id
string | image
string | input
string | system
string | output
string | boxes
list | box_texts
list | detail
dict |
|---|---|---|---|---|---|---|---|
2105.11174-zh-1
|
2105.11174-zh-1.jpg
|
<FigureHere>
Figure 1: The overall framework of Retrieved En-
hanced Model for Commonsense Generation.
理受益。我们通过提出检索增强 T5 (RE-T5) 来扩展此想法，该技术为原始 T5 配备了可训练的检索器，用于根据给定的概念选择原型句子。同时，参考(Zhou et al.,2021),我们为CommonGen 设计了一个预训练任务，该任务继续在从外部语料库中提取的伪概念集上对RE-T5 进行预训练。在这个预训练阶段，我们还使用了猎犬
形式上，给定一个概念集$X=$ $\{ x_1, x_2, \ldots , x_n\}$ ,其中$x_i$ 表示第$i$个概念， $n$是概念的数量，我们的目标是生成令牌$Y$ = $\{ y_1, y_2, \ldots , y_m\}$ 的自然语言输出，该输出使用$X$中的所有给定概念来描述我们日常生活中的常见场景
## 2.1 Retrieval
由于外部语料库具有大量的场景知识来描述概念之间的关系$\overline{\mathrm{E}}($Fan et al.,2020),因此我们检索与输入概念相关的句子以帮助模型执行更好的常识推理。首先，给定一个输入概念集，我们从外部语料库中提取输入$X$ 中至少包含两个概念的所有句子作为候选集$z$ 。然后，我们设计了两个检索模型，匹配检索器和可训练检索器，以进一步检索$k$原型句子$Z=\{z_1,z_2,\ldots,z_k\},Z\subseteq\mathcal{Z}$作为 RE-T5 的辅助输人上下文
Matching Retriever 匹配的检索器首先根据所包含概念的数量对候选句子进行排序。然后，它简单地从包含最多概念的句子作为辅助输入开始对$k$句子进行采样
Trainable Retriever 为了从句子候选集中检索更多有用的句子，我们设计了一个可训练的检索器，该检索器预测分数以对这些候选进行排名，然后选择 top- $k$句子作为附加上下文。评分器基于 BERT (Devlin et al., 2019) 构建， 这是一种通常用于语言理解的预训练语言模型。给定一个概念集$X$和一个候选句子$z_i$, 我们的可训练检索器首先将它们连接成一个文本输入：
$$[\mathrm C\mathrm L\mathrm S]X[\mathrm S\mathrm E\mathrm P]z_i[\mathrm S\mathrm E\mathrm P]$$
其中[CLS] 和[SEP] 是 BERT 中的特殊符号
我们将其传递给 BERT,该 BERT 为每个输入令牌生成一个输出向量。我们将对应于[CLS] 的输出向量用作输入序列$\mathbb{E}($表示为 $c)$ 的聚合表示，并通过 sigmoid 激活将其转换为线性层，以获得二进制分类输出$y_c$ 。
$$y_{c}=\sigma(W_{c}\boldsymbol{c}+b_{c})$$
(1)
其中$W_c$是投影矩阵，$b_c$是偏置
要训练此检索器，对于 CommonGen 训练集中的每个概念集，我们将其配对句子用作正例，并从训练集中随机抽取另一个句子作为反例。然后，我们对这种二元分类采用交叉熵损失。将选择得分最高的顶级$k$句子作为辅助输人$Z$
我们将描述如何在 CommonGen 预训练和
微调阶段中使用这两个检索器
### 2.2 Pre-training
为了增强模型的常识推理能力，我们为 RE-T5设计了一个预训练任务，该任务类似于原始的 CommonGen 任务。更详细地说，给定来自外部语料库的句子，我们首先使用 spaCy$\boxed{\mathrm{F}}$ (Honnibal et al.,2020)用词性标记句子，并提取动词，名词和专有名词作为伪概念短语。然后，我们仅在 ConceptNet$\underline{\mathrm{E}}($Speer et al. 2017)中保留短语，并删除 CommonGen 的测试集中出现的概念集。我们使用原始句子作为目标句子，并在给定伪概念的情况下构造了使用 RE-T5 生成该句子的预训练任务
由于针对伪概念的提取方法，在检索原型句子时，对于预训练数据中的每个概念集，我们都有一个很大的候选集$z$ ,候选句子数量过多。这导致使用可训练寻回器的推理时间很长。因此，由于速度的考虑以及在预训练中引人了一定程度的随机性，我们使用匹配检索器检索$k$句子作为辅助输入$Z$
|
[
[
0.313232421875,
0.07598876953125,
0.397216796875,
0.07781982421875,
0.397216796875,
0.08905029296875,
0.313232421875,
0.0872802734375
],
[
0.5224609375,
0.07659912109375,
0.8818359375,
0.07659912109375,
0.8818359375,
0.09027099609375,
0.5224609375,
0.09027099609375
],
[
0.333251953125,
0.0872802734375,
0.3837890625,
0.08905029296875,
0.382080078125,
0.1015625,
0.33154296875,
0.09979248046875
],
[
0.264404296875,
0.09442138671875,
0.297119140625,
0.09442138671875,
0.297119140625,
0.1015625,
0.264404296875,
0.1015625
],
[
0.515625,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10687255859375,
0.515625,
0.10687255859375
],
[
0.403076171875,
0.10394287109375,
0.4609375,
0.10394287109375,
0.4609375,
0.1181640625,
0.403076171875,
0.1181640625
],
[
0.5146484375,
0.1092529296875,
0.8818359375,
0.10809326171875,
0.8818359375,
0.122314453125,
0.5146484375,
0.12353515625
],
[
0.154541015625,
0.11639404296875,
0.29296875,
0.11639404296875,
0.29296875,
0.1300048828125,
0.154541015625,
0.1300048828125
],
[
0.399658203125,
0.11517333984375,
0.462646484375,
0.1175537109375,
0.4609375,
0.1318359375,
0.39794921875,
0.1300048828125
],
[
0.51708984375,
0.1270751953125,
0.73876953125,
0.1270751953125,
0.73876953125,
0.13720703125,
0.51708984375,
0.13720703125
],
[
0.183837890625,
0.13720703125,
0.2568359375,
0.13720703125,
0.2568359375,
0.1484375,
0.183837890625,
0.1484375
],
[
0.4248046875,
0.133056640625,
0.47021484375,
0.133056640625,
0.47021484375,
0.1436767578125,
0.4248046875,
0.1436767578125
],
[
0.4306640625,
0.144287109375,
0.465087890625,
0.144287109375,
0.465087890625,
0.152587890625,
0.4306640625,
0.152587890625
],
[
0.2216796875,
0.152587890625,
0.2301025390625,
0.1591796875,
0.2225341796875,
0.163330078125,
0.2149658203125,
0.157958984375
],
[
0.5205078125,
0.1484375,
0.880859375,
0.1484375,
0.880859375,
0.162109375,
0.5205078125,
0.162109375
],
[
0.4013671875,
0.157958984375,
0.46337890625,
0.157958984375,
0.46337890625,
0.169189453125,
0.4013671875,
0.169189453125
],
[
0.405517578125,
0.168701171875,
0.459228515625,
0.17041015625,
0.45751953125,
0.185302734375,
0.40478515625,
0.182861328125
],
[
0.5146484375,
0.1644287109375,
0.880859375,
0.1644287109375,
0.880859375,
0.1776123046875,
0.5146484375,
0.1776123046875
],
[
0.193115234375,
0.171630859375,
0.32568359375,
0.171630859375,
0.32568359375,
0.1859130859375,
0.193115234375,
0.1859130859375
],
[
0.515625,
0.18115234375,
0.8818359375,
0.18115234375,
0.8818359375,
0.19482421875,
0.515625,
0.19482421875
],
[
0.246826171875,
0.189453125,
0.27197265625,
0.189453125,
0.27197265625,
0.201904296875,
0.246826171875,
0.201904296875
],
[
0.5146484375,
0.19775390625,
0.8818359375,
0.19775390625,
0.8818359375,
0.21142578125,
0.5146484375,
0.21142578125
],
[
0.193115234375,
0.207275390625,
0.32568359375,
0.207275390625,
0.32568359375,
0.221435546875,
0.193115234375,
0.221435546875
],
[
0.5146484375,
0.2120361328125,
0.8818359375,
0.213134765625,
0.8818359375,
0.22802734375,
0.5146484375,
0.226806640625
],
[
0.2626953125,
0.22802734375,
0.3046875,
0.2298583984375,
0.30322265625,
0.2423095703125,
0.26123046875,
0.240478515625
],
[
0.515625,
0.2298583984375,
0.88330078125,
0.2298583984375,
0.88330078125,
0.243408203125,
0.515625,
0.243408203125
],
[
0.5146484375,
0.243408203125,
0.87744140625,
0.2452392578125,
0.87744140625,
0.260009765625,
0.5146484375,
0.25830078125
],
[
0.15283203125,
0.251708984375,
0.36279296875,
0.251708984375,
0.36279296875,
0.265380859375,
0.15283203125,
0.265380859375
],
[
0.5146484375,
0.262451171875,
0.880859375,
0.262451171875,
0.880859375,
0.275634765625,
0.5146484375,
0.275634765625
],
[
0.22412109375,
0.273681640625,
0.2880859375,
0.273681640625,
0.2880859375,
0.28076171875,
0.22412109375,
0.28076171875
],
[
0.515625,
0.276611328125,
0.580078125,
0.279052734375,
0.57861328125,
0.29345703125,
0.5146484375,
0.29150390625
],
[
0.33154296875,
0.2880859375,
0.341796875,
0.2880859375,
0.341796875,
0.295654296875,
0.33154296875,
0.295654296875
],
[
0.3046875,
0.2998046875,
0.35693359375,
0.2998046875,
0.35693359375,
0.314208984375,
0.3046875,
0.314208984375
],
[
0.6171875,
0.300537109375,
0.78271484375,
0.300537109375,
0.78271484375,
0.314208984375,
0.6171875,
0.314208984375
],
[
0.33154296875,
0.314208984375,
0.343505859375,
0.3154296875,
0.341796875,
0.325927734375,
0.330078125,
0.32470703125
],
[
0.2476806640625,
0.33203125,
0.4248046875,
0.33203125,
0.4248046875,
0.345703125,
0.2476806640625,
0.345703125
],
[
0.515625,
0.327880859375,
0.8759765625,
0.327880859375,
0.8759765625,
0.341552734375,
0.515625,
0.341552734375
],
[
0.5322265625,
0.34326171875,
0.880859375,
0.34326171875,
0.880859375,
0.35693359375,
0.5322265625,
0.35693359375
],
[
0.515625,
0.361083984375,
0.8818359375,
0.361083984375,
0.8818359375,
0.374755859375,
0.515625,
0.374755859375
],
[
0.2476806640625,
0.374755859375,
0.426513671875,
0.374755859375,
0.426513671875,
0.3896484375,
0.2476806640625,
0.3896484375
],
[
0.515625,
0.3759765625,
0.8818359375,
0.3759765625,
0.8818359375,
0.3896484375,
0.515625,
0.3896484375
],
[
0.295654296875,
0.3955078125,
0.3720703125,
0.3955078125,
0.3720703125,
0.40673828125,
0.295654296875,
0.40673828125
],
[
0.5146484375,
0.392578125,
0.880859375,
0.392578125,
0.880859375,
0.40625,
0.5146484375,
0.40625
],
[
0.5146484375,
0.407958984375,
0.8095703125,
0.4091796875,
0.8095703125,
0.42333984375,
0.5146484375,
0.4228515625
],
[
0.12005615234375,
0.421630859375,
0.489501953125,
0.421630859375,
0.489501953125,
0.435302734375,
0.12005615234375,
0.435302734375
],
[
0.623046875,
0.42333984375,
0.771484375,
0.42041015625,
0.771484375,
0.438232421875,
0.623046875,
0.441162109375
],
[
0.85791015625,
0.424560546875,
0.88330078125,
0.424560546875,
0.88330078125,
0.438232421875,
0.85791015625,
0.438232421875
],
[
0.12005615234375,
0.43701171875,
0.447509765625,
0.43701171875,
0.447509765625,
0.44775390625,
0.12005615234375,
0.44775390625
],
[
0.5146484375,
0.44482421875,
0.78271484375,
0.4453125,
0.78271484375,
0.460205078125,
0.5146484375,
0.458984375
],
[
0.5341796875,
0.4619140625,
0.88330078125,
0.4619140625,
0.88330078125,
0.4755859375,
0.5341796875,
0.4755859375
],
[
0.51708984375,
0.47802734375,
0.8818359375,
0.47802734375,
0.8818359375,
0.4912109375,
0.51708984375,
0.4912109375
],
[
0.12005615234375,
0.485107421875,
0.48779296875,
0.485107421875,
0.48779296875,
0.498779296875,
0.12005615234375,
0.498779296875
],
[
0.5146484375,
0.493408203125,
0.880859375,
0.493408203125,
0.880859375,
0.50830078125,
0.5146484375,
0.50830078125
],
[
0.12176513671875,
0.501953125,
0.48779296875,
0.501953125,
0.48779296875,
0.515625,
0.12176513671875,
0.515625
],
[
0.515625,
0.51025390625,
0.8818359375,
0.51025390625,
0.8818359375,
0.52392578125,
0.515625,
0.52392578125
],
[
0.12005615234375,
0.51708984375,
0.48779296875,
0.5166015625,
0.48779296875,
0.52978515625,
0.12005615234375,
0.53076171875
],
[
0.515625,
0.52685546875,
0.8818359375,
0.52685546875,
0.8818359375,
0.54052734375,
0.515625,
0.54052734375
],
[
0.12176513671875,
0.53369140625,
0.48779296875,
0.53369140625,
0.48779296875,
0.54736328125,
0.12176513671875,
0.54736328125
],
[
0.5146484375,
0.54345703125,
0.55517578125,
0.54150390625,
0.5556640625,
0.55419921875,
0.515625,
0.5556640625
],
[
0.12005615234375,
0.5498046875,
0.486083984375,
0.5498046875,
0.486083984375,
0.5634765625,
0.12005615234375,
0.5634765625
],
[
0.5322265625,
0.55859375,
0.8818359375,
0.55859375,
0.8818359375,
0.572265625,
0.5322265625,
0.572265625
],
[
0.12176513671875,
0.56640625,
0.48779296875,
0.56640625,
0.48779296875,
0.580078125,
0.12176513671875,
0.580078125
],
[
0.5146484375,
0.576171875,
0.75390625,
0.576171875,
0.75390625,
0.5888671875,
0.5146484375,
0.5888671875
],
[
0.12005615234375,
0.58203125,
0.48779296875,
0.58203125,
0.48779296875,
0.595703125,
0.12005615234375,
0.595703125
],
[
0.12176513671875,
0.599609375,
0.2301025390625,
0.599609375,
0.2301025390625,
0.61083984375,
0.12176513671875,
0.61083984375
],
[
0.51123046875,
0.599609375,
0.66162109375,
0.60107421875,
0.66162109375,
0.615234375,
0.51123046875,
0.6142578125
],
[
0.1368408203125,
0.615234375,
0.44921875,
0.615234375,
0.44921875,
0.62890625,
0.1368408203125,
0.62890625
],
[
0.444091796875,
0.62060546875,
0.44921875,
0.62060546875,
0.44921875,
0.6259765625,
0.444091796875,
0.6259765625
],
[
0.515625,
0.62158203125,
0.88330078125,
0.62158203125,
0.88330078125,
0.63525390625,
0.515625,
0.63525390625
],
[
0.118408203125,
0.6318359375,
0.490234375,
0.63134765625,
0.490234375,
0.6455078125,
0.118408203125,
0.646484375
],
[
0.515625,
0.63720703125,
0.8818359375,
0.63720703125,
0.8818359375,
0.65087890625,
0.515625,
0.65087890625
],
[
0.12005615234375,
0.64794921875,
0.48779296875,
0.64794921875,
0.48779296875,
0.6611328125,
0.12005615234375,
0.6611328125
],
[
0.51318359375,
0.65185546875,
0.8818359375,
0.65283203125,
0.8818359375,
0.66748046875,
0.51318359375,
0.66650390625
],
[
0.1175537109375,
0.6611328125,
0.489501953125,
0.6630859375,
0.489501953125,
0.68115234375,
0.1175537109375,
0.6787109375
],
[
0.515625,
0.66845703125,
0.8818359375,
0.66845703125,
0.8818359375,
0.681640625,
0.515625,
0.681640625
],
[
0.12005615234375,
0.6787109375,
0.48779296875,
0.6796875,
0.48779296875,
0.69482421875,
0.12005615234375,
0.693359375
],
[
0.5146484375,
0.68603515625,
0.8818359375,
0.68603515625,
0.8818359375,
0.69970703125,
0.5146484375,
0.69970703125
],
[
0.118408203125,
0.6953125,
0.267822265625,
0.69677734375,
0.267822265625,
0.71142578125,
0.118408203125,
0.71044921875
],
[
0.51708984375,
0.70361328125,
0.880859375,
0.70361328125,
0.880859375,
0.71435546875,
0.51708984375,
0.71435546875
],
[
0.5146484375,
0.71728515625,
0.87890625,
0.71728515625,
0.87890625,
0.73095703125,
0.5146484375,
0.73095703125
],
[
0.118408203125,
0.72705078125,
0.2418212890625,
0.72705078125,
0.2418212890625,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.515625,
0.73388671875,
0.8818359375,
0.73388671875,
0.8818359375,
0.7490234375,
0.515625,
0.7490234375
],
[
0.12176513671875,
0.75,
0.48779296875,
0.75,
0.48779296875,
0.76318359375,
0.12176513671875,
0.76318359375
],
[
0.515625,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76416015625,
0.515625,
0.76416015625
],
[
0.12176513671875,
0.7666015625,
0.489501953125,
0.7666015625,
0.489501953125,
0.77978515625,
0.12176513671875,
0.77978515625
],
[
0.515625,
0.76708984375,
0.8818359375,
0.76708984375,
0.8818359375,
0.78076171875,
0.515625,
0.78076171875
],
[
0.1234130859375,
0.78369140625,
0.486083984375,
0.78369140625,
0.486083984375,
0.79443359375,
0.1234130859375,
0.79443359375
],
[
0.5146484375,
0.783203125,
0.8017578125,
0.783203125,
0.8017578125,
0.79638671875,
0.5146484375,
0.79638671875
],
[
0.12005615234375,
0.798828125,
0.48779296875,
0.798828125,
0.48779296875,
0.8125,
0.12005615234375,
0.8125
],
[
0.5341796875,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.5341796875,
0.8125
],
[
0.12005615234375,
0.81396484375,
0.48779296875,
0.81396484375,
0.48779296875,
0.82763671875,
0.12005615234375,
0.82763671875
],
[
0.51708984375,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.51708984375,
0.82763671875
],
[
0.12005615234375,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.5146484375,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.84423828125,
0.5146484375,
0.84423828125
],
[
0.118408203125,
0.84716796875,
0.48779296875,
0.84716796875,
0.48779296875,
0.86083984375,
0.118408203125,
0.86083984375
],
[
0.5146484375,
0.84716796875,
0.880859375,
0.84716796875,
0.880859375,
0.86083984375,
0.5146484375,
0.86083984375
],
[
0.12176513671875,
0.86328125,
0.489501953125,
0.86328125,
0.489501953125,
0.876953125,
0.12176513671875,
0.876953125
],
[
0.515625,
0.86328125,
0.880859375,
0.86328125,
0.880859375,
0.876953125,
0.515625,
0.876953125
],
[
0.12005615234375,
0.87890625,
0.489501953125,
0.87890625,
0.489501953125,
0.8935546875,
0.12005615234375,
0.8935546875
],
[
0.5146484375,
0.87890625,
0.8818359375,
0.87744140625,
0.8818359375,
0.892578125,
0.5146484375,
0.8935546875
],
[
0.12176513671875,
0.896484375,
0.2283935546875,
0.896484375,
0.2283935546875,
0.90771484375,
0.12176513671875,
0.90771484375
],
[
0.515625,
0.8955078125,
0.74072265625,
0.8955078125,
0.74072265625,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"CommonGen",
"MatchingRetriever匹配的检索器首先根据",
"Dataset",
"ample",
"所包含概念的数量对候选句子进行排序。然",
"External",
"后，它简单地从包含最多概念的句子作为辅助",
"Input concept set (X)",
"Corpora",
"输入开始对k句子进行采样",
"Xi,X2, . Xn",
"simply",
"iselect",
"",
"TrainableRetriever为了从句子候选集中检",
"Candidate",
"Set(Z)",
"索更多有用的句子，我们设计了一个可训练的",
"Matching Retriever",
"检索器，该检索器预测分数以对这些候选进行",
"OR",
"排名，然后选择top-k句子作为附加上下文。",
"Trainable Retriever",
"评分器基于BERT(Devlin etal.，2019）构建，",
"retrieve",
"这是一种通常用于语言理解的预训练语言模",
"型。给定一个概念集X和一个候选句子zi",
"Retrieved prototype sentences (2)",
"我们的可训练检索器首先将它们连接成一个文",
"Z1-Z2.,Z",
"本输入：",
"+",
"(X,z)",
"[CLS]X[SEP]zi[SEP]",
"+",
"Retrieval Enhanced Generator",
"其中[CLS]和[SEP】是BERT中的特殊符号",
"我们将其传递给BERT，该BERT为每个",
"输入令牌生成一个输出向量。我们将对应于",
"Natural language output ()",
"[CLS]的输出向量用作输入序列E（表示为c）",
"Y1.Y2.., Ym",
"的聚合表示，并通过sigmoid激活将其转换为",
"线性层，以获得二进制分类输出yc。",
"Figure 1: The overall framework of Retrieved En-",
"Ye = a(Wec+ be)",
"(1)",
"hanced Modelfor Commonsense Generation.",
"其中Wc是投影矩阵，bc是偏置",
"要训练此检索器，对于CommonGen训练",
"集中的每个概念集，我们将其配对句子用作正",
"理受益。我们通过提出检索增强T5（RE-T5)",
"例，并从训练集中随机抽取另一个句子作为反",
"来扩展此想法，该技术为原始T5配备了可训",
"例。然后，我们对这种二元分类采用交叉摘损",
"练的检索器，用于根据给定的概念选择原型句",
"失。将选择得分最高的顶级k句子作为辅助输",
"子。同时，参考（Zhouetal.，2021），我们为",
"人Z",
"CommonGen 设计了一个预训练任务，该任务",
"我们将描述如何在CommonGen预训练和",
"继续在从外部语料库中提取的伪概念集上对",
"微调阶段中使用这两个检索器",
"RE-T5进行预训练。在这个预训练阶段，我们",
"还使用了猎犬",
"2.2Pre-training",
"形式上，给定一个概念集X",
"",
"为了增强模型的常识推理能力，我们为RE-T5",
"α1,2,...,n，其中表示第i个概念，",
"设计了一个预训练任务，该任务类似于原始",
"n是概念的数量，我们的目标是生成令牌",
"的CommonGen任务。更详细地说，给定来",
"Y={y1,y2，...，ym）的自然语言输出，该输",
"自外部语料库的句子，我们首先使用spaCvlF",
"出使用X中的所有给定概念来描述我们日常",
"(Honnibaletal.，2020）用词性标记句子，并",
"生活中的常见场景",
"提取动词，名词和专有名词作为伪概念短语。",
"然后，我们仅在ConceptNetE(Speeretal.",
"2.1 Retrieval",
"2017）中保留短语，并删除CommonGen的测",
"由于外部语料库具有大量的场景知识来描述",
"试集中出现的概念集。我们使用原始句子作为",
"概念之间的关系E(Fanetal.，2020），因此我",
"目标句子，并在给定伪概念的情况下构造了使",
"们检索与输入概念相关的句子以帮助模型执",
"用RE-T5生成该句子的预训练任务",
"行更好的常识推理。首先，给定一个输入概念",
"由于针对伪概念的提取方法，在检索原型句",
"集，我们从外部语料库中提取输入X中至少",
"子时，对于预训练数据中的每个概念集，我们",
"包含两个概念的所有句子作为候选集2。然",
"都有一个很大的候选集2，候选句子数量过",
"后，我们设计了两个检索模型，匹配检索器",
"多。这导致使用可训练寻回器的推理时间很",
"和可训练检索器，以进一步检索k原型句子",
"长。因此，由于速度的考虑以及在预训练中引",
"Z={21,22,··,z},ZZ作为RE-T5的辅",
"人了一定程度的随机性，我们使用匹配检索器",
"助输入上下文",
"检索k句子作为辅助输入Z"
] |
{
"arxiv_id": "2105.11174",
"language": "zh",
"page_idx": 1
}
|
||
2105.11174-zh-2
|
2105.11174-zh-2.jpg
|
<table>
<tbody>
<tr>
<th>Model</th>
<th>BLEU- 4</th>
<th>CIDEr</th>
<th>SPICE</th>
<th>SPICE $\langle$v1.0</th>
</tr>
<tr>
<td>GPT- 2 Radford et al. ,2019</td>
<td>26.833</td>
<td>12.187</td>
<td>23.567</td>
<td>25.90</td>
</tr>
<tr>
<td>BERT- Gen (Bao et al., 2020)</td>
<td>23.468</td>
<td>12.606</td>
<td>24.822</td>
<td>27.30</td>
</tr>
<tr>
<td>UniLM $^{\prime }$Dong et al. , 2019</td>
<td>30.616</td>
<td>14.889</td>
<td>27.429</td>
<td>30.20</td>
</tr>
<tr>
<td>BART ( Lewis et al. , 2020) </td>
<td>31.827</td>
<td>13.976</td>
<td>27.995</td>
<td>30.60</td>
</tr>
<tr>
<td>T5- base Raffel et al. , 2020</td>
<td>18.546</td>
<td>9.399</td>
<td>19.871</td>
<td>22.00</td>
</tr>
<tr>
<td>T5-large Raffel et al. , 2020</td>
<td>31.962</td>
<td>15.128</td>
<td>28.855</td>
<td>31.60</td>
</tr>
<tr>
<td>EKI- BART Fan et al., 2020</td>
<td>35.945</td>
<td>16.999</td>
<td>29.583</td>
<td>32.40</td>
</tr>
<tr>
<td>KG- BART Liu et al. , 2021</td>
<td>33.867</td>
<td>16.927</td>
<td>29.634</td>
<td>32.70</td>
</tr>
<tr>
<td>CALM( T5- base) $^{\prime }$Zhou et al. , 2021</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>33.00</td>
</tr>
<tr>
<td>RE- T5 (ours</td>
<td>40.863</td>
<td>17.663</td>
<td>31.079</td>
<td>34.30</td>
</tr>
</tbody>
</table>
Table 2: Test results on CommonGen benchmark. All results except CALM are based on the latest
human references(v1.1). v1.0 indicates evaluation with old evaluation protocol.$^2$
检索后，RE-T5 将输入概念和检索到的原型句子的串联作为输入，并将原始句子作为输出
### 2.3 Fine-tuning
在微调阶段，我们使用可训练的检索器对候选集$\mathcal{Z}$中的句子进行评分，并选择顶部的$k$句子作为附加上下文$Z$ 。与预训练类似，RE-T5将输入概念和检索到的原型句子的串联作为输人，并将原始句子作为输出。
# 3 Experiments 3.1 Experiments Settings
Baselines 我们将 RE-T5 与几种基准系统进$\begin{aligned}\text{行了比较。GPT-2,BERT-Gen,UniLM,BART}\end{aligned}$ 和 T5 是在$\overline{\mathrm{E}}($Lin et al., 2020) 中测试的预训练语言模型。它们都在 CommonGen 训练集上进行了微调，以串联的概念作为输入，以描述句子作为输出。EKI-BART$\underline{\mathrm{E}}($Fan et al. 2020)是 CommonGen 的检索和生成框架，他们使用简单的检索器来增强预训练的 BARTE 2021) 在编码器和解码器端都使用知识图增强BART, 并继续使用掩码概念令牌生成任务预训练 BART。CALM (Zhou et al., 2021) 设计了几种自监督策略，鼓励模型专注于以概念为中心的信息
Implementation Details 我们在微调中采用T5-base 作为生成模型，BERT-base 作为可训练的检索器。我们使用 HuggingfaceTransformer (Wolf et al.,2020)进行模型实现。对于预训练阶段，我们使用 AdamW 优化器 (Loshchilov and Hutter, 2019) ,初始学习率为 2e-6, 权重衰减为0。01,亚当·1e-6预热分数为 0。01。模型预训练 3个 epoch, 批量大小为 16,梯度累积 4 批。对于微调，使用AdamW 优化模型，初始学习率为 5e-5, 批量为 64, 梯度累积为 3, 预热分数为0。01,并训练了 20 个时代。同时，使用 AdamW 优化器以初始学习率 2e-5, 批量大小 64 对 BERT-base 评分器进行了优化，并对模型进行了 3 个历元的训练。对于检索到的句子数$k$ ,我们实验选择 3 。所有实验均使用具有 32GB 内存的4V100 进行3.2 Results
Dataset CommonGen 是一个基准数据集， 旨在诊断模型是否具有生成常识推理的能力 (Lin et al., 2020) 。此数据集包含用于培训/开发/测试的 32,651/993/1,497 个概念集，相应句子的数量为 67,389/4,018/7,644。我们使用 BLEU$\underline{\mathrm{E}}($Papineni et al.,2002), ROUGEE(Lin,2004) , METEOR$\underline{\mathrm{E}}($Banerjee and Lavie, 2005) , CIDEr$\boxed{\mathrm{E}}($Vedantam et al., 2015)和 SPICEE(Anderson et al., 2016) 作为评估指标。由于 SPICE 与人类评估$\operatorname{E}(\operatorname{Lin})$ et al., 2020) 的相关性最大，因此我们将 SPICE 作为主要指标
External Corpora 为了与 CommonGen 数据集的分布保持一致，我们使用 VATEXE (Wang et al., 2019) , Activity$\boxed{\mathrm{E}}($Krishna et al., 2017),SNLIE(Bowman et al.,2015)和MNLIE(Williams et al., 2018) 作为外部语料库。我们从这些语料库中抽取 500k 个句子来构建我们的预训练数据集。同时，这些数据集也被用作我们检索模块的句子池。对于预训练和微调，CommonGen 目标中出现的所有句子都不用作检索句子候选
$^2$https://inklab.usc.edu/CommonGen/leaderboard.
表 2显示了 CommonGen 测试集上不同方法的结果。RE-T5 在所有指标上都大大优于以前的所有方法，并设置了新的技术水平。RE-T5将预训练语言模型的生成灵活性与基于检索
html
|
[
[
0.173828125,
0.0748291015625,
0.231689453125,
0.0748291015625,
0.231689453125,
0.08905029296875,
0.173828125,
0.08905029296875
],
[
0.57080078125,
0.0748291015625,
0.62890625,
0.0748291015625,
0.62890625,
0.08905029296875,
0.57080078125,
0.08905029296875
],
[
0.64404296875,
0.0748291015625,
0.70361328125,
0.0748291015625,
0.70361328125,
0.08905029296875,
0.64404296875,
0.08905029296875
],
[
0.7197265625,
0.0748291015625,
0.82958984375,
0.0748291015625,
0.82958984375,
0.08905029296875,
0.7197265625,
0.08905029296875
],
[
0.57275390625,
0.09027099609375,
0.62744140625,
0.09027099609375,
0.62744140625,
0.105712890625,
0.57275390625,
0.105712890625
],
[
0.48779296875,
0.07598876953125,
0.55517578125,
0.07598876953125,
0.55517578125,
0.0872802734375,
0.48779296875,
0.0872802734375
],
[
0.17626953125,
0.0914306640625,
0.411376953125,
0.0914306640625,
0.411376953125,
0.1051025390625,
0.17626953125,
0.1051025390625
],
[
0.491943359375,
0.0914306640625,
0.54833984375,
0.0914306640625,
0.54833984375,
0.105712890625,
0.491943359375,
0.105712890625
],
[
0.6455078125,
0.0914306640625,
0.7021484375,
0.0914306640625,
0.7021484375,
0.105712890625,
0.6455078125,
0.105712890625
],
[
0.75,
0.0914306640625,
0.7958984375,
0.0914306640625,
0.7958984375,
0.105712890625,
0.75,
0.105712890625
],
[
0.17626953125,
0.105712890625,
0.4130859375,
0.10687255859375,
0.4130859375,
0.1217041015625,
0.17626953125,
0.12054443359375
],
[
0.491943359375,
0.10687255859375,
0.54833984375,
0.10687255859375,
0.54833984375,
0.1217041015625,
0.491943359375,
0.1217041015625
],
[
0.57275390625,
0.10687255859375,
0.62890625,
0.10687255859375,
0.62890625,
0.1217041015625,
0.57275390625,
0.1217041015625
],
[
0.6455078125,
0.10687255859375,
0.7021484375,
0.10687255859375,
0.7021484375,
0.1217041015625,
0.6455078125,
0.1217041015625
],
[
0.75,
0.10687255859375,
0.796875,
0.10687255859375,
0.796875,
0.1217041015625,
0.75,
0.1217041015625
],
[
0.491943359375,
0.122314453125,
0.54833984375,
0.122314453125,
0.54833984375,
0.13720703125,
0.491943359375,
0.13720703125
],
[
0.57275390625,
0.122314453125,
0.62744140625,
0.122314453125,
0.62744140625,
0.13720703125,
0.57275390625,
0.13720703125
],
[
0.6455078125,
0.122314453125,
0.7021484375,
0.122314453125,
0.7021484375,
0.1383056640625,
0.6455078125,
0.1383056640625
],
[
0.75,
0.12353515625,
0.796875,
0.12353515625,
0.796875,
0.1383056640625,
0.75,
0.1383056640625
],
[
0.6455078125,
0.13720703125,
0.7021484375,
0.1395263671875,
0.7001953125,
0.155029296875,
0.64404296875,
0.152587890625
],
[
0.177978515625,
0.12353515625,
0.39208984375,
0.12353515625,
0.39208984375,
0.1383056640625,
0.177978515625,
0.1383056640625
],
[
0.490234375,
0.1395263671875,
0.54833984375,
0.1395263671875,
0.54833984375,
0.15380859375,
0.490234375,
0.15380859375
],
[
0.57275390625,
0.1395263671875,
0.62744140625,
0.1395263671875,
0.62744140625,
0.15380859375,
0.57275390625,
0.15380859375
],
[
0.75,
0.1395263671875,
0.796875,
0.1395263671875,
0.796875,
0.155029296875,
0.75,
0.155029296875
],
[
0.177978515625,
0.155029296875,
0.4013671875,
0.155029296875,
0.4013671875,
0.168701171875,
0.177978515625,
0.168701171875
],
[
0.177978515625,
0.14013671875,
0.3896484375,
0.14013671875,
0.3896484375,
0.15380859375,
0.177978515625,
0.15380859375
],
[
0.491943359375,
0.1561279296875,
0.54833984375,
0.1561279296875,
0.54833984375,
0.17041015625,
0.491943359375,
0.17041015625
],
[
0.5751953125,
0.1561279296875,
0.623046875,
0.1561279296875,
0.623046875,
0.17041015625,
0.5751953125,
0.17041015625
],
[
0.6455078125,
0.1561279296875,
0.7021484375,
0.1561279296875,
0.7021484375,
0.17041015625,
0.6455078125,
0.17041015625
],
[
0.75,
0.1561279296875,
0.796875,
0.1561279296875,
0.796875,
0.17041015625,
0.75,
0.17041015625
],
[
0.177978515625,
0.17041015625,
0.40478515625,
0.17041015625,
0.40478515625,
0.185302734375,
0.177978515625,
0.185302734375
],
[
0.491943359375,
0.171630859375,
0.54833984375,
0.171630859375,
0.54833984375,
0.1859130859375,
0.491943359375,
0.1859130859375
],
[
0.57275390625,
0.171630859375,
0.62744140625,
0.171630859375,
0.62744140625,
0.1859130859375,
0.57275390625,
0.1859130859375
],
[
0.6455078125,
0.171630859375,
0.70361328125,
0.171630859375,
0.70361328125,
0.1859130859375,
0.6455078125,
0.1859130859375
],
[
0.75,
0.171630859375,
0.796875,
0.171630859375,
0.796875,
0.1859130859375,
0.75,
0.1859130859375
],
[
0.177978515625,
0.18701171875,
0.4130859375,
0.18701171875,
0.4130859375,
0.20068359375,
0.177978515625,
0.20068359375
],
[
0.491943359375,
0.18701171875,
0.54833984375,
0.18701171875,
0.54833984375,
0.203125,
0.491943359375,
0.203125
],
[
0.57275390625,
0.188232421875,
0.62744140625,
0.188232421875,
0.62744140625,
0.203125,
0.57275390625,
0.203125
],
[
0.6455078125,
0.188232421875,
0.70361328125,
0.188232421875,
0.70361328125,
0.203125,
0.6455078125,
0.203125
],
[
0.75,
0.1859130859375,
0.796875,
0.188232421875,
0.7958984375,
0.2037353515625,
0.748046875,
0.201904296875
],
[
0.177978515625,
0.2037353515625,
0.40478515625,
0.2037353515625,
0.40478515625,
0.21728515625,
0.177978515625,
0.21728515625
],
[
0.491943359375,
0.2037353515625,
0.54833984375,
0.2037353515625,
0.54833984375,
0.218505859375,
0.491943359375,
0.218505859375
],
[
0.57275390625,
0.2037353515625,
0.62744140625,
0.2037353515625,
0.62744140625,
0.218505859375,
0.57275390625,
0.218505859375
],
[
0.6455078125,
0.2037353515625,
0.7021484375,
0.2037353515625,
0.7021484375,
0.218505859375,
0.6455078125,
0.218505859375
],
[
0.75,
0.203125,
0.796875,
0.203125,
0.796875,
0.218505859375,
0.75,
0.218505859375
],
[
0.17626953125,
0.2203369140625,
0.466796875,
0.2203369140625,
0.466796875,
0.2340087890625,
0.17626953125,
0.2340087890625
],
[
0.515625,
0.221435546875,
0.5263671875,
0.2244873046875,
0.5224609375,
0.232177734375,
0.51123046875,
0.2286376953125
],
[
0.75,
0.2197265625,
0.796875,
0.2197265625,
0.796875,
0.235107421875,
0.75,
0.235107421875
],
[
0.17626953125,
0.2381591796875,
0.287109375,
0.2381591796875,
0.287109375,
0.2493896484375,
0.17626953125,
0.2493896484375
],
[
0.49365234375,
0.2381591796875,
0.54833984375,
0.2381591796875,
0.54833984375,
0.2493896484375,
0.49365234375,
0.2493896484375
],
[
0.57421875,
0.2381591796875,
0.63037109375,
0.2381591796875,
0.63037109375,
0.2493896484375,
0.57421875,
0.2493896484375
],
[
0.646484375,
0.2381591796875,
0.7001953125,
0.2381591796875,
0.7001953125,
0.2493896484375,
0.646484375,
0.2493896484375
],
[
0.748046875,
0.2369384765625,
0.796875,
0.2369384765625,
0.796875,
0.251708984375,
0.748046875,
0.251708984375
],
[
0.118408203125,
0.26416015625,
0.88330078125,
0.26416015625,
0.88330078125,
0.27783203125,
0.118408203125,
0.27783203125
],
[
0.6337890625,
0.2410888671875,
0.64208984375,
0.2410888671875,
0.64208984375,
0.2452392578125,
0.6337890625,
0.2452392578125
],
[
0.12005615234375,
0.279052734375,
0.71728515625,
0.279052734375,
0.71728515625,
0.292236328125,
0.12005615234375,
0.292236328125
],
[
0.1376953125,
0.317626953125,
0.486083984375,
0.317626953125,
0.486083984375,
0.33203125,
0.1376953125,
0.33203125
],
[
0.5205078125,
0.318359375,
0.8818359375,
0.318359375,
0.8818359375,
0.33203125,
0.5205078125,
0.33203125
],
[
0.118408203125,
0.334228515625,
0.486083984375,
0.334228515625,
0.486083984375,
0.347412109375,
0.118408203125,
0.347412109375
],
[
0.5146484375,
0.334228515625,
0.8818359375,
0.333251953125,
0.8818359375,
0.347412109375,
0.5146484375,
0.3486328125
],
[
0.515625,
0.349853515625,
0.8818359375,
0.350830078125,
0.8818359375,
0.365234375,
0.515625,
0.364013671875
],
[
0.118408203125,
0.361083984375,
0.2626953125,
0.363525390625,
0.2626953125,
0.377685546875,
0.118408203125,
0.3759765625
],
[
0.5146484375,
0.366455078125,
0.8818359375,
0.36767578125,
0.8818359375,
0.3818359375,
0.5146484375,
0.38134765625
],
[
0.12005615234375,
0.385498046875,
0.48779296875,
0.385498046875,
0.48779296875,
0.3984375,
0.12005615234375,
0.3984375
],
[
0.515625,
0.383056640625,
0.8818359375,
0.38427734375,
0.8818359375,
0.3984375,
0.515625,
0.39794921875
],
[
0.12005615234375,
0.40087890625,
0.48779296875,
0.40087890625,
0.48779296875,
0.41455078125,
0.12005615234375,
0.41455078125
],
[
0.5146484375,
0.399658203125,
0.880859375,
0.399658203125,
0.880859375,
0.413330078125,
0.5146484375,
0.413330078125
],
[
0.12176513671875,
0.416259765625,
0.489501953125,
0.416259765625,
0.489501953125,
0.429931640625,
0.12176513671875,
0.429931640625
],
[
0.51318359375,
0.41455078125,
0.8818359375,
0.4150390625,
0.8818359375,
0.429931640625,
0.51318359375,
0.4287109375
],
[
0.12005615234375,
0.432861328125,
0.48779296875,
0.432861328125,
0.48779296875,
0.446533203125,
0.12005615234375,
0.446533203125
],
[
0.5146484375,
0.431640625,
0.8818359375,
0.431640625,
0.8818359375,
0.4453125,
0.5146484375,
0.4453125
],
[
0.12005615234375,
0.44775390625,
0.34765625,
0.448974609375,
0.34765625,
0.463134765625,
0.12005615234375,
0.4619140625
],
[
0.5146484375,
0.446533203125,
0.880859375,
0.44775390625,
0.880859375,
0.4619140625,
0.5146484375,
0.46142578125
],
[
0.5146484375,
0.4619140625,
0.8818359375,
0.463134765625,
0.8818359375,
0.47802734375,
0.5146484375,
0.476806640625
],
[
0.1175537109375,
0.47021484375,
0.268798828125,
0.47021484375,
0.268798828125,
0.485107421875,
0.1175537109375,
0.485107421875
],
[
0.5146484375,
0.478515625,
0.8818359375,
0.48095703125,
0.8818359375,
0.49462890625,
0.5146484375,
0.4921875
],
[
0.1158447265625,
0.48681640625,
0.341796875,
0.488037109375,
0.341796875,
0.505859375,
0.1158447265625,
0.5048828125
],
[
0.515625,
0.49462890625,
0.8818359375,
0.495361328125,
0.8818359375,
0.51025390625,
0.515625,
0.5087890625
],
[
0.1259765625,
0.51123046875,
0.484375,
0.51123046875,
0.484375,
0.52490234375,
0.1259765625,
0.52490234375
],
[
0.515625,
0.51123046875,
0.8818359375,
0.5126953125,
0.8818359375,
0.52685546875,
0.515625,
0.525390625
],
[
0.12005615234375,
0.52685546875,
0.48779296875,
0.525390625,
0.48779296875,
0.54052734375,
0.12005615234375,
0.54150390625
],
[
0.5146484375,
0.52685546875,
0.60791015625,
0.529296875,
0.60791015625,
0.54345703125,
0.5146484375,
0.54150390625
],
[
0.118408203125,
0.5419921875,
0.486083984375,
0.5419921875,
0.486083984375,
0.5556640625,
0.118408203125,
0.5556640625
],
[
0.5205078125,
0.54736328125,
0.8818359375,
0.54736328125,
0.8818359375,
0.56103515625,
0.5205078125,
0.56103515625
],
[
0.118408203125,
0.55810546875,
0.489501953125,
0.55859375,
0.489501953125,
0.57373046875,
0.118408203125,
0.572265625
],
[
0.5146484375,
0.5634765625,
0.8818359375,
0.5634765625,
0.8818359375,
0.57763671875,
0.5146484375,
0.57763671875
],
[
0.118408203125,
0.576171875,
0.490234375,
0.57470703125,
0.490234375,
0.5888671875,
0.118408203125,
0.59033203125
],
[
0.515625,
0.580078125,
0.8818359375,
0.580078125,
0.8818359375,
0.59326171875,
0.515625,
0.59326171875
],
[
0.12005615234375,
0.59130859375,
0.47265625,
0.59130859375,
0.47265625,
0.60546875,
0.12005615234375,
0.60546875
],
[
0.474365234375,
0.5986328125,
0.48779296875,
0.5986328125,
0.48779296875,
0.6025390625,
0.474365234375,
0.6025390625
],
[
0.5146484375,
0.595703125,
0.880859375,
0.595703125,
0.880859375,
0.609375,
0.5146484375,
0.609375
],
[
0.118408203125,
0.60693359375,
0.489501953125,
0.60791015625,
0.489501953125,
0.62255859375,
0.118408203125,
0.62158203125
],
[
0.5146484375,
0.61083984375,
0.8818359375,
0.60986328125,
0.8818359375,
0.62451171875,
0.5146484375,
0.6259765625
],
[
0.118408203125,
0.62353515625,
0.490234375,
0.62255859375,
0.490234375,
0.63720703125,
0.118408203125,
0.63818359375
],
[
0.515625,
0.62744140625,
0.8818359375,
0.62744140625,
0.8818359375,
0.64111328125,
0.515625,
0.64111328125
],
[
0.12005615234375,
0.64013671875,
0.489501953125,
0.64013671875,
0.489501953125,
0.65380859375,
0.12005615234375,
0.65380859375
],
[
0.515625,
0.6435546875,
0.8818359375,
0.6435546875,
0.8818359375,
0.65673828125,
0.515625,
0.65673828125
],
[
0.12005615234375,
0.65625,
0.489501953125,
0.65625,
0.489501953125,
0.66943359375,
0.12005615234375,
0.66943359375
],
[
0.515625,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.67333984375,
0.515625,
0.67333984375
],
[
0.12005615234375,
0.6728515625,
0.489501953125,
0.6728515625,
0.489501953125,
0.68603515625,
0.12005615234375,
0.68603515625
],
[
0.51708984375,
0.67578125,
0.8818359375,
0.67578125,
0.8818359375,
0.689453125,
0.51708984375,
0.689453125
],
[
0.118408203125,
0.68701171875,
0.2333984375,
0.6884765625,
0.2333984375,
0.703125,
0.118408203125,
0.70166015625
],
[
0.515625,
0.6923828125,
0.8818359375,
0.6923828125,
0.8818359375,
0.70703125,
0.515625,
0.70703125
],
[
0.1259765625,
0.71435546875,
0.489501953125,
0.71435546875,
0.489501953125,
0.72802734375,
0.1259765625,
0.72802734375
],
[
0.515625,
0.708984375,
0.8818359375,
0.708984375,
0.8818359375,
0.72265625,
0.515625,
0.72265625
],
[
0.118408203125,
0.72998046875,
0.489501953125,
0.728515625,
0.489501953125,
0.74365234375,
0.118408203125,
0.74462890625
],
[
0.515625,
0.724609375,
0.8818359375,
0.724609375,
0.8818359375,
0.73828125,
0.515625,
0.73828125
],
[
0.5146484375,
0.7392578125,
0.8818359375,
0.74072265625,
0.8818359375,
0.7548828125,
0.5146484375,
0.75439453125
],
[
0.12005615234375,
0.74658203125,
0.48779296875,
0.74658203125,
0.48779296875,
0.76025390625,
0.12005615234375,
0.76025390625
],
[
0.12176513671875,
0.7626953125,
0.48779296875,
0.7626953125,
0.48779296875,
0.77685546875,
0.12176513671875,
0.77685546875
],
[
0.515625,
0.75732421875,
0.8818359375,
0.75732421875,
0.8818359375,
0.77099609375,
0.515625,
0.77099609375
],
[
0.5146484375,
0.771484375,
0.8818359375,
0.771484375,
0.8818359375,
0.78515625,
0.5146484375,
0.78515625
],
[
0.12005615234375,
0.77783203125,
0.489501953125,
0.77783203125,
0.489501953125,
0.7919921875,
0.12005615234375,
0.7919921875
],
[
0.51318359375,
0.7880859375,
0.880859375,
0.7880859375,
0.880859375,
0.8017578125,
0.51318359375,
0.8017578125
],
[
0.12176513671875,
0.79443359375,
0.48779296875,
0.79443359375,
0.48779296875,
0.80810546875,
0.12176513671875,
0.80810546875
],
[
0.12005615234375,
0.81005859375,
0.48779296875,
0.81005859375,
0.48779296875,
0.82470703125,
0.12005615234375,
0.82470703125
],
[
0.5146484375,
0.8046875,
0.611328125,
0.8046875,
0.611328125,
0.8193359375,
0.5146484375,
0.8193359375
],
[
0.12176513671875,
0.82666015625,
0.489501953125,
0.82666015625,
0.489501953125,
0.84033203125,
0.12176513671875,
0.84033203125
],
[
0.51318359375,
0.82470703125,
0.62646484375,
0.82470703125,
0.62646484375,
0.8388671875,
0.51318359375,
0.8388671875
],
[
0.12176513671875,
0.84326171875,
0.48779296875,
0.84326171875,
0.48779296875,
0.85693359375,
0.12176513671875,
0.85693359375
],
[
0.515625,
0.8466796875,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.515625,
0.86083984375
],
[
0.118408203125,
0.85986328125,
0.30322265625,
0.85986328125,
0.30322265625,
0.87353515625,
0.118408203125,
0.87353515625
],
[
0.515625,
0.86328125,
0.8818359375,
0.86328125,
0.8818359375,
0.876953125,
0.515625,
0.876953125
],
[
0.1368408203125,
0.8818359375,
0.48779296875,
0.8828125,
0.48779296875,
0.89794921875,
0.1368408203125,
0.896484375
],
[
0.5146484375,
0.87890625,
0.88330078125,
0.87744140625,
0.88330078125,
0.892578125,
0.5146484375,
0.8935546875
],
[
0.12005615234375,
0.896484375,
0.154541015625,
0.896484375,
0.154541015625,
0.90771484375,
0.12005615234375,
0.90771484375
],
[
0.515625,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"Model",
"CIDEr",
"SPICE",
"SPICE(v1.0)",
"12.187",
"BLEU-4",
"GPT-2 (Radford et al., 2019)",
"26.833",
"23.567",
"25.90",
"BERT-Gen (Bao et al., 2020)",
"23.468",
"12.606",
"24.822",
"27.30",
"30.616",
"14.889",
"27.429",
"30.20",
"27.995",
"UniLM (Dong et al., 2019)",
"31.827",
"13.976",
"30.60",
"T5-base (Raffel et al., 2020)",
"BART (Lewis et al., 2020)",
"18.546",
"9.399",
"19.871",
"22.00",
"T5-large (Raffel et al., 2020)",
"31.962",
"15.128",
"28.855",
"31.60",
"EKI-BART (Fan et al.. 2020)",
"35.945",
"16.999",
"29.583",
"32.40",
"KG-BART (Liu et al., 2021)",
"33.867",
"16.927",
"29.634",
"32.70",
"CALM(T5-base) (Zhou et al., 2021)",
".",
"33.00",
"RE-T5 (ours)",
"40.863",
"17.663",
"31.079",
"34.30",
"Table 2: Test results on CommonGen benchmark. All results except CALM are based on the latest",
"",
"human references(vl.1). v1.0 indicates evaluation with old evaluation protocol. 2",
"检索后，RE-T5将输入概念和检索到的原型",
"Baselines我们将RE-T5与几种基准系统进",
"句子的串联作为输入，并将原始句子作为输出",
"行了比较。GPT-2,BERT-Gen,UniLM,BART",
"和T5是在E(Linetal.,2020）中测试的预训",
"2.3Fine-tuning",
"练语言模型。它们都在CommonGen训练集",
"在微调阶段，我们使用可训练的检索器对候选",
"上进行了微调，以串联的概念作为输人，以",
"集2中的句子进行评分，并选择顶部的k句",
"描述句子作为输出。EKI-BARTE(Fan etal.",
"子作为附加上下文Z。与预训练类似，RE-T5",
"2020）是CommonGen的检索和生成框架，他",
"将输入概念和检索到的原型句子的串联作为输",
"们使用简单的检索器来增强预训练的BARTE",
"人，并将原始句子作为输出。",
"(Lewis et al., 2020) 。KG-BART (Liu et al.,",
"2021）在编码器和解码器端都使用知识图增强",
"3Experiments",
"BART，并继续使用掩码概念令牌生成任务预",
"3.1 Experiments Settings",
"训练BART。CALM(Zhouetal.,2021)设计",
"DatasetCommonGen是一个基准数据集",
"了几种自监督策略，鼓励模型专注于以概念为",
"旨在诊断模型是否具有生成常识推理的能",
"中心的信息",
"力（Linetal.，2020）。此数据集包含用于",
"ImplementationDetails我们在微调中采用",
"培训/开发/测试的32,651/993/1,497个概念",
"T5-base作为生成模型，BERT-base作为可",
"集，相应句子的数量为67,389/4,018/7.644。",
"训练的检索器。我们使用HuggingfaceTrans-",
"我们使用BLEUE(Papineni et al.，2002)",
"",
"former（Wolfetal.，2020）进行模型实现。",
"ROUGEE(Lin,2004)METEORE(Banerjee",
"对于预训练阶段，我们使用AdamW优化",
"and Lavie, 2005) , CIDErE(Vedantam et al.,",
"器（LoshchilovandHutter，2019），初始学",
"2015)和SPICEE(Andersonetal.,2016)作",
"习率为2e-6，权重衰减为0。01，亚当：1e-6",
"为评估指标。由于SPICE与人类评估E(Lin",
"预热分数为0。01。模型预训练3个epoch，批",
"etal.,2020)的相关性最大，因此我们将SPICE",
"量大小为16，梯度累积4批。对于微调，使用",
"作为主要指标",
"AdamW优化模型，初始学习率为5e-5，批量",
"External Corpora为了与 CommonGen 数",
"为64，梯度累积为3，预热分数为0。01，并训",
"据集的分布保持一致，我们使用VATEXE",
"练了20个时代。同时，使用AdamW优化器以",
"初始学习率2e-5，批量大小64对BERT-base",
"(Wang et al., 2019) , ActivityE(Krishna et al.,",
"2017)，SNLIE(Bowman etal.,2015)和",
"评分器进行了优化，并对模型进行了3个历元",
"的训练。对于检索到的句子数k，我们实验",
"MNLIE（Williamsetal.，2018）作为外部语料",
"选择3。所有实验均使用具有32GB内存的",
"库。我们从这些语料库中抽取500k个句子来",
"构建我们的预训练数据集。同时，这些数据集",
"4V100进行",
"也被用作我们检索模块的句子池。对于预训练",
"3.2 Results",
"和微调，CommonGen目标中出现的所有句子",
"表2显示了CommonGen测试集上不同方法",
"都不用作检索句子候选",
"的结果。RE-T5在所有指标上都大大优于以前",
"2https://inklab.usc.edu/CommonGen/leaderboard.",
"的所有方法，并设置了新的技术水平。RE-T5",
"html",
"将预训练语言模型的生成灵活性与基于检索"
] |
{
"arxiv_id": "2105.11174",
"language": "zh",
"page_idx": 2
}
|
||
2210.15851-zh-1
|
2210.15851-zh-1.jpg
|
# 使用
通用表示和交叉映射改进零快照多语言翻译
Shuhao Gu$^{1,2},\textbf{Yang Feng}^{1,2*}$
$^{1}$ Key Laboratory of Intelligent Information Processing,
Institute of Computing Technology, Chinese Academy of Sciences (ICT/CAS)
$^{\hat{2}}$University of Chinese Academy of Sciences
$\{$ gushuhao19b,fengyang $\}@$ict.ac.cn
### $\mathbf{Abstract}$
The many-to-many multilingual neural machine translation can translate between language pairs unseen during training, i.e., zeroshot translation. Improving zero-shot translation requires the model to learn universal representations and cross-mapping relationships to transfer the knowledge learned on the supervised directions to the zero-shot directions. In this work, we propose the state mover's distance based on the optimal theory to model the difference of the representations output by the encoder. Then, we bridge the gap between the semantic-equivalent representations of different languages at the token level by minimizing the proposed distance to learn universal representations. Besides, we propose an agreement-based training scheme, which can help the model make consistent predictions based on the semantic-equivalent sentences to learn universal cross-mapping relationships for all translation directions. The experimental results on diverse multilingual datasets show that our method can improve consistently compared with the baseline system and other contrast methods. The analysis proves that our method can better align the semantic space and improve the prediction consistency.
系，这些特征和关系可以将不同语言的语义等效句子映射到目标语言的特定空间。通过这种方式，该模型可以将在监督翻译方向上学习的知识转移到零拍翻译方向。然而，现有的模型结构和训练方案不能保证通用表示和交叉映射，因为缺乏明确的约束。具体地，编码器可以将不同的语言映射到不同的语义子空间，解码器可以针对不同的源语言学习不同的映射关系，尤其是当模型具有高容量时
许多研究人员已经尝试解决这个问题。Pham et al.(2019) 建议将编码器的输出压缩为一致数量的状态，以仅编码与语言无关的功能。Arivazhagan et al. (2019) 添加正则化损失，以最大化源句子和目标句子的句子表示之间的相似性。Pan et al.(2021)提出了对比学习方案，以最大程度地减少相似句子的句子表示差距，并最大程度地减少无关句子的句子表示差距。上述所有工作都试图最大程度地减少句子级别上不同语言的表示差异，这给 NMT 带来了两个问题。首先，这些工作通常通过最大池化或平均获得编码器输出的句子级表示，这可能会忽略句子长度，单词对齐关系和其他令牌级信息。其次，正则化句子表示与 NMT 模型的工作范式不匹配，因为解码器直接对整个状态序列而不是句子表示进行交叉关注。此外，上述所有工作都集中在编码器端，无法帮助学习解码器的通用映射关系
鉴于上述情况，我们提出了一种学习通用表示和交叉映射的方法，以提高零拍翻译性能。基于最优运输理论，我们提出了状态移动器的距离 (SMD),以在令牌级别对两个状态序列的差异进行建模。为了将来自不同语言的语义等效句子映射到语义空间的同一位置，我们添加了一个辅助损失，以最小化源句子和目标句子的 SMD。此外，我们提出了一种基于协议的培训方案，以学习具有相同目标语言的翻译方向的通用映射关系。我们将源句子和目标句子混合在一起，得到一个伪句子。然后，解码器分别以该伪句子和相应的源句子或目标句子为条件进行预测。我们试图通过最小化两个输出分布的 KL 散度来提高预测的一致性。在不同
### 1 $\textbf{Introduction}$
多对多多多语言神经机器翻译(NMT)E(Ha et al., 2016; Firat et al., 2016; Johnson et al., 2017; Gu et al., 2018; Fan et al., 2020; Zhang et al., 2020a) 模型可以在单个模型中支持多个翻译方向。共享编码器将输入的句子编码到语义空间，然后共享解码器从该空间解码以生成目标语言的翻译。这种范式允许模型在训练期间看不见的语言对之间进行翻译，即。e。,零拍翻译。零镜头翻译可以提高推理效率，使模型需要更少的双语训练数据。执行零翻译需要通用表示来编码与语言无关的特征和交叉映射关
$^*$Corresponding author: Yang Feng.
Reproducible code:
https://github.com/ictnlp/Zero-
MNMT.
|
[
[
0.47607421875,
0.0926513671875,
0.52294921875,
0.0926513671875,
0.52294921875,
0.10809326171875,
0.47607421875,
0.10809326171875
],
[
0.27197265625,
0.1134033203125,
0.724609375,
0.1134033203125,
0.724609375,
0.127685546875,
0.27197265625,
0.127685546875
],
[
0.382080078125,
0.1436767578125,
0.62646484375,
0.142578125,
0.62646484375,
0.1602783203125,
0.382080078125,
0.160888671875
],
[
0.276123046875,
0.1591796875,
0.7255859375,
0.160888671875,
0.7255859375,
0.1787109375,
0.276123046875,
0.177001953125
],
[
0.185546875,
0.1787109375,
0.81689453125,
0.1787109375,
0.81689453125,
0.193603515625,
0.185546875,
0.193603515625
],
[
0.313232421875,
0.191162109375,
0.68701171875,
0.19482421875,
0.68701171875,
0.2120361328125,
0.313232421875,
0.208984375
],
[
0.35107421875,
0.2120361328125,
0.6572265625,
0.2120361328125,
0.6572265625,
0.226806640625,
0.35107421875,
0.226806640625
],
[
0.2626953125,
0.254150390625,
0.341796875,
0.254150390625,
0.341796875,
0.268310546875,
0.2626953125,
0.268310546875
],
[
0.5146484375,
0.254638671875,
0.880859375,
0.254638671875,
0.880859375,
0.268310546875,
0.5146484375,
0.268310546875
],
[
0.51318359375,
0.26953125,
0.87890625,
0.26953125,
0.87890625,
0.283203125,
0.51318359375,
0.283203125
],
[
0.1435546875,
0.279052734375,
0.459228515625,
0.27783203125,
0.459228515625,
0.292236328125,
0.1435546875,
0.29345703125
],
[
0.515625,
0.287353515625,
0.87890625,
0.287353515625,
0.87890625,
0.300537109375,
0.515625,
0.300537109375
],
[
0.14697265625,
0.29443359375,
0.455078125,
0.29443359375,
0.455078125,
0.3046875,
0.14697265625,
0.3046875
],
[
0.14697265625,
0.30810546875,
0.456787109375,
0.30810546875,
0.456787109375,
0.32177734375,
0.14697265625,
0.32177734375
],
[
0.51123046875,
0.302734375,
0.87890625,
0.3017578125,
0.87890625,
0.3154296875,
0.51123046875,
0.31640625
],
[
0.14697265625,
0.322509765625,
0.459228515625,
0.322509765625,
0.459228515625,
0.336181640625,
0.14697265625,
0.336181640625
],
[
0.51318359375,
0.319580078125,
0.880859375,
0.319580078125,
0.880859375,
0.333251953125,
0.51318359375,
0.333251953125
],
[
0.145263671875,
0.336181640625,
0.459228515625,
0.33740234375,
0.459228515625,
0.350830078125,
0.145263671875,
0.349853515625
],
[
0.51318359375,
0.336181640625,
0.875,
0.336181640625,
0.875,
0.3466796875,
0.51318359375,
0.3466796875
],
[
0.14697265625,
0.3515625,
0.45751953125,
0.3515625,
0.45751953125,
0.365234375,
0.14697265625,
0.365234375
],
[
0.51318359375,
0.350830078125,
0.87890625,
0.350830078125,
0.87890625,
0.364013671875,
0.51318359375,
0.364013671875
],
[
0.145263671875,
0.364013671875,
0.453369140625,
0.366455078125,
0.453369140625,
0.380126953125,
0.145263671875,
0.377685546875
],
[
0.5146484375,
0.3681640625,
0.880859375,
0.3681640625,
0.880859375,
0.3818359375,
0.5146484375,
0.3818359375
],
[
0.1485595703125,
0.38134765625,
0.456787109375,
0.38134765625,
0.456787109375,
0.391357421875,
0.1485595703125,
0.391357421875
],
[
0.5146484375,
0.38427734375,
0.76171875,
0.38427734375,
0.76171875,
0.39794921875,
0.5146484375,
0.39794921875
],
[
0.1485595703125,
0.3955078125,
0.456787109375,
0.3955078125,
0.456787109375,
0.40625,
0.1485595703125,
0.40625
],
[
0.14697265625,
0.407958984375,
0.45751953125,
0.407958984375,
0.45751953125,
0.421630859375,
0.14697265625,
0.421630859375
],
[
0.5322265625,
0.402587890625,
0.880859375,
0.402587890625,
0.880859375,
0.416259765625,
0.5322265625,
0.416259765625
],
[
0.14697265625,
0.4228515625,
0.456787109375,
0.4228515625,
0.456787109375,
0.435791015625,
0.14697265625,
0.435791015625
],
[
0.51318359375,
0.419189453125,
0.880859375,
0.419189453125,
0.880859375,
0.432861328125,
0.51318359375,
0.432861328125
],
[
0.145263671875,
0.435302734375,
0.456787109375,
0.435791015625,
0.456787109375,
0.449462890625,
0.145263671875,
0.448974609375
],
[
0.51318359375,
0.43701171875,
0.87744140625,
0.43701171875,
0.87744140625,
0.44775390625,
0.51318359375,
0.44775390625
],
[
0.14697265625,
0.45068359375,
0.45751953125,
0.45068359375,
0.45751953125,
0.46435546875,
0.14697265625,
0.46435546875
],
[
0.5146484375,
0.451904296875,
0.880859375,
0.451904296875,
0.880859375,
0.465576171875,
0.5146484375,
0.465576171875
],
[
0.1435546875,
0.46435546875,
0.459228515625,
0.463134765625,
0.459228515625,
0.47802734375,
0.1435546875,
0.478515625
],
[
0.51318359375,
0.466064453125,
0.880859375,
0.46728515625,
0.880859375,
0.482177734375,
0.51318359375,
0.48095703125
],
[
0.145263671875,
0.478515625,
0.456787109375,
0.47802734375,
0.456787109375,
0.4912109375,
0.145263671875,
0.4921875
],
[
0.5146484375,
0.48388671875,
0.880859375,
0.48388671875,
0.880859375,
0.49755859375,
0.5146484375,
0.49755859375
],
[
0.145263671875,
0.4921875,
0.45751953125,
0.493408203125,
0.45751953125,
0.50732421875,
0.145263671875,
0.505859375
],
[
0.5146484375,
0.49951171875,
0.880859375,
0.49951171875,
0.880859375,
0.51318359375,
0.5146484375,
0.51318359375
],
[
0.1485595703125,
0.5087890625,
0.455078125,
0.5087890625,
0.455078125,
0.51953125,
0.1485595703125,
0.51953125
],
[
0.14697265625,
0.521484375,
0.456787109375,
0.521484375,
0.456787109375,
0.53515625,
0.14697265625,
0.53515625
],
[
0.51318359375,
0.5166015625,
0.880859375,
0.5166015625,
0.880859375,
0.52978515625,
0.51318359375,
0.52978515625
],
[
0.14697265625,
0.53759765625,
0.453369140625,
0.53759765625,
0.453369140625,
0.54736328125,
0.14697265625,
0.54736328125
],
[
0.51318359375,
0.5322265625,
0.880859375,
0.5322265625,
0.880859375,
0.5458984375,
0.51318359375,
0.5458984375
],
[
0.14697265625,
0.55029296875,
0.456787109375,
0.55029296875,
0.456787109375,
0.56396484375,
0.14697265625,
0.56396484375
],
[
0.51318359375,
0.54638671875,
0.87890625,
0.54638671875,
0.87890625,
0.56005859375,
0.51318359375,
0.56005859375
],
[
0.145263671875,
0.56396484375,
0.45751953125,
0.56396484375,
0.45751953125,
0.57763671875,
0.145263671875,
0.57763671875
],
[
0.51123046875,
0.5625,
0.880859375,
0.5634765625,
0.880859375,
0.57763671875,
0.51123046875,
0.57666015625
],
[
0.14697265625,
0.5791015625,
0.456787109375,
0.5791015625,
0.456787109375,
0.5927734375,
0.14697265625,
0.5927734375
],
[
0.51123046875,
0.5791015625,
0.880859375,
0.580078125,
0.880859375,
0.59423828125,
0.51123046875,
0.59326171875
],
[
0.14697265625,
0.59423828125,
0.456787109375,
0.59423828125,
0.456787109375,
0.60498046875,
0.14697265625,
0.60498046875
],
[
0.5146484375,
0.5966796875,
0.880859375,
0.5966796875,
0.880859375,
0.60986328125,
0.5146484375,
0.60986328125
],
[
0.145263671875,
0.60693359375,
0.45751953125,
0.60546875,
0.45751953125,
0.619140625,
0.145263671875,
0.62060546875
],
[
0.5146484375,
0.6123046875,
0.880859375,
0.6123046875,
0.880859375,
0.6259765625,
0.5146484375,
0.6259765625
],
[
0.145263671875,
0.619140625,
0.456787109375,
0.62060546875,
0.456787109375,
0.63525390625,
0.145263671875,
0.63427734375
],
[
0.515625,
0.6298828125,
0.87890625,
0.6298828125,
0.87890625,
0.64013671875,
0.515625,
0.64013671875
],
[
0.1485595703125,
0.6357421875,
0.456787109375,
0.6357421875,
0.456787109375,
0.646484375,
0.1485595703125,
0.646484375
],
[
0.14697265625,
0.6494140625,
0.382080078125,
0.6494140625,
0.382080078125,
0.6630859375,
0.14697265625,
0.6630859375
],
[
0.51318359375,
0.64453125,
0.880859375,
0.64453125,
0.880859375,
0.65771484375,
0.51318359375,
0.65771484375
],
[
0.5146484375,
0.66015625,
0.880859375,
0.66015625,
0.880859375,
0.67333984375,
0.5146484375,
0.67333984375
],
[
0.1175537109375,
0.67578125,
0.145263671875,
0.67578125,
0.145263671875,
0.68701171875,
0.1175537109375,
0.68701171875
],
[
0.1502685546875,
0.6748046875,
0.2626953125,
0.6748046875,
0.2626953125,
0.689453125,
0.1502685546875,
0.689453125
],
[
0.51318359375,
0.6767578125,
0.67919921875,
0.6767578125,
0.67919921875,
0.6904296875,
0.51318359375,
0.6904296875
],
[
0.53076171875,
0.69482421875,
0.8818359375,
0.6953125,
0.8818359375,
0.71044921875,
0.53076171875,
0.708984375
],
[
0.12005615234375,
0.70068359375,
0.484375,
0.70068359375,
0.484375,
0.71435546875,
0.12005615234375,
0.71435546875
],
[
0.118408203125,
0.71630859375,
0.48779296875,
0.71630859375,
0.48779296875,
0.72998046875,
0.118408203125,
0.72998046875
],
[
0.51123046875,
0.71142578125,
0.87890625,
0.71142578125,
0.87890625,
0.724609375,
0.51123046875,
0.724609375
],
[
0.1175537109375,
0.732421875,
0.48779296875,
0.732421875,
0.48779296875,
0.7451171875,
0.1175537109375,
0.7451171875
],
[
0.51318359375,
0.728515625,
0.87744140625,
0.728515625,
0.87744140625,
0.7392578125,
0.51318359375,
0.7392578125
],
[
0.118408203125,
0.75,
0.486083984375,
0.75,
0.486083984375,
0.76025390625,
0.118408203125,
0.76025390625
],
[
0.51318359375,
0.74462890625,
0.880859375,
0.74462890625,
0.880859375,
0.75830078125,
0.51318359375,
0.75830078125
],
[
0.118408203125,
0.76416015625,
0.48779296875,
0.76416015625,
0.48779296875,
0.77783203125,
0.118408203125,
0.77783203125
],
[
0.51123046875,
0.7587890625,
0.87890625,
0.7587890625,
0.87890625,
0.7724609375,
0.51123046875,
0.7724609375
],
[
0.118408203125,
0.78076171875,
0.486083984375,
0.78076171875,
0.486083984375,
0.79443359375,
0.118408203125,
0.79443359375
],
[
0.51318359375,
0.77685546875,
0.880859375,
0.77685546875,
0.880859375,
0.79052734375,
0.51318359375,
0.79052734375
],
[
0.118408203125,
0.79638671875,
0.486083984375,
0.79638671875,
0.486083984375,
0.81005859375,
0.118408203125,
0.81005859375
],
[
0.51123046875,
0.79150390625,
0.87890625,
0.79150390625,
0.87890625,
0.8046875,
0.51123046875,
0.8046875
],
[
0.1175537109375,
0.8134765625,
0.484375,
0.8134765625,
0.484375,
0.82666015625,
0.1175537109375,
0.82666015625
],
[
0.51318359375,
0.80859375,
0.880859375,
0.80859375,
0.880859375,
0.822265625,
0.51318359375,
0.822265625
],
[
0.118408203125,
0.8291015625,
0.486083984375,
0.8291015625,
0.486083984375,
0.8427734375,
0.118408203125,
0.8427734375
],
[
0.51318359375,
0.826171875,
0.880859375,
0.826171875,
0.880859375,
0.8388671875,
0.51318359375,
0.8388671875
],
[
0.12005615234375,
0.8466796875,
0.483642578125,
0.8466796875,
0.483642578125,
0.85693359375,
0.12005615234375,
0.85693359375
],
[
0.51318359375,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.118408203125,
0.86083984375,
0.48779296875,
0.86083984375,
0.48779296875,
0.87451171875,
0.118408203125,
0.87451171875
],
[
0.51123046875,
0.85546875,
0.880859375,
0.85693359375,
0.880859375,
0.87158203125,
0.51123046875,
0.87060546875
],
[
0.51318359375,
0.87353515625,
0.880859375,
0.87353515625,
0.880859375,
0.88720703125,
0.51318359375,
0.88720703125
],
[
0.145263671875,
0.8818359375,
0.361083984375,
0.8818359375,
0.361083984375,
0.8955078125,
0.145263671875,
0.8955078125
],
[
0.130126953125,
0.89453125,
0.273681640625,
0.89453125,
0.273681640625,
0.90771484375,
0.130126953125,
0.90771484375
],
[
0.51123046875,
0.8896484375,
0.87744140625,
0.8896484375,
0.87744140625,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.300537109375,
0.8955078125,
0.484375,
0.89453125,
0.484375,
0.90625,
0.300537109375,
0.90673828125
],
[
0.1175537109375,
0.90673828125,
0.16796875,
0.90673828125,
0.16796875,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.51318359375,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"使用",
"通用表示和交叉映射改进零快照多语言翻译",
"Shuhao Gu',2, Yang Feng,2*",
"I Key Laboratory of Intelligent Information Processing,",
"Institute of Computing Technology, Chinese Academy of Sciences (ICT/CAS)",
"2 University of Chinese Academy of Sciences",
"[ gushuhao19b,fengyang ↓ @ict.ac.cn",
"Abstract",
"系，这些特征和关系可以将不同语言的语义等",
"效句子映射到目标语言的特定空间。通过这种",
"The many-to-many multilingual neural ma-",
"方式，该模型可以将在监督翻译方向上学习的",
"chine translation can translate between lan",
"guage pairs unseen during training, i.e., zero",
"知识转移到零拍翻译方向。然而，现有的模型",
"shot translation. Improving zero-shot transla-",
"结构和训练方案不能保证通用表示和交叉映",
"tion requires the model to learn universal repre-",
"射，因为缺乏明确的约束。具体地，编码器口",
"sentations and cross-mapping relationships to",
"以将不同的语言映射到不同的语义子空间，解",
"transfer the knowledge learned on the super",
"码器可以针对不同的源语言学习不同的映射关",
"vised directions to the zero-shot directions.In",
"系，尤其是当模型具有高容量时",
"this work, weproposethe state mover's dis-",
"tance based on the optimal theory to model",
"许多研究人员已经尝试解决这个问题。Pham",
"the difference of the representations output by",
"etal.(2019)建议将编码器的输出压缩为一致数",
"the encoder. Then, we bridge the gap between",
"量的状态，以仅编码与语言无关的功能。Ari",
"the semantic-equivalent representations of dif",
"vazhaganetal.（2019）添加正则化损失，以最大",
"ferent languages at the token level by mini-",
"化源句子和目标句子的句子表示之间的相似",
"mizing the proposed distance to learn univer",
"性。Panetal.(2021)提出了对比学习方案，以",
"sal representations. Besides, we propose an",
"最大程度地减少相似句子的句子表示差距，并",
"agreement-basedtraining scheme,which can",
"help the model make consistent predictions",
"最大程度地减少无关句子的句子表示差距。上",
"based on the semantic-equivalent sentences",
"述所有工作都试图最大程度地减少句子级别",
"to learn universal cross-mapping relationships",
"上不同语言的表示差异，这给NMT带来了两",
"for all translation directions. The experimental",
"个问题。首先，这些工作通常通过最大池化或",
"results on diverse multilingual datasets show",
"平均获得编码器输出的句子级表示，这可能会",
"thatourmethodcanimproveconsistentlycom",
"忽略句子长度，单词对齐关系和其他令牌级信",
"pared with the baseline system and other con-",
"息。其次，正则化句子表示与NMT模型的工",
"trast methods.The analysis proves that our",
"作范式不匹配，因为解码器直接对整个状态序",
"method canbetteralignthesemanticspaceand",
"improve the prediction consistency",
"列而不是句子表示进行交叉关注。此外，上述",
"所有工作都集中在编码器端，无法帮助学习解",
"1",
"Introduction",
"码器的通用映射关系",
"鉴于上述情况，我们提出了一种学习通用表",
"多对多多多语言神经机器翻译（NMT）E(Ha",
"et al., 2016; Firat et al., 2016; Johnson et al., 2017;",
"示和交叉映射的方法，以提高零拍翻译性能。",
"Gu et al., 2018; Fan et al., 2020; Zhang et al.,",
"基于最优运输理论，我们提出了状态移动器的",
"2020a）模型可以在单个模型中支持多个翻译",
"距离（SMD），以在令牌级别对两个状态序列",
"方向。共享编码器将输入的句子编码到语义空",
"的差异进行建模。为了将来自不同语言的语义",
"间，然后共享解码器从该空间解码以生成目标",
"等效句子映射到语义空间的同一位置，我们添",
"语言的翻译。这种范式允许模型在训练期间看",
"加了一个辅助损失，以最小化源句子和目标句",
"不见的语言对之间进行翻译，即。e。，零拍翻",
"子的SMD。此外，我们提出了一种基于协议的",
"译。零镜头翻译可以提高推理效率，使模型需",
"培训方案，以学习具有相同目标语言的翻译方",
"要更少的双语训练数据。执行零翻译需要通用",
"向的通用映射关系。我们将源句子和目标句子",
"表示来编码与语言无关的特征和交叉映射关",
"混合在一起，得到一个伪句子。然后，解码器",
"分别以该伪句子和相应的源句子或目标句子为",
"*Corresponding author: Yang Feng",
"Reproducible code:",
"条件进行预测。我们试图通过最小化两个输出",
"https://github.com/ictnlp/Zero",
"MNMT.",
"分布的KL散度来提高预测的一致性。在不同"
] |
{
"arxiv_id": "2210.15851",
"language": "zh",
"page_idx": 1
}
|
||
2210.15851-zh-2
|
2210.15851-zh-2.jpg
|
的多语言数据集上的实验结果表明，我们的方法可以带来 2???BLEU 改进强基线系统，并始终优于其他对比方法。分析证明，我们的方法可以更好地对齐语义空间，提高预测一致性
## 2 $\textbf{Background}$
在本节中，我们将简要介绍 TRANSFORMER $\boxed{\mathrm{E}}$ (Vaswani et al., 2017) 模型以及多对多多多语言翻译。
## 2.1 $\textbf{The transformer}$
我们将符号的输入序列表示为$x=$ $( x_1, \ldots , x_{nx})$ ,将地面实况序列表示为 $y=$ $(y_1,\ldots,y_{ny})$ 。变压器模型基于编码器-解码器架构。编码器由$N$相同的层组成。每层有两个子层。第一个是多头自关注子层，第二个是全连接的前馈网络。两个子层之后都是残差连接操作和层归一化操作。输入序列$x$将首先转换为向量序列。然后，该矢量序列将被馈送到编码器中，并且第$N$层的输出将被视为源状态序列。我们将其表示为$H_x$ 。解码器也由 $N$ 相同的层组成。除了每个编码器层中的相同类型的两个子层之外，还在它们之间插人了交叉关注子层，该交叉关注子层在编码器的输出上执行多头关注。我们可以得到由源句子和$k-1$ 先前目标词条件的第$k$个目标词的预测概率。通过在老师强迫下最小化基本事实序列的交叉熵损失来优化模型：
$$\mathcal{L}_{CE}=-\frac{1}{n_y}\sum_{k=1}^{n_y}\log p(y_k|\boldsymbol{y}_{<k},\boldsymbol{x};\theta),$$
(1)
其中$n_y$是目标句子的长度，$\theta$表示模型参数
## 2.2 Multilingual Translation
我们定义$L=\{l_1,\ldots,l_M\}$,其中$L$是训练阶段涉及的$M$语言的集合。在 Johnson et al. (2017)之后，我们共享所有语言的所有模型参数。在$\overline{\mathrm{ELiu~et~al.(2020)}}$之后，我们分别在源句子和目标句子的开头添加一个特定的语言$\operatorname{ID}$标记，以指示该语言
# 3 $\textbf{Method}$
我们方法的主要思想是帮助编码器输出所有语言的通用表示，并帮助解码器将语义等效表示从不同语言映射到目标语言的空间。我们提出了两种方法来实现这一目标。第一种是直接弥合携带相同语义的状态序列之间的差距。第二种是强制解码器根据语义等价的句子做出一致的预测。图 1显示了整个训练方案
<FigureHere>
Figure 1: The training scheme of our method. $x$ and $y$ denote a pair of translations; $H_x$ and $H_y$ denote the corresponding state sequences. $z$ is the pseudo sentence by mixuping $x$ and $y.$ 'Dec' denotes the decoder and there is only one decoder in the model. 'stop-grad' denotes the stop-gradient operation during back propagation. $\mathcal{L}_CE,\mathcal{L}_{OT}$ , and $\mathcal{L}_AT$ denote the cross entropy loss, optimal transport loss, and agreement-based train- $\operatorname{ing}\log\log.$
### 3.1 0ptimal Transport
地球移动器的距离基于最优运输理论$\overline{\mathrm{E}}($Villani, 2009; Peyré et al.,2019) , 地球移动器的距离 (EMD) 测量将概率质量从一个分布运输到另一个分布的最小成本。假设有两个概率分布$\mu$和$\mu^\prime$,定义为：
$$\mu=\{(\boldsymbol{w}_{i},m_{i})\}_{i=1}^{n},\quad s.t.\sum_{i}m_{i}=1;\\\mu^{\prime}=\{(\boldsymbol{w}_{j}^{\prime},m_{j}^{\prime})\}_{j=1}^{n^{\prime}},\quad s.t.\sum_{j}m_{j}^{\prime}=1,$$
(2)
其中每个数据点$w_i\in\mathbb{R}^d$具有概率质量$m_i$ $(m_{i}>0)。\mu$中有$n$个数据点。我们定义了一个成本函数$c(\boldsymbol{w}_i,\boldsymbol{w}_j^{\prime})$ ,它确定了两点$\boldsymbol{w}_i$和$w_i^{\prime}$之间的单位成本。如上所述，EMD 定义为：
$$\begin{aligned}\mathcal{D}(\mu,\mu^{\prime})&=\operatorname*{min}_{T\geq0}\sum_{i,j}\boldsymbol{T}_{ij}c(\boldsymbol{w}_{i},\boldsymbol{w}_{j}^{\prime})\\s.t.&\sum_{j=1}^{n^{\prime}}\boldsymbol{T}_{ij}=m_{i},\forall i\in\{1,\ldots,n\},\\&\sum_{i=1}^{n}\boldsymbol{T}_{ij}=m_{j}^{\prime},\forall j\in\{1,\ldots,n^{\prime}\}.\end{aligned}$$
(3)
$T_{ij}$表示从$\mu$传输到$\mu^\prime$的质量
状态移动器的距离遵循 EMD,我们定义状态移动器的距离(SMD)以测量两个状态序列之间的最小“旅行成本”。给定一对翻译$x=(x_1,\ldots,x_{nx})$和$\boldsymbol y=(y_1,\ldots,y_{ny})$ ,我们可以在将它们馈送到编码器后得到它们相应的
|
[
[
0.12005615234375,
0.0885009765625,
0.483642578125,
0.0885009765625,
0.483642578125,
0.09857177734375,
0.12005615234375,
0.09857177734375
],
[
0.5576171875,
0.09027099609375,
0.59423828125,
0.09027099609375,
0.59423828125,
0.09979248046875,
0.5576171875,
0.09979248046875
],
[
0.7900390625,
0.0885009765625,
0.80419921875,
0.0885009765625,
0.80419921875,
0.0968017578125,
0.7900390625,
0.0968017578125
],
[
0.118408203125,
0.1015625,
0.484375,
0.1015625,
0.484375,
0.11517333984375,
0.118408203125,
0.11517333984375
],
[
0.515625,
0.10394287109375,
0.5615234375,
0.105712890625,
0.56103515625,
0.12054443359375,
0.5146484375,
0.1181640625
],
[
0.5810546875,
0.10687255859375,
0.62451171875,
0.10687255859375,
0.62451171875,
0.1181640625,
0.5810546875,
0.1181640625
],
[
0.1175537109375,
0.1181640625,
0.483642578125,
0.1181640625,
0.483642578125,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.779296875,
0.1217041015625,
0.8212890625,
0.1217041015625,
0.8212890625,
0.133056640625,
0.779296875,
0.133056640625
],
[
0.118408203125,
0.1353759765625,
0.466796875,
0.1353759765625,
0.466796875,
0.1484375,
0.118408203125,
0.1484375
],
[
0.51904296875,
0.13720703125,
0.5634765625,
0.13720703125,
0.5634765625,
0.1519775390625,
0.51904296875,
0.1519775390625
],
[
0.58251953125,
0.13720703125,
0.62890625,
0.13720703125,
0.62890625,
0.1519775390625,
0.58251953125,
0.1519775390625
],
[
0.724609375,
0.14013671875,
0.7314453125,
0.14013671875,
0.7314453125,
0.144287109375,
0.724609375,
0.144287109375
],
[
0.78076171875,
0.149658203125,
0.82958984375,
0.1519775390625,
0.82861328125,
0.16748046875,
0.779296875,
0.1650390625
],
[
0.5458984375,
0.15673828125,
0.55322265625,
0.15673828125,
0.55322265625,
0.162109375,
0.5458984375,
0.162109375
],
[
0.60107421875,
0.1561279296875,
0.623046875,
0.1561279296875,
0.623046875,
0.163330078125,
0.60107421875,
0.163330078125
],
[
0.6708984375,
0.15673828125,
0.69775390625,
0.15673828125,
0.69775390625,
0.1644287109375,
0.6708984375,
0.1644287109375
],
[
0.1158447265625,
0.1650390625,
0.151123046875,
0.1650390625,
0.151123046875,
0.1776123046875,
0.1158447265625,
0.1776123046875
],
[
0.1485595703125,
0.163330078125,
0.26025390625,
0.1650390625,
0.26025390625,
0.179931640625,
0.1485595703125,
0.1776123046875
],
[
0.1502685546875,
0.168701171875,
0.154541015625,
0.168701171875,
0.154541015625,
0.17333984375,
0.1502685546875,
0.17333984375
],
[
0.5205078125,
0.169189453125,
0.611328125,
0.171630859375,
0.611328125,
0.1859130859375,
0.5205078125,
0.18408203125
],
[
0.65576171875,
0.17333984375,
0.736328125,
0.17333984375,
0.736328125,
0.188232421875,
0.65576171875,
0.188232421875
],
[
0.5205078125,
0.185302734375,
0.611328125,
0.185302734375,
0.611328125,
0.1953125,
0.5205078125,
0.1953125
],
[
0.77685546875,
0.18115234375,
0.87158203125,
0.18115234375,
0.87158203125,
0.1923828125,
0.77685546875,
0.1923828125
],
[
0.118408203125,
0.1923828125,
0.486083984375,
0.1923828125,
0.486083984375,
0.2060546875,
0.118408203125,
0.2060546875
],
[
0.59765625,
0.201904296875,
0.78662109375,
0.201904296875,
0.78662109375,
0.215576171875,
0.59765625,
0.215576171875
],
[
0.8076171875,
0.19775390625,
0.81689453125,
0.191162109375,
0.82373046875,
0.196533203125,
0.8154296875,
0.203125
],
[
0.1175537109375,
0.2078857421875,
0.484375,
0.2078857421875,
0.484375,
0.221435546875,
0.1175537109375,
0.221435546875
],
[
0.65576171875,
0.21435546875,
0.67529296875,
0.21435546875,
0.67529296875,
0.22802734375,
0.65576171875,
0.22802734375
],
[
0.73046875,
0.21142578125,
0.7421875,
0.2197265625,
0.73046875,
0.2286376953125,
0.71875,
0.2203369140625
],
[
0.118408203125,
0.2257080078125,
0.162841796875,
0.2257080078125,
0.162841796875,
0.2369384765625,
0.118408203125,
0.2369384765625
],
[
0.51123046875,
0.2423095703125,
0.880859375,
0.2423095703125,
0.880859375,
0.255859375,
0.51123046875,
0.255859375
],
[
0.11419677734375,
0.2529296875,
0.2939453125,
0.2529296875,
0.2939453125,
0.2666015625,
0.11419677734375,
0.2666015625
],
[
0.51123046875,
0.257080078125,
0.87890625,
0.257080078125,
0.87890625,
0.270751953125,
0.51123046875,
0.270751953125
],
[
0.118408203125,
0.275634765625,
0.455078125,
0.275634765625,
0.455078125,
0.289306640625,
0.118408203125,
0.289306640625
],
[
0.51318359375,
0.271484375,
0.8818359375,
0.271484375,
0.8818359375,
0.28515625,
0.51318359375,
0.28515625
],
[
0.450927734375,
0.28076171875,
0.459228515625,
0.28076171875,
0.459228515625,
0.2861328125,
0.450927734375,
0.2861328125
],
[
0.51123046875,
0.28515625,
0.8818359375,
0.28515625,
0.8818359375,
0.298583984375,
0.51123046875,
0.298583984375
],
[
0.11419677734375,
0.292236328125,
0.478515625,
0.29150390625,
0.478515625,
0.305908203125,
0.11419677734375,
0.30712890625
],
[
0.51318359375,
0.2998046875,
0.87890625,
0.2998046875,
0.87890625,
0.3134765625,
0.51318359375,
0.3134765625
],
[
0.1158447265625,
0.30810546875,
0.486083984375,
0.30712890625,
0.486083984375,
0.32177734375,
0.1158447265625,
0.322509765625
],
[
0.509765625,
0.312255859375,
0.880859375,
0.3134765625,
0.880859375,
0.327880859375,
0.509765625,
0.32666015625
],
[
0.1175537109375,
0.322509765625,
0.486083984375,
0.32470703125,
0.486083984375,
0.33837890625,
0.1175537109375,
0.336181640625
],
[
0.51318359375,
0.327880859375,
0.880859375,
0.327880859375,
0.880859375,
0.341552734375,
0.51318359375,
0.341552734375
],
[
0.118408203125,
0.34033203125,
0.486083984375,
0.34033203125,
0.486083984375,
0.35400390625,
0.118408203125,
0.35400390625
],
[
0.5146484375,
0.34326171875,
0.880859375,
0.34326171875,
0.880859375,
0.35400390625,
0.5146484375,
0.35400390625
],
[
0.1175537109375,
0.355224609375,
0.484375,
0.355224609375,
0.484375,
0.3681640625,
0.1175537109375,
0.3681640625
],
[
0.51123046875,
0.358154296875,
0.568359375,
0.358154296875,
0.568359375,
0.3681640625,
0.51123046875,
0.3681640625
],
[
0.1175537109375,
0.371826171875,
0.483642578125,
0.371826171875,
0.483642578125,
0.385498046875,
0.1175537109375,
0.385498046875
],
[
0.1175537109375,
0.388427734375,
0.484375,
0.388427734375,
0.484375,
0.402099609375,
0.1175537109375,
0.402099609375
],
[
0.51318359375,
0.399658203125,
0.5556640625,
0.399658203125,
0.5556640625,
0.410888671875,
0.51318359375,
0.410888671875
],
[
0.55517578125,
0.3984375,
0.7080078125,
0.3984375,
0.7080078125,
0.412109375,
0.55517578125,
0.412109375
],
[
0.1175537109375,
0.405029296875,
0.486083984375,
0.405029296875,
0.486083984375,
0.418701171875,
0.1175537109375,
0.418701171875
],
[
0.1175537109375,
0.42041015625,
0.484375,
0.42041015625,
0.484375,
0.43408203125,
0.1175537109375,
0.43408203125
],
[
0.51318359375,
0.421630859375,
0.8818359375,
0.421630859375,
0.8818359375,
0.435302734375,
0.51318359375,
0.435302734375
],
[
0.118408203125,
0.43701171875,
0.486083984375,
0.43701171875,
0.486083984375,
0.45068359375,
0.118408203125,
0.45068359375
],
[
0.509765625,
0.435791015625,
0.880859375,
0.43701171875,
0.880859375,
0.451904296875,
0.509765625,
0.45068359375
],
[
0.118408203125,
0.453125,
0.486083984375,
0.453125,
0.486083984375,
0.466064453125,
0.118408203125,
0.466064453125
],
[
0.51123046875,
0.45361328125,
0.87744140625,
0.45361328125,
0.87744140625,
0.46728515625,
0.51123046875,
0.46728515625
],
[
0.118408203125,
0.47021484375,
0.483642578125,
0.47021484375,
0.483642578125,
0.48095703125,
0.118408203125,
0.48095703125
],
[
0.51123046875,
0.468505859375,
0.87890625,
0.4697265625,
0.87890625,
0.48388671875,
0.51123046875,
0.482666015625
],
[
0.118408203125,
0.485107421875,
0.489501953125,
0.485107421875,
0.489501953125,
0.498779296875,
0.118408203125,
0.498779296875
],
[
0.51123046875,
0.486328125,
0.6533203125,
0.486328125,
0.6533203125,
0.49951171875,
0.51123046875,
0.49951171875
],
[
0.118408203125,
0.5029296875,
0.480224609375,
0.5029296875,
0.480224609375,
0.51318359375,
0.118408203125,
0.51318359375
],
[
0.1175537109375,
0.51708984375,
0.484375,
0.51708984375,
0.484375,
0.53076171875,
0.1175537109375,
0.53076171875
],
[
0.5341796875,
0.51708984375,
0.8271484375,
0.515625,
0.8271484375,
0.5361328125,
0.5341796875,
0.5380859375
],
[
0.118408203125,
0.533203125,
0.27197265625,
0.533203125,
0.27197265625,
0.54638671875,
0.118408203125,
0.54638671875
],
[
0.744140625,
0.54052734375,
0.75390625,
0.54052734375,
0.75390625,
0.54736328125,
0.744140625,
0.54736328125
],
[
0.85986328125,
0.5458984375,
0.87890625,
0.5458984375,
0.87890625,
0.5546875,
0.85986328125,
0.5546875
],
[
0.53564453125,
0.552734375,
0.84033203125,
0.552734375,
0.84033203125,
0.57080078125,
0.53564453125,
0.57080078125
],
[
0.235107421875,
0.55859375,
0.496337890625,
0.57177734375,
0.49365234375,
0.60400390625,
0.231689453125,
0.59130859375
],
[
0.270263671875,
0.5654296875,
0.28369140625,
0.5654296875,
0.28369140625,
0.57177734375,
0.270263671875,
0.57177734375
],
[
0.16455078125,
0.57373046875,
0.2166748046875,
0.57763671875,
0.2149658203125,
0.59130859375,
0.162109375,
0.58837890625
],
[
0.748046875,
0.57177734375,
0.767578125,
0.57177734375,
0.767578125,
0.5849609375,
0.748046875,
0.5849609375
],
[
0.2376708984375,
0.5830078125,
0.2939453125,
0.5927734375,
0.28955078125,
0.60693359375,
0.2333984375,
0.59716796875
],
[
0.52880859375,
0.60107421875,
0.87890625,
0.6015625,
0.87890625,
0.6162109375,
0.52880859375,
0.615234375
],
[
0.118408203125,
0.6181640625,
0.47265625,
0.6181640625,
0.47265625,
0.6318359375,
0.118408203125,
0.6318359375
],
[
0.51123046875,
0.6181640625,
0.87890625,
0.6181640625,
0.87890625,
0.6318359375,
0.51123046875,
0.6318359375
],
[
0.52880859375,
0.63525390625,
0.880859375,
0.63525390625,
0.880859375,
0.6484375,
0.52880859375,
0.6484375
],
[
0.1158447265625,
0.646484375,
0.35693359375,
0.64794921875,
0.35693359375,
0.662109375,
0.1158447265625,
0.6611328125
],
[
0.5146484375,
0.6494140625,
0.8818359375,
0.6494140625,
0.8818359375,
0.66455078125,
0.5146484375,
0.66455078125
],
[
0.118408203125,
0.67041015625,
0.486083984375,
0.67041015625,
0.486083984375,
0.68408203125,
0.118408203125,
0.68408203125
],
[
0.118408203125,
0.68603515625,
0.484375,
0.68603515625,
0.484375,
0.69970703125,
0.118408203125,
0.69970703125
],
[
0.53076171875,
0.68310546875,
0.78271484375,
0.68310546875,
0.78271484375,
0.70068359375,
0.53076171875,
0.70068359375
],
[
0.118408203125,
0.70166015625,
0.484375,
0.70166015625,
0.484375,
0.71533203125,
0.118408203125,
0.71533203125
],
[
0.62109375,
0.69775390625,
0.65576171875,
0.69775390625,
0.65576171875,
0.7060546875,
0.62109375,
0.7060546875
],
[
0.65576171875,
0.70166015625,
0.67822265625,
0.70166015625,
0.67822265625,
0.71533203125,
0.65576171875,
0.71533203125
],
[
0.118408203125,
0.71875,
0.486083984375,
0.71875,
0.486083984375,
0.732421875,
0.118408203125,
0.732421875
],
[
0.60205078125,
0.71875,
0.62451171875,
0.71875,
0.62451171875,
0.73095703125,
0.60205078125,
0.73095703125
],
[
0.59619140625,
0.728515625,
0.84619140625,
0.728515625,
0.84619140625,
0.75048828125,
0.59619140625,
0.75048828125
],
[
0.55322265625,
0.7353515625,
0.5810546875,
0.7353515625,
0.5810546875,
0.74658203125,
0.55322265625,
0.74658203125
],
[
0.8564453125,
0.7353515625,
0.8818359375,
0.73828125,
0.87744140625,
0.751953125,
0.85205078125,
0.7490234375
],
[
0.8486328125,
0.7412109375,
0.8564453125,
0.7412109375,
0.8564453125,
0.7451171875,
0.8486328125,
0.7451171875
],
[
0.59619140625,
0.7490234375,
0.62744140625,
0.7451171875,
0.63037109375,
0.7626953125,
0.599609375,
0.765625
],
[
0.12005615234375,
0.736328125,
0.483642578125,
0.736328125,
0.483642578125,
0.74658203125,
0.12005615234375,
0.74658203125
],
[
0.1175537109375,
0.75048828125,
0.307373046875,
0.75048828125,
0.307373046875,
0.76416015625,
0.1175537109375,
0.76416015625
],
[
0.59619140625,
0.77099609375,
0.8427734375,
0.77099609375,
0.8427734375,
0.7998046875,
0.59619140625,
0.7998046875
],
[
0.603515625,
0.765625,
0.623046875,
0.765625,
0.623046875,
0.77978515625,
0.603515625,
0.77978515625
],
[
0.1158447265625,
0.78076171875,
0.151123046875,
0.78076171875,
0.151123046875,
0.79345703125,
0.1158447265625,
0.79345703125
],
[
0.1435546875,
0.77978515625,
0.2225341796875,
0.77978515625,
0.2225341796875,
0.79443359375,
0.1435546875,
0.79443359375
],
[
0.599609375,
0.79443359375,
0.62646484375,
0.79443359375,
0.62646484375,
0.80810546875,
0.599609375,
0.80810546875
],
[
0.118408203125,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82177734375,
0.118408203125,
0.82177734375
],
[
0.118408203125,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.51123046875,
0.82373046875,
0.75244140625,
0.82373046875,
0.75244140625,
0.83740234375,
0.51123046875,
0.83740234375
],
[
0.118408203125,
0.84130859375,
0.484375,
0.84130859375,
0.484375,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.509765625,
0.84033203125,
0.87890625,
0.84033203125,
0.87890625,
0.85400390625,
0.509765625,
0.85400390625
],
[
0.12005615234375,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.12005615234375,
0.87060546875
],
[
0.51123046875,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.12005615234375,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.12005615234375,
0.8857421875
],
[
0.51318359375,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.118408203125,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51318359375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90380859375,
0.51318359375,
0.90380859375
],
[
0.118408203125,
0.90478515625,
0.399658203125,
0.90478515625,
0.399658203125,
0.91845703125,
0.118408203125,
0.91845703125
],
[
0.515625,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.515625,
0.91845703125
]
] |
[
"的多语言数据集上的实验结果表明，我们的方",
"LAT",
"LO",
"法可以带来2???BLEU改进强基线系统，并始",
"P(z|x)",
"P(zly)",
"终优于其他对比方法。分析证明，我们的方法",
"P(ylx)",
"可以更好地对齐语义空间，提高预测一致性",
"Dec",
"[Dec",
"-",
"Dec",
"t",
"T",
"grad",
"2",
"Background",
"",
"<FR>I vais fine.",
"Encoder",
"z~mixup(x,y)",
"<FR>Jevaisbien.",
"在本节中，我们将简要介绍TRANSFORMERE",
"<FR>Je vais bien,<EN>Iam fine.",
"2",
"(Vaswanietal.,2017)模型以及多对多多多语言",
"y",
"f",
"翻译。",
"Figure I: The training scheme of our method. and",
"2.1 The transformer",
"y denote a pair of translations; H, and Hy denote",
"我们将符号的输人序列表示为α",
"the corresponding state sequences. z is the pseudo sen-",
"",
"tence by mixuping and y . 'Dec' denotes the decoder",
"(C1,...,ana），将地面实况序列表示为y",
"and there is only one decoder in the model. 'stop-grad'",
"(y1，...，yny）。变压器模型基于编码器-解码器",
"denotes the stop-gradient operation during back propa",
"架构。编码器由N相同的层组成。每层有两",
"gation. LcE, CoT , and CAT denote the cross entropy",
"个子层。第一个是多头自关注子层，第二个是",
"loss,optimaltransportloss,andagreement-basedtrain",
"全连接的前馈网络。两个子层之后都是残差连",
"ingloss.",
"接操作和层归一化操作。输入序列&将首先转",
"换为向量序列。然后，该矢量序列将被馈送到",
"3.1",
"Optimal Transport",
"编码器中，并且第N层的输出将被视为源状",
"态序列。我们将其表示为H。解码器也由N",
"地球移动器的距离基于最优运输理论E(Vil-",
"相同的层组成。除了每个编码器层中的相同类",
"lani,2009;Peyréetal.,2019），地球移动器的距",
"型的两个子层之外，还在它们之间插人了交叉",
"离（EMD）测量将概率质量从一个分布运输到",
"关注子层，该交叉关注子层在编码器的输出上",
"另一个分布的最小成本。假设有两个概率分布",
"执行多头关注。我们可以得到由源句子和k-1",
"u和，定义为：",
"先前目标词条件的第k个目标词的预测概率",
"通过在老师强迫下最小化基本事实序列的交叉",
"μ= (wi,mi)=1, s.t.mi= 1;",
"摘损失来优化模型：",
"2.",
"(2)",
"μ = [(wj,m'))g=1,s.t.mj = 1,",
"1Z1og (ul<k,;0), (1)",
"y",
"LCE",
"j",
"y k=1",
"其中每个数据点wiERd具有概率质量mi",
"其中nu是目标句子的长度，表示模型参数",
"(mi>0）。μ中有n个数据点。我们定义了",
"个成本函数c(wi，w'），它确定了两点w和",
"2.2Multilingual Translation",
"w,之间的单位成本。如上所述，EMD定义为：",
"我们定义L={l1,·,lM}，其中L是训练",
"阶段涉及的M语言的集合。在Johnsonetal.",
"D(μ, μ) =min Tijc(wi, w)",
"(2017)之后，我们共享所有语言的所有模型参",
"T>0",
"i.j",
"数。在ELiuetal.(2020)之后，我们分别在源",
"n'",
"ET, =mi, ie (1,...n],",
"s.t.",
"(3)",
"",
"j=1",
"句子和自标句子的开头添加一个特定的语言",
"ID标记，以指示该语言",
"ZTu m., j 1....",
"n",
"3",
"Method",
"i=1",
"我们方法的主要思想是帮助编码器输出所有语",
"言的通用表示，并帮助解码器将语义等效表示",
"Tii表示从μ传输到u的质量",
"从不同语言映射到目标语言的空间。我们提出",
"状态移动器的距离遵循EMD，我们定义状",
"了两种方法来实现这一目标。第一种是直接弥",
"态移动器的距离（SMD）以测量两个状态序",
"合携带相同语义的状态序列之间的差距。第二",
"列之间的最小“旅行成本”。给定一对翻译",
"种是强制解码器根据语义等价的句子做出一致",
"&=（ai,..,ana）和y=（yi,...,Yny），我们",
"的预测。图1显示了整个训练方案",
"可以在将它们馈送到编码器后得到它们相应的"
] |
{
"arxiv_id": "2210.15851",
"language": "zh",
"page_idx": 2
}
|
||
2210.15851-zh-4
|
2210.15851-zh-4.jpg
|
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>Language Pairs</th>
<th>Size</th>
</tr>
<tr>
<td>$\mathbf{IWSLT}$</td>
<td>En$\leftrightarrow \{$ De,It,Nl,Ro$\}$</td>
<td>$1.79M$</td>
</tr>
<tr>
<td>IWSLT- b</td>
<td>Nl$\leftrightarrow$De$\leftrightarrow$En$\leftrightarrow$It$\leftrightarrow$Ro</td>
<td>$1.79M$</td>
</tr>
<tr>
<td>$PC-6$</td>
<td>En$\leftrightarrow \{$ Kk, Tr, Ro, Cs, Ru $\}$</td>
<td>7.9M</td>
</tr>
<tr>
<td>OPUS- 7</td>
<td>En$\leftrightarrow \{$ De,Fr,Nl,Ru,Zh,Ar$\}$</td>
<td>$11.6M$</td>
</tr>
</tbody>
</table>
Table 1: The statics of our datasets.
的较长句子，使它们的长度相等。由于这两个句子是翻译对，因此它们的句子长度通常很接近，截断不会显着减少较长句子的长度，并且不会增强解码器学习较短的输出。我们将截断的句子表示为$x^{\prime}$和$y^{\prime}$ ,其长度表示为$n^{\prime}$ 。然后我们可以生成$z$为：
$$z=g\odot x'+(1-\boldsymbol{g})\odot\boldsymbol{y}',$$
(11)
其中$g\in\{0,1\}^{n^{\prime}},\odot$表示元素的乘积。$g$中的每个元素都是从伯努利$(\lambda)$中采样的，其中参数 λ 是从 Beta $(\alpha,\beta)$ 中采样的，而$\alpha$ 和β是两个超参数。$z$中确定翻译方向的语言标记来自$x$或$y$
目标函数与公式E9 类似，我们根据模型预测
分布的 KL 散度定义了另一个对称损失：
$$\begin{aligned}\mathcal{L}_{AT}=&\frac{1}{2n'}\sum_{k=1}^{n'}KL\left(p(z_{k}|\boldsymbol{z}_{<k},\boldsymbol{H}_{\boldsymbol{x}})||p(z_{k}|\boldsymbol{z}_{<k},\boldsymbol{H}_{\boldsymbol{y}})\right)\\&+KL\left(p(z_{k}|\boldsymbol{z}_{<k},\boldsymbol{H}_{\boldsymbol{y}})||p(z_{k}|\boldsymbol{z}_{<k},\boldsymbol{H}_{\boldsymbol{x}})\right).\end{aligned}$$
为方便起见，我们省略了模型参数
## 3.3 The Final Loss
最终损失由三部分组成，交叉熵损失(方程回1),基于 SMD 的最佳传输损失(方程$\overline{\mathrm{E}}$9)和基于协议的训练的 KL 发散损失(方程$\overline{\mathrm{E}}$12 ):
(13)
$$\mathcal{L}=\mathcal{L}_{CE}+\gamma_{1}|x|\mathcal{L}_{OT}+\gamma_{2}\mathcal{L}_{AT}$$
两个正则化控制 MAX。由于$\mathcal{L}_{OT}$是在句子级别上计算的，而其他两个损失是在令牌级别上计算的，因此我们将平均序列长度$|x|$乘以$\mathcal{L}_{OT}$。在这三个损失中，第一项主导了模型的参数更新，并决定了模型的性能。后两个正则化损失项仅稍微修改了梯度的方向。因为第一个损失项不依赖于$H_y$,所以我们将停止梯度操作应用于$H_y$ (图$\tilde{\mathbb{E}}$1 ),这意味着梯度不会通过$H_y$到达编码器
# 4 Experiments
4.1 Data Preparation
我们在以下多语言数据集上进行实验： 计信息在表 1 中。我们把更多的细节在附录
IWSLT17 (Cettolo et al., 2017) 我们模拟了两个场景。第一个 (IWSLT) 是 English-pivot, 我们只保留从/到英语的平行句子。第二个(IWSLTb) 具有枢轴链，其中两种语言通过枢轴语言链连接。每个平移方向大约有0。22M 句对。这两个场景都有八个监督平移方向和十二个零拍平移方向。我们使用官方验证和测试集
PC-6 从 PC-32 语料库中提取 PC-6 数据集 (Lin et al.,2020)。不同语言对的数据量不平衡，从0。12M 至 1。84M。此数据集有十个监督和二十个零拍翻译方向。我们将从 WMT16??? 收集的验证和测试集用于监督方向。从 WikiMatrix[E] (Schwenk et al., 2021) 中提取零触发验证和测试集，每个集包含约 1K???K 句子对
OPUS-7OPUS-7 数据集从 OPUS-100 语料库中提取 (Zhang et al.,2020a)。语言对来自不同的语言家族，并且具有显着差异。此数据集有 12个监督翻译方向和 30 个零拍翻译方向。我们使用 Zhang et al.(2020a)发布的标准验证和测试集。为方便起见，我们将零镜头测试集与相同的目标语言连接起来
我们使用斯坦福分词器$\overline{\mathrm{E}}($Tseng et al., 2005; Monroe et al. , 2014)来分割阿拉伯语和中文，并
Moses 工具包E(Koehn et al., 2007) 来标记其他语言。此外，执行 32K 的集成操作以学习 BPEE(Sennrich et al., 2016) 。
## 4.2 Systems
我们使用名为 Fairseq-py $\boxed{\mathrm{E}}($Ott et al.,2019) 的开源工具包作为我们的 Transformer 系统。我们实施以下系统：
·零射(ZS)仅使用交叉熵损失(方程$\mathbb{E}$1)训练的基线系统。然后直接在零射测试集上对模型进行测试
$\cdot$数据透视转换(PivT)(Cheng et al.,2017)与ZS 相同的转换模型。该模型首先将源语言转换为枢轴语言，然后生成目标语言
$\bullet$句子表示对齐(SRA)$\operatorname{E}($Arivazhagan et al., 2019)此方法添加了正则化损失，以最大程度地减少源和目标句子表示的差异。
$$\mathcal{L}=\mathcal{L}_{CE}+\gamma Dis(Enc(s),Enc(t)),$$
(14)
其中“Dis”表示距离函数，“Enc($\cdot)”表示句$ 子表示。我们使用平均句子表示和欧几里得距离函数，因为我们发现它们的效果更好。我们将超参数$\gamma$从 0.1 更改为 1 来调整性能
IWSLT17, PC-6 和 OPUS-7。训练集的简要统 ·SoftmaxForcing (SF) $\boxed{\mathrm{E}}$(Pham et al.,2019) 此
方法使解码器能够通过添加额外的损失来从自
|
[
[
0.1234130859375,
0.08197021484375,
0.1822509765625,
0.0843505859375,
0.181396484375,
0.09857177734375,
0.12176513671875,
0.0968017578125
],
[
0.2032470703125,
0.08489990234375,
0.306396484375,
0.08489990234375,
0.306396484375,
0.0968017578125,
0.2032470703125,
0.0968017578125
],
[
0.4306640625,
0.0843505859375,
0.468505859375,
0.0843505859375,
0.468505859375,
0.0968017578125,
0.4306640625,
0.0968017578125
],
[
0.51123046875,
0.08489990234375,
0.880859375,
0.08612060546875,
0.880859375,
0.1009521484375,
0.51123046875,
0.09979248046875
],
[
0.1275634765625,
0.097412109375,
0.1822509765625,
0.097412109375,
0.1822509765625,
0.11224365234375,
0.1275634765625,
0.11224365234375
],
[
0.429931640625,
0.0885009765625,
0.43408203125,
0.0885009765625,
0.43408203125,
0.09320068359375,
0.429931640625,
0.09320068359375
],
[
0.2032470703125,
0.09857177734375,
0.361083984375,
0.09857177734375,
0.361083984375,
0.11224365234375,
0.2032470703125,
0.11224365234375
],
[
0.426513671875,
0.097412109375,
0.476806640625,
0.097412109375,
0.476806640625,
0.11224365234375,
0.426513671875,
0.11224365234375
],
[
0.51318359375,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.2032470703125,
0.11224365234375,
0.3837890625,
0.11224365234375,
0.3837890625,
0.1258544921875,
0.2032470703125,
0.1258544921875
],
[
0.42822265625,
0.10986328125,
0.47607421875,
0.11224365234375,
0.474365234375,
0.1270751953125,
0.426513671875,
0.12469482421875
],
[
0.5146484375,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.5146484375,
0.1318359375
],
[
0.1292724609375,
0.1270751953125,
0.166259765625,
0.1270751953125,
0.166259765625,
0.1395263671875,
0.1292724609375,
0.1395263671875
],
[
0.1292724609375,
0.1134033203125,
0.19482421875,
0.1134033203125,
0.19482421875,
0.12469482421875,
0.1292724609375,
0.12469482421875
],
[
0.208251953125,
0.127685546875,
0.391357421875,
0.127685546875,
0.391357421875,
0.1383056640625,
0.208251953125,
0.1383056640625
],
[
0.4306640625,
0.1270751953125,
0.47021484375,
0.1270751953125,
0.47021484375,
0.1395263671875,
0.4306640625,
0.1395263671875
],
[
0.426513671875,
0.1395263671875,
0.474365234375,
0.141357421875,
0.47265625,
0.152587890625,
0.4248046875,
0.15087890625
],
[
0.51318359375,
0.1353759765625,
0.880859375,
0.1353759765625,
0.880859375,
0.1484375,
0.51318359375,
0.1484375
],
[
0.2032470703125,
0.14013671875,
0.420654296875,
0.14013671875,
0.420654296875,
0.15380859375,
0.2032470703125,
0.15380859375
],
[
0.1292724609375,
0.14013671875,
0.187255859375,
0.14013671875,
0.187255859375,
0.1519775390625,
0.1292724609375,
0.1519775390625
],
[
0.51318359375,
0.149658203125,
0.880859375,
0.15087890625,
0.880859375,
0.1650390625,
0.51318359375,
0.1644287109375
],
[
0.181396484375,
0.1650390625,
0.4189453125,
0.16748046875,
0.4189453125,
0.18115234375,
0.181396484375,
0.1787109375
],
[
0.5146484375,
0.16748046875,
0.880859375,
0.16748046875,
0.880859375,
0.18115234375,
0.5146484375,
0.18115234375
],
[
0.5146484375,
0.182861328125,
0.8251953125,
0.182861328125,
0.8251953125,
0.196533203125,
0.5146484375,
0.196533203125
],
[
0.12005615234375,
0.2060546875,
0.486083984375,
0.2060546875,
0.486083984375,
0.2161865234375,
0.12005615234375,
0.2161865234375
],
[
0.51123046875,
0.201904296875,
0.87890625,
0.201904296875,
0.87890625,
0.215576171875,
0.51123046875,
0.215576171875
],
[
0.118408203125,
0.2197265625,
0.484375,
0.2197265625,
0.484375,
0.2327880859375,
0.118408203125,
0.2327880859375
],
[
0.51318359375,
0.218505859375,
0.880859375,
0.218505859375,
0.880859375,
0.232177734375,
0.51318359375,
0.232177734375
],
[
0.118408203125,
0.236328125,
0.484375,
0.236328125,
0.484375,
0.2493896484375,
0.118408203125,
0.2493896484375
],
[
0.509765625,
0.2327880859375,
0.87890625,
0.2340087890625,
0.87890625,
0.2476806640625,
0.509765625,
0.2464599609375
],
[
0.12005615234375,
0.254150390625,
0.483642578125,
0.254150390625,
0.483642578125,
0.26416015625,
0.12005615234375,
0.26416015625
],
[
0.509765625,
0.2493896484375,
0.880859375,
0.248779296875,
0.880859375,
0.26318359375,
0.509765625,
0.26416015625
],
[
0.118408203125,
0.268310546875,
0.48779296875,
0.268310546875,
0.48779296875,
0.281982421875,
0.118408203125,
0.281982421875
],
[
0.51123046875,
0.2666015625,
0.880859375,
0.2666015625,
0.880859375,
0.27978515625,
0.51123046875,
0.27978515625
],
[
0.12005615234375,
0.2861328125,
0.29296875,
0.2861328125,
0.29296875,
0.29638671875,
0.12005615234375,
0.29638671875
],
[
0.51123046875,
0.281982421875,
0.87744140625,
0.281982421875,
0.87744140625,
0.295654296875,
0.51123046875,
0.295654296875
],
[
0.5146484375,
0.298583984375,
0.8037109375,
0.298583984375,
0.8037109375,
0.312255859375,
0.5146484375,
0.312255859375
],
[
0.19482421875,
0.31005859375,
0.411376953125,
0.308837890625,
0.411376953125,
0.32373046875,
0.19482421875,
0.32470703125
],
[
0.4189453125,
0.3134765625,
0.441650390625,
0.3134765625,
0.441650390625,
0.320556640625,
0.4189453125,
0.320556640625
],
[
0.44580078125,
0.314208984375,
0.45166015625,
0.314208984375,
0.45166015625,
0.319580078125,
0.44580078125,
0.319580078125
],
[
0.450927734375,
0.31005859375,
0.48779296875,
0.31005859375,
0.48779296875,
0.322509765625,
0.450927734375,
0.322509765625
],
[
0.5146484375,
0.317626953125,
0.87890625,
0.317626953125,
0.87890625,
0.330810546875,
0.5146484375,
0.330810546875
],
[
0.118408203125,
0.334228515625,
0.48779296875,
0.334228515625,
0.48779296875,
0.3486328125,
0.118408203125,
0.3486328125
],
[
0.51123046875,
0.333251953125,
0.87744140625,
0.333251953125,
0.87744140625,
0.3466796875,
0.51123046875,
0.3466796875
],
[
0.118408203125,
0.350830078125,
0.48779296875,
0.350830078125,
0.48779296875,
0.364013671875,
0.118408203125,
0.364013671875
],
[
0.51318359375,
0.349853515625,
0.880859375,
0.349853515625,
0.880859375,
0.363525390625,
0.51318359375,
0.363525390625
],
[
0.118408203125,
0.366455078125,
0.48779296875,
0.366455078125,
0.48779296875,
0.380126953125,
0.118408203125,
0.380126953125
],
[
0.509765625,
0.364013671875,
0.880859375,
0.365234375,
0.880859375,
0.380126953125,
0.509765625,
0.37890625
],
[
0.12005615234375,
0.38427734375,
0.486083984375,
0.38427734375,
0.486083984375,
0.394287109375,
0.12005615234375,
0.394287109375
],
[
0.51123046875,
0.38134765625,
0.87744140625,
0.38134765625,
0.87744140625,
0.394287109375,
0.51123046875,
0.394287109375
],
[
0.118408203125,
0.396728515625,
0.19482421875,
0.3984375,
0.19482421875,
0.413330078125,
0.118408203125,
0.410888671875
],
[
0.51318359375,
0.39794921875,
0.87890625,
0.39794921875,
0.87890625,
0.410888671875,
0.51318359375,
0.410888671875
],
[
0.12005615234375,
0.41455078125,
0.484375,
0.41455078125,
0.484375,
0.427490234375,
0.12005615234375,
0.427490234375
],
[
0.51123046875,
0.41455078125,
0.69775390625,
0.41455078125,
0.69775390625,
0.427490234375,
0.51123046875,
0.427490234375
],
[
0.118408203125,
0.43115234375,
0.43408203125,
0.43115234375,
0.43408203125,
0.44482421875,
0.118408203125,
0.44482421875
],
[
0.53076171875,
0.432861328125,
0.87890625,
0.432861328125,
0.87890625,
0.446533203125,
0.53076171875,
0.446533203125
],
[
0.202392578125,
0.453125,
0.2275390625,
0.453125,
0.2275390625,
0.46435546875,
0.202392578125,
0.46435546875
],
[
0.51318359375,
0.448974609375,
0.880859375,
0.448974609375,
0.880859375,
0.4619140625,
0.51318359375,
0.4619140625
],
[
0.181396484375,
0.457275390625,
0.214111328125,
0.463134765625,
0.2091064453125,
0.4755859375,
0.17626953125,
0.4697265625
],
[
0.12005615234375,
0.465576171875,
0.21240234375,
0.476806640625,
0.208251953125,
0.4921875,
0.1158447265625,
0.48095703125
],
[
0.197265625,
0.460205078125,
0.88330078125,
0.46142578125,
0.88330078125,
0.485107421875,
0.197265625,
0.48388671875
],
[
0.19140625,
0.48388671875,
0.2301025390625,
0.48388671875,
0.2301025390625,
0.496337890625,
0.19140625,
0.496337890625
],
[
0.5146484375,
0.482666015625,
0.87744140625,
0.482666015625,
0.87744140625,
0.493408203125,
0.5146484375,
0.493408203125
],
[
0.51123046875,
0.495361328125,
0.765625,
0.49755859375,
0.765625,
0.5126953125,
0.51123046875,
0.51025390625
],
[
0.17626953125,
0.501953125,
0.484375,
0.501953125,
0.484375,
0.515625,
0.17626953125,
0.515625
],
[
0.453369140625,
0.5185546875,
0.484375,
0.5185546875,
0.484375,
0.53076171875,
0.453369140625,
0.53076171875
],
[
0.51123046875,
0.53076171875,
0.62451171875,
0.53076171875,
0.62451171875,
0.548828125,
0.51123046875,
0.548828125
],
[
0.118408203125,
0.54345703125,
0.393798828125,
0.54345703125,
0.393798828125,
0.55712890625,
0.118408203125,
0.55712890625
],
[
0.51123046875,
0.56005859375,
0.880859375,
0.55859375,
0.880859375,
0.57373046875,
0.51123046875,
0.57470703125
],
[
0.11419677734375,
0.56787109375,
0.280517578125,
0.568359375,
0.280517578125,
0.5830078125,
0.11419677734375,
0.58203125
],
[
0.51318359375,
0.57666015625,
0.8818359375,
0.57666015625,
0.8818359375,
0.59033203125,
0.51318359375,
0.59033203125
],
[
0.1175537109375,
0.58837890625,
0.486083984375,
0.58837890625,
0.486083984375,
0.6015625,
0.1175537109375,
0.6015625
],
[
0.5146484375,
0.59423828125,
0.64990234375,
0.59423828125,
0.64990234375,
0.60498046875,
0.5146484375,
0.60498046875
],
[
0.1175537109375,
0.60498046875,
0.486083984375,
0.60498046875,
0.486083984375,
0.6181640625,
0.1175537109375,
0.6181640625
],
[
0.5087890625,
0.61083984375,
0.880859375,
0.61083984375,
0.880859375,
0.62451171875,
0.5087890625,
0.62451171875
],
[
0.118408203125,
0.62158203125,
0.486083984375,
0.62158203125,
0.486083984375,
0.63525390625,
0.118408203125,
0.63525390625
],
[
0.51123046875,
0.6259765625,
0.880859375,
0.62744140625,
0.880859375,
0.642578125,
0.51123046875,
0.64013671875
],
[
0.1796875,
0.646484375,
0.489501953125,
0.646484375,
0.489501953125,
0.6611328125,
0.1796875,
0.6611328125
],
[
0.51318359375,
0.64453125,
0.60693359375,
0.64453125,
0.60693359375,
0.65625,
0.51318359375,
0.65625
],
[
0.509765625,
0.6611328125,
0.87890625,
0.6611328125,
0.87890625,
0.6748046875,
0.509765625,
0.6748046875
],
[
0.1175537109375,
0.67041015625,
0.483642578125,
0.67041015625,
0.483642578125,
0.68408203125,
0.1175537109375,
0.68408203125
],
[
0.51123046875,
0.6787109375,
0.880859375,
0.6787109375,
0.880859375,
0.6923828125,
0.51123046875,
0.6923828125
],
[
0.1175537109375,
0.68603515625,
0.484375,
0.68603515625,
0.484375,
0.69970703125,
0.1175537109375,
0.69970703125
],
[
0.509765625,
0.693359375,
0.78857421875,
0.69482421875,
0.78857421875,
0.708984375,
0.509765625,
0.7080078125
],
[
0.12005615234375,
0.70361328125,
0.484375,
0.70361328125,
0.484375,
0.71728515625,
0.12005615234375,
0.71728515625
],
[
0.51123046875,
0.71337890625,
0.880859375,
0.71337890625,
0.880859375,
0.72705078125,
0.51123046875,
0.72705078125
],
[
0.1175537109375,
0.7197265625,
0.484375,
0.71875,
0.484375,
0.73291015625,
0.1175537109375,
0.73388671875
],
[
0.51318359375,
0.72998046875,
0.880859375,
0.72998046875,
0.880859375,
0.74365234375,
0.51318359375,
0.74365234375
],
[
0.118408203125,
0.736328125,
0.486083984375,
0.736328125,
0.486083984375,
0.75,
0.118408203125,
0.75
],
[
0.51318359375,
0.74658203125,
0.779296875,
0.74658203125,
0.779296875,
0.76025390625,
0.51318359375,
0.76025390625
],
[
0.12005615234375,
0.75439453125,
0.468505859375,
0.75439453125,
0.468505859375,
0.76416015625,
0.12005615234375,
0.76416015625
],
[
0.1175537109375,
0.76708984375,
0.484375,
0.7685546875,
0.484375,
0.783203125,
0.1175537109375,
0.7822265625
],
[
0.118408203125,
0.78515625,
0.484375,
0.78515625,
0.484375,
0.798828125,
0.118408203125,
0.798828125
],
[
0.5439453125,
0.78369140625,
0.82373046875,
0.78369140625,
0.82373046875,
0.79736328125,
0.5439453125,
0.79736328125
],
[
0.8310546875,
0.7890625,
0.8447265625,
0.7890625,
0.8447265625,
0.79345703125,
0.8310546875,
0.79345703125
],
[
0.84228515625,
0.78515625,
0.880859375,
0.78515625,
0.880859375,
0.79638671875,
0.84228515625,
0.79638671875
],
[
0.1175537109375,
0.80029296875,
0.28369140625,
0.80029296875,
0.28369140625,
0.81396484375,
0.1175537109375,
0.81396484375
],
[
0.11419677734375,
0.826171875,
0.264404296875,
0.82666015625,
0.264404296875,
0.84130859375,
0.11419677734375,
0.84033203125
],
[
0.51123046875,
0.82080078125,
0.880859375,
0.82177734375,
0.880859375,
0.8359375,
0.51123046875,
0.8349609375
],
[
0.51318359375,
0.83837890625,
0.880859375,
0.83837890625,
0.880859375,
0.8515625,
0.51318359375,
0.8515625
],
[
0.1158447265625,
0.85107421875,
0.301513671875,
0.8515625,
0.301513671875,
0.8662109375,
0.1158447265625,
0.865234375
],
[
0.51318359375,
0.8525390625,
0.880859375,
0.85400390625,
0.880859375,
0.8681640625,
0.51318359375,
0.86767578125
],
[
0.118408203125,
0.87451171875,
0.486083984375,
0.87451171875,
0.486083984375,
0.88525390625,
0.118408203125,
0.88525390625
],
[
0.5146484375,
0.87158203125,
0.8212890625,
0.87158203125,
0.8212890625,
0.8818359375,
0.5146484375,
0.8818359375
],
[
0.1158447265625,
0.88720703125,
0.489501953125,
0.88818359375,
0.489501953125,
0.90283203125,
0.1158447265625,
0.90185546875
],
[
0.50732421875,
0.8896484375,
0.8818359375,
0.8896484375,
0.8818359375,
0.90283203125,
0.50732421875,
0.90283203125
],
[
0.118408203125,
0.90478515625,
0.47265625,
0.90478515625,
0.47265625,
0.91845703125,
0.118408203125,
0.91845703125
],
[
0.51123046875,
0.90673828125,
0.87744140625,
0.90673828125,
0.87744140625,
0.91748046875,
0.51123046875,
0.91748046875
]
] |
[
"Dataset",
"Language Pairs",
"Size",
"IWSLT17（Cettoloetal.，2017）我们模拟了两个",
"IWSLT",
"",
"En ( De, It, NI, Ro )",
"1.79M",
"场景。第一个（IWSLT）是English-pivot，我们",
"Nl De En It Ro",
"1.79M",
"只保留从/到英语的平行句子。第二个（IWSLT",
"PC-6",
"IWSLT-b",
"En <→1 Kk, Tr, Ro, Cs, Ru",
"7.9M",
"11.6M",
"b）具有枢轴链，其中两种语言通过枢轴语言链",
"En { De, Fr, Nl, Ru, Zh, Ar }",
"OPUS-7",
"连接。每个平移方向大约有0。22M句对。这",
"Table l: The statics of our datasets.",
"两个场景都有八个监督平移方向和十二个零拍",
"平移方向。我们使用官方验证和测试集",
"的较长句子，使它们的长度相等。由于这两个",
"PC-6从PC-32语料库中提取PC-6数据集（Lin",
"句子是翻译对，因此它们的句子长度通常很接",
"etal.,2020）。不同语言对的数据量不平衡，从",
"近，截断不会显着减少较长句子的长度，并且",
"0。12M至1。84M。此数据集有十个监督和二十",
"不会增强解码器学习较短的输出。我们将截断",
"个零拍翻译方向。我们将从WMT16???收集的",
"的句子表示为α'和y，其长度表示为n。然",
"验证和测试集用于监督方向。从WikiMatrixE",
"后我们可以生成之为：",
"(Schwenketal.，2021）中提取零触发验证和测",
"试集，每个集包含约IK???K句子对",
"z=goa+(1-g)oy,",
"",
"",
"(11)",
"OPUS-7OPUS-7数据集从OPUS-100语料库中",
"其中gE[0,1]n，O表示元素的乘积。g中的",
"提取（Zhangetal.2020a）。语言对来自不同的",
"每个元素都是从伯努利（）中采样的，其中参",
"语言家族，并且具有显着差异。此数据集有12",
"数入是从Beta（α,β）中采样的，而α和β是",
"个监督翻译方向和30个零拍翻译方向。我们",
"两个超参数。之中确定翻译方向的语言标记来",
"使用Zhangetal.（2020a）发布的标准验证和测",
"自a或y",
"试集。为方便起见，我们将零镜头测试集与相",
"目标函数与公式E9类似，我们根据模型预测",
"同的目标语言连接起来",
"分布的KL散度定义了另一个对称损失：",
"我们使用斯坦福分词器E(Tsengetal.,2005",
"n'",
"Monroeetal.2014)来分割阿拉伯语和中文，并",
"1",
"CAT=2n",
"KL(p(2|≥<,H)p(zkz<k,Hy)使用Moses工具包(Koehn et al.,2007)来标",
"k=1",
"记其他语言。此外，执行32K的集成操作以学",
"习BPEE(Sennrich et al.,2016)。",
"+ KL (p(zkz<k, Hy)llp(zkz<k, Ha)) .",
"(12)",
"4.2Systems",
"为方便起见，我们省略了模型参数",
"我们使用名为Fairseq-pyE(Ottetal.,2019）的",
"3.3The Final Loss",
"开源工具包作为我们的Transformer系统。我",
"最终损失由三部分组成，交叉摘损失（方程E",
"们实施以下系统：",
"1），基于SMD的最佳传输损失（方程E9）和",
"·零射（ZS）仅使用交叉熵损失（方程E1）训",
"基于协议的训练的KL发散损失（方程E12）：",
"练的基线系统。然后直接在零射测试集上对模",
"C=LCE+α|CoT +2CAT(13)",
"型进行测试",
"·数据透视转换（PivT）（Chengetal.，2017）与",
"两个正则化控制MAX。由于CoT是在句子级",
"ZS相同的转换模型。该模型首先将源语言转",
"别上计算的，而其他两个损失是在令牌级别",
"换为枢轴语言，然后生成目标语言",
"上计算的，因此我们将平均序列长度α乘以",
"·句子表示对齐（SRA）E（Arivazhaganetal.，",
"LoOT。在这三个损失中，第一项主导了模型的",
"2019）此方法添加了正则化损失，以最大程度",
"参数更新，并决定了模型的性能。后两个正则",
"地减少源和目标句子表示的差异。",
"化损失项仅稍微修改了梯度的方向。因为第",
"个损失项不依赖于Hy，所以我们将停止梯度",
"操作应用于Hy（图E1），这意味着梯度不会",
"L = LcE +Dis(Enc(s),Enc(t)),",
"",
"(14)",
"通过Hy到达编码器",
"4Experiments",
"其中“Dis”表示距离函数，“Enc（·）”表示句",
"子表示。我们使用平均句子表示和欧几里得距",
"4.1DataPreparation",
"离函数，因为我们发现它们的效果更好。我们",
"我们在以下多语言数据集上进行实验：",
"将超参数从0.1更改为1来调整性能",
"IWSLT17，PC-6和OPUS-7。训练集的简要统",
"·SoftmaxForcing(SF)El(Pham et al.,2019)此",
"计信息在表1中。我们把更多的细节在附录",
"方法使解码器能够通过添加额外的损失来从自"
] |
{
"arxiv_id": "2210.15851",
"language": "zh",
"page_idx": 4
}
|
||
1708.07038-zh-0
|
1708.07038-zh-0.jpg
|
Non-linear Convolution Filters for CNN-based Learning
Georgios Zoumpourlis Alexandros Doumanoglou Nicholas Vretos Petros Daras Information Technologies Institute, Center for Research and Technology Hellas, Greece
$6^{\mathrm{th}}$ km Charilaou-Thermi Road, Thessaloniki, Greece
$\{$ zoump.giorgos, aldoum, vretos, daras $\}$ @iti.gr
## 1. Introduction
卷积神经网络(CNN)已被证明可以在各种计算机视觉任务 (例如图像分类)上实现最先进的结果。正如卷积的概念，用于模仿视觉皮层中神经元的功能方面， 对于理解它们的成功至关重要。典型的卷积层是线性系统，因为它们的输出是其输入的仿射变换。由于它们的线性性质，它们缺乏表达可能的非线性的能力，这些
$\operatorname{Abstract}$
During the last years, Convolutional Neural Networks (CNNs) have achieved state-of-the-art performance in image classification. Their architectures have largely drawn inspiration by models of the primate visual system. However, while recent research results of neuroscience prove the existence of non-linear operations in the response of complex visual cells, little effort has been devoted to extend the convolution technique to non-linear forms. Typical convolutional layers are $\operatorname*{linear}$systems, hence their expressiveness is limited. To overcome this, various non-linearities have been used as activation functions inside CNNs, while also many pooling strategies have been applied. We address the issue of developing a convolution method in the context of a computational model of the visual cortex, exploring quadratic forms through the Volterra kernels. Such forms, constituting a more rich function space, are used as approximations of the response profile of visual cells. Our proposed second-order convolution is tested on CIFAR-10 and CIFAR-100. We show that a network which combines linear and non-linear filters in its convolutional layers, can outperform networks that use standard linear filters with the same architecture, yielding results competitive with the state-of-the-art on these datasets.
非线性实际上可能出现在初级视觉皮层[25]中复杂细胞的反应中。因此，我们声称他们的表现力是有限的。为了克服这个问题，各种非线性已被用作 CNN 内部的激活函数，同时也应用了许多池化策略。利用神经科学的研究成果，几乎没有努力探索将卷积技术扩展到非线性形式的新计算模型，这些模型证明了视觉细胞[31] [24]响应中存在非线性运算。人类视觉皮层的复杂性表明，关于卷积运算的应用方式，需要通过 CNN 来弥合差距。这些差距之一是对高阶模型的探索在这项工作中，我们研究了通过应用 Volterra 的理论 [32] 采用替代卷积方案来提高 CNN 的学习能力的可能性，该理论已用于研究非线性生理系统，使其适应空间域。考虑到卷积运算，而不是仅对线性项求和以计算滤波器对数据补丁的响应，我们建议还对输入数据补丁的所有元素对之间的乘法相互作用产生的非线性项进行求和。通过二阶形式转换输入，我们的目标是使它们更可分离。这样，就创建了在选择性和不变性方面具有更丰富属性的卷积滤波器。拟议工作的新颖性是：
$\bullet$ The incorporation of a biologically plausible non-
linear convolution scheme in the functionality of
CNNs
$\bullet$ The derivation of the equations that describe the
forward and backward pass during the training
process of this flter type
$\bullet$ A CUDA-based implementation of our method as a non-linear convolutional layer's module in $\mathop{\mathrm{Torch}}7^{1}[5]$
本文的其余部分组织如下：在 2 节中，概述了相关工作。在 3 节中，描述了所提出的方法，理论上基于
Hubel 和 Wiesel [13]所描述的那样，它们的体系结构 Volterra 的计算方法，并在数学上解释了训练的概念， 在很大程度上受到灵长类动物视觉系统模型的启发。而在 4 节中给出了我们方案的实际实现的描述。在 5
节中绘制了 CIFAR-10 和 CIFAR-100 数据集的实验结
果，最后在 6 节中总结了该论文
$^{1}$http://torch.ch/
|
[
[
0.205078125,
0.1319580078125,
0.765625,
0.1337890625,
0.765625,
0.1514892578125,
0.205078125,
0.149658203125
],
[
0.1298828125,
0.1925048828125,
0.84326171875,
0.1925048828125,
0.84326171875,
0.20703125,
0.1298828125,
0.20703125
],
[
0.1217041015625,
0.21337890625,
0.8486328125,
0.21337890625,
0.8486328125,
0.2279052734375,
0.1217041015625,
0.2279052734375
],
[
0.261474609375,
0.23291015625,
0.708984375,
0.2342529296875,
0.708984375,
0.248779296875,
0.261474609375,
0.2474365234375
],
[
0.325927734375,
0.25830078125,
0.6435546875,
0.25830078125,
0.6435546875,
0.271484375,
0.325927734375,
0.271484375
],
[
0.238525390625,
0.30810546875,
0.31201171875,
0.30810546875,
0.31201171875,
0.322509765625,
0.238525390625,
0.322509765625
],
[
0.50341796875,
0.30810546875,
0.890625,
0.309326171875,
0.890625,
0.323974609375,
0.50341796875,
0.322509765625
],
[
0.5048828125,
0.324462890625,
0.8896484375,
0.324462890625,
0.8896484375,
0.33837890625,
0.5048828125,
0.33837890625
],
[
0.1004638671875,
0.341552734375,
0.46728515625,
0.341552734375,
0.46728515625,
0.354736328125,
0.1004638671875,
0.354736328125
],
[
0.50634765625,
0.341552734375,
0.8896484375,
0.341552734375,
0.8896484375,
0.3515625,
0.50634765625,
0.3515625
],
[
0.08251953125,
0.356689453125,
0.46728515625,
0.356689453125,
0.46728515625,
0.37060546875,
0.08251953125,
0.37060546875
],
[
0.5048828125,
0.354736328125,
0.890625,
0.354736328125,
0.890625,
0.36865234375,
0.5048828125,
0.36865234375
],
[
0.08087158203125,
0.371826171875,
0.468017578125,
0.37060546875,
0.468017578125,
0.385009765625,
0.08087158203125,
0.386474609375
],
[
0.50341796875,
0.37060546875,
0.890625,
0.37060546875,
0.890625,
0.3837890625,
0.50341796875,
0.3837890625
],
[
0.080078125,
0.386962890625,
0.46728515625,
0.386962890625,
0.46728515625,
0.40087890625,
0.080078125,
0.40087890625
],
[
0.50341796875,
0.385009765625,
0.890625,
0.385009765625,
0.890625,
0.3984375,
0.50341796875,
0.3984375
],
[
0.08251953125,
0.404052734375,
0.468017578125,
0.404052734375,
0.468017578125,
0.4140625,
0.08251953125,
0.4140625
],
[
0.5048828125,
0.40087890625,
0.890625,
0.40087890625,
0.890625,
0.4140625,
0.5048828125,
0.4140625
],
[
0.08251953125,
0.419189453125,
0.464111328125,
0.419189453125,
0.464111328125,
0.429931640625,
0.08251953125,
0.429931640625
],
[
0.50341796875,
0.4140625,
0.890625,
0.415283203125,
0.890625,
0.429931640625,
0.50341796875,
0.4287109375
],
[
0.08087158203125,
0.43310546875,
0.46728515625,
0.43310546875,
0.46728515625,
0.4462890625,
0.08087158203125,
0.4462890625
],
[
0.50341796875,
0.4287109375,
0.890625,
0.429931640625,
0.890625,
0.4443359375,
0.50341796875,
0.443115234375
],
[
0.080078125,
0.447509765625,
0.46728515625,
0.447509765625,
0.46728515625,
0.4609375,
0.080078125,
0.4609375
],
[
0.50341796875,
0.44580078125,
0.8896484375,
0.44580078125,
0.8896484375,
0.458984375,
0.50341796875,
0.458984375
],
[
0.08087158203125,
0.462158203125,
0.46728515625,
0.462158203125,
0.46728515625,
0.47607421875,
0.08087158203125,
0.47607421875
],
[
0.5048828125,
0.460205078125,
0.890625,
0.460205078125,
0.890625,
0.473388671875,
0.5048828125,
0.473388671875
],
[
0.080078125,
0.477783203125,
0.468017578125,
0.477783203125,
0.468017578125,
0.4912109375,
0.080078125,
0.4912109375
],
[
0.5048828125,
0.47607421875,
0.890625,
0.47607421875,
0.890625,
0.4892578125,
0.5048828125,
0.4892578125
],
[
0.08087158203125,
0.492431640625,
0.468017578125,
0.492431640625,
0.468017578125,
0.505859375,
0.08087158203125,
0.505859375
],
[
0.5048828125,
0.4912109375,
0.890625,
0.4912109375,
0.890625,
0.5048828125,
0.5048828125,
0.5048828125
],
[
0.08087158203125,
0.50830078125,
0.46728515625,
0.50830078125,
0.46728515625,
0.521484375,
0.08087158203125,
0.521484375
],
[
0.5048828125,
0.505859375,
0.890625,
0.505859375,
0.890625,
0.51953125,
0.5048828125,
0.51953125
],
[
0.080078125,
0.52294921875,
0.468017578125,
0.521484375,
0.468017578125,
0.5361328125,
0.080078125,
0.537109375
],
[
0.5048828125,
0.52099609375,
0.890625,
0.52099609375,
0.890625,
0.5341796875,
0.5048828125,
0.5341796875
],
[
0.08087158203125,
0.537109375,
0.465576171875,
0.537109375,
0.465576171875,
0.55029296875,
0.08087158203125,
0.55029296875
],
[
0.5048828125,
0.5361328125,
0.890625,
0.5361328125,
0.890625,
0.5498046875,
0.5048828125,
0.5498046875
],
[
0.08087158203125,
0.55322265625,
0.465576171875,
0.55322265625,
0.465576171875,
0.56640625,
0.08087158203125,
0.56640625
],
[
0.5048828125,
0.5517578125,
0.8896484375,
0.5517578125,
0.8896484375,
0.56494140625,
0.5048828125,
0.56494140625
],
[
0.08087158203125,
0.5673828125,
0.465576171875,
0.5673828125,
0.465576171875,
0.58056640625,
0.08087158203125,
0.58056640625
],
[
0.50634765625,
0.568359375,
0.80224609375,
0.568359375,
0.80224609375,
0.5791015625,
0.50634765625,
0.5791015625
],
[
0.08087158203125,
0.583984375,
0.465576171875,
0.583984375,
0.465576171875,
0.59765625,
0.08087158203125,
0.59765625
],
[
0.08087158203125,
0.5986328125,
0.4697265625,
0.5986328125,
0.4697265625,
0.6123046875,
0.08087158203125,
0.6123046875
],
[
0.5146484375,
0.59326171875,
0.88623046875,
0.5947265625,
0.88623046875,
0.6123046875,
0.5146484375,
0.611328125
],
[
0.08087158203125,
0.6142578125,
0.468017578125,
0.6142578125,
0.468017578125,
0.62744140625,
0.08087158203125,
0.62744140625
],
[
0.53515625,
0.611328125,
0.89208984375,
0.611328125,
0.89208984375,
0.625,
0.53515625,
0.625
],
[
0.08087158203125,
0.62890625,
0.468017578125,
0.62890625,
0.468017578125,
0.642578125,
0.08087158203125,
0.642578125
],
[
0.53515625,
0.626953125,
0.5810546875,
0.626953125,
0.5810546875,
0.6416015625,
0.53515625,
0.6416015625
],
[
0.08087158203125,
0.64306640625,
0.468017578125,
0.64306640625,
0.468017578125,
0.6572265625,
0.08087158203125,
0.6572265625
],
[
0.080078125,
0.6591796875,
0.468017578125,
0.6591796875,
0.468017578125,
0.67236328125,
0.080078125,
0.67236328125
],
[
0.51708984375,
0.6591796875,
0.890625,
0.6591796875,
0.890625,
0.67236328125,
0.51708984375,
0.67236328125
],
[
0.08087158203125,
0.67236328125,
0.46728515625,
0.673828125,
0.46728515625,
0.68798828125,
0.08087158203125,
0.6875
],
[
0.53369140625,
0.673828125,
0.89208984375,
0.6748046875,
0.89208984375,
0.689453125,
0.53369140625,
0.68798828125
],
[
0.08087158203125,
0.689453125,
0.468017578125,
0.689453125,
0.468017578125,
0.70263671875,
0.08087158203125,
0.70263671875
],
[
0.53369140625,
0.6875,
0.71826171875,
0.68798828125,
0.71826171875,
0.70556640625,
0.53369140625,
0.705078125
],
[
0.080078125,
0.70263671875,
0.188720703125,
0.7041015625,
0.188720703125,
0.71826171875,
0.080078125,
0.71728515625
],
[
0.5185546875,
0.7216796875,
0.89208984375,
0.7216796875,
0.89208984375,
0.73486328125,
0.5185546875,
0.73486328125
],
[
0.5361328125,
0.7373046875,
0.890625,
0.7373046875,
0.890625,
0.75048828125,
0.5361328125,
0.75048828125
],
[
0.080078125,
0.75390625,
0.21728515625,
0.7548828125,
0.21728515625,
0.76953125,
0.080078125,
0.76806640625
],
[
0.53369140625,
0.75,
0.62353515625,
0.751953125,
0.62353515625,
0.7666015625,
0.53369140625,
0.7646484375
],
[
0.101318359375,
0.7822265625,
0.468017578125,
0.7822265625,
0.468017578125,
0.79541015625,
0.101318359375,
0.79541015625
],
[
0.521484375,
0.78076171875,
0.8896484375,
0.7822265625,
0.8896484375,
0.796875,
0.521484375,
0.79541015625
],
[
0.08087158203125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.5048828125,
0.79736328125,
0.8896484375,
0.79736328125,
0.8896484375,
0.81103515625,
0.5048828125,
0.81103515625
],
[
0.08087158203125,
0.81103515625,
0.46728515625,
0.8125,
0.46728515625,
0.82373046875,
0.08087158203125,
0.82275390625
],
[
0.50341796875,
0.8125,
0.88818359375,
0.8125,
0.88818359375,
0.82568359375,
0.50341796875,
0.82568359375
],
[
0.08087158203125,
0.8271484375,
0.4697265625,
0.8271484375,
0.4697265625,
0.84033203125,
0.08087158203125,
0.84033203125
],
[
0.50341796875,
0.82763671875,
0.89208984375,
0.82763671875,
0.89208984375,
0.84130859375,
0.50341796875,
0.84130859375
],
[
0.08251953125,
0.84228515625,
0.468017578125,
0.84228515625,
0.468017578125,
0.85595703125,
0.08251953125,
0.85595703125
],
[
0.5048828125,
0.84228515625,
0.890625,
0.84228515625,
0.890625,
0.85595703125,
0.5048828125,
0.85595703125
],
[
0.080078125,
0.85595703125,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.869140625
],
[
0.50244140625,
0.85595703125,
0.736328125,
0.85693359375,
0.736328125,
0.8720703125,
0.50244140625,
0.87060546875
],
[
0.08251953125,
0.87255859375,
0.468017578125,
0.87255859375,
0.468017578125,
0.88623046875,
0.08251953125,
0.88623046875
],
[
0.08087158203125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.51953125,
0.88720703125,
0.62744140625,
0.88720703125,
0.62744140625,
0.9013671875,
0.51953125,
0.9013671875
]
] |
[
"Non-linear Convolution Filters for CNN-based Learning",
"Georgios ZoumpourlisAlexandros DoumanoglouNicholas VretosPetros Daras",
"Information Technologies Institute, Center for Research and Technology Hellas, Greece",
"6thkm Charilaou-ThermiRoad, Thessaloniki,Greece",
"[ zoump.giorgos, aldoum, vretos, daras ↓ @iti.gr",
"Abstract",
"非线性实际上可能出现在初级视觉皮层[25]】中复杂细",
"胞的反应中。因此，我们声称他们的表现力是有限的",
"During the last years, Convolutional Neural Net-",
"为了克服这个问题，各种非线性已被用作CNN内部的",
"works (CNNs) have achieved state-of-the-art perfor-",
"激活函数，同时也应用了许多池化策略。利用神经科学",
"mance in image classification. Their architectures have",
"的研究成果，几乎没有努力探索将卷积技术扩展到非",
"largely drawn inspiration by models of the primate vi-",
"线性形式的新计算模型，这些模型证明了视觉细胞[31]",
"sual system.However,while recent research results of",
"[24]响应中存在非线性运算。人类视觉皮层的复杂性表",
"neuroscience prove the existence of non-linear opera,",
"明，关于卷积运算的应用方式，需要通过CNN来弥合",
"tions in the response of complex visual cells, little effort",
"差距。这些差距之一是对高阶模型的探索在这项工作",
"has been devoted to extend the convolution technique",
"中，我们研究了通过应用Volterra的理论[32]采用替",
"to non-linear forms. Typical convolutional layers are",
"代卷积方案来提高CNN的学习能力的可能性，该理论",
"linear systems, hence their expressiveness is limited. To",
"已用于研究非线性生理系统，使其适应空间域。考虑",
"overcome this, various non-linearities have been used",
"到卷积运算，而不是仅对线性项求和以计算滤波器对",
"as activation functions inside CNNs, while also many",
"数据补丁的响应，我们建议还对输入数据补丁的所有",
"pooling strategies have been applied. We address the",
"元素对之间的乘法相互作用产生的非线性项进行求和。",
"issue of developing a convolution method in the con",
"通过二阶形式转换输入，我们的目标是使它们更可分",
"text of a computational model of the visual cortex, ex",
"离。这样，就创建了在选择性和不变性方面具有更丰富",
"ploring quadratic forms through the Volterra kernels",
"属性的卷积滤波器。拟议工作的新颖性是：",
"Such forms, constituting a more rich function space.",
"are used as approximations of the response profile of",
". The incorporation of a biologically plausible non",
"visual cells. Our proposed second-order convolution is",
"linear convolution scheme in the functionality of",
"tested on CIFAR-10 and CIFAR-100. We show that a",
"CNNS",
"network which combines linear and non-linear filters in",
"its convolutional layers, can outperform networks that",
". The derivation of the equations that describe the",
"use standard linear filters with the same architecture,",
"forward and backward pass during the training",
"yielding results competitive with the state-of-the-art on",
"process of this filter type",
"these datasets.",
": A CUDA-based implementation of our method",
"as a non-linear convolutional layer's module in",
"1. Introduction",
"Torch7 1 [5]",
"卷积神经网络（CNN）已被证明可以在各种计算机",
"本文的其余部分组织如下：在2节中，概述了相关",
"视觉任务（例如图像分类）上实现最先进的结果。正如",
"工作。在3节中，描述了所提出的方法，理论上基于",
"Hubel和Wiesel「13]所描述的那样，它们的体系结构",
"Volterra的计算方法，并在数学上解释了训练的概念",
"在很大程度上受到灵长类动物视觉系统模型的启发。。",
"而在4节中给出了我们方案的实际实现的描述。在5",
"卷积的概念，用于模仿视觉皮层中神经元的功能方面，",
"节中绘制了CIFAR-10和CIFAR-100数据集的实验结",
"对于理解它们的成功至关重要。典型的卷积层是线性",
"果，最后在6节中总结了该论文",
"系统，因为它们的输出是其输人的仿射变换。由于它们",
"的线性性质，它们缺乏表达可能的非线性的能力，这些",
"1http://torch.ch/"
] |
{
"arxiv_id": "1708.07038",
"language": "zh",
"page_idx": 0
}
|
||
1708.07038-zh-1
|
1708.07038-zh-1.jpg
|
# 2. Related Work
第一个受生物启发的神经网络之一是福岛的Neocognitron [8] ,它是 CNN 的前身，由 LeCun 等人介绍。在[6]中。卷积层是 CNN 的核心构建块。CNN 的早期实现在其卷积层中使用了预定义的 Gabor 滤波器。这类过滤器可以非常准确地模拟在初级视觉皮层(V1)[21] 中发现的简单细胞的特性。这种类型的视觉细胞具有响应曲线，其特征在于感受场内的空间求和。寻找简单细胞的最佳空间刺激[7] 是一个基于其兴奋区和抑制区[23]的空间排列的过程。然而，这并不适用于复杂的视觉细胞。此外，我们无法通过找到它们的最佳刺激来准确描述它们的特性。到目前为止，大多数CNN 实现都忽略了这一事实，因为它们已经适应了线性类型的卷积滤波器，试图在其功能中应用定量而非定性的变化。他等人。[10]提出了残差网络(ResNets), 该网络具有与其正常卷积层平行的快捷方式连接，作为增加模型参数时消失/爆炸梯度和硬优化问题的解决方案 (i。e。添加更多图层)。Zagoruyko&Komodakis [34]表明，宽 ResNets 可以通过层数超过 ResNets, 从而将兴趣转移到增加每层过滤器的数量上。另外，对于专注于创建具有更多卷积层或更多过滤器的网络的工作，我们评估使用非线性和线性项作为卷积内核的近似值以提高 CNN 性能的影响除了 ResNet 之外，ImageNetChallenge [27]还通过以新方式使用卷积层的方法实现了非常低的错误率，从而增强了它们的表示能力。Lin 等人。[20] 提出了“网络中的网络(NIN)”, 作为对典型过滤器存在的低抽象级别的补救。他们建立了具有更复杂结构的微神经网络，而不是传统的卷积滤波器，后者是一种广义的线性模型，可以在感受场内提取数据。为了将输入数据映射到输出，他们使用多层感知器作为非线性函数逼近器，他们将其称为“ mlpconv”层。通过以与 CNN 类似的方式在输人上滑动微网络来获得输出特征图。Szegedy 等人。[30] 以“ Inception 模块”的形式引入了新的组织级别，该模块使用可变大小的过滤器来捕获不同大小的不同视觉模式，并逼近最佳稀疏结构。谢等人。[33] 提出了一种利用“Inception”模型的拆分-转换-合并策略的方法， 该方法执行一组转换，每个转换都在低维嵌入上，其输出通过求和进行聚合。[19]的作者基于大脑中大量的递归突触，提出了使用递归神经网络进行图像分类的方法。他们证明，与前馈架构相比，在卷积层中插人循环连接可提高结果。他们的工作是将源自神经科学的机制在生物学上合理地整合到 CNN 中在 [28] 中，提出了一种玻尔兹曼学习算法，其中特征交互用于将隐藏单元变成高阶特征检测器。在[22]中，提出了一种将这种学习算法应用于高阶玻尔兹曼机的有效方法，使它们在计算上易于处理实际问题。在[1]中，Bergstra , 线性滤波器的输入输出函数为：
等人。创建了一个神经激活模型，该模型通过合并二阶相互作用并使用替代非线性作为激活函数，在数据集上显示出改进的泛化。在[2]中，尝试将二次形式分析和解释为接受场。在他们的研究中，发现二次形式可用于建模非线性感受野，因为它们遵循初级视觉皮层中复杂细胞的某些特性。这些特性包括对边缘的响应，
相移不变性，方向选择性，非正交抑制，末端抑制和侧抑制。与标准线性形式相反，在二次和其他非线性形式中，最佳刺激不能提供其属性的完整描述。结果表明对于最佳刺激，没有不变性，而对于其他一般的次优刺激，可能存在许多不变性，这些不变性可能很大，但缺乏简单的解释。。虽然最优刺激与滤波器的不变性无关， 但它的邻域是在更宽松的变换不变性下研究的。结果表明，适当的二次形式可以证明相移和方向变化的不变性。从前面的讨论中，我们得出的结论是，将非线性形式用于卷积层可能是计算机视觉的合理未来方向
# $3.{\mathrm{Proposed~Method}}$
如前所述，所提出的方法利用 Volterra 内核理论来提供利用在感受场中发生的非线性运算的方法。到目前为止，据我们所知，非线性主要是通过激活函数和CNN 不同层之间的池化操作来利用的。然而，这种非线性可能是编码视觉系统内部过程的一种方法，而不是存在于感受野区域中的方法。
我们的方法遵循 CNN 的典型工作流程，通过排列不同目的的层(卷积，池化，激活函数，批处理归一化， 退出，完全连接等。),而非线性卷积层实际上可以在所有现有架构中插入。然而，由于涉及的可训练参数的增加，应注意整个过程的复杂性。为此，还提供了4节中的 CUDA 实现
# $3. 1. \textbf{ Volterra- based convolution}$
Volterra 级数模型是连续函数的近似序列，用于表示非线性动力学系统的输入输出关系，使用多项式函数展开。它们的方程式可以由无限阶项组成，但是基于它们的实际实现使用截断版本，保留了高达$r$阶的项。与线性卷积类似，基于 Volterra 的卷积使用内核来过滤输入数据。一阶 Volterra 内核包含滤波器线性部的系数。二阶核表示两个输入元素之间的二次相互作用的系数。通常，$r$阶的内核表示$r$输人元素之间的非线性交互对响应的权重。在计算机视觉领域，Volterra 内核以前已在 [17] 中用于人脸识别，有效地用作非线性函数的近似值
# 3.2. Forward pass
对于我们提出的卷积，我们采用了二阶 Volterra 级数。给定一个带有$n$元素$(n=k_h\cdot k_w)$ $I\in\mathbb{R}^{k_h\times k_w}$,并将其整形为向量
$$\boldsymbol{x}=\begin{bmatrix}x_1&x_2&\cdots&x_n\end{bmatrix}^T$$
(1)
(2)
$$y(\boldsymbol{x})=\sum_{i=1}^n\begin{pmatrix}w_1^ix_i\end{pmatrix}+b$$
其中$w_{1}^{i}$是 XMATHX 的权重卷
此函数以以下二次形式扩展：
|
[
[
0.080078125,
0.09027099609375,
0.226318359375,
0.09027099609375,
0.226318359375,
0.10479736328125,
0.080078125,
0.10479736328125
],
[
0.5048828125,
0.09344482421875,
0.8896484375,
0.09344482421875,
0.8896484375,
0.106689453125,
0.5048828125,
0.106689453125
],
[
0.5048828125,
0.10797119140625,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.5048828125,
0.121826171875
],
[
0.1004638671875,
0.117431640625,
0.468017578125,
0.117431640625,
0.468017578125,
0.1307373046875,
0.1004638671875,
0.1307373046875
],
[
0.08087158203125,
0.1307373046875,
0.465576171875,
0.1307373046875,
0.465576171875,
0.14453125,
0.08087158203125,
0.14453125
],
[
0.50634765625,
0.125,
0.88623046875,
0.125,
0.88623046875,
0.1351318359375,
0.50634765625,
0.1351318359375
],
[
0.5048828125,
0.1383056640625,
0.8896484375,
0.1383056640625,
0.8896484375,
0.1514892578125,
0.5048828125,
0.1514892578125
],
[
0.08251953125,
0.146484375,
0.468017578125,
0.146484375,
0.468017578125,
0.160400390625,
0.08251953125,
0.160400390625
],
[
0.50634765625,
0.1546630859375,
0.8896484375,
0.1546630859375,
0.8896484375,
0.1654052734375,
0.50634765625,
0.1654052734375
],
[
0.08251953125,
0.1622314453125,
0.46728515625,
0.1622314453125,
0.46728515625,
0.175537109375,
0.08251953125,
0.175537109375
],
[
0.50341796875,
0.1685791015625,
0.88818359375,
0.1685791015625,
0.88818359375,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.08251953125,
0.1787109375,
0.46728515625,
0.1787109375,
0.46728515625,
0.189453125,
0.08251953125,
0.189453125
],
[
0.5048828125,
0.18310546875,
0.88818359375,
0.18310546875,
0.88818359375,
0.1962890625,
0.5048828125,
0.1962890625
],
[
0.08087158203125,
0.1925048828125,
0.468017578125,
0.1925048828125,
0.468017578125,
0.205810546875,
0.08087158203125,
0.205810546875
],
[
0.5048828125,
0.199462890625,
0.88818359375,
0.199462890625,
0.88818359375,
0.210205078125,
0.5048828125,
0.210205078125
],
[
0.08251953125,
0.2083740234375,
0.4697265625,
0.2083740234375,
0.4697265625,
0.2215576171875,
0.08251953125,
0.2215576171875
],
[
0.5048828125,
0.21337890625,
0.89208984375,
0.21337890625,
0.89208984375,
0.2266845703125,
0.5048828125,
0.2266845703125
],
[
0.0849609375,
0.2247314453125,
0.465576171875,
0.2247314453125,
0.465576171875,
0.23486328125,
0.0849609375,
0.23486328125
],
[
0.50244140625,
0.2291259765625,
0.861328125,
0.2291259765625,
0.861328125,
0.242431640625,
0.50244140625,
0.242431640625
],
[
0.08251953125,
0.238037109375,
0.46728515625,
0.238037109375,
0.46728515625,
0.251953125,
0.08251953125,
0.251953125
],
[
0.08416748046875,
0.25244140625,
0.46728515625,
0.25244140625,
0.46728515625,
0.266357421875,
0.08416748046875,
0.266357421875
],
[
0.50244140625,
0.2587890625,
0.68310546875,
0.2587890625,
0.68310546875,
0.272705078125,
0.50244140625,
0.272705078125
],
[
0.08251953125,
0.268310546875,
0.46728515625,
0.268310546875,
0.46728515625,
0.2822265625,
0.08251953125,
0.2822265625
],
[
0.08251953125,
0.2841796875,
0.46728515625,
0.2841796875,
0.46728515625,
0.29736328125,
0.08251953125,
0.29736328125
],
[
0.52392578125,
0.2822265625,
0.8896484375,
0.28271484375,
0.8896484375,
0.29736328125,
0.52392578125,
0.296630859375
],
[
0.08087158203125,
0.296630859375,
0.468017578125,
0.29736328125,
0.468017578125,
0.311279296875,
0.08087158203125,
0.31005859375
],
[
0.50341796875,
0.298583984375,
0.88818359375,
0.298583984375,
0.88818359375,
0.3125,
0.50341796875,
0.3125
],
[
0.08251953125,
0.313232421875,
0.4697265625,
0.313232421875,
0.4697265625,
0.326904296875,
0.08251953125,
0.326904296875
],
[
0.5048828125,
0.314453125,
0.890625,
0.314453125,
0.890625,
0.32763671875,
0.5048828125,
0.32763671875
],
[
0.08416748046875,
0.330078125,
0.46728515625,
0.330078125,
0.46728515625,
0.34033203125,
0.08416748046875,
0.34033203125
],
[
0.50341796875,
0.328857421875,
0.890625,
0.328857421875,
0.890625,
0.34228515625,
0.50341796875,
0.34228515625
],
[
0.08416748046875,
0.34521484375,
0.46728515625,
0.34521484375,
0.46728515625,
0.35595703125,
0.08416748046875,
0.35595703125
],
[
0.5048828125,
0.3447265625,
0.890625,
0.3447265625,
0.890625,
0.35791015625,
0.5048828125,
0.35791015625
],
[
0.08416748046875,
0.359130859375,
0.468017578125,
0.359130859375,
0.468017578125,
0.37255859375,
0.08416748046875,
0.37255859375
],
[
0.5048828125,
0.359130859375,
0.72119140625,
0.359130859375,
0.72119140625,
0.37255859375,
0.5048828125,
0.37255859375
],
[
0.08251953125,
0.373779296875,
0.46728515625,
0.373779296875,
0.46728515625,
0.386962890625,
0.08251953125,
0.386962890625
],
[
0.52392578125,
0.375,
0.8896484375,
0.375,
0.8896484375,
0.38818359375,
0.52392578125,
0.38818359375
],
[
0.08251953125,
0.389404296875,
0.468017578125,
0.389404296875,
0.468017578125,
0.40283203125,
0.08251953125,
0.40283203125
],
[
0.5048828125,
0.389404296875,
0.890625,
0.389404296875,
0.890625,
0.40283203125,
0.5048828125,
0.40283203125
],
[
0.08416748046875,
0.404052734375,
0.468017578125,
0.404052734375,
0.468017578125,
0.417236328125,
0.08416748046875,
0.417236328125
],
[
0.5048828125,
0.404052734375,
0.890625,
0.404052734375,
0.890625,
0.417236328125,
0.5048828125,
0.417236328125
],
[
0.08416748046875,
0.42041015625,
0.46728515625,
0.42041015625,
0.46728515625,
0.43115234375,
0.08416748046875,
0.43115234375
],
[
0.50634765625,
0.42041015625,
0.88818359375,
0.42041015625,
0.88818359375,
0.43115234375,
0.50634765625,
0.43115234375
],
[
0.08251953125,
0.43505859375,
0.465576171875,
0.43505859375,
0.465576171875,
0.448974609375,
0.08251953125,
0.448974609375
],
[
0.5048828125,
0.43505859375,
0.8896484375,
0.43505859375,
0.8896484375,
0.448974609375,
0.5048828125,
0.448974609375
],
[
0.08087158203125,
0.449462890625,
0.46728515625,
0.449462890625,
0.46728515625,
0.46337890625,
0.08087158203125,
0.46337890625
],
[
0.50341796875,
0.449462890625,
0.6298828125,
0.449462890625,
0.6298828125,
0.46337890625,
0.50341796875,
0.46337890625
],
[
0.08251953125,
0.466552734375,
0.46728515625,
0.466552734375,
0.46728515625,
0.4765625,
0.08251953125,
0.4765625
],
[
0.08251953125,
0.479248046875,
0.44677734375,
0.479248046875,
0.44677734375,
0.492431640625,
0.08251953125,
0.492431640625
],
[
0.50341796875,
0.474853515625,
0.74853515625,
0.474853515625,
0.74853515625,
0.488037109375,
0.50341796875,
0.488037109375
],
[
0.443603515625,
0.484130859375,
0.453369140625,
0.484130859375,
0.453369140625,
0.488037109375,
0.443603515625,
0.488037109375
],
[
0.08087158203125,
0.494384765625,
0.468017578125,
0.494384765625,
0.468017578125,
0.50830078125,
0.08087158203125,
0.50830078125
],
[
0.521484375,
0.496826171875,
0.890625,
0.49755859375,
0.890625,
0.51220703125,
0.521484375,
0.51123046875
],
[
0.08251953125,
0.51025390625,
0.46728515625,
0.51025390625,
0.46728515625,
0.5234375,
0.08251953125,
0.5234375
],
[
0.5048828125,
0.51318359375,
0.8896484375,
0.51318359375,
0.8896484375,
0.5263671875,
0.5048828125,
0.5263671875
],
[
0.08251953125,
0.5244140625,
0.468017578125,
0.5244140625,
0.468017578125,
0.53857421875,
0.08251953125,
0.53857421875
],
[
0.5048828125,
0.52880859375,
0.8896484375,
0.52880859375,
0.8896484375,
0.54248046875,
0.5048828125,
0.54248046875
],
[
0.08087158203125,
0.54052734375,
0.468017578125,
0.54052734375,
0.468017578125,
0.5537109375,
0.08087158203125,
0.5537109375
],
[
0.50634765625,
0.54541015625,
0.8896484375,
0.54541015625,
0.8896484375,
0.55615234375,
0.50634765625,
0.55615234375
],
[
0.08251953125,
0.55615234375,
0.468017578125,
0.55615234375,
0.468017578125,
0.5693359375,
0.08251953125,
0.5693359375
],
[
0.5048828125,
0.556640625,
0.890625,
0.55810546875,
0.890625,
0.57275390625,
0.5048828125,
0.5712890625
],
[
0.08087158203125,
0.5693359375,
0.46728515625,
0.568359375,
0.46728515625,
0.58349609375,
0.08087158203125,
0.583984375
],
[
0.5048828125,
0.57568359375,
0.88818359375,
0.57568359375,
0.88818359375,
0.5859375,
0.5048828125,
0.5859375
],
[
0.08251953125,
0.5869140625,
0.46728515625,
0.5869140625,
0.46728515625,
0.59765625,
0.08251953125,
0.59765625
],
[
0.50244140625,
0.5869140625,
0.890625,
0.58837890625,
0.890625,
0.60302734375,
0.50244140625,
0.6015625
],
[
0.0882568359375,
0.599609375,
0.46728515625,
0.599609375,
0.46728515625,
0.61279296875,
0.0882568359375,
0.61279296875
],
[
0.50341796875,
0.60302734375,
0.8896484375,
0.60302734375,
0.8896484375,
0.6162109375,
0.50341796875,
0.6162109375
],
[
0.08251953125,
0.6171875,
0.465576171875,
0.6171875,
0.465576171875,
0.62744140625,
0.08251953125,
0.62744140625
],
[
0.50341796875,
0.61865234375,
0.890625,
0.61865234375,
0.890625,
0.6318359375,
0.50341796875,
0.6318359375
],
[
0.08251953125,
0.630859375,
0.468017578125,
0.630859375,
0.468017578125,
0.64453125,
0.08251953125,
0.64453125
],
[
0.5048828125,
0.6337890625,
0.890625,
0.6337890625,
0.890625,
0.64794921875,
0.5048828125,
0.64794921875
],
[
0.08251953125,
0.64599609375,
0.468017578125,
0.64599609375,
0.468017578125,
0.6591796875,
0.08251953125,
0.6591796875
],
[
0.5009765625,
0.64794921875,
0.619140625,
0.6484375,
0.619140625,
0.66357421875,
0.5009765625,
0.662109375
],
[
0.08416748046875,
0.662109375,
0.465576171875,
0.662109375,
0.465576171875,
0.67236328125,
0.08416748046875,
0.67236328125
],
[
0.08251953125,
0.67529296875,
0.46728515625,
0.67529296875,
0.46728515625,
0.689453125,
0.08251953125,
0.689453125
],
[
0.50341796875,
0.6748046875,
0.64794921875,
0.6748046875,
0.64794921875,
0.68798828125,
0.50341796875,
0.68798828125
],
[
0.08251953125,
0.6923828125,
0.464111328125,
0.6923828125,
0.464111328125,
0.70263671875,
0.08251953125,
0.70263671875
],
[
0.52392578125,
0.69775390625,
0.890625,
0.69775390625,
0.890625,
0.7109375,
0.52392578125,
0.7109375
],
[
0.08251953125,
0.70849609375,
0.46728515625,
0.70849609375,
0.46728515625,
0.71826171875,
0.08251953125,
0.71826171875
],
[
0.5048828125,
0.71337890625,
0.88623046875,
0.71337890625,
0.88623046875,
0.7265625,
0.5048828125,
0.7265625
],
[
0.08416748046875,
0.72265625,
0.46728515625,
0.72265625,
0.46728515625,
0.73291015625,
0.08416748046875,
0.73291015625
],
[
0.5009765625,
0.7216796875,
0.8203125,
0.724609375,
0.8203125,
0.74560546875,
0.5009765625,
0.7421875
],
[
0.08251953125,
0.736328125,
0.46728515625,
0.736328125,
0.46728515625,
0.75,
0.08251953125,
0.75
],
[
0.08416748046875,
0.75048828125,
0.46728515625,
0.75048828125,
0.46728515625,
0.7646484375,
0.08416748046875,
0.7646484375
],
[
0.6044921875,
0.76025390625,
0.77294921875,
0.75830078125,
0.77294921875,
0.77978515625,
0.6044921875,
0.7822265625
],
[
0.86669921875,
0.76123046875,
0.89208984375,
0.76123046875,
0.89208984375,
0.77880859375,
0.86669921875,
0.77880859375
],
[
0.0849609375,
0.76806640625,
0.465576171875,
0.76806640625,
0.465576171875,
0.77880859375,
0.0849609375,
0.77880859375
],
[
0.08087158203125,
0.78076171875,
0.465576171875,
0.78076171875,
0.465576171875,
0.79443359375,
0.08087158203125,
0.79443359375
],
[
0.08251953125,
0.796875,
0.46728515625,
0.796875,
0.46728515625,
0.81005859375,
0.08251953125,
0.81005859375
],
[
0.5048828125,
0.79345703125,
0.73779296875,
0.79345703125,
0.73779296875,
0.806640625,
0.5048828125,
0.806640625
],
[
0.08251953125,
0.81103515625,
0.465576171875,
0.81103515625,
0.465576171875,
0.82470703125,
0.08251953125,
0.82470703125
],
[
0.08251953125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.08251953125,
0.84033203125
],
[
0.67626953125,
0.8212890625,
0.69287109375,
0.8212890625,
0.69287109375,
0.830078125,
0.67626953125,
0.830078125
],
[
0.61669921875,
0.830078125,
0.77783203125,
0.8291015625,
0.77783203125,
0.8466796875,
0.61669921875,
0.84765625
],
[
0.86767578125,
0.83203125,
0.890625,
0.83203125,
0.890625,
0.8466796875,
0.86767578125,
0.8466796875
],
[
0.08660888671875,
0.84326171875,
0.465576171875,
0.84326171875,
0.465576171875,
0.85400390625,
0.08660888671875,
0.85400390625
],
[
0.6748046875,
0.84228515625,
0.69775390625,
0.8466796875,
0.69287109375,
0.8623046875,
0.6689453125,
0.85693359375
],
[
0.08087158203125,
0.85693359375,
0.465576171875,
0.85595703125,
0.465576171875,
0.869140625,
0.08087158203125,
0.87060546875
],
[
0.08251953125,
0.87255859375,
0.468017578125,
0.87255859375,
0.468017578125,
0.88623046875,
0.08251953125,
0.88623046875
],
[
0.50341796875,
0.87060546875,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50341796875,
0.88525390625
],
[
0.08416748046875,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.08416748046875,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.705078125,
0.88720703125,
0.705078125,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"2. Related Work",
"相移不变性，方向选择性，非正交抑制，末端抑制和侧",
"抑制。与标准线性形式相反，在二次和其他非线性形式",
"第一个受生物启发的神经网络之一是福岛的",
"Neocognitron[8]，它是CNN的前身，由LeCun等人",
"中，最佳刺激不能提供其属性的完整描述。结果表明",
"对于最佳刺激，没有不变性，而对于其他一般的次优刺",
"介绍。在6]中。卷积层是 CNN的核心构建块。CNN",
"激，可能存在许多不变性，这些不变性可能很大，但缺",
"的早期实现在其卷积层中使用了预定义的Gabor滤波",
"乏简单的解释。。虽然最优刺激与滤波器的不变性无关",
"器。这类过滤器可以非常准确地模拟在初级视觉皮层",
"但它的邻域是在更宽松的变换不变性下研究的。结果",
"（V1）[21]中发现的简单细胞的特性。这种类型的视觉",
"表明，适当的二次形式可以证明相移和方向变化的不",
"细胞具有响应曲线，其特征在于感受场内的空间求和。",
"变性。从前面的讨论中，我们得出的结论是，将非线性",
"寻找简单细胞的最佳空间刺激7是一个基于其兴奋区",
"形式用于卷积层可能是计算机视觉的合理未来方向",
"和抑制区[23]的空间排列的过程。然而，这并不适用",
"于复杂的视觉细胞。此外，我们无法通过找到它们的最",
"3. Proposed Method",
"佳刺激来准确描述它们的特性。到目前为止，大多数",
"CNN实现都忽略了这一事实，因为它们已经适应了线",
"如前所述，所提出的方法利用Volterra内核理论来",
"性类型的卷积滤波器，试图在其功能中应用定量而非",
"提供利用在感受场中发生的非线性运算的方法。到目",
"定性的变化。他等人。[10]提出了残差网络（ResNets),",
"前为止，据我们所知，非线性主要是通过激活函数和",
"该网络具有与其正常卷积层平行的快捷方式连接，作",
"CNN不同层之间的池化操作来利用的。然而，这种非",
"为增加模型参数时消失/爆炸梯度和硬优化问题的解决",
"线性可能是编码视觉系统内部过程的一种方法，而不",
"方案（i。e。添加更多图层）。Zagoruyko&Komodakis",
"是存在于感受野区域中的方法。",
"[34]表明，宽ResNets可以通过层数超过ResNets，从",
"我们的方法遵循CNN的典型工作流程，通过排列",
"而将兴趣转移到增加每层过滤器的数量上。另外，对于",
"不同目的的层（卷积，池化，激活函数，批处理归一化，",
"专注于创建具有更多卷积层或更多过滤器的网络的工",
"退出，完全连接等。），而非线性卷积层实际上可以在",
"作，我们评估使用非线性和线性项作为卷积内核的近",
"所有现有架构中插入。然而，由于涉及的可训练参数的",
"似值以提高CNN性能的影响除了ResNet之外，Im-",
"增加，应注意整个过程的复杂性。为此，还提供了4节",
"ageNetChallenge[27]还通过以新方式使用卷积层的方",
"中的CUDA实现",
"法实现了非常低的错误率，从而增强了它们的表示能",
"力。Lin等人。[20]提出了“网络中的网络（NIN）",
"3.1. Volterra-based convolution",
"",
"作为对典型过滤器存在的低抽象级别的补救。他们建",
"Volterra级数模型是连续函数的近似序列，用于表",
"立了具有更复杂结构的微神经网络，而不是传统的卷",
"示非线性动力学系统的输人输出关系，使用多项式函",
"积滤波器，后者是一种广义的线性模型，可以在感受",
"数展开。它们的方程式可以由无限阶项组成，但是基于",
"场内提取数据。为了将输人数据映射到输出，他们使",
"它们的实际实现使用截断版本，保留了高达阶的项。",
"用多层感知器作为非线性函数逼近器，他们将其称为",
"与线性卷积类似，基于Volterra 的卷积使用内核来过",
"“mlpconv”层。通过以与CNN类似的方式在输人上",
"滤输入数据。一阶Volterra内核包含滤波器线性部分",
"滑动微网络来获得输出特征图。Szegedy等人。30以",
"的系数。二阶核表示两个输人元素之间的二次相互作",
"Inception模块”的形式引入了新的组织级别，该模",
"用的系数。通常，r阶的内核表示r输人元素之间的非",
"块使用可变天小的过滤器来捕获不同天小的不同视觉",
"线性交互对响应的权重。在计算机视觉领域，Volterra",
"模式，并逼近最佳稀疏结构。谢等人。[33]提出了一种",
"内核以前已在[17]中用于人脸识别，有效地用作非线",
"利用“Inception”模型的拆分-转换-合并策略的方法，",
"性函数的近似值",
"该方法执行一组转换，每个转换都在低维嵌入上，其输",
"出通过求和进行聚合。[19]的作者基于大脑中大量的递",
"3.2. Forward pass",
"归突触，提出了使用递归神经网络进行图像分类的方",
"对于我们提出的卷积，我们采用了二阶Volterra级",
"法。他们证明，与前馈架构相比，在卷积层中插人循环",
"数。给定一个带有n元素（n=kh·kw）的输人补丁",
"连接可提高结果。他们的工作是将源自神经科学的机",
"IeRhaxh，并将其整形为向量aEIR\"：",
"制在生物学上合理地整合到CNN中在[28]中，提出",
"了一种玻尔兹曼学习算法，其中特征交互用于将隐藏",
"-1 2.. n",
"(1)",
"单元变成高阶特征检测器。在22中，提出了一种将",
"这种学习算法应用于高阶玻尔兹曼机的有效方法，使",
"它们在计算上易于处理实际问题。在[1]中，Bergstra",
"，线性滤波器的输入输出函数为：",
"等人。创建了一个神经激活模型，该模型通过合并二阶",
"相互作用并使用替代非线性作为激活函数，在数据集",
"n",
"g(c) =(wiai) +b",
"(2)",
"上显示出改进的泛化。在2中，尝试将二次形式分析",
"",
"和解释为接受场。在他们的研究中，发现二次形式可",
"用于建模非线性感受野，因为它们遵循初级视觉皮层",
"其中w是XMATHX的权重卷积。在我们的方法中，",
"中复杂细胞的某些特性。这些特性包括对边缘的响应，",
"此函数以以下二次形式扩展："
] |
{
"arxiv_id": "1708.07038",
"language": "zh",
"page_idx": 1
}
|
||
1708.07038-zh-2
|
1708.07038-zh-2.jpg
|
(3)
$$y(\boldsymbol{x})=\sum\limits_{i=1}^n\left(w_1^ix_i\right)+\sum\limits_{i=1}^n\sum\limits_{j=i}^n\left(w_2^{i,j}x_ix_j\right)+b$$
其中$w_{2}^{i,j}$是滤波器二阶项的权重。为了避免考虑每对输入元素$(x_i,x_j)$的交互项的两倍，我们对包含其权重的矩阵$w_2$采用上三角形式，因此二阶内核的可训练参数的数量为$n(n+1)/2$ 。计算基于 Volterra 的阶数为$r$的滤波器的参数总数$n_V$的泛型类型为：
$$n_V=\frac{(n+r)!}{n!r!}$$
(4)
以更紧凑的形式，(3)写为：
(5)
$$y(\boldsymbol{x})=\underbrace{\boldsymbol{x}^T\boldsymbol{w}_2\boldsymbol{x}}_{\text{quadratic term}}+\underbrace{\boldsymbol{w}_1^T\boldsymbol{x}}_{\text{linear term}}+b$$
<NEXT\_nLINE>
$$\boldsymbol{w}_2=\begin{bmatrix}w_2^{1,1}&w_2^{1,2}&\cdots&w_2^{1,n}\\0&w_2^{2,2}&\cdots&w_2^{2,n}\\\vdots&\vdots&\ddots&\vdots\\0&0&\cdots&w_2^{n,n}\end{bmatrix}$$
(6)
包含二次项的系数$w_{2}^{i,j}$ ,并且：
$${\boldsymbol w_{1}}^{T}=\begin{bmatrix}w_{1}^{1}&w_{1}^{2}&\cdots&w_{1}^{n}\end{bmatrix}$$
(7)
包含线性项的系数$w_1^i$ 。因此，建议的卷积输出可以
重写为：
$$y(\boldsymbol{x})=\begin{bmatrix}w_2^{1,1}\\w_2^{1,2}\\w_2^{1,3}\\\vdots\\w_2^{n,n}\end{bmatrix}^T\begin{bmatrix}x_1x_1\\x_1x_2\\x_1x_3\\\vdots\\x_nx_n\end{bmatrix}+\begin{bmatrix}w_1^1\\w_1^2\\w_1^3\\\vdots\\w_1^n\end{bmatrix}^T\begin{bmatrix}x_1\\x_2\\x_3\\\vdots\\x_n\end{bmatrix}+b$$
(8
请注意，权重$w_2^{i,j}$的上标$(i,j)$表示与交互的输入元
素$x_i$和$x_j$的空间位置的对应关系。
## 3.3. Backward pass
基于 Volterra 的卷积的后向传递方程的推导是通过使经典的反向传播方案适应上述(3)的输人输出函数来完成的。为了训练 Volterra 内核的权重，我们需要计算层输出$y(x)$相对于权重$w_1^i$和$w_2^{i,j}$的梯度。为了传播错误，我们必须计算层输出$\tilde{y(x)}$相对于输入$x_i$ 的梯度。因此，$\frac {\partial y}{\partial w_1^i}$ , $\frac {\partial y}{\partial w_2^{i, j}}$和$\frac{\partial y}{\partial x_i}$是将用于优化基于Volterra 的卷积层的权重参数并最小化网络损耗的术语。反向传播的数学方程式如下：
$$\dfrac{\partial y}{\partial w_1^i}=x_i\quad\dfrac{\partial y}{\partial w_2^{i,j}}=x_ix_j$$
(9)
$$\dfrac{\partial y}{\partial x_i}=w_1^i+\sum\limits_{k=1}^i\left(w_2^{k,i}x_k\right)+\sum\limits_{k=i}^n\left(w_2^{i,k}x_k\right)$$
(10)
$\begin{array}{ll}{4.}&{{\mathrm{Quadratic~convolution~filter~implementa}}}\\{{\mathrm{tion}}}\end{array}$
为了试验非线性卷积滤波器，我们使用了 Torch7 科学框架。基于 Volterra 的卷积被实现为与 CUDA 后端集成的模块，用于 Torch7 的神经网络(cunn)包。在 Torch7 中编写模块主要包括实现模块的前向传递(3)以及计算模块的梯度($\frac\partial E{\partial\boldsymbol{w}}$ 和$\frac\partial E{\partial\boldsymbol{x}}$),这些梯度用于反向传播。我们用$E$表示网络准则函数定义的误差，并将$\frac{\partial E}{\partial\boldsymbol{w}}$称为层的权重梯度，将$\frac{\partial E}{\partial\boldsymbol{x}}$称为层的输人梯度。为了在 CUDA 中实现正向传递，我们使用标准的 im2col [3]模式将数据补丁展开为列，然后使用基于 Volterra 的过滤器权重进行矩阵乘法。im2col 操作由 CUDA 内核并行进行，而对于矩阵乘法，我们使用了完善的 CUDABLAS 函数。随后，计算权重梯度在某种程度上类似于计算前向通过。再一次，im2col 操作作为 CUDA 内核在输入图像上执行，其输出矩阵与前一层的输入梯度相乘，得到$\frac{\partial E}{\partial w}$。在基于 Volterra 的卷积层中，最昂贵的操作是计算输入梯度。如前所述， 与线性卷积相反，在线性卷积中，输入梯度与提供的输人无关，我们层的输入梯度与输人相关。因此，为了计算输入梯度矩阵，首先我们计算一个包含输出相对于输入的梯度的展开矩阵。然后使用 CUDABLAS 函数将此矩阵与上一层的输入梯度相乘。最后，适当的逆col2imCUDA 内核聚合操作会导致基于 Volterra 的层的输入梯度$\frac{\partial E}{\partial x}$的最终矩阵
所提出的卷积方案与线性卷积之间的主要区别在于， 在我们的例子中，$\frac{\partial y}{\partial x_i}$是依赖于$x_i$的函数。这意味着， 与标准过滤器相比，该术语对于特征图的每个补丁都是不同的，导致额外的计算成本，当错误必须传播到网络中前面的可训练层时。此成本与$H_o\cdot W_o$成比例， 其中$H_o$和$W_o$分别是图层输出特征图的高度和宽度。我们层的代码可在 http://vcl.iti.gr/volterra 。
## 5. Experiments
我们在两个基准数据集上测量我们提出的基于Volterra 的卷积的性能：CIFAR-10 和 CIFAR-100[15] ,在配备 Inteli7-5820KCPU, 64GBRAM 和 NvidiaTi- $\operatorname{tanXGPU}$的 PC 上运行我们的实验。基于 Volterra 的卷积层在 Torch7 中实现。我们首先描述实验设置，然后我们展示了定量分析，在参数，分类误差和列车损失方面，对于所提出的方法
# 5.1. CNN architecture selection
如 4 节中所述，在 CNN 的多个层中使用建议的卷积，在反向传播期间引入了额外的计算开销。因此，我
|
[
[
0.58642578125,
0.1029052734375,
0.609375,
0.1029052734375,
0.609375,
0.117431640625,
0.58642578125,
0.117431640625
],
[
0.72607421875,
0.10101318359375,
0.75341796875,
0.10101318359375,
0.75341796875,
0.117431640625,
0.72607421875,
0.117431640625
],
[
0.7685546875,
0.111083984375,
0.814453125,
0.1142578125,
0.81396484375,
0.126953125,
0.76708984375,
0.12371826171875
],
[
0.86767578125,
0.111083984375,
0.890625,
0.111083984375,
0.890625,
0.1256103515625,
0.86767578125,
0.1256103515625
],
[
0.61767578125,
0.11297607421875,
0.64892578125,
0.11297607421875,
0.64892578125,
0.125,
0.61767578125,
0.125
],
[
0.58251953125,
0.11553955078125,
0.61669921875,
0.12060546875,
0.61279296875,
0.1363525390625,
0.57861328125,
0.1307373046875
],
[
0.1732177734375,
0.125,
0.2900390625,
0.12249755859375,
0.2900390625,
0.141357421875,
0.1732177734375,
0.143310546875
],
[
0.268798828125,
0.125,
0.4345703125,
0.121826171875,
0.4345703125,
0.1427001953125,
0.268798828125,
0.1458740234375
],
[
0.7197265625,
0.11932373046875,
0.7587890625,
0.11932373046875,
0.7587890625,
0.1363525390625,
0.7197265625,
0.1363525390625
],
[
0.125,
0.126953125,
0.181396484375,
0.126953125,
0.181396484375,
0.141357421875,
0.125,
0.141357421875
],
[
0.426513671875,
0.1307373046875,
0.44677734375,
0.1307373046875,
0.44677734375,
0.1383056640625,
0.426513671875,
0.1383056640625
],
[
0.44287109375,
0.1287841796875,
0.46728515625,
0.1287841796875,
0.46728515625,
0.14013671875,
0.44287109375,
0.14013671875
],
[
0.26953125,
0.1363525390625,
0.27685546875,
0.1363525390625,
0.27685546875,
0.1395263671875,
0.26953125,
0.1395263671875
],
[
0.17724609375,
0.1427001953125,
0.2017822265625,
0.1427001953125,
0.2017822265625,
0.1546630859375,
0.17724609375,
0.1546630859375
],
[
0.273681640625,
0.143310546875,
0.32177734375,
0.143310546875,
0.32177734375,
0.1546630859375,
0.273681640625,
0.1546630859375
],
[
0.5390625,
0.15087890625,
0.830078125,
0.1578369140625,
0.82861328125,
0.1875,
0.53759765625,
0.179931640625
],
[
0.101318359375,
0.169189453125,
0.468017578125,
0.17041015625,
0.468017578125,
0.184326171875,
0.101318359375,
0.18310546875
],
[
0.859375,
0.1654052734375,
0.890625,
0.1654052734375,
0.890625,
0.177978515625,
0.859375,
0.177978515625
],
[
0.54345703125,
0.17236328125,
0.57666015625,
0.17236328125,
0.57666015625,
0.1849365234375,
0.54345703125,
0.1849365234375
],
[
0.740234375,
0.17236328125,
0.75634765625,
0.17236328125,
0.75634765625,
0.179931640625,
0.740234375,
0.179931640625
],
[
0.62841796875,
0.179931640625,
0.65625,
0.179931640625,
0.65625,
0.1925048828125,
0.62841796875,
0.1925048828125
],
[
0.736328125,
0.179931640625,
0.76123046875,
0.179931640625,
0.76123046875,
0.1925048828125,
0.736328125,
0.1925048828125
],
[
0.08251953125,
0.186279296875,
0.46728515625,
0.186279296875,
0.46728515625,
0.199462890625,
0.08251953125,
0.199462890625
],
[
0.0849609375,
0.20263671875,
0.465576171875,
0.20263671875,
0.465576171875,
0.21337890625,
0.0849609375,
0.21337890625
],
[
0.5048828125,
0.20263671875,
0.53369140625,
0.20263671875,
0.53369140625,
0.2108154296875,
0.5048828125,
0.2108154296875
],
[
0.5400390625,
0.2008056640625,
0.88623046875,
0.2008056640625,
0.88623046875,
0.2139892578125,
0.5400390625,
0.2139892578125
],
[
0.08251953125,
0.216552734375,
0.468017578125,
0.216552734375,
0.468017578125,
0.2298583984375,
0.08251953125,
0.2298583984375
],
[
0.53369140625,
0.2183837890625,
0.56787109375,
0.2183837890625,
0.56787109375,
0.2310791015625,
0.53369140625,
0.2310791015625
],
[
0.08087158203125,
0.2310791015625,
0.398681640625,
0.2310791015625,
0.398681640625,
0.2442626953125,
0.08087158203125,
0.2442626953125
],
[
0.52392578125,
0.24365234375,
0.890625,
0.24365234375,
0.890625,
0.2568359375,
0.52392578125,
0.2568359375
],
[
0.2646484375,
0.2568359375,
0.324462890625,
0.2568359375,
0.324462890625,
0.271484375,
0.2646484375,
0.271484375
],
[
0.50341796875,
0.2568359375,
0.890625,
0.2568359375,
0.890625,
0.271484375,
0.50341796875,
0.271484375
],
[
0.223876953125,
0.267578125,
0.26318359375,
0.267578125,
0.26318359375,
0.279052734375,
0.223876953125,
0.279052734375
],
[
0.44677734375,
0.266357421875,
0.468017578125,
0.266357421875,
0.468017578125,
0.279052734375,
0.44677734375,
0.279052734375
],
[
0.281005859375,
0.2734375,
0.313720703125,
0.2734375,
0.313720703125,
0.285888671875,
0.281005859375,
0.285888671875
],
[
0.50634765625,
0.274658203125,
0.8896484375,
0.274658203125,
0.8896484375,
0.285400390625,
0.50634765625,
0.285400390625
],
[
0.5048828125,
0.28857421875,
0.890625,
0.28857421875,
0.890625,
0.3017578125,
0.5048828125,
0.3017578125
],
[
0.08087158203125,
0.29736328125,
0.28759765625,
0.29736328125,
0.28759765625,
0.311279296875,
0.08087158203125,
0.311279296875
],
[
0.50244140625,
0.302978515625,
0.890625,
0.302978515625,
0.890625,
0.31640625,
0.50244140625,
0.31640625
],
[
0.221435546875,
0.322509765625,
0.279296875,
0.324462890625,
0.27783203125,
0.34033203125,
0.2197265625,
0.33837890625
],
[
0.306396484375,
0.323974609375,
0.39794921875,
0.32080078125,
0.398681640625,
0.339111328125,
0.30810546875,
0.34228515625
],
[
0.5048828125,
0.318115234375,
0.890625,
0.318115234375,
0.890625,
0.33203125,
0.5048828125,
0.33203125
],
[
0.1519775390625,
0.32568359375,
0.2091064453125,
0.32568359375,
0.2091064453125,
0.34033203125,
0.1519775390625,
0.34033203125
],
[
0.28515625,
0.328857421875,
0.30810546875,
0.328857421875,
0.30810546875,
0.337158203125,
0.28515625,
0.337158203125
],
[
0.44677734375,
0.326904296875,
0.468017578125,
0.326904296875,
0.468017578125,
0.339111328125,
0.44677734375,
0.339111328125
],
[
0.5048828125,
0.333251953125,
0.890625,
0.333251953125,
0.890625,
0.3466796875,
0.5048828125,
0.3466796875
],
[
0.20751953125,
0.34521484375,
0.291748046875,
0.34521484375,
0.291748046875,
0.356689453125,
0.20751953125,
0.356689453125
],
[
0.306396484375,
0.34521484375,
0.371826171875,
0.34521484375,
0.371826171875,
0.356689453125,
0.306396484375,
0.356689453125
],
[
0.50341796875,
0.3466796875,
0.890625,
0.347900390625,
0.890625,
0.3623046875,
0.50341796875,
0.361083984375
],
[
0.5048828125,
0.363037109375,
0.890625,
0.363037109375,
0.890625,
0.376953125,
0.5048828125,
0.376953125
],
[
0.08416748046875,
0.371826171875,
0.2115478515625,
0.371826171875,
0.2115478515625,
0.3818359375,
0.08416748046875,
0.3818359375
],
[
0.50341796875,
0.376953125,
0.890625,
0.37744140625,
0.890625,
0.392578125,
0.50341796875,
0.391357421875
],
[
0.2059326171875,
0.392578125,
0.250732421875,
0.392578125,
0.250732421875,
0.417236328125,
0.2059326171875,
0.417236328125
],
[
0.2548828125,
0.395751953125,
0.293212890625,
0.39013671875,
0.298095703125,
0.412353515625,
0.2607421875,
0.417236328125
],
[
0.298095703125,
0.396484375,
0.3857421875,
0.39013671875,
0.38818359375,
0.415283203125,
0.301513671875,
0.421630859375
],
[
0.50341796875,
0.393310546875,
0.890625,
0.393310546875,
0.890625,
0.4072265625,
0.50341796875,
0.4072265625
],
[
0.30224609375,
0.41357421875,
0.37744140625,
0.41357421875,
0.37744140625,
0.43505859375,
0.30224609375,
0.43505859375
],
[
0.50732421875,
0.4091796875,
0.890625,
0.4091796875,
0.890625,
0.42236328125,
0.50732421875,
0.42236328125
],
[
0.2254638671875,
0.41845703125,
0.23779296875,
0.41845703125,
0.23779296875,
0.429931640625,
0.2254638671875,
0.429931640625
],
[
0.253173828125,
0.417236328125,
0.293212890625,
0.4091796875,
0.2998046875,
0.429931640625,
0.261474609375,
0.438232421875
],
[
0.44677734375,
0.429931640625,
0.468017578125,
0.429931640625,
0.468017578125,
0.442626953125,
0.44677734375,
0.442626953125
],
[
0.50634765625,
0.425537109375,
0.8896484375,
0.425537109375,
0.8896484375,
0.436279296875,
0.50634765625,
0.436279296875
],
[
0.303955078125,
0.438232421875,
0.36767578125,
0.438232421875,
0.36767578125,
0.453857421875,
0.303955078125,
0.453857421875
],
[
0.160888671875,
0.43115234375,
0.1968994140625,
0.43115234375,
0.1968994140625,
0.443115234375,
0.160888671875,
0.443115234375
],
[
0.268798828125,
0.439453125,
0.28515625,
0.439453125,
0.28515625,
0.453857421875,
0.268798828125,
0.453857421875
],
[
0.50341796875,
0.439453125,
0.890625,
0.439453125,
0.890625,
0.45263671875,
0.50341796875,
0.45263671875
],
[
0.291748046875,
0.4443359375,
0.298095703125,
0.4443359375,
0.298095703125,
0.449462890625,
0.291748046875,
0.449462890625
],
[
0.2254638671875,
0.43994140625,
0.23779296875,
0.43994140625,
0.23779296875,
0.451904296875,
0.2254638671875,
0.451904296875
],
[
0.301513671875,
0.455078125,
0.3857421875,
0.458251953125,
0.384033203125,
0.4765625,
0.2998046875,
0.473388671875
],
[
0.5185546875,
0.453857421875,
0.890625,
0.453857421875,
0.890625,
0.46728515625,
0.5185546875,
0.46728515625
],
[
0.26708984375,
0.458984375,
0.28515625,
0.458984375,
0.28515625,
0.472900390625,
0.26708984375,
0.472900390625
],
[
0.2059326171875,
0.458984375,
0.240234375,
0.455810546875,
0.24267578125,
0.4716796875,
0.20751953125,
0.474853515625
],
[
0.5048828125,
0.4697265625,
0.88818359375,
0.4697265625,
0.88818359375,
0.48291015625,
0.5048828125,
0.48291015625
],
[
0.10296630859375,
0.48486328125,
0.325927734375,
0.48486328125,
0.325927734375,
0.5,
0.10296630859375,
0.5
],
[
0.5048828125,
0.48291015625,
0.8896484375,
0.484130859375,
0.8896484375,
0.498779296875,
0.5048828125,
0.49755859375
],
[
0.50634765625,
0.5,
0.8896484375,
0.5,
0.8896484375,
0.51318359375,
0.50634765625,
0.51318359375
],
[
0.17236328125,
0.5146484375,
0.2254638671875,
0.50830078125,
0.2296142578125,
0.52783203125,
0.176513671875,
0.5341796875
],
[
0.27294921875,
0.517578125,
0.294189453125,
0.517578125,
0.294189453125,
0.52978515625,
0.27294921875,
0.52978515625
],
[
0.330810546875,
0.517578125,
0.369384765625,
0.517578125,
0.369384765625,
0.53271484375,
0.330810546875,
0.53271484375
],
[
0.44677734375,
0.517578125,
0.468017578125,
0.517578125,
0.468017578125,
0.52978515625,
0.44677734375,
0.52978515625
],
[
0.5048828125,
0.5146484375,
0.890625,
0.5146484375,
0.890625,
0.52783203125,
0.5048828125,
0.52783203125
],
[
0.30810546875,
0.51953125,
0.336669921875,
0.51953125,
0.336669921875,
0.52783203125,
0.30810546875,
0.52783203125
],
[
0.2254638671875,
0.51806640625,
0.25732421875,
0.51806640625,
0.25732421875,
0.52978515625,
0.2254638671875,
0.52978515625
],
[
0.5048828125,
0.52880859375,
0.89208984375,
0.52880859375,
0.89208984375,
0.54248046875,
0.5048828125,
0.54248046875
],
[
0.1004638671875,
0.54248046875,
0.468017578125,
0.54345703125,
0.468017578125,
0.55810546875,
0.1004638671875,
0.556640625
],
[
0.50341796875,
0.54443359375,
0.890625,
0.54443359375,
0.890625,
0.55810546875,
0.50341796875,
0.55810546875
],
[
0.08087158203125,
0.5595703125,
0.1396484375,
0.5595703125,
0.1396484375,
0.57373046875,
0.08087158203125,
0.57373046875
],
[
0.50244140625,
0.55810546875,
0.890625,
0.5595703125,
0.890625,
0.57373046875,
0.50244140625,
0.57275390625
],
[
0.50341796875,
0.57373046875,
0.701171875,
0.57373046875,
0.701171875,
0.58837890625,
0.50341796875,
0.58837890625
],
[
0.521484375,
0.5888671875,
0.890625,
0.59033203125,
0.890625,
0.60498046875,
0.521484375,
0.60400390625
],
[
0.5009765625,
0.60107421875,
0.65673828125,
0.60400390625,
0.65673828125,
0.6220703125,
0.5009765625,
0.61865234375
],
[
0.64892578125,
0.60400390625,
0.890625,
0.60498046875,
0.890625,
0.619140625,
0.64892578125,
0.61865234375
],
[
0.1805419921875,
0.60595703125,
0.20751953125,
0.61865234375,
0.1871337890625,
0.64453125,
0.16015625,
0.6318359375
],
[
0.1650390625,
0.60595703125,
0.19287109375,
0.60595703125,
0.19287109375,
0.6171875,
0.1650390625,
0.6171875
],
[
0.2344970703125,
0.60693359375,
0.271240234375,
0.60693359375,
0.271240234375,
0.6171875,
0.2344970703125,
0.6171875
],
[
0.318603515625,
0.609375,
0.330810546875,
0.609375,
0.330810546875,
0.61279296875,
0.318603515625,
0.61279296875
],
[
0.36279296875,
0.60595703125,
0.3896484375,
0.60595703125,
0.3896484375,
0.6171875,
0.36279296875,
0.6171875
],
[
0.314453125,
0.6220703125,
0.334228515625,
0.6220703125,
0.334228515625,
0.6337890625,
0.314453125,
0.6337890625
],
[
0.5048828125,
0.6220703125,
0.890625,
0.6220703125,
0.890625,
0.63525390625,
0.5048828125,
0.63525390625
],
[
0.370849609375,
0.625,
0.3857421875,
0.625,
0.3857421875,
0.63525390625,
0.370849609375,
0.63525390625
],
[
0.1805419921875,
0.62744140625,
0.2034912109375,
0.64013671875,
0.1846923828125,
0.6611328125,
0.160888671875,
0.6484375
],
[
0.23779296875,
0.62353515625,
0.27294921875,
0.62353515625,
0.27294921875,
0.63525390625,
0.23779296875,
0.63525390625
],
[
0.101318359375,
0.642578125,
0.1544189453125,
0.642578125,
0.1544189453125,
0.6572265625,
0.101318359375,
0.6572265625
],
[
0.24267578125,
0.64306640625,
0.26318359375,
0.64306640625,
0.26318359375,
0.6513671875,
0.24267578125,
0.6513671875
],
[
0.318603515625,
0.64306640625,
0.324462890625,
0.64306640625,
0.324462890625,
0.6494140625,
0.318603515625,
0.6494140625
],
[
0.400390625,
0.64306640625,
0.428955078125,
0.64306640625,
0.428955078125,
0.65478515625,
0.400390625,
0.65478515625
],
[
0.50634765625,
0.63818359375,
0.88818359375,
0.63818359375,
0.88818359375,
0.6484375,
0.50634765625,
0.6484375
],
[
0.373291015625,
0.64599609375,
0.38232421875,
0.64599609375,
0.38232421875,
0.6494140625,
0.373291015625,
0.6494140625
],
[
0.438720703125,
0.64794921875,
0.4453125,
0.64794921875,
0.4453125,
0.65283203125,
0.438720703125,
0.65283203125
],
[
0.44677734375,
0.64599609375,
0.46728515625,
0.64599609375,
0.46728515625,
0.65576171875,
0.44677734375,
0.65576171875
],
[
0.50341796875,
0.65087890625,
0.8896484375,
0.65087890625,
0.8896484375,
0.6640625,
0.50341796875,
0.6640625
],
[
0.178955078125,
0.66552734375,
0.1895751953125,
0.66552734375,
0.1895751953125,
0.67529296875,
0.178955078125,
0.67529296875
],
[
0.373291015625,
0.66748046875,
0.381591796875,
0.66748046875,
0.381591796875,
0.67041015625,
0.373291015625,
0.67041015625
],
[
0.5048828125,
0.66748046875,
0.890625,
0.66748046875,
0.890625,
0.68115234375,
0.5048828125,
0.68115234375
],
[
0.16259765625,
0.68310546875,
0.19775390625,
0.68310546875,
0.19775390625,
0.69580078125,
0.16259765625,
0.69580078125
],
[
0.2303466796875,
0.6845703125,
0.275390625,
0.6845703125,
0.275390625,
0.6943359375,
0.2303466796875,
0.6943359375
],
[
0.366943359375,
0.68505859375,
0.38818359375,
0.68505859375,
0.38818359375,
0.693359375,
0.366943359375,
0.693359375
],
[
0.5048828125,
0.68310546875,
0.82666015625,
0.68310546875,
0.82666015625,
0.6962890625,
0.5048828125,
0.6962890625
],
[
0.50244140625,
0.71044921875,
0.6396484375,
0.71044921875,
0.6396484375,
0.7236328125,
0.50244140625,
0.7236328125
],
[
0.318603515625,
0.68603515625,
0.325927734375,
0.68603515625,
0.325927734375,
0.68994140625,
0.318603515625,
0.68994140625
],
[
0.08251953125,
0.7119140625,
0.468017578125,
0.7119140625,
0.468017578125,
0.7265625,
0.08251953125,
0.7265625
],
[
0.08087158203125,
0.72900390625,
0.3349609375,
0.72900390625,
0.3349609375,
0.7421875,
0.08087158203125,
0.7421875
],
[
0.52392578125,
0.734375,
0.8896484375,
0.734375,
0.8896484375,
0.74755859375,
0.52392578125,
0.74755859375
],
[
0.07843017578125,
0.751953125,
0.2344970703125,
0.75244140625,
0.2344970703125,
0.767578125,
0.07843017578125,
0.7666015625
],
[
0.5048828125,
0.75,
0.88818359375,
0.75,
0.88818359375,
0.76318359375,
0.5048828125,
0.76318359375
],
[
0.51708984375,
0.7666015625,
0.88623046875,
0.7666015625,
0.88623046875,
0.7763671875,
0.51708984375,
0.7763671875
],
[
0.101318359375,
0.7763671875,
0.468017578125,
0.7763671875,
0.468017578125,
0.79052734375,
0.101318359375,
0.79052734375
],
[
0.50341796875,
0.77978515625,
0.890625,
0.77978515625,
0.890625,
0.79345703125,
0.50341796875,
0.79345703125
],
[
0.08087158203125,
0.79052734375,
0.468017578125,
0.79150390625,
0.468017578125,
0.8046875,
0.08087158203125,
0.8037109375
],
[
0.50244140625,
0.79345703125,
0.890625,
0.79443359375,
0.890625,
0.8095703125,
0.50244140625,
0.80810546875
],
[
0.08087158203125,
0.806640625,
0.468017578125,
0.806640625,
0.468017578125,
0.82080078125,
0.08087158203125,
0.82080078125
],
[
0.50341796875,
0.81005859375,
0.890625,
0.81005859375,
0.890625,
0.82373046875,
0.50341796875,
0.82373046875
],
[
0.08087158203125,
0.82275390625,
0.46728515625,
0.82275390625,
0.46728515625,
0.8359375,
0.08087158203125,
0.8359375
],
[
0.50341796875,
0.82568359375,
0.6845703125,
0.82568359375,
0.6845703125,
0.8388671875,
0.50341796875,
0.8388671875
],
[
0.08251953125,
0.8369140625,
0.468017578125,
0.8369140625,
0.468017578125,
0.85107421875,
0.08251953125,
0.85107421875
],
[
0.0767822265625,
0.8486328125,
0.471435546875,
0.8486328125,
0.471435546875,
0.87255859375,
0.0767822265625,
0.87255859375
],
[
0.5009765625,
0.8486328125,
0.74658203125,
0.8486328125,
0.74658203125,
0.8623046875,
0.5009765625,
0.8623046875
],
[
0.08087158203125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.52392578125,
0.8720703125,
0.8896484375,
0.8720703125,
0.8896484375,
0.88525390625,
0.52392578125,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.31787109375,
0.88720703125,
0.31787109375,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.5048828125,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5048828125,
0.90087890625
]
] |
[
"dy",
"ay",
"Ritj",
"(9)",
"=Ti",
"awi",
"Z(wiri)+)",
"ZE(w'aiai)+b",
"awij",
"y(α)=",
"",
"(3)",
"",
"-1",
"i=l j=i",
" =ui +Z(ua)+≥(uar)",
"其中ws，是滤波器二阶项的权重。为了避免考虑每",
"(10)",
"Ori",
"",
"k=1",
"k=i",
"对输人元素（αi，a）的交互项的两倍，我们对包含其权",
"重的矩阵W2采用上三角形式，因此二阶内核的可训练",
"4",
"Quadratic convolution filter implementa",
"参数的数量为n(n+1)/2。计算基于Volterra的阶数",
"tion",
"为r的滤波器的参数总数nv的泛型类型为：",
"为了试验非线性卷积滤波器，我们使用了Torch7科",
"(n+r)!",
"学框架。基于Volterra 的卷积被实现为与CUDA后",
"nv=",
"(4)",
"n!r!",
"端集成的模块，用于Torch7的神经网络（cunn）包。",
"在Torch7中编写模块主要包括实现模块的前向传递",
"以更紧凑的形式，（3）写为：",
"（3）以及计算模块的梯度（和E），这些梯度",
"aTw2",
"wiTa+b",
"用于反向传播。我们用E表示网络准则函数定义的误",
"y(a) =",
"",
"(5)",
"差，并将称为层的权重梯度，将称为层的输",
"quadratic term",
"linear term",
"人梯度。为了在CUDA中实现正向传递，我们使用标",
"准的im2col[3]模式将数据补丁展开为列，然后使用基",
"<NEXTnLINE>",
"于Volterra的过滤器权重进行矩阵乘法。im2col操作",
"[w",
"wf.2",
"... g.",
"由CUDA内核并行进行，而对于矩阵乘法，我们使用",
".w2.n",
"了完善的CUDABLAS函数。随后，计算权重梯度在",
"0",
"wg.2",
"(6)",
"某种程度上类似于计算前向通过。再一次，im2col操作",
".",
"W2:",
"1",
"作为CUDA内核在输入图像上执行，其输出矩阵与前",
"",
":",
". w]",
"层的输人梯度相乘，得到E。在基于Volterra的",
"0",
"L 0",
"卷积层中，最昂贵的操作是计算输入梯度。如前所述",
"包含二次项的系数w，并且：",
"与线性卷积相反，在线性卷积中，输人梯度与提供的输",
"人无关，我们层的输入梯度与输人相关。因此，为了计",
"wiT=",
"w?",
"wn",
"(7)",
"算输人梯度矩阵，首先我们计算一个包含输出相对于",
"..",
"wi",
"输入的梯度的展开矩阵。然后使用CUDABLAS函数",
"包含线性项的系数wi。因此，建议的卷积输出可以",
"将此矩阵与上一层的输人梯度相乘。最后，适当的逆",
"重写为：",
"col2imCUDA内核聚合操作会导致基于Volterra的层",
"的输入梯度的最终矩阵",
"所提出的卷积方案与线性卷积之间的主要区别在于，",
"在我们的例子中，",
"是依赖于的函数。这意味着，",
"",
"w2'",
"Ciai",
"3",
"a1",
"w",
"与标准过滤器相比，该术语对于特征图的每个补丁都",
"2",
"",
"122",
"y(α) =",
"1a",
"I",
"+b",
"是不同的，导致额外的计算成本，当错误必须传播到",
"+",
"",
"(8)",
"网络中前面的可训练层时。此成本与H。·W。成比例",
":",
"-",
"其中H。和W。分别是图层输出特征图的高度和宽度。",
"wn,r",
"anan",
"n",
"我们层的代码可在http://vcl.iti.gr/volterra。",
"5.Experiments",
"",
"请注意，权重wij的上标（i,）表示与交互的输人元",
"素和，的空间位置的对应关系。",
"我们在两个基准数据集上测量我们提出的基于",
"3.3.Backward pass",
"Volterra的卷积的性能：CIFAR-10和CIFAR-100[15",
"在配备Inteli7-5820KCPU，64GBRAM和NvidiaTi",
"基于Volterra的卷积的后向传递方程的推导是通过",
"tanXGPU的PC上运行我们的实验。基于Volterra的",
"使经典的反向传播方案适应上述（3）的输人输出函数",
"卷积层在Torch7中实现。我们首先描述实验设置，然",
"来完成的。为了训练Volterra内核的权重，我们需要",
"后我们展示了定量分析，在参数，分类误差和列车损失",
"计算层输出y(α)相对于权重wi和wz的梯度。为了",
"方面，对于所提出的方法",
"传播错误，我们必须计算层输出y(a）相对于输人&i",
"的梯度。因此，，和是将用于优化基于",
"5.1. CNN architecture selection",
"Volterra的卷积层的权重参数并最小化网络损耗的术",
"如4节中所述，在CNN的多个层中使用建议的卷",
"语。反向传播的数学方程式如下：",
"积，在反向传播期间引人了额外的计算开销。因此，我"
] |
{
"arxiv_id": "1708.07038",
"language": "zh",
"page_idx": 2
}
|
||
1708.07038-zh-3
|
1708.07038-zh-3.jpg
|
<table>
<tbody>
<tr>
<th>Network stage</th>
<th>Output size</th>
<th>28, N= 4, k= 10)</th>
</tr>
<tr>
<td>Initial convolution</td>
<td>$32\times32$</td>
<td>Vormalization $3\times 3$ , 16 $\cdot k$</td>
</tr>
<tr>
<td>Group 1</td>
<td>$32\times32$</td>
<td>$16\cdot k$ $\times N$blocks $16\cdot k$</td>
</tr>
<tr>
<td>Group 2</td>
<td>$16\times16$</td>
<td>$32\cdot k$ $\times N$blocks $32\cdot k$</td>
</tr>
<tr>
<td>$\operatorname{Group}3$</td>
<td>$8\times8$</td>
<td>$64\cdot k$ $\times N$blocks $64\cdot k$</td>
</tr>
<tr>
<td>Pooling</td>
<td>$8\times8$ $8\times8$ $1\times1$</td>
<td>Vormalization ReLU Pooling, 8$\times$8</td>
</tr>
<tr>
<td>Classifier</td>
<td> </td>
<td>${\mathrm{nnected~10(100)}}$ $\sqrt{\mathrm{Iax}}$ :10(100) </td>
</tr>
</tbody>
</table>
们限制自己仅在 CNN 模型的第一个卷积层中测试此类过滤器。我们选择 WideResNet [34] 的现代架构，它主要由一个卷积层组成，然后是 3 个卷积组和一个分类器。如果$d$是这样一个网络的深度，那么每个卷积组包含
$N=(d-4)/6$卷积块。在一个组中，每个卷积层的滤波器的数量由扩展因子$k$控制。在我们的架构中，我们遵循上述规则，进行了三个更改：
[label=)]we insert a Batch Normalization layer in the start of the network we change the number of the first convolutional layer's output channels, from 16 to 16$\cdot k$ (i.e., equal to the number of the first group's output channels) and we change the shortcut of the first block in the frst group, into an identity mapping, as a consequence of our second change.
