35]。更具体地说，行为数据（e.g，单击并查看）涉", "nism to capture the relation dependencies carried by neighbors", "及一组用户=（u)和一组项目I={i)。我们将用户-项目交", "in hyperbolic space. Moreover, a novel angle constraint is intro-", "互建模为二分图Gb=((u,i)luEu,ieI)，其中每个(u,i)对", "duced forhierarchicalrelations,which is ableto betterpreserve", "表示用户u已与项自i进行了交互", "items' attributes information in the embedding space and thus", "知识图谱(KG)。KG是真实世界事实的集合，例如项目属性、", "improvetherepresentational capacityandexpressiveness.", "概念或外部常识。令T为三元组集，8为实体集，R为关系", ". Gated Aggregation with Dual Embeddings. Unlike previous KG-", "集，它涉及规范和逆方向上的关系（e.g组成和组成）。令", "oriented aggregation strategies, we view user-item interactions", "KG为异质图Gk=((h,r,t)/h,tE8,rER)，其中每个三元组" ]

{ "arxiv_id": "2204.04959", "language": "zh", "page_idx": 1 }

[ [ "0.08826", "0.108", "0.4836", "0.108", "0.4836", "0.1218", "0.08826", "0.1218" ], [ "0.5186", "0.108", "0.9126", "0.108", "0.9126", "0.1218", "0.5186", "0.1218" ], [ "0.08826", "0.1218", "0.4805", "0.1218", "0.4805", "0.1351", "0.08826", "0.1351" ], [ "0.5186", "0.1218", "0.697", "0.1218", "0.697", "0.1351", "0.5186", "0.1351" ], [ "0.0866", "0.1364", "0.4805", "0.1364", "0.4805", "0.1497", "0.0866", "0.1497" ], [ "0.5337", "0.1351", "0.9126", "0.1364", "0.9126", "0.1497", "0.5337", "0.1483" ], [ "0.0866", "0.1497", "0.4805", "0.1497", "0.4805", "0.1635", "0.0866", "0.1635" ], [ "0.5186", "0.1497", "0.9126", "0.1497", "0.9126", "0.1635", "0.5186", "0.1635" ], [ "0.08826", "0.1635", "0.482", "0.1635", "0.482", "0.1768", "0.08826", "0.1768" ], [ "0.5215", "0.1654", "0.911", "0.1654", "0.911", "0.1755", "0.5215", "0.1755" ], [ "0.08826", "0.178", "0.246", "0.178", "0.246", "0.1881", "0.08826", "0.1881" ], [ "0.5195", "0.178", "0.9126", "0.178", "0.9126", "0.1913", "0.5195", "0.1913" ], [ "0.0866", "0.1957", "0.4805", "0.1957", "0.4805", "0.209", "0.0866", "0.209" ], [ "0.5186", "0.1925", "0.9126", "0.1925", "0.9126", "0.2058", "0.5186", "0.2058" ], [ "0.0866", "0.209", "0.4558", "0.209", "0.4558", "0.2229", "0.0866", "0.2229" ], [ "0.5195", "0.207", "0.9004", "0.207", "0.9004", "0.2172", "0.5195", "0.2172" ], [ "0.5186", "0.2291", "0.9126", "0.2291", "0.9126", "0.2424", "0.5186", "0.2424" ], [ "0.1177", "0.2424", "0.2573", "0.2424", "0.2573", "0.2524", "0.1177", "0.2524" ], [ "0.5186", "0.2424", "0.9126", "0.2424", "0.9126", "0.2556", "0.5186", "0.2556" ], [ "0.5186", "0.2556", "0.9126", "0.2556", "0.9126", "0.2695", "0.5186", "0.2695" ], [ "0.08826", "0.2437", "0.1152", "0.2437", "0.1152", "0.252", "0.08826", "0.252" ], [ "0.08826", "0.26", "0.4795", "0.26", "0.4795", "0.2708", "0.08826", "0.2708" ], [ "0.0866", "0.2727", "0.4805", "0.2727", "0.4805", "0.286", "0.0866", "0.286" ], [ "0.5186", "0.2683", "0.9126", "0.2695", "0.9126", "0.2827", "0.5186", "0.2822" ], [ "0.5186", "0.2822", "0.9126", "0.2842", "0.9126", "0.2974", "0.5186", "0.2954" ], [ "0.08826", "0.2886", "0.4795", "0.2886", "0.4795", "0.2986", "0.08826", "0.2986" ], [ "0.0866", "0.3005", "0.4795", "0.3005", "0.4795", "0.3145", "0.0866", "0.3145" ], [ "0.5186", "0.2974", "0.9126", "0.2974", "0.9126", "0.3113", "0.5186", "0.3113" ], [ "0.08826", "0.3145", "0.4795", "0.315", "0.4795", "0.327", "0.08826", "0.3257" ], [ "0.523", "0.3113", "0.9126", "0.3113", "0.9126", "0.3245", "0.523", "0.3245" ], [ "0.0866", "0.3289", "0.4795", "0.3289", "0.4795", "0.3423", "0.0866", "0.3423" ], [ "0.5215", "0.327", "0.911", "0.327", "0.911", "0.3372", "0.5215", "0.3372" ], [ "0.08496", "0.3416", "0.482", "0.3423", "0.482", "0.356", "0.08496", "0.3547" ], [ "0.5186", "0.339", "0.9126", "0.339", "0.9126", "0.3528", "0.5186", "0.3528" ], [ "0.08496", "0.3547", "0.2646", "0.3567", "0.2646", "0.3706", "0.08496", "0.3687" ], [ "0.5186", "0.3516", "0.9126", "0.3516", "0.9126", "0.3655", "0.5186", "0.3655" ], [ "0.5195", "0.3674", "0.9126", "0.3674", "0.9126", "0.3806", "0.5195", "0.3806" ], [ "0.5186", "0.3806", "0.9126", "0.3806", "0.9126", "0.3945", "0.5186", "0.3945" ], [ "0.08496", "0.3882", "0.2427", "0.3882", "0.2427", "0.4016", "0.08496", "0.4016" ], [ "0.517", "0.3933", "0.834", "0.3933", "0.834", "0.4072", "0.517", "0.4072" ], [ "0.08905", "0.4077", "0.4795", "0.4077", "0.4795", "0.4185", "0.08905", "0.4185" ], [ "0.08826", "0.4204", "0.4805", "0.4204", "0.4805", "0.4343", "0.08826", "0.4343" ], [ "0.5645", "0.4153", "0.865", "0.4172", "0.865", "0.4463", "0.5645", "0.4443" ], [ "0.863", "0.4268", "0.8735", "0.4268", "0.8735", "0.4312", "0.863", "0.4312" ], [ "0.8945", "0.4248", "0.91", "0.4248", "0.91", "0.4343", "0.8945", "0.4343" ], [ "0.08826", "0.4343", "0.4795", "0.4343", "0.4795", "0.4475", "0.08826", "0.4475" ], [ "0.08826", "0.4475", "0.4805", "0.4475", "0.4805", "0.461", "0.08826", "0.461" ], [ "0.517", "0.4507", "0.911", "0.452", "0.911", "0.4653", "0.517", "0.464" ], [ "0.08905", "0.4634", "0.478", "0.4634", "0.478", "0.4734", "0.08905", "0.4734" ], [ "0.517", "0.4666", "0.9126", "0.4673", "0.9126", "0.4817", "0.517", "0.481" ], [ "0.0866", "0.476", "0.482", "0.476", "0.482", "0.4893", "0.0866", "0.4893" ], [ "0.0866", "0.4893", "0.4575", "0.4893", "0.4575", "0.5024", "0.0866", "0.5024" ], [ "0.517", "0.4841", "0.91", "0.481", "0.91", "0.4988", "0.517", "0.502" ], [ "0.08417", "0.5024", "0.4805", "0.502", "0.4805", "0.515", "0.08417", "0.5166" ], [ "0.5156", "0.4988", "0.914", "0.5", "0.914", "0.5176", "0.5156", "0.5166" ], [ "0.08826", "0.5195", "0.10297", "0.5195", "0.10297", "0.53", "0.08826", "0.53" ], [ "0.5215", "0.518", "0.911", "0.518", "0.911", "0.529", "0.5215", "0.529" ], [ "0.165", "0.536", "0.3938", "0.5327", "0.3938", "0.5537", "0.165", "0.5566" ], [ "0.5195", "0.532", "0.9126", "0.532", "0.9126", "0.5425", "0.5195", "0.5425" ], [ "0.4631", "0.539", "0.482", "0.539", "0.482", "0.552", "0.4631", "0.552" ], [ "0.517", "0.5444", "0.911", "0.5425", "0.911", "0.56", "0.517", "0.5625" ], [ "0.10046", "0.56", "0.482", "0.56", "0.482", "0.5737", "0.10046", "0.5737" ], [ "0.5195", "0.561", "0.914", "0.561", "0.914", "0.5747", "0.5195", "0.5747" ], [ "0.08905", "0.5757", "0.4617", "0.5757", "0.4617", "0.586", "0.08905", "0.586" ], [ "0.09644", "0.565", "0.10297", "0.565", "0.10297", "0.5693", "0.09644", "0.5693" ], [ "0.5215", "0.577", "0.9126", "0.577", "0.9126", "0.587", "0.5215", "0.587" ], [ "0.08826", "0.589", "0.4795", "0.589", "0.4795", "0.5996", "0.08826", "0.5996" ], [ "0.517", "0.5903", "0.789", "0.5884", "0.789", "0.603", "0.517", "0.605" ], [ "0.08826", "0.6016", "0.4805", "0.6016", "0.4805", "0.6157", "0.08826", "0.6157" ], [ "0.5337", "0.605", "0.9126", "0.605", "0.9126", "0.6187", "0.5337", "0.6187" ], [ "0.08905", "0.617", "0.4795", "0.617", "0.4795", "0.6274", "0.08905", "0.6274" ], [ "0.517", "0.6187", "0.9126", "0.6187", "0.9126", "0.632", "0.517", "0.632" ], [ "0.08826", "0.63", "0.4314", "0.63", "0.4314", "0.643", "0.08826", "0.643" ], [ "0.5195", "0.634", "0.5557", "0.634", "0.5557", "0.646", "0.5195", "0.646" ], [ "0.3096", "0.643", "0.3816", "0.646", "0.3816", "0.66", "0.3096", "0.6577" ], [ "0.179", "0.653", "0.3064", "0.653", "0.3064", "0.6665", "0.179", "0.6665" ], [ "0.4631", "0.653", "0.482", "0.653", "0.482", "0.6665", "0.4631", "0.6665" ], [ "0.556", "0.6514", "0.744", "0.662", "0.7417", "0.688", "0.5537", "0.678" ], [ "0.2974", "0.6577", "0.3865", "0.66", "0.3865", "0.675", "0.2974", "0.6724" ], [ "0.6562", "0.659", "0.6685", "0.661", "0.6665", "0.6655", "0.6543", "0.6636" ], [ "0.722", "0.659", "0.869", "0.657", "0.869", "0.674", "0.722", "0.6753" ], [ "0.8965", "0.662", "0.91", "0.662", "0.91", "0.672", "0.8965", "0.672" ], [ "0.1209", "0.6724", "0.4453", "0.681", "0.4436", "0.707", "0.11926", "0.6987" ], [ "0.6436", "0.66", "0.6562", "0.66", "0.6562", "0.664", "0.6436", "0.664" ], [ "0.3865", "0.678", "0.429", "0.678", "0.429", "0.688", "0.3865", "0.688" ], [ "0.4631", "0.683", "0.482", "0.683", "0.482", "0.696", "0.4631", "0.696" ], [ "0.5186", "0.6895", "0.911", "0.6895", "0.911", "0.7026", "0.5186", "0.7026" ], [ "0.0866", "0.7056", "0.3423", "0.7056", "0.3423", "0.7197", "0.0866", "0.7197" ], [ "0.5195", "0.705", "0.911", "0.705", "0.911", "0.7153", "0.5195", "0.7153" ], [ "0.1953", "0.7227", "0.4143", "0.7227", "0.4143", "0.7373", "0.1953", "0.7373" ], [ "0.5195", "0.7173", "0.911", "0.7173", "0.911", "0.731", "0.5195", "0.731" ], [ "0.1487", "0.731", "0.1993", "0.733", "0.1993", "0.744", "0.1487", "0.742" ], [ "0.464", "0.733", "0.4805", "0.733", "0.4805", "0.742", "0.464", "0.742" ], [ "0.5195", "0.733", "0.9126", "0.733", "0.9126", "0.7437", "0.5195", "0.7437" ], [ "0.2402", "0.7393", "0.3782", "0.7393", "0.3782", "0.7524", "0.2402", "0.7524" ], [ "0.5215", "0.747", "0.911", "0.747", "0.911", "0.757", "0.5215", "0.757" ], [ "0.5195", "0.7603", "0.911", "0.7603", "0.911", "0.771", "0.5195", "0.771" ], [ "0.08417", "0.7676", "0.353", "0.768", "0.353", "0.782", "0.08417", "0.781" ], [ "0.5195", "0.7734", "0.91", "0.7734", "0.91", "0.784", "0.5195", "0.784" ], [ "0.08905", "0.787", "0.4795", "0.787", "0.4795", "0.7974", "0.08905", "0.7974" ], [ "0.517", "0.785", "0.9126", "0.786", "0.9126", "0.8", "0.517", "0.799" ], [ "0.0866", "0.799", "0.4795", "0.8", "0.4795", "0.813", "0.0866", "0.8125" ], [ "0.5186", "0.8", "0.9126", "0.8", "0.9126", "0.813", "0.5186", "0.813" ], [ "0.08826", "0.815", "0.4795", "0.815", "0.4795", "0.8257", "0.08826", "0.8257" ], [ "0.5186", "0.815", "0.562", "0.815", "0.562", "0.827", "0.5186", "0.827" ], [ "0.08826", "0.829", "0.478", "0.829", "0.478", "0.839", "0.08826", "0.839" ], [ "0.5337", "0.827", "0.9126", "0.8276", "0.9126", "0.8413", "0.5337", "0.8403" ], [ "0.08826", "0.8433", "0.4805", "0.8433", "0.4805", "0.854", "0.08826", "0.854" ], [ "0.5215", "0.8433", "0.911", "0.8433", "0.911", "0.854", "0.5215", "0.854" ], [ "0.0866", "0.856", "0.4805", "0.856", "0.4805", "0.869", "0.0866", "0.869" ], [ "0.5186", "0.856", "0.9126", "0.856", "0.9126", "0.869", "0.5186", "0.869" ], [ "0.08496", "0.8687", "0.4805", "0.869", "0.4805", "0.8833", "0.08496", "0.882" ], [ "0.5186", "0.869", "0.9126", "0.869", "0.9126", "0.8833", "0.5186", "0.8833" ], [ "0.08826", "0.8853", "0.4795", "0.8853", "0.4795", "0.896", "0.08826", "0.896" ], [ "0.5195", "0.8853", "0.911", "0.8853", "0.911", "0.896", "0.5195", "0.896" ] ]

[ "(h,r,t)ET表示头实体h和尾实体t之间存在关系r。例如，", "在KG中设计了一个新的聚合方案，因为以前的基于欧几里得", "三胞胎（牛仔裤，品牌，Levi's）表示牛仔裤的品牌是Levi's。", "的方法在双曲空间中不可行", "因为我们假设所有项目在KG中都显示为实体(i。e。，IC8", "同时，我们认为以前的方法无法正确表征项目，因为它们没", "，所有现有的知识感知推荐系统E[44,45,47]所做的共同假", "有将层次关系与其余部分区分开来，并且仅以粗粒度对KG关", "设，我们可以将用户-项目图中的项目与KG中的实体链接起", "系进行建模。具体来说，KG中的层次关系通常携带实体的属", "来，为交互提供辅助语义", "性信息（e.g品牌和面料），而非层次关系只能指示实体之间", "任务说明。给定用户-项目图Gb和KGGk，我们的知识感知", "的相关性（e.g,匹配并类似于）。这促使我们在嵌人空间中为", "建议的任务是预测用户采用她从未参与过的项目的可能性", "分层三元组引人一个额外的约束，以提高表示能力和表现力", "3.2.1HyperbolicRelation-TransitiveAggregation.我们首先考", "METHODOLOGY", "虑双曲空间中KG的聚合方案。由于一个项目i可以涉及多个", "KG三元组，因此其邻域在一定程度上反映了i与其连接实体", "3", "在本节中，我们首先介绍有关双曲空间的基本概念，然后详细", "介绍HAKG的两个关键组件，i.e.层次感知建模和具有双嵌入", "之间的关系相似性。形式上，给定KG中的目标项i，聚合KG", "中i邻居的本地信息可以揭示i的相关知识关联。因此，我们", "的门控聚合。前者将用户和项目以及实体和关系嵌入双曲线", "空间。它在KG上采用了一种新颖的双曲线关系感知聚合，并", "使用Ni=((r,t)l(i,r,t)EGk)来表示邻域实体和KG中项i的", "且还引入了角度约束来学习嵌入空间中项目的属性语义。后", "一阶关系，并建议整合邻域实体的关系上下文以生成项i的知", "者利用双重项目嵌人分别对来自用户-项目图的高阶协作信号", "识表示：1。先前的研究E47]表明，基于节点的聚合无法区", "和来自KG的知识关联进行编码，并通过门机制提取有用的语", "分关系路径，从而无法保留路径携带的关系依赖性。因此，有", "义以学习高质量的用户表示", "必要将关系路径集成到表示中。此外，基于欧几里得传播的", "方法主要使用均值聚合来平均邻居信息。双曲空间中均值聚", "集的类似物是Frechet均值E[11]，但是它没有封闭形式的解。", "3.1Preliminaries", "相反，我们将均值聚合器中的关系上下文建模为：", "双曲线几何[3]是具有恒定负曲率的非欧几里得几何。与欧几", "里得空间相比，双曲几何覆盖的空间量呈指数增长，而不是多", "e(\"=expo ((r)en, g,(e( e)", "", "(5)", "项式w.r.t.半径。此属性使我们能够有效地捕获双曲线空间中", "用户-项目交互和KG的底层层次结构。在这项工作中，我们使", "其中④是Mobius加法运算，log(）和exp(）分别是上述指数", "用具有恒定曲率-c=-1的PoincarxMATHXSQe球模型来学", "映射和对数映射，e(o）是实体t在双曲空间中的ID嵌入。对", "习双曲线嵌人，因为它对梯度优化的可行性E[34,54]。下面，", "我们首先介绍PoincarxMATHXSQe球模型的必要数学基础", "于KG中的每个三元组（i,r,t），我们通过将关系r建模为e(0)", "Poincare球&切线空间。Poincare球模型的定义域为：Rn中", "的向量转换来设计关系转换e(0）er。为了避免双曲空间中", "的", "的复杂均值运算，我们首先将每个关系上下文映射到切线空", "B=(x.n):x+.+<)", "间，因为这是欧几单得近似执行得最好的地方（cf.图E3）；然", "(1)", "后，我们利用指数映射将其映射回双曲空间，以获得表示e(1)", "B上z点的切线空间TB是n维欧儿里得空间，它最接近z", "。因此，聚合器能够将关系消息集成到双曲线空间中的目标表", "周围的B，这对于在双曲空间国[7]中执行聚合操作很有用", "", "示中，并通过映射操作避免复杂的双曲线均值池。类似地，我", "指数映射&对数映射。双曲空间和切线空间可以通过指数和", "们可以获得每个KG实体eE8的表示e(1)", "对数映射来桥接。具体来说，指数映射可以将切线空问TB映", "我们进一步堆叠更多的聚合层，以探索项目的高阶知识关", "射到双曲空间B，而对数映射则相反地将B映射到TB。特", "联。从技术上讲，我们递归地将1层之后的i项的知识表示表", "别是，下面定义了z点处两个映射的封闭形式表达式。", "示为：", "IIxll _) x", "expz(x) = z tanh(", "(2)", "el) =expo(mZ(r)en", "(-1/z/2xl", "", "N loge(t-1) (e(f-1) 由er))", "(6)", "1og(y) = (1- IZ/) anh-(l-z @ yl)/-2每", "-", "-z@y", "(3)", "3.2.2AngleConstraintofHierarchicaltriplets.接下来，我们专", "where represents Mobius addition [12] :", "注于对具有层次关系的项目的属性信息进行建模。为了简化我", "(1 + 2(x, y) + llyl/2)x + (1 - x12)y", "们的讨论，我们假设KG中的所有关系都聚集成两种类型（i.e", "x@y=", "(4)", "分层和非分层），并且关系的类型是先验给出的。事实上，大", "1 + 2(x y) + x/21lyl/2", "多数KGs[2,6]中都有明确的关系类型。当此类信息不可用时，", "有两个替代选项。一种是通过广泛使用的Krackhardt评分标", "3.2Hierarchy-AwareModeling", "准E[27]来推断关系的“层次性”；另一种是简单地假设项目", "与以前的基于传播的研究F[37.43.45，471不同，后者只是将用", "连接的实体充当项目的属性，因此，这些关系是分层的。我们", "户项图和KG嵌入到欧几里得空间中，我们的目标是捕获数据", "将在4.3.2节中报告的实验表明，两个选项之间的差异可以忽", "的非欧几里得潜在解部结构。以图1(b）所示的Alibaba-iFishion", "略不计", "数据集为例，用户-项目交互的度分布揭示了一种具有幂律分", "如前所述，层次关系可以揭示项目的属性。但是，上述聚合", "布的底层分层树状结构。同时，KG还表现出分层模式[2,6]。", "方案无法在嵌入空间中保留此语义，因为它只是将所有相邻信", "但是，由于高失真E[12,34]，无法在欧几里得空间中正确捕", "息聚合为表示形式。为了明确地对项目的特征信息进行建模，", "获数据的分层属性。因此，我们将用户和项目以及实体和关系", "我们考虑了双曲包含锥E[13]，它是一族凸锥，可以对具有嵌", "嵌入到双曲线空间中，其中可以自然地保留层次结构。我们还", "套角锥的层次结构进行建模。从技术上讲，我们使用Cx表示" ]

{ "arxiv_id": "2204.04959", "language": "zh", "page_idx": 2 }

[ [ "0.08496", "0.1067", "0.4805", "0.108", "0.4805", "0.1218", "0.08496", "0.1206" ], [ "0.523", "0.1061", "0.9062", "0.1061", "0.9062", "0.1193", "0.523", "0.1193" ], [ "0.08496", "0.1206", "0.4805", "0.1218", "0.4805", "0.1351", "0.08496", "0.1338" ], [ "0.6787", "0.1218", "0.9043", "0.1218", "0.9043", "0.1351", "0.6787", "0.1351" ], [ "0.08826", "0.137", "0.4805", "0.137", "0.4805", "0.1477", "0.08826", "0.1477" ], [ "0.738", "0.137", "0.778", "0.137", "0.778", "0.1483", "0.738", "0.1483" ], [ "0.8037", "0.1351", "0.8423", "0.137", "0.841", "0.1497", "0.8022", "0.1477" ], [ "0.8667", "0.137", "0.903", "0.137", "0.903", "0.1483", "0.8667", "0.1483" ], [ "0.6006", "0.1383", "0.642", "0.1383", "0.642", "0.1497", "0.6006", "0.1497" ], [ "0.5317", "0.1401", "0.5864", "0.1427", "0.585", "0.1572", "0.5312", "0.1547" ], [ "0.6006", "0.1477", "0.642", "0.1497", "0.6406", "0.1635", "0.5986", "0.161" ], [ "0.8022", "0.1477", "0.8433", "0.1477", "0.8433", "0.1622", "0.8022", "0.1622" ], [ "0.7427", "0.1483", "0.7812", "0.1483", "0.7812", "0.1622", "0.7427", "0.1622" ], [ "0.0866", "0.1483", "0.327", "0.1509", "0.327", "0.1642", "0.0866", "0.1622" ], [ "0.865", "0.1497", "0.902", "0.1497", "0.902", "0.161", "0.865", "0.161" ], [ "0.5293", "0.1559", "0.589", "0.1559", "0.589", "0.1674", "0.5293", "0.1674" ], [ "0.5986", "0.1578", "0.6763", "0.1604", "0.6763", "0.1749", "0.5986", "0.1724" ], [ "0.7295", "0.161", "0.7793", "0.161", "0.7793", "0.1724", "0.7295", "0.1724" ], [ "0.789", "0.1622", "0.8423", "0.1622", "0.8423", "0.1724", "0.789", "0.1724" ], [ "0.8447", "0.1578", "0.903", "0.161", "0.902", "0.1755", "0.8433", "0.1724" ], [ "0.1936", "0.1686", "0.3677", "0.1686", "0.3677", "0.1818", "0.1936", "0.1818" ], [ "0.455", "0.1692", "0.4836", "0.1692", "0.4836", "0.1843", "0.455", "0.1843" ], [ "0.6006", "0.1768", "0.6543", "0.1768", "0.6543", "0.1881", "0.6006", "0.1881" ], [ "0.7427", "0.1768", "0.7793", "0.1768", "0.7793", "0.1881", "0.7427", "0.1881" ], [ "0.8037", "0.1768", "0.8423", "0.1768", "0.8423", "0.1881", "0.8037", "0.1881" ], [ "0.8667", "0.1755", "0.9043", "0.1755", "0.9043", "0.1895", "0.8667", "0.1895" ], [ "0.5986", "0.1863", "0.6626", "0.1881", "0.6626", "0.2026", "0.5986", "0.2008" ], [ "0.5254", "0.178", "0.589", "0.1799", "0.589", "0.1945", "0.5254", "0.1925" ], [ "0.763", "0.1881", "0.7812", "0.1881", "0.7812", "0.202", "0.763", "0.202" ], [ "0.8228", "0.1881", "0.8423", "0.1881", "0.8423", "0.2008", "0.8228", "0.2008" ], [ "0.5415", "0.1945", "0.5767", "0.1945", "0.5767", "0.207", "0.5415", "0.207" ], [ "0.8906", "0.1895", "0.902", "0.1895", "0.902", "0.1995", "0.8906", "0.1995" ], [ "0.08496", "0.1957", "0.2942", "0.1957", "0.2942", "0.209", "0.08496", "0.209" ], [ "0.6006", "0.2008", "0.6543", "0.2008", "0.6543", "0.2122", "0.6006", "0.2122" ], [ "0.739", "0.2008", "0.7817", "0.2008", "0.7817", "0.2122", "0.739", "0.2122" ], [ "0.778", "0.2008", "0.8423", "0.2008", "0.8423", "0.2122", "0.778", "0.2122" ], [ "0.8613", "0.2008", "0.903", "0.2008", "0.903", "0.2122", "0.8613", "0.2122" ], [ "0.0866", "0.2134", "0.4805", "0.2134", "0.4805", "0.2267", "0.0866", "0.2267" ], [ "0.08496", "0.2267", "0.4805", "0.226", "0.4805", "0.2393", "0.08496", "0.2405" ], [ "0.08826", "0.2424", "0.4795", "0.2424", "0.4795", "0.2524", "0.08826", "0.2524" ], [ "0.517", "0.238", "0.9126", "0.2393", "0.9126", "0.2524", "0.517", "0.252" ], [ "0.08826", "0.2556", "0.4795", "0.2556", "0.4795", "0.2664", "0.08826", "0.2664" ], [ "0.5186", "0.252", "0.911", "0.2524", "0.911", "0.2664", "0.5186", "0.2651" ], [ "0.0866", "0.2683", "0.482", "0.2683", "0.482", "0.2822", "0.0866", "0.2822" ], [ "0.5195", "0.2683", "0.911", "0.2683", "0.911", "0.279", "0.5195", "0.279" ], [ "0.0866", "0.2822", "0.4805", "0.2822", "0.4805", "0.2954", "0.0866", "0.2954" ], [ "0.5186", "0.281", "0.9126", "0.281", "0.9126", "0.2942", "0.5186", "0.2942" ], [ "0.0866", "0.2966", "0.4805", "0.2966", "0.4805", "0.31", "0.0866", "0.31" ], [ "0.5195", "0.2966", "0.911", "0.2966", "0.911", "0.307", "0.5195", "0.307" ], [ "0.0866", "0.3093", "0.1642", "0.3113", "0.1642", "0.3225", "0.0866", "0.3208" ], [ "0.5186", "0.3093", "0.9126", "0.3093", "0.9126", "0.3225", "0.5186", "0.3225" ], [ "0.2385", "0.327", "0.3284", "0.3289", "0.3284", "0.3435", "0.2385", "0.3416" ], [ "0.5186", "0.3225", "0.9126", "0.3225", "0.9126", "0.336", "0.5186", "0.336" ], [ "0.4558", "0.33", "0.482", "0.33", "0.482", "0.3423", "0.4558", "0.3423" ], [ "0.517", "0.3352", "0.9126", "0.336", "0.9126", "0.3499", "0.517", "0.3484" ], [ "0.0866", "0.3499", "0.4805", "0.3499", "0.4805", "0.363", "0.0866", "0.363" ], [ "0.5195", "0.3516", "0.911", "0.3516", "0.911", "0.3623", "0.5195", "0.3623" ], [ "0.08417", "0.361", "0.1487", "0.363", "0.1487", "0.3774", "0.08417", "0.3757" ], [ "0.5186", "0.3643", "0.8", "0.3643", "0.8", "0.3774", "0.5186", "0.3774" ], [ "0.08417", "0.3894", "0.2566", "0.3894", "0.2566", "0.4028", "0.08417", "0.4028" ], [ "0.5195", "0.3901", "0.9126", "0.3901", "0.9126", "0.401", "0.5195", "0.401" ], [ "0.0866", "0.4077", "0.4805", "0.4077", "0.4805", "0.4216", "0.0866", "0.4216" ], [ "0.5195", "0.404", "0.911", "0.404", "0.911", "0.414", "0.5195", "0.414" ], [ "0.0866", "0.4216", "0.4805", "0.4216", "0.4805", "0.435", "0.0866", "0.435" ], [ "0.5195", "0.4185", "0.911", "0.4185", "0.911", "0.4287", "0.5195", "0.4287" ], [ "0.5186", "0.4312", "0.9126", "0.4312", "0.9126", "0.4443", "0.5186", "0.4443" ], [ "0.08826", "0.4375", "0.4795", "0.4375", "0.4795", "0.4475", "0.08826", "0.4475" ], [ "0.08826", "0.4495", "0.4805", "0.4495", "0.4805", "0.4634", "0.08826", "0.4634" ], [ "0.5186", "0.4443", "0.9126", "0.4443", "0.9126", "0.4583", "0.5186", "0.4583" ], [ "0.08496", "0.464", "0.4805", "0.464", "0.4805", "0.4792", "0.08496", "0.4792" ], [ "0.5186", "0.459", "0.6685", "0.459", "0.6685", "0.473", "0.5186", "0.473" ], [ "0.08417", "0.4792", "0.4805", "0.4797", "0.4805", "0.4944", "0.08417", "0.4937" ], [ "0.5186", "0.4841", "0.9126", "0.4841", "0.9126", "0.4976", "0.5186", "0.4976" ], [ "0.08826", "0.4956", "0.1609", "0.4956", "0.1609", "0.507", "0.08826", "0.507" ], [ "0.517", "0.4956", "0.9126", "0.4968", "0.9126", "0.511", "0.517", "0.5103" ], [ "0.08496", "0.513", "0.4795", "0.515", "0.4795", "0.53", "0.08496", "0.528" ], [ "0.5195", "0.512", "0.911", "0.512", "0.911", "0.523", "0.5195", "0.523" ], [ "0.5186", "0.5244", "0.898", "0.5244", "0.898", "0.5386", "0.5186", "0.5386" ], [ "0.08826", "0.531", "0.478", "0.531", "0.478", "0.5415", "0.08826", "0.5415" ], [ "0.0866", "0.5435", "0.4805", "0.5435", "0.4805", "0.5566", "0.0866", "0.5566" ], [ "0.5186", "0.5503", "0.914", "0.5503", "0.914", "0.5645", "0.5186", "0.5645" ], [ "0.0866", "0.5566", "0.4805", "0.5566", "0.4805", "0.571", "0.0866", "0.571" ], [ "0.5186", "0.563", "0.914", "0.563", "0.914", "0.577", "0.5186", "0.577" ], [ "0.08826", "0.571", "0.4805", "0.571", "0.4805", "0.584", "0.08826", "0.584" ], [ "0.5186", "0.577", "0.914", "0.5776", "0.914", "0.5913", "0.5186", "0.5903" ], [ "0.0866", "0.584", "0.482", "0.584", "0.482", "0.5977", "0.0866", "0.5977" ], [ "0.08496", "0.5967", "0.4805", "0.5977", "0.4805", "0.6113", "0.08496", "0.61" ], [ "0.5215", "0.5933", "0.9126", "0.5933", "0.9126", "0.604", "0.5215", "0.604" ], [ "0.517", "0.604", "0.9126", "0.605", "0.9126", "0.619", "0.517", "0.6187" ], [ "0.08826", "0.6123", "0.4805", "0.6123", "0.4805", "0.6255", "0.08826", "0.6255" ], [ "0.5186", "0.6187", "0.9126", "0.6187", "0.9126", "0.632", "0.5186", "0.632" ], [ "0.0866", "0.6255", "0.4805", "0.6255", "0.4805", "0.6396", "0.0866", "0.6396" ], [ "0.5186", "0.6333", "0.914", "0.6333", "0.914", "0.6465", "0.5186", "0.6465" ], [ "0.0866", "0.6396", "0.482", "0.6396", "0.482", "0.653", "0.0866", "0.653" ], [ "0.517", "0.646", "0.9126", "0.6465", "0.9126", "0.66", "0.517", "0.659" ], [ "0.08496", "0.654", "0.3245", "0.654", "0.3245", "0.664", "0.08496", "0.664" ], [ "0.5195", "0.662", "0.9126", "0.662", "0.9126", "0.6724", "0.5195", "0.6724" ], [ "0.517", "0.674", "0.914", "0.675", "0.914", "0.688", "0.517", "0.6875" ], [ "0.0866", "0.681", "0.2443", "0.681", "0.2443", "0.6914", "0.0866", "0.6914" ], [ "0.5186", "0.688", "0.914", "0.688", "0.914", "0.702", "0.5186", "0.702" ], [ "0.0866", "0.6978", "0.482", "0.6978", "0.482", "0.711", "0.0866", "0.711" ], [ "0.5195", "0.704", "0.911", "0.704", "0.911", "0.714", "0.5195", "0.714" ], [ "0.08496", "0.711", "0.2998", "0.712", "0.2998", "0.726", "0.08496", "0.7246" ], [ "0.5186", "0.7163", "0.9126", "0.7163", "0.9126", "0.73", "0.5186", "0.73" ], [ "0.08496", "0.728", "0.4805", "0.728", "0.4805", "0.742", "0.08496", "0.742" ], [ "0.5195", "0.73", "0.914", "0.73", "0.914", "0.7437", "0.5195", "0.7437" ], [ "0.0989", "0.742", "0.3865", "0.742", "0.3865", "0.756", "0.0989", "0.756" ], [ "0.5195", "0.7456", "0.911", "0.7456", "0.911", "0.756", "0.5195", "0.756" ], [ "0.08496", "0.756", "0.4805", "0.756", "0.4805", "0.7695", "0.08496", "0.7695" ], [ "0.5215", "0.759", "0.911", "0.759", "0.911", "0.7695", "0.5215", "0.7695" ], [ "0.10046", "0.771", "0.482", "0.771", "0.482", "0.784", "0.10046", "0.784" ], [ "0.5186", "0.7715", "0.914", "0.7715", "0.914", "0.785", "0.5186", "0.785" ], [ "0.0972", "0.7827", "0.4224", "0.7827", "0.4224", "0.797", "0.0972", "0.797" ], [ "0.5186", "0.784", "0.914", "0.784", "0.914", "0.7974", "0.5186", "0.7974" ], [ "0.0866", "0.7974", "0.482", "0.7974", "0.482", "0.811", "0.0866", "0.811" ], [ "0.5186", "0.8", "0.9126", "0.8", "0.9126", "0.813", "0.5186", "0.813" ], [ "0.10046", "0.813", "0.3743", "0.813", "0.3743", "0.8237", "0.10046", "0.8237" ], [ "0.5186", "0.813", "0.9126", "0.813", "0.9126", "0.827", "0.5186", "0.827" ], [ "0.5186", "0.827", "0.9126", "0.827", "0.9126", "0.8403", "0.5186", "0.8403" ], [ "0.08826", "0.8384", "0.1323", "0.8384", "0.1323", "0.8496", "0.08826", "0.8496" ], [ "0.1259", "0.836", "0.312", "0.837", "0.312", "0.8516", "0.1259", "0.851" ], [ "0.5195", "0.8384", "0.5664", "0.8403", "0.5645", "0.855", "0.5186", "0.853" ], [ "0.0866", "0.856", "0.4795", "0.856", "0.4795", "0.869", "0.0866", "0.869" ], [ "0.08496", "0.869", "0.4795", "0.869", "0.4795", "0.8833", "0.08496", "0.8833" ], [ "0.08826", "0.8853", "0.4795", "0.8853", "0.4795", "0.896", "0.08826", "0.896" ], [ "0.5156", "0.882", "0.6895", "0.8833", "0.6895", "0.898", "0.5156", "0.8965" ] ]

[ "通过这样做，协作和知识的互补信息被分别编码在最终表示", "Table 1: Statistics of the datasets used in our experiments.", "中。与以前的工作不同，我们采用余弦相似度从项目的行为和", "Alibaba-iFashionYelp2018Last-FM", "属性方面分别预测用户ui采用项目i的可能性，然后采用两", "114,737", "45,919", "23,566", "# Users", "User-Item", "# Items", "45,538", "30,040", "个预测的总和作为最终预测得分gui：", "48,123", "Interaction", "# Interactions", "1,781,093", "1,185,068", "3,034,796", "ui = cos(eu,éi) + cos(eu,ei)", "(14)", "#Entities", "59,156", "90,961", "58,266", "# Relations", "Knowledge", "51", "42", "Graph", "9", "3.5ModelOptimization", "# Triplets", "279,155", "1,853,704", "464,567", "我们选择对比损失[33]来优化HAKG。与广泛使用的BPR损", "失[35]相比，它通过引人更多的负样本和惩罚无信息的样本", "来缓解收敛问题。具体来说，对于用户-项目图（i.e.，正对）捕", "域，大小和稀疏性方面有所不同。在之前的工作[45]之后，我", "获的每个交互（u,i），我们随机采样Mu|未观察到的项目", "们收集KG中项目的两跳邻居实体，为每个数据集构建项目知", "以与用户u一起形成负对，表示为Mu，并最大化正对之间的", "识图。此外，对于KG中关系的层次结构类型，在不失一般性", "相似性，同时最小化负对的相似性m：<NEXMATHLINE>。通", "的情况下，我们简单地将项目连接关系视为层次结构，因为它", "过结合预测损失和角度损失，我们最小化以下目标函数来学", "们能够揭示项目的属性。此外，为了保证数据质量，我们采用", "习模型参数", "10核设置E[45]，i.e,保留至少十个交互的用户和项目，并过", "=L+ACH", "滤掉少于十个三元组中涉及的KG实体。我们使用与先前研究", "(15)", "相同的数据分区E[45,47]进行比较（i.e,所有数据集的训练，", "其中入是一个超参数，用于控制公式国（?？）中定义的角度损", "验证和测试集的比例分别为80%，10%和10%）。表1总结了我", "失的权重", "们实验中使用的三个数据集的总体统计数据", "3.6ModelAnalysis", "4.1.2EvaluationMetrics.我们采用全排名策略E[28,47]来评", "3.6.1Model Size.先前的研究[47]已经证实，丢弃非线性特", "估性能。具体来说，我们将用户以前未采用的所有项目视为阴", "征变换不仅可以产生更好的性能，而且可以减少允余参数。因", "性，并将测试集中的相关项自目视为阳性。为了评估top-K建", "议的有效性，我们采用了两种广泛使用的评估协议E[18,28]", "此，在HAKG的聚合方案中，我们丢弃了非线性激活函数和", "可学习的变换度量。HAKG的模型参数包括：（1）用户，项目，", "recall@K和ndcg@K，其中默认为K=20。我们报告测试集", "KG实体和关系(e),e(),e),erluu,ieIe&,rR)的", "中所有用户的平均指标", "ID嵌人；（2）用于方程E（10）中使用的信息门控聚合的转换", "4.1.3Baselines.为了证明HAKG的有效性，我们将其与最先", "参数W1.W2", "进的方法进行了比较，包括无KG方法（MF），基于嵌人的方法", "3.6.2TimeComplexity:HAKG的时间成本主要来自聚合方案。", "CKE和UGRec），基于传播的方法（KGNN-LS，KGAT，CKAN", "和KGIN）以及基于双曲线的方法（Hyper-<NEXE和LKGR）", "在KG上的聚合中，更新双曲线空间中的知识项嵌人的计算", "复杂度为O(IGkldL），其中IGkl，d和L表示KG三元组的", "4.1.4ParameterSettings.我们在Pytorch中实现了HAKG，并", "数量，嵌入大小和层数。在用户-项目图上的聚合中，协作项", "发布了我们的实现2（代码，参数设置和训练日志）以促进", "目嵌人和用户嵌入的计算复杂度为O(IGbldL），其中IGl是", "可重复性。为了进行公平的比较，我们将ID嵌入的大小固定", "交互数。对于层次感知建模，角度约束的成本为O(III），其", "中II是项数。此外，欧几里得空间和双曲空间之间的映射", "为64，并使用AdamE[24]优化器优化所有模型，其中批处", "理大小固定为4096。Xavier初始化器[15]用于初始化模型参", "成本为O(IGbl+IGkl）。因此，整个训练时期的时间复杂度为", "数。我们对超参数进行网格搜索：在10-4,10-3,10-2)之间调", "O(IGkldL+IGbldL）。在相同的实验设置（i.e,相同的嵌人大小", "整学习速率，在{10-5,10-4,，10-2,10-1)中搜索角度损失入", "和传播层）下，HAKG具有与三种代表性的基于传播的方法", "的权重，并在(1,2,3)中调整GNN层L以实现基于传播的方", "KGAT，CKAN和KGIN相当的复杂性", "法。对于每个用户的负样本数Mul和对比损失的余量m，我", "们分别为Alibaba-iFashion，Yelp2018和Last-FM数据集设置为", "4EXPERIMENTS", "[200,400,400)和(0.6,0.8,0.7)。此外，由于双曲空间中的优化", "我们提供了经验结果来证明我们提出的HAKG框架的有效性。", "实际上具有挑战性，因此我们改为在原点的切线空间中定义", "实验旨在回答以下三个研究问题：", "所有参数，使用标准欧几里得技术优化嵌人，并使用指数映射", ". RQ1: How does HAKG perform, compared with the state-of-the", "来恢复双曲参数E[6]。所有基线方法的参数都经过仔细调整，", "art knowledge-aware recommendation models?", "以实现最佳性能。具体来说，对于KGAT，我们使用隐藏大小", ". RQ2: How do different components of HAKG (i.e., hierarchical", "64,32,16)将深度设置为三，并使用MF的预训练ID嵌入作为", "modeling.gated propagation,dual embeddings, and the depth of", "初始化；对于UGRec，我们将铰链损失的边距值设置为1。5；", "propagation layers) affect the performance of HAKG?", "对于CKAN，KGNN-LS和LKGR，我们将邻域的大小设置为16；", ". RQ3: Can HAKG provide meaningful insights on the hierarchical", "对于KGIN，我们将意图数固定为4；对于Hyper-Know，我们", "modellingbothfrom structures and relations?", "将曲率设置为-1。此外，对所有方法执行早期停止策略，如果", "测试集上的召回@20在连续10个时期内没有增加，则i.e.过", "4.1", "Experimental Settings", "早停止", "4.1.1DatasetDescription.我们利用三个基准数据集来评估", "HAKG的性能：Alibaba-iFashion，Yelp2018和Last-FM。这三个", "数据集在最先进的方法E[8，31，4547]中被广泛采用，并且在", "2https://github.com/Scottdyt/HAKG" ]

{ "arxiv_id": "2204.04959", "language": "zh", "page_idx": 4 }

[ [ "0.08826", "0.07764", "0.3743", "0.07764", "0.3743", "0.0884", "0.08826", "0.0884" ], [ "0.642", "0.07574", "0.911", "0.077", "0.911", "0.0884", "0.642", "0.0871" ], [ "0.0866", "0.108", "0.4805", "0.108", "0.4805", "0.1218", "0.0866", "0.1218" ], [ "0.5156", "0.1061", "0.7114", "0.1067", "0.7114", "0.1218", "0.5156", "0.1206" ], [ "0.08826", "0.1218", "0.4805", "0.1218", "0.4805", "0.1351", "0.08826", "0.1351" ], [ "0.517", "0.1256", "0.9126", "0.1256", "0.9126", "0.1395", "0.517", "0.1395" ], [ "0.10455", "0.137", "0.4795", "0.137", "0.4795", "0.1477", "0.10455", "0.1477" ], [ "0.5186", "0.1395", "0.914", "0.1395", "0.914", "0.1528", "0.5186", "0.1528" ], [ "0.0866", "0.1497", "0.4805", "0.1497", "0.4805", "0.1635", "0.0866", "0.1635" ], [ "0.517", "0.1497", "0.915", "0.1497", "0.915", "0.1674", "0.517", "0.1674" ], [ "0.08826", "0.1654", "0.478", "0.1654", "0.478", "0.1755", "0.08826", "0.1755" ], [ "0.5186", "0.1674", "0.9126", "0.1674", "0.9126", "0.1812", "0.5186", "0.1812" ], [ "0.08826", "0.1787", "0.4795", "0.1787", "0.4795", "0.1895", "0.08826", "0.1895" ], [ "0.5186", "0.1812", "0.9126", "0.1812", "0.9126", "0.1945", "0.5186", "0.1945" ], [ "0.08905", "0.1932", "0.4795", "0.1932", "0.4795", "0.2039", "0.08905", "0.2039" ], [ "0.517", "0.1957", "0.726", "0.1957", "0.726", "0.209", "0.517", "0.209" ], [ "0.08826", "0.207", "0.478", "0.207", "0.478", "0.2172", "0.08826", "0.2172" ], [ "0.08826", "0.2203", "0.4795", "0.2203", "0.4795", "0.2311", "0.08826", "0.2311" ], [ "0.517", "0.2216", "0.876", "0.2216", "0.876", "0.2349", "0.517", "0.2349" ], [ "0.08826", "0.2329", "0.4412", "0.2329", "0.4412", "0.2468", "0.08826", "0.2468" ], [ "0.5186", "0.2393", "0.947", "0.2393", "0.947", "0.2524", "0.5186", "0.2524" ], [ "0.10297", "0.2468", "0.4795", "0.2468", "0.4795", "0.26", "0.10297", "0.26" ], [ "0.0866", "0.26", "0.4805", "0.26", "0.4805", "0.2734", "0.0866", "0.2734" ], [ "0.5195", "0.2551", "0.914", "0.2551", "0.914", "0.2651", "0.5195", "0.2651" ], [ "0.5195", "0.2676", "0.763", "0.2676", "0.763", "0.281", "0.5195", "0.281" ], [ "0.08905", "0.276", "0.4795", "0.276", "0.4795", "0.286", "0.08905", "0.286" ], [ "0.08826", "0.2886", "0.4805", "0.2886", "0.4805", "0.3018", "0.08826", "0.3018" ], [ "0.6006", "0.2942", "0.831", "0.2942", "0.831", "0.308", "0.6006", "0.308" ], [ "0.0866", "0.3005", "0.4795", "0.3005", "0.4795", "0.3145", "0.0866", "0.3145" ], [ "0.08826", "0.3176", "0.478", "0.3176", "0.478", "0.3276", "0.08826", "0.3276" ], [ "0.0866", "0.3289", "0.4795", "0.3289", "0.4795", "0.3423", "0.0866", "0.3423" ], [ "0.536", "0.3276", "0.5947", "0.3276", "0.5947", "0.3423", "0.536", "0.3423" ], [ "0.6313", "0.33", "0.705", "0.33", "0.705", "0.3416", "0.6313", "0.3416" ], [ "0.08826", "0.3423", "0.4805", "0.3423", "0.4805", "0.356", "0.08826", "0.356" ], [ "0.535", "0.3484", "0.5537", "0.3484", "0.5537", "0.363", "0.535", "0.363" ], [ "0.6313", "0.3499", "0.7915", "0.351", "0.7915", "0.3643", "0.6313", "0.363" ], [ "0.0866", "0.358", "0.3962", "0.358", "0.3962", "0.3718", "0.0866", "0.3718" ], [ "0.536", "0.363", "0.5557", "0.363", "0.5557", "0.3774", "0.536", "0.3774" ], [ "0.6313", "0.363", "0.7935", "0.3643", "0.7935", "0.379", "0.6313", "0.3774" ], [ "0.10297", "0.3718", "0.4795", "0.3718", "0.4795", "0.385", "0.10297", "0.385" ], [ "0.5376", "0.3774", "0.596", "0.3774", "0.596", "0.3926", "0.5376", "0.3926" ], [ "0.6313", "0.379", "0.884", "0.379", "0.884", "0.3926", "0.6313", "0.3926" ], [ "0.0866", "0.385", "0.482", "0.385", "0.482", "0.3984", "0.0866", "0.3984" ], [ "0.536", "0.3977", "0.556", "0.3977", "0.556", "0.4124", "0.536", "0.4124" ], [ "0.6323", "0.3977", "0.749", "0.3977", "0.749", "0.411", "0.6323", "0.411" ], [ "0.536", "0.411", "0.5786", "0.411", "0.5786", "0.4268", "0.536", "0.4268" ], [ "0.08905", "0.401", "0.4795", "0.401", "0.4795", "0.411", "0.08905", "0.411" ], [ "0.08905", "0.414", "0.478", "0.414", "0.478", "0.4248", "0.08905", "0.4248" ], [ "0.6323", "0.4124", "0.8936", "0.4124", "0.8936", "0.4255", "0.6323", "0.4255" ], [ "0.08826", "0.4268", "0.4805", "0.4268", "0.4805", "0.44", "0.08826", "0.44" ], [ "0.536", "0.4312", "0.6113", "0.428", "0.613", "0.4426", "0.5376", "0.4458" ], [ "0.6323", "0.433", "0.7817", "0.433", "0.7817", "0.4463", "0.6323", "0.4463" ], [ "0.08496", "0.4395", "0.4805", "0.44", "0.4805", "0.4539", "0.08496", "0.4526" ], [ "0.536", "0.4475", "0.621", "0.4426", "0.6216", "0.457", "0.5376", "0.4622" ], [ "0.6313", "0.4475", "0.841", "0.4475", "0.841", "0.461", "0.6313", "0.461" ], [ "0.08826", "0.4558", "0.4795", "0.4558", "0.4795", "0.4666", "0.08826", "0.4666" ], [ "0.08826", "0.4697", "0.1969", "0.4697", "0.1969", "0.4797", "0.08826", "0.4797" ], [ "0.5195", "0.4968", "0.5317", "0.4968", "0.5317", "0.505", "0.5195", "0.505" ], [ "0.5293", "0.4968", "0.55", "0.4968", "0.55", "0.505", "0.5293", "0.505" ], [ "0.546", "0.4956", "0.669", "0.4956", "0.669", "0.506", "0.546", "0.506" ], [ "0.1136", "0.509", "0.4795", "0.509", "0.4795", "0.523", "0.1136", "0.523" ], [ "0.5195", "0.5146", "0.911", "0.5146", "0.911", "0.5244", "0.5195", "0.5244" ], [ "0.1259", "0.523", "0.4983", "0.523", "0.4983", "0.536", "0.1259", "0.536" ], [ "0.5195", "0.528", "0.911", "0.528", "0.911", "0.5386", "0.5195", "0.5386" ], [ "0.1259", "0.535", "0.4795", "0.535", "0.4795", "0.549", "0.1259", "0.549" ], [ "0.5195", "0.5415", "0.911", "0.5415", "0.911", "0.552", "0.5195", "0.552" ], [ "0.125", "0.55", "0.2834", "0.5503", "0.2834", "0.5645", "0.125", "0.563" ], [ "0.5186", "0.5537", "0.6772", "0.5537", "0.6772", "0.5674", "0.5186", "0.5674" ], [ "0.6855", "0.556", "0.911", "0.556", "0.911", "0.5664", "0.6855", "0.5664" ], [ "0.11273", "0.563", "0.482", "0.563", "0.482", "0.5776", "0.11273", "0.5776" ], [ "0.5186", "0.5674", "0.9126", "0.5674", "0.9126", "0.5806", "0.5186", "0.5806" ], [ "0.125", "0.5776", "0.492", "0.577", "0.492", "0.5903", "0.125", "0.5913" ], [ "0.5186", "0.582", "0.9126", "0.582", "0.9126", "0.595", "0.5186", "0.595" ], [ "0.125", "0.5923", "0.5063", "0.5923", "0.5063", "0.606", "0.125", "0.606" ], [ "0.5186", "0.595", "0.911", "0.5967", "0.911", "0.608", "0.5186", "0.607" ], [ "0.125", "0.605", "0.3022", "0.605", "0.3022", "0.6187", "0.125", "0.6187" ], [ "0.5195", "0.6113", "0.911", "0.6113", "0.911", "0.622", "0.5195", "0.622" ], [ "0.1111", "0.619", "0.4805", "0.619", "0.4805", "0.6333", "0.1111", "0.6333" ], [ "0.517", "0.622", "0.9126", "0.6226", "0.9126", "0.636", "0.517", "0.6353" ], [ "0.125", "0.632", "0.4795", "0.6333", "0.4795", "0.6465", "0.125", "0.646" ], [ "0.5186", "0.637", "0.914", "0.637", "0.914", "0.651", "0.5186", "0.651" ], [ "0.1259", "0.648", "0.1814", "0.648", "0.1814", "0.659", "0.1259", "0.659" ], [ "0.5195", "0.653", "0.8896", "0.653", "0.8896", "0.6636", "0.5195", "0.6636" ], [ "0.517", "0.678", "0.746", "0.678", "0.746", "0.6914", "0.517", "0.6914" ], [ "0.517", "0.6924", "0.9126", "0.6934", "0.9126", "0.711", "0.517", "0.7104" ], [ "0.08496", "0.6987", "0.1168", "0.6987", "0.1168", "0.7104", "0.08496", "0.7104" ], [ "0.1111", "0.6987", "0.3381", "0.6987", "0.3381", "0.709", "0.1111", "0.709" ], [ "0.517", "0.709", "0.9126", "0.711", "0.9126", "0.726", "0.517", "0.7236" ], [ "0.08496", "0.7173", "0.392", "0.7173", "0.392", "0.731", "0.08496", "0.731" ], [ "0.517", "0.7246", "0.9126", "0.726", "0.9126", "0.7393", "0.517", "0.738" ], [ "0.0866", "0.735", "0.4805", "0.7363", "0.4805", "0.7476", "0.0866", "0.747" ], [ "0.5186", "0.7407", "0.911", "0.7407", "0.911", "0.754", "0.5186", "0.754" ], [ "0.08826", "0.7505", "0.4805", "0.7505", "0.4805", "0.765", "0.08826", "0.765" ], [ "0.5215", "0.756", "0.911", "0.756", "0.911", "0.7666", "0.5215", "0.7666" ], [ "0.0866", "0.7666", "0.4697", "0.7646", "0.4697", "0.779", "0.0866", "0.781" ], [ "0.517", "0.7646", "0.914", "0.7646", "0.914", "0.782", "0.517", "0.782" ], [ "0.08826", "0.781", "0.4805", "0.781", "0.4805", "0.7954", "0.08826", "0.7954" ], [ "0.5186", "0.7827", "0.911", "0.7827", "0.911", "0.7935", "0.5186", "0.7935" ], [ "0.0866", "0.799", "0.482", "0.799", "0.482", "0.8125", "0.0866", "0.8125" ], [ "0.5156", "0.7944", "0.8384", "0.7944", "0.8384", "0.8125", "0.5156", "0.8125" ], [ "0.08826", "0.8145", "0.4795", "0.8145", "0.4795", "0.8247", "0.08826", "0.8247" ], [ "0.08826", "0.827", "0.4805", "0.827", "0.4805", "0.8413", "0.08826", "0.8413" ], [ "0.517", "0.8237", "0.6626", "0.8237", "0.6626", "0.837", "0.517", "0.837" ], [ "0.0866", "0.8413", "0.4805", "0.8413", "0.4805", "0.856", "0.0866", "0.856" ], [ "0.5186", "0.8413", "0.911", "0.8413", "0.911", "0.855", "0.5186", "0.855" ], [ "0.0866", "0.856", "0.4805", "0.856", "0.4805", "0.869", "0.0866", "0.869" ], [ "0.5195", "0.857", "0.911", "0.857", "0.911", "0.867", "0.5195", "0.867" ], [ "0.0866", "0.869", "0.4805", "0.869", "0.4805", "0.8833", "0.0866", "0.8833" ], [ "0.5186", "0.869", "0.9126", "0.869", "0.9126", "0.8833", "0.5186", "0.8833" ], [ "0.08826", "0.8853", "0.2279", "0.8853", "0.2279", "0.896", "0.08826", "0.896" ], [ "0.5186", "0.882", "0.5645", "0.884", "0.564", "0.8984", "0.517", "0.8965" ] ]

[ "Conditional Generation Net for Medication Recommendatior", "WwW'22,April 2529,2022, Virtual Event, Lyon,France", "生来说是耗时的，对于没有经验的医生来说是容易出错的。为", "2.2EHR&DDIGraph", "了解决这个问题，敦促可以帮助医生做出决策的自动药物推荐", "Ge=(M,8e）和Gd=(M,8a)分别表示EHR图和DDI图，其", "由于临床价值，药物推荐吸引了越来越多的研究兴趣。已经", "中8e是R中所有可能的药物组合，8d是已知的DDI。形式", "提出了一系列基于深度学习的用药推荐方法[27,28]，可以将", "上，我们使用邻接矩阵Ae,AdeRIMIxIM来说明图的构造。", "其分为两类：基于实例的模型[7.18，29]仅使用惠者当前的诊", "Ae[i,j]=1表示i和j药物出现在同一次就诊中。对于Ad，仅", "断和程序进行推荐，而忽略了患者的纵向病史。以这种方式，", "考虑成对DDI，Aa[i,]=1表示第i和第j种药物相互作用。", "基于实例的模型无法考虑历史疾病发展过程。为了解决这个", "Ge和 Gd对所有患者都是相同的", "问题，纵向模型【14.19.27.28]被设计为利用纵向患者病史并", "捕获时间依赖性。现有的纵向模型通常包括两个阶段，首先将", "2.3Medication Recommendation Problem", "已知信息聚合为患者级表示，然后根据它进行药物推荐", "给定患者当前的诊断D程序P.，历史访问信息[Vi,V2..，V-i]", "现有工作的一个问题是，它们没有明确地模拟同一患者的", "药物建议之间的关系。然而，在临床实践中，对同一患者的建", "以及EHR和DDI图Ge和Ga：目标是训练模型（i。e。，COGNet)，", "可以为该患者推荐合适的药物组合M", "议密切相关。例如，对于惠有慢性病的患者，他们可能一生都", "在使用同一种药物。如图E1所示，我们对MIMIC-IⅡI数据集", "Table 1: Notations used in COGNet", "进行了统计分析。对于每次访问，我们计算历史和Jaccard中", "出现的药物在当前药物和过去药物之间的比例。我们可以看", "到，在大多数访问中，以前推荐了很大一部分处方药。受此启", " Notation", "Description", "发，我们重新考虑从药物层面的角度利用历史信息。这里的挑", "R", "Electronic Health Records", "战是如何准确地确定历史药物目前是否仍然相关", "V", "Record for a single patient", "在本文中，我们提出了一种基于编码器-解码器的生成网络，", "D,P,M", "Diagnoses, Procedure and Medication set", "以顺序方式生产适当的药物，称为条件生成网络（COGNet）。", "G*", "EHR or DDI Graph", "8*,A*", "建议的COGNet由基本模型和复制模块组成。基本模型仅根据", "惠者本次就诊时的健康状况进行推荐；复制模块在建模时引", "Edge set and adjacency matrix of graph G", "入历史就诊信息。与基本模型不同，在生成每种药物时，复制", "E*ER/|xs", "Origin embedding tables", "模块决定是从历史推荐中复制药物还是预测新的药物。在公", "EgERIMIxs", "Graph embedding for medications", "共数据集上的实验证明了该模型的有效性。我们总结了我们", "的主要贡献如下：", "3", "", "FRAMEWORK", ":We propose a medication recommendation model, COGNet", "图3说明了COGNet的体系结构。我们采用基于编码器-解码", "which introduces a novel copy-or-predict mechanism. COGNet", "器的生成网络，根据患者当前和历史健康状况生成适当的药", "can leverage historical recommendations to produce a more", "物组合。在编码器中，我们对历史就诊的所有医疗代码进行编", "accuraterecommendation", "码(i。e。, V, V2, ...", "，-1）表示惠者的历史健康状况", ": We develop a novel hierarchical selection mechanism, which", "并对t一次就诊的诊断和程序代码进行编码，以表示患者当前", "choosesthereusablemedicinestocopyfrombothmedication-", "的健康状况。在解码器中，我们将逐一生成药物以组成推荐药", "level and visit-level perspectives. This increases the intepretabil-", "物组合。在每个解码步骤中，解码器结合诊断，程序和生成的", "ity of the proposed COGNet.", "药物的信息，以推荐下一种药物。此外，如果当前访问中的某", ".We conduct comprehensive experiments on a public dataset", "些疾病与过去的访问保持一致，则复制模块将直接从历史药", "MIMIC-III to demonstrate the effectiveness of the proposed", "物组合中复制相应的药物。我们将首先描述基于编码器-解码", "COGNet.", "器的基本生成模型，然后在以下小节中详细介绍复制模块。", "3.1InputRepresentations", "我们设计了三个嵌人表，EaERlDIxs，EpeRIPIxs和EmE", "2", "PROBLEMFORMULATION", "RIMIxs，其中每行是一个嵌入向量，用于不同的诊断代码，程", "2.1 Electrical Health Records (EHR)", "序代码或药物代码。这里s表示嵌人空间的维数。对于每个", "EHR的基本单位是患者，每个患者由多次就诊组成。令R=", "诊断集D;（iE[1,t]），我们首先通过嵌人矩阵Ea将其每个", "(V(i))-表示N患者的纵向EHR。每位患者至少去过一次", "元素dEDi转换为s维向量d。然后，我们可以获得诊断集", "医院，第i位患者可以表示为一系列多元观察结果V(i)", "D;ERID,Ixs的表示。对于每个程序集P;（tE[1,t]）和药物", "[(),(,.,]。这里T(i)是第i位患者的就诊次数。", "集M;（je[1,t-1]），我们也可以分别通过嵌入矩阵Ep和", "为了简化表示法，我们忽略索引i，并使用单个患者描述", "Em获得它们的表示P;eRlPilxs和M,eRIMlxs", "我们的方法，然后将患者表示为V=[Vi.V2，*，V。令", "D=(di,d2,*,dD)表示诊断集，P=(P12,·*·PIP}表示", "3.2BasicModel", "程序集，M=(m1,m2,,mM)表示药物集。|Dl，Pl和", "在基本模型中，我们仅根据当前就诊中患者的健康状况推荐", "IMI分别表示所有可能的诊断，程序和药物的数量。然后，患", "药物组合。基本模型是编码器-解码器生成模型。它由四个模", "者的每次就诊都可以用Vt=Dt,Pt,M)表示，其中DCD", "块组成：诊断编码器，程序编码器，用药图编码器和药物组合", "，PCP和MCM", "解码器" ]

{ "arxiv_id": "2202.06588", "language": "zh", "page_idx": 2 }

[ [ "0.0866", "0.07574", "0.3774", "0.07574", "0.3774", "0.089", "0.0866", "0.089" ], [ "0.6406", "0.0745", "0.911", "0.07574", "0.911", "0.089", "0.6406", "0.0884" ], [ "0.0866", "0.108", "0.3513", "0.108", "0.3513", "0.1218", "0.0866", "0.1218" ], [ "0.5215", "0.108", "0.9126", "0.108", "0.9126", "0.1218", "0.5215", "0.1218" ], [ "0.2385", "0.1276", "0.3284", "0.1276", "0.3284", "0.1427", "0.2385", "0.1427" ], [ "0.4631", "0.1288", "0.482", "0.1288", "0.482", "0.1414", "0.4631", "0.1414" ], [ "0.5215", "0.1237", "0.834", "0.1237", "0.834", "0.1338", "0.5215", "0.1338" ], [ "0.08496", "0.1497", "0.4805", "0.1509", "0.4805", "0.1642", "0.08496", "0.1635" ], [ "0.527", "0.1445", "0.911", "0.1445", "0.911", "0.159", "0.527", "0.159" ], [ "0.08905", "0.1654", "0.165", "0.1654", "0.165", "0.1768", "0.08905", "0.1768" ], [ "0.5195", "0.1674", "0.9126", "0.1674", "0.9126", "0.1818", "0.5195", "0.1818" ], [ "0.11273", "0.1812", "0.482", "0.1812", "0.482", "0.1945", "0.11273", "0.1945" ], [ "0.5195", "0.1849", "0.626", "0.1849", "0.626", "0.1957", "0.5195", "0.1957" ], [ "0.125", "0.1957", "0.4805", "0.1945", "0.4805", "0.2084", "0.125", "0.209" ], [ "0.6216", "0.2052", "0.806", "0.2058", "0.806", "0.2203", "0.6216", "0.2197" ], [ "0.886", "0.2058", "0.915", "0.2058", "0.915", "0.2203", "0.886", "0.2203" ], [ "0.1259", "0.2229", "0.4805", "0.2229", "0.4805", "0.2361", "0.1259", "0.2361" ], [ "0.1259", "0.209", "0.4836", "0.209", "0.4836", "0.2229", "0.1259", "0.2229" ], [ "0.5312", "0.2267", "0.9126", "0.2267", "0.9126", "0.2412", "0.5312", "0.2412" ], [ "0.1259", "0.2361", "0.482", "0.2361", "0.482", "0.25", "0.1259", "0.25" ], [ "0.5195", "0.2424", "0.9126", "0.2424", "0.9126", "0.2556", "0.5195", "0.2556" ], [ "0.1259", "0.2507", "0.4739", "0.2507", "0.4739", "0.2644", "0.1259", "0.2644" ], [ "0.5186", "0.2568", "0.5786", "0.2568", "0.5786", "0.2683", "0.5186", "0.2683" ], [ "0.1111", "0.2632", "0.4805", "0.2632", "0.4805", "0.2766", "0.1111", "0.2766" ], [ "0.125", "0.2766", "0.4836", "0.2778", "0.4836", "0.2922", "0.125", "0.2917" ], [ "0.5186", "0.2842", "0.568", "0.2842", "0.568", "0.2954", "0.5186", "0.2954" ], [ "0.5605", "0.2842", "0.6724", "0.2842", "0.6724", "0.2974", "0.5605", "0.2974" ], [ "0.1274", "0.2922", "0.4795", "0.2922", "0.4795", "0.303", "0.1274", "0.303" ], [ "0.1259", "0.3062", "0.4805", "0.3062", "0.4805", "0.3193", "0.1259", "0.3193" ], [ "0.5186", "0.3005", "0.9126", "0.3005", "0.9126", "0.3145", "0.5186", "0.3145" ], [ "0.1259", "0.3193", "0.4656", "0.3193", "0.4656", "0.3333", "0.1259", "0.3333" ], [ "0.5215", "0.3164", "0.911", "0.3164", "0.911", "0.327", "0.5215", "0.327" ], [ "0.0866", "0.3352", "0.4795", "0.336", "0.4795", "0.3499", "0.0866", "0.3484" ], [ "0.5215", "0.3308", "0.911", "0.3308", "0.911", "0.3416", "0.5215", "0.3416" ], [ "0.0866", "0.3484", "0.2607", "0.3499", "0.2607", "0.363", "0.0866", "0.3623" ], [ "0.5195", "0.3447", "0.911", "0.3447", "0.911", "0.3547", "0.5195", "0.3547" ], [ "0.0989", "0.3591", "0.4844", "0.358", "0.4844", "0.379", "0.0989", "0.3801" ], [ "0.5195", "0.358", "0.911", "0.358", "0.911", "0.3687", "0.5195", "0.3687" ], [ "0.0825", "0.3738", "0.4004", "0.3774", "0.4004", "0.3945", "0.0825", "0.3901" ], [ "0.5195", "0.3718", "0.5474", "0.3718", "0.5474", "0.3833", "0.5195", "0.3833" ], [ "0.5337", "0.3838", "0.914", "0.3838", "0.914", "0.3977", "0.5337", "0.3977" ], [ "0.1896", "0.3984", "0.3774", "0.3984", "0.3774", "0.4167", "0.1896", "0.4167" ], [ "0.517", "0.3977", "0.914", "0.3984", "0.914", "0.4124", "0.517", "0.411" ], [ "0.4631", "0.404", "0.482", "0.404", "0.482", "0.4167", "0.4631", "0.4167" ], [ "0.5195", "0.414", "0.911", "0.414", "0.911", "0.4248", "0.5195", "0.4248" ], [ "0.08496", "0.4204", "0.4795", "0.4216", "0.4795", "0.4363", "0.08496", "0.435" ], [ "0.5195", "0.428", "0.911", "0.428", "0.911", "0.4382", "0.5195", "0.4382" ], [ "0.08826", "0.4382", "0.179", "0.4382", "0.179", "0.4495", "0.08826", "0.4495" ], [ "0.517", "0.44", "0.9126", "0.44", "0.9126", "0.4539", "0.517", "0.4539" ], [ "0.10455", "0.4507", "0.4805", "0.4507", "0.4805", "0.464", "0.10455", "0.464" ], [ "0.517", "0.4526", "0.9126", "0.4539", "0.9126", "0.4673", "0.517", "0.4666" ], [ "0.0866", "0.4653", "0.2418", "0.4653", "0.2418", "0.4792", "0.0866", "0.4792" ], [ "0.5195", "0.4673", "0.9126", "0.4673", "0.9126", "0.481", "0.5195", "0.481" ], [ "0.1584", "0.4849", "0.4084", "0.4849", "0.4084", "0.4988", "0.1584", "0.4988" ], [ "0.517", "0.481", "0.9126", "0.4797", "0.9126", "0.4937", "0.517", "0.4944" ], [ "0.464", "0.488", "0.4805", "0.488", "0.4805", "0.4976", "0.464", "0.4976" ], [ "0.5186", "0.4956", "0.9126", "0.4956", "0.9126", "0.509", "0.5186", "0.509" ], [ "0.0866", "0.505", "0.2484", "0.505", "0.2484", "0.518", "0.0866", "0.518" ], [ "0.5195", "0.5103", "0.607", "0.5103", "0.607", "0.5215", "0.5195", "0.5215" ], [ "0.1013", "0.5195", "0.4805", "0.5195", "0.4805", "0.5327", "0.1013", "0.5327" ], [ "0.517", "0.529", "0.9126", "0.53", "0.9126", "0.5444", "0.517", "0.5435" ], [ "0.08826", "0.535", "0.165", "0.535", "0.165", "0.547", "0.08826", "0.547" ], [ "0.5186", "0.5454", "0.9126", "0.5454", "0.9126", "0.5596", "0.5186", "0.5596" ], [ "0.156", "0.5537", "0.4102", "0.5537", "0.4102", "0.5684", "0.156", "0.5684" ], [ "0.4631", "0.5547", "0.482", "0.5547", "0.482", "0.5674", "0.4631", "0.5674" ], [ "0.5186", "0.5596", "0.5728", "0.5596", "0.5728", "0.571", "0.5186", "0.571" ], [ "0.08496", "0.5757", "0.2467", "0.5757", "0.2467", "0.589", "0.08496", "0.589" ], [ "0.576", "0.579", "0.8555", "0.579", "0.8555", "0.5933", "0.576", "0.5933" ], [ "0.8677", "0.5854", "0.88", "0.5854", "0.88", "0.589", "0.8677", "0.589" ], [ "0.8823", "0.5806", "0.9126", "0.5806", "0.9126", "0.5933", "0.8823", "0.5933" ], [ "0.10297", "0.5923", "0.4805", "0.5923", "0.4805", "0.606", "0.10297", "0.606" ], [ "0.576", "0.6016", "0.8545", "0.603", "0.8545", "0.6206", "0.576", "0.619" ], [ "0.8823", "0.605", "0.9126", "0.607", "0.911", "0.619", "0.8813", "0.617" ], [ "0.5156", "0.6235", "0.9126", "0.625", "0.9126", "0.6426", "0.5156", "0.6416" ], [ "0.08826", "0.606", "0.2133", "0.606", "0.2133", "0.619", "0.08826", "0.619" ], [ "0.2378", "0.6274", "0.33", "0.6274", "0.33", "0.6396", "0.2378", "0.6396" ], [ "0.4631", "0.627", "0.4836", "0.627", "0.4836", "0.64", "0.4631", "0.64" ], [ "0.0866", "0.6465", "0.2361", "0.6465", "0.2361", "0.66", "0.0866", "0.66" ], [ "0.5186", "0.6416", "0.9126", "0.6416", "0.9126", "0.655", "0.5186", "0.655" ], [ "0.5195", "0.657", "0.79", "0.657", "0.79", "0.6675", "0.5195", "0.6675" ], [ "0.0866", "0.6675", "0.4795", "0.6675", "0.4795", "0.681", "0.0866", "0.681" ], [ "0.705", "0.674", "0.8037", "0.674", "0.8037", "0.6924", "0.705", "0.6924" ], [ "0.0866", "0.6816", "0.482", "0.6816", "0.482", "0.696", "0.0866", "0.696" ], [ "0.6177", "0.6875", "0.709", "0.683", "0.71", "0.6987", "0.619", "0.7026" ], [ "0.886", "0.6875", "0.9126", "0.6875", "0.9126", "0.7007", "0.886", "0.7007" ], [ "0.0866", "0.6963", "0.4795", "0.6963", "0.4795", "0.7104", "0.0866", "0.7104" ], [ "0.746", "0.696", "0.7656", "0.696", "0.7656", "0.709", "0.746", "0.709" ], [ "0.08826", "0.7104", "0.4805", "0.7104", "0.4805", "0.7236", "0.08826", "0.7236" ], [ "0.5156", "0.7153", "0.7344", "0.714", "0.7344", "0.728", "0.5156", "0.729" ], [ "0.08826", "0.7246", "0.1168", "0.7246", "0.1168", "0.7363", "0.08826", "0.7363" ], [ "0.10455", "0.7373", "0.4805", "0.7373", "0.4805", "0.7505", "0.10455", "0.7505" ], [ "0.5186", "0.7363", "0.9126", "0.7363", "0.9126", "0.75", "0.5186", "0.75" ], [ "0.08826", "0.7524", "0.4795", "0.7524", "0.4795", "0.763", "0.08826", "0.763" ], [ "0.517", "0.7505", "0.9126", "0.75", "0.9126", "0.7646", "0.517", "0.765" ], [ "0.08496", "0.763", "0.482", "0.7646", "0.482", "0.779", "0.08496", "0.778" ], [ "0.5186", "0.7676", "0.9126", "0.7676", "0.9126", "0.781", "0.5186", "0.781" ], [ "0.08417", "0.778", "0.1854", "0.78", "0.1854", "0.7944", "0.08417", "0.7925" ], [ "0.5186", "0.784", "0.621", "0.784", "0.621", "0.7954", "0.5186", "0.7954" ], [ "0.2361", "0.7944", "0.33", "0.7944", "0.33", "0.806", "0.2361", "0.806" ], [ "0.4558", "0.7944", "0.482", "0.7944", "0.482", "0.8066", "0.4558", "0.8066" ], [ "0.535", "0.7974", "0.9126", "0.7974", "0.9126", "0.811", "0.535", "0.811" ], [ "0.08826", "0.8125", "0.482", "0.8125", "0.482", "0.8257", "0.08826", "0.8257" ], [ "0.5186", "0.8125", "0.614", "0.8125", "0.614", "0.8237", "0.5186", "0.8237" ], [ "0.664", "0.827", "0.767", "0.827", "0.767", "0.8413", "0.664", "0.8413" ], [ "0.886", "0.8257", "0.9126", "0.8276", "0.911", "0.8413", "0.884", "0.839" ], [ "0.5146", "0.8477", "0.6406", "0.8433", "0.642", "0.8613", "0.5156", "0.8657" ], [ "0.08826", "0.8276", "0.2852", "0.8276", "0.2852", "0.8384", "0.08826", "0.8384" ], [ "0.1569", "0.8486", "0.411", "0.8486", "0.411", "0.8623", "0.1569", "0.8623" ], [ "0.455", "0.8496", "0.482", "0.8496", "0.482", "0.8623", "0.455", "0.8623" ], [ "0.08826", "0.8687", "0.4805", "0.8687", "0.4805", "0.882", "0.08826", "0.882" ], [ "0.535", "0.8687", "0.911", "0.8687", "0.911", "0.882", "0.535", "0.882" ], [ "0.08826", "0.8853", "0.4795", "0.8853", "0.4795", "0.896", "0.08826", "0.896" ], [ "0.5156", "0.882", "0.9126", "0.8833", "0.9126", "0.898", "0.5156", "0.8965" ] ]

[ "Conditional Generation Net for Medication Recommendation", "wwW22April229, 222, Virtual vent, yon,Fran", "还可以通过过程编码器获得过程表示P：", "中起辅助作用。为此，我们调整患者的健康状况，以对未发现", "P, = Encp(Pr)", "(5)", "的疾病进行自适应建模，以指导下一个药物推荐。", "3.2.3MedicationGraphEncoder.此编码器旨在对两种药物关", "M, =LayerNorm(M,+MH(M,Dt,D)+MH(M, Pt,P)) (12)", "系进行建模", "最后，我们使用M，的最后一行（i。e。，M.i-i）通过MLP层", ". Medication EHR co-occurrence relationship : Some medica-", "预测第i种药物。", "tions are often prescribed together for better efficacy.For", "Prg=Softmax(M,i-Wg+bg)", "(13)", "bination used to treat acute gastritis. Hence, modeling these", "example, ranitidine and sucralfate are the popular drug com-", "其中WaERsxIMI和bgERIMI是可学习的参数。Pr。表示", "co-occurrence relations can help the decoder to recommend", "词汇表中所有药物的概率，其中最大概率的药物是步骤i中的", "the drug according to the partially generated medications.", "预测药物", ".DDI : As mentioned above, some medications have the DDI", "and can not be used together. When recommending the drug.", "3.3", "Copy Module", "the decoder should avoid that it is conflict with thepast rec", "ommended drugs. Hence, modeling this relation can help to", "在基本模型部分，我们介绍了基于编码器-解码器的模型，以", "recommend a safe and effective medication combination", "根据患者当前的健康状况生成药物推荐结果。它不考虑患者的", "受E[19]的启发，我们使用图卷积网络E[12]基于药物EHR和", "历史访问信息。在本节中，我们设计了一个新颖的复制模块来", "DDI图对两种关系进行建模", "扩展基本模型，该模块首先比较当前和历史就诊的健康状况", "给定输人药物特征XERIMIXs和药物图形邻接矩阵AE", "然后根据病情变化复制可重复使用的药物以开出当前就诊的", "RIMIxIMI，GCN(）将获得新的药物表示如下：", "处方", "由于患者可能有多次就诊，因此我们使用分层选择机制在", "GCN(X, A) = α(O-± AO- x)", "每个解码步骤进行复制过程。首先，我们使用访问级别选择，", "(6)", "通过比较他们的健康状况来选择类似的访问。然后，在特定的", "A=A+I，I是单位矩阵，O是MATHX矩阵的对角度。e。", "解码步骤中，我们使用药物级别选择从所选就诊的处方中挑选", "Oii =Z, Aij)", "特定药物，并将其添加到推荐结果中。最后，我们重复上述过", "首先，我们使用两层GCN基于EHR图邻接矩阵Ae对药物", "程，迭代复制可重复使用的药物，形成完整的推荐用药清单。", "EHR共现关系进行建模。", "请注意，选择过程像注意机制E[26]一样是“软”的，并为所", "Ge = GCN(ReLU(GCN(Em, Ae)W9, Ae)", "有访问/药物分配不同的概率以突出显示选择。Wewilltakethei", "(7)", "-thmedicationrecommendationofthet-thvisit(i。e。，Mti）为例", "，其中wg是可学习参数", "说明复制过程", "然后，我们使用另一个两层GCN基于DDI图Aa对DDI关", "3.3.1Visit-level Selection：我们首先使用两个门控聚合层，通", "系进行建模。", "过分别聚合其诊断和过程表示来编码所有访问的访问级别运", "Gd = GCN(ReLU(GCN(Em, Aa)Wg, Aa)", "(8)", "行状况：", "，其中W9是可学习参数", "ug = Softmax(tanh(D,w + b,)W? + b2) T D", "", "(14)", "最后，我们融合了两个药物关系表示Ge和Gd，以获得关", "up = Softmax(tanh(P'w, +b,)wp + bp)Tp,", "(15)", "其中W,eR$x，bleRo，b2eR和THMAXtrainable。然后", "系感知的药物表示。", "Eg = Ge- AGd", "(9)", "，其中入是可学习参数", "我们通过测量过去的j访问（1≤j≤t-1）与当前的t访问", "之间的相似性来计算其访问级别选择得分。", "3.2.4MedicationCombinationDecoder.解码器将为t-次访问", "og.od+of.op", "逐一推荐药物。例如，在第i次解码步骤中，解码器将根据部", "cj = Softmax(", "(16)", "分生成的药物组合(Mt,o,..,Mt,i-1)和第t次访问中的患者", "Vs", "健康状况推荐下一种药物Mti(i。e。、D和P)。解码过程", "cjeR表示第j次访问的选择得分", "如下", "首先，我们通过原始嵌入矩阵Em和关系表示矩阵Eg将所", "3.3.2Medication-levelSelection.我们使用公式（12）中制定的", "有生成的药物转换为向量，并可以获得部分生成的药物组合", "隐藏状态M.i-1，该状态对诊断，程序和推荐药物的信息进", "Mm和M9的两个表示。我们融合两个表示以获得关系感知的", "行了全面编码，以确定哪些历史药物mEU-IM,在当前情", "药物组合表示。", "况下可重复使用", "M, = Mm +M", "(10)", "我们首先通过第3节中的类似过程获得所有过去访问的药", "然后，我们使用多头自我注意机制（在公式（2）中定义）来", "物表示。2。1：", "M,= Encm(M)", "(17)", "其中 M,eRIMiIxs", "捕获推荐药物之间的相互作用。", "M, = LayerNorm(M, + MH(M,, M,, M.)", "(11)", "直观地，药物推荐任务旨在生成可以覆盖患者所有疾病的药", "然后，我们使用隐藏状态Mt.i-1作为查询向量来计算沿", "物推荐。因此，惠者的疾病和程序表示D，和P，在解码过程", "药物维度的选择得分。第j次访问Mi,k（1≤j≤t-1和" ]

{ "arxiv_id": "2202.06588", "language": "zh", "page_idx": 4 }

[ [ "0.1755", "0.0956", "0.8228", "0.0956", "0.8228", "0.10986", "0.1755", "0.10986" ], [ "0.351", "0.1354", "0.6514", "0.1354", "0.6514", "0.1484", "0.351", "0.1484" ], [ "0.2957", "0.1509", "0.705", "0.152", "0.705", "0.1663", "0.2957", "0.165" ], [ "0.3242", "0.1687", "0.678", "0.1687", "0.678", "0.1829", "0.3242", "0.1829" ], [ "0.4458", "0.1853", "0.5557", "0.1853", "0.5557", "0.1965", "0.4458", "0.1965" ], [ "0.3105", "0.2019", "0.7036", "0.2019", "0.7036", "0.2156", "0.3105", "0.2156" ], [ "0.2627", "0.2542", "0.3418", "0.2542", "0.3418", "0.2683", "0.2627", "0.2683" ], [ "0.513", "0.2546", "0.8774", "0.2546", "0.8774", "0.2683", "0.513", "0.2683" ], [ "0.513", "0.2708", "0.879", "0.2708", "0.879", "0.2852", "0.513", "0.2852" ], [ "0.1453", "0.2832", "0.4575", "0.2832", "0.4575", "0.2976", "0.1453", "0.2976" ], [ "0.513", "0.2874", "0.881", "0.2874", "0.881", "0.3005", "0.513", "0.3005" ], [ "0.1453", "0.2964", "0.4568", "0.2964", "0.4568", "0.31", "0.1453", "0.31" ], [ "0.513", "0.3027", "0.6978", "0.3027", "0.6978", "0.3164", "0.513", "0.3164" ], [ "0.1453", "0.3123", "0.4592", "0.3123", "0.4592", "0.326", "0.1453", "0.326" ], [ "0.532", "0.3196", "0.881", "0.3196", "0.881", "0.3333", "0.532", "0.3333" ], [ "0.147", "0.3267", "0.4575", "0.3267", "0.4575", "0.3403", "0.147", "0.3403" ], [ "0.147", "0.3403", "0.4568", "0.3403", "0.4568", "0.354", "0.147", "0.354" ], [ "0.5146", "0.3374", "0.879", "0.3374", "0.879", "0.3474", "0.5146", "0.3474" ], [ "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" ], [ "0.1436", "0.3694", "0.4575", "0.3682", "0.4575", "0.3818", "0.1436", "0.383" ], [ "0.511", "0.3665", "0.881", "0.3677", "0.881", "0.3818", "0.511", "0.3813" ], [ "0.1453", "0.3843", "0.4575", "0.3843", "0.4575", "0.398", "0.1453", "0.398" ], [ "0.5146", "0.3843", "0.881", "0.3843", "0.881", "0.398", "0.5146", "0.398" ], [ "0.147", "0.398", "0.4568", "0.398", "0.4568", "0.411", "0.147", "0.411" ], [ "0.5146", "0.3997", "0.881", "0.3997", "0.881", "0.4133", "0.5146", "0.4133" ], [ "0.1453", "0.411", "0.4575", "0.411", "0.4575", "0.4246", "0.1453", "0.4246" ], [ "0.5146", "0.4175", "0.879", "0.4175", "0.879", "0.4275", "0.5146", "0.4275" ], [ "0.147", "0.4246", "0.4568", "0.4258", "0.4568", "0.4395", "0.147", "0.4382" ], [ "0.513", "0.4316", "0.881", "0.4316", "0.881", "0.4453", "0.513", "0.4453" ], [ "0.1453", "0.4407", "0.4575", "0.4407", "0.4575", "0.4536", "0.1453", "0.4536" ], [ "0.1453", "0.4531", "0.4575", "0.4536", "0.4575", "0.4685", "0.1453", "0.4673" ], [ "0.513", "0.449", "0.881", "0.449", "0.881", "0.462", "0.513", "0.462" ], [ "0.1436", "0.466", "0.4575", "0.4685", "0.4575", "0.4822", "0.1436", "0.4797" ], [ "0.513", "0.4656", "0.881", "0.4656", "0.881", "0.4785", "0.513", "0.4785" ], [ "0.147", "0.4827", "0.4568", "0.4827", "0.4568", "0.4963", "0.147", "0.4963" ], [ "0.511", "0.4785", "0.882", "0.4797", "0.882", "0.4946", "0.511", "0.4934" ], [ "0.1453", "0.4963", "0.4592", "0.4963", "0.4592", "0.5103", "0.1453", "0.5103" ], [ "0.511", "0.4963", "0.66", "0.4963", "0.66", "0.5103", "0.511", "0.5103" ], [ "0.1453", "0.511", "0.2007", "0.511", "0.2007", "0.5254", "0.1453", "0.5254" ], [ "0.532", "0.5127", "0.881", "0.5127", "0.881", "0.5254", "0.532", "0.5254" ], [ "0.513", "0.5293", "0.881", "0.5293", "0.881", "0.542", "0.513", "0.542" ], [ "0.11584", "0.5405", "0.1453", "0.5405", "0.1453", "0.552", "0.11584", "0.552" ], [ "0.1503", "0.538", "0.2627", "0.539", "0.2627", "0.554", "0.1503", "0.5527" ], [ "0.513", "0.5444", "0.881", "0.5444", "0.881", "0.558", "0.513", "0.558" ], [ "0.511", "0.56", "0.879", "0.56", "0.879", "0.5737", "0.511", "0.5737" ], [ "0.1184", "0.5664", "0.4878", "0.5664", "0.4878", "0.58", "0.1184", "0.58" ], [ "0.11755", "0.582", "0.4844", "0.582", "0.4844", "0.5957", "0.11755", "0.5957" ], [ "0.513", "0.5776", "0.8774", "0.5776", "0.8774", "0.5884", "0.513", "0.5884" ], [ "0.1184", "0.5986", "0.4844", "0.5986", "0.4844", "0.6123", "0.1184", "0.6123" ], [ "0.511", "0.5933", "0.882", "0.5933", "0.882", "0.607", "0.511", "0.607" ], [ "0.511", "0.608", "0.881", "0.6094", "0.881", "0.6235", "0.511", "0.6226" ], [ "0.11755", "0.6143", "0.486", "0.6133", "0.486", "0.627", "0.11755", "0.6274" ], [ "0.1184", "0.6313", "0.4878", "0.6313", "0.4878", "0.6445", "0.1184", "0.6445" ], [ "0.511", "0.626", "0.7173", "0.626", "0.7173", "0.6396", "0.511", "0.6396" ], [ "0.11755", "0.6455", "0.4844", "0.6455", "0.4844", "0.659", "0.11755", "0.659" ], [ "0.532", "0.641", "0.7407", "0.641", "0.7407", "0.655", "0.532", "0.655" ], [ "0.1184", "0.663", "0.4844", "0.663", "0.4844", "0.677", "0.1184", "0.677" ], [ "0.532", "0.6577", "0.882", "0.6577", "0.882", "0.673", "0.532", "0.673" ], [ "0.12006", "0.6787", "0.4844", "0.6787", "0.4844", "0.6924", "0.12006", "0.6924" ], [ "0.5474", "0.6733", "0.882", "0.675", "0.882", "0.6895", "0.5474", "0.6885" ], [ "0.11755", "0.6934", "0.4844", "0.6934", "0.4844", "0.707", "0.11755", "0.707" ], [ "0.5474", "0.6895", "0.7314", "0.6895", "0.7314", "0.703", "0.5474", "0.703" ], [ "0.531", "0.703", "0.8833", "0.705", "0.8833", "0.7227", "0.531", "0.72" ], [ "0.11755", "0.7104", "0.4844", "0.7104", "0.4844", "0.7236", "0.11755", "0.7236" ], [ "0.11755", "0.728", "0.3257", "0.728", "0.3257", "0.741", "0.11755", "0.741" ], [ "0.5483", "0.7227", "0.882", "0.7227", "0.882", "0.737", "0.5483", "0.737" ], [ "0.5474", "0.737", "0.858", "0.7383", "0.858", "0.753", "0.5474", "0.752" ], [ "0.1368", "0.7437", "0.4844", "0.7437", "0.4844", "0.7573", "0.1368", "0.7573" ], [ "0.531", "0.753", "0.882", "0.7544", "0.882", "0.7686", "0.531", "0.767" ], [ "0.1184", "0.7603", "0.486", "0.7603", "0.486", "0.774", "0.1184", "0.774" ], [ "0.5483", "0.771", "0.882", "0.771", "0.882", "0.785", "0.5483", "0.785" ], [ "0.1184", "0.777", "0.4844", "0.777", "0.4844", "0.7905", "0.1184", "0.7905" ], [ "0.5483", "0.786", "0.879", "0.786", "0.879", "0.8", "0.5483", "0.8" ], [ "0.1184", "0.792", "0.486", "0.792", "0.486", "0.8057", "0.1184", "0.8057" ], [ "0.55", "0.8027", "0.882", "0.8027", "0.882", "0.818", "0.55", "0.818" ], [ "0.12006", "0.81", "0.471", "0.81", "0.471", "0.821", "0.12006", "0.821" ], [ "0.4685", "0.8135", "0.4785", "0.8135", "0.4785", "0.818", "0.4685", "0.818" ], [ "0.12006", "0.8247", "0.4844", "0.8247", "0.4844", "0.8384", "0.12006", "0.8384" ], [ "0.5483", "0.8193", "0.764", "0.8193", "0.764", "0.833", "0.5483", "0.833" ], [ "0.11755", "0.8403", "0.486", "0.8403", "0.486", "0.854", "0.11755", "0.854" ], [ "0.509", "0.8457", "0.7983", "0.8467", "0.7983", "0.861", "0.509", "0.86" ], [ "0.12006", "0.858", "0.3745", "0.858", "0.3745", "0.868", "0.12006", "0.868" ], [ "0.1377", "0.8745", "0.4844", "0.8745", "0.4844", "0.8853", "0.1377", "0.8853" ], [ "0.511", "0.8716", "0.881", "0.8726", "0.881", "0.887", "0.511", "0.8857" ], [ "0.1184", "0.8896", "0.486", "0.8896", "0.486", "0.903", "0.1184", "0.903" ], [ "0.511", "0.8896", "0.881", "0.8896", "0.881", "0.903", "0.511", "0.903" ], [ "0.11755", "0.904", "0.4878", "0.905", "0.4878", "0.9194", "0.11755", "0.9185" ], [ "0.511", "0.9062", "0.882", "0.9062", "0.882", "0.9194", "0.511", "0.9194" ] ]

[ "揭示多注释过程：检香注释数量和实例难度对模型性能的影响", "Pritam Kadasi and Mayank Singh", "Department of Computer Science and Engineering", "Indian Institute of Technology Gandhinagar", "Gujarat, India", "(pritam.k,singh.mayank @iitgn.ac.in", "Abstract", "有效利用可用资源。虽然一些研究(Wanetal.", "2023；Zhangetal.，2021）已经提供了关于找到", "The NLP community has long advocated for", "最佳注释者数量的见解和建议，但这个问题的", "the construction of multi-annotator datasets to", "最终解决方案尚未实现", "better capture the nuances of language inter-", "另一个挑战是每个实例可用的注释数量有", "pretation, subjectivity, and ambiguity. This pa", "per conducts a retrospective study to show how", "限，通常不超过6一10，即使有大量招募的注", "performance scores can vary when a dataset", "释者（Plank，2022）。此限制源于大量实例所", "expands from a single annotation per instance", "需的大量注释工作。因此，当在此类数据集", "to multiple annotations. We propose a novel", "上训练模型时，它们仅捕获注释器池的一小", "multi-annotator simulation process to generate", "部分的意见和信息。此外，某些数据集尚未", "datasets with varying annotation budgets. We", "发布注释器特定的标签或未建立到单个注释", "show that similar datasets with the same anno", "器 (Nie et al., 2020; Jigsaw, 2018; Davidson et al.,", "tation budget can lead to varying performance", "gains. Our findings challenge the popular be-", "2017）的映射。然而，这种趋势正在逐渐转", "lief that models trained on multi-annotation ex", "变，人们越来越认识到应该提供注释器级标", "amples always lead to better performance than", "签 (Prabhakaran et al., 2021; Basile et al., 2021;", "models trained on single or few-annotation ex-", "Denton et al., 2021)", "amples.", "本研究旨在通过模拟多注释过程来解决缺少", "注释器特定标签的挑战。通过这项研究，我们", "1", "Introduction", "提供了有关如何包含更多注释者可以引入模", "型性能变化的见解，并确定了影响这种变化的", "创建数据集的过程通常涉及时间，资源和预算", "等实际限制，这些限制限制了可用于收集注释", "因素。考虑到以前的研究（Swayamdiptaetal.", "的注释者或专家的数量(Shengetal.,2008）。因", "2020)强调了单个实例难度对模型性能的影响，", "我们研究了添加更多注释如何改变实例的难度", "此，在收集的数据中，每个实例普遍存在单个", "或几个标签（取决于注释器的有限数量）。然", "级别，从而影响模型性能", "而，这些数据集上的训练模型对其泛化能力", "总之，我们的主要贡献是：", "提出了挑战，主要是因为数据缺乏多样性。由", ": We propose a novel multi-annotator simula-", "于训练数据（Basileetal.，2021;Plank，2022）缺", "tion process to address the issue of missing", "乏不同的视角和变化，模型可能难以学习稳健", "annotator-specific labels", ". We demonstrate, that increasing the number", "的表示形式，并且无法有效地推广(Nieetal..", "2020; Meissner et al., 2021)", "of annotations per instance does not necessar-", "ily result in significant performance gains.", "为了应对这些挑战，NLP社区强调了利用多", ". We also demonstrate, that altering the num-", "注释器数据集（Davanietal.，2022）的优势，还", "ber of annotations per instance has a notice-", "强调了发布多注释器数据集和相关信息（文化", "able impact on the difficulty of instances as", "和人口统计等。）(Sap et al.,2022;Hershcovich", "perceived by the model and consequently af-", "etal.2022）。然而，这种方法引人了自已的", "", "系列挑战。使用多个注释器收集数据需要大量", "fects the model performance.", "时间、注释预算和注释器专业知识，以确保创", "2The Multi-annotated Dataset", "建具有不同视角的高质量数据集", "此外，在注释预算有限的情况下，确定给定", "在实际情况下，注释过程从雇用一个或多个注", "约束内注释器的最佳数量变得至关重要。这", "释者开始，这些注释者对数据集中的每个实例", "不仅有助于节省注释时间和预算，还可以确保", "进行注释。为了增强真实标签分布的表示，我" ]

{ "arxiv_id": "2310.14572", "language": "zh", "page_idx": 1 }

[ [ "0.1184", "0.0861", "0.4844", "0.0861", "0.4844", "0.0998", "0.1184", "0.0998" ], [ "0.532", "0.0861", "0.881", "0.0861", "0.881", "0.10095", "0.532", "0.10095" ], [ "0.11755", "0.10156", "0.4836", "0.10156", "0.4836", "0.1152", "0.11755", "0.1152" ], [ "0.546", "0.10095", "0.882", "0.10156", "0.882", "0.1164", "0.546", "0.1152" ], [ "0.1184", "0.11816", "0.4844", "0.11816", "0.4844", "0.1318", "0.1184", "0.1318" ], [ "0.5474", "0.11816", "0.763", "0.11816", "0.763", "0.1318", "0.5474", "0.1318" ], [ "0.1184", "0.1354", "0.486", "0.1354", "0.486", "0.1484", "0.1184", "0.1484" ], [ "0.531", "0.133", "0.882", "0.1354", "0.882", "0.1484", "0.531", "0.1467" ], [ "0.1184", "0.1509", "0.413", "0.1509", "0.413", "0.1644", "0.1184", "0.1644" ], [ "0.5474", "0.1497", "0.882", "0.1497", "0.882", "0.1633", "0.5474", "0.1633" ], [ "0.1351", "0.1663", "0.4844", "0.165", "0.4844", "0.1799", "0.1351", "0.1812" ], [ "0.5483", "0.1663", "0.7", "0.1663", "0.7", "0.1799", "0.5483", "0.1799" ], [ "0.11755", "0.1812", "0.4836", "0.1812", "0.4836", "0.1953", "0.11755", "0.1953" ], [ "0.1184", "0.199", "0.4844", "0.199", "0.4844", "0.2131", "0.1184", "0.2131" ], [ "0.513", "0.1953", "0.843", "0.1953", "0.843", "0.209", "0.513", "0.209" ], [ "0.11755", "0.2131", "0.4844", "0.2156", "0.4844", "0.2299", "0.11755", "0.228" ], [ "0.5156", "0.2162", "0.5615", "0.2162", "0.5615", "0.228", "0.5156", "0.228" ], [ "0.607", "0.2131", "0.777", "0.2144", "0.777", "0.2286", "0.607", "0.228" ], [ "0.1351", "0.2322", "0.486", "0.2322", "0.486", "0.2452", "0.1351", "0.2452" ], [ "0.613", "0.231", "0.8037", "0.231", "0.8037", "0.2446", "0.613", "0.2446" ], [ "0.11755", "0.2465", "0.4844", "0.2452", "0.4844", "0.26", "0.11755", "0.2612" ], [ "0.1276", "0.2363", "0.1377", "0.2363", "0.1377", "0.2434", "0.1276", "0.2434" ], [ "0.613", "0.2465", "0.869", "0.2465", "0.869", "0.2612", "0.613", "0.2612" ], [ "0.1184", "0.2632", "0.4836", "0.2632", "0.4836", "0.2778", "0.1184", "0.2778" ], [ "0.614", "0.2632", "0.806", "0.2632", "0.806", "0.2766", "0.614", "0.2766" ], [ "0.5156", "0.2778", "0.5767", "0.2798", "0.575", "0.2944", "0.5146", "0.2922" ], [ "0.617", "0.2756", "0.804", "0.2778", "0.804", "0.2957", "0.617", "0.2935" ], [ "0.5205", "0.2935", "0.79", "0.2944", "0.79", "0.3088", "0.5205", "0.308" ], [ "0.1184", "0.2798", "0.486", "0.2798", "0.486", "0.2935", "0.1184", "0.2935" ], [ "0.1184", "0.2957", "0.4844", "0.2957", "0.4844", "0.31", "0.1184", "0.31" ], [ "0.1184", "0.3113", "0.4844", "0.3113", "0.4844", "0.3247", "0.1184", "0.3247" ], [ "0.517", "0.3088", "0.7314", "0.308", "0.7314", "0.326", "0.517", "0.3267" ], [ "0.11755", "0.3267", "0.4844", "0.326", "0.4844", "0.3403", "0.11755", "0.3416" ], [ "0.519", "0.326", "0.6978", "0.326", "0.6978", "0.339", "0.519", "0.339" ], [ "0.1184", "0.3425", "0.4836", "0.3425", "0.4836", "0.3557", "0.1184", "0.3557" ], [ "0.5225", "0.3457", "0.5366", "0.3457", "0.5366", "0.357", "0.5225", "0.357" ], [ "0.5693", "0.3425", "0.76", "0.3416", "0.76", "0.3557", "0.5693", "0.357" ], [ "0.1184", "0.3594", "0.4844", "0.3594", "0.4844", "0.3723", "0.1184", "0.3723" ], [ "0.5205", "0.3599", "0.54", "0.3599", "0.54", "0.3735", "0.5205", "0.3735" ], [ "0.571", "0.3594", "0.7188", "0.3594", "0.7188", "0.3723", "0.571", "0.3723" ], [ "0.1184", "0.376", "0.486", "0.376", "0.486", "0.3896", "0.1184", "0.3896" ], [ "0.519", "0.3772", "0.54", "0.3772", "0.54", "0.3901", "0.519", "0.3901" ], [ "0.5996", "0.376", "0.876", "0.376", "0.876", "0.3896", "0.5996", "0.3896" ], [ "0.1184", "0.3926", "0.378", "0.3926", "0.378", "0.4062", "0.1184", "0.4062" ], [ "0.5205", "0.3914", "0.54", "0.3914", "0.54", "0.4062", "0.5205", "0.4062" ], [ "0.5996", "0.3938", "0.879", "0.3938", "0.879", "0.4038", "0.5996", "0.4038" ], [ "0.5425", "0.408", "0.66", "0.408", "0.66", "0.4216", "0.5425", "0.4216" ], [ "0.2167", "0.4229", "0.2477", "0.4229", "0.2477", "0.434", "0.2167", "0.434" ], [ "0.2585", "0.4229", "0.287", "0.4229", "0.287", "0.434", "0.2585", "0.434" ], [ "0.315", "0.4192", "0.361", "0.4216", "0.3594", "0.4358", "0.3132", "0.434" ], [ "0.5225", "0.4258", "0.5366", "0.4258", "0.5366", "0.437", "0.5225", "0.437" ], [ "0.571", "0.4233", "0.6323", "0.4233", "0.6323", "0.4382", "0.571", "0.4382" ], [ "0.2225", "0.4329", "0.2477", "0.4329", "0.2477", "0.4465", "0.2225", "0.4465" ], [ "0.2603", "0.434", "0.2854", "0.434", "0.2854", "0.4453", "0.2603", "0.4453" ], [ "0.3198", "0.434", "0.3552", "0.434", "0.3552", "0.4465", "0.3198", "0.4465" ], [ "0.5225", "0.4412", "0.5366", "0.4412", "0.5366", "0.4531", "0.5225", "0.4531" ], [ "0.571", "0.4395", "0.677", "0.4395", "0.677", "0.4536", "0.571", "0.4536" ], [ "0.2585", "0.4448", "0.2854", "0.4448", "0.2854", "0.456", "0.2585", "0.456" ], [ "0.2217", "0.4448", "0.2468", "0.4448", "0.2468", "0.4573", "0.2217", "0.4573" ], [ "0.3208", "0.4448", "0.3552", "0.4448", "0.3552", "0.4573", "0.3208", "0.4573" ], [ "0.5693", "0.456", "0.6724", "0.4536", "0.6724", "0.4685", "0.5693", "0.4702" ], [ "0.1838", "0.4453", "0.2107", "0.4453", "0.2107", "0.4573", "0.1838", "0.4573" ], [ "0.5156", "0.4573", "0.538", "0.4573", "0.538", "0.4697", "0.5156", "0.4697" ], [ "0.2167", "0.4739", "0.2494", "0.4739", "0.2494", "0.485", "0.2167", "0.485" ], [ "0.2551", "0.4739", "0.287", "0.4739", "0.287", "0.4863", "0.2551", "0.4863" ], [ "0.3225", "0.4739", "0.361", "0.4739", "0.361", "0.485", "0.3225", "0.485" ], [ "0.5146", "0.4714", "0.6055", "0.4714", "0.6055", "0.4863", "0.5146", "0.4863" ], [ "0.5146", "0.4868", "0.6187", "0.485", "0.6187", "0.5005", "0.5146", "0.503" ], [ "0.1301", "0.513", "0.2225", "0.513", "0.2225", "0.528", "0.1301", "0.528" ], [ "0.2688", "0.5103", "0.3938", "0.5127", "0.3938", "0.528", "0.2688", "0.5254" ], [ "0.531", "0.5225", "0.882", "0.5225", "0.882", "0.536", "0.531", "0.536" ], [ "0.3535", "0.539", "0.392", "0.542", "0.3862", "0.572", "0.3477", "0.5684" ], [ "0.3787", "0.536", "0.3862", "0.536", "0.3862", "0.5435", "0.3787", "0.5435" ], [ "0.513", "0.539", "0.882", "0.539", "0.882", "0.5527", "0.513", "0.5527" ], [ "0.2024", "0.5474", "0.2284", "0.5474", "0.2284", "0.5586", "0.2024", "0.5586" ], [ "0.2418", "0.55", "0.2627", "0.55", "0.2627", "0.558", "0.2418", "0.558" ], [ "0.3955", "0.5464", "0.4114", "0.5464", "0.4114", "0.5547", "0.3955", "0.5547" ], [ "0.2494", "0.5547", "0.2627", "0.5654", "0.251", "0.5737", "0.236", "0.564" ], [ "0.3938", "0.5557", "0.4114", "0.5557", "0.4114", "0.5654", "0.3938", "0.5654" ], [ "0.4492", "0.557", "0.465", "0.557", "0.465", "0.564", "0.4492", "0.564" ], [ "0.5146", "0.5547", "0.881", "0.5547", "0.881", "0.5684", "0.5146", "0.5684" ], [ "0.3628", "0.564", "0.382", "0.5664", "0.3787", "0.576", "0.361", "0.5737" ], [ "0.1184", "0.5654", "0.1453", "0.5654", "0.1453", "0.5767", "0.1184", "0.5767" ], [ "0.3208", "0.5654", "0.351", "0.5654", "0.351", "0.5767", "0.3208", "0.5767" ], [ "0.3938", "0.5654", "0.4114", "0.5654", "0.4114", "0.5737", "0.3938", "0.5737" ], [ "0.451", "0.5654", "0.4668", "0.5654", "0.4668", "0.5723", "0.451", "0.5723" ], [ "0.1646", "0.568", "0.1914", "0.568", "0.1914", "0.5747", "0.1646", "0.5747" ], [ "0.2083", "0.5586", "0.2275", "0.5586", "0.2275", "0.5684", "0.2083", "0.5684" ], [ "0.2822", "0.5684", "0.2896", "0.5664", "0.293", "0.5723", "0.2854", "0.5747" ], [ "0.5146", "0.5693", "0.881", "0.571", "0.881", "0.585", "0.5146", "0.5845" ], [ "0.2032", "0.585", "0.2317", "0.585", "0.2317", "0.5967", "0.2032", "0.5967" ], [ "0.3586", "0.583", "0.3855", "0.586", "0.382", "0.5996", "0.3552", "0.5967" ], [ "0.2418", "0.589", "0.2747", "0.589", "0.2747", "0.5933", "0.2418", "0.5933" ], [ "0.3972", "0.5913", "0.4072", "0.5913", "0.4072", "0.5957", "0.3972", "0.5957" ], [ "0.5146", "0.586", "0.8774", "0.586", "0.8774", "0.601", "0.5146", "0.601" ], [ "0.513", "0.6025", "0.881", "0.6025", "0.881", "0.616", "0.513", "0.616" ], [ "0.1628", "0.6123", "0.4417", "0.6123", "0.4417", "0.626", "0.1628", "0.626" ], [ "0.513", "0.619", "0.879", "0.619", "0.879", "0.633", "0.513", "0.633" ], [ "0.511", "0.6484", "0.55", "0.6484", "0.55", "0.661", "0.511", "0.661" ], [ "0.5425", "0.6465", "0.66", "0.6484", "0.66", "0.663", "0.5425", "0.661" ], [ "0.1184", "0.6562", "0.1335", "0.6562", "0.1335", "0.665", "0.1184", "0.665" ], [ "0.1511", "0.654", "0.4307", "0.654", "0.4307", "0.6675", "0.1511", "0.6675" ], [ "0.1486", "0.6714", "0.215", "0.6714", "0.215", "0.683", "0.1486", "0.683" ], [ "0.511", "0.673", "0.5576", "0.673", "0.5576", "0.687", "0.511", "0.687" ], [ "0.553", "0.673", "0.629", "0.675", "0.629", "0.6895", "0.553", "0.687" ], [ "0.1184", "0.6953", "0.486", "0.6953", "0.486", "0.709", "0.1184", "0.709" ], [ "0.511", "0.6934", "0.879", "0.6934", "0.879", "0.707", "0.511", "0.707" ], [ "0.1184", "0.7104", "0.4844", "0.7104", "0.4844", "0.7236", "0.1184", "0.7236" ], [ "0.5146", "0.7114", "0.881", "0.7114", "0.881", "0.7246", "0.5146", "0.7246" ], [ "0.1184", "0.727", "0.4836", "0.727", "0.4836", "0.7407", "0.1184", "0.7407" ], [ "0.513", "0.727", "0.8716", "0.727", "0.8716", "0.741", "0.513", "0.741" ], [ "0.867", "0.7324", "0.873", "0.7324", "0.873", "0.737", "0.867", "0.737" ], [ "0.11755", "0.7437", "0.486", "0.7437", "0.486", "0.7573", "0.11755", "0.7573" ], [ "0.511", "0.7437", "0.882", "0.7437", "0.882", "0.7573", "0.511", "0.7573" ], [ "0.1184", "0.7603", "0.486", "0.7603", "0.486", "0.774", "0.1184", "0.774" ], [ "0.513", "0.759", "0.882", "0.759", "0.882", "0.774", "0.513", "0.774" ], [ "0.1184", "0.7754", "0.486", "0.7754", "0.486", "0.789", "0.1184", "0.789" ], [ "0.513", "0.7754", "0.881", "0.7754", "0.881", "0.789", "0.513", "0.789" ], [ "0.1184", "0.7915", "0.4844", "0.7915", "0.4844", "0.8047", "0.1184", "0.8047" ], [ "0.513", "0.7915", "0.882", "0.7915", "0.882", "0.8057", "0.513", "0.8057" ], [ "0.1184", "0.808", "0.4844", "0.808", "0.4844", "0.822", "0.1184", "0.822" ], [ "0.513", "0.808", "0.881", "0.808", "0.881", "0.822", "0.513", "0.822" ], [ "0.1184", "0.8247", "0.486", "0.8247", "0.486", "0.8384", "0.1184", "0.8384" ], [ "0.51", "0.8237", "0.879", "0.8237", "0.879", "0.8374", "0.51", "0.8374" ], [ "0.1184", "0.8403", "0.486", "0.8403", "0.486", "0.854", "0.1184", "0.854" ], [ "0.513", "0.839", "0.881", "0.839", "0.881", "0.854", "0.513", "0.854" ], [ "0.11755", "0.8555", "0.2007", "0.854", "0.2007", "0.868", "0.11755", "0.8706" ], [ "0.513", "0.855", "0.882", "0.855", "0.882", "0.869", "0.513", "0.869" ], [ "0.1368", "0.8716", "0.486", "0.8716", "0.486", "0.8857", "0.1368", "0.8857" ], [ "0.513", "0.8726", "0.6035", "0.8726", "0.6035", "0.8843", "0.513", "0.8843" ], [ "0.1511", "0.887", "0.486", "0.888", "0.486", "0.903", "0.1511", "0.902" ], [ "0.531", "0.888", "0.881", "0.888", "0.881", "0.902", "0.531", "0.902" ], [ "0.1528", "0.905", "0.4575", "0.905", "0.4575", "0.9185", "0.1528", "0.9185" ], [ "0.513", "0.905", "0.879", "0.905", "0.879", "0.9185", "0.513", "0.9185" ] ]