由于乘法交互项$x_ix_j$ ,第一个更改对于防止基于Volterra 的卷积的输出爆炸至关重要。在我们的实验中，在该层中应用的仿射变换$y=\gamma\hat{x}+\beta$的参数$\gamma$稳定为值$0<\gamma<1$ 。选择第二个变化，以便当在第一卷积层中应用基于 Volterra 的卷积时，有足够的非线性滤波器要学习，从而产生特征丰富的信号。进行第三个更改是因为当块的输入和输出通道相等时，则其快捷方式是标识映射，以便将其输入添加到其输出中，而无需通过使用卷积层来调整快捷方式中的特征通道。在这种情况下，第一卷积层的信号原封不动地流过第一组块的快捷方式
我们实验中使用的模型如表 1 所示。为了评估在每个数据集上应用基于 Volterra 的卷积的影响，我们测试了两个版本的通用 CNN 模型。作为基线模型的第一个版本不使用任何非线性卷积滤波器。另一个版本在
第一卷积层中包含非线性滤波器，在网络的所有卷积
组中包含线性滤波器
# 5.2. Experimental setup
在我们所有的实验中，我们使用随机梯度下降(SGD),动量设置为 0.9,交叉熵损失，批量大小为 128 , 训练我们的网络 220 个时期。将 Dropout 设置为 0.3 , 并按照[10]中的方式进行权重初始化。实验中使用的学习率和权重衰减策略如表 2 所示。对于CIFAR-10 和 CIFAR-100, 应用于火车和测试集数据的数据预处理操作是减去通道平均值，然后除以在火车集上计算的通道标准偏差。我们使用适度的数据增强， 使用 50% 概率的水平翻转和每个图像侧 4 个像素的反射填充，随机裁剪$32\times32$大小
# $5. 3. \textbf{CIFAR- 10 and CIFAR- 100}$
CIFAR-10 和 CIFAR-100 数据集包含 60.000 32×32常见对象类别的 RGB 图像(e。g。,动物、车辆等。),其中火车集有 50.000 , 测试集有 10.000 图像。CIFAR-10有 10 个类，CIFAR-100 有 100 个类。所有班级都有相等数量的火车和测试样本。在 CIFAR-10 中，我们基
<table>
<tbody>
<tr>
<th>$\operatorname{Epoch}$</th>
<th>Learning rate</th>
<th>Weight $\operatorname{decay}$</th>
</tr>
<tr>
<td>1-59</td>
<td>0.1</td>
<td>0.0005</td>
</tr>
<tr>
<td>60-119</td>
<td>0.02</td>
<td>0.0005</td>
</tr>
<tr>
<td>120-159 </td>
<td>0.004</td>
<td>0.0005</td>
</tr>
<tr>
<td>160-199 </td>
<td>0.0008</td>
<td>0.0005</td>
</tr>
<tr>
<td>200- 220 </td>
<td>0.0008</td>
<td>0</td>
</tr>
</tbody>
</table>
Table $2^{\lfloor\frac{1}{\text{Learning rate and weight decay strategy used in}}}$
our experiments.
|
[
[
0.26708984375,
0.0921630859375,
0.373291015625,
0.09344482421875,
0.373291015625,
0.10797119140625,
0.26708984375,
0.106689453125
],
[
0.39453125,
0.0921630859375,
0.48193359375,
0.0921630859375,
0.48193359375,
0.106689453125,
0.39453125,
0.106689453125
],
[
0.5048828125,
0.0921630859375,
0.69921875,
0.0921630859375,
0.69921875,
0.106689453125,
0.5048828125,
0.106689453125
],
[
0.52783203125,
0.1142578125,
0.6787109375,
0.1142578125,
0.6787109375,
0.1275634765625,
0.52783203125,
0.1275634765625
],
[
0.2548828125,
0.12249755859375,
0.3896484375,
0.12249755859375,
0.3896484375,
0.1363525390625,
0.2548828125,
0.1363525390625
],
[
0.41015625,
0.12249755859375,
0.468017578125,
0.12249755859375,
0.468017578125,
0.136962890625,
0.41015625,
0.136962890625
],
[
0.53759765625,
0.1337890625,
0.6708984375,
0.1337890625,
0.6708984375,
0.147705078125,
0.53759765625,
0.147705078125
],
[
0.289306640625,
0.1590576171875,
0.3544921875,
0.1610107421875,
0.35302734375,
0.179931640625,
0.28759765625,
0.177978515625
],
[
0.5390625,
0.154052734375,
0.6259765625,
0.154052734375,
0.6259765625,
0.1685791015625,
0.5390625,
0.1685791015625
],
[
0.41015625,
0.1610107421875,
0.468017578125,
0.1610107421875,
0.468017578125,
0.1767578125,
0.41015625,
0.1767578125
],
[
0.488525390625,
0.1622314453125,
0.529296875,
0.1641845703125,
0.52783203125,
0.1767578125,
0.48779296875,
0.1749267578125
],
[
0.6396484375,
0.1622314453125,
0.71826171875,
0.1622314453125,
0.71826171875,
0.1767578125,
0.6396484375,
0.1767578125
],
[
0.5400390625,
0.173583984375,
0.6259765625,
0.173583984375,
0.6259765625,
0.1881103515625,
0.5400390625,
0.1881103515625
],
[
0.5400390625,
0.1932373046875,
0.62744140625,
0.1932373046875,
0.62744140625,
0.2083740234375,
0.5400390625,
0.2083740234375
],
[
0.2900390625,
0.2020263671875,
0.3544921875,
0.2020263671875,
0.3544921875,
0.2171630859375,
0.2900390625,
0.2171630859375
],
[
0.41015625,
0.2020263671875,
0.468017578125,
0.2020263671875,
0.468017578125,
0.216552734375,
0.41015625,
0.216552734375
],
[
0.48779296875,
0.2008056640625,
0.53125,
0.20263671875,
0.529296875,
0.2171630859375,
0.486083984375,
0.21533203125
],
[
0.6396484375,
0.2020263671875,
0.71826171875,
0.2020263671875,
0.71826171875,
0.216552734375,
0.6396484375,
0.216552734375
],
[
0.5400390625,
0.21337890625,
0.6259765625,
0.21337890625,
0.6259765625,
0.2279052734375,
0.5400390625,
0.2279052734375
],
[
0.5400390625,
0.23291015625,
0.62744140625,
0.23291015625,
0.62744140625,
0.2474365234375,
0.5400390625,
0.2474365234375
],
[
0.291748046875,
0.240478515625,
0.3544921875,
0.242431640625,
0.35302734375,
0.25830078125,
0.2900390625,
0.255615234375
],
[
0.416748046875,
0.2412109375,
0.461669921875,
0.2412109375,
0.461669921875,
0.2568359375,
0.416748046875,
0.2568359375
],
[
0.490234375,
0.2412109375,
0.529296875,
0.24365234375,
0.52783203125,
0.255615234375,
0.488525390625,
0.25390625
],
[
0.6396484375,
0.2412109375,
0.71826171875,
0.2412109375,
0.71826171875,
0.255615234375,
0.6396484375,
0.255615234375
],
[
0.53759765625,
0.25244140625,
0.62744140625,
0.25244140625,
0.62744140625,
0.267578125,
0.53759765625,
0.267578125
],
[
0.416748046875,
0.272705078125,
0.4599609375,
0.272705078125,
0.4599609375,
0.28857421875,
0.416748046875,
0.28857421875
],
[
0.52783203125,
0.272705078125,
0.6787109375,
0.272705078125,
0.6787109375,
0.287353515625,
0.52783203125,
0.287353515625
],
[
0.293212890625,
0.290283203125,
0.351318359375,
0.292236328125,
0.3505859375,
0.306884765625,
0.291748046875,
0.304931640625
],
[
0.416748046875,
0.292236328125,
0.4599609375,
0.292236328125,
0.4599609375,
0.306884765625,
0.416748046875,
0.306884765625
],
[
0.580078125,
0.292236328125,
0.6259765625,
0.292236328125,
0.6259765625,
0.306884765625,
0.580078125,
0.306884765625
],
[
0.416748046875,
0.3125,
0.4599609375,
0.3125,
0.4599609375,
0.326904296875,
0.416748046875,
0.326904296875
],
[
0.52294921875,
0.311279296875,
0.685546875,
0.311279296875,
0.685546875,
0.32568359375,
0.52294921875,
0.32568359375
],
[
0.5146484375,
0.333251953125,
0.69140625,
0.333251953125,
0.69140625,
0.3466796875,
0.5146484375,
0.3466796875
],
[
0.28759765625,
0.341552734375,
0.35546875,
0.341552734375,
0.35546875,
0.35595703125,
0.28759765625,
0.35595703125
],
[
0.54150390625,
0.35107421875,
0.66259765625,
0.3515625,
0.66259765625,
0.3662109375,
0.54150390625,
0.365478515625
],
[
0.295654296875,
0.36865234375,
0.6689453125,
0.36865234375,
0.6689453125,
0.3818359375,
0.295654296875,
0.3818359375
],
[
0.08251953125,
0.40771484375,
0.468017578125,
0.40771484375,
0.468017578125,
0.421630859375,
0.08251953125,
0.421630859375
],
[
0.50341796875,
0.40771484375,
0.89208984375,
0.40771484375,
0.89208984375,
0.421630859375,
0.50341796875,
0.421630859375
],
[
0.08087158203125,
0.42236328125,
0.46728515625,
0.42236328125,
0.46728515625,
0.4375,
0.08087158203125,
0.4375
],
[
0.50244140625,
0.421630859375,
0.65283203125,
0.42236328125,
0.65283203125,
0.4375,
0.50244140625,
0.436279296875
],
[
0.08416748046875,
0.438232421875,
0.46728515625,
0.438232421875,
0.46728515625,
0.451904296875,
0.08416748046875,
0.451904296875
],
[
0.080078125,
0.451904296875,
0.4697265625,
0.45263671875,
0.4697265625,
0.46728515625,
0.080078125,
0.466552734375
],
[
0.50341796875,
0.449462890625,
0.69287109375,
0.449462890625,
0.69287109375,
0.464111328125,
0.50341796875,
0.464111328125
],
[
0.080078125,
0.466552734375,
0.131591796875,
0.468505859375,
0.1298828125,
0.484130859375,
0.07843017578125,
0.481689453125
],
[
0.52294921875,
0.473388671875,
0.890625,
0.473388671875,
0.890625,
0.4873046875,
0.52294921875,
0.4873046875
],
[
0.1004638671875,
0.484130859375,
0.468017578125,
0.484130859375,
0.468017578125,
0.498779296875,
0.1004638671875,
0.498779296875
],
[
0.5009765625,
0.4873046875,
0.89208984375,
0.488037109375,
0.89208984375,
0.50244140625,
0.5009765625,
0.501953125
],
[
0.08087158203125,
0.5,
0.468017578125,
0.5,
0.468017578125,
0.51318359375,
0.08087158203125,
0.51318359375
],
[
0.5048828125,
0.50390625,
0.89208984375,
0.50390625,
0.89208984375,
0.517578125,
0.5048828125,
0.517578125
],
[
0.08087158203125,
0.51513671875,
0.330810546875,
0.51513671875,
0.330810546875,
0.52880859375,
0.08087158203125,
0.52880859375
],
[
0.5048828125,
0.51953125,
0.89208984375,
0.51953125,
0.89208984375,
0.53271484375,
0.5048828125,
0.53271484375
],
[
0.11358642578125,
0.5361328125,
0.468017578125,
0.5361328125,
0.468017578125,
0.5498046875,
0.11358642578125,
0.5498046875
],
[
0.50341796875,
0.53271484375,
0.8896484375,
0.5341796875,
0.8896484375,
0.548828125,
0.50341796875,
0.54736328125
],
[
0.11358642578125,
0.55029296875,
0.468017578125,
0.55029296875,
0.468017578125,
0.564453125,
0.11358642578125,
0.564453125
],
[
0.50341796875,
0.5498046875,
0.890625,
0.5498046875,
0.890625,
0.56298828125,
0.50341796875,
0.56298828125
],
[
0.11358642578125,
0.56640625,
0.46728515625,
0.56640625,
0.46728515625,
0.580078125,
0.11358642578125,
0.580078125
],
[
0.50341796875,
0.564453125,
0.890625,
0.564453125,
0.890625,
0.57763671875,
0.50341796875,
0.57763671875
],
[
0.111083984375,
0.580078125,
0.4697265625,
0.58056640625,
0.4697265625,
0.59521484375,
0.111083984375,
0.5947265625
],
[
0.5048828125,
0.5791015625,
0.88818359375,
0.5791015625,
0.88818359375,
0.59228515625,
0.5048828125,
0.59228515625
],
[
0.11273193359375,
0.5966796875,
0.4697265625,
0.5966796875,
0.4697265625,
0.611328125,
0.11273193359375,
0.611328125
],
[
0.5048828125,
0.5947265625,
0.890625,
0.5947265625,
0.890625,
0.60791015625,
0.5048828125,
0.60791015625
],
[
0.11273193359375,
0.60986328125,
0.468017578125,
0.6123046875,
0.468017578125,
0.62548828125,
0.11273193359375,
0.62353515625
],
[
0.50244140625,
0.609375,
0.73193359375,
0.60791015625,
0.73193359375,
0.62255859375,
0.50244140625,
0.62353515625
],
[
0.11273193359375,
0.626953125,
0.4697265625,
0.626953125,
0.4697265625,
0.6416015625,
0.11273193359375,
0.6416015625
],
[
0.11273193359375,
0.64013671875,
0.1690673828125,
0.64306640625,
0.1683349609375,
0.6572265625,
0.111083984375,
0.65380859375
],
[
0.50244140625,
0.63623046875,
0.74853515625,
0.63623046875,
0.74853515625,
0.6494140625,
0.50244140625,
0.6494140625
],
[
0.521484375,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.521484375,
0.673828125
],
[
0.08905029296875,
0.6748046875,
0.46728515625,
0.6748046875,
0.46728515625,
0.689453125,
0.08905029296875,
0.689453125
],
[
0.50341796875,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.689453125,
0.50341796875,
0.689453125
],
[
0.080078125,
0.68798828125,
0.468017578125,
0.689453125,
0.468017578125,
0.7041015625,
0.080078125,
0.70263671875
],
[
0.50341796875,
0.68994140625,
0.89208984375,
0.68994140625,
0.89208984375,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.08251953125,
0.705078125,
0.468017578125,
0.705078125,
0.468017578125,
0.7197265625,
0.08251953125,
0.7197265625
],
[
0.5048828125,
0.705078125,
0.890625,
0.705078125,
0.890625,
0.7197265625,
0.5048828125,
0.7197265625
],
[
0.080078125,
0.7197265625,
0.468017578125,
0.71826171875,
0.468017578125,
0.73291015625,
0.080078125,
0.734375
],
[
0.50341796875,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.734375,
0.50341796875,
0.734375
],
[
0.08087158203125,
0.73486328125,
0.468017578125,
0.73486328125,
0.468017578125,
0.74853515625,
0.08087158203125,
0.74853515625
],
[
0.08087158203125,
0.75048828125,
0.46728515625,
0.75048828125,
0.46728515625,
0.7646484375,
0.08087158203125,
0.7646484375
],
[
0.564453125,
0.76220703125,
0.61767578125,
0.76220703125,
0.61767578125,
0.7763671875,
0.564453125,
0.7763671875
],
[
0.63134765625,
0.76220703125,
0.8408203125,
0.76220703125,
0.8408203125,
0.7763671875,
0.63134765625,
0.7763671875
],
[
0.564453125,
0.77978515625,
0.61669921875,
0.77978515625,
0.61669921875,
0.79541015625,
0.564453125,
0.79541015625
],
[
0.08251953125,
0.76513671875,
0.46728515625,
0.76513671875,
0.46728515625,
0.77880859375,
0.08251953125,
0.77880859375
],
[
0.08251953125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.08251953125,
0.79443359375
],
[
0.6689453125,
0.77978515625,
0.69775390625,
0.77978515625,
0.69775390625,
0.796875,
0.6689453125,
0.796875
],
[
0.7646484375,
0.78076171875,
0.81640625,
0.78076171875,
0.81640625,
0.79541015625,
0.7646484375,
0.79541015625
],
[
0.08087158203125,
0.79541015625,
0.468017578125,
0.79541015625,
0.468017578125,
0.8095703125,
0.08087158203125,
0.8095703125
],
[
0.55615234375,
0.798828125,
0.62353515625,
0.798828125,
0.62353515625,
0.81298828125,
0.55615234375,
0.81298828125
],
[
0.66650390625,
0.7998046875,
0.701171875,
0.7998046875,
0.701171875,
0.8125,
0.66650390625,
0.8125
],
[
0.7646484375,
0.798828125,
0.81640625,
0.798828125,
0.81640625,
0.81298828125,
0.7646484375,
0.81298828125
],
[
0.08087158203125,
0.81005859375,
0.465576171875,
0.81005859375,
0.465576171875,
0.82373046875,
0.08087158203125,
0.82373046875
],
[
0.5537109375,
0.8173828125,
0.62744140625,
0.8173828125,
0.62744140625,
0.83203125,
0.5537109375,
0.83203125
],
[
0.66259765625,
0.8173828125,
0.705078125,
0.8173828125,
0.705078125,
0.83349609375,
0.66259765625,
0.83349609375
],
[
0.7646484375,
0.8173828125,
0.814453125,
0.8173828125,
0.814453125,
0.83349609375,
0.7646484375,
0.83349609375
],
[
0.080078125,
0.82568359375,
0.1968994140625,
0.82568359375,
0.1968994140625,
0.8388671875,
0.080078125,
0.8388671875
],
[
0.5537109375,
0.8359375,
0.62744140625,
0.8359375,
0.62744140625,
0.85107421875,
0.5537109375,
0.85107421875
],
[
0.65869140625,
0.8359375,
0.708984375,
0.8359375,
0.708984375,
0.85107421875,
0.65869140625,
0.85107421875
],
[
0.7646484375,
0.8359375,
0.814453125,
0.8359375,
0.814453125,
0.85107421875,
0.7646484375,
0.85107421875
],
[
0.08087158203125,
0.85498046875,
0.46728515625,
0.85595703125,
0.46728515625,
0.87060546875,
0.08087158203125,
0.869140625
],
[
0.1004638671875,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.1004638671875,
0.85498046875
],
[
0.5537109375,
0.85498046875,
0.62841796875,
0.85498046875,
0.62841796875,
0.869140625,
0.5537109375,
0.869140625
],
[
0.65869140625,
0.85498046875,
0.708984375,
0.85498046875,
0.708984375,
0.869140625,
0.65869140625,
0.869140625
],
[
0.77783203125,
0.85791015625,
0.7939453125,
0.8515625,
0.80126953125,
0.86279296875,
0.7861328125,
0.869140625
],
[
0.08251953125,
0.8720703125,
0.45751953125,
0.8720703125,
0.45751953125,
0.88525390625,
0.08251953125,
0.88525390625
],
[
0.50244140625,
0.86865234375,
0.89208984375,
0.869140625,
0.89208984375,
0.88720703125,
0.50244140625,
0.88623046875
],
[
0.455078125,
0.87548828125,
0.464111328125,
0.87548828125,
0.464111328125,
0.8798828125,
0.455078125,
0.8798828125
],
[
0.080078125,
0.88525390625,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.080078125,
0.8994140625
],
[
0.50341796875,
0.88720703125,
0.61376953125,
0.88720703125,
0.61376953125,
0.89794921875,
0.50341796875,
0.89794921875
]
] |
[
"Network stage",
"Output size",
"Model (d=28, N=4, k=10)",
"Batch Normalization",
"Initial convolution",
"32 ×32",
"Conv 3 x 3, 16 -k",
"Group 1",
"3×3,16.k",
"32×32",
"Conv",
"×N blocks",
"3 × 3,16.k",
"3 ×3,32 k",
"Group 2",
"16 × 16",
"Conv",
"xN blocks",
"3 × 3,32 k",
"3 × 3,64 k",
"Group 3",
"8×8",
"Conv",
"×N blocks",
"3×3,64k",
"8×8",
"Batch Normalization",
"Pooling",
"8×8",
"ReLU",
"1×1",
"Average Pooling, 8 × 8",
"Fully-Connected 10(100)",
"Classifier",
"SoftMax 10(100)",
"Table 1.Network architecture used in our experiments.",
"们限制自已仅在CNN模型的第一个卷积层中测试此",
"第一卷积层中包含非线性滤波器，在网络的所有卷积",
"类过滤器。我们选择WideResNet[34]的现代架构，它",
"组中包含线性滤波器",
"主要由一个卷积层组成，然后是3个卷积组和一个分",
"类器。如果d是这样一个网络的深度，那么每个卷积",
"5.2. Experimental setup",
"组包含",
"在我们所有的实验中，我们使用随机梯度下降",
"N=（d一4)/6卷积块。在一个组中，每个卷积层的",
"（SGD），动量设置为0.9，交叉摘损失，批量大小",
"滤波器的数量由扩展因子k控制。在我们的架构中，我",
"为128，训练我们的网络220个时期。将Dropout设",
"们遵循上述规则，进行了三个更改：",
"置为0.3，并按照[10]中的方式进行权重初始化。实",
"[label-)]we insert a Batch Normalization layer in",
"验中使用的学习率和权重衰减策略如表2所示。对于",
"the start of the network we change the number",
"CIFAR-10和CIFAR-100，应用于火车和测试集数据的",
"of the first convolutional layer's output channels,",
"数据预处理操作是减去通道平均值，然后除以在火车",
"from 16 to 16 .k (i.e., equal to the number of the",
"集上计算的通道标准偏差。我们使用适度的数据增强",
"first group's output channels) and we change the",
"使用50%概率的水平翻转和每个图像侧4个像素的反",
"shortcut ofthefirstblock in thefirstgroup,intoan",
"射填充，随机裁剪32×32大小",
"identity mapping, as a consequence of our second",
"change.",
"5.3. CIFAR-10 and CIFAR-100",
"CIFAR-10和CIFAR-100数据集包含60.00032×32",
"由于乘法交互项i让，第一个更改对于防止基于",
"常见对象类别的RGB图像（e。g。、动物、车辆等。），其",
"Volterra的卷积的输出爆炸至关重要。在我们的实验",
"中火车集有50.000，测试集有10.000图像。CIFAR-10",
"中，在该层中应用的仿射变换y=+β的参数稳",
"有10个类，CIFAR-100有100个类。所有班级都有相",
"定为值0<<1。选择第二个变化，以便当在第一卷",
"等数量的火车和测试样本。在CIFAR-10中，我们基",
"积层中应用基于Volterra的卷积时，有足够的非线性",
"滤波器要学习，从而产生特征丰富的信号。进行第三个",
"Epoch",
"Learning rate Weight decay",
"159",
"更改是因为当块的输人和输出通道相等时，则其快捷",
"方式是标识映射，以便将其输人添加到其输出中，而无",
"0.1",
"0.0005",
"需通过使用卷积层来调整快捷方式中的特征通道。在",
"60 -119",
"0.02",
"0.0005",
"这种情况下，第一卷积层的信号原封不动地流过第一",
"120159",
"0.004",
"0.0005",
"组块的快捷方式",
"160-199",
"0.0008",
"0.0005",
"个数据集上应用基于Volterra的卷积的影响，我们测",
"我们实验中使用的模型如表1所示。为了评估在每",
"200-220",
"0.0008",
"0",
"试了两个版本的通用CNN模型。作为基线模型的第",
"Table 2.Learning rate and weight decay strategy used in",
"",
"个版本不使用任何非线性卷积滤波器。另一个版本在",
"our experiments."
] |
{
"arxiv_id": "1708.07038",
"language": "zh",
"page_idx": 3
}
|
||
1708.07038-zh-4
|
1708.07038-zh-4.jpg
|
<table>
<tbody>
<tr>
<th>Network</th>
<th>Depth</th>
<th>$\#$ Parameters</th>
<th>CIFAR-10</th>
<th>CIFAR-100</th>
</tr>
<tr>
<td>NIN [20]</td>
<td> </td>
<td> </td>
<td>8.81</td>
<td> </td>
</tr>
<tr>
<td>DSN [18]</td>
<td>3</td>
<td> </td>
<td>7.97</td>
<td>34.57</td>
</tr>
<tr>
<td>All- CNN [ 29] </td>
<td>9</td>
<td>1.3M</td>
<td>7.25</td>
<td> </td>
</tr>
<tr>
<td rowspan="2">ResNet with Stochastic Depth [12]</td>
<td>110</td>
<td>1.7M</td>
<td>5.23</td>
<td>24.58</td>
</tr>
<tr>
<td>1202</td>
<td>10.2M</td>
<td>4.91</td>
<td> </td>
</tr>
<tr>
<td>pre-act Resnet [11]</td>
<td>1001</td>
<td>10.2M</td>
<td>4.62</td>
<td>22.71</td>
</tr>
<tr>
<td>Wide ResNet [34]</td>
<td>40</td>
<td>55.8M</td>
<td>3.80</td>
<td>18.30</td>
</tr>
<tr>
<td>PyramidNet[9]</td>
<td>110</td>
<td>28.3M</td>
<td>3.77</td>
<td>18.29</td>
</tr>
<tr>
<td>Wide-DelugeNet[16]</td>
<td>146</td>
<td>20.2M</td>
<td>3.76</td>
<td>19.02</td>
</tr>
<tr>
<td>OrthoReg on Wide ResNet [26]</td>
<td>28</td>
<td> </td>
<td>3.69</td>
<td>18.56</td>
</tr>
<tr>
<td>Steerable CNNs [4]</td>
<td>14</td>
<td>9.1M</td>
<td>3.65</td>
<td>18.82</td>
</tr>
<tr>
<td>ResNeXt [ 33] </td>
<td>29</td>
<td>68.1M</td>
<td>3.58</td>
<td>17.31</td>
</tr>
<tr>
<td>Wide ResNet with Singular Value Bounding [14]</td>
<td>28</td>
<td>36.5M</td>
<td>3.52</td>
<td>18.32</td>
</tr>
<tr>
<td>Oriented Response Net [35]</td>
<td>28</td>
<td>18.4M</td>
<td>3.52</td>
<td>19.22</td>
</tr>
<tr>
<td>Baseline Wide ResNet</td>
<td>28</td>
<td>36.6M</td>
<td>3.62</td>
<td>18.29</td>
</tr>
<tr>
<td>ResNet</td>
<td>28</td>
<td>36.7M</td>
<td>3.51</td>
<td>18.24</td>
</tr>
</tbody>
</table>
(horizonta)
Table 3.
${\text{flipping, padding and 32}\times32}$cropping).
于 Volterra 的 WideResNet 产生了 3.51% 的测试错误， 这显示了我们使用基线模型获得的 3.62% 错误的改进， 从而在此数据集上设置了最新技术。在 CIFAR-100 中， 我们基于 Volterra 的 WideResNet 产生的测试错误为$18. 24\%$ , 这表明我们使用基准模型获得的 $18.29\%$ 错
误有所改善。我们在 CIFAR-100 上的结果仅被 [33] 超越，因为它们的模型使用了大量的参数。当通过非线性卷积滤波器提取时，馈送到卷积组的特征使网络避免了过拟合。这可以通过我们的模型在 CIFAR-100 上的损失图来推断，如图 2 所示。基线宽 ResNet 虽然比基于 Volterra 的宽 ResNet 具有更低的损耗，但产生更高的测试误差。我们的基于 Volterra 的 WideResNets 仅比基线对应物多 0.05% 个参数。表 3 中提供了这些数据集的最佳方法的摘要
<FigureHere>
Output
Output
<FigureHere>
Figure 2. Train loss on CIFAR-100.
Figure 1. Structure of the proposed CNN model (left) and
a typical convolutional block (right).
|
[
[
0.15283203125,
0.0921630859375,
0.2115478515625,
0.0921630859375,
0.2115478515625,
0.106689453125,
0.15283203125,
0.106689453125
],
[
0.476318359375,
0.0921630859375,
0.52294921875,
0.0921630859375,
0.52294921875,
0.106689453125,
0.476318359375,
0.106689453125
],
[
0.5439453125,
0.0921630859375,
0.63232421875,
0.0921630859375,
0.63232421875,
0.106689453125,
0.5439453125,
0.106689453125
],
[
0.65625,
0.09344482421875,
0.7197265625,
0.09344482421875,
0.7197265625,
0.10479736328125,
0.65625,
0.10479736328125
],
[
0.744140625,
0.09344482421875,
0.81640625,
0.09344482421875,
0.81640625,
0.10479736328125,
0.744140625,
0.10479736328125
],
[
0.1544189453125,
0.111083984375,
0.2115478515625,
0.111083984375,
0.2115478515625,
0.12249755859375,
0.1544189453125,
0.12249755859375
],
[
0.67236328125,
0.10919189453125,
0.70361328125,
0.10919189453125,
0.70361328125,
0.125,
0.67236328125,
0.125
],
[
0.498291015625,
0.11553955078125,
0.50244140625,
0.11553955078125,
0.50244140625,
0.121826171875,
0.498291015625,
0.121826171875
],
[
0.5849609375,
0.11553955078125,
0.59228515625,
0.11553955078125,
0.59228515625,
0.121826171875,
0.5849609375,
0.121826171875
],
[
0.67236328125,
0.1256103515625,
0.705078125,
0.1256103515625,
0.705078125,
0.141357421875,
0.67236328125,
0.141357421875
],
[
0.76123046875,
0.126953125,
0.79736328125,
0.126953125,
0.79736328125,
0.1395263671875,
0.76123046875,
0.1395263671875
],
[
0.495849609375,
0.1275634765625,
0.5087890625,
0.1275634765625,
0.5087890625,
0.1395263671875,
0.495849609375,
0.1395263671875
],
[
0.1544189453125,
0.1275634765625,
0.214111328125,
0.1275634765625,
0.214111328125,
0.1395263671875,
0.1544189453125,
0.1395263671875
],
[
0.5849609375,
0.1307373046875,
0.59228515625,
0.1307373046875,
0.59228515625,
0.136962890625,
0.5849609375,
0.136962890625
],
[
0.1568603515625,
0.1427001953125,
0.240234375,
0.1427001953125,
0.240234375,
0.1572265625,
0.1568603515625,
0.1572265625
],
[
0.49267578125,
0.1427001953125,
0.5107421875,
0.1427001953125,
0.5107421875,
0.1558837890625,
0.49267578125,
0.1558837890625
],
[
0.57177734375,
0.143310546875,
0.60546875,
0.143310546875,
0.60546875,
0.1558837890625,
0.57177734375,
0.1558837890625
],
[
0.67333984375,
0.143310546875,
0.70361328125,
0.143310546875,
0.70361328125,
0.1558837890625,
0.67333984375,
0.1558837890625
],
[
0.5703125,
0.1578369140625,
0.60888671875,
0.160400390625,
0.60693359375,
0.175537109375,
0.56884765625,
0.173583984375
],
[
0.67236328125,
0.1578369140625,
0.7060546875,
0.1610107421875,
0.70361328125,
0.1749267578125,
0.6689453125,
0.1717529296875
],
[
0.490234375,
0.160400390625,
0.515625,
0.160400390625,
0.515625,
0.17236328125,
0.490234375,
0.17236328125
],
[
0.1544189453125,
0.160400390625,
0.369384765625,
0.160400390625,
0.369384765625,
0.173583984375,
0.1544189453125,
0.173583984375
],
[
0.76123046875,
0.1610107421875,
0.79736328125,
0.1610107421875,
0.79736328125,
0.173583984375,
0.76123046875,
0.173583984375
],
[
0.486083984375,
0.17236328125,
0.51953125,
0.175537109375,
0.51708984375,
0.1900634765625,
0.483642578125,
0.1875
],
[
0.56787109375,
0.175537109375,
0.609375,
0.175537109375,
0.609375,
0.1900634765625,
0.56787109375,
0.1900634765625
],
[
0.67333984375,
0.1767578125,
0.70166015625,
0.1767578125,
0.70166015625,
0.189453125,
0.67333984375,
0.189453125
],
[
0.484375,
0.1912841796875,
0.51708984375,
0.1912841796875,
0.51708984375,
0.20703125,
0.484375,
0.20703125
],
[
0.7607421875,
0.1912841796875,
0.79736328125,
0.1912841796875,
0.79736328125,
0.20703125,
0.7607421875,
0.20703125
],
[
0.56787109375,
0.1925048828125,
0.609375,
0.1925048828125,
0.609375,
0.20703125,
0.56787109375,
0.20703125
],
[
0.1544189453125,
0.1925048828125,
0.273681640625,
0.1925048828125,
0.273681640625,
0.20703125,
0.1544189453125,
0.20703125
],
[
0.67333984375,
0.1932373046875,
0.70361328125,
0.1932373046875,
0.70361328125,
0.205810546875,
0.67333984375,
0.205810546875
],
[
0.156005859375,
0.208984375,
0.265625,
0.208984375,
0.265625,
0.2235107421875,
0.156005859375,
0.2235107421875
],
[
0.49267578125,
0.208984375,
0.51123046875,
0.208984375,
0.51123046875,
0.2235107421875,
0.49267578125,
0.2235107421875
],
[
0.56787109375,
0.208984375,
0.609375,
0.208984375,
0.609375,
0.2235107421875,
0.56787109375,
0.2235107421875
],
[
0.67236328125,
0.210205078125,
0.70361328125,
0.210205078125,
0.70361328125,
0.222900390625,
0.67236328125,
0.222900390625
],
[
0.76123046875,
0.2083740234375,
0.79833984375,
0.2083740234375,
0.79833984375,
0.2235107421875,
0.76123046875,
0.2235107421875
],
[
0.6708984375,
0.2247314453125,
0.70361328125,
0.2247314453125,
0.70361328125,
0.240478515625,
0.6708984375,
0.240478515625
],
[
0.76123046875,
0.2247314453125,
0.79833984375,
0.2247314453125,
0.79833984375,
0.240478515625,
0.76123046875,
0.240478515625
],
[
0.488525390625,
0.2259521484375,
0.51708984375,
0.2259521484375,
0.51708984375,
0.240478515625,
0.488525390625,
0.240478515625
],
[
0.15283203125,
0.2259521484375,
0.253173828125,
0.2259521484375,
0.253173828125,
0.240478515625,
0.15283203125,
0.240478515625
],
[
0.56787109375,
0.2259521484375,
0.609375,
0.2259521484375,
0.609375,
0.240478515625,
0.56787109375,
0.240478515625
],
[
0.56787109375,
0.2412109375,
0.609375,
0.2412109375,
0.609375,
0.2568359375,
0.56787109375,
0.2568359375
],
[
0.6708984375,
0.2412109375,
0.705078125,
0.2412109375,
0.705078125,
0.2568359375,
0.6708984375,
0.2568359375
],
[
0.488525390625,
0.242431640625,
0.51708984375,
0.242431640625,
0.51708984375,
0.2568359375,
0.488525390625,
0.2568359375
],
[
0.1544189453125,
0.242431640625,
0.28515625,
0.242431640625,
0.28515625,
0.2568359375,
0.1544189453125,
0.2568359375
],
[
0.76318359375,
0.24365234375,
0.79736328125,
0.24365234375,
0.79736328125,
0.255615234375,
0.76318359375,
0.255615234375
],
[
0.1544189453125,
0.2587890625,
0.348876953125,
0.2587890625,
0.348876953125,
0.272705078125,
0.1544189453125,
0.272705078125
],
[
0.491943359375,
0.2587890625,
0.51123046875,
0.2587890625,
0.51123046875,
0.2734375,
0.491943359375,
0.2734375
],
[
0.67236328125,
0.25830078125,
0.70361328125,
0.25830078125,
0.70361328125,
0.2734375,
0.67236328125,
0.2734375
],
[
0.76318359375,
0.260009765625,
0.79736328125,
0.260009765625,
0.79736328125,
0.272705078125,
0.76318359375,
0.272705078125
],
[
0.56884765625,
0.274658203125,
0.60693359375,
0.274658203125,
0.60693359375,
0.290283203125,
0.56884765625,
0.290283203125
],
[
0.67236328125,
0.274658203125,
0.705078125,
0.274658203125,
0.705078125,
0.290283203125,
0.67236328125,
0.290283203125
],
[
0.76123046875,
0.274658203125,
0.79833984375,
0.274658203125,
0.79833984375,
0.290283203125,
0.76123046875,
0.290283203125
],
[
0.49267578125,
0.27587890625,
0.51123046875,
0.27587890625,
0.51123046875,
0.2890625,
0.49267578125,
0.2890625
],
[
0.15283203125,
0.27587890625,
0.275390625,
0.27587890625,
0.275390625,
0.290283203125,
0.15283203125,
0.290283203125
],
[
0.6708984375,
0.291748046875,
0.70361328125,
0.291748046875,
0.70361328125,
0.306884765625,
0.6708984375,
0.306884765625
],
[
0.76123046875,
0.291748046875,
0.79736328125,
0.291748046875,
0.79736328125,
0.306884765625,
0.76123046875,
0.306884765625
],
[
0.56787109375,
0.292236328125,
0.609375,
0.292236328125,
0.609375,
0.306884765625,
0.56787109375,
0.306884765625
],
[
0.1544189453125,
0.29345703125,
0.238525390625,
0.29345703125,
0.238525390625,
0.304931640625,
0.1544189453125,
0.304931640625
],
[
0.491943359375,
0.292236328125,
0.51123046875,
0.292236328125,
0.51123046875,
0.30615234375,
0.491943359375,
0.30615234375
],
[
0.1544189453125,
0.30810546875,
0.45751953125,
0.309326171875,
0.45751953125,
0.323974609375,
0.1544189453125,
0.322509765625
],
[
0.6708984375,
0.30810546875,
0.70361328125,
0.30810546875,
0.70361328125,
0.323974609375,
0.6708984375,
0.323974609375
],
[
0.76123046875,
0.30810546875,
0.79833984375,
0.30810546875,
0.79833984375,
0.323974609375,
0.76123046875,
0.323974609375
],
[
0.56640625,
0.309326171875,
0.609375,
0.309326171875,
0.609375,
0.323974609375,
0.56640625,
0.323974609375
],
[
0.491943359375,
0.309326171875,
0.51123046875,
0.309326171875,
0.51123046875,
0.322509765625,
0.491943359375,
0.322509765625
],
[
0.6708984375,
0.322509765625,
0.7060546875,
0.32568359375,
0.70361328125,
0.341552734375,
0.66845703125,
0.33837890625
],
[
0.76123046875,
0.324462890625,
0.79833984375,
0.324462890625,
0.79833984375,
0.34033203125,
0.76123046875,
0.34033203125
],
[
0.491943359375,
0.32568359375,
0.51123046875,
0.32568359375,
0.51123046875,
0.34033203125,
0.491943359375,
0.34033203125
],
[
0.1544189453125,
0.32568359375,
0.325927734375,
0.32568359375,
0.325927734375,
0.339111328125,
0.1544189453125,
0.339111328125
],
[
0.56787109375,
0.32568359375,
0.609375,
0.32568359375,
0.609375,
0.34033203125,
0.56787109375,
0.34033203125
],
[
0.67236328125,
0.34228515625,
0.705078125,
0.34228515625,
0.705078125,
0.35791015625,
0.67236328125,
0.35791015625
],
[
0.76123046875,
0.34228515625,
0.79833984375,
0.34228515625,
0.79833984375,
0.35791015625,
0.76123046875,
0.35791015625
],
[
0.491943359375,
0.343505859375,
0.51123046875,
0.343505859375,
0.51123046875,
0.35791015625,
0.491943359375,
0.35791015625
],
[
0.1544189453125,
0.343505859375,
0.294189453125,
0.343505859375,
0.294189453125,
0.356689453125,
0.1544189453125,
0.356689453125
],
[
0.56640625,
0.343505859375,
0.611328125,
0.343505859375,
0.611328125,
0.35791015625,
0.56640625,
0.35791015625
],
[
0.156005859375,
0.35986328125,
0.33251953125,
0.35986328125,
0.33251953125,
0.373779296875,
0.156005859375,
0.373779296875
],
[
0.491943359375,
0.35986328125,
0.5107421875,
0.35986328125,
0.5107421875,
0.373779296875,
0.491943359375,
0.373779296875
],
[
0.56787109375,
0.35986328125,
0.611328125,
0.35986328125,
0.611328125,
0.375,
0.56787109375,
0.375
],
[
0.67236328125,
0.359130859375,
0.70361328125,
0.359130859375,
0.70361328125,
0.375,
0.67236328125,
0.375
],
[
0.76318359375,
0.361083984375,
0.79736328125,
0.361083984375,
0.79736328125,
0.373779296875,
0.76318359375,
0.373779296875
],
[
0.08087158203125,
0.375732421875,
0.1356201171875,
0.375732421875,
0.1356201171875,
0.386962890625,
0.08087158203125,
0.386962890625
],
[
0.1422119140625,
0.375732421875,
0.89208984375,
0.375732421875,
0.89208984375,
0.389404296875,
0.1422119140625,
0.389404296875
],
[
0.080078125,
0.389404296875,
0.34716796875,
0.389404296875,
0.34716796875,
0.40283203125,
0.080078125,
0.40283203125
],
[
0.08416748046875,
0.4287109375,
0.468017578125,
0.4287109375,
0.468017578125,
0.443115234375,
0.08416748046875,
0.443115234375
],
[
0.50341796875,
0.4287109375,
0.890625,
0.4287109375,
0.890625,
0.443115234375,
0.50341796875,
0.443115234375
],
[
0.08087158203125,
0.4443359375,
0.468017578125,
0.4443359375,
0.468017578125,
0.458251953125,
0.08087158203125,
0.458251953125
],
[
0.50341796875,
0.4443359375,
0.890625,
0.4443359375,
0.890625,
0.458251953125,
0.50341796875,
0.458251953125
],
[
0.08251953125,
0.460205078125,
0.4697265625,
0.460205078125,
0.4697265625,
0.473388671875,
0.08251953125,
0.473388671875
],
[
0.50244140625,
0.458984375,
0.89208984375,
0.460205078125,
0.89208984375,
0.473388671875,
0.50244140625,
0.472900390625
],
[
0.08087158203125,
0.474853515625,
0.468017578125,
0.474853515625,
0.468017578125,
0.488037109375,
0.08087158203125,
0.488037109375
],
[
0.50634765625,
0.474853515625,
0.890625,
0.474853515625,
0.890625,
0.488037109375,
0.50634765625,
0.488037109375
],
[
0.08251953125,
0.4912109375,
0.46728515625,
0.4912109375,
0.46728515625,
0.501953125,
0.08251953125,
0.501953125
],
[
0.50244140625,
0.488037109375,
0.890625,
0.4892578125,
0.890625,
0.50390625,
0.50244140625,
0.50244140625
],
[
0.50341796875,
0.5048828125,
0.890625,
0.5048828125,
0.890625,
0.51806640625,
0.50341796875,
0.51806640625
],
[
0.5048828125,
0.521484375,
0.8896484375,
0.521484375,
0.8896484375,
0.5322265625,
0.5048828125,
0.5322265625
],
[
0.1544189453125,
0.53271484375,
0.1846923828125,
0.53271484375,
0.1846923828125,
0.54541015625,
0.1544189453125,
0.54541015625
],
[
0.50341796875,
0.53515625,
0.890625,
0.53515625,
0.890625,
0.548828125,
0.50341796875,
0.548828125
],
[
0.16259765625,
0.5517578125,
0.1748046875,
0.5517578125,
0.1748046875,
0.56298828125,
0.16259765625,
0.56298828125
],
[
0.50244140625,
0.5498046875,
0.6650390625,
0.5498046875,
0.6650390625,
0.56298828125,
0.50244140625,
0.56298828125
],
[
0.125,
0.568359375,
0.214111328125,
0.568359375,
0.214111328125,
0.5791015625,
0.125,
0.5791015625
],
[
0.295654296875,
0.5791015625,
0.325927734375,
0.5791015625,
0.325927734375,
0.59130859375,
0.295654296875,
0.59130859375
],
[
0.1405029296875,
0.59521484375,
0.19775390625,
0.59521484375,
0.19775390625,
0.60595703125,
0.1405029296875,
0.60595703125
],
[
0.30810546875,
0.609375,
0.314453125,
0.609375,
0.314453125,
0.6123046875,
0.30810546875,
0.6123046875
],
[
0.148681640625,
0.62744140625,
0.1895751953125,
0.62744140625,
0.1895751953125,
0.6396484375,
0.148681640625,
0.6396484375
],
[
0.265625,
0.6220703125,
0.35546875,
0.6220703125,
0.35546875,
0.63330078125,
0.265625,
0.63330078125
],
[
0.54736328125,
0.6298828125,
0.56787109375,
0.6298828125,
0.56787109375,
0.642578125,
0.54736328125,
0.642578125
],
[
0.294189453125,
0.6484375,
0.326904296875,
0.6484375,
0.326904296875,
0.66015625,
0.294189453125,
0.66015625
],
[
0.69677734375,
0.64453125,
0.82666015625,
0.64453125,
0.82666015625,
0.65771484375,
0.69677734375,
0.65771484375
],
[
0.148681640625,
0.65478515625,
0.1895751953125,
0.65478515625,
0.1895751953125,
0.66552734375,
0.148681640625,
0.66552734375
],
[
0.54736328125,
0.66357421875,
0.5703125,
0.66357421875,
0.5703125,
0.6767578125,
0.54736328125,
0.6767578125
],
[
0.69677734375,
0.6591796875,
0.78759765625,
0.6591796875,
0.78759765625,
0.67041015625,
0.69677734375,
0.67041015625
],
[
0.148681640625,
0.68115234375,
0.1895751953125,
0.68115234375,
0.1895751953125,
0.6923828125,
0.148681640625,
0.6923828125
],
[
0.281982421875,
0.67529296875,
0.339111328125,
0.67529296875,
0.339111328125,
0.68603515625,
0.281982421875,
0.68603515625
],
[
0.2646484375,
0.70068359375,
0.35693359375,
0.70068359375,
0.35693359375,
0.7109375,
0.2646484375,
0.7109375
],
[
0.375732421875,
0.70068359375,
0.418212890625,
0.70263671875,
0.416748046875,
0.7138671875,
0.374267578125,
0.7119140625
],
[
0.54833984375,
0.70068359375,
0.56787109375,
0.70068359375,
0.56787109375,
0.71337890625,
0.54833984375,
0.71337890625
],
[
0.125,
0.71337890625,
0.2132568359375,
0.71337890625,
0.2132568359375,
0.7236328125,
0.125,
0.7236328125
],
[
0.52294921875,
0.71630859375,
0.57421875,
0.71630859375,
0.57421875,
0.75390625,
0.52294921875,
0.75390625
],
[
0.295654296875,
0.72802734375,
0.326904296875,
0.72802734375,
0.326904296875,
0.7392578125,
0.295654296875,
0.7392578125
],
[
0.15283203125,
0.74072265625,
0.1846923828125,
0.74072265625,
0.1846923828125,
0.751953125,
0.15283203125,
0.751953125
],
[
0.52685546875,
0.744140625,
0.54736328125,
0.744140625,
0.54736328125,
0.76953125,
0.52685546875,
0.76953125
],
[
0.291748046875,
0.7548828125,
0.330810546875,
0.7548828125,
0.330810546875,
0.7666015625,
0.291748046875,
0.7666015625
],
[
0.1405029296875,
0.7666015625,
0.19775390625,
0.7666015625,
0.19775390625,
0.77783203125,
0.1405029296875,
0.77783203125
],
[
0.54736328125,
0.771484375,
0.5703125,
0.771484375,
0.5703125,
0.7861328125,
0.54736328125,
0.7861328125
],
[
0.279296875,
0.77978515625,
0.340576171875,
0.7822265625,
0.340576171875,
0.79345703125,
0.279296875,
0.79150390625
],
[
0.1282958984375,
0.79345703125,
0.2091064453125,
0.79345703125,
0.2091064453125,
0.8037109375,
0.1282958984375,
0.8037109375
],
[
0.54736328125,
0.8095703125,
0.57177734375,
0.8095703125,
0.57177734375,
0.8212890625,
0.54736328125,
0.8212890625
],
[
0.148681640625,
0.8193359375,
0.188720703125,
0.8193359375,
0.188720703125,
0.83056640625,
0.148681640625,
0.83056640625
],
[
0.324462890625,
0.8173828125,
0.33251953125,
0.8173828125,
0.33251953125,
0.8212890625,
0.324462890625,
0.8212890625
],
[
0.16259765625,
0.8388671875,
0.1748046875,
0.8388671875,
0.1748046875,
0.8515625,
0.16259765625,
0.8515625
],
[
0.3056640625,
0.8388671875,
0.31787109375,
0.8388671875,
0.31787109375,
0.85302734375,
0.3056640625,
0.85302734375
],
[
0.54833984375,
0.84521484375,
0.56884765625,
0.84521484375,
0.56884765625,
0.85595703125,
0.54833984375,
0.85595703125
],
[
0.1519775390625,
0.85302734375,
0.188720703125,
0.85498046875,
0.1871337890625,
0.8671875,
0.1502685546875,
0.86572265625
],
[
0.293212890625,
0.85400390625,
0.330078125,
0.85400390625,
0.330078125,
0.86572265625,
0.293212890625,
0.86572265625
],
[
0.5986328125,
0.85400390625,
0.625,
0.85400390625,
0.625,
0.86572265625,
0.5986328125,
0.86572265625
],
[
0.6396484375,
0.85400390625,
0.6640625,
0.85400390625,
0.6640625,
0.86572265625,
0.6396484375,
0.86572265625
],
[
0.6787109375,
0.85400390625,
0.70361328125,
0.85400390625,
0.70361328125,
0.86572265625,
0.6787109375,
0.86572265625
],
[
0.72119140625,
0.85595703125,
0.74169921875,
0.85595703125,
0.74169921875,
0.8642578125,
0.72119140625,
0.8642578125
],
[
0.7607421875,
0.85595703125,
0.78173828125,
0.85595703125,
0.78173828125,
0.8642578125,
0.7607421875,
0.8642578125
],
[
0.7998046875,
0.85595703125,
0.82177734375,
0.85595703125,
0.82177734375,
0.8642578125,
0.7998046875,
0.8642578125
],
[
0.8388671875,
0.85595703125,
0.861328125,
0.85595703125,
0.861328125,
0.8642578125,
0.8388671875,
0.8642578125
],
[
0.68896484375,
0.86572265625,
0.73388671875,
0.86572265625,
0.73388671875,
0.8798828125,
0.68896484375,
0.8798828125
],
[
0.560546875,
0.85595703125,
0.58251953125,
0.85595703125,
0.58251953125,
0.8642578125,
0.560546875,
0.8642578125
],
[
0.080078125,
0.8720703125,
0.148681640625,
0.87060546875,
0.148681640625,
0.8818359375,
0.080078125,
0.88330078125
],
[
0.143798828125,
0.87060546875,
0.4697265625,
0.87060546875,
0.4697265625,
0.8837890625,
0.143798828125,
0.8837890625
],
[
0.07843017578125,
0.88525390625,
0.32275390625,
0.88525390625,
0.32275390625,
0.8984375,
0.07843017578125,
0.8984375
],
[
0.57421875,
0.88623046875,
0.6357421875,
0.88623046875,
0.6357421875,
0.89794921875,
0.57421875,
0.89794921875
],
[
0.63623046875,
0.88623046875,
0.81396484375,
0.88623046875,
0.81396484375,
0.896484375,
0.63623046875,
0.896484375
]
] |
[
"Network",
"Depth",
"#Parameters",
"CIFAR-10",
"CIFAR-100",
"NIN [20]",
"8.81",
":1:",
"-",
"7.97",
"34.57",
"3",
"DSN [18]",
"-",
"All-CNN [29]",
"9",
"1.3M",
"7.25",
"1.7M",
"5.23",
"110",
"ResNet with Stochastic Depth [12]",
"24.58",
"1202",
"10.2M",
"4.91",
"1001",
"22.71",
"10.2M",
"pre-act Resnet [11]",
"4.62",
"Wide ResNet [34]",
"40",
"55.8M",
"3.80",
"18.30",
"3.77",
"18.29",
"110",
"PyramidNet [9]",
"28.3M",
"20.2M",
"3.76",
"146",
"Wide-DelugeNet [16]",
"19.02",
"OrthoReg on Wide ResNet [26]",
"28",
"3.69",
"18.56",
"9.1M",
"3.65",
"18.82",
"14",
"Steerable CNNs [4]",
"3.58",
"17.31",
"68.1M",
"ResNeXt [33]",
"29",
"Wide ResNet with Singular Value Bounding [14]",
"3.52",
"18.32",
"36.5M",
"28",
"3.52",
"19.22",
"28",
"Oriented Response Net [35]",
"18.4M",
"3.62",
"18.29",
"28",
"Baseline Wide ResNet",
"36.6M",
"Volterra-based Wide ResNet",
"28",
"36.7M",
"3.51",
"18.24",
"Table 3.",
"Test set classification error results on CIFAR-10 and CIFAR-100, using moderate data augmentation (horizontal",
"flipping, padding and 32 × 32 cropping)",
"于Volterra的WideResNet产生了3.51%的测试错误，",
"误有所改善。我们在CIFAR-100上的结果仅被[33]超",
"这显示了我们使用基线模型获得的3.62%错误的改进，",
"越，因为它们的模型使用了大量的参数。当通过非线性",
"从而在此数据集上设置了最新技术。在CIFAR-100中，",
"卷积滤波器提取时，馈送到卷积组的特征使网络避免",
"我们基于Volterra的WideResNet产生的测试错误为",
"了过拟合。这可以通过我们的模型在CIFAR-100上的",
"18.24%，这表明我们使用基准模型获得的18.29%错",
"损失图来推断，如图2所示。基线宽ResNet虽然比基",
"于Volterra的宽ResNet具有更低的损耗，但产生更高",
"的测试误差。我们的基于Volterra的WideResNets仅",
"Input",
"比基线对应物多0.05%个参数。表3中提供了这些数",
"7",
"据集的最佳方法的摘要",
"Batch Normalization",
"Input",
"Convolution",
"",
"Group 1",
"Batch Normalization",
"-4.1",
"ReLU",
"Volterra-based WRN",
"Group2",
"-4.2 ",
"BaselineWRN",
"Group3",
"Convolution",
"BatchNormalization",
"Identity",
"4.3",
"Batch Normalization",
"",
"ReLU",
"ReLU",
"1607",
"Dropout",
"AvgPooling",
"-4.5",
"Convolution",
"Fully-Connected",
"-4.6",
"SoftMax",
"",
"7",
"v",
"-4.7",
"Output",
"Output",
"140",
"150",
"160",
"170",
"180",
"190",
"200",
"Epoch",
"130",
"Figure 1..",
"Structure of the proposed CNN model (left) and",
"a typical convolutional block (right).",
"Figure 2.",
"Train loss on CIFAR-100."
] |
{
"arxiv_id": "1708.07038",
"language": "zh",
"page_idx": 4
}
|
||
1806.00778-zh-0
|
1806.00778-zh-0.jpg
|
# Multi-Cast Attention Networks for Retrieval-based Question Answering and Response Prediction
Yi Tay
Nanyang Technological University
Singapore
ytay$017@$e.ntu.edu.sg
Siu Cheung Hui
Nanyang Technological University
$\operatorname{Singapore}$
asschui@ntu.edu.sg
Luu Anh Tuan
Institute for Infocomm Research
Singapore
at.luu@i2r.a-star.edu.sg
# 1 INTRODUCTION
$\mathbf{ABSTRACT}$
Attention is typically used to select informative sub-phrases that are used for prediction. This paper investigates the novel use of attention as a form of feature augmentation, i.e, casted attention . We propose Multi-Cast Attention Networks (MCAN), a new attention mechanism and general model architecture for a potpourri of ranking tasks in the conversational modeling and question answering domains. Our approach performs a series of soft attention operations, each time casting a scalar feature upon the inner word embeddings. The key idea is to provide a real-valued hint (feature) to a subsequent encoder layer and is targeted at improving the representation learning process. There are several advantages to this design, e.g., it allows an arbitrary number of attention mechanisms to be casted, allowing for multiple attention types (e.g., co-attention, intra-attention) and attention variants (e.g., alignment-pooling, max-pooling, meanpooling) to be executed simultaneously. This not only eliminates the costly need to tune the nature of the co-attention layer, but also provides greater extents of explainability to practitioners. Via extensive experiments on four well-known benchmark datasets, we show that MCAN achieves stateof-the-art performance. On the Ubuntu Dialogue Corpus, MCAN outperforms existing state-of-the-art models by 9% . MCAN also achieves the best performing score to date on the well-studied TrecQA dataset.
文档查询对之间的文本相关性建模是信息检索(IR)研究的核心。直观地说，这使得现实生活中的应用程序种类繁多，从标准的网络搜索到自动聊天机器人。关键思想是，这些系统学习文档-查询对之间的评分功能，提供候选人的排名列表作为输出。这种 IR 系统的相当一部分集中在短文本文档上，例如。g。,回答基于事实的问题或在基于聊天的系统的上下文中选择最佳响应。基于检索的响应和问答(QA)系统的应用是通用的，可能作为一个强大的独立的特定领域的系统或更大的， 通用的聊天系统，如 Alexa 的关键组件。本文提出了一种适用于此类任务的通用神经排序模型
神经网络(或深度学习)在基于检索的系统[8,27,29,39,52] 中引起了相当大的关注。值得注意的是，许多基准测试的主要最新系统现在都是神经模型，几乎完全放弃了传统的特征工程技术。在这些系统中，卷积或递归网络具有最新技术，例如神经注意力[1,25,39],从而在标准基准上获得了非常有竞争力的结果。注意力的关键思想是只提取对预测有用的最相关的信息。在文本数据的上下文中，注意力学习根据文档中的单词和子短语的重要性来加权它们。同样，共同关注机制[5, 28, 50, 54] 是一种关注机制，可以学习文档和查询方面的成对联合关注
注意力是传统上使用的，通常被想象为特征提取器。它的行为可以被认为是一种动态的池化形式，因为它学会了选择和组合不同的单词来形成最终的文档表示。本文将注意力重新想象为特征增强方法的一种形式。投射注意力的目的不是组合学习或汇集，而是为后续层提供提示。据我们所知，这是一种在神经排名模型中利用注意力的新方法。我们首先描述它的优点，以及它如何处理今天设计的现有模型的弱点
通常，注意一次应用于句子[25,39] 。学习最终表示，然后将其传递给预测层。在处理序列对的上下文中，应用了共同注意力，并学习了每个句子的最终表示形式[5,28,54] 。适用于许多现有模型的一个明显缺点是，它们通常仅限于一种注意力变体。在使用一个或多个注意电话的情况下(e。g。、共同注意力和内部注意力等。),级联通常用于融合表示[20,28] 。不幸的是，这会使每个调用的表示大小增加一倍，从而在后续层中产生成本
渴望多个关注电话的理由是直观的。在[20,28]中，同时使用了共同关注和内部关注，因为它们都提供了文档对的不同视图，从而学习了可用于预测的高质量表示。因此，这可以显著提高性能。此外，Co-Attention 也有不同的风格，可以与提取最大均值池[5,54]或基于对齐的池[3,20,28]一起使用。每种共同注意类型都会产生不同的文档表示形式。在最大池化中，根据单词对其他文本序列的最大贡献来提取信号。均值池计算其对整个句子的贡献。对齐池是另一种共同关注的方式，它将语义相似的子短语对齐在一起。因此，不同的池运算符提供了不同的句子对视图。这通常被调整为超参数，即。
# $\mathbf{KEYWORDS}$
## Deep Learning; Information Retrieval; Question Answering; Conversation Modeling
ACM Reference Format:
Yi Tay, Luu Anh Tuan, and Siu Cheung Hui. 2024. Multi-Cast Attention Networks for Retrieval-based Question Answering and Response Prediction . In Proceedings of ACM Conference (Con- $ference'17).$ ACM, New York, NY, USA, 9 pages. https://doi.org/ 10.1145/nnnnnnn.nnnnnn
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the frst page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org.
$Conference’ 17$, July 2017, $Washington$, $DC$, $USA$
$\begin{array}{l}{\odot2024\mathrm{~Association~for~Computing~Machinery.}}\\{\mathrm{ACM~ISBN~978-x-xxxxx-xxxx-x/YY/MM...\$15.00}}\end{array}$
https://doi.org/10.1145/nnnnnn.nnnnnn
|
[
[
0.11273193359375,
0.10418701171875,
0.88623046875,
0.10479736328125,
0.88623046875,
0.12249755859375,
0.11273193359375,
0.121826171875
],
[
0.26953125,
0.1300048828125,
0.73046875,
0.1300048828125,
0.73046875,
0.15087890625,
0.26953125,
0.15087890625
],
[
0.2017822265625,
0.1666259765625,
0.261474609375,
0.1666259765625,
0.261474609375,
0.1817626953125,
0.2017822265625,
0.1817626953125
],
[
0.435546875,
0.1654052734375,
0.56396484375,
0.1673583984375,
0.56396484375,
0.1817626953125,
0.435546875,
0.179931640625
],
[
0.701171875,
0.1673583984375,
0.83642578125,
0.1673583984375,
0.83642578125,
0.1817626953125,
0.701171875,
0.1817626953125
],
[
0.1094970703125,
0.184326171875,
0.3544921875,
0.184326171875,
0.3544921875,
0.1976318359375,
0.1094970703125,
0.1976318359375
],
[
0.38232421875,
0.184326171875,
0.619140625,
0.184326171875,
0.619140625,
0.1976318359375,
0.38232421875,
0.1976318359375
],
[
0.64453125,
0.18310546875,
0.890625,
0.18310546875,
0.890625,
0.1962890625,
0.64453125,
0.1962890625
],
[
0.193603515625,
0.1976318359375,
0.26953125,
0.199462890625,
0.26953125,
0.2139892578125,
0.193603515625,
0.212158203125
],
[
0.463134765625,
0.1988525390625,
0.53759765625,
0.1988525390625,
0.53759765625,
0.21337890625,
0.463134765625,
0.21337890625
],
[
0.73193359375,
0.1962890625,
0.80615234375,
0.199462890625,
0.8056640625,
0.21533203125,
0.73046875,
0.212158203125
],
[
0.1519775390625,
0.21337890625,
0.310546875,
0.2139892578125,
0.310546875,
0.2291259765625,
0.1519775390625,
0.2279052734375
],
[
0.41259765625,
0.21337890625,
0.58837890625,
0.2139892578125,
0.58837890625,
0.2291259765625,
0.41259765625,
0.2279052734375
],
[
0.69775390625,
0.21337890625,
0.8388671875,
0.2139892578125,
0.8388671875,
0.2291259765625,
0.69775390625,
0.2279052734375
],
[
0.51953125,
0.2374267578125,
0.5361328125,
0.240478515625,
0.53173828125,
0.250732421875,
0.515625,
0.2474365234375
],
[
0.54736328125,
0.2374267578125,
0.69921875,
0.2374267578125,
0.69921875,
0.250732421875,
0.54736328125,
0.250732421875
],
[
0.5361328125,
0.2412109375,
0.54345703125,
0.2412109375,
0.54345703125,
0.246826171875,
0.5361328125,
0.246826171875
],
[
0.08905029296875,
0.2392578125,
0.1871337890625,
0.2392578125,
0.1871337890625,
0.25,
0.08905029296875,
0.25
],
[
0.08660888671875,
0.255615234375,
0.48193359375,
0.255615234375,
0.48193359375,
0.26953125,
0.08660888671875,
0.26953125
],
[
0.521484375,
0.255615234375,
0.91259765625,
0.255615234375,
0.91259765625,
0.26953125,
0.521484375,
0.26953125
],
[
0.08660888671875,
0.271484375,
0.4794921875,
0.271484375,
0.4794921875,
0.2822265625,
0.08660888671875,
0.2822265625
],
[
0.51708984375,
0.268310546875,
0.91259765625,
0.26953125,
0.91259765625,
0.28271484375,
0.51708984375,
0.2822265625
],
[
0.0849609375,
0.2841796875,
0.48046875,
0.2841796875,
0.48046875,
0.29736328125,
0.0849609375,
0.29736328125
],
[
0.5185546875,
0.2841796875,
0.91259765625,
0.2841796875,
0.91259765625,
0.29736328125,
0.5185546875,
0.29736328125
],
[
0.08251953125,
0.29736328125,
0.48193359375,
0.296630859375,
0.48193359375,
0.311279296875,
0.08251953125,
0.3125
],
[
0.51953125,
0.29736328125,
0.91259765625,
0.29736328125,
0.91259765625,
0.311279296875,
0.51953125,
0.311279296875
],
[
0.08660888671875,
0.313232421875,
0.48046875,
0.313232421875,
0.48046875,
0.323974609375,
0.08660888671875,
0.323974609375
],
[
0.51953125,
0.313232421875,
0.9111328125,
0.313232421875,
0.9111328125,
0.323974609375,
0.51953125,
0.323974609375
],
[
0.0849609375,
0.32568359375,
0.48193359375,
0.32568359375,
0.48193359375,
0.339111328125,
0.0849609375,
0.339111328125
],
[
0.51708984375,
0.32568359375,
0.9140625,
0.32568359375,
0.9140625,
0.339111328125,
0.51708984375,
0.339111328125
],
[
0.08416748046875,
0.33837890625,
0.4794921875,
0.339111328125,
0.4794921875,
0.354248046875,
0.08416748046875,
0.352783203125
],
[
0.5185546875,
0.339111328125,
0.91259765625,
0.339111328125,
0.91259765625,
0.352783203125,
0.5185546875,
0.352783203125
],
[
0.08660888671875,
0.354248046875,
0.4794921875,
0.354248046875,
0.4794921875,
0.367431640625,
0.08660888671875,
0.367431640625
],
[
0.51953125,
0.354248046875,
0.91259765625,
0.354248046875,
0.91259765625,
0.367431640625,
0.51953125,
0.367431640625
],
[
0.0849609375,
0.367431640625,
0.483642578125,
0.367431640625,
0.483642578125,
0.380615234375,
0.0849609375,
0.380615234375
],
[
0.5185546875,
0.367431640625,
0.91259765625,
0.367431640625,
0.91259765625,
0.380615234375,
0.5185546875,
0.380615234375
],
[
0.08416748046875,
0.380126953125,
0.48193359375,
0.380615234375,
0.48193359375,
0.39453125,
0.08416748046875,
0.393310546875
],
[
0.5185546875,
0.380615234375,
0.740234375,
0.380615234375,
0.740234375,
0.39453125,
0.5185546875,
0.39453125
],
[
0.08660888671875,
0.395751953125,
0.48193359375,
0.395751953125,
0.48193359375,
0.4091796875,
0.08660888671875,
0.4091796875
],
[
0.53515625,
0.39453125,
0.91259765625,
0.39453125,
0.91259765625,
0.40771484375,
0.53515625,
0.40771484375
],
[
0.0849609375,
0.4091796875,
0.48193359375,
0.4091796875,
0.48193359375,
0.42236328125,
0.0849609375,
0.42236328125
],
[
0.51953125,
0.40771484375,
0.91259765625,
0.40771484375,
0.91259765625,
0.421630859375,
0.51953125,
0.421630859375
],
[
0.0849609375,
0.42236328125,
0.483642578125,
0.42236328125,
0.483642578125,
0.436279296875,
0.0849609375,
0.436279296875
],
[
0.51953125,
0.423583984375,
0.9111328125,
0.423583984375,
0.9111328125,
0.434326171875,
0.51953125,
0.434326171875
],
[
0.0849609375,
0.436279296875,
0.48193359375,
0.436279296875,
0.48193359375,
0.449462890625,
0.0849609375,
0.449462890625
],
[
0.51953125,
0.4375,
0.9111328125,
0.4375,
0.9111328125,
0.447509765625,
0.51953125,
0.447509765625
],
[
0.0849609375,
0.45068359375,
0.48193359375,
0.45068359375,
0.48193359375,
0.464111328125,
0.0849609375,
0.464111328125
],
[
0.51953125,
0.449462890625,
0.91259765625,
0.449462890625,
0.91259765625,
0.46337890625,
0.51953125,
0.46337890625
],
[
0.0849609375,
0.464111328125,
0.48193359375,
0.464111328125,
0.48193359375,
0.479248046875,
0.0849609375,
0.479248046875
],
[
0.5185546875,
0.462158203125,
0.91259765625,
0.464111328125,
0.91259765625,
0.477783203125,
0.5185546875,
0.47607421875
],
[
0.0849609375,
0.477783203125,
0.483642578125,
0.477783203125,
0.483642578125,
0.4912109375,
0.0849609375,
0.4912109375
],
[
0.521484375,
0.479248046875,
0.9111328125,
0.479248046875,
0.9111328125,
0.4892578125,
0.521484375,
0.4892578125
],
[
0.08660888671875,
0.49365234375,
0.4794921875,
0.49365234375,
0.4794921875,
0.50390625,
0.08660888671875,
0.50390625
],
[
0.5185546875,
0.4912109375,
0.9140625,
0.4912109375,
0.9140625,
0.5048828125,
0.5185546875,
0.5048828125
],
[
0.08416748046875,
0.5048828125,
0.48193359375,
0.50390625,
0.48193359375,
0.51806640625,
0.08416748046875,
0.51953125
],
[
0.51708984375,
0.50390625,
0.9140625,
0.5048828125,
0.9140625,
0.51953125,
0.51708984375,
0.51806640625
],
[
0.08660888671875,
0.51953125,
0.48046875,
0.51953125,
0.48046875,
0.53271484375,
0.08660888671875,
0.53271484375
],
[
0.51953125,
0.52099609375,
0.60693359375,
0.52099609375,
0.60693359375,
0.53076171875,
0.51953125,
0.53076171875
],
[
0.08660888671875,
0.53271484375,
0.48046875,
0.53271484375,
0.48046875,
0.546875,
0.08660888671875,
0.546875
],
[
0.53173828125,
0.5322265625,
0.9140625,
0.53271484375,
0.9140625,
0.54736328125,
0.53173828125,
0.546875
],
[
0.0849609375,
0.546875,
0.48046875,
0.546875,
0.48046875,
0.56103515625,
0.0849609375,
0.56103515625
],
[
0.5185546875,
0.546875,
0.91259765625,
0.546875,
0.91259765625,
0.56005859375,
0.5185546875,
0.56005859375
],
[
0.0849609375,
0.56005859375,
0.4794921875,
0.56005859375,
0.4794921875,
0.57373046875,
0.0849609375,
0.57373046875
],
[
0.51708984375,
0.56005859375,
0.91259765625,
0.56103515625,
0.91259765625,
0.57470703125,
0.51708984375,
0.57373046875
],
[
0.08416748046875,
0.57373046875,
0.48193359375,
0.57470703125,
0.48193359375,
0.58837890625,
0.08416748046875,
0.5869140625
],
[
0.51953125,
0.5771484375,
0.9111328125,
0.5771484375,
0.9111328125,
0.5869140625,
0.51953125,
0.5869140625
],
[
0.0849609375,
0.58837890625,
0.30810546875,
0.58837890625,
0.30810546875,
0.6015625,
0.0849609375,
0.6015625
],
[
0.51953125,
0.58837890625,
0.900390625,
0.58837890625,
0.900390625,
0.6015625,
0.51953125,
0.6015625
],
[
0.51953125,
0.6015625,
0.91259765625,
0.6015625,
0.91259765625,
0.61572265625,
0.51953125,
0.61572265625
],
[
0.08660888671875,
0.61572265625,
0.1953125,
0.61572265625,
0.1953125,
0.62890625,
0.08660888671875,
0.62890625
],
[
0.5185546875,
0.6162109375,
0.85693359375,
0.6162109375,
0.85693359375,
0.6298828125,
0.5185546875,
0.6298828125
],
[
0.0849609375,
0.6337890625,
0.48046875,
0.6337890625,
0.48046875,
0.64794921875,
0.0849609375,
0.64794921875
],
[
0.53515625,
0.6298828125,
0.91259765625,
0.6298828125,
0.91259765625,
0.64306640625,
0.53515625,
0.64306640625
],
[
0.08251953125,
0.646484375,
0.2418212890625,
0.6484375,
0.2418212890625,
0.662109375,
0.08251953125,
0.66015625
],
[
0.5185546875,
0.64306640625,
0.91259765625,
0.64306640625,
0.91259765625,
0.6572265625,
0.5185546875,
0.6572265625
],
[
0.51953125,
0.65771484375,
0.91259765625,
0.65771484375,
0.91259765625,
0.671875,
0.51953125,
0.671875
],
[
0.0849609375,
0.66845703125,
0.2303466796875,
0.66845703125,
0.2303466796875,
0.681640625,
0.0849609375,
0.681640625
],
[
0.51953125,
0.67041015625,
0.91259765625,
0.67041015625,
0.91259765625,
0.6845703125,
0.51953125,
0.6845703125
],
[
0.0849609375,
0.68115234375,
0.48046875,
0.68115234375,
0.48046875,
0.69580078125,
0.0849609375,
0.69580078125
],
[
0.51708984375,
0.6845703125,
0.9140625,
0.68505859375,
0.9140625,
0.69970703125,
0.51708984375,
0.69873046875
],
[
0.08660888671875,
0.69580078125,
0.48046875,
0.69580078125,
0.48046875,
0.70556640625,
0.08660888671875,
0.70556640625
],
[
0.5185546875,
0.69970703125,
0.9140625,
0.69970703125,
0.9140625,
0.71337890625,
0.5185546875,
0.71337890625
],
[
0.0849609375,
0.70703125,
0.48046875,
0.70703125,
0.48046875,
0.72021484375,
0.0849609375,
0.72021484375
],
[
0.5185546875,
0.7119140625,
0.9140625,
0.7119140625,
0.9140625,
0.72607421875,
0.5185546875,
0.72607421875
],
[
0.0849609375,
0.7197265625,
0.48193359375,
0.7197265625,
0.48193359375,
0.73291015625,
0.0849609375,
0.73291015625
],
[
0.08416748046875,
0.73193359375,
0.250732421875,
0.73291015625,
0.250732421875,
0.744140625,
0.08416748046875,
0.74365234375
],
[
0.51953125,
0.72802734375,
0.60693359375,
0.72802734375,
0.60693359375,
0.7392578125,
0.51953125,
0.7392578125
],
[
0.53369140625,
0.7412109375,
0.91259765625,
0.7412109375,
0.91259765625,
0.7548828125,
0.53369140625,
0.7548828125
],
[
0.08416748046875,
0.75830078125,
0.48046875,
0.7587890625,
0.48046875,
0.77294921875,
0.08416748046875,
0.771484375
],
[
0.51953125,
0.75390625,
0.9140625,
0.75390625,
0.9140625,
0.767578125,
0.51953125,
0.767578125
],
[
0.238525390625,
0.77099609375,
0.47802734375,
0.76953125,
0.47802734375,
0.78271484375,
0.238525390625,
0.7841796875
],
[
0.51708984375,
0.7666015625,
0.91259765625,
0.767578125,
0.91259765625,
0.7822265625,
0.51708984375,
0.78076171875
],
[
0.2156982421875,
0.7763671875,
0.226318359375,
0.7763671875,
0.226318359375,
0.77978515625,
0.2156982421875,
0.77978515625
],
[
0.08660888671875,
0.78076171875,
0.476318359375,
0.78076171875,
0.476318359375,
0.79150390625,
0.08660888671875,
0.79150390625
],
[
0.5185546875,
0.7822265625,
0.91259765625,
0.7822265625,
0.91259765625,
0.79541015625,
0.5185546875,
0.79541015625
],
[
0.08660888671875,
0.79052734375,
0.47802734375,
0.79052734375,
0.47802734375,
0.80029296875,
0.08660888671875,
0.80029296875
],
[
0.51953125,
0.796875,
0.9140625,
0.796875,
0.9140625,
0.81005859375,
0.51953125,
0.81005859375
],
[
0.08660888671875,
0.8017578125,
0.48046875,
0.80029296875,
0.48046875,
0.8125,
0.08660888671875,
0.81298828125
],
[
0.08660888671875,
0.771484375,
0.2181396484375,
0.771484375,
0.2181396484375,
0.7822265625,
0.08660888671875,
0.7822265625
],
[
0.08416748046875,
0.80810546875,
0.48046875,
0.81005859375,
0.48046875,
0.82470703125,
0.08416748046875,
0.82275390625
],
[
0.5185546875,
0.8095703125,
0.91259765625,
0.81005859375,
0.91259765625,
0.82373046875,
0.5185546875,
0.82275390625
],
[
0.08660888671875,
0.82275390625,
0.4794921875,
0.82275390625,
0.4794921875,
0.83349609375,
0.08660888671875,
0.83349609375
],
[
0.5185546875,
0.82373046875,
0.91259765625,
0.82373046875,
0.91259765625,
0.8369140625,
0.5185546875,
0.8369140625
],
[
0.08905029296875,
0.833984375,
0.4794921875,
0.833984375,
0.4794921875,
0.84130859375,
0.08905029296875,
0.84130859375
],
[
0.08660888671875,
0.84228515625,
0.240234375,
0.84228515625,
0.240234375,
0.85302734375,
0.08660888671875,
0.85302734375
],
[
0.51953125,
0.8388671875,
0.9111328125,
0.8388671875,
0.9111328125,
0.849609375,
0.51953125,
0.849609375
],
[
0.08660888671875,
0.85302734375,
0.37744140625,
0.85302734375,
0.37744140625,
0.86279296875,
0.08660888671875,
0.86279296875
],
[
0.5185546875,
0.8515625,
0.91259765625,
0.8515625,
0.91259765625,
0.86572265625,
0.5185546875,
0.86572265625
],
[
0.08416748046875,
0.861328125,
0.346435546875,
0.86279296875,
0.346435546875,
0.87744140625,
0.08416748046875,
0.87548828125
],
[
0.51708984375,
0.8642578125,
0.91259765625,
0.86572265625,
0.91259765625,
0.87890625,
0.51708984375,
0.87744140625
],
[
0.0882568359375,
0.87353515625,
0.36279296875,
0.87353515625,
0.36279296875,
0.8837890625,
0.0882568359375,
0.8837890625
],
[
0.08660888671875,
0.8837890625,
0.33251953125,
0.8837890625,
0.33251953125,
0.89453125,
0.08660888671875,
0.89453125
]
] |
[
"Multi-Cast Attention Networks for Retrieval-based Question",
"Answering and Response Prediction",
"Yi Tay",
"Luu Anh Tuan",
"Siu Cheung Hui",
"Nanyang Technological University",
"Institutefor InfocommResearch",
"Nanyang Technological University",
"Singapore",
"Singapore",
"Singapore",
"ytay017@e.ntu.edu.sg",
"at.luu@i2r.a-star.edu.sg",
"asschui@ntu.edu.sg",
"1",
"INTRODUCTION",
"",
"ABSTRACT",
"Attention is typically used to select informative sub-phrases",
"文档查询对之间的文本相关性建模是信息检索（IR）研究的",
"that are used for prediction. This paper investigates the",
"核心。直观地说，这使得现实生活中的应用程序种类繁多，从",
"novel use of attention as a form of feature augmentation,",
"标准的网络搜索到自动聊天机器人。关键思想是，这些系统学",
"i.e, casted attention . We propose Multi-Cast Attention Net-",
"习文档-查询对之间的评分功能，提供候选人的排名列表作为",
"works (MCAN), a new attention mechanism and general",
"输出。这种IR系统的相当一部分集中在短文本文档上，例如。",
"model architecture for a potpourri of ranking tasks in the",
"g。，回答基于事实的问题或在基于聊天的系统的上下文中选",
"conversational modeling and question answering domains",
"择最佳响应。基于检索的响应和问答（QA）系统的应用是通",
"Our approach performs a series of soft attention operations",
"用的，可能作为一个强大的独立的特定领域的系统或更大的，",
"each time casting a scalar feature upon the inner word em-",
"通用的聊天系统，如Alexa的关键组件。本文提出了一种适",
"beddings. The key idea is to provide a real-valued hint (fea-",
"用于此类任务的通用神经排序模型",
"ture) to a subsequent encoder layer and is targeted at improv-",
"神经网络（或深度学习)在基于检索的系统[8,27,29,39,52]",
"ing the representation learning process. There are several ad-",
"中引起了相当大的关注。值得注意的是，许多基准测试的主要",
"vantages to this design, e.g., it allows an arbitrary number of",
"最新系统现在都是神经模型，几乎完全放弃了传统的特征工",
"attention mechanisms to be casted, allowing for multiple at-",
"程技术。在这些系统中，卷积或递归网络具有最新技术，例如",
"tention types (e.g., co-attention, intra-attention) and atten-",
"神经注意力[1，25，39]，从而在标准基准上获得了非常有竞",
"tion variants (e.g., alignment-pooling, max-pooling, mean-",
"争力的结果。注意力的关键思想是只提取对预测有用的最相",
"pooling) to be executed simultaneously. This not only elimi-",
"关的信息。在文本数据的上下文中，注意力学习根据文档中",
"nates the costly need totunethe nature of the co-attention",
"的单词和子短语的重要性来加权它们。同样，共同关注机制",
"layer, but also provides greater extents of explainability to",
"[5，28，50,54]是一种关注机制，可以学习文档和查询方面的",
"practitioners. Via extensive experiments on four well-known",
"成对联合关注",
"benchmark datasets, we show that MCAN achieves state",
"注意力是传统上使用的，通常被想象为特征提取器。它的行",
"of-the-art performance. On the Ubuntu Dialogue Corpus,",
"为可以被认为是一种动态的池化形式，因为它学会了选择和",
"MCAN outperforms existing state-of-the-art models by 9%",
"组合不同的单词来形成最终的文档表示。本文将注意力重新",
"MCAN also achieves the best performing score to date on",
"想象为特征增强方法的一种形式。投射注意力的目的不是组",
"the well-studied TrecQA dataset.",
"合学习或汇集，而是为后续层提供提示。据我们所知，这是",
"种在神经排名模型中利用注意力的新方法。我们首先描述它",
"KEYWORDS",
"的优点，以及它如何处理今天设计的现有模型的弱点",
"Deep Learning; Information Retrieval; Question Answering;",
"通常，注意一次应用于句子[25,39]。学习最终表示，然后",
"Conversation Modeling",
"将其传递给预测层。在处理序列对的上下文中，应用了共同注",
"意力，并学习了每个句子的最终表示形式[5，28，54]】。适用于",
"ACM Reference Format:",
"许多现有模型的一个明显缺点是，它们通常仅限于一种注意",
"Yi Tay, Luu Anh Tuan, and Siu Cheung Hui. 2024. Multi-Cast",
"力变体。在使用一个或多个注意电话的情况下（e。g。、共同",
"AttentionNetworksforRetrieval-basedQuestionAnsweringand",
"注意力和内部注意力等。），级联通常用于融合表示[20，28]。",
"Response Prediction : In Proceedings of ACM Conference (Con",
"不幸的是，这会使每个调用的表示大小增加一倍，从而在后续",
"ference'17). ACM, New York, NY, USA, 9 pages. https://doi.org/",
"10.1145/nnnnnnn.nnnnnnn",
"层中产生成本",
"渴望多个关注电话的理由是直观的。在[20，28]中，同时",
"Permission to make digital or hard copies of all or part of this work",
"使用了共同关注和内部关注，因为它们都提供了文档对的不",
"use is granted without fee provided that",
"同视图，从而学习了可用于预测的高质量表示。因此，这可以",
"",
"copies are not made or distributed for profit or commercial advantag",
"显著提高性能。此外，Co-Attention也有不同的风格，可以",
"and that copies bear this notice and the full citation on the first",
"与提取最大均值池[5,54]或基于对齐的池[3，20,28]一起使",
"page. Copyrights for components of this work owned by others than",
"for personal or classro",
"ACM must be honored.Abstracting with credit is permitted, To copy",
"用。每种共同注意类型都会产生不同的文档表示形式。在最大",
"otherwise, or republish, to post on servers or to redistribute to lists,",
"池化中，根据单词对其他文本序列的最大贡献来提取信号。均",
"requirespriorspecificpermissionand/orafee.Requestpermissions",
"frompermissions@acm.org",
"值池计算其对整个句子的贡献。对齐池是另一种共同关注的",
"Conference'17,July 2017,Washington,DC,USA",
"方式，它将语义相似的子短语对齐在一起。因此，不同的池运",
"2024 Association for Computing Machinery",
"算符提供了不同的句子对视图。这通常被调整为超参数，即。",
"ACM ISBN 978-x-xxxx-xxxx-x/YY/MM...$15.00",
"https://doi.org/10.1145/nnnnnnn.nnnnnnn"
] |
{
"arxiv_id": "1806.00778",
"language": "zh",
"page_idx": 0
}
|
||
1806.00778-zh-1
|
1806.00778-zh-1.jpg
|
e。, 执行体系结构工程以找到在每个问题域和数据集上效果最
好的最佳变体
我们的方法旨在服务于两个重要目的-(1)通过使注意力能够被调用任意$k$ 次而几乎没有任何后果，从而消除了对该组件的体系结构工程的需求，并且(2)同时通过对多个视图进行建模来提高性能通过多次注意调用。因此，我们的方法与多头注意力具有相似的精神，尽管效率很高。为此，我们引入了 Multi-CastAttentionNetworks(MCAN), 这是一种新的深度学习架构，用于问答和对话建模领域的任务。在我们的方法中，注意力被投射，与大多数其他将其用作汇集操作的作品相反。我们多次投射共同注意力，每次返回一个压缩的标量特征，该特征重新附加到原始单词表示。关键的直觉是，压缩可以实现多个注意力调用的可扩展投射，旨在为后续层提供不仅具有全局知识而且具有跨句子知识的提示。直观地，当将这些增强的嵌入传递到组成编码器(例如长短期记忆编码器)中时，LSTM 可以从此提示中受益，并相应地改变其表示学习过程。
## 1.1 0ur Contributions
总而言之，这项工作的主要贡献是：
$\bullet$ For the first time, we propose a new paradigm of utilizing attentions not as a pooling operator but as a form of feature augmentation. We propose an overall architecture, Multi-Cast Attention Networks (MCAN) for generic sequence pair modeling.
$\bullet$ We evaluate our proposed model on four benchmark
tasks, i. e. , Dialogue Reply Prediction ( Ubuntu dialogue corpus), Factoid Question Answering (TrecQA), Community Question Answering (QatarLiving forums from SemEval 2016) and Tweet Reply Prediction (Customer support). On Ubuntu dialogue corpus, MCAN outperforms the existing state-of-the-art models by 9% . MCAN also achieves the best performing score of 0.838 MAP and 0.904 MRR on the well-studied TrecQA dataset.
$\bullet$ We provide a comprehensive and in-depth analysis of the inner workings of our proposed MCAN model. We show that the casted attention features are interpretable and are capable of learning (1) a neural adaptation of word overlap and (2) a differentiation of evidence and anti-evidence words/patterns.
2 MULTI-CAST ATTENTION NETWORKS
在本节中，我们将描述我们提出的 MCAN 模型。我们模型的输人是两个文本序列，我们将其表示为查询$q$和文档$d$ 。在我们的问题中，查询文档可以推广到不同的问题领域，例如问答或消息响应预测。图 1 说明了问答检索的整体模型架构
## 2.1 Input Encoder
文档和查询输入作为独热编码向量传入。由$W_e\in\mathbb{R}^{d\times|V|}$参数化的单词嵌人层将每个单词转换为密集的单词表示$w\in\mathbb{R}^d$ $V$是词汇表中所有单词的集合
2.1.1 Highway Encoder.每个字嵌人都通过一个公路编码器层。公路网[31]是门控非线性变换层，控制信息流到后续层。许多作品采用投影层，该投影层经过训练以代替原始单
<FigureHere>
Figure 1: Illustration of our proposed Multi-Cast Attention Networks $( \textbf{ Best viewed in color }) .$ MCAN is a wide multiheaded attention architecture that utilizes compression functions and attention as features. Example is given for questionanswer retrieval. Input Encoding layer is ommitted for clarity.
词嵌人。这不仅节省了计算成本，而且减少了可训练参数的数量。我们的工作扩展了这个投影层以使用高速公路编码器。这样做的直觉很简单，我。e。, 高速公路编码器可以解释为数据驱动的字过滤器。因此，我们可以想象他们以参数方式学习哪些单词倾向于对手头的任务重要而不重要。例如，过滤停止词和通常对预测没有太大贡献的词。与本质上门控的递归模型类似，此高速公路编码器层控制(每个单词的)多少信息流向后续层
令$H(.)$和$T(.)$分别为具有 ReLU 和 sigmoid 激活函数的
单层仿射变换。单个公路网络层定义为：
$$y=H(x,W_H)\cdot T(x,W_T)+(1-T(x,W_T))\cdot x$$
(1)
其中$W_H,W_T\in\mathbb{R}^{r\times d}$ 。值得注意的是，仿射变换的维度可能与输人向量的大小不同。在这种情况下，使用附加的非线性变换将$x$投影到相同的维度
## 2.2 $\textbf{Co- Attention}$
共同关注[50]是一种成对关注机制，可以共同关注文本序列对。在本节中，我们介绍注意的四个变体，即。e。(1)最大池，(2)均值池，(3)对齐池，最后(4)内注意(或自我注) 意)。共同注意的第一步是学习两个序列中每个单词之间的亲和力(或相似性)矩阵。跟随 Parikh 等人。[20],我们采用以下公式来学习亲和矩阵。
$$s_{ij}=F(q_i)^TF(d_j)$$
(2)
其中$F(.)$是一个函数，例如多层感知器 (MLP)。其他形式的共同关注也是可能的，例如$s_ij=q_i^{\top}Md_j$和$s_ij=F([q_i;d_j])$。
|
[
[
0.0849609375,
0.10797119140625,
0.48046875,
0.10797119140625,
0.48046875,
0.121826171875,
0.0849609375,
0.121826171875
],
[
0.0882568359375,
0.12249755859375,
0.176513671875,
0.12249755859375,
0.176513671875,
0.1337890625,
0.0882568359375,
0.1337890625
],
[
0.708984375,
0.12249755859375,
0.72216796875,
0.12249755859375,
0.72216796875,
0.1319580078125,
0.708984375,
0.1319580078125
],
[
0.10296630859375,
0.1363525390625,
0.48046875,
0.1363525390625,
0.48046875,
0.149658203125,
0.10296630859375,
0.149658203125
],
[
0.65673828125,
0.1319580078125,
0.76708984375,
0.1331787109375,
0.76708984375,
0.14453125,
0.65673828125,
0.143310546875
],
[
0.08660888671875,
0.149658203125,
0.48046875,
0.149658203125,
0.48046875,
0.1634521484375,
0.08660888671875,
0.1634521484375
],
[
0.7099609375,
0.1458740234375,
0.72216796875,
0.1458740234375,
0.72216796875,
0.1546630859375,
0.7099609375,
0.1546630859375
],
[
0.69677734375,
0.1572265625,
0.73876953125,
0.1572265625,
0.73876953125,
0.1673583984375,
0.69677734375,
0.1673583984375
],
[
0.08660888671875,
0.1634521484375,
0.48046875,
0.1634521484375,
0.48046875,
0.1767578125,
0.08660888671875,
0.1767578125
],
[
0.59716796875,
0.17041015625,
0.60888671875,
0.17041015625,
0.60888671875,
0.179931640625,
0.59716796875,
0.179931640625
],
[
0.63232421875,
0.17041015625,
0.646484375,
0.17041015625,
0.646484375,
0.179931640625,
0.63232421875,
0.179931640625
],
[
0.68115234375,
0.17041015625,
0.69677734375,
0.17041015625,
0.69677734375,
0.179931640625,
0.68115234375,
0.179931640625
],
[
0.740234375,
0.17041015625,
0.7548828125,
0.17041015625,
0.7548828125,
0.179931640625,
0.740234375,
0.179931640625
],
[
0.78369140625,
0.17041015625,
0.79833984375,
0.17041015625,
0.79833984375,
0.179931640625,
0.78369140625,
0.179931640625
],
[
0.8349609375,
0.1717529296875,
0.8447265625,
0.1717529296875,
0.8447265625,
0.1787109375,
0.8349609375,
0.1787109375
],
[
0.0882568359375,
0.1787109375,
0.47802734375,
0.1787109375,
0.47802734375,
0.189453125,
0.0882568359375,
0.189453125
],
[
0.57666015625,
0.1787109375,
0.71337890625,
0.179931640625,
0.71337890625,
0.1912841796875,
0.57666015625,
0.1900634765625
],
[
0.7294921875,
0.1787109375,
0.86376953125,
0.179931640625,
0.86376953125,
0.1912841796875,
0.7294921875,
0.1900634765625
],
[
0.08660888671875,
0.1900634765625,
0.47802734375,
0.1900634765625,
0.47802734375,
0.203857421875,
0.08660888671875,
0.203857421875
],
[
0.54736328125,
0.1912841796875,
0.5537109375,
0.1912841796875,
0.5537109375,
0.1944580078125,
0.54736328125,
0.1944580078125
],
[
0.60302734375,
0.1962890625,
0.64208984375,
0.1962890625,
0.64208984375,
0.2171630859375,
0.60302734375,
0.2171630859375
],
[
0.78759765625,
0.1944580078125,
0.830078125,
0.1944580078125,
0.830078125,
0.2197265625,
0.78759765625,
0.2197265625
],
[
0.6806640625,
0.2020263671875,
0.6953125,
0.2020263671875,
0.6953125,
0.210205078125,
0.6806640625,
0.210205078125
],
[
0.74267578125,
0.20263671875,
0.74658203125,
0.20263671875,
0.74658203125,
0.210205078125,
0.74267578125,
0.210205078125
],
[
0.68310546875,
0.212158203125,
0.74609375,
0.210205078125,
0.74658203125,
0.2259521484375,
0.6845703125,
0.2279052734375
],
[
0.0882568359375,
0.2052001953125,
0.48046875,
0.2052001953125,
0.48046875,
0.2183837890625,
0.0882568359375,
0.2183837890625
],
[
0.0849609375,
0.2171630859375,
0.48046875,
0.2197265625,
0.48046875,
0.23291015625,
0.0849609375,
0.2310791015625
],
[
0.525390625,
0.2266845703125,
0.5908203125,
0.2266845703125,
0.5908203125,
0.238037109375,
0.525390625,
0.238037109375
],
[
0.68701171875,
0.2279052734375,
0.74609375,
0.2279052734375,
0.74609375,
0.2392578125,
0.68701171875,
0.2392578125
],
[
0.76123046875,
0.2235107421875,
0.90185546875,
0.2247314453125,
0.90185546875,
0.2392578125,
0.76123046875,
0.238037109375
],
[
0.08660888671875,
0.231689453125,
0.4794921875,
0.23291015625,
0.4794921875,
0.246826171875,
0.08660888671875,
0.24560546875
],
[
0.58837890625,
0.2291259765625,
0.66259765625,
0.2291259765625,
0.66259765625,
0.240478515625,
0.58837890625,
0.240478515625
],
[
0.54833984375,
0.2392578125,
0.56787109375,
0.2392578125,
0.56787109375,
0.251953125,
0.54833984375,
0.251953125
],
[
0.625,
0.2392578125,
0.6435546875,
0.2392578125,
0.6435546875,
0.251953125,
0.625,
0.251953125
],
[
0.70751953125,
0.2412109375,
0.72216796875,
0.2412109375,
0.72216796875,
0.250732421875,
0.70751953125,
0.250732421875
],
[
0.78173828125,
0.2412109375,
0.79833984375,
0.2412109375,
0.79833984375,
0.250732421875,
0.78173828125,
0.250732421875
],
[
0.86669921875,
0.236083984375,
0.87744140625,
0.2442626953125,
0.86669921875,
0.25244140625,
0.85546875,
0.2442626953125
],
[
0.0882568359375,
0.246826171875,
0.4794921875,
0.246826171875,
0.4794921875,
0.260009765625,
0.0882568359375,
0.260009765625
],
[
0.5458984375,
0.255126953125,
0.57177734375,
0.255126953125,
0.57177734375,
0.261962890625,
0.5458984375,
0.261962890625
],
[
0.62158203125,
0.255126953125,
0.64892578125,
0.255126953125,
0.64892578125,
0.261962890625,
0.62158203125,
0.261962890625
],
[
0.69921875,
0.255126953125,
0.7236328125,
0.255126953125,
0.7236328125,
0.261962890625,
0.69921875,
0.261962890625
],
[
0.78125,
0.255615234375,
0.8037109375,
0.255615234375,
0.8037109375,
0.260009765625,
0.78125,
0.260009765625
],
[
0.8505859375,
0.255126953125,
0.88134765625,
0.255126953125,
0.88134765625,
0.261962890625,
0.8505859375,
0.261962890625
],
[
0.6259765625,
0.26318359375,
0.6435546875,
0.26318359375,
0.6435546875,
0.276611328125,
0.6259765625,
0.276611328125
],
[
0.78173828125,
0.26318359375,
0.7998046875,
0.26318359375,
0.7998046875,
0.276611328125,
0.78173828125,
0.276611328125
],
[
0.708984375,
0.264404296875,
0.72216796875,
0.264404296875,
0.72216796875,
0.27587890625,
0.708984375,
0.27587890625
],
[
0.08660888671875,
0.260009765625,
0.4794921875,
0.260009765625,
0.4794921875,
0.2734375,
0.08660888671875,
0.2734375
],
[
0.55126953125,
0.266357421875,
0.564453125,
0.266357421875,
0.564453125,
0.27587890625,
0.55126953125,
0.27587890625
],
[
0.85888671875,
0.26513671875,
0.87158203125,
0.26513671875,
0.87158203125,
0.27587890625,
0.85888671875,
0.27587890625
],
[
0.0882568359375,
0.27587890625,
0.4794921875,
0.27587890625,
0.4794921875,
0.285888671875,
0.0882568359375,
0.285888671875
],
[
0.5400390625,
0.279052734375,
0.57177734375,
0.279052734375,
0.57177734375,
0.285888671875,
0.5400390625,
0.285888671875
],
[
0.70361328125,
0.279052734375,
0.7294921875,
0.279052734375,
0.7294921875,
0.287353515625,
0.70361328125,
0.287353515625
],
[
0.779296875,
0.276611328125,
0.8056640625,
0.279052734375,
0.8037109375,
0.287353515625,
0.77783203125,
0.285400390625
],
[
0.8505859375,
0.279052734375,
0.88134765625,
0.279052734375,
0.88134765625,
0.285888671875,
0.8505859375,
0.285888671875
],
[
0.08660888671875,
0.287353515625,
0.48046875,
0.28857421875,
0.48046875,
0.3017578125,
0.08660888671875,
0.300537109375
],
[
0.6337890625,
0.281005859375,
0.65185546875,
0.281005859375,
0.65185546875,
0.285400390625,
0.6337890625,
0.285400390625
],
[
0.7099609375,
0.290283203125,
0.72216796875,
0.290283203125,
0.72216796875,
0.300537109375,
0.7099609375,
0.300537109375
],
[
0.78515625,
0.290283203125,
0.7998046875,
0.290283203125,
0.7998046875,
0.300537109375,
0.78515625,
0.300537109375
],
[
0.625,
0.291748046875,
0.63818359375,
0.287353515625,
0.6435546875,
0.29736328125,
0.6298828125,
0.3017578125
],
[
0.5556640625,
0.292236328125,
0.56201171875,
0.292236328125,
0.56201171875,
0.29736328125,
0.5556640625,
0.29736328125
],
[
0.86279296875,
0.292236328125,
0.87109375,
0.292236328125,
0.87109375,
0.29736328125,
0.86279296875,
0.29736328125
],
[
0.08660888671875,
0.3017578125,
0.48046875,
0.3017578125,
0.48046875,
0.31494140625,
0.08660888671875,
0.31494140625
],
[
0.6005859375,
0.302978515625,
0.61376953125,
0.302978515625,
0.61376953125,
0.30615234375,
0.6005859375,
0.30615234375
],
[
0.61669921875,
0.3017578125,
0.63232421875,
0.3017578125,
0.63232421875,
0.309326171875,
0.61669921875,
0.309326171875
],
[
0.83642578125,
0.3037109375,
0.84326171875,
0.300537109375,
0.84619140625,
0.304931640625,
0.8388671875,
0.306884765625
],
[
0.0849609375,
0.31494140625,
0.4794921875,
0.31494140625,
0.4794921875,
0.328857421875,
0.0849609375,
0.328857421875
],
[
0.60546875,
0.32080078125,
0.615234375,
0.32080078125,
0.615234375,
0.32568359375,
0.60546875,
0.32568359375
],
[
0.08660888671875,
0.330810546875,
0.125,
0.330810546875,
0.125,
0.3447265625,
0.08660888671875,
0.3447265625
],
[
0.6396484375,
0.33203125,
0.646484375,
0.33203125,
0.646484375,
0.335205078125,
0.6396484375,
0.335205078125
],
[
0.62841796875,
0.341552734375,
0.65185546875,
0.341552734375,
0.65185546875,
0.349853515625,
0.62841796875,
0.349853515625
],
[
0.7861328125,
0.337158203125,
0.7939453125,
0.343505859375,
0.7890625,
0.349853515625,
0.779296875,
0.343505859375
],
[
0.85693359375,
0.34228515625,
0.865234375,
0.34228515625,
0.865234375,
0.3466796875,
0.85693359375,
0.3466796875
],
[
0.1298828125,
0.35595703125,
0.281005859375,
0.356689453125,
0.281005859375,
0.371826171875,
0.1298828125,
0.37060546875
],
[
0.5595703125,
0.34228515625,
0.5703125,
0.34228515625,
0.5703125,
0.34521484375,
0.5595703125,
0.34521484375
],
[
0.08660888671875,
0.35791015625,
0.13232421875,
0.35791015625,
0.13232421875,
0.369384765625,
0.08660888671875,
0.369384765625
],
[
0.08660888671875,
0.375732421875,
0.313720703125,
0.375732421875,
0.313720703125,
0.389404296875,
0.08660888671875,
0.389404296875
],
[
0.64794921875,
0.37890625,
0.68115234375,
0.380615234375,
0.6806640625,
0.392578125,
0.646484375,
0.39013671875
],
[
0.75732421875,
0.380615234375,
0.78125,
0.380615234375,
0.78125,
0.389404296875,
0.75732421875,
0.389404296875
],
[
0.111083984375,
0.393310546875,
0.48046875,
0.393310546875,
0.48046875,
0.4072265625,
0.111083984375,
0.4072265625
],
[
0.12335205078125,
0.4072265625,
0.48193359375,
0.40771484375,
0.48193359375,
0.421630859375,
0.12335205078125,
0.42041015625
],
[
0.5185546875,
0.412353515625,
0.91259765625,
0.412353515625,
0.91259765625,
0.425537109375,
0.5185546875,
0.425537109375
],
[
0.12335205078125,
0.421630859375,
0.48193359375,
0.421630859375,
0.48193359375,
0.43505859375,
0.12335205078125,
0.43505859375
],
[
0.51708984375,
0.425537109375,
0.91259765625,
0.425537109375,
0.91259765625,
0.439453125,
0.51708984375,
0.439453125
],
[
0.12335205078125,
0.43505859375,
0.48193359375,
0.434326171875,
0.48193359375,
0.447509765625,
0.12335205078125,
0.448974609375
],
[
0.5185546875,
0.44140625,
0.91259765625,
0.44140625,
0.91259765625,
0.451904296875,
0.5185546875,
0.451904296875
],
[
0.12335205078125,
0.448974609375,
0.336669921875,
0.449462890625,
0.336669921875,
0.46337890625,
0.12335205078125,
0.462158203125
],
[
0.5185546875,
0.453857421875,
0.9140625,
0.453857421875,
0.9140625,
0.46728515625,
0.5185546875,
0.46728515625
],
[
0.111083984375,
0.4609375,
0.48193359375,
0.462158203125,
0.48193359375,
0.4765625,
0.111083984375,
0.47607421875
],
[
0.5185546875,
0.4697265625,
0.9111328125,
0.4697265625,
0.9111328125,
0.479736328125,
0.5185546875,
0.479736328125
],
[
0.125,
0.4765625,
0.483642578125,
0.4765625,
0.483642578125,
0.490478515625,
0.125,
0.490478515625
],
[
0.125,
0.490478515625,
0.48193359375,
0.4892578125,
0.48193359375,
0.50390625,
0.125,
0.5048828125
],
[
0.1258544921875,
0.5048828125,
0.48046875,
0.5048828125,
0.48046875,
0.51806640625,
0.1258544921875,
0.51806640625
],
[
0.5185546875,
0.5087890625,
0.91259765625,
0.51025390625,
0.91259765625,
0.5234375,
0.5185546875,
0.52294921875
],
[
0.125,
0.517578125,
0.48046875,
0.517578125,
0.48046875,
0.53076171875,
0.125,
0.53076171875
],
[
0.51953125,
0.5234375,
0.91259765625,
0.5234375,
0.91259765625,
0.537109375,
0.51953125,
0.537109375
],
[
0.125,
0.5322265625,
0.48193359375,
0.5322265625,
0.48193359375,
0.54541015625,
0.125,
0.54541015625
],
[
0.51708984375,
0.5361328125,
0.91259765625,
0.537109375,
0.91259765625,
0.55029296875,
0.51708984375,
0.5498046875
],
[
0.125,
0.54541015625,
0.48193359375,
0.54541015625,
0.48193359375,
0.5595703125,
0.125,
0.5595703125
],
[
0.51953125,
0.55322265625,
0.91259765625,
0.55322265625,
0.91259765625,
0.56298828125,
0.51953125,
0.56298828125
],
[
0.1258544921875,
0.56103515625,
0.48046875,
0.56103515625,
0.48046875,
0.5712890625,
0.1258544921875,
0.5712890625
],
[
0.51708984375,
0.56298828125,
0.91259765625,
0.564453125,
0.91259765625,
0.5791015625,
0.51708984375,
0.57763671875
],
[
0.1258544921875,
0.57470703125,
0.4794921875,
0.57470703125,
0.4794921875,
0.58544921875,
0.1258544921875,
0.58544921875
],
[
0.5185546875,
0.5791015625,
0.91259765625,
0.5791015625,
0.91259765625,
0.59228515625,
0.5185546875,
0.59228515625
],
[
0.1258544921875,
0.5869140625,
0.181396484375,
0.5888671875,
0.181396484375,
0.60107421875,
0.1258544921875,
0.5986328125
],
[
0.5185546875,
0.59228515625,
0.9140625,
0.59228515625,
0.9140625,
0.60595703125,
0.5185546875,
0.60595703125
],
[
0.111083984375,
0.60107421875,
0.483642578125,
0.60107421875,
0.483642578125,
0.6142578125,
0.111083984375,
0.6142578125
],
[
0.51953125,
0.60791015625,
0.564453125,
0.60791015625,
0.564453125,
0.619140625,
0.51953125,
0.619140625
],
[
0.1258544921875,
0.61572265625,
0.48046875,
0.61572265625,
0.48046875,
0.62890625,
0.1258544921875,
0.62890625
],
[
0.53515625,
0.62060546875,
0.91259765625,
0.62060546875,
0.91259765625,
0.6337890625,
0.53515625,
0.6337890625
],
[
0.1258544921875,
0.62890625,
0.483642578125,
0.62890625,
0.483642578125,
0.642578125,
0.1258544921875,
0.642578125
],
[
0.51953125,
0.6337890625,
0.775390625,
0.6337890625,
0.775390625,
0.64794921875,
0.51953125,
0.64794921875
],
[
0.1258544921875,
0.64453125,
0.48046875,
0.64453125,
0.48046875,
0.65478515625,
0.1258544921875,
0.65478515625
],
[
0.125,
0.6572265625,
0.48193359375,
0.6572265625,
0.48193359375,
0.67041015625,
0.125,
0.67041015625
],
[
0.580078125,
0.65576171875,
0.85107421875,
0.65576171875,
0.85107421875,
0.66943359375,
0.580078125,
0.66943359375
],
[
0.89208984375,
0.65771484375,
0.91259765625,
0.65771484375,
0.91259765625,
0.67041015625,
0.89208984375,
0.67041015625
],
[
0.1258544921875,
0.67236328125,
0.410888671875,
0.67236328125,
0.410888671875,
0.68310546875,
0.1258544921875,
0.68310546875
],
[
0.5185546875,
0.67822265625,
0.9140625,
0.67822265625,
0.9140625,
0.69580078125,
0.5185546875,
0.69580078125
],
[
0.11517333984375,
0.69775390625,
0.461669921875,
0.69775390625,
0.461669921875,
0.7109375,
0.11517333984375,
0.7109375
],
[
0.5185546875,
0.69580078125,
0.91259765625,
0.69580078125,
0.91259765625,
0.708984375,
0.5185546875,
0.708984375
],
[
0.5185546875,
0.71044921875,
0.6689453125,
0.71044921875,
0.6689453125,
0.72021484375,
0.5185546875,
0.72021484375
],
[
0.08660888671875,
0.70068359375,
0.11517333984375,
0.70068359375,
0.11517333984375,
0.708984375,
0.08660888671875,
0.708984375
],
[
0.08660888671875,
0.71533203125,
0.48046875,
0.71533203125,
0.48046875,
0.72900390625,
0.08660888671875,
0.72900390625
],
[
0.08660888671875,
0.72998046875,
0.48193359375,
0.72998046875,
0.48193359375,
0.74365234375,
0.08660888671875,
0.74365234375
],
[
0.51953125,
0.73779296875,
0.54345703125,
0.73779296875,
0.54345703125,
0.7470703125,
0.51953125,
0.7470703125
],
[
0.54150390625,
0.7392578125,
0.5458984375,
0.7392578125,
0.5458984375,
0.744140625,
0.54150390625,
0.744140625
],
[
0.56396484375,
0.7373046875,
0.66845703125,
0.7373046875,
0.66845703125,
0.74755859375,
0.56396484375,
0.74755859375
],
[
0.08660888671875,
0.74365234375,
0.48193359375,
0.74365234375,
0.48193359375,
0.7568359375,
0.08660888671875,
0.7568359375
],
[
0.08660888671875,
0.75830078125,
0.45751953125,
0.75830078125,
0.45751953125,
0.771484375,
0.08660888671875,
0.771484375
],
[
0.521484375,
0.755859375,
0.9111328125,
0.755859375,
0.9111328125,
0.7666015625,
0.521484375,
0.7666015625
],
[
0.5185546875,
0.76806640625,
0.9140625,
0.76806640625,
0.9140625,
0.7822265625,
0.5185546875,
0.7822265625
],
[
0.08416748046875,
0.78515625,
0.2484130859375,
0.78515625,
0.2484130859375,
0.798828125,
0.08416748046875,
0.798828125
],
[
0.5185546875,
0.78076171875,
0.91259765625,
0.78076171875,
0.91259765625,
0.79443359375,
0.5185546875,
0.79443359375
],
[
0.5185546875,
0.796875,
0.91259765625,
0.796875,
0.91259765625,
0.81005859375,
0.5185546875,
0.81005859375
],
[
0.08660888671875,
0.80322265625,
0.48193359375,
0.7998046875,
0.48193359375,
0.8173828125,
0.08660888671875,
0.82080078125
],
[
0.5185546875,
0.81005859375,
0.91259765625,
0.81005859375,
0.91259765625,
0.82373046875,
0.5185546875,
0.82373046875
],
[
0.0849609375,
0.8173828125,
0.48193359375,
0.81640625,
0.48193359375,
0.830078125,
0.0849609375,
0.83056640625
],
[
0.5185546875,
0.82275390625,
0.689453125,
0.82373046875,
0.689453125,
0.8369140625,
0.5185546875,
0.8359375
],
[
0.0849609375,
0.83203125,
0.279296875,
0.83203125,
0.279296875,
0.84521484375,
0.0849609375,
0.84521484375
],
[
0.66015625,
0.84521484375,
0.76953125,
0.84521484375,
0.76953125,
0.85986328125,
0.66015625,
0.85986328125
],
[
0.89208984375,
0.84521484375,
0.9140625,
0.84521484375,
0.9140625,
0.85986328125,
0.89208984375,
0.85986328125
],
[
0.10455322265625,
0.85400390625,
0.48046875,
0.85400390625,
0.48046875,
0.8671875,
0.10455322265625,
0.8671875
],
[
0.08660888671875,
0.86865234375,
0.48046875,
0.86865234375,
0.48046875,
0.8818359375,
0.08660888671875,
0.8818359375
],
[
0.51708984375,
0.8662109375,
0.91259765625,
0.8671875,
0.91259765625,
0.8818359375,
0.51708984375,
0.880859375
],
[
0.08660888671875,
0.88330078125,
0.4794921875,
0.88330078125,
0.4794921875,
0.89306640625,
0.08660888671875,
0.89306640625
],
[
0.515625,
0.880859375,
0.89892578125,
0.8818359375,
0.89892578125,
0.896484375,
0.515625,
0.89599609375
],
[
0.89453125,
0.88720703125,
0.90283203125,
0.88720703125,
0.90283203125,
0.892578125,
0.89453125,
0.892578125
]
] |
[
"e。，执行体系结构工程以找到在每个问题域和数据集上效果最",
"好的最佳变体",
"↑",
"我们的方法旨在服务于两个重要目的-（1）通过使注意力能",
"MeanMax Pooling + Highway Laser",
"够被调用任意k次而几乎没有任何后果，从而消除了对该组",
"合",
"L.STM Enodes",
"件的体系结构工程的需求，并且（2）同时通过对多个视图进",
"合",
"合",
"合",
"金",
"金",
"食",
"行建模来提高性能通过多次注意调用。因此，我们的方法与多",
"coocoocooo.coo..c.coo",
"aoooaqq.coo.ooq..coq.ooo",
"头注意力具有相似的精神，尽管效率很高。。为此，我们引人",
"",
"Soupste",
"Ssamne",
"M",
"",
"Altention Features",
"了Multi-CastAttentionNetworks(MCAN)，这是一种新的深",
"度学习架构，用于问答和对话建模领域的任务。在我们的方法",
"1000000",
"000000",
"0o可o",
"中，注意力被投射，与大多数其他将其用作汇集操作的作品相",
"000000",
"+",
"+",
"金",
"合",
"i",
"反。我们多次投射共同注意力，每次返回一个压缩的标量特",
"Compressig",
"Compressio",
"Compreso",
"ompressi",
"Compressio",
"+",
"+",
"↑",
"征，该特征重新附加到原始单词表示。关键的直觉是，压缩可",
"合",
"",
"以实现多个注意力调用的可扩展投射，旨在为后续层提供不",
"Intra-Attent",
"Max Pool",
"Mcan Fo",
"xea-Aenti",
"仅具有全局知识而且具有跨句子知识的提示。直观地，当将这",
"ntPoo",
"+",
"+",
"4",
"",
"",
"些增强的嵌入传递到组成编码器（例如长短期记忆编码器）中",
"",
"",
"",
"时，LSTM可以从此提示中受益，并相应地改变其表示学习",
"",
"过程。",
"",
"",
"",
"",
"Our Contributions",
".",
"1.1",
"总而言之，这项工作的主要贡献是：",
"Question",
"Answe",
". For the first time, we propose a new paradigm of utiliz.",
"ing attentions not as a pooling operator but as a form",
"Figure 1: Illustration of our proposed Multi-Cast Attention",
"of feature augmentation. We propose an overall archi-",
"Networks ( Best viewed in color ). MCAN is a wide multi-",
"tecture, Multi-Cast Attention Networks (MCAN)for",
"headed attention architecture that utilizes compressionfunc-",
"generic sequence pair modeling.",
"tions and attention as features. Example is given for question-",
". We evaluate our proposed model on four benchmark",
"answer retrieval.Input Encodinglayer is ommitted for clarity.",
"tasks, i.e., Dialogue Reply Prediction (Ubuntu dialogue",
"corpus), Factoid Question Answering (TrecQA), Com-",
"munity Question Answering (QatarLiving forums from",
"词嵌入。这不仅节省了计算成本，而且减少了可训练参数的数",
"SemEval 2016) and Tweet Reply Prediction (Customer",
"量。我们的工作扩展了这个投影层以使用高速公路编码器。这",
"support). On Ubuntu dialogue corpus, MCAN out-",
"样做的直觉很简单，我。e。，高速公路编码器可以解释为数据",
"performs the existing state-of-the-art models by 9%",
"驱动的字过滤器。因此，我们可以想象他们以参数方式学习哪",
".MCAN also achieves the best performing score of",
"些单词倾向于对手头的任务重要而不重要。例如，过滤停止词",
"0.838MAPand0.904MRRonthewell-studiedTrecQA",
"和通常对预测没有太大贡献的词。与本质上门控的递归模型",
"dataset.",
"类似，此高速公路编码器层控制（每个单词的）多少信息流向",
". We provide a comprehensive and in-depth analysis of",
"后续层",
"the inner workings of our proposed MCAN model.",
"令H(）和T(）分别为具有ReLU和sigmoid激活函数的",
"We show that the casted attention features are in-",
"单层仿射变换。单个公路网络层定义为：",
"terpretable and are capable of learning(l)a neural",
"adaptation of word overlap and (2) a differentiation of",
"y = H(x, WH) -T(x, WT) + (1 - T(x, WT)) x",
"(1)",
"evidence and anti-evidence words/patterns",
"其中WH,WrERrxd。值得注意的是，仿射变换的维度可能与",
"MULTI-CAST ATTENTION NETWORKS",
"输人向量的大小不同。在这种情况下，使用附加的非线性变换",
"将x投影到相同的维度",
"2",
"在本节中，我们将描述我们提出的MCAN模型。我们模型的",
"输人是两个文本序列，我们将其表示为查询q和文档d。在",
"2.2",
"",
"Co-Attention",
"我们的问题中，查询文档可以推广到不同的问题领域，例如问",
"答或消息响应预测。图1说明了问答检索的整体模型架构",
"共同关注[50]是一种成对关注机制，可以共同关注文本序列",
"对。在本节中，我们介绍注意的四个变体，即。e。，（1）最大",
"2.1Input Encoder",
"池，（2）均值池，（3）对齐池，最后（4）内注意（或自我注",
"意）。共同注意的第一步是学习两个序列中每个单词之间的亲",
"文档和查询输入作为独热编码向量传入。由WeeRdxIVI参数",
"和力（或相似性）矩阵。跟随Parikh等人。[20]，我们采用",
"化的单词嵌入层将每个单词转换为密集的单词表示weRd。",
"以下公式来学习亲和矩阵。",
"V是词汇表中所有单词的集合",
"sij = F(qi)TF(d,)",
"(2)",
"2.1.1HighwayEncoder。每个字嵌人都通过一个公路编",
"码器层。公路网[31]是门控非线性变换层，控制信息流到后",
"其中F(）是一个函数，例如多层感知器（MLP)。其他形式的",
"续层。许多作品采用投影层，该投影层经过训练以代替原始单",
"共同关注也是可能的，例如sij=qTMd,和sij=F([qi;djl)",
""
] |
{
"arxiv_id": "1806.00778",
"language": "zh",
"page_idx": 1
}
|
||
1806.00778-zh-2
|
1806.00778-zh-2.jpg
|
2.2.1 Extractive Pooling.提取共同注意力的最常见变体
是最大池共同注意力，它根据每个单词对其他文本序列的最
大影响来关注每个单词。
(3
$$q'=Soft(\max_{col}(s))^{\top}q\:\mathrm{and}\:d'=Soft(\max_{row}(s))^{\top}d$$
,其中$q^\prime,d^{\prime}$分别是$q$和$d$的共同注意表示。软(。)是 Softmax
运算符。或者，也可以使用矩阵 s 的平均行和列池化：
(4)
$$q'=Soft(mean(s))^\top q\:\mathrm{and}\:d'=Soft(mean(s))^\top d$$
但是，每个池化运算符都有不同的影响，可以直观地理解为： 最大池化根据每个单词在其他文本中所有单词的最大重要性来选择每个单词。均值池是一种更健康的比较，根据单词对其他文本的整体影响来关注单词。这通常与数据集相关，被视为超参数，并进行调整以查看哪个在保留的集合上表现最佳
2.2.2 Alignment-Pooling.基于软对齐的池化也已用于学习共同注意表示[20] 。然而，软对齐的关键区别在于它重新排列了序列对，而标准的共同注意只是学会了对重要单词进行加权和评分。然后按如下方式学习共同注意表示：
$$d_i':=\sum_{j=1}^{\ell_q}\frac{exp(s_{ij})}{\sum_{k=1}^{\ell_q}exp(s_{ik})}q_j\:\text{and}\:q_j':=\sum_{i=1}^{\ell_d}\frac{exp(s_{ij})}{\sum_{k=1}^{\ell_d}exp(s_{kj})}d_i$$
其中$d_i^\prime$ 是$q$ 中与 $d_i$ 柔和对齐的子短语。直观上$,d_i^\prime$ 是$\left\{q_j\right\}_{j=1}^{\ell_q}$
的加权和，选择$q$中最相关的部分来表示$d_i$
最近提出了
2.2.3 $Intra- Attention.$ Intra-Attention 或 Self-Attention 来学习意识到长期依赖关系的表示形式。这通常被表述为相对于自身的共同关注 (或对齐) 操作。在这种情况下，我们将内部关注独立地应用于文档和查询。为了简单起见，我们将它们称为$x$,而不是$q$或$d$ 。注意内函数定义为：
$$x_i':=\sum_{j=1}^\ell\frac{exp(s_{ij})}{\sum_{k=1}^\ell exp(s_{ik})}x_j$$
(6)
其中$x_i^{\prime}$是$x_j$的注意内表示
# 2.3 Multi-Cast Attention
在这一点上，很容易进行几个观察。首先，每种注意力机制都为模型提供了不同的风味。其次，注意力被用来通过重新加权或重新对齐来改变原始表示。因此，大多数神经架构仅使用一种类型的共同注意或对齐函数[5,20] 。但是，这需要调整正确的模型体系结构，并且可能会错过使用多种共同注意机制带来的好处。因此，我们的工作将每个注意力操作都转换为单词级功能
2.3.1 Casted Attention.令$x$为$q$或$d$,而$\bar{x}$是应用共同关注或软关注对齐后$x\text{ 的表示 }^{1}$。共注意运算符的注意特征是：
(7)
(8)
$$f_{c}=F_{c}([\bar{x};x])\\f_{m}=F_{c}(\bar{x}\odot x)\\f_{s}=F_{c}(\bar{x}-x)$$
(9)
其中$\odot$是 Hadamard 乘积，[.;.] 是串联运算符。$F_{\mathrm{c}}(.)$ 是一个压缩函数，用于将特征缩减为标量。直观地说，这里实现的
${\overline{{^{1}\text{We omit sub}}}}$scripts for clarity.
是，我们通过比较共同注意之前和之后的表示来对共同注意的影响进行建模。对于软注意力对齐，这里需要注意的是$x$和$\bar{x}$ (尽管长度相等)具有“交换”语义。换句话说，在 $q$ 的情况下，$\bar{q}$实际上包含$d$的对齐表示。最后，使用多个比较运算符 (减法，串联和乘法运算符)是为了捕获多个视角，并受到ESIM 模型[3]的启发
2.3.2 Compression Function.本节定义$F_c(.)$所使用的压
缩函数。压缩的基本原理简单直观-我们不想用高维向量膨胀后续层，从而导致后续层中的参数成本。我们研究了三个压缩函数的使用，它们能够将$n$维向量简化为标量。
$\bullet$ Sum (SM) Function is a non-parameterized function
that sums the entire vector, returning a scalar as an
output.
$$F(x)=\sum_i^nx_i\:,\:\forall x_i\in x$$
(10)
$\bullet$ Neural Network (NN) is a fully-connected layer that converts each $n$ dimensional feature vector as follows:
$$F(x)=ReLU(W_c(x)+b_c).$$
(11)
where $W_c\times\mathbb{R}^{n\times1}$ and $b_c\in\mathbb{R}$ are the parameters of the
FC layer.
$\bullet$ Factorization Machines (FM) are general purpose ma-
chine learning techniques that accept a real-valued fea-
ture vector $x\in\mathbb{R}^n$ and return a scalar output.
(12)
$$F(x)=w_0+\sum_{i=1}^nw_i\:x_i+\sum_{i=1}^n\sum_{j=i+1}^n\langle v_i,v_j\rangle\:x_i\:x_j$$
where $w_0\in\mathbb{R},w_i\in\mathbb{R}^n$ and $\{v_1,\cdots v_n\}\in\mathbb{R}^n\times k$ are the parameters of the FM model. FMs are expressive models that capture pairwise interactions between features using factorized parameters. $k$ is the number of factors of the FM model. For more details, we refer interested readers to [24] .
请注意，我们不会在多个注意力投射之间共享参数，因为每个注意力投射都旨在对不同的视图进行建模。我们的实验分别在模型名称 MCAN (SM), MCAN (NN) 和 MCAN (FM) 下报告了上述变体
2.3.3 Multi-Cast .我们的体系结构背后的关键思想是促进$k$ 注意调用 (或强制转换),每个强制转换都用一个真正有价值的注意提示来增强原始单词嵌人。我们正式描述了多播注意力机制。对于每个查询文档对，我们应用(1)具有均值池的共同注意力(2)具有最大池的共同注意力和(3)具有对齐池的共同注意力。此外，我们将 Intra-Attention 分别应用于查询和文档。每个注意力投射都会产生三个标量(每个单词), 这些标量与单词嵌人连接在一起。最终铸造的特征向量是$z\in\mathbb{R}^{12}$ 。因此，对于每个单词$w_i$ ,新单词表示形式变为$\bar{w}_{i}=[w_{i};z_{i}]$
2.4 Long Short-Term Memory Encoder
接下来，将引起注意的单词表示$\bar{w}_1,\bar{w}_2,\ldots\bar{w}_\ell$传递到顺序编码
器层。我们采用标准的香草长短期记忆(LSTM)编码器：
(13)
$$h_i=LSTM(u,i),\forall i\in[1,\ldots\ell]$$
|
[
[
0.10455322265625,
0.10919189453125,
0.1544189453125,
0.10919189453125,
0.1544189453125,
0.11932373046875,
0.10455322265625,
0.11932373046875
],
[
0.1544189453125,
0.10919189453125,
0.47802734375,
0.10919189453125,
0.47802734375,
0.11932373046875,
0.1544189453125,
0.11932373046875
],
[
0.51953125,
0.10919189453125,
0.9111328125,
0.10919189453125,
0.9111328125,
0.11932373046875,
0.51953125,
0.11932373046875
],
[
0.08660888671875,
0.121826171875,
0.48193359375,
0.121826171875,
0.48193359375,
0.1351318359375,
0.08660888671875,
0.1351318359375
],
[
0.5185546875,
0.121826171875,
0.9140625,
0.121826171875,
0.9140625,
0.1351318359375,
0.5185546875,
0.1351318359375
],
[
0.0882568359375,
0.136962890625,
0.240234375,
0.136962890625,
0.240234375,
0.147705078125,
0.0882568359375,
0.147705078125
],
[
0.5185546875,
0.136962890625,
0.9111328125,
0.136962890625,
0.9111328125,
0.147705078125,
0.5185546875,
0.147705078125
],
[
0.131591796875,
0.154052734375,
0.439453125,
0.154052734375,
0.439453125,
0.1673583984375,
0.131591796875,
0.1673583984375
],
[
0.5185546875,
0.149658203125,
0.91259765625,
0.149658203125,
0.91259765625,
0.1634521484375,
0.5185546875,
0.1634521484375
],
[
0.438720703125,
0.160400390625,
0.455810546875,
0.160400390625,
0.455810546875,
0.1641845703125,
0.438720703125,
0.1641845703125
],
[
0.4599609375,
0.1578369140625,
0.47802734375,
0.1578369140625,
0.47802734375,
0.1666259765625,
0.4599609375,
0.1666259765625
],
[
0.2017822265625,
0.1641845703125,
0.2197265625,
0.169189453125,
0.2156982421875,
0.1767578125,
0.19775390625,
0.17236328125
],
[
0.5185546875,
0.1634521484375,
0.91259765625,
0.1634521484375,
0.91259765625,
0.1767578125,
0.5185546875,
0.1767578125
],
[
0.08905029296875,
0.179931640625,
0.48193359375,
0.18115234375,
0.48193359375,
0.1944580078125,
0.08905029296875,
0.1932373046875
],
[
0.51708984375,
0.175537109375,
0.6611328125,
0.1767578125,
0.6611328125,
0.1900634765625,
0.51708984375,
0.189453125
],
[
0.08660888671875,
0.1944580078125,
0.4306640625,
0.1944580078125,
0.4306640625,
0.2083740234375,
0.08660888671875,
0.2083740234375
],
[
0.5361328125,
0.199462890625,
0.5888671875,
0.199462890625,
0.5888671875,
0.2108154296875,
0.5361328125,
0.2108154296875
],
[
0.5810546875,
0.1988525390625,
0.9140625,
0.199462890625,
0.9140625,
0.21337890625,
0.5810546875,
0.212158203125
],
[
0.1258544921875,
0.21337890625,
0.44775390625,
0.21337890625,
0.44775390625,
0.2279052734375,
0.1258544921875,
0.2279052734375
],
[
0.441162109375,
0.21533203125,
0.48046875,
0.21533203125,
0.48046875,
0.2266845703125,
0.441162109375,
0.2266845703125
],
[
0.5185546875,
0.2139892578125,
0.91259765625,
0.2139892578125,
0.91259765625,
0.2247314453125,
0.5185546875,
0.2247314453125
],
[
0.19775390625,
0.2266845703125,
0.214111328125,
0.2266845703125,
0.214111328125,
0.2342529296875,
0.19775390625,
0.2342529296875
],
[
0.51953125,
0.2266845703125,
0.91259765625,
0.2266845703125,
0.91259765625,
0.240478515625,
0.51953125,
0.240478515625
],
[
0.08660888671875,
0.2392578125,
0.48193359375,
0.2392578125,
0.48193359375,
0.25244140625,
0.08660888671875,
0.25244140625
],
[
0.5185546875,
0.240478515625,
0.82177734375,
0.240478515625,
0.82177734375,
0.25390625,
0.5185546875,
0.25390625
],
[
0.0849609375,
0.25244140625,
0.48046875,
0.25390625,
0.48046875,
0.267578125,
0.0849609375,
0.266357421875
],
[
0.54345703125,
0.25830078125,
0.91259765625,
0.25830078125,
0.91259765625,
0.271484375,
0.54345703125,
0.271484375
],
[
0.08660888671875,
0.267578125,
0.48046875,
0.267578125,
0.48046875,
0.281005859375,
0.08660888671875,
0.281005859375
],
[
0.5556640625,
0.270751953125,
0.9140625,
0.271484375,
0.9140625,
0.285888671875,
0.5556640625,
0.285400390625
],
[
0.08660888671875,
0.281005859375,
0.48193359375,
0.281005859375,
0.48193359375,
0.294189453125,
0.08660888671875,
0.294189453125
],
[
0.55810546875,
0.287353515625,
0.60693359375,
0.287353515625,
0.60693359375,
0.298583984375,
0.55810546875,
0.298583984375
],
[
0.0882568359375,
0.296630859375,
0.455078125,
0.296630859375,
0.455078125,
0.306884765625,
0.0882568359375,
0.306884765625
],
[
0.69921875,
0.300537109375,
0.7099609375,
0.309326171875,
0.69921875,
0.317626953125,
0.68896484375,
0.309326171875
],
[
0.101318359375,
0.314453125,
0.48046875,
0.31494140625,
0.48046875,
0.330078125,
0.101318359375,
0.328857421875
],
[
0.64208984375,
0.313232421875,
0.7890625,
0.31494140625,
0.7890625,
0.330078125,
0.64208984375,
0.32763671875
],
[
0.8837890625,
0.31494140625,
0.91259765625,
0.31494140625,
0.91259765625,
0.32763671875,
0.8837890625,
0.32763671875
],
[
0.08660888671875,
0.330078125,
0.48046875,
0.330078125,
0.48046875,
0.343505859375,
0.08660888671875,
0.343505859375
],
[
0.69921875,
0.324462890625,
0.71142578125,
0.333984375,
0.69921875,
0.343505859375,
0.68701171875,
0.333984375
],
[
0.0882568359375,
0.34521484375,
0.4794921875,
0.34521484375,
0.4794921875,
0.35595703125,
0.0882568359375,
0.35595703125
],
[
0.54345703125,
0.3466796875,
0.91259765625,
0.3466796875,
0.91259765625,
0.35986328125,
0.54345703125,
0.35986328125
],
[
0.0849609375,
0.356689453125,
0.418212890625,
0.35791015625,
0.418212890625,
0.371826171875,
0.0849609375,
0.37060546875
],
[
0.5556640625,
0.359130859375,
0.9111328125,
0.35986328125,
0.9111328125,
0.373779296875,
0.5556640625,
0.37255859375
],
[
0.1298828125,
0.376953125,
0.1478271484375,
0.376953125,
0.1478271484375,
0.39013671875,
0.1298828125,
0.39013671875
],
[
0.1544189453125,
0.37890625,
0.221435546875,
0.3818359375,
0.2197265625,
0.396484375,
0.15283203125,
0.393310546875
],
[
0.33251953125,
0.375732421875,
0.35302734375,
0.380615234375,
0.34716796875,
0.393310546875,
0.326904296875,
0.38818359375
],
[
0.34716796875,
0.37890625,
0.423095703125,
0.3818359375,
0.42236328125,
0.396484375,
0.346435546875,
0.393310546875
],
[
0.6337890625,
0.380615234375,
0.7939453125,
0.380615234375,
0.7939453125,
0.39453125,
0.6337890625,
0.39453125
],
[
0.8837890625,
0.380615234375,
0.9140625,
0.380615234375,
0.9140625,
0.393310546875,
0.8837890625,
0.393310546875
],
[
0.0931396484375,
0.386962890625,
0.134033203125,
0.389404296875,
0.13232421875,
0.40478515625,
0.09234619140625,
0.40283203125
],
[
0.12744140625,
0.38818359375,
0.181396484375,
0.3984375,
0.176513671875,
0.41357421875,
0.1217041015625,
0.40283203125
],
[
0.2344970703125,
0.39013671875,
0.348876953125,
0.39013671875,
0.348876953125,
0.40478515625,
0.2344970703125,
0.40478515625
],
[
0.439453125,
0.389404296875,
0.48046875,
0.389404296875,
0.48046875,
0.404052734375,
0.439453125,
0.404052734375
],
[
0.12744140625,
0.399658203125,
0.188720703125,
0.404052734375,
0.1871337890625,
0.419189453125,
0.1258544921875,
0.415283203125
],
[
0.324462890625,
0.395751953125,
0.443603515625,
0.392578125,
0.4453125,
0.4140625,
0.325927734375,
0.417236328125
],
[
0.17724609375,
0.40283203125,
0.1846923828125,
0.40283203125,
0.1846923828125,
0.4072265625,
0.17724609375,
0.4072265625
],
[
0.1805419921875,
0.40283203125,
0.238525390625,
0.40283203125,
0.238525390625,
0.4140625,
0.1805419921875,
0.4140625
],
[
0.55810546875,
0.40283203125,
0.91259765625,
0.40283203125,
0.91259765625,
0.416748046875,
0.55810546875,
0.416748046875
],
[
0.08416748046875,
0.421630859375,
0.473876953125,
0.423583984375,
0.473876953125,
0.44140625,
0.08416748046875,
0.439453125
],
[
0.55615234375,
0.416748046875,
0.619140625,
0.41845703125,
0.619140625,
0.43115234375,
0.55615234375,
0.4287109375
],
[
0.54345703125,
0.43115234375,
0.9140625,
0.43115234375,
0.9140625,
0.4443359375,
0.54345703125,
0.4443359375
],
[
0.0882568359375,
0.443115234375,
0.369384765625,
0.443115234375,
0.369384765625,
0.453857421875,
0.0882568359375,
0.453857421875
],
[
0.55810546875,
0.44580078125,
0.9140625,
0.44580078125,
0.9140625,
0.458984375,
0.55810546875,
0.458984375
],
[
0.10455322265625,
0.45703125,
0.176513671875,
0.45703125,
0.176513671875,
0.468505859375,
0.10455322265625,
0.468505859375
],
[
0.55615234375,
0.458251953125,
0.86279296875,
0.458984375,
0.86279296875,
0.472900390625,
0.55615234375,
0.4716796875
],
[
0.10296630859375,
0.4765625,
0.48046875,
0.4765625,
0.48046875,
0.490478515625,
0.10296630859375,
0.490478515625
],
[
0.66015625,
0.479248046875,
0.67724609375,
0.479248046875,
0.67724609375,
0.492431640625,
0.66015625,
0.492431640625
],
[
0.72216796875,
0.4765625,
0.77294921875,
0.4765625,
0.77294921875,
0.50048828125,
0.72216796875,
0.50048828125
],
[
0.576171875,
0.48486328125,
0.74658203125,
0.48486328125,
0.74658203125,
0.505859375,
0.576171875,
0.505859375
],
[
0.7646484375,
0.4873046875,
0.8505859375,
0.4892578125,
0.8505859375,
0.50390625,
0.7646484375,
0.501953125
],
[
0.8837890625,
0.4892578125,
0.91259765625,
0.4892578125,
0.91259765625,
0.501953125,
0.8837890625,
0.501953125
],
[
0.74853515625,
0.496826171875,
0.765625,
0.496826171875,
0.765625,
0.50048828125,
0.74853515625,
0.50048828125
],
[
0.7255859375,
0.50244140625,
0.77685546875,
0.50244140625,
0.77685546875,
0.5146484375,
0.7255859375,
0.5146484375
],
[
0.65869140625,
0.50390625,
0.6787109375,
0.50390625,
0.6787109375,
0.5146484375,
0.65869140625,
0.5146484375
],
[
0.0882568359375,
0.492431640625,
0.4794921875,
0.492431640625,
0.4794921875,
0.50244140625,
0.0882568359375,
0.50244140625
],
[
0.0882568359375,
0.505859375,
0.48046875,
0.505859375,
0.48046875,
0.5166015625,
0.0882568359375,
0.5166015625
],
[
0.0882568359375,
0.51953125,
0.4794921875,
0.51953125,
0.4794921875,
0.52978515625,
0.0882568359375,
0.52978515625
],
[
0.5537109375,
0.521484375,
0.9140625,
0.51953125,
0.9140625,
0.537109375,
0.5537109375,
0.5390625
],
[
0.0849609375,
0.53076171875,
0.39453125,
0.53076171875,
0.39453125,
0.54443359375,
0.0849609375,
0.54443359375
],
[
0.55810546875,
0.537109375,
0.9140625,
0.537109375,
0.9140625,
0.55029296875,
0.55810546875,
0.55029296875
],
[
0.253173828125,
0.55322265625,
0.330810546875,
0.5546875,
0.330810546875,
0.5693359375,
0.253173828125,
0.5673828125
],
[
0.5595703125,
0.5517578125,
0.9111328125,
0.5517578125,
0.9111328125,
0.5625,
0.5595703125,
0.5625
],
[
0.1993408203125,
0.56298828125,
0.256591796875,
0.556640625,
0.2607421875,
0.57568359375,
0.2034912109375,
0.58203125
],
[
0.2442626953125,
0.5537109375,
0.253173828125,
0.5537109375,
0.253173828125,
0.55810546875,
0.2442626953125,
0.55810546875
],
[
0.4599609375,
0.56298828125,
0.48193359375,
0.56298828125,
0.48193359375,
0.5771484375,
0.4599609375,
0.5771484375
],
[
0.55810546875,
0.56494140625,
0.91259765625,
0.56494140625,
0.91259765625,
0.5791015625,
0.55810546875,
0.5791015625
],
[
0.348876953125,
0.568359375,
0.361083984375,
0.568359375,
0.361083984375,
0.5771484375,
0.348876953125,
0.5771484375
],
[
0.2344970703125,
0.57080078125,
0.3544921875,
0.568359375,
0.3544921875,
0.5869140625,
0.2344970703125,
0.5888671875
],
[
0.55810546875,
0.5791015625,
0.9140625,
0.5791015625,
0.9140625,
0.59228515625,
0.55810546875,
0.59228515625
],
[
0.08416748046875,
0.59326171875,
0.265625,
0.5947265625,
0.265625,
0.60791015625,
0.08416748046875,
0.60693359375
],
[
0.55810546875,
0.5947265625,
0.6611328125,
0.5947265625,
0.6611328125,
0.60498046875,
0.55810546875,
0.60498046875
],
[
0.521484375,
0.611328125,
0.9111328125,
0.611328125,
0.9111328125,
0.6220703125,
0.521484375,
0.6220703125
],
[
0.0882568359375,
0.62255859375,
0.11358642578125,
0.62255859375,
0.11358642578125,
0.630859375,
0.0882568359375,
0.630859375
],
[
0.1258544921875,
0.62060546875,
0.30224609375,
0.62060546875,
0.30224609375,
0.6337890625,
0.1258544921875,
0.6337890625
],
[
0.5185546875,
0.62255859375,
0.91259765625,
0.62255859375,
0.91259765625,
0.63623046875,
0.5185546875,
0.63623046875
],
[
0.5185546875,
0.63720703125,
0.9140625,
0.63720703125,
0.9140625,
0.65087890625,
0.5185546875,
0.65087890625
],
[
0.1094970703125,
0.62353515625,
0.1298828125,
0.6220703125,
0.131591796875,
0.6298828125,
0.111083984375,
0.6318359375
],
[
0.0882568359375,
0.64013671875,
0.48046875,
0.64013671875,
0.48046875,
0.65087890625,
0.0882568359375,
0.65087890625
],
[
0.0882568359375,
0.65380859375,
0.4794921875,
0.65380859375,
0.4794921875,
0.6640625,
0.0882568359375,
0.6640625
],
[
0.521484375,
0.65283203125,
0.63623046875,
0.65283203125,
0.63623046875,
0.66357421875,
0.521484375,
0.66357421875
],
[
0.08660888671875,
0.66650390625,
0.468017578125,
0.66650390625,
0.468017578125,
0.6796875,
0.08660888671875,
0.6796875
],
[
0.08660888671875,
0.6796875,
0.48046875,
0.6796875,
0.48046875,
0.693359375,
0.08660888671875,
0.693359375
],
[
0.53515625,
0.6748046875,
0.91259765625,
0.6748046875,
0.91259765625,
0.68505859375,
0.53515625,
0.68505859375
],
[
0.5185546875,
0.6875,
0.91259765625,
0.6875,
0.91259765625,
0.70068359375,
0.5185546875,
0.70068359375
],
[
0.0882568359375,
0.69580078125,
0.4794921875,
0.69580078125,
0.4794921875,
0.70556640625,
0.0882568359375,
0.70556640625
],
[
0.5185546875,
0.70068359375,
0.91259765625,
0.70068359375,
0.91259765625,
0.7138671875,
0.5185546875,
0.7138671875
],
[
0.0882568359375,
0.71044921875,
0.4794921875,
0.71044921875,
0.4794921875,
0.72021484375,
0.0882568359375,
0.72021484375
],
[
0.5185546875,
0.7138671875,
0.91259765625,
0.7138671875,
0.91259765625,
0.72802734375,
0.5185546875,
0.72802734375
],
[
0.0882568359375,
0.7236328125,
0.1478271484375,
0.7236328125,
0.1478271484375,
0.73486328125,
0.0882568359375,
0.73486328125
],
[
0.51708984375,
0.72802734375,
0.91259765625,
0.72900390625,
0.91259765625,
0.7421875,
0.51708984375,
0.7412109375
],
[
0.10296630859375,
0.74365234375,
0.1544189453125,
0.74365234375,
0.1544189453125,
0.75390625,
0.10296630859375,
0.75390625
],
[
0.1519775390625,
0.744140625,
0.4794921875,
0.744140625,
0.4794921875,
0.7548828125,
0.1519775390625,
0.7548828125
],
[
0.51953125,
0.74365234375,
0.9111328125,
0.74365234375,
0.9111328125,
0.75390625,
0.51953125,
0.75390625
],
[
0.08660888671875,
0.7568359375,
0.48046875,
0.7568359375,
0.48046875,
0.77099609375,
0.08660888671875,
0.77099609375
],
[
0.51953125,
0.755859375,
0.91259765625,
0.755859375,
0.91259765625,
0.76953125,
0.51953125,
0.76953125
],
[
0.0882568359375,
0.77294921875,
0.12335205078125,
0.77294921875,
0.12335205078125,
0.7841796875,
0.0882568359375,
0.7841796875
],
[
0.51953125,
0.771484375,
0.9111328125,
0.771484375,
0.9111328125,
0.7822265625,
0.51953125,
0.7822265625
],
[
0.5185546875,
0.7822265625,
0.9140625,
0.78271484375,
0.9140625,
0.79736328125,
0.5185546875,
0.796875
],
[
0.2418212890625,
0.79248046875,
0.330078125,
0.79248046875,
0.330078125,
0.8037109375,
0.2418212890625,
0.8037109375
],
[
0.4599609375,
0.79150390625,
0.48193359375,
0.79150390625,
0.48193359375,
0.8046875,
0.4599609375,
0.8046875
],
[
0.51953125,
0.7998046875,
0.59619140625,
0.7998046875,
0.59619140625,
0.81103515625,
0.51953125,
0.81103515625
],
[
0.23779296875,
0.81005859375,
0.328369140625,
0.81005859375,
0.328369140625,
0.8212890625,
0.23779296875,
0.8212890625
],
[
0.459228515625,
0.8095703125,
0.48193359375,
0.8095703125,
0.48193359375,
0.82275390625,
0.459228515625,
0.82275390625
],
[
0.2418212890625,
0.82763671875,
0.328369140625,
0.82763671875,
0.328369140625,
0.8388671875,
0.2418212890625,
0.8388671875
],
[
0.4599609375,
0.82763671875,
0.48193359375,
0.82763671875,
0.48193359375,
0.84033203125,
0.4599609375,
0.84033203125
],
[
0.51708984375,
0.82470703125,
0.84326171875,
0.82470703125,
0.84326171875,
0.83837890625,
0.51708984375,
0.83837890625
],
[
0.08660888671875,
0.8466796875,
0.468017578125,
0.8466796875,
0.468017578125,
0.85986328125,
0.08660888671875,
0.85986328125
],
[
0.5185546875,
0.84228515625,
0.9140625,
0.84228515625,
0.9140625,
0.85595703125,
0.5185546875,
0.85595703125
],
[
0.08660888671875,
0.861328125,
0.48046875,
0.861328125,
0.48046875,
0.875,
0.08660888671875,
0.875
],
[
0.51708984375,
0.85595703125,
0.8896484375,
0.85595703125,
0.8896484375,
0.869140625,
0.51708984375,
0.869140625
],
[
0.625,
0.87744140625,
0.8056640625,
0.87744140625,
0.8056640625,
0.8916015625,
0.625,
0.8916015625
],
[
0.8837890625,
0.87744140625,
0.91259765625,
0.87744140625,
0.91259765625,
0.89013671875,
0.8837890625,
0.89013671875
],
[
0.0849609375,
0.880859375,
0.2646484375,
0.88330078125,
0.2646484375,
0.896484375,
0.0849609375,
0.89453125
]
] |
[
"2.2.1",
"ExtractivePooling，提取共同注意力的最常见变体",
"是，我们通过比较共同注意之前和之后的表示来对共同注意",
"是最大池共同注意力，它根据每个单词对其他文本序列的最",
"的影响进行建模。对于软注意力对齐，这里需要注意的是x和",
"大影响来关注每个单词。",
"x（尽管长度相等）具有“交换”语义。换句话说，在9的情",
"q' = Soft(max(s)Tg and d' = Soft(max(s)Td",
"况下，9实际上包含d的对齐表示。最后，使用多个比较运算",
"",
"(3",
"col",
"符（减法，串联和乘法运算符）是为了捕获多个视角，并受到",
"，其中q',d分别是q和d的共同注意表示。软（。）是Softmax",
"ESIM模型[3]的启发",
"运算符。或者，也可以使用矩阵s的平均行和列池化：",
"2.3.2",
"CompressionFunction，本节定义Fc（）所使用的压",
"q' = Soft(mean(s))Tq and d'= Soft(mean(s))Td",
"(4)",
"缩函数。压缩的基本原理简单直观-我们不想用高维向量膨胀",
"col",
"后续层，从而导致后续层中的参数成本。我们研究了三个压缩",
"但是，每个池化运算符都有不同的影响，可以直观地理解为：",
"函数的使用，它们能够将n维向量简化为标量。",
"最大池化根据每个单词在其他文本中所有单词的最大重要性",
". Sum (SM) Function is a non-parameterized function",
"来选择每个单词。均值池是一种更健康的比较，根据单词对其",
"that sums the entire vector, returning a scalar as an",
"他文本的整体影响来关注单词。这通常与数据集相关，被视为",
"output.",
"超参数，并进行调整以查看哪个在保留的集合上表现最佳",
"",
"2.2.2Alignment-Pooling·基于软对齐的池化也已用于学",
"F(x)=xi, VxiEx",
"(10)",
"习共同注意表示[20]。然而，软对齐的关键区别在于它重新",
"人",
"排列了序列对，而标准的共同注意只是学会了对重要单词进",
". Neural Network (NN) is a fully-connected layer that",
"行加权和评分。然后按如下方式学习共同注意表示：",
"converts each n dimensional feature vector as follows:",
"fg",
"exp(sij)",
"td",
"exp(sij)",
"F(x) = ReLU(W(x) + bc),",
"(11)",
"dj :=",
"Zsla",
"qj and qj := 2",
"di (5)",
"B12k=1",
"台,exp(sk)",
"",
"exp(sik)",
"where We × Rnx1 and bc R are the parameters of the",
"其中d是q中与di柔和对齐的子短语。直观上，d 是(q",
"FC layer.",
". Factorization Machines (FM) are general purpose ma-",
"的加权和，选择9中最相关的部分来表示d",
"chine learning techniques that accept a real-valued fea-",
"最近提出了",
"ture vector x e Rn and return a scalar output.",
"2.2.3Intra-Attention.Intra-Attention 或 Self-Attention",
"n",
"",
"F(x) = wo + Zwi xi +",
" <ui.oj) xi xj",
"(12)",
"",
"i=1 j=i+1",
"i=1",
"来学习意识到长期依赖关系的表示形式。这通常被表述为相",
"对于自身的共同关注（或对齐）操作。在这种情况下，我们将",
"内部关注独立地应用于文档和查询。为了简单起见，我们将它",
"where wo e R, wi e R\" and (ui,..- n) e Rnxk are the",
"们称为x，而不是g或d。注意内函数定义为：",
"parameters of the FM model. FMs are expressive mod-",
"exp(sij)",
"elsthatcapturepairwiseinteractionsbetweenfeatures",
"x =",
"P",
"(6)",
"using factorized parameters. k is the number of factors",
"Xj",
"台Zk1 exp(sik)",
"of the FM model. For more details, we refer interested",
"其中x是x，的注意内表示",
"readersto24]",
"请注意，我们不会在多个注意力投射之间共享参数，因为每个",
"2.3",
"Multi-CastAttention",
"注意力投射都旨在对不同的视图进行建模。我们的实验分别",
"在模型名称MCAN（SM)，MCAN（NN）和MCAN（FM)",
"",
"在这一点上，很容易进行几个观察。首先，每种注意力机制都",
"为模型提供了不同的风味。其次，注意力被用来通过重新加权",
"下报告了上述变体",
"或重新对齐来改变原始表示。因此，大多数神经架构仅使用",
"种类型的共同注意或对齐函数[5，20]。但是，这需要调整正",
"2.3.3Multi-Cast，我们的体系结构背后的关键思想是促",
"进k注意调用（或强制转换），每个强制转换都用一个真正有",
"确的模型体系结构，并且可能会错过使用多种共同注意机制",
"价值的注意提示来增强原始单词嵌人。我们正式描述了多播",
"带来的好处。因此，我们的工作将每个注意力操作都转换为单",
"注意力机制。对于每个查询文档对，我们应用（1）具有均值",
"词级功能",
"池的共同注意力（2）具有最大池的共同注意力和（3）具有对",
"2.3.1",
"CastedAttention令x为q或d，而x是应用共",
"齐池的共同注意力。此外，我们将Intra-Attention分别应用",
"同关注或软关注对齐后x的表示1。共注意运算符的注意特",
"于查询和文档。每个注意力投射都会产生三个标量（每个单",
"征是：",
"词），这些标量与单词嵌入连接在一起。最终铸造的特征向量",
"是zeR12。因此，对于每个单词wi，新单词表示形式变为",
"fe = Fe([x;xl)",
"(7)",
"wi=[wi;zi]",
"fm=Fc(xOx)",
"(8)",
"fs = Fc(x - x)",
"(9)",
"2.4Long Short-Term Memory Encoder",
"其中是Hadamard乘积，［]是串联运算符。F(）是",
"接下来，将引起注意的单词表示w1,w2..we传递到顺序编码",
"个压缩函数，用于将特征缩减为标量。直观地说，这里实现的",
"器层。我们采用标准的香草长短期记忆（LSTM）编码器：",
"h, = LSTM(u,i), Vi e [1,...]",
"(13)",
"IWe omit subscripts for clarity."
] |
{
"arxiv_id": "1806.00778",
"language": "zh",
"page_idx": 2
}
|
||
1806.00778-zh-3
|
1806.00778-zh-3.jpg
|
,其中$\ell$表示序列的最大长度。值得注意的是，LSTM 的参数本质上是“siamese”,在文档和查询之间共享权重。关键思想是 LSTM 编码器通过使用非线性变换作为门控函数来学习知道顺序依赖性的表示。由于 LSTM 是标准的神经构建块，因此我们省略了技术细节，以简洁。因此，将注意力作为该层之前的功能的关键思想是，它为 LSTM 编码器提供了提示，这些提示提供了诸如(1)长期和全局句子知识以及(2)句子对(文档和查询)
2.4.1 Pooling Operation.最后，将池化函数应用于每个
句子的隐藏状态$\{h_1\ldots h_\ell\}$ ,将序列转换为固定维表示。
$$h=MeanMax[h_1\ldots h_\ell]$$
(14)
我们采用 MeanMax 池运算符，它将平均池和最大池的结果连接在一起。我们发现这始终比单独使用 max 或 mean 池更好
2.5 Prediction Layer and Optimization
最后，给定文档-查询对的固定维表示形式，我们将它们的串联传递到两层$h$维高速公路网络中。我们模型的最终预测层计算如下：
(15)
$$y_{out}=H_2(H_1([x_q;x_d;x_q\odot x_d;x_q-x_d]))$$
其中$H_1(.),H_2(.)$是具有 ReLU 激活的公路网络层。然后将输
出传递到最终的线性 softmax 层。
(16)
$$y_{pred}=softmax(W_F\cdot y_{out}+b_F)$$
其中$W_F\in\mathbb{R}^{h\times2}$和$b_F\in\mathbb{R}^2$ 。然后使用具有 L2 正则化的标准
多类交叉熵损失来训练网络
$$J(\theta)=-\sum_{i=1}^N\left[y_i\log\hat{y}_i+(1-y_i)\log(1-\hat{y}_i)\right]+\lambda||\theta||_{L2}$$
(17
其中$\theta$是网络的参数。$\hat{y}$是网络的输出。$||\theta||_L2$是 L2 正则化，
$\lambda$是正则化的权重
# 3 EMPIRICAL EVALUATION
在我们的实验中，我们旨在回答以下研究问题(RQs):
(1) RQ1 - Does our proposed approach achieve state-of-
the-art performance on question answering and conversation modeling tasks? What is the relative improvement over well-established baselines?
(2) RQ2 - What are the impacts of architectural design
on performance? Is the LSTM encoder necessary to make use of the casted features? Does all the variations of co-attention contribute to the overall model performance?
(3) RQ3- Can we explain the inner workings of our pro-
posed model? Can we interpret the casted attention
features?
3.1 Experiment 1- Dialogue Predictiom
在第一个任务中，我们评估我们的模型成功预测对话中下一
个回复的能力
3.1.1 Dataset and Evaluation Metric.对于此实验，我们利用大型且众所周知的大型 Ubuntu 对话语料库 (UDC)[15] 。我们使用 Xu 等人提供的相同测试拆分。[51]。在此任务中， 目标是将句子与其回复匹配。在[46]之后，该任务主要利用每个对话中的最后两个话语，预测后者是否遵循前者。训练集由一百万个具有 1:1正负比的消息-响应对组成。开发和测试集具有 9:1 比率。在[46,51]之后，我们使用召回 @k $(R_n@K$ )的评估指标，该指标指示$n$候选项的前$k$结果中是否存在基本事实。使用的四个评估指标是$R_2\textcircled{a}$1、$R_{10}\textcircled{a}$1、$R_{10}\textcircled{a}$2 和$R_{10}\textcircled{a}5$
3.1.2 Competitive Baselines and Implementation Details. 我们与大量竞争基线进行比较，例如。g。, MLP, DeepMatch [16] , ARC- I/ ARC- II [ 9] , CNTN [ 21] , MatchPyramid [ 19] ,香草 LSTM, 注意力池 LSTM [5] , MV-LSTM [37] ,最后是最先进的知识增强的混合神经网络(KEHN)。可以在[46] 上找到基线的详细说明。由于测试拆分是相同的，我们直接从[46]报告结果。为了公平比较，我们将 MCAN 中的 LSTM 编码器大小设置为$d=100$,这使其等于[46]中的模型。我们使用 Adam 优化器 [13] 优化 MCAN, 初始学习率为 3$\times10^{-4}$ 。将 0.2 的辍学率应用于除单词嵌人层以外的所有层。序列被动态截断或填充到其批处理最大值(具有 50 令牌的硬限制)。我们用预训练的 GloVe 嵌人初始化单词嵌入层
3.1.3 Experimental Results.表 1 报告了我们的实验结果。显然，我们观察到所有 MCAN 模型都比现有的最先进模型获得了巨大的性能提升。更具体地说，所有指标的改进都比 KEHNN 更好。相对于 AP-LSTM 和 MV-LSTM 等强基线的性能提升更大，在$R_{10}\textcircled{a}$1 方面达到了 15%的提升。这确定了 MCAN 模型的有效性。总体而言，MCAN (FM) 和MCAN (NN) 在性能方面具有可比性。MCAN(SM) 略低于MCAN(FM) 和 MCAN(NN)。然而，它的性能仍然大大高于现有的最先进的模型。
<table>
<tbody>
<tr>
<th> </th>
<th>$R_2\textcircled{a}1$</th>
<th>$R_{10}\textcircled{a}1$</th>
<th>$R_{10}\textcircled{a}2$</th>
<th>$R_{10}\textcircled{a}5$</th>
</tr>
<tr>
<td>MLP</td>
<td>0.651</td>
<td>0.256</td>
<td>0.380</td>
<td>0.703</td>
</tr>
<tr>
<td>DeepMatch</td>
<td>0.593</td>
<td>0.345</td>
<td>0.376</td>
<td>0.693</td>
</tr>
<tr>
<td>ARC- I</td>
<td>0.665</td>
<td>0.221</td>
<td>0.360</td>
<td>0.684</td>
</tr>
<tr>
<td>ARC- II</td>
<td>0.736</td>
<td>0.380</td>
<td>0.534</td>
<td>0.777</td>
</tr>
<tr>
<td>CNTN</td>
<td>0.743</td>
<td>0.349</td>
<td>0.512</td>
<td>0.797</td>
</tr>
<tr>
<td>MatchP vramid</td>
<td>0.743</td>
<td>0.420</td>
<td>0.554</td>
<td>0.786</td>
</tr>
<tr>
<td>LSTM</td>
<td>0.725</td>
<td>0.361</td>
<td>0.494</td>
<td>0.801</td>
</tr>
<tr>
<td>AP- LSTM</td>
<td>0.758</td>
<td>0.381</td>
<td>0.545</td>
<td>0.801</td>
</tr>
<tr>
<td>MV- LSTM</td>
<td>0.767</td>
<td>0.410</td>
<td>0.565</td>
<td>0.800</td>
</tr>
<tr>
<td>KEHNN</td>
<td>0.786</td>
<td>0.460</td>
<td>0.591</td>
<td>0.819</td>
</tr>
<tr>
<td>MCAN $\overline{\mathrm{SM}}$</td>
<td>0.831</td>
<td>0.548</td>
<td>0.682</td>
<td>0.873</td>
</tr>
<tr>
<td>MCAN $^{\prime }$NN</td>
<td>0.833</td>
<td>0.549</td>
<td>0.686</td>
<td>0.875</td>
</tr>
<tr>
<td>MCAN (FM)</td>
<td>0.834</td>
<td>0.551</td>
<td>0.684</td>
<td>0.875</td>
</tr>
</tbody>
</table>
Table $1: \textbf{Performance Comparison on Ubuntu Dialogue Cor- }$
pus. Best result is in boldface and second best is underlined.
3.2 Experiment 2-Factoid Question Answering
Factoid 问题回答是回答基于事实的问题的任务。在此任务中，
目标是提供给定问题的答案的排名列表
|
[
[
0.0906982421875,
0.10797119140625,
0.48046875,
0.10797119140625,
0.48046875,
0.121826171875,
0.0906982421875,
0.121826171875
],
[
0.5361328125,
0.10919189453125,
0.576171875,
0.10919189453125,
0.576171875,
0.11932373046875,
0.5361328125,
0.11932373046875
],
[
0.583984375,
0.10797119140625,
0.91259765625,
0.10797119140625,
0.91259765625,
0.121826171875,
0.583984375,
0.121826171875
],
[
0.08660888671875,
0.121826171875,
0.4794921875,
0.121826171875,
0.4794921875,
0.1351318359375,
0.08660888671875,
0.1351318359375
],
[
0.5185546875,
0.121826171875,
0.91259765625,
0.121826171875,
0.91259765625,
0.1351318359375,
0.5185546875,
0.1351318359375
],
[
0.0849609375,
0.1337890625,
0.48046875,
0.1351318359375,
0.48046875,
0.1483154296875,
0.0849609375,
0.147705078125
],
[
0.51953125,
0.136962890625,
0.91259765625,
0.136962890625,
0.91259765625,
0.147705078125,
0.51953125,
0.147705078125
],
[
0.08660888671875,
0.149658203125,
0.48046875,
0.149658203125,
0.48046875,
0.1634521484375,
0.08660888671875,
0.1634521484375
],
[
0.51953125,
0.149658203125,
0.9140625,
0.149658203125,
0.9140625,
0.1634521484375,
0.51953125,
0.1634521484375
],
[
0.0882568359375,
0.1654052734375,
0.47802734375,
0.1654052734375,
0.47802734375,
0.175537109375,
0.0882568359375,
0.175537109375
],
[
0.51953125,
0.1634521484375,
0.91259765625,
0.1634521484375,
0.91259765625,
0.1767578125,
0.51953125,
0.1767578125
],
[
0.08905029296875,
0.1787109375,
0.47802734375,
0.1787109375,
0.47802734375,
0.189453125,
0.08905029296875,
0.189453125
],
[
0.52392578125,
0.18115234375,
0.53173828125,
0.18115234375,
0.53173828125,
0.1849365234375,
0.52392578125,
0.1849365234375
],
[
0.53125,
0.1787109375,
0.9111328125,
0.1787109375,
0.9111328125,
0.189453125,
0.53125,
0.189453125
],
[
0.08660888671875,
0.1912841796875,
0.48046875,
0.1912841796875,
0.48046875,
0.2052001953125,
0.08660888671875,
0.2052001953125
],
[
0.5185546875,
0.1900634765625,
0.91259765625,
0.1912841796875,
0.91259765625,
0.2052001953125,
0.5185546875,
0.203857421875
],
[
0.08660888671875,
0.20703125,
0.1732177734375,
0.20703125,
0.1732177734375,
0.2183837890625,
0.08660888671875,
0.2183837890625
],
[
0.51953125,
0.2052001953125,
0.9140625,
0.2052001953125,
0.9140625,
0.2183837890625,
0.51953125,
0.2183837890625
],
[
0.5185546875,
0.2183837890625,
0.9140625,
0.2183837890625,
0.9140625,
0.231689453125,
0.5185546875,
0.231689453125
],
[
0.10455322265625,
0.2291259765625,
0.1544189453125,
0.2291259765625,
0.1544189453125,
0.240478515625,
0.10455322265625,
0.240478515625
],
[
0.1519775390625,
0.2266845703125,
0.4794921875,
0.2279052734375,
0.4794921875,
0.2412109375,
0.1519775390625,
0.240478515625
],
[
0.515625,
0.23291015625,
0.56201171875,
0.23291015625,
0.56201171875,
0.2474365234375,
0.515625,
0.2474365234375
],
[
0.0849609375,
0.242431640625,
0.44287109375,
0.242431640625,
0.44287109375,
0.255615234375,
0.0849609375,
0.255615234375
],
[
0.5361328125,
0.2568359375,
0.57177734375,
0.2568359375,
0.57177734375,
0.268310546875,
0.5361328125,
0.268310546875
],
[
0.583984375,
0.255615234375,
0.91259765625,
0.255615234375,
0.91259765625,
0.26953125,
0.583984375,
0.26953125
],
[
0.2091064453125,
0.266357421875,
0.358642578125,
0.266357421875,
0.358642578125,
0.27978515625,
0.2091064453125,
0.27978515625
],
[
0.451904296875,
0.267578125,
0.48046875,
0.267578125,
0.48046875,
0.27978515625,
0.451904296875,
0.27978515625
],
[
0.51708984375,
0.26953125,
0.91259765625,
0.268310546875,
0.91259765625,
0.28271484375,
0.51708984375,
0.2841796875
],
[
0.51953125,
0.2841796875,
0.91259765625,
0.2841796875,
0.91259765625,
0.29736328125,
0.51953125,
0.29736328125
],
[
0.0882568359375,
0.29345703125,
0.47802734375,
0.29345703125,
0.47802734375,
0.3037109375,
0.0882568359375,
0.3037109375
],
[
0.52685546875,
0.296630859375,
0.91259765625,
0.29736328125,
0.91259765625,
0.3125,
0.52685546875,
0.311279296875
],
[
0.0882568359375,
0.306884765625,
0.473876953125,
0.306884765625,
0.473876953125,
0.317626953125,
0.0882568359375,
0.317626953125
],
[
0.525390625,
0.302978515625,
0.53173828125,
0.302978515625,
0.53173828125,
0.30810546875,
0.525390625,
0.30810546875
],
[
0.5185546875,
0.31005859375,
0.91259765625,
0.311279296875,
0.91259765625,
0.32568359375,
0.5185546875,
0.324462890625
],
[
0.521484375,
0.32568359375,
0.91259765625,
0.32568359375,
0.91259765625,
0.339111328125,
0.521484375,
0.339111328125
],
[
0.08416748046875,
0.333251953125,
0.41015625,
0.333984375,
0.41015625,
0.348388671875,
0.08416748046875,
0.347900390625
],
[
0.51953125,
0.339111328125,
0.91259765625,
0.339111328125,
0.91259765625,
0.352783203125,
0.51953125,
0.352783203125
],
[
0.0882568359375,
0.3515625,
0.4794921875,
0.3515625,
0.4794921875,
0.365478515625,
0.0882568359375,
0.365478515625
],
[
0.5185546875,
0.352783203125,
0.91259765625,
0.352783203125,
0.91259765625,
0.3662109375,
0.5185546875,
0.3662109375
],
[
0.0849609375,
0.3662109375,
0.48046875,
0.3662109375,
0.48046875,
0.380126953125,
0.0849609375,
0.380126953125
],
[
0.5185546875,
0.3662109375,
0.9111328125,
0.3662109375,
0.9111328125,
0.380126953125,
0.5185546875,
0.380126953125
],
[
0.08660888671875,
0.380126953125,
0.1544189453125,
0.3818359375,
0.1544189453125,
0.393310546875,
0.08660888671875,
0.391357421875
],
[
0.525390625,
0.386474609375,
0.53173828125,
0.386474609375,
0.53173828125,
0.389404296875,
0.525390625,
0.389404296875
],
[
0.52783203125,
0.380615234375,
0.9140625,
0.380615234375,
0.9140625,
0.39453125,
0.52783203125,
0.39453125
],
[
0.5185546875,
0.39453125,
0.9140625,
0.39453125,
0.9140625,
0.40771484375,
0.5185546875,
0.40771484375
],
[
0.1568603515625,
0.40478515625,
0.408447265625,
0.40478515625,
0.408447265625,
0.419189453125,
0.1568603515625,
0.419189453125
],
[
0.450927734375,
0.40478515625,
0.48046875,
0.40478515625,
0.48046875,
0.417236328125,
0.450927734375,
0.417236328125
],
[
0.5185546875,
0.410400390625,
0.81884765625,
0.410400390625,
0.81884765625,
0.42041015625,
0.5185546875,
0.42041015625
],
[
0.0849609375,
0.427978515625,
0.48046875,
0.4287109375,
0.48046875,
0.443115234375,
0.0849609375,
0.442626953125
],
[
0.53515625,
0.43310546875,
0.576171875,
0.43310546875,
0.576171875,
0.4443359375,
0.53515625,
0.4443359375
],
[
0.57421875,
0.436279296875,
0.5810546875,
0.436279296875,
0.5810546875,
0.43994140625,
0.57421875,
0.43994140625
],
[
0.5849609375,
0.431884765625,
0.91259765625,
0.431884765625,
0.91259765625,
0.44580078125,
0.5849609375,
0.44580078125
],
[
0.0849609375,
0.44140625,
0.303955078125,
0.442626953125,
0.303955078125,
0.45703125,
0.0849609375,
0.455810546875
],
[
0.51708984375,
0.4443359375,
0.9140625,
0.44580078125,
0.9140625,
0.460205078125,
0.51708984375,
0.458984375
],
[
0.5185546875,
0.460205078125,
0.91259765625,
0.460205078125,
0.91259765625,
0.473388671875,
0.5185546875,
0.473388671875
],
[
0.1805419921875,
0.466552734375,
0.384033203125,
0.464111328125,
0.384033203125,
0.481689453125,
0.1805419921875,
0.484130859375
],
[
0.451904296875,
0.466552734375,
0.48046875,
0.468505859375,
0.4794921875,
0.481689453125,
0.450927734375,
0.479736328125
],
[
0.5185546875,
0.473388671875,
0.91259765625,
0.473388671875,
0.91259765625,
0.4873046875,
0.5185546875,
0.4873046875
],
[
0.08660888671875,
0.492431640625,
0.48193359375,
0.492431640625,
0.48193359375,
0.5068359375,
0.08660888671875,
0.5068359375
],
[
0.51953125,
0.4892578125,
0.9111328125,
0.4892578125,
0.9111328125,
0.5,
0.51953125,
0.5
],
[
0.5185546875,
0.50048828125,
0.9140625,
0.50048828125,
0.9140625,
0.5146484375,
0.5185546875,
0.5146484375
],
[
0.08905029296875,
0.51025390625,
0.26318359375,
0.51025390625,
0.26318359375,
0.52099609375,
0.08905029296875,
0.52099609375
],
[
0.51708984375,
0.5146484375,
0.9111328125,
0.5146484375,
0.9111328125,
0.52880859375,
0.51708984375,
0.52880859375
],
[
0.51708984375,
0.52783203125,
0.9111328125,
0.52978515625,
0.9111328125,
0.54345703125,
0.51708984375,
0.54150390625
],
[
0.1666259765625,
0.54248046875,
0.181396484375,
0.54248046875,
0.181396484375,
0.55322265625,
0.1666259765625,
0.55322265625
],
[
0.160888671875,
0.5498046875,
0.483642578125,
0.5498046875,
0.483642578125,
0.5712890625,
0.160888671875,
0.5712890625
],
[
0.51953125,
0.54443359375,
0.65673828125,
0.54443359375,
0.65673828125,
0.5546875,
0.51953125,
0.5546875
],
[
0.1070556640625,
0.5537109375,
0.1650390625,
0.55615234375,
0.1650390625,
0.568359375,
0.1070556640625,
0.56640625
],
[
0.1650390625,
0.5625,
0.188720703125,
0.5673828125,
0.181396484375,
0.58349609375,
0.158447265625,
0.57763671875
],
[
0.66845703125,
0.580078125,
0.8798828125,
0.580078125,
0.8798828125,
0.5947265625,
0.66845703125,
0.5947265625
],
[
0.0849609375,
0.5869140625,
0.483642578125,
0.58837890625,
0.483642578125,
0.60302734375,
0.0849609375,
0.6015625
],
[
0.54833984375,
0.59326171875,
0.59326171875,
0.59326171875,
0.59326171875,
0.60791015625,
0.54833984375,
0.60791015625
],
[
0.6689453125,
0.59521484375,
0.70751953125,
0.59521484375,
0.70751953125,
0.60693359375,
0.6689453125,
0.60693359375
],
[
0.7255859375,
0.59521484375,
0.7646484375,
0.59521484375,
0.7646484375,
0.60693359375,
0.7255859375,
0.60693359375
],
[
0.78125,
0.59521484375,
0.82177734375,
0.59521484375,
0.82177734375,
0.60693359375,
0.78125,
0.60693359375
],
[
0.8349609375,
0.5947265625,
0.87841796875,
0.5947265625,
0.87841796875,
0.609375,
0.8349609375,
0.609375
],
[
0.0849609375,
0.60302734375,
0.205078125,
0.60302734375,
0.205078125,
0.6162109375,
0.0849609375,
0.6162109375
],
[
0.66845703125,
0.60791015625,
0.708984375,
0.60791015625,
0.708984375,
0.62255859375,
0.66845703125,
0.62255859375
],
[
0.7236328125,
0.60791015625,
0.765625,
0.60791015625,
0.765625,
0.62255859375,
0.7236328125,
0.62255859375
],
[
0.779296875,
0.60791015625,
0.82275390625,
0.60791015625,
0.82275390625,
0.62255859375,
0.779296875,
0.62255859375
],
[
0.8349609375,
0.60791015625,
0.87841796875,
0.60791015625,
0.87841796875,
0.62255859375,
0.8349609375,
0.62255859375
],
[
0.66845703125,
0.6220703125,
0.7099609375,
0.6220703125,
0.7099609375,
0.63623046875,
0.66845703125,
0.63623046875
],
[
0.5556640625,
0.609375,
0.63134765625,
0.609375,
0.63134765625,
0.62060546875,
0.5556640625,
0.62060546875
],
[
0.7236328125,
0.6220703125,
0.7646484375,
0.6220703125,
0.7646484375,
0.63623046875,
0.7236328125,
0.63623046875
],
[
0.779296875,
0.6220703125,
0.82177734375,
0.6220703125,
0.82177734375,
0.63623046875,
0.779296875,
0.63623046875
],
[
0.8349609375,
0.6220703125,
0.87841796875,
0.6220703125,
0.87841796875,
0.63623046875,
0.8349609375,
0.63623046875
],
[
0.1168212890625,
0.630859375,
0.351318359375,
0.630859375,
0.351318359375,
0.64453125,
0.1168212890625,
0.64453125
],
[
0.5556640625,
0.62255859375,
0.60107421875,
0.62255859375,
0.60107421875,
0.6337890625,
0.5556640625,
0.6337890625
],
[
0.0849609375,
0.6318359375,
0.1168212890625,
0.6318359375,
0.1168212890625,
0.64453125,
0.0849609375,
0.64453125
],
[
0.5537109375,
0.63525390625,
0.60888671875,
0.63525390625,
0.60888671875,
0.6494140625,
0.5537109375,
0.6494140625
],
[
0.66650390625,
0.63525390625,
0.708984375,
0.63525390625,
0.708984375,
0.6494140625,
0.66650390625,
0.6494140625
],
[
0.7236328125,
0.63525390625,
0.765625,
0.63525390625,
0.765625,
0.6494140625,
0.7236328125,
0.6494140625
],
[
0.779296875,
0.63525390625,
0.82177734375,
0.63525390625,
0.82177734375,
0.6494140625,
0.779296875,
0.6494140625
],
[
0.8349609375,
0.63623046875,
0.87744140625,
0.63623046875,
0.87744140625,
0.65087890625,
0.8349609375,
0.65087890625
],
[
0.0849609375,
0.6494140625,
0.438720703125,
0.6494140625,
0.438720703125,
0.66357421875,
0.0849609375,
0.66357421875
],
[
0.66650390625,
0.6494140625,
0.708984375,
0.6494140625,
0.708984375,
0.6640625,
0.66650390625,
0.6640625
],
[
0.7236328125,
0.6494140625,
0.765625,
0.6494140625,
0.765625,
0.6640625,
0.7236328125,
0.6640625
],
[
0.779296875,
0.6494140625,
0.82177734375,
0.6494140625,
0.82177734375,
0.6640625,
0.779296875,
0.6640625
],
[
0.8349609375,
0.6494140625,
0.87744140625,
0.6494140625,
0.87744140625,
0.6640625,
0.8349609375,
0.6640625
],
[
0.65673828125,
0.66357421875,
0.708984375,
0.66357421875,
0.708984375,
0.67822265625,
0.65673828125,
0.67822265625
],
[
0.5556640625,
0.65087890625,
0.60107421875,
0.65087890625,
0.60107421875,
0.662109375,
0.5556640625,
0.662109375
],
[
0.7236328125,
0.66357421875,
0.765625,
0.66357421875,
0.765625,
0.67822265625,
0.7236328125,
0.67822265625
],
[
0.779296875,
0.66357421875,
0.82177734375,
0.66357421875,
0.82177734375,
0.67822265625,
0.779296875,
0.67822265625
],
[
0.8349609375,
0.66357421875,
0.87744140625,
0.66357421875,
0.87744140625,
0.67822265625,
0.8349609375,
0.67822265625
],
[
0.5556640625,
0.66552734375,
0.66015625,
0.66552734375,
0.66015625,
0.67822265625,
0.5556640625,
0.67822265625
],
[
0.101318359375,
0.66845703125,
0.483642578125,
0.66845703125,
0.483642578125,
0.681640625,
0.101318359375,
0.681640625
],
[
0.66650390625,
0.6767578125,
0.708984375,
0.6767578125,
0.708984375,
0.6923828125,
0.66650390625,
0.6923828125
],
[
0.7236328125,
0.6767578125,
0.7646484375,
0.6767578125,
0.7646484375,
0.69140625,
0.7236328125,
0.69140625
],
[
0.779296875,
0.6767578125,
0.82177734375,
0.6767578125,
0.82177734375,
0.69140625,
0.779296875,
0.69140625
],
[
0.8349609375,
0.67822265625,
0.87744140625,
0.67822265625,
0.87744140625,
0.6923828125,
0.8349609375,
0.6923828125
],
[
0.5556640625,
0.6787109375,
0.6005859375,
0.6787109375,
0.6005859375,
0.68994140625,
0.5556640625,
0.68994140625
],
[
0.12744140625,
0.68310546875,
0.47802734375,
0.68310546875,
0.47802734375,
0.693359375,
0.12744140625,
0.693359375
],
[
0.66650390625,
0.69140625,
0.708984375,
0.69140625,
0.708984375,
0.70556640625,
0.66650390625,
0.70556640625
],
[
0.7236328125,
0.69140625,
0.765625,
0.69140625,
0.765625,
0.70556640625,
0.7236328125,
0.70556640625
],
[
0.779296875,
0.69140625,
0.82177734375,
0.69140625,
0.82177734375,
0.70556640625,
0.779296875,
0.70556640625
],
[
0.8349609375,
0.69140625,
0.87744140625,
0.69140625,
0.87744140625,
0.70556640625,
0.8349609375,
0.70556640625
],
[
0.5556640625,
0.6923828125,
0.62744140625,
0.6923828125,
0.62744140625,
0.7041015625,
0.5556640625,
0.7041015625
],
[
0.1258544921875,
0.6962890625,
0.48193359375,
0.6962890625,
0.48193359375,
0.71044921875,
0.1258544921875,
0.71044921875
],
[
0.55224609375,
0.70263671875,
0.63134765625,
0.705078125,
0.63134765625,
0.7197265625,
0.55224609375,
0.71728515625
],
[
0.66845703125,
0.705078125,
0.708984375,
0.705078125,
0.708984375,
0.7197265625,
0.66845703125,
0.7197265625
],
[
0.7236328125,
0.705078125,
0.765625,
0.705078125,
0.765625,
0.7197265625,
0.7236328125,
0.7197265625
],
[
0.779296875,
0.705078125,
0.82177734375,
0.705078125,
0.82177734375,
0.7197265625,
0.779296875,
0.7197265625
],
[
0.8349609375,
0.70263671875,
0.87744140625,
0.705078125,
0.8759765625,
0.72021484375,
0.833984375,
0.71826171875
],
[
0.1258544921875,
0.71044921875,
0.373291015625,
0.71044921875,
0.373291015625,
0.7236328125,
0.1258544921875,
0.7236328125
],
[
0.5537109375,
0.71728515625,
0.61376953125,
0.7197265625,
0.61279296875,
0.734375,
0.55224609375,
0.73193359375
],
[
0.66650390625,
0.71826171875,
0.708984375,
0.71826171875,
0.708984375,
0.73291015625,
0.66650390625,
0.73291015625
],
[
0.7236328125,
0.71826171875,
0.7646484375,
0.71826171875,
0.7646484375,
0.734375,
0.7236328125,
0.734375
],
[
0.779296875,
0.71728515625,
0.8203125,
0.7197265625,
0.81884765625,
0.734375,
0.77783203125,
0.73193359375
],
[
0.83642578125,
0.71826171875,
0.87841796875,
0.71826171875,
0.87841796875,
0.734375,
0.83642578125,
0.734375
],
[
0.10455322265625,
0.72607421875,
0.4794921875,
0.72607421875,
0.4794921875,
0.736328125,
0.10455322265625,
0.736328125
],
[
0.55224609375,
0.73193359375,
0.640625,
0.734375,
0.640625,
0.74853515625,
0.55224609375,
0.7470703125
],
[
0.6689453125,
0.73486328125,
0.7060546875,
0.73486328125,
0.7060546875,
0.7470703125,
0.6689453125,
0.7470703125
],
[
0.7255859375,
0.73486328125,
0.7646484375,
0.73486328125,
0.7646484375,
0.7470703125,
0.7255859375,
0.7470703125
],
[
0.779296875,
0.73291015625,
0.82177734375,
0.73291015625,
0.82177734375,
0.74755859375,
0.779296875,
0.74755859375
],
[
0.83837890625,
0.73486328125,
0.87841796875,
0.73486328125,
0.87841796875,
0.7470703125,
0.83837890625,
0.7470703125
],
[
0.55224609375,
0.744140625,
0.6435546875,
0.7470703125,
0.6435546875,
0.76123046875,
0.55224609375,
0.7587890625
],
[
0.12335205078125,
0.7373046875,
0.48046875,
0.7373046875,
0.48046875,
0.75048828125,
0.12335205078125,
0.75048828125
],
[
0.7236328125,
0.744140625,
0.7646484375,
0.7470703125,
0.76318359375,
0.76220703125,
0.72216796875,
0.76025390625
],
[
0.8349609375,
0.744140625,
0.87744140625,
0.7470703125,
0.8759765625,
0.76220703125,
0.833984375,
0.76025390625
],
[
0.779296875,
0.7470703125,
0.82177734375,
0.7470703125,
0.82177734375,
0.76123046875,
0.779296875,
0.76123046875
],
[
0.1258544921875,
0.751953125,
0.48193359375,
0.751953125,
0.48193359375,
0.76513671875,
0.1258544921875,
0.76513671875
],
[
0.66845703125,
0.7470703125,
0.708984375,
0.7470703125,
0.708984375,
0.76123046875,
0.66845703125,
0.76123046875
],
[
0.7255859375,
0.76025390625,
0.76318359375,
0.76025390625,
0.76318359375,
0.77587890625,
0.7255859375,
0.77587890625
],
[
0.83642578125,
0.76025390625,
0.87841796875,
0.76025390625,
0.87841796875,
0.77587890625,
0.83642578125,
0.77587890625
],
[
0.66845703125,
0.76123046875,
0.708984375,
0.76123046875,
0.708984375,
0.77587890625,
0.66845703125,
0.77587890625
],
[
0.5556640625,
0.76123046875,
0.6435546875,
0.76123046875,
0.6435546875,
0.77294921875,
0.5556640625,
0.77294921875
],
[
0.78125,
0.76123046875,
0.82177734375,
0.76123046875,
0.82177734375,
0.77587890625,
0.78125,
0.77587890625
],
[
0.12744140625,
0.767578125,
0.48046875,
0.767578125,
0.48046875,
0.77783203125,
0.12744140625,
0.77783203125
],
[
0.1258544921875,
0.78076171875,
0.21728515625,
0.77880859375,
0.21728515625,
0.79052734375,
0.1258544921875,
0.79248046875
],
[
0.51708984375,
0.7763671875,
0.9140625,
0.7763671875,
0.9140625,
0.79150390625,
0.51708984375,
0.79150390625
],
[
0.10296630859375,
0.79345703125,
0.48193359375,
0.79345703125,
0.48193359375,
0.806640625,
0.10296630859375,
0.806640625
],
[
0.515625,
0.79150390625,
0.9111328125,
0.79052734375,
0.9111328125,
0.8037109375,
0.515625,
0.8046875
],
[
0.1258544921875,
0.806640625,
0.48046875,
0.806640625,
0.48046875,
0.82080078125,
0.1258544921875,
0.82080078125
],
[
0.1258544921875,
0.8212890625,
0.1871337890625,
0.8212890625,
0.1871337890625,
0.83349609375,
0.1258544921875,
0.83349609375
],
[
0.1282958984375,
0.849609375,
0.4150390625,
0.849609375,
0.4150390625,
0.86279296875,
0.1282958984375,
0.86279296875
],
[
0.5361328125,
0.8486328125,
0.91259765625,
0.849609375,
0.91259765625,
0.8642578125,
0.5361328125,
0.86279296875
],
[
0.51953125,
0.8515625,
0.54345703125,
0.8515625,
0.54345703125,
0.85986328125,
0.51953125,
0.85986328125
],
[
0.08660888671875,
0.85107421875,
0.13232421875,
0.85107421875,
0.13232421875,
0.8623046875,
0.08660888671875,
0.8623046875
],
[
0.08660888671875,
0.8671875,
0.4697265625,
0.8671875,
0.4697265625,
0.880859375,
0.08660888671875,
0.880859375
],
[
0.51708984375,
0.8671875,
0.9140625,
0.8671875,
0.9140625,
0.880859375,
0.51708984375,
0.880859375
],
[
0.0882568359375,
0.8818359375,
0.1748046875,
0.8818359375,
0.1748046875,
0.89306640625,
0.0882568359375,
0.89306640625
],
[
0.51953125,
0.88330078125,
0.7685546875,
0.88330078125,
0.7685546875,
0.89306640625,
0.51953125,
0.89306640625
]
] |
[
"。其中表示序列的最大长度。值得注意的是，LSTM的参数",
"3.1.1",
"Datasetand EvaluationMetric对于此实验，我们",
"本质上是“siamese\"，在文档和查询之间共享权重。关键思想",
"利用大型且众所周知的大型Ubuntu对话语料库（UDC）[15]",
"是LSTM编码器通过使用非线性变换作为门控函数来学习知",
"。我们使用Xu等人提供的相同测试拆分。[51]。在此任务中，",
"道顺序依赖性的表示。由于LSTM是标准的神经构建块，因",
"目标是将句子与其回复匹配。在[46]之后，该任务主要利用每",
"此我们省略了技术细节，以简洁。因此，将注意力作为该层之",
"个对话中的最后两个话语，预测后者是否遵循前者。训练集由",
"前的功能的关键思想是，它为LSTM编码器提供了提示，这",
"",
"百万个具有1：1正负比的消息-响应对组成。开发和测试集",
"些提示提供了诸如（1）长期和全局句子知识以及（2）句子对",
"具有9：1比率。在[46，51]之后，我们使用召回@k（Rn@K",
"（文档和查询）",
"）的评估指标，该指标指示n候选项的前k结果中是否存在",
"基本事实。使用的四个评估指标是R2@1、R10@1、R10@2和",
"2.4.1",
"PoolingOperation.最后，将池化函数应用于每个",
"R10@5",
"句子的隐藏状态h1..he），将序列转换为固定维表示。",
"3.1.2",
"Competitive Baselines and Implementation Details .",
"h = MeanMax[hi ... he]",
"(14)",
"我们与大量竞争基线进行比较，例如。g。，MLP，DeepMatch",
"[16] ,ARC-I/ARC-II [9] ,CNTN [21] ,MatchPyramid [19]",
"我们采用MeanMax池运算符，它将平均池和最大池的结果连",
"香草LSTM，注意力池LSTM[5]，MV-LSTM[37]，最后",
"接在一起。我们发现这始终比单独使用max或mean池更好",
"",
"是最先进的知识增强的混合神经网络（KEHN）。可以在[46]",
"上找到基线的详细说明。由于测试拆分是相同的，我们直接从",
"2.5Prediction Layer and Optimization",
"[46]报告结果。为了公平比较，我们将MCAN中的LSTM",
"最后，给定文档-查询对的固定维表示形式，我们将它们的串",
"编码器大小设置为d=100，这使其等于[46]中的模型。我们",
"联传递到两层h维高速公路网络中。我们模型的最终预测层",
"使用Adam优化器[13]优化MCAN，初始学习率为3×10-4",
"计算如下：",
"",
"将0.2的锻学率应用于除单词嵌人层以外的所有层。序列被",
"动态截断或填充到其批处理最大值（具有50令牌的硬限制）。",
"Yout = H2(Hi([xqixd;Xq xd;Xq -xal))",
"(15)",
"我们用预训练的GloVe嵌人初始化单词嵌入层",
"其中Hi(),H2(）是具有ReLU激活的公路网络层。然后将输",
"3.1.3",
"",
"ExperimentalResults。表1报告了我们的实验结",
"出传递到最终的线性softmax层。",
"果。显然，我们观察到所有MCAN模型都比现有的最先进",
"模型获得了巨大的性能提升。更具体地说，所有指标的改进",
"Ypred = softmax(WF yout + br)",
"(16)",
"都比KEHNN更好。相对于AP-LSTM和MV-LSTM等强",
"其中WFERhx2和bpR²。然后使用具有L2正则化的标准",
"基线的性能提升更大，在R10@1方面达到了15%的提升。这",
"确定了MCAN模型的有效性。总体而言，MCAN（FM）和",
"多类交叉损失来训练网络",
"MCAN（NN）在性能方面具有可比性。MCAN(SM）略低于",
"MCAN(FM）和MCAN(NN)。然而，它的性能仍然大大高于",
"N",
"Z[yi log9i + (1-yi) log(1-9i)] + lll2 (17)",
"现有的最先进的模型。",
"J(0) = -",
"E",
"R2@1 R10@1R10@2R10@5",
"其中θ是网络的参数。9是网络的输出。I10L2是L2正则化，",
"MLP",
"0.651",
"0.256",
"0.380",
"0.703",
"入是正则化的权重",
"0.593",
"0.345",
"0.376",
"0.693",
"0.665",
"DeepMatch",
"0.221",
"0.360",
"0.684",
"EMPIRICAL EVALUATION",
"ARC-I",
"3",
"ARC-II",
"0.736",
"0.380",
"0.534",
"0.777",
"在我们的实验中，我们旨在回答以下研究问题（RQs)：",
"0.743",
"0.349",
"0.512",
"0.797",
"0.743",
"CNTN",
"0.420",
"0.554",
"0.786",
"MatchPyramid",
"(1) RQ1 - Does our proposed approach achieve state-of-",
"0.725",
"0.361",
"0.494",
"0.801",
"LSTM",
"the-artperformanceon question answering and conver",
"0.758",
"0.381",
"0.545",
"0.801",
"AP-LSTM",
"sation modeling tasks? What is the relative improve-",
"MV-LSTM",
"0.767",
"0.410",
"0.565",
"0.800",
"ment over well-established baselines?",
"KEHNN",
"0.786",
"0.460",
"0.591",
"0.819",
"(2)RQ2-Whataretheimpactsofarchitecturaldesign",
"MCAN (SM)",
"0.831",
"0.548",
"0.682",
"0.873",
"MCAN (NN)",
"on performance? Is the LSTM encoder necessary to",
"0.549",
"0.875",
"0.686",
"make use of the casted features? Does all the varia-",
"0.833",
"0.551",
"0.875",
"0.834",
"MCAN (FM)",
"0.684",
"tions of co-attention contribute to the overall model",
"performance?",
"Table 1: Performance Comparison on Ubuntu Dialogue Cor-",
"(3) RQ3 -Can we explain the inner workings of our pro-",
"pus. Best result is in boldface and second best is underlined.",
"posed model? Can we interpret the casted attention",
"features?",
"Experiment 1-Dialogue Prediction",
"2Experiment2-FactoidQuestionAnswering",
"3.2",
"3.1",
"在第一个任务中，我们评估我们的模型成功预测对话中下",
"Factoid问题回答是回答基于事实的问题的任务。在此任务中，",
"个回复的能力",
"目标是提供给定问题的答案的排名列表"
] |
{
"arxiv_id": "1806.00778",
"language": "zh",
"page_idx": 3
}
|
||
2209.06995-zh-2
|
2209.06995-zh-2.jpg
|
据进行聚类，然后用预定义的启发法 (Müller et al., 2022) 贪婪地从每个聚类中选择样本，这样就无法控制所选样本之间的距离，从而不能充分提升样本的多样性。此外，在冷启动情况下，利用 PLM 的知识进行样品选择至关重要。虽然有几种方法利用预训练嵌入 (Hacohen et al., 2022; Chang et al., 2021)或掩码语言建模(MLM)损失 (Yuan et al.,2020) 来辅助数据选择，但预训练和微调任务之间的不匹配会损害其有效性。综上所述，用 PLMs 设计有效的冷启动数据选择方法仍然具有挑战性
我们提出了 PATRON $^{1}$,这是一种为 PLM 量身定制的基于提示的数据选择策略，以解决上述挑战。为了在冷启动设置下不访问任何标记数据的情况下估计模型不确定性，PATRON 利用提示符 (Gao et al.,2021a),该提示符将分类任务转换为具有自定义模板和版本控制器的完形填空式任务，通过预测[MASK]令牌的表面名称，为未标记数据生成任务感知伪标签。通过这种方式，我们还弥合了预训练和下行任务之间的差距，并从 PLM 中提取特定于任务的知识，以促进数据选择
但是，此类伪标签的一个重要问题是，即使在校准后，它们也可能不准确且有偏差$\overline{\mathrm{E}}($Zhao et al.,2021)。为了弥补这一点，我们进一步提出了不确定性传播，首先根据嵌入空间中的核相似性来测量样本之间的相关性，然后将其预测不确定性传播给其邻居。因此，仅当样本本身及其邻居的预测不确定性较高时，样本才会具有较高的传播不确定性，这表明该模型对于该样本周围的局部区域不太确定
为了选择一批不同的样本，我们超越了现有技术，提出了一种名为 partition-thenrewrite$\boxed{\mathrm{E}}($ PTR )的两阶段方法，该方法最初是针对组合优化$\overline{\mathrm{E}}($Chen and Tian, 2019) 提出的，用于动态调整每个簇内的选定样本。具体来说，我们首先使用 K-Means 聚类来划分未标记的数据，并从每个聚类中选择一个样本来初始化我们的解决方案。然后，我们基于$k$ -最近邻(kNN)构建邻居图，以对所选数据之间的邻域关系进行编码，并显式控制它们之间的距离。之后，我们添加一个额外的正则化项， 以防止每个聚类中的选定样本与其邻居聚类中的样本过于接近。我们将上述过程迭代了几轮，以逐步完善我们的解决方案，并促进数据选择的多样性
值得注意的是，PATRON 可以自然地应用于各种设置，包括香草微调，基于提示的学习，半监督学习和标准多轮 AL, 以提高 PLM 微调的
$^{1}\underline{\mathrm{P}}$ rompt-based d at a selection fo r few-sh o t PLM fi n
$\bar{\text{e-tuning.}}$
数据效率。我们总结了我们工作的主要贡献如下：(1) 冷启动数据选择范例 PATRON ,用于消除标签稀缺性问题以进行少量 PLM 微调；(2) 基于提示的不确定性传播方法来查询大多数信息样本；(3)分区然后重写(PTR)策略，通过在 AL 期间演示 3 个基本 XX 上的 THX 标签来平衡现有样本的多样性和信息性。(4)。$2\%-6。9\%$平均
### 2 $\textbf{Related Work}$
少拍语言模型微调。我们的方法与 NLP 中其他标签有效的学习范式密切相关，例如冷启动微调$\overline{\mathrm{E}}($Zhang et al., 2020b; Shnarch et al., 2022), 基于提示的学习$^{2}$ E(Gao et al., 2021a; Schick and Schütze, 202la,b; Min et al., 2022; Zhang et al., 2022c; Hu et al., 2022) , 半监督学习$\overline{\mathrm{E}}($Du et al., 2021; Wang et al., 2021b; Xie et al., 2020; Xu et al., 2023) 以及许多其他方法， 但是-与我们的方法正交-他们假设给出了一小部分标记数据并设计了更好的训练策略。相反， 我们的目标是从未标记的语料库中选择最有价值的实例，该实例与上述方法正交并可以与之结合，以进一步提高标签效率，如$\overline{\mathrm{E}}5.2$ 和 5.3节所示
训练数据选择。设计更好的策略来选择性地注释训练数据并不是一个新的研究课题。最重要的研究方向之一在于主动学习 (Zhang et al., 2020a; Schröder et al., 2022; Yu et al., 2022) , 它提高了深度 NLP 模型的标签效率。但是，它们中的大多数都需要大量干净的标签，以便在数据选择之前首先训练模型$\overline{\mathrm{E}}($Ash et al., 2020; Zhang and Plank, 2021)。不同的是，我们的目标是在没有给出初始标记数据的情况下，以最小的监督促进训练数据选择
这种冷启动数据选择的思想已应用于图像分类$\overline{\mathrm{E}}($Wang et al., 2021a; Hacohen et al., 2022) 和语音处理$\boxed{\mathrm{E}}($Park et al.,2022),但尚未针对NLP 域进行充分探索。对于此设置，Chang et al. (2021)专注于从嵌入空间中选择数据，但无法利用 PLM 中特定于任务的知识。Yuan et al (2020) 使用传销损失作为不确定性度量的代理，Liu et al. (2021a); Su et al. (2022) 研究亿级语言模型的少量样本选择$\overline{\mathrm{E}}($Brown et al., 2020) ,但主要集中在上下文学习上。与上述方法不同，我们的目标是利用提示来促进样本选择， 并设计额外的技术(i。e。、不确定性传播和PTR), $以 提 高 少 拍 PLM$ 微调的性能
$^2$In this work, we refer prompt-based learning to Fixed-
prompt PLM Tuning mentioned in (Liu et al., 2021b) .