[ "们可以选择通过招募额外的注释者来扩展此过", ": Subsequently, we randomize these annota", "程。我们继续这个迭代过程，直到超出注释预", "tion lists using a consistent random seed", "算，或者我们观察到模型在预测真实标签分布", "across instances. [Lines 5-6]", "方面的性能饱和。因此，我们获得了分配给此", ".Next,we select thefirstk annotations from", "多注释数据集中每个实例的多个注释", "each randomized list, creating the dataset sub-", "在本研究论文中，多注释器数据集D被正", "set Dk . [Lines 4-8]", "式表征为三元组D=（X,A,Y）。集合X表", "示N文本实例，表示为C1,a2,...,C。集", "Algorithm 1Creation of Annotator Datasets", "A对应于M注释器，表示为a1,a2,.·,aM", "Input:", "X : set of N instances", "注释矩阵Y捕获注释，其中行由X索引，", "CL : list of C class labels", "列由 A索引。具体来说，Y=Y[X;A]=", "", "LC : label counts of shape N × C", "Y[α1,2,..,n;ai,a2,..,aM]。更简单地说", "M : number of annotators", "Output:", "D' = {D1, D2, ..., DM)", "1:AL←GETANNOTATIONLISTO", "条目Y[ci;ai]存储由注释器ai分配给实例ai", "的标签yi,j。此外，注释器集As包括k注释", "器，其中定义了1≤k≤M。因此，限制为", "2: Initialize an empty set D'", "Ak的D的子集表示为Dk=（X,Ab,Y'），其", "3: for k ← 1 to M do", "中Y'=YIX:Al。本文将D称为数据集子", "4:", "Initialize an empty list Y", "集，每个实例具有k注释。图1说明了一个玩", "5: ", "for i← 1 to N do", "具多注释器数据集，展示了M注释器和N实", "6:", "SL ←RANDOMSHUFFLE(AL[i])", "例及其包含2和k注释器的子集", "7:", "Choose first k annotations from AL", "and add it to Y", "[91,1", "91,2", "y1.M", "8:", "end for", "3/2,1", "32,2", "92,M", "9:", "D←(X,Y'", "93,2", "93.1", "93,M", "Add Dk to D", "Y-", "10:", "LUN,1", "UN.2", "UN,MJ", "I1: end for", "12: Return D'", "A2 = [a1,d2]", "Ak = (al,a2,..ak)", "通过采用此算法，我们可以为每个实例生成", "E", "", "k注释，从而解决了现有多注释器数据集中注", "[91,1", "91,2", "91.2", "92", "32,2", "92.", "释器特定标签的限制。通过使用不同的随机种", "y3,1", "D2=", "X,AR", "33,2", "43.4", "X.A", "92.1", "D", "子或参数重复该算法，我们可以创建多个数据", "YN.1", "UN", "N", "LA", "集子集Dk，每个子集包含每个实例的k注释，", "这种灵活性可以生成不同的子集，扩大了我们", "Figure 1: A Toy Multi-Annotator Dataset", "研究中可以探索和分析的多注释器场景的范围", "4", "Experiments", "3.", "Simulating the Multi-annotation", "Process", "4.1", "Datasets", "根据我们目前的知识，值得注意的是，现有的", "我们选择了ChaosNLI数据集（Nieetal2020)", "多注释器数据集通常不包括注释器特定的标", "进行研究，因为它在公开可用的数据集(Plank,", "签。相反，可用信息仅限于每个实例(Nieetal.", "2022）中每个实例包含最多的注释（=100）", "", "2020;Jigsaw,2018;Davidson et al.,2017)的标", "ChaosNLI是一种自然语言推理（NLI）任务数", "签分布。例如，在每个实例具有M注释和三", "据集，以其高歧义性而闻名。此外，ChaosNLI", "个可能的标签的情况下，标签分布通常由列表", "数据集包括子数据集，即ChaosNLI-S和", "[p,,rl表示，其中p、g和r是总和为M的正", "ChaosNLI-M，它们分别是从SNLI(Bowman", "整数。为了解决此约束，我们为多注释器方案", "et al.,2015）和MNLI 匹配的（Williams et al.,", "引人了一个模拟过程，该过程利用了实例级标", "2018）的开发集中提取的子集。另一个子数据", "签分布。我们提出的方法（参见算法1）包含", "集 ChaosNLI-α是根据以下 AbductiveNLI的", "以下步骤：", "整个开发集创建的，称为α-NLI（Bhagavatula", ": Initially, we generate a list of annotations", "et al., 2019)", "for each instance by considering the actual", "ChaosNLI数据集由4,645个实例组成，每个", "instance-level label distribution. [Line 1]", "实例都带有100个新注释。此外，该数据集已" ]

{ "arxiv_id": "2310.14572", "language": "zh", "page_idx": 2 }

[ [ "0.3105", "0.0831", "0.3745", "0.0831", "0.3745", "0.0932", "0.3105", "0.0932" ], [ "0.659", "0.0831", "0.723", "0.0831", "0.723", "0.0932", "0.659", "0.0932" ], [ "0.1368", "0.09265", "0.168", "0.09265", "0.168", "0.10394", "0.1368", "0.10394" ], [ "0.3074", "0.09265", "0.3704", "0.09265", "0.3704", "0.1027", "0.3074", "0.1027" ], [ "0.6616", "0.09265", "0.7246", "0.09265", "0.7246", "0.1027", "0.6616", "0.1027" ], [ "0.785", "0.09265", "0.848", "0.09265", "0.848", "0.1027", "0.785", "0.1027" ], [ "0.4182", "0.0944", "0.4768", "0.0944", "0.4768", "0.10156", "0.4182", "0.10156" ], [ "0.2065", "0.0944", "0.2612", "0.0944", "0.2612", "0.10156", "0.2065", "0.10156" ], [ "0.3535", "0.0998", "0.3728", "0.1027", "0.3704", "0.11224", "0.351", "0.10925" ], [ "0.5415", "0.0944", "0.5977", "0.0944", "0.5977", "0.10156", "0.5415", "0.10156" ], [ "0.2007", "0.10156", "0.22", "0.10156", "0.22", "0.10986", "0.2007", "0.10986" ], [ "0.3047", "0.10156", "0.3225", "0.10156", "0.3225", "0.111", "0.3047", "0.111" ], [ "0.4702", "0.10156", "0.4878", "0.10156", "0.4878", "0.10986", "0.4702", "0.10986" ], [ "0.592", "0.10095", "0.6113", "0.10394", "0.6094", "0.111", "0.5884", "0.1081" ], [ "0.7793", "0.10156", "0.7983", "0.10156", "0.7983", "0.111", "0.7793", "0.111" ], [ "0.839", "0.10156", "0.8564", "0.10156", "0.8564", "0.10986", "0.839", "0.10986" ], [ "0.4106", "0.1027", "0.43", "0.1027", "0.43", "0.111", "0.4106", "0.111" ], [ "0.251", "0.1027", "0.2703", "0.1027", "0.2703", "0.111", "0.251", "0.111" ], [ "0.531", "0.1027", "0.5483", "0.1027", "0.5483", "0.111", "0.531", "0.111" ], [ "0.656", "0.1027", "0.6733", "0.1027", "0.6733", "0.111", "0.656", "0.111" ], [ "0.714", "0.1027", "0.735", "0.1027", "0.735", "0.111", "0.714", "0.111" ], [ "0.1234", "0.10925", "0.178", "0.111", "0.178", "0.1223", "0.1234", "0.12054" ], [ "0.1855", "0.11224", "0.2822", "0.11224", "0.2822", "0.1223", "0.1855", "0.1223" ], [ "0.288", "0.11224", "0.3862", "0.11224", "0.3862", "0.1223", "0.288", "0.1223" ], [ "0.3955", "0.11224", "0.4417", "0.11224", "0.4417", "0.1223", "0.3955", "0.1223" ], [ "0.455", "0.11224", "0.5015", "0.11224", "0.5015", "0.1223", "0.455", "0.1223" ], [ "0.509", "0.11224", "0.565", "0.11224", "0.565", "0.1223", "0.509", "0.1223" ], [ "0.567", "0.11224", "0.6265", "0.11224", "0.6265", "0.1223", "0.567", "0.1223" ], [ "0.638", "0.11224", "0.756", "0.11224", "0.756", "0.1223", "0.638", "0.1223" ], [ "0.763", "0.11224", "0.8154", "0.11224", "0.8154", "0.1223", "0.763", "0.1223" ], [ "0.8193", "0.1152", "0.827", "0.1152", "0.827", "0.11816", "0.8193", "0.11816" ], [ "0.8237", "0.11224", "0.873", "0.11224", "0.873", "0.1223", "0.8237", "0.1223" ], [ "0.1873", "0.1217", "0.241", "0.1217", "0.241", "0.1318", "0.1873", "0.1318" ], [ "0.2351", "0.1217", "0.2822", "0.1217", "0.2822", "0.1318", "0.2351", "0.1318" ], [ "0.2896", "0.1217", "0.3862", "0.1217", "0.3862", "0.1318", "0.2896", "0.1318" ], [ "0.3972", "0.1194", "0.4417", "0.1217", "0.44", "0.133", "0.3955", "0.1312" ], [ "0.455", "0.1217", "0.5015", "0.1217", "0.5015", "0.133", "0.455", "0.133" ], [ "0.511", "0.1217", "0.567", "0.1217", "0.567", "0.1318", "0.511", "0.1318" ], [ "0.575", "0.1217", "0.6245", "0.1217", "0.6245", "0.1318", "0.575", "0.1318" ], [ "0.635", "0.1217", "0.69", "0.1217", "0.69", "0.1318", "0.635", "0.1318" ], [ "0.7", "0.1217", "0.7515", "0.1217", "0.7515", "0.1318", "0.7", "0.1318" ], [ "0.763", "0.1217", "0.8154", "0.1217", "0.8154", "0.1318", "0.763", "0.1318" ], [ "0.825", "0.1217", "0.873", "0.1217", "0.873", "0.1318", "0.825", "0.1318" ], [ "0.236", "0.1277", "0.2805", "0.1312", "0.2795", "0.1426", "0.2351", "0.1395" ], [ "0.287", "0.1277", "0.388", "0.13", "0.388", "0.1443", "0.287", "0.1426" ], [ "0.1318", "0.1289", "0.1738", "0.1312", "0.1721", "0.1437", "0.1301", "0.1414" ], [ "0.3972", "0.1289", "0.44", "0.1312", "0.4382", "0.1426", "0.3955", "0.1401" ], [ "0.51", "0.1289", "0.567", "0.1312", "0.567", "0.1426", "0.51", "0.1401" ], [ "0.565", "0.1289", "0.6265", "0.1312", "0.6265", "0.1426", "0.565", "0.1401" ], [ "0.7", "0.1289", "0.7515", "0.1312", "0.75", "0.1426", "0.6987", "0.1401" ], [ "0.76", "0.1289", "0.8154", "0.1312", "0.8154", "0.1426", "0.76", "0.1401" ], [ "0.825", "0.1289", "0.873", "0.1312", "0.8716", "0.1426", "0.8237", "0.1401" ], [ "0.1873", "0.13", "0.2393", "0.13", "0.2393", "0.1414", "0.1873", "0.1414" ], [ "0.1368", "0.1223", "0.1704", "0.1223", "0.1704", "0.1312", "0.1368", "0.1312" ], [ "0.4568", "0.1312", "0.5015", "0.1312", "0.5015", "0.1426", "0.4568", "0.1426" ], [ "0.636", "0.13", "0.69", "0.13", "0.69", "0.1414", "0.636", "0.1414" ], [ "0.1838", "0.1372", "0.2822", "0.1401", "0.2822", "0.155", "0.1838", "0.152" ], [ "0.69", "0.1383", "0.7515", "0.1401", "0.7515", "0.152", "0.69", "0.1497" ], [ "0.764", "0.1383", "0.8096", "0.1401", "0.8076", "0.152", "0.763", "0.1497" ], [ "0.827", "0.1383", "0.873", "0.1401", "0.8716", "0.152", "0.825", "0.1497" ], [ "0.3972", "0.1395", "0.4417", "0.1414", "0.44", "0.152", "0.3955", "0.1497" ], [ "0.2896", "0.1395", "0.3862", "0.1401", "0.3862", "0.152", "0.2896", "0.1509" ], [ "0.6396", "0.1401", "0.69", "0.1401", "0.69", "0.1509", "0.6396", "0.1509" ], [ "0.4568", "0.1414", "0.5015", "0.1414", "0.5015", "0.152", "0.4568", "0.152" ], [ "0.1377", "0.1414", "0.1721", "0.1414", "0.1721", "0.1497", "0.1377", "0.1497" ], [ "0.5146", "0.1426", "0.5635", "0.1426", "0.5635", "0.1497", "0.5146", "0.1497" ], [ "0.574", "0.1414", "0.6265", "0.1414", "0.6265", "0.152", "0.574", "0.152" ], [ "0.1234", "0.1497", "0.2837", "0.1497", "0.2837", "0.1603", "0.1234", "0.1603" ], [ "0.2822", "0.152", "0.3862", "0.152", "0.3862", "0.1592", "0.2822", "0.1592" ], [ "0.392", "0.152", "0.4534", "0.152", "0.4534", "0.1592", "0.392", "0.1592" ], [ "0.451", "0.152", "0.565", "0.152", "0.565", "0.1592", "0.451", "0.1592" ], [ "0.565", "0.152", "0.6245", "0.152", "0.6245", "0.1592", "0.565", "0.1592" ], [ "0.6396", "0.152", "0.69", "0.152", "0.69", "0.1592", "0.6396", "0.1592" ], [ "0.69", "0.152", "0.754", "0.152", "0.754", "0.1592", "0.69", "0.1592" ], [ "0.7617", "0.152", "0.8096", "0.152", "0.8096", "0.1592", "0.7617", "0.1592" ], [ "0.8286", "0.1509", "0.875", "0.1509", "0.875", "0.1609", "0.8286", "0.1609" ], [ "0.11584", "0.1704", "0.881", "0.1716", "0.881", "0.1859", "0.11584", "0.1853" ], [ "0.11755", "0.1882", "0.881", "0.1882", "0.881", "0.199", "0.11755", "0.199" ], [ "0.1184", "0.2031", "0.8774", "0.2031", "0.8774", "0.2131", "0.1184", "0.2131" ], [ "0.11755", "0.2156", "0.882", "0.2156", "0.882", "0.2286", "0.11755", "0.2286" ], [ "0.11755", "0.2299", "0.881", "0.2299", "0.881", "0.2434", "0.11755", "0.2434" ], [ "0.1184", "0.2452", "0.3997", "0.2452", "0.3997", "0.2559", "0.1184", "0.2559" ], [ "0.1184", "0.2832", "0.486", "0.2832", "0.486", "0.2964", "0.1184", "0.2964" ], [ "0.511", "0.284", "0.881", "0.284", "0.881", "0.2976", "0.511", "0.2976" ], [ "0.126", "0.2986", "0.4844", "0.2998", "0.4844", "0.3135", "0.126", "0.3123" ], [ "0.511", "0.2986", "0.881", "0.2998", "0.881", "0.3135", "0.511", "0.3123" ], [ "0.1184", "0.3164", "0.486", "0.3164", "0.486", "0.33", "0.1184", "0.33" ], [ "0.513", "0.3164", "0.825", "0.3164", "0.825", "0.33", "0.513", "0.33" ], [ "0.11755", "0.3308", "0.4836", "0.3308", "0.4836", "0.3445", "0.11755", "0.3445" ], [ "0.513", "0.3486", "0.5415", "0.3486", "0.5415", "0.3599", "0.513", "0.3599" ], [ "0.555", "0.3474", "0.8076", "0.3474", "0.8076", "0.361", "0.555", "0.361" ], [ "0.538", "0.3516", "0.552", "0.3499", "0.553", "0.3552", "0.54", "0.357" ], [ "0.12006", "0.3499", "0.2275", "0.3499", "0.2275", "0.361", "0.12006", "0.361" ], [ "0.1377", "0.3677", "0.4836", "0.3677", "0.4836", "0.3777", "0.1377", "0.3777" ], [ "0.552", "0.364", "0.846", "0.3652", "0.846", "0.379", "0.552", "0.3777" ], [ "0.11584", "0.3801", "0.4895", "0.3813", "0.4895", "0.3955", "0.11584", "0.3943" ], [ "0.552", "0.3813", "0.7446", "0.3801", "0.7446", "0.3943", "0.552", "0.3955" ], [ "0.11755", "0.398", "0.486", "0.398", "0.486", "0.411", "0.11755", "0.411" ], [ "0.5146", "0.405", "0.881", "0.405", "0.881", "0.4187", "0.5146", "0.4187" ], [ "0.11755", "0.4146", "0.486", "0.4146", "0.486", "0.4275", "0.11755", "0.4275" ], [ "0.511", "0.4204", "0.881", "0.4204", "0.881", "0.434", "0.511", "0.434" ], [ "0.11755", "0.43", "0.4844", "0.43", "0.4844", "0.4436", "0.11755", "0.4436" ], [ "0.511", "0.4358", "0.879", "0.4358", "0.879", "0.4495", "0.511", "0.4495" ], [ "0.1184", "0.4465", "0.486", "0.4465", "0.486", "0.4602", "0.1184", "0.4602" ], [ "0.5156", "0.4536", "0.879", "0.4536", "0.879", "0.4644", "0.5156", "0.4644" ], [ "0.12177", "0.4656", "0.1276", "0.4656", "0.1276", "0.4739", "0.12177", "0.4739" ], [ "0.511", "0.4673", "0.881", "0.4685", "0.881", "0.4827", "0.511", "0.4822" ], [ "0.1368", "0.4797", "0.4844", "0.4797", "0.4844", "0.4934", "0.1368", "0.4934" ], [ "0.5146", "0.485", "0.881", "0.485", "0.881", "0.4988", "0.5146", "0.4988" ], [ "0.1184", "0.4963", "0.486", "0.4963", "0.486", "0.5103", "0.1184", "0.5103" ], [ "0.513", "0.502", "0.881", "0.502", "0.881", "0.5156", "0.513", "0.5156" ], [ "0.1184", "0.5127", "0.4902", "0.5127", "0.4902", "0.5254", "0.1184", "0.5254" ], [ "0.513", "0.517", "0.881", "0.517", "0.881", "0.531", "0.513", "0.531" ], [ "0.1184", "0.53", "0.4878", "0.53", "0.4878", "0.5405", "0.1184", "0.5405" ], [ "0.513", "0.533", "0.881", "0.533", "0.881", "0.5464", "0.513", "0.5464" ], [ "0.1184", "0.5444", "0.486", "0.5444", "0.486", "0.558", "0.1184", "0.558" ], [ "0.511", "0.5474", "0.881", "0.549", "0.881", "0.5635", "0.511", "0.5625" ], [ "0.1184", "0.561", "0.4836", "0.561", "0.4836", "0.572", "0.1184", "0.572" ], [ "0.511", "0.564", "0.881", "0.5654", "0.881", "0.58", "0.511", "0.579" ], [ "0.11755", "0.5767", "0.486", "0.5767", "0.486", "0.5903", "0.11755", "0.5903" ], [ "0.511", "0.58", "0.882", "0.5806", "0.882", "0.5957", "0.511", "0.594" ], [ "0.11755", "0.593", "0.4878", "0.593", "0.4878", "0.6055", "0.11755", "0.6055" ], [ "0.511", "0.5967", "0.881", "0.597", "0.881", "0.6123", "0.511", "0.611" ], [ "0.1184", "0.6094", "0.486", "0.6094", "0.486", "0.6226", "0.1184", "0.6226" ], [ "0.511", "0.6133", "0.881", "0.6133", "0.881", "0.627", "0.511", "0.627" ], [ "0.1184", "0.626", "0.486", "0.626", "0.486", "0.6396", "0.1184", "0.6396" ], [ "0.511", "0.629", "0.879", "0.629", "0.879", "0.6426", "0.511", "0.6426" ], [ "0.11755", "0.641", "0.4878", "0.641", "0.4878", "0.655", "0.11755", "0.655" ], [ "0.513", "0.6465", "0.8774", "0.6465", "0.8774", "0.6567", "0.513", "0.6567" ], [ "0.1184", "0.6577", "0.486", "0.6577", "0.486", "0.6714", "0.1184", "0.6714" ], [ "0.5146", "0.662", "0.881", "0.662", "0.881", "0.676", "0.5146", "0.676" ], [ "0.1184", "0.673", "0.4844", "0.673", "0.4844", "0.686", "0.1184", "0.686" ], [ "0.513", "0.677", "0.882", "0.6787", "0.882", "0.6924", "0.513", "0.6904" ], [ "0.1184", "0.6895", "0.4878", "0.6895", "0.4878", "0.703", "0.1184", "0.703" ], [ "0.511", "0.6934", "0.837", "0.6934", "0.837", "0.707", "0.511", "0.707" ], [ "0.1184", "0.706", "0.4878", "0.706", "0.4878", "0.7197", "0.1184", "0.7197" ], [ "0.51", "0.7104", "0.9507", "0.709", "0.9507", "0.7227", "0.51", "0.7236" ], [ "0.11755", "0.7217", "0.486", "0.7217", "0.486", "0.7354", "0.11755", "0.7354" ], [ "0.5146", "0.728", "0.614", "0.728", "0.614", "0.7393", "0.5146", "0.7393" ], [ "0.1184", "0.737", "0.4844", "0.737", "0.4844", "0.7505", "0.1184", "0.7505" ], [ "0.534", "0.745", "0.881", "0.745", "0.881", "0.756", "0.534", "0.756" ], [ "0.11584", "0.753", "0.486", "0.7544", "0.486", "0.767", "0.11584", "0.7666" ], [ "0.511", "0.7603", "0.881", "0.7603", "0.881", "0.774", "0.511", "0.774" ], [ "0.1184", "0.7695", "0.486", "0.7695", "0.486", "0.783", "0.1184", "0.783" ], [ "0.5146", "0.778", "0.879", "0.778", "0.879", "0.788", "0.5146", "0.788" ], [ "0.1184", "0.786", "0.4844", "0.786", "0.4844", "0.8", "0.1184", "0.8" ], [ "0.513", "0.792", "0.881", "0.792", "0.881", "0.8057", "0.513", "0.8057" ], [ "0.1184", "0.802", "0.4844", "0.802", "0.4844", "0.8154", "0.1184", "0.8154" ], [ "0.513", "0.808", "0.881", "0.808", "0.881", "0.822", "0.513", "0.822" ], [ "0.1184", "0.8184", "0.1545", "0.8184", "0.1545", "0.8306", "0.1184", "0.8306" ], [ "0.513", "0.8247", "0.879", "0.8247", "0.879", "0.8384", "0.513", "0.8384" ], [ "0.1394", "0.836", "0.4878", "0.836", "0.4878", "0.8496", "0.1394", "0.8496" ], [ "0.513", "0.8413", "0.879", "0.8413", "0.879", "0.855", "0.513", "0.855" ], [ "0.11755", "0.8516", "0.4844", "0.8516", "0.4844", "0.865", "0.11755", "0.865" ], [ "0.511", "0.857", "0.881", "0.857", "0.881", "0.8706", "0.511", "0.8706" ], [ "0.1368", "0.879", "0.4895", "0.88", "0.4895", "0.898", "0.1368", "0.8965" ], [ "0.5146", "0.8745", "0.881", "0.8745", "0.881", "0.8853", "0.5146", "0.8853" ], [ "0.11755", "0.8945", "0.4902", "0.8945", "0.4902", "0.9077", "0.11755", "0.9077" ], [ "0.511", "0.8896", "0.879", "0.8896", "0.879", "0.903", "0.511", "0.903" ], [ "0.1184", "0.9067", "0.1931", "0.9067", "0.1931", "0.9185", "0.1184", "0.9185" ], [ "0.513", "0.905", "0.8833", "0.905", "0.8833", "0.9185", "0.513", "0.9185" ] ]

[ "Min. Accurac", "Max. Accuracy", "Model", "ChaosNLI-M", "ChaosNLI-M", "ChaosNLI-a", "ChaosNLIa", "ChaosNLI-S", "LD", "ChaosNLI-S", "ML", "ML", "LD", "LD", "ML", "LD", "ML", "LD", "ML", "ML", "LD", "RoBERTa", "0.647 (1)0.647 (1)", "0.558(1)0.558(1)", "0.695 (1)", "0.695 (2)", "0.75(100)", "0.741 (20)", "0.719(80)0.731(100)", "0.734 (30)", "", "0.73 (30)", "0.647 (1)", "0.643 (1)", "0.564 (1)0.561 (1)", "0.647(2)", "0.648 (1)", "0.743 (100)", "0.77 (100)", "0.744 (100)", "0.751(80)", "0.679 (30)", "0.685 (30)", "0.639 (1)", "0.568（)0.568(1)", "ALBERT", "0.668 (1)", "0.796(100)", "0.737(100)", "0.751(90)", "0.695 (100)", "0.695 (90)", "0.639 (1)", "XLNet", "0.668 (1)", "0.706 (100)", "0.643()0.643(1)", "0.769(100)", "0.613 (3)", "0.616(3)", "0.598 (6)", "0.579(1)0.579(1)", "0.751 (90)", "0.585 (6)", "BERT", "0.753 (90)", "0.757 (100)", "DistilBERT0.632 (1)0.632 (1)", "0.533(1)0.533(1)", "0.582(70)", "0.584 (90)0.724 (100)", "0.73(100)", "0.692(80)", "0.682(90)", "0.608 (3)", "0.61 (3)", "Table 2: The performance of various models in both the ML and LD settings is presented in this table. Values", "indicate accuracy,and values inbraces indicatek.Thevalueshighlighted inboldindicatethe optimal number of", "annotatorswheretheperformancereachesitspeakcomparedtothemaximumannotationbudgetallocated(1oo)", "Conversely, the highlighted values in the minimum accuracy column indicate the lowest performance achieved", "compared to the minimum budget allocated (1). This information provides insights into the impact of the number", "ofannotatorsonthemodel'sperformance", "分别。不同数据集之间性能增益的显著差异进", "特定任务选择适当数据集的重要性。此外，这", "步强调了增加注释对性能改进的影响是不一", "些发现强调了确定最佳注释者数量的重要性，", "致的。它强调需要仔细分析和理解每个数据集", "因为模型的性能随着注释的增加而变化", "和模型组合的具体特征，以确定注释数量和性", "5.2", "Does the number of annotations", "", "能之间的关系", "为了解释观察到的复杂行为，我们使用", "influence the difficulty of instances as", "V-Information (Ethayarajh et al., 2022)。V-", "perceived by the model?", "information是一种度量，它量化了模型根据", "为了研究这个问题，我们采用了Swayamdipta", "给定输入预测输出的难易程度。V-信息越高，", "etal.（2020）提出的数据集制图的概念，该概", "模型越容易预测给定输入的输出。此外，V-信", "念利用训练动力学来根据实例的（1）置信", "息不能为负，除非模型过拟合等。（请参见$？？", "度来区分实例，该置信度以跨时期的正确标", "", "签的平均概率来衡量，以及（2）可变性，由", "图3提供了跨五个不同PLM的三个数据集", "上述置信度的方差表示。此分析会生成一个", "的V信息分数的可视化表示。正如预期的那", "数据集映射，该映射可识别三个不同的困难", "样，对于ChaosNLI-S和ChaosNLI-M数据集，", "区域：易于学习，难以学习以及相对于训练", "V信息得分更高。在基于LD的性能评估中，", "模型不明确的实例。易于学习（e）实例始", "表现出较高V信息得分的模型也往往会产生", "终表现出高置信度和低可变性，表明该模型", "较高的准确性得分。例如，在ChaosNLI-S数", "可以对其进行正确分类。另一方面，难以学", "据集的准确性方面，RoBERTa优于其他模型", "习的(h）实例具有低置信度和低可变性，表", "（性能相似的XLNet除外）。对于ChaosNLI-S", "明模型在多个时期内难以正确地对它们进行", "数据集，从k=20开始的V信息分数的饱和", "一致分类。模糊(a）实例在真实标签的预测", "有效地解释了在20个注释之后观察到的基于", "概率中显示出高可变性。我们通过合并其他", "LD的精度饱和，如图E2所示。这种现象表明，", "注释来调查这些类别之间的过渡比例。例", "在从实例中提取有价值的见解方面，模型达到", "如，e→a表示所有转换中从易学类别到歧", "了附加注释提供递减回报的地步。因此，超", "义类别的转换所占的比例。Thisprovidesvalu", "过此阈值后，模型的性能将不再显著提高。对", "ableinsightsintotheunderlyingfactorsthatcon", "于ChaosNLI-α数据集，除RoBERTa和XLNet", "tributetotheobservedimprovementsorlackthereofinthemodel", "（V-InformationE[0,0.25]，相对较低）外，所", "sperformance", "有模型的√信息得分3均约为零。这意味着向", "图4说明了ChaosNLI-S和ChaosNLI-M数", "ChaosNLI-α数据集添加更多注释不会在输入", "据集中的一个有趣模式：随着注释数量的增", "和输出标签分布之间建立明确的关系。这一观", "加，很大一部分训练实例从e类别过渡。例", "察表明，对于数据集的这一特定变体，模型可", "如，在1到10个注释之间的所有转换中，超", "能依赖于所提供注释以外的因素来做出准确的", "过60个%涉及从→e类别移动的实例。但是，", "预测", "超过10个注释后，从a类别转换到e的实例", "上述发现表明，在相同预算下训练时，并非", "比例并未显示出大幅增加。另一方面，从e一→", "所有数据集都能产生相似的性能，这强调了为", "a类别的反向过渡是第二常见的过渡，平均比", "3we used same hyperparameters for all k's due to which", "例为20%。随着添加更多注释，从→e的过", "models for k ≤ 3 overfitted resulting in negative V -", "渡与从→a的过渡之间的比例差异变得更加", "Information", "明显（至少29%）。在ChaosNLI-M数据集中，" ]

{ "arxiv_id": "2310.14572", "language": "zh", "page_idx": 4 }

[ [ "0.1748", "0.0341", "0.763", "0.0341", "0.763", "0.04736", "0.1748", "0.04736" ], [ "0.1772", "0.1016", "0.739", "0.1042", "0.739", "0.1256", "0.1772", "0.1237" ], [ "0.1847", "0.1515", "0.3857", "0.1528", "0.3857", "0.1674", "0.1847", "0.1666" ], [ "0.1847", "0.1674", "0.482", "0.1674", "0.482", "0.1812", "0.1847", "0.1812" ], [ "0.1847", "0.1812", "0.3308", "0.1812", "0.3308", "0.1945", "0.1847", "0.1945" ], [ "0.1847", "0.1957", "0.343", "0.1957", "0.343", "0.209", "0.1847", "0.209" ], [ "0.1847", "0.209", "0.5146", "0.209", "0.5146", "0.2229", "0.1847", "0.2229" ], [ "0.4631", "0.2588", "0.5376", "0.2588", "0.5376", "0.2734", "0.4631", "0.2734" ], [ "0.2336", "0.2942", "0.7656", "0.2954", "0.7656", "0.31", "0.2336", "0.3093" ], [ "0.2345", "0.31", "0.63", "0.31", "0.63", "0.324", "0.2345", "0.324" ], [ "0.625", "0.31", "0.7646", "0.31", "0.7646", "0.324", "0.625", "0.324" ], [ "0.2336", "0.324", "0.7656", "0.324", "0.7656", "0.3372", "0.2336", "0.3372" ], [ "0.2345", "0.3384", "0.7646", "0.3384", "0.7646", "0.3516", "0.2345", "0.3516" ], [ "0.2336", "0.3516", "0.7656", "0.3516", "0.7656", "0.3655", "0.2336", "0.3655" ], [ "0.2336", "0.3655", "0.7646", "0.3655", "0.7646", "0.379", "0.2336", "0.379" ], [ "0.2345", "0.3801", "0.7656", "0.3801", "0.7656", "0.3933", "0.2345", "0.3933" ], [ "0.2345", "0.3933", "0.7656", "0.3933", "0.7656", "0.4072", "0.2345", "0.4072" ], [ "0.2345", "0.4072", "0.7656", "0.4072", "0.7656", "0.4204", "0.2345", "0.4204" ], [ "0.2336", "0.4192", "0.7656", "0.4204", "0.7656", "0.435", "0.2336", "0.4343" ], [ "0.232", "0.433", "0.7656", "0.435", "0.7656", "0.449", "0.232", "0.4463" ], [ "0.2336", "0.4475", "0.7656", "0.4475", "0.7656", "0.4622", "0.2336", "0.4622" ], [ "0.2336", "0.4622", "0.7656", "0.4622", "0.7656", "0.476", "0.2336", "0.476" ], [ "0.2336", "0.476", "0.7656", "0.476", "0.7656", "0.4893", "0.2336", "0.4893" ], [ "0.232", "0.488", "0.7656", "0.4893", "0.7656", "0.504", "0.232", "0.5024" ], [ "0.2336", "0.504", "0.7656", "0.504", "0.7656", "0.5176", "0.2336", "0.5176" ], [ "0.2336", "0.5176", "0.7656", "0.5176", "0.7656", "0.532", "0.2336", "0.532" ], [ "0.2336", "0.532", "0.7656", "0.532", "0.7656", "0.5454", "0.2336", "0.5454" ], [ "0.2336", "0.5454", "0.7646", "0.5454", "0.7646", "0.5596", "0.2336", "0.5596" ], [ "0.2336", "0.5596", "0.7656", "0.5596", "0.7656", "0.5728", "0.2336", "0.5728" ], [ "0.2336", "0.5737", "0.7656", "0.5737", "0.7656", "0.587", "0.2336", "0.587" ], [ "0.2336", "0.586", "0.763", "0.586", "0.763", "0.5996", "0.2336", "0.5996" ], [ "0.2345", "0.601", "0.7656", "0.601", "0.7656", "0.6143", "0.2345", "0.6143" ], [ "0.2336", "0.6143", "0.7656", "0.6143", "0.7656", "0.6274", "0.2336", "0.6274" ], [ "0.2336", "0.627", "0.7656", "0.6274", "0.7656", "0.6426", "0.2336", "0.6416" ], [ "0.2336", "0.6426", "0.7646", "0.6426", "0.7646", "0.656", "0.2336", "0.656" ], [ "0.2336", "0.656", "0.767", "0.656", "0.767", "0.6694", "0.2336", "0.6694" ], [ "0.2336", "0.6694", "0.3447", "0.6694", "0.3447", "0.683", "0.2336", "0.683" ], [ "0.1724", "0.7363", "0.1896", "0.7363", "0.1896", "0.75", "0.1724", "0.75" ], [ "0.2091", "0.7373", "0.3162", "0.7373", "0.3162", "0.7505", "0.2091", "0.7505" ], [ "0.1748", "0.7715", "0.2035", "0.7715", "0.2035", "0.784", "0.1748", "0.784" ], [ "0.2173", "0.7715", "0.5596", "0.7715", "0.5596", "0.785", "0.2173", "0.785" ], [ "0.1765", "0.8", "0.8228", "0.8", "0.8228", "0.813", "0.1765", "0.813" ], [ "0.1748", "0.8125", "0.822", "0.813", "0.822", "0.8276", "0.1748", "0.827" ], [ "0.1765", "0.8276", "0.822", "0.8276", "0.822", "0.8413", "0.1765", "0.8413" ], [ "0.1765", "0.8423", "0.8228", "0.8423", "0.8228", "0.857", "0.1765", "0.857" ], [ "0.1765", "0.856", "0.4412", "0.856", "0.4412", "0.8706", "0.1765", "0.8706" ], [ "0.451", "0.856", "0.824", "0.856", "0.824", "0.869", "0.451", "0.869" ], [ "0.1732", "0.869", "0.822", "0.869", "0.822", "0.8833", "0.1732", "0.8833" ], [ "0.1765", "0.884", "0.8228", "0.884", "0.8228", "0.898", "0.1765", "0.898" ], [ "0.1748", "0.898", "0.824", "0.898", "0.824", "0.911", "0.1748", "0.911" ], [ "0.1765", "0.911", "0.794", "0.911", "0.794", "0.925", "0.1765", "0.925" ], [ "0.492", "0.9478", "0.5073", "0.9478", "0.5073", "0.962", "0.492", "0.962" ] ]

[ "To appear at the ICLR 2024 Workshop on Representational Alignment (Re-Align)", "使用语言建模识别和解释非线性人类概念表示", "Wanqian Bao&Uri Hasson", "Center for Mind/Brain Sciences (CIMeC)", "University of Trento", "Rovereto,38068,Italy", "[email protected],[email protected]", "Abstract", "The question of whether people's experience in the world shapes conceptual", "representation and lexical semantics is longstanding.", "Word-association,", "feature-listing and similarity rating tasks are methods that aim to address", "this question but ultimately require a subjective interpretation of the latent", "dimensions or clusters identified. In this study, we introduce a supervised", "representational-alignment method that （i )determines whether two", "groups of individuals share the same basis of a certain category, and", "( ii ) explains in what respects they differ. In applying this method,", "we show that congenital blindness induces conceptual reorganization in", "both a-modal and sensory-related verbal domains, and we identify the", "associated semantic shifts.We first apply supervised feature-pruning", "to a language model (GloVe) to optimize prediction accuracy of human", "similarity judgments from word embeddings. Pruning identifies one subset", "of retained GloVe features that optimizes prediction of judgments made", "by sighted individuals and another subset that optimizes judgments made", "by blind.A linear probing analysis then interprets the latent semantics", "of these feature-subsets by learning a mapping from the retained GloVe", "features to 65 interpretable semantic dimensions.We applied this approach", "to seven semantic domains, including verbs related to motion, sight, touch,", "and amodal verbs related to knowledge acquisition. We find that blind", "individuals more strongly associate social and cognitive meanings to verbs", "related to motion or those communicating non-speech vocal utterances", "(e.g., whimper, moan). Conversely, for amodal verbs, they demonstrate", "much sparser information. Finally, for some verbs, representations of blind", "and sighted are highly similar.The study presents a formal approach", "for studying interindividual differences in word meaning, and the first", "demonstration of how blindness impacts conceptual representation of", "everyday verbs.", "1", "Introduction", "1.1", "Conceptual representation in congenitally blind", "单词是否在不同语言之间或同一语言的不同说话者之间捕获相似含义的问题对于理解语言", "与概念表示之间的关系至关重要。Thompsonetal.（2020）使用一种计算建模方法来表明", "非专业的日常单词的表示形式可能在不同文化之间存在很大差异。。这个问题也强烈适用于", "使用相同语言的人。获取亚文化词典（e。g。，俚语）显然需要学习，并且随着时间的推移与", "含义的变化相关(e。g。，？）。但是，", "一个人的个人经历是否会对诸如“触摸”，“看到”和", "“学习”之类的日常单词产生不同的含义？解决这个问题的一种方法是研究盲人单词含义的", "本质，直接评估世界上的感官体验是否会影响单词含义的知识。事实上，这个问题已经引起", "认知心理学家几十年的兴趣（最近的评论，见Mamusetal.，2023；？）。这项工作产生了大", "量证据，表明盲人使用语言的方式表明对视觉世界的了解以及与视觉相关的单词的使用", "1" ]

{ "arxiv_id": "2403.06204", "language": "zh", "page_idx": 0 }

[ [ "0.1765", "0.0341", "0.763", "0.0341", "0.763", "0.04736", "0.1765", "0.04736" ], [ "0.1765", "0.1061", "0.8228", "0.1061", "0.8228", "0.1193", "0.1765", "0.1193" ], [ "0.1748", "0.11554", "0.4844", "0.11743", "0.4844", "0.1351", "0.1748", "0.1332" ], [ "0.4739", "0.1187", "0.6562", "0.1193", "0.6562", "0.1338", "0.4739", "0.1332" ], [ "0.649", "0.1193", "0.822", "0.1187", "0.822", "0.132", "0.649", "0.1332" ], [ "0.1748", "0.1332", "0.8228", "0.1332", "0.8228", "0.1465", "0.1748", "0.1465" ], [ "0.1765", "0.1465", "0.8228", "0.1465", "0.8228", "0.1604", "0.1765", "0.1604" ], [ "0.1765", "0.1604", "0.8228", "0.1604", "0.8228", "0.1736", "0.1765", "0.1736" ], [ "0.1765", "0.1736", "0.8228", "0.1736", "0.8228", "0.1881", "0.1765", "0.1881" ], [ "0.1765", "0.1875", "0.8203", "0.1875", "0.8203", "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", "0.778", "0.2166", "0.778", "0.2299", "0.1765", "0.2299" ], [ "0.1732", "0.2456", "0.619", "0.2468", "0.619", "0.2615", "0.1732", "0.26" ], [ "0.1772", "0.2727", "0.8228", "0.2727", "0.8228", "0.286", "0.1772", "0.286" ], [ "0.1748", "0.286", "0.822", "0.286", "0.822", "0.2998", "0.1748", "0.2998" ], [ "0.1765", "0.2998", "0.8228", "0.2998", "0.8228", "0.3132", "0.1765", "0.3132" ], [ "0.1772", "0.3132", "0.8228", "0.3132", "0.8228", "0.327", "0.1772", "0.327" ], [ "0.1772", "0.3276", "0.8228", "0.3276", "0.8228", "0.3416", "0.1772", "0.3416" ], [ "0.1765", "0.3416", "0.8228", "0.3416", "0.8228", "0.3547", "0.1765", "0.3547" ], [ "0.1765", "0.3547", "0.8228", "0.3547", "0.8228", "0.3687", "0.1765", "0.3687" ], [ "0.1748", "0.3674", "0.8228", "0.3687", "0.8228", "0.3833", "0.1748", "0.3818" ], [ "0.1772", "0.3833", "0.8228", "0.3833", "0.8228", "0.3965", "0.1772", "0.3965" ], [ "0.1772", "0.3965", "0.8228", "0.3965", "0.8228", "0.4104", "0.1772", "0.4104" ], [ "0.1772", "0.411", "0.8228", "0.411", "0.8228", "0.4248", "0.1772", "0.4248" ], [ "0.1765", "0.4248", "0.822", "0.4248", "0.822", "0.4382", "0.1765", "0.4382" ], [ "0.1765", "0.4382", "0.8228", "0.4382", "0.8228", "0.452", "0.1765", "0.452" ], [ "0.1772", "0.452", "0.8228", "0.452", "0.8228", "0.4653", "0.1772", "0.4653" ], [ "0.1765", "0.4666", "0.8228", "0.4666", "0.8228", "0.4797", "0.1765", "0.4797" ], [ "0.1772", "0.4797", "0.822", "0.4797", "0.822", "0.4937", "0.1772", "0.4937" ], [ "0.1772", "0.4937", "0.8228", "0.4937", "0.8228", "0.507", "0.1772", "0.507" ], [ "0.1772", "0.509", "0.8203", "0.509", "0.8203", "0.5195", "0.1772", "0.5195" ], [ "0.1748", "0.5195", "0.8228", "0.5215", "0.8228", "0.536", "0.1748", "0.534" ], [ "0.1748", "0.535", "0.3718", "0.535", "0.3718", "0.549", "0.1748", "0.549" ], [ "0.1748", "0.5693", "0.1912", "0.5693", "0.1912", "0.584", "0.1748", "0.584" ], [ "0.1936", "0.5747", "0.2009", "0.5747", "0.2009", "0.5776", "0.1936", "0.5776" ], [ "0.2091", "0.5713", "0.2769", "0.5713", "0.2769", "0.5835", "0.2091", "0.5835" ], [ "0.1748", "0.601", "0.2075", "0.601", "0.2075", "0.6123", "0.1748", "0.6123" ], [ "0.2051", "0.604", "0.2115", "0.604", "0.2115", "0.6094", "0.2051", "0.6094" ], [ "0.2173", "0.6016", "0.4395", "0.6016", "0.4395", "0.6123", "0.2173", "0.6123" ], [ "0.1765", "0.6255", "0.8228", "0.6255", "0.8228", "0.6396", "0.1765", "0.6396" ], [ "0.1765", "0.64", "0.824", "0.64", "0.824", "0.654", "0.1765", "0.654" ], [ "0.1765", "0.653", "0.8228", "0.653", "0.8228", "0.6665", "0.1765", "0.6665" ], [ "0.1772", "0.6675", "0.8228", "0.6675", "0.8228", "0.6816", "0.1772", "0.6816" ], [ "0.1765", "0.6816", "0.8228", "0.6816", "0.8228", "0.6963", "0.1765", "0.6963" ], [ "0.1732", "0.6924", "0.8228", "0.6924", "0.8228", "0.7104", "0.1732", "0.7104" ], [ "0.1765", "0.709", "0.8228", "0.709", "0.8228", "0.7227", "0.1765", "0.7227" ], [ "0.179", "0.7485", "0.2467", "0.752", "0.246", "0.7666", "0.1772", "0.763" ], [ "0.5156", "0.7485", "0.5864", "0.7505", "0.5864", "0.765", "0.5156", "0.763" ], [ "0.3743", "0.756", "0.3816", "0.7505", "0.3865", "0.755", "0.38", "0.7603" ], [ "0.2974", "0.763", "0.3718", "0.765", "0.3718", "0.776", "0.2974", "0.7734" ], [ "0.71", "0.759", "0.716", "0.759", "0.716", "0.765", "0.71", "0.765" ], [ "0.2998", "0.7734", "0.3718", "0.7734", "0.3718", "0.784", "0.2998", "0.784" ], [ "0.6284", "0.768", "0.7114", "0.768", "0.7114", "0.779", "0.6284", "0.779" ], [ "0.3228", "0.782", "0.3513", "0.785", "0.3489", "0.7954", "0.3218", "0.7935" ], [ "0.626", "0.779", "0.71", "0.779", "0.71", "0.786", "0.626", "0.786" ], [ "0.642", "0.786", "0.693", "0.786", "0.693", "0.797", "0.642", "0.797" ], [ "0.3308", "0.799", "0.3506", "0.799", "0.3506", "0.811", "0.3308", "0.811" ], [ "0.6626", "0.7974", "0.681", "0.7974", "0.681", "0.811", "0.6626", "0.811" ], [ "0.478", "0.8145", "0.4844", "0.811", "0.4885", "0.8184", "0.482", "0.8213" ], [ "0.6055", "0.8145", "0.613", "0.8145", "0.613", "0.8213", "0.6055", "0.8213" ], [ "0.8022", "0.8145", "0.8096", "0.8145", "0.8096", "0.8184", "0.8022", "0.8184" ], [ "0.6357", "0.8213", "0.808", "0.8213", "0.808", "0.8354", "0.6357", "0.8354" ], [ "0.2009", "0.8228", "0.2524", "0.8228", "0.2524", "0.834", "0.2009", "0.834" ], [ "0.407", "0.8228", "0.4795", "0.8228", "0.4795", "0.8335", "0.407", "0.8335" ], [ "0.3162", "0.8276", "0.361", "0.8276", "0.361", "0.8384", "0.3162", "0.8384" ], [ "0.5195", "0.8164", "0.63", "0.8228", "0.6284", "0.8403", "0.5186", "0.834" ], [ "0.1887", "0.832", "0.267", "0.832", "0.267", "0.8423", "0.1887", "0.8423" ], [ "0.398", "0.832", "0.482", "0.832", "0.482", "0.8423", "0.398", "0.8423" ], [ "0.528", "0.832", "0.603", "0.832", "0.603", "0.8423", "0.528", "0.8423" ], [ "0.626", "0.832", "0.7017", "0.832", "0.7017", "0.8423", "0.626", "0.8423" ], [ "0.7363", "0.832", "0.8", "0.832", "0.8", "0.8423", "0.7363", "0.8423" ], [ "0.71", "0.8354", "0.721", "0.8354", "0.721", "0.839", "0.71", "0.839" ], [ "0.3179", "0.837", "0.3555", "0.837", "0.3555", "0.8477", "0.3179", "0.8477" ], [ "0.3896", "0.8335", "0.3962", "0.8335", "0.3962", "0.837", "0.3896", "0.837" ], [ "0.1871", "0.8413", "0.2695", "0.8413", "0.2695", "0.8516", "0.1871", "0.8516" ], [ "0.3938", "0.8413", "0.4844", "0.8413", "0.4844", "0.853", "0.3938", "0.853" ], [ "0.5376", "0.8413", "0.5947", "0.8413", "0.5947", "0.853", "0.5376", "0.853" ], [ "0.6543", "0.8413", "0.6807", "0.8413", "0.6807", "0.8496", "0.6543", "0.8496" ], [ "0.7485", "0.8413", "0.785", "0.8413", "0.785", "0.8516", "0.7485", "0.8516" ], [ "0.1765", "0.8926", "0.824", "0.8926", "0.824", "0.906", "0.1765", "0.906" ], [ "0.1765", "0.9077", "0.71", "0.9077", "0.71", "0.9185", "0.1765", "0.9185" ], [ "0.4927", "0.9487", "0.5063", "0.9487", "0.5063", "0.9604", "0.4927", "0.9604" ] ]

[ "To appear at the ICLR 2024 Workshop on Representational Alignment (Re-Align)", "然而，在概念组织中，盲人和有远见的区别仍然是悬而未决的问题。已使用多种方法来量化", "盲人的语义知识，包括产生定义（e。g。，", "？），单词关联（e。g。，", "？），事件描述（Mamus", "etal.，2023）和颜色的相似性判断（e。g。，？）等。最近，Bednyetal.（2019）获得了不同类", "型动词的相似性判断，并使用代表性相似性分析和其他描述性技术检查了数据。一个重要", "的发现是，对于某些动词类型，两组之间的相似性判断相似。特别是与视觉刺激相关的动", "词（例如。g。，“闪耀”，“眼”）以一致的方式判断。有趣的是，Lewis et al.（2019)随后", "表明，这些相似性判断是由语言模型（word2vec；？）产生的相似性值（的代理）很好地预测", "的。。作者建议盲人和有视力的人由于对语言的相似接触而学习相似的表示。然而，这些先", "前的研究无法正式，准确地解释盲人和有远见的概念表示在表示日常单词时如何不同", "1.2Current approach: supervised learning of representation", "当前的研究旨在确定盲人和有视力的人在表示动词含义时是否使用相似的信息。为了解决", "这个问题，我们采用了两步计算建模方法，每一步都使用交叉验证进行验证。在步骤1中", "我们使用监督修剪来微调语言模型以预测人类相似性判断。在步骤2中，我们使用探测任", "务检查微调模型特征空间中潜在的信息，研究保留的向量（单词嵌人）和其中包含的信息。", "总而言之，我们研究了基础向量和在经过微调以预测盲人或视力行为的语言模型（的子空", "间）中形成的表示形式。我们在这里对逻辑进行非正式的描述。方法中提供了详细信息，附", "录中给出了更正式的算法演示。图1中总结了工作流，我们在这里引用了该图中的元素。首", "先，我们首先修剪GloVe词嵌入（1）以改善人类相似性判断的预测（2）。这会产生一个基", "于较小的保留特征集的嵌人矩阵（3）。这是针对七个不同类别的动词类型的先天盲人和有视", "力的控件提供的判断分别进行的（预期结果，这些结果为盲人和有视力的人产生了不同的保", "留特征集）。修剪后，我们使用探测任务（Belinkov，2022）检查保留的特征集中的信息。探", "测任务基于偏最小二乘回归（PLSR）函数，并将Binderetal.（2016）提供的数据集作为目", "标，其中包括在65个维度上评级的535个单词（5）。（Chersonietal.，2021）显示，可以", "使用PLSR从其GloVe功能预测这535个单词的人类评分。相反，我们仅使用保留的特征", "（4）和65个人类注释的维度（5）来学习这535个单词的嵌人之间的映射，从而为每个单", "词产生65个维度的预测（6)。此步骤使用为盲人和有视力者保留的特征单独执行。至关重", "要的是，通过将预测值（6）与地面实况值（5）进行比较，探测任务可以量化映射中使用的", "保留特征集中65个语义维度的相对显着性。此外，使用探测任务，我们还可以确定65个", "Binder特征中的哪些是针对盲人和视力不同的预测。在显示它们在交叉验证中的有效表现", "之后，我们使用修剪和探测", "2 ", "", "Method", "2.1", "", "Datasets andword embeddings", "对于相似性判断，我们使用了Bednyetal.（2019）公开提供的数据。25个先天盲人和22个", "有视力的人以1到7的等级对动词对的语义相似性进行了评分（7个最相似）。材料为119", "个动词，分为七类（我们遵循原始术语）。每个类别包含15个动词，除了一个类别包含14", "个。这些动词类别是：感知视觉(e。g。，看，看)，感知触摸(e。g。，轻拍，摩擦)，感知非模", "态（e。g。，学习，发现)，光的发射(e。g。，闪烁，发光)，动画声音的发射（e。g。，呻吟，咆", "哮)，发出无生命的声音(e。g。、喻喻声、哔哔声）和运动动词的方式(e。g。、浮动、反弹）。", "附录中给出了完整的列表。在参与者中对相似性评级进行了标准化，以创建相似性矩阵，并", "Pruning", "Probing", "2", "Human Similarity", "5", "Judgments forN", "535probewords", "verbs", "human-annotatedor", "65features", "Tin", "Iin", "3", "4", "P", "supervisedout535x65prediction", "Full GloVe", "Retained GloVe", "Supervised", "subsetaof cloven", "embeddings for N", "embeddingsforN", "featuresfor535", "Probingusing)", "matrixfromQ", "", "Pruning", "", "verbs (Rfeatures)", "verbs (Q features)", "probewords", "PLSR", "features", "Figure 1: Analysis workflow consisting of supervised pruning and supervised probing. Solid", "linesindicateinputdatasetsand dashedlinesindicateproductsof analyses", "2" ]

{ "arxiv_id": "2403.06204", "language": "zh", "page_idx": 1 }

[ [ "0.1748", "0.0341", "0.761", "0.03284", "0.761", "0.04736", "0.1748", "0.0486" ], [ "0.1748", "0.1061", "0.8203", "0.1061", "0.8203", "0.1193", "0.1748", "0.1193" ], [ "0.1765", "0.1193", "0.8", "0.1193", "0.8", "0.1332", "0.1765", "0.1332" ], [ "0.1765", "0.1401", "0.8228", "0.1401", "0.8228", "0.154", "0.1765", "0.154" ], [ "0.1772", "0.154", "0.822", "0.154", "0.822", "0.1674", "0.1772", "0.1674" ], [ "0.1772", "0.1674", "0.822", "0.1674", "0.822", "0.1812", "0.1772", "0.1812" ], [ "0.1772", "0.1818", "0.822", "0.1818", "0.822", "0.1957", "0.1772", "0.1957" ], [ "0.1765", "0.1945", "0.8", "0.1957", "0.8", "0.2102", "0.1765", "0.209" ], [ "0.1772", "0.2102", "0.8228", "0.2102", "0.8228", "0.2235", "0.1772", "0.2235" ], [ "0.1772", "0.2235", "0.4722", "0.2235", "0.4722", "0.2374", "0.1772", "0.2374" ], [ "0.1748", "0.2456", "0.822", "0.2456", "0.822", "0.2588", "0.1748", "0.2588" ], [ "0.1765", "0.2588", "0.822", "0.2588", "0.822", "0.2734", "0.1765", "0.2734" ], [ "0.1748", "0.2715", "0.8203", "0.2715", "0.8203", "0.286", "0.1748", "0.286" ], [ "0.1732", "0.286", "0.5986", "0.286", "0.5986", "0.2998", "0.1732", "0.2998" ], [ "0.1724", "0.318", "0.357", "0.3208", "0.357", "0.334", "0.1724", "0.332" ], [ "0.1748", "0.3467", "0.3628", "0.3467", "0.3628", "0.361", "0.1748", "0.361" ], [ "0.1772", "0.3718", "0.822", "0.3718", "0.822", "0.385", "0.1772", "0.385" ], [ "0.1765", "0.3833", "0.822", "0.3838", "0.822", "0.3984", "0.1765", "0.3977" ], [ "0.1765", "0.3984", "0.8203", "0.3984", "0.8203", "0.4124", "0.1765", "0.4124" ], [ "0.1772", "0.4136", "0.822", "0.4136", "0.822", "0.4268", "0.1772", "0.4268" ], [ "0.1765", "0.4268", "0.822", "0.4268", "0.822", "0.44", "0.1765", "0.44" ], [ "0.1765", "0.44", "0.822", "0.4414", "0.822", "0.455", "0.1765", "0.4539" ], [ "0.1765", "0.455", "0.8203", "0.455", "0.8203", "0.4685", "0.1765", "0.4685" ], [ "0.1772", "0.4685", "0.822", "0.4685", "0.822", "0.4817", "0.1772", "0.4817" ], [ "0.1772", "0.4841", "0.6626", "0.4841", "0.6626", "0.4944", "0.1772", "0.4944" ], [ "0.6587", "0.4841", "0.8203", "0.4841", "0.8203", "0.4944", "0.6587", "0.4944" ], [ "0.1772", "0.4968", "0.822", "0.4968", "0.822", "0.5103", "0.1772", "0.5103" ], [ "0.179", "0.511", "0.8203", "0.511", "0.8203", "0.5215", "0.179", "0.5215" ], [ "0.1772", "0.5234", "0.822", "0.5234", "0.822", "0.537", "0.1772", "0.537" ], [ "0.1772", "0.537", "0.822", "0.537", "0.822", "0.5503", "0.1772", "0.5503" ], [ "0.1772", "0.5503", "0.822", "0.5503", "0.822", "0.5645", "0.1772", "0.5645" ], [ "0.1772", "0.565", "0.822", "0.565", "0.822", "0.579", "0.1772", "0.579" ], [ "0.1772", "0.5806", "0.7173", "0.5806", "0.7173", "0.5913", "0.1772", "0.5913" ], [ "0.1732", "0.6094", "0.584", "0.6113", "0.584", "0.6255", "0.1732", "0.6235" ], [ "0.1748", "0.6353", "0.8203", "0.6353", "0.8203", "0.6484", "0.1748", "0.6484" ], [ "0.179", "0.6514", "0.8203", "0.6514", "0.8203", "0.662", "0.179", "0.662" ], [ "0.1772", "0.662", "0.8203", "0.662", "0.8203", "0.6753", "0.1772", "0.6753" ], [ "0.1765", "0.678", "0.822", "0.678", "0.822", "0.6914", "0.1765", "0.6914" ], [ "0.1772", "0.6924", "0.822", "0.6924", "0.822", "0.7056", "0.1772", "0.7056" ], [ "0.179", "0.7085", "0.2075", "0.7085", "0.2075", "0.7163", "0.179", "0.7163" ], [ "0.1772", "0.726", "0.8228", "0.726", "0.8228", "0.7393", "0.1772", "0.7393" ], [ "0.1772", "0.7407", "0.822", "0.7407", "0.822", "0.754", "0.1772", "0.754" ], [ "0.1772", "0.754", "0.822", "0.754", "0.822", "0.7676", "0.1772", "0.7676" ], [ "0.179", "0.7695", "0.8203", "0.7695", "0.8203", "0.78", "0.179", "0.78" ], [ "0.1765", "0.782", "0.822", "0.781", "0.822", "0.7954", "0.1765", "0.797" ], [ "0.1772", "0.7954", "0.822", "0.7954", "0.822", "0.8096", "0.1772", "0.8096" ], [ "0.1772", "0.81", "0.8228", "0.81", "0.8228", "0.8237", "0.1772", "0.8237" ], [ "0.1772", "0.8237", "0.822", "0.8237", "0.822", "0.837", "0.1772", "0.837" ], [ "0.1772", "0.839", "0.5728", "0.839", "0.5728", "0.8496", "0.1772", "0.8496" ], [ "0.1732", "0.869", "0.3586", "0.8706", "0.3586", "0.8853", "0.1732", "0.884" ], [ "0.1732", "0.896", "0.8228", "0.8945", "0.8228", "0.912", "0.1732", "0.9136" ], [ "0.1765", "0.911", "0.8228", "0.911", "0.8228", "0.925", "0.1765", "0.925" ], [ "0.4927", "0.9497", "0.5063", "0.9497", "0.5063", "0.961", "0.4927", "0.961" ] ]

[ "To appear at the ICLR 2024 Workshop on Representational Alignment (Re-Align)", "针对每个动词类别分别计算矩阵，并对该组进行平均。此外，在我们的分析中，我们排除了", "三名未完成评分任务的参与者。这最终产生了24位盲人和20位有远见的参与者的数据", "对于探测任务，我们使用了Binderetal.（2016）收集的语义注释。通过在线调查，他们获", "得了535个单词的人类特征评级。参与者指出了每个单词与65个不同语义属性的链接程", "度，涵盖了14个总体领域的含义。对于某些单词，某些属性不适用。例如，时间和持续时", "间的维度不适用于“鞋”之类的词。。在这些情况下，参与者表示零。原始数据集包含来自", "14个类别的65个维度的535个单词，但是，“使用”一词出现了两次：一次作为动词，", "次作为形容词。在我们的分析中，无论是使用一个还是另一个，结果都没有改变，因此我们", "删除了一个，并分析了剩余的534个单词", "对于单词嵌人，我们使用GloVe（Penningtonetal.，2014）作为其嵌人在预测单词相似性判", "断（e。g。，？）。我们使用了来自Wikipedia2014+Gigaword5语料库的300维GloVe向量", "其中包含大约400,000个无大小写的词汇条目，总共有60亿个令牌。我们提取了Bedny", "etal.（2019）和Binderetal.（2016）中使用的那些词的嵌人", "2.2Supervisedpruning", "2.2.1 Pruning algorithm", "我们使用了有监督的修剪算法，该算法是基于包装器的顺序特征选择过程，可优化人类相似", "性矩阵与从嵌人空间生成的矩阵之间的对齐。它的目标是识别一个减少的特征子集，以便", "当该子集用于生成Word×Word相似性矩阵时，生成的矩阵产生与人类相似性判断的最大", "拟合使用余弦相似度从单词嵌入计算单词之间的成对相似度，而使用Spearman的Rho秩", "相关系数（p）根据Richie&Bhatia（2021）计算两个相似度矩阵之间的拟合。修剪有两种", "方法。在第一种方法中，通过考虑与每个动词类别相关的整套相似性判断来应用修剪。这标", "识了优化该类别中动词的相似性判断预测的特征子集。例如，如果一个类别由15个动词组", "成，则有105个独特的相似性判断来监督修剪过程。第二种方法使用交叉验证来证明修剪", "确实可以识别有意义的维度，如样本外泛化所示。在每个折叠中，一", "一个单词被指定为自标", "并且测试集仅包括那些涉及目标单词的相似性判断。训练集由类别中的所有其他动词对组", "成。例如，在具有15个单词的动词类别中，测试集包括自标单词出现的14个单词对。通", "过使用完整的300GloVe特征集确定匹配的人类数据的预测准确性来建立测试集的基线性", "能。将此基线与仅使用从训练集中保留的特征时的性能进行比较。作为一个额外的控制，在", "每种情况下，我们还构建了一个随机特征集，其大小与保留集的大小相匹配，并将其用于测", "试集。这两种方法都应用于与Bednyetal.（2019）中的七个动词类型中的每一个相关联的", "相似性矩阵，这是由盲人和有视力的人产生的（i。e。，共14个不同的时间）", "2.2.2Overlap between feature sets retained by pruning", "因为修剪识别特征的子集，所以对于修剪算法的任何两个结果，可以比较两组保留特征之", "间的相似性。此计算对通过修剪保留的特征集的索引进行操作。作为关联的指标，我们使用", "了两组U.V的般子系数：般子（U，V）=（2*IUnVD/(IUI+IVD）。鉴于我们比较了大小", "经常变化的集合，在这种情况下，我们通过从较大集合中选择n排名靠前的特征来调整较", "大集合以匹配较小集合的大小，其中n是较小集合的大小。这确保了两组对分母项的贡献", "相等", "重要的是要记住，两个不同的语义域（e。g。，域A和B）表面上具有不同的含义（i。e。，跨", "两个域的单词之间的相似性较低）仍然可以通过修剪收敛在完全相同的特征集上。只要保留", "的特征对于域A和B中的单词始终具有不同的值，就会发生这种情况。换句话说，修剪可", "以识别形成用于表征域结构的基础集的特征，而相似性评级则由跨对象的这些特征的值确", "定。为了说明，考虑两个域中的单词，A和B(e。g。，两种类型的名词)，其中修剪仅保留两", "个域的特征a、b、c、d。对于域A中的单词，（a，b）的值始终高于（c，d)，其余方差说", "明域内的相似性。相反，对于域B中的单词，（a，b）的值始终低于（c，d)。尽管A和B", "中对象的相似性大于跨域的相似性，但两个域都依赖于完全相同的特征。因此，是否需要相", "同的特征来解释两个域的相似性结构与域的相似性无关。", "2.3Supervised probing", "遵循先前的工作（e。g。，，Chersonietal.，2021:Utsumi,2020；？），我们使用偏最小二乘回", "归（PLSR）来预测535个单词的数据集的65个绑定维度，我们还为其提取了GloVe嵌入。", "3" ]

{ "arxiv_id": "2403.06204", "language": "zh", "page_idx": 2 }

[ [ "0.1765", "0.0341", "0.763", "0.0341", "0.763", "0.0486", "0.1765", "0.0486" ], [ "0.4019", "0.1061", "0.4673", "0.1061", "0.4673", "0.1206", "0.4019", "0.1206" ], [ "0.5933", "0.1061", "0.642", "0.1061", "0.642", "0.1206", "0.5933", "0.1206" ], [ "0.6567", "0.1061", "0.71", "0.1061", "0.71", "0.1206", "0.6567", "0.1206" ], [ "0.7305", "0.1061", "0.814", "0.1061", "0.814", "0.1206", "0.7305", "0.1206" ], [ "0.835", "0.1067", "0.8667", "0.1067", "0.8667", "0.1193", "0.835", "0.1193" ], [ "0.4805", "0.125", "0.5664", "0.127", "0.5664", "0.1427", "0.4805", "0.1401" ], [ "0.482", "0.108", "0.5786", "0.108", "0.5786", "0.1193", "0.482", "0.1193" ], [ "0.415", "0.1256", "0.4495", "0.1256", "0.4495", "0.1414", "0.415", "0.1414" ], [ "0.607", "0.1256", "0.6284", "0.1256", "0.6284", "0.1401", "0.607", "0.1401" ], [ "0.665", "0.1256", "0.701", "0.1256", "0.701", "0.1414", "0.665", "0.1414" ], [ "0.7236", "0.127", "0.806", "0.127", "0.806", "0.1414", "0.7236", "0.1414" ], [ "0.841", "0.127", "0.864", "0.127", "0.864", "0.1401", "0.841", "0.1401" ], [ "0.562", "0.1288", "0.5728", "0.1288", "0.5728", "0.1383", "0.562", "0.1383" ], [ "0.1847", "0.1276", "0.2852", "0.1276", "0.2852", "0.1395", "0.1847", "0.1395" ], [ "0.808", "0.13", "0.814", "0.13", "0.814", "0.137", "0.808", "0.137" ], [ "0.183", "0.1395", "0.267", "0.1414", "0.267", "0.1559", "0.183", "0.154" ], [ "0.482", "0.1401", "0.5645", "0.1401", "0.5645", "0.1547", "0.482", "0.1547" ], [ "0.7236", "0.1401", "0.806", "0.1401", "0.806", "0.1547", "0.7236", "0.1547" ], [ "0.6045", "0.1414", "0.6323", "0.1414", "0.6323", "0.1559", "0.6045", "0.1559" ], [ "0.4167", "0.1414", "0.4495", "0.1414", "0.4495", "0.154", "0.4167", "0.154" ], [ "0.6665", "0.1414", "0.699", "0.1414", "0.699", "0.154", "0.6665", "0.154" ], [ "0.808", "0.1459", "0.814", "0.1459", "0.814", "0.1515", "0.808", "0.1515" ], [ "0.8384", "0.1414", "0.8667", "0.1414", "0.8667", "0.1559", "0.8384", "0.1559" ], [ "0.415", "0.1528", "0.452", "0.1559", "0.4495", "0.1692", "0.4143", "0.1674" ], [ "0.607", "0.1547", "0.6284", "0.1547", "0.6284", "0.1692", "0.607", "0.1692" ], [ "0.6665", "0.1547", "0.699", "0.1547", "0.699", "0.1674", "0.6665", "0.1674" ], [ "0.7236", "0.1547", "0.8057", "0.1547", "0.8057", "0.1692", "0.7236", "0.1692" ], [ "0.841", "0.1547", "0.863", "0.1547", "0.863", "0.1686", "0.841", "0.1686" ], [ "0.4836", "0.1559", "0.564", "0.1559", "0.564", "0.1674", "0.4836", "0.1674" ], [ "0.1854", "0.1559", "0.273", "0.1559", "0.273", "0.1674", "0.1854", "0.1674" ], [ "0.808", "0.159", "0.814", "0.159", "0.814", "0.1654", "0.808", "0.1654" ], [ "0.1847", "0.1686", "0.2998", "0.1686", "0.2998", "0.1818", "0.1847", "0.1818" ], [ "0.482", "0.1686", "0.568", "0.1686", "0.568", "0.1831", "0.482", "0.1831" ], [ "0.607", "0.1686", "0.6274", "0.1686", "0.6274", "0.1831", "0.607", "0.1831" ], [ "0.722", "0.1686", "0.773", "0.1686", "0.773", "0.1831", "0.722", "0.1831" ], [ "0.7656", "0.1686", "0.806", "0.1686", "0.806", "0.1831", "0.7656", "0.1831" ], [ "0.841", "0.1686", "0.863", "0.1686", "0.863", "0.1818", "0.841", "0.1818" ], [ "0.6665", "0.1692", "0.699", "0.1692", "0.699", "0.1818", "0.6665", "0.1818" ], [ "0.4167", "0.1692", "0.4495", "0.1692", "0.4495", "0.1818", "0.4167", "0.1818" ], [ "0.564", "0.1736", "0.5728", "0.1736", "0.5728", "0.1787", "0.564", "0.1787" ], [ "0.808", "0.1724", "0.814", "0.1724", "0.814", "0.1787", "0.808", "0.1787" ], [ "0.4805", "0.1799", "0.5645", "0.1818", "0.5645", "0.1976", "0.4805", "0.1957" ], [ "0.7656", "0.1799", "0.806", "0.1818", "0.8057", "0.1976", "0.7646", "0.1957" ], [ "0.415", "0.1812", "0.452", "0.1843", "0.4495", "0.1976", "0.4143", "0.1957" ], [ "0.183", "0.1818", "0.3733", "0.1812", "0.3733", "0.1957", "0.183", "0.1963" ], [ "0.607", "0.1818", "0.6274", "0.1818", "0.6274", "0.1963", "0.607", "0.1963" ], [ "0.665", "0.1818", "0.701", "0.1818", "0.701", "0.1976", "0.665", "0.1976" ], [ "0.722", "0.1818", "0.7715", "0.1818", "0.7715", "0.1963", "0.722", "0.1963" ], [ "0.808", "0.1875", "0.8145", "0.1875", "0.8145", "0.1932", "0.808", "0.1932" ], [ "0.841", "0.1831", "0.863", "0.1831", "0.863", "0.1963", "0.841", "0.1963" ], [ "0.482", "0.1932", "0.564", "0.1957", "0.564", "0.2108", "0.482", "0.209" ], [ "0.665", "0.1945", "0.7017", "0.1976", "0.699", "0.2122", "0.664", "0.2102" ], [ "0.4167", "0.1957", "0.4495", "0.1976", "0.4478", "0.2102", "0.415", "0.2084" ], [ "0.1847", "0.1963", "0.384", "0.1963", "0.384", "0.2102", "0.1847", "0.2102" ], [ "0.609", "0.1963", "0.6274", "0.1963", "0.6274", "0.2102", "0.609", "0.2102" ], [ "0.7236", "0.1963", "0.806", "0.1963", "0.806", "0.2108", "0.7236", "0.2108" ], [ "0.808", "0.2008", "0.814", "0.2008", "0.814", "0.207", "0.808", "0.207" ], [ "0.841", "0.1963", "0.8613", "0.1963", "0.8613", "0.2102", "0.841", "0.2102" ], [ "0.1847", "0.209", "0.2385", "0.2108", "0.2378", "0.226", "0.183", "0.2235" ], [ "0.609", "0.2102", "0.6274", "0.2102", "0.6274", "0.2247", "0.609", "0.2247" ], [ "0.722", "0.2102", "0.818", "0.2102", "0.818", "0.2247", "0.722", "0.2247" ], [ "0.4836", "0.2108", "0.5645", "0.2108", "0.5645", "0.2229", "0.4836", "0.2229" ], [ "0.4167", "0.2108", "0.4495", "0.2108", "0.4495", "0.2235", "0.4167", "0.2235" ], [ "0.6665", "0.2108", "0.699", "0.2108", "0.699", "0.2235", "0.6665", "0.2235" ], [ "0.841", "0.2108", "0.863", "0.2108", "0.863", "0.2235", "0.841", "0.2235" ], [ "0.1748", "0.2405", "0.824", "0.2405", "0.824", "0.254", "0.1748", "0.254" ], [ "0.1748", "0.254", "0.824", "0.254", "0.824", "0.2676", "0.1748", "0.2676" ], [ "0.1748", "0.2676", "0.824", "0.2676", "0.824", "0.281", "0.1748", "0.281" ], [ "0.1732", "0.281", "0.824", "0.2798", "0.824", "0.2942", "0.1732", "0.2954" ], [ "0.1765", "0.2954", "0.8228", "0.2954", "0.8228", "0.3093", "0.1765", "0.3093" ], [ "0.1748", "0.3093", "0.824", "0.3093", "0.824", "0.3225", "0.1748", "0.3225" ], [ "0.1748", "0.3225", "0.824", "0.3225", "0.824", "0.336", "0.1748", "0.336" ], [ "0.1765", "0.3384", "0.2712", "0.3384", "0.2712", "0.3499", "0.1765", "0.3499" ], [ "0.1772", "0.3926", "0.8203", "0.3926", "0.8203", "0.4028", "0.1772", "0.4028" ], [ "0.1748", "0.4048", "0.8228", "0.4048", "0.8228", "0.4185", "0.1748", "0.4185" ], [ "0.1765", "0.4185", "0.824", "0.4185", "0.824", "0.432", "0.1765", "0.432" ], [ "0.1765", "0.433", "0.824", "0.433", "0.824", "0.4463", "0.1765", "0.4463" ], [ "0.1765", "0.4463", "0.6235", "0.4463", "0.6235", "0.4602", "0.1765", "0.4602" ], [ "0.1748", "0.4673", "0.822", "0.4673", "0.822", "0.481", "0.1748", "0.481" ], [ "0.1748", "0.4797", "0.8228", "0.4797", "0.8228", "0.4944", "0.1748", "0.4944" ], [ "0.1765", "0.4956", "0.8228", "0.4956", "0.8228", "0.509", "0.1765", "0.509" ], [ "0.1765", "0.509", "0.824", "0.509", "0.824", "0.523", "0.1765", "0.523" ], [ "0.1748", "0.523", "0.824", "0.523", "0.824", "0.536", "0.1748", "0.536" ], [ "0.1748", "0.536", "0.8228", "0.535", "0.8228", "0.55", "0.1748", "0.5503" ], [ "0.1765", "0.5503", "0.8228", "0.5503", "0.8228", "0.5645", "0.1765", "0.5645" ], [ "0.1748", "0.5645", "0.8228", "0.5645", "0.8228", "0.579", "0.1748", "0.579" ], [ "0.1765", "0.579", "0.4836", "0.579", "0.4836", "0.5923", "0.1765", "0.5923" ], [ "0.1765", "0.5996", "0.8228", "0.5996", "0.8228", "0.613", "0.1765", "0.613" ], [ "0.1748", "0.613", "0.8228", "0.613", "0.8228", "0.6274", "0.1748", "0.6274" ], [ "0.1805", "0.631", "0.1896", "0.631", "0.1896", "0.636", "0.1805", "0.636" ], [ "0.1854", "0.627", "0.824", "0.627", "0.824", "0.64", "0.1854", "0.64" ], [ "0.1748", "0.6416", "0.8228", "0.6416", "0.8228", "0.655", "0.1748", "0.655" ], [ "0.1765", "0.655", "0.8228", "0.655", "0.8228", "0.6685", "0.1765", "0.6685" ], [ "0.1765", "0.6685", "0.8228", "0.6685", "0.8228", "0.6816", "0.1765", "0.6816" ], [ "0.1748", "0.681", "0.822", "0.681", "0.822", "0.696", "0.1748", "0.696" ], [ "0.1748", "0.6963", "0.8228", "0.6963", "0.8228", "0.7104", "0.1748", "0.7104" ], [ "0.1765", "0.7104", "0.2427", "0.7104", "0.2427", "0.7246", "0.1765", "0.7246" ], [ "0.1748", "0.7603", "0.1896", "0.7603", "0.1896", "0.771", "0.1748", "0.771" ], [ "0.1854", "0.7603", "0.273", "0.7603", "0.273", "0.7715", "0.1854", "0.7715" ], [ "0.1732", "0.797", "0.7344", "0.797", "0.7344", "0.811", "0.1732", "0.811" ], [ "0.1765", "0.8276", "0.824", "0.8276", "0.824", "0.8413", "0.1765", "0.8413" ], [ "0.1748", "0.8403", "0.8228", "0.8413", "0.8228", "0.856", "0.1748", "0.855" ], [ "0.1765", "0.856", "0.824", "0.856", "0.824", "0.869", "0.1765", "0.869" ], [ "0.1765", "0.869", "0.822", "0.869", "0.822", "0.8833", "0.1765", "0.8833" ], [ "0.1765", "0.884", "0.8228", "0.884", "0.8228", "0.898", "0.1765", "0.898" ], [ "0.1765", "0.898", "0.8228", "0.898", "0.8228", "0.911", "0.1765", "0.911" ], [ "0.1772", "0.9136", "0.5156", "0.9136", "0.5156", "0.924", "0.1772", "0.924" ], [ "0.492", "0.9487", "0.509", "0.9487", "0.509", "0.962", "0.492", "0.962" ] ]

[ "To appear at the ICLR 2024 Workshop on Representational Alignment (Re-Align)", "CB Base", "CB F.", "S Base", "S Retained", "S F.", "0.39 ( 0.32", "CB Retained", "0.41", "74", "0.47", "0.46 { 0.37", "84", "喜", "Perc.Amodal", "", "Perc. Sight", "0.63 { 0.48", "0.68 0.50", "118", "0.56", "0.64", "", "113", "0.16", "76", "0.14", "0.38 { 0.08", "58", "0.400.19", "Perc. Touch", "", "Emission.Light", "0.62 {0.54", "99", "0.54 {", "0.41", "80", "0.42", "0.56", "", "", "0.26 (0.27", "0.28", "0.25", "Emission.Animate Sound", "67", "0.33", "0.37", "", "59", "0.09 { 0.08", "0.13", "0.10", "Emission. Inanimate Sound", "46", "0.06 { 0.07", "", "57", "Motion", "75", "0.40 (0.18 ]", "0.29/0.22", "0.35", "0.33", "60", "Table l: Out of sample generalization of supervised pruning. Base (Baseline) indicates", "prediction when using all 300 GloVe features. Retained indicates prediction using features", "learned from training set. CB = Congenitally Blind ; S = Sighted. F. (Features) indicates", "average number of features retained from training set.. Numbers in curly brackets in the", "“Retained' columns define chance and indicate prediction performance when using randomly", "selected feature sets matched in size to those retained from the training set. Values in bold", "indicate better prediction than baseline, or similar prediction to baseline that surpasses the", "chance value.", "我们遵循完全相同的过程，将其分别应用于通过监督修剪发现的14个保留的特征集（七个", "动词类型为盲人/视力）中的每一个。我们强调，对于每个动词类别，我们仅使用在修剪步", "骤中为该类别识别的GloVe特征(i。e。，我们不使用所有300个功能)。此预测的结果允", "许确定不同域保留的集合是否编码不同的语义。用于PLSR的保留特征是在考虑到整套人", "类相似性判断的情况下产生的(i。e。，在交叉验证上下文之外)", "我们将PLSR应用于保留集，采用留一法交叉验证程序（跟随？）。Binder的数据集包括", "535个单词，在删除“使用”一词的双重使用后，剩下534个单词。。对于每个折叠，我们", "将533个单词指定为训练集，将单个遗漏的单词指定为测试集。在训练集上训练PLSR，然", "后将该模型应用于测试词，预测其65个特征值。请注意，当应用于所有534个单词时，这", "将生成一个534×65堆叠矩阵，包括所有单词，我们将其称为留一交叉验证（LOOCV）堆", "叠矩阵65；LSM65。堆叠矩阵是由534个不同的回归产生的。我们进一步注意到，在某些", "分析中，我们通过将65个维度合并到14个更广泛的语义区域中，将534×65矩阵压缩为", "534×14，从而生成LSMi14。将65个特征合并到14个域是根据Binderetal.（2016）中", "指定的分配实现的（请参见它们的表1、2）", "预测矩阵中的数据可作为评估针对盲人和视力以及不同动词类型修剪的嵌入的语义内容差", "异的基础。我们执行三个分析：首先，按列（特征）考虑LSM65，我们将每列（n=534", "", "的预测值与Binder的地面实况数据相关联。这是Chersonietal.（2021）使用的预测每个", "特征的准确性的指标。这些数据也被平均到14个较大的语义域中。其次，按行（字）考虑", "LSM65，我们将每行的预测值与地面实况数据相关联。第三，为了确定盲人和有远见的修", "剪嵌人的预测精度是否存在差异，我们还将精度定义为LSM65中每个单元格的预测值与地", "面真值之间的绝对偏差。以单词（N=534）为分析单位，这些按单元格的预测误差可以比", "较14个领域中每个领域的盲人和有远见者的预测准确性。我们使用配对T检验，以单词为", "分析单位", "3", "Results", "3.1 Preliminary evaluation: out of sample generalization of pruning solution", "表1显示了使用所有GloVe特征（基线）或通过修剪训练集（“保留”）学习的特征时测试", "集数据的预测。修剪通常会成功推广：通过修剪学习的功能子集超过了基准性能，或者仅", "使用总共300个功能中的20-30个%来保持基准性能。在三种情况下，我们没有找到概括：", "对于发射-无生命的声音（视力和盲人）和发射-动画声音（盲人）。在后一种情况下，全集", "保留集和随机集的性能都非常相似（～。27)。请注意，表1中报告的所有数据均基于针对", "由14个数据点组成的测试折叠分区计算的预测性能，这可能会限制灵敏度。在下面的分析", "中，我们在同的完整数据集上应用和评估修剪", "4." ]

{ "arxiv_id": "2403.06204", "language": "zh", "page_idx": 3 }

[ [ "0.1765", "0.0341", "0.763", "0.0341", "0.763", "0.04736", "0.1765", "0.04736" ], [ "0.4004", "0.1048", "0.4697", "0.1067", "0.4697", "0.1218", "0.4004", "0.1193" ], [ "0.7573", "0.1061", "0.806", "0.1061", "0.806", "0.1206", "0.7573", "0.1206" ], [ "0.8203", "0.1067", "0.8545", "0.1067", "0.8545", "0.1193", "0.8203", "0.1193" ], [ "0.5933", "0.108", "0.6465", "0.108", "0.6465", "0.1193", "0.5933", "0.1193" ], [ "0.482", "0.108", "0.5786", "0.108", "0.5786", "0.1187", "0.482", "0.1187" ], [ "0.661", "0.108", "0.7417", "0.108", "0.7417", "0.1193", "0.661", "0.1193" ], [ "0.1814", "0.125", "0.286", "0.1256", "0.286", "0.1401", "0.1814", "0.1395" ], [ "0.4656", "0.11237", "0.478", "0.11237", "0.478", "0.11554", "0.4656", "0.11554" ], [ "0.415", "0.1256", "0.451", "0.1256", "0.451", "0.1414", "0.415", "0.1414" ], [ "0.513", "0.1256", "0.5474", "0.1256", "0.5474", "0.1414", "0.513", "0.1414" ], [ "0.601", "0.1256", "0.636", "0.1256", "0.636", "0.1414", "0.601", "0.1414" ], [ "0.683", "0.1256", "0.7183", "0.1256", "0.7183", "0.1414", "0.683", "0.1414" ], [ "0.7715", "0.127", "0.7935", "0.127", "0.7935", "0.1401", "0.7715", "0.1401" ], [ "0.826", "0.127", "0.8545", "0.127", "0.8545", "0.1395", "0.826", "0.1395" ], [ "0.183", "0.1395", "0.2656", "0.1414", "0.2656", "0.1559", "0.183", "0.154" ], [ "0.6855", "0.1395", "0.7173", "0.1414", "0.716", "0.1547", "0.6846", "0.1528" ], [ "0.7695", "0.1401", "0.796", "0.1401", "0.796", "0.154", "0.7695", "0.154" ], [ "0.5146", "0.1414", "0.544", "0.1414", "0.544", "0.154", "0.5146", "0.154" ], [ "0.4167", "0.1414", "0.4495", "0.1414", "0.4495", "0.154", "0.4167", "0.154" ], [ "0.6045", "0.1414", "0.6357", "0.1414", "0.6357", "0.154", "0.6045", "0.154" ], [ "0.824", "0.1414", "0.853", "0.1414", "0.853", "0.154", "0.824", "0.154" ], [ "0.4167", "0.1547", "0.4495", "0.1547", "0.4495", "0.1674", "0.4167", "0.1674" ], [ "0.5146", "0.1547", "0.546", "0.1547", "0.546", "0.1674", "0.5146", "0.1674" ], [ "0.6045", "0.1547", "0.6357", "0.1547", "0.6357", "0.1674", "0.6045", "0.1674" ], [ "0.6855", "0.1547", "0.7173", "0.1547", "0.7173", "0.1674", "0.6855", "0.1674" ], [ "0.7715", "0.1547", "0.7935", "0.1547", "0.7935", "0.1686", "0.7715", "0.1686" ], [ "0.83", "0.1515", "0.8506", "0.1559", "0.846", "0.1692", "0.826", "0.1654" ], [ "0.4167", "0.1654", "0.451", "0.1686", "0.4495", "0.1831", "0.4143", "0.1812" ], [ "0.1854", "0.1559", "0.2712", "0.1559", "0.2712", "0.1674", "0.1854", "0.1674" ], [ "0.5146", "0.1666", "0.546", "0.1686", "0.544", "0.1818", "0.513", "0.1799" ], [ "0.6855", "0.1666", "0.7173", "0.1686", "0.716", "0.1818", "0.6846", "0.1799" ], [ "0.603", "0.1686", "0.6357", "0.1686", "0.6357", "0.1812", "0.603", "0.1812" ], [ "0.1854", "0.1686", "0.2998", "0.1686", "0.2998", "0.1818", "0.1854", "0.1818" ], [ "0.773", "0.1686", "0.7935", "0.1686", "0.7935", "0.1818", "0.773", "0.1818" ], [ "0.8286", "0.1686", "0.8486", "0.1686", "0.8486", "0.1818", "0.8286", "0.1818" ], [ "0.183", "0.1818", "0.3733", "0.1812", "0.3733", "0.1957", "0.183", "0.1963" ], [ "0.4167", "0.1831", "0.4495", "0.1831", "0.4495", "0.1957", "0.4167", "0.1957" ], [ "0.5146", "0.1831", "0.546", "0.1831", "0.546", "0.1957", "0.5146", "0.1957" ], [ "0.6045", "0.1831", "0.6357", "0.1831", "0.6357", "0.1957", "0.6045", "0.1957" ], [ "0.6855", "0.1831", "0.7173", "0.1831", "0.7173", "0.1957", "0.6855", "0.1957" ], [ "0.773", "0.1818", "0.7935", "0.1818", "0.7935", "0.1957", "0.773", "0.1957" ], [ "0.83", "0.1831", "0.8486", "0.1831", "0.8486", "0.1963", "0.83", "0.1963" ], [ "0.1847", "0.1963", "0.384", "0.1963", "0.384", "0.2102", "0.1847", "0.2102" ], [ "0.4167", "0.1963", "0.4495", "0.1963", "0.4495", "0.209", "0.4167", "0.209" ], [ "0.5146", "0.1963", "0.546", "0.1963", "0.546", "0.209", "0.5146", "0.209" ], [ "0.6045", "0.1963", "0.6357", "0.1963", "0.6357", "0.209", "0.6045", "0.209" ], [ "0.6855", "0.1963", "0.7173", "0.1963", "0.7173", "0.209", "0.6855", "0.209" ], [ "0.773", "0.1963", "0.7935", "0.1963", "0.7935", "0.2102", "0.773", "0.2102" ], [ "0.83", "0.1963", "0.8486", "0.1963", "0.8486", "0.209", "0.83", "0.209" ], [ "0.513", "0.209", "0.5474", "0.209", "0.5474", "0.2247", "0.513", "0.2247" ], [ "0.601", "0.209", "0.636", "0.209", "0.636", "0.2247", "0.601", "0.2247" ], [ "0.6846", "0.209", "0.7197", "0.209", "0.7197", "0.2247", "0.6846", "0.2247" ], [ "0.773", "0.2102", "0.7935", "0.2102", "0.7935", "0.2247", "0.773", "0.2247" ], [ "0.83", "0.2102", "0.8506", "0.2102", "0.8506", "0.2247", "0.83", "0.2247" ], [ "0.4167", "0.2108", "0.4495", "0.2108", "0.4495", "0.2235", "0.4167", "0.2235" ], [ "0.1847", "0.2102", "0.2385", "0.2102", "0.2385", "0.2247", "0.1847", "0.2247" ], [ "0.1748", "0.238", "0.8228", "0.2393", "0.8228", "0.254", "0.1748", "0.2524" ], [ "0.1765", "0.254", "0.8013", "0.254", "0.8013", "0.2676", "0.1765", "0.2676" ], [ "0.1732", "0.3372", "0.6284", "0.3384", "0.6284", "0.3528", "0.1732", "0.3516" ], [ "0.1772", "0.3838", "0.8203", "0.3838", "0.8203", "0.3945", "0.1772", "0.3945" ], [ "0.1765", "0.3965", "0.8228", "0.3965", "0.8228", "0.4104", "0.1765", "0.4104" ], [ "0.1765", "0.4104", "0.822", "0.4104", "0.822", "0.4236", "0.1765", "0.4236" ], [ "0.1765", "0.4248", "0.8228", "0.4248", "0.8228", "0.4382", "0.1765", "0.4382" ], [ "0.1765", "0.44", "0.2418", "0.44", "0.2418", "0.4507", "0.1765", "0.4507" ], [ "0.1765", "0.459", "0.824", "0.459", "0.824", "0.473", "0.1765", "0.473" ], [ "0.1765", "0.473", "0.824", "0.473", "0.824", "0.486", "0.1765", "0.486" ], [ "0.1748", "0.486", "0.824", "0.4849", "0.824", "0.5", "0.1748", "0.5005" ], [ "0.1765", "0.5005", "0.8228", "0.5005", "0.8228", "0.5146", "0.1765", "0.5146" ], [ "0.1765", "0.5146", "0.822", "0.5146", "0.822", "0.528", "0.1765", "0.528" ], [ "0.1765", "0.529", "0.8228", "0.529", "0.8228", "0.5425", "0.1765", "0.5425" ], [ "0.1765", "0.5425", "0.5522", "0.5425", "0.5522", "0.556", "0.1765", "0.556" ], [ "0.1765", "0.563", "0.824", "0.563", "0.824", "0.577", "0.1765", "0.577" ], [ "0.1748", "0.5757", "0.8228", "0.577", "0.8228", "0.5913", "0.1748", "0.5903" ], [ "0.1765", "0.5913", "0.824", "0.5913", "0.824", "0.605", "0.1765", "0.605" ], [ "0.1765", "0.606", "0.8228", "0.606", "0.8228", "0.619", "0.1765", "0.619" ], [ "0.1765", "0.6187", "0.8228", "0.6187", "0.8228", "0.632", "0.1765", "0.632" ], [ "0.1748", "0.632", "0.8096", "0.632", "0.8096", "0.646", "0.1748", "0.646" ], [ "0.1748", "0.653", "0.819", "0.653", "0.819", "0.6665", "0.1748", "0.6665" ], [ "0.1772", "0.6675", "0.8228", "0.6675", "0.8228", "0.681", "0.1772", "0.681" ], [ "0.1765", "0.6816", "0.822", "0.6816", "0.822", "0.696", "0.1765", "0.696" ], [ "0.1772", "0.6978", "0.8203", "0.6978", "0.8203", "0.7085", "0.1772", "0.7085" ], [ "0.1772", "0.711", "0.822", "0.711", "0.822", "0.7217", "0.1772", "0.7217" ], [ "0.1765", "0.7236", "0.8228", "0.7236", "0.8228", "0.7373", "0.1765", "0.7373" ], [ "0.1765", "0.7363", "0.822", "0.7363", "0.822", "0.75", "0.1765", "0.75" ], [ "0.1772", "0.7524", "0.8228", "0.7524", "0.8228", "0.763", "0.1772", "0.763" ], [ "0.1765", "0.765", "0.5024", "0.765", "0.5024", "0.779", "0.1765", "0.779" ], [ "0.1765", "0.786", "0.8228", "0.786", "0.8228", "0.8", "0.1765", "0.8" ], [ "0.1772", "0.8", "0.8228", "0.8", "0.8228", "0.813", "0.1772", "0.813" ], [ "0.1772", "0.813", "0.8228", "0.813", "0.8228", "0.827", "0.1772", "0.827" ], [ "0.1765", "0.8276", "0.8228", "0.8276", "0.8228", "0.8413", "0.1765", "0.8413" ], [ "0.1748", "0.8403", "0.822", "0.8413", "0.822", "0.855", "0.1748", "0.854" ], [ "0.1748", "0.856", "0.8228", "0.856", "0.8228", "0.869", "0.1748", "0.869" ], [ "0.1765", "0.8687", "0.822", "0.869", "0.822", "0.884", "0.1765", "0.8833" ], [ "0.1748", "0.882", "0.822", "0.8833", "0.822", "0.898", "0.1748", "0.8965" ], [ "0.1748", "0.896", "0.824", "0.8965", "0.824", "0.911", "0.1748", "0.91" ], [ "0.1765", "0.911", "0.4697", "0.911", "0.4697", "0.925", "0.1765", "0.925" ], [ "0.492", "0.9478", "0.5073", "0.9478", "0.5073", "0.962", "0.492", "0.962" ] ]

[ "To appear at the ICLR 2024 Workshop on Representational Alignment (Re-Align)", "CB Base", "CB F.", "S F.", "S Base", "CB Retained", "SRetained", "Perc. Amodal", "", "0.43", "0.74", "0.48", "0.77", "25", "126", "Perc. Sight", "0.85", "113", "0.81", "0.56", "0.69", "149", "0.13", "0.64", "0.14", "0.69", "60", "59", "0.64", "Perc. Touch", "0.88", "0.85", "0.49", "Emission.Light", "55", "94", "Emission.Animate Sound", "0.25", "0.69", "0.32", "0.76", "88", "62", "Emission. Inanimate Sound", "0.10", "0.60", "0.16", "0.64", "32", "32", "0.68", "0.35", "0.73", "57", "32", "0.36", "Motion", "Table 2: Comparison of prediction accuracy (Spearman's rho) when using full feature set", "(baseline) or retained features sets (retained). In this application, there was no test set.", "3.2Correspondence between feature-sets retained by pruning", "表2显示了在交叉验证框架的上下文之外进行的应用于单个动词数据集的修剪过程的结果", "虽然从机器学习的角度来看，这可能被视为过度拟合，但我们的目标是获得一组特定人类", "相似性判断的最有效预测因子。此修剪建立了所有后续分析中使用的保留特征集。从表中", "可以明显看出，在这种情况下，修剪的影响是巨大的，在某些情况下，将预测精度提高了四", "倍或六倍", "数据还表明，从修剪的GloVe嵌人预测人类相似性判断可以接近噪声上限。Bedny等人。", "（2019）提供了上限估计：他们从单独的（第三）视力人群中获得了相似性等级，并评估了该", "组的等级与实验视力人群和盲人人群的等级之间的相关性。在某些情况下，例如对于发光", "动词，观察到高相关性（0。91,0。视力和盲人为93），表明不同群体的相似性评级具有很高", "的可靠性。在这种情况下，监督修剪实现了0的相关性。85和0。88分别。相反，对于运", "动动词，发现较低的行为相关性（0。75,0。72)，有监督修剪导致0。73和0。68。这表明", "对于特定的语义类别，来自修剪特征的预测接近上限", "除此之外，我们注意到以下内容。首先，保留更多特征并不一定会产生更好的预测准确性：", "对于Perception-Amodal动词和Emission-Light动词，盲人和有远见者之间保留的特征数", "量差异很大，但预测准确性（基线和保留）相似。其次，保留更多功能的修剪解决方案并不", "稀疏。我们压缩了每个保留的嵌人矩阵，并将其与原始大小（Ziv&Lempel，1977）进行了比", "较。我们发现，在所有解决方案中，压缩比几乎恰好为0。4（范围：0.39－0.42）。这表明", "不同的保留特征集包含每个维度的相似信息。因此，具有大量特征的保留集包含更多信息", "Dice系数分析（图2a）表明，在对角线中发现了最高的系数，该系数索引了相同动词类型", "近视和失明的情况。最高值(O。89）被发现用于发光动词(S_e_1，CB_e_1)。然而，值得", "注意的是，对于感知-非模态、运动和发射-无生命声音，重叠较弱。有一种情况是动词类型", "之间有很强的重叠，在Perception-Amodal和Perception-Sight动词的高Dice值中可以看", "到，但仅适用于视力（Dice=0.42）。重要的是，对于盲人来说，情况并非如此，这些动词", "类型实际上没有重叠（Dice=0.04）。这表明，对于有视力的人，感知-非模态动词与更强", "的视觉成分相关联。。许多殷子系数都很低，在0到0的范围内。05，其中一个达到零。运", "动和发射-无生命声音显示特别低的值。总而言之，分析通过修剪保留的特征的重叠强烈表", "明，不同的动词类别映射到不同的嵌人维度上", "图2b显示没有稳定的特征核心倾向于在修剪解决方案中保留。对于300个GloVe特征中", "的每一个，我们计算了它在14个不同的保留集合中出现的频率。如图所示，14个解决方案", "中的9个以上没有出现任何特征，而300个解决方案中只有16个出现在7-9个解决方案", "中。因此，几乎不支持与所有动词类别相关的一组核心功能。有趣的是，GloVe的300个功", "能中有19个没有保留在14个修剪应用程序中。为了理解它们包含的含义，我们从Binder", "etal.（2016）的数据集中确定了前20个单词，这些特征产生了最高的激活总和。获得的单", "词是：选举，部长，银行家，选民，被捕，政治人物，尖叫声，政党，机场，枪击，外交官", "讲话，集会，商人，玻璃，大教堂，驾驶员，扩音器，指挥官，陪审团（前五个得分远高于", "其他得分）。可以理解的是，其中大多数是与社会，法律和秩序，群体和社会状况有关的术", "语，它们确实与我们使用的动词类别无关", "5" ]

{ "arxiv_id": "2403.06204", "language": "zh", "page_idx": 4 }

[ [ "0.0825", "0.09344", "0.468", "0.09344", "0.468", "0.1067", "0.0825", "0.1067" ], [ "0.505", "0.09344", "0.8906", "0.09344", "0.8906", "0.1067", "0.505", "0.1067" ], [ "0.0825", "0.108", "0.468", "0.108", "0.468", "0.1225", "0.0825", "0.1225" ], [ "0.5034", "0.1067", "0.8906", "0.1067", "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", "0.1237", "0.8896", "0.137", "0.5034", "0.1364" ], [ "0.0825", "0.1383", "0.4673", "0.1383", "0.4673", "0.1515", "0.0825", "0.1515" ], [ "0.505", "0.1383", "0.8906", "0.1383", "0.8906", "0.1515", "0.505", "0.1515" ], [ "0.0825", "0.154", "0.4673", "0.154", "0.4673", "0.1674", "0.0825", "0.1674" ], [ "0.5034", "0.154", "0.8906", "0.154", "0.8906", "0.1674", "0.5034", "0.1674" ], [ "0.0825", "0.1686", "0.468", "0.1686", "0.468", "0.1818", "0.0825", "0.1818" ], [ "0.5034", "0.1686", "0.6504", "0.1686", "0.6504", "0.1818", "0.5034", "0.1818" ], [ "0.0825", "0.1849", "0.455", "0.1849", "0.455", "0.1957", "0.0825", "0.1957" ], [ "0.5254", "0.1849", "0.8906", "0.1849", "0.8906", "0.1957", "0.5254", "0.1957" ], [ "0.1013", "0.1989", "0.468", "0.1989", "0.468", "0.2122", "0.1013", "0.2122" ], [ "0.5034", "0.1989", "0.892", "0.1989", "0.892", "0.2122", "0.5034", "0.2122" ], [ "0.08417", "0.2134", "0.468", "0.2134", "0.468", "0.2267", "0.08417", "0.2267" ], [ "0.505", "0.2134", "0.8906", "0.2134", "0.8906", "0.2267", "0.505", "0.2267" ], [ "0.0825", "0.2291", "0.468", "0.2291", "0.468", "0.2424", "0.0825", "0.2424" ], [ "0.5034", "0.2291", "0.8906", "0.2291", "0.8906", "0.2424", "0.5034", "0.2424" ], [ "0.0825", "0.2437", "0.468", "0.2437", "0.468", "0.2568", "0.0825", "0.2568" ], [ "0.5034", "0.2437", "0.8906", "0.2437", "0.8906", "0.2568", "0.5034", "0.2568" ], [ "0.0825", "0.2588", "0.468", "0.2588", "0.468", "0.2727", "0.0825", "0.2727" ], [ "0.5024", "0.2583", "0.8906", "0.2588", "0.8906", "0.2734", "0.5024", "0.2727" ], [ "0.0809", "0.2727", "0.468", "0.2734", "0.468", "0.2886", "0.0809", "0.2874" ], [ "0.505", "0.276", "0.8896", "0.276", "0.8896", "0.286", "0.505", "0.286" ], [ "0.08417", "0.289", "0.468", "0.289", "0.468", "0.303", "0.08417", "0.303" ], [ "0.5034", "0.289", "0.8906", "0.289", "0.8906", "0.303", "0.5034", "0.303" ], [ "0.0801", "0.3037", "0.468", "0.303", "0.468", "0.3164", "0.0801", "0.3176" ], [ "0.505", "0.3062", "0.6665", "0.3062", "0.6665", "0.3164", "0.505", "0.3164" ], [ "0.0825", "0.3193", "0.4673", "0.3193", "0.4673", "0.3333", "0.0825", "0.3333" ], [ "0.524", "0.3193", "0.8906", "0.3193", "0.8906", "0.3333", "0.524", "0.3333" ], [ "0.08417", "0.3352", "0.4673", "0.3352", "0.4673", "0.3484", "0.08417", "0.3484" ], [ "0.505", "0.334", "0.8896", "0.334", "0.8896", "0.348", "0.505", "0.348" ], [ "0.08417", "0.3516", "0.4673", "0.3516", "0.4673", "0.3623", "0.08417", "0.3623" ], [ "0.505", "0.3499", "0.8896", "0.3499", "0.8896", "0.363", "0.505", "0.363" ], [ "0.0809", "0.3643", "0.468", "0.3643", "0.468", "0.3774", "0.0809", "0.3774" ], [ "0.505", "0.3662", "0.8784", "0.3662", "0.8784", "0.377", "0.505", "0.377" ], [ "0.0825", "0.3801", "0.468", "0.3801", "0.468", "0.3933", "0.0825", "0.3933" ], [ "0.5034", "0.3801", "0.8896", "0.3801", "0.8896", "0.3933", "0.5034", "0.3933" ], [ "0.0801", "0.3933", "0.468", "0.3945", "0.468", "0.4092", "0.0801", "0.4077" ], [ "0.5034", "0.3945", "0.8906", "0.3945", "0.8906", "0.4077", "0.5034", "0.4077" ], [ "0.08417", "0.411", "0.1299", "0.411", "0.1299", "0.4224", "0.08417", "0.4224" ], [ "0.505", "0.4104", "0.7817", "0.4104", "0.7817", "0.4236", "0.505", "0.4236" ], [ "0.1013", "0.4268", "0.4673", "0.4268", "0.4673", "0.4375", "0.1013", "0.4375" ], [ "0.0801", "0.44", "0.468", "0.4395", "0.468", "0.4526", "0.0801", "0.4539" ], [ "0.501", "0.4382", "0.6006", "0.4363", "0.6006", "0.4507", "0.501", "0.4526" ], [ "0.0801", "0.4539", "0.468", "0.4558", "0.468", "0.4697", "0.0801", "0.4673" ], [ "0.524", "0.464", "0.8906", "0.464", "0.8906", "0.4778", "0.524", "0.4778" ], [ "0.0809", "0.4702", "0.468", "0.4702", "0.468", "0.4841", "0.0809", "0.4841" ], [ "0.5034", "0.4792", "0.8896", "0.4792", "0.8896", "0.4924", "0.5034", "0.4924" ], [ "0.0825", "0.4873", "0.4673", "0.4873", "0.4673", "0.4976", "0.0825", "0.4976" ], [ "0.5034", "0.4937", "0.8906", "0.4944", "0.8906", "0.5083", "0.5034", "0.507" ], [ "0.0825", "0.5005", "0.468", "0.5005", "0.468", "0.5146", "0.0825", "0.5146" ], [ "0.505", "0.509", "0.8906", "0.509", "0.8906", "0.523", "0.505", "0.523" ], [ "0.0825", "0.5166", "0.4673", "0.5166", "0.4673", "0.53", "0.0825", "0.53" ], [ "0.0809", "0.531", "0.4673", "0.531", "0.4673", "0.5444", "0.0809", "0.5444" ], [ "0.505", "0.526", "0.888", "0.526", "0.888", "0.536", "0.505", "0.536" ], [ "0.5034", "0.539", "0.8906", "0.539", "0.8906", "0.553", "0.5034", "0.553" ], [ "0.0809", "0.5454", "0.468", "0.5454", "0.468", "0.5596", "0.0809", "0.5596" ], [ "0.5034", "0.5547", "0.8896", "0.5547", "0.8896", "0.5684", "0.5034", "0.5684" ], [ "0.0809", "0.561", "0.2942", "0.561", "0.2942", "0.5747", "0.0809", "0.5747" ], [ "0.505", "0.5684", "0.6846", "0.5684", "0.6846", "0.582", "0.505", "0.582" ], [ "0.0801", "0.5903", "0.2345", "0.5903", "0.2345", "0.604", "0.0801", "0.604" ], [ "0.5034", "0.5933", "0.6587", "0.5933", "0.6587", "0.607", "0.5034", "0.607" ], [ "0.10046", "0.6157", "0.468", "0.6157", "0.468", "0.629", "0.10046", "0.629" ], [ "0.523", "0.617", "0.8906", "0.617", "0.8906", "0.631", "0.523", "0.631" ], [ "0.0809", "0.63", "0.468", "0.63", "0.468", "0.643", "0.0809", "0.643" ], [ "0.5034", "0.632", "0.8936", "0.632", "0.8936", "0.646", "0.5034", "0.646" ], [ "0.0825", "0.646", "0.468", "0.646", "0.468", "0.659", "0.0825", "0.659" ], [ "0.5034", "0.6465", "0.8945", "0.6465", "0.8945", "0.66", "0.5034", "0.66" ], [ "0.0825", "0.662", "0.4656", "0.662", "0.4656", "0.6724", "0.0825", "0.6724" ], [ "0.5024", "0.661", "0.7876", "0.661", "0.7876", "0.675", "0.5024", "0.675" ], [ "0.0825", "0.675", "0.4673", "0.675", "0.4673", "0.688", "0.0825", "0.688" ], [ "0.0825", "0.6914", "0.4656", "0.6914", "0.4656", "0.705", "0.0825", "0.705" ], [ "0.5884", "0.6924", "0.8105", "0.6924", "0.8105", "0.7026", "0.5884", "0.7026" ], [ "0.812", "0.6943", "0.8594", "0.6943", "0.8594", "0.702", "0.812", "0.702" ], [ "0.8677", "0.6914", "0.8906", "0.6914", "0.8906", "0.704", "0.8677", "0.704" ], [ "0.0825", "0.7056", "0.468", "0.7056", "0.468", "0.7197", "0.0825", "0.7197" ], [ "0.808", "0.696", "0.814", "0.696", "0.814", "0.7007", "0.808", "0.7007" ], [ "0.0825", "0.7227", "0.3774", "0.7227", "0.3774", "0.733", "0.0825", "0.733" ], [ "0.505", "0.7197", "0.8906", "0.7197", "0.8906", "0.733", "0.505", "0.733" ], [ "0.1013", "0.7363", "0.468", "0.7363", "0.468", "0.75", "0.1013", "0.75" ], [ "0.5034", "0.735", "0.523", "0.735", "0.523", "0.7476", "0.5034", "0.7476" ], [ "0.0809", "0.752", "0.4673", "0.752", "0.4673", "0.765", "0.0809", "0.765" ], [ "0.501", "0.757", "0.777", "0.759", "0.777", "0.7734", "0.501", "0.7715" ], [ "0.08417", "0.768", "0.4673", "0.768", "0.4673", "0.779", "0.08417", "0.779" ], [ "0.0809", "0.782", "0.468", "0.782", "0.468", "0.7954", "0.0809", "0.7954" ], [ "0.524", "0.782", "0.8906", "0.782", "0.8906", "0.7954", "0.524", "0.7954" ], [ "0.0825", "0.797", "0.468", "0.797", "0.468", "0.81", "0.0825", "0.81" ], [ "0.505", "0.799", "0.8906", "0.799", "0.8906", "0.8096", "0.505", "0.8096" ], [ "0.0809", "0.811", "0.259", "0.811", "0.259", "0.8247", "0.0809", "0.8247" ], [ "0.5034", "0.811", "0.8906", "0.811", "0.8906", "0.8247", "0.5034", "0.8247" ], [ "0.1013", "0.827", "0.468", "0.827", "0.468", "0.8403", "0.1013", "0.8403" ], [ "0.505", "0.827", "0.8906", "0.827", "0.8906", "0.8403", "0.505", "0.8403" ], [ "0.08417", "0.8423", "0.468", "0.8423", "0.468", "0.856", "0.08417", "0.856" ], [ "0.505", "0.8433", "0.8896", "0.8433", "0.8896", "0.854", "0.505", "0.854" ], [ "0.08496", "0.8594", "0.4673", "0.8594", "0.4673", "0.869", "0.08496", "0.869" ], [ "0.505", "0.858", "0.536", "0.858", "0.536", "0.869", "0.505", "0.869" ], [ "0.0809", "0.8726", "0.468", "0.8726", "0.468", "0.886", "0.0809", "0.886" ], [ "0.523", "0.8706", "0.8896", "0.872", "0.8896", "0.886", "0.523", "0.8853" ], [ "0.0809", "0.887", "0.468", "0.887", "0.468", "0.901", "0.0809", "0.901" ], [ "0.5034", "0.887", "0.8906", "0.887", "0.8906", "0.901", "0.5034", "0.901" ] ]

[ "差异是否正在增加E[13]，如图。1。已经研究了各种", "[93]，标签平滑[72]，批量归一化E[30]，虚拟对抗", "策略，以通过权重正则化E[19]，退出E[7]，混合E", "训练E[51]，洪泛EXMATHX。然而，这些策略主要", "93，标签平滑72批量归一化E30等来减轻过", "是为监督或半监督学习而设计的，这些策略不太适合", "度拟合。然而，所有这些策略都是为监督和半监督学习", "通常涉及域间隙和无监督损失的域自适应学习。我们", "而设计的，其中训练数据和测试数据通常具有非常相", "设计了一种对抗攻击技术，可以有效地减轻领域自适", "似的分布。对于域自适应学习，由于忽略了不同域的数", "应学习中的过度拟合", "据之间广泛存在的域间隙，因此它们不能很好地适应", "对抗性攻击：已在各种安全问题中研究了对抗性攻", "我们设计了一种鲁棒的域自适应技术，该技术引入", "击。例如，【73]显示对抗性样本很容易混淆CNN模型。", "了一种新颖的傅立叶对抗攻击（FAA）技术，以减轻无", "以下工作通过快速梯度符号14]，最小对抗扰动[53】，", "监督域自适应中的过拟合。FAA通过生成对抗性样本", "通用对抗扰动[52]，无梯度攻击【1]，可转移对抗样本", "来缓解过度拟合，以防止监督和无监督UDA损失的过", "生成【43】等从不同方面改进了对抗攻击。对抗性攻击", "度最小化，如图所示。1。具体来说，FAA将训练图", "也已应用于其他任务，例如。g。，[96，51]采用对抗性", "像分解为多个频率分量（FC），并且仅干扰捕获很少语", "样本来减轻监督和半监督学寸中的过度拟合，79生成", "义信息的FC。与限制扰动噪声大小以保持图像语义完", "对抗性样本以进行数据增强，而【42，89]增强可转移特", "整的传统攻击不同，FAA允许在其生成的对抗性样本", "征以实现域发散最小化", "中进行大量扰动，但对图像语义的修改最小。此功能对", "大多数现有的对抗攻击方法通常会限制扰动噪声的", "于无监督域自适应至关重要，无监督域自适应通常涉", "大小，以最小程度地修改图像语义。然而，这种生成的", "及清晰的域间隙，因此需要具有较大扰动的对抗性样", "对抗性样本不能很好地解决领域自适应学习中的过度", "本。通过在训练中引入FAA生成的对抗性样本，网络", "拟合，这通常涉及公平大小的领域差距。我们设计了", "可以继续“随机游走”，避免过度拟合和漂移到具有平", "种创新的傅里叶对抗攻击技术，该技术允许在没有大", "坦损失景观E[6，36，40]的区域，从而导致更强大的域", "小约束的情况下生成对抗样本，但对图像语义的修改", "自适应", "最小，更多细节将在随后的小节中描述", "这项工作的贡献可以概括为三个方面。首先，我们", "确定无监督域适应中的过度拟合问题，并引人对抗性", "3.Method", "攻击，通过防止训练目标过度最小化来缓解过度拟合。", "我们通过傅立叶对抗攻击实现了强大的域自适应，", "其次，我们设计了一种创新的傅里叶对抗攻击（FAA）", "如图所示。2。训练分为两个阶段，即攻击阶段和防", "技术来生成新的对抗样本，以减轻域适应中的过度拟", "御阶段。给定训练图像，攻击阶段学习识别具有有限语", "合。FAA是通用的，可以有效地处理有监督的源损失", "义信息的正确FC，这些信息允许大幅度的扰动噪声。", "和无监督的目标损失。第三，我们在语义分割，对象检", "测和图像分类方面对多个计算机视觉任务进行了广泛", "它还学以最小的图像语义修改生成对抗性样本（具", "有可扰动的FC）。在防御阶段，生成的对抗性样本通", "的实验。所有实验均表明，我们的方法可以缓解过度拟", "过防止训练损失过度最小化来减轻过度拟合，更多细", "合并始终如一地改善域自适应", "节将在随后的小节中描述", "2. Related Works", "3.1. Task Definition", "域自适应：域自适应已被广泛研究以减轻数据注释", "我们专注于无监督域自适应（UDA）的问题。给定", "约束。大多数现有作品可大致分为三类。第一类是基于", "标记的源数据【Xs，Y和未标记的目标数据Xt，", "对抗性的训练，它采用鉴别器来对齐特征，输出或潜在", "我们的目标是学习在Xt上表现良好的任务模型F。", "空间46.78，48.75,67.24.80.76.25,88.95，18，28", "基线模型仅使用标记的源数据进行训练：", "中的源域和目标域。第二类是基于图像转换的，它调整", "图像外观以减轻域间隙[41，92，90，94]。第三类是基于", "Csup(Xs.Ys;F) =l(F(Xs),Ys),", "", "(1)", "自我训练的，它预测伪标签或最小化摘以指导从目标", "", "样本99.98.49.16.26，17进行送代学习", "其中l(）表示与精度相关的损失e.g.，标准交叉摘损", "域自适应涉及两种典型的训练损失，即对标记源数", "失", "据的有监督损失和对未标记目标数据的无监督损失。最", "3.2. Fourier Adversarial Attacking", "先进的方法倾向于过度最小化两种类型的损失，这直", "接导致具有次优适应的偏离模型，如图所示。1。我", "我们提出的傅里叶对抗攻击（FAA）生成对抗样本", "们设计了一种健壮的域自适应技术，通过防止损失过", "来攻击训练损失，以减轻域适应中的过度拟合，如图。", "度最小化来解决这个问题", "2。在对抗性样本生成中，它首先采用傅立叶变换将输", "网络训练中的过度拟合：过度拟合是深度网络训练", "人图像分解为多个频率分量（FC），然后向非语义FC", "中的常见现象，在深度学习和计算机视觉社区中得到", "注入扰动，这充许大幅度的扰动，但对图像语义的修改", "厂广泛的研究2，5，54，62，84，82。天多数现有作", "最小", "品都通过某些正则化策略来解决过拟合问题，例如权", "傅立叶分解：受JPEG[81，58]和频域学习[27，33]", "重衰减E[19]，退出E[71]，11正则化E[74]，混合E", "的启发，我们将图像转换为频率空间，并将其分解" ]

{ "arxiv_id": "2106.02874", "language": "zh", "page_idx": 1 }

[ [ "0.1013", "0.09216", "0.468", "0.09216", "0.468", "0.1061", "0.1013", "0.1061" ], [ "0.5024", "0.09155", "0.8906", "0.09216", "0.8906", "0.1067", "0.5024", "0.1061" ], [ "0.0801", "0.1067", "0.3865", "0.1042", "0.3865", "0.1218", "0.0801", "0.125" ], [ "0.505", "0.108", "0.789", "0.108", "0.789", "0.1218", "0.505", "0.1218" ], [ "0.1936", "0.1288", "0.353", "0.1332", "0.353", "0.1509", "0.1936", "0.1465" ], [ "0.5034", "0.125", "0.8906", "0.125", "0.8906", "0.1395", "0.5034", "0.1395" ], [ "0.4468", "0.1364", "0.468", "0.1364", "0.468", "0.1483", "0.4468", "0.1483" ], [ "0.5312", "0.1414", "0.5947", "0.1414", "0.5947", "0.1515", "0.5312", "0.1515" ], [ "0.5034", "0.1547", "0.8325", "0.1547", "0.8325", "0.1692", "0.5034", "0.1692" ], [ "0.08905", "0.1666", "0.0972", "0.1666", "0.0972", "0.1718", "0.08905", "0.1718" ], [ "0.0948", "0.1622", "0.468", "0.1622", "0.468", "0.1755", "0.0948", "0.1755" ], [ "0.0809", "0.1755", "0.4387", "0.1755", "0.4387", "0.1895", "0.0809", "0.1895" ], [ "0.511", "0.1718", "0.749", "0.1718", "0.749", "0.1818", "0.511", "0.1818" ], [ "0.5107", "0.1875", "0.546", "0.1875", "0.546", "0.1989", "0.5107", "0.1989" ], [ "0.546", "0.1863", "0.8896", "0.1863", "0.8896", "0.1963", "0.546", "0.1963" ], [ "0.10046", "0.1925", "0.4631", "0.1925", "0.4631", "0.2058", "0.10046", "0.2058" ], [ "0.2197", "0.2134", "0.327", "0.2166", "0.327", "0.2343", "0.2197", "0.2311" ], [ "0.5107", "0.2184", "0.527", "0.2184", "0.527", "0.2279", "0.5107", "0.2279" ], [ "0.546", "0.2166", "0.7837", "0.2166", "0.7837", "0.2299", "0.546", "0.2299" ], [ "0.5146", "0.2311", "0.528", "0.2349", "0.523", "0.2456", "0.509", "0.2412" ], [ "0.4468", "0.2197", "0.468", "0.2197", "0.468", "0.2329", "0.4468", "0.2329" ], [ "0.546", "0.2311", "0.6914", "0.2311", "0.6914", "0.2443", "0.546", "0.2443" ], [ "0.0784", "0.2443", "0.4697", "0.2412", "0.4697", "0.26", "0.0784", "0.2632" ], [ "0.513", "0.2468", "0.5293", "0.25", "0.527", "0.26", "0.5107", "0.2568" ], [ "0.5293", "0.25", "0.5415", "0.2474", "0.5435", "0.2524", "0.5312", "0.2551" ], [ "0.5474", "0.2468", "0.7837", "0.2468", "0.7837", "0.26", "0.5474", "0.26" ], [ "0.09644", "0.2632", "0.464", "0.2632", "0.464", "0.2734", "0.09644", "0.2734" ], [ "0.513", "0.2632", "0.527", "0.2632", "0.527", "0.2727", "0.513", "0.2727" ], [ "0.5293", "0.2644", "0.544", "0.2644", "0.544", "0.2715", "0.5293", "0.2715" ], [ "0.544", "0.26", "0.6914", "0.2615", "0.6914", "0.276", "0.544", "0.2747" ], [ "0.0866", "0.2664", "0.0989", "0.2664", "0.0989", "0.2727", "0.0866", "0.2727" ], [ "0.09314", "0.276", "0.4673", "0.276", "0.4673", "0.289", "0.09314", "0.289" ], [ "0.511", "0.2766", "0.5825", "0.2766", "0.5825", "0.2886", "0.511", "0.2886" ], [ "0.0809", "0.2903", "0.4673", "0.2903", "0.4673", "0.3037", "0.0809", "0.3037" ], [ "0.5063", "0.2922", "0.601", "0.2903", "0.601", "0.3037", "0.5063", "0.3062" ], [ "0.0809", "0.3062", "0.4673", "0.3062", "0.4673", "0.3193", "0.0809", "0.3193" ], [ "0.0768", "0.327", "0.2279", "0.33", "0.2279", "0.348", "0.0768", "0.3447" ], [ "0.505", "0.334", "0.8906", "0.334", "0.8906", "0.348", "0.505", "0.348" ], [ "0.1013", "0.3542", "0.468", "0.3542", "0.468", "0.3674", "0.1013", "0.3674" ], [ "0.505", "0.3499", "0.888", "0.3499", "0.888", "0.363", "0.505", "0.363" ], [ "0.0809", "0.3687", "0.468", "0.3687", "0.468", "0.3818", "0.0809", "0.3818" ], [ "0.5024", "0.363", "0.8896", "0.3643", "0.8896", "0.3774", "0.5024", "0.377" ], [ "0.0809", "0.3833", "0.4673", "0.3833", "0.4673", "0.3965", "0.0809", "0.3965" ], [ "0.5024", "0.379", "0.888", "0.379", "0.888", "0.3926", "0.5024", "0.3926" ], [ "0.505", "0.3945", "0.8906", "0.3945", "0.8906", "0.4077", "0.505", "0.4077" ], [ "0.0825", "0.401", "0.2418", "0.401", "0.2418", "0.411", "0.0825", "0.411" ], [ "0.0972", "0.4092", "0.4714", "0.4092", "0.4714", "0.43", "0.0972", "0.43" ], [ "0.5034", "0.4104", "0.8906", "0.4104", "0.8906", "0.4236", "0.5034", "0.4236" ], [ "0.0801", "0.428", "0.468", "0.4287", "0.468", "0.443", "0.0801", "0.4426" ], [ "0.505", "0.4268", "0.5605", "0.4268", "0.5605", "0.4375", "0.505", "0.4375" ], [ "0.0809", "0.4443", "0.4673", "0.4443", "0.4673", "0.4583", "0.0809", "0.4583" ], [ "0.576", "0.4507", "0.819", "0.4507", "0.819", "0.464", "0.576", "0.464" ], [ "0.0825", "0.461", "0.4656", "0.461", "0.4656", "0.4717", "0.0825", "0.4717" ], [ "0.0809", "0.4766", "0.4272", "0.4766", "0.4272", "0.4873", "0.0809", "0.4873" ], [ "0.584", "0.4717", "0.818", "0.4717", "0.818", "0.4849", "0.584", "0.4849" ], [ "0.1871", "0.5005", "0.4631", "0.5005", "0.4631", "0.515", "0.1871", "0.515" ], [ "0.505", "0.4988", "0.8896", "0.4988", "0.8896", "0.512", "0.505", "0.512" ], [ "0.0809", "0.509", "0.1805", "0.507", "0.1805", "0.521", "0.0809", "0.523" ], [ "0.1847", "0.5166", "0.2157", "0.5166", "0.2157", "0.529", "0.1847", "0.529" ], [ "0.2998", "0.5176", "0.3718", "0.5176", "0.3718", "0.529", "0.2998", "0.529" ], [ "0.505", "0.5146", "0.8906", "0.5146", "0.8906", "0.528", "0.505", "0.528" ], [ "0.5034", "0.529", "0.6665", "0.529", "0.6665", "0.5425", "0.5034", "0.5425" ], [ "0.2296", "0.536", "0.4714", "0.535", "0.4714", "0.55", "0.2296", "0.5503" ], [ "0.524", "0.5444", "0.8906", "0.5444", "0.8906", "0.558", "0.524", "0.558" ], [ "0.0809", "0.5645", "0.4714", "0.5645", "0.4714", "0.5776", "0.0809", "0.5776" ], [ "0.505", "0.5596", "0.8906", "0.5596", "0.8906", "0.5728", "0.505", "0.5728" ], [ "0.0809", "0.5776", "0.4656", "0.58", "0.4656", "0.5947", "0.0809", "0.5923" ], [ "0.505", "0.5747", "0.8896", "0.5747", "0.8896", "0.5884", "0.505", "0.5884" ], [ "0.0809", "0.5947", "0.468", "0.5947", "0.468", "0.608", "0.0809", "0.608" ], [ "0.505", "0.5913", "0.8896", "0.5913", "0.8896", "0.6016", "0.505", "0.6016" ], [ "0.0809", "0.61", "0.468", "0.61", "0.468", "0.6235", "0.0809", "0.6235" ], [ "0.5034", "0.605", "0.8906", "0.605", "0.8906", "0.6187", "0.5034", "0.6187" ], [ "0.0801", "0.6235", "0.392", "0.6255", "0.392", "0.6396", "0.0801", "0.637" ], [ "0.505", "0.622", "0.8228", "0.622", "0.8228", "0.632", "0.505", "0.632" ], [ "0.0989", "0.638", "0.3733", "0.6396", "0.3733", "0.654", "0.0989", "0.653" ], [ "0.5254", "0.6353", "0.8906", "0.6353", "0.8906", "0.6484", "0.5254", "0.6484" ], [ "0.505", "0.651", "0.888", "0.651", "0.888", "0.664", "0.505", "0.664" ], [ "0.1111", "0.6675", "0.4192", "0.6675", "0.4192", "0.681", "0.1111", "0.681" ], [ "0.4429", "0.6685", "0.4673", "0.6685", "0.4673", "0.6797", "0.4429", "0.6797" ], [ "0.505", "0.664", "0.8896", "0.664", "0.8896", "0.678", "0.505", "0.678" ], [ "0.4224", "0.6724", "0.4453", "0.6724", "0.4453", "0.6753", "0.4224", "0.6753" ], [ "0.1177", "0.681", "0.1316", "0.681", "0.1316", "0.6895", "0.1177", "0.6895" ], [ "0.5034", "0.6787", "0.8896", "0.6797", "0.8896", "0.6943", "0.5034", "0.6934" ], [ "0.0801", "0.6987", "0.33", "0.6987", "0.33", "0.712", "0.0801", "0.712" ], [ "0.505", "0.6963", "0.585", "0.6963", "0.585", "0.7085", "0.505", "0.7085" ], [ "0.10046", "0.7217", "0.468", "0.7217", "0.468", "0.735", "0.10046", "0.735" ], [ "0.5024", "0.7246", "0.6406", "0.7246", "0.6406", "0.738", "0.5024", "0.738" ], [ "0.0801", "0.735", "0.4673", "0.7363", "0.4673", "0.7505", "0.0801", "0.75" ], [ "0.0809", "0.7505", "0.4673", "0.7505", "0.4673", "0.7646", "0.0809", "0.7646" ], [ "0.523", "0.7476", "0.8896", "0.7476", "0.8896", "0.761", "0.523", "0.761" ], [ "0.0825", "0.7676", "0.468", "0.7676", "0.468", "0.781", "0.0825", "0.781" ], [ "0.5034", "0.7646", "0.8906", "0.7646", "0.8906", "0.778", "0.5034", "0.778" ], [ "0.0825", "0.781", "0.3733", "0.781", "0.3733", "0.7944", "0.0825", "0.7944" ], [ "0.505", "0.781", "0.863", "0.781", "0.863", "0.7915", "0.505", "0.7915" ], [ "0.10046", "0.797", "0.468", "0.797", "0.468", "0.81", "0.10046", "0.81" ], [ "0.5024", "0.802", "0.609", "0.803", "0.609", "0.8174", "0.5024", "0.8164" ], [ "0.0809", "0.8125", "0.286", "0.8125", "0.286", "0.8257", "0.0809", "0.8257" ], [ "0.523", "0.827", "0.8896", "0.827", "0.8896", "0.8403", "0.523", "0.8403" ], [ "0.2181", "0.839", "0.3284", "0.8384", "0.3284", "0.853", "0.2181", "0.854" ], [ "0.4468", "0.8403", "0.468", "0.8403", "0.468", "0.853", "0.4468", "0.853" ], [ "0.5063", "0.8433", "0.8896", "0.8433", "0.8896", "0.854", "0.5063", "0.854" ], [ "0.2279", "0.851", "0.2361", "0.854", "0.2336", "0.857", "0.2255", "0.854" ], [ "0.5063", "0.857", "0.8906", "0.857", "0.8906", "0.8706", "0.5063", "0.8706" ], [ "0.10046", "0.872", "0.4673", "0.872", "0.4673", "0.8853", "0.10046", "0.8853" ], [ "0.5024", "0.8706", "0.8896", "0.872", "0.8896", "0.886", "0.5024", "0.8853" ], [ "0.0801", "0.886", "0.468", "0.887", "0.468", "0.901", "0.0801", "0.8994" ], [ "0.5024", "0.887", "0.8906", "0.887", "0.8906", "0.901", "0.5024", "0.901" ] ]

[ "使用扰动的FCfe，我们将它们全部转换回（通过", "Algorithm1Theproposed Robust Domain Adaptation", "傅立叶逆变换）以获得对抗性样本&FAA：", "via Fourier adversarial attacking (FAA)", "FAA = F-1(C(fc),", "Require: Input data X ; task model F ; Fourier at-", "(3)", "tacker A", "Ensure: Learnt parameters of task model F", "", "其中C(）表示D;N）的逆过程，该过程通过对所", "有元素进行求和来将分解的zfc“合成”回完整的", "1:foriter =l to MarIterdo", "2:", "Samplea batch of X uniformly from the dataset", "傅里叶分解和对抗性攻击可以组合形成FAA如下：", "FAA=A(r),", "3:", "Defending phase (minimize loss):", "4:", "(4)", "Update F via Eq. 6", "其中A拍摄图像并通过FAA输出对抗性样本αFAA", "5:", "", "Attacking phase (maximize loss):", "如上所述。A具有两个带有可学习参数的子模块i.e", "6:", "", "Update A via Eq. 5", "", "门模块G和用于空间加权图S生成的单层神经网", "7: end for", "络。A中的其余操作是确定性的，例如傅立叶变换及其", "8:return F", "逆（F(）和F-1）），分解和重组（D(;N）和C(））", "3.3. FAA Training", "体来说，模型F可以用于语义分割，对象检测或图像", "建议的FAA涉及三种类型的损失，包括要攻击的任", "分类任务。数据可以是带有监督源损失的标记源数据", "务损失（i.e.，本工作中的UDA训练损失），限制可扰", "也可以是带有无监督目标损失的未标记目标数据，例", "动FC比例的门相关损失以及旨在最小化对图像语义", "如对抗性损失[75，48，80，89，55，42，29，66]，自训", "练损失E[99，98，41，83，37，90]或摘损失E[80，15]等。", "的攻击影响的重建损失", "给定输人图像XCRH×W×3，任务模型F和傅立", "以下是有监督损失和基于自我训练的无监督损失，以", "叶攻击者A，我们用C(X；A,F)表示任务损失。请注", "供参考：", "意，任务损失是指本工作中的UDA培训损失，但也可", "CsuP(Xs;A, F) = l(F(A(X), Y)", "以是其他类型的损失，例如有监督或无监督的培训损", "失。与门相关的损耗和重建损耗由以下公式定义", "Cun(Xt; A, F) = l(F(A(Xt)), Yt)", "Z(G(zfc)), if Z(G(zf) > N * p,", "其中l）表示与精度相关的损失，e.g.表示交叉摘损", "Lgat(X;A) =", "10,", "otherwise,", "失；A(X）和A(Xt）是受FAA干扰的输人，Yt是未", "标记目标数据的伪标签", "Lrec(X; A) = [R(X) - R(A(X)l,", "总而言之，我们的健壮域自适应具有双向训练框架，", "其中G(zfc）是A中的门控过程，如公式E2中所述。", "包括攻击阶段和防御阶段，如图。2。在攻击阶段，任", "R是一个带通滤波器，用于获取捕获语义信息（e.g.", "务模型F是固定的，攻击者A生成对抗性样本以增加", "结构和形状）的中频内容27]，因此一致性损失可以", "训练损失。在防御阶段，修复A并更新F以减少训练", "确保所选FC包含最少的语义信息。P是一个超参数，", "损失。这两个阶段以另一种方式进行，并形成一个双向", "它通过N*p限制了可扰动FC的最大数量", "培训框架（对抗性培训）。详情请参考算法1", "FAA的总体培训目标由以下公式制定：", "这个双向训练框架使用FAA来防止UDA损失的过", "度最小化，迫使它们围绕一个小的值振荡。换句话说", "max C(X;A, F) - Lgat(X;A) - Lrec(X;A)", "(5)", "它确保了任务模型F可以继续“随机游走”并漂移到", "", "A", "具有平坦损耗景观的区域E[6，36，40]，从而导致鲁棒", "3.4.RobustDomain Adaptation", "的域自适应", "第3节。2和3。3描述通过FAA生成对抗性样本", "4.Experiments", "的“攻击阶段”。本节介绍了建议的鲁棒域自适应技术，", "其中FAA用于在“防御阶段”中减轻UDA中的过拟", "本节介绍了实验，包括数据集和实现细节，语义分", "合。具体来说，我们应用FAA生成的对抗性样本来防", "割的域自适应（具有消融研究），对象检测和图像分类", "止域适应中UDA训练损失的过度最小化", "任务，以及讨论，更多细节将在随后的小节中描述", "给定任务模型F和傅立叶攻击者A，任务模型F", "4.1. Datasets", "的训练目标由以下公式表示：", "适应语义分割：我们考虑了两个从合成到真实的分", "min C(X;A, F)", "(6)", "割任务：1)GTA561→城市景观10和2)SYNTHIA", "", "[63]→城市景观。GTA5包含24,966合成图像，并与", "攻击（i.e.，方程式E5）和防御（i.e.，方程式E6", "Cityscapes共享19类别。对于SYNTHIA，我们使用", "）的优化功能都是通用的，适用于各种任务和数据。具", "“SYNTHIA-RAND-CITYSCAPES”，其中包含9,400" ]

{ "arxiv_id": "2106.02874", "language": "zh", "page_idx": 3 }

[ [ "0.1184", "0.0873", "0.482", "0.0873", "0.482", "0.0974", "0.1184", "0.0974" ], [ "0.1184", "0.1027", "0.4836", "0.1027", "0.4836", "0.1164", "0.1184", "0.1164" ], [ "0.1184", "0.11816", "0.4844", "0.11816", "0.4844", "0.1318", "0.1184", "0.1318" ], [ "0.1184", "0.1354", "0.4844", "0.1354", "0.4844", "0.1484", "0.1184", "0.1484" ], [ "0.11755", "0.1497", "0.4836", "0.1497", "0.4836", "0.1633", "0.11755", "0.1633" ], [ "0.1184", "0.1675", "0.4836", "0.1675", "0.4836", "0.1812", "0.1184", "0.1812" ], [ "0.11755", "0.1829", "0.1823", "0.1829", "0.1823", "0.1978", "0.11755", "0.1978" ], [ "0.11584", "0.2073", "0.486", "0.2073", "0.486", "0.2203", "0.11584", "0.2203" ], [ "0.1318", "0.2227", "0.4878", "0.2227", "0.4878", "0.2369", "0.1318", "0.2369" ], [ "0.1301", "0.2393", "0.4475", "0.2393", "0.4475", "0.253", "0.1301", "0.253" ], [ "0.11584", "0.2542", "0.486", "0.2546", "0.486", "0.2695", "0.11584", "0.2683" ], [ "0.1301", "0.2725", "0.486", "0.2725", "0.486", "0.2861", "0.1301", "0.2861" ], [ "0.1293", "0.2874", "0.465", "0.288", "0.465", "0.3027", "0.1293", "0.3018" ], [ "0.11755", "0.3047", "0.4878", "0.3047", "0.4878", "0.3196", "0.11755", "0.3196" ], [ "0.1301", "0.3218", "0.4878", "0.3218", "0.4878", "0.335", "0.1301", "0.335" ], [ "0.1301", "0.3362", "0.486", "0.3362", "0.486", "0.3499", "0.1301", "0.3499" ], [ "0.1318", "0.354", "0.4727", "0.354", "0.4727", "0.3677", "0.1318", "0.3677" ], [ "0.11584", "0.3706", "0.486", "0.3706", "0.486", "0.3843", "0.11584", "0.3843" ], [ "0.1301", "0.3855", "0.4844", "0.3855", "0.4844", "0.3984", "0.1301", "0.3984" ], [ "0.1301", "0.4026", "0.4844", "0.4026", "0.4844", "0.4163", "0.1301", "0.4163" ], [ "0.1318", "0.4187", "0.4358", "0.4187", "0.4358", "0.4316", "0.1318", "0.4316" ], [ "0.534", "0.4407", "0.879", "0.4407", "0.879", "0.4536", "0.534", "0.4536" ], [ "0.11755", "0.4465", "0.1511", "0.4465", "0.1511", "0.4578", "0.11755", "0.4578" ], [ "0.1486", "0.4495", "0.1545", "0.4495", "0.1545", "0.4548", "0.1486", "0.4548" ], [ "0.1503", "0.4465", "0.2747", "0.4465", "0.2747", "0.4602", "0.1503", "0.4602" ], [ "0.5146", "0.4573", "0.881", "0.4573", "0.881", "0.4702", "0.5146", "0.4702" ], [ "0.1184", "0.4702", "0.2913", "0.4702", "0.2913", "0.484", "0.1184", "0.484" ], [ "0.513", "0.4727", "0.882", "0.4727", "0.882", "0.4863", "0.513", "0.4863" ], [ "0.51", "0.4868", "0.882", "0.488", "0.882", "0.503", "0.51", "0.502" ], [ "0.513", "0.505", "0.882", "0.505", "0.882", "0.5186", "0.513", "0.5186" ], [ "0.513", "0.521", "0.858", "0.521", "0.858", "0.5337", "0.513", "0.5337" ], [ "0.513", "0.5444", "0.881", "0.5444", "0.881", "0.558", "0.513", "0.558" ], [ "0.513", "0.561", "0.881", "0.561", "0.881", "0.5747", "0.513", "0.5747" ], [ "0.513", "0.5776", "0.879", "0.5776", "0.879", "0.5884", "0.513", "0.5884" ], [ "0.513", "0.5933", "0.881", "0.5933", "0.881", "0.607", "0.513", "0.607" ], [ "0.51", "0.608", "0.881", "0.608", "0.881", "0.6216", "0.51", "0.6216" ], [ "0.5146", "0.626", "0.881", "0.626", "0.881", "0.6396", "0.5146", "0.6396" ], [ "0.513", "0.641", "0.881", "0.641", "0.881", "0.655", "0.513", "0.655" ], [ "0.513", "0.6577", "0.882", "0.6577", "0.882", "0.6714", "0.513", "0.6714" ], [ "0.5156", "0.676", "0.879", "0.676", "0.879", "0.686", "0.5156", "0.686" ], [ "0.513", "0.6904", "0.879", "0.6904", "0.879", "0.7036", "0.513", "0.7036" ], [ "0.511", "0.705", "0.879", "0.705", "0.879", "0.7188", "0.511", "0.7188" ], [ "0.513", "0.7227", "0.8613", "0.7227", "0.8613", "0.733", "0.513", "0.733" ], [ "0.513", "0.737", "0.881", "0.737", "0.881", "0.752", "0.513", "0.752" ], [ "0.513", "0.7544", "0.881", "0.7544", "0.881", "0.767", "0.513", "0.767" ], [ "0.513", "0.771", "0.8774", "0.771", "0.8774", "0.781", "0.513", "0.781" ], [ "0.5146", "0.786", "0.881", "0.786", "0.881", "0.8", "0.5146", "0.8" ], [ "0.5146", "0.8027", "0.623", "0.8027", "0.623", "0.814", "0.5146", "0.814" ], [ "0.511", "0.8496", "0.8345", "0.8496", "0.8345", "0.8643", "0.511", "0.8643" ], [ "0.513", "0.8896", "0.879", "0.8896", "0.879", "0.903", "0.513", "0.903" ], [ "0.513", "0.905", "0.5903", "0.905", "0.5903", "0.9194", "0.513", "0.9194" ] ]

[ "(RobustModelC压缩）。首先：我们利用深度", "神经网络的不确定性来量化训练样本的难度", "这是由通过模型修剪获得的原始大型网络的多", "个子网络的样本预测方差给出的。其次，我们", "利用这种特定于样本的度量来平滑和正则化不", "同的压缩技术族。这项工作的主要贡献可以总", "结如下：", ".We perform a comprehensive analysis to evalu-", "ate the OOD generalization ability and robust-", "ness of compressed models for NLU tasks.", ".We further analyze plausible reasons for the", "low generalizability of compressed models and", "demonstrate connections to shortcut learning", ". We propose a mitigation framework for regu-", "larizing model compression, termed as RMC,", "which smooths the knowledge distillation train-", "ing based on the estimated sample difficulties.", ". We perform experiments to demonstrate that our", "RMCframeworkimproves OODgeneralization", "while not sacrificing the standard in-distribution", "task performance on multiple NLU tasks.", "为了减轻一般PLM的偏见，有一些初", "2", "", "Related Work", "步的工作，包括专家产品E(Clarketal.，2019;", "快捷方式学习和缓解。", "Heetal.,2019;Sanhetal.,2021），重新加权E", "(Schuster et al., 2019; Yaghoobzadeh et al., 2019;", "Utamaetal.,2020），对抗性训练E(Staceyetal.", "2020），后验正则化E（Chengetal.，2021）等", "在模型压缩中的鲁棒性。当前评估模型压缩", "性能的实践主要集中在标准基准性能（Zhu", "etal..2020:Wangetal..2021）。在计算机视觉领", "域，先前的工作表明压缩模型在压缩识别样本", "(CIE)E(Hookeretal.,2019)中表现不佳，并且", "压缩放大了对某些人口统计学El(Hookeretal.,", "2020）的算法偏见。与我们最相似的工作是两", "个并发工作E(Xu et al.,2021a;Liet al.,2021)", "它们研究了超出自然语言理解任务标准基准的", "压缩模型的性能。然而，这两项工作主要集中", "在评估压缩模型相对于对抗性攻击场景的鲁棒", "性，即。e。TextFooler（Jinetal..2020）和统", "对抗框架（Lietal.,2021）。相反，我们在OOD", "测试集中全面描述了BERT压缩的鲁棒性，以", "探索压缩技术的OOD通用性。此外，我们使", "用此健壮性分析的见解来设计可概括且健壮的", "模型压缩框架", "3Are Compressed Models Robust?", "我们进行了全面的分析，以评估压缩语言模型", "的鲁棒性" ]

{ "arxiv_id": "2110.08419", "language": "zh", "page_idx": 2 }

[ [ "0.1142", "0.0849", "0.3594", "0.0861", "0.3594", "0.10095", "0.1142", "0.0998" ], [ "0.511", "0.0849", "0.881", "0.0873", "0.881", "0.10095", "0.511", "0.0986" ], [ "0.1184", "0.1081", "0.486", "0.1081", "0.486", "0.1217", "0.1184", "0.1217" ], [ "0.5264", "0.1027", "0.881", "0.1027", "0.881", "0.1164", "0.5264", "0.1164" ], [ "0.1184", "0.1217", "0.1738", "0.12354", "0.1721", "0.1383", "0.11755", "0.136" ], [ "0.5264", "0.1194", "0.881", "0.1194", "0.881", "0.133", "0.5264", "0.133" ], [ "0.5264", "0.1354", "0.881", "0.1354", "0.881", "0.1484", "0.5264", "0.1484" ], [ "0.1184", "0.1426", "0.4844", "0.1426", "0.4844", "0.1561", "0.1184", "0.1561" ], [ "0.5283", "0.1509", "0.881", "0.1509", "0.881", "0.1644", "0.5283", "0.1644" ], [ "0.1184", "0.158", "0.486", "0.158", "0.486", "0.1716", "0.1184", "0.1716" ], [ "0.5264", "0.1663", "0.879", "0.1663", "0.879", "0.1799", "0.5264", "0.1799" ], [ "0.12006", "0.1758", "0.4836", "0.1758", "0.4836", "0.1859", "0.12006", "0.1859" ], [ "0.5283", "0.1829", "0.881", "0.1829", "0.881", "0.1965", "0.5283", "0.1965" ], [ "0.1184", "0.1907", "0.4836", "0.1907", "0.4836", "0.2037", "0.1184", "0.2037" ], [ "0.525", "0.199", "0.879", "0.1978", "0.879", "0.212", "0.525", "0.2131" ], [ "0.11584", "0.2048", "0.4844", "0.206", "0.4844", "0.2203", "0.11584", "0.2197" ], [ "0.5283", "0.2131", "0.8193", "0.2156", "0.8193", "0.2299", "0.5283", "0.228" ], [ "0.1184", "0.2227", "0.4844", "0.2227", "0.4844", "0.2363", "0.1184", "0.2363" ], [ "0.517", "0.2369", "0.531", "0.2369", "0.531", "0.2434", "0.517", "0.2434" ], [ "0.525", "0.2328", "0.882", "0.2328", "0.882", "0.2477", "0.525", "0.2477" ], [ "0.1184", "0.2393", "0.4844", "0.2393", "0.4844", "0.253", "0.1184", "0.253" ], [ "0.11755", "0.2546", "0.486", "0.2546", "0.486", "0.2683", "0.11755", "0.2683" ], [ "0.5283", "0.2494", "0.882", "0.2494", "0.882", "0.2632", "0.5283", "0.2632" ], [ "0.5264", "0.2642", "0.881", "0.2654", "0.881", "0.2798", "0.5264", "0.279" ], [ "0.1184", "0.2708", "0.4844", "0.2708", "0.4844", "0.284", "0.1184", "0.284" ], [ "0.11755", "0.2852", "0.4836", "0.2852", "0.4836", "0.2986", "0.11755", "0.2986" ], [ "0.5264", "0.2808", "0.879", "0.2808", "0.879", "0.2944", "0.5264", "0.2944" ], [ "0.1184", "0.3018", "0.471", "0.3018", "0.471", "0.3154", "0.1184", "0.3154" ], [ "0.5264", "0.2976", "0.8774", "0.2964", "0.8774", "0.3113", "0.5264", "0.3123" ], [ "0.4685", "0.308", "0.4802", "0.308", "0.4802", "0.3123", "0.4685", "0.3123" ], [ "0.1184", "0.3184", "0.4878", "0.3184", "0.4878", "0.332", "0.1184", "0.332" ], [ "0.5283", "0.3135", "0.881", "0.3135", "0.881", "0.3267", "0.5283", "0.3267" ], [ "0.5264", "0.3289", "0.881", "0.33", "0.881", "0.3445", "0.5264", "0.3433" ], [ "0.12006", "0.3362", "0.1721", "0.3362", "0.1721", "0.3474", "0.12006", "0.3474" ], [ "0.11755", "0.3516", "0.4836", "0.3516", "0.4836", "0.3652", "0.11755", "0.3652" ], [ "0.5283", "0.3467", "0.771", "0.3467", "0.771", "0.361", "0.5283", "0.361" ], [ "0.1184", "0.3694", "0.4844", "0.3694", "0.4844", "0.383", "0.1184", "0.383" ], [ "0.531", "0.3723", "0.881", "0.3723", "0.881", "0.386", "0.531", "0.386" ], [ "0.11755", "0.3855", "0.4878", "0.3855", "0.4878", "0.3984", "0.11755", "0.3984" ], [ "0.513", "0.3896", "0.882", "0.3896", "0.882", "0.4026", "0.513", "0.4026" ], [ "0.11755", "0.402", "0.486", "0.402", "0.486", "0.415", "0.11755", "0.415" ], [ "0.511", "0.4038", "0.681", "0.405", "0.681", "0.4192", "0.511", "0.4187" ], [ "0.1184", "0.4187", "0.4844", "0.4187", "0.4844", "0.4316", "0.1184", "0.4316" ], [ "0.513", "0.4329", "0.5557", "0.4329", "0.5557", "0.4448", "0.513", "0.4448" ], [ "0.555", "0.4312", "0.6943", "0.4316", "0.6943", "0.4465", "0.555", "0.4453" ], [ "0.1184", "0.4507", "0.4895", "0.4507", "0.4895", "0.4644", "0.1184", "0.4644" ], [ "0.11755", "0.434", "0.4836", "0.434", "0.4836", "0.4478", "0.11755", "0.4478" ], [ "0.513", "0.4531", "0.881", "0.4531", "0.881", "0.466", "0.513", "0.466" ], [ "0.12006", "0.466", "0.4106", "0.466", "0.4106", "0.4797", "0.12006", "0.4797" ], [ "0.5146", "0.4697", "0.881", "0.4697", "0.881", "0.4827", "0.5146", "0.4827" ], [ "0.1184", "0.485", "0.4844", "0.485", "0.4844", "0.4988", "0.1184", "0.4988" ], [ "0.513", "0.485", "0.881", "0.485", "0.881", "0.4988", "0.513", "0.4988" ], [ "0.1184", "0.5005", "0.4844", "0.5005", "0.4844", "0.514", "0.1184", "0.514" ], [ "0.511", "0.502", "0.879", "0.502", "0.879", "0.5156", "0.511", "0.5156" ], [ "0.11755", "0.517", "0.4844", "0.517", "0.4844", "0.531", "0.11755", "0.531" ], [ "0.513", "0.5186", "0.881", "0.5186", "0.881", "0.532", "0.513", "0.532" ], [ "0.12006", "0.535", "0.4844", "0.535", "0.4844", "0.546", "0.12006", "0.546" ], [ "0.511", "0.533", "0.879", "0.533", "0.879", "0.5464", "0.511", "0.5464" ], [ "0.1184", "0.55", "0.4844", "0.55", "0.4844", "0.5635", "0.1184", "0.5635" ], [ "0.5146", "0.552", "0.8774", "0.552", "0.8774", "0.5625", "0.5146", "0.5625" ], [ "0.12006", "0.568", "0.4836", "0.568", "0.4836", "0.5776", "0.12006", "0.5776" ], [ "0.511", "0.564", "0.881", "0.5654", "0.881", "0.58", "0.511", "0.579" ], [ "0.1184", "0.58", "0.4844", "0.58", "0.4844", "0.5933", "0.1184", "0.5933" ], [ "0.511", "0.5806", "0.879", "0.5806", "0.879", "0.594", "0.511", "0.594" ], [ "0.11755", "0.5967", "0.486", "0.5967", "0.486", "0.61", "0.11755", "0.61" ], [ "0.511", "0.5986", "0.881", "0.5986", "0.881", "0.6123", "0.511", "0.6123" ], [ "0.1184", "0.6143", "0.4878", "0.6143", "0.4878", "0.6274", "0.1184", "0.6274" ], [ "0.513", "0.6143", "0.881", "0.6143", "0.881", "0.6274", "0.513", "0.6274" ], [ "0.513", "0.6313", "0.879", "0.6313", "0.879", "0.6445", "0.513", "0.6445" ], [ "0.11584", "0.64", "0.3208", "0.64", "0.3208", "0.654", "0.11584", "0.654" ], [ "0.513", "0.6465", "0.879", "0.6465", "0.879", "0.66", "0.513", "0.66" ], [ "0.1184", "0.661", "0.4878", "0.661", "0.4878", "0.675", "0.1184", "0.675" ], [ "0.51", "0.661", "0.881", "0.661", "0.881", "0.675", "0.51", "0.675" ], [ "0.1184", "0.6777", "0.486", "0.6777", "0.486", "0.6914", "0.1184", "0.6914" ], [ "0.513", "0.6787", "0.881", "0.6787", "0.881", "0.6924", "0.513", "0.6924" ], [ "0.11755", "0.6924", "0.486", "0.6924", "0.486", "0.706", "0.11755", "0.706" ], [ "0.513", "0.695", "0.881", "0.695", "0.881", "0.708", "0.513", "0.708" ], [ "0.11755", "0.7104", "0.1663", "0.7104", "0.1663", "0.7246", "0.11755", "0.7246" ], [ "0.511", "0.709", "0.881", "0.7114", "0.881", "0.7246", "0.511", "0.7227" ], [ "0.511", "0.7256", "0.66", "0.727", "0.66", "0.741", "0.511", "0.7407" ], [ "0.11584", "0.7363", "0.4878", "0.7363", "0.4878", "0.75", "0.11584", "0.75" ], [ "0.1318", "0.753", "0.4895", "0.753", "0.4895", "0.7666", "0.1318", "0.7666" ], [ "0.509", "0.753", "0.7734", "0.753", "0.7734", "0.7666", "0.509", "0.7666" ], [ "0.1301", "0.7686", "0.4878", "0.767", "0.4878", "0.781", "0.1301", "0.782" ], [ "0.5146", "0.775", "0.879", "0.775", "0.879", "0.789", "0.5146", "0.789" ], [ "0.1318", "0.785", "0.4878", "0.785", "0.4878", "0.799", "0.1318", "0.799" ], [ "0.513", "0.7915", "0.881", "0.7915", "0.881", "0.8047", "0.513", "0.8047" ], [ "0.1318", "0.8027", "0.486", "0.8027", "0.486", "0.8135", "0.1318", "0.8135" ], [ "0.511", "0.807", "0.879", "0.807", "0.879", "0.821", "0.511", "0.821" ], [ "0.1318", "0.818", "0.4895", "0.818", "0.4895", "0.8306", "0.1318", "0.8306" ], [ "0.511", "0.8237", "0.879", "0.8237", "0.879", "0.8374", "0.511", "0.8374" ], [ "0.1335", "0.835", "0.486", "0.835", "0.486", "0.8457", "0.1335", "0.8457" ], [ "0.5156", "0.8403", "0.881", "0.8403", "0.881", "0.854", "0.5156", "0.854" ], [ "0.1318", "0.851", "0.2468", "0.851", "0.2468", "0.8633", "0.1318", "0.8633" ], [ "0.513", "0.8555", "0.881", "0.8555", "0.881", "0.869", "0.513", "0.869" ], [ "0.1394", "0.869", "0.4878", "0.869", "0.4878", "0.883", "0.1394", "0.883" ], [ "0.511", "0.8706", "0.882", "0.8716", "0.882", "0.8857", "0.511", "0.8853" ], [ "0.12006", "0.8843", "0.4836", "0.8843", "0.4836", "0.8945", "0.12006", "0.8945" ], [ "0.509", "0.882", "0.875", "0.883", "0.875", "0.905", "0.509", "0.904" ], [ "0.12006", "0.8955", "0.482", "0.8955", "0.482", "0.9062", "0.12006", "0.9062" ], [ "0.1184", "0.9067", "0.486", "0.9067", "0.486", "0.9175", "0.1184", "0.9175" ], [ "0.51", "0.904", "0.8833", "0.904", "0.8833", "0.9214", "0.51", "0.9214" ] ]

[ "3.1 Compression Techniques", ".FEVER(Thorne et al., 2018) : This is a fact", "我们考虑两个流行的压缩族，即知识蒸馏和模", "verification dataset. Recent studies indicate that", "型修剪", "there are strong shortcuts in the claims (Utama", "et al., 2020).To facilitate the robustness", "知识蒸馏：这里的目标是通过使用知识蒸馏来", "and generalization evaluation of fact verification", "模仿较大的教师模型的行为来训练小型模型E", "models, two symmetric test sets (i.e., Sym vl", "（Hintonetal..2015）。在这项工作中，我们专", "and Sym v2) were created, where bias exists", "注于任务不可知的蒸馏。特别是，我们考虑从", "in the symmetric pairs (Schuster et al., 2019)", "BERT-碱蒸馏的DistilBERT(Sanh etal.,2019)", ". Both 00D test sets have 712 samples.", "和MiniLM（Wangetal.,2020）。为了进行公", "", "QQP : The task is to predict whether a pair of", "平的比较，我们选择具有相似容量的压缩模型", "（本工作中的66M参数）。为了评估压缩技术", "questions is semantically equivalent. We con-", "sider the OOD test set PAWS-qqp, which con-", "对模型鲁棒性的影响，我们还考虑了类似容量", "较小的模型，而不使用知识蒸馏。这些是通过", "tains 677 test samples generated from QQP cor", "简单的截断获得的，我们保留了大模型的第", "pus (Zhang et al., 2019; Yang et al., 2019)", "", "个6层，并通过从头开始预训练一个较小的6", "Besides, we also consider the PAWS-wiki OOD", "test set, which consists of 8, 000 test samples", "层模型", "迭代幅度修剪：这是一个特定于任务的非结构", "generated from Wikipedia pages.", "化修剪方法（Sanhetal.,2020）。在每个下游任", "对于所有三个任务，我们都将准确性用作评", "务的微调过程中，删除具有最低幅度的权重，", "估指标，并在分布内开发集和OOD测试集上", "直到修剪后的模型达到目标稀疏性。请注意，", "评估压缩模型的性能", "我们使用标准的修剪技术，而不是基于LTH", "3.3", "Evaluation Setup", "etal.,2020）。我们还考虑了不同的修剪比率，", "的修剪（彩票假设），该技术使用重绕E(Chen", "在这项工作中，我们使用无套管BERT-base作", "以获得具有不同稀疏程度的修剪模型", "为教师网络，并研究其压缩变体的鲁棒性。最", "结构化修剪：此方法系列基于以下假设：注意", "终模型由基于BERT的编码器（或其压缩变", "头E(Prasanna et al., 2020; Voita et al., 2019; Bian", "体）组成，带有分类头（池化输出顶部的线", "etal.,2021;Chenetal.,2021）中存在允余。我们", "性层）。最近的研究表明，学习率和训练时", "还考虑特定于任务的修剪。在每个任务的微调", "间等因素可能会对鲁棒性产生重大影响（Tu", "过程中，它会根据它们对模型预测的重要性来", "etal.，2020）。特别是，增加训练时期可以帮助", "修剪整个注意力。请参阅Sec。A在附录中了", "改善OOD测试集的泛化。在这项工作中，我", "解更多详情。我们总共修剪了20%个注意力", "们将重点放在与未压缩老师相比的压缩模型的", "头(i。e。，28注意头)。进一步修剪会增加稀疏", "relativerobustness上，而不是它们的绝对精度。", "性，并显着降低模型在分布内开发集上的性能", "为了进行公平的比较，我们统一了所有模型的", "实验设置。我们使用具有权重衰减（Loshchilov", "3.2Evaluation Datasets", "andHutter，2017）的Adam优化器，其中学习", "为了评估上一节介绍的压缩模型的鲁棒性，我", "率固定为2e-5，我们在所有数据集上训练5", "们使用了三个NLU任务，包括MNLI，FEVER", "个时期的所有模型。我们使用PyTorch进行实", "和QQP1。请参阅Sec。B在附录中了解更多", "验，并使用Huggingface模型池中的预训练模", "详情。", "型（Wolfetal.2019）。我们报告了所有实验三", "次运行的平均结果", ". MNLI (Williams et al., 2018) : This is a nat-", "ural language inference task. In this work, we", "3.4Relative Robustness Metric", "report the accuracy metric on the matched sub-", "正如我们稍后演示的那样，随着压缩比或模型", "set. We use HANS (McCoy et al., 2019) as the", "稀疏性的增加，对于域内和OOD测试集，较", "adversarial test set,which contains 30,000 syn-", "小模型的性能都会下降。为了比较分布内任", "thetic samples. Models that exploit shortcut fea-", "务性能和OOD泛化性之间的差距，我们定义", "tureshavebeen showntoperformpoorlyonthe", "了一个新的指标来衡量压缩模型相对于未压", "HANS test set.", "缩BERT库（教师）的性能差距。首先，我们计", "'MNLI, FEVER, and QQP are the three most widely used", "算分布式开发集和OOD测试集之间的精度差", "datasets to evaluate the shortcut learning/bias behavior and", "距，作为BERT-base的FFooD（用△BERTbas", "OOD generalization of PLMs in the literature(Tu et al.", "2020; He et al.,2019;Clark et al.,2019; Schuster et al.,2019)", "表示)；及其压缩变体（用△compressed表示)。其" ]

{ "arxiv_id": "2110.08419", "language": "zh", "page_idx": 3 }

[ [ "0.33", "0.0861", "0.3687", "0.0861", "0.3687", "0.0986", "0.33", "0.0986" ], [ "0.4592", "0.0849", "0.509", "0.0849", "0.509", "0.0998", "0.4592", "0.0998" ], [ "0.635", "0.0861", "0.6665", "0.0861", "0.6665", "0.0986", "0.635", "0.0986" ], [ "0.236", "0.0968", "0.4165", "0.0998", "0.4165", "0.11755", "0.236", "0.114" ], [ "0.4048", "0.0998", "0.559", "0.10156", "0.559", "0.1164", "0.4048", "0.114" ], [ "0.555", "0.0998", "0.706", "0.1027", "0.706", "0.11755", "0.555", "0.114" ], [ "0.693", "0.0998", "0.8237", "0.10156", "0.8237", "0.11755", "0.693", "0.1152" ], [ "0.2451", "0.1164", "0.2837", "0.11816", "0.2822", "0.1312", "0.2435", "0.1289" ], [ "0.1704", "0.1027", "0.22", "0.1027", "0.22", "0.114", "0.1704", "0.114" ], [ "0.1704", "0.11816", "0.236", "0.11816", "0.236", "0.13", "0.1704", "0.13" ], [ "0.293", "0.11816", "0.3225", "0.11816", "0.3225", "0.13", "0.293", "0.13" ], [ "0.3315", "0.11816", "0.3628", "0.11816", "0.3628", "0.1312", "0.3315", "0.1312" ], [ "0.4106", "0.11816", "0.4417", "0.11816", "0.4417", "0.1312", "0.4106", "0.1312" ], [ "0.4475", "0.11816", "0.4768", "0.11816", "0.4768", "0.1312", "0.4475", "0.1312" ], [ "0.4878", "0.11816", "0.517", "0.11816", "0.517", "0.1312", "0.4878", "0.1312" ], [ "0.5615", "0.11816", "0.5938", "0.11816", "0.5938", "0.1312", "0.5615", "0.1312" ], [ "0.608", "0.11816", "0.638", "0.11816", "0.638", "0.1312", "0.608", "0.1312" ], [ "0.663", "0.11816", "0.692", "0.11816", "0.692", "0.1312", "0.663", "0.1312" ], [ "0.1704", "0.1342", "0.2032", "0.1342", "0.2032", "0.1467", "0.1704", "0.1467" ], [ "0.2435", "0.1342", "0.2854", "0.1342", "0.2854", "0.1478", "0.2435", "0.1478" ], [ "0.293", "0.1342", "0.3242", "0.1342", "0.3242", "0.1467", "0.293", "0.1467" ], [ "0.3333", "0.1342", "0.3628", "0.1342", "0.3628", "0.1467", "0.3333", "0.1467" ], [ "0.3704", "0.133", "0.4417", "0.133", "0.4417", "0.1478", "0.3704", "0.1478" ], [ "0.4475", "0.1342", "0.4768", "0.1342", "0.4768", "0.1467", "0.4475", "0.1467" ], [ "0.4878", "0.1342", "0.519", "0.1342", "0.519", "0.1467", "0.4878", "0.1467" ], [ "0.5225", "0.1342", "0.559", "0.1342", "0.559", "0.1467", "0.5225", "0.1467" ], [ "0.555", "0.1342", "0.592", "0.1342", "0.592", "0.1467", "0.555", "0.1467" ], [ "0.608", "0.1342", "0.638", "0.1342", "0.638", "0.1467", "0.608", "0.1467" ], [ "0.665", "0.1342", "0.693", "0.1342", "0.693", "0.1467", "0.665", "0.1467" ], [ "0.7036", "0.1342", "0.739", "0.1342", "0.739", "0.1467", "0.7036", "0.1467" ], [ "0.7656", "0.1342", "0.802", "0.1342", "0.802", "0.1467", "0.7656", "0.1467" ], [ "0.2418", "0.1455", "0.288", "0.1455", "0.288", "0.1603", "0.2418", "0.1603" ], [ "0.3704", "0.1455", "0.4417", "0.1455", "0.4417", "0.1603", "0.3704", "0.1603" ], [ "0.519", "0.1455", "0.5938", "0.1455", "0.5938", "0.1603", "0.519", "0.1603" ], [ "0.293", "0.1467", "0.3225", "0.1467", "0.3225", "0.1592", "0.293", "0.1592" ], [ "0.1704", "0.1467", "0.2024", "0.1467", "0.2024", "0.1592", "0.1704", "0.1592" ], [ "0.3315", "0.1467", "0.3628", "0.1467", "0.3628", "0.158", "0.3315", "0.158" ], [ "0.4475", "0.1467", "0.4768", "0.1467", "0.4768", "0.158", "0.4475", "0.158" ], [ "0.4878", "0.1467", "0.519", "0.1467", "0.519", "0.1592", "0.4878", "0.1592" ], [ "0.608", "0.1467", "0.638", "0.1467", "0.638", "0.1592", "0.608", "0.1592" ], [ "0.663", "0.1467", "0.693", "0.1467", "0.693", "0.1592", "0.663", "0.1592" ], [ "0.702", "0.1467", "0.739", "0.1467", "0.739", "0.1592", "0.702", "0.1592" ], [ "0.764", "0.1467", "0.7983", "0.1467", "0.7983", "0.1592", "0.764", "0.1592" ], [ "0.293", "0.1561", "0.3225", "0.158", "0.3208", "0.1704", "0.2896", "0.1675" ], [ "0.555", "0.1561", "0.592", "0.158", "0.5903", "0.1704", "0.553", "0.1687" ], [ "0.2418", "0.1567", "0.287", "0.1567", "0.287", "0.1716", "0.2418", "0.1716" ], [ "0.1704", "0.158", "0.2032", "0.158", "0.2032", "0.1704", "0.1704", "0.1704" ], [ "0.3315", "0.158", "0.3628", "0.158", "0.3628", "0.1692", "0.3315", "0.1692" ], [ "0.372", "0.158", "0.44", "0.158", "0.44", "0.1692", "0.372", "0.1692" ], [ "0.4475", "0.158", "0.4785", "0.158", "0.4785", "0.1704", "0.4475", "0.1704" ], [ "0.4878", "0.158", "0.517", "0.158", "0.517", "0.1704", "0.4878", "0.1704" ], [ "0.5225", "0.158", "0.559", "0.158", "0.559", "0.1692", "0.5225", "0.1692" ], [ "0.608", "0.158", "0.6396", "0.158", "0.6396", "0.1704", "0.608", "0.1704" ], [ "0.663", "0.158", "0.693", "0.158", "0.693", "0.1704", "0.663", "0.1704" ], [ "0.702", "0.158", "0.738", "0.158", "0.738", "0.1704", "0.702", "0.1704" ], [ "0.764", "0.158", "0.8003", "0.158", "0.8003", "0.1704", "0.764", "0.1704" ], [ "0.2435", "0.1687", "0.2854", "0.1704", "0.2837", "0.1829", "0.2418", "0.1812" ], [ "0.7036", "0.1687", "0.738", "0.1704", "0.7363", "0.1829", "0.702", "0.1812" ], [ "0.3342", "0.1692", "0.3628", "0.1692", "0.3628", "0.1812", "0.3342", "0.1812" ], [ "0.1704", "0.1692", "0.2032", "0.1692", "0.2032", "0.1818", "0.1704", "0.1818" ], [ "0.4895", "0.1692", "0.517", "0.1692", "0.517", "0.1812", "0.4895", "0.1812" ], [ "0.6094", "0.1692", "0.638", "0.1692", "0.638", "0.1818", "0.6094", "0.1818" ], [ "0.663", "0.1692", "0.693", "0.1692", "0.693", "0.1818", "0.663", "0.1818" ], [ "0.7656", "0.1704", "0.8003", "0.1704", "0.8003", "0.1818", "0.7656", "0.1818" ], [ "0.4492", "0.1704", "0.4768", "0.1704", "0.4768", "0.1818", "0.4492", "0.1818" ], [ "0.294", "0.1716", "0.3208", "0.1716", "0.3208", "0.1799", "0.294", "0.1799" ], [ "0.3745", "0.1716", "0.409", "0.1716", "0.409", "0.1799", "0.3745", "0.1799" ], [ "0.4055", "0.1716", "0.4382", "0.1716", "0.4382", "0.1799", "0.4055", "0.1799" ], [ "0.525", "0.1716", "0.5557", "0.1716", "0.5557", "0.1799", "0.525", "0.1799" ], [ "0.561", "0.1716", "0.5903", "0.1716", "0.5903", "0.1799", "0.561", "0.1799" ], [ "0.11584", "0.1912", "0.882", "0.1912", "0.882", "0.2048", "0.11584", "0.2048" ], [ "0.11584", "0.206", "0.881", "0.206", "0.881", "0.2197", "0.11584", "0.2197" ], [ "0.11584", "0.2197", "0.831", "0.2197", "0.831", "0.2328", "0.11584", "0.2328" ], [ "0.465", "0.2494", "0.5146", "0.2494", "0.5146", "0.2642", "0.465", "0.2642" ], [ "0.6396", "0.2494", "0.6733", "0.2494", "0.6733", "0.2654", "0.6396", "0.2654" ], [ "0.5996", "0.2642", "0.706", "0.2673", "0.706", "0.2832", "0.5996", "0.2798" ], [ "0.336", "0.2505", "0.3745", "0.2505", "0.3745", "0.2632", "0.336", "0.2632" ], [ "0.1663", "0.2666", "0.2167", "0.2666", "0.2167", "0.2808", "0.1663", "0.2808" ], [ "0.2451", "0.2666", "0.378", "0.2666", "0.378", "0.2798", "0.2451", "0.2798" ], [ "0.3687", "0.2666", "0.567", "0.2666", "0.567", "0.2808", "0.3687", "0.2808" ], [ "0.559", "0.2673", "0.601", "0.2673", "0.601", "0.279", "0.559", "0.279" ], [ "0.708", "0.2654", "0.8296", "0.2673", "0.8296", "0.282", "0.708", "0.2798" ], [ "0.3376", "0.282", "0.3687", "0.284", "0.367", "0.2964", "0.336", "0.2944" ], [ "0.4148", "0.282", "0.4492", "0.282", "0.4492", "0.2976", "0.4148", "0.2976" ], [ "0.455", "0.2832", "0.4844", "0.2832", "0.4844", "0.2957", "0.455", "0.2957" ], [ "0.4937", "0.2832", "0.5225", "0.2832", "0.5225", "0.2964", "0.4937", "0.2964" ], [ "0.5693", "0.2832", "0.601", "0.2832", "0.601", "0.2976", "0.5693", "0.2976" ], [ "0.615", "0.2832", "0.6465", "0.2832", "0.6465", "0.2976", "0.615", "0.2976" ], [ "0.669", "0.2832", "0.6978", "0.2832", "0.6978", "0.2964", "0.669", "0.2964" ], [ "0.2527", "0.284", "0.288", "0.284", "0.288", "0.2964", "0.2527", "0.2964" ], [ "0.168", "0.284", "0.2334", "0.284", "0.2334", "0.2957", "0.168", "0.2957" ], [ "0.3005", "0.284", "0.3284", "0.284", "0.3284", "0.2957", "0.3005", "0.2957" ], [ "0.392", "0.2874", "0.3997", "0.2874", "0.3997", "0.2944", "0.392", "0.2944" ], [ "0.168", "0.2998", "0.2351", "0.2998", "0.2351", "0.3113", "0.168", "0.3113" ], [ "0.2544", "0.2998", "0.2854", "0.2998", "0.2854", "0.3113", "0.2544", "0.3113" ], [ "0.3005", "0.2998", "0.33", "0.2998", "0.33", "0.3113", "0.3005", "0.3113" ], [ "0.339", "0.2998", "0.3687", "0.2998", "0.3687", "0.3113", "0.339", "0.3113" ], [ "0.378", "0.2998", "0.4475", "0.2998", "0.4475", "0.3113", "0.378", "0.3113" ], [ "0.455", "0.2998", "0.4844", "0.2998", "0.4844", "0.3113", "0.455", "0.3113" ], [ "0.4937", "0.2998", "0.523", "0.2998", "0.523", "0.3113", "0.4937", "0.3113" ], [ "0.529", "0.2998", "0.5996", "0.2998", "0.5996", "0.3113", "0.529", "0.3113" ], [ "0.615", "0.2998", "0.644", "0.2998", "0.644", "0.3123", "0.615", "0.3123" ], [ "0.669", "0.2998", "0.6987", "0.2998", "0.6987", "0.3113", "0.669", "0.3113" ], [ "0.7095", "0.2998", "0.748", "0.2998", "0.748", "0.3113", "0.7095", "0.3113" ], [ "0.7715", "0.2998", "0.8076", "0.2998", "0.8076", "0.3113", "0.7715", "0.3113" ], [ "0.2544", "0.31", "0.2854", "0.3123", "0.2837", "0.3247", "0.2527", "0.3225" ], [ "0.378", "0.31", "0.4165", "0.3123", "0.4148", "0.3237", "0.3762", "0.3218" ], [ "0.4114", "0.31", "0.4475", "0.3123", "0.4458", "0.3247", "0.4106", "0.3225" ], [ "0.455", "0.3113", "0.4836", "0.3113", "0.4836", "0.3237", "0.455", "0.3237" ], [ "0.4954", "0.3113", "0.523", "0.3113", "0.523", "0.3237", "0.4954", "0.3237" ], [ "0.669", "0.3113", "0.6987", "0.3113", "0.6987", "0.3225", "0.669", "0.3225" ], [ "0.7095", "0.3113", "0.7446", "0.3113", "0.7446", "0.3225", "0.7095", "0.3225" ], [ "0.771", "0.3113", "0.8076", "0.3113", "0.8076", "0.3237", "0.771", "0.3237" ], [ "0.2988", "0.3123", "0.3284", "0.3123", "0.3284", "0.3237", "0.2988", "0.3237" ], [ "0.1696", "0.3123", "0.2183", "0.3123", "0.2183", "0.3237", "0.1696", "0.3237" ], [ "0.339", "0.3123", "0.3687", "0.3123", "0.3687", "0.3237", "0.339", "0.3237" ], [ "0.531", "0.3123", "0.5996", "0.3123", "0.5996", "0.3237", "0.531", "0.3237" ], [ "0.615", "0.3123", "0.6455", "0.3123", "0.6455", "0.3237", "0.615", "0.3237" ], [ "0.669", "0.3206", "0.6987", "0.3225", "0.6978", "0.335", "0.6675", "0.3333" ], [ "0.771", "0.3206", "0.8076", "0.3225", "0.806", "0.335", "0.769", "0.3333" ], [ "0.339", "0.3218", "0.3704", "0.3247", "0.367", "0.335", "0.336", "0.3333" ], [ "0.2988", "0.3218", "0.3284", "0.3247", "0.3267", "0.335", "0.2957", "0.3333" ], [ "0.4114", "0.3218", "0.4475", "0.3237", "0.4458", "0.3362", "0.4106", "0.3342" ], [ "0.7095", "0.3218", "0.7466", "0.3237", "0.7446", "0.3362", "0.708", "0.3342" ], [ "0.455", "0.3225", "0.4844", "0.3225", "0.4844", "0.335", "0.455", "0.335" ], [ "0.378", "0.3225", "0.4165", "0.3225", "0.4165", "0.3342", "0.378", "0.3342" ], [ "0.4954", "0.3225", "0.523", "0.3225", "0.523", "0.3342", "0.4954", "0.3342" ], [ "0.531", "0.3225", "0.5996", "0.3225", "0.5996", "0.3342", "0.531", "0.3342" ], [ "0.614", "0.3225", "0.6455", "0.3225", "0.6455", "0.335", "0.614", "0.335" ], [ "0.2527", "0.3237", "0.2854", "0.3237", "0.2854", "0.335", "0.2527", "0.335" ], [ "0.1696", "0.3237", "0.2418", "0.3237", "0.2418", "0.3342", "0.1696", "0.3342" ], [ "0.168", "0.335", "0.2435", "0.335", "0.2435", "0.3467", "0.168", "0.3467" ], [ "0.2527", "0.335", "0.2854", "0.335", "0.2854", "0.3467", "0.2527", "0.3467" ], [ "0.2988", "0.335", "0.3284", "0.335", "0.3284", "0.3467", "0.2988", "0.3467" ], [ "0.34", "0.335", "0.3687", "0.335", "0.3687", "0.3467", "0.34", "0.3467" ], [ "0.378", "0.335", "0.4458", "0.335", "0.4458", "0.3467", "0.378", "0.3467" ], [ "0.455", "0.335", "0.4836", "0.335", "0.4836", "0.3467", "0.455", "0.3467" ], [ "0.4954", "0.335", "0.523", "0.335", "0.523", "0.3467", "0.4954", "0.3467" ], [ "0.531", "0.335", "0.5977", "0.335", "0.5977", "0.3467", "0.531", "0.3467" ], [ "0.615", "0.335", "0.6455", "0.335", "0.6455", "0.3467", "0.615", "0.3467" ], [ "0.669", "0.335", "0.6987", "0.335", "0.6987", "0.3467", "0.669", "0.3467" ], [ "0.7095", "0.335", "0.7466", "0.335", "0.7466", "0.3467", "0.7095", "0.3467" ], [ "0.7715", "0.335", "0.8076", "0.335", "0.8076", "0.3467", "0.7715", "0.3467" ], [ "0.11584", "0.3557", "0.882", "0.3557", "0.882", "0.3706", "0.11584", "0.3706" ], [ "0.11584", "0.3706", "0.882", "0.3706", "0.882", "0.3843", "0.11584", "0.3843" ], [ "0.11755", "0.3843", "0.636", "0.3843", "0.636", "0.398", "0.11755", "0.398" ], [ "0.1301", "0.4233", "0.178", "0.4233", "0.178", "0.4382", "0.1301", "0.4382" ], [ "0.2494", "0.4258", "0.339", "0.4258", "0.339", "0.437", "0.2494", "0.437" ], [ "0.3435", "0.4258", "0.3804", "0.4258", "0.3804", "0.437", "0.3435", "0.437" ], [ "0.3855", "0.4258", "0.4282", "0.4258", "0.4282", "0.437", "0.3855", "0.437" ], [ "0.4358", "0.4258", "0.471", "0.4258", "0.471", "0.4382", "0.4358", "0.4382" ], [ "0.513", "0.4246", "0.879", "0.4246", "0.879", "0.4395", "0.513", "0.4395" ], [ "0.2805", "0.437", "0.3074", "0.4407", "0.3032", "0.4548", "0.2761", "0.452" ], [ "0.378", "0.4287", "0.3862", "0.4287", "0.3862", "0.4316", "0.378", "0.4316" ], [ "0.1301", "0.4407", "0.1957", "0.4407", "0.1957", "0.452", "0.1301", "0.452" ], [ "0.3477", "0.4407", "0.3787", "0.4407", "0.3787", "0.4531", "0.3477", "0.4531" ], [ "0.3914", "0.4407", "0.4224", "0.4407", "0.4224", "0.4531", "0.3914", "0.4531" ], [ "0.5146", "0.4412", "0.881", "0.4412", "0.881", "0.4548", "0.5146", "0.4548" ], [ "0.1293", "0.456", "0.2377", "0.4573", "0.2377", "0.4685", "0.1293", "0.4673" ], [ "0.2805", "0.4573", "0.3064", "0.4573", "0.3064", "0.4685", "0.2805", "0.4685" ], [ "0.3477", "0.4573", "0.378", "0.4573", "0.378", "0.4685", "0.3477", "0.4685" ], [ "0.3914", "0.4548", "0.4224", "0.4573", "0.4207", "0.4685", "0.3896", "0.466" ], [ "0.4358", "0.4573", "0.471", "0.4573", "0.471", "0.4685", "0.4358", "0.4685" ], [ "0.513", "0.4578", "0.881", "0.4578", "0.881", "0.4714", "0.513", "0.4714" ], [ "0.2822", "0.4656", "0.3074", "0.4685", "0.3032", "0.4822", "0.2778", "0.4785" ], [ "0.3477", "0.4673", "0.378", "0.4673", "0.378", "0.4797", "0.3477", "0.4797" ], [ "0.3914", "0.4685", "0.4224", "0.4685", "0.4224", "0.4797", "0.3914", "0.4797" ], [ "0.4358", "0.4685", "0.471", "0.4685", "0.471", "0.4797", "0.4358", "0.4797" ], [ "0.513", "0.4739", "0.879", "0.4739", "0.879", "0.4868", "0.513", "0.4868" ], [ "0.392", "0.4785", "0.4224", "0.4785", "0.4224", "0.4905", "0.392", "0.4905" ], [ "0.1351", "0.4797", "0.2377", "0.4797", "0.2377", "0.4912", "0.1351", "0.4912" ], [ "0.1335", "0.4685", "0.2377", "0.4685", "0.2377", "0.4785", "0.1335", "0.4785" ], [ "0.2805", "0.4797", "0.3064", "0.4797", "0.3064", "0.4922", "0.2805", "0.4922" ], [ "0.3477", "0.4797", "0.378", "0.4797", "0.378", "0.4912", "0.3477", "0.4912" ], [ "0.4358", "0.4797", "0.471", "0.4797", "0.471", "0.4912", "0.4358", "0.4912" ], [ "0.513", "0.488", "0.882", "0.4893", "0.882", "0.5044", "0.513", "0.503" ], [ "0.11584", "0.5005", "0.4878", "0.5005", "0.4878", "0.514", "0.11584", "0.514" ], [ "0.511", "0.506", "0.881", "0.506", "0.881", "0.5195", "0.511", "0.5195" ], [ "0.11584", "0.5156", "0.4878", "0.5156", "0.4878", "0.5293", "0.11584", "0.5293" ], [ "0.513", "0.5225", "0.882", "0.5225", "0.882", "0.5376", "0.513", "0.5376" ], [ "0.11755", "0.5293", "0.4895", "0.5293", "0.4895", "0.542", "0.11755", "0.542" ], [ "0.11755", "0.5435", "0.4878", "0.5435", "0.4878", "0.557", "0.11755", "0.557" ], [ "0.513", "0.538", "0.882", "0.538", "0.882", "0.552", "0.513", "0.552" ], [ "0.11755", "0.557", "0.4878", "0.557", "0.4878", "0.571", "0.11755", "0.571" ], [ "0.513", "0.5527", "0.8833", "0.554", "0.8833", "0.5684", "0.513", "0.568" ], [ "0.1142", "0.571", "0.2957", "0.571", "0.2957", "0.5845", "0.1142", "0.5845" ], [ "0.5283", "0.5693", "0.882", "0.5693", "0.882", "0.5845", "0.5283", "0.5845" ], [ "0.513", "0.586", "0.881", "0.586", "0.881", "0.5996", "0.513", "0.5996" ], [ "0.513", "0.601", "0.881", "0.6016", "0.881", "0.616", "0.513", "0.615" ], [ "0.11755", "0.619", "0.4844", "0.618", "0.4844", "0.633", "0.11755", "0.6343" ], [ "0.513", "0.618", "0.881", "0.618", "0.881", "0.632", "0.513", "0.632" ], [ "0.1125", "0.6274", "0.4836", "0.6357", "0.4836", "0.6577", "0.1125", "0.6494" ], [ "0.4148", "0.632", "0.4785", "0.6343", "0.4785", "0.6455", "0.4148", "0.6436" ], [ "0.5146", "0.6343", "0.882", "0.6343", "0.882", "0.6484", "0.5146", "0.6484" ], [ "0.1184", "0.653", "0.486", "0.653", "0.486", "0.6675", "0.1184", "0.6675" ], [ "0.513", "0.651", "0.6675", "0.651", "0.6675", "0.6646", "0.513", "0.6646" ], [ "0.1184", "0.6694", "0.4844", "0.6694", "0.4844", "0.683", "0.1184", "0.683" ], [ "0.511", "0.6685", "0.881", "0.6704", "0.881", "0.685", "0.511", "0.683" ], [ "0.12006", "0.685", "0.4844", "0.685", "0.4844", "0.6987", "0.12006", "0.6987" ], [ "0.511", "0.684", "0.881", "0.686", "0.881", "0.7007", "0.511", "0.6987" ], [ "0.11755", "0.7017", "0.4844", "0.7017", "0.4844", "0.7163", "0.11755", "0.7163" ], [ "0.513", "0.7017", "0.881", "0.7017", "0.881", "0.7163", "0.513", "0.7163" ], [ "0.11755", "0.7163", "0.486", "0.7173", "0.486", "0.7324", "0.11755", "0.731" ], [ "0.513", "0.7173", "0.879", "0.7173", "0.879", "0.731", "0.513", "0.731" ], [ "0.1184", "0.733", "0.2284", "0.733", "0.2284", "0.7476", "0.1184", "0.7476" ], [ "0.5146", "0.7354", "0.881", "0.7354", "0.881", "0.749", "0.5146", "0.749" ], [ "0.513", "0.7505", "0.881", "0.7505", "0.881", "0.7656", "0.513", "0.7656" ], [ "0.513", "0.767", "0.881", "0.767", "0.881", "0.781", "0.513", "0.781" ], [ "0.11755", "0.774", "0.1486", "0.774", "0.1486", "0.785", "0.11755", "0.785" ], [ "0.1604", "0.774", "0.382", "0.774", "0.382", "0.7876", "0.1604", "0.7876" ], [ "0.513", "0.783", "0.882", "0.782", "0.882", "0.7964", "0.513", "0.7974" ], [ "0.1184", "0.802", "0.4878", "0.802", "0.4878", "0.8154", "0.1184", "0.8154" ], [ "0.513", "0.8", "0.881", "0.8", "0.881", "0.8135", "0.513", "0.8135" ], [ "0.1184", "0.818", "0.486", "0.818", "0.486", "0.8306", "0.1184", "0.8306" ], [ "0.511", "0.8154", "0.881", "0.814", "0.881", "0.829", "0.511", "0.83" ], [ "0.11755", "0.8345", "0.486", "0.8345", "0.486", "0.8486", "0.11755", "0.8486" ], [ "0.513", "0.832", "0.882", "0.832", "0.882", "0.8457", "0.513", "0.8457" ], [ "0.11755", "0.8496", "0.1755", "0.8496", "0.1755", "0.8643", "0.11755", "0.8643" ], [ "0.5146", "0.847", "0.881", "0.847", "0.881", "0.861", "0.5146", "0.861" ], [ "0.5146", "0.8633", "0.669", "0.8633", "0.669", "0.8774", "0.5146", "0.8774" ], [ "0.1184", "0.8726", "0.4836", "0.8726", "0.4836", "0.8857", "0.1184", "0.8857" ], [ "0.1184", "0.888", "0.4878", "0.888", "0.4878", "0.903", "0.1184", "0.903" ], [ "0.529", "0.8896", "0.8833", "0.8926", "0.8833", "0.91", "0.529", "0.9067" ], [ "0.11755", "0.905", "0.4844", "0.905", "0.4844", "0.9185", "0.11755", "0.9185" ], [ "0.51", "0.9062", "0.6616", "0.9062", "0.6616", "0.9194", "0.51", "0.9194" ] ]

[ "MNLI", "FEVER", "QQP", "#Param DEV HANS Fbias", "DEVSym1 Sym2 Fbias", "DEV pawsurikipawsqp", "Fbias Average Frvias", "109M", "Sparsity", "BERT-base", "84.2", "59.8", "86.2", "58.9", "64.5", "90.9", "48.9", "34.7", "20 %", "87.2M", "84.4", "55.5", "1.182 86.5", "57.0", "64.6", "1.045", "90.7", "47.2", "33.5", "1.037", "1.088", "65.4M", "1.204 86.4", "1.05190.5", "84.0", "40 %", "54.7", "57.2", "64.0", "46.6", "32.4", "1.049", "1.101", "83.4", "90.2", "43.6M", "60 %", "52.8", "1.26686.3", "56.9", "63.3", "1.068", "45.9", "31.8", "1.061", "1.132", "32.7M", "1.065", "52.2", "70 %", "63.3", "45.4", "30.7", "1.127", "56.6", "81.8", "1.249", "85.9", "1.063", "89.5", "Table l: Accuracy comparison (in percent) and relative bias Fbias (the smaller the better) for models with iterative", "magnitude pruning with different levels of sparsity. The last column indicates the average Fbias values over three", "tasks. Pruned models have relatively higher degradation in OOD test set compared to the development set.", "FEVER", "QQP", "pawsuriki pawsgp", "MNLI", "Sparsity", "# Param DEV HANS", "SFbias DEV Syml Sym2Foias", "DEV", "Fbias Average Fbias", "59.8", "86.2", "58.9", "64.5", "90.9", "48.9", "34.7", "109M", "BERT-base", "84.2", "-", "DistilBERT", "66M", "82.3", "51.2", "1.28984.5", "51.9", "60.4", "1.18389.9", "48.1", "34.6", "1.006", "1.159", "66M", "1.309", "84.2", "53.4", "60.7", "31.0", "1.039", "1.162", "83.1", "MiniLM", "51.4", "1.13789.9", "46.8", "32.4", "1.155", "51.6", "80.8", "84.4", "1.056", "52.6", "1.247", "60.4", "1.16390.0", "46.0", "66M", "Truncated-16", "Pretrained-16", "66M", "81.6", "52.2", "1.22985.8", "54.7", "62.6", "1.11590.0", "46.4", "33.9", "1.045", "1.130", "Table 2: Accuracy comparison (in percent) and relative bias Fbias (the smaller the better) of compressed models", "with knowledge distillation. Distilled models have relatively higher degradation in OOD test set compared to the", "development set.Except BERT-base, all othermodels have66M parameters.", "Models", "Attemtion heads", "DEV", "HANS", "Fbias", "压缩模型即使在60%稀疏度2时也保留了分", "144", "", "BERT-base", "84.2", "59.8", "布内集合的准确性。然而，OOD测试集上的", "BERT-116heads-v1", "116", "84.1", "55.5", "1.172", "泛化精度有很大的下降。这表明开发集无法捕", "116", "84.2", "53.7", "1.250", "获修剪模型的通用性。其次，随着稀疏性的增", "55.3", "BERT-116heads-v3", "BERT-116heads-v2", "116", "84.0", "1.176", "加，OOD测试集上的泛化精度大大降低，同", "Table 3: Accuracy comparison (in percent) and relative", "时下降到MNLI等任务的随机猜测。第三，在", "bias Fbias (the smaller the better) of compressed mod-", "高度稀疏的情况下（例如。g。70%），开发和", "els with structured pruning.Pruned models have rela-", "tively higher degradation in OOD test set compared to", "OOD测试集性能都受到显著影响。通常，我", "the development set. All compressed models have been", "们在表E1中观察到所有稀疏水平的Fbias>1", " pruned 28 attention heads.", "请注意，我们限制了70%的最大稀疏性，在", "这之后，即使在开发集（Liangetal.,2021）上", "训练也是不稳定的，性能会显著下降。与以前", "次，我们将相对偏差计算为压缩模型相对于", "的情况一样，与开发集相比，OOD测试集的精", "BERT-base的精度差距之比：Fbias=会mERrb", "compressed", "度有很大的下降（例如。g。，7.6%vs1.9%分别", "这里Fbias>1表示压缩模型比BERT-base更", "为MNLI任务降级）", "有偏差，偏差程度在Fbias的较大值中捕获。由", "知识蒸馏：与修剪类似，与蒸馏模型的分销开", "于FEVER有两个OOD测试集，因此我们使用", "发集相比，我们在OOD测试集中观察到更高", "syml和sym2的整体精度来计算Fbias。同样", "的精度下降。以MNLI上的DistilBERT性能为", "QQP的OOD精度是PAWS-wiki和PAWS-qqp", "例，与OOD测试集中的8.6%下降相比，开发", "上的总体精度", "集中的1.9%精度下降。这也可以在表2中得", "到验证，其中所有Fbias值都大于1，表明所", "有蒸馏模型的鲁棒性都不如BERT基。另一个", "3.5", "Experimental Observations", "有趣的观察是，蒸馏模型，我。e。，DistilBERT", "我们报告了表E1中迭代幅度修剪，表2中", "和MiniLM，与预训练模型相比，Fbias有更高", "知识提炼和表E3中结构化修剪的准确性和相", "的偏差，i。e。，Pretrained-16和Truncated-16，因", "对偏差度量Fbias的性能。我们有以下主要观", "为我们在表2中比较了它们的平均Fbias值。", "察结果", "这表明压缩过程在蒸馏模型的低通用性和鲁棒", "性中起着重要作用。", "迭代幅度修剪：首先，对于轻度和中级稀疏性", "修剪后的模型在发行版开发集上具有可比的，", "2Here, 60 % sparsity indicates that 40 % parameters are", "有时甚至更好的性能。以FEVER为例，其中", "remaining after pruning." ]

{ "arxiv_id": "2110.08419", "language": "zh", "page_idx": 4 }

[ [ "0.1993", "0.1351", "0.774", "0.1351", "0.774", "0.1497", "0.1993", "0.1497" ], [ "0.343", "0.1578", "0.626", "0.1578", "0.626", "0.1724", "0.343", "0.1724" ], [ "0.353", "0.2039", "0.4617", "0.2039", "0.4617", "0.2216", "0.353", "0.2216" ], [ "0.5195", "0.2039", "0.636", "0.202", "0.636", "0.2197", "0.5195", "0.2216" ], [ "0.6978", "0.2039", "0.834", "0.2058", "0.834", "0.2203", "0.6978", "0.2184" ], [ "0.2263", "0.2216", "0.367", "0.2203", "0.367", "0.2393", "0.2263", "0.2405" ], [ "0.1316", "0.2052", "0.2942", "0.202", "0.2942", "0.2197", "0.1316", "0.2229" ], [ "0.4248", "0.2216", "0.581", "0.2203", "0.581", "0.238", "0.4248", "0.2393" ], [ "0.6406", "0.2247", "0.74", "0.2247", "0.74", "0.2361", "0.6406", "0.2361" ], [ "0.2998", "0.2424", "0.407", "0.2424", "0.407", "0.2556", "0.2998", "0.2556" ], [ "0.4534", "0.2412", "0.671", "0.2424", "0.671", "0.2568", "0.4534", "0.2556" ], [ "0.326", "0.26", "0.6445", "0.2588", "0.6445", "0.2734", "0.326", "0.2747" ], [ "0.2361", "0.307", "0.3145", "0.3093", "0.3137", "0.3276", "0.2345", "0.3257" ], [ "0.5034", "0.3113", "0.619", "0.31", "0.619", "0.3245", "0.5034", "0.3257" ], [ "0.524", "0.3276", "0.888", "0.3276", "0.888", "0.3416", "0.524", "0.3416" ], [ "0.0989", "0.3435", "0.468", "0.3435", "0.468", "0.3567", "0.0989", "0.3567" ], [ "0.505", "0.3447", "0.888", "0.3447", "0.888", "0.3547", "0.505", "0.3547" ], [ "0.0809", "0.3567", "0.4673", "0.3567", "0.4673", "0.3706", "0.0809", "0.3706" ], [ "0.505", "0.358", "0.8896", "0.358", "0.8896", "0.3718", "0.505", "0.3718" ], [ "0.0809", "0.3738", "0.4673", "0.3738", "0.4673", "0.387", "0.0809", "0.387" ], [ "0.5034", "0.3738", "0.892", "0.3738", "0.892", "0.387", "0.5034", "0.387" ], [ "0.0809", "0.3882", "0.468", "0.3882", "0.468", "0.4016", "0.0809", "0.4016" ], [ "0.5034", "0.3882", "0.652", "0.3882", "0.652", "0.4016", "0.5034", "0.4016" ], [ "0.0801", "0.404", "0.4673", "0.404", "0.4673", "0.4172", "0.0801", "0.4172" ], [ "0.523", "0.404", "0.8906", "0.404", "0.8906", "0.4172", "0.523", "0.4172" ], [ "0.0809", "0.4185", "0.4673", "0.4185", "0.4673", "0.432", "0.0809", "0.432" ], [ "0.5063", "0.4204", "0.8896", "0.4204", "0.8896", "0.4312", "0.5063", "0.4312" ], [ "0.0809", "0.4343", "0.468", "0.4343", "0.468", "0.4475", "0.0809", "0.4475" ], [ "0.505", "0.4343", "0.8906", "0.4343", "0.8906", "0.4475", "0.505", "0.4475" ], [ "0.0801", "0.449", "0.4673", "0.4475", "0.4673", "0.4622", "0.0801", "0.4634" ], [ "0.5034", "0.4495", "0.8906", "0.4495", "0.8906", "0.4634", "0.5034", "0.4634" ], [ "0.0809", "0.464", "0.468", "0.464", "0.468", "0.4778", "0.0809", "0.4778" ], [ "0.5034", "0.464", "0.634", "0.464", "0.634", "0.4778", "0.5034", "0.4778" ], [ "0.0801", "0.4792", "0.4673", "0.4792", "0.4673", "0.4937", "0.0801", "0.4937" ], [ "0.523", "0.4797", "0.8906", "0.4797", "0.8906", "0.4944", "0.523", "0.4944" ], [ "0.0809", "0.4944", "0.468", "0.4944", "0.468", "0.5083", "0.0809", "0.5083" ], [ "0.505", "0.4956", "0.8906", "0.4956", "0.8906", "0.509", "0.505", "0.509" ], [ "0.0809", "0.509", "0.4673", "0.509", "0.4673", "0.523", "0.0809", "0.523" ], [ "0.505", "0.512", "0.8896", "0.512", "0.8896", "0.523", "0.505", "0.523" ], [ "0.0801", "0.5234", "0.4673", "0.523", "0.4673", "0.537", "0.0801", "0.5386" ], [ "0.505", "0.526", "0.8906", "0.526", "0.8906", "0.539", "0.505", "0.539" ], [ "0.0809", "0.539", "0.4697", "0.539", "0.4697", "0.553", "0.0809", "0.553" ], [ "0.505", "0.5415", "0.8896", "0.5415", "0.8896", "0.5547", "0.505", "0.5547" ], [ "0.0809", "0.5547", "0.468", "0.5547", "0.468", "0.5684", "0.0809", "0.5684" ], [ "0.5034", "0.556", "0.8906", "0.556", "0.8906", "0.5693", "0.5034", "0.5693" ], [ "0.0809", "0.5693", "0.468", "0.5693", "0.468", "0.5835", "0.0809", "0.5835" ], [ "0.5034", "0.571", "0.8896", "0.571", "0.8896", "0.584", "0.5034", "0.584" ], [ "0.0809", "0.5854", "0.468", "0.5854", "0.468", "0.5986", "0.0809", "0.5986" ], [ "0.5034", "0.586", "0.8906", "0.586", "0.8906", "0.5996", "0.5034", "0.5996" ], [ "0.0809", "0.5996", "0.2769", "0.5996", "0.2769", "0.613", "0.0809", "0.613" ], [ "0.5034", "0.601", "0.8906", "0.601", "0.8906", "0.6143", "0.5034", "0.6143" ], [ "0.5034", "0.616", "0.6587", "0.616", "0.6587", "0.63", "0.5034", "0.63" ], [ "0.523", "0.631", "0.8906", "0.631", "0.8906", "0.6445", "0.523", "0.6445" ], [ "0.0801", "0.648", "0.2157", "0.648", "0.2157", "0.662", "0.0801", "0.662" ], [ "0.505", "0.6465", "0.8906", "0.6465", "0.8906", "0.66", "0.505", "0.66" ], [ "0.505", "0.662", "0.8906", "0.662", "0.8906", "0.6753", "0.505", "0.6753" ], [ "0.10046", "0.674", "0.4478", "0.674", "0.4478", "0.6875", "0.10046", "0.6875" ], [ "0.505", "0.6787", "0.8896", "0.6787", "0.8896", "0.6895", "0.505", "0.6895" ], [ "0.0809", "0.688", "0.468", "0.688", "0.468", "0.7026", "0.0809", "0.7026" ], [ "0.5034", "0.6914", "0.8906", "0.6924", "0.8906", "0.707", "0.5034", "0.7056" ], [ "0.0825", "0.704", "0.468", "0.704", "0.468", "0.7173", "0.0825", "0.7173" ], [ "0.505", "0.7085", "0.8906", "0.7085", "0.8906", "0.7217", "0.505", "0.7217" ], [ "0.0825", "0.72", "0.4673", "0.72", "0.4673", "0.731", "0.0825", "0.731" ], [ "0.5034", "0.7227", "0.8906", "0.7217", "0.8906", "0.735", "0.5034", "0.7363" ], [ "0.0809", "0.735", "0.2607", "0.735", "0.2607", "0.7485", "0.0809", "0.7485" ], [ "0.505", "0.738", "0.8906", "0.738", "0.8906", "0.752", "0.505", "0.752" ], [ "0.10297", "0.75", "0.468", "0.75", "0.468", "0.763", "0.10297", "0.763" ], [ "0.5024", "0.7524", "0.8906", "0.752", "0.8906", "0.7666", "0.5024", "0.7676" ], [ "0.0825", "0.765", "0.4697", "0.765", "0.4697", "0.779", "0.0825", "0.779" ], [ "0.5024", "0.768", "0.8906", "0.768", "0.8906", "0.782", "0.5024", "0.782" ], [ "0.0809", "0.78", "0.468", "0.78", "0.468", "0.7935", "0.0809", "0.7935" ], [ "0.5034", "0.784", "0.8906", "0.784", "0.8906", "0.7974", "0.5034", "0.7974" ], [ "0.0809", "0.7954", "0.468", "0.7954", "0.468", "0.8096", "0.0809", "0.8096" ], [ "0.5034", "0.799", "0.8906", "0.799", "0.8906", "0.8125", "0.5034", "0.8125" ], [ "0.0809", "0.811", "0.2932", "0.811", "0.2932", "0.8247", "0.0809", "0.8247" ], [ "0.5034", "0.813", "0.892", "0.813", "0.892", "0.8276", "0.5034", "0.8276" ], [ "0.10046", "0.827", "0.468", "0.827", "0.468", "0.8403", "0.10046", "0.8403" ], [ "0.501", "0.8276", "0.7573", "0.829", "0.7573", "0.8433", "0.501", "0.8423" ], [ "0.0825", "0.8413", "0.4673", "0.8413", "0.4673", "0.855", "0.0825", "0.855" ], [ "0.523", "0.8447", "0.6953", "0.8447", "0.6953", "0.858", "0.523", "0.858" ], [ "0.0809", "0.857", "0.468", "0.857", "0.468", "0.8706", "0.0809", "0.8706" ], [ "0.0809", "0.872", "0.468", "0.872", "0.468", "0.8853", "0.0809", "0.8853" ], [ "0.5156", "0.8706", "0.8906", "0.872", "0.8906", "0.886", "0.5156", "0.8853" ], [ "0.0825", "0.887", "0.468", "0.887", "0.468", "0.901", "0.0825", "0.901" ], [ "0.535", "0.887", "0.892", "0.887", "0.892", "0.901", "0.535", "0.901" ] ]

[ "Bridging Unsupervised and Supervised Depth from Focus", "via All-in-Focus Supervision", "Ren Wang 1", "Yu-Lun Liu 1", "Yu-Hao Huang ", "Yu-Lin Chang 1", "Ning-Hsu Wang 1,2", "Chia-Ping Chen 1", "Kevin Jou1", "1 MediaTek Inc.", "? National Tsing Hua University", "https://github.com/albert100121/AiFDepthNet", "Abstract", "确地配准深度图", "为解决此问题而提出的一些无监督学习方法[9，42", "Depth estimation is a long-lasting yet important", "were。他们通常使用图像重建损失和一致性损失，而", "task in computer vision. Most of the previous works", "不需要地面实况深度数据。他们还可以通过直接使用", "try to estimate depth from input images and assume", "现实世界的立体图像进行训练来减轻领域差距，而无", "images are all-in-focus (AiF), which is less common", "需相应的注册深度图", "in real-world applications. On the other hand, a few", "深度估计的另一个相对未被探索的线索是散焦模糊。", "works take defocus blur into account and consider it as", "焦点深度（或散焦）的任务旨在从焦点堆栈i.e.估计场", "another cue for depth estimation. In this paper, we pro", "景的深度，该图像是由同一相机在不同焦点位置拍摄", "pose a method to estimate not only a depth map but", "的。这允许消费者自动对焦单目相机在没有额外硬件", "an AiF image from a set of images with different focus", "的情况下估计深度", "positions (known as a focal stack). We design a shared", "基于常规优化的离焦深度方法E[37，36,26]估计每", "architecture to exploit the relationship between depth", "个像素的清晰度水平，并且经常遭受textureless对象", "and AiF estimation : As a result , the proposed method", "或光圈问题。深度学习技术「11，23有助于克服这些", "can be trained either supervisedly with ground truth", "问题，但需要监督learning的地面实况深度数据。使", "depth, or unsupervisedly with AiF images as supervi-", "用注册的深度图检索焦点堆栈既困难又耗时，更不用", "sory signals . We show in various experiments that our", "说ToF传感器E[11]等硬件解决方案获得的不完美深", "method outperforms the state-of-the-art methods both", "度数据了。可以在具有合成深度mapsE[23]的合成数", "quantitatively and qualitatively , and also has higher", "据集上合成散焦模糊。However，itisstillquestionable", "efficiency in inference time", "whether the thin lens synthesis model could represent", "real-world optics well.", "在本文中，我们提出了一种新颖的方法来从输人焦", "1.Introduction", "点堆栈中联合估计深度和全聚焦（AiF）图像。我们利用", "这些twotasks之间的关系，设计一个共享的公共网络。", "深度估计是几十年来最基本的计算机视觉问题之", "此外，所提出的网络可以用地面真相深度图进行监督", "许多下游任务，如增强现实（AR）、虚拟现实（VR）和", "训练，也可以只用地面真相AiF图像进行非监督训练。", "自动驾驶，高度依赖这一研究课题。最近，它还enabled", "与高质量的labeled深度相比，获取AiFimages相对于", "越来越多的智能手机摄影应用程序，例如景深调整，背", "easier，因为可以使用更小的光圈alongwith更长的曝", "景替换和拍照后改变焦点", "光时间捕获AiF图像。However，collectingthecorre", "因此，深度感测已成为捕获设备的基本组件。由于需", "sponding focal stack of an AiF image might be difficult", "要专门的硬件，因此诸如飞行时间（ToF）和structured", "because of the focus breathing phenomenon, where a", "light之类的主动深度感测解决方案通常既昂贵又耗电。", "camera's field of view changes as the lens moves. To", "被动技术，如双目或多视角立体声，成本和功率效率更", "address this problem, we use synthetic data without", "高，但在无文本区域容易出错", "such effects during training, and apply a calibration", "基于深度学习的立体匹配methods通过直接从输人", "process on real data during testing.", "图像学习深度估计，以数据驱动的方式解决了这个问", "我们的贡献有三个方面：", "题。然而，它们需要大量高质量的成对训练数据，这些", "数据的获取既耗时又昂贵。当训练数据不完美时，它们", ". When trained supervisedly, our method outper.", "也会受到影响：合成和不切实际的输入图像，或者不准", "forms the state-of-the-art methods in various com-" ]

{ "arxiv_id": "2108.10843", "language": "zh", "page_idx": 0 }

[ [ "0.1111", "0.09216", "0.4248", "0.09155", "0.4248", "0.1061", "0.1111", "0.1067" ], [ "0.5034", "0.09216", "0.8896", "0.09216", "0.8896", "0.1061", "0.5034", "0.1061" ], [ "0.5024", "0.108", "0.8896", "0.108", "0.8896", "0.1218", "0.5024", "0.1218" ], [ "0.0948", "0.11743", "0.468", "0.11743", "0.468", "0.1307", "0.0948", "0.1307" ], [ "0.5034", "0.1237", "0.8896", "0.1237", "0.8896", "0.137", "0.5034", "0.137" ], [ "0.11273", "0.1332", "0.468", "0.1332", "0.468", "0.1465", "0.11273", "0.1465" ], [ "0.505", "0.1383", "0.8906", "0.1383", "0.8906", "0.1515", "0.505", "0.1515" ], [ "0.11273", "0.1465", "0.4656", "0.1465", "0.4656", "0.1604", "0.11273", "0.1604" ], [ "0.505", "0.154", "0.8906", "0.154", "0.8906", "0.1674", "0.505", "0.1674" ], [ "0.11273", "0.1635", "0.3782", "0.1635", "0.3782", "0.1768", "0.11273", "0.1768" ], [ "0.5034", "0.1666", "0.7656", "0.1674", "0.7656", "0.1818", "0.5034", "0.1812" ], [ "0.0948", "0.1875", "0.4673", "0.1875", "0.4673", "0.202", "0.0948", "0.202" ], [ "0.524", "0.1831", "0.892", "0.1831", "0.892", "0.1963", "0.524", "0.1963" ], [ "0.1111", "0.2026", "0.4673", "0.2026", "0.4673", "0.2166", "0.1111", "0.2166" ], [ "0.505", "0.1989", "0.8906", "0.1989", "0.8906", "0.2122", "0.505", "0.2122" ], [ "0.1111", "0.2166", "0.468", "0.2172", "0.468", "0.2317", "0.1111", "0.2311" ], [ "0.5024", "0.2108", "0.8906", "0.2134", "0.8906", "0.2279", "0.5024", "0.226" ], [ "0.1111", "0.2291", "0.1528", "0.2329", "0.1503", "0.25", "0.10864", "0.2456" ], [ "0.505", "0.2291", "0.8896", "0.2291", "0.8896", "0.2424", "0.505", "0.2424" ], [ "0.5034", "0.2437", "0.8906", "0.2437", "0.8906", "0.2568", "0.5034", "0.2568" ], [ "0.0784", "0.2615", "0.2279", "0.2615", "0.2279", "0.276", "0.0784", "0.276" ], [ "0.5034", "0.2588", "0.8906", "0.2588", "0.8906", "0.2727", "0.5034", "0.2727" ], [ "0.505", "0.2747", "0.8906", "0.2747", "0.8906", "0.2886", "0.505", "0.2886" ], [ "0.0801", "0.286", "0.2566", "0.286", "0.2566", "0.2998", "0.0801", "0.2998" ], [ "0.5034", "0.2874", "0.8906", "0.2886", "0.8906", "0.303", "0.5034", "0.3018" ], [ "0.10046", "0.3093", "0.4673", "0.3093", "0.4673", "0.3225", "0.10046", "0.3225" ], [ "0.505", "0.3037", "0.8906", "0.3037", "0.8906", "0.3176", "0.505", "0.3176" ], [ "0.0809", "0.324", "0.4673", "0.324", "0.4673", "0.3372", "0.0809", "0.3372" ], [ "0.505", "0.3208", "0.888", "0.3208", "0.888", "0.3308", "0.505", "0.3308" ], [ "0.0809", "0.339", "0.468", "0.339", "0.468", "0.3528", "0.0809", "0.3528" ], [ "0.505", "0.334", "0.8906", "0.334", "0.8906", "0.348", "0.505", "0.348" ], [ "0.0825", "0.3542", "0.468", "0.3542", "0.468", "0.3674", "0.0825", "0.3674" ], [ "0.505", "0.351", "0.8677", "0.351", "0.8677", "0.361", "0.505", "0.361" ], [ "0.5034", "0.3623", "0.8906", "0.363", "0.8906", "0.3774", "0.5034", "0.377" ], [ "0.0825", "0.3694", "0.468", "0.3694", "0.468", "0.3833", "0.0825", "0.3833" ], [ "0.5034", "0.3774", "0.7646", "0.379", "0.7646", "0.3933", "0.5034", "0.3926" ], [ "0.0809", "0.3838", "0.4673", "0.3838", "0.4673", "0.3977", "0.0809", "0.3977" ], [ "0.0825", "0.3997", "0.4673", "0.3997", "0.4673", "0.4136", "0.0825", "0.4136" ], [ "0.5024", "0.401", "0.7256", "0.4016", "0.7256", "0.4167", "0.5024", "0.4153" ], [ "0.0825", "0.4153", "0.4673", "0.4153", "0.4673", "0.4287", "0.0825", "0.4287" ], [ "0.0809", "0.4287", "0.4673", "0.4287", "0.4673", "0.443", "0.0809", "0.443" ], [ "0.523", "0.4255", "0.892", "0.4255", "0.892", "0.44", "0.523", "0.44" ], [ "0.0825", "0.4443", "0.468", "0.4443", "0.468", "0.459", "0.0825", "0.459" ], [ "0.505", "0.4414", "0.892", "0.4414", "0.892", "0.455", "0.505", "0.455" ], [ "0.5024", "0.4539", "0.8906", "0.4558", "0.8906", "0.4702", "0.5024", "0.4685" ], [ "0.0825", "0.4602", "0.4673", "0.4602", "0.4673", "0.4734", "0.0825", "0.4734" ], [ "0.0809", "0.4749", "0.4673", "0.4749", "0.4673", "0.488", "0.0809", "0.488" ], [ "0.5034", "0.4697", "0.8906", "0.4702", "0.8906", "0.4849", "0.5034", "0.4841" ], [ "0.0809", "0.4905", "0.4673", "0.4905", "0.4673", "0.504", "0.0809", "0.504" ], [ "0.5034", "0.486", "0.888", "0.486", "0.888", "0.5", "0.5034", "0.5" ], [ "0.0809", "0.506", "0.4673", "0.506", "0.4673", "0.5195", "0.0809", "0.5195" ], [ "0.505", "0.502", "0.8906", "0.502", "0.8906", "0.515", "0.505", "0.515" ], [ "0.08417", "0.5215", "0.4673", "0.5215", "0.4673", "0.532", "0.08417", "0.532" ], [ "0.505", "0.5166", "0.888", "0.5166", "0.888", "0.53", "0.505", "0.53" ], [ "0.0809", "0.535", "0.4673", "0.535", "0.4673", "0.549", "0.0809", "0.549" ], [ "0.5063", "0.5327", "0.8896", "0.5327", "0.8896", "0.5435", "0.5063", "0.5435" ], [ "0.0809", "0.5503", "0.4673", "0.5503", "0.4673", "0.5645", "0.0809", "0.5645" ], [ "0.505", "0.547", "0.888", "0.547", "0.888", "0.561", "0.505", "0.561" ], [ "0.0809", "0.565", "0.2385", "0.565", "0.2385", "0.58", "0.0809", "0.58" ], [ "0.5034", "0.56", "0.8906", "0.561", "0.8906", "0.5757", "0.5034", "0.5747" ], [ "0.5034", "0.577", "0.8906", "0.577", "0.8906", "0.5913", "0.5034", "0.5913" ], [ "0.0784", "0.5884", "0.3162", "0.589", "0.3162", "0.604", "0.0784", "0.603" ], [ "0.5034", "0.5923", "0.892", "0.5923", "0.892", "0.606", "0.5034", "0.606" ], [ "0.10046", "0.6113", "0.4673", "0.6113", "0.4673", "0.6255", "0.10046", "0.6255" ], [ "0.5024", "0.606", "0.888", "0.607", "0.888", "0.622", "0.5024", "0.6206" ], [ "0.5024", "0.6206", "0.8906", "0.622", "0.8906", "0.636", "0.5024", "0.6353" ], [ "0.0801", "0.627", "0.468", "0.6274", "0.468", "0.6426", "0.0801", "0.6416" ], [ "0.0809", "0.643", "0.468", "0.643", "0.468", "0.657", "0.0809", "0.657" ], [ "0.5034", "0.6396", "0.625", "0.6396", "0.625", "0.6494", "0.5034", "0.6494" ], [ "0.0801", "0.6577", "0.468", "0.657", "0.468", "0.672", "0.0801", "0.6724" ], [ "0.501", "0.664", "0.601", "0.664", "0.601", "0.6787", "0.501", "0.6787" ], [ "0.0809", "0.674", "0.468", "0.674", "0.468", "0.6875", "0.0809", "0.6875" ], [ "0.0809", "0.688", "0.468", "0.688", "0.468", "0.7026", "0.0809", "0.7026" ], [ "0.523", "0.69", "0.8906", "0.69", "0.8906", "0.704", "0.523", "0.704" ], [ "0.0825", "0.704", "0.468", "0.704", "0.468", "0.7173", "0.0825", "0.7173" ], [ "0.505", "0.704", "0.8906", "0.704", "0.8906", "0.7183", "0.505", "0.7183" ], [ "0.0809", "0.7183", "0.468", "0.7183", "0.468", "0.732", "0.0809", "0.732" ], [ "0.505", "0.7197", "0.8906", "0.7197", "0.8906", "0.733", "0.505", "0.733" ], [ "0.0809", "0.7344", "0.4673", "0.7344", "0.4673", "0.7476", "0.0809", "0.7476" ], [ "0.5034", "0.735", "0.8906", "0.735", "0.8906", "0.7485", "0.5034", "0.7485" ], [ "0.0809", "0.75", "0.468", "0.75", "0.468", "0.763", "0.0809", "0.763" ], [ "0.5024", "0.7485", "0.888", "0.75", "0.888", "0.7646", "0.5024", "0.763" ], [ "0.0809", "0.7646", "0.468", "0.7646", "0.468", "0.778", "0.0809", "0.778" ], [ "0.501", "0.7646", "0.8896", "0.7646", "0.8896", "0.778", "0.501", "0.778" ], [ "0.0809", "0.779", "0.468", "0.779", "0.468", "0.7925", "0.0809", "0.7925" ], [ "0.5034", "0.78", "0.8906", "0.78", "0.8906", "0.7935", "0.5034", "0.7935" ], [ "0.0809", "0.7944", "0.468", "0.7944", "0.468", "0.808", "0.0809", "0.808" ], [ "0.5024", "0.7944", "0.888", "0.7944", "0.888", "0.808", "0.5024", "0.808" ], [ "0.0809", "0.8096", "0.4673", "0.8096", "0.4673", "0.8237", "0.0809", "0.8237" ], [ "0.5034", "0.8125", "0.5884", "0.8125", "0.5884", "0.8237", "0.5034", "0.8237" ], [ "0.0784", "0.8237", "0.4575", "0.8247", "0.4575", "0.839", "0.0784", "0.8384" ], [ "0.5034", "0.8335", "0.615", "0.8335", "0.615", "0.8477", "0.5034", "0.8477" ], [ "0.0801", "0.8486", "0.2051", "0.8486", "0.2051", "0.8623", "0.0801", "0.8623" ], [ "0.523", "0.857", "0.8906", "0.857", "0.8906", "0.8706", "0.523", "0.8706" ], [ "0.0989", "0.8706", "0.468", "0.872", "0.468", "0.886", "0.0989", "0.8853" ], [ "0.505", "0.872", "0.888", "0.872", "0.888", "0.886", "0.505", "0.886" ], [ "0.0809", "0.887", "0.4697", "0.887", "0.4697", "0.901", "0.0809", "0.901" ], [ "0.5", "0.882", "0.8906", "0.882", "0.8906", "0.903", "0.5", "0.903" ] ]