|
[
[
0.12005615234375,
0.0885009765625,
0.486083984375,
0.0885009765625,
0.486083984375,
0.09857177734375,
0.12005615234375,
0.09857177734375
],
[
0.51318359375,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.51318359375,
0.1009521484375
],
[
0.118408203125,
0.10272216796875,
0.483642578125,
0.10272216796875,
0.483642578125,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.5146484375,
0.10272216796875,
0.87890625,
0.10272216796875,
0.87890625,
0.11639404296875,
0.5146484375,
0.11639404296875
],
[
0.118408203125,
0.12054443359375,
0.48193359375,
0.12054443359375,
0.48193359375,
0.1312255859375,
0.118408203125,
0.1312255859375
],
[
0.51318359375,
0.12054443359375,
0.87744140625,
0.12054443359375,
0.87744140625,
0.1312255859375,
0.51318359375,
0.1312255859375
],
[
0.118408203125,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.118408203125,
0.1484375
],
[
0.51123046875,
0.1341552734375,
0.87744140625,
0.1341552734375,
0.87744140625,
0.1478271484375,
0.51123046875,
0.1478271484375
],
[
0.118408203125,
0.1519775390625,
0.484375,
0.1519775390625,
0.484375,
0.1650390625,
0.118408203125,
0.1650390625
],
[
0.51318359375,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.1650390625,
0.51318359375,
0.1650390625
],
[
0.12005615234375,
0.16748046875,
0.484375,
0.16748046875,
0.484375,
0.18115234375,
0.12005615234375,
0.18115234375
],
[
0.51123046875,
0.166259765625,
0.87890625,
0.166259765625,
0.87890625,
0.179931640625,
0.51123046875,
0.179931640625
],
[
0.118408203125,
0.182861328125,
0.48779296875,
0.1817626953125,
0.48779296875,
0.196533203125,
0.118408203125,
0.19775390625
],
[
0.51318359375,
0.182861328125,
0.8818359375,
0.182861328125,
0.8818359375,
0.196533203125,
0.51318359375,
0.196533203125
],
[
0.118408203125,
0.198974609375,
0.48779296875,
0.198974609375,
0.48779296875,
0.2120361328125,
0.118408203125,
0.2120361328125
],
[
0.5087890625,
0.19775390625,
0.6396484375,
0.1995849609375,
0.6396484375,
0.213134765625,
0.5087890625,
0.21142578125
],
[
0.1175537109375,
0.215576171875,
0.483642578125,
0.215576171875,
0.483642578125,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.118408203125,
0.232177734375,
0.484375,
0.232177734375,
0.484375,
0.2452392578125,
0.118408203125,
0.2452392578125
],
[
0.50732421875,
0.2381591796875,
0.6689453125,
0.2392578125,
0.6689453125,
0.254150390625,
0.50732421875,
0.2529296875
],
[
0.118408203125,
0.2476806640625,
0.393798828125,
0.2476806640625,
0.393798828125,
0.26123046875,
0.118408203125,
0.26123046875
],
[
0.1351318359375,
0.26416015625,
0.486083984375,
0.26318359375,
0.486083984375,
0.27783203125,
0.1351318359375,
0.279052734375
],
[
0.51318359375,
0.270751953125,
0.87890625,
0.270751953125,
0.87890625,
0.283935546875,
0.51318359375,
0.283935546875
],
[
0.118408203125,
0.28076171875,
0.48779296875,
0.28076171875,
0.48779296875,
0.29443359375,
0.118408203125,
0.29443359375
],
[
0.51123046875,
0.2861328125,
0.880859375,
0.287353515625,
0.880859375,
0.3017578125,
0.51123046875,
0.300537109375
],
[
0.118408203125,
0.297607421875,
0.486083984375,
0.297607421875,
0.486083984375,
0.311279296875,
0.118408203125,
0.311279296875
],
[
0.5146484375,
0.303955078125,
0.880859375,
0.303955078125,
0.880859375,
0.317626953125,
0.5146484375,
0.317626953125
],
[
0.1175537109375,
0.3134765625,
0.486083984375,
0.312255859375,
0.486083984375,
0.32666015625,
0.1175537109375,
0.327880859375
],
[
0.509765625,
0.318359375,
0.87890625,
0.318359375,
0.87890625,
0.33203125,
0.509765625,
0.33203125
],
[
0.12005615234375,
0.330078125,
0.486083984375,
0.330078125,
0.486083984375,
0.34326171875,
0.12005615234375,
0.34326171875
],
[
0.51318359375,
0.336181640625,
0.880859375,
0.336181640625,
0.880859375,
0.349853515625,
0.51318359375,
0.349853515625
],
[
0.118408203125,
0.345703125,
0.486083984375,
0.345703125,
0.486083984375,
0.359375,
0.118408203125,
0.359375
],
[
0.51123046875,
0.352783203125,
0.880859375,
0.352783203125,
0.880859375,
0.366455078125,
0.51123046875,
0.366455078125
],
[
0.12005615234375,
0.363525390625,
0.483642578125,
0.363525390625,
0.483642578125,
0.37353515625,
0.12005615234375,
0.37353515625
],
[
0.51318359375,
0.3681640625,
0.880859375,
0.3681640625,
0.880859375,
0.3818359375,
0.51318359375,
0.3818359375
],
[
0.12005615234375,
0.377685546875,
0.486083984375,
0.377685546875,
0.486083984375,
0.391357421875,
0.12005615234375,
0.391357421875
],
[
0.51123046875,
0.383056640625,
0.880859375,
0.38427734375,
0.880859375,
0.3984375,
0.51123046875,
0.39794921875
],
[
0.118408203125,
0.394287109375,
0.48193359375,
0.394287109375,
0.48193359375,
0.405029296875,
0.118408203125,
0.405029296875
],
[
0.51123046875,
0.3984375,
0.880859375,
0.40087890625,
0.880859375,
0.4150390625,
0.51123046875,
0.413330078125
],
[
0.118408203125,
0.410400390625,
0.48779296875,
0.410400390625,
0.48779296875,
0.42333984375,
0.118408203125,
0.42333984375
],
[
0.5146484375,
0.416259765625,
0.880859375,
0.416259765625,
0.880859375,
0.429931640625,
0.5146484375,
0.429931640625
],
[
0.1175537109375,
0.42578125,
0.301513671875,
0.424560546875,
0.301513671875,
0.439453125,
0.1175537109375,
0.440673828125
],
[
0.51123046875,
0.432861328125,
0.87890625,
0.432861328125,
0.87890625,
0.446533203125,
0.51123046875,
0.446533203125
],
[
0.1368408203125,
0.443603515625,
0.48779296875,
0.443603515625,
0.48779296875,
0.457275390625,
0.1368408203125,
0.457275390625
],
[
0.51318359375,
0.448974609375,
0.87744140625,
0.448974609375,
0.87744140625,
0.4619140625,
0.51318359375,
0.4619140625
],
[
0.118408203125,
0.460205078125,
0.486083984375,
0.460205078125,
0.486083984375,
0.473876953125,
0.118408203125,
0.473876953125
],
[
0.509765625,
0.46435546875,
0.880859375,
0.463134765625,
0.880859375,
0.47802734375,
0.509765625,
0.478515625
],
[
0.1175537109375,
0.4755859375,
0.486083984375,
0.474365234375,
0.486083984375,
0.4892578125,
0.1175537109375,
0.490478515625
],
[
0.51318359375,
0.482177734375,
0.56689453125,
0.482177734375,
0.56689453125,
0.493408203125,
0.51318359375,
0.493408203125
],
[
0.118408203125,
0.4921875,
0.484375,
0.4921875,
0.484375,
0.505859375,
0.118408203125,
0.505859375
],
[
0.51123046875,
0.50048828125,
0.87890625,
0.50048828125,
0.87890625,
0.51416015625,
0.51123046875,
0.51416015625
],
[
0.118408203125,
0.50830078125,
0.484375,
0.50830078125,
0.484375,
0.521484375,
0.118408203125,
0.521484375
],
[
0.51318359375,
0.51708984375,
0.880859375,
0.51708984375,
0.880859375,
0.53076171875,
0.51318359375,
0.53076171875
],
[
0.118408203125,
0.52392578125,
0.486083984375,
0.52392578125,
0.486083984375,
0.53759765625,
0.118408203125,
0.53759765625
],
[
0.51318359375,
0.533203125,
0.880859375,
0.533203125,
0.880859375,
0.54638671875,
0.51318359375,
0.54638671875
],
[
0.1175537109375,
0.5390625,
0.484375,
0.5390625,
0.484375,
0.552734375,
0.1175537109375,
0.552734375
],
[
0.51123046875,
0.548828125,
0.880859375,
0.548828125,
0.880859375,
0.5625,
0.51123046875,
0.5625
],
[
0.118408203125,
0.5556640625,
0.484375,
0.5556640625,
0.484375,
0.5693359375,
0.118408203125,
0.5693359375
],
[
0.51318359375,
0.5654296875,
0.8818359375,
0.5654296875,
0.8818359375,
0.5791015625,
0.51318359375,
0.5791015625
],
[
0.118408203125,
0.572265625,
0.376220703125,
0.572265625,
0.376220703125,
0.5859375,
0.118408203125,
0.5859375
],
[
0.5146484375,
0.58056640625,
0.880859375,
0.58056640625,
0.880859375,
0.59423828125,
0.5146484375,
0.59423828125
],
[
0.1368408203125,
0.59033203125,
0.484375,
0.59033203125,
0.484375,
0.60400390625,
0.1368408203125,
0.60400390625
],
[
0.51123046875,
0.5966796875,
0.87890625,
0.5966796875,
0.87890625,
0.60986328125,
0.51123046875,
0.60986328125
],
[
0.1175537109375,
0.60546875,
0.489501953125,
0.60498046875,
0.489501953125,
0.619140625,
0.1175537109375,
0.62060546875
],
[
0.509765625,
0.6123046875,
0.87744140625,
0.61328125,
0.87744140625,
0.62744140625,
0.509765625,
0.626953125
],
[
0.1175537109375,
0.62060546875,
0.486083984375,
0.62158203125,
0.486083984375,
0.6357421875,
0.1175537109375,
0.63525390625
],
[
0.51318359375,
0.6298828125,
0.8818359375,
0.6298828125,
0.8818359375,
0.6435546875,
0.51318359375,
0.6435546875
],
[
0.118408203125,
0.63818359375,
0.486083984375,
0.63818359375,
0.486083984375,
0.65185546875,
0.118408203125,
0.65185546875
],
[
0.51318359375,
0.6455078125,
0.73486328125,
0.6455078125,
0.73486328125,
0.6591796875,
0.51318359375,
0.6591796875
],
[
0.118408203125,
0.65380859375,
0.486083984375,
0.65283203125,
0.486083984375,
0.66748046875,
0.118408203125,
0.66845703125
],
[
0.5283203125,
0.66455078125,
0.87890625,
0.66455078125,
0.87890625,
0.677734375,
0.5283203125,
0.677734375
],
[
0.118408203125,
0.67041015625,
0.484375,
0.67041015625,
0.484375,
0.68408203125,
0.118408203125,
0.68408203125
],
[
0.12005615234375,
0.68603515625,
0.484375,
0.68603515625,
0.484375,
0.69970703125,
0.12005615234375,
0.69970703125
],
[
0.51318359375,
0.68115234375,
0.880859375,
0.68115234375,
0.880859375,
0.6953125,
0.51318359375,
0.6953125
],
[
0.51123046875,
0.69677734375,
0.87890625,
0.69677734375,
0.87890625,
0.71044921875,
0.51123046875,
0.71044921875
],
[
0.118408203125,
0.703125,
0.484375,
0.703125,
0.484375,
0.71630859375,
0.118408203125,
0.71630859375
],
[
0.118408203125,
0.71875,
0.484375,
0.71875,
0.484375,
0.732421875,
0.118408203125,
0.732421875
],
[
0.51318359375,
0.71435546875,
0.87890625,
0.71435546875,
0.87890625,
0.72802734375,
0.51318359375,
0.72802734375
],
[
0.118408203125,
0.7353515625,
0.484375,
0.7353515625,
0.484375,
0.7490234375,
0.118408203125,
0.7490234375
],
[
0.51318359375,
0.72998046875,
0.880859375,
0.72998046875,
0.880859375,
0.74365234375,
0.51318359375,
0.74365234375
],
[
0.5146484375,
0.7451171875,
0.880859375,
0.7451171875,
0.880859375,
0.7587890625,
0.5146484375,
0.7587890625
],
[
0.118408203125,
0.751953125,
0.484375,
0.751953125,
0.484375,
0.765625,
0.118408203125,
0.765625
],
[
0.118408203125,
0.76708984375,
0.483642578125,
0.76708984375,
0.483642578125,
0.77783203125,
0.118408203125,
0.77783203125
],
[
0.51123046875,
0.7626953125,
0.880859375,
0.7626953125,
0.880859375,
0.775390625,
0.51123046875,
0.775390625
],
[
0.51123046875,
0.77685546875,
0.87890625,
0.77685546875,
0.87890625,
0.79052734375,
0.51123046875,
0.79052734375
],
[
0.12005615234375,
0.783203125,
0.48779296875,
0.783203125,
0.48779296875,
0.79638671875,
0.12005615234375,
0.79638671875
],
[
0.1175537109375,
0.798828125,
0.486083984375,
0.798828125,
0.486083984375,
0.8125,
0.1175537109375,
0.8125
],
[
0.51123046875,
0.79443359375,
0.880859375,
0.79443359375,
0.880859375,
0.80810546875,
0.51123046875,
0.80810546875
],
[
0.53076171875,
0.81005859375,
0.880859375,
0.81005859375,
0.880859375,
0.82373046875,
0.53076171875,
0.82373046875
],
[
0.12005615234375,
0.81640625,
0.2275390625,
0.81640625,
0.2275390625,
0.82763671875,
0.12005615234375,
0.82763671875
],
[
0.51318359375,
0.82666015625,
0.87890625,
0.82666015625,
0.87890625,
0.84033203125,
0.51318359375,
0.84033203125
],
[
0.1368408203125,
0.8330078125,
0.486083984375,
0.8330078125,
0.486083984375,
0.8466796875,
0.1368408203125,
0.8466796875
],
[
0.51318359375,
0.8427734375,
0.880859375,
0.8427734375,
0.880859375,
0.85546875,
0.51318359375,
0.85546875
],
[
0.118408203125,
0.8486328125,
0.48779296875,
0.8486328125,
0.48779296875,
0.8623046875,
0.118408203125,
0.8623046875
],
[
0.1175537109375,
0.8642578125,
0.483642578125,
0.8642578125,
0.483642578125,
0.87744140625,
0.1175537109375,
0.87744140625
],
[
0.51123046875,
0.859375,
0.80029296875,
0.859375,
0.80029296875,
0.87255859375,
0.51123046875,
0.87255859375
],
[
0.139404296875,
0.892578125,
0.48779296875,
0.8935546875,
0.48779296875,
0.90771484375,
0.139404296875,
0.90673828125
],
[
0.5322265625,
0.892578125,
0.87890625,
0.8935546875,
0.87890625,
0.90673828125,
0.5322265625,
0.90625
],
[
0.118408203125,
0.90625,
0.17041015625,
0.9091796875,
0.1695556640625,
0.92041015625,
0.1175537109375,
0.91748046875
],
[
0.51318359375,
0.90625,
0.84033203125,
0.90625,
0.84033203125,
0.91943359375,
0.51318359375,
0.91943359375
]
] |
[
"据进行聚类，然后用预定义的启发法（Miller",
"数据效率。我们总结了我们工作的主要贡献如",
"etal.,2022)贪婪地从每个聚类中选择样本，这",
"下：（1）冷启动数据选择范例PATRON，用于消",
"样就无法控制所选样本之间的距离，从而不",
"除标签稀缺性问题以进行少量PLM微调：（2",
"能充分提升样本的多样性。此外，在冷启动情",
"基于提示的不确定性传播方法来查询大多数信",
"况下，利用PLM的知识进行样品选择至关重",
"息样本；（3）分区然后重写（PTR）策略，通",
"要。虽然有几种方法利用预训练嵌人（Hacohen",
"过在AL期间演示3个基本XX上的THX标",
"etal.，2022；Changetal.，2021）或掩码语言建模",
"签来平衡现有样本的多样性和信息性。（4）。",
"(MLM)损失（Yuanetal.,2020）来辅助数据选",
"2%-6。9%平均",
"择，但预训练和微调任务之间的不匹配会损害",
"其有效性。综上所述，用PLMs设计有效的冷",
"2Related Work",
"启动数据选择方法仍然具有挑战性",
"我们提出了PATRON！，这是一种为PLM量",
"少拍语言模型微调。我们的方法与NLP中",
"身定制的基于提示的数据选择策略，以解决上",
"其他标签有效的学习范式密切相关，例如冷",
"述挑战。为了在冷启动设置下不访问任何标记",
"启动微调E(Zhangetal.,2020b;Shnarchetal.,",
"数据的情况下估计模型不确定性，PATRON利",
"2022），基于提示的学习2E(Ga0etal.2021a",
"用提示符（Gaoetal.,2021a），该提示符将分类",
"Schick and Schuitze, 202la,b; Min et al., 2022;",
"任务转换为具有自定义模板和版本控制器的完",
"Zhang et al.,2022c;Hu et al.,2022)，半监督",
"形填空式任务，通过预测［MASKI今牌的表面",
"学习E(Du et al., 2021; Wang et al., 2021b; Xie",
"名称，为未标记数据生成任务感知伪标签。通",
"etal.,2020;Xuetal.,2023）以及许多其他方法",
"过这种方式：我们还弥合了预训练和下行任务",
"但是-与我们的方法正交-他们假设给出了一小",
"之间的差距，并从PLM中提取特定于任务的",
"部分标记数据并设计了更好的训练策略。相反，",
"知识，以促进数据选择",
"我们的目标是从未标记的语料库中选择最有价",
"但是，此类伪标签的一个重要问题是，即使",
"值的实例，该实例与上述方法正交并可以与之",
"在校准后，它们也可能不准确且有偏差E(Zhao",
"结合，以进一步提高标签效率，如E5.2和5.3",
"etal,2021）。为了弥补这一点，我们进一步提",
"节所示",
"出了不确定性传播，首先根据嵌入空间中的核",
"训练数据选择。设计更好的策略来选择性地",
"相似性来测量样本之间的相关性，然后将其预",
"注释训练数据并不是一个新的研究课题。最重",
"测不确定性传播给其邻居。因此，仅当样本本",
"要的研究方向之一在于主动学习(Zhangetal.",
"身及其邻居的预测不确定性较高时，样本才会",
"2020a; Schroder et al., 2022; Yu et al., 2022), 它",
"具有较高的传播不确定性，这表明该模型对于",
"提高了深度NLP模型的标签效率。但是，它",
"该样本周围的局部区域不太确定",
"们中的大多数都需要大量干净的标签，以便在",
"为了选择一批不同的样本，我们超越了",
"数据选择之前首先训练模型E(Ashetal.,2020",
"现有技术，提出了一种名为partition-then-",
"ZhangandPlank，2021）。不同的是，我们的目",
"rewriteE（PTR）的两阶段方法，该方法最",
"标是在没有给出初始标记数据的情况下，以最",
"初是针对组合优化E(ChenandTian，2019）提",
"小的监督促进训练数据选择",
"出的，用于动态调整每个簇内的选定样本。具",
"这种冷启动数据选择的思想已应用于图像",
"体来说，我们首先使用K-Means聚类来划分未",
"标记的数据，并从每个聚类中选择一个样本来",
"分类E(Wang et al., 2021a; Hacohen et al., 2022)",
"和语音处理E(Parketal，2022），但尚未针对",
"初始化我们的解决方案。然后，我们基于k-最",
"近邻（kNN）构建邻居图，以对所选数据之间",
"NLP域进行充分探索。对于此设置，Changetal.",
"的邻域关系进行编码，并显式控制它们之间的",
"(2021）专注于从嵌入空间中选择数据，但无法",
"利用PLM中特定于任务的知识。Yuan etal.",
"距离。之后，我们添加一个额外的正则化项，",
"以防止每个聚类中的选定样本与其邻居聚类",
"(2020）使用传销损失作为不确定性度量的代",
"理，Liu et al.(202la); Su et al.(2022)研究亿级",
"中的样本过于接近。我们将上述过程迭代了几",
"轮，以逐步完善我们的解决方案，并促进数据",
"语言模型的少量样本选择E(Brownetal,2020)",
"但主要集中在上下文学习上。与上述方法不",
"选择的多样性",
"同，我们的目标是利用提示来促进样本选择，",
"值得注意的是，PATRON可以自然地应用于",
"并设计额外的技术(i。e。、不确定性传播和",
"各种设置，包括香草微调，基于提示的学习，半",
"监督学习和标准多轮AL，以提高PLM微调的",
"PTR)，以提高少拍PLM微调的性能",
"'P rompt-based d at a selection fo r few-sh o t PLM fi n",
"2In this work, we refer prompt-based learning to Fixed",
"e-tuning",
"prompt PLM Tuning mentioned in (Liu et al., 2021b)"
] |
{
"arxiv_id": "2209.06995",
"language": "zh",
"page_idx": 2
}
|
||
2209.06995-zh-3
|
2209.06995-zh-3.jpg
|
# 3 Backgrounds 3.1 Problem Formulation
表示为每个有限的预算 XMAXXMAX $\mathcal{D}_u=$ $\{x_j\}_{j=1}^U$。接下来，获取的样本被标记并从$\mathcal{D}_u$ 移动到$\mathcal{D}_l$ 。然后，我们使用$\mathcal{D}_l$对预训练的语言模型$f(\cdot;\theta)$进行微调，以最大程度地提高下游分类任务的性能。上述步骤可以是单轮(Chang et al., 2021; Hacohen et al., 2022)(在本例中为$b=|B|$ ),也可以重复多轮(Yuan et al., 2020)( $b=|B|/|$Rounds| ), 直到达到预算$|B|$
3.2 Prompt-based Learning for PLMs
已经提出了提示方法，通过应用完形填空式任务来微调 PLM (Schick and Schütze, 2021a,b) 来弥合预训练和微调阶段之间的差距。形式上， 提示中有两个关键组件：预定义的模板$\tau$和版本控制器$\mathcal{V}$。对于每个输人示例$x$ ,它将被包含一段自然语言文本以及[MASK]令牌的模板包裹，然后再送人 PLM $\mathcal{M}$。然后，使用语言转换器$\mathcal{V}$将任务标签$y$映射到词汇表中的单个单词$\mathcal{V}(y)$。以二元情感分类为例，对于输入句子$x$ ,模板 $\mathcal{T}$可以是 $\underline\mathcal{T}(x)=[x\:.$ It was [MASK]. ] , 而正面和负面情感的语言器可以分别是“good ”和“terrible”
使用模板和语言转换器，我们可以通过掩码语言建模(MLM)计算标签集$\mathcal{Y}$上的概率分布，如下所示
$$\begin{aligned}p\left(y\mid x\right)&=p\left([\mathsf{MASK}]=\mathcal{V}(y)\mid\mathcal{T}(x)\right)\\&=\frac{\exp\left(\boldsymbol{w}_{\mathcal{V}(y)}^{T}\boldsymbol{h}_{[\mathsf{MASK}]}\right)}{\sum_{y^{\prime}\in\mathcal{Y}}\exp\left(\boldsymbol{w}_{\mathcal{V}(y^{\prime})}^{T}\boldsymbol{h}_{[\mathsf{MASK}]}\right)}\end{aligned}$$
(1)
其中$h_\text{[MASK] 是[MASK] 标记的隐藏嵌入，}w_{\mathcal{V}(y)}$ 表示来自$\mathcal{M}$的标签字$\mathcal{V}(y)$的嵌入。由于这些令牌的嵌入已在具有 MLM 目标的预训练期间进行了优化，因此提示的使用缩小了预训练和微调之间的差距。换句话说，在使 PLM 适应新任务时，提示可作为先验知识的来源
# 4 Methodology
在本节中，我们介绍了利用提示进行冷启动数据选择的方法 PATRON 。我们首先介绍如何利用提示进行冷启动场景下的不确定性估计。利用估计的不确定性，我们提出了两个关键设计，即不确定性传播和分区重写(PTR)策略， 以平衡样本选择的信息量和多样性。
4.1 Uncertainty Estimation with Prompts
我们首先描述如何估计未标记数据的不确定性以促进 PATRON 。给定没有标记数据的预训练语言模型(PLM) $\mathcal{M}$ ,我们利用提示生成伪标
签$^{3}$以进行不确定性估计。根据公式$\mathbb{E}$1,我们能够根据[MASK]令牌的预测，获得每个样本$x$上不同标签词的出现概率
但是，直接采用此概率可能会出现问题，因为PLM 会遇到错误校准问题$\overline{\mathrm{E}}($Zhao et al., 2021; Hu et al., 2022) , $i.e.$,标签字可能会出现不同的频率，从而使其中一些不太可能被预测。因此，方程 1 中的预测和估计的不确定性可能是有偏差的
意识到这一点，我们采用(Hu et al.,2022) 中的方法来计算标签词的上下文先验。我们首先为每个类$i$选择具有最高$p(y_i|x)$熵的$k$样本作为
$$\mathcal{S}=\bigcup_{i\in\{1,2,\ldots,c\}}\mathop{\mathrm{Top-k}}_{x\in\mathcal{D}_{u}}p(y_{i}|x).$$
(2)
来构造支持集$\mathcal{S}$,然后，将上下文化的先验近
似为
$$P(v)\approx\frac{1}{|\mathcal{S}|}\sum_{x\in\mathcal{S}}P_{\mathcal{M}}\left(\text{[MASK]}=v\mid x\right),$$
(3)
# $\begin{array}{cc}4.2&\text{Uncertainty Propagation for Data Utility}\\&\textbf{Estimation}\end{array}$
尽管我们已经减轻了基于提示的伪标签的偏差，但由于在零拍设置下监管不足，此类伪标签仍然可能不准确。在这种情况下，直接使用方程$\mathbb{E}??$ 中的不确定性进行样本选择会产生次优结果，因为它可能对异常值敏感，异常值自然具有较大的模型不确定性，但对模型学习$\overline{\mathrm{E}}($Karamcheti et al., 2021)的益处较小
为了解决这个问题，我们利用嵌人空间中的核相似性来测量数据点之间的相关性并传播模型不确定性：对于每个数据点$x$ ,我们首先使用SimCSEE(Gao et al., 2021b) $^4$作为$z=g(x;\theta)$ 来计算其嵌人，并基于其计算 $K$ -近邻。然后， 我们选择径向基函数 (RBF) (Scholkopf et al., 1997)作为两个数据点$x_i$和$x_j$的相似性度量， 表示为
$$\kappa\left(x_{i},x_{j}\right)=\exp\left(-\rho\left\|\boldsymbol{z}_{i}-\boldsymbol{z}_{j}\right\|_{2}^{2}\right),$$
,其中$z_i$是来自 SimCSE 的$x_i$的嵌入$,\rho$是控
(4)
制权重传播的超参数。形式上，$x$的传播不确
$$\hat{u}_{\mathrm{prop}}(x)=u(x)+\frac{\sum_{x_{i}\in\mathcal{X}_{\mathrm{KNN}}(x)}\kappa(x,x_{i})\cdot u(x_{i})}{|\mathcal{X}_{\mathrm{KNN}}(x)|}.$$
我们强调，只有当样本对其自身及其邻居具有较高的不确定性时，才会导致较高的传播不确定性，这表明 PLM 不确定样本周围的周围
$^{3}$In this study, we use the manual prompts and verbalizers from existing works (Hu et al., 2022; Schick and Schütze, 2021a) due to their simplicity and competitive performance.
$^{4}$Notably, we use the version of princeton-nlp/
unsup- simcse- roberta- base as the encoder
|
[
[
0.11419677734375,
0.08612060546875,
0.267822265625,
0.08612060546875,
0.267822265625,
0.1009521484375,
0.11419677734375,
0.1009521484375
],
[
0.51318359375,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.51318359375,
0.09979248046875
],
[
0.1175537109375,
0.10809326171875,
0.334228515625,
0.10809326171875,
0.334228515625,
0.1217041015625,
0.1175537109375,
0.1217041015625
],
[
0.51318359375,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.51318359375,
0.1181640625,
0.755859375,
0.1181640625,
0.755859375,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.1175537109375,
0.1300048828125,
0.478515625,
0.127685546875,
0.478515625,
0.141357421875,
0.1175537109375,
0.1436767578125
],
[
0.5322265625,
0.1353759765625,
0.880859375,
0.1353759765625,
0.880859375,
0.1484375,
0.5322265625,
0.1484375
],
[
0.1175537109375,
0.1455078125,
0.484375,
0.1436767578125,
0.484375,
0.1602783203125,
0.1175537109375,
0.162109375
],
[
0.51123046875,
0.1519775390625,
0.87744140625,
0.1519775390625,
0.87744140625,
0.162109375,
0.51123046875,
0.162109375
],
[
0.12005615234375,
0.163330078125,
0.483642578125,
0.163330078125,
0.483642578125,
0.17333984375,
0.12005615234375,
0.17333984375
],
[
0.51318359375,
0.16748046875,
0.87890625,
0.16748046875,
0.87890625,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.1175537109375,
0.177001953125,
0.483642578125,
0.177001953125,
0.483642578125,
0.190673828125,
0.1175537109375,
0.190673828125
],
[
0.51123046875,
0.1817626953125,
0.87744140625,
0.1817626953125,
0.87744140625,
0.1953125,
0.51123046875,
0.1953125
],
[
0.118408203125,
0.19482421875,
0.486083984375,
0.19482421875,
0.486083984375,
0.2078857421875,
0.118408203125,
0.2078857421875
],
[
0.51318359375,
0.1995849609375,
0.87890625,
0.1995849609375,
0.87890625,
0.213134765625,
0.51318359375,
0.213134765625
],
[
0.1175537109375,
0.210205078125,
0.48779296875,
0.208984375,
0.48779296875,
0.223876953125,
0.1175537109375,
0.2244873046875
],
[
0.51123046875,
0.2161865234375,
0.58447265625,
0.2161865234375,
0.58447265625,
0.22802734375,
0.51123046875,
0.22802734375
],
[
0.118408203125,
0.226806640625,
0.486083984375,
0.226806640625,
0.486083984375,
0.240478515625,
0.118408203125,
0.240478515625
],
[
0.53076171875,
0.232177734375,
0.87890625,
0.232177734375,
0.87890625,
0.2452392578125,
0.53076171875,
0.2452392578125
],
[
0.1158447265625,
0.2410888671875,
0.48779296875,
0.2423095703125,
0.48779296875,
0.257080078125,
0.1158447265625,
0.255859375
],
[
0.51318359375,
0.2476806640625,
0.8818359375,
0.2476806640625,
0.8818359375,
0.26123046875,
0.51318359375,
0.26123046875
],
[
0.118408203125,
0.260009765625,
0.166259765625,
0.260009765625,
0.166259765625,
0.271484375,
0.118408203125,
0.271484375
],
[
0.5146484375,
0.26416015625,
0.8818359375,
0.26416015625,
0.8818359375,
0.279052734375,
0.5146484375,
0.279052734375
],
[
0.1158447265625,
0.2861328125,
0.432373046875,
0.2861328125,
0.432373046875,
0.2998046875,
0.1158447265625,
0.2998046875
],
[
0.51318359375,
0.281982421875,
0.5517578125,
0.281982421875,
0.5517578125,
0.29443359375,
0.51318359375,
0.29443359375
],
[
0.580078125,
0.29443359375,
0.6962890625,
0.29443359375,
0.6962890625,
0.312255859375,
0.580078125,
0.312255859375
],
[
0.68603515625,
0.29443359375,
0.81005859375,
0.297607421875,
0.81005859375,
0.312255859375,
0.68603515625,
0.308837890625
],
[
0.85791015625,
0.29638671875,
0.8818359375,
0.29638671875,
0.8818359375,
0.311279296875,
0.85791015625,
0.311279296875
],
[
0.118408203125,
0.30712890625,
0.484375,
0.30712890625,
0.484375,
0.320556640625,
0.118408203125,
0.320556640625
],
[
0.615234375,
0.3134765625,
0.74072265625,
0.30810546875,
0.7421875,
0.32470703125,
0.6171875,
0.330078125
],
[
0.1175537109375,
0.322509765625,
0.484375,
0.322509765625,
0.484375,
0.336181640625,
0.1175537109375,
0.336181640625
],
[
0.51123046875,
0.32666015625,
0.880859375,
0.327880859375,
0.880859375,
0.342529296875,
0.51123046875,
0.341552734375
],
[
0.12005615234375,
0.34033203125,
0.484375,
0.34033203125,
0.484375,
0.350830078125,
0.12005615234375,
0.350830078125
],
[
0.51318359375,
0.345703125,
0.5517578125,
0.345703125,
0.5517578125,
0.358154296875,
0.51318359375,
0.358154296875
],
[
0.118408203125,
0.355712890625,
0.484375,
0.355712890625,
0.484375,
0.369384765625,
0.118408203125,
0.369384765625
],
[
0.611328125,
0.35693359375,
0.640625,
0.3623046875,
0.63623046875,
0.37353515625,
0.60693359375,
0.3681640625
],
[
0.54150390625,
0.364013671875,
0.60205078125,
0.366455078125,
0.60107421875,
0.38134765625,
0.5400390625,
0.37890625
],
[
0.62744140625,
0.359375,
0.86376953125,
0.364013671875,
0.86376953125,
0.385498046875,
0.62744140625,
0.380126953125
],
[
0.603515625,
0.366455078125,
0.66748046875,
0.377685546875,
0.66162109375,
0.3984375,
0.59765625,
0.38720703125
],
[
0.85400390625,
0.36767578125,
0.87890625,
0.36767578125,
0.87890625,
0.37890625,
0.85400390625,
0.37890625
],
[
0.118408203125,
0.38720703125,
0.484375,
0.38720703125,
0.484375,
0.40087890625,
0.118408203125,
0.40087890625
],
[
0.1175537109375,
0.37060546875,
0.483642578125,
0.37060546875,
0.483642578125,
0.38427734375,
0.1175537109375,
0.38427734375
],
[
0.118408203125,
0.402099609375,
0.484375,
0.402587890625,
0.484375,
0.41748046875,
0.118408203125,
0.416259765625
],
[
0.51318359375,
0.40625,
0.54150390625,
0.40625,
0.54150390625,
0.41748046875,
0.51318359375,
0.41748046875
],
[
0.55517578125,
0.40380859375,
0.87744140625,
0.405029296875,
0.87744140625,
0.419189453125,
0.55517578125,
0.418701171875
],
[
0.5400390625,
0.410400390625,
0.55322265625,
0.410400390625,
0.55322265625,
0.413330078125,
0.5400390625,
0.413330078125
],
[
0.118408203125,
0.419189453125,
0.484375,
0.419189453125,
0.484375,
0.432861328125,
0.118408203125,
0.432861328125
],
[
0.55517578125,
0.4228515625,
0.6396484375,
0.4228515625,
0.6396484375,
0.43408203125,
0.55517578125,
0.43408203125
],
[
0.118408203125,
0.435791015625,
0.484375,
0.435791015625,
0.484375,
0.449462890625,
0.118408203125,
0.449462890625
],
[
0.51318359375,
0.438232421875,
0.880859375,
0.439453125,
0.880859375,
0.45361328125,
0.51318359375,
0.453125
],
[
0.1175537109375,
0.451904296875,
0.48779296875,
0.451904296875,
0.48779296875,
0.465576171875,
0.1175537109375,
0.465576171875
],
[
0.51123046875,
0.454833984375,
0.880859375,
0.4560546875,
0.880859375,
0.47021484375,
0.51123046875,
0.4697265625
],
[
0.12005615234375,
0.4697265625,
0.484375,
0.4697265625,
0.484375,
0.479736328125,
0.12005615234375,
0.479736328125
],
[
0.51318359375,
0.47265625,
0.880859375,
0.47265625,
0.880859375,
0.486328125,
0.51318359375,
0.486328125
],
[
0.12005615234375,
0.48388671875,
0.258544921875,
0.48388671875,
0.258544921875,
0.49755859375,
0.12005615234375,
0.49755859375
],
[
0.51318359375,
0.48681640625,
0.880859375,
0.488037109375,
0.880859375,
0.5029296875,
0.51318359375,
0.501953125
],
[
0.1368408203125,
0.50048828125,
0.486083984375,
0.50048828125,
0.486083984375,
0.51416015625,
0.1368408203125,
0.51416015625
],
[
0.51318359375,
0.5048828125,
0.880859375,
0.5048828125,
0.880859375,
0.5185546875,
0.51318359375,
0.5185546875
],
[
0.1175537109375,
0.5166015625,
0.484375,
0.515625,
0.484375,
0.52978515625,
0.1175537109375,
0.53076171875
],
[
0.515625,
0.52099609375,
0.880859375,
0.52099609375,
0.880859375,
0.53369140625,
0.515625,
0.53369140625
],
[
0.12005615234375,
0.53369140625,
0.2275390625,
0.53369140625,
0.2275390625,
0.5458984375,
0.12005615234375,
0.5458984375
],
[
0.5146484375,
0.5361328125,
0.82958984375,
0.5361328125,
0.82958984375,
0.5498046875,
0.5146484375,
0.5498046875
],
[
0.173828125,
0.5556640625,
0.4130859375,
0.5556640625,
0.4130859375,
0.5693359375,
0.173828125,
0.5693359375
],
[
0.5322265625,
0.552734375,
0.87890625,
0.552734375,
0.87890625,
0.56640625,
0.5322265625,
0.56640625
],
[
0.276123046875,
0.57373046875,
0.39794921875,
0.57373046875,
0.39794921875,
0.58837890625,
0.276123046875,
0.58837890625
],
[
0.51318359375,
0.5693359375,
0.880859375,
0.5693359375,
0.880859375,
0.5830078125,
0.51318359375,
0.5830078125
],
[
0.466796875,
0.57763671875,
0.486083984375,
0.57763671875,
0.486083984375,
0.58740234375,
0.466796875,
0.58740234375
],
[
0.51318359375,
0.58447265625,
0.87744140625,
0.58447265625,
0.87744140625,
0.59716796875,
0.51318359375,
0.59716796875
],
[
0.2476806640625,
0.59326171875,
0.4248046875,
0.59326171875,
0.4248046875,
0.60791015625,
0.2476806640625,
0.60791015625
],
[
0.51123046875,
0.599609375,
0.880859375,
0.60107421875,
0.880859375,
0.615234375,
0.51123046875,
0.6142578125
],
[
0.1158447265625,
0.61767578125,
0.486083984375,
0.6181640625,
0.486083984375,
0.6357421875,
0.1158447265625,
0.63525390625
],
[
0.51318359375,
0.6162109375,
0.87744140625,
0.61767578125,
0.87744140625,
0.6318359375,
0.51318359375,
0.63134765625
],
[
0.1175537109375,
0.63525390625,
0.484375,
0.6328125,
0.484375,
0.64794921875,
0.1175537109375,
0.6494140625
],
[
0.51318359375,
0.63427734375,
0.880859375,
0.63427734375,
0.880859375,
0.64794921875,
0.51318359375,
0.64794921875
],
[
0.12176513671875,
0.65283203125,
0.483642578125,
0.65283203125,
0.483642578125,
0.6630859375,
0.12176513671875,
0.6630859375
],
[
0.5146484375,
0.6494140625,
0.880859375,
0.6494140625,
0.880859375,
0.6630859375,
0.5146484375,
0.6630859375
],
[
0.118408203125,
0.66748046875,
0.484375,
0.66748046875,
0.484375,
0.68115234375,
0.118408203125,
0.68115234375
],
[
0.51318359375,
0.66650390625,
0.5693359375,
0.66650390625,
0.5693359375,
0.68115234375,
0.51318359375,
0.68115234375
],
[
0.118408203125,
0.68310546875,
0.484375,
0.68310546875,
0.484375,
0.69677734375,
0.118408203125,
0.69677734375
],
[
0.56494140625,
0.68408203125,
0.880859375,
0.68408203125,
0.880859375,
0.69873046875,
0.56494140625,
0.69873046875
],
[
0.118408203125,
0.69970703125,
0.429931640625,
0.69970703125,
0.429931640625,
0.71337890625,
0.118408203125,
0.71337890625
],
[
0.5283203125,
0.70361328125,
0.880859375,
0.70361328125,
0.880859375,
0.71875,
0.5283203125,
0.71875
],
[
0.1435546875,
0.72021484375,
0.26611328125,
0.724609375,
0.264404296875,
0.74072265625,
0.1427001953125,
0.736328125
],
[
0.51318359375,
0.7197265625,
0.87890625,
0.72021484375,
0.87890625,
0.7353515625,
0.51318359375,
0.73388671875
],
[
0.118408203125,
0.72705078125,
0.1502685546875,
0.72705078125,
0.1502685546875,
0.7353515625,
0.118408203125,
0.7353515625
],
[
0.5146484375,
0.73681640625,
0.64208984375,
0.73681640625,
0.64208984375,
0.75048828125,
0.5146484375,
0.75048828125
],
[
0.1175537109375,
0.7451171875,
0.483642578125,
0.7451171875,
0.483642578125,
0.7587890625,
0.1175537109375,
0.7587890625
],
[
0.654296875,
0.7451171875,
0.8759765625,
0.7490234375,
0.8759765625,
0.77392578125,
0.654296875,
0.77099609375
],
[
0.118408203125,
0.76318359375,
0.484375,
0.76318359375,
0.484375,
0.77685546875,
0.118408203125,
0.77685546875
],
[
0.51318359375,
0.7626953125,
0.65576171875,
0.76025390625,
0.65576171875,
0.77490234375,
0.51318359375,
0.77685546875
],
[
0.72705078125,
0.771484375,
0.8037109375,
0.771484375,
0.8037109375,
0.7861328125,
0.72705078125,
0.7861328125
],
[
0.1175537109375,
0.779296875,
0.483642578125,
0.779296875,
0.483642578125,
0.7919921875,
0.1175537109375,
0.7919921875
],
[
0.85791015625,
0.783203125,
0.8818359375,
0.783203125,
0.8818359375,
0.79736328125,
0.85791015625,
0.79736328125
],
[
0.118408203125,
0.79443359375,
0.484375,
0.79443359375,
0.484375,
0.80810546875,
0.118408203125,
0.80810546875
],
[
0.51318359375,
0.7998046875,
0.87890625,
0.7998046875,
0.87890625,
0.8134765625,
0.51318359375,
0.8134765625
],
[
0.1175537109375,
0.81005859375,
0.490234375,
0.81103515625,
0.490234375,
0.826171875,
0.1175537109375,
0.82470703125
],
[
0.51123046875,
0.81787109375,
0.87744140625,
0.81787109375,
0.87744140625,
0.82763671875,
0.51123046875,
0.82763671875
],
[
0.118408203125,
0.82763671875,
0.4013671875,
0.82763671875,
0.4013671875,
0.84130859375,
0.118408203125,
0.84130859375
],
[
0.51318359375,
0.83203125,
0.87890625,
0.83203125,
0.87890625,
0.845703125,
0.51318359375,
0.845703125
],
[
0.1158447265625,
0.8525390625,
0.459228515625,
0.85498046875,
0.459228515625,
0.869140625,
0.1158447265625,
0.86767578125
],
[
0.53564453125,
0.859375,
0.87744140625,
0.859375,
0.87744140625,
0.869140625,
0.53564453125,
0.869140625
],
[
0.1175537109375,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.5146484375,
0.87060546875,
0.87744140625,
0.87060546875,
0.87744140625,
0.88134765625,
0.5146484375,
0.88134765625
],
[
0.51123046875,
0.8798828125,
0.86962890625,
0.8798828125,
0.86962890625,
0.8935546875,
0.51123046875,
0.8935546875
],
[
0.118408203125,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.5322265625,
0.8935546875,
0.880859375,
0.8935546875,
0.880859375,
0.90673828125,
0.5322265625,
0.90673828125
],
[
0.118408203125,
0.90478515625,
0.48779296875,
0.90478515625,
0.48779296875,
0.91845703125,
0.118408203125,
0.91845703125
],
[
0.51123046875,
0.90673828125,
0.7958984375,
0.90625,
0.7958984375,
0.91748046875,
0.51123046875,
0.91845703125
]
] |
[
"3Backgrounds",
"签3以进行不确定性估计。根据公式E1，我",
"3.1ProblemFormulation",
"们能够根据［MASK】令牌的预测，获得每个样",
"本上不同标签词的出现概率",
"表示为每个有限的预算XMAXXMAXD",
"但是，直接采用此概率可能会出现问题，因为",
"[aj}=1。接下来，获取的样本被标记并从Du",
"PLM会遇到错误校准问题F(Zhaoetal2021",
"移动到D。然后，我们使用D，对预训练的",
"Huetal.,2022），i.e.，标签字可能会出现不同",
"语言模型f（;の）进行微调，以最大程度地提",
"的频率，从而使其中一些不太可能被预测。因",
"高下游分类任务的性能。上述步骤可以是单",
"此，方程1中的预测和估计的不确定性可能是",
"轮 (Chang et al., 2021; Hacohen et al., 2022)(在",
"有偏差的",
"本例中为b=B），也可以重复多轮（Yuan",
"意识到这一点，我们采用（Huetal.,2022)中",
"etal.,2020）（b=B//Roundsl），直到达到预",
"的方法来计算标签词的上下文先验。我们首先",
"算|B",
"为每个类i选择具有最高p(yile）的样本",
"3.2Prompt-based Learning for PLMs",
"作为",
"s=U",
"Top-k p(yila).",
"(2)",
"已经提出了提示方法，通过应用完形填空式任",
"ie[1,...] reDu",
"务来微调PLM（SchickandSchuitze,202la,b）来",
"来构造支持集S，然后，将上下文化的先验近",
"弥合预训练和微调阶段之间的差距。形式上",
"似为",
"提示中有两个关键组件：预定义的模板T和版",
"1",
"P(u) ~",
"PM(IMASKI =|2)，(",
"ISIEs",
"(3)",
"含一段自然语言文本以及［MASK】令牌的模板",
"本控制器V。对于每个输人示例，它将被包",
"包裹，然后再送人PLMM。然后，使用语言转",
"4.2",
"Uncertainty Propagation for Data Utility",
"",
"换器V将任务标签y映射到词汇表中的单个单",
"Estimation",
"词V(y）。以二元情感分类为例，对于输入句子",
"尽管我们已经减轻了基于提示的伪标签的偏",
"c，模板T可以是T(α）=[c.Itwas[MASK]］,",
"差，但由于在零拍设置下监管不足，此类伪标",
"而正面和负面情感的语言器可以分别是“good",
"签仍然可能不准确。在这种情况下，直接使用",
"”和“terrible”",
"方程E?？中的不确定性进行样本选择会产生",
"使用模板和语言转换器，我们可以通过掩码",
"次优结果，因为它可能对异常值敏感，异常值",
"语言建模（MLM）计算标签集上的概率分",
"自然具有较大的模型不确定性，但对模型学",
"布，如下所示",
"习E(Karamchetietal.，2021）的益处较小",
"p(y / a) =p([MASKJ = V(y) / T(r))",
"为了解决这个问题，我们利用嵌入空间中的",
"exp (w(u)hMASK)",
"核相似性来测量数据点之间的相关性并传播模",
"(1)",
"型不确定性：对于每个数据点C，我们首先使用",
"Zy'ey exp (w(y)hEMASK))",
"SimCSEE(Gao et al.,2021b)4作为 z = g(c;0)",
"其中h[MASKI是［MASK】标记的隐藏嵌人，WV(u)",
"来计算其嵌入，并基于其计算K-近邻。然后",
"表示来自M的标签字V(y)的嵌入。由于这些",
"我们选择径向基函数（RBF）（Scholkopf etal.",
"令牌的嵌人已在具有MLM自标的预训练期间",
"1997）作为两个数据点i和α的相似性度量，",
"进行了优化，因此提示的使用缩小了预训练和",
"表示为",
"微调之间的差距。换句话说，在使PLM适应",
"r (ci,c) =exp(-p zi-zjll2)，(4)",
"新任务时，提示可作为先验知识的来源",
"其中zi是来自SimCSE的i的嵌人，p是控",
"Methodology",
"制权重传播的超参数。形式上，&的传播不确",
"4",
"定性可以表示为",
"在本节中，我们介绍了利用提示进行冷启动数",
"ErExkn() (a, i) u(2i)",
"据选择的方法PATRON。我们首先介绍如何利",
"uprop(r) = u(r) +",
"KNN(α))",
"用提示进行冷启动场景下的不确定性估计。利",
"(5)",
"用估计的不确定性，我们提出了两个关键设",
"我们强调，只有当样本对其自身及其邻居具",
"计，即不确定性传播和分区重写（PTR）策略，",
"有较高的不确定性时：才会导致较高的传播不",
"以平衡样本选择的信息量和多样性。",
"确定性，这表明PLM不确定样本周围的周围",
"4.1 Uncertainty Estimation with Prompts",
"In this study,we usethe manual prompts and verbalizers",
"我们首先描述如何估计未标记数据的不确定性",
"from existing works (Hu et al., 2022; Schick and Schiutze",
"202la) due to their simplicity and competitive performance",
"以促进PATRON。给定没有标记数据的预训练",
"4Notably,we use the version of princeton-nlp/",
"语言模型（PLM）M，我们利用提示生成伪标",
"unsup-simcse-roberta-base as theencoder"
] |
{
"arxiv_id": "2209.06995",
"language": "zh",
"page_idx": 3
}
|
||
2209.06995-zh-4
|
2209.06995-zh-4.jpg
|
<FigureHere>
Figure 2: The illustration of the overall procedure for PATRON .
区域。在这种情况下，主动注释此类样本将对
PLM 最有益
### 4.3 Partition-then-rewrite( PTR)for Diversity-Promoting Data Selection
<table>
<tbody>
<tr>
<td>$\textbf{rithm 1: Procedures of the data se- }$ on strategy PATRON.</td>
</tr>
<tr>
<td>$\mathbf{ut:}$ Unlabeled samples $\mathcal{X}_u;$Pre-trained LM $\mathcal{M}=f(\cdot;\theta)$ , number of acquired samples $B$ ,the number of iterations $T\left(T=2\mathrm{~in~this}\right)$ ark</td>
</tr>
<tr>
<td>tep 1: Uncertainty Propagation for Utility stimation.</td>
</tr>
<tr>
<td>Calculate uncertainty for samples $x\in\mathcal{X}_u$ with ompts based on Eq. (??).</td>
</tr>
<tr>
<td>11 L 2 $\operatorname{tep}2:$Predict-then-propagate(PTR) for Diversity</td>
</tr>
<tr>
<td>$\operatorname{tep}2:$Predict-then-propagate (PTR) for Diversity ${\mathrm{omoting~Selection}}.$</td>
</tr>
<tr>
<td>Run K- Means on $\mathcal{X} _{u}$ with $k=B$ until mergence.</td>
</tr>
<tr>
<td>Select initial sample set $Q^{(0)}$ based on Eq.(6).</td>
</tr>
<tr>
<td>$d0$ 2c. Building the additional KNN graph to obtain ${\mathcal{X} }$ with Fa (7)</td>
</tr>
<tr>
<td>2d.Update$Q^{\mathrm{b}}$ by optimizing the selected sample within each cluster $\tilde{q}$ with Eq.(8).</td>
</tr>
<tr>
<td>tput: The final selected labeled data $Q^{(T)}$</td>
</tr>
</tbody>
</table>
添加了一个额外的正则化项，以防止相邻集群
中的样本过于接近：
$$\begin{aligned}&\tilde{q}_{i}=\operatorname*{argmax}_{x_{j}\in\mathcal{C}_{i}}(\hat{u}_{\mathrm{prop}}(x_{j})-\beta\left\|\boldsymbol{z}_{j}-\bar{\boldsymbol{z}}_{i}\right\|_{2}^{2}\\&-\gamma\sum_{q_{k}\in\mathcal{X}_{\mathrm{c-knn},i}}\left[m-\left\|\boldsymbol{z}_{j}-\boldsymbol{z}_{k}\right\|_{2}^{2}\right]_{+}),\\\end{aligned}$$
现代 AL 方法通常不是一次查询一个样本，而是查询一批样本以提高查询效率。在这种情况下，查询样本而不考虑其相关性将导致具有有限性能增益的冗余查询集$\overline{\mathrm{E}}($Ein-Dor et al., 2020)。我们现在介绍我们的 PTR 策略，用于以估计的不确定性为基础的多样性促进样本选择
用分区初始化选择。由于先前的研究$\overline{\mathrm{E}}$(Aharoni and Goldberg, 2020) 揭示了 PLM 隐式地学习按主题聚类的句子表示，即使没有进行微调，我们首先采用 K-Means 聚类将未标记的池$\mathcal{D}_u$根据其嵌入划分为不同的簇，并强制覆盖选定样本的不同主题。我们按照现有的作品$\mathbb{E}$ (Chang et al., 2021; Hacohen et al., 2022) 设置等于 $b$ 的簇数，表示为$\mathcal{C} _i$ $( 1\leq i\leq b) ^5$。然后，我们使用贪婪方法从$\mathcal{C}_i$中选择一个样本$q_i$ , 以将所选数据池 $Q$ 初始化为
$q_{i}= \operatorname* { argmax} _{{T_{i}\in \mathcal{C} _{i}}}\left ( \hat{u} _{\mathrm{prop}}( x_{j}) - \beta \left \| \boldsymbol{z}_{j}- \bar{z} _{i}\right \| _{2}^{2}\right ) $, (6) ,其中$\bar{z}_i=\frac1{|\mathcal{C}_i|}\sum_{x_j\in\mathcal{C}_i}\boldsymbol{z}_j$是群集$i$的质心，而$\beta$是超参数。通过这种方式，我们的目标是选择具有较高传播不确定性的数据点，而不是远离大多数数据点，以平衡不确定性和多样性
带重写的示例细化。尽管前面的步骤试图在每个聚类中选择信息量最大的样本，但它们无法对不同聚类中样本之间的关系进行建模。因此，样本仍然可以非常接近相邻簇中的其他选定样本，从而导致有限的总体多样性。为了解决这个问题，我们构建了一个额外的 KNN 图， 以从其他集群中检索最近的查询样本
(7)
$\mathcal{X}_{\mathrm{c-KNN,}i}=$KNN$(q_i,\mathcal{Q}).$
请注意，我们使用 c-KNN 表示集群级 KNN , 以与 4.2 节中的样本级 KNN 区分开。为了更新选定的池 $Q$ ,对于集群 $i$ ,我们向公式$\mathbb{E}$6
(8)
其中$\gamma$是惩罚项的权重$,m=0.5$是预定义的余量，$[\cdot]_+=\max(\cdot,0)$是函数。为了解释正则化项，我们认为，当相邻簇中选定样本之间的距离小于$m$时，正则化将大于0，以阻止它们一起被选择
我们将上述重写步骤运行几次，直到收敛以获得最终集$\mathcal{Q}=\{\tilde{q}_i\}_{i=1}^b$,这通常需要 2-3 次迭代°C。PATRON的整个过程在算法 1 中
# 5 Experiments
### 5.1 Experiment Setup
<NEXT\_LINE。遵循先前的工作$\overline{\mathrm{E}}($Yuan et al., 2020; Schröder et al., 2022), 我们在实验中使用了六个文本分类任务：IMDB$\boxed{\mathrm{F}}($Maas et al.
$^{5}$Here we use one-round AL for better illustration. We provide the details for adapting PTR to the multi-round AL setting in Appendix D .
$^{6}$The efficiency analysis of PATRON is in Appendix E.
|
[
[
0.16796875,
0.0926513671875,
0.21826171875,
0.0926513671875,
0.21826171875,
0.10272216796875,
0.16796875,
0.10272216796875
],
[
0.173828125,
0.10394287109375,
0.279541015625,
0.105712890625,
0.279541015625,
0.1175537109375,
0.173828125,
0.11517333984375
],
[
0.16796875,
0.11639404296875,
0.2301025390625,
0.11639404296875,
0.2301025390625,
0.127685546875,
0.16796875,
0.127685546875
],
[
0.75,
0.11224365234375,
0.8017578125,
0.11224365234375,
0.8017578125,
0.119384765625,
0.75,
0.119384765625
],
[
0.17626953125,
0.1300048828125,
0.35107421875,
0.1300048828125,
0.35107421875,
0.13720703125,
0.17626953125,
0.13720703125
],
[
0.7958984375,
0.1395263671875,
0.81689453125,
0.1395263671875,
0.81689453125,
0.1436767578125,
0.7958984375,
0.1436767578125
],
[
0.1695556640625,
0.146728515625,
0.2091064453125,
0.146728515625,
0.2091064453125,
0.157958984375,
0.1695556640625,
0.157958984375
],
[
0.330078125,
0.149658203125,
0.35107421875,
0.149658203125,
0.35107421875,
0.15673828125,
0.330078125,
0.15673828125
],
[
0.78857421875,
0.146728515625,
0.8251953125,
0.146728515625,
0.8251953125,
0.15380859375,
0.78857421875,
0.15380859375
],
[
0.231689453125,
0.162109375,
0.306396484375,
0.162109375,
0.306396484375,
0.1728515625,
0.231689453125,
0.1728515625
],
[
0.300537109375,
0.162109375,
0.35693359375,
0.162109375,
0.35693359375,
0.1728515625,
0.300537109375,
0.1728515625
],
[
0.16796875,
0.18701171875,
0.370361328125,
0.18701171875,
0.370361328125,
0.19775390625,
0.16796875,
0.19775390625
],
[
0.4130859375,
0.1859130859375,
0.55322265625,
0.18701171875,
0.55322265625,
0.198974609375,
0.4130859375,
0.19775390625
],
[
0.63623046875,
0.18701171875,
0.7958984375,
0.18701171875,
0.7958984375,
0.19775390625,
0.63623046875,
0.19775390625
],
[
0.2822265625,
0.208984375,
0.71533203125,
0.208984375,
0.71533203125,
0.22265625,
0.2822265625,
0.22265625
],
[
0.118408203125,
0.240478515625,
0.486083984375,
0.240478515625,
0.486083984375,
0.254638671875,
0.118408203125,
0.254638671875
],
[
0.5146484375,
0.243408203125,
0.83642578125,
0.24462890625,
0.83642578125,
0.25830078125,
0.5146484375,
0.257080078125
],
[
0.1158447265625,
0.254638671875,
0.21826171875,
0.257080078125,
0.21826171875,
0.271484375,
0.1158447265625,
0.26953125
],
[
0.5224609375,
0.260009765625,
0.7080078125,
0.260009765625,
0.7080078125,
0.273681640625,
0.5224609375,
0.273681640625
],
[
0.5263671875,
0.27490234375,
0.8212890625,
0.27490234375,
0.8212890625,
0.2880859375,
0.5263671875,
0.2880859375
],
[
0.1175537109375,
0.287353515625,
0.162841796875,
0.287353515625,
0.162841796875,
0.298583984375,
0.1175537109375,
0.298583984375
],
[
0.1561279296875,
0.28515625,
0.4306640625,
0.28515625,
0.4306640625,
0.298583984375,
0.1561279296875,
0.298583984375
],
[
0.5751953125,
0.2880859375,
0.84033203125,
0.2880859375,
0.84033203125,
0.3017578125,
0.5751953125,
0.3017578125
],
[
0.156982421875,
0.300537109375,
0.438232421875,
0.3017578125,
0.438232421875,
0.31640625,
0.156982421875,
0.3154296875
],
[
0.5751953125,
0.298583984375,
0.84814453125,
0.298583984375,
0.84814453125,
0.312255859375,
0.5751953125,
0.312255859375
],
[
0.5751953125,
0.3134765625,
0.6171875,
0.3134765625,
0.6171875,
0.32470703125,
0.5751953125,
0.32470703125
],
[
0.1175537109375,
0.320556640625,
0.484375,
0.320556640625,
0.484375,
0.334228515625,
0.1175537109375,
0.334228515625
],
[
0.5283203125,
0.322509765625,
0.8076171875,
0.32373046875,
0.8076171875,
0.33740234375,
0.5283203125,
0.336181640625
],
[
0.118408203125,
0.33740234375,
0.486083984375,
0.33740234375,
0.486083984375,
0.350830078125,
0.118408203125,
0.350830078125
],
[
0.53662109375,
0.3349609375,
0.611328125,
0.336181640625,
0.611328125,
0.347412109375,
0.53662109375,
0.3466796875
],
[
0.5283203125,
0.345703125,
0.84228515625,
0.345703125,
0.84228515625,
0.359375,
0.5283203125,
0.359375
],
[
0.1175537109375,
0.3515625,
0.483642578125,
0.3515625,
0.483642578125,
0.365234375,
0.1175537109375,
0.365234375
],
[
0.53662109375,
0.359375,
0.705078125,
0.359375,
0.705078125,
0.372314453125,
0.53662109375,
0.372314453125
],
[
0.12005615234375,
0.369384765625,
0.483642578125,
0.369384765625,
0.483642578125,
0.383056640625,
0.12005615234375,
0.383056640625
],
[
0.53076171875,
0.37060546875,
0.84228515625,
0.37060546875,
0.84228515625,
0.385498046875,
0.53076171875,
0.385498046875
],
[
0.1175537109375,
0.385498046875,
0.483642578125,
0.385498046875,
0.483642578125,
0.3984375,
0.1175537109375,
0.3984375
],
[
0.52880859375,
0.383056640625,
0.85400390625,
0.383056640625,
0.85400390625,
0.396728515625,
0.52880859375,
0.396728515625
],
[
0.118408203125,
0.40087890625,
0.486083984375,
0.40087890625,
0.486083984375,
0.4150390625,
0.118408203125,
0.4150390625
],
[
0.5400390625,
0.396728515625,
0.6689453125,
0.396728515625,
0.6689453125,
0.40673828125,
0.5400390625,
0.40673828125
],
[
0.5283203125,
0.40625,
0.7900390625,
0.40625,
0.7900390625,
0.419189453125,
0.5283203125,
0.419189453125
],
[
0.1175537109375,
0.41748046875,
0.139404296875,
0.41748046875,
0.139404296875,
0.431640625,
0.1175537109375,
0.431640625
],
[
0.5400390625,
0.42041015625,
0.61376953125,
0.42041015625,
0.61376953125,
0.43115234375,
0.5400390625,
0.43115234375
],
[
0.118408203125,
0.435302734375,
0.48779296875,
0.435302734375,
0.48779296875,
0.448974609375,
0.118408203125,
0.448974609375
],
[
0.52880859375,
0.43115234375,
0.84228515625,
0.43115234375,
0.84228515625,
0.44482421875,
0.52880859375,
0.44482421875
],
[
0.5283203125,
0.4423828125,
0.68115234375,
0.441162109375,
0.68115234375,
0.4560546875,
0.5283203125,
0.457275390625
],
[
0.1175537109375,
0.45068359375,
0.486083984375,
0.45068359375,
0.486083984375,
0.46435546875,
0.1175537109375,
0.46435546875
],
[
0.5576171875,
0.45361328125,
0.81689453125,
0.45361328125,
0.81689453125,
0.46728515625,
0.5576171875,
0.46728515625
],
[
0.12005615234375,
0.46728515625,
0.484375,
0.46728515625,
0.484375,
0.48095703125,
0.12005615234375,
0.48095703125
],
[
0.56689453125,
0.466064453125,
0.7314453125,
0.466064453125,
0.7314453125,
0.479736328125,
0.56689453125,
0.479736328125
],
[
0.5556640625,
0.4755859375,
0.82958984375,
0.47802734375,
0.82958984375,
0.495361328125,
0.5556640625,
0.493408203125
],
[
0.118408203125,
0.482666015625,
0.484375,
0.482666015625,
0.484375,
0.496337890625,
0.118408203125,
0.496337890625
],
[
0.5634765625,
0.4912109375,
0.8212890625,
0.4912109375,
0.8212890625,
0.5048828125,
0.5634765625,
0.5048828125
],
[
0.1175537109375,
0.498779296875,
0.483642578125,
0.498779296875,
0.483642578125,
0.5126953125,
0.1175537109375,
0.5126953125
],
[
0.5283203125,
0.50830078125,
0.8095703125,
0.505859375,
0.8095703125,
0.52099609375,
0.5283203125,
0.5224609375
],
[
0.1175537109375,
0.51416015625,
0.486083984375,
0.51318359375,
0.486083984375,
0.52783203125,
0.1175537109375,
0.529296875
],
[
0.118408203125,
0.53076171875,
0.484375,
0.53076171875,
0.484375,
0.54443359375,
0.118408203125,
0.54443359375
],
[
0.1158447265625,
0.54345703125,
0.486083984375,
0.54345703125,
0.486083984375,
0.56103515625,
0.1158447265625,
0.56103515625
],
[
0.51123046875,
0.55419921875,
0.87890625,
0.55419921875,
0.87890625,
0.56787109375,
0.51123046875,
0.56787109375
],
[
0.1175537109375,
0.5634765625,
0.486083984375,
0.5625,
0.486083984375,
0.57666015625,
0.1175537109375,
0.57763671875
],
[
0.5146484375,
0.57177734375,
0.6689453125,
0.57177734375,
0.6689453125,
0.5849609375,
0.5146484375,
0.5849609375
],
[
0.11419677734375,
0.5791015625,
0.385498046875,
0.57666015625,
0.385498046875,
0.59423828125,
0.11419677734375,
0.5966796875
],
[
0.53076171875,
0.5849609375,
0.8310546875,
0.5849609375,
0.8310546875,
0.60546875,
0.53076171875,
0.60546875
],
[
0.1275634765625,
0.5966796875,
0.432373046875,
0.5966796875,
0.432373046875,
0.6142578125,
0.1275634765625,
0.6142578125
],
[
0.43408203125,
0.599609375,
0.480224609375,
0.5966796875,
0.483642578125,
0.61083984375,
0.4375,
0.6142578125
],
[
0.5810546875,
0.6015625,
0.623046875,
0.6015625,
0.623046875,
0.615234375,
0.5810546875,
0.615234375
],
[
0.181396484375,
0.61328125,
0.22412109375,
0.61328125,
0.22412109375,
0.62353515625,
0.181396484375,
0.62353515625
],
[
0.85791015625,
0.6142578125,
0.880859375,
0.6142578125,
0.880859375,
0.626953125,
0.85791015625,
0.626953125
],
[
0.11419677734375,
0.62158203125,
0.490234375,
0.62158203125,
0.490234375,
0.6494140625,
0.11419677734375,
0.6494140625
],
[
0.62451171875,
0.62353515625,
0.6455078125,
0.62353515625,
0.6455078125,
0.6357421875,
0.62451171875,
0.6357421875
],
[
0.6826171875,
0.62451171875,
0.71728515625,
0.62451171875,
0.71728515625,
0.6357421875,
0.6826171875,
0.6357421875
],
[
0.71728515625,
0.62353515625,
0.76171875,
0.62353515625,
0.76171875,
0.63818359375,
0.71728515625,
0.63818359375
],
[
0.60205078125,
0.64111328125,
0.66650390625,
0.642578125,
0.66650390625,
0.65380859375,
0.60205078125,
0.65283203125
],
[
0.58837890625,
0.62890625,
0.60205078125,
0.62890625,
0.60205078125,
0.63525390625,
0.58837890625,
0.63525390625
],
[
0.1175537109375,
0.64453125,
0.484375,
0.64453125,
0.484375,
0.65771484375,
0.1175537109375,
0.65771484375
],
[
0.1175537109375,
0.66015625,
0.486083984375,
0.6611328125,
0.486083984375,
0.67578125,
0.1175537109375,
0.6748046875
],
[
0.51318359375,
0.65478515625,
0.87890625,
0.65478515625,
0.87890625,
0.66845703125,
0.51318359375,
0.66845703125
],
[
0.5146484375,
0.67138671875,
0.880859375,
0.67138671875,
0.880859375,
0.68603515625,
0.5146484375,
0.68603515625
],
[
0.118408203125,
0.677734375,
0.466796875,
0.677734375,
0.466796875,
0.69140625,
0.118408203125,
0.69140625
],
[
0.1351318359375,
0.6923828125,
0.483642578125,
0.693359375,
0.483642578125,
0.7080078125,
0.1351318359375,
0.70703125
],
[
0.5146484375,
0.6884765625,
0.880859375,
0.6884765625,
0.880859375,
0.70166015625,
0.5146484375,
0.70166015625
],
[
0.51123046875,
0.70361328125,
0.8818359375,
0.70361328125,
0.8818359375,
0.71728515625,
0.51123046875,
0.71728515625
],
[
0.118408203125,
0.71044921875,
0.486083984375,
0.71044921875,
0.486083984375,
0.7236328125,
0.118408203125,
0.7236328125
],
[
0.5283203125,
0.72021484375,
0.60546875,
0.72021484375,
0.60546875,
0.732421875,
0.5283203125,
0.732421875
],
[
0.118408203125,
0.72705078125,
0.484375,
0.72705078125,
0.484375,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.53076171875,
0.7353515625,
0.880859375,
0.736328125,
0.880859375,
0.75048828125,
0.53076171875,
0.75
],
[
0.118408203125,
0.7421875,
0.486083984375,
0.7421875,
0.486083984375,
0.755859375,
0.118408203125,
0.755859375
],
[
0.51318359375,
0.7529296875,
0.880859375,
0.7529296875,
0.880859375,
0.76708984375,
0.51318359375,
0.76708984375
],
[
0.118408203125,
0.7587890625,
0.486083984375,
0.7587890625,
0.486083984375,
0.7724609375,
0.118408203125,
0.7724609375
],
[
0.51123046875,
0.7666015625,
0.84619140625,
0.7666015625,
0.84619140625,
0.78369140625,
0.51123046875,
0.78369140625
],
[
0.1175537109375,
0.77392578125,
0.484375,
0.77490234375,
0.484375,
0.7890625,
0.1175537109375,
0.7880859375
],
[
0.118408203125,
0.79150390625,
0.4130859375,
0.79150390625,
0.4130859375,
0.8046875,
0.118408203125,
0.8046875
],
[
0.51318359375,
0.802734375,
0.5263671875,
0.802734375,
0.5263671875,
0.8125,
0.51318359375,
0.8125
],
[
0.5458984375,
0.80029296875,
0.66015625,
0.80029296875,
0.66015625,
0.8154296875,
0.5458984375,
0.8154296875
],
[
0.20654296875,
0.80712890625,
0.393798828125,
0.8046875,
0.393798828125,
0.82177734375,
0.20654296875,
0.82373046875
],
[
0.462646484375,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82080078125,
0.462646484375,
0.82080078125
],
[
0.118408203125,
0.82373046875,
0.484375,
0.82373046875,
0.484375,
0.83740234375,
0.118408203125,
0.83740234375
],
[
0.51123046875,
0.826171875,
0.5556640625,
0.826171875,
0.5556640625,
0.83740234375,
0.51123046875,
0.83740234375
],
[
0.5556640625,
0.826171875,
0.70361328125,
0.826171875,
0.70361328125,
0.8388671875,
0.5556640625,
0.8388671875
],
[
0.1175537109375,
0.83837890625,
0.484375,
0.8388671875,
0.484375,
0.85400390625,
0.1175537109375,
0.8525390625
],
[
0.51123046875,
0.845703125,
0.87890625,
0.845703125,
0.87890625,
0.85986328125,
0.51123046875,
0.85986328125
],
[
0.1175537109375,
0.85546875,
0.486083984375,
0.85546875,
0.486083984375,
0.869140625,
0.1175537109375,
0.869140625
],
[
0.51318359375,
0.8623046875,
0.880859375,
0.8623046875,
0.880859375,
0.8759765625,
0.51318359375,
0.8759765625
],
[
0.1376953125,
0.88134765625,
0.48779296875,
0.88134765625,
0.48779296875,
0.89453125,
0.1376953125,
0.89453125
],
[
0.51318359375,
0.87744140625,
0.880859375,
0.876953125,
0.880859375,
0.89111328125,
0.51318359375,
0.892578125
],
[
0.1175537109375,
0.8935546875,
0.486083984375,
0.8935546875,
0.486083984375,
0.90673828125,
0.1175537109375,
0.90673828125
],
[
0.118408203125,
0.90673828125,
0.264404296875,
0.90673828125,
0.264404296875,
0.91748046875,
0.118408203125,
0.91748046875
],
[
0.5322265625,
0.90380859375,
0.86962890625,
0.90478515625,
0.86962890625,
0.91943359375,
0.5322265625,
0.91845703125
]
] |
[
"Sentencex",
"Best movie of this year.",
"Prompt T(α)",
"initialization",
"Bestmovieofthisyear.Itwas[MASK]",
"neo",
"PLM",
"u(x)",
"selectior",
"Pseudo Label",
"Uncertainty",
"1.Uncertainty Estimation with Prompts",
"2.Uncertainty Propagation",
"3. Partition-then-rewrite (PTR)",
"Figure 2: The illustration of the overall procedure for PATRON .",
"区域。在这种情况下，主动注释此类样本将对",
"Algorithm 1:Procedures of the data se-",
"PLM最有益",
"lection strategy PATRON.",
"Input: Unlabeled samples X,:Pre-trained LM",
"4.3",
"Partition-then-rewrite（PTR)for",
"M = f(-; 0) , number of acquired samples",
"Diversity-Promoting Data Selection",
"B, the number of iterations T (T =2 in this",
"work).",
"现代AL方法通常不是一次查询一个样本，而",
"I/ Step 1 : Uncertainty Propagation for Utility",
"是查询一批样本以提高查询效率。在这种情",
"Estimation.",
"la . Calculate uncertainty for samples e Xu with",
"况下查询样本而不考虑其相关性将导致具",
"prompts based on Eq. (??).",
"有有限性能增益的余查询集E(Ein-Doretal.",
"1b . Estimate uncertainty iuprop with Eq. (4) and (5).",
"2020）。我们现在介绍我们的PTR策略，用于",
"I/ Step 2 : Predict-then-propagate (PTR) for Diversity",
"以估计的不确定性为基础的多样性促进样本选",
"Promoting Selection.",
"2a . Run K-Means on Xu with k = B until",
"择",
"convergence",
"用分区初始化选择。由于先前的研究E(Aha-",
"2b . Select initial sample set Q(0) based on Eq. (6).",
"for t = 1, 2,..., T do",
"roniandGoldberg，2020）揭示了PLM隐式地学",
"2c. Building the additional KNN graph to",
"习按主题聚类的句子表示，即使没有进行微",
"obtain Xe-KNN with Eq. (7)",
"2d . Update Q(t) by optimizing the selected",
"调，我们首先采用K-Means聚类将未标记的池",
"sample within each cluster q with Eq. (8).",
"Du根据其嵌人划分为不同的簇，并强制覆盖",
"Output: The final selected labeled data Q(T)",
"选定样本的不同主题。我们按照现有的作品E",
"(Chang et al.,2021;Hacohen et al.,2022)设置等",
"于b的簇数，表示为C，（1≤i<b）５。然",
"添加了一个额外的正则化项，以防止相邻集群",
"后，我们使用贪婪方法从C中选择一个样本",
"中的样本过于接近：",
"qi，以将所选数据池Q初始化为",
"Qi =argmax (uprop(aj)-βzj -zill2",
"qi = argmax (uprop(ci) -βIzj - zll2",
")，(6)",
"rjEC,",
"ajEC,",
"(8)",
"，其中=e是群集的质心，而",
">",
"m-",
"lz-",
"qEXe-knn,",
"Y",
"β是超参数。通过这种方式，我们的目标是选",
"择具有较高传播不确定性的数据点，而不是远",
"其中是惩罚项的权重，m=0.5是预定义的",
"余量，[]+=max(,O)是函数。为了解释正则",
"离大多数数据点，以平衡不确定性和多样性",
"带重写的示例细化。尽管前面的步骤试图在",
"化项，我们认为，当相邻簇中选定样本之间的",
"距离小于㎡时，正则化将大于0，以阻止它们",
"每个聚类中选择信息量最大的样本，但它们无",
"起被选择",
"法对不同聚类中样本之间的关系进行建模。因",
"我们将上述重写步骤运行几次，直到收敛以",
"此，样本仍然可以非常接近相邻簇中的其他选",
"获得最终集={9-1，这通常需要2-3次",
"定样本，从而导致有限的总体多样性。为了解",
"选代６。PATRON 的整个过程在算法 1中",
"决这个问题，我们构建了一个额外的KNN图",
"以从其他集群中检索最近的查询样本",
"5",
"Experiments",
"Xe-KNN,i = KNN(qi, Q).",
"(7)",
"请注意，我们使用c-KNN表示集群级KNN",
"5.1",
"Experiment Setup",
"以与4.2节中的样本级KNN区分开。为了更",
"<NEXT_LINE。遵循先前的工作E(Yuanetal.",
"新选定的池Q，对于集群i，我们向公式E6",
"2020；Schroderetal.，2022），我们在实验中使",
"5Here we use one-round AL for better illustration. We",
"用了六个文本分类任务：IMDBE(Maasetal.,",
"provide the details for adapting PTR to the multi-round AL",
"setting in Appendix D.",
"6The efficiency analysis of PATRON is in Appendix E."
] |
{
"arxiv_id": "2209.06995",
"language": "zh",
"page_idx": 4
}
|
||
1907.11499-zh-0
|
1907.11499-zh-0.jpg
|
# $\textbf{Weakly Supervised Domain Detection}$
Yumo Xu and Mirella Lapata
Institute for Language, Cognition and Computation
School of Informatics, University of Edinburgh
10 Crichton Street, Edinburgh EH8 9AB
yumo.xußed.ac.uk,mlapêinf.ed.ac.uk
## 1 Introduction
文本分类是自然语言处理中的一项基本任务， 已被发现在广泛的应用中很有用，包括使用户能够识别网站上的内容的搜索引擎，情感和社交媒体分析，客户关系管理系统以及垃圾邮件检测。在过去的几年里，文本分类主要被建模为一个有监督的学习问题(例如。g。, Kim 2014; McCallum and Nigam 1998; Iyyer et al. 2015 ), 必须为其收集适当标记的数据。这样的数据通常是依赖于域的(i。e。, 涵盖特定主题，如与“商业”或“医学”相关的主题)和使用来自一个领域的数据训练的分类器很可能在另一个领域表现不佳。例如，短语“the mouse died quickly" 可能表示在描述手持式指向设备的客户评论中的负面情绪，或者在描述对啮齿动物进行的实验室实验时的正面情绪。随着神经网络等数据饥渴的机器学习技术的兴起，处
# Abstract
In this paper we introduce domain detection as a new natural language processing task. We argue that the ability to detect textual segments which are domain-heavy, i.e., sentences or phrases which are representative of and provide evidence for a given domain could enhance the robustness and portability of various text classification applications. We propose an encoderdetector framework for domain detection and bootstrap classifiers with multiple instance learning (MIL). The model is hierarchically organized and suited to multilabel classification. We demonstrate that despite learning with minimal supervision, our model can be applied to text spans of different granularities, languages, and genres. We also showcase the potential of domain detection for text summarization.
理各种领域$^{1}$的能力变得更加相关，并将其应用于大量文本媒体，从新闻文章到 Twitter, 博客文章，医学期刊，Reddit 评论和议会辩论 (Kim, 2014; Yang et al., 2016; Conneau et al., 2017;Zhang et al., 2016)
当训练数据可用于其中一个或几个域时，如何最好地处理多个域的问题在文献中引起了极大的兴趣。域自适应领域 (Jiang and Zhai, 2007; Blitzer et al., 2006; Daume III, 2007; Finkel and Manning, 2009; Lu et al., 2016) 旨在使用从具有足够标记数据的源域传输的知识，改善目标域中很少或没有标记数据的预测函数的学习。另一行工作 (Li and Zong, 2008; Wu and Huang, 2015; Chen and Cardie, 2018) 假设多个域可能存在标记数据，但数量不足以为其中一个或多个域训练分类器。多域文本分类的目的是利用所有可用资源，以便同时提高跨域的系统性能
在本文中，我们研究了如何获取特定领域的数据的问题，以便能够开发文本分类工具以及更多的领域感知应用程序，例如摘要，问题回答和信息提取。我们将此任务称为域检测，并假设一个相当常见的设置，其中语料库集合的域是已知的，目的是识别重域的文本段，即。e。,为给定领域提供证据的文档、句子或短语
可以将域检测表述为多标签分类问题，其中训练模型以识别句子，短语或单词级别的域证据。根据定义，域检测将需要使用细粒度的域标签来训练数据，从而增加了注释负担；我们必须首先为训练域检测器和建模我们关心的任务提供标签。在本文中，我们从多示例学习(MIL; Keeler and Rumelhart 1992 ) 的角度考虑
$^{\mathrm{l}}$The term“domain”has been permissively used in the
literature to describe (a) a collection of documents related to a particular topic such as user-reviews in Amazon for a product category (e.g., books, movies ),(b) a type of information source (e.g., twitter, news articles ), and (c) various fields of knowledge (e.g., Medicine, Law, Sport ). In this paper we adopt the latter definition of domains, however, nothing in our approach precludes applying it to different domain labels
|
[
[
0.30908203125,
0.0831298828125,
0.6943359375,
0.0843505859375,
0.6943359375,
0.09857177734375,
0.30908203125,
0.097412109375
],
[
0.37451171875,
0.141357421875,
0.6337890625,
0.141357421875,
0.6337890625,
0.1561279296875,
0.37451171875,
0.1561279296875
],
[
0.295654296875,
0.157958984375,
0.70947265625,
0.157958984375,
0.70947265625,
0.1728515625,
0.295654296875,
0.1728515625
],
[
0.312255859375,
0.174560546875,
0.6943359375,
0.174560546875,
0.6943359375,
0.188232421875,
0.312255859375,
0.188232421875
],
[
0.341796875,
0.190673828125,
0.66650390625,
0.190673828125,
0.66650390625,
0.2037353515625,
0.341796875,
0.2037353515625
],
[
0.297119140625,
0.208984375,
0.7080078125,
0.2078857421875,
0.7080078125,
0.221435546875,
0.297119140625,
0.22265625
],
[
0.264404296875,
0.29345703125,
0.343505859375,
0.295654296875,
0.343505859375,
0.31005859375,
0.264404296875,
0.30810546875
],
[
0.5146484375,
0.29443359375,
0.8818359375,
0.29443359375,
0.8818359375,
0.30810546875,
0.5146484375,
0.30810546875
],
[
0.5146484375,
0.311279296875,
0.8818359375,
0.312255859375,
0.8818359375,
0.325927734375,
0.5146484375,
0.32470703125
],
[
0.154541015625,
0.328857421875,
0.447509765625,
0.328857421875,
0.447509765625,
0.342529296875,
0.154541015625,
0.342529296875
],
[
0.515625,
0.327880859375,
0.8818359375,
0.327880859375,
0.8818359375,
0.341552734375,
0.515625,
0.341552734375
],
[
0.1561279296875,
0.34326171875,
0.45166015625,
0.34326171875,
0.45166015625,
0.35693359375,
0.1561279296875,
0.35693359375
],
[
0.5146484375,
0.344482421875,
0.88330078125,
0.344482421875,
0.88330078125,
0.358154296875,
0.5146484375,
0.358154296875
],
[
0.156982421875,
0.35693359375,
0.44921875,
0.35693359375,
0.44921875,
0.36767578125,
0.156982421875,
0.36767578125
],
[
0.5146484375,
0.35986328125,
0.7021484375,
0.35986328125,
0.7021484375,
0.37353515625,
0.5146484375,
0.37353515625
],
[
0.154541015625,
0.369384765625,
0.45166015625,
0.37060546875,
0.45166015625,
0.385498046875,
0.154541015625,
0.38427734375
],
[
0.5322265625,
0.374755859375,
0.88330078125,
0.3759765625,
0.88330078125,
0.39013671875,
0.5322265625,
0.3896484375
],
[
0.1561279296875,
0.385498046875,
0.44921875,
0.385498046875,
0.44921875,
0.3984375,
0.1561279296875,
0.3984375
],
[
0.5146484375,
0.391357421875,
0.8818359375,
0.392578125,
0.8818359375,
0.40673828125,
0.5146484375,
0.40625
],
[
0.156982421875,
0.40087890625,
0.450927734375,
0.40087890625,
0.450927734375,
0.410888671875,
0.156982421875,
0.410888671875
],
[
0.1561279296875,
0.413330078125,
0.450927734375,
0.413330078125,
0.450927734375,
0.427001953125,
0.1561279296875,
0.427001953125
],
[
0.515625,
0.4091796875,
0.8818359375,
0.4091796875,
0.8818359375,
0.4228515625,
0.515625,
0.4228515625
],
[
0.1561279296875,
0.427001953125,
0.450927734375,
0.427490234375,
0.450927734375,
0.441162109375,
0.1561279296875,
0.440673828125
],
[
0.515625,
0.42578125,
0.88330078125,
0.42578125,
0.88330078125,
0.439453125,
0.515625,
0.439453125
],
[
0.1561279296875,
0.441162109375,
0.44921875,
0.441162109375,
0.44921875,
0.4560546875,
0.1561279296875,
0.4560546875
],
[
0.51318359375,
0.441162109375,
0.880859375,
0.441162109375,
0.880859375,
0.454833984375,
0.51318359375,
0.454833984375
],
[
0.1561279296875,
0.4560546875,
0.450927734375,
0.4560546875,
0.450927734375,
0.4697265625,
0.1561279296875,
0.4697265625
],
[
0.515625,
0.4560546875,
0.8818359375,
0.457275390625,
0.8818359375,
0.471435546875,
0.515625,
0.47021484375
],
[
0.156982421875,
0.47021484375,
0.44921875,
0.47021484375,
0.44921875,
0.48388671875,
0.156982421875,
0.48388671875
],
[
0.51318359375,
0.473876953125,
0.8818359375,
0.473876953125,
0.8818359375,
0.48681640625,
0.51318359375,
0.48681640625
],
[
0.156982421875,
0.486328125,
0.44580078125,
0.486328125,
0.44580078125,
0.496337890625,
0.156982421875,
0.496337890625
],
[
0.51318359375,
0.486328125,
0.88330078125,
0.488037109375,
0.88330078125,
0.505859375,
0.51318359375,
0.50439453125
],
[
0.154541015625,
0.498779296875,
0.44921875,
0.49755859375,
0.44921875,
0.5126953125,
0.154541015625,
0.51318359375
],
[
0.51318359375,
0.50439453125,
0.88330078125,
0.5048828125,
0.88330078125,
0.51953125,
0.51318359375,
0.5185546875
],
[
0.156982421875,
0.51416015625,
0.447509765625,
0.51416015625,
0.447509765625,
0.52490234375,
0.156982421875,
0.52490234375
],
[
0.1561279296875,
0.52685546875,
0.450927734375,
0.52685546875,
0.450927734375,
0.54052734375,
0.1561279296875,
0.54052734375
],
[
0.515625,
0.521484375,
0.8818359375,
0.521484375,
0.8818359375,
0.53515625,
0.515625,
0.53515625
],
[
0.156982421875,
0.54150390625,
0.44921875,
0.54150390625,
0.44921875,
0.5546875,
0.156982421875,
0.5546875
],
[
0.515625,
0.53759765625,
0.8818359375,
0.53759765625,
0.8818359375,
0.55029296875,
0.515625,
0.55029296875
],
[
0.1561279296875,
0.5556640625,
0.450927734375,
0.5556640625,
0.450927734375,
0.5693359375,
0.1561279296875,
0.5693359375
],
[
0.51708984375,
0.5546875,
0.880859375,
0.5546875,
0.880859375,
0.5654296875,
0.51708984375,
0.5654296875
],
[
0.1561279296875,
0.5693359375,
0.44921875,
0.5693359375,
0.44921875,
0.5830078125,
0.1561279296875,
0.5830078125
],
[
0.53564453125,
0.57177734375,
0.87890625,
0.57177734375,
0.87890625,
0.58203125,
0.53564453125,
0.58203125
],
[
0.1561279296875,
0.58447265625,
0.364501953125,
0.58447265625,
0.364501953125,
0.59716796875,
0.1561279296875,
0.59716796875
],
[
0.515625,
0.5859375,
0.880859375,
0.5859375,
0.880859375,
0.599609375,
0.515625,
0.599609375
],
[
0.51708984375,
0.6015625,
0.880859375,
0.6015625,
0.880859375,
0.615234375,
0.51708984375,
0.615234375
],
[
0.515625,
0.61767578125,
0.880859375,
0.61767578125,
0.880859375,
0.63134765625,
0.515625,
0.63134765625
],
[
0.1175537109375,
0.62744140625,
0.151123046875,
0.62744140625,
0.151123046875,
0.6396484375,
0.1175537109375,
0.6396484375
],
[
0.1427001953125,
0.6259765625,
0.26611328125,
0.626953125,
0.26611328125,
0.64111328125,
0.1427001953125,
0.64013671875
],
[
0.515625,
0.63525390625,
0.880859375,
0.63525390625,
0.880859375,
0.6484375,
0.515625,
0.6484375
],
[
0.12005615234375,
0.65380859375,
0.489501953125,
0.65380859375,
0.489501953125,
0.66748046875,
0.12005615234375,
0.66748046875
],
[
0.515625,
0.65185546875,
0.880859375,
0.65185546875,
0.880859375,
0.6650390625,
0.515625,
0.6650390625
],
[
0.12176513671875,
0.67041015625,
0.48779296875,
0.67041015625,
0.48779296875,
0.68408203125,
0.12176513671875,
0.68408203125
],
[
0.51318359375,
0.66650390625,
0.880859375,
0.66650390625,
0.880859375,
0.6796875,
0.51318359375,
0.6796875
],
[
0.12005615234375,
0.68701171875,
0.484375,
0.68701171875,
0.484375,
0.69775390625,
0.12005615234375,
0.69775390625
],
[
0.5341796875,
0.68408203125,
0.880859375,
0.68408203125,
0.880859375,
0.69775390625,
0.5341796875,
0.69775390625
],
[
0.12176513671875,
0.703125,
0.48779296875,
0.703125,
0.48779296875,
0.71630859375,
0.12176513671875,
0.71630859375
],
[
0.515625,
0.69970703125,
0.8818359375,
0.69970703125,
0.8818359375,
0.71337890625,
0.515625,
0.71337890625
],
[
0.12005615234375,
0.71875,
0.48779296875,
0.71875,
0.48779296875,
0.732421875,
0.12005615234375,
0.732421875
],
[
0.5146484375,
0.71630859375,
0.8818359375,
0.71630859375,
0.8818359375,
0.72998046875,
0.5146484375,
0.72998046875
],
[
0.1368408203125,
0.7353515625,
0.490234375,
0.7353515625,
0.490234375,
0.75,
0.1368408203125,
0.75
],
[
0.5146484375,
0.732421875,
0.8818359375,
0.732421875,
0.8818359375,
0.7451171875,
0.5146484375,
0.7451171875
],
[
0.118408203125,
0.75,
0.489501953125,
0.75,
0.489501953125,
0.76318359375,
0.118408203125,
0.76318359375
],
[
0.51318359375,
0.74755859375,
0.8818359375,
0.74755859375,
0.8818359375,
0.76123046875,
0.51318359375,
0.76123046875
],
[
0.12176513671875,
0.765625,
0.48779296875,
0.765625,
0.48779296875,
0.779296875,
0.12176513671875,
0.779296875
],
[
0.5146484375,
0.76416015625,
0.8818359375,
0.76416015625,
0.8818359375,
0.77783203125,
0.5146484375,
0.77783203125
],
[
0.12176513671875,
0.783203125,
0.489501953125,
0.783203125,
0.489501953125,
0.79638671875,
0.12176513671875,
0.79638671875
],
[
0.51123046875,
0.779296875,
0.8818359375,
0.779296875,
0.8818359375,
0.7919921875,
0.51123046875,
0.7919921875
],
[
0.12176513671875,
0.798828125,
0.48779296875,
0.798828125,
0.48779296875,
0.8125,
0.12176513671875,
0.8125
],
[
0.5341796875,
0.80419921875,
0.88330078125,
0.8046875,
0.88330078125,
0.8193359375,
0.5341796875,
0.818359375
],
[
0.1234130859375,
0.81640625,
0.470947265625,
0.81640625,
0.470947265625,
0.82666015625,
0.1234130859375,
0.82666015625
],
[
0.5146484375,
0.8193359375,
0.880859375,
0.8193359375,
0.880859375,
0.830078125,
0.5146484375,
0.830078125
],
[
0.12005615234375,
0.83056640625,
0.490234375,
0.83056640625,
0.490234375,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.51318359375,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.84423828125,
0.51318359375,
0.84423828125
],
[
0.118408203125,
0.84716796875,
0.48779296875,
0.84716796875,
0.48779296875,
0.86083984375,
0.118408203125,
0.86083984375
],
[
0.5146484375,
0.8466796875,
0.8818359375,
0.8466796875,
0.8818359375,
0.85693359375,
0.5146484375,
0.85693359375
],
[
0.12176513671875,
0.86328125,
0.48779296875,
0.86328125,
0.48779296875,
0.876953125,
0.12176513671875,
0.876953125
],
[
0.51318359375,
0.85791015625,
0.8818359375,
0.85693359375,
0.8818359375,
0.87060546875,
0.51318359375,
0.87158203125
],
[
0.5146484375,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8828125,
0.5146484375,
0.8828125
],
[
0.12176513671875,
0.8798828125,
0.48779296875,
0.8798828125,
0.48779296875,
0.8935546875,
0.12176513671875,
0.8935546875
],
[
0.5146484375,
0.88525390625,
0.87890625,
0.88525390625,
0.87890625,
0.8955078125,
0.5146484375,
0.8955078125
],
[
0.12176513671875,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.12176513671875,
0.9091796875
],
[
0.5146484375,
0.89794921875,
0.8623046875,
0.896484375,
0.8623046875,
0.90771484375,
0.5146484375,
0.9091796875
]
] |
[
"Weakly Supervised Domain Detection",
"Yumo Xu and Mirella Lapata",
"Institute for Language, Cognition and Computation",
"School of Informatics, University of Edinburgh",
"10 Crichton Street, Edinburgh EH8 9AB",
"yumo.xu@ed.ac.uk,mlap@inf.ed.ac.uk",
"Abstract",
"理各种领域1的能力变得更加相关，并将其",
"应用于大量文本媒体，从新闻文章到Twitter",
"In this paper we introduce domain detec",
"博客文章，医学期刊，Reddit评论和议会辩",
"tion as a new natural language processing",
"论 (Kim, 2014; Yang et al., 2016; Conneau et al.",
"task.We argue that the ability to detect",
"2017; Zhang et al., 2016)",
"textual segments which are domain-heavy,",
"当训练数据可用于其中一个或几个域时，如",
"i.e., sentences or phrases which are rep-",
"何最好地处理多个域的问题在文献中引起了极",
"resentativeof andprovideevidencefora",
"given domain could enhance the robustness",
"大的兴趣。域自适应领域（JiangandZhai，2007;",
"and portability of various text classifica-",
"Blitzer et al., 2006; Daume III, 2007; Finkel and",
"tion applications. We propose an encoder-",
"Manning，2009;Luetal.,2016）旨在使用从具有",
"detector framework for domain detection",
"足够标记数据的源域传输的知识，改善目标",
"and bootstrap classifiers with multiple in-",
"域中很少或没有标记数据的预测函数的学习。",
"stance learning (MIL).The model is hier",
"另一行工作(Li and Zong,2008; Wu and Huang,",
"archically organized and suited to multil-",
"2015；ChenandCardie，2018）假设多个域可能",
"abel classification.We demonstrate that de-",
"spite learning with minimal supervision, our",
"存在标记数据，但数量不足以为其中一个或多",
"model can be applied to text spans of differ-",
"个域训练分类器。多域文本分类的目的是利用",
"ent granularities, languages, and genres. We",
"所有可用资源，以便同时提高跨域的系统性能",
"also showcase the potential of domain de-",
"在本文中，我们研究了如何获取特定领域的",
"tection for text summarization.",
"数据的问题，以便能够开发文本分类工具以及",
"更多的领域感知应用程序，例如摘要，问题回",
"答和信息提取。我们将此任务称为域检测，并",
"1",
"Introduction",
"假设一个相当常见的设置，其中语料库集合的",
"文本分类是自然语言处理中的一项基本任务，",
"域是已知的，目的是识别重域的文本段，即。",
"已被发现在广泛的应用中很有用，包括使用户",
"e。，为给定领域提供证据的文档、句子或短语",
"能够识别网站上的内容的搜索引擎，情感和社",
"可以将域检测表述为多标签分类问题，其中",
"交媒体分析，客户关系管理系统以及垃圾邮件",
"训练模型以识别句子，短语或单词级别的域证",
"检测。在过去的几年里，文本分类主要被建模为",
"据。根据定义，域检测将需要使用细粒度的域",
"个有监督的学习问题（例如。g。，Kim2014;",
"标签来训练数据，从而增加了注释负担；我们",
"McCallum and Nigam 1998; Iyyer et al. 2015),",
"必须首先为训练域检测器和建模我们关心的",
"必须为其收集适当标记的数据。这样的数据通",
"任务提供标签。在本文中，我们从多示例学习",
"常是依赖于域的(i。e。，涵盖特定主题，如",
"（MIL；KeelerandRumelhart1992）的角度考虑",
"与“商业”或“医学”相关的主题）和使用来",
"'The term “\"domain\" has been permissively used in the",
"自一个领域的数据训练的分类器很可能在另",
"literature to describe (a) a collection of documents relatedto a",
"个领域表现不佳。例如，短语“the mousedied",
"particular topic such as user-reviews in Amazon for a produc",
"quickly”可能表示在描述手持式指向设备的",
"category (e.g.,books,movies),(b)a type of information",
"客户评论中的负面情绪，或者在描述对啮齿",
"source (e.g., twitter, news articles ), and (c) various fields of",
"knowledge(e.g.,Medicine,Law,Sport).In thispaper we",
"动物进行的实验室实验时的正面情绪。随着神",
"adoptthelatterdefinitionofdomains,however,nothing in ou",
"经网络等数据饥渴的机器学习技术的兴起，处",
"approachprecludes applying itto different domainlabels."
] |
{
"arxiv_id": "1907.11499",
"language": "zh",
"page_idx": 0
}
|
||
1907.11499-zh-1
|
1907.11499-zh-1.jpg
|
了细粒度域检测问题，并开发了很少有人参与的域模型。我们的模型不需要从单独标记的细分中学习，而是只需要文档级监督和可选的先验领域知识，并学习内省地判断构成细分的领域。重要的是，我们不需要文档级域注释，因为我们通过利用从维基百科获取的信息通过远程监督获得这些注释
我们的域检测框架包括两个神经网络模块； 编码器学习单词和句子的表示以及先验域信息，如果后者可用(e。g。,域定义),而检测器为单词、句子和文档生成特定于域的分数。我们获得了一个分段级域预测器，该预测器使用分层的，基于注意力的神经体系结构 (Vaswani et al., 2017) 在文档级标签上进行了端到端训练。我们使用英语和中文进行域检测实验，并使用自动和基于人的评估来衡量系统性能。实验结果表明，尽管从弱监督中学习，我们的模型优于几个强基线，并且在语言和文本类型上都很健壮。我们还展示了我们的模型在文本摘要中的应用潜力
我们在这项工作中的贡献有三个方面；我们提出了域检测，作为一个新的细粒度多标签学习问题，我们认为这将有利于域感知 NLP 工具的开发；我们在多实例学习的背景下引人了弱监督编码器-检测器模型；并证明它可以在不修改的情况下跨语言和文本类型应用。
## 2 Related Work
我们的工作位于多个研究领域的交汇处，包括领域自适应，表示学习，多实例学习和主题建模。我们在下面回顾相关工作
Domain Adaptation 已经提出了多种域自适应方法 (Jiang and Zhai, 2007; Arnold et al., 2007; Pan et al., 2010) 来处理监督模型面临的新域中注释数据的缺乏。Daume III and Marcu (2006) 建议学习三个独立的模型，一个特定于源域， 一个特定于目标域，第三个表示域一般信息。在 Daume III (2007) 中进一步引人了一种简单而有效的功能增强技术，该技术随后在分层贝叶斯框架内对 Finkel and Manning (2009) 进行了重铸。最近，Lu et al.(2016)提出了一个用于域自适应的通用正则化框架，而 CamachoCollados and Navigli (2017) 将域信息集成到词汇资源中。文本分类中的一种流行方法学习跨多个域不变的特征，同时显式建模每个域的单个特征 (Chen and Cardie, 2018; Wu and Huang, 2015; Bousmalis et al., 2016)
类似于域自适应，我们的检测任务还可以识别不同域的最具判别性的特征。然而，虽然适应旨在通过转移知识使模型更具可移植性，但检测侧重于域本身，并识别为其语义提供最佳
证据的文本段，从而允许创建具有显式域标签
的数据集，可以进一步应用域适应技术
Multiple Instance Learning 多实例学习(MIL)处理标签与实例组或包(在我们的例子中为文档)相关联的问题，而实例标签(段级域标签)未被观察到。然后，任务是通过推断包 (Keeler and Rumelhart, 1992; Dietterich et al., 1997; Maron and Ratan, 1998) 的标签或联合实例和包(Zhou et al., 2009; Wei et al., 2014; Kotzias et al., 2015) 的标签来进行聚合实例级预测。我们的域检测模型是后一种变体的一个例子
初始 MIIL 模型，在包标签和实例标签之间采用了相对强的一致性假设。例如，在二进制分类中，只有当袋子的所有实例都为正 (Dietterich et al., 1997; Maron and Ratan, 1998; Zhang et al., 2002; Andrews and Hofmann, 2004; Carbonetto et al.,2008)。随后通过调查预测组合(Weidmann et al., 2003; Zhou et al., 2009) 放宽了假设
在 NLP 中，多实例学习主要应用于情感分析。Kotzias et al.(2015) 使用由预先训练的分层 CNN (Denil et al., 2014) 获得的句子向量作为 MIL 目标下的特征，该目标仅对实例对包分类的贡献进行平均 (i。e。,正面/负面文档情绪)。Pappas and Popescu-Belis (2014) 采用多实例回归模型，将情绪得分分配给特定的产品方面，使用预测的加权求和。最近，Angelidis and Lapata (2018) 提出了 MILNET , 这是一种用于情感分析的多实例学习网络模型。他们采用注意力机制来灵活地权衡预测，并识别情绪化的文本片段(i。e。、句子或从句)
我们与以前基于 MIL 的工作不同，设计了一种具有自我关注和非循环结构的编码模块，该模块特别适合于长文档的高效建模。与 MILNET (Angelidis and Lapata, 2018)相比，我们的方法推广到任意粒度的段；它引人了一个实例评分函数，该函数支持多标签而不是二进制分类，并考虑了先验知识(e。g。,域定义),以更好地通知模型的预测
Topic Modeling 主题模型是围绕文档集合的语义由潜在变量控制的思想构建的。因此，目的是揭示这些潜在的变量-主题-塑造文档集合的含义。潜在狄利克雷分配(LDA; Blei et al. 2003 )是最著名的主题模型之一。在 LDA 中， 文档是使用$\overline{\mathbb{E}}K$ 主题的混合概率生成的，这些主题又以单词分布为特征。文档中的单词是通过根据主题分布重复采样主题并根据所选主题选择单词来生成的
尽管大多数主题模型都是无监督的，但某
|
[
[
0.124267578125,
0.07659912109375,
0.489501953125,
0.07659912109375,
0.489501953125,
0.09027099609375,
0.124267578125,
0.09027099609375
],
[
0.5146484375,
0.07659912109375,
0.8818359375,
0.07659912109375,
0.8818359375,
0.09027099609375,
0.5146484375,
0.09027099609375
],
[
0.12005615234375,
0.09320068359375,
0.48779296875,
0.09320068359375,
0.48779296875,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.5146484375,
0.09320068359375,
0.8271484375,
0.09320068359375,
0.8271484375,
0.10687255859375,
0.5146484375,
0.10687255859375
],
[
0.12005615234375,
0.10986328125,
0.48779296875,
0.10986328125,
0.48779296875,
0.12353515625,
0.12005615234375,
0.12353515625
],
[
0.52490234375,
0.119384765625,
0.880859375,
0.1217041015625,
0.880859375,
0.135986328125,
0.52490234375,
0.1341552734375
],
[
0.118408203125,
0.1258544921875,
0.486083984375,
0.1258544921875,
0.486083984375,
0.1395263671875,
0.118408203125,
0.1395263671875
],
[
0.118408203125,
0.14013671875,
0.486083984375,
0.14013671875,
0.486083984375,
0.15380859375,
0.118408203125,
0.15380859375
],
[
0.51318359375,
0.1383056640625,
0.8818359375,
0.1383056640625,
0.8818359375,
0.1519775390625,
0.51318359375,
0.1519775390625
],
[
0.12005615234375,
0.157958984375,
0.486083984375,
0.157958984375,
0.486083984375,
0.168701171875,
0.12005615234375,
0.168701171875
],
[
0.515625,
0.15380859375,
0.8818359375,
0.15380859375,
0.8818359375,
0.16748046875,
0.515625,
0.16748046875
],
[
0.12005615234375,
0.17333984375,
0.287109375,
0.17333984375,
0.287109375,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.5146484375,
0.17041015625,
0.87890625,
0.17041015625,
0.87890625,
0.18115234375,
0.5146484375,
0.18115234375
],
[
0.51318359375,
0.185302734375,
0.880859375,
0.185302734375,
0.880859375,
0.198974609375,
0.51318359375,
0.198974609375
],
[
0.1376953125,
0.191162109375,
0.489501953125,
0.191162109375,
0.489501953125,
0.201904296875,
0.1376953125,
0.201904296875
],
[
0.12005615234375,
0.2060546875,
0.489501953125,
0.2060546875,
0.489501953125,
0.2197265625,
0.12005615234375,
0.2197265625
],
[
0.51318359375,
0.203125,
0.8818359375,
0.203125,
0.8818359375,
0.2161865234375,
0.51318359375,
0.2161865234375
],
[
0.12005615234375,
0.22265625,
0.489501953125,
0.22265625,
0.489501953125,
0.236328125,
0.12005615234375,
0.236328125
],
[
0.5146484375,
0.218505859375,
0.880859375,
0.218505859375,
0.880859375,
0.232177734375,
0.5146484375,
0.232177734375
],
[
0.12005615234375,
0.2381591796875,
0.489501953125,
0.2381591796875,
0.489501953125,
0.251708984375,
0.12005615234375,
0.251708984375
],
[
0.515625,
0.236328125,
0.880859375,
0.236328125,
0.880859375,
0.2464599609375,
0.515625,
0.2464599609375
],
[
0.12005615234375,
0.254150390625,
0.48779296875,
0.254150390625,
0.48779296875,
0.267333984375,
0.12005615234375,
0.267333984375
],
[
0.5146484375,
0.25048828125,
0.8818359375,
0.25048828125,
0.8818359375,
0.26416015625,
0.5146484375,
0.26416015625
],
[
0.12176513671875,
0.270751953125,
0.48779296875,
0.270751953125,
0.48779296875,
0.283935546875,
0.12176513671875,
0.283935546875
],
[
0.5146484375,
0.267333984375,
0.5517578125,
0.267333984375,
0.5517578125,
0.27978515625,
0.5146484375,
0.27978515625
],
[
0.12005615234375,
0.2861328125,
0.48779296875,
0.2861328125,
0.48779296875,
0.2998046875,
0.12005615234375,
0.2998046875
],
[
0.53076171875,
0.283935546875,
0.8818359375,
0.283935546875,
0.8818359375,
0.297607421875,
0.53076171875,
0.297607421875
],
[
0.12005615234375,
0.302734375,
0.48779296875,
0.302734375,
0.48779296875,
0.31640625,
0.12005615234375,
0.31640625
],
[
0.515625,
0.300537109375,
0.87890625,
0.300537109375,
0.87890625,
0.311279296875,
0.515625,
0.311279296875
],
[
0.12176513671875,
0.318359375,
0.489501953125,
0.318359375,
0.489501953125,
0.33203125,
0.12176513671875,
0.33203125
],
[
0.5146484375,
0.31640625,
0.880859375,
0.31640625,
0.880859375,
0.330078125,
0.5146484375,
0.330078125
],
[
0.12005615234375,
0.334228515625,
0.48779296875,
0.334228515625,
0.48779296875,
0.347412109375,
0.12005615234375,
0.347412109375
],
[
0.51123046875,
0.330810546875,
0.8818359375,
0.33203125,
0.8818359375,
0.3466796875,
0.51123046875,
0.345703125
],
[
0.12005615234375,
0.350830078125,
0.489501953125,
0.350830078125,
0.489501953125,
0.364013671875,
0.12005615234375,
0.364013671875
],
[
0.51318359375,
0.3486328125,
0.880859375,
0.3486328125,
0.880859375,
0.3623046875,
0.51318359375,
0.3623046875
],
[
0.12005615234375,
0.365234375,
0.48779296875,
0.365234375,
0.48779296875,
0.37890625,
0.12005615234375,
0.37890625
],
[
0.51123046875,
0.363525390625,
0.8818359375,
0.364013671875,
0.8818359375,
0.37890625,
0.51123046875,
0.377685546875
],
[
0.12005615234375,
0.383056640625,
0.2476806640625,
0.383056640625,
0.2476806640625,
0.396728515625,
0.12005615234375,
0.396728515625
],
[
0.515625,
0.38134765625,
0.8818359375,
0.38134765625,
0.8818359375,
0.394287109375,
0.515625,
0.394287109375
],
[
0.1376953125,
0.399658203125,
0.48779296875,
0.399658203125,
0.48779296875,
0.413330078125,
0.1376953125,
0.413330078125
],
[
0.5146484375,
0.39794921875,
0.55322265625,
0.39794921875,
0.55322265625,
0.410400390625,
0.5146484375,
0.410400390625
],
[
0.12005615234375,
0.4150390625,
0.489501953125,
0.4150390625,
0.489501953125,
0.4287109375,
0.12005615234375,
0.4287109375
],
[
0.5322265625,
0.41455078125,
0.8818359375,
0.41455078125,
0.8818359375,
0.427490234375,
0.5322265625,
0.427490234375
],
[
0.12005615234375,
0.43115234375,
0.490234375,
0.43115234375,
0.490234375,
0.44482421875,
0.12005615234375,
0.44482421875
],
[
0.51318359375,
0.429931640625,
0.8818359375,
0.429931640625,
0.8818359375,
0.443603515625,
0.51318359375,
0.443603515625
],
[
0.12176513671875,
0.44775390625,
0.489501953125,
0.44775390625,
0.489501953125,
0.46142578125,
0.12176513671875,
0.46142578125
],
[
0.51318359375,
0.4453125,
0.880859375,
0.4453125,
0.880859375,
0.458984375,
0.51318359375,
0.458984375
],
[
0.118408203125,
0.463134765625,
0.48779296875,
0.463134765625,
0.48779296875,
0.476806640625,
0.118408203125,
0.476806640625
],
[
0.5146484375,
0.4619140625,
0.8818359375,
0.4619140625,
0.8818359375,
0.4755859375,
0.5146484375,
0.4755859375
],
[
0.12005615234375,
0.479736328125,
0.43994140625,
0.479736328125,
0.43994140625,
0.493408203125,
0.12005615234375,
0.493408203125
],
[
0.5146484375,
0.476806640625,
0.8818359375,
0.478515625,
0.8818359375,
0.4921875,
0.5146484375,
0.490478515625
],
[
0.515625,
0.49462890625,
0.8818359375,
0.49462890625,
0.8818359375,
0.50830078125,
0.515625,
0.50830078125
],
[
0.118408203125,
0.50830078125,
0.156982421875,
0.50830078125,
0.156982421875,
0.52099609375,
0.118408203125,
0.52099609375
],
[
0.151123046875,
0.50830078125,
0.274658203125,
0.50830078125,
0.274658203125,
0.521484375,
0.151123046875,
0.521484375
],
[
0.515625,
0.51025390625,
0.8818359375,
0.51025390625,
0.8818359375,
0.52392578125,
0.515625,
0.52392578125
],
[
0.51318359375,
0.52685546875,
0.880859375,
0.52685546875,
0.880859375,
0.54052734375,
0.51318359375,
0.54052734375
],
[
0.12005615234375,
0.533203125,
0.489501953125,
0.533203125,
0.489501953125,
0.54638671875,
0.12005615234375,
0.54638671875
],
[
0.12005615234375,
0.54736328125,
0.486083984375,
0.54736328125,
0.486083984375,
0.56103515625,
0.12005615234375,
0.56103515625
],
[
0.515625,
0.54443359375,
0.880859375,
0.54443359375,
0.880859375,
0.5546875,
0.515625,
0.5546875
],
[
0.51708984375,
0.55859375,
0.8818359375,
0.55859375,
0.8818359375,
0.572265625,
0.51708984375,
0.572265625
],
[
0.12005615234375,
0.5654296875,
0.359375,
0.5654296875,
0.359375,
0.5791015625,
0.12005615234375,
0.5791015625
],
[
0.51708984375,
0.576171875,
0.8818359375,
0.576171875,
0.8818359375,
0.5888671875,
0.51708984375,
0.5888671875
],
[
0.124267578125,
0.59033203125,
0.489501953125,
0.59033203125,
0.489501953125,
0.60400390625,
0.124267578125,
0.60400390625
],
[
0.51318359375,
0.59033203125,
0.796875,
0.59033203125,
0.796875,
0.60400390625,
0.51318359375,
0.60400390625
],
[
0.12176513671875,
0.60546875,
0.489501953125,
0.60546875,
0.489501953125,
0.619140625,
0.12176513671875,
0.619140625
],
[
0.53076171875,
0.60693359375,
0.87890625,
0.60693359375,
0.87890625,
0.62060546875,
0.53076171875,
0.62060546875
],
[
0.12005615234375,
0.62158203125,
0.489501953125,
0.62158203125,
0.489501953125,
0.63525390625,
0.12005615234375,
0.63525390625
],
[
0.5146484375,
0.62353515625,
0.880859375,
0.62353515625,
0.880859375,
0.63720703125,
0.5146484375,
0.63720703125
],
[
0.12005615234375,
0.6357421875,
0.48779296875,
0.6357421875,
0.48779296875,
0.6494140625,
0.12005615234375,
0.6494140625
],
[
0.5146484375,
0.64013671875,
0.880859375,
0.64013671875,
0.880859375,
0.65380859375,
0.5146484375,
0.65380859375
],
[
0.12005615234375,
0.65380859375,
0.489501953125,
0.65380859375,
0.489501953125,
0.66748046875,
0.12005615234375,
0.66748046875
],
[
0.51318359375,
0.65478515625,
0.880859375,
0.65478515625,
0.880859375,
0.66845703125,
0.51318359375,
0.66845703125
],
[
0.130126953125,
0.66845703125,
0.489501953125,
0.66943359375,
0.489501953125,
0.68408203125,
0.130126953125,
0.68310546875
],
[
0.515625,
0.6728515625,
0.8818359375,
0.6728515625,
0.8818359375,
0.68603515625,
0.515625,
0.68603515625
],
[
0.12005615234375,
0.68603515625,
0.48779296875,
0.68603515625,
0.48779296875,
0.69970703125,
0.12005615234375,
0.69970703125
],
[
0.51318359375,
0.68701171875,
0.8818359375,
0.6884765625,
0.8818359375,
0.703125,
0.51318359375,
0.70166015625
],
[
0.12005615234375,
0.703125,
0.483642578125,
0.703125,
0.483642578125,
0.71337890625,
0.12005615234375,
0.71337890625
],
[
0.51123046875,
0.70166015625,
0.884765625,
0.703125,
0.884765625,
0.72021484375,
0.51123046875,
0.7197265625
],
[
0.12005615234375,
0.71875,
0.48779296875,
0.71875,
0.48779296875,
0.732421875,
0.12005615234375,
0.732421875
],
[
0.515625,
0.7216796875,
0.69873046875,
0.7216796875,
0.69873046875,
0.7353515625,
0.515625,
0.7353515625
],
[
0.12176513671875,
0.7353515625,
0.486083984375,
0.7353515625,
0.486083984375,
0.7490234375,
0.12176513671875,
0.7490234375
],
[
0.12005615234375,
0.75048828125,
0.48779296875,
0.75048828125,
0.48779296875,
0.76416015625,
0.12005615234375,
0.76416015625
],
[
0.51904296875,
0.7490234375,
0.880859375,
0.7490234375,
0.880859375,
0.7626953125,
0.51904296875,
0.7626953125
],
[
0.12176513671875,
0.7685546875,
0.48779296875,
0.7685546875,
0.48779296875,
0.779296875,
0.12176513671875,
0.779296875
],
[
0.515625,
0.7666015625,
0.87744140625,
0.7666015625,
0.87744140625,
0.77685546875,
0.515625,
0.77685546875
],
[
0.12005615234375,
0.783203125,
0.489501953125,
0.783203125,
0.489501953125,
0.79638671875,
0.12005615234375,
0.79638671875
],
[
0.5146484375,
0.7822265625,
0.8818359375,
0.7822265625,
0.8818359375,
0.7958984375,
0.5146484375,
0.7958984375
],
[
0.12005615234375,
0.798828125,
0.489501953125,
0.798828125,
0.489501953125,
0.8125,
0.12005615234375,
0.8125
],
[
0.51318359375,
0.79736328125,
0.880859375,
0.79736328125,
0.880859375,
0.81103515625,
0.51318359375,
0.81103515625
],
[
0.118408203125,
0.8134765625,
0.489501953125,
0.8154296875,
0.489501953125,
0.8291015625,
0.118408203125,
0.82666015625
],
[
0.5146484375,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.5146484375,
0.82763671875
],
[
0.118408203125,
0.83056640625,
0.337646484375,
0.83056640625,
0.337646484375,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.51708984375,
0.830078125,
0.8818359375,
0.830078125,
0.8818359375,
0.84326171875,
0.51708984375,
0.84326171875
],
[
0.1376953125,
0.84716796875,
0.489501953125,
0.84716796875,
0.489501953125,
0.86083984375,
0.1376953125,
0.86083984375
],
[
0.515625,
0.8466796875,
0.8818359375,
0.8466796875,
0.8818359375,
0.85986328125,
0.515625,
0.85986328125
],
[
0.12176513671875,
0.8642578125,
0.489501953125,
0.8642578125,
0.489501953125,
0.87744140625,
0.12176513671875,
0.87744140625
],
[
0.51318359375,
0.8623046875,
0.880859375,
0.8623046875,
0.880859375,
0.8759765625,
0.51318359375,
0.8759765625
],
[
0.12176513671875,
0.8798828125,
0.489501953125,
0.8798828125,
0.489501953125,
0.8935546875,
0.12176513671875,
0.8935546875
],
[
0.5146484375,
0.87744140625,
0.66162109375,
0.87744140625,
0.66162109375,
0.89111328125,
0.5146484375,
0.89111328125
],
[
0.12176513671875,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.12176513671875,
0.9091796875
],
[
0.5322265625,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.5322265625,
0.9091796875
]
] |
[
"了细粒度域检测问题，并开发了很少有人参与",
"证据的文本段，从而允许创建具有显式域标签",
"的域模型。我们的模型不需要从单独标记的细",
"的数据集，可以进一步应用域适应技术",
"分中学习，而是只需要文档级监督和可选的先",
"MultipleInstanceLearning多实例学习",
"验领域知识，并学习内省地判断构成细分的领",
"域。重要的是，我们不需要文档级域注释，因",
"（MIL）处理标签与实例组或包（在我们的例",
"为我们通过利用从维基百科获取的信息通过远",
"子中为文档）相关联的问题，而实例标签（段",
"程监督获得这些注释",
"级域标签）未被观察到。然后：任务是通过",
"推断包（KeelerandRumelhart，1992；Dietterich",
"我们的域检测框架包括两个神经网络模块；",
"编码器学习单词和句子的表示以及先验域信",
"etal.,1997;MaronandRatan，1998）的标签或联",
"息，如果后者可用（e。g。，域定义），而检测器",
"合实例和包(Zhouetal.,2009；Weietal.,2014;",
"为单词、句子和文档生成特定于域的分数。我",
"Kotziasetal..2015）的标签来进行聚合实例级",
"们获得了一个分段级域预测器，该预测器使用",
"预测。我们的域检测模型是后一种变体的一个",
"分层的，基于注意力的神经体系结构（Vaswani",
"例子",
"etal.，2017)在文档级标签上进行了端到端训",
"初始MIL模型，在包标签和实例标签之间",
"练。我们使用英语和中文进行域检测实验，并",
"采用了相对强的一致性假设。例如，在二进制",
"使用自动和基于人的评估来衡量系统性能。实",
"分类中，只有当袋子的所有实例都为正(Diet-",
"验结果表明，尽管从弱监督中学习，我们的模",
"terich et al., 1997; Maron and Ratan, 1998; Zhang",
"型优于几个强基线，并且在语言和文本类型上",
"et al., 2002; Andrews and Hofmann, 2004; Car-",
"都很健壮。我们还展示了我们的模型在文本摘",
"bonettoetal.，2008）。随后通过调查预测组合",
"要中的应用潜力",
"(Weidmannetal.,2003;Zhou etal.,2009)放宽了",
"我们在这项工作中的贡献有三个方面；我们",
"假设",
"提出了域检测，作为一个新的细粒度多标签学",
"在NLP中，多实例学习主要应用于情感分",
"习问题，我们认为这将有利于域感知NLP工",
"析。Kotzias etal.（2015）使用由预先训练的分",
"具的开发；我们在多实例学习的背景下引入了",
"层CNN(Deniletal.,2014）获得的句子向量作",
"弱监督编码器-检测器模型；并证明它可以在",
"为MIL目标下的特征，该目标仅对实例对包",
"不修改的情况下跨语言和文本类型应用。",
"分类的贡献进行平均（i。e。，正面/负面文档",
"情绪）。PappasandPopescu-Belis(2014）采用多",
"2",
"Related Work",
"实例回归模型，将情绪得分分配给特定的产品",
"方面，使用预测的加权求和。最近，Angelidis",
"我们的工作位于多个研究领域的交汇处，包括",
"领域自适应，表示学习，多实例学习和主题建",
"andLapata（2018）提出了MILNET：这是一种",
"用于情感分析的多实例学习网络模型。他们采",
"模。我们在下面回顾相关工作",
"用注意力机制来灵活地权衡预测，并识别情绪",
"DomainAdaptation，已经提出了多种域自适",
"化的文本片段(i。e。、句子或从句)",
"应方法(Jiang and Zhai,2007;Arnold et al.,2007;",
"我们与以前基于MIL的工作不同，设计了一",
"Panetal.,2010)来处理监督模型面临的新域中",
"种具有自我关注和非循环结构的编码模块，该",
"注释数据的缺乏。DaumeIⅢII and Marcu(2006)",
"模块特别适合于长文档的高效建模。与MIL-",
"建议学习三个独立的模型，一个特定于源域，",
"NET（AngelidisandLapata，2018）相比，我们的",
"一个特定于目标域，第三个表示域一般信息。",
"方法推广到任意粒度的段；它引入了一个实例",
"在DaumeⅢ(2007)中进一步引人了一种简单",
"评分函数，该函数支持多标签而不是二进制分",
"而有效的功能增强技术，该技术随后在分层贝",
"类，并考虑了先验知识(e。g。，域定义)，以",
"叶斯框架内对FinkelandManning（2009）进行",
"更好地通知模型的预测",
"了重铸。最近，Luetal.(2016)提出了一个用",
"于域自适应的通用正则化框架，而Camacho-",
"TopicModeling主题模型是围绕文档集合的",
"ColladosandNavigli（2017）将域信息集成到词",
"语义由潜在变量控制的思想构建的。因此，目",
"汇资源中。文本分类中的一种流行方法学习跨",
"的是揭示这些潜在的变量-主题-塑造文档集合",
"多个域不变的特征，同时显式建模每个域的单",
"的含义。潜在狄利克雷分配(LDA；Blei et al",
"个特征（ChenandCardie,2018;WuandHuang，",
"2003）是最著名的主题模型之一。在LDA中，",
"2015; Bousmalis et al., 2016)",
"文档是使用EK主题的混合概率生成的，这些",
"类似于域自适应，我们的检测任务还可以识",
"主题又以单词分布为特征。文档中的单词是通",
"别不同域的最具判别性的特征。然而，虽然适",
"过根据主题分布重复采样主题并根据所选主题",
"应旨在通过转移知识使模型更具可移植性，但",
"选择单词来生成的",
"检测侧重于域本身，并识别为其语义提供最佳",
"尽管大多数主题模型都是无监督的，但某"
] |
{
"arxiv_id": "1907.11499",
"language": "zh",
"page_idx": 1
}
|
||
1907.11499-zh-3
|
1907.11499-zh-3.jpg
|
$g$ ,然后将其作为输入提供给句子级编码器， 该编码器输出上下文化表示$H$及其注意力权重$\beta$
在这项工作中，我们的目标是对相当长的文档(e。g。,维基百科文章；有关详细信息， 请参阅第 6 节)。因此，我们的编码器基于Transformer 架构 (Vaswani et al.,2017) ,这是最近提出的高效模型，在机器翻译 (Vaswani et al., 2017) 和问答 (Yu et al., 2018) 中实现了最先进的性能。Transformer 旨在减少顺序计算的基本约束，这是基于递归神经网络的大多数体系结构的基础。它消除了复发，有利于应用自我注意机制，该机制直接对句子中所有单词之间的关系进行建模，而不管它们的位置如何
Self-Attentive Encoder 如图E2 所示 , Transformer 是一个非递归框架，包括$m$ E相同的层。通过使用添加到输入嵌人的位置编码来表示序列中每个令牌的(相对或绝对)位置的信息(请参见图 2 的底部)。我们用$X$表示句子中的位置增强输入。我们的模型在单词和句子编码器中使用了四层。前三层与 Transformer $(m=3)$中的层相同，包括一个多头自关注子层和一个位置全连接的前馈网络。最后一层只是一个多头自我关注层，为后续操作产生注意力权重
单头注意将三个参数作为 Transformer (Vaswani et al.,2017)中的输入：查询矩阵，键(Vaswani et al.,2017)中的输入：查询矩阵，键(Vaswani et al.,2017)中的输入：查询矩阵，键矩阵和值矩阵。这三个矩阵是相同的，并且等于字编码器第一层的输入[E]X。单头注意力的输出通过以下方式计算：
$$\operatorname{head}(\boldsymbol X,\boldsymbol X,\boldsymbol X)=\operatorname{softmax}(\frac{\boldsymbol X\boldsymbol X^\intercal}{\sqrt{d_x}})\boldsymbol X.$$
(1)
多头注意力允许在不同位置共同关注来自不同表示子空间的信息。这是通过首先对输人应用不同的线性投影，然后将它们连接起来来完成的：
$$\mathrm{head}^{(k)}=\mathrm{head}(\boldsymbol{X}\boldsymbol{W}_{1}^{(k)},\boldsymbol{X}\boldsymbol{W}_{2}^{(k)},\boldsymbol{X}\boldsymbol{W}_{3}^{(k)})\\\mathrm{multi-head=concat(head^{(1)},..,head^{(r)})}\boldsymbol{W}_{4}$$
应用于每个位置：$^2$
,其中我们为单词和句子编码器采用四个头$(r=4)$。Transformer 中的第二个子层(见图2 )是一个完全连接的前馈网络，分别相同地
$$\mathrm{FFN}(\boldsymbol{x})=\max(0,\boldsymbol{xW}_5)\boldsymbol{W}_6.$$
(4)
在通过前三层对输入嵌人进行顺序编码之后，我们获得了上下文化的单词表示$Z\in$
$^{2}$We omit here the bias term for the sake of simplicity.
$\mathbb{R}^{d_z\times n}$。基于$\mathbb{E}\boldsymbol{Z}$ ,单词编码器中的最后一个多头注意力层针对每个句子产生一组注意力矩阵$\mathcal{A}=\{\boldsymbol{A}^{(k)}\}_{k=1}^r$,其中$\boldsymbol{A}^(k)\in\mathbb{R}^n\times n$ 。因此，在衡量单词对句子的贡献时，e。g。,就领域得分和表示形式而言，我们可以有选择地关注集合$\mathcal{A}=\{\boldsymbol{A}^{(k)}\}_{k=1}^r:$
(5)
$$\alpha=\operatorname{softmax}(\frac{1}{\sqrt{nr}}\sum\limits_{k}^{r}\sum\limits_{\ell}^{n}A_{\star,\ell}^{(k)})$$
中的显著单词，其中 softmax 函数输出单词上
的显著性分布：
(6)
$$\mathrm{softmax}(\boldsymbol{a}_\ell)=\frac{e^{\boldsymbol{a}_\ell}}{\sum_{\boldsymbol{a}_{\ell'}\in\boldsymbol{a}}e^{\boldsymbol{a}_{\ell'}}}$$
并获得句子嵌入$g=Z\alpha$
同样，我们采用另一种自我关注的编码器来获取上下文化的句子表示$H\in\mathbb{R}^{d_h\times s}$。最后一层输出多头注意力得分矩阵$\mathcal{B}=\{\boldsymbol{B}^{(k)}\}_{k=1}^r$ (带有 $B^(k)\in\mathbb{R}^{s\times s})$, 我们将句子显着性计算为：
(7)
$$\boldsymbol{\beta}=\mathrm{softmax}(\frac{1}{\sqrt{sr}}\sum_{k}^{r}\sum_{j}^{s}\boldsymbol{B}_{\star,j}^{(k)}).$$
Prior Information 除了文档(及其域标签)之外，我们可能还具有有关域的一些先验知识， 例如。g。,它的一般语义内容和与之相关的各种主题。例如，我们可能期望“生活方式”领域的文章不会谈论导弹或战争，因为这些是“军事”领域的经常性主题。如前所述，在本文中， 我们假设我们有域定义$\overline {\mathrm{E} }\mathcal{U}$ 用几句话表示为先验知识。域定义与 WoRDENC 和 SENTENC 共享参数，并在定义矩阵$U\in\mathbb{R}^d_h\times C$中编码
从直觉上讲，确定单词的域可能比句子的域更难；由于更长和更具表现力，句子比含义通常依赖于支持上下文的单词提供了更多与领域相关的线索。因此，我们仅将域定义$\operatorname{E}\boldsymbol{U}$ 注人到单词检测器中
## 5 The Detector Module
基于 DETNET 级别的分数，最终采用了与单
词，句子和文档相对应的三个检测器：
$$P=\mathrm{WORDDET}(Z,U)$$
$Q^{instc}=\begin{bmatrix}\boldsymbol{q}_1^{instc};\ldots;\boldsymbol{q}_s^{instc}\end{bmatrix}$where $\boldsymbol q^instc=\boldsymbol{P}\alpha$
$$\begin{aligned}\boldsymbol{I}_{s}\\\boldsymbol{Q}&=\mathrm{SENTDET}(\boldsymbol{Q}^{instc},\boldsymbol{H})\\\tilde{\boldsymbol{y}}&=\mathrm{DocDET}(\boldsymbol{Q},\boldsymbol{\beta})\end{aligned}$$
WORDDET 首先使用词汇语义信息$\overline{\mathrm{E}}Z$ 和先前(领域)知识$\mathbb{E}U$ 来产生 XMATHX 域分数； XMATHX
|
[
[
0.1175537109375,
0.07598876953125,
0.491943359375,
0.0736083984375,
0.491943359375,
0.0914306640625,
0.1175537109375,
0.09320068359375
],
[
0.51123046875,
0.0748291015625,
0.86572265625,
0.07659912109375,
0.86572265625,
0.0914306640625,
0.51123046875,
0.08905029296875
],
[
0.85986328125,
0.08013916015625,
0.880859375,
0.08013916015625,
0.880859375,
0.0872802734375,
0.85986328125,
0.0872802734375
],
[
0.12176513671875,
0.09320068359375,
0.489501953125,
0.09320068359375,
0.489501953125,
0.10687255859375,
0.12176513671875,
0.10687255859375
],
[
0.5146484375,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10687255859375,
0.5146484375,
0.10687255859375
],
[
0.12176513671875,
0.10986328125,
0.156982421875,
0.10986328125,
0.156982421875,
0.1217041015625,
0.12176513671875,
0.1217041015625
],
[
0.51318359375,
0.10687255859375,
0.88330078125,
0.105712890625,
0.88330078125,
0.12353515625,
0.51318359375,
0.12469482421875
],
[
0.1376953125,
0.1270751953125,
0.486083984375,
0.1270751953125,
0.486083984375,
0.13720703125,
0.1376953125,
0.13720703125
],
[
0.515625,
0.1258544921875,
0.8818359375,
0.1258544921875,
0.8818359375,
0.1395263671875,
0.515625,
0.1395263671875
],
[
0.12176513671875,
0.141357421875,
0.490234375,
0.141357421875,
0.490234375,
0.1561279296875,
0.12176513671875,
0.1561279296875
],
[
0.51318359375,
0.14013671875,
0.880859375,
0.14013671875,
0.880859375,
0.15380859375,
0.51318359375,
0.15380859375
],
[
0.12176513671875,
0.1591796875,
0.486083984375,
0.1591796875,
0.486083984375,
0.169189453125,
0.12176513671875,
0.169189453125
],
[
0.509765625,
0.15380859375,
0.7138671875,
0.1561279296875,
0.7138671875,
0.17333984375,
0.509765625,
0.171630859375
],
[
0.118408203125,
0.17333984375,
0.486083984375,
0.17333984375,
0.486083984375,
0.18701171875,
0.118408203125,
0.18701171875
],
[
0.118408203125,
0.188232421875,
0.486083984375,
0.188232421875,
0.486083984375,
0.201904296875,
0.118408203125,
0.201904296875
],
[
0.57275390625,
0.193603515625,
0.69189453125,
0.193603515625,
0.69189453125,
0.2078857421875,
0.57275390625,
0.2078857421875
],
[
0.6962890625,
0.188232421875,
0.70947265625,
0.188232421875,
0.70947265625,
0.196533203125,
0.6962890625,
0.196533203125
],
[
0.7255859375,
0.193603515625,
0.7314453125,
0.193603515625,
0.7314453125,
0.198974609375,
0.7255859375,
0.198974609375
],
[
0.767578125,
0.19482421875,
0.82373046875,
0.196533203125,
0.82373046875,
0.2078857421875,
0.767578125,
0.2060546875
],
[
0.8076171875,
0.1953125,
0.8154296875,
0.1953125,
0.8154296875,
0.1995849609375,
0.8076171875,
0.1995849609375
],
[
0.67919921875,
0.201904296875,
0.72119140625,
0.196533203125,
0.7255859375,
0.213134765625,
0.68408203125,
0.218505859375
],
[
0.85986328125,
0.196533203125,
0.8818359375,
0.196533203125,
0.8818359375,
0.207275390625,
0.85986328125,
0.207275390625
],
[
0.12005615234375,
0.2060546875,
0.48779296875,
0.2060546875,
0.48779296875,
0.2197265625,
0.12005615234375,
0.2197265625
],
[
0.759765625,
0.203125,
0.77099609375,
0.203125,
0.77099609375,
0.207275390625,
0.759765625,
0.207275390625
],
[
0.72705078125,
0.213134765625,
0.74658203125,
0.213134765625,
0.74658203125,
0.22265625,
0.72705078125,
0.22265625
],
[
0.75390625,
0.2120361328125,
0.771484375,
0.2120361328125,
0.771484375,
0.221435546875,
0.75390625,
0.221435546875
],
[
0.12176513671875,
0.221435546875,
0.48779296875,
0.221435546875,
0.48779296875,
0.235107421875,
0.12176513671875,
0.235107421875
],
[
0.5146484375,
0.23095703125,
0.8818359375,
0.232177734375,
0.8818359375,
0.2464599609375,
0.5146484375,
0.2452392578125
],
[
0.12005615234375,
0.2381591796875,
0.489501953125,
0.2381591796875,
0.489501953125,
0.251708984375,
0.12005615234375,
0.251708984375
],
[
0.12176513671875,
0.254150390625,
0.486083984375,
0.254150390625,
0.486083984375,
0.267333984375,
0.12176513671875,
0.267333984375
],
[
0.5146484375,
0.248779296875,
0.63232421875,
0.248779296875,
0.63232421875,
0.262451171875,
0.5146484375,
0.262451171875
],
[
0.12176513671875,
0.270751953125,
0.48193359375,
0.270751953125,
0.48193359375,
0.28076171875,
0.12176513671875,
0.28076171875
],
[
0.74462890625,
0.270751953125,
0.77099609375,
0.270751953125,
0.77099609375,
0.283203125,
0.74462890625,
0.283203125
],
[
0.59033203125,
0.27978515625,
0.705078125,
0.27978515625,
0.705078125,
0.29443359375,
0.59033203125,
0.29443359375
],
[
0.85791015625,
0.27978515625,
0.88330078125,
0.27978515625,
0.88330078125,
0.29345703125,
0.85791015625,
0.29345703125
],
[
0.12005615234375,
0.2861328125,
0.468505859375,
0.2861328125,
0.468505859375,
0.2998046875,
0.12005615234375,
0.2998046875
],
[
0.7119140625,
0.2861328125,
0.80615234375,
0.2861328125,
0.80615234375,
0.3046875,
0.7119140625,
0.3046875
],
[
0.1234130859375,
0.311279296875,
0.486083984375,
0.311279296875,
0.486083984375,
0.32470703125,
0.1234130859375,
0.32470703125
],
[
0.5205078125,
0.319580078125,
0.53662109375,
0.319580078125,
0.53662109375,
0.32666015625,
0.5205078125,
0.32666015625
],
[
0.53076171875,
0.31640625,
0.7314453125,
0.31640625,
0.7314453125,
0.330810546875,
0.53076171875,
0.330810546875
],
[
0.12005615234375,
0.327880859375,
0.48779296875,
0.327880859375,
0.48779296875,
0.341552734375,
0.12005615234375,
0.341552734375
],
[
0.53564453125,
0.334228515625,
0.880859375,
0.334228515625,
0.880859375,
0.344482421875,
0.53564453125,
0.344482421875
],
[
0.12176513671875,
0.34326171875,
0.489501953125,
0.34326171875,
0.489501953125,
0.35693359375,
0.12176513671875,
0.35693359375
],
[
0.51318359375,
0.345703125,
0.88330078125,
0.345703125,
0.88330078125,
0.363525390625,
0.51318359375,
0.363525390625
],
[
0.12005615234375,
0.359375,
0.48779296875,
0.359375,
0.48779296875,
0.372314453125,
0.12005615234375,
0.372314453125
],
[
0.5146484375,
0.3623046875,
0.86572265625,
0.364013671875,
0.86572265625,
0.37890625,
0.5146484375,
0.377197265625
],
[
0.12005615234375,
0.3759765625,
0.48779296875,
0.3759765625,
0.48779296875,
0.3896484375,
0.12005615234375,
0.3896484375
],
[
0.861328125,
0.37060546875,
0.875,
0.372314453125,
0.873046875,
0.377197265625,
0.85986328125,
0.374755859375
],
[
0.51123046875,
0.37890625,
0.8818359375,
0.380126953125,
0.8818359375,
0.394287109375,
0.51123046875,
0.393798828125
],
[
0.12176513671875,
0.391357421875,
0.489501953125,
0.391357421875,
0.489501953125,
0.405029296875,
0.12176513671875,
0.405029296875
],
[
0.5146484375,
0.3984375,
0.5439453125,
0.3984375,
0.5439453125,
0.410888671875,
0.5146484375,
0.410888671875
],
[
0.12176513671875,
0.4091796875,
0.484375,
0.4091796875,
0.484375,
0.419189453125,
0.12176513671875,
0.419189453125
],
[
0.118408203125,
0.42333984375,
0.48779296875,
0.42333984375,
0.48779296875,
0.43701171875,
0.118408203125,
0.43701171875
],
[
0.6943359375,
0.42578125,
0.7060546875,
0.42578125,
0.7060546875,
0.435302734375,
0.6943359375,
0.435302734375
],
[
0.57275390625,
0.432861328125,
0.67822265625,
0.432861328125,
0.67822265625,
0.44482421875,
0.57275390625,
0.44482421875
],
[
0.71533203125,
0.4287109375,
0.82861328125,
0.427490234375,
0.82861328125,
0.443603515625,
0.71533203125,
0.44482421875
],
[
0.118408203125,
0.439453125,
0.48779296875,
0.438232421875,
0.48779296875,
0.453125,
0.118408203125,
0.45361328125
],
[
0.85986328125,
0.43408203125,
0.87890625,
0.43408203125,
0.87890625,
0.443603515625,
0.85986328125,
0.443603515625
],
[
0.68115234375,
0.440673828125,
0.7255859375,
0.440673828125,
0.7255859375,
0.4560546875,
0.68115234375,
0.4560546875
],
[
0.72119140625,
0.44775390625,
0.767578125,
0.449462890625,
0.765625,
0.4619140625,
0.7197265625,
0.460205078125
],
[
0.118408203125,
0.4560546875,
0.486083984375,
0.4560546875,
0.486083984375,
0.4697265625,
0.118408203125,
0.4697265625
],
[
0.12005615234375,
0.473876953125,
0.175537109375,
0.473876953125,
0.175537109375,
0.485107421875,
0.12005615234375,
0.485107421875
],
[
0.51318359375,
0.47021484375,
0.8818359375,
0.471435546875,
0.8818359375,
0.486328125,
0.51318359375,
0.485107421875
],
[
0.1376953125,
0.488037109375,
0.489501953125,
0.488037109375,
0.489501953125,
0.501953125,
0.1376953125,
0.501953125
],
[
0.515625,
0.488037109375,
0.8818359375,
0.488037109375,
0.8818359375,
0.501953125,
0.515625,
0.501953125
],
[
0.12005615234375,
0.50439453125,
0.489501953125,
0.50439453125,
0.489501953125,
0.51708984375,
0.12005615234375,
0.51708984375
],
[
0.5146484375,
0.50439453125,
0.8818359375,
0.50439453125,
0.8818359375,
0.5185546875,
0.5146484375,
0.5185546875
],
[
0.12005615234375,
0.521484375,
0.484375,
0.521484375,
0.484375,
0.5322265625,
0.12005615234375,
0.5322265625
],
[
0.515625,
0.51953125,
0.88330078125,
0.51953125,
0.88330078125,
0.533203125,
0.515625,
0.533203125
],
[
0.12005615234375,
0.53515625,
0.486083984375,
0.53515625,
0.486083984375,
0.548828125,
0.12005615234375,
0.548828125
],
[
0.5146484375,
0.5361328125,
0.88330078125,
0.5361328125,
0.88330078125,
0.5498046875,
0.5146484375,
0.5498046875
],
[
0.12005615234375,
0.5517578125,
0.330078125,
0.5517578125,
0.330078125,
0.5654296875,
0.12005615234375,
0.5654296875
],
[
0.515625,
0.5517578125,
0.8818359375,
0.5517578125,
0.8818359375,
0.5654296875,
0.515625,
0.5654296875
],
[
0.515625,
0.568359375,
0.88330078125,
0.568359375,
0.88330078125,
0.58203125,
0.515625,
0.58203125
],
[
0.361083984375,
0.57763671875,
0.41650390625,
0.57763671875,
0.41650390625,
0.5927734375,
0.361083984375,
0.5927734375
],
[
0.14111328125,
0.5849609375,
0.364501953125,
0.5849609375,
0.364501953125,
0.599609375,
0.14111328125,
0.599609375
],
[
0.41064453125,
0.58740234375,
0.486083984375,
0.58740234375,
0.486083984375,
0.5986328125,
0.41064453125,
0.5986328125
],
[
0.51318359375,
0.58447265625,
0.88330078125,
0.5830078125,
0.88330078125,
0.59716796875,
0.51318359375,
0.5986328125
],
[
0.36865234375,
0.59423828125,
0.405517578125,
0.59423828125,
0.405517578125,
0.60986328125,
0.36865234375,
0.60986328125
],
[
0.51318359375,
0.5966796875,
0.85791015625,
0.5966796875,
0.85791015625,
0.6142578125,
0.51318359375,
0.6142578125
],
[
0.139404296875,
0.62158203125,
0.486083984375,
0.62158203125,
0.486083984375,
0.63525390625,
0.139404296875,
0.63525390625
],
[
0.5341796875,
0.61767578125,
0.88330078125,
0.61767578125,
0.88330078125,
0.63134765625,
0.5341796875,
0.63134765625
],
[
0.515625,
0.6328125,
0.880859375,
0.6328125,
0.880859375,
0.646484375,
0.515625,
0.646484375
],
[
0.12176513671875,
0.6396484375,
0.48779296875,
0.6396484375,
0.48779296875,
0.65283203125,
0.12176513671875,
0.65283203125
],
[
0.51318359375,
0.64794921875,
0.880859375,
0.64794921875,
0.880859375,
0.6611328125,
0.51318359375,
0.6611328125
],
[
0.1234130859375,
0.65673828125,
0.486083984375,
0.65673828125,
0.486083984375,
0.66748046875,
0.1234130859375,
0.66748046875
],
[
0.515625,
0.6650390625,
0.8818359375,
0.6650390625,
0.8818359375,
0.6787109375,
0.515625,
0.6787109375
],
[
0.1234130859375,
0.6728515625,
0.16455078125,
0.6728515625,
0.16455078125,
0.68408203125,
0.1234130859375,
0.68408203125
],
[
0.515625,
0.68115234375,
0.64404296875,
0.68115234375,
0.64404296875,
0.69482421875,
0.515625,
0.69482421875
],
[
0.118408203125,
0.69482421875,
0.37451171875,
0.69482421875,
0.37451171875,
0.71630859375,
0.118408203125,
0.71630859375
],
[
0.35107421875,
0.69482421875,
0.489501953125,
0.69482421875,
0.489501953125,
0.71630859375,
0.35107421875,
0.71630859375
],
[
0.515625,
0.71044921875,
0.52880859375,
0.71044921875,
0.52880859375,
0.71875,
0.515625,
0.71875
],
[
0.5263671875,
0.71044921875,
0.54833984375,
0.71044921875,
0.54833984375,
0.71875,
0.5263671875,
0.71875
],
[
0.54736328125,
0.70703125,
0.73486328125,
0.7080078125,
0.73486328125,
0.72265625,
0.54736328125,
0.7216796875
],
[
0.11248779296875,
0.71875,
0.489501953125,
0.72021484375,
0.489501953125,
0.73828125,
0.11248779296875,
0.736328125
],
[
0.515625,
0.73291015625,
0.8818359375,
0.73291015625,
0.8818359375,
0.74658203125,
0.515625,
0.74658203125
],
[
0.1368408203125,
0.75048828125,
0.489501953125,
0.75048828125,
0.489501953125,
0.76416015625,
0.1368408203125,
0.76416015625
],
[
0.5146484375,
0.75,
0.8154296875,
0.75,
0.8154296875,
0.76318359375,
0.5146484375,
0.76318359375
],
[
0.1259765625,
0.755859375,
0.1368408203125,
0.755859375,
0.1368408203125,
0.76123046875,
0.1259765625,
0.76123046875
],
[
0.118408203125,
0.7666015625,
0.486083984375,
0.7666015625,
0.486083984375,
0.77978515625,
0.118408203125,
0.77978515625
],
[
0.70361328125,
0.775390625,
0.884765625,
0.775390625,
0.884765625,
0.7890625,
0.70361328125,
0.7890625
],
[
0.12005615234375,
0.783203125,
0.48779296875,
0.783203125,
0.48779296875,
0.79638671875,
0.12005615234375,
0.79638671875
],
[
0.12005615234375,
0.798828125,
0.279541015625,
0.79638671875,
0.279541015625,
0.81103515625,
0.12005615234375,
0.8134765625
],
[
0.515625,
0.79443359375,
0.88671875,
0.79443359375,
0.88671875,
0.8125,
0.515625,
0.8125
],
[
0.6767578125,
0.81787109375,
0.884765625,
0.81640625,
0.884765625,
0.83056640625,
0.6767578125,
0.83203125
],
[
0.189697265625,
0.82763671875,
0.489501953125,
0.82763671875,
0.489501953125,
0.8427734375,
0.189697265625,
0.8427734375
],
[
0.72705078125,
0.83837890625,
0.884765625,
0.83837890625,
0.884765625,
0.8515625,
0.72705078125,
0.8515625
],
[
0.1376953125,
0.85693359375,
0.484375,
0.85693359375,
0.484375,
0.87060546875,
0.1376953125,
0.87060546875
],
[
0.51318359375,
0.86083984375,
0.88330078125,
0.8623046875,
0.88330078125,
0.876953125,
0.51318359375,
0.8759765625
],
[
0.12005615234375,
0.87353515625,
0.489501953125,
0.87353515625,
0.489501953125,
0.88720703125,
0.12005615234375,
0.88720703125
],
[
0.5146484375,
0.8798828125,
0.88330078125,
0.8798828125,
0.88330078125,
0.8935546875,
0.5146484375,
0.8935546875
],
[
0.14111328125,
0.89453125,
0.465087890625,
0.89794921875,
0.465087890625,
0.9111328125,
0.14111328125,
0.90771484375
],
[
0.5146484375,
0.896484375,
0.59423828125,
0.896484375,
0.59423828125,
0.90771484375,
0.5146484375,
0.90771484375
]
] |
[
"9，然后将其作为输入提供给句子级编码器，",
"Rdxn。基于Ez，单词编码器中的最后",
"",
"该编码器输出上下文化表示H及其注意力权",
"个多头注意力层针对每个句子产生一组注意力",
"重β",
"矩阵A=[A()=1 ，其中 A(h) IRnxn。因",
"在这项工作中，我们的目标是对相当长的",
"此，在衡量单词对句子的贡献时，e。g。，就领",
"文档(e。g。，维基百科文章；有关详细信息，",
"域得分和表示形式而言，我们可以有选择地关",
"请参阅第6节。因此，我们的编码器基于",
"注集合 A= A())K=1 :",
"Transformer架构（Vaswanietal.，2017），这是",
"最近提出的高效模型，在机器翻译（Vaswani",
"α = softmax(-",
"1",
"1",
"A",
"L",
"Vnr",
"(5)",
"etal.,2017）和问答（Yuetal.，2018）中实现了最",
"",
"k",
"l",
"先进的性能。Transformer在减少顺序计算的",
"中的显著单词，其中softmax函数输出单词上",
"基本约束，这是基于递归神经网络的大多数体",
"系结构的基础。它消除了复发，有利于应用自",
"的显著性分布：",
"我注意机制该机制直接对句子中所有单词之",
"eal",
"softmax(ae) =",
"(6)",
"间的关系进行建模，而不管它们的位置如何",
"Lavcaeae",
"Self-AttentiveEncoder如图E2所示，Trans-",
"",
"并获得句子嵌入g=Zα",
"former是一个非递归框架，包括mE相同的",
"同样，我们采用另一种自我关注的编码器来",
"层。通过使用添加到输入嵌入的位置编码来表",
"获取上下文化的句子表示HERdrxs。最后",
"示序列中每个令牌的（相对或绝对）位置的信",
"一层输出多头注意力得分矩阵B={B()",
"息（请参见图2的底部）。我们用X表示句",
"",
"（带有B()ERsxs），我们将句子显着性计算",
"子中的位置增强输入。我们的模型在单词和句",
"为:",
"子编码器中使用了四层。前三层与Transformel",
"（㎡=3）中的层相同，包括一个多头自关注子",
"1",
"β= softmax(",
"B():",
"层和一个位置全连接的前馈网络。最后一层只",
"(7",
"Vsr",
"kj",
"是一个多头自我关注层，为后续操作产生注意",
"力权重",
"PriorInformation除了文档（及其域标签）之",
"单头注意将三个参数作为Transformer",
"外，我们可能还具有有关域的一些先验知识，",
"(Vaswanietal.,2017）中的输人：查询矩阵，键",
"例如。g。，它的一般语义内容和与之相关的各",
"矩阵和值矩阵。这三个矩阵是相同的，并且等",
"种主题。例如，我们可能期望“生活方式”领域",
"于字编码器第一层的输入EX。单头注意力",
"的文章不会谈论导弹或战争，因为这些是“军",
"的输出通过以下方式计算：",
"事”领域的经常性主题。如前所述，在本文中，",
"我们假设我们有域定义E用几句话表示为先",
"XXT",
"head(X, X, X) = softmax(",
")X. (1)",
"验知识。域定义与WORDENC和SENTENC共",
"Vd.",
"享参数，并在定义矩阵UERdhxC 中编码",
"多头注意力允许在不同位置共同关注来自不",
"从直觉上讲，确定单词的域可能比句子的域",
"更难；由于更长和更具表现力，句子比含义通",
"同表示子空间的信息。这是通过首先对输入应",
"常依赖于支持上下文的单词提供了更多与领域",
"用不同的线性投影，然后将它们连接起来来完",
"相关的线索。因此，我们仅将域定义EU注人",
"成的：",
"到单词检测器中",
"head() = head(xw(), Xw(k",
"(), Xw(*) (2)",
"5",
"",
"The Detector Module",
"multi-head=concat(head(1), .,head()W4 (3)",
"基于DETNET级别的分数，最终采用了与单",
"其中我们为单词和句子编码器采用四个头",
"词，句子和文档相对应的三个检测器：",
"",
"（r=4）。Transformer中的第二个子层（见图",
"P = WORDDET(Z,U)",
"2）是一个完全连接的前馈网络，分别相同地",
"应用于每个位置：2",
"Qinstc - [qinstc,...; qinste] where qinstc - Pa",
"Q = SENTDET(Qinstc,H)",
"FFN(α) = max(O, aWs)W6.(4)",
"y = DOcDET(Q,β)",
"在通过前三层对输人嵌入进行顺序编码之",
"WORDDET首先使用词汇语义信息EZ和先",
"后，我们获得了上下文化的单词表示ZE",
"前（领域）知识EU来产生XMATHX域分数；",
"2We omit herethe bias termfor the sake of simplicity",
"XMATHX"
] |
{
"arxiv_id": "1907.11499",
"language": "zh",
"page_idx": 3
}
|
||
1907.11499-zh-4
|
1907.11499-zh-4.jpg
|
<FigureHere>
Figure 3: Domain predictions for words and sentences; the instance-bag relation applies to wordssentences (red shadow) and sentences-documents (green shadow). Squares denote representations of words or sentences, and circles are domain scores.
Word Detector 我们的第一个检测器产生单词域分数。对于一个句子，我们使用其自身的上下文单词语义信息获得一个自评分矩阵$\mathbb{E}$ $\overline{P^{self}}:$
$$P^{self}=\tanh(W_zZ).$$
(8)
与 Angelidis and Lapata (2018) 中使用的表示形式相反，我们从上下文化表示形式生成实例分数，i。e。, $Z$ 。由于 softmax 函数通常倾向于单模式输出，因此我们采用$\tanh(\cdot)\in(-1,1)$ 作为域评分函数，以根据多标签方案定制 MIL。如前所述，我们采用域定义作为单词级别的先验信息，并通过以下方式计算先验分数：
(9)
$$P^{prior}=\tanh(\max(0,U^\intercal W_u)Z)$$
,其中$W_u\in\mathbb{R}^{d_u\times d_z}$将先验信息$U$投射到输人语义空间上。先验得分矩阵$P^{prior}$捕获域定义和句子内容之间的相互作用
在这项工作中，我们灵活地将评分组件与门集成在一起，如图$\boxed{\mathrm{E}}$3 所示。关键思想是通过以下方式学习先验门$\operatorname{E\Gamma}$平衡方程$\operatorname{E}(8)$ 和$\operatorname{E}(9)$ ):
(10)
$$\Gamma=\gamma\sigma(\boldsymbol{W}_{g,p}[\boldsymbol{Z},\boldsymbol{P}^{self},\boldsymbol{P}^{prior}])\\\boldsymbol{P}=\boldsymbol{\Gamma}\odot\boldsymbol{P}^{prior}+(\boldsymbol{J}-\boldsymbol{\Gamma})\odot\boldsymbol{P}^{self}$$
(11)
其中$J$是全 1 矩阵，$P\in\mathbb{R}^{C\times n}$是单词级的最终域得分矩阵；$\odot$E表示元素乘法 XMATHX- $4。\sigma(\cdot)\in(0,1)$是 sigmoid 函数，$\Gamma\in(0,\gamma)$是具有缩放因子$\overline{\mathrm{E}}\gamma$ 的先验门，这是一个控制先验信息和实例的整体效果的超参数。$^{3}$
${^3}$Initially, we expected to balance these effects by purely relying on the learned function without a scaling factor. This, however, led to poor performance.
Sentence Detector 第二个检测器基于来自句子编码器的信号，先验信息和单词实例来识别具有领域重语义的句子。我们再次通过以下方式获得自评分矩阵$\mathbb{E}Q^{self}:$
$$Q^{self}=\tanh(\boldsymbol{W}_h\boldsymbol{H}).$$
(12)
从句子级信号计算句子分数后，我们从单个单词估计域分数。我们通过在方程(5), $q^{instc}=P\alpha$中重用$\alpha$来做到这一点。在为每个句子收集$q^instc$之后，我们获得$Q^instc\in$ $\mathbb{R}^{C\times s}$作为完整的实例得分矩阵
类似于单词级检测器(请参见公式$\mathbb{E}(10)).$ 我们采用句子级向上门$\Lambda$将领域得分从下游单词实例动态传播到句子包：
(13)
$$\boldsymbol{\Lambda}=\lambda\sigma(\boldsymbol{W}_\ell[\boldsymbol{H},\boldsymbol{Q}^{instc},\boldsymbol{Q}^{self}])\\\boldsymbol{Q}=\boldsymbol{\Lambda}\odot\boldsymbol{Q}^{instc}+(\boldsymbol{J}-\boldsymbol{\Lambda})\odot\boldsymbol{Q}^{self}$$
(14)
其中$Q$是最终句子得分矩阵。
Document Detector 文档级域分数基于句子显著性分布$\beta$ (请参见公式(7)),并计算为句子分数的加权平均值：
$$\tilde{y}=Q\beta.$$
(15)
我们仅使用文档级监督进行$C$域中的多标签
学习。形式上，我们的训练目标是：
$$\mathcal{L}=\min-\dfrac{1}{N}\sum_{i}^{N}\sum_{c}^{C}\log(1+e^{-\tilde{\boldsymbol{y}}_{c}^{(i)}\boldsymbol{y}_{c}^{(i)}})$$
其中$N$是训练集大小。在测试时，我们将域划分为一个相关的集合和一个不相关的集合，用于看不见的样本。由于域评分函数是$\tanh ( \cdot ) \in ( - 1, 1)$ , 因此我们使用 0的阈值来校准$\tilde{y}$ 。
### 6 $\textbf{Experimental Setup}$
Datasets DETNET 在从维基百科$^{5}$创建的两个数据集上进行了英语和中文训练。$^{6}$维基百科文章是根据代表知识领域定义特征的类别层次结构组织的。我们递归地收集维基百科页面，首先根据它们与域名的匹配确定根类别。然后我们获得它们的子类别，这些子类别的子类别，等等。我们将与类别关联的所有页面视为其根类别的域的代表
$^{4}$If$\forall c\in[1,C]:\tilde{\boldsymbol{y}}_{c}<0$ holds, we set $\tilde{\boldsymbol{y}}_{c^{*}}=1$ and select
$c^*$ as $c^*=\arg\max_c\tilde{\boldsymbol{y}}_c$ to produce a positive prediction.
$^{5}$http://static.wikipedia.org/downloads/2008-06/en $^{6}\bar{\text{Available at https://github.com/yumoxu/detnet}}$
|
[
[
0.51708984375,
0.07659912109375,
0.8818359375,
0.07659912109375,
0.8818359375,
0.09027099609375,
0.51708984375,
0.09027099609375
],
[
0.51708984375,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10687255859375,
0.51708984375,
0.10687255859375
],
[
0.32421875,
0.1051025390625,
0.382080078125,
0.1051025390625,
0.382080078125,
0.11224365234375,
0.32421875,
0.11224365234375
],
[
0.33154296875,
0.1134033203125,
0.3720703125,
0.1134033203125,
0.3720703125,
0.12353515625,
0.33154296875,
0.12353515625
],
[
0.5146484375,
0.1092529296875,
0.880859375,
0.1092529296875,
0.880859375,
0.122314453125,
0.5146484375,
0.122314453125
],
[
0.359375,
0.1258544921875,
0.41064453125,
0.1258544921875,
0.41064453125,
0.133056640625,
0.359375,
0.133056640625
],
[
0.51318359375,
0.1258544921875,
0.73779296875,
0.122314453125,
0.73779296875,
0.13720703125,
0.51318359375,
0.14013671875
],
[
0.341796875,
0.133056640625,
0.35107421875,
0.133056640625,
0.35107421875,
0.135986328125,
0.341796875,
0.135986328125
],
[
0.3720703125,
0.1353759765625,
0.399658203125,
0.1353759765625,
0.399658203125,
0.1436767578125,
0.3720703125,
0.1436767578125
],
[
0.313232421875,
0.1478271484375,
0.376220703125,
0.1478271484375,
0.376220703125,
0.157958984375,
0.313232421875,
0.157958984375
],
[
0.27783203125,
0.1300048828125,
0.2939453125,
0.1300048828125,
0.2939453125,
0.1383056640625,
0.27783203125,
0.1383056640625
],
[
0.60791015625,
0.15087890625,
0.7900390625,
0.1484375,
0.7900390625,
0.17041015625,
0.60791015625,
0.1728515625
],
[
0.8486328125,
0.1561279296875,
0.88330078125,
0.1561279296875,
0.88330078125,
0.168701171875,
0.8486328125,
0.168701171875
],
[
0.340087890625,
0.17041015625,
0.366943359375,
0.17041015625,
0.366943359375,
0.1787109375,
0.340087890625,
0.1787109375
],
[
0.39794921875,
0.17041015625,
0.43994140625,
0.168701171875,
0.43994140625,
0.177001953125,
0.39794921875,
0.1787109375
],
[
0.210693359375,
0.18408203125,
0.264404296875,
0.18701171875,
0.2626953125,
0.19775390625,
0.2091064453125,
0.19482421875
],
[
0.2418212890625,
0.171630859375,
0.2568359375,
0.171630859375,
0.2568359375,
0.179931640625,
0.2418212890625,
0.179931640625
],
[
0.307373046875,
0.190673828125,
0.32568359375,
0.193603515625,
0.32421875,
0.19775390625,
0.306396484375,
0.19482421875
],
[
0.53564453125,
0.18701171875,
0.880859375,
0.18701171875,
0.880859375,
0.19775390625,
0.53564453125,
0.19775390625
],
[
0.21240234375,
0.19775390625,
0.240966796875,
0.19775390625,
0.240966796875,
0.204833984375,
0.21240234375,
0.204833984375
],
[
0.51708984375,
0.201904296875,
0.88330078125,
0.201904296875,
0.88330078125,
0.215576171875,
0.51708984375,
0.215576171875
],
[
0.515625,
0.2161865234375,
0.8818359375,
0.2161865234375,
0.8818359375,
0.23095703125,
0.515625,
0.23095703125
],
[
0.197265625,
0.232177734375,
0.2216796875,
0.232177734375,
0.2216796875,
0.2392578125,
0.197265625,
0.2392578125
],
[
0.515625,
0.2327880859375,
0.88330078125,
0.2327880859375,
0.88330078125,
0.2476806640625,
0.515625,
0.2476806640625
],
[
0.51123046875,
0.2464599609375,
0.765625,
0.248779296875,
0.765625,
0.2666015625,
0.51123046875,
0.26416015625
],
[
0.12005615234375,
0.262451171875,
0.48779296875,
0.262451171875,
0.48779296875,
0.275634765625,
0.12005615234375,
0.275634765625
],
[
0.5341796875,
0.2666015625,
0.8818359375,
0.2666015625,
0.8818359375,
0.27978515625,
0.5341796875,
0.27978515625
],
[
0.12005615234375,
0.27783203125,
0.48779296875,
0.27783203125,
0.48779296875,
0.29150390625,
0.12005615234375,
0.29150390625
],
[
0.5146484375,
0.283203125,
0.8818359375,
0.283203125,
0.8818359375,
0.29638671875,
0.5146484375,
0.29638671875
],
[
0.118408203125,
0.29345703125,
0.486083984375,
0.29345703125,
0.486083984375,
0.30712890625,
0.118408203125,
0.30712890625
],
[
0.5146484375,
0.297607421875,
0.7421875,
0.298583984375,
0.7421875,
0.3134765625,
0.5146484375,
0.312255859375
],
[
0.12005615234375,
0.311279296875,
0.48779296875,
0.311279296875,
0.48779296875,
0.32470703125,
0.12005615234375,
0.32470703125
],
[
0.118408203125,
0.325927734375,
0.484375,
0.325927734375,
0.484375,
0.339111328125,
0.118408203125,
0.339111328125
],
[
0.57861328125,
0.327880859375,
0.8212890625,
0.327880859375,
0.8212890625,
0.342529296875,
0.57861328125,
0.342529296875
],
[
0.84814453125,
0.330078125,
0.88330078125,
0.330078125,
0.88330078125,
0.342529296875,
0.84814453125,
0.342529296875
],
[
0.54150390625,
0.3486328125,
0.82373046875,
0.347412109375,
0.82373046875,
0.365234375,
0.54150390625,
0.366455078125
],
[
0.84814453125,
0.3515625,
0.88330078125,
0.3515625,
0.88330078125,
0.364013671875,
0.84814453125,
0.364013671875
],
[
0.1259765625,
0.36767578125,
0.486083984375,
0.36767578125,
0.486083984375,
0.38134765625,
0.1259765625,
0.38134765625
],
[
0.118408203125,
0.3818359375,
0.484375,
0.3818359375,
0.484375,
0.3955078125,
0.118408203125,
0.3955078125
],
[
0.515625,
0.38134765625,
0.75,
0.38134765625,
0.75,
0.394287109375,
0.515625,
0.394287109375
],
[
0.12176513671875,
0.399658203125,
0.48779296875,
0.399658203125,
0.48779296875,
0.413330078125,
0.12176513671875,
0.413330078125
],
[
0.5205078125,
0.40673828125,
0.8818359375,
0.40673828125,
0.8818359375,
0.42041015625,
0.5205078125,
0.42041015625
],
[
0.12176513671875,
0.41455078125,
0.17626953125,
0.41455078125,
0.17626953125,
0.42578125,
0.12176513671875,
0.42578125
],
[
0.2166748046875,
0.42578125,
0.3896484375,
0.4287109375,
0.3896484375,
0.446533203125,
0.2166748046875,
0.443603515625
],
[
0.515625,
0.42333984375,
0.8818359375,
0.42333984375,
0.8818359375,
0.43701171875,
0.515625,
0.43701171875
],
[
0.46337890625,
0.431640625,
0.489501953125,
0.431640625,
0.489501953125,
0.44482421875,
0.46337890625,
0.44482421875
],
[
0.515625,
0.439453125,
0.7060546875,
0.439453125,
0.7060546875,
0.453125,
0.515625,
0.453125
],
[
0.12176513671875,
0.454833984375,
0.489501953125,
0.454833984375,
0.489501953125,
0.468505859375,
0.12176513671875,
0.468505859375
],
[
0.66162109375,
0.4697265625,
0.73486328125,
0.4697265625,
0.73486328125,
0.48388671875,
0.66162109375,
0.48388671875
],
[
0.84814453125,
0.468505859375,
0.884765625,
0.468505859375,
0.884765625,
0.48388671875,
0.84814453125,
0.48388671875
],
[
0.12176513671875,
0.48681640625,
0.486083984375,
0.48681640625,
0.486083984375,
0.50048828125,
0.12176513671875,
0.50048828125
],
[
0.12176513671875,
0.47021484375,
0.484375,
0.47021484375,
0.484375,
0.48388671875,
0.12176513671875,
0.48388671875
],
[
0.12005615234375,
0.5029296875,
0.489501953125,
0.5029296875,
0.489501953125,
0.5166015625,
0.12005615234375,
0.5166015625
],
[
0.51318359375,
0.49951171875,
0.8818359375,
0.49951171875,
0.8818359375,
0.51318359375,
0.51318359375,
0.51318359375
],
[
0.12005615234375,
0.51953125,
0.489501953125,
0.51953125,
0.489501953125,
0.533203125,
0.12005615234375,
0.533203125
],
[
0.5146484375,
0.5166015625,
0.796875,
0.5166015625,
0.796875,
0.52978515625,
0.5146484375,
0.52978515625
],
[
0.12176513671875,
0.53515625,
0.48779296875,
0.53515625,
0.48779296875,
0.548828125,
0.12176513671875,
0.548828125
],
[
0.51123046875,
0.54052734375,
0.8427734375,
0.5224609375,
0.8486328125,
0.57763671875,
0.51708984375,
0.595703125
],
[
0.6396484375,
0.54345703125,
0.69189453125,
0.54345703125,
0.69189453125,
0.5546875,
0.6396484375,
0.5546875
],
[
0.12176513671875,
0.5517578125,
0.43994140625,
0.5517578125,
0.43994140625,
0.5654296875,
0.12176513671875,
0.5654296875
],
[
0.6572265625,
0.5634765625,
0.66748046875,
0.5634765625,
0.66748046875,
0.56787109375,
0.6572265625,
0.56787109375
],
[
0.8369140625,
0.55810546875,
0.87890625,
0.55810546875,
0.87890625,
0.5693359375,
0.8369140625,
0.5693359375
],
[
0.17041015625,
0.576171875,
0.438232421875,
0.57763671875,
0.438232421875,
0.595703125,
0.17041015625,
0.59326171875
],
[
0.4375,
0.5830078125,
0.459228515625,
0.5830078125,
0.459228515625,
0.59033203125,
0.4375,
0.59033203125
],
[
0.4609375,
0.58056640625,
0.48779296875,
0.58056640625,
0.48779296875,
0.59326171875,
0.4609375,
0.59326171875
],
[
0.45751953125,
0.58447265625,
0.46337890625,
0.58447265625,
0.46337890625,
0.59033203125,
0.45751953125,
0.59033203125
],
[
0.432373046875,
0.58447265625,
0.4375,
0.58447265625,
0.4375,
0.5888671875,
0.432373046875,
0.5888671875
],
[
0.133544921875,
0.60498046875,
0.489501953125,
0.60498046875,
0.489501953125,
0.62255859375,
0.133544921875,
0.62255859375
],
[
0.515625,
0.599609375,
0.8818359375,
0.599609375,
0.8818359375,
0.61328125,
0.515625,
0.61328125
],
[
0.1275634765625,
0.615234375,
0.139404296875,
0.615234375,
0.139404296875,
0.6181640625,
0.1275634765625,
0.6181640625
],
[
0.5146484375,
0.615234375,
0.8818359375,
0.615234375,
0.8818359375,
0.62890625,
0.5146484375,
0.62890625
],
[
0.515625,
0.6318359375,
0.8818359375,
0.6318359375,
0.8818359375,
0.6455078125,
0.515625,
0.6455078125
],
[
0.12005615234375,
0.62060546875,
0.489501953125,
0.62060546875,
0.489501953125,
0.63818359375,
0.12005615234375,
0.63818359375
],
[
0.12176513671875,
0.64013671875,
0.358642578125,
0.64013671875,
0.358642578125,
0.65380859375,
0.12176513671875,
0.65380859375
],
[
0.5146484375,
0.64794921875,
0.8818359375,
0.64794921875,
0.8818359375,
0.6611328125,
0.5146484375,
0.6611328125
],
[
0.1376953125,
0.65625,
0.486083984375,
0.65625,
0.486083984375,
0.66943359375,
0.1376953125,
0.66943359375
],
[
0.51123046875,
0.6630859375,
0.60205078125,
0.66015625,
0.603515625,
0.6787109375,
0.51318359375,
0.681640625
],
[
0.118408203125,
0.67138671875,
0.486083984375,
0.67138671875,
0.486083984375,
0.68505859375,
0.118408203125,
0.68505859375
],
[
0.1234130859375,
0.689453125,
0.489501953125,
0.689453125,
0.489501953125,
0.703125,
0.1234130859375,
0.703125
],
[
0.51708984375,
0.69677734375,
0.5283203125,
0.69677734375,
0.5283203125,
0.70361328125,
0.51708984375,
0.70361328125
],
[
0.54736328125,
0.69140625,
0.72314453125,
0.6923828125,
0.72314453125,
0.70703125,
0.54736328125,
0.7060546875
],
[
0.5341796875,
0.69775390625,
0.54736328125,
0.69775390625,
0.54736328125,
0.70166015625,
0.5341796875,
0.70166015625
],
[
0.12176513671875,
0.70703125,
0.1376953125,
0.70703125,
0.1376953125,
0.71728515625,
0.12176513671875,
0.71728515625
],
[
0.51904296875,
0.7197265625,
0.8818359375,
0.7197265625,
0.8818359375,
0.73291015625,
0.51904296875,
0.73291015625
],
[
0.160400390625,
0.73095703125,
0.4130859375,
0.728515625,
0.4130859375,
0.74658203125,
0.160400390625,
0.7490234375
],
[
0.453369140625,
0.73388671875,
0.48779296875,
0.73388671875,
0.48779296875,
0.74658203125,
0.453369140625,
0.74658203125
],
[
0.51708984375,
0.7353515625,
0.880859375,
0.7353515625,
0.880859375,
0.7490234375,
0.51708984375,
0.7490234375
],
[
0.145263671875,
0.75048828125,
0.4306640625,
0.7490234375,
0.4306640625,
0.76953125,
0.145263671875,
0.771484375
],
[
0.453369140625,
0.7548828125,
0.48779296875,
0.7548828125,
0.48779296875,
0.76708984375,
0.453369140625,
0.76708984375
],
[
0.515625,
0.751953125,
0.8818359375,
0.751953125,
0.8818359375,
0.765625,
0.515625,
0.765625
],
[
0.515625,
0.7685546875,
0.87744140625,
0.7685546875,
0.87744140625,
0.779296875,
0.515625,
0.779296875
],
[
0.1175537109375,
0.779296875,
0.489501953125,
0.77978515625,
0.489501953125,
0.79736328125,
0.1175537109375,
0.79638671875
],
[
0.5146484375,
0.783203125,
0.880859375,
0.783203125,
0.880859375,
0.79638671875,
0.5146484375,
0.79638671875
],
[
0.12005615234375,
0.798828125,
0.48779296875,
0.798828125,
0.48779296875,
0.8125,
0.12005615234375,
0.8125
],
[
0.515625,
0.7998046875,
0.8818359375,
0.7998046875,
0.8818359375,
0.8134765625,
0.515625,
0.8134765625
],
[
0.118408203125,
0.8154296875,
0.48779296875,
0.8154296875,
0.48779296875,
0.8291015625,
0.118408203125,
0.8291015625
],
[
0.5146484375,
0.8154296875,
0.8818359375,
0.8154296875,
0.8818359375,
0.8291015625,
0.5146484375,
0.8291015625
],
[
0.118408203125,
0.83056640625,
0.48779296875,
0.83056640625,
0.48779296875,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.5146484375,
0.83203125,
0.7001953125,
0.83203125,
0.7001953125,
0.845703125,
0.5146484375,
0.845703125
],
[
0.1175537109375,
0.845703125,
0.4248046875,
0.84326171875,
0.4248046875,
0.86083984375,
0.1175537109375,
0.86328125
],
[
0.5341796875,
0.85498046875,
0.88330078125,
0.85546875,
0.88330078125,
0.87060546875,
0.5341796875,
0.869140625
],
[
0.14111328125,
0.8681640625,
0.486083984375,
0.869140625,
0.486083984375,
0.88427734375,
0.14111328125,
0.8828125
],
[
0.51318359375,
0.8681640625,
0.85791015625,
0.869140625,
0.85791015625,
0.8828125,
0.51318359375,
0.8818359375
],
[
0.118408203125,
0.8828125,
0.48779296875,
0.8828125,
0.48779296875,
0.896484375,
0.118408203125,
0.896484375
],
[
0.53564453125,
0.8828125,
0.84033203125,
0.8828125,
0.84033203125,
0.896484375,
0.53564453125,
0.896484375
],
[
0.12005615234375,
0.896484375,
0.32666015625,
0.896484375,
0.32666015625,
0.91015625,
0.12005615234375,
0.91015625
],
[
0.5341796875,
0.89453125,
0.8251953125,
0.8955078125,
0.8251953125,
0.9091796875,
0.5341796875,
0.90771484375
]
] |
[
"SentenceDetector第二个检测器基于来自句",
"子编码器的信号，先验信息和单词实例来识别",
"Document",
"Score",
"具有领域重语义的句子。我们再次通过以下方",
"Sentence",
"式获得自评分矩阵EQself：",
"",
"Score",
"O",
"",
"Qself = tanh(WhH).",
"(12)",
"H",
"Oosel",
"WordC",
"Q",
"",
"从句子级信号计算子分数后，我们从单",
"Score",
"个单词估计域分数。我们通过在方程（5）",
"qinstc =Pα中重用 α来做到这一点。在为",
"",
"每个句子收集qinstc之后，我们获得QinstcE",
"RC×s作为完整的实例得分矩阵",
"Figure 3: Domain predictions for words and sen-",
"类似于单词级检测器（请参见公式E（10））",
"tences; the instance-bag relation applies to words-",
"我们采用句子级向上门△将领域得分从下游",
"sentences (red shadow) and sentences-documents",
"单词实例动态传播到句子包：",
"(green shadow). Squares denote representations of",
"words or sentences, and circles are domain scores",
"A = Ao(We[H,Qinstc, Qselfj)",
"(13)",
"Q= AoQinstc +(J - A)oQself",
"(14)",
"WordDetector我们的第一个检测器产生单",
"词域分数。对于一个句子，我们使用其自身的",
"其中Q是最终句子得分矩阵。",
"上下文单词语义信息获得一个自评分矩阵E",
"DocumentDetector文档级域分数基于句子",
"pself :",
"pself = tanh(W,Z).",
"显著性分布β（请参见公式（7）），并计算为",
"(8)",
"句子分数的加权平均值：",
"与AngelidisandLapata（2018）中使用的表示形",
"y = Qβ.",
"(15)",
"数，i。e。，Z。由于softmax函数通常倾向于单",
"式相反，我们从上下文化表示形式生成实例分",
"模式输出，因此我们采用tanh(）E（-1,1）作",
"我们仅使用文档级监督进行C域中的多标签",
"为域评分函数，以根据多标签方案定制MIL。",
"学习。形式上，我们的训练目标是：",
"如前所述，我们采用域定义作为单词级别的先",
"Cmm-a+,",
"NC",
"验信息，并通过以下方式计算先验分数：",
"",
"（16)",
"pprior = tanh(max(o,UTWu)z)",
"",
"(9)",
"",
"",
"其中WuERduxdz将先验信息U投射到输",
"其中N是训练集大小。在测试时，我们将",
"",
"域划分为一个相关的集合和一个不相关的集",
"合，用于看不见的样本。由于域评分函数是",
"人语义空间上。先验得分矩阵Pprior捕获域定",
"义和句子内容之间的相互作用",
"tanh(）E（-1,1），因此我们使用0的阈值来",
"在这项工作中，我们灵活地将评分组件与门",
"校准9。4",
"集成在一起，如图E3所示。关键思想是通过以",
"下方式学习先验门ET平衡方程E（8）和E（9",
"b",
"Experimental Setup",
"",
"):",
"DatasetsDETNET在从维基百科5创建的两",
"I = o(Wg,p[Z, pself, pprior])",
"(10)",
"个数据集上进行了英语和中文训练。6维基百",
"P=Io pprior +(J-)oPself",
"(11)",
"科文章是根据代表知识领域定义特征的类别",
"层次结构组织的。我们递归地收集维基百科页",
"其中J是全1矩阵，PERCxn是单词级的最",
"面，首先根据它们与域名的匹配确定根类别。",
"终域得分矩阵；E表示元素乘法XMATHX-",
"然后我们获得它们的子类别，这些子类别的子",
"4。α() E(0,1)是 sigmoid函数，TE(0,)是",
"类别，等等。我们将与类别关联的所有页面视",
"具有缩放因子的先验门，这是一个控制先",
"为其根类别的域的代表",
"验信息和实例的整体效果的超参数。3",
"*If Vc e [1, C] : ye < 0 holds, we set ye+ = 1 and select",
"Initially, we expected to balance these effects by purely",
"c\"as c*= arg maxyto produce a positive prediction.",
"relying on the learned function without a scaling factor. This,",
"http://static.wikipedia.org/downloads/2008-06/en",
"however, led to poor performance.",
"Available at https://github.com/yumoxu/detnet"
] |
{
"arxiv_id": "1907.11499",
"language": "zh",
"page_idx": 4
}
|
||
2110.09771-zh-0
|
2110.09771-zh-0.jpg
|
关于具有核和神经函数逼近的无奖励 RL: 单代理 MDP 和马
# 尔可夫博弈
ShuangQiu* Jieping Ye$^\dagger$ Zhaoran Wang$^{\ddagger }$ Zhuoran Yang$^{8}$
May 18, 2024
$\mathbf{Abstract}$
To achieve sample efficiency in reinforcement learning (RL), it necessitates efficiently exploring the underlying environment. Under the offline setting, addressing the exploration challenge lies in collecting an offline dataset with sufficient coverage. Motivated by such a challenge, we study the reward-free RL problem, where an agent aims to thoroughly explore the environment without any pre-specified reward function. Then, given any extrinsic reward, the agent computes the policy via a planning algorithm with offline data collected in the exploration phase. Moreover, we tackle this problem under the context of function approximation, leveraging powerful function approximators. Specifically, we propose to explore via an optimistic variant of the value-iteration algorithm incorporating kernel and neural function approximations, where we adopt the associated exploration bonus as the exploration reward. Moreover, we design exploration and planning algorithms for both single-agent MDPs and zero-sum Markov games and prove that our methods can achieve $\widetilde{\mathcal{O}}(1/\varepsilon^2)$ sample complexity for generating a $\varepsilon$-suboptimal policy or $\varepsilon$ -approximate Nash equilibrium when given an arbitrary extrinsic reward. To the best of our knowledge, we establish the first provably efficient reward-free RL algorithm with kernel and neural function approximators.
## 1 $\textbf{Introduction}$
虽然具有函数逼近的强化学习 (RL)在经验上取得了巨大的成功 (Mnih et al., 2015; Silver et al., 2016,2017; Vinyals et al., 2019) ,但其应用主要是通过与未知环境的大量交互来实现的，尤其是当状态空间很大并且使用了诸如神经网络之类的函数逼近器时。为了实现样本效率，任何 RL 算法都需要显式或隐式地准确学习过渡模型，这带来了高效探索的需要
在离线 RL 的设置下，代理旨在仅从先验收集的离线数据集中学习最佳策略，而无需与环境进行任何交互。因此，收集的离线数据集应具有由最佳策略生成的轨迹的足够覆盖范围。然而，在现实世界的 RL 应用中，奖励函数通常是由学习者根据领域知识设计的。学习者可能有一组奖励函数可供选择或使用奖励设计(Laud, 2004; Grzes, 2017) 的自适应算法。在这种情况下，通常希望收集
$^{*}$University of Michigan. Email: qiush@umich.edu $^{\dagger}$University of Michigan. Email: jpye@umich.edu. $^{\dagger}$Northwestern University. Email: zhaoranwang@gmail.com.
$^{6}$Princeton University. Email: zy6@princeton.edu.
1
|
[
[
0.134033203125,
0.147705078125,
0.86767578125,
0.147705078125,
0.86767578125,
0.1654052734375,
0.134033203125,
0.1654052734375
],
[
0.4306640625,
0.175537109375,
0.57177734375,
0.175537109375,
0.57177734375,
0.1962890625,
0.4306640625,
0.1962890625
],
[
0.21728515625,
0.2197265625,
0.330078125,
0.2197265625,
0.330078125,
0.2374267578125,
0.21728515625,
0.2374267578125
],
[
0.35693359375,
0.2203369140625,
0.4599609375,
0.2183837890625,
0.4599609375,
0.236083984375,
0.35693359375,
0.238037109375
],
[
0.49267578125,
0.2203369140625,
0.6259765625,
0.2203369140625,
0.6259765625,
0.23486328125,
0.49267578125,
0.23486328125
],
[
0.66015625,
0.2197265625,
0.78759765625,
0.2197265625,
0.78759765625,
0.2374267578125,
0.66015625,
0.2374267578125
],
[
0.453369140625,
0.2235107421875,
0.484375,
0.2235107421875,
0.484375,
0.2310791015625,
0.453369140625,
0.2310791015625
],
[
0.328369140625,
0.2247314453125,
0.3544921875,
0.2247314453125,
0.3544921875,
0.231689453125,
0.328369140625,
0.231689453125
],
[
0.44287109375,
0.25390625,
0.55615234375,
0.25390625,
0.55615234375,
0.268310546875,
0.44287109375,
0.268310546875
],
[
0.465576171875,
0.323974609375,
0.53515625,
0.323974609375,
0.53515625,
0.335205078125,
0.465576171875,
0.335205078125
],
[
0.1854248046875,
0.347900390625,
0.83837890625,
0.347900390625,
0.83837890625,
0.361083984375,
0.1854248046875,
0.361083984375
],
[
0.160888671875,
0.3642578125,
0.83837890625,
0.3642578125,
0.83837890625,
0.37744140625,
0.160888671875,
0.37744140625
],
[
0.16015625,
0.380126953125,
0.8388671875,
0.380126953125,
0.8388671875,
0.39453125,
0.16015625,
0.39453125
],
[
0.16015625,
0.397705078125,
0.83837890625,
0.397705078125,
0.83837890625,
0.410888671875,
0.16015625,
0.410888671875
],
[
0.16015625,
0.41357421875,
0.8388671875,
0.41357421875,
0.8388671875,
0.427978515625,
0.16015625,
0.427978515625
],
[
0.160888671875,
0.429931640625,
0.83837890625,
0.429931640625,
0.83837890625,
0.4443359375,
0.160888671875,
0.4443359375
],
[
0.158447265625,
0.4443359375,
0.8388671875,
0.4462890625,
0.8388671875,
0.4609375,
0.158447265625,
0.458984375
],
[
0.160888671875,
0.46337890625,
0.8408203125,
0.46337890625,
0.8408203125,
0.477783203125,
0.160888671875,
0.477783203125
],
[
0.16015625,
0.479248046875,
0.8388671875,
0.479736328125,
0.8388671875,
0.494384765625,
0.16015625,
0.49365234375
],
[
0.16015625,
0.496826171875,
0.83642578125,
0.496826171875,
0.83642578125,
0.51025390625,
0.16015625,
0.51025390625
],
[
0.160888671875,
0.51318359375,
0.8388671875,
0.51318359375,
0.8388671875,
0.52783203125,
0.160888671875,
0.52783203125
],
[
0.16015625,
0.52978515625,
0.8388671875,
0.52978515625,
0.8388671875,
0.54345703125,
0.16015625,
0.54345703125
],
[
0.16015625,
0.546875,
0.83837890625,
0.546875,
0.83837890625,
0.56005859375,
0.16015625,
0.56005859375
],
[
0.160888671875,
0.56298828125,
0.83837890625,
0.56298828125,
0.83837890625,
0.5771484375,
0.160888671875,
0.5771484375
],
[
0.16015625,
0.580078125,
0.461669921875,
0.580078125,
0.461669921875,
0.59326171875,
0.16015625,
0.59326171875
],
[
0.11358642578125,
0.6220703125,
0.1683349609375,
0.619140625,
0.1690673828125,
0.63525390625,
0.11517333984375,
0.63720703125
],
[
0.160888671875,
0.6220703125,
0.3095703125,
0.6220703125,
0.3095703125,
0.63623046875,
0.160888671875,
0.63623046875
],
[
0.11767578125,
0.6572265625,
0.8798828125,
0.6572265625,
0.8798828125,
0.671875,
0.11767578125,
0.671875
],
[
0.1168212890625,
0.67529296875,
0.88134765625,
0.67529296875,
0.88134765625,
0.689453125,
0.1168212890625,
0.689453125
],
[
0.1192626953125,
0.6943359375,
0.88134765625,
0.6943359375,
0.88134765625,
0.70849609375,
0.1192626953125,
0.70849609375
],
[
0.1168212890625,
0.7138671875,
0.6337890625,
0.7138671875,
0.6337890625,
0.72802734375,
0.1168212890625,
0.72802734375
],
[
0.1446533203125,
0.73291015625,
0.88134765625,
0.73291015625,
0.88134765625,
0.7470703125,
0.1446533203125,
0.7470703125
],
[
0.1168212890625,
0.751953125,
0.88134765625,
0.751953125,
0.88134765625,
0.76513671875,
0.1168212890625,
0.76513671875
],
[
0.11767578125,
0.77099609375,
0.88134765625,
0.77099609375,
0.88134765625,
0.7841796875,
0.11767578125,
0.7841796875
],
[
0.1192626953125,
0.7890625,
0.88134765625,
0.7890625,
0.88134765625,
0.80322265625,
0.1192626953125,
0.80322265625
],
[
0.1380615234375,
0.81494140625,
0.47216796875,
0.81494140625,
0.47216796875,
0.8291015625,
0.1380615234375,
0.8291015625
],
[
0.1380615234375,
0.830078125,
0.465576171875,
0.830078125,
0.465576171875,
0.84326171875,
0.1380615234375,
0.84326171875
],
[
0.136474609375,
0.84228515625,
0.52783203125,
0.84326171875,
0.52783203125,
0.85791015625,
0.136474609375,
0.85693359375
],
[
0.136474609375,
0.85791015625,
0.473876953125,
0.85791015625,
0.473876953125,
0.8720703125,
0.136474609375,
0.8720703125
],
[
0.491943359375,
0.89794921875,
0.50732421875,
0.89794921875,
0.50732421875,
0.912109375,
0.491943359375,
0.912109375
]
] |
[
"关于具有核和神经函数逼近的无奖励RL：单代理MDP和马",
"尔可夫博奔",
"Shuang Qiu*",
"Jieping Yet",
"Zhaoran Wangt",
"Zhuoran Yangs",
"",
"",
"May 18, 2024",
"Abstract",
"To achieve sample efficiency in reinforcement learning (RL), it necessitates efficiently explor-",
"ing the underlying environment. Under the offline setting, addressing the exploration challenge",
"lies in collecting an offline dataset with sufficient coverage. Motivated by such a challenge, we",
"study the reward-free RL problem, where an agent aims to thoroughly explore the environment",
"without any pre-specified reward function. Then, given any extrinsic reward, the agent com-",
"putes the policy via a planning algorithm with offline data collected in the exploration phase.",
"Moreover, we tackle this problem under the context of function approximation, leveraging pow-",
"erful function approximators. Specifically, we propose to explore via an optimistic variant of",
"the value-iteration algorithm incorporating kernel and neural function approximations, where",
"we adopt the associated exploration bonus as the exploration reward. Moreover, we design ex",
"ploration and planning algorithms for both single-agent MDPs and zero-sum Markov games and",
"prove that our methods can achieve O(1/2) sample complexity for generating a -suboptimal",
"policy or -approximate Nash equilibrium when given an arbitrary extrinsic reward. To the",
"best of our knowledge, we establish the first provably efficient reward-free RL algorithm with",
"kernel and neural function approximators.",
"1",
"Introduction",
"虽然具有函数逼近的强化学习（RL）在经验上取得了巨大的成功（Mnihetal.，2015；Silveretal.",
"2016，2017：Vinyalsetal.，2019），但其应用主要是通过与未知环境的大量交互来实现的，尤其是",
"当状态空间很大并且使用了诸如神经网络之类的函数逼近器时。为了实现样本效率，任何RL算法",
"都需要显式或隐式地准确学习过渡模型，这带来了高效探索的需要",
"在离线RL的设置下，代理旨在仅从先验收集的离线数据集中学习最佳策略，而无需与环境进",
"行任何交互。因此，收集的离线数据集应具有由最佳策略生成的轨迹的足够覆盖范围。然而，在现",
"实世界的RL应用中，奖励函数通常是由学习者根据领域知识设计的。学习者可能有一组奖励函数",
"可供选择或使用奖励设计（Laud，2004；Grzes，2017）的自适应算法。在这种情况下，通常希望收集",
"*University of Michigan.Email:qiush@umich.edu",
"fUniversityofMichigan.Email:jpye@umich.edu.",
"+Northwestern University.Email:zhaoranwanggmail.com",
"$Princeton University. Email: zy6@princeton.edu.",
"1"
] |
{
"arxiv_id": "2110.09771",
"language": "zh",
"page_idx": 0
}
|
||
2110.09771-zh-1
|
2110.09771-zh-1.jpg
|
一个离线数据集，该数据集涵盖与一组奖励函数相关的所有可能的最佳轨迹。有了这样一个良性的
离线数据集，对于任何任意的奖励函数，RL 代理都有足够的信息来估计相应的策略
要以原则性的方式研究这样的问题，我们将重点放在无奖励 RL 的框架上，该框架由探索阶段和计划阶段组成。具体来说，在探索阶段，代理在不访问任何预先指定的奖励的情况下与环境进行交互，并为随后的计划阶段收集经验轨迹。在计划阶段，使用在探索阶段收集的离线数据，代理在给定外部奖励函数时计算最佳策略，而无需与环境进行进一步的交互
最近，许多工作都集中在设计可证明的样本有效的无奖励 RL 算法上。对于单代理表格情况Jin et al. (2020a); Kaufmann et al. (2020); Ménard et al. (2020); Zhang et al. (2020) 实现了用于获得$\varepsilon$次优策略的 $\tilde{\mathcal{O}}($poly$(H,|\mathcal{S}|,|\mathcal{A}|)/\varepsilon^2)$ 样本复杂性，其中 $|\mathcal{S}|,|\mathcal{A}|$ 分别是状态和动作空间的大小。鉴于较大的动作和状态空间，作品 Zanette et al. (2020b); Wang et al. (2020a) 通过对单代理马尔可夫决策过程 (MDP) 应用线性函数逼近，从理论上分析了无奖励 RL, 从而实现了$\tilde{\mathcal{O}}($poly$(H,\mathfrak{o})/\varepsilon^2)$ 样本复杂度$\delta$表示特征空间的维度。然而，RL 算法结合非线性函数逼近器，如核函数逼近器和神经函数逼近器，在各种应用问题上都取得了巨大的经验成功 (e。g。, Duan et al. (2016); Silver et al. (2016,2017); Wang et al. (2018); Vinyals et al. (2019)),这要归功于它们的表达能力。另一方面，尽管在 Bai and Jin (2020); Liu et al. (2020)中已经研究了表格情况下多人马尔可夫游戏的无奖励 RL 算法，但仍然缺乏从理论上研究具有函数逼近的多智能体场景的工作。因此，以下问题仍然悬而未决：
Can we design provably efficient reward-free RL algorithms with kernel and neural function
approximations for both single-agent MDPs and Markov games?