[ "parisons, while our method also runs faster.", "canbegenerated后捕获。利用这一特性，通过适当", "的manipulating对那些光场图像[28]进行捕获后重新", ": To our knowledge, the proposed method is the", "聚焦canalsobeaccomplished。由于光场相机提供来", "first that can learn depth estimation unsupervis-", "自不同poses的多视图information，因此可以帮助", "edlyfrom only AiF images and performsfavorably", "manycomputationalphotographyapplications。随着", "against the state-of-the-art methods.", "视图数量的增加，可以检索更多信息", ": Domain gaps can be mitigated by our method with", "由于采样资源有限，光场图像经常遭受低空间分辨", "test-time optimization on real-world data, espe", "率。程etal．[7]和金etal．[17]proposed将深度", "cially when ground truth depth data are not avail", "学习应用于从光场图像构建超分辨率图像，以覆盖这", "able.", "种情况。Depth estimation could also be achieved by", "leveraging the abundant information from light field", "2. Related Work", "images [20, 12, 13, 8, 38, 21, 45, 29, 31] . Levoy et al.", "[20]proposed利用核极平面图像（EPIs)，其中contain", "2.1. Depth from Focus", "both空间和角度信息，用于深度估计from光场图像。", "深度估计是基本的computervisiontaskthataims", "Heberetal．[12]proposed，以学习4D光场与其相", "touse不同线索，例如图像序列中的颜色，语义，立体", "应的4D深度场之间的端到端映射，并伴有高阶正则", "或the差异，以预测or融合深度图[5,40]。以前的大多", "化细化。Byfollowing[12]，they然后设计了一个编码", "数作品assume都认为输入图像是全对焦的，而在现实", "器-解码器架构，从光场图像13中提取儿何信息。", "些作品E[15,33]专注于提供具有地面实况深度或AIF", "世界的场景中，imagesareusually被认为是在后台或a", "图像的光场数据集，以供进一步使用", "shallowdepth-of-field(DoF）中散焦的。Nevertheless,", "some方法详细阐述了深度估计with散焦图像。基于", "2.4. Realistic Data Synthesis", "常规优化的方法[37，36]proposedtodirectly估计距", "焦点stacks,andavariantapproach[26]tries的深度，", "Deep learning significantly improves the quality of", "以生成索引图，其中每个像素都分配给the焦点位", "computer vision tasks with a large amount of training", "data但是，collecting现实世界的数据通常成本高昂", "置 leading to the maximal sharpness 。 Chen et al. [6]", "found相对模糊与视差的关系，andmakeuseof it增", "且耗时。因此，许多作品insteadtarget将现实数据合", "强鲁棒性of匹配。来自焦点扫描视频的深度[18]目标", "成为provide sufficient data for training。Barron et al", "从具有连续焦点位置的图像估计深度。Recently，deep", "[3]proposed通过分层谊染方案withaz-buffer合成散", "基于学习的方法[11.24可以更精确地对模糊性进行建", "焦图像。Wadhwaetal.[39]结合从a双像素相机计", "模，并获得更好的深度质量。Ontheotherhand，some", "算的personsegmentationmaskwitha深度图生成浅", "works[2,4,35]usedeep learningtechniques to消除单", "景深图像。Guretal．[10]use一个可微点扩散函数", "个images的散焦模糊", "即thePSF卷积层，合成逼真的散焦images，并从散", "焦网络训练一个深度，生成深度maps。Herrmannet", "2.2. Multi-focus Image Fusion", "al. [14] proposed a learning-based approach for autofo-", "Although real-world images usually have defocus", "cus and recently provided a real dataset for training", "However, the dataset was not open to the public at the", "blur,most computer vision applications are supposed", "to cooperate with all-in-focus (AiF) images. Therefore,", "submission time.", "more and more attention is given on shallow-DoF im-", "3. Method", "age deblurring and multi-focus image fusion . A multi", "focus图像融合方法[41]appliesLaplacian对不同尺度", "我们在本节中描述所提出的方法。Sec。3.1给出了我", "图像的融合。詹etal「43]proposed一种引导滤波器，", "们方法的概述。然后，在Sec中说明了用于bothof深度", "有助于多聚焦图像融合过程中的边缘保持。Nejatiet", "估计和AiF图像重建的共享网络。3.2。Theattention", "al．[27]learn通过池化决策的相对锐度measurement", "mechanism that bridges the two tasks is introduced in", "andproduceapixel-levelscoremap的稀疏表示。提", "Sec. 3.3. We then depict the core concept to turn AiF", "出了一种无监督学习方法[32]was，用于融合多曝", "images into supervisory signals for unsupervised depth", "光或多聚焦图像以生成highdymamicrange（HDR）or", "estimation in Sec. 3.4. Finally, Sec. 3.5 shows the loss", "an AiF image。刘 et al.[22] proposed to learn a fo", "functions for training with depth supervision and AiF", "cus map and a segmentation map through deep neural", "supervision.", "networks, and fuse images by integrating these maps", "3.1.Overview", "2.3. Light Field", "对焦深度旨在通过散焦提示从焦点堆栈恢复深度。", "光场相机捕获空间分布的光线。通过重新煊染", "如图所示。E1，给定具有逐渐变化的焦点位置Pe", "through数字化multi-viewimages，可变孔径pictures", "RHxWx1xF的F图像的焦点堆栈SERHxW×3xF" ]

{ "arxiv_id": "2108.10843", "language": "zh", "page_idx": 1 }

[ [ "0.609", "0.113", "0.6543", "0.113", "0.6543", "0.1477", "0.609", "0.1477" ], [ "0.841", "0.1092", "0.847", "0.1092", "0.847", "0.11237", "0.841", "0.11237" ], [ "0.7256", "0.1225", "0.744", "0.1225", "0.744", "0.1364", "0.7256", "0.1364" ], [ "0.5293", "0.1288", "0.5664", "0.1288", "0.5664", "0.1459", "0.5293", "0.1459" ], [ "0.1569", "0.1351", "0.2524", "0.1351", "0.2524", "0.1465", "0.1569", "0.1465" ], [ "0.669", "0.137", "0.7114", "0.13", "0.716", "0.1459", "0.6733", "0.1528" ], [ "0.79", "0.1509", "0.8857", "0.1509", "0.8857", "0.1654", "0.79", "0.1654" ], [ "0.09235", "0.1622", "0.0972", "0.1622", "0.0972", "0.1674", "0.09235", "0.1674" ], [ "0.2378", "0.1749", "0.2974", "0.1749", "0.2974", "0.1895", "0.2378", "0.1895" ], [ "0.3367", "0.1749", "0.3962", "0.1749", "0.3962", "0.1895", "0.3367", "0.1895" ], [ "0.4968", "0.1831", "0.511", "0.1831", "0.511", "0.1932", "0.4968", "0.1932" ], [ "0.2852", "0.1995", "0.3464", "0.202", "0.3447", "0.2172", "0.2834", "0.2153" ], [ "0.536", "0.2203", "0.568", "0.2203", "0.568", "0.2361", "0.536", "0.2361" ], [ "0.1405", "0.2279", "0.2615", "0.2279", "0.2615", "0.2424", "0.1405", "0.2424" ], [ "0.7197", "0.2247", "0.746", "0.2247", "0.746", "0.2374", "0.7197", "0.2374" ], [ "0.669", "0.238", "0.7256", "0.2329", "0.728", "0.2468", "0.6724", "0.252" ], [ "0.806", "0.2551", "0.869", "0.2551", "0.869", "0.2695", "0.806", "0.2695" ], [ "0.0809", "0.2842", "0.8906", "0.2842", "0.8906", "0.2974", "0.0809", "0.2974" ], [ "0.0801", "0.3005", "0.8906", "0.3005", "0.8906", "0.3145", "0.0801", "0.3145" ], [ "0.0801", "0.315", "0.8906", "0.315", "0.8906", "0.3289", "0.0801", "0.3289" ], [ "0.0768", "0.3308", "0.669", "0.33", "0.669", "0.3447", "0.0768", "0.3452" ], [ "0.08905", "0.3726", "0.4673", "0.3726", "0.4673", "0.3865", "0.08905", "0.3865" ], [ "0.5195", "0.3694", "0.892", "0.3706", "0.892", "0.3882", "0.5195", "0.387" ], [ "0.1095", "0.3882", "0.468", "0.3882", "0.468", "0.4016", "0.1095", "0.4016" ], [ "0.501", "0.385", "0.892", "0.385", "0.892", "0.4028", "0.501", "0.4028" ], [ "0.0768", "0.3977", "0.4126", "0.3977", "0.4126", "0.4185", "0.0768", "0.4185" ], [ "0.0825", "0.3894", "0.1168", "0.3894", "0.1168", "0.401", "0.0825", "0.401" ], [ "0.5034", "0.4028", "0.8906", "0.4028", "0.8906", "0.4167", "0.5034", "0.4167" ], [ "0.5063", "0.4192", "0.5923", "0.4192", "0.5923", "0.4312", "0.5063", "0.4312" ], [ "0.0784", "0.4268", "0.2893", "0.428", "0.2893", "0.4426", "0.0784", "0.4414" ], [ "0.58", "0.4363", "0.8145", "0.4382", "0.8145", "0.4602", "0.58", "0.4583" ], [ "0.869", "0.443", "0.8906", "0.443", "0.8906", "0.457", "0.869", "0.457" ], [ "0.10046", "0.4507", "0.468", "0.4507", "0.468", "0.4653", "0.10046", "0.4653" ], [ "0.0825", "0.4673", "0.4673", "0.4673", "0.4673", "0.481", "0.0825", "0.481" ], [ "0.523", "0.4653", "0.8906", "0.4666", "0.8906", "0.481", "0.523", "0.4797" ], [ "0.0809", "0.4817", "0.4673", "0.4817", "0.4673", "0.4956", "0.0809", "0.4956" ], [ "0.5195", "0.4817", "0.8906", "0.4817", "0.8906", "0.4956", "0.5195", "0.4956" ], [ "0.0809", "0.4968", "0.4673", "0.4968", "0.4673", "0.5103", "0.0809", "0.5103" ], [ "0.501", "0.4924", "0.649", "0.4924", "0.649", "0.513", "0.501", "0.513" ], [ "0.09314", "0.512", "0.4673", "0.512", "0.4673", "0.526", "0.09314", "0.526" ], [ "0.638", "0.515", "0.7563", "0.5176", "0.7563", "0.536", "0.638", "0.534" ], [ "0.8667", "0.5195", "0.8936", "0.5195", "0.8936", "0.537", "0.8667", "0.537" ], [ "0.0809", "0.528", "0.468", "0.528", "0.468", "0.5415", "0.0809", "0.5415" ], [ "0.0801", "0.5415", "0.2443", "0.5425", "0.2443", "0.5566", "0.0801", "0.556" ], [ "0.5024", "0.5435", "0.539", "0.5435", "0.539", "0.5596", "0.5024", "0.5596" ], [ "0.1013", "0.558", "0.4673", "0.558", "0.4673", "0.5713", "0.1013", "0.5713" ], [ "0.5825", "0.5713", "0.653", "0.5713", "0.653", "0.5947", "0.5825", "0.5947" ], [ "0.683", "0.5674", "0.7793", "0.5674", "0.7793", "0.582", "0.683", "0.582" ], [ "0.8667", "0.5747", "0.8936", "0.5747", "0.8936", "0.5913", "0.8667", "0.5913" ], [ "0.613", "0.58", "0.6274", "0.58", "0.6274", "0.587", "0.613", "0.587" ], [ "0.665", "0.5854", "0.6807", "0.5854", "0.6807", "0.5923", "0.665", "0.5923" ], [ "0.6587", "0.586", "0.806", "0.5854", "0.806", "0.603", "0.6587", "0.604" ], [ "0.0809", "0.5728", "0.468", "0.5728", "0.468", "0.586", "0.0809", "0.586" ], [ "0.0809", "0.587", "0.468", "0.587", "0.468", "0.6016", "0.0809", "0.6016" ], [ "0.0809", "0.603", "0.468", "0.603", "0.468", "0.616", "0.0809", "0.616" ], [ "0.5215", "0.608", "0.892", "0.6094", "0.892", "0.627", "0.5215", "0.6255" ], [ "0.0825", "0.617", "0.468", "0.617", "0.468", "0.631", "0.0825", "0.631" ], [ "0.5034", "0.625", "0.744", "0.627", "0.744", "0.6416", "0.5034", "0.6396" ], [ "0.0809", "0.634", "0.2443", "0.634", "0.2443", "0.648", "0.0809", "0.648" ], [ "0.589", "0.6465", "0.8037", "0.648", "0.8037", "0.6685", "0.589", "0.6675" ], [ "0.8677", "0.651", "0.8906", "0.651", "0.8906", "0.6655", "0.8677", "0.6655" ], [ "0.0801", "0.657", "0.2852", "0.657", "0.2852", "0.6704", "0.0801", "0.6704" ], [ "0.524", "0.675", "0.892", "0.675", "0.892", "0.688", "0.524", "0.688" ], [ "0.0989", "0.681", "0.468", "0.6753", "0.468", "0.6934", "0.0989", "0.6987" ], [ "0.505", "0.69", "0.8906", "0.69", "0.8906", "0.704", "0.505", "0.704" ], [ "0.09644", "0.6963", "0.468", "0.6963", "0.468", "0.7104", "0.09644", "0.7104" ], [ "0.5024", "0.704", "0.8936", "0.705", "0.8936", "0.7197", "0.5024", "0.7183" ], [ "0.0801", "0.7104", "0.468", "0.711", "0.468", "0.726", "0.0801", "0.7246" ], [ "0.5034", "0.72", "0.892", "0.72", "0.892", "0.7344", "0.5034", "0.7344" ], [ "0.0809", "0.7266", "0.2754", "0.7266", "0.2754", "0.7407", "0.0809", "0.7407" ], [ "0.10046", "0.7407", "0.4673", "0.7407", "0.4673", "0.755", "0.10046", "0.755" ], [ "0.5034", "0.7363", "0.574", "0.7363", "0.574", "0.7476", "0.5034", "0.7476" ], [ "0.0784", "0.755", "0.4004", "0.7524", "0.4004", "0.771", "0.0784", "0.773" ], [ "0.523", "0.75", "0.8906", "0.7505", "0.8906", "0.765", "0.523", "0.7646" ], [ "0.5034", "0.765", "0.886", "0.765", "0.886", "0.779", "0.5034", "0.779" ], [ "0.21", "0.78", "0.3367", "0.782", "0.3367", "0.8003", "0.21", "0.799" ], [ "0.4468", "0.786", "0.468", "0.786", "0.468", "0.8", "0.4468", "0.8" ], [ "0.5034", "0.781", "0.892", "0.781", "0.892", "0.7944", "0.5034", "0.7944" ], [ "0.5034", "0.797", "0.888", "0.797", "0.888", "0.81", "0.5034", "0.81" ], [ "0.0784", "0.811", "0.1168", "0.811", "0.1168", "0.8276", "0.0784", "0.8276" ], [ "0.524", "0.81", "0.8906", "0.81", "0.8906", "0.8247", "0.524", "0.8247" ], [ "0.5034", "0.8257", "0.8906", "0.8257", "0.8906", "0.839", "0.5034", "0.839" ], [ "0.1283", "0.839", "0.1969", "0.839", "0.1969", "0.869", "0.1283", "0.869" ], [ "0.2336", "0.839", "0.3857", "0.839", "0.3857", "0.854", "0.2336", "0.854" ], [ "0.505", "0.8413", "0.8906", "0.8413", "0.8906", "0.855", "0.505", "0.855" ], [ "0.4468", "0.8486", "0.468", "0.8486", "0.468", "0.8623", "0.4468", "0.8623" ], [ "0.2361", "0.8594", "0.4084", "0.8594", "0.4084", "0.8726", "0.2361", "0.8726" ], [ "0.5024", "0.857", "0.8896", "0.857", "0.8896", "0.8706", "0.5024", "0.8706" ], [ "0.2133", "0.8643", "0.2402", "0.8643", "0.2402", "0.872", "0.2133", "0.872" ], [ "0.505", "0.872", "0.892", "0.872", "0.892", "0.886", "0.505", "0.886" ], [ "0.0801", "0.886", "0.3245", "0.886", "0.3245", "0.901", "0.0801", "0.901" ], [ "0.5034", "0.886", "0.523", "0.886", "0.523", "0.901", "0.5034", "0.901" ] ]

[ "广", "", "C", "00)", "Focal stacks", "MAiF", "All-in-focus I", ":", "Encoder", "Decoder", "M", "3Dconv", "so)", "Focus Position P", "&", "Mdepth", "Depth D", "Figure l: An overview of the proposed method. Given a stack of images with varving focus positions, i.e., focal", "stack, our model first produces an intermediate attention map M . The intermediate attention map can be shared", "between depth estimation and all-in-focus (AiF) image reconstruction. With different normalization functions, the", "attention map can be further manipulated to generate either depth or AiF results.", "我们的方法通过共享网络产生多个注意力表示。然", "深度关注Mdepth也可以解释为概率分布，因为", "可以使用这些注意力图生成该场景的深度图DE", "softplus函数确保了非负性并将Mdepth归一化为有效", "RHxWxI和全对焦（AiF）图像IERHxWx3", "后，", "的概率分布。然后可以通过以下方式得出每个像素的", "预期深度值：", "3.2. Network Architecture", "Di,j,1 = (Mdepth P),j1,t.", "(3)", "DDFFE[11]使用2DConvNets来解决清晰度测量", "问题，这是聚焦深度方面的主要挑战之一。另一方面，", "对于AiF图像重建，我们执行类似的过程，除了归", "DefocusNet[24lapplies是一个全局池化层，作为几个", "化函数从softplus更改为softmax。也就是说，AIF", "权重共享2DConvNets之间的通信工具。此架构的另", "注意MAiF获得为：", "个目标是允许任意大小的焦点堆栈。但是，由于2D", "MAiF = o(M),", "(4)", "卷积和简单全局池化的限制，可能无法有效捕获堆栈", "维度上的一些重要信息", "其中", "因此，我们采用Inception3DE[1]作为模型的骨干。", "MAF.t", "exp(Mi,j.1,t)", "(5)", "", "J", "Zh=1 exp(Mi,j,1,n)", "如图所示。E1，我们的模型是由3D卷积组成的编码", "器-解码器网络。通过3D卷积，可以更好地跨帧捕获散", "焦线索，从而促进深度估计和AiF图像重建的任务。此", "然后可以将AIF注意MAiF用于AIF图像重建。", "外，我们的模型还可以处理焦点堆栈与arbitrarysizes", "每个像素的预期AiF图像表示为：", "归因于3D卷积的性质", "Iij,h = E1(MAiF.S)i,j,k,t.", "(6)", "3.3. Attention Mechanism", "我们分别采用thesoftmax和softplus归一化s进", "我们网络的输出是一个中间注意MERHxW×1xF", "行深度和AIF图像估计的原因是为了解决稀疏焦点堆", "中间注意力M的潜在期望是，它应该反映每个焦点", "栈的问题。By sparse, we mean that the stack size is", "位置导致最大锐度的概率。然后它可以有益于深度估", "small, and the focus positions inside the focal stack are", "计和AIF图像重建的both", "对于深度估计，我们建议通过 softplus归一化将中", "not dense", "间注意力M归一化为深度注意力Mdepth？", "对于AiF图像重建，最好沿着堆栈维度选择最清", "晰的像素。在稀疏堆栈中混合多个像素通常无济于事", "Mdepth = s(M),", "(1)", "Therefore, we leverage the softmax function to pursue", "the peaking phenomenon to extract the clearest pixels.", "其中", "对于深度估计，thesoftmax归一化会导致稀疏焦点", "堆栈的严重量化，因为它只是选择最大清晰度的最近", "Mant", "In (1 + exp(Mi,j.1,t))", "焦点位置。另一方面，the softplus归一化导致更平坦", "(2)", "- ln (1 +exp(Mi.j.1.n))", "的分布，因此可以通过在稀疏焦点位置之间进行插值", ">n", "来预测更准确的深度。Fig。2显示了此设计选择的效", "softplus函数是ReLU的平滑版本", "果" ]

{ "arxiv_id": "2108.10843", "language": "zh", "page_idx": 2 }

[ [ "0.5034", "0.09216", "0.8896", "0.09216", "0.8896", "0.1061", "0.5034", "0.1061" ], [ "0.156", "0.1067", "0.2296", "0.1067", "0.2296", "0.1187", "0.156", "0.1187" ], [ "0.33", "0.1067", "0.4192", "0.1067", "0.4192", "0.1187", "0.33", "0.1187" ], [ "0.5024", "0.1061", "0.8906", "0.1067", "0.8906", "0.1218", "0.5024", "0.1206" ], [ "0.505", "0.1237", "0.591", "0.1237", "0.591", "0.1351", "0.505", "0.1351" ], [ "0.2018", "0.13", "0.246", "0.13", "0.246", "0.137", "0.2018", "0.137" ], [ "0.2157", "0.137", "0.2615", "0.137", "0.2615", "0.1445", "0.2157", "0.1445" ], [ "0.2876", "0.132", "0.3137", "0.132", "0.3137", "0.1718", "0.2876", "0.1718" ], [ "0.6685", "0.1459", "0.7236", "0.1509", "0.721", "0.1654", "0.665", "0.1604" ], [ "0.778", "0.1445", "0.835", "0.1497", "0.8325", "0.1666", "0.7754", "0.161" ], [ "0.5435", "0.1559", "0.66", "0.1528", "0.66", "0.1704", "0.5435", "0.1736" ], [ "0.11273", "0.1497", "0.1217", "0.1497", "0.1217", "0.154", "0.11273", "0.154" ], [ "0.7466", "0.159", "0.7656", "0.159", "0.7656", "0.1674", "0.7466", "0.1674" ], [ "0.869", "0.1572", "0.8906", "0.1572", "0.8906", "0.1704", "0.869", "0.1704" ], [ "0.6855", "0.1654", "0.7075", "0.1654", "0.7075", "0.178", "0.6855", "0.178" ], [ "0.794", "0.1654", "0.8164", "0.1654", "0.8164", "0.1787", "0.794", "0.1787" ], [ "0.3447", "0.1749", "0.4272", "0.1749", "0.4272", "0.1863", "0.3447", "0.1863" ], [ "0.501", "0.1901", "0.539", "0.1901", "0.539", "0.2058", "0.501", "0.2058" ], [ "0.10046", "0.202", "0.1299", "0.202", "0.1299", "0.2456", "0.10046", "0.2456" ], [ "0.2876", "0.2026", "0.3137", "0.2026", "0.3137", "0.2456", "0.2876", "0.2456" ], [ "0.7363", "0.2134", "0.7876", "0.2184", "0.785", "0.2343", "0.734", "0.2291" ], [ "0.08496", "0.2052", "0.0989", "0.2052", "0.0989", "0.2349", "0.08496", "0.2349" ], [ "0.6006", "0.2229", "0.6787", "0.2247", "0.6787", "0.2393", "0.6006", "0.2374" ], [ "0.697", "0.2172", "0.7075", "0.2247", "0.701", "0.2299", "0.6914", "0.2235" ], [ "0.697", "0.2349", "0.701", "0.2349", "0.699", "0.2424", "0.6953", "0.2424" ], [ "0.751", "0.2343", "0.774", "0.2343", "0.774", "0.2468", "0.751", "0.2468" ], [ "0.859", "0.2468", "0.892", "0.2468", "0.892", "0.262", "0.859", "0.262" ], [ "0.0866", "0.2361", "0.09644", "0.2361", "0.09644", "0.2437", "0.0866", "0.2437" ], [ "0.1602", "0.25", "0.2385", "0.25", "0.2385", "0.26", "0.1602", "0.26" ], [ "0.3447", "0.25", "0.423", "0.25", "0.423", "0.26", "0.3447", "0.26" ], [ "0.6895", "0.2583", "0.7017", "0.2583", "0.7017", "0.2676", "0.6895", "0.2676" ], [ "0.734", "0.254", "0.7817", "0.2568", "0.7812", "0.2727", "0.732", "0.2695" ], [ "0.5947", "0.2644", "0.6772", "0.2644", "0.6772", "0.279", "0.5947", "0.279" ], [ "0.7036", "0.2664", "0.7197", "0.2664", "0.7197", "0.281", "0.7036", "0.281" ], [ "0.6846", "0.276", "0.6953", "0.2708", "0.705", "0.281", "0.694", "0.2874" ], [ "0.746", "0.2734", "0.7695", "0.2734", "0.7695", "0.286", "0.746", "0.286" ], [ "0.0809", "0.2842", "0.4673", "0.2842", "0.4673", "0.2942", "0.0809", "0.2942" ], [ "0.709", "0.2827", "0.722", "0.2827", "0.722", "0.2917", "0.709", "0.2917" ], [ "0.0801", "0.2974", "0.468", "0.2974", "0.468", "0.3113", "0.0801", "0.3113" ], [ "0.505", "0.3037", "0.8906", "0.3037", "0.8906", "0.3176", "0.505", "0.3176" ], [ "0.0784", "0.3113", "0.468", "0.3125", "0.468", "0.327", "0.0784", "0.3257" ], [ "0.5034", "0.3193", "0.8037", "0.3193", "0.8037", "0.3333", "0.5034", "0.3333" ], [ "0.0784", "0.327", "0.468", "0.3276", "0.468", "0.3423", "0.0784", "0.3416" ], [ "0.0809", "0.3435", "0.468", "0.3435", "0.468", "0.3567", "0.0809", "0.3567" ], [ "0.58", "0.3447", "0.812", "0.3447", "0.812", "0.3623", "0.58", "0.3623" ], [ "0.859", "0.3452", "0.8936", "0.3452", "0.8936", "0.361", "0.859", "0.361" ], [ "0.0809", "0.358", "0.4673", "0.358", "0.4673", "0.3718", "0.0809", "0.3718" ], [ "0.0801", "0.3726", "0.468", "0.3726", "0.468", "0.3865", "0.0801", "0.3865" ], [ "0.5034", "0.3738", "0.739", "0.3738", "0.739", "0.387", "0.5034", "0.387" ], [ "0.0784", "0.3882", "0.4697", "0.3882", "0.4697", "0.4016", "0.0784", "0.4016" ], [ "0.0801", "0.4028", "0.468", "0.4028", "0.468", "0.4172", "0.0801", "0.4172" ], [ "0.5024", "0.401", "0.625", "0.4016", "0.625", "0.4167", "0.5024", "0.4153" ], [ "0.0784", "0.4153", "0.1847", "0.4185", "0.1847", "0.433", "0.0784", "0.43" ], [ "0.524", "0.4268", "0.8896", "0.4268", "0.8896", "0.44", "0.524", "0.44" ], [ "0.5063", "0.4426", "0.8906", "0.4426", "0.8906", "0.4558", "0.5063", "0.4558" ], [ "0.0784", "0.4602", "0.4346", "0.461", "0.4346", "0.476", "0.0784", "0.4749" ], [ "0.5034", "0.457", "0.892", "0.457", "0.892", "0.4702", "0.5034", "0.4702" ], [ "0.5024", "0.4717", "0.8896", "0.4717", "0.8896", "0.486", "0.5024", "0.486" ], [ "0.1013", "0.4849", "0.468", "0.4841", "0.468", "0.4988", "0.1013", "0.5" ], [ "0.501", "0.486", "0.8906", "0.4873", "0.8906", "0.502", "0.501", "0.5005" ], [ "0.0825", "0.5005", "0.468", "0.5005", "0.468", "0.5146", "0.0825", "0.5146" ], [ "0.5034", "0.502", "0.572", "0.502", "0.572", "0.5166", "0.5034", "0.5166" ], [ "0.0825", "0.5176", "0.4673", "0.5176", "0.4673", "0.528", "0.0825", "0.528" ], [ "0.0825", "0.531", "0.468", "0.531", "0.468", "0.5444", "0.0825", "0.5444" ], [ "0.505", "0.528", "0.607", "0.528", "0.607", "0.539", "0.505", "0.539" ], [ "0.0809", "0.5454", "0.468", "0.5454", "0.468", "0.5596", "0.0809", "0.5596" ], [ "0.523", "0.55", "0.888", "0.55", "0.888", "0.563", "0.523", "0.563" ], [ "0.0825", "0.5625", "0.4558", "0.5625", "0.4558", "0.5728", "0.0825", "0.5728" ], [ "0.5034", "0.565", "0.8906", "0.565", "0.8906", "0.579", "0.5034", "0.579" ], [ "0.1013", "0.577", "0.468", "0.577", "0.468", "0.5903", "0.1013", "0.5903" ], [ "0.5034", "0.5806", "0.6846", "0.5806", "0.6846", "0.5947", "0.5034", "0.5947" ], [ "0.0825", "0.5923", "0.468", "0.5923", "0.468", "0.606", "0.0825", "0.606" ], [ "0.0809", "0.607", "0.468", "0.607", "0.468", "0.6206", "0.0809", "0.6206" ], [ "0.5034", "0.605", "0.6914", "0.605", "0.6914", "0.6187", "0.5034", "0.6187" ], [ "0.0809", "0.622", "0.468", "0.6206", "0.468", "0.6353", "0.0809", "0.636" ], [ "0.523", "0.6274", "0.8906", "0.6274", "0.8906", "0.6416", "0.523", "0.6416" ], [ "0.0825", "0.637", "0.4673", "0.637", "0.4673", "0.651", "0.0825", "0.651" ], [ "0.5024", "0.6416", "0.8906", "0.6426", "0.8906", "0.657", "0.5024", "0.656" ], [ "0.0809", "0.651", "0.4673", "0.6514", "0.4673", "0.6665", "0.0809", "0.6655" ], [ "0.5034", "0.6577", "0.892", "0.6577", "0.892", "0.672", "0.5034", "0.672" ], [ "0.0809", "0.6675", "0.468", "0.6675", "0.468", "0.6816", "0.0809", "0.6816" ], [ "0.5024", "0.674", "0.892", "0.674", "0.892", "0.6875", "0.5024", "0.6875" ], [ "0.0809", "0.683", "0.468", "0.683", "0.468", "0.6963", "0.0809", "0.6963" ], [ "0.5034", "0.6875", "0.8423", "0.688", "0.8423", "0.7026", "0.5034", "0.702" ], [ "0.0801", "0.6978", "0.1642", "0.6978", "0.1642", "0.712", "0.0801", "0.712" ], [ "0.505", "0.712", "0.7236", "0.712", "0.7236", "0.7266", "0.505", "0.7266" ], [ "0.0784", "0.7227", "0.2263", "0.7236", "0.2263", "0.738", "0.0784", "0.7373" ], [ "0.524", "0.7363", "0.8906", "0.7363", "0.8906", "0.75", "0.524", "0.75" ], [ "0.10046", "0.7485", "0.4673", "0.7485", "0.4673", "0.762", "0.10046", "0.762" ], [ "0.505", "0.7505", "0.8906", "0.7505", "0.8906", "0.765", "0.505", "0.765" ], [ "0.0809", "0.7646", "0.1887", "0.7646", "0.1887", "0.778", "0.0809", "0.778" ], [ "0.505", "0.765", "0.892", "0.765", "0.892", "0.78", "0.505", "0.78" ], [ "0.5024", "0.781", "0.892", "0.781", "0.892", "0.7944", "0.5024", "0.7944" ], [ "0.1365", "0.7915", "0.4143", "0.7905", "0.4143", "0.808", "0.1365", "0.8096" ], [ "0.4468", "0.7935", "0.4673", "0.7935", "0.4673", "0.806", "0.4468", "0.806" ], [ "0.5024", "0.7944", "0.892", "0.7954", "0.892", "0.81", "0.5024", "0.8096" ], [ "0.505", "0.811", "0.892", "0.811", "0.892", "0.8247", "0.505", "0.8247" ], [ "0.0809", "0.8228", "0.2998", "0.8228", "0.2998", "0.836", "0.0809", "0.836" ], [ "0.505", "0.827", "0.8906", "0.827", "0.8906", "0.8403", "0.505", "0.8403" ], [ "0.10046", "0.8384", "0.468", "0.8384", "0.468", "0.8516", "0.10046", "0.8516" ], [ "0.5034", "0.8413", "0.8896", "0.8413", "0.8896", "0.855", "0.5034", "0.855" ], [ "0.0809", "0.854", "0.2632", "0.854", "0.2632", "0.867", "0.0809", "0.867" ], [ "0.505", "0.857", "0.8906", "0.857", "0.8906", "0.8706", "0.505", "0.8706" ], [ "0.505", "0.872", "0.8896", "0.872", "0.8896", "0.8853", "0.505", "0.8853" ], [ "0.1969", "0.8833", "0.3506", "0.8833", "0.3506", "0.8965", "0.1969", "0.8965" ], [ "0.4468", "0.8833", "0.4697", "0.8833", "0.4697", "0.8965", "0.4468", "0.8965" ], [ "0.5024", "0.8853", "0.892", "0.886", "0.892", "0.901", "0.5024", "0.8994" ] ]

[ "其中Igt表示地面实况AiF图像。此外，我们还鼓励", "Supervised", "Unsupervised", "使用E[9]中的边缘感知加权使深度图局部平滑。平滑", "损失定义为：", "Depth At", "AiFAttention", "canqoia", "aDi,j.i", "ODi.j.1", "Lsmooth = E[W", "0", "W", "(9)", "ar", "Qy", "Stack dimension", "其中", "Aimaaoid", "Kiqqoid", "ali,j.k", "", "W_= exp (", "T", "", "ar", "(10)", "0", "Stackdimensior", "Stackdimensior", "X", "ali,j,k", "Wy = exp (", ">", "?", "Oy", "Figure2:The effectof normalization onattention", "k", "maps. The first row shows that if we adopt the same", "和入是基于地面实况AiF图像的边缘加权的超参数。", "softmax normalization for depth and AiF attention,", "然后，我们的无监督深度估计的总损失为：", "both of them tend to be a fatter distribution in super", "vised learning and face a peaking phenomenon in unsu-", "Lunsupervised = LAiF + QLsmooth,", "(11)", "pervised learning (AiF supervision) The second row", "shows that with separate normalization as described in", "，其中α表示平滑损失的重要性", "Sec. 3.3 , each of depth and AiF attention becomes", "a proper distribution in both supervised and unsuper-", "4.Evaluation", "vised learning.", "在本节中，我们描述了实验中使用的数据集，并报告", "了消融研究以及与距焦点深度的最新方法的比较。在", "3.4.Towards UnsupervisedDepthEstimation", "补充材料中，我们提供了AiF图像重建，andvisual", "comparisons on DDFF 12-Scene E[11] , Middlebury", "与我们的方法类似，DefocusNet[24]还通过中间", "StereoDatasetsE[34]和DefocusNet数据集E[24]的", "散焦图结合了深度估计和AiF图像重建的任务。他们", "其他结果", "的中间散焦图需要监督信号，可以通过从地面实况深", "度计算the混淆圆来得出。此外，他们还建议在中间散", "4.1.Datasets", "焦图之后使用DepthNet来预测输出深度。因此，他们", "在定量实验和视觉比较中总共使用了五个数据集", "的方法只能在地面实况深度数据可用时进行监督训练", "Table1中总结了它们的描述和统计信息。有关更多详", "相反，我们建议使用中间注意力M来桥接两个任", "细信息，请参阅补充材料", "务，这不需要中间监督信号fromgroundtruthdepth", "data。此外，在M之后，are没有可学习的参数。我", "4.2. Evaluation Metrics", "们只使用固定的归一化函数来生成输出深度和AIF图", "在本文中，我们使用以下指标评估定量结果：平均", "像。因此，即使地面真相深度数据不可用，我们仍然可", "绝对误差（MAE），均方误差（MSE），均方根误差", "以训练共享网络仅通过来自地面真相AiF图像的监督", "（RMSE），对数均方根误差（logRMS），相对绝对误差", "信号生成中间注意力M。也就是说，我们的方法可以", "（Abs。rel。），相对平方误差（Sqr。rel。），颠簸（颠簸），", "在有或没有地面实况深度数据的情况下进行监督或无", "使用=1.25的准确性，每栈推理时间（秒。）", "限制地训练", "4.3.ImplementationDetails", "3.5. Training Loss", "我们在PyTorch[30]中实现了我们的方法。我们的", "对于有监督的深度估计，我们的模型使用简单的L1", "模型是使用Adam优化器E[19]（β1=0.9，β2=0.999", "损失进行训练：", "）从头开始训练的，学习率为10-4。对于DDFF12-", "Scene[11]、4D Light Field Dataset [15] ,FlyingTh-", "Lsupervised = Ldepth = E[D - Dgtlli],", "(7)", "ings3D[25]和theDefocusNetdataset[24]，我", "们分别使用10、10、15和5作为输人堆栈大小。在", "，其中Dgt代表地面实况深度", "unsupervisedlearning中，平滑损失a的权重在所有", "对于无监督深度估计，我们的模型也可以通过AiF", "实验中都设置为0.002。我们在训练期间应用随机空间", "监督的L1损失进行训练：", "变换（翻转，裁剪，旋转）和随机颜色抖动（亮度，对", "比度和伽玛】进行数据增强。除DDFF12-Scene【11]", "LAiF=E[I-Igtli],", "(8)", "外，随机裁剪后的训练补丁大小为256×256。Because" ]

{ "arxiv_id": "2108.10843", "language": "zh", "page_idx": 3 }

[ [ "0.1993", "0.1067", "0.2222", "0.1067", "0.2222", "0.1193", "0.1993", "0.1193" ], [ "0.2793", "0.154", "0.722", "0.154", "0.722", "0.1718", "0.2793", "0.1718" ], [ "0.2141", "0.2267", "0.6396", "0.226", "0.6396", "0.2437", "0.2141", "0.2443" ], [ "0.63", "0.2291", "0.785", "0.2291", "0.785", "0.2424", "0.63", "0.2424" ], [ "0.4207", "0.2412", "0.5786", "0.2424", "0.5786", "0.2568", "0.4207", "0.2556" ], [ "0.2998", "0.2568", "0.705", "0.2568", "0.705", "0.2708", "0.2998", "0.2708" ], [ "0.4631", "0.308", "0.5376", "0.308", "0.5376", "0.3225", "0.4631", "0.3225" ], [ "0.2345", "0.3416", "0.761", "0.3416", "0.761", "0.356", "0.2345", "0.356" ], [ "0.2336", "0.3547", "0.7656", "0.3547", "0.7656", "0.3694", "0.2336", "0.3694" ], [ "0.2336", "0.3694", "0.763", "0.3694", "0.763", "0.3833", "0.2336", "0.3833" ], [ "0.232", "0.3833", "0.761", "0.3838", "0.761", "0.3977", "0.232", "0.3965" ], [ "0.232", "0.3958", "0.7656", "0.3965", "0.7656", "0.411", "0.232", "0.4104" ], [ "0.232", "0.411", "0.763", "0.411", "0.763", "0.4248", "0.232", "0.4248" ], [ "0.232", "0.4248", "0.7607", "0.4248", "0.7607", "0.4382", "0.232", "0.4382" ], [ "0.232", "0.4375", "0.7656", "0.4382", "0.7656", "0.4526", "0.232", "0.452" ], [ "0.2345", "0.4526", "0.7656", "0.4526", "0.7656", "0.4666", "0.2345", "0.4666" ], [ "0.2345", "0.4666", "0.761", "0.4666", "0.761", "0.4797", "0.2345", "0.4797" ], [ "0.2336", "0.4797", "0.7646", "0.4797", "0.7646", "0.4944", "0.2336", "0.4944" ], [ "0.2336", "0.4937", "0.763", "0.4937", "0.763", "0.507", "0.2336", "0.507" ], [ "0.232", "0.506", "0.7646", "0.507", "0.7646", "0.5215", "0.232", "0.521" ], [ "0.232", "0.521", "0.7646", "0.5215", "0.7646", "0.536", "0.232", "0.535" ], [ "0.232", "0.534", "0.4102", "0.536", "0.4102", "0.55", "0.232", "0.5474" ], [ "0.1724", "0.5776", "0.3179", "0.58", "0.3179", "0.5947", "0.1724", "0.5923" ], [ "0.1765", "0.6123", "0.8228", "0.6123", "0.8228", "0.6255", "0.1765", "0.6255" ], [ "0.1732", "0.625", "0.8203", "0.625", "0.8203", "0.638", "0.1732", "0.638" ], [ "0.1748", "0.6396", "0.8228", "0.6396", "0.8228", "0.653", "0.1748", "0.653" ], [ "0.1765", "0.653", "0.822", "0.653", "0.822", "0.6665", "0.1765", "0.6665" ], [ "0.1765", "0.6675", "0.822", "0.6675", "0.822", "0.681", "0.1765", "0.681" ], [ "0.1765", "0.681", "0.486", "0.681", "0.486", "0.6943", "0.1765", "0.6943" ], [ "0.1772", "0.702", "0.8228", "0.702", "0.8228", "0.7153", "0.1772", "0.7153" ], [ "0.1748", "0.7153", "0.8203", "0.7153", "0.8203", "0.729", "0.1748", "0.729" ], [ "0.1765", "0.73", "0.8228", "0.73", "0.8228", "0.7437", "0.1765", "0.7437" ], [ "0.1765", "0.7456", "0.5664", "0.7456", "0.5664", "0.756", "0.1765", "0.756" ], [ "0.1765", "0.765", "0.824", "0.765", "0.824", "0.779", "0.1765", "0.779" ], [ "0.1765", "0.779", "0.822", "0.779", "0.822", "0.7925", "0.1765", "0.7925" ], [ "0.1748", "0.7925", "0.824", "0.7925", "0.824", "0.806", "0.1748", "0.806" ], [ "0.1765", "0.806", "0.824", "0.806", "0.824", "0.8193", "0.1765", "0.8193" ], [ "0.1765", "0.821", "0.8228", "0.821", "0.8228", "0.834", "0.1765", "0.834" ], [ "0.1765", "0.834", "0.8228", "0.834", "0.8228", "0.8477", "0.1765", "0.8477" ], [ "0.1765", "0.8477", "0.822", "0.8477", "0.822", "0.8613", "0.1765", "0.8613" ], [ "0.1765", "0.8623", "0.824", "0.8623", "0.824", "0.8755", "0.1765", "0.8755" ], [ "0.1765", "0.8774", "0.2418", "0.8774", "0.2418", "0.888", "0.1765", "0.888" ], [ "0.1732", "0.896", "0.822", "0.8965", "0.822", "0.911", "0.1732", "0.91" ], [ "0.1748", "0.911", "0.8228", "0.91", "0.8228", "0.925", "0.1748", "0.9253" ], [ "0.4927", "0.9497", "0.5063", "0.9497", "0.5063", "0.9604", "0.4927", "0.9604" ] ]

[ "=1", "贝叶斯非参数的流，分布式变分推理", "Trevor CampbellJulian Straub2 John W. Fisher III?", "2JonathanP.Howl", "1LIDS, 2CSAIL, MIT", "[ tdjc@, jstraub@csail., fisher@csail. , jhow@ mit.edu", "Abstract", "This paper presents a methodology for creating streaming, distributed infer", "ence algorithms for Bayesian nonparametric (BNP) models. In the proposed", "framework, processing nodes receive a sequence of data minibatches, com-", "pute a variational posteriorfor each, and make asynchronous streaming up", "dates to a central model. In contrast to previous algorithms, the proposed", "framework is truly streaming, distributed, asynchronous, learning-rate-free.", "and truncation-free. The key challenge in developing the framework, aris", "ing from the fact that BNP models do not impose an inherent ordering", "on their components, is finding the correspondence between minibatch and", "central BNP posterior components before performing each update. To ad", "dress this, the paper develops a combinatorial optimization problem over", "component correspondences, and provides an efficient solution technique.", "The paper concludes with an application of the methodology to the DP", "mixture model, with experimental results demonstrating its practical scal-", "abilityandperformance.", "1Introduction", "贝叶斯非参数（BNP）随机过程是流先验-它们的独特之处在于，它们在概率意义上指定潜", "在模型的复杂性应随着观察到的数据量的增加而增加。此属性在许多数据分析问题中都具", "有常识-例如，人们希望在阅读106文档后在文档语料库中遇到的主题比阅读10文档后遇", "到的主题要多得多-并且在具有无限，持久的数据流的设置中变得至关重要。虽然它们的固", "定参数表亲可用于推断先验已知大小E[1，2]的数据集的模型复杂性，但对于流数据设置中", "模型复杂性增长的概念，此类先验是沉默的", "贝叶斯非参数也自然适合于数据处理的并行化，由于可交换性和条件独立性，它们通常通", "过deFinetti定理表现出来。例如，来自中国餐厅进程[3]的标签将呈现在。i。d。E通过对", "基础Dirichlet过程（DP）随机度量进行条件调整，并从IndianBuffet过程E[4]中分配特", "征。i。d。通过对底层过程（BP）随机测度的条件化", "鉴于这些属性，人们可能会期望有大量的BNP推理算法可以解决与并行化和流相关的挑", "战。但是，以前的工作仅针对参数模型E5，6]一致地解决了这两个设置，并且直到最近才", "针对BNP分别解决了这两个设置。在流式设置中，[7]和[8]使用顺序变分近似为DP混", "合模型开发了流式推理。随机变分推理E[9]和相关方法E[10-13]通常被认为是流算法，但", "是它们的性能取决于学习率的选择以及具有已知，固定大小的先验数据集E[5]。在变分方", "法之外，这是本文的重点，对于BNPE[14，15]，存在精确的并行MCMC方法；使用此类", "方法的权衡是，它们从后验而不是分布本身提供样本，并且有关评估收敛性的结果仍然有", "限。E[16]还开发了用于推理的顺序粒子过滤器，但是这些过滤器存在粒子简并性和指数遗", "忘的问题", "BNP的流式分布式设置带来的主要挑战是组件识别的组合问题。大多数BNP模型都包含", "一组无限的潜在“组件”的概念（例如。g。群集在DP混合模型中），并且不对组件施加固", "1" ]

{ "arxiv_id": "1510.09161", "language": "zh", "page_idx": 0 }

[ [ "0.4167", "0.1237", "0.4468", "0.1237", "0.4468", "0.132", "0.4167", "0.132" ], [ "0.2427", "0.1401", "0.2688", "0.1401", "0.2688", "0.1445", "0.2427", "0.1445" ], [ "0.664", "0.161", "0.7075", "0.161", "0.7075", "0.1718", "0.664", "0.1718" ], [ "0.7524", "0.1604", "0.7974", "0.1604", "0.7974", "0.1718", "0.7524", "0.1718" ], [ "0.694", "0.1895", "0.7256", "0.1895", "0.7256", "0.1976", "0.694", "0.1976" ], [ "0.6284", "0.1642", "0.638", "0.1642", "0.638", "0.1686", "0.6284", "0.1686" ], [ "0.701", "0.2058", "0.74", "0.2058", "0.74", "0.2166", "0.701", "0.2166" ], [ "0.1936", "0.2153", "0.2075", "0.2153", "0.2075", "0.2235", "0.1936", "0.2235" ], [ "0.2402", "0.2134", "0.2932", "0.2134", "0.2932", "0.2235", "0.2402", "0.2235" ], [ "0.1912", "0.2235", "0.25", "0.2235", "0.25", "0.2343", "0.1912", "0.2343" ], [ "0.3594", "0.2235", "0.4805", "0.2235", "0.4805", "0.2343", "0.3594", "0.2343" ], [ "0.4968", "0.2235", "0.789", "0.2235", "0.789", "0.2343", "0.4968", "0.2343" ], [ "0.1748", "0.2361", "0.822", "0.2361", "0.822", "0.25", "0.1748", "0.25" ], [ "0.1765", "0.2664", "0.822", "0.2664", "0.822", "0.2798", "0.1765", "0.2798" ], [ "0.1765", "0.2798", "0.822", "0.2798", "0.822", "0.2935", "0.1765", "0.2935" ], [ "0.1748", "0.2922", "0.824", "0.2922", "0.824", "0.31", "0.1748", "0.31" ], [ "0.1772", "0.3125", "0.822", "0.3125", "0.822", "0.3257", "0.1772", "0.3257" ], [ "0.1765", "0.327", "0.8145", "0.327", "0.8145", "0.3403", "0.1765", "0.3403" ], [ "0.1772", "0.348", "0.8228", "0.348", "0.8228", "0.361", "0.1772", "0.361" ], [ "0.1772", "0.3623", "0.822", "0.3623", "0.822", "0.3757", "0.1772", "0.3757" ], [ "0.1772", "0.3757", "0.822", "0.3757", "0.822", "0.3894", "0.1772", "0.3894" ], [ "0.1765", "0.3894", "0.822", "0.3882", "0.822", "0.4016", "0.1765", "0.4028" ], [ "0.1772", "0.4028", "0.8228", "0.4028", "0.8228", "0.4167", "0.1772", "0.4167" ], [ "0.1765", "0.4153", "0.822", "0.4167", "0.822", "0.43", "0.1765", "0.4287" ], [ "0.1772", "0.4312", "0.8228", "0.4312", "0.8228", "0.4443", "0.1772", "0.4443" ], [ "0.1765", "0.4443", "0.822", "0.4443", "0.822", "0.4583", "0.1765", "0.4583" ], [ "0.1765", "0.4583", "0.6465", "0.4583", "0.6465", "0.4717", "0.1765", "0.4717" ], [ "0.1748", "0.4956", "0.6895", "0.4956", "0.6895", "0.5103", "0.1748", "0.5103" ], [ "0.1765", "0.5264", "0.8228", "0.5264", "0.8228", "0.5405", "0.1765", "0.5405" ], [ "0.1772", "0.5415", "0.822", "0.5415", "0.822", "0.5547", "0.1772", "0.5547" ], [ "0.1765", "0.5547", "0.8228", "0.5547", "0.8228", "0.5684", "0.1765", "0.5684" ], [ "0.1765", "0.5684", "0.8228", "0.5684", "0.8228", "0.582", "0.1765", "0.582" ], [ "0.1772", "0.5835", "0.8228", "0.5835", "0.8228", "0.5967", "0.1772", "0.5967" ], [ "0.1772", "0.5986", "0.5474", "0.5986", "0.5474", "0.6094", "0.1772", "0.6094" ], [ "0.179", "0.619", "0.8203", "0.619", "0.8203", "0.63", "0.179", "0.63" ], [ "0.1765", "0.631", "0.822", "0.631", "0.822", "0.6445", "0.1765", "0.6445" ], [ "0.1912", "0.646", "0.822", "0.646", "0.822", "0.659", "0.1912", "0.659" ], [ "0.1772", "0.66", "0.2336", "0.66", "0.2336", "0.672", "0.1772", "0.672" ], [ "0.1748", "0.6924", "0.3896", "0.6924", "0.3896", "0.7056", "0.1748", "0.7056" ], [ "0.1765", "0.7197", "0.822", "0.7197", "0.822", "0.733", "0.1765", "0.733" ], [ "0.1772", "0.733", "0.822", "0.733", "0.822", "0.7476", "0.1772", "0.7476" ], [ "0.1772", "0.747", "0.822", "0.747", "0.822", "0.761", "0.1772", "0.761" ], [ "0.1765", "0.761", "0.822", "0.761", "0.822", "0.776", "0.1765", "0.776" ], [ "0.1772", "0.7734", "0.822", "0.7744", "0.822", "0.789", "0.1772", "0.7886" ], [ "0.1765", "0.7905", "0.8228", "0.7905", "0.8228", "0.8047", "0.1765", "0.8047" ], [ "0.1765", "0.8066", "0.8228", "0.8066", "0.8228", "0.821", "0.1765", "0.821" ], [ "0.1765", "0.821", "0.824", "0.821", "0.824", "0.8354", "0.1765", "0.8354" ], [ "0.1765", "0.836", "0.2632", "0.836", "0.2632", "0.8467", "0.1765", "0.8467" ], [ "0.1765", "0.855", "0.822", "0.856", "0.822", "0.8706", "0.1765", "0.869" ], [ "0.1765", "0.869", "0.822", "0.869", "0.822", "0.8833", "0.1765", "0.8833" ], [ "0.1772", "0.8853", "0.8203", "0.8853", "0.8203", "0.896", "0.1772", "0.896" ], [ "0.1969", "0.911", "0.7686", "0.911", "0.7686", "0.925", "0.1969", "0.925" ], [ "0.4885", "0.9526", "0.5024", "0.9478", "0.509", "0.9585", "0.4958", "0.9634" ] ]

[ "Postaro", "nfraEPe", "Central Node", "Node", "(3r+2)", "", "Data Streur", "(a)", "Retrieve", "data/prior", "(b)Performinference", "(c)Perform componentID(d)Updatethemodel", "Figure 1: The four main steps of the algorithm that is run asynchronously on each processing node.", "有的排序。因此，为了组合来自多个处理器的有关组件的信息，必须首先找到组件之间的对", "应关系。即使对于中等大小的模型，蛮力搜索也是棘手的-对于两组大小为K1和K2的组", "件，存在（Ki+K2）可能的对应关系。此外，还没有一种方法来评估BNP模型的组件对应关", "系的质量。在MCMC文献中已经研究了这个问题，它被称为“标签切换问题”，但是过去", "的解决方案技术通常是特定于模型的，并且仅限于在非常简单的混合模型上使用E[17，18]", "本文介绍了一种为贝叶斯非参数模型创建流，分布式推理算法的方法。在建议的框架中（如", "图1中的单个节点A所示），处理节点接收一系列数据小批次，为每个数据小批次计算变", "分后验，并使用从组件识别优化获得的映射对中心模型进行异步流更新。这项工作的主要", "贡献如下。首先，我们开发了一个小批量后验分解，它激发了二个适合贝叶斯非参数的无学", "习速率的流媒体分布式框架。然后，我们通过最大化组件匹配的概率来推导组件识别优化", "问题。我们证明了BNP先验正则化了优化中的模型复杂性；二个有趣的副作用是，无论小", "批量变分推理方案是否被截断，所提出的算法都是无截断的。最后，我们基于Jensen不等", "式1为Dirichlet过程先验提供了一个有效的可计算正则化边界。本文最后将该方法应用于", "DP混合模型，实验结果证明了该方法在实践中的可扩展性和性能", "2Streaming, distributed Bayesian nonparametric inference", "所提出的框架是由后验分解驱动的，该后验分解将在2.1节中讨论，涉及一组处理节点，这", "些节点可以异步访问中心变分后验近似（如图1中的单个节点所示。数据以小批量序列的", "形式提供给每个处理节点。当处理节点接收到小批量数据时，它获得中心后验（图1a），并", "将其用作先验，计算小批量变分后验近似（图1b）。当小批推理完成时，节点然后执行小", "批后验和当前中央后验之间的组件识别，考虑其他处理节点可能进行的修改（图1c）。最", "后，它将小批量后验合并到中央变分后验中（图1d）", "在以下各节中，我们使用DP混合物E3作为推理框架技术开发的指导示例。但是，需要", "强调的是，本文中的材料推广到许多其他BNP模型，例如分层DP（HDP）主题模型E[19", "BP潜在特征模型E[20]和Pitman-Yor（PY）混合物E[21]（有关更多详细信息，请参", "见附录）", "2.1Posterior decomposition", "考虑DP混合模型E[3]，具有聚类参数，分配z和观测数据y。对于每个处理节点进行", "的每个异步更新，将数据集分为三个子集y=y。UyiUym进行分析。当处理节点接收到m", "个inibatch数据ym时，它向中央处理节点查询基本后验p(0,zolyo），它将用作小批量推理", "的先验。推理完成后，它再次向中央处理节点查询中间后验p(0,zo，ziyo,yi），该节点说明", "了自小批推理开始以来来自其他处理节点的异步更新。每个子集yr、rE{o,i,m)具有N,", "观测值(yri}=1，并且每个变量zrjEN将yri分配给群集参数z。给定の和z的独立", "性，然后通过先前的中心密度9和之，以及给定潜在参数的数据的条件独立性，贝叶斯规", "则得出θ和的后验分解，给定y，XMATHX。但是，有两个问题阻碍了分解规则（？？）", "的直接应用：", "未知的组件对应关系：由于通常很难准确地找到minibatch后验p(，m，zlyo，m），因此", "需要近似方法。此外，由于（？？）需要密度相乘，因此难以使用基于采样的方法，建议采用", "变分方法。典型的平均场变分技术引人了后验参数的人工排序，从而打破了对称性，这对于", "'Regularization bounds for other popular BNP priors may be found in the supplement", "2" ]

{ "arxiv_id": "1510.09161", "language": "zh", "page_idx": 1 }

[ [ "0.1772", "0.1061", "0.8228", "0.1061", "0.8228", "0.1193", "0.1772", "0.1193" ], [ "0.1765", "0.1206", "0.3545", "0.1206", "0.3545", "0.1307", "0.1765", "0.1307" ], [ "0.1772", "0.1401", "0.822", "0.1401", "0.822", "0.154", "0.1772", "0.154" ], [ "0.1765", "0.154", "0.822", "0.154", "0.822", "0.1674", "0.1765", "0.1674" ], [ "0.1765", "0.1674", "0.824", "0.1674", "0.824", "0.1818", "0.1765", "0.1818" ], [ "0.1765", "0.1818", "0.714", "0.1818", "0.714", "0.1957", "0.1765", "0.1957" ], [ "0.1748", "0.2134", "0.4805", "0.2134", "0.4805", "0.2267", "0.1748", "0.2267" ], [ "0.1765", "0.238", "0.8228", "0.238", "0.8228", "0.252", "0.1765", "0.252" ], [ "0.1765", "0.2524", "0.232", "0.2524", "0.232", "0.2644", "0.1765", "0.2644" ], [ "0.4265", "0.2695", "0.6855", "0.2664", "0.6855", "0.2886", "0.4265", "0.2917" ], [ "0.535", "0.2922", "0.5557", "0.2922", "0.5557", "0.305", "0.535", "0.305" ], [ "0.312", "0.3062", "0.71", "0.2998", "0.71", "0.3208", "0.312", "0.327" ], [ "0.5337", "0.3245", "0.558", "0.3245", "0.558", "0.3333", "0.5337", "0.3333" ], [ "0.585", "0.3384", "0.609", "0.3384", "0.609", "0.3499", "0.585", "0.3499" ], [ "0.2255", "0.3499", "0.774", "0.3435", "0.774", "0.3674", "0.2255", "0.3738" ], [ "0.8013", "0.3528", "0.8228", "0.3528", "0.8228", "0.3655", "0.8013", "0.3655" ], [ "0.585", "0.3674", "0.6094", "0.3674", "0.6094", "0.379", "0.585", "0.379" ], [ "0.2524", "0.3933", "0.749", "0.387", "0.749", "0.4124", "0.2524", "0.4185" ], [ "0.572", "0.3882", "0.584", "0.3801", "0.5947", "0.3882", "0.5825", "0.3965" ], [ "0.572", "0.4124", "0.596", "0.4124", "0.596", "0.4236", "0.572", "0.4236" ], [ "0.1887", "0.4287", "0.8228", "0.4287", "0.8228", "0.4426", "0.1887", "0.4426" ], [ "0.1748", "0.4414", "0.8228", "0.4426", "0.8228", "0.457", "0.1748", "0.4558" ], [ "0.1772", "0.4583", "0.697", "0.4583", "0.697", "0.4685", "0.1772", "0.4685" ], [ "0.1765", "0.4792", "0.822", "0.4792", "0.822", "0.4924", "0.1765", "0.4924" ], [ "0.1765", "0.4924", "0.8105", "0.4924", "0.8105", "0.506", "0.1765", "0.506" ], [ "0.1748", "0.506", "0.8203", "0.506", "0.8203", "0.5195", "0.1748", "0.5195" ], [ "0.1765", "0.518", "0.822", "0.5195", "0.822", "0.534", "0.1765", "0.5327" ], [ "0.1772", "0.534", "0.822", "0.534", "0.822", "0.5474", "0.1772", "0.5474" ], [ "0.1896", "0.5474", "0.822", "0.5474", "0.822", "0.561", "0.1896", "0.561" ], [ "0.1765", "0.56", "0.822", "0.561", "0.822", "0.5757", "0.1765", "0.5747" ], [ "0.1765", "0.5747", "0.822", "0.5747", "0.822", "0.589", "0.1765", "0.589" ], [ "0.1732", "0.587", "0.6357", "0.5884", "0.6357", "0.603", "0.1732", "0.6016" ], [ "0.3545", "0.6094", "0.501", "0.605", "0.501", "0.6235", "0.3545", "0.6274" ], [ "0.2418", "0.6187", "0.343", "0.6187", "0.343", "0.63", "0.2418", "0.63" ], [ "0.501", "0.616", "0.7573", "0.613", "0.7573", "0.632", "0.501", "0.6353" ], [ "0.3774", "0.625", "0.5195", "0.622", "0.5195", "0.6396", "0.3774", "0.6426" ], [ "0.308", "0.632", "0.3145", "0.632", "0.3145", "0.638", "0.308", "0.638" ], [ "0.3628", "0.6353", "0.3774", "0.6353", "0.3774", "0.6426", "0.3628", "0.6426" ], [ "0.5", "0.646", "0.5215", "0.646", "0.5215", "0.6577", "0.5", "0.6577" ], [ "0.8022", "0.651", "0.824", "0.651", "0.824", "0.664", "0.8022", "0.664" ], [ "0.3096", "0.657", "0.7173", "0.653", "0.7173", "0.675", "0.3096", "0.6787" ], [ "0.5", "0.678", "0.5215", "0.678", "0.5215", "0.686", "0.5", "0.686" ], [ "0.3513", "0.6914", "0.5645", "0.69", "0.5645", "0.705", "0.3513", "0.7056" ], [ "0.1772", "0.7134", "0.8228", "0.7134", "0.8228", "0.7266", "0.1772", "0.7266" ], [ "0.1772", "0.73", "0.8203", "0.73", "0.8203", "0.7407", "0.1772", "0.7407" ], [ "0.1896", "0.742", "0.822", "0.742", "0.822", "0.756", "0.1896", "0.756" ], [ "0.1772", "0.756", "0.822", "0.756", "0.822", "0.7695", "0.1772", "0.7695" ], [ "0.1765", "0.7695", "0.4763", "0.7695", "0.4763", "0.7827", "0.1765", "0.7827" ], [ "0.429", "0.7905", "0.4763", "0.7905", "0.4763", "0.8003", "0.429", "0.8003" ], [ "0.3162", "0.806", "0.4143", "0.806", "0.4143", "0.8174", "0.3162", "0.8174" ], [ "0.4387", "0.8003", "0.6855", "0.8003", "0.6855", "0.8213", "0.4387", "0.8213" ], [ "0.411", "0.81", "0.4192", "0.81", "0.4192", "0.8145", "0.411", "0.8145" ], [ "0.384", "0.8174", "0.3906", "0.8228", "0.384", "0.8276", "0.3774", "0.8228" ], [ "0.4387", "0.821", "0.4673", "0.821", "0.4673", "0.832", "0.4387", "0.832" ], [ "0.8013", "0.821", "0.826", "0.821", "0.826", "0.8384", "0.8013", "0.8384" ], [ "0.3865", "0.837", "0.5786", "0.837", "0.5786", "0.851", "0.3865", "0.851" ], [ "0.4272", "0.854", "0.634", "0.854", "0.634", "0.867", "0.4272", "0.867" ], [ "0.1765", "0.875", "0.7036", "0.875", "0.7036", "0.888", "0.1765", "0.888" ], [ "0.574", "0.898", "0.6562", "0.8994", "0.6562", "0.914", "0.574", "0.912" ], [ "0.66", "0.898", "0.7534", "0.898", "0.7534", "0.912", "0.66", "0.912" ], [ "0.2303", "0.904", "0.2769", "0.9077", "0.2737", "0.922", "0.2279", "0.9175" ], [ "0.2876", "0.8984", "0.384", "0.896", "0.384", "0.911", "0.2876", "0.914" ], [ "0.3938", "0.904", "0.564", "0.906", "0.564", "0.9204", "0.3938", "0.9185" ], [ "0.8013", "0.906", "0.8228", "0.906", "0.8228", "0.9194", "0.8013", "0.9194" ], [ "0.6567", "0.91", "0.7817", "0.9077", "0.7817", "0.9253", "0.6567", "0.928" ], [ "0.3896", "0.9077", "0.3962", "0.9077", "0.3962", "0.9136", "0.3896", "0.9136" ], [ "0.5767", "0.911", "0.6045", "0.9136", "0.603", "0.927", "0.574", "0.924" ], [ "0.492", "0.9478", "0.5073", "0.9478", "0.5073", "0.962", "0.492", "0.962" ], [ "0.281", "0.912", "0.3857", "0.909", "0.3857", "0.924", "0.281", "0.927" ] ]

[ "使用密度乘法E6]正确组合后验至关重要。。因此，使用具有平均场变分近似的（？？）需", "要首先解决组件识别问题", "未知的模型大小：虽然以前的后验合并程序需要在小批次后验和中央后验E[5，6]的组件之", "间进行1对1匹配，但贝叶斯非参数后验打破了这一假设。实际上，来自同一非参数混合", "模型的数据集9。，yi和9m可以由相同，不相交或重叠的聚类参数集生成。换句话说，在", "解决组件识别问题并合并小批后验之前，无法确定唯一后验组件的全局数量", "2.2Variational componentidentification", "的合并原因是，假设我们在小批量分解（？？）中具有以下平均场指数族先验和近似变分后", "验密度，", "p(0k) = h(0)en% T(x)-A(no) Vk E N", "K", "p(0, olyo) ~ Qo(0, 0) = Co(z0) II h(Oh)eT(0x)-A(non)", "k=1", "Km", "p(0,zm,20l9m, 0) = Qm(0, zm,z0) = (m(zm)o(20) II h(0k)enmx T(0u)-A(mk)", "(1)", "k=1", "p(0, zi, zol i; o0) = g;(0, i, z0) = C(2)So(20) II h(0k)en T(0x)-A(mn),", "", "k=1", "其中S(），rE[o,i,m)是聚类标签zr的分类分布的乘积，目标是使用THX后验。这", "在图3a中得到了证明，该图显示了忽略组件识别的合成实验（在4节中描述）的结果。随", "着并行线程数量的增加，会产生更多的匹配错误，从而导致模型质量下降", "要解决此问题，请首先注意9m和qi的第一个K。组件没有问题；这些组件可以直接合并", "因为它们各自对应于9。的K。组件。因此，组件识别问题简化为找到小批量后验的最后", "个Km=Km-K。组件与中间后验的最后一个K,=K,-K。组件之间的对应关系。为了简", "化符号（并且不失一般性），固定中间后验9i的组件顺序，并将α：[Km]→[K，+Km】定义", "为从小批量后验组件k到更新的中央后验组件α(k)的1对1映射，其中[K]=[1,..,K]", "第一个K。组件没有排序歧义的事实可以表示为α(k)=kVkE[K。l。请注意，合并后的", "最大组件数为K，+K，因为小批后验中的每个最后K组件可能对应于中间后验中的新", "组件。将三个变分近似（1）代人（？？）后，组件识别优化的目标是找到产生最大更新后验", "归一化常数i的1对1映射*。e。匹配具有相似密度的组件，", "_[p(z, 2mz0)", "g*←argmax", "jo(0,z0)-1qm(0, zm, z0)qi(0, zi,z0)", "Jep(2i|20)p(2m/z0)", "a", "Z", "Km", "(2)", "s.t. qm(0, zm) = C(zm) IIh(g(k)enmxT(o(k)-A(mma)", "k=1", "o(k) = k, Vk e[K,],o 1-to-1", "其中Sm（zm）是这样的分布Pca（zmi=α(k)）=Psm（zmi=k）。取目标的对数并利用平均", "场解耦可以将目标分离为两个项的总和：一个表示组件之间的匹配质量（0上的积分），另", "个表示最终模型大小（z上的总和）。虽然第一个术语以封闭形式提供，但第二个术语通", "常不是。因此，利用对数函数的凹性，Jensen不等式产生了一个下界，可以用来代替棘手的", "原始目标，从而导致最终的组件识别优化：", "K+K\"", "*←argmax", "Z A (n) + E [1og p(zi m 0)]", "", "6", "k=1", "(3)", "s.t. n =Nik +nmh -nok", "(k) = k Vk E [K],o 1-to-1.", "在补充中可以找到更详细的优化推导。E表示分布S。（zo)S（z)sm（zm）和", "nmg-1(k)", "kEα([Kml)", "Trk=", "Nrkk≤K,", "Vr e(o,i, m],nmk =", "(4)", "k(Kmi)，", "", "no", "3", "nk>," ]

{ "arxiv_id": "1510.09161", "language": "zh", "page_idx": 2 }

[ [ "0.179", "0.1061", "0.822", "0.1061", "0.822", "0.1193", "0.179", "0.1193" ], [ "0.1772", "0.1193", "0.822", "0.1193", "0.822", "0.1338", "0.1772", "0.1338" ], [ "0.1765", "0.1332", "0.7134", "0.1332", "0.7134", "0.1433", "0.1765", "0.1433" ], [ "0.1765", "0.1528", "0.8228", "0.154", "0.8228", "0.1686", "0.1765", "0.1674" ], [ "0.1772", "0.1692", "0.819", "0.1692", "0.819", "0.1799", "0.1772", "0.1799" ], [ "0.1772", "0.1843", "0.8203", "0.1843", "0.8203", "0.1945", "0.1772", "0.1945" ], [ "0.1748", "0.1945", "0.822", "0.1957", "0.822", "0.2102", "0.1748", "0.209" ], [ "0.1765", "0.2102", "0.8203", "0.2102", "0.8203", "0.2235", "0.1765", "0.2235" ], [ "0.179", "0.226", "0.822", "0.226", "0.822", "0.2361", "0.179", "0.2361" ], [ "0.1765", "0.238", "0.642", "0.238", "0.642", "0.252", "0.1765", "0.252" ], [ "0.1748", "0.2715", "0.4355", "0.2715", "0.4355", "0.2854", "0.1748", "0.2854" ], [ "0.1748", "0.2986", "0.824", "0.2974", "0.824", "0.3125", "0.1748", "0.3132" ], [ "0.1772", "0.3125", "0.822", "0.3125", "0.822", "0.3257", "0.1772", "0.3257" ], [ "0.1772", "0.3257", "0.822", "0.3257", "0.822", "0.339", "0.1772", "0.339" ], [ "0.1772", "0.339", "0.822", "0.339", "0.822", "0.3542", "0.1772", "0.3542" ], [ "0.1765", "0.3528", "0.407", "0.3542", "0.407", "0.3687", "0.1765", "0.3674" ], [ "0.259", "0.3865", "0.3945", "0.3865", "0.3945", "0.3997", "0.259", "0.3997" ], [ "0.3906", "0.387", "0.822", "0.387", "0.822", "0.4016", "0.3906", "0.4016" ], [ "0.2957", "0.4016", "0.343", "0.4016", "0.343", "0.4092", "0.2957", "0.4092" ], [ "0.1929", "0.3882", "0.2607", "0.3882", "0.2607", "0.3997", "0.1929", "0.3997" ], [ "0.1772", "0.4224", "0.8228", "0.4224", "0.8228", "0.4363", "0.1772", "0.4363" ], [ "0.1748", "0.4363", "0.513", "0.4375", "0.513", "0.452", "0.1748", "0.4507" ], [ "0.1724", "0.473", "0.638", "0.4734", "0.638", "0.488", "0.1724", "0.4873" ], [ "0.1748", "0.506", "0.822", "0.507", "0.822", "0.521", "0.1748", "0.5195" ], [ "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", "0.5474", "0.8228", "0.5474", "0.8228", "0.561", "0.1772", "0.561" ], [ "0.1772", "0.5625", "0.4043", "0.5625", "0.4043", "0.5757", "0.1772", "0.5757" ], [ "0.1732", "0.595", "0.4143", "0.595", "0.4143", "0.6094", "0.1732", "0.6094" ], [ "0.1748", "0.6226", "0.824", "0.622", "0.824", "0.636", "0.1748", "0.637" ], [ "0.1765", "0.637", "0.2009", "0.637", "0.2009", "0.6494", "0.1765", "0.6494" ], [ "0.3694", "0.661", "0.6274", "0.66", "0.6274", "0.6816", "0.3694", "0.683" ], [ "0.8022", "0.6665", "0.8228", "0.6665", "0.8228", "0.681", "0.8022", "0.681" ], [ "0.528", "0.6816", "0.5557", "0.6816", "0.5557", "0.6934", "0.528", "0.6934" ], [ "0.1953", "0.705", "0.8228", "0.705", "0.8228", "0.7197", "0.1953", "0.7197" ], [ "0.1765", "0.7183", "0.822", "0.7183", "0.822", "0.733", "0.1765", "0.733" ], [ "0.1772", "0.733", "0.822", "0.733", "0.822", "0.7476", "0.1772", "0.7476" ], [ "0.1765", "0.7476", "0.822", "0.7476", "0.822", "0.762", "0.1765", "0.762" ], [ "0.1772", "0.7646", "0.8228", "0.7646", "0.8228", "0.778", "0.1772", "0.778" ], [ "0.1772", "0.78", "0.822", "0.78", "0.822", "0.7935", "0.1772", "0.7935" ], [ "0.1765", "0.7944", "0.822", "0.7944", "0.822", "0.808", "0.1765", "0.808" ], [ "0.1772", "0.8096", "0.822", "0.8096", "0.822", "0.8228", "0.1772", "0.8228" ], [ "0.1765", "0.8228", "0.822", "0.8228", "0.822", "0.836", "0.1765", "0.836" ], [ "0.1765", "0.8384", "0.822", "0.8384", "0.822", "0.8516", "0.1765", "0.8516" ], [ "0.1772", "0.853", "0.8228", "0.853", "0.8228", "0.866", "0.1772", "0.866" ], [ "0.1765", "0.866", "0.524", "0.866", "0.524", "0.88", "0.1765", "0.88" ], [ "0.1896", "0.896", "0.8203", "0.898", "0.8203", "0.9155", "0.1896", "0.9136" ], [ "0.4927", "0.955", "0.5024", "0.9497", "0.5073", "0.955", "0.4983", "0.9604" ] ]

[ "下的期望，其中α（[Kml）表示映射α的范围。（4）中的定义确保在后验rE(i,m,o不包", "含特定组件k时使用先前的no。优化（3）的直觉是，它将查找具有高相似性的组件对应", "关系（通过对数分割函数）与正则化项2结合在最终更新的后验模型大小上", "尽管其动机来自Dirichlet过程混合物，但组件识别优化（3）并不特定于此模型。实际上：", "推导不依赖于Dirichlet过程混合物的任何特定属性；优化适用于具有一组“组件”e和一", "组组合“指标”之的任何贝叶斯非参数模型。例如，优化适用于具有主题词分布和局部到", "全局主题对应的分层Dirichlet过程主题模型E[10]，以及具有特征θ和二进制分配向量", "z的beta过程潜在特征模型E[4]。组件识别优化（3）中目标的形式反映了这种普遍性", "为广将所提出的流式分布式方法应用于特定模型，只需要一个黑盒变分推理算法来计算形", "式为（1）的后验，以及一种计算或限制目标中期望的方法（3）。", "2.3Updating the central posterior", "要更新中央后验，节点首先将其锁定并通过（3）求解α*。锁定可防止其他节点求解（3）", "或修改中央后验，但不会阻止其他节点读取中央后验，获取小批量或进行推理；4节中的合", "成实验表明，这在实践中不会产生明显的时间损失。然后，处理节点将α*及其小批量变分", "后验传递到中央处理节点，在中央处理节点中，使用乘积分解（？？）在（？？）中找到更新", "的中央变分后验q，并带有参数", " Kimax g*(k) ", "(,S(z) = (i(zi)So(zo)Sm (zm),mk = nik +nmk-nok. (5)", "kEKm", "K=max", "最后，该节点解锁中央后验，接收新小批量的下一个处理节点将使用来自中央节点的上述", "K、C(z）和nk作为它们的K。、S（zo）和nok", "3Application to the Dirichletprocess mixture model", "（3）目标中的期望通常难以以封闭形式计算；因此，可以在其位置使用合适的下限。本节", "介绍了Dirichlet过程的这样一个界，并讨论了使用开发的界将所提出的推理框架应用于", "Dirichlet过程混合模型。至关重要的是，下界分解使得优化（3）成为最大权重的二分匹配", "问题。匈牙利算法可以在多项式时间[22]中解决此类问题，从而在建议的流式分布式框架", "中导致易于处理的组件识别步骤", "3.1 Regularization lower bound", "对于浓度参数为α>0的Dirichlet过程，p(，zm，%）是可交换分区概率函数（EPPF)E", "[23]", "p(zi, 2m,z0) α QlkKI-1 II (nk - 1)I,", "(6)", "kEK", "其中nk是分配给簇XNETHX。考虑到变分分布Sr(zr），rE{i,m,o)和", "MATHX_MATHX_1上的XMTHX_MATHX_1分别是pz，zm，z）的XMTHX_7的", "中央分类分布的乘积，C的p(z，zm，%）和MATHX_5的5可以分别。图2演示了在具有", "N=100数据点的合成实验中，对于各种DP浓度参数值αE「10-3,1031，下界的行为。", "真正则化logEc[p()]是通过104样本的样本近似来计算的。在图2a中，簇K的数量是变", "化的，对称分类标签权重设置为。该图显示了两个重要现象。首先，界限随着K→0的", "增加而增加；换句话说，它优先考虑更少的，更大的簇，这是典型的BNP“丰富的变得更", "丰富”的属性。其次，约束作为K→N的行为取决于浓度参数α-随着α的增加，更多的", "簇是首选。在图2b中，簇数K固定为10，分类标签权重从具有参数E10-31031的对", "称Dirichlet分布中采样。该图表明，在高标记不确定性的情况下，界限不会显着降低，并", "且对于低标记不确定性几乎是精确的。总体而言，图2a证明了所提出的下界表现出与真实", "正则化相似的行为，支持其在优化中的使用（3）", "2This is equivalent to the KL-divergence regularization -KL o(zo)C;(zi)S%(zm) p(zi, zm, zo)", "X" ]

{ "arxiv_id": "1510.09161", "language": "zh", "page_idx": 3 }

[ [ "0.429", "0.1237", "0.46", "0.1237", "0.46", "0.127", "0.429", "0.127" ], [ "0.2632", "0.2184", "0.2712", "0.2184", "0.2712", "0.2229", "0.2632", "0.2229" ], [ "0.5215", "0.2166", "0.544", "0.2166", "0.544", "0.2279", "0.5215", "0.2279" ], [ "0.3635", "0.2229", "0.4019", "0.2197", "0.4028", "0.2311", "0.3652", "0.2343" ], [ "0.6436", "0.2229", "0.6685", "0.2229", "0.6685", "0.2299", "0.6436", "0.2299" ], [ "0.71", "0.2197", "0.7417", "0.2197", "0.7417", "0.2279", "0.71", "0.2279" ], [ "0.3513", "0.2299", "0.3628", "0.2299", "0.3628", "0.2343", "0.3513", "0.2343" ], [ "0.5986", "0.2247", "0.6313", "0.2267", "0.63", "0.2374", "0.597", "0.2349" ], [ "0.1748", "0.2361", "0.8228", "0.2374", "0.8228", "0.252", "0.1748", "0.2507" ], [ "0.1765", "0.25", "0.824", "0.25", "0.824", "0.2632", "0.1765", "0.2632" ], [ "0.1748", "0.262", "0.2532", "0.2644", "0.2532", "0.276", "0.1748", "0.2734" ], [ "0.1748", "0.2917", "0.576", "0.2917", "0.576", "0.305", "0.1748", "0.305" ], [ "0.1765", "0.3176", "0.8228", "0.3176", "0.8228", "0.3308", "0.1765", "0.3308" ], [ "0.1732", "0.33", "0.826", "0.33", "0.826", "0.351", "0.1732", "0.351" ], [ "0.1748", "0.3467", "0.5156", "0.348", "0.5156", "0.3694", "0.1748", "0.3687" ], [ "0.509", "0.3591", "0.5156", "0.3591", "0.5156", "0.3623", "0.509", "0.3623" ], [ "0.5186", "0.3542", "0.822", "0.3542", "0.822", "0.3674", "0.5186", "0.3674" ], [ "0.1765", "0.3674", "0.806", "0.3674", "0.806", "0.3818", "0.1765", "0.3818" ], [ "0.179", "0.3894", "0.824", "0.3894", "0.824", "0.4104", "0.179", "0.4104" ], [ "0.1772", "0.4185", "0.822", "0.4185", "0.822", "0.432", "0.1772", "0.432" ], [ "0.1765", "0.432", "0.822", "0.432", "0.822", "0.4458", "0.1765", "0.4458" ], [ "0.1772", "0.4458", "0.822", "0.4458", "0.822", "0.459", "0.1772", "0.459" ], [ "0.1765", "0.459", "0.8228", "0.459", "0.8228", "0.473", "0.1765", "0.473" ], [ "0.1772", "0.4734", "0.8096", "0.4734", "0.8096", "0.4873", "0.1772", "0.4873" ], [ "0.1772", "0.4873", "0.8228", "0.4873", "0.8228", "0.5005", "0.1772", "0.5005" ], [ "0.1765", "0.5005", "0.8228", "0.5005", "0.8228", "0.5146", "0.1765", "0.5146" ], [ "0.1765", "0.515", "0.5513", "0.515", "0.5513", "0.529", "0.1765", "0.529" ], [ "0.38", "0.5386", "0.615", "0.5386", "0.615", "0.553", "0.38", "0.553" ], [ "0.8022", "0.5386", "0.824", "0.5386", "0.824", "0.553", "0.8022", "0.553" ], [ "0.4714", "0.5547", "0.4763", "0.5547", "0.4763", "0.561", "0.4714", "0.561" ], [ "0.1724", "0.58", "0.3179", "0.5835", "0.3179", "0.5986", "0.1724", "0.595" ], [ "0.1765", "0.6123", "0.822", "0.6123", "0.822", "0.6255", "0.1765", "0.6255" ], [ "0.1765", "0.6255", "0.822", "0.6255", "0.822", "0.6396", "0.1765", "0.6396" ], [ "0.1748", "0.6396", "0.8228", "0.6396", "0.8228", "0.654", "0.1748", "0.654" ], [ "0.1912", "0.653", "0.822", "0.653", "0.822", "0.6665", "0.1912", "0.6665" ], [ "0.1748", "0.6675", "0.8228", "0.6665", "0.8228", "0.681", "0.1748", "0.6816" ], [ "0.1765", "0.681", "0.8228", "0.681", "0.8228", "0.696", "0.1765", "0.696" ], [ "0.1765", "0.6943", "0.3782", "0.696", "0.3782", "0.709", "0.1765", "0.7085" ], [ "0.1772", "0.7173", "0.8228", "0.7173", "0.8228", "0.731", "0.1772", "0.731" ], [ "0.1732", "0.728", "0.4495", "0.73", "0.4495", "0.7476", "0.1732", "0.7456" ], [ "0.437", "0.7266", "0.824", "0.73", "0.824", "0.7476", "0.437", "0.744" ], [ "0.1765", "0.744", "0.8228", "0.744", "0.8228", "0.7583", "0.1765", "0.7583" ], [ "0.1765", "0.759", "0.822", "0.759", "0.822", "0.773", "0.1765", "0.773" ], [ "0.1765", "0.771", "0.8228", "0.7715", "0.8228", "0.786", "0.1765", "0.785" ], [ "0.1765", "0.785", "0.822", "0.785", "0.822", "0.8", "0.1765", "0.8" ], [ "0.1772", "0.8", "0.8228", "0.8", "0.8228", "0.813", "0.1772", "0.813" ], [ "0.1772", "0.813", "0.8228", "0.813", "0.8228", "0.827", "0.1772", "0.827" ], [ "0.1772", "0.827", "0.8228", "0.827", "0.8228", "0.8413", "0.1772", "0.8413" ], [ "0.1772", "0.8413", "0.8228", "0.8413", "0.8228", "0.855", "0.1772", "0.855" ], [ "0.1765", "0.855", "0.822", "0.855", "0.822", "0.8687", "0.1765", "0.8687" ], [ "0.1765", "0.869", "0.822", "0.869", "0.822", "0.8833", "0.1765", "0.8833" ], [ "0.1748", "0.8833", "0.822", "0.8833", "0.822", "0.8965", "0.1748", "0.8965" ], [ "0.1772", "0.898", "0.8228", "0.898", "0.8228", "0.911", "0.1772", "0.911" ], [ "0.1765", "0.911", "0.339", "0.911", "0.339", "0.925", "0.1765", "0.925" ], [ "0.492", "0.9478", "0.5073", "0.9478", "0.5073", "0.962", "0.492", "0.962" ] ]

[ "LCwerBou", "500", "G.cou.", "P of Clusters", "tainty.a", "un1000", "P", "Chust9165", "Figure 2: The Dirichlet process regularization and lower bound, with (2a ) fully uncertain labelling", "and varying number of clusters, and ( 2b ) the number of clusters fixed with varying labelling", "uncertainty.", "3.2Solving the component identification optimization", "鉴于目标中的正则化（?？）和组件匹配分数（3）都分解为每个kE[K，+K的", "项之和，因此可以使用匹配分数矩阵ReR(Ki+Km)(Ki+Km）和选择器变量Xe", "[0,1)(Ki+K)x(Ki+Km）重写目标。设置Xk", "..", "=1表示小批量后验中的组件k与中间", "后验中的组件j匹配（i。e。α(k）=j），提供使用（4）和（??）定义的分数Rkj作为", "Rkj =A(nij+nmk-ioj)+(1-eu+mk+5o)loga+logT(max[2,tij + tmk+to;l)(7)", "优化（3）可以根据X和R的属性重写为<NEXX_THLINE-1。约束Xkk=1VkE[K。", "将X的上部K。×K。块固定为I（由于第一个K。组件直接匹配的事实），将非对角线块", "固定为0。将X，R表示为X，R的右下（K+Km）×（K+Km）块，剩余的优化问", "题是具有成本矩阵-R'的X'上的线性分配问题，可以使用匈牙利4算法MAXXO。请注", "意，如果Km=K。或K，=K。，这意味着不需要解决匹配问题-直接匹配小批后验的第", "个K。组件，并将最后一个K㎡设置为新组件。在该框架的实际实现中，随着观察到更多", "的数据，通常会以递减的速度发现新的群集，因此解决的匹配问题的数量也逐渐减少。通过", "查找X*的非零元素，找到最终的最佳组件映射α*", "o*(k) ←argmaxXk, Vk E[Km]", "(8)", "", "4Experiments", "在本节中，建议的推理框架是在具有正态-逆-Wishart（NIW）先验的DP高斯混合上进行评", "估的。我们将与标准变分推理耦合的流，分布式过程10-1（SDA-DP）和10-1（XMATH", "XsubMAP1）与五种最先进的推理算法进行了比较：记忆化在线变分推理（moVBE[13]", "随机在线变分推理（SVI)E[9，学习率X。先验是手动设置的，所有方法都是随机初", "始化的。使用多次通过数据的方法（e。g。moVB，SVI）被允许这样做。允许moVB进", "行出生/死亡移动，而SVI/Batch具有固定的截断。所有实验均在具有24个CPU内核和", "12GiBRAM的计算机上进行", "合成：此数据集由100,000个二维向量组成，这些向量是从具有100个簇的高斯混合模型和", "NIW(μo,ko, o,vo)先验 μo =0,ko", "=10-3，亚。=1和Vo=4生成的。算法给出了真实", "的NIW先验、DP浓度α=5和大小为50的小批。SDA-DP小批量推理被截断为K=50", "组件，所有其他算法被截断为K=200组件。图3显示了30次试验的实验结果，这些结果", "说明了SDA-DP的许多重要特性。首先，忽略组件识别问题会导致模型质量随着并行线程", "数量的增加而降低，因为会产生更多的匹配错误（图3a）。其次，如果使用建议的优化适", "当地考虑了组件标识，则增加并行线程的数量可以减少执行时间，但不会影响最终的模型", "质量（图3b）。第三，SDA-DP（具有40个线程）在显着减少的时间内收敛到与比较算法", "相同的最终测试日志似然（图3c）。第四，每个组件识别优化通常需要～10-5秒，因此匹", "配占总计算的时间不到一毫秒，并且不会显着影响整体计算时间（图3d）。第五，大多数", "组件匹配问题在前80个小批量更新（总共2.000个）中得到解决-之后，真正的集群都被发", "现了，处理节点对这些集群做出了贡献，而不是创建新的集群，根据3.2节末尾的讨论（图", "3e）最后，增加的并行化可以有利于发现正确数量的集群；只有一个线程，早期犯的错误是", "建立在并持续存在，而更多的线程有更多的组件识别问题得到解决，因此有更多的机会发", "现正确的集群（图3f）", "5" ]

{ "arxiv_id": "1510.09161", "language": "zh", "page_idx": 4 }

[ [ "0.1013", "0.09155", "0.4673", "0.09155", "0.4673", "0.1048", "0.1013", "0.1048" ], [ "0.505", "0.09155", "0.8896", "0.09155", "0.8896", "0.1016", "0.505", "0.1016" ], [ "0.1013", "0.108", "0.468", "0.108", "0.468", "0.1218", "0.1013", "0.1218" ], [ "0.5034", "0.1042", "0.8906", "0.1042", "0.8906", "0.11743", "0.5034", "0.11743" ], [ "0.5034", "0.11743", "0.892", "0.11743", "0.892", "0.1307", "0.5034", "0.1307" ], [ "0.10046", "0.1237", "0.1609", "0.1256", "0.1609", "0.137", "0.10046", "0.1351" ], [ "0.5024", "0.132", "0.8823", "0.132", "0.8823", "0.1459", "0.5024", "0.1459" ], [ "0.0825", "0.1383", "0.468", "0.1383", "0.468", "0.1515", "0.0825", "0.1515" ], [ "0.5073", "0.1497", "0.5513", "0.1497", "0.5513", "0.1642", "0.5073", "0.1642" ], [ "0.589", "0.1515", "0.6543", "0.1515", "0.6543", "0.1635", "0.589", "0.1635" ], [ "0.697", "0.1528", "0.9067", "0.1528", "0.9067", "0.1635", "0.697", "0.1635" ], [ "0.511", "0.1622", "0.5317", "0.1622", "0.5317", "0.1755", "0.511", "0.1755" ], [ "0.10297", "0.1547", "0.4673", "0.1547", "0.4673", "0.1654", "0.10297", "0.1654" ], [ "0.5664", "0.1622", "0.6807", "0.1635", "0.6807", "0.1749", "0.5664", "0.1736" ], [ "0.7197", "0.1642", "0.7344", "0.1642", "0.7344", "0.1749", "0.7197", "0.1749" ], [ "0.8228", "0.1635", "0.847", "0.1635", "0.847", "0.1755", "0.8228", "0.1755" ], [ "0.513", "0.1736", "0.5337", "0.1736", "0.5337", "0.1863", "0.513", "0.1863" ], [ "0.8228", "0.1736", "0.846", "0.1736", "0.846", "0.1863", "0.8228", "0.1863" ], [ "0.6094", "0.1749", "0.636", "0.1749", "0.636", "0.1863", "0.6094", "0.1863" ], [ "0.7197", "0.1749", "0.734", "0.1749", "0.734", "0.1863", "0.7197", "0.1863" ], [ "0.513", "0.1843", "0.5337", "0.1843", "0.5337", "0.1976", "0.513", "0.1976" ], [ "0.6055", "0.1843", "0.6406", "0.1863", "0.6396", "0.1976", "0.6045", "0.1957" ], [ "0.7183", "0.1843", "0.7363", "0.1843", "0.7363", "0.1989", "0.7183", "0.1989" ], [ "0.10297", "0.1849", "0.4673", "0.1849", "0.4673", "0.1957", "0.10297", "0.1957" ], [ "0.1013", "0.1686", "0.4697", "0.1686", "0.4697", "0.1818", "0.1013", "0.1818" ], [ "0.8228", "0.1849", "0.846", "0.1849", "0.846", "0.1989", "0.8228", "0.1989" ], [ "0.5146", "0.1963", "0.5337", "0.1963", "0.5337", "0.209", "0.5146", "0.209" ], [ "0.609", "0.1963", "0.6357", "0.1963", "0.6357", "0.209", "0.609", "0.209" ], [ "0.7183", "0.1963", "0.7363", "0.1963", "0.7363", "0.2102", "0.7183", "0.2102" ], [ "0.8228", "0.1963", "0.846", "0.1963", "0.846", "0.209", "0.8228", "0.209" ], [ "0.513", "0.2084", "0.5337", "0.2084", "0.5337", "0.2203", "0.513", "0.2203" ], [ "0.10046", "0.1976", "0.468", "0.1989", "0.468", "0.2122", "0.10046", "0.2108" ], [ "0.589", "0.2084", "0.6543", "0.2102", "0.6543", "0.2216", "0.589", "0.2197" ], [ "0.7173", "0.2084", "0.7363", "0.2084", "0.7363", "0.2216", "0.7173", "0.2216" ], [ "0.8228", "0.2084", "0.846", "0.2084", "0.846", "0.2203", "0.8228", "0.2203" ], [ "0.589", "0.2184", "0.6562", "0.2203", "0.6562", "0.2317", "0.589", "0.2299" ], [ "0.10046", "0.2122", "0.468", "0.2134", "0.468", "0.2279", "0.10046", "0.2267" ], [ "0.513", "0.2197", "0.5337", "0.2197", "0.5337", "0.2317", "0.513", "0.2317" ], [ "0.7197", "0.2203", "0.734", "0.2203", "0.734", "0.2317", "0.7197", "0.2317" ], [ "0.8228", "0.2197", "0.846", "0.2197", "0.846", "0.2317", "0.8228", "0.2317" ], [ "0.1013", "0.2291", "0.2214", "0.2291", "0.2214", "0.2437", "0.1013", "0.2437" ], [ "0.513", "0.2299", "0.5337", "0.2299", "0.5337", "0.2437", "0.513", "0.2437" ], [ "0.591", "0.2311", "0.6543", "0.2311", "0.6543", "0.2424", "0.591", "0.2424" ], [ "0.7256", "0.2267", "0.739", "0.2374", "0.726", "0.2468", "0.7134", "0.2361" ], [ "0.824", "0.2311", "0.846", "0.2311", "0.846", "0.2437", "0.824", "0.2437" ], [ "0.5146", "0.2424", "0.535", "0.2424", "0.535", "0.2551", "0.5146", "0.2551" ], [ "0.5825", "0.2437", "0.665", "0.2437", "0.665", "0.254", "0.5825", "0.254" ], [ "0.7197", "0.2437", "0.7344", "0.2437", "0.7344", "0.254", "0.7197", "0.254" ], [ "0.824", "0.2424", "0.846", "0.2424", "0.846", "0.2551", "0.824", "0.2551" ], [ "0.5146", "0.252", "0.5435", "0.254", "0.54", "0.2676", "0.513", "0.2651" ], [ "0.7183", "0.254", "0.7363", "0.254", "0.7363", "0.2676", "0.7183", "0.2676" ], [ "0.8228", "0.252", "0.847", "0.2551", "0.8447", "0.2683", "0.8203", "0.2651" ], [ "0.591", "0.2551", "0.6562", "0.2551", "0.6562", "0.2651", "0.591", "0.2651" ], [ "0.0801", "0.2556", "0.2279", "0.2556", "0.2279", "0.2695", "0.0801", "0.2695" ], [ "0.513", "0.2644", "0.535", "0.2644", "0.535", "0.2766", "0.513", "0.2766" ], [ "0.5864", "0.2644", "0.6567", "0.2664", "0.6567", "0.2778", "0.5864", "0.276" ], [ "0.7197", "0.2651", "0.734", "0.2651", "0.734", "0.2766", "0.7197", "0.2766" ], [ "0.8228", "0.2644", "0.846", "0.2644", "0.846", "0.2766", "0.8228", "0.2766" ], [ "0.7173", "0.276", "0.7363", "0.276", "0.7363", "0.289", "0.7173", "0.289" ], [ "0.822", "0.276", "0.8486", "0.276", "0.8486", "0.2886", "0.822", "0.2886" ], [ "0.6055", "0.2766", "0.642", "0.2766", "0.642", "0.2886", "0.6055", "0.2886" ], [ "0.0784", "0.279", "0.4673", "0.281", "0.4673", "0.2954", "0.0784", "0.2935" ], [ "0.5146", "0.2766", "0.539", "0.2766", "0.539", "0.289", "0.5146", "0.289" ], [ "0.556", "0.2874", "0.6895", "0.2874", "0.6895", "0.3005", "0.556", "0.3005" ], [ "0.7173", "0.2874", "0.7363", "0.2874", "0.7363", "0.3005", "0.7173", "0.3005" ], [ "0.822", "0.2886", "0.8486", "0.2886", "0.8486", "0.3005", "0.822", "0.3005" ], [ "0.5537", "0.2986", "0.6895", "0.2986", "0.6895", "0.3125", "0.5537", "0.3125" ], [ "0.5146", "0.2886", "0.539", "0.2886", "0.539", "0.3005", "0.5146", "0.3005" ], [ "0.513", "0.2998", "0.5435", "0.2998", "0.5435", "0.3113", "0.513", "0.3113" ], [ "0.7197", "0.2998", "0.7344", "0.2998", "0.7344", "0.31", "0.7197", "0.31" ], [ "0.8203", "0.2998", "0.8486", "0.2998", "0.8486", "0.3113", "0.8203", "0.3113" ], [ "0.5415", "0.3037", "0.5483", "0.3037", "0.5483", "0.307", "0.5415", "0.307" ], [ "0.0825", "0.305", "0.4656", "0.305", "0.4656", "0.315", "0.0825", "0.315" ], [ "0.0801", "0.3176", "0.4673", "0.3176", "0.4673", "0.3308", "0.0801", "0.3308" ], [ "0.5034", "0.324", "0.8906", "0.3245", "0.8906", "0.339", "0.5034", "0.3384" ], [ "0.0825", "0.3352", "0.4673", "0.3352", "0.4673", "0.3452", "0.0825", "0.3452" ], [ "0.5034", "0.3403", "0.869", "0.3403", "0.869", "0.3542", "0.5034", "0.3542" ], [ "0.0809", "0.3484", "0.4697", "0.3484", "0.4697", "0.3623", "0.0809", "0.3623" ], [ "0.0801", "0.363", "0.468", "0.3643", "0.468", "0.3774", "0.0801", "0.377" ], [ "0.626", "0.358", "0.7974", "0.3726", "0.7935", "0.4048", "0.6216", "0.3901" ], [ "0.755", "0.3655", "0.79", "0.3655", "0.79", "0.3806", "0.755", "0.3806" ], [ "0.0801", "0.379", "0.468", "0.379", "0.468", "0.3926", "0.0801", "0.3926" ], [ "0.0809", "0.3933", "0.468", "0.3933", "0.468", "0.4072", "0.0809", "0.4072" ], [ "0.7695", "0.3933", "0.79", "0.4092", "0.7715", "0.4224", "0.751", "0.4072" ], [ "0.869", "0.3882", "0.8936", "0.3914", "0.8906", "0.406", "0.8667", "0.4028" ], [ "0.0809", "0.4092", "0.4673", "0.4092", "0.4673", "0.4224", "0.0809", "0.4224" ], [ "0.601", "0.4072", "0.7573", "0.406", "0.7573", "0.4204", "0.601", "0.4216" ], [ "0.7534", "0.414", "0.789", "0.414", "0.789", "0.4312", "0.7534", "0.4312" ], [ "0.0825", "0.4236", "0.468", "0.4248", "0.468", "0.4382", "0.0825", "0.4375" ], [ "0.0809", "0.4395", "0.468", "0.4395", "0.468", "0.4526", "0.0809", "0.4526" ], [ "0.523", "0.4343", "0.8896", "0.4343", "0.8896", "0.4475", "0.523", "0.4475" ], [ "0.0809", "0.4539", "0.468", "0.4539", "0.468", "0.4673", "0.0809", "0.4673" ], [ "0.5034", "0.4495", "0.669", "0.4495", "0.669", "0.4634", "0.5034", "0.4634" ], [ "0.0801", "0.4673", "0.4673", "0.4697", "0.4673", "0.483", "0.0801", "0.481" ], [ "0.5024", "0.476", "0.6006", "0.476", "0.6006", "0.4905", "0.5024", "0.4905" ], [ "0.0809", "0.4841", "0.468", "0.4841", "0.468", "0.4976", "0.0809", "0.4976" ], [ "0.0825", "0.5005", "0.1478", "0.5005", "0.1478", "0.512", "0.0825", "0.512" ], [ "0.505", "0.5024", "0.701", "0.5024", "0.701", "0.513", "0.505", "0.513" ], [ "0.0784", "0.521", "0.3022", "0.5215", "0.3022", "0.536", "0.0784", "0.535" ], [ "0.5034", "0.5244", "0.8906", "0.5244", "0.8906", "0.5386", "0.5034", "0.5386" ], [ "0.5024", "0.539", "0.892", "0.539", "0.892", "0.553", "0.5024", "0.553" ], [ "0.0809", "0.5454", "0.468", "0.5454", "0.468", "0.5596", "0.0809", "0.5596" ], [ "0.0825", "0.56", "0.468", "0.56", "0.468", "0.5737", "0.0825", "0.5737" ], [ "0.5024", "0.5547", "0.8906", "0.5547", "0.8906", "0.5684", "0.5024", "0.5684" ], [ "0.5034", "0.571", "0.8896", "0.571", "0.8896", "0.584", "0.5034", "0.584" ], [ "0.0825", "0.577", "0.4673", "0.577", "0.4673", "0.587", "0.0825", "0.587" ], [ "0.0809", "0.5903", "0.468", "0.5903", "0.468", "0.604", "0.0809", "0.604" ], [ "0.505", "0.5854", "0.8906", "0.5854", "0.8906", "0.5986", "0.505", "0.5986" ], [ "0.0801", "0.604", "0.468", "0.605", "0.468", "0.619", "0.0801", "0.6187" ], [ "0.501", "0.5996", "0.8896", "0.5996", "0.8896", "0.613", "0.501", "0.613" ], [ "0.0809", "0.6206", "0.468", "0.6206", "0.468", "0.634", "0.0809", "0.634" ], [ "0.505", "0.617", "0.8896", "0.617", "0.8896", "0.6274", "0.505", "0.6274" ], [ "0.0825", "0.6353", "0.468", "0.634", "0.468", "0.648", "0.0825", "0.6484" ], [ "0.5034", "0.631", "0.8906", "0.631", "0.8906", "0.6445", "0.5034", "0.6445" ], [ "0.0809", "0.651", "0.4673", "0.651", "0.4673", "0.664", "0.0809", "0.664" ], [ "0.5034", "0.646", "0.8906", "0.646", "0.8906", "0.659", "0.5034", "0.659" ], [ "0.501", "0.6577", "0.892", "0.659", "0.892", "0.677", "0.501", "0.6753" ], [ "0.0825", "0.6655", "0.4656", "0.6655", "0.4656", "0.6787", "0.0825", "0.6787" ], [ "0.0825", "0.681", "0.4631", "0.681", "0.4631", "0.6943", "0.0825", "0.6943" ], [ "0.5034", "0.6753", "0.8906", "0.6753", "0.8906", "0.6895", "0.5034", "0.6895" ], [ "0.5034", "0.69", "0.8906", "0.69", "0.8906", "0.704", "0.5034", "0.704" ], [ "0.0801", "0.7085", "0.2115", "0.7085", "0.2115", "0.7227", "0.0801", "0.7227" ], [ "0.5034", "0.7056", "0.892", "0.7056", "0.892", "0.7197", "0.5034", "0.7197" ], [ "0.5024", "0.7197", "0.8896", "0.72", "0.8896", "0.735", "0.5024", "0.7344" ], [ "0.0784", "0.731", "0.468", "0.728", "0.468", "0.7456", "0.0784", "0.7485" ], [ "0.5034", "0.7363", "0.892", "0.7363", "0.892", "0.75", "0.5034", "0.75" ], [ "0.0809", "0.7476", "0.468", "0.7476", "0.468", "0.761", "0.0809", "0.761" ], [ "0.0784", "0.7583", "0.4722", "0.7603", "0.4722", "0.78", "0.0784", "0.778" ], [ "0.505", "0.7524", "0.7036", "0.7524", "0.7036", "0.763", "0.505", "0.763" ], [ "0.5024", "0.765", "0.892", "0.7666", "0.892", "0.781", "0.5024", "0.78" ], [ "0.0801", "0.778", "0.4697", "0.778", "0.4697", "0.7954", "0.0801", "0.7954" ], [ "0.505", "0.782", "0.8906", "0.782", "0.8906", "0.7954", "0.505", "0.7954" ], [ "0.0784", "0.7915", "0.4697", "0.7944", "0.4697", "0.813", "0.0784", "0.81" ], [ "0.5034", "0.797", "0.8906", "0.797", "0.8906", "0.81", "0.5034", "0.81" ], [ "0.0809", "0.811", "0.4673", "0.811", "0.4673", "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" ], [ "0.505", "0.827", "0.888", "0.827", "0.888", "0.8403", "0.505", "0.8403" ], [ "0.0809", "0.8413", "0.468", "0.8413", "0.468", "0.855", "0.0809", "0.855" ], [ "0.5024", "0.8403", "0.8906", "0.8413", "0.8906", "0.855", "0.5024", "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" ], [ "0.505", "0.872", "0.8906", "0.872", "0.8906", "0.886", "0.505", "0.886" ], [ "0.0825", "0.888", "0.1299", "0.888", "0.1299", "0.8994", "0.0825", "0.8994" ], [ "0.5034", "0.887", "0.7485", "0.887", "0.7485", "0.901", "0.5034", "0.901" ], [ "0.4805", "0.925", "0.4927", "0.925", "0.4927", "0.9375", "0.4805", "0.9375" ] ]

[ "SatMAE++, that achieves impressive performance", "Table 1.Sentinel-2 bands description.The multi-spectral", "on both optical as well as multi-spectral satellite", "data contains three different GSD resolutions ranging from", "10mto 60m.Different band combinations can beutilized to", "imagery", "highlight specific information in the multi-spectral image", ": We perform extensive experimentation on six", "Band", "Description", "GSD (m)Central Wavelength (nm)", "B1", "datasets to validate the efficacy of our multi-scale", "Ultra Blue (Aerosol)", "60", "443", "B2", "490", "Blue", "10", "B3", "Green", "10", "new state-of-the-art on all six datasets.On the", "pre-training framework. Our SATMAE++ sets a", "560", "B4", "Red", "10", "665", "B5", "downstream task of land cover classification, our", "Red Edge 1", "20", "705", "Red Edge 2", "SatMAE++ achieves an absolute gain of 3.6 % over", "B6", "20", "740", "the baseline [5] :", "B7", "Red Edge 3", "20", "783", "B8", "Near Infra Red", "10", "842", "B8A", "20", "865", "Red Edge 4", "2. Related Work", "B9 ", "WaterVapor", "60", "940", "60", "1375", "Cirrus", "2.1. Representation Learning for Satellite Imagery", "B10", "Short Wave Infra Red 1", "20", "1610", "Short Wave Infra Red 2", "B11", "B12", "20", "2190", "", "最近，自监督学习已被广泛用作各种遥感任务的预训", "练步骤，使模型能够从未标记的数据中学习丰富的特", "位置编码转换器利用位置编码来学习补丁标记的空间", "征表示。GASSL2和SeCo116采用对比学习策略来", "位置。通常，变压器使用正弦位置编码，如下所示：", "展示不同下游任务预训练的有效性。Neumannet。al。", "[18]探索遥感图像域内表示学习的思想。Mendietaet。", "U(pos,2) = sin(前f)", "k", "al。[17]研究了工作中的持续学习策略和掩模图像建模", "技术。SatMAE5]将掩模图像建模用于多光谱，时间", "", "(1)", "和光学遥感数据。ScaleMAE[21]为光学卫星图像引人", "Vy(pos, 2i + 1) = cos(", "2%", "了基于地面采样距离（GSD）的位置编码，以对遥感", "数据中存在的比例信息进行编码。虽然ScaleMAE在", "其中pos是位置，i是特征尺寸索引，d是可能的总", "光学卫星数据上显示出了有希望的结果，但其复杂的", "位置，2是一个大常数", "GSD位置编码限制了其用于多光谱数据的能力。相反", "4. Method", "我们简化了在光学和多光谱卫星数据中提取多尺度信", "息的想法", "4.1.BaselineFramework", "2.2. Multi-scale Information", "我们采用最近的SatMAE[5]框架作为我们的基线。", "SatMAE遵循基于香草MAE的RGB数据架构，如", "视觉图像通常包含各种大小和比例的对象。因此，希望", "利用多尺度信息来学习更好的语义表示。卷积神经网络", "3节所述。为了处理多频谱数据，SatMAE基于多频", "谱信道的GSD执行信道分组。为了使训练稳定并减少", "和变换器1.7.9，10.14.15.19，20.23采用了多尺度", "信息，并在分类，检测和分割任务上显示出有希望的结", "自我关注操作的内存需求，将输入的空间大小减小为", "果。ConvMAE[8]引人了分层掩码策略，以学习混合卷", "96×96，而不是224×224[5]的标准输人大小。因此，", "积变换器编码器的多尺度特征。Point-M2AE[25]提出", "补丁大小减少到8。然后它创建单独的补丁嵌入层并在", "了一种用于3D点云的多尺度自动编码器。ScaleMAE", "空间维度中连接它们。SatMAE修改了MAE的位置", "[21]提出了一种基于拉普拉斯算子的解码器来学习多", "编码以合并光谱信息。为此，为每个通道组创建单独", "的编码，并将其连接到apos,i，ypos，位置编码，以便最", "尺度信息。我们对该方法进行了合理化，并引人了基于", "可扩展卷积的上采样块，以在多个尺度上重建特征图。", "终编码维度为D。之后，它将特征传递给变压器编码", "器。与香草MAE类似，解码器采用变压器编码器的输", "3. Background", "出，将可见令牌放置到其原始位置，并附加可学习的掩", "码令牌以获得N令牌。随后，将频谱编码添加到补丁", "屏蔽自动编码器（MAE）采用输人IERCxHxW，并", "令牌并馈送到一系列变压器块。最后，利用均方误差", "将图像调整为具有补丁大小P的非重叠补丁的序列S", "，其中 SERNxP℃和 N=（号）是补丁的数量。", "（MSE）损失来衡量重建质量", "限制：尽管上述基准框架在RGB和多光谱数据上运行", "利用补丁嵌入层f：RP℃→RD获得嵌人令牌序列", "良好，但是它没有利用遥感数据中存在的多尺度信息。", "SERN×D。N标记的一部分m被屏蔽，其余的1-m", "此外，上述基线框架努力在具有多个尺度级别的图像", "标记被输人到变换器编码器，该编码器利用位置嵌入", "的域之间进行泛化。此外，ScaleMAE[21]表明，缩放", "来捕获图像中补丁的空间信息。在解码器中，可见令牌", "信息可以通过基于GSD的位置嵌入进行编码。但是", "被放回其原始序列位置，并附加可学习掩码令牌以获", "ScaleMAE引人的基于GSD的位置嵌只能与每个通", "得N令牌。位置嵌入被添加到N令牌中，并被馈送", "道[21]具有相同GSD分辨率的RGB数据一起使用。", "到一系列转换器块。解码器最终利用均方误差损失重", "因此，需要重新考虑SatMAE框架的设计，以学习不", "建图像", "受单一数据模态约束的多尺度信息", "2" ]

{ "arxiv_id": "2403.05419", "language": "zh", "page_idx": 1 }

[ [ "0.7417", "0.09344", "0.7686", "0.09344", "0.7686", "0.1016", "0.7417", "0.1016" ], [ "0.1503", "0.1042", "0.1936", "0.1042", "0.1936", "0.11426", "0.1503", "0.11426" ], [ "0.7812", "0.11426", "0.822", "0.11426", "0.822", "0.1256", "0.7812", "0.1256" ], [ "0.384", "0.13", "0.5923", "0.137", "0.5884", "0.1932", "0.38", "0.1875" ], [ "0.3381", "0.1401", "0.411", "0.1427", "0.4102", "0.1755", "0.3367", "0.1736" ], [ "0.8384", "0.1433", "0.88", "0.1433", "0.88", "0.154", "0.8384", "0.154" ], [ "0.1912", "0.1509", "0.232", "0.1509", "0.232", "0.1622", "0.1912", "0.1622" ], [ "0.2615", "0.1445", "0.3145", "0.1338", "0.3284", "0.1724", "0.2754", "0.1831" ], [ "0.5728", "0.1515", "0.5825", "0.1515", "0.5825", "0.1692", "0.5728", "0.1692" ], [ "0.524", "0.1642", "0.5293", "0.1642", "0.5293", "0.1718", "0.524", "0.1718" ], [ "0.569", "0.1654", "0.574", "0.1654", "0.574", "0.1724", "0.569", "0.1724" ], [ "0.6313", "0.1692", "0.6807", "0.1718", "0.6787", "0.1843", "0.63", "0.1818" ], [ "0.5186", "0.1718", "0.5293", "0.1718", "0.5293", "0.1755", "0.5186", "0.1755" ], [ "0.429", "0.1674", "0.4429", "0.1674", "0.4429", "0.1718", "0.429", "0.1718" ], [ "0.286", "0.1768", "0.2957", "0.1799", "0.29", "0.1901", "0.281", "0.1875" ], [ "0.2222", "0.1976", "0.2507", "0.1976", "0.2507", "0.209", "0.2222", "0.209" ], [ "0.6216", "0.1925", "0.687", "0.1945", "0.687", "0.209", "0.6216", "0.207" ], [ "0.267", "0.1989", "0.308", "0.1989", "0.308", "0.2734", "0.267", "0.2734" ], [ "0.339", "0.1995", "0.3694", "0.202", "0.3652", "0.2412", "0.335", "0.2393" ], [ "0.4805", "0.2102", "0.4968", "0.207", "0.5", "0.2153", "0.4844", "0.2197" ], [ "0.6357", "0.2058", "0.6724", "0.2058", "0.6724", "0.2172", "0.6357", "0.2172" ], [ "0.4795", "0.2166", "0.5034", "0.2166", "0.5034", "0.2374", "0.4795", "0.2374" ], [ "0.569", "0.2267", "0.581", "0.2267", "0.581", "0.2349", "0.569", "0.2349" ], [ "0.589", "0.2299", "0.6216", "0.2299", "0.6216", "0.2424", "0.589", "0.2424" ], [ "0.5947", "0.2405", "0.6216", "0.2405", "0.6216", "0.2524", "0.5947", "0.2524" ], [ "0.361", "0.2615", "0.505", "0.262", "0.505", "0.2734", "0.361", "0.2727" ], [ "0.6216", "0.2632", "0.689", "0.2651", "0.689", "0.2798", "0.6216", "0.2778" ], [ "0.6396", "0.2747", "0.675", "0.279", "0.6724", "0.289", "0.636", "0.2854" ], [ "0.4355", "0.2917", "0.4468", "0.2917", "0.4468", "0.303", "0.4355", "0.303" ], [ "0.468", "0.2935", "0.4902", "0.2935", "0.4902", "0.3062", "0.468", "0.3062" ], [ "0.4673", "0.3018", "0.4983", "0.3062", "0.4958", "0.318", "0.464", "0.3145" ], [ "0.2876", "0.3145", "0.3015", "0.3145", "0.3015", "0.324", "0.2876", "0.324" ], [ "0.3218", "0.3176", "0.3447", "0.3176", "0.3447", "0.3289", "0.3218", "0.3289" ], [ "0.3186", "0.3276", "0.3464", "0.3276", "0.3464", "0.3403", "0.3186", "0.3403" ], [ "0.7427", "0.3257", "0.857", "0.3276", "0.857", "0.339", "0.7427", "0.3372" ], [ "0.0801", "0.3591", "0.8906", "0.3599", "0.8906", "0.375", "0.0801", "0.3738" ], [ "0.0809", "0.3738", "0.8896", "0.3738", "0.8896", "0.387", "0.0809", "0.387" ], [ "0.0809", "0.3882", "0.892", "0.3882", "0.892", "0.4016", "0.0809", "0.4016" ], [ "0.0801", "0.4016", "0.892", "0.4016", "0.892", "0.4153", "0.0801", "0.4153" ], [ "0.0801", "0.4153", "0.892", "0.4153", "0.892", "0.4287", "0.0801", "0.4287" ], [ "0.0809", "0.43", "0.8906", "0.43", "0.8906", "0.443", "0.0809", "0.443" ], [ "0.0801", "0.443", "0.5186", "0.443", "0.5186", "0.457", "0.0801", "0.457" ], [ "0.0809", "0.4749", "0.2793", "0.4749", "0.2793", "0.488", "0.0809", "0.488" ], [ "0.5024", "0.4734", "0.8906", "0.4749", "0.8906", "0.4893", "0.5024", "0.488" ], [ "0.5034", "0.4905", "0.8906", "0.4905", "0.8906", "0.504", "0.5034", "0.504" ], [ "0.0809", "0.4976", "0.468", "0.4976", "0.468", "0.511", "0.0809", "0.511" ], [ "0.5034", "0.505", "0.892", "0.505", "0.892", "0.518", "0.5034", "0.518" ], [ "0.08417", "0.513", "0.468", "0.513", "0.468", "0.5264", "0.08417", "0.5264" ], [ "0.5034", "0.518", "0.892", "0.5195", "0.892", "0.534", "0.5034", "0.5327" ], [ "0.0825", "0.528", "0.4673", "0.528", "0.4673", "0.5415", "0.0825", "0.5415" ], [ "0.5024", "0.534", "0.796", "0.534", "0.796", "0.5474", "0.5024", "0.5474" ], [ "0.0809", "0.5425", "0.4673", "0.5425", "0.4673", "0.556", "0.0809", "0.556" ], [ "0.524", "0.5503", "0.892", "0.5503", "0.892", "0.5645", "0.524", "0.5645" ], [ "0.0809", "0.558", "0.468", "0.558", "0.468", "0.5713", "0.0809", "0.5713" ], [ "0.505", "0.565", "0.8774", "0.565", "0.8774", "0.579", "0.505", "0.579" ], [ "0.0809", "0.5728", "0.468", "0.5728", "0.468", "0.586", "0.0809", "0.586" ], [ "0.5034", "0.5806", "0.8906", "0.5806", "0.8906", "0.5947", "0.5034", "0.5947" ], [ "0.08417", "0.5884", "0.468", "0.5884", "0.468", "0.6016", "0.08417", "0.6016" ], [ "0.5034", "0.595", "0.8896", "0.595", "0.8896", "0.6094", "0.5034", "0.6094" ], [ "0.0809", "0.604", "0.468", "0.604", "0.468", "0.617", "0.0809", "0.617" ], [ "0.505", "0.6113", "0.8896", "0.6113", "0.8896", "0.625", "0.505", "0.625" ], [ "0.0825", "0.6187", "0.4697", "0.6187", "0.4697", "0.632", "0.0825", "0.632" ], [ "0.505", "0.6255", "0.8896", "0.6255", "0.8896", "0.6396", "0.505", "0.6396" ], [ "0.0809", "0.6333", "0.4673", "0.6333", "0.4673", "0.6465", "0.0809", "0.6465" ], [ "0.5034", "0.6416", "0.8896", "0.6416", "0.8896", "0.655", "0.5034", "0.655" ], [ "0.0809", "0.648", "0.4656", "0.648", "0.4656", "0.661", "0.0809", "0.661" ], [ "0.505", "0.656", "0.8906", "0.656", "0.8906", "0.6694", "0.505", "0.6694" ], [ "0.0809", "0.662", "0.3906", "0.6636", "0.3906", "0.678", "0.0809", "0.677" ], [ "0.505", "0.672", "0.8906", "0.672", "0.8906", "0.685", "0.505", "0.685" ], [ "0.08417", "0.6787", "0.4673", "0.6787", "0.4673", "0.6924", "0.08417", "0.6924" ], [ "0.0809", "0.69", "0.468", "0.69", "0.468", "0.7085", "0.0809", "0.7085" ], [ "0.5024", "0.685", "0.8906", "0.686", "0.8906", "0.7007", "0.5024", "0.6997" ], [ "0.5024", "0.6997", "0.6094", "0.702", "0.6094", "0.7163", "0.5024", "0.714" ], [ "0.0825", "0.709", "0.4673", "0.709", "0.4673", "0.7227", "0.0825", "0.7227" ], [ "0.0809", "0.7246", "0.4673", "0.7246", "0.4673", "0.738", "0.0809", "0.738" ], [ "0.6216", "0.728", "0.773", "0.7266", "0.773", "0.744", "0.6216", "0.7456" ], [ "0.869", "0.73", "0.8906", "0.73", "0.8906", "0.744", "0.869", "0.744" ], [ "0.0825", "0.7393", "0.468", "0.7393", "0.468", "0.7524", "0.0825", "0.7524" ], [ "0.0809", "0.7505", "0.468", "0.7505", "0.468", "0.768", "0.0809", "0.768" ], [ "0.505", "0.7583", "0.8906", "0.7583", "0.8906", "0.7715", "0.505", "0.7715" ], [ "0.0825", "0.771", "0.4656", "0.771", "0.4656", "0.781", "0.0825", "0.781" ], [ "0.505", "0.773", "0.8906", "0.773", "0.8906", "0.786", "0.505", "0.786" ], [ "0.0825", "0.786", "0.464", "0.786", "0.464", "0.797", "0.0825", "0.797" ], [ "0.505", "0.787", "0.8906", "0.787", "0.8906", "0.8003", "0.505", "0.8003" ], [ "0.0809", "0.8", "0.2778", "0.8", "0.2778", "0.813", "0.0809", "0.813" ], [ "0.505", "0.803", "0.8906", "0.803", "0.8906", "0.8164", "0.505", "0.8164" ], [ "0.5034", "0.8184", "0.888", "0.8184", "0.888", "0.832", "0.5034", "0.832" ], [ "0.0825", "0.834", "0.1259", "0.834", "0.1259", "0.8447", "0.0825", "0.8447" ], [ "0.1323", "0.8354", "0.3882", "0.8354", "0.3882", "0.845", "0.1323", "0.845" ], [ "0.505", "0.8335", "0.8906", "0.8335", "0.8906", "0.8467", "0.505", "0.8467" ], [ "0.0768", "0.8496", "0.468", "0.851", "0.468", "0.8726", "0.0768", "0.872" ], [ "0.12335", "0.837", "0.1299", "0.837", "0.1299", "0.8423", "0.12335", "0.8423" ], [ "0.505", "0.8496", "0.5605", "0.8496", "0.5605", "0.8613", "0.505", "0.8613" ], [ "0.0768", "0.8643", "0.4722", "0.8643", "0.4722", "0.888", "0.0768", "0.888" ], [ "0.6216", "0.8613", "0.7715", "0.8594", "0.7715", "0.881", "0.6216", "0.8833" ], [ "0.8677", "0.8643", "0.8906", "0.8643", "0.8906", "0.8774", "0.8677", "0.8774" ], [ "0.0801", "0.881", "0.4722", "0.88", "0.4722", "0.9014", "0.0801", "0.903" ], [ "0.505", "0.887", "0.892", "0.887", "0.892", "0.901", "0.505", "0.901" ], [ "0.478", "0.924", "0.4944", "0.924", "0.4944", "0.938", "0.478", "0.938" ] ]

[ "(H,W", "(4H, 4W)", "(2H,2W)", "HH", "", "(4H,4W)", "(2H,2W)", "中", "", "", "", "Linear", "", "", "1", "(H, W)", "Upsample", "paeipna", "srupoa", "", "Block", "", "Li", "MSE", "Loss", "Loss=a,L,+a,L,+aL", "Upsample", "Block", "L", "L1", "Loss", "L3", "L1", "Loss", "Multi-scaleReconstruction", "Figure 1.Ilustration of Mask Autoencoder (MAE)framework for SatMAE++.The input image having spatial resolution", "of (4H,4W) is downsampled twice to obtain the images of resolution (2H,2W) and (H, W), respectively. We then feed the", "image with resolution (H, W) to the MAE similar to the SatMAE [5] framework. The decoder reconstructs the image at the", "resolution (H, W) and apply MSE loss to measure reconstruction quality. The reconstructed output is projected back to the", "feature space and is upsampled through upsampling blocks to obtain features at (2H,2W) and (4H, 4W) resolutions. The", "upsampled outputs are projected back to the image space and Li loss is utilized to penalize the reconstructions at higher", "resolutions. The overall loss is the weighted mean of all the losses.", "4.2.Overall Architecture", "MAE首先利用补丁嵌人层来修补输入图像。在多光谱", "输人的情况下，单独的补丁嵌入层用于不同的频带通", "图。1说明了SatMAE++的整体框架，该框架克服", "道组。之后，沿着空间维度连接不同组的补丁令牌。我", "了基线框架SatMAE[5]和其他最新方法的局限性，用", "们屏蔽补丁令牌的75个%，类似于其他MAE方法。", "于多光谱（fMoW-Sentinel）以及RGB数据（fMoW-", "然后，将位置编码添加到可见补丁标记中", "RGB）的多尺度预训练。我们在三个尺度级别中的大多", "在[5]之后，我们使用不依赖于GSD信息的通用位", "数级别上获取输入图像，并将最低尺度级别的图像馈", "置编码（如公式E1所示）。可见补丁令牌被馈送到", "送到SatMAE框架。基本框架获取输人图像，应用补", "系列转换器块，这些转换器块产生编码的可见特征。与", "丁嵌入和掩码操作，并将可见令牌提供给转换器编码", "SatMAE类似，解码器从变换器编码器获取编码的可", "器。随后，解码器获取编码器的输出，并重建与最低比", "见特征，并应用线性投影以减少嵌人尺寸。然后，将可", "例级别输入具有相同空间尺寸的图像。来自SatMAE", "见特征放回其原始索引位置，并将可学习掩码标记附", "模型的重建输出被上采样块用来在更高的尺度水平上", "加到可见标记。之后，将RGB或多光谱位置编码添加", "执行细粒度重建。在更高尺度上的重建鼓励模型学习", "到补丁令牌中。最终，补丁令牌被馈送到解码器变换", "多尺度表示，从而提高各种下游任务的性能", "器，并且最终的投影层将解码的特征映射回空间域。将", "上采样块：图。2显示了上采样块的架构。它采用输", "人特征XERC×HxW，并将它们通过转置卷积层，以", "解码图像F整形为原始输入尺寸，并利用均方误差来", "计算重建质量。", "对特征的空间分辨率进行上采样。我们将上采样的特", "征归一化，然后应用泄漏的relu激活操作。然后，将", "L=(F-I)?", "(2)", "由两个卷积层组成的残差块用于局部特征增强。通过", "使用线性投影层将增强的特征XeRC×2Hx2W投影回", "从解码器获得重建的输人后，我们使用线性投影将重", "空间域，并利用平均绝对误差来计算输入和重建图像", "建的图像映射回特征空间。然后，我们利用转置卷积以", "之间的重建误差。接下来，我们详细讨论在两个和三个", "(2H×2W）的分辨率对特征图进行上采样。上采样的", "尺度级别的多尺度重建过程", "特征映射是从由两个卷积层组成的残差块传递的。最", "后，我们将特征投影回图像空间以获得缩放的重建图", "4.2.1", "Reconstruction at TwoScaleLevels", "像F，并应用L1损失来分析模型在更高比例下的重", "让IERCxHxW成为MAE的输人图像。我们拍摄分", "", "建性能。", "辨率为RC×21x2W的图像1，并将其缩小以获得尺", "L2=F-i", "(3)", "寸为RCXHxW的图像1，图像I被馈送到MAE中，", "同样的超分辨率方法，我们利用较高比例的L1损失，", "3" ]

{ "arxiv_id": "2403.05419", "language": "zh", "page_idx": 2 }

[ [ "0.1217", "0.0947", "0.2197", "0.0966", "0.2197", "0.1111", "0.1217", "0.1092" ], [ "0.5034", "0.089", "0.892", "0.089", "0.892", "0.1029", "0.5034", "0.1029" ], [ "0.5034", "0.1042", "0.8906", "0.1048", "0.8906", "0.11615", "0.5034", "0.11554" ], [ "0.2115", "0.11743", "0.2173", "0.11743", "0.2173", "0.137", "0.2115", "0.137" ], [ "0.5024", "0.11743", "0.8906", "0.11615", "0.8906", "0.1307", "0.5024", "0.132" ], [ "0.1168", "0.1288", "0.12335", "0.1276", "0.1259", "0.1364", "0.11926", "0.1383" ], [ "0.11926", "0.137", "0.12335", "0.137", "0.12335", "0.1459", "0.11926", "0.1459" ], [ "0.5034", "0.1332", "0.8896", "0.1332", "0.8896", "0.1433", "0.5034", "0.1433" ], [ "0.5024", "0.1445", "0.888", "0.1459", "0.888", "0.159", "0.5024", "0.1578" ], [ "0.1847", "0.1528", "0.1993", "0.1528", "0.1993", "0.2267", "0.1847", "0.2267" ], [ "0.2059", "0.1515", "0.2197", "0.1515", "0.2197", "0.1849", "0.2059", "0.1849" ], [ "0.501", "0.1578", "0.6685", "0.159", "0.6685", "0.1736", "0.501", "0.1724" ], [ "0.255", "0.1654", "0.2754", "0.1654", "0.2754", "0.2108", "0.255", "0.2108" ], [ "0.2566", "0.1666", "0.3423", "0.1547", "0.3545", "0.2084", "0.2688", "0.2197" ], [ "0.1708", "0.1787", "0.1805", "0.1787", "0.1805", "0.1989", "0.1708", "0.1989" ], [ "0.1209", "0.1849", "0.134", "0.1849", "0.134", "0.1932", "0.1209", "0.1932" ], [ "0.2361", "0.1863", "0.2443", "0.1863", "0.2443", "0.1913", "0.2361", "0.1913" ], [ "0.505", "0.1863", "0.576", "0.1863", "0.576", "0.2008", "0.505", "0.2008" ], [ "0.706", "0.1863", "0.778", "0.1863", "0.778", "0.2008", "0.706", "0.2008" ], [ "0.826", "0.1875", "0.9043", "0.1875", "0.9043", "0.1989", "0.826", "0.1989" ], [ "0.3162", "0.1976", "0.33", "0.1989", "0.3284", "0.209", "0.3145", "0.2084" ], [ "0.5146", "0.1995", "0.5933", "0.202", "0.5933", "0.2166", "0.5146", "0.214" ], [ "0.7017", "0.1995", "0.7793", "0.202", "0.7793", "0.2166", "0.7017", "0.214" ], [ "0.8423", "0.202", "0.884", "0.202", "0.884", "0.2166", "0.8423", "0.2166" ], [ "0.282", "0.2084", "0.2942", "0.2026", "0.3015", "0.2134", "0.2893", "0.2184" ], [ "0.5156", "0.2172", "0.6177", "0.2172", "0.6177", "0.2317", "0.5156", "0.2317" ], [ "0.705", "0.2184", "0.778", "0.2184", "0.778", "0.2299", "0.705", "0.2299" ], [ "0.8423", "0.214", "0.884", "0.2166", "0.8823", "0.2317", "0.841", "0.2299" ], [ "0.7017", "0.2299", "0.7793", "0.2317", "0.7793", "0.2468", "0.7017", "0.2443" ], [ "0.8447", "0.2291", "0.8813", "0.2311", "0.88", "0.2468", "0.8433", "0.2443" ], [ "0.6807", "0.2437", "0.8037", "0.2456", "0.8037", "0.2632", "0.6807", "0.2615" ], [ "0.517", "0.2329", "0.5864", "0.2329", "0.5864", "0.2443", "0.517", "0.2443" ], [ "0.5156", "0.2468", "0.6167", "0.2468", "0.6167", "0.2615", "0.5156", "0.2615" ], [ "0.847", "0.2474", "0.8784", "0.2474", "0.8784", "0.26", "0.847", "0.26" ], [ "0.7036", "0.26", "0.7812", "0.2632", "0.7793", "0.2778", "0.7017", "0.2747" ], [ "0.846", "0.2615", "0.8813", "0.2615", "0.8813", "0.2766", "0.846", "0.2766" ], [ "0.0809", "0.2734", "0.4673", "0.2734", "0.4673", "0.2874", "0.0809", "0.2874" ], [ "0.5156", "0.262", "0.6235", "0.262", "0.6235", "0.2766", "0.5156", "0.2766" ], [ "0.5146", "0.2747", "0.603", "0.2766", "0.603", "0.2917", "0.5146", "0.289" ], [ "0.705", "0.2747", "0.7793", "0.2778", "0.778", "0.2922", "0.7036", "0.289" ], [ "0.8423", "0.2766", "0.884", "0.2766", "0.884", "0.2917", "0.8423", "0.2917" ], [ "0.0801", "0.2886", "0.468", "0.2886", "0.468", "0.3018", "0.0801", "0.3018" ], [ "0.7036", "0.2903", "0.7793", "0.2922", "0.7793", "0.307", "0.7036", "0.305" ], [ "0.8423", "0.2903", "0.8823", "0.2922", "0.8813", "0.308", "0.841", "0.3062" ], [ "0.0809", "0.3018", "0.4673", "0.3018", "0.4673", "0.315", "0.0809", "0.315" ], [ "0.5156", "0.2922", "0.614", "0.2922", "0.614", "0.307", "0.5156", "0.307" ], [ "0.5107", "0.307", "0.6587", "0.308", "0.6587", "0.3225", "0.5107", "0.3213" ], [ "0.7036", "0.3062", "0.7812", "0.3093", "0.7793", "0.324", "0.7017", "0.3208" ], [ "0.8423", "0.308", "0.884", "0.308", "0.884", "0.3225", "0.8423", "0.3225" ], [ "0.0801", "0.315", "0.4673", "0.315", "0.4673", "0.3289", "0.0801", "0.3289" ], [ "0.0801", "0.33", "0.4697", "0.33", "0.4697", "0.3435", "0.0801", "0.3435" ], [ "0.0801", "0.3435", "0.10706", "0.3435", "0.10706", "0.3547", "0.0801", "0.3547" ], [ "0.5024", "0.3516", "0.706", "0.3528", "0.706", "0.3674", "0.5024", "0.3662" ], [ "0.0809", "0.375", "0.4126", "0.375", "0.4126", "0.3882", "0.0809", "0.3882" ], [ "0.505", "0.3774", "0.8906", "0.3774", "0.8906", "0.3914", "0.505", "0.3914" ], [ "0.505", "0.3926", "0.83", "0.3926", "0.83", "0.406", "0.505", "0.406" ], [ "0.1993", "0.4016", "0.3506", "0.4028", "0.3506", "0.4172", "0.1993", "0.4167" ], [ "0.4468", "0.404", "0.468", "0.404", "0.468", "0.4172", "0.4468", "0.4172" ], [ "0.5034", "0.4092", "0.8906", "0.4092", "0.8906", "0.4224", "0.5034", "0.4224" ], [ "0.5063", "0.4236", "0.8906", "0.4236", "0.8906", "0.4375", "0.5063", "0.4375" ], [ "0.5034", "0.4395", "0.7974", "0.4395", "0.7974", "0.4526", "0.5034", "0.4526" ], [ "0.5024", "0.4539", "0.8906", "0.455", "0.8906", "0.4697", "0.5024", "0.4685" ], [ "0.0809", "0.4702", "0.4019", "0.4702", "0.4019", "0.4841", "0.0809", "0.4841" ], [ "0.505", "0.4702", "0.892", "0.4702", "0.892", "0.4841", "0.505", "0.4841" ], [ "0.5034", "0.4849", "0.8906", "0.4849", "0.8906", "0.5", "0.5034", "0.5" ], [ "0.0825", "0.4956", "0.4656", "0.4956", "0.4656", "0.509", "0.0825", "0.509" ], [ "0.505", "0.5005", "0.8906", "0.5005", "0.8906", "0.5146", "0.505", "0.5146" ], [ "0.0825", "0.5103", "0.468", "0.5103", "0.468", "0.5234", "0.0825", "0.5234" ], [ "0.5034", "0.5166", "0.6855", "0.5166", "0.6855", "0.53", "0.5034", "0.53" ], [ "0.0801", "0.523", "0.468", "0.5195", "0.468", "0.5405", "0.0801", "0.5435" ], [ "0.505", "0.532", "0.8906", "0.532", "0.8906", "0.5454", "0.505", "0.5454" ], [ "0.0809", "0.5405", "0.4673", "0.5405", "0.4673", "0.5537", "0.0809", "0.5537" ], [ "0.0784", "0.5503", "0.4714", "0.5503", "0.4714", "0.5713", "0.0784", "0.5713" ], [ "0.505", "0.5474", "0.892", "0.5474", "0.892", "0.561", "0.505", "0.561" ], [ "0.5034", "0.5625", "0.8906", "0.5625", "0.8906", "0.5757", "0.5034", "0.5757" ], [ "0.0801", "0.571", "0.1871", "0.565", "0.1887", "0.579", "0.0809", "0.584" ], [ "0.183", "0.571", "0.468", "0.5693", "0.468", "0.584", "0.183", "0.5854" ], [ "0.5034", "0.5776", "0.888", "0.5776", "0.888", "0.5913", "0.5034", "0.5913" ], [ "0.0809", "0.586", "0.468", "0.586", "0.468", "0.5996", "0.0809", "0.5996" ], [ "0.5034", "0.5923", "0.8906", "0.5923", "0.8906", "0.606", "0.5034", "0.606" ], [ "0.0809", "0.601", "0.468", "0.601", "0.468", "0.6143", "0.0809", "0.6143" ], [ "0.5034", "0.608", "0.8906", "0.608", "0.8906", "0.622", "0.5034", "0.622" ], [ "0.0825", "0.616", "0.4673", "0.616", "0.4673", "0.63", "0.0825", "0.63" ], [ "0.5034", "0.6226", "0.8906", "0.6226", "0.8906", "0.636", "0.5034", "0.636" ], [ "0.0809", "0.631", "0.468", "0.631", "0.468", "0.6445", "0.0809", "0.6445" ], [ "0.5034", "0.638", "0.892", "0.638", "0.892", "0.6514", "0.5034", "0.6514" ], [ "0.0809", "0.6445", "0.468", "0.646", "0.468", "0.66", "0.0809", "0.659" ], [ "0.5034", "0.654", "0.7817", "0.654", "0.7817", "0.6675", "0.5034", "0.6675" ], [ "0.0825", "0.661", "0.468", "0.661", "0.468", "0.675", "0.0825", "0.675" ], [ "0.0809", "0.6753", "0.4673", "0.6753", "0.4673", "0.6895", "0.0809", "0.6895" ], [ "0.5034", "0.681", "0.88", "0.681", "0.88", "0.696", "0.5034", "0.696" ], [ "0.08417", "0.6934", "0.3179", "0.6934", "0.3179", "0.704", "0.08417", "0.704" ], [ "0.505", "0.705", "0.892", "0.705", "0.892", "0.7197", "0.505", "0.7197" ], [ "0.2303", "0.714", "0.327", "0.7236", "0.3218", "0.755", "0.2255", "0.7456" ], [ "0.505", "0.72", "0.8906", "0.72", "0.8906", "0.735", "0.505", "0.735" ], [ "0.2018", "0.728", "0.2303", "0.73", "0.2296", "0.741", "0.2009", "0.7393" ], [ "0.2255", "0.7373", "0.2507", "0.7373", "0.2507", "0.7456", "0.2255", "0.7456" ], [ "0.4468", "0.738", "0.4697", "0.738", "0.4697", "0.7505", "0.4468", "0.7505" ], [ "0.5034", "0.7363", "0.8906", "0.7363", "0.8906", "0.75", "0.5034", "0.75" ], [ "0.2524", "0.744", "0.2695", "0.744", "0.2695", "0.752", "0.2524", "0.752" ], [ "0.1854", "0.755", "0.3677", "0.756", "0.3677", "0.771", "0.1854", "0.7695" ], [ "0.505", "0.752", "0.8896", "0.752", "0.8896", "0.765", "0.505", "0.765" ], [ "0.5034", "0.7666", "0.888", "0.7666", "0.888", "0.78", "0.5034", "0.78" ], [ "0.0768", "0.784", "0.2181", "0.785", "0.2181", "0.8", "0.0768", "0.799" ], [ "0.505", "0.7827", "0.8896", "0.7827", "0.8896", "0.7935", "0.505", "0.7935" ], [ "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", "0.8906", "0.811", "0.8906", "0.8247", "0.505", "0.8247" ], [ "0.0825", "0.827", "0.4656", "0.827", "0.4656", "0.8403", "0.0825", "0.8403" ], [ "0.5024", "0.827", "0.8896", "0.827", "0.8896", "0.8403", "0.5024", "0.8403" ], [ "0.0825", "0.8413", "0.468", "0.8413", "0.468", "0.855", "0.0825", "0.855" ], [ "0.5024", "0.8413", "0.8896", "0.8413", "0.8896", "0.855", "0.5024", "0.855" ], [ "0.0809", "0.857", "0.452", "0.857", "0.452", "0.8706", "0.0809", "0.8706" ], [ "0.505", "0.856", "0.8906", "0.856", "0.8906", "0.869", "0.505", "0.869" ], [ "0.0809", "0.872", "0.468", "0.872", "0.468", "0.8853", "0.0809", "0.8853" ], [ "0.505", "0.872", "0.892", "0.872", "0.892", "0.8853", "0.505", "0.8853" ], [ "0.0809", "0.887", "0.1152", "0.887", "0.1152", "0.8994", "0.0809", "0.8994" ], [ "0.5034", "0.8853", "0.8447", "0.886", "0.8447", "0.901", "0.5034", "0.8994" ], [ "0.4795", "0.925", "0.4927", "0.925", "0.4927", "0.936", "0.4795", "0.936" ] ]

[ "Upsample Block", "Table 2. Comparison of state-of-the-art methods on the", "validationset of fMoW-RGBdataset.*representsthatwe", "OP", "reproduce the results using the public codebase provided", "-", "-", "bythe authorsand pre-train themodelfor 8o0 epochs on", "fMoW-RGB, which are consistent to the performance re", "ose", "e+on", "ported in ScaleMAE [21]", "AUOD", "", "", "IX", "+", "Method", "Backbone", "Topl Acc.", "", "GASSL[2]", "ResNet-50", "71.55", "", "MoCo-V2 [11]", "ResNet-50", "64.34", "ViT-Large", "68.4", "Conv ViT-Large", "MAE[12]", "ConvMAE [8]", "74.1", "ViT-Large", "77.9", "Figure 2. The illustration of upsample block used in Sat-", "ScaleMAE [2]]", "SatMAE[5]", "ViT-Large", "77.84", "MAE++ framework. Input features X are upsampled by", "ViT-Large", "73.87", "utilizing the transpose convolution operation. Afterwards,", "SatMAE * [5]", "SatMAE++(Ours)", "ViT-Large", "78.14", "a residual block which is composed of two convolution lay-", "ers is employed to enhance the upsampled features given as", "X.", "5.1. Pre-training Datasets", "以便模型可以学习重建实际图像值。总损耗为：", "我们利用了两个公开的大规模数据集，用于在多光谱", "和RGB卫星数据上对视觉转换器进行预训练", "Loss=Q1L1+Q2L2", "(4)", "fMoW-RGB：世界功能地图（fMoW）【4]是高分辨率", "卫星图像的大规模公开数据集。数据集分为62个分类", "任务，包括约363k训练和53k测试图像", "fMoW-Sentinel：SatMAE[5]优化并扩展了用于分类任", "4.2.2Reconstruction at Three Scale Levels", "务的fMoW-RGB，并包括图像的Sentinel-2数据。与", "fMoW-RGB类似，此数据集有62个类类别。数据集", "对于多光谱数据，由于与RGB数据相比，模型输入的", "包含更多数量的图像，包括712874个训练、84939个", "分辨率较小，因此我们以三个比例级别重建图像（RGB", "验证和84966个测试图像", "使用224×224像素）。在这里，我们以RC×4H×4W", "重建结果：我们在图中展示了fMoW-Sentinel数据集", "的更高分辨率拍摄图像1。然后，我们对图像I进", "行两次下采样，以分别获得图像1EIRCX2Hx2W和", "上的多尺度重建结果。3。通过采用多尺度预训练，", "我们观察到图像重建质量的提高。此外，我们将我们的", "IERCxHxW", "。如上一节所述，我们以（H,W）的空", "方法的重建结果与图中的基线方法（SatMAE）进行了", "间分辨率重建图像F。之后，应用线性投影层将数据", "比较。4。SatMAE在可见斑块上的重建质量比掩蔽", "投影到特征空间并将其输人到上采样块。上采样块使", "斑块差得多。然而，我们的多尺度方法提高了可见和掩", "用转置卷积将空间分辨率提高两倍，然后将其馈送到", "蔽斑块的重建质量。与SatMAE相比，在实现良好重", "残差卷积块以获得具有空间分辨率（2H,2W）的特征", "建的同时（参见图。3和4），我们观察到我们在某些", "F。将特征F输人到另一个上采样块，以获得尺寸为", "情况下重建细粒度结构细节的斗争方法", "(4H,4W）的特征F。特征F和F都被投影回图像空", "间，并应用L1损失来测量重建特征质量。总损失是三", "5.2. Pre-training and Finetuning on fMoW-RGB", "个损失的加权平均值，如下所示：", "预训练：与SatMAE[5]和ScaleMAE[21]类似，我", "F-", "们在fMoW-RGB数据集上预训练ViT-Large[6]模型。", "L3=", "n", "(5)", "模型的配置与[5]i中使用的配置相同。e。，输人图像", "=1", "Loss=Q1L1+α2L2+Q3L3", "的空间分辨率为（224×224），补丁大小为16。在预训", "练期间，我们将图像的短边调整为448像素，然后从", "5.Experiments", "调整大小的图像中随机裁剪448×448像素。然后通过", "使用双线性插值将尺寸为448×448像素的图像下采样", "在本节中，我们首先讨论主流数据集。然后，我们解释", "为224×224像素的较低分辨率。然后，我们将大小为", "用于主流基准的预训练和微调程序。之后，我们讨论了", "224×224像素的图像输人到模型中。我们在预训练实", "用于迁移学习实验的数据集及其微调方法。。我们通过", "验中使用AdamW优化器和余弦学习率调度器。初始", "针对SatMAE++的简化的多尺度MAE预训练进", "学习速率设置为7e-4，单GPU使用64的批处理大小。", "步讨论了学习表示的性能，并给出了下游任务的微调", "类似于??？在预训练模型时屏蔽补丁的%。我们利用8", "结果", "个NVIDIAV100GPU来训练800个时代的模型", "4" ]

{ "arxiv_id": "2403.05419", "language": "zh", "page_idx": 3 }

[ [ "0.1871", "0.0947", "0.3096", "0.09534", "0.3096", "0.10986", "0.1871", "0.1092" ], [ "0.596", "0.0947", "0.7036", "0.09534", "0.7036", "0.10986", "0.596", "0.1092" ], [ "0.357", "0.108", "0.3757", "0.108", "0.3757", "0.1218", "0.357", "0.1218" ], [ "0.7656", "0.1067", "0.785", "0.1067", "0.785", "0.1218", "0.7656", "0.1218" ], [ "0.2385", "0.1225", "0.2573", "0.1225", "0.2573", "0.137", "0.2385", "0.137" ], [ "0.6465", "0.1225", "0.6665", "0.1225", "0.6665", "0.137", "0.6465", "0.137" ], [ "0.1528", "0.132", "0.1708", "0.132", "0.1708", "0.1465", "0.1528", "0.1465" ], [ "0.4756", "0.1332", "0.4944", "0.1332", "0.4944", "0.1465", "0.4756", "0.1465" ], [ "0.5605", "0.1332", "0.58", "0.1332", "0.58", "0.1465", "0.5605", "0.1465" ], [ "0.1356", "0.3176", "0.1896", "0.3176", "0.1896", "0.332", "0.1356", "0.332" ], [ "0.2607", "0.3384", "0.2932", "0.3384", "0.2932", "0.351", "0.2607", "0.351" ], [ "0.4558", "0.3384", "0.513", "0.3384", "0.513", "0.3528", "0.4558", "0.3528" ], [ "0.669", "0.3372", "0.706", "0.3372", "0.706", "0.3528", "0.669", "0.3528" ], [ "0.0809", "0.3718", "0.888", "0.3718", "0.888", "0.385", "0.0809", "0.385" ], [ "0.0801", "0.385", "0.689", "0.385", "0.689", "0.3984", "0.0801", "0.3984" ], [ "0.697", "0.387", "0.8906", "0.387", "0.8906", "0.3977", "0.697", "0.3977" ], [ "0.0809", "0.3997", "0.8896", "0.3997", "0.8896", "0.4136", "0.0809", "0.4136" ], [ "0.0784", "0.4124", "0.3938", "0.4136", "0.3938", "0.428", "0.0784", "0.4268" ], [ "0.0825", "0.4507", "0.4673", "0.4507", "0.4673", "0.461", "0.0825", "0.461" ], [ "0.5034", "0.4458", "0.8906", "0.4458", "0.8906", "0.4602", "0.5034", "0.4602" ], [ "0.5034", "0.461", "0.625", "0.461", "0.625", "0.4717", "0.5034", "0.4717" ], [ "0.625", "0.4602", "0.8896", "0.4602", "0.8896", "0.4734", "0.625", "0.4734" ], [ "0.5034", "0.476", "0.8896", "0.476", "0.8896", "0.486", "0.5034", "0.486" ], [ "0.0825", "0.4634", "0.468", "0.4634", "0.468", "0.4766", "0.0825", "0.4766" ], [ "0.0825", "0.4792", "0.4673", "0.4792", "0.4673", "0.4924", "0.0825", "0.4924" ], [ "0.5024", "0.488", "0.892", "0.488", "0.892", "0.502", "0.5024", "0.502" ], [ "0.09644", "0.4956", "0.4656", "0.4956", "0.4656", "0.506", "0.09644", "0.506" ], [ "0.5024", "0.5024", "0.5522", "0.5024", "0.5522", "0.5146", "0.5024", "0.5146" ], [ "0.0809", "0.509", "0.468", "0.509", "0.468", "0.523", "0.0809", "0.523" ], [ "0.0825", "0.5244", "0.1152", "0.5244", "0.1152", "0.537", "0.0825", "0.537" ], [ "0.694", "0.5264", "0.7686", "0.529", "0.7686", "0.5435", "0.694", "0.5415" ], [ "0.8013", "0.529", "0.8774", "0.529", "0.8774", "0.5435", "0.8013", "0.5435" ], [ "0.0809", "0.539", "0.468", "0.539", "0.468", "0.553", "0.0809", "0.553" ], [ "0.5186", "0.53", "0.5786", "0.53", "0.5786", "0.5415", "0.5186", "0.5415" ], [ "0.6953", "0.5425", "0.767", "0.5444", "0.767", "0.558", "0.6953", "0.556" ], [ "0.8164", "0.5435", "0.8594", "0.5435", "0.8594", "0.558", "0.8164", "0.558" ], [ "0.0809", "0.5537", "0.468", "0.5537", "0.468", "0.5684", "0.0809", "0.5684" ], [ "0.5195", "0.5454", "0.5923", "0.5454", "0.5923", "0.5566", "0.5195", "0.5566" ], [ "0.5186", "0.558", "0.649", "0.5596", "0.649", "0.5737", "0.5186", "0.5728" ], [ "0.694", "0.556", "0.767", "0.558", "0.767", "0.5737", "0.694", "0.5713" ], [ "0.8164", "0.558", "0.859", "0.558", "0.859", "0.5737", "0.8164", "0.5737" ], [ "0.0809", "0.5693", "0.464", "0.5693", "0.464", "0.5835", "0.0809", "0.5835" ], [ "0.5195", "0.5737", "0.619", "0.5737", "0.619", "0.5884", "0.5195", "0.5884" ], [ "0.6953", "0.5737", "0.767", "0.5737", "0.767", "0.5884", "0.6953", "0.5884" ], [ "0.8164", "0.5737", "0.857", "0.5737", "0.857", "0.589", "0.8164", "0.589" ], [ "0.0809", "0.5854", "0.468", "0.5854", "0.468", "0.5986", "0.0809", "0.5986" ], [ "0.5186", "0.5903", "0.66", "0.5903", "0.66", "0.604", "0.5186", "0.604" ], [ "0.694", "0.5884", "0.7686", "0.5903", "0.7686", "0.605", "0.694", "0.603" ], [ "0.8164", "0.5903", "0.859", "0.5903", "0.859", "0.605", "0.8164", "0.605" ], [ "0.0809", "0.5996", "0.468", "0.5996", "0.468", "0.613", "0.0809", "0.613" ], [ "0.5186", "0.606", "0.6406", "0.606", "0.6406", "0.6206", "0.5186", "0.6206" ], [ "0.694", "0.603", "0.7686", "0.606", "0.767", "0.622", "0.693", "0.6187" ], [ "0.8164", "0.606", "0.857", "0.606", "0.857", "0.6206", "0.8164", "0.6206" ], [ "0.0825", "0.6157", "0.4673", "0.6157", "0.4673", "0.629", "0.0825", "0.629" ], [ "0.5186", "0.6187", "0.607", "0.6206", "0.607", "0.6353", "0.5186", "0.6333" ], [ "0.693", "0.617", "0.7695", "0.622", "0.7686", "0.637", "0.6914", "0.6333" ], [ "0.8164", "0.6206", "0.859", "0.6206", "0.859", "0.6353", "0.8164", "0.6353" ], [ "0.0825", "0.63", "0.468", "0.63", "0.468", "0.643", "0.0825", "0.643" ], [ "0.5195", "0.6353", "0.619", "0.6353", "0.619", "0.6494", "0.5195", "0.6494" ], [ "0.694", "0.632", "0.7686", "0.636", "0.767", "0.6514", "0.693", "0.648" ], [ "0.8164", "0.6353", "0.857", "0.6353", "0.857", "0.6494", "0.8164", "0.6494" ], [ "0.0801", "0.643", "0.3406", "0.6445", "0.3406", "0.659", "0.0801", "0.6577" ], [ "0.694", "0.6484", "0.7686", "0.6514", "0.767", "0.6675", "0.693", "0.664" ], [ "0.8145", "0.651", "0.8594", "0.651", "0.8594", "0.6655", "0.8145", "0.6655" ], [ "0.1422", "0.6665", "0.4673", "0.6665", "0.4673", "0.681", "0.1422", "0.681" ], [ "0.517", "0.6514", "0.661", "0.6514", "0.661", "0.6655", "0.517", "0.6655" ], [ "0.0809", "0.6675", "0.1136", "0.6675", "0.1136", "0.6797", "0.0809", "0.6797" ], [ "0.1209", "0.683", "0.1847", "0.683", "0.1847", "0.6943", "0.1209", "0.6943" ], [ "0.0801", "0.705", "0.468", "0.704", "0.468", "0.7183", "0.0801", "0.7197" ], [ "0.505", "0.6997", "0.8906", "0.6997", "0.8906", "0.7134", "0.505", "0.7134" ], [ "0.5034", "0.714", "0.8906", "0.7134", "0.8906", "0.728", "0.5034", "0.729" ], [ "0.0825", "0.72", "0.4673", "0.72", "0.4673", "0.7344", "0.0825", "0.7344" ], [ "0.0801", "0.735", "0.4673", "0.735", "0.4673", "0.7485", "0.0801", "0.7485" ], [ "0.5034", "0.73", "0.892", "0.73", "0.892", "0.744", "0.5034", "0.744" ], [ "0.08417", "0.7505", "0.468", "0.7505", "0.468", "0.7646", "0.08417", "0.7646" ], [ "0.5034", "0.7456", "0.8906", "0.7456", "0.8906", "0.759", "0.5034", "0.759" ], [ "0.5024", "0.759", "0.831", "0.759", "0.831", "0.773", "0.5024", "0.773" ], [ "0.0825", "0.7666", "0.4673", "0.7666", "0.4673", "0.78", "0.0825", "0.78" ], [ "0.0825", "0.7827", "0.4656", "0.7827", "0.4656", "0.7935", "0.0825", "0.7935" ], [ "0.505", "0.778", "0.8906", "0.778", "0.8906", "0.7915", "0.505", "0.7915" ], [ "0.0801", "0.7954", "0.4673", "0.7944", "0.4673", "0.8096", "0.0801", "0.81" ], [ "0.5034", "0.7935", "0.8896", "0.7935", "0.8896", "0.8066", "0.5034", "0.8066" ], [ "0.0809", "0.811", "0.4673", "0.811", "0.4673", "0.8247", "0.0809", "0.8247" ], [ "0.5034", "0.808", "0.892", "0.808", "0.892", "0.8213", "0.5034", "0.8213" ], [ "0.0809", "0.8257", "0.4673", "0.8257", "0.4673", "0.839", "0.0809", "0.839" ], [ "0.505", "0.8237", "0.8906", "0.8237", "0.8906", "0.837", "0.505", "0.837" ], [ "0.0825", "0.8413", "0.4673", "0.8413", "0.4673", "0.855", "0.0825", "0.855" ], [ "0.505", "0.8384", "0.892", "0.8384", "0.892", "0.853", "0.505", "0.853" ], [ "0.0801", "0.857", "0.4673", "0.856", "0.4673", "0.869", "0.0801", "0.8706" ], [ "0.505", "0.855", "0.5864", "0.855", "0.5864", "0.866", "0.505", "0.866" ], [ "0.0809", "0.872", "0.468", "0.872", "0.468", "0.8853", "0.0809", "0.8853" ], [ "0.5034", "0.872", "0.8906", "0.872", "0.8906", "0.8853", "0.5034", "0.8853" ], [ "0.0809", "0.886", "0.468", "0.886", "0.468", "0.901", "0.0809", "0.901" ], [ "0.5024", "0.886", "0.8906", "0.886", "0.8906", "0.901", "0.5024", "0.901" ], [ "0.478", "0.924", "0.4944", "0.924", "0.4944", "0.938", "0.478", "0.938" ] ]

[ "Multi-scaleInput", "Reconstruction", "4x", "4x", "2x", "2x", "1x", "1x", "1x", "SatMAE", "Ours", "SatMAE", "Ours", "Figure 3. SatMAE++ reconstruction results at multi-scale level. Examples from fMoW-Sentinel dataset are shown here", "For illustration, we show the RGB channels of the multi-spectral data here.The images are", "econstructedatresolutionsof", "(H,W), (2H,2W), and (4H,4W),respectively.We observe that the proposed model provide better reconstruction results", "compared to SatMAE at resolution of (H, W) :", "微调：我们通过以端到端的方式加载预训练权重来微", "Table 3. Finetuning results on the validation set of fMoW-", "Sentinel dataset.", "* represents that we reproduce the re-", "sultsusingthepubliclyavailablecodebaseprovidedbythe", "调ViT-Large模型。初始学习率设置为le-3，单GPU", "使用批量大小8。我们将AdamW优化器与余弦调度器", "authors and pre-train the model on fMoW-Sentinel for 50", "起使用，并保持与5类似的增强和权重衰减配置", "epochs.", "我们在8个NVIDIAV100GPU上微调了50个时期的", "模型", "Backbone", "Topl Acc.", "讨论：我们在Tab中展示了我们方法的微调结果。2。", "Method", "ViT-Base", "50.45", "最先进的 ScaleMAE[21]显示了77的改进性能。9通", "MoCo-V3", "MoCo-V3+Group", "ViT-Base", "51.33", "过利用基于GSD的复杂位置编码和拉普拉斯解码器", "SatMAE * [5]", "ViT-Base", "60.01", "在SatMAE[5]上使用%。但是，我们的方法通过提供", "SatMAE++(Ours)", "ViT-Base", "62.69", "78的分数而优于最先进的ScaleMAE[21]。14%不使", "ScaleMAE *[21]", "ViT-Large", "42.21", "用复杂的基于GSD的位置编码。与基线方法[5]相比，", "SatMAE [5]", "ViT-Large", "61.48", "我们的方法提供了4的绝对增益。27%超过再现结果", "SatMAE * [5]", "ViT-Large", "61.61", "和0。3%超过了论文中报告的数字。", "ViT-Large", "63.23", "Pre-training and Finetuning on fMoW-", "SatMAE++(Ours)", "5.3.", "Sentinel", "预训练：在[5]之后，我们在图像大小为96×96且补", "样图像。将具有96×96空间分辨率的调整大小后的图", "像输人到模型中。该模型在8个NVIDIAV100GPU上", "丁大小为8的fMoW-Sentinel数据集上预训练ViT-", "Large模型。我们采用SatMAE+Group+IM策略对多", "进行了预训练，基本学习率为1e-4，批处理大小为8，", "光谱数据进行通道分组和独立掩蔽。Sentinel-2图像的", "持续50个时期。优化器、学习速率调度器、权重衰减、", "掩码比率的其余设置保持与[5]中列出的相同。", "所有13个波段都没有被利用，我们在预训练和微调期", "间丢弃了通道B1、B9和B10。我们遵循5的通道分", "微调：在微调阶段，我们加载ViT-Large模型的预训", "组方法，并分别创建（i）包含B2，B3，B4，B8通道", "练权重，并将其微调30个时期。我们使用2e-4的基本", "的组，（ii）由B5，B6，B7，B8A通道组成的组，以及", "学习率，96×96像素的输人大小和8的补丁大小。在", "(ii）B11，B12通道的组。选择的组具有相同的GSD", "[5]之后，我们利用AdamW优化器，余弦调度器和相", "分辨率。在预训练阶段，将图像的短边调整为384像", "同的数据增强。我们采用8个NVIDIAV100GPU对模", "素，并使用随机裁剪获得大小为384×384像素的图", "型进行微调", "像。然后，我们对384×384大小的图像进行插值，以", "讨论：我们报告了Tab中fMoW-Sentinel数据集的最新", "分别获得两个大小为192×192和96×96像素的下采", "比较。3。我们采用多尺度预训练的方法比SatMAE", "5" ]

{ "arxiv_id": "2403.05419", "language": "zh", "page_idx": 4 }

[ [ "0.12335", "0.113", "0.8486", "0.11426", "0.8486", "0.132", "0.12335", "0.1307" ], [ "0.2361", "0.1799", "0.7344", "0.1799", "0.7344", "0.1945", "0.2361", "0.1945" ], [ "0.08496", "0.1963", "0.1316", "0.1945", "0.1323", "0.2108", "0.0866", "0.2134" ], [ "0.2418", "0.2058", "0.3218", "0.2084", "0.3218", "0.226", "0.2418", "0.2235" ], [ "0.5024", "0.2102", "0.884", "0.2102", "0.884", "0.2247", "0.5024", "0.2247" ], [ "0.5024", "0.226", "0.8823", "0.226", "0.8823", "0.2405", "0.5024", "0.2405" ], [ "0.1217", "0.2361", "0.4412", "0.2361", "0.4412", "0.25", "0.1217", "0.25" ], [ "0.501", "0.2393", "0.8857", "0.2405", "0.8857", "0.2551", "0.501", "0.254" ], [ "0.1217", "0.252", "0.4412", "0.252", "0.4412", "0.2651", "0.1217", "0.2651" ], [ "0.501", "0.2556", "0.7974", "0.2556", "0.7974", "0.2695", "0.501", "0.2695" ], [ "0.1217", "0.2664", "0.4395", "0.2664", "0.4395", "0.2798", "0.1217", "0.2798" ], [ "0.1217", "0.2822", "0.4412", "0.2822", "0.4412", "0.2954", "0.1217", "0.2954" ], [ "0.5024", "0.2778", "0.8857", "0.2778", "0.8857", "0.2922", "0.5024", "0.2922" ], [ "0.1217", "0.2974", "0.4412", "0.2974", "0.4412", "0.3113", "0.1217", "0.3113" ], [ "0.5024", "0.2954", "0.884", "0.2954", "0.884", "0.3062", "0.5024", "0.3062" ], [ "0.1209", "0.3113", "0.4395", "0.3113", "0.4395", "0.3245", "0.1209", "0.3245" ], [ "0.5024", "0.3093", "0.8857", "0.3093", "0.8857", "0.3225", "0.5024", "0.3225" ], [ "0.1209", "0.3257", "0.4395", "0.3257", "0.4395", "0.339", "0.1209", "0.339" ], [ "0.501", "0.324", "0.884", "0.324", "0.884", "0.3372", "0.501", "0.3372" ], [ "0.1217", "0.3416", "0.4412", "0.3416", "0.4412", "0.356", "0.1217", "0.356" ], [ "0.501", "0.339", "0.714", "0.339", "0.714", "0.3528", "0.501", "0.3528" ], [ "0.1209", "0.3567", "0.4395", "0.3567", "0.4395", "0.3706", "0.1209", "0.3706" ], [ "0.5024", "0.361", "0.8857", "0.361", "0.8857", "0.375", "0.5024", "0.375" ], [ "0.1209", "0.3726", "0.2688", "0.3726", "0.2688", "0.3865", "0.1209", "0.3865" ], [ "0.2769", "0.3726", "0.4412", "0.3726", "0.4412", "0.3865", "0.2769", "0.3865" ], [ "0.5024", "0.377", "0.8823", "0.377", "0.8823", "0.3901", "0.5024", "0.3901" ], [ "0.1209", "0.387", "0.4412", "0.387", "0.4412", "0.401", "0.1209", "0.401" ], [ "0.5024", "0.3926", "0.884", "0.3926", "0.884", "0.406", "0.5024", "0.406" ], [ "0.1209", "0.4028", "0.4412", "0.4028", "0.4412", "0.4167", "0.1209", "0.4167" ], [ "0.5024", "0.4072", "0.8823", "0.4072", "0.8823", "0.4204", "0.5024", "0.4204" ], [ "0.1217", "0.4172", "0.4412", "0.4172", "0.4412", "0.4312", "0.1217", "0.4312" ], [ "0.5034", "0.4224", "0.884", "0.4224", "0.884", "0.4363", "0.5034", "0.4363" ], [ "0.1217", "0.433", "0.4395", "0.433", "0.4395", "0.4463", "0.1217", "0.4463" ], [ "0.5024", "0.4375", "0.8857", "0.4375", "0.8857", "0.4507", "0.5024", "0.4507" ], [ "0.11926", "0.4475", "0.4395", "0.4475", "0.4395", "0.461", "0.11926", "0.461" ], [ "0.5", "0.452", "0.8857", "0.4507", "0.8857", "0.4653", "0.5", "0.4666" ], [ "0.1217", "0.4634", "0.4412", "0.4634", "0.4412", "0.4766", "0.1217", "0.4766" ], [ "0.501", "0.4666", "0.8823", "0.4673", "0.8823", "0.4817", "0.501", "0.481" ], [ "0.1217", "0.4792", "0.4412", "0.4792", "0.4412", "0.4924", "0.1217", "0.4924" ], [ "0.5034", "0.483", "0.8823", "0.483", "0.8823", "0.4968", "0.5034", "0.4968" ], [ "0.1209", "0.4937", "0.4412", "0.4937", "0.4412", "0.507", "0.1209", "0.507" ], [ "0.5034", "0.5", "0.884", "0.5", "0.884", "0.5103", "0.5034", "0.5103" ], [ "0.1217", "0.5083", "0.4412", "0.5083", "0.4412", "0.5215", "0.1217", "0.5215" ], [ "0.5", "0.511", "0.694", "0.512", "0.694", "0.5264", "0.5", "0.526" ], [ "0.1209", "0.5234", "0.4395", "0.5234", "0.4395", "0.537", "0.1209", "0.537" ], [ "0.1217", "0.5386", "0.4412", "0.5386", "0.4412", "0.553", "0.1217", "0.553" ], [ "0.5024", "0.535", "0.884", "0.535", "0.884", "0.549", "0.5024", "0.549" ], [ "0.12335", "0.5537", "0.4412", "0.5537", "0.4412", "0.5674", "0.12335", "0.5674" ], [ "0.5024", "0.5503", "0.8823", "0.5503", "0.8823", "0.5645", "0.5024", "0.5645" ], [ "0.1217", "0.5684", "0.4412", "0.5684", "0.4412", "0.582", "0.1217", "0.582" ], [ "0.5156", "0.5664", "0.884", "0.5664", "0.884", "0.58", "0.5156", "0.58" ], [ "0.1217", "0.584", "0.4412", "0.584", "0.4412", "0.5977", "0.1217", "0.5977" ], [ "0.5024", "0.579", "0.8857", "0.58", "0.8857", "0.5947", "0.5024", "0.5933" ], [ "0.1209", "0.5986", "0.4395", "0.5986", "0.4395", "0.6123", "0.1209", "0.6123" ], [ "0.5024", "0.595", "0.8857", "0.595", "0.8857", "0.61", "0.5024", "0.61" ], [ "0.1209", "0.6143", "0.3513", "0.6143", "0.3513", "0.6274", "0.1209", "0.6274" ], [ "0.501", "0.6094", "0.8857", "0.61", "0.8857", "0.625", "0.501", "0.6235" ], [ "0.501", "0.6255", "0.8823", "0.625", "0.8823", "0.64", "0.501", "0.6416" ], [ "0.501", "0.64", "0.6724", "0.6416", "0.6724", "0.656", "0.501", "0.655" ], [ "0.08905", "0.654", "0.2222", "0.654", "0.2222", "0.6685", "0.08905", "0.6685" ], [ "0.5024", "0.6636", "0.884", "0.6636", "0.884", "0.678", "0.5024", "0.678" ], [ "0.0907", "0.677", "0.4739", "0.677", "0.4739", "0.6914", "0.0907", "0.6914" ], [ "0.5", "0.678", "0.8857", "0.6753", "0.8857", "0.6924", "0.5", "0.6943" ], [ "0.09235", "0.6924", "0.4722", "0.6924", "0.4722", "0.7056", "0.09235", "0.7056" ], [ "0.5024", "0.6934", "0.8857", "0.6934", "0.8857", "0.707", "0.5024", "0.707" ], [ "0.0907", "0.707", "0.4722", "0.707", "0.4722", "0.72", "0.0907", "0.72" ], [ "0.5024", "0.709", "0.8857", "0.709", "0.8857", "0.7236", "0.5024", "0.7236" ], [ "0.08905", "0.7217", "0.4722", "0.7217", "0.4722", "0.735", "0.08905", "0.735" ], [ "0.5024", "0.7246", "0.884", "0.7246", "0.884", "0.738", "0.5024", "0.738" ], [ "0.0907", "0.738", "0.4722", "0.738", "0.4722", "0.7485", "0.0907", "0.7485" ], [ "0.5024", "0.7393", "0.8857", "0.7393", "0.8857", "0.7524", "0.5024", "0.7524" ], [ "0.0907", "0.754", "0.4722", "0.754", "0.4722", "0.7646", "0.0907", "0.7646" ], [ "0.501", "0.7524", "0.8857", "0.755", "0.8857", "0.768", "0.501", "0.7666" ], [ "0.08826", "0.7666", "0.4722", "0.7676", "0.4722", "0.782", "0.08826", "0.781" ], [ "0.5034", "0.7695", "0.8813", "0.7695", "0.8813", "0.7827", "0.5034", "0.7827" ], [ "0.1111", "0.7905", "0.4714", "0.7905", "0.4714", "0.8037", "0.1111", "0.8037" ], [ "0.5024", "0.785", "0.884", "0.785", "0.884", "0.799", "0.5024", "0.799" ], [ "0.08826", "0.803", "0.4739", "0.8037", "0.4739", "0.8184", "0.08826", "0.8174" ], [ "0.501", "0.799", "0.8857", "0.8", "0.8857", "0.8145", "0.501", "0.813" ], [ "0.08905", "0.8164", "0.4714", "0.8164", "0.4714", "0.8306", "0.08905", "0.8306" ], [ "0.5024", "0.815", "0.8857", "0.815", "0.8857", "0.83", "0.5024", "0.83" ], [ "0.08826", "0.829", "0.4722", "0.83", "0.4722", "0.8433", "0.08826", "0.8423" ], [ "0.501", "0.829", "0.8857", "0.83", "0.8857", "0.8447", "0.501", "0.8433" ], [ "0.08826", "0.8423", "0.33", "0.8423", "0.33", "0.856", "0.08826", "0.856" ], [ "0.505", "0.8447", "0.8857", "0.8447", "0.8857", "0.8594", "0.505", "0.8594" ], [ "0.5034", "0.86", "0.8857", "0.86", "0.8857", "0.875", "0.5034", "0.875" ], [ "0.08826", "0.866", "0.4739", "0.8657", "0.4739", "0.88", "0.08826", "0.881" ], [ "0.08905", "0.88", "0.4739", "0.88", "0.4739", "0.893", "0.08905", "0.893" ], [ "0.501", "0.875", "0.6665", "0.875", "0.6665", "0.888", "0.501", "0.888" ], [ "0.08826", "0.89", "0.3489", "0.8926", "0.3489", "0.907", "0.08826", "0.905" ] ]

[ "Robustness and Accuracy Could Be Reconcilable by (Proper) Definition", "Tianyu Pang 12 Min Lin? Xiao Yang1 Jun Zhul Shuicheng Yan 2", "。3in", "Abstract", "止（Riceetal.,2020;Zhangetal.,2020），实例重新", "加权(Balaji etal.,2019;Zhang et al.,2021a)，并", "The trade-off between robustness and accu-", "利用额外的数据（Alayracetal.，2019；Carmonetal.，", "racy has been widely studied in the adver-", "2019;Hendrycksetal.,2019），仅举几例", "sarial literature. Although still controver", "sial, the prevailing view is that this trade-", "然而，其他发现显示相反。Stutzetal.（2019）和Yang", "off is inherent, either empirically or theoret-", "etal.（2020）认为，可以通过流形分析或局部Lipschitz", "ically. Thus, we dig for the origin of this", "函数来实现鲁棒性和准确性。Rozsaetal.（2016）和", "trade-off in adversarial training and find that", "Gilmeretal.（2018）还揭示了更好的泛化有助于在玩", "it may stem from the improperly defined ro-", "具和大规模数据集上的鲁棒性", "bust error, which imposes an inductive bias", "鉴于双方的论点，健壮性与准确性的权衡是否本质上", "of local invariance-", "an overcorrection to-", "存在很多争议。但是在考虑它的存在之前，让我们首", "wards 'smoothness. Given this, we advo-", "先就稳健模型应该如何表现达成共识一稳健性的定义。", "cate employing local equivariance to describe", "最受欢迎的是Madryetal.（2018），他根据鲁棒误差定", "the ideal behavior of a robust model, lead-", "义了鲁棒性，该误差被公式化为局部最大化的损失函", "ing to a self-consistent robust error named", "数。这个定义在对抗训练（AT）中被广泛采用（Shafahi", "SCORE. By definition, SCORE facilitates", "etal.,2019；Wongetal.，2020）。此外，还有一些其", "the reconciliation between robustness and ac-", "他的定义。例如，Szegedyetal.（2014）将鲁棒性定义", "curacy, while still handling the worst-case un-", "为翻转预测标签所需的最小扰动，这在几个对抗性攻", "certainty via robust optimization. By simply", "击研究中得到了应用（Moosavi-Dezfoolietal..2016:", "substituting KL divergence with variants of", "Carlini and Wagner, 2017)", "distance metrics, SCORE can be efficiently", "minimized. Empirically, our models achieve", "然后，我们尝试描述鲁棒性-准确性权衡如何源于先前", "top-rank performance on RobustBench under", "定义的鲁棒误差。回想一下，在监督学习中，我们有", "AutoAttack. Besides, SCORE provides in-", "个联合数据分布pa(a,y），以及一个用于标签y的", "structive insights for explaining the overfit-", "判别模型pe(y），该模型以输人为条件。在标准", "ting phenomenon and semantic input gradi", "设置中，我们通过最小化pa(yla）和pe(c）之间的", "ents observed on robust models.", "预期KL差异来获得准确的模型，i。e。，标准误差", "Rstandard(0)（见式（1))w。r。t。参数0。最优解0*", "满足po*(yr)=pa(yc)", "1.Introduction", "在对抗性设置中，Madryetal.（2018）建议最小化鲁棒", "对抗性鲁棒性和清洁精度之间的权衡已被广泛观察", "误差RMadry(0）（请参见公式E(3）)，以训练可靠的", "到E(Schmidt et al.,2018; Su et al.,2018; Zhang et al.,", "模型。与标准误差相比，鲁棒误差包含一个内部最大", "2019；Wangetal.，2020a;b）。在一些简单的情况下", "化问题，找到使pa(yla）和pe(ylr）之间的KL发散最", "这种权衡甚至被证明是可证明存在的（Tsipras et al.,", "大化的点EB(r)，其中B(α)是a周围的一组允许", "2019:Nakkiran.2019:Raghunathan et al..2020:Ja", "点。最小化鲁棒误差w。r。t。θ对局部不变性施加了感", "vanmardetal..2020Yuetal.,2021）。因此，已经", "应偏差：对于VrEB(r），鼓励pe(yr)等于pa(yr)", "提出了许多策略来缓解这种权衡，例如。g。，提前停", "。这种局部不变的偏差使得学习模型倾向于过度平滑", "1 Dept.of Comp.Sci. and Tech.,Institute for Al.", "正如在以前的工作中所观察到的（Stutzetal.，2020;", "BNRist Center, THBI Lab, Tsinghua-Bosch Joint Center", "Chenetal.,2021b）。因此，通常pe*（yla）pa(ylc）", "for ML, Tsinghua University.2 Sea AI Lab, Singapore .", "为mingRMadry(0）的最佳0*。InFig。1，我们观察", "Correspondence to: Jun Zhu< [email protected] >,", "到即使在玩具示例上，p*(yla）也不会收敛到pa(ylα）", "Shuicheng Yan < [email protected] >.", "。pa(yl）不是最优鲁棒模型的结论。r。t。pa(ylr）本", "身似乎违反直觉，表明RMadry(0）本质上不支持鲁棒", "Proceedings of the 39th International Conference on Ma-", "chine Learning, Baltimore, Maryland, USA, PMLR 162,", "性和准确性之间的协调", "2022. Copyright 2022 by the author(s)." ]

{ "arxiv_id": "2202.10103", "language": "zh", "page_idx": 0 }

[ [ "0.3594", "0.05936", "0.615", "0.05936", "0.615", "0.06946", "0.3594", "0.06946" ], [ "0.09314", "0.089", "0.1422", "0.0884", "0.1447", "0.178", "0.09644", "0.1787" ], [ "0.2998", "0.09784", "0.3162", "0.09784", "0.3162", "0.113", "0.2998", "0.113" ], [ "0.3015", "0.1067", "0.3218", "0.108", "0.3179", "0.1515", "0.2974", "0.1509" ], [ "0.1177", "0.11554", "0.125", "0.1206", "0.1177", "0.1256", "0.1111", "0.1206" ], [ "0.3284", "0.11237", "0.3381", "0.11237", "0.3381", "0.1332", "0.3284", "0.1332" ], [ "0.3308", "0.1332", "0.3367", "0.1332", "0.3367", "0.1414", "0.3308", "0.1414" ], [ "0.7344", "0.1364", "0.74", "0.1427", "0.7344", "0.1459", "0.7295", "0.1395" ], [ "0.4656", "0.1465", "0.4697", "0.1465", "0.4697", "0.1528", "0.4656", "0.1528" ], [ "0.7305", "0.1459", "0.74", "0.1459", "0.74", "0.159", "0.7305", "0.159" ], [ "0.3367", "0.1654", "0.3635", "0.1654", "0.3635", "0.1686", "0.3367", "0.1686" ], [ "0.4272", "0.1654", "0.4495", "0.1654", "0.4495", "0.1686", "0.4272", "0.1686" ], [ "0.607", "0.1654", "0.6357", "0.1654", "0.6357", "0.1686", "0.607", "0.1686" ], [ "0.705", "0.1654", "0.7173", "0.1654", "0.7173", "0.1686", "0.705", "0.1686" ], [ "0.7427", "0.1654", "0.8506", "0.1654", "0.8506", "0.1692", "0.7427", "0.1692" ], [ "0.1283", "0.1724", "0.1912", "0.1724", "0.1912", "0.1799", "0.1283", "0.1799" ], [ "0.1929", "0.1718", "0.2656", "0.1718", "0.2656", "0.1818", "0.1929", "0.1818" ], [ "0.1896", "0.1843", "0.1978", "0.1843", "0.1978", "0.1875", "0.1896", "0.1875" ], [ "0.3015", "0.1901", "0.3145", "0.1849", "0.3228", "0.1963", "0.3096", "0.202" ], [ "0.1356", "0.1976", "0.2427", "0.1976", "0.2427", "0.2084", "0.1356", "0.2084" ], [ "0.304", "0.1963", "0.3245", "0.1963", "0.3245", "0.2317", "0.304", "0.2317" ], [ "0.1299", "0.2058", "0.267", "0.2084", "0.267", "0.2267", "0.1299", "0.2247" ], [ "0.3308", "0.2153", "0.3367", "0.2153", "0.3367", "0.2203", "0.3308", "0.2203" ], [ "0.134", "0.2235", "0.2418", "0.2235", "0.2418", "0.2374", "0.134", "0.2374" ], [ "0.3057", "0.2317", "0.3186", "0.2317", "0.3186", "0.2424", "0.3057", "0.2424" ], [ "0.1323", "0.238", "0.2336", "0.2374", "0.2336", "0.2488", "0.1323", "0.25" ], [ "0.7236", "0.2329", "0.7427", "0.2299", "0.7466", "0.2474", "0.728", "0.2507" ], [ "0.415", "0.252", "0.4412", "0.252", "0.4412", "0.2588", "0.415", "0.2588" ], [ "0.5376", "0.252", "0.5825", "0.252", "0.5825", "0.2588", "0.5376", "0.2588" ], [ "0.7363", "0.252", "0.8105", "0.252", "0.8105", "0.2588", "0.7363", "0.2588" ], [ "0.8096", "0.252", "0.853", "0.252", "0.853", "0.2588", "0.8096", "0.2588" ], [ "0.6772", "0.254", "0.7197", "0.254", "0.7197", "0.2583", "0.6772", "0.2583" ], [ "0.4885", "0.2524", "0.5107", "0.2524", "0.5107", "0.2568", "0.4885", "0.2568" ], [ "0.3757", "0.2615", "0.415", "0.2615", "0.415", "0.2715", "0.3757", "0.2715" ], [ "0.511", "0.2615", "0.5513", "0.2615", "0.5513", "0.2715", "0.511", "0.2715" ], [ "0.6436", "0.2588", "0.687", "0.262", "0.6855", "0.2734", "0.642", "0.2708" ], [ "0.778", "0.2615", "0.8228", "0.2615", "0.8228", "0.2727", "0.778", "0.2727" ], [ "0.773", "0.2715", "0.83", "0.2715", "0.83", "0.2822", "0.773", "0.2822" ], [ "0.1209", "0.2632", "0.2688", "0.2632", "0.2688", "0.2778", "0.1209", "0.2778" ], [ "0.558", "0.2766", "0.638", "0.2766", "0.638", "0.2874", "0.558", "0.2874" ], [ "0.08826", "0.2998", "0.8857", "0.2986", "0.8857", "0.3132", "0.08826", "0.3145" ], [ "0.08826", "0.3145", "0.8857", "0.3145", "0.8857", "0.3276", "0.08826", "0.3276" ], [ "0.0907", "0.3276", "0.8857", "0.3276", "0.8857", "0.3416", "0.0907", "0.3416" ], [ "0.08826", "0.3403", "0.886", "0.3416", "0.886", "0.3547", "0.08826", "0.3542" ], [ "0.08905", "0.356", "0.886", "0.356", "0.886", "0.3694", "0.08905", "0.3694" ], [ "0.0866", "0.3687", "0.5786", "0.3694", "0.5786", "0.3838", "0.0866", "0.3833" ], [ "0.0907", "0.4216", "0.4739", "0.4216", "0.4739", "0.435", "0.0907", "0.435" ], [ "0.501", "0.4216", "0.8857", "0.4216", "0.8857", "0.435", "0.501", "0.435" ], [ "0.0907", "0.4363", "0.4739", "0.4363", "0.4739", "0.4495", "0.0907", "0.4495" ], [ "0.501", "0.4363", "0.884", "0.4363", "0.884", "0.4495", "0.501", "0.4495" ], [ "0.09235", "0.452", "0.4714", "0.452", "0.4714", "0.4653", "0.09235", "0.4653" ], [ "0.501", "0.4507", "0.884", "0.4507", "0.884", "0.4653", "0.501", "0.4653" ], [ "0.0907", "0.4666", "0.4739", "0.4666", "0.4739", "0.481", "0.0907", "0.481" ], [ "0.501", "0.4666", "0.884", "0.4666", "0.884", "0.4797", "0.501", "0.4797" ], [ "0.0907", "0.4817", "0.4739", "0.4817", "0.4739", "0.4956", "0.0907", "0.4956" ], [ "0.501", "0.481", "0.884", "0.481", "0.884", "0.4944", "0.501", "0.4944" ], [ "0.08826", "0.4944", "0.4739", "0.4968", "0.4739", "0.511", "0.08826", "0.509" ], [ "0.501", "0.4976", "0.5195", "0.4976", "0.5195", "0.5103", "0.501", "0.5103" ], [ "0.0907", "0.511", "0.4739", "0.511", "0.4739", "0.526", "0.0907", "0.526" ], [ "0.08826", "0.5264", "0.4739", "0.526", "0.4739", "0.5405", "0.08826", "0.5415" ], [ "0.5", "0.53", "0.7715", "0.53", "0.7715", "0.5444", "0.5", "0.5444" ], [ "0.0907", "0.5425", "0.4722", "0.5425", "0.4722", "0.556", "0.0907", "0.556" ], [ "0.0907", "0.5566", "0.464", "0.5566", "0.464", "0.571", "0.0907", "0.571" ], [ "0.5024", "0.5566", "0.8857", "0.5566", "0.8857", "0.571", "0.5024", "0.571" ], [ "0.5024", "0.5728", "0.884", "0.5728", "0.884", "0.586", "0.5024", "0.586" ], [ "0.0907", "0.58", "0.4739", "0.58", "0.4739", "0.5933", "0.0907", "0.5933" ], [ "0.5", "0.586", "0.8857", "0.587", "0.8857", "0.6016", "0.5", "0.601" ], [ "0.08905", "0.595", "0.4722", "0.595", "0.4722", "0.6094", "0.08905", "0.6094" ], [ "0.501", "0.603", "0.812", "0.603", "0.812", "0.616", "0.501", "0.616" ], [ "0.08905", "0.6094", "0.4714", "0.608", "0.4714", "0.6226", "0.08905", "0.6235" ], [ "0.08905", "0.625", "0.4722", "0.6235", "0.4722", "0.638", "0.08905", "0.6396" ], [ "0.5", "0.634", "0.634", "0.634", "0.634", "0.648", "0.5", "0.648" ], [ "0.0907", "0.64", "0.4722", "0.64", "0.4722", "0.654", "0.0907", "0.654" ], [ "0.08905", "0.655", "0.4722", "0.655", "0.4722", "0.6694", "0.08905", "0.6694" ], [ "0.5024", "0.6577", "0.8823", "0.6577", "0.8823", "0.672", "0.5024", "0.672" ], [ "0.0907", "0.6704", "0.4722", "0.6704", "0.4722", "0.6846", "0.0907", "0.6846" ], [ "0.5", "0.672", "0.884", "0.672", "0.884", "0.686", "0.5", "0.686" ], [ "0.09235", "0.685", "0.4722", "0.685", "0.4722", "0.6987", "0.09235", "0.6987" ], [ "0.517", "0.688", "0.8857", "0.688", "0.8857", "0.702", "0.517", "0.702" ], [ "0.09235", "0.7007", "0.4722", "0.7007", "0.4722", "0.714", "0.09235", "0.714" ], [ "0.5", "0.702", "0.7686", "0.7026", "0.7686", "0.7173", "0.5", "0.7163" ], [ "0.0907", "0.7163", "0.4722", "0.7163", "0.4722", "0.73", "0.0907", "0.73" ], [ "0.0866", "0.726", "0.4763", "0.728", "0.4763", "0.7485", "0.0866", "0.747" ], [ "0.08905", "0.744", "0.4739", "0.7456", "0.4739", "0.7603", "0.08905", "0.759" ], [ "0.5186", "0.7505", "0.884", "0.7505", "0.884", "0.765", "0.5186", "0.765" ], [ "0.08905", "0.761", "0.1569", "0.761", "0.1569", "0.776", "0.08905", "0.776" ], [ "0.0907", "0.7827", "0.4739", "0.7827", "0.4739", "0.797", "0.0907", "0.797" ], [ "0.0866", "0.7944", "0.4756", "0.7954", "0.4756", "0.8164", "0.0866", "0.815" ], [ "0.5024", "0.797", "0.8813", "0.797", "0.8813", "0.811", "0.5024", "0.811" ], [ "0.08826", "0.8125", "0.4722", "0.813", "0.4722", "0.8276", "0.08826", "0.827" ], [ "0.5", "0.8096", "0.886", "0.811", "0.886", "0.829", "0.5", "0.827" ], [ "0.0907", "0.829", "0.4722", "0.829", "0.4722", "0.8423", "0.0907", "0.8423" ], [ "0.5", "0.827", "0.8857", "0.8276", "0.8857", "0.8413", "0.5", "0.8403" ], [ "0.08905", "0.8433", "0.4722", "0.8433", "0.4722", "0.858", "0.08905", "0.858" ], [ "0.4983", "0.8413", "0.796", "0.8433", "0.796", "0.858", "0.4983", "0.856" ], [ "0.08905", "0.8594", "0.2484", "0.8594", "0.2484", "0.8726", "0.08905", "0.8726" ], [ "0.5215", "0.866", "0.886", "0.866", "0.886", "0.88", "0.5215", "0.88" ], [ "0.08905", "0.882", "0.4722", "0.882", "0.4722", "0.896", "0.08905", "0.896" ], [ "0.5", "0.88", "0.8857", "0.88", "0.8857", "0.893", "0.5", "0.893" ], [ "0.501", "0.893", "0.574", "0.893", "0.574", "0.905", "0.501", "0.905" ] ]

[ "Title Suppressed Due to ExcessiveSize", "", "?", "e", "", "688", "", "", "", "80", "", "6于", "?", "", "3 0123456", "543240", "12345678", "", "?", "Construction of pd(a.y):", "", "pa(y=0 =g,pa(y=1)=:", ".5", "pa(rtg = 0) ~ N(-1,4);", "R", "pa(rly = 1) ~ N(1, 1).", "8a", "456", "34567", "343210123", "B45613", "45678", "", "10 steps", "50steps", "100steps", "500steps", "(converged)", "pa(ylr)pe(ylr)", "Training process", "Figure 1.Toy demo on training pe(yl)with RMadry(0)and RscoRe(0),respectively.We consider1-d binary classification", "where E R and y E (0,1] , and the perturbed set B(r) = (r'r' - | ≤ 1) . Totally 60,000 training input-label pairs", "are sampled from the data distribution pa(r, y) , as detailed in the left panel. The model pe(yl) is a shallow MLP with", "two hidden layers and a sigmoid output. We use Adam optimizer and train for 500 steps to ensure convergence. The", "small gap between the converged pe(y|) and pd(ylr) in the SCORE case is caused by limited model capacity. Note that", "in toy demo we assume pa(yl) is accessible, which is not true in practice.", "正式解决上述不一致促使我们正确地重新定义鲁棒错", "取代KL发散的有效性，这些度量是根据对SCORE", "误。具体来说，我们将局部不变性的归纳偏差替换为", "的分析开发的。我们改进了AutoAttackE（Croceand", "局部等方差，即。e。，VaEB(r)，pe(yc'）被鼓励点", "Hein，2020）下最先进的AT方法，并在RobustBenchF", "明智地坚持pa(ylc），导致自一致鲁棒误差（SCORE)", "(Croceetal.,2020)上的CIFAR-10和CIFAR-100的", "的定义为RsCORE(O）（见等式（4）)。与鲁棒误差相比，", "排行榜上使用1MDDPM生成的数据实现了一流的性", "SCORE在其内部问题中找到了使pa(yla'）和pe(yr'）", "能", "之间的KL发散最大化的点a。SCORE将最优解", "pg(ya）与pa(ylr)(i。e。，自洽性，如图所示。1),", "2. Self-Consistent Robust Error", "同时在有限样本情况下保持鲁棒优化（Wald，1945）的", "范式，如图。2。更多细节可以在Sec中找到。2", "根据Madryetal.（2018），鲁棒性连接到鲁棒性误差", "的特定定义，这意味着更健壮的模型是实现较低鲁棒", "如何优化SCORE？请注意，我们只能在玩具箱中对", "性误差的模型。1在本节中，我们首先重新审视以前的", "pa(yα）进行闭式访问。在实践中，通过现成的一阶优", "鲁棒性定义，然后提出一种自洽的鲁棒误差", "化器最小化SCORE(e。g。，SGD或Adam）需要估计", "V，logpa(yla）（请参见附录EB.2）。这个任务可以分", "2.1. Preliminaries", "解为分数匹配任务（Vincent，2011）分别估计数据分", "数Vlogpa(α）和Vlogpa(ly）。然而，我们最初的", "在监督学习中，训练集（(r,y'))-i由Ni组成。i。d", "实验表明，估计的数据分数具有很高的方差，因此在", "从联合数据分布pa(a,y）中采样的输人标签对（ri,yi）", "判别学习过程中很好地采用它们是不容易的。幸运的", "令pe(yr）为由参数化的判别模型。它被训练为", "是，如Sec中所述。E3，我们发现，通过将KL散度", "通过最小化标准误差来匹配pa(ya）：", "替换为满足距离公理（对称性，三角不等式）的任何度", "量D，我们可以得出基于距离的SCORERBcORE(O)", "的上界和下界，并在不知道√logpa(ya）的情况下将", "Rstandard(0) =Epa() [KL (pd(y|c)pe(yl)) ,(1)", "其最小化", "在秒内。4，我们通过Pinsker不等式弥合了基于距", "离的SCORERScORE(O)与以前使用的基于KL的目", "其中KL(PIQ)表示两个分布P和Q之间的KL散度", "标（如RMadry（の））之间的差距。这种联系激励我们", "由于数据分布pd独立于，因此最小化Rstandard(の)", "对稳健模型的一些众所周知的现象提出有启发性的解", "W。r。t。等价于最小化交叉摘损失（Friedmanetal.，", "释，例如。g。，过拟合（Riceetal.，2020）和语义梯", "2001），其中最优解为pg*(ya）=pa(ya)", "度 (Ilyas et al., 2019)", "1Here, robustness refers to its differentiable surrogate", "在秒内。5，我们验证了用基于距离的度量（及其变体）", "form: In Sec. 5.1 we will discuss its 0-1 definition used for", "evaluation." ]

{ "arxiv_id": "2202.10103", "language": "zh", "page_idx": 1 }

[ [ "0.2854", "0.09265", "0.7153", "0.0944", "0.7153", "0.11224", "0.2854", "0.10986" ], [ "0.2091", "0.1437", "0.5283", "0.1437", "0.5283", "0.158", "0.2091", "0.158" ], [ "0.519", "0.1437", "0.791", "0.1437", "0.791", "0.158", "0.519", "0.158" ], [ "0.2627", "0.1644", "0.7407", "0.1644", "0.7407", "0.1787", "0.2627", "0.1787" ], [ "0.3132", "0.1812", "0.6987", "0.1812", "0.6987", "0.1948", "0.3132", "0.1948" ], [ "0.2627", "0.2542", "0.3418", "0.2542", "0.3418", "0.2683", "0.2627", "0.2683" ], [ "0.5146", "0.2546", "0.8774", "0.2546", "0.8774", "0.2683", "0.5146", "0.2683" ], [ "0.513", "0.2708", "0.881", "0.2708", "0.881", "0.284", "0.513", "0.284" ], [ "0.147", "0.2778", "0.4575", "0.2778", "0.4575", "0.2915", "0.147", "0.2915" ], [ "0.513", "0.2861", "0.881", "0.2861", "0.881", "0.2998", "0.513", "0.2998" ], [ "0.147", "0.2935", "0.4575", "0.2935", "0.4575", "0.3071", "0.147", "0.3071" ], [ "0.147", "0.3071", "0.4575", "0.3071", "0.4575", "0.3206", "0.147", "0.3206" ], [ "0.513", "0.3027", "0.642", "0.3027", "0.642", "0.3164", "0.513", "0.3164" ], [ "0.1453", "0.3218", "0.4592", "0.3218", "0.4592", "0.335", "0.1453", "0.335" ], [ "0.531", "0.3218", "0.881", "0.3218", "0.881", "0.335", "0.531", "0.335" ], [ "0.1453", "0.335", "0.4575", "0.3362", "0.4575", "0.3499", "0.1453", "0.3486" ], [ "0.513", "0.3374", "0.882", "0.3374", "0.882", "0.3508", "0.513", "0.3508" ], [ "0.147", "0.3508", "0.4575", "0.3508", "0.4575", "0.364", "0.147", "0.364" ], [ "0.513", "0.3528", "0.882", "0.3528", "0.882", "0.3665", "0.513", "0.3665" ], [ "0.147", "0.364", "0.4592", "0.364", "0.4592", "0.3777", "0.147", "0.3777" ], [ "0.5146", "0.3694", "0.8774", "0.3694", "0.8774", "0.3801", "0.5146", "0.3801" ], [ "0.1453", "0.379", "0.4575", "0.379", "0.4575", "0.3926", "0.1453", "0.3926" ], [ "0.513", "0.3855", "0.882", "0.386", "0.882", "0.3997", "0.513", "0.3984" ], [ "0.147", "0.3926", "0.4575", "0.3926", "0.4575", "0.4062", "0.147", "0.4062" ], [ "0.5156", "0.405", "0.565", "0.405", "0.565", "0.412", "0.5156", "0.412" ], [ "0.5728", "0.402", "0.881", "0.402", "0.881", "0.415", "0.5728", "0.415" ], [ "0.5635", "0.4062", "0.5767", "0.4062", "0.5767", "0.4104", "0.5635", "0.4104" ], [ "0.1427", "0.4062", "0.455", "0.4067", "0.455", "0.4204", "0.1427", "0.4192" ], [ "0.147", "0.4216", "0.4575", "0.4216", "0.4575", "0.4353", "0.147", "0.4353" ], [ "0.5146", "0.4187", "0.881", "0.4187", "0.881", "0.4316", "0.5146", "0.4316" ], [ "0.1453", "0.4353", "0.4592", "0.4353", "0.4592", "0.449", "0.1453", "0.449" ], [ "0.513", "0.434", "0.879", "0.434", "0.879", "0.4478", "0.513", "0.4478" ], [ "0.1453", "0.4495", "0.4568", "0.4495", "0.4568", "0.4631", "0.1453", "0.4631" ], [ "0.513", "0.449", "0.881", "0.4495", "0.881", "0.4644", "0.513", "0.4631" ], [ "0.147", "0.4631", "0.4575", "0.4631", "0.4575", "0.478", "0.147", "0.478" ], [ "0.5146", "0.466", "0.881", "0.466", "0.881", "0.4797", "0.5146", "0.4797" ], [ "0.1453", "0.478", "0.4575", "0.478", "0.4575", "0.4912", "0.1453", "0.4912" ], [ "0.5156", "0.4822", "0.8774", "0.4822", "0.8774", "0.4954", "0.5156", "0.4954" ], [ "0.1486", "0.4934", "0.455", "0.4934", "0.455", "0.5044", "0.1486", "0.5044" ], [ "0.513", "0.4988", "0.881", "0.4988", "0.881", "0.5127", "0.513", "0.5127" ], [ "0.1453", "0.506", "0.4568", "0.506", "0.4568", "0.5195", "0.1453", "0.5195" ], [ "0.1453", "0.5195", "0.4575", "0.5215", "0.4575", "0.535", "0.1453", "0.533" ], [ "0.5156", "0.514", "0.881", "0.514", "0.881", "0.5293", "0.5156", "0.5293" ], [ "0.147", "0.535", "0.4575", "0.535", "0.4575", "0.549", "0.147", "0.549" ], [ "0.5156", "0.531", "0.881", "0.531", "0.881", "0.5444", "0.5156", "0.5444" ], [ "0.1486", "0.5503", "0.4568", "0.5503", "0.4568", "0.561", "0.1486", "0.561" ], [ "0.51", "0.5464", "0.839", "0.546", "0.839", "0.56", "0.51", "0.561" ], [ "0.1453", "0.5625", "0.4575", "0.564", "0.4575", "0.5776", "0.1453", "0.576" ], [ "0.531", "0.5635", "0.879", "0.564", "0.879", "0.579", "0.531", "0.5776" ], [ "0.1453", "0.5776", "0.455", "0.5776", "0.455", "0.5913", "0.1453", "0.5913" ], [ "0.511", "0.579", "0.882", "0.58", "0.882", "0.594", "0.511", "0.5933" ], [ "0.147", "0.5913", "0.4575", "0.5913", "0.4575", "0.605", "0.147", "0.605" ], [ "0.511", "0.5957", "0.881", "0.5967", "0.881", "0.611", "0.511", "0.61" ], [ "0.147", "0.6055", "0.4575", "0.6055", "0.4575", "0.619", "0.147", "0.619" ], [ "0.513", "0.6133", "0.881", "0.6133", "0.881", "0.627", "0.513", "0.627" ], [ "0.1486", "0.6216", "0.4568", "0.6216", "0.4568", "0.632", "0.1486", "0.632" ], [ "0.147", "0.633", "0.4568", "0.633", "0.4568", "0.6465", "0.147", "0.6465" ], [ "0.513", "0.6274", "0.879", "0.6274", "0.879", "0.641", "0.513", "0.641" ], [ "0.1486", "0.6494", "0.4575", "0.6494", "0.4575", "0.66", "0.1486", "0.66" ], [ "0.5146", "0.6455", "0.882", "0.6455", "0.882", "0.659", "0.5146", "0.659" ], [ "0.1503", "0.6646", "0.4575", "0.6646", "0.4575", "0.675", "0.1503", "0.675" ], [ "0.513", "0.661", "0.881", "0.661", "0.881", "0.675", "0.513", "0.675" ], [ "0.1453", "0.677", "0.4575", "0.677", "0.4575", "0.6904", "0.1453", "0.6904" ], [ "0.511", "0.677", "0.881", "0.6777", "0.881", "0.6914", "0.511", "0.6904" ], [ "0.147", "0.6914", "0.4568", "0.6914", "0.4568", "0.705", "0.147", "0.705" ], [ "0.5146", "0.6934", "0.881", "0.6934", "0.881", "0.707", "0.5146", "0.707" ], [ "0.147", "0.705", "0.4592", "0.705", "0.4592", "0.7188", "0.147", "0.7188" ], [ "0.513", "0.7104", "0.881", "0.7104", "0.881", "0.7236", "0.513", "0.7236" ], [ "0.1453", "0.7197", "0.4575", "0.7197", "0.4575", "0.733", "0.1453", "0.733" ], [ "0.513", "0.7256", "0.881", "0.7256", "0.881", "0.7393", "0.513", "0.7393" ], [ "0.147", "0.734", "0.4575", "0.734", "0.4575", "0.7476", "0.147", "0.7476" ], [ "0.147", "0.7476", "0.4575", "0.7476", "0.4575", "0.761", "0.147", "0.761" ], [ "0.5156", "0.7437", "0.879", "0.7437", "0.879", "0.7544", "0.5156", "0.7544" ], [ "0.147", "0.7627", "0.4575", "0.7627", "0.4575", "0.7754", "0.147", "0.7754" ], [ "0.51", "0.7583", "0.879", "0.7573", "0.879", "0.7715", "0.51", "0.7725" ], [ "0.147", "0.7754", "0.4575", "0.7754", "0.4575", "0.789", "0.147", "0.789" ], [ "0.531", "0.777", "0.881", "0.777", "0.881", "0.7905", "0.531", "0.7905" ], [ "0.1427", "0.788", "0.4417", "0.7905", "0.4417", "0.8047", "0.1427", "0.8027" ], [ "0.511", "0.7915", "0.879", "0.7915", "0.879", "0.8047", "0.511", "0.8047" ], [ "0.511", "0.807", "0.881", "0.808", "0.881", "0.8223", "0.511", "0.822" ], [ "0.11584", "0.8164", "0.1511", "0.8164", "0.1511", "0.8276", "0.11584", "0.8276" ], [ "0.1486", "0.814", "0.3997", "0.8154", "0.3997", "0.83", "0.1486", "0.829" ], [ "0.5146", "0.8247", "0.881", "0.8247", "0.881", "0.8384", "0.5146", "0.8384" ], [ "0.12177", "0.843", "0.4844", "0.843", "0.4844", "0.8525", "0.12177", "0.8525" ], [ "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", "0.881", "0.857", "0.881", "0.8706", "0.513", "0.8706" ], [ "0.11755", "0.8726", "0.4878", "0.8726", "0.4878", "0.8857", "0.11755", "0.8857" ], [ "0.513", "0.8726", "0.881", "0.8726", "0.881", "0.887", "0.513", "0.887" ], [ "0.1184", "0.8896", "0.486", "0.8896", "0.486", "0.903", "0.1184", "0.903" ], [ "0.513", "0.8896", "0.881", "0.8896", "0.881", "0.903", "0.513", "0.903" ], [ "0.11755", "0.905", "0.486", "0.905", "0.486", "0.9185", "0.11755", "0.9185" ], [ "0.513", "0.905", "0.881", "0.905", "0.881", "0.9185", "0.513", "0.9185" ] ]

[ "与目标语言就绪任务适配器的跨语言传输", "MarinelaParovielAlanAnsell", "Ivan VulielAnnaKorhonen", "1Language Technology Lab, TAL, University of Cambridge", "(mp939,aja63,iv250,alk23 } @cam.ac.uk", "Abstract", "的多语言表示空间使（Piresetal.,2019;Ketal", "2020）的跨语言传输成为可能，也就是说，以", "Adapters have emerged as a modular and", "完全缺乏这些任务的培训数据的语言以合理的", "parameter-efficient approach to (zero-shot)", "cross-lingual transfer. The established MAD", "准确度执行任务", "X framework employs separate language and", "通常通过对高资源源语言中特定于任务的", "task adapters which can be arbitrarily com-", "数据的预训练MMT进行微调来执行零镜头跨", "bined to perform the transfer of any task to", "语言传输（i。e。，通常是英语)，然后直接应", "any target language. Subsequently, BAD-", "用它以目标语言进行任务预测。在标准设置", "X, an extension of the MAD-X framework,", "中，模型关于目标语言的知识仅在预训练阶段", "achieves improved transfer at the cost of MAD", "Artetx", "etal.,2020)期间获得。为了提高传输", "P", "X's modularity by creating ‘bilingual' adapters", "specific to the source-target language pair.", "性能，可以在任务微调之前对目标语言（Ponti", "In this work, we aim to take the best of", "etal.，2020;Pfeifferetal.，2020b）中的未标记数", "both worlds by (i) fine-tuning task adapters", "据进行微调。尽管如此，在这种情况下，目标", "adapted to the target language(s) (so-called", "语言的性能低于源语言，这种差异被称为跨语", "‘target language-ready(TLR) adapters) to", "言迁移间隙(Huetal.，2020）。至关重要的是", "maintain high transfer performance,but(ii)", "对于最需要这种转移的语言，（Joshietal.,2020)", "without sacrificing the highly modular design", "of MAD-X. The main idea of ‘target language-", "：i的转移差距往往会增加。e。，对于低资源", "ready' adapters is to resolve the training-vs", "目标语言，以及在类型上更远离源语言的语言", "inference discrepancy of MAD-X:thetask", "(例如。g。,English)(Lauscher etal.,2020)", "adapter “sees' the target language adapter for", "适配器(Rebuffi et al.,2017;Houlsby et al.,", "the very first time during inference, and thus", "2019）已成为帮助零触发跨语言传输（Pfeiffer", "might not be fully compatible with it.We", "et al., 2020b; Ustun et al., 2022a; Ansell et al.,", "address this mismatch by exposing the task", "2021；Parovicetal.，2022）的重要方法。它们", "adapterto thetargetlanguageadapterduring", "training, and empirically validate several vari", "提供了几个好处：（i）为目标语言提供额外的", "ants of the idea: in the simplestform, we al-", "表示能力；（i）与全模型微调相比，参数效", "ernate between using the source and target", "率更高，因为它们允许大型MMT的参数保持", "language adapters during task adapter train-", "不变，从而保留了MMT在预训练期间获得的", "ing, which can be generalized to cycling over", "多语言知识。它们还（ii）在学习和存储知识", "any set of language adapters. We evaluate dif-", "(Pfeifferetal.，2020a）的不同方面方面提供了模", "ferent TLR-based transfer configurations with", "块化：此属性使它们能够以有利的方式组合", "varying degrees of generality across a suite of", "standard cross-lingual benchmarks, and find", "以实现更好的性能，并且以前经过微调的模块", "that the most general (and thus most modular)", "(e。g。，语言适配器）在不同的应用程序中重用", "configuration consistently outperforms MAD", "已建立的基于适配器的跨语言传输框架", "X and BAD-X on most tasks and languages.", "MAD-X（Pfeifferetal，2020b）训练单独的语言", "适配器（LA）和任务适配器（TA），然后可以", "1", "Introduction and Motivation", "将其任意组合以将任何任务传输到任何语言。", "多语言NLP的最新进展主要由大型多语言转", "尽管具有高度模块化的设计，主要来自专用的", "换器模型（MMT）驱动，例如mBERT（Devlin", "每语言和每任务适配器，MAD-X的TA缺乏对", "et al., 2019), XLM-R (Conneau et al., 2020) 和", "感兴趣的目标语言的“适应性”：i。e。，它的助", "mT5（Xueetal.，2021），这些模型已在100多", "教完全与目标语言无关。更准确地说，在任务", "种语言的未标记数据上进行了训练。他们共享", "微调期间，MAD-XTA仅暴露于源语言LA，并" ]

{ "arxiv_id": "2306.02767", "language": "zh", "page_idx": 1 }

[ [ "0.513", "0.0861", "0.881", "0.0861", "0.881", "0.0998", "0.513", "0.0998" ], [ "0.3208", "0.0944", "0.3284", "0.0944", "0.3284", "0.0986", "0.3208", "0.0986" ], [ "0.287", "0.10095", "0.3594", "0.1027", "0.3594", "0.114", "0.287", "0.11224" ], [ "0.5146", "0.1027", "0.881", "0.1027", "0.881", "0.1164", "0.5146", "0.1164" ], [ "0.3225", "0.1164", "0.3267", "0.1164", "0.3257", "0.1223", "0.3208", "0.1217" ], [ "0.511", "0.1164", "0.882", "0.1152", "0.882", "0.133", "0.511", "0.1342" ], [ "0.3005", "0.1289", "0.3535", "0.1289", "0.3535", "0.1401", "0.3005", "0.1401" ], [ "0.5146", "0.1342", "0.879", "0.1342", "0.879", "0.1484", "0.5146", "0.1484" ], [ "0.2494", "0.1538", "0.403", "0.1538", "0.403", "0.1687", "0.2494", "0.1687" ], [ "0.5156", "0.1509", "0.881", "0.1509", "0.881", "0.1644", "0.5156", "0.1644" ], [ "0.525", "0.1644", "0.881", "0.165", "0.881", "0.1829", "0.525", "0.1818" ], [ "0.2544", "0.1799", "0.3477", "0.1799", "0.3477", "0.1912", "0.2544", "0.1912" ], [ "0.5146", "0.1829", "0.881", "0.1829", "0.881", "0.1965", "0.5146", "0.1965" ], [ "0.2544", "0.1996", "0.351", "0.1996", "0.351", "0.2114", "0.2544", "0.2114" ], [ "0.511", "0.199", "0.879", "0.199", "0.879", "0.212", "0.511", "0.212" ], [ "0.287", "0.2144", "0.3105", "0.2144", "0.3105", "0.2214", "0.287", "0.2214" ], [ "0.5156", "0.2162", "0.858", "0.2162", "0.858", "0.2268", "0.5156", "0.2268" ], [ "0.2527", "0.2257", "0.351", "0.2257", "0.351", "0.2369", "0.2527", "0.2369" ], [ "0.5146", "0.231", "0.881", "0.231", "0.881", "0.2446", "0.5146", "0.2446" ], [ "0.5146", "0.2465", "0.879", "0.2465", "0.879", "0.2612", "0.5146", "0.2612" ], [ "0.2896", "0.2583", "0.3594", "0.2583", "0.3594", "0.2683", "0.2896", "0.2683" ], [ "0.513", "0.2625", "0.879", "0.2625", "0.879", "0.2756", "0.513", "0.2756" ], [ "0.3074", "0.284", "0.339", "0.284", "0.339", "0.2957", "0.3074", "0.2957" ], [ "0.5146", "0.279", "0.882", "0.279", "0.882", "0.2935", "0.5146", "0.2935" ], [ "0.513", "0.2935", "0.879", "0.2944", "0.879", "0.3088", "0.513", "0.308" ], [ "0.288", "0.3071", "0.357", "0.3088", "0.357", "0.3206", "0.288", "0.3184" ], [ "0.513", "0.31", "0.879", "0.3113", "0.879", "0.326", "0.513", "0.3247" ], [ "0.3198", "0.3247", "0.3267", "0.3247", "0.3267", "0.3308", "0.3198", "0.3308" ], [ "0.513", "0.3267", "0.708", "0.3267", "0.708", "0.3403", "0.513", "0.3403" ], [ "0.2688", "0.335", "0.378", "0.335", "0.378", "0.3467", "0.2688", "0.3467" ], [ "0.3032", "0.3508", "0.345", "0.3508", "0.345", "0.3582", "0.3032", "0.3582" ], [ "0.511", "0.3582", "0.55", "0.3582", "0.55", "0.3694", "0.511", "0.3694" ], [ "0.5425", "0.3557", "0.721", "0.3582", "0.721", "0.3723", "0.5425", "0.3706" ], [ "0.5146", "0.3843", "0.881", "0.3843", "0.881", "0.398", "0.5146", "0.398" ], [ "0.11755", "0.3914", "0.4878", "0.3914", "0.4878", "0.405", "0.11755", "0.405" ], [ "0.513", "0.3997", "0.879", "0.3997", "0.879", "0.4133", "0.513", "0.4133" ], [ "0.1184", "0.4067", "0.486", "0.4067", "0.486", "0.4204", "0.1184", "0.4204" ], [ "0.513", "0.415", "0.881", "0.4146", "0.881", "0.4287", "0.513", "0.43" ], [ "0.1184", "0.4216", "0.486", "0.4216", "0.486", "0.4353", "0.1184", "0.4353" ], [ "0.1184", "0.4353", "0.4878", "0.4353", "0.4878", "0.449", "0.1184", "0.449" ], [ "0.513", "0.4316", "0.881", "0.4316", "0.881", "0.4453", "0.513", "0.4453" ], [ "0.11755", "0.449", "0.4895", "0.449", "0.4895", "0.462", "0.11755", "0.462" ], [ "0.5146", "0.449", "0.8774", "0.449", "0.8774", "0.462", "0.5146", "0.462" ], [ "0.11755", "0.4631", "0.4878", "0.4631", "0.4878", "0.4768", "0.11755", "0.4768" ], [ "0.511", "0.4631", "0.882", "0.462", "0.882", "0.4768", "0.511", "0.478" ], [ "0.11584", "0.478", "0.4878", "0.4768", "0.4878", "0.4905", "0.11584", "0.4912" ], [ "0.511", "0.4797", "0.879", "0.4797", "0.879", "0.4934", "0.511", "0.4934" ], [ "0.11755", "0.4912", "0.4878", "0.4922", "0.4878", "0.506", "0.11755", "0.505" ], [ "0.513", "0.4954", "0.881", "0.4954", "0.881", "0.5103", "0.513", "0.5103" ], [ "0.1184", "0.506", "0.486", "0.506", "0.486", "0.5195", "0.1184", "0.5195" ], [ "0.513", "0.511", "0.882", "0.511", "0.882", "0.5254", "0.513", "0.5254" ], [ "0.11755", "0.521", "0.486", "0.521", "0.486", "0.5337", "0.11755", "0.5337" ], [ "0.513", "0.528", "0.879", "0.528", "0.879", "0.5415", "0.513", "0.5415" ], [ "0.1184", "0.5337", "0.486", "0.533", "0.486", "0.5464", "0.1184", "0.5474" ], [ "0.511", "0.542", "0.879", "0.5435", "0.879", "0.558", "0.511", "0.557" ], [ "0.1184", "0.549", "0.4878", "0.549", "0.4878", "0.5625", "0.1184", "0.5625" ], [ "0.1184", "0.564", "0.486", "0.564", "0.486", "0.5747", "0.1184", "0.5747" ], [ "0.513", "0.56", "0.881", "0.56", "0.881", "0.5737", "0.513", "0.5737" ], [ "0.12006", "0.579", "0.486", "0.579", "0.486", "0.589", "0.12006", "0.589" ], [ "0.513", "0.5767", "0.879", "0.5767", "0.879", "0.5903", "0.513", "0.5903" ], [ "0.1184", "0.5913", "0.4634", "0.5913", "0.4634", "0.605", "0.1184", "0.605" ], [ "0.513", "0.5933", "0.881", "0.5933", "0.881", "0.607", "0.513", "0.607" ], [ "0.5146", "0.6094", "0.881", "0.6094", "0.881", "0.6226", "0.5146", "0.6226" ], [ "0.5146", "0.6245", "0.882", "0.6245", "0.882", "0.638", "0.5146", "0.638" ], [ "0.1368", "0.6357", "0.471", "0.6357", "0.471", "0.6494", "0.1368", "0.6494" ], [ "0.511", "0.6396", "0.881", "0.64", "0.881", "0.655", "0.511", "0.654" ], [ "0.11755", "0.651", "0.4836", "0.651", "0.4836", "0.6646", "0.11755", "0.6646" ], [ "0.5156", "0.6562", "0.876", "0.6562", "0.876", "0.6694", "0.5156", "0.6694" ], [ "0.11584", "0.6675", "0.4878", "0.6685", "0.4878", "0.683", "0.11584", "0.6816" ], [ "0.5146", "0.6733", "0.881", "0.6733", "0.881", "0.687", "0.5146", "0.687" ], [ "0.11755", "0.684", "0.4844", "0.683", "0.4844", "0.6978", "0.11755", "0.6987" ], [ "0.5156", "0.6895", "0.881", "0.6895", "0.881", "0.703", "0.5156", "0.703" ], [ "0.11755", "0.709", "0.4836", "0.709", "0.4836", "0.7227", "0.11755", "0.7227" ], [ "0.511", "0.7036", "0.882", "0.705", "0.882", "0.7197", "0.511", "0.7188" ], [ "0.11755", "0.7246", "0.482", "0.7246", "0.482", "0.7383", "0.11755", "0.7383" ], [ "0.5146", "0.7217", "0.7363", "0.7217", "0.7363", "0.7354", "0.5146", "0.7354" ], [ "0.1184", "0.741", "0.4844", "0.741", "0.4844", "0.755", "0.1184", "0.755" ], [ "0.513", "0.742", "0.879", "0.7437", "0.879", "0.7583", "0.513", "0.7573" ], [ "0.1184", "0.7573", "0.4844", "0.7573", "0.4844", "0.771", "0.1184", "0.771" ], [ "0.513", "0.7603", "0.881", "0.7603", "0.881", "0.774", "0.513", "0.774" ], [ "0.1184", "0.774", "0.4844", "0.774", "0.4844", "0.7876", "0.1184", "0.7876" ], [ "0.517", "0.7754", "0.881", "0.7754", "0.881", "0.7905", "0.517", "0.7905" ], [ "0.12006", "0.789", "0.4844", "0.789", "0.4844", "0.8027", "0.12006", "0.8027" ], [ "0.5146", "0.792", "0.881", "0.792", "0.881", "0.8057", "0.5146", "0.8057" ], [ "0.1184", "0.8057", "0.486", "0.8057", "0.486", "0.8193", "0.1184", "0.8193" ], [ "0.513", "0.808", "0.881", "0.808", "0.881", "0.822", "0.513", "0.822" ], [ "0.1184", "0.8223", "0.4358", "0.8223", "0.4358", "0.836", "0.1184", "0.836" ], [ "0.511", "0.8223", "0.881", "0.8237", "0.881", "0.8384", "0.511", "0.8374" ], [ "0.1351", "0.8403", "0.4836", "0.8403", "0.4836", "0.854", "0.1351", "0.854" ], [ "0.513", "0.8403", "0.881", "0.8403", "0.881", "0.854", "0.513", "0.854" ], [ "0.1184", "0.857", "0.4878", "0.857", "0.4878", "0.8706", "0.1184", "0.8706" ], [ "0.5146", "0.858", "0.586", "0.858", "0.586", "0.869", "0.5146", "0.869" ], [ "0.1184", "0.8726", "0.4844", "0.8726", "0.4844", "0.8857", "0.1184", "0.8857" ], [ "0.534", "0.8813", "0.881", "0.8813", "0.881", "0.8945", "0.534", "0.8945" ], [ "0.11755", "0.887", "0.486", "0.888", "0.486", "0.903", "0.11755", "0.902" ], [ "0.513", "0.8936", "0.881", "0.8936", "0.881", "0.9067", "0.513", "0.9067" ], [ "0.1184", "0.905", "0.486", "0.905", "0.486", "0.9185", "0.1184", "0.9185" ], [ "0.511", "0.9062", "0.575", "0.9077", "0.575", "0.9194", "0.511", "0.9175" ] ]