回答上述问题的主要挑战在于如何将非线性逼近器适当地集成到无奖励 RL 的框架中，以及如何通过设计适合这种逼近的探索奖励和奖金来激励探索。在本文中，我们通过应对这些挑战为上述问题提供了肯定的答案。我们的贡献总结如下：
贡献。在本文中，我们首先针对单代理 MDP 设置提出了具有核和神经函数逼近的可证明样本和计算高效的无奖励 RL 算法。我们的探索算法是最小二乘值迭代算法的乐观变体，结合了核函数和神经函数逼近器，该算法采用关联的 (缩放的) 奖金作为探索奖励。在计划阶段，我们的方法实现了$\widetilde{\mathcal{O}}(1/\varepsilon^2)$ 样本复杂性，从而为任意外部奖励函数生成了$\varepsilon$ 次优策略。此外，我们将所提出的单代理设置方法扩展到零和马尔可夫博弈设置，使得该算法可以实现$\widetilde{\mathcal{O}}(1/\varepsilon^2)$样本复杂度来生成策略对， 该策略对是$\varepsilon$ -近似纳什均衡。特别是，在马尔可夫博弈的计划阶段，我们的算法仅涉及找到可以有效求解的 Q 函数形成的矩阵博弈的纳什均衡，这是独立的兴趣。我们方法的样本复杂性与表格或线性函数逼近设置的现有工作中的$\tilde{\mathcal{O}}(1/\varepsilon^2)$结果相匹配。据我们所知，我们建立了第一个可证明有效的无奖励 RL 算法，该算法具有用于单代理和多代理设置的内核和神经功能逼近器
相关工作。已经有很多工作集中在为单代理和多代理 RL 问题设计可证明有效的无奖励 RL 算法。对于单代理场景，Jin et al. (2020a)将表格设置的无奖励 RL 形式化，并以$\widetilde{\mathcal{O}}($poly$(H,|\mathcal{S}|,|\mathcal{A}|)/\varepsilon^2)$ 样本复杂度为所提出的算法提供理论分析，以实现$\varepsilon$次优策略。在最近的几部作品(Kaufmann et al., 2020; Ménard et al., 2020; Zhang et al., 2020) 中，表格设置的样本复杂性得到了进一步改
善。最近，Zanette et al. (2020b); Wang et al. (2020a) 从线性函数逼近的角度研究了无回报 RL。对于多主体设置，Bai and Jin (2020)研究了表格情况下零和马尔可夫博弈的无奖励探索。Liu et al. (2020)进一步提出了适用于多人一般和游戏的可证明的无奖励 RL 算法
我们的工作也与研究具有函数逼近的 RL 算法的一系列工作密切相关。有大量的著作 (Yang and Wang, 2019, 2020; Cai et al., 2019; Zanette et al., 2020a; Jin et al., 2020b; Wang et al., 2019;
2
|
[
[
0.11767578125,
0.0947265625,
0.8798828125,
0.0947265625,
0.8798828125,
0.10797119140625,
0.11767578125,
0.10797119140625
],
[
0.1192626953125,
0.11297607421875,
0.771484375,
0.11297607421875,
0.771484375,
0.126953125,
0.1192626953125,
0.126953125
],
[
0.146240234375,
0.1319580078125,
0.88134765625,
0.1319580078125,
0.88134765625,
0.1458740234375,
0.146240234375,
0.1458740234375
],
[
0.11767578125,
0.15087890625,
0.88134765625,
0.15087890625,
0.88134765625,
0.1641845703125,
0.11767578125,
0.1641845703125
],
[
0.1192626953125,
0.169189453125,
0.88134765625,
0.169189453125,
0.88134765625,
0.18310546875,
0.1192626953125,
0.18310546875
],
[
0.1168212890625,
0.1875,
0.65185546875,
0.1881103515625,
0.65185546875,
0.20263671875,
0.1168212890625,
0.2020263671875
],
[
0.1446533203125,
0.20703125,
0.88232421875,
0.20703125,
0.88232421875,
0.2215576171875,
0.1446533203125,
0.2215576171875
],
[
0.11517333984375,
0.2247314453125,
0.88134765625,
0.2259521484375,
0.88134765625,
0.240478515625,
0.11517333984375,
0.2392578125
],
[
0.11767578125,
0.2442626953125,
0.8798828125,
0.2442626953125,
0.8798828125,
0.25830078125,
0.11767578125,
0.25830078125
],
[
0.11767578125,
0.26318359375,
0.8798828125,
0.26318359375,
0.8798828125,
0.276611328125,
0.11767578125,
0.276611328125
],
[
0.11767578125,
0.2822265625,
0.88232421875,
0.2822265625,
0.88232421875,
0.296630859375,
0.11767578125,
0.296630859375
],
[
0.1168212890625,
0.3017578125,
0.88134765625,
0.3017578125,
0.88134765625,
0.31494140625,
0.1168212890625,
0.31494140625
],
[
0.1168212890625,
0.3193359375,
0.8798828125,
0.3193359375,
0.8798828125,
0.333251953125,
0.1168212890625,
0.333251953125
],
[
0.11517333984375,
0.339111328125,
0.88232421875,
0.339111328125,
0.88232421875,
0.352783203125,
0.11517333984375,
0.352783203125
],
[
0.1168212890625,
0.356689453125,
0.8798828125,
0.356689453125,
0.8798828125,
0.371826171875,
0.1168212890625,
0.371826171875
],
[
0.11767578125,
0.376953125,
0.88134765625,
0.376953125,
0.88134765625,
0.39013671875,
0.11767578125,
0.39013671875
],
[
0.11767578125,
0.395751953125,
0.2320556640625,
0.395751953125,
0.2320556640625,
0.4091796875,
0.11767578125,
0.4091796875
],
[
0.143798828125,
0.4267578125,
0.85693359375,
0.4267578125,
0.85693359375,
0.43994140625,
0.143798828125,
0.43994140625
],
[
0.2442626953125,
0.443115234375,
0.75244140625,
0.442626953125,
0.75244140625,
0.460205078125,
0.2442626953125,
0.4609375
],
[
0.146240234375,
0.47607421875,
0.8798828125,
0.47607421875,
0.8798828125,
0.4892578125,
0.146240234375,
0.4892578125
],
[
0.1192626953125,
0.494384765625,
0.88134765625,
0.494384765625,
0.88134765625,
0.50830078125,
0.1192626953125,
0.50830078125
],
[
0.1192626953125,
0.51318359375,
0.48193359375,
0.51318359375,
0.48193359375,
0.5263671875,
0.1192626953125,
0.5263671875
],
[
0.11767578125,
0.53515625,
0.88134765625,
0.53515625,
0.88134765625,
0.548828125,
0.11767578125,
0.548828125
],
[
0.11767578125,
0.55322265625,
0.8798828125,
0.55322265625,
0.8798828125,
0.56640625,
0.11767578125,
0.56640625
],
[
0.11767578125,
0.57275390625,
0.8798828125,
0.57275390625,
0.8798828125,
0.5859375,
0.11767578125,
0.5859375
],
[
0.1168212890625,
0.5888671875,
0.88134765625,
0.59033203125,
0.88134765625,
0.60498046875,
0.1168212890625,
0.60400390625
],
[
0.11767578125,
0.60986328125,
0.88134765625,
0.60986328125,
0.88134765625,
0.62353515625,
0.11767578125,
0.62353515625
],
[
0.11767578125,
0.62890625,
0.88134765625,
0.62890625,
0.88134765625,
0.642578125,
0.11767578125,
0.642578125
],
[
0.11767578125,
0.64794921875,
0.88134765625,
0.64794921875,
0.88134765625,
0.6611328125,
0.11767578125,
0.6611328125
],
[
0.11767578125,
0.66650390625,
0.88134765625,
0.66650390625,
0.88134765625,
0.6796875,
0.11767578125,
0.6796875
],
[
0.11767578125,
0.6845703125,
0.79345703125,
0.6845703125,
0.79345703125,
0.69775390625,
0.11767578125,
0.69775390625
],
[
0.1192626953125,
0.7119140625,
0.8798828125,
0.7119140625,
0.8798828125,
0.72265625,
0.1192626953125,
0.72265625
],
[
0.1168212890625,
0.72802734375,
0.88232421875,
0.72900390625,
0.88232421875,
0.74365234375,
0.1168212890625,
0.7421875
],
[
0.11517333984375,
0.7470703125,
0.8798828125,
0.74755859375,
0.8798828125,
0.76220703125,
0.11517333984375,
0.76123046875
],
[
0.11517333984375,
0.767578125,
0.88134765625,
0.767578125,
0.88134765625,
0.78076171875,
0.11517333984375,
0.78076171875
],
[
0.11767578125,
0.78515625,
0.88134765625,
0.78515625,
0.88134765625,
0.798828125,
0.11767578125,
0.798828125
],
[
0.11767578125,
0.8046875,
0.8798828125,
0.8046875,
0.8798828125,
0.818359375,
0.11767578125,
0.818359375
],
[
0.1168212890625,
0.82373046875,
0.6708984375,
0.82373046875,
0.6708984375,
0.8369140625,
0.1168212890625,
0.8369140625
],
[
0.1446533203125,
0.84228515625,
0.88134765625,
0.84228515625,
0.88134765625,
0.85595703125,
0.1446533203125,
0.85595703125
],
[
0.1168212890625,
0.861328125,
0.8798828125,
0.861328125,
0.8798828125,
0.875,
0.1168212890625,
0.875
],
[
0.491943359375,
0.8984375,
0.50732421875,
0.8984375,
0.50732421875,
0.91357421875,
0.491943359375,
0.91357421875
]
] |
[
"一个离线数据集，该数据集涵盖与一组奖励函数相关的所有可能的最佳轨迹。有了这样一个良性的",
"离线数据集，对于任何任意的奖励函数，RL代理都有足够的信息来估计相应的策略",
"要以原则性的方式研究这样的问题，我们将重点放在无奖励RL的框架上，该框架由探索阶段",
"和计划阶段组成。具体来说，在探索阶段，代理在不访问任何预先指定的奖励的情况下与环境进行",
"交互，并为随后的计划阶段收集经验轨迹。在计划阶段，使用在探索阶段收集的离线数据，代理在",
"给定外部奖励函数时计算最佳策略，而无需与环境进行进一步的交互",
"最近，许多工作都集中在设计可证明的样本有效的无奖励RL算法上。对于单代理表格情况，",
"Jin et al.(2020a);Kaufmann et al.(2020);Ménard et al.(2020);Zhang et al.(2020)实现了用于获",
"得ε次优策略的(poly(H,JS，AD)/ε）样本复杂性，其中S，A|分别是状态和动作空间的大小。",
"鉴于较大的动作和状态空间，作品Zanetteetal.（2020b)Wangetal.（2020a）通过对单代理马尔可",
"夫决策过程（MDP）应用线性函数逼近，从理论上分析了无奖励RL，从而实现了O(poly(H,o)/e2）",
"样本复杂度表示特征空间的维度。然而，RL算法结合非线性函数逼近器，如核函数逼近器和",
"神经函数逼近器，在各种应用问题上都取得了巨大的经验成功（e。g。，Duanetal.（2016)；Silver",
"etal.（2016，2017)；Wangetal.（2018)；Vinyalsetal.（2019）），这要归功于它们的表达能力。另一",
"方面，尽管在BaiandJin（2020)；Liuetal.（2020）中已经研究了表格情况下多人马尔可夫游戏的",
"无奖励RL算法，但仍然缺乏从理论上研究具有函数逼近的多智能体场景的工作。因此，以下问题",
"仍然悬而未决：",
"Can we design provably efficient reward-free RL algorithms with kernel and neural function",
"approximations for both single-agent MDPs and Markoy games?",
"回答上述问题的主要挑战在于如何将非线性逼近器适当地集成到无奖励RL的框架中，以及如",
"何通过设计适合这种逼近的探索奖励和奖金来激励探索。在本文中，我们通过应对这些挑战为上述",
"问题提供了肯定的答案。我们的贡献总结如下：",
"贡献。在本文中，我们首先针对单代理MDP设置提出了具有核和神经函数逼近的可证明样本和计",
"算高效的无奖励RL算法。我们的探索算法是最小二乘值迭代算法的乐观变体，结合了核函数和神",
"经函数逼近器，该算法采用关联的（缩放的）奖金作为探索奖励。在计划阶段，我们的方法实现了",
"O(1/2）样本复杂性，从而为任意外部奖励函数生成了次优策略。此外，我们将所提出的单代理",
"设置方法扩展到零和马尔可夫博奔设置，使得该算法可以实现(1/ε2）样本复杂度来生成策略对",
"该策略对是ε-近似纳什均衡。特别是，在马尔可夫博奔的计划阶段，我们的算法仅涉及找到可以",
"有效求解的Q函数形成的矩阵博奔的纳什均衡，这是独立的兴趣。我们方法的样本复杂性与表格",
"或线性函数逼近设置的现有工作中的①(1e2）结果相匹配。据我们所知，我们建立了第一个可证明",
"有效的无奖励RL算法，该算法具有用于单代理和多代理设置的内核和神经功能逼近器",
"相关工作。已经有很多工作集中在为单代理和多代理RL问题设计可证明有效的无奖励RL算法",
"对于单代理场景，Jinetal.（2020a）将表格设置的无奖励RL形式化，并以O(poly(H,SI，A)/e²)",
"样本复杂度为所提出的算法提供理论分析，以实现次优策略。在最近的几部作品（Kaufmann",
"etal.，2020；Menardetal.，2020；Zhangetal.，2020）中，表格设置的样本复杂性得到了进一步改",
"善。最近，Zanetteetal.（2020b)；Wangetal.（2020a）从线性函数逼近的角度研究了无回报RL。对",
"于多主体设置，BaiandJin（2020）研究了表格情况下零和马尔可夫博奔的无奖励探索。Liuetal",
"（2020）进一步提出了适用于多人一般和游戏的可证明的无奖励RL算法",
"我们的工作也与研究具有函数逼近的RL算法的一系列工作密切相关。有大量的著作（Yang",
"and Wang, 2019, 2020; Cai et al., 2019; Zanette et al., 2020a; Jin et al., 2020b; Wang et al., 2019:",
"2"
] |
{
"arxiv_id": "2110.09771",
"language": "zh",
"page_idx": 1
}
|
||
2110.09771-zh-2
|
2110.09771-zh-2.jpg
|
Ayoub et al., 2020; Zhou et al., 2020a; Kakade et al., 2020) 用 (广义) 线性函数逼近研究不同的RL 问题。此外，Wang et al. (2020b) 提出了一种具有一般函数逼近的乐观最小二乘迭代 (LSVI) 算法。我们的工作更多地与内核化上下文土匪 (Srinivas et al., 2009; Valko et al., 2013; Chowdhury and Gopalan, 2017) , 神经上下文土匪 (Zhou et al., 2020b) 以及具有内核和神经函数近似 (Yang et al., 2020) 的基于奖励的在线 RL 的问题有关。。具体来说，(Srinivas et al., 2009; Valko et al., 2013; Chowdhury and Gopalan, 2017) 提出了用于内核化上下文土匪的算法，并构造了相应的置信上限 (UCB) 奖金以鼓励探索。对于神经上下文土匪，Zhou et al. (2020b) 提出了一种基于神经网络的算法，名为 NeuralUCB, 该算法使用神经随机特征映射来构造 UCB 进行探索。从这些作品中改编 UCB 奖金结构，最近的工作 Yang et al. (2020)研究了具有核和神经功能近似的在线 RL 间题的乐观 LSVI 算法，该算法涵盖了上下文强盗问题作为特例。但是，这些先前的工作仅研究了使用奖励反馈执行探索的设置，不能直接应用于无奖励 RL 问题。受上述工作的启发，我们的工作将核和神经函数逼近的思想扩展到了无奖励的单代理 MDP 和零和马尔可夫游戏
# 2 Preliminaries
在本节中，我们介绍本文的基本符号和问题背景
# 2.1 Markov Decision Process
考虑具有由元组$(\mathcal{S},\mathcal{A},H,\mathbb{P},r)$定义的 XMATHX 状态 XDP 的情节单代理 X1, 其中$\mathcal{S}$表示状态空间，$\mathcal{A}$是代理的动作空间，$H$是每个情节的长度，$\mathbb{P}=\{\mathbb{P}_h\}_h=1^H$是具有$\mathbb{P}_h(s^\prime|s,a)$的转换模型$h$ -th 步骤中的转换概率。具体来说，我们假设真正的过渡模型$\mathbb{P}$对于需要无奖励探索的代理是未知的。代理的策略是概率分布$\pi=\{\pi_h\}_h=1^H$的集合，其中$\pi_h:\mathcal{S}\mapsto\Delta_\mathcal{A}^{\color{red}{1}}$和$\Delta_\mathcal{A}$表示在空间$\mathcal{A}$上定义的概率单纯形
对于特定的策略$\{\pi_h\}_{h=1}^H$和奖励函数$\{r_h\}_{h=1}^H$,在过渡模型$\{\mathbb{P}_h\}_h=1^H$下，我们将第$h$步的关联值函数$V_h^\pi(s,r):\mathcal{S}\mapsto\mathbb{R}$定义为$V_h^\pi(s,r):=\mathbb{E}[\sum_{h^{\prime}=h}^Hr_{h^{\prime}}(s_{h^{\prime}},a_{h^{\prime}})\mid s_h=s,\pi,\mathbb{P}]$ 。相应的动作值函数$(\mathbb{Q}$函数) $Q_h^\pi:\mathcal{S}\times\mathcal{A}\mapsto\mathbb{R}$进一步定义为$Q_h^\pi(s,a,r):=\mathbb{E}[\sum_{h^{\prime}=h}^Hr_{h^{\prime}}(s_{h^{\prime}},a_{h^{\prime}})\mid s_h=s,a_h=a,\pi,\mathbb{P}]$ 。因此，我们有$V_h^\pi(s,r)=\langle Q_h^\pi(s,\cdot,r),\pi_h(s)\rangle_{\mathcal{A}}^2$和$Q_h^\pi(s,a,r)=r_h(s,a)+\langle\mathbb{P}_h(\cdot|s,a),V_{h+1}^\pi(\cdot,r)\rangle_{\mathcal{S}}$的Bellman 方程，其中我们让$\langle\cdot,\cdot\rangle_S,\langle\cdot,\cdot\rangle_{\mathcal{A}}$表示空间$\mathcal{S},\mathcal{A}$上的内积。上述 Bellman 方程适用于所有设置$V_{H+1}^\pi(s)=0,\forall s\in\mathcal{S}$的$h\in[H]$ 。在本文的其余部分中，为了简化符号，我们为任何转移概率$\mathbb{P}_h$和值函数$V(\cdot,r)$重写$\langle\mathbb{P}_h(\cdot|s,a),V_{h+1}(\cdot,r)\rangle_S=\mathbb{P}_hV_{h+1}(s,a,r)$ 。此外，我们将$\pi_r^*$表示为最优策略 w。r。t。$r$ ,使得$\pi_r^*$最大化$V_1^\pi(s_1,r)^{\color{red}3}$。然后，我们定义$Q_h^*(s,a,r):=Q_h^{\pi_r^*}(s,a,r)$ 和$V_{h}^{*}(s,r):=V_{h}^{\pi_{r}^{*}}(s,r)$ 。我们说$\tilde{\pi}$是一个$\varepsilon$次优策略，如果它满足
$$V_1^*(s_1,r)-V_1^{\widetilde{\pi}}(s_1,r)\leq\varepsilon.$$
$^1$If a policy $\pi_h$ is deterministic, for simplicity, we slightly abuse the notion by letting $\pi_h:\mathcal{S}\mapsto\mathcal{A}.$ A similar
notation is also defined for the Markov game setting.
2When a policy $\pi _{h}$ is deterministic, we have the value function defined as $V_{h}^{\pi }( s, r) = Q_{h}^{\pi }( s, \pi _{h}( s) , r)$ .
$^{3}$WLOG, we assume the agent starts from a fixed state $s_1$ at $h=1.$ We also make the same assumption for the
Markov game setting.
3
|
[
[
0.1168212890625,
0.0921630859375,
0.88134765625,
0.09344482421875,
0.88134765625,
0.10797119140625,
0.1168212890625,
0.106689453125
],
[
0.11767578125,
0.11297607421875,
0.88134765625,
0.11297607421875,
0.88134765625,
0.126953125,
0.11767578125,
0.126953125
],
[
0.1168212890625,
0.1307373046875,
0.88134765625,
0.1319580078125,
0.88134765625,
0.146484375,
0.1168212890625,
0.1458740234375
],
[
0.11517333984375,
0.1483154296875,
0.88134765625,
0.15087890625,
0.88134765625,
0.1654052734375,
0.11517333984375,
0.1634521484375
],
[
0.1168212890625,
0.169189453125,
0.8798828125,
0.169189453125,
0.8798828125,
0.18310546875,
0.1168212890625,
0.18310546875
],
[
0.11767578125,
0.1881103515625,
0.88134765625,
0.1881103515625,
0.88134765625,
0.2020263671875,
0.11767578125,
0.2020263671875
],
[
0.12091064453125,
0.20703125,
0.88134765625,
0.20703125,
0.88134765625,
0.2215576171875,
0.12091064453125,
0.2215576171875
],
[
0.1168212890625,
0.2259521484375,
0.88134765625,
0.2259521484375,
0.88134765625,
0.240478515625,
0.1168212890625,
0.240478515625
],
[
0.11767578125,
0.2442626953125,
0.88134765625,
0.2442626953125,
0.88134765625,
0.25830078125,
0.11767578125,
0.25830078125
],
[
0.1168212890625,
0.261962890625,
0.88134765625,
0.26318359375,
0.88134765625,
0.27783203125,
0.1168212890625,
0.276611328125
],
[
0.11767578125,
0.2822265625,
0.88134765625,
0.2822265625,
0.88134765625,
0.296630859375,
0.11767578125,
0.296630859375
],
[
0.11767578125,
0.3017578125,
0.7236328125,
0.3017578125,
0.7236328125,
0.31494140625,
0.11767578125,
0.31494140625
],
[
0.11517333984375,
0.3447265625,
0.318603515625,
0.3447265625,
0.318603515625,
0.359130859375,
0.11517333984375,
0.359130859375
],
[
0.11767578125,
0.380126953125,
0.490234375,
0.380126953125,
0.490234375,
0.393310546875,
0.11767578125,
0.393310546875
],
[
0.11517333984375,
0.417236328125,
0.414306640625,
0.417236328125,
0.414306640625,
0.431884765625,
0.11517333984375,
0.431884765625
],
[
0.1168212890625,
0.4443359375,
0.88134765625,
0.44580078125,
0.88134765625,
0.460205078125,
0.1168212890625,
0.458984375
],
[
0.11767578125,
0.464111328125,
0.88232421875,
0.464111328125,
0.88232421875,
0.479248046875,
0.11767578125,
0.479248046875
],
[
0.11358642578125,
0.481689453125,
0.88134765625,
0.48291015625,
0.88134765625,
0.49755859375,
0.11358642578125,
0.496826171875
],
[
0.1168212890625,
0.50048828125,
0.88134765625,
0.501953125,
0.88134765625,
0.5166015625,
0.1168212890625,
0.51513671875
],
[
0.1192626953125,
0.521484375,
0.2418212890625,
0.521484375,
0.2418212890625,
0.53515625,
0.1192626953125,
0.53515625
],
[
0.1422119140625,
0.537109375,
0.88232421875,
0.53857421875,
0.88232421875,
0.55615234375,
0.1422119140625,
0.5546875
],
[
0.11767578125,
0.5595703125,
0.88134765625,
0.5595703125,
0.88134765625,
0.57373046875,
0.11767578125,
0.57373046875
],
[
0.11358642578125,
0.57470703125,
0.8837890625,
0.57568359375,
0.8837890625,
0.59326171875,
0.11358642578125,
0.59228515625
],
[
0.1192626953125,
0.5966796875,
0.88134765625,
0.5966796875,
0.88134765625,
0.611328125,
0.1192626953125,
0.611328125
],
[
0.11358642578125,
0.6123046875,
0.88232421875,
0.61279296875,
0.88232421875,
0.630859375,
0.11358642578125,
0.6298828125
],
[
0.11767578125,
0.6337890625,
0.88134765625,
0.6337890625,
0.88134765625,
0.6484375,
0.11767578125,
0.6484375
],
[
0.1168212890625,
0.65283203125,
0.8798828125,
0.65283203125,
0.8798828125,
0.66748046875,
0.1168212890625,
0.66748046875
],
[
0.11767578125,
0.671875,
0.88134765625,
0.671875,
0.88134765625,
0.68603515625,
0.11767578125,
0.68603515625
],
[
0.11767578125,
0.68994140625,
0.650390625,
0.68994140625,
0.650390625,
0.705078125,
0.11767578125,
0.705078125
],
[
0.40283203125,
0.72607421875,
0.59716796875,
0.72607421875,
0.59716796875,
0.74072265625,
0.40283203125,
0.74072265625
],
[
0.13232421875,
0.74560546875,
0.88134765625,
0.74755859375,
0.88134765625,
0.76513671875,
0.13232421875,
0.76318359375
],
[
0.11517333984375,
0.76220703125,
0.464111328125,
0.76220703125,
0.464111328125,
0.77587890625,
0.11517333984375,
0.77587890625
],
[
0.1356201171875,
0.7744140625,
0.80126953125,
0.7763671875,
0.80126953125,
0.79150390625,
0.1356201171875,
0.7890625
],
[
0.1356201171875,
0.78857421875,
0.88134765625,
0.7890625,
0.88134765625,
0.8037109375,
0.1356201171875,
0.80322265625
],
[
0.11517333984375,
0.80322265625,
0.2607421875,
0.8046875,
0.2607421875,
0.8193359375,
0.11517333984375,
0.8173828125
],
[
0.491943359375,
0.89794921875,
0.50732421875,
0.89794921875,
0.50732421875,
0.912109375,
0.491943359375,
0.912109375
]
] |
[
"Ayoubetal.，2020；Zhouetal.，2020a；Kakadeetal.，2020）用（广义）线性函数逼近研究不同的",
"RL问题。此外，Wangetal.（2020b）提出了一种具有一般函数逼近的乐观最小二乘迭代（LSVI)",
"算法。我们的工作更多地与内核化上下文土匪（Srinivasetal.，2009；Valkoetal.，2013；Chowdhury",
"andGopalan，2017），神经上下文土匪（Zhouetal.，2020b）以及具有内核和神经函数近似（Yang",
"etal.，2020）的基于奖励的在线RL的问题有关。具体来说，（Srinivasetal.，2009；Valkoetal.",
"2013；ChowdhuryandGopalan，2017）提出了用于内核化上下文土匪的算法，并构造了相应的置信",
"上限（UCB）奖金以鼓励探索。对于神经上下文土匪，Zhouetal.（2020b）提出了一种基于神经网",
"络的算法，名为NeuralUCB，该算法使用神经随机特征映射来构造UCB进行探索。从这些作品中",
"改编UCB奖金结构，最近的工作Yangetal.（2020）研究了具有核和神经功能近似的在线RL问",
"题的乐观LSVI算法，该算法涵盖了上下文强盗问题作为特例。但是，这些先前的工作仅研究了使",
"用奖励反馈执行探索的设置，不能直接应用于无奖励RL问题。受上述工作的启发，我们的工作将",
"核和神经函数逼近的思想扩展到了无奖励的单代理MDP和零和马尔可夫游戏",
"2Preliminaries",
"在本节中，我们介绍本文的基本符号和问题背景",
"2.1 Markov Decision Process",
"考虑具有由元组（S,A,H,P,r）定义的XMATHX状态XDP的情节单代理X1，其中S表示状态",
"空间，A是代理的动作空间，H是每个情节的长度，P={Ph)h-1是具有Ph(s|s,α）的转换模型h",
"-th步骤中的转换概率。具体来说，我们假设真正的过渡模型P对于需要无奖励探索的代理是未知",
"的。代理的策略是概率分布π=元h)H=的集合，其中元h：S→△A1和△A表示在空间A上定",
"义的概率单纯形",
"对于特定的策略【元h)H-1和奖励函数{rh)H-1，在过渡模型{Ph)h=1下，我们将第h步的关联",
"值函数Vh（s,）：S→R定义为Vh（s,r）=EZh=hrh(sh'ah）lsh=s,元,P）。相应的动作值函数",
"(Q 函数）Q:S× A→R 进一步定义为 Q(s,a,r):=E[h=hrh(sh',ah)ISh=s,ah =a,π,P] 。",
"因此，我们有 Vh(s,r)=(Q(s,,r),元h(s))A 2和 Q(s,a,r)=rh(s,a)+(Ph(s,a),Vh+i(,r))s的",
"Bellman方程，其中我们让《，>s，<，>A表示空间S，A上的内积。上述Bellman方程适用于",
"所有设置V+1(s)=0,VsES的hE[H]。在本文的其余部分中，为了简化符号，我们为任何转",
"移概率Ph和值函数V(,r）重写(Ph(s,a),Vh+i(,r))s=PhVh+1(s,a,r）。此外，我们将元表示",
"为最优策略w。r。t。r，使得元*最大化V(si,r）3。然后，我们定义Q(s,a,r)：=Qhr(s,a,r）",
"和V(s,r）：=V(s,r）。我们说元是一个ε次优策略，如果它满足",
"V*(si,r) -Vi(si,r) ≤e.",
"'f a policy h is deterministic, for simplicity, we slightly abuse the notion by letting Th : S A. A similar",
"notation is also defined for the Markov game setting",
"2When a policy h is deterministic, we have the value function defined as V (s, r) = Qh(s, h(s),r)",
"3WLOG, we assume the agent starts from a fixed state si at h = 1 . We also make the same assumption for the",
"Markov game setting.",
"3"
] |
{
"arxiv_id": "2110.09771",
"language": "zh",
"page_idx": 2
}
|
||
2110.09771-zh-3
|
2110.09771-zh-3.jpg
|
2.2 Zero-Sum Markov Game
在本文中，我们还考虑了具有 XMATHX 状态$S$的玩家 XMATHX 动作$S$采取以元组$(\mathcal{S},\mathcal{A},\mathcal{B},H,\mathbb{P},r)$
为特征的情节零和马尔可夫游戏，其中$\mathcal{S}$表示状态空间，$\mathcal{A}$和$\mathcal{B}$是两个玩家的动作空间，$H$是每个情节。同样，我们假设转换模型$\mathbb{P}=\{\mathbb{P}_h\}_{h=1}^H$对两个玩家都是未知的。播放器 1 的策略是概率分布$\pi=\{\pi_h\}_{h=1}^H$与$\pi:S\mapsto\Delta_A$的集合。类似地，播放器 2 的策略是概率分布$\nu=\{\nu_h\}_h=1^H$和$\nu:\mathcal{S}\mapsto\Delta_{\mathcal{B}}$的集合。这里$\Delta_\mathcal{A}$和$\Delta_\mathcal{B}$是在空间$\mathcal{A}$和$\mathcal{B}$上定义的概率单形
对于特定的策略对$\pi$和$\nu$以及奖励函数$\{r_h\}_h\in[H]$,在过渡模型$\{\mathbb{P}_h\}_h\in[H]$下，我们将$h$步骤中的值函数$V_h^{\pi,\nu}(s,r):\mathcal{S}\mapsto\mathbb{R}$定义为$V_h^{\bar{\pi},\nu}(s,r):=\mathbb{E}[\sum_{h^{\prime}=h}^Hr_{h^{\prime}}(s_{h^{\prime}},a_{h^{\prime}},b_{h^{\prime}})\mid s_h=s,\pi,\nu,\mathbb{P}]$ 。我们进一步将相应的动作值函数(Q 函数)$Q_h^{\pi,\nu}:\mathcal{S}\times\mathcal{A}\times\mathcal{B}\mapsto\mathbb{R}$定义为$Q_h^{\pi,\nu}(s,a,b,r):=$ $\mathbb{E}[\sum_{h^{\prime}=h}^Hr_{h^{\prime}}(s_{h^{\prime}},a_{h^{\prime}},b_{h^{\prime}})\mid s_h=s,a_h=a,b_h=b,\pi,\mathbb{P}]$ 。因此，我们将 Bellman 方程称为$V_h^{\pi,\nu}(s,r)=$ $\mathbb{E}_{a\sim\pi_h,b\sim\nu_h}[Q_h^{\pi,\nu}(s,a,b,r)]$ 和$Q_h^{\pi , \nu }( s, a, b, r) = r_h( s, a, b) + \mathbb{P} _hV_{h+ 1}^{\pi , \nu }( s, a, b, r)$ ,其中，为了便于表示， 我们还将$\mathbb{P}_hV_{h+1}^{\pi,\nu}(s,a,b,r)=\left\langle\mathbb{P}_h(\cdot|s,a,b),V_{h+1}^{\pi,\nu}(\cdot,r)\right\rangle_{\mathcal{S}}$
$\operatorname*{max}_{\pi}\operatorname*{min}_{\nu}V_{1}^{\pi,\nu}(s_{1},r)=\operatorname*{min}_{\nu}\operatorname*{max}_{\pi}V_{1}^{\pi,\nu}(s_{1},r)$ 。为简单起见，我们让$V_h^\dagger(s,r)=V_{h}^{\pi^\dagger,\nu^\dagger}(s,r)$和$Q_{h}^{\dagger}(s,a,b,r)=Q_{h}^{\pi^{\dagger,\nu^{\dagger}}}(s,a,b,r)$在$h$步中表示 NE $(\pi^\dagger,\nu^\dagger)$下的值函数和 Q 函数。我们进一步将具有策略$\pi$的播放器 1 的最佳响应定义为 br$(\pi):=\operatorname{argmin}_\nu V_1^{\pi,\nu}(s_1,r)$ ,将具有策略$\nu$的播放器 2的最佳响应定义为 br$(\nu):=\operatorname{argmax}_\pi V_1^{\pi,\nu}(s_1,r)$ 。因此，如果$(\widetilde{\pi},\widetilde{\nu})$满足
$$V_{1}^{\mathrm{br}(\widetilde{\nu}),\widetilde{\nu}}(s_{1},r)-V_{1}^{\widetilde{\pi},\mathrm{br}(\widetilde{\pi})}(s_{1},r)\leq\varepsilon,$$
,其中$V_1^{\mathrm{br}(\widetilde{\nu}),\widetilde{\nu}}(s_1,r)\geq V_1^{\dagger}(s_1,r)\geq V_1^{\widetilde{\pi},\mathrm{br}(\widetilde{\pi})}(s_1,r)$始终成立，则说$(\widetilde{\pi},\widetilde{\nu})$是$\varepsilon$近似的 NE。另一方面，我们让$V_1^* ( s, r) = \max _{\pi , \nu }V_1^{\pi , \nu }( s, r)$ ,即当$h=1$ 时的最大值函数。然后，我们得到了$h$步骤$V_{h}^{*}(s,r)$和$Q_h^{*}(s,a,b,r)$的相关值函数和 Q 函数
2.3 Reproducing Kernel Hilbert Space
我们研究了基于再生核希尔伯特空间 (RKHS) 的核函数逼近。由于概念的轻微滥用，我们将$\mathcal{Z}=\mathcal{S}\times\mathcal{A}$用于单代理 MDP 设置，将$\mathcal{Z}=\mathcal{S}\times\mathcal{A}\times\mathcal{B}$用于零和游戏设置，以便$z=(s,a)\in\mathcal{Z}$或$z=(s,a,b)\in\mathcal{Z}$用于不同情况。我们假设空间$z$是逼近函数的输入空间，其中$z$是$\mathbb{R}^d$上的紧致空间。如果有将$(s,a)$或$(s,a,b)$嵌人到空间$\mathbb{R}^d$中的预处理方法，也可以实现这一点。我们让$\mathcal{H}$ 是在具有内核函数 ker$:z\times z\mapsto\mathbb{R}$的空间$z$上定义的 RKHS。我们进一步将 RKHS $\mathcal{H}$上的内积定义为$\langle\cdot,\cdot\rangle_{\mathcal{H}}:\mathcal{H}\times\mathcal{H}\mapsto\mathbb{R}$和范数$\|\cdot\|_{\mathcal{H}}:\mathcal{H}\mapsto\mathbb{R}$ 。我们在 RKHS $\mathcal{H}$上有一个特征图$\phi:\mathcal{Z}\mapsto\mathcal{H}$ ,并为$f\in\mathcal{H}$定义了函数$f(z):=\langle f,\phi(z)\rangle_\mathcal{H}$ 。然后将内核定义为
$$\ker(z,z'):=\langle\phi(z),\phi(z')\rangle_{\mathcal H},\quad\forall z,z'\in\mathcal Z.$$
我们假设$\sup_z\in\mathcal{Z}\ker(z,z)\leq1$使得任何$z\in\mathcal{Z}$的$\|\phi(z)\|_\mathcal{H}\leq1$
2.4 Overparameterized Neural Network
本文进一步考虑了利用超参数化神经网络的函数逼近器。超参数神经网络最近在理论和实践中都引起了很多关注 (Neyshabur et al., 2018; Allen-Zhu et al., 2018; Arora et al., 2019; Gao et al., 2019
4
|
[
[
0.1168212890625,
0.09344482421875,
0.41259765625,
0.09344482421875,
0.41259765625,
0.10797119140625,
0.1168212890625,
0.10797119140625
],
[
0.11767578125,
0.121826171875,
0.90283203125,
0.121826171875,
0.90283203125,
0.1363525390625,
0.11767578125,
0.1363525390625
],
[
0.11767578125,
0.141357421875,
0.88134765625,
0.141357421875,
0.88134765625,
0.1546630859375,
0.11767578125,
0.1546630859375
],
[
0.11767578125,
0.1590576171875,
0.88134765625,
0.1590576171875,
0.88134765625,
0.173583984375,
0.11767578125,
0.173583984375
],
[
0.1168212890625,
0.1767578125,
0.88232421875,
0.1767578125,
0.88232421875,
0.1944580078125,
0.1168212890625,
0.1944580078125
],
[
0.11517333984375,
0.1956787109375,
0.6953125,
0.1962890625,
0.6953125,
0.2108154296875,
0.11517333984375,
0.210205078125
],
[
0.143798828125,
0.21533203125,
0.88134765625,
0.21533203125,
0.88134765625,
0.23291015625,
0.143798828125,
0.23291015625
],
[
0.11767578125,
0.23486328125,
0.88134765625,
0.23486328125,
0.88134765625,
0.25,
0.11767578125,
0.25
],
[
0.11767578125,
0.25244140625,
0.88134765625,
0.25244140625,
0.88134765625,
0.267578125,
0.11767578125,
0.267578125
],
[
0.11358642578125,
0.268310546875,
0.8837890625,
0.26953125,
0.8837890625,
0.290283203125,
0.11358642578125,
0.2890625
],
[
0.1168212890625,
0.290283203125,
0.88232421875,
0.290283203125,
0.88232421875,
0.30810546875,
0.1168212890625,
0.30810546875
],
[
0.11517333984375,
0.306884765625,
0.53759765625,
0.31005859375,
0.53759765625,
0.32763671875,
0.11517333984375,
0.324462890625
],
[
0.1422119140625,
0.32568359375,
0.8837890625,
0.32568359375,
0.8837890625,
0.349853515625,
0.1422119140625,
0.349853515625
],
[
0.1168212890625,
0.349853515625,
0.88134765625,
0.349853515625,
0.88134765625,
0.367431640625,
0.1168212890625,
0.367431640625
],
[
0.11517333984375,
0.369384765625,
0.8837890625,
0.369384765625,
0.8837890625,
0.39013671875,
0.11517333984375,
0.39013671875
],
[
0.11767578125,
0.393310546875,
0.88232421875,
0.393310546875,
0.88232421875,
0.40771484375,
0.11767578125,
0.40771484375
],
[
0.11767578125,
0.410400390625,
0.689453125,
0.410400390625,
0.689453125,
0.427978515625,
0.11767578125,
0.427978515625
],
[
0.361083984375,
0.439453125,
0.640625,
0.44140625,
0.640625,
0.4697265625,
0.361083984375,
0.46728515625
],
[
0.11358642578125,
0.479736328125,
0.8837890625,
0.479736328125,
0.8837890625,
0.50390625,
0.11358642578125,
0.50390625
],
[
0.11767578125,
0.505859375,
0.88134765625,
0.505859375,
0.88134765625,
0.52099609375,
0.11767578125,
0.52099609375
],
[
0.11767578125,
0.5244140625,
0.494384765625,
0.5244140625,
0.494384765625,
0.5390625,
0.11767578125,
0.5390625
],
[
0.1168212890625,
0.5625,
0.5087890625,
0.5625,
0.5087890625,
0.5771484375,
0.1168212890625,
0.5771484375
],
[
0.11767578125,
0.59033203125,
0.88134765625,
0.59033203125,
0.88134765625,
0.60498046875,
0.11767578125,
0.60498046875
],
[
0.1192626953125,
0.60791015625,
0.8798828125,
0.60791015625,
0.8798828125,
0.62255859375,
0.1192626953125,
0.62255859375
],
[
0.1168212890625,
0.62744140625,
0.8798828125,
0.626953125,
0.8798828125,
0.6416015625,
0.1168212890625,
0.642578125
],
[
0.11767578125,
0.646484375,
0.88134765625,
0.646484375,
0.88134765625,
0.66015625,
0.11767578125,
0.66015625
],
[
0.11767578125,
0.66552734375,
0.88134765625,
0.66552734375,
0.88134765625,
0.6796875,
0.11767578125,
0.6796875
],
[
0.1168212890625,
0.6845703125,
0.88134765625,
0.68310546875,
0.88134765625,
0.69775390625,
0.1168212890625,
0.69873046875
],
[
0.1168212890625,
0.7041015625,
0.6298828125,
0.7041015625,
0.6298828125,
0.71826171875,
0.1168212890625,
0.71826171875
],
[
0.340576171875,
0.734375,
0.65283203125,
0.73486328125,
0.65283203125,
0.75244140625,
0.340576171875,
0.751953125
],
[
0.11767578125,
0.77099609375,
0.6044921875,
0.77099609375,
0.6044921875,
0.78857421875,
0.11767578125,
0.78857421875
],
[
0.11358642578125,
0.80810546875,
0.51953125,
0.806640625,
0.51953125,
0.82470703125,
0.11358642578125,
0.82568359375
],
[
0.11767578125,
0.83837890625,
0.8798828125,
0.83837890625,
0.8798828125,
0.8515625,
0.11767578125,
0.8515625
],
[
0.1168212890625,
0.85693359375,
0.8798828125,
0.85693359375,
0.8798828125,
0.87060546875,
0.1168212890625,
0.87060546875
],
[
0.491943359375,
0.89794921875,
0.5087890625,
0.89794921875,
0.5087890625,
0.91357421875,
0.491943359375,
0.91357421875
]
] |
[
"2.2 Zero-Sum Markov Game",
"在本文中，我们还考虑了具有XMATHX状态S的玩家XMATHX动作S采取以元组（S,A,B,H,P,r",
"为特征的情节零和马尔可夫游戏，其中S表示状态空间，A和B是两个玩家的动作空间，H是每",
"个情节。同样，我们假设转换模型P={Ph}H-对两个玩家都是未知的。播放器1的策略是概率",
"分布元=【元h)h-1与元：S→△A的集合。类似地，播放器2的策略是概率分布=(Vh)h-1和",
"V：S→△B的集合。这里△A和△B是在空间A和B上定义的概率单形",
"对于特定的策略对元和以及奖励函数【rh}he[H]，在过渡模型[Ph}he[H]下，我们将h步",
"骤中的值函数Vh(s,r)：S→R定义为Vh(s,r）=E[Zh=hrh(sh',ah',bh)|sh=S,,V,P]",
"。我们进一步将相应的动作值函数（Q函数）Qn：S×A×B→R定义为QT(s,a,b,r）:=",
"E[ZH=hrh(sh',ah,bh)|sh=8,ah=a,bh=b,元,P]。因此，我们将Bellman方程称为VT\"(s,r)=",
"Ea~πh,b~vn[Qh\"(s,a,b,r)]和 Qh\"(s,a,b,r)=rh(s,a,b)+PhVh+i(s,a,b,r)，其中，为了便于表示,",
"我们还将PhV+i(s,a,b,r)=<Ph(-Is,a,b),Vh+i(,r)s",
"我们将纳什均衡（NE）（元t,ut）定义为maxmin,V\"(si）的解，其中我们有Vtt（s1,r）=",
"maxminV(s1,r）=min,maxV(s1,r）。为简单起见，我们让Vt(s,r)=Vt(s,r）和",
"Qt(s,a,b,r)=Qf,ut(s,a,b,r)在 h步中表示NE（πt,ut）下的值函数和Q函数。我们进一步将具",
"有策略元的播放器1的最佳响应定义为br(元）：=argmin，Vi（s1,r），将具有策略v的播放器2",
"的最佳响应定义为br(u）：=argmaxV(si,r）。因此，如果（元,）满足",
"Vbri(s1, r) - vi,br()(s1,r) ≤e,",
"，其中vbr(s1,r)≥Vt(s1,r)≥V元,br()(s1,r)始终成立，则说（元,D)是近似的NE。另一方",
"面，我们让V*(s,r）=max元,vVu（s,r），即当h=1时的最大值函数。然后，我们得到了h步骤",
"V(s,r)和Q(s,a,b,r)的相关值函数和Q函数",
"2.3Reproducing Kernel Hilbert Space",
"我们研究了基于再生核希尔伯特空间（RKHS）的核函数逼近。由于概念的轻微滥用，我们将",
"Z=S×A用于单代理MDP设置，将Z=S×A×B用于零和游戏设置，以便z=（sa)EZ或",
"z=（s,a,b)Ez用于不同情况。我们假设空间z是逼近函数的输入空间，其中z是Rd上的紧致",
"空间。如果有将（s,a）或（s,a,b）嵌人到空间Rd中的预处理方法，也可以实现这一点。我们让H",
"是在具有内核函数ker：Z×ZR的空间Z上定义的RKHS。我们进一步将RKHSH上的内积",
"定义为《)H：H×H→R和范数I·H：H→R。我们在RKHSH上有一个特征图Φ：ZHH",
"，并为fEH定义了函数f(z):=<f,Φ(2))H。然后将内核定义为",
"ker(z,z2) := (Φ(z), Φ(z))H,Vz, 2' E z",
"我们假设 supzezker(z,z)≤1使得任何的 ll()1",
"2.4 Overparameterized Neural Network",
"本文进一步考虑了利用超参数化神经网络的函数逼近器。超参数神经网络最近在理论和实践中都引",
"起了很多关注 (Neyshabur et al.,2018; Allen-Zhu et al., 2018; Arora et al., 2019; Gao et al.,2019",
"4"
] |
{
"arxiv_id": "2110.09771",
"language": "zh",
"page_idx": 3
}
|
||
2110.09771-zh-4
|
2110.09771-zh-4.jpg
|
Bai and Lee, 2019) 。具 体 来 说 , 在 我 们 的 工 作 中 , 我 们 有 一 个 带 有 2$m$神经元和权重$(\boldsymbol{v},W)$的两
层神经网络$f( \cdot ; b, W) : \mathcal{Z} \mapsto \mathbb{R}$ ,
(1)
$$f(z;\boldsymbol{v},W)=\frac{1}{\sqrt{2m}}\sum_{i=1}^{2m}v_i\cdot\operatorname{act}(W_i^\top z),$$
,其中 act 是激活函数，$\boldsymbol{v}=[v_1,\cdots,v_{2m}]^\top$和$W=[W_1,W_2,\cdots,W_{2m}]$ 。在这里，我们假设$z=(s,a)$或$z=(s,a,b)$与$z\in\mathcal{Z}$满足$\|z\|_2=1$ , i。e。, $z$在$\mathbb{R}^d$中的单位超球面上归一化。令$W^{(0)}$为$W$的初始值，$v^{(0)}$为$v$的初始化。上述模型的初始化步骤执行如下：我们让$v_i\sim\operatorname*{Unif}(\{-1,1\})$和$W_i^{(0)}\sim N(0,I_d/d)$为所有$i\in[m]$ ,其中$I_d$是$\mathbb{R}^d\times d$中的单位矩阵，而$\begin{aligned}v_{i}^{(0)}=-v_{i-m}^{(0)},W_{i}^{(0)}=W_{i-m}^{(0)}\text{为所有 }i\in\{m+1,2m\}\text{。这里我们让 }N(0,I_{d}/d)\text{ 表示高斯分布。在}\\\text{本文中，我们将 }v\text{ 固定为 }v^{(0)},\text{ 并且为了便于理论分析，我们仅学习 }W\text{。因此，我们用 }f(z;W)\end{aligned}$ 表示$f( z; , \boldsymbol{v}, W)$ , 以简化表示法。泛的研究，例如。g。, Gao et al. (2019); Bai and Lee (2019) 。当模型被过度参数化时，我。e。, $m$足够大，我们可以用神经切线核 (NTK) (Jacot et al., 2018) 来表征训练这种神经网络的动力学。这里我们定义
$$\varphi(z;W):=[\nabla_{W_1}f(z;W)^\top,\cdots,\nabla_{W_{2m}}f(z;W)^\top]^\top,$$
(2)
,其中我们让$\nabla_{W_i}f(z;W)$是一个列向量，使得$\varphi(z;W)\in\mathbb{R}^{2md}$ 。因此，以$W^{(0)}$初始化$W$的随
机性为条件，我们进一步定义了内核
$$\ker_{m}(z,z')=\langle\varphi(z;W^{(0)}),\varphi(z';W^{(0)})\rangle,\forall z,z'\in\mathcal{Z}.$$
此外，我们考虑模型$f(z,W)$在初始值$W^(0)$处的线性化，使得我们有$f_\mathrm{lin}(z;W):=f(z;W^{(0)})+$ $\langle\varphi(z;W^{(0)}),W-W^{(0)}\rangle$。此外，以下等式成立$:f_\mathrm{lin}(z;W)=\langle\varphi(z;W^{(0)}),W-W^{(0)}\rangle$ 自 $f(z;W^{(0)})=0$ 通过初始化方案。我们可以看到，线性化函数$f_\mathrm{lin}(z;W)$是 RKHS 上的一个函数，内核为 ker$_m(z,z^{\prime})$ 。当使用$m\to\infty$对模型进行过参数化时，内核$\ker_m(z,z^{\prime})$收敛到 NTK 内核，该内核定义为$\mathrm{ker}_{\mathrm{ntk}}= \mathbb{E} _{\omega \sim N( 0, I_d/ d) }[ \mathrm{act}^{\prime }( \omega ^\top z) \cdot \mathrm{act}^{\prime }( \omega ^\top z^{\prime }) \cdot z^\top z^{\prime }]$ ,其中 act'是激活函数 act 的导数
# 3 Single-Agent MDP Setting
在本节中，我们在具有内核和神经函数逼近的单代理 MDP
绍了我们的理论结果
3.1 Kernel Function Approximation
我们提出的方法由具有给定外在奖励函数的无奖励探索阶段和计划阶段组成。在算法 1 和算法 2
中总结了勘探阶段和计划阶段。
具体来说，探索算法是具有函数逼近的值迭代算法的乐观变体。在算法 1 中，我们使用$Q_h^k$ 和$V_h^k$表示探索奖励的乐观 Q 函数和值函数。在探索阶段，代理不会访问真正的奖励函数，并根据值函数$V_h^k$确定的策略$\{\pi_h^k\}_{(h,k)\in[H]\times[K]}$探索$K$剧集的环境，并收集轨迹$\{s_h^k,a_h^k\}_{(h,k)\in[H]\times[K]}$ 以进行后续计划阶段。因此，我们不是直接近似 Q 函数，而是针对任何$(s,a)\in\mathcal{S}\times\mathcal{A}$ ,通过
5
|
[
[
0.1168212890625,
0.0921630859375,
0.88134765625,
0.09344482421875,
0.88134765625,
0.10797119140625,
0.1168212890625,
0.106689453125
],
[
0.11767578125,
0.11297607421875,
0.465576171875,
0.11297607421875,
0.465576171875,
0.126953125,
0.11767578125,
0.126953125
],
[
0.471435546875,
0.1458740234375,
0.5087890625,
0.1458740234375,
0.5087890625,
0.1610107421875,
0.471435546875,
0.1610107421875
],
[
0.50341796875,
0.14453125,
0.525390625,
0.14453125,
0.525390625,
0.1572265625,
0.50341796875,
0.1572265625
],
[
0.5,
0.15283203125,
0.64453125,
0.1546630859375,
0.64453125,
0.173583984375,
0.5,
0.1717529296875
],
[
0.85888671875,
0.1558837890625,
0.8837890625,
0.1558837890625,
0.8837890625,
0.173583984375,
0.85888671875,
0.173583984375
],
[
0.459228515625,
0.1634521484375,
0.529296875,
0.1666259765625,
0.52783203125,
0.1849365234375,
0.45751953125,
0.1817626953125
],
[
0.3544921875,
0.1572265625,
0.453369140625,
0.1572265625,
0.453369140625,
0.1717529296875,
0.3544921875,
0.1717529296875
],
[
0.11358642578125,
0.199462890625,
0.88232421875,
0.1988525390625,
0.88232421875,
0.216552734375,
0.11358642578125,
0.2171630859375
],
[
0.11767578125,
0.2203369140625,
0.86669921875,
0.2203369140625,
0.86669921875,
0.2342529296875,
0.11767578125,
0.2342529296875
],
[
0.861328125,
0.2235107421875,
0.87841796875,
0.2235107421875,
0.87841796875,
0.2310791015625,
0.861328125,
0.2310791015625
],
[
0.11767578125,
0.238037109375,
0.88134765625,
0.238037109375,
0.88134765625,
0.25244140625,
0.11767578125,
0.25244140625
],
[
0.11358642578125,
0.251953125,
0.8857421875,
0.251953125,
0.8857421875,
0.27587890625,
0.11358642578125,
0.27587890625
],
[
0.11358642578125,
0.270751953125,
0.8837890625,
0.271484375,
0.8837890625,
0.29345703125,
0.11358642578125,
0.292236328125
],
[
0.11767578125,
0.292236328125,
0.8837890625,
0.292236328125,
0.8837890625,
0.31005859375,
0.11767578125,
0.31005859375
],
[
0.1168212890625,
0.314453125,
0.8798828125,
0.313232421875,
0.8798828125,
0.32763671875,
0.1168212890625,
0.328857421875
],
[
0.11767578125,
0.333984375,
0.8837890625,
0.333984375,
0.8837890625,
0.348388671875,
0.11767578125,
0.348388671875
],
[
0.1168212890625,
0.352783203125,
0.88134765625,
0.352783203125,
0.88134765625,
0.3662109375,
0.1168212890625,
0.3662109375
],
[
0.1168212890625,
0.371826171875,
0.2254638671875,
0.371826171875,
0.2254638671875,
0.385009765625,
0.1168212890625,
0.385009765625
],
[
0.301513671875,
0.404052734375,
0.689453125,
0.40283203125,
0.689453125,
0.42041015625,
0.301513671875,
0.421630859375
],
[
0.85693359375,
0.40478515625,
0.8837890625,
0.40478515625,
0.8837890625,
0.42236328125,
0.85693359375,
0.42236328125
],
[
0.1217041015625,
0.438232421875,
0.88134765625,
0.438232421875,
0.88134765625,
0.45263671875,
0.1217041015625,
0.45263671875
],
[
0.1168212890625,
0.455810546875,
0.40185546875,
0.458251953125,
0.40185546875,
0.472900390625,
0.1168212890625,
0.47021484375
],
[
0.306396484375,
0.490478515625,
0.69287109375,
0.490478515625,
0.69287109375,
0.50830078125,
0.306396484375,
0.50830078125
],
[
0.1168212890625,
0.5234375,
0.8857421875,
0.5234375,
0.8857421875,
0.54150390625,
0.1168212890625,
0.54150390625
],
[
0.1192626953125,
0.54443359375,
0.88134765625,
0.54443359375,
0.88134765625,
0.5595703125,
0.1192626953125,
0.5595703125
],
[
0.11767578125,
0.564453125,
0.88232421875,
0.564453125,
0.88232421875,
0.57763671875,
0.11767578125,
0.57763671875
],
[
0.11767578125,
0.58349609375,
0.88134765625,
0.58349609375,
0.88134765625,
0.59765625,
0.11767578125,
0.59765625
],
[
0.111083984375,
0.5986328125,
0.8056640625,
0.5966796875,
0.8056640625,
0.6171875,
0.111083984375,
0.619140625
],
[
0.11358642578125,
0.642578125,
0.47802734375,
0.64453125,
0.47802734375,
0.66357421875,
0.11358642578125,
0.6611328125
],
[
0.11767578125,
0.6787109375,
0.8798828125,
0.6787109375,
0.8798828125,
0.693359375,
0.11767578125,
0.693359375
],
[
0.1168212890625,
0.69873046875,
0.27783203125,
0.69873046875,
0.27783203125,
0.7119140625,
0.1168212890625,
0.7119140625
],
[
0.1168212890625,
0.736328125,
0.48193359375,
0.736328125,
0.48193359375,
0.75048828125,
0.1168212890625,
0.75048828125
],
[
0.1192626953125,
0.7666015625,
0.88134765625,
0.7666015625,
0.88134765625,
0.7763671875,
0.1192626953125,
0.7763671875
],
[
0.11767578125,
0.7841796875,
0.35546875,
0.7841796875,
0.35546875,
0.79736328125,
0.11767578125,
0.79736328125
],
[
0.146240234375,
0.80322265625,
0.8798828125,
0.80322265625,
0.8798828125,
0.81640625,
0.146240234375,
0.81640625
],
[
0.11767578125,
0.82080078125,
0.88134765625,
0.82080078125,
0.88134765625,
0.83544921875,
0.11767578125,
0.83544921875
],
[
0.111083984375,
0.833984375,
0.8837890625,
0.8359375,
0.8837890625,
0.861328125,
0.111083984375,
0.859375
],
[
0.11767578125,
0.859375,
0.88134765625,
0.859375,
0.88134765625,
0.87255859375,
0.11767578125,
0.87255859375
],
[
0.491943359375,
0.89794921875,
0.50732421875,
0.89794921875,
0.50732421875,
0.91357421875,
0.491943359375,
0.91357421875
]
] |
[
"BaiandLee，2019）。具体来说，在我们的工作中，我们有一个带有2m神经元和权重（u,W）的两",
"层神经网络f(;b,W）：2→R，可以表示为",
"1",
"2m",
"vi act(WTz),",
"(1)",
"V2m=1",
"f(z; v, W) =",
"，其中act是激活函数，=[u1,..，2m]T和W=[Wi,W2，·,W2m］。在这里，我们假设",
"z=(s,a)或z=（s,a,b)与zEz满足z2=1，i。e。，z在Rd中的单位超球面上归",
"",
"化。令W(）为W的初始值，(）为的初始化。上述模型的初始化步骤执行如下：我们让",
"vi.～Unif(-1,1)和w)～N(0,Ia/d)为所有ie[m)，其中Ia是Rdxd中的单位矩阵，而",
"u)=-um，w()=0m为所有ie(m+1,2m)。这里我们让N(0,Ia/d)表示高斯分布。在",
"本文中，我们将固定为(O），并且为了便于理论分析，我们仅学习W。因此，我们用f(z;W)",
"表示f（z；，,W），以简化表示法。这种神经网络模型在最近关于神经网络分析的论文中得到了广",
"泛的研究，例如。g。，Gaoetal.（2019)；BaiandLee（2019）。当模型被过度参数化时，我。e。",
"m足够大，我们可以用神经切线核（NTK）（Jacotetal.，2018）来表征训练这种神经网络的动力学。",
"这里我们定义",
"p(z; W) := [Vwif(z; W) T,..- , Vw2mf(z; W)TjT,",
"(2)",
"其中我们让Vwf(z;W）是一个列向量，使得(z;W）ER2md。因此，以W(O)初始化W的随",
"机性为条件，我们进一步定义了内核",
"kerm(z,z2') = (0(z; W(0)), (z; W(0)), Vz,z' e z.",
"此外，我们考虑模型f(z,W）在初始值W(O)处的线性化，使得我们有fin(z;W）：=f(z;W(0)）+",
"((z;W(0)),W-W(0))。此外,以下等式成立:fiin(z;W)=《(z;W(0)),W-W(0))自f(z;W(0))=0",
"通过初始化方案。我们可以看到，线性化函数fiin（z;W）是RKHS上的一个函数，内核为kerm（z，z）",
"。当使用m→o对模型进行过参数化时，内核kermz,z）收敛到NTK内核，该内核定义为",
"kerntk=Ew~N(0,la/d)[act'(wTz)·act'(wTz\"）·zTz]，其中act'是激活函数act 的导数",
"3Single-Agent MDP Setting",
"在本节中，我们在具有内核和神经函数逼近的单代理MDP设置下介绍我们的方法。然后，我们介",
"绍了我们的理论结果",
"3.1 Kernel Function Approximation",
"我们提出的方法由具有给定外在奖励函数的无奖励探索阶段和计划阶段组成。在算法1和算法2",
"中总结了勘探阶段和计划阶段。",
"具体来说，探索算法是具有函数逼近的值选代算法的乐观变体。在算法1中，我们使用Q",
"和Vk表示探索奖励的乐观Q函数和值函数。在探索阶段，代理不会访问真正的奖励函数，并根",
"据值函数V确定的策略(元h(a,)[][K]探索K剧集的环境，并收集轨迹（sh,%)(h,)e[H]x[K]",
"以进行后续计划阶段。因此，我们不是直接近似Q函数，而是针对任何（s,a）ES×A，通过",
"5"
] |
{
"arxiv_id": "2110.09771",
"language": "zh",
"page_idx": 4
}
|
||
2303.07771-zh-0
|
2303.07771-zh-0.jpg
|
$\underline{\text{Published at ICLR 2023 Workshop on Domain Generalization}}$
血细胞形态学中用于稳健单细胞分类的不平衡域泛化
Rao Muhammad Umer$^1$, Armin Gruber$^1,2$, Sayedali Shetab Boushehri$^1,3$, Christian Metak$^1.$
Carsten Marr$^{1}$
$\{$ christian.metak, carsten.marr $\}$ @helmholtz-muenchen.de
$\begin{array}{c}{\text{Institute of AI for Health, Helmholtz Zentrum München- German Research Center for}}\end{array}$
Environmental Health, Neuherberg 85764, Germany.$^1$
Laboratory of Leukemia Diagnostics, Department of Medicine III, University Hospital,
LMU Munich, Munich, Germany.$^2$
Data Science, Pharmaceutical Research and Early Development Informatics (pREDi),
Roche Innovation Center Munich, Germany.$^3$
# Abstract
Accurate morphological classification of white blood cells (WBCs) is an important step in the diagnosis of leukemia, a disease in which nonfunctional blast cells accumulate in the bone marrow. Recently, deep convolutional neural networks (CNNs) have been successfully used to classify leukocytes by training them on single-cell images from a specific domain. Most CNN models assume that the distributions of the training and test data are similar, i.e., the data are independently and identically distributed. Therefore, they are not robust to different staining procedures, magnifications, resolutions, scanners, or imaging protocols, as well as variations in clinical centers or patient cohorts. In addition, domain-specific data imbalances affect the $\dot{\text{generalization performance of classifiers.}}$Here, we train a robust CNN for WBC classification by addressing cross-domain data imbalance and domain shifts. To this end, we use two loss functions and demonstrate their effectiveness in out-of-distribution (OOD) generalization. Our approach achieves the best F1 macro score compared to other existing methods and is able to consider rare cell types. This is the first demonstration of imbalanced domain generalization in hematological cytomorphology and paves the way for robust single cell classification methods for the application in laboratories and clinics.
1 Introduction
血液学涉及血液，血液形成组织和血液相关疾病的研究。血液病的精确和早期诊断对于成功治疗至关重要。几十年来，对染色的外周血(PB)和骨髓(BM)样本中的血细胞进行显微镜检查和分类一直是诊断血液恶性肿瘤的关键步骤。使用光学显微镜进行细胞形态学检查$\overline{\mathrm{E}}($Walter et al., 2022)仍然是血液学诊断的骨干之一，通常代表检查的第一步，并指导其他方法，例如免疫表型，细胞遗传学和分子遗传学。在形态学检查期间，血液样本由血液学家进行显微镜评估，并筛选是否存在可以指示白血病等疾病的非典型细胞群。通常，每个样本至少有 200 个细胞必须根据当前的临床指南进行分类。这种在显微镜下的手动评估和分类可能是繁琐，重复和耗时的。此外，它在很大程度上依赖于训练有素和经验丰富的工作人员，并且由于人为因素 i.e., 有限的观察者内部/观察者之间的可重复性而容易发生变化深度学习 (DL) 模型在解决各个领域的实际分类任务方面显示出巨大的潜力，例如计算机视觉$\overline{\mathrm{E}}($He et al., 2016) , 自然语言处理$\boxed{\mathrm{E}}($Brown et al., 2020) 以及包括医学成像$\underline{\mathrm{E}}($Matek et al., 2019; Wang et al., 2019) 和组织病理学$\mathbb{E}($Komura & Ishikawa, 2018; Hägele et al., 2020) 在内的医疗保健。它们通常是在隐含的假设下开发和测试的，即火车和测试数据是独立且相同地从同一分布(IID)中提取的。。推广到看不见的测试域对人类来说是很自然的， 但对机器来说却很有挑战性。域泛化 (DG)旨在通过利用来自不同域的数据来训练对分布变化具有鲁棒性的域不变模型$\boxed{\mathrm{F}}($Zhou et al.,2022; Wang et al.,2022)。然而，来自多个不同域的实际数据通常表现出不平衡的标签分布 i.e., 少数类在一个域中包含非常大量的样本， 而在另一个域中只有少数样本或没有样本(参见图。1)。此外，严重不平衡的数据分布是
1
|
[
[
0.176513671875,
0.034088134765625,
0.61767578125,
0.034088134765625,
0.61767578125,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.17724609375,
0.10479736328125,
0.81884765625,
0.10479736328125,
0.81884765625,
0.12249755859375,
0.17724609375,
0.12249755859375
],
[
0.188720703125,
0.1514892578125,
0.84619140625,
0.1514892578125,
0.84619140625,
0.1654052734375,
0.188720703125,
0.1654052734375
],
[
0.1871337890625,
0.1685791015625,
0.281982421875,
0.1685791015625,
0.281982421875,
0.179931640625,
0.1871337890625,
0.179931640625
],
[
0.1895751953125,
0.18115234375,
0.61279296875,
0.18115234375,
0.61279296875,
0.1944580078125,
0.1895751953125,
0.1944580078125
],
[
0.1829833984375,
0.1932373046875,
0.8056640625,
0.1932373046875,
0.8056640625,
0.20703125,
0.1829833984375,
0.20703125
],
[
0.1846923828125,
0.208984375,
0.56787109375,
0.208984375,
0.56787109375,
0.222900390625,
0.1846923828125,
0.222900390625
],
[
0.1829833984375,
0.2235107421875,
0.80126953125,
0.2235107421875,
0.80126953125,
0.2374267578125,
0.1829833984375,
0.2374267578125
],
[
0.1846923828125,
0.236083984375,
0.4306640625,
0.2374267578125,
0.4306640625,
0.250732421875,
0.1846923828125,
0.25
],
[
0.1846923828125,
0.251953125,
0.79736328125,
0.251953125,
0.79736328125,
0.26513671875,
0.1846923828125,
0.26513671875
],
[
0.1846923828125,
0.266357421875,
0.5048828125,
0.266357421875,
0.5048828125,
0.27978515625,
0.1846923828125,
0.27978515625
],
[
0.463134765625,
0.31640625,
0.53759765625,
0.31640625,
0.53759765625,
0.330810546875,
0.463134765625,
0.330810546875
],
[
0.2344970703125,
0.347900390625,
0.7646484375,
0.347900390625,
0.7646484375,
0.361083984375,
0.2344970703125,
0.361083984375
],
[
0.2344970703125,
0.3623046875,
0.765625,
0.3623046875,
0.765625,
0.375732421875,
0.2344970703125,
0.375732421875
],
[
0.2344970703125,
0.375732421875,
0.765625,
0.375732421875,
0.765625,
0.389404296875,
0.2344970703125,
0.389404296875
],
[
0.233642578125,
0.39013671875,
0.7548828125,
0.389404296875,
0.7548828125,
0.40283203125,
0.233642578125,
0.404052734375
],
[
0.2344970703125,
0.40478515625,
0.76318359375,
0.40478515625,
0.76318359375,
0.415283203125,
0.2344970703125,
0.415283203125
],
[
0.2320556640625,
0.417236328125,
0.7646484375,
0.417236328125,
0.7646484375,
0.43115234375,
0.2320556640625,
0.43115234375
],
[
0.233642578125,
0.43115234375,
0.7646484375,
0.43115234375,
0.7646484375,
0.4443359375,
0.233642578125,
0.4443359375
],
[
0.233642578125,
0.44580078125,
0.765625,
0.44580078125,
0.765625,
0.458984375,
0.233642578125,
0.458984375
],
[
0.233642578125,
0.458984375,
0.7646484375,
0.458251953125,
0.7646484375,
0.472900390625,
0.233642578125,
0.473388671875
],
[
0.233642578125,
0.472900390625,
0.765625,
0.472900390625,
0.765625,
0.486083984375,
0.233642578125,
0.486083984375
],
[
0.233642578125,
0.4873046875,
0.765625,
0.48486328125,
0.765625,
0.498779296875,
0.233642578125,
0.50048828125
],
[
0.2344970703125,
0.50048828125,
0.765625,
0.50048828125,
0.765625,
0.5146484375,
0.2344970703125,
0.5146484375
],
[
0.233642578125,
0.5146484375,
0.7646484375,
0.5146484375,
0.7646484375,
0.52783203125,
0.233642578125,
0.52783203125
],
[
0.2344970703125,
0.52880859375,
0.765625,
0.52880859375,
0.765625,
0.54248046875,
0.2344970703125,
0.54248046875
],
[
0.233642578125,
0.54248046875,
0.7646484375,
0.54248046875,
0.7646484375,
0.55615234375,
0.233642578125,
0.55615234375
],
[
0.233642578125,
0.55615234375,
0.765625,
0.55615234375,
0.765625,
0.5693359375,
0.233642578125,
0.5693359375
],
[
0.233642578125,
0.568359375,
0.76318359375,
0.568359375,
0.76318359375,
0.58203125,
0.233642578125,
0.58203125
],
[
0.233642578125,
0.58203125,
0.765625,
0.58349609375,
0.765625,
0.59765625,
0.233642578125,
0.5966796875
],
[
0.2320556640625,
0.59521484375,
0.404296875,
0.59765625,
0.404296875,
0.611328125,
0.2320556640625,
0.609375
],
[
0.1732177734375,
0.63818359375,
0.2034912109375,
0.63818359375,
0.2034912109375,
0.65087890625,
0.1732177734375,
0.65087890625
],
[
0.205078125,
0.63720703125,
0.31787109375,
0.63818359375,
0.31787109375,
0.65283203125,
0.205078125,
0.6513671875
],
[
0.176513671875,
0.66845703125,
0.82275390625,
0.66845703125,
0.82275390625,
0.681640625,
0.176513671875,
0.681640625
],
[
0.1732177734375,
0.681640625,
0.82177734375,
0.68115234375,
0.82177734375,
0.69580078125,
0.1732177734375,
0.6962890625
],
[
0.1748046875,
0.6943359375,
0.82177734375,
0.69580078125,
0.82177734375,
0.71044921875,
0.1748046875,
0.708984375
],
[
0.176513671875,
0.708984375,
0.57421875,
0.708984375,
0.57421875,
0.72265625,
0.176513671875,
0.72265625
],
[
0.57666015625,
0.71044921875,
0.82275390625,
0.71044921875,
0.82275390625,
0.7236328125,
0.57666015625,
0.7236328125
],
[
0.17724609375,
0.72607421875,
0.82177734375,
0.72607421875,
0.82177734375,
0.736328125,
0.17724609375,
0.736328125
],
[
0.176513671875,
0.73779296875,
0.82275390625,
0.73779296875,
0.82275390625,
0.751953125,
0.176513671875,
0.751953125
],
[
0.176513671875,
0.751953125,
0.82275390625,
0.751953125,
0.82275390625,
0.76513671875,
0.176513671875,
0.76513671875
],
[
0.176513671875,
0.76513671875,
0.82275390625,
0.76513671875,
0.82275390625,
0.77880859375,
0.176513671875,
0.77880859375
],
[
0.17724609375,
0.77978515625,
0.814453125,
0.77978515625,
0.814453125,
0.79345703125,
0.17724609375,
0.79345703125
],
[
0.176513671875,
0.7998046875,
0.82177734375,
0.7998046875,
0.82177734375,
0.81298828125,
0.176513671875,
0.81298828125
],
[
0.176513671875,
0.814453125,
0.82421875,
0.814453125,
0.82421875,
0.82763671875,
0.176513671875,
0.82763671875
],
[
0.176513671875,
0.82763671875,
0.82177734375,
0.82763671875,
0.82177734375,
0.84130859375,
0.176513671875,
0.84130859375
],
[
0.176513671875,
0.84326171875,
0.8203125,
0.84326171875,
0.8203125,
0.85400390625,
0.176513671875,
0.85400390625
],
[
0.176513671875,
0.85498046875,
0.82275390625,
0.85595703125,
0.82275390625,
0.869140625,
0.176513671875,
0.86865234375
],
[
0.1748046875,
0.8671875,
0.82275390625,
0.869140625,
0.82275390625,
0.8837890625,
0.1748046875,
0.8818359375
],
[
0.1748046875,
0.8818359375,
0.82177734375,
0.88330078125,
0.82177734375,
0.89794921875,
0.1748046875,
0.896484375
],
[
0.176513671875,
0.89794921875,
0.82177734375,
0.89794921875,
0.82177734375,
0.9111328125,
0.176513671875,
0.9111328125
],
[
0.176513671875,
0.9111328125,
0.82275390625,
0.9111328125,
0.82275390625,
0.9248046875,
0.176513671875,
0.9248046875
],
[
0.490234375,
0.94873046875,
0.50732421875,
0.94873046875,
0.50732421875,
0.9619140625,
0.490234375,
0.9619140625
]
] |
[
"Published at ICLR 2023 Workshop on Domain Generalization",
"血细胞形态学中用于稳健单细胞分类的不平衡域泛化",
"Rao Muhammad Umerl, Armin Gruberl,2, Sayedali Shetab Boushehril,3, Christian Metak!",
"CarstenMarr",
"( christian.metak, carsten.marr ↓ @helmholtz-muenchen.de",
"Institute of AI for Health, Helmholtz Zentrum Munchen - German Research Center for",
"Environmental Health, Neuherberg 85764, Germany.1",
"Laboratory of Leukemia Diagnostics, Department of Medicine III, University Hospital",
"LMU Munich, Munich, Germany.2",
"Data Science, Pharmaceutical Research and Early Development Informatics (pREDi),",
"Roche Innovation Center Munich, Germany.",
"Abstract",
"Accurate morphological classification of white blood cells (WBCs) is an im-",
"portant step in the diagnosis of leukemia, a disease in which nonfunctional",
"blast cells accumulate in the bone marrow. Recently, deep convolutional",
"neural networks (CNNs)have been successfully used to classifyleukocyte",
"bytraining them on single-cell imagesfrom a specific domain.Most CNN",
"models assume that the distributions of the training and test data are simi",
"lar, i.e., the data are independently and identically distributed. Therefore,",
"they are not robust to different staining procedures, magnifications, resolu-",
"tions, scanners, or imaging protocols, as well as variations in clinical centers",
"or patient cohorts. In addition, domain-specific data imbalances affect the",
"generalization performance of classifiers.Here,we train a robust CNN",
"for WBC classification by addressing cross-domain data imbalance and do-",
"main shifts. To this end, we use two loss functions and demonstrate their",
"effectiveness in out-of-distribution (OOD) generalization. Our approach",
"achieves the best F1 macro score compared to other existing methods and",
"is able to consider rare cell types. This is the first demonstration of imbal-",
"anced domain generalization in hematological cytomorphology and paves",
"the way for robust single cell classification methods for the application in",
"laboratories and clinics.",
"1",
"Introduction",
"血液学涉及血液，血液形成组织和血液相关疾病的研究。血液病的精确和早期诊断对于成",
"功治疗至关重要。几十年来，对染色的外周血（PB）和骨髓（BM）样本中的血细胞进行显",
"微镜检查和分类一直是诊断血液恶性肿瘤的关键步骤。使用光学显微镜进行细胞形态学检",
"查E（Walteretal.,2022)仍然是血液学诊断的骨干之一",
"，通常代表检查的第一步，并指导",
"其他方法，例如免疫表型，细胞遗传学和分子遗传学。在形态学检查期间，血液样本由血液",
"学家进行显微镜评估，并筛选是否存在可以指示白血病等疾病的非典型细胞群。通常，每个",
"样本至少有200个细胞必须根据当前的临床指南进行分类。这种在显微镜下的手动评估和",
"分类可能是繁琐，重复和耗时的。此外，它在很大程度上依赖于训练有素和经验丰富的工作",
"人员，并且由于人为因素i.e.，有限的观察者内部/观察者之间的可重复性而容易发生变化",
"深度学习（DL）模型在解决各个领域的实际分类任务方面显示出巨大的潜力，例如计算机",
"视觉E(Heetal.，2016），自然语言处理E(Brownetal.，2020）以及包括医学成像E(Matek",
"et al.,2019;Wangetal.,2019)和组织病理学E(Komura&Ishikawa,2018;Hageleetal.",
"2020）在内的医疗保健。它们通常是在隐含的假设下开发和测试的，即火车和测试数据是独",
"立且相同地从同一分布（IID）中提取的。。推广到看不见的测试域对人类来说是很自然的，",
"但对机器来说却很有挑战性。域泛化（DG）旨在通过利用来自不同域的数据来训练对分布",
"变化具有鲁棒性的域不变模型E(Zhouetal.,2022；Wangetal.，2022）。然而，来自多个不",
"同域的实际数据通常表现出不平衡的标签分布i.e.，少数类在一个域中包含非常大量的样本，",
"而在另一个域中只有少数样本或没有样本（参见图。1）。此外，严重不平衡的数据分布是",
"1"
] |
{
"arxiv_id": "2303.07771",
"language": "zh",
"page_idx": 0
}
|
||
2303.07771-zh-4
|
2303.07771-zh-4.jpg
|
$\underline{\text{Published at ICLR 2023 Workshop on Domain Generalization}}$
3.2 Evaluation Metrics and Results
由于数据集不平衡，我们的目标是在看不见的测试域(OOD)上获得较高的 F1 宏观得分，这使得每个类的权重相等，而不管其基数如何。除此之外，我们还报告了 F1-micro 分数，它给出了每个班级的加权平均值我们在表 1 和表 2 中报告
<FigureHere>
( b) Target ( INT$\_$20) domain testset ( unseen)
了通过对源域 (Acevedo\_20 和Matek\_19)验证集和看不见的测试域 (INT\_20)进行五重交叉验证确定的不平衡 DG 分类结果。我们将我们的模型与一个像 ERME(Vapnik, 1999) 这样的香草训练模型$\mathbb{E}(1)$ 以及两种当前的 SOTADG 方法进行了比较，例如 DANNE$( Ganin$ et al., 2016) 和 CORALE(Sun $\&$ Saenko, 2016)。我们训练模型的两种不同变体，一种是对编码器部分(表示形式 z)和 MLP 分类器部分进行耦合训练，另一种(表示为 Ours $^{+}$)是对编码器部分和编码器部分进行解耦训练。分类器部分使用类平衡采样。我们在 INT\_20 测试集上实现了最佳的 F1-marco 得分，而DG 方法 DANN 由于在训练过程中假设了数据平衡而比香草ERM 模型表现更差。InFig。3 ,我们还将模型的混淆矩阵与ERM 方法进行了比较。我们的模型通过识别少数类别的样本 Figure 3: Confusion matrices show an improved clas-
表现更好，而 ERM 错过了少数
$\textbf{类别的样本，例如非典型淋巴细胞}_{\mathrm{Cyte\_atypical~class~with~our~method~compared~to~the~stan-}}$
dard ERM model.
4 Conclusion
我们开发了一个强大的 CNN 模型，用于血液细胞形态学中的非分布泛化，解决了三个主要挑战：数据不平衡，域转移和缺失类。我们展示了如何通过优化潜在特征空间中的损失函数和网络的输出日志，针对不同的领域改进现有的预训练深度模型。我们的工作展示了如何解决血液学单一 WBC 分类中的生物学，流行病学和技术变异性，以训练强大的基于 AI 的细胞分类器，为其在临床环境中的安全和生产性使用铺平道路。
References
Andrea Acevedo, Anna Merino, Santiago Alférez, Ángel Molina, Laura Boldú, and José
Rodellar. A dataset of microscopic peripheral blood cell images for development of auto-
matic recognition systems. Data in brief, 30, 2020.
Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla
Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, IIya Sutskever, and Dario Amodei. Language models are few-shot learners. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (eds.), Advances in Neural Information Processing Systems, volume 33, pp. 1877-1901, 2020.
|
[
[
0.176513671875,
0.034088134765625,
0.619140625,
0.034088134765625,
0.619140625,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.1732177734375,
0.10479736328125,
0.443603515625,
0.10479736328125,
0.443603515625,
0.11871337890625,
0.1732177734375,
0.11871337890625
],
[
0.1748046875,
0.1319580078125,
0.82275390625,
0.1307373046875,
0.82275390625,
0.1458740234375,
0.1748046875,
0.146484375
],
[
0.1748046875,
0.146484375,
0.82275390625,
0.146484375,
0.82275390625,
0.160400390625,
0.1748046875,
0.160400390625
],
[
0.176513671875,
0.160400390625,
0.82421875,
0.160400390625,
0.82421875,
0.1749267578125,
0.176513671875,
0.1749267578125
],
[
0.1748046875,
0.1749267578125,
0.400390625,
0.1749267578125,
0.400390625,
0.1881103515625,
0.1748046875,
0.1881103515625
],
[
0.178955078125,
0.1881103515625,
0.398681640625,
0.1881103515625,
0.398681640625,
0.2020263671875,
0.178955078125,
0.2020263671875
],
[
0.5390625,
0.1962890625,
0.580078125,
0.1962890625,
0.580078125,
0.203857421875,
0.5390625,
0.203857421875
],
[
0.68115234375,
0.1944580078125,
0.701171875,
0.1962890625,
0.69921875,
0.205810546875,
0.6806640625,
0.203857421875
],
[
0.176513671875,
0.2020263671875,
0.400390625,
0.2020263671875,
0.400390625,
0.21533203125,
0.176513671875,
0.21533203125
],
[
0.1748046875,
0.2139892578125,
0.398681640625,
0.21533203125,
0.398681640625,
0.2298583984375,
0.1748046875,
0.2291259765625
],
[
0.17724609375,
0.2298583984375,
0.400390625,
0.2298583984375,
0.400390625,
0.24365234375,
0.17724609375,
0.24365234375
],
[
0.17724609375,
0.2442626953125,
0.386474609375,
0.2442626953125,
0.386474609375,
0.255126953125,
0.17724609375,
0.255126953125
],
[
0.176513671875,
0.255126953125,
0.398681640625,
0.255615234375,
0.398681640625,
0.270751953125,
0.176513671875,
0.26953125
],
[
0.176513671875,
0.270751953125,
0.400390625,
0.270751953125,
0.400390625,
0.2841796875,
0.176513671875,
0.2841796875
],
[
0.1748046875,
0.28271484375,
0.400390625,
0.2841796875,
0.400390625,
0.298583984375,
0.1748046875,
0.29736328125
],
[
0.178955078125,
0.298583984375,
0.400390625,
0.298583984375,
0.400390625,
0.3125,
0.178955078125,
0.3125
],
[
0.50634765625,
0.30615234375,
0.52685546875,
0.30810546875,
0.525390625,
0.31494140625,
0.5048828125,
0.313232421875
],
[
0.1748046875,
0.3125,
0.398681640625,
0.3125,
0.398681640625,
0.32568359375,
0.1748046875,
0.32568359375
],
[
0.176513671875,
0.32763671875,
0.398681640625,
0.32763671875,
0.398681640625,
0.33837890625,
0.176513671875,
0.33837890625
],
[
0.176513671875,
0.34033203125,
0.326904296875,
0.34033203125,
0.326904296875,
0.354248046875,
0.176513671875,
0.354248046875
],
[
0.3203125,
0.341552734375,
0.398681640625,
0.341552734375,
0.398681640625,
0.352783203125,
0.3203125,
0.352783203125
],
[
0.43701171875,
0.33837890625,
0.78173828125,
0.33837890625,
0.78173828125,
0.3515625,
0.43701171875,
0.3515625
],
[
0.176513671875,
0.354248046875,
0.400390625,
0.354248046875,
0.400390625,
0.367431640625,
0.176513671875,
0.367431640625
],
[
0.5458984375,
0.354736328125,
0.5849609375,
0.354736328125,
0.5849609375,
0.365478515625,
0.5458984375,
0.365478515625
],
[
0.68115234375,
0.35595703125,
0.69921875,
0.35595703125,
0.69921875,
0.3642578125,
0.68115234375,
0.3642578125
],
[
0.17724609375,
0.367431640625,
0.393798828125,
0.367431640625,
0.393798828125,
0.380615234375,
0.17724609375,
0.380615234375
],
[
0.57421875,
0.37744140625,
0.58642578125,
0.37744140625,
0.58642578125,
0.380615234375,
0.57421875,
0.380615234375
],
[
0.6611328125,
0.37744140625,
0.66845703125,
0.37744140625,
0.66845703125,
0.380615234375,
0.6611328125,
0.380615234375
],
[
0.52392578125,
0.386474609375,
0.53173828125,
0.386474609375,
0.53173828125,
0.389404296875,
0.52392578125,
0.389404296875
],
[
0.74267578125,
0.393310546875,
0.74658203125,
0.393310546875,
0.74658203125,
0.3984375,
0.74267578125,
0.3984375
],
[
0.176513671875,
0.3818359375,
0.400390625,
0.3818359375,
0.400390625,
0.395751953125,
0.176513671875,
0.395751953125
],
[
0.176513671875,
0.395751953125,
0.398681640625,
0.395751953125,
0.398681640625,
0.4091796875,
0.176513671875,
0.4091796875
],
[
0.176513671875,
0.4091796875,
0.400390625,
0.4091796875,
0.400390625,
0.42236328125,
0.176513671875,
0.42236328125
],
[
0.1748046875,
0.421630859375,
0.400390625,
0.42236328125,
0.400390625,
0.4375,
0.1748046875,
0.436279296875
],
[
0.176513671875,
0.4375,
0.400390625,
0.4375,
0.400390625,
0.45068359375,
0.176513671875,
0.45068359375
],
[
0.1748046875,
0.448974609375,
0.400390625,
0.449462890625,
0.400390625,
0.464111328125,
0.1748046875,
0.46337890625
],
[
0.583984375,
0.458984375,
0.6396484375,
0.458984375,
0.6396484375,
0.466552734375,
0.583984375,
0.466552734375
],
[
0.73046875,
0.458984375,
0.7490234375,
0.458984375,
0.7490234375,
0.466552734375,
0.73046875,
0.466552734375
],
[
0.708984375,
0.4609375,
0.73046875,
0.4609375,
0.73046875,
0.464111328125,
0.708984375,
0.464111328125
],
[
0.176513671875,
0.464111328125,
0.400390625,
0.464111328125,
0.400390625,
0.477783203125,
0.176513671875,
0.477783203125
],
[
0.496826171875,
0.464111328125,
0.74853515625,
0.46533203125,
0.74853515625,
0.4873046875,
0.496826171875,
0.486083984375
],
[
0.50341796875,
0.460205078125,
0.51318359375,
0.460205078125,
0.51318359375,
0.464111328125,
0.50341796875,
0.464111328125
],
[
0.1732177734375,
0.4765625,
0.40283203125,
0.479248046875,
0.40283203125,
0.492431640625,
0.1732177734375,
0.490478515625
],
[
0.191162109375,
0.492431640625,
0.400390625,
0.492431640625,
0.400390625,
0.505859375,
0.191162109375,
0.505859375
],
[
0.1748046875,
0.50390625,
0.400390625,
0.5048828125,
0.400390625,
0.51953125,
0.1748046875,
0.51806640625
],
[
0.459228515625,
0.50048828125,
0.76123046875,
0.50048828125,
0.76123046875,
0.5146484375,
0.459228515625,
0.5146484375
],
[
0.1732177734375,
0.51806640625,
0.40185546875,
0.51953125,
0.40185546875,
0.5341796875,
0.1732177734375,
0.53271484375
],
[
0.406005859375,
0.52783203125,
0.41259765625,
0.52783203125,
0.41259765625,
0.5322265625,
0.406005859375,
0.5322265625
],
[
0.414306640625,
0.52587890625,
0.82421875,
0.52587890625,
0.82421875,
0.5390625,
0.414306640625,
0.5390625
],
[
0.17724609375,
0.5341796875,
0.40283203125,
0.5341796875,
0.40283203125,
0.54736328125,
0.17724609375,
0.54736328125
],
[
0.1707763671875,
0.54052734375,
0.830078125,
0.5517578125,
0.830078125,
0.5771484375,
0.1707763671875,
0.56494140625
],
[
0.414306640625,
0.54052734375,
0.826171875,
0.54052734375,
0.826171875,
0.5537109375,
0.414306640625,
0.5537109375
],
[
0.414306640625,
0.5673828125,
0.5458984375,
0.5673828125,
0.5458984375,
0.58056640625,
0.414306640625,
0.58056640625
],
[
0.176513671875,
0.58837890625,
0.1895751953125,
0.58837890625,
0.1895751953125,
0.5986328125,
0.176513671875,
0.5986328125
],
[
0.20751953125,
0.58544921875,
0.303955078125,
0.5869140625,
0.303955078125,
0.6015625,
0.20751953125,
0.599609375
],
[
0.17724609375,
0.6220703125,
0.8203125,
0.6220703125,
0.8203125,
0.6318359375,
0.17724609375,
0.6318359375
],
[
0.176513671875,
0.6337890625,
0.82177734375,
0.6337890625,
0.82177734375,
0.64794921875,
0.176513671875,
0.64794921875
],
[
0.176513671875,
0.64794921875,
0.82177734375,
0.64794921875,
0.82177734375,
0.6611328125,
0.176513671875,
0.6611328125
],
[
0.176513671875,
0.6611328125,
0.82177734375,
0.6611328125,
0.82177734375,
0.67529296875,
0.176513671875,
0.67529296875
],
[
0.1748046875,
0.67529296875,
0.625,
0.67529296875,
0.625,
0.689453125,
0.1748046875,
0.689453125
],
[
0.1732177734375,
0.71337890625,
0.265625,
0.71630859375,
0.265625,
0.73095703125,
0.1732177734375,
0.72802734375
],
[
0.176513671875,
0.7392578125,
0.82421875,
0.7392578125,
0.82421875,
0.75390625,
0.176513671875,
0.75390625
],
[
0.191162109375,
0.75390625,
0.826171875,
0.75390625,
0.826171875,
0.76806640625,
0.191162109375,
0.76806640625
],
[
0.1895751953125,
0.76806640625,
0.5537109375,
0.76806640625,
0.5537109375,
0.7822265625,
0.1895751953125,
0.7822265625
],
[
0.1748046875,
0.7998046875,
0.82275390625,
0.7998046875,
0.82275390625,
0.81298828125,
0.1748046875,
0.81298828125
],
[
0.1895751953125,
0.81298828125,
0.82275390625,
0.81298828125,
0.82275390625,
0.82763671875,
0.1895751953125,
0.82763671875
],
[
0.19287109375,
0.82763671875,
0.82275390625,
0.82763671875,
0.82275390625,
0.84130859375,
0.19287109375,
0.84130859375
],
[
0.19287109375,
0.84130859375,
0.82275390625,
0.84130859375,
0.82275390625,
0.85498046875,
0.19287109375,
0.85498046875
],
[
0.1895751953125,
0.85400390625,
0.82275390625,
0.85498046875,
0.82275390625,
0.869140625,
0.1895751953125,
0.86865234375
],
[
0.191162109375,
0.869140625,
0.82421875,
0.869140625,
0.82421875,
0.88330078125,
0.191162109375,
0.88330078125
],
[
0.1895751953125,
0.88330078125,
0.82177734375,
0.88330078125,
0.82177734375,
0.896484375,
0.1895751953125,
0.896484375
],
[
0.1895751953125,
0.89794921875,
0.82177734375,
0.89794921875,
0.82177734375,
0.9111328125,
0.1895751953125,
0.9111328125
],
[
0.191162109375,
0.91015625,
0.2296142578125,
0.912109375,
0.2279052734375,
0.923828125,
0.1895751953125,
0.921875
],
[
0.491943359375,
0.94775390625,
0.50732421875,
0.94775390625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Published at ICLR 2023 Workshop on Domain Generalization",
"3.2Evaluation Metrics and Results",
"由于数据集不平衡，我们的目标是在看不见的测试域（OOD）上获得较",
"高的F1宏观得分，这使得每个类的权重相等，而不管其基数如何。除",
"此之外，我们还报告了F1-micro分数，它给出了每个班级的加权平均值",
"我们在表1和表2中报告",
"了通过对源域（Acevedo_20和",
"ERMmode",
"Ours",
"Matek_19）验证集和看不见的",
"测试域（INT_20）进行五重交",
"叉验证确定的不平衡DG分类",
"结果。。我们将我们的模型与",
"个像ERME（Vapnik，1999）这",
"样的香草训练模型E(1）以及两",
"种当前的SOTADG方法进行",
"了比较，例如 DANNE(Ganin",
"",
"et al.,2016)和 CORALE(Sun",
"&Saenko2016）：我们训练模",
"型的两种不同变体，一",
"一种是对编",
"(a) Source (Matek_19, Acevedo_20) domain val-set",
"码器部分（表示形式z）和MLP",
"ERMmode",
"Ours",
"分类器部分进行耦合训练，另一",
"",
"",
"",
"",
"种（表示为Ours+）是对编码",
"器部分和编码器部分进行解耦",
"训练。分类器部分使用类平衡采",
"样。我们在INT_20测试集上实",
"现了最佳的F1-marco得分，而",
"DG方法DANN由于在训练过",
"",
"",
"..",
"程中假设了数据平衡而比香草",
"",
"",
"ERM模型表现更差。InFig。3",
"我们还将模型的混淆矩阵与",
"ERM方法进行了比较。我们的",
"(b) Target (INT_20) domain testset (unseen)",
"模型通过识别少数类别的样本",
"",
"Figure 3:Confusion matrices show an improved clas-",
"表现更好，而ERM错过了少数",
"类别的样本，例如非典型淋巴细胞yte_atypicalclasswithourmethodcomparedtothestan",
"sification of cells from the lowly populated lympho-",
"dard ERM model.",
"4.",
"Conclusion",
"我们开发了一个强大的CNN模型，用于血液细胞形态学中的非分布泛化，解决了三个主要",
"挑战：数据不平衡，域转移和缺失类。我们展示了如何通过优化潜在特征空间中的损失函数",
"和网络的输出日志，针对不同的领域改进现有的预训练深度模型。我们的工作展示了如何",
"解决血液学单一WBC分类中的生物学，流行病学和技术变异性，以训练强大的基于AI的",
"细胞分类器，为其在临床环境中的安全和生产性使用铺平道路。",
"References",
"Andrea Acevedo, Anna Merino, Santiago Alferez, Angel Molina, Laura Boldi, and Jose",
"Rodellar. A dataset of microscopic peripheral blood cell images for development of auto-",
"matic recognition systems. Data in brief, 30, 2020.",
"Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla",
"Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini",
"Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya",
"Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric",
"Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner,",
"Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are",
"few-shot learners. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin",
"(eds.), Advances in Neural Information Processing Systems, volume 33, pp. 1877-1901",
"2020.",
"5"
] |
{
"arxiv_id": "2303.07771",
"language": "zh",
"page_idx": 4
}
|
||
1911.10519-zh-2
|
1911.10519-zh-2.jpg
|
<table>
<tbody>
<tr>
<th>Map</th>
<th>Start Boundary</th>
<th>EndI Boundary</th>
</tr>
<tr>
<td>Map 1</td>
<td>(0, -5, 0)</td>
<td>(10, 20, 6)</td>
</tr>
<tr>
<td>Map 2</td>
<td>0. 11 0</td>
<td>20, 5, 6 1 </td>
</tr>
<tr>
<td>Map 3</td>
<td>(0,-5,0)</td>
<td>10, 20, 6) 1</td>
</tr>
<tr>
<td>Map 4</td>
<td>(0, 0 1 11</td>
<td>26, 20, 10)</td>
</tr>
</tbody>
</table>
3.3
Cost Function
由于生成的路线的随机性，存在无人机可能与障碍物碰撞并且无法继续运动的可能性，因此无法到达目标位置。因此，需要在路线规划问题中引人各种成本，例如燃料成本，急转弯成本和不完整路线成本。所涉及的各种成本可以用[13]来说明：
$\begin{array}{l}\bullet C_{fuel}:\mathrm{The~cost~of~fuel~which~depends~on~length~of~the~route}\\\mathrm{followed.~This~cost~is~less~when~the~route~followed~has~a}\end{array}$ smaller length, which in turn leads to less consumption of fuel and lower time to reach the goal.
$\bullet C_{diverqence}:$This cost is governed by frequency of sharp turns present in route. Lower the sharp turns the smoother and stable the route is.
$\bullet C_{gap}:$This cost corresponds to the separation among the goal and the route's end when the UAVs fail to reach the goal location. This situation could arise due to a significant obstacle which lies in between the route to goal. This cost corresponds to the highest interest amongst all the costs and assigned with the highest priority. This cost should have a zero value in the optimal solution.
到目标位置的生成路线$R_{(x,y,z)}$可以表示为从源位置开始的
一系列点，如下所示：
(6)
$$R_{x,y,z}=\{S,a_1,a_2,a_3,a_4,...a_n,G\}$$
其中，$S$表示起始位置；$G$表示目标位置；$a_1,a_2,...a_n$表示
路线中出现的点
为了确定燃料的成本，假定无人机的速度在路线规划期间是恒定的。因此，燃料的成本可以通过使用无人机行驶的总距离来获得。燃料成本用公式 7 表示
$$C_{fuel}=\sum(D_{x,y,z})$$
(7)
当从$a_j-1$到$a_j$的方向变化与从$a_j$到$a_j+1$的方向变化不匹配时，$^{\prime}$由于急转弯而产生的成本。它可以通过搜索所获得的路线中的情况来计算，在两者之间的角度有相当大的转弯量。此成本将随着计划路线中急转弯的增加而增加。此成本可以从公式 8 中推断出
$$C_{divergence}=\text{number of turning points in route}$$
(8)
现在最重要的成本，我。e。$,C_{gap}$是使用欧几里得距离计算的。所需的两个点是路线的终点和目标位置。如果端点和目标位置与此成本相同，则将其设置为0，否则将借助方程式 9 确定
$$c_{gap}=\sqrt{(X_{end}-T_x)^2+(Y_{end}-T_y)^2+(Z_{end}-T_z)^2}$$
$\begin{array}{l}(X_{end},Y_{end},Z_{end})\\\end{array}$表示生成的路由的端点；$(T_x,T_y,T_z)$ 表
示目标位置
可以使用公式 10 计算路线的总成本
$$C_{total}=P_1\times C_{fuel}+P_2\times C_{divergence}+P_3\times C_{gap}$$
总成本函数有助于从路线集中实现最佳路线
这里$P_1,P_2,P_3$是实验参数 , 其值是在问题描述的实验和约束的帮助下启动的。在将更大的路由长度作为首选项的情况下， 将为$P_1$分配更高的值。类似地$,P_2$向路线平滑度分配倾斜度。从$P_3$ ,它是确保计划的路线达到目标。无人机倾向于仅包含有关其位置的信息，并且对于随后的增量，它们可能会改变其在附近小区之一中的位置。因此，无人机没有任何关于它将在什么时间和地点面对障碍物的先验知识；这就是为什么无人机在遇到障碍物时稳定地继续修改其路径的原因
3.4 Trajectory generation for multiple UAVs
由 SSA 生成的轨迹可能包括一些急转弯，这些急转弯实际上对于 UAV 来说是不可行的。因此，平滑生成的路线变得至关重要。无人机的移动和速度可以使用车辆移动性的多项式函数来表示。多项式函数涉及时间的分量。由五次多项式生成连续路线，这表明一阶导数是一致的。可以有效地确定多项式导数， 以这种方式获得活动的结果。用方程 11 表示传统的五阶时间多项式函数[24]
$$S(t)=a_5t^5+a_4t^4+a_3t^3+a_2t^2+a_1t^1+a_0$$
(11)
在生成轨迹时，在初始阶段获得子点。Sub-pointcanbeconsideredasaned
为了产生平滑的路线，通过基于五阶多项式的轨迹拟合策略
[24]利用无人机通过子点的速度和运动
在 3D 环境中使用 SSA 进行多无人机航路规划的主要目的
是：
$\bullet$to localize the UAVs and destination location.
$\bullet$to plan a route between multiple UAVs and their correspond- $\begin{array}{c}\text{ing targets and simultaneously ensure that no UAV collides}\\\end{array}$ with each other or any obstacle.
### 4 Salp Swarm Algorithm
Salp 是一种具有透明桶形身体的海洋生物。它是 Salpidae 家族的一部分。它的运动是通过身体抽水来控制的，并作为前进的动力。SSA 的灵感来自于在水中觅食和导航时 salps 的蜂拥而至的行为。$\operatorname{salp}$的形状如图 4(a)所示。深海中的骤降形成的群体被称为骤降链。此链如图 4(b)所示。该算法首先由Mirjalili 等人开发。al。2017年[19]。选择 SSA 进行路线规划的主要原因是其简单性，并且该算法的灵感来自海洋中寻找食物的自然导航和觅食行为。关于多 UAV 路线规划，可以将$\operatorname{salp}$链的觅食阶段视为对环境中不同目标的搜索，并且可以将导航阶段视为连接环境中的点以获得所生成路径长度的航路点为了理解 Salp 链，开发了一个数学模型。人口分为 2 个部
分-领导者和追随者。领队萨尔普出现在链条的最前面，而其余的萨尔普则被认为是追随者。前面的齐射护送着蜂群，而其余的齐射则彼此跟随并领导(直接或间接)
与其他基于群的技术类似，salps 的位置在$n$维搜索空间中定义，其中$n$表示给定问题的变量数量。因此，所有 salps 的位置都保存在名为$x$的二维矩阵中。假设在搜索空间中存在一$\overline{\text{个名为 }F\text{ 的食物源作为蜂群的目标}}$
可以从以下方程式中观察到 SSA 的完整工作
公式 12 更新引线的位置：
$$x_j^1=\left\{\begin{array}{cc}F_j+c_1((ub_j-lb_j)c_2+lb_j)&c_3\ge0\\F_j-c_1((ub_j-lb_j)c_2+lb_j)&c_3<0\end{array}\right.$$
(12)
其中 $x_j^1=$领队的位置；$F_j=$食物来源的位置；$ub_j$ 和 $lb_j$
是$j^{th}$维度的上下界$;c_1,c_2$ 、$c_3$是均匀随机数
在公式 13 中，$c_1$在探索$\&$开发阶段之间保持平衡：
$$c_1=2e^{-(\frac{4l}{L})^2}$$
(13)
其中，$l:$当前迭代$,L:$最大迭代次数
探索被定义为算法试图探索搜索空间的阶段。避免本地解决
方案发生在此阶段。勘探之后是开发阶段，在该阶段中，主要
关注的是即兴创作在勘探阶段中探索的解决方案
|
[
[
0.07135009765625,
0.06353759765625,
0.298828125,
0.06414794921875,
0.298828125,
0.07781982421875,
0.07135009765625,
0.07659912109375
],
[
0.5341796875,
0.06414794921875,
0.9267578125,
0.0653076171875,
0.9267578125,
0.07659912109375,
0.5341796875,
0.07598876953125
],
[
0.1351318359375,
0.08197021484375,
0.173828125,
0.0843505859375,
0.172119140625,
0.0968017578125,
0.133544921875,
0.09442138671875
],
[
0.51904296875,
0.07781982421875,
0.9267578125,
0.07781982421875,
0.9267578125,
0.0885009765625,
0.51904296875,
0.0885009765625
],
[
0.19140625,
0.08489990234375,
0.3046875,
0.08489990234375,
0.3046875,
0.0955810546875,
0.19140625,
0.0955810546875
],
[
0.3046875,
0.08489990234375,
0.41650390625,
0.08489990234375,
0.41650390625,
0.0955810546875,
0.3046875,
0.0955810546875
],
[
0.51904296875,
0.08905029296875,
0.9267578125,
0.08905029296875,
0.9267578125,
0.09979248046875,
0.51904296875,
0.09979248046875
],
[
0.51708984375,
0.09979248046875,
0.927734375,
0.09979248046875,
0.927734375,
0.1134033203125,
0.51708984375,
0.1134033203125
],
[
0.1351318359375,
0.1092529296875,
0.1822509765625,
0.11102294921875,
0.181396484375,
0.12353515625,
0.133544921875,
0.1217041015625
],
[
0.193115234375,
0.11102294921875,
0.255126953125,
0.11102294921875,
0.255126953125,
0.1217041015625,
0.193115234375,
0.1217041015625
],
[
0.316650390625,
0.11102294921875,
0.38623046875,
0.11102294921875,
0.38623046875,
0.1217041015625,
0.316650390625,
0.1217041015625
],
[
0.51708984375,
0.11102294921875,
0.927734375,
0.11102294921875,
0.927734375,
0.12469482421875,
0.51708984375,
0.12469482421875
],
[
0.1351318359375,
0.122314453125,
0.183837890625,
0.122314453125,
0.183837890625,
0.1341552734375,
0.1351318359375,
0.1341552734375
],
[
0.189697265625,
0.122314453125,
0.2493896484375,
0.122314453125,
0.2493896484375,
0.1341552734375,
0.189697265625,
0.1341552734375
],
[
0.316650390625,
0.122314453125,
0.38037109375,
0.122314453125,
0.38037109375,
0.1341552734375,
0.316650390625,
0.1341552734375
],
[
0.51904296875,
0.12469482421875,
0.61376953125,
0.12469482421875,
0.61376953125,
0.1353759765625,
0.51904296875,
0.1353759765625
],
[
0.60693359375,
0.12353515625,
0.927734375,
0.12353515625,
0.927734375,
0.13720703125,
0.60693359375,
0.13720703125
],
[
0.1351318359375,
0.1318359375,
0.183837890625,
0.1341552734375,
0.1822509765625,
0.1484375,
0.133544921875,
0.146728515625
],
[
0.189697265625,
0.1341552734375,
0.25439453125,
0.1341552734375,
0.25439453125,
0.1455078125,
0.189697265625,
0.1455078125
],
[
0.318115234375,
0.1341552734375,
0.38623046875,
0.1341552734375,
0.38623046875,
0.1455078125,
0.318115234375,
0.1455078125
],
[
0.51708984375,
0.1353759765625,
0.9296875,
0.1353759765625,
0.9296875,
0.1484375,
0.51708984375,
0.1484375
],
[
0.1351318359375,
0.146728515625,
0.185546875,
0.146728515625,
0.185546875,
0.157958984375,
0.1351318359375,
0.157958984375
],
[
0.193115234375,
0.146728515625,
0.2493896484375,
0.146728515625,
0.2493896484375,
0.157958984375,
0.193115234375,
0.157958984375
],
[
0.316650390625,
0.146728515625,
0.393798828125,
0.146728515625,
0.393798828125,
0.157958984375,
0.316650390625,
0.157958984375
],
[
0.51904296875,
0.1484375,
0.8212890625,
0.1484375,
0.8212890625,
0.1591796875,
0.51904296875,
0.1591796875
],
[
0.5205078125,
0.177001953125,
0.5439453125,
0.177001953125,
0.5439453125,
0.185302734375,
0.5205078125,
0.185302734375
],
[
0.5615234375,
0.174560546875,
0.8251953125,
0.174560546875,
0.8251953125,
0.188232421875,
0.5615234375,
0.188232421875
],
[
0.0679931640625,
0.188232421875,
0.09906005859375,
0.1859130859375,
0.10076904296875,
0.19775390625,
0.0697021484375,
0.1995849609375
],
[
0.1175537109375,
0.189453125,
0.2091064453125,
0.189453125,
0.2091064453125,
0.1995849609375,
0.1175537109375,
0.1995849609375
],
[
0.51708984375,
0.196533203125,
0.927734375,
0.196533203125,
0.927734375,
0.210205078125,
0.51708984375,
0.210205078125
],
[
0.07135009765625,
0.21142578125,
0.480224609375,
0.21142578125,
0.480224609375,
0.221435546875,
0.07135009765625,
0.221435546875
],
[
0.51708984375,
0.2078857421875,
0.927734375,
0.2078857421875,
0.927734375,
0.221435546875,
0.51708984375,
0.221435546875
],
[
0.07135009765625,
0.221435546875,
0.48193359375,
0.221435546875,
0.48193359375,
0.235107421875,
0.07135009765625,
0.235107421875
],
[
0.51708984375,
0.2197265625,
0.9296875,
0.2197265625,
0.9296875,
0.2327880859375,
0.51708984375,
0.2327880859375
],
[
0.07135009765625,
0.2340087890625,
0.48193359375,
0.2340087890625,
0.48193359375,
0.2476806640625,
0.07135009765625,
0.2476806640625
],
[
0.51708984375,
0.232177734375,
0.927734375,
0.232177734375,
0.927734375,
0.2452392578125,
0.51708984375,
0.2452392578125
],
[
0.07135009765625,
0.2464599609375,
0.45751953125,
0.2464599609375,
0.45751953125,
0.257080078125,
0.07135009765625,
0.257080078125
],
[
0.51708984375,
0.243408203125,
0.92529296875,
0.243408203125,
0.92529296875,
0.257080078125,
0.51708984375,
0.257080078125
],
[
0.51708984375,
0.255859375,
0.927734375,
0.255859375,
0.927734375,
0.26953125,
0.51708984375,
0.26953125
],
[
0.0621337890625,
0.270751953125,
0.483642578125,
0.270751953125,
0.483642578125,
0.283935546875,
0.0621337890625,
0.283935546875
],
[
0.51708984375,
0.267333984375,
0.623046875,
0.268310546875,
0.623046875,
0.27978515625,
0.51708984375,
0.279052734375
],
[
0.06463623046875,
0.27978515625,
0.48193359375,
0.28076171875,
0.48193359375,
0.29443359375,
0.06463623046875,
0.29345703125
],
[
0.0697021484375,
0.292236328125,
0.48193359375,
0.292236328125,
0.48193359375,
0.305908203125,
0.0697021484375,
0.305908203125
],
[
0.06634521484375,
0.303955078125,
0.32568359375,
0.3046875,
0.32568359375,
0.318359375,
0.06634521484375,
0.317626953125
],
[
0.5556640625,
0.302734375,
0.86376953125,
0.303955078125,
0.86376953125,
0.318359375,
0.5556640625,
0.317626953125
],
[
0.8984375,
0.305908203125,
0.927734375,
0.305908203125,
0.927734375,
0.318359375,
0.8984375,
0.318359375
],
[
0.06378173828125,
0.318359375,
0.1435546875,
0.322509765625,
0.1427001953125,
0.334228515625,
0.0621337890625,
0.330078125
],
[
0.14111328125,
0.3154296875,
0.48193359375,
0.31640625,
0.48193359375,
0.330810546875,
0.14111328125,
0.330078125
],
[
0.0697021484375,
0.330078125,
0.480224609375,
0.330078125,
0.480224609375,
0.34033203125,
0.0697021484375,
0.34033203125
],
[
0.53564453125,
0.33203125,
0.99658203125,
0.33203125,
0.99658203125,
0.342529296875,
0.53564453125,
0.342529296875
],
[
0.07135009765625,
0.342529296875,
0.2301025390625,
0.342529296875,
0.2301025390625,
0.352783203125,
0.07135009765625,
0.352783203125
],
[
0.51904296875,
0.342529296875,
0.9296875,
0.342529296875,
0.9296875,
0.355712890625,
0.51904296875,
0.355712890625
],
[
0.060455322265625,
0.352783203125,
0.483642578125,
0.350830078125,
0.483642578125,
0.365234375,
0.060455322265625,
0.36767578125
],
[
0.51904296875,
0.352783203125,
0.7734375,
0.352783203125,
0.7734375,
0.366455078125,
0.51904296875,
0.366455078125
],
[
0.0697021484375,
0.364013671875,
0.48193359375,
0.364013671875,
0.48193359375,
0.377685546875,
0.0697021484375,
0.377685546875
],
[
0.5341796875,
0.366455078125,
0.927734375,
0.366455078125,
0.927734375,
0.380126953125,
0.5341796875,
0.380126953125
],
[
0.0679931640625,
0.3759765625,
0.48193359375,
0.3759765625,
0.48193359375,
0.3896484375,
0.0679931640625,
0.3896484375
],
[
0.51708984375,
0.37890625,
0.5439453125,
0.37890625,
0.5439453125,
0.39013671875,
0.51708984375,
0.39013671875
],
[
0.0697021484375,
0.388427734375,
0.483642578125,
0.388427734375,
0.483642578125,
0.402099609375,
0.0697021484375,
0.402099609375
],
[
0.0697021484375,
0.399658203125,
0.48193359375,
0.399658203125,
0.48193359375,
0.413330078125,
0.0697021484375,
0.413330078125
],
[
0.51123046875,
0.40380859375,
0.82958984375,
0.40380859375,
0.82958984375,
0.41455078125,
0.51123046875,
0.41455078125
],
[
0.07135009765625,
0.413330078125,
0.48193359375,
0.413330078125,
0.48193359375,
0.42333984375,
0.07135009765625,
0.42333984375
],
[
0.5087890625,
0.41455078125,
0.927734375,
0.41455078125,
0.927734375,
0.427490234375,
0.5087890625,
0.427490234375
],
[
0.0697021484375,
0.424560546875,
0.300537109375,
0.424560546875,
0.300537109375,
0.435302734375,
0.0697021484375,
0.435302734375
],
[
0.5146484375,
0.427001953125,
0.9296875,
0.42578125,
0.9296875,
0.439453125,
0.5146484375,
0.440673828125
],
[
0.51708984375,
0.438232421875,
0.74072265625,
0.438232421875,
0.74072265625,
0.451904296875,
0.51708984375,
0.451904296875
],
[
0.08477783203125,
0.446533203125,
0.483642578125,
0.446533203125,
0.483642578125,
0.46435546875,
0.08477783203125,
0.46435546875
],
[
0.08477783203125,
0.4619140625,
0.210693359375,
0.4619140625,
0.210693359375,
0.47265625,
0.08477783203125,
0.47265625
],
[
0.51708984375,
0.47265625,
0.5322265625,
0.47265625,
0.5322265625,
0.482177734375,
0.51708984375,
0.482177734375
],
[
0.5556640625,
0.471435546875,
0.73291015625,
0.471435546875,
0.73291015625,
0.485107421875,
0.5556640625,
0.485107421875
],
[
0.156982421875,
0.488037109375,
0.393798828125,
0.488037109375,
0.393798828125,
0.501953125,
0.156982421875,
0.501953125
],
[
0.45751953125,
0.488037109375,
0.48193359375,
0.488037109375,
0.48193359375,
0.501953125,
0.45751953125,
0.501953125
],
[
0.515625,
0.49462890625,
0.9296875,
0.495361328125,
0.9296875,
0.5087890625,
0.515625,
0.50830078125
],
[
0.08563232421875,
0.5048828125,
0.48193359375,
0.505859375,
0.48193359375,
0.51953125,
0.08563232421875,
0.5185546875
],
[
0.51904296875,
0.5087890625,
0.9267578125,
0.5087890625,
0.9267578125,
0.51953125,
0.51904296875,
0.51953125
],
[
0.07135009765625,
0.5185546875,
0.17626953125,
0.5185546875,
0.17626953125,
0.52978515625,
0.07135009765625,
0.52978515625
],
[
0.51708984375,
0.51953125,
0.927734375,
0.51953125,
0.927734375,
0.533203125,
0.51708984375,
0.533203125
],
[
0.08563232421875,
0.529296875,
0.480224609375,
0.529296875,
0.480224609375,
0.5419921875,
0.08563232421875,
0.5419921875
],
[
0.515625,
0.53076171875,
0.9296875,
0.53076171875,
0.9296875,
0.54443359375,
0.515625,
0.54443359375
],
[
0.0697021484375,
0.5419921875,
0.483642578125,
0.5419921875,
0.483642578125,
0.5556640625,
0.0697021484375,
0.5556640625
],
[
0.51708984375,
0.54345703125,
0.9296875,
0.54345703125,
0.9296875,
0.55712890625,
0.51708984375,
0.55712890625
],
[
0.0697021484375,
0.55419921875,
0.2880859375,
0.55419921875,
0.2880859375,
0.56787109375,
0.0697021484375,
0.56787109375
],
[
0.51708984375,
0.5546875,
0.9296875,
0.5546875,
0.9296875,
0.568359375,
0.51708984375,
0.568359375
],
[
0.5205078125,
0.568359375,
0.927734375,
0.568359375,
0.927734375,
0.5791015625,
0.5205078125,
0.5791015625
],
[
0.20654296875,
0.58056640625,
0.344970703125,
0.58056640625,
0.344970703125,
0.595703125,
0.20654296875,
0.595703125
],
[
0.4609375,
0.5830078125,
0.483642578125,
0.5830078125,
0.483642578125,
0.59326171875,
0.4609375,
0.59326171875
],
[
0.51904296875,
0.5791015625,
0.9296875,
0.5791015625,
0.9296875,
0.5927734375,
0.51904296875,
0.5927734375
],
[
0.51904296875,
0.59130859375,
0.9267578125,
0.59130859375,
0.9267578125,
0.6015625,
0.51904296875,
0.6015625
],
[
0.08563232421875,
0.60107421875,
0.48193359375,
0.60107421875,
0.48193359375,
0.6142578125,
0.08563232421875,
0.6142578125
],
[
0.51708984375,
0.6015625,
0.9267578125,
0.6015625,
0.9267578125,
0.615234375,
0.51708984375,
0.615234375
],
[
0.07135009765625,
0.61328125,
0.483642578125,
0.61328125,
0.483642578125,
0.626953125,
0.07135009765625,
0.626953125
],
[
0.5341796875,
0.61328125,
0.9296875,
0.61328125,
0.9296875,
0.626953125,
0.5341796875,
0.626953125
],
[
0.07135009765625,
0.6259765625,
0.48193359375,
0.6259765625,
0.48193359375,
0.6396484375,
0.07135009765625,
0.6396484375
],
[
0.51708984375,
0.62451171875,
0.927734375,
0.62451171875,
0.927734375,
0.63818359375,
0.51708984375,
0.63818359375
],
[
0.0697021484375,
0.63818359375,
0.480224609375,
0.63818359375,
0.480224609375,
0.65185546875,
0.0697021484375,
0.65185546875
],
[
0.51708984375,
0.63720703125,
0.9296875,
0.63720703125,
0.9296875,
0.65087890625,
0.51708984375,
0.65087890625
],
[
0.07135009765625,
0.6494140625,
0.177978515625,
0.6494140625,
0.177978515625,
0.6611328125,
0.07135009765625,
0.6611328125
],
[
0.51708984375,
0.6494140625,
0.78076171875,
0.6494140625,
0.78076171875,
0.6630859375,
0.51708984375,
0.6630859375
],
[
0.53564453125,
0.66015625,
0.9267578125,
0.66015625,
0.9267578125,
0.67333984375,
0.53564453125,
0.67333984375
],
[
0.51708984375,
0.67138671875,
0.9296875,
0.6728515625,
0.9296875,
0.68701171875,
0.51708984375,
0.68603515625
],
[
0.10833740234375,
0.68603515625,
0.444091796875,
0.68310546875,
0.444091796875,
0.70068359375,
0.10833740234375,
0.70361328125
],
[
0.45751953125,
0.6904296875,
0.478515625,
0.6904296875,
0.478515625,
0.69775390625,
0.45751953125,
0.69775390625
],
[
0.51708984375,
0.68505859375,
0.923828125,
0.68505859375,
0.923828125,
0.69873046875,
0.51708984375,
0.69873046875
],
[
0.450927734375,
0.6923828125,
0.45751953125,
0.6923828125,
0.45751953125,
0.6953125,
0.450927734375,
0.6953125
],
[
0.51904296875,
0.69677734375,
0.75390625,
0.69677734375,
0.75390625,
0.71044921875,
0.51904296875,
0.71044921875
],
[
0.08734130859375,
0.71337890625,
0.48193359375,
0.71337890625,
0.48193359375,
0.72705078125,
0.08734130859375,
0.72705078125
],
[
0.53662109375,
0.71044921875,
0.82958984375,
0.71044921875,
0.82958984375,
0.72021484375,
0.53662109375,
0.72021484375
],
[
0.5341796875,
0.7197265625,
0.7001953125,
0.7197265625,
0.7001953125,
0.73291015625,
0.5341796875,
0.73291015625
],
[
0.07135009765625,
0.72705078125,
0.48193359375,
0.72705078125,
0.48193359375,
0.73681640625,
0.07135009765625,
0.73681640625
],
[
0.0697021484375,
0.73681640625,
0.48193359375,
0.73681640625,
0.48193359375,
0.75048828125,
0.0697021484375,
0.75048828125
],
[
0.07135009765625,
0.751953125,
0.08734130859375,
0.751953125,
0.08734130859375,
0.76123046875,
0.07135009765625,
0.76123046875
],
[
0.59619140625,
0.75732421875,
0.61376953125,
0.75732421875,
0.61376953125,
0.765625,
0.59619140625,
0.765625
],
[
0.60693359375,
0.75439453125,
0.85791015625,
0.7548828125,
0.85791015625,
0.76953125,
0.60693359375,
0.7685546875
],
[
0.55517578125,
0.76025390625,
0.59033203125,
0.76025390625,
0.59033203125,
0.775390625,
0.55517578125,
0.775390625
],
[
0.59375,
0.7666015625,
0.85791015625,
0.7666015625,
0.85791015625,
0.78076171875,
0.59375,
0.78076171875
],
[
0.8984375,
0.7626953125,
0.927734375,
0.7626953125,
0.927734375,
0.77490234375,
0.8984375,
0.77490234375
],
[
0.0814208984375,
0.79150390625,
0.1275634765625,
0.79150390625,
0.1275634765625,
0.8017578125,
0.0814208984375,
0.8017578125
],
[
0.1427001953125,
0.7880859375,
0.48193359375,
0.7880859375,
0.48193359375,
0.8017578125,
0.1427001953125,
0.8017578125
],
[
0.5341796875,
0.79443359375,
0.927734375,
0.7958984375,
0.927734375,
0.80859375,
0.5341796875,
0.80810546875
],
[
0.51708984375,
0.80859375,
0.8505859375,
0.80859375,
0.8505859375,
0.822265625,
0.51708984375,
0.822265625
],
[
0.08477783203125,
0.8154296875,
0.483642578125,
0.81640625,
0.483642578125,
0.83056640625,
0.08477783203125,
0.830078125
],
[
0.5322265625,
0.82177734375,
0.87744140625,
0.82177734375,
0.87744140625,
0.8349609375,
0.5322265625,
0.8349609375
],
[
0.0679931640625,
0.8291015625,
0.14697265625,
0.82666015625,
0.14697265625,
0.84130859375,
0.0679931640625,
0.84326171875
],
[
0.08734130859375,
0.84033203125,
0.32568359375,
0.84033203125,
0.32568359375,
0.85400390625,
0.08734130859375,
0.85400390625
],
[
0.67333984375,
0.8515625,
0.767578125,
0.84326171875,
0.77099609375,
0.85693359375,
0.6767578125,
0.865234375
],
[
0.8984375,
0.85107421875,
0.9296875,
0.85107421875,
0.9296875,
0.8623046875,
0.8984375,
0.8623046875
],
[
0.53564453125,
0.86767578125,
0.8037109375,
0.86767578125,
0.8037109375,
0.88134765625,
0.53564453125,
0.88134765625
],
[
0.0831298828125,
0.87890625,
0.48193359375,
0.87890625,
0.48193359375,
0.8935546875,
0.0831298828125,
0.8935546875
],
[
0.5341796875,
0.8798828125,
0.9296875,
0.8798828125,
0.9296875,
0.8935546875,
0.5341796875,
0.8935546875
],
[
0.51904296875,
0.8935546875,
0.9267578125,
0.8935546875,
0.9267578125,
0.90380859375,
0.51904296875,
0.90380859375
],
[
0.08734130859375,
0.90380859375,
0.37451171875,
0.90380859375,
0.37451171875,
0.91748046875,
0.08734130859375,
0.91748046875
],
[
0.51904296875,
0.90625,
0.83642578125,
0.90625,
0.83642578125,
0.91650390625,
0.51904296875,
0.91650390625
]
] |
[
"Table1 3DMap Boundary Representation",
"这里P1.P2,P3是实验参数，其值是在问题描述的实验和约",
"Map",
"束的帮助下启动的。在将更大的路由长度作为首选项的情况下",
"Start Boundary",
"End Boundary",
"将为P1分配更高的值。类似地，P2向路线平滑度分配倾斜度。",
"从P3，它是确保计划的路线达到目标。无人机倾向于仅包含",
"Map1",
"(0, -5, 0)",
"(10, 20, 6)",
"有关其位置的信息，并且对于随后的增量，它们可能会改变其",
"Map 2",
"(0, 0, 0)",
"(20, 5, 6)",
"在附近小区之",
"一中的位置。因此，无人机没有任何关于它将在",
"Map3",
"(0, -5, 0)",
"(10,20,6)",
"什么时间和地点面对障碍物的先验知识；这就是为什么无人机",
"Map 4",
"(0, 0, 0)",
"(26,20,10)",
"在遇到障碍物时稳定地继续修改其路径的原因",
"3.4",
"Trajectory generation for multiple UAVs",
"3.3",
"Cost Function",
"由SSA生成的轨迹可能包括一些急转弯，这些急转弯实际上",
"由于生成的路线的随机性，存在无人机可能与障碍物碰撞并且",
"对于UAV来说是不可行的。因此，平滑生成的路线变得至关",
"无法继续运动的可能性，因此无法到达目标位置。因此，需要",
"重要。无人机的移动和速度可以使用车辆移动性的多项式函数",
"在路线规划问题中引人各种成本，例如燃料成本，急转弯成本",
"来表示。多项式函数涉及时间的分量。由五次多项式生成连续",
"和不完整路线成本。所涉及的各种成本可以用[13]来说明：",
"路线，这表明一阶导数是一致的。可以有效地确定多项式导数",
"以这种方式获得活动的结果。用方程11表示传统的五阶时间",
".Cfuel : The cost of fuel which depends on length of the route",
"多项式函数[24]",
"followed. This cost is less when the route followed has a",
"smaller length, which in turn leads to less consumption of",
"fuel and lower time to reach the goal.",
"S(t)=ast5+a4t4+a3t3+a2t2+aitl+ao",
"(11)",
".Cdivergenc",
"e : This cost is governed by frequency of sharp",
"turns present in route.Lower the sharp turns the smoother",
"在生成轨迹时，在初始阶段获得子点。Sub-pointcanbeconsideredasanec",
"and stable the route is.",
"为了产生平滑的路线，通过基于五阶多项式的轨迹拟合策略",
".Cgap : This cost corresponds to the separation among the",
"[24]利用无人机通过子点的速度和运动",
"goal and the route's end when the UAVs fail to reach the",
"在3D环境中使用 SSA进行多无人机航路规划的主要目的",
"goal location. This situation could arise due to a significant",
"是：",
"obstacle which lies in between the route to goal. This cost",
"corresponds to the highest interest amongst all the costs and",
"oto localize the UAVs and destination location.",
"assigned with the highest priority.This cost should have a",
"to plan a route between multiple UAVs and their correspond-",
"zero value in the optimal solution",
"ing targets and simultaneously ensure that no UAV collides",
"with each other or any obstacle.",
"到目标位置的生成路线 R(#,9;2）可以表示为从源位置开始的",
"系列点，如下所示：",
"4.",
"Salp Swarm Algorithm",
"Ra,y,z = {S,al,a2,a3,a4,.an,G]",
"(6)",
"Salp是一种具有透明桶形身体的海洋生物。它是Salpidae家",
"其中，S表示起始位置；G表示目标位置；a1,a2...an表示",
"族的一部分。它的运动是通过身体抽水来控制的，并作为前进",
"路线中出现的点",
"的动力。SSA 的灵感来自于在水中觅食和导航时 salps 的蜂",
"为了确定燃料的成本，假定无人机的速度在路线规划期间是",
"拥而至的行为。salp的形状如图4(a）所示。深海中的骤降形",
"恒定的。因此，燃料的成本可以通过使用无人机行驶的总距离",
"成的群体被称为骤降链。此链如图4（b）所示。该算法首先由",
"来获得。燃料成本用公式7表示",
"Mirjalili等人开发。al。2017年[19]。选择SSA进行路线规",
"划的主要原因是其简单性，并且该算法的灵感来自海洋中寻找",
"Cfuel = (Dr,y,z)",
"(7)",
"食物的自然导航和觅食行为。关于多UAV路线规划，可以将",
"salp链的觅食阶段视为对环境中不同目标的搜索，并且可以将",
"当从ai-1到ai的方向变化与从ai到ai1的方向变化不",
"导航阶段视为连接环境中的点以获得所生成路径长度的航路点",
"匹配时，由于急转弯而产生的成本。它可以通过搜索所获得的",
"为了理解Salp链，开发了二个数学模型。人口分为2个部",
"路线中的情况来计算，在两者之间的角度有相当大的转弯量。",
"分-领导者和追随者。领队萨尔普出现在链条的最前面，而其余",
"此成本将随着计划路线中急转弯的增加而增加。此成本可以从",
"的萨尔普则被认为是追随者。前面的齐射护送着蜂群，而其余",
"公式8中推断出",
"的齐射则彼此跟随并领导（直接或间接）",
"与其他基于群的技术类似，salps的位置在n维搜索空间中",
"定义，其中n表示给定问题的变量数量。因此，所有 salps的",
"Cdivergence = number of turning points in route",
"(8",
"位置都保存在名为的二维矩阵中。假设在搜索空间中存在一",
"",
"个名为F的食物源作为蜂群的目标",
"现在最重要的成本，我。e。，gap是使用欧几里得距离计算",
"可以从以下方程式中观察到SSA的完整工作",
"公式12更新引线的位置",
"的。所需的两个点是路线的终点和目标位置。如果端点和目标",
"位置与此成本相同，则将其设置为0，否则将借助方程式9确",
"定",
"",
"Fj+ci((ubj-lb,)c2+lb,) c3≥0",
"cj=",
" F, - ci(ubj - lb,)c2 + lb,) c3 <0",
"(12)",
"Cgap=",
"(Xend - T)2 +(Yend - Ty)2 +(Zend - Tz)2 (9)",
"其中a,=领队的位置；F,=食物来源的位置；ub,和lbi",
"是jth维度的上下界；c1、C2、c3是均匀随机数",
"（Xend，Yend，Zend）表示生成的路由的端点；（Tr，Ty，T）表",
"在公式13中，C1在探索&开发阶段之间保持平衡",
"示自标位置",
"可以使用公式10计算路线的总成本",
"C1 = 2e-()2",
"(13)",
"其中，1：当前送代，L：最大送代次数",
"Ctotal = Pi × Cfuel + P2 × Cdivergence + P3 × Cgap (10)",
"探索被定义为算法试图探索搜索空间的阶段。避免本地解决",
"方案发生在此阶段。勘探之后是开发阶段，在该阶段中，主要",
"总成本函数有助于从路线集中实现最佳路线",
"关注的是即兴创作在勘探阶段中探索的解决方案"
] |
{
"arxiv_id": "1911.10519",
"language": "zh",
"page_idx": 2
}
|
||
2303.13868-zh-0
|
2303.13868-zh-0.jpg
|
Physically Adversarial Infrared Patches with Learnable Shapes and Locations
Xingxing Wei$^* $, Jie Yu, Yao Huang
Institute of Artificial Intelligence, Beihang University, Beijing, 100191, China
xxwei@buaa.edu.cn, sy$2106137@$buaa.edu.cn, y\_huang@buaa.edu.cn
## Abstract
Owing to the extensive application of infrared object detectors in the safety-critical tasks, it is necessary to evaluate their robustness against adversarial examples in the real world. However, current few physical infrared attacks are complicated to implement in practical application because of their complex transformation from digital world to physical world. To address this issue, in this paper, we propose a physically feasible infrared attack method called "adversarial infrared patches". Considering the imaging mechanism of infrared cameras by capturing objects' thermal radiation, adversarial infrared patches conduct attacks by attaching a patch of thermal insulation materials on the target object to manipulate its thermal distribution. To enhance adversarial attacks, we present a novel aggregation regularization to guide the simultaneous learning for the patch' shape and location on the target object. Thus, a simple gradient-based optimization can be adapted to solve for them. We verify adversarial infrared patches in different object detection tasks with various object detectors. Experimental results show that our method achieves more than 90% Attack Success Rate (ASR) versus the pedestrian detector and vehicle detector in the physical environment, where the objects are captured in different angles, distances, postures, and scenes. More importantly, adversarial infrared patch is easy to implement, and it only needs 0.5 hours to be constructed in the physical world, which verifies its effectiveness and efficiency.
## 1. Introduction
深度神经网络(DNN)在各种视觉任务中显示出很有前途的性能，包括对象检测[14],分类[8],人脸识别[16]和自动驾驶[15] 。但是，通常已知 DNN 容易受到对抗性示例[2,5,26] ,i 的攻击。e。,人类察觉不到的扰动输人会欺骗基于 DNNs 的系统给出错误的
$^{*}$Corresponding author
<FigureHere>
Figure 1. The generation process of adversarial infrared patches. We see the pedestrian cannot be detected after the infrared patches are pasted on the pedestrian in the physical world.
预测。此外，这些对抗性的例子可以在物理世界中被利用。在这种情况下，一种广泛使用的技术称为对抗性补丁[1,4,24,28] ,该技术已通过生成精心设计的贴纸[4,23]或通过在眼镜框[17,22]上添加特定纹理来成功应用于交通标志检测。对抗性补丁的成功引起了人们的关注，因为它们对现实世界中部署的基于 DNN 的系统构成了巨大威胁
如今，以其在恶劣环境中的强大抗干扰能力而闻名， 热红外图像中的目标检测已广泛应用于许多安全关键任务，例如安全监控[19] ,遥感[25] 等。因此，有必要评估红外目标探测器的物理对抗鲁棒性。但是，上述对抗性补丁在红外图像中不能很好地工作，因为它们取决于从 RGB 外观的角度产生的对抗性扰动。红外摄像机无法捕获这些扰动，红外摄像机通过对物体的热辐射[20] 进行编码来执行成像。尽管最近很少有人提出解决这个问题的工作，但它们也有自己的局限性。例如，基于轻小灯泡[30] 纸板的对抗性灯泡在现实世界中实现起来很复杂，并且由于产生热源而不隐身。基于大规模 QR 码的对抗性服装 [29] 通过利用隔热材料覆盖物体的热辐射来提高隐蔽性，但仍然具有从数字到物理世界的复杂转换，因此在现实世界中不容易实现
在本文中，我们提出了一种物理上隐身且易于实现的红外攻击方法，称为“对抗性红外补丁”。考虑到红外摄像机通过捕捉物体热辐射的成像机理，我们还将隔热材料附着在目标物体上，以操纵其热分布，从而确保隐蔽性。但与通过复杂的 QR 码图案对抗服装 [29] 不同，我们利用简单的补丁来裁剪隔热材料，然后调整补丁在目标物体上的形状和位置以进行攻击。与对抗
|
[
[
0.09722900390625,
0.1351318359375,
0.87353515625,
0.1351318359375,
0.87353515625,
0.149658203125,
0.09722900390625,
0.149658203125
],
[
0.326904296875,
0.184326171875,
0.646484375,
0.184326171875,
0.646484375,
0.1988525390625,
0.326904296875,
0.1988525390625
],
[
0.1641845703125,
0.2008056640625,
0.80615234375,
0.2008056640625,
0.80615234375,
0.21533203125,
0.1641845703125,
0.21533203125
],
[
0.25732421875,
0.2203369140625,
0.7060546875,
0.2203369140625,
0.7060546875,
0.2342529296875,
0.25732421875,
0.2342529296875
],
[
0.23779296875,
0.268310546875,
0.314453125,
0.268310546875,
0.314453125,
0.285888671875,
0.23779296875,
0.285888671875
],
[
0.1004638671875,
0.3017578125,
0.46728515625,
0.3017578125,
0.46728515625,
0.31494140625,
0.1004638671875,
0.31494140625
],
[
0.080078125,
0.318115234375,
0.468017578125,
0.318115234375,
0.468017578125,
0.33203125,
0.080078125,
0.33203125
],
[
0.08087158203125,
0.333984375,
0.468017578125,
0.333984375,
0.468017578125,
0.347900390625,
0.08087158203125,
0.347900390625
],
[
0.08087158203125,
0.348388671875,
0.46728515625,
0.348388671875,
0.46728515625,
0.3623046875,
0.08087158203125,
0.3623046875
],
[
0.74169921875,
0.349853515625,
0.75732421875,
0.349853515625,
0.75732421875,
0.359130859375,
0.74169921875,
0.359130859375
],
[
0.080078125,
0.3642578125,
0.468017578125,
0.3642578125,
0.468017578125,
0.37744140625,
0.080078125,
0.37744140625
],
[
0.080078125,
0.37890625,
0.468017578125,
0.37744140625,
0.468017578125,
0.392578125,
0.080078125,
0.393310546875
],
[
0.64794921875,
0.3837890625,
0.7294921875,
0.3837890625,
0.7294921875,
0.39453125,
0.64794921875,
0.39453125
],
[
0.54345703125,
0.392578125,
0.62353515625,
0.392578125,
0.62353515625,
0.40283203125,
0.54345703125,
0.40283203125
],
[
0.73193359375,
0.386962890625,
0.74169921875,
0.386962890625,
0.74169921875,
0.392578125,
0.73193359375,
0.392578125
],
[
0.08087158203125,
0.39453125,
0.4697265625,
0.39453125,
0.4697265625,
0.40771484375,
0.08087158203125,
0.40771484375
],
[
0.76953125,
0.393310546875,
0.83642578125,
0.393310546875,
0.83642578125,
0.40087890625,
0.76953125,
0.40087890625
],
[
0.8408203125,
0.395751953125,
0.84619140625,
0.395751953125,
0.84619140625,
0.40087890625,
0.8408203125,
0.40087890625
],
[
0.84619140625,
0.395751953125,
0.85107421875,
0.395751953125,
0.85107421875,
0.40087890625,
0.84619140625,
0.40087890625
],
[
0.08087158203125,
0.4091796875,
0.468017578125,
0.4091796875,
0.468017578125,
0.42236328125,
0.08087158203125,
0.42236328125
],
[
0.50244140625,
0.410888671875,
0.89208984375,
0.410888671875,
0.89208984375,
0.4248046875,
0.50244140625,
0.4248046875
],
[
0.08087158203125,
0.423583984375,
0.468017578125,
0.423583984375,
0.468017578125,
0.4375,
0.08087158203125,
0.4375
],
[
0.50341796875,
0.4248046875,
0.890625,
0.4248046875,
0.890625,
0.438232421875,
0.50341796875,
0.438232421875
],
[
0.08087158203125,
0.438232421875,
0.46728515625,
0.438232421875,
0.46728515625,
0.45263671875,
0.08087158203125,
0.45263671875
],
[
0.50341796875,
0.438232421875,
0.89208984375,
0.438232421875,
0.89208984375,
0.45263671875,
0.50341796875,
0.45263671875
],
[
0.08087158203125,
0.453857421875,
0.4697265625,
0.453857421875,
0.4697265625,
0.46728515625,
0.08087158203125,
0.46728515625
],
[
0.5009765625,
0.451904296875,
0.6044921875,
0.45068359375,
0.6044921875,
0.46533203125,
0.5009765625,
0.466552734375
],
[
0.08251953125,
0.47021484375,
0.46728515625,
0.47021484375,
0.46728515625,
0.48095703125,
0.08251953125,
0.48095703125
],
[
0.50341796875,
0.477783203125,
0.890625,
0.477783203125,
0.890625,
0.4912109375,
0.50341796875,
0.4912109375
],
[
0.080078125,
0.48486328125,
0.468017578125,
0.48291015625,
0.468017578125,
0.496826171875,
0.080078125,
0.498779296875
],
[
0.5048828125,
0.49365234375,
0.89208984375,
0.49365234375,
0.89208984375,
0.5068359375,
0.5048828125,
0.5068359375
],
[
0.08087158203125,
0.5,
0.468017578125,
0.5,
0.468017578125,
0.51318359375,
0.08087158203125,
0.51318359375
],
[
0.5048828125,
0.50830078125,
0.890625,
0.50830078125,
0.890625,
0.521484375,
0.5048828125,
0.521484375
],
[
0.080078125,
0.5146484375,
0.46728515625,
0.5146484375,
0.46728515625,
0.52783203125,
0.080078125,
0.52783203125
],
[
0.080078125,
0.52880859375,
0.46728515625,
0.52880859375,
0.46728515625,
0.54248046875,
0.080078125,
0.54248046875
],
[
0.50341796875,
0.5234375,
0.890625,
0.5244140625,
0.890625,
0.5361328125,
0.50341796875,
0.53515625
],
[
0.5048828125,
0.53857421875,
0.890625,
0.53857421875,
0.890625,
0.5517578125,
0.5048828125,
0.5517578125
],
[
0.08251953125,
0.54541015625,
0.46728515625,
0.54541015625,
0.46728515625,
0.5595703125,
0.08251953125,
0.5595703125
],
[
0.5048828125,
0.55322265625,
0.890625,
0.55322265625,
0.890625,
0.56640625,
0.5048828125,
0.56640625
],
[
0.08087158203125,
0.56005859375,
0.468017578125,
0.56005859375,
0.468017578125,
0.57373046875,
0.08087158203125,
0.57373046875
],
[
0.5048828125,
0.568359375,
0.65185546875,
0.568359375,
0.65185546875,
0.58203125,
0.5048828125,
0.58203125
],
[
0.08087158203125,
0.57470703125,
0.46728515625,
0.57470703125,
0.46728515625,
0.58837890625,
0.08087158203125,
0.58837890625
],
[
0.52392578125,
0.58544921875,
0.8896484375,
0.58544921875,
0.8896484375,
0.5986328125,
0.52392578125,
0.5986328125
],
[
0.08251953125,
0.59130859375,
0.465576171875,
0.59130859375,
0.465576171875,
0.6015625,
0.08251953125,
0.6015625
],
[
0.50341796875,
0.5986328125,
0.8896484375,
0.5986328125,
0.8896484375,
0.6123046875,
0.50341796875,
0.6123046875
],
[
0.08087158203125,
0.60498046875,
0.468017578125,
0.60498046875,
0.468017578125,
0.61865234375,
0.08087158203125,
0.61865234375
],
[
0.08251953125,
0.619140625,
0.468017578125,
0.619140625,
0.468017578125,
0.63330078125,
0.08251953125,
0.63330078125
],
[
0.50341796875,
0.6142578125,
0.890625,
0.6142578125,
0.890625,
0.62744140625,
0.50341796875,
0.62744140625
],
[
0.5048828125,
0.62890625,
0.89208984375,
0.62890625,
0.89208984375,
0.642578125,
0.5048828125,
0.642578125
],
[
0.08087158203125,
0.63525390625,
0.46728515625,
0.63525390625,
0.46728515625,
0.6484375,
0.08087158203125,
0.6484375
],
[
0.50341796875,
0.64306640625,
0.8896484375,
0.64306640625,
0.8896484375,
0.6572265625,
0.50341796875,
0.6572265625
],
[
0.08087158203125,
0.65087890625,
0.46728515625,
0.65087890625,
0.46728515625,
0.6640625,
0.08087158203125,
0.6640625
],
[
0.08087158203125,
0.66552734375,
0.46728515625,
0.66552734375,
0.46728515625,
0.6787109375,
0.08087158203125,
0.6787109375
],
[
0.50341796875,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.50341796875,
0.673828125
],
[
0.5048828125,
0.6748046875,
0.890625,
0.6748046875,
0.890625,
0.68798828125,
0.5048828125,
0.68798828125
],
[
0.08087158203125,
0.68115234375,
0.468017578125,
0.68115234375,
0.468017578125,
0.6943359375,
0.08087158203125,
0.6943359375
],
[
0.50341796875,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.50341796875,
0.7041015625
],
[
0.08251953125,
0.69775390625,
0.465576171875,
0.69775390625,
0.465576171875,
0.70849609375,
0.08251953125,
0.70849609375
],
[
0.08087158203125,
0.7109375,
0.46728515625,
0.7109375,
0.46728515625,
0.724609375,
0.08087158203125,
0.724609375
],
[
0.5048828125,
0.70556640625,
0.890625,
0.70556640625,
0.890625,
0.7197265625,
0.5048828125,
0.7197265625
],
[
0.080078125,
0.724609375,
0.393798828125,
0.7265625,
0.393798828125,
0.74072265625,
0.080078125,
0.73779296875
],
[
0.50341796875,
0.7197265625,
0.8896484375,
0.7197265625,
0.8896484375,
0.73291015625,
0.50341796875,
0.73291015625
],
[
0.50341796875,
0.734375,
0.88818359375,
0.73486328125,
0.88818359375,
0.74853515625,
0.50341796875,
0.74755859375
],
[
0.5048828125,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.7646484375,
0.5048828125,
0.7646484375
],
[
0.5048828125,
0.76513671875,
0.8896484375,
0.76513671875,
0.8896484375,
0.77880859375,
0.5048828125,
0.77880859375
],
[
0.07843017578125,
0.77294921875,
0.21728515625,
0.7734375,
0.21728515625,
0.78857421875,
0.07843017578125,
0.787109375
],
[
0.50244140625,
0.77880859375,
0.8798828125,
0.77978515625,
0.8798828125,
0.79443359375,
0.50244140625,
0.79345703125
],
[
0.10296630859375,
0.7998046875,
0.46728515625,
0.7998046875,
0.46728515625,
0.81298828125,
0.10296630859375,
0.81298828125
],
[
0.52294921875,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.52294921875,
0.81005859375
],
[
0.08251953125,
0.81494140625,
0.46728515625,
0.81494140625,
0.46728515625,
0.8291015625,
0.08251953125,
0.8291015625
],
[
0.5048828125,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.5048828125,
0.82470703125
],
[
0.08087158203125,
0.830078125,
0.46728515625,
0.8291015625,
0.46728515625,
0.84326171875,
0.08087158203125,
0.8447265625
],
[
0.5048828125,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.5048828125,
0.84033203125
],
[
0.08251953125,
0.84521484375,
0.468017578125,
0.84521484375,
0.468017578125,
0.859375,
0.08251953125,
0.859375
],
[
0.50341796875,
0.84033203125,
0.88818359375,
0.84130859375,
0.88818359375,
0.85595703125,
0.50341796875,
0.85498046875
],
[
0.08087158203125,
0.85986328125,
0.468017578125,
0.85986328125,
0.468017578125,
0.87353515625,
0.08087158203125,
0.87353515625
],
[
0.50341796875,
0.85693359375,
0.88623046875,
0.85693359375,
0.88623046875,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.50341796875,
0.87060546875,
0.8896484375,
0.8720703125,
0.8896484375,
0.88623046875,
0.50341796875,
0.88525390625
],
[
0.098876953125,
0.88818359375,
0.23779296875,
0.88818359375,
0.23779296875,
0.9013671875,
0.098876953125,
0.9013671875
],
[
0.5048828125,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5048828125,
0.90087890625
]
] |
[
"Physically Adversarial Infrared Patches with Learnable Shapes and Locations",
"Xingxing Wei,Jie Yu,Yao Huang",
"Institute of Artificial Intelligence,Beihang University,Beijing,10019l,China",
"xxwei@buaa.edu.cn, sy2106137@buaa.edu.cn,y_huang@buaa.edu.cr",
"Abstract",
"Owing to the extensive application of infrared ob",
"ject detectors in the safety-critical tasks, it is neces",
"sary to evaluate their robustness against adversarial",
"examples in the real world. However, current few phys.",
"sticl",
"ical infrared attacks are complicated to implement in",
"practical application because of their complex transfor-",
"Patch in physical ",
"Original Object",
"o1",
"mation from digital world to physical world.To ad-",
"PerturbedOb",
"e",
"",
"dress this issue, in this paper, we propose a physically",
"Figure 1. The generation process of adversarial infrared",
"feasible infrared attack method called \"adversarial in",
"patches. We see the pedestrian cannot be detected after",
"frared patches\" Considering the imaging mechanism",
"the infrared patches are pasted on the pedestrian in the",
"of infrared cameras by capturing objects' thermal ra-",
"physical world.",
"diation,adversarial infrared patches conduct attacks",
"预测。此外，这些对抗性的例子可以在物理世界中被",
"by attaching a patch of thermal insulation materials",
"利用。在这种情况下，一种广泛使用的技术称为对抗",
"on the target object to manipulate its thermal distri.",
"性补丁1,4,24,28］，该技术已通过生成精心设计的贴",
"bution. To enhance adversarial attacks, we present a",
"novel aggregation regularization to guide the simulta",
"纸「4.23]或通过在眼镜框「17.221上添加特定纹理来成",
"功应用于交通标志检测。对抗性补丁的成功引起了人",
"neous learning for the patch' shape and location on",
"们的关注，因为它们对现实世界中部署的基于DNN的",
"the target object. Thus, a simple gradient-based opti",
"系统构成了巨大威胁",
"mization can be adapted to solve for them. We verify",
"如今，以其在恶劣环境中的强大抗干扰能力而闻名，",
"adversarial infrared patches in different object detec",
"热红外图像中的目标检测已广泛应用于许多安全关键",
"tion tasks with various object detectors. Experimental",
"results show that our method achieves more than 90 %",
"任务，例如安全监控【19]，遥感[25]等。因此，有必",
"要评估红外目标探测器的物理对抗鲁棒性。但是，上述",
"Attack Success Rate (ASR) versus the pedestrian detec",
"对抗性补丁在红外图像中不能很好地工作，因为它们",
"tor and vehicle detector in the physical environment,",
"where the objects are captured in different angles, dis.",
"取决于从RGB外观的角度产生的对抗性扰动。红外摄",
"像机无法捕获这些扰动，红外摄像机通过对物体的热",
"tances, postures, and scenes. More importantly, adver-",
"辐射[20]进行编码来执行成像。尽管最近很少有人提",
"sarial infrared patch is easyto implement,and it only",
"needs 0.5 hours to be constructed in the physical world,",
"出解决这个问题的工作，但它们也有自己已的局限性。例",
"which verifies its effectiveness and efficiency",
"如，基于轻小灯泡[30】纸板的对抗性灯泡在现实世界",
"中实现起来很复杂，并且由于产生热源而不隐身。基于",
"大规模QR码的对抗性服装【29]通过利用隔热材料覆",
"盖物体的热辐射来提高隐蔽性，但仍然具有从数字到",
"1.Introduction",
"物理世界的复杂转换，因此在现实世界中不容易实现",
"深度神经网络（DNN）在各种视觉任务中显示出很",
"在本文中，我们提出了一种物理上隐身且易于实现",
"有前途的性能，包括对象检测[14]，分类[8]，人脸识",
"的红外攻击方法，称为“对抗性红外补丁”。考虑到红",
"别[16]和自动驾驶[15]。但是，通常已知DNN容易",
"外摄像机通过捕捉物体热辐射的成像机理，我们还将",
"受到对抗性示例[2,5,26]，i的攻击。。e。，人类察觉",
"隔热材料附着在目标物体上，以操纵其热分布，从而确",
"不到的扰动输人会欺骗基于DNNs的系统给出错误的",
"保隐蔽性。但与通过复杂的QR码图案对抗服装[29",
"不同，我们利用简单的补丁来裁剪隔热材料，然后调整",
"*Corresponding author",
"补丁在目标物体上的形状和位置以进行攻击。与对抗"
] |
{
"arxiv_id": "2303.13868",
"language": "zh",
"page_idx": 0
}
|
||
2303.13868-zh-3
|
2303.13868-zh-3.jpg
|
,其中较大的$C(M)$表示更好的聚合度
显然，$\mathcal{L}_agg$计算的所有操作都是可微的，可以在基于梯度的优化方法中应用。我们的目标是减少$\mathcal{L}_{agg}$ 使红外补丁簇足够大，以便掩码$M$中的点聚集在可用位置上
## 3.2.2 Binary regularization
如前所述，$M$是一个二进制矩阵。为了实现这一目标我们使用 MSE (均方误差)来构造正则化，如下所示：
$$\mathcal{L}_{MSE}(\mathcal{H}(M),\boldsymbol{I})=\frac{1}{hw}\sum_{j=1}^{h}\sum_{k=1}^{w}(\mathcal{H}(M_{jk})-1)^{2},$$
其中$I$是一个矩阵，其每个元素为 1。这里的函数$\mathcal{H}$ 类似于 ReLU 函数。$\mathcal{H}(M_{jk})=1$如果$M_jk\leq V_{thre}$和$\mathcal{H}(M_{jk})=M_{jk}$ 如果 $M_jk\geq V_{thre}$。在 ReLU 函数中， 值等于零的点的导数被设置为零，而在函数$\mathcal{H}$中，值等于$V_{thre}$的点的导数也被设置为零，因此我们可以使$\mathcal{H}$像 ReLU 函数一样可微分。通过最小化方程(8), 我们只测量一个与$M$中大于阈值$V_{thre}$的像素值之间的距离
此外，由于补丁应该是目标对象上的局部区域，因此我们在$M$上添加了稀疏限制。由于$L_0$ -norm 是不可微的，因此我们将其替换为$L_1$ -norm。通过最小化$||M||_1$ ,我们可以选择尽可能稀疏的像素，以使行人在图像中消失。最终的正则化定义如下：
$$\mathcal{L}_{binary}(M)=||M||_{1}+\alpha\cdot\mathcal{L}_{MSE}(\mathcal{H}(M),\boldsymbol{I}),$$
(9
其中$\alpha$是一个加权因子。我们的目标是最小化$\mathcal{L}_binary$
,以便我们优化稀疏和二进制$M$
3.2.3 Total loss function
现在，我们将上述三个损失合并为我们的总损失，并通过加权因子$\lambda_1$和$\lambda_2$在损失之间保持平衡。最终的目标函数由下式给出：
$$\mathcal{L}_{obj}=\mathcal{L}_{attack}(f(\boldsymbol{x}_{adv}))+\lambda_{1}\mathcal{L}_{binary}(\boldsymbol{M})+\lambda_{2}\mathcal{L}_{agg}(\boldsymbol{M})$$
(10) (11)
$$M^*=\arg\min_M\mathcal{L}_{obj}(f(\boldsymbol{x}_{adv})),$$
,其中加权因子可以根据经验获得，以满足我们的要求。在实现中，为了保证红外贴片贴附在目标物体上，应将其限制在红外图像中目标物体的区域内。详细地说，给定一个可用的二进制$M_{obj}$作为对象掩码来定位目标对象，我们可以处理方程(2)中的$M=M\cdot M_{obj}$来执行以下优化。$M_{obj}$可以通过对象检测或分割获得
## 3.2.4 Optimization
为了求解方程 (10 ), 我们使用具有动量 [11] 的梯度下降算法。令$t$表示第$t$次迭代，$x_t^{adv},M_t$分别表示第$t$次迭代中的对抗图像和红外掩模，则：
(12)
$$\boldsymbol{x}_{t+1}^{adv}=\boldsymbol{x}\odot(1-M_{t+1})+\hat{\boldsymbol{x}}\odot M_{t+1},$$
$$M_{t+1}=(M_t-\epsilon\cdot g_{t+1})\odot M_{obj},$$
(13)
$t$中的某些元素。为了平滑这些元素，我们在优化过程
中微调梯度如下：
(14)
$$g_{t+1}=g_{t+1}\odot Norm(M_t*\mathcal{K}_{gau}),$$
其中$Norm(\cdot)$是归一化，$\mathcal{K}_gau$是高斯核，*是卷积。方程( 14 )考虑元素的上下文信息来去除异常点，使形状平滑
# 3.3. Analysis for Cover Image
物体的红外热像由物体发出的热辐射决定。基于这一特点，我们在目标的特定位置放置了一块隔热材料以减少热辐射，从而改变了物体红外辐射的分布，以欺骗探测器
<FigureHere>
Figure 4. The comparison of infrared images taken by the infrared camera between (b) normal object and (c) object with patches made by the (a) aerogel material.
在实践中，由于其低热辐射和低热发射率，我们选择气凝胶材料作为隔热材料。材料及其隔热效果见图 4 。我们可以看到物体上的材料可以屏蔽目标物体的大部分热辐射，并在一定程度上减少相应区域的热辐射。我们发现与红外图像中具有材料的区域对应的灰度值几乎相同，因此将这些灰度值设置为$\hat{x}$
## 3.4. The Overall Algorithm
算法 1 中总结了生成对抗性红外补丁的整体算法。给定行人检测器和红外图像，算法 1 将输出对抗性红外补丁和对抗性图像。图 5 给出了优化掩模$M$的工艺。我们看到面具逐渐收敛到一个稳定的形状和位置
# 4. Experiments
在本节中，我们首先在数字世界中针对行人检测器对我们的方法进行消融研究，然后分别给出物理世界中针对行人检测器和车辆检测器的攻击性能。最后给出了抗辩的结果。更多的实验可以在补充材料中找到
## $4. 1. \textbf{ Simulation of Physical Attacks}$
Dataset:我们使用 FLIRADAS 数据集 $^{\mathrm{l}}$来模拟物理攻击。与 [29] 和 [30] 类似，我们选择包含行人的图像作为我们的数据集。训练集包括 7873 个图像，测试集包括 2027 个图像。然后，我们选择目标模型可以成
$^{1}$https://www.flir.com/oem/adas/ adas-dataset-form/
|
[
[
0.080078125,
0.09344482421875,
0.365234375,
0.0921630859375,
0.365234375,
0.1060791015625,
0.080078125,
0.106689453125
],
[
0.58251953125,
0.09344482421875,
0.8447265625,
0.09344482421875,
0.8447265625,
0.106689453125,
0.58251953125,
0.106689453125
],
[
0.8408203125,
0.0997314453125,
0.84716796875,
0.0997314453125,
0.84716796875,
0.1029052734375,
0.8408203125,
0.1029052734375
],
[
0.859375,
0.0947265625,
0.8896484375,
0.0947265625,
0.8896484375,
0.1060791015625,
0.859375,
0.1060791015625
],
[
0.1004638671875,
0.106689453125,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.1004638671875,
0.12060546875
],
[
0.5009765625,
0.11553955078125,
0.8896484375,
0.11614990234375,
0.8896484375,
0.1300048828125,
0.5009765625,
0.1287841796875
],
[
0.08087158203125,
0.12249755859375,
0.45751953125,
0.12371826171875,
0.45751953125,
0.136962890625,
0.08087158203125,
0.1363525390625
],
[
0.08251953125,
0.1383056640625,
0.46728515625,
0.1383056640625,
0.46728515625,
0.1514892578125,
0.08251953125,
0.1514892578125
],
[
0.5048828125,
0.1331787109375,
0.625,
0.1331787109375,
0.625,
0.143310546875,
0.5048828125,
0.143310546875
],
[
0.08087158203125,
0.15087890625,
0.131591796875,
0.15283203125,
0.1298828125,
0.1673583984375,
0.080078125,
0.1654052734375
],
[
0.576171875,
0.160400390625,
0.86279296875,
0.1590576171875,
0.86279296875,
0.173583984375,
0.576171875,
0.1749267578125
],
[
0.85888671875,
0.160400390625,
0.8896484375,
0.160400390625,
0.8896484375,
0.17236328125,
0.85888671875,
0.17236328125
],
[
0.07843017578125,
0.1849365234375,
0.28759765625,
0.1849365234375,
0.28759765625,
0.1988525390625,
0.07843017578125,
0.1988525390625
],
[
0.50244140625,
0.186279296875,
0.890625,
0.1875,
0.890625,
0.2020263671875,
0.50244140625,
0.2008056640625
],
[
0.08251953125,
0.2083740234375,
0.468017578125,
0.2083740234375,
0.468017578125,
0.2215576171875,
0.08251953125,
0.2215576171875
],
[
0.50341796875,
0.20263671875,
0.890625,
0.20263671875,
0.890625,
0.216552734375,
0.50341796875,
0.216552734375
],
[
0.08087158203125,
0.2235107421875,
0.468017578125,
0.2235107421875,
0.468017578125,
0.2374267578125,
0.08087158203125,
0.2374267578125
],
[
0.50341796875,
0.2197265625,
0.56884765625,
0.2197265625,
0.56884765625,
0.2298583984375,
0.50341796875,
0.2298583984375
],
[
0.240234375,
0.24560546875,
0.47216796875,
0.255126953125,
0.471435546875,
0.2841796875,
0.238525390625,
0.274658203125
],
[
0.50244140625,
0.2412109375,
0.73779296875,
0.24365234375,
0.73779296875,
0.25830078125,
0.50244140625,
0.255615234375
],
[
0.273681640625,
0.248779296875,
0.2998046875,
0.248779296875,
0.2998046875,
0.255615234375,
0.273681640625,
0.255615234375
],
[
0.098876953125,
0.25830078125,
0.236083984375,
0.2568359375,
0.236083984375,
0.271484375,
0.098876953125,
0.272705078125
],
[
0.2303466796875,
0.267578125,
0.2459716796875,
0.267578125,
0.2459716796875,
0.274658203125,
0.2303466796875,
0.274658203125
],
[
0.2418212890625,
0.267578125,
0.283447265625,
0.267578125,
0.283447265625,
0.279052734375,
0.2418212890625,
0.279052734375
],
[
0.521484375,
0.26513671875,
0.8896484375,
0.266357421875,
0.8896484375,
0.281005859375,
0.521484375,
0.27978515625
],
[
0.268798828125,
0.274658203125,
0.318603515625,
0.274658203125,
0.318603515625,
0.285400390625,
0.268798828125,
0.285400390625
],
[
0.5146484375,
0.2822265625,
0.88818359375,
0.2822265625,
0.88818359375,
0.29541015625,
0.5146484375,
0.29541015625
],
[
0.08087158203125,
0.29541015625,
0.468017578125,
0.29541015625,
0.468017578125,
0.309326171875,
0.08087158203125,
0.309326171875
],
[
0.50244140625,
0.29541015625,
0.890625,
0.296630859375,
0.890625,
0.311279296875,
0.50244140625,
0.31005859375
],
[
0.08251953125,
0.311279296875,
0.468017578125,
0.311279296875,
0.468017578125,
0.324462890625,
0.08251953125,
0.324462890625
],
[
0.50341796875,
0.313232421875,
0.56884765625,
0.313232421875,
0.56884765625,
0.324462890625,
0.50341796875,
0.324462890625
],
[
0.08087158203125,
0.32568359375,
0.468017578125,
0.32568359375,
0.468017578125,
0.34033203125,
0.08087158203125,
0.34033203125
],
[
0.08251953125,
0.34228515625,
0.46728515625,
0.34228515625,
0.46728515625,
0.352783203125,
0.08251953125,
0.352783203125
],
[
0.080078125,
0.354736328125,
0.468017578125,
0.35595703125,
0.468017578125,
0.369384765625,
0.080078125,
0.36865234375
],
[
0.08087158203125,
0.371826171875,
0.468017578125,
0.371826171875,
0.468017578125,
0.385009765625,
0.08087158203125,
0.385009765625
],
[
0.08087158203125,
0.386474609375,
0.468017578125,
0.386474609375,
0.468017578125,
0.399658203125,
0.08087158203125,
0.399658203125
],
[
0.08087158203125,
0.40283203125,
0.1298828125,
0.40283203125,
0.1298828125,
0.4140625,
0.08087158203125,
0.4140625
],
[
0.101318359375,
0.41845703125,
0.46728515625,
0.41845703125,
0.46728515625,
0.4287109375,
0.101318359375,
0.4287109375
],
[
0.68896484375,
0.4267578125,
0.69775390625,
0.4267578125,
0.69775390625,
0.431884765625,
0.68896484375,
0.431884765625
],
[
0.8095703125,
0.427978515625,
0.81884765625,
0.427978515625,
0.81884765625,
0.43505859375,
0.8095703125,
0.43505859375
],
[
0.50341796875,
0.44140625,
0.8896484375,
0.44140625,
0.8896484375,
0.451904296875,
0.50341796875,
0.451904296875
],
[
0.08087158203125,
0.431884765625,
0.468017578125,
0.431884765625,
0.468017578125,
0.44580078125,
0.08087158203125,
0.44580078125
],
[
0.08251953125,
0.4462890625,
0.468017578125,
0.4462890625,
0.468017578125,
0.460205078125,
0.08251953125,
0.460205078125
],
[
0.50244140625,
0.45263671875,
0.89208984375,
0.453857421875,
0.89208984375,
0.46728515625,
0.50244140625,
0.466552734375
],
[
0.080078125,
0.460205078125,
0.468017578125,
0.4609375,
0.468017578125,
0.47607421875,
0.080078125,
0.474853515625
],
[
0.5048828125,
0.4697265625,
0.81201171875,
0.4697265625,
0.81201171875,
0.479736328125,
0.5048828125,
0.479736328125
],
[
0.08087158203125,
0.4765625,
0.366943359375,
0.4765625,
0.366943359375,
0.490478515625,
0.08087158203125,
0.490478515625
],
[
0.10540771484375,
0.50048828125,
0.465576171875,
0.50048828125,
0.465576171875,
0.5146484375,
0.10540771484375,
0.5146484375
],
[
0.52392578125,
0.501953125,
0.8896484375,
0.501953125,
0.8896484375,
0.51220703125,
0.52392578125,
0.51220703125
],
[
0.50244140625,
0.51513671875,
0.890625,
0.51513671875,
0.890625,
0.52880859375,
0.50244140625,
0.52880859375
],
[
0.080078125,
0.521484375,
0.465576171875,
0.5234375,
0.465576171875,
0.53857421875,
0.080078125,
0.5361328125
],
[
0.5087890625,
0.53515625,
0.521484375,
0.53515625,
0.521484375,
0.5390625,
0.5087890625,
0.5390625
],
[
0.515625,
0.52978515625,
0.89208984375,
0.53076171875,
0.89208984375,
0.54443359375,
0.515625,
0.54345703125
],
[
0.080078125,
0.537109375,
0.31787109375,
0.537109375,
0.31787109375,
0.55029296875,
0.080078125,
0.55029296875
],
[
0.5048828125,
0.54443359375,
0.890625,
0.54443359375,
0.890625,
0.55810546875,
0.5048828125,
0.55810546875
],
[
0.5048828125,
0.56005859375,
0.89208984375,
0.56005859375,
0.89208984375,
0.57373046875,
0.5048828125,
0.57373046875
],
[
0.080078125,
0.57080078125,
0.268798828125,
0.57080078125,
0.268798828125,
0.583984375,
0.080078125,
0.583984375
],
[
0.5048828125,
0.57568359375,
0.78173828125,
0.57568359375,
0.78173828125,
0.5888671875,
0.5048828125,
0.5888671875
],
[
0.08087158203125,
0.59326171875,
0.46728515625,
0.59326171875,
0.46728515625,
0.60693359375,
0.08087158203125,
0.60693359375
],
[
0.50341796875,
0.60107421875,
0.71826171875,
0.60107421875,
0.71826171875,
0.6142578125,
0.50341796875,
0.6142578125
],
[
0.08087158203125,
0.609375,
0.46728515625,
0.609375,
0.46728515625,
0.62255859375,
0.08087158203125,
0.62255859375
],
[
0.08087158203125,
0.625,
0.221435546875,
0.625,
0.221435546875,
0.63818359375,
0.08087158203125,
0.63818359375
],
[
0.52392578125,
0.62548828125,
0.8896484375,
0.62548828125,
0.8896484375,
0.63623046875,
0.52392578125,
0.63623046875
],
[
0.5048828125,
0.6396484375,
0.8896484375,
0.6396484375,
0.8896484375,
0.65283203125,
0.5048828125,
0.65283203125
],
[
0.08087158203125,
0.64794921875,
0.47216796875,
0.64794921875,
0.47216796875,
0.6611328125,
0.08087158203125,
0.6611328125
],
[
0.50341796875,
0.65283203125,
0.88818359375,
0.65283203125,
0.88818359375,
0.66650390625,
0.50341796875,
0.66650390625
],
[
0.438720703125,
0.66357421875,
0.46728515625,
0.66357421875,
0.46728515625,
0.67529296875,
0.438720703125,
0.67529296875
],
[
0.5048828125,
0.66943359375,
0.8798828125,
0.66943359375,
0.8798828125,
0.68310546875,
0.5048828125,
0.68310546875
],
[
0.1683349609375,
0.67529296875,
0.381591796875,
0.6767578125,
0.381591796875,
0.6943359375,
0.1683349609375,
0.693359375
],
[
0.435546875,
0.67822265625,
0.468017578125,
0.67822265625,
0.468017578125,
0.68994140625,
0.435546875,
0.68994140625
],
[
0.253173828125,
0.68994140625,
0.265625,
0.68994140625,
0.265625,
0.6962890625,
0.253173828125,
0.6962890625
],
[
0.08660888671875,
0.705078125,
0.46728515625,
0.705078125,
0.46728515625,
0.71533203125,
0.08660888671875,
0.71533203125
],
[
0.50244140625,
0.69970703125,
0.6396484375,
0.69970703125,
0.6396484375,
0.71337890625,
0.50244140625,
0.71337890625
],
[
0.08087158203125,
0.71826171875,
0.46728515625,
0.71826171875,
0.46728515625,
0.73193359375,
0.08087158203125,
0.73193359375
],
[
0.52294921875,
0.724609375,
0.890625,
0.724609375,
0.890625,
0.73779296875,
0.52294921875,
0.73779296875
],
[
0.08251953125,
0.734375,
0.465576171875,
0.734375,
0.465576171875,
0.744140625,
0.08251953125,
0.744140625
],
[
0.50341796875,
0.7392578125,
0.890625,
0.7392578125,
0.890625,
0.75244140625,
0.50341796875,
0.75244140625
],
[
0.08087158203125,
0.74853515625,
0.468017578125,
0.74853515625,
0.468017578125,
0.76220703125,
0.08087158203125,
0.76220703125
],
[
0.50244140625,
0.75390625,
0.8896484375,
0.75390625,
0.8896484375,
0.767578125,
0.50244140625,
0.767578125
],
[
0.080078125,
0.76220703125,
0.468017578125,
0.76318359375,
0.468017578125,
0.77783203125,
0.080078125,
0.7763671875
],
[
0.5048828125,
0.771484375,
0.87744140625,
0.771484375,
0.87744140625,
0.7822265625,
0.5048828125,
0.7822265625
],
[
0.08087158203125,
0.77880859375,
0.450927734375,
0.77880859375,
0.450927734375,
0.79248046875,
0.08087158203125,
0.79248046875
],
[
0.5048828125,
0.79541015625,
0.77685546875,
0.79541015625,
0.77685546875,
0.8095703125,
0.5048828125,
0.8095703125
],
[
0.080078125,
0.81005859375,
0.2320556640625,
0.81005859375,
0.2320556640625,
0.82373046875,
0.080078125,
0.82373046875
],
[
0.5087890625,
0.818359375,
0.8896484375,
0.818359375,
0.8896484375,
0.83203125,
0.5087890625,
0.83203125
],
[
0.08416748046875,
0.83544921875,
0.46728515625,
0.83544921875,
0.46728515625,
0.84521484375,
0.08416748046875,
0.84521484375
],
[
0.5048828125,
0.83544921875,
0.88818359375,
0.83544921875,
0.88818359375,
0.84521484375,
0.5048828125,
0.84521484375
],
[
0.08087158203125,
0.8486328125,
0.468017578125,
0.8486328125,
0.468017578125,
0.8623046875,
0.08087158203125,
0.8623046875
],
[
0.50341796875,
0.8486328125,
0.890625,
0.8486328125,
0.890625,
0.8623046875,
0.50341796875,
0.8623046875
],
[
0.080078125,
0.8642578125,
0.36279296875,
0.8642578125,
0.36279296875,
0.87744140625,
0.080078125,
0.87744140625
],
[
0.50244140625,
0.8642578125,
0.8896484375,
0.8642578125,
0.8896484375,
0.87744140625,
0.50244140625,
0.87744140625
],
[
0.1405029296875,
0.88525390625,
0.439453125,
0.88623046875,
0.439453125,
0.90380859375,
0.1405029296875,
0.90283203125
],
[
0.431396484375,
0.88818359375,
0.46728515625,
0.88818359375,
0.46728515625,
0.90087890625,
0.431396484375,
0.90087890625
],
[
0.51953125,
0.88623046875,
0.84716796875,
0.88818359375,
0.84716796875,
0.90283203125,
0.51953125,
0.90087890625
]
] |
[
"，其中较大的C（M）表示更好的聚合度",
"Mt+i =(Mt-e·gt+1)o Mobi,",
"",
"(13)",
"显然，Cagg计算的所有操作都是可微的，可以在基",
"t中的某些元素。为了平滑这些元素，我们在优化过程",
"于梯度的优化方法中应用。我们的目标是减少Ca99",
"使红外补丁簇足够大，以便掩码M中的点聚集在可用",
"中微调梯度如下：",
"位置上",
"gt+i=gt+1Norm(Mt*Kgau)，",
"(14)",
"3.2.2Binary regularization",
"其中Norm(）是归一化，Kgau是高斯核，*是卷积。",
"如前所述，M是一个二进制矩阵。为了实现这一目标，",
"方程（14）考虑元素的上下文信息来去除异常点，使",
"我们使用MSE（均方误差）来构造正则化，如下所示：",
"形状平滑",
"Z(H(Mjk)-1), (8)",
"3.3. Analysis for Cover Image",
"h",
"CMSE(H(M),I)=",
"",
"hwZ",
"物体的红外热像由物体发出的热辐射决定。基于这",
"j=1k=1",
"特点，我们在目标的特定位置放置了一块隔热材料",
"其中I是一个矩阵，其每个元素为1。这里的函数H",
"以减少热辐射，从而改变了物体红外辐射的分布，以欺",
"类似于ReLU函数。H(Mik）=1如果Mik≤Vthre和",
"骗探测器",
"H(Mik）=Mik如果Mik≥Vthre。在ReLU函数中，",
"值等于零的点的导数被设置为零，而在函数孔中，值",
"等于Vthre的点的导数也被设置为零，因此我们可以使",
"H像ReLU函数一样可微分。通过最小化方程（8）",
"我们只测量一个与M中大于阈值Vthre的像素值之间",
"的距离",
"此外，由于补丁应该是自标对象上的局部区域，因",
"(b)",
"(c)",
"Figure4.The comparison of infrared imagestakenbythe",
"此我们在M上添加了稀疏限制。由于Lo-norm是不",
"可微的，因此我们将其替换为Li-norm。通过最小化",
"infrared camera between (b) normal object and (c) object",
"M1，我们可以选择尽可能稀疏的像素，以使行人",
"with patches made bythe (a)aerogel material.",
"在图像中消失。最终的正则化定义如下：",
"Cbinary(M)=I|MIli +Q·LMSE(H(M),I),(9)",
"在实践中，由于其低热辐射和低热发射率，我们选",
"择气凝胶材料作为隔热材料。材料及其隔热效果见图4",
"其中α是一个加权因子。我们的目标是最小化Cbinary",
"",
"我们可以看到物体上的材料可以屏蔽目标物体的大",
"：以便我们优化稀疏和二进制M",
"部分热辐射，并在一定程度上减少相应区域的热辐射。",
"我们发现与红外图像中具有材料的区域对应的灰度值",
"3.2.3Total loss function",
"几乎相同，因此将这些灰度值设置为",
"现在，我们将上述三个损失合并为我们的总损失，并通",
"3.4. The Overall Algorithm",
"过加权因子入1和入2在损失之间保持平衡。最终的目",
"标函数由下式给出：",
"算法1中总结了生成对抗性红外补丁的整体算法。",
"给定行人检测器和红外图像，算法1将输出对抗性红",
"Lobi=Lattack(f(Cadu))+iLbinary(M)+A2Lagg(M)",
"外补丁和对抗性图像。图5给出了优化掩模M的工",
"(10)",
"艺。我们看到面具逐渐收敛到一个稳定的形状和位置",
"M* = arg min Lobj(f(aadu)),",
"(11)",
"M",
"其中加权因子可以根据经验获得，以满足我们的要求。",
"4. Experiments",
"在实现中，为了保证红外贴片贴附在目标物体上，应将",
"在本节中，我们首先在数字世界中针对行人检测器",
"其限制在红外图像中目标物体的区域内。详细地说，给",
"对我们的方法进行消融研究，然后分别给出物理世界",
"定一个可用的二进制Mobi作为对象掩码来定位目标",
"中针对行人检测器和车辆检测器的攻击性能。最后给",
"对象，我们可以处理方程（2）中的M=M·Mobi来",
"出了抗辩的结果。更多的实验可以在补充材料中找到",
"执行以下优化。Mobi可以通过对象检测或分割获得",
"4.1. Simulation of Physical Attacks",
"3.2.4Optimization",
"Dataset：我们使用FLIRADAS数据集1来模拟物",
"为了求解方程（10），我们使用具有动量111的梯度下",
"理攻击。与29和30类似，我们选择包含行人的图",
"降算法。令t表示第t次迭代，aadu，Mt分别表示第",
"像作为我们的数据集。训练集包括7873个图像，测试",
"t次迭代中的对抗图像和红外掩模，则：",
"集包括2027个图像。然后，我们选择目标模型可以成",
"=o(1-Mt+1)+oMt+1,",
"(12)",
"Ihttps://www.flir.com/oem/adas/ adas-dataset-form/"
] |
{
"arxiv_id": "2303.13868",
"language": "zh",
"page_idx": 3
}
|
||
2303.13868-zh-4
|
2303.13868-zh-4.jpg
|
<FigureHere>
Figure 5. The optimization process of mask $M$ for every 5 iterations, where we can see the mask gradually becomes
aggregated and stable.
<FigureHere>
Figure 6. Comparisons with SOTA infrared attacks in dig-
ital world.
<table>
<tbody>
<tr>
<td>$\overline{\mathrm{Algc}}$ Patc</td>
<td>rithm 1 Optimization for Adversarial Infrared $\underline{{\mathrm{ches}}}$</td>
</tr>
<tr>
<td>$\operatorname{Inp}$ $Ou$</td>
<td>ut: Pedestrian detector $f(\cdot)$,clean infrared image $\boldsymbol{x}$ , cover image $\hat{\boldsymbol{x}}$ , restriction region $M_obj$ , max iteration $T$, max size $\varepsilon_{max}$, threshold $s_{thr}.$ tput: Infrared patch $M^*$, and adversarial exam</td>
</tr>
<tr>
<td>M THL </td>
<td>ple adv 汁</td>
</tr>
<tr>
<td>1: 2</td>
<td>$m1t1g$ 1| , $q_{0}= 0$ .</td>
</tr>
<tr>
<td>3: 4:</td>
<td>.01 $t=0$to$T$do Compute the gradient $g_t+1$ according to Eq.(??)</td>
</tr>
<tr>
<td>5:</td>
<td>Finetune the gradient $g_t+1$ according to Eq.(14)</td>
</tr>
<tr>
<td>6:</td>
<td>Undate the mask $\because M.$ $F_{a1}/13$ rlinc $\mathsf{f}_{\mathrm{i}}0$</td>
</tr>
<tr>
<td>7:</td>
<td>Compute rdins Eq. ( 12 $t_{i}0$</td>
</tr>
<tr>
<td>8:</td>
<td>Perform the prediction $y=$ = $f( \boldsymbol{x}_{t+ 1}^{adv}; \theta )$ , and se- lect the top-1 bounding box $(b_1,s_1)$</td>
</tr>
<tr>
<td>9:</td>
<td>if $\|M_{t}$ then $S1$ $s_{th}$ amm $||_{1}$ max</td>
</tr>
<tr>
<td>10:</td>
<td>$M^{*}$ $M_{t}$ 和 $a.$ break 二</td>
</tr>
<tr>
<td>11: </td>
<td>end if</td>
</tr>
<tr>
<td>12:</td>
<td>end for</td>
</tr>
<tr>
<td>13:</td>
<td>$\mu$ $1117m$ 7</td>
</tr>
</tbody>
</table>
<FigureHere>
Figure 7. The experimental setting of ablation study for
the infrared patches' shapes and locations.
12。30%。这些结果验证了我们方法的优势。图 6(b)
# 给出了一个定性示例
Table 1. Ablation study for infrared patches' shapes and
locations.
<FigureHere>
功检测到的高概率的测试图像作为最终图像进行攻击。
因此，干净的 AP 是 100%
Target detector :对于行人检测任务，我们还选择了基于 YOLOv3 [14] 的检测器，因为它的速度很高， 与[29] 和[30] 相同。我们选择官方预训练的权重作为初始化权重，然后在 FLIRADAS 训练数据集上重新训练模型
Metrics:攻击成功率(ASR)和平均精度(AP)用于评估攻击性能。ASR 表示所有测试图像中成功攻击的图像的比率。AP 是通过测量精确召回 (PR)曲线下的区域来计算的
4.1.2 Effects of shapes and locations
4.1.1 Comparisons with SOTA methods
为了研究不同形状和位置的影响，我们在这里进行消融研究。具体来说，我们首先通过求解方程( 10)生成对抗性红外贴片，以实现给定行人的最佳形状和位置 (图 7 中的左子图)。然后，我们进行另外两个实验。一种是在使用随机位置(图 7 中右子图的顶行)时固定最佳形状，然后计算这五个随机位置下的平均 ASR。此设置测试最佳位置对攻击的影响。另一个是固定最佳位置，同时改变不同的补丁的形状(图 7 中右子图的底行)。这些形状可以是正方形，水平矩形，垂直矩形，三角形，菱形等。然后我们计算这五个形状下的平
我们首先测试了我们的方法在数字世界中的攻击性能，并与同一数据集中的对抗性服装 [29] 和对抗性灯泡[30]进行了比较。图$\color{red}6(\mathfrak a)$中绘制了 Precision-Recall 曲线，还列出了相应的 AP 值，我们看到行人检测器的AP 下降到 12。05% 从 100% 攻击后由我们的方法，这是明显低于对抗性灯泡”35。88%, 略低于对抗性服装的
|
[
[
0.2320556640625,
0.09344482421875,
0.273681640625,
0.09344482421875,
0.273681640625,
0.10101318359375,
0.2320556640625,
0.10101318359375
],
[
0.30810546875,
0.09344482421875,
0.338134765625,
0.09344482421875,
0.338134765625,
0.10101318359375,
0.30810546875,
0.10101318359375
],
[
0.3857421875,
0.09344482421875,
0.41015625,
0.09344482421875,
0.41015625,
0.10101318359375,
0.3857421875,
0.10101318359375
],
[
0.4599609375,
0.09344482421875,
0.484375,
0.09344482421875,
0.484375,
0.10101318359375,
0.4599609375,
0.10101318359375
],
[
0.53759765625,
0.09344482421875,
0.57177734375,
0.09344482421875,
0.57177734375,
0.10101318359375,
0.53759765625,
0.10101318359375
],
[
0.615234375,
0.09344482421875,
0.64794921875,
0.09344482421875,
0.64794921875,
0.10101318359375,
0.615234375,
0.10101318359375
],
[
0.76953125,
0.09344482421875,
0.7890625,
0.09344482421875,
0.7890625,
0.10101318359375,
0.76953125,
0.10101318359375
],
[
0.69384765625,
0.0947265625,
0.71142578125,
0.0947265625,
0.71142578125,
0.0985107421875,
0.69384765625,
0.0985107421875
],
[
0.156005859375,
0.0947265625,
0.1895751953125,
0.0947265625,
0.1895751953125,
0.0985107421875,
0.156005859375,
0.0985107421875
],
[
0.326904296875,
0.111083984375,
0.346435546875,
0.111083984375,
0.346435546875,
0.1256103515625,
0.326904296875,
0.1256103515625
],
[
0.406005859375,
0.10986328125,
0.42236328125,
0.10986328125,
0.42236328125,
0.1256103515625,
0.406005859375,
0.1256103515625
],
[
0.484375,
0.111083984375,
0.498291015625,
0.111083984375,
0.498291015625,
0.121826171875,
0.484375,
0.121826171875
],
[
0.080078125,
0.1641845703125,
0.89208984375,
0.1641845703125,
0.89208984375,
0.177978515625,
0.080078125,
0.177978515625
],
[
0.250732421875,
0.11236572265625,
0.268798828125,
0.11236572265625,
0.268798828125,
0.1256103515625,
0.250732421875,
0.1256103515625
],
[
0.07843017578125,
0.177978515625,
0.2303466796875,
0.175537109375,
0.2303466796875,
0.189453125,
0.07843017578125,
0.1912841796875
],
[
0.08087158203125,
0.2083740234375,
0.468017578125,
0.2083740234375,
0.468017578125,
0.2215576171875,
0.08087158203125,
0.2215576171875
],
[
0.08251953125,
0.2247314453125,
0.1380615234375,
0.2247314453125,
0.1380615234375,
0.236083984375,
0.08251953125,
0.236083984375
],
[
0.08416748046875,
0.240478515625,
0.468017578125,
0.2412109375,
0.468017578125,
0.255615234375,
0.08416748046875,
0.255126953125
],
[
0.10540771484375,
0.255126953125,
0.468017578125,
0.255615234375,
0.468017578125,
0.270751953125,
0.10540771484375,
0.26953125
],
[
0.1070556640625,
0.271484375,
0.4248046875,
0.271484375,
0.4248046875,
0.285888671875,
0.1070556640625,
0.285888671875
],
[
0.0849609375,
0.285400390625,
0.468017578125,
0.285400390625,
0.468017578125,
0.2998046875,
0.0849609375,
0.2998046875
],
[
0.69677734375,
0.2890625,
0.73779296875,
0.2890625,
0.73779296875,
0.2998046875,
0.69677734375,
0.2998046875
],
[
0.8037109375,
0.291748046875,
0.8447265625,
0.291748046875,
0.8447265625,
0.298583984375,
0.8037109375,
0.298583984375
],
[
0.108642578125,
0.2998046875,
0.1683349609375,
0.302978515625,
0.1666259765625,
0.317626953125,
0.1070556640625,
0.314453125
],
[
0.5810546875,
0.3017578125,
0.59619140625,
0.3017578125,
0.59619140625,
0.311279296875,
0.5810546875,
0.311279296875
],
[
0.7607421875,
0.302978515625,
0.77392578125,
0.302978515625,
0.77392578125,
0.311279296875,
0.7607421875,
0.311279296875
],
[
0.08660888671875,
0.31494140625,
0.3798828125,
0.31640625,
0.3798828125,
0.330810546875,
0.08660888671875,
0.330078125
],
[
0.50244140625,
0.318115234375,
0.89208984375,
0.318115234375,
0.89208984375,
0.33203125,
0.50244140625,
0.33203125
],
[
0.0849609375,
0.330078125,
0.46728515625,
0.33203125,
0.46728515625,
0.3466796875,
0.0849609375,
0.3447265625
],
[
0.50341796875,
0.333251953125,
0.57177734375,
0.333251953125,
0.57177734375,
0.3447265625,
0.50341796875,
0.3447265625
],
[
0.0882568359375,
0.361083984375,
0.24267578125,
0.361083984375,
0.24267578125,
0.375,
0.0882568359375,
0.375
],
[
0.08905029296875,
0.380126953125,
0.10455322265625,
0.380126953125,
0.10455322265625,
0.389404296875,
0.08905029296875,
0.389404296875
],
[
0.1217041015625,
0.376953125,
0.46728515625,
0.376953125,
0.46728515625,
0.391357421875,
0.1217041015625,
0.391357421875
],
[
0.79150390625,
0.397705078125,
0.80126953125,
0.397705078125,
0.80126953125,
0.4091796875,
0.79150390625,
0.4091796875
],
[
0.0882568359375,
0.40771484375,
0.10540771484375,
0.40771484375,
0.10540771484375,
0.421630859375,
0.0882568359375,
0.421630859375
],
[
0.12335205078125,
0.40771484375,
0.46728515625,
0.40771484375,
0.46728515625,
0.421630859375,
0.12335205078125,
0.421630859375
],
[
0.08905029296875,
0.439453125,
0.10455322265625,
0.439453125,
0.10455322265625,
0.449462890625,
0.08905029296875,
0.449462890625
],
[
0.125,
0.4375,
0.4453125,
0.4375,
0.4453125,
0.451904296875,
0.125,
0.451904296875
],
[
0.0882568359375,
0.45263671875,
0.10540771484375,
0.45263671875,
0.10540771484375,
0.466552734375,
0.0882568359375,
0.466552734375
],
[
0.12335205078125,
0.451904296875,
0.381591796875,
0.451904296875,
0.381591796875,
0.466552734375,
0.12335205078125,
0.466552734375
],
[
0.833984375,
0.458251953125,
0.85302734375,
0.458251953125,
0.85302734375,
0.46533203125,
0.833984375,
0.46533203125
],
[
0.85107421875,
0.458984375,
0.85546875,
0.458984375,
0.85546875,
0.464111328125,
0.85107421875,
0.464111328125
],
[
0.12335205078125,
0.46728515625,
0.4697265625,
0.46728515625,
0.4697265625,
0.481689453125,
0.12335205078125,
0.481689453125
],
[
0.6298828125,
0.4609375,
0.640625,
0.4609375,
0.640625,
0.464111328125,
0.6298828125,
0.464111328125
],
[
0.08905029296875,
0.4697265625,
0.10455322265625,
0.4697265625,
0.10455322265625,
0.479736328125,
0.08905029296875,
0.479736328125
],
[
0.125,
0.48291015625,
0.3798828125,
0.48291015625,
0.3798828125,
0.496826171875,
0.125,
0.496826171875
],
[
0.08905029296875,
0.5,
0.10455322265625,
0.5,
0.10455322265625,
0.51025390625,
0.08905029296875,
0.51025390625
],
[
0.1217041015625,
0.49755859375,
0.414306640625,
0.49755859375,
0.414306640625,
0.51220703125,
0.1217041015625,
0.51220703125
],
[
0.136474609375,
0.51025390625,
0.400390625,
0.51123046875,
0.400390625,
0.52978515625,
0.136474609375,
0.52880859375
],
[
0.50341796875,
0.505859375,
0.890625,
0.505859375,
0.890625,
0.51953125,
0.50341796875,
0.51953125
],
[
0.08251953125,
0.51318359375,
0.10455322265625,
0.51318359375,
0.10455322265625,
0.52587890625,
0.08251953125,
0.52587890625
],
[
0.50341796875,
0.51953125,
0.78515625,
0.52099609375,
0.78515625,
0.5341796875,
0.50341796875,
0.53271484375
],
[
0.08251953125,
0.52880859375,
0.10455322265625,
0.52880859375,
0.10455322265625,
0.54150390625,
0.08251953125,
0.54150390625
],
[
0.12335205078125,
0.52880859375,
0.1690673828125,
0.52880859375,
0.1690673828125,
0.54052734375,
0.12335205078125,
0.54052734375
],
[
0.08087158203125,
0.54150390625,
0.16259765625,
0.54150390625,
0.16259765625,
0.55615234375,
0.08087158203125,
0.55615234375
],
[
0.50244140625,
0.5498046875,
0.890625,
0.55029296875,
0.890625,
0.56494140625,
0.50244140625,
0.564453125
],
[
0.07843017578125,
0.55615234375,
0.226318359375,
0.556640625,
0.226318359375,
0.57470703125,
0.07843017578125,
0.57373046875
],
[
0.50341796875,
0.56640625,
0.65283203125,
0.56640625,
0.65283203125,
0.58056640625,
0.50341796875,
0.58056640625
],
[
0.5009765625,
0.59228515625,
0.89208984375,
0.59326171875,
0.89208984375,
0.60791015625,
0.5009765625,
0.60693359375
],
[
0.08087158203125,
0.60302734375,
0.46728515625,
0.60302734375,
0.46728515625,
0.6162109375,
0.08087158203125,
0.6162109375
],
[
0.50341796875,
0.609375,
0.56787109375,
0.609375,
0.56787109375,
0.62060546875,
0.50341796875,
0.62060546875
],
[
0.08087158203125,
0.61865234375,
0.273681640625,
0.61865234375,
0.273681640625,
0.6318359375,
0.08087158203125,
0.6318359375
],
[
0.08660888671875,
0.6337890625,
0.468017578125,
0.6337890625,
0.468017578125,
0.64794921875,
0.08660888671875,
0.64794921875
],
[
0.6708984375,
0.6337890625,
0.7490234375,
0.6337890625,
0.7490234375,
0.64599609375,
0.6708984375,
0.64599609375
],
[
0.78125,
0.6337890625,
0.814453125,
0.6337890625,
0.814453125,
0.646484375,
0.78125,
0.646484375
],
[
0.77783203125,
0.64794921875,
0.81396484375,
0.64794921875,
0.81396484375,
0.66357421875,
0.77783203125,
0.66357421875
],
[
0.56201171875,
0.63525390625,
0.65625,
0.63525390625,
0.65625,
0.646484375,
0.56201171875,
0.646484375
],
[
0.0849609375,
0.6484375,
0.468017578125,
0.6484375,
0.468017578125,
0.66357421875,
0.0849609375,
0.66357421875
],
[
0.0849609375,
0.6640625,
0.46728515625,
0.6640625,
0.46728515625,
0.67822265625,
0.0849609375,
0.67822265625
],
[
0.60107421875,
0.66552734375,
0.615234375,
0.66552734375,
0.615234375,
0.6767578125,
0.60107421875,
0.6767578125
],
[
0.765625,
0.66357421875,
0.82666015625,
0.66357421875,
0.82666015625,
0.67822265625,
0.765625,
0.67822265625
],
[
0.08251953125,
0.6787109375,
0.468017578125,
0.6787109375,
0.468017578125,
0.693359375,
0.08251953125,
0.693359375
],
[
0.705078125,
0.68115234375,
0.71728515625,
0.68115234375,
0.71728515625,
0.6923828125,
0.705078125,
0.6923828125
],
[
0.765625,
0.67822265625,
0.82666015625,
0.67822265625,
0.82666015625,
0.693359375,
0.765625,
0.693359375
],
[
0.70361328125,
0.6923828125,
0.7197265625,
0.6923828125,
0.7197265625,
0.70703125,
0.70361328125,
0.70703125
],
[
0.7646484375,
0.693359375,
0.8310546875,
0.693359375,
0.8310546875,
0.70849609375,
0.7646484375,
0.70849609375
],
[
0.60302734375,
0.69580078125,
0.61669921875,
0.69580078125,
0.61669921875,
0.70703125,
0.60302734375,
0.70703125
],
[
0.08251953125,
0.69580078125,
0.1298828125,
0.69580078125,
0.1298828125,
0.70703125,
0.08251953125,
0.70703125
],
[
0.08416748046875,
0.70849609375,
0.468017578125,
0.708984375,
0.468017578125,
0.7236328125,
0.08416748046875,
0.72265625
],
[
0.08251953125,
0.724609375,
0.46728515625,
0.724609375,
0.46728515625,
0.73779296875,
0.08251953125,
0.73779296875
],
[
0.50341796875,
0.72802734375,
0.5439453125,
0.72802734375,
0.5439453125,
0.7392578125,
0.50341796875,
0.7392578125
],
[
0.54150390625,
0.7265625,
0.779296875,
0.7265625,
0.779296875,
0.74072265625,
0.54150390625,
0.74072265625
],
[
0.08087158203125,
0.7392578125,
0.46728515625,
0.7392578125,
0.46728515625,
0.75244140625,
0.08087158203125,
0.75244140625
],
[
0.080078125,
0.75390625,
0.1953125,
0.75390625,
0.1953125,
0.767578125,
0.080078125,
0.767578125
],
[
0.50341796875,
0.75048828125,
0.8896484375,
0.75,
0.8896484375,
0.7646484375,
0.50341796875,
0.76513671875
],
[
0.5048828125,
0.7666015625,
0.88818359375,
0.7666015625,
0.88818359375,
0.77978515625,
0.5048828125,
0.77978515625
],
[
0.5048828125,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.5048828125,
0.79443359375
],
[
0.08087158203125,
0.7890625,
0.125,
0.7890625,
0.125,
0.80029296875,
0.08087158203125,
0.80029296875
],
[
0.12335205078125,
0.79248046875,
0.1298828125,
0.79248046875,
0.1298828125,
0.796875,
0.12335205078125,
0.796875
],
[
0.13232421875,
0.78857421875,
0.38232421875,
0.78857421875,
0.38232421875,
0.8017578125,
0.13232421875,
0.8017578125
],
[
0.5048828125,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.5048828125,
0.81005859375
],
[
0.08251953125,
0.81103515625,
0.468017578125,
0.81103515625,
0.468017578125,
0.82470703125,
0.08251953125,
0.82470703125
],
[
0.5185546875,
0.8125,
0.88818359375,
0.8125,
0.88818359375,
0.82275390625,
0.5185546875,
0.82275390625
],
[
0.08087158203125,
0.82568359375,
0.46728515625,
0.82568359375,
0.46728515625,
0.84033203125,
0.08087158203125,
0.84033203125
],
[
0.50244140625,
0.82568359375,
0.890625,
0.82470703125,
0.890625,
0.8388671875,
0.50244140625,
0.84033203125
],
[
0.08087158203125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.08087158203125,
0.85498046875
],
[
0.5048828125,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.08251953125,
0.85595703125,
0.468017578125,
0.85595703125,
0.468017578125,
0.869140625,
0.08251953125,
0.869140625
],
[
0.50634765625,
0.85791015625,
0.8896484375,
0.85791015625,
0.8896484375,
0.86865234375,
0.50634765625,
0.86865234375
],
[
0.08087158203125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.5048828125,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.5048828125,
0.88525390625
],
[
0.080078125,
0.88525390625,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.080078125,
0.8994140625
],
[
0.50244140625,
0.88623046875,
0.890625,
0.88525390625,
0.890625,
0.8994140625,
0.50244140625,
0.90087890625
]
] |
[
"5iteratior",
"10itera",
"5iter",
"20ite",
"25iterat",
"30itera",
"40it",
"35.f",
"niha",
"公",
"i",
"",
"Figure 5. The optimization process of mask M for every 5 iterations, where we can see the mask gradually becomes",
"",
"aggregated and stable.",
"Algorithm 1 Optimization for Adversarial Infrared",
"Patches",
"Input:Pedestrian detector f(),clean infrared image",
",cover image ,restriction region Mobj,max",
"iteration T , max size emar , threshold sthr",
"Output: Infrared patch M* , and adversarial exam-",
"Clean image",
"Infrared patch",
"ple aadu",
"(a)",
"(b)",
"1: Initialize Mo randomly in [0, 1] , go = 0 .",
"Figure 6. Comparisons with SOTA infrared attacks in dig-",
"2: Restrict Mo in the object region: Mo = MooMobj",
"ital world.",
"3: for t = 0 to T do",
"4:",
"Compute the gradient gt+1 according to Eq.(??)",
"-",
"5:",
"Finetune the gradient gt+1 according to Eq.(14)",
"6:",
"Update the mask Mt+1 according to Eq.(13)",
"7:",
"Compute cadi according to Eq.(12)",
"",
"",
"Perform the prediction y = f(cadi;0) , and se-",
"",
"8:",
"lect the top-1 bounding box (bi, si)",
"9:",
"if si ≤ sthr and /Mt+illi ≤emar then",
"M* = Mt+1 ; aadu = ai ; break;",
"Figure 7. The experimental setting of ablation study for",
"10:",
"the infrared patches' shapes and locations.",
"11:",
"end if",
"12: end for",
"12。30%。这些结果验证了我们方法的优势。图6（b）",
"13: return M*, aadu",
"给出了一个定性示例",
"Table 1. Ablation study for infrared patches' shapes and",
"功检测到的高概率的测试图像作为最终图像进行攻击。",
"locations.",
"因此，干净的AP是100%",
"Targetdetector：对于行人检测任务，我们还选择",
"Our shape",
"ASR",
"0%",
"Our location",
"了基于YOLOv3[14]的检测器，因为它的速度很高，",
"与29]和[30]］相同。我们选择官方预训练的权重作为",
"V",
"54.16 %",
"初始化权重，然后在FLIRADAS训练数据集上重新训",
"V",
"59.68 %",
"V",
"93.60 %",
"V",
"练模型",
"Metrics：攻击成功率（ASR）和平均精度（AP）用",
"于评估攻击性能。ASR表示所有测试图像中成功攻击",
"4.1.2 ",
"Effects of shapes and locations",
"的图像的比率。AP是通过测量精确召回（PR）曲线下",
"的区域来计算的",
"为了研究不同形状和位置的影响，我们在这里进行消",
"融研究。具体来说，我们首先通过求解方程（10）生",
"成对抗性红外贴片，以实现给定行人的最佳形状和位",
"4.1.1",
"",
"Comparisons with SOTA methods",
"置（图7中的左子图。然后，我们进行另外两个实验。",
"我们首先测试了我们的方法在数字世界中的攻击性",
"种是在使用随机位置（图7中右子图的顶行）时固",
"能，并与同一数据集中的对抗性服装[29]和对抗性灯",
"定最佳形状，然后计算这五个随机位置下的平均ASR。",
"泡[30进行了比较。图6（a）中绘制了Precision-Recall",
"此设置测试最佳位置对攻击的影响。另一个是固定最",
"曲线，还列出了相应的AP值，我们看到行人检测器的",
"佳位置，同时改变不同的补丁的形状（图了中右子图",
"AP下降到12。05%从100%攻击后由我们的方法，这",
"的底行）。这些形状可以是正方形，水平矩形，垂直矩",
"是明显低于对抗性灯泡’35。88%，略低于对抗性服装的",
"形，三角形，菱形等。然后我们计算这五个形状下的平"
] |
{
"arxiv_id": "2303.13868",
"language": "zh",
"page_idx": 4
}
|
||
2005.02013-zh-0
|
2005.02013-zh-0.jpg
|
## 受控释义生成的神经句法预排序
Tanya Goyal and Greg Durrett Department of Computer Science The University of Texas at Austin
tanyagoyal@utexas.edu, gdurrett@cs.utexas.edu
# $\operatorname{Abstract}$
Paraphrasing natural language sentences is a multifaceted process: it might involve replacing individual words or short phrases, local rearrangement of content, or high-level restructuring like topicalization or passivization. Past approaches struggle to cover this space of paraphrase possibilities in an interpretable manner. Our work, inspired by pre-ordering literature in ma- $\bar{\text{chine translation, uses syntactic transfor-}}$ mations to softly“reorder”the source sentence and guide our neural paraphrasing model. First, given an input sentence, we derive a set of feasible syntactic rearrangements using an encoder-decoder model. This model operates over a partially lexical, partially syntactic view of the sentence and can reorder big chunks. Next, we use each proposed rearrangement to produce a sequence of position embeddings, which encourages our final encoder-decoder paraphrase model to attend to the source words in a particular order. Our evaluation, both automatic and human, shows that the proposed system retains the quality of the baseline approaches while giving a substantial increase in the diversity of the generated paraphrases.$^1$
## 1 Introduction
释 义 生 成 ( McKeown, 1983; Barzilay and Lee, 2003)最近引起了人们的极大兴趣，包括大规模数据集收集和管理 (Lan et al., 2017; Wieting and Gimpel, 2018) 以及诸如深度生成模型(Gupta et al., 2018; Li et al., 2019) 之类的建模进步。如果释义模型公开了可以操纵以产生各种释义 (Iyyer et al., 2018; Chen et al., 2019b; Park et al., 2019) 的控制机制，则证明它们特别有用，这使得这些模型可用于数据增强(Yu
$^{1}$Data and code are available at https://github.com/
tagoyal/sow-reap-paraphrasing
<FigureHere>
Figure 1: Overview of our paraphrase model. First, we choose various pairs of constituents to abstract away in the source sentence, then use a neural transducer to generate possible reorderings of the abstracted sentences. From these, we construct a guide reordering of the input sentence which then informs the generation of output paraphrases.
et al., 2018) 和对抗性示例生成 (Iyyer et al., 2018)。然而，涉及句法控制机制的现有方法不能有效地覆盖释义可能性的空间。使用覆盖解析树 (Iyyer et al., 2018) 顶部的语法模板是不灵活的，使用完全指定的示例句子(Chen et al., 2019b) 会带来如何有效检索此类句子的问题。对于特定的输入句子，使用这些过去的方法来枚举对该句子有意义的一组重新排序是具有挑战性的
在本文中，我们提出了一种两阶段方法来解决这些限制，如图 1 所示。首先，我们使用编码器-解码器模型 (Sow ,用于源顺序重写)在输人句子的各种抽象版本上应用转换操作。这些转换产生了单词和成分的可能的重新排序，可以将其组合以获得输入句子的多个可行的重新排列。每个重新排列都指定了我们应该访问源句子单词的顺序；请注意，这样的顺序可能会鼓励模型被动化(在主题之前访问对象),主题化或重新排序子句。这些顺序通过位置嵌入的方式为我们的编码器-解码器释义模型 (Reap , 用于重排感知释义)进行编码，这些位置嵌入被添加到源句子编码中以指定所需的生成顺序(请参见图$\overline{\mathrm{E}}2$ )。整个工作流程的灵感来自机器翻译 (Xia and McCord, 2004; Collins et al., 2005)中的预购文献；但是，我们的设置明确要求娱乐与不同释义现象相对应的各种可能的
|
[
[
0.33154296875,
0.0831298828125,
0.6689453125,
0.0843505859375,
0.6689453125,
0.1015625,
0.33154296875,
0.1009521484375
],
[
0.366943359375,
0.13720703125,
0.64208984375,
0.13720703125,
0.64208984375,
0.1519775390625,
0.366943359375,
0.1519775390625
],
[
0.359375,
0.15380859375,
0.64990234375,
0.15380859375,
0.64990234375,
0.168701171875,
0.359375,
0.168701171875
],
[
0.355224609375,
0.168701171875,
0.6533203125,
0.17041015625,
0.6533203125,
0.185302734375,
0.355224609375,
0.182861328125
],
[
0.295654296875,
0.1859130859375,
0.7119140625,
0.18408203125,
0.7119140625,
0.201904296875,
0.295654296875,
0.2037353515625
],
[
0.2626953125,
0.26318359375,
0.343505859375,
0.265380859375,
0.343505859375,
0.28076171875,
0.2626953125,
0.279052734375
],
[
0.53662109375,
0.276611328125,
0.603515625,
0.276611328125,
0.603515625,
0.287353515625,
0.53662109375,
0.287353515625
],
[
0.6337890625,
0.27783203125,
0.654296875,
0.27783203125,
0.654296875,
0.28515625,
0.6337890625,
0.28515625
],
[
0.7060546875,
0.279052734375,
0.71728515625,
0.279052734375,
0.71728515625,
0.283935546875,
0.7060546875,
0.283935546875
],
[
0.76904296875,
0.279052734375,
0.779296875,
0.279052734375,
0.779296875,
0.281982421875,
0.76904296875,
0.281982421875
],
[
0.84619140625,
0.279052734375,
0.8564453125,
0.279052734375,
0.8564453125,
0.281982421875,
0.84619140625,
0.281982421875
],
[
0.70947265625,
0.290283203125,
0.86572265625,
0.290283203125,
0.86572265625,
0.300537109375,
0.70947265625,
0.300537109375
],
[
0.60205078125,
0.27978515625,
0.63232421875,
0.27978515625,
0.63232421875,
0.283935546875,
0.60205078125,
0.283935546875
],
[
0.14697265625,
0.29345703125,
0.4609375,
0.29345703125,
0.4609375,
0.30712890625,
0.14697265625,
0.30712890625
],
[
0.5517578125,
0.29345703125,
0.55908203125,
0.29345703125,
0.55908203125,
0.298583984375,
0.5517578125,
0.298583984375
],
[
0.60791015625,
0.29443359375,
0.6630859375,
0.295654296875,
0.6630859375,
0.305908203125,
0.60791015625,
0.3046875
],
[
0.145263671875,
0.30810546875,
0.4609375,
0.30810546875,
0.4609375,
0.32177734375,
0.145263671875,
0.32177734375
],
[
0.5634765625,
0.30712890625,
0.5751953125,
0.303955078125,
0.57666015625,
0.308837890625,
0.56494140625,
0.311279296875
],
[
0.62890625,
0.308837890625,
0.6337890625,
0.308837890625,
0.6337890625,
0.314208984375,
0.62890625,
0.314208984375
],
[
0.74072265625,
0.30810546875,
0.85205078125,
0.30810546875,
0.85205078125,
0.318359375,
0.74072265625,
0.318359375
],
[
0.5576171875,
0.31640625,
0.56494140625,
0.31640625,
0.56494140625,
0.320556640625,
0.5576171875,
0.320556640625
],
[
0.59423828125,
0.318359375,
0.72314453125,
0.31640625,
0.72314453125,
0.334228515625,
0.59423828125,
0.336181640625
],
[
0.73046875,
0.32177734375,
0.8486328125,
0.32177734375,
0.8486328125,
0.33203125,
0.73046875,
0.33203125
],
[
0.55908203125,
0.328857421875,
0.59619140625,
0.328857421875,
0.59619140625,
0.336181640625,
0.55908203125,
0.336181640625
],
[
0.14697265625,
0.32177734375,
0.4609375,
0.32177734375,
0.4609375,
0.3349609375,
0.14697265625,
0.3349609375
],
[
0.724609375,
0.334228515625,
0.8193359375,
0.334228515625,
0.8193359375,
0.33837890625,
0.724609375,
0.33837890625
],
[
0.81787109375,
0.333251953125,
0.84814453125,
0.333251953125,
0.84814453125,
0.34033203125,
0.81787109375,
0.34033203125
],
[
0.14697265625,
0.350830078125,
0.4609375,
0.350830078125,
0.4609375,
0.364013671875,
0.14697265625,
0.364013671875
],
[
0.1485595703125,
0.336181640625,
0.459228515625,
0.336181640625,
0.459228515625,
0.349853515625,
0.1485595703125,
0.349853515625
],
[
0.515625,
0.358154296875,
0.603515625,
0.358154296875,
0.603515625,
0.369384765625,
0.515625,
0.369384765625
],
[
0.59765625,
0.35693359375,
0.88330078125,
0.35693359375,
0.88330078125,
0.37060546875,
0.59765625,
0.37060546875
],
[
0.145263671875,
0.363525390625,
0.4609375,
0.364013671875,
0.4609375,
0.37890625,
0.145263671875,
0.377685546875
],
[
0.51318359375,
0.369384765625,
0.884765625,
0.37060546875,
0.884765625,
0.385498046875,
0.51318359375,
0.38427734375
],
[
0.14697265625,
0.37890625,
0.4609375,
0.377685546875,
0.4609375,
0.391357421875,
0.14697265625,
0.392578125
],
[
0.515625,
0.38720703125,
0.8818359375,
0.38720703125,
0.8818359375,
0.39794921875,
0.515625,
0.39794921875
],
[
0.14697265625,
0.393798828125,
0.459228515625,
0.393798828125,
0.459228515625,
0.40673828125,
0.14697265625,
0.40673828125
],
[
0.5146484375,
0.399658203125,
0.88671875,
0.399658203125,
0.88671875,
0.413330078125,
0.5146484375,
0.413330078125
],
[
0.14697265625,
0.40673828125,
0.4609375,
0.40673828125,
0.4609375,
0.42041015625,
0.14697265625,
0.42041015625
],
[
0.51318359375,
0.41455078125,
0.884765625,
0.41455078125,
0.884765625,
0.427490234375,
0.51318359375,
0.427490234375
],
[
0.1485595703125,
0.4228515625,
0.459228515625,
0.4228515625,
0.459228515625,
0.432861328125,
0.1485595703125,
0.432861328125
],
[
0.51318359375,
0.4287109375,
0.88330078125,
0.4287109375,
0.88330078125,
0.4423828125,
0.51318359375,
0.4423828125
],
[
0.14697265625,
0.435302734375,
0.4609375,
0.435302734375,
0.4609375,
0.448974609375,
0.14697265625,
0.448974609375
],
[
0.51123046875,
0.441162109375,
0.85205078125,
0.443603515625,
0.85205078125,
0.457275390625,
0.51123046875,
0.454833984375
],
[
0.145263671875,
0.448974609375,
0.459228515625,
0.449462890625,
0.459228515625,
0.46435546875,
0.145263671875,
0.463134765625
],
[
0.14697265625,
0.46435546875,
0.4609375,
0.46435546875,
0.4609375,
0.47802734375,
0.14697265625,
0.47802734375
],
[
0.145263671875,
0.476806640625,
0.4609375,
0.478515625,
0.4609375,
0.4921875,
0.145263671875,
0.490478515625
],
[
0.14697265625,
0.4921875,
0.459228515625,
0.4921875,
0.459228515625,
0.505859375,
0.14697265625,
0.505859375
],
[
0.51123046875,
0.488037109375,
0.880859375,
0.4892578125,
0.880859375,
0.50439453125,
0.51123046875,
0.5029296875
],
[
0.14697265625,
0.50732421875,
0.459228515625,
0.50732421875,
0.459228515625,
0.52099609375,
0.14697265625,
0.52099609375
],
[
0.51318359375,
0.50439453125,
0.88330078125,
0.5048828125,
0.88330078125,
0.51953125,
0.51318359375,
0.5185546875
],
[
0.14697265625,
0.52099609375,
0.459228515625,
0.52099609375,
0.459228515625,
0.53369140625,
0.14697265625,
0.53369140625
],
[
0.515625,
0.521484375,
0.88330078125,
0.521484375,
0.88330078125,
0.53515625,
0.515625,
0.53515625
],
[
0.14697265625,
0.53515625,
0.4609375,
0.53515625,
0.4609375,
0.548828125,
0.14697265625,
0.548828125
],
[
0.515625,
0.53759765625,
0.88330078125,
0.53759765625,
0.88330078125,
0.5517578125,
0.515625,
0.5517578125
],
[
0.14697265625,
0.5498046875,
0.459228515625,
0.5498046875,
0.459228515625,
0.5634765625,
0.14697265625,
0.5634765625
],
[
0.515625,
0.552734375,
0.8818359375,
0.552734375,
0.8818359375,
0.56787109375,
0.515625,
0.56787109375
],
[
0.14697265625,
0.5634765625,
0.4609375,
0.5634765625,
0.4609375,
0.57763671875,
0.14697265625,
0.57763671875
],
[
0.5146484375,
0.5693359375,
0.8818359375,
0.5693359375,
0.8818359375,
0.5830078125,
0.5146484375,
0.5830078125
],
[
0.14697265625,
0.57666015625,
0.459228515625,
0.57666015625,
0.459228515625,
0.59033203125,
0.14697265625,
0.59033203125
],
[
0.515625,
0.58447265625,
0.880859375,
0.58447265625,
0.880859375,
0.59716796875,
0.515625,
0.59716796875
],
[
0.14697265625,
0.59130859375,
0.4609375,
0.59130859375,
0.4609375,
0.60498046875,
0.14697265625,
0.60498046875
],
[
0.14697265625,
0.60546875,
0.4609375,
0.60546875,
0.4609375,
0.619140625,
0.14697265625,
0.619140625
],
[
0.5146484375,
0.60107421875,
0.8818359375,
0.60107421875,
0.8818359375,
0.615234375,
0.5146484375,
0.615234375
],
[
0.1485595703125,
0.62158203125,
0.45751953125,
0.62158203125,
0.45751953125,
0.6318359375,
0.1485595703125,
0.6318359375
],
[
0.51708984375,
0.619140625,
0.62451171875,
0.619140625,
0.62451171875,
0.63134765625,
0.51708984375,
0.63134765625
],
[
0.14697265625,
0.63427734375,
0.459228515625,
0.63427734375,
0.459228515625,
0.64794921875,
0.14697265625,
0.64794921875
],
[
0.5341796875,
0.6357421875,
0.88330078125,
0.6357421875,
0.88330078125,
0.65087890625,
0.5341796875,
0.65087890625
],
[
0.14697265625,
0.64794921875,
0.45751953125,
0.64794921875,
0.45751953125,
0.6611328125,
0.14697265625,
0.6611328125
],
[
0.51318359375,
0.65185546875,
0.88330078125,
0.65283203125,
0.88330078125,
0.66748046875,
0.51318359375,
0.66650390625
],
[
0.14697265625,
0.6630859375,
0.4609375,
0.6630859375,
0.4609375,
0.6767578125,
0.14697265625,
0.6767578125
],
[
0.5146484375,
0.66845703125,
0.8818359375,
0.66943359375,
0.8818359375,
0.68408203125,
0.5146484375,
0.68310546875
],
[
0.14697265625,
0.6767578125,
0.2880859375,
0.6767578125,
0.2880859375,
0.6904296875,
0.14697265625,
0.6904296875
],
[
0.51708984375,
0.68603515625,
0.8818359375,
0.68603515625,
0.8818359375,
0.69970703125,
0.51708984375,
0.69970703125
],
[
0.151123046875,
0.70361328125,
0.267822265625,
0.705078125,
0.267822265625,
0.7197265625,
0.151123046875,
0.71875
],
[
0.51318359375,
0.70068359375,
0.880859375,
0.70068359375,
0.880859375,
0.71435546875,
0.51318359375,
0.71435546875
],
[
0.51708984375,
0.71875,
0.8818359375,
0.71875,
0.8818359375,
0.732421875,
0.51708984375,
0.732421875
],
[
0.1259765625,
0.705078125,
0.1351318359375,
0.71142578125,
0.1259765625,
0.71728515625,
0.1175537109375,
0.71142578125
],
[
0.12005615234375,
0.72998046875,
0.48779296875,
0.72998046875,
0.48779296875,
0.74462890625,
0.12005615234375,
0.74462890625
],
[
0.51318359375,
0.73291015625,
0.880859375,
0.73291015625,
0.880859375,
0.74658203125,
0.51318359375,
0.74658203125
],
[
0.12005615234375,
0.74658203125,
0.489501953125,
0.74658203125,
0.489501953125,
0.76025390625,
0.12005615234375,
0.76025390625
],
[
0.515625,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76416015625,
0.515625,
0.76416015625
],
[
0.12005615234375,
0.7626953125,
0.490234375,
0.76123046875,
0.490234375,
0.775390625,
0.12005615234375,
0.77685546875
],
[
0.515625,
0.7666015625,
0.8818359375,
0.7666015625,
0.8818359375,
0.77978515625,
0.515625,
0.77978515625
],
[
0.118408203125,
0.77783203125,
0.48779296875,
0.77685546875,
0.48779296875,
0.79150390625,
0.118408203125,
0.7919921875
],
[
0.515625,
0.7822265625,
0.880859375,
0.7822265625,
0.880859375,
0.79638671875,
0.515625,
0.79638671875
],
[
0.1175537109375,
0.79345703125,
0.486083984375,
0.79345703125,
0.486083984375,
0.80712890625,
0.1175537109375,
0.80712890625
],
[
0.51318359375,
0.79736328125,
0.88330078125,
0.798828125,
0.88330078125,
0.8134765625,
0.51318359375,
0.8125
],
[
0.118408203125,
0.81103515625,
0.484375,
0.81103515625,
0.484375,
0.82470703125,
0.118408203125,
0.82470703125
],
[
0.5146484375,
0.81396484375,
0.880859375,
0.81396484375,
0.880859375,
0.82763671875,
0.5146484375,
0.82763671875
],
[
0.118408203125,
0.82666015625,
0.489501953125,
0.826171875,
0.489501953125,
0.84033203125,
0.118408203125,
0.84130859375
],
[
0.5146484375,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.84423828125,
0.5146484375,
0.84423828125
],
[
0.12005615234375,
0.84326171875,
0.48779296875,
0.84326171875,
0.48779296875,
0.85693359375,
0.12005615234375,
0.85693359375
],
[
0.51123046875,
0.845703125,
0.88330078125,
0.8466796875,
0.88330078125,
0.86083984375,
0.51123046875,
0.85986328125
],
[
0.12005615234375,
0.859375,
0.489501953125,
0.859375,
0.489501953125,
0.87353515625,
0.12005615234375,
0.87353515625
],
[
0.51318359375,
0.86328125,
0.880859375,
0.86328125,
0.880859375,
0.876953125,
0.51318359375,
0.876953125
],
[
0.1376953125,
0.8818359375,
0.489501953125,
0.8828125,
0.489501953125,
0.89794921875,
0.1376953125,
0.896484375
],
[
0.5146484375,
0.87890625,
0.8818359375,
0.87890625,
0.8818359375,
0.892578125,
0.5146484375,
0.892578125
],
[
0.118408203125,
0.896484375,
0.3359375,
0.896484375,
0.3359375,
0.91015625,
0.118408203125,
0.91015625
],
[
0.5146484375,
0.8955078125,
0.880859375,
0.8955078125,
0.880859375,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"受控释义生成的神经句法预排序",
"Tanya Goyal and Greg Durrett",
"Department of Computer Science",
"The University of Texas at Austin",
"tanyagoyal@utexas.edu, gdurrett@cs.utexas.edu",
"Abstract",
"Source Order",
"itine",
"ar",
"A",
"",
"The game was won by the Clippers",
"ou",
"Paraphrasing natural language sentences",
"S",
"XNp won YNP",
"is a multifaceted process: it might in-",
"Vp",
"",
"Transformerseq2seq",
"R",
"Ywon y xcan",
"32",
"on the gamx",
"volve replacing individual words or short",
"ipperswonthe",
"game",
"high-level restructuring like topicalization",
"phrases, local rearrangement of content, or",
"Figure l:",
"Overview of our paraphrase model.",
"or passivization. Past approaches struggle",
"First, we choose various pairs of constituents to ab-",
"to cover this space of paraphrase possibili-",
"stract away in the source sentence, then use a neu",
"ties in an interpretable manner. Our work,",
"ral transducer to generate possible reorderings of",
"inspired by pre-ordering literature in ma-",
"the abstracted sentences. From these, we construct",
"chine translation, uses syntactic transfor-",
"a guide reordering of the input sentence which then",
"mations to softly \"reorder\" the source sen-",
"informs the generation of output paraphrases.",
"tence and guide our neural paraphrasing",
"model. First, given an input sentence, we",
"derive a set of feasible syntactic rearrange-",
"ments using an encoder-decoder model.",
"etal.，2018）和对抗性示例生成（lyyeretal.，",
"This model operates over a partially lexi-",
"2018）。然而，涉及句法控制机制的现有方法",
"cal, partially syntactic view of the sentence",
"不能有效地覆盖释义可能性的空间。。使用覆",
"and can reorder big chunks. Next, we use",
"盖解析树（Iyyeretal.，2018）顶部的语法模板",
"each proposed rearrangement to produce",
"是不灵活的，使用完全指定的示例句子（Chen",
"a sequence of position embeddings, which",
"etal.,2019b）会带来如何有效检索此类句子的",
"encourages our final encoder-decoder para.",
"问题。对于特定的输入句子，使用这些过去的",
"phrase model to attend to the source words",
"in a particular order. Our evaluation, both",
"方法来枚举对该句子有意义的一组重新排序是",
"automatic and human, shows that the pro",
"具有挑战性的",
"posed system retains the quality of the",
"在本文中，我们提出了一种两阶段方法来解",
"baseline approaches while giving a substan.",
"决这些限制，如图1所示。首先，我们使用编码",
"tial increase in the diversity of the gener-",
"器-解码器模型（Sow，用于源顺序重写）在输",
"ated paraphrases.",
"人句子的各种抽象版本上应用转换操作。这些",
"Introduction",
"转换产生了单词和成分的可能的重新排序，可",
"以将其组合以获得输入句子的多个可行的重新",
"1",
"释义生成(McKeown,1983;Barzilay and Lee,",
"排列。每个重新排列都指定了我们应该访问源",
"2003）最近引起了人们的极大兴趣，包括大规",
"句子单词的顺序；请注意，这样的顺序可能会",
"模数据集收集和管理（Lanetal.2017；Wiet-",
"鼓励模型被动化（在主题之前访问对象），主题",
"ingandGimpel，2018）以及诸如深度生成模型",
"化或重新排序子句。这些顺序通过位置嵌人的",
"(Guptaetal..2018:Lietal..2019）之类的建模",
"方式为我们的编码器-解码器释义模型（Reap，",
"进步。如果释义模型公开了可以操纵以产生各",
"用于重排感知释义）进行编码，这些位置嵌入",
"种释义(Iyyer et al.,2018;Chen et al.,2019b;",
"被添加到源句子编码中以指定所需的生成顺序",
"Parketal.2019）的控制机制，则证明它们特",
"（请参见图E2）。整个工作流程的灵感来自机",
"别有用，这使得这些模型可用于数据增强（Yu",
"器翻译 (Xia and McCord, 2004; Collins et al.",
"IData and code are available at https:/ /github.com/",
"2005）中的预购文献；但是，我们的设置明确",
"tagoyal/sow-reap-paraphrasing",
"要求娱乐与不同释义现象相对应的各种可能的"
] |
{
"arxiv_id": "2005.02013",
"language": "zh",
"page_idx": 0
}
|
||
2005.02013-zh-1
|
2005.02013-zh-1.jpg
|
<FigureHere>
Figure 2: Rearrangement aware paraphrasing ( Reap $){\mathrm{model.}}{\mathrm{The}}{\mathrm{~gray~area~corresponds~to}}$ the standard transformer encoder-decoder system. Our model adds position embeddings corresponding to the target reordering to encoder outputs. The decoder attends over these augmented encodings during both training and inference.
## 排序
我们在大规模英语释义数据集 ParaNMT- 50M (Wieting and Gimpel, 2018) 上训练和评估我们的方法。结果表明，我们的方法产生了更多不同的释义，同时保留了强基线模型所表现出的质量。我们进一步证明了所提出的基于语法的转换过程为输入句子生成了一组可行的重排。最后，我们表明位置嵌入提供了一种简单而有效的方式来编码重新排序信息，并且生成的释义与所需的重新排序输入具有很高的一致性
### 2 Method
给定一个输入句子$x=\{x_1,x_2,\ldots,x_n\}$ ,我们的目标是生成一组结构上不同的释义$Y=$ $\{\boldsymbol{y}^1,\boldsymbol{y}^2,\ldots,\boldsymbol{y}^k\}$ 。我们通过首先为输入句子$R=\{\boldsymbol{r}^1,\boldsymbol{r}^2,\ldots,\boldsymbol{r}^k\}$生成$k$不同的重新排序来实现这一点，这些重新排序指导每个相应$y$ 的生成顺序。每个重新排序都表示为源句子索引的排列
我们的方法以序列到序列模型为中心，该模型可以生成大致尊重输入令牌特定顺序的释义。形式上，这是一个模型$P(\boldsymbol{y}\mid x,\boldsymbol{r})$ 。首先， 我们假设可以访问目标重排序集$R$ ,并在 2.2节中描述了这种重排感知释义模型(Reap)。然后，在 2.3 节中，我们概述了我们的重新排序方法，包括源顺序重写(Sow)模型，该模型在推理$(x\to R)$期间产生适合给定输入句子$x$的重新排序集
## 2.1 Base Model
本工作中讨论的模型建立在标准的序列到序列转换器模型 (Vaswani et al.,2017)上，该模型
使用堆叠的自关注层来编码输入令牌$x$并解码相应的目标序列$y$ 。这个模型是在图 2 的灰色块中绘制的。在整个工作中，我们使用字节对编码(BPE)(Sennrich et al.,2016)来标记我们的输入和输出句子。这些模型以标准方式训练，使用教师强制最大化目标序列的对数似然。此外，为了确保解码器不会在解码过程的每个步骤中重复处理相同的输入令牌，我们包括一个覆盖损失项，如 See et al.(2017) 中提出的
请注意，由于变压器模型的体系结构是非递归的，因此它将位置嵌人添加到输入单词嵌入中，以指示$x$和$y$中单词的正确顺序(请参见图 2 )。在这项工作中，我们建议使用一组额外的位置嵌入来指示生成期间所需的单词顺序，如下所述
2.2 Rearrangement aware Paraphrasing
Model ( Reap )
让$r=\{r_1,r_2,\ldots,r_n\}$指示与输入令牌$x$相对应的目标重新排序。我们希望模型在生成最终输出释义时大致按照此指定顺序处理令牌。例如，在图 1 的示例中，重新排序指定在生成释义时，模型应在输出中生成与快船相关的内容之前生成与游戏相关的内容。在这种情况下，基于正在应用的重新排列，模型很可能会在其生成中使用钝化，尽管这不是严格执行的
我们的模型$P(\boldsymbol{y}\mid\boldsymbol{x},\boldsymbol{r})$的体系结构如图 2所示。考虑在编码器中具有$M$层和在解码器中具有$N$层的堆叠的编码器-解码器架构。我们通过一组附加的位置嵌人$PE_{\boldsymbol{r}}$使目标重新排序$r$可访问此变压器模型。我们使用正弦函数来构造以下 Vaswani et al. (2017)
令$E_M=$encoder$_M(x)$为编码器的$M^th$ (最后)层的输出。特殊用途的位置嵌人被添加到此层的输出(见图 2): $E=E_M+PE_r$ 。请注意，这些与在输入层添加的标准位置嵌入是分开的；这样的嵌入也用于我们的模型中，以编码源句子的原始顺序。变压器解码器模型在计算注意力时出现在$E$上，并且位置嵌入的存在应鼓励生成遵循所需的顺序$r$ ,同时仍符合解码器语言模型。我们在 4.3 节中的实验表明，这种位置嵌入方法能够成功地指导释义的生成，同时取决于输入句子的语义以及所需的顺序
## 2.3 Sentence Reordering
我们现在概述了生成这些所需重新排序$r$的方法。我们通过在句子的句法抽象的各个级别上使用 Sow 模型预测短语重排来做到这一点。我们将多个这样的短语级重排组合在一起，以获得一组$R$的句子级重排。这是使用自顶向
|
[
[
0.5146484375,
0.07598876953125,
0.880859375,
0.07598876953125,
0.880859375,
0.08905029296875,
0.5146484375,
0.08905029296875
],
[
0.16455078125,
0.08197021484375,
0.2333984375,
0.0843505859375,
0.2333984375,
0.0955810546875,
0.16455078125,
0.09320068359375
],
[
0.231689453125,
0.0885009765625,
0.359375,
0.0885009765625,
0.359375,
0.1015625,
0.231689453125,
0.1015625
],
[
0.18896484375,
0.097412109375,
0.2333984375,
0.097412109375,
0.2333984375,
0.1051025390625,
0.18896484375,
0.1051025390625
],
[
0.5146484375,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10809326171875,
0.5146484375,
0.10809326171875
],
[
0.162841796875,
0.1092529296875,
0.2359619140625,
0.1092529296875,
0.2359619140625,
0.12054443359375,
0.162841796875,
0.12054443359375
],
[
0.2435302734375,
0.1092529296875,
0.364501953125,
0.1092529296875,
0.364501953125,
0.122314453125,
0.2435302734375,
0.122314453125
],
[
0.5146484375,
0.1092529296875,
0.880859375,
0.1092529296875,
0.880859375,
0.122314453125,
0.5146484375,
0.122314453125
],
[
0.385498046875,
0.1175537109375,
0.46337890625,
0.1175537109375,
0.46337890625,
0.12890625,
0.385498046875,
0.12890625
],
[
0.301513671875,
0.12353515625,
0.30908203125,
0.12353515625,
0.30908203125,
0.1270751953125,
0.301513671875,
0.1270751953125
],
[
0.2451171875,
0.1312255859375,
0.36279296875,
0.1312255859375,
0.36279296875,
0.141357421875,
0.2451171875,
0.141357421875
],
[
0.515625,
0.1258544921875,
0.8818359375,
0.1258544921875,
0.8818359375,
0.1395263671875,
0.515625,
0.1395263671875
],
[
0.133544921875,
0.1318359375,
0.235107421875,
0.1318359375,
0.235107421875,
0.142578125,
0.133544921875,
0.142578125
],
[
0.393798828125,
0.133056640625,
0.46337890625,
0.133056640625,
0.46337890625,
0.144287109375,
0.393798828125,
0.144287109375
],
[
0.515625,
0.14013671875,
0.880859375,
0.14013671875,
0.880859375,
0.15380859375,
0.515625,
0.15380859375
],
[
0.27783203125,
0.1561279296875,
0.32666015625,
0.1561279296875,
0.32666015625,
0.16748046875,
0.27783203125,
0.16748046875
],
[
0.397216796875,
0.1561279296875,
0.459228515625,
0.1561279296875,
0.459228515625,
0.16748046875,
0.397216796875,
0.16748046875
],
[
0.5146484375,
0.1561279296875,
0.8818359375,
0.15673828125,
0.8818359375,
0.171630859375,
0.5146484375,
0.17041015625
],
[
0.5146484375,
0.1728515625,
0.880859375,
0.1728515625,
0.880859375,
0.1859130859375,
0.5146484375,
0.1859130859375
],
[
0.1561279296875,
0.179931640625,
0.2301025390625,
0.1817626953125,
0.2301025390625,
0.193603515625,
0.1561279296875,
0.191162109375
],
[
0.2435302734375,
0.179931640625,
0.358642578125,
0.18115234375,
0.358642578125,
0.1923828125,
0.2435302734375,
0.191162109375
],
[
0.1368408203125,
0.191162109375,
0.231689453125,
0.1923828125,
0.231689453125,
0.2037353515625,
0.1368408203125,
0.203125
],
[
0.515625,
0.189453125,
0.88330078125,
0.189453125,
0.88330078125,
0.203125,
0.515625,
0.203125
],
[
0.25439453125,
0.1953125,
0.345947265625,
0.1953125,
0.345947265625,
0.1995849609375,
0.25439453125,
0.1995849609375
],
[
0.515625,
0.204833984375,
0.88330078125,
0.204833984375,
0.88330078125,
0.218505859375,
0.515625,
0.218505859375
],
[
0.15869140625,
0.21435546875,
0.235107421875,
0.215576171875,
0.235107421875,
0.226806640625,
0.15869140625,
0.2257080078125
],
[
0.2509765625,
0.2161865234375,
0.358642578125,
0.2161865234375,
0.358642578125,
0.226806640625,
0.2509765625,
0.226806640625
],
[
0.39208984375,
0.21435546875,
0.462646484375,
0.2161865234375,
0.462646484375,
0.22802734375,
0.39208984375,
0.2257080078125
],
[
0.515625,
0.22265625,
0.57080078125,
0.22265625,
0.57080078125,
0.2340087890625,
0.515625,
0.2340087890625
],
[
0.5341796875,
0.2381591796875,
0.88330078125,
0.2381591796875,
0.88330078125,
0.251708984375,
0.5341796875,
0.251708984375
],
[
0.118408203125,
0.24462890625,
0.478515625,
0.24462890625,
0.478515625,
0.25830078125,
0.118408203125,
0.25830078125
],
[
0.474365234375,
0.2493896484375,
0.480224609375,
0.2493896484375,
0.480224609375,
0.254638671875,
0.474365234375,
0.254638671875
],
[
0.118408203125,
0.2587890625,
0.48779296875,
0.2587890625,
0.48779296875,
0.2724609375,
0.118408203125,
0.2724609375
],
[
0.515625,
0.254150390625,
0.8818359375,
0.254150390625,
0.8818359375,
0.267333984375,
0.515625,
0.267333984375
],
[
0.1175537109375,
0.271484375,
0.483642578125,
0.271484375,
0.483642578125,
0.28515625,
0.1175537109375,
0.28515625
],
[
0.515625,
0.270751953125,
0.8818359375,
0.270751953125,
0.8818359375,
0.283935546875,
0.515625,
0.283935546875
],
[
0.12005615234375,
0.287353515625,
0.489501953125,
0.287353515625,
0.489501953125,
0.300537109375,
0.12005615234375,
0.300537109375
],
[
0.515625,
0.2861328125,
0.8818359375,
0.2861328125,
0.8818359375,
0.2998046875,
0.515625,
0.2998046875
],
[
0.12005615234375,
0.3017578125,
0.48779296875,
0.3017578125,
0.48779296875,
0.3154296875,
0.12005615234375,
0.3154296875
],
[
0.5146484375,
0.300537109375,
0.880859375,
0.300537109375,
0.880859375,
0.314208984375,
0.5146484375,
0.314208984375
],
[
0.12005615234375,
0.317626953125,
0.48779296875,
0.317626953125,
0.48779296875,
0.327880859375,
0.12005615234375,
0.327880859375
],
[
0.515625,
0.319580078125,
0.62646484375,
0.319580078125,
0.62646484375,
0.330810546875,
0.515625,
0.330810546875
],
[
0.12005615234375,
0.330078125,
0.411376953125,
0.330078125,
0.411376953125,
0.34326171875,
0.12005615234375,
0.34326171875
],
[
0.51123046875,
0.34326171875,
0.8505859375,
0.345703125,
0.8505859375,
0.35986328125,
0.51123046875,
0.358154296875
],
[
0.55322265625,
0.359375,
0.6826171875,
0.361083984375,
0.6826171875,
0.3759765625,
0.55322265625,
0.37353515625
],
[
0.12005615234375,
0.371826171875,
0.156982421875,
0.371826171875,
0.156982421875,
0.38427734375,
0.12005615234375,
0.38427734375
],
[
0.51318359375,
0.38134765625,
0.88330078125,
0.38134765625,
0.88330078125,
0.3984375,
0.51318359375,
0.3984375
],
[
0.1351318359375,
0.38720703125,
0.490234375,
0.385986328125,
0.490234375,
0.40087890625,
0.1351318359375,
0.402099609375
],
[
0.515625,
0.39794921875,
0.88330078125,
0.39794921875,
0.88330078125,
0.410888671875,
0.515625,
0.410888671875
],
[
0.12005615234375,
0.40380859375,
0.48779296875,
0.40380859375,
0.48779296875,
0.41748046875,
0.12005615234375,
0.41748046875
],
[
0.12005615234375,
0.419189453125,
0.48779296875,
0.419189453125,
0.48779296875,
0.432861328125,
0.12005615234375,
0.432861328125
],
[
0.515625,
0.41455078125,
0.88330078125,
0.41455078125,
0.88330078125,
0.427490234375,
0.515625,
0.427490234375
],
[
0.5146484375,
0.429931640625,
0.880859375,
0.429931640625,
0.880859375,
0.443603515625,
0.5146484375,
0.443603515625
],
[
0.12176513671875,
0.435791015625,
0.489501953125,
0.435791015625,
0.489501953125,
0.449462890625,
0.12176513671875,
0.449462890625
],
[
0.515625,
0.4453125,
0.8818359375,
0.4453125,
0.8818359375,
0.458984375,
0.515625,
0.458984375
],
[
0.12005615234375,
0.453125,
0.48779296875,
0.453125,
0.48779296875,
0.466064453125,
0.12005615234375,
0.466064453125
],
[
0.515625,
0.4619140625,
0.8818359375,
0.4619140625,
0.8818359375,
0.4755859375,
0.515625,
0.4755859375
],
[
0.118408203125,
0.468505859375,
0.486083984375,
0.468505859375,
0.486083984375,
0.482177734375,
0.118408203125,
0.482177734375
],
[
0.51904296875,
0.47802734375,
0.8818359375,
0.47802734375,
0.8818359375,
0.4912109375,
0.51904296875,
0.4912109375
],
[
0.12176513671875,
0.485107421875,
0.48779296875,
0.485107421875,
0.48779296875,
0.498779296875,
0.12176513671875,
0.498779296875
],
[
0.12176513671875,
0.50048828125,
0.48779296875,
0.50048828125,
0.48779296875,
0.51416015625,
0.12176513671875,
0.51416015625
],
[
0.5146484375,
0.495361328125,
0.87890625,
0.495361328125,
0.87890625,
0.505859375,
0.5146484375,
0.505859375
],
[
0.5322265625,
0.51025390625,
0.884765625,
0.51025390625,
0.884765625,
0.52490234375,
0.5322265625,
0.52490234375
],
[
0.1234130859375,
0.5185546875,
0.47021484375,
0.5185546875,
0.47021484375,
0.529296875,
0.1234130859375,
0.529296875
],
[
0.118408203125,
0.5322265625,
0.160400390625,
0.5322265625,
0.160400390625,
0.54736328125,
0.118408203125,
0.54736328125
],
[
0.515625,
0.52685546875,
0.8818359375,
0.52685546875,
0.8818359375,
0.54052734375,
0.515625,
0.54052734375
],
[
0.51708984375,
0.54345703125,
0.88330078125,
0.54345703125,
0.88330078125,
0.55712890625,
0.51708984375,
0.55712890625
],
[
0.118408203125,
0.56103515625,
0.225830078125,
0.56103515625,
0.225830078125,
0.576171875,
0.118408203125,
0.576171875
],
[
0.5146484375,
0.55810546875,
0.880859375,
0.55810546875,
0.880859375,
0.57177734375,
0.5146484375,
0.57177734375
],
[
0.51318359375,
0.57373046875,
0.880859375,
0.57373046875,
0.880859375,
0.58740234375,
0.51318359375,
0.58740234375
],
[
0.12005615234375,
0.58837890625,
0.489501953125,
0.58837890625,
0.489501953125,
0.6025390625,
0.12005615234375,
0.6025390625
],
[
0.51318359375,
0.59033203125,
0.80029296875,
0.59033203125,
0.80029296875,
0.60400390625,
0.51318359375,
0.60400390625
],
[
0.12176513671875,
0.60400390625,
0.48779296875,
0.60400390625,
0.48779296875,
0.61767578125,
0.12176513671875,
0.61767578125
],
[
0.5322265625,
0.60693359375,
0.8818359375,
0.60693359375,
0.8818359375,
0.62060546875,
0.5322265625,
0.62060546875
],
[
0.12005615234375,
0.619140625,
0.48779296875,
0.6181640625,
0.48779296875,
0.6328125,
0.12005615234375,
0.63427734375
],
[
0.5146484375,
0.62158203125,
0.880859375,
0.62353515625,
0.880859375,
0.63720703125,
0.5146484375,
0.63525390625
],
[
0.118408203125,
0.63427734375,
0.48779296875,
0.63427734375,
0.48779296875,
0.64794921875,
0.118408203125,
0.64794921875
],
[
0.515625,
0.6396484375,
0.8818359375,
0.6396484375,
0.8818359375,
0.65283203125,
0.515625,
0.65283203125
],
[
0.118408203125,
0.65087890625,
0.489501953125,
0.65185546875,
0.489501953125,
0.66650390625,
0.118408203125,
0.6650390625
],
[
0.51318359375,
0.65380859375,
0.880859375,
0.65380859375,
0.880859375,
0.66748046875,
0.51318359375,
0.66748046875
],
[
0.118408203125,
0.66748046875,
0.486083984375,
0.66748046875,
0.486083984375,
0.68115234375,
0.118408203125,
0.68115234375
],
[
0.515625,
0.67138671875,
0.8818359375,
0.67138671875,
0.8818359375,
0.68603515625,
0.515625,
0.68603515625
],
[
0.12176513671875,
0.68603515625,
0.193115234375,
0.68603515625,
0.193115234375,
0.69775390625,
0.12176513671875,
0.69775390625
],
[
0.51708984375,
0.689453125,
0.880859375,
0.689453125,
0.880859375,
0.69970703125,
0.51708984375,
0.69970703125
],
[
0.1376953125,
0.70166015625,
0.48779296875,
0.70166015625,
0.48779296875,
0.71533203125,
0.1376953125,
0.71533203125
],
[
0.515625,
0.70361328125,
0.8818359375,
0.70361328125,
0.8818359375,
0.71728515625,
0.515625,
0.71728515625
],
[
0.12005615234375,
0.71728515625,
0.48779296875,
0.71728515625,
0.48779296875,
0.73095703125,
0.12005615234375,
0.73095703125
],
[
0.51318359375,
0.71875,
0.880859375,
0.71875,
0.880859375,
0.732421875,
0.51318359375,
0.732421875
],
[
0.12176513671875,
0.732421875,
0.489501953125,
0.73291015625,
0.489501953125,
0.74755859375,
0.12176513671875,
0.74658203125
],
[
0.515625,
0.736328125,
0.8818359375,
0.736328125,
0.8818359375,
0.75,
0.515625,
0.75
],
[
0.12176513671875,
0.75,
0.489501953125,
0.75,
0.489501953125,
0.76318359375,
0.12176513671875,
0.76318359375
],
[
0.515625,
0.751953125,
0.8818359375,
0.751953125,
0.8818359375,
0.765625,
0.515625,
0.765625
],
[
0.12176513671875,
0.765625,
0.489501953125,
0.765625,
0.489501953125,
0.779296875,
0.12176513671875,
0.779296875
],
[
0.51318359375,
0.7666015625,
0.8818359375,
0.76708984375,
0.8818359375,
0.7822265625,
0.51318359375,
0.78076171875
],
[
0.12005615234375,
0.7822265625,
0.48779296875,
0.7822265625,
0.48779296875,
0.7958984375,
0.12005615234375,
0.7958984375
],
[
0.51318359375,
0.78515625,
0.55517578125,
0.78515625,
0.55517578125,
0.798828125,
0.51318359375,
0.798828125
],
[
0.12176513671875,
0.79736328125,
0.48779296875,
0.79736328125,
0.48779296875,
0.81103515625,
0.12176513671875,
0.81103515625
],
[
0.5146484375,
0.81103515625,
0.56494140625,
0.81103515625,
0.56494140625,
0.822265625,
0.5146484375,
0.822265625
],
[
0.55908203125,
0.80859375,
0.72900390625,
0.81103515625,
0.72900390625,
0.826171875,
0.55908203125,
0.82373046875
],
[
0.118408203125,
0.82763671875,
0.273681640625,
0.830078125,
0.273681640625,
0.84423828125,
0.118408203125,
0.8427734375
],
[
0.118408203125,
0.8125,
0.489501953125,
0.8134765625,
0.489501953125,
0.82763671875,
0.118408203125,
0.82666015625
],
[
0.515625,
0.83056640625,
0.8818359375,
0.83056640625,
0.8818359375,
0.845703125,
0.515625,
0.845703125
],
[
0.515625,
0.84716796875,
0.8818359375,
0.84716796875,
0.8818359375,
0.86083984375,
0.515625,
0.86083984375
],
[
0.12176513671875,
0.85791015625,
0.1695556640625,
0.85791015625,
0.1695556640625,
0.869140625,
0.12176513671875,
0.869140625
],
[
0.162841796875,
0.85791015625,
0.26123046875,
0.85791015625,
0.26123046875,
0.869140625,
0.162841796875,
0.869140625
],
[
0.515625,
0.8623046875,
0.88330078125,
0.86328125,
0.88330078125,
0.87744140625,
0.515625,
0.876953125
],
[
0.12005615234375,
0.87744140625,
0.48779296875,
0.87890625,
0.48779296875,
0.8935546875,
0.12005615234375,
0.892578125
],
[
0.5146484375,
0.8798828125,
0.8818359375,
0.8798828125,
0.8818359375,
0.8935546875,
0.5146484375,
0.8935546875
],
[
0.118408203125,
0.8955078125,
0.486083984375,
0.8955078125,
0.486083984375,
0.9091796875,
0.118408203125,
0.9091796875
],
[
0.5146484375,
0.8955078125,
0.8818359375,
0.8955078125,
0.8818359375,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"使用堆叠的自关注层来编码输入令牌&并解",
"NewEncoder",
"000O",
"OutputE",
"码相应的目标序列y。这个模型是在图2的",
"Target Orderr",
"0082",
"灰色块中绘制的。在整个工作中，我们使用字",
"Output tokens y",
"",
"0",
"节对编码（BPE）（Sennrichetal.，2016）来标",
"Encoder OutputEM",
"Thegame",
"记我们的输入和输出句子。这些模型以标准方",
"Encoder",
"Decoder",
"式训练，使用教师强制最大化目标序列的对数",
"似然。此外，为了确保解码器不会在解码过程",
"Original Order",
"D3",
"Token embeddings",
"的每个步骤中重复处理相同的输人令牌，我们",
"",
"包括一个覆盖损失项，如Seeetal.（2017）中",
"Input tokens x",
"Clippers won the game",
"Bos)(The",
"提出的",
"请注意，由于变压器模型的体系结构是非递",
"Figure 2: Rearrangement aware paraphrasing",
"5",
"Reap ) model.The gray area corresponds to",
"归的，因此它将位置嵌入添加到输人单词嵌入",
"the standard transformer encoder-decoder system",
"中，以指示&和y中单词的正确顺序（请参",
"Our model adds position embeddings correspond-",
"见图2）。在这项工作中，我们建议使用一组",
"ing to the target reordering to encoder outputs.",
"额外的位置嵌入来指示生成期间所需的单词顺",
"The decoder attends over these augmented encod",
"序，如下所述",
"ings during both training and inference.",
"2.2Rearrangement aware Paraphrasing",
"Model (Reap)",
"排序",
"让r={ri,r2,...,rn)指示与输入令牌相",
"我们在大规模英语释义数据集ParaNMT-",
"对应的目标重新排序。我们希望模型在生成最",
"50M（WietingandGimpel,2018）上训练和评",
"估我们的方法。结果表明，我们的方法产生了",
"终输出释义时大致按照此指定顺序处理令牌。",
"例如，在图1的示例中，重新排序指定在生",
"更多不同的释义，同时保留了强基线模型所表",
"成释义时，模型应在输出中生成与快船相关的",
"现出的质量。我们进一步证明了所提出的基于",
"内容之前生成与游戏相关的内容。在这种情况",
"语法的转换过程为输入句子生成了一组可行的",
"下，基于正在应用的重新排列，模型很可能会",
"重排。最后，我们表明位置嵌人提供了一种简",
"单而有效的方式来编码重新排序信息，并且生",
"在其生成中使用钝化，尽管这不是严格执行的",
"我们的模型P(y「c,r）的体系结构如图2",
"成的释义与所需的重新排序输入具有很高的",
"致性",
"所示。考虑在编码器中具有M层和在解码器",
"中具有N层的堆叠的编码器-解码器架构。我",
"2 Method",
"们通过一组附加的位置嵌入PE使目标重新",
"排序可访问此变压器模型。我们使用正弦函",
"给定一个输人句子=【α1,c2，，an］，我",
"数来构造以下Vaswanietal.（2017）",
"们的目标是生成一组结构上不同的释义Y=",
"令EM=encoderM(α）为编码器的Mth（最",
"{yl,y²,...,yh}。我们通过首先为输人句子",
"后）层的输出。特殊用途的位置嵌入被添加到",
"R =rl,r2,...,rk1生成 k 不同的重新排序",
"此层的输出（见图2）：E=EM+PEr。请",
"来实现这一点，这些重新排序指导每个相应9",
"注意，这些与在输入层添加的标准位置嵌入是",
"的生成顺序。每个重新排序都表示为源句子索",
"分开的；这样的嵌入也用于我们的模型中，以",
"引的排列",
"编码源句子的原始顺序。变压器解码器模型在",
"我们的方法以序列到序列模型为中心，该模",
"计算注意力时出现在E上，并且位置嵌人的",
"型可以生成大致尊重输人令牌特定顺序的释",
"存在应鼓励生成遵循所需的顺序，同时仍符",
"义。形式上，这是一个模型P(yla,r)。首先，",
"合解码器语言模型。我们在4.3节中的实验表",
"我们假设可以访问目标重排序集R，并在2.2",
"明，这种位置嵌入方法能够成功地指导释义的",
"节中描述了这种重排感知释义模型（Reap）。",
"生成，同时取决于输入句子的语义以及所需的",
"然后，在2.3节中，我们概述了我们的重新排",
"顺序",
"序方法，包括源顺序重写（Sow）模型，该模",
"2.3",
"Sentence Reordering",
"子的重新排序集",
"型在推理（αc→R）期间产生适合给定输入句",
"我们现在概述了生成这些所需重新排序的",
"方法。我们通过在句子的句法抽象的各个级别",
"2.1",
"Base Model",
"上使用Sow模型预测短语重排来做到这一点。",
"本工作中讨论的模型建立在标准的序列到序列",
"我们将多个这样的短语级重排组合在一起，以",
"转换器模型（Vaswanietal.，2017）上，该模型",
"获得一组R的句子级重排。这是使用自顶向"
] |
{
"arxiv_id": "2005.02013",
"language": "zh",
"page_idx": 1
}
|
||
2005.02013-zh-2
|
2005.02013-zh-2.jpg
|
<FigureHere>
Figure 3: Overview of the source sentence rearrangement workflow for one level of recursion at the root node. First, candidate tree segment pairs contained within the input node are selected. A transduction operation is applied over the abstracted phrase, giving the reordering 4 5 1 2 3 for the case shown in red, then the process recursively continues for each abstracted node. This results in a reordering for the full source sentence; the reordering indices serve as additional input to the Reap model.
下的方法完成的，从解析树的根节点开始。算
法 1 中概述了整个递归过程。
得每个抽象短语的重排(步骤 2)。最后，这种顶级重排与抽象短语中递归构造的短语重排相结合，以获得句子级重排(步骤 3)
2.3.1
Step 1: SelectSegmentPairs 我们首先选择构成 seq2seq 模型输入的短语元组。短语元组$(t,A,B)$由子树$t$组成，其中提取了成分$A$和$B$ (由它们的语法类别代替)。例如，在图 3中，以红色圈出的 S $_0$, S 和 VP $_2$节点形成一个短语元组。$A$和$B$的多个不同组合是可能的。2
<table>
<tbody>
<tr>
<td>$\operatorname{lgorithm}1$Reorder$(t)$</td>
<td> </td>
</tr>
<tr>
<td>$\mathcal{T}=$SelectSegmentPairs $(t)$ $\mathcal{R}=$InitializeBeam (size$=k)$</td>
<td>/// $\operatorname{Step}1$</td>
</tr>
<tr>
<td>e$(z,r_{a},r_{b})$ $r=$Combine $\operatorname{score}(r)$ = $\operatorname { score} ( z)$ + $\Re$pelgb$(r$,score$(r))$ end for end for return $\mathcal{R}$ </td>
<td>/// $\operatorname{Step}3$ $\operatorname{score}(r_a)$</td>
</tr>
</tbody>
</table>
# 2.3.2
Step 2: ReorderPhrase 接下来，我们获得每个短语元组$(t,A,B)$的重新排列。我们首先形成一个输入，该输入由$t$的产量组成，其中$A$和$B$被抽象出来；e。g。IfSIwillVP,showninredinFigure 3。我们使用序列到序列模型 (Sow 模型),该模型将此字符串作为输入并产生相应的输出序列。然后，我们在输入和生成的输出序列之间执行单词级对齐(使用 GloVe 嵌入之间的余弦相似性),以获得必须应用于输入序列的重排。$^{3}$输出序列的对数概率作为此重排的分数
递归算法的一个步骤有三个主要步骤：图 3显示了一次迭代的总体工作流程(这里，选择句子的根节点进行说明)。首先，我们选择尊重解析树边界的输入短语的子短语对，其中每对由不重叠的短语组成(步骤 1)。由于目的是学习通用语法控制的重排，因此我们抽象出两个子短语，并将它们替换为非终止符号，仅保留组成标签信息。例如，我们在图$\overline{\mathrm{E}}3$ 中显示了三个短语对，可以将其抽象化以产生句子的简化形式。然后，我们使用 seq2seq 模型来获
$^2$In order to limit the number of such pairs, we employ a threshold on the fraction of non-abstracted words remaining in the phrase, outlined in more detail $\begin{array}{c}\text{in the Appendix.}\\\end{array}$
$^{3}$We experimented with a pointer network to predict $\begin{array}{c}\mathrm{indices~directly;~however,~the~approach~of~generate~and}\\\end{array}$ then align post hoc resulted in a much more stable ${\mathrm{model}}.$
|
[
[
0.1318359375,
0.08013916015625,
0.32568359375,
0.08197021484375,
0.32568359375,
0.0955810546875,
0.1318359375,
0.09320068359375
],
[
0.420654296875,
0.0885009765625,
0.450927734375,
0.0885009765625,
0.450927734375,
0.09979248046875,
0.420654296875,
0.09979248046875
],
[
0.5634765625,
0.09027099609375,
0.623046875,
0.09027099609375,
0.623046875,
0.1015625,
0.5634765625,
0.1015625
],
[
0.68603515625,
0.0914306640625,
0.755859375,
0.0914306640625,
0.755859375,
0.1015625,
0.68603515625,
0.1015625
],
[
0.77490234375,
0.0914306640625,
0.873046875,
0.0914306640625,
0.873046875,
0.1015625,
0.77490234375,
0.1015625
],
[
0.387939453125,
0.10272216796875,
0.41064453125,
0.10272216796875,
0.41064453125,
0.1134033203125,
0.387939453125,
0.1134033203125
],
[
0.1485595703125,
0.09320068359375,
0.310546875,
0.09320068359375,
0.310546875,
0.10687255859375,
0.1485595703125,
0.10687255859375
],
[
0.443359375,
0.10394287109375,
0.47607421875,
0.10394287109375,
0.47607421875,
0.11224365234375,
0.443359375,
0.11224365234375
],
[
0.12176513671875,
0.10986328125,
0.3359375,
0.11102294921875,
0.3359375,
0.12469482421875,
0.12176513671875,
0.12353515625
],
[
0.54248046875,
0.10986328125,
0.63037109375,
0.10986328125,
0.63037109375,
0.1175537109375,
0.54248046875,
0.1175537109375
],
[
0.6826171875,
0.10809326171875,
0.75830078125,
0.10809326171875,
0.75830078125,
0.1181640625,
0.6826171875,
0.1181640625
],
[
0.7841796875,
0.10986328125,
0.875,
0.10986328125,
0.875,
0.1217041015625,
0.7841796875,
0.1217041015625
],
[
0.151123046875,
0.119384765625,
0.298828125,
0.1217041015625,
0.298828125,
0.1353759765625,
0.151123046875,
0.133056640625
],
[
0.4609375,
0.12054443359375,
0.483642578125,
0.12054443359375,
0.483642578125,
0.12890625,
0.4609375,
0.12890625
],
[
0.48779296875,
0.1175537109375,
0.515625,
0.12054443359375,
0.51318359375,
0.1312255859375,
0.484375,
0.127685546875
],
[
0.5693359375,
0.1341552734375,
0.62646484375,
0.1341552734375,
0.62646484375,
0.144287109375,
0.5693359375,
0.144287109375
],
[
0.68603515625,
0.1353759765625,
0.74462890625,
0.1353759765625,
0.74462890625,
0.1455078125,
0.68603515625,
0.1455078125
],
[
0.78271484375,
0.13720703125,
0.85205078125,
0.13720703125,
0.85205078125,
0.1484375,
0.78271484375,
0.1484375
],
[
0.5087890625,
0.1395263671875,
0.5224609375,
0.1395263671875,
0.5224609375,
0.149658203125,
0.5087890625,
0.149658203125
],
[
0.1376953125,
0.1436767578125,
0.2418212890625,
0.1436767578125,
0.2418212890625,
0.155029296875,
0.1376953125,
0.155029296875
],
[
0.3896484375,
0.142578125,
0.39794921875,
0.142578125,
0.39794921875,
0.146728515625,
0.3896484375,
0.146728515625
],
[
0.235107421875,
0.1519775390625,
0.301513671875,
0.1519775390625,
0.301513671875,
0.171630859375,
0.235107421875,
0.171630859375
],
[
0.58837890625,
0.1602783203125,
0.615234375,
0.1602783203125,
0.615234375,
0.171630859375,
0.58837890625,
0.171630859375
],
[
0.587890625,
0.171630859375,
0.63623046875,
0.177001953125,
0.6337890625,
0.188232421875,
0.58447265625,
0.1817626953125
],
[
0.62451171875,
0.169189453125,
0.63818359375,
0.169189453125,
0.63818359375,
0.17578125,
0.62451171875,
0.17578125
],
[
0.7080078125,
0.1728515625,
0.8154296875,
0.1728515625,
0.8154296875,
0.18408203125,
0.7080078125,
0.18408203125
],
[
0.387939453125,
0.17578125,
0.5283203125,
0.17578125,
0.5283203125,
0.1859130859375,
0.387939453125,
0.1859130859375
],
[
0.26123046875,
0.17333984375,
0.32568359375,
0.17333984375,
0.32568359375,
0.185302734375,
0.26123046875,
0.185302734375
],
[
0.59619140625,
0.185302734375,
0.62890625,
0.185302734375,
0.62890625,
0.196533203125,
0.59619140625,
0.196533203125
],
[
0.6650390625,
0.1817626953125,
0.8564453125,
0.1817626953125,
0.8564453125,
0.1953125,
0.6650390625,
0.1953125
],
[
0.20654296875,
0.193603515625,
0.268798828125,
0.193603515625,
0.268798828125,
0.20068359375,
0.20654296875,
0.20068359375
],
[
0.372802734375,
0.1859130859375,
0.5400390625,
0.1859130859375,
0.5400390625,
0.196533203125,
0.372802734375,
0.196533203125
],
[
0.297119140625,
0.193603515625,
0.318115234375,
0.193603515625,
0.318115234375,
0.20068359375,
0.297119140625,
0.20068359375
],
[
0.204833984375,
0.20068359375,
0.298828125,
0.20068359375,
0.298828125,
0.2078857421875,
0.204833984375,
0.2078857421875
],
[
0.185546875,
0.215576171875,
0.2509765625,
0.215576171875,
0.2509765625,
0.226806640625,
0.185546875,
0.226806640625
],
[
0.443359375,
0.2120361328125,
0.5439453125,
0.21728515625,
0.54248046875,
0.23095703125,
0.441650390625,
0.2257080078125
],
[
0.3837890625,
0.2197265625,
0.391357421875,
0.223876953125,
0.38623046875,
0.2286376953125,
0.3779296875,
0.223876953125
],
[
0.5859375,
0.2161865234375,
0.84033203125,
0.21728515625,
0.84033203125,
0.2286376953125,
0.5859375,
0.22802734375
],
[
0.441650390625,
0.22802734375,
0.489501953125,
0.22802734375,
0.489501953125,
0.2381591796875,
0.441650390625,
0.2381591796875
],
[
0.5859375,
0.2286376953125,
0.78076171875,
0.2298583984375,
0.78076171875,
0.243408203125,
0.5859375,
0.2423095703125
],
[
0.3525390625,
0.23095703125,
0.397216796875,
0.23095703125,
0.397216796875,
0.2423095703125,
0.3525390625,
0.2423095703125
],
[
0.151123046875,
0.2340087890625,
0.345947265625,
0.2298583984375,
0.345947265625,
0.251708984375,
0.151123046875,
0.255859375
],
[
0.16455078125,
0.2369384765625,
0.189697265625,
0.2369384765625,
0.189697265625,
0.2410888671875,
0.16455078125,
0.2410888671875
],
[
0.339111328125,
0.2392578125,
0.4013671875,
0.240478515625,
0.4013671875,
0.25048828125,
0.339111328125,
0.2493896484375
],
[
0.444091796875,
0.2369384765625,
0.48193359375,
0.2369384765625,
0.48193359375,
0.248779296875,
0.444091796875,
0.248779296875
],
[
0.313232421875,
0.24462890625,
0.32666015625,
0.24462890625,
0.32666015625,
0.2476806640625,
0.313232421875,
0.2476806640625
],
[
0.33154296875,
0.24462890625,
0.339111328125,
0.24462890625,
0.339111328125,
0.248779296875,
0.33154296875,
0.248779296875
],
[
0.118408203125,
0.268310546875,
0.88330078125,
0.268310546875,
0.88330078125,
0.281982421875,
0.118408203125,
0.281982421875
],
[
0.118408203125,
0.283203125,
0.88330078125,
0.283203125,
0.88330078125,
0.29638671875,
0.118408203125,
0.29638671875
],
[
0.118408203125,
0.29638671875,
0.8818359375,
0.29638671875,
0.8818359375,
0.311279296875,
0.118408203125,
0.311279296875
],
[
0.118408203125,
0.311279296875,
0.88330078125,
0.311279296875,
0.88330078125,
0.32470703125,
0.118408203125,
0.32470703125
],
[
0.118408203125,
0.32666015625,
0.73486328125,
0.32666015625,
0.73486328125,
0.33740234375,
0.118408203125,
0.33740234375
],
[
0.1234130859375,
0.365234375,
0.489501953125,
0.365234375,
0.489501953125,
0.37890625,
0.1234130859375,
0.37890625
],
[
0.5146484375,
0.365234375,
0.8818359375,
0.365234375,
0.8818359375,
0.37890625,
0.5146484375,
0.37890625
],
[
0.1175537109375,
0.38134765625,
0.35693359375,
0.380126953125,
0.35693359375,
0.394287109375,
0.1175537109375,
0.3955078125
],
[
0.5146484375,
0.380126953125,
0.8818359375,
0.38134765625,
0.8818359375,
0.3955078125,
0.5146484375,
0.394287109375
],
[
0.5146484375,
0.39794921875,
0.796875,
0.39794921875,
0.796875,
0.410888671875,
0.5146484375,
0.410888671875
],
[
0.118408203125,
0.413330078125,
0.31494140625,
0.41455078125,
0.31494140625,
0.4287109375,
0.118408203125,
0.427490234375
],
[
0.118408203125,
0.432861328125,
0.44580078125,
0.43408203125,
0.44580078125,
0.448974609375,
0.118408203125,
0.44775390625
],
[
0.5146484375,
0.432861328125,
0.55322265625,
0.432861328125,
0.55322265625,
0.4453125,
0.5146484375,
0.4453125
],
[
0.118408203125,
0.45068359375,
0.489501953125,
0.45068359375,
0.489501953125,
0.46435546875,
0.118408203125,
0.46435546875
],
[
0.51318359375,
0.4560546875,
0.8818359375,
0.457275390625,
0.8818359375,
0.471435546875,
0.51318359375,
0.47021484375
],
[
0.5146484375,
0.47265625,
0.8818359375,
0.473876953125,
0.8818359375,
0.488037109375,
0.5146484375,
0.48681640625
],
[
0.1376953125,
0.48388671875,
0.36279296875,
0.48388671875,
0.36279296875,
0.49755859375,
0.1376953125,
0.49755859375
],
[
0.41650390625,
0.48388671875,
0.483642578125,
0.48388671875,
0.483642578125,
0.498779296875,
0.41650390625,
0.498779296875
],
[
0.51318359375,
0.490478515625,
0.88330078125,
0.490478515625,
0.88330078125,
0.50439453125,
0.51318359375,
0.50439453125
],
[
0.1368408203125,
0.49951171875,
0.3779296875,
0.49951171875,
0.3779296875,
0.51318359375,
0.1368408203125,
0.51318359375
],
[
0.5146484375,
0.5048828125,
0.884765625,
0.505859375,
0.884765625,
0.52099609375,
0.5146484375,
0.51953125
],
[
0.1368408203125,
0.515625,
0.301513671875,
0.515625,
0.301513671875,
0.529296875,
0.1368408203125,
0.529296875
],
[
0.5146484375,
0.5224609375,
0.8818359375,
0.5224609375,
0.8818359375,
0.5361328125,
0.5146484375,
0.5361328125
],
[
0.1561279296875,
0.53076171875,
0.382080078125,
0.53076171875,
0.382080078125,
0.5458984375,
0.1561279296875,
0.5458984375
],
[
0.414794921875,
0.5322265625,
0.484375,
0.5322265625,
0.484375,
0.54638671875,
0.414794921875,
0.54638671875
],
[
0.5283203125,
0.53759765625,
0.8818359375,
0.5380859375,
0.8818359375,
0.552734375,
0.5283203125,
0.5517578125
],
[
0.156982421875,
0.54736328125,
0.385498046875,
0.54736328125,
0.385498046875,
0.5625,
0.156982421875,
0.5625
],
[
0.44921875,
0.54736328125,
0.491943359375,
0.54736328125,
0.491943359375,
0.5634765625,
0.44921875,
0.5634765625
],
[
0.5146484375,
0.55419921875,
0.5810546875,
0.5517578125,
0.58251953125,
0.56787109375,
0.515625,
0.5693359375
],
[
0.15283203125,
0.5625,
0.2225341796875,
0.56396484375,
0.2225341796875,
0.5791015625,
0.15283203125,
0.57666015625
],
[
0.1561279296875,
0.57666015625,
0.385498046875,
0.57666015625,
0.385498046875,
0.59423828125,
0.1561279296875,
0.59423828125
],
[
0.44921875,
0.580078125,
0.490234375,
0.580078125,
0.490234375,
0.595703125,
0.44921875,
0.595703125
],
[
0.51318359375,
0.58837890625,
0.5556640625,
0.58837890625,
0.5556640625,
0.60400390625,
0.51318359375,
0.60400390625
],
[
0.15283203125,
0.59423828125,
0.2225341796875,
0.5966796875,
0.2225341796875,
0.61083984375,
0.15283203125,
0.609375
],
[
0.1561279296875,
0.61083984375,
0.365234375,
0.61083984375,
0.365234375,
0.6259765625,
0.1561279296875,
0.6259765625
],
[
0.51318359375,
0.6142578125,
0.880859375,
0.6142578125,
0.880859375,
0.62744140625,
0.51318359375,
0.62744140625
],
[
0.172119140625,
0.626953125,
0.359375,
0.62744140625,
0.359375,
0.642578125,
0.172119140625,
0.64111328125
],
[
0.41650390625,
0.62744140625,
0.486083984375,
0.62744140625,
0.486083984375,
0.642578125,
0.41650390625,
0.642578125
],
[
0.51318359375,
0.6298828125,
0.8818359375,
0.63134765625,
0.8818359375,
0.6455078125,
0.51318359375,
0.64453125
],
[
0.173828125,
0.6435546875,
0.2435302734375,
0.6435546875,
0.2435302734375,
0.65771484375,
0.173828125,
0.65771484375
],
[
0.2392578125,
0.6455078125,
0.2939453125,
0.6455078125,
0.2939453125,
0.65625,
0.2392578125,
0.65625
],
[
0.287109375,
0.64453125,
0.491943359375,
0.64453125,
0.491943359375,
0.65771484375,
0.287109375,
0.65771484375
],
[
0.51708984375,
0.64794921875,
0.8818359375,
0.64794921875,
0.8818359375,
0.6611328125,
0.51708984375,
0.6611328125
],
[
0.17041015625,
0.6591796875,
0.340087890625,
0.6611328125,
0.340087890625,
0.6787109375,
0.17041015625,
0.6767578125
],
[
0.515625,
0.6630859375,
0.88330078125,
0.6630859375,
0.88330078125,
0.677734375,
0.515625,
0.677734375
],
[
0.154541015625,
0.6787109375,
0.2149658203125,
0.6787109375,
0.2149658203125,
0.693359375,
0.154541015625,
0.693359375
],
[
0.51318359375,
0.677734375,
0.8818359375,
0.6787109375,
0.8818359375,
0.693359375,
0.51318359375,
0.6923828125
],
[
0.1351318359375,
0.69482421875,
0.198974609375,
0.69482421875,
0.198974609375,
0.708984375,
0.1351318359375,
0.708984375
],
[
0.5146484375,
0.69482421875,
0.880859375,
0.69482421875,
0.880859375,
0.7080078125,
0.5146484375,
0.7080078125
],
[
0.133544921875,
0.708984375,
0.2216796875,
0.708984375,
0.2216796875,
0.72705078125,
0.133544921875,
0.72705078125
],
[
0.515625,
0.71142578125,
0.8818359375,
0.71142578125,
0.8818359375,
0.724609375,
0.515625,
0.724609375
],
[
0.515625,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7412109375,
0.515625,
0.7412109375
],
[
0.5146484375,
0.7421875,
0.880859375,
0.7421875,
0.880859375,
0.755859375,
0.5146484375,
0.755859375
],
[
0.1368408203125,
0.75,
0.490234375,
0.7490234375,
0.490234375,
0.76318359375,
0.1368408203125,
0.76416015625
],
[
0.515625,
0.7587890625,
0.88330078125,
0.7587890625,
0.88330078125,
0.77392578125,
0.515625,
0.77392578125
],
[
0.12176513671875,
0.7666015625,
0.48779296875,
0.7666015625,
0.48779296875,
0.77978515625,
0.12176513671875,
0.77978515625
],
[
0.118408203125,
0.7822265625,
0.48779296875,
0.78076171875,
0.48779296875,
0.7958984375,
0.118408203125,
0.79638671875
],
[
0.5146484375,
0.77685546875,
0.755859375,
0.77685546875,
0.755859375,
0.79052734375,
0.5146484375,
0.79052734375
],
[
0.12176513671875,
0.798828125,
0.48779296875,
0.798828125,
0.48779296875,
0.8125,
0.12176513671875,
0.8125
],
[
0.118408203125,
0.81396484375,
0.486083984375,
0.81396484375,
0.486083984375,
0.82763671875,
0.118408203125,
0.82763671875
],
[
0.53076171875,
0.81005859375,
0.884765625,
0.8125,
0.884765625,
0.826171875,
0.53076171875,
0.82373046875
],
[
0.12005615234375,
0.83056640625,
0.489501953125,
0.83056640625,
0.489501953125,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.51318359375,
0.826171875,
0.8818359375,
0.826171875,
0.8818359375,
0.8359375,
0.51318359375,
0.8359375
],
[
0.51123046875,
0.8359375,
0.88330078125,
0.8359375,
0.88330078125,
0.849609375,
0.51123046875,
0.849609375
],
[
0.12176513671875,
0.84716796875,
0.489501953125,
0.84716796875,
0.489501953125,
0.86083984375,
0.12176513671875,
0.86083984375
],
[
0.51123046875,
0.84716796875,
0.62890625,
0.8486328125,
0.62890625,
0.85986328125,
0.51123046875,
0.859375
],
[
0.118408203125,
0.86328125,
0.486083984375,
0.86328125,
0.486083984375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.53076171875,
0.85791015625,
0.884765625,
0.85986328125,
0.884765625,
0.87451171875,
0.53076171875,
0.87255859375
],
[
0.51318359375,
0.87158203125,
0.884765625,
0.87158203125,
0.884765625,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.1234130859375,
0.87744140625,
0.48779296875,
0.87744140625,
0.48779296875,
0.892578125,
0.1234130859375,
0.892578125
],
[
0.51318359375,
0.88427734375,
0.88671875,
0.88427734375,
0.88671875,
0.8984375,
0.51318359375,
0.8984375
],
[
0.12005615234375,
0.89453125,
0.489501953125,
0.89453125,
0.489501953125,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.51318359375,
0.896484375,
0.5615234375,
0.896484375,
0.5615234375,
0.90771484375,
0.51318359375,
0.90771484375
]
] |
[
"REORDER(SO): Recursively reorder",
"S",
"Sow Input",
"SowOutput",
"Source reordering",
"SBAR",
"constituents to get final ordering",
"PRPVP",
"FOR A,B E SELECTSEGMENTPAIRS(SO)",
"SBARIwillcarryNI",
"SBAR NPIcarry",
"SBARIwll carryNP",
"REORDERPHRASE (SO,A,B)",
"MD",
"(VP)",
"IfSIwill VP",
"IwillVPifs",
"SIwilVP",
"NP",
"SBAR I willcarry NP",
"OR",
"HSIIYP",
"",
"",
"中",
"REORDERPHRASE:",
"SELECTSEGMENTPAIRS:",
"REORDER(A)",
"Sow",
"Use seq2seq model to reorder phrase",
"If SIwill carry",
"Choose constituents to abstract",
"brella",
"671234",
"REORDER(B)",
"Final parar",
"",
"Ifit continuestorainan umbrellais whatiwill carry",
"#88",
"Iwillcarryanumbrellaifrain continues",
"Derived",
"on",
"itror",
"reorderings",
"REAP",
"",
"",
"Figure 3: Overview of the source sentence rearrangement workflow for one level of recursion at the root",
"node. First, candidate tree segment pairs contained within the input node are selected. A transduction",
"operation is applied over the abstracted phrase, giving the reordering 4 5 1 2 3 for the case shown in red,",
"then the process recursively continues for each abstracted node. This results in a reordering for the full",
"source sentence; the reordering indices serve as additional input to the Reap model.",
"下的方法完成的，从解析树的根节点开始。算",
"得每个抽象短语的重排（步骤2）。最后，这种",
"法1中概述了整个递归过程。",
"顶级重排与抽象短语中递归构造的短语重排相",
"结合，以获得句子级重排（步骤3）",
"Algorithm 1 Reorder (t)",
"Input: Sub-tree t of the input parse tree",
"2.3.1",
"Output: Top-k list of reorderings for t 's yield",
"Step1：SelectSegmentPairs我们首先选择构",
"成seq2seq模型输人的短语元组。短语元组",
"T = SelectSegmentPairs (t)",
"/ / Step 1",
"(t,A,B)由子树t组成，其中提取了成分A和",
"R = InitializeBeam (size = k)",
"B（由它们的语法类别代替）。例如，在图3",
"for (A,B) in T do",
"中，以红色圈出的So，S和VP2节点形成",
"z = ReorderPhrase (t, A, B)",
"/ / Step 2",
"个短语元组。A和B的多个不同组合是可",
"RA(1, ..., k) = Reorder (tA)",
"//k",
"能的。2",
"orderings",
"RB(1,..., k) = Reorder (tB)",
"//k",
"2.3.2",
"orderings",
"for ra,rb in RA × RB do",
"Step2：ReorderPhrase接下来，我们获得",
"r = Combine (z,ra, rb)",
"// Step 3",
"每个短语元组（t,A,B）的重新排列。我们",
"score(r)",
"",
" score(z) + score(ra) +",
"首先形成一个输入，该输入由t的产量组",
"eorg(sh) (r, score(r)",
"成，其中A和B被抽象出来；e。g。IfSI",
"end for",
"willVP,showninredinFigure3。我们使用序",
"end for",
"列到序列模型（Sow模型），该模型将此字符",
"return R",
"串作为输入并产生相应的输出序列。然后，我",
"们在输入和生成的输出序列之间执行单词级对",
"齐（使用GloVe嵌入之间的余弦相似性），以",
"递归算法的一个步骤有三个主要步骤：图3",
"获得必须应用于输人序列的重排。3输出序列",
"显示了一次送代的总体工作流程（这里，选择",
"句子的根节点进行说明）。首先，我们选择尊",
"的对数概率作为此重排的分数",
"重解析树边界的输人短语的子短语对，其中每",
"对由不重叠的短语组成（步骤1)。由于目的是",
"2In order to limit the number of such pairs, we",
"学习通用语法控制的重排，因此我们抽象出两",
"employathresholdonthefraction of non-abstracted",
"words remaining in the phrase, outlined in more detail",
"个子短语，并将它们替换为非终止符号，仅保",
"in the Appendix.",
"留组成标签信息。例如，我们在图E3中显示",
"3We experimented with a pointer network to predict",
"indices directly; however, the approach of generate and",
"了三个短语对，可以将其抽象化以产生句子的",
"then align post hoc resulted in a much more stable",
"简化形式。然后，我们使用seq2seq模型来获",
"model."
] |
{
"arxiv_id": "2005.02013",
"language": "zh",
"page_idx": 2
}
|
||
2005.02013-zh-3
|
2005.02013-zh-3.jpg
|
Sow model Sow 模型是序列到序列模型$P( y^{\prime }\mid x^{\prime }, o)$ , $\text{遵 循 E2. 1 节 中 的 转 换 器 框 架 。}^4$ $x^{\prime}$和$y^{\prime}$都使用词块词汇进行编码；此外，与POS 标签和组成标签(对于非终端)相对应的嵌入被添加到输入嵌入中。例如，在图 3 中，如果 S 是$(x^\prime,y^{\prime})$的示例 ,则 SI 将 VP 和 I 将 VP 对。虽然不是正式要求，但算法 1 确保在这些序列中总是有两个非末端标签。$o$是一个取值为 Monotone 或 Flip 的变量。这将对首选项进行编码，以使两个抽象节点保持相同的顺序或在输出中“翻转”它们。$^5o$ 在 模 型 中 进 行 编 码 , 并具有用于单调的$\left\{\ldots0,0,1,0,\ldots2,0\ldots\right\}$和用于翻转的$\{\ldots0,0,2,0,\ldots1,0\ldots\}$形式的附加位置编码，其中非零位置对应于短语中抽象的非终端的位置。Sow Model 的这些位置嵌入的处理方式类似于 Reap 模型的$r$嵌入。在推理过程中，我们同时使用单调重排和翻转重排为每个短语元组生成两个重新排序，每种类型一个
我们在 3 节中描述了此模型的训练
2.3.3
Step 3: Combine 上一步给出了子树$t$的重新排列。为了从中获得句子级别的重新排列，我们首先在子树$t_A$和$t_B$上递归应用 Reorder 算法，该算法返回每个子树的前 k 个重新排列。我们迭代每个重排对$(r_a,r_b)$ ,将这些重新排序应用于抽象短语$A$和$B$ 。这在图 3 的左侧进行了说明。通过对所涉及的所有短语级重排取平均值来对由此获得的句子级表示进行评分
# 3 Data and Training
我们在通过回译 CzEng (Bojar et al., 2016) 语料库的捷克语句子构建的 ParaNMT-50M 释义数据集 (Wieting and Gimpel, 2018) 上训练和评估我们的模型。我们过滤此数据集以删除较短的句子(少于8个标记),低质量的释义对 (由数据集中包含的翻译分数量化)和表现出低重新排序的示例(由基于每个单词的位置的重新排序分数量化源及其在目标句子中的对齐单词)。这给我们留下了超过 350k 配对的释义对
## 3.1 Training Data for Reap
为了训练我们的 Reap 模型(在 2.2 节中概述), 我们采用现有的释义对$(x,y^*)$ ,并根据它们与目标句子的对齐方式得出源句子标记的伪地
$\overset{4}{\operatorname*{\mathrm{See~Appendix~for~Sow~model~architecture~diagram.}}}$ $^{5}$In syntactic translation systems, rules similarly can
be divided by whether they preserve order or invert it
(Wu, 1997) .
<FigureHere>
Figure 4: Paraphrase sentence pair and its aligned tuples $A\to B,C$ and $A^\prime\to B^{\prime},C^{\prime}.$ These produce the training data for the Sow Model .
面实况重排$r^*$ 。为了获得这些重新排列，我们首先获得源和目标句子中所有标记的上下文嵌人(Devlin et al.,2019)。我们遵循 Lerner and Petrov (2013) 中概述的策略，并在遍历依赖关系树时执行重新排序。从源句子的根节点开始，我们根据目标句子中相应对齐单词的顺序来确定头部及其子代之间的顺序(独立于其他决策)。我们继续递归遍历以获得句子级别的重新排列。这反映了 2.3 节中的重新排列策略，该策略在选区解析树上而不是依赖关系解析上运行
给定三元组$(x,r^*,y^*)$ ,我们可以训练我们的 Reap 模型，以生成对伪地面真相重新排序的最终释义条件
## 3.2 Training Data for Sow
ParaNMT-50M 数据集包含句子级释义对。但是，为了训练我们的 Sow 模型(在 2.3 节中概述),我们需要在其中看到带有语法抽象的短语级释义。我们使用以下过程从 ParaNMT- 50M 数据集中提取这些内容，如图 4 所示。我们遵循 Zhang et al. (2020) ,并计算句子中所有成分对及其释义之间的短语对齐分数。$^{6}$从这组短语对齐分数中，我们计算了短语之间的部分一对一映射(图$\overline{\mathrm{E}}$4 中的彩色形状);也就是说，并非所有短语都对齐，而是一对一对齐的子集。最后，我们在语法翻译 (Galley et al. 2004)中提取与规则对齐类似的对齐块：当对齐短语$A$和$A^\prime$分别包含对齐短语对$(B,C)$和$(B^{\prime},C^{\prime})$时，我们可以提取对齐元组$(t_A,B,C)$ 和$\left(t_{A^{\prime}},B^{\prime},C^{\prime}\right)$。抽象出短语$\left(B,C\right)$和$\left(B^\prime,C^{\prime}\right)$ ,以构造短语级换能器的训练数据，包括监督$o=$Monotone 或 Flip 。使用上述对齐策略， 我们能够获得超过？? 万个对齐的短语对
### 4 Evaluation
Setup 由于我们的主要目标是评估模型生成各种释义的能力，因此我们获得了一组释义， 并将其与其他方法产生的释义集进行了比较。
${{\frac{6}{\mathrm{The~score~is~computed~using~a~weighted~mean~of}}}}$ the contextual similarity between individual words in the phrases, where the weights are determined by the corpus-level inverse-document frequency of the words. Details in the Appendix.
|
[
[
0.1275634765625,
0.07598876953125,
0.486083984375,
0.07598876953125,
0.486083984375,
0.08905029296875,
0.1275634765625,
0.08905029296875
],
[
0.55322265625,
0.0789794921875,
0.83642578125,
0.08013916015625,
0.83642578125,
0.09442138671875,
0.55322265625,
0.09320068359375
],
[
0.12005615234375,
0.09320068359375,
0.48779296875,
0.09320068359375,
0.48779296875,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.1234130859375,
0.11102294921875,
0.486083984375,
0.11102294921875,
0.486083984375,
0.1217041015625,
0.1234130859375,
0.1217041015625
],
[
0.56689453125,
0.11102294921875,
0.8193359375,
0.11224365234375,
0.8193359375,
0.1270751953125,
0.56689453125,
0.1258544921875
],
[
0.12005615234375,
0.1258544921875,
0.48779296875,
0.1258544921875,
0.48779296875,
0.1395263671875,
0.12005615234375,
0.1395263671875
],
[
0.12176513671875,
0.141357421875,
0.489501953125,
0.141357421875,
0.489501953125,
0.155029296875,
0.12176513671875,
0.155029296875
],
[
0.51318359375,
0.141357421875,
0.8818359375,
0.141357421875,
0.8818359375,
0.155029296875,
0.51318359375,
0.155029296875
],
[
0.118408203125,
0.15673828125,
0.48779296875,
0.1561279296875,
0.48779296875,
0.17041015625,
0.118408203125,
0.171630859375
],
[
0.51123046875,
0.15380859375,
0.8818359375,
0.155029296875,
0.8818359375,
0.169189453125,
0.51123046875,
0.168701171875
],
[
0.118408203125,
0.1728515625,
0.48779296875,
0.171630859375,
0.48779296875,
0.1859130859375,
0.118408203125,
0.18701171875
],
[
0.5146484375,
0.17041015625,
0.79248046875,
0.169189453125,
0.79248046875,
0.18115234375,
0.5146484375,
0.1817626953125
],
[
0.12005615234375,
0.189453125,
0.48779296875,
0.189453125,
0.48779296875,
0.203125,
0.12005615234375,
0.203125
],
[
0.118408203125,
0.2037353515625,
0.489501953125,
0.2060546875,
0.489501953125,
0.2197265625,
0.118408203125,
0.21728515625
],
[
0.515625,
0.210205078125,
0.88330078125,
0.210205078125,
0.88330078125,
0.223876953125,
0.515625,
0.223876953125
],
[
0.12005615234375,
0.221435546875,
0.489501953125,
0.221435546875,
0.489501953125,
0.235107421875,
0.12005615234375,
0.235107421875
],
[
0.51318359375,
0.226806640625,
0.880859375,
0.226806640625,
0.880859375,
0.240478515625,
0.51318359375,
0.240478515625
],
[
0.12005615234375,
0.2381591796875,
0.489501953125,
0.2381591796875,
0.489501953125,
0.251708984375,
0.12005615234375,
0.251708984375
],
[
0.51318359375,
0.2423095703125,
0.880859375,
0.2423095703125,
0.880859375,
0.255859375,
0.51318359375,
0.255859375
],
[
0.12005615234375,
0.254150390625,
0.489501953125,
0.254150390625,
0.489501953125,
0.267333984375,
0.12005615234375,
0.267333984375
],
[
0.515625,
0.260009765625,
0.880859375,
0.260009765625,
0.880859375,
0.270751953125,
0.515625,
0.270751953125
],
[
0.12005615234375,
0.270751953125,
0.48779296875,
0.270751953125,
0.48779296875,
0.283935546875,
0.12005615234375,
0.283935546875
],
[
0.5146484375,
0.27490234375,
0.8818359375,
0.27490234375,
0.8818359375,
0.2880859375,
0.5146484375,
0.2880859375
],
[
0.12005615234375,
0.2861328125,
0.48779296875,
0.2861328125,
0.48779296875,
0.2998046875,
0.12005615234375,
0.2998046875
],
[
0.51318359375,
0.289306640625,
0.880859375,
0.289306640625,
0.880859375,
0.302734375,
0.51318359375,
0.302734375
],
[
0.12005615234375,
0.3017578125,
0.48779296875,
0.3017578125,
0.48779296875,
0.3154296875,
0.12005615234375,
0.3154296875
],
[
0.5146484375,
0.30712890625,
0.88330078125,
0.30712890625,
0.88330078125,
0.320556640625,
0.5146484375,
0.320556640625
],
[
0.12176513671875,
0.319580078125,
0.486083984375,
0.319580078125,
0.486083984375,
0.330078125,
0.12176513671875,
0.330078125
],
[
0.51318359375,
0.32177734375,
0.8818359375,
0.32373046875,
0.8818359375,
0.33740234375,
0.51318359375,
0.3349609375
],
[
0.118408203125,
0.333251953125,
0.486083984375,
0.333251953125,
0.486083984375,
0.3466796875,
0.118408203125,
0.3466796875
],
[
0.5146484375,
0.339111328125,
0.8818359375,
0.339111328125,
0.8818359375,
0.352783203125,
0.5146484375,
0.352783203125
],
[
0.12176513671875,
0.3515625,
0.48779296875,
0.3515625,
0.48779296875,
0.3623046875,
0.12176513671875,
0.3623046875
],
[
0.515625,
0.35693359375,
0.8818359375,
0.35693359375,
0.8818359375,
0.36767578125,
0.515625,
0.36767578125
],
[
0.1234130859375,
0.366455078125,
0.1561279296875,
0.366455078125,
0.1561279296875,
0.37890625,
0.1234130859375,
0.37890625
],
[
0.515625,
0.372314453125,
0.58837890625,
0.372314453125,
0.58837890625,
0.38427734375,
0.515625,
0.38427734375
],
[
0.1376953125,
0.383056640625,
0.418212890625,
0.383056640625,
0.418212890625,
0.396728515625,
0.1376953125,
0.396728515625
],
[
0.5341796875,
0.38720703125,
0.8818359375,
0.38720703125,
0.8818359375,
0.40087890625,
0.5341796875,
0.40087890625
],
[
0.51318359375,
0.402099609375,
0.8818359375,
0.402587890625,
0.8818359375,
0.41748046875,
0.51318359375,
0.416259765625
],
[
0.12005615234375,
0.4091796875,
0.162109375,
0.4091796875,
0.162109375,
0.4228515625,
0.12005615234375,
0.4228515625
],
[
0.515625,
0.419189453125,
0.6455078125,
0.419189453125,
0.6455078125,
0.432861328125,
0.515625,
0.432861328125
],
[
0.12005615234375,
0.427490234375,
0.48779296875,
0.427490234375,
0.48779296875,
0.441162109375,
0.12005615234375,
0.441162109375
],
[
0.12005615234375,
0.443603515625,
0.489501953125,
0.443603515625,
0.489501953125,
0.457275390625,
0.12005615234375,
0.457275390625
],
[
0.5087890625,
0.44482421875,
0.744140625,
0.4453125,
0.744140625,
0.460205078125,
0.5087890625,
0.458984375
],
[
0.118408203125,
0.460205078125,
0.48779296875,
0.458984375,
0.48779296875,
0.473876953125,
0.118408203125,
0.474365234375
],
[
0.5146484375,
0.468505859375,
0.87890625,
0.468505859375,
0.87890625,
0.478515625,
0.5146484375,
0.478515625
],
[
0.12005615234375,
0.476806640625,
0.48779296875,
0.476806640625,
0.48779296875,
0.490478515625,
0.12005615234375,
0.490478515625
],
[
0.515625,
0.482666015625,
0.8818359375,
0.482666015625,
0.8818359375,
0.496337890625,
0.515625,
0.496337890625
],
[
0.12005615234375,
0.4921875,
0.48779296875,
0.4921875,
0.48779296875,
0.505859375,
0.12005615234375,
0.505859375
],
[
0.5146484375,
0.498779296875,
0.8818359375,
0.498779296875,
0.8818359375,
0.5126953125,
0.5146484375,
0.5126953125
],
[
0.118408203125,
0.50830078125,
0.48779296875,
0.50830078125,
0.48779296875,
0.521484375,
0.118408203125,
0.521484375
],
[
0.515625,
0.5166015625,
0.8818359375,
0.5166015625,
0.8818359375,
0.52685546875,
0.515625,
0.52685546875
],
[
0.12005615234375,
0.52490234375,
0.48779296875,
0.52490234375,
0.48779296875,
0.5380859375,
0.12005615234375,
0.5380859375
],
[
0.5146484375,
0.5322265625,
0.88330078125,
0.5322265625,
0.88330078125,
0.5458984375,
0.5146484375,
0.5458984375
],
[
0.12176513671875,
0.54052734375,
0.48779296875,
0.54052734375,
0.48779296875,
0.55419921875,
0.12176513671875,
0.55419921875
],
[
0.5146484375,
0.54638671875,
0.880859375,
0.54638671875,
0.880859375,
0.56005859375,
0.5146484375,
0.56005859375
],
[
0.12176513671875,
0.5546875,
0.15869140625,
0.55712890625,
0.1561279296875,
0.572265625,
0.12005615234375,
0.57080078125
],
[
0.515625,
0.5634765625,
0.8818359375,
0.5634765625,
0.8818359375,
0.57666015625,
0.515625,
0.57666015625
],
[
0.118408203125,
0.58447265625,
0.318115234375,
0.58447265625,
0.318115234375,
0.59716796875,
0.118408203125,
0.59716796875
],
[
0.515625,
0.580078125,
0.8818359375,
0.580078125,
0.8818359375,
0.59326171875,
0.515625,
0.59326171875
],
[
0.5146484375,
0.59423828125,
0.880859375,
0.59423828125,
0.880859375,
0.60791015625,
0.5146484375,
0.60791015625
],
[
0.12005615234375,
0.60986328125,
0.489501953125,
0.60986328125,
0.489501953125,
0.62353515625,
0.12005615234375,
0.62353515625
],
[
0.5146484375,
0.61083984375,
0.8818359375,
0.61083984375,
0.8818359375,
0.62451171875,
0.5146484375,
0.62451171875
],
[
0.118408203125,
0.6259765625,
0.489501953125,
0.62451171875,
0.489501953125,
0.6396484375,
0.118408203125,
0.64013671875
],
[
0.5146484375,
0.62744140625,
0.880859375,
0.62744140625,
0.880859375,
0.64111328125,
0.5146484375,
0.64111328125
],
[
0.12005615234375,
0.64111328125,
0.48779296875,
0.64111328125,
0.48779296875,
0.65478515625,
0.12005615234375,
0.65478515625
],
[
0.5146484375,
0.642578125,
0.880859375,
0.6435546875,
0.880859375,
0.65673828125,
0.5146484375,
0.65625
],
[
0.12176513671875,
0.6591796875,
0.48779296875,
0.6591796875,
0.48779296875,
0.6728515625,
0.12176513671875,
0.6728515625
],
[
0.515625,
0.66015625,
0.8818359375,
0.66015625,
0.8818359375,
0.67333984375,
0.515625,
0.67333984375
],
[
0.12005615234375,
0.6748046875,
0.48779296875,
0.6748046875,
0.48779296875,
0.6884765625,
0.12005615234375,
0.6884765625
],
[
0.5146484375,
0.67578125,
0.88330078125,
0.67578125,
0.88330078125,
0.689453125,
0.5146484375,
0.689453125
],
[
0.12005615234375,
0.69140625,
0.484375,
0.69140625,
0.484375,
0.70166015625,
0.12005615234375,
0.70166015625
],
[
0.5146484375,
0.6923828125,
0.884765625,
0.6923828125,
0.884765625,
0.7060546875,
0.5146484375,
0.7060546875
],
[
0.12005615234375,
0.70703125,
0.48779296875,
0.70703125,
0.48779296875,
0.72021484375,
0.12005615234375,
0.72021484375
],
[
0.515625,
0.71044921875,
0.880859375,
0.71044921875,
0.880859375,
0.72021484375,
0.515625,
0.72021484375
],
[
0.12005615234375,
0.7236328125,
0.484375,
0.7236328125,
0.484375,
0.73388671875,
0.12005615234375,
0.73388671875
],
[
0.5146484375,
0.724609375,
0.88330078125,
0.724609375,
0.88330078125,
0.73828125,
0.5146484375,
0.73828125
],
[
0.12005615234375,
0.7392578125,
0.489501953125,
0.7392578125,
0.489501953125,
0.7529296875,
0.12005615234375,
0.7529296875
],
[
0.495361328125,
0.7451171875,
0.50390625,
0.74365234375,
0.50537109375,
0.74658203125,
0.496337890625,
0.7490234375
],
[
0.515625,
0.7421875,
0.8388671875,
0.7421875,
0.8388671875,
0.7529296875,
0.515625,
0.7529296875
],
[
0.12176513671875,
0.75732421875,
0.156982421875,
0.75732421875,
0.156982421875,
0.7685546875,
0.12176513671875,
0.7685546875
],
[
0.51123046875,
0.7685546875,
0.6455078125,
0.7685546875,
0.6455078125,
0.7822265625,
0.51123046875,
0.7822265625
],
[
0.1158447265625,
0.78076171875,
0.35693359375,
0.7822265625,
0.35693359375,
0.79638671875,
0.1158447265625,
0.7958984375
],
[
0.51904296875,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.8056640625,
0.51904296875,
0.8056640625
],
[
0.1234130859375,
0.8046875,
0.48779296875,
0.8046875,
0.48779296875,
0.8154296875,
0.1234130859375,
0.8154296875
],
[
0.5146484375,
0.80859375,
0.88330078125,
0.81005859375,
0.88330078125,
0.82373046875,
0.5146484375,
0.822265625
],
[
0.12005615234375,
0.8193359375,
0.489501953125,
0.8193359375,
0.489501953125,
0.83447265625,
0.12005615234375,
0.83447265625
],
[
0.5146484375,
0.826171875,
0.88330078125,
0.826171875,
0.88330078125,
0.8388671875,
0.5146484375,
0.8388671875
],
[
0.1234130859375,
0.8359375,
0.48779296875,
0.8359375,
0.48779296875,
0.849609375,
0.1234130859375,
0.849609375
],
[
0.53076171875,
0.8466796875,
0.884765625,
0.8486328125,
0.884765625,
0.86328125,
0.53076171875,
0.86083984375
],
[
0.139404296875,
0.85791015625,
0.489501953125,
0.85986328125,
0.489501953125,
0.87451171875,
0.139404296875,
0.87255859375
],
[
0.51123046875,
0.85986328125,
0.880859375,
0.85986328125,
0.880859375,
0.87353515625,
0.51123046875,
0.87353515625
],
[
0.139404296875,
0.87060546875,
0.48779296875,
0.87158203125,
0.48779296875,
0.8857421875,
0.139404296875,
0.88525390625
],
[
0.51123046875,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.118408203125,
0.8828125,
0.489501953125,
0.88427734375,
0.489501953125,
0.8984375,
0.118408203125,
0.89794921875
],
[
0.5146484375,
0.88427734375,
0.88330078125,
0.88427734375,
0.88330078125,
0.8984375,
0.5146484375,
0.8984375
],
[
0.118408203125,
0.896484375,
0.20068359375,
0.89794921875,
0.20068359375,
0.9091796875,
0.118408203125,
0.90771484375
],
[
0.515625,
0.896484375,
0.6826171875,
0.896484375,
0.6826171875,
0.91015625,
0.515625,
0.91015625
]
] |
[
"SowmodelSow模型是序列到序列模型",
"If it continues to rain I will carry an umbrella",
"P(ylα',o），遵循E2.1节中的转换器框架。4",
"和y都使用词块词汇进行编码；此外，与",
"I will carry an umbrella if rain continues",
"POS标签和组成标签（对于非终端）相对应的",
"嵌入被添加到输人嵌入中。例如，在图3中，如",
"Figure 4: Paraphrase sentence pair and its aligned",
"果S是（α',y'）的示例，则SI将VP和I将VP",
"tuples A → B, C and A' -→ B',C\". These produce",
"对。虽然不是正式要求，但算法1确保在这些",
"the training data for the Sow Model .",
"序列中总是有两个非末端标签。0是一个取值",
"为Monotone或Flip的变量。这将对首选项进",
"面实况重排r*。为了获得这些重新排列，我",
"行编码，以使两个抽象节点保持相同的顺序或",
"们首先获得源和目标句子中所有标记的上下文",
"在输出中“翻转”它们。5o在模型中进行编码，",
"嵌人（Devlinetal.，2019）。我们遵循Lerner",
"并具有用于单调的.0,0,1,0，.2,0.和",
"andPetrov（2013）中概述的策略，并在遍历依",
"用于翻转的..0,0,2,0,·.1,0.·形式的附",
"赖关系树时执行重新排序。从源句子的根节点",
"加位置编码，其中非零位置对应于短语中抽象",
"开始，我们根据目标句子中相应对齐单词的顺",
"的非终端的位置。Sow Model的这些位置嵌人",
"序来确定头部及其子代之间的顺序（独立于其",
"的处理方式类似于Reap模型的r嵌人。在推",
"他决策）。我们继续递归遍历以获得句子级别",
"理过程中，我们同时使用单调重排和翻转重排",
"的重新排列。这反映了2.3节中的重新排列策",
"为每个短语元组生成两个重新排序，每种类型",
"略，该策略在选区解析树上而不是依赖关系解",
"一个",
"析上运行",
"我们在3节中描述了此模型的训练",
"给定三元组（α,r*，y*），我们可以训练我们",
"的Reap模型，以生成对伪地面真相重新排序",
"2.3.3",
"的最终释义条件",
"Step3:Combine上一步给出了子树t的重新",
"排列。为了从中获得句子级别的重新排列，我",
"3.2TrainingData forSow",
"们首先在子树tA和tB上递归应用Reorder",
"ParaNMT-50M数据集包含句子级释义对。但",
"算法，该算法返回每个子树的前k个重新排",
"是，为了训练我们的Sow模型（在2.3节中",
"列。我们迭代每个重排对（ra，rb），将这些重",
"概述），我们需要在其中看到带有语法抽象的",
"新排序应用于抽象短语A和B。这在图3的",
"短语级释义。我们使用以下过程从ParaNMT",
"左侧进行了说明。通过对所涉及的所有短语级",
"50M数据集中提取这些内容，如图4所示。我",
"重排取平均值来对由此获得的句子级表示进行",
"们遵循Zhangetal.（2020），并计算句子中所",
"评分",
"有成分对及其释义之间的短语对齐分数。6从",
"3Data and Training",
"这组短语对齐分数中，我们计算了短语之间的",
"部分一对一映射（图E4中的彩色形状）；也就",
"我们在通过回译CzEng（Bojaretal.，2016）语",
"是说，并非所有短语都对齐，而是一对一对齐",
"料库的捷克语句子构建的ParaNMT-50M释",
"的子集。最后，我们在语法翻译（Galleyetal.",
"义数据集（WietingandGimpel，2018）上训练",
"2004）中提取与规则对齐类似的对齐块：当对",
"和评估我们的模型。我们过滤此数据集以删除",
"齐短语A和A'分别包含对齐短语对（B,C)和",
"较短的句子（少于8个标记），低质量的释义",
"(B',C\"）时，我们可以提取对齐元组（tA,B,C)",
"对（由数据集中包含的翻译分数量化）和表现",
"和（tA,B',C'\"）。抽象出短语（B,)和（B',C\")",
"出低重新排序的示例（由基于每个单词的位置",
"，以构造短语级换能器的训练数据，包括监督",
"的重新排序分数量化源及其在目标句子中的对",
"o=Monotone或Flip。使用上述对齐策略，",
"齐单词）。这给我们留下了超过350k配对的释",
"",
"我们能够获得超过？？？万个对齐的短语对",
"义对",
"4Evaluation",
"3.1Training Data for Reap",
"Setup由于我们的主要目标是评估模型生成",
"为了训练我们的Reap模型（在2.2节中概述）",
"各种释义的能力，因此我们获得了一组释义，",
"我们采用现有的释义对（α,y*），并根据它们",
"并将其与其他方法产生的释义集进行了比较。",
"与目标句子的对齐方式得出源句子标记的伪地",
"6The score is computed using a weighted mean of",
"4See Appendix for Sow model architecture diagram.",
"the contextual similarity between individual words in",
"5In syntactic translation systems, rules similarly can",
"the phrases, where the weights are determined by the",
"be divided by whether they preserve order or invert it",
"corpus-level inverse-document frequency of the words.",
"(Wu, 1997)",
"Details in the Appendix."
] |
{
"arxiv_id": "2005.02013",
"language": "zh",
"page_idx": 3
}
|
||
2005.02013-zh-4
|
2005.02013-zh-4.jpg
|
<table>
<tbody>
<tr>
<th rowspan="2">Model</th>
<th colspan="4">oracle qualıty (over $\because10$ 1: no reiection) $\uparrow$</th>
<th> </th>
<th colspan="2">pairwise diversity $r~($post-rejectic</th>
</tr>
<tr>
<th>BLEU</th>
<th>ROUGE-1</th>
<th>$\operatorname{ROUGE-2}$</th>
<th>$\operatorname{ROUGE-L}$</th>
<th>$\%$ $\operatorname{rejected}$</th>
<th>self-BLEU $\downarrow$</th>
<th>$\operatorname{self-WER}\uparrow$</th>
</tr>
<tr>
<td>$\operatorname{copv-input}$</td>
<td>18.4</td>
<td>54.4</td>
<td>27.2</td>
<td>49.2</td>
<td>0</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td rowspan="3">${\mathrm{IIIP}}$ SCPN Transformer seq2seq +diverse-decoding</td>
<td>21.3</td>
<td>53.2</td>
<td>30.3</td>
<td>51.0</td>
<td>40.6</td>
<td>35.9</td>
<td>63.4</td>
</tr>
<tr>
<td>32.8</td>
<td>63.1</td>
<td>41.4</td>
<td>63.3</td>
<td>12.7</td>
<td>50.7</td>
<td>35.4</td>
</tr>
<tr>
<td>24.8</td>
<td>56.8</td>
<td>33.2</td>
<td>56.4</td>
<td>21.3</td>
<td>34.2</td>
<td>58.1</td>
</tr>
<tr>
<td rowspan="2">Sow- Reap (LSTM) Sow- Reap</td>
<td>27.0</td>
<td>57.9</td>
<td>34.8</td>
<td>57.5</td>
<td>31.7</td>
<td>46.2</td>
<td>53.9</td>
</tr>
<tr>
<td>30.9</td>
<td>62.3</td>
<td>40.2</td>
<td>61.7</td>
<td>15.9</td>
<td>38.0</td>
<td>57.9</td>
</tr>
</tbody>
</table>
Table 1: Quality and diversity metrics for the different models. Our proposed approach outperforms other diverse models (SCPN and diverse decoding) in terms of all the quality metrics. These models exhibit higher diversity, but with many more rejected paraphrases, indicating that these models more freely generate bad paraphrases.
为了获得 10 个释义，我们首先使用$\overline{\mathrm{E}}$2.3 节中的 Sow 方法计算一组 10 个不同的重新排序$\boldsymbol{r}^1,\ldots,\boldsymbol{r}^{10}$ ,然后使用 Reap 为每个生成 1 个最佳释义。我们使用 top- $k$解码来生成与重新排序相对应的最终释义集。我们的评估是通过 ParaNMT-50M 的 10k 示例完成的
### 4.1 Quantitative Evaluatiom
in the test set, we select $\hat{y}^{best}$ that achieves the best (oracle) sentence-level score with the ground truth paraphrase $\boldsymbol y.$ The corpus level evaluation is performed using pairs $( \hat{\boldsymbol{y}} ^{best}, \boldsymbol{y})$ .
2. Diversity:We calculate BLEU or WER be-
tween all pairs $(\boldsymbol{y}^i,\boldsymbol{y}^j)$ generated by a single model on a single sentence, then macroaverage these values at a corpus-level.
除了这些度量之外，我们还使用 Wieting et al. (2017)提出的释义相似性模型来计算生成的输出相对于输入的释义分数。与 Iyyer et al. (2018)类似，我们使用此分数过滤掉低质量的释义。我们根据此标准报告所有模型的拒绝率。请注意，我们的多样性度量是在过滤后计算的，因为通过包含语义不同的无意义释义候选项很容易获得高多样性
表 1 概述了不同型号的性能。结果表明，我们提出的模型在所有质量指标上都大大优于SCPN 模型。$^{8}$此外，我们的 LSTM 模型还击败了 SCPN 模型的性能，表明质量的提高不能完全归因于变压器的使用。我们的完整模型(带有重新排列)的质量也与香草 seq2seq 模型(没有重新排列)的质量相当。这表明，包含来自基于语法的神经换能器的重排不会损害质量，同时会大大提高分集性能
SCPN 模型具有 40.6% 的高拒绝分数。这表明，在用于为每个句子生成释义的 10 模板中， 平均而言，4 不适合给定的句子，因此被拒绝。另一方面，对于我们的模型，生成的释义中只有 15.9% 被拒绝，这意味着产生的重排通常是有意义的。这与香草 seq2seq 模型所表现出的$12.7\%$拒绝率相当，该模型不以任何语法或重新排列为条件，因此永远不必遵循不适当的结构
Baselines 我们将我们的模型与先前工作(Iyyer et al., 2018) 中提出的句法控制释义网络 (SCPN) 模型进行比较。它根据预先枚举的语法模板列表生成 10 不同的释义输出。这种方法已被证明优于其他以可解释的中间结构(Chen et al., 2019b) 为条件的释义方法。此外， 我们还报告了以下基准模型的结果：i) 精确输出输入句子的复制输入模型。ii) 香草 seq2seq 模型，该模型使用与 2.1 节相同的变压器编码器-解码器体系结构，但不以任何目标重排为条件。我们使用 top- $k$采样 (Fan et al., 2018) 从该模型生成 10 释义。$^{7}$ iii) 使用上述转换器 seq2seq 模型的多样化解码模型，在生成过程中具有多样化的解码 (Kumar et al., 2019) 。在这里，诱导的多样性是不受控制的，旨在最大化指标，如不同的 n-gram 和生成的句子之间的编辑距离。iv) 我们模型的 LSTM 版本， 其中 Reap 模型使用注意 (Bahdanau et al., 2014)的 LSTM 并复制 (See et al.,2017)而不是变压器。我们仍然使用基于转换器的短语转换器来获取源句子的重新排序，并且仍然在LSTM 注意力中使用位置编码
与 Cho et al. (2019)类似，我们报告了两种
类型的度量：
1.Quality:Given$k$generated paraphrases
$Y=\{\boldsymbol{y}^1,\boldsymbol{y}^2\ldots\boldsymbol{y}^k\}$ for each input sentence $^{7}$Prior work (Wang et al., 2019; Li et al., 2019) has
shown that such a transformer-based model provides a strong baseline and outperforms previous LSTM-based $(\bar{\text{Hasan et al., 2016) and VAE-based }}($Gupta et al., 2018) approaches.
$^{8}$The difference in performance between our proposed model and baseline models is statistically significant according to a paired bootstrap test.
|
[
[
0.2216796875,
0.07598876953125,
0.62060546875,
0.07659912109375,
0.62060546875,
0.0914306640625,
0.2216796875,
0.09027099609375
],
[
0.71728515625,
0.07659912109375,
0.95947265625,
0.07659912109375,
0.95947265625,
0.0914306640625,
0.71728515625,
0.0914306640625
],
[
0.34765625,
0.0926513671875,
0.424072265625,
0.09442138671875,
0.424072265625,
0.1092529296875,
0.34765625,
0.10687255859375
],
[
0.438232421875,
0.0926513671875,
0.515625,
0.09442138671875,
0.515625,
0.1092529296875,
0.438232421875,
0.10687255859375
],
[
0.5341796875,
0.0926513671875,
0.6171875,
0.09442138671875,
0.6171875,
0.1092529296875,
0.5341796875,
0.10687255859375
],
[
0.62744140625,
0.0955810546875,
0.8134765625,
0.0955810546875,
0.8134765625,
0.1092529296875,
0.62744140625,
0.1092529296875
],
[
0.85400390625,
0.0968017578125,
0.9326171875,
0.0968017578125,
0.9326171875,
0.10809326171875,
0.85400390625,
0.10809326171875
],
[
0.370361328125,
0.11224365234375,
0.405517578125,
0.11224365234375,
0.405517578125,
0.127685546875,
0.370361328125,
0.127685546875
],
[
0.279541015625,
0.09442138671875,
0.333251953125,
0.09442138671875,
0.333251953125,
0.1092529296875,
0.279541015625,
0.1092529296875
],
[
0.2939453125,
0.1134033203125,
0.32666015625,
0.1134033203125,
0.32666015625,
0.1258544921875,
0.2939453125,
0.1258544921875
],
[
0.4609375,
0.11224365234375,
0.495361328125,
0.11224365234375,
0.495361328125,
0.127685546875,
0.4609375,
0.127685546875
],
[
0.55908203125,
0.1134033203125,
0.59033203125,
0.1134033203125,
0.59033203125,
0.1258544921875,
0.55908203125,
0.1258544921875
],
[
0.66162109375,
0.1134033203125,
0.6826171875,
0.1134033203125,
0.6826171875,
0.1270751953125,
0.66162109375,
0.1270751953125
],
[
0.22412109375,
0.1258544921875,
0.270263671875,
0.1258544921875,
0.270263671875,
0.13720703125,
0.22412109375,
0.13720703125
],
[
0.19482421875,
0.114013671875,
0.270263671875,
0.114013671875,
0.270263671875,
0.1258544921875,
0.19482421875,
0.1258544921875
],
[
0.2939453125,
0.1258544921875,
0.328369140625,
0.1258544921875,
0.328369140625,
0.1383056640625,
0.2939453125,
0.1383056640625
],
[
0.3720703125,
0.1258544921875,
0.403076171875,
0.1258544921875,
0.403076171875,
0.1383056640625,
0.3720703125,
0.1383056640625
],
[
0.4609375,
0.1258544921875,
0.495361328125,
0.1258544921875,
0.495361328125,
0.1383056640625,
0.4609375,
0.1383056640625
],
[
0.5576171875,
0.1258544921875,
0.59033203125,
0.1258544921875,
0.59033203125,
0.1383056640625,
0.5576171875,
0.1383056640625
],
[
0.6572265625,
0.1258544921875,
0.6884765625,
0.1258544921875,
0.6884765625,
0.1383056640625,
0.6572265625,
0.1383056640625
],
[
0.75390625,
0.1258544921875,
0.78662109375,
0.1258544921875,
0.78662109375,
0.1383056640625,
0.75390625,
0.1383056640625
],
[
0.87890625,
0.1258544921875,
0.91162109375,
0.1258544921875,
0.91162109375,
0.1383056640625,
0.87890625,
0.1383056640625
],
[
0.1275634765625,
0.135986328125,
0.27197265625,
0.1383056640625,
0.27197265625,
0.1519775390625,
0.1275634765625,
0.149658203125
],
[
0.2939453125,
0.135986328125,
0.328369140625,
0.1383056640625,
0.32666015625,
0.15087890625,
0.29296875,
0.1484375
],
[
0.65576171875,
0.135986328125,
0.68994140625,
0.135986328125,
0.68994140625,
0.1519775390625,
0.65576171875,
0.1519775390625
],
[
0.87890625,
0.13720703125,
0.91162109375,
0.13720703125,
0.91162109375,
0.149658203125,
0.87890625,
0.149658203125
],
[
0.462646484375,
0.1383056640625,
0.49365234375,
0.1383056640625,
0.49365234375,
0.149658203125,
0.462646484375,
0.149658203125
],
[
0.3720703125,
0.1383056640625,
0.403076171875,
0.1383056640625,
0.403076171875,
0.149658203125,
0.3720703125,
0.149658203125
],
[
0.5576171875,
0.1383056640625,
0.59033203125,
0.1383056640625,
0.59033203125,
0.15087890625,
0.5576171875,
0.15087890625
],
[
0.75390625,
0.1383056640625,
0.78662109375,
0.1383056640625,
0.78662109375,
0.149658203125,
0.75390625,
0.149658203125
],
[
0.139404296875,
0.149658203125,
0.273681640625,
0.149658203125,
0.273681640625,
0.163330078125,
0.139404296875,
0.163330078125
],
[
0.2939453125,
0.149658203125,
0.32666015625,
0.149658203125,
0.32666015625,
0.162109375,
0.2939453125,
0.162109375
],
[
0.3720703125,
0.149658203125,
0.403076171875,
0.149658203125,
0.403076171875,
0.162109375,
0.3720703125,
0.162109375
],
[
0.462646484375,
0.149658203125,
0.49365234375,
0.149658203125,
0.49365234375,
0.162109375,
0.462646484375,
0.162109375
],
[
0.5576171875,
0.149658203125,
0.59033203125,
0.149658203125,
0.59033203125,
0.162109375,
0.5576171875,
0.162109375
],
[
0.65576171875,
0.149658203125,
0.6884765625,
0.149658203125,
0.6884765625,
0.162109375,
0.65576171875,
0.162109375
],
[
0.75390625,
0.149658203125,
0.78662109375,
0.149658203125,
0.78662109375,
0.162109375,
0.75390625,
0.162109375
],
[
0.87744140625,
0.149658203125,
0.91015625,
0.149658203125,
0.91015625,
0.162109375,
0.87744140625,
0.162109375
],
[
0.29296875,
0.1650390625,
0.330078125,
0.168701171875,
0.32666015625,
0.1817626953125,
0.28955078125,
0.1787109375
],
[
0.370361328125,
0.166259765625,
0.40478515625,
0.166259765625,
0.40478515625,
0.1817626953125,
0.370361328125,
0.1817626953125
],
[
0.459228515625,
0.166259765625,
0.495361328125,
0.166259765625,
0.495361328125,
0.1817626953125,
0.459228515625,
0.1817626953125
],
[
0.5556640625,
0.1650390625,
0.59423828125,
0.168701171875,
0.591796875,
0.182861328125,
0.55322265625,
0.179931640625
],
[
0.654296875,
0.166259765625,
0.68994140625,
0.166259765625,
0.68994140625,
0.1817626953125,
0.654296875,
0.1817626953125
],
[
0.75244140625,
0.166259765625,
0.78857421875,
0.166259765625,
0.78857421875,
0.1817626953125,
0.75244140625,
0.1817626953125
],
[
0.87744140625,
0.166259765625,
0.91162109375,
0.166259765625,
0.91162109375,
0.1817626953125,
0.87744140625,
0.1817626953125
],
[
0.2939453125,
0.179931640625,
0.328369140625,
0.179931640625,
0.328369140625,
0.1923828125,
0.2939453125,
0.1923828125
],
[
0.1368408203125,
0.16748046875,
0.270263671875,
0.16748046875,
0.270263671875,
0.18115234375,
0.1368408203125,
0.18115234375
],
[
0.370361328125,
0.179931640625,
0.40478515625,
0.179931640625,
0.40478515625,
0.1923828125,
0.370361328125,
0.1923828125
],
[
0.4609375,
0.179931640625,
0.495361328125,
0.179931640625,
0.495361328125,
0.1923828125,
0.4609375,
0.1923828125
],
[
0.5576171875,
0.179931640625,
0.59033203125,
0.179931640625,
0.59033203125,
0.1923828125,
0.5576171875,
0.1923828125
],
[
0.6572265625,
0.179931640625,
0.6884765625,
0.179931640625,
0.6884765625,
0.1923828125,
0.6572265625,
0.1923828125
],
[
0.75390625,
0.179931640625,
0.78662109375,
0.179931640625,
0.78662109375,
0.1923828125,
0.75390625,
0.1923828125
],
[
0.87744140625,
0.179931640625,
0.91015625,
0.179931640625,
0.91015625,
0.1923828125,
0.87744140625,
0.1923828125
],
[
0.118408203125,
0.2078857421875,
0.8818359375,
0.2078857421875,
0.8818359375,
0.221435546875,
0.118408203125,
0.221435546875
],
[
0.198974609375,
0.18115234375,
0.270263671875,
0.18115234375,
0.270263671875,
0.1923828125,
0.198974609375,
0.1923828125
],
[
0.118408203125,
0.221435546875,
0.884765625,
0.221435546875,
0.884765625,
0.236328125,
0.118408203125,
0.236328125
],
[
0.118408203125,
0.236328125,
0.8818359375,
0.236328125,
0.8818359375,
0.2493896484375,
0.118408203125,
0.2493896484375
],
[
0.118408203125,
0.25048828125,
0.35693359375,
0.25048828125,
0.35693359375,
0.26416015625,
0.118408203125,
0.26416015625
],
[
0.12005615234375,
0.290283203125,
0.48779296875,
0.289306640625,
0.48779296875,
0.303955078125,
0.12005615234375,
0.3046875
],
[
0.53564453125,
0.289306640625,
0.88330078125,
0.289306640625,
0.88330078125,
0.303955078125,
0.53564453125,
0.303955078125
],
[
0.12005615234375,
0.305908203125,
0.486083984375,
0.305908203125,
0.486083984375,
0.319580078125,
0.12005615234375,
0.319580078125
],
[
0.53662109375,
0.30712890625,
0.88330078125,
0.30712890625,
0.88330078125,
0.320556640625,
0.53662109375,
0.320556640625
],
[
0.118408203125,
0.320556640625,
0.489501953125,
0.320556640625,
0.489501953125,
0.33837890625,
0.118408203125,
0.33837890625
],
[
0.53662109375,
0.322509765625,
0.88330078125,
0.322509765625,
0.88330078125,
0.33740234375,
0.53662109375,
0.33740234375
],
[
0.12005615234375,
0.339111328125,
0.48779296875,
0.339111328125,
0.48779296875,
0.352783203125,
0.12005615234375,
0.352783203125
],
[
0.53564453125,
0.33740234375,
0.88330078125,
0.339111328125,
0.88330078125,
0.35400390625,
0.53564453125,
0.3515625
],
[
0.12005615234375,
0.355712890625,
0.48779296875,
0.355712890625,
0.48779296875,
0.369384765625,
0.12005615234375,
0.369384765625
],
[
0.5341796875,
0.352783203125,
0.62109375,
0.355224609375,
0.62109375,
0.369384765625,
0.5341796875,
0.36767578125
],
[
0.1175537109375,
0.369384765625,
0.42236328125,
0.369384765625,
0.42236328125,
0.38427734375,
0.1175537109375,
0.38427734375
],
[
0.51318359375,
0.37060546875,
0.884765625,
0.37060546875,
0.884765625,
0.385498046875,
0.51318359375,
0.385498046875
],
[
0.53564453125,
0.38720703125,
0.88671875,
0.38720703125,
0.88671875,
0.402099609375,
0.53564453125,
0.402099609375
],
[
0.12005615234375,
0.399658203125,
0.15283203125,
0.399658203125,
0.15283203125,
0.410888671875,
0.12005615234375,
0.410888671875
],
[
0.166259765625,
0.3984375,
0.36279296875,
0.3984375,
0.36279296875,
0.412109375,
0.166259765625,
0.412109375
],
[
0.53662109375,
0.402587890625,
0.88671875,
0.402587890625,
0.88671875,
0.41748046875,
0.53662109375,
0.41748046875
],
[
0.1275634765625,
0.419189453125,
0.48779296875,
0.419189453125,
0.48779296875,
0.432861328125,
0.1275634765625,
0.432861328125
],
[
0.53662109375,
0.419189453125,
0.8427734375,
0.419189453125,
0.8427734375,
0.432861328125,
0.53662109375,
0.432861328125
],
[
0.12005615234375,
0.435791015625,
0.48779296875,
0.435791015625,
0.48779296875,
0.449462890625,
0.12005615234375,
0.449462890625
],
[
0.515625,
0.43701171875,
0.880859375,
0.43701171875,
0.880859375,
0.45068359375,
0.515625,
0.45068359375
],
[
0.12005615234375,
0.451904296875,
0.484375,
0.451904296875,
0.484375,
0.465576171875,
0.12005615234375,
0.465576171875
],
[
0.515625,
0.453125,
0.8818359375,
0.453125,
0.8818359375,
0.46728515625,
0.515625,
0.46728515625
],
[
0.12005615234375,
0.468505859375,
0.48779296875,
0.468505859375,
0.48779296875,
0.482177734375,
0.12005615234375,
0.482177734375
],
[
0.5146484375,
0.4697265625,
0.87890625,
0.4697265625,
0.87890625,
0.482666015625,
0.5146484375,
0.482666015625
],
[
0.12176513671875,
0.48388671875,
0.486083984375,
0.48388671875,
0.486083984375,
0.49755859375,
0.12176513671875,
0.49755859375
],
[
0.51318359375,
0.485107421875,
0.880859375,
0.485107421875,
0.880859375,
0.498779296875,
0.51318359375,
0.498779296875
],
[
0.12005615234375,
0.50048828125,
0.484375,
0.50048828125,
0.484375,
0.51416015625,
0.12005615234375,
0.51416015625
],
[
0.515625,
0.501953125,
0.8818359375,
0.501953125,
0.8818359375,
0.515625,
0.515625,
0.515625
],
[
0.12005615234375,
0.51708984375,
0.484375,
0.51708984375,
0.484375,
0.53076171875,
0.12005615234375,
0.53076171875
],
[
0.515625,
0.5185546875,
0.8818359375,
0.5185546875,
0.8818359375,
0.5322265625,
0.515625,
0.5322265625
],
[
0.12005615234375,
0.5322265625,
0.48779296875,
0.5322265625,
0.48779296875,
0.5458984375,
0.12005615234375,
0.5458984375
],
[
0.51318359375,
0.533203125,
0.8818359375,
0.533203125,
0.8818359375,
0.54638671875,
0.51318359375,
0.54638671875
],
[
0.12005615234375,
0.548828125,
0.48779296875,
0.548828125,
0.48779296875,
0.5634765625,
0.12005615234375,
0.5634765625
],
[
0.51318359375,
0.548828125,
0.71533203125,
0.548828125,
0.71533203125,
0.5634765625,
0.51318359375,
0.5634765625
],
[
0.118408203125,
0.56396484375,
0.486083984375,
0.56396484375,
0.486083984375,
0.57763671875,
0.118408203125,
0.57763671875
],
[
0.5341796875,
0.568359375,
0.88330078125,
0.568359375,
0.88330078125,
0.58203125,
0.5341796875,
0.58203125
],
[
0.118408203125,
0.580078125,
0.486083984375,
0.580078125,
0.486083984375,
0.59326171875,
0.118408203125,
0.59326171875
],
[
0.515625,
0.58447265625,
0.8818359375,
0.58447265625,
0.8818359375,
0.59716796875,
0.515625,
0.59716796875
],
[
0.12176513671875,
0.5966796875,
0.489501953125,
0.5966796875,
0.489501953125,
0.61083984375,
0.12176513671875,
0.61083984375
],
[
0.5146484375,
0.599609375,
0.8818359375,
0.599609375,
0.8818359375,
0.6142578125,
0.5146484375,
0.6142578125
],
[
0.12005615234375,
0.61328125,
0.489501953125,
0.61328125,
0.489501953125,
0.626953125,
0.12005615234375,
0.626953125
],
[
0.515625,
0.615234375,
0.8818359375,
0.615234375,
0.8818359375,
0.6298828125,
0.515625,
0.6298828125
],
[
0.12005615234375,
0.62890625,
0.491943359375,
0.62890625,
0.491943359375,
0.642578125,
0.12005615234375,
0.642578125
],
[
0.5146484375,
0.6318359375,
0.8818359375,
0.63134765625,
0.8818359375,
0.6455078125,
0.5146484375,
0.646484375
],
[
0.12005615234375,
0.64453125,
0.489501953125,
0.64453125,
0.489501953125,
0.6591796875,
0.12005615234375,
0.6591796875
],
[
0.51318359375,
0.646484375,
0.8818359375,
0.64794921875,
0.8818359375,
0.662109375,
0.51318359375,
0.6611328125
],
[
0.118408203125,
0.66015625,
0.484375,
0.66015625,
0.484375,
0.67333984375,
0.118408203125,
0.67333984375
],
[
0.51123046875,
0.6630859375,
0.880859375,
0.6630859375,
0.880859375,
0.677734375,
0.51123046875,
0.677734375
],
[
0.12005615234375,
0.677734375,
0.490234375,
0.677734375,
0.490234375,
0.69140625,
0.12005615234375,
0.69140625
],
[
0.5146484375,
0.68115234375,
0.88330078125,
0.68115234375,
0.88330078125,
0.69482421875,
0.5146484375,
0.69482421875
],
[
0.12005615234375,
0.693359375,
0.489501953125,
0.693359375,
0.489501953125,
0.70703125,
0.12005615234375,
0.70703125
],
[
0.5146484375,
0.69775390625,
0.75390625,
0.69775390625,
0.75390625,
0.71142578125,
0.5146484375,
0.71142578125
],
[
0.118408203125,
0.7080078125,
0.48779296875,
0.708984375,
0.48779296875,
0.7236328125,
0.118408203125,
0.72265625
],
[
0.52880859375,
0.71337890625,
0.88330078125,
0.71435546875,
0.88330078125,
0.728515625,
0.52880859375,
0.72802734375
],
[
0.12005615234375,
0.7255859375,
0.489501953125,
0.7255859375,
0.489501953125,
0.7392578125,
0.12005615234375,
0.7392578125
],
[
0.5146484375,
0.728515625,
0.8818359375,
0.72998046875,
0.8818359375,
0.74462890625,
0.5146484375,
0.74365234375
],
[
0.12005615234375,
0.7421875,
0.489501953125,
0.7421875,
0.489501953125,
0.755859375,
0.12005615234375,
0.755859375
],
[
0.515625,
0.74658203125,
0.88330078125,
0.74658203125,
0.88330078125,
0.76123046875,
0.515625,
0.76123046875
],
[
0.12005615234375,
0.75830078125,
0.36279296875,
0.75830078125,
0.36279296875,
0.771484375,
0.12005615234375,
0.771484375
],
[
0.5146484375,
0.76123046875,
0.880859375,
0.76123046875,
0.880859375,
0.775390625,
0.5146484375,
0.775390625
],
[
0.139404296875,
0.77490234375,
0.48779296875,
0.77490234375,
0.48779296875,
0.7880859375,
0.139404296875,
0.7880859375
],
[
0.515625,
0.77783203125,
0.8818359375,
0.77783203125,
0.8818359375,
0.7919921875,
0.515625,
0.7919921875
],
[
0.12176513671875,
0.79150390625,
0.2216796875,
0.79150390625,
0.2216796875,
0.802734375,
0.12176513671875,
0.802734375
],
[
0.515625,
0.79443359375,
0.8818359375,
0.79443359375,
0.8818359375,
0.80810546875,
0.515625,
0.80810546875
],
[
0.118408203125,
0.80712890625,
0.489501953125,
0.80712890625,
0.489501953125,
0.82177734375,
0.118408203125,
0.82177734375
],
[
0.5146484375,
0.81005859375,
0.8818359375,
0.81005859375,
0.8818359375,
0.82373046875,
0.5146484375,
0.82373046875
],
[
0.139404296875,
0.822265625,
0.489501953125,
0.822265625,
0.489501953125,
0.83740234375,
0.139404296875,
0.83740234375
],
[
0.515625,
0.82666015625,
0.8818359375,
0.82666015625,
0.8818359375,
0.84033203125,
0.515625,
0.84033203125
],
[
0.1376953125,
0.845703125,
0.489501953125,
0.84716796875,
0.489501953125,
0.8623046875,
0.1376953125,
0.85986328125
],
[
0.5146484375,
0.84326171875,
0.53662109375,
0.84326171875,
0.53662109375,
0.85546875,
0.5146484375,
0.85546875
],
[
0.118408203125,
0.85986328125,
0.48779296875,
0.85986328125,
0.48779296875,
0.87353515625,
0.118408203125,
0.87353515625
],
[
0.1175537109375,
0.87158203125,
0.489501953125,
0.87060546875,
0.489501953125,
0.88525390625,
0.1175537109375,
0.8857421875
],
[
0.5322265625,
0.869140625,
0.884765625,
0.87158203125,
0.884765625,
0.8857421875,
0.5322265625,
0.88427734375
],
[
0.12005615234375,
0.88427734375,
0.489501953125,
0.88427734375,
0.489501953125,
0.89794921875,
0.12005615234375,
0.89794921875
],
[
0.51318359375,
0.88427734375,
0.884765625,
0.88427734375,
0.884765625,
0.89794921875,
0.51318359375,
0.89794921875
],
[
0.118408203125,
0.8955078125,
0.2418212890625,
0.896484375,
0.2418212890625,
0.90771484375,
0.118408203125,
0.90673828125
],
[
0.51318359375,
0.896484375,
0.8193359375,
0.896484375,
0.8193359375,
0.91015625,
0.51318359375,
0.91015625
]
] |
[
"Model oracle quality (over 10 sentences, no rejection)↑",
"pairwise diversity (post-rejection)",
"ROUGE-1",
"ROUGE-2",
"ROUGE-L",
"% rejected self-BLEU +",
"self-WER↑",
"54.4",
"BLEU",
"18.4",
"27.2",
"49.2",
"0",
"SCPN",
"copy-input",
"21.3",
"53.2",
"30.3",
"51.0",
"40.6",
"35.9",
"63.4",
"Transformer seq2seq",
"32.8",
"12.7",
"35.4",
"41.4",
"63.1",
"63.3",
"50.7",
"+ diverse-decoding",
"24.8",
"56.8",
"33.2",
"56.4",
"21.3",
"34.2",
"58.1",
"27.0",
"57.9",
"34.8",
"57.5",
"31.7",
"46.2",
"53.9",
"30.9",
"Sow-Reap (LSTM)",
"62.3",
"40.2",
"61.7",
"15.9",
"38.0",
"57.9",
"Table 1: Quality and diversity metrics for the different models. Our proposed approach outperforms",
"Sow-Reap",
"other diverse models (SCPN and diverse decoding) in terms of all the quality metrics. These models",
"exhibit higher diversity, but with many more rejected paraphrases, indicating that these models more",
"freely generate bad paraphrases.",
"为了获得10个释义，我们首先使用E2.3节",
"in the test set, we select ybest that achieves",
"中的Sow方法计算一组10个不同的重新排序",
"the best (oracle) sentence-level score with",
"rl,...,rlo，然后使用 Reap为每个生成1个",
"the ground truth paraphrase y . The corpus",
"最佳释义。我们使用top-k解码来生成与重",
"level evaluation is performed using pairs",
"新排序相对应的最终释义集。我们的评估是通",
"(ybest, y) :",
"过ParaNMT-50M的10k示例完成的",
"2. Diversity : We calculate BLEU or WER be-",
"tween all pairs (y',y') generated by a sin-",
"4.1",
"Quantitative Evaluation",
"gle model on a single sentence, then macro-",
"Baselines我们将我们的模型与先前工作",
"average these values at a corpus-level.",
"(Iyyeretal.，2018）中提出的句法控制释义网",
"除了这些度量之外，我们还使用Wietingetal.",
"络（SCPN）模型进行比较。它根据预先枚举",
"(2017)提出的释义相似性模型来计算生成的",
"的语法模板列表生成10不同的释义输出。这",
"输出相对于输人的释义分数。与 Iyyer et al",
"种方法已被证明优于其他以可解释的中间结构",
"（2018）类似，我们使用此分数过滤掉低质量的",
"(Chenetal.,2019b）为条件的释义方法。此外",
"释义。我们根据此标准报告所有模型的拒绝",
"我们还报告了以下基准模型的结果：i）精确输",
"率。请注意，我们的多样性度量是在过滤后计",
"出输入句子的复制输入模型。ii）香草seg2seq",
"算的，因为通过包含语义不同的无意义释义候",
"模型，该模型使用与2.1节相同的变压器编码",
"选项很容易获得高多样性",
"器-解码器体系结构，但不以任何目标重排为",
"表1概述了不同型号的性能。结果表明，我",
"条件。我们使用top-k采样（Fanetal.,2018)",
"们提出的模型在所有质量指标上都大大优于",
"从该模型生成10释义。7ii）使用上述转换",
"SCPN模型。8此外，我们的LSTM模型还击",
"器seq2seq模型的多样化解码模型，在生成过",
"败了SCPN模型的性能，表明质量的提高不",
"程中具有多样化的解码（Kumaretal.，2019）。",
"能完全归因于变压器的使用。我们的完整模型",
"在这里，诱导的多样性是不受控制的，旨在最",
"（带有重新排列）的质量也与香草seq2seq模型",
"大化指标，如不同的n-gram和生成的句子之",
"（没有重新排列）的质量相当。这表明，包含",
"间的编辑距离。iv）我们模型的LSTM版本，",
"来自基于语法的神经换能器的重排不会损害质",
"其中Reap模型使用注意（Bahdanauetal.,",
"量，同时会大大提高分集性能",
"2014）的LSTM并复制（Seeetal.，2017）而",
"SCPN模型具有40.6%的高拒绝分数。这表",
"不是变压器。我们仍然使用基于转换器的短语",
"明，在用于为每个句子生成释义的10模板中，",
"转换器来获取源句子的重新排序，并且仍然在",
"平均而言，4不适合给定的句子，因此被拒绝。",
"LSTM注意力中使用位置编码",
"另一方面，对于我们的模型，生成的释义中只",
"与Choetal.（2019）类似，我们报告了两种",
"有15.9%被拒绝，这意味着产生的重排通常是",
"类型的度量：",
"有意义的。这与香草seq2seq模型所表现出的",
"1. Quality : Given k generated paraphrases",
"12.7%拒绝率相当，该模型不以任何语法或重",
"Y = [y', y?.. yh] for each input sentence",
"新排列为条件，因此永远不必遵循不适当的结",
"7Prior work (Wang et al., 2019; Li et al., 2019) has",
"构",
"shown that such a transformer-based model provides a",
"strong baseline and outperforms previous LSTM-based",
"8The difference in performance between our pro-",
"(Hasan et al., 2016) and VAE-based (Gupta et al.,",
"posed model and baseline models is statistically sig-",
"2018)approaches.",
"nificant according to a paired bootstrap test."
] |
{
"arxiv_id": "2005.02013",
"language": "zh",
"page_idx": 4
}
|
||
2303.03169-zh-0
|
2303.03169-zh-0.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2023}}$
# Lipschitz 神经网络的统一代数透视
Alexandre Araujo*1, Aaron Havens*2, Blaise Delattre$^3,4$, Alexandre Allauzen$^3,5$ and Bin Hu$^2$
$^{1}$INRIA, Ecole Normale Supérieure, CNRS, PSL University, Paris, France
$^{2}$CSL$\&$ECE,University of Illinois Urbana-Champaign, IL,USA
$^{3}$ Miles Team, LAMSADE, Université Paris-Dauphine, PSL University, Paris, France
$^{4}$Foxstream, Vaulx-en-Velin, France
$^{5}$ESPCI PSL, Paris, France
$\operatorname{Abstract}$
Important research efforts have focused on the design and training of neural networks with a controlled Lipschitz constant. The goal is to increase and sometimes guarantee the robustness against adversarial attacks. Recent promising techniques draw inspirations from different backgrounds to design 1- Lipschitz neural networks, just to name a few: convex poten- tial layers derive from the discretization of continuous dynamical systems, $\begin{array}{c}\mathrm{Almost-Orthogonal-Layer~proposes~a~tailored~method~for~matrix~rescaling.}\\\end{array}$ However, it is today important to consider the recent and promising contributions in the feld under a common theoretical lens to better design new and improved layers. This paper introduces a novel algebraic perspective unifying various types of 1-Lipschitz neural networks, including the ones previously mentioned, along with methods based on orthogonality and spectral methods. Interestingly, we show that many existing techniques can be derived and generalized via fınding analytical solutions of a common semidefnite programming (SDP) condition. We also prove that AOL biases the scaled weight to the ones which are close to the set of orthogonal matrices in a certain mathematical manner. Moreover, our algebraic condition, combined with the Gershgorin circle theorem, readily leads to new and diverse parameterizations for 1-Lipschitz network layers. Our approach, called SDP-based Lipschitz Layers (SLL), allows us to design non-trivial yet efficient generalization of convex potential layers. Finally, the comprehensive set of experiments on image classification shows that SLLs outperform previous approaches on certified robust accuracy. Code is available at github.com/araujoalexandre/Lipschitz-SLL-Networks
(10/26/2023){:}Erratum is added in Appendix D. This is an updated version that fxes an implementation issue in the previous version. Due to that implementation issue, the original numerical results in our original ICLR paper are not accurate. We elaborate on the issue and provides some fix in $\tilde{\mathrm{Appendix~D}}.$
1
$\operatorname{Introduction}$
如今，深度神经网络的鲁棒性是建立对现实生活应用决策的信心的巨大挑战。应对这一挑战需要保证对抗性攻击的预测稳定性。在这种情况下，神经网络的 Lipschitz 常数是许多最新进展的核心关键属性。与分类器的余量一起，此属性使我们能够证明对最坏情况下的对抗性扰动的鲁棒性。该认证基于一个稳定的球体，在该球体中，对于球体内部的任何扰动， 决策都保持不变 (?)
1-Lipschitz 层的设计提供了一种成功的方法来为整个神经网络强制执行此属性。为此，已经设计了许多不同的技术，例如频谱归一化$\mathbb{E}(??)$ ,正交参数化$\mathbb{E}(????)$ ,凸势层 (CPL)$\mathbb{E}$
$^*$ Equal contribution.
1
|
[
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.176513671875,
0.10418701171875,
0.59326171875,
0.10418701171875,
0.59326171875,
0.125,
0.176513671875,
0.125
],
[
0.1854248046875,
0.1514892578125,
0.84716796875,
0.1514892578125,
0.84716796875,
0.1666259765625,
0.1854248046875,
0.1666259765625
],
[
0.181396484375,
0.17236328125,
0.71826171875,
0.173583984375,
0.71826171875,
0.1881103515625,
0.181396484375,
0.1875
],
[
0.1805419921875,
0.1849365234375,
0.66015625,
0.1875,
0.66015625,
0.2052001953125,
0.1805419921875,
0.20263671875
],
[
0.181396484375,
0.2020263671875,
0.79150390625,
0.203857421875,
0.79150390625,
0.2183837890625,
0.181396484375,
0.216552734375
],
[
0.181396484375,
0.2139892578125,
0.44775390625,
0.2171630859375,
0.44775390625,
0.23486328125,
0.181396484375,
0.231689453125
],
[
0.181396484375,
0.2291259765625,
0.38818359375,
0.231689453125,
0.38818359375,
0.246826171875,
0.181396484375,
0.24365234375
],
[
0.461669921875,
0.302978515625,
0.5390625,
0.302978515625,
0.5390625,
0.32080078125,
0.461669921875,
0.32080078125
],
[
0.233642578125,
0.335205078125,
0.7646484375,
0.335205078125,
0.7646484375,
0.348388671875,
0.233642578125,
0.348388671875
],
[
0.233642578125,
0.348388671875,
0.765625,
0.348388671875,
0.765625,
0.3623046875,
0.233642578125,
0.3623046875
],
[
0.233642578125,
0.363037109375,
0.7646484375,
0.363037109375,
0.7646484375,
0.376953125,
0.233642578125,
0.376953125
],
[
0.233642578125,
0.375732421875,
0.765625,
0.375732421875,
0.765625,
0.39013671875,
0.233642578125,
0.39013671875
],
[
0.233642578125,
0.39013671875,
0.7646484375,
0.39013671875,
0.7646484375,
0.404052734375,
0.233642578125,
0.404052734375
],
[
0.2320556640625,
0.40283203125,
0.7646484375,
0.404052734375,
0.7646484375,
0.41845703125,
0.2320556640625,
0.417236328125
],
[
0.233642578125,
0.416748046875,
0.76318359375,
0.417236328125,
0.76318359375,
0.431884765625,
0.233642578125,
0.43115234375
],
[
0.233642578125,
0.431884765625,
0.7646484375,
0.431884765625,
0.7646484375,
0.44580078125,
0.233642578125,
0.44580078125
],
[
0.2320556640625,
0.4443359375,
0.765625,
0.44580078125,
0.765625,
0.460205078125,
0.2320556640625,
0.458984375
],
[
0.2344970703125,
0.4609375,
0.76123046875,
0.4609375,
0.76123046875,
0.4716796875,
0.2344970703125,
0.4716796875
],
[
0.233642578125,
0.473388671875,
0.765625,
0.473388671875,
0.765625,
0.4873046875,
0.233642578125,
0.4873046875
],
[
0.233642578125,
0.4873046875,
0.765625,
0.4873046875,
0.765625,
0.50048828125,
0.233642578125,
0.50048828125
],
[
0.2320556640625,
0.5,
0.7646484375,
0.50048828125,
0.7646484375,
0.51513671875,
0.2320556640625,
0.5146484375
],
[
0.2344970703125,
0.51513671875,
0.765625,
0.51513671875,
0.765625,
0.52880859375,
0.2344970703125,
0.52880859375
],
[
0.233642578125,
0.52880859375,
0.765625,
0.52783203125,
0.765625,
0.54248046875,
0.233642578125,
0.54345703125
],
[
0.2344970703125,
0.54248046875,
0.76708984375,
0.54248046875,
0.76708984375,
0.55615234375,
0.2344970703125,
0.55615234375
],
[
0.233642578125,
0.55615234375,
0.7646484375,
0.55615234375,
0.7646484375,
0.57080078125,
0.233642578125,
0.57080078125
],
[
0.236083984375,
0.5712890625,
0.76318359375,
0.5712890625,
0.76318359375,
0.58203125,
0.236083984375,
0.58203125
],
[
0.2344970703125,
0.583984375,
0.7646484375,
0.583984375,
0.7646484375,
0.59765625,
0.2344970703125,
0.59765625
],
[
0.2344970703125,
0.59765625,
0.7646484375,
0.59765625,
0.7646484375,
0.6123046875,
0.2344970703125,
0.6123046875
],
[
0.233642578125,
0.611328125,
0.7646484375,
0.611328125,
0.7646484375,
0.62548828125,
0.233642578125,
0.62548828125
],
[
0.233642578125,
0.62548828125,
0.7646484375,
0.62548828125,
0.7646484375,
0.6396484375,
0.233642578125,
0.6396484375
],
[
0.2320556640625,
0.63818359375,
0.7646484375,
0.63818359375,
0.7646484375,
0.65283203125,
0.2320556640625,
0.65283203125
],
[
0.2320556640625,
0.65283203125,
0.7255859375,
0.6513671875,
0.7255859375,
0.66650390625,
0.2320556640625,
0.66748046875
],
[
0.2344970703125,
0.673828125,
0.76318359375,
0.673828125,
0.76318359375,
0.6875,
0.2344970703125,
0.6875
],
[
0.233642578125,
0.68798828125,
0.7646484375,
0.68798828125,
0.7646484375,
0.7021484375,
0.233642578125,
0.7021484375
],
[
0.233642578125,
0.7021484375,
0.765625,
0.7021484375,
0.765625,
0.71533203125,
0.233642578125,
0.71533203125
],
[
0.2344970703125,
0.71533203125,
0.7646484375,
0.71533203125,
0.7646484375,
0.72900390625,
0.2344970703125,
0.72900390625
],
[
0.2344970703125,
0.73095703125,
0.325927734375,
0.73095703125,
0.325927734375,
0.7421875,
0.2344970703125,
0.7421875
],
[
0.1732177734375,
0.76806640625,
0.2034912109375,
0.76806640625,
0.2034912109375,
0.78076171875,
0.1732177734375,
0.78076171875
],
[
0.200927734375,
0.767578125,
0.31787109375,
0.76806640625,
0.31787109375,
0.78271484375,
0.200927734375,
0.7822265625
],
[
0.176513671875,
0.798828125,
0.82421875,
0.798828125,
0.82421875,
0.8125,
0.176513671875,
0.8125
],
[
0.176513671875,
0.81298828125,
0.82275390625,
0.81298828125,
0.82275390625,
0.8271484375,
0.176513671875,
0.8271484375
],
[
0.176513671875,
0.8271484375,
0.82275390625,
0.8271484375,
0.82275390625,
0.84033203125,
0.176513671875,
0.84033203125
],
[
0.1748046875,
0.84033203125,
0.82421875,
0.84033203125,
0.82421875,
0.85400390625,
0.1748046875,
0.85400390625
],
[
0.176513671875,
0.85595703125,
0.3203125,
0.85595703125,
0.3203125,
0.8662109375,
0.176513671875,
0.8662109375
],
[
0.1748046875,
0.875,
0.82275390625,
0.87353515625,
0.82275390625,
0.88818359375,
0.1748046875,
0.8896484375
],
[
0.176513671875,
0.8896484375,
0.82421875,
0.8896484375,
0.82421875,
0.90283203125,
0.176513671875,
0.90283203125
],
[
0.1993408203125,
0.91015625,
0.343017578125,
0.912109375,
0.343017578125,
0.92529296875,
0.1993408203125,
0.923828125
],
[
0.491943359375,
0.94775390625,
0.50732421875,
0.94775390625,
0.50732421875,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2023",
"Lipschitz神经网络的统一代数透视",
"Alexandre Araujo*1, Aaron Havens*2, Blaise Delattre3,4, Alexandre Allauzen3,5 and Bin Hu?",
"1INRIA, Ecole Normale Superieure, CNRS, PSL University, Paris, France",
"2 CSL & ECE, University of Ilinois Urbana-Champaign, IL, USA",
"3 Miles Team, LAMSADE, Université Paris-Dauphine, PSL University, Paris, France",
"4 Foxstream, Vaulx-en-Velin, France",
"5ESPCIPSL,Paris,France",
"Abstract",
"Important research efforts have focused on the design and training of neu-",
"ral networks with a controlled Lipschitz constant. The goal is to increase",
"and sometimes guarantee the robustness against adversarial attacks. Re-",
"cent promising techniques draw inspirations from different backgrounds to",
"design 1-Lipschitz neural networks, just to name a few: convex poten-",
"tial layers derive from the discretization of continuous dynamical systems,",
"Almost-Orthogonal-Layer proposes a tailored method for matrix rescaling",
"However, it is today important to consider the recent and promising con-",
"tributions in the field under a common theoretical lens to better design",
"new and improved layers.This paper introduces a novel algebraic per",
"spective unifying various types of 1-Lipschitz neural networks, including",
"the ones previously mentioned, along with methods based on orthogonality",
"and spectral methods. Interestingly, we show that many existing tech-",
"niques can be derived and generalized via finding analytical solutions of a",
"common semidefinite programming (SDP) condition. We also prove that",
"AOL biases the scaled weight to the ones which are close to the set of",
"orthogonal matrices in a certain mathematical manner. Moreover, our al-",
"gebraiccondition,combined with the Gershgorin circle theorem,readily",
"leads to new and diverse parameterizations for 1-Lipschitz network layers.",
"Our approach, called SDP-based Lipschitz Layers (SLL), allows us to design",
"non-trivial yet efficient generalization of convex potential layers. Finally",
"the comprehensive set of experiments on image classification shows that",
"SLLs outperform previous approaches on certified robust accuracy. Code",
"is available at github.com/araujoalexandre/Lipschitz-SLL-Networks",
"(10/26/2023): Erratum is added in Appendix D. This is an updated ver-",
"sion that fixes an implementation issue in the previous version. Due to that",
"implementation issue, the original numerical results in our original ICLR",
"paper are not accurate. We elaborate on the issue and provides some fix in",
"Appendix D.",
"1",
"Introduction",
"如今，深度神经网络的鲁棒性是建立对现实生活应用决策的信心的巨大挑战。应对这一挑",
"战需要保证对抗性攻击的预测稳定性。在这种情况下，神经网络的Lipschitz常数是许多最",
"新进展的核心关键属性。与分类器的余量一起，此属性使我们能够证明对最坏情况下的对",
"抗性扰动的鲁棒性。该认证基于一个稳定的球体，在该球体中，对于球体内部的任何扰动，",
"决策都保持不变（？）",
"1-Lipschitz层的设计提供了一种成功的方法来为整个神经网络强制执行此属性。为此，已经",
"设计了许多不同的技术，例如频谱归一化E（??），正交参数化E(?????），凸势层（CPL)E",
"*Equal contribution.",
"1"
] |
{
"arxiv_id": "2303.03169",
"language": "zh",
"page_idx": 0
}
|
||
2303.03169-zh-1
|
2303.03169-zh-1.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2023}}$
(?)和几乎正交层 (AOL)E(?) 。虽然所有这些技术都有相同的目标，但它们的动机和派生可能会有很大的不同，从而提供不同的解决方案。尽管如此，他们的原始实验比较未能真正了解其独特的性能，健全性以及最终可能的互补性。因此，一个问题成为深人分析和未来发展的障碍：
Are there common principles underlying the developments of 1-Lipschitz Layers?
在本文中，我们提出了一种基于统一半定编程(SDP)方法的新颖观点来回答这个问题。我们介绍了一种常见的代数条件，这些条件是各种类型的方法的基础，例如频谱归一化，基于正交性的方法，AOL 和 CPL。我们的主要见解是，该条件可以表述为一个统一且简单的SDP 问题，并且通过找到该 SDP 的“分析解决方案”,系统地产生了 1-Lipschitz 体系结构的开发。我们的主要贡献总结如下。
$\bullet$ We provide a unifying algebraic perspective for 1-Lipschitz network layers by showing that
existing techniques such as spectral normalization, orthogonal parameterization, AOL, and CPL can all be recast as a solution of the same simple SDP condition (Theorem 1 and related discussions). Consequently, any new analytical solutions of our proposed SDP condition will immediately lead to new 1-Lipschitz network structures.
$\bullet$ Built upon the above algebraic viewpoint, we give a rigorous mathematical interpreta-
tion for AOL explaining how this method promotes“almost orthogonality”in training
(Theorem 2).
$\bullet$ Based on our SDPs, a new family of 1-Lipschitz network structures termed as SDP-based
Lipschitz layers (SLL) has been developed. Specifically, we apply the Gershgorin circle theorem to obtain some new SDP solutions, leading to non-trivial extensions of CPL (Theorem 3). We derive new SDP conditions to characterize SLL in a very general form (Theorem 4).
$\bullet$ Finally, we show, by a comprehensive set of experiments, that our new SDP-based Lips-
chitz layers outperform previous approaches on certified robust accuracy.
我们的工作受到？的启发，?为给定神经网络的 Lipschitz 常数的数值估计开发了 SDP 条件。一个主要的区别是，我们专注于“分析 SDP 解决方案”,可用于表征 1-Lipschitz 网络结构
2 Related Work
近年来，认证方法一直是可信机器学习发展的核心，尤其是深度学习。随机平滑$\overline{\mathbb{E}}(??)$ 是提供可证明的鲁棒性保证的首批防御之一。该方法简单地通过智能引入随机噪声来扩展给定的分类器，以增强分类器的鲁棒性。尽管此方法提供了有趣的认证鲁棒性级别，但它具有重要的缺点，例如推理的高计算成本以及从信息论的角度得出的一些不可能的结果$\mathbb{E}(??)$
证明分类器鲁棒性的另一种方法是控制其 Lipschitz 常数(??) 。主要思想是通过上边界分类器的边缘来导出特征空间中的认证半径。有关更多详细信息，请参见？的命题 1。该半径以及网络的 Lipschitz 常数可以证明鲁棒性。为了减小 Lipschitz 常数并具有非平凡的认证精度，? 和？都通过计算全局 Lipschitz 常数的界限来提高余量，但是，这些界限已被证明是宽松的。?不是对全局 Lipschitz 常数进行上限限制，而是利用$\log$al 信息来对 Lipschitz 常数进行更严格的限制。另一方面，其他作品，而不是上界的局部或全局 Lipschitz,设计的神经网络架构，可证明是 1-Lipschitz。在这个方向上的第一种方法之一是用它的光谱范数 (??) 归一化每一层。通过构造，每一层都是 1-Lipschitz。后来，大量的研究用正交矩阵代替了归一化的权重矩阵。通过添加梯度保存$\mathbb{E} ( ? ? ? ? )$ ,对光谱归一化方法进行了改进。这些方法在训练过程中通过正交性约束参数。具体来说，可以使用 Cayley 变换来约束权重$\mathbb{E}(?)$ ,并且以类似的方式，SOC$\overline{\mathbb{E}}(?)$ 使用倾斜对称矩阵的指数对其层进行参数化，使其正交。为了降低成本，?、? 和？在傅立叶域中正交化它们的卷积核
最近$,\operatorname{\overline{E}?}$ 的一项工作从动力学系统的角度研究了 Lipschitz 网络。从残差网络的连续视图开始，他们表明使用 Cayley 变换(?)和 SOC(?)的参数化分别对应于连续流的两个特定
2
|
[
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.048614501953125,
0.176513671875,
0.048614501953125
],
[
0.176513671875,
0.1060791015625,
0.8203125,
0.1060791015625,
0.8203125,
0.11932373046875,
0.176513671875,
0.11932373046875
],
[
0.17724609375,
0.11932373046875,
0.82177734375,
0.11932373046875,
0.82177734375,
0.1331787109375,
0.17724609375,
0.1331787109375
],
[
0.17724609375,
0.1331787109375,
0.82177734375,
0.1331787109375,
0.82177734375,
0.146484375,
0.17724609375,
0.146484375
],
[
0.176513671875,
0.147705078125,
0.2459716796875,
0.147705078125,
0.2459716796875,
0.1578369140625,
0.176513671875,
0.1578369140625
],
[
0.2115478515625,
0.186279296875,
0.78759765625,
0.186279296875,
0.78759765625,
0.199462890625,
0.2115478515625,
0.199462890625
],
[
0.17724609375,
0.2247314453125,
0.82275390625,
0.2247314453125,
0.82275390625,
0.238037109375,
0.17724609375,
0.238037109375
],
[
0.1748046875,
0.2374267578125,
0.82275390625,
0.238037109375,
0.82275390625,
0.25244140625,
0.1748046875,
0.251953125
],
[
0.17724609375,
0.25244140625,
0.82177734375,
0.25244140625,
0.82177734375,
0.266357421875,
0.17724609375,
0.266357421875
],
[
0.1748046875,
0.26513671875,
0.82275390625,
0.266357421875,
0.82275390625,
0.281005859375,
0.1748046875,
0.27978515625
],
[
0.176513671875,
0.281005859375,
0.428955078125,
0.281005859375,
0.428955078125,
0.294189453125,
0.176513671875,
0.294189453125
],
[
0.1732177734375,
0.3017578125,
0.82177734375,
0.3017578125,
0.82177734375,
0.31494140625,
0.1732177734375,
0.31494140625
],
[
0.193603515625,
0.317626953125,
0.8203125,
0.317626953125,
0.8203125,
0.32763671875,
0.193603515625,
0.32763671875
],
[
0.193603515625,
0.330810546875,
0.8203125,
0.330810546875,
0.8203125,
0.341552734375,
0.193603515625,
0.341552734375
],
[
0.19287109375,
0.343505859375,
0.82275390625,
0.343505859375,
0.82275390625,
0.35791015625,
0.19287109375,
0.35791015625
],
[
0.1895751953125,
0.35595703125,
0.69287109375,
0.356689453125,
0.69287109375,
0.37060546875,
0.1895751953125,
0.369384765625
],
[
0.1732177734375,
0.373779296875,
0.8203125,
0.373779296875,
0.8203125,
0.386962890625,
0.1732177734375,
0.386962890625
],
[
0.1895751953125,
0.386474609375,
0.82421875,
0.386962890625,
0.82421875,
0.401611328125,
0.1895751953125,
0.40087890625
],
[
0.19287109375,
0.401611328125,
0.289306640625,
0.401611328125,
0.289306640625,
0.41357421875,
0.19287109375,
0.41357421875
],
[
0.1732177734375,
0.417236328125,
0.82421875,
0.417236328125,
0.82421875,
0.43115234375,
0.1732177734375,
0.43115234375
],
[
0.191162109375,
0.431884765625,
0.82275390625,
0.431884765625,
0.82275390625,
0.44580078125,
0.191162109375,
0.44580078125
],
[
0.19287109375,
0.44580078125,
0.82275390625,
0.44580078125,
0.82275390625,
0.460205078125,
0.19287109375,
0.460205078125
],
[
0.191162109375,
0.458251953125,
0.82275390625,
0.458984375,
0.82275390625,
0.473388671875,
0.191162109375,
0.472900390625
],
[
0.19287109375,
0.473388671875,
0.2900390625,
0.473388671875,
0.2900390625,
0.48486328125,
0.19287109375,
0.48486328125
],
[
0.17236328125,
0.4873046875,
0.8203125,
0.4892578125,
0.8203125,
0.50390625,
0.17236328125,
0.501953125
],
[
0.1895751953125,
0.50244140625,
0.7138671875,
0.50244140625,
0.7138671875,
0.5166015625,
0.1895751953125,
0.5166015625
],
[
0.176513671875,
0.5244140625,
0.82275390625,
0.5244140625,
0.82275390625,
0.53857421875,
0.176513671875,
0.53857421875
],
[
0.176513671875,
0.53857421875,
0.82275390625,
0.53857421875,
0.82275390625,
0.5517578125,
0.176513671875,
0.5517578125
],
[
0.1748046875,
0.55322265625,
0.2091064453125,
0.55322265625,
0.2091064453125,
0.56494140625,
0.1748046875,
0.56494140625
],
[
0.1732177734375,
0.61279296875,
0.328369140625,
0.61279296875,
0.328369140625,
0.62744140625,
0.1732177734375,
0.62744140625
],
[
0.1748046875,
0.6591796875,
0.82275390625,
0.66015625,
0.82275390625,
0.673828125,
0.1748046875,
0.67236328125
],
[
0.176513671875,
0.673828125,
0.4697265625,
0.6748046875,
0.4697265625,
0.68798828125,
0.176513671875,
0.6875
],
[
0.47216796875,
0.6748046875,
0.82177734375,
0.6748046875,
0.82177734375,
0.68798828125,
0.47216796875,
0.68798828125
],
[
0.176513671875,
0.68798828125,
0.82275390625,
0.68798828125,
0.82275390625,
0.7021484375,
0.176513671875,
0.7021484375
],
[
0.17724609375,
0.7041015625,
0.8056640625,
0.7041015625,
0.8056640625,
0.7138671875,
0.17724609375,
0.7138671875
],
[
0.176513671875,
0.72265625,
0.8203125,
0.7236328125,
0.8203125,
0.7373046875,
0.176513671875,
0.736328125
],
[
0.176513671875,
0.7373046875,
0.82275390625,
0.7373046875,
0.82275390625,
0.75048828125,
0.176513671875,
0.75048828125
],
[
0.17724609375,
0.75244140625,
0.82177734375,
0.75244140625,
0.82177734375,
0.76318359375,
0.17724609375,
0.76318359375
],
[
0.176513671875,
0.76513671875,
0.82177734375,
0.76513671875,
0.82177734375,
0.77880859375,
0.176513671875,
0.77880859375
],
[
0.17724609375,
0.77978515625,
0.82177734375,
0.77978515625,
0.82177734375,
0.79345703125,
0.17724609375,
0.79345703125
],
[
0.17724609375,
0.79248046875,
0.82275390625,
0.79248046875,
0.82275390625,
0.80615234375,
0.17724609375,
0.80615234375
],
[
0.1748046875,
0.8046875,
0.82177734375,
0.80615234375,
0.82177734375,
0.82080078125,
0.1748046875,
0.8193359375
],
[
0.176513671875,
0.82080078125,
0.82177734375,
0.82080078125,
0.82177734375,
0.833984375,
0.176513671875,
0.833984375
],
[
0.17724609375,
0.833984375,
0.82177734375,
0.833984375,
0.82177734375,
0.84765625,
0.17724609375,
0.84765625
],
[
0.176513671875,
0.84765625,
0.82177734375,
0.84765625,
0.82177734375,
0.861328125,
0.176513671875,
0.861328125
],
[
0.17724609375,
0.8623046875,
0.82275390625,
0.8623046875,
0.82275390625,
0.87548828125,
0.17724609375,
0.87548828125
],
[
0.176513671875,
0.87548828125,
0.6845703125,
0.87548828125,
0.6845703125,
0.8896484375,
0.176513671875,
0.8896484375
],
[
0.176513671875,
0.89599609375,
0.82177734375,
0.896484375,
0.82177734375,
0.9111328125,
0.176513671875,
0.91015625
],
[
0.17724609375,
0.91357421875,
0.8203125,
0.91357421875,
0.8203125,
0.923828125,
0.17724609375,
0.923828125
],
[
0.490234375,
0.94873046875,
0.50732421875,
0.94873046875,
0.50732421875,
0.9619140625,
0.490234375,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2023",
"（?）和几乎正交层（AOL）E（?）。虽然所有这些技术都有相同的目标，但它们的动机和派生",
"可能会有很大的不同，从而提供不同的解决方案。尽管如此，他们的原始实验比较未能真正",
"了解其独特的性能，健全性以及最终可能的互补性。因此，一个问题成为深人分析和未来发",
"展的障碍：",
"Are there common principles underlying the developments of 1-Lipschitz Layers?",
"在本文中，我们提出了一种基于统一半定编程（SDP）方法的新颖观点来回答这个问题。我",
"们介绍了一种常见的代数条件，这些条件是各种类型的方法的基础，例如频谱归一化，基",
"于正交性的方法，AOL和CPL。我们的主要见解是，该条件可以表述为一个统一且简单的",
"SDP问题，并且通过找到该SDP的“分析解决方案”，系统地产生了1-Lipschitz体系结构",
"的开发。我们的主要贡献总结如下。",
".We provide a unifying algebraic perspectivefor 1-Lipschitz network layers by showing that",
"existingtechniques such as spectralnormalization,orthogonal parameterization,AOL",
"andCPLcanallberecast asasolutionofthesamesimpleSDPcondition(Theorem",
"and related discussions). Consequently, any new analytical solutions of our proposed SDP",
"condition will immediatelylead to new1-Lipschitznetwork structures",
". Built upon the above algebraic viewpoint, we give a rigorous mathematical interpreta",
"tion for AOL explaining how this method promotes “almost orthogonality\" in training",
"(Theorem 2)",
": Based on our SDPs, a new family of 1-Lipschitz network structures termed as SDP-based",
"Lipschitz layers (SLL) has been developed. Specifically, we apply the Gershgorin circle",
"theorem to obtain some new SDP solutions, leading to non-trivial extensions of CPL",
"(Theorem 3 ). We derive new SDP conditions to characterize SLL in a very general form",
"(Theorem4)",
". Finally, we show, by a comprehensive set of experiments, that our new SDP-based Lips.",
"chitz layers outperform previous approaches on certified robust accuracy.",
"我们的工作受到？的启发，？为给定神经网络的Lipschitz常数的数值估计开发了SDP条",
"件。一个主要的区别是，我们专注于“分析SDP解决方案”，可用于表征1-Lipschitz网络",
"结构",
"2 Related Work",
"近年来，认证方法一直是可信机器学习发展的核心，尤其是深度学习。随机平滑E（？?）是",
"提供可证明的鲁棒性保证的首批防御之一",
"。该方法简单地通过智能引人随机噪声来扩展给",
"定的分类器，以增强分类器的鲁棒性。尽管此方法提供了有趣的认证鲁棒性级别，但它具有",
"重要的缺点，例如推理的高计算成本以及从信息论的角度得出的一些不可能的结果E（？？）",
"证明分类器鲁棒性的另一种方法是控制其Lipschitz常数（??）。主要思想是通过上边界分",
"类器的边缘来导出特征空间中的认证半径。有关更多详细信息，请参见？的命题1。该半径",
"以及网络的Lipschitz常数可以证明鲁棒性。为了减小Lipschitz常数并具有非平凡的认证",
"精度，？和？都通过计算全局Lipschitz常数的界限来提高余量，但是，这些界限已被证明",
"是宽松的。？不是对全局Lipschitz常数进行上限限制，而是利用local信息来对Lipschitz",
"常数进行更严格的限制。另一方面，其他作品，而不是上界的局部或全局Lipschitz，设计",
"的神经网络架构，可证明是1-Lipschitz。在这个方向上的第一种方法之一是用它的光谱范",
"数（？?）归一化每一层。通过构造，每一层都是1-Lipschitz。后来，大量的研究用正交矩阵",
"代替了归一化的权重矩阵。通过添加梯度保存E（?????），对光谱归一化方法进行了改进",
"这些方法在训练过程中通过正交性约束参数。具体来说，可以使用Cayley变换来约束权",
"重E(?），并且以类似的方式，SOCE(?）使用倾斜对称矩阵的指数对其层进行参数化，使",
"其正交。为了降低成本，？、？和？在傅立叶域中正交化它们的卷积核",
"最近，E？的一项工作从动力学系统的角度研究了Lipschitz网络。从残差网络的连续视图",
"开始，他们表明使用Cayley变换（?）和SOC（？）的参数化分别对应于连续流的两个特定",
"2"
] |
{
"arxiv_id": "2303.03169",
"language": "zh",
"page_idx": 1
}
|
||
2303.03169-zh-2
|
2303.03169-zh-2.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2023}}$
离散化方案。此外，从凸势流中导出了一个新层，以确保 1-Lipschitz$\overline{\mathrm{E}}$属性$^{1}:$
(1)
$$z=x-\frac{2}{\|W\|_2^2}W\sigma(W^\top x+b),$$
其中$\|W\|_2$是权重矩阵$W$的频谱范数，$\sigma$是 ReLU 激活函数。一般来说，正交层的训练可能是昂贵的。Cayley 方法涉及矩阵求逆，SOC 的实现需要 SVD 或迭代泰勒展开。CPL 方法可以更有效，尽管仍然需要$\|W\|_2$的计算
最近的一项工作，几乎正交层 (AOL)(?)提出了一个中间立场：一种新的归一化，它通过支持正交性来使层 1-Lipschitz。完全连接的 AOL 层定义为$z=WDx+b$ ,其中$D\text{ 是 }^2$给出的对角矩阵：
$$D=\mathrm{diag}\left(\sum_j|W^\top W|_{ij}\right)^{-\frac{1}{2}}$$
(2)
他们证明了这一层是 1-Lipschitz,并且他们根据经验表明，在训练后，该层的雅可比行列式
(相对于 $x$ )几乎是正交的，因此有助于训练
另一个灵感来源是凸程序在神经网络鲁棒性认证中的应用 (???????)。最相关的工作是？, 它利用控制理论(?)的二次约束方法来$\overline{\mathrm{E}}$制定 SDP,以数值估计神经网络的全局 Lipschitz 常数。对于训练相对较小的 Lipschitz 网络(?),可以数值求解此类 SDP。然而，由于现有SDP 求解器的限制，在将此类方法部署到具有大数据集的深度学习问题时，可扩展性一直是一个问题。我们的重点是 Lipschitz 网络结构的设计，我们通过分析解决 SDP 来避免可扩展性问题
# 3 Background
Notation. $n\times n$单位矩阵和$n\times n$零矩阵分别表示为$I_n$和$0_n$ 。当维度从上下文中清除时，将省略下标。当矩阵$P$为负半定(定)时，我们将使用符号$P\preceq(\prec)0$ 。当矩阵$P$为正半定(定)时，我们将使用符号$P\succeq(\succ)0$ 。令$e_i$表示其$i$条目为 1 且所有其他条目为0的向量。给定标量$\{a_i\}_i=1^n$的集合，我们使用符号$\deg(a_i)$表示$n\times n$对角矩阵，其$(i,i)$ 条目为$a_i$。对于矩阵$A$,以下符号 $A^\mathrm{T},\|A\|_2$, tr$(A),\sigma_\mathrm{min}(A),\|A\|_F$ 和 $\rho(A)$ 分别代表其转置，最大奇异值，亦线，最小奇异佰，Frobenius 范数和光谱半径
Lipschitz functions. 函数$f:\mathbb{R}^n\to\mathbb{R}^m$相对于$\ell_2$范数是$L$ -Lipschitz,如果它满足所有$x,y\in\mathbb{R}^n$的$\|f(x)-f(y)\|\leq L\|x-y\|$ ,其中$\|\cdot\|$代表$\ell_2$范数。一个重要的事实是可以根据其 Lipschitz 常数 (?) 来证明神经网络的鲁棒性。在本文中，我们感兴趣的情况下，$L=1$。具体来说，我们考虑 1-Lipschitz 神经网络的训练。如果神经网络的每一层都是 1-Lipschitz, 那么整个神经网络也是 1-Lipschitz。Lipschitz 常数也满足三角不等式，因此凸组合将保留 1-Lipschitz 性质
Matrix cones: Positive semidefiniteness and diagonal dominance.
令$S^n$表示所有$n\times n$
实对称矩阵的集合。令$S_+^n\subset S^n$为所有$n\times n$对称正半定矩阵的集合。众所周知，$S_+^n$是$S^n$中的一个闭尖凸锥。通过跟踪内积，$S_+^n$也是自对偶的。考虑两个对称矩阵$A$和$B$,使得$A\succeq B\in S^n$ ,则我们有$A-B\in S_+^n$ ,“并且 tr$(A-B)$提供了$A$和$B$之间的距离度量。此外，我们有$\|A-B\|_F\leq$tr$(A-B)$。最后，所有具有非负对角条目的 $n\times n$ 实对称对角占优矩阵的集合由$D^n$表示。已知$D^n$形成一个封闭的，尖的，全锥(?)。根据 Gershgorin 圆定理(?),我们知道$D^n\subset S_+^n$。还已知$D^n$小于$S_+^n$ (?)。对于任何$A\in D^n$ ,我们有$A_{ii}\geq\sum_{j:j\neq i}|A_{ij}|$ 。重要的是需要$A_{ii}\geq0$ ,并且实对称对角占优矩阵的集合本身不是圆锥
4 An Algebraic Unification of 1-Lipschitz Layers
在本节中，我们通过开发表征 Lipschitz 属性的通用 SDP 条件，为各种 1-Lipschitz 层(光谱归一化，正交化，AOL 和 CP$\tilde{\mathcal{L}})$提供统一的代数视角。基于我们的代数观点，我们还提出了一种新的数学解释，解释了 AOL 如何在训练中促进正交性
$^{1}$We reverse the transposition from the original layer to have a consistent notation in the rest of
the article.
$^2$For simplicity, we assume all the columns of $W$ have at least one non-zero entry. Then (2) is
well defined.
3
|
[
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.048614501953125,
0.176513671875,
0.048614501953125
],
[
0.176513671875,
0.10479736328125,
0.74267578125,
0.10418701171875,
0.74267578125,
0.11871337890625,
0.176513671875,
0.11932373046875
],
[
0.4755859375,
0.1275634765625,
0.48193359375,
0.1331787109375,
0.476318359375,
0.1383056640625,
0.468017578125,
0.1319580078125
],
[
0.49267578125,
0.1319580078125,
0.60546875,
0.1331787109375,
0.60546875,
0.147705078125,
0.49267578125,
0.146484375
],
[
0.80224609375,
0.1337890625,
0.82421875,
0.1337890625,
0.82421875,
0.147705078125,
0.80224609375,
0.147705078125
],
[
0.451904296875,
0.1427001953125,
0.496826171875,
0.1427001953125,
0.496826171875,
0.1572265625,
0.451904296875,
0.1572265625
],
[
0.393798828125,
0.1363525390625,
0.439453125,
0.1363525390625,
0.439453125,
0.147705078125,
0.393798828125,
0.147705078125
],
[
0.1829833984375,
0.169189453125,
0.1895751953125,
0.169189453125,
0.1895751953125,
0.17236328125,
0.1829833984375,
0.17236328125
],
[
0.188720703125,
0.1641845703125,
0.82275390625,
0.1641845703125,
0.82275390625,
0.177978515625,
0.188720703125,
0.177978515625
],
[
0.17724609375,
0.1787109375,
0.82275390625,
0.1787109375,
0.82275390625,
0.1925048828125,
0.17724609375,
0.1925048828125
],
[
0.176513671875,
0.1925048828125,
0.50732421875,
0.1925048828125,
0.50732421875,
0.205810546875,
0.176513671875,
0.205810546875
],
[
0.176513671875,
0.21337890625,
0.82177734375,
0.21337890625,
0.82177734375,
0.2266845703125,
0.176513671875,
0.2266845703125
],
[
0.176513671875,
0.2259521484375,
0.82275390625,
0.2266845703125,
0.82275390625,
0.2412109375,
0.176513671875,
0.240478515625
],
[
0.17724609375,
0.242431640625,
0.279296875,
0.242431640625,
0.279296875,
0.25390625,
0.17724609375,
0.25390625
],
[
0.396240234375,
0.261474609375,
0.5703125,
0.261474609375,
0.5703125,
0.279052734375,
0.396240234375,
0.279052734375
],
[
0.80224609375,
0.264404296875,
0.82421875,
0.264404296875,
0.82421875,
0.27783203125,
0.80224609375,
0.27783203125
],
[
0.176513671875,
0.287353515625,
0.82275390625,
0.287353515625,
0.82275390625,
0.300537109375,
0.176513671875,
0.300537109375
],
[
0.17724609375,
0.302978515625,
0.494384765625,
0.302978515625,
0.494384765625,
0.313232421875,
0.17724609375,
0.313232421875
],
[
0.17724609375,
0.322509765625,
0.82275390625,
0.322509765625,
0.82275390625,
0.3359375,
0.17724609375,
0.3359375
],
[
0.17724609375,
0.3359375,
0.82275390625,
0.3359375,
0.82275390625,
0.349853515625,
0.17724609375,
0.349853515625
],
[
0.17724609375,
0.35107421875,
0.82177734375,
0.35107421875,
0.82177734375,
0.3642578125,
0.17724609375,
0.3642578125
],
[
0.1748046875,
0.3623046875,
0.82177734375,
0.3642578125,
0.82177734375,
0.37744140625,
0.1748046875,
0.375732421875
],
[
0.176513671875,
0.37744140625,
0.82177734375,
0.37744140625,
0.82177734375,
0.391357421875,
0.176513671875,
0.391357421875
],
[
0.176513671875,
0.393310546875,
0.259033203125,
0.393310546875,
0.259033203125,
0.40478515625,
0.176513671875,
0.40478515625
],
[
0.1732177734375,
0.425537109375,
0.310546875,
0.425537109375,
0.310546875,
0.43994140625,
0.1732177734375,
0.43994140625
],
[
0.1805419921875,
0.458251953125,
0.2459716796875,
0.458251953125,
0.2459716796875,
0.468505859375,
0.1805419921875,
0.468505859375
],
[
0.268798828125,
0.45703125,
0.8203125,
0.45703125,
0.8203125,
0.47021484375,
0.268798828125,
0.47021484375
],
[
0.176513671875,
0.47021484375,
0.82177734375,
0.47021484375,
0.82177734375,
0.484130859375,
0.176513671875,
0.484130859375
],
[
0.17724609375,
0.484130859375,
0.82177734375,
0.484130859375,
0.82177734375,
0.49755859375,
0.17724609375,
0.49755859375
],
[
0.176513671875,
0.498779296875,
0.82177734375,
0.498779296875,
0.82177734375,
0.51220703125,
0.176513671875,
0.51220703125
],
[
0.17724609375,
0.51220703125,
0.82177734375,
0.51220703125,
0.82177734375,
0.5263671875,
0.17724609375,
0.5263671875
],
[
0.176513671875,
0.52587890625,
0.69287109375,
0.52587890625,
0.69287109375,
0.5390625,
0.176513671875,
0.5390625
],
[
0.181396484375,
0.5546875,
0.82177734375,
0.5546875,
0.82177734375,
0.568359375,
0.181396484375,
0.568359375
],
[
0.1748046875,
0.568359375,
0.8203125,
0.5673828125,
0.8203125,
0.58203125,
0.1748046875,
0.58349609375
],
[
0.17724609375,
0.58349609375,
0.82177734375,
0.58349609375,
0.82177734375,
0.5966796875,
0.17724609375,
0.5966796875
],
[
0.178955078125,
0.5966796875,
0.82177734375,
0.5966796875,
0.82177734375,
0.60986328125,
0.178955078125,
0.60986328125
],
[
0.176513671875,
0.609375,
0.82177734375,
0.60986328125,
0.82177734375,
0.625,
0.176513671875,
0.62353515625
],
[
0.1748046875,
0.62353515625,
0.414306640625,
0.625,
0.414306640625,
0.6396484375,
0.1748046875,
0.63818359375
],
[
0.178955078125,
0.65380859375,
0.6435546875,
0.65380859375,
0.6435546875,
0.66748046875,
0.178955078125,
0.66748046875
],
[
0.65625,
0.65771484375,
0.66259765625,
0.65771484375,
0.66259765625,
0.6611328125,
0.65625,
0.6611328125
],
[
0.6640625,
0.65380859375,
0.82177734375,
0.65380859375,
0.82177734375,
0.66748046875,
0.6640625,
0.66748046875
],
[
0.17724609375,
0.66845703125,
0.82177734375,
0.66845703125,
0.82177734375,
0.681640625,
0.17724609375,
0.681640625
],
[
0.176513671875,
0.68310546875,
0.82177734375,
0.68310546875,
0.82177734375,
0.6962890625,
0.176513671875,
0.6962890625
],
[
0.1748046875,
0.6962890625,
0.8203125,
0.6962890625,
0.8203125,
0.71044921875,
0.1748046875,
0.71044921875
],
[
0.176513671875,
0.7138671875,
0.82177734375,
0.7138671875,
0.82177734375,
0.72802734375,
0.176513671875,
0.72802734375
],
[
0.17724609375,
0.7265625,
0.82177734375,
0.7265625,
0.82177734375,
0.7412109375,
0.17724609375,
0.7412109375
],
[
0.1748046875,
0.7392578125,
0.82275390625,
0.73779296875,
0.82275390625,
0.755859375,
0.1748046875,
0.7568359375
],
[
0.176513671875,
0.7568359375,
0.82275390625,
0.7568359375,
0.82275390625,
0.771484375,
0.176513671875,
0.771484375
],
[
0.1732177734375,
0.7890625,
0.5986328125,
0.79150390625,
0.5986328125,
0.80615234375,
0.1732177734375,
0.8037109375
],
[
0.176513671875,
0.8212890625,
0.82177734375,
0.8212890625,
0.82177734375,
0.83544921875,
0.176513671875,
0.83544921875
],
[
0.17724609375,
0.83544921875,
0.82177734375,
0.83544921875,
0.82177734375,
0.8486328125,
0.17724609375,
0.8486328125
],
[
0.1748046875,
0.849609375,
0.62744140625,
0.849609375,
0.62744140625,
0.86279296875,
0.1748046875,
0.86279296875
],
[
0.193603515625,
0.87255859375,
0.82421875,
0.87255859375,
0.82421875,
0.88623046875,
0.193603515625,
0.88623046875
],
[
0.176513671875,
0.88623046875,
0.2442626953125,
0.88623046875,
0.2442626953125,
0.89794921875,
0.176513671875,
0.89794921875
],
[
0.193603515625,
0.89794921875,
0.82421875,
0.8994140625,
0.82421875,
0.9140625,
0.193603515625,
0.912109375
],
[
0.176513671875,
0.91357421875,
0.256591796875,
0.91357421875,
0.256591796875,
0.923828125,
0.176513671875,
0.923828125
],
[
0.491943359375,
0.94873046875,
0.5087890625,
0.94873046875,
0.5087890625,
0.9619140625,
0.491943359375,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2023",
"离散化方案。此外，从凸势流中导出了一个新层，以确保1-LipschitzE属性1：",
"2",
"Wo(WTa+b)",
"(1)",
"IIWI12",
"Z=a",
"",
"其中IWIl2是权重矩阵W的频谱范数，α是ReLU激活函数。一般来说，正交层的训练",
"可能是昂贵的。Cayley方法涉及矩阵求逆，SOC的实现需要SVD或迭代泰勒展开。CPL",
"方法可以更有效，尽管仍然需要IWI2的计算",
"最近的一项工作，几乎正交层（AOL）（?）提出了一个中间立场：一种新的归一化，它通过",
"支持正交性来使层1-Lipschitz。完全连接的AOL层定义为=WDa+b，其中D是2给",
"出的对角矩阵：",
"D = diag (Z, IWTWlij",
"(2)",
"他们证明了这一层是1-Lipschitz，并且他们根据经验表明，在训练后，该层的雅可比行列式",
"（相对于）儿乎是正交的，因此有助于训练",
"另一个灵感来源是凸程序在神经网络鲁棒性认证中的应用（??????）。最相关的工作是？",
"它利用控制理论（?）的二次约束方法来E制定SDP，以数值估计神经网络的全局Lipschitz",
"常数。对于训练相对较小的Lipschitz网络（?），可以数值求解此类SDP。然而，由于现有",
"SDP求解器的限制，在将此类方法部署到具有大数据集的深度学习问题时，可扩展性一直",
"是一个问题。我们的重点是Lipschitz网络结构的设计，我们通过分析解决SDP来避免可",
"扩展性问题",
"3Background",
"Notation",
"n×n单位矩阵和n×n零矩阵分别表示为In和On。当维度从上下文中清除",
"时，将省略下标。当矩阵P为负半定（定）时，我们将使用符号P≤（)0。当矩阵P为",
"正半定（定）时，我们将使用符号P≥（)0。令ei表示其i-条目为1且所有其他条目为",
"0的向量。给定标量(ai}n=1的集合，我们使用符号diag(ai）表示n×n对角矩阵，其（i,i）",
"条目为ai。对于矩阵A，以下符号AT，All2，tr(A），αmin(A），IAllF和p(A)分别代",
"表其转置，最大奇异值，迹线，最小奇异值，Frobenius范数和光谱半径",
"Lipschitzfunctions.函数f：Rn→Rm相对于l2范数是L-Lipschitz，如果它满足所",
"有,yER的llf(α)-f(y)Il≤La-yll，其中 ll-l代表2范数。一个重要的事实是",
"可以根据其Lipschitz常数（?）来证明神经网络的鲁棒性。在本文中，我们感兴趣的情况",
"下，L=1。具体来说，我们考虑1-Lipschitz神经网络的训练。如果神经网络的每一层都",
"是1-Lipschitz，那么整个神经网络也是1-Lipschitz。Lipschitz常数也满足三角不等式，因",
"此凸组合将保留1-Lipschitz性质",
"Matrix cones: Positive semidefiniteness and diagonal dominance.",
"",
"令Sn表示所有nxn",
"实对称矩阵的集合。令SCSn为所有n×n对称正半定矩阵的集合。众所周知，S是",
"Sn中的一个闭尖凸锥。通过跟踪内积，S也是自对偶的。考虑两个对称矩阵A和B，使",
"得ABE Sn，则我们有 A-B ESn，并且 tr(A-B)提供了 A和 B之间的距离度量",
"此外，我们有IA-BIlE≤tr(A-B）。最后，所有具有非负对角条目的n×n实对称对角",
"占优矩阵的集合由Dn表示。已知Dn形成一个封闭的，尖的，全锥（?）。根据Gershgorin",
"圆定理（？），我们知道DnCS。还已知Dn小于S（？）。对于任何AEDn，我们有",
"Ai≥ZjilAil。重要的是需要Aii≥0，并且实对称对角占优矩阵的集合本身不是圆锥",
"4An Algebraic Unification of 1-Lipschitz Layers",
"在本节中，我们通过开发表征Lipschitz属性的通用SDP条件，为各种1-Lipschitz层（光",
"谱归一化，正交化，AOL和CPL）提供统一的代数视角。基于我们的代数观点，我们还提",
"出了一种新的数学解释，解释了AOL如何在训练中促进正交性",
"1We reverse the transposition from the original layer to have a consistent notation in the rest of",
"the article",
"2For simplicity, we assume all the columns of W have at least one non-zero entry. Then (2) is",
"well defined.",
"3"
] |
{
"arxiv_id": "2303.03169",
"language": "zh",
"page_idx": 2
}
|
||
2303.03169-zh-3
|
2303.03169-zh-3.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2023}}$
4.1 The unifying Algebraic Condition
首先，我们提出了一个代数条件，该条件可用于统一 SN, AOL 和 CPL 等现有技术的发展。
我们的主要定理形式化如下
Theorem 1. For any weight matrix $W\in\mathbb{R}^{m\times n}$, if there exists a nonsingular diagonal matrix
$T$ such that $W^{\mathrm{T}}W-T\preceq0$, then the two following statements hold true.
1. The mapping $g(x)=WT^{-\frac12}x+b$ is 1-Lipschitz.
2. The mapping $h(x)=x-2WT^{-1}\sigma(W^{\mathrm{T}}x+b)$ is 1-Lipschitz if $\sigma$ is ReLU, tanh or
sigmoid.
附录中提供了上述定理的证明和一些相关的控制理论解释。该定理允许我们在两种重要情况下使用$T$的各种选择来设计不同的 1-Lipschitz 层：对于语句 1 的线性变换，以及对于语句 2 的残差和非线性块。此外，对于任何给定的权重矩阵$W$ ,条件$W^\mathrm{T}W\preceq T$在$T$中是线性的，因此可以视为具有决策变量$T$ 的 SDP 条件。为了强调该定理的重要性，我们建议通过为 SDP 条件$W^\mathrm{T}W\preceq T$选择特定的$T$来推导用于设计 1-Lipschitz 层的现有方法。然后自动获取 1-Lipschitz 属性
$\bullet$ Spectral Normalization (SN) corresponds to an almost trivial choice if we notice
${\mathrm{that~}}W^{\mathrm{T} }W\preceq \| W_{1}^{\mathrm{T} }W\| _{2}\dot{I} \preceq ^{\prime }\| W\| _{2}^{2}\dot{I}$ .Hence with$T= \| W\| _{2}^{2}I$ ,we build the SN
$\operatorname { layer} g( x) = WT^{- \frac 12}x+ b= \frac 1{\| W\| _{2}}\bar{W} x+ b$ .
· The Orthogonality-based parameterization is obtained by setting $T=I$ and enforc-
$\begin{array}{l}{\mathrm{ing~the~equality~}W^{\mathrm{T}}W=T=I\:.}&{\mathrm{Then~obviously~}g(x)=Wx+b\:\mathrm{is~1-Lipschitz.}}\\{\mathrm{AOL~formula~can~be~derived~by~letting~}T\:=\:\mathrm{diag}(\sum_{i=1}^{n}|W^{\mathrm{T}}W|_{ij})\:.}&{\mathrm{With~this}}\end{array}$ choice, we have $T-W^{\mathrm{T}}W\in\boldsymbol{D}^{n}\subset\boldsymbol{S}_{+}^{n}$, hence $W^{\mathrm{T}}\dot{W}\preceq T.$ Then Statement 1 in Theorem 1 implies that the AOL layer, written as $g(x)=WT^{-\frac12}x+b$ ,is $1{\mathrm{-Lipschitz.}}^{3}$
$\bullet$ CPL follows the same SN choice $T=\|W\|_2^2I$ , but with Statement 2 of Theorem 1
.Hence we derive a different function $h(x)=x-\frac2{\|W\|_{2}^{2}}W\sigma(W^{\mathrm{T}}x+b)$ which is also
1-Lipschitz.
上述讨论说明了在同一理论框架内表达所有这些方法的好处，为我们提供了一种新的工具来表征不同方法之间的相似性。例如，SN 和 CPL 共享$T=\|W\|_2^2I$的相同选择。它们之间的区别在于使用哪个语句。因此，CPL 可以被视为 SN 的“剩余版本”。显然，剩余网络结构允许 CPL 比 SN 更有效地解决梯度消失问题。使用相同的方法，我们可以很容易地从我们的统一代数条件中推断出基于正交性的参数化和 AOL 的“残差”对应物是什么。对于基于正交性的参数化，如果我们通过 SOC 和 ECO 等方法强制执行$W^\mathrm{T}W=T=I$ ,则函数$h(x)=x-2W\sigma(W^{\mathrm{T}}x+b)$为 1-Lipschitz (通过定理 1 中的语句 2)。最后，如果我们选择$T=$diag$\left(\sum_j=1^n|W^{\mathrm{T}}W|_{ij}\right)$,那么函数 $h(x)=x-2W\operatorname{diag}\left(\sum_{j=1}^n|W^{\mathrm{T}}W|_{ij}\right)^{-1}\sigma(W^{\mathrm{T}}x+b)$ 也是 1-Lipschitz。因此，从现有的 1-Lipschitz 网络结构中创建新类是很简单的
定理 1 的另一个重要结果是关于新层开发。SDP 条件$W^\mathrm{T}W-T\preceq\underline{0}$的任何新的非奇异对角解$T$都会立即以$g(x)=WT^-\frac12x+b$或$h(x)=x-2WT^{-1}\sigma(W^\mathrm{T}x+b)$的形式导致新的1-Lipschitz 网络结构。因此，可以将 1-Lipschitz 网络结构的发展重新表述为寻找具有非奇$\hat{\text{异对角 }T\text{ 的矩阵不等式 }W^\text{T}W\preceq T\text{ 的解析解。事实上，Gershgorin 圆定理可以帮助系统地}}$ 改进$T$的现有选择。在$\operatorname{E}$5 节中，我们将讨论$T$的此类新选择以及改进 CPL 的相关应用。在这一点上，值得注意的是，要开发深度 Lipschitz 网络，重要的是要有$T$的解析公式。$T$ 的解析公式将能够快速计算$WT^{-\frac12}$或$WT^{-1}$
定理 1 在建立 1-Lipschitz 网络层与代数条件$W^\mathrm{T}W\preceq T$之间的连接方面非常强大。接下来，我们将更仔细地研究这个代数条件，并提供一个新的数学解释，解释 AOL 如何生成“几乎正交”的权重
$^{3}$For ease of exposition, our main paper always assumes that all the columns of $W$ have at least
one non-zero entry such that (2) is well defined. To drop this assumption, we can use a variant of Theorem 1 which replaces the SDP condition with a bilinear matrix inequality condition. We will discuss this point in the appendix.
4
|
[
[
0.176513671875,
0.034088134765625,
0.50732421875,
0.034088134765625,
0.50732421875,
0.048614501953125,
0.176513671875,
0.048614501953125
],
[
0.176513671875,
0.1060791015625,
0.45751953125,
0.1060791015625,
0.45751953125,
0.11932373046875,
0.176513671875,
0.11932373046875
],
[
0.176513671875,
0.1307373046875,
0.82275390625,
0.1307373046875,
0.82275390625,
0.14453125,
0.176513671875,
0.14453125
],
[
0.176513671875,
0.1458740234375,
0.371826171875,
0.1458740234375,
0.371826171875,
0.1590576171875,
0.176513671875,
0.1590576171875
],
[
0.1748046875,
0.160400390625,
0.82421875,
0.160400390625,
0.82421875,
0.177978515625,
0.1748046875,
0.177978515625
],
[
0.1732177734375,
0.1749267578125,
0.70361328125,
0.175537109375,
0.70361328125,
0.1900634765625,
0.1732177734375,
0.189453125
],
[
0.2115478515625,
0.203857421875,
0.5810546875,
0.20263671875,
0.5810546875,
0.2171630859375,
0.2115478515625,
0.2183837890625
],
[
0.2115478515625,
0.2279052734375,
0.82421875,
0.2279052734375,
0.82421875,
0.242431640625,
0.2115478515625,
0.242431640625
],
[
0.2344970703125,
0.24365234375,
0.291748046875,
0.24365234375,
0.291748046875,
0.255126953125,
0.2344970703125,
0.255126953125
],
[
0.17724609375,
0.266357421875,
0.82275390625,
0.266357421875,
0.82275390625,
0.27978515625,
0.17724609375,
0.27978515625
],
[
0.176513671875,
0.281005859375,
0.82275390625,
0.281005859375,
0.82275390625,
0.294189453125,
0.176513671875,
0.294189453125
],
[
0.176513671875,
0.294189453125,
0.82275390625,
0.292236328125,
0.82275390625,
0.30615234375,
0.176513671875,
0.30810546875
],
[
0.176513671875,
0.30810546875,
0.82177734375,
0.30810546875,
0.82177734375,
0.3212890625,
0.176513671875,
0.3212890625
],
[
0.1748046875,
0.32080078125,
0.82275390625,
0.3212890625,
0.82275390625,
0.3359375,
0.1748046875,
0.335205078125
],
[
0.176513671875,
0.3359375,
0.381591796875,
0.3359375,
0.381591796875,
0.349853515625,
0.176513671875,
0.349853515625
],
[
0.214111328125,
0.361083984375,
0.8203125,
0.361083984375,
0.8203125,
0.375,
0.214111328125,
0.375
],
[
0.233642578125,
0.375732421875,
0.82177734375,
0.375732421875,
0.82177734375,
0.39013671875,
0.233642578125,
0.39013671875
],
[
0.2296142578125,
0.38818359375,
0.5390625,
0.39013671875,
0.5390625,
0.410888671875,
0.2296142578125,
0.4091796875
],
[
0.2156982421875,
0.41357421875,
0.82177734375,
0.41357421875,
0.82177734375,
0.427978515625,
0.2156982421875,
0.427978515625
],
[
0.2320556640625,
0.427978515625,
0.80224609375,
0.427978515625,
0.80224609375,
0.44140625,
0.2320556640625,
0.44140625
],
[
0.2156982421875,
0.44580078125,
0.82421875,
0.44580078125,
0.82421875,
0.460205078125,
0.2156982421875,
0.460205078125
],
[
0.233642578125,
0.462158203125,
0.82275390625,
0.462158203125,
0.82275390625,
0.4765625,
0.233642578125,
0.4765625
],
[
0.2320556640625,
0.479248046875,
0.82666015625,
0.479248046875,
0.82666015625,
0.496826171875,
0.2320556640625,
0.496826171875
],
[
0.2344970703125,
0.49560546875,
0.330810546875,
0.49560546875,
0.330810546875,
0.5068359375,
0.2344970703125,
0.5068359375
],
[
0.214111328125,
0.51123046875,
0.82275390625,
0.51220703125,
0.82275390625,
0.5263671875,
0.214111328125,
0.52587890625
],
[
0.2296142578125,
0.5244140625,
0.82666015625,
0.5244140625,
0.82666015625,
0.54541015625,
0.2296142578125,
0.54541015625
],
[
0.2344970703125,
0.54443359375,
0.314453125,
0.54443359375,
0.314453125,
0.55615234375,
0.2344970703125,
0.55615234375
],
[
0.17724609375,
0.57080078125,
0.82177734375,
0.57080078125,
0.82177734375,
0.583984375,
0.17724609375,
0.583984375
],
[
0.17724609375,
0.583984375,
0.82177734375,
0.583984375,
0.82177734375,
0.59765625,
0.17724609375,
0.59765625
],
[
0.1748046875,
0.59521484375,
0.82421875,
0.5947265625,
0.82421875,
0.6123046875,
0.1748046875,
0.61279296875
],
[
0.17724609375,
0.6123046875,
0.82275390625,
0.6123046875,
0.82275390625,
0.62548828125,
0.17724609375,
0.62548828125
],
[
0.1748046875,
0.625,
0.82177734375,
0.62548828125,
0.82177734375,
0.64013671875,
0.1748046875,
0.6396484375
],
[
0.17724609375,
0.63818359375,
0.82275390625,
0.63818359375,
0.82275390625,
0.65283203125,
0.17724609375,
0.65283203125
],
[
0.1748046875,
0.65380859375,
0.82275390625,
0.65380859375,
0.82275390625,
0.66748046875,
0.1748046875,
0.66748046875
],
[
0.17236328125,
0.671875,
0.72802734375,
0.671875,
0.72802734375,
0.6923828125,
0.17236328125,
0.6923828125
],
[
0.71826171875,
0.67529296875,
0.82275390625,
0.67529296875,
0.82275390625,
0.68798828125,
0.71826171875,
0.68798828125
],
[
0.176513671875,
0.693359375,
0.73876953125,
0.693359375,
0.73876953125,
0.70703125,
0.176513671875,
0.70703125
],
[
0.176513671875,
0.7138671875,
0.82275390625,
0.7138671875,
0.82275390625,
0.72900390625,
0.176513671875,
0.72900390625
],
[
0.176513671875,
0.72998046875,
0.82177734375,
0.72998046875,
0.82177734375,
0.744140625,
0.176513671875,
0.744140625
],
[
0.17724609375,
0.744140625,
0.82177734375,
0.744140625,
0.82177734375,
0.75830078125,
0.17724609375,
0.75830078125
],
[
0.17724609375,
0.75830078125,
0.82275390625,
0.75830078125,
0.82275390625,
0.771484375,
0.17724609375,
0.771484375
],
[
0.176513671875,
0.77294921875,
0.81884765625,
0.77294921875,
0.81884765625,
0.78271484375,
0.176513671875,
0.78271484375
],
[
0.176513671875,
0.7861328125,
0.82275390625,
0.7861328125,
0.82275390625,
0.7998046875,
0.176513671875,
0.7998046875
],
[
0.1748046875,
0.8017578125,
0.5,
0.80029296875,
0.5,
0.814453125,
0.1748046875,
0.81494140625
],
[
0.17724609375,
0.82373046875,
0.82275390625,
0.82373046875,
0.82275390625,
0.8369140625,
0.17724609375,
0.8369140625
],
[
0.176513671875,
0.8369140625,
0.82275390625,
0.8369140625,
0.82275390625,
0.8515625,
0.176513671875,
0.8515625
],
[
0.176513671875,
0.8515625,
0.313720703125,
0.8515625,
0.313720703125,
0.86572265625,
0.176513671875,
0.86572265625
],
[
0.193603515625,
0.8720703125,
0.8203125,
0.87353515625,
0.8203125,
0.88818359375,
0.193603515625,
0.88623046875
],
[
0.176513671875,
0.88623046875,
0.73046875,
0.88623046875,
0.73046875,
0.8994140625,
0.176513671875,
0.8994140625
],
[
0.7294921875,
0.88818359375,
0.82177734375,
0.88818359375,
0.82177734375,
0.8984375,
0.7294921875,
0.8984375
],
[
0.176513671875,
0.8994140625,
0.82275390625,
0.8994140625,
0.82275390625,
0.91357421875,
0.176513671875,
0.91357421875
],
[
0.176513671875,
0.912109375,
0.40283203125,
0.91357421875,
0.40283203125,
0.9248046875,
0.176513671875,
0.923828125
],
[
0.49267578125,
0.95068359375,
0.50732421875,
0.95068359375,
0.50732421875,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2023",
"4.1The unifying Algebraic Condition",
"首先，我们提出了一个代数条件，该条件可用于统一SN，AOL和CPL等现有技术的发展。",
"我们的主要定理形式化如下",
"Theorem 1. For any weight matrix W e Rmxn , if there exists a nonsingular diagonal matrix",
"T such that WTw - T = o , then the two following statements hold true.",
"1. Themapping g(r)=WT-α+b is 1-Lipschitz",
"2. The mapping h(r) = - 2WT-lo(WTa + b) is 1 -Lipschitz if g is ReLU, tanh or",
"sigmoid",
"附录中提供了上述定理的证明和一些相关的控制理论解释。该定理允许我们在两种重要情",
"况下使用T的各种选择来设计不同的1-Lipschitz层：对于语句1的线性变换，以及对于语",
"句2的残差和非线性块。此外，对于任何给定的权重矩阵W，条件WTWT在T中是",
"线性的，因此可以视为具有决策变量T的SDP条件。为了强调该定理的重要性，我们建议",
"通过为SDP条件WTW=T选择特定的T来推导用于设计1-Lipschitz层的现有方法。然",
"后自动获取1-Lipschitz属性",
": Spectral Normalization (SN) corresponds to an almost trivial choice if we notice",
"that WTW 3 WTWl2I IIWI2I. Hence with T = IWIl2I , we build the SN",
"layer g(a) = WT-2r +b= WlWa +b .",
".. The Orthogonality-based parameterization is obtained by setting T = I and enforc-",
"ing the equality WTW = T = I . Then obviously g(α) = Wa + b is 1-Lipschitz.",
". AOL formula can be derived by letting T = diag(Zr-ilWTWlz) : With this",
"choice, we have T - WTW E Dn C S\" , hence WTW T. Then Statement",
"1 in Theorem 1 implies that the AOL layer, written as g(c) = WT-2α +b , is",
"1-Lipschitz.3",
": CPL follows the same SN choice T =W2I , but with Statement 2 of Theorem 1",
". Hence we derive a different function h(a) - -Wi Wo(WTr + b) which is also",
"1-Lipschitz",
"上述讨论说明了在同一理论框架内表达所有这些方法的好处，为我们提供了一种新的工具",
"来表征不同方法之间的相似性。例如，SN和CPL共享T=IWII的相同选择。它们之间",
"的区别在于使用哪个语句。因此，CPL可以被视为SN的“剩余版本”。显然，剩余网络结",
"构允许CPL比SN更有效地解决梯度消失问题。使用相同的方法，我们可以很容易地从我",
"们的统一代数条件中推断出基于正交性的参数化和AOL的“残差”对应物是什么。对于基",
"于正交性的参数化，如果我们通过SOC和ECO等方法强制执行WTW=T=I，则函数",
"h(α）=a-2Wα(Wα+b）为1-Lipschitz（通过定理1中的语句2）。最后，如果我们选择",
"T=diag (,= WTWla),那么函数 h(a)=a-2Wdiag(,=1 /WTWla)",
"a(Wia+b)",
"也是1-Lipschitz。因此，从现有的1-Lipschitz网络结构中创建新类是很简单的",
"定理1的另一个重要结果是关于新层开发。SDP条件WTW-T≥0的任何新的非奇异对",
"角解T都会立即以g(α)=WT-+b或h(c)=α-2WT-1α(WTα+b)的形式导致新的",
"1-Lipschitz网络结构。因此，可以将1-Lipschitz网络结构的发展重新表述为寻找具有非奇",
"异对角T的矩阵不等式WTW≥T的解析解。事实上，Gershgorin圆定理可以帮助系统地",
"改进T的现有选择。在F5节中我们将讨论T的此类新选择以及改进CPL的相关应用",
"在这一点上，值得注意的是，要开发深度Lipschitz网络，重要的是要有T的解析公式。T",
"的解析公式将能够快速计算WT-或WT-1",
"定理1在建立1-Lipschitz网络层与代数条件WTW≤T之间的连接方面非常强大。接下",
"来，我们将更仔细地研究这个代数条件，并提供一个新的数学解释，解释AOL如何生成",
"“几乎正交”的权重",
"3For ease of exposition, our main paper always assumes that all the columns of w have at least",
"one non-zero entry such that (2) is well defined. To drop this assumption, we can u",
"se a variant of",
"Theorem 1 which replaces the SDP condition with a bilinear matrix inequality condition. We will",
"discuss this point in the appendix.",
"4"
] |
{
"arxiv_id": "2303.03169",
"language": "zh",
"page_idx": 3
}
|
||
2303.03169-zh-4
|
2303.03169-zh-4.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2023}}$
Remark 1. The proof of Statement 2 in Theorem 1 relies on (?, Lemma 1) , which requires the activation function $\sigma$ to be slope-restricted on [0,1]. Therefore, Statement 2 cannot be applied to the case with $\sigma$ being the GroupSort activation function (?). In contrast, Statement 1 can be used to build neural networks with any activation functions which are 1-Lipschitz.
4.2 A New Mathematical Interpretation for AOL
在？中，可以观察到 AOL 可以学习“几乎正交”的权重，从而克服了梯度消失问题。事实上，在 AOL 中使用的$T$的选择在特定的数学意义上是最优的，正如用下一个定理形式化的那样
Theorem 2. Given any $W\in\mathbb{R}^{m\times n}$ which does not have zero columns, define the set $T=$ $\left\{T:T\text{is nonsingular diagonal, and }T-W^{\mathrm{T}}W\in\boldsymbol{D}^{n}\right\}.$ Then the choice of $T$ for the AOL ${\mathrm{method~actually~satisfies}}$
$$T=\mathrm{diag}(\sum_{j=1}^{n}|W^{\mathrm{T}}W|_{ij})=\arg\min_{T\in T}\mathrm{tr}(I-T^{-\frac{1}{2}}W^{\mathrm{T}}WT^{-\frac{1}{2}})=\arg\min_{T\in T}\|T^{-\frac{1}{2}}W^{\mathrm{T}}WT^{-\frac{1}{2}}-I\|_{F}.$$
我们将上述结果的证明推迟到附录中。在这里，我们为上述结果提供一些解释。显然，数量$\|T^{-\frac12}W^\mathrm{T}WT^{-\frac12}-I\|_F$为缩放后的权重矩阵$WT^{-\frac12}$与$n\times n$正交矩阵集之间的距离提供了度量。如果$\left\|T^{-\frac12}W^\mathrm{T}WT^{-\frac12}-I\right\|_F=0$,则缩放后的权重$WT^-\frac12$ 是正交的。如果$\|T^{-\frac12}W^\mathrm{T}WT^{-\frac12}-I\|_F$很小，则意味着$WT^-\frac12$ “几乎正交”并且接近正交矩阵集。由于我们需要$W^\mathrm{T}W-T\preceq0$ ,因此我们知道$I-T^{-\frac12}W^\mathrm{T}WT^{-\frac12}$是一个正半定矩阵，其轨迹提供了一个替代度量，用于量化$WT^{-\frac12}$与正交矩阵集之间的距离。重要的是，我们有以下不等式：
$$\|T^{-\frac{1}{2}}W^{\mathrm{T}}WT^{-\frac{1}{2}}-I\|_{F}\leq\mathrm{tr}(I-T^{-\frac{1}{2}}W^{\mathrm{T}}WT^{-\frac{1}{2}}).$$
如果 tr$(I-T^{-\frac12}W^\mathrm{T}WT^{-\frac12})$很小，则$\|T^-\frac12W^\mathrm{T}WT^{-\frac12}-I\|_F$也很小，并且$WT^-\frac12$接近正交矩阵的集合。因此，对定理 2 的一种解释是，在满足$T-W^\mathrm{T}W\in D^n$的所有非奇异对角缩放矩阵$T$中，选择 AOL 中使用的$T$使缩放后的权重矩阵$WT^-\frac12$最接近正交矩阵集。这为 AOL 如何生成 “几乎正交”权重提供了新的数学解释
AOL 的一个潜在问题是$D^n$通常比$S_+^n$小得多，并且与定理 1 中的原始条件$T-W^\mathrm{T}W\in S_+^n$ 相比，条件$T-W^\mathrm{T}W\in D^n$可能过于保守。如果我们表示集合$\hat{T}=\begin{cases}T:T\text{is nonsingular diagonal, and }T-W^\mathrm{T}W\in\mathcal{S}_+^n\}&,\text{那么}\end{cases}$ 我们有$\arg\min_{T\in\boldsymbol{T}}$tr$(I-T^{-\frac12}W^\mathrm{T}WT^{-\frac12})\leq\arg\min_{T\in\boldsymbol{T}}$tr$(I-T^{-\frac12}W^\mathrm{T}WT^{-\frac12})$ 和$\arg\min_{T\in\hat{\boldsymbol{T}}}\|T^{-\frac12W^\mathrm{T}WT^{-\frac12}-I\|_F\leq\arg\min_{T\in\boldsymbol{T}}\|T^{-\frac12}W^\mathrm{T}\bar{WT^{-\frac12}}-I\|_F\text{。这导致了 }T\text{ 的}}$ 有趣的替代选择，可以进一步促进正交性：
(3)
$$T=\underset{T\in\hat{\boldsymbol{T}}}{\arg\min}\|T^{-\frac{1}{2}}W^{\mathrm{T}}WT^{-\frac{1}{2}}-I\|_{F}\quad\underset{T\in\hat{\boldsymbol{T}}}{\mathrm{or}}\quad T=\underset{T\in\hat{\boldsymbol{T}}}{\arg\min}\mathop{\mathrm{tr}}(I-T^{-\frac{1}{2}}W^{\mathrm{T}}WT^{-\frac{1}{2}})$$
尽管 (3) 可以作为小玩具示例上的凸程序来解决，但将$T$的这种选择用于大规模问题是不切实际的。我们希望我们上面的理论讨论将激发更多关于开发$T$促进正交性的新实践选择的未来研究
5 Extensions of CPL: The Power of Gershgorin circle theorem
在本节中，我们通过为我们的条件$W^\mathrm{T}W\preceq T$提供新的分析解决方案，将原始 CPL 层 (5) 扩展为新的 1-Lipschitz 网络结构族。我们将这个一般的层系列称为基于 SDP 的 Lipschitz 层(SLL),因为条件$W^\mathrm{T}W\preceq T$可以视为决策变量$T$的 SDP。首先，我们通过使用定理1应用$T$的更一般选择来扩展现有的 CPL (方程 (1))。从定理 1之后的讨论中，我们已经知道我们可以使用$T=$diag$(\sum_j=1^n|W^{\mathrm{T}}W|_{ij})$ 的选择来代替原始选择 $T=\|W\|_2^2I$ 。在本节中，我们将通过$T$的更一般的选择来加强 CPL,该选择基于 Gershgorin 圆定理的特殊版本。具体来说，我们将应用(?,Corollary 6.1.6)来显示以下结果。
5
|
[
[
0.176513671875,
0.0328369140625,
0.50634765625,
0.0328369140625,
0.50634765625,
0.046722412109375,
0.176513671875,
0.046722412109375
],
[
0.176513671875,
0.1060791015625,
0.82177734375,
0.1060791015625,
0.82177734375,
0.11932373046875,
0.176513671875,
0.11932373046875
],
[
0.176513671875,
0.11932373046875,
0.82177734375,
0.11932373046875,
0.82177734375,
0.1331787109375,
0.176513671875,
0.1331787109375
],
[
0.1732177734375,
0.1319580078125,
0.82177734375,
0.1337890625,
0.82177734375,
0.147705078125,
0.1732177734375,
0.1458740234375
],
[
0.1732177734375,
0.1458740234375,
0.82275390625,
0.146484375,
0.82275390625,
0.1610107421875,
0.1732177734375,
0.160400390625
],
[
0.176513671875,
0.1622314453125,
0.2607421875,
0.1622314453125,
0.2607421875,
0.173583984375,
0.176513671875,
0.173583984375
],
[
0.1748046875,
0.1925048828125,
0.5400390625,
0.1925048828125,
0.5400390625,
0.205810546875,
0.1748046875,
0.205810546875
],
[
0.1748046875,
0.2183837890625,
0.82275390625,
0.2183837890625,
0.82275390625,
0.231689453125,
0.1748046875,
0.231689453125
],
[
0.176513671875,
0.23291015625,
0.8203125,
0.23291015625,
0.8203125,
0.246826171875,
0.176513671875,
0.246826171875
],
[
0.176513671875,
0.2474365234375,
0.2254638671875,
0.2474365234375,
0.2254638671875,
0.2587890625,
0.176513671875,
0.2587890625
],
[
0.1748046875,
0.26318359375,
0.81640625,
0.26318359375,
0.81640625,
0.27783203125,
0.1748046875,
0.27783203125
],
[
0.17724609375,
0.27978515625,
0.82275390625,
0.27978515625,
0.82275390625,
0.29345703125,
0.17724609375,
0.29345703125
],
[
0.1748046875,
0.29345703125,
0.3544921875,
0.29345703125,
0.3544921875,
0.306884765625,
0.1748046875,
0.306884765625
],
[
0.17236328125,
0.323974609375,
0.830078125,
0.323974609375,
0.830078125,
0.3447265625,
0.17236328125,
0.3447265625
],
[
0.24267578125,
0.341552734375,
0.271240234375,
0.341552734375,
0.271240234375,
0.35791015625,
0.24267578125,
0.35791015625
],
[
0.378173828125,
0.34033203125,
0.404296875,
0.34033203125,
0.404296875,
0.348388671875,
0.378173828125,
0.348388671875
],
[
0.619140625,
0.341552734375,
0.6435546875,
0.341552734375,
0.6435546875,
0.349853515625,
0.619140625,
0.349853515625
],
[
0.176513671875,
0.373779296875,
0.82275390625,
0.373779296875,
0.82275390625,
0.386962890625,
0.176513671875,
0.386962890625
],
[
0.176513671875,
0.38818359375,
0.82275390625,
0.38818359375,
0.82275390625,
0.40283203125,
0.176513671875,
0.40283203125
],
[
0.17236328125,
0.401611328125,
0.82421875,
0.40283203125,
0.82421875,
0.42041015625,
0.17236328125,
0.419189453125
],
[
0.1732177734375,
0.41845703125,
0.82421875,
0.419189453125,
0.82421875,
0.4375,
0.1732177734375,
0.436279296875
],
[
0.1748046875,
0.43505859375,
0.82421875,
0.436279296875,
0.82421875,
0.453857421875,
0.1748046875,
0.45263671875
],
[
0.176513671875,
0.453857421875,
0.82275390625,
0.453857421875,
0.82275390625,
0.468505859375,
0.176513671875,
0.468505859375
],
[
0.176513671875,
0.466552734375,
0.2156982421875,
0.4697265625,
0.214111328125,
0.481689453125,
0.1748046875,
0.479248046875
],
[
0.32177734375,
0.490478515625,
0.67626953125,
0.490478515625,
0.67626953125,
0.50830078125,
0.32177734375,
0.50830078125
],
[
0.1732177734375,
0.5166015625,
0.82275390625,
0.517578125,
0.82275390625,
0.53515625,
0.1732177734375,
0.5341796875
],
[
0.176513671875,
0.5341796875,
0.82177734375,
0.5341796875,
0.82177734375,
0.54736328125,
0.176513671875,
0.54736328125
],
[
0.17724609375,
0.5498046875,
0.82275390625,
0.5498046875,
0.82275390625,
0.56298828125,
0.17724609375,
0.56298828125
],
[
0.176513671875,
0.56298828125,
0.5986328125,
0.56298828125,
0.5986328125,
0.5771484375,
0.176513671875,
0.5771484375
],
[
0.176513671875,
0.58544921875,
0.82275390625,
0.58544921875,
0.82275390625,
0.5986328125,
0.176513671875,
0.5986328125
],
[
0.176513671875,
0.599609375,
0.82275390625,
0.599609375,
0.82275390625,
0.6142578125,
0.176513671875,
0.6142578125
],
[
0.176513671875,
0.6171875,
0.82275390625,
0.6171875,
0.82275390625,
0.6318359375,
0.176513671875,
0.6318359375
],
[
0.1732177734375,
0.6337890625,
0.82421875,
0.6337890625,
0.82421875,
0.6513671875,
0.1732177734375,
0.6513671875
],
[
0.1732177734375,
0.65087890625,
0.82421875,
0.65087890625,
0.82421875,
0.671875,
0.1732177734375,
0.671875
],
[
0.176513671875,
0.66943359375,
0.47802734375,
0.66943359375,
0.47802734375,
0.68310546875,
0.176513671875,
0.68310546875
],
[
0.2197265625,
0.6923828125,
0.82421875,
0.693359375,
0.82421875,
0.7109375,
0.2197265625,
0.71044921875
],
[
0.26708984375,
0.71044921875,
0.2998046875,
0.71044921875,
0.2998046875,
0.7216796875,
0.26708984375,
0.7216796875
],
[
0.56787109375,
0.7119140625,
0.5947265625,
0.7119140625,
0.5947265625,
0.72021484375,
0.56787109375,
0.72021484375
],
[
0.178955078125,
0.73193359375,
0.8203125,
0.73193359375,
0.8203125,
0.7421875,
0.178955078125,
0.7421875
],
[
0.176513671875,
0.74365234375,
0.82177734375,
0.744140625,
0.82177734375,
0.75830078125,
0.176513671875,
0.7568359375
],
[
0.17724609375,
0.76025390625,
0.2607421875,
0.76025390625,
0.2607421875,
0.771484375,
0.17724609375,
0.771484375
],
[
0.17236328125,
0.79248046875,
0.7158203125,
0.79443359375,
0.7158203125,
0.8095703125,
0.17236328125,
0.806640625
],
[
0.1748046875,
0.82373046875,
0.82275390625,
0.82470703125,
0.82275390625,
0.8388671875,
0.1748046875,
0.83837890625
],
[
0.176513671875,
0.8388671875,
0.8203125,
0.8388671875,
0.8203125,
0.85302734375,
0.176513671875,
0.85302734375
],
[
0.1748046875,
0.85302734375,
0.82177734375,
0.85302734375,
0.82177734375,
0.8662109375,
0.1748046875,
0.8662109375
],
[
0.176513671875,
0.8671875,
0.82177734375,
0.8671875,
0.82177734375,
0.880859375,
0.176513671875,
0.880859375
],
[
0.1748046875,
0.880859375,
0.82177734375,
0.880859375,
0.82177734375,
0.89599609375,
0.1748046875,
0.89599609375
],
[
0.1748046875,
0.89599609375,
0.82275390625,
0.89794921875,
0.82275390625,
0.912109375,
0.1748046875,
0.91015625
],
[
0.1748046875,
0.91015625,
0.65673828125,
0.9111328125,
0.65673828125,
0.9248046875,
0.1748046875,
0.923828125
],
[
0.491943359375,
0.94775390625,
0.50732421875,
0.94775390625,
0.50732421875,
0.96337890625,
0.491943359375,
0.96337890625
]
] |
[
"Published as a conference paper at ICLR 2023",
"Remark 1. The proof of Statement 2 in Theorem 1 relies on (?, Lemma 1) , which requires",
"the activation function to be slope-restricted on [0, 1] . Therefore, Statement 2 cannot",
"be applied to the case with being the GroupSort activation function (?) :In contrast,",
"Statement 1 can be used to build neural networks with any activation functions which are",
"1-Lipschitz.",
"4.2ANewMathematicalInterpretationforAOL",
"在？中，可以观察到AOL可以学习“几乎正交”的权重，从而克服了梯度消失问题。事实",
"上，在AOL中使用的T的选择在特定的数学意义上是最优的，正如用下一个定理形式化",
"的那样",
"Theorem 2. Given any W Rmxn which does not have zero columns, define the set T",
"[T : Tis nonsingular diagonal, and T- WTW E Dn↓ . Then the choice of T for the AOL",
"method actually satisfies",
"T= diag(IWTWli)= arg mintr(I -T--wTWT-) = argmin|T-WTWT-- - IlF.",
"j=1",
"TET",
"TET",
"我们将上述结果的证明推迟到附录中。在这里，我们为上述结果提供一些解释。显然，数",
"量IT-WTWT--IF为缩放后的权重矩阵WT-与n×n正交矩阵集之间的距离",
"提供了度量。如果IT-WTWT--IlF=0，则缩放后的权重WT一是正交的。如果",
"IT-WTWT--IF很小，则意味着WT-“几乎正交”并且接近正交矩阵集。由于我",
"们需要WTW-T≥，因此我们知道I-T-WTWT-是一个正半定矩阵，其轨迹提",
"供了一个替代度量，用于量化WT一与正交矩阵集之间的距离。重要的是，我们有以下不",
"等式：",
"IT-WTWT--IF≤tr(I-T-WTWT-).",
"如果tr(I-T-WTWT-）很小，则IT-WTWT-IF也很小，并且WT-接近正",
"交矩阵的集合。因此，对定理2的一种解释是，在满足T-WTWEDn的所有非奇异对",
"角缩放矩阵T中，选择AOL中使用的T使缩放后的权重矩阵WT一最接近正交矩阵集。",
"这为AOL如何生成“几乎正交”权重提供了新的数学解释",
"AOL的一个潜在问题是Dn通常比S小得多，并且与定理1中的原始",
"条件T-WTWES相比，条件T-WTWEDn可能过于保守。如",
"果我们表示集合T={T：Tisnonsingulardiagonal,andT-WTWES，那么",
"我们有 argminTeTtr(I-T--WTWT-）≤argminTeTtr(I-T-WTWT-）和",
"argminTelT--WTWT--IlF≤argminTeTlT-WTWT--IF。这导致了T的",
"有趣的替代选择，可以进一步促进正交性：",
"T=argminT-WTWT--Il or T=argmintr(I-T--WTWT-)(3)",
"TET",
"TET",
"尽管（3）可以作为小玩具示例上的凸程序来解决，但将T的这种选择用于大规模问题是不",
"切实际的。我们希望我们上面的理论讨论将激发更多关于开发T促进正交性的新实践选择",
"的未来研究",
"5ExtensionsofCPL:ThePowerofGershgorincircletheorem",
"在本节中，我们通过为我们的条件WTW=T提供新的分析解决方案，将原始CPL层（5)",
"扩展为新的1-Lipschitz网络结构族。我们将这个一般的层系列称为基于SDP的Lipschitz",
"层（SLL），因为条件WTW≤T可以视为决策变量T的SDP。首先，我们通过使用定理",
"1应用T的更一般选择来扩展现有的CPL（方程（1））。从定理1之后的讨论中，我们已",
"经知道我们可以使用T=diag(Zi=iIWTWli）的选择来代替原始选择T=IWI。在本",
"节中，我们将通过T的更一般的选择来加强CPL，该选择基于Gershgorin圆定理的特殊",
"版本。具体来说，我们将应用（?，Corollary6.1.6）来显示以下结果。",
"5"
] |
{
"arxiv_id": "2303.03169",
"language": "zh",
"page_idx": 4
}
|
||
2305.08732-zh-1
|
2305.08732-zh-1.jpg
|
训练语料库中的知识。同样，Bhagavatula et al. (2020)通过使用背诵增强生成从模型权重中检索来利用大型语言模型(>10B)中的隐式知识。这些研究肯定了 PLM 封装了大量的知识， 具有未开发的潜力，而在我们的论文中，我们追求一种更普遍适用但简单的解决方案，以充分利用 PLM 中的 NLP 知识
为了满足这一需求，我们引入了知识反刍， 以帮助模型在处理知识密集型任务时进行深思熟虑的思考。类似于动物如何反刍食物以获得更好的消化和吸收-通过将食物从胃中反流回口腔进行额外的咀嚼-我们的目标是通过让模型首先审查存储在其参数中的相关知识，然后整合这些知识来模拟这一过程更好地处理相关任务。详细地说，我们建议通过简单地添加“据我所知”来激励模型回忆潜在知识，从而在任务引导的提示下进行知识回顾。随后，我们通过 FFN 整合知识，明确利用潜在知识来帮助解决下游任务，因为FFN 在 PLMs (Wang et al., 2022) 中起着至关重要的作用
我们将建议的知识反思应用于各种 PLM, 包括 RoBERTa (Liu et al., 2019) , DeBERTa (He et al.,2021)。我们还将知识反刍转移到大语言GPT-3(175B) (Brown et al.,2020) 。在六个常识推理任务和 GLUE 基准上的实验结果表明，所提出的简单方法可以获得性能增益，甚至优于检索外部知识的基线。最后，我们总结了这项工作的贡献如下：
$\cdot$ We propose a novel approach of Knowledge Rumination to better utilize the knowledge stored in the parameters, which is model agnostic and can be applied to any PLMs
$\bullet$ Experimental results demonstrate that the proposed approach can successfully elicit related knowledge for both small and large PLMs, yielding better performance on six commonsense tasks and GLUE benchmarks. $\bullet$ Comprehensive empirical analysis indicates that still a large underestimated amount of knowledge can be retrieved from PLM's model weights, and our work takes a small step in this direction.
### 2 $\textbf{Related Work and Background}$
Extracting Knowledge from PLMs 先前的研究表明，PLM 隐式包含大量知识。Petroni et al. (2019)已经表明，通过将知识库(KB)关系转换为自然语言模板，可以在知识库 (KB)完 我们以一个代表性的任务，多项选择常识推成任务中使用此类语言模型。基于这一发现，理为例，阐明知识反刍的细节，它可以简单
研究人员试图将 PLM 视为知识库。一些研究(Bosselut et al., 2019; West et al., 2022; Hao et al., 2022)使用 PLM 自动构建知识图。同时，其他一些 (Shwartz et al., 2020; Li et al., 2022) 发现 PLM 拥有的知识可用于增强模型在下游任务中的性能。迄今为止，有几项工作$\mathbb{E}$(Wang et al., 2023; Zelikman et al., 2022; Bhagavatula et al., 2020) 尝试利用 PLM 生成自由文本推理以进行推理。我们的方法与以前的工作不同， 我们的目标是增强模型对它已经知道的内容的理解，以提高性能
Knowledge-Enhanced Models 研究人员求助
于外部资源，以促进模型处理知识密集型情况的能力。一个方向是将问题放在 KB 中 ,并使用问题和检索到的知识进行推理 (Yasunaga et al., 2022,2021; Zhang et al., 2022; Sun et al., 2019; Yao et al., 2022; Lv et al., 2020; Lin et al., 2019) 。由于预先训练的模型也可以被视为知识存储， 最近的几项研究，包括 Self-talk(Shwartz et al., 2020)、Rainier (Liu et al., 2022a) 、GKP (Liu et al., 2022b) 、 ElicitKnowledge (Li et al., 2022) ,建议将大型语言模型(e。g。, GPT-3)作为外部来源，为下游任务获取知识。相反，我们的方法与依赖外部资源(如知识库(KB)或语言模型 (LM)) 不同。。相反，我们专注于充分利用模型本身获得的潜在知识。还有一些工作将问题分解为子问题，并要求模型回答每个子问题，例如最小到最大提示$\overline{\mathrm{E}}$(Zhou et al.,2023)。然而，即使这些方法遇到了我们提出的问题， 其中模型可能在其参数内拥有子问题的答案， 但无法提供正确的响应。我们方法背后的基本直觉是，当前利用预训练语言模型(PLM)的强大功能的方法尚未完全利用模型参数中的知识
请注意，我们的方法与 Self-talk (Shwartz et al., 2020) 最接近，但具有管理参数知识的附加功能(例如前馈网络中的嵌入)。这种能力在一定程度上拓宽了学术思想的范围
### 3 Knowledge Rumination
在本节中，我们将介绍知识反刍的技术细节， 以挖掘 PLM (§3.1)的潜力。给定 PLM $G$ ,我们首先冻结模型参数并设计特定于任务的提示$(\S3.2)$,以指导模型审查其存储的有关任务和输入的知识(知识审查)。然后，我们在调整下游任务(知识整合)期间整合模型的潜在知识(§3.3)
## 3.1 Model Architecture
|
[
[
0.12176513671875,
0.0885009765625,
0.483642578125,
0.0885009765625,
0.483642578125,
0.09857177734375,
0.12176513671875,
0.09857177734375
],
[
0.51318359375,
0.0872802734375,
0.8818359375,
0.0872802734375,
0.8818359375,
0.1009521484375,
0.51318359375,
0.1009521484375
],
[
0.118408203125,
0.10272216796875,
0.48779296875,
0.10272216796875,
0.48779296875,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.51123046875,
0.10272216796875,
0.87890625,
0.10272216796875,
0.87890625,
0.11639404296875,
0.51123046875,
0.11639404296875
],
[
0.12005615234375,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.12005615234375,
0.1318359375
],
[
0.51318359375,
0.12054443359375,
0.87890625,
0.12054443359375,
0.87890625,
0.1312255859375,
0.51318359375,
0.1312255859375
],
[
0.118408203125,
0.1341552734375,
0.486083984375,
0.1341552734375,
0.486083984375,
0.1478271484375,
0.118408203125,
0.1478271484375
],
[
0.51318359375,
0.1353759765625,
0.8818359375,
0.1353759765625,
0.8818359375,
0.1484375,
0.51318359375,
0.1484375
],
[
0.118408203125,
0.1519775390625,
0.48779296875,
0.1519775390625,
0.48779296875,
0.1650390625,
0.118408203125,
0.1650390625
],
[
0.51123046875,
0.15087890625,
0.880859375,
0.15087890625,
0.880859375,
0.1644287109375,
0.51123046875,
0.1644287109375
],
[
0.118408203125,
0.16748046875,
0.48779296875,
0.16748046875,
0.48779296875,
0.18115234375,
0.118408203125,
0.18115234375
],
[
0.51318359375,
0.16748046875,
0.880859375,
0.16748046875,
0.880859375,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.118408203125,
0.182861328125,
0.3359375,
0.182861328125,
0.3359375,
0.196533203125,
0.118408203125,
0.196533203125
],
[
0.51123046875,
0.1817626953125,
0.87890625,
0.1817626953125,
0.87890625,
0.1953125,
0.51123046875,
0.1953125
],
[
0.1351318359375,
0.1995849609375,
0.48779296875,
0.1995849609375,
0.48779296875,
0.213134765625,
0.1351318359375,
0.213134765625
],
[
0.51123046875,
0.198974609375,
0.87744140625,
0.198974609375,
0.87744140625,
0.2120361328125,
0.51123046875,
0.2120361328125
],
[
0.1175537109375,
0.215576171875,
0.486083984375,
0.215576171875,
0.486083984375,
0.2286376953125,
0.1175537109375,
0.2286376953125
],
[
0.51123046875,
0.215576171875,
0.880859375,
0.215576171875,
0.880859375,
0.2286376953125,
0.51123046875,
0.2286376953125
],
[
0.118408203125,
0.232177734375,
0.489501953125,
0.232177734375,
0.489501953125,
0.2452392578125,
0.118408203125,
0.2452392578125
],
[
0.51123046875,
0.232177734375,
0.880859375,
0.232177734375,
0.880859375,
0.2452392578125,
0.51123046875,
0.2452392578125
],
[
0.1175537109375,
0.2464599609375,
0.489501953125,
0.248779296875,
0.489501953125,
0.262451171875,
0.1175537109375,
0.260009765625
],
[
0.509765625,
0.2464599609375,
0.66015625,
0.2476806640625,
0.66015625,
0.262451171875,
0.509765625,
0.26123046875
],
[
0.118408203125,
0.26416015625,
0.484375,
0.26416015625,
0.484375,
0.27783203125,
0.118408203125,
0.27783203125
],
[
0.515625,
0.270751953125,
0.87890625,
0.270751953125,
0.87890625,
0.28515625,
0.515625,
0.28515625
],
[
0.118408203125,
0.28076171875,
0.48779296875,
0.28076171875,
0.48779296875,
0.29443359375,
0.118408203125,
0.29443359375
],
[
0.51318359375,
0.289306640625,
0.880859375,
0.289306640625,
0.880859375,
0.302734375,
0.51318359375,
0.302734375
],
[
0.118408203125,
0.29638671875,
0.484375,
0.29638671875,
0.484375,
0.31005859375,
0.118408203125,
0.31005859375
],
[
0.5146484375,
0.3046875,
0.880859375,
0.3046875,
0.880859375,
0.318359375,
0.5146484375,
0.318359375
],
[
0.118408203125,
0.312255859375,
0.484375,
0.312255859375,
0.484375,
0.325927734375,
0.118408203125,
0.325927734375
],
[
0.1175537109375,
0.32666015625,
0.484375,
0.327880859375,
0.484375,
0.341552734375,
0.1175537109375,
0.34033203125
],
[
0.515625,
0.322509765625,
0.87890625,
0.322509765625,
0.87890625,
0.333251953125,
0.515625,
0.333251953125
],
[
0.51318359375,
0.33740234375,
0.8818359375,
0.33740234375,
0.8818359375,
0.350830078125,
0.51318359375,
0.350830078125
],
[
0.118408203125,
0.344482421875,
0.486083984375,
0.344482421875,
0.486083984375,
0.358154296875,
0.118408203125,
0.358154296875
],
[
0.51318359375,
0.352783203125,
0.8818359375,
0.352783203125,
0.8818359375,
0.366455078125,
0.51318359375,
0.366455078125
],
[
0.118408203125,
0.361083984375,
0.48193359375,
0.361083984375,
0.48193359375,
0.371826171875,
0.118408203125,
0.371826171875
],
[
0.51318359375,
0.369384765625,
0.8818359375,
0.369384765625,
0.8818359375,
0.383056640625,
0.51318359375,
0.383056640625
],
[
0.1175537109375,
0.3759765625,
0.486083984375,
0.377197265625,
0.486083984375,
0.391357421875,
0.1175537109375,
0.39013671875
],
[
0.51123046875,
0.38427734375,
0.87890625,
0.38427734375,
0.87890625,
0.39794921875,
0.51123046875,
0.39794921875
],
[
0.118408203125,
0.394287109375,
0.39208984375,
0.394287109375,
0.39208984375,
0.405029296875,
0.118408203125,
0.405029296875
],
[
0.51123046875,
0.402099609375,
0.87890625,
0.402099609375,
0.87890625,
0.4150390625,
0.51123046875,
0.4150390625
],
[
0.1351318359375,
0.4091796875,
0.484375,
0.407958984375,
0.484375,
0.4228515625,
0.1351318359375,
0.42333984375
],
[
0.51318359375,
0.41748046875,
0.87890625,
0.41748046875,
0.87890625,
0.43115234375,
0.51318359375,
0.43115234375
],
[
0.118408203125,
0.42578125,
0.486083984375,
0.42578125,
0.486083984375,
0.439453125,
0.118408203125,
0.439453125
],
[
0.5283203125,
0.43408203125,
0.87744140625,
0.43408203125,
0.87744140625,
0.448974609375,
0.5283203125,
0.448974609375
],
[
0.1175537109375,
0.441162109375,
0.483642578125,
0.441162109375,
0.483642578125,
0.454833984375,
0.1175537109375,
0.454833984375
],
[
0.51318359375,
0.449462890625,
0.880859375,
0.449462890625,
0.880859375,
0.463134765625,
0.51318359375,
0.463134765625
],
[
0.1175537109375,
0.457763671875,
0.483642578125,
0.457763671875,
0.483642578125,
0.471435546875,
0.1175537109375,
0.471435546875
],
[
0.51318359375,
0.466064453125,
0.87890625,
0.466064453125,
0.87890625,
0.479736328125,
0.51318359375,
0.479736328125
],
[
0.1175537109375,
0.473876953125,
0.483642578125,
0.473876953125,
0.483642578125,
0.48681640625,
0.1175537109375,
0.48681640625
],
[
0.51318359375,
0.482177734375,
0.87890625,
0.482177734375,
0.87890625,
0.495361328125,
0.51318359375,
0.495361328125
],
[
0.118408203125,
0.490478515625,
0.486083984375,
0.490478515625,
0.486083984375,
0.50439453125,
0.118408203125,
0.50439453125
],
[
0.5146484375,
0.49755859375,
0.880859375,
0.49755859375,
0.880859375,
0.51123046875,
0.5146484375,
0.51123046875
],
[
0.1175537109375,
0.505859375,
0.486083984375,
0.505859375,
0.486083984375,
0.51953125,
0.1175537109375,
0.51953125
],
[
0.5146484375,
0.51416015625,
0.880859375,
0.51416015625,
0.880859375,
0.52783203125,
0.5146484375,
0.52783203125
],
[
0.1175537109375,
0.5224609375,
0.25439453125,
0.5224609375,
0.25439453125,
0.5361328125,
0.1175537109375,
0.5361328125
],
[
0.51318359375,
0.529296875,
0.8818359375,
0.52978515625,
0.8818359375,
0.54443359375,
0.51318359375,
0.54345703125
],
[
0.1351318359375,
0.5498046875,
0.486083984375,
0.5498046875,
0.486083984375,
0.5634765625,
0.1351318359375,
0.5634765625
],
[
0.51318359375,
0.54638671875,
0.8818359375,
0.54638671875,
0.8818359375,
0.56005859375,
0.51318359375,
0.56005859375
],
[
0.15283203125,
0.56640625,
0.48779296875,
0.56787109375,
0.48779296875,
0.58056640625,
0.15283203125,
0.580078125
],
[
0.51318359375,
0.5625,
0.8818359375,
0.5625,
0.8818359375,
0.576171875,
0.51318359375,
0.576171875
],
[
0.51123046875,
0.57666015625,
0.880859375,
0.57763671875,
0.880859375,
0.59130859375,
0.51123046875,
0.59033203125
],
[
0.154541015625,
0.5830078125,
0.489501953125,
0.5830078125,
0.489501953125,
0.5966796875,
0.154541015625,
0.5966796875
],
[
0.154541015625,
0.5986328125,
0.438232421875,
0.5986328125,
0.438232421875,
0.6123046875,
0.154541015625,
0.6123046875
],
[
0.51318359375,
0.59423828125,
0.880859375,
0.59423828125,
0.880859375,
0.60791015625,
0.51318359375,
0.60791015625
],
[
0.51318359375,
0.60986328125,
0.880859375,
0.60986328125,
0.880859375,
0.62353515625,
0.51318359375,
0.62353515625
],
[
0.1376953125,
0.626953125,
0.48779296875,
0.626953125,
0.48779296875,
0.64013671875,
0.1376953125,
0.64013671875
],
[
0.51318359375,
0.62744140625,
0.5341796875,
0.62744140625,
0.5341796875,
0.64013671875,
0.51318359375,
0.64013671875
],
[
0.154541015625,
0.64453125,
0.48779296875,
0.64453125,
0.48779296875,
0.65771484375,
0.154541015625,
0.65771484375
],
[
0.53076171875,
0.642578125,
0.880859375,
0.64111328125,
0.880859375,
0.65625,
0.53076171875,
0.65673828125
],
[
0.15283203125,
0.65771484375,
0.48779296875,
0.66015625,
0.48779296875,
0.6748046875,
0.15283203125,
0.6728515625
],
[
0.51318359375,
0.6591796875,
0.87890625,
0.6591796875,
0.87890625,
0.6728515625,
0.51318359375,
0.6728515625
],
[
0.15283203125,
0.67578125,
0.48779296875,
0.67578125,
0.48779296875,
0.689453125,
0.15283203125,
0.689453125
],
[
0.5146484375,
0.67578125,
0.880859375,
0.67578125,
0.880859375,
0.689453125,
0.5146484375,
0.689453125
],
[
0.154541015625,
0.6923828125,
0.483642578125,
0.6923828125,
0.483642578125,
0.7060546875,
0.154541015625,
0.7060546875
],
[
0.5146484375,
0.69140625,
0.8251953125,
0.69140625,
0.8251953125,
0.705078125,
0.5146484375,
0.705078125
],
[
0.1376953125,
0.72021484375,
0.489501953125,
0.72021484375,
0.489501953125,
0.73388671875,
0.1376953125,
0.73388671875
],
[
0.509765625,
0.7197265625,
0.75146484375,
0.7197265625,
0.75146484375,
0.73291015625,
0.509765625,
0.73291015625
],
[
0.15283203125,
0.736328125,
0.48779296875,
0.736328125,
0.48779296875,
0.75048828125,
0.15283203125,
0.75048828125
],
[
0.51318359375,
0.74462890625,
0.880859375,
0.74462890625,
0.880859375,
0.75830078125,
0.51318359375,
0.75830078125
],
[
0.15283203125,
0.7529296875,
0.48779296875,
0.7529296875,
0.48779296875,
0.7666015625,
0.15283203125,
0.7666015625
],
[
0.5146484375,
0.76123046875,
0.880859375,
0.76123046875,
0.880859375,
0.77490234375,
0.5146484375,
0.77490234375
],
[
0.154541015625,
0.7685546875,
0.48779296875,
0.7685546875,
0.48779296875,
0.7822265625,
0.154541015625,
0.7822265625
],
[
0.51318359375,
0.775390625,
0.87744140625,
0.775390625,
0.87744140625,
0.7890625,
0.51318359375,
0.7890625
],
[
0.15283203125,
0.78515625,
0.30908203125,
0.78369140625,
0.30908203125,
0.798828125,
0.15283203125,
0.7998046875
],
[
0.51318359375,
0.79345703125,
0.880859375,
0.79345703125,
0.880859375,
0.80712890625,
0.51318359375,
0.80712890625
],
[
0.1158447265625,
0.81396484375,
0.42236328125,
0.81396484375,
0.42236328125,
0.82763671875,
0.1158447265625,
0.82763671875
],
[
0.51123046875,
0.80859375,
0.87890625,
0.80859375,
0.87890625,
0.822265625,
0.51123046875,
0.822265625
],
[
0.5146484375,
0.82373046875,
0.87890625,
0.82470703125,
0.87890625,
0.8388671875,
0.5146484375,
0.83837890625
],
[
0.12005615234375,
0.84130859375,
0.486083984375,
0.84130859375,
0.486083984375,
0.85498046875,
0.12005615234375,
0.85498046875
],
[
0.5146484375,
0.8427734375,
0.59619140625,
0.8427734375,
0.59619140625,
0.85400390625,
0.5146484375,
0.85400390625
],
[
0.118408203125,
0.85693359375,
0.484375,
0.85693359375,
0.484375,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.509765625,
0.8662109375,
0.7138671875,
0.86767578125,
0.7138671875,
0.8818359375,
0.509765625,
0.88134765625
],
[
0.118408203125,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.1175537109375,
0.8896484375,
0.486083984375,
0.8896484375,
0.486083984375,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.51123046875,
0.8896484375,
0.87890625,
0.8896484375,
0.87890625,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.12005615234375,
0.90625,
0.490234375,
0.90625,
0.490234375,
0.91943359375,
0.12005615234375,
0.91943359375
],
[
0.51123046875,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91943359375,
0.51123046875,
0.91943359375
]
] |
[
"训练语料库中的知识。同样，Bhagavatulaetal.",
"研究人员试图将PLM视为知识库。一些研究",
"(2020)通过使用背诵增强生成从模型权重中检",
"(Bosselut et al., 2019; West et al., 2022; Hao et al.",
"索来利用大型语言模型（>10B）中的隐式知",
"2022）使用PLM自动构建知识图。同时，其",
"识。这些研究肯定了PLM封装了大量的知识，",
"他一些(Shwartzet al.,2020;Lietal.,2022)发",
"具有未开发的潜力，而在我们的论文中，我们",
"现PLM拥有的知识可用于增强模型在下游任",
"追求一种更普遍适用但简单的解决方案，以充",
"务中的性能。迄今为止，有几项工作El(Wang",
"分利用PLM中的NLP知识",
"et al., 2023; Zelikman et al., 2022; Bhagavatula",
"为了满足这一需求，我们引人了知识反，",
"etal.,2020)尝试利用PLM生成自由文本推理",
"以帮助模型在处理知识密集型任务时进行深",
"以进行推理。我们的方法与以前的工作不同，",
"思熟虑的思考。类似于动物如何反台食物以获",
"我们的目标是增强模型对它已经知道的内容的",
"得更好的消化和吸收-通过将食物从胃中反流",
"理解，以提高性能",
"回口腔进行额外的咀嚼-我们的目标是通过让",
"Knowledge-EnhancedModels研究人员求助",
"模型首先审查存储在其参数中的相关知识，然",
"于外部资源，以促进模型处理知识密集型情况",
"后整合这些知识来模拟这一过程更好地处理相",
"的能力。一个方向是将问题放在KB中，并使用",
"关任务。详细地说，我们建议通过简单地添加",
"“据我所知”来激励模型回忆潜在知识，从而",
"问题和检索到的知识进行推理（Yasunagaetal.",
"2022, 2021; Zhang et al., 2022; Sun et al., 2019;",
"在任务引导的提示下进行知识回顾。随后，我",
"Yao et al., 2022; Lv et al., 2020; Lin et al., 2019) ",
"们通过FFN整合知识，明确利用潜在知识来",
"由于预先训练的模型也可以被视为知识存储，",
"帮助解决下游任务，因为FFN在PLMs（Wang",
"最近的几项研究，包括Self-talk（Shwartzetal.",
"etal.，2022）中起着至关重要的作用",
"2020)Rainier (Liu et al., 2022a)GKP (Liu",
"我们将建议的知识反思应用于各种PLM",
"et al., 2022b)、ElicitKnowledge (Li et al., 2022)",
"包括RoBERTa(Liu et al.,2019)，DeBERTa(He",
"建议将大型语言模型(e。g。，GPT-3)作为外",
"etal.,2021）。我们还将知识反转移到大语言",
"部来源，为下游任务获取知识。相反，我们的",
"GPT-3(175B）（Brownetal.,2020）。在六个常识",
"方法与依赖外部资源（如知识库（KB）或语言",
"推理任务和GLUE基准上的实验结果表明，所",
"模型（LM)）不同。。相反，我们专注于充分利",
"提出的简单方法可以获得性能增益，甚至优于",
"用模型本身获得的潜在知识。还有一些工作将",
"检索外部知识的基线。最后，我们总结了这项",
"问题分解为子问题，并要求模型回答每个子问",
"工作的贡献如下：",
"题，例如最小到最大提示E(Zhouetal.,2023）。",
": We propose a novel approach of Knowledge",
"然而，即使这些方法遇到了我们提出的问题，",
"Rumination to better utilize the knowledge",
"其中模型可能在其参数内拥有子问题的答案，",
"但无法提供正确的响应。我们方法背后的基本",
"stored in the parameters, which is model ag-",
"nostic and can be applied to any PLMs",
"直觉是，当前利用预训练语言模型（PLM）的",
"强大功能的方法尚未完全利用模型参数中的知",
". Experimental results demonstrate that the",
"识",
"proposed approach can successfully elicit re-",
"请注意，我们的方法与Self-talk（Shwartz",
"lated knowledge for both small and large",
"etal.，2020）最接近，但具有管理参数知识的",
"PLMs, yielding better performance on six",
"附加功能（例如前馈网络中的嵌入）。这种能",
"commonsense tasks and GLUE benchmarks.",
"力在一定程度上拓宽了学术思想的范围",
": Comprehensive empirical analysis indicates",
"3KnowledgeRumination",
"that still a large underestimated amount of",
"在本节中，我们将介绍知识反的技术细节，",
"knowledge can be retrieved from PLM's",
"以挖掘PLM（S3.1）的潜力。给定PLMG，我",
"model weights, and our work takes a small",
"们首先冻结模型参数并设计特定于任务的提示",
"step in this direction.",
"（s3.2），以指导模型审查其存储的有关任务和",
"2Related Work and Background",
"输入的知识（知识审查）。然后，我们在调整",
"下游任务（知识整合）期间整合模型的潜在知",
"ExtractingKnowledgefromPLMs先前的研",
"识（83.3）",
"究表明，PLM隐式包含大量知识。Petronietal.",
"3.1 Model Architecture",
"(2019）已经表明，通过将知识库（KB）关系转",
"换为自然语言模板，可以在知识库（KB）完",
"我们以一个代表性的任务，多项选择常识推",
"成任务中使用此类语言模型。基于这一发现，",
"理为例，阐明知识反的细节，它可以简单"
] |
{
"arxiv_id": "2305.08732",
"language": "zh",
"page_idx": 1
}
|
||
2305.08732-zh-3
|
2305.08732-zh-3.jpg
|
用“[CLS]”的输出作为分数(以下等式中的 FFN 中的两个线性层模拟为键值网络 $K$ 和$V$ $f_{\mathrm{cls}}$)来计算每个提及$m\in M$的相关性得分： , 我们采用两个不同的线性层将信息$r$投影到
(5)
$$\rho_m=f_{\mathrm{cls}}\left(G_{\mathrm{enc}}([q;m])\right)$$
我们以前 2 个相关性得分$\rho$作为目标提及。“提及相关性评分”的详情见附录 A.5。实际上， 除了这些常识或上下文知识之外，重要的是要注意 PLM 还存储其他类型的知识，包括技能知识和任务知识$\overline{\mathrm{E}}($Wang et al.,2022)。因此， 我们假设模型可以获取有关任务本身的潜在知识，因此我们进一步构建任务提示以鼓励模型反映其参数中编码的技能知识
总的来说，我们使用三种不同类型的任务引
导提示以及三个感兴趣的领域来探测 PLM:
$\bullet$ Background Prompt: is designed to help the model to think about the background, the prompt is As far as I know [MASK].
$\cdot$ Mention Prompt: is used to elicit memories
of mentions, the formation is About <Men-
tion>, I know [MASK].
$\bullet$ Task Prompt: is designed to help the model
reminisce memories of the task. For exam-
ple, for the sentiment analysis, the prompt is
About sentiment analysis, I know [MASK]
我们在任务引导提示中放置了几个“[MASK]”, 并且“[MASK]”的长度是不同任务的超参数
我们的方法可以应用于不同的 PLM。对于像 RoBERTa 这样的编码器样式模型，我们利用“[MASK]”的隐藏状态$h_{[\mathrm{MASK}]}$作为潜在的知识来进行反思。以下等式中的$f_\mathrm{mask}$表示从模型的输出中获取$h_{[\mathrm{MASK}]}$ 。
$$r=f_{\mathrm{mask}}(G_{\theta}\left([q;p_{q}]\right))$$
(6)
以下部分将说明如何利用这些知识。
## 3.3 Knowledge Consolidation with FFN Neuron Augmentation
为了加强其理解，该模型应重新消化(注入)$q$ 的知识$r$ ,类似于动物如何再次咀嚼食物。然而，在哪里注人 PLMs 仍然是一个具有挑战性的问题，表明潜在的工作来调查如何保持模型的知识。先前的研究$\overline{\mathrm{E}}($Dai et al., 2022; Wang et al.,2022)发现前馈网络(FFN)作为知识神经元或技能神经元起作用，说明 FFN 可以存储事实信息并编码特定于任务的技能。受这些发现的启发，我们将$r$合并到 FFN 中，就像以前的工作(Yao et al.,2022)一样。在这里，我们选择 Top-1 层来重新测试(注人)知识。假设
匹配层的向量空间：
(7) (8)
$$\begin{aligned}\phi_k&=W_k\cdot r\\\phi_v&=W_v\cdot r\end{aligned}$$
,其中$W_k$和$W_v$表示两个线性层的权重(XMATHX 的中间是 XMATHX\_。两个矩阵$W_k$ 和$W_v$随机初始化，并将在训练期间更新。我们通过将投影的知识连接到线性层的末尾来扩展 FFN, 并获得扩展的$K_E,V_E$ 。计算可以描述如下：
$$\begin{array}{c}FFN(H)=f(H\cdot K_E)\cdot V_E\\=f(H\cdot[\phi_k:K])\cdot[\phi_v:V]\end{array}$$
(9)
$H$表示自关注模块的输出隐藏状态。该模型将
在反刍知识的帮助下回答这个问题
# 4 Experiments
## 4.1 Dataset
我们针对常识推理基准的六个知识密集型任务评估了建议的方法：CommonsenseQA (THQA) $\left(\mathrm{Talmor~et~al.,2019)~,~SocialIQA~(Sap~et~al.,2019)}\right)$ , PhysicalQA (PIQA)。我们遵循先前作品(Ya-1) sunaga et al.,2022) 使用的数据拆分。同时，为了更好地理解任务提示在我们的方法中的作用以及在预训练语言模型中学习的技能知识E (Wang et al.,2022),我们考虑了GLUE 基准$\overline{\mathrm{F}}$ (Wang et al., 2018) 的任务 , 包括单句任务 ( SST, CoLA),推理任务(QNLI,RTE)以及相似性和释义任务(STS-B,MRPC)。我们在附录 A.7中提供数据集详细信息
# 4.2 Baselines
我们选择 RoBERTa\_large (Liu et al., 2019) 和DeBERTa\_large$\boxed{\mathrm{F}}($He et al., 2021) 作为中型语言模型的骨干模型，并将性能与以下强大的外部知识增强方法进行比较：QA-GNN (Yasunaga et al., 2021) , GreaseLM (Zhang et al., 2022) 和DragonXMATHX\_。有关基线模型的更多详细信息，请参见附录 A.3.2。对于 GLUE 基准测试中的任务，除了 RoBERTa\_large 之外，我们还将我们的模型与提示学习方法 LM-BFFE (Gao et al., 2021)进行了比较。LM-BFF 提出了一种基于提示的微调方法和一种改进的策略，用于动态和选择性地将演示合并到每个上下文中。在我们的论文中，我们简单地使用人类策划的提示(由 LM-BFF 提供),并利用RoBERT$a_l$arge作为骨干
|
[
[
0.118408203125,
0.0872802734375,
0.484375,
0.0872802734375,
0.484375,
0.097412109375,
0.118408203125,
0.097412109375
],
[
0.490234375,
0.08905029296875,
0.50732421875,
0.08905029296875,
0.50732421875,
0.0968017578125,
0.490234375,
0.0968017578125
],
[
0.509765625,
0.0885009765625,
0.8759765625,
0.0885009765625,
0.8759765625,
0.09857177734375,
0.509765625,
0.09857177734375
],
[
0.12005615234375,
0.10272216796875,
0.48193359375,
0.10272216796875,
0.48193359375,
0.11639404296875,
0.12005615234375,
0.11639404296875
],
[
0.5146484375,
0.105712890625,
0.53076171875,
0.105712890625,
0.53076171875,
0.1134033203125,
0.5146484375,
0.1134033203125
],
[
0.52880859375,
0.10272216796875,
0.87890625,
0.10272216796875,
0.87890625,
0.11639404296875,
0.52880859375,
0.11639404296875
],
[
0.5146484375,
0.1181640625,
0.66748046875,
0.1181640625,
0.66748046875,
0.1318359375,
0.5146484375,
0.1318359375
],
[
0.208251953125,
0.1312255859375,
0.391357421875,
0.1312255859375,
0.391357421875,
0.1455078125,
0.208251953125,
0.1455078125
],
[
0.405517578125,
0.1341552734375,
0.453369140625,
0.135986328125,
0.453369140625,
0.1436767578125,
0.405517578125,
0.141357421875
],
[
0.4609375,
0.133056640625,
0.486083984375,
0.133056640625,
0.486083984375,
0.1455078125,
0.4609375,
0.1455078125
],
[
0.6337890625,
0.15087890625,
0.72705078125,
0.15087890625,
0.72705078125,
0.162109375,
0.6337890625,
0.162109375
],
[
0.8564453125,
0.149658203125,
0.8818359375,
0.149658203125,
0.8818359375,
0.1644287109375,
0.8564453125,
0.1644287109375
],
[
0.1175537109375,
0.1602783203125,
0.483642578125,
0.1602783203125,
0.483642578125,
0.17333984375,
0.1175537109375,
0.17333984375
],
[
0.63037109375,
0.169189453125,
0.72900390625,
0.16748046875,
0.72900390625,
0.1817626953125,
0.63037109375,
0.18408203125
],
[
0.85791015625,
0.169189453125,
0.8818359375,
0.169189453125,
0.8818359375,
0.18408203125,
0.85791015625,
0.18408203125
],
[
0.12005615234375,
0.1776123046875,
0.48193359375,
0.1776123046875,
0.48193359375,
0.188232421875,
0.12005615234375,
0.188232421875
],
[
0.118408203125,
0.193603515625,
0.484375,
0.193603515625,
0.484375,
0.207275390625,
0.118408203125,
0.207275390625
],
[
0.5224609375,
0.204833984375,
0.52880859375,
0.204833984375,
0.52880859375,
0.208984375,
0.5224609375,
0.208984375
],
[
0.53076171875,
0.20068359375,
0.880859375,
0.20068359375,
0.880859375,
0.21435546875,
0.53076171875,
0.21435546875
],
[
0.118408203125,
0.208984375,
0.48779296875,
0.208984375,
0.48779296875,
0.22265625,
0.118408203125,
0.22265625
],
[
0.51123046875,
0.2161865234375,
0.87890625,
0.2161865234375,
0.87890625,
0.2298583984375,
0.51123046875,
0.2298583984375
],
[
0.118408203125,
0.2257080078125,
0.48779296875,
0.2257080078125,
0.48779296875,
0.2392578125,
0.118408203125,
0.2392578125
],
[
0.5146484375,
0.2340087890625,
0.880859375,
0.2340087890625,
0.880859375,
0.24462890625,
0.5146484375,
0.24462890625
],
[
0.118408203125,
0.2410888671875,
0.486083984375,
0.2410888671875,
0.486083984375,
0.254638671875,
0.118408203125,
0.254638671875
],
[
0.51318359375,
0.2493896484375,
0.8759765625,
0.2493896484375,
0.8759765625,
0.260009765625,
0.51318359375,
0.260009765625
],
[
0.12005615234375,
0.25830078125,
0.486083984375,
0.25830078125,
0.486083984375,
0.271484375,
0.12005615234375,
0.271484375
],
[
0.51123046875,
0.26416015625,
0.87890625,
0.26416015625,
0.87890625,
0.27783203125,
0.51123046875,
0.27783203125
],
[
0.118408203125,
0.273681640625,
0.35693359375,
0.273681640625,
0.35693359375,
0.287353515625,
0.118408203125,
0.287353515625
],
[
0.51318359375,
0.28076171875,
0.57666015625,
0.28076171875,
0.57666015625,
0.295654296875,
0.51318359375,
0.295654296875
],
[
0.1368408203125,
0.289306640625,
0.483642578125,
0.289306640625,
0.483642578125,
0.302734375,
0.1368408203125,
0.302734375
],
[
0.12005615234375,
0.305908203125,
0.459228515625,
0.305908203125,
0.459228515625,
0.319580078125,
0.12005615234375,
0.319580078125
],
[
0.5810546875,
0.30810546875,
0.81201171875,
0.31005859375,
0.81201171875,
0.32470703125,
0.5810546875,
0.322509765625
],
[
0.83642578125,
0.32177734375,
0.85791015625,
0.32177734375,
0.85791015625,
0.328857421875,
0.83642578125,
0.328857421875
],
[
0.85400390625,
0.319580078125,
0.880859375,
0.319580078125,
0.880859375,
0.33203125,
0.85400390625,
0.33203125
],
[
0.59375,
0.32666015625,
0.8154296875,
0.32666015625,
0.8154296875,
0.34033203125,
0.59375,
0.34033203125
],
[
0.1376953125,
0.333251953125,
0.486083984375,
0.333251953125,
0.486083984375,
0.3466796875,
0.1376953125,
0.3466796875
],
[
0.15283203125,
0.3486328125,
0.484375,
0.3486328125,
0.484375,
0.3623046875,
0.15283203125,
0.3623046875
],
[
0.5146484375,
0.358154296875,
0.880859375,
0.358154296875,
0.880859375,
0.371826171875,
0.5146484375,
0.371826171875
],
[
0.15283203125,
0.365234375,
0.4248046875,
0.365234375,
0.4248046875,
0.37890625,
0.15283203125,
0.37890625
],
[
0.51123046875,
0.37353515625,
0.7900390625,
0.37353515625,
0.7900390625,
0.38720703125,
0.51123046875,
0.38720703125
],
[
0.139404296875,
0.391357421875,
0.486083984375,
0.391357421875,
0.486083984375,
0.405029296875,
0.139404296875,
0.405029296875
],
[
0.154541015625,
0.407958984375,
0.490234375,
0.407958984375,
0.490234375,
0.421630859375,
0.154541015625,
0.421630859375
],
[
0.509765625,
0.40380859375,
0.6591796875,
0.40380859375,
0.6591796875,
0.418701171875,
0.509765625,
0.418701171875
],
[
0.49951171875,
0.410888671875,
0.509765625,
0.40625,
0.5146484375,
0.410888671875,
0.50390625,
0.416259765625
],
[
0.154541015625,
0.424560546875,
0.330078125,
0.424560546875,
0.330078125,
0.438232421875,
0.154541015625,
0.438232421875
],
[
0.51123046875,
0.427490234375,
0.62060546875,
0.429931640625,
0.62060546875,
0.44482421875,
0.51123046875,
0.4423828125
],
[
0.1376953125,
0.45068359375,
0.48779296875,
0.45068359375,
0.48779296875,
0.46435546875,
0.1376953125,
0.46435546875
],
[
0.51318359375,
0.45361328125,
0.87744140625,
0.45361328125,
0.87744140625,
0.46435546875,
0.51318359375,
0.46435546875
],
[
0.151123046875,
0.46728515625,
0.490234375,
0.465576171875,
0.490234375,
0.478515625,
0.151123046875,
0.48095703125
],
[
0.5146484375,
0.468505859375,
0.8818359375,
0.468505859375,
0.8818359375,
0.482177734375,
0.5146484375,
0.482177734375
],
[
0.154541015625,
0.482666015625,
0.489501953125,
0.482666015625,
0.489501953125,
0.496337890625,
0.154541015625,
0.496337890625
],
[
0.51123046875,
0.48388671875,
0.87890625,
0.48388671875,
0.87890625,
0.49755859375,
0.51123046875,
0.49755859375
],
[
0.154541015625,
0.498779296875,
0.47265625,
0.498779296875,
0.47265625,
0.5126953125,
0.154541015625,
0.5126953125
],
[
0.51708984375,
0.5029296875,
0.87744140625,
0.5029296875,
0.87744140625,
0.51318359375,
0.51708984375,
0.51318359375
],
[
0.51318359375,
0.51708984375,
0.87890625,
0.5166015625,
0.87890625,
0.52978515625,
0.51318359375,
0.53076171875
],
[
0.118408203125,
0.525390625,
0.48779296875,
0.525390625,
0.48779296875,
0.5390625,
0.118408203125,
0.5390625
],
[
0.5146484375,
0.533203125,
0.880859375,
0.533203125,
0.880859375,
0.54638671875,
0.5146484375,
0.54638671875
],
[
0.118408203125,
0.54150390625,
0.478515625,
0.54150390625,
0.478515625,
0.5546875,
0.118408203125,
0.5546875
],
[
0.51318359375,
0.548828125,
0.8818359375,
0.548828125,
0.8818359375,
0.5625,
0.51318359375,
0.5625
],
[
0.1351318359375,
0.55712890625,
0.483642578125,
0.55712890625,
0.483642578125,
0.57080078125,
0.1351318359375,
0.57080078125
],
[
0.509765625,
0.56396484375,
0.8818359375,
0.5634765625,
0.8818359375,
0.57763671875,
0.509765625,
0.5791015625
],
[
0.118408203125,
0.57470703125,
0.486083984375,
0.57470703125,
0.486083984375,
0.58837890625,
0.118408203125,
0.58837890625
],
[
0.51123046875,
0.58056640625,
0.880859375,
0.58056640625,
0.880859375,
0.59423828125,
0.51123046875,
0.59423828125
],
[
0.1175537109375,
0.5888671875,
0.486083984375,
0.59033203125,
0.486083984375,
0.60498046875,
0.1175537109375,
0.60400390625
],
[
0.51123046875,
0.595703125,
0.8818359375,
0.5966796875,
0.8818359375,
0.61083984375,
0.51123046875,
0.60986328125
],
[
0.1175537109375,
0.60498046875,
0.486083984375,
0.60546875,
0.486083984375,
0.62060546875,
0.1175537109375,
0.619140625
],
[
0.51318359375,
0.61083984375,
0.8818359375,
0.6123046875,
0.8818359375,
0.626953125,
0.51318359375,
0.6259765625
],
[
0.1158447265625,
0.6181640625,
0.366943359375,
0.62255859375,
0.366943359375,
0.64013671875,
0.1158447265625,
0.6357421875
],
[
0.51318359375,
0.62744140625,
0.69775390625,
0.62890625,
0.69775390625,
0.6435546875,
0.51318359375,
0.642578125
],
[
0.2149658203125,
0.65087890625,
0.38623046875,
0.65185546875,
0.38623046875,
0.66650390625,
0.2149658203125,
0.6650390625
],
[
0.46337890625,
0.65283203125,
0.48779296875,
0.65283203125,
0.48779296875,
0.6650390625,
0.46337890625,
0.6650390625
],
[
0.51123046875,
0.65625,
0.63232421875,
0.65673828125,
0.63232421875,
0.67138671875,
0.51123046875,
0.67041015625
],
[
0.1175537109375,
0.6796875,
0.399658203125,
0.6796875,
0.399658203125,
0.693359375,
0.1175537109375,
0.693359375
],
[
0.51123046875,
0.6787109375,
0.87890625,
0.6787109375,
0.87890625,
0.6923828125,
0.51123046875,
0.6923828125
],
[
0.51318359375,
0.6953125,
0.8818359375,
0.6953125,
0.8818359375,
0.708984375,
0.51318359375,
0.708984375
],
[
0.1158447265625,
0.70703125,
0.44580078125,
0.70703125,
0.44580078125,
0.72021484375,
0.1158447265625,
0.72021484375
],
[
0.515625,
0.7119140625,
0.880859375,
0.7119140625,
0.880859375,
0.7255859375,
0.515625,
0.7255859375
],
[
0.156982421875,
0.7236328125,
0.3359375,
0.7236328125,
0.3359375,
0.73681640625,
0.156982421875,
0.73681640625
],
[
0.51123046875,
0.72705078125,
0.8818359375,
0.72802734375,
0.8818359375,
0.7421875,
0.51123046875,
0.7412109375
],
[
0.1175537109375,
0.74365234375,
0.486083984375,
0.74365234375,
0.486083984375,
0.75732421875,
0.1175537109375,
0.75732421875
],
[
0.51318359375,
0.74462890625,
0.880859375,
0.74462890625,
0.880859375,
0.75830078125,
0.51318359375,
0.75830078125
],
[
0.1175537109375,
0.7587890625,
0.486083984375,
0.7587890625,
0.486083984375,
0.7724609375,
0.1175537109375,
0.7724609375
],
[
0.51318359375,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.118408203125,
0.77685546875,
0.484375,
0.77685546875,
0.484375,
0.78759765625,
0.118408203125,
0.78759765625
],
[
0.51318359375,
0.77685546875,
0.87744140625,
0.77685546875,
0.87744140625,
0.78759765625,
0.51318359375,
0.78759765625
],
[
0.118408203125,
0.7919921875,
0.486083984375,
0.7919921875,
0.486083984375,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.51318359375,
0.7919921875,
0.880859375,
0.7919921875,
0.880859375,
0.8056640625,
0.51318359375,
0.8056640625
],
[
0.118408203125,
0.80859375,
0.486083984375,
0.80859375,
0.486083984375,
0.822265625,
0.118408203125,
0.822265625
],
[
0.51123046875,
0.80810546875,
0.8818359375,
0.80712890625,
0.8818359375,
0.82177734375,
0.51123046875,
0.822265625
],
[
0.118408203125,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.51123046875,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.51123046875,
0.83837890625
],
[
0.118408203125,
0.84130859375,
0.486083984375,
0.84130859375,
0.486083984375,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.51318359375,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.51318359375,
0.85498046875
],
[
0.118408203125,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51123046875,
0.85693359375,
0.87890625,
0.85693359375,
0.87890625,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.118408203125,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.51123046875,
0.87451171875,
0.87890625,
0.87451171875,
0.87890625,
0.88525390625,
0.51123046875,
0.88525390625
],
[
0.12005615234375,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.12005615234375,
0.90283203125
],
[
0.51318359375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.118408203125,
0.90625,
0.48779296875,
0.90625,
0.48779296875,
0.91943359375,
0.118408203125,
0.91943359375
],
[
0.509765625,
0.90380859375,
0.7109375,
0.90478515625,
0.7109375,
0.91943359375,
0.509765625,
0.91845703125
]
] |
[
"用ICLSI\"的输出作为分数（以下等式中的",
"",
"FFN中的两个线性层模拟为键值网络K和V",
"fels）来计算每个提及mEM的相关性得分：",
"",
"我们采用两个不同的线性层将信息r投影到",
"匹配层的向量空间：",
"Pm = fels (Genc([q; ml))",
"",
"(5)",
"Φk=Wh-r",
"(7)",
"我们以前2个相关性得分p作为目标提及。“提",
"Φ,=Wu.r",
"(8)",
"及相关性评分”的详情见附录A.5：实际上",
"除了这些常识或上下文知识之外，重要的是要",
"",
"其中Wk和W表示两个线性层的权重",
"注意PLM还存储其他类型的知识，包括技能",
"(XMATHX的中间是XMATHX_。两个矩阵Wk",
"知识和任务知识E(Wangetal.，2022）。因此，",
"和W，随机初始化，并将在训练期间更新。我",
"我们假设模型可以获取有关任务本身的潜在知",
"们通过将投影的知识连接到线性层的未尾来扩",
"识，因此我们进一步构建任务提示以鼓励模型",
"展FFN，并获得扩展的KE，VE。计算可以描",
"反映其参数中编码的技能知识",
"述如下：",
"总的来说，我们使用三种不同类型的任务引",
"导提示以及三个感兴趣的领域来探测PLM：",
"FFN(H)=f(H·KE)·VE",
"",
"(9)",
"= f(H . [Φ : K]) - [Φ, : V]",
"• Background Prompt : is designed to help",
"the model to think about the background, the",
"H表示自关注模块的输出隐藏状态。该模型将",
"prompt is As far as I know [MASK].",
"在反台知识的帮助下回答这个问题",
": Mention Prompt : is used to elicit memories",
"of mentions, the formation is About <Men-",
"4Experiments",
"",
"tion>, I know [MASK].",
"4.1Dataset",
": Task Prompt : is designed to help the model",
"我们针对常识推理基准的六个知识密集型任务",
"reminiscememories ofthetask.For exam-",
"评估了建议的方法：CommonsenseQA（THQA）",
"ple, for the sentiment analysis, the prompt is",
"(Talmor et al., 2019) , SociallQA (Sap et al., 2019)",
"About sentiment analysis, I know [MASK]",
"PhysicalQA（PIQA）。我们遵循先前作品（Ya",
"sunagaetal.,2022）使用的数据拆分。同时，为",
"我们在任务引导提示中放置了几个“[MASK]”，",
"了更好地理解任务提示在我们的方法中的作",
"并且“[MASK]”的长度是不同任务的超参数",
"用以及在预训练语言模型中学习的技能知识E",
"我们的方法可以应用于不同的PLM。对于",
"（Wangetal.,2022），我们考虑了GLUE基准E",
"像RoBERTa这样的编码器样式模型，我们利",
"(Wangetal.,2018）的任务，包括单句任务（SST，",
"用“[MASK]”的隐藏状态h[MASK】作为潜在",
"CoLA），推理任务（QNLI，RTE）以及相似性",
"的知识来进行反思。以下等式中的fmask表示",
"和释义任务（STS-B，MRPC)。我们在附录A.7",
"从模型的输出中获取h[MASK]。",
"中提供数据集详细信息",
"r = fmask(Ge ([q; Pal))",
"(6)",
"4.2Baselines",
"以下部分将说明如何利用这些知识。",
"我们选择RoBERTalarge（Liuetal.，2019）和",
"DeBERTa_largeE(Heetal.，2021）作为中型语",
"3.3Knowledge Consolidation with FFN",
"言模型的骨干模型，并将性能与以下强大的外",
"Neuron Augmentation",
"部知识增强方法进行比较：QA-GNN（Yasunaga",
"为了加强其理解，该模型应重新消化（注入）9",
"et al., 2021)，GreaseLM (Zhang et al.,2022)和",
"的知识，类似于动物如何再次咀嚼食物。然",
"DragonXMATHX_。有关基线模型的更多详细",
"而，在哪单注入PLMs仍然是一个具有挑战性",
"信息，请参见附录A.3.2。对于GLUE基准",
"的问题，表明潜在的工作来调查如何保持模型",
"测试中的任务，除了RoBERTa_large之外，我",
"的知识。先前的研究E(Daietal.2022；Wang",
"们还将我们的模型与提示学习方法LM-BFFE",
"etal.,2022）发现前馈网络（FFN）作为知识神",
"(Gaoetal.，2021）进行了比较。LM-BFF提出",
"经元或技能神经元起作用，说明FFN可以存",
"了一种基于提示的微调方法和一种改进的策",
"储事实信息并编码特定于任务的技能。受这些",
"略，用于动态和选择性地将演示合并到每个",
"发现的启发，我们将r合并到FFN中，就像以",
"上下文中。在我们的论文中，我们简单地使用",
"前的工作（Yaoetal.2022）一样。在这里，我",
"人类策划的提示（由LM-BFF提供），并利用",
"们选择Top-1层来重新测试（注人）知识。假设",
"RoBERTa_large作为骨干"
] |
{
"arxiv_id": "2305.08732",
"language": "zh",
"page_idx": 3
}
|
||
2305.08732-zh-4
|
2305.08732-zh-4.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th> </th>
<th>$CSQA$</th>
<th>SocialIQA</th>
<th>aNLI</th>
<th>$OBQA$</th>
<th>$PIQA$</th>
<th>HellaSwag</th>
</tr>
<tr>
<td rowspan="2">$\mathbf{Basic}$ :Model</td>
<td>$\mathrm{RoBERTa}^{*}$</td>
<td>68.7</td>
<td>75.9</td>
<td>82.7</td>
<td>64.9</td>
<td>79.4</td>
<td>82.3</td>
</tr>
<tr>
<td>DeBERTa</td>
<td>72.4</td>
<td>77.2</td>
<td>86.0</td>
<td>74.6</td>
<td>81.1</td>
<td>89.0</td>
</tr>
<tr>
<td> </td>
<td>$\mathrm{QAGNN}^{*}$</td>
<td>73.4</td>
<td>75.7</td>
<td>83.0</td>
<td>67.8</td>
<td>79.6</td>
<td>82.6</td>
</tr>
<tr>
<td rowspan="2">$\boldsymbol{L}\boldsymbol{\Lambda}\boldsymbol{l}.$ Knowledge</td>
<td>GreaseLM*</td>
<td>74.2</td>
<td>75.5</td>
<td>83.3</td>
<td>66.9</td>
<td>79.6</td>
<td>82.8</td>
</tr>
<tr>
<td>$\mathrm{Dragon}^{*}$</td>
<td>76.0</td>
<td>76.8</td>
<td>84.0</td>
<td>72.0</td>
<td>81.1</td>
<td>85.2</td>
</tr>
<tr>
<td>Rumination</td>
<td>RumiRoBERTa</td>
<td>70.3</td>
<td>77.7</td>
<td>86.1</td>
<td>70.0</td>
<td>80.8</td>
<td>85.8</td>
</tr>
<tr>
<td>$\mathbf{Model}$</td>
<td>RumiDeBERTa</td>
<td>74.3</td>
<td>78.4</td>
<td>86.7</td>
<td>76.0</td>
<td>81.9</td>
<td>89.5</td>
</tr>
</tbody>
</table>
Table 1: Accuracy on downstream commonsense reasoning tasks. Scores of the methods marked with * are taken from Yasunaga et al. (2022) . As the official tests for CSQA, PIOA and HellaSwag are hidden, here we report the in-house Dev (IHdev) and Test (IHtest) accuracy, following the data split in Yasunaga et al. (2022) .
<table>
<tbody>
<tr>
<th> </th>
<th> </th>
<th>$SST-2$</th>
<th>$SST-5$</th>
<th>CoLA</th>
<th>$RTE$</th>
<th>$MRPC$ $f1$</th>
<th>QNLI</th>
<th>$STS-B$ pear</th>
<th>$AVG$</th>
</tr>
<tr>
<td rowspan="2">Basic :Model</td>
<td>RoBERTa</td>
<td>95.00</td>
<td>58.70</td>
<td>62.60</td>
<td>80.90</td>
<td>91.40</td>
<td>93.30</td>
<td>91.90</td>
<td>81.97</td>
</tr>
<tr>
<td>LM- BFF</td>
<td>95.41</td>
<td>60.67</td>
<td>69.27</td>
<td>86.28</td>
<td>92.76</td>
<td>94.60</td>
<td>92.00</td>
<td>84.42</td>
</tr>
<tr>
<td>Rumination</td>
<td>RumiRoBERTa</td>
<td>95.75</td>
<td>60.85</td>
<td>68.57</td>
<td>85.92</td>
<td>93.92</td>
<td>94.84</td>
<td>92.23</td>
<td>84.58</td>
</tr>
<tr>
<td>Model</td>
<td>RumiLM- BFP</td>
<td>96.21</td>
<td>61.17</td>
<td>70.85</td>
<td>86.64</td>
<td>93.73</td>
<td>94.64</td>
<td>92.05</td>
<td>85.04</td>
</tr>
</tbody>
</table>
Table 2: Supervised GLUE benchmark results. Here, we report the results on the validation set following LM-
BFF (Gao et al., 2021). The prompt used here is the same as LM-BFF
## 4.3 Experiment Implementation
在带有任务引导提示的知识回顾阶段，模型的主干被冻结，我们仅更新前置的可训练连续令牌 (前缀提示)。在实现 DeBERTa 时，由于复杂的注意力机制，我们只是冻结整个模型，不添加前缀令牌。对于常识推理任务，我们将提到的提示和背景提示结合起来，对于 GLUE 基准测试，我们发现任务提示更有用。更多细节可以在附录 A.2 和 A.3 中找到
方面，RumiDeBERTa 在大多数任务上的表现都优于 Dragon。但是，使用外部知识的模型在CommonsenseQA 任务上表现最好。假设一些常识推理数据集是来自预先存在的知识库。例如，CommonsenseQA 派生自 ConceptNet。上述观察结果表明：1) 存储在参数中的知识是健壮的，并且在微调期间需要显式激活。2)从外部来源检索知识的模型的性能受到知识来源的质量和相关性的影响，而我们的反刍方法可以产生更相关的知识
如表$\mathbb{E}$2 所示，GLUE 基准测试的结果表明知识反刍方法优于基本微调模型 RoBERTa 和基于提示的方法 LM-BFF, LM-BFF 的% 平均提高了+1, RoBERTa 的% 平均提高了3。与微调和即时学习相比，这些性能提升凸显了知识反刍方法的有效性
## 4.4 Main Results
4.5 Out-of-Distribution (OOD) Performance
我们在表$\overline{\mathbb{E}}$1 中列出了常识推理任务的结果， 在表$\overline{\mathrm{E}}$2 中列出了 GLUE 基准测试的结果。所提出的技术，称为知识反刍，在大多数数据集上展示了卓越的性能。值得注意的是 ,它的性能优于朴素的基线，并获得了与包含外部知识的基线更好或相当的性能。如表所示 ,所提出的知识反刍方法 RumiRoBERTa 和 RumiDeBERTa 在六个常识推理任务上显示出更好的性能。结果表明，RumiRoBERTa 和 RumiDeBERTa 始终优于现有的语言模型(RoBERTa 和 DeBERTa),与RoBERTa 和 DeBERTa 相比，CSQA 的% 绝对准确性显着提高了2。这些结果表明，存储在 PLM 参数中的知识仍然可以进一步利用。此外，可以观察到 RumiRoBERTa 和结合外部知识的方法具有可比的结果。值得注意的是，在 SocialIQA, aNLI 和 HellaSwag 上，RumiRoBERTa 甚至优于预先训练的，知识增强的模型 Dragon。另一
为了更好地说明知识反刍方法的广泛适用性和泛化能力，我们扩展了我们的评估，以结合在分布外(OOD)测试集上的性能。表 3对应用于 RoBERTa 和 RumiRoBERTa 模型的微调技术的 OOD 性能进行了比较研究。总的来说，与传统的微调模型相比，RumiRoBERTa 在 OOD 测试中表现出卓越的性能。值得注意的是，当 RumiRoBERTa 接受 OBQA 培训并随后在 HellaSwag 和 PIQA 上进行测试时，它比RoBERTa 取得了 5% 的优势。这种性能的提高
|
[
[
0.5693359375,
0.08612060546875,
0.61865234375,
0.0885009765625,
0.6171875,
0.10394287109375,
0.568359375,
0.1015625
],
[
0.76416015625,
0.0872802734375,
0.85400390625,
0.09027099609375,
0.85205078125,
0.1051025390625,
0.76318359375,
0.1015625
],
[
0.62890625,
0.08905029296875,
0.68701171875,
0.08905029296875,
0.68701171875,
0.10394287109375,
0.62890625,
0.10394287109375
],
[
0.40478515625,
0.08905029296875,
0.459228515625,
0.08905029296875,
0.459228515625,
0.10394287109375,
0.40478515625,
0.10394287109375
],
[
0.474365234375,
0.09027099609375,
0.55517578125,
0.09027099609375,
0.55517578125,
0.1015625,
0.474365234375,
0.1015625
],
[
0.7021484375,
0.08905029296875,
0.75146484375,
0.08905029296875,
0.75146484375,
0.10394287109375,
0.7021484375,
0.10394287109375
],
[
0.4130859375,
0.10986328125,
0.450927734375,
0.10986328125,
0.450927734375,
0.1258544921875,
0.4130859375,
0.1258544921875
],
[
0.496337890625,
0.10986328125,
0.5341796875,
0.10986328125,
0.5341796875,
0.1258544921875,
0.496337890625,
0.1258544921875
],
[
0.57421875,
0.10986328125,
0.61279296875,
0.10986328125,
0.61279296875,
0.1258544921875,
0.57421875,
0.1258544921875
],
[
0.640625,
0.10986328125,
0.67822265625,
0.10986328125,
0.67822265625,
0.1258544921875,
0.640625,
0.1258544921875
],
[
0.7080078125,
0.10986328125,
0.74658203125,
0.10986328125,
0.74658203125,
0.1258544921875,
0.7080078125,
0.1258544921875
],
[
0.7900390625,
0.10986328125,
0.8271484375,
0.10986328125,
0.8271484375,
0.1258544921875,
0.7900390625,
0.1258544921875
],
[
0.1561279296875,
0.12054443359375,
0.2493896484375,
0.12054443359375,
0.2493896484375,
0.1318359375,
0.1561279296875,
0.1318359375
],
[
0.270263671875,
0.11224365234375,
0.353515625,
0.11224365234375,
0.353515625,
0.12353515625,
0.270263671875,
0.12353515625
],
[
0.270263671875,
0.1270751953125,
0.349365234375,
0.1270751953125,
0.349365234375,
0.141357421875,
0.270263671875,
0.141357421875
],
[
0.4130859375,
0.1270751953125,
0.45166015625,
0.1270751953125,
0.45166015625,
0.142578125,
0.4130859375,
0.142578125
],
[
0.496337890625,
0.1270751953125,
0.53564453125,
0.1270751953125,
0.53564453125,
0.142578125,
0.496337890625,
0.142578125
],
[
0.57421875,
0.1270751953125,
0.61376953125,
0.1270751953125,
0.61376953125,
0.142578125,
0.57421875,
0.142578125
],
[
0.6396484375,
0.1270751953125,
0.67919921875,
0.1270751953125,
0.67919921875,
0.142578125,
0.6396484375,
0.142578125
],
[
0.7060546875,
0.1270751953125,
0.74462890625,
0.1270751953125,
0.74462890625,
0.142578125,
0.7060546875,
0.142578125
],
[
0.78857421875,
0.12469482421875,
0.82861328125,
0.1270751953125,
0.8271484375,
0.1436767578125,
0.78662109375,
0.141357421875
],
[
0.270263671875,
0.149658203125,
0.349365234375,
0.149658203125,
0.349365234375,
0.1644287109375,
0.270263671875,
0.1644287109375
],
[
0.4130859375,
0.1478271484375,
0.453369140625,
0.15087890625,
0.450927734375,
0.1650390625,
0.411376953125,
0.163330078125
],
[
0.496337890625,
0.149658203125,
0.53564453125,
0.149658203125,
0.53564453125,
0.1650390625,
0.496337890625,
0.1650390625
],
[
0.57421875,
0.149658203125,
0.61376953125,
0.149658203125,
0.61376953125,
0.1650390625,
0.57421875,
0.1650390625
],
[
0.6396484375,
0.149658203125,
0.67822265625,
0.149658203125,
0.67822265625,
0.1650390625,
0.6396484375,
0.1650390625
],
[
0.70947265625,
0.149658203125,
0.74658203125,
0.149658203125,
0.74658203125,
0.1650390625,
0.70947265625,
0.1650390625
],
[
0.78857421875,
0.149658203125,
0.82861328125,
0.149658203125,
0.82861328125,
0.1650390625,
0.78857421875,
0.1650390625
],
[
0.151123046875,
0.15673828125,
0.187255859375,
0.15673828125,
0.187255859375,
0.1728515625,
0.151123046875,
0.1728515625
],
[
0.4130859375,
0.1650390625,
0.45166015625,
0.1650390625,
0.45166015625,
0.18115234375,
0.4130859375,
0.18115234375
],
[
0.496337890625,
0.1650390625,
0.5341796875,
0.1650390625,
0.5341796875,
0.18115234375,
0.496337890625,
0.18115234375
],
[
0.57421875,
0.1650390625,
0.61376953125,
0.1650390625,
0.61376953125,
0.18115234375,
0.57421875,
0.18115234375
],
[
0.640625,
0.1650390625,
0.67822265625,
0.1650390625,
0.67822265625,
0.18115234375,
0.640625,
0.18115234375
],
[
0.7080078125,
0.1650390625,
0.74658203125,
0.1650390625,
0.74658203125,
0.18115234375,
0.7080078125,
0.18115234375
],
[
0.78857421875,
0.1650390625,
0.82861328125,
0.1650390625,
0.82861328125,
0.18115234375,
0.78857421875,
0.18115234375
],
[
0.151123046875,
0.1728515625,
0.240966796875,
0.174560546875,
0.240966796875,
0.189453125,
0.151123046875,
0.18701171875
],
[
0.270263671875,
0.166259765625,
0.361083984375,
0.1644287109375,
0.361083984375,
0.179931640625,
0.270263671875,
0.1817626953125
],
[
0.4130859375,
0.18115234375,
0.45166015625,
0.18115234375,
0.45166015625,
0.196533203125,
0.4130859375,
0.196533203125
],
[
0.496337890625,
0.18115234375,
0.53564453125,
0.18115234375,
0.53564453125,
0.196533203125,
0.496337890625,
0.196533203125
],
[
0.57421875,
0.18115234375,
0.61376953125,
0.18115234375,
0.61376953125,
0.196533203125,
0.57421875,
0.196533203125
],
[
0.6396484375,
0.18115234375,
0.67919921875,
0.18115234375,
0.67919921875,
0.196533203125,
0.6396484375,
0.196533203125
],
[
0.7080078125,
0.18115234375,
0.74658203125,
0.18115234375,
0.74658203125,
0.196533203125,
0.7080078125,
0.196533203125
],
[
0.7900390625,
0.18115234375,
0.82861328125,
0.18115234375,
0.82861328125,
0.196533203125,
0.7900390625,
0.196533203125
],
[
0.4130859375,
0.2037353515625,
0.450927734375,
0.2037353515625,
0.450927734375,
0.2203369140625,
0.4130859375,
0.2203369140625
],
[
0.268798828125,
0.1817626953125,
0.337646484375,
0.1817626953125,
0.337646484375,
0.196533203125,
0.268798828125,
0.196533203125
],
[
0.496337890625,
0.2037353515625,
0.5341796875,
0.2037353515625,
0.5341796875,
0.2197265625,
0.496337890625,
0.2197265625
],
[
0.57421875,
0.2037353515625,
0.61279296875,
0.2037353515625,
0.61279296875,
0.2197265625,
0.57421875,
0.2197265625
],
[
0.6396484375,
0.2037353515625,
0.67919921875,
0.2037353515625,
0.67919921875,
0.2197265625,
0.6396484375,
0.2197265625
],
[
0.7060546875,
0.2037353515625,
0.74658203125,
0.2037353515625,
0.74658203125,
0.2197265625,
0.7060546875,
0.2197265625
],
[
0.78857421875,
0.2037353515625,
0.82861328125,
0.2037353515625,
0.82861328125,
0.2197265625,
0.78857421875,
0.2197265625
],
[
0.270263671875,
0.204833984375,
0.3896484375,
0.204833984375,
0.3896484375,
0.218505859375,
0.270263671875,
0.218505859375
],
[
0.15283203125,
0.2060546875,
0.2418212890625,
0.2060546875,
0.2418212890625,
0.21728515625,
0.15283203125,
0.21728515625
],
[
0.4130859375,
0.2197265625,
0.450927734375,
0.2197265625,
0.450927734375,
0.235107421875,
0.4130859375,
0.235107421875
],
[
0.496337890625,
0.2197265625,
0.53564453125,
0.2197265625,
0.53564453125,
0.235107421875,
0.496337890625,
0.235107421875
],
[
0.640625,
0.2197265625,
0.67919921875,
0.2197265625,
0.67919921875,
0.235107421875,
0.640625,
0.235107421875
],
[
0.78857421875,
0.2197265625,
0.82861328125,
0.2197265625,
0.82861328125,
0.235107421875,
0.78857421875,
0.235107421875
],
[
0.270263671875,
0.2203369140625,
0.3896484375,
0.2203369140625,
0.3896484375,
0.2340087890625,
0.270263671875,
0.2340087890625
],
[
0.151123046875,
0.2203369140625,
0.20654296875,
0.2203369140625,
0.20654296875,
0.235107421875,
0.151123046875,
0.235107421875
],
[
0.5751953125,
0.221435546875,
0.611328125,
0.221435546875,
0.611328125,
0.2340087890625,
0.5751953125,
0.2340087890625
],
[
0.7080078125,
0.221435546875,
0.74462890625,
0.221435546875,
0.74462890625,
0.2340087890625,
0.7080078125,
0.2340087890625
],
[
0.1158447265625,
0.2493896484375,
0.880859375,
0.25048828125,
0.880859375,
0.265380859375,
0.1158447265625,
0.26416015625
],
[
0.1175537109375,
0.26416015625,
0.880859375,
0.26416015625,
0.880859375,
0.27783203125,
0.1175537109375,
0.27783203125
],
[
0.1158447265625,
0.27978515625,
0.7841796875,
0.27978515625,
0.7841796875,
0.29345703125,
0.1158447265625,
0.29345703125
],
[
0.59033203125,
0.31005859375,
0.66015625,
0.3134765625,
0.6591796875,
0.327880859375,
0.58837890625,
0.32470703125
],
[
0.7314453125,
0.31005859375,
0.81201171875,
0.3134765625,
0.81005859375,
0.327880859375,
0.73046875,
0.32470703125
],
[
0.418212890625,
0.312255859375,
0.466796875,
0.312255859375,
0.466796875,
0.32666015625,
0.418212890625,
0.32666015625
],
[
0.359375,
0.312255859375,
0.408935546875,
0.312255859375,
0.408935546875,
0.32666015625,
0.359375,
0.32666015625
],
[
0.47607421875,
0.312255859375,
0.52490234375,
0.312255859375,
0.52490234375,
0.32666015625,
0.47607421875,
0.32666015625
],
[
0.5400390625,
0.3134765625,
0.57666015625,
0.3134765625,
0.57666015625,
0.325927734375,
0.5400390625,
0.325927734375
],
[
0.67529296875,
0.312255859375,
0.72119140625,
0.312255859375,
0.72119140625,
0.32666015625,
0.67529296875,
0.32666015625
],
[
0.82861328125,
0.3134765625,
0.8671875,
0.3134765625,
0.8671875,
0.32470703125,
0.82861328125,
0.32470703125
],
[
0.2392578125,
0.33203125,
0.310546875,
0.33203125,
0.310546875,
0.3466796875,
0.2392578125,
0.3466796875
],
[
0.361083984375,
0.33203125,
0.405517578125,
0.33203125,
0.405517578125,
0.3466796875,
0.361083984375,
0.3466796875
],
[
0.4189453125,
0.33203125,
0.46337890625,
0.33203125,
0.46337890625,
0.3466796875,
0.4189453125,
0.3466796875
],
[
0.478515625,
0.33203125,
0.52294921875,
0.33203125,
0.52294921875,
0.3466796875,
0.478515625,
0.3466796875
],
[
0.53564453125,
0.33203125,
0.580078125,
0.33203125,
0.580078125,
0.3466796875,
0.53564453125,
0.3466796875
],
[
0.603515625,
0.33203125,
0.64990234375,
0.33203125,
0.64990234375,
0.3466796875,
0.603515625,
0.3466796875
],
[
0.67529296875,
0.33203125,
0.7197265625,
0.33203125,
0.7197265625,
0.3466796875,
0.67529296875,
0.3466796875
],
[
0.75,
0.33203125,
0.79443359375,
0.33203125,
0.79443359375,
0.3466796875,
0.75,
0.3466796875
],
[
0.82373046875,
0.33203125,
0.86962890625,
0.33203125,
0.86962890625,
0.3466796875,
0.82373046875,
0.3466796875
],
[
0.1351318359375,
0.34033203125,
0.2216796875,
0.34033203125,
0.2216796875,
0.3515625,
0.1351318359375,
0.3515625
],
[
0.2392578125,
0.3466796875,
0.306396484375,
0.3466796875,
0.306396484375,
0.361083984375,
0.2392578125,
0.361083984375
],
[
0.361083984375,
0.344482421875,
0.40478515625,
0.3466796875,
0.403076171875,
0.3623046875,
0.359375,
0.35986328125
],
[
0.4189453125,
0.3466796875,
0.46337890625,
0.3466796875,
0.46337890625,
0.361083984375,
0.4189453125,
0.361083984375
],
[
0.478515625,
0.3466796875,
0.52294921875,
0.3466796875,
0.52294921875,
0.3623046875,
0.478515625,
0.3623046875
],
[
0.53564453125,
0.3466796875,
0.580078125,
0.3466796875,
0.580078125,
0.3623046875,
0.53564453125,
0.3623046875
],
[
0.603515625,
0.3466796875,
0.64990234375,
0.3466796875,
0.64990234375,
0.3623046875,
0.603515625,
0.3623046875
],
[
0.67529296875,
0.3466796875,
0.7197265625,
0.3466796875,
0.7197265625,
0.361083984375,
0.67529296875,
0.361083984375
],
[
0.75,
0.3466796875,
0.79443359375,
0.3466796875,
0.79443359375,
0.3623046875,
0.75,
0.3623046875
],
[
0.82373046875,
0.3466796875,
0.869140625,
0.3466796875,
0.869140625,
0.3623046875,
0.82373046875,
0.3623046875
],
[
0.4189453125,
0.364013671875,
0.46337890625,
0.366455078125,
0.462646484375,
0.3818359375,
0.418212890625,
0.380126953125
],
[
0.478515625,
0.366455078125,
0.52294921875,
0.366455078125,
0.52294921875,
0.3818359375,
0.478515625,
0.3818359375
],
[
0.53564453125,
0.366455078125,
0.580078125,
0.366455078125,
0.580078125,
0.3818359375,
0.53564453125,
0.3818359375
],
[
0.603515625,
0.366455078125,
0.64990234375,
0.366455078125,
0.64990234375,
0.3818359375,
0.603515625,
0.3818359375
],
[
0.67529296875,
0.364013671875,
0.7197265625,
0.366455078125,
0.71875,
0.3818359375,
0.67333984375,
0.380126953125
],
[
0.75,
0.366455078125,
0.79443359375,
0.366455078125,
0.79443359375,
0.3818359375,
0.75,
0.3818359375
],
[
0.82373046875,
0.366455078125,
0.86962890625,
0.366455078125,
0.86962890625,
0.3818359375,
0.82373046875,
0.3818359375
],
[
0.359375,
0.36767578125,
0.40478515625,
0.36767578125,
0.40478515625,
0.3818359375,
0.359375,
0.3818359375
],
[
0.1318359375,
0.36767578125,
0.214111328125,
0.36767578125,
0.214111328125,
0.37890625,
0.1318359375,
0.37890625
],
[
0.240966796875,
0.3681640625,
0.349365234375,
0.3681640625,
0.349365234375,
0.380126953125,
0.240966796875,
0.380126953125
],
[
0.359375,
0.3818359375,
0.40478515625,
0.3818359375,
0.40478515625,
0.396728515625,
0.359375,
0.396728515625
],
[
0.4189453125,
0.3818359375,
0.46337890625,
0.3818359375,
0.46337890625,
0.396728515625,
0.4189453125,
0.396728515625
],
[
0.478515625,
0.3818359375,
0.52294921875,
0.3818359375,
0.52294921875,
0.396728515625,
0.478515625,
0.396728515625
],
[
0.53564453125,
0.3818359375,
0.580078125,
0.3818359375,
0.580078125,
0.396728515625,
0.53564453125,
0.396728515625
],
[
0.603515625,
0.3818359375,
0.64990234375,
0.3818359375,
0.64990234375,
0.396728515625,
0.603515625,
0.396728515625
],
[
0.67529296875,
0.3818359375,
0.7197265625,
0.3818359375,
0.7197265625,
0.396728515625,
0.67529296875,
0.396728515625
],
[
0.75,
0.380126953125,
0.79443359375,
0.3818359375,
0.79248046875,
0.39794921875,
0.748046875,
0.3955078125
],
[
0.82373046875,
0.3818359375,
0.869140625,
0.3818359375,
0.869140625,
0.396728515625,
0.82373046875,
0.396728515625
],
[
0.240966796875,
0.383056640625,
0.340087890625,
0.383056640625,
0.340087890625,
0.394287109375,
0.240966796875,
0.394287109375
],
[
0.133544921875,
0.383056640625,
0.181396484375,
0.383056640625,
0.181396484375,
0.394287109375,
0.133544921875,
0.394287109375
],
[
0.1175537109375,
0.410888671875,
0.880859375,
0.410888671875,
0.880859375,
0.424560546875,
0.1175537109375,
0.424560546875
],
[
0.1175537109375,
0.42578125,
0.599609375,
0.42578125,
0.599609375,
0.439453125,
0.1175537109375,
0.439453125
],
[
0.1175537109375,
0.465576171875,
0.38623046875,
0.466064453125,
0.38623046875,
0.479736328125,
0.1175537109375,
0.478515625
],
[
0.5146484375,
0.466064453125,
0.880859375,
0.466064453125,
0.880859375,
0.479736328125,
0.5146484375,
0.479736328125
],
[
0.118408203125,
0.48681640625,
0.486083984375,
0.48681640625,
0.486083984375,
0.50048828125,
0.118408203125,
0.50048828125
],
[
0.51123046875,
0.482177734375,
0.880859375,
0.48095703125,
0.880859375,
0.495361328125,
0.51123046875,
0.496337890625
],
[
0.12005615234375,
0.5029296875,
0.483642578125,
0.5029296875,
0.483642578125,
0.5166015625,
0.12005615234375,
0.5166015625
],
[
0.509765625,
0.49755859375,
0.8818359375,
0.496337890625,
0.8818359375,
0.51123046875,
0.509765625,
0.5126953125
],
[
0.118408203125,
0.5185546875,
0.483642578125,
0.5185546875,
0.483642578125,
0.533203125,
0.118408203125,
0.533203125
],
[
0.51318359375,
0.51416015625,
0.880859375,
0.51416015625,
0.880859375,
0.52783203125,
0.51318359375,
0.52783203125
],
[
0.118408203125,
0.53369140625,
0.484375,
0.53369140625,
0.484375,
0.54736328125,
0.118408203125,
0.54736328125
],
[
0.51123046875,
0.529296875,
0.8818359375,
0.52978515625,
0.8818359375,
0.54443359375,
0.51123046875,
0.54345703125
],
[
0.1175537109375,
0.55029296875,
0.486083984375,
0.5498046875,
0.486083984375,
0.56396484375,
0.1175537109375,
0.5654296875
],
[
0.51318359375,
0.54638671875,
0.880859375,
0.54638671875,
0.880859375,
0.56005859375,
0.51318359375,
0.56005859375
],
[
0.1158447265625,
0.56640625,
0.484375,
0.56640625,
0.484375,
0.580078125,
0.1158447265625,
0.580078125
],
[
0.51318359375,
0.5625,
0.880859375,
0.5625,
0.880859375,
0.576171875,
0.51318359375,
0.576171875
],
[
0.51123046875,
0.57763671875,
0.880859375,
0.57666015625,
0.880859375,
0.59130859375,
0.51123046875,
0.5927734375
],
[
0.1175537109375,
0.58447265625,
0.486083984375,
0.58447265625,
0.486083984375,
0.59716796875,
0.1175537109375,
0.59716796875
],
[
0.1175537109375,
0.59716796875,
0.353515625,
0.59716796875,
0.353515625,
0.6123046875,
0.1175537109375,
0.6123046875
],
[
0.5146484375,
0.59423828125,
0.880859375,
0.59423828125,
0.880859375,
0.60791015625,
0.5146484375,
0.60791015625
],
[
0.51318359375,
0.60986328125,
0.66162109375,
0.60986328125,
0.66162109375,
0.62451171875,
0.51318359375,
0.62451171875
],
[
0.1175537109375,
0.6259765625,
0.268798828125,
0.6259765625,
0.268798828125,
0.6396484375,
0.1175537109375,
0.6396484375
],
[
0.5322265625,
0.62744140625,
0.880859375,
0.62744140625,
0.880859375,
0.64111328125,
0.5322265625,
0.64111328125
],
[
0.118408203125,
0.646484375,
0.484375,
0.646484375,
0.484375,
0.66015625,
0.118408203125,
0.66015625
],
[
0.51123046875,
0.642578125,
0.87890625,
0.642578125,
0.87890625,
0.65625,
0.51123046875,
0.65625
],
[
0.1175537109375,
0.662109375,
0.484375,
0.662109375,
0.484375,
0.67578125,
0.1175537109375,
0.67578125
],
[
0.51123046875,
0.66015625,
0.880859375,
0.65771484375,
0.880859375,
0.67138671875,
0.51123046875,
0.67333984375
],
[
0.118408203125,
0.6796875,
0.486083984375,
0.6796875,
0.486083984375,
0.693359375,
0.118408203125,
0.693359375
],
[
0.51318359375,
0.67578125,
0.880859375,
0.67578125,
0.880859375,
0.6904296875,
0.51318359375,
0.6904296875
],
[
0.118408203125,
0.6953125,
0.486083984375,
0.6953125,
0.486083984375,
0.708984375,
0.118408203125,
0.708984375
],
[
0.51318359375,
0.6923828125,
0.8818359375,
0.6923828125,
0.8818359375,
0.7060546875,
0.51318359375,
0.7060546875
],
[
0.12005615234375,
0.71142578125,
0.486083984375,
0.71142578125,
0.486083984375,
0.724609375,
0.12005615234375,
0.724609375
],
[
0.51318359375,
0.7080078125,
0.6630859375,
0.7080078125,
0.6630859375,
0.7216796875,
0.51318359375,
0.7216796875
],
[
0.1175537109375,
0.7255859375,
0.484375,
0.72705078125,
0.484375,
0.7412109375,
0.1175537109375,
0.74072265625
],
[
0.51318359375,
0.73828125,
0.5556640625,
0.73828125,
0.5556640625,
0.75,
0.51318359375,
0.75
],
[
0.55517578125,
0.736328125,
0.87890625,
0.73681640625,
0.87890625,
0.751953125,
0.55517578125,
0.75048828125
],
[
0.118408203125,
0.7421875,
0.486083984375,
0.7421875,
0.486083984375,
0.755859375,
0.118408203125,
0.755859375
],
[
0.118408203125,
0.76025390625,
0.486083984375,
0.76025390625,
0.486083984375,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.51318359375,
0.7587890625,
0.880859375,
0.7587890625,
0.880859375,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.118408203125,
0.775390625,
0.48779296875,
0.775390625,
0.48779296875,
0.7890625,
0.118408203125,
0.7890625
],
[
0.51318359375,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.7890625,
0.51318359375,
0.7890625
],
[
0.118408203125,
0.79150390625,
0.486083984375,
0.79052734375,
0.486083984375,
0.8046875,
0.118408203125,
0.8056640625
],
[
0.51318359375,
0.79052734375,
0.8818359375,
0.79150390625,
0.8818359375,
0.8056640625,
0.51318359375,
0.8046875
],
[
0.1175537109375,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82177734375,
0.1175537109375,
0.82177734375
],
[
0.51318359375,
0.80810546875,
0.880859375,
0.80810546875,
0.880859375,
0.82177734375,
0.51318359375,
0.82177734375
],
[
0.1175537109375,
0.82373046875,
0.48779296875,
0.822265625,
0.48779296875,
0.83740234375,
0.1175537109375,
0.83837890625
],
[
0.51318359375,
0.82373046875,
0.880859375,
0.82373046875,
0.880859375,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.118408203125,
0.84033203125,
0.48779296875,
0.84033203125,
0.48779296875,
0.85400390625,
0.118408203125,
0.85400390625
],
[
0.51123046875,
0.84033203125,
0.87890625,
0.84033203125,
0.87890625,
0.85400390625,
0.51123046875,
0.85400390625
],
[
0.118408203125,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51318359375,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.118408203125,
0.87255859375,
0.489501953125,
0.87255859375,
0.489501953125,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.51123046875,
0.87255859375,
0.87890625,
0.87255859375,
0.87890625,
0.8857421875,
0.51123046875,
0.8857421875
],
[
0.118408203125,
0.88818359375,
0.48779296875,
0.88818359375,
0.48779296875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51318359375,
0.88818359375,
0.880859375,
0.88818359375,
0.880859375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.12005615234375,
0.90478515625,
0.470947265625,
0.90478515625,
0.470947265625,
0.91845703125,
0.12005615234375,
0.91845703125
],
[
0.468505859375,
0.9091796875,
0.48193359375,
0.9091796875,
0.48193359375,
0.9130859375,
0.468505859375,
0.9130859375
],
[
0.51318359375,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"aNLI",
"HellaSwag",
"OBQA",
"CSQA",
"SociallQA",
"PIQA",
"68.7",
"75.9",
"82.7",
"64.9",
"79.4",
"82.3",
"BasicModel",
"RoBERTa*",
"DeBERTa",
"72.4",
"77.2",
"86.0",
"74.6",
"81.1",
"89.0",
"QAGNN*",
"73.4",
"75.7",
"83.0",
"67.8",
"79.6",
"82.6",
"Ext.",
"74.2",
"75.5",
"83.3",
"66.9",
"79.6",
"82.8",
"Knowledge",
"GreaseLM*",
"76.0",
"76.8",
"84.0",
"72.0",
"81.1",
"85.2",
"70.3",
"Dragon*",
"77.7",
"86.1",
"70.0",
"80.8",
"85.8",
"RumiRoBERTa",
"Rumination",
"74.3",
"78.4",
"76.0",
"89.5",
"RumiDeBERTa",
"Model",
"86.7",
"81.9",
"Table l: Accuracy on downstream commonsense reasoning tasks. Scores of the methods marked with * are taken",
"from Yasunaga et al. (2022) . As the official tests for CSQA, PIOA and HellaSwag are hidden, here we report the",
"in-house Dev (IHdev) and Test (IHtest) accuracy, following the data split in Yasunaga et al. (2022) ",
"MRPC f1",
"STS-B pear",
"SST-5",
"SST-2",
"CoLA",
"RTE",
"QNLI",
"AVG",
"RoBERTa",
"95.00",
"58.70",
"62.60",
"80.90",
"91.40",
"93.30",
"91.90",
"81.97",
"Basic Model",
"LM-BFF",
"95.41",
"60.67",
"69.27",
"86.28",
"92.76",
"94.60",
"92.00",
"84.42",
"60.85",
"68.57",
"85.92",
"93.92",
"94.84",
"92.23",
"84.58",
"95.75",
"Rumination",
"RumiRoBERTa",
"96.21",
"61.17",
"70.85",
"86.64",
"93.73",
"94.64",
"92.05",
"85.04",
"RumiLM-BFF",
"Model",
"Table 2: Supervised GLUE benchmark results. Here, we report the results on the validation set following LM",
"BFF (Gao et al., 2021) . The prompt used here is the same as LM-BFF.",
"4.3ExperimentImplementation",
"方面，RumiDeBERTa在大多数任务上的表现",
"在带有任务引导提示的知识回顾阶段，模型的",
"都优于Dragon。但是，使用外部知识的模型在",
"主干被冻结，我们仅更新前置的可训练连续令",
"CommonsenseQA任务上表现最好。假设一些",
"牌（前缀提示）。在实现DeBERTa时，由于复",
"常识推理数据集是来自预先存在的知识库。例",
"杂的注意力机制，我们只是冻结整个模型，不",
"如，CommonsenseQA派生自ConceptNet。上述",
"添加前缀令牌。对于常识推理任务，我们将提",
"观察结果表明：1）存储在参数中的知识是健",
"到的提示和背景提示结合起来，对于GLUE基",
"壮的，并且在微调期间需要显式激活。2）从外",
"部来源检索知识的模型的性能受到知识来源的",
"准测试，我们发现任务提示更有用。更多细节",
"可以在附录A.2和A.3中找到",
"质量和相关性的影响，而我们的反方法可以",
"产生更相关的知识",
"4.4Main Results",
"如表E2所示，GLUE基准测试的结果表明，",
"我们在表1中列出了常识推理任务的结果，",
"知识反各方法优于基本微调模型RoBERTa和",
"在表E2 中列出了GLUE基准测试的结果。所",
"基于提示的方法LM-BFF，LM-BFF的%平均",
"提出的技术，称为知识反，在大多数数据集上",
"提高了+1，RoBERTa的%平均提高了3。与微",
"展示了卓越的性能。值得注意的是，它的性能优",
"调和即时学习相比，这些性能提升凸显了知识",
"于朴素的基线，并获得了与包含外部知识的基",
"反方法的有效性",
"线更好或相当的性能。如表所示，所提出的知识",
"4.5",
"Out-of-Distribution (OOD)Performance",
"反含方法RumiRoBERTa和RumiDeBERTa在",
"六个常识推理任务上显示出更好的性能。结果",
"为了更好地说明知识反方法的广泛适用性",
"表明，RumiRoBERTa和RumiDeBERTa始终优",
"和泛化能力，我们扩展了我们的评估，以结",
"于现有的语言模型（RoBERTa和DeBERTa），与",
"合在分布外（OOD）测试集上的性能。表3",
"RoBERTa和DeBERTa相比，CSQA的%绝对准",
"对应用于RoBERTa和RumiRoBERTa模型的",
"确性显着提高了2。这些结果表明，存储在PLM",
"微调技术的OOD性能进行了比较研究。总的",
"参数中的知识仍然可以进一步利用。此外，可以",
"来说，与传统的微调模型相比，RumiRoBERTa",
"观察到RumiRoBERTa和结合外部知识的方法",
"在OOD测试中表现出卓越的性能。值得注意",
"具有可比的结果。值得注意的是，在SociallQA，",
"的是，当RumiRoBERTa接受OBQA培训并随",
"aNLI和HellaSwag上，RumiRoBERTa甚至优",
"后在HellaSwag和PIQA上进行测试时，它比",
"于预先训练的，知识增强的模型Dragon。另-",
"",
"RoBERTa取得了5%的优势。这种性能的提高"
] |
{
"arxiv_id": "2305.08732",
"language": "zh",
"page_idx": 4
}
|
||
2109.11295-zh-1
|
2109.11295-zh-1.jpg
|
预训练语言模型的动态知识提炼
$\textbf{Lei Li}^{\dagger }, \textbf{Yankai Lin}^{§ }, \textbf{Shuhuai Ren}^{\dagger }, \textbf{Peng Li}^{8}, \textbf{Jie Zhou}^{\S }, \textbf{Xu Sun}^{\text{Lei Li}\dagger }, \textbf{Yankai Lin}^{§ }, \textbf{Shuhuai Ren}^{\dagger }, \textbf{Peng Li}^{§ }, \textbf{Jie Zhou}^{§ }, \textbf{Xu Sun}^{\text{Lei Li}\dagger }, \textbf{Yankai Lin}^{§ }, \textbf{Shuhuai Ren}^{\dagger }, \textbf{Peng Li}^{§ }, \textbf{Jie Zhou}^{§ }, \textbf{Xu Su}$
$^{\dagger}$MOE Key Laboratory of Computational Linguistics, School of EECS, Peking University
$^{6}$Pattern Recognition Center, WeChat AI, Tencent Inc., China
{ lilei, shuhuai\_ren } @stu.pku.edu.cn xusun@pku.edu.cn
$\{ yankailin, patrickpli, withtomzhou $ Qtecent.com$
## $\mathbf{Abstract}$
Knowledge distillation (KD) has been proved effective for compressing large-scale pretrained language models. However, existing methods conduct KD statically, e.g., the student model aligns its output distribution to that of a selected teacher model on the pre-defined training dataset. In this paper, we explore whether a dynamic knowledge distillation that empowers the student to adjust the learning procedure according to its competency, regarding the student performance and learning efficiency. We explore the dynamical adjustments on three aspects: teacher model adoption, data selection, and KD objective adaptation. Experimental results show that (1) proper selection of teacher model can boost the performance of student model; (2) conducting KD with 10 % informative instances achieves comparable performance while greatly accelerates the training; (3) the student performance can be boosted by adjusting the supervision contribution of different alignment objective. We find dynamic knowledge distillation is promising and provide discussions on potential future directions towards more efficient KD methods.
型的输出概率分布对齐。也就是说，KD 的以下三个方面是预先指定的，在学习过程中保持不变：(1) 要学习的教师模型 (学习目标);(2) 用于查询教师的数据 (学习材料);(3) 目标函数和相应的权重 (学习方法)。但是，随着学生能力在培训阶段的发展，预先定义这些学习设置并使其保持不变尤其不合理。静态地进行 KD 可能会导致(1)从太大的老师那里进行不合格的学习，(2)在学生已经掌握的实例上进行重复学习，以及(3)在不必要的排列上进行次优学习。这促使我们去探索一个有趣的问题：一个动态的 KD 框架 , 考虑到学生在培训过程中的能力演变，是否能带来好处，关于学生的表现和学习效率？
在本文中，我们提出了一个动态知识蒸馏$\boxed{\mathrm{E}}$ (DynamicKD) 框架，该框架试图使学生能够根据其能力调整学习过程。具体来说， 受到主动学习 (Settles,2009)的成功启发，我们将预测不确定性。g。,预测分类概率分布的熵，作为学生能力的代理。我们努力回答以下研究问题：(RQ1)随着学生的发展，哪个老师适合学习？(RQ2)在整个KD阶段，哪些数据实际上对学生模型有用？(RQ3)在KD 过程中，最佳学习目标是否会改变？特别是， 我们首先探讨了教师规模对动态知识提炼的影响。其次，我们探讨动态选择学生不确定 KD 的实例是否可以带来更好的性能和培训效率的权衡。第三，我们探讨了从预测概率分布的对齐和整个 KD 阶段的中间表示的对齐来动态调整监督是否可以提高性能
我们的实验结果表明：(1)具有更多层的较大教师模型可能会培养出更差的学生。我们表明，根据学生的能力选择合适的教师模型可以提高成绩。(2)我们可以只使用根据学生预测不确定性选择的 10 个% 信息实例来实现可比的性能。随着学生变得更强壮，这些情况也会在培训过程中发展。(3)我们可以通过动态调整教师模型的不同对齐目标的监督来提高学生的成绩
我们的观察证明了当前静态 KD 框架的局限性。提出的基于不确定性的动态 KD 框架只是
## 1 $\textbf{Introduction}$
知识提炼(KD) (Hinton et al., 2015) 旨在将知识从大型教师模型转移到小型学生模型。近年来， 它已被广泛使用$\overline{\mathrm{E}}$(Sanh et al., 2019; Jiao et al., 2020; Sun et al., 2019) 来压缩像 BERT$\boxed{\mathrm{F}}($Devlin et al., 2019) 和 RoBERTa$\underline{\mathrm{E}}($Liu et al., 2019) 这样的大规模预训练语言模型$\overline{\mathrm{E}}($PLM)。通过知识蒸馏，我们可以获得一个小得多的模型，具有可比的性能，同时大大降低了内存使用量并加速了模型推理
虽然简单有效，但现有方法通常静态地执行KD 学习过程，例如。g。, 学生模型将其输出概率分布与整个预定义语料库上选定教师模
$^{1}$Our code is available at https://github.com/
lancopku/DynamicKD
|
[
[
0.333251953125,
0.0955810546875,
0.66650390625,
0.0955810546875,
0.66650390625,
0.10986328125,
0.333251953125,
0.10986328125
],
[
0.2216796875,
0.1341552734375,
0.779296875,
0.1341552734375,
0.779296875,
0.1484375,
0.2216796875,
0.1484375
],
[
0.14111328125,
0.149658203125,
0.85986328125,
0.15087890625,
0.85986328125,
0.1650390625,
0.14111328125,
0.1644287109375
],
[
0.25439453125,
0.166259765625,
0.75,
0.16748046875,
0.75,
0.1817626953125,
0.25439453125,
0.18115234375
],
[
0.173828125,
0.18408203125,
0.8427734375,
0.18408203125,
0.8427734375,
0.19775390625,
0.173828125,
0.19775390625
],
[
0.20654296875,
0.203125,
0.8095703125,
0.203125,
0.8095703125,
0.2161865234375,
0.20654296875,
0.2161865234375
],
[
0.2626953125,
0.254150390625,
0.340087890625,
0.254150390625,
0.340087890625,
0.268310546875,
0.2626953125,
0.268310546875
],
[
0.51318359375,
0.254638671875,
0.880859375,
0.254638671875,
0.880859375,
0.268310546875,
0.51318359375,
0.268310546875
],
[
0.51318359375,
0.270751953125,
0.880859375,
0.270751953125,
0.880859375,
0.283935546875,
0.51318359375,
0.283935546875
],
[
0.145263671875,
0.281982421875,
0.45751953125,
0.281982421875,
0.45751953125,
0.295654296875,
0.145263671875,
0.295654296875
],
[
0.5146484375,
0.287353515625,
0.880859375,
0.287353515625,
0.880859375,
0.300537109375,
0.5146484375,
0.300537109375
],
[
0.145263671875,
0.29638671875,
0.45751953125,
0.29638671875,
0.45751953125,
0.311279296875,
0.145263671875,
0.311279296875
],
[
0.51318359375,
0.302734375,
0.880859375,
0.302734375,
0.880859375,
0.31640625,
0.51318359375,
0.31640625
],
[
0.1435546875,
0.311279296875,
0.45751953125,
0.31005859375,
0.45751953125,
0.32470703125,
0.1435546875,
0.325927734375
],
[
0.51318359375,
0.319580078125,
0.880859375,
0.319580078125,
0.880859375,
0.333251953125,
0.51318359375,
0.333251953125
],
[
0.14697265625,
0.325927734375,
0.45751953125,
0.325927734375,
0.45751953125,
0.339111328125,
0.14697265625,
0.339111328125
],
[
0.145263671875,
0.339111328125,
0.45751953125,
0.339111328125,
0.45751953125,
0.352783203125,
0.145263671875,
0.352783203125
],
[
0.51318359375,
0.3349609375,
0.880859375,
0.3349609375,
0.880859375,
0.3486328125,
0.51318359375,
0.3486328125
],
[
0.14697265625,
0.352783203125,
0.456787109375,
0.352783203125,
0.456787109375,
0.366455078125,
0.14697265625,
0.366455078125
],
[
0.51123046875,
0.350830078125,
0.87890625,
0.350830078125,
0.87890625,
0.364013671875,
0.51123046875,
0.364013671875
],
[
0.145263671875,
0.3681640625,
0.456787109375,
0.3681640625,
0.456787109375,
0.3818359375,
0.145263671875,
0.3818359375
],
[
0.5146484375,
0.3681640625,
0.880859375,
0.3681640625,
0.880859375,
0.3818359375,
0.5146484375,
0.3818359375
],
[
0.1435546875,
0.38134765625,
0.45751953125,
0.380126953125,
0.45751953125,
0.394287109375,
0.1435546875,
0.3955078125
],
[
0.51123046875,
0.383056640625,
0.87744140625,
0.383056640625,
0.87744140625,
0.396728515625,
0.51123046875,
0.396728515625
],
[
0.14697265625,
0.396728515625,
0.45751953125,
0.396728515625,
0.45751953125,
0.410400390625,
0.14697265625,
0.410400390625
],
[
0.51318359375,
0.399658203125,
0.87890625,
0.399658203125,
0.87890625,
0.413330078125,
0.51318359375,
0.413330078125
],
[
0.14697265625,
0.410888671875,
0.45751953125,
0.410888671875,
0.45751953125,
0.424560546875,
0.14697265625,
0.424560546875
],
[
0.5146484375,
0.4150390625,
0.880859375,
0.4150390625,
0.880859375,
0.4287109375,
0.5146484375,
0.4287109375
],
[
0.14697265625,
0.424560546875,
0.459228515625,
0.424560546875,
0.459228515625,
0.438232421875,
0.14697265625,
0.438232421875
],
[
0.51318359375,
0.431640625,
0.87890625,
0.431640625,
0.87890625,
0.4453125,
0.51318359375,
0.4453125
],
[
0.145263671875,
0.439453125,
0.456787109375,
0.439453125,
0.456787109375,
0.453125,
0.145263671875,
0.453125
],
[
0.51123046875,
0.44775390625,
0.87890625,
0.44775390625,
0.87890625,
0.46142578125,
0.51123046875,
0.46142578125
],
[
0.14697265625,
0.454833984375,
0.455078125,
0.454833984375,
0.455078125,
0.465576171875,
0.14697265625,
0.465576171875
],
[
0.1485595703125,
0.4697265625,
0.456787109375,
0.4697265625,
0.456787109375,
0.479736328125,
0.1485595703125,
0.479736328125
],
[
0.5146484375,
0.465576171875,
0.615234375,
0.465576171875,
0.615234375,
0.476806640625,
0.5146484375,
0.476806640625
],
[
0.14697265625,
0.482177734375,
0.45751953125,
0.482177734375,
0.45751953125,
0.495361328125,
0.14697265625,
0.495361328125
],
[
0.52880859375,
0.48095703125,
0.880859375,
0.482177734375,
0.880859375,
0.496337890625,
0.52880859375,
0.495361328125
],
[
0.1485595703125,
0.49755859375,
0.455078125,
0.49755859375,
0.455078125,
0.50830078125,
0.1485595703125,
0.50830078125
],
[
0.51318359375,
0.496337890625,
0.880859375,
0.49755859375,
0.880859375,
0.5126953125,
0.51318359375,
0.51123046875
],
[
0.14697265625,
0.51025390625,
0.45751953125,
0.51025390625,
0.45751953125,
0.52392578125,
0.14697265625,
0.52392578125
],
[
0.51123046875,
0.51416015625,
0.87890625,
0.51416015625,
0.87890625,
0.52783203125,
0.51123046875,
0.52783203125
],
[
0.1485595703125,
0.525390625,
0.453369140625,
0.525390625,
0.453369140625,
0.5361328125,
0.1485595703125,
0.5361328125
],
[
0.5146484375,
0.52978515625,
0.8818359375,
0.52978515625,
0.8818359375,
0.54443359375,
0.5146484375,
0.54443359375
],
[
0.145263671875,
0.5380859375,
0.455078125,
0.53759765625,
0.455078125,
0.55029296875,
0.145263671875,
0.5517578125
],
[
0.14697265625,
0.552734375,
0.456787109375,
0.552734375,
0.456787109375,
0.56640625,
0.14697265625,
0.56640625
],
[
0.5146484375,
0.54736328125,
0.880859375,
0.54736328125,
0.880859375,
0.5625,
0.5146484375,
0.5625
],
[
0.14697265625,
0.568359375,
0.456787109375,
0.568359375,
0.456787109375,
0.5791015625,
0.14697265625,
0.5791015625
],
[
0.5146484375,
0.5634765625,
0.880859375,
0.5634765625,
0.880859375,
0.57666015625,
0.5146484375,
0.57666015625
],
[
0.145263671875,
0.5791015625,
0.45751953125,
0.580078125,
0.45751953125,
0.59423828125,
0.145263671875,
0.59326171875
],
[
0.515625,
0.5791015625,
0.880859375,
0.5791015625,
0.880859375,
0.5927734375,
0.515625,
0.5927734375
],
[
0.14697265625,
0.595703125,
0.45751953125,
0.595703125,
0.45751953125,
0.609375,
0.14697265625,
0.609375
],
[
0.51123046875,
0.59326171875,
0.87890625,
0.59326171875,
0.87890625,
0.60693359375,
0.51123046875,
0.60693359375
],
[
0.14697265625,
0.609375,
0.456787109375,
0.609375,
0.456787109375,
0.62255859375,
0.14697265625,
0.62255859375
],
[
0.51123046875,
0.609375,
0.880859375,
0.60986328125,
0.880859375,
0.62451171875,
0.51123046875,
0.62353515625
],
[
0.145263671875,
0.62353515625,
0.456787109375,
0.62353515625,
0.456787109375,
0.63720703125,
0.145263671875,
0.63720703125
],
[
0.5146484375,
0.626953125,
0.87744140625,
0.626953125,
0.87744140625,
0.64013671875,
0.5146484375,
0.64013671875
],
[
0.51123046875,
0.642578125,
0.87744140625,
0.642578125,
0.87744140625,
0.65625,
0.51123046875,
0.65625
],
[
0.5146484375,
0.66015625,
0.880859375,
0.66015625,
0.880859375,
0.67333984375,
0.5146484375,
0.67333984375
],
[
0.1158447265625,
0.66748046875,
0.151123046875,
0.66748046875,
0.151123046875,
0.6787109375,
0.1158447265625,
0.6787109375
],
[
0.1485595703125,
0.66748046875,
0.2626953125,
0.66748046875,
0.2626953125,
0.6796875,
0.1485595703125,
0.6796875
],
[
0.5146484375,
0.67578125,
0.880859375,
0.67578125,
0.880859375,
0.689453125,
0.5146484375,
0.689453125
],
[
0.1175537109375,
0.69140625,
0.486083984375,
0.69140625,
0.486083984375,
0.705078125,
0.1175537109375,
0.705078125
],
[
0.51123046875,
0.6904296875,
0.87890625,
0.6904296875,
0.87890625,
0.70361328125,
0.51123046875,
0.70361328125
],
[
0.12005615234375,
0.71044921875,
0.489501953125,
0.71044921875,
0.489501953125,
0.72021484375,
0.12005615234375,
0.72021484375
],
[
0.51318359375,
0.7080078125,
0.880859375,
0.7080078125,
0.880859375,
0.7216796875,
0.51318359375,
0.7216796875
],
[
0.118408203125,
0.724609375,
0.48779296875,
0.724609375,
0.48779296875,
0.73828125,
0.118408203125,
0.73828125
],
[
0.51318359375,
0.724609375,
0.71533203125,
0.724609375,
0.71533203125,
0.73828125,
0.51318359375,
0.73828125
],
[
0.1175537109375,
0.74072265625,
0.486083984375,
0.74072265625,
0.486083984375,
0.75439453125,
0.1175537109375,
0.75439453125
],
[
0.53076171875,
0.7421875,
0.880859375,
0.7421875,
0.880859375,
0.755859375,
0.53076171875,
0.755859375
],
[
0.118408203125,
0.75732421875,
0.48779296875,
0.75732421875,
0.48779296875,
0.77099609375,
0.118408203125,
0.77099609375
],
[
0.51123046875,
0.75732421875,
0.87890625,
0.75732421875,
0.87890625,
0.77099609375,
0.51123046875,
0.77099609375
],
[
0.118408203125,
0.7724609375,
0.48779296875,
0.7724609375,
0.48779296875,
0.7861328125,
0.118408203125,
0.7861328125
],
[
0.51318359375,
0.77490234375,
0.880859375,
0.77490234375,
0.880859375,
0.7880859375,
0.51318359375,
0.7880859375
],
[
0.1175537109375,
0.7880859375,
0.48779296875,
0.78759765625,
0.48779296875,
0.8017578125,
0.1175537109375,
0.802734375
],
[
0.509765625,
0.7890625,
0.87890625,
0.7890625,
0.87890625,
0.802734375,
0.509765625,
0.802734375
],
[
0.118408203125,
0.80712890625,
0.484375,
0.80712890625,
0.484375,
0.81787109375,
0.118408203125,
0.81787109375
],
[
0.51318359375,
0.80712890625,
0.8818359375,
0.80712890625,
0.8818359375,
0.82080078125,
0.51318359375,
0.82080078125
],
[
0.118408203125,
0.822265625,
0.2283935546875,
0.822265625,
0.2283935546875,
0.83447265625,
0.118408203125,
0.83447265625
],
[
0.51318359375,
0.822265625,
0.880859375,
0.822265625,
0.880859375,
0.8359375,
0.51318359375,
0.8359375
],
[
0.1368408203125,
0.83740234375,
0.486083984375,
0.83740234375,
0.486083984375,
0.85107421875,
0.1368408203125,
0.85107421875
],
[
0.51318359375,
0.83837890625,
0.8818359375,
0.83837890625,
0.8818359375,
0.8515625,
0.51318359375,
0.8515625
],
[
0.1158447265625,
0.8515625,
0.484375,
0.8525390625,
0.484375,
0.86767578125,
0.1158447265625,
0.8662109375
],
[
0.51318359375,
0.85498046875,
0.880859375,
0.85498046875,
0.880859375,
0.8681640625,
0.51318359375,
0.8681640625
],
[
0.118408203125,
0.869140625,
0.48779296875,
0.869140625,
0.48779296875,
0.8828125,
0.118408203125,
0.8828125
],
[
0.51318359375,
0.87158203125,
0.568359375,
0.87158203125,
0.568359375,
0.8828125,
0.51318359375,
0.8828125
],
[
0.1376953125,
0.892578125,
0.484375,
0.8935546875,
0.484375,
0.90771484375,
0.1376953125,
0.90673828125
],
[
0.52880859375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.52880859375,
0.90283203125
],
[
0.118408203125,
0.90625,
0.280517578125,
0.90478515625,
0.280517578125,
0.91845703125,
0.118408203125,
0.91943359375
],
[
0.51318359375,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"预训练语言模型的动态知识提炼",
"Lei Lit, Yankai Lin', Shuhuai Rent, Peng Lis, Jie Zhous, Xu Sun",
"fMOE Key Laboratory of Computational Linguistics, School of EECS,Peking University",
"Pattern Recognition Center, WeChat AI, Tencent Inc., China",
"( lilei, shuhuai ren J @stu.pku.edu.cn xusun@pku.edu.cn",
"( yankailin, patrickpli, withtomzhou @tecent.com",
"Abstract",
"型的输出概率分布对齐。也就是说,KD的以下",
"三个方面是预先指定的，在学习过程中保持不",
"Knowledge distillation (KD) has been proved",
"变：(1）要学习的教师模型（学习目标);(2）用于",
"effective for compressing large-scale pre-",
"查询教师的数据（学习材料);(3）目标函数和相",
"trained language models. However, existing",
"应的权重(学习方法)。但是，随着学生能力在",
"methods conduct KD statically, e.g., the stu-",
"dent model aligns its output distribution to that",
"培训阶段的发展，预先定义这些学习设置并使",
"of a selected teacher model on the pre-defined",
"其保持不变尤其不合理。静态地进行KD可能",
"training dataset.In this paper, we explore",
"会导致（1）从太大的老师那里进行不合格的",
"whether a dynamic knowledge distillation that",
"学习，（2）在学生已经掌握的实例上进行重复",
"empowers the student to adjust the learning",
"学习，以及（3）在不必要的排列上进行次优学",
"procedure according to its competency, regard",
"习。这促使我们去探索一个有趣的问题：一个",
"ing the student performance and learning effi-",
"动态的KD框架，考虑到学生在培训过程中的",
"ciency. We explore the dynamical adjustments",
"能力演变，是否能带来好处，关于学生的表现",
"onthreeaspects:teachermodeladoption,data",
"selection,andKDobjectiveadaptation.Ex",
"和学习效率？",
"perimental results show that (1) proper selec-",
"在本文中，我们提出了一个动态知识蒸",
"tion ofteacher model canboosttheperfor",
"馏E（DynamicKD）框架，该框架试图使学",
"mance of student model; (2) conducting KD",
"生能够根据其能力调整学习过程。具体来说",
"with10%informativeinstancesachievescom",
"受到主动学习（Settles，2009）的成功启发，我",
"parable performance while greatly accelerates",
"the training; (3) the student performance can",
"们将预测不确定性。g。，预测分类概率分布",
"beboosted by adjusting thesupervision con",
"的熵，作为学生能力的代理。我们努力回答以",
"tribution of different alignment objective. We",
"下研究问题：（RQ1）随着学生的发展，哪个",
"find dynamic knowledge distillation is promis-",
"老师适合学习？（RQ2）在整个KD阶段，哪",
"ing and provide discussions on potential future",
"些数据实际上对学生模型有用？（RQ3）在KD",
"directions towards more efficient KD methods.",
"过程中，最佳学习目标是否会改变？特别是",
"我们首先探讨了教师规模对动态知识提炼的影",
"响。其次，我们探讨动态选择学生不确定KD",
"1",
"Introduction",
"的实例是否可以带来更好的性能和培训效率的",
"知识提炼（KD)（Hintonetal.2015）旨在将知识",
"权衡。第三，我们探讨了从预测概率分布的对",
"从大型教师模型转移到小型学生模型。近年来，",
"齐和整个KD阶段的中间表示的对齐来动态调",
"它已被广泛使用E(Sanhetal.，2019;Jiaoetal.，",
"整监督是否可以提高性能",
"2020;Sunetal.,2019)来压缩像BERTE(Devlin",
"我们的实验结果表明：（1）具有更多层的较",
"etal.,2019)和RoBERTaE(Liuetal.,2019)这样",
"大教师模型可能会培养出更差的学生。我们表",
"的大规模预训练语言模型E(PLM)。通过知识",
"明，根据学生的能力选择合适的教师模型可以",
"蒸馏，我们可以获得一个小得多的模型，具有",
"提高成绩。(2）我们可以只使用根据学生预测",
"可比的性能，同时大大降低了内存使用量并加",
"不确定性选择的10个%信息实例来实现可比",
"速了模型推理",
"的性能。随着学生变得更强壮，这些情况也会",
"虽然简单有效，但现有方法通常静态地执行",
"在培训过程中发展。(3)我们可以通过动态调",
"KD学习过程，例如。g。，学生模型将其输出",
"整教师模型的不同对齐目标的监督来提高学生",
"概率分布与整个预定义语料库上选定教师模",
"的成绩",
"'Our code is available at https://github.com/",
"我们的观察证明了当前静态KD框架的局限",
"lancopku/DynamicKD",
"性。提出的基于不确定性的动态KD框架只是"
] |
{
"arxiv_id": "2109.11295",
"language": "zh",
"page_idx": 1
}
|
||
2109.11295-zh-3
|
2109.11295-zh-3.jpg
|
<FigureHere>
Figure 2: Performance gain of distilled student model with various layer sizes. The teacher model is BERT BASE with 12 layers.
$\mathrm{BERT}_{\mathrm{LARGE}}$老师明显优于小型 BERT$_\mathrm{~BASE}$老师模型，但 BERT $_{\mathrm{BASE}}$老师提炼的学生模型在所有三个数据集上都取得了更好的性能。这种现象是违反直觉的，因为更大的老师应该为学生模型提供更好的监督信号。我们认为，关于教师模型的大小，有两个可能的因素会导致绩效下降：
(1) 随着教师模型变得更大并且对其预测$\boxed{\mathrm{E}}$ (Guo et al., 2017; Desai and Durrett, 2020) 更有信心，教师模型的预测日志变得不那么软，这降低了通过软目标进行知识转移的效果。我们发现，较小的$\tau$也会导致学生模型的性能下降，这表明较少软化的教师预测会降低学生的成绩。$^{2}$
(2)教师和学生模型之间的容量差距随着教师的变大而增加。学生模型的能力无法与大型教师模型的能力相匹配，这削弱了KD 的表现
为了探索这些因素的综合影响，我们提取了具有不同层的学生模型，并在图$\overline{\mathrm{E}}$2 中绘制了与直接训练学生模型而不进行蒸馏相比的性能增益。可以发现，通过减少学生人数，教师模型的更好监督可以提高绩效，而随着差距变得越来越大，两个反作用因素占主导地位，从而降低了绩效。我们注意到，Mirzadeh et al. (2020)在使用卷积网络的计算机视觉任务中也观察到了这种现象，表明这是一个普遍的问题，需要更深入的研究。请注意，BERT $_{\mathrm{BASE}}$和$\mathrm{BERT}_{\mathrm{LARGE}}$也不同于隐藏大小的数量，关于隐藏大小的实验，其中也存在这种现象，相应的结果可以在附录 C 中找到
3.1.2 Uncertainty-based Teacher Adoption
我们的初步观察表明，为 KD 选择合适的教师模型对学生的表现很重要。虽然一旦设置了老师和学生，能力差距是一个固有的问题，但我们很好奇，在培训期间根据学生的能力动态查询合适的老师是否可以充分利用教师模型。在不失一般性的情况下，我们进行 KD 从具有不
$^{2}$Refer to Appendix B for details
同数量的 Transformer 层的两个教师模型中训练学生模型。我们假设在最初的培训阶段，学生可以更多地依赖小老师模型，而在大老师变得更强时转向大老师进行更准确的监督。具体来说，我们建议利用学生预测不确定性作为能力的代理，受到主动学习 (Settles,2009)中成功应用的启发，并设计两种基于不确定性的教师采用策略：
硬选择：一个批次中的实例根据学生预测不确定性进行排序，i。e。, 预测类分布的熵。然后将实例平均分为学生最不确定的实例和模型最自信的实例。对于不确定的部分，小老师会询问监督信号，而大老师会为学生有信心的实例提供软标签
软选择：来自两个教师的相应 KD 损失权重在实例级别进行了软调整。形式上，给定两个教师模型$T_1$ $\operatorname{E}($在我们的例子中为 BERT $_\text{BASE})$ 和$T_2\operatorname{E}($BERT $_\mathrm{LARGE})$,我们可以将多教师设置中的 KD 目标重写为：。$w_1$和$w_2$控制着两位老师对监督的相对贡献。当学生对培训实例不确定时，我们会自适应地降低大老师的监督。采用预测不确定性作为学生能力的度量，例如$x:$
$$u_x=\text{Entropy}\left(\sigma\left(S\left(x\right)\right)\right)$$
(4)
,其中$\sigma$是归一化函数，e。g, 用于将 logit 向量映射到概率分布的 softmax 函数。$w_1$和$w_2$ 调整如下：
(5)
$$w_1=\dfrac{u_x}{U},\quad w_2=1-\dfrac{u_x}{U}$$
其中$U$ 是一个归一化因子，它将权重重新缩放为[0,1] 。这样 , 学生在不确定当前实例时将更多地关注小老师，而在对大老师的预测充满信心时将更多地依赖大老师。
# 3.13 Experiments
Settings 我们进行实验，在 RTE, CoLA 和IMDB 上从 BERT $_\text{BASE 和 BERT }_{\text{LARGE 中提取 6}}$ 层学生模型，遵循E3.1.1 节中的探测分析设置Results 提出的选择策略的结果如表 1 所示。我们观察到，硬选择策略总体上达到了 65。3准确性，优于直接从两个教师模型的集合中学习。这表明，通过选择合适的教师模型来学习， 所提出的策略是有效的。软选择策略的表现也优于基准，而硬选择策略的表现几乎没有落后于基准。。我们将其归因于两位老师提供的监督具有不同的柔软度，因此可能会混淆学生模型
## 3.2 Dynamic Data Selection
我们要探讨的第二个研究问题是，哪些数据将更有利于学生的表现。随着蒸馏的进行，学生
|
[
[
0.1376953125,
0.08612060546875,
0.15869140625,
0.08905029296875,
0.1561279296875,
0.0955810546875,
0.1351318359375,
0.0926513671875
],
[
0.297119140625,
0.08489990234375,
0.334228515625,
0.08489990234375,
0.334228515625,
0.163330078125,
0.297119140625,
0.163330078125
],
[
0.515625,
0.0885009765625,
0.87744140625,
0.0885009765625,
0.87744140625,
0.09857177734375,
0.515625,
0.09857177734375
],
[
0.1368408203125,
0.097412109375,
0.15869140625,
0.097412109375,
0.15869140625,
0.10809326171875,
0.1368408203125,
0.10809326171875
],
[
0.31982421875,
0.10272216796875,
0.32666015625,
0.10809326171875,
0.31982421875,
0.11224365234375,
0.312255859375,
0.10687255859375
],
[
0.51318359375,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.12176513671875,
0.114013671875,
0.16455078125,
0.114013671875,
0.16455078125,
0.1591796875,
0.12176513671875,
0.1591796875
],
[
0.1234130859375,
0.10809326171875,
0.156982421875,
0.10272216796875,
0.162109375,
0.1181640625,
0.1275634765625,
0.12353515625
],
[
0.51318359375,
0.1175537109375,
0.87890625,
0.1181640625,
0.87890625,
0.1318359375,
0.51318359375,
0.1312255859375
],
[
0.313232421875,
0.127685546875,
0.32421875,
0.127685546875,
0.32421875,
0.1395263671875,
0.313232421875,
0.1395263671875
],
[
0.5146484375,
0.1353759765625,
0.880859375,
0.1353759765625,
0.880859375,
0.1484375,
0.5146484375,
0.1484375
],
[
0.1292724609375,
0.152587890625,
0.156982421875,
0.1561279296875,
0.1561279296875,
0.1644287109375,
0.1275634765625,
0.160888671875
],
[
0.30322265625,
0.15673828125,
0.32421875,
0.15380859375,
0.328369140625,
0.1650390625,
0.307373046875,
0.168701171875
],
[
0.515625,
0.152587890625,
0.87890625,
0.152587890625,
0.87890625,
0.163330078125,
0.515625,
0.163330078125
],
[
0.139404296875,
0.1650390625,
0.1561279296875,
0.1650390625,
0.1561279296875,
0.17578125,
0.139404296875,
0.17578125
],
[
0.5146484375,
0.16748046875,
0.87744140625,
0.16748046875,
0.87744140625,
0.1776123046875,
0.5146484375,
0.1776123046875
],
[
0.139404296875,
0.1817626953125,
0.154541015625,
0.1817626953125,
0.154541015625,
0.1859130859375,
0.139404296875,
0.1859130859375
],
[
0.1695556640625,
0.18408203125,
0.285400390625,
0.1859130859375,
0.285400390625,
0.1995849609375,
0.1695556640625,
0.19775390625
],
[
0.355224609375,
0.188232421875,
0.450927734375,
0.188232421875,
0.450927734375,
0.1953125,
0.355224609375,
0.1953125
],
[
0.51318359375,
0.182861328125,
0.880859375,
0.182861328125,
0.880859375,
0.196533203125,
0.51318359375,
0.196533203125
],
[
0.5146484375,
0.20068359375,
0.61279296875,
0.20068359375,
0.61279296875,
0.21142578125,
0.5146484375,
0.21142578125
],
[
0.1175537109375,
0.208984375,
0.484375,
0.208984375,
0.484375,
0.22265625,
0.1175537109375,
0.22265625
],
[
0.51318359375,
0.215576171875,
0.880859375,
0.215576171875,
0.880859375,
0.2286376953125,
0.51318359375,
0.2286376953125
],
[
0.118408203125,
0.223876953125,
0.486083984375,
0.223876953125,
0.486083984375,
0.2369384765625,
0.118408203125,
0.2369384765625
],
[
0.51123046875,
0.232177734375,
0.87890625,
0.232177734375,
0.87890625,
0.2452392578125,
0.51123046875,
0.2452392578125
],
[
0.1175537109375,
0.240478515625,
0.246826171875,
0.2392578125,
0.246826171875,
0.25048828125,
0.1175537109375,
0.251708984375
],
[
0.51123046875,
0.2476806640625,
0.87890625,
0.2476806640625,
0.87890625,
0.26123046875,
0.51123046875,
0.26123046875
],
[
0.515625,
0.26416015625,
0.880859375,
0.26416015625,
0.880859375,
0.27783203125,
0.515625,
0.27783203125
],
[
0.118408203125,
0.27978515625,
0.48779296875,
0.27978515625,
0.48779296875,
0.29443359375,
0.118408203125,
0.29443359375
],
[
0.51318359375,
0.27978515625,
0.880859375,
0.27978515625,
0.880859375,
0.29345703125,
0.51318359375,
0.29345703125
],
[
0.118408203125,
0.29638671875,
0.486083984375,
0.29638671875,
0.486083984375,
0.31005859375,
0.118408203125,
0.31005859375
],
[
0.51318359375,
0.29638671875,
0.603515625,
0.29638671875,
0.603515625,
0.30810546875,
0.51318359375,
0.30810546875
],
[
0.1175537109375,
0.311279296875,
0.484375,
0.311279296875,
0.484375,
0.32470703125,
0.1175537109375,
0.32470703125
],
[
0.51123046875,
0.312255859375,
0.880859375,
0.312255859375,
0.880859375,
0.325927734375,
0.51123046875,
0.325927734375
],
[
0.12005615234375,
0.330078125,
0.484375,
0.330078125,
0.484375,
0.34033203125,
0.12005615234375,
0.34033203125
],
[
0.51318359375,
0.327880859375,
0.880859375,
0.327880859375,
0.880859375,
0.341552734375,
0.51318359375,
0.341552734375
],
[
0.118408203125,
0.344482421875,
0.486083984375,
0.344482421875,
0.486083984375,
0.358154296875,
0.118408203125,
0.358154296875
],
[
0.51123046875,
0.342529296875,
0.8818359375,
0.34326171875,
0.8818359375,
0.358154296875,
0.51123046875,
0.35693359375
],
[
0.1175537109375,
0.359375,
0.484375,
0.359375,
0.484375,
0.372314453125,
0.1175537109375,
0.372314453125
],
[
0.51318359375,
0.35986328125,
0.880859375,
0.35986328125,
0.880859375,
0.374755859375,
0.51318359375,
0.374755859375
],
[
0.118408203125,
0.377197265625,
0.16455078125,
0.377197265625,
0.16455078125,
0.391357421875,
0.118408203125,
0.391357421875
],
[
0.51318359375,
0.3759765625,
0.880859375,
0.3759765625,
0.880859375,
0.3896484375,
0.51318359375,
0.3896484375
],
[
0.1351318359375,
0.393798828125,
0.48779296875,
0.393798828125,
0.48779296875,
0.40673828125,
0.1351318359375,
0.40673828125
],
[
0.5146484375,
0.394287109375,
0.87890625,
0.394287109375,
0.87890625,
0.405029296875,
0.5146484375,
0.405029296875
],
[
0.118408203125,
0.410888671875,
0.48193359375,
0.410888671875,
0.48193359375,
0.421630859375,
0.118408203125,
0.421630859375
],
[
0.51318359375,
0.4091796875,
0.880859375,
0.4091796875,
0.880859375,
0.4228515625,
0.51318359375,
0.4228515625
],
[
0.1175537109375,
0.42578125,
0.484375,
0.42578125,
0.484375,
0.439453125,
0.1175537109375,
0.439453125
],
[
0.51123046875,
0.424560546875,
0.87890625,
0.424560546875,
0.87890625,
0.438232421875,
0.51123046875,
0.438232421875
],
[
0.118408203125,
0.443603515625,
0.483642578125,
0.443603515625,
0.483642578125,
0.45361328125,
0.118408203125,
0.45361328125
],
[
0.509765625,
0.441162109375,
0.5380859375,
0.4423828125,
0.53662109375,
0.45361328125,
0.5087890625,
0.453125
],
[
0.60205078125,
0.457275390625,
0.791015625,
0.457275390625,
0.791015625,
0.47021484375,
0.60205078125,
0.47021484375
],
[
0.85791015625,
0.457763671875,
0.880859375,
0.457763671875,
0.880859375,
0.47021484375,
0.85791015625,
0.47021484375
],
[
0.1175537109375,
0.474365234375,
0.484375,
0.474365234375,
0.484375,
0.488037109375,
0.1175537109375,
0.488037109375
],
[
0.12005615234375,
0.460205078125,
0.48193359375,
0.460205078125,
0.48193359375,
0.47021484375,
0.12005615234375,
0.47021484375
],
[
0.51708984375,
0.47802734375,
0.5322265625,
0.48095703125,
0.5283203125,
0.4892578125,
0.51318359375,
0.485107421875
],
[
0.52490234375,
0.47802734375,
0.880859375,
0.47802734375,
0.880859375,
0.4921875,
0.52490234375,
0.4921875
],
[
0.1175537109375,
0.490478515625,
0.1822509765625,
0.488037109375,
0.183837890625,
0.5029296875,
0.118408203125,
0.5048828125
],
[
0.51318359375,
0.4921875,
0.880859375,
0.49462890625,
0.880859375,
0.5087890625,
0.51318359375,
0.50732421875
],
[
0.1368408203125,
0.50732421875,
0.486083984375,
0.50732421875,
0.486083984375,
0.52099609375,
0.1368408203125,
0.52099609375
],
[
0.5146484375,
0.5126953125,
0.59423828125,
0.5126953125,
0.59423828125,
0.5224609375,
0.5146484375,
0.5224609375
],
[
0.1175537109375,
0.5224609375,
0.484375,
0.521484375,
0.484375,
0.5361328125,
0.1175537109375,
0.53759765625
],
[
0.6455078125,
0.52978515625,
0.66650390625,
0.52978515625,
0.66650390625,
0.5419921875,
0.6455078125,
0.5419921875
],
[
0.7734375,
0.533203125,
0.791015625,
0.533203125,
0.791015625,
0.54150390625,
0.7734375,
0.54150390625
],
[
0.76416015625,
0.5361328125,
0.7734375,
0.5361328125,
0.7734375,
0.54052734375,
0.76416015625,
0.54052734375
],
[
0.599609375,
0.53369140625,
0.68603515625,
0.54345703125,
0.68115234375,
0.5625,
0.59423828125,
0.552734375
],
[
0.6767578125,
0.5380859375,
0.76318359375,
0.53515625,
0.76416015625,
0.5498046875,
0.67822265625,
0.552734375
],
[
0.85546875,
0.5380859375,
0.880859375,
0.5380859375,
0.880859375,
0.55029296875,
0.85546875,
0.55029296875
],
[
0.755859375,
0.54638671875,
0.77490234375,
0.54638671875,
0.77490234375,
0.55419921875,
0.755859375,
0.55419921875
],
[
0.771484375,
0.54638671875,
0.79443359375,
0.54638671875,
0.79443359375,
0.55712890625,
0.771484375,
0.55712890625
],
[
0.1368408203125,
0.55712890625,
0.484375,
0.55712890625,
0.484375,
0.57080078125,
0.1368408203125,
0.57080078125
],
[
0.12005615234375,
0.54150390625,
0.483642578125,
0.54150390625,
0.483642578125,
0.5517578125,
0.12005615234375,
0.5517578125
],
[
0.509765625,
0.56396484375,
0.880859375,
0.5654296875,
0.880859375,
0.580078125,
0.509765625,
0.5791015625
],
[
0.118408203125,
0.572265625,
0.484375,
0.572265625,
0.484375,
0.5859375,
0.118408203125,
0.5859375
],
[
0.51123046875,
0.58056640625,
0.87744140625,
0.58056640625,
0.87744140625,
0.59423828125,
0.51123046875,
0.59423828125
],
[
0.12005615234375,
0.59033203125,
0.484375,
0.59033203125,
0.484375,
0.60107421875,
0.12005615234375,
0.60107421875
],
[
0.51318359375,
0.59716796875,
0.880859375,
0.59716796875,
0.880859375,
0.61083984375,
0.51318359375,
0.61083984375
],
[
0.1175537109375,
0.60400390625,
0.484375,
0.60400390625,
0.484375,
0.61767578125,
0.1175537109375,
0.61767578125
],
[
0.51318359375,
0.6142578125,
0.75830078125,
0.6142578125,
0.75830078125,
0.62744140625,
0.51318359375,
0.62744140625
],
[
0.12005615234375,
0.62255859375,
0.483642578125,
0.62255859375,
0.483642578125,
0.6328125,
0.12005615234375,
0.6328125
],
[
0.118408203125,
0.63720703125,
0.484375,
0.63720703125,
0.484375,
0.65087890625,
0.118408203125,
0.65087890625
],
[
0.51123046875,
0.63818359375,
0.67529296875,
0.63818359375,
0.67529296875,
0.65185546875,
0.51123046875,
0.65185546875
],
[
0.1175537109375,
0.65185546875,
0.483642578125,
0.65185546875,
0.483642578125,
0.6650390625,
0.1175537109375,
0.6650390625
],
[
0.515625,
0.65625,
0.880859375,
0.65673828125,
0.880859375,
0.67138671875,
0.515625,
0.67041015625
],
[
0.118408203125,
0.66943359375,
0.48193359375,
0.66943359375,
0.48193359375,
0.6796875,
0.118408203125,
0.6796875
],
[
0.51123046875,
0.67333984375,
0.8818359375,
0.67333984375,
0.8818359375,
0.68701171875,
0.51123046875,
0.68701171875
],
[
0.118408203125,
0.68505859375,
0.484375,
0.68505859375,
0.484375,
0.69873046875,
0.118408203125,
0.69873046875
],
[
0.51318359375,
0.6884765625,
0.880859375,
0.6884765625,
0.880859375,
0.70166015625,
0.51318359375,
0.70166015625
],
[
0.118408203125,
0.70166015625,
0.484375,
0.70166015625,
0.484375,
0.71533203125,
0.118408203125,
0.71533203125
],
[
0.1158447265625,
0.71630859375,
0.484375,
0.71728515625,
0.484375,
0.732421875,
0.1158447265625,
0.73095703125
],
[
0.515625,
0.7119140625,
0.8818359375,
0.71337890625,
0.8818359375,
0.72802734375,
0.515625,
0.72705078125
],
[
0.118408203125,
0.73388671875,
0.486083984375,
0.73388671875,
0.486083984375,
0.74755859375,
0.118408203125,
0.74755859375
],
[
0.51318359375,
0.72998046875,
0.8818359375,
0.72998046875,
0.8818359375,
0.74365234375,
0.51318359375,
0.74365234375
],
[
0.1158447265625,
0.7490234375,
0.32568359375,
0.7490234375,
0.32568359375,
0.7626953125,
0.1158447265625,
0.7626953125
],
[
0.5146484375,
0.7451171875,
0.880859375,
0.7451171875,
0.880859375,
0.7587890625,
0.5146484375,
0.7587890625
],
[
0.515625,
0.76318359375,
0.8759765625,
0.76318359375,
0.8759765625,
0.77392578125,
0.515625,
0.77392578125
],
[
0.1158447265625,
0.77783203125,
0.468505859375,
0.77783203125,
0.468505859375,
0.79150390625,
0.1158447265625,
0.79150390625
],
[
0.5146484375,
0.77783203125,
0.880859375,
0.77783203125,
0.880859375,
0.79150390625,
0.5146484375,
0.79150390625
],
[
0.118408203125,
0.798828125,
0.486083984375,
0.798828125,
0.486083984375,
0.8125,
0.118408203125,
0.8125
],
[
0.51318359375,
0.79443359375,
0.880859375,
0.79443359375,
0.880859375,
0.80810546875,
0.51318359375,
0.80810546875
],
[
0.51318359375,
0.81005859375,
0.880859375,
0.81005859375,
0.880859375,
0.82373046875,
0.51318359375,
0.82373046875
],
[
0.118408203125,
0.81640625,
0.483642578125,
0.81640625,
0.483642578125,
0.82666015625,
0.118408203125,
0.82666015625
],
[
0.118408203125,
0.83056640625,
0.486083984375,
0.83056640625,
0.486083984375,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.51318359375,
0.82763671875,
0.87744140625,
0.82763671875,
0.87744140625,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.118408203125,
0.8466796875,
0.486083984375,
0.8466796875,
0.486083984375,
0.85986328125,
0.118408203125,
0.85986328125
],
[
0.51318359375,
0.84423828125,
0.5322265625,
0.84423828125,
0.5322265625,
0.85400390625,
0.51318359375,
0.85400390625
],
[
0.118408203125,
0.86328125,
0.486083984375,
0.86328125,
0.486083984375,
0.876953125,
0.118408203125,
0.876953125
],
[
0.509765625,
0.86767578125,
0.74658203125,
0.86767578125,
0.74658203125,
0.88134765625,
0.509765625,
0.88134765625
],
[
0.118408203125,
0.87890625,
0.486083984375,
0.87890625,
0.486083984375,
0.892578125,
0.118408203125,
0.892578125
],
[
0.51123046875,
0.8896484375,
0.87890625,
0.8896484375,
0.87890625,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1376953125,
0.90380859375,
0.341796875,
0.90478515625,
0.341796875,
0.91943359375,
0.1376953125,
0.91845703125
],
[
0.515625,
0.90673828125,
0.87890625,
0.90673828125,
0.87890625,
0.91748046875,
0.515625,
0.91748046875
]
] |
[
"0.60",
"",
"同数量的Transformer层的两个教师模型中训",
"0.55",
"",
"练学生模型。我们假设在最初的培训阶段，学",
"目",
"0.50",
"生可以更多地依赖小老师模型，而在大老师变",
"",
"得更强时转向大老师进行更准确的监督。具体",
"% 0.30",
"1O",
"来说，我们建议利用学生预测不确定性作为能",
"0.25",
"力的代理，受到主动学习（Settles.2009）中成",
"0.20",
"ONumberof Student Layer",
"lumherofStudentLave",
"功应用的启发，并设计两种基于不确定性的教",
"师采用策略：",
"Figure 2: Performance gain of distilled student model",
"硬选择：一个批次中的实例根据学生预测不确",
"with various layer sizes.The teacher model is BERT",
"定性进行排序，i。e。，预测类分布的熵。然后",
"BAsE with 12 layers.",
"将实例平均分为学生最不确定的实例和模型最",
"自信的实例。对于不确定的部分，小老师会询",
"BERTLARGE老师明显优于小型BERTBASE老师",
"问监督信号，而大老师会为学生有信心的实例",
"模型，但BERTBASE老师提炼的学生模型在所",
"提供软标签",
"有三个数据集上都取得了更好的性能。这种现",
"软选择：来自两个教师的相应KD损失权重在",
"象是违反直觉的，因为更大的老师应该为学生",
"实例级别进行了软调整。形式上，给定两个教",
"模型提供更好的监督信号。我们认为，关于教",
"师模型TIE（在我们的例子中为BERTBASE）",
"师模型的大小，有两个可能的因素会导致绩效",
"和T2E（BERTLARGE），我们可以将多教师设",
"下降：",
"置中的KD目标重写为：。Wi和W2控制着两位",
"（1）随着教师模型变得更大并且对其预测E",
"老师对监督的相对贡献。当学生对培训实例不",
"(Guoetal.,2017:Desai andDurrett,2020)更有",
"确定时，我们会自适应地降低大老师的监督。",
"信心，教师模型的预测日志变得不那么软，这",
"采用预测不确定性作为学生能力的度量，例如",
"降低了通过软自标进行知识转移的效果。我们",
"α:",
"ur = Entropy (α (S (α))",
"(4)",
"降，这表明较少软化的教师预测会降低学生的",
"发现，较小的T也会导致学生模型的性能下",
"",
"其中α是归一化函数，e。g，用于将logit向",
"成绩。2",
"量映射到概率分布的softmax函数。W1和w2",
"（2）教师和学生模型之间的容量差距随着教",
"调整如下：",
"师的变大而增加。学生模型的能力无法与大型",
"ur",
"ua",
"",
"W1=”",
"W2=1-",
"(5)",
"",
"U",
"为了探索这些因素的综合影响，我们提取了",
"教师模型的能力相匹配，这削弱了KD的表现",
"其中U是一个归一化因子，它将权重重新缩",
"具有不同层的学生模型，并在图E2中绘制了",
"放为[0,1]。这样，学生在不确定当前实例时",
"与直接训练学生模型而不进行蒸留相比的性能",
"将更多地关注小老师，而在对大老师的预测充",
"增益。可以发现，通过减少学生人数，教师模",
"满信心时将更多地依赖大老师。",
"型的更好监督可以提高绩效，而随着差距变得",
"越来越大，两个反作用因素占主导地位，从而",
"3.1.3Experiments",
"降低了绩效。。我们注意到，Mirzadeh etal",
"Settings我们进行实验，在RTE，CoLA和",
"(2020）在使用卷积网络的计算机视觉任务中也",
"IMDB上从BERTBASE和BERTLARGE中提取6",
"观察到了这种现象，表明这是一个普遍的问",
"层学生模型，遵循E3.1.1节中的探测分析设置",
"题，需要更深入的研究。请注意，BERTBASE和",
"BERTLARGE也不同于隐藏大小的数量，关于隐",
"Results提出的选择策略的结果如表1所示。",
"藏大小的实验，其中也存在这种现象，相应的",
"我们观察到，硬选择策略总体上达到了65。3",
"结果可以在附录C中找到",
"准确性，优于直接从两个教师模型的集合中学",
"习。这表明，通过选择合适的教师模型来学习",
"3.1.2Uncertainty-based Teacher Adoption",
"所提出的策略是有效的。软选择策略的表现也",
"我们的初步观察表明，为KD选择合适的教师",
"优于基准，而硬选择策略的表现几乎没有落后",
"于基准。。我们将其归因于两位老师提供的监",
"模型对学生的表现很重要。虽然一日设置了老",
"师和学生，能力差距是一个固有的问题，但我",
"督具有不同的柔软度，因此可能会混淆学生模",
"们很好奇，在培训期间根据学生的能力动态查",
"型",
"询合适的老师是否可以充分利用教师模型。在",
"3.2Dynamic Data Selection",
"不失一般性的情况下，我们进行KD从具有不",
"我们要探讨的第二个研究问题是，哪些数据将",
"2Refer to Appendix B for details",
"更有利于学生的表现。随着蒸馏的进行，学生"
] |
{
"arxiv_id": "2109.11295",
"language": "zh",
"page_idx": 3
}
|
||
2109.11295-zh-4
|
2109.11295-zh-4.jpg
|
<table>
<tbody>
<tr>
<th>Method</th>
<th>$\#$FLOPs</th>
<th>SST-5</th>
<th>IMDB</th>
<th>MRPC M</th>
<th>MNLI- m/ r mr</th>
</tr>
<tr>
<td>$\mathbf{BERT}$ (Teacher) BASE</td>
<td> </td>
<td>52.0</td>
<td>89.1</td>
<td>86.8</td>
<td>84.0/84.4</td>
</tr>
<tr>
<td>Vanilla KD</td>
<td>$45.1B$</td>
<td>47.4</td>
<td>86.8</td>
<td>80.2</td>
<td>81.7/82.0</td>
</tr>
<tr>
<td>Random</td>
<td>$22.6B$</td>
<td>46.8</td>
<td>86.4</td>
<td>79.7</td>
<td>81.4/81.6</td>
</tr>
<tr>
<td>Uncertainty- Entropy</td>
<td>$28.2B$</td>
<td>46.7</td>
<td>86.8</td>
<td>79.4</td>
<td>81.5/82.0</td>
</tr>
<tr>
<td>Uncertainty-Margin</td>
<td>$28.2B$</td>
<td>46.6</td>
<td>86.8</td>
<td>79.4</td>
<td>81.4/81.9</td>
</tr>
<tr>
<td>Uncertainty-LC</td>
<td>$28.2B$</td>
<td>46.5</td>
<td>86.8</td>
<td>79.4</td>
<td>81.4/81.9</td>
</tr>
<tr>
<td>$\Delta$</td>
<td> </td>
<td>0.6</td>
<td>0.0</td>
<td>0.5</td>
<td>-0.2/0.0</td>
</tr>
</tbody>
</table>
Table 2: Dynamic data selection results with $r$ set to 0.5. Results are averaged of 3 seeds on the validation set. $\Delta$ denotes the minimal performance degradation of different selection strategies compares to vanilla KD
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>$\#$Train</th>
<th>$\#$Aug Train</th>
<th>$\#Dev$</th>
<th>$\#$Test</th>
<th>$\#Class$</th>
</tr>
<tr>
<td>SST-5</td>
<td>8.8k</td>
<td>$176k$</td>
<td>$1.1k$</td>
<td>2.2k</td>
<td>5</td>
</tr>
<tr>
<td>IMDB</td>
<td>$20k$</td>
<td>$400k$</td>
<td>$5k$</td>
<td>$25k$</td>
<td>2</td>
</tr>
<tr>
<td>MNLI</td>
<td>$393k$</td>
<td>786,0k</td>
<td>$20k$</td>
<td>$20k$</td>
<td>3</td>
</tr>
<tr>
<td>MRPC</td>
<td>3.7k</td>
<td>$74k$</td>
<td>$0.4k$</td>
<td>$1.7\mathbf{k}$</td>
<td>2</td>
</tr>
<tr>
<td>$RTE$</td>
<td>2.5k</td>
<td>$50k$</td>
<td>0.3k</td>
<td>$3k$</td>
<td>2</td>
</tr>
<tr>
<td>CoLA</td>
<td>$8.5k$</td>
<td>$170k$</td>
<td>$1k$</td>
<td>$1k$</td>
<td>2</td>
</tr>
</tbody>
</table>
Table 3: Statistics of datasets. # Aug Train denotes the number of the augmented training dataset following Jiao et al. (2020) .
变得越来越强壮，因此可以消除对学生掌握的实例的重复学习。如果有这样的实例对于学生模型的学习是至关重要的，我们是否可以只对这些实例进行 KD, 以提高学习效率？此外，重要的实例是保持静态还是随着学生模型的发展而发展？这些问题促使我们探索动态选择实例的效果。
3.2.1 Uncertainty-based Data Selection
我们建议根据学生预测的不确定性积极选择信息实例，其灵感来自主动学习的成功实践E (Settles,2009)。形式上，给定一个批次中的$N$ 实例，对于每个实例$x$ ,在学生模型的类标签$y$上对应的输出类概率分布为$P(y\mid x)=$ $\sigma\left(S\left(x\right)\right)$。我们使用以下策略以可忽略的计算开销计算$x$的不确定性得分$u_x:$
熵$\boxed{\mathrm{E}}($Settles,2009),用于测量学生预测分布的
不确定性：
(6)
$$u_{x}=-\sum_{y}P\left(y\mid x\right)\log P\left(y\mid x\right).$$
边距，计算为第一个和第二个最可能的类$y_1^*$
和 $y_2^*$之间的边距：
(7)
$$u_x=P\left(y_1^*\mid x\right)-P\left(y_2^*\mid x\right).$$
最小置信度$\mathbb{E}$(LC),表示模型关于预测类
$\hat{y}=\arg\max_yP(y\mid x)$的不确定性：
$$u_{x}=1-P\left(\hat{y}\mid x\right)$$
(8)
我们根据其预测不确定性对实例进行批量排序，并且仅选择顶部的$N\times r$实例来查询教师模型，其中$r\in(0,1]$是控制要查询的数量的选择比率。请注意，在像 IMDB 这样的二进制分类任务中，使用上述策略的选定子集是相同的。我们还设计了一个随机策略，随机选择$N\times r$实例，作为评估选择策略有效性的基准
## 3.2.2 Experiments
Settings 我们对两个情感分类数据集IMDB$\boxed{\mathrm{F}}($Maas et al.,2011)和 SST-5$\boxed{\mathrm{E}}($Socher et al., 2013) 以及包括 MRPC$\underline{\mathbb{F}}$(Dolan and Brockett, 2005) 和 MNLI$\underline{\mathrm{E}}($Williams et al., 2018) 的自然语言推理任务进行调查实验。数据集的统计信息和实现细节可以分别在表 3 和 D 中找到。我们将准确性报告为所有评估任务的性能度量。此外，我们还提供了相应的计算 FLOP 来比较学习效率。更详细地说，我们将 KDt 的计算成本$C$分为三个部分：学生转发$F_s$,教师转发$F_t$和学生反向$B_s$,用于更新参数。请注意，$F_s\approx B_s\ll F_t$ ,因为教师模型通常比学生模型大得多。通过仅从学生最不确定的$N\times r$ 实例中主动学习，成本降低为：
$$C'=N\times F_s+N\times r\times B_s+N\times r\times F_t.$$
例如，6层学生 BERT 模型的计算 FLOP 数量为11.3 B, 而 12 层 BERT 教师模型的计算 FLOP 数量为 22.5 B$\boxed{\mathrm{E}}$(Jiao et al., 2020)。当$r$设置为0.1时，总 KD 成本从 45.1 B 减少到 14.7 B
Results with Original Dataset 将$r$ 设置为0.5时的结果列于表$\overline{\mathbb{E}}$2 。总体而言，可以发现，与VanillaKD 相比，选择实例只会导致学生成绩的下降可以忽略不计，这表明了基于不确定性的策略的有效性。有趣的是，随机策略的表现与基于不确定性的策略非常接近，我们将其归因于基础信息数据空间也可以被随机选择的实例覆盖。此外，我们注意到，在具有较大训练数据大小的任务上，性能下降较小。例如，选择具有预测熵的信息实例会导致0。2 在由 393 k 训练实例组成的 MNLI 数据集上的准确性下降，而导致0。使用 3.7 k 实例的 MRPC 上的 8个性能下降。一个可能的原因是，对于微小的数据集，底层数据分布没有被训练数据很好地覆盖，因此进一步下采样训练数据会导致更大的性能差距。为了验证这一点，我们转向使用增强技术丰富原始训练数据集的设置
|
[
[
0.1275634765625,
0.0831298828125,
0.17041015625,
0.08489990234375,
0.1695556640625,
0.1009521484375,
0.1259765625,
0.09857177734375
],
[
0.235107421875,
0.0872802734375,
0.478515625,
0.0872802734375,
0.478515625,
0.097412109375,
0.235107421875,
0.097412109375
],
[
0.52294921875,
0.0885009765625,
0.5751953125,
0.0885009765625,
0.5751953125,
0.09979248046875,
0.52294921875,
0.09979248046875
],
[
0.7119140625,
0.0885009765625,
0.767578125,
0.0885009765625,
0.767578125,
0.09979248046875,
0.7119140625,
0.09979248046875
],
[
0.7734375,
0.08489990234375,
0.8759765625,
0.0872802734375,
0.8759765625,
0.1015625,
0.7734375,
0.09979248046875
],
[
0.1259765625,
0.09979248046875,
0.231689453125,
0.1015625,
0.231689453125,
0.1134033203125,
0.1259765625,
0.11102294921875
],
[
0.5693359375,
0.08905029296875,
0.71875,
0.08905029296875,
0.71875,
0.09979248046875,
0.5693359375,
0.09979248046875
],
[
0.340087890625,
0.09979248046875,
0.361083984375,
0.105712890625,
0.353515625,
0.11639404296875,
0.333251953125,
0.10986328125
],
[
0.3779296875,
0.09979248046875,
0.4013671875,
0.10394287109375,
0.397216796875,
0.11517333984375,
0.372802734375,
0.11102294921875
],
[
0.424072265625,
0.1015625,
0.480224609375,
0.1015625,
0.480224609375,
0.11224365234375,
0.424072265625,
0.11224365234375
],
[
0.52294921875,
0.1051025390625,
0.5615234375,
0.1051025390625,
0.5615234375,
0.1175537109375,
0.52294921875,
0.1175537109375
],
[
0.599609375,
0.105712890625,
0.62744140625,
0.105712890625,
0.62744140625,
0.119384765625,
0.599609375,
0.119384765625
],
[
0.6826171875,
0.105712890625,
0.71533203125,
0.105712890625,
0.71533203125,
0.1181640625,
0.6826171875,
0.1181640625
],
[
0.736328125,
0.105712890625,
0.765625,
0.105712890625,
0.765625,
0.1181640625,
0.736328125,
0.1181640625
],
[
0.78662109375,
0.105712890625,
0.81689453125,
0.105712890625,
0.81689453125,
0.1181640625,
0.78662109375,
0.1181640625
],
[
0.84814453125,
0.10687255859375,
0.85791015625,
0.10687255859375,
0.85791015625,
0.1181640625,
0.84814453125,
0.1181640625
],
[
0.295654296875,
0.11102294921875,
0.32421875,
0.114013671875,
0.32080078125,
0.12469482421875,
0.29296875,
0.1217041015625
],
[
0.334228515625,
0.11102294921875,
0.359375,
0.11102294921875,
0.359375,
0.122314453125,
0.334228515625,
0.122314453125
],
[
0.372802734375,
0.11102294921875,
0.399658203125,
0.11102294921875,
0.399658203125,
0.12353515625,
0.372802734375,
0.12353515625
],
[
0.2435302734375,
0.11224365234375,
0.274658203125,
0.11224365234375,
0.274658203125,
0.12054443359375,
0.2435302734375,
0.12054443359375
],
[
0.297119140625,
0.1015625,
0.322509765625,
0.1015625,
0.322509765625,
0.1134033203125,
0.297119140625,
0.1134033203125
],
[
0.1292724609375,
0.1134033203125,
0.185546875,
0.1134033203125,
0.185546875,
0.12353515625,
0.1292724609375,
0.12353515625
],
[
0.424072265625,
0.11224365234375,
0.480224609375,
0.11224365234375,
0.480224609375,
0.12353515625,
0.424072265625,
0.12353515625
],
[
0.67919921875,
0.11639404296875,
0.71533203125,
0.1181640625,
0.7138671875,
0.1312255859375,
0.67822265625,
0.12890625
],
[
0.74462890625,
0.1175537109375,
0.767578125,
0.1175537109375,
0.767578125,
0.1312255859375,
0.74462890625,
0.1312255859375
],
[
0.78857421875,
0.114013671875,
0.8193359375,
0.1175537109375,
0.81689453125,
0.1341552734375,
0.78515625,
0.1312255859375
],
[
0.599609375,
0.1181640625,
0.63037109375,
0.1181640625,
0.63037109375,
0.133056640625,
0.599609375,
0.133056640625
],
[
0.5224609375,
0.1181640625,
0.5634765625,
0.1181640625,
0.5634765625,
0.1300048828125,
0.5224609375,
0.1300048828125
],
[
0.84814453125,
0.1181640625,
0.85986328125,
0.1181640625,
0.85986328125,
0.1312255859375,
0.84814453125,
0.1312255859375
],
[
0.1275634765625,
0.1270751953125,
0.172119140625,
0.1270751953125,
0.172119140625,
0.1383056640625,
0.1275634765625,
0.1383056640625
],
[
0.295654296875,
0.127685546875,
0.322509765625,
0.127685546875,
0.322509765625,
0.14013671875,
0.295654296875,
0.14013671875
],
[
0.3359375,
0.127685546875,
0.361083984375,
0.127685546875,
0.361083984375,
0.1395263671875,
0.3359375,
0.1395263671875
],
[
0.372802734375,
0.127685546875,
0.399658203125,
0.127685546875,
0.399658203125,
0.14013671875,
0.372802734375,
0.14013671875
],
[
0.424072265625,
0.127685546875,
0.480224609375,
0.127685546875,
0.480224609375,
0.1395263671875,
0.424072265625,
0.1395263671875
],
[
0.59375,
0.127685546875,
0.63037109375,
0.1312255859375,
0.62744140625,
0.1455078125,
0.59033203125,
0.142578125
],
[
0.79248046875,
0.127685546875,
0.8193359375,
0.1312255859375,
0.8154296875,
0.1455078125,
0.78857421875,
0.142578125
],
[
0.6689453125,
0.1300048828125,
0.71533203125,
0.1300048828125,
0.71533203125,
0.144287109375,
0.6689453125,
0.144287109375
],
[
0.2418212890625,
0.12890625,
0.27783203125,
0.12890625,
0.27783203125,
0.14013671875,
0.2418212890625,
0.14013671875
],
[
0.5224609375,
0.1312255859375,
0.5634765625,
0.1312255859375,
0.5634765625,
0.144287109375,
0.5224609375,
0.144287109375
],
[
0.73876953125,
0.1312255859375,
0.765625,
0.1312255859375,
0.765625,
0.1436767578125,
0.73876953125,
0.1436767578125
],
[
0.84814453125,
0.1312255859375,
0.85791015625,
0.1312255859375,
0.85791015625,
0.1436767578125,
0.84814453125,
0.1436767578125
],
[
0.372802734375,
0.133056640625,
0.399658203125,
0.1353759765625,
0.39794921875,
0.1484375,
0.370361328125,
0.146728515625
],
[
0.2418212890625,
0.1341552734375,
0.27783203125,
0.135986328125,
0.276123046875,
0.1484375,
0.240966796875,
0.146728515625
],
[
0.2939453125,
0.135986328125,
0.322509765625,
0.135986328125,
0.322509765625,
0.1478271484375,
0.2939453125,
0.1478271484375
],
[
0.334228515625,
0.1353759765625,
0.361083984375,
0.1353759765625,
0.361083984375,
0.1478271484375,
0.334228515625,
0.1478271484375
],
[
0.4248046875,
0.135986328125,
0.480224609375,
0.135986328125,
0.480224609375,
0.1478271484375,
0.4248046875,
0.1478271484375
],
[
0.2939453125,
0.142578125,
0.32666015625,
0.1455078125,
0.32080078125,
0.17041015625,
0.2880859375,
0.16748046875
],
[
0.5205078125,
0.1436767578125,
0.56689453125,
0.1436767578125,
0.56689453125,
0.157958984375,
0.5205078125,
0.157958984375
],
[
0.1292724609375,
0.13720703125,
0.231689453125,
0.13720703125,
0.231689453125,
0.1478271484375,
0.1292724609375,
0.1478271484375
],
[
0.1275634765625,
0.144287109375,
0.2301025390625,
0.146728515625,
0.2301025390625,
0.1602783203125,
0.1275634765625,
0.157958984375
],
[
0.2418212890625,
0.144287109375,
0.27783203125,
0.146728515625,
0.276123046875,
0.1591796875,
0.240966796875,
0.15673828125
],
[
0.3359375,
0.144287109375,
0.361083984375,
0.1478271484375,
0.358642578125,
0.1591796875,
0.333251953125,
0.1561279296875
],
[
0.59765625,
0.144287109375,
0.62890625,
0.144287109375,
0.62890625,
0.15673828125,
0.59765625,
0.15673828125
],
[
0.6884765625,
0.144287109375,
0.71728515625,
0.144287109375,
0.71728515625,
0.15673828125,
0.6884765625,
0.15673828125
],
[
0.73486328125,
0.144287109375,
0.767578125,
0.144287109375,
0.767578125,
0.15673828125,
0.73486328125,
0.15673828125
],
[
0.78857421875,
0.144287109375,
0.81787109375,
0.144287109375,
0.81787109375,
0.15673828125,
0.78857421875,
0.15673828125
],
[
0.8447265625,
0.1436767578125,
0.861328125,
0.1436767578125,
0.861328125,
0.17041015625,
0.8447265625,
0.17041015625
],
[
0.4248046875,
0.146728515625,
0.480224609375,
0.146728515625,
0.480224609375,
0.157958984375,
0.4248046875,
0.157958984375
],
[
0.372802734375,
0.1455078125,
0.4013671875,
0.1455078125,
0.4013671875,
0.157958984375,
0.372802734375,
0.157958984375
],
[
0.240966796875,
0.155029296875,
0.27783203125,
0.155029296875,
0.27783203125,
0.16748046875,
0.240966796875,
0.16748046875
],
[
0.334228515625,
0.155029296875,
0.361083984375,
0.155029296875,
0.361083984375,
0.16748046875,
0.334228515625,
0.16748046875
],
[
0.372802734375,
0.155029296875,
0.399658203125,
0.155029296875,
0.399658203125,
0.16748046875,
0.372802734375,
0.16748046875
],
[
0.59619140625,
0.155029296875,
0.63037109375,
0.155029296875,
0.63037109375,
0.17041015625,
0.59619140625,
0.17041015625
],
[
0.424072265625,
0.1561279296875,
0.480224609375,
0.1561279296875,
0.480224609375,
0.16748046875,
0.424072265625,
0.16748046875
],
[
0.1292724609375,
0.1561279296875,
0.208251953125,
0.1561279296875,
0.208251953125,
0.16748046875,
0.1292724609375,
0.16748046875
],
[
0.6884765625,
0.1561279296875,
0.71728515625,
0.1561279296875,
0.71728515625,
0.169189453125,
0.6884765625,
0.169189453125
],
[
0.796875,
0.1561279296875,
0.81787109375,
0.1561279296875,
0.81787109375,
0.17041015625,
0.796875,
0.17041015625
],
[
0.73486328125,
0.15673828125,
0.765625,
0.15673828125,
0.765625,
0.169189453125,
0.73486328125,
0.169189453125
],
[
0.5224609375,
0.15673828125,
0.55322265625,
0.15673828125,
0.55322265625,
0.169189453125,
0.5224609375,
0.169189453125
],
[
0.59619140625,
0.16748046875,
0.63037109375,
0.16748046875,
0.63037109375,
0.182861328125,
0.59619140625,
0.182861328125
],
[
0.68115234375,
0.16748046875,
0.71728515625,
0.16748046875,
0.71728515625,
0.182861328125,
0.68115234375,
0.182861328125
],
[
0.84814453125,
0.169189453125,
0.85791015625,
0.169189453125,
0.85791015625,
0.18115234375,
0.84814453125,
0.18115234375
],
[
0.748046875,
0.17041015625,
0.76416015625,
0.17041015625,
0.76416015625,
0.18115234375,
0.748046875,
0.18115234375
],
[
0.5224609375,
0.169189453125,
0.5634765625,
0.169189453125,
0.5634765625,
0.182861328125,
0.5224609375,
0.182861328125
],
[
0.80029296875,
0.17041015625,
0.81689453125,
0.17041015625,
0.81689453125,
0.18115234375,
0.80029296875,
0.18115234375
],
[
0.42822265625,
0.171630859375,
0.478515625,
0.171630859375,
0.478515625,
0.182861328125,
0.42822265625,
0.182861328125
],
[
0.1292724609375,
0.1728515625,
0.14111328125,
0.1728515625,
0.14111328125,
0.182861328125,
0.1292724609375,
0.182861328125
],
[
0.295654296875,
0.17333984375,
0.32080078125,
0.17333984375,
0.32080078125,
0.1817626953125,
0.295654296875,
0.1817626953125
],
[
0.340087890625,
0.17333984375,
0.358642578125,
0.17333984375,
0.358642578125,
0.1817626953125,
0.340087890625,
0.1817626953125
],
[
0.3720703125,
0.171630859375,
0.399658203125,
0.171630859375,
0.399658203125,
0.182861328125,
0.3720703125,
0.182861328125
],
[
0.1175537109375,
0.196533203125,
0.489501953125,
0.196533203125,
0.489501953125,
0.210205078125,
0.1175537109375,
0.210205078125
],
[
0.51123046875,
0.19775390625,
0.880859375,
0.19775390625,
0.880859375,
0.21142578125,
0.51123046875,
0.21142578125
],
[
0.1175537109375,
0.21142578125,
0.486083984375,
0.21142578125,
0.486083984375,
0.2244873046875,
0.1175537109375,
0.2244873046875
],
[
0.51123046875,
0.2120361328125,
0.87890625,
0.2120361328125,
0.87890625,
0.2257080078125,
0.51123046875,
0.2257080078125
],
[
0.118408203125,
0.2257080078125,
0.48779296875,
0.2257080078125,
0.48779296875,
0.2392578125,
0.118408203125,
0.2392578125
],
[
0.51318359375,
0.22802734375,
0.634765625,
0.22802734375,
0.634765625,
0.2381591796875,
0.51318359375,
0.2381591796875
],
[
0.1175537109375,
0.2381591796875,
0.47265625,
0.2392578125,
0.47265625,
0.254150390625,
0.1175537109375,
0.2529296875
],
[
0.5146484375,
0.270751953125,
0.55517578125,
0.270751953125,
0.55517578125,
0.281982421875,
0.5146484375,
0.281982421875
],
[
0.54736328125,
0.26953125,
0.67333984375,
0.26953125,
0.67333984375,
0.283203125,
0.54736328125,
0.283203125
],
[
0.118408203125,
0.283203125,
0.484375,
0.283203125,
0.484375,
0.29638671875,
0.118408203125,
0.29638671875
],
[
0.52490234375,
0.29150390625,
0.87890625,
0.29150390625,
0.87890625,
0.305908203125,
0.52490234375,
0.305908203125
],
[
0.118408203125,
0.298583984375,
0.486083984375,
0.298583984375,
0.486083984375,
0.312255859375,
0.118408203125,
0.312255859375
],
[
0.509765625,
0.305908203125,
0.880859375,
0.30712890625,
0.880859375,
0.32177734375,
0.509765625,
0.320556640625
],
[
0.118408203125,
0.3154296875,
0.484375,
0.3154296875,
0.484375,
0.328857421875,
0.118408203125,
0.328857421875
],
[
0.51318359375,
0.32373046875,
0.8818359375,
0.32373046875,
0.8818359375,
0.33740234375,
0.51318359375,
0.33740234375
],
[
0.1175537109375,
0.330078125,
0.486083984375,
0.330810546875,
0.486083984375,
0.344482421875,
0.1175537109375,
0.34326171875
],
[
0.51123046875,
0.339111328125,
0.87890625,
0.34033203125,
0.87890625,
0.355224609375,
0.51123046875,
0.35400390625
],
[
0.118408203125,
0.347412109375,
0.486083984375,
0.347412109375,
0.486083984375,
0.361083984375,
0.118408203125,
0.361083984375
],
[
0.51318359375,
0.35693359375,
0.87890625,
0.35693359375,
0.87890625,
0.37060546875,
0.51318359375,
0.37060546875
],
[
0.118408203125,
0.364013671875,
0.486083984375,
0.364013671875,
0.486083984375,
0.377685546875,
0.118408203125,
0.377685546875
],
[
0.51318359375,
0.372314453125,
0.8818359375,
0.372314453125,
0.8818359375,
0.385986328125,
0.51318359375,
0.385986328125
],
[
0.118408203125,
0.38134765625,
0.1796875,
0.38134765625,
0.1796875,
0.392578125,
0.118408203125,
0.392578125
],
[
0.51123046875,
0.38720703125,
0.87890625,
0.38720703125,
0.87890625,
0.40087890625,
0.51123046875,
0.40087890625
],
[
0.1175537109375,
0.4091796875,
0.43994140625,
0.4091796875,
0.43994140625,
0.4228515625,
0.1175537109375,
0.4228515625
],
[
0.51318359375,
0.40625,
0.87744140625,
0.40625,
0.87744140625,
0.416259765625,
0.51318359375,
0.416259765625
],
[
0.5146484375,
0.42041015625,
0.87890625,
0.42041015625,
0.87890625,
0.43408203125,
0.5146484375,
0.43408203125
],
[
0.1175537109375,
0.43115234375,
0.484375,
0.43115234375,
0.484375,
0.44482421875,
0.1175537109375,
0.44482421875
],
[
0.51123046875,
0.435791015625,
0.87890625,
0.435791015625,
0.87890625,
0.449462890625,
0.51123046875,
0.449462890625
],
[
0.118408203125,
0.446533203125,
0.48779296875,
0.446533203125,
0.48779296875,
0.460205078125,
0.118408203125,
0.460205078125
],
[
0.5146484375,
0.45361328125,
0.87890625,
0.45361328125,
0.87890625,
0.46435546875,
0.5146484375,
0.46435546875
],
[
0.118408203125,
0.463134765625,
0.486083984375,
0.463134765625,
0.486083984375,
0.476806640625,
0.118408203125,
0.476806640625
],
[
0.51318359375,
0.468505859375,
0.87890625,
0.468505859375,
0.87890625,
0.482177734375,
0.51318359375,
0.482177734375
],
[
0.118408203125,
0.478515625,
0.48779296875,
0.478515625,
0.48779296875,
0.4921875,
0.118408203125,
0.4921875
],
[
0.51318359375,
0.486328125,
0.87744140625,
0.486328125,
0.87744140625,
0.496337890625,
0.51318359375,
0.496337890625
],
[
0.118408203125,
0.495361328125,
0.486083984375,
0.495361328125,
0.486083984375,
0.5087890625,
0.118408203125,
0.5087890625
],
[
0.51318359375,
0.501953125,
0.75830078125,
0.501953125,
0.75830078125,
0.515625,
0.51318359375,
0.515625
],
[
0.118408203125,
0.51123046875,
0.484375,
0.51123046875,
0.484375,
0.52490234375,
0.118408203125,
0.52490234375
],
[
0.118408203125,
0.52783203125,
0.376220703125,
0.52783203125,
0.376220703125,
0.54150390625,
0.118408203125,
0.54150390625
],
[
0.5224609375,
0.53369140625,
0.87890625,
0.53369140625,
0.87890625,
0.54736328125,
0.5224609375,
0.54736328125
],
[
0.118408203125,
0.54443359375,
0.484375,
0.54443359375,
0.484375,
0.55810546875,
0.118408203125,
0.55810546875
],
[
0.1175537109375,
0.55859375,
0.20068359375,
0.56103515625,
0.20068359375,
0.576171875,
0.1175537109375,
0.57373046875
],
[
0.515625,
0.57080078125,
0.87890625,
0.57080078125,
0.87890625,
0.58056640625,
0.515625,
0.58056640625
],
[
0.51318359375,
0.5849609375,
0.87744140625,
0.5849609375,
0.87744140625,
0.595703125,
0.51318359375,
0.595703125
],
[
0.16796875,
0.59130859375,
0.465087890625,
0.59130859375,
0.465087890625,
0.6123046875,
0.16796875,
0.6123046875
],
[
0.46337890625,
0.5966796875,
0.486083984375,
0.5966796875,
0.486083984375,
0.60693359375,
0.46337890625,
0.60693359375
],
[
0.5146484375,
0.60107421875,
0.880859375,
0.60107421875,
0.880859375,
0.6142578125,
0.5146484375,
0.6142578125
],
[
0.235107421875,
0.61083984375,
0.25439453125,
0.61083984375,
0.25439453125,
0.62451171875,
0.235107421875,
0.62451171875
],
[
0.51318359375,
0.61767578125,
0.85546875,
0.61767578125,
0.85546875,
0.63134765625,
0.51318359375,
0.63134765625
],
[
0.1175537109375,
0.64013671875,
0.486083984375,
0.642578125,
0.486083984375,
0.65673828125,
0.1175537109375,
0.65478515625
],
[
0.5146484375,
0.6455078125,
0.8818359375,
0.646484375,
0.8818359375,
0.6611328125,
0.5146484375,
0.66015625
],
[
0.118408203125,
0.6591796875,
0.264404296875,
0.6591796875,
0.264404296875,
0.6728515625,
0.118408203125,
0.6728515625
],
[
0.51318359375,
0.66455078125,
0.880859375,
0.66455078125,
0.880859375,
0.677734375,
0.51318359375,
0.677734375
],
[
0.51318359375,
0.6796875,
0.880859375,
0.6796875,
0.880859375,
0.693359375,
0.51318359375,
0.693359375
],
[
0.18896484375,
0.69140625,
0.4189453125,
0.69140625,
0.4189453125,
0.7060546875,
0.18896484375,
0.7060546875
],
[
0.462646484375,
0.69140625,
0.490234375,
0.69482421875,
0.48779296875,
0.705078125,
0.459228515625,
0.70166015625
],
[
0.51123046875,
0.69482421875,
0.880859375,
0.69482421875,
0.880859375,
0.7080078125,
0.51123046875,
0.7080078125
],
[
0.5146484375,
0.7119140625,
0.880859375,
0.7119140625,
0.880859375,
0.7255859375,
0.5146484375,
0.7255859375
],
[
0.420654296875,
0.69677734375,
0.444091796875,
0.69677734375,
0.444091796875,
0.70068359375,
0.420654296875,
0.70068359375
],
[
0.118408203125,
0.7255859375,
0.48779296875,
0.7255859375,
0.48779296875,
0.7392578125,
0.118408203125,
0.7392578125
],
[
0.51318359375,
0.72705078125,
0.880859375,
0.72802734375,
0.880859375,
0.7412109375,
0.51318359375,
0.74072265625
],
[
0.1175537109375,
0.74365234375,
0.3955078125,
0.74365234375,
0.3955078125,
0.75732421875,
0.1175537109375,
0.75732421875
],
[
0.515625,
0.7451171875,
0.87890625,
0.7451171875,
0.87890625,
0.755859375,
0.515625,
0.755859375
],
[
0.51123046875,
0.7587890625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.51123046875,
0.7724609375
],
[
0.2275390625,
0.775390625,
0.37451171875,
0.77490234375,
0.37451171875,
0.7890625,
0.2275390625,
0.79052734375
],
[
0.46337890625,
0.77685546875,
0.489501953125,
0.77685546875,
0.489501953125,
0.7880859375,
0.46337890625,
0.7880859375
],
[
0.51318359375,
0.77685546875,
0.880859375,
0.77685546875,
0.880859375,
0.79052734375,
0.51318359375,
0.79052734375
],
[
0.51318359375,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.8056640625,
0.51318359375,
0.8056640625
],
[
0.12005615234375,
0.81005859375,
0.484375,
0.81005859375,
0.484375,
0.82080078125,
0.12005615234375,
0.82080078125
],
[
0.51318359375,
0.81005859375,
0.87744140625,
0.81005859375,
0.87744140625,
0.82080078125,
0.51318359375,
0.82080078125
],
[
0.118408203125,
0.82470703125,
0.48779296875,
0.82470703125,
0.48779296875,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.51318359375,
0.82666015625,
0.87744140625,
0.82666015625,
0.87744140625,
0.83740234375,
0.51318359375,
0.83740234375
],
[
0.1175537109375,
0.84130859375,
0.484375,
0.84130859375,
0.484375,
0.85498046875,
0.1175537109375,
0.85498046875
],
[
0.51123046875,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.51123046875,
0.85498046875
],
[
0.118408203125,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51123046875,
0.85693359375,
0.87890625,
0.85693359375,
0.87890625,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.12005615234375,
0.87451171875,
0.484375,
0.87451171875,
0.484375,
0.88525390625,
0.12005615234375,
0.88525390625
],
[
0.5146484375,
0.87451171875,
0.87890625,
0.87451171875,
0.87890625,
0.88525390625,
0.5146484375,
0.88525390625
],
[
0.118408203125,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51123046875,
0.8896484375,
0.87744140625,
0.8896484375,
0.87744140625,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1175537109375,
0.90380859375,
0.483642578125,
0.90478515625,
0.483642578125,
0.91845703125,
0.1175537109375,
0.91748046875
],
[
0.51318359375,
0.90673828125,
0.80615234375,
0.90673828125,
0.80615234375,
0.91748046875,
0.51318359375,
0.91748046875
]
] |
[
"Method",
"#FLOPsSST-5IMDB MRPC MNLI-m /mm",
"Dataset",
"#Dev",
"#Test#Class",
"BERTsAsE (Teacher)",
"#Train#AugTrain",
"89.1",
"86.8",
"84.0/84.4",
"SST-5",
"8.8k",
"176k",
"1.1k",
"2.2k",
"5",
"47.4",
"86.8",
"80.2",
"45.1B",
"52.0",
"Vanilla KD",
"81.7/82.0",
"400k",
"5k",
"25k",
"20k",
"IMDB",
"2",
"Random",
"46.8",
"86.4",
"79.7",
"81.4/81.6",
"393k",
"20k",
"786,0k",
"22.6B",
"MNLI",
"20k",
"3",
"79.4",
"28.2B",
"46.7",
"86.8",
"81.5/82.0",
"465",
"MRPC",
"Uncertainty-Entropy",
"Uncertainty-Margin",
"28.2B",
"86.8",
"3.7k",
"74k",
"0.4k",
"1.7k",
"22",
"81.4/81.9",
"79.4",
"28.2B",
"86.8",
"79.4",
"2.5k",
"81.4/81.9",
"Uncertainty-LC",
"50k",
"3k",
"0.3k",
"RTE",
"8.5k",
"170k",
"2",
"1k",
"CoLA",
"1k",
" 0.2 /0.0",
"△",
"0.6",
"0.0",
" 0.5",
"Table 2: Dynamic data selection results with r set to",
"Table 3: Statistics of datasets. # Aug Train denotes",
"0.5 . Results are averaged of 3 seeds on the validation",
"the number of the augmented training dataset following",
"set. △ denotes the minimal performance degradation of",
"Jiao et al. (2020).",
"different selection strategies compares to vanilla KD",
"3.2.2",
"Experiments",
"变得越来越强壮，因此可以消除对学生掌握的",
"Settings我们对两个情感分类数据集",
"实例的重复学习。如果有这样的实例对于学生",
"IMDBE(Maas et al., 2011) 和 SST-5E(Socher",
"模型的学习是至关重要的，我们是否可以只对",
"etal.,2013)以及包括MRPCE(DolanandBrock-",
"这些实例进行KD，以提高学习效率？此外，重",
"ett,2005）和MNLIE(Williams et al.,2018)的自",
"要的实例是保持静态还是随着学生模型的发展",
"然语言推理任务进行调查实验。数据集的统计",
"而发展？这些问题促使我们探索动态选择实例",
"信息和实现细节可以分别在表3和D中找到。",
"的效果。",
"我们将准确性报告为所有评估任务的性能度",
"3.2.1Uncertainty-based Data Selection",
"量。此外，我们还提供了相应的计算FLOP来",
"比较学习效率。更详细地说，我们将KDt的计",
"我们建议根据学生预测的不确定性积极选择",
"算成本C分为三个部分：学生转发F。，教师",
"信息实例，其灵感来自主动学习的成功实践E",
"转发F和学生反向B。：用于更新参数。请注",
"(Settles,2009）。形式上，给定一个批次中的N",
"意，F～B。《Ft，因为教师模型通常比学生",
"实例，对于每个实例，在学生模型的类标",
"模型大得多。通过仅从学生最不确定的N×T",
"签y上对应的输出类概率分布为P(yla)=",
"实例中主动学习，成本降低为：",
"α（S(α)）。我们使用以下策略以可忽略的计算",
"开销计算α的不确定性得分u：",
"C'=N×F+N×r×Bs+N×r×Ft. (9)",
"滴E(Settles,2009），用于测量学生预测分布的",
"不确定性：",
"例如，6层学生BERT模型的计算FLOP数量为",
"11.3B，而12层BERT教师模型的计算FLOF",
"ur =-P(y [ a) log P(y | a).",
"(6)",
"数量为22.5BE(Jiaoetal.,2020）。当r设置为",
"y",
"0.1时，总KD成本从45.1B减少到14.7B",
"边距，计算为第一个和第二个最可能的类9",
"ResultswithOriginalDataset将r设置为0.5",
"和y2之间的边距：",
"时的结果列于表E2。总体而言，可以发现，与",
"VanillaKD相比，选择实例只会导致学生成绩",
"ur = P(y [ a)-P(y2 [ α).",
"(7)",
"的下降可以忽略不计，这表明了基于不确定性",
"的策略的有效性。有趣的是，随机策略的表现",
"",
"最小置信度E（LC），表示模型关于预测类",
"与基于不确定性的策略非常接近，我们将其归",
"g=argmaxyP(y「)的不确定性：",
"因于基础信息数据空间也可以被随机选择的实",
"例覆盖。此外，我们注意到，在具有较大训练",
"ur=1-P(g|)",
"(8)",
"数据大小的任务上，性能下降较小。例如，选",
"择具有预测熵的信息实例会导致0。2在由393",
"我们根据其预测不确定性对实例进行批量排",
"k训练实例组成的MNLI数据集上的准确性下",
"序，并且仅选择顶部的N×r实例来查询教",
"降，而导致0。使用3.7k实例的MRPC上的8",
"师模型，其中rE(O,1)是控制要查询的数量",
"个性能下降。一个可能的原因是，对于微小的",
"的选择比率。请注意，在像IMDB这样的二进",
"数据集，底层数据分布没有被训练数据很好地",
"制分类任务中，使用上述策略的选定子集是相",
"覆盖，因此进一步下采样训练数据会导致更大",
"同的。我们还设计了一个随机策略，随机选择",
"的性能差距。为了验证这一点，我们转向使用",
"N×r实例，作为评估选择策略有效性的基准",
"增强技术丰富原始训练数据集的设置"
] |
{
"arxiv_id": "2109.11295",
"language": "zh",
"page_idx": 4
}
|
||
2306.04340-zh-0
|
2306.04340-zh-0.jpg
|
# Co-evolving Graph Reasoning Network for $\textbf{Emotion- Cause Pair Extraction}$
Bowen Xing$^1,2$ and Ivor W. Tsang$^2,1$
$^{1}$ Australian Artificial Intelligence Institute, University of Technology Sydney
$^{2}{\mathrm{~Centre~for~Frontier~AI~Research,~A*STAR}}$
bwxing714@gmail.com ivor\_tsang@cfar.a-star.edu.sg
$\mathbf{Abstract.~Emotion-Cause~Pair~Extraction~(ECPE)~aims~to~extract~all}$ emotion clauses and their corresponding cause clauses from a document. Existing approaches tackle this task through multi-task learning (MTL) framework in which the two subtasks provide indicative clues for ECPE. However, the previous MTL framework considers only one round of multitask reasoning and ignores the reverse feedbacks from ECPE to the subtasks. Besides, its multi-task reasoning only relies on semantics-level interactions, which cannot capture the explicit dependencies, and both the encoder sharing and multi-task hidden states concatenations can hardly capture the causalities. To solve these issues, we first put forward a new MTL framework based on Co-evolving Reasoning. It (1) models the bidirectional feedbacks between ECPE and its subtasks; (2) allows the three tasks to evolve together and prompt each other recurrently; (3) integrates prediction-level interactions to capture explicit dependencies. Then we propose a novel multi-task relational graph (MRG) to suffciently exploit the causal relations. Finally, we propose a Co-evolving Graph Reasoning Network (CGR-Net) that implements our MTL framework and conducts Co-evolving Reasoning on MRG. Experimental results show that our model achieves new state-of-the-art performance, and further analysis confirms the advantages of our method.
Keywords: Multi-Task Learning· Relational Graph Reasoning·Emotion-
Cause Extraction $\cdot$ Natural Language Processing.
## 1 Introduction
情感原因对提取 (ECPE) 是自然语言处理/人工智能领域中一项新的具有挑战性的任务。它旨在从原始文档中自动提取所有情感子句和相应的原因子句，这对于实际应用[26]具有重要价值。考虑一个文档“[在学生$]_1,[$的记忆中，他经常为他们支付$]_2,[$的学费，这是值得尊敬和感动的$]_3$ 。”,第三个子句表达了一种情感，这种情感是由第二个子句触发的，因此这两个子句形成了情感-原因对。直观地，检测表达原因和情绪的子句，即原因提取(CE)和情绪提取(EE) 是 ECPE 的两个子任务。因此，最近的模型[10,13,1,36]实现了图中所示的多任务学习框架。1(a),引人 CE 和 EE 为 ECPE 提供指示性线索
尽管以前的多任务学习 (MTL) 框架取得了可喜的进展，但根据我们的观察，它仍然存在一些问题，这些问题阻碍了 ECPE 和 CE/EE 之间的多任务推
|
[
[
0.256591796875,
0.1458740234375,
0.7490234375,
0.1458740234375,
0.7490234375,
0.1634521484375,
0.256591796875,
0.1634521484375
],
[
0.3203125,
0.1666259765625,
0.6845703125,
0.1673583984375,
0.6845703125,
0.1849365234375,
0.3203125,
0.184326171875
],
[
0.369384765625,
0.21337890625,
0.6337890625,
0.2108154296875,
0.6337890625,
0.2291259765625,
0.369384765625,
0.2310791015625
],
[
0.24267578125,
0.2392578125,
0.7587890625,
0.240478515625,
0.7587890625,
0.255126953125,
0.24267578125,
0.25390625
],
[
0.35546875,
0.2568359375,
0.370849609375,
0.2568359375,
0.370849609375,
0.264404296875,
0.35546875,
0.264404296875
],
[
0.363525390625,
0.255126953125,
0.64892578125,
0.255126953125,
0.64892578125,
0.268310546875,
0.363525390625,
0.268310546875
],
[
0.3056640625,
0.26953125,
0.476318359375,
0.26953125,
0.476318359375,
0.28271484375,
0.3056640625,
0.28271484375
],
[
0.46728515625,
0.268310546875,
0.69921875,
0.26953125,
0.69921875,
0.2841796875,
0.46728515625,
0.28271484375
],
[
0.26708984375,
0.318115234375,
0.73876953125,
0.318115234375,
0.73876953125,
0.328857421875,
0.26708984375,
0.328857421875
],
[
0.265625,
0.330810546875,
0.736328125,
0.330810546875,
0.736328125,
0.3447265625,
0.265625,
0.3447265625
],
[
0.2646484375,
0.34521484375,
0.73876953125,
0.34521484375,
0.73876953125,
0.359130859375,
0.2646484375,
0.359130859375
],
[
0.2646484375,
0.359130859375,
0.73779296875,
0.359130859375,
0.73779296875,
0.37255859375,
0.2646484375,
0.37255859375
],
[
0.2646484375,
0.37255859375,
0.74169921875,
0.37255859375,
0.74169921875,
0.386474609375,
0.2646484375,
0.386474609375
],
[
0.2646484375,
0.386962890625,
0.740234375,
0.386962890625,
0.740234375,
0.40087890625,
0.2646484375,
0.40087890625
],
[
0.265625,
0.40087890625,
0.740234375,
0.40087890625,
0.740234375,
0.4140625,
0.265625,
0.4140625
],
[
0.2646484375,
0.4140625,
0.73876953125,
0.4140625,
0.73876953125,
0.427978515625,
0.2646484375,
0.427978515625
],
[
0.2646484375,
0.4287109375,
0.73876953125,
0.4287109375,
0.73876953125,
0.442626953125,
0.2646484375,
0.442626953125
],
[
0.26708984375,
0.443115234375,
0.73779296875,
0.443115234375,
0.73779296875,
0.453857421875,
0.26708984375,
0.453857421875
],
[
0.2646484375,
0.455810546875,
0.740234375,
0.455810546875,
0.740234375,
0.4697265625,
0.2646484375,
0.4697265625
],
[
0.265625,
0.47021484375,
0.73876953125,
0.47021484375,
0.73876953125,
0.48095703125,
0.265625,
0.48095703125
],
[
0.265625,
0.48486328125,
0.73779296875,
0.48486328125,
0.73779296875,
0.49560546875,
0.265625,
0.49560546875
],
[
0.2646484375,
0.49755859375,
0.73779296875,
0.49755859375,
0.73779296875,
0.51123046875,
0.2646484375,
0.51123046875
],
[
0.2646484375,
0.51123046875,
0.740234375,
0.51123046875,
0.740234375,
0.5244140625,
0.2646484375,
0.5244140625
],
[
0.2646484375,
0.5244140625,
0.740234375,
0.52587890625,
0.740234375,
0.5390625,
0.2646484375,
0.53857421875
],
[
0.2646484375,
0.5390625,
0.73876953125,
0.5390625,
0.73876953125,
0.55322265625,
0.2646484375,
0.55322265625
],
[
0.265625,
0.55322265625,
0.740234375,
0.55322265625,
0.740234375,
0.56640625,
0.265625,
0.56640625
],
[
0.265625,
0.5673828125,
0.740234375,
0.5673828125,
0.740234375,
0.58056640625,
0.265625,
0.58056640625
],
[
0.2646484375,
0.58056640625,
0.52783203125,
0.58056640625,
0.52783203125,
0.5947265625,
0.2646484375,
0.5947265625
],
[
0.265625,
0.60791015625,
0.765625,
0.60791015625,
0.765625,
0.6220703125,
0.265625,
0.6220703125
],
[
0.2646484375,
0.62060546875,
0.5986328125,
0.6220703125,
0.5986328125,
0.63623046875,
0.2646484375,
0.63525390625
],
[
0.21728515625,
0.6611328125,
0.381591796875,
0.6611328125,
0.381591796875,
0.6748046875,
0.21728515625,
0.6748046875
],
[
0.2197265625,
0.69140625,
0.78515625,
0.69140625,
0.78515625,
0.705078125,
0.2197265625,
0.705078125
],
[
0.2197265625,
0.70556640625,
0.78515625,
0.70556640625,
0.78515625,
0.7197265625,
0.2197265625,
0.7197265625
],
[
0.2181396484375,
0.72021484375,
0.78515625,
0.72021484375,
0.78515625,
0.73486328125,
0.2181396484375,
0.73486328125
],
[
0.22216796875,
0.7373046875,
0.78173828125,
0.7373046875,
0.78173828125,
0.74755859375,
0.22216796875,
0.74755859375
],
[
0.2320556640625,
0.75048828125,
0.78369140625,
0.75048828125,
0.78369140625,
0.7646484375,
0.2320556640625,
0.7646484375
],
[
0.2197265625,
0.7666015625,
0.78515625,
0.7666015625,
0.78515625,
0.77978515625,
0.2197265625,
0.77978515625
],
[
0.2181396484375,
0.77978515625,
0.78515625,
0.7822265625,
0.78515625,
0.79541015625,
0.2181396484375,
0.79345703125
],
[
0.2197265625,
0.796875,
0.685546875,
0.796875,
0.685546875,
0.81005859375,
0.2197265625,
0.81005859375
],
[
0.2459716796875,
0.81103515625,
0.78515625,
0.81103515625,
0.78515625,
0.82470703125,
0.2459716796875,
0.82470703125
],
[
0.2197265625,
0.8271484375,
0.78369140625,
0.8271484375,
0.78369140625,
0.84033203125,
0.2197265625,
0.84033203125
]
] |
[
"Co-evolving Graph Reasoning Network for",
"Emotion-Cause Pair Extraction",
"Bowen Xingl,2 and Ivor W. Tsang,1",
"1 Australian Artificial Intelligence Institute, University of Technology Sydney",
"2",
"Centre forFrontier AI Research, A*STAR",
"bwxing714@gmail.com",
"ivor_tsang@cfar.a-star.edu.sg",
"Abstract.Emotion-CausePairExtraction(ECPE)aimstoextractall",
"emotion clauses and their corresponding cause clauses from a document",
"Existing approaches tackle this task through multi-task learning (MTL)",
"framework in which the two subtasks provide indicative clues for ECPE.",
"However, the previous MTL framework considers only one round of multi-",
"task reasoning and ignores the reverse feedbacks from ECPE to the sub-",
"tasks. Besides, its multi-task reasoning only relies on semantics-level in-",
"teractions, which cannot capture the explicit dependencies, and both the",
"encoder sharing and multi-task hidden states concatenations can hardly",
"capturethe causalities.To solve theseissues,we first put forward a new",
"MTL framework based on Co-evolving Reasoning. It (1) models the bidi-",
"rectionalfeedbacksbetweenECPE anditssubtasks;(2)allowsthethree",
"taskstoevolvetogetherandprompteachotherrecurrently;(3)integrates",
"prediction-level interactions to capture explicit dependencies, Then we",
"propose a novel multi-task relational graph (MRG) to sufficiently exploit",
"the causal relations. Finally, we propose a Co-evolving Graph Reasoning",
"Network (CGR-Net) that implements our MTL framework and conducts",
"Co-evolving Reasoning on MRG. Experimental results show that our",
"model achieves new state-of-the-art performance, and further analysis",
"confirms the advantages of our method.",
"Keywords: Multi-Task LearningRelational Graph ReasoningEmotion",
"Cause Extraction Natural Language Processing.",
"1Introduction",
"情感原因对提取（ECPE）是自然语言处理/人工智能领域中一项新的具有挑战",
"性的任务。它旨在从原始文档中自动提取所有情感子句和相应的原因子句，这对",
"于实际应用[26]具有重要价值。考虑一个文档“【在学生]1，「的记忆中，他经",
"常为他们支付2，1的学费，这是值得尊敬和感动的3。”，第三个子表达了",
"-种情感，这种情感是由第二个子句触发的，因此这两个子句形成了情感-原因",
"对。直观地，检测表达原因和情绪的子句，即原因提取（CE）和情绪提取（EE），",
"是ECPE的两个子任务。因此，最近的模型[10,13,1,36]实现了图中所示的多",
"任务学习框架。1（a)，引人CE和EE为ECPE提供指示性线索",
"尽管以前的多任务学习（MTL）框架取得了可喜的进展，但根据我们的观",
"察，它仍然存在一些问题，这些问题阻碍了ECPE和CE/EE之间的多任务推"
] |
{
"arxiv_id": "2306.04340",
"language": "zh",
"page_idx": 0
}
|
||
2306.04340-zh-1
|
2306.04340-zh-1.jpg
|
2
Bowen & Ivor.
<FigureHere>
Fig.1. Comparison of the previous MTL framework and our MTL framework for
$ECPE.$
理。首先，以前的工作只考虑从 CE/EE 到 ECPE 的单向消息。CE/EE 的预测可能由于其不可靠的语义而不正确。在这种情况下，来自 CE/EE 的虚假信息可能会误导 ECPE。其次，以前作品中的单轮多任务推理过程是不能胜任的，考虑到由于情感和原因|12|固有的模糊性和微妙性，机器很难像人类一样理解情感、原因及其因果关系。第三，在以前的工作中，ECPE 和 CE/EE 之间的多任务推理仅通过隐式语义级交互来实现，例如共享编码器。一方面，这与人类的直觉以及 ECPE 和 CE/EE 之间的因果关系不一致，两者都基于预测或标签；另一方面，与预测信息相比，语义中传达的指示性信息是隐含的，并且相对不足。
针对上述问题，我们提出了一种新的基于协同进化推理的 MTL 框架，如图。1 (b)。首先，除了从 CE/EE 到 ECPE 的单向消息外，我们的 MTL 框架还对从 ECPE 到 CE/EE 的反向反馈进行了建模。如果 ECPE 预测正确，则传输给 CE/EE 的指示性信息可以改善它们。两个改进的子任务可以进一步反向促进 ECPE。如果 ECPE 预测不正确，则传输给 CE/EE 的有缺陷的信息可以作为他们的反馈，并使他们重新考虑为 ECPE 提供更好的信息，从而可以进一步提示 CE/EE。其次，为了实现这种良性循环，我们设计了递归多任务推理机制。这样，三个任务的知识就可以逐渐一起进化，相互提示。第三，我们建议通过引人两种预测级交互来利用预测中的相关性：预测-预测和预测-语义交互。通过这种方式，预测(估计的标签分布)传达的明确相关信息可以在我们的 MTL 框架中流动，然后促进共同进化推理。语义可以在预测的每一步得到直接的反馈，然后重新考虑改进
此外，先前作品中采用的 MTL 序列结构仅基于共享编码器和多任务隐藏状态级联，这几乎无法捕获因果关系。这促使我们寻求一种更有效的方法来充分利用 ECPE 和 CE/EE 之间的因果关系。为此，我们设计了一种新颖的多任务关系图 (MRG),其中有三组节点源自文档中的子句，分别对应于 ECPE, CE 和EE。此外，我们设计了不同的关系类型，这些关系类型对应于 ECPE 和 CE/EE 之间的因果关系
为了实现我们的 MTL 框架并在 MRG 上进行协同进化推理，我们提出了一种协同进化图推理网络(CGR-Net),其核心是多任务关系图转换(MRGT)单
|
[
[
0.21728515625,
0.11614990234375,
0.2459716796875,
0.11614990234375,
0.2459716796875,
0.1287841796875,
0.21728515625,
0.1287841796875
],
[
0.273681640625,
0.117431640625,
0.366943359375,
0.117431640625,
0.366943359375,
0.1287841796875,
0.273681640625,
0.1287841796875
],
[
0.55615234375,
0.2734375,
0.580078125,
0.2734375,
0.580078125,
0.281005859375,
0.55615234375,
0.281005859375
],
[
0.580078125,
0.291748046875,
0.5849609375,
0.291748046875,
0.5849609375,
0.296630859375,
0.580078125,
0.296630859375
],
[
0.47216796875,
0.31005859375,
0.483642578125,
0.31005859375,
0.483642578125,
0.313232421875,
0.47216796875,
0.313232421875
],
[
0.5556640625,
0.32568359375,
0.60302734375,
0.326904296875,
0.60302734375,
0.34228515625,
0.5556640625,
0.341552734375
],
[
0.2197265625,
0.35986328125,
0.78515625,
0.35986328125,
0.78515625,
0.373779296875,
0.2197265625,
0.373779296875
],
[
0.2197265625,
0.373779296875,
0.265625,
0.373779296875,
0.265625,
0.3837890625,
0.2197265625,
0.3837890625
],
[
0.2197265625,
0.416748046875,
0.78515625,
0.416748046875,
0.78515625,
0.429931640625,
0.2197265625,
0.429931640625
],
[
0.221435546875,
0.43115234375,
0.78173828125,
0.43115234375,
0.78173828125,
0.4443359375,
0.221435546875,
0.4443359375
],
[
0.221435546875,
0.4462890625,
0.78369140625,
0.4462890625,
0.78369140625,
0.460205078125,
0.221435546875,
0.460205078125
],
[
0.221435546875,
0.4609375,
0.78369140625,
0.4609375,
0.78369140625,
0.474853515625,
0.221435546875,
0.474853515625
],
[
0.221435546875,
0.477783203125,
0.78173828125,
0.477783203125,
0.78173828125,
0.488037109375,
0.221435546875,
0.488037109375
],
[
0.221435546875,
0.4912109375,
0.78369140625,
0.4912109375,
0.78369140625,
0.5048828125,
0.221435546875,
0.5048828125
],
[
0.221435546875,
0.50830078125,
0.78173828125,
0.50830078125,
0.78173828125,
0.51806640625,
0.221435546875,
0.51806640625
],
[
0.233642578125,
0.52294921875,
0.77685546875,
0.52294921875,
0.77685546875,
0.5361328125,
0.233642578125,
0.5361328125
],
[
0.2442626953125,
0.53857421875,
0.78369140625,
0.53857421875,
0.78369140625,
0.5517578125,
0.2442626953125,
0.5517578125
],
[
0.2181396484375,
0.55322265625,
0.78369140625,
0.55322265625,
0.78369140625,
0.56640625,
0.2181396484375,
0.56640625
],
[
0.2181396484375,
0.56640625,
0.78369140625,
0.5673828125,
0.78369140625,
0.58203125,
0.2181396484375,
0.58056640625
],
[
0.2197265625,
0.58349609375,
0.78369140625,
0.58349609375,
0.78369140625,
0.5966796875,
0.2197265625,
0.5966796875
],
[
0.2197265625,
0.59765625,
0.78515625,
0.59765625,
0.78515625,
0.611328125,
0.2197265625,
0.611328125
],
[
0.2197265625,
0.6123046875,
0.78173828125,
0.6123046875,
0.78173828125,
0.62548828125,
0.2197265625,
0.62548828125
],
[
0.2181396484375,
0.626953125,
0.78369140625,
0.62744140625,
0.78369140625,
0.642578125,
0.2181396484375,
0.6416015625
],
[
0.2197265625,
0.64306640625,
0.78369140625,
0.64306640625,
0.78369140625,
0.6572265625,
0.2197265625,
0.6572265625
],
[
0.221435546875,
0.6591796875,
0.78369140625,
0.6591796875,
0.78369140625,
0.67236328125,
0.221435546875,
0.67236328125
],
[
0.2181396484375,
0.673828125,
0.78369140625,
0.67236328125,
0.78369140625,
0.6875,
0.2181396484375,
0.68798828125
],
[
0.2197265625,
0.68798828125,
0.78369140625,
0.68798828125,
0.78369140625,
0.7021484375,
0.2197265625,
0.7021484375
],
[
0.2197265625,
0.7041015625,
0.334228515625,
0.7041015625,
0.334228515625,
0.71728515625,
0.2197265625,
0.71728515625
],
[
0.24267578125,
0.71826171875,
0.78369140625,
0.7197265625,
0.78369140625,
0.734375,
0.24267578125,
0.73291015625
],
[
0.2197265625,
0.73486328125,
0.78515625,
0.73486328125,
0.78515625,
0.74853515625,
0.2197265625,
0.74853515625
],
[
0.221435546875,
0.75048828125,
0.78515625,
0.75048828125,
0.78515625,
0.7646484375,
0.221435546875,
0.7646484375
],
[
0.221435546875,
0.76513671875,
0.78515625,
0.76513671875,
0.78515625,
0.77880859375,
0.221435546875,
0.77880859375
],
[
0.2197265625,
0.78076171875,
0.7861328125,
0.78076171875,
0.7861328125,
0.79443359375,
0.2197265625,
0.79443359375
],
[
0.2197265625,
0.79541015625,
0.334228515625,
0.79541015625,
0.334228515625,
0.8095703125,
0.2197265625,
0.8095703125
],
[
0.2459716796875,
0.81103515625,
0.77783203125,
0.81103515625,
0.77783203125,
0.82470703125,
0.2459716796875,
0.82470703125
],
[
0.2197265625,
0.8271484375,
0.78369140625,
0.8271484375,
0.78369140625,
0.84033203125,
0.2197265625,
0.84033203125
]
] |
[
"2",
"Bowen & Ivor",
"motior",
"福",
"E",
"Corsvning",
"Fig.1. Comparison of the previous MTL framework and our MTL framework for",
"ECPE.",
"理。首先，以前的工作只考虑从CE/EE到ECPE的单向消息。CE/EE的预测",
"可能由于其不可靠的语义而不正确。在这种情况下，来自CE/EE的虚假信息可",
"能会误导ECPE。其次，以前作品中的单轮多任务推理过程是不能胜任的，考",
"虑到由于情感和原因[12]固有的模糊性和微妙性，机器很难像人类一样理解情",
"感、原因及其因果关系。第三，在以前的工作中，ECPE和CE/EE之间的多任",
"务推理仅通过隐式语义级交互来实现，例如共享编码器。一方面，这与人类的直",
"觉以及ECPE和CE/EE之间的因果关系不一致，两者都基于预测或标签；另",
"方面，与预测信息相比，语义中传达的指示性信息是隐含的，并且相对不足。",
"针对上述问题，我们提出了一种新的基于协同进化推理的MTL框架，如图。",
"1（b)。首先，除了从CE/EE到ECPE的单向消息外，我们的MTL框架还对",
"从ECPE到CE/EE的反向反馈进行了建模。如果ECPE预测正确，则传输",
"给CE/EE的指示性信息可以改善它们。两个改进的子任务可以进一步反向促",
"进ECPE。如果ECPE预测不正确，则传输给CE/EE的有缺陷的信息可以作",
"为他们的反馈，并使他们重新考虑为ECPE提供更好的信息，从而可以进一步",
"提示CE/EE。其次，为了实现这种良性循环，我们设计了递归多任务推理机制。",
"这样，三个任务的知识就可以逐渐一起进化，相互提示。第三，我们建议通过引",
"入两种预测级交互来利用预测中的相关性：预测-预测和预测-语义交互。通过这",
"种方式，预测（估计的标签分布）传达的明确相关信息可以在我们的MTL框架",
"中流动，然后促进共同进化推理。语义可以在预测的每一步得到直接的反馈，然",
"后重新考虑改进",
"此外，先前作品中采用的MTL序列结构仅基于共享编码器和多任务隐藏状",
"态级联，这几乎无法捕获因果关系。这促使我们寻求一种更有效的方法来充分利",
"用ECPE和CE/EE之间的因果关系。为此，我们设计了一种新颖的多任务关",
"系图（MRG），其中有三组节点源自文档中的子句，分别对应于ECPE，CE和",
"EE。此外，我们设计了不同的关系类型，这些关系类型对应于ECPE和CE/EE",
"之间的因果关系",
"为了实现我们的MTL框架并在MRG上进行协同进化推理，我们提出了",
"种协同进化图推理网络（CGR-Net），其核心是多任务关系图转换（MRGT）单"
] |
{
"arxiv_id": "2306.04340",
"language": "zh",
"page_idx": 1
}
|
||
2306.04340-zh-2
|
2306.04340-zh-2.jpg
|
Co-evolving Graph Reasoning Network for Emotion-Cause Pair Extraction 3
元。CGR-Net 首先生成特定于任务的隐藏状态，并生成初始估计的标签分布。然后，MRGT 单元递归地将三个任务的隐藏状态和标签分布作为输人，然后分三个步骤更新它们：标签分布的投影，关系局部图变换和非局部自变换。最后， ECPE 在最后一步的预测被用来提取潜在的情绪-原因对。并且我们设计了一个基于逻辑约束的线束损失，迫使这三个任务在协同进化推理的良性循环中逐渐相互促进
总而言之，我们的贡献有三个方面：
(1)我们提出了一个新的基于共同进化推理和 MRG 的 MTL 框架，以充分利用相关性和因果关系。据我们所知，我们的 MTL 框架是第一个允许 ECPE 和CE/EE 递归相互促进的框架，MRG 是 ECPE 的第一个 MTL 图结构。
(2)为了实现我们的 MTL 框架并在 MRG 上进行共同进化推理，我们提出了一
种新颖的 CGR-Net, 其核心是多任务关系图转换单元。
(3)在基准数据集上的实验结果表明，我们的 CGR-Net 的性能明显优于现有的最新模型。并进一步分析证明了 CGR-Net 不同组件的有效性和 MRG 的优越性
## 2 Related Works
情感原因提取(ECE)[18,4,21,20,14,16,15,9,35]是一项长期任务，其目标是提取文档中给定情感表达的原因。但是，这就要求必须对情绪进行人工注释，这就限制了实际应用。因此，最近[26] 提出了 ECPE 任务和两步解决方案，而错误传播可能从第一步到第二步发生。为此，最近的工作提出了统一的端到端模型[10,23,2,25,11,25,6,13,5]来解决 MTL 框架中的 ECPE
[10]提出了一个模型，该模型集成了情感-原因对，交互和预测的 2D 表示。[23]从排名的角度处理 ECPE,并采用基于内核的相对位置嵌入进行排名。[37] 提出了一种标记方案，用于编码情感-原因对中情感子句和原因子句之间的距离。在此基础上，[13] 提出了一种标签分布细化方法，该方法根据预定义的规则使用CE 和 EE 中的标签来调整 ECPE 的输出标签分布。但是，细化方法不参与模型训练，只对评价的输出进行工作
最近，多粒度语义感知图模型(MGSAG) [1]将细粒度和粗粒度语义特征结合在一起，旨在解决子句语义的距离限制。而基于矩阵胶囊的多粒度框架 (MaCa) [36]引人了矩阵胶囊来获取更细粒度的子句对特征，聚类每个子句对的关系
与以前的工作不同，我们(1)提出了共同进化推理，它允许三个任务逐渐且充分地相互促进；(2)引人预测-预测和预测-语义交互来对显式相关性进行建模， 并为可以重新考虑以改进的语义提供反馈；(3)通过设计新颖的 MRG 来有效利用偶然关系
# 3 Methodology
在深入研究 MRG 和 CGR-Net 之前，我们首先介绍工作中的任务制定。我们将ECPE 转换为标签分类任务，并使用以原因为中心的标记方案 [37]。每个子句$x_{i}$都有一个两元组标记$y_i^t=(y_i^{t,c},y_i^{t,d})\in\mathcal{C}_t$,其中$y_i^t,c\in\{\mathbb{C},\mathbb{O}\}$表示$x_i$是否是原因子句，$y_i^{t,d}\in\{-\gamma,...,-1,0,1,...,\gamma,\bot\}$表示$x_i$与其触发的情感子句之间的距离，而$^,\perp^{\prime}$始终与”O’相关联，表示 $x_i$ 是非原因子句。$\gamma$ 是一个控制情感-原因对最大跨度的超参数。因此，ECPE (标签)完全具有$|\mathcal{C}_t|=2(\gamma+1)$类。至于
|
[
[
0.233642578125,
0.11614990234375,
0.73046875,
0.11614990234375,
0.73046875,
0.1300048828125,
0.233642578125,
0.1300048828125
],
[
0.73046875,
0.11932373046875,
0.77294921875,
0.11932373046875,
0.77294921875,
0.126953125,
0.73046875,
0.126953125
],
[
0.771484375,
0.11871337890625,
0.7861328125,
0.11871337890625,
0.7861328125,
0.1275634765625,
0.771484375,
0.1275634765625
],
[
0.2197265625,
0.1483154296875,
0.78515625,
0.1483154296875,
0.78515625,
0.1622314453125,
0.2197265625,
0.1622314453125
],
[
0.2181396484375,
0.1622314453125,
0.78369140625,
0.1634521484375,
0.78369140625,
0.177978515625,
0.2181396484375,
0.1767578125
],
[
0.221435546875,
0.1767578125,
0.78369140625,
0.1787109375,
0.78369140625,
0.1925048828125,
0.221435546875,
0.1900634765625
],
[
0.2197265625,
0.1932373046875,
0.78369140625,
0.1932373046875,
0.78369140625,
0.20703125,
0.2197265625,
0.20703125
],
[
0.2197265625,
0.208984375,
0.78515625,
0.208984375,
0.78515625,
0.222900390625,
0.2197265625,
0.222900390625
],
[
0.221435546875,
0.2259521484375,
0.285888671875,
0.2259521484375,
0.285888671875,
0.236083984375,
0.221435546875,
0.236083984375
],
[
0.2442626953125,
0.2392578125,
0.495849609375,
0.2392578125,
0.495849609375,
0.25244140625,
0.2442626953125,
0.25244140625
],
[
0.2197265625,
0.25390625,
0.78515625,
0.25390625,
0.78515625,
0.267578125,
0.2197265625,
0.267578125
],
[
0.22216796875,
0.270751953125,
0.78369140625,
0.270751953125,
0.78369140625,
0.281005859375,
0.22216796875,
0.281005859375
],
[
0.2197265625,
0.2841796875,
0.7236328125,
0.2841796875,
0.7236328125,
0.29736328125,
0.2197265625,
0.29736328125
],
[
0.2197265625,
0.2998046875,
0.78515625,
0.2998046875,
0.78515625,
0.313232421875,
0.2197265625,
0.313232421875
],
[
0.2197265625,
0.314453125,
0.60888671875,
0.314453125,
0.60888671875,
0.32763671875,
0.2197265625,
0.32763671875
],
[
0.2181396484375,
0.328857421875,
0.78173828125,
0.328857421875,
0.78173828125,
0.34228515625,
0.2181396484375,
0.34228515625
],
[
0.2197265625,
0.34521484375,
0.78369140625,
0.34521484375,
0.78369140625,
0.359130859375,
0.2197265625,
0.359130859375
],
[
0.21728515625,
0.385009765625,
0.40185546875,
0.385009765625,
0.40185546875,
0.399658203125,
0.21728515625,
0.399658203125
],
[
0.2197265625,
0.417236328125,
0.78515625,
0.417236328125,
0.78515625,
0.43115234375,
0.2197265625,
0.43115234375
],
[
0.2197265625,
0.43310546875,
0.78369140625,
0.43310546875,
0.78369140625,
0.4462890625,
0.2197265625,
0.4462890625
],
[
0.221435546875,
0.447509765625,
0.78515625,
0.447509765625,
0.78515625,
0.4609375,
0.221435546875,
0.4609375
],
[
0.221435546875,
0.46337890625,
0.78515625,
0.46337890625,
0.78515625,
0.4765625,
0.221435546875,
0.4765625
],
[
0.2181396484375,
0.477783203125,
0.650390625,
0.477783203125,
0.650390625,
0.4912109375,
0.2181396484375,
0.4912109375
],
[
0.2459716796875,
0.49365234375,
0.78369140625,
0.49365234375,
0.78369140625,
0.5068359375,
0.2459716796875,
0.5068359375
],
[
0.22216796875,
0.51025390625,
0.78369140625,
0.51025390625,
0.78369140625,
0.52099609375,
0.22216796875,
0.52099609375
],
[
0.221435546875,
0.52587890625,
0.78173828125,
0.52587890625,
0.78173828125,
0.5361328125,
0.221435546875,
0.5361328125
],
[
0.2197265625,
0.5390625,
0.78369140625,
0.5390625,
0.78369140625,
0.55322265625,
0.2197265625,
0.55322265625
],
[
0.2197265625,
0.5537109375,
0.78515625,
0.5537109375,
0.78515625,
0.5673828125,
0.2197265625,
0.5673828125
],
[
0.221435546875,
0.57080078125,
0.461669921875,
0.57080078125,
0.461669921875,
0.58056640625,
0.221435546875,
0.58056640625
],
[
0.2442626953125,
0.583984375,
0.78369140625,
0.583984375,
0.78369140625,
0.5986328125,
0.2442626953125,
0.5986328125
],
[
0.2197265625,
0.599609375,
0.78369140625,
0.599609375,
0.78369140625,
0.61279296875,
0.2197265625,
0.61279296875
],
[
0.221435546875,
0.6162109375,
0.75341796875,
0.6162109375,
0.75341796875,
0.626953125,
0.221435546875,
0.626953125
],
[
0.2459716796875,
0.6298828125,
0.78369140625,
0.6298828125,
0.78369140625,
0.64306640625,
0.2459716796875,
0.64306640625
],
[
0.22216796875,
0.646484375,
0.78173828125,
0.646484375,
0.78173828125,
0.6572265625,
0.22216796875,
0.6572265625
],
[
0.2197265625,
0.66015625,
0.78515625,
0.66015625,
0.78515625,
0.673828125,
0.2197265625,
0.673828125
],
[
0.221435546875,
0.6767578125,
0.316162109375,
0.6767578125,
0.316162109375,
0.68798828125,
0.221435546875,
0.68798828125
],
[
0.214111328125,
0.7109375,
0.3857421875,
0.7138671875,
0.3857421875,
0.73193359375,
0.214111328125,
0.72900390625
],
[
0.2197265625,
0.74755859375,
0.78369140625,
0.74755859375,
0.78369140625,
0.76123046875,
0.2197265625,
0.76123046875
],
[
0.21728515625,
0.76220703125,
0.78369140625,
0.76220703125,
0.78369140625,
0.77587890625,
0.21728515625,
0.77587890625
],
[
0.2181396484375,
0.77978515625,
0.78369140625,
0.77978515625,
0.78369140625,
0.79345703125,
0.2181396484375,
0.79345703125
],
[
0.2181396484375,
0.79345703125,
0.78515625,
0.79443359375,
0.78515625,
0.8125,
0.2181396484375,
0.81103515625
],
[
0.2197265625,
0.81103515625,
0.78369140625,
0.81103515625,
0.78369140625,
0.82568359375,
0.2197265625,
0.82568359375
],
[
0.2197265625,
0.8271484375,
0.78369140625,
0.8271484375,
0.78369140625,
0.84033203125,
0.2197265625,
0.84033203125
]
] |
[
"Co-evolving Graph Reasoning Network for Emotion-Cause Pair Extraction",
"",
"3",
"元。CGR-Net首先生成特定于任务的隐藏状态，并生成初始估计的标签分布。",
"然后，MRGT单元递归地将三个任务的隐藏状态和标签分布作为输人，然后分",
"三个步骤更新它们：标签分布的投影，关系局部图变换和非局部自变换。最后",
"ECPE在最后一步的预测被用来提取潜在的情绪-原因对。并且我们设计了一个",
"基于逻辑约束的线束损失，迫使这三个任务在协同进化推理的良性循环中逐渐",
"相互促进",
"总而言之，我们的贡献有三个方面：",
"1）我们提出了一个新的基于共同进化推理和MRG的MTL框架，以充分利",
"用相关性和因果关系。据我们所知，我们的MTL框架是第一个充许ECPE和",
"CE/EE递归相互促进的框架，MRG是ECPE的第一个MTL图结构。",
"（2）为了实现我们的MTL框架并在MRG上进行共同进化推理，我们提出了一",
"种新颖的CGR-Net，其核心是多任务关系图转换单元。",
"（3）在基准数据集上的实验结果表明，我们的CGR-Net的性能明显优于现有的",
"最新模型。并进一步分析证明了CGR-Net不同组件的有效性和MRG的优越性",
"2 Related Works",
"情感原因提取（ECE）[18,4,21,20,14,16,15,9,35]是一项长期任务，其目标是",
"提取文档中给定情感表达的原因。但是，这就要求必须对情绪进行人工注释，这",
"就限制了实际应用。因此，最近[26]提出了ECPE任务和两步解决方案，而错",
"误传播可能从第一步到第二步发生。为此，最近的工作提出了统一的端到端模",
"型[10,23,2,25,11,25,6,13,5]来解决MTL框架中的ECPE",
"[10]提出了一个模型，该模型集成了情感-原因对，交互和预测的2D表示。",
"23从排名的角度处理ECPE，并采用基于内核的相对位置嵌入进行排名。37",
"提出厂一种标记方案，用于缩编码情感-原因对中情感子句和原因子句之间的距离",
"在此基础上，[13]提出了一种标签分布细化方法，该方法根据预定义的规则使用",
"CE和EE中的标签来调整ECPE的输出标签分布。但是，细化方法不参与模",
"型训练，只对评价的输出进行工作",
"最近，多粒度语义感知图模型（MGSAG[1]将细粒度和粗粒度语义特征结合",
"在一起，旨在解决子句语义的距离限制。而基于矩阵胶囊的多粒度框架（MaCa）",
"36引人了矩阵胶囊来获取更细粒度的子句对特征，聚类每个子句对的关系",
"与以前的工作不同，我们（1）提出了共同进化推理，它允许三个任务逐渐且",
"充分地相互促进；（2）引入预测-预测和预测-语义交互来对显式相关性进行建模",
"并为可以重新考虑以改进的语义提供反馈；（3）通过设计新颖的MRG来有效",
"利用偶然关系",
"3Methodology",
"在深人研究MRG和CGR-Net之前，我们首先介绍工作中的任务制定。我们将",
"ECPE转换为标签分类任务，并使用以原因为中心的标记方案[37]。每个子句",
"ri都有一个两元组标记yt=（ytyta)ECt，其中ytce[C,O)表示ai是否是",
"原因子句，{--1,0,1…,}表示与其触发的情感子句之间的",
"距离，而”工始终与'O”相关联，表示i是非原因子句。是一个控制情感-原",
"因对最大跨度的超参数。因此，ECPE（标签）完全具有Ct=2(+1)类。至于"
] |
{
"arxiv_id": "2306.04340",
"language": "zh",
"page_idx": 2
}
|
||
2306.04340-zh-3
|
2306.04340-zh-3.jpg
|
4 Bowen $\&$ Ivor.
Table 1. Relation types in MRG, w.l.o.g. $\gamma = 2. I_t( i)$ indicates node $i$ is a cause (c)
node, tag (t) node or emotion (e) node.“, denotes the set of [-2,-1,0,1,2].
<table>
<tbody>
<tr>
<th>$\sum_{k=1}^{\infty}$</th>
<th>$పి$</th>
<th>出</th>
<th>8</th>
<th>花</th>
<th>$岀$</th>
<th> -</th>
<th>$\overrightarrow{\ddot{\varrho}}$ </th>
<th>$\underline{\frac{2}{2}}$</th>
<th>$\Overrightarrow{\varrho}$</th>
<th>$\stackrel{\gamma}{\textcircled{2}}$</th>
<th>$\begin{array}{c}{\underrightarrow{\Lambda}}\\{\stackrel{1}{\ddot{0}}}\\\end{array}$</th>
<th>$\begin{array}{c}{\overrightarrow{5}}\\{\overrightarrow{5}}\\{\overrightarrow{5}}\end{array}$</th>
<th>$\sum_{i=1}^{\infty}\sum_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=2}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}i_{i=1}^{\infty}$</th>
<th>$\Overrightarrow{\ddot{\varpi}}$</th>
<th>$\sum_{i=1}^{\infty}\frac{1}{2i}$</th>
</tr>
<tr>
<td>$I_t( i)$</td>
<td> C</td>
<td>t</td>
<td>e</td>
<td>C</td>
<td>t</td>
<td>t</td>
<td>t</td>
<td>t</td>
<td>t</td>
<td>t</td>
<td>e</td>
<td>e</td>
<td>e</td>
<td>e</td>
<td>e</td>
</tr>
<tr>
<td>$I_t$ 7</td>
<td>C</td>
<td>t</td>
<td>e</td>
<td>t</td>
<td>C</td>
<td>e</td>
<td>e</td>
<td>e</td>
<td>e</td>
<td>e</td>
<td>t</td>
<td>t</td>
<td>t</td>
<td>t</td>
<td>e</td>
</tr>
<tr>
<td>rdisl $T$</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>0</td>
<td>0</td>
<td>-2</td>
<td>-1</td>
<td>0</td>
<td>1</td>
<td>2</td>
<td>-2 </td>
<td>-1</td>
<td>0</td>
<td>1</td>
<td>2</td>
</tr>
</tbody>
</table>
CE(原因)和 EE(情感),它们都被公式化为二进制分类任务：$y_i^c\in\mathcal{C}_c=\{1,0\}$
和$y_{i}^{e}\in \mathcal{C} _{e}= \{ 1, 0\}$ 。
3.1 Constructing a MRG from a Document
在本文中，我们设计了一个多任务关系图(MRG) $\mathcal{G}=(\mathcal{V},\mathcal{E},\mathcal{R})$ ,以通过对三个任务(原因，标签和情感)的自我和相互交互进行建模来利用因果关系。文档$\mathcal{D}$中的每个子句$x_i$分别导出三个节点$c_i,t_i$和$e_i$,分别用于原因，标签和情感，因此$|\mathcal{V}|=3n$ 。边$(i,j,r_{ij})\in\mathcal{E}$表示从节点$i$到节点$j$的信息传播，$r_ij\in\mathcal{R}$ 是边的关系类型。请注意，节点$i$和节点$j$可能对应于不同的任务。我们定义了三种规则来确定 MRG 中两个节点之间的连接：
方向：$(j,i,r_{ji})\in\mathcal{E}$如果$(i,j,r_{ij})\in\mathcal{E}$ 。在 MRG 中，两个节点之间的信息传播是双向的。这保证了标签和原因/情绪之间的双向相关性。本地连接： $\forall(i,j,r_{ij}),|$rdis$(i,j)|\leq\gamma$ ,其中 rdis$(i,j)$表示$\mathcal{D}$中节点$i$和节点$j$的子句之间的相对距离。一般来说，关于语篇|7|的衔接和连贯，两个遥远的从句具有因果关系的概率相对较小。因此，MRG 中的边基于本地连接，在这项工作中，我们约束$\mathcal{D}$中两个连接节点的子句的相对距离范围为$-\gamma$到$\gamma$ ,与 ECPE 标签的跨度范围一致
因果关系的关系类型：表 1 列出了 MRG 中的关系类型。MRG 的一个例子如图所示。2。为了捕获自任务局部上下文依赖关系，我们定义了$r_{ij}= cc$ , $r_{ij}=tt$和$r_{ij}=ee$分别对原因，标签和情感的局部自转换进行建模。至于任务间的交互，首先，关于标签任务的方案，原因和标签任务之间有两个明确的因果关系：(1)如果$y_i^c=1$ , 那么$y_i^t,c=$C ,反之亦然；(2) 如果$y_i^c=0$ ,那么$y_i^t,c=$O ,反之亦然。为了在 MRG 中对这些因果关系进行建模，我们定义了$r_{ij}=ct$ 和$r_{ij}=tc$,以实现原因和标签之间的相互转换。此外，标签与情感任务之间存在四种因果关系。首先，如果$y_i^{t,c}=\mathbb{C}$,则$x_i-\gamma\sim x_{i+\gamma}$中至少有一个情感子$\begin{array}{c}{\text{句：}y_{i-\gamma}^{e}=1\mathrm{~or~}y_{i-\gamma+1}^{e}=1\mathrm{~or~...~or~}y_{i+\gamma-1}^{e}=1\mathrm{~or~}y_{i+\gamma}^{e}=1\text{ 。其次，如果}}\\{x_{i-\gamma}\sim x_{i+\gamma}\text{ 中没有情感子句：}y_{i-\gamma}^{e}=0\mathrm{~and~}y_{i-\gamma+1}^{e}=0\mathrm{~and~}\ldots\mathrm{and~}y_{i+\gamma-1}^{e}=}\end{array}$ 0 and $y_i+\gamma^e=0$ ,则$y_i^{t,c}=$O 和$y_i^{t,d}=\perp$。第三，如果$y_i^{t,d}=m$ ,则$y_i+m^e=1$ 。相反，$y_i+m^e=1$不能推导出$y_i^{t,d}=m$ ,但直观上如果$p(y_i+m^e=1)$增加， $p(y_{i}^{t,d}=m)$ 也应该增加。第四，如果$y_i+m^e=0$,则$y_i^{t,d}\neq m$。相反，尽管$y_i^{t,d}\neq m$不能推导$y_i+m^e=0$ ,但直观上如果$p(y_i^{t,d}=m)$减小，则$p(y_i+m^e=0)$ 应该增加。为了在 MRG 中对这些tag - emotion因果关系进行建模，我们定义了一组由 $r_ij= I_t( i) I_t( j) :$ rdis$(ij)$ 表示的关系，一些实例如图。2。例如， $r_{ij}=et:-2$表示从节点$i$ (情感节点)到节点$j$ (标签节点)的关系以及节点$i$与节点$j$之间的相对距离
|
[
[
0.2156982421875,
0.11614990234375,
0.2484130859375,
0.11614990234375,
0.2484130859375,
0.1287841796875,
0.2156982421875,
0.1287841796875
],
[
0.27294921875,
0.117431640625,
0.366943359375,
0.117431640625,
0.366943359375,
0.1287841796875,
0.27294921875,
0.1287841796875
],
[
0.2181396484375,
0.14453125,
0.78369140625,
0.14453125,
0.78369140625,
0.1590576171875,
0.2181396484375,
0.1590576171875
],
[
0.2181396484375,
0.1590576171875,
0.7197265625,
0.1590576171875,
0.7197265625,
0.17236328125,
0.2181396484375,
0.17236328125
],
[
0.28759765625,
0.1932373046875,
0.2998046875,
0.1932373046875,
0.2998046875,
0.20703125,
0.28759765625,
0.20703125
],
[
0.47216796875,
0.1944580078125,
0.47802734375,
0.1944580078125,
0.47802734375,
0.199462890625,
0.47216796875,
0.199462890625
],
[
0.5048828125,
0.1925048828125,
0.51708984375,
0.2020263671875,
0.5048828125,
0.2108154296875,
0.49267578125,
0.2020263671875
],
[
0.52294921875,
0.1944580078125,
0.55810546875,
0.1944580078125,
0.55810546875,
0.210205078125,
0.52294921875,
0.210205078125
],
[
0.5498046875,
0.1962890625,
0.59228515625,
0.1944580078125,
0.59326171875,
0.20703125,
0.55126953125,
0.208984375
],
[
0.66650390625,
0.1962890625,
0.7060546875,
0.1962890625,
0.7060546875,
0.20703125,
0.66650390625,
0.20703125
],
[
0.5146484375,
0.1988525390625,
0.521484375,
0.1988525390625,
0.521484375,
0.203857421875,
0.5146484375,
0.203857421875
],
[
0.34716796875,
0.1956787109375,
0.384033203125,
0.2020263671875,
0.3798828125,
0.21337890625,
0.343017578125,
0.20703125
],
[
0.47216796875,
0.199462890625,
0.47802734375,
0.199462890625,
0.47802734375,
0.2083740234375,
0.47216796875,
0.2083740234375
],
[
0.705078125,
0.199462890625,
0.71728515625,
0.199462890625,
0.71728515625,
0.203857421875,
0.705078125,
0.203857421875
],
[
0.3857421875,
0.2020263671875,
0.39794921875,
0.2020263671875,
0.39794921875,
0.210205078125,
0.3857421875,
0.210205078125
],
[
0.3349609375,
0.2008056640625,
0.348876953125,
0.2008056640625,
0.348876953125,
0.210205078125,
0.3349609375,
0.210205078125
],
[
0.416748046875,
0.203857421875,
0.423095703125,
0.203857421875,
0.423095703125,
0.208984375,
0.416748046875,
0.208984375
],
[
0.443603515625,
0.203857421875,
0.44775390625,
0.203857421875,
0.44775390625,
0.208984375,
0.443603515625,
0.208984375
],
[
0.281005859375,
0.2183837890625,
0.3095703125,
0.2183837890625,
0.3095703125,
0.23291015625,
0.281005859375,
0.23291015625
],
[
0.441162109375,
0.2235107421875,
0.44775390625,
0.2235107421875,
0.44775390625,
0.2291259765625,
0.441162109375,
0.2291259765625
],
[
0.471435546875,
0.2235107421875,
0.47802734375,
0.2235107421875,
0.47802734375,
0.2291259765625,
0.471435546875,
0.2291259765625
],
[
0.5,
0.2235107421875,
0.50634765625,
0.2235107421875,
0.50634765625,
0.2291259765625,
0.5,
0.2291259765625
],
[
0.52783203125,
0.2235107421875,
0.5361328125,
0.2235107421875,
0.5361328125,
0.2291259765625,
0.52783203125,
0.2291259765625
],
[
0.27783203125,
0.2298583984375,
0.313720703125,
0.2298583984375,
0.313720703125,
0.242431640625,
0.27783203125,
0.242431640625
],
[
0.35546875,
0.231689453125,
0.40283203125,
0.231689453125,
0.40283203125,
0.242431640625,
0.35546875,
0.242431640625
],
[
0.404296875,
0.23291015625,
0.423095703125,
0.23291015625,
0.423095703125,
0.240478515625,
0.404296875,
0.240478515625
],
[
0.439453125,
0.2310791015625,
0.451904296875,
0.236083984375,
0.44775390625,
0.242431640625,
0.435546875,
0.2374267578125
],
[
0.471435546875,
0.23486328125,
0.47802734375,
0.23486328125,
0.47802734375,
0.240478515625,
0.471435546875,
0.240478515625
],
[
0.498291015625,
0.23486328125,
0.50732421875,
0.23486328125,
0.50732421875,
0.240478515625,
0.498291015625,
0.240478515625
],
[
0.265625,
0.24365234375,
0.32275390625,
0.24365234375,
0.32275390625,
0.25390625,
0.265625,
0.25390625
],
[
0.441162109375,
0.2442626953125,
0.44775390625,
0.2442626953125,
0.44775390625,
0.25,
0.441162109375,
0.25
],
[
0.4697265625,
0.2442626953125,
0.47802734375,
0.2442626953125,
0.47802734375,
0.25,
0.4697265625,
0.25
],
[
0.5,
0.24560546875,
0.50634765625,
0.24560546875,
0.50634765625,
0.248779296875,
0.5,
0.248779296875
],
[
0.21728515625,
0.285400390625,
0.78173828125,
0.287353515625,
0.78173828125,
0.3017578125,
0.21728515625,
0.2998046875
],
[
0.4150390625,
0.24560546875,
0.423095703125,
0.24560546875,
0.423095703125,
0.250732421875,
0.4150390625,
0.250732421875
],
[
0.2197265625,
0.302978515625,
0.35693359375,
0.302978515625,
0.35693359375,
0.31640625,
0.2197265625,
0.31640625
],
[
0.21728515625,
0.33837890625,
0.59228515625,
0.33837890625,
0.59228515625,
0.3515625,
0.21728515625,
0.3515625
],
[
0.2197265625,
0.3623046875,
0.78369140625,
0.3623046875,
0.78369140625,
0.375732421875,
0.2197265625,
0.375732421875
],
[
0.2197265625,
0.376953125,
0.78369140625,
0.376953125,
0.78369140625,
0.39013671875,
0.2197265625,
0.39013671875
],
[
0.21728515625,
0.39013671875,
0.78369140625,
0.391357421875,
0.78369140625,
0.406005859375,
0.21728515625,
0.40478515625
],
[
0.2181396484375,
0.4072265625,
0.78369140625,
0.4072265625,
0.78369140625,
0.421630859375,
0.2181396484375,
0.421630859375
],
[
0.2197265625,
0.421630859375,
0.78369140625,
0.421630859375,
0.78369140625,
0.43505859375,
0.2197265625,
0.43505859375
],
[
0.221435546875,
0.4375,
0.5537109375,
0.4375,
0.5537109375,
0.45068359375,
0.221435546875,
0.45068359375
],
[
0.2442626953125,
0.451904296875,
0.78515625,
0.45068359375,
0.78515625,
0.46533203125,
0.2442626953125,
0.466552734375
],
[
0.2197265625,
0.466552734375,
0.78515625,
0.468505859375,
0.78515625,
0.481689453125,
0.2197265625,
0.479736328125
],
[
0.221435546875,
0.48291015625,
0.78369140625,
0.48291015625,
0.78369140625,
0.496826171875,
0.221435546875,
0.496826171875
],
[
0.221435546875,
0.49755859375,
0.78369140625,
0.49755859375,
0.78369140625,
0.51123046875,
0.221435546875,
0.51123046875
],
[
0.2197265625,
0.51123046875,
0.78369140625,
0.51220703125,
0.78369140625,
0.5263671875,
0.2197265625,
0.52587890625
],
[
0.2197265625,
0.52783203125,
0.78369140625,
0.52783203125,
0.78369140625,
0.54248046875,
0.2197265625,
0.54248046875
],
[
0.2197265625,
0.54443359375,
0.31787109375,
0.54443359375,
0.31787109375,
0.55615234375,
0.2197265625,
0.55615234375
],
[
0.2442626953125,
0.55810546875,
0.78515625,
0.55810546875,
0.78515625,
0.5712890625,
0.2442626953125,
0.5712890625
],
[
0.2181396484375,
0.5712890625,
0.775390625,
0.57373046875,
0.775390625,
0.5869140625,
0.2181396484375,
0.58544921875
],
[
0.2197265625,
0.58837890625,
0.78369140625,
0.58837890625,
0.78369140625,
0.6015625,
0.2197265625,
0.6015625
],
[
0.221435546875,
0.60302734375,
0.78369140625,
0.60302734375,
0.78369140625,
0.6162109375,
0.221435546875,
0.6162109375
],
[
0.2197265625,
0.6171875,
0.78173828125,
0.6171875,
0.78173828125,
0.630859375,
0.2197265625,
0.630859375
],
[
0.233642578125,
0.63330078125,
0.78369140625,
0.6337890625,
0.78369140625,
0.64794921875,
0.233642578125,
0.646484375
],
[
0.2197265625,
0.6484375,
0.78369140625,
0.6484375,
0.78369140625,
0.662109375,
0.2197265625,
0.662109375
],
[
0.221435546875,
0.6640625,
0.78369140625,
0.6640625,
0.78369140625,
0.6787109375,
0.221435546875,
0.6787109375
],
[
0.2181396484375,
0.6787109375,
0.38232421875,
0.6787109375,
0.38232421875,
0.6923828125,
0.2181396484375,
0.6923828125
],
[
0.396240234375,
0.6787109375,
0.78369140625,
0.6787109375,
0.78369140625,
0.693359375,
0.396240234375,
0.693359375
],
[
0.2156982421875,
0.689453125,
0.7861328125,
0.69140625,
0.7861328125,
0.7119140625,
0.2156982421875,
0.71044921875
],
[
0.214111328125,
0.70703125,
0.7890625,
0.70849609375,
0.7890625,
0.72998046875,
0.214111328125,
0.72900390625
],
[
0.21728515625,
0.7265625,
0.7900390625,
0.7265625,
0.7900390625,
0.74755859375,
0.21728515625,
0.74755859375
],
[
0.2197265625,
0.74755859375,
0.78369140625,
0.74755859375,
0.78369140625,
0.76220703125,
0.2197265625,
0.76220703125
],
[
0.2181396484375,
0.7646484375,
0.7861328125,
0.7666015625,
0.7861328125,
0.78076171875,
0.2181396484375,
0.77880859375
],
[
0.221435546875,
0.77978515625,
0.78515625,
0.77978515625,
0.78515625,
0.79443359375,
0.221435546875,
0.79443359375
],
[
0.221435546875,
0.796875,
0.78515625,
0.796875,
0.78515625,
0.81005859375,
0.221435546875,
0.81005859375
],
[
0.2181396484375,
0.81005859375,
0.78515625,
0.80810546875,
0.78515625,
0.82568359375,
0.2181396484375,
0.82763671875
],
[
0.2181396484375,
0.8271484375,
0.41259765625,
0.8271484375,
0.41259765625,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"4",
"Bowen& Ivor",
"Table 1. Relation types in MRG, w.l.o.g. = 2 . It(i) indicates node i is a cause (c)",
"node, tag (t) node or emotion (e) node. \" denotes the set of [-2, -1, O, 1, 2]",
"D",
"1",
"o",
"",
"80",
"批地",
"",
"花",
"",
"",
"8",
"8",
"T",
"",
"It(℃)",
"T",
"t",
"t",
"t",
"It(G)",
"te",
"t",
"C",
"e",
"e",
"rdis(i,j)",
"(",
"2",
"",
"CE（原因）和EE（情感），它们都被公式化为二进制分类任务：yEC={1,O)",
"U",
"和 yiECe={1,0]。",
"3.1 Constructing a MRG from a Document",
"在本文中，我们设计了一个多任务关系图（MRG）9=（V,8,R），以通过对三个",
"任务（原因，标签和情感）的自我和相互交互进行建模来利用因果关系。。文档",
"D中的每个子句ai分别导出三个节点ci，t和ei，分别用于原因，标签和情",
"感，因此V=3n。边（i,j,ri)E表示从节点i到节点j的信息传播，TiiER",
"是边的关系类型。请注意，节点讠和节点可能对应于不同的任务。我们定义了",
"三种规则来确定MRG中两个节点之间的连接：",
"方向：（i,i,rj）E&如果（i,j,ri）E&。在MRG中，两个节点之间的信",
"息传播是双向的。这保证了标签和原因/情绪之间的双向相关性。本地连接",
"V(i,j,ri),rdisi,j)≤，其中rdis(i,j）表示D中节点i和节点j的子句之间",
"的相对距离。一般来说，关于语篇[7]的衔接和连贯，两个遥远的从句具有因果",
"关系的概率相对较小。因此，MRG中的边基于本地连接，在这项工作中，我们",
"约束D中两个连接节点的子句的相对距离范围为-到，与ECPE标签的",
"跨度范围一致",
"因果关系的关系类型：表1列出了MRG中的关系类型。MRG的一个例子",
"如图所示。2。为了捕获自任务局部上下文依赖关系，我们定义了ri=cc",
"ri=t和ri=ee分别对原因，标签和情感的局部自转换进行建模。至于任务",
"间的交互，首先，关于标签任务的方案，原因和标签任务之间有两个明确的因果",
"关系：(1)如果y=1，那么ytc=C，反之亦然；（2)如果y=0，那么yc=C",
"反之亦然。为了在MRG中对这些因果关系进行建模，我们定义了rij=ct",
"和ri=tc，以实现原因和标签之间的相互转换。此外，标签与情感任务之间存",
"在四种因果关系。首先，如果yi=C，则ci-～i+中至少有一个情感子",
"句：y-=1ory-%",
"=1orory+-1=1ory+=1。其次，如果",
"i-~i+中没有情感子句：yi~=0and yi-~+1=0and...and yi+-1=",
"0and y+=0，则yc=0和 ytid=l。第三，如果yid=m，则yi+m=1",
"。相反，yi+m=1不能推导出yid=m，但直观上如果p(yi+m=1)增加，",
"p(yt,d=m）也应该增加。第四，如果y+m=O，则ytd≠m。相反，尽管",
"ytd≠m不能推导yi+m=0，但直观上如果p(ytd=m）减小，则p(yi+m=0)",
"应该增加。为了在MRG中对这些tag-emotion因果关系进行建模，我们定",
"义了一组由rii=It(i)It(i）：rdis(ii)表示的关系，一些实例如图。2。例如",
"Ti=et：一2表示从节点（情感节点）到节点i（标签节点）的关系以及节点",
"与节点j之间的相对距离"
] |
{
"arxiv_id": "2306.04340",
"language": "zh",
"page_idx": 3
}
|
||
2306.04340-zh-4
|
2306.04340-zh-4.jpg
|
5
Co-evolving Graph Reasoning Network for Emotion-Cause Pair Extraction
<FigureHere>
$\mathbf{Fig.2.~An~example~of~MRG~(~\gamma=2~).~W.L.O.G,~only~the~edges~directed~into~}c_{i},t_{i}~$and
$e_i$ are illustrated. And self-loops are not shown for simplification.
在 MRG 中，每个任务间关系都对应一个细粒度的相对距离，与标记方案的定义一致。因此，沿着这些关系的任务间图转换可以实现更充分和明确的多任务推理
# 3.2 CGR-Net
我们的协同进化图推理网络(CGR-Net)的整体架构如图。3(1)。它由三个部分组成：分层编码，初始估计和共同进化推理。接下来，我们描述这三个组件的过程。
# Hierarchical Encoding
Word-level Clause Encoding 子句编码的目的是为每个子句生成包含单词级依赖关系的表示。在前面的工作之后，每个子句都被输入到 BERT[8]编码器中， 然后将[CLS] 令牌的最后一个隐藏状态作为子句表示。现在我们获得$\mathcal{D}$的子句表示序列：$H=(h_0,...,h_n)$。
Multi-$task\textit{ Clause- level Document Encoding 在 本 文 中 , 我 们 利 用 BiLSTM [ 17] }$ 通过对子句间依赖关系进行建模来生成上下文敏感的子句隐藏状态。为了获得三个任务的特定于任务的子句隐藏状态，我们分别在$H$上应用三个 BiLSTM 来分别获得原因，标签和情感的初始子句隐藏状态：$H_c^0=\left(h_{c,1}^0,\ldots,h_{c,n}^0\right)$, $H_{t}^{0}=\left(h_{t,1}^{0},...,h_{t,n}^{0}\right)$和$H_e^0=\left(h_{e,1}^{0},...,h_{e,n}^{0}\right)$
Initial Estimation 由于 MRGT 单元将上一步中预测的三个任务的标签分布作为输入，因此$H_c^0,H_t^0$和$H_e^0$分别被馈送到原因解码器，标签解码器和情
|
[
[
0.2344970703125,
0.117431640625,
0.72802734375,
0.117431640625,
0.72802734375,
0.1275634765625,
0.2344970703125,
0.1275634765625
],
[
0.77392578125,
0.117431640625,
0.7861328125,
0.117431640625,
0.7861328125,
0.1275634765625,
0.77392578125,
0.1275634765625
],
[
0.351318359375,
0.149658203125,
0.35546875,
0.1514892578125,
0.351318359375,
0.1558837890625,
0.34716796875,
0.154052734375
],
[
0.58837890625,
0.179931640625,
0.61669921875,
0.179931640625,
0.61669921875,
0.1944580078125,
0.58837890625,
0.1944580078125
],
[
0.496826171875,
0.2083740234375,
0.5146484375,
0.2108154296875,
0.51123046875,
0.2171630859375,
0.494384765625,
0.2139892578125
],
[
0.6044921875,
0.208984375,
0.62353515625,
0.208984375,
0.62353515625,
0.216552734375,
0.6044921875,
0.216552734375
],
[
0.39794921875,
0.236083984375,
0.42236328125,
0.236083984375,
0.42236328125,
0.2474365234375,
0.39794921875,
0.2474365234375
],
[
0.54736328125,
0.23486328125,
0.57421875,
0.23486328125,
0.57421875,
0.246826171875,
0.54736328125,
0.246826171875
],
[
0.39794921875,
0.24560546875,
0.42236328125,
0.24560546875,
0.42236328125,
0.25244140625,
0.39794921875,
0.25244140625
],
[
0.5498046875,
0.246826171875,
0.57177734375,
0.246826171875,
0.57177734375,
0.25390625,
0.5498046875,
0.25390625
],
[
0.60107421875,
0.271484375,
0.6259765625,
0.2734375,
0.625,
0.281005859375,
0.6005859375,
0.279052734375
],
[
0.496826171875,
0.2998046875,
0.521484375,
0.292236328125,
0.52685546875,
0.3037109375,
0.50244140625,
0.311279296875
],
[
0.59619140625,
0.294189453125,
0.61767578125,
0.300537109375,
0.61376953125,
0.309326171875,
0.59228515625,
0.3037109375
],
[
0.2181396484375,
0.35791015625,
0.7861328125,
0.359130859375,
0.7861328125,
0.373779296875,
0.2181396484375,
0.37255859375
],
[
0.2181396484375,
0.373779296875,
0.64794921875,
0.37255859375,
0.64794921875,
0.386474609375,
0.2181396484375,
0.386962890625
],
[
0.2442626953125,
0.41845703125,
0.78369140625,
0.41845703125,
0.78369140625,
0.431884765625,
0.2442626953125,
0.431884765625
],
[
0.2197265625,
0.431884765625,
0.78369140625,
0.43310546875,
0.78369140625,
0.447509765625,
0.2197265625,
0.4462890625
],
[
0.2197265625,
0.447509765625,
0.271240234375,
0.447509765625,
0.271240234375,
0.462158203125,
0.2197265625,
0.462158203125
],
[
0.2181396484375,
0.490478515625,
0.2484130859375,
0.490478515625,
0.2484130859375,
0.50244140625,
0.2181396484375,
0.50244140625
],
[
0.26708984375,
0.4912109375,
0.34716796875,
0.4912109375,
0.34716796875,
0.50244140625,
0.26708984375,
0.50244140625
],
[
0.2197265625,
0.52099609375,
0.78515625,
0.52099609375,
0.78515625,
0.5341796875,
0.2197265625,
0.5341796875
],
[
0.2197265625,
0.53515625,
0.78369140625,
0.53515625,
0.78369140625,
0.548828125,
0.2197265625,
0.548828125
],
[
0.2181396484375,
0.5498046875,
0.2607421875,
0.5498046875,
0.2607421875,
0.56494140625,
0.2181396484375,
0.56494140625
],
[
0.223876953125,
0.59130859375,
0.408447265625,
0.59228515625,
0.408447265625,
0.60693359375,
0.223876953125,
0.60595703125
],
[
0.226318359375,
0.626953125,
0.78515625,
0.626953125,
0.78515625,
0.64013671875,
0.226318359375,
0.64013671875
],
[
0.2197265625,
0.6416015625,
0.78173828125,
0.6416015625,
0.78173828125,
0.65478515625,
0.2197265625,
0.65478515625
],
[
0.2197265625,
0.6572265625,
0.78515625,
0.6572265625,
0.78515625,
0.67041015625,
0.2197265625,
0.67041015625
],
[
0.2181396484375,
0.671875,
0.426513671875,
0.671875,
0.426513671875,
0.68505859375,
0.2181396484375,
0.68505859375
],
[
0.2254638671875,
0.70703125,
0.78369140625,
0.70703125,
0.78369140625,
0.72021484375,
0.2254638671875,
0.72021484375
],
[
0.2197265625,
0.72265625,
0.78515625,
0.72265625,
0.78515625,
0.736328125,
0.2197265625,
0.736328125
],
[
0.221435546875,
0.7373046875,
0.78515625,
0.7373046875,
0.78515625,
0.75048828125,
0.221435546875,
0.75048828125
],
[
0.2197265625,
0.75244140625,
0.78759765625,
0.75244140625,
0.78759765625,
0.767578125,
0.2197265625,
0.767578125
],
[
0.2181396484375,
0.7666015625,
0.52685546875,
0.767578125,
0.52685546875,
0.78515625,
0.2181396484375,
0.7841796875
],
[
0.2254638671875,
0.81103515625,
0.78369140625,
0.81103515625,
0.78369140625,
0.82470703125,
0.2254638671875,
0.82470703125
],
[
0.2197265625,
0.8271484375,
0.78369140625,
0.8271484375,
0.78369140625,
0.84033203125,
0.2197265625,
0.84033203125
]
] |
[
"Co-evolvingGraphReasoningNetworkforEmotion-CausePairExtraction",
"5",
".",
"CE2",
"te:-1",
"et:-1",
"t",
"Cet:0",
"to",
"te:0",
"t:1",
"Ce:2",
"et:2",
"Fig.2. An example of MRG (=2 ). W.L.O.G, only the edges directed into ci,t, and",
"eiare illustrated.And self-loops are not shown for simplification",
"在MRG中，每个任务间关系都对应一个细粒度的相对距离，与标记方案的",
"定义一致。因此，沿着这些关系的任务间图转换可以实现更充分和明确的多任",
"务推理",
"3.2",
"CGR-Net",
"我们的协同进化图推理网络（CGR-Net）的整体架构如图。3（1)。它由三个部",
"分组成：分层编码，初始估计和共同进化推理。接下来，我们描述这三个组件的",
"过程。",
"Hierarchical Encoding",
"Word-levelClauseEncoding子句编码的目的是为每个子句生成包含单词级依",
"赖关系的表示。在前面的工作之后，每个子句都被输人到BERT[8]编码器中",
"然后将[CLS］令牌的最后一个隐藏状态作为子句表示。现在我们获得D的子句",
"表示序列：H=（ho...hn）。",
"Multi-taskClause-levelDocumentEncoding在本文中，我们利用BiLSTM17",
"通过对子句间依赖关系进行建模来生成上下文敏感的子句隐藏状态。为了获得",
"三个任务的特定于任务的子句隐藏状态，我们分别在H上应用三个BLSTM",
"来分别获得原因，标签和情感的初始子句隐藏状态：H°=（h.1,,ho.n），",
"Ho= (h,i...,ht.n)和Hg= (h,i,..he,n)",
"InitialEstimation，由于MRGT单元将上一步中预测的三个任务的标签分",
"布作为输入，因此H°，H和H°分别被馈送到原因解码器，标签解码器和情"
] |
{
"arxiv_id": "2306.04340",
"language": "zh",
"page_idx": 4
}
|
||
1806.06029-zh-0
|
1806.06029-zh-0.jpg
|
# 单镜头无监督跨域转换
Sagie Benaim$^{1}$ and Lior Wolf$^{1,2}$
$^{\mathrm{l}}$The School of Computer Science, Tel Aviv University , Israel
$^{2}$Facebook AI Research
# Abstract
Given a single image $x$ from domain $A$ and a set of images from domain $B$ our task is to generate the analogous of $x$ in $B.$ We argue that this task could be a key AI capability that underlines the ability of cognitive agents to act in the world and present empirical evidence that the existing unsupervised domain translation methods fail on this task. Our method follows a two step process. First, a variational autoencoder for domain $B$ is trained. Then, given the new sample $x$ , we create a variational autoencoder for domain $A$ by adapting the layers that are close to the image in order to directly fit $x$ ,and only indirectly adapt the other layers. Our experiments indicate that the new method does as well, when trained on one sample $x$ , as the existing domain transfer methods, when these enjoy a multitude of training samples from domain $A.$ Our code is made publicly available at https://github.com/sagiebenaim/OneShotTranslation .
### $\textbf{1}$ $\textbf{Introduction}$
智能代理积累知识的直观范例的简化如下。所获得的知识由一个模型捕获，该模型保留了以前看到的样本，并且还能够通过混合观察到的样本来生成新样本。代理通过暴露于一系列对象来不断学习。每当观察到新样本时，代理将使用内部模型生成与观察到的样本类似的虚拟样本，并将观察到的对象和混合对象进行比较，以更新内部模型
感知混合框架$\mathbb{E}[?]$ 的此变体需要多个算法解决方案。一个主要的挑战是“学习悖论”的一个具体案例，即。e。,如何学习它还不知道的东西，或者，在上面的范例中，如果观察到的样本是看不见的，并且可能与已经观察到的任何东西都有很大的不同，那么如何构建类似的心理图像
从计算上讲，此生成步骤需要解决我们称之为一次性无监督跨域翻译的任务：给定来自未知域$A$的单个样本$x$和许多样本，或者几乎等效地，域$B$的模型，生成类似于$x$的样本$y\in B$ 。尽管已经有大量的研究致力于无监督域翻译，其中提供了来自域$A$的许多样本，但据我们所知，文献并未涉及一次性案例
要清楚的是，由于部分文献可能将这些类型的任务称为零拍学习，因此除了要映射的图像$x$ 之外，我们在$A$中没有给出任何训练图像。例如，考虑$\overline{\mathbb{E}}[?]$中提出的将斑马映射到马的任务。给定许多斑马和马的训练图像，现有方法可以很好地执行此任务。然而，似乎完全有可能将单个斑马图像映射到类似的马图像，即使没有看到任何其他斑马图像
我们提出的称为 OST (一次转换) 的方法不对称地使用两个域，并采用两个步骤。首先，为域$B$构造了一个变分自编码器。这使我们能够有效地从域$B$编码样本，并基于随机潜在空间向量生成新样本。为了鼓励通用性，我们进一步增加了$B$ 的样本产生的轻微旋转和随机水平平移
在第二阶段，克隆变分自动编码器以创建两个副本，它们共享编码器的顶层和解码器的底层，一个用于$B$中的样本，一个用于$A$中的样本$x$ 。对自动编码器进行了重建损失以及单样本单向圆度损失的训练。域$B$中的样本继续像第一步一样训练自己的副本，同时更新共32nd Conference on Neural Information Processing Systems (NIPS 2018), Montréal, Canada.
|
[
[
0.361083984375,
0.1275634765625,
0.6396484375,
0.1275634765625,
0.6396484375,
0.1483154296875,
0.361083984375,
0.1483154296875
],
[
0.393798828125,
0.2008056640625,
0.6044921875,
0.199462890625,
0.6044921875,
0.2139892578125,
0.393798828125,
0.21533203125
],
[
0.294189453125,
0.2279052734375,
0.7060546875,
0.2279052734375,
0.7060546875,
0.242431640625,
0.294189453125,
0.242431640625
],
[
0.42236328125,
0.242431640625,
0.57666015625,
0.242431640625,
0.57666015625,
0.255615234375,
0.42236328125,
0.255615234375
],
[
0.463134765625,
0.29541015625,
0.53759765625,
0.29541015625,
0.53759765625,
0.31005859375,
0.463134765625,
0.31005859375
],
[
0.2344970703125,
0.326904296875,
0.75732421875,
0.326904296875,
0.75732421875,
0.34033203125,
0.2344970703125,
0.34033203125
],
[
0.238525390625,
0.341552734375,
0.765625,
0.341552734375,
0.765625,
0.354736328125,
0.238525390625,
0.354736328125
],
[
0.233642578125,
0.354248046875,
0.7646484375,
0.354736328125,
0.7646484375,
0.36865234375,
0.233642578125,
0.367431640625
],
[
0.233642578125,
0.3662109375,
0.7646484375,
0.367431640625,
0.7646484375,
0.3818359375,
0.233642578125,
0.380615234375
],
[
0.233642578125,
0.3818359375,
0.76123046875,
0.38330078125,
0.76123046875,
0.39453125,
0.233642578125,
0.393310546875
],
[
0.2344970703125,
0.396484375,
0.7646484375,
0.396484375,
0.7646484375,
0.410400390625,
0.2344970703125,
0.410400390625
],
[
0.236083984375,
0.410888671875,
0.76318359375,
0.410888671875,
0.76318359375,
0.421630859375,
0.236083984375,
0.421630859375
],
[
0.236083984375,
0.4248046875,
0.76318359375,
0.4248046875,
0.76318359375,
0.43505859375,
0.236083984375,
0.43505859375
],
[
0.236083984375,
0.438232421875,
0.76318359375,
0.438232421875,
0.76318359375,
0.448974609375,
0.236083984375,
0.448974609375
],
[
0.236083984375,
0.45263671875,
0.76123046875,
0.45263671875,
0.76123046875,
0.46337890625,
0.236083984375,
0.46337890625
],
[
0.233642578125,
0.46533203125,
0.7646484375,
0.46337890625,
0.7646484375,
0.4765625,
0.233642578125,
0.479248046875
],
[
0.2344970703125,
0.477783203125,
0.73876953125,
0.479248046875,
0.73876953125,
0.492431640625,
0.2344970703125,
0.4912109375
],
[
0.1732177734375,
0.52099609375,
0.316162109375,
0.52099609375,
0.316162109375,
0.5341796875,
0.1732177734375,
0.5341796875
],
[
0.17724609375,
0.5517578125,
0.8203125,
0.5517578125,
0.8203125,
0.5625,
0.17724609375,
0.5625
],
[
0.176513671875,
0.564453125,
0.82177734375,
0.564453125,
0.82177734375,
0.57763671875,
0.176513671875,
0.57763671875
],
[
0.176513671875,
0.57763671875,
0.82275390625,
0.57763671875,
0.82275390625,
0.59130859375,
0.176513671875,
0.59130859375
],
[
0.176513671875,
0.59228515625,
0.6845703125,
0.59228515625,
0.6845703125,
0.60595703125,
0.176513671875,
0.60595703125
],
[
0.178955078125,
0.6142578125,
0.810546875,
0.6142578125,
0.810546875,
0.625,
0.178955078125,
0.625
],
[
0.80810546875,
0.6171875,
0.81640625,
0.6171875,
0.81640625,
0.6220703125,
0.80810546875,
0.6220703125
],
[
0.17724609375,
0.625,
0.82177734375,
0.625,
0.82177734375,
0.63818359375,
0.17724609375,
0.63818359375
],
[
0.176513671875,
0.6416015625,
0.82177734375,
0.6416015625,
0.82177734375,
0.6513671875,
0.176513671875,
0.6513671875
],
[
0.176513671875,
0.65478515625,
0.25732421875,
0.65478515625,
0.25732421875,
0.66650390625,
0.176513671875,
0.66650390625
],
[
0.17724609375,
0.6748046875,
0.82275390625,
0.6748046875,
0.82275390625,
0.68798828125,
0.17724609375,
0.68798828125
],
[
0.176513671875,
0.6875,
0.82177734375,
0.6875,
0.82177734375,
0.70068359375,
0.176513671875,
0.70068359375
],
[
0.17724609375,
0.7021484375,
0.82275390625,
0.7021484375,
0.82275390625,
0.71533203125,
0.17724609375,
0.71533203125
],
[
0.176513671875,
0.71728515625,
0.451904296875,
0.71728515625,
0.451904296875,
0.72802734375,
0.176513671875,
0.72802734375
],
[
0.178955078125,
0.73779296875,
0.82275390625,
0.73779296875,
0.82275390625,
0.74853515625,
0.178955078125,
0.74853515625
],
[
0.176513671875,
0.74853515625,
0.82177734375,
0.75048828125,
0.82177734375,
0.7646484375,
0.176513671875,
0.76220703125
],
[
0.17724609375,
0.7646484375,
0.82177734375,
0.7646484375,
0.82177734375,
0.77783203125,
0.17724609375,
0.77783203125
],
[
0.17724609375,
0.77978515625,
0.69677734375,
0.77978515625,
0.69677734375,
0.79052734375,
0.17724609375,
0.79052734375
],
[
0.1748046875,
0.79736328125,
0.82275390625,
0.798828125,
0.82275390625,
0.8125,
0.1748046875,
0.81103515625
],
[
0.176513671875,
0.8125,
0.82177734375,
0.8125,
0.82177734375,
0.82568359375,
0.176513671875,
0.82568359375
],
[
0.17724609375,
0.82763671875,
0.82177734375,
0.82763671875,
0.82177734375,
0.83837890625,
0.17724609375,
0.83837890625
],
[
0.176513671875,
0.84130859375,
0.2418212890625,
0.84130859375,
0.2418212890625,
0.8515625,
0.176513671875,
0.8515625
],
[
0.176513671875,
0.85986328125,
0.82275390625,
0.85986328125,
0.82275390625,
0.87353515625,
0.176513671875,
0.87353515625
],
[
0.17724609375,
0.87548828125,
0.3798828125,
0.87548828125,
0.3798828125,
0.88623046875,
0.17724609375,
0.88623046875
],
[
0.3798828125,
0.875,
0.82177734375,
0.875,
0.82177734375,
0.88818359375,
0.3798828125,
0.88818359375
],
[
0.176513671875,
0.88720703125,
0.82275390625,
0.88720703125,
0.82275390625,
0.9013671875,
0.176513671875,
0.9013671875
],
[
0.1748046875,
0.9248046875,
0.7236328125,
0.9248046875,
0.7236328125,
0.93798828125,
0.1748046875,
0.93798828125
]
] |
[
"单镜头无监督跨域转换",
"Sagie Benaim'and LiorWolfl.2",
"I The School of Computer Science,Tel Aviv University, Israel",
"2Facebook AI Research",
"Abstract",
"Given a single image from domain A and a set of images from domain B",
"our task is to generate the analogous of in B.We argue that this task could",
"be a key AI capability that underlines the ability of cognitive agents to act in",
"the world and present empirical evidence that the existing unsupervised domain",
"translationmethodsfail onthistask.Ourmethodfollowsa two stepprocess.First",
"a variational autoencoder for domain B is trained. Then, given the new sample",
"α,wecreateavariationalautoencoderfordomainAbyadaptingthelayersthat",
"areclosetotheimageinordertodirectlyfit,andonlyindirectlyadaptthe",
"otherlayers.Ourexperimentsindicatethatthenewmethoddoesaswell,when",
"trained onone sample,astheexistingdomaintransfermethods,whenthese",
"enjoy a multitude of training samples from domain A.Our code is made publicly",
"availableathttps://github.com/sagiebenaim/OneShotTranslation",
"1Introduction",
"智能代理积累知识的直观范例的简化如下。所获得的知识由一个模型捕获，该模型保留了",
"以前看到的样本，并且还能够通过混合观察到的样本来生成新样本。代理通过暴露于一系",
"列对象来不断学习。每当观察到新样本时，代理将使用内部模型生成与观察到的样本类似",
"的虚拟样本，并将观察到的对象和混合对象进行比较，以更新内部模型",
"感知混合框架E[?的此变体需要多个算法解决方案。一个主要的挑战是“学习悸论”的",
"",
"个具体案例，即。e。，如何学习它还不知道的东西，或者，在上面的范例中，如果观察到的",
"样本是看不见的，并且可能与已经观察到的任何东西都有很大的不同，那么如何构建类似",
"的心理图像",
"从计算上讲，此生成步骤需要解决我们称之为一次性无监督跨域翻译的任务：给定来自未",
"知域A的单个样本和许多样本，或者几乎等效地，域B的模型，生成类似于的样本",
"yEB。尽管已经有大量的研究致力于无监督域翻译，其中提供了来自域A的许多样本，但",
"据我们所知，文献并未涉及一次性案例",
"要清楚的是，由于部分文献可能将这些类型的任务称为零拍学习，因此除了要映射的图像",
"之外，我们在A中没有给出任何训练图像。例如，考虑E[?]中提出的将斑马映射到马的任",
"务。给定许多斑马和马的训练图像，现有方法可以很好地执行此任务。然而，似乎完全有可",
"能将单个斑马图像映射到类似的马图像，即使没有看到任何其他斑马图像",
"我们提出的称为OST（一次转换）的方法不对称地使用两个域，并采用两个步骤。首先，为",
"域B构造了一个变分自编码器。这使我们能够有效地从域B编码样本，并基于随机潜在空",
"间向量生成新样本。为了鼓励通用性，我们进一步增加了B的样本产生的轻微旋转和随机",
"水平平移",
"在第二阶段，克隆变分自动编码器以创建两个副本，它们共享编码器的顶层和解码器的底",
"层，一个用于B中的样本，",
"一个用于A中的样本。对自动编码器进行了重建损失以及单",
"样本单向圆度损失的训练。域B中的样本继续像第一步一样训练自已的副本，同时更新共",
"32nd Conference on Neural Information Processing Systems (NIPS 2018), Montreal, Canada"
] |
{
"arxiv_id": "1806.06029",
"language": "zh",
"page_idx": 0
}
|
||
1806.06029-zh-1
|
1806.06029-zh-1.jpg
|
享层和非共享层。示例$x$中的渐变仅更新非共享层，而不更新共享层。这样，$B$的自动编码器由$x$通过圆度损失在域$B$的非共享层上产生的损失进行调整，并通过后续的共享层适应$B$的样本。这允许共享层逐渐适应新的样本$x$ ,但可以防止在此单个样本上过度拟合。如前所述，将增强应用于$B$ ,也应用于$x$以增加稳定性
我们进行了各种各样的实验，并证明了 OST 在低射场景中优于现有算法。在大多数数据集上，该方法还通过单个训练示例提供了与其他方法针对整个域$A$图像集获得的精度相当的精度。这一成功揭示了无监督域转换的潜在机制，因为在一次性情况下，域$A$中样本间相关性的约束不适用
# 2 $\textbf{Previous Work}$
无监督域转换方法接收两组样本，每个域一个，并学习一个函数，该函数在一个域中的样本与另一个域中的类似样本之间进行映射[2,2,2,2,2,2,2,2,2,2,2]。这样的方法是无监督的，因为两个集合是完全不成对的
可以根据多个线索恢复域之间的映射。首先，域之间的共享对象可以作为监督样本。早期的无监督跨语言词典翻译方法$\mathbb{E}[?,?,?,?]$ 就是这种情况，该方法通过考虑语言间的编辑距离来识别国际单词 (“computer”,“computadora”,“kompüter”)或其他具有共享词源的单词。这些词被用作引导映射过程的种子集
第二个线索是对象关系。通常认为，在域$A$中的对象之间的成对相似性在域$B$转换后得以保留。这在[?]中利用了类之间的 L2 距离。在无监督的单词到单词翻译$\mathbb{E}[?,?,?,?]$的工作中，每种语言中单词之间的关系都由单词向量E[?] 编码，并且通过一种语言的向量到第二种语言的向量的线性变换可以很好地近似翻译
第三个线索是内部对象关系。如果域$A$的对象是复杂的并且包含多个部分，则可以预期，在映射之后，域$B$中的对应物将具有类似的部分排列。通过检查E[?] 中图像两半之间的距离来证明这一点，并且它也是无监督 NLP 翻译方法的基础，该方法可以在观察到不匹配的语料库$\mathbb{E}[?]$之后将一种语言的句子翻译成另一种语言的句子
捕获这些内部对象关系的另一种方法是为两个域构造单独的自动编码器，这两个域共享许多权重[E[?,?]。假设低级图像属性(例如纹理和颜色)是特定于域的，并且中级和顶级属性对于两个域都是通用的
第三个线索也通过用于执行跨域映射的神经网络结构隐式地体现 (在自动编码器体系结构和其他方法中)$\mathbb{E}[?]$。网络的容量限制了可能解决方案的空间和使用的相对较浅的网络，它们的体系结构决定了解决方案的形式。与确保生成的图像来自目标域的 GANE[?] 约束一起，并进一步受到圆度约束$\mathbb{E}[2,2,?]$的限制，消除了映射中的许多歧义
在一次转换的上下文中，不可能在$B$中找到或生成与给定$x\in A$的类似物，因为未定义两个域之间的域不变距离。人们可以尝试使用通用距离，如感知距离，但这将使工作半监督如$\mathbb{E}[?,?]$ (这些方法也不是一次性的)。由于域 $A$ 中没有对象间关系，因此可以使用的唯一提示是第三种类型
我们已经尝试比较$x$内的各种图像部分，从而推广了$\overline{\mathrm{E}}[?]$的图像半部分解决方案。然而这并不工作。相反，我们的工作依赖于域$A$的中级表示类似于$B$的假设，如上所述，这是基于自动编码器的跨域翻译工作$\mathbb{E}[2,?]$的基本假设
# 3 One-Shot Translation
在无监督跨域翻译问题中，为学习算法提供了来自两个域$A$和$B$的未标记数据集。目标是学习一个函数$T$ ,该函数将域$A$中的样本映射到域$B$中的模拟样本。在基于自动编码器的映射技术[2]中，学习了两个编码器/解码器。我们用$E_A(E_B)$表示域$A(B)$的编码器用$G_A$ ($G_B$ )表示解码器。为了将域$A$中的样本$x$转换为域$B$ ,采用$A$的编码器和$B$的解码器，i。e。$,T_AB=G_B\circ E_A$
通过在两个自动编码器之间共享层来给出对翻译形式的强约束。编码器的较低层和解码器的顶层是特定于域的且不共享。编码器的顶层和解码器的底层是共享的。这种共享在两个域的编码上强制采用相同的结构，对于翻译的成功至关重要
具体来说，我们编写$E_A=E^S\circ E_A^U$、$E_B=E^S\circ E_B^U$、$G_A=G_A^U\circ G^S$ 和$G_B=G_B^U\circ G^S$,
其中上标$S$和$U$分别表示共享部分和非共享部分，下标表示域。此结构如图。 1
2
|
[
[
0.176513671875,
0.0921630859375,
0.82275390625,
0.09344482421875,
0.82275390625,
0.106689453125,
0.176513671875,
0.1060791015625
],
[
0.176513671875,
0.106689453125,
0.82177734375,
0.106689453125,
0.82177734375,
0.12060546875,
0.176513671875,
0.12060546875
],
[
0.176513671875,
0.12060546875,
0.82275390625,
0.12060546875,
0.82275390625,
0.1337890625,
0.176513671875,
0.1337890625
],
[
0.176513671875,
0.1337890625,
0.5400390625,
0.1337890625,
0.5400390625,
0.147705078125,
0.176513671875,
0.147705078125
],
[
0.17724609375,
0.1546630859375,
0.82177734375,
0.1546630859375,
0.82177734375,
0.1685791015625,
0.17724609375,
0.1685791015625
],
[
0.178955078125,
0.1685791015625,
0.82177734375,
0.1685791015625,
0.82177734375,
0.1817626953125,
0.178955078125,
0.1817626953125
],
[
0.176513671875,
0.18310546875,
0.82177734375,
0.18310546875,
0.82177734375,
0.1962890625,
0.176513671875,
0.1962890625
],
[
0.17724609375,
0.1988525390625,
0.3056640625,
0.1988525390625,
0.3056640625,
0.208984375,
0.17724609375,
0.208984375
],
[
0.1732177734375,
0.2310791015625,
0.3349609375,
0.2310791015625,
0.3349609375,
0.24560546875,
0.1732177734375,
0.24560546875
],
[
0.178955078125,
0.26318359375,
0.82177734375,
0.26318359375,
0.82177734375,
0.2734375,
0.178955078125,
0.2734375
],
[
0.17724609375,
0.274658203125,
0.82177734375,
0.274658203125,
0.82177734375,
0.28857421875,
0.17724609375,
0.28857421875
],
[
0.17724609375,
0.2890625,
0.419189453125,
0.2890625,
0.419189453125,
0.2998046875,
0.17724609375,
0.2998046875
],
[
0.178955078125,
0.311279296875,
0.8203125,
0.311279296875,
0.8203125,
0.3212890625,
0.178955078125,
0.3212890625
],
[
0.176513671875,
0.323974609375,
0.82275390625,
0.323974609375,
0.82275390625,
0.337158203125,
0.176513671875,
0.337158203125
],
[
0.17724609375,
0.337158203125,
0.82275390625,
0.337158203125,
0.82275390625,
0.35107421875,
0.17724609375,
0.35107421875
],
[
0.176513671875,
0.352783203125,
0.435546875,
0.352783203125,
0.435546875,
0.363037109375,
0.176513671875,
0.363037109375
],
[
0.178955078125,
0.373779296875,
0.8203125,
0.373779296875,
0.8203125,
0.3837890625,
0.178955078125,
0.3837890625
],
[
0.17724609375,
0.385009765625,
0.82177734375,
0.385009765625,
0.82177734375,
0.3984375,
0.17724609375,
0.3984375
],
[
0.178955078125,
0.40087890625,
0.82177734375,
0.40087890625,
0.82177734375,
0.410888671875,
0.178955078125,
0.410888671875
],
[
0.17724609375,
0.415283203125,
0.5009765625,
0.415283203125,
0.5009765625,
0.425537109375,
0.17724609375,
0.425537109375
],
[
0.176513671875,
0.434326171875,
0.82421875,
0.434326171875,
0.82421875,
0.447509765625,
0.176513671875,
0.447509765625
],
[
0.176513671875,
0.447509765625,
0.82177734375,
0.447509765625,
0.82177734375,
0.4609375,
0.176513671875,
0.4609375
],
[
0.176513671875,
0.4609375,
0.82275390625,
0.4609375,
0.82275390625,
0.474853515625,
0.176513671875,
0.474853515625
],
[
0.17724609375,
0.4765625,
0.5908203125,
0.4765625,
0.5908203125,
0.4873046875,
0.17724609375,
0.4873046875
],
[
0.17724609375,
0.49755859375,
0.8203125,
0.49755859375,
0.8203125,
0.50830078125,
0.17724609375,
0.50830078125
],
[
0.17724609375,
0.51025390625,
0.82275390625,
0.51025390625,
0.82275390625,
0.5234375,
0.17724609375,
0.5234375
],
[
0.176513671875,
0.5244140625,
0.3544921875,
0.5244140625,
0.3544921875,
0.53515625,
0.176513671875,
0.53515625
],
[
0.176513671875,
0.54345703125,
0.82275390625,
0.54345703125,
0.82275390625,
0.556640625,
0.176513671875,
0.556640625
],
[
0.17724609375,
0.55810546875,
0.82177734375,
0.55810546875,
0.82177734375,
0.5712890625,
0.17724609375,
0.5712890625
],
[
0.17724609375,
0.5712890625,
0.82177734375,
0.5712890625,
0.82177734375,
0.58544921875,
0.17724609375,
0.58544921875
],
[
0.1748046875,
0.58544921875,
0.68896484375,
0.58544921875,
0.68896484375,
0.5986328125,
0.1748046875,
0.5986328125
],
[
0.17724609375,
0.60791015625,
0.8203125,
0.60791015625,
0.8203125,
0.61865234375,
0.17724609375,
0.61865234375
],
[
0.17724609375,
0.619140625,
0.82177734375,
0.619140625,
0.82177734375,
0.63330078125,
0.17724609375,
0.63330078125
],
[
0.17724609375,
0.63525390625,
0.82275390625,
0.63525390625,
0.82275390625,
0.64599609375,
0.17724609375,
0.64599609375
],
[
0.176513671875,
0.6494140625,
0.32275390625,
0.64794921875,
0.32275390625,
0.65771484375,
0.176513671875,
0.66015625
],
[
0.176513671875,
0.66845703125,
0.82177734375,
0.66845703125,
0.82177734375,
0.681640625,
0.176513671875,
0.681640625
],
[
0.17724609375,
0.681640625,
0.82177734375,
0.681640625,
0.82177734375,
0.69580078125,
0.17724609375,
0.69580078125
],
[
0.176513671875,
0.69580078125,
0.54150390625,
0.69580078125,
0.54150390625,
0.708984375,
0.176513671875,
0.708984375
],
[
0.1732177734375,
0.72998046875,
0.390625,
0.72998046875,
0.390625,
0.74365234375,
0.1732177734375,
0.74365234375
],
[
0.176513671875,
0.76123046875,
0.82275390625,
0.76123046875,
0.82275390625,
0.7744140625,
0.176513671875,
0.7744140625
],
[
0.1748046875,
0.7734375,
0.82177734375,
0.7744140625,
0.82177734375,
0.78857421875,
0.1748046875,
0.787109375
],
[
0.176513671875,
0.78857421875,
0.82177734375,
0.78857421875,
0.82177734375,
0.8017578125,
0.176513671875,
0.8017578125
],
[
0.17724609375,
0.8017578125,
0.82275390625,
0.8017578125,
0.82275390625,
0.81640625,
0.17724609375,
0.81640625
],
[
0.176513671875,
0.81640625,
0.406982421875,
0.81640625,
0.406982421875,
0.830078125,
0.176513671875,
0.830078125
],
[
0.17724609375,
0.8369140625,
0.82275390625,
0.8369140625,
0.82275390625,
0.85107421875,
0.17724609375,
0.85107421875
],
[
0.17724609375,
0.85107421875,
0.82177734375,
0.85107421875,
0.82177734375,
0.8642578125,
0.17724609375,
0.8642578125
],
[
0.17724609375,
0.86572265625,
0.6005859375,
0.86572265625,
0.6005859375,
0.87548828125,
0.17724609375,
0.87548828125
],
[
0.17724609375,
0.8837890625,
0.826171875,
0.8837890625,
0.826171875,
0.89794921875,
0.17724609375,
0.89794921875
],
[
0.176513671875,
0.8984375,
0.7548828125,
0.8984375,
0.7548828125,
0.912109375,
0.176513671875,
0.912109375
],
[
0.48779296875,
0.93994140625,
0.50244140625,
0.9345703125,
0.5087890625,
0.9443359375,
0.494384765625,
0.95068359375
]
] |
[
"享层和非共享层。示例中的渐变仅更新非共享层，而不更新共享层。这样，B的自动编码",
"器由通过圆度损失在域B的非共享层上产生的损失进行调整，并通过后续的共享层适应",
"B的样本。这允许共享层逐渐适应新的样本&，但可以防止在此单个样本上过度拟合。如前",
"所述，将增强应用于B，也应用于以增加稳定性",
"我们进行了各种各样的实验，并证明了OST在低射场景中优于现有算法。在大多数数据集",
"上，该方法还通过单个训练示例提供了与其他方法针对整个域A图像集获得的精度相当的",
"精度。这一成功揭示了无监督域转换的潜在机制，因为在一次性情况下，域A中样本间相",
"关性的纳束不适用",
"2Previous Work",
"无监督域转换方法接收两组样本，每个域一个，并学习一个函数，该函数在一个域中的样本",
"与另一个域中的类似样本之间进行映射[?,？,？,？,？,？,？,,？,？,?】。这样的方法是无监督",
"的，因为两个集合是完全不成对的",
"可以根据多个线索恢复域之间的映射。首先，域之间的共享对象可以作为监督样本。早期的",
"无监督跨语言词典翻译方法E[？，？，？?]就是这种情况，该方法通过考虑语言间的编辑距离",
"来识别国际单词（“computer\"，“computadora”，“kompiter”）或其他具有共享词源的单词。",
"这些词被用作引导映射过程的种了集",
"第二个线索是对象关系。通常认为：在域A中的对象之间的成对相似性在域B转换后得以",
"保留。这在[?]中利用了类之间的L2距离。在无监督的单词到单词翻译E[?,?,?,?]的工作",
"中，每种语言中单词之间的关系都由单词向量E？编码，并且通过一种语言的向量到第二",
"种语言的向量的线性变换可以很好地近似翻译",
"第三个线索是内部对象关系。如果域A的对象是复杂的并且包含多个部分，则可以预期，在",
"映射之后，域B中的对应物将具有类似的部分排列。通过检查E[?]中图像两半之间的距离",
"来证明这一点，并且它也是无监督NLP翻译方法的基础，该方法可以在观察到不匹配的语",
"料库E?之后将一种语言的句子翻译成另一种语言的句子",
"捕获这些内部对象关系的另一种方法是为两个域构造单独的自动编码器，这两个域共享许",
"多权重E[？，？]。假设低级图像属性（例如纹理和颜色）是特定于域的，并且中级和顶级属",
"性对于两个域都是通用的",
"第三个线索也通过用于执行跨域映射的神经网络结构隐式地体现（在自动编码器体系结构",
"和其他方法中)E[?]。网络的容量限制了可能解决方案的空间和使用的相对较浅的网络，它",
"们的体系结构决定了解决方案的形式。与确保生成的图像来自目标域的GANE[?]约束一",
"起，并进一步受到圆度约束E[?,?,?]的限制，消除了映射中的许多歧义",
"在一次转换的上下文中，不可能在B中找到或生成与给定aEA的类似物，因为未定义两",
"个域之间的域不变距离。人们可以尝试使用通用距离，如感知距离，但这将使工作半监督",
"如E?，？（这些方法也不是一次性的）。由于域A中没有对象间关系，因此可以使用的唯",
"一提示是第三种类型",
"我们已经尝试比较内的各种图像部分，从而推广了E[？?]的图像半部分解决方案。然而，",
"这并不工作。相反，我们的工作依赖于域A的中级表示类似于B的假设，如上所述，这是",
"基于自动编码器的跨域翻译工作E[?,?]的基本假设",
"3One-Shot Translation",
"在无监督跨域翻译问题中，为学习算法提供了来自两个域A和B的未标记数据集。目标是",
"学习一个函数T，该函数将域A中的样本映射到域B中的模拟样本。在基于自动编码器的",
"映射技术［?]中，学习了两个编码器/解码器。我们用EA（EB）表示域A（B）的编码器",
"用GA（GB）表示解码器。为了将域A中的样本转换为域B，采用A的编码器和B的",
"解码器，i。e。,TAB=GBoEA",
"通过在两个自动编码器之间共享层来给出对翻译形式的强约束。编码器的较低层和解码器",
"的顶层是特定于域的且不共享。编码器的顶层和解码器的底层是共享的。这种共享在两个",
"域的编码上强制采用相同的结构，对于翻译的成功至关重要",
"具体来说，我们编写EA=ESEU、EB=ES。EU、GA=GUGS和GB=GUGS，",
"其中上标S和U分别表示共享部分和非共享部分，下标表示域。此结构如图。1",
"2"
] |
{
"arxiv_id": "1806.06029",
"language": "zh",
"page_idx": 1
}
|
||
1806.06029-zh-2
|
1806.06029-zh-2.jpg
|
<FigureHere>
Figure 1: Illustration of the two phases of training. (Phase I): Augmented samples from domain $\bar{B},\bar{P}(\Lambda)$ , are used to train a variational autoencoder for domain $\tilde{B}.R_{BB}$ denotes the space of reconstructed samples from $P(\Lambda).$ (Phase II): the variational autoencoder of phase I is cloned, while sharing the weights of part of the encoder $(E^S)$ and part of the decoder $(G^{\hat{S}}).$ These shared parts, marked with a snowflake, are frozen with respect to the sample $x.$ For both phase I and phase II, we train a discriminator $D_B$ to ensure that the generated image belong to the distribution of domain $B.P(x)$ and $P(\Lambda)$ are translated to a common feature space, $C_E$, using $E_A^U$ and $E_B^U$ respectively $C$ (resp$C_G$ )is the space of features, constructed after passing $C_E$ (resp $C^{\prime})$ through the common encoder $E^{S}($resp common decoder $G^S).R_{AB}$ denotes the subspace of samples in $B$ constructed from $P(x)$ , which is generated by augmenting $x.R_AA$ denotes the space of reconstructed samples $\begin{array}{c}\mathrm{from~}P(x).R_{ABA}\text{ denotes the subspace of samples in }A\text{ constructed by translating }P(x)\end{array}$to domain $B$ and then back to $A.$
除了参与两个自动编码器的网络之外，还在两个阶段中训练了对抗性鉴别器$D_B$,以便对域$B$进行建模。域$A$在低射学习的情况下不包含足够的真实示例，此外，由于任务是从$A$映射到$B$,因此不太需要域$A$鉴别器。当映射$x$(在扩充后，使用转换$T$映射到$B$ )时，鉴别器$D_B$用于提供对抗信号
## 3.1 Phase One of Training
在第一阶段，我们采用来自域 B 的图像的训练集$\Lambda$ ,并为该域训练变分自动编码器。该方法采用增强算子，该算子由图像的小随机旋转和水平平移组成。我们用$P(\Lambda)$表示通过随机增加每个样本$s\in\Lambda$构建的训练集
使用以下损失：
(1)
(2)
$$\begin{aligned}&\mathcal{L}_{REC_{B}}=\sum_{s\in P(\Lambda)}\|G_{B}(E_{B}(s))-s\|_{1}\\&\mathcal{L}_{VAE_{B}}=\sum_{s\in P(\Lambda)}\mathrm{KL}(E_{B}\circ P(\Lambda)||\mathcal{N}(0,I))\\&\mathcal{L}_{\mathrm{GAN}_{B}}=\sum_{s\in P(\Lambda)}-\ell(\overline{D_{B}}(G_{B}(E_{B}(s))),0)\\&\mathcal{L}_{\mathrm{D}_{B}}=\sum_{s\in P(\Lambda)}+\ell(D_{B}(\overline{G_{B}}(\overline{E_{B}}(s))),0)+\ell(D_{B}(s),1)\end{aligned}$$
(3)
(4)
其中前三个损失分别是重建损失，变分损失和发电机上的对抗损失，第四个损失是 GAN 鉴别器的损失，其中我们使用条形表示$G_B$在此损失的反向传播期间未更新。$\ell$可以是二元交叉熵或最小二乘损失，$\ell ( x, y) = ( x- y) ^2$ [?]
在第一阶段训练$E_B$和$G_B$时，以下损失被最小化：
(5)
$$\mathcal{L}^{I}=\mathcal{L}_{REC_{B}}+\alpha_{1}\mathcal{L}_{VAE_{B}}+\alpha_{2}L_{\mathrm{GAN}}$$
3
|
[
[
0.65185546875,
0.11553955078125,
0.6953125,
0.117431640625,
0.69384765625,
0.1287841796875,
0.650390625,
0.126953125
],
[
0.66845703125,
0.1458740234375,
0.6806640625,
0.1458740234375,
0.6806640625,
0.1546630859375,
0.66845703125,
0.1546630859375
],
[
0.5556640625,
0.17041015625,
0.5703125,
0.17041015625,
0.5703125,
0.179931640625,
0.5556640625,
0.179931640625
],
[
0.576171875,
0.169189453125,
0.65673828125,
0.169189453125,
0.65673828125,
0.18115234375,
0.576171875,
0.18115234375
],
[
0.6650390625,
0.17041015625,
0.67724609375,
0.17041015625,
0.67724609375,
0.179931640625,
0.6650390625,
0.179931640625
],
[
0.73388671875,
0.1717529296875,
0.78125,
0.1717529296875,
0.78125,
0.1817626953125,
0.73388671875,
0.1817626953125
],
[
0.23779296875,
0.17236328125,
0.261474609375,
0.17236328125,
0.261474609375,
0.179931640625,
0.23779296875,
0.179931640625
],
[
0.2197265625,
0.173583984375,
0.2279052734375,
0.173583984375,
0.2279052734375,
0.177978515625,
0.2197265625,
0.177978515625
],
[
0.271240234375,
0.1749267578125,
0.281982421875,
0.1749267578125,
0.281982421875,
0.1787109375,
0.271240234375,
0.1787109375
],
[
0.396240234375,
0.173583984375,
0.443603515625,
0.173583984375,
0.443603515625,
0.18115234375,
0.396240234375,
0.18115234375
],
[
0.2156982421875,
0.199462890625,
0.2254638671875,
0.199462890625,
0.2254638671875,
0.2052001953125,
0.2156982421875,
0.2052001953125
],
[
0.328369140625,
0.199462890625,
0.343017578125,
0.199462890625,
0.343017578125,
0.2083740234375,
0.328369140625,
0.2083740234375
],
[
0.55224609375,
0.1988525390625,
0.564453125,
0.1988525390625,
0.564453125,
0.2052001953125,
0.55224609375,
0.2052001953125
],
[
0.66650390625,
0.20263671875,
0.67626953125,
0.1962890625,
0.68310546875,
0.2020263671875,
0.6748046875,
0.208984375
],
[
0.325927734375,
0.2279052734375,
0.34228515625,
0.2279052734375,
0.34228515625,
0.236083984375,
0.325927734375,
0.236083984375
],
[
0.65625,
0.2259521484375,
0.69384765625,
0.2259521484375,
0.69384765625,
0.236083984375,
0.65625,
0.236083984375
],
[
0.77392578125,
0.236083984375,
0.79833984375,
0.236083984375,
0.79833984375,
0.24365234375,
0.77392578125,
0.24365234375
],
[
0.3203125,
0.27783203125,
0.38232421875,
0.27783203125,
0.38232421875,
0.292236328125,
0.3203125,
0.292236328125
],
[
0.6435546875,
0.27783203125,
0.71142578125,
0.27783203125,
0.71142578125,
0.292236328125,
0.6435546875,
0.292236328125
],
[
0.1732177734375,
0.314453125,
0.253173828125,
0.3125,
0.253173828125,
0.32568359375,
0.1732177734375,
0.32763671875
],
[
0.2442626953125,
0.3125,
0.82421875,
0.313232421875,
0.82421875,
0.32763671875,
0.2442626953125,
0.326904296875
],
[
0.1748046875,
0.326904296875,
0.82421875,
0.32763671875,
0.82421875,
0.34228515625,
0.1748046875,
0.341552734375
],
[
0.1748046875,
0.341552734375,
0.82275390625,
0.341552734375,
0.82275390625,
0.354736328125,
0.1748046875,
0.354736328125
],
[
0.1748046875,
0.354736328125,
0.82177734375,
0.354736328125,
0.82177734375,
0.369384765625,
0.1748046875,
0.369384765625
],
[
0.1748046875,
0.369384765625,
0.82275390625,
0.369384765625,
0.82275390625,
0.38330078125,
0.1748046875,
0.38330078125
],
[
0.1748046875,
0.3837890625,
0.82421875,
0.3837890625,
0.82421875,
0.397705078125,
0.1748046875,
0.397705078125
],
[
0.176513671875,
0.396484375,
0.82177734375,
0.396484375,
0.82177734375,
0.410888671875,
0.176513671875,
0.410888671875
],
[
0.1748046875,
0.410888671875,
0.82275390625,
0.410888671875,
0.82275390625,
0.4248046875,
0.1748046875,
0.4248046875
],
[
0.1748046875,
0.4248046875,
0.82421875,
0.4248046875,
0.82421875,
0.439453125,
0.1748046875,
0.439453125
],
[
0.1732177734375,
0.4375,
0.82421875,
0.438232421875,
0.82421875,
0.45263671875,
0.1732177734375,
0.451904296875
],
[
0.1748046875,
0.45263671875,
0.82421875,
0.45263671875,
0.82421875,
0.466552734375,
0.1748046875,
0.466552734375
],
[
0.176513671875,
0.46728515625,
0.325927734375,
0.46728515625,
0.325927734375,
0.477783203125,
0.176513671875,
0.477783203125
],
[
0.1748046875,
0.50390625,
0.82177734375,
0.50390625,
0.82177734375,
0.517578125,
0.1748046875,
0.517578125
],
[
0.176513671875,
0.5166015625,
0.82275390625,
0.5166015625,
0.82275390625,
0.53076171875,
0.176513671875,
0.53076171875
],
[
0.176513671875,
0.52978515625,
0.82275390625,
0.52978515625,
0.82275390625,
0.54443359375,
0.176513671875,
0.54443359375
],
[
0.176513671875,
0.54443359375,
0.373291015625,
0.54443359375,
0.373291015625,
0.55810546875,
0.176513671875,
0.55810546875
],
[
0.1748046875,
0.57568359375,
0.214111328125,
0.57568359375,
0.214111328125,
0.5869140625,
0.1748046875,
0.5869140625
],
[
0.2115478515625,
0.57373046875,
0.374267578125,
0.57470703125,
0.374267578125,
0.5888671875,
0.2115478515625,
0.58837890625
],
[
0.176513671875,
0.60107421875,
0.82177734375,
0.60107421875,
0.82177734375,
0.6142578125,
0.176513671875,
0.6142578125
],
[
0.176513671875,
0.61279296875,
0.82177734375,
0.61279296875,
0.82177734375,
0.62744140625,
0.176513671875,
0.62744140625
],
[
0.176513671875,
0.62744140625,
0.419189453125,
0.62744140625,
0.419189453125,
0.6416015625,
0.176513671875,
0.6416015625
],
[
0.1748046875,
0.646484375,
0.281005859375,
0.64794921875,
0.281005859375,
0.662109375,
0.1748046875,
0.6611328125
],
[
0.3095703125,
0.673828125,
0.38232421875,
0.673828125,
0.38232421875,
0.68798828125,
0.3095703125,
0.68798828125
],
[
0.3857421875,
0.671875,
0.55810546875,
0.671875,
0.55810546875,
0.689453125,
0.3857421875,
0.689453125
],
[
0.80224609375,
0.673828125,
0.826171875,
0.673828125,
0.826171875,
0.6875,
0.80224609375,
0.6875
],
[
0.37744140625,
0.689453125,
0.4248046875,
0.689453125,
0.4248046875,
0.7041015625,
0.37744140625,
0.7041015625
],
[
0.306396484375,
0.71044921875,
0.3896484375,
0.7119140625,
0.3896484375,
0.72802734375,
0.306396484375,
0.72607421875
],
[
0.3798828125,
0.708984375,
0.5986328125,
0.708984375,
0.5986328125,
0.7265625,
0.3798828125,
0.7265625
],
[
0.80224609375,
0.7109375,
0.826171875,
0.7109375,
0.826171875,
0.72607421875,
0.80224609375,
0.72607421875
],
[
0.37744140625,
0.7265625,
0.4248046875,
0.7265625,
0.4248046875,
0.7412109375,
0.37744140625,
0.7412109375
],
[
0.313720703125,
0.74853515625,
0.38232421875,
0.74853515625,
0.38232421875,
0.76318359375,
0.313720703125,
0.76318359375
],
[
0.384033203125,
0.74365234375,
0.5947265625,
0.74365234375,
0.5947265625,
0.76513671875,
0.384033203125,
0.76513671875
],
[
0.80224609375,
0.74853515625,
0.826171875,
0.74853515625,
0.826171875,
0.76220703125,
0.80224609375,
0.76220703125
],
[
0.37744140625,
0.7646484375,
0.426513671875,
0.7646484375,
0.426513671875,
0.77880859375,
0.37744140625,
0.77880859375
],
[
0.328369140625,
0.7841796875,
0.38232421875,
0.7861328125,
0.381591796875,
0.8017578125,
0.326904296875,
0.7998046875
],
[
0.384033203125,
0.7822265625,
0.69384765625,
0.7822265625,
0.69384765625,
0.80322265625,
0.384033203125,
0.80322265625
],
[
0.80224609375,
0.7861328125,
0.826171875,
0.7861328125,
0.826171875,
0.7998046875,
0.80224609375,
0.7998046875
],
[
0.37744140625,
0.8017578125,
0.426513671875,
0.8017578125,
0.426513671875,
0.81640625,
0.37744140625,
0.81640625
],
[
0.176513671875,
0.82373046875,
0.82275390625,
0.82373046875,
0.82275390625,
0.8369140625,
0.176513671875,
0.8369140625
],
[
0.1748046875,
0.8359375,
0.82421875,
0.8369140625,
0.82421875,
0.8515625,
0.1748046875,
0.85107421875
],
[
0.17724609375,
0.8515625,
0.495849609375,
0.8515625,
0.495849609375,
0.86572265625,
0.17724609375,
0.86572265625
],
[
0.1748046875,
0.87255859375,
0.5400390625,
0.87255859375,
0.5400390625,
0.88623046875,
0.1748046875,
0.88623046875
],
[
0.369384765625,
0.89013671875,
0.6298828125,
0.892578125,
0.6298828125,
0.9111328125,
0.369384765625,
0.9091796875
],
[
0.80224609375,
0.89306640625,
0.826171875,
0.89306640625,
0.826171875,
0.90771484375,
0.80224609375,
0.90771484375
],
[
0.494384765625,
0.93603515625,
0.50634765625,
0.93603515625,
0.50634765625,
0.94775390625,
0.494384765625,
0.94775390625
]
] |
[
"RAA/RABA",
"GA",
"CE",
"ESOCAGS",
"C",
"Real/Fake",
"ES",
"E",
"",
"Real/Fake",
"20",
"G",
"ED",
"",
"RBB",
"RAB/RB",
"RAB/R",
"(Phase I)",
"(Phase II)",
"Figure l:",
"Illustration of the two phases of training. (Phase I): Augmented samples from domain",
"B, P(A), are used to train a variational autoencoder for domain B: RBB denotes the space of",
"reconstructed samples from P(A) . (Phase II): the variational autoencoder of phase Iis cloned, while",
"sharing the weights of part of the encoder (Es ) and part of the decoder (Gs ). These shared parts",
"marked with a snowflake, are frozen with respect to the sample . For both phase I and phase II",
"we train a discriminator DB to ensure that the generated image belong to the distribution of domain",
"B . P() and P(A) are translated to a common feature space, Ce , using E and Eu respectively",
"C (resp CG ) is the space of features, constructed after passing Ce (resp C) through the common",
"encoder ES (resp common decoder Gs ). RAB denotes the subspace of samples in B constructed",
"from P(), which is generated by augmenting . RAA denotes the space of reconstructed samples",
"from P(r) :. RABA denotes the subspace of samples in A constructed by translating P(c) to domain",
"B and then back to A.",
"除了参与两个自动编码器的网络之外，还在两个阶段中训练了对抗性鉴别器DB，以便对域",
"B进行建模。域A在低射学习的情况下不包含足够的真实示例，此外，由于任务是从A映",
"射到B，因此不太需要域A鉴别器。当映射（在扩充后，使用转换T映射到B）时，鉴",
"别器DB用于提供对抗信号",
"3.1",
"Phase One of Training",
"在第一阶段，我们采用来自域B的图像的训练集A，并为该域训练变分自动编码器。该方",
"法采用增强算子，该算子由图像的小随机旋转和水平平移组成。我们用P（A）表示通过随机",
"增加每个样本sEA构建的训练集",
"使用以下损失：",
"CRECB=",
" GB(EB(s)) - slli",
"(1)",
"SEP(A)",
"CVAEB=",
" KL(EB P(A)W(0, I))",
"(2)",
"SEP(A)",
"CGANB=",
"Z -l(DB(Gb(EB(s),0)",
"(3)",
"SEP(A)",
"LDB=",
"Z +e(Db(GB(EB(s), 0) + l(DB(s), 1)",
"(4)",
"sEP(A)",
"其中前三个损失分别是重建损失，变分损失和发电机上的对抗损失，第四个损失是GAN鉴",
"别器的损失，其中我们使用条形表示GB在此损失的反向传播期间未更新。l可以是二元交",
"叉或最小二乘损失，l(,y)=(α-y)²[?]",
"在第一阶段训练EB和GB时，以下损失被最小化：",
"C' = CRECB +Q1CVAEB +Q2LGAN",
"(5)",
"3"
] |
{
"arxiv_id": "1806.06029",
"language": "zh",
"page_idx": 2
}
|
||
1908.04992-zh-0
|
1908.04992-zh-0.jpg
|
Memory-Based Neighbourhood Embedding for Visual Recognition
Suichan Li$^{1,2*}$ Dapeng Chen$^{3\dagger}$ Bin Liu$^{1,2\dagger}$ Nenghai Yu$^{1,2}$ Rui Zhao$^{3}$
$^{1}$School of Information Science and Technology, University of Science and Technology of China
$^{2}$ Key Laboratory of Electromagnetic Space Information, the Chinese Academy of Sciences
$^{3}$SenseTime Research
lsc1230@ mail. ustc. edu. cn, $\{$ chendapeng, zhaorui $\}$ @ sensetime. com, $\{$ flowice, ynh $\}$ @ ustc. edu. cn
# $\operatorname{Abstract}$
Learning discriminative image feature embeddings is of great importance to visual recognition. To achieve better feature embeddings, most current methods focus on designing different network structures or loss functions, and the estimated feature embeddings are usually only related to the input images. In this paper, we propose Memory-based Neighbourhood Embedding ( MNE) to enhance a general CNN feature by consider- ing its neighbourhood. The method aims to solve two critical problems, i.e., how to acquire more relevant neighbours in the network training and how to aggregate the neighbourhood information for a more discriminative embedding. We frst augment an episodic memory module into the network, which can provide more relevant neighbours for both training and testing. Then the neighbours are organized in a tree graph with the target instance as the root node. The neighbourhood information is gradually aggregated to the root node in a bottom-up manner, and aggregation weights are supervised by the class relationships between the nodes. We apply MNE on image search and few shot learning tasks. Extensive ablation studies demonstrate the effectiveness of each component, and our method significantly outperforms the state-of-theart approaches.
# 1. Introduction
将图像的语义信息编码为特征嵌人是视觉识别的核心要求。希望将来自相同或相关类的图像映射到流形上的附近点，这对于许多应用程序至关重要，例如少拍学习$\mathbb{E}[41,31,37]$ ,视觉搜索$\mathbb{E}[7,13,45]$ ,面部/人识别$\mathbb{E}[20,26,43,2]$ 和细粒度检索$\mathbb{E}[19,18]$。使用理想
$^{*}$This work is done when Suichan Li is an intern at SenseTime.
$^{\dagger}$D. Chen and B. Liu are the co-corresponding authors.
<FigureHere>
Figure 1: A case to explain the basic idea of the proposed method. The red nodes and blue nodes represent the feature embeddings of two classes respectively. The original feature embeddings may not be able to separate the samples of different classes as the two nodes connected by the red line. While our approach, taking the feature and its neighbourhood graph into account, can modify the current feature embeddings with their context structure and output the embeddings more consistent with their labels.
的特征嵌人，分类任务可以减少到最近邻问题，而检索
任务将通过检查图像间的关系变得更容易
为了改进深度神经网络的特征嵌入，最先进的方法主要集中在设计网络结构上。例如，VGGNets$\underline{\mathrm{E}}[30]$和GoogleNet$\underline{\mathbb{E}}[38]$建议增加网络深度可以提高特征嵌入的质量，而 ResNet$\underline{\mathrm{E}}[10]$表明，添加基于身份的跳过连接可以帮助学习更深，更强大的网络。同时，还存在一系列方法，这些方法通过施加不同的损失函数来监督特征嵌人，包括对比损失$\mathbb{E}[8]$ ,三重态损失$\mathbb{E}[45]$ 和四重态损失E[3]。Chenetal。[2] 在一个图像批次内构建了一个图，用 CRF 形成了一种群损。这些损失函数的主要目标是减少类内变化，同时扩大类间变化
尽管这些努力不断带来更好的特征嵌入，但大多数都是基于单个图像来估计特征嵌入，而丰富的上下文信息被忽略了。当同一类别的个体发生剧烈的外观变化，或者不同类别之间的视觉差异较小时，嵌入单个图像以分离不同类别的样本变得非常困难。这种情况的例子如图所示。。1 a, 其中两个类的特征嵌人可以彼此
|
[
[
0.1519775390625,
0.1319580078125,
0.81640625,
0.1331787109375,
0.81640625,
0.15087890625,
0.1519775390625,
0.149658203125
],
[
0.11358642578125,
0.1817626953125,
0.256591796875,
0.177978515625,
0.256591796875,
0.1956787109375,
0.11358642578125,
0.199462890625
],
[
0.273681640625,
0.1817626953125,
0.4345703125,
0.1787109375,
0.4345703125,
0.1962890625,
0.273681640625,
0.199462890625
],
[
0.451904296875,
0.1817626953125,
0.5703125,
0.1787109375,
0.5703125,
0.1962890625,
0.451904296875,
0.199462890625
],
[
0.5888671875,
0.18115234375,
0.7255859375,
0.1787109375,
0.7255859375,
0.1976318359375,
0.5888671875,
0.199462890625
],
[
0.76123046875,
0.18310546875,
0.8544921875,
0.18310546875,
0.8544921875,
0.1976318359375,
0.76123046875,
0.1976318359375
],
[
0.08905029296875,
0.222900390625,
0.89208984375,
0.2235107421875,
0.89208984375,
0.2412109375,
0.08905029296875,
0.240478515625
],
[
0.1070556640625,
0.240478515625,
0.87158203125,
0.2412109375,
0.87158203125,
0.2587890625,
0.1070556640625,
0.25830078125
],
[
0.39453125,
0.2568359375,
0.58642578125,
0.261474609375,
0.58642578125,
0.279052734375,
0.39453125,
0.274658203125
],
[
0.1707763671875,
0.27978515625,
0.810546875,
0.27978515625,
0.810546875,
0.294189453125,
0.1707763671875,
0.294189453125
],
[
0.236083984375,
0.32763671875,
0.314453125,
0.32763671875,
0.314453125,
0.34521484375,
0.236083984375,
0.34521484375
],
[
0.79150390625,
0.354736328125,
0.82275390625,
0.354736328125,
0.82275390625,
0.367431640625,
0.79150390625,
0.367431640625
],
[
0.098876953125,
0.361083984375,
0.468017578125,
0.361083984375,
0.468017578125,
0.375732421875,
0.098876953125,
0.375732421875
],
[
0.79833984375,
0.359130859375,
0.83251953125,
0.373779296875,
0.82177734375,
0.389404296875,
0.78759765625,
0.375
],
[
0.81787109375,
0.365478515625,
0.83251953125,
0.365478515625,
0.83251953125,
0.37255859375,
0.81787109375,
0.37255859375
],
[
0.08087158203125,
0.376953125,
0.468017578125,
0.376953125,
0.468017578125,
0.39013671875,
0.08087158203125,
0.39013671875
],
[
0.080078125,
0.391357421875,
0.465576171875,
0.391357421875,
0.465576171875,
0.40478515625,
0.080078125,
0.40478515625
],
[
0.82177734375,
0.40087890625,
0.8388671875,
0.40087890625,
0.8388671875,
0.4140625,
0.82177734375,
0.4140625
],
[
0.08251953125,
0.4091796875,
0.46728515625,
0.4091796875,
0.46728515625,
0.419189453125,
0.08251953125,
0.419189453125
],
[
0.8408203125,
0.40771484375,
0.85302734375,
0.410400390625,
0.85107421875,
0.415283203125,
0.8388671875,
0.41357421875
],
[
0.82861328125,
0.415283203125,
0.86376953125,
0.421630859375,
0.861328125,
0.429931640625,
0.826171875,
0.423583984375
],
[
0.080078125,
0.42236328125,
0.4697265625,
0.42236328125,
0.4697265625,
0.436279296875,
0.080078125,
0.436279296875
],
[
0.8388671875,
0.4287109375,
0.8486328125,
0.431884765625,
0.84619140625,
0.4375,
0.83642578125,
0.434326171875
],
[
0.08087158203125,
0.439453125,
0.46728515625,
0.439453125,
0.46728515625,
0.449462890625,
0.08087158203125,
0.449462890625
],
[
0.080078125,
0.45068359375,
0.468017578125,
0.451904296875,
0.468017578125,
0.466552734375,
0.080078125,
0.46533203125
],
[
0.68701171875,
0.455078125,
0.76708984375,
0.455078125,
0.76708984375,
0.462158203125,
0.68701171875,
0.462158203125
],
[
0.7900390625,
0.453857421875,
0.88232421875,
0.45263671875,
0.88232421875,
0.46337890625,
0.7900390625,
0.464111328125
],
[
0.63134765625,
0.455810546875,
0.765625,
0.462158203125,
0.7646484375,
0.477783203125,
0.6298828125,
0.4716796875
],
[
0.50732421875,
0.458251953125,
0.5498046875,
0.45263671875,
0.55224609375,
0.46337890625,
0.5107421875,
0.468505859375
],
[
0.57421875,
0.458251953125,
0.60302734375,
0.458251953125,
0.60302734375,
0.462158203125,
0.57421875,
0.462158203125
],
[
0.78125,
0.458251953125,
0.7939453125,
0.458251953125,
0.7939453125,
0.466552734375,
0.78125,
0.466552734375
],
[
0.7939453125,
0.46337890625,
0.85107421875,
0.46337890625,
0.85107421875,
0.473388671875,
0.7939453125,
0.473388671875
],
[
0.5458984375,
0.46337890625,
0.58642578125,
0.464111328125,
0.58642578125,
0.4716796875,
0.5458984375,
0.47021484375
],
[
0.08087158203125,
0.46728515625,
0.46728515625,
0.46728515625,
0.46728515625,
0.48095703125,
0.08087158203125,
0.48095703125
],
[
0.08251953125,
0.484130859375,
0.46728515625,
0.484130859375,
0.46728515625,
0.494384765625,
0.08251953125,
0.494384765625
],
[
0.50341796875,
0.4873046875,
0.89208984375,
0.4873046875,
0.89208984375,
0.50048828125,
0.50341796875,
0.50048828125
],
[
0.08087158203125,
0.49755859375,
0.4697265625,
0.49755859375,
0.4697265625,
0.51123046875,
0.08087158203125,
0.51123046875
],
[
0.50244140625,
0.501953125,
0.890625,
0.501953125,
0.890625,
0.51513671875,
0.50244140625,
0.51513671875
],
[
0.080078125,
0.51123046875,
0.468017578125,
0.51318359375,
0.468017578125,
0.52783203125,
0.080078125,
0.52587890625
],
[
0.50341796875,
0.517578125,
0.890625,
0.517578125,
0.890625,
0.52783203125,
0.50341796875,
0.52783203125
],
[
0.080078125,
0.52783203125,
0.468017578125,
0.5263671875,
0.468017578125,
0.54150390625,
0.080078125,
0.54248046875
],
[
0.50244140625,
0.52978515625,
0.890625,
0.52978515625,
0.890625,
0.54345703125,
0.50244140625,
0.54345703125
],
[
0.080078125,
0.54345703125,
0.468017578125,
0.54345703125,
0.468017578125,
0.556640625,
0.080078125,
0.556640625
],
[
0.50244140625,
0.54345703125,
0.89208984375,
0.54345703125,
0.89208984375,
0.556640625,
0.50244140625,
0.556640625
],
[
0.080078125,
0.5595703125,
0.4697265625,
0.5595703125,
0.4697265625,
0.57275390625,
0.080078125,
0.57275390625
],
[
0.50341796875,
0.55810546875,
0.89208984375,
0.55810546875,
0.89208984375,
0.5712890625,
0.50341796875,
0.5712890625
],
[
0.08087158203125,
0.57373046875,
0.468017578125,
0.57373046875,
0.468017578125,
0.5869140625,
0.08087158203125,
0.5869140625
],
[
0.5048828125,
0.57275390625,
0.890625,
0.57275390625,
0.890625,
0.58349609375,
0.5048828125,
0.58349609375
],
[
0.080078125,
0.5888671875,
0.4697265625,
0.5888671875,
0.4697265625,
0.60302734375,
0.080078125,
0.60302734375
],
[
0.50244140625,
0.58544921875,
0.890625,
0.58544921875,
0.890625,
0.5986328125,
0.50244140625,
0.5986328125
],
[
0.50244140625,
0.5986328125,
0.861328125,
0.5986328125,
0.861328125,
0.6123046875,
0.50244140625,
0.6123046875
],
[
0.08251953125,
0.60498046875,
0.46728515625,
0.60498046875,
0.46728515625,
0.61572265625,
0.08251953125,
0.61572265625
],
[
0.080078125,
0.61865234375,
0.4697265625,
0.61865234375,
0.4697265625,
0.63330078125,
0.080078125,
0.63330078125
],
[
0.080078125,
0.6337890625,
0.468017578125,
0.6337890625,
0.468017578125,
0.64794921875,
0.080078125,
0.64794921875
],
[
0.5048828125,
0.62890625,
0.890625,
0.62890625,
0.890625,
0.642578125,
0.5048828125,
0.642578125
],
[
0.50341796875,
0.64453125,
0.79736328125,
0.64453125,
0.79736328125,
0.65771484375,
0.50341796875,
0.65771484375
],
[
0.08251953125,
0.65087890625,
0.465576171875,
0.65087890625,
0.465576171875,
0.6611328125,
0.08251953125,
0.6611328125
],
[
0.080078125,
0.6640625,
0.4697265625,
0.6640625,
0.4697265625,
0.67822265625,
0.080078125,
0.67822265625
],
[
0.52392578125,
0.6591796875,
0.890625,
0.6591796875,
0.890625,
0.673828125,
0.52392578125,
0.673828125
],
[
0.08087158203125,
0.6787109375,
0.4697265625,
0.6787109375,
0.4697265625,
0.6923828125,
0.08087158203125,
0.6923828125
],
[
0.5048828125,
0.6748046875,
0.890625,
0.6748046875,
0.890625,
0.68798828125,
0.5048828125,
0.68798828125
],
[
0.50244140625,
0.68798828125,
0.890625,
0.689453125,
0.890625,
0.7041015625,
0.50244140625,
0.70263671875
],
[
0.08251953125,
0.69580078125,
0.46728515625,
0.69580078125,
0.46728515625,
0.70556640625,
0.08251953125,
0.70556640625
],
[
0.08087158203125,
0.7109375,
0.465576171875,
0.7109375,
0.465576171875,
0.7216796875,
0.08087158203125,
0.7216796875
],
[
0.50341796875,
0.705078125,
0.890625,
0.705078125,
0.890625,
0.71826171875,
0.50341796875,
0.71826171875
],
[
0.50341796875,
0.71826171875,
0.89208984375,
0.7197265625,
0.89208984375,
0.734375,
0.50341796875,
0.73291015625
],
[
0.080078125,
0.724609375,
0.19287109375,
0.724609375,
0.19287109375,
0.73779296875,
0.080078125,
0.73779296875
],
[
0.5048828125,
0.73486328125,
0.8896484375,
0.73486328125,
0.8896484375,
0.74853515625,
0.5048828125,
0.74853515625
],
[
0.5048828125,
0.75048828125,
0.88818359375,
0.75048828125,
0.88818359375,
0.76513671875,
0.5048828125,
0.76513671875
],
[
0.08087158203125,
0.767578125,
0.21728515625,
0.767578125,
0.21728515625,
0.7822265625,
0.08087158203125,
0.7822265625
],
[
0.50341796875,
0.76513671875,
0.88818359375,
0.7646484375,
0.88818359375,
0.77880859375,
0.50341796875,
0.77978515625
],
[
0.5048828125,
0.77978515625,
0.890625,
0.77978515625,
0.890625,
0.79443359375,
0.5048828125,
0.79443359375
],
[
0.1004638671875,
0.79248046875,
0.468017578125,
0.79248046875,
0.468017578125,
0.80615234375,
0.1004638671875,
0.80615234375
],
[
0.50341796875,
0.79541015625,
0.8798828125,
0.79541015625,
0.8798828125,
0.8095703125,
0.50341796875,
0.8095703125
],
[
0.08251953125,
0.8095703125,
0.46728515625,
0.8095703125,
0.46728515625,
0.8193359375,
0.08251953125,
0.8193359375
],
[
0.52392578125,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.52392578125,
0.82470703125
],
[
0.08416748046875,
0.82275390625,
0.468017578125,
0.82275390625,
0.468017578125,
0.8359375,
0.08416748046875,
0.8359375
],
[
0.50341796875,
0.82470703125,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.50341796875,
0.8388671875
],
[
0.08087158203125,
0.8369140625,
0.468017578125,
0.8369140625,
0.468017578125,
0.85107421875,
0.08087158203125,
0.85107421875
],
[
0.5048828125,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.08251953125,
0.85302734375,
0.468017578125,
0.85302734375,
0.468017578125,
0.8662109375,
0.08251953125,
0.8662109375
],
[
0.5048828125,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.098876953125,
0.875,
0.465576171875,
0.875,
0.465576171875,
0.88818359375,
0.098876953125,
0.88818359375
],
[
0.5048828125,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.5048828125,
0.88525390625
],
[
0.09722900390625,
0.88720703125,
0.43310546875,
0.88818359375,
0.43310546875,
0.8994140625,
0.09722900390625,
0.8984375
],
[
0.5048828125,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.5048828125,
0.90087890625
]
] |
[
"Memory-Based Neighbourhood Embedding for Visual Recognition",
"Suichan Li 1,2 *",
"Dapeng Chen 3 t",
"Bin Liu 1,2t",
"Nenghai Yu 1,2",
"Rui Zhao 3",
"1 School of Information Science and Technology, University of Science and Technology of China",
"2 Key Laboratory of Electromagnetic Space Information, the Chinese Academy of Sciences",
"3 SenseTime Research",
"Isc1230@mail.ustc.edu.cn, { chendapeng,zhaorui } @sensetime.com, { fowice,ynh } @ustc.edu.cn",
"Abstract",
".",
"Learning discriminative image feature embeddings is",
"",
"",
"of great importance to visual recognition. To achieve",
"better feature embeddings, most current methods fo-",
".",
"cus on designing different network structures or loss",
"福",
"Q.",
"functions, and the estimated feature embeddings are",
"",
"usually only relatedtotheinput images. In this paper.",
"we propose Memory-based Neighbourhood Embedding",
"osedmethodove",
"Theobtainedgraph",
"(b)Theneighbourhoodgraph",
"(a)The.c",
"alteat",
"(c)",
"embeddings",
"beddings",
"(MNE) to enhance a general CNN feature by consider-",
"ing its neighbourhood.The method aims to solvetwo",
"Figure l: A case to explain the basic idea of the proposed",
"critical problems, i.e., how to acquire more relevant",
"method. The red nodes and blue nodes represent the fea-",
"neighbours in the network training and how to aggre-",
"ture embeddings of two classes respectively.The original",
"gate the neighbourhood information for a more dis-",
"feature embeddings may not be able to separate the sam-",
"criminative embedding. We first augment an episodic",
"ples of different classes as the two nodes connected by the",
"memory module into the network, which can provide",
"red line. While our approach, taking the feature and its",
"more relevant neighbours for both training and test-",
"neighbourhood graph intoaccount,canmodifythecurrent",
"ing. Then the neighbours are organized in a tree graph",
"feature embeddings with their context structure and out-",
"put the embeddings more consistent with their labels.",
"with the target instance as the root node. The neigh-",
"bourhood information is gradually aggregated to the",
"root node in a bottom-up manner, and aggregation",
"的特征嵌入，分类任务可以减少到最近邻问题，而检索",
"任务将通过检查图像间的关系变得更容易",
"weights are supervised by the class relationships be-",
"tween the nodes. We apply MNE on image search",
"为了改进深度神经网络的特征嵌人，最先进的方法",
"and few shot learning tasks. Extensive ablation studies",
"主要集中在设计网络结构上。例如，VGGNetsE[30]和",
"GoogleNetE[38]建议增加网络深度可以提高特征嵌人",
"demonstratethe effectiveness of each component,and",
"ourmethod significantlyoutperforms the state-of-the",
"的质量，而ResNetE[10]表明，添加基于身份的跳过",
"连接可以帮助学习更深，更强大的网络。同时，还存",
"art approaches.",
"在一系列方法，这些方法通过施加不同的损失函数来",
"监督特征嵌入，包括对比损失E[8]，三重态损失E[45]",
"1.Introduction",
"和四重态损失E[3]。Chenetal。[2]在一个图像批次",
"内构建了一个图，用CRF形成了一种群损。这些损失",
"将图像的语义信息编码为特征嵌人是视觉识别的核",
"函数的主要目标是减少类内变化，同时扩大类间变化",
"心要求。希望将来自相同或相关类的图像映射到流形",
"尽管这些努力不断带来更好的特征嵌人，但大多数",
"上的附近点，这对于许多应用程序至关重要，例如少拍",
"都是基于单个图像来估计特征嵌人，而丰富的上下文",
"学习[41,31,37]，视觉搜索E[713，45]，面部/人识",
"信息被忽略了。当同一类别的个体发生剧烈的外观变",
"别E[20,26,43,2]和细粒度检索E[19，18]。使用理想",
"化，或者不同类别之间的视觉差异较小时，嵌人单个图",
"*This work is done when Suichan Li is an intern at SenseTime",
"像以分离不同类别的样本变得非常困难。这种情况的",
"tD. Chen and B. Liu are the co-corresponding authors.",
"例子如图所示。。1a，其中两个类的特征嵌人可以彼此"
] |
{
"arxiv_id": "1908.04992",
"language": "zh",
"page_idx": 0
}
|
||
1908.04992-zh-1
|
1908.04992-zh-1.jpg
|
非常接近。然而，如果考虑到特征嵌入及其邻域，两个紧密的特征嵌人可以被分类为两个不同的聚类，并且与其邻居的亲和力可以反过来用于修改原始特征嵌人， 从而实现更多的区别特征，如图 3 所示。1 b 和图。1 c
为了有效地增强特征嵌入并充分利用身份注释，我们提出了基于内存的邻居嵌人(MNE)来利用邻居之间的丰富特征和关系。对于传统的神经网络，基于批处理的训练策略确定训练样本只能观察同一批图像的特征。为了获取更多相关的上下文信息，我们使用情景记忆来增强网络，该记忆存储许多图像实例的必要特征和标签，并且在训练期间根据新计算的 CNN 特征来更新记忆。对于图像批处理中的每个图像，我们从内存中检索邻居并将其组织为树图，其中目标实例对应于根节点。基于树结构，我们提出了一种迭代聚合策略，将邻居的信息逐渐传播到根节点。通过这种方式，靠近目标实例的邻居可以比远程实例对目标实例的嵌人产生更大的影响。我们还观察到，如果两个节点属于同一类，则它们之间的聚合权重更大，这有助于在每个类中产生更紧凑的邻域嵌入。因此，引人了一个新的注意力模块来预测聚合权重，并且我们通过节点级的类关系来监督该模块。经过特征聚合的迭代，我们从根节点获得目标实例的邻域嵌人
本文的贡献可以概括为三个方面：(1)我们通过将情节记忆增强到深度神经网络来利用邻域信息进行特征嵌入。内存可以提供更相关的邻居并支持端到端训练。(2) 提出了一种迭代特征聚合策略，对邻域内的信息进行。该策略将邻域组织为树状图，并将相关信息逐渐传播到根节点。(3)为特征聚合引人了一个新的注意力模块，该模块还受到节点语义关系的监督，以更好地分离不同类的特征嵌人。我们将 MNE 应用于图像搜索和少量的镜头学习任务。广泛的消融研究验证了这三种成分的有效性，我们的方法明显优于最先进的方法
# 2. Related Work
所提出的方法旨在利用上下文进行特征学习。它与图神经网络(GNN)有关，它从图结构中聚合信息，并与转导式少射学习中的方法有一些相似之处
基于上下文的特征学习。最近，研究人员开始利用深度神经网络的上下文信息来学习更好的特征嵌入或改进图像间的相似性。Turcot 等人。[39] 建议通过合并相邻图像的有用特征来增强图像的词袋表示。Iscen 等人。[11]通过稀疏线性系统求解器对局部图像区域的描述符进行扩散，以优化图像检索的排名分数。Donoser 等人。[5] 分析了多种扩散机制，推导了检索应用范围内迭代扩散过程的通用框架。尽管这些方法试图涉及更多的图像进行特征学习，但它们需要预先计算特征或只能在一个图像批次中获取特征。为了解决这个问题，我们用一个情景记忆模块来增强神经网络。还存在一些在网络中添加内存模块的工作。Sprechmann 等人。[32] 扩展了带有记忆的神经网络，并将旧的训练示例存储在内存中以进行参数调整。Sukhbaatar 等人。[34]在大型外部存储器上提出了一种递归神经网络架
构，允许在输出符号之前从存储器中多次读取。在我们的例子中，我们从每个训练图像的情景记忆模块中检索邻居，并聚合邻域信息以实现更具判别性的特征嵌人
图形神经网络。图神经网络(GNN)是CNN 从常规欧几里得数据到图的直接扩展。遵循表示学习的思想， 提出了 DeepWalk$\underline{\mathrm{E}}[21]$ 通过将 SkipGram 模型$\underline{\mathrm{E}}[17]$ 与图随机游走相结合来生成图嵌入。GraghSAGE$\boxed{\mathrm{E}}[9]$ 建议以归纳方式计算节点表示。它为每个节点采样一个固定大小的邻域，然后执行简单的特征聚合，如均值池、最大池化和 LSTM。与估计所有节点的图嵌人的GraphSAGE 不同，我们的方法只对目标实例感兴趣。然后可以在树中有效地操作特征聚合，并将信息逐渐传播到根节点。为了改善特征聚合，Petar 等人。[40] 提出了一种基于注意力的架构来对图结构数据执行节点分类，其中注意力权重是隐式学习的。与[40]相比， 我们的方法充分利用了内存中图像的注释，并通过节点级的类关系来监督注意力权重
传递性少射学习。给定一个具有不同类的标记数据集， 少射学习的目标是为一组看不见的新类训练分类器，对于这些新类，只有几个标记的示例可用。与原始的少发学习问题相比，转导的少发任务同时为所有测试实例提供了信息，这使我们可以利用未标记的测试实例。为了解决这个问题，TPN$\mathbb{E}[15]$使用邻域图将标签从标记的实例扩展到未标记的实例，其中邻域结构由高斯相似性函数构造。我们的方法旨在通过特征聚合来增强目标节点的特征嵌入，而不是执行标签传播。FEATE [46]通过采用自我注意机制将嵌人从任务不可知转换为特定于任务。特别是，它直接通过注意力权重选择相关实例，然后组合它们的变换特征以获得新的特征嵌人。我们不使用线性特征组合，而是使用从内存中检索到的邻居构造树图，然后将特征沿树自适应地聚合到根节点
# 3. Methodology
我们的目标是通过检查实例及其邻域之间的关系来增强单个图像实例的嵌人。为了有效地获取特征空间中更相关的邻居，我们用情景记忆增强了神经网络，它提供了大量实例的必要特征和标签。给定从 CNN 主干中提取的目标特征，我们从内存中检索其邻居并将其组织为树状图。最终的嵌人是通过基于注意力的迭代聚合策略在树上获得的。特别是，我们动态修剪树的叶节点，并通过监督邻域中的成对关系来学习注意力权重。整体框架如图所示。2
# 3.1. Episodic Memory
大多数时候，一个图像的特征嵌人已经可以编码其语义信息以进行视觉识别。或者，我们也可以通过检查实例的相邻特征来估计实例的标签，这是因为具有非常相似特征的图像通常属于相同或相关的类。为了允许快速获取更多相关的邻居，同时保持标准深度模型的高性能和良好的泛化，我们添加了一个情景记忆模块
|
[
[
0.08251953125,
0.09344482421875,
0.4697265625,
0.09344482421875,
0.4697265625,
0.106689453125,
0.08251953125,
0.106689453125
],
[
0.476318359375,
0.0985107421875,
0.484375,
0.0985107421875,
0.484375,
0.10162353515625,
0.476318359375,
0.10162353515625
],
[
0.50341796875,
0.09344482421875,
0.8896484375,
0.09344482421875,
0.8896484375,
0.106689453125,
0.50341796875,
0.106689453125
],
[
0.08416748046875,
0.10986328125,
0.46728515625,
0.10986328125,
0.46728515625,
0.12060546875,
0.08416748046875,
0.12060546875
],
[
0.50341796875,
0.10986328125,
0.88818359375,
0.10986328125,
0.88818359375,
0.12060546875,
0.50341796875,
0.12060546875
],
[
0.08416748046875,
0.12371826171875,
0.468017578125,
0.12371826171875,
0.468017578125,
0.136962890625,
0.08416748046875,
0.136962890625
],
[
0.50341796875,
0.125,
0.8896484375,
0.125,
0.8896484375,
0.1351318359375,
0.50341796875,
0.1351318359375
],
[
0.08087158203125,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.1514892578125,
0.08087158203125,
0.1514892578125
],
[
0.50634765625,
0.14013671875,
0.5185546875,
0.14013671875,
0.5185546875,
0.149658203125,
0.50634765625,
0.149658203125
],
[
0.5048828125,
0.154052734375,
0.890625,
0.154052734375,
0.890625,
0.1673583984375,
0.5048828125,
0.1673583984375
],
[
0.101318359375,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.101318359375,
0.1817626953125
],
[
0.50341796875,
0.1685791015625,
0.890625,
0.1685791015625,
0.890625,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.08251953125,
0.184326171875,
0.46728515625,
0.184326171875,
0.46728515625,
0.1976318359375,
0.08251953125,
0.1976318359375
],
[
0.50341796875,
0.184326171875,
0.88818359375,
0.184326171875,
0.88818359375,
0.1976318359375,
0.50341796875,
0.1976318359375
],
[
0.080078125,
0.1988525390625,
0.46728515625,
0.1976318359375,
0.46728515625,
0.2108154296875,
0.080078125,
0.212158203125
],
[
0.50634765625,
0.199462890625,
0.88623046875,
0.199462890625,
0.88623046875,
0.210205078125,
0.50634765625,
0.210205078125
],
[
0.08251953125,
0.21533203125,
0.46728515625,
0.21533203125,
0.46728515625,
0.2259521484375,
0.08251953125,
0.2259521484375
],
[
0.50341796875,
0.21337890625,
0.8896484375,
0.21337890625,
0.8896484375,
0.2266845703125,
0.50341796875,
0.2266845703125
],
[
0.08251953125,
0.2291259765625,
0.468017578125,
0.2291259765625,
0.468017578125,
0.242431640625,
0.08251953125,
0.242431640625
],
[
0.5048828125,
0.2291259765625,
0.890625,
0.2291259765625,
0.890625,
0.242431640625,
0.5048828125,
0.242431640625
],
[
0.08251953125,
0.24560546875,
0.46728515625,
0.24560546875,
0.46728515625,
0.255615234375,
0.08251953125,
0.255615234375
],
[
0.5048828125,
0.24365234375,
0.8896484375,
0.24365234375,
0.8896484375,
0.2568359375,
0.5048828125,
0.2568359375
],
[
0.08251953125,
0.2587890625,
0.468017578125,
0.2587890625,
0.468017578125,
0.272705078125,
0.08251953125,
0.272705078125
],
[
0.5048828125,
0.2587890625,
0.88818359375,
0.2587890625,
0.88818359375,
0.272705078125,
0.5048828125,
0.272705078125
],
[
0.08251953125,
0.274658203125,
0.46728515625,
0.274658203125,
0.46728515625,
0.28857421875,
0.08251953125,
0.28857421875
],
[
0.5048828125,
0.274658203125,
0.8896484375,
0.274658203125,
0.8896484375,
0.28857421875,
0.5048828125,
0.28857421875
],
[
0.08251953125,
0.291748046875,
0.46728515625,
0.291748046875,
0.46728515625,
0.3017578125,
0.08251953125,
0.3017578125
],
[
0.50634765625,
0.291748046875,
0.88623046875,
0.291748046875,
0.88623046875,
0.3017578125,
0.50634765625,
0.3017578125
],
[
0.08416748046875,
0.30615234375,
0.46728515625,
0.30615234375,
0.46728515625,
0.31640625,
0.08416748046875,
0.31640625
],
[
0.50341796875,
0.304931640625,
0.8896484375,
0.304931640625,
0.8896484375,
0.318115234375,
0.50341796875,
0.318115234375
],
[
0.08416748046875,
0.3212890625,
0.465576171875,
0.3212890625,
0.465576171875,
0.33203125,
0.08416748046875,
0.33203125
],
[
0.5048828125,
0.3193359375,
0.88818359375,
0.3193359375,
0.88818359375,
0.333251953125,
0.5048828125,
0.333251953125
],
[
0.08251953125,
0.335205078125,
0.468017578125,
0.335205078125,
0.468017578125,
0.348388671875,
0.08251953125,
0.348388671875
],
[
0.50341796875,
0.335205078125,
0.8896484375,
0.335205078125,
0.8896484375,
0.348388671875,
0.50341796875,
0.348388671875
],
[
0.0849609375,
0.3515625,
0.451904296875,
0.3515625,
0.451904296875,
0.3623046875,
0.0849609375,
0.3623046875
],
[
0.50341796875,
0.348388671875,
0.73046875,
0.348388671875,
0.73046875,
0.3623046875,
0.50341796875,
0.3623046875
],
[
0.449462890625,
0.354736328125,
0.463134765625,
0.354736328125,
0.463134765625,
0.359130859375,
0.449462890625,
0.359130859375
],
[
0.080078125,
0.3642578125,
0.46728515625,
0.363037109375,
0.46728515625,
0.376953125,
0.080078125,
0.37744140625
],
[
0.5048828125,
0.3642578125,
0.88818359375,
0.3642578125,
0.88818359375,
0.37744140625,
0.5048828125,
0.37744140625
],
[
0.080078125,
0.37744140625,
0.468017578125,
0.37890625,
0.468017578125,
0.393310546875,
0.080078125,
0.392578125
],
[
0.5048828125,
0.380615234375,
0.8896484375,
0.380615234375,
0.8896484375,
0.391357421875,
0.5048828125,
0.391357421875
],
[
0.08087158203125,
0.39453125,
0.468017578125,
0.39453125,
0.468017578125,
0.40771484375,
0.08087158203125,
0.40771484375
],
[
0.5048828125,
0.39453125,
0.890625,
0.39453125,
0.890625,
0.40771484375,
0.5048828125,
0.40771484375
],
[
0.08251953125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.08251953125,
0.423583984375
],
[
0.5048828125,
0.410400390625,
0.8896484375,
0.410400390625,
0.8896484375,
0.423583984375,
0.5048828125,
0.423583984375
],
[
0.08087158203125,
0.4248046875,
0.2442626953125,
0.425537109375,
0.2442626953125,
0.439453125,
0.08087158203125,
0.438232421875
],
[
0.5048828125,
0.4248046875,
0.8896484375,
0.425537109375,
0.8896484375,
0.439453125,
0.5048828125,
0.438232421875
],
[
0.10296630859375,
0.442626953125,
0.465576171875,
0.442626953125,
0.465576171875,
0.45263671875,
0.10296630859375,
0.45263671875
],
[
0.5048828125,
0.43994140625,
0.890625,
0.43994140625,
0.890625,
0.453857421875,
0.5048828125,
0.453857421875
],
[
0.08087158203125,
0.455810546875,
0.46728515625,
0.455810546875,
0.46728515625,
0.4697265625,
0.08087158203125,
0.4697265625
],
[
0.50244140625,
0.453857421875,
0.8896484375,
0.455078125,
0.8896484375,
0.4697265625,
0.50244140625,
0.468505859375
],
[
0.08251953125,
0.47021484375,
0.468017578125,
0.47021484375,
0.468017578125,
0.484130859375,
0.08251953125,
0.484130859375
],
[
0.5048828125,
0.47021484375,
0.8896484375,
0.47021484375,
0.8896484375,
0.484130859375,
0.5048828125,
0.484130859375
],
[
0.08251953125,
0.4873046875,
0.46728515625,
0.4873046875,
0.46728515625,
0.49755859375,
0.08251953125,
0.49755859375
],
[
0.50341796875,
0.48486328125,
0.890625,
0.484130859375,
0.890625,
0.49755859375,
0.50341796875,
0.498779296875
],
[
0.08251953125,
0.501953125,
0.468017578125,
0.501953125,
0.468017578125,
0.51513671875,
0.08251953125,
0.51513671875
],
[
0.50634765625,
0.50244140625,
0.8896484375,
0.50244140625,
0.8896484375,
0.51318359375,
0.50634765625,
0.51318359375
],
[
0.08251953125,
0.5166015625,
0.468017578125,
0.5166015625,
0.468017578125,
0.52978515625,
0.08251953125,
0.52978515625
],
[
0.50634765625,
0.517578125,
0.88818359375,
0.517578125,
0.88818359375,
0.52783203125,
0.50634765625,
0.52783203125
],
[
0.08416748046875,
0.53271484375,
0.46728515625,
0.53271484375,
0.46728515625,
0.54345703125,
0.08416748046875,
0.54345703125
],
[
0.50341796875,
0.53076171875,
0.890625,
0.53076171875,
0.890625,
0.54443359375,
0.50341796875,
0.54443359375
],
[
0.08251953125,
0.54541015625,
0.468017578125,
0.54541015625,
0.468017578125,
0.5595703125,
0.08251953125,
0.5595703125
],
[
0.5048828125,
0.54541015625,
0.8896484375,
0.54541015625,
0.8896484375,
0.5595703125,
0.5048828125,
0.5595703125
],
[
0.08251953125,
0.5625,
0.468017578125,
0.5625,
0.468017578125,
0.57275390625,
0.08251953125,
0.57275390625
],
[
0.50341796875,
0.5595703125,
0.8896484375,
0.56005859375,
0.8896484375,
0.57470703125,
0.50341796875,
0.57373046875
],
[
0.08251953125,
0.57763671875,
0.455078125,
0.57763671875,
0.455078125,
0.58837890625,
0.08251953125,
0.58837890625
],
[
0.50341796875,
0.5771484375,
0.5556640625,
0.5771484375,
0.5556640625,
0.59130859375,
0.50341796875,
0.59130859375
],
[
0.07843017578125,
0.60498046875,
0.2279052734375,
0.60400390625,
0.2279052734375,
0.61865234375,
0.07843017578125,
0.619140625
],
[
0.50244140625,
0.599609375,
0.640625,
0.60302734375,
0.640625,
0.62060546875,
0.50244140625,
0.6171875
],
[
0.10455322265625,
0.6318359375,
0.46728515625,
0.6318359375,
0.46728515625,
0.642578125,
0.10455322265625,
0.642578125
],
[
0.52392578125,
0.62890625,
0.890625,
0.62890625,
0.890625,
0.642578125,
0.52392578125,
0.642578125
],
[
0.08251953125,
0.64599609375,
0.46728515625,
0.64599609375,
0.46728515625,
0.6591796875,
0.08251953125,
0.6591796875
],
[
0.50341796875,
0.64453125,
0.8896484375,
0.64453125,
0.8896484375,
0.65771484375,
0.50341796875,
0.65771484375
],
[
0.08416748046875,
0.66015625,
0.390625,
0.66015625,
0.390625,
0.673828125,
0.08416748046875,
0.673828125
],
[
0.50341796875,
0.65771484375,
0.8896484375,
0.66015625,
0.8896484375,
0.673828125,
0.50341796875,
0.671875
],
[
0.08251953125,
0.67529296875,
0.468017578125,
0.67529296875,
0.468017578125,
0.689453125,
0.08251953125,
0.689453125
],
[
0.50341796875,
0.6748046875,
0.8896484375,
0.6748046875,
0.8896484375,
0.68798828125,
0.50341796875,
0.68798828125
],
[
0.08251953125,
0.69140625,
0.468017578125,
0.69140625,
0.468017578125,
0.705078125,
0.08251953125,
0.705078125
],
[
0.50341796875,
0.68798828125,
0.89208984375,
0.689453125,
0.89208984375,
0.70263671875,
0.50341796875,
0.7021484375
],
[
0.08251953125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.08251953125,
0.7197265625
],
[
0.50341796875,
0.705078125,
0.890625,
0.705078125,
0.890625,
0.71826171875,
0.50341796875,
0.71826171875
],
[
0.08416748046875,
0.72265625,
0.468017578125,
0.72265625,
0.468017578125,
0.73291015625,
0.08416748046875,
0.73291015625
],
[
0.5048828125,
0.7216796875,
0.88818359375,
0.7216796875,
0.88818359375,
0.73193359375,
0.5048828125,
0.73193359375
],
[
0.08087158203125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.75,
0.08087158203125,
0.75
],
[
0.5048828125,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.5048828125,
0.74853515625
],
[
0.08251953125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.08251953125,
0.76513671875
],
[
0.50244140625,
0.74853515625,
0.69384765625,
0.75,
0.69384765625,
0.7646484375,
0.50244140625,
0.76318359375
],
[
0.0849609375,
0.76806640625,
0.465576171875,
0.76806640625,
0.465576171875,
0.77880859375,
0.0849609375,
0.77880859375
],
[
0.5009765625,
0.77294921875,
0.67626953125,
0.7744140625,
0.67626953125,
0.7890625,
0.5009765625,
0.787109375
],
[
0.08416748046875,
0.78271484375,
0.465576171875,
0.78271484375,
0.465576171875,
0.79345703125,
0.08416748046875,
0.79345703125
],
[
0.08087158203125,
0.79541015625,
0.46728515625,
0.796875,
0.46728515625,
0.81005859375,
0.08087158203125,
0.8095703125
],
[
0.52392578125,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.52392578125,
0.81005859375
],
[
0.08251953125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.08251953125,
0.82470703125
],
[
0.5048828125,
0.81298828125,
0.8896484375,
0.81298828125,
0.8896484375,
0.82373046875,
0.5048828125,
0.82373046875
],
[
0.08087158203125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.08087158203125,
0.84033203125
],
[
0.50244140625,
0.8271484375,
0.8896484375,
0.8271484375,
0.8896484375,
0.84033203125,
0.50244140625,
0.84033203125
],
[
0.08251953125,
0.84326171875,
0.46728515625,
0.84326171875,
0.46728515625,
0.85400390625,
0.08251953125,
0.85400390625
],
[
0.50244140625,
0.84130859375,
0.890625,
0.84228515625,
0.890625,
0.85595703125,
0.50244140625,
0.85498046875
],
[
0.08251953125,
0.85693359375,
0.46728515625,
0.85595703125,
0.46728515625,
0.869140625,
0.08251953125,
0.87060546875
],
[
0.50244140625,
0.85693359375,
0.8896484375,
0.85595703125,
0.8896484375,
0.869140625,
0.50244140625,
0.87060546875
],
[
0.08087158203125,
0.87060546875,
0.468017578125,
0.8720703125,
0.468017578125,
0.88623046875,
0.08087158203125,
0.88525390625
],
[
0.50341796875,
0.87255859375,
0.890625,
0.87255859375,
0.890625,
0.88623046875,
0.50341796875,
0.88623046875
],
[
0.08087158203125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.5048828125,
0.8896484375,
0.51953125,
0.8896484375,
0.51953125,
0.8994140625,
0.5048828125,
0.8994140625
]
] |
[
"非常接近。然而，如果考虑到特征嵌入及其邻域，两个",
"",
"构，允许在输出符号之前从存储器中多次读取。在我们",
"紧密的特征嵌入可以被分类为两个不同的聚类，并且",
"的例子中，我们从每个训练图像的情景记忆模块中检",
"与其邻居的亲和力可以反过来用于修改原始特征嵌入，",
"索邻居，并聚合邻域信息以实现更具判别性的特征嵌",
"从而实现更多的区别特征，如图3所示。。1b和图。1",
"入",
"图形神经网络。图神经网络（GNN）是CNN从常规",
"为了有效地增强特征嵌人并充分利用身份注释，我",
"欧几里得数据到图的直接扩展。遵循表示学习的思想",
"们提出了基于内存的邻居嵌人（MNE）来利用邻居之",
"提出了DeepWalkE[2]通过将SkipGram模型E[17",
"间的丰富特征和关系。对于传统的神经网络，基于批处",
"与图随机游走相结合来生成图嵌人。GraghSAGEE9",
"理的训练策略确定训练样本只能观察同一批图像的特",
"建议以归纳方式计算节点表示。它为每个节点采样一",
"征。为了获取更多相关的上下文信息，我们使用情景记",
"个固定大小的邻域，然后执行简单的特征聚合，如均值",
"忆来增强网络，该记忆存储许多图像实例的必要特征",
"池、最大池化和LSTM。与估计所有节点的图嵌入的",
"和标签，并且在训练期间根据新计算的CNN特征来更",
"GraphSAGE不同，我们的方法只对目标实例感兴趣",
"新记忆。对于图像批处理中的每个图像，我们从内存中",
"然后可以在树中有效地操作特征聚合，并将信息逐渐",
"检索邻居并将其组织为树图，其中自标实例对应于根",
"传播到根节点。为厂改善特征聚合，Petar等人。40",
"节点。基于树结构，我们提出了一种迭代聚合策略，将",
"提出了一种基于注意力的架构来对图结构数据执行节",
"邻居的信息逐渐传播到根节点。通过这种方式，靠近目",
"点分类，其中注意力权重是隐式学习的。与【40]相比",
"标实例的邻居可以比远程实例对目标实例的嵌人产生",
"我们的方法充分利用了内存中图像的注释，并通过节",
"更大的影响。我们还观察到，如果两个节点属于同",
"点级的类关系来监督注意力权重",
"",
"类，则它们之间的聚合权重更大，这有助于在每个类中",
"传递性少射学习。给定一个具有不同类的标记数据集",
"产生更紧凑的邻域嵌入。因此，引入了一个新的注意力",
"少射学习的目标是为一组看不见的新类训练分类器，对",
"模块来预测聚合权重，并且我们通过节点级的类关系",
"于这些新类，只有几个标记的示例可用。与原始的少发",
"来监督该模块。经过特征聚合的迭代，我们从根节点获",
"学习问题相比，转导的少发任务同时为所有测试实例",
"得目标实例的邻域嵌入",
"提供了信息，这使我们可以利用未标记的测试实例。为",
"本文的贡献可以概括为三个方面：（1）我们通过将",
"了解决这个问题，TPNE[15]使用邻域图将标签从标记",
"情节记忆增强到深度神经网络来利用邻域信息进行特",
"的实例扩展到未标记的实例，其中邻域结构由高斯相",
"征嵌入。内存可以提供更相关的邻居并支持端到端训",
"似性函数构造。我们的方法旨在通过特征聚合来增强",
"练。（2）提出了一种送代特征聚合策略，对邻域内的信",
"目标节点的特征嵌入，而不是执行标签传播。FEATE",
"息进行。该策略将邻域组织为树状图，并将相关信息逐",
"46通过采用自我注意机制将嵌入从任务不可知转换",
"渐传播到根节点。（3）为特征聚合引入了一个新的注意",
"为特定于任务。特别是，它直接通过注意力权重选择相",
"力模块，该模块还受到节点语义关系的监督，以更好地",
"关实例，然后组合它们的变换特征以获得新的特征嵌",
"分离不同类的特征嵌人。我们将MNE应用于图像搜索",
"人。我们不使用线性特征组合，而是使用从内存中检索",
"和少量的镜头学习任务。广泛的消融研究验证了这三",
"到的邻居构造树图，然后将特征沿树自适应地聚合到",
"种成分的有效性，我们的方法明显优于最先进的方法",
"根节点",
"2.RelatedWork",
"3. Methodology",
"所提出的方法旨在利用上下文进行特征学习。它与",
"我们的目标是通过检查实例及其邻域之间的关系来",
"图神经网络（GNN）有关，它从图结构中聚合信息，并",
"增强单个图像实例的嵌人。为了有效地获取特征空间",
"与转导式少射学习中的方法有一些相似之处",
"中更相关的邻居，我们用情景记忆增强了神经网络，它",
"基于上下文的特征学习。最近，研究人员开始利用深度",
"提供了大量实例的必要特征和标签。给定从CNN主干",
"神经网络的上下文信息来学习更好的特征嵌入或改进",
"中提取的目标特征，我们从内存中检索其邻居并将其",
"图像间的相似性。Turcot等人。[39]建议通过合并相",
"组织为树状图。最终的嵌入是通过基于注意力的迭代",
"邻图像的有用特征来增强图像的词袋表示。Iscen等人。",
"聚合策略在树上获得的。特别是，我们动态修剪树的叶",
"[11]通过稀疏线性系统求解器对局部图像区域的描述",
"节点，并通过监督邻域中的成对关系来学习注意力权",
"符进行扩散，以优化图像检索的排名分数。Donoser等",
"重。整体框架如图所示。2",
"人。5分析了多种扩散机制，推导了检索应用范围",
"3.1.EpisodicMemory",
"内选代扩散过程的通用框架。尽管这些方法试图涉及",
"更多的图像进行特征学习，但它们需要预先计算特征",
"大多数时候，一个图像的特征嵌人已经可以编码其",
"或只能在一个图像批次中获取特征。为了解决这个问",
"语义信息以进行视觉识别。或者，我们也可以通过检查",
"题，我们用一个情景记忆模块来增强神经网络。还存",
"实例的相邻特征来估计实例的标签，这是因为具有非",
"在一些在网络中添加内存模块的工作。Sprechmamm等",
"常相似特征的图像通常属于相同或相关的类。为了允",
"人。【32]扩展了带有记忆的神经网络，并将旧的训练",
"许快速获取更多相关的邻居，同时保持标准深度模型",
"示例存储在内存中以进行参数调整。Sukhbaatar等人。",
"的高性能和良好的泛化，我们添加了一个情景记忆模",
"[34]在大型外部存储器上提出了一种递归神经网络架",
"块"
] |
{
"arxiv_id": "1908.04992",
"language": "zh",
"page_idx": 1
}
|
||
1908.04992-zh-3
|
1908.04992-zh-3.jpg
|
<FigureHere>
Figure 3: Illustration of the proposed ASA- $h$ module. We estimate the current feature of a node by aggregating the previous feature and the features of its child nodes. The attention weights are supervised by whether two nodes belong to the same class.
将图中的信息逐渐传播到目标实例。具体来说，节点$u$
的第$h$个特征聚合可以表示为：
$$f_u^h=\mathrm{Aggregate}(f_u^{h-1},\mathcal{C}(u)),$$
(3)
其中$\mathcal{C}(u)$是节点$u$的子节点，$f_u^0$由原始特征初始化。直观上，靠近目标实例的节点应该对最终嵌人有更大的影响。因此，我们在动态图形上执行功能聚合，如Alg 中所示。2。每次，我们更新树中所有分支节点的特征，然后删除叶节点以形成新树。最后，树仅保留一个节点，其特征是邻域嵌入。我们对邻域嵌人施加交叉熵损失：
(4,
$$\mathcal{L}_C(t)=-\sum_{i=1}^Iy_{i,t}\log\left(\frac{\exp(\boldsymbol{w}_i^\top\boldsymbol{f}_t^\prime)}{\sum_{j=1}^I\exp(\boldsymbol{w}_j^\top\boldsymbol{f}_t^\prime)}\right),$$
,其中$y_i,t$是$y_{i,t}=1$的索引标签，如果图像$t$属于$i$
th 类，否则$y_i,t=0$ 。总共有$I$类
3.3. Aggregation with Supervised Attention
像均值池化和最大池化这样的一般聚合策略不能确定哪些邻居更重要。为了自适应地聚合来自同一类的特征，这对于视觉识别至关重要，我们提出了一个名为 ASA 的网络模块，以将特征聚合为具有增强的Attention
图中所有父节点的$h$聚合由同一模块完成(如图。2 ),用 ASA- $h$表示。在该模块中，我们在子节点上引人注意力权重，然后通过以下方式指定聚合：
$$f_u^h=\boldsymbol{W}_A^h\left(\boldsymbol{f}_u^{h-1}+\sum_{v\in\mathcal{C}(u)}a_{u,v}^h\boldsymbol{f}_v^{h-1}\right)+\boldsymbol{b}_A^h,$$
(5,
其中$W_{A}^{h},b_{A}^{h}$是特征转换的参数，而$a_{u,*}^{h}$是注意力权重。子节点$f_v^{h-1}$的特征嵌入如果来自同一个类，则需要将其映射为接近父节点$f_u^{h-1}$的特征嵌人，因此，注意力权重$a_{u,v}^h$需要很高。与大多数隐式学习网络中注意力权重的方法不同，我们利用内存中的标签信息来监督注意力模块。特别是，注意力权重被设计为与两个节点属于同一类的概率成比例：
(6)
$$a_{u,v}^h=p_{u,v}^h/\sum_{k\in\mathcal{C}(u)}p_{u,k}^h.$$
概率$p_{u,v}^h$是通过以下步骤从先前的特征嵌入$f_u^{h-1}$和
$f_v^{h-1}$估计的：
$$d_{u,v}^{h}=\boldsymbol{W}_{D}^{h}(\boldsymbol{f}_{u}^{h-1}-\boldsymbol{f}_{v}^{h-1})+\boldsymbol{b}_{D}^{h},\\p_{u,v}^{h}=\sigma(\boldsymbol{W}_{S}^{h}(\boldsymbol{d}_{u,v}^{h}\circ\boldsymbol{d}_{u,v}^{h})+\boldsymbol{b}_{S}^{h}),$$
(7)
,其中$W_D^h$,。使用 $d_{u,v}^h$, 我们首先执行元素平方，然后将获得的向量投影为标量值，最后使用 sigmoid 函数$\sigma$将标量值归一化为 (0,1)以内。我们监督父节点及其所有子节点之间的概率：
$$\mathcal{L}_{P}^{h}(u)=-\sum_{v\in\mathcal{C}(u)}[y_{u,v}\log p_{u,v}^{h}+(1-y_{u,v})\log(1-p_{u,v}^{h})],$$
(8)
其中$u\in\mathcal{B}^h$是$h$树图的分支节点集。如果节点$u$和$v$属于同一类，则为$y_u,v=1$ ,否则为$y_u,v=0$ 。Fig。3显示了所提出的 ASA- $h$模块的详细架构
总的来说，我们的 MNE 是通过两种损失函数来学习的。一个是多类交叉熵损失(方程$\color{red}{\mathrm{E}}\color{blue}{\mathrm{E}}\color{red}{\mathrm{4}}$),它施加在邻域嵌人上，另一个是二进制交叉熵损失(方程$\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{8}}$), 它监督每个树节点之间的成对概率
## 4. Implementation
提出的基于内存的邻域嵌人(MNE)应用于两个识别任务，i.e. , 图像搜索和转换式少射学习。现在，我们相应地介绍这两个任务的 CNN 骨干选择，内存构建和训练细节
图像搜索。遵循$\mathbb{E}[36]$ 中的实践，我们采用 ResNet-50$\mathbb{F}$ [10]作为骨干网，并将最后一个向下采样块的步幅从 2更改为 1, 这有助于获得更细粒度的特征。我们预先训练 CNN 骨干网对训练集中的所有类进行分类，然后使用预先训练的网络提取整个训练集的特征，并在训练阶段使用它们初始化情景记忆
在训练中，我们通过随机水平翻转和随机擦除$\overline{\mathbb{E}}[48]$ 来增强输入图像，然后使用 Adam$\underline{\mathbb{E}}[12]$优化整个模型。由于 CNN 骨干网已经过预先训练，因此 CNN 骨干网的初始学习率分别为$10^{-5}$,模型其他部分的初始学习率分别为 10$^{-4}$。我们将学习率衰减 0。20 个历元后获得 1 个，40 个历元后获得最终模型。在测试中我们利用 CNN 主干提取图库图像的特征，并将其增强到内存中。使用此内存，我们估计所有查询和图库图像的邻域嵌入，并利用嵌人执行图像搜索
传递性少射学习。为了与现有方法进行公平的比较，我们采用了广泛使用的四层卷积网络[6,31]作为骨干。它包含 4 个块。每个块都有一个内核大小为 3 的卷积层，一个批处理归一化层，一个重新激活层和一个最大池化层。我们在训练集上对具有交叉熵损失的骨干网
|
[
[
0.08251953125,
0.0921630859375,
0.4697265625,
0.0921630859375,
0.4697265625,
0.1060791015625,
0.08251953125,
0.1060791015625
],
[
0.63134765625,
0.10418701171875,
0.76708984375,
0.10418701171875,
0.76708984375,
0.111083984375,
0.63134765625,
0.111083984375
],
[
0.85693359375,
0.10479736328125,
0.8759765625,
0.10479736328125,
0.8759765625,
0.12060546875,
0.85693359375,
0.12060546875
],
[
0.521484375,
0.10986328125,
0.5439453125,
0.10986328125,
0.5439453125,
0.11871337890625,
0.521484375,
0.11871337890625
],
[
0.080078125,
0.106689453125,
0.303955078125,
0.10797119140625,
0.303955078125,
0.121826171875,
0.080078125,
0.12060546875
],
[
0.5087890625,
0.121826171875,
0.5595703125,
0.121826171875,
0.5595703125,
0.136962890625,
0.5087890625,
0.136962890625
],
[
0.7939453125,
0.11614990234375,
0.82275390625,
0.11614990234375,
0.82275390625,
0.12371826171875,
0.7939453125,
0.12371826171875
],
[
0.69140625,
0.1300048828125,
0.7197265625,
0.1287841796875,
0.7197265625,
0.1363525390625,
0.69140625,
0.136962890625
],
[
0.7646484375,
0.126953125,
0.78125,
0.126953125,
0.78125,
0.1363525390625,
0.7646484375,
0.1363525390625
],
[
0.62158203125,
0.1319580078125,
0.6435546875,
0.1319580078125,
0.6435546875,
0.1351318359375,
0.62158203125,
0.1351318359375
],
[
0.1707763671875,
0.1319580078125,
0.37744140625,
0.1319580078125,
0.37744140625,
0.149658203125,
0.1707763671875,
0.149658203125
],
[
0.44677734375,
0.1351318359375,
0.468017578125,
0.1351318359375,
0.468017578125,
0.147705078125,
0.44677734375,
0.147705078125
],
[
0.80224609375,
0.136962890625,
0.81640625,
0.14013671875,
0.81201171875,
0.149658203125,
0.79833984375,
0.146484375
],
[
0.85888671875,
0.1351318359375,
0.8759765625,
0.1351318359375,
0.8759765625,
0.1641845703125,
0.85888671875,
0.1641845703125
],
[
0.69384765625,
0.14013671875,
0.7197265625,
0.14013671875,
0.7197265625,
0.1483154296875,
0.69384765625,
0.1483154296875
],
[
0.619140625,
0.14013671875,
0.64453125,
0.14013671875,
0.64453125,
0.1483154296875,
0.619140625,
0.1483154296875
],
[
0.76123046875,
0.143310546875,
0.77392578125,
0.1351318359375,
0.78515625,
0.143310546875,
0.77294921875,
0.1514892578125
],
[
0.734375,
0.1514892578125,
0.7490234375,
0.1514892578125,
0.7490234375,
0.160400390625,
0.734375,
0.160400390625
],
[
0.765625,
0.1514892578125,
0.775390625,
0.1483154296875,
0.779296875,
0.1546630859375,
0.76953125,
0.1578369140625
],
[
0.65869140625,
0.154052734375,
0.6748046875,
0.154052734375,
0.6748046875,
0.1610107421875,
0.65869140625,
0.1610107421875
],
[
0.57421875,
0.15283203125,
0.5947265625,
0.15283203125,
0.5947265625,
0.1610107421875,
0.57421875,
0.1610107421875
],
[
0.60107421875,
0.1546630859375,
0.62158203125,
0.1546630859375,
0.62158203125,
0.1590576171875,
0.60107421875,
0.1590576171875
],
[
0.62353515625,
0.1572265625,
0.65673828125,
0.1572265625,
0.65673828125,
0.160400390625,
0.62353515625,
0.160400390625
],
[
0.69921875,
0.1572265625,
0.71728515625,
0.1572265625,
0.71728515625,
0.160400390625,
0.69921875,
0.160400390625
],
[
0.861328125,
0.160400390625,
0.87353515625,
0.160400390625,
0.87353515625,
0.1717529296875,
0.861328125,
0.1717529296875
],
[
0.76953125,
0.1610107421875,
0.779296875,
0.1610107421875,
0.779296875,
0.1685791015625,
0.76953125,
0.1685791015625
],
[
0.08251953125,
0.160400390625,
0.468017578125,
0.160400390625,
0.468017578125,
0.173583984375,
0.08251953125,
0.173583984375
],
[
0.7685546875,
0.169189453125,
0.779296875,
0.169189453125,
0.779296875,
0.1787109375,
0.7685546875,
0.1787109375
],
[
0.810546875,
0.1641845703125,
0.82275390625,
0.1641845703125,
0.82275390625,
0.1749267578125,
0.810546875,
0.1749267578125
],
[
0.08251953125,
0.1749267578125,
0.46728515625,
0.1749267578125,
0.46728515625,
0.1881103515625,
0.08251953125,
0.1881103515625
],
[
0.85888671875,
0.169189453125,
0.875,
0.169189453125,
0.875,
0.1925048828125,
0.85888671875,
0.1925048828125
],
[
0.58251953125,
0.18115234375,
0.63134765625,
0.1925048828125,
0.62353515625,
0.21337890625,
0.57421875,
0.2020263671875
],
[
0.65185546875,
0.1900634765625,
0.7060546875,
0.1881103515625,
0.70751953125,
0.20703125,
0.65283203125,
0.208984375
],
[
0.73193359375,
0.1900634765625,
0.77294921875,
0.1900634765625,
0.77294921875,
0.20703125,
0.73193359375,
0.20703125
],
[
0.84716796875,
0.1962890625,
0.8857421875,
0.1962890625,
0.8857421875,
0.2171630859375,
0.84716796875,
0.2171630859375
],
[
0.7890625,
0.1976318359375,
0.82666015625,
0.1976318359375,
0.82666015625,
0.2052001953125,
0.7890625,
0.2052001953125
],
[
0.5146484375,
0.2008056640625,
0.56396484375,
0.20263671875,
0.56201171875,
0.2171630859375,
0.51318359375,
0.21533203125
],
[
0.08251953125,
0.1912841796875,
0.46728515625,
0.1912841796875,
0.46728515625,
0.2020263671875,
0.08251953125,
0.2020263671875
],
[
0.5810546875,
0.203857421875,
0.5888671875,
0.203857421875,
0.5888671875,
0.20703125,
0.5810546875,
0.20703125
],
[
0.5888671875,
0.203857421875,
0.61767578125,
0.2020263671875,
0.61767578125,
0.205810546875,
0.5888671875,
0.2083740234375
],
[
0.08087158203125,
0.2197265625,
0.451904296875,
0.2197265625,
0.451904296875,
0.23291015625,
0.08087158203125,
0.23291015625
],
[
0.08251953125,
0.2052001953125,
0.468017578125,
0.2052001953125,
0.468017578125,
0.2183837890625,
0.08251953125,
0.2183837890625
],
[
0.64453125,
0.2247314453125,
0.68115234375,
0.2310791015625,
0.67724609375,
0.24365234375,
0.640625,
0.2374267578125
],
[
0.69140625,
0.222900390625,
0.826171875,
0.2259521484375,
0.826171875,
0.2474365234375,
0.69140625,
0.2442626953125
],
[
0.515625,
0.2298583984375,
0.57275390625,
0.2298583984375,
0.57275390625,
0.240478515625,
0.515625,
0.240478515625
],
[
0.861328125,
0.2266845703125,
0.87744140625,
0.2266845703125,
0.87744140625,
0.2412109375,
0.861328125,
0.2412109375
],
[
0.08251953125,
0.23486328125,
0.465576171875,
0.23486328125,
0.465576171875,
0.248779296875,
0.08251953125,
0.248779296875
],
[
0.74267578125,
0.2442626953125,
0.7548828125,
0.2442626953125,
0.7548828125,
0.250732421875,
0.74267578125,
0.250732421875
],
[
0.08251953125,
0.251953125,
0.1380615234375,
0.251953125,
0.1380615234375,
0.26318359375,
0.08251953125,
0.26318359375
],
[
0.7900390625,
0.255126953125,
0.80810546875,
0.255126953125,
0.80810546875,
0.264404296875,
0.7900390625,
0.264404296875
],
[
0.69775390625,
0.264404296875,
0.72607421875,
0.264404296875,
0.72607421875,
0.271484375,
0.69775390625,
0.271484375
],
[
0.7236328125,
0.26513671875,
0.74267578125,
0.264404296875,
0.74267578125,
0.268310546875,
0.7236328125,
0.26953125
],
[
0.19287109375,
0.274658203125,
0.205078125,
0.274658203125,
0.205078125,
0.2841796875,
0.19287109375,
0.2841796875
],
[
0.10540771484375,
0.281005859375,
0.26318359375,
0.285400390625,
0.261474609375,
0.306884765625,
0.10455322265625,
0.302978515625
],
[
0.293212890625,
0.27783203125,
0.378173828125,
0.27783203125,
0.378173828125,
0.29541015625,
0.293212890625,
0.29541015625
],
[
0.44677734375,
0.28857421875,
0.46728515625,
0.28857421875,
0.46728515625,
0.300537109375,
0.44677734375,
0.300537109375
],
[
0.50341796875,
0.291748046875,
0.88623046875,
0.291748046875,
0.88623046875,
0.304931640625,
0.50341796875,
0.304931640625
],
[
0.50341796875,
0.302978515625,
0.89208984375,
0.304931640625,
0.89208984375,
0.3193359375,
0.50341796875,
0.317626953125
],
[
0.261474609375,
0.292236328125,
0.408447265625,
0.292236328125,
0.408447265625,
0.317626953125,
0.261474609375,
0.317626953125
],
[
0.1854248046875,
0.3037109375,
0.2099609375,
0.3037109375,
0.2099609375,
0.31494140625,
0.1854248046875,
0.31494140625
],
[
0.50341796875,
0.318115234375,
0.8896484375,
0.318115234375,
0.8896484375,
0.33203125,
0.50341796875,
0.33203125
],
[
0.08087158203125,
0.32568359375,
0.4697265625,
0.32568359375,
0.4697265625,
0.339111328125,
0.08087158203125,
0.339111328125
],
[
0.50341796875,
0.33203125,
0.8896484375,
0.333251953125,
0.8896484375,
0.3466796875,
0.50341796875,
0.34521484375
],
[
0.08087158203125,
0.341552734375,
0.33251953125,
0.341552734375,
0.33251953125,
0.35595703125,
0.08087158203125,
0.35595703125
],
[
0.50341796875,
0.34521484375,
0.66650390625,
0.3466796875,
0.66650390625,
0.35986328125,
0.50341796875,
0.359130859375
],
[
0.080078125,
0.3662109375,
0.416748046875,
0.3662109375,
0.416748046875,
0.380126953125,
0.080078125,
0.380126953125
],
[
0.10296630859375,
0.39013671875,
0.46728515625,
0.39013671875,
0.46728515625,
0.40087890625,
0.10296630859375,
0.40087890625
],
[
0.5009765625,
0.389404296875,
0.890625,
0.38818359375,
0.890625,
0.40283203125,
0.5009765625,
0.404052734375
],
[
0.08251953125,
0.404052734375,
0.468017578125,
0.404052734375,
0.468017578125,
0.417236328125,
0.08251953125,
0.417236328125
],
[
0.5,
0.40283203125,
0.89453125,
0.40283203125,
0.89453125,
0.42041015625,
0.5,
0.42041015625
],
[
0.080078125,
0.41845703125,
0.468017578125,
0.417236328125,
0.468017578125,
0.43115234375,
0.080078125,
0.431884765625
],
[
0.50244140625,
0.419189453125,
0.810546875,
0.419189453125,
0.810546875,
0.43310546875,
0.50244140625,
0.43310546875
],
[
0.08251953125,
0.434326171875,
0.468017578125,
0.434326171875,
0.468017578125,
0.447509765625,
0.08251953125,
0.447509765625
],
[
0.525390625,
0.43505859375,
0.890625,
0.43505859375,
0.890625,
0.448974609375,
0.525390625,
0.448974609375
],
[
0.08251953125,
0.45068359375,
0.148681640625,
0.45068359375,
0.148681640625,
0.4609375,
0.08251953125,
0.4609375
],
[
0.5048828125,
0.45068359375,
0.890625,
0.45068359375,
0.890625,
0.464111328125,
0.5048828125,
0.464111328125
],
[
0.101318359375,
0.464111328125,
0.468017578125,
0.464111328125,
0.468017578125,
0.477783203125,
0.101318359375,
0.477783203125
],
[
0.50341796875,
0.46533203125,
0.89208984375,
0.46533203125,
0.89208984375,
0.479248046875,
0.50341796875,
0.479248046875
],
[
0.08087158203125,
0.479248046875,
0.46728515625,
0.479248046875,
0.46728515625,
0.492431640625,
0.08087158203125,
0.492431640625
],
[
0.5048828125,
0.48095703125,
0.7490234375,
0.48095703125,
0.7490234375,
0.494384765625,
0.5048828125,
0.494384765625
],
[
0.08251953125,
0.494384765625,
0.39794921875,
0.494384765625,
0.39794921875,
0.50830078125,
0.08251953125,
0.50830078125
],
[
0.50244140625,
0.51220703125,
0.6650390625,
0.51220703125,
0.6650390625,
0.52587890625,
0.50244140625,
0.52587890625
],
[
0.1094970703125,
0.51953125,
0.4697265625,
0.51953125,
0.4697265625,
0.537109375,
0.1094970703125,
0.537109375
],
[
0.52392578125,
0.537109375,
0.890625,
0.537109375,
0.890625,
0.5517578125,
0.52392578125,
0.5517578125
],
[
0.08087158203125,
0.5498046875,
0.468017578125,
0.5498046875,
0.468017578125,
0.564453125,
0.08087158203125,
0.564453125
],
[
0.50341796875,
0.55322265625,
0.890625,
0.55322265625,
0.890625,
0.56640625,
0.50341796875,
0.56640625
],
[
0.08087158203125,
0.56494140625,
0.468017578125,
0.56494140625,
0.468017578125,
0.58349609375,
0.08087158203125,
0.58349609375
],
[
0.50341796875,
0.56640625,
0.89208984375,
0.5673828125,
0.89208984375,
0.58203125,
0.50341796875,
0.58056640625
],
[
0.08251953125,
0.58203125,
0.468017578125,
0.58203125,
0.468017578125,
0.5966796875,
0.08251953125,
0.5966796875
],
[
0.5048828125,
0.583984375,
0.5849609375,
0.583984375,
0.5849609375,
0.59521484375,
0.5048828125,
0.59521484375
],
[
0.08087158203125,
0.5986328125,
0.468017578125,
0.5986328125,
0.468017578125,
0.6123046875,
0.08087158203125,
0.6123046875
],
[
0.5048828125,
0.5986328125,
0.890625,
0.5986328125,
0.890625,
0.6123046875,
0.5048828125,
0.6123046875
],
[
0.08087158203125,
0.61279296875,
0.468017578125,
0.61279296875,
0.468017578125,
0.626953125,
0.08087158203125,
0.626953125
],
[
0.5048828125,
0.61279296875,
0.89208984375,
0.61279296875,
0.89208984375,
0.626953125,
0.5048828125,
0.626953125
],
[
0.08251953125,
0.62890625,
0.468017578125,
0.62890625,
0.468017578125,
0.642578125,
0.08251953125,
0.642578125
],
[
0.50341796875,
0.62890625,
0.8896484375,
0.62890625,
0.8896484375,
0.642578125,
0.50341796875,
0.642578125
],
[
0.08251953125,
0.64306640625,
0.2998046875,
0.64306640625,
0.2998046875,
0.6572265625,
0.08251953125,
0.6572265625
],
[
0.50341796875,
0.64306640625,
0.890625,
0.64306640625,
0.890625,
0.6572265625,
0.50341796875,
0.6572265625
],
[
0.5048828125,
0.6591796875,
0.89208984375,
0.6591796875,
0.89208984375,
0.67236328125,
0.5048828125,
0.67236328125
],
[
0.191162109375,
0.66943359375,
0.35693359375,
0.66943359375,
0.35693359375,
0.6875,
0.191162109375,
0.6875
],
[
0.44677734375,
0.67236328125,
0.468017578125,
0.67236328125,
0.468017578125,
0.68505859375,
0.44677734375,
0.68505859375
],
[
0.50341796875,
0.6748046875,
0.71728515625,
0.6748046875,
0.71728515625,
0.68798828125,
0.50341796875,
0.68798828125
],
[
0.279296875,
0.689453125,
0.32275390625,
0.689453125,
0.32275390625,
0.70068359375,
0.279296875,
0.70068359375
],
[
0.52294921875,
0.68994140625,
0.88818359375,
0.68994140625,
0.88818359375,
0.7041015625,
0.52294921875,
0.7041015625
],
[
0.5048828125,
0.705078125,
0.890625,
0.705078125,
0.890625,
0.7197265625,
0.5048828125,
0.7197265625
],
[
0.08087158203125,
0.7138671875,
0.468017578125,
0.7138671875,
0.468017578125,
0.72802734375,
0.08087158203125,
0.72802734375
],
[
0.50244140625,
0.71826171875,
0.8896484375,
0.7197265625,
0.8896484375,
0.734375,
0.50244140625,
0.73291015625
],
[
0.08416748046875,
0.73291015625,
0.17724609375,
0.73291015625,
0.17724609375,
0.744140625,
0.08416748046875,
0.744140625
],
[
0.5048828125,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.5048828125,
0.74853515625
],
[
0.1666259765625,
0.755859375,
0.375732421875,
0.755859375,
0.375732421875,
0.77099609375,
0.1666259765625,
0.77099609375
],
[
0.50341796875,
0.75,
0.890625,
0.75,
0.890625,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.44775390625,
0.76953125,
0.468017578125,
0.76953125,
0.468017578125,
0.77978515625,
0.44775390625,
0.77978515625
],
[
0.50341796875,
0.76513671875,
0.88623046875,
0.7646484375,
0.88623046875,
0.77880859375,
0.50341796875,
0.77978515625
],
[
0.1666259765625,
0.7763671875,
0.381591796875,
0.7763671875,
0.381591796875,
0.79052734375,
0.1666259765625,
0.79052734375
],
[
0.5048828125,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.5048828125,
0.79443359375
],
[
0.5048828125,
0.79541015625,
0.890625,
0.79541015625,
0.890625,
0.8095703125,
0.5048828125,
0.8095703125
],
[
0.0849609375,
0.8017578125,
0.46728515625,
0.80322265625,
0.46728515625,
0.8173828125,
0.0849609375,
0.81640625
],
[
0.50341796875,
0.81005859375,
0.78125,
0.81005859375,
0.78125,
0.82373046875,
0.50341796875,
0.82373046875
],
[
0.08416748046875,
0.8193359375,
0.46728515625,
0.8193359375,
0.46728515625,
0.830078125,
0.08416748046875,
0.830078125
],
[
0.5048828125,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.5048828125,
0.84033203125
],
[
0.08251953125,
0.83349609375,
0.46728515625,
0.83349609375,
0.46728515625,
0.8466796875,
0.08251953125,
0.8466796875
],
[
0.5048828125,
0.84130859375,
0.89208984375,
0.84130859375,
0.89208984375,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.08251953125,
0.8486328125,
0.285888671875,
0.8486328125,
0.285888671875,
0.8623046875,
0.08251953125,
0.8623046875
],
[
0.5048828125,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.08087158203125,
0.869140625,
0.465576171875,
0.87060546875,
0.465576171875,
0.88818359375,
0.08087158203125,
0.88720703125
],
[
0.5048828125,
0.87255859375,
0.890625,
0.87255859375,
0.890625,
0.88623046875,
0.5048828125,
0.88623046875
],
[
0.44775390625,
0.88818359375,
0.468017578125,
0.88818359375,
0.468017578125,
0.90087890625,
0.44775390625,
0.90087890625
],
[
0.50244140625,
0.88720703125,
0.8896484375,
0.88720703125,
0.8896484375,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"将图中的信息逐渐传播到目标实例。具体来说，节点u",
"ATTENTIONWEIGHTGENERATION",
"LOSS",
"",
"的第h个特征聚合可以表示为：",
"eareetings",
"siqmoid",
"",
"",
"",
"fh = Aggregate(fh-1,C(u),",
"(3)",
"广",
"日",
"",
"一",
"?",
"广",
"福",
"",
"",
"",
"",
"",
"图",
"",
"其中C(u）是节点u的子节点，fu由原始特征初始化。",
"一",
"国",
"直观上，靠近目标实例的节点应该对最终嵌人有更大",
"",
"Slcenain",
"elemsewise",
"proetion",
"ategen",
"normalizat",
"childenadgs",
"的影响。因此，我们在动态图形上执行功能聚合，如",
"",
"Ornieeti",
"特征，然后删除叶节点以形成新树。最后，树仅保留",
"Alg中所示。2。每次，我们更新树中所有分支节点的",
"weightedo",
"AGREATONnsprsete.",
"！",
"fh",
"个节点，其特征是邻域嵌入。我们对邻域嵌人施加交叉",
"fh-",
"摘损失：",
"",
"Th",
"eh",
"I",
"Cc(t) = - yit log",
"exp(wT f')",
"(4)",
"Figure 3: Illustration of the proposed ASA- h module",
"We estimate the current feature of a node by aggregating",
"(Ef-1exp(w, ft))",
"i=1",
"the previous feature and the features of its child nodes.",
"其中it是yit=1的索引标签，如果图像t属于i",
"The attention weights are supervised by whether two nodes",
"th类，否则yi,t=0。总共有I类",
"belong to the same class",
"3.3. Aggregation with Supervised Attention",
"像均值池化和最大池化这样的一般聚合策略不能",
"，其中uEBh是h树图的分支节点集。如果节点u和",
"确定哪些邻居更重要。为了自适应地聚合来自同一类",
"属于同一类，则为yu,u=1，否则为yu,=0。Fig。",
"的特征，这对于视觉识别至关重要，我们提出了一个",
"3显示了所提出的ASA-h模块的详细架构",
"名为ASA的网络模块，以将特征聚合为具有增强的",
"总的来说，我们的MNE是通过两种损失函数来学",
"Attention",
"习的。一个是多类交叉熵损失（方程E4），它施加在",
"图中所有父节点的h聚合由同一模块完成（如图。",
"邻域嵌入上，另一个是二进制交叉损失（方程E8），",
"2），用ASA-h表示。在该模块中，我们在子节点上引",
"它监督每个树节点之间的成对概率",
"入注意力权重，然后通过以下方式指定聚合：",
"4.Implementation",
"fh = Wh(fh-1 +Zuec(u)ah,fh-1) +bA, (5)",
"提出的基于内存的邻域嵌入（MNE）应用于两个识",
"其中Wh,bh是特征转换的参数，而ah*是注意力权",
"别任务，i.e.，图像搜索和转换式少射学习。现在，我",
"重。子节点于h-1的特征嵌人如果来自同一个类，则需",
"们相应地介绍这两个任务的CNN骨干选择，内存构建",
"要将其映射为接近父节点fh-1的特征嵌人，因此，注",
"和训练细节",
"意力权重ah，需要很高。与大多数隐式学习网络中注",
"图像搜索。遵循E[36]中的实践，我们采用ResNet-50E",
"意力权重的方法不同，我们利用内存中的标签信息来",
"[10]作为骨干网，并将最后一个向下采样块的步幅从2",
"监督注意力模块。特别是，注意力权重被设计为与两个",
"更改为1，这有助于获得更细粒度的特征。我们预先训",
"节点属于同一类的概率成比例：",
"练CNN骨干网对训练集中的所有类进行分类，然后使",
"用预先训练的网络提取整个训练集的特征，并在训练",
"ah,u=ph,u/ Z ph,k.",
"(6)",
"阶段使用它们初始化情景记忆",
"kEc(u)",
"在训练中，我们通过随机水平翻转和随机擦除E[48]",
"来增强输入图像，然后使用AdamE[12]优化整个模",
"概率ph，是通过以下步骤从先前的特征嵌人fh-1和",
"型。由于CNN骨干网已经过预先训练，因此CNN骨",
"fh-1估计的：",
"干网的初始学习率分别为10-5，模型其他部分的初始",
"dh,=Wb(fh-1-fh-1)+bb",
"学习率分别为10-4。我们将学习率衰减0。20个历",
"(7)",
"元后获得1个，40个历元后获得最终模型。在测试中",
"ph,= a(Ws(dh.uo dh.)+bs),",
"我们利用CNN主干提取图库图像的特征，并将其增强",
"到内存中。使用此内存，我们估计所有查询和图库图像",
"其中Wb，。使用dhw，我们首先执行元素平方，然",
"的邻域嵌入，并利用嵌入执行图像搜索",
"后将获得的向量投影为标量值，最后使用sigmoid函",
"传递性少射学习。为了与现有方法进行公平的比较，我",
"数将标量值归一化为（0,1）以内。我们监督父节点",
"们采用了广泛使用的四层卷积网络「6，311作为骨干。",
"及其所有子节点之间的概率：",
"它包含4个块。每个块都有一个内核大小为3的卷积",
"Cp(u)=-Zwec(u)[yu, log ph,+(1-yu,) log(1-ph,u)],",
"层，一个批处理归一化层，一个重新激活层和一个最大",
"(8)",
"池化层。我们在训练集上对具有交叉摘损失的骨干网"
] |
{
"arxiv_id": "1908.04992",
"language": "zh",
"page_idx": 3
}
|
||
2003.11142-zh-0
|
2003.11142-zh-0.jpg
|
# BigNAS: Scaling Up Neural Architecture Search with Big Single-Stage Models
Jiahui Yu$^1,2$,Pengchong Jin$^1$,Hanxiao Liu$^1$,Gabriel Bender$^1$
Pieter-Jan Kindermans$^1$,Mingxing Tan$^1$,Thomas Huang$^2$,Xiaodan Song$^1$,
Ruoming Pang$^1$, and Quoc Le$^1$
$^{1}\text{Google Brain}\\^{2}\text{University of Illinois at Urbana-Champaign}$
${\mathrm{jiahuiyu@google.com}}$
$\mathbf{Abstract.~Neural~architecture~search~(NAS)~has~shown~promising~ree}$ sults discovering models that are both accurate and fast. For NAS, training a one-shot model has become a popular strategy to rank the relative quality of different architectures ( child models ) using a single set of shared weights. However, while one-shot model weights can effectively rank different network architectures, the absolute accuracies from these shared weights are typically far below those obtained from stand-alone training. To compensate, existing methods assume that the weights must be retrained, finetuned, or otherwise post-processed after the search is completed. These steps significantly increase the compute requirements and complexity of the architecture search and model deployment. In this work, we propose BigNAS, an approach that challenges the conventional wisdom that post-processing of the weights is necessary to get good prediction accuracies. Without extra retraining or post-processing steps , we are able to train a single set of shared weights on ImageNet and use these weights to obtain child models whose sizes range from 200 to 1000 MFLOPs. Our discovered model family, BigNASModels, achieve top-1 accuracies ranging from $76.5\%$ to 80.9% , surpassing state-of-the-art models in this range including EfficientNets and Once-for-All networks without extra retraining or post-processing. We present ablative study and analysis to further understand the proposed BigNASModels.
Keywords: Efficient Neural Architecture Search, AutoML
## 1 Introduction
设计准确高效的网络架构对于边缘设备上的深度学习至关重要。如果将单个神经网络架构部署在较慢的设备 [38] 上，则可能需要超过一个数量级的推理时间。此外，由于硬件和设备驱动程序[34]的差异，即使是具有相似总体速度的两个设备$(e.g.$,由不同制造商制造的电话 CPU)也可能支持非常不同的网络架构。这使得它不仅可以搜索针对特定设备优化的不同大小的架构，还可以确保这些模型可以有效地部署
过去，为了优化单个设备和延迟目标的网络架构$\operatorname { E} [ 32]$ ,神经架构搜索(NAS) 方法[28,39,40]已被证明是有效的。虽然早期的$\dot{\mathrm{NAS}}$方法对于大多数
|
[
[
0.223876953125,
0.146484375,
0.78125,
0.146484375,
0.78125,
0.1610107421875,
0.223876953125,
0.1610107421875
],
[
0.334228515625,
0.169189453125,
0.67236328125,
0.169189453125,
0.67236328125,
0.184326171875,
0.334228515625,
0.184326171875
],
[
0.279296875,
0.2139892578125,
0.72119140625,
0.2139892578125,
0.72119140625,
0.2291259765625,
0.279296875,
0.2291259765625
],
[
0.2320556640625,
0.2298583984375,
0.76953125,
0.2298583984375,
0.76953125,
0.2442626953125,
0.2320556640625,
0.2442626953125
],
[
0.3896484375,
0.24560546875,
0.611328125,
0.2442626953125,
0.611328125,
0.2587890625,
0.3896484375,
0.260009765625
],
[
0.44677734375,
0.267578125,
0.5556640625,
0.270751953125,
0.5556640625,
0.28857421875,
0.44677734375,
0.285400390625
],
[
0.3505859375,
0.28271484375,
0.65283203125,
0.285888671875,
0.65283203125,
0.300537109375,
0.3505859375,
0.29736328125
],
[
0.428955078125,
0.300537109375,
0.57666015625,
0.300537109375,
0.57666015625,
0.314453125,
0.428955078125,
0.314453125
],
[
0.2646484375,
0.34521484375,
0.73876953125,
0.3466796875,
0.73876953125,
0.361083984375,
0.2646484375,
0.35986328125
],
[
0.265625,
0.363037109375,
0.736328125,
0.363037109375,
0.736328125,
0.373779296875,
0.265625,
0.373779296875
],
[
0.2646484375,
0.375732421875,
0.740234375,
0.375732421875,
0.740234375,
0.389404296875,
0.2646484375,
0.389404296875
],
[
0.265625,
0.389404296875,
0.74169921875,
0.389404296875,
0.74169921875,
0.40283203125,
0.265625,
0.40283203125
],
[
0.265625,
0.40478515625,
0.73779296875,
0.40478515625,
0.73779296875,
0.415283203125,
0.265625,
0.415283203125
],
[
0.2646484375,
0.417236328125,
0.73876953125,
0.417236328125,
0.73876953125,
0.43115234375,
0.2646484375,
0.43115234375
],
[
0.2646484375,
0.43115234375,
0.73779296875,
0.43115234375,
0.73779296875,
0.4443359375,
0.2646484375,
0.4443359375
],
[
0.2646484375,
0.44580078125,
0.740234375,
0.44580078125,
0.740234375,
0.458984375,
0.2646484375,
0.458984375
],
[
0.2646484375,
0.458251953125,
0.740234375,
0.458251953125,
0.740234375,
0.4716796875,
0.2646484375,
0.4716796875
],
[
0.265625,
0.472900390625,
0.73876953125,
0.472900390625,
0.73876953125,
0.486083984375,
0.265625,
0.486083984375
],
[
0.2646484375,
0.486083984375,
0.740234375,
0.486083984375,
0.740234375,
0.5,
0.2646484375,
0.5
],
[
0.265625,
0.5,
0.73876953125,
0.5,
0.73876953125,
0.51318359375,
0.265625,
0.51318359375
],
[
0.265625,
0.5146484375,
0.73779296875,
0.5146484375,
0.73779296875,
0.52783203125,
0.265625,
0.52783203125
],
[
0.2646484375,
0.5263671875,
0.73876953125,
0.52783203125,
0.73876953125,
0.54248046875,
0.2646484375,
0.54150390625
],
[
0.265625,
0.54150390625,
0.73876953125,
0.54150390625,
0.73876953125,
0.5546875,
0.265625,
0.5546875
],
[
0.2646484375,
0.5537109375,
0.73876953125,
0.5546875,
0.73876953125,
0.5693359375,
0.2646484375,
0.568359375
],
[
0.265625,
0.57080078125,
0.73779296875,
0.57080078125,
0.73779296875,
0.58056640625,
0.265625,
0.58056640625
],
[
0.265625,
0.58349609375,
0.740234375,
0.58349609375,
0.740234375,
0.5966796875,
0.265625,
0.5966796875
],
[
0.265625,
0.59765625,
0.740234375,
0.59765625,
0.740234375,
0.611328125,
0.265625,
0.611328125
],
[
0.265625,
0.611328125,
0.73876953125,
0.611328125,
0.73876953125,
0.625,
0.265625,
0.625
],
[
0.2646484375,
0.625,
0.6953125,
0.625,
0.6953125,
0.63818359375,
0.2646484375,
0.63818359375
],
[
0.265625,
0.65283203125,
0.65625,
0.65283203125,
0.65625,
0.66357421875,
0.265625,
0.66357421875
],
[
0.2181396484375,
0.69140625,
0.2418212890625,
0.69140625,
0.2418212890625,
0.7041015625,
0.2181396484375,
0.7041015625
],
[
0.25,
0.68994140625,
0.378173828125,
0.69140625,
0.378173828125,
0.70556640625,
0.25,
0.705078125
],
[
0.2197265625,
0.72021484375,
0.78369140625,
0.72021484375,
0.78369140625,
0.734375,
0.2197265625,
0.734375
],
[
0.2197265625,
0.736328125,
0.78515625,
0.736328125,
0.78515625,
0.75,
0.2197265625,
0.75
],
[
0.2197265625,
0.751953125,
0.78369140625,
0.751953125,
0.78369140625,
0.76513671875,
0.2197265625,
0.76513671875
],
[
0.2197265625,
0.7666015625,
0.78515625,
0.7666015625,
0.78515625,
0.78076171875,
0.2197265625,
0.78076171875
],
[
0.2181396484375,
0.78076171875,
0.7861328125,
0.77978515625,
0.7861328125,
0.79345703125,
0.2181396484375,
0.79443359375
],
[
0.2197265625,
0.796875,
0.366943359375,
0.796875,
0.366943359375,
0.81005859375,
0.2197265625,
0.81005859375
],
[
0.24267578125,
0.81103515625,
0.78515625,
0.81005859375,
0.78515625,
0.82470703125,
0.24267578125,
0.82568359375
],
[
0.2181396484375,
0.8271484375,
0.78515625,
0.8271484375,
0.78515625,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"BigNAS: Scaling Up Neural Architecture Search",
"with Big Single-Stage Models",
"Jiahui Yul,2, Pengchong Jin', Hanxiao Liul, Gabriel Benderl",
"Pieter-Jan Kindermans', Mingxing Tan', Thomas Huang?, Xiaodan Song'",
"Ruoming Pang', and Quoc Le",
"1 Google Brain",
"2University of Ilinois at Urbana-Champaign",
"jiahuiyu@google.com",
"Abstract. Neural architecture search (NAS) has shown promising re-",
"sults discovering models that are both accurate and fast.For NAS, train-",
"ing a one-shot model has become a popular strategy to rank the relative",
"quality of different architectures （child models ) using a single set of",
"shared weights.However, while one-shot model weights can effectively",
"rank different network architectures, the absolute accuracies from these",
"shared weights are typically far below those obtained from stand-alone",
"training. To compensate, existing methods assume that the weights must",
"be retrained, finetuned, or otherwise post-processed after the search is",
"completed. These steps significantly increase the compute requirements",
"and complexity of the architecture search and model deployment. In this",
"work, we propose BigNAS, an approach that challenges the conventional",
"wisdom that post-processing of the weights is necessary to get good pre-",
"diction accuracies. Without extra retraining or post-processing steps,",
"we are able to train a single set of shared weights on ImageNet and use",
"these weights to obtain child models whose sizes range from 200 to 1000",
"MFLOPs.Our discovered modelfamily.BigNASModels,achievetop-1",
"accuracies ranging from 76.5 % to 80.9 % , surpassing state-of-the-art",
"models in this range including EfficientNets and Once-for-All networks",
"without extra retraining or post-processing. We present ablative study",
"and analysis to further understand the proposed BigNASModels.",
"Keywords:Efficient Neural Architecture Search,AutoML",
"1",
"Introduction",
"设计准确高效的网络架构对于边缘设备上的深度学习至关重要。如果将单个神",
"经网络架构部署在较慢的设备[38]上，则可能需要超过一个数量级的推理时间。",
"此外，由于硬件和设备驱动程序[34]的差异，即使是具有相似总体速度的两个",
"设备（e.g.，由不同制造商制造的电话CPU）也可能支持非常不同的网络架构。",
"这使得它不仅可以搜索针对特定设备优化的不同大小的架构，还可以确保这些",
"模型可以有效地部署",
"过去，为了优化单个设备和延迟目标的网络架构[32]，神经架构搜索",
"(NAS）方法[28,39,40]已被证明是有效的。虽然早期的NAS方法对于大多数"
] |
{
"arxiv_id": "2003.11142",
"language": "zh",
"page_idx": 0
}
|
||
2003.11142-zh-1
|
2003.11142-zh-1.jpg
|
2
J. Yu et al.
<FigureHere>
Fig.1: Comparison with several existing workflows. We use nested squares to denote models with shared weights, and use the size of the square to denote the size of each model. Workflow in the middle refers the concurrent work from [5] , where submodels are sequentially induced through progressive distillation and channel sorting. We simultaneously train all child models in a single-stage model with proposed modifications, and deploy them without retraining or finetuning.
从业者来说非常昂贵，但最近基于权重共享的高效 NAS 方法将搜索成本降低了[2,23,26,36]个数量级。这些方法的工作原理是训练一个超级网络，然后识别通过网络的路径-其操作的子集-在满足用户指定的特定硬件设备的延迟约束的同时提供最佳的准确性。这种方法的优点是我们可以训练超级网络，然后使用它从用户定义的搜索空间对许多不同的候选架构进行排名
然而，从这个超级网络获得的预测的绝对精度通常远低于以独立方式从头开始训练的模型[2] 。因此，通常假设需要对超级网络的权重进行大量的后处理以获得高质量的模型部署精度。例如，一个建议的解决方案是为每个感兴趣的设备和每个感兴趣的延迟预算重新训练一个单独的模型[6,34] 。但是，这会产生很大的开销，特别是在部署方案数量很大的情况下。第二种解决方案是在训练完成后对权重进行后处理；例如，使用针对一次性网络 [5] 提出的渐进式收缩启发式算法。然而，这个后处理步骤使模型训练管道变得复杂。此外，来自“一劳永逸”网络[5]的子模型仍然需要通过其他时期$(e.g.$, ImageNet 上的 75 个时期)进行微调，以实现最佳精度
在这项工作中，我们重新评估了普遍的信念，即为了获得有竞争力的准确性必须对共享权重进行重新训练或后处理。我们提出了几种技术来弥合具有共享参数的小型和大型子模型之间不同的初始化和学习动态之间的差距。通过这些技术，我们能够训练一个单阶段模型：一个单一的模型，我们可以直接切片高质量的子模型，而无需任何额外的后处理
我们搜索一个大型单阶段模型，该模型包含小型子模型(???MFLOP,与MobileNetV3 相当) 和大型子模型(???GFLOP,与 EfficientNets 相当)。与现有的一次性方法$\mathbb{E}[2,4,23,26,36]$ 不同 , 我们训练的单阶段模型提供了更广泛的模型容量覆盖范围，更重要的是，所有子模型都经过了训练，以便在搜索阶段结束时同时达到出色的性能。然后可以通过简单的从粗到精的选择策略进行架构选择。一旦选择了架构，我们可以通过简单地将单阶段模型切片以进行即时部署来获得子模型 w。r。t。给定的约束，如内存占用和/或运行时延迟。工作流程如图 1 所示
简化的 BigNAS 工作流的成功依赖于一个目标：如何训练高质量的单阶段模型？该目标本身具有挑战性。例如，我们发现，如果一个大的单阶段模型没有正确初始化，训练损失就会爆炸；在训练过程中，大的孩子模型在小的孩子模型之前就开始过度拟合；从经验上看，更大的孩子模型往往在训练数据上过度拟合
|
[
[
0.21728515625,
0.11614990234375,
0.2548828125,
0.11614990234375,
0.2548828125,
0.1300048828125,
0.21728515625,
0.1300048828125
],
[
0.27294921875,
0.117431640625,
0.346435546875,
0.117431640625,
0.346435546875,
0.1275634765625,
0.27294921875,
0.1275634765625
],
[
0.346435546875,
0.1514892578125,
0.428955078125,
0.1514892578125,
0.428955078125,
0.1622314453125,
0.346435546875,
0.1622314453125
],
[
0.486083984375,
0.15087890625,
0.60302734375,
0.1514892578125,
0.60302734375,
0.1634521484375,
0.486083984375,
0.1622314453125
],
[
0.2320556640625,
0.1572265625,
0.3095703125,
0.160400390625,
0.3095703125,
0.1717529296875,
0.2320556640625,
0.1685791015625
],
[
0.64892578125,
0.1514892578125,
0.78173828125,
0.1514892578125,
0.78173828125,
0.1622314453125,
0.64892578125,
0.1622314453125
],
[
0.491943359375,
0.1685791015625,
0.75732421875,
0.1666259765625,
0.7587890625,
0.2298583984375,
0.49267578125,
0.231689453125
],
[
0.238525390625,
0.184326171875,
0.294189453125,
0.184326171875,
0.294189453125,
0.1944580078125,
0.238525390625,
0.1944580078125
],
[
0.36767578125,
0.179931640625,
0.38818359375,
0.179931640625,
0.38818359375,
0.1925048828125,
0.36767578125,
0.1925048828125
],
[
0.293212890625,
0.1944580078125,
0.3095703125,
0.1988525390625,
0.306396484375,
0.2052001953125,
0.2900390625,
0.2008056640625
],
[
0.2484130859375,
0.216552734375,
0.293212890625,
0.216552734375,
0.293212890625,
0.2279052734375,
0.2484130859375,
0.2279052734375
],
[
0.346435546875,
0.216552734375,
0.359375,
0.216552734375,
0.359375,
0.2247314453125,
0.346435546875,
0.2247314453125
],
[
0.2197265625,
0.24560546875,
0.78515625,
0.24560546875,
0.78515625,
0.2587890625,
0.2197265625,
0.2587890625
],
[
0.2181396484375,
0.260009765625,
0.78515625,
0.260009765625,
0.78515625,
0.2734375,
0.2181396484375,
0.2734375
],
[
0.2181396484375,
0.27587890625,
0.78369140625,
0.27587890625,
0.78369140625,
0.2890625,
0.2181396484375,
0.2890625
],
[
0.221435546875,
0.290283203125,
0.78369140625,
0.290283203125,
0.78369140625,
0.3037109375,
0.221435546875,
0.3037109375
],
[
0.2181396484375,
0.30615234375,
0.7861328125,
0.30615234375,
0.7861328125,
0.3193359375,
0.2181396484375,
0.3193359375
],
[
0.2181396484375,
0.32080078125,
0.78369140625,
0.3212890625,
0.78369140625,
0.3359375,
0.2181396484375,
0.335205078125
],
[
0.221435546875,
0.37060546875,
0.78173828125,
0.37060546875,
0.78173828125,
0.380615234375,
0.221435546875,
0.380615234375
],
[
0.223876953125,
0.386474609375,
0.78173828125,
0.386474609375,
0.78173828125,
0.396484375,
0.223876953125,
0.396484375
],
[
0.221435546875,
0.40087890625,
0.78173828125,
0.40087890625,
0.78173828125,
0.410888671875,
0.221435546875,
0.410888671875
],
[
0.221435546875,
0.4140625,
0.78515625,
0.4140625,
0.78515625,
0.427978515625,
0.221435546875,
0.427978515625
],
[
0.2197265625,
0.429931640625,
0.609375,
0.429931640625,
0.609375,
0.443115234375,
0.2197265625,
0.443115234375
],
[
0.2442626953125,
0.44580078125,
0.78515625,
0.44580078125,
0.78515625,
0.458984375,
0.2442626953125,
0.458984375
],
[
0.221435546875,
0.462158203125,
0.78369140625,
0.462158203125,
0.78369140625,
0.472900390625,
0.221435546875,
0.472900390625
],
[
0.221435546875,
0.47607421875,
0.78515625,
0.47607421875,
0.78515625,
0.4892578125,
0.221435546875,
0.4892578125
],
[
0.22216796875,
0.492431640625,
0.78125,
0.492431640625,
0.78125,
0.50244140625,
0.22216796875,
0.50244140625
],
[
0.221435546875,
0.50830078125,
0.78369140625,
0.50830078125,
0.78369140625,
0.51806640625,
0.221435546875,
0.51806640625
],
[
0.2197265625,
0.521484375,
0.78515625,
0.521484375,
0.78515625,
0.53515625,
0.2197265625,
0.53515625
],
[
0.221435546875,
0.5361328125,
0.77685546875,
0.5361328125,
0.77685546875,
0.5498046875,
0.221435546875,
0.5498046875
],
[
0.2197265625,
0.5517578125,
0.78369140625,
0.5517578125,
0.78369140625,
0.56494140625,
0.2197265625,
0.56494140625
],
[
0.221435546875,
0.568359375,
0.463134765625,
0.568359375,
0.463134765625,
0.5791015625,
0.221435546875,
0.5791015625
],
[
0.2442626953125,
0.58349609375,
0.78515625,
0.58349609375,
0.78515625,
0.5966796875,
0.2442626953125,
0.5966796875
],
[
0.2197265625,
0.59765625,
0.78515625,
0.59765625,
0.78515625,
0.611328125,
0.2197265625,
0.611328125
],
[
0.221435546875,
0.6142578125,
0.78173828125,
0.6142578125,
0.78173828125,
0.625,
0.221435546875,
0.625
],
[
0.2197265625,
0.62744140625,
0.78515625,
0.62744140625,
0.78515625,
0.6416015625,
0.2197265625,
0.6416015625
],
[
0.2197265625,
0.64306640625,
0.498291015625,
0.64306640625,
0.498291015625,
0.6572265625,
0.2197265625,
0.6572265625
],
[
0.24267578125,
0.6591796875,
0.78515625,
0.65771484375,
0.78515625,
0.67236328125,
0.24267578125,
0.673828125
],
[
0.2181396484375,
0.6748046875,
0.78515625,
0.6748046875,
0.78515625,
0.68798828125,
0.2181396484375,
0.68798828125
],
[
0.221435546875,
0.69140625,
0.78369140625,
0.69140625,
0.78369140625,
0.7021484375,
0.221435546875,
0.7021484375
],
[
0.2197265625,
0.705078125,
0.78515625,
0.705078125,
0.78515625,
0.71826171875,
0.2197265625,
0.71826171875
],
[
0.22216796875,
0.7216796875,
0.78173828125,
0.7216796875,
0.78173828125,
0.73193359375,
0.22216796875,
0.73193359375
],
[
0.2181396484375,
0.73291015625,
0.78369140625,
0.734375,
0.78369140625,
0.74853515625,
0.2181396484375,
0.74755859375
],
[
0.2197265625,
0.75,
0.78515625,
0.75,
0.78515625,
0.76318359375,
0.2197265625,
0.76318359375
],
[
0.2197265625,
0.7666015625,
0.324462890625,
0.7666015625,
0.324462890625,
0.7763671875,
0.2197265625,
0.7763671875
],
[
0.24267578125,
0.77978515625,
0.78369140625,
0.77978515625,
0.78369140625,
0.79345703125,
0.24267578125,
0.79345703125
],
[
0.221435546875,
0.796875,
0.78369140625,
0.796875,
0.78369140625,
0.81005859375,
0.221435546875,
0.81005859375
],
[
0.221435546875,
0.81298828125,
0.78173828125,
0.81298828125,
0.78173828125,
0.82373046875,
0.221435546875,
0.82373046875
],
[
0.2197265625,
0.8271484375,
0.78369140625,
0.8271484375,
0.78369140625,
0.84033203125,
0.2197265625,
0.84033203125
]
] |
[
"2",
"J.Yuetal.",
"One-shotmodel",
"Progressive shrinking",
"Retrain and deploy",
"Single-stagemodel (ours",
"",
"Deploy instar",
"0",
"ning",
"Distilation",
"",
"Fig.l: Comparison with several existing workflows. We use nested squares to",
"denote models with shared weights, and use the size of the square to denote the",
"size of each model. Workflow in the middle refers the concurrent work from [5]",
"where submodels are sequentially induced through progressive distillation and",
"channel sorting. We simultaneously train all child models in a single-stage model",
"with proposed modifications, and deploy them without retraining or finetuning",
"从业者来说非常昂贵，但最近基于权重共享的高效NAS方法将搜索成本降低",
"了12，23，26，36个数量级。这些方法的工作原理是训练一个超级网络，然后识别",
"通过网络的路径-其操作的子集-在满足用户指定的特定硬件设备的延迟约束的后",
"时提供最佳的准确性。这种方法的优点是我们可以训练超级网络，然后使用它",
"从用户定义的搜索空间对许多不同的候选架构进行排名",
"然而，从这个超级网络获得的预测的绝对精度通常远低于以独立方式从头开",
"始训练的模型2。因此，通常假设需要对超级网络的权重进行大量的后处理",
"以获得高质量的模型部署精度。例如，一个建议的解决方案是为每个感兴趣的",
"设备和每个感兴趣的延迟预算重新训练一个单独的模型16.34。但是，这会产",
"生很大的开销，特别是在部署方案数量很大的情况下。第二种解决方案是在训",
"练完成后对权重进行后处理；例如，使用针对一次性网络[5]提出的渐进式收缩",
"启发式算法。然而，这个后处理步骤使模型训练管道变得复杂。此外，来自“",
"劳永逸”网络[5]的子模型仍然需要通过其他时期（e.9.，ImageNet上的75个",
"时期）进行微调，以实现最佳精度",
"在这项工作中，我们重新评估了普遍的信念，即为了获得有竞争力的准确性，",
"必须对共享权重进行重新训练或后处理。我们提出了几种技术来弥合具有共享",
"参数的小型和大型子模型之间不同的初始化和学习动态之间的差距。通过这些",
"技术，我们能够训练一个单阶段模型：一个单一的模型，我们可以直接切片高质",
"量的子模型，而无需任何额外的后处理",
"我们搜索一个大型单阶段模型，该模型包含小型子模型（???MFLOP，与",
"MobileNetV3相当）和大型子模型（???GFLOP，与EfficientNets相当）。与现",
"有的一次性方法F2.4，23，26，36不同，我们训练的单阶段模型提供了更广泛的",
"模型容量覆盖范围，更重要的是，所有子模型都经过了训练，以便在搜索阶段结",
"束时同时达到出色的性能。然后可以通过简单的从粗到精的选择策略进行架构",
"选择。一旦选择了架构，我们可以通过简单地将单阶段模型切片以进行即时部",
"署来获得子模型W。I。t。给定的约束，如内存占用和/或运行时延迟。工作流",
"程如图1所示",
"简化的BigNAS工作流的成功依赖于一个目标：如何训练高质量的单阶段模",
"型？该目标本身具有挑战性。例如，我们发现，如果一个大的单阶段模型没有正",
"确初始化，训练损失就会爆炸；在训练过程中，大的孩子模型在小的孩子模型之",
"前就开始过度拟合；从经验上看，更大的孩子模型往往在训练数据上过度拟合"
] |
{
"arxiv_id": "2003.11142",
"language": "zh",
"page_idx": 1
}
|
||
2003.11142-zh-2
|
2003.11142-zh-2.jpg
|
BigNAS: Neural Architecture Search with Big Single-Stage Models 3
得更多。为了应对这些挑战，我们系统地研究和重新审视独立网络的传统训练技术，并将其应用于训练体重共享单阶段模型。通过提出的技术，我们能够在ImageNet 上训练高质量的单阶段模型 , 并获得一系列子模型，这些子模型同时超过 200 至 1000MFLOP 范围内的所有最新模型，包括 EfficientNetsB0-B2(1。$6\%$在 400MFLOP 下更准确), 无需在搜索完成后重新训练或微调子模型。例如，我们的一个孩子模型达到了 80。9%在 1GFLOP 时的 top-1 精度 (4× 的计算量比 ResNet-50 少)
## 2 Related Work
早期的 NAS 方法 [21,22,28,39,40] 从头开始(在较小的代理任务上)训练数千个候选体系结构，并将其验证性能用作对算法的反馈，该算法学会专注于搜索空间中最有希望的区域。最近的工作试图通过训练一个单一的过度参数化的一次性模型来摊销成本。搜索空间中的每个体系结构仅使用一次性模型中的操作的子集；通过使用共享权重来估计其相对精度，可以有效地对这些子模型进行排名$\mathbb{E}[2,4,6,19,23,26,34-36]$
作为补充方向，提出了资源感知的 NAS 方法，以同时最大化预测精度并最
小化资源需求，例如延迟，FLOP 或内存占用$\mathbb{E}[3,5,10,31,32,34,36]$
所有上述方法都需要两阶段训练：一旦确定了最佳架构(通过代理任务或使用一次性模型),就必须从头开始重新训练以获得具有更高精度的最终模型。在大多数现有的工作中，单个搜索实验一次仅针对单个资源预算或狭窄范围的资源预算
为了缓解这些问题，[5]在我们的工作中同时提出了渐进式训练方法 (OFA), 这个想法是预先训练一个完整的网络，然后逐步提炼它以获得较小的网络。此外，需要信道排序过程来逐步构建较小的网络。然而，在我们提出的 BigNAS 中，单阶段模型中的所有子模型都是同时训练的，允许小型和大型网络的学习相互受益。在训练过程中，我们始终为子模型在每个层中保留较低索引的通道在每个阶段中保留较低索引的层，从而消除了排序过程。我们的 BigNAS 能够处理更广泛的模型(从 200个 MFLOP 到 1个 GFLOP), 并更好地覆盖各种部署方案和各种资源预算
我们的工作与可纤薄网络$\mathbb{E}[36-38]$ 在训练可用于许多子模型的单个共享权重集方面具有高级相似性。但是，尽管纤薄的网络专门用于仅改变通道的数量，但我们能够处理更大的空间，同时搜索许多体系结构维度(内核和通道大小，网络深度，输人分辨率),从而包含并优于手动设计的扩展启发式在EfficientNets [ 33] 中
3 Architecture Search with Single-Stage Models
我们提出的方法包括两个步骤：
1. We train a big single-stage model from which we can directly sample or slice
different architectures as child models for instant inference and deployment. In contrast to previous works, our training is single-stage. In other words: the trained model weights from a search can be directly used for deployment, without any need to retrain them from scratch $(e.g.[2,4,10,23,26,31,36]$ ) or otherwise post-process them $(~e.g.,[5]~).$
|
[
[
0.2900390625,
0.11614990234375,
0.7890625,
0.11614990234375,
0.7890625,
0.1300048828125,
0.2900390625,
0.1300048828125
],
[
0.2197265625,
0.1483154296875,
0.78515625,
0.1483154296875,
0.78515625,
0.1622314453125,
0.2197265625,
0.1622314453125
],
[
0.221435546875,
0.1654052734375,
0.78369140625,
0.1654052734375,
0.78369140625,
0.175537109375,
0.221435546875,
0.175537109375
],
[
0.221435546875,
0.179931640625,
0.78173828125,
0.179931640625,
0.78173828125,
0.1900634765625,
0.221435546875,
0.1900634765625
],
[
0.221435546875,
0.1932373046875,
0.78369140625,
0.1932373046875,
0.78369140625,
0.20703125,
0.221435546875,
0.20703125
],
[
0.2197265625,
0.208984375,
0.78515625,
0.208984375,
0.78515625,
0.222900390625,
0.2197265625,
0.222900390625
],
[
0.2197265625,
0.2235107421875,
0.78369140625,
0.2235107421875,
0.78369140625,
0.2374267578125,
0.2197265625,
0.2374267578125
],
[
0.2181396484375,
0.2392578125,
0.378173828125,
0.2392578125,
0.378173828125,
0.25244140625,
0.2181396484375,
0.25244140625
],
[
0.21728515625,
0.274658203125,
0.39208984375,
0.274658203125,
0.39208984375,
0.2890625,
0.21728515625,
0.2890625
],
[
0.221435546875,
0.304931640625,
0.78173828125,
0.304931640625,
0.78173828125,
0.31494140625,
0.221435546875,
0.31494140625
],
[
0.2197265625,
0.318115234375,
0.78515625,
0.318115234375,
0.78515625,
0.33203125,
0.2197265625,
0.33203125
],
[
0.2197265625,
0.333984375,
0.77392578125,
0.333984375,
0.77392578125,
0.347900390625,
0.2197265625,
0.347900390625
],
[
0.2197265625,
0.348388671875,
0.78515625,
0.348388671875,
0.78515625,
0.3623046875,
0.2197265625,
0.3623046875
],
[
0.2197265625,
0.363037109375,
0.78369140625,
0.363037109375,
0.78369140625,
0.376953125,
0.2197265625,
0.376953125
],
[
0.2197265625,
0.37890625,
0.4306640625,
0.37890625,
0.4306640625,
0.392578125,
0.2197265625,
0.392578125
],
[
0.2442626953125,
0.393310546875,
0.78515625,
0.393310546875,
0.78515625,
0.4072265625,
0.2442626953125,
0.4072265625
],
[
0.2181396484375,
0.4091796875,
0.70751953125,
0.4091796875,
0.70751953125,
0.42236328125,
0.2181396484375,
0.42236328125
],
[
0.24267578125,
0.4248046875,
0.78515625,
0.4248046875,
0.78515625,
0.438232421875,
0.24267578125,
0.438232421875
],
[
0.221435546875,
0.439453125,
0.78515625,
0.439453125,
0.78515625,
0.45263671875,
0.221435546875,
0.45263671875
],
[
0.2181396484375,
0.45263671875,
0.7861328125,
0.453857421875,
0.7861328125,
0.468505859375,
0.2181396484375,
0.46728515625
],
[
0.2197265625,
0.47021484375,
0.26708984375,
0.47021484375,
0.26708984375,
0.481689453125,
0.2197265625,
0.481689453125
],
[
0.2442626953125,
0.48486328125,
0.7861328125,
0.48486328125,
0.7861328125,
0.498779296875,
0.2442626953125,
0.498779296875
],
[
0.2197265625,
0.5,
0.78515625,
0.5,
0.78515625,
0.51318359375,
0.2197265625,
0.51318359375
],
[
0.221435546875,
0.5166015625,
0.78369140625,
0.5166015625,
0.78369140625,
0.5263671875,
0.221435546875,
0.5263671875
],
[
0.2197265625,
0.52880859375,
0.78369140625,
0.52978515625,
0.78369140625,
0.54345703125,
0.2197265625,
0.54248046875
],
[
0.2197265625,
0.54443359375,
0.78369140625,
0.54443359375,
0.78369140625,
0.55810546875,
0.2197265625,
0.55810546875
],
[
0.2197265625,
0.56005859375,
0.78369140625,
0.56005859375,
0.78369140625,
0.57373046875,
0.2197265625,
0.57373046875
],
[
0.221435546875,
0.57470703125,
0.7861328125,
0.57470703125,
0.7861328125,
0.58837890625,
0.221435546875,
0.58837890625
],
[
0.2197265625,
0.59033203125,
0.3798828125,
0.59033203125,
0.3798828125,
0.60107421875,
0.2197265625,
0.60107421875
],
[
0.2418212890625,
0.60400390625,
0.78369140625,
0.60400390625,
0.78369140625,
0.6171875,
0.2418212890625,
0.6171875
],
[
0.2181396484375,
0.619140625,
0.78173828125,
0.619140625,
0.78173828125,
0.63330078125,
0.2181396484375,
0.63330078125
],
[
0.2197265625,
0.63720703125,
0.78173828125,
0.63720703125,
0.78173828125,
0.64794921875,
0.2197265625,
0.64794921875
],
[
0.2197265625,
0.65087890625,
0.78369140625,
0.65087890625,
0.78369140625,
0.6640625,
0.2197265625,
0.6640625
],
[
0.2181396484375,
0.66552734375,
0.363525390625,
0.66552734375,
0.363525390625,
0.6787109375,
0.2181396484375,
0.6787109375
],
[
0.21728515625,
0.7021484375,
0.69921875,
0.7021484375,
0.69921875,
0.71630859375,
0.21728515625,
0.71630859375
],
[
0.221435546875,
0.73193359375,
0.435546875,
0.73193359375,
0.435546875,
0.7421875,
0.221435546875,
0.7421875
],
[
0.223876953125,
0.75,
0.78173828125,
0.75,
0.78173828125,
0.76318359375,
0.223876953125,
0.76318359375
],
[
0.2467041015625,
0.7666015625,
0.78173828125,
0.7666015625,
0.78173828125,
0.77978515625,
0.2467041015625,
0.77978515625
],
[
0.2467041015625,
0.7822265625,
0.78515625,
0.7822265625,
0.78515625,
0.79541015625,
0.2467041015625,
0.79541015625
],
[
0.2459716796875,
0.79541015625,
0.78173828125,
0.796875,
0.78173828125,
0.81103515625,
0.2459716796875,
0.81005859375
],
[
0.2484130859375,
0.8125,
0.78515625,
0.8125,
0.78515625,
0.82568359375,
0.2484130859375,
0.82568359375
],
[
0.2467041015625,
0.82763671875,
0.5537109375,
0.82763671875,
0.5537109375,
0.84130859375,
0.2467041015625,
0.84130859375
]
] |
[
"BigNAS: Neural Architecture Search with Big Single-Stage Models3",
"得更多。为了应对这些挑战，我们系统地研究和重新审视独立网络的传统训练",
"技术，并将其应用于训练体重共享单阶段模型。通过提出的技术：我们能够在",
"ImageNet上训练高质量的单阶段模型，并获得一系列子模型，这些子模型同时",
"超过200至1000MFLOP范围内的所有最新模型，包括EficientNetsB0-B2(1.",
"6%在400MFLOP下更准确)，无需在搜索完成后重新训练或微调子模型。例",
"如，我们的一个孩子模型达到了80。9%在1GFLOP时的top-1精度（4×的计",
"算量比ResNet-50少）",
"2Related Work",
"早期的NAS方法121.22.28.39.40从头开始（在较小的代理任务上）训练数于",
"个候选体系结构，并将其验证性能用作对算法的反馈，该算法学会专注于搜索",
"空间中最有希望的区域。最近的工作试图通过训练一个单一的过度参数化的",
"次性模型来摊销成本。搜索空间中的每个体系结构仅使用一次性模型中的操作",
"的子集；通过使用共享权重来估计其相对精度，可以有效地对这些子模型进行",
"排名E[2,4,6,19,23,26,34-36]",
"作为补充方向，提出了资源感知的NAS方法，以同时最大化预测精度并最",
"小化资源需求，例如延迟，FLOP或内存占用E[3,5,10,31,32,34,36]",
"所有上述方法都需要两阶段训练：一旦确定了最佳架构（通过代理任务或使",
"用一次性模型），就必须从头开始重新训练以获得具有更高精度的最终模型。在",
"大多数现有的工作中，单个搜索实验一次仅针对单个资源预算或狭窄范围的资",
"源预算",
"为了缓解这些问题，[5]在我们的工作中同时提出了渐进式训练方法（OFA)。",
"这个想法是预先训练一个完整的网络，然后逐步提炼它以获得较小的网络。此",
"外，需要信道排序过程来逐步构建较小的网络。然而，在我们提出的BigNAS",
"中，单阶段模型中的所有子模型都是同时训练的，允许小型和大型网络的学习",
"相互受益。在训练过程中，我们始终为子模型在每个层中保留较低索引的通道",
"在每个阶段中保留较低索引的层，从而消除了排序过程。我们的BigNAS能够",
"处理更广泛的模型（从200个MFLOP到1个GFLOP），并更好地覆盖各种部",
"署方案和各种资源预算",
"我们的工作与可纤薄网络E[36-38]在训练可用于许多子模型的单个共享",
"权重集方面具有高级相似性。但是，尽管纤薄的网络专门用于仅改变通道的",
"数量，但我们能够处理更天的空间，同时搜索许多体系结构维度（内核和通",
"道大小，网络深度，输入分辨率），从而包含并优于手动设计的扩展启发式在",
"EfficientNets [33]中",
"3Architecture Search with Single-Stage Models",
"我们提出的方法包括两个步骤：",
"1. We train a big single-stage model from which we can directly sample or slice",
"different architectures as child models for instant inference and deployment",
"In contrast to previous works, our training is single-stage. In other words:",
"the trained model weights from a search can be directly used for deployment",
"without any need to retrain them from scratch (e.g.[2,4,10,23,26,31,36])",
"or otherwise post-process them (e.g., [5] )."
] |
{
"arxiv_id": "2003.11142",
"language": "zh",
"page_idx": 2
}
|
||
2003.11142-zh-3
|
2003.11142-zh-3.jpg
|
4 J. Yu et al.
2. We select architectures using a simple coarse-to-fine selection method to find
the most accurate model under the given resource constraints $(e.g.$,FLOPs,
memory footprint and/or runtime latency budgets on different devices).
下面，我们将首先从五个方面系统地研究如何训练高质量的单阶段模型：训练时的网络采样，就地蒸馏，网络初始化，收敛行为和正则化。然后，我们将提出一种从粗到细的方法来进行有效的资源感知架构选择
3.1 Training a High-Quality Single-Stage Model
由于小型和大型儿童模型的独特初始化和学习动态，训练高质量的单阶段模型非常重要且非常重要。在本节中，我们首先概括了最初由$|\mathbb{E}[37]$ 引入的两种技术，以同时训练一组具有不同通道号的高质量网络，并表明两者都可以扩展以处理更大的空间，在该空间中，可以共同搜索体系结构维度，包括内核大小，通道号，输人分辨率，网络深度。然后，我们提出了三种额外的技术来解决小型和大型儿童模型的不同初始化和学习动态
三明治规则。在每个训练步骤中，给定一小批数据，夹心规则 [37] 对最小子模型、最大 (全) 子模型和$N$随机抽样子模型 (我们实验中的$N=2$ )进行抽样。然后，在更新单阶段模型的权重之前，聚合所有采样子模型的梯度。由于我们的搜索空间中包含多个体系结构维度，因此“最小”子模型是具有最低输入分辨率，最薄宽度，最浅深度和最小内核大小的子模型(每个反向残差块$\mathbb{E}[29]$ 中的深度卷积的内核)。动机是通过在所有子模型中同时提高性能下限(最小的子模型)和性能上限(最大的子模型)来同时改善我们搜索空间中的所有子模型
就地蒸馏。在单阶段模型的训练过程中，就地蒸馏[37]采用最大可能的子模型 (完整模型) 预测的软标签来监督所有其他子模型。在我们的训练环境中， 就地蒸馏的好处是免费的，因为由于三明治规则，我们总是可以在每个梯度更新步骤中访问最大子模型的预测。Wenotethatallchildmodelsareonlytrainedwiththeinplacedivisionloss,startingfromthefirsttrainingsteptotheendofthetraining。为简单起见，我们的实验中未使用温度超参数或蒸馏/目标损失[E[15] 的混合物
在训练期间，输入图像被随机裁剪为初步数据增强步骤。当将高分辨率教师模型提炼为低分辨率学生模型时，我们发现将相同的图像补丁提供给教师和学生是有帮助的。在我们的数据准备中，我们首先随机裁剪具有固定分辨率的图像(在 ImageNet 上我们使用 224),然后将双三次插值应用于同一补丁，以将其转换为所有目标分辨率$(e.g.$、192、288、320 等。)。在这种情况下，最大的儿童模型(教师)预测的软标签与其他儿童模型(学生)看到的输入更兼容。因此，这可以作为一个更准确的蒸馏信号。我们的初步结果表明，与采样不同的补丁相比，这导致$\sim0.3\%$ 改善了儿童模型的平均 top-1 准确性
初始化。当我们第一次尝试训练更大更深的单阶段模型时，我们发现训练非常不稳定，并且当我们使用为训练正常神经网络而优化的学习率时，训练损失激增。当我们将学习率降低到其原始值的 30% 时，训练开始工作，但是这种配置导致更差的结果$(\sim1.0\%$ top-1 精度在 ImageNet 上下降)
虽然稳定模型训练通常是一个复杂且开放式的问题，但我们发现在这种情况下，对我们的设置进行简单的更改就足以稳定训练。由于我们搜索空间中的所有子模型都是残差网络，因此我们通过在每个残差块的最后一个BatchNormalization$\underline{\mathrm{E}}[20]$层中设置可学习的缩放系数$\gamma=0$ ,将每个残差块的输出 (在跳过连接之前) 初始化为全零张量，以确保每个残差块前后的方差相
|
[
[
0.21728515625,
0.117431640625,
0.268798828125,
0.117431640625,
0.268798828125,
0.1287841796875,
0.21728515625,
0.1287841796875
],
[
0.271240234375,
0.117431640625,
0.346435546875,
0.117431640625,
0.346435546875,
0.1287841796875,
0.271240234375,
0.1287841796875
],
[
0.223876953125,
0.1483154296875,
0.7861328125,
0.1483154296875,
0.7861328125,
0.1622314453125,
0.223876953125,
0.1622314453125
],
[
0.2467041015625,
0.1641845703125,
0.78369140625,
0.1641845703125,
0.78369140625,
0.177978515625,
0.2467041015625,
0.177978515625
],
[
0.2467041015625,
0.1787109375,
0.75732421875,
0.1787109375,
0.75732421875,
0.1925048828125,
0.2467041015625,
0.1925048828125
],
[
0.2467041015625,
0.205810546875,
0.78369140625,
0.205810546875,
0.78369140625,
0.216552734375,
0.2467041015625,
0.216552734375
],
[
0.2197265625,
0.2183837890625,
0.78173828125,
0.2197265625,
0.78173828125,
0.23291015625,
0.2197265625,
0.231689453125
],
[
0.2197265625,
0.2342529296875,
0.609375,
0.2342529296875,
0.609375,
0.2474365234375,
0.2197265625,
0.2474365234375
],
[
0.2181396484375,
0.272705078125,
0.62841796875,
0.272705078125,
0.62841796875,
0.285888671875,
0.2181396484375,
0.285888671875
],
[
0.2197265625,
0.29736328125,
0.78515625,
0.29736328125,
0.78515625,
0.311279296875,
0.2197265625,
0.311279296875
],
[
0.2197265625,
0.313232421875,
0.78369140625,
0.313232421875,
0.78369140625,
0.326904296875,
0.2197265625,
0.326904296875
],
[
0.2197265625,
0.32763671875,
0.78369140625,
0.32763671875,
0.78369140625,
0.341552734375,
0.2197265625,
0.341552734375
],
[
0.221435546875,
0.343505859375,
0.78369140625,
0.343505859375,
0.78369140625,
0.356689453125,
0.221435546875,
0.356689453125
],
[
0.22216796875,
0.35986328125,
0.78369140625,
0.35986328125,
0.78369140625,
0.37060546875,
0.22216796875,
0.37060546875
],
[
0.2197265625,
0.37255859375,
0.495849609375,
0.373779296875,
0.495849609375,
0.386962890625,
0.2197265625,
0.386474609375
],
[
0.24267578125,
0.386962890625,
0.78515625,
0.38818359375,
0.78515625,
0.40283203125,
0.24267578125,
0.401611328125
],
[
0.2197265625,
0.404052734375,
0.78369140625,
0.404052734375,
0.78369140625,
0.417236328125,
0.2197265625,
0.417236328125
],
[
0.221435546875,
0.42041015625,
0.78173828125,
0.42041015625,
0.78173828125,
0.43115234375,
0.221435546875,
0.43115234375
],
[
0.221435546875,
0.434326171875,
0.78369140625,
0.434326171875,
0.78369140625,
0.447509765625,
0.221435546875,
0.447509765625
],
[
0.221435546875,
0.45068359375,
0.78173828125,
0.45068359375,
0.78173828125,
0.4609375,
0.221435546875,
0.4609375
],
[
0.221435546875,
0.464111328125,
0.78515625,
0.464111328125,
0.78515625,
0.477783203125,
0.221435546875,
0.477783203125
],
[
0.2197265625,
0.479248046875,
0.77294921875,
0.479248046875,
0.77294921875,
0.492431640625,
0.2197265625,
0.492431640625
],
[
0.2459716796875,
0.49560546875,
0.78173828125,
0.49560546875,
0.78173828125,
0.505859375,
0.2459716796875,
0.505859375
],
[
0.2197265625,
0.5087890625,
0.78369140625,
0.5087890625,
0.78369140625,
0.52294921875,
0.2197265625,
0.52294921875
],
[
0.2197265625,
0.5244140625,
0.78369140625,
0.5244140625,
0.78369140625,
0.53857421875,
0.2197265625,
0.53857421875
],
[
0.2197265625,
0.5390625,
0.78369140625,
0.5390625,
0.78369140625,
0.55322265625,
0.2197265625,
0.55322265625
],
[
0.2197265625,
0.5546875,
0.78515625,
0.5546875,
0.78515625,
0.568359375,
0.2197265625,
0.568359375
],
[
0.2197265625,
0.5693359375,
0.76123046875,
0.5693359375,
0.76123046875,
0.58349609375,
0.2197265625,
0.58349609375
],
[
0.2442626953125,
0.58544921875,
0.78515625,
0.58544921875,
0.78515625,
0.5986328125,
0.2442626953125,
0.5986328125
],
[
0.2197265625,
0.60107421875,
0.78369140625,
0.60107421875,
0.78369140625,
0.6142578125,
0.2197265625,
0.6142578125
],
[
0.2181396484375,
0.6142578125,
0.78369140625,
0.61572265625,
0.78369140625,
0.62890625,
0.2181396484375,
0.62744140625
],
[
0.2181396484375,
0.62890625,
0.78515625,
0.6298828125,
0.78515625,
0.64453125,
0.2181396484375,
0.64306640625
],
[
0.2197265625,
0.64599609375,
0.78369140625,
0.64599609375,
0.78369140625,
0.6591796875,
0.2197265625,
0.6591796875
],
[
0.22216796875,
0.662109375,
0.78369140625,
0.662109375,
0.78369140625,
0.67236328125,
0.22216796875,
0.67236328125
],
[
0.2197265625,
0.67529296875,
0.78515625,
0.67529296875,
0.78515625,
0.689453125,
0.2197265625,
0.689453125
],
[
0.221435546875,
0.68994140625,
0.65869140625,
0.68994140625,
0.65869140625,
0.7041015625,
0.221435546875,
0.7041015625
],
[
0.2442626953125,
0.70556640625,
0.78515625,
0.70556640625,
0.78515625,
0.7197265625,
0.2442626953125,
0.7197265625
],
[
0.2197265625,
0.72021484375,
0.78515625,
0.72021484375,
0.78515625,
0.734375,
0.2197265625,
0.734375
],
[
0.221435546875,
0.7373046875,
0.78369140625,
0.7373046875,
0.78369140625,
0.74755859375,
0.221435546875,
0.74755859375
],
[
0.2197265625,
0.751953125,
0.65185546875,
0.751953125,
0.65185546875,
0.76513671875,
0.2197265625,
0.76513671875
],
[
0.2459716796875,
0.7666015625,
0.78515625,
0.7666015625,
0.78515625,
0.77978515625,
0.2459716796875,
0.77978515625
],
[
0.221435546875,
0.78271484375,
0.78369140625,
0.78271484375,
0.78369140625,
0.79345703125,
0.221435546875,
0.79345703125
],
[
0.2197265625,
0.796875,
0.78173828125,
0.796875,
0.78173828125,
0.81005859375,
0.2197265625,
0.81005859375
],
[
0.2197265625,
0.81103515625,
0.78369140625,
0.81103515625,
0.78369140625,
0.82470703125,
0.2197265625,
0.82470703125
],
[
0.2181396484375,
0.8271484375,
0.78173828125,
0.8271484375,
0.78173828125,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"4",
"J.Yu etal.",
"2. We select architectures using a simple coarse-to-fine selection method to find",
"the most accurate model under the given resource constraints ( e.g., FLOPs,",
"memory footprint and/or runtime latency budgets on different devices)",
"下面，我们将首先从五个方面系统地研究如何训练高质量的单阶段模型：训",
"练时的网络采样，就地蒸馏，网络初始化，收敛行为和正则化。然后，我们将提",
"出一种从粗到细的方法来进行有效的资源感知架构选择",
"3.1:Training a High-Quality Single-Stage Model",
"由于小型和大型儿童模型的独特初始化和学习动态，训练高质量的单阶段模型",
"非常重要且非常重要。在本节中，我们首先概括了最初由E[37]引入的两种技",
"术，以同时训练一组具有不同通道号的高质量网络，并表明两者都可以扩展以",
"处理更大的空间，在该空间中，可以共同搜索体系结构维度，包括内核大小，通",
"道号，输人分辨率，网络深度。然后，我们提出了三种额外的技术来解决小型和",
"大型儿童模型的不同初始化和学习动态",
"三明治规则。在每个训练步骤中，给定一小批数据，夹心规则[37]对最小子",
"模型、最大（全）子模型和N随机抽样子模型（我们实验中的N=2）进行抽",
"样。然后，在更新单阶段模型的权重之前，聚合所有采样子模型的梯度。由于我",
"们的搜索空间中包含多个体系结构维度，因此“最小”子模型是具有最低输入分",
"辨率，最薄宽度，最浅深度和最小内核大小的子模型（每个反向残差块E29中",
"的深度卷积的内核）。动机是通过在所有子模型中同时提高性能下限（最小的子",
"模型）和性能上限（最大的子模型）来同时改善我们搜索空间中的所有子模型",
"就地蒸馏。在单阶段模型的训练过程中，就地蒸留37采用最大可能的子",
"模型（完整模型）预测的软标签来监督所有其他子模型。在我们的训练环境中",
"就地蒸馏的好处是免费的，因为由于三明治规则，我们总是可以在每个梯度更",
"新步骤中访问最大子模型的预测。Wenotethatallchildmodelsareonlytrainedwith-",
"theinplacedivisionloss,startingfromthefirsttrainingsteptotheendofthetraining。为",
"简单起见，我们的实验中未使用温度超参数或蒸馏/目标损失E[15]的混合物",
"在训练期间，输入图像被随机裁剪为初步数据增强步骤。当将高分辨率教师",
"模型提炼为低分辨率学生模型时，我们发现将相同的图像补丁提供给教师和学",
"生是有帮助的。在我们的数据准备中，我们首先随机裁剪具有固定分辨率的图",
"像（在ImageNet上我们使用224），然后将双三次插值应用于同一补丁，以将",
"其转换为所有目标分辨率（e.9.、192、288、320等。）。在这种情况下，最大的",
"儿童模型（教师）预测的软标签与其他儿童模型（学生）看到的输人更兼容。因",
"此，这可以作为一个更准确的蒸馏信号。我们的初步结果表明，与采样不同的补",
"丁相比，这导致～0.3%改善了儿童模型的平均top-1准确性",
"初始化。当我们第一次尝试训练更大更深的单阶段模型时，我们发现训练非",
"常不稳定，并且当我们使用为训练正常神经网络而优化的学习率时，训练损失",
"激增。当我们将学习率降低到其原始值的30%时：训练开始工作，但是这种配",
"置导致更差的结果（~1.0%top-1精度在ImageNet上下降）",
"虽然稳定模型训练通常是一个复杂且开放式的问题，但我们发现在这种",
"情况下，对我们的设置进行简单的更改就足以稳定训练。由于我们搜索空",
"间中的所有子模型都是残差网络，因此我们通过在每个残差块的最后一个",
"BatchNormalizationE[20]层中设置可学习的缩放系数=0，将每个残差块的",
"输出（在跳过连接之前）初始化为全零张量，以确保每个残差块前后的方差相"
] |
{
"arxiv_id": "2003.11142",
"language": "zh",
"page_idx": 3
}
|
||
2003.11142-zh-4
|
2003.11142-zh-4.jpg
|
BigNAS: Neural Architecture Search with Big Single-Stage Models 5
同，而与扇人无关。此初始化最初在[9]中提到，并通过$\sim0.2\%$在其设置中提高了准确性，但在我们的设置中更为关键(通过$\sim1.0\%$ 进行改进)。当分辨率或通道不同时，我们还在每个阶段过渡中添加了一个跳过连接(如果需要，使用$2\times2$平均池化和/或$1\times1$卷积)来显式构建标识映射 [13]
收敛行为。在实践中，我们发现大的孩子模型收敛速度更快，而小的孩子模型收敛速度更慢。图 2a 显示了单阶段模型训练过程中的典型学习曲线，其中我们绘制了小型和大型子模型随时间变化的验证精度。这揭示了一个困境：在训练步骤$t$ ,当大孩子模型的性能达到峰值时，小孩子模型没有得到充分训练； 在训练步骤$t^{\prime}$,当小孩子模型具有更好的性能时，大孩子模型已经过拟合
<FigureHere>
Fig.2: On the left, we show typical accuracy curves during the training process for both small and big child models. It reveals a common dilemma in training big single-stage models: at training step $t$ when the performance of big child models peaks, the small child models are not fully-trained; and at training step $t^\prime$ when the small child models have better performance, the big child models already overfitted. On the right, we plot the simple modified learning rate schedules with constant ending to address this issue.
为了解决这个问题，我们将重点放在学习率时间表上。我们首先在图 2b 中绘制了 MobileNet 系列[16,17,29] 、MNasNets [32]和 EfficientNets [33]的优化和广泛使用的指数衰减学习速率调度。我们介绍了一个简单的修改这个学习率的时间表，命名为指数衰减与恒定的结束，它有一个恒定的学习率在训练结束时，当它达到 5% 的初始学习率 (图 2b )。它带来两个好处。首先，在最后的学习率略大的情况下，小型儿童模型学习速度更快。其次，随着权重的振荡， 末端的恒定学习率缓解了大子模型的过拟合
正则化。根据经验，在比较训练/验证损失时，我们发现大的子模型倾向于过度拟合训练数据，而小的子模型倾向于不拟合。在以前的工作中，Bender 等人。[2] 将相同的权重衰减应用于所有子模型，无论它们是小还是大。为了防止较大网络的过度拟合，对于 EfficientNets, Tan 等人。[33] 发现对较大的神经网络使用较大的丢失率$\mathbb{E}[30]$ 率很有帮助。在我们训练大型单阶段模型的背景下， 由于具有共享参数的小型儿童模型和大型儿童模型之间的相互作用，这变得更
|
[
[
0.289306640625,
0.11553955078125,
0.734375,
0.11553955078125,
0.734375,
0.1287841796875,
0.289306640625,
0.1287841796875
],
[
0.77392578125,
0.11871337890625,
0.7861328125,
0.11871337890625,
0.7861328125,
0.126953125,
0.77392578125,
0.126953125
],
[
0.2181396484375,
0.147705078125,
0.78515625,
0.146484375,
0.78515625,
0.1610107421875,
0.2181396484375,
0.1622314453125
],
[
0.2197265625,
0.1634521484375,
0.78515625,
0.1634521484375,
0.78515625,
0.1767578125,
0.2197265625,
0.1767578125
],
[
0.2197265625,
0.1767578125,
0.78369140625,
0.1767578125,
0.78369140625,
0.1900634765625,
0.2197265625,
0.1900634765625
],
[
0.2181396484375,
0.1932373046875,
0.646484375,
0.1932373046875,
0.646484375,
0.20703125,
0.2181396484375,
0.20703125
],
[
0.2442626953125,
0.210205078125,
0.78515625,
0.210205078125,
0.78515625,
0.2235107421875,
0.2442626953125,
0.2235107421875
],
[
0.2197265625,
0.2247314453125,
0.78515625,
0.2247314453125,
0.78515625,
0.238037109375,
0.2197265625,
0.238037109375
],
[
0.221435546875,
0.2412109375,
0.78369140625,
0.2412109375,
0.78369140625,
0.251953125,
0.221435546875,
0.251953125
],
[
0.221435546875,
0.255126953125,
0.78759765625,
0.255126953125,
0.78759765625,
0.268310546875,
0.221435546875,
0.268310546875
],
[
0.2181396484375,
0.268310546875,
0.74609375,
0.26953125,
0.74609375,
0.2841796875,
0.2181396484375,
0.28271484375
],
[
0.285888671875,
0.335205078125,
0.334228515625,
0.335205078125,
0.334228515625,
0.34228515625,
0.285888671875,
0.34228515625
],
[
0.28759765625,
0.34228515625,
0.31787109375,
0.34228515625,
0.31787109375,
0.3466796875,
0.28759765625,
0.3466796875
],
[
0.318603515625,
0.34228515625,
0.330078125,
0.34521484375,
0.326904296875,
0.348388671875,
0.31787109375,
0.34521484375
],
[
0.52392578125,
0.352783203125,
0.53125,
0.352783203125,
0.53125,
0.359130859375,
0.52392578125,
0.359130859375
],
[
0.24267578125,
0.3447265625,
0.2607421875,
0.3447265625,
0.2607421875,
0.35986328125,
0.24267578125,
0.35986328125
],
[
0.2303466796875,
0.35791015625,
0.2646484375,
0.35791015625,
0.2646484375,
0.43115234375,
0.2303466796875,
0.43115234375
],
[
0.378173828125,
0.386474609375,
0.386474609375,
0.386474609375,
0.386474609375,
0.389404296875,
0.378173828125,
0.389404296875
],
[
0.2459716796875,
0.4267578125,
0.25732421875,
0.4267578125,
0.25732421875,
0.438232421875,
0.2459716796875,
0.438232421875
],
[
0.2467041015625,
0.44140625,
0.253173828125,
0.439453125,
0.25732421875,
0.4462890625,
0.250732421875,
0.448974609375
],
[
0.289306640625,
0.455078125,
0.306396484375,
0.455078125,
0.306396484375,
0.462158203125,
0.289306640625,
0.462158203125
],
[
0.33251953125,
0.455078125,
0.400390625,
0.455078125,
0.400390625,
0.4697265625,
0.33251953125,
0.4697265625
],
[
0.56396484375,
0.455078125,
0.583984375,
0.455078125,
0.583984375,
0.462158203125,
0.56396484375,
0.462158203125
],
[
0.611328125,
0.453857421875,
0.689453125,
0.453857421875,
0.689453125,
0.468505859375,
0.611328125,
0.468505859375
],
[
0.69921875,
0.455078125,
0.72802734375,
0.455078125,
0.72802734375,
0.462158203125,
0.69921875,
0.462158203125
],
[
0.455078125,
0.45703125,
0.48193359375,
0.45703125,
0.48193359375,
0.4609375,
0.455078125,
0.4609375
],
[
0.42724609375,
0.45703125,
0.450927734375,
0.45703125,
0.450927734375,
0.4609375,
0.42724609375,
0.4609375
],
[
0.60107421875,
0.45703125,
0.61279296875,
0.455078125,
0.615234375,
0.460205078125,
0.60302734375,
0.462158203125
],
[
0.73193359375,
0.45703125,
0.75732421875,
0.45703125,
0.75732421875,
0.4609375,
0.73193359375,
0.4609375
],
[
0.35302734375,
0.477783203125,
0.375732421875,
0.477783203125,
0.375732421875,
0.492431640625,
0.35302734375,
0.492431640625
],
[
0.62744140625,
0.4765625,
0.654296875,
0.4765625,
0.654296875,
0.49365234375,
0.62744140625,
0.49365234375
],
[
0.2197265625,
0.5048828125,
0.78515625,
0.5048828125,
0.78515625,
0.51953125,
0.2197265625,
0.51953125
],
[
0.21728515625,
0.51806640625,
0.7861328125,
0.51953125,
0.7861328125,
0.5341796875,
0.21728515625,
0.53271484375
],
[
0.21728515625,
0.5341796875,
0.78369140625,
0.5341796875,
0.78369140625,
0.54736328125,
0.21728515625,
0.54736328125
],
[
0.2181396484375,
0.55029296875,
0.78515625,
0.55029296875,
0.78515625,
0.564453125,
0.2181396484375,
0.564453125
],
[
0.2181396484375,
0.564453125,
0.78173828125,
0.564453125,
0.78173828125,
0.57763671875,
0.2181396484375,
0.57763671875
],
[
0.2181396484375,
0.58056640625,
0.7861328125,
0.58056640625,
0.7861328125,
0.5947265625,
0.2181396484375,
0.5947265625
],
[
0.2197265625,
0.5966796875,
0.52294921875,
0.5966796875,
0.52294921875,
0.60986328125,
0.2197265625,
0.60986328125
],
[
0.2442626953125,
0.64306640625,
0.78515625,
0.64306640625,
0.78515625,
0.6572265625,
0.2442626953125,
0.6572265625
],
[
0.2197265625,
0.6591796875,
0.78515625,
0.6591796875,
0.78515625,
0.67236328125,
0.2197265625,
0.67236328125
],
[
0.221435546875,
0.673828125,
0.78369140625,
0.673828125,
0.78369140625,
0.6875,
0.221435546875,
0.6875
],
[
0.2197265625,
0.689453125,
0.78515625,
0.689453125,
0.78515625,
0.70263671875,
0.2197265625,
0.70263671875
],
[
0.2181396484375,
0.70263671875,
0.78515625,
0.7041015625,
0.78515625,
0.71826171875,
0.2181396484375,
0.71728515625
],
[
0.2197265625,
0.7197265625,
0.78369140625,
0.7197265625,
0.78369140625,
0.73291015625,
0.2197265625,
0.73291015625
],
[
0.2197265625,
0.73486328125,
0.529296875,
0.73486328125,
0.529296875,
0.74853515625,
0.2197265625,
0.74853515625
],
[
0.2459716796875,
0.75048828125,
0.78515625,
0.75048828125,
0.78515625,
0.7646484375,
0.2459716796875,
0.7646484375
],
[
0.2197265625,
0.7666015625,
0.78515625,
0.7666015625,
0.78515625,
0.77978515625,
0.2197265625,
0.77978515625
],
[
0.221435546875,
0.78076171875,
0.78515625,
0.78076171875,
0.78515625,
0.79443359375,
0.221435546875,
0.79443359375
],
[
0.2197265625,
0.79541015625,
0.78515625,
0.79541015625,
0.78515625,
0.8095703125,
0.2197265625,
0.8095703125
],
[
0.2197265625,
0.81103515625,
0.78369140625,
0.81103515625,
0.78369140625,
0.82470703125,
0.2197265625,
0.82470703125
],
[
0.2197265625,
0.8271484375,
0.78515625,
0.8271484375,
0.78515625,
0.84033203125,
0.2197265625,
0.84033203125
]
] |
[
"BigNAS: Neural Architecture Search with Big Single-Stage Models",
"5",
"同，而与扇人无关。此初始化最初在[9]中提到，并通过～0.2%在其设置中提",
"高了准确性，但在我们的设置中更为关键（通过～1.0%进行改进）。当分辨率",
"或通道不同时，我们还在每个阶段过渡中添加了一个跳过连接（如果需要，使用",
"2×2平均池化和/或1×1卷积）来显式构建标识映射[13]",
"收敛行为。在实践中，我们发现大的孩子模型收敛速度更快，而小的孩子模",
"型收敛速度更慢。图2a显示了单阶段模型训练过程中的典型学习曲线，其中",
"我们绘制了小型和大型子模型随时间变化的验证精度。这揭示了一个困境：在",
"训练步骤t，当大孩子模型的性能达到峰值时，小孩子模型没有得到充分训练；",
"在训练步骤t，当小孩子模型具有更好的性能时，大孩子模型已经过拟合",
"SmallChildMode",
"inehildM",
"nde",
"0.2",
"0.80",
"L",
"4",
"0.6",
"",
"20000",
"tomberefrong sec",
"20000",
"amberofooningsSeooo",
"100000",
":12000",
"100000",
"000",
"12000",
"(a)",
"(b)",
"Fig. 2: On the left, we show typical accuracy curves during the training process",
"for both small and big child models. It reveals a common dilemma in training big",
"single-stage models: at training step t when the performance of big child models",
"peaks, the small child models are not fully-trained; and at training step t' when",
"the small child models have better performance, the big child models already",
"overfitted. On the right, we plot the simple modified learning rate schedules",
"with constant ending to address this issue.",
"为了解决这个问题，我们将重点放在学习率时间表上。我们首先在图2b中",
"绘制了MobileNet系列[16,17,29]、MNasNets[32]和EfficientNets[33]的优",
"化和广泛使用的指数衰减学习速率调度。我们介绍了一个简单的修改这个学习",
"率的时间表，命名为指数衰减与恒定的结束，它有一个恒定的学习率在训练结",
"束时，当它达到5%的初始学习率（图2b）。它带来两个好处。首先，在最后",
"的学习率略大的情况下，小型儿童模型学习速度更快。其次，随着权重的振荡",
"末端的恒定学习率缓解了大子模型的过拟合",
"正则化。根据经验，在比较训练/验证损失时，我们发现大的子模型倾向于",
"过度拟合训练数据，而小的子模型倾向于不拟合。在以前的工作中，Bender等",
"人。[2]将相同的权重衰减应用于所有子模型，无论它们是小还是大。为了防止",
"较大网络的过度拟合，对于EfficientNets，Tan等人。[33]发现对较大的神经网",
"络使用较大的丢失率E[30]率很有帮助。在我们训练大型单阶段模型的背景下",
"由于具有共享参数的小型儿童模型和大型儿童模型之间的相互作用，这变得更"
] |
{
"arxiv_id": "2003.11142",
"language": "zh",
"page_idx": 4
}
|
||
2310.05095-zh-1
|
2310.05095-zh-1.jpg
|
AI 牛成的文本检测器有多可靠？使用规避软提示的评估框架
Tharindu Kumarage Paras Sheth Raha Moraffah Joshua Garland Huan Liu
Arizona State University
$\{$ kskumara, psheth5, rmoraffa, jtgarlan, huanliu $\}$ @asu.edu
# Abstract
In recent years, there has been a rapid proliferation of AI-generated text, primarily driven by the release of powerful pre-trained language models (PLMs). To address the issue of misuse associated with AI-generated text, various high-performing detectors have been developed, including the OpenAI detector and the Stanford DetectGPT. In our study, we ask how reliable these detectors are. We answer the question by designing a novel approach that can prompt any PLM to generate text that evades these high-performing detectors. The proposed approach suggests a universal evasive prompt, a novel type of soft prompt, which guides PLMs in producing human-like" text that can mislead the detectors. The novel universal evasive prompt is achieved in two steps: First, we create an evasive soft prompt tailored to a specific PLM through prompt tuning; and then, we leverage the transferability of soft prompts to transfer the learned evasive soft prompt from one PLM to another. Employing multiple PLMs in various writing tasks, we conduct extensive experiments to evaluate the efficacy of the evasive soft prompts in their evasion of state-of-the-art $detectors.$
### 1 Introduction
基于变压器的预训练语言模型(PLM)的最新进展，例如 PaLM$\boxed{\mathrm{F}}($Thoppilan et al.,2022) , GPT3。5$\overline{\mathrm{E}}($Ouyang et al.,2022) 和 GPT4$\underline{\mathrm{E}}($OpenAI,2023) 大大提高了自然语言生成 (NLG) 功能。结果，在各种写作任务中，包括摘要，论文写作，学术和科学写作以及新闻业，高度引人注目的 AI 生成文本激增。虽然人工智能生成的文本可能令人印象深刻，但它也带来了潜在的滥用风险，例如学术欺诈和传播人工智能生成的错误信息 (Dergaa et al., 2023; Kreps et al.,2022)。因此，为了打击与 AI 生成文本相关的滥用，我们观察到自动 AI 生成文本检测器的出现，其中包括 GPTZero$\underline{\mathrm{E}}($Tian,2023)
<FigureHere>
Figure 1: Evasive soft prompt guiding PLM to evade
AI-generated-text detector
TurnitinAI 文本检测器$^{1}$等商业产品，以及开源方法 DetectGPT$\boxed{\mathrm{E}}($Mitchell et al.,2023)和OpenAIGPT-2 检测器$\overline{\mathrm{E}}$(Solaiman et al., 2019) 的微调版本
尽管 AI 生成的文本检测器在使用现有数据集进行标准评估时表现出令人印象深刻的性能，但在实际场景中部署时，其功效可能会受到挑战。例如，恶意用户可能会操纵 PLM 以某种方式生成文本，从而误导检测器将其分类为人类书写的文本。。这种情况可能会在实际应用中导致不利后果，如学术欺诈检测或人工智能生成的新闻识别。因此，在这种情况下， 必须评估当前 AI 生成的文本检测器的可靠性。因此，在本研究中，我们旨在回答现有最先进的 AI 生成文本检测器的可靠性问题
为了回答这个问题，我们提出了一种规避软提示，这是一种新颖的软提示形式，专门用于规避 AI 生成的文本检测器。如图 1 所示，当与标准输入提示结合使用时，这种回避的软提示有效地引导 PLM 生成令人信服地类似于人类书写内容的文本，从而逃避了 AI 生成文本检测器的检测。鉴于近年来 PLM 的快速发展，开发一个可以评估当前和未来 PLM 上检测器可靠性的更通用的框架非常关键。为了促进这一点，我们将我们的回避软提示方案设计为“通用”, 使任何 PLM 都可以轻松采用它。拟议的通用回避软提示是通过两步过程实现的。首先，我们使用提示调整(PT)$\boxed{\mathrm{E}}($Lester et al., 2021a) 开发了针对特定 PLM 量身定制的回避软提示。然后，利用软提示的可传递性， 我们有效地将获得的回避软提示从一个 PLM
$^{1}$https://www.turnitin.com/solutions/ai-writing
|
[
[
0.183837890625,
0.0955810546875,
0.8134765625,
0.0955810546875,
0.8134765625,
0.10986328125,
0.183837890625,
0.10986328125
],
[
0.1435546875,
0.1455078125,
0.85986328125,
0.1455078125,
0.85986328125,
0.1602783203125,
0.1435546875,
0.1602783203125
],
[
0.4013671875,
0.160888671875,
0.60205078125,
0.162109375,
0.60205078125,
0.177001953125,
0.4013671875,
0.17578125
],
[
0.2091064453125,
0.1787109375,
0.8037109375,
0.1787109375,
0.8037109375,
0.193603515625,
0.2091064453125,
0.193603515625
],
[
0.26123046875,
0.25048828125,
0.341796875,
0.2529296875,
0.341796875,
0.268310546875,
0.26123046875,
0.2666015625
],
[
0.5224609375,
0.25830078125,
0.5859375,
0.25830078125,
0.5859375,
0.27783203125,
0.5224609375,
0.27783203125
],
[
0.68115234375,
0.273681640625,
0.76416015625,
0.273681640625,
0.76416015625,
0.305908203125,
0.68115234375,
0.305908203125
],
[
0.69775390625,
0.270751953125,
0.75146484375,
0.270751953125,
0.75146484375,
0.28076171875,
0.69775390625,
0.28076171875
],
[
0.5205078125,
0.27783203125,
0.60546875,
0.27783203125,
0.60546875,
0.2880859375,
0.5205078125,
0.2880859375
],
[
0.6337890625,
0.281982421875,
0.6455078125,
0.281982421875,
0.6455078125,
0.28515625,
0.6337890625,
0.28515625
],
[
0.80029296875,
0.281982421875,
0.81689453125,
0.281982421875,
0.81689453125,
0.290283203125,
0.80029296875,
0.290283203125
],
[
0.8671875,
0.29150390625,
0.87158203125,
0.29150390625,
0.87158203125,
0.29638671875,
0.8671875,
0.29638671875
],
[
0.1435546875,
0.283935546875,
0.459228515625,
0.283203125,
0.459228515625,
0.297607421875,
0.1435546875,
0.298583984375
],
[
0.8486328125,
0.292236328125,
0.861328125,
0.292236328125,
0.861328125,
0.29638671875,
0.8486328125,
0.29638671875
],
[
0.14697265625,
0.300537109375,
0.456787109375,
0.300537109375,
0.456787109375,
0.311279296875,
0.14697265625,
0.311279296875
],
[
0.5341796875,
0.2998046875,
0.58837890625,
0.2998046875,
0.58837890625,
0.31005859375,
0.5341796875,
0.31005859375
],
[
0.5400390625,
0.305908203125,
0.5810546875,
0.305908203125,
0.5810546875,
0.317626953125,
0.5400390625,
0.317626953125
],
[
0.7060546875,
0.300537109375,
0.73779296875,
0.300537109375,
0.73779296875,
0.30810546875,
0.7060546875,
0.30810546875
],
[
0.14697265625,
0.314208984375,
0.45751953125,
0.314208984375,
0.45751953125,
0.32470703125,
0.14697265625,
0.32470703125
],
[
0.1435546875,
0.32666015625,
0.459228515625,
0.325927734375,
0.459228515625,
0.34033203125,
0.1435546875,
0.341552734375
],
[
0.14697265625,
0.341552734375,
0.45751953125,
0.341552734375,
0.45751953125,
0.355224609375,
0.14697265625,
0.355224609375
],
[
0.51318359375,
0.336181640625,
0.880859375,
0.336181640625,
0.880859375,
0.349853515625,
0.51318359375,
0.349853515625
],
[
0.145263671875,
0.355224609375,
0.45751953125,
0.355224609375,
0.45751953125,
0.3681640625,
0.145263671875,
0.3681640625
],
[
0.5146484375,
0.350830078125,
0.68994140625,
0.350830078125,
0.68994140625,
0.361083984375,
0.5146484375,
0.361083984375
],
[
0.145263671875,
0.369384765625,
0.45751953125,
0.369384765625,
0.45751953125,
0.383056640625,
0.145263671875,
0.383056640625
],
[
0.145263671875,
0.38427734375,
0.456787109375,
0.38427734375,
0.456787109375,
0.39794921875,
0.145263671875,
0.39794921875
],
[
0.5283203125,
0.391357421875,
0.880859375,
0.392578125,
0.880859375,
0.40673828125,
0.5283203125,
0.40625
],
[
0.145263671875,
0.39794921875,
0.45751953125,
0.39794921875,
0.45751953125,
0.410888671875,
0.145263671875,
0.410888671875
],
[
0.51123046875,
0.407958984375,
0.87890625,
0.407958984375,
0.87890625,
0.4228515625,
0.51123046875,
0.4228515625
],
[
0.1485595703125,
0.41455078125,
0.455078125,
0.41455078125,
0.455078125,
0.424560546875,
0.1485595703125,
0.424560546875
],
[
0.14697265625,
0.427001953125,
0.45751953125,
0.427001953125,
0.45751953125,
0.440673828125,
0.14697265625,
0.440673828125
],
[
0.51318359375,
0.424560546875,
0.8818359375,
0.424560546875,
0.8818359375,
0.438232421875,
0.51318359375,
0.438232421875
],
[
0.1485595703125,
0.4423828125,
0.456787109375,
0.4423828125,
0.456787109375,
0.453125,
0.1485595703125,
0.453125
],
[
0.51123046875,
0.441162109375,
0.60791015625,
0.439453125,
0.60791015625,
0.45361328125,
0.51123046875,
0.4560546875
],
[
0.145263671875,
0.454833984375,
0.45751953125,
0.454833984375,
0.45751953125,
0.468505859375,
0.145263671875,
0.468505859375
],
[
0.53076171875,
0.457763671875,
0.880859375,
0.457763671875,
0.880859375,
0.471435546875,
0.53076171875,
0.471435546875
],
[
0.145263671875,
0.4697265625,
0.45751953125,
0.4697265625,
0.45751953125,
0.482666015625,
0.145263671875,
0.482666015625
],
[
0.5146484375,
0.474365234375,
0.880859375,
0.474365234375,
0.880859375,
0.488037109375,
0.5146484375,
0.488037109375
],
[
0.145263671875,
0.482666015625,
0.45751953125,
0.482177734375,
0.45751953125,
0.496337890625,
0.145263671875,
0.49755859375
],
[
0.145263671875,
0.495361328125,
0.45751953125,
0.49755859375,
0.45751953125,
0.51123046875,
0.145263671875,
0.5087890625
],
[
0.5146484375,
0.4921875,
0.87744140625,
0.4921875,
0.87744140625,
0.5029296875,
0.5146484375,
0.5029296875
],
[
0.14697265625,
0.5126953125,
0.45751953125,
0.5126953125,
0.45751953125,
0.525390625,
0.14697265625,
0.525390625
],
[
0.51318359375,
0.50732421875,
0.87890625,
0.50732421875,
0.87890625,
0.52099609375,
0.51318359375,
0.52099609375
],
[
0.145263671875,
0.525390625,
0.45751953125,
0.525390625,
0.45751953125,
0.5390625,
0.145263671875,
0.5390625
],
[
0.5146484375,
0.52490234375,
0.87890625,
0.52490234375,
0.87890625,
0.53515625,
0.5146484375,
0.53515625
],
[
0.145263671875,
0.54052734375,
0.45751953125,
0.54052734375,
0.45751953125,
0.55419921875,
0.145263671875,
0.55419921875
],
[
0.5146484375,
0.5390625,
0.880859375,
0.5390625,
0.880859375,
0.552734375,
0.5146484375,
0.552734375
],
[
0.14697265625,
0.5546875,
0.45751953125,
0.5546875,
0.45751953125,
0.568359375,
0.14697265625,
0.568359375
],
[
0.51318359375,
0.55419921875,
0.87890625,
0.55419921875,
0.87890625,
0.56787109375,
0.51318359375,
0.56787109375
],
[
0.14697265625,
0.568359375,
0.45751953125,
0.568359375,
0.45751953125,
0.58203125,
0.14697265625,
0.58203125
],
[
0.51318359375,
0.5693359375,
0.8818359375,
0.57080078125,
0.8818359375,
0.5849609375,
0.51318359375,
0.58447265625
],
[
0.14697265625,
0.5830078125,
0.45751953125,
0.5830078125,
0.45751953125,
0.5966796875,
0.14697265625,
0.5966796875
],
[
0.51318359375,
0.58740234375,
0.880859375,
0.58740234375,
0.880859375,
0.60107421875,
0.51318359375,
0.60107421875
],
[
0.1435546875,
0.5966796875,
0.459228515625,
0.595703125,
0.459228515625,
0.60986328125,
0.1435546875,
0.61083984375
],
[
0.51318359375,
0.6015625,
0.880859375,
0.6015625,
0.880859375,
0.615234375,
0.51318359375,
0.615234375
],
[
0.1485595703125,
0.61328125,
0.456787109375,
0.61328125,
0.456787109375,
0.62353515625,
0.1485595703125,
0.62353515625
],
[
0.14697265625,
0.6259765625,
0.456787109375,
0.6259765625,
0.456787109375,
0.6396484375,
0.14697265625,
0.6396484375
],
[
0.5146484375,
0.62060546875,
0.796875,
0.62060546875,
0.796875,
0.63134765625,
0.5146484375,
0.63134765625
],
[
0.145263671875,
0.6396484375,
0.45751953125,
0.6396484375,
0.45751953125,
0.65283203125,
0.145263671875,
0.65283203125
],
[
0.5341796875,
0.6357421875,
0.880859375,
0.6357421875,
0.880859375,
0.6494140625,
0.5341796875,
0.6494140625
],
[
0.14697265625,
0.65478515625,
0.21240234375,
0.65478515625,
0.21240234375,
0.66650390625,
0.14697265625,
0.66650390625
],
[
0.51318359375,
0.65283203125,
0.87890625,
0.65283203125,
0.87890625,
0.66650390625,
0.51318359375,
0.66650390625
],
[
0.51318359375,
0.66845703125,
0.880859375,
0.66845703125,
0.880859375,
0.681640625,
0.51318359375,
0.681640625
],
[
0.1158447265625,
0.68505859375,
0.145263671875,
0.68505859375,
0.145263671875,
0.69677734375,
0.1158447265625,
0.69677734375
],
[
0.151123046875,
0.68505859375,
0.2626953125,
0.68505859375,
0.2626953125,
0.69677734375,
0.151123046875,
0.69677734375
],
[
0.51708984375,
0.68505859375,
0.8818359375,
0.68505859375,
0.8818359375,
0.69873046875,
0.51708984375,
0.69873046875
],
[
0.51318359375,
0.70166015625,
0.87890625,
0.70166015625,
0.87890625,
0.71533203125,
0.51318359375,
0.71533203125
],
[
0.118408203125,
0.71142578125,
0.486083984375,
0.71142578125,
0.486083984375,
0.724609375,
0.118408203125,
0.724609375
],
[
0.5146484375,
0.71728515625,
0.880859375,
0.71728515625,
0.880859375,
0.73095703125,
0.5146484375,
0.73095703125
],
[
0.118408203125,
0.72705078125,
0.491943359375,
0.72705078125,
0.491943359375,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.51318359375,
0.73291015625,
0.87890625,
0.73291015625,
0.87890625,
0.74658203125,
0.51318359375,
0.74658203125
],
[
0.1175537109375,
0.7421875,
0.489501953125,
0.74365234375,
0.489501953125,
0.75830078125,
0.1175537109375,
0.75732421875
],
[
0.51318359375,
0.75,
0.880859375,
0.75,
0.880859375,
0.76318359375,
0.51318359375,
0.76318359375
],
[
0.1175537109375,
0.7587890625,
0.486083984375,
0.7587890625,
0.486083984375,
0.7724609375,
0.1175537109375,
0.7724609375
],
[
0.51123046875,
0.76416015625,
0.87890625,
0.76416015625,
0.87890625,
0.77783203125,
0.51123046875,
0.77783203125
],
[
0.1175537109375,
0.77490234375,
0.486083984375,
0.775390625,
0.486083984375,
0.79052734375,
0.1175537109375,
0.7890625
],
[
0.51123046875,
0.77978515625,
0.87744140625,
0.78076171875,
0.87744140625,
0.7958984375,
0.51123046875,
0.79443359375
],
[
0.1175537109375,
0.79150390625,
0.483642578125,
0.79150390625,
0.483642578125,
0.8046875,
0.1175537109375,
0.8046875
],
[
0.51123046875,
0.7958984375,
0.880859375,
0.79736328125,
0.880859375,
0.8125,
0.51123046875,
0.81005859375
],
[
0.12005615234375,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82177734375,
0.12005615234375,
0.82177734375
],
[
0.51123046875,
0.8134765625,
0.880859375,
0.8134765625,
0.880859375,
0.82666015625,
0.51123046875,
0.82666015625
],
[
0.118408203125,
0.82470703125,
0.486083984375,
0.82470703125,
0.486083984375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.51318359375,
0.830078125,
0.880859375,
0.830078125,
0.880859375,
0.84326171875,
0.51318359375,
0.84326171875
],
[
0.12005615234375,
0.84130859375,
0.486083984375,
0.84130859375,
0.486083984375,
0.85498046875,
0.12005615234375,
0.85498046875
],
[
0.51123046875,
0.845703125,
0.87744140625,
0.845703125,
0.87744140625,
0.859375,
0.51123046875,
0.859375
],
[
0.118408203125,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.5146484375,
0.8623046875,
0.87744140625,
0.8623046875,
0.87744140625,
0.8759765625,
0.5146484375,
0.8759765625
],
[
0.1175537109375,
0.87255859375,
0.484375,
0.87255859375,
0.484375,
0.88720703125,
0.1175537109375,
0.88720703125
],
[
0.51123046875,
0.87890625,
0.880859375,
0.876953125,
0.880859375,
0.89013671875,
0.51123046875,
0.892578125
],
[
0.118408203125,
0.88818359375,
0.484375,
0.88818359375,
0.484375,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.118408203125,
0.90478515625,
0.486083984375,
0.90478515625,
0.486083984375,
0.91845703125,
0.118408203125,
0.91845703125
],
[
0.5341796875,
0.90380859375,
0.8154296875,
0.90478515625,
0.8154296875,
0.91943359375,
0.5341796875,
0.91845703125
]
] |
[
"AI生成的文本检测器有多可靠？使用规避软提示的评估框架",
"Tharindu Kumarage Paras Sheth Raha Moraffah Joshua Garland Huan Liu",
"Arizona State University",
"{ kskumara, psheth5, rmoraffa, jtgarlan, huanliu } @asu.edu",
"Abstract",
"opipa",
"",
"In a rapidly",
"EvasiveSoftPrompt",
"",
"囍",
"I",
"In recent years, there has been a rapid prolif-",
"Nrit",
"eration of AI-generated text,primarily driven",
"Writea news",
"article...",
"crisis..",
"by the release of powerful pre-trainedlan-",
"guage models (PLMs). To address the is-",
"sue of misuse associated with Al-generated",
"Figure 1: Evasive soft prompt guiding PLM to evade",
"text, various high-performing detectors have",
"AI-generated-text detector",
"been developed, including the OpenAI detec-",
"tor and the Stanford DetectGPT. In our study,",
"TurnitinAI文本检测器丨等商业产品，以及",
"we ask how reliable these detectors are. We",
"开源方法DetectGPTE(Mitchell etal.,2023）和",
"answerthequestionbydesigninganovelap",
"proach that can prompt any PLM to generate",
"OpenAIGPT-2检测器E(Solaiman etal.,2019)",
"text that evades these high-performingdetec-",
"的微调版本",
"tors. The proposed approach suggests a uni-",
"尽管AI生成的文本检测器在使用现有数据",
"versal evasive prompt, a novel type of soft",
"集进行标准评估时表现出令人印象深刻的性",
"prompt, which guides PLMs in producing",
"\"human-like\" text that can mislead the detec-",
"能，但在实际场景中部署时，其功效可能会受",
"tors.The novel universal evasive prompt is",
"到挑战。例如，恶意用户可能会操纵PLM以",
"achieved in two steps: First, we create an eva-",
"某种方式生成文本，从而误导检测器将其分类",
"sive soft prompt tailored to a specific PLM",
"为人类书写的文本。。这种情况可能会在实际",
"through prompt tuning; and then, we leverage",
"应用中导致不利后果，如学术欺诈检测或人工",
"the transferability of soft prompts to transfer",
"智能生成的新闻识别。因此，在这种情况下，",
"the learned evasive soft prompt from one PLM",
"必须评估当前AI生成的文本检测器的可靠性。",
"to another. Employing multiple PLMs in vari-",
"因此，在本研究中，我们旨在回答现有最先进",
"ouswritingtasks,weconductextensiveexper",
"iments to evaluate the efficacy of the evasive",
"的AI生成文本检测器的可靠性问题",
"soft prompts in their evasion of state-of-the-art",
"为了回答这个问题，我们提出了一种规避软",
"detectors.",
"提示，这是一种新颖的软提示形式，专门用于",
"规避AI生成的文本检测器。如图1所示，当",
"1",
"Introduction",
"与标准输入提示结合使用时，这种回避的软",
"提示有效地引导PLM生成令人信服地类似于",
"基于变压器的预训练语言模型（PLM）的最",
"人类书写内容的文本，从而逃避了AI生成文",
"新进展，例如PaLME（Thoppilanetal.，2022），",
"本检测器的检测。鉴于近年来PLM的快速发",
"GPT3。5E(Ouyang et al.,2022)和 GPT4E(Ope-",
"展，开发一个可以评估当前和未来PLM上检",
"nAI,2023）大大提高了自然语言生成（NLG）功",
"测器可靠性的更通用的框架非常关键。为了促",
"能。结果，在各种写作任务中，包括摘要，论",
"进这一点，我们将我们的回避软提示方案设计",
"文写作，学术和科学写作以及新闻业，高度引",
"为“通用”，使任何PLM都可以轻松采用它。",
"人注目的AI生成文本激增。。虽然人工智能",
"拟议的通用回避软提示是通过两步过程实现",
"生成的文本可能令人印象深刻，但它也带来",
"的。首先，我们使用提示调整（PT）E(Lester",
"了潜在的滥用风险，例如学术欺诈和传播人工",
"etal.2021a)开发了针对特定PLM量身定制的",
"智能生成的错误信息（Dergaaetal.，2023;Kreps",
"回避软提示。然后，利用软提示的可传递性",
"etal.,2022）。因此，为了打击与AI生成文本",
"我们有效地将获得的回避软提示从一个PLM",
"相关的滥用，我们观察到自动AI生成文本检",
"测器的出现，其中包括GPTZeroE(Tian，2023)",
"'https://www.turnitin.com/solutions/ai-writing"
] |
{
"arxiv_id": "2310.05095",
"language": "zh",
"page_idx": 1
}
|
||
2310.05095-zh-2
|
2310.05095-zh-2.jpg
|
转移到另一个。我们将 AI 生成的文本检测器
的 EvasiveSoftPrompts 框架称为 EScaPe
我们在不同的 PLM 和不同的写作任务上进行的实验提供了 EScaPe 在逃避最先进的检测器方面的功效的经验证据。此外，我们在可转移性方面的实验证明了在多个突出的 PLM 上成功应用了回避软提示。此外，我们介绍了$\mathbf{EScaPe}$在不同的最先进的检测器之间传输的有效性。我们的发现强调了进一步研究在设计更可靠的检测机制以打击与 AI 生成的文本相关的滥用方面的重要性。总之，我们的研究做出了以下主要贡献：
1. We propose the framework EScaPe
which introduces the concept of evasive soft prompts—a novel type of soft prompt specifically designed to guide PLMs in evading state-of-the-art AI-generated-text detectors.
2. We demonstrate the transferability of evasive
soft prompts learned through the EScaPe framework across different PLMs, making it a universal evasive soft prompt to assess the reliability of AI-generated-text detectors on any current or future PLM.
3. We conducted extensive experiments by em-
ploying a wide range of PLMs in various realworld writing tasks to demonstrate the effectiveness of the proposed framework, consequently highlighting the vulnerability of the existing AI-generated-text detectors.
## 2 $\textbf{Related Work}$
2.1 $\textbf{Detection of AI- Generated Text}$
AI 文本检测的任务通常被认为是二进制分类任务，有两类：“人类编写的”和“AI 编写的”。“在早期，探索了几种基于监督学习的分类方法来检测 AI 生成的文本，例如逻辑回归和 SVC$\underline{\mathrm{E}}($Ippolito et al.,2019)。相反，GLTRE (Gehrmann et al., 2019) 采用一组简单的统计测试来确定输入文本序列是否是 AI 生成的，从而使该方法成为零。近年来，基于 PLM 的微调探测器已经成为最先进的技术，包括 OpenAI 的 GPT2 探测器 (Solaiman et al., 2019; Jawahar et al., 2020; Zellers et al., 2019; Kumarage et al., 2023)。随着较新的大型语言模型的快速发展人们越来越重视少发或零发检测的功能以及这些检测器的可解释性$\overline {\mathrm{E} }( \text{Mitrovi\' {c}etal. , 2023) }$ 。一些新的探测器包括 GPTZero$\boxed{\mathrm{F}}($Tian, 2023) 和 OpenAI 的探测器$\overline{\mathrm{E}}$(Kirchner et al., 2023) 等商业产品。最近一种名为 DetectGPT$\underline{\mathrm{E}}($Mitchell
et al., 2023) 的高性能零触发检测方法基于以下假设：对 AI 生成的文本进行较小的重写将显示出比原始样本更低的令牌日志概率。在PLM 生成的文本上加水印也是一种令人兴奋的方法，在研究界引起了关注(Kirchenbauer et al.,2023)。但是，它假设 AI 生成器本身支持水印的实现，这降低了该方法的实用性
2.2 $\textbf{AI- generated- text Detector Evasior}$
最近很少有研究调查释义作为逃避 AI 生成的文本检测技术的功效。(Sadasivan et al., 2023; Krishna et al., 2023) 表明，对文本进行释义会极大地破坏 AI 生成的文本检测器的性能，从而引起人们对此类检测方法可靠性的担忧。但是，由于以下原因，我们的工作与释义有很大不同：1)我们旨在根据生成文本的原始 PLM 的功能评估现有检测器的可靠性。在释义中， 使用辅助 PLM 来重新表达原始 PLM 的文本以逃避检测，从而导致两步过程，并且 2) 释义攻击评估为检测器的 I 型和 II 型错误提供了统一的分数。但是，必须区分这两种类型的错误，以验证检测器的可靠性。在现实世界的场景中，最可能的情况将涉及II型错误，其中恶意行为者试图生成可以逃避检测器并导致假阴性的 AI 文本。我们的研究明确地侧重于模拟这种情况并评估 II 型错误
## 2.3 $\textbf{Soft Prompt Tuning}$
快速调整是一种广泛使用的方法，用于指导PLM 生成所需的输出$\boxed{\mathrm{E}}$(Jiang et al., 2020) 。GPT-3 在提示调整方面取得了早期成功，使用定制提示在多个任务上实现了卓越的性能 (Brown et al., 2020) 。这导致了对硬提示的广泛研究，硬提示是在离散空间中手动或自动制作的提示 (Mishra et al.,2021;Gao et al., 2020)。同时，研究人员探索了软提示的潜力 (Liu et al.,2023)。与硬提示不同，软提示位于连续嵌入空间中。因此，可以使用特定于任务的监督直接训练软提示 (Wu and Shi, 2022; Gu et al., 2022)。用于软提示的著名方法，例如提示调整(PT)$\operatorname{E}($Lester et al.,2021b)和 P 调整$\boxed{\mathrm{F}}($Liu et al.,2022),已获得与下游任务的PLM 的全参数微调相当的性能。此外，最近的工作已经提出了跨任务和跨 PLM 架构的软提示的可移植性$\overline{\mathrm{E}}($Su et al., 2022; Vu et al., 2021)
### 3 $\textbf{Methodology}$
图 2 说明了生成通用回避软提示的过程，该过程涉及两个主要步骤：回避软提示学习和回避软提示传递。在第一步中，我们学习特定冻结PLM 的回避软提示 (源 PLM- $PLM^\mathrm{s}$ )。第二
|
[
[
0.118408203125,
0.0872802734375,
0.48779296875,
0.0872802734375,
0.48779296875,
0.1009521484375,
0.118408203125,
0.1009521484375
],
[
0.51318359375,
0.0872802734375,
0.8759765625,
0.0872802734375,
0.8759765625,
0.097412109375,
0.51318359375,
0.097412109375
],
[
0.118408203125,
0.10272216796875,
0.4306640625,
0.10272216796875,
0.4306640625,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.51318359375,
0.10272216796875,
0.87890625,
0.10272216796875,
0.87890625,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.1376953125,
0.12054443359375,
0.486083984375,
0.12054443359375,
0.486083984375,
0.1312255859375,
0.1376953125,
0.1312255859375
],
[
0.51123046875,
0.1181640625,
0.880859375,
0.119384765625,
0.880859375,
0.133056640625,
0.51123046875,
0.1318359375
],
[
0.118408203125,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.118408203125,
0.1484375
],
[
0.51123046875,
0.1353759765625,
0.880859375,
0.1353759765625,
0.880859375,
0.1484375,
0.51123046875,
0.1484375
],
[
0.118408203125,
0.1519775390625,
0.486083984375,
0.1519775390625,
0.486083984375,
0.1650390625,
0.118408203125,
0.1650390625
],
[
0.51123046875,
0.15087890625,
0.880859375,
0.149658203125,
0.880859375,
0.1644287109375,
0.51123046875,
0.1650390625
],
[
0.1175537109375,
0.16748046875,
0.48779296875,
0.166259765625,
0.48779296875,
0.18115234375,
0.1175537109375,
0.1817626953125
],
[
0.51318359375,
0.16748046875,
0.880859375,
0.16748046875,
0.880859375,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.12005615234375,
0.18408203125,
0.484375,
0.18408203125,
0.484375,
0.19775390625,
0.12005615234375,
0.19775390625
],
[
0.51123046875,
0.182861328125,
0.8447265625,
0.182861328125,
0.8447265625,
0.196533203125,
0.51123046875,
0.196533203125
],
[
0.118408203125,
0.1995849609375,
0.484375,
0.1995849609375,
0.484375,
0.213134765625,
0.118408203125,
0.213134765625
],
[
0.12005615234375,
0.21728515625,
0.486083984375,
0.21728515625,
0.486083984375,
0.22802734375,
0.12005615234375,
0.22802734375
],
[
0.509765625,
0.2120361328125,
0.8369140625,
0.2120361328125,
0.8369140625,
0.2257080078125,
0.509765625,
0.2257080078125
],
[
0.118408203125,
0.232177734375,
0.48779296875,
0.232177734375,
0.48779296875,
0.2452392578125,
0.118408203125,
0.2452392578125
],
[
0.51318359375,
0.2340087890625,
0.87890625,
0.2340087890625,
0.87890625,
0.2476806640625,
0.51318359375,
0.2476806640625
],
[
0.118408203125,
0.2493896484375,
0.484375,
0.2493896484375,
0.484375,
0.260009765625,
0.118408203125,
0.260009765625
],
[
0.51318359375,
0.251708984375,
0.87744140625,
0.251708984375,
0.87744140625,
0.262451171875,
0.51318359375,
0.262451171875
],
[
0.118408203125,
0.26416015625,
0.273681640625,
0.26416015625,
0.273681640625,
0.27783203125,
0.118408203125,
0.27783203125
],
[
0.51318359375,
0.2666015625,
0.87890625,
0.2666015625,
0.87890625,
0.27978515625,
0.51318359375,
0.27978515625
],
[
0.51123046875,
0.28076171875,
0.87744140625,
0.28076171875,
0.87744140625,
0.29443359375,
0.51123046875,
0.29443359375
],
[
0.130126953125,
0.29150390625,
0.408935546875,
0.290283203125,
0.408935546875,
0.3046875,
0.130126953125,
0.305908203125
],
[
0.40478515625,
0.292236328125,
0.47021484375,
0.292236328125,
0.47021484375,
0.303955078125,
0.40478515625,
0.303955078125
],
[
0.5146484375,
0.2998046875,
0.87890625,
0.2998046875,
0.87890625,
0.31005859375,
0.5146484375,
0.31005859375
],
[
0.154541015625,
0.30810546875,
0.48779296875,
0.30810546875,
0.48779296875,
0.32177734375,
0.154541015625,
0.32177734375
],
[
0.51123046875,
0.3134765625,
0.87890625,
0.3134765625,
0.87890625,
0.32666015625,
0.51123046875,
0.32666015625
],
[
0.154541015625,
0.32470703125,
0.489501953125,
0.32470703125,
0.489501953125,
0.33837890625,
0.154541015625,
0.33837890625
],
[
0.51318359375,
0.330810546875,
0.8818359375,
0.330810546875,
0.8818359375,
0.344482421875,
0.51318359375,
0.344482421875
],
[
0.15283203125,
0.33837890625,
0.48779296875,
0.34033203125,
0.48779296875,
0.355224609375,
0.15283203125,
0.352783203125
],
[
0.51123046875,
0.345703125,
0.8818359375,
0.3466796875,
0.8818359375,
0.361083984375,
0.51123046875,
0.35986328125
],
[
0.154541015625,
0.355712890625,
0.47265625,
0.355712890625,
0.47265625,
0.369384765625,
0.154541015625,
0.369384765625
],
[
0.5146484375,
0.363525390625,
0.880859375,
0.363525390625,
0.880859375,
0.377197265625,
0.5146484375,
0.377197265625
],
[
0.130126953125,
0.38427734375,
0.48779296875,
0.38427734375,
0.48779296875,
0.39794921875,
0.130126953125,
0.39794921875
],
[
0.5146484375,
0.37890625,
0.8818359375,
0.37890625,
0.8818359375,
0.392578125,
0.5146484375,
0.392578125
],
[
0.15283203125,
0.399658203125,
0.48779296875,
0.3984375,
0.48779296875,
0.413330078125,
0.15283203125,
0.41455078125
],
[
0.515625,
0.3955078125,
0.87744140625,
0.3955078125,
0.87744140625,
0.40625,
0.515625,
0.40625
],
[
0.51318359375,
0.410888671875,
0.880859375,
0.410888671875,
0.880859375,
0.424560546875,
0.51318359375,
0.424560546875
],
[
0.1561279296875,
0.41748046875,
0.48779296875,
0.41748046875,
0.48779296875,
0.427490234375,
0.1561279296875,
0.427490234375
],
[
0.51318359375,
0.427001953125,
0.880859375,
0.427001953125,
0.880859375,
0.440673828125,
0.51318359375,
0.440673828125
],
[
0.15283203125,
0.432861328125,
0.489501953125,
0.432861328125,
0.489501953125,
0.446533203125,
0.15283203125,
0.446533203125
],
[
0.51123046875,
0.443603515625,
0.87890625,
0.443603515625,
0.87890625,
0.457275390625,
0.51123046875,
0.457275390625
],
[
0.1561279296875,
0.449462890625,
0.486083984375,
0.449462890625,
0.486083984375,
0.460205078125,
0.1561279296875,
0.460205078125
],
[
0.154541015625,
0.46435546875,
0.3525390625,
0.46435546875,
0.3525390625,
0.47802734375,
0.154541015625,
0.47802734375
],
[
0.5146484375,
0.46142578125,
0.8759765625,
0.46142578125,
0.8759765625,
0.471435546875,
0.5146484375,
0.471435546875
],
[
0.51318359375,
0.473876953125,
0.880859375,
0.474365234375,
0.880859375,
0.4892578125,
0.51318359375,
0.488037109375
],
[
0.1292724609375,
0.490478515625,
0.49365234375,
0.4912109375,
0.49365234375,
0.505859375,
0.1292724609375,
0.5048828125
],
[
0.51123046875,
0.4921875,
0.71875,
0.4921875,
0.71875,
0.505859375,
0.51123046875,
0.505859375
],
[
0.15283203125,
0.50830078125,
0.489501953125,
0.50732421875,
0.489501953125,
0.521484375,
0.15283203125,
0.5224609375
],
[
0.5087890625,
0.5185546875,
0.71728515625,
0.52099609375,
0.71728515625,
0.53515625,
0.5087890625,
0.533203125
],
[
0.1561279296875,
0.52490234375,
0.49365234375,
0.52490234375,
0.49365234375,
0.5380859375,
0.1561279296875,
0.5380859375
],
[
0.154541015625,
0.54052734375,
0.490234375,
0.54052734375,
0.490234375,
0.5546875,
0.154541015625,
0.5546875
],
[
0.5146484375,
0.5419921875,
0.880859375,
0.5419921875,
0.880859375,
0.5556640625,
0.5146484375,
0.5556640625
],
[
0.1561279296875,
0.55712890625,
0.489501953125,
0.55712890625,
0.489501953125,
0.57080078125,
0.1561279296875,
0.57080078125
],
[
0.51123046875,
0.55712890625,
0.880859375,
0.55810546875,
0.880859375,
0.572265625,
0.51123046875,
0.57177734375
],
[
0.154541015625,
0.572265625,
0.42236328125,
0.572265625,
0.42236328125,
0.5859375,
0.154541015625,
0.5859375
],
[
0.5146484375,
0.57470703125,
0.8818359375,
0.57470703125,
0.8818359375,
0.58837890625,
0.5146484375,
0.58837890625
],
[
0.5146484375,
0.59130859375,
0.880859375,
0.59130859375,
0.880859375,
0.60498046875,
0.5146484375,
0.60498046875
],
[
0.1158447265625,
0.60107421875,
0.154541015625,
0.60107421875,
0.154541015625,
0.61328125,
0.1158447265625,
0.61328125
],
[
0.1502685546875,
0.60107421875,
0.274658203125,
0.60107421875,
0.274658203125,
0.6142578125,
0.1502685546875,
0.6142578125
],
[
0.51318359375,
0.60791015625,
0.880859375,
0.60791015625,
0.880859375,
0.62158203125,
0.51318359375,
0.62158203125
],
[
0.1175537109375,
0.6259765625,
0.411376953125,
0.6259765625,
0.411376953125,
0.6396484375,
0.1175537109375,
0.6396484375
],
[
0.51318359375,
0.62353515625,
0.880859375,
0.62353515625,
0.880859375,
0.63720703125,
0.51318359375,
0.63720703125
],
[
0.515625,
0.6396484375,
0.880859375,
0.6396484375,
0.880859375,
0.65283203125,
0.515625,
0.65283203125
],
[
0.118408203125,
0.64794921875,
0.48779296875,
0.64794921875,
0.48779296875,
0.6611328125,
0.118408203125,
0.6611328125
],
[
0.51123046875,
0.65380859375,
0.87890625,
0.65380859375,
0.87890625,
0.66845703125,
0.51123046875,
0.66845703125
],
[
0.1175537109375,
0.6630859375,
0.484375,
0.6630859375,
0.484375,
0.6767578125,
0.1175537109375,
0.6767578125
],
[
0.51318359375,
0.67138671875,
0.8818359375,
0.67138671875,
0.8818359375,
0.68505859375,
0.51318359375,
0.68505859375
],
[
0.118408203125,
0.6796875,
0.484375,
0.6796875,
0.484375,
0.693359375,
0.118408203125,
0.693359375
],
[
0.51123046875,
0.68603515625,
0.87890625,
0.68603515625,
0.87890625,
0.69970703125,
0.51123046875,
0.69970703125
],
[
0.118408203125,
0.6953125,
0.486083984375,
0.6953125,
0.486083984375,
0.708984375,
0.118408203125,
0.708984375
],
[
0.5146484375,
0.70361328125,
0.87744140625,
0.70361328125,
0.87744140625,
0.71435546875,
0.5146484375,
0.71435546875
],
[
0.1175537109375,
0.71142578125,
0.486083984375,
0.71044921875,
0.486083984375,
0.724609375,
0.1175537109375,
0.7255859375
],
[
0.51318359375,
0.7197265625,
0.880859375,
0.7197265625,
0.880859375,
0.73291015625,
0.51318359375,
0.73291015625
],
[
0.118408203125,
0.72705078125,
0.483642578125,
0.72705078125,
0.483642578125,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.51318359375,
0.7353515625,
0.880859375,
0.7353515625,
0.880859375,
0.7490234375,
0.51318359375,
0.7490234375
],
[
0.118408203125,
0.74462890625,
0.484375,
0.74462890625,
0.484375,
0.75830078125,
0.118408203125,
0.75830078125
],
[
0.51123046875,
0.75048828125,
0.880859375,
0.751953125,
0.880859375,
0.765625,
0.51123046875,
0.76416015625
],
[
0.118408203125,
0.76025390625,
0.48779296875,
0.76025390625,
0.48779296875,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.51318359375,
0.7685546875,
0.880859375,
0.7685546875,
0.880859375,
0.7822265625,
0.51318359375,
0.7822265625
],
[
0.118408203125,
0.77685546875,
0.48779296875,
0.77685546875,
0.48779296875,
0.79052734375,
0.118408203125,
0.79052734375
],
[
0.51318359375,
0.78369140625,
0.880859375,
0.78369140625,
0.880859375,
0.79736328125,
0.51318359375,
0.79736328125
],
[
0.118408203125,
0.79150390625,
0.484375,
0.79150390625,
0.484375,
0.8046875,
0.118408203125,
0.8046875
],
[
0.1175537109375,
0.80712890625,
0.484375,
0.80810546875,
0.484375,
0.822265625,
0.1175537109375,
0.82177734375
],
[
0.5146484375,
0.8017578125,
0.875,
0.8017578125,
0.875,
0.8125,
0.5146484375,
0.8125
],
[
0.118408203125,
0.82470703125,
0.489501953125,
0.82470703125,
0.489501953125,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.50732421875,
0.82470703125,
0.66015625,
0.8291015625,
0.66015625,
0.8466796875,
0.50732421875,
0.8427734375
],
[
0.12005615234375,
0.84130859375,
0.484375,
0.84130859375,
0.484375,
0.85498046875,
0.12005615234375,
0.85498046875
],
[
0.118408203125,
0.85693359375,
0.484375,
0.85693359375,
0.484375,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51123046875,
0.85693359375,
0.87890625,
0.85693359375,
0.87890625,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.1427001953125,
0.87255859375,
0.48779296875,
0.87158203125,
0.48779296875,
0.8857421875,
0.1427001953125,
0.88720703125
],
[
0.51318359375,
0.87451171875,
0.87744140625,
0.87451171875,
0.87744140625,
0.88525390625,
0.51318359375,
0.88525390625
],
[
0.118408203125,
0.8896484375,
0.48779296875,
0.8896484375,
0.48779296875,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51318359375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51318359375,
0.90283203125
],
[
0.1175537109375,
0.90478515625,
0.489501953125,
0.90380859375,
0.489501953125,
0.91845703125,
0.1175537109375,
0.91943359375
],
[
0.509765625,
0.90478515625,
0.87890625,
0.90478515625,
0.87890625,
0.91845703125,
0.509765625,
0.91845703125
]
] |
[
"转移到另一个。我们将AI生成的文本检测器",
"etal.2023）的高性能零触发检测方法基于以下",
"的EvasiveSoftPrompts框架称为EScaPe",
"假设：对AI生成的文本进行较小的重写将",
"我们在不同的PLM和不同的写作任务上进",
"显示出比原始样本更低的令牌日志概率。在",
"行的实验提供了EScaPe在逃避最先进的检测",
"PLM生成的文本上加水印也是一种令人兴奋",
"器方面的功效的经验证据。此外，我们在可转",
"的方法，在研究界引起了关注（Kirchenbauer",
"移性方面的实验证明了在多个突出的PLM上",
"etal,2023）。但是，它假设AI生成器本身支",
"成功应用了回避软提示。此外，我们介绍了",
"持水印的实现，这降低了该方法的实用性",
"EScaPe在不同的最先进的检测器之间传输的",
"有效性。我们的发现强调了进一步研究在设计",
"2.2AI-generated-text Detector Evasion",
"更可靠的检测机制以打击与AI生成的文本相",
"最近很少有研究调查释义作为逃避AI生成的",
"关的滥用方面的重要性。总之，我们的研究做",
"文本检测技术的功效。（Sadasivanetal.，2023",
"出了以下主要贡献：",
"Krishnaetal.，2023）表明，对文本进行释义会",
"极大地破坏AI生成的文本检测器的性能，从",
"1.We proposethe framework",
"EScaPe",
"而引起人们对此类检测方法可靠性的担忧。但",
"which introduces the concept of evasive soft",
"是，由于以下原因，我们的工作与释义有很大",
"prompts-a novel type of soft prompt specif-",
"不同：1）我们旨在根据生成文本的原始PLM",
"ically designed to guide PLMs in evading",
"的功能评估现有检测器的可靠性。在释义中，",
"state-of-the-art AI-generated-text detectors.",
"使用辅助PLM来重新表达原始PLM的文本",
"2. We demonstrate the transferability of evasive",
"以逃避检测，从而导致两步过程，并且2）释",
"soft prompts learned through the EScaPe",
"义攻击评估为检测器的I型和I型错误提供了",
"统一的分数。但是，必须区分这两种类型的错",
"frameworkacrossdifferentPLMs,makingit",
"误，以验证检测器的可靠性。在现实世界的场",
"a universal evasive soft prompt to assess the",
"景中，最可能的情况将涉及Ⅱ型错误，其中恶",
"reliability of Al-generated-textdetectorson",
"any current or future PLM.",
"意行为者试图生成可以逃避检测器并导致假阴",
"性的AI文本。我们的研究明确地侧重于模拟",
"3. We conducted extensive experiments by em-",
"这种情况并评估Ⅱ型错误",
"ploying a wide range of PLMs in various real-",
"2.3Soft Prompt Tuning",
"world writing tasks to demonstrate the effec-",
"tiveness of the proposed framework, conse-",
"快速调整是一种广泛使用的方法，用于指导",
"quently highlighting the vulnerability of the",
"PLM生成所需的输出E(Jiangetal.，2020）。",
"existing Al-generated-text detectors.",
"GPT-3在提示调整方面取得了早期成功，使",
"用定制提示在多个任务上实现了卓越的性",
"2",
"Related Work",
"能(Brownetal.，2020）。这导致了对硬提示",
"2.1 Detection of AI-Generated Text",
"的广泛研究，硬提示是在离散空间中手动或",
"自动制作的提示（Mishraetal.，2021;Gaoetal.,",
"AI文本检测的任务通常被认为是二进制分类",
"2020）。同时，研究人员探索了软提示的潜",
"任务，有两类：“人类编写的”和“AI编写",
"力（Liuetal.,2023）。与硬提示不同，软提示位",
"的”。“在早期，探索了几种基于监督学习的分",
"于连续嵌人空间中。因此，可以使用特定于任",
"类方法来检测AI生成的文本，例如逻辑回归",
"务的监督直接训练软提示（WuandShi.2022",
"和 SVCE(Ippolito etal.,2019）。相反，GLTRE",
"Guetal.，2022）。用于软提示的著名方法，例",
"(Gehrmannetal.,2019）采用一组简单的统计测",
"如提示调整（PT）E(Lesteretal.,2021b）和P",
"试来确定输人文本序列是否是AI生成的，从",
"调整E(Liuetal.,2022），已获得与下游任务的",
"而使该方法成为零。近年来，基于PLM的微调",
"PLM的全参数微调相当的性能。此外，最近的",
"探测器已经成为最先进的技术，包括OpenAI",
"工作已经提出了跨任务和跨PLM架构的软提",
"的GPT2探测器（Solaiman etal.,2019;Jawahar",
"et al., 2020; Zellers et al., 2019; Kumarage et al.",
"示的可移植性F(Suetal.2022；Vuetal.,2021）",
"2023）。随着较新的大型语言模型的快速发展，",
"3Methodology",
"人们越来越重视少发或零发检测的功能以及",
"这些检测器的可解释性E(Mitrovicetal.，2023)",
"图2说明了生成通用回避软提示的过程，该过",
"一些新的探测器包括GPTZeroE(Tian，2023)",
"程涉及两个主要步骤：回避软提示学习和回避",
"和OpenAI的探测器E(Kirchneretal.，2023）等",
"软提示传递。在第一步中，我们学习特定冻结",
"商业产品。最近一种名为DetectGPTE(Mitchell",
"PLM的回避软提示（源PLM-PLMS）。第二"
] |
{
"arxiv_id": "2310.05095",
"language": "zh",
"page_idx": 2
}
|
||
2310.05095-zh-4
|
2310.05095-zh-4.jpg
|
$\bullet$ Reward is the reward function that aids the RL agent in performing actions more aligned towards the goal- making the generated text looks human-written to the detector. Given the language model $PLM^s$ , we initialize the policy $\pi$ as $PLM^s$ and then fine-tune only the evasive soft prompt embedding $P^\mathrm{s}$ to perform the task well using RL. We fit a reward model $f$ using the loss:
(3)
$$loss(f)=\mathbb{E}_{\left(Y,\{\hat{z}\}_i\right)}\left[\log\frac{e^{f(Y,\hat{z})}}{\sum_ie^{f(Y,\hat{z}_i)}}\right]$$
where $Y$ is the text generated by $PLM^s$ for the prompt $X$ , and $\hat{z}$ is the detector output. To keep $\pi$ from moving too far from $PLM^s$,we leverage a penalty with expectation $\beta$KL$( \pi , PLM^s)$ ,(where KL denotes the KL divergence) modifying the RL reward as,
(4)
$$\begin{aligned}F(y,\hat{z})&=f(y,\hat{z})-\beta\mathrm{KL}(\pi,PLM^{s})\\&=f(y,\hat{z})-\beta\log\frac{\pi(y\mid x)}{PLM^{s}(y\mid x)}\end{aligned}$$
(5)
Here $x\in X$ and $y\in Y$ (single instance of an input prompt and generated text). Adding the KL term adds two benefits. First, it prevents the policy from moving too far from the range where $f$ is valid, and second, it enforces coherence and topicality in the generated text.
· Action of the RL agent is updating the $k$ tokens of the evasive soft prompt $P^s$ based on the policy objective formulated as,
(6)
$$J(\lambda)=\mathbb{E}_{\pi_{\lambda}}[\sum_{i=1}^{k}F_{i}],$$
where $\lambda$ is the language model's parameters related to the soft prompt embedding, and $F_i$ is the reward for input $y_i$ and reward $\hat{z}_i.$ We use the gradient of this objective after each iteration to update the evasive soft prompt.
根据定义的奖励信号优化回避软提示，该模型
最终输出被检测器检测为人为书写的文本
## 3.2 Evasive Soft Prompt Transfer
在学习了$PLM^s$上的回避软提示之后，我们将经过训练的回避软提示转移到目标$PLM^t$ 的语义空间中。在软提示$\overline{\mathrm{E}}($Su et al., 2022) 的跨体系结构可传递性工作之后，我们训练投影仪将回避的软提示从源 PLM 有效地映射到目标 PLM
正式地，我们将源$PLM^s$的软提示定义为$P^s= p_1^s, . . . , p_k^s$ ; $P^s\in \mathbb{R} ^{e_s\times k}$。投影仪$g_\theta$旨在将$P^{s}$投影到$P^t\in\mathbb{R}^{e_t\times k}$,表示目标$PLM^t$的语义空间。这里，$e_s$和$e_t$分别表示源和目标PLM 的输入嵌入维度。我们使用两层前馈神经网络对投影仪函数$g_\mathrm{\theta}$进行参数化：
$$\begin{matrix}P^t=g_\theta(P^s)=W_2(\sigma(P^sW_1+b_1))+b_2&(7)\end{matrix}$$
在方程(7)中，$W_1\in \mathbb{R} _h^e\times e_s$ $, W_2\in \mathbb{R} ^{e_t\times e_h}$ 是可训练矩阵，$b_1\in\mathbb{R}^e_h$ 、$b_2\in\mathbb{R}^e_t$是偏置，$\sigma$ 是非线性激活函数。这里$e_h$表示投影仪网络的隐藏层大小。最后，我们使用我们的 RL 框架来训练上述跨模型投影参数，以进行回避软提示学习。但是，鉴于我们现在正在从已经学习的回避软提示$P^s$初始化$P^t$ ,因此可以通过较少的迭代来学习目标$PLM^t$的回避软提示
### 4 $\textbf{Experiment Setup}$
在这里，我们描述了用于验证我们的框架的实验设置，包括 AI 生成器 (PLM),编写任务(数据集),检测设置，基线以及 EScaPe 的实现细节，以支持可重复性
4.1 AI-Text Generators (PLMs)
我们在 HuggingFace 上提供的众多开源 PLM 上评估我们的框架。具体来说，我们选择了当前最先进的 PLM ,例如 LLaMA(7B;E(Touvron et al., 2023) ), Falcon ( 7B;$\underline{\mathrm{E}}($Almazrouei et al., 2023) ),以及已建立的强大 PLM , 例如 GPTNeoX (20B;$\overline{\mathbb{E}}$(Black et al.,2022))和 OPT(2。7B; (Zhang et al.,2022) )。在我们的实验中，我们以两种方式使用上述 PLM。首先，对于零触发语言生成，我们为分析中考虑的不同写作任务生成 AI 文本。其次，我们将它们用作框架回避软提示学习任务的基础 PLM。我们在这两种情况下都采用了类似的语言生成参数，将top-p 设置为 0。96 和温度为 0。9
## 4.2 $\textbf{Writing Tasks}$
在我们的实验中，我们分析了我们的框架如何与不同的 AI 相关的写作任务一起工作，这些任务有可能被滥用。我们专注于三个主要写作任务：新闻写作，学术写作和创意写作
对于新闻写作任务，我们结合了两个数据集，以收集来自 CNN,《华盛顿邮报》和 BBC 等知名来源的人工撰写的新闻文章。我们从TuringBench 数据集$\overline{\mathbb{E}}$(Uchendu et al.,2021) 获得了CNN 和《华盛顿邮报》的文章，从 Xsum 数据集$\overline{\mathbb{E}}($Narayan et al.,2018)获得了 BBC 的文章。为了表示学术论文，我们使用了 SQuAD 数据集$\boxed{\mathrm{F}}($Rajpurkar et al.,2016),并从上下文
|
[
[
0.1158447265625,
0.08612060546875,
0.489501953125,
0.08612060546875,
0.489501953125,
0.09979248046875,
0.1158447265625,
0.09979248046875
],
[
0.52880859375,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.52880859375,
0.09979248046875
],
[
0.1318359375,
0.10272216796875,
0.490234375,
0.10272216796875,
0.490234375,
0.11639404296875,
0.1318359375,
0.11639404296875
],
[
0.5087890625,
0.09857177734375,
0.8818359375,
0.097412109375,
0.8818359375,
0.1181640625,
0.5087890625,
0.119384765625
],
[
0.1318359375,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.1318359375,
0.1318359375
],
[
0.51123046875,
0.11517333984375,
0.8818359375,
0.11517333984375,
0.8818359375,
0.133056640625,
0.51123046875,
0.133056640625
],
[
0.1318359375,
0.1353759765625,
0.489501953125,
0.1353759765625,
0.489501953125,
0.1484375,
0.1318359375,
0.1484375
],
[
0.51318359375,
0.1353759765625,
0.880859375,
0.1353759765625,
0.880859375,
0.1484375,
0.51318359375,
0.1484375
],
[
0.130126953125,
0.15087890625,
0.489501953125,
0.149658203125,
0.489501953125,
0.1644287109375,
0.130126953125,
0.1650390625
],
[
0.51123046875,
0.149658203125,
0.87744140625,
0.149658203125,
0.87744140625,
0.163330078125,
0.51123046875,
0.163330078125
],
[
0.130126953125,
0.1650390625,
0.48779296875,
0.166259765625,
0.48779296875,
0.18115234375,
0.130126953125,
0.179931640625
],
[
0.5146484375,
0.16748046875,
0.80419921875,
0.16748046875,
0.80419921875,
0.18115234375,
0.5146484375,
0.18115234375
],
[
0.130126953125,
0.1817626953125,
0.483642578125,
0.1817626953125,
0.483642578125,
0.1953125,
0.130126953125,
0.1953125
],
[
0.130126953125,
0.19775390625,
0.486083984375,
0.198974609375,
0.486083984375,
0.213134765625,
0.130126953125,
0.2120361328125
],
[
0.52880859375,
0.1923828125,
0.8818359375,
0.19482421875,
0.8818359375,
0.2120361328125,
0.52880859375,
0.210205078125
],
[
0.1318359375,
0.2161865234375,
0.193115234375,
0.2161865234375,
0.193115234375,
0.22802734375,
0.1318359375,
0.22802734375
],
[
0.52880859375,
0.22265625,
0.87890625,
0.22265625,
0.87890625,
0.240478515625,
0.52880859375,
0.240478515625
],
[
0.359375,
0.2381591796875,
0.41650390625,
0.2381591796875,
0.41650390625,
0.2529296875,
0.359375,
0.2529296875
],
[
0.154541015625,
0.24462890625,
0.306396484375,
0.2529296875,
0.3046875,
0.271484375,
0.15283203125,
0.26318359375
],
[
0.5146484375,
0.2410888671875,
0.880859375,
0.2410888671875,
0.880859375,
0.254638671875,
0.5146484375,
0.254638671875
],
[
0.466796875,
0.248779296875,
0.489501953125,
0.254150390625,
0.483642578125,
0.2666015625,
0.4609375,
0.26123046875
],
[
0.341796875,
0.260009765625,
0.42822265625,
0.2529296875,
0.4306640625,
0.270751953125,
0.344970703125,
0.27783203125
],
[
0.5146484375,
0.25830078125,
0.880859375,
0.25830078125,
0.880859375,
0.271484375,
0.5146484375,
0.271484375
],
[
0.51318359375,
0.273681640625,
0.880859375,
0.273681640625,
0.880859375,
0.287353515625,
0.51318359375,
0.287353515625
],
[
0.130126953125,
0.2861328125,
0.486083984375,
0.2861328125,
0.486083984375,
0.2998046875,
0.130126953125,
0.2998046875
],
[
0.51318359375,
0.290283203125,
0.87890625,
0.290283203125,
0.87890625,
0.300537109375,
0.51318359375,
0.300537109375
],
[
0.1318359375,
0.303955078125,
0.489501953125,
0.303955078125,
0.489501953125,
0.317626953125,
0.1318359375,
0.317626953125
],
[
0.51123046875,
0.305908203125,
0.880859375,
0.305908203125,
0.880859375,
0.319580078125,
0.51123046875,
0.319580078125
],
[
0.1318359375,
0.318359375,
0.48779296875,
0.318359375,
0.48779296875,
0.33203125,
0.1318359375,
0.33203125
],
[
0.51318359375,
0.32177734375,
0.880859375,
0.32177734375,
0.880859375,
0.3349609375,
0.51318359375,
0.3349609375
],
[
0.130126953125,
0.3349609375,
0.489501953125,
0.334228515625,
0.489501953125,
0.3486328125,
0.130126953125,
0.349853515625
],
[
0.51123046875,
0.33740234375,
0.869140625,
0.33740234375,
0.869140625,
0.3515625,
0.51123046875,
0.3515625
],
[
0.1318359375,
0.350830078125,
0.48779296875,
0.350830078125,
0.48779296875,
0.365234375,
0.1318359375,
0.365234375
],
[
0.1318359375,
0.3681640625,
0.306396484375,
0.3681640625,
0.306396484375,
0.3818359375,
0.1318359375,
0.3818359375
],
[
0.51123046875,
0.364013671875,
0.70361328125,
0.366455078125,
0.70361328125,
0.38134765625,
0.51123046875,
0.37890625
],
[
0.14697265625,
0.394287109375,
0.42822265625,
0.394287109375,
0.42822265625,
0.4091796875,
0.14697265625,
0.4091796875
],
[
0.462646484375,
0.394287109375,
0.489501953125,
0.394287109375,
0.489501953125,
0.407958984375,
0.462646484375,
0.407958984375
],
[
0.51318359375,
0.391357421875,
0.880859375,
0.391357421875,
0.880859375,
0.405029296875,
0.51318359375,
0.405029296875
],
[
0.51123046875,
0.40673828125,
0.87744140625,
0.40673828125,
0.87744140625,
0.42041015625,
0.51123046875,
0.42041015625
],
[
0.202392578125,
0.413330078125,
0.455078125,
0.42578125,
0.45166015625,
0.451904296875,
0.198974609375,
0.439453125
],
[
0.365234375,
0.41455078125,
0.42822265625,
0.41455078125,
0.42822265625,
0.4287109375,
0.365234375,
0.4287109375
],
[
0.462646484375,
0.42333984375,
0.486083984375,
0.42333984375,
0.486083984375,
0.435791015625,
0.462646484375,
0.435791015625
],
[
0.509765625,
0.421630859375,
0.880859375,
0.4228515625,
0.880859375,
0.43701171875,
0.509765625,
0.435791015625
],
[
0.51318359375,
0.439453125,
0.71728515625,
0.439453125,
0.71728515625,
0.453125,
0.51318359375,
0.453125
],
[
0.1318359375,
0.4560546875,
0.486083984375,
0.4560546875,
0.486083984375,
0.4697265625,
0.1318359375,
0.4697265625
],
[
0.51123046875,
0.46435546875,
0.77734375,
0.465576171875,
0.77734375,
0.479736328125,
0.51123046875,
0.478515625
],
[
0.1318359375,
0.47265625,
0.486083984375,
0.47265625,
0.486083984375,
0.486328125,
0.1318359375,
0.486328125
],
[
0.1318359375,
0.488037109375,
0.489501953125,
0.488037109375,
0.489501953125,
0.501953125,
0.1318359375,
0.501953125
],
[
0.51123046875,
0.48681640625,
0.87890625,
0.48681640625,
0.87890625,
0.50048828125,
0.51123046875,
0.50048828125
],
[
0.1318359375,
0.5048828125,
0.48779296875,
0.5048828125,
0.48779296875,
0.5185546875,
0.1318359375,
0.5185546875
],
[
0.5146484375,
0.5048828125,
0.8759765625,
0.5048828125,
0.8759765625,
0.515625,
0.5146484375,
0.515625
],
[
0.133544921875,
0.52099609375,
0.48779296875,
0.52099609375,
0.48779296875,
0.53369140625,
0.133544921875,
0.53369140625
],
[
0.5146484375,
0.51953125,
0.87890625,
0.51953125,
0.87890625,
0.533203125,
0.5146484375,
0.533203125
],
[
0.1318359375,
0.53759765625,
0.361083984375,
0.53759765625,
0.361083984375,
0.55029296875,
0.1318359375,
0.55029296875
],
[
0.51318359375,
0.53369140625,
0.87890625,
0.53369140625,
0.87890625,
0.54736328125,
0.51318359375,
0.54736328125
],
[
0.51318359375,
0.5517578125,
0.8818359375,
0.5517578125,
0.8818359375,
0.5654296875,
0.51318359375,
0.5654296875
],
[
0.1175537109375,
0.5625,
0.48779296875,
0.5625,
0.48779296875,
0.57666015625,
0.1175537109375,
0.57666015625
],
[
0.51123046875,
0.56640625,
0.87890625,
0.56640625,
0.87890625,
0.580078125,
0.51123046875,
0.580078125
],
[
0.1318359375,
0.5791015625,
0.489501953125,
0.5791015625,
0.489501953125,
0.5927734375,
0.1318359375,
0.5927734375
],
[
0.51318359375,
0.58447265625,
0.880859375,
0.58447265625,
0.880859375,
0.59716796875,
0.51318359375,
0.59716796875
],
[
0.130126953125,
0.59423828125,
0.337646484375,
0.59326171875,
0.337646484375,
0.60791015625,
0.130126953125,
0.609375
],
[
0.51318359375,
0.5986328125,
0.87890625,
0.5986328125,
0.87890625,
0.6123046875,
0.51318359375,
0.6123046875
],
[
0.333251953125,
0.6181640625,
0.353515625,
0.6181640625,
0.353515625,
0.63134765625,
0.333251953125,
0.63134765625
],
[
0.5146484375,
0.615234375,
0.880859375,
0.615234375,
0.880859375,
0.62890625,
0.5146484375,
0.62890625
],
[
0.2283935546875,
0.63134765625,
0.387939453125,
0.63134765625,
0.387939453125,
0.65185546875,
0.2283935546875,
0.65185546875
],
[
0.46337890625,
0.63525390625,
0.48779296875,
0.63525390625,
0.48779296875,
0.64794921875,
0.46337890625,
0.64794921875
],
[
0.51318359375,
0.6318359375,
0.880859375,
0.6318359375,
0.880859375,
0.6455078125,
0.51318359375,
0.6455078125
],
[
0.330078125,
0.65087890625,
0.355224609375,
0.65087890625,
0.355224609375,
0.6630859375,
0.330078125,
0.6630859375
],
[
0.5146484375,
0.6484375,
0.880859375,
0.6484375,
0.880859375,
0.662109375,
0.5146484375,
0.662109375
],
[
0.5146484375,
0.66455078125,
0.880859375,
0.66455078125,
0.880859375,
0.677734375,
0.5146484375,
0.677734375
],
[
0.130126953125,
0.67041015625,
0.484375,
0.6728515625,
0.484375,
0.68701171875,
0.130126953125,
0.68505859375
],
[
0.509765625,
0.6787109375,
0.78515625,
0.677734375,
0.78515625,
0.6953125,
0.509765625,
0.69677734375
],
[
0.1318359375,
0.6884765625,
0.484375,
0.6884765625,
0.484375,
0.70166015625,
0.1318359375,
0.70166015625
],
[
0.1292724609375,
0.70361328125,
0.489501953125,
0.703125,
0.489501953125,
0.71728515625,
0.1292724609375,
0.71875
],
[
0.51123046875,
0.70703125,
0.6708984375,
0.70703125,
0.6708984375,
0.72021484375,
0.51123046875,
0.72021484375
],
[
0.1318359375,
0.72021484375,
0.48779296875,
0.72021484375,
0.48779296875,
0.73388671875,
0.1318359375,
0.73388671875
],
[
0.51318359375,
0.72705078125,
0.87890625,
0.72705078125,
0.87890625,
0.74072265625,
0.51318359375,
0.74072265625
],
[
0.1318359375,
0.7353515625,
0.361083984375,
0.736328125,
0.361083984375,
0.75048828125,
0.1318359375,
0.75
],
[
0.515625,
0.74365234375,
0.880859375,
0.74365234375,
0.880859375,
0.75732421875,
0.515625,
0.75732421875
],
[
0.118408203125,
0.76123046875,
0.48779296875,
0.76123046875,
0.48779296875,
0.77490234375,
0.118408203125,
0.77490234375
],
[
0.51318359375,
0.7587890625,
0.87890625,
0.7587890625,
0.87890625,
0.7724609375,
0.51318359375,
0.7724609375
],
[
0.118408203125,
0.77685546875,
0.44921875,
0.77685546875,
0.44921875,
0.79052734375,
0.118408203125,
0.79052734375
],
[
0.515625,
0.77783203125,
0.82275390625,
0.77783203125,
0.82275390625,
0.7880859375,
0.515625,
0.7880859375
],
[
0.53076171875,
0.7919921875,
0.880859375,
0.7919921875,
0.880859375,
0.8056640625,
0.53076171875,
0.8056640625
],
[
0.1175537109375,
0.80419921875,
0.397216796875,
0.80419921875,
0.397216796875,
0.81787109375,
0.1175537109375,
0.81787109375
],
[
0.51318359375,
0.80810546875,
0.880859375,
0.80810546875,
0.880859375,
0.82177734375,
0.51318359375,
0.82177734375
],
[
0.118408203125,
0.82373046875,
0.484375,
0.82373046875,
0.484375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.51318359375,
0.82470703125,
0.87890625,
0.82470703125,
0.87890625,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.1175537109375,
0.84033203125,
0.483642578125,
0.84033203125,
0.483642578125,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.509765625,
0.84033203125,
0.880859375,
0.8388671875,
0.880859375,
0.85400390625,
0.509765625,
0.85498046875
],
[
0.118408203125,
0.85693359375,
0.484375,
0.85693359375,
0.484375,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.51123046875,
0.85693359375,
0.87890625,
0.85693359375,
0.87890625,
0.87060546875,
0.51123046875,
0.87060546875
],
[
0.1175537109375,
0.87255859375,
0.483642578125,
0.87255859375,
0.483642578125,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.51318359375,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.118408203125,
0.8896484375,
0.483642578125,
0.8896484375,
0.483642578125,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51318359375,
0.88818359375,
0.880859375,
0.88720703125,
0.880859375,
0.90185546875,
0.51318359375,
0.90283203125
],
[
0.118408203125,
0.90625,
0.181396484375,
0.90625,
0.181396484375,
0.91748046875,
0.118408203125,
0.91748046875
],
[
0.5146484375,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91943359375,
0.5146484375,
0.91943359375
]
] |
[
". Reward is the reward function that aids the RL",
"正式地，我们将源PLMS的软提示定义为",
"agent in performing actions more aligned to-",
"Ps= pi,.. %;pse Resxk。投影仪 ge 旨在",
"wards the goal - making the generated text looks",
"将Ps投影到PtERetxk，表示目标PLMt的",
"human-written to the detector. Given the lan-",
"语义空间。这里，es和et分别表示源和目标",
"guage model PLMs, we initialize the policy",
"PLM的输嵌入维度。我们使用两层前馈神",
" as PLMs and then fine-tune only the evasive",
"经网络对投影仪函数9e进行参数化：",
"soft prompt embedding Ps to perform the task",
"well using RL. We fit a reward model f using",
"pt = ge(Ps) = W2(o(PsWi + bi)) + b2 (7)",
"the loss:",
"在方程(7)中, Wi Rehxes、W2 Retxen",
"ef(Y,2)",
"loss(f) = E(y,[3).)",
"是可训练矩阵，b1EReh、b2ERet是偏置，α",
"(3)",
"E, ef(Y,2)",
"是非线性激活函数。这里eh表示投影仪网络",
"的隐藏层大小。最后，我们使用我们的RL框",
"where Y is the text generated by PLMs for the",
"架来训练上述跨模型投影参数，以进行回避软",
"prompt X , and 2 is the detector output. To keep",
"提示学习。但是，鉴于我们现在正在从已经学",
"πfrommovingtoofarfromPLMs,welever",
"习的回避软提示Ps初始化Pt，因此可以通过",
"age a penalty with expectation βKL(π, PLMs)",
"较少的代来学习目标PLMt的回避软提示",
", (where KL denotes the KL divergence) modi-",
"fying the RL reward as,",
"4Experiment Setup",
"F(y,2) = f(y, 2) -βKL(π, PLM)",
"(4)",
"在这里，我们描述了用于验证我们的框架的实",
"验设置，包括AI生成器（PLM），编写任务",
"= f(y, 2) - βlog PLM-(y/ a)",
"π(y / α)",
"(5)",
"（数据集），检测设置，基线以及EScaPe的实",
"现细节，以支持可重复性",
"Here E X and y E Y (single instance of an",
"4.1 AI-Text Generators (PLMs)",
"input prompt and generated text). Adding the",
"KL term adds two benefits. First, it prevents the",
"我们在HuggingFace上提供的众多开源PLM",
"policy from moving too far from the range where",
"上评估我们的框架。具体来说，我们选择了当",
"f is valid, and second, it enforces coherence and",
"前最先进的PLM，例如LLaMA（7B;E(Touvron",
"topicality in the generated text.",
"et al., 2023) ), Falcon (7B;E(Almazrouei et al..",
"2023）），以及已建立的强大PLM，例如GPT",
": Action of the RL agent is updating the k tokens",
"NeoX（20B;E(Black etal.,2022)）和OPT（2",
"of the evasive soft prompt Ps based on the pol-",
"7B；（Zhangetal,2022)）。在我们的实验中，我",
"icy objective formulated as,",
"们以两种方式使用上述PLM。首先，对于零触",
"k",
"发语言生成，我们为分析中考虑的不同写作任",
"J() = Ex[Z Fil,",
"(6)",
"务生成AI文本。其次，我们将它们用作框架",
"i=1",
"回避软提示学习任务的基础PLM。我们在这",
"两种情况下都采用了类似的语言生成参数，将",
"where ^ is the language model's parameters re",
"top-p设置为0。96和温度为0。9",
"lated to the soft prompt embedding, and F; is the",
"reward for input yi and reward zi.We use the",
"4.2 Writing Tasks",
"gradient of this objective after each iteration to",
"在我们的实验中，我们分析了我们的框架如何",
"update the evasive soft prompt.",
"与不同的AI相关的写作任务一起工作，这些",
"根据定义的奖励信号优化回避软提示，该模型",
"任务有可能被滥用。我们专注于三个主要写作",
"最终输出被检测器检测为人为书写的文本",
"任务：新闻写作，学术写作和创意写作",
"对于新闻写作任务，我们结合了两个数据",
"3.2Evasive Soft Prompt Transfer",
"集，以收集来自CNN，《华盛顿邮报》和BBC",
"在学习了PLMS上的回避软提示之后，我们",
"等知名来源的人工撰写的新闻文章。我们从",
"将经过训练的回避软提示转移到目标PLMt",
"TuringBench数据集E(Uchenduet al.，2021）获",
"的语义空间中。在软提示E(Suetal.,2022)的",
"得了CNN和《华盛顿邮报》的文章，从Xsum",
"跨体系结构可传递性工作之后，我们训练投影",
"数据集E(Narayanetal.，2018）获得了BBC的",
"仪将回避的软提示从源PLM有效地映射到目",
"文章。为了表示学术论文，我们使用了SQuAD",
"标PLM",
"数据集E(Rajpurkaretal.，2016），并从上下文"
] |
{
"arxiv_id": "2310.05095",
"language": "zh",
"page_idx": 4
}
|
||
2305.05392-zh-0
|
2305.05392-zh-0.jpg
|
Sharpness-Aware Minimization Alone can Improve Adversarial Robustness
$\textbf{Zeming Wei}^{*1}\textbf{ Jingyu Zhu }^{*1}\textbf{ Yihao Zhang }^{*\text{ 1}}$
???。3in
Abstract
Sharpness-Aware Minimization (SAM) is an effective method for improving generalization ability by regularizing loss sharpness. In this paper, we explore SAM in the context of adversarial robustness. We find that using only SAM can achieve superior adversarial robustness without sacrificing clean accuracy compared to standard training, which is an unexpected benefit. We also discuss the relation between SAM and adversarial training (AT), a popular method for improving the adversarial robustness of DNNs. In particular, we show that SAM and AT differ in terms of perturbation strength, leading to different accuracy and robustness trade-offs. We provide theoretical evidence for these claims in a simplified model. Finally, while AT suffers from decreased clean accuracy and computational overhead, we suggest that SAM can be regarded as a lightweight substitute for AT under certain requirements. Code is available at https:// github. com/weizeming/SAM\_AT .
## $1. \textbf{Introduction}$
锐度感知最小化 (SAM)E(Foret et al.,2020) 是一种新颖的训练框架，可通过同时最小化损失值和损失锐度来改善模型泛化。SAM 的目标是最小化参数周围的锐度，可以表示为
(1)
$$\max\limits_{\|\epsilon\|\leq\rho}L(w+\epsilon)+\lambda\|w\|_2^2,$$
,其中$L$ 是损失函数，$w$ 是模型的参数，$\|w\|_2^2$是正则化项，$\rho$控制权重扰动的大小。直观上，更大的$\rho$会导致更强的重量扰动，并推动模型找到更平坦的损失景观。到目前为止，SAM 已成为增强机器学习模型自然准确性性能的强大工具
在本文中，我们旨在通过对抗性鲁棒性的角度来探索SAM。具体来说，我们研究了 SAM 的鲁棒性，以防御
$^{*}$Equal contribution$^{\text{1}}$Peking University. Correspondence to:
$\bar{\text{Zeming Wei<weizeming@stu.pku.edu.cn>}}$
$2^{nd}AdvML\textit{Frontiers workshop at 40}^{th}\textit{International Conference}$ $onMachine\textit{ Learning, Honolulu, Hawaii, USA. PMLR 202, 2023. }$ Copyright 2023 by the author(s).
对抗性示例，这些示例是具有小扰动的自然示例，这些扰动会误导模型产生不正确的预测E(Szegedy et al., 2013; Goodfellow et al., 2014)。对抗性示例的发现引起了人们对关键领域应用程序$\overline{\mathrm{E}}$(Ma et al., 2020) 的安全性的严重关注，并在防御方面引起了很多研究关注。目前，对抗训练(AT)$\boxed{\mathrm{F}}$(Madry et al.,2017)已被证明是最有效的方法$\boxed{\mathrm{F}}($Athalye et al., 2018) 在各种防御方法中提高深度神经网络(DNN)的对抗鲁棒性。然而尽管在提高对抗鲁棒性方面取得了成功，但对抗训练中仍然存在一些缺陷，例如自然准确性降低$\overline{\mathrm{E}}$(Tsipras et al., 2018) , 计算开销$\overline{\mathrm{E}}$(Shafahi et al., 2019) , 类公平性$\boxed{\mathrm{F}}($Xu et al., 2021; Wei et al., 2023a) 以及缺乏形式保证$\boxed{\mathrm{F}}($Wang et al., 2021; Zhang et al., 2023)
令人惊讶的是，我们发现使用 SAM 训练的模型比使用标准方法训练的模型表现出更高的对抗鲁棒性，这是一个意想不到的好处。此外，当将 SAM 与 AT 进行比较时，SAM 具有计算成本较低且自然精度不降低的优势。基于上述讨论，我们在本文中提出了两个研究问题 (RQ):
$\bullet$ RQ1: Why does SAM improve adversarial robustness
compared to standard training?
$\bullet$ RQ2: Can SAM be used as a lightweight substitution
for adversarial training?
要回答上述两个问题，我们首先在对抗鲁棒性方面提供对 SAM 的全面了解。具体来说，我们介绍了 SAM 和 AT 之间的内在联系 , 它们都在训练阶段应用对抗性数据增强来消除自然示例中的非鲁棒性特征[E](Tsipras et al.,2018) )。因此，SAM 和 AT 都可以有效地增强训练模型的鲁棒性，从而提高鲁棒泛化能力。但是，我们也注意到，SAM 和 AT 之间仍然存在一些差异。例如，SAM 隐式添加对抗性扰动，而 AT 显式应用扰动。此外，SAM 和 AT 训练期间的扰动(攻击)强度不同， 导致在自然和稳健的准确性权衡方面产生不同的结果此外，我们在简化的数据模型中用理论证据验证了所提出的经验理解。遵循基于鲁棒和非鲁棒特征分解$\overline{\mathrm{E}}$ (Tsipras et al.,2018)的数据分布，我们表明 SAM 和 AT 都可以通过在鲁棒特征上偏向更多权重来提高训练模型的鲁棒性。此外，我们还表明，SAM 需要更大的扰动预算才能实现与 AT 相当的鲁棒性，这验证了我们的假设，即 SAM 的扰动强度低于 AT
最后，我们在基准数据集上进行了实验，以验证我们的理解。我们发现，使用 SAM 训练的模型在对抗鲁棒性方面确实明显优于标准训练的模型，并且还表现出
|
[
[
0.111083984375,
0.11553955078125,
0.865234375,
0.11553955078125,
0.865234375,
0.1300048828125,
0.111083984375,
0.1300048828125
],
[
0.313720703125,
0.1787109375,
0.65869140625,
0.1787109375,
0.65869140625,
0.1932373046875,
0.313720703125,
0.1932373046875
],
[
0.08660888671875,
0.1932373046875,
0.1519775390625,
0.1956787109375,
0.1519775390625,
0.210205078125,
0.08660888671875,
0.2083740234375
],
[
0.2418212890625,
0.2052001953125,
0.32177734375,
0.20703125,
0.32177734375,
0.2247314453125,
0.2418212890625,
0.222900390625
],
[
0.5009765625,
0.2083740234375,
0.8857421875,
0.208984375,
0.8857421875,
0.2235107421875,
0.5009765625,
0.222900390625
],
[
0.1217041015625,
0.2298583984375,
0.44287109375,
0.2298583984375,
0.44287109375,
0.2442626953125,
0.1217041015625,
0.2442626953125
],
[
0.50244140625,
0.2247314453125,
0.8837890625,
0.2247314453125,
0.8837890625,
0.238037109375,
0.50244140625,
0.238037109375
],
[
0.12091064453125,
0.2442626953125,
0.44287109375,
0.24560546875,
0.44287109375,
0.260009765625,
0.12091064453125,
0.2587890625
],
[
0.5009765625,
0.2392578125,
0.88232421875,
0.2392578125,
0.88232421875,
0.25244140625,
0.5009765625,
0.25244140625
],
[
0.12091064453125,
0.260009765625,
0.441162109375,
0.261474609375,
0.441162109375,
0.27587890625,
0.12091064453125,
0.274658203125
],
[
0.50244140625,
0.255126953125,
0.8857421875,
0.255126953125,
0.8857421875,
0.268310546875,
0.50244140625,
0.268310546875
],
[
0.12091064453125,
0.274658203125,
0.438720703125,
0.274658203125,
0.438720703125,
0.28857421875,
0.12091064453125,
0.28857421875
],
[
0.50244140625,
0.26953125,
0.8857421875,
0.26953125,
0.8857421875,
0.28271484375,
0.50244140625,
0.28271484375
],
[
0.50341796875,
0.285400390625,
0.8857421875,
0.285400390625,
0.8857421875,
0.298583984375,
0.50341796875,
0.298583984375
],
[
0.1217041015625,
0.291748046875,
0.439453125,
0.291748046875,
0.439453125,
0.304931640625,
0.1217041015625,
0.304931640625
],
[
0.50341796875,
0.2998046875,
0.88232421875,
0.2998046875,
0.88232421875,
0.313232421875,
0.50341796875,
0.313232421875
],
[
0.12091064453125,
0.30615234375,
0.441162109375,
0.30615234375,
0.441162109375,
0.3193359375,
0.12091064453125,
0.3193359375
],
[
0.50244140625,
0.31494140625,
0.88232421875,
0.31494140625,
0.88232421875,
0.328857421875,
0.50244140625,
0.328857421875
],
[
0.1217041015625,
0.3212890625,
0.441162109375,
0.3212890625,
0.441162109375,
0.335205078125,
0.1217041015625,
0.335205078125
],
[
0.1217041015625,
0.3359375,
0.439453125,
0.3359375,
0.439453125,
0.349853515625,
0.1217041015625,
0.349853515625
],
[
0.50244140625,
0.330078125,
0.8857421875,
0.330078125,
0.8857421875,
0.343505859375,
0.50244140625,
0.343505859375
],
[
0.50341796875,
0.34521484375,
0.8837890625,
0.34521484375,
0.8837890625,
0.359130859375,
0.50341796875,
0.359130859375
],
[
0.1217041015625,
0.352783203125,
0.438720703125,
0.352783203125,
0.438720703125,
0.363037109375,
0.1217041015625,
0.363037109375
],
[
0.50244140625,
0.35986328125,
0.8837890625,
0.35986328125,
0.8837890625,
0.373779296875,
0.50244140625,
0.373779296875
],
[
0.1217041015625,
0.3662109375,
0.439453125,
0.3662109375,
0.439453125,
0.380126953125,
0.1217041015625,
0.380126953125
],
[
0.5009765625,
0.373779296875,
0.8857421875,
0.375732421875,
0.8857421875,
0.389404296875,
0.5009765625,
0.386962890625
],
[
0.12091064453125,
0.3818359375,
0.439453125,
0.3818359375,
0.439453125,
0.395751953125,
0.12091064453125,
0.395751953125
],
[
0.50244140625,
0.39013671875,
0.7861328125,
0.39013671875,
0.7861328125,
0.404052734375,
0.50244140625,
0.404052734375
],
[
0.12091064453125,
0.396484375,
0.441162109375,
0.396484375,
0.441162109375,
0.410400390625,
0.12091064453125,
0.410400390625
],
[
0.1192626953125,
0.412353515625,
0.439453125,
0.412353515625,
0.439453125,
0.425537109375,
0.1192626953125,
0.425537109375
],
[
0.50244140625,
0.410888671875,
0.88232421875,
0.41357421875,
0.88232421875,
0.4267578125,
0.50244140625,
0.4248046875
],
[
0.12091064453125,
0.4267578125,
0.441162109375,
0.4267578125,
0.441162109375,
0.43994140625,
0.12091064453125,
0.43994140625
],
[
0.50341796875,
0.429931640625,
0.88232421875,
0.429931640625,
0.88232421875,
0.43994140625,
0.50341796875,
0.43994140625
],
[
0.1217041015625,
0.442626953125,
0.441162109375,
0.442626953125,
0.441162109375,
0.455810546875,
0.1217041015625,
0.455810546875
],
[
0.515625,
0.443115234375,
0.8837890625,
0.443115234375,
0.8837890625,
0.45703125,
0.515625,
0.45703125
],
[
0.12335205078125,
0.458984375,
0.439453125,
0.458984375,
0.439453125,
0.4697265625,
0.12335205078125,
0.4697265625
],
[
0.50244140625,
0.458984375,
0.8857421875,
0.458984375,
0.8857421875,
0.472900390625,
0.50244140625,
0.472900390625
],
[
0.12091064453125,
0.4716796875,
0.441162109375,
0.47021484375,
0.441162109375,
0.48486328125,
0.12091064453125,
0.486083984375
],
[
0.50244140625,
0.473388671875,
0.8857421875,
0.473388671875,
0.8857421875,
0.4873046875,
0.50244140625,
0.4873046875
],
[
0.12091064453125,
0.4873046875,
0.441162109375,
0.4873046875,
0.441162109375,
0.50048828125,
0.12091064453125,
0.50048828125
],
[
0.50341796875,
0.490478515625,
0.56201171875,
0.490478515625,
0.56201171875,
0.501953125,
0.50341796875,
0.501953125
],
[
0.1217041015625,
0.50390625,
0.435546875,
0.50390625,
0.435546875,
0.5146484375,
0.1217041015625,
0.5146484375
],
[
0.51708984375,
0.51025390625,
0.8857421875,
0.51123046875,
0.8857421875,
0.52587890625,
0.51708984375,
0.5244140625
],
[
0.1217041015625,
0.517578125,
0.441162109375,
0.517578125,
0.441162109375,
0.53076171875,
0.1217041015625,
0.53076171875
],
[
0.53173828125,
0.5263671875,
0.73779296875,
0.5263671875,
0.73779296875,
0.54052734375,
0.53173828125,
0.54052734375
],
[
0.1217041015625,
0.53271484375,
0.393798828125,
0.5322265625,
0.393798828125,
0.54541015625,
0.1217041015625,
0.546875
],
[
0.5185546875,
0.548828125,
0.8837890625,
0.548828125,
0.8837890625,
0.5625,
0.5185546875,
0.5625
],
[
0.53173828125,
0.564453125,
0.69140625,
0.564453125,
0.69140625,
0.57763671875,
0.53173828125,
0.57763671875
],
[
0.08905029296875,
0.5771484375,
0.221435546875,
0.5771484375,
0.221435546875,
0.59033203125,
0.08905029296875,
0.59033203125
],
[
0.50341796875,
0.58837890625,
0.88232421875,
0.58837890625,
0.88232421875,
0.5986328125,
0.50341796875,
0.5986328125
],
[
0.08905029296875,
0.6015625,
0.473876953125,
0.6015625,
0.473876953125,
0.61572265625,
0.08905029296875,
0.61572265625
],
[
0.5009765625,
0.6015625,
0.8857421875,
0.6015625,
0.8857421875,
0.61572265625,
0.5009765625,
0.61572265625
],
[
0.0906982421875,
0.6171875,
0.47216796875,
0.6171875,
0.47216796875,
0.630859375,
0.0906982421875,
0.630859375
],
[
0.5009765625,
0.6171875,
0.8857421875,
0.6171875,
0.8857421875,
0.630859375,
0.5009765625,
0.630859375
],
[
0.0906982421875,
0.63330078125,
0.473876953125,
0.63330078125,
0.473876953125,
0.646484375,
0.0906982421875,
0.646484375
],
[
0.50244140625,
0.6337890625,
0.88232421875,
0.6337890625,
0.88232421875,
0.64453125,
0.50244140625,
0.64453125
],
[
0.0882568359375,
0.646484375,
0.205078125,
0.64794921875,
0.205078125,
0.662109375,
0.0882568359375,
0.6611328125
],
[
0.5009765625,
0.64599609375,
0.8857421875,
0.646484375,
0.8857421875,
0.6611328125,
0.5009765625,
0.66015625
],
[
0.5009765625,
0.6611328125,
0.8857421875,
0.662109375,
0.8857421875,
0.6767578125,
0.5009765625,
0.67529296875
],
[
0.1854248046875,
0.66748046875,
0.373291015625,
0.66552734375,
0.373291015625,
0.7021484375,
0.1854248046875,
0.7041015625
],
[
0.451904296875,
0.673828125,
0.473876953125,
0.673828125,
0.473876953125,
0.6875,
0.451904296875,
0.6875
],
[
0.50244140625,
0.67822265625,
0.8857421875,
0.67822265625,
0.8857421875,
0.69140625,
0.50244140625,
0.69140625
],
[
0.50244140625,
0.6923828125,
0.8857421875,
0.6923828125,
0.8857421875,
0.70556640625,
0.50244140625,
0.70556640625
],
[
0.101318359375,
0.70849609375,
0.473876953125,
0.70849609375,
0.473876953125,
0.7216796875,
0.101318359375,
0.7216796875
],
[
0.50244140625,
0.70849609375,
0.8837890625,
0.70849609375,
0.8837890625,
0.7216796875,
0.50244140625,
0.7216796875
],
[
0.0906982421875,
0.72265625,
0.473876953125,
0.72265625,
0.473876953125,
0.736328125,
0.0906982421875,
0.736328125
],
[
0.5009765625,
0.72265625,
0.87744140625,
0.72265625,
0.87744140625,
0.736328125,
0.5009765625,
0.736328125
],
[
0.08905029296875,
0.736328125,
0.473876953125,
0.7373046875,
0.473876953125,
0.75048828125,
0.08905029296875,
0.75
],
[
0.50244140625,
0.744140625,
0.8857421875,
0.744140625,
0.8857421875,
0.75830078125,
0.50244140625,
0.75830078125
],
[
0.0906982421875,
0.75244140625,
0.47216796875,
0.75244140625,
0.47216796875,
0.7666015625,
0.0906982421875,
0.7666015625
],
[
0.5,
0.7587890625,
0.8857421875,
0.76123046875,
0.8857421875,
0.7744140625,
0.5,
0.77294921875
],
[
0.08905029296875,
0.76806640625,
0.250732421875,
0.76806640625,
0.250732421875,
0.77880859375,
0.08905029296875,
0.77880859375
],
[
0.5009765625,
0.7744140625,
0.88232421875,
0.7744140625,
0.88232421875,
0.78857421875,
0.5009765625,
0.78857421875
],
[
0.0906982421875,
0.79248046875,
0.47216796875,
0.79248046875,
0.47216796875,
0.80322265625,
0.0906982421875,
0.80322265625
],
[
0.50244140625,
0.79248046875,
0.8837890625,
0.79248046875,
0.8837890625,
0.80322265625,
0.50244140625,
0.80322265625
],
[
0.08905029296875,
0.80615234375,
0.473876953125,
0.80615234375,
0.473876953125,
0.8193359375,
0.08905029296875,
0.8193359375
],
[
0.5009765625,
0.80615234375,
0.8857421875,
0.80615234375,
0.8857421875,
0.8193359375,
0.5009765625,
0.8193359375
],
[
0.5009765625,
0.8193359375,
0.8857421875,
0.82080078125,
0.8857421875,
0.83544921875,
0.5009765625,
0.833984375
],
[
0.108642578125,
0.8291015625,
0.471435546875,
0.8291015625,
0.471435546875,
0.84326171875,
0.108642578125,
0.84326171875
],
[
0.0882568359375,
0.84130859375,
0.35693359375,
0.84228515625,
0.35693359375,
0.85693359375,
0.0882568359375,
0.85595703125
],
[
0.50244140625,
0.8359375,
0.7490234375,
0.8359375,
0.7490234375,
0.849609375,
0.50244140625,
0.849609375
],
[
0.50244140625,
0.85791015625,
0.8857421875,
0.85791015625,
0.8857421875,
0.8720703125,
0.50244140625,
0.8720703125
],
[
0.0882568359375,
0.86572265625,
0.471435546875,
0.8671875,
0.471435546875,
0.8818359375,
0.0882568359375,
0.8798828125
],
[
0.5009765625,
0.87353515625,
0.8857421875,
0.87353515625,
0.8857421875,
0.88720703125,
0.5009765625,
0.88720703125
],
[
0.0906982421875,
0.8798828125,
0.47216796875,
0.8798828125,
0.47216796875,
0.89306640625,
0.0906982421875,
0.89306640625
],
[
0.0882568359375,
0.89306640625,
0.283447265625,
0.89306640625,
0.283447265625,
0.9072265625,
0.0882568359375,
0.9072265625
],
[
0.50244140625,
0.89013671875,
0.88232421875,
0.89013671875,
0.88232421875,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"Sharpness-Aware Minimization Alone can Improve Adversarial Robustness",
"Zeming Wei * 1 Jingyu Zhu * 1 Yihao Zhang * 1",
"???。3in",
"Abstract",
"对抗性示例，这些示例是具有小扰动的自然示例，这",
"Sharpness-Aware Minimization (SAM) is an ef-",
"些扰动会误导模型产生不正确的预测E(Szegedyetal.，",
"fective method for improving generalization abil-",
"2013；Goodfellowetal.，2014）。对抗性示例的发现引",
"ity by regularizing loss sharpness. In this paper,",
"起了人们对关键领域应用程序E(Maetal.，2020)的安",
"we explore SAM in the context of adversarial",
"全性的严重关注，并在防御方面引起了很多研究关注。",
"目前，对抗训练（AT）E(Madryetal.，2017）已被证明",
"robustness. We find that using only SAM can",
"是最有效的方法E（Athalyeetal.，2018）在各种防御方",
"achieve superior adversarial robustness without",
"法中提高深度神经网络（DNN）的对抗鲁棒性。然而",
"sacrificing clean accuracy compared to standard",
"training, which is an unexpected benefit. We also",
"尽管在提高对抗鲁棒性方面取得了成功，但对抗训练",
"中仍然存在一些缺陷，例如自然准确性降低E(Tsipras",
"discusstherelationbetween SAMandadversar",
"etal.，2018），计算开销E(Shafahietal.，2019），类公平",
"ial training (AT), a popular method for improving",
"性E（Xuetal.,2021;Weietal.,2023a）以及缺乏形式保",
"the adversarial robustness of DNNs. In particu",
"证E(Wang et al., 2021; Zhang et al., 2023)",
"lar, we show that SAM and AT differ in terms",
"of perturbation strength, leading to different ac",
"令人惊牙的是，我们发现使用SAM训练的模型比使用",
"curacy and robustness trade-offs.We provide",
"标准方法训练的模型表现出更高的对抗鲁棒性，这是",
"theoretical evidence for these claims in a simpli-",
"个意想不到的好处。此外，当将SAM与AT进行比",
"fied model.Finally,while AT suffersfromde",
"较时，SAM具有计算成本较低且自然精度不降低的优",
"creased clean accuracy and computational over",
"势。基于上述讨论，我们在本文中提出了两个研究问",
"head, we suggest that SAM can be regarded as",
"题(RQ)：",
"a lightweight substitute for AT under certain re",
".RQI: Why does SAMimprove adversarial robustness",
"quirements.Code is available at https : / /",
"compared to standard training?",
"github.com/weizeming/sAM_AT",
": RQ2 : Can SAM be used as a lightweight substitution",
"for adversarial training?",
"1.Introduction",
"要回答上述两个问题，我们首先在对抗鲁棒性方面提",
"锐度感知最小化（SAM)E(Foretetal.，2020）是一种新",
"供对SAM的全面了解。具体来说，我们介绍了SAM",
"颖的训练框架，可通过同时最小化损失值和损失锐度",
"和AT之间的内在联系，它们都在训练阶段应用对抗性",
"来改善模型泛化。SAM的目标是最小化参数周围的锐",
"数据增强来消除自然示例中的非鲁棒性特征F(Tsipras",
"度，可以表示为",
"etal,2018)）。因此，SAM和AT都可以有效地增强训",
"练模型的鲁棒性，从而提高鲁棒泛化能力。但是，我",
", (w + )+ l,.",
"(1)",
"们也注意到，SAM和AT之间仍然存在一些差异。例",
"如，SAM隐式添加对抗性扰动，而AT显式应用扰动。",
"其中L是损失函数，W是模型的参数，Ⅱw是正则",
"此外，SAM和AT训练期间的扰动（攻击）强度不同",
"化项，p控制权重扰动的大小。直观上，更大的p会导",
"导致在自然和稳健的准确性权衡方面产生不同的结果",
"致更强的重量扰动，并推动模型找到更平坦的损失景",
"此外，我们在简化的数据模型中用理论证据验证了所",
"观。到目前为止，SAM已成为增强机器学习模型自然",
"提出的经验理解。遵循基于鲁棒和非鲁棒特征分解E",
"准确性性能的强大工具",
"(Tsiprasetal.，2018）的数据分布，我们表明SAM和AT",
"在本文中，我们旨在通过对抗性鲁棒性的角度来探索",
"都可以通过在鲁棒特征上偏向更多权重来提高训练模",
"SAM。具体来说，我们研究了SAM的鲁棒性，以防御",
"型的鲁棒性。此外，我们还表明，SAM需要更大的扰",
"动预算才能实现与AT相当的鲁棒性，这验证了我们的",
"Equal contribution 'Peking University.Correspondence to",
"Zeming Wei< weizeming@stu.pku.edu.cn >.",
"假设，即SAM的扰动强度低于AT",
"最后，我们在基准数据集上进行了实验，以验证我们",
"2nd AdvML Frontiers workshop at 4oth International Conference",
"的理解。我们发现，使用SAM训练的模型在对抗鲁棒",
"on Machine Learning,Honolulu, Hawai, USA. PMLR 202,2023.",
"Copyright 2023 by the author(s).",
"性方面确实明显优于标准训练的模型，并且还表现出"
] |
{
"arxiv_id": "2305.05392",
"language": "zh",
"page_idx": 0
}
|
||
2305.05392-zh-1
|
2305.05392-zh-1.jpg
|
Sharpness-Aware Minimization Alone can Improve Adversarial Robustness
更好的自然准确性。综上所述，我们的经验和理论理
解可以回答 RQ1
值得注意的是，SAM 和 AT 之间的健壮性仍然存在很大差距。然而，AT 的自然准确性始终低于标准训练更不用说 SAM 了。同时，SAM 在计算成本方面也优于 AT。因此，我们最终回答 RQ2 的结论是，在以下要求下，SAM 可以被认为是 AT 在提高对抗鲁棒性方面的轻量级替代品：(1)不损失自然准确性；(2)计算成本不会显着增加
总而言之，我们在本文中的主要贡献是：
1. We point out that using SAM alone can notably en-
hance adversarial robustness without sacrificing clean accuracy compared to standard training, which is an unexpected benefit.
2. We provide both empirical and theoretical explana-
tions to clarify how SAM can enhance adversarial robustness. In particular, we discuss the relation between SAM and AT and demonstrate that they improve adversarial robustness by eliminating nonrobust features. However, they differ in perturbation strengths, which leads to different trade-offs between natural and robust accuracy.
3. We conducted experiments on benchmark datasets to
verify our proposed insight. We also suggest that SAM can be considered a lightweight substitute for AT under certain requirements.
# 2. Background and related work
2.1. Sharpness awareness minimization (SAM)
为了处理传统机器学习算法中的不良泛化问题(Hochreiter& Schmidhuber, 1994; 1997) 分别尝试在损失景观中搜索平坦极小值和惩罚锐度，在泛化中获得了良好的效果$\boxed{\mathrm{E}}($Keskar et al.,2016; Neyshabur et al., 2017; Dziugaite & Roy, 2017)。受此启发，一系列作品着重于在损失景观中使用平坦度或清晰度的概念，以确保更好的泛化，例如。g。熵-SGD(Chaudhari et al., 2019) 与随机加权平均 (SWA) (Izmailov et al., 2018) 。锐度感知最小化 (SAM) (Foret et al.,2020) 也属于这一类，它同时最小化损耗值和损耗锐度，如(1)中所述从理论上讲，SAM 的良好泛化能力是由以下事实保证的：在高概率中，以下不等式成立：
(2)
$$L_{\mathscr{D}}(\boldsymbol{w})\leq\max_{\|\boldsymbol{\epsilon}\|_2\leq\rho}L_{\mathcal{S}}(\boldsymbol{w}+\boldsymbol{\epsilon})+h\left(\|\boldsymbol{w}\|_2^2/\rho^2\right),$$
其中 set $\mathcal{S}$是从分布$\mathscr{D}$生成的$,h:\mathbb{R}_+\to\mathbb{R}_+$是一个
严格递增的函数
SAM 在其他研究领域也有许多应用，例如语言模型(Bahri et al., 2021) 和流体动力学 (Jetly et al., 2022) , 显
示了 SAM 在各个领域的可扩展性。此外，SAM 算法的许多改进如自适应 SAM(ASAM) (Kwon et al., 2021) 、高效 SAM(ESAM) (Du et al.,2021) , LookSAM (Liu et al.,2022)、稀疏 SAM(SSAM) (Mi et al.,2022) , FisherSAM (Kim et al.,2022) 和 FSAM (Zhong et al., 2022) ,进一步改进了 SAM 和
## 2.2. Adversarial robustness
自从发现对抗性示例E(Szegedy et al.,2013; Goodfellow et al.,2014)以来，对抗性鲁棒性和对抗性训练已成为热门的研究主题，它发现 DNN 很容易被对抗性示例愚弄，从而做出错误的决策，而对抗性示例是通过在正常示例中添加小扰动而制作的。恶意对手可以通过制作对抗性示例来进行对抗性攻击$\mathbb{E}($Chen et al., 2023b; Wei et al.,2023b) , 这会导致有关 DNN 部署的严重安全问题。到目前为止，已经提出了许多防御方法(Papernot et al., 2016; Xie et al., 2019; Bai et al., 2019; Mo et al., 2022; Chen et al., 2023a) , 其中对抗训练 (AT) (Madry et al., 2017; Wang et al., 2019) 被认为是最有前途的防御对抗攻击的方法，可以用
(3)
$$\min_{w}\mathbb{E}_{(x,y)\sim\mathcal{D}}\max_{\|\delta\|\leq\epsilon}L(w;x+\delta,y),$$
来表示，其中$\mathcal{D}$是数据分布，$\epsilon$是扰动的余量，$w$的函数是。g。交叉熵损失)。对于内部最大化过程，通常使用投影梯度下降(PGD)攻击来生成对抗性示例：
$$x^{t+1}=\Pi_{\mathcal{B}(x,\epsilon)}(x^t+\alpha\cdot\mathrm{sign}(\nabla_x\ell(\theta;x^t,y))),$$
,其中$\Pi$将对抗性示例投影到扰动边界$\mathcal{B}(x,\epsilon)=\{x^\prime:$
$\|x'-x\|_p\leq\epsilon\}$上$,\alpha$ 表示梯度上升的步长
虽然有效地提高了对抗鲁棒性，但对抗训练暴露了一些缺陷，例如计算开销$\overline{\mathrm{E}}$(Shafahi et al.,2019),类公平性$\boxed{\mathrm{F}}($Xu et al., 2021; Wei et al., 2023a) , 其中自然精度降低$\overline{\mathbb{E}}$(Tsipras et al., 2018; Wang & Wang, 2023) 已成为主要问题。证明了鲁棒性和自然精度$\overline{\mathrm{E}}($Tsipras et al., 2018)之间存在内在的权衡，这可以解释为什么 AT 会大大降低标准精度
在对抗鲁棒性的背景下，也有一些作品试图在对抗训练中引入平坦的损失景观 (Wu et al.,2020; Yu et al., 2022a;b)。最具代表性的是对抗权重扰动(AWP)E (Wu et al., 2020) , 它同时在示例和特征空间上添加扰动，以在对抗训练上应用锐度感知最小化。然而，AWP 也受到自然精度下降的影响。。此外，平坦的损失景观可以带来更好的稳健性的原因还没有得到很好的解释据我们所知，我们是第一个发现 SAM 和 AT 之间的内在联系的人，我们揭示了 SAM 可以通过隐式地将更多的权重偏置在健壮的特征上来提高对抗性
# 3. Empirical understanding
在本节中，我们介绍了我们提出的关于 SAM 和 AT 之间关系的经验理解，这可以解释 SAM 如何提高对抗性
|
[
[
0.25,
0.05682373046875,
0.7236328125,
0.058074951171875,
0.7236328125,
0.07135009765625,
0.25,
0.0706787109375
],
[
0.09234619140625,
0.08837890625,
0.47216796875,
0.08837890625,
0.47216796875,
0.0985107421875,
0.09234619140625,
0.0985107421875
],
[
0.50244140625,
0.08837890625,
0.8837890625,
0.08837890625,
0.8837890625,
0.0985107421875,
0.50244140625,
0.0985107421875
],
[
0.08905029296875,
0.10162353515625,
0.2059326171875,
0.10162353515625,
0.2059326171875,
0.11553955078125,
0.08905029296875,
0.11553955078125
],
[
0.5009765625,
0.10162353515625,
0.8857421875,
0.10162353515625,
0.8857421875,
0.11553955078125,
0.5009765625,
0.11553955078125
],
[
0.50244140625,
0.117431640625,
0.8857421875,
0.117431640625,
0.8857421875,
0.1307373046875,
0.50244140625,
0.1307373046875
],
[
0.08905029296875,
0.125,
0.473876953125,
0.125,
0.473876953125,
0.1383056640625,
0.08905029296875,
0.1383056640625
],
[
0.5009765625,
0.1319580078125,
0.8857421875,
0.1319580078125,
0.8857421875,
0.1458740234375,
0.5009765625,
0.1458740234375
],
[
0.0906982421875,
0.141357421875,
0.47216796875,
0.141357421875,
0.47216796875,
0.1514892578125,
0.0906982421875,
0.1514892578125
],
[
0.498291015625,
0.1458740234375,
0.8857421875,
0.146484375,
0.8857421875,
0.1610107421875,
0.498291015625,
0.160400390625
],
[
0.0906982421875,
0.1546630859375,
0.473876953125,
0.1546630859375,
0.473876953125,
0.1685791015625,
0.0906982421875,
0.1685791015625
],
[
0.5,
0.1622314453125,
0.67626953125,
0.1622314453125,
0.67626953125,
0.175537109375,
0.5,
0.175537109375
],
[
0.09234619140625,
0.1717529296875,
0.47216796875,
0.1717529296875,
0.47216796875,
0.1817626953125,
0.09234619140625,
0.1817626953125
],
[
0.0906982421875,
0.184326171875,
0.47216796875,
0.184326171875,
0.47216796875,
0.1976318359375,
0.0906982421875,
0.1976318359375
],
[
0.0906982421875,
0.199462890625,
0.473876953125,
0.199462890625,
0.473876953125,
0.21337890625,
0.0906982421875,
0.21337890625
],
[
0.5009765625,
0.1944580078125,
0.6953125,
0.1944580078125,
0.6953125,
0.2083740234375,
0.5009765625,
0.2083740234375
],
[
0.0906982421875,
0.216552734375,
0.2034912109375,
0.216552734375,
0.2034912109375,
0.2266845703125,
0.0906982421875,
0.2266845703125
],
[
0.50341796875,
0.2171630859375,
0.8837890625,
0.2171630859375,
0.8837890625,
0.2310791015625,
0.50341796875,
0.2310791015625
],
[
0.0906982421875,
0.238037109375,
0.373291015625,
0.238037109375,
0.373291015625,
0.251953125,
0.0906982421875,
0.251953125
],
[
0.50341796875,
0.2342529296875,
0.8837890625,
0.2342529296875,
0.8837890625,
0.2442626953125,
0.50341796875,
0.2442626953125
],
[
0.5009765625,
0.2474365234375,
0.8837890625,
0.2474365234375,
0.8837890625,
0.261474609375,
0.5009765625,
0.261474609375
],
[
0.1004638671875,
0.26513671875,
0.47216796875,
0.26513671875,
0.47216796875,
0.279052734375,
0.1004638671875,
0.279052734375
],
[
0.50341796875,
0.264404296875,
0.8837890625,
0.264404296875,
0.8837890625,
0.274658203125,
0.50341796875,
0.274658203125
],
[
0.1192626953125,
0.27978515625,
0.47216796875,
0.27978515625,
0.47216796875,
0.29345703125,
0.1192626953125,
0.29345703125
],
[
0.50341796875,
0.27978515625,
0.8837890625,
0.27978515625,
0.8837890625,
0.290283203125,
0.50341796875,
0.290283203125
],
[
0.12091064453125,
0.296630859375,
0.473876953125,
0.29541015625,
0.473876953125,
0.31005859375,
0.12091064453125,
0.311279296875
],
[
0.5009765625,
0.29345703125,
0.8857421875,
0.29345703125,
0.8857421875,
0.306884765625,
0.5009765625,
0.306884765625
],
[
0.1217041015625,
0.313232421875,
0.2484130859375,
0.3125,
0.2484130859375,
0.323974609375,
0.1217041015625,
0.324462890625
],
[
0.50341796875,
0.31005859375,
0.8837890625,
0.31005859375,
0.8837890625,
0.32080078125,
0.50341796875,
0.32080078125
],
[
0.5009765625,
0.322509765625,
0.8857421875,
0.323974609375,
0.8857421875,
0.337158203125,
0.5009765625,
0.3359375
],
[
0.1004638671875,
0.335205078125,
0.473876953125,
0.335205078125,
0.473876953125,
0.348388671875,
0.1004638671875,
0.348388671875
],
[
0.5,
0.337158203125,
0.8857421875,
0.33837890625,
0.8857421875,
0.3515625,
0.5,
0.35107421875
],
[
0.12335205078125,
0.3515625,
0.47216796875,
0.3515625,
0.47216796875,
0.3623046875,
0.12335205078125,
0.3623046875
],
[
0.5,
0.3515625,
0.8837890625,
0.352783203125,
0.8837890625,
0.367431640625,
0.5,
0.3662109375
],
[
0.12335205078125,
0.367431640625,
0.471435546875,
0.367431640625,
0.471435546875,
0.37744140625,
0.12335205078125,
0.37744140625
],
[
0.5009765625,
0.36865234375,
0.8857421875,
0.36865234375,
0.8857421875,
0.3818359375,
0.5009765625,
0.3818359375
],
[
0.1192626953125,
0.37890625,
0.47216796875,
0.380126953125,
0.47216796875,
0.39453125,
0.1192626953125,
0.393310546875
],
[
0.50341796875,
0.385009765625,
0.6787109375,
0.385009765625,
0.6787109375,
0.395751953125,
0.50341796875,
0.395751953125
],
[
0.12091064453125,
0.39453125,
0.47216796875,
0.395751953125,
0.47216796875,
0.4091796875,
0.12091064453125,
0.40771484375
],
[
0.576171875,
0.40478515625,
0.81396484375,
0.40478515625,
0.81396484375,
0.42236328125,
0.576171875,
0.42236328125
],
[
0.86279296875,
0.406005859375,
0.8857421875,
0.406005859375,
0.8857421875,
0.419189453125,
0.86279296875,
0.419189453125
],
[
0.82421875,
0.410888671875,
0.8408203125,
0.410888671875,
0.8408203125,
0.415283203125,
0.82421875,
0.415283203125
],
[
0.12091064453125,
0.410400390625,
0.47216796875,
0.410400390625,
0.47216796875,
0.423583984375,
0.12091064453125,
0.423583984375
],
[
0.58642578125,
0.419189453125,
0.5986328125,
0.419189453125,
0.5986328125,
0.425537109375,
0.58642578125,
0.425537109375
],
[
0.6708984375,
0.417236328125,
0.70751953125,
0.417236328125,
0.70751953125,
0.427978515625,
0.6708984375,
0.427978515625
],
[
0.12091064453125,
0.425537109375,
0.47216796875,
0.4248046875,
0.47216796875,
0.438232421875,
0.12091064453125,
0.439453125
],
[
0.12091064453125,
0.43994140625,
0.3056640625,
0.44140625,
0.3056640625,
0.455078125,
0.12091064453125,
0.453857421875
],
[
0.50341796875,
0.43505859375,
0.8837890625,
0.43505859375,
0.8837890625,
0.448974609375,
0.50341796875,
0.448974609375
],
[
0.50244140625,
0.45068359375,
0.8857421875,
0.45068359375,
0.8857421875,
0.464111328125,
0.50244140625,
0.464111328125
],
[
0.1004638671875,
0.46337890625,
0.473876953125,
0.46337890625,
0.473876953125,
0.4765625,
0.1004638671875,
0.4765625
],
[
0.5009765625,
0.46533203125,
0.8486328125,
0.46533203125,
0.8486328125,
0.479248046875,
0.5009765625,
0.479248046875
],
[
0.12091064453125,
0.479248046875,
0.473876953125,
0.479248046875,
0.473876953125,
0.49365234375,
0.12091064453125,
0.49365234375
],
[
0.12091064453125,
0.492431640625,
0.473876953125,
0.49365234375,
0.473876953125,
0.50830078125,
0.12091064453125,
0.5068359375
],
[
0.53759765625,
0.5,
0.88623046875,
0.501953125,
0.88623046875,
0.51953125,
0.53759765625,
0.517578125
],
[
0.1217041015625,
0.5087890625,
0.325927734375,
0.5087890625,
0.325927734375,
0.52294921875,
0.1217041015625,
0.52294921875
],
[
0.5087890625,
0.52783203125,
0.515625,
0.52783203125,
0.515625,
0.5322265625,
0.5087890625,
0.5322265625
],
[
0.5146484375,
0.52294921875,
0.88623046875,
0.52294921875,
0.88623046875,
0.5361328125,
0.5146484375,
0.5361328125
],
[
0.5009765625,
0.5361328125,
0.8037109375,
0.537109375,
0.8037109375,
0.5517578125,
0.5009765625,
0.55029296875
],
[
0.08905029296875,
0.54443359375,
0.365234375,
0.54443359375,
0.365234375,
0.55810546875,
0.08905029296875,
0.55810546875
],
[
0.50244140625,
0.5595703125,
0.88134765625,
0.5595703125,
0.88134765625,
0.57275390625,
0.50244140625,
0.57275390625
],
[
0.08905029296875,
0.57080078125,
0.4150390625,
0.57080078125,
0.4150390625,
0.58056640625,
0.08905029296875,
0.58056640625
],
[
0.5009765625,
0.57373046875,
0.8857421875,
0.57470703125,
0.8857421875,
0.5888671875,
0.5009765625,
0.58837890625
],
[
0.0906982421875,
0.59326171875,
0.4755859375,
0.59326171875,
0.4755859375,
0.60693359375,
0.0906982421875,
0.60693359375
],
[
0.5009765625,
0.59033203125,
0.8857421875,
0.59033203125,
0.8857421875,
0.60400390625,
0.5009765625,
0.60400390625
],
[
0.0906982421875,
0.60986328125,
0.47216796875,
0.60986328125,
0.47216796875,
0.62060546875,
0.0906982421875,
0.62060546875
],
[
0.5009765625,
0.60595703125,
0.8857421875,
0.60595703125,
0.8857421875,
0.619140625,
0.5009765625,
0.619140625
],
[
0.0906982421875,
0.62255859375,
0.473876953125,
0.62255859375,
0.473876953125,
0.63623046875,
0.0906982421875,
0.63623046875
],
[
0.5009765625,
0.62060546875,
0.8837890625,
0.62060546875,
0.8837890625,
0.6337890625,
0.5009765625,
0.6337890625
],
[
0.0906982421875,
0.63818359375,
0.47216796875,
0.63818359375,
0.47216796875,
0.6513671875,
0.0906982421875,
0.6513671875
],
[
0.5,
0.6337890625,
0.8857421875,
0.63525390625,
0.8857421875,
0.6494140625,
0.5,
0.6484375
],
[
0.08905029296875,
0.65283203125,
0.47216796875,
0.65283203125,
0.47216796875,
0.66650390625,
0.08905029296875,
0.66650390625
],
[
0.5009765625,
0.65087890625,
0.63232421875,
0.65087890625,
0.63232421875,
0.6640625,
0.5009765625,
0.6640625
],
[
0.08905029296875,
0.66748046875,
0.473876953125,
0.66845703125,
0.473876953125,
0.681640625,
0.08905029296875,
0.68115234375
],
[
0.50244140625,
0.6748046875,
0.8837890625,
0.6748046875,
0.8837890625,
0.68505859375,
0.50244140625,
0.68505859375
],
[
0.0906982421875,
0.6845703125,
0.47216796875,
0.6845703125,
0.47216796875,
0.69775390625,
0.0906982421875,
0.69775390625
],
[
0.5009765625,
0.68798828125,
0.8837890625,
0.68798828125,
0.8837890625,
0.7021484375,
0.5009765625,
0.7021484375
],
[
0.08905029296875,
0.69873046875,
0.4755859375,
0.69873046875,
0.4755859375,
0.7119140625,
0.08905029296875,
0.7119140625
],
[
0.5,
0.70263671875,
0.8857421875,
0.7021484375,
0.8857421875,
0.71630859375,
0.5,
0.71728515625
],
[
0.0906982421875,
0.71337890625,
0.459228515625,
0.71337890625,
0.459228515625,
0.7265625,
0.0906982421875,
0.7265625
],
[
0.4599609375,
0.71728515625,
0.468017578125,
0.71728515625,
0.468017578125,
0.7216796875,
0.4599609375,
0.7216796875
],
[
0.5,
0.71728515625,
0.8857421875,
0.71826171875,
0.8857421875,
0.73193359375,
0.5,
0.73095703125
],
[
0.08905029296875,
0.72900390625,
0.468017578125,
0.72900390625,
0.468017578125,
0.7421875,
0.08905029296875,
0.7421875
],
[
0.50244140625,
0.734375,
0.8857421875,
0.734375,
0.8857421875,
0.74755859375,
0.50244140625,
0.74755859375
],
[
0.0906982421875,
0.751953125,
0.473876953125,
0.751953125,
0.473876953125,
0.76513671875,
0.0906982421875,
0.76513671875
],
[
0.50244140625,
0.75,
0.8857421875,
0.75,
0.8857421875,
0.76318359375,
0.50244140625,
0.76318359375
],
[
0.09234619140625,
0.76806640625,
0.339111328125,
0.76806640625,
0.339111328125,
0.77880859375,
0.09234619140625,
0.77880859375
],
[
0.50244140625,
0.76318359375,
0.8759765625,
0.76318359375,
0.8759765625,
0.7763671875,
0.50244140625,
0.7763671875
],
[
0.5009765625,
0.7861328125,
0.88232421875,
0.7861328125,
0.88232421875,
0.7998046875,
0.5009765625,
0.7998046875
],
[
0.12744140625,
0.8037109375,
0.473876953125,
0.8037109375,
0.473876953125,
0.818359375,
0.12744140625,
0.818359375
],
[
0.50244140625,
0.8017578125,
0.8837890625,
0.8017578125,
0.8837890625,
0.81494140625,
0.50244140625,
0.81494140625
],
[
0.1993408203125,
0.81494140625,
0.24267578125,
0.81494140625,
0.24267578125,
0.8271484375,
0.1993408203125,
0.8271484375
],
[
0.5009765625,
0.81494140625,
0.7958984375,
0.81640625,
0.7958984375,
0.83056640625,
0.5009765625,
0.830078125
],
[
0.0906982421875,
0.8388671875,
0.47216796875,
0.8388671875,
0.47216796875,
0.85302734375,
0.0906982421875,
0.85302734375
],
[
0.5,
0.849609375,
0.73779296875,
0.8515625,
0.73779296875,
0.8662109375,
0.5,
0.8642578125
],
[
0.09234619140625,
0.85595703125,
0.205078125,
0.85595703125,
0.205078125,
0.8662109375,
0.09234619140625,
0.8662109375
],
[
0.08905029296875,
0.876953125,
0.47216796875,
0.876953125,
0.47216796875,
0.89013671875,
0.08905029296875,
0.89013671875
],
[
0.5009765625,
0.876953125,
0.88232421875,
0.876953125,
0.88232421875,
0.89013671875,
0.5009765625,
0.89013671875
],
[
0.08905029296875,
0.892578125,
0.473876953125,
0.892578125,
0.473876953125,
0.90576171875,
0.08905029296875,
0.90576171875
],
[
0.5009765625,
0.892578125,
0.8857421875,
0.892578125,
0.8857421875,
0.90576171875,
0.5009765625,
0.90576171875
]
] |
[
"Sharpness-Aware Minimization Alone can Improve Adversarial Robustness",
"更好的自然准确性。综上所述，我们的经验和理论理",
"示了SAM在各个领域的可扩展性。此外，SAM算法",
"解可以回答RQ1",
"的许多改进如自适应SAM(ASAM）（Kwonetal,2021)",
"、高效SAM(ESAM)(Duetal.,2021)、LookSAM(Liu",
"值得注意的是，SAM和AT之间的健壮性仍然存在很",
"et al.,2022)、稀疏 SAM(SSAM)(Miet al.,2022)、Fish-",
"大差距。然而，AT的自然准确性始终低于标准训练",
"erSAM (Kim et al., 2022) 和 FSAM (Zhong et al., 2022)",
"更不用说SAM了。同时，SAM在计算成本方面也优",
"，进一步改进了SAM和",
"于AT。因此，我们最终回答RO2的结论是，在以下要",
"求下，SAM可以被认为是AT在提高对抗鲁棒性方面",
"的轻量级替代品：（1）不损失自然准确性；（2）计算成",
"2.2.Adversarial robustness",
"本不会显着增加",
"自从发现对抗性示例ESzegedyetal.，2013Goodfellow",
"总而言之，我们在本文中的主要贡献是：",
"etal.，2014）以来，对抗性鲁棒性和对抗性训练已成为",
"热门的研究主题，它发现DNN很容易被对抗性示例愚",
"1. We point out that using SAM alone can notably en.",
"弄，从而做出错误的决策，而对抗性示例是通过在正常",
"hance adversarial robustness without sacrificing clean",
"示例中添加小扰动而制作的。恶意对手可以通过制作",
"accuracy compared to standard training, which is an",
"对抗性示例来进行对抗性攻击E(Chenetal.，2023b；Wei",
"unexpectedbenefit.",
"etal.，2023b），这会导致有关DNN部署的严重安全问",
"题。到目前为止，已经提出了许多防御方法(Papernot",
"2. We provide both empirical and theoretical explana-",
"et al., 2016; Xie et al., 2019; Bai et al., 2019; Mo et al.,",
"tions to clarify howSAM can enhance adversarial",
"2022;Chenetal.，2023a），其中对抗训练（AT）（Madry",
"robustness.Inparticular,wediscuss the relation",
"etal.，2017；Wangetal.，2019）被认为是最有前途的防御",
"between SAM and AT and demonstrate that they",
"对抗攻击的方法，可以用",
"improve adversarial robustness by eliminating non",
"min E(r,y)~D max L(w; r + d, y),",
"(3)",
"",
"robust features. However, they differ in perturbation",
"w",
"I0l≤e",
"strengths,which leads to differenttrade-offsbetween",
"natural and robust accuracy.",
"来表示，其中D是数据分布，E是扰动的余量，W的函",
"数是。g。交叉损失)。对于内部最大化过程，通常使",
"3. We conducted experiments on benchmark datasets to",
"用投影梯度下降（PGD）攻击来生成对抗性示例：",
"verify our proposed insight. We also suggest that",
"SAM can be considered a lightweight substitute for",
"rt+1 = IB(a,e)(at + Q · sign(Vrl(0; at, y), (4)",
"AT under certain requirements.",
"",
"其中II将对抗性示例投影到扰动边界B(α,e)=c：",
"a-p≤e上，α表示梯度上升的步长",
"2. Background and related work",
"虽然有效地提高了对抗鲁棒性，但对抗训练暴露了一",
"2.1.Sharpness awareness minimization (SAM)",
"些缺陷，例如计算开销E(Shafahietal.，2019），类公平",
"为了处理传统机器学习算法中的不良泛化问题，",
"性E(Xuetal.,2021;Weietal.2023a），其中自然精度",
"（Hochreiter&Schmidhuber，1994；1997）分别尝试在损",
"降低E（Tsiprasetal.,2018;Wang&Wang，2023）已成为",
"失景观中搜索平坦极小值和惩罚锐度，在泛化中获",
"主要问题。证明了鲁棒性和自然精度E(Tsipras et al.,",
"得了良好的效果E（Keskaretal.，2016;Neyshaburetal.，",
"2018)之间存在内在的权衡，这可以解释为什么AT会",
"2017;Dziugaite&Roy，2017）。受此启发，一系列作品",
"大大降低标准精度",
"着重于在损失景观中使用平坦度或清晰度的概念，以",
"在对抗鲁棒性的背景下，也有一些作品试图在对抗训",
"确保更好的泛化，例如。g。炳-SGD（Chaudharietal.",
"练中引人平坦的损失景观（Wuetal.，2020；Yuetal.",
"2019）与随机加权平均（SWA）（Izmailovetal.，2018）。",
"2022a;b）。最具代表性的是对抗权重扰动（AWP）E",
"锐度感知最小化（SAM）（Foretetal.，2020）也属于这",
"-",
"(Wuetal.2020），它同时在示例和特征空间上添加扰",
"类，它同时最小化损耗值和损耗锐度，如（1）中所述",
"动，以在对抗训练上应用锐度感知最小化。然而，AWP",
"从理论上讲，SAM的良好泛化能力是由以下事实保证",
"也受到自然精度下降的影响。。此外，平坦的损失景观",
"的：在高概率中，以下不等式成立",
"可以带来更好的稳健性的原因还没有得到很好的解释",
"据我们所知，我们是第一个发现SAM和AT之间的内",
"L(w) ≤ max Ls(w+e)+h (llwl2/p2), (2)",
"在联系的人，我们揭示了SAM可以通过隐式地将更多",
"Ilell2≤p",
"的权重偏置在健壮的特征上来提高对抗性",
"其中setS是从分布の生成的，h：R+→R+是一个",
"3.Empirical understanding",
"严格递增的函数",
"SAM在其他研究领域也有许多应用，例如语言模型",
"在本节中，我们介绍了我们提出的关于SAM和AT之",
"(Bahrietal.,2021）和流体动力学（Jetlyetal.，2022），显",
"间关系的经验理解，这可以解释SAM如何提高对抗性"
] |
{
"arxiv_id": "2305.05392",
"language": "zh",
"page_idx": 1
}
|
||
2305.05392-zh-3
|
2305.05392-zh-3.jpg
|
Sharpness-Aware Minimization Alone can Improve Adversarial Robustness
Table $l.$ Natural and Robust Accuracy evaluation on CIFAR-100 dataset.
<table>
<tbody>
<tr>
<th> </th>
<th rowspan="2">Natural Accuracy</th>
<th colspan="2">$\ell_{\infty}$ -Robust Accuracy</th>
<th colspan="2">$\ell_{2}$ -Robust Accuracy</th>
</tr>
<tr>
<th>Method</th>
<th>$\epsilon=1/255$</th>
<th>$\epsilon=$ 2/255 </th>
<th>16/255 $\epsilon=$</th>
<th>32/255 $\epsilon=$</th>
</tr>
<tr>
<td>$ST$</td>
<td>76.9</td>
<td>13.6</td>
<td>1.7</td>
<td>44.5</td>
<td>21.2</td>
</tr>
<tr>
<td>$\operatorname { SAM} ($ $\rho = 0. 1$ )</td>
<td>78.0</td>
<td>19.6</td>
<td>3.0</td>
<td>51.5</td>
<td>27.2</td>
</tr>
<tr>
<td>$\operatorname { SAM} ($ $\rho = 0. 2$ )</td>
<td>78.5</td>
<td>23.1</td>
<td>4.2</td>
<td>54.2</td>
<td>31.3</td>
</tr>
<tr>
<td>$\operatorname { SAM} ($ $\rho = 0. 4$ )</td>
<td>78.7</td>
<td>28.3</td>
<td>6.5</td>
<td>57.0</td>
<td>36.2</td>
</tr>
<tr>
<td>$\operatorname{AT}\left(\:\ell_{\infty}-\epsilon=1/255\:\right)$ </td>
<td>73.1</td>
<td>60.4</td>
<td>46.6</td>
<td>67.4</td>
<td>61.5</td>
</tr>
<tr>
<td>AT$\left ( \:\ell _{\infty }\right )$ $- \epsilon = 2/ 255$ )</td>
<td>70.1</td>
<td>60.3</td>
<td>50.6</td>
<td>65.7</td>
<td>60.6</td>
</tr>
<tr>
<td>AT$\left ( \ell _{\infty }\right )$ $- \epsilon = 4/ 255$ )</td>
<td>66.2</td>
<td>59.3</td>
<td>52.0</td>
<td>62.8</td>
<td>58.9</td>
</tr>
<tr>
<td>$AT\left(\:\ell_{\infty}\right.$ $-\epsilon=$ =8/255 ) </td>
<td>60.4</td>
<td>55.1</td>
<td>50.4</td>
<td>57.0</td>
<td>54.3</td>
</tr>
<tr>
<td>${\mathrm{AT}}\left(\:\ell_{2}-\epsilon=16$, /255 )</td>
<td>74.8</td>
<td>52.8</td>
<td>31.4</td>
<td>66.3</td>
<td>57.7</td>
</tr>
<tr>
<td>$\operatorname{AT}\left(\:\ell_{2}-\epsilon=32/255\:\right)$</td>
<td>73.2</td>
<td>57.4</td>
<td>40.6</td>
<td>67.1</td>
<td>61.0</td>
</tr>
<tr>
<td>$\mathop{\mathrm{AT}}\left(\:\ell_{2}-\epsilon=64/255\:\right)$</td>
<td>70.7</td>
<td>58.1</td>
<td>45.9</td>
<td>66.1</td>
<td>60.7</td>
</tr>
<tr>
<td>${\mathrm{AT}}\left(\:\ell_{2}-\epsilon=128$, /255</td>
<td>67.4</td>
<td>58.2</td>
<td>48.7</td>
<td>63.9</td>
<td>60.4</td>
</tr>
</tbody>
</table>
Theorem 4.4. In the classification problem above, denote the best parameter for SAM training $w_1^{SAM}\cdot Suppose$ that $\epsilon$ $is$ $small$, $we$ have $w_{1}^{SAM}\approx w_{1}^{* }+ \frac 23\tilde {w_{1}^{* }\epsilon ^{2}}$ .
与标准训练类似，还存在一个由对抗性训练返回的最
佳参数$w_1^{AT}$,可以用以下定理表示：
Theorem 4.2 (Adversarial training). In the classification problem above, under adversarial training with perturbation bound $\epsilon<\eta$ , the adversarial optimal parameter value
## 4.5. Relation between SAM and AT
(10)
$$w_1^{AT}=\frac{\ln p-\ln(1-p)}{2(\eta-\epsilon)}.$$
我们进一步讨论 AT 和 SAM 之间不同的攻击(扰动)强度。回想一下，在我们在$\mathbb{E}$3 节中的经验理解中，SAM 的扰动比 AT 更温和，这可以解释为 SAM 在鲁棒性-准确性权衡中更多地关注自然准确性而较少地关注鲁棒性。因此，为了达到相同的鲁棒性水平 (这是由对特征$x_1$的依赖性来衡量的，i。e。$w_1$)的大小，SAM 需要更大的扰动范围，而对于 AT, $x$上的较小扰动就足够了。从理论上讲，以下定理验证了我们的陈述：
我们可以看到$w_1$已乘以$\frac\eta{\eta-\epsilon},$,这增加了对分类器的强大功能$x_1$的依赖性。这表明，与标准训练的模型相比，对抗训练的模型更加注重鲁棒性，从而提高了对抗鲁棒性
4.4. Sharpness-Aware Minimization (SAM)
Theorem 4.5. Denote the perturbation range $\epsilon \textit{of AT and}$ SAM as $\epsilon_{AT}$ and $\epsilon_{SAM}$, respectively. Then, when both methods return the same parameter w 1 , we have the following relation between $\epsilon_{AT}$ and $\epsilon_SAM:$
现在我们考虑 SAM 的情况。回想一下，SAM 的优化
目标是
$$\mathbb{E}_{(x,y)\sim\mathcal{D}}\left[\max_{|\delta|\leq\epsilon}\mathcal{L}(x,y;w+\delta)\right].$$
(11)
(13)
$$2+\frac{3}{\epsilon_{SAM}^2}\approx\frac{2\eta}{\epsilon_{AT}}$$
我们首先通过证明用 SAM 训练的参数$w_1$也大于$w_1^*$ 来解释为什么 SAM 可以提高对抗鲁棒性，如下所示： Theorem $4.3\left(\text{Sharpness-aware minimization}\right).In$ the clas- $sificationproblemabove,thebestparameterforSAMtrain$ $\begin{array}{l}{suyuanonprovemabove,}\\{ingw_{1}^{SAM}satisfhes~that}\end{array}$
$$w_1^{SAM}>w_1^*.$$
(12)
从定理$\boxed{\mathrm{F}}4.5$,我们可以识别 AT 和 SAM 的不同扰动强度。从定理 4.5 可以很容易地得出，当(13)成立时$,\epsilon_{SAM}$大于$\epsilon_AT$,因为我们假设$\eta$是一个小的正， $\epsilon$在定理$\boxed{\mathrm{E}}4.4$ 和 $\epsilon_{AT}<\eta$ 在定理$\boxed{\mathrm{F}}4.2$ 。因此，要在强大的功能$x_1$上获得相同的权重$w_1$,只需选择比$\epsilon_{SAM}$小得多的$\epsilon_{AT}$。另一方面，在相同的扰动范围$\epsilon_{AT}=\epsilon_{SAM}$下，在 AT 下训练的模型具有比 SAM 更大的参数$w_1$,因此它着重于更高的鲁棒性，但降低了更多的自然精度
所有校样均可在附录 A 中找到。综上所述，我们可以得出结论，AT 利用显式和强扰动去噪非稳健特征，而SAM 利用隐式和中等扰动。这与我们在 3 节中的经验
从定理 4.2 和 4.3 我们可以看到$w_{1}^{AT}$和$w_{1}^{SAM}$都大于$w_{1}^{* }$ , 这表明 SAM 和 AT 都提高了训练模型的鲁棒性。但是，与对抗性训练相比，定性关系不足以量化 SAM 实现了多少鲁棒性，我们尝试通过以下定理定量估计$w_{1}^{SAM}$来进一步发展：
|
[
[
0.250732421875,
0.05682373046875,
0.7236328125,
0.05682373046875,
0.7236328125,
0.0706787109375,
0.250732421875,
0.0706787109375
],
[
0.268798828125,
0.09344482421875,
0.701171875,
0.09344482421875,
0.701171875,
0.10797119140625,
0.268798828125,
0.10797119140625
],
[
0.471435546875,
0.11236572265625,
0.615234375,
0.11236572265625,
0.615234375,
0.126953125,
0.471435546875,
0.126953125
],
[
0.66015625,
0.11236572265625,
0.79736328125,
0.11236572265625,
0.79736328125,
0.126953125,
0.66015625,
0.126953125
],
[
0.2059326171875,
0.12060546875,
0.2607421875,
0.12060546875,
0.2607421875,
0.1319580078125,
0.2059326171875,
0.1319580078125
],
[
0.325927734375,
0.11614990234375,
0.4453125,
0.11932373046875,
0.4453125,
0.1351318359375,
0.325927734375,
0.1319580078125
],
[
0.47216796875,
0.126953125,
0.53759765625,
0.126953125,
0.53759765625,
0.141357421875,
0.47216796875,
0.141357421875
],
[
0.54833984375,
0.1287841796875,
0.62158203125,
0.1287841796875,
0.62158203125,
0.14013671875,
0.54833984375,
0.14013671875
],
[
0.63818359375,
0.1287841796875,
0.7197265625,
0.1287841796875,
0.7197265625,
0.14013671875,
0.63818359375,
0.14013671875
],
[
0.736328125,
0.1287841796875,
0.81640625,
0.1287841796875,
0.81640625,
0.14013671875,
0.736328125,
0.14013671875
],
[
0.370849609375,
0.147705078125,
0.406005859375,
0.147705078125,
0.406005859375,
0.1634521484375,
0.370849609375,
0.1634521484375
],
[
0.48193359375,
0.147705078125,
0.51708984375,
0.147705078125,
0.51708984375,
0.1634521484375,
0.48193359375,
0.1634521484375
],
[
0.57275390625,
0.147705078125,
0.60107421875,
0.147705078125,
0.60107421875,
0.1641845703125,
0.57275390625,
0.1641845703125
],
[
0.6640625,
0.147705078125,
0.69775390625,
0.147705078125,
0.69775390625,
0.1634521484375,
0.6640625,
0.1634521484375
],
[
0.7587890625,
0.147705078125,
0.7939453125,
0.147705078125,
0.7939453125,
0.1634521484375,
0.7587890625,
0.1634521484375
],
[
0.370849609375,
0.1634521484375,
0.406005859375,
0.1634521484375,
0.406005859375,
0.1787109375,
0.370849609375,
0.1787109375
],
[
0.2197265625,
0.1483154296875,
0.2459716796875,
0.1483154296875,
0.2459716796875,
0.1622314453125,
0.2197265625,
0.1622314453125
],
[
0.48193359375,
0.1634521484375,
0.51708984375,
0.1634521484375,
0.51708984375,
0.1787109375,
0.48193359375,
0.1787109375
],
[
0.66259765625,
0.1634521484375,
0.69775390625,
0.1634521484375,
0.69775390625,
0.1787109375,
0.66259765625,
0.1787109375
],
[
0.7607421875,
0.1634521484375,
0.7939453125,
0.1634521484375,
0.7939453125,
0.1787109375,
0.7607421875,
0.1787109375
],
[
0.57421875,
0.1641845703125,
0.60302734375,
0.1641845703125,
0.60302734375,
0.1787109375,
0.57421875,
0.1787109375
],
[
0.176513671875,
0.1641845703125,
0.289306640625,
0.1641845703125,
0.289306640625,
0.1787109375,
0.176513671875,
0.1787109375
],
[
0.1732177734375,
0.1767578125,
0.2900390625,
0.1787109375,
0.2900390625,
0.1962890625,
0.1732177734375,
0.1944580078125
],
[
0.370849609375,
0.177978515625,
0.404296875,
0.177978515625,
0.404296875,
0.1932373046875,
0.370849609375,
0.1932373046875
],
[
0.48193359375,
0.177978515625,
0.515625,
0.177978515625,
0.515625,
0.1932373046875,
0.48193359375,
0.1932373046875
],
[
0.6640625,
0.177978515625,
0.69775390625,
0.177978515625,
0.69775390625,
0.1932373046875,
0.6640625,
0.1932373046875
],
[
0.7607421875,
0.177978515625,
0.7939453125,
0.177978515625,
0.7939453125,
0.1932373046875,
0.7607421875,
0.1932373046875
],
[
0.370849609375,
0.1932373046875,
0.404296875,
0.1932373046875,
0.404296875,
0.208984375,
0.370849609375,
0.208984375
],
[
0.57421875,
0.1787109375,
0.6005859375,
0.1787109375,
0.6005859375,
0.1925048828125,
0.57421875,
0.1925048828125
],
[
0.48046875,
0.1932373046875,
0.51708984375,
0.1932373046875,
0.51708984375,
0.208984375,
0.48046875,
0.208984375
],
[
0.66259765625,
0.1932373046875,
0.69775390625,
0.1932373046875,
0.69775390625,
0.208984375,
0.66259765625,
0.208984375
],
[
0.7587890625,
0.1932373046875,
0.7939453125,
0.1932373046875,
0.7939453125,
0.208984375,
0.7587890625,
0.208984375
],
[
0.57421875,
0.1944580078125,
0.6005859375,
0.1944580078125,
0.6005859375,
0.2083740234375,
0.57421875,
0.2083740234375
],
[
0.176513671875,
0.1944580078125,
0.289306640625,
0.1944580078125,
0.289306640625,
0.208984375,
0.176513671875,
0.208984375
],
[
0.370849609375,
0.2139892578125,
0.404296875,
0.2139892578125,
0.404296875,
0.2298583984375,
0.370849609375,
0.2298583984375
],
[
0.48193359375,
0.2139892578125,
0.51708984375,
0.2139892578125,
0.51708984375,
0.2298583984375,
0.48193359375,
0.2298583984375
],
[
0.5703125,
0.2139892578125,
0.60546875,
0.2139892578125,
0.60546875,
0.2298583984375,
0.5703125,
0.2298583984375
],
[
0.66259765625,
0.2139892578125,
0.69775390625,
0.2139892578125,
0.69775390625,
0.2298583984375,
0.66259765625,
0.2298583984375
],
[
0.7587890625,
0.2139892578125,
0.7939453125,
0.2139892578125,
0.7939453125,
0.2298583984375,
0.7587890625,
0.2298583984375
],
[
0.370849609375,
0.2291259765625,
0.404296875,
0.2291259765625,
0.404296875,
0.2442626953125,
0.370849609375,
0.2442626953125
],
[
0.158447265625,
0.21533203125,
0.306396484375,
0.21533203125,
0.306396484375,
0.2291259765625,
0.158447265625,
0.2291259765625
],
[
0.5703125,
0.2291259765625,
0.60546875,
0.2291259765625,
0.60546875,
0.2442626953125,
0.5703125,
0.2442626953125
],
[
0.66259765625,
0.2291259765625,
0.69775390625,
0.2291259765625,
0.69775390625,
0.2442626953125,
0.66259765625,
0.2442626953125
],
[
0.483642578125,
0.2310791015625,
0.515625,
0.2310791015625,
0.515625,
0.24365234375,
0.483642578125,
0.24365234375
],
[
0.158447265625,
0.2298583984375,
0.306396484375,
0.2298583984375,
0.306396484375,
0.2442626953125,
0.158447265625,
0.2442626953125
],
[
0.76123046875,
0.2310791015625,
0.79345703125,
0.2310791015625,
0.79345703125,
0.24365234375,
0.76123046875,
0.24365234375
],
[
0.370849609375,
0.2442626953125,
0.404296875,
0.2442626953125,
0.404296875,
0.260009765625,
0.370849609375,
0.260009765625
],
[
0.48193359375,
0.2442626953125,
0.51708984375,
0.2442626953125,
0.51708984375,
0.260009765625,
0.48193359375,
0.260009765625
],
[
0.5703125,
0.2442626953125,
0.60546875,
0.2442626953125,
0.60546875,
0.260009765625,
0.5703125,
0.260009765625
],
[
0.66259765625,
0.2442626953125,
0.69775390625,
0.2442626953125,
0.69775390625,
0.260009765625,
0.66259765625,
0.260009765625
],
[
0.7607421875,
0.2442626953125,
0.7939453125,
0.2442626953125,
0.7939453125,
0.260009765625,
0.7607421875,
0.260009765625
],
[
0.48046875,
0.2587890625,
0.515625,
0.2587890625,
0.515625,
0.274658203125,
0.48046875,
0.274658203125
],
[
0.158447265625,
0.24560546875,
0.3056640625,
0.24560546875,
0.3056640625,
0.2587890625,
0.158447265625,
0.2587890625
],
[
0.158447265625,
0.261474609375,
0.306396484375,
0.261474609375,
0.306396484375,
0.274658203125,
0.158447265625,
0.274658203125
],
[
0.370849609375,
0.261474609375,
0.404296875,
0.261474609375,
0.404296875,
0.2734375,
0.370849609375,
0.2734375
],
[
0.57177734375,
0.261474609375,
0.6044921875,
0.261474609375,
0.6044921875,
0.2734375,
0.57177734375,
0.2734375
],
[
0.6640625,
0.261474609375,
0.69677734375,
0.261474609375,
0.69677734375,
0.2734375,
0.6640625,
0.2734375
],
[
0.7587890625,
0.2587890625,
0.7939453125,
0.2587890625,
0.7939453125,
0.274658203125,
0.7587890625,
0.274658203125
],
[
0.370849609375,
0.281005859375,
0.404296875,
0.281005859375,
0.404296875,
0.296630859375,
0.370849609375,
0.296630859375
],
[
0.48046875,
0.281005859375,
0.51708984375,
0.281005859375,
0.51708984375,
0.296630859375,
0.48046875,
0.296630859375
],
[
0.66259765625,
0.281005859375,
0.69775390625,
0.281005859375,
0.69775390625,
0.296630859375,
0.66259765625,
0.296630859375
],
[
0.57177734375,
0.28271484375,
0.6044921875,
0.28271484375,
0.6044921875,
0.29541015625,
0.57177734375,
0.29541015625
],
[
0.1568603515625,
0.2822265625,
0.30810546875,
0.2822265625,
0.30810546875,
0.29541015625,
0.1568603515625,
0.29541015625
],
[
0.7607421875,
0.28271484375,
0.79345703125,
0.28271484375,
0.79345703125,
0.29541015625,
0.7607421875,
0.29541015625
],
[
0.370849609375,
0.29541015625,
0.404296875,
0.29541015625,
0.404296875,
0.3125,
0.370849609375,
0.3125
],
[
0.48193359375,
0.294189453125,
0.5185546875,
0.29736328125,
0.515625,
0.313232421875,
0.48046875,
0.311279296875
],
[
0.66259765625,
0.296630859375,
0.69677734375,
0.296630859375,
0.69677734375,
0.3125,
0.66259765625,
0.3125
],
[
0.57177734375,
0.29736328125,
0.6044921875,
0.29736328125,
0.6044921875,
0.31005859375,
0.57177734375,
0.31005859375
],
[
0.158447265625,
0.29736328125,
0.30810546875,
0.29736328125,
0.30810546875,
0.311279296875,
0.158447265625,
0.311279296875
],
[
0.7607421875,
0.29736328125,
0.79345703125,
0.29736328125,
0.79345703125,
0.31005859375,
0.7607421875,
0.31005859375
],
[
0.156005859375,
0.311279296875,
0.30810546875,
0.3125,
0.30810546875,
0.326904296875,
0.156005859375,
0.32568359375
],
[
0.370849609375,
0.311279296875,
0.404296875,
0.311279296875,
0.404296875,
0.326904296875,
0.370849609375,
0.326904296875
],
[
0.48193359375,
0.311279296875,
0.515625,
0.311279296875,
0.515625,
0.326904296875,
0.48193359375,
0.326904296875
],
[
0.5703125,
0.311279296875,
0.6044921875,
0.311279296875,
0.6044921875,
0.326904296875,
0.5703125,
0.326904296875
],
[
0.6640625,
0.313232421875,
0.6953125,
0.313232421875,
0.6953125,
0.32568359375,
0.6640625,
0.32568359375
],
[
0.7607421875,
0.311279296875,
0.7939453125,
0.311279296875,
0.7939453125,
0.326904296875,
0.7607421875,
0.326904296875
],
[
0.370849609375,
0.326904296875,
0.406005859375,
0.326904296875,
0.406005859375,
0.34228515625,
0.370849609375,
0.34228515625
],
[
0.48193359375,
0.326904296875,
0.51708984375,
0.326904296875,
0.51708984375,
0.34228515625,
0.48193359375,
0.34228515625
],
[
0.5703125,
0.326904296875,
0.60546875,
0.326904296875,
0.60546875,
0.34228515625,
0.5703125,
0.34228515625
],
[
0.66259765625,
0.326904296875,
0.69775390625,
0.326904296875,
0.69775390625,
0.34228515625,
0.66259765625,
0.34228515625
],
[
0.7587890625,
0.326904296875,
0.7939453125,
0.326904296875,
0.7939453125,
0.34228515625,
0.7587890625,
0.34228515625
],
[
0.5,
0.371826171875,
0.8837890625,
0.37255859375,
0.8837890625,
0.386962890625,
0.5,
0.386474609375
],
[
0.15283203125,
0.32763671875,
0.313720703125,
0.32763671875,
0.313720703125,
0.341552734375,
0.15283203125,
0.341552734375
],
[
0.09234619140625,
0.373779296875,
0.47216796875,
0.373779296875,
0.47216796875,
0.386962890625,
0.09234619140625,
0.386962890625
],
[
0.08905029296875,
0.38818359375,
0.351318359375,
0.386962890625,
0.351318359375,
0.401611328125,
0.08905029296875,
0.40283203125
],
[
0.496826171875,
0.38330078125,
0.88623046875,
0.38330078125,
0.88623046875,
0.404052734375,
0.496826171875,
0.404052734375
],
[
0.494384765625,
0.39453125,
0.7861328125,
0.396484375,
0.7861328125,
0.421630859375,
0.494384765625,
0.419189453125
],
[
0.08905029296875,
0.4072265625,
0.473876953125,
0.4072265625,
0.473876953125,
0.42041015625,
0.08905029296875,
0.42041015625
],
[
0.0882568359375,
0.42236328125,
0.473876953125,
0.42236328125,
0.473876953125,
0.436279296875,
0.0882568359375,
0.436279296875
],
[
0.0882568359375,
0.438232421875,
0.4755859375,
0.438232421875,
0.4755859375,
0.451904296875,
0.0882568359375,
0.451904296875
],
[
0.5009765625,
0.43505859375,
0.7509765625,
0.43505859375,
0.7509765625,
0.448974609375,
0.5009765625,
0.448974609375
],
[
0.188720703125,
0.4609375,
0.36279296875,
0.451904296875,
0.365234375,
0.477783203125,
0.191162109375,
0.4873046875
],
[
0.50341796875,
0.460205078125,
0.8837890625,
0.460205078125,
0.8837890625,
0.47021484375,
0.50341796875,
0.47021484375
],
[
0.443603515625,
0.4697265625,
0.473876953125,
0.4697265625,
0.473876953125,
0.481689453125,
0.443603515625,
0.481689453125
],
[
0.275390625,
0.479248046875,
0.336669921875,
0.479248046875,
0.336669921875,
0.49365234375,
0.275390625,
0.49365234375
],
[
0.50244140625,
0.473388671875,
0.8857421875,
0.473388671875,
0.8857421875,
0.4873046875,
0.50244140625,
0.4873046875
],
[
0.50244140625,
0.488037109375,
0.8857421875,
0.488037109375,
0.8857421875,
0.50244140625,
0.50244140625,
0.50244140625
],
[
0.50244140625,
0.50390625,
0.8857421875,
0.50390625,
0.8857421875,
0.517578125,
0.50244140625,
0.517578125
],
[
0.0882568359375,
0.517578125,
0.473876953125,
0.51953125,
0.473876953125,
0.537109375,
0.0882568359375,
0.53515625
],
[
0.50341796875,
0.52099609375,
0.88232421875,
0.52099609375,
0.88232421875,
0.53076171875,
0.50341796875,
0.53076171875
],
[
0.0906982421875,
0.53857421875,
0.473876953125,
0.53857421875,
0.473876953125,
0.5517578125,
0.0906982421875,
0.5517578125
],
[
0.5,
0.5341796875,
0.8837890625,
0.53271484375,
0.8837890625,
0.54736328125,
0.5,
0.548828125
],
[
0.0906982421875,
0.55322265625,
0.473876953125,
0.55322265625,
0.473876953125,
0.5673828125,
0.0906982421875,
0.5673828125
],
[
0.50244140625,
0.548828125,
0.8837890625,
0.548828125,
0.8837890625,
0.5625,
0.50244140625,
0.5625
],
[
0.08905029296875,
0.5693359375,
0.156005859375,
0.5693359375,
0.156005859375,
0.58056640625,
0.08905029296875,
0.58056640625
],
[
0.50341796875,
0.56494140625,
0.833984375,
0.56494140625,
0.833984375,
0.5791015625,
0.50341796875,
0.5791015625
],
[
0.5009765625,
0.58544921875,
0.88623046875,
0.58544921875,
0.88623046875,
0.5986328125,
0.5009765625,
0.5986328125
],
[
0.08905029296875,
0.60107421875,
0.393798828125,
0.60107421875,
0.393798828125,
0.6142578125,
0.08905029296875,
0.6142578125
],
[
0.5,
0.599609375,
0.8837890625,
0.599609375,
0.8837890625,
0.6142578125,
0.5,
0.6142578125
],
[
0.5009765625,
0.6162109375,
0.88623046875,
0.6162109375,
0.88623046875,
0.6298828125,
0.5009765625,
0.6298828125
],
[
0.0906982421875,
0.62353515625,
0.473876953125,
0.62353515625,
0.473876953125,
0.63720703125,
0.0906982421875,
0.63720703125
],
[
0.5,
0.6298828125,
0.77294921875,
0.630859375,
0.77294921875,
0.64599609375,
0.5,
0.64453125
],
[
0.08905029296875,
0.63818359375,
0.1405029296875,
0.63818359375,
0.1405029296875,
0.65283203125,
0.08905029296875,
0.65283203125
],
[
0.191162109375,
0.6611328125,
0.2115478515625,
0.6572265625,
0.2156982421875,
0.66943359375,
0.1953125,
0.673828125
],
[
0.2344970703125,
0.66015625,
0.373291015625,
0.66015625,
0.373291015625,
0.6748046875,
0.2344970703125,
0.6748046875
],
[
0.443603515625,
0.6611328125,
0.473876953125,
0.6611328125,
0.473876953125,
0.673828125,
0.443603515625,
0.673828125
],
[
0.67236328125,
0.66552734375,
0.705078125,
0.66552734375,
0.705078125,
0.68115234375,
0.67236328125,
0.68115234375
],
[
0.17724609375,
0.671875,
0.268798828125,
0.671875,
0.268798828125,
0.68310546875,
0.17724609375,
0.68310546875
],
[
0.6953125,
0.67041015625,
0.7490234375,
0.6611328125,
0.7548828125,
0.68115234375,
0.701171875,
0.68994140625
],
[
0.85302734375,
0.6748046875,
0.88818359375,
0.6748046875,
0.88818359375,
0.68994140625,
0.85302734375,
0.68994140625
],
[
0.6611328125,
0.67822265625,
0.708984375,
0.689453125,
0.701171875,
0.70849609375,
0.65283203125,
0.6962890625
],
[
0.6298828125,
0.67529296875,
0.6611328125,
0.67529296875,
0.6611328125,
0.6875,
0.6298828125,
0.6875
],
[
0.0882568359375,
0.689453125,
0.47216796875,
0.68994140625,
0.47216796875,
0.705078125,
0.0882568359375,
0.7041015625
],
[
0.72119140625,
0.68603515625,
0.75341796875,
0.68603515625,
0.75341796875,
0.69873046875,
0.72119140625,
0.69873046875
],
[
0.0906982421875,
0.705078125,
0.46728515625,
0.705078125,
0.46728515625,
0.71826171875,
0.0906982421875,
0.71826171875
],
[
0.5009765625,
0.71630859375,
0.8837890625,
0.71728515625,
0.8837890625,
0.73193359375,
0.5009765625,
0.73095703125
],
[
0.0906982421875,
0.72607421875,
0.473876953125,
0.72607421875,
0.473876953125,
0.736328125,
0.0906982421875,
0.736328125
],
[
0.50244140625,
0.73291015625,
0.8837890625,
0.73291015625,
0.8837890625,
0.7470703125,
0.50244140625,
0.7470703125
],
[
0.0882568359375,
0.7392578125,
0.4755859375,
0.73779296875,
0.4755859375,
0.75244140625,
0.0882568359375,
0.75390625
],
[
0.08416748046875,
0.75,
0.25244140625,
0.74755859375,
0.25244140625,
0.76953125,
0.08416748046875,
0.771484375
],
[
0.5009765625,
0.74853515625,
0.8857421875,
0.74853515625,
0.8857421875,
0.76318359375,
0.5009765625,
0.76318359375
],
[
0.5009765625,
0.76318359375,
0.8857421875,
0.76318359375,
0.8857421875,
0.77783203125,
0.5009765625,
0.77783203125
],
[
0.2303466796875,
0.77294921875,
0.330078125,
0.7744140625,
0.330078125,
0.796875,
0.2303466796875,
0.79443359375
],
[
0.443603515625,
0.77978515625,
0.473876953125,
0.77978515625,
0.473876953125,
0.79248046875,
0.443603515625,
0.79248046875
],
[
0.50244140625,
0.77880859375,
0.8837890625,
0.77880859375,
0.8837890625,
0.79248046875,
0.50244140625,
0.79248046875
],
[
0.5009765625,
0.79443359375,
0.8837890625,
0.79443359375,
0.8837890625,
0.80810546875,
0.5009765625,
0.80810546875
],
[
0.5,
0.8095703125,
0.8837890625,
0.806640625,
0.8837890625,
0.8212890625,
0.5,
0.82373046875
],
[
0.08905029296875,
0.82763671875,
0.473876953125,
0.82763671875,
0.473876953125,
0.84521484375,
0.08905029296875,
0.84521484375
],
[
0.5009765625,
0.82275390625,
0.88232421875,
0.82373046875,
0.88232421875,
0.83837890625,
0.5009765625,
0.8369140625
],
[
0.5,
0.8369140625,
0.61767578125,
0.83837890625,
0.61767578125,
0.85302734375,
0.5,
0.8515625
],
[
0.0882568359375,
0.8447265625,
0.471435546875,
0.84228515625,
0.471435546875,
0.85986328125,
0.0882568359375,
0.8623046875
],
[
0.0906982421875,
0.861328125,
0.473876953125,
0.861328125,
0.473876953125,
0.875,
0.0906982421875,
0.875
],
[
0.50244140625,
0.861328125,
0.8837890625,
0.861328125,
0.8837890625,
0.875,
0.50244140625,
0.875
],
[
0.08905029296875,
0.875,
0.47216796875,
0.87548828125,
0.47216796875,
0.89013671875,
0.08905029296875,
0.8896484375
],
[
0.50244140625,
0.876953125,
0.8857421875,
0.876953125,
0.8857421875,
0.89013671875,
0.50244140625,
0.89013671875
],
[
0.0882568359375,
0.88818359375,
0.2467041015625,
0.88818359375,
0.2467041015625,
0.90576171875,
0.0882568359375,
0.90576171875
],
[
0.5009765625,
0.8916015625,
0.8857421875,
0.8916015625,
0.8857421875,
0.90478515625,
0.5009765625,
0.90478515625
]
] |
[
"Sharpness-Aware Minimization Alone can Improve Adversarial Robustness",
"Table 1. Natural and Robust Accuracy evaluation on CIFAR-100 dataset.",
"lo -Robust Accuracy",
"l2 -Robust Accuracy",
"Method",
" Natural Accuracy",
"=1/255",
"E=2/255",
"E=16/255",
"E=32/255",
"76.9",
"13.6",
"1.7",
"44.5",
"21.2",
"78.0",
"ST",
"19.6",
"51.5",
"27.2",
"3.0",
"SAM(p =0.1)",
"SAM(p= 0.2)",
"78.5",
"23.1",
"54.2",
"31.3",
"78.7",
"4.2",
"28.3",
"57.0",
"36.2",
"6.5",
"SAM(p = 0.4)",
"73.1",
"60.4",
"46.6",
"67.4",
"61.5",
"70.1",
"AT(l。 - e = 1/255)",
"50.6",
"65.7",
"60.3",
"AT (l。 - E = 2/255)",
"60.6",
"66.2",
"59.3",
"52.0",
"62.8",
"58.9",
"55.1",
"AT(l。 -E = 4/255)",
"AT (l。 -E = 8/255)",
"60.4",
"50.4",
"57.0",
"54.3",
"74.8",
"52.8",
"66.3",
"31.4",
"AT(l2=E=16/255)",
"57.7",
"73.2",
"57.4",
"67.1",
"40.6",
"AT (l2 - E= 32/255)",
"61.0",
"AT(l2 - E= 64/255)",
"70.7",
"58.1",
"45.9",
"66.1",
"60.7",
"67.4",
"58.2",
"48.7",
"63.9",
"60.4",
"Theorem 4.4. In the classification problem above, denote",
"AT(l2 -E = 128/255)",
"与标准训练类似，还存在一个由对抗性训练返回的最",
"佳参数WAT，可以用以下定理表示：",
"the best parameter for SAM training wSAMSuppose that",
"Eis small, we have wsAM ~ wi +gwie.",
"Theorem 4.2 (Adversarial training). In the classification",
"problem above, under adversarial training with perturba-",
"tion bound e < n, the adversarial optimal parameter value",
"4.5. Relation between SAM and AT",
"wAT _ Inp- In(1 -p)",
"我们进一步讨论AT和SAM之间不同的攻击（扰动）强",
"(10)",
"2(n -e)",
"度。回想一下，在我们在E3节中的经验理解中，SAM",
"的扰动比AT更温和，这可以解释为SAM在鲁棒性-准",
"确性权衡中更多地关注自然准确性而较少地关注鲁棒",
"我们可以看到wi已乘以，一，，这增加了对分类器的",
"性。：因此，为了达到相同的鲁棒性水平（这是由对特征",
"强大功能1的依赖性。这表明，与标准训练的模型相",
"ci的依赖性来衡量的，i。e。wi）的大小，SAM需要",
"比，对抗训练的模型更加注重鲁棒性，从而提高了对",
"更大的扰动范围，而对于AT，上的较小扰动就足够",
"抗鲁棒性",
"了。从理论上讲，以下定理验证了我们的陈述：",
"Theorem 4.5. Denote the perturbation range E of AT and",
"4.4. Sharpness-Aware Minimization (SAM)",
"SAM as EAT and esAM , respectively. Then, when both",
"methods return the same parameter wi , we have the fol-",
"现在我们考虑SAM的情况。回想一下，SAM的优化",
"lowing relationbetween EATand eSAM :",
"目标是",
"E",
"max L(r,y; w+ )",
"(11)",
"3",
"(,)~D 18|≤e",
"~2m",
"(13)",
"ESAM",
"2 +",
"我们首先通过证明用SAM训练的参数wi也大于w",
"EAT",
"来解释为什么SAM可以提高对抗鲁棒性，如下所示：",
"从定理E4.5，我们可以识别AT和SAM的不同扰动",
"Theorem4.3(Sharpness-awareminimization).Intheclas-",
"强度。从定理4.5可以很容易地得出，当（13）成立",
"sification problem above, the best parameterfor SAM train-",
"ing wSAM satisfies that",
"时，ESAM大于EAT，因为我们假设n是一个小的正，",
"E在定理E4.4和EAT<n在定理E4.2。因此，要在",
"wSAM >wi.",
"(12)",
"强大的功能&1上获得相同的权重W1，只需选择比",
"ESAM小得多的EAT。另一方面，在相同的扰动范围",
"EAT=ESAM下，在AT下训练的模型具有比SAM更",
"从定理4.2和4.3我们可以看到wAT和wSAM都大于",
"大的参数w1，因此它着重于更高的鲁棒性，但降低了",
"更多的自然精度",
"Wi，这表明SAM和AT都提高了训练模型的鲁棒性。",
"但是，与对抗性训练相比，定性关系不足以量化SAM",
"所有校样均可在附录A中找到。综上所述，我们可以",
"实现了多少鲁棒性，我们尝试通过以下定理定量估计",
"得出结论，AT利用显式和强扰动去噪非稳健特征，而",
"WSAM来进一步发展：",
"SAM利用隐式和中等扰