[ "体。首先，所谓的TASK-MULTITLR变体在源", "T", "Add&Norm", "语言和可用于所考虑任务的所有目标语言集上", "", "运行（e。g。，MasakhaNER数据集中表示的所", "NLITA", "有语言）。其次，ALL-MULTITLR变体将源语言", "Choose LA: step % (K+1)", "与多个任务的数据集中的所有目标语言结合在", "一起（例如。g。，以MasakhaNER表示的所有", "O:English LA", "语言，以AmericasNLI表示的所有语言，等等。", "1:Target1LA", ")；稍后请参见83。这些变体提高了模块化和", "...", "参数效率，并且在每个任务上都与MAD-X", "K:TargetKLA", "样具有模块化和参数效率：需要一个TA来处", "理到任何目标语言的传输。同时，与MAD-X不", "Add&Norm", "同，它们可以接触到它们将用于的多种目标语", "FFN", "言所产生的表示形式。在微调时处理多个LA", "可能会使TA整体上更加健壮：多语言可能会", "Add&Norm", "充当正则化，迫使TA专注于更通用的跨语言", "A", "功能 (Ansell et al., 2021)", "Multi-Head Attentior", "个个1", "3", "Experimental Setup", "评估任务和语言。我们在一套标准的跨语言传", "Figure 1: A general multilingual task adapter (TA) tar-", "输基准上全面评估我们的TLR适配器框架。它", "get language-ready (TLR) module at one MMT layer,", "们涵盖四个不同的任务系列（NER，DP，NLI", "showing the language adapters (LAs) for English as the", "source language and K target languages along with the", "和QA），共有六个不同的数据集和35种不", "NLI TA. The TA is trained by cycling over the K + 1", "同的目标语言，涵盖了低资源和高资源语言", "LAs associated with the K + 1 languages. For a given", "的类型和地理上不同的语言样本。对于NER，", "stepnumber, onlythe LA step% (K+1)is switched", "我们使用MasakhaNER数据集（Adelanietal.", "on and the forward pass goes through that LA. Setting", "2021），其中包含来自非洲大陆的10种低资源", "K = O results in the original MAD-X setup, where only", "语言。3对于DP，我们使用通用依赖项2。7", "the source LA is switched on, while a bilingual TLR", "(Zeman et al.,2020)，并从BAD-x(Parovic et al..", "variant is given by K = 1.Setting K = 1 and re", "2022)继承了10种类型多样的低资源目标语言", "moving the English LA formulates the TARGET -only", "TLR variant. See $2.2 for the descriptions of all the", "集。对于NLI，我们依赖于AmericasNLI数据", "variants. The same adapter configuration(s), but with", "集（Ebrahimietal.，2022），其中包含来自美洲", "different parameters, are added at each MMT layer.", "的10种低资源语言，以及来自XNLI（Conneau", "etal.，2018）的语言子集。最后，对于QA，我们", "使用XQuAD（Artetxe et al.,2020)和TyDiQA-", "此过程在TA的微调阶段立即向TA公开", "GoldP(Clarketal.2020）中的语言子集。选择", "种目标语言（双语TLRTA）或多种目标语言", "了XNLI，XQuAD和TyDiQA-GoldP的子集", "（多语言TLRTA），从而使其更好地装备（i。e。，", "以将（i）低资源语言（Joshietal.,2020）与（i）", "准备好）用于推理阶段，其中TA与单个LLA", "专用于（i。e。，‘MAD-X’)LA很容易获得。附", "TLR变体。虽然BILINGUALTA微调自然来自", "录A中的表5提供了所有任务、数据集和语", "BAD-X，并且似乎适合在固定的L。和Lt对之", "言及其语言代码的完整概述", "间进行传输，但最好仅在LtLA之上训练TA。", "基础MMT。我们使用mBERT报告所有任务", "这样的仅TARGET的TLRTA对于那些LA已", "的结果，并在104种语言的Wikipedia上进行", "经在足够的语料库上接受过训练的高资源语言", "了预训练(Devlinetal.,2019）。先前的工作建", "可能特别有效，以至于将它们与L，配对是有", "议将mBERT作为真正的低资源语言（Pfeiffer", "害的。对于也远离L。或缺乏足够的词汇重叠", "etal.,2021b；Anselletal.,2021）的性能更好的", "的资源较高的Lt-S，这尤其可以检测到", "MMT。为了验证我们的TLR适配器的健壮性", "TARGET和BILINGUALTLRTA需要为每个", "我们还将XLM-R（Conneauetal.，2020）用于任", "感兴趣的Lt训练专用TA，这使得它们的计算", "务的子集", "效率低于MAD-X，并且它们总体上引人了更", "3We exclude Amharic from our experiments as it uses a", "多参数。使用MULTILINGUALTLRTA可减轻此", "script not supported by mBERT, resulting in 9 NER target", "开销。我们考虑MULTILINGUALTA的两个变", "languages" ]

{ "arxiv_id": "2306.02767", "language": "zh", "page_idx": 3 }

[ [ "0.2477", "0.0873", "0.2805", "0.0873", "0.2805", "0.0986", "0.2477", "0.0986" ], [ "0.3198", "0.0873", "0.3435", "0.0873", "0.3435", "0.0998", "0.3198", "0.0998" ], [ "0.3855", "0.0873", "0.413", "0.0873", "0.413", "0.0998", "0.3855", "0.0998" ], [ "0.444", "0.0873", "0.471", "0.0873", "0.471", "0.0998", "0.444", "0.0998" ], [ "0.614", "0.0944", "0.6323", "0.0944", "0.6323", "0.10156", "0.614", "0.10156" ], [ "0.7524", "0.0944", "0.7886", "0.0944", "0.7886", "0.10156", "0.7524", "0.10156" ], [ "0.1276", "0.10394", "0.189", "0.10394", "0.189", "0.114", "0.1276", "0.114" ], [ "0.2585", "0.10394", "0.272", "0.10394", "0.272", "0.114", "0.2585", "0.114" ], [ "0.3914", "0.10394", "0.4072", "0.10394", "0.4072", "0.114", "0.3914", "0.114" ], [ "0.451", "0.10394", "0.465", "0.10394", "0.465", "0.114", "0.451", "0.114" ], [ "0.763", "0.1069", "0.818", "0.1069", "0.818", "0.11755", "0.763", "0.11755" ], [ "0.2568", "0.1134", "0.2761", "0.1134", "0.2761", "0.1271", "0.2568", "0.1271" ], [ "0.3257", "0.1051", "0.339", "0.1051", "0.339", "0.114", "0.3257", "0.114" ], [ "0.3225", "0.1134", "0.3418", "0.1134", "0.3418", "0.1271", "0.3225", "0.1271" ], [ "0.392", "0.114", "0.4055", "0.114", "0.4055", "0.1247", "0.392", "0.1247" ], [ "0.4475", "0.1134", "0.4685", "0.1134", "0.4685", "0.1259", "0.4475", "0.1259" ], [ "0.763", "0.1152", "0.8228", "0.1152", "0.8228", "0.1259", "0.763", "0.1259" ], [ "0.3704", "0.1259", "0.4207", "0.1259", "0.4207", "0.1372", "0.3704", "0.1372" ], [ "0.1276", "0.1152", "0.1704", "0.1152", "0.1704", "0.1259", "0.1276", "0.1259" ], [ "0.44", "0.1259", "0.471", "0.1259", "0.471", "0.1372", "0.44", "0.1372" ], [ "0.764", "0.1259", "0.831", "0.1259", "0.831", "0.133", "0.764", "0.133" ], [ "0.2393", "0.1271", "0.288", "0.1271", "0.288", "0.1372", "0.2393", "0.1372" ], [ "0.1293", "0.1271", "0.2065", "0.1271", "0.2065", "0.1372", "0.1293", "0.1372" ], [ "0.3015", "0.1271", "0.3535", "0.1271", "0.3535", "0.1372", "0.3015", "0.1372" ], [ "0.5474", "0.1354", "0.553", "0.1354", "0.553", "0.1414", "0.5474", "0.1414" ], [ "0.7656", "0.1318", "0.839", "0.1318", "0.839", "0.1426", "0.7656", "0.1426" ], [ "0.2527", "0.136", "0.2778", "0.136", "0.2778", "0.1484", "0.2527", "0.1484" ], [ "0.1276", "0.136", "0.2258", "0.1383", "0.2258", "0.1484", "0.1276", "0.1467" ], [ "0.318", "0.136", "0.345", "0.136", "0.345", "0.1478", "0.318", "0.1478" ], [ "0.388", "0.1383", "0.4106", "0.1383", "0.4106", "0.1484", "0.388", "0.1484" ], [ "0.4434", "0.136", "0.4702", "0.136", "0.4702", "0.1484", "0.4434", "0.1484" ], [ "0.2544", "0.1467", "0.2761", "0.1467", "0.2761", "0.1592", "0.2544", "0.1592" ], [ "0.315", "0.1467", "0.3477", "0.1467", "0.3477", "0.1592", "0.315", "0.1592" ], [ "0.769", "0.1414", "0.831", "0.1414", "0.831", "0.152", "0.769", "0.152" ], [ "0.1276", "0.1478", "0.1948", "0.1478", "0.1948", "0.1592", "0.1276", "0.1592" ], [ "0.3855", "0.1478", "0.4114", "0.1478", "0.4114", "0.1592", "0.3855", "0.1592" ], [ "0.444", "0.1478", "0.4668", "0.1478", "0.4668", "0.1603", "0.444", "0.1603" ], [ "0.531", "0.1561", "0.5576", "0.1561", "0.5576", "0.1787", "0.531", "0.1787" ], [ "0.147", "0.1729", "0.4517", "0.1733", "0.4517", "0.1882", "0.147", "0.187" ], [ "0.5474", "0.1818", "0.553", "0.1818", "0.553", "0.187", "0.5474", "0.187" ], [ "0.1184", "0.2131", "0.4844", "0.2131", "0.4844", "0.2268", "0.1184", "0.2268" ], [ "0.11755", "0.228", "0.4844", "0.2286", "0.4844", "0.2423", "0.11755", "0.2411" ], [ "0.581", "0.2351", "0.601", "0.2351", "0.601", "0.2423", "0.581", "0.2423" ], [ "0.629", "0.2363", "0.825", "0.2363", "0.825", "0.2434", "0.629", "0.2434" ], [ "0.1184", "0.2434", "0.4844", "0.2446", "0.4844", "0.2588", "0.1184", "0.2583" ], [ "0.693", "0.2434", "0.711", "0.2434", "0.711", "0.2505", "0.693", "0.2505" ], [ "0.11755", "0.2612", "0.4836", "0.2612", "0.4836", "0.275", "0.11755", "0.275" ], [ "0.1184", "0.2756", "0.486", "0.2766", "0.486", "0.2915", "0.1184", "0.2903" ], [ "0.511", "0.2778", "0.879", "0.2778", "0.879", "0.2915", "0.511", "0.2915" ], [ "0.1184", "0.2935", "0.4844", "0.2935", "0.4844", "0.3071", "0.1184", "0.3071" ], [ "0.5146", "0.2922", "0.879", "0.2922", "0.879", "0.3027", "0.5146", "0.3027" ], [ "0.1184", "0.31", "0.4844", "0.31", "0.4844", "0.3237", "0.1184", "0.3237" ], [ "0.511", "0.306", "0.882", "0.306", "0.882", "0.3196", "0.511", "0.3196" ], [ "0.513", "0.3206", "0.8286", "0.3206", "0.8286", "0.3342", "0.513", "0.3342" ], [ "0.12006", "0.3267", "0.4844", "0.3267", "0.4844", "0.3403", "0.12006", "0.3403" ], [ "0.1184", "0.3416", "0.4844", "0.3416", "0.4844", "0.3557", "0.1184", "0.3557" ], [ "0.1184", "0.3582", "0.4878", "0.3582", "0.4878", "0.3718", "0.1184", "0.3718" ], [ "0.513", "0.361", "0.879", "0.361", "0.879", "0.3748", "0.513", "0.3748" ], [ "0.12006", "0.3748", "0.4844", "0.3748", "0.4844", "0.3884", "0.12006", "0.3884" ], [ "0.513", "0.3777", "0.86", "0.3777", "0.86", "0.3914", "0.513", "0.3914" ], [ "0.11755", "0.3901", "0.4844", "0.3901", "0.4844", "0.4038", "0.11755", "0.4038" ], [ "0.11584", "0.405", "0.4844", "0.4062", "0.4844", "0.4204", "0.11584", "0.4192" ], [ "0.5264", "0.408", "0.5483", "0.408", "0.5483", "0.415", "0.5264", "0.415" ], [ "0.544", "0.405", "0.744", "0.405", "0.744", "0.4187", "0.544", "0.4187" ], [ "0.1184", "0.4229", "0.419", "0.4229", "0.419", "0.4358", "0.1184", "0.4358" ], [ "0.51", "0.4092", "0.5264", "0.4038", "0.531", "0.412", "0.5146", "0.4175" ], [ "0.5146", "0.43", "0.879", "0.4312", "0.879", "0.4453", "0.5146", "0.4448" ], [ "0.12006", "0.4453", "0.486", "0.4453", "0.486", "0.459", "0.12006", "0.459" ], [ "0.5146", "0.4478", "0.881", "0.4478", "0.881", "0.4614", "0.5146", "0.4614" ], [ "0.1184", "0.4614", "0.4844", "0.4614", "0.4844", "0.4744", "0.1184", "0.4744" ], [ "0.511", "0.4631", "0.881", "0.4631", "0.881", "0.4768", "0.511", "0.4768" ], [ "0.1184", "0.478", "0.4878", "0.478", "0.4878", "0.4912", "0.1184", "0.4912" ], [ "0.511", "0.4797", "0.879", "0.4797", "0.879", "0.4934", "0.511", "0.4934" ], [ "0.12006", "0.4934", "0.4878", "0.4934", "0.4878", "0.5073", "0.12006", "0.5073" ], [ "0.513", "0.4963", "0.882", "0.4963", "0.882", "0.5103", "0.513", "0.5103" ], [ "0.11755", "0.5103", "0.4844", "0.5103", "0.4844", "0.524", "0.11755", "0.524" ], [ "0.51", "0.5127", "0.881", "0.5127", "0.881", "0.5254", "0.51", "0.5254" ], [ "0.11755", "0.5254", "0.4878", "0.525", "0.4878", "0.539", "0.11755", "0.5405" ], [ "0.511", "0.528", "0.881", "0.528", "0.881", "0.5415", "0.511", "0.5415" ], [ "0.12006", "0.542", "0.486", "0.542", "0.486", "0.5557", "0.12006", "0.5557" ], [ "0.5156", "0.5435", "0.881", "0.5435", "0.881", "0.557", "0.5156", "0.557" ], [ "0.11755", "0.558", "0.486", "0.557", "0.486", "0.572", "0.11755", "0.5723" ], [ "0.513", "0.56", "0.881", "0.56", "0.881", "0.5737", "0.513", "0.5737" ], [ "0.12006", "0.5747", "0.486", "0.5747", "0.486", "0.5884", "0.12006", "0.5884" ], [ "0.511", "0.5747", "0.881", "0.576", "0.881", "0.5903", "0.511", "0.589" ], [ "0.11755", "0.589", "0.4836", "0.589", "0.4836", "0.6025", "0.11755", "0.6025" ], [ "0.513", "0.5913", "0.879", "0.5913", "0.879", "0.605", "0.513", "0.605" ], [ "0.12006", "0.607", "0.4878", "0.607", "0.4878", "0.6206", "0.12006", "0.6206" ], [ "0.513", "0.6094", "0.876", "0.6094", "0.876", "0.619", "0.513", "0.619" ], [ "0.1184", "0.6226", "0.4844", "0.6226", "0.4844", "0.6357", "0.1184", "0.6357" ], [ "0.511", "0.6235", "0.8774", "0.6235", "0.8774", "0.637", "0.511", "0.637" ], [ "0.1184", "0.6396", "0.486", "0.6396", "0.486", "0.653", "0.1184", "0.653" ], [ "0.5146", "0.641", "0.879", "0.641", "0.879", "0.655", "0.5146", "0.655" ], [ "0.1184", "0.655", "0.4685", "0.655", "0.4685", "0.6685", "0.1184", "0.6685" ], [ "0.5146", "0.6567", "0.881", "0.6567", "0.881", "0.6704", "0.5146", "0.6704" ], [ "0.11755", "0.677", "0.4836", "0.677", "0.4836", "0.6904", "0.11755", "0.6904" ], [ "0.513", "0.6733", "0.534", "0.6733", "0.534", "0.686", "0.513", "0.686" ], [ "0.12006", "0.6934", "0.486", "0.6934", "0.486", "0.707", "0.12006", "0.707" ], [ "0.511", "0.6934", "0.879", "0.6934", "0.879", "0.707", "0.511", "0.707" ], [ "0.12006", "0.7104", "0.4844", "0.7104", "0.4844", "0.7236", "0.12006", "0.7236" ], [ "0.513", "0.7104", "0.8774", "0.7104", "0.8774", "0.7236", "0.513", "0.7236" ], [ "0.12006", "0.727", "0.4878", "0.727", "0.4878", "0.7407", "0.12006", "0.7407" ], [ "0.513", "0.728", "0.881", "0.728", "0.881", "0.741", "0.513", "0.741" ], [ "0.11755", "0.741", "0.4844", "0.742", "0.4844", "0.7573", "0.11755", "0.756" ], [ "0.513", "0.7437", "0.879", "0.7437", "0.879", "0.7583", "0.513", "0.7583" ], [ "0.11755", "0.7573", "0.4844", "0.7583", "0.4844", "0.7725", "0.11755", "0.7715" ], [ "0.511", "0.7583", "0.881", "0.759", "0.881", "0.774", "0.511", "0.7725" ], [ "0.1184", "0.775", "0.486", "0.775", "0.486", "0.788", "0.1184", "0.788" ], [ "0.511", "0.775", "0.879", "0.775", "0.879", "0.788", "0.511", "0.788" ], [ "0.1184", "0.7915", "0.486", "0.7915", "0.486", "0.8047", "0.1184", "0.8047" ], [ "0.513", "0.792", "0.879", "0.792", "0.879", "0.8057", "0.513", "0.8057" ], [ "0.11755", "0.807", "0.4878", "0.807", "0.4878", "0.821", "0.11755", "0.821" ], [ "0.513", "0.808", "0.8774", "0.808", "0.8774", "0.822", "0.513", "0.822" ], [ "0.1184", "0.8223", "0.4878", "0.8223", "0.4878", "0.836", "0.1184", "0.836" ], [ "0.51", "0.8237", "0.881", "0.8237", "0.881", "0.8374", "0.51", "0.8374" ], [ "0.1184", "0.8384", "0.486", "0.8384", "0.486", "0.8525", "0.1184", "0.8525" ], [ "0.51", "0.8403", "0.879", "0.8403", "0.879", "0.854", "0.51", "0.854" ], [ "0.11755", "0.854", "0.4878", "0.8525", "0.4878", "0.8677", "0.11755", "0.868" ], [ "0.513", "0.8555", "0.882", "0.8555", "0.882", "0.869", "0.513", "0.869" ], [ "0.531", "0.8716", "0.882", "0.8716", "0.882", "0.8857", "0.531", "0.8857" ], [ "0.1351", "0.879", "0.486", "0.8813", "0.486", "0.8955", "0.1351", "0.8936" ], [ "0.1184", "0.8936", "0.4878", "0.8936", "0.4878", "0.9067", "0.1184", "0.9067" ], [ "0.511", "0.888", "0.882", "0.887", "0.882", "0.902", "0.511", "0.903" ], [ "0.11755", "0.905", "0.3005", "0.905", "0.3005", "0.9185", "0.11755", "0.9185" ], [ "0.511", "0.905", "0.881", "0.905", "0.881", "0.9185", "0.511", "0.9185" ] ]

[ "NER", "DP", "NLI", "QA", "Ave", "models", "BatchSize", "8", "32", "16", "MAD-X", "10 ", "8", "10", "5", "15", "TARGET", "2·10-5", "Epochs", "10-4", "BILINGUAL", "5.10-5", "Learning Rate", "5.10-5", "50", "TASK-MULTI", "250", "EvalFreq.(steps)", "250", "625", "625", "F1", "LAS", "ALL-MULTI", "Eval Metric", "Acc", "F1", "350", "Table 1: Hyperparameters for different tasks.", "", "语言适配器。我们训练最少100个时期或", "100.000个步骤的LA，批处理大小为8，学", "NER", "DPAmericasNLIXNLIXQuADTyDiQA", "习率为5.10-5，最大序列长度为256。4对于", "asks", "低资源语言，我们每1,000步评估一次LA，对", "于高资源语言，每5000步评估一次LA，并选", "Figure 2: The average scores of MAD-X, TARGET", "择产生最低困惑的LA，并在保留的单语数据", "BILINGUAL,TASK-MULTIand ALL-MULTIvariants", "的5%上进行评估（1%用于高资源语言）。对", "on NER (F1), DP (LAS), AmericasNLI (acc), XNLI", "(acc), XQuAD (FI) and TyDiQA (F1) datasets.", "于BAD-X基线，我们直接使用（Parovicetal.,", "2022）中的双语LA。在Pfeifferetal.(2020b）之", "后，适配器缩减因子（i。e。，MMT的隐藏大", "体，除了英语LA之外，我们还对来自所有任", "小与适配器的瓶颈大小之间的比率）对于所有", "务和数据集的所有评估语言的LA进行循环", "LA为2。对于MAD-XLA，我们使用高效的", "Pfeiffer适配器配置（Pfeifferetal.，2020a）和可", "", "Results and Discussion", "逆适配器，而BAD-XLA不包括它们。", "4", "主要结果。mBERT对所有任务和所有语言的", "任务适配器。我们通过将TA堆叠在相应LA", "主要结果如表2所示，平均值如图2所示。", "的顶部来微调TA（请参见图1）。在微调期间，", "XLM-R的其他结果见附录B。作为一个普遍", "MMT的参数和所有LA的参数都被冻结。与先", "的趋势，我们观察到所有建议的TLR变体在所", "前的工作(Pfeifferetal.,2020b)(i。e。，d=48", "有任务中的大多数目标语言上都优于MAD-X", ")，并且像LA一样，它们使用Pfeiffer配置。", "。除了在所有任务上达到更高的平均值之外，", "表1中列出了不同任务的超参数，这些参数", "TLR框架的最佳每任务变体在以下目标语言", "也是从以前的工作中借用的。此外，我们在训", "上超过了MAD-X：9/9（NER），10/10(DP）", "练QATA时使用提前停止4(i。e。，当F1得", "10/10(AmericasNLI), 6/6 (XNLI), 4/4(XQuAD)", "分在连续四个评估周期内没有增加时，我们", "和5/5（TyDiQA）目标语言。我们还证明，在", "停止训练）。我们使用英语SOuADvl。1个用", "两个任务（DP，AmericasNLI）上，我们可以", "于TyDiQA-GoldP的训练数据（Rajpurkaretal.,", "轻松获得BAD-XLA的模块化程度要低得多的", "2016），因为（i）它比TyDiQA的本机训练集", "BAD-X实现了收益。总之，来自TableE2的全", "要大得多，并且（ii）我们在初步实验中观察", "面结果集证实了TLR适配器在一系列（类型", "到目标语言的性能高于TyDiQA的训练数据", "多样）目标语言和数据集中的有效性和多功能", "传输设置：详细信息。在我们所有的迁移实验", "性", "中，源语言L。都固定为英语，我们评估了82.2", "跨任务和TLR变体的结果细分。在NER和DP", "中描述的不同变体。对于MAD-X基线，我们", "上，我们观察到非常相似的结果趋势。重要的", "依赖于它的’MAD-Xv2。O'变体，将适配器放", "是，最模块化的ALL-MULTI变体提供了最高", "在Transformer的最后一层，已发现可以提高", "的整体性能：e。g。，它达到了F1的平均分", "整个(Pfeifferetal.,2021b）板的传输性能。对于", "数69。86NER任务中的%，同时比MAD-X高", "TASK-MULTITLR变体，以及使用英语LA，我", "出1。平均9个%和所有9种目标语言。DP", "们使用该特定数据集中所有评估语言的LA对", "任务中还显示了该变体的明显收益。TARGET", "TA进行微调。例如，对于DP，它跨越10种", "和BILINGUAL变体在大多数语言中也获得了", "语言，而对于NLI，我们用AmericasNLI的10", "收益，其中BILINGUAL在这两种语言中更强。", "种语言微调一个单独的TASK-MULTITLR，另", "然而，与ALL-MULTI相比，它们的总体效用", "一个用于XNLI语言。对于ALL-MULTITLR变", "较低，因为它们的性能较低，模块化程度较低", "在AmericasNLI上，所有TLR变体都比", "4For some low-resource languages with small corpora 100", "epochs leads to under-training, so the minimum number of", "MAD-X显示出可观的收益，从而实现了E5-", "training steps is set to 30,000.", "6%更高的平均准确性。它们在所有10种目标" ]

{ "arxiv_id": "2306.02767", "language": "zh", "page_idx": 4 }

[ [ "0.3333", "0.0968", "0.669", "0.0956", "0.669", "0.1134", "0.3333", "0.114" ], [ "0.157", "0.136", "0.843", "0.1342", "0.843", "0.152", "0.157", "0.1538" ], [ "0.2603", "0.1526", "0.7446", "0.1526", "0.7446", "0.1663", "0.2603", "0.1663" ], [ "0.2032", "0.1633", "0.797", "0.165", "0.797", "0.1829", "0.2032", "0.1812" ], [ "0.1604", "0.1799", "0.8423", "0.1812", "0.8423", "0.1953", "0.1604", "0.1948" ], [ "0.3728", "0.1953", "0.6323", "0.1953", "0.6323", "0.209", "0.3728", "0.209" ], [ "0.236", "0.2102", "0.7734", "0.2102", "0.7734", "0.2239", "0.236", "0.2239" ], [ "0.4702", "0.2434", "0.532", "0.2434", "0.532", "0.2546", "0.4702", "0.2546" ], [ "0.1184", "0.2583", "0.8833", "0.2583", "0.8833", "0.2715", "0.1184", "0.2715" ], [ "0.12006", "0.2708", "0.882", "0.2708", "0.882", "0.284", "0.12006", "0.284" ], [ "0.12006", "0.2852", "0.882", "0.2852", "0.882", "0.2957", "0.12006", "0.2957" ], [ "0.12006", "0.2964", "0.8833", "0.2964", "0.8833", "0.31", "0.12006", "0.31" ], [ "0.12006", "0.3088", "0.8833", "0.3088", "0.8833", "0.3225", "0.12006", "0.3225" ], [ "0.1184", "0.3218", "0.8833", "0.3218", "0.8833", "0.3362", "0.1184", "0.3362" ], [ "0.12006", "0.3342", "0.882", "0.3342", "0.882", "0.3474", "0.12006", "0.3474" ], [ "0.1184", "0.3486", "0.8833", "0.3486", "0.8833", "0.3623", "0.1184", "0.3623" ], [ "0.12006", "0.3623", "0.882", "0.3623", "0.882", "0.3723", "0.12006", "0.3723" ], [ "0.1184", "0.3735", "0.3418", "0.3748", "0.3418", "0.3884", "0.1184", "0.3872" ], [ "0.1184", "0.401", "0.7095", "0.401", "0.7095", "0.4146", "0.1184", "0.4146" ], [ "0.2217", "0.4246", "0.378", "0.4258", "0.378", "0.4407", "0.2217", "0.4395" ], [ "0.538", "0.4312", "0.561", "0.4312", "0.561", "0.4382", "0.538", "0.4382" ], [ "0.777", "0.4275", "0.827", "0.4312", "0.825", "0.4395", "0.775", "0.4358" ], [ "0.802", "0.4412", "0.8154", "0.4412", "0.8154", "0.4453", "0.802", "0.4453" ], [ "0.8154", "0.4412", "0.8555", "0.4412", "0.8555", "0.4453", "0.8154", "0.4453" ], [ "0.7363", "0.4507", "0.7715", "0.4507", "0.7715", "0.4536", "0.7363", "0.4536" ], [ "0.771", "0.449", "0.8833", "0.449", "0.8833", "0.456", "0.771", "0.456" ], [ "0.1234", "0.456", "0.486", "0.456", "0.486", "0.466", "0.1234", "0.466" ], [ "0.742", "0.459", "0.7617", "0.459", "0.7617", "0.462", "0.742", "0.462" ], [ "0.777", "0.4573", "0.876", "0.4573", "0.876", "0.4644", "0.777", "0.4644" ], [ "0.12006", "0.4685", "0.4878", "0.4685", "0.4878", "0.4822", "0.12006", "0.4822" ], [ "0.727", "0.4656", "0.781", "0.4656", "0.781", "0.4727", "0.727", "0.4727" ], [ "0.623", "0.4797", "0.7197", "0.4797", "0.7197", "0.4868", "0.623", "0.4868" ], [ "0.76", "0.481", "0.797", "0.481", "0.797", "0.485", "0.76", "0.485" ], [ "0.7773", "0.4912", "0.8364", "0.4912", "0.8364", "0.4988", "0.7773", "0.4988" ], [ "0.12006", "0.4827", "0.486", "0.4827", "0.486", "0.4963", "0.12006", "0.4963" ], [ "0.12006", "0.4976", "0.4878", "0.4976", "0.4878", "0.511", "0.12006", "0.511" ], [ "0.538", "0.4934", "0.559", "0.4934", "0.559", "0.502", "0.538", "0.502" ], [ "0.727", "0.4988", "0.894", "0.4988", "0.894", "0.509", "0.727", "0.509" ], [ "0.12006", "0.511", "0.4878", "0.511", "0.4878", "0.5254", "0.12006", "0.5254" ], [ "0.729", "0.509", "0.889", "0.509", "0.889", "0.5166", "0.729", "0.5166" ], [ "0.721", "0.5156", "0.7676", "0.5156", "0.7676", "0.5254", "0.721", "0.5254" ], [ "0.575", "0.5215", "0.592", "0.5215", "0.592", "0.5254", "0.575", "0.5254" ], [ "0.708", "0.5215", "0.7153", "0.5166", "0.721", "0.521", "0.714", "0.5254" ], [ "0.7886", "0.525", "0.8867", "0.525", "0.8867", "0.535", "0.7886", "0.535" ], [ "0.6055", "0.528", "0.6187", "0.528", "0.6187", "0.532", "0.6055", "0.532" ], [ "0.11755", "0.524", "0.4878", "0.525", "0.4878", "0.539", "0.11755", "0.538" ], [ "0.7246", "0.531", "0.869", "0.533", "0.869", "0.5444", "0.7246", "0.542" ], [ "0.7305", "0.528", "0.75", "0.528", "0.75", "0.532", "0.7305", "0.532" ], [ "0.1184", "0.538", "0.471", "0.538", "0.471", "0.552", "0.1184", "0.552" ], [ "0.623", "0.5464", "0.769", "0.5464", "0.769", "0.557", "0.623", "0.557" ], [ "0.12006", "0.554", "0.4836", "0.554", "0.4836", "0.564", "0.12006", "0.564" ], [ "0.12006", "0.5664", "0.4878", "0.5664", "0.4878", "0.58", "0.12006", "0.58" ], [ "0.511", "0.5664", "0.881", "0.5664", "0.881", "0.58", "0.511", "0.58" ], [ "0.12006", "0.582", "0.4895", "0.582", "0.4895", "0.5957", "0.12006", "0.5957" ], [ "0.511", "0.5806", "0.8833", "0.5806", "0.8833", "0.594", "0.511", "0.594" ], [ "0.12177", "0.5967", "0.4895", "0.5967", "0.4895", "0.61", "0.12177", "0.61" ], [ "0.5146", "0.597", "0.5767", "0.597", "0.5767", "0.6094", "0.5146", "0.6094" ], [ "0.12177", "0.6123", "0.4878", "0.6123", "0.4878", "0.6226", "0.12177", "0.6226" ], [ "0.12006", "0.6245", "0.4878", "0.6245", "0.4878", "0.638", "0.12006", "0.638" ], [ "0.5146", "0.6245", "0.882", "0.6245", "0.882", "0.638", "0.5146", "0.638" ], [ "0.12006", "0.6396", "0.4878", "0.6396", "0.4878", "0.653", "0.12006", "0.653" ], [ "0.513", "0.637", "0.882", "0.637", "0.882", "0.651", "0.513", "0.651" ], [ "0.12177", "0.654", "0.486", "0.654", "0.486", "0.6646", "0.12177", "0.6646" ], [ "0.5156", "0.653", "0.8833", "0.653", "0.8833", "0.6665", "0.5156", "0.6665" ], [ "0.12006", "0.6675", "0.4895", "0.6675", "0.4895", "0.681", "0.12006", "0.681" ], [ "0.513", "0.665", "0.8833", "0.6665", "0.8833", "0.681", "0.513", "0.6797" ], [ "0.1184", "0.681", "0.4878", "0.681", "0.4878", "0.695", "0.1184", "0.695" ], [ "0.513", "0.681", "0.882", "0.6797", "0.882", "0.695", "0.513", "0.6953" ], [ "0.12006", "0.6953", "0.4895", "0.6953", "0.4895", "0.709", "0.12006", "0.709" ], [ "0.5146", "0.6953", "0.882", "0.6953", "0.882", "0.709", "0.5146", "0.709" ], [ "0.1184", "0.7104", "0.4895", "0.7104", "0.4895", "0.7236", "0.1184", "0.7236" ], [ "0.5146", "0.7104", "0.8833", "0.7104", "0.8833", "0.7236", "0.5146", "0.7236" ], [ "0.1184", "0.7227", "0.4878", "0.7236", "0.4878", "0.7383", "0.1184", "0.737" ], [ "0.5156", "0.7256", "0.615", "0.7256", "0.615", "0.7363", "0.5156", "0.7363" ], [ "0.12177", "0.7383", "0.4895", "0.7383", "0.4895", "0.752", "0.12177", "0.752" ], [ "0.532", "0.7393", "0.8833", "0.7393", "0.8833", "0.753", "0.532", "0.753" ], [ "0.12177", "0.7544", "0.486", "0.7544", "0.486", "0.764", "0.12177", "0.764" ], [ "0.513", "0.753", "0.882", "0.753", "0.882", "0.7666", "0.513", "0.7666" ], [ "0.12177", "0.7686", "0.4878", "0.7686", "0.4878", "0.7793", "0.12177", "0.7793" ], [ "0.5146", "0.7686", "0.882", "0.7686", "0.882", "0.782", "0.5146", "0.782" ], [ "0.1184", "0.781", "0.3938", "0.781", "0.3938", "0.7944", "0.1184", "0.7944" ], [ "0.5156", "0.7837", "0.881", "0.7837", "0.881", "0.7944", "0.5156", "0.7944" ], [ "0.1394", "0.7974", "0.486", "0.7974", "0.486", "0.808", "0.1394", "0.808" ], [ "0.5146", "0.7964", "0.8833", "0.7964", "0.8833", "0.81", "0.5146", "0.81" ], [ "0.12006", "0.811", "0.4844", "0.811", "0.4844", "0.822", "0.12006", "0.822" ], [ "0.5146", "0.81", "0.882", "0.81", "0.882", "0.8237", "0.5146", "0.8237" ], [ "0.1234", "0.826", "0.4878", "0.826", "0.4878", "0.836", "0.1234", "0.836" ], [ "0.5156", "0.826", "0.881", "0.826", "0.881", "0.836", "0.5156", "0.836" ], [ "0.12006", "0.8384", "0.4895", "0.8384", "0.4895", "0.8516", "0.12006", "0.8516" ], [ "0.513", "0.8384", "0.8833", "0.8384", "0.8833", "0.8516", "0.513", "0.8516" ], [ "0.12177", "0.854", "0.4878", "0.854", "0.4878", "0.8643", "0.12177", "0.8643" ], [ "0.513", "0.8525", "0.881", "0.854", "0.881", "0.865", "0.513", "0.8643" ], [ "0.5156", "0.8677", "0.8833", "0.8677", "0.8833", "0.8813", "0.5156", "0.8813" ], [ "0.1511", "0.8843", "0.2761", "0.8843", "0.2761", "0.8945", "0.1511", "0.8945" ], [ "0.5156", "0.882", "0.882", "0.882", "0.882", "0.8955", "0.5156", "0.8955" ], [ "0.1453", "0.8945", "0.4106", "0.8955", "0.4106", "0.909", "0.1453", "0.9077" ], [ "0.513", "0.8945", "0.8833", "0.8955", "0.8833", "0.91", "0.513", "0.909" ] ]

[ "多模态表征学习的思维图软提示", "Juncheng Yangl,3,*, Zuchao Lil.2,t, Shuai Xie4,*, Wei Yu', Shijun Lil,t, Bo Dul,?", "1School of Computer Science, Wuhan University, Wuhan, Hubei, China", "2National Engineering Research Center for Multimedia Software, Wuhan, Hubei, China", "3School of Electronic Information Engineering,Henan Polytechnic Institute,Nanyang,Henan, China", "4JD Explore Academy, Beijing, China", "[ yjuncheng,zcli-charlie,yuwei,shjli,dubo ] @whu.edu.cn, [email protected]", "Abstract", "The chain-of-thought technique has been received well in multi-modal tasks. It is a step-by-step linear reasoning process that", "adjusts the length of the chain to improve the performance of generated prompts. However, human thought processes are", "predominantly non-linear, as they encompass multiple aspects simultaneously and employ dynamic adjustment and updating", "mechanisms. Therefore, we propose a novel Aggregation-Graph-of-Thought (AGoT) mechanism for soft-prompt tuning in", "multi-modal representation learning. The proposed AGoT models the human thought process not only as a chain but also", "models each step as a reasoning aggregation graph to cope with the overlooked multiple aspects of thinking in single-step", "reasoning. This turns the entire reasoning process into prompt aggregation and prompt flow operations. Experiments show", "that our multi-modal model enhanced with AGoT soft-prompting achieves good results in several tasks such as text-image", "retrieval,visual question answering,and image recognition.In addition,we demonstrate that it has good domaingeneralization", "performance due to better reasoning.", "Keywords: chain-of-thought, aggregation-graph-of-thought, multi-modal, representation learning", "1.Introduction", "Input", "CoTprompts", "nos", "onthetab", "aaatas", "ome books on the bookshelf", "多模态表示学习包括视觉，文本和声音等不同模", "ar", "mebooksandaclockon", "态数据的融合，这对于多模态任务至关重要，例如", "thebookshelf", "(a) Ordinary step-by-ste", "f-7homg", "AGoTprompts", "文本到语音合成E(Renetal.,2019），视听语音识", "别E（Afourasetal.，2018），文本图像检索E(Gabeur", "Input", "There are two kinds of things on the table", "etal.,2022;Yuetal.，2022），视觉语言识别E(Zhou", "Therearesomebooksandaclock onthe", "bookshelf", "", "制", "different types of books", "", "etal.,2022a）和视觉问答(Li etal.，2022；Zheng", "and a black clock on the bookshelf.", "here", "etal.，2021）。在视觉语言模型中，即时学习作为", "(b) Aggregation-Graph-of-Thought", "种有效的策略被广泛采用（Zhuetal.2023:Geetal.", "2023）。提示最早应用于NLP领域，其作用是对预", "Figure l: Comparison between (a) ordinary step-by", "先训练好的语言模型进行提示，帮助其更好地理解", "step Chain-of-Thought and (b) Aggregation-Graph-of-", "人类问题。最早的提示直接与原始文本拼接成离散", "Thought.", "字符，在训练过程中保持不变，称为硬提示（Schick", "and Schitze,2021;Han etal.,2022）。这后来被开发", "推理。从最初的“桌子上有一些东西”到“书架上", "成软提示（Guetal.，2022;Liuetal.，2022），可以", "有一些书”的描述，我们可以看到声明的内容变得", "在训练期间根据上下文语义进行调整。在多模式", "更加充分。最后，短语“书架上有一些书和一个时", "任务中，CoOpE(Zhouetal.,2022b）首先将CLIPE", "钟”完成了对图像的相对完整的描述。这种逐步推", "(Radfordetal.2021)中的固定令牌提示符转换为可", "理提高了模型的视觉理解性能。CoT-PT(Geetal.,", "学习的令牌嵌人。CoCoOp（Zhouetal.，2022a)在", "2023）率先将CoT适应于视觉领域-语言任务，结合", "CoOp中添加了实例图像特征，KgCoOp（Yaoetal.，", "视觉和CoT提示来调整模型，并在下游任务中取", "2023）使用固定文本提示和可学习提示作为监督信", "得良好的效果", "号来约束CLIP模型的训练。然而，这些方法依赖", "尽管CoT技术考虑了推理过程中逐渐加深的理", "于单一的语言提示来指导视觉理解，但没有意识到", "解过程，但由于忽略了可以从多个角度进行理解的", "语言的一个基本但非常重要的特征，即推理，这使", "事实，因此它无法充分利用推理的好处来进行多模", "得多模态模型能够快速适应新的场景。", "态理解。针对这一特点，我们提出了一种新的CoT", "思维链（CoT）技术模拟大脑的推理过程并模", "推理方法，称为聚合思维图（AGoT）。如图1(b)", "其行为，以提高模型的泛化能力。对于复杂的问题", "所示，在输人图片时,AGoT中的每一步，从“桌子上", "不是直接提供答案，而是使用一串推理句子来生成", "有两种东西”到“书架上有一些书和一个钟”再到", "答案。如图E1（a）所示，由CoT给出包含一些书", "“书架上有一些不同类型的书和一个黑色的钟”，推", "籍和一个时钟的图片，然后逐步对图片的内容进行", "理过程的每一步都比图更合理1（a）由于理解了多", "个方面的聚合。很明显，在每个推理节点上，AGoT", "EqualContribution", "聚合了多个层面，而CoT提示只考虑一个层面。从", "+ Correspondence to: Zuchao Li, Shijun Li.", "最右边的模拟结果中可以观察到，在聚合了多个级" ]

{ "arxiv_id": "2404.04538", "language": "zh", "page_idx": 0 }

[ [ "0.12006", "0.0748", "0.4878", "0.0748", "0.4878", "0.0885", "0.12006", "0.0885" ], [ "0.513", "0.076", "0.882", "0.0766", "0.882", "0.0903", "0.513", "0.08905" ], [ "0.12006", "0.0903", "0.4878", "0.0903", "0.4878", "0.10394", "0.12006", "0.10394" ], [ "0.513", "0.0903", "0.8833", "0.0903", "0.8833", "0.10394", "0.513", "0.10394" ], [ "0.12006", "0.1057", "0.486", "0.1057", "0.486", "0.1164", "0.12006", "0.1164" ], [ "0.5146", "0.1057", "0.879", "0.1057", "0.879", "0.1164", "0.5146", "0.1164" ], [ "0.12177", "0.11816", "0.4895", "0.11816", "0.4895", "0.1318", "0.12177", "0.1318" ], [ "0.511", "0.11755", "0.882", "0.11816", "0.882", "0.1318", "0.511", "0.1312" ], [ "0.12177", "0.133", "0.4895", "0.1318", "0.4895", "0.1455", "0.12177", "0.1467" ], [ "0.5146", "0.133", "0.881", "0.1342", "0.881", "0.1455", "0.5146", "0.1443" ], [ "0.12177", "0.1478", "0.4895", "0.1478", "0.4895", "0.1609", "0.12177", "0.1609" ], [ "0.513", "0.1455", "0.881", "0.1455", "0.881", "0.1592", "0.513", "0.1592" ], [ "0.12177", "0.1633", "0.486", "0.1633", "0.486", "0.1733", "0.12177", "0.1733" ], [ "0.5156", "0.1633", "0.881", "0.1633", "0.881", "0.1733", "0.5156", "0.1733" ], [ "0.1301", "0.1758", "0.4895", "0.1758", "0.4895", "0.1895", "0.1301", "0.1895" ], [ "0.513", "0.1758", "0.882", "0.1758", "0.882", "0.1895", "0.513", "0.1895" ], [ "0.12006", "0.1907", "0.3552", "0.1907", "0.3552", "0.2037", "0.12006", "0.2037" ], [ "0.517", "0.1912", "0.6484", "0.1912", "0.6484", "0.2019", "0.517", "0.2019" ], [ "0.1351", "0.2037", "0.4895", "0.2037", "0.4895", "0.2173", "0.1351", "0.2173" ], [ "0.12006", "0.2185", "0.4895", "0.2185", "0.4895", "0.2322", "0.12006", "0.2322" ], [ "0.5225", "0.2185", "0.882", "0.2197", "0.882", "0.234", "0.5225", "0.2328" ], [ "0.1797", "0.234", "0.4878", "0.234", "0.4878", "0.2446", "0.1797", "0.2446" ], [ "0.5146", "0.234", "0.882", "0.234", "0.882", "0.2477", "0.5146", "0.2477" ], [ "0.12006", "0.2465", "0.4895", "0.2465", "0.4895", "0.26", "0.12006", "0.26" ], [ "0.1234", "0.2363", "0.1838", "0.2363", "0.1838", "0.2434", "0.1234", "0.2434" ], [ "0.513", "0.2488", "0.8833", "0.2488", "0.8833", "0.2625", "0.513", "0.2625" ], [ "0.12177", "0.2625", "0.4878", "0.2625", "0.4878", "0.2725", "0.12177", "0.2725" ], [ "0.5146", "0.2632", "0.881", "0.2632", "0.881", "0.2766", "0.5146", "0.2766" ], [ "0.12006", "0.275", "0.4895", "0.275", "0.4895", "0.288", "0.12006", "0.288" ], [ "0.5146", "0.279", "0.882", "0.279", "0.882", "0.2893", "0.5146", "0.2893" ], [ "0.12006", "0.2893", "0.4895", "0.2893", "0.4895", "0.3027", "0.12006", "0.3027" ], [ "0.5156", "0.2935", "0.8696", "0.2935", "0.8696", "0.304", "0.5156", "0.304" ], [ "0.12177", "0.3047", "0.486", "0.3047", "0.486", "0.3154", "0.12177", "0.3154" ], [ "0.5146", "0.306", "0.882", "0.306", "0.882", "0.3196", "0.5146", "0.3196" ], [ "0.12177", "0.3196", "0.4878", "0.3196", "0.4878", "0.33", "0.12177", "0.33" ], [ "0.5146", "0.3218", "0.881", "0.3218", "0.881", "0.332", "0.5146", "0.332" ], [ "0.12006", "0.332", "0.4878", "0.332", "0.4878", "0.3457", "0.12006", "0.3457" ], [ "0.5146", "0.3342", "0.882", "0.3342", "0.882", "0.3474", "0.5146", "0.3474" ], [ "0.12177", "0.3467", "0.4902", "0.3467", "0.4902", "0.3599", "0.12177", "0.3599" ], [ "0.5156", "0.3499", "0.882", "0.3499", "0.882", "0.3599", "0.5156", "0.3599" ], [ "0.1184", "0.3599", "0.4878", "0.3599", "0.4878", "0.3735", "0.1184", "0.3735" ], [ "0.513", "0.3623", "0.8833", "0.3635", "0.8833", "0.3772", "0.513", "0.376" ], [ "0.12177", "0.376", "0.2032", "0.376", "0.2032", "0.3872", "0.12177", "0.3872" ], [ "0.517", "0.3777", "0.881", "0.3777", "0.881", "0.3884", "0.517", "0.3884" ], [ "0.1368", "0.3896", "0.361", "0.3896", "0.361", "0.4026", "0.1368", "0.4026" ], [ "0.5146", "0.3896", "0.882", "0.3901", "0.882", "0.4038", "0.5146", "0.4026" ], [ "0.5146", "0.405", "0.882", "0.405", "0.882", "0.4187", "0.5146", "0.4187" ], [ "0.1368", "0.411", "0.4895", "0.4104", "0.4895", "0.4246", "0.1368", "0.4258" ], [ "0.5225", "0.4233", "0.529", "0.4233", "0.529", "0.427", "0.5225", "0.427" ], [ "0.525", "0.4204", "0.879", "0.4204", "0.879", "0.4312", "0.525", "0.4312" ], [ "0.1528", "0.4275", "0.4878", "0.4275", "0.4878", "0.4382", "0.1528", "0.4382" ], [ "0.5146", "0.4316", "0.881", "0.4329", "0.881", "0.4465", "0.5146", "0.4453" ], [ "0.1511", "0.4407", "0.398", "0.4407", "0.398", "0.4536", "0.1511", "0.4536" ], [ "0.5156", "0.4478", "0.882", "0.4478", "0.882", "0.4614", "0.5156", "0.4614" ], [ "0.1368", "0.4631", "0.4895", "0.4631", "0.4895", "0.4768", "0.1368", "0.4768" ], [ "0.5156", "0.4631", "0.881", "0.4631", "0.881", "0.4739", "0.5156", "0.4739" ], [ "0.1528", "0.478", "0.4878", "0.478", "0.4878", "0.4912", "0.1528", "0.4912" ], [ "0.5146", "0.4756", "0.8833", "0.4756", "0.8833", "0.4893", "0.5146", "0.4893" ], [ "0.1528", "0.4922", "0.4878", "0.4922", "0.4878", "0.506", "0.1528", "0.506" ], [ "0.5146", "0.4905", "0.882", "0.4905", "0.882", "0.5044", "0.5146", "0.5044" ], [ "0.1511", "0.505", "0.4895", "0.506", "0.4895", "0.521", "0.1511", "0.5195" ], [ "0.5146", "0.505", "0.882", "0.505", "0.882", "0.5186", "0.5146", "0.5186" ], [ "0.1528", "0.5225", "0.3074", "0.5225", "0.3074", "0.533", "0.1528", "0.533" ], [ "0.5156", "0.5186", "0.882", "0.5186", "0.882", "0.532", "0.5156", "0.532" ], [ "0.5146", "0.5337", "0.565", "0.5337", "0.565", "0.546", "0.5146", "0.546" ], [ "0.1368", "0.5435", "0.4878", "0.5435", "0.4878", "0.557", "0.1368", "0.557" ], [ "0.1528", "0.558", "0.4878", "0.558", "0.4878", "0.572", "0.1528", "0.572" ], [ "0.613", "0.5625", "0.7715", "0.5654", "0.7715", "0.583", "0.613", "0.58" ], [ "0.1561", "0.5723", "0.4844", "0.5723", "0.4844", "0.583", "0.1561", "0.583" ], [ "0.1528", "0.585", "0.4878", "0.585", "0.4878", "0.5986", "0.1528", "0.5986" ], [ "0.511", "0.5933", "0.738", "0.5957", "0.738", "0.61", "0.511", "0.608" ], [ "0.1528", "0.5996", "0.2477", "0.6016", "0.2477", "0.6133", "0.1528", "0.611" ], [ "0.5146", "0.616", "0.881", "0.616", "0.881", "0.63", "0.5146", "0.63" ], [ "0.5156", "0.6313", "0.881", "0.6313", "0.881", "0.6455", "0.5156", "0.6455" ], [ "0.22", "0.641", "0.382", "0.641", "0.382", "0.6562", "0.22", "0.6562" ], [ "0.5156", "0.6465", "0.882", "0.6465", "0.882", "0.66", "0.5156", "0.66" ], [ "0.5146", "0.66", "0.6826", "0.66", "0.6826", "0.6733", "0.5146", "0.6733" ], [ "0.1234", "0.6685", "0.4072", "0.6685", "0.4072", "0.6816", "0.1234", "0.6816" ], [ "0.3997", "0.6694", "0.4878", "0.6694", "0.4878", "0.6797", "0.3997", "0.6797" ], [ "0.12177", "0.684", "0.486", "0.684", "0.486", "0.695", "0.12177", "0.695" ], [ "0.671", "0.684", "0.791", "0.681", "0.791", "0.695", "0.671", "0.6978" ], [ "0.5786", "0.6953", "0.6465", "0.6953", "0.6465", "0.707", "0.5786", "0.707" ], [ "0.86", "0.695", "0.885", "0.695", "0.885", "0.707", "0.86", "0.707" ], [ "0.6514", "0.6978", "0.8237", "0.6978", "0.8237", "0.7197", "0.6514", "0.7197" ], [ "0.12006", "0.697", "0.4878", "0.697", "0.4878", "0.7104", "0.12006", "0.7104" ], [ "0.12006", "0.7114", "0.4878", "0.7114", "0.4878", "0.7246", "0.12006", "0.7246" ], [ "0.12006", "0.7256", "0.486", "0.7256", "0.486", "0.7363", "0.12006", "0.7363" ], [ "0.5146", "0.7285", "0.879", "0.7285", "0.879", "0.7393", "0.5146", "0.7393" ], [ "0.12177", "0.741", "0.4878", "0.741", "0.4878", "0.752", "0.12177", "0.752" ], [ "0.5156", "0.7446", "0.882", "0.7446", "0.882", "0.755", "0.5156", "0.755" ], [ "0.12006", "0.7544", "0.4895", "0.7544", "0.4895", "0.767", "0.12006", "0.767" ], [ "0.513", "0.7573", "0.882", "0.7583", "0.882", "0.7695", "0.513", "0.7686" ], [ "0.12006", "0.7686", "0.4878", "0.7686", "0.4878", "0.782", "0.12006", "0.782" ], [ "0.513", "0.771", "0.881", "0.771", "0.881", "0.7837", "0.513", "0.7837" ], [ "0.1184", "0.782", "0.4878", "0.781", "0.4878", "0.796", "0.1184", "0.7964" ], [ "0.5146", "0.7837", "0.881", "0.785", "0.881", "0.799", "0.5146", "0.7974" ], [ "0.12006", "0.7964", "0.43", "0.7964", "0.43", "0.81", "0.12006", "0.81" ], [ "0.5146", "0.8", "0.881", "0.8", "0.881", "0.8135", "0.5146", "0.8135" ], [ "0.1368", "0.81", "0.4878", "0.81", "0.4878", "0.8237", "0.1368", "0.8237" ], [ "0.513", "0.814", "0.876", "0.814", "0.876", "0.8276", "0.513", "0.8276" ], [ "0.12006", "0.8247", "0.4878", "0.8247", "0.4878", "0.8384", "0.12006", "0.8384" ], [ "0.1184", "0.8374", "0.4878", "0.8384", "0.4878", "0.8525", "0.1184", "0.8516" ], [ "0.513", "0.8457", "0.7773", "0.8457", "0.7773", "0.8594", "0.513", "0.8594" ], [ "0.12177", "0.855", "0.486", "0.855", "0.486", "0.865", "0.12177", "0.865" ], [ "0.12006", "0.8677", "0.4878", "0.8677", "0.4878", "0.8813", "0.12006", "0.8813" ], [ "0.513", "0.8677", "0.882", "0.8677", "0.882", "0.8813", "0.513", "0.8813" ], [ "0.12006", "0.882", "0.4878", "0.882", "0.4878", "0.8955", "0.12006", "0.8955" ], [ "0.5146", "0.882", "0.882", "0.882", "0.882", "0.8955", "0.5146", "0.8955" ], [ "0.12006", "0.8965", "0.4895", "0.8965", "0.4895", "0.91", "0.12006", "0.91" ], [ "0.5146", "0.8965", "0.882", "0.8965", "0.882", "0.91", "0.5146", "0.91" ] ]

[ "别的方面之后，AGoT提示提供了对问题的更全面", "可学习提示和手工制作提示之间的差异来增强模", "的了解。具体来说，AGoT将CoT中的每个推理步", "型的泛化。DualCoOp（Sunetal.,2022）提出使用双", "骤建模为推理聚合图，以应对单步推理中被忽视的", "上下文优化作为统一框架来解决多标签识别任务", "多视图思维问题。图中的每个聚合节点也充当CoT", "ProGrad（Zhuetal.，2023）更新提示的方式使其梯", "内的一个节点。多个节点相互连接形成一个链，每", "度与一般知识保持一致，而ProDA（Luetal.，2022）", "个节点的表示形式从前一个节点、多个子节点和视", "侧重于学习提示先验分布。这些方法利用语言提示", "觉特征聚合而成。然后将得到的节点表示传递给下", "来协助视觉任务。然而，他们未能复制人类固有的", "个。我们设计了一个流量控制器来控制前一个节", "推理过程，从而错过了通过视觉语言推理获得的深", "点和当前节点之间的信息流程度", "层视觉理解的优势", "我们使用强基线模型CLIP作为基础模型，并在", "视觉语言理解任务，文本图像检索（Flickr30kE", "Chain-of-ThoughtCoT技术在大型语言模型", "retal.,2015)，MSCOCOEl(Chenetal.,", "(LLM）少射推理和微调方面表现出色。Wei等人。", "2015））和视觉问答（VQAv2E(Shenetal.，2022)", "Plummei", "(Wei etal.,2022)在LLM中率先使用CoT作为离", "）中进行了实验。。与CLIP+5相比，我们的AGoT", "散线索学习，效果良好。自洽性-CoT（Wangetal.，", "方法取得了更好的结果。70%，+5。40%，+19。91%", "2023）通过对答案使用多数投票，显著提高了CoT", "在每个任务中，并且与CLIP+CoT相比也显示出改", "方法的性能。STaR（Zelikmanetal.，2022）提出了", "进，+1。70%，+0。80%，+0。88%分别。此外，由于", "种提升方法，即使是中小型模型也可以使用CoT。", "我们的AGoT可以用作软提示方法，它可以提高视", "AutoCoT（Zhangetal.，2023a）对各种问题进行采样", "觉分类任务的泛化能力。我们在各种视觉分类任务", "并自动生成推理链。Shi等人。(Shietal.,2023）评", "上对其进行了评估，发现与以前的最新方法相比，", "估CoT解决多语种小学数学问题的强大多语种推", "AGoT实现了更好的域泛化能力，证明了AGoT方", "理能力。ToT（Yaoetal.,2024）允许LLM通过考虑", "法的有效性", "许多不同的推理和自我评估路径来选择行动方案。", "总之，我们的贡献有三个方面：", "CoT在多模态任务中也有广泛的应用。Multimodal", "CoT(Zhangetal,2023b）将语言和视觉模态集成到", "we propose a novel Aggregation-Graph-of-", "-", "一个两阶段框架中，使用多模态信息更好地推断答", "Thought mechanism for soft-prompt tuning in", "案。他等人。(Heetal.，2023）建议通过潜在的空", "multi-modal representation learning", "间扩散过程生成与语言思想一致的有效图像特征。", ": To cope with the overlooked multiple aspects of", "CoT-PT（Geetal2023）使用CoT提示调整视觉语", "thinking in single-step reasoning, we model the", "言模型。VCoT(Roseetal.,2023）将视觉增强集成", "reasoning step as an aggregation graph, and turn", "到推理任务中，使用CoT提示和视觉效果递归地", "the whole process into a prompt aggregation and", "弥合顺序数据中的逻辑空白。然而，这种方法是面", "promptflow operation.", "向生成性任务的，而我们的方法解决了下游的歧视", "性任务", ": AGoT exhibits strong multi-modal representation", "learning ability and achieves good results on 18", "3.Methodology", "datasets such as text-image retrieval,VQA,and", "image classification and has good domain general", "3.1.Contrastive Learning", "izationability", "为了执行多模态表示学习，这项工作遵循成对文本", "和图像中的对比学习。与CLIP一致，令&为输人", "2.Related Work", "图像，t为提示和相应类形成的文本序列。该模型", "计算每个类ck的概率：", "Multi-modal Representation Learning", "多模态表", "示学习具有大量基于视觉语言编码结构的模型。", "ecos(G(t),H(r))/T", "p(chr) =", "(1)", "EKtecos(t),H(a)/r*", "CLIP是使用Transformer (Vaswani et al.,2017)的文", "本编码器，而图像编码器是ResNet(Heetal.,2016)", "或VisionTransformer(Dosovitskiyetal.2021)。该", "其中cos（）表示余弦相似度：H）表示来自图像", "模型在4亿个图像-文本对上进行了训练，并在30", "编码器的图像表示，G（t）表示来自连接提示和类", "个数据集上表现良好。ALIGNE(Jiaetal.，2021)是", "描述句子的文本编码器的文本表示tK的总数", "另一种典型的视觉语言模型，它使用对比学习来训", "为K。对比学习的目标是使用有限的数据对学习", "练超过十亿个噜杂的图像-文本对。它在MSCOCO", "广义图像编码器H(）和文本编码器g(），从而使", "和Flickr30k数据集上实现了最先进的性能", "模型具有跨模态语义空间，并可以更好地转移到下", "这些多模态基础模型广泛用于下游任务，例如", "游任务，例如文本图像检索，VQA和细粒度分类", "图像分类，视频文本检索，跟踪，图像字幕和点云", "理解。CoOp通过将CLIP中“a的照片”的固定文", "3.2.Prompt Learning and CoT", "本描述转换为可学习的令牌嵌入来增强模型的泛", "化性能。为了解决在看不见的数据上性能不佳的", "CoOp首先将CLIP模型中的固定提示（例如“a的", "挑战，CoCoOp将实例功能合并到文本描述中，从", "照片”）转换为可学习的嵌入hP，从而进行软提", "而进一步提高了模型的性能。KgCoOp通过最小化", "示。这克服了在手动创建提示时对技能和工程经验" ]

{ "arxiv_id": "2404.04538", "language": "zh", "page_idx": 1 }

[ [ "0.2167", "0.079", "0.2301", "0.079", "0.2301", "0.0885", "0.2167", "0.0885" ], [ "0.3552", "0.08014", "0.372", "0.08014", "0.372", "0.08905", "0.3552", "0.08905" ], [ "0.178", "0.0974", "0.2048", "0.0974", "0.2048", "0.10925", "0.178", "0.10925" ], [ "0.3198", "0.0986", "0.3342", "0.0986", "0.3342", "0.1081", "0.3198", "0.1081" ], [ "0.3896", "0.0968", "0.4014", "0.10095", "0.3955", "0.10925", "0.3838", "0.1051" ], [ "0.69", "0.11755", "0.7188", "0.11755", "0.7188", "0.1289", "0.69", "0.1289" ], [ "0.2468", "0.0998", "0.2678", "0.0998", "0.2678", "0.1081", "0.2468", "0.1081" ], [ "0.602", "0.1259", "0.6675", "0.1259", "0.6675", "0.136", "0.602", "0.136" ], [ "0.66", "0.1271", "0.729", "0.1289", "0.729", "0.1401", "0.66", "0.1383" ], [ "0.3896", "0.1538", "0.409", "0.1538", "0.409", "0.1621", "0.3896", "0.1621" ], [ "0.3586", "0.1277", "0.3687", "0.1277", "0.3687", "0.133", "0.3586", "0.133" ], [ "0.3208", "0.155", "0.3418", "0.155", "0.3418", "0.1633", "0.3208", "0.1633" ], [ "0.4902", "0.155", "0.519", "0.155", "0.519", "0.1633", "0.4902", "0.1633" ], [ "0.5615", "0.155", "0.5825", "0.155", "0.5825", "0.1621", "0.5615", "0.1621" ], [ "0.357", "0.187", "0.3728", "0.1912", "0.3687", "0.2007", "0.3525", "0.1965" ], [ "0.2477", "0.1567", "0.266", "0.1567", "0.266", "0.1603", "0.2477", "0.1603" ], [ "0.2167", "0.1907", "0.2284", "0.1907", "0.2284", "0.1996", "0.2167", "0.1996" ], [ "0.534", "0.1936", "0.5425", "0.1936", "0.5425", "0.199", "0.534", "0.199" ], [ "0.1914", "0.2173", "0.251", "0.2173", "0.251", "0.2286", "0.1914", "0.2286" ], [ "0.33", "0.2173", "0.3914", "0.2173", "0.3914", "0.2286", "0.33", "0.2286" ], [ "0.504", "0.2173", "0.565", "0.2173", "0.565", "0.2286", "0.504", "0.2286" ], [ "0.735", "0.2197", "0.75", "0.2197", "0.75", "0.2286", "0.735", "0.2286" ], [ "0.1427", "0.2452", "0.1755", "0.2452", "0.1755", "0.257", "0.1427", "0.257" ], [ "0.126", "0.257", "0.189", "0.257", "0.189", "0.2673", "0.126", "0.2673" ], [ "0.5728", "0.257", "0.6113", "0.257", "0.6113", "0.2683", "0.5728", "0.2683" ], [ "0.5684", "0.2683", "0.617", "0.2683", "0.617", "0.279", "0.5684", "0.279" ], [ "0.659", "0.2632", "0.6963", "0.2632", "0.6963", "0.275", "0.659", "0.275" ], [ "0.748", "0.288", "0.8423", "0.288", "0.8423", "0.2998", "0.748", "0.2998" ], [ "0.1184", "0.3154", "0.881", "0.3154", "0.881", "0.3289", "0.1184", "0.3289" ], [ "0.1184", "0.33", "0.8403", "0.33", "0.8403", "0.3433", "0.1184", "0.3433" ], [ "0.12006", "0.3552", "0.4895", "0.3552", "0.4895", "0.3682", "0.12006", "0.3682" ], [ "0.5146", "0.3552", "0.882", "0.3552", "0.882", "0.3682", "0.5146", "0.3682" ], [ "0.1184", "0.3682", "0.4902", "0.3677", "0.4902", "0.3818", "0.1184", "0.383" ], [ "0.5146", "0.3694", "0.882", "0.3694", "0.882", "0.383", "0.5146", "0.383" ], [ "0.12177", "0.3855", "0.2107", "0.3855", "0.2107", "0.3955", "0.12177", "0.3955" ], [ "0.5156", "0.3843", "0.882", "0.3843", "0.882", "0.398", "0.5156", "0.398" ], [ "0.5146", "0.398", "0.882", "0.398", "0.882", "0.411", "0.5146", "0.411" ], [ "0.2627", "0.4062", "0.4106", "0.4038", "0.4106", "0.4187", "0.2627", "0.4204" ], [ "0.5156", "0.411", "0.882", "0.411", "0.882", "0.4258", "0.5156", "0.4258" ], [ "0.168", "0.4187", "0.2377", "0.4187", "0.2377", "0.43", "0.168", "0.43" ], [ "0.4668", "0.4187", "0.4895", "0.4187", "0.4895", "0.4287", "0.4668", "0.4287" ], [ "0.2418", "0.4246", "0.4375", "0.4204", "0.4375", "0.437", "0.2418", "0.4412" ], [ "0.2468", "0.4246", "0.2747", "0.4246", "0.2747", "0.4316", "0.2468", "0.4316" ], [ "0.5156", "0.4258", "0.769", "0.4258", "0.769", "0.4395", "0.5156", "0.4395" ], [ "0.12177", "0.4507", "0.4878", "0.4507", "0.4878", "0.4656", "0.12177", "0.4656" ], [ "0.5205", "0.4578", "0.8833", "0.4578", "0.8833", "0.4727", "0.5205", "0.4727" ], [ "0.12177", "0.466", "0.1377", "0.466", "0.1377", "0.4768", "0.12177", "0.4768" ], [ "0.5146", "0.4727", "0.881", "0.4727", "0.881", "0.4863", "0.5146", "0.4863" ], [ "0.1377", "0.4797", "0.4878", "0.4797", "0.4878", "0.4934", "0.1377", "0.4934" ], [ "0.5283", "0.4863", "0.8833", "0.485", "0.8833", "0.4995", "0.5283", "0.5005" ], [ "0.12006", "0.4934", "0.4878", "0.4934", "0.4878", "0.5073", "0.12006", "0.5073" ], [ "0.513", "0.5005", "0.881", "0.5005", "0.881", "0.514", "0.513", "0.514" ], [ "0.12177", "0.5073", "0.4878", "0.5073", "0.4878", "0.521", "0.12177", "0.521" ], [ "0.1184", "0.521", "0.4895", "0.5215", "0.4895", "0.536", "0.1184", "0.535" ], [ "0.5146", "0.5156", "0.882", "0.5156", "0.882", "0.5293", "0.5146", "0.5293" ], [ "0.5156", "0.53", "0.8833", "0.53", "0.8833", "0.5435", "0.5156", "0.5435" ], [ "0.12006", "0.536", "0.4148", "0.536", "0.4148", "0.55", "0.12006", "0.55" ], [ "0.5146", "0.5435", "0.6533", "0.5435", "0.6533", "0.557", "0.5146", "0.557" ], [ "0.236", "0.558", "0.419", "0.5557", "0.419", "0.571", "0.236", "0.5723" ], [ "0.465", "0.5693", "0.4895", "0.5693", "0.4895", "0.582", "0.465", "0.582" ], [ "0.679", "0.5684", "0.7", "0.5684", "0.7", "0.5806", "0.679", "0.5806" ], [ "0.2351", "0.576", "0.4458", "0.5693", "0.4475", "0.5903", "0.236", "0.5967" ], [ "0.671", "0.5776", "0.791", "0.5806", "0.791", "0.5986", "0.671", "0.5957" ], [ "0.1411", "0.571", "0.2107", "0.571", "0.2107", "0.582", "0.1411", "0.582" ], [ "0.2225", "0.579", "0.2477", "0.579", "0.2477", "0.583", "0.2225", "0.583" ], [ "0.613", "0.583", "0.681", "0.583", "0.681", "0.594", "0.613", "0.594" ], [ "0.679", "0.597", "0.7", "0.597", "0.7", "0.6055", "0.679", "0.6055" ], [ "0.12177", "0.605", "0.4534", "0.605", "0.4534", "0.615", "0.12177", "0.615" ], [ "0.5767", "0.61", "0.818", "0.611", "0.818", "0.626", "0.5767", "0.6245" ], [ "0.1351", "0.6177", "0.486", "0.6177", "0.486", "0.632", "0.1351", "0.632" ], [ "0.8613", "0.6123", "0.885", "0.6123", "0.885", "0.6245", "0.8613", "0.6245" ], [ "0.1184", "0.632", "0.4878", "0.632", "0.4878", "0.6455", "0.1184", "0.6455" ], [ "0.687", "0.629", "0.6987", "0.629", "0.6987", "0.638", "0.687", "0.638" ], [ "0.596", "0.6357", "0.771", "0.6357", "0.771", "0.6577", "0.596", "0.6577" ], [ "0.12006", "0.648", "0.1931", "0.648", "0.1931", "0.659", "0.12006", "0.659" ], [ "0.679", "0.6567", "0.705", "0.6567", "0.705", "0.6685", "0.679", "0.6685" ], [ "0.2258", "0.6685", "0.4324", "0.6675", "0.4324", "0.681", "0.2258", "0.6816" ], [ "0.12427", "0.6797", "0.1873", "0.6797", "0.1873", "0.695", "0.12427", "0.695" ], [ "0.1957", "0.684", "0.492", "0.675", "0.4937", "0.6997", "0.1973", "0.709" ], [ "0.529", "0.6787", "0.885", "0.6787", "0.885", "0.6934", "0.529", "0.6934" ], [ "0.2065", "0.6885", "0.2334", "0.6885", "0.2334", "0.6953", "0.2065", "0.6953" ], [ "0.5225", "0.686", "0.529", "0.686", "0.529", "0.6904", "0.5225", "0.6904" ], [ "0.517", "0.6934", "0.76", "0.6934", "0.76", "0.707", "0.517", "0.707" ], [ "0.531", "0.708", "0.882", "0.708", "0.882", "0.7217", "0.531", "0.7217" ], [ "0.12006", "0.7144", "0.3896", "0.7144", "0.3896", "0.728", "0.12006", "0.728" ], [ "0.5156", "0.7227", "0.8833", "0.7227", "0.8833", "0.7363", "0.5156", "0.7363" ], [ "0.5156", "0.737", "0.882", "0.737", "0.882", "0.7505", "0.5156", "0.7505" ], [ "0.12006", "0.745", "0.1696", "0.745", "0.1696", "0.7573", "0.12006", "0.7573" ], [ "0.1628", "0.745", "0.424", "0.745", "0.424", "0.759", "0.1628", "0.759" ], [ "0.5146", "0.75", "0.882", "0.75", "0.882", "0.763", "0.5146", "0.763" ], [ "0.1184", "0.767", "0.4878", "0.767", "0.4878", "0.781", "0.1184", "0.781" ], [ "0.5146", "0.7656", "0.7925", "0.7656", "0.7925", "0.7793", "0.5146", "0.7793" ], [ "0.1234", "0.782", "0.486", "0.782", "0.486", "0.796", "0.1234", "0.796" ], [ "0.571", "0.792", "0.8286", "0.792", "0.8286", "0.8057", "0.571", "0.8057" ], [ "0.86", "0.792", "0.885", "0.792", "0.885", "0.8047", "0.86", "0.8047" ], [ "0.12006", "0.81", "0.4878", "0.81", "0.4878", "0.8237", "0.12006", "0.8237" ], [ "0.12006", "0.7974", "0.4844", "0.7974", "0.4844", "0.808", "0.12006", "0.808" ], [ "0.12177", "0.826", "0.4844", "0.826", "0.4844", "0.836", "0.12177", "0.836" ], [ "0.5156", "0.8237", "0.881", "0.8237", "0.881", "0.8374", "0.5156", "0.8374" ], [ "0.12006", "0.8384", "0.4895", "0.8384", "0.4895", "0.8516", "0.12006", "0.8516" ], [ "0.5146", "0.8384", "0.8833", "0.8384", "0.8833", "0.8516", "0.5146", "0.8516" ], [ "0.12177", "0.8525", "0.4878", "0.8525", "0.4878", "0.866", "0.12177", "0.866" ], [ "0.5156", "0.8525", "0.8833", "0.8525", "0.8833", "0.866", "0.5156", "0.866" ], [ "0.1184", "0.866", "0.4878", "0.865", "0.4878", "0.88", "0.1184", "0.8813" ], [ "0.5146", "0.868", "0.881", "0.868", "0.881", "0.879", "0.5146", "0.879" ], [ "0.12177", "0.8813", "0.4878", "0.8813", "0.4878", "0.8945", "0.12177", "0.8945" ], [ "0.5156", "0.88", "0.8833", "0.8813", "0.8833", "0.8955", "0.5156", "0.8945" ], [ "0.12006", "0.8955", "0.4895", "0.8955", "0.4895", "0.909", "0.12006", "0.909" ], [ "0.511", "0.8955", "0.882", "0.8955", "0.882", "0.909", "0.511", "0.909" ] ]

[ "Gi", "G2", "ni1", "n", "n", "Text", "ni2", "[CLASS]", "→Encoder", "n2", "N", "n23", "nz3", "nzR", "+", ".1", "+", "+", "MetaNet", "MetaNet", "MetaNet", "T1", "Flow", "Controller", "Image", "Encoder", "H(x)", "Similarity Scores", "Figure 2: Ilustration of the proposed AGoT. It learns a high-quality soft prompt with prompt aggregation ( blue", ") and prompt flow operation ( gray ) for multi-view thinking and adaptation to complex multi-modal tasks.", "的需求，并在图像识别，细粒度分类等14个数据", "表示与由MetaNet提取的视觉信息组合，然后流到", "集中的整体性能方面超过了CLIP。每个类的概率", "具有流控制器的下一个子图。经过乙推理步骤后", "可以表示为：", "生成最终的AGoT提示表示S(Gz），并与类标签", "[CLASS]结合，为文本编码器构建最终的输人提", "ecos(9([h,tx),H(z)/T", "示。最后，采用类CLIP多模态对比学习网络，对", "p(ckr)=", "(2)", "ZKecos(G(hP,t:),H()/+", "K", "相似度最高的图像进行分类或描述", "其中[h，ts]表示可学习提示嵌入和文本嵌人的串", "PromptAggregation为了更好地从不同角度探", "联", "索提示信息，我们为AGoT的每个推理步骤提出了", "为了解决CoOp在看不见的类上表现不佳的问", "-种提示聚合方法，以生成高质量的提示表示。详", "题，CoCoOp进一步使用简单的MetaNet网络为每", "细地说，在每个子图G，中，R高斯初始化的元提", "个图像生成条件令牌，然后将此图像侧令牌与文本", "提示相结合，动态地适应每个图像实例的变化，克", "示学习器充当子节点n，WeightNet从图像表示", "H(r）中学习聚合边权重E：中心节点提示表示", "服了类转移的问题。预测概率计算如下：", "8(N）可以表示为：", "ecos(G([h,M(r),tx),H(r)/T", "(3)", "R", "K ecos((h,M(a),t:),H(m)/r", "e(nir)* Eir", "p(ckr) =", "K", "&(N)= ", "r-1", "其中M（a）是来自MetaNet的条件今牌的表示", "where Eir = WeightNet,r(H(a)", "CoT-PT为视觉语言任务中的快速学习引人了", "(5)", "CoT。它利用图像特征进行软提示学习，预测概率", "R", "subject to Eir = 1,", "计算如下：", "r=1", "ecos(G([cot(h),M(),t),H(r)/", "p(chla)", "E,ecos(lo(h),M(),),()/, (4)", "其中WeightNetir（）实现为3层MLP网络，其", "TK", "", "中ReLU作为子图G,的激活函数", "如CoCoOp中所述，它添加了一个MetaNet网络", "其中cot是用于提示学习的CoT模块", "来对图像编码器生成的视觉特征进行编码，并作为", "偏差项添加到文本提示中。我们遵循这种设计，并", "3.3.", "Aggregation-Graph-of-Thought", "引人了二个MetaNet网络，将视觉信息聚合到中央", "CoT由于其卓越的推理能力而在各种任务中取得", "节点表示中。此聚合过程可以表示为：", "了良好的结果。但是，它不能在单个推理步骤下从", "E(G) = &(N) + MetaNet;(H()).", "(6)", "制了多模态理解中推理的进一步帮助。为了解决", "多个视图中的不同实例收集信息。这种局限性限", "这个问题，我们利用图神经网络中的聚合机制，并", "PromptFlow在AGoT中，我们将问题的解决分", "使用子图链G={G1,...,Gz}来构造AGoT。每", "解为一步一步的推理过程，其中每个步骤都中继来", "个子图G=（N,E）是建立在一组子节点n=", "自先前步骤的信息，以逐步加深理解。直观上，随", "nii.nir}上的有向加权图，它从R不同的视图", "着任务复杂性的增加，需要更复杂的推理图链。为", "中学习元提示，并以学习到的权重E，作为边聚合", "了解决这个问题，我们引人了一个动态提示信息流", "到中心节点N，如图2所示。此外，聚合的提示", "控制器。具体来说，给定输人图像，编码的图像特" ]

{ "arxiv_id": "2404.04538", "language": "zh", "page_idx": 2 }

[ [ "0.1436", "0.0778", "0.2284", "0.0778", "0.2284", "0.08905", "0.1436", "0.08905" ], [ "0.2468", "0.079", "0.3005", "0.079", "0.3005", "0.0903", "0.2468", "0.0903" ], [ "0.318", "0.079", "0.3804", "0.079", "0.3804", "0.0903", "0.318", "0.0903" ], [ "0.3955", "0.079", "0.4634", "0.079", "0.4634", "0.0903", "0.3955", "0.0903" ], [ "0.519", "0.076", "0.882", "0.076", "0.882", "0.08905", "0.519", "0.08905" ], [ "0.511", "0.0885", "0.8833", "0.08905", "0.8833", "0.10394", "0.511", "0.1027" ], [ "0.1704", "0.0974", "0.2024", "0.0974", "0.2024", "0.11224", "0.1704", "0.11224" ], [ "0.2544", "0.0986", "0.2913", "0.0986", "0.2913", "0.111", "0.2544", "0.111" ], [ "0.33", "0.0986", "0.3687", "0.0986", "0.3687", "0.111", "0.33", "0.111" ], [ "0.4106", "0.0986", "0.4492", "0.0986", "0.4492", "0.111", "0.4106", "0.111" ], [ "0.513", "0.10394", "0.882", "0.10394", "0.882", "0.11755", "0.513", "0.11755" ], [ "0.3315", "0.11816", "0.3687", "0.11816", "0.3687", "0.1312", "0.3315", "0.1312" ], [ "0.4106", "0.11816", "0.4492", "0.11816", "0.4492", "0.1312", "0.4106", "0.1312" ], [ "0.5156", "0.11816", "0.8833", "0.11816", "0.8833", "0.1318", "0.5156", "0.1318" ], [ "0.409", "0.1277", "0.451", "0.13", "0.4492", "0.1455", "0.4072", "0.1437" ], [ "0.2435", "0.1194", "0.3005", "0.1194", "0.3005", "0.1312", "0.2435", "0.1312" ], [ "0.1646", "0.1318", "0.2024", "0.1318", "0.2024", "0.1437", "0.1646", "0.1437" ], [ "0.2468", "0.1318", "0.2988", "0.1318", "0.2988", "0.1437", "0.2468", "0.1437" ], [ "0.33", "0.1318", "0.3687", "0.1318", "0.3687", "0.1443", "0.33", "0.1443" ], [ "0.513", "0.1318", "0.882", "0.1318", "0.882", "0.1455", "0.513", "0.1455" ], [ "0.2494", "0.1443", "0.294", "0.1443", "0.294", "0.1592", "0.2494", "0.1592" ], [ "0.33", "0.1455", "0.3687", "0.1455", "0.3687", "0.158", "0.33", "0.158" ], [ "0.4106", "0.1455", "0.4492", "0.1455", "0.4492", "0.158", "0.4106", "0.158" ], [ "0.5146", "0.1455", "0.8833", "0.1467", "0.8833", "0.1603", "0.5146", "0.1592" ], [ "0.2418", "0.1621", "0.3015", "0.1644", "0.3005", "0.1787", "0.241", "0.177" ], [ "0.409", "0.1621", "0.4517", "0.1644", "0.451", "0.1787", "0.4072", "0.177" ], [ "0.5156", "0.1609", "0.8833", "0.1609", "0.8833", "0.1746", "0.5156", "0.1746" ], [ "0.5146", "0.1733", "0.6963", "0.1758", "0.6963", "0.1895", "0.5146", "0.187" ], [ "0.33", "0.165", "0.3687", "0.165", "0.3687", "0.1776", "0.33", "0.1776" ], [ "0.1646", "0.1776", "0.2024", "0.1776", "0.2024", "0.1907", "0.1646", "0.1907" ], [ "0.2468", "0.1787", "0.2988", "0.1787", "0.2988", "0.1907", "0.2468", "0.1907" ], [ "0.33", "0.1787", "0.3687", "0.1787", "0.3687", "0.1907", "0.33", "0.1907" ], [ "0.4106", "0.1787", "0.4492", "0.1787", "0.4492", "0.1907", "0.4106", "0.1907" ], [ "0.2527", "0.1912", "0.294", "0.1912", "0.294", "0.2048", "0.2527", "0.2048" ], [ "0.33", "0.1912", "0.3687", "0.1912", "0.3687", "0.2037", "0.33", "0.2037" ], [ "0.4106", "0.1912", "0.4492", "0.1912", "0.4492", "0.2037", "0.4106", "0.2037" ], [ "0.2451", "0.209", "0.3005", "0.2114", "0.2988", "0.2257", "0.2435", "0.2239" ], [ "0.33", "0.212", "0.3687", "0.212", "0.3687", "0.2239", "0.33", "0.2239" ], [ "0.4106", "0.2114", "0.4492", "0.2114", "0.4492", "0.2239", "0.4106", "0.2239" ], [ "0.5225", "0.212", "0.882", "0.212", "0.882", "0.2227", "0.5225", "0.2227" ], [ "0.1646", "0.2239", "0.2024", "0.2239", "0.2024", "0.2363", "0.1646", "0.2363" ], [ "0.2468", "0.2245", "0.2988", "0.2245", "0.2988", "0.2363", "0.2468", "0.2363" ], [ "0.33", "0.2245", "0.3687", "0.2245", "0.3687", "0.2363", "0.33", "0.2363" ], [ "0.4106", "0.2239", "0.4492", "0.2239", "0.4492", "0.2363", "0.4106", "0.2363" ], [ "0.517", "0.2268", "0.8833", "0.2268", "0.8833", "0.2369", "0.517", "0.2369" ], [ "0.251", "0.2363", "0.2957", "0.2363", "0.2957", "0.2505", "0.251", "0.2505" ], [ "0.33", "0.2369", "0.3687", "0.2369", "0.3687", "0.2494", "0.33", "0.2494" ], [ "0.4106", "0.2369", "0.4492", "0.2369", "0.4492", "0.2494", "0.4106", "0.2494" ], [ "0.513", "0.2393", "0.885", "0.2393", "0.885", "0.253", "0.513", "0.253" ], [ "0.5146", "0.253", "0.882", "0.253", "0.882", "0.2666", "0.5146", "0.2666" ], [ "0.11755", "0.2642", "0.4895", "0.2642", "0.4895", "0.2778", "0.11755", "0.2778" ], [ "0.513", "0.2666", "0.881", "0.2666", "0.881", "0.2798", "0.513", "0.2798" ], [ "0.1184", "0.2798", "0.4895", "0.2798", "0.4895", "0.2935", "0.1184", "0.2935" ], [ "0.513", "0.2798", "0.882", "0.2808", "0.882", "0.2957", "0.513", "0.2944" ], [ "0.1184", "0.2944", "0.2527", "0.2944", "0.2527", "0.308", "0.1184", "0.308" ], [ "0.5146", "0.2944", "0.881", "0.2944", "0.881", "0.308", "0.5146", "0.308" ], [ "0.5156", "0.31", "0.885", "0.31", "0.885", "0.3237", "0.5156", "0.3237" ], [ "0.5283", "0.3237", "0.881", "0.3237", "0.881", "0.3374", "0.5283", "0.3374" ], [ "0.1604", "0.3333", "0.2418", "0.3333", "0.2418", "0.3445", "0.1604", "0.3445" ], [ "0.294", "0.332", "0.3477", "0.332", "0.3477", "0.3467", "0.294", "0.3467" ], [ "0.3997", "0.3333", "0.4492", "0.3333", "0.4492", "0.3445", "0.3997", "0.3445" ], [ "0.5156", "0.3384", "0.8833", "0.3384", "0.8833", "0.3516", "0.5156", "0.3516" ], [ "0.4048", "0.3508", "0.444", "0.3508", "0.444", "0.3665", "0.4048", "0.3665" ], [ "0.511", "0.3508", "0.882", "0.3516", "0.882", "0.3665", "0.511", "0.3652" ], [ "0.3015", "0.3528", "0.34", "0.3528", "0.34", "0.3652", "0.3015", "0.3652" ], [ "0.1838", "0.3516", "0.215", "0.3516", "0.215", "0.3665", "0.1838", "0.3665" ], [ "0.2913", "0.3706", "0.3494", "0.3723", "0.3477", "0.3872", "0.2896", "0.3855" ], [ "0.513", "0.3652", "0.882", "0.3665", "0.882", "0.3801", "0.513", "0.379" ], [ "0.4048", "0.3723", "0.4434", "0.3723", "0.4434", "0.3855", "0.4048", "0.3855" ], [ "0.4048", "0.3855", "0.4434", "0.3855", "0.4434", "0.398", "0.4048", "0.398" ], [ "0.5146", "0.3801", "0.8833", "0.3801", "0.8833", "0.3938", "0.5146", "0.3938" ], [ "0.1738", "0.386", "0.2217", "0.386", "0.2217", "0.398", "0.1738", "0.398" ], [ "0.2957", "0.386", "0.346", "0.386", "0.346", "0.398", "0.2957", "0.398" ], [ "0.403", "0.3967", "0.444", "0.3984", "0.4434", "0.4133", "0.4014", "0.411" ], [ "0.513", "0.3943", "0.882", "0.3943", "0.882", "0.408", "0.513", "0.408" ], [ "0.5146", "0.4092", "0.6675", "0.4092", "0.6675", "0.4229", "0.5146", "0.4229" ], [ "0.2988", "0.3984", "0.3435", "0.3984", "0.3435", "0.4133", "0.2988", "0.4133" ], [ "0.2913", "0.4163", "0.3494", "0.4187", "0.3477", "0.4329", "0.2896", "0.4312" ], [ "0.4048", "0.4187", "0.4434", "0.4187", "0.4434", "0.4312", "0.4048", "0.4312" ], [ "0.178", "0.4316", "0.22", "0.4316", "0.22", "0.4453", "0.178", "0.4453" ], [ "0.2957", "0.4329", "0.346", "0.4329", "0.346", "0.4448", "0.2957", "0.4448" ], [ "0.4048", "0.4316", "0.4434", "0.4316", "0.4434", "0.4448", "0.4048", "0.4448" ], [ "0.2988", "0.4448", "0.3435", "0.4448", "0.3435", "0.459", "0.2988", "0.459" ], [ "0.4048", "0.4453", "0.4434", "0.4453", "0.4434", "0.4578", "0.4048", "0.4578" ], [ "0.519", "0.4448", "0.882", "0.4448", "0.882", "0.4578", "0.519", "0.4578" ], [ "0.2913", "0.462", "0.351", "0.4644", "0.3494", "0.4785", "0.2896", "0.4768" ], [ "0.4048", "0.4644", "0.4434", "0.4644", "0.4434", "0.4768", "0.4048", "0.4768" ], [ "0.5156", "0.459", "0.8833", "0.459", "0.8833", "0.4727", "0.5156", "0.4727" ], [ "0.1738", "0.4756", "0.2217", "0.478", "0.22", "0.4922", "0.1721", "0.4905" ], [ "0.2957", "0.4768", "0.346", "0.4785", "0.345", "0.4905", "0.294", "0.488" ], [ "0.4048", "0.478", "0.4434", "0.478", "0.4434", "0.4905", "0.4048", "0.4905" ], [ "0.5156", "0.4739", "0.8833", "0.4739", "0.8833", "0.4868", "0.5156", "0.4868" ], [ "0.3005", "0.4912", "0.3418", "0.4912", "0.3418", "0.503", "0.3005", "0.503" ], [ "0.4055", "0.4912", "0.4434", "0.4912", "0.4434", "0.503", "0.4055", "0.503" ], [ "0.513", "0.4863", "0.885", "0.4868", "0.885", "0.502", "0.513", "0.5005" ], [ "0.5146", "0.502", "0.8833", "0.502", "0.8833", "0.5156", "0.5146", "0.5156" ], [ "0.1184", "0.5195", "0.4878", "0.5195", "0.4878", "0.533", "0.1184", "0.533" ], [ "0.5156", "0.5166", "0.882", "0.5166", "0.882", "0.53", "0.5156", "0.53" ], [ "0.11755", "0.533", "0.3745", "0.5337", "0.3745", "0.5474", "0.11755", "0.5464" ], [ "0.513", "0.5293", "0.8833", "0.53", "0.8833", "0.5444", "0.513", "0.5435" ], [ "0.5146", "0.546", "0.5425", "0.546", "0.5425", "0.558", "0.5146", "0.558" ], [ "0.1234", "0.5664", "0.4878", "0.5664", "0.4878", "0.58", "0.1234", "0.58" ], [ "0.11755", "0.579", "0.4878", "0.58", "0.4878", "0.594", "0.11755", "0.5933" ], [ "0.517", "0.5776", "0.8833", "0.579", "0.8833", "0.5933", "0.517", "0.593" ], [ "0.11755", "0.5933", "0.4902", "0.594", "0.4902", "0.6094", "0.11755", "0.608" ], [ "0.517", "0.5957", "0.882", "0.5957", "0.882", "0.6055", "0.517", "0.6055" ], [ "0.1184", "0.6094", "0.2612", "0.6094", "0.2612", "0.6226", "0.1184", "0.6226" ], [ "0.5156", "0.608", "0.882", "0.608", "0.882", "0.6216", "0.5156", "0.6216" ], [ "0.5156", "0.6216", "0.882", "0.6216", "0.882", "0.6357", "0.5156", "0.6357" ], [ "0.5146", "0.637", "0.882", "0.637", "0.882", "0.651", "0.5146", "0.651" ], [ "0.168", "0.654", "0.336", "0.654", "0.336", "0.6675", "0.168", "0.6675" ], [ "0.5146", "0.651", "0.8833", "0.651", "0.8833", "0.6646", "0.5146", "0.6646" ], [ "0.513", "0.6646", "0.6323", "0.6646", "0.6323", "0.6777", "0.513", "0.6777" ], [ "0.12006", "0.655", "0.1503", "0.655", "0.1503", "0.6675", "0.12006", "0.6675" ], [ "0.12427", "0.6816", "0.4878", "0.6816", "0.4878", "0.6953", "0.12427", "0.6953" ], [ "0.1184", "0.697", "0.4895", "0.697", "0.4895", "0.7104", "0.1184", "0.7104" ], [ "0.5156", "0.7017", "0.5483", "0.7017", "0.5483", "0.7134", "0.5156", "0.7134" ], [ "0.5415", "0.7007", "0.686", "0.7007", "0.686", "0.7144", "0.5415", "0.7144" ], [ "0.12006", "0.712", "0.486", "0.712", "0.486", "0.7227", "0.12006", "0.7227" ], [ "0.11755", "0.7236", "0.4878", "0.7246", "0.4878", "0.7393", "0.11755", "0.7383" ], [ "0.517", "0.7236", "0.8833", "0.7246", "0.8833", "0.7393", "0.517", "0.7383" ], [ "0.12177", "0.7393", "0.4895", "0.7393", "0.4895", "0.753", "0.12177", "0.753" ], [ "0.5146", "0.7393", "0.8833", "0.7393", "0.8833", "0.753", "0.5146", "0.753" ], [ "0.12006", "0.7544", "0.4895", "0.7544", "0.4895", "0.767", "0.12006", "0.767" ], [ "0.5156", "0.7544", "0.8833", "0.7544", "0.8833", "0.767", "0.5156", "0.767" ], [ "0.1184", "0.7666", "0.4878", "0.767", "0.4878", "0.782", "0.1184", "0.781" ], [ "0.513", "0.7666", "0.882", "0.767", "0.882", "0.782", "0.513", "0.781" ], [ "0.1184", "0.782", "0.4902", "0.782", "0.4902", "0.796", "0.1184", "0.796" ], [ "0.5146", "0.782", "0.881", "0.782", "0.881", "0.796", "0.5146", "0.796" ], [ "0.12006", "0.7964", "0.492", "0.7964", "0.492", "0.81", "0.12006", "0.81" ], [ "0.513", "0.7964", "0.8833", "0.7964", "0.8833", "0.81", "0.513", "0.81" ], [ "0.12006", "0.81", "0.4895", "0.81", "0.4895", "0.8237", "0.12006", "0.8237" ], [ "0.513", "0.81", "0.882", "0.81", "0.882", "0.8237", "0.513", "0.8237" ], [ "0.12006", "0.826", "0.4878", "0.826", "0.4878", "0.836", "0.12006", "0.836" ], [ "0.5156", "0.826", "0.881", "0.826", "0.881", "0.836", "0.5156", "0.836" ], [ "0.1184", "0.8384", "0.4878", "0.8384", "0.4878", "0.8516", "0.1184", "0.8516" ], [ "0.5146", "0.8384", "0.8833", "0.8384", "0.8833", "0.8516", "0.5146", "0.8516" ], [ "0.12177", "0.854", "0.4878", "0.854", "0.4878", "0.8643", "0.12177", "0.8643" ], [ "0.5146", "0.8525", "0.882", "0.8525", "0.882", "0.866", "0.5146", "0.866" ], [ "0.1184", "0.8677", "0.4895", "0.8677", "0.4895", "0.8813", "0.1184", "0.8813" ], [ "0.513", "0.8677", "0.882", "0.8677", "0.882", "0.8813", "0.513", "0.8813" ], [ "0.12177", "0.883", "0.4878", "0.883", "0.4878", "0.8936", "0.12177", "0.8936" ], [ "0.5146", "0.88", "0.8833", "0.8813", "0.8833", "0.8955", "0.5146", "0.8945" ], [ "0.12006", "0.8955", "0.3167", "0.8955", "0.3167", "0.909", "0.12006", "0.909" ], [ "0.5146", "0.898", "0.881", "0.898", "0.881", "0.9077", "0.5146", "0.9077" ] ]

[ "Training data", "Method", "Flickr30k", "MSCOCO", "VQA在VQA中，我们还对0进行采样。25%0。", "5%和0。75来自（Geetal.,2023）之后的训练集", "0 %", "CLIP", "83.00", "53.30", "的%。结果见表2。与图像到文本检索类似，我", "82.80", "53.50", "们的方法在所有设置中都优于CLIP，CoCoOp和", "55.80", "CoCoOp", "0.5%", "CoT-PT", "83.50", "CoT-PT。AGoT获得0。51%,0。59%和0。88%分", "AGoT", "84.30", "57.10", "别比CoT-PT增加。实验表明，由于更好的推理和", "CoCoOp", "56.40", "多模态表示学习，我们的AGoT方法还可以使相对", "更复杂的VQA任务受益", "84.50", "1.0 %", "CoT-PT", "85.10", "56.70", "AGoT", "87.60", "57.60", "CoCoOp", "85.00", "57.00", "CrossLabelGeneralization除了在多模态理解", "2.0 %", "CoT-PT", "86.00", "57.90", "任务上进行评估外，基于先前工作的结论E（Zhou", "AGoT", "88.70", "58.70", "etal.，2022a），更好的提示方法可以导致更好的域", "泛化能力。因此，我们首先在交叉标签泛化设置上", "Table 1: Comparison between CLIP, CoCoOp, CoT", "进行了实验。具体来说，我们在11个数据集上使用", "PT, and our AGoT methods on the Flickr30k and", "AGoT进行了实验，其中每个数据集分为基类和新", "MSCOCO datasets.", "类，以表示源域和目标域。为了训练模型，我们考", "虑了每个类16个样本，并设置了推理步骤Z=5", "表3中总结的结果表明，AGoT优于其他模型", "Training data", "Method", "VQAv2", "例如CLIP，CoOp，CoCoOp，ProGrad，KgCoOp和", "11.83", "CoT-PT。具体来说，AToT实现了5的精度提升。", "CLIP", "0 %", "CoCoOp", "98%,3。08%,1。85%,1。52%,0。68%和0。58%分", "27.23", "29.13", "别在这些强基线上。这些实验表明，当遇到新数据", "0.25 %", "CoT-PT", "29.64", "时，AGoT表现出很强的泛化性能，这是由思想推", "理的多个方面引起的", "AGoT", "CoCoOp", "29.51", "0.5 %", "CoT-PT", "30.72", "AGoT", "31.31", "CrossDatasetGeneralization除了跨标签泛化之", "CoCoOp", "30.76", "外，我们还探讨了跨数据集泛化设置。如表4所", "0.75 %", "CoT-PT", "30.86", "示，AGoT的性能比CoOp高出0。61ImageNet上", "AGoT", "31.74", "的%（源代码）。此外，当将模型应用于10个数据", "集的目标域时，我们的结果表明0。96%对以前最", "Table 2: Comparison between CLIP, CoCoOp, CoT-PT,", "先进的KgCoOp的改进。这些实验发现也为与普通", "andourAGoTontheVQAv2dataset.", "CoT相比，AGoT的强大可转移性提供了坚实的证", "据。", "Baselines我们将我们提出的方法与以下最先进", "的模型进行了比较：带有固定手工制作提示的零拍", "DomainGeneralization表5显示了域泛化的结", "CLIP“一张照片”（类)，CoOp，CoCoOp，ProGrad，", "果。我们在ImageNet上训练的模型在其他四个不", "KgCoOp和 CoT-PT", "同类型的ImageNet数据集上进行了测试，我们的", "方法优于CLIP，UPT，CoCoOp，CoOp，ProGrad", "KgCoOp。这证实了CoT对领域泛化确实有很好的", "Results and Analysis", "效果，而AGoT是对CoT的有效改进，带来了更强", "的推理泛化能力", "4.2.", "Text-ImageRetrieval在本实验中，我们在两个", "图像字幕数据集上评估了我们的模型：Flickr30k和", "4.3.", "AblationStudy", "MSCOCO。我们使用O的设置训练和测试了我们的", "模型。5%,1。0%和2。0%的训练数据，并在测试集", "ReasoningSteps为了演示AGoT中推理步骤的", "上对其进行评估。从表1可以看出，我们的方法在", "影响，我们使用训练比为2的Flickr30K数据集对", "所有测试中都优于CLIP、CoCoOp和 CoT-PT。在", "图像-文本检索任务进行了实验。0%。具体来说，我", "Flickr3Ok和MSCOCO数据集上，当训练数据仅由", "们将推理步骤从2改变为9。如图E3所示，结果", "0。5%的总数据，AGoT执行了0。80%和1。30%", "表明，当纪元设置为80并且推理步骤设置为3时", "优于目前的SOTA方法CoT-PT；当训练数据为1。", "AGoT达到最佳性能。这可以归因于以下事实：当", "0%的总数据，AGoT执行了2。5%和0。9%优于", "数据大小较小时，AGoT中的较长的推理步骤可能", "CoT-PT：而当训练数据为2。0%的总数据，AGoT", "不稳定，并且容易出现过拟合。因此，增加推理步", "执行了1。7%和0。8%分别优于CoT-PT。这些改", "骤并不总是能提高性能。如表E6所示，继续增加", "进表明，我们提出的AGoT对于多模态表示学习比", "训练样本的比例，当训练样本比例达到5。0%，推", "CoT更有效，CoT是以前的最新技术。此外，我们", "理步骤5的性能优于推理步骤3，实现了最佳性能", "提出的新颖的软提示有助于跨模态表示学习，而无", "同样，当样本比例增加到8时，也得出了同样的结", "需调整CLIP的预训练参数", "论。0%。这表明随着数据集变得相对更大，任务变" ]

{ "arxiv_id": "2404.04538", "language": "zh", "page_idx": 4 }

[ [ "0.232", "0.1445", "0.7715", "0.1459", "0.7715", "0.1635", "0.232", "0.1622" ], [ "0.3718", "0.1666", "0.6323", "0.1674", "0.6323", "0.1849", "0.3718", "0.1843" ], [ "0.3447", "0.2153", "0.66", "0.2153", "0.66", "0.2291", "0.3447", "0.2291" ], [ "0.312", "0.2405", "0.6895", "0.2412", "0.6895", "0.2556", "0.312", "0.2551" ], [ "0.3586", "0.2556", "0.642", "0.2556", "0.642", "0.2695", "0.3586", "0.2695" ], [ "0.3635", "0.2708", "0.6396", "0.2708", "0.6396", "0.2842", "0.3635", "0.2842" ], [ "0.2688", "0.3257", "0.5605", "0.3257", "0.5605", "0.336", "0.2688", "0.336" ], [ "0.5605", "0.3245", "0.739", "0.3245", "0.739", "0.3384", "0.5605", "0.3384" ], [ "0.2656", "0.3384", "0.739", "0.3384", "0.739", "0.3528", "0.2656", "0.3528" ], [ "0.2646", "0.3528", "0.739", "0.3528", "0.739", "0.3662", "0.2646", "0.3662" ], [ "0.2656", "0.3662", "0.739", "0.3662", "0.739", "0.3801", "0.2656", "0.3801" ], [ "0.2656", "0.3801", "0.738", "0.3801", "0.738", "0.3933", "0.2656", "0.3933" ], [ "0.2656", "0.3933", "0.739", "0.3933", "0.739", "0.4072", "0.2656", "0.4072" ], [ "0.2656", "0.4077", "0.739", "0.4077", "0.739", "0.4216", "0.2656", "0.4216" ], [ "0.2656", "0.4216", "0.739", "0.4216", "0.739", "0.435", "0.2656", "0.435" ], [ "0.2656", "0.435", "0.739", "0.435", "0.739", "0.449", "0.2656", "0.449" ], [ "0.2688", "0.4539", "0.327", "0.4539", "0.327", "0.461", "0.2688", "0.461" ], [ "0.3325", "0.4507", "0.738", "0.4507", "0.738", "0.461", "0.3325", "0.461" ], [ "0.267", "0.4653", "0.71", "0.4653", "0.71", "0.476", "0.267", "0.476" ], [ "0.2157", "0.504", "0.3782", "0.505", "0.3782", "0.521", "0.2157", "0.5195" ], [ "0.2197", "0.5386", "0.785", "0.5386", "0.785", "0.552", "0.2197", "0.552" ], [ "0.2197", "0.5537", "0.7837", "0.5537", "0.7837", "0.5674", "0.2197", "0.5674" ], [ "0.2214", "0.571", "0.7837", "0.571", "0.7837", "0.5806", "0.2214", "0.5806" ], [ "0.2214", "0.5854", "0.7876", "0.5854", "0.7876", "0.5986", "0.2214", "0.5986" ], [ "0.2197", "0.5996", "0.694", "0.5996", "0.694", "0.613", "0.2197", "0.613" ], [ "0.2443", "0.6143", "0.785", "0.6143", "0.785", "0.6274", "0.2443", "0.6274" ], [ "0.2214", "0.63", "0.7837", "0.63", "0.7837", "0.643", "0.2214", "0.643" ], [ "0.2214", "0.6465", "0.7837", "0.6465", "0.7837", "0.657", "0.2214", "0.657" ], [ "0.2197", "0.66", "0.785", "0.66", "0.785", "0.674", "0.2197", "0.674" ], [ "0.2197", "0.6753", "0.785", "0.6753", "0.785", "0.69", "0.2197", "0.69" ], [ "0.2214", "0.6924", "0.6436", "0.6924", "0.6436", "0.7026", "0.2214", "0.7026" ], [ "0.2427", "0.705", "0.7837", "0.705", "0.7837", "0.7183", "0.2427", "0.7183" ], [ "0.2197", "0.72", "0.785", "0.72", "0.785", "0.7344", "0.2197", "0.7344" ], [ "0.2214", "0.7373", "0.7817", "0.7373", "0.7817", "0.7476", "0.2214", "0.7476" ], [ "0.2197", "0.752", "0.7837", "0.752", "0.7837", "0.765", "0.2197", "0.765" ], [ "0.2197", "0.7666", "0.7837", "0.7666", "0.7837", "0.78", "0.2197", "0.78" ], [ "0.2222", "0.7827", "0.7837", "0.7827", "0.7837", "0.7935", "0.2222", "0.7935" ], [ "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" ], [ "0.2181", "0.827", "0.3677", "0.827", "0.3677", "0.8403", "0.2181", "0.8403" ] ]

[ "Multi-Instance Learning by Treating Instances", "As Non-I.I.D. Samples", "Zhi-Hua Zhou, Yu-Yin Sun, and Yu-Feng Li", "National Key Laboratory for Novel Software Technology,", "Nanjing University, Nanjing 210093, China", "[zhouzh, sunyy, liyf] @lamda.nju.edu.cn", "Abstract.Previousstudiesonmulti-instano", "celearningtypicallytreated", "instances in the bags as independently and identically distributed : The", "instances in a bag, however, are rarely independent in real tasks, and", "a better performance can be expected if the instances are treated in", "an non-i.i.d. way that exploits relations among instances. In this paper", "we propose two simple yet effective methods. In the first method, we", "explicitly map every bag to an undirected graph and design a graph", "kernel for distinguishing the positive and negative bags. In the second", "method, we implicitly construct graphs by deriving affinity matrices and", "proposea", "nefficientgraphkernelconsideringthecliqueinformation.The", "effectiveness of theproposedmethodsarevalidated by experiments", "1Introduction", "在多实例学习11]中，每个训练示例都是一袋实例。如果包至少包含一个正实", "例，则为正，否则为负。尽管训练袋的标签是已知的，但是，袋中实例的标签是", "未知的。自标是构建一个学习者来对看不见的袋子进行分类。多实例学习已被", "发现在不同领域有用，例如图像分类[6,7]，图像检索[35]；文本分类[2,24]，", "计算机安全[22]，面部检测[27,32]，计算机辅助医疗诊断[12]等", "多实例学习的一个突出优势主要在于许多真实对象具有固有的结构，通过采", "用多实例表示，我们能够比简单地使用平面单实例表示更自然地表示此类对象", "并捕获更多信息。例如，假设我们可以将图像划分为儿个部分。与将整个图像表", "示为单实例相反，如果我们将每个部分表示为一个实例，则分区信息由多实例", "表示捕获；并且如果分区是有意义的（e。g。，每个部分对应一个显著区域)，多", "实例表示捕获的附加信息可能有助于使学习任务更容易处理", "在任何地方应用多实例学习技术显然不是一个好主意，因为如果单实例表示", "就足够了，那么使用多实例表示只会使百合花镀金。即使在对象具有固有结构", "的任务上，我们也应该记住，多实例表示的强大之处在于其捕获某些结构信息", "的能力。但是，正如Zhou和Xu[36]所指出的那样，以前关于多实例学习的研", "究通常将包中的实例视为独立且相同的分布；这忽略了以下事实：实例之间的", "关系传达了重要的结构信息。再次考虑上述图像任务，将不同的图像部分视为", "相互关联的样本显然比将它们视为不相关的样本更有意义。实际上，包中的实", "例很少是独立的，如果在非i中处理实例，则可以预期更好的性能。i。d。利用", "实例之间关系的方式" ]

{ "arxiv_id": "0807.1997", "language": "zh", "page_idx": 0 }

[ [ "0.2443", "0.1483", "0.785", "0.1483", "0.785", "0.1622", "0.2443", "0.1622" ], [ "0.2214", "0.1654", "0.7817", "0.1654", "0.7817", "0.1755", "0.2214", "0.1755" ], [ "0.2214", "0.1799", "0.7837", "0.1799", "0.7837", "0.1901", "0.2214", "0.1901" ], [ "0.2181", "0.1932", "0.4004", "0.1932", "0.4004", "0.207", "0.2181", "0.207" ], [ "0.2443", "0.214", "0.785", "0.214", "0.785", "0.2279", "0.2443", "0.2279" ], [ "0.2197", "0.2299", "0.778", "0.2299", "0.778", "0.2437", "0.2197", "0.2437" ], [ "0.2173", "0.2974", "0.392", "0.2974", "0.392", "0.3113", "0.2173", "0.3113" ], [ "0.2197", "0.3567", "0.785", "0.3567", "0.785", "0.3706", "0.2197", "0.3706" ], [ "0.2296", "0.3718", "0.785", "0.3718", "0.785", "0.3865", "0.2296", "0.3865" ], [ "0.2197", "0.3865", "0.6772", "0.3865", "0.6772", "0.3997", "0.2197", "0.3997" ], [ "0.2443", "0.4077", "0.785", "0.4077", "0.785", "0.4216", "0.2443", "0.4216" ], [ "0.2197", "0.4224", "0.785", "0.4224", "0.785", "0.4363", "0.2197", "0.4363" ], [ "0.2181", "0.4363", "0.785", "0.4375", "0.785", "0.452", "0.2181", "0.4507" ], [ "0.2214", "0.4526", "0.785", "0.4526", "0.785", "0.4666", "0.2214", "0.4666" ], [ "0.2222", "0.4697", "0.7837", "0.4697", "0.7837", "0.4797", "0.2222", "0.4797" ], [ "0.2214", "0.4841", "0.7837", "0.4841", "0.7837", "0.4976", "0.2214", "0.4976" ], [ "0.2214", "0.5", "0.7837", "0.5", "0.7837", "0.5103", "0.2214", "0.5103" ], [ "0.2181", "0.513", "0.785", "0.512", "0.785", "0.526", "0.2181", "0.5264" ], [ "0.2197", "0.529", "0.785", "0.529", "0.785", "0.5425", "0.2197", "0.5425" ], [ "0.2197", "0.5435", "0.7837", "0.5435", "0.7837", "0.5566", "0.2197", "0.5566" ], [ "0.2197", "0.5566", "0.785", "0.558", "0.785", "0.5713", "0.2197", "0.571" ], [ "0.2181", "0.5737", "0.7837", "0.5737", "0.7837", "0.587", "0.2181", "0.587" ], [ "0.2197", "0.589", "0.786", "0.589", "0.786", "0.603", "0.2197", "0.603" ], [ "0.2197", "0.604", "0.626", "0.604", "0.626", "0.617", "0.2197", "0.617" ], [ "0.2427", "0.6235", "0.7837", "0.6235", "0.7837", "0.637", "0.2427", "0.637" ], [ "0.2197", "0.638", "0.785", "0.6396", "0.785", "0.654", "0.2197", "0.653" ], [ "0.2181", "0.654", "0.7837", "0.654", "0.7837", "0.6675", "0.2181", "0.6675" ], [ "0.2222", "0.672", "0.7817", "0.672", "0.7817", "0.6816", "0.2222", "0.6816" ], [ "0.2345", "0.686", "0.7837", "0.686", "0.7837", "0.6997", "0.2345", "0.6997" ], [ "0.2214", "0.7007", "0.7837", "0.7007", "0.7837", "0.714", "0.2214", "0.714" ], [ "0.2214", "0.7163", "0.7817", "0.7163", "0.7817", "0.7266", "0.2214", "0.7266" ], [ "0.2214", "0.732", "0.7837", "0.732", "0.7837", "0.742", "0.2214", "0.742" ], [ "0.2427", "0.75", "0.7837", "0.752", "0.7837", "0.765", "0.2427", "0.763" ], [ "0.2214", "0.7666", "0.785", "0.7666", "0.785", "0.78", "0.2214", "0.78" ], [ "0.2214", "0.781", "0.785", "0.781", "0.785", "0.7944", "0.2214", "0.7944" ], [ "0.2197", "0.797", "0.785", "0.797", "0.785", "0.81", "0.2197", "0.81" ], [ "0.2197", "0.8125", "0.785", "0.8125", "0.785", "0.8257", "0.2197", "0.8257" ], [ "0.2197", "0.827", "0.5786", "0.827", "0.5786", "0.8403", "0.2197", "0.8403" ] ]

[ "在本文中，我们提出了两种不将实例视为i的多实例学习方法。i。d。样品。", "我们的基本思想是将每个包视为要作为一个整体处理的实体，并将实例视为该", "实体的相互关联的组件。实验表明，我们提出的方法实现了与最先进的多示例", "学习方法高度竞争的性能", "本文的其余部分组织如下。我们在第2节中简要回顾了相关工作，在第3节", "中提出了新方法，在第4节中报告了我们的实验，最后在第5节中总结了论文", "2Related Work", "在过去的十年中，已经开发了许多多实例学习方法。仅举几例，多元密度[16]", "k-最近邻算法引用-NN[29]，决策树RELIC[22]和MITI[4]，神经网络", "BP-MIP和RBF-MIP[33]，规则学习算法BooRXIPPLINXMI", "许多研究人员已经研究了用于多示例学习的内核方法。Gartner等人。[14]", "通过将每个包视为一组特征向量，然后直接应用集合内核来定义MI内核。", "Andrews等人。[2]提出的mi-SVM和MI-SVM。mi-SVM尝试识别实例的最", "大边距超平面，但要遵守以下约束：每个正包的至少一个实例位于正半空间中，", "而所有负包的实例都位于负半空间中：MI-SVM尝试通过将包中“最积极实例”", "的边距视为该包的边距来识别包的最大边距超平面。Cheung和Kwok[8]认为", "最积极实例的符号而不是价值很重要。他们定义了一个损失函数：充许袋子和", "实例参与优化过程，并使用格式良好的约束凹凸程序来执行优化。后来，Kwok", "和Cheung[15]通过将生成模型纳入内核设计中，设计了边缘化的多实例内核。", "Chen和Wang[7]提出了DD-SVM方法，该方法使用DiverseDensity[16]来学", "习一组实例原型，然后将包映射到基于实例原型的特征空间。Zhou和Xu[36]", "提出了MissSVM方法，将阴性袋的实例作为标记示例，而将阳性袋的实例作为", "具有正约束的未标记示例。Wang等人。[28]通过将每个袋子表示为基于未标记", "数据得出的混合模型的一些聚合后验来提出PPMM内核", "除了分类之外，还研究了多实例回归[1,21]，并定义了不同版本的广义多实", "例学习[30,23]。标准多实例学习与广义多实例学习的主要区别在于，在标准多", "实例学习中有一个单一的概念，如果有一个实例满足这个概念，那么一个包就是", "正的：而在广义多实例学习30.23中有多个概念，只有当所有的概念都满足时", "个包才是正的(i。e。，包中包含来自每个概念的实例）。最近，关于多示例半", "监督学习[19]，多示例主动学习[24]和多示例多标签学习[37]的研究也已被报", "道。在本文中，我们主要研究标准的多实例学习11并将表明我们的方法也", "适用于多实例回归。实际上，也可以将我们的建议扩展到多实例学习的其他变体", "Zhou和Xu[36]指出，不应将包中的实例视为i。i。d。样本，本文提供了一", "种解决方案。我们的基本思想是将每个袋子视为一个整体来处理。有其他方法", "可以实现这个想法，而在本文中，我们将每个袋子视为一个图形。McGovern和", "Jensen[17]将多实例学习作为处理关系数据的工具，其中每个实例都以图形形", "式给出。在这里，我们正在研究命题数据，没有自然图。与将实例作为图相比，", "我们将每个包视为图，将每个实例视为图中的节点" ]

{ "arxiv_id": "0807.1997", "language": "zh", "page_idx": 1 }

[ [ "0.2239", "0.2247", "0.7754", "0.2247", "0.7754", "0.238", "0.2239", "0.238" ], [ "0.3447", "0.2644", "0.3628", "0.2644", "0.3628", "0.2715", "0.3447", "0.2715" ], [ "0.5073", "0.2615", "0.5254", "0.2615", "0.5254", "0.2708", "0.5073", "0.2708" ], [ "0.6465", "0.2644", "0.664", "0.2676", "0.661", "0.2734", "0.6436", "0.2708" ], [ "0.5215", "0.2734", "0.539", "0.2734", "0.539", "0.281", "0.5215", "0.281" ], [ "0.4968", "0.2798", "0.5146", "0.2727", "0.5195", "0.2822", "0.5024", "0.289" ], [ "0.6567", "0.2798", "0.665", "0.286", "0.6567", "0.2922", "0.649", "0.286" ], [ "0.517", "0.2966", "0.5317", "0.2966", "0.5317", "0.3037", "0.517", "0.3037" ], [ "0.6465", "0.2966", "0.665", "0.2966", "0.665", "0.3093", "0.6465", "0.3093" ], [ "0.4844", "0.2974", "0.501", "0.2974", "0.501", "0.305", "0.4844", "0.305" ], [ "0.3694", "0.3005", "0.3782", "0.2954", "0.384", "0.3005", "0.3743", "0.3062" ], [ "0.2181", "0.3245", "0.785", "0.3257", "0.785", "0.3403", "0.2181", "0.339" ], [ "0.2197", "0.3416", "0.739", "0.3416", "0.739", "0.3516", "0.2197", "0.3516" ], [ "0.2173", "0.3833", "0.4836", "0.3833", "0.4836", "0.3977", "0.2173", "0.3977" ], [ "0.2197", "0.4172", "0.7837", "0.4172", "0.7837", "0.4312", "0.2197", "0.4312" ], [ "0.2214", "0.4343", "0.7837", "0.4343", "0.7837", "0.4443", "0.2214", "0.4443" ], [ "0.2181", "0.4475", "0.6587", "0.4475", "0.6587", "0.461", "0.2181", "0.461" ], [ "0.2443", "0.4622", "0.785", "0.4622", "0.785", "0.476", "0.2443", "0.476" ], [ "0.2197", "0.4778", "0.7837", "0.4778", "0.7837", "0.4912", "0.2197", "0.4912" ], [ "0.2214", "0.4937", "0.7837", "0.4937", "0.7837", "0.507", "0.2214", "0.507" ], [ "0.2222", "0.507", "0.785", "0.507", "0.785", "0.5215", "0.2222", "0.5215" ], [ "0.2197", "0.5234", "0.7817", "0.523", "0.7817", "0.537", "0.2197", "0.5386" ], [ "0.2197", "0.5386", "0.7837", "0.5386", "0.7837", "0.553", "0.2197", "0.553" ], [ "0.2181", "0.552", "0.7837", "0.553", "0.7837", "0.5674", "0.2181", "0.5664" ], [ "0.2443", "0.5693", "0.7817", "0.5693", "0.7817", "0.5835", "0.2443", "0.5835" ], [ "0.2197", "0.584", "0.7837", "0.584", "0.7837", "0.5977", "0.2197", "0.5977" ], [ "0.2214", "0.5996", "0.7812", "0.5996", "0.7812", "0.61", "0.2214", "0.61" ], [ "0.2214", "0.6143", "0.7837", "0.6143", "0.7837", "0.6274", "0.2214", "0.6274" ], [ "0.2222", "0.631", "0.7837", "0.631", "0.7837", "0.6416", "0.2222", "0.6416" ], [ "0.2214", "0.6445", "0.7837", "0.6445", "0.7837", "0.6577", "0.2214", "0.6577" ], [ "0.2214", "0.659", "0.7837", "0.659", "0.7837", "0.6724", "0.2214", "0.6724" ], [ "0.2222", "0.675", "0.7837", "0.675", "0.7837", "0.688", "0.2222", "0.688" ], [ "0.2214", "0.69", "0.785", "0.69", "0.785", "0.704", "0.2214", "0.704" ], [ "0.2214", "0.705", "0.785", "0.705", "0.785", "0.7183", "0.2214", "0.7183" ], [ "0.2197", "0.72", "0.773", "0.72", "0.773", "0.7344", "0.2197", "0.7344" ], [ "0.246", "0.735", "0.785", "0.735", "0.785", "0.7485", "0.246", "0.7485" ], [ "0.2214", "0.7505", "0.7837", "0.7505", "0.7837", "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", "0.785", "0.781", "0.785", "0.7954", "0.2214", "0.7954" ], [ "0.2239", "0.7974", "0.7817", "0.7974", "0.7817", "0.808", "0.2239", "0.808" ], [ "0.2214", "0.811", "0.7837", "0.811", "0.7837", "0.8247", "0.2214", "0.8247" ], [ "0.2197", "0.827", "0.7837", "0.827", "0.7837", "0.8403", "0.2197", "0.8403" ] ]

[ "Fig.1. Example images with six marked patches each corresponding to an instance", "", "口", "", "", "A", "", "国", "口", "", "", "Fig.2. If we do not consider the relations among the instances, the three bags are", "similar to each other sincethey have identical number of very similar instances", "3The Proposed Methods", "在本节中，我们提出了MIGraph和miGraph方法。MIGraph方法将每个包显", "式映射到一个无向图，并使用新的图内核来区分正负包。miGraph方法通过导", "出亲和矩阵来隐式构造图，并定义了考虑集团信息的有效图核", "在介绍细节之前，我们给出多实例学习的正式定义如下。令表示实例", "空间。给定一个数据集{(Xi,yi)，.，(X,yi)，.，（XN,yN))，其中X，=", "[ai,,aii，,inC称为袋子，yiE={-l,+l}是X,的标签", "目标是生成一个学习者来对看不见的袋子进行分类。这里&iE是实例", "[il，..，ail，.，aijd]”，Til是ai在第l个属性处的值，N是训练包数，n", "是X，中的实例数，d是属性数。如果存在gE{1,.·，ni}，使得aig是正实例", "则X，是正包，因此yi=+1；否则yi=-1。然而，指数g的具体值是未知的", "我们首先解释我们对所提出方法的直觉。在这里，我们使用图3所示的三个", "示例图像。1用于说明。为简单起见，我们在每个图中显示六个标记的补丁", "并假设每个图像对应一个袋子，每个补工对应袋子中的一个实例，并日标记的具", "有相同颜色的补丁非常相似（真实案例当然更复杂，但要点与插图相似）。如果", "实例被视为独立样本，则图。1可以抽象为图。2：这是以前的多实例学习研", "究的典型方式，显然这三个包彼此相似，因为它们包含相同数量的非常相似的实", "例。但是，如果我们考虑实例之间的关系，我们可以发现，在前两个袋子中，蓝", "色标记彼此非常接近，而在第三个袋子中，蓝色标记分散在橙色标记中，因此前", "两个袋子应该比第三个袋子更相似。在这种情况下，图。1可以用图抽象。3。", "很明显，图中的抽象。3比图中更可取。2。这里的本质是，属于同一类别的", "袋子的关系结构相对更相似，而属于不同类别的袋子的关系结构相对更不相似", "现在我们描述MIGraph方法。第一步是为每个包构建一个图。受[26]的启", "发，它表明-graph有助于发现数据的底层流形结构，在这里我们为每个包建", "立一个e-graph。这个过程很简单。对于包X，，我们将其每个实例视为节点", "然后，我们计算每对节点的距离，e。g。、aiu和aiu。如果aiu和Ciu之间的距", "离小于预设國值E，则在这两个节点之间建立一条边，其中边的权重表示两个节", "点的亲和力（在实验中我们使用非零距离的归一化倒数作为亲和力值）。可以使", "用许多距离度量来计算距离。根据流形性质[26]，i。e。，一个小的局部区域大" ]

{ "arxiv_id": "0807.1997", "language": "zh", "page_idx": 2 }

[ [ "0.5073", "0.1547", "0.523", "0.1547", "0.523", "0.1635", "0.5073", "0.1635" ], [ "0.492", "0.1692", "0.517", "0.1692", "0.517", "0.178", "0.492", "0.178" ], [ "0.5195", "0.1686", "0.5376", "0.1686", "0.5376", "0.1724", "0.5195", "0.1724" ], [ "0.2181", "0.2184", "0.786", "0.2184", "0.786", "0.2317", "0.2181", "0.2317" ], [ "0.2197", "0.2343", "0.7837", "0.2343", "0.7837", "0.2443", "0.2197", "0.2443" ], [ "0.2181", "0.2474", "0.281", "0.2474", "0.281", "0.2588", "0.2181", "0.2588" ], [ "0.2197", "0.2874", "0.7837", "0.2874", "0.7837", "0.3005", "0.2197", "0.3005" ], [ "0.2214", "0.3018", "0.7837", "0.3018", "0.7837", "0.315", "0.2214", "0.315" ], [ "0.2214", "0.3176", "0.7837", "0.3176", "0.7837", "0.3308", "0.2214", "0.3308" ], [ "0.2181", "0.332", "0.786", "0.3308", "0.786", "0.3484", "0.2181", "0.3499" ], [ "0.2214", "0.3499", "0.615", "0.3499", "0.615", "0.3643", "0.2214", "0.3643" ], [ "0.5034", "0.3818", "0.5107", "0.3818", "0.5107", "0.387", "0.5034", "0.387" ], [ "0.505", "0.377", "0.642", "0.3757", "0.642", "0.3945", "0.505", "0.3958" ], [ "0.4878", "0.3833", "0.4944", "0.3833", "0.4944", "0.387", "0.4878", "0.387" ], [ "0.761", "0.385", "0.7876", "0.385", "0.7876", "0.4028", "0.761", "0.4028" ], [ "0.4756", "0.3926", "0.597", "0.3945", "0.597", "0.4124", "0.4756", "0.4104" ], [ "0.339", "0.3865", "0.4697", "0.3865", "0.4697", "0.401", "0.339", "0.401" ], [ "0.5884", "0.3926", "0.6406", "0.3977", "0.638", "0.4136", "0.585", "0.4077" ], [ "0.2214", "0.4216", "0.785", "0.4216", "0.785", "0.435", "0.2214", "0.435" ], [ "0.2214", "0.4375", "0.2974", "0.4375", "0.2974", "0.449", "0.2214", "0.449" ], [ "0.2427", "0.452", "0.7837", "0.452", "0.7837", "0.4653", "0.2427", "0.4653" ], [ "0.2197", "0.4666", "0.7817", "0.4653", "0.7817", "0.4797", "0.2197", "0.481" ], [ "0.2214", "0.4817", "0.785", "0.4817", "0.785", "0.4956", "0.2214", "0.4956" ], [ "0.2173", "0.4956", "0.66", "0.4968", "0.66", "0.511", "0.2173", "0.5103" ], [ "0.2443", "0.512", "0.7837", "0.512", "0.7837", "0.526", "0.2443", "0.526" ], [ "0.2214", "0.528", "0.785", "0.528", "0.785", "0.5425", "0.2214", "0.5425" ], [ "0.2197", "0.5415", "0.785", "0.5425", "0.785", "0.5566", "0.2197", "0.556" ], [ "0.2197", "0.558", "0.3325", "0.558", "0.3325", "0.5713", "0.2197", "0.5713" ], [ "0.2197", "0.5835", "0.785", "0.5835", "0.785", "0.5967", "0.2197", "0.5967" ], [ "0.2197", "0.5986", "0.7876", "0.5986", "0.7876", "0.613", "0.2197", "0.613" ], [ "0.2181", "0.6123", "0.4836", "0.613", "0.4836", "0.6274", "0.2181", "0.627" ], [ "0.343", "0.6353", "0.66", "0.6353", "0.66", "0.6665", "0.343", "0.6665" ], [ "0.429", "0.6636", "0.6587", "0.66", "0.6587", "0.688", "0.429", "0.6914" ], [ "0.761", "0.6675", "0.7876", "0.6675", "0.7876", "0.685", "0.761", "0.685" ], [ "0.2214", "0.6987", "0.785", "0.6987", "0.785", "0.7134", "0.2214", "0.7134" ], [ "0.2173", "0.712", "0.398", "0.7134", "0.398", "0.728", "0.2173", "0.7266" ], [ "0.5063", "0.738", "0.5884", "0.738", "0.5884", "0.7524", "0.5063", "0.7524" ], [ "0.3489", "0.747", "0.4412", "0.747", "0.4412", "0.761", "0.3489", "0.761" ], [ "0.763", "0.747", "0.786", "0.747", "0.786", "0.761", "0.763", "0.761" ], [ "0.4495", "0.756", "0.6436", "0.757", "0.6436", "0.7715", "0.4495", "0.771" ], [ "0.2427", "0.784", "0.785", "0.784", "0.785", "0.799", "0.2427", "0.799" ], [ "0.2173", "0.797", "0.3506", "0.799", "0.3506", "0.8164", "0.2173", "0.8145" ], [ "0.361", "0.827", "0.642", "0.827", "0.642", "0.8413", "0.361", "0.8413" ], [ "0.761", "0.8247", "0.7876", "0.8247", "0.7876", "0.8423", "0.761", "0.8423" ] ]

[ "口", "广门", "", "Fig.3. If we consider the relations among the instances, the first two bags are more", "similar than thethird bag.Here,the solid lines highlight thehigh affinity among similar", "instances", "约是一个欧几里得空间，我们使用欧几里得距离来建立E-图。当涉及分类属性", "时，我们使用VDM（值差度量）[25]作为补充。详细地说，假设第一个属性", "是分类的，而其余的（d一i）属性是标准化为[0,1]的连续属性。我们可以使用", "(Zh=1VDM(a1,h,α2,h)+h=i+1a1,h-a2,h/2)1/2来测量1和2之间的距", "离。这里，两个值z1和22之间的VDM距离可以通过", "", "[Nz.z21.e_ Nz,22.c", "2", "(1)", "c=1Nz,z1", "VDM (z1, 22) =", "Nz,z2", "来计算，其中Nz.z表示在Z上持有值z的训练示例的数量，XMATHXMA2_1", "的X，表示", "将训练袋映射到一组图后，我们可以有很多选项来构建分类器。例如，我们", "可以构建一个采用图编辑距离[18]的k-最近邻分类器，或者我们可以设计一个", "图内核[13]来捕获图之间的相似性，然后通过SVM等内核机器解决分类问题。", "MIGraph方法采用第二种方式，我们的图核的思想如图。4", "简而言之，要测量图左侧所示的两个袋子之间的相似性。4，我们使用节点", "内核（i。e。，knode）来合并节点传递的信息，使用边缘内核（i。e。，kedge）来", "合并边所传达的信息，并将其聚合以获得最终的图核（i。e。，kG）。形式上，我", "们定义kG如下", "Definition 1. Given two multi-instance bags X; and X, which are presented as", "graphs Gh([chu]n=i, [ehjmi), h = i,j , where nh and mh are the number of", "nodes and edges in Gh , respectively.", "kc(X, X.) =Eμ D, hnode(aia jb)", "+DE, hedge(eia, ejb),", "(2)", "where knode and kedge are positive semidefinite kernels. To avoid numerical prob-", "lem, k is normalized to", "kG(X, X,)", "kG(X, X,) :", "(3)", "Vkc(X,,X)/kc(X,,X))", "knode和kedge可以通过多种方式定义。在这里，我们简单地将使用高斯RBF", "核的 knode定义为", "knode(cia,Cjb)=exp(-laia-jbl),", "(4)" ]

{ "arxiv_id": "0807.1997", "language": "zh", "page_idx": 3 }

[ [ "0.523", "0.2058", "0.625", "0.2058", "0.625", "0.2172", "0.523", "0.2172" ], [ "0.3308", "0.289", "0.6685", "0.289", "0.6685", "0.303", "0.3308", "0.303" ], [ "0.2279", "0.324", "0.7837", "0.324", "0.7837", "0.3372", "0.2279", "0.3372" ], [ "0.2181", "0.3372", "0.6953", "0.3372", "0.6953", "0.3516", "0.2181", "0.3516" ], [ "0.2427", "0.3542", "0.7812", "0.3542", "0.7812", "0.3674", "0.2427", "0.3674" ], [ "0.2181", "0.3687", "0.7837", "0.3674", "0.7837", "0.3818", "0.2181", "0.3833" ], [ "0.2214", "0.385", "0.7812", "0.385", "0.7812", "0.3958", "0.2214", "0.3958" ], [ "0.2181", "0.3977", "0.7812", "0.3984", "0.7812", "0.4136", "0.2181", "0.4124" ], [ "0.2336", "0.4136", "0.7837", "0.4136", "0.7837", "0.428", "0.2336", "0.428" ], [ "0.2214", "0.4287", "0.7837", "0.4287", "0.7837", "0.4426", "0.2214", "0.4426" ], [ "0.2197", "0.4443", "0.7837", "0.4443", "0.7837", "0.4583", "0.2197", "0.4583" ], [ "0.2197", "0.459", "0.7563", "0.459", "0.7563", "0.4734", "0.2197", "0.4734" ], [ "0.2443", "0.4749", "0.7837", "0.4749", "0.7837", "0.488", "0.2443", "0.488" ], [ "0.2197", "0.488", "0.7837", "0.4893", "0.7837", "0.504", "0.2197", "0.5024" ], [ "0.2214", "0.504", "0.7837", "0.504", "0.7837", "0.518", "0.2214", "0.518" ], [ "0.2197", "0.5195", "0.3882", "0.5195", "0.3882", "0.5327", "0.2197", "0.5327" ], [ "0.2418", "0.5327", "0.785", "0.534", "0.785", "0.549", "0.2418", "0.5474" ], [ "0.2214", "0.549", "0.7837", "0.55", "0.7837", "0.5645", "0.2214", "0.563" ], [ "0.2214", "0.5645", "0.785", "0.5645", "0.785", "0.5776", "0.2214", "0.5776" ], [ "0.2197", "0.58", "0.343", "0.58", "0.343", "0.5933", "0.2197", "0.5933" ], [ "0.2443", "0.5947", "0.7837", "0.5947", "0.7837", "0.608", "0.2443", "0.608" ], [ "0.2214", "0.61", "0.7837", "0.61", "0.7837", "0.6235", "0.2214", "0.6235" ], [ "0.2197", "0.6235", "0.7837", "0.625", "0.7837", "0.6396", "0.2197", "0.638" ], [ "0.2214", "0.64", "0.7837", "0.64", "0.7837", "0.654", "0.2214", "0.654" ], [ "0.2197", "0.656", "0.562", "0.656", "0.562", "0.6694", "0.2197", "0.6694" ], [ "0.2173", "0.677", "0.7876", "0.678", "0.7876", "0.6924", "0.2173", "0.6914" ], [ "0.2173", "0.6934", "0.4019", "0.6924", "0.4019", "0.707", "0.2173", "0.7085" ], [ "0.3308", "0.7173", "0.6562", "0.7104", "0.6567", "0.7373", "0.3325", "0.744" ], [ "0.763", "0.7266", "0.786", "0.7266", "0.786", "0.7407", "0.763", "0.7407" ], [ "0.4673", "0.735", "0.621", "0.7344", "0.621", "0.752", "0.4673", "0.7524" ], [ "0.2197", "0.7603", "0.7876", "0.7603", "0.7876", "0.778", "0.2197", "0.778" ], [ "0.2181", "0.776", "0.3423", "0.7764", "0.3423", "0.7915", "0.2181", "0.7905" ], [ "0.2181", "0.7944", "0.785", "0.7974", "0.785", "0.8164", "0.2181", "0.813" ], [ "0.2361", "0.813", "0.786", "0.813", "0.786", "0.8276", "0.2361", "0.8276" ], [ "0.2378", "0.829", "0.4697", "0.829", "0.4697", "0.839", "0.2378", "0.839" ] ]

[ "e kernel +edgekerne", "Fig. 4. Illustration of the graph kernel in MIGraph", "因此方程的第一部分E2正是使用高斯RBF核的MI核[14]。通过分别用", "eia和eib替换aia和aib，kedge也以类似于公式E4的形式定义", "这里的关键是如何定义描述边缘的特征向量。本文对于连接包X，的节点it", "和aiu的边，我们将其定义为[du,Pu,du,Pul'，其中du是节点aiu的度，即连接", "u与其他节点的边数。请注意，它已通过将其除以与X，相对应的图中的边总", "数进行了归一化。d，是节点aiu的度，其定义类似。pu定义为pu=wu/Zwu，", "其中分子是将aiu连接到aiu的边的权重；Wu,*是将&iu连接到X，中任何", "节点的边的权重之和，因此所有分母是。很明显，Pu传达了与节点au的连接", "对于节点&iu有多重要（或不重要）的信息。Pu的定义类似于节点aiu。这里", "的直觉是，如果边缘的结束节点的属性（e。g。，高度节点或低度节点）类似", "等式E2中定义的kG是正定核，可用于任何类型的图。kG(Xi，X）的计算", "复杂度为O(ninj+mimz）。kG显然满足图形内核定义应考虑的所有四个主要", "属性[5]。1我们上面的设计非常简单，但是在下一节中我们可以看到所提出的", "MIGraph方法相当有效", "MIGraph的一个缺陷是kG的计算复杂度是O(nini+mim），由边的数量", "主导。对于包含大量实例的包，将存在大量边缘，并且MIGraph将难以执行。", "因此，需要一种计算成本较小的方法。为此，我们提出了一种简单、高效但有效", "的miGraph方法", "对于包X，我们可以计算其实例之间的距离，并通过将距离与阈值进行", "比较来得出亲和矩阵Wi。例如，如果实例aia和aiu之间的距离小于，则", "Wi在a行和u列的元素wau设置为1，否则为0。对于X，有许多方法可以", "导出Wi。在本文中，我们使用高斯距离计算距离，并将设置为袋子中的平", "均距离。miGraph的密钥，内核kg，定义如下", "Definition 2. Given two multi-instance bags X and Xj which contains ni and", "ninstances,respectively.", "kg(X, X) - En-, WaWak(aia, ib)", "(5)", "Ea= Wia Er=, Wib", "where Wia = 1/n=1 wau , Wjb = 1/En=1 wbu , and k(aia,&jb) is defined as", "similar as Eq. 4.", "I We have tried to apply some existing graph kernels directly but unfortunately the", "results were not good. Due to the page limit, comparison with different graph kernels", "willbereportedinalongerversion," ]

{ "arxiv_id": "0807.1997", "language": "zh", "page_idx": 4 }

[ [ "0.1323", "0.132", "0.839", "0.132", "0.839", "0.1497", "0.1323", "0.1497" ], [ "0.3015", "0.1812", "0.6685", "0.1818", "0.6685", "0.1995", "0.3015", "0.1989" ], [ "0.2754", "0.2008", "0.697", "0.2008", "0.697", "0.2153", "0.2754", "0.2153" ], [ "0.1732", "0.2153", "0.7974", "0.2166", "0.7974", "0.2343", "0.1732", "0.2329" ], [ "0.3594", "0.2374", "0.614", "0.2374", "0.614", "0.252", "0.3594", "0.252" ], [ "0.2378", "0.2842", "0.3137", "0.286", "0.3137", "0.3037", "0.2378", "0.3018" ], [ "0.5034", "0.2886", "0.886", "0.2886", "0.886", "0.3018", "0.5034", "0.3018" ], [ "0.505", "0.3037", "0.8906", "0.3037", "0.8906", "0.3176", "0.505", "0.3176" ], [ "0.10046", "0.318", "0.468", "0.318", "0.468", "0.332", "0.10046", "0.332" ], [ "0.5034", "0.3193", "0.7563", "0.3193", "0.7563", "0.3333", "0.5034", "0.3333" ], [ "0.0809", "0.3352", "0.468", "0.3352", "0.468", "0.3484", "0.0809", "0.3484" ], [ "0.0809", "0.3499", "0.4673", "0.3499", "0.4673", "0.363", "0.0809", "0.363" ], [ "0.5186", "0.3484", "0.8906", "0.3484", "0.8906", "0.3623", "0.5186", "0.3623" ], [ "0.0784", "0.3643", "0.4673", "0.363", "0.4673", "0.3774", "0.0784", "0.379" ], [ "0.535", "0.3623", "0.892", "0.3623", "0.892", "0.377", "0.535", "0.377" ], [ "0.0801", "0.3801", "0.468", "0.3801", "0.468", "0.3945", "0.0801", "0.3945" ], [ "0.536", "0.379", "0.8906", "0.379", "0.8906", "0.3926", "0.536", "0.3926" ], [ "0.0801", "0.3945", "0.4673", "0.3945", "0.4673", "0.4077", "0.0801", "0.4077" ], [ "0.5337", "0.3933", "0.8906", "0.3926", "0.8906", "0.4072", "0.5337", "0.4077" ], [ "0.0801", "0.4104", "0.468", "0.4104", "0.468", "0.4236", "0.0801", "0.4236" ], [ "0.535", "0.4092", "0.892", "0.4092", "0.892", "0.4224", "0.535", "0.4224" ], [ "0.0825", "0.4268", "0.4673", "0.4268", "0.4673", "0.4375", "0.0825", "0.4375" ], [ "0.536", "0.4248", "0.8906", "0.4248", "0.8906", "0.4382", "0.536", "0.4382" ], [ "0.0809", "0.4414", "0.468", "0.4414", "0.468", "0.452", "0.0809", "0.452" ], [ "0.535", "0.4395", "0.892", "0.4395", "0.892", "0.4526", "0.535", "0.4526" ], [ "0.0801", "0.455", "0.468", "0.455", "0.468", "0.4697", "0.0801", "0.4697" ], [ "0.5376", "0.455", "0.8906", "0.455", "0.8906", "0.4653", "0.5376", "0.4653" ], [ "0.0784", "0.4697", "0.4697", "0.4685", "0.4697", "0.483", "0.0784", "0.4841" ], [ "0.536", "0.4685", "0.8906", "0.4685", "0.8906", "0.4817", "0.536", "0.4817" ], [ "0.0801", "0.4849", "0.4673", "0.4849", "0.4673", "0.5", "0.0801", "0.5" ], [ "0.535", "0.4841", "0.8906", "0.4841", "0.8906", "0.4976", "0.535", "0.4976" ], [ "0.0801", "0.5005", "0.4673", "0.5005", "0.4673", "0.5146", "0.0801", "0.5146" ], [ "0.5376", "0.5005", "0.8896", "0.5005", "0.8896", "0.511", "0.5376", "0.511" ], [ "0.0972", "0.5146", "0.468", "0.5166", "0.468", "0.53", "0.0972", "0.528" ], [ "0.536", "0.5146", "0.892", "0.5146", "0.892", "0.528", "0.536", "0.528" ], [ "0.0809", "0.531", "0.4673", "0.531", "0.4673", "0.5444", "0.0809", "0.5444" ], [ "0.535", "0.53", "0.8906", "0.53", "0.8906", "0.5435", "0.535", "0.5435" ], [ "0.0801", "0.547", "0.468", "0.547", "0.468", "0.56", "0.0801", "0.56" ], [ "0.536", "0.547", "0.648", "0.547", "0.648", "0.5566", "0.536", "0.5566" ], [ "0.0801", "0.56", "0.468", "0.56", "0.468", "0.5737", "0.0801", "0.5737" ], [ "0.5195", "0.571", "0.892", "0.571", "0.892", "0.584", "0.5195", "0.584" ], [ "0.0809", "0.577", "0.4697", "0.577", "0.4697", "0.5903", "0.0809", "0.5903" ], [ "0.0784", "0.589", "0.4697", "0.5913", "0.4697", "0.605", "0.0784", "0.603" ], [ "0.535", "0.5854", "0.892", "0.5854", "0.892", "0.5986", "0.535", "0.5986" ], [ "0.0784", "0.606", "0.468", "0.607", "0.468", "0.6206", "0.0784", "0.619" ], [ "0.535", "0.6016", "0.892", "0.6016", "0.892", "0.6157", "0.535", "0.6157" ], [ "0.0801", "0.622", "0.4697", "0.622", "0.4697", "0.6353", "0.0801", "0.6353" ], [ "0.536", "0.617", "0.892", "0.617", "0.892", "0.631", "0.536", "0.631" ], [ "0.0801", "0.636", "0.4673", "0.636", "0.4673", "0.6494", "0.0801", "0.6494" ], [ "0.5376", "0.6333", "0.8896", "0.6333", "0.8896", "0.643", "0.5376", "0.643" ], [ "0.0809", "0.6514", "0.4697", "0.6514", "0.4697", "0.6655", "0.0809", "0.6655" ], [ "0.535", "0.6465", "0.892", "0.6465", "0.892", "0.66", "0.535", "0.66" ], [ "0.0801", "0.6655", "0.468", "0.6665", "0.468", "0.681", "0.0801", "0.6797" ], [ "0.536", "0.661", "0.8906", "0.661", "0.8906", "0.675", "0.536", "0.675" ], [ "0.536", "0.6753", "0.892", "0.6753", "0.892", "0.6895", "0.536", "0.6895" ], [ "0.0809", "0.683", "0.4673", "0.683", "0.4673", "0.6934", "0.0809", "0.6934" ], [ "0.0801", "0.6963", "0.468", "0.6963", "0.468", "0.7104", "0.0801", "0.7104" ], [ "0.535", "0.6914", "0.8906", "0.6914", "0.8906", "0.705", "0.535", "0.705" ], [ "0.5337", "0.705", "0.892", "0.7056", "0.892", "0.72", "0.5337", "0.7197" ], [ "0.0784", "0.712", "0.468", "0.7104", "0.468", "0.7236", "0.0784", "0.726" ], [ "0.5317", "0.72", "0.6562", "0.7217", "0.6562", "0.7363", "0.5317", "0.735" ], [ "0.0809", "0.7266", "0.4697", "0.7266", "0.4697", "0.7407", "0.0809", "0.7407" ], [ "0.0801", "0.741", "0.367", "0.741", "0.367", "0.755", "0.0801", "0.755" ], [ "0.523", "0.75", "0.8906", "0.7505", "0.8906", "0.765", "0.523", "0.7646" ], [ "0.505", "0.7666", "0.8896", "0.7666", "0.8896", "0.78", "0.505", "0.78" ], [ "0.5034", "0.781", "0.8906", "0.781", "0.8906", "0.7944", "0.5034", "0.7944" ], [ "0.0809", "0.787", "0.2173", "0.787", "0.2173", "0.8003", "0.0809", "0.8003" ], [ "0.5034", "0.797", "0.8906", "0.797", "0.8906", "0.81", "0.5034", "0.81" ], [ "0.10297", "0.81", "0.468", "0.81", "0.468", "0.8237", "0.10297", "0.8237" ], [ "0.5034", "0.811", "0.8906", "0.811", "0.8906", "0.8247", "0.5034", "0.8247" ], [ "0.0809", "0.8257", "0.468", "0.8247", "0.468", "0.839", "0.0809", "0.8403" ], [ "0.505", "0.827", "0.8906", "0.827", "0.8906", "0.8403", "0.505", "0.8403" ], [ "0.08417", "0.8413", "0.468", "0.8413", "0.468", "0.855", "0.08417", "0.855" ], [ "0.5073", "0.8413", "0.8906", "0.8413", "0.8906", "0.855", "0.5073", "0.855" ], [ "0.0809", "0.857", "0.4673", "0.857", "0.4673", "0.8706", "0.0809", "0.8706" ], [ "0.5034", "0.857", "0.8906", "0.857", "0.8906", "0.8706", "0.5034", "0.8706" ], [ "0.10864", "0.872", "0.4673", "0.872", "0.4673", "0.8853", "0.10864", "0.8853" ], [ "0.5034", "0.872", "0.892", "0.872", "0.892", "0.8853", "0.5034", "0.8853" ], [ "0.0801", "0.8853", "0.468", "0.886", "0.468", "0.901", "0.0801", "0.8994" ], [ "0.5024", "0.8853", "0.892", "0.886", "0.892", "0.901", "0.5024", "0.8994" ] ]

[ "Crowd Counting on Images with Scale Variation and Isolated Clusters", "Haoyue Bai Song Wen S.-H. Gary Chan", "Department of Computer Science and Engineering", "The Hong Kong University of Science and Technology, Hong Kong, China", "[ hbaiaa, swenaa, gchan @cse.ust.hk", "Abstract", "象数量，而不是检测，识别和计数整个图像中的对象", "尽管最近取得了进展，但精确的人群计数仍然具有挑", "Crowd counting is to estimate the number of objects", "战性。这主要是由于以下两个因素：", "(e.g., people or vehicles) in an image of unconstrained", "congested scenes. Designing a general crowd count-", ". Large variation in object scale: Due to the per-", "ing algorithm applicable to a wide range of crowd im-", "spective and distance of the camera with respec-", "ages is challenging, mainly due to the possibly large", "tive to an object, two objects of the same phys-", "variation in object scales and the presence of many", "ical size would appear as different size (or scale)", "isolated small clusters. Previous approaches based on", "in an image, the so-called “scale variation.\" In a", "convolution operations withmulti-branch architecture", "crowded scene, objects may have large scale vari-", "areeffectivefor onlysomenarrowbands of scales,and", "ation. Some works have studied this scale vari-", "have not captured the long-range contextual relation-", "ation issue and achieved encouraging results us-", "ship due to isolated clustering. To address that, we", "ing multi-branch architecture with different filter", "propose SACANet , a novel s cale- a daptive long-", "sizes [36, 26] However, they are often based", "range c ontext- a ware net work for crowd counting.", "on fixed kernel size, which is only sensitivewithin", "SACANet consists of three major modules: the pyra-", "a few narrow bands of receptive fields in the full", "mid contextual module which extracts long-range con-", "spectrum, and hence cannot generally support di-", "textual information and enlarges the receptive field,", "verse scenarios.", "a scale-adaptive self-attention multi-branch module to", ". Isolated small clusters of objects: An image often", "attain high scale sensitivity and detection accuracy of", "isolated clusters, and a hierarchical fusion module to", "has isolated object clusters, i.e., some targets are", "fuse multi-level self-attention features.With group", "located far from the crowded mass. These isolated", "normalization, SACANet achieves better optimality in", "clusters may be many but small in size, e.g., one", "the training process. We have conducted extensive ex-", "ortwoobjectsinmanyscattered clusters.Convo-", "periments using the VisDrone2019 People dataset, the", "lutional operations with local receptive fields often", "VisDrone2019 Vehicle dataset, and some other chal-", "cannot capture properly long-range contextual in-", "formation, and hence have difficulty in modeling", "lengingbenchmarks.As comparedwiththe state-of-", "the-art methods, SACANet is shown to be effective,", "such isolated small clusters. This leads to unsat-", "isfactory results, especially for many such small", "especially for extremely crowded conditions with di-", "isolated clusters.", "verse scales and scattered clusters, and achieves much", "lower MAE as compared with baselines.", "无人机或监控摄像机捕获的人群场景表现出广泛", "的尺度变化和对象聚类。我们在表1中说明了Vis-", "Drone2019数据集[38]中带有车辆的三个典型图像。", "1. Introduction", "我们在第一行中显示原始图像；在第二行中显示其相", "人群计数是在不受约束的拥挤环境中估计物体（例", "应的密度图；在第三行中，根据其长度和高度的平均", "如人或车辆）的数量，其中图像通常由监视摄像机或无", "值，以像素为单位的对象的比例（对象大小）分布。为", "人机（UAV）拍摄。人群计数因其在公共安全、拥堵监", "了理解小的聚类效果，我们取N最近的邻居与对象的", "测、交通管理等方面的应用而受到广泛关注[30]，[11]", "距离，并计算它们的平均值，对于一些小的N（比如", "一种有前途的人群计数方法是使用基于密度图回归", "1-5）。较高的平均值意味着对象位于一个小的孤立簇", "的卷积神经网络（CNN），该方法估计每单位像素的对", "中，其大小不超过质量的Nk，反之亦然。因此，在" ]

{ "arxiv_id": "1909.03839", "language": "zh", "page_idx": 0 }

[ [ "0.0825", "0.0947", "0.4673", "0.0947", "0.4673", "0.1048", "0.0825", "0.1048" ], [ "0.5034", "0.09344", "0.8896", "0.09344", "0.8896", "0.1067", "0.5034", "0.1067" ], [ "0.08417", "0.10986", "0.4673", "0.10986", "0.4673", "0.1206", "0.08417", "0.1206" ], [ "0.5034", "0.108", "0.8906", "0.108", "0.8906", "0.1218", "0.5034", "0.1218" ], [ "0.0825", "0.1237", "0.4084", "0.1237", "0.4084", "0.137", "0.0825", "0.137" ], [ "0.505", "0.1237", "0.7812", "0.1237", "0.7812", "0.137", "0.505", "0.137" ], [ "0.10046", "0.1395", "0.468", "0.1395", "0.468", "0.1528", "0.10046", "0.1528" ], [ "0.5024", "0.1445", "0.7183", "0.1459", "0.7183", "0.1604", "0.5024", "0.159" ], [ "0.0809", "0.154", "0.468", "0.154", "0.468", "0.1674", "0.0809", "0.1674" ], [ "0.0809", "0.1692", "0.468", "0.1692", "0.468", "0.1831", "0.0809", "0.1831" ], [ "0.5254", "0.1704", "0.886", "0.1704", "0.886", "0.1812", "0.5254", "0.1812" ], [ "0.0801", "0.1843", "0.4656", "0.1849", "0.4656", "0.1989", "0.0801", "0.1976" ], [ "0.5034", "0.1831", "0.8896", "0.1843", "0.8896", "0.1976", "0.5034", "0.1963" ], [ "0.0784", "0.1976", "0.468", "0.1989", "0.468", "0.2122", "0.0784", "0.2108" ], [ "0.5034", "0.1989", "0.8906", "0.1989", "0.8906", "0.2122", "0.5034", "0.2122" ], [ "0.0801", "0.2134", "0.4697", "0.214", "0.4697", "0.2279", "0.0801", "0.2267" ], [ "0.505", "0.2153", "0.8896", "0.2153", "0.8896", "0.226", "0.505", "0.226" ], [ "0.0809", "0.2291", "0.1953", "0.2291", "0.1953", "0.2424", "0.0809", "0.2424" ], [ "0.505", "0.2291", "0.8906", "0.2291", "0.8906", "0.2424", "0.505", "0.2424" ], [ "0.10046", "0.2456", "0.4656", "0.2456", "0.4656", "0.2588", "0.10046", "0.2588" ], [ "0.505", "0.2443", "0.8906", "0.2443", "0.8906", "0.2583", "0.505", "0.2583" ], [ "0.0809", "0.26", "0.468", "0.26", "0.468", "0.2734", "0.0809", "0.2734" ], [ "0.5034", "0.2588", "0.8906", "0.2588", "0.8906", "0.2727", "0.5034", "0.2727" ], [ "0.0809", "0.276", "0.468", "0.276", "0.468", "0.289", "0.0809", "0.289" ], [ "0.505", "0.2747", "0.8906", "0.2747", "0.8906", "0.2886", "0.505", "0.2886" ], [ "0.08496", "0.2922", "0.4673", "0.2922", "0.4673", "0.303", "0.08496", "0.303" ], [ "0.5034", "0.289", "0.8906", "0.289", "0.8906", "0.303", "0.5034", "0.303" ], [ "0.0801", "0.3062", "0.468", "0.3062", "0.468", "0.3193", "0.0801", "0.3193" ], [ "0.5063", "0.3062", "0.888", "0.3062", "0.888", "0.3164", "0.5063", "0.3164" ], [ "0.0801", "0.318", "0.1478", "0.3208", "0.1478", "0.3352", "0.0801", "0.3333" ], [ "0.505", "0.3193", "0.8896", "0.3193", "0.8896", "0.3333", "0.505", "0.3333" ], [ "0.505", "0.3352", "0.8906", "0.3352", "0.8906", "0.3484", "0.505", "0.3484" ], [ "0.0801", "0.3516", "0.2263", "0.3516", "0.2263", "0.3655", "0.0801", "0.3655" ], [ "0.505", "0.351", "0.6855", "0.351", "0.6855", "0.3643", "0.505", "0.3643" ], [ "0.1013", "0.3774", "0.4673", "0.3774", "0.4673", "0.3914", "0.1013", "0.3914" ], [ "0.5024", "0.377", "0.8896", "0.377", "0.8896", "0.3901", "0.5024", "0.3901" ], [ "0.0825", "0.3945", "0.4673", "0.3945", "0.4673", "0.4048", "0.0825", "0.4048" ], [ "0.5312", "0.3933", "0.7114", "0.3945", "0.7114", "0.4092", "0.5312", "0.4077" ], [ "0.0825", "0.4077", "0.468", "0.4077", "0.468", "0.4216", "0.0825", "0.4216" ], [ "0.0809", "0.4236", "0.2712", "0.4236", "0.2712", "0.4375", "0.0809", "0.4375" ], [ "0.523", "0.4185", "0.8906", "0.4185", "0.8906", "0.432", "0.523", "0.432" ], [ "0.505", "0.4343", "0.892", "0.4343", "0.892", "0.4475", "0.505", "0.4475" ], [ "0.0784", "0.4495", "0.3015", "0.449", "0.3015", "0.4634", "0.0784", "0.464" ], [ "0.505", "0.449", "0.8906", "0.449", "0.8906", "0.4622", "0.505", "0.4622" ], [ "0.505", "0.464", "0.892", "0.464", "0.892", "0.4778", "0.505", "0.4778" ], [ "0.1013", "0.476", "0.4673", "0.476", "0.4673", "0.486", "0.1013", "0.486" ], [ "0.5034", "0.4792", "0.8906", "0.4792", "0.8906", "0.4924", "0.5034", "0.4924" ], [ "0.0801", "0.4873", "0.468", "0.488", "0.468", "0.5024", "0.0801", "0.502" ], [ "0.5034", "0.4944", "0.8906", "0.4944", "0.8906", "0.5083", "0.5034", "0.5083" ], [ "0.0809", "0.504", "0.4673", "0.504", "0.4673", "0.5176", "0.0809", "0.5176" ], [ "0.5034", "0.511", "0.8906", "0.511", "0.8906", "0.5215", "0.5034", "0.5215" ], [ "0.0866", "0.5234", "0.0989", "0.5234", "0.0989", "0.53", "0.0866", "0.53" ], [ "0.0972", "0.521", "0.4656", "0.521", "0.4656", "0.531", "0.0972", "0.531" ], [ "0.505", "0.526", "0.569", "0.526", "0.569", "0.537", "0.505", "0.537" ], [ "0.0825", "0.534", "0.4673", "0.534", "0.4673", "0.5474", "0.0825", "0.5474" ], [ "0.523", "0.5405", "0.8906", "0.5405", "0.8906", "0.5537", "0.523", "0.5537" ], [ "0.0784", "0.5474", "0.4656", "0.549", "0.4656", "0.5625", "0.0784", "0.561" ], [ "0.5034", "0.5547", "0.8906", "0.5547", "0.8906", "0.5684", "0.5034", "0.5684" ], [ "0.08417", "0.5645", "0.4673", "0.5645", "0.4673", "0.5776", "0.08417", "0.5776" ], [ "0.505", "0.5693", "0.892", "0.5693", "0.892", "0.5835", "0.505", "0.5835" ], [ "0.0825", "0.579", "0.4673", "0.579", "0.4673", "0.5923", "0.0825", "0.5923" ], [ "0.5034", "0.5854", "0.5312", "0.5854", "0.5312", "0.5977", "0.5034", "0.5977" ], [ "0.0825", "0.5947", "0.4673", "0.5947", "0.4673", "0.608", "0.0825", "0.608" ], [ "0.0825", "0.6113", "0.4656", "0.6113", "0.4656", "0.622", "0.0825", "0.622" ], [ "0.5034", "0.607", "0.761", "0.607", "0.761", "0.6206", "0.5034", "0.6206" ], [ "0.0825", "0.625", "0.4673", "0.625", "0.4673", "0.638", "0.0825", "0.638" ], [ "0.524", "0.63", "0.892", "0.63", "0.892", "0.643", "0.524", "0.643" ], [ "0.0809", "0.64", "0.468", "0.64", "0.468", "0.654", "0.0809", "0.654" ], [ "0.5034", "0.646", "0.8896", "0.646", "0.8896", "0.659", "0.5034", "0.659" ], [ "0.0809", "0.655", "0.468", "0.655", "0.468", "0.6685", "0.0809", "0.6685" ], [ "0.5034", "0.659", "0.8896", "0.66", "0.8896", "0.674", "0.5034", "0.6724" ], [ "0.0809", "0.6704", "0.2051", "0.6704", "0.2051", "0.6846", "0.0809", "0.6846" ], [ "0.5024", "0.675", "0.8896", "0.675", "0.8896", "0.688", "0.5024", "0.688" ], [ "0.0809", "0.6963", "0.3774", "0.6963", "0.3774", "0.7104", "0.0809", "0.7104" ], [ "0.505", "0.6914", "0.8896", "0.6914", "0.8896", "0.705", "0.505", "0.705" ], [ "0.5034", "0.7056", "0.8906", "0.7056", "0.8906", "0.7197", "0.5034", "0.7197" ], [ "0.1013", "0.72", "0.468", "0.72", "0.468", "0.7344", "0.1013", "0.7344" ], [ "0.505", "0.7227", "0.5864", "0.7227", "0.5864", "0.7344", "0.505", "0.7344" ], [ "0.08417", "0.7373", "0.4673", "0.7373", "0.4673", "0.7476", "0.08417", "0.7476" ], [ "0.524", "0.7363", "0.8906", "0.7363", "0.8906", "0.75", "0.524", "0.75" ], [ "0.0825", "0.7524", "0.4656", "0.7524", "0.4656", "0.763", "0.0825", "0.763" ], [ "0.5034", "0.752", "0.8906", "0.752", "0.8906", "0.765", "0.5034", "0.765" ], [ "0.0825", "0.7666", "0.468", "0.7666", "0.468", "0.78", "0.0825", "0.78" ], [ "0.505", "0.768", "0.888", "0.768", "0.888", "0.779", "0.505", "0.779" ], [ "0.0825", "0.7827", "0.183", "0.7827", "0.183", "0.7935", "0.0825", "0.7935" ], [ "0.5063", "0.7827", "0.888", "0.7827", "0.888", "0.7935", "0.5063", "0.7935" ], [ "0.10046", "0.797", "0.468", "0.797", "0.468", "0.81", "0.10046", "0.81" ], [ "0.5034", "0.797", "0.8906", "0.797", "0.8906", "0.81", "0.5034", "0.81" ], [ "0.0801", "0.811", "0.4673", "0.8125", "0.4673", "0.8257", "0.0801", "0.8247" ], [ "0.505", "0.813", "0.888", "0.813", "0.888", "0.8237", "0.505", "0.8237" ], [ "0.0809", "0.827", "0.4673", "0.827", "0.4673", "0.8403", "0.0809", "0.8403" ], [ "0.5034", "0.827", "0.693", "0.827", "0.693", "0.8403", "0.5034", "0.8403" ], [ "0.0825", "0.8433", "0.4656", "0.8433", "0.4656", "0.854", "0.0825", "0.854" ], [ "0.0809", "0.857", "0.4673", "0.857", "0.4673", "0.8706", "0.0809", "0.8706" ], [ "0.5034", "0.851", "0.831", "0.851", "0.831", "0.8613", "0.5034", "0.8613" ], [ "0.08417", "0.8735", "0.4673", "0.8735", "0.4673", "0.884", "0.08417", "0.884" ], [ "0.523", "0.8726", "0.888", "0.8726", "0.888", "0.886", "0.523", "0.886" ], [ "0.08417", "0.8896", "0.4673", "0.8896", "0.4673", "0.8994", "0.08417", "0.8994" ], [ "0.5024", "0.887", "0.8896", "0.887", "0.8896", "0.901", "0.5024", "0.901" ] ]

[ "将多级尺度和上下文特征结合起来，生成高精度的密", "态感受野大小，无法扩展以解决人群场景中的大范围", "度图。SACANet还使用组归一化，通过减少小批量大", "变化。此外，当前的方法不能充分利用远程上下文信", "小的影响来实现更高的最优性和更好的收敛性", "息，并且难以对孤立的小集群进行建模", "我们已经对基于无人机和基于监控摄像", "2.3. Drone-based Scenarios", "机的人群数据集进行了广泛的实验，用于不", "受限制的人群计数。我们使用来自四个数", "无人机传感器和监控摄像头都可以捕获人群场景", "据集的图像：VisDrone2019People&Vehicle，Shang", "但是基于无人机的场景的人群分析对于智能城市应用", "haiTechA&ShanghaiTechB。与最先进的方法相比，", "而言更加灵活E[19]。但是，以前的工作更多地集中在", "SACANet在所有四个具有挑战性的基准测试中都实", "基于监控摄像头的人群计数上。据我们所知，基于无人", "现了卓越的性能", "机的人群计数尚未得到充分探索，而且还缺乏公开的", "本文的其余部分组织如下。在2节中，我们回顾了", "大规模多样化数据集。此外，与监控摄像机拍摄的图像", "相关的工作。然后，我们在E3节中介绍了SACANet", "相比，基于无人机的人群图像的孤立小簇问题更加严", "的详细信息，包括其金字塔上下文模块，比例自适应", "重[24]。在本文中，我们将VisDrone2019具有挑战", "自我注意模块，分层融合，组归一化和自标函数。我们", "性的数据集E[37]修改为两个大规模的多样化的基于", "在4节中讨论了实验设置并说明了结果，并在5节中", "无人机的人群计数数据集，这些新拆分的数据集可以", "得出结论", "促进领域。此外，我们解决了两个主要的具有挑战性的", "规模变化和孤立的小集群问题，用于在端到端框架内", "2. Related Work", "进行不受约束的人群计数", "在本节中，我们将从三个主要方向讨论人群计数方", "3. SACANet for Crowds with Scale Variation", "法的相关工作：传统的人群计数算法（第E2.1节），基", "and Isolated Clusters", "于深度学习的方法（第E2.2节）和基于无人机的场景", "的人群计数（第E2.3节）。", "在本节中，我们将介绍SACANet的详细信息，这", "是一种新颖的scal-adaptive远程context-a软", "2.1. Traditional Approaches", "件net用于人群计数的工作。SACANet由三个组件组", "成：金字塔上下文模块（第E3.1节），比例自适应自我", "早期的人群计数方法通常基于具有手工制作功能的", "关注模块（第E3.2节）和分层融合模块（第E3.3节）。", "检测模型，即。e。，他们利用行人或身体部位检测器来", "此外，我们在E3.4节中描述了如何利用组归一化来促", "检测单个物体，并计算整个图像中的数量[25]、[17]", "进收敛并实现更好的最优性。自标函数的细节在3.5", "", "然而，这些基于检测的方法的性能在高度拥挤的场", "节中讨论", "景中严重下降。一些研究人员试图使用具有HOG和", "在图E1中，我们显示了SACANet的体系结构。上", "SIFT等低级特征的基于回归的方法来计算全局数E[4", "半部分介绍了SACANet的三个组件，我们在底部（黄", "。即使依赖于低级特征，这些方法也能为全局计数估计", "色箭头下方）更详细地描述了比例自适应自我注意模", "获得更好的结果。为了整合空间信息，研究人员提出了", "块。", "基于密度图回归的方法，即测量人群场景中某个区域", "每单位像素的人数。正如在E14中讨论的那样，这项", "3.1. Pyramid Contextual Module", "工作是第一个提供具有线性映射算法的基于密度图回", "金字塔上下文模块旨在提取远程上下文信息并扩大", "归的人群计数方法的工作。然后，另一项工作通过随机", "感受野。为了与以前的作品E[15]E[5]E[20] 进行公", "森林回归对其进行了改进，以学习非线性映射并获得", "平的比较，我们的方法将具有出色可转移性的截短的", "更好的性能E[23]", "VGG-16E[28]作为骨于。我们提取VGG-16的前十层", "2.2. Deep Learning-based Approaches", "只有三个池化层，以平衡大的有效感受野和密度图分", "辨率。此外，仅对较高层采用多分支可以在训练期间提", "最近，研究人员采用了基于深度学习的方法，而不", "高内存效率", "是依靠手工制作的特征来生成高质量的密度图：并实", "金字塔上下文模块由特征提取器和上下文融合组成。", "现了精确的人群计数E3：E27L。这些方法可以应", "我们将输入图像下采样为原始分辨率的1/4，然后将", "用于计算不同类型的对象（i。e。、车辆和细胞）而不是", "这两种输入以共享权重馈送到VGG-16前端。在将生", "人159", "成的特征图的通道输入到第二个比例自适应自我注意", "研究人员提出了多列卷积神经网络，每列具有不同", "模块之前，我们对其进行混洗。此组件可以扩大感受野", "的内核大小，以解决规模变化问题[36]。Switching-", "并捕获远程上下文信息。编码器部分还设计为便于下", "CNN将基于补丁的交换块附加到多列结构上，并更好", "游尺度自适应自注意处理。", "地处理每列的特定缩放范围26。HydraCNN利用具", "有多个尺度的图像补丁金字塔进行人群估计[22]。但", "3.2.Scale-AdaptiveSelf-AttentionModule", "是：具有不适当的感受野大小的计数网络会给多尺度", "缩放自适应自我关注模块是为了适应广泛的尺度", "自标带来不平衡的焦点。。此外，这些方法仅依赖于静", "并通过远程上下文信息检测孤立的集群。这部分由三" ]

{ "arxiv_id": "1909.03839", "language": "zh", "page_idx": 2 }

[ [ "0.1772", "0.09344", "0.2524", "0.101", "0.25", "0.1187", "0.1748", "0.1111" ], [ "0.2427", "0.0966", "0.3718", "0.0966", "0.3718", "0.10986", "0.2427", "0.10986" ], [ "0.4453", "0.09784", "0.4697", "0.09784", "0.4697", "0.10986", "0.4453", "0.10986" ], [ "0.5034", "0.09216", "0.8896", "0.09216", "0.8896", "0.1061", "0.5034", "0.1061" ], [ "0.5063", "0.108", "0.8896", "0.108", "0.8896", "0.1218", "0.5063", "0.1218" ], [ "0.0809", "0.11615", "0.4673", "0.11743", "0.4673", "0.132", "0.0809", "0.1307" ], [ "0.505", "0.125", "0.888", "0.125", "0.888", "0.1351", "0.505", "0.1351" ], [ "0.0825", "0.1332", "0.468", "0.1332", "0.468", "0.1465", "0.0825", "0.1465" ], [ "0.5034", "0.1383", "0.892", "0.1383", "0.892", "0.1528", "0.5034", "0.1528" ], [ "0.0825", "0.1483", "0.4558", "0.1483", "0.4558", "0.1622", "0.0825", "0.1622" ], [ "0.5063", "0.1547", "0.751", "0.1547", "0.751", "0.1654", "0.5063", "0.1654" ], [ "0.5254", "0.1704", "0.8896", "0.1704", "0.8896", "0.1812", "0.5254", "0.1812" ], [ "0.0801", "0.178", "0.4182", "0.1768", "0.4182", "0.1913", "0.0801", "0.1925" ], [ "0.505", "0.1843", "0.8906", "0.1843", "0.8906", "0.1976", "0.505", "0.1976" ], [ "0.1013", "0.2026", "0.468", "0.2026", "0.468", "0.2166", "0.1013", "0.2166" ], [ "0.5034", "0.1989", "0.8906", "0.1989", "0.8906", "0.2122", "0.5034", "0.2122" ], [ "0.0809", "0.2172", "0.468", "0.2172", "0.468", "0.2311", "0.0809", "0.2311" ], [ "0.5034", "0.2122", "0.8896", "0.2134", "0.8896", "0.2279", "0.5034", "0.2267" ], [ "0.0825", "0.2317", "0.4697", "0.2317", "0.4697", "0.2456", "0.0825", "0.2456" ], [ "0.5034", "0.2291", "0.8896", "0.2291", "0.8896", "0.2424", "0.5034", "0.2424" ], [ "0.0809", "0.2474", "0.4673", "0.2474", "0.4673", "0.2615", "0.0809", "0.2615" ], [ "0.505", "0.2456", "0.888", "0.2456", "0.888", "0.2556", "0.505", "0.2556" ], [ "0.0809", "0.262", "0.468", "0.262", "0.468", "0.276", "0.0809", "0.276" ], [ "0.5063", "0.2615", "0.7695", "0.2615", "0.7695", "0.2715", "0.5063", "0.2715" ], [ "0.0825", "0.2778", "0.3896", "0.2778", "0.3896", "0.2917", "0.0825", "0.2917" ], [ "0.7134", "0.2734", "0.826", "0.2874", "0.8228", "0.3018", "0.71", "0.2886" ], [ "0.5073", "0.2854", "0.7173", "0.2854", "0.7173", "0.2986", "0.5073", "0.2986" ], [ "0.739", "0.2966", "0.7485", "0.2966", "0.7485", "0.3037", "0.739", "0.3037" ], [ "0.0801", "0.303", "0.468", "0.303", "0.468", "0.3164", "0.0801", "0.3164" ], [ "0.714", "0.3037", "0.8364", "0.3176", "0.8325", "0.3352", "0.71", "0.3213" ], [ "0.834", "0.307", "0.898", "0.3113", "0.8965", "0.3245", "0.8325", "0.3208" ], [ "0.505", "0.315", "0.7183", "0.315", "0.7183", "0.3289", "0.505", "0.3289" ], [ "0.10046", "0.3257", "0.468", "0.3257", "0.468", "0.339", "0.10046", "0.339" ], [ "0.738", "0.3257", "0.749", "0.3257", "0.749", "0.3372", "0.738", "0.3372" ], [ "0.863", "0.327", "0.871", "0.327", "0.871", "0.336", "0.863", "0.336" ], [ "0.863", "0.3308", "0.8906", "0.3352", "0.886", "0.3484", "0.859", "0.3435" ], [ "0.0825", "0.3423", "0.4673", "0.3423", "0.4673", "0.3528", "0.0825", "0.3528" ], [ "0.5254", "0.3447", "0.8906", "0.3452", "0.8906", "0.3599", "0.5254", "0.3591" ], [ "0.0809", "0.356", "0.4673", "0.356", "0.4673", "0.3694", "0.0809", "0.3694" ], [ "0.505", "0.361", "0.892", "0.361", "0.892", "0.3757", "0.505", "0.3757" ], [ "0.0809", "0.3706", "0.468", "0.3706", "0.468", "0.385", "0.0809", "0.385" ], [ "0.505", "0.377", "0.8906", "0.377", "0.8906", "0.3901", "0.505", "0.3901" ], [ "0.0784", "0.385", "0.281", "0.3838", "0.281", "0.3984", "0.0784", "0.3997" ], [ "0.5024", "0.3914", "0.634", "0.3914", "0.634", "0.4048", "0.5024", "0.4048" ], [ "0.1447", "0.3984", "0.3367", "0.3984", "0.3367", "0.4312", "0.1447", "0.4312" ], [ "0.255", "0.404", "0.298", "0.404", "0.298", "0.414", "0.255", "0.414" ], [ "0.3245", "0.4048", "0.4224", "0.4048", "0.4224", "0.4167", "0.3245", "0.4167" ], [ "0.524", "0.4077", "0.8906", "0.4077", "0.8906", "0.4216", "0.524", "0.4216" ], [ "0.0989", "0.4192", "0.1487", "0.4192", "0.1487", "0.4312", "0.0989", "0.4312" ], [ "0.4453", "0.4204", "0.4673", "0.4204", "0.4673", "0.433", "0.4453", "0.433" ], [ "0.5034", "0.4204", "0.8906", "0.4216", "0.8906", "0.4363", "0.5034", "0.435" ], [ "0.1969", "0.433", "0.2296", "0.433", "0.2296", "0.44", "0.1969", "0.44" ], [ "0.1691", "0.4255", "0.4043", "0.4236", "0.4043", "0.452", "0.1691", "0.4539" ], [ "0.505", "0.4375", "0.8906", "0.4375", "0.8906", "0.4507", "0.505", "0.4507" ], [ "0.0989", "0.4526", "0.468", "0.4539", "0.468", "0.4685", "0.0989", "0.4673" ], [ "0.5254", "0.455", "0.8896", "0.455", "0.8896", "0.4653", "0.5254", "0.4653" ], [ "0.0809", "0.4697", "0.4673", "0.4697", "0.4673", "0.483", "0.0809", "0.483" ], [ "0.5034", "0.4685", "0.8906", "0.4685", "0.8906", "0.4817", "0.5034", "0.4817" ], [ "0.0825", "0.486", "0.4673", "0.486", "0.4673", "0.4968", "0.0825", "0.4968" ], [ "0.5034", "0.4841", "0.8906", "0.4841", "0.8906", "0.4976", "0.5034", "0.4976" ], [ "0.0809", "0.5", "0.468", "0.5", "0.468", "0.513", "0.0809", "0.513" ], [ "0.5024", "0.4968", "0.888", "0.4988", "0.888", "0.513", "0.5024", "0.511" ], [ "0.0809", "0.515", "0.3096", "0.515", "0.3096", "0.529", "0.0809", "0.529" ], [ "0.5024", "0.512", "0.888", "0.513", "0.888", "0.528", "0.5024", "0.5264" ], [ "0.10046", "0.53", "0.468", "0.53", "0.468", "0.5435", "0.10046", "0.5435" ], [ "0.5034", "0.529", "0.8896", "0.529", "0.8896", "0.5425", "0.5034", "0.5425" ], [ "0.0825", "0.5454", "0.468", "0.5454", "0.468", "0.5596", "0.0825", "0.5596" ], [ "0.5034", "0.5444", "0.7363", "0.5444", "0.7363", "0.558", "0.5034", "0.558" ], [ "0.0825", "0.561", "0.1381", "0.561", "0.1381", "0.5728", "0.0825", "0.5728" ], [ "0.5024", "0.5693", "0.6665", "0.571", "0.6665", "0.5854", "0.5024", "0.584" ], [ "0.125", "0.5757", "0.327", "0.563", "0.3325", "0.6157", "0.1299", "0.6274" ], [ "0.1691", "0.5806", "0.1887", "0.5806", "0.1887", "0.5933", "0.1691", "0.5933" ], [ "0.326", "0.577", "0.4739", "0.577", "0.4739", "0.6255", "0.326", "0.6255" ], [ "0.357", "0.5806", "0.38", "0.5806", "0.38", "0.5933", "0.357", "0.5933" ], [ "0.0866", "0.595", "0.1422", "0.595", "0.1422", "0.606", "0.0866", "0.606" ], [ "0.524", "0.5947", "0.8906", "0.5947", "0.8906", "0.608", "0.524", "0.608" ], [ "0.1691", "0.61", "0.1929", "0.61", "0.1929", "0.622", "0.1691", "0.622" ], [ "0.3545", "0.61", "0.384", "0.61", "0.384", "0.622", "0.3545", "0.622" ], [ "0.505", "0.6094", "0.8906", "0.6094", "0.8906", "0.6226", "0.505", "0.6226" ], [ "0.4468", "0.625", "0.4673", "0.625", "0.4673", "0.637", "0.4468", "0.637" ], [ "0.505", "0.625", "0.8896", "0.625", "0.8896", "0.638", "0.505", "0.638" ], [ "0.5034", "0.6396", "0.8906", "0.6396", "0.8906", "0.653", "0.5034", "0.653" ], [ "0.0809", "0.6274", "0.29", "0.629", "0.29", "0.64", "0.0809", "0.6396" ], [ "0.10046", "0.6426", "0.468", "0.6426", "0.468", "0.656", "0.10046", "0.656" ], [ "0.0809", "0.6577", "0.4673", "0.6577", "0.4673", "0.672", "0.0809", "0.672" ], [ "0.5034", "0.654", "0.8906", "0.654", "0.8906", "0.6685", "0.5034", "0.6685" ], [ "0.0809", "0.672", "0.4673", "0.672", "0.4673", "0.685", "0.0809", "0.685" ], [ "0.5034", "0.6694", "0.8896", "0.6694", "0.8896", "0.683", "0.5034", "0.683" ], [ "0.0825", "0.6875", "0.4673", "0.6875", "0.4673", "0.7007", "0.0825", "0.7007" ], [ "0.5034", "0.685", "0.8896", "0.685", "0.8896", "0.6987", "0.5034", "0.6987" ], [ "0.0825", "0.7026", "0.4673", "0.7026", "0.4673", "0.7163", "0.0825", "0.7163" ], [ "0.5034", "0.7007", "0.8037", "0.7007", "0.8037", "0.714", "0.5034", "0.714" ], [ "0.0825", "0.7173", "0.4673", "0.7173", "0.4673", "0.731", "0.0825", "0.731" ], [ "0.524", "0.7163", "0.8906", "0.7163", "0.8906", "0.73", "0.524", "0.73" ], [ "0.0825", "0.7344", "0.468", "0.7344", "0.468", "0.7476", "0.0825", "0.7476" ], [ "0.5034", "0.731", "0.8906", "0.731", "0.8906", "0.744", "0.5034", "0.744" ], [ "0.0825", "0.7485", "0.468", "0.7485", "0.468", "0.762", "0.0825", "0.762" ], [ "0.5034", "0.747", "0.8906", "0.747", "0.8906", "0.7603", "0.5034", "0.7603" ], [ "0.0809", "0.7646", "0.3586", "0.7646", "0.3586", "0.778", "0.0809", "0.778" ], [ "0.505", "0.761", "0.8906", "0.761", "0.8906", "0.7744", "0.505", "0.7744" ], [ "0.5034", "0.7764", "0.831", "0.7764", "0.831", "0.7905", "0.5034", "0.7905" ], [ "0.0825", "0.7905", "0.4224", "0.7905", "0.4224", "0.8003", "0.0825", "0.8003" ], [ "0.5024", "0.803", "0.8716", "0.803", "0.8716", "0.8164", "0.5024", "0.8164" ], [ "0.10046", "0.811", "0.4673", "0.811", "0.4673", "0.8247", "0.10046", "0.8247" ], [ "0.0825", "0.827", "0.4656", "0.827", "0.4656", "0.8403", "0.0825", "0.8403" ], [ "0.523", "0.827", "0.8906", "0.827", "0.8906", "0.8403", "0.523", "0.8403" ], [ "0.0784", "0.8403", "0.4429", "0.8413", "0.4429", "0.855", "0.0784", "0.854" ], [ "0.505", "0.8413", "0.8896", "0.8413", "0.8896", "0.855", "0.505", "0.855" ], [ "0.10046", "0.857", "0.468", "0.857", "0.468", "0.8706", "0.10046", "0.8706" ], [ "0.5024", "0.857", "0.8906", "0.857", "0.8906", "0.8706", "0.5024", "0.8706" ], [ "0.0801", "0.8706", "0.468", "0.872", "0.468", "0.886", "0.0801", "0.8853" ], [ "0.509", "0.8774", "0.5215", "0.8774", "0.5215", "0.882", "0.509", "0.882" ], [ "0.5156", "0.872", "0.8906", "0.8706", "0.8906", "0.8853", "0.5156", "0.886" ], [ "0.0809", "0.887", "0.4697", "0.887", "0.4697", "0.901", "0.0809", "0.901" ], [ "0.5034", "0.887", "0.8906", "0.887", "0.8906", "0.901", "0.5034", "0.901" ] ]

[ "LE=N", "F(X;Q) - YI2.", "(3)", "类别0（行人）和类别1（人）合并为一个数据集进行", "人群计数，新的注释位置是原始边界框的头部点。我们", "其中α表示模型参数，N表示像素数，X表示输人图", "重新排列数据拆分，并过滤掉注释目标对象数量小于", "像，Y是其地面实况，F(X；α）是生成的密度图。我们", "10的情况。最后，该数据集由2392个训练样本、329", "可以通过总结估计的人群密度图来预测人群计数结果", "个验证样本和626个测试样本组成", "VisDrone2019车辆「38]。在对VisDrone2019People", "4.Experiments and IllustrativeResults", "数据集进行类似的修改和重新排列步骤之后，我们将", "在本节中，我们在E4.1节中描述了评估指标和比", "类别4（汽车），类别5（货车），类别6（卡车）和类", "较方案。在4.2节中介绍了四个数据集和地面实况生", "别9（公共汽车）合并为一个数据集，用于车辆人群计", "成方法的描述。第4.3节显示了我们的培训详细信息。", "数。我们得到了3953个训练样本，364个验证样本和", "4.4节详细介绍了人员和车辆数据集的定性和定量分", "986个测试样本。新的车辆注释位置是原始边界框的中", "析。此外，我们在4.5节中进行消融研究，并在4.6", "心点。这两种新的注释操作定义如下", "节中比较SACANet在无约束情况下的应用", "bbwidth.bbtopl,", "People[cols, rows] =[bbieft +", "2", "4.1. Evaluation Metrics and Comparison Schemes", "bbwidth,bbtop+", "bbheight", "Vehicle[cols, rows] =[bbieft +", "我们使用变异系数（CV）来测量度标度变化E[1L。", "2", "2", "L(6)", "对于孤立小集群的级别，我们在对其应用K-means算", "上海科技A[36]。上海科技A数据集包括482张", "法后，利用Dunn有效性指数（DVI）来计算距离分", "人群图像，总数为241，677人。每个图像的计数从33", "布E[21]。两个评估指标定义如下：CV=α/μ，其中", "到3139不等。此数据集从Internet随机抓取，并且具", "α是标准偏差，比是平均值，", "有不固定的分辨率", "r rea, n.", "min", "i-i", "上海科技B[36]。ShanghaiTechB数据集有716张", "DVI=", "(4)", "固定分辨率的图像，这些图像是由固定摄像机从繁忙", "max", "<maxkvamen. (l, - a,l", "的购物街上拍摄的。每个图像的计数从12到578不等", "DVI计算最短簇间距离除以最大簇内距离。在我们", "我们通过使用高斯核（归一化为1）模糊车辆注", "的实验中，我们应用K-means算法[7]（我们实验中", "释的每个头部或中心并考虑每个数据集中所有图像", "的K=2）对平均N=2最近距离数据进行预处理", "的空间分布来生成地面实况。对于稀疏人群数据集", "然后计算DVI值。DVI值越大，簇间距离越大，簇内", "VisDrone2019People，VisDrone2019Vehicle和 Shang", "距离越短，孤立的小尺寸簇越多", "haiTechB，我们使用固定内核α=15作为生成方法", "对于总体人群计数结果，使用两个指标进行评估E", "对于ShanghaiTechA拥挤的数据集，我们采用几何自", "[33]，平均绝对误差（MAE）和均方误差（MSE），定", "适应内核方法生成地面实况[36]", "义如下：", "4.3. Training Details", "--~1α- l MSE-", "N", "NAZC-CP", "N", "MAE=", "在我们的实验中，我们使用具有来自ImageNet分类", "1", "-1", "挑战数据集E[13]的预训练权重的VGG-16主干。对", "(5)", "于其他层，我们使用均值为零、标准差为1的高斯分布", "的随机权重进行初始化。我们使用Adam优化器[12]", "其中N是测试图像的总数C", "我们将我们的方法与VisDrone2019数据集上的三", "种方案进行了比较：VGG-16E[28]，MCNNE[36]和", "初始学习率为le-4。对于ShanghaiTechB数据集，我", "CSRNetE[15]。VGG-16是一个强大的骨于，我们直接", "们随机翻转一些样本以增加原始训练数据。对于具有", "将其修改为人群计数网络。多列卷积神经网络（MCNN）", "更高分辨率的数据集，我们将输入图像的大小调整为", "是一种具有多分支架构的著名人群计数方法。我们用", "不超过768×1024，但保持相同的纵横比", "三个分支在我们的框架中实现它，这与我们的相同，以", "在测试步骤中，我们直接测试具有固定分辨率的", "便进行公平的比较。CSRNet是用于拥塞场景理解的最", "ShanghaiTechB数据集的完整图像。对于分辨率不固", "先进的方法之一，它利用扩展的卷积来扩大感受野。我", "定的其他数据集，我们在将图像输入网络之前调整其", "们使用与我们相同的实验设置来实现它", "大小。通过对整个图像求和来给出每个图像的估计总", "计数。我们实现了基于PyTorch的 SACANet", "4.2.Datasetsand GroundTruth Generation", "4.4. Experiments on the VisDrone2019 Datasets", "我们在四个具有挑战性的人群计数数据集上评估", "了我们的方法：VisDrone2019People&车辆和Shang-", "我们将原始VisDrone2019挑战数据集修改为两个", "haiTechA&B。四个数据集的统计信息如表2所示", "人群计数数据集（VisDrone2019车辆&人员），并", "VisDrone2019人[38]。我们将原始VisDrone2019", "在两个新数据集上进行广泛的实验。详情见表2", "对象检测数据集[37]与目标的边界框一起修改为人群", "", "我们实现了一个强大的VGG-16网络和最先进的", "计数注释。原始VisDrone2019数据集包含11个类别。", "方法MCNN[36]和CSRNet[15]，并在同一数" ]

{ "arxiv_id": "1909.03839", "language": "zh", "page_idx": 4 }

[ [ "0.3162", "0.2827", "0.3718", "0.281", "0.3718", "0.2886", "0.3162", "0.2903" ], [ "0.08826", "0.3018", "0.911", "0.3018", "0.911", "0.3164", "0.08826", "0.3164" ], [ "0.0866", "0.3176", "0.9126", "0.3176", "0.9126", "0.3308", "0.0866", "0.3308" ], [ "0.0866", "0.3333", "0.4795", "0.3333", "0.4795", "0.3467", "0.0866", "0.3467" ], [ "0.4478", "0.375", "0.6055", "0.377", "0.6055", "0.3901", "0.4478", "0.3882" ], [ "0.7236", "0.375", "0.8667", "0.3757", "0.8667", "0.3901", "0.7236", "0.3894" ], [ "0.3896", "0.3958", "0.4248", "0.3965", "0.4207", "0.4797", "0.3857", "0.4792" ], [ "0.1626", "0.377", "0.3096", "0.377", "0.3096", "0.3901", "0.1626", "0.3901" ], [ "0.392", "0.3965", "0.4102", "0.3818", "0.4265", "0.3945", "0.4084", "0.4092" ], [ "0.3652", "0.4072", "0.3882", "0.4072", "0.3882", "0.481", "0.3652", "0.481" ], [ "0.0948", "0.4167", "0.1054", "0.4167", "0.1054", "0.4363", "0.0948", "0.4363" ], [ "0.652", "0.4236", "0.6665", "0.4224", "0.671", "0.464", "0.6562", "0.4653" ], [ "0.0972", "0.4382", "0.10297", "0.4382", "0.10297", "0.4666", "0.0972", "0.4666" ], [ "0.3938", "0.473", "0.4192", "0.473", "0.4192", "0.4912", "0.3938", "0.4912" ], [ "0.558", "0.4749", "0.5645", "0.4749", "0.5645", "0.4778", "0.558", "0.4778" ], [ "0.3896", "0.4912", "0.4182", "0.4912", "0.4182", "0.504", "0.3896", "0.504" ], [ "0.29", "0.504", "0.298", "0.504", "0.298", "0.5083", "0.29", "0.5083" ], [ "0.2133", "0.512", "0.259", "0.512", "0.259", "0.523", "0.2133", "0.523" ], [ "0.505", "0.512", "0.5483", "0.512", "0.5483", "0.523", "0.505", "0.523" ], [ "0.773", "0.512", "0.819", "0.512", "0.819", "0.5234", "0.773", "0.5234" ], [ "0.4028", "0.529", "0.5947", "0.529", "0.5947", "0.5435", "0.4028", "0.5435" ], [ "0.699", "0.53", "0.859", "0.53", "0.859", "0.5435", "0.699", "0.5435" ], [ "0.0866", "0.5435", "0.9126", "0.5435", "0.9126", "0.558", "0.0866", "0.558" ], [ "0.12335", "0.531", "0.3179", "0.531", "0.3179", "0.5444", "0.12335", "0.5444" ], [ "0.08496", "0.56", "0.6445", "0.56", "0.6445", "0.5737", "0.08496", "0.5737" ], [ "0.728", "0.6724", "0.7563", "0.6704", "0.7573", "0.681", "0.7295", "0.683" ], [ "0.7017", "0.7505", "0.7812", "0.7524", "0.7812", "0.7676", "0.7017", "0.765" ], [ "0.1259", "0.8335", "0.1912", "0.8335", "0.1912", "0.8447", "0.1259", "0.8447" ], [ "0.3179", "0.834", "0.3743", "0.834", "0.3743", "0.8447", "0.3179", "0.8447" ], [ "0.5415", "0.837", "0.55", "0.837", "0.55", "0.8413", "0.5415", "0.8413" ], [ "0.6978", "0.8335", "0.7686", "0.8335", "0.7686", "0.8447", "0.6978", "0.8447" ], [ "0.0866", "0.853", "0.911", "0.853", "0.911", "0.867", "0.0866", "0.867" ], [ "0.08826", "0.869", "0.914", "0.869", "0.914", "0.8833", "0.08826", "0.8833" ], [ "0.08496", "0.884", "0.9126", "0.8833", "0.9126", "0.898", "0.08496", "0.8984" ], [ "0.0866", "0.8994", "0.8037", "0.8994", "0.8037", "0.9136", "0.0866", "0.9136" ] ]

[ "-66", "Fig. 5: Snapshots of RicMonk performing multiple backward brachiation cycles. First row corresponds to a zB", "maneuver (left to right), while second and third rows correspond to consecutive FB maneuvers (left to right). Green", "and red dots indicate start and end points respectively", "Velocity [rad/sl vs Time []", "Torque [Nm] vs Time []", "83849", "Position [radl vs Time []", "", "[s/", "pe.1", "noi", "S", "7.5", "H", "10.0", "6", "Time [s]", "Time Js", "Time [s]", "(b)Generalized velocityplot", "(c) Optimum input plot", "Fig: 6: Plots for the multiple backward brachiation maneuvers - one ZB maneuver, and two FB maneuvers in", "(a) Generalized position plot", "succession. Vertical dotted lines separate swing phases and grasp and releases", "3.7", "=6.1s", "=6.3s", "=7.2", "K", "t=8.5s", "Fig. 7: Snapshots of RicMonk performing multiple brachiation motions in the presence of mass uncertainty that", "is loaded asymmetrically. First row corresponds to a zF maneuver (left to right), while second and third rows", "correspond to consecutive BF maneuvers (left to right). In row 2, second and third images illustrate the robot arm", "stuck due to a disturbance in motion. Green and red dots indicate start and end points respectively" ]

{ "arxiv_id": "2403.15762", "language": "zh", "page_idx": 4 }

[ [ "0.1462", "0.0505", "0.728", "0.0505", "0.728", "0.0644", "0.1462", "0.0644" ], [ "0.1602", "0.127", "0.839", "0.13", "0.839", "0.1477", "0.1602", "0.1445" ], [ "0.3218", "0.154", "0.6772", "0.154", "0.6772", "0.1674", "0.3218", "0.1674" ], [ "0.1462", "0.1932", "0.3105", "0.1932", "0.3105", "0.207", "0.1462", "0.207" ], [ "0.7427", "0.1945", "0.8545", "0.1945", "0.8545", "0.2084", "0.7427", "0.2084" ], [ "0.1462", "0.209", "0.281", "0.209", "0.281", "0.2229", "0.1462", "0.2229" ], [ "0.74", "0.2102", "0.8545", "0.2102", "0.8545", "0.2235", "0.74", "0.2235" ], [ "0.1478", "0.2247", "0.5703", "0.2247", "0.5703", "0.238", "0.1478", "0.238" ], [ "0.1478", "0.2393", "0.3857", "0.2393", "0.3857", "0.254", "0.1478", "0.254" ], [ "0.1447", "0.2524", "0.3137", "0.254", "0.3137", "0.2683", "0.1447", "0.2676" ], [ "0.1462", "0.2778", "0.2754", "0.2778", "0.2754", "0.2922", "0.1462", "0.2922" ], [ "0.66", "0.281", "0.851", "0.281", "0.851", "0.2917", "0.66", "0.2917" ], [ "0.1478", "0.2954", "0.4167", "0.2954", "0.4167", "0.31", "0.1478", "0.31" ], [ "0.1462", "0.31", "0.4944", "0.3113", "0.4944", "0.3289", "0.1462", "0.3276" ], [ "0.1462", "0.3276", "0.29", "0.33", "0.29", "0.3447", "0.1462", "0.3423" ], [ "0.46", "0.4048", "0.54", "0.4048", "0.54", "0.4167", "0.46", "0.4167" ], [ "0.1772", "0.428", "0.819", "0.428", "0.819", "0.4414", "0.1772", "0.4414" ], [ "0.1772", "0.443", "0.819", "0.443", "0.819", "0.457", "0.1772", "0.457" ], [ "0.179", "0.4583", "0.822", "0.4583", "0.822", "0.4717", "0.179", "0.4717" ], [ "0.1772", "0.4734", "0.8203", "0.4734", "0.8203", "0.4873", "0.1772", "0.4873" ], [ "0.179", "0.488", "0.8203", "0.488", "0.8203", "0.5024", "0.179", "0.5024" ], [ "0.1772", "0.504", "0.8203", "0.504", "0.8203", "0.518", "0.1772", "0.518" ], [ "0.179", "0.5195", "0.819", "0.5195", "0.819", "0.5327", "0.179", "0.5327" ], [ "0.1765", "0.5327", "0.818", "0.5327", "0.818", "0.547", "0.1765", "0.547" ], [ "0.1772", "0.549", "0.819", "0.549", "0.819", "0.5625", "0.1772", "0.5625" ], [ "0.1805", "0.565", "0.819", "0.565", "0.819", "0.5757", "0.1805", "0.5757" ], [ "0.1772", "0.5776", "0.822", "0.579", "0.822", "0.5933", "0.1772", "0.5923" ], [ "0.1772", "0.5933", "0.576", "0.5933", "0.576", "0.607", "0.1772", "0.607" ], [ "0.1772", "0.613", "0.819", "0.6143", "0.819", "0.629", "0.1772", "0.6274" ], [ "0.1447", "0.6577", "0.3022", "0.6577", "0.3022", "0.672", "0.1447", "0.672" ], [ "0.1462", "0.6875", "0.853", "0.688", "0.853", "0.7026", "0.1462", "0.702" ], [ "0.1478", "0.705", "0.851", "0.705", "0.851", "0.7183", "0.1478", "0.7183" ], [ "0.1462", "0.7236", "0.853", "0.7236", "0.853", "0.7373", "0.1462", "0.7373" ], [ "0.1462", "0.7407", "0.853", "0.7407", "0.853", "0.754", "0.1462", "0.754" ], [ "0.1462", "0.757", "0.853", "0.757", "0.853", "0.771", "0.1462", "0.771" ], [ "0.1462", "0.7734", "0.853", "0.7734", "0.853", "0.787", "0.1462", "0.787" ], [ "0.1462", "0.7915", "0.853", "0.7915", "0.853", "0.8047", "0.1462", "0.8047" ], [ "0.1487", "0.808", "0.851", "0.808", "0.851", "0.8213", "0.1487", "0.8213" ], [ "0.1462", "0.8257", "0.853", "0.8257", "0.853", "0.839", "0.1462", "0.839" ], [ "0.1478", "0.8447", "0.8506", "0.8447", "0.8506", "0.855", "0.1478", "0.855" ], [ "0.1478", "0.86", "0.853", "0.86", "0.853", "0.8735", "0.1478", "0.8735" ], [ "0.1478", "0.879", "0.851", "0.879", "0.851", "0.8896", "0.1478", "0.8896" ], [ "0.492", "0.914", "0.5073", "0.914", "0.5073", "0.9287", "0.492", "0.9287" ] ]

[ "ICML2022 Workshop on Adaptive Experimental Design and Active Learning in the Real World", "Online Active Learning for Soft Sensor Development using", "Semi-SupervisedAutoencoders", "Davide Cacciarelli", "[email protected]", "Murat Kulahci", "[email protected]", "Department of Applied Mathematics and Computer Science", "Technical University of Denmark", "Kgs.Lyngby,Denmark", "John Tyssedal", "[email protected]", "DepartmentofMathematical Sciences", "Norwegian University of Science and Technology", "Trondheim, Norway", "Abstract", "Data-driven soft sensors are extensively used in industrial and chemical processes to pre-", "dict hard-to-measure process variables whose real value is difficult to track during routine", "operations. The regression models used by these sensors often require a large number of", "labeled examples, yet obtaining the label information can be very expensive given the high", "time and cost required by quality inspections. In this context, active learning methods can", "be highly beneficial as they can suggest the most informative labels to query. However,", "most of the active learning strategies proposed for regression focus on the offline setting.", "In this work, we adapt some of these approaches to the stream-based scenario and show", "how they can be used to select the most informative data points. We also demonstrate how", "to use a semi-supervised architecture based on orthogonal autoencoders tolearn salient", "features in a lower dimensional space. The Tennessee Eastman Process is used to compare", "the predictive performance of the proposed approaches.", "Keywords: Active Learning, Semi-Supervised Learning, Linear Regression, Autoencoder.", "1.Introduction", "在工业操作中，软传感器经常用于实时预测难以测量的过程变量，以及支持系统备份策略，", "假设分析，传感器验证和故障诊断（Fortunaetal.，2007）。软传感器分为两种类型：模型", "驱动的传感器，用于描述正常运行条件下过程的理想稳态；数据驱动的传感器，用于更好", "地近似真实过程条件（Kadlecetal.，2009）。训练软传感器开发中使用的回归模型需要许多", "标记的观测值，但在工业环境中，数据通常仅以未标记的形式丰富获取产品信息可能既昂", "贵又耗时，因为它可能需要人类专家的干预或使用昂贵的测试设备。因此，主动学习对于", "减少实现引人注目的预测性能所需的标签数量变得越来越有用。。基于主动学习的抽样方案", "使用一些评估标准来评估未标记数据点的信息量，并优先考虑用于构建模型的最有用实例", "的标记。可以根据如何将未标记的实例馈送到学习者中，然后选择由oracle（Settles，2009)", "标记的方式来识别三个宏方案。第一种情况称为成员香合成，它充许学习者查询合成生", "成的实例的标签，而不是从流程分布中采样的标签。第二种情况是基于流的主动学习，也", "称为选择性采样。它表示一种情况，在这种情况下，实例是按顺序绘制的，学习者必须立即", "1" ]

{ "arxiv_id": "2212.13067", "language": "zh", "page_idx": 0 }

[ [ "0.1478", "0.1187", "0.8506", "0.1187", "0.8506", "0.132", "0.1478", "0.132" ], [ "0.1487", "0.137", "0.4314", "0.137", "0.4314", "0.1477", "0.1487", "0.1477" ], [ "0.501", "0.137", "0.8486", "0.137", "0.8486", "0.1477", "0.501", "0.1477" ], [ "0.1487", "0.154", "0.3987", "0.154", "0.3987", "0.1642", "0.1487", "0.1642" ], [ "0.4004", "0.1547", "0.4192", "0.1547", "0.4192", "0.1622", "0.4004", "0.1622" ], [ "0.4429", "0.1547", "0.4631", "0.1547", "0.4631", "0.1622", "0.4429", "0.1622" ], [ "0.5063", "0.154", "0.8506", "0.154", "0.8506", "0.1642", "0.5063", "0.1642" ], [ "0.1478", "0.1692", "0.4878", "0.1692", "0.4878", "0.1831", "0.1478", "0.1831" ], [ "0.486", "0.1692", "0.6763", "0.1692", "0.6763", "0.1831", "0.486", "0.1831" ], [ "0.1748", "0.1875", "0.8545", "0.1875", "0.8545", "0.2008", "0.1748", "0.2008" ], [ "0.1487", "0.2039", "0.853", "0.2039", "0.853", "0.2172", "0.1487", "0.2172" ], [ "0.1478", "0.2203", "0.851", "0.2203", "0.851", "0.2343", "0.1478", "0.2343" ], [ "0.1462", "0.238", "0.853", "0.238", "0.853", "0.252", "0.1462", "0.252" ], [ "0.1462", "0.2551", "0.1993", "0.2551", "0.1993", "0.2695", "0.1462", "0.2695" ], [ "0.1447", "0.2954", "0.2942", "0.2954", "0.2942", "0.3093", "0.1447", "0.3093" ], [ "0.1478", "0.3213", "0.851", "0.3213", "0.851", "0.336", "0.1478", "0.336" ], [ "0.1462", "0.3403", "0.851", "0.3403", "0.851", "0.3542", "0.1462", "0.3542" ], [ "0.1447", "0.356", "0.839", "0.356", "0.839", "0.3706", "0.1447", "0.3706" ], [ "0.4004", "0.3984", "0.5986", "0.3984", "0.5986", "0.4167", "0.4004", "0.4167" ], [ "0.4795", "0.414", "0.5063", "0.414", "0.5063", "0.432", "0.4795", "0.432" ], [ "0.1462", "0.4458", "0.853", "0.4463", "0.853", "0.461", "0.1462", "0.4602" ], [ "0.4346", "0.4666", "0.596", "0.4749", "0.5923", "0.5176", "0.4307", "0.509" ], [ "0.4126", "0.4893", "0.4468", "0.4893", "0.4468", "0.5005", "0.4126", "0.5005" ], [ "0.4453", "0.4944", "0.4927", "0.507", "0.4844", "0.526", "0.437", "0.5146" ], [ "0.83", "0.4893", "0.853", "0.4893", "0.853", "0.504", "0.83", "0.504" ], [ "0.1478", "0.536", "0.851", "0.536", "0.851", "0.55", "0.1478", "0.55" ], [ "0.1478", "0.552", "0.7686", "0.552", "0.7686", "0.565", "0.1478", "0.565" ], [ "0.1462", "0.587", "0.384", "0.587", "0.384", "0.601", "0.1462", "0.601" ], [ "0.1478", "0.6123", "0.851", "0.6123", "0.851", "0.6255", "0.1478", "0.6255" ], [ "0.1447", "0.629", "0.853", "0.63", "0.853", "0.6445", "0.1447", "0.643" ], [ "0.1447", "0.646", "0.851", "0.6465", "0.851", "0.661", "0.1447", "0.66" ], [ "0.1478", "0.6655", "0.853", "0.6655", "0.853", "0.6787", "0.1478", "0.6787" ], [ "0.1462", "0.681", "0.492", "0.681", "0.492", "0.6943", "0.1462", "0.6943" ], [ "0.3823", "0.7134", "0.613", "0.7134", "0.613", "0.728", "0.3823", "0.728" ], [ "0.831", "0.7153", "0.853", "0.7153", "0.853", "0.73", "0.831", "0.73" ], [ "0.1462", "0.747", "0.853", "0.7476", "0.853", "0.762", "0.1462", "0.761" ], [ "0.1462", "0.7646", "0.853", "0.7646", "0.853", "0.779", "0.1462", "0.779" ], [ "0.1462", "0.782", "0.6685", "0.782", "0.6685", "0.7954", "0.1462", "0.7954" ], [ "0.1462", "0.8174", "0.384", "0.8174", "0.384", "0.832", "0.1462", "0.832" ], [ "0.1478", "0.8423", "0.853", "0.8423", "0.853", "0.857", "0.1478", "0.857" ], [ "0.1478", "0.86", "0.8545", "0.86", "0.8545", "0.8735", "0.1478", "0.8735" ], [ "0.1462", "0.8774", "0.8545", "0.8774", "0.8545", "0.8916", "0.1462", "0.8916" ], [ "0.492", "0.914", "0.5073", "0.914", "0.5073", "0.9287", "0.492", "0.9287" ] ]

[ "决定是保留实例并查询其标签还是丢弃它。第三种也是最后一种情况是基于池的主动学习", "它描述了一种情况，即一次性收集大", "已的数据并提供给学习者，学习者可以对所有", "数据点进行排名并选择信息最丰", "富", "改护", "虽然最近儿年许多研究人员一直在研究主动", "学习，但基于池的主动学习分类受到了最多", "的关注（Caietal.,2013)", "在这项工作中，我们专注于基于流的主动学习（CacciarelliandKulahci，2023），它代表", "了一项更困难的任务，因为学习者在决定查询哪些标签之前无法观察所有可用的观察结果。", "我们认为，这种情况准确地反映了大批量生产过程，在这些过程中，样品的处理速度非常", "快，标签不再可检索。对于具有类似属性的所有工业过程，应考虑并优先考虑基于流的主", "动学习", "2. Background", "在回归建模中，我们尝试学习一个函数f：αERP→yER来预测与其他过程变量αERP", "相关的质量特性或难以测量的变量yER。因此，对于许多主动学习方法（Caietal.,2013)", "，我们假设带有n观察值的标记数据集=(ai，yi)n-1,ciERP,yER}最初可用于拟合", "f(c; β) = Biai =βT a", "1=0", "类型的线性回归模型，其中ao=1是截距项，c和i=1，,P是p过程变量。通过最小化", "之(u- (a)", "=", "ni=l", "(1)", "给出的平方误差损失来估计参数在建立了初始模型之后，我们的目标是通过评估未标记的", "数据点来获取其他观察值，直到满足预算约束为止。下面介绍了一些常见的方法", "2.1 Mahalanobis Distance", "如果我们仅检查特征空间，则在为训练集收集实例时可能追求的理想属性是确保观察值", "之间的多样性。在统计过程控制（SPC）中广泛用于检测异常数据点（Hotelling，1947）的", "HotellingT2控制图可以实现。就像我们在SPC中所做的那样，在这种情况下，我们使用", "马氏距离（方程2）来测量新的未标记实例与当前训练集L中的观察值之间的差异。新的", "未标记实例α的HotellingT2统计量计算为", "T2(α)= (α-α)Ts-1(α-)", "(2)", "，其中ε和S分别对应于C的样本均值向量和样本协方差矩阵。Geetal.（2014）已将此", "方法扩展为主成分回归（PCR）模型，他提出了取决于HotellingT²统计量和平方预测误", "差的采样指数。在这种情况下，采样函数仅由argmacaT2(α）表示", "2.2 Query By Committee", "虽然以前的方法仅考虑特征空间，但委员会查询（QBC）会尝试评估有关响应的不确定性。", "最初针对分类问题引入的此策略已由Burbidgeetal.（1997）扩展到回归任务。主要的直觉", "是，通过构建在原始训练集B(K）=（f1，f2，.…，fk)的引导副本上训练的回归模型的集合，", "2" ]

{ "arxiv_id": "2212.13067", "language": "zh", "page_idx": 1 }

[ [ "0.1478", "0.1187", "0.851", "0.1187", "0.851", "0.132", "0.1478", "0.132" ], [ "0.1478", "0.1351", "0.5435", "0.1351", "0.5435", "0.1483", "0.1478", "0.1483" ], [ "0.4224", "0.1666", "0.6313", "0.1686", "0.63", "0.2203", "0.4207", "0.2184" ], [ "0.433", "0.1718", "0.4805", "0.1572", "0.492", "0.178", "0.4436", "0.1925" ], [ "0.38", "0.1787", "0.4314", "0.1787", "0.4314", "0.1932", "0.38", "0.1932" ], [ "0.4167", "0.1843", "0.4429", "0.1881", "0.4395", "0.1976", "0.4143", "0.1925" ], [ "0.83", "0.1787", "0.853", "0.1787", "0.853", "0.1932", "0.83", "0.1932" ], [ "0.1462", "0.2229", "0.851", "0.2229", "0.851", "0.2374", "0.1462", "0.2374" ], [ "0.1447", "0.2405", "0.7876", "0.2405", "0.7876", "0.254", "0.1447", "0.254" ], [ "0.1447", "0.2747", "0.4084", "0.276", "0.4084", "0.2903", "0.1447", "0.289" ], [ "0.1478", "0.2998", "0.853", "0.2998", "0.853", "0.3132", "0.1478", "0.3132" ], [ "0.1462", "0.3176", "0.853", "0.3176", "0.853", "0.3308", "0.1462", "0.3308" ], [ "0.1462", "0.3352", "0.851", "0.3352", "0.851", "0.3484", "0.1462", "0.3484" ], [ "0.1478", "0.351", "0.683", "0.351", "0.683", "0.3643", "0.1478", "0.3643" ], [ "0.3367", "0.377", "0.6587", "0.3865", "0.6567", "0.4216", "0.335", "0.4124" ], [ "0.3757", "0.3977", "0.423", "0.4072", "0.4182", "0.4255", "0.3708", "0.4172" ], [ "0.1544", "0.443", "0.165", "0.443", "0.165", "0.4475", "0.1544", "0.4475" ], [ "0.1602", "0.4375", "0.853", "0.4363", "0.853", "0.4507", "0.1602", "0.452" ], [ "0.1462", "0.4526", "0.7695", "0.4539", "0.7695", "0.4685", "0.1462", "0.4673" ], [ "0.343", "0.481", "0.653", "0.481", "0.653", "0.512", "0.343", "0.512" ], [ "0.4028", "0.5024", "0.429", "0.5024", "0.429", "0.5166", "0.4028", "0.5166" ], [ "0.603", "0.5005", "0.6313", "0.5005", "0.6313", "0.5176", "0.603", "0.5176" ], [ "0.5884", "0.5215", "0.6445", "0.5215", "0.6445", "0.537", "0.5884", "0.537" ], [ "0.83", "0.5234", "0.853", "0.5234", "0.853", "0.5386", "0.83", "0.5386" ], [ "0.452", "0.532", "0.5884", "0.529", "0.5884", "0.547", "0.452", "0.55" ], [ "0.607", "0.539", "0.6313", "0.5425", "0.6274", "0.5566", "0.603", "0.5537" ], [ "0.455", "0.561", "0.607", "0.561", "0.607", "0.5757", "0.455", "0.5757" ], [ "0.1462", "0.5933", "0.851", "0.5933", "0.851", "0.607", "0.1462", "0.607" ], [ "0.1462", "0.61", "0.653", "0.61", "0.653", "0.6274", "0.1462", "0.6274" ], [ "0.1447", "0.6494", "0.3708", "0.6494", "0.3708", "0.664", "0.1447", "0.664" ], [ "0.1462", "0.677", "0.853", "0.677", "0.853", "0.69", "0.1462", "0.69" ], [ "0.1462", "0.6934", "0.851", "0.6934", "0.851", "0.707", "0.1462", "0.707" ], [ "0.1478", "0.711", "0.853", "0.711", "0.853", "0.7246", "0.1478", "0.7246" ], [ "0.1462", "0.728", "0.853", "0.728", "0.853", "0.741", "0.1462", "0.741" ], [ "0.1478", "0.7456", "0.853", "0.7456", "0.853", "0.759", "0.1478", "0.759" ], [ "0.1462", "0.762", "0.853", "0.762", "0.853", "0.776", "0.1462", "0.776" ], [ "0.1478", "0.78", "0.8545", "0.78", "0.8545", "0.7935", "0.1478", "0.7935" ], [ "0.1447", "0.7954", "0.406", "0.797", "0.406", "0.811", "0.1447", "0.81" ], [ "0.4167", "0.829", "0.584", "0.829", "0.584", "0.8423", "0.4167", "0.8423" ], [ "0.8286", "0.8276", "0.8555", "0.8276", "0.8555", "0.8447", "0.8286", "0.8447" ], [ "0.1478", "0.86", "0.853", "0.86", "0.853", "0.8735", "0.1478", "0.8735" ], [ "0.1478", "0.877", "0.853", "0.877", "0.853", "0.89", "0.1478", "0.89" ], [ "0.492", "0.914", "0.5073", "0.914", "0.5073", "0.9287", "0.492", "0.9287" ] ]

[ "我们可以近似预测方差的分布。一且建立了集成，我们就可以测量委员会成员对每个未标", "记的观察&所做预测的方差。此方差或歧义计算为", "2( (a)", "LA", "a(α) =", "", "(3)", "，其中yk(）是集合成员做出的预测的平均值。然后，采样函数简单地变为argmaaaa(a）", "。关键的直觉是，如果许多模型不同意与实例关联的标签，则该实例是模棱两可的", "2.3 Expected Model Change", "由Caietal.（2013）引入，预期的模型更改（EMC）建议查询未标记的示例，该示例将导", "致当前模型参数的最大更改，如果我们知道其标签。模型变化被测量为当前模型参数与在", "扩大的训练集+=CU（α+，y+）上拟合模型后获得的参数之间的差。损失的梯度用于估计", "模型变化。考虑到增强的训练集L+，公式1中所示的损失函数变为", "=(u-(a)+(u+- (a+)2", "ni=l", "", "其中最后一项（以下称为ea+（β））表示使用在L上训练的模型测量的损失与在L+上", "训练的模型之间的差异。边际损失l+（β）相对于新点&+中的参数β的导数由", "e = la(P) = 2(y+ - (a+)(t)", "ap", "ap", "βTa+", "(4)", "= 2(y+- f(α+))", "ap", "= 2(y+- f(α+))α+", "给出。由于我们不知道&+，y+的真实标签，它将被替换为预测fi（+）由引导集成的成", "员。最后给出了argmaca=1ll(fi(a）-f(a))all的采样函数", "3. Proposed Approach", "鉴于不可能对未标记的实例进行实时排名并确定性地优化采样标准，我们建议利用未标记", "的数据对传入数据点的信息量施加阈值或控制上限（UCL）。可以通过观察过程一段时间而", "不对产品信息y进行采样，或者使用已经以历史数据库H=[(αi),a,ERP}形式提供的数", "据来获取未标记的数据池。基于池的主动学习和在线主动学习之间的主要区别在于，与孔", "有关的观察结果的标签无法再查询，因为它们仅以数字形式存在，并且关联的物理部件或", "组件不再可用。H中的数据用于估计方程2、3和4中的标准所采用的统计量的分布。在这", "项研究中，我们采用了高斯核的核密度估计。然后通过指定适当的采样率α来确定UCL。", "对于给定的标准丁，值定义为", "P(J(α) ≥ UCL) = α", "(5)", "根据特定标准J（Cacciarellietal.，2022，2023），使用从方程式5获得的UCL，我们应该", "只收集Q%最具信息性的数据点。在这项工作中，我们使用马氏距离，模糊性和预期的模", "3" ]

{ "arxiv_id": "2212.13067", "language": "zh", "page_idx": 2 }

[ [ "0.1447", "0.11743", "0.5728", "0.1193", "0.5728", "0.1332", "0.1447", "0.1307" ], [ "0.581", "0.1187", "0.853", "0.1187", "0.853", "0.132", "0.581", "0.132" ], [ "0.1487", "0.1364", "0.3594", "0.1364", "0.3594", "0.1465", "0.1487", "0.1465" ], [ "0.653", "0.137", "0.851", "0.137", "0.851", "0.1477", "0.653", "0.1477" ], [ "0.1478", "0.1528", "0.2737", "0.1528", "0.2737", "0.1635", "0.1478", "0.1635" ], [ "0.2917", "0.1547", "0.3757", "0.1547", "0.3757", "0.1622", "0.2917", "0.1622" ], [ "0.6787", "0.1515", "0.8506", "0.154", "0.8506", "0.1654", "0.6787", "0.1635" ], [ "0.1478", "0.1674", "0.3513", "0.1674", "0.3513", "0.1812", "0.1478", "0.1812" ], [ "0.652", "0.1749", "0.6787", "0.1749", "0.6787", "0.1787", "0.652", "0.1787" ], [ "0.693", "0.1692", "0.853", "0.1692", "0.853", "0.1831", "0.693", "0.1831" ], [ "0.1478", "0.1863", "0.259", "0.1863", "0.259", "0.1995", "0.1478", "0.1995" ], [ "0.6895", "0.1895", "0.7563", "0.1895", "0.7563", "0.1963", "0.6895", "0.1963" ], [ "0.761", "0.1875", "0.8486", "0.1875", "0.8486", "0.1989", "0.761", "0.1989" ], [ "0.1478", "0.2026", "0.3628", "0.2026", "0.3628", "0.2166", "0.1478", "0.2166" ], [ "0.3594", "0.2084", "0.3635", "0.2084", "0.3635", "0.2134", "0.3594", "0.2134" ], [ "0.621", "0.2084", "0.625", "0.2084", "0.625", "0.2134", "0.621", "0.2134" ], [ "0.6396", "0.2052", "0.851", "0.2052", "0.851", "0.2153", "0.6396", "0.2153" ], [ "0.625", "0.2203", "0.853", "0.2203", "0.853", "0.2343", "0.625", "0.2343" ], [ "0.3757", "0.2102", "0.3857", "0.2102", "0.3857", "0.2134", "0.3757", "0.2134" ], [ "0.1487", "0.2216", "0.3652", "0.2216", "0.3652", "0.2317", "0.1487", "0.2317" ], [ "0.1462", "0.2374", "0.853", "0.2374", "0.853", "0.2507", "0.1462", "0.2507" ], [ "0.1478", "0.2551", "0.853", "0.2551", "0.853", "0.2683", "0.1478", "0.2683" ], [ "0.1478", "0.2727", "0.853", "0.2727", "0.853", "0.286", "0.1478", "0.286" ], [ "0.1478", "0.289", "0.853", "0.289", "0.853", "0.303", "0.1478", "0.303" ], [ "0.1462", "0.3062", "0.8545", "0.3062", "0.8545", "0.3193", "0.1462", "0.3193" ], [ "0.1462", "0.324", "0.4314", "0.324", "0.4314", "0.3372", "0.1462", "0.3372" ], [ "0.3733", "0.3623", "0.6235", "0.3623", "0.6235", "0.3757", "0.3733", "0.3757" ], [ "0.4944", "0.4048", "0.505", "0.4048", "0.505", "0.4167", "0.4944", "0.4167" ], [ "0.3105", "0.4204", "0.353", "0.4204", "0.353", "0.4312", "0.3105", "0.4312" ], [ "0.4983", "0.4172", "0.5024", "0.4172", "0.5024", "0.4255", "0.4983", "0.4255" ], [ "0.492", "0.4255", "0.5063", "0.4255", "0.5063", "0.4443", "0.492", "0.4443" ], [ "0.6436", "0.4204", "0.6846", "0.4204", "0.6846", "0.4312", "0.6436", "0.4312" ], [ "0.3096", "0.4622", "0.3506", "0.4622", "0.3506", "0.4766", "0.3096", "0.4766" ], [ "0.501", "0.464", "0.5415", "0.4622", "0.5435", "0.4766", "0.5024", "0.4792" ], [ "0.642", "0.4634", "0.681", "0.4634", "0.681", "0.4792", "0.642", "0.4792" ], [ "0.638", "0.5195", "0.671", "0.5215", "0.669", "0.5327", "0.636", "0.531" ], [ "0.6274", "0.5327", "0.6846", "0.5327", "0.6846", "0.5444", "0.6274", "0.5444" ], [ "0.6445", "0.5645", "0.6807", "0.5645", "0.6807", "0.58", "0.6445", "0.58" ], [ "0.286", "0.5933", "0.7114", "0.5923", "0.7114", "0.607", "0.286", "0.608" ], [ "0.1748", "0.632", "0.8506", "0.632", "0.8506", "0.6426", "0.1748", "0.6426" ], [ "0.1478", "0.6484", "0.853", "0.6484", "0.853", "0.662", "0.1478", "0.662" ], [ "0.1462", "0.6655", "0.851", "0.6655", "0.851", "0.6787", "0.1462", "0.6787" ], [ "0.1447", "0.681", "0.853", "0.6816", "0.853", "0.6963", "0.1447", "0.696" ], [ "0.1478", "0.7007", "0.8447", "0.7007", "0.8447", "0.714", "0.1478", "0.714" ], [ "0.841", "0.7056", "0.8506", "0.7056", "0.8506", "0.711", "0.841", "0.711" ], [ "0.1462", "0.7173", "0.8545", "0.7173", "0.8545", "0.731", "0.1462", "0.731" ], [ "0.1462", "0.7344", "0.54", "0.7344", "0.54", "0.7476", "0.1462", "0.7476" ], [ "0.1462", "0.778", "0.3015", "0.778", "0.3015", "0.7915", "0.1462", "0.7915" ], [ "0.1478", "0.808", "0.851", "0.808", "0.851", "0.8213", "0.1478", "0.8213" ], [ "0.1447", "0.8237", "0.8545", "0.8247", "0.8545", "0.839", "0.1447", "0.8384" ], [ "0.1462", "0.8423", "0.853", "0.8423", "0.853", "0.856", "0.1462", "0.856" ], [ "0.1462", "0.86", "0.8545", "0.86", "0.8545", "0.8735", "0.1462", "0.8735" ], [ "0.1462", "0.877", "0.2361", "0.877", "0.2361", "0.8916", "0.1462", "0.8916" ], [ "0.492", "0.914", "0.5073", "0.914", "0.5073", "0.9287", "0.492", "0.9287" ] ]

[ "型变化作为抽样标准了来测试基于流的主动学习例程。", "在开始主动学习例程并收集其他观", "察结果之前：我们还建议道", "网络来使用平监督体系结", "构。通过半监督", "，我", "如何提取相关特征，这", "些特征可能比原始输入特行", "", "文献中已经证明可以", "用半监督（Fru", "万法", "已经提出了", "自动编码器来处理故障分类", "", "", "中的半监督学习。最近，还", "etal.，2021）中研究了自动编", "", "在软传感器应用程序（Yuan", "码器，但尚未评估它们对基于流的主动学习方案的贡献。在这项工作中，我们建议使用半监", "督架构，如图1所示。正交自编码器（OAE）用于特征提取，然后将编码的特征用作线性", "回归模型中的预测因子。OAE是一种自动编码器网络，可最大程度地减少正交损耗，该正", "交损耗由平方重构损耗和正交正则化项（Wangetal.，2019；CacciarelliandKulahci，2022）", "组成。正则化由参数入加权，可鼓励网络在其瓶颈中学习不相关的特征。这对于缓解回归", "建模阶段的多重共线性问题特别有益", "Ortho-Loss:(x,z,)=x-2+zz-1", "0", "Encoder", "", "", "Decoder", "XERP", "ZERK", "RERP", "Linear", "Regression", "DER", "Figure l: Semi-supervised architecture based on OAE.", "半监督模型的主要优点是提取的特征比原始过程变量更具表现力。但是，如果瓶颈的维", "度低于输入特征之一，则主动学习还有一个额外的好处。实际上，由于提供的大多数主动", "学习方法都是基于模型的，因此需要初始数量的标签。特别是QBC和EMC，采用线性回", "归模型的预测来选择应查询的数据点。为了唯一地确定回归模型的系数，我们需要的观测", "数大于要估计的参数β的数量。这个初始的观察集通常是随机收集的（Caietal.，2013）", "", "因此，通过降低参数β的维数，我们将能够预测主动收集阶段，并在相同的实验成本下获", "得更稳健的估计。完整的主动学习例程报告附录A", "4. Experiments", "田纳西州伊士曼过程（TEP）被认为是测试过程控制方法（Ricker，1996:Capacietal.，2019）", "的黄金标准基准，最近，它也被用于验证主动学习和软传感器开发方法（Zhuetal.，2015;", "Grbicetal.,2013）。在回归建模中用作预测变量的变量与Zhuetal.（2015）和Grbicetal.", "(2013）使用的16个受控过程变量相同，并且连续响应是Stream9E，这是一种属于吹扫流", "的成分测量", "4" ]

{ "arxiv_id": "2212.13067", "language": "zh", "page_idx": 3 }

[ [ "0.3882", "0.1256", "0.4631", "0.1256", "0.4631", "0.1332", "0.3882", "0.1332" ], [ "0.7983", "0.125", "0.83", "0.125", "0.83", "0.1332", "0.7983", "0.1332" ], [ "0.3857", "0.1332", "0.4844", "0.1351", "0.4844", "0.1465", "0.3857", "0.1445" ], [ "0.7837", "0.1338", "0.831", "0.1338", "0.831", "0.1459", "0.7837", "0.1459" ], [ "0.179", "0.1401", "0.1854", "0.1401", "0.1854", "0.1465", "0.179", "0.1465" ], [ "0.8", "0.1414", "0.831", "0.1459", "0.8286", "0.1578", "0.7974", "0.154" ], [ "0.7915", "0.1477", "0.7983", "0.1477", "0.7983", "0.1528", "0.7915", "0.1528" ], [ "0.785", "0.1547", "0.8022", "0.1547", "0.8022", "0.1635", "0.785", "0.1635" ], [ "0.8013", "0.1559", "0.8286", "0.1559", "0.8286", "0.1642", "0.8013", "0.1642" ], [ "0.5186", "0.1572", "0.5317", "0.1572", "0.5317", "0.161", "0.5186", "0.161" ], [ "0.1772", "0.1604", "0.1929", "0.1642", "0.1887", "0.1749", "0.1732", "0.1704" ], [ "0.1765", "0.1875", "0.1912", "0.1875", "0.1912", "0.1957", "0.1765", "0.1957" ], [ "0.1626", "0.2122", "0.1854", "0.2122", "0.1854", "0.2329", "0.1626", "0.2329" ], [ "0.5034", "0.214", "0.513", "0.214", "0.513", "0.226", "0.5034", "0.226" ], [ "0.179", "0.26", "0.1871", "0.262", "0.1847", "0.2651", "0.1772", "0.2644" ], [ "0.1765", "0.3062", "0.1912", "0.3062", "0.1912", "0.3132", "0.1765", "0.3132" ], [ "0.38", "0.3213", "0.3865", "0.327", "0.38", "0.3333", "0.3733", "0.3276" ], [ "0.4314", "0.3213", "0.4453", "0.3213", "0.4453", "0.332", "0.4314", "0.332" ], [ "0.546", "0.3225", "0.5596", "0.3225", "0.5596", "0.3308", "0.546", "0.3308" ], [ "0.6045", "0.3225", "0.619", "0.3225", "0.619", "0.3308", "0.6045", "0.3308" ], [ "0.2573", "0.3225", "0.2737", "0.3225", "0.2737", "0.3308", "0.2573", "0.3308" ], [ "0.2051", "0.324", "0.2141", "0.324", "0.2141", "0.3308", "0.2051", "0.3308" ], [ "0.2998", "0.3289", "0.3865", "0.3289", "0.3865", "0.3403", "0.2998", "0.3403" ], [ "0.642", "0.3289", "0.732", "0.33", "0.732", "0.3416", "0.642", "0.3403" ], [ "0.7812", "0.324", "0.786", "0.324", "0.786", "0.33", "0.7812", "0.33" ], [ "0.2197", "0.3484", "0.4355", "0.3499", "0.4355", "0.3643", "0.2197", "0.363" ], [ "0.6006", "0.3484", "0.746", "0.3499", "0.746", "0.3643", "0.6006", "0.363" ], [ "0.1462", "0.3718", "0.853", "0.3718", "0.853", "0.3865", "0.1462", "0.3865" ], [ "0.1462", "0.387", "0.853", "0.387", "0.853", "0.4016", "0.1462", "0.4016" ], [ "0.1462", "0.406", "0.851", "0.406", "0.851", "0.4204", "0.1462", "0.4204" ], [ "0.1462", "0.4236", "0.492", "0.4236", "0.492", "0.4375", "0.1462", "0.4375" ], [ "0.1748", "0.461", "0.853", "0.461", "0.853", "0.4749", "0.1748", "0.4749" ], [ "0.1478", "0.4778", "0.851", "0.4778", "0.851", "0.4924", "0.1478", "0.4924" ], [ "0.1478", "0.4944", "0.8506", "0.4944", "0.8506", "0.509", "0.1478", "0.509" ], [ "0.1487", "0.512", "0.851", "0.512", "0.851", "0.5264", "0.1487", "0.5264" ], [ "0.1487", "0.53", "0.853", "0.53", "0.853", "0.5435", "0.1487", "0.5435" ], [ "0.1478", "0.547", "0.853", "0.547", "0.853", "0.561", "0.1478", "0.561" ], [ "0.1447", "0.563", "0.853", "0.5645", "0.853", "0.579", "0.1447", "0.5776" ], [ "0.1462", "0.579", "0.853", "0.5806", "0.853", "0.595", "0.1462", "0.5933" ], [ "0.1462", "0.5977", "0.851", "0.5977", "0.851", "0.6123", "0.1462", "0.6123" ], [ "0.1447", "0.637", "0.2852", "0.637", "0.2852", "0.651", "0.1447", "0.651" ], [ "0.1478", "0.6636", "0.853", "0.6636", "0.853", "0.678", "0.1478", "0.678" ], [ "0.1462", "0.681", "0.853", "0.681", "0.853", "0.6943", "0.1462", "0.6943" ], [ "0.1478", "0.6978", "0.853", "0.6978", "0.853", "0.711", "0.1478", "0.711" ], [ "0.1478", "0.7153", "0.851", "0.7153", "0.851", "0.729", "0.1478", "0.729" ], [ "0.1462", "0.731", "0.8486", "0.731", "0.8486", "0.744", "0.1462", "0.744" ], [ "0.1462", "0.75", "0.343", "0.75", "0.343", "0.763", "0.1462", "0.763" ], [ "0.4944", "0.9155", "0.5063", "0.9155", "0.5063", "0.9253", "0.4944", "0.9253" ] ]

[ "InputFeatures", "RND", "ExtractedFeatures", "HOT", "L", "QBC", "-", "十", "EMC", "YAE", "0.9", "0.8", "", "ASW", "03", "0.3", "", "60", "20", "30", "30", "20", "Labeled Examples", "Labeled Examples", "", "(a) Semi-Supervised Learning", "(b) Active Learning", "Figure 2: (a) shows the learning curves with random sampling using the original process", "variables and the features extracted by the OAE and (b) shows the learning curves of", "different active learning methods: random (RND), Hotelling T2 (HOT), query by committee", "(QBC), and expected model change (EMC).", "图2a显示了建议的半监督架构如何提高预测性能。数据在两种情况下都是随机抽样的，", "但两个线性模型是使用原始过程变量和OAE提取的特征进行拟合的。为了确保两条学习曲", "线之间的可比性，当收集的观察值超过16时，我们拟合了第一个模型。我们认为这种改进", "是由于OAE能够在其编码特征中表达数据中的非线性关系，并且对于提取的特征，我们具", "有相同数量的观测值，以估计较少数量的参数。在图2b中，我们尝试通过使用提出的主动", "学习策略来改善半监督结果。很明显，在推荐信息最丰富的数据点方面，EMC和QBC始", "终优于被动随机方法。相反，马氏距离似乎会恶化预测性能。我们认为这可能是由于具有高", "T2统计量的数据点可能是异常值，其包含在训练集中最终会降低性能。应该注意的是，在", "图2b中，所有的采样策略都使用OAE提取的特征。实验设置和培训细节在附录B中报告", "5. Conclusion", "工业数据通常仅在未标记的情况下可用，因为质量检查和手动注释任务既昂贵又耗时。在", "这项工作中，我们提出了一种基于OAE的半监督模型，用于提取相关特征并减少多重共线", "性。最重要的是，我们回顾并调整了一些最广泛使用的线性回归主动学习策略的在线设置。", "分析表明，如何正确使用历史数据并考虑预期响应，可以更快地减少预测误差。对于未来", "的研究，我们将考虑更先进的架构，如LSTM自动编码器或变压器，以获得考虑到数据中", "的时间依赖性的编码特征", "5" ]

{ "arxiv_id": "2212.13067", "language": "zh", "page_idx": 4 }

[ [ "0.33", "0.1332", "0.642", "0.1332", "0.642", "0.1477", "0.33", "0.1477" ], [ "0.1805", "0.1559", "0.7935", "0.1559", "0.7935", "0.1704", "0.1805", "0.1704" ], [ "0.1447", "0.2039", "0.2769", "0.202", "0.2769", "0.2197", "0.1447", "0.2216" ], [ "0.3145", "0.2052", "0.4558", "0.2052", "0.4558", "0.2197", "0.3145", "0.2197" ], [ "0.5063", "0.2052", "0.619", "0.2052", "0.619", "0.2197", "0.5063", "0.2197" ], [ "0.6763", "0.2052", "0.8286", "0.2052", "0.8286", "0.2197", "0.6763", "0.2197" ], [ "0.1772", "0.2235", "0.2385", "0.2235", "0.2385", "0.238", "0.1772", "0.238" ], [ "0.3145", "0.2247", "0.451", "0.2247", "0.451", "0.238", "0.3145", "0.238" ], [ "0.4944", "0.2247", "0.6284", "0.2247", "0.6284", "0.238", "0.4944", "0.238" ], [ "0.6787", "0.2247", "0.819", "0.2247", "0.819", "0.238", "0.6787", "0.238" ], [ "0.282", "0.2632", "0.4558", "0.2632", "0.4558", "0.2778", "0.282", "0.2778" ], [ "0.5522", "0.2615", "0.6665", "0.2615", "0.6665", "0.279", "0.5522", "0.279" ], [ "0.3022", "0.2827", "0.4355", "0.2827", "0.4355", "0.2966", "0.3022", "0.2966" ], [ "0.524", "0.2827", "0.6914", "0.2827", "0.6914", "0.2966", "0.524", "0.2966" ], [ "0.2378", "0.3435", "0.3145", "0.3435", "0.3145", "0.361", "0.2378", "0.361" ], [ "0.5024", "0.3467", "0.8545", "0.3467", "0.8545", "0.3599", "0.5024", "0.3599" ], [ "0.10046", "0.379", "0.468", "0.379", "0.468", "0.3926", "0.10046", "0.3926" ], [ "0.0809", "0.3945", "0.4673", "0.3945", "0.4673", "0.4077", "0.0809", "0.4077" ], [ "0.5034", "0.3894", "0.6396", "0.3894", "0.6396", "0.404", "0.5034", "0.404" ], [ "0.0784", "0.4092", "0.4697", "0.4077", "0.4697", "0.4224", "0.0784", "0.4236" ], [ "0.523", "0.4153", "0.8906", "0.4153", "0.8906", "0.4287", "0.523", "0.4287" ], [ "0.0809", "0.4248", "0.468", "0.4248", "0.468", "0.4382", "0.0809", "0.4382" ], [ "0.505", "0.43", "0.8906", "0.43", "0.8906", "0.443", "0.505", "0.443" ], [ "0.0801", "0.44", "0.468", "0.44", "0.468", "0.4539", "0.0801", "0.4539" ], [ "0.5034", "0.4458", "0.8906", "0.4458", "0.8906", "0.459", "0.5034", "0.459" ], [ "0.0801", "0.455", "0.468", "0.455", "0.468", "0.4685", "0.0801", "0.4685" ], [ "0.505", "0.461", "0.8896", "0.461", "0.8896", "0.4749", "0.505", "0.4749" ], [ "0.0801", "0.4702", "0.4656", "0.4702", "0.4656", "0.4841", "0.0801", "0.4841" ], [ "0.505", "0.476", "0.8037", "0.476", "0.8037", "0.4893", "0.505", "0.4893" ], [ "0.818", "0.4766", "0.892", "0.4749", "0.892", "0.4893", "0.818", "0.4912" ], [ "0.0809", "0.4849", "0.468", "0.4849", "0.468", "0.4988", "0.0809", "0.4988" ], [ "0.5034", "0.4893", "0.8906", "0.4905", "0.8906", "0.505", "0.5034", "0.504" ], [ "0.0801", "0.5005", "0.468", "0.5005", "0.468", "0.5146", "0.0801", "0.5146" ], [ "0.505", "0.506", "0.8906", "0.506", "0.8906", "0.5195", "0.505", "0.5195" ], [ "0.0809", "0.5166", "0.4673", "0.5166", "0.4673", "0.53", "0.0809", "0.53" ], [ "0.505", "0.5215", "0.636", "0.5215", "0.636", "0.535", "0.505", "0.535" ], [ "0.0784", "0.53", "0.468", "0.529", "0.468", "0.5435", "0.0784", "0.5444" ], [ "0.524", "0.535", "0.8906", "0.535", "0.8906", "0.549", "0.524", "0.549" ], [ "0.0809", "0.5454", "0.468", "0.5454", "0.468", "0.5596", "0.0809", "0.5596" ], [ "0.505", "0.5503", "0.8906", "0.5503", "0.8906", "0.5645", "0.505", "0.5645" ], [ "0.0809", "0.56", "0.468", "0.56", "0.468", "0.5737", "0.0809", "0.5737" ], [ "0.5034", "0.5645", "0.888", "0.5664", "0.888", "0.5806", "0.5034", "0.579" ], [ "0.0801", "0.5757", "0.468", "0.5757", "0.468", "0.589", "0.0801", "0.589" ], [ "0.505", "0.5806", "0.8896", "0.5806", "0.8896", "0.5947", "0.505", "0.5947" ], [ "0.0801", "0.5903", "0.468", "0.5903", "0.468", "0.604", "0.0801", "0.604" ], [ "0.5034", "0.5967", "0.8906", "0.5967", "0.8906", "0.61", "0.5034", "0.61" ], [ "0.0801", "0.605", "0.468", "0.605", "0.468", "0.6187", "0.0801", "0.6187" ], [ "0.5034", "0.6113", "0.8906", "0.6113", "0.8906", "0.625", "0.5034", "0.625" ], [ "0.0809", "0.6226", "0.4673", "0.6226", "0.4673", "0.6333", "0.0809", "0.6333" ], [ "0.505", "0.627", "0.8906", "0.627", "0.8906", "0.64", "0.505", "0.64" ], [ "0.0809", "0.636", "0.4697", "0.636", "0.4697", "0.6494", "0.0809", "0.6494" ], [ "0.505", "0.6426", "0.8896", "0.6426", "0.8896", "0.656", "0.505", "0.656" ], [ "0.0801", "0.6514", "0.4673", "0.6514", "0.4673", "0.6655", "0.0801", "0.6655" ], [ "0.5034", "0.657", "0.8896", "0.657", "0.8896", "0.6704", "0.5034", "0.6704" ], [ "0.0801", "0.6655", "0.468", "0.6665", "0.468", "0.681", "0.0801", "0.6797" ], [ "0.505", "0.6724", "0.892", "0.6724", "0.892", "0.686", "0.505", "0.686" ], [ "0.0801", "0.6797", "0.4673", "0.681", "0.4673", "0.696", "0.0801", "0.6943" ], [ "0.505", "0.6895", "0.7646", "0.6895", "0.7646", "0.6997", "0.505", "0.6997" ], [ "0.0825", "0.6978", "0.4656", "0.6978", "0.4656", "0.7085", "0.0825", "0.7085" ], [ "0.0801", "0.711", "0.4697", "0.711", "0.4697", "0.7246", "0.0801", "0.7246" ], [ "0.527", "0.7085", "0.8813", "0.7085", "0.8813", "0.7217", "0.527", "0.7217" ], [ "0.0809", "0.726", "0.468", "0.726", "0.468", "0.7393", "0.0809", "0.7393" ], [ "0.5186", "0.731", "0.892", "0.731", "0.892", "0.744", "0.5186", "0.744" ], [ "0.0801", "0.741", "0.468", "0.741", "0.468", "0.755", "0.0801", "0.755" ], [ "0.535", "0.747", "0.892", "0.747", "0.892", "0.7603", "0.535", "0.7603" ], [ "0.0809", "0.7583", "0.4673", "0.7583", "0.4673", "0.768", "0.0809", "0.768" ], [ "0.536", "0.761", "0.892", "0.761", "0.892", "0.7744", "0.536", "0.7744" ], [ "0.0801", "0.7715", "0.4697", "0.7715", "0.4697", "0.786", "0.0801", "0.786" ], [ "0.536", "0.7764", "0.8906", "0.7764", "0.8906", "0.7905", "0.536", "0.7905" ], [ "0.0801", "0.787", "0.468", "0.787", "0.468", "0.8003", "0.0801", "0.8003" ], [ "0.535", "0.7915", "0.892", "0.7915", "0.892", "0.806", "0.535", "0.806" ], [ "0.0809", "0.8037", "0.4673", "0.8037", "0.4673", "0.8145", "0.0809", "0.8145" ], [ "0.5376", "0.808", "0.8896", "0.808", "0.8896", "0.8184", "0.5376", "0.8184" ], [ "0.0801", "0.8174", "0.4656", "0.8174", "0.4656", "0.8306", "0.0801", "0.8306" ], [ "0.5337", "0.8213", "0.892", "0.8213", "0.892", "0.8354", "0.5337", "0.8354" ], [ "0.0801", "0.8335", "0.468", "0.8335", "0.468", "0.8467", "0.0801", "0.8467" ], [ "0.535", "0.837", "0.892", "0.837", "0.892", "0.851", "0.535", "0.851" ], [ "0.5337", "0.8516", "0.808", "0.8516", "0.808", "0.8657", "0.5337", "0.8657" ], [ "0.09644", "0.8643", "0.2181", "0.8643", "0.2181", "0.875", "0.09644", "0.875" ], [ "0.0972", "0.8755", "0.468", "0.8755", "0.468", "0.8896", "0.0972", "0.8896" ], [ "0.517", "0.872", "0.892", "0.872", "0.892", "0.8853", "0.517", "0.8853" ], [ "0.0784", "0.886", "0.4617", "0.887", "0.4617", "0.901", "0.0784", "0.8994" ], [ "0.5337", "0.8853", "0.892", "0.886", "0.892", "0.901", "0.5337", "0.8994" ], [ "0.482", "0.928", "0.492", "0.928", "0.492", "0.935", "0.482", "0.935" ] ]

[ "Factorized Higher-Order CNNs", "with an Application to Spatio-Temporal Emotion Estimation", "Jean Kossaifi* t", "Antoine Toisoul*", "Adrian Bulat", "Yannis Panagakist", "NVIDIA", "Samsung AI Center", "Samsung AI Center", "University of Athens", "Timothy Hospedales", "Maja Pantict", "Samsung AI Center", "Imperial College London", "Abstract", "tion datasets, AffectNet, SEWA, and AFEW-VA.", "Training deep neural networks with spatio-temporal", "(i.e., 3D) or multidimensional convolutions of higher-", "1.Introduction", "order is computationally challenging due to millions of", "随着深度卷积神经网络的空前成功，人们一直在寻", "unknown parameters across dozens of layers. To al-", "求训练更深层次的网络。然而，虽然更深的神经网络在", "leviate this, one approach is to apply low-rank tensor", "经过适当训练后可以提供更好的性能，但这种深度也", "decompositions to convolution kernels in order to com-", "会转化为内存和计算量大的模型，通常具有数千万个", "press the network and reduce its number of parame-", "参数。在训练训练高阶卷积网时尤其如此", "e。g。视", "ters. Alternatively, new convolutional blocks, such as", "频上的三阶（3D)。但是，此类模型对于在时空域中执", "MobileNet, can be directly designed for efficiency. In", "行预测是必需的，并且在许多应用中至关重要，包括动", "this paper, we unify these two approaches by propos-", "作识别和情感识别", "ing a tensor factorization framework for efficient mul-", "在本文中，我们与传统方法不同，提出了一种新颖", "tidimensional (separable) convolutions of higher-order.", "的因式分解多维卷积块，该卷积块通过有效使用数据", "Interestingly, the proposed framework enables a novel", "中的结构来实现卓越的性能。此外，我们的模型可以首", "higher-order transduction, allowing to train a network", "先在图像域上进行训练，并无缝扩展以将性能转移到", "on a given domain (e.g., 2D images or N-dimensional", "时域。通过我们提出的块的结构，这种新颖的转导方法", "data in general) and using transduction to general-", "成为可能。此外，它允许大幅减少参数的数量，同时提", "ize tohigher-orderdata such as videos(or(N+K)", "高性能和计算效率，它可以应用于现有的时空网络架", "dimensional data in general), capturing for instance", "构，如ResNet3D【47]。我们的方法利用CP张量分", "temporal dynamics while preserving the learnt spa-", "解，以便分别学习3D卷积的解开的空间和时间信息。", "tial information. We apply the proposed methodology,", "这大大提高了准确性，同时减少了时空架构的参数数", "coined CP-Higher-Order Convolution (HO-CPConv ),", "量，并极大地促进了视频数据的训练", "to spatio-temporalfacial emotion analysis.Most ex-", "isting facial affect models focus on static imagery and", "In summary, we make the following contributions:", "discard all temporal information. This is due to the", ". We show that many of deep nets architectural", "above-mentioned burden of training 3D convolutional", "improvements, such as MobileNet or ResNet 's", "nets and the lack of large bodies of video data anno-", "Bottleneck blocks, are in fact drawn from the", "tated by experts. We address both issues with our", "same larger family of tensor decomposition meth-", "proposed framework. Initial training is first done on", "ods (Section 3 ). We propose a general framework", "static imagery beforeusing transductiontogeneralize", "unifvingtensordecompositionsandefficientarchi-", "to the temporal domain. We demonstrate superior per-", "tectures, showing how these efficient architectures", "formance on three challenging large scale affect estima-", "can be derived by applying tensor decomposition", "to convolutional kernels (Section 3.4 )", "*Joint first authors.", "f Jean Kossaifi, Yannis Panagakis and Maja Pantic were with", ". Using this framework, we propose factorized", "Samsung AI Center, Cambridge and Imperial College London.", "higher-order convolutional neural networks, that", "1" ]

{ "arxiv_id": "1906.06196", "language": "zh", "page_idx": 0 }

[ [ "0.1316", "0.09216", "0.2769", "0.09216", "0.2769", "0.1029", "0.1316", "0.1029" ], [ "0.536", "0.09344", "0.5923", "0.09344", "0.5923", "0.101", "0.536", "0.101" ], [ "0.6006", "0.09155", "0.728", "0.09155", "0.728", "0.1016", "0.6006", "0.1016" ], [ "0.2793", "0.11426", "0.2998", "0.11426", "0.2998", "0.1218", "0.2793", "0.1218" ], [ "0.826", "0.11426", "0.8555", "0.11426", "0.8555", "0.1218", "0.826", "0.1218" ], [ "0.273", "0.1206", "0.312", "0.1206", "0.312", "0.1307", "0.273", "0.1307" ], [ "0.8228", "0.1206", "0.863", "0.1206", "0.863", "0.1307", "0.8228", "0.1307" ], [ "0.0809", "0.2507", "0.8896", "0.2507", "0.8896", "0.2651", "0.0809", "0.2651" ], [ "0.0809", "0.2664", "0.8906", "0.2664", "0.8906", "0.2798", "0.0809", "0.2798" ], [ "0.0809", "0.2798", "0.8906", "0.2798", "0.8906", "0.2935", "0.0809", "0.2935" ], [ "0.0801", "0.2954", "0.8906", "0.2966", "0.8906", "0.3113", "0.0801", "0.31" ], [ "0.0809", "0.3132", "0.1365", "0.3132", "0.1365", "0.3245", "0.0809", "0.3245" ], [ "0.1136", "0.3542", "0.464", "0.3542", "0.464", "0.3674", "0.1136", "0.3674" ], [ "0.505", "0.3542", "0.8906", "0.3542", "0.8906", "0.3674", "0.505", "0.3674" ], [ "0.1111", "0.3674", "0.4673", "0.3674", "0.4673", "0.3806", "0.1111", "0.3806" ], [ "0.5034", "0.3687", "0.8906", "0.3687", "0.8906", "0.3818", "0.5034", "0.3818" ], [ "0.1136", "0.3838", "0.4673", "0.3838", "0.4673", "0.3977", "0.1136", "0.3977" ], [ "0.5034", "0.3838", "0.7183", "0.3838", "0.7183", "0.3984", "0.5034", "0.3984" ], [ "0.11273", "0.3997", "0.267", "0.3997", "0.267", "0.4136", "0.11273", "0.4136" ], [ "0.5293", "0.406", "0.8906", "0.4072", "0.8906", "0.4216", "0.5293", "0.4204" ], [ "0.09314", "0.4185", "0.468", "0.4192", "0.468", "0.4343", "0.09314", "0.433" ], [ "0.8447", "0.4136", "0.8545", "0.4136", "0.8545", "0.4172", "0.8447", "0.4172" ], [ "0.505", "0.4224", "0.8906", "0.4224", "0.8906", "0.4363", "0.505", "0.4363" ], [ "0.1111", "0.435", "0.468", "0.435", "0.468", "0.449", "0.1111", "0.449" ], [ "0.5024", "0.4375", "0.8906", "0.4375", "0.8906", "0.4507", "0.5024", "0.4507" ], [ "0.1136", "0.452", "0.4673", "0.452", "0.4673", "0.4622", "0.1136", "0.4622" ], [ "0.505", "0.4539", "0.8896", "0.4539", "0.8896", "0.464", "0.505", "0.464" ], [ "0.1111", "0.4634", "0.2507", "0.4653", "0.2507", "0.4792", "0.1111", "0.4766" ], [ "0.5063", "0.4685", "0.8906", "0.4685", "0.8906", "0.4817", "0.5063", "0.4817" ], [ "0.0948", "0.486", "0.4673", "0.486", "0.4673", "0.5", "0.0948", "0.5" ], [ "0.505", "0.483", "0.8906", "0.483", "0.8906", "0.4968", "0.505", "0.4968" ], [ "0.1111", "0.5005", "0.4656", "0.5005", "0.4656", "0.5146", "0.1111", "0.5146" ], [ "0.505", "0.4976", "0.8906", "0.4976", "0.8906", "0.511", "0.505", "0.511" ], [ "0.1111", "0.5166", "0.468", "0.515", "0.468", "0.529", "0.1111", "0.53" ], [ "0.505", "0.512", "0.8906", "0.512", "0.8906", "0.526", "0.505", "0.526" ], [ "0.1136", "0.5327", "0.1772", "0.5327", "0.1772", "0.5444", "0.1136", "0.5444" ], [ "0.505", "0.528", "0.8906", "0.528", "0.8906", "0.5415", "0.505", "0.5415" ], [ "0.505", "0.5415", "0.8906", "0.5425", "0.8906", "0.5566", "0.505", "0.556" ], [ "0.0948", "0.553", "0.4673", "0.553", "0.4673", "0.5664", "0.0948", "0.5664" ], [ "0.505", "0.558", "0.892", "0.558", "0.892", "0.5713", "0.505", "0.5713" ], [ "0.1111", "0.5664", "0.468", "0.5674", "0.468", "0.5806", "0.1111", "0.58" ], [ "0.5034", "0.5737", "0.8447", "0.5737", "0.8447", "0.584", "0.5034", "0.584" ], [ "0.1152", "0.5835", "0.4656", "0.5835", "0.4656", "0.5933", "0.1152", "0.5933" ], [ "0.523", "0.5903", "0.8906", "0.5903", "0.8906", "0.604", "0.523", "0.604" ], [ "0.1111", "0.5967", "0.2917", "0.5967", "0.2917", "0.61", "0.1111", "0.61" ], [ "0.5034", "0.604", "0.892", "0.605", "0.892", "0.619", "0.5034", "0.6187" ], [ "0.505", "0.6206", "0.8906", "0.6206", "0.8906", "0.634", "0.505", "0.634" ], [ "0.0801", "0.627", "0.3586", "0.627", "0.3586", "0.6416", "0.0801", "0.6416" ], [ "0.509", "0.637", "0.8906", "0.637", "0.8906", "0.648", "0.509", "0.648" ], [ "0.505", "0.651", "0.8906", "0.651", "0.8906", "0.664", "0.505", "0.664" ], [ "0.1054", "0.66", "0.4673", "0.66", "0.4673", "0.674", "0.1054", "0.674" ], [ "0.505", "0.6655", "0.8906", "0.6655", "0.8906", "0.6787", "0.505", "0.6787" ], [ "0.0825", "0.6753", "0.4673", "0.6753", "0.4673", "0.6895", "0.0825", "0.6895" ], [ "0.505", "0.6797", "0.8906", "0.6797", "0.8906", "0.6943", "0.505", "0.6943" ], [ "0.0809", "0.69", "0.468", "0.69", "0.468", "0.704", "0.0809", "0.704" ], [ "0.5024", "0.6943", "0.6177", "0.6943", "0.6177", "0.7085", "0.5024", "0.7085" ], [ "0.0801", "0.704", "0.4673", "0.705", "0.4673", "0.7197", "0.0801", "0.7183" ], [ "0.524", "0.712", "0.8906", "0.712", "0.8906", "0.726", "0.524", "0.726" ], [ "0.0809", "0.72", "0.468", "0.72", "0.468", "0.7344", "0.0809", "0.7344" ], [ "0.505", "0.728", "0.8906", "0.728", "0.8906", "0.741", "0.505", "0.741" ], [ "0.0825", "0.735", "0.468", "0.735", "0.468", "0.7485", "0.0825", "0.7485" ], [ "0.505", "0.7437", "0.8906", "0.7437", "0.8906", "0.757", "0.505", "0.757" ], [ "0.08496", "0.7505", "0.468", "0.7505", "0.468", "0.7646", "0.08496", "0.7646" ], [ "0.505", "0.7583", "0.8906", "0.7583", "0.8906", "0.7715", "0.505", "0.7715" ], [ "0.0825", "0.765", "0.468", "0.765", "0.468", "0.779", "0.0825", "0.779" ], [ "0.5024", "0.773", "0.8906", "0.7734", "0.8906", "0.787", "0.5024", "0.786" ], [ "0.0809", "0.781", "0.468", "0.781", "0.468", "0.7944", "0.0809", "0.7944" ], [ "0.5034", "0.7886", "0.8906", "0.7886", "0.8906", "0.803", "0.5034", "0.803" ], [ "0.0809", "0.7954", "0.468", "0.7954", "0.468", "0.8096", "0.0809", "0.8096" ], [ "0.5034", "0.803", "0.8906", "0.803", "0.8906", "0.8174", "0.5034", "0.8174" ], [ "0.0809", "0.811", "0.4656", "0.811", "0.4656", "0.8247", "0.0809", "0.8247" ], [ "0.5024", "0.8164", "0.8896", "0.8174", "0.8896", "0.832", "0.5024", "0.8306" ], [ "0.0801", "0.8257", "0.468", "0.8247", "0.468", "0.839", "0.0801", "0.8403" ], [ "0.505", "0.8335", "0.8906", "0.8335", "0.8906", "0.8467", "0.505", "0.8467" ], [ "0.0809", "0.8413", "0.433", "0.8413", "0.433", "0.855", "0.0809", "0.855" ], [ "0.505", "0.8486", "0.888", "0.8486", "0.888", "0.8623", "0.505", "0.8623" ], [ "0.1013", "0.857", "0.4673", "0.857", "0.4673", "0.8706", "0.1013", "0.8706" ], [ "0.5034", "0.8623", "0.751", "0.863", "0.751", "0.8774", "0.5034", "0.877" ], [ "0.0825", "0.872", "0.468", "0.872", "0.468", "0.8853", "0.0825", "0.8853" ], [ "0.0809", "0.887", "0.468", "0.887", "0.468", "0.901", "0.0809", "0.901" ], [ "0.528", "0.8853", "0.8906", "0.886", "0.8906", "0.901", "0.528", "0.8994" ], [ "0.478", "0.9224", "0.4944", "0.9224", "0.4944", "0.9395", "0.478", "0.9395" ] ]

[ "Static Separable (2D)Convolutior", "Spatio-Tempo", "al (3D)SeparableConvolution", "Stati", "Tempor", "Prediction", "Prediction", "Figure 1: Overview of our method , here represented for a single channel of a single input. We start by training a", "2D CNN with our proposed factorized convolutional block on static images (left) : We then apply transduction to", "extend the model from the static to the spatio-temporal domain (right) : The pretrained spatial factors (blue and", "red) are first kept fixed, before jointly fine-tuning all the parameters once the temporal factors (green) have been", "trained.", "leverage efficient general multi-dimensional convo", "用张量方法进行有效卷积的方法，可以压缩或重新制", "lutions. These achieve the same performance with", "定它们的速度。其次，直接制定有效神经架构的方法，", "a fraction of the parameters and floating point op-", "例如。g。，使用可分离的卷积", "erations (Section 4).", "Tensormethodsfor efficientdeep networks张量", ".Finally,we propose a novel mechanism called", "", "方法的属性【14，41，17]使它们成为深度学习的首选。", "higher-order transduction which can be employed", "除了对深度神经网络E[3]的性质进行理论研究外，还", "to convert our model,trained in N dimensions,to", "特别在重新参数化现有层F49]的背景下对它们进行", "N+Kdimensions.", "了研究。。这种重新参数化的一个目标是节省参数空", ":We show that our factorized higher-order networks", "间【7]。例如，[31]建议将全连接层的权重矩阵重塑为", "outperform existing works on static affect esti-", "具有张量-火车（TT)E32结构的高阶张量。在后续", "mation onthe AffectNet, SEWAand AFEW-VA", "工作E[4]中，相同的策略也应用于卷积层。完全连接", "datasets.", "的层和平坦化层可以一起删除，并替换为张量回归层E", "[21]。这些通过从高阶激活张量到任意阶输出张量的", ": Using transduction on the static models, we also", "低秩多线性映射来表示输出。通过应用张量收缩[201，", "demonstrate state-of-the-art results for continuous", "在保持多线性结构的同时，也可以节省参数空间", "facialemotion analvsisfromvideoonbothSEWA", "张量重新参数化的另一个优点是计算速度加快。特", "and AFEW-VA datasets.", "别是，张量分解是从卷积核获得可分离滤波器的有效", "方法。这些可分离的卷积是由E[36]在滤波器组的背景", "2. Background and related work", "下在计算机视觉中提出的。121首先将此概念应用于深", "度学习，并建议使用可分离卷积来利用跨通道的穴余。", "Multidimensionalconvolutions出现在不同领域的", "[1，27]建议将CP分解直接应用于预训练的2D卷积", "几个数学模型中。它们是卷积神经网络（CNN）[2628", "层的（4维）内核，使它们可分离。通过微调来补偿所", "的基石，使它们能够通过减轻维度的诅咒[34]来有效", "产生的性能损失", "地从高维数据中学习。然而，CNN在计算上要求很高，", "使用Tucker分解代替CP分解预训练网络的卷积层", "卷积的成本在训练和推理过程中都占主导地位。因此，", "也可以获得有效的卷积重写E[15】。这允许将卷积重写", "人们对提高多维卷积的效率越来越感兴趣。已经提出", "为1×1卷积，然后使用较小的内核和另一个1×1卷", "了几种有效的卷积实现。例如，通过将卷积核转换为", "积进行常规卷积。在这种情况下，卷积核的空间维度保", "Toeplitz矩阵，可以将2D卷积有效地实现为矩阵乘", "持不变，只有输入和输出通道被压缩。同样，通过微调", "法。但是，此过程需要跨Toeplitz矩阵中的不同矩阵", "整个网络来补偿性能损失。最后，【46]建议删除卷积层", "列多次复制内核值，从而增加了内存需求。由于构建列", "中的几余，并将其表示为具有较少参数的两个卷积层", "矩阵所需的空间，通过im2col方法实现卷积也是内存", "的组成。每个2D滤波器由秩的总和近似-1矩阵。由", "密集型的。对于移动或嵌人式设备，这些内存要求可能", "于这种受限的设置，可以使用SVD轻松获得封闭形式", "过高，从而阻碍了在资源有限的平台中部署CNN", "的解决方案。在这里，我们统一了上述工作，并提出了", "通常，大多数现有的有效卷积尝试都是孤立的，目", "因式分解的高阶（可分离的）卷积", "前没有统一的框架来研究它们。特别是我们对两个不", "同的工作分支感兴趣，我们将在下一步回顾。首先，利", "Eficientneuralnetworks虽然自从使用张量分解", "2" ]

{ "arxiv_id": "1906.06196", "language": "zh", "page_idx": 1 }

[ [ "0.5034", "0.09155", "0.886", "0.09155", "0.886", "0.1048", "0.5034", "0.1048" ], [ "0.524", "0.1187", "0.888", "0.1187", "0.888", "0.132", "0.524", "0.132" ], [ "0.5034", "0.132", "0.8906", "0.1332", "0.8906", "0.1477", "0.5034", "0.1465" ], [ "0.09235", "0.1477", "0.4534", "0.1477", "0.4534", "0.161", "0.09235", "0.161" ], [ "0.505", "0.1509", "0.8896", "0.1509", "0.8896", "0.161", "0.505", "0.161" ], [ "0.505", "0.1642", "0.8896", "0.1642", "0.8896", "0.178", "0.505", "0.178" ], [ "0.505", "0.1787", "0.8896", "0.1787", "0.8896", "0.1925", "0.505", "0.1925" ], [ "0.0809", "0.1895", "0.2754", "0.1895", "0.2754", "0.2026", "0.0809", "0.2026" ], [ "0.5063", "0.1932", "0.8896", "0.1932", "0.8896", "0.207", "0.5063", "0.207" ], [ "0.1013", "0.2108", "0.468", "0.2108", "0.468", "0.226", "0.1013", "0.226" ], [ "0.505", "0.209", "0.8906", "0.209", "0.8906", "0.2229", "0.505", "0.2229" ], [ "0.0825", "0.2291", "0.4656", "0.2291", "0.4656", "0.2393", "0.0825", "0.2393" ], [ "0.505", "0.226", "0.568", "0.226", "0.568", "0.2374", "0.505", "0.2374" ], [ "0.0809", "0.2424", "0.468", "0.2424", "0.468", "0.2556", "0.0809", "0.2556" ], [ "0.527", "0.2468", "0.8906", "0.2474", "0.8906", "0.262", "0.527", "0.2615" ], [ "0.0809", "0.2568", "0.3145", "0.2568", "0.3145", "0.2708", "0.0809", "0.2708" ], [ "0.505", "0.2632", "0.8906", "0.2632", "0.8906", "0.2766", "0.505", "0.2766" ], [ "0.1666", "0.2827", "0.282", "0.2827", "0.282", "0.2974", "0.1666", "0.2974" ], [ "0.505", "0.279", "0.8906", "0.279", "0.8906", "0.2922", "0.505", "0.2922" ], [ "0.2361", "0.2922", "0.2524", "0.2974", "0.2467", "0.307", "0.2303", "0.3018" ], [ "0.255", "0.2917", "0.4673", "0.2886", "0.4673", "0.3093", "0.255", "0.3125" ], [ "0.5024", "0.2935", "0.8896", "0.2935", "0.8896", "0.307", "0.5024", "0.307" ], [ "0.1993", "0.2942", "0.2197", "0.2986", "0.2157", "0.3113", "0.1953", "0.307" ], [ "0.0809", "0.2954", "0.1708", "0.2954", "0.1708", "0.307", "0.0809", "0.307" ], [ "0.1683", "0.2974", "0.183", "0.2922", "0.1887", "0.3018", "0.1732", "0.307" ], [ "0.2157", "0.303", "0.2279", "0.303", "0.2279", "0.307", "0.2157", "0.307" ], [ "0.1683", "0.3093", "0.282", "0.3093", "0.282", "0.3193", "0.1683", "0.3193" ], [ "0.505", "0.308", "0.8896", "0.308", "0.8896", "0.3225", "0.505", "0.3225" ], [ "0.505", "0.324", "0.888", "0.324", "0.888", "0.3384", "0.505", "0.3384" ], [ "0.10046", "0.332", "0.255", "0.332", "0.255", "0.3452", "0.10046", "0.3452" ], [ "0.1136", "0.3528", "0.2793", "0.3528", "0.2793", "0.385", "0.1136", "0.385" ], [ "0.5293", "0.3484", "0.8906", "0.3484", "0.8906", "0.3623", "0.5293", "0.3623" ], [ "0.1136", "0.363", "0.2769", "0.361", "0.2769", "0.3894", "0.1136", "0.3914" ], [ "0.505", "0.3643", "0.8906", "0.3643", "0.8906", "0.3774", "0.505", "0.3774" ], [ "0.2507", "0.3662", "0.4631", "0.3655", "0.4631", "0.377", "0.2507", "0.3774" ], [ "0.0809", "0.3655", "0.1209", "0.3674", "0.11926", "0.379", "0.0801", "0.377" ], [ "0.1871", "0.3706", "0.2059", "0.3706", "0.2059", "0.375", "0.1871", "0.375" ], [ "0.2361", "0.3687", "0.2507", "0.3687", "0.2507", "0.3757", "0.2361", "0.3757" ], [ "0.5034", "0.377", "0.8906", "0.3774", "0.8906", "0.3926", "0.5034", "0.3914" ], [ "0.10046", "0.3984", "0.468", "0.3984", "0.468", "0.4167", "0.10046", "0.4167" ], [ "0.505", "0.3933", "0.8906", "0.3933", "0.8906", "0.4072", "0.505", "0.4072" ], [ "0.3896", "0.4124", "0.4043", "0.4124", "0.4043", "0.4167", "0.3896", "0.4167" ], [ "0.505", "0.4092", "0.8906", "0.4092", "0.8906", "0.4224", "0.505", "0.4224" ], [ "0.0784", "0.4153", "0.3447", "0.4136", "0.3447", "0.4312", "0.0784", "0.433" ], [ "0.5063", "0.4255", "0.8896", "0.4255", "0.8896", "0.4363", "0.5063", "0.4363" ], [ "0.10046", "0.433", "0.3694", "0.433", "0.3694", "0.4463", "0.10046", "0.4463" ], [ "0.505", "0.4395", "0.5376", "0.4395", "0.5376", "0.452", "0.505", "0.452" ], [ "0.1478", "0.4583", "0.2737", "0.4583", "0.2737", "0.4685", "0.1478", "0.4685" ], [ "0.134", "0.4666", "0.3594", "0.464", "0.3594", "0.486", "0.134", "0.488" ], [ "0.527", "0.4622", "0.8896", "0.4634", "0.8896", "0.4778", "0.527", "0.4766" ], [ "0.0809", "0.4697", "0.1405", "0.473", "0.1396", "0.4873", "0.0801", "0.4841" ], [ "0.3545", "0.4749", "0.3635", "0.4749", "0.3635", "0.4792", "0.3545", "0.4792" ], [ "0.4004", "0.4734", "0.464", "0.4734", "0.464", "0.4841", "0.4004", "0.4841" ], [ "0.5034", "0.4792", "0.8896", "0.4792", "0.8896", "0.4924", "0.5034", "0.4924" ], [ "0.1381", "0.486", "0.2737", "0.486", "0.2737", "0.4968", "0.1381", "0.4968" ], [ "0.4468", "0.4988", "0.468", "0.4988", "0.468", "0.512", "0.4468", "0.512" ], [ "0.505", "0.4944", "0.8896", "0.4944", "0.8896", "0.5083", "0.505", "0.5083" ], [ "0.505", "0.509", "0.8906", "0.509", "0.8906", "0.523", "0.505", "0.523" ], [ "0.10046", "0.5264", "0.468", "0.5264", "0.468", "0.5405", "0.10046", "0.5405" ], [ "0.5034", "0.523", "0.8896", "0.5234", "0.8896", "0.5386", "0.5034", "0.537" ], [ "0.0809", "0.5425", "0.468", "0.5425", "0.468", "0.556", "0.0809", "0.556" ], [ "0.5034", "0.539", "0.8906", "0.539", "0.8906", "0.553", "0.5034", "0.553" ], [ "0.5024", "0.5503", "0.892", "0.5503", "0.892", "0.5684", "0.5024", "0.5684" ], [ "0.0809", "0.5566", "0.2917", "0.5566", "0.2917", "0.571", "0.0809", "0.571" ], [ "0.501", "0.563", "0.892", "0.563", "0.892", "0.584", "0.501", "0.584" ], [ "0.4983", "0.5776", "0.8936", "0.5776", "0.8936", "0.6016", "0.4983", "0.6016" ], [ "0.5186", "0.5933", "0.5474", "0.5933", "0.5474", "0.601", "0.5186", "0.601" ], [ "0.569", "0.5947", "0.589", "0.5947", "0.589", "0.5986", "0.569", "0.5986" ], [ "0.601", "0.5947", "0.6235", "0.5947", "0.6235", "0.5977", "0.601", "0.5977" ], [ "0.5", "0.6016", "0.8945", "0.603", "0.8945", "0.6235", "0.5", "0.6226" ], [ "0.1462", "0.616", "0.1805", "0.616", "0.1805", "0.634", "0.1462", "0.634" ], [ "0.2255", "0.613", "0.25", "0.613", "0.25", "0.6206", "0.2255", "0.6206" ], [ "0.2157", "0.619", "0.2222", "0.619", "0.2222", "0.6226", "0.2157", "0.6226" ], [ "0.2484", "0.619", "0.3022", "0.622", "0.3022", "0.632", "0.2484", "0.63" ], [ "0.3137", "0.617", "0.451", "0.617", "0.451", "0.632", "0.3137", "0.632" ], [ "0.1953", "0.6235", "0.2035", "0.6235", "0.2035", "0.629", "0.1953", "0.629" ], [ "0.2239", "0.6235", "0.255", "0.6206", "0.2566", "0.629", "0.2255", "0.632" ], [ "0.5034", "0.6235", "0.5376", "0.6235", "0.5376", "0.636", "0.5034", "0.636" ], [ "0.2091", "0.6255", "0.2173", "0.6226", "0.2197", "0.6255", "0.2115", "0.629" ], [ "0.0809", "0.6206", "0.1209", "0.6206", "0.1209", "0.631", "0.0809", "0.631" ], [ "0.1283", "0.63", "0.152", "0.63", "0.152", "0.6416", "0.1283", "0.6416" ], [ "0.1929", "0.6333", "0.2524", "0.6333", "0.2524", "0.64", "0.1929", "0.64" ], [ "0.3145", "0.632", "0.33", "0.632", "0.33", "0.6416", "0.3145", "0.6416" ], [ "0.361", "0.6445", "0.406", "0.6445", "0.406", "0.656", "0.361", "0.656" ], [ "0.3015", "0.66", "0.3464", "0.66", "0.3464", "0.6675", "0.3015", "0.6675" ], [ "0.2778", "0.674", "0.3162", "0.674", "0.3162", "0.681", "0.2778", "0.681" ], [ "0.505", "0.7373", "0.8896", "0.7373", "0.8896", "0.7476", "0.505", "0.7476" ], [ "0.5024", "0.75", "0.8896", "0.75", "0.8896", "0.763", "0.5024", "0.763" ], [ "0.5024", "0.7666", "0.7715", "0.765", "0.7715", "0.779", "0.5024", "0.78" ], [ "0.0809", "0.778", "0.468", "0.778", "0.468", "0.7925", "0.0809", "0.7925" ], [ "0.0801", "0.7925", "0.468", "0.7915", "0.468", "0.806", "0.0801", "0.8066" ], [ "0.0801", "0.808", "0.468", "0.808", "0.468", "0.8213", "0.0801", "0.8213" ], [ "0.5254", "0.8096", "0.888", "0.81", "0.888", "0.8247", "0.5254", "0.8237" ], [ "0.0809", "0.8237", "0.1283", "0.8237", "0.1283", "0.8354", "0.0809", "0.8354" ], [ "0.5034", "0.8228", "0.892", "0.8228", "0.892", "0.8403", "0.5034", "0.8403" ], [ "0.505", "0.8413", "0.8906", "0.8413", "0.8906", "0.855", "0.505", "0.855" ], [ "0.1013", "0.855", "0.4631", "0.855", "0.4631", "0.8687", "0.1013", "0.8687" ], [ "0.505", "0.857", "0.8906", "0.857", "0.8906", "0.8706", "0.505", "0.8706" ], [ "0.505", "0.872", "0.8906", "0.872", "0.8906", "0.8853", "0.505", "0.8853" ], [ "0.1544", "0.881", "0.3938", "0.879", "0.3938", "0.901", "0.1544", "0.903" ], [ "0.4468", "0.886", "0.468", "0.886", "0.468", "0.8984", "0.4468", "0.8984" ], [ "0.5034", "0.887", "0.8906", "0.887", "0.8906", "0.901", "0.5034", "0.901" ], [ "0.478", "0.924", "0.4944", "0.924", "0.4944", "0.938", "0.478", "0.938" ] ]

[ "3.4.Efficient architectures in a tensor framework", "虽然张量分解已经在数学领域进行了数十年的探索", "并且在深度学习的背景下进行了多年的探索，但它们", "Figure 2: Illustration of a 2D Kruskal convolution.", "经常以不同的形式被重新发现和重新引入。在这单，我", "们在张量分解的视角下重新审视流行的深度神经网络", "架构。具体来说，我们展示了如何通过将张量分解应用", "3.3. Tucker convolutions", "于其内核来从常规卷积中获得这些块。在实践中，批量", "如前所述，我们考虑卷积F=★W。然而，代替", "归一化层和非线性被插人中间卷积之间，以促进从头", "Kruskal结构，我们现在假设内核W上的低秩Tucker", "开始学习", "结构（可以通过应用Tucker分解容易地获得）并产生", "ResNetBottleneckblock[9]在他们关于深度残差", "有效的公式E[15]。我们可以写：", "网络的开创性工作中引人了一个块，创造的瓶颈块。它", "Ro-1 R1-1 Ra-1 Ra-", "包括一系列1×1卷积，以减少通道数，较小的常规", "", "Z gro,rrara u(T)u(CuH)uw)", "（3×3）卷积，以及另一个1×1卷积，以将秩恢复到", "", "W(t, s,j.i) =", "T", "", "ro=0 1=0 r2=0r3=0", "所需的输出通道数。基于3.3节中得出的等价性，可", "以直接将其视为将Tucker分解应用于常规卷积的内核", "插回卷积，我们得到：", "2LE222E", "ResNextandXceptionResNextE[48构建在此", "2222222", "瓶颈架构上，正如我们已经展示的那样，这等效于将", "Gro.rUST)u)UHuwXitut", "Fr.u.r =", "", "", "Tucker分解应用于卷积内核。为了进一步降低秩，输", "我们可以通过编写H=×2U(H)×U(W).来进一", "出被表示为这样的瓶颈的总和，具有较低的秩。这可以", "ro", "使用分组卷积48]有效地重新表述。同时，[2]提出", "步将沿空间维度的因素吸收到核心", "了一种类似的方法，但在分组深度卷积之后没有1×1", "在这种情况下，上面的表达式简化为：", "卷积", "CHWRo-1Ri-1", "ZEZZ Hro.n.U(T", "MobileNetv1MobileNetv1[10]使用由深度可分", "Fty,a=", "", "Xk,i+yi+a", "离卷积（卷积的空间部分）和1×1卷积组成的构建", "k=1j=ii=1 ro=0 r1=0", "(3)", "块来调整输出通道的数量。这可以很容易地从CP分", "解（第E3.2节）中获得，如下所示：首先，我们按照", "换句话说，这等效于首先转换通道数，然后在从秩", "公式E2中的详细说明编写卷积权重张量，其秩等于", "返回到目标通道数之前应用（小）卷积。这可以通过重", "输入通道的数量i。e。R=C。通过组合两个空间1", "D卷积U(H）和U(W），可以获得第一个深度可分离的", "新排列公式3中的项来看出：", "卷积。这导致单个空间因子U(S)ERHXWXR，使得", "u=uuw)，然后，1×1卷积由其余因子的矩", "i.i.r", "", "", "阵乘积U(F）=U(T）(U(C)TeRT×C 给出。。如图4", "UT)", "Ri-", "", "Hn1j.", "u)x(k,i+yi+a)", "", "", "所示", "", "Ftu.r", "ro=0", "11r=0", ":=1", "1x1conv", "Hxw.con", "1x1 cony", "Figure 4:MobileNet blocks are a special case of CP", "convolutions , without the first convolution, and with", "spatial factors are combined into one.", "Figure 3: Illustration of a Tucker convolution expressed", "as a series of small efficient convolutions. Note that this", "is the approach taken byResNet for theBottleneck", "MobileNetv2MobileNetv2E[39]采用类似的方法", "blocks.", "将空间因子分组为一个空间因子U(S)εRHxWxR，如", "前面针对MobileNet的情况所述。然而，其他因素没", "简而言之，这简化为以下表达式，也如图E3所示：", "有受到影响。在这种情况下，分解的等级对应于每个层", "的扩展因子×输人通道的数量。这导致两个1×1卷", "F= ((α ×o U(C)*9) ×o U(T)", "(4)", "积和一个3×3深度可分离卷积。最后，内核权重张量", "4" ]

{ "arxiv_id": "1906.06196", "language": "zh", "page_idx": 3 }

[ [ "0.3132", "0.0849", "0.69", "0.0849", "0.69", "0.0998", "0.3132", "0.0998" ], [ "0.2913", "0.11755", "0.721", "0.11755", "0.721", "0.1354", "0.2913", "0.1354" ], [ "0.2568", "0.1318", "0.7314", "0.1318", "0.7314", "0.1526", "0.2568", "0.1526" ], [ "0.723", "0.136", "0.742", "0.136", "0.742", "0.1443", "0.723", "0.1443" ], [ "0.3064", "0.1497", "0.7", "0.1526", "0.7", "0.1704", "0.3064", "0.1675" ], [ "0.1755", "0.1663", "0.8296", "0.1687", "0.8296", "0.187", "0.1755", "0.1853" ], [ "0.3284", "0.1818", "0.677", "0.1841", "0.677", "0.2048", "0.3284", "0.2031" ], [ "0.1948", "0.2031", "0.817", "0.2031", "0.817", "0.2203", "0.1948", "0.2203" ], [ "0.2007", "0.2197", "0.8076", "0.2214", "0.8076", "0.2363", "0.2007", "0.234" ], [ "0.2627", "0.2632", "0.3435", "0.2654", "0.3435", "0.2808", "0.2627", "0.279" ], [ "0.5146", "0.2654", "0.882", "0.2666", "0.882", "0.2808", "0.5146", "0.2798" ], [ "0.513", "0.2808", "0.881", "0.2808", "0.881", "0.2944", "0.513", "0.2944" ], [ "0.147", "0.2935", "0.461", "0.2935", "0.461", "0.3071", "0.147", "0.3071" ], [ "0.5146", "0.2964", "0.881", "0.2964", "0.881", "0.31", "0.5146", "0.31" ], [ "0.1486", "0.3088", "0.4575", "0.3088", "0.4575", "0.3196", "0.1486", "0.3196" ], [ "0.5146", "0.3142", "0.882", "0.3142", "0.882", "0.328", "0.5146", "0.328" ], [ "0.147", "0.3218", "0.461", "0.3218", "0.461", "0.335", "0.147", "0.335" ], [ "0.1453", "0.335", "0.4592", "0.3362", "0.4592", "0.3499", "0.1453", "0.3486" ], [ "0.5156", "0.33", "0.5728", "0.33", "0.5728", "0.3445", "0.5156", "0.3445" ], [ "0.147", "0.3499", "0.4592", "0.3499", "0.4592", "0.364", "0.147", "0.364" ], [ "0.534", "0.3457", "0.882", "0.3457", "0.882", "0.3599", "0.534", "0.3599" ], [ "0.147", "0.364", "0.461", "0.364", "0.461", "0.3777", "0.147", "0.3777" ], [ "0.5146", "0.3623", "0.8833", "0.3623", "0.8833", "0.3772", "0.5146", "0.3772" ], [ "0.147", "0.379", "0.461", "0.379", "0.461", "0.3926", "0.147", "0.3926" ], [ "0.5156", "0.3777", "0.8833", "0.3777", "0.8833", "0.3926", "0.5156", "0.3926" ], [ "0.147", "0.3914", "0.4575", "0.3914", "0.4575", "0.405", "0.147", "0.405" ], [ "0.517", "0.3943", "0.882", "0.3943", "0.882", "0.408", "0.517", "0.408" ], [ "0.1486", "0.408", "0.4592", "0.408", "0.4592", "0.4187", "0.1486", "0.4187" ], [ "0.5156", "0.411", "0.882", "0.411", "0.882", "0.4246", "0.5156", "0.4246" ], [ "0.147", "0.4216", "0.461", "0.4216", "0.461", "0.4353", "0.147", "0.4353" ], [ "0.517", "0.427", "0.8833", "0.427", "0.8833", "0.4407", "0.517", "0.4407" ], [ "0.147", "0.4353", "0.461", "0.4353", "0.461", "0.449", "0.147", "0.449" ], [ "0.5156", "0.4436", "0.8833", "0.4436", "0.8833", "0.4573", "0.5156", "0.4573" ], [ "0.147", "0.4495", "0.4592", "0.4495", "0.4592", "0.4631", "0.147", "0.4631" ], [ "0.147", "0.4631", "0.4592", "0.4631", "0.4592", "0.478", "0.147", "0.478" ], [ "0.513", "0.4578", "0.8833", "0.459", "0.8833", "0.4739", "0.513", "0.4727" ], [ "0.147", "0.478", "0.4592", "0.478", "0.4592", "0.4912", "0.147", "0.4912" ], [ "0.5146", "0.4744", "0.882", "0.4744", "0.882", "0.4893", "0.5146", "0.4893" ], [ "0.147", "0.4922", "0.4592", "0.4922", "0.4592", "0.506", "0.147", "0.506" ], [ "0.5146", "0.4912", "0.882", "0.4912", "0.882", "0.505", "0.5146", "0.505" ], [ "0.1453", "0.506", "0.461", "0.506", "0.461", "0.5195", "0.1453", "0.5195" ], [ "0.5156", "0.5073", "0.881", "0.5073", "0.881", "0.521", "0.5156", "0.521" ], [ "0.1436", "0.521", "0.461", "0.5195", "0.461", "0.533", "0.1436", "0.5337" ], [ "0.5156", "0.524", "0.881", "0.524", "0.881", "0.5376", "0.5156", "0.5376" ], [ "0.147", "0.5337", "0.4592", "0.5337", "0.4592", "0.5474", "0.147", "0.5474" ], [ "0.5146", "0.539", "0.885", "0.538", "0.885", "0.5527", "0.5146", "0.554" ], [ "0.147", "0.549", "0.461", "0.549", "0.461", "0.5625", "0.147", "0.5625" ], [ "0.5146", "0.5557", "0.7188", "0.5557", "0.7188", "0.5693", "0.5146", "0.5693" ], [ "0.147", "0.5635", "0.461", "0.5635", "0.461", "0.5776", "0.147", "0.5776" ], [ "0.147", "0.5767", "0.461", "0.5767", "0.461", "0.5903", "0.147", "0.5903" ], [ "0.532", "0.572", "0.8833", "0.572", "0.8833", "0.586", "0.532", "0.586" ], [ "0.1453", "0.5913", "0.4592", "0.589", "0.4592", "0.604", "0.1453", "0.6055" ], [ "0.517", "0.5884", "0.885", "0.5884", "0.885", "0.6016", "0.517", "0.6016" ], [ "0.147", "0.6055", "0.461", "0.6055", "0.461", "0.619", "0.147", "0.619" ], [ "0.5156", "0.604", "0.8833", "0.604", "0.8833", "0.618", "0.5156", "0.618" ], [ "0.1486", "0.6216", "0.4592", "0.6216", "0.4592", "0.632", "0.1486", "0.632" ], [ "0.5146", "0.619", "0.882", "0.619", "0.882", "0.6343", "0.5146", "0.6343" ], [ "0.1453", "0.6343", "0.461", "0.633", "0.461", "0.648", "0.1453", "0.6484" ], [ "0.5146", "0.6353", "0.881", "0.6353", "0.881", "0.6484", "0.5146", "0.6484" ], [ "0.147", "0.6484", "0.461", "0.6484", "0.461", "0.662", "0.147", "0.662" ], [ "0.5156", "0.652", "0.882", "0.652", "0.882", "0.665", "0.5156", "0.665" ], [ "0.1453", "0.661", "0.3687", "0.662", "0.3687", "0.676", "0.1453", "0.675" ], [ "0.5146", "0.6675", "0.881", "0.6675", "0.881", "0.681", "0.5146", "0.681" ], [ "0.1528", "0.6895", "0.266", "0.6904", "0.266", "0.705", "0.1528", "0.7036" ], [ "0.517", "0.685", "0.882", "0.685", "0.882", "0.6987", "0.517", "0.6987" ], [ "0.11584", "0.6953", "0.126", "0.6895", "0.1351", "0.697", "0.12427", "0.703" ], [ "0.5146", "0.7007", "0.882", "0.6997", "0.882", "0.7144", "0.5146", "0.7153" ], [ "0.12006", "0.7153", "0.4895", "0.7144", "0.4895", "0.7285", "0.12006", "0.73" ], [ "0.5146", "0.7163", "0.882", "0.7163", "0.882", "0.731", "0.5146", "0.731" ], [ "0.12006", "0.7324", "0.4902", "0.7324", "0.4902", "0.745", "0.12006", "0.745" ], [ "0.5146", "0.733", "0.882", "0.733", "0.882", "0.7466", "0.5146", "0.7466" ], [ "0.1184", "0.7476", "0.4902", "0.7476", "0.4902", "0.761", "0.1184", "0.761" ], [ "0.513", "0.7476", "0.885", "0.749", "0.885", "0.763", "0.513", "0.7627" ], [ "0.11755", "0.763", "0.492", "0.7627", "0.492", "0.777", "0.11755", "0.778" ], [ "0.513", "0.763", "0.881", "0.764", "0.881", "0.7793", "0.513", "0.778" ], [ "0.1184", "0.78", "0.4878", "0.78", "0.4878", "0.7935", "0.1184", "0.7935" ], [ "0.513", "0.781", "0.882", "0.781", "0.882", "0.7944", "0.513", "0.7944" ], [ "0.12006", "0.796", "0.4895", "0.7944", "0.4895", "0.8086", "0.12006", "0.81" ], [ "0.5146", "0.796", "0.831", "0.7964", "0.831", "0.811", "0.5146", "0.81" ], [ "0.12177", "0.8125", "0.4878", "0.8125", "0.4878", "0.826", "0.12177", "0.826" ], [ "0.529", "0.821", "0.8833", "0.821", "0.8833", "0.8345", "0.529", "0.8345" ], [ "0.1024", "0.832", "0.2544", "0.8345", "0.2544", "0.8486", "0.1024", "0.8467" ], [ "0.5483", "0.836", "0.885", "0.8374", "0.885", "0.8516", "0.5483", "0.851" ], [ "0.1024", "0.8457", "0.2747", "0.8467", "0.2747", "0.861", "0.1024", "0.86" ], [ "0.09906", "0.8555", "0.4902", "0.858", "0.4902", "0.876", "0.09906", "0.8735" ], [ "0.552", "0.854", "0.839", "0.854", "0.839", "0.8677", "0.552", "0.8677" ], [ "0.11755", "0.8716", "0.4902", "0.8706", "0.4902", "0.8853", "0.11755", "0.8857" ], [ "0.1184", "0.8843", "0.4902", "0.8843", "0.4902", "0.8984", "0.1184", "0.8984" ], [ "0.5283", "0.879", "0.885", "0.879", "0.885", "0.8926", "0.5283", "0.8926" ], [ "0.12006", "0.8965", "0.1646", "0.8965", "0.1646", "0.9077", "0.12006", "0.9077" ], [ "0.55", "0.8955", "0.885", "0.8955", "0.885", "0.91", "0.55", "0.91" ] ]

[ "我们在文本摘要上取得了什么成就？", "Dandan Huangl,2*, Leyang Cuil,2,3*, Sen Yang1,2*,", "Guangsheng Baoi2, Kun Wang, Jun Xie, Yue Zhangl", "1.2", "1 School of Engineering, Westlake University", "2 Institute of Advanced Technology, Westlake Institute for Advanced Study", "3 Zhejiang University, 4 Tencent SPPD", "[ huangdandan, cuileyang, yangsen, baoguangsheng } @westlake.edu.cn,", "[email protected], [email protected],[email protected]", "Abstract", "于深度学习的兴起，已经研究了提取E(Cheng", "and Lapata,2016:Xuand Durrett,2019)和抽", "Deep learning has led to significant im-", "象/Fl(Nallapati et al., 2016: Lewis et al.. 2019:", "provement in text summarization with var", "Balachandranetal.，2020）摘要系统的神经体", "ious methods investigated and improved", "ROUGE scores reported over the years.", "系结构", "However, gaps still exist between sum-", "尽管多年来已经在GigawordE(Graffetal.,", "maries produced by automatic summariz-", "2003)，NYTE(Grusky et al.，2018)和", "ers and human professionals. Aiming to", "CNN/DME(Hermann etal.,2015）等标准基", "gain more understanding of summarization", "准上报告了ROUGE分数的提高，但人们普", "systems with respect totheir strengths and", "遍认为，机器生成的摘要的质量仍然远远落后", "limits on a fine-grained syntactic and se-", "于人类书面的摘要。作为部分原因，ROUGE", "mantic level, we consult the Multidimen-", "被证明不足以作为摘要质量评估的精确指标E", "sional Quality Metric 1 (MQM) and quan-", "tify 8 major sources of errors on 10 rep", "(Liu and Liu, 2008; Bohm et al., 2019)。在", "resentative summarization models manu-", "研究文献中，人类评估已作为补充（Narayan", "ally. Primarily, we find that 1) under sim-", "etal.，2018）进行。然而，伴随着ROUGE分", "ilar settings, extractive summarizers are", "数的人类评估报告在范围和覆盖范围上都是有", "in general better than their abstractive", "限的。。在细粒度的层面上，仍然不确定我们", "counterparts thanks to strength in faith-", "总体上取得了什么成就，以及每个里程碑技术", "fulness and factual-consistency; 2) mile-", "带来了什么根本性的变化", "stone techniques such as copy, coverage", "and hybrid extractive/abstractive methods", "我们的目标是通过量化代表性模型的主要误", "do bring specific improvements but also", "差来源来解决上述问题。特别是，在MQME", "demonstrate limitations; 3) pre-training", "(Mariana，2014）之后，我们在Accuracy和", "techniques,and in particular sequence-to-", "Fluency方面设计了8个指标。根据每个指", "sequence pre-training, are highly effective", "标，通过测试集上的总体错误计数来分析模", "for improving text summarization, with", "型，因此与现有的手动评估报告相比，我们的", "BART giving the best results.", "评估可以更加翔实和客观。我们称这组度量", "Introduction", "为PolyTope。使用PolyTope，我们手动评", "I", "估了10个文本摘要器的结构，包括Lead-3", "自动文本摘要由于其实际重要性而受到不断", "TextRankE(Mihalcea and Tarau, 2004)，具", "的研究关注。现有方法可分为抽取式(Dorr", "有AttentionE(Rushetal.，2015）的序列到序", "et al., 2003; Mihalcea and Tarau, 2004; Nalla-", "列,SummaRuNNerE(Nallapati et al.,2017),", "patietal.,2017）和抽象式(JingandMcKe-", "Point-GeneratorE(Tu et al., 2016; See et al.,", "own, 2000; Rush et al., 2015; See et al., 2017)", "2017), Bottom-UpE(Gehrmann et al., 2018)", "两种方法，前者直接从原文中选取短语和句", "，BertSumExX。根据经验，我们发现：", "子作为摘要，后者利用词汇词合成删节。由", "1. Preneural vs Neural: Traditional rule-", "* Equal contribution.", "based methods are still strong baselines", "t Corresponding author.", "1 MQM is a framework for declaring and describing hu-", "given powerful neural architectures.", "man writing quality which stipulates a hierarchical list-", "ing of error types restricted to human writing and trans-", "2. Extractive vs Abstractive: Under simi-", "lation.", "lar settings, extractive approaches outper-" ]

{ "arxiv_id": "2010.04529", "language": "zh", "page_idx": 0 }

[ [ "0.1561", "0.0766", "0.4895", "0.0766", "0.4895", "0.09143", "0.1561", "0.09143" ], [ "0.5146", "0.076", "0.882", "0.0766", "0.882", "0.09143", "0.5146", "0.0903" ], [ "0.1561", "0.0932", "0.4902", "0.0932", "0.4902", "0.1069", "0.1561", "0.1069" ], [ "0.5146", "0.0932", "0.882", "0.0932", "0.882", "0.1069", "0.5146", "0.1069" ], [ "0.1561", "0.10925", "0.4937", "0.10925", "0.4937", "0.1223", "0.1561", "0.1223" ], [ "0.5146", "0.10925", "0.882", "0.10925", "0.882", "0.1223", "0.5146", "0.1223" ], [ "0.1545", "0.1259", "0.4902", "0.1259", "0.4902", "0.1395", "0.1545", "0.1395" ], [ "0.513", "0.1259", "0.882", "0.1259", "0.882", "0.1395", "0.513", "0.1395" ], [ "0.1561", "0.1426", "0.1838", "0.1426", "0.1838", "0.155", "0.1561", "0.155" ], [ "0.5156", "0.1401", "0.882", "0.1414", "0.882", "0.1561", "0.5156", "0.155" ], [ "0.513", "0.1567", "0.882", "0.158", "0.882", "0.1716", "0.513", "0.1704" ], [ "0.1318", "0.1675", "0.4902", "0.1675", "0.4902", "0.1812", "0.1318", "0.1812" ], [ "0.511", "0.1729", "0.879", "0.1729", "0.879", "0.1859", "0.511", "0.1859" ], [ "0.1561", "0.1841", "0.4902", "0.1841", "0.4902", "0.1978", "0.1561", "0.1978" ], [ "0.5156", "0.1895", "0.8423", "0.1895", "0.8423", "0.2031", "0.5156", "0.2031" ], [ "0.157", "0.1996", "0.4878", "0.1996", "0.4878", "0.2131", "0.157", "0.2131" ], [ "0.157", "0.2173", "0.4844", "0.2173", "0.4844", "0.228", "0.157", "0.228" ], [ "0.523", "0.2173", "0.8833", "0.2173", "0.8833", "0.231", "0.523", "0.231" ], [ "0.1561", "0.2322", "0.4895", "0.2322", "0.4895", "0.2465", "0.1561", "0.2465" ], [ "0.5156", "0.234", "0.8833", "0.234", "0.8833", "0.2477", "0.5156", "0.2477" ], [ "0.1561", "0.2477", "0.4895", "0.2477", "0.4895", "0.2625", "0.1561", "0.2625" ], [ "0.511", "0.2488", "0.881", "0.2488", "0.881", "0.2632", "0.511", "0.2632" ], [ "0.1561", "0.2632", "0.4902", "0.2632", "0.4902", "0.2778", "0.1561", "0.2778" ], [ "0.5156", "0.2666", "0.882", "0.2666", "0.882", "0.2798", "0.5156", "0.2798" ], [ "0.1561", "0.2798", "0.4902", "0.2798", "0.4902", "0.2935", "0.1561", "0.2935" ], [ "0.5146", "0.282", "0.8833", "0.282", "0.8833", "0.2957", "0.5146", "0.2957" ], [ "0.1545", "0.2957", "0.4895", "0.2944", "0.4895", "0.3088", "0.1545", "0.31" ], [ "0.517", "0.2986", "0.882", "0.2986", "0.882", "0.3123", "0.517", "0.3123" ], [ "0.1545", "0.3123", "0.2284", "0.3123", "0.2284", "0.3237", "0.1545", "0.3237" ], [ "0.5146", "0.3123", "0.881", "0.3123", "0.881", "0.3267", "0.5146", "0.3267" ], [ "0.519", "0.332", "0.881", "0.332", "0.881", "0.3425", "0.519", "0.3425" ], [ "0.1335", "0.3384", "0.4902", "0.3384", "0.4902", "0.3516", "0.1335", "0.3516" ], [ "0.5156", "0.3467", "0.882", "0.3467", "0.882", "0.3599", "0.5156", "0.3599" ], [ "0.1561", "0.354", "0.4895", "0.354", "0.4895", "0.3677", "0.1561", "0.3677" ], [ "0.5156", "0.3623", "0.882", "0.3623", "0.882", "0.376", "0.5156", "0.376" ], [ "0.1561", "0.3694", "0.4878", "0.3694", "0.4878", "0.383", "0.1561", "0.383" ], [ "0.5146", "0.3772", "0.882", "0.3777", "0.882", "0.3926", "0.5146", "0.3914" ], [ "0.1561", "0.386", "0.486", "0.386", "0.486", "0.3997", "0.1561", "0.3997" ], [ "0.5146", "0.3938", "0.882", "0.3938", "0.882", "0.4067", "0.5146", "0.4067" ], [ "0.1561", "0.4026", "0.4878", "0.4026", "0.4878", "0.4163", "0.1561", "0.4163" ], [ "0.5156", "0.4104", "0.882", "0.4104", "0.882", "0.4233", "0.5156", "0.4233" ], [ "0.1561", "0.4187", "0.486", "0.4187", "0.486", "0.4316", "0.1561", "0.4316" ], [ "0.5156", "0.4258", "0.882", "0.427", "0.882", "0.4412", "0.5156", "0.4407" ], [ "0.1561", "0.4353", "0.4878", "0.4353", "0.4878", "0.449", "0.1561", "0.449" ], [ "0.5156", "0.4436", "0.881", "0.4436", "0.881", "0.4573", "0.5156", "0.4573" ], [ "0.1561", "0.4507", "0.486", "0.4507", "0.486", "0.4644", "0.1561", "0.4644" ], [ "0.5156", "0.459", "0.882", "0.459", "0.882", "0.4727", "0.5156", "0.4727" ], [ "0.1587", "0.4685", "0.486", "0.4685", "0.486", "0.4785", "0.1587", "0.4785" ], [ "0.513", "0.4739", "0.882", "0.4744", "0.882", "0.488", "0.513", "0.4868" ], [ "0.157", "0.4827", "0.345", "0.4827", "0.345", "0.4963", "0.157", "0.4963" ], [ "0.5156", "0.4912", "0.882", "0.4912", "0.882", "0.505", "0.5156", "0.505" ], [ "0.1377", "0.5083", "0.4878", "0.5083", "0.4878", "0.5215", "0.1377", "0.5215" ], [ "0.5156", "0.509", "0.5356", "0.509", "0.5356", "0.5186", "0.5156", "0.5186" ], [ "0.11755", "0.5215", "0.4895", "0.524", "0.4895", "0.538", "0.11755", "0.536" ], [ "0.529", "0.524", "0.8833", "0.525", "0.8833", "0.539", "0.529", "0.538" ], [ "0.12006", "0.538", "0.4878", "0.538", "0.4878", "0.552", "0.12006", "0.552" ], [ "0.513", "0.5405", "0.882", "0.5405", "0.882", "0.554", "0.513", "0.554" ], [ "0.12006", "0.5557", "0.4358", "0.5557", "0.4358", "0.5693", "0.12006", "0.5693" ], [ "0.5146", "0.5547", "0.8833", "0.5557", "0.8833", "0.571", "0.5146", "0.5693" ], [ "0.5146", "0.572", "0.882", "0.572", "0.882", "0.585", "0.5146", "0.585" ], [ "0.1184", "0.583", "0.2761", "0.583", "0.2761", "0.597", "0.1184", "0.597" ], [ "0.517", "0.5884", "0.8833", "0.5884", "0.8833", "0.6025", "0.517", "0.6025" ], [ "0.126", "0.608", "0.4878", "0.608", "0.4878", "0.6216", "0.126", "0.6216" ], [ "0.5156", "0.605", "0.882", "0.605", "0.882", "0.618", "0.5156", "0.618" ], [ "0.12006", "0.6235", "0.4895", "0.6235", "0.4895", "0.637", "0.12006", "0.637" ], [ "0.5156", "0.6216", "0.8833", "0.6216", "0.8833", "0.6353", "0.5156", "0.6353" ], [ "0.1184", "0.6396", "0.4895", "0.6396", "0.4895", "0.653", "0.1184", "0.653" ], [ "0.5156", "0.637", "0.882", "0.637", "0.882", "0.651", "0.5156", "0.651" ], [ "0.12006", "0.6562", "0.4878", "0.6562", "0.4878", "0.6694", "0.12006", "0.6694" ], [ "0.517", "0.653", "0.882", "0.653", "0.882", "0.6665", "0.517", "0.6665" ], [ "0.12177", "0.673", "0.4895", "0.673", "0.4895", "0.686", "0.12177", "0.686" ], [ "0.5156", "0.6694", "0.679", "0.6694", "0.679", "0.683", "0.5156", "0.683" ], [ "0.12006", "0.687", "0.4878", "0.687", "0.4878", "0.7007", "0.12006", "0.7007" ], [ "0.534", "0.686", "0.882", "0.686", "0.882", "0.7007", "0.534", "0.7007" ], [ "0.1184", "0.703", "0.486", "0.703", "0.486", "0.7163", "0.1184", "0.7163" ], [ "0.5146", "0.703", "0.881", "0.703", "0.881", "0.7163", "0.5146", "0.7163" ], [ "0.1335", "0.72", "0.4878", "0.72", "0.4878", "0.734", "0.1335", "0.734" ], [ "0.5146", "0.7188", "0.882", "0.7188", "0.882", "0.7324", "0.5146", "0.7324" ], [ "0.126", "0.7246", "0.1368", "0.7246", "0.1368", "0.73", "0.126", "0.73" ], [ "0.11755", "0.7354", "0.287", "0.7363", "0.287", "0.7505", "0.11755", "0.75" ], [ "0.5146", "0.734", "0.881", "0.734", "0.881", "0.7476", "0.5146", "0.7476" ], [ "0.5146", "0.7505", "0.882", "0.75", "0.882", "0.764", "0.5146", "0.7656" ], [ "0.1276", "0.7603", "0.4895", "0.7603", "0.4895", "0.775", "0.1276", "0.775" ], [ "0.5146", "0.7656", "0.881", "0.7656", "0.881", "0.7793", "0.5146", "0.7793" ], [ "0.12006", "0.777", "0.4878", "0.777", "0.4878", "0.7905", "0.12006", "0.7905" ], [ "0.5146", "0.782", "0.879", "0.782", "0.879", "0.796", "0.5146", "0.796" ], [ "0.12006", "0.7935", "0.4895", "0.7935", "0.4895", "0.807", "0.12006", "0.807" ], [ "0.5146", "0.799", "0.879", "0.799", "0.879", "0.8135", "0.5146", "0.8135" ], [ "0.1184", "0.8086", "0.4836", "0.8086", "0.4836", "0.8223", "0.1184", "0.8223" ], [ "0.511", "0.814", "0.881", "0.8135", "0.881", "0.8276", "0.511", "0.829" ], [ "0.1318", "0.8237", "0.4878", "0.8247", "0.4878", "0.839", "0.1318", "0.8384" ], [ "0.5146", "0.8306", "0.882", "0.8306", "0.882", "0.844", "0.5146", "0.844" ], [ "0.12177", "0.8413", "0.4878", "0.8413", "0.4878", "0.855", "0.12177", "0.855" ], [ "0.5146", "0.847", "0.882", "0.847", "0.882", "0.861", "0.5146", "0.861" ], [ "0.12177", "0.857", "0.4902", "0.857", "0.4902", "0.8706", "0.12177", "0.8706" ], [ "0.5146", "0.8623", "0.882", "0.8633", "0.882", "0.8774", "0.5146", "0.877" ], [ "0.12006", "0.8726", "0.4878", "0.8726", "0.4878", "0.887", "0.12006", "0.887" ], [ "0.5156", "0.88", "0.882", "0.88", "0.882", "0.8936", "0.5156", "0.8936" ], [ "0.09906", "0.891", "0.3896", "0.8945", "0.3896", "0.913", "0.09906", "0.91" ], [ "0.517", "0.8955", "0.882", "0.8955", "0.882", "0.909", "0.517", "0.909" ] ]

[ "form abstractive models in general. The", "，复制E(Guetal.,2016），覆盖E(Tuetal.,", "main shortcoming is unnecessity for ex-", "2016;See etal.,2017)，平滑E(Milleretal.,", "tractive models, and omission and in-", "2019)和预训练E(Lewisetal.,2019;Liuand", "trinsic hallucination for abstractive mod-", "Lapata，2019）之类的技术，以增强摘要。混", "els.", "合抽象和提取方法采用两步法，包括内容选", "择和文本生成（Gehrmannet al.，2018；Hsu", "3. Milestone Techniques: Copy works effec-", "etal.,2018:Celikyilmazetal.,2018），实现了", "tively in reproducing details. It also re-", "比ROUGE中的端到端模型更高的性能。", "duces duplication on the word level but", "tends to cause redundancy to a certair", "Analysis of Summarization 基于 ROUGE", "degree. Coverage solves repetition errors", "分析摘要系统的工作很多。Lapata 和Barzi-", "by a large margin, but shows limits in", "lay（2005）探讨了机器生成摘要中“一致性”", "faithful content generation. Hybrid ex-", "的基本方面。Zhang等人。（2018）分析了抽", "tractive/abstractive models reflect the rel-", "象系统，而Kedzie等人。（2018）和Zhong等", "ative strengths and weaknesses of the two", "人。（2019）在提取汇总中搜索有效的体系结", "methods.", "构。Kryscinski等人。（2019）从允余性、相", "关性和信息量三个方面评估了摘要的整体质", "4. Pre-training: Pre-training is highly effec-", "量。以上所有都依赖于自动评估指标。我们的", "tive for summarization, and even achieves", "工作与这些努力是一致的，因为我们对各个", "a better content selection capability with-", "方面进行了精细的评估。与上述工作不同，我", "out copy and coverage mechanisms. Par", "们使用人类评估而不是自动评估。事实上，上", "ticularly, joint pre-training combining", "述分析工作在得出丰富结论的同时，也暴露出", "text understanding and generation gives", "自动化工具包的不足。自动评估的质量经常受", "the most salient advantage, with the", "到研究界（Novikovaetal.,2017;Zopf,2018)", "BART model achieving by far the state", "的批评，因为它既没有渗透到基于生成的文", "of-the-artresults onboth automaticand", "本（LiuandLiu，2008）的整体质量中，也没", "our human evaluations.", "有与人类判断（Kryscinskietal.，2019）相关", "我们发布了测试集，其中包括10个基于", "联", "PolyTope的系统输出及其手动标记的错误，", "还进行了分析工作，通过人类评估（Narayan", "以及一个用户友好的评估工具包，可帮助将来", "et al.,2018;Liu and Lapata,2019)增强了", "对评估方法和自动摘要系统2进行研究", "ROUGE。这样的工作报告粗粒度的人类评", "估分数，通常包括2到3个方面，如信息，流", "2Related Work", "利和简洁。最近，Maynezetal.（2020）对500", "ExtractiveSummarization基于统计方法的", "篇文章进行了5个神经抽象模型的人体评估。", "早期工作E(Neto etal.,2002;Mihalcea and", "他们的主要目标是验证抽象模型的真实性和真", "Tarau，2004）利用专业知识手动设计功能或规", "实性。相比之下，我们在8个错误度量上评估", "则。基于神经架构的最新工作将摘要视为单", "了基于规则的基线和提取/抽象汇总器，其中", "词或句子级别的分类问题，并通过计算句子表", "包括真实性和真实性", "示形式来解决它E(Chengand Lapata，2016;", "我们的工作还与人类评估的研究有关，以进", "Nallapati et al., 2017: Xu and Durrett, 2019)", "行总结。为此，PyramidE(Nenkova andPas", "最近，Zhongetal.（2020）采用文档级功能", "sonneau，2004）根据其系统输出和多个引用对", "", "来重新排序提取摘要", "摘要器进行评分。注释者被要求识别语义意义", "的最小内容单元，然后通过计算有多少引用摘", "Abstractive Summarization Jing 和 McKe-", "要包含这个单元来将每个单元与一个权重相", "ownE(2000）提出了一种基于剪切粘贴的抽象", "关联。摘要的分数是根据单位的数量和权重计", "摘要器，该摘要器将提取的片段编辑并合并", "算的。除金字塔外，还有基于排名（Narayar", "为连贯的句子。Rush等人。，（2015）提出了", "etal.，2018），最佳-最差缩放E（Kiritchenko", "一种用于抽象摘要的序列到序列架构。随后，", "and Mohammad,2017)和问答E(Clarkeand", "使用了Transformer，并通过ROUGE分数E", "Lapata，2010）的人类评估指标。上述方法为每", "(Duanetal.，2019）超越了传统的抽象摘要器。", "个汇总输出分配一个分数。与这些方法相比，", "还研究了诸如AMR解析E(Liuetal.，2015)", "我们的基于错误计数的指标是由MQM用于", "2 https:/ /github.com/hddbang/PolyTope", "人类写作的，并且更加细粒度和信息丰富。我" ]

{ "arxiv_id": "2010.04529", "language": "zh", "page_idx": 1 }

[ [ "0.3257", "0.0736", "0.44", "0.0736", "0.44", "0.08435", "0.3257", "0.08435" ], [ "0.623", "0.0736", "0.742", "0.0736", "0.742", "0.08435", "0.623", "0.08435" ], [ "0.677", "0.0831", "0.7515", "0.0849", "0.7515", "0.0968", "0.677", "0.0944" ], [ "0.2091", "0.076", "0.2568", "0.076", "0.2568", "0.0831", "0.2091", "0.0831" ], [ "0.1436", "0.0849", "0.1957", "0.0849", "0.1957", "0.0956", "0.1436", "0.0956" ], [ "0.2688", "0.0849", "0.382", "0.0849", "0.382", "0.0956", "0.2688", "0.0956" ], [ "0.3896", "0.0849", "0.4417", "0.0849", "0.4417", "0.0956", "0.3896", "0.0956" ], [ "0.4358", "0.0849", "0.4978", "0.0849", "0.4978", "0.0956", "0.4358", "0.0956" ], [ "0.5903", "0.08435", "0.671", "0.0849", "0.671", "0.0968", "0.5903", "0.0956" ], [ "0.7446", "0.0849", "0.8613", "0.0849", "0.8613", "0.0956", "0.7446", "0.0956" ], [ "0.553", "0.0861", "0.5767", "0.0861", "0.5767", "0.0944", "0.553", "0.0944" ], [ "0.509", "0.0861", "0.5366", "0.0861", "0.5366", "0.0944", "0.509", "0.0944" ], [ "0.1721", "0.0968", "0.2317", "0.0968", "0.2317", "0.1081", "0.1721", "0.1081" ], [ "0.272", "0.0944", "0.309", "0.0968", "0.3074", "0.10925", "0.2703", "0.1069" ], [ "0.3342", "0.0968", "0.3687", "0.0968", "0.3687", "0.1081", "0.3342", "0.1081" ], [ "0.3955", "0.0968", "0.4324", "0.0968", "0.4324", "0.1081", "0.3955", "0.1081" ], [ "0.455", "0.0968", "0.4895", "0.0968", "0.4895", "0.1081", "0.455", "0.1081" ], [ "0.613", "0.0968", "0.65", "0.0968", "0.65", "0.1081", "0.613", "0.1081" ], [ "0.6978", "0.0968", "0.7314", "0.0968", "0.7314", "0.1081", "0.6978", "0.1081" ], [ "0.7656", "0.0968", "0.804", "0.0968", "0.804", "0.1081", "0.7656", "0.1081" ], [ "0.818", "0.0974", "0.858", "0.0974", "0.858", "0.10925", "0.818", "0.10925" ], [ "0.3342", "0.1057", "0.3687", "0.1081", "0.367", "0.1194", "0.3333", "0.11755" ], [ "0.504", "0.0974", "0.5825", "0.0974", "0.5825", "0.10925", "0.504", "0.10925" ], [ "0.6978", "0.1057", "0.7305", "0.1081", "0.729", "0.1194", "0.6963", "0.11755" ], [ "0.818", "0.1057", "0.8564", "0.10986", "0.854", "0.12054", "0.8154", "0.1164" ], [ "0.2737", "0.1081", "0.309", "0.1081", "0.309", "0.1194", "0.2737", "0.1194" ], [ "0.1721", "0.1081", "0.2317", "0.1081", "0.2317", "0.11816", "0.1721", "0.11816" ], [ "0.3955", "0.1081", "0.4324", "0.1081", "0.4324", "0.1194", "0.3955", "0.1194" ], [ "0.4534", "0.1081", "0.4878", "0.1081", "0.4878", "0.1194", "0.4534", "0.1194" ], [ "0.5054", "0.10925", "0.5366", "0.10925", "0.5366", "0.11755", "0.5054", "0.11755" ], [ "0.5483", "0.10925", "0.581", "0.10925", "0.581", "0.11755", "0.5483", "0.11755" ], [ "0.614", "0.1081", "0.6484", "0.1081", "0.6484", "0.1194", "0.614", "0.1194" ], [ "0.7676", "0.10925", "0.8003", "0.10925", "0.8003", "0.11755", "0.7676", "0.11755" ], [ "0.1704", "0.11816", "0.2334", "0.11816", "0.2334", "0.13", "0.1704", "0.13" ], [ "0.266", "0.11816", "0.309", "0.11816", "0.309", "0.13", "0.266", "0.13" ], [ "0.3333", "0.11816", "0.3687", "0.11816", "0.3687", "0.13", "0.3333", "0.13" ], [ "0.3938", "0.11816", "0.4307", "0.11816", "0.4307", "0.13", "0.3938", "0.13" ], [ "0.4534", "0.11816", "0.4902", "0.11816", "0.4902", "0.13", "0.4534", "0.13" ], [ "0.5015", "0.11816", "0.5825", "0.11816", "0.5825", "0.13", "0.5015", "0.13" ], [ "0.6094", "0.11816", "0.6484", "0.11816", "0.6484", "0.13", "0.6094", "0.13" ], [ "0.6963", "0.11755", "0.7305", "0.11755", "0.7305", "0.13", "0.6963", "0.13" ], [ "0.764", "0.11816", "0.8003", "0.11816", "0.8003", "0.13", "0.764", "0.13" ], [ "0.817", "0.11816", "0.8564", "0.11816", "0.8564", "0.13", "0.817", "0.13" ], [ "0.1184", "0.1426", "0.8833", "0.1426", "0.8833", "0.1561", "0.1184", "0.1561" ], [ "0.12006", "0.1567", "0.8345", "0.1567", "0.8345", "0.1704", "0.12006", "0.1704" ], [ "0.12006", "0.1859", "0.486", "0.1859", "0.486", "0.1996", "0.12006", "0.1996" ], [ "0.5146", "0.1859", "0.8833", "0.1859", "0.8833", "0.1996", "0.5146", "0.1996" ], [ "0.1184", "0.2019", "0.486", "0.2007", "0.486", "0.2144", "0.1184", "0.2156" ], [ "0.5146", "0.2019", "0.8833", "0.2019", "0.8833", "0.2156", "0.5146", "0.2156" ], [ "0.1234", "0.2197", "0.486", "0.2197", "0.486", "0.2299", "0.1234", "0.2299" ], [ "0.5146", "0.2173", "0.817", "0.2173", "0.817", "0.2322", "0.5146", "0.2322" ], [ "0.12177", "0.2351", "0.4878", "0.2351", "0.4878", "0.2488", "0.12177", "0.2488" ], [ "0.523", "0.2434", "0.882", "0.2434", "0.882", "0.257", "0.523", "0.257" ], [ "0.12006", "0.2505", "0.4878", "0.2505", "0.4878", "0.2642", "0.12006", "0.2642" ], [ "0.517", "0.2588", "0.882", "0.2588", "0.882", "0.2725", "0.517", "0.2725" ], [ "0.1184", "0.2654", "0.2393", "0.2666", "0.2393", "0.2808", "0.1184", "0.2798" ], [ "0.513", "0.275", "0.881", "0.275", "0.881", "0.288", "0.513", "0.288" ], [ "0.11755", "0.2935", "0.22", "0.2935", "0.22", "0.308", "0.11755", "0.308" ], [ "0.5156", "0.2903", "0.881", "0.2903", "0.881", "0.304", "0.5156", "0.304" ], [ "0.5146", "0.306", "0.854", "0.3071", "0.854", "0.3218", "0.5146", "0.3206" ], [ "0.1184", "0.3184", "0.486", "0.3184", "0.486", "0.332", "0.1184", "0.332" ], [ "0.1184", "0.335", "0.4878", "0.3362", "0.4878", "0.3508", "0.1184", "0.3499" ], [ "0.5264", "0.332", "0.8833", "0.332", "0.8833", "0.3467", "0.5264", "0.3467" ], [ "0.12006", "0.3528", "0.2494", "0.3528", "0.2494", "0.3665", "0.12006", "0.3665" ], [ "0.5156", "0.3486", "0.882", "0.3486", "0.882", "0.3623", "0.5156", "0.3623" ], [ "0.5156", "0.364", "0.882", "0.364", "0.882", "0.3777", "0.5156", "0.3777" ], [ "0.1184", "0.379", "0.3284", "0.379", "0.3284", "0.3926", "0.1184", "0.3926" ], [ "0.5146", "0.379", "0.854", "0.3801", "0.854", "0.3943", "0.5146", "0.3938" ], [ "0.1234", "0.401", "0.4878", "0.401", "0.4878", "0.4146", "0.1234", "0.4146" ], [ "0.5205", "0.405", "0.882", "0.405", "0.882", "0.4192", "0.5205", "0.4192" ], [ "0.12177", "0.4175", "0.4878", "0.4175", "0.4878", "0.4312", "0.12177", "0.4312" ], [ "0.517", "0.4229", "0.882", "0.4229", "0.882", "0.4358", "0.517", "0.4358" ], [ "0.12006", "0.4329", "0.4878", "0.4329", "0.4878", "0.4465", "0.12006", "0.4465" ], [ "0.5156", "0.4382", "0.8833", "0.4382", "0.8833", "0.452", "0.5156", "0.452" ], [ "0.1184", "0.449", "0.486", "0.449", "0.486", "0.462", "0.1184", "0.462" ], [ "0.5146", "0.4536", "0.8833", "0.4536", "0.8833", "0.4673", "0.5146", "0.4673" ], [ "0.1184", "0.4631", "0.4878", "0.4644", "0.4878", "0.4785", "0.1184", "0.478" ], [ "0.5156", "0.4714", "0.5415", "0.4714", "0.5415", "0.484", "0.5156", "0.484" ], [ "0.12177", "0.481", "0.476", "0.481", "0.476", "0.4946", "0.12177", "0.4946" ], [ "0.5483", "0.4946", "0.8833", "0.4946", "0.8833", "0.5083", "0.5483", "0.5083" ], [ "0.12427", "0.505", "0.4878", "0.5073", "0.4878", "0.521", "0.12427", "0.5186" ], [ "0.5146", "0.511", "0.882", "0.511", "0.882", "0.525", "0.5146", "0.525" ], [ "0.1184", "0.5225", "0.4878", "0.5225", "0.4878", "0.536", "0.1184", "0.536" ], [ "0.5146", "0.527", "0.881", "0.527", "0.881", "0.5405", "0.5146", "0.5405" ], [ "0.1184", "0.5376", "0.4878", "0.538", "0.4878", "0.552", "0.1184", "0.5503" ], [ "0.5146", "0.5435", "0.882", "0.5435", "0.882", "0.557", "0.5146", "0.557" ], [ "0.12177", "0.5547", "0.4878", "0.5547", "0.4878", "0.5684", "0.12177", "0.5684" ], [ "0.5146", "0.558", "0.8345", "0.56", "0.8345", "0.5737", "0.5146", "0.572" ], [ "0.1184", "0.571", "0.486", "0.571", "0.486", "0.5845", "0.1184", "0.5845" ], [ "0.12006", "0.5874", "0.3123", "0.5874", "0.3123", "0.601", "0.12006", "0.601" ], [ "0.5205", "0.583", "0.882", "0.5845", "0.882", "0.5986", "0.5205", "0.597" ], [ "0.517", "0.601", "0.882", "0.601", "0.882", "0.6143", "0.517", "0.6143" ], [ "0.1234", "0.611", "0.4878", "0.611", "0.4878", "0.6245", "0.1234", "0.6245" ], [ "0.517", "0.618", "0.879", "0.618", "0.879", "0.629", "0.517", "0.629" ], [ "0.12006", "0.629", "0.4878", "0.629", "0.4878", "0.6426", "0.12006", "0.6426" ], [ "0.5146", "0.633", "0.882", "0.633", "0.882", "0.6465", "0.5146", "0.6465" ], [ "0.12177", "0.6445", "0.4895", "0.6445", "0.4895", "0.6577", "0.12177", "0.6577" ], [ "0.5146", "0.6484", "0.791", "0.6484", "0.791", "0.662", "0.5146", "0.662" ], [ "0.12177", "0.661", "0.4844", "0.661", "0.4844", "0.6714", "0.12177", "0.6714" ], [ "0.12006", "0.677", "0.4878", "0.677", "0.4878", "0.6904", "0.12006", "0.6904" ], [ "0.513", "0.6797", "0.55", "0.6797", "0.55", "0.6895", "0.513", "0.6895" ], [ "0.5415", "0.676", "0.7173", "0.676", "0.7173", "0.6895", "0.5415", "0.6895" ], [ "0.12177", "0.695", "0.486", "0.695", "0.486", "0.705", "0.12177", "0.705" ], [ "0.5146", "0.7007", "0.881", "0.7007", "0.881", "0.7144", "0.5146", "0.7144" ], [ "0.1184", "0.708", "0.3477", "0.709", "0.3477", "0.7236", "0.1184", "0.7227" ], [ "0.5156", "0.7188", "0.882", "0.7188", "0.882", "0.7324", "0.5156", "0.7324" ], [ "0.1234", "0.7354", "0.4895", "0.7354", "0.4895", "0.749", "0.1234", "0.749" ], [ "0.5146", "0.734", "0.882", "0.733", "0.882", "0.7476", "0.5146", "0.749" ], [ "0.12177", "0.7505", "0.4895", "0.7505", "0.4895", "0.764", "0.12177", "0.764" ], [ "0.5156", "0.7505", "0.882", "0.7505", "0.882", "0.764", "0.5156", "0.764" ], [ "0.12177", "0.7666", "0.4895", "0.7666", "0.4895", "0.78", "0.12177", "0.78" ], [ "0.517", "0.7666", "0.882", "0.7666", "0.882", "0.78", "0.517", "0.78" ], [ "0.12006", "0.783", "0.4878", "0.783", "0.4878", "0.7964", "0.12006", "0.7964" ], [ "0.5146", "0.783", "0.882", "0.783", "0.882", "0.7964", "0.5146", "0.7964" ], [ "0.12006", "0.799", "0.4895", "0.799", "0.4895", "0.8125", "0.12006", "0.8125" ], [ "0.513", "0.799", "0.791", "0.799", "0.791", "0.8125", "0.513", "0.8125" ], [ "0.1184", "0.8135", "0.3132", "0.814", "0.3132", "0.829", "0.1184", "0.8276" ], [ "0.531", "0.8135", "0.882", "0.814", "0.882", "0.829", "0.531", "0.8276" ], [ "0.5146", "0.8306", "0.882", "0.8306", "0.882", "0.844", "0.5146", "0.844" ], [ "0.12006", "0.8413", "0.3376", "0.8413", "0.3376", "0.855", "0.12006", "0.855" ], [ "0.5146", "0.8457", "0.8833", "0.8467", "0.8833", "0.861", "0.5146", "0.86" ], [ "0.1234", "0.8633", "0.4895", "0.8633", "0.4895", "0.877", "0.1234", "0.877" ], [ "0.5156", "0.8633", "0.8833", "0.8633", "0.8833", "0.877", "0.5156", "0.877" ], [ "0.12177", "0.879", "0.4963", "0.879", "0.4963", "0.8926", "0.12177", "0.8926" ], [ "0.5146", "0.879", "0.882", "0.879", "0.882", "0.8926", "0.5146", "0.8926" ], [ "0.12177", "0.8955", "0.4878", "0.8955", "0.4878", "0.909", "0.12177", "0.909" ], [ "0.5156", "0.8955", "0.882", "0.8955", "0.882", "0.909", "0.5156", "0.909" ] ]

[ "Extractive Methods", "Abstractive Methods", "Bottom-Up", "Methods", "ROUGE", "Lead-3TextRank", "Summa", "BertExt", "PG-Coverage", "BertAbsBART", "PG", "S2S", "ROUGE-1", "39.20", "40.20", "39.60", "43.25", "39.53", "41.22", "42.13", "44.16", "17.56", "31.3336.44", "18.68", "21.28", "15.70", "ROUGE-2", "16.20", "20.24", "11.81", "15.66", "17.28", "19.60", "ROUGE-L", "35.50", "36.44", "35.30", "39.63", "28.8033.42", "36.38", "38.34", "39.18", "40.90", "Table 1: ROUGE scores of 10 summarizers on CNN/DM Dataset (non-anonymous version). We get the", "score of Lead-3 and TextRank from Nallapati et al. (2017) and Zhou et al. (2018) , respectively.", "们在6节中的图3中显示了更多评估指标之", "的选择和重新排列，采用了目标到来源的关注", "间的经验对比。最近，EStiennonetal.（2020)", "模块来捕获来自每个编码器隐藏状态的信息。", "使用人类评估作为训练自动摘要器的奖励，与", "我们遵循See等人的实施。（2017）。", "使用参考摘要训练的模型相比，报告了显着的", "Pointer-Generator See et al.(2017)引人", "改进。他们的工作还证明了人类评估在文本摘", "了指针网络（Vinyalsetal.，2015），以解决", "要中的有用性。", "seq2seq模型倾向于不准确地再现事实细节的", "3 Models", "问题。该方法既可以通过生成器从词汇表中生", "成单词，又可以通过指针从源中复制内容。", "我们重新实施并评估了10种具有代表性和影", "响力的方法。他们公开报道的ROUGEF1分", "Pointer-Generator-with-Coverage参见等。", "数如表1所示。", "(2017）使用覆盖机制（Tuetal.，2016）避免重", "复问题。该机制计算覆盖向量作为注意力机制", "3.1Extractive Methods", "的额外输入，以加强对不同位置的注意力。", "Lead-3Lead-3是常用的基线，它简单地选择", "Bottom-UpGehrmann等人。(2018）提出", "前三个句子作为摘要。它被最近的工作用作标", "了一种两步法，首先选择潜在的输出词，然后", "准基线（Cheng and Lapata,2016;Gehrmann", "基于指针生成器网络生成摘要。Bottom-Up被", "etal.，2018）。从直觉上讲，新闻领域文章的", "选为集成了提取和抽象方法的混合模型的代", "前三句很可能是其摘要，因此Lead-3的结果", "表。", "可以是人类撰写的摘要的高度可靠的近似值。", "BertSumExtAbs BertSumExtAbsE(Liu", "TextRank TextRank (Mihalcea and Tarau,", "andLapata，2019）采用与BertSumExt相同", "2004）是一种基于图排序模型的无监督关键文", "的编码器，以及具有随机初始化参数的6层", "本单元选择方法（Pageetal，1998）。它通过", "Transformer解码器。它被选为具有预训练上", "计算句子之间的相似度并相应地生成加权图来", "下文化句子表示的神经抽象模型的代表。", "定义“推荐”。提取权重高的句子作为摘要。它", "被选为统计模型的代表。", "BARTBARTE(Lewisetal.,2019)联合预", "训练seq2seq模型，结合双向编码器和自回归", "SummaRuNNer SummaRuNNer (Nallapati", "解码器，而不是仅对编码器进行预训练。在汇总", "etal.,2017）是一个代表性的神经提取模型，它", "数据集上进行了进一步的微调，它在ROUGE", "从输人中选择完整的句子作为摘要。它首先使", "分数方面达到了当前最先进的结果", "用分层的BiGRU对输入进行编码，然后从左", "到右扫描输入句子。通过所有先前选择的加权", "4", "Evaluation Method", "和生成累积的摘要表示，将其馈送到逻辑分类", "我们通过使用ROUGE（Lin，2004）进行自动", "器中以对摘要进行最终预测。", "评分和PolyTope进行人工评分来分析系统性", "BertSumExt BertSumExt (Liu and Lapata,", "能。ROUGE已被大多数总结工作所采用。它", "2019）将预先训练的BERT（Devlinetal.,", "是一种基于召回的度量，用于计算系统输出和", "2019）作为句子编码器，并将附加的Trans-", "人类摘要之间的词汇重叠。特别是，ROUGE-1", "former作为文档编码器。基于句子表示的分类", "基于unigram重叠，ROUGE-2基于bigram", "器用于句子选择。它利用来自微调BERT的", "ROUGE-L基于最长的公共子序列", "知识来生成更好的摘要。", "PolyTope是一种基于多维质量度量", "（MQM）E(Mariana，2014）的面向错误的细", "3.2Abstractive Methods", "粒度人类评估方法。特别是，它由8种问题类", "Seq2SeqwithAttentionRush等人首先将序", "型E4.1节)，8种语法标签（E4.2节）和一", "列到序列模型结构用于抽象摘要。（2015）。", "组严重性规则（E4.3节）组成，用于定位错误", "为了允许有效和自由的文本生成，而不是简单", "并自动计算测试文档的总体得分。如图E3所" ]

{ "arxiv_id": "2010.04529", "language": "zh", "page_idx": 2 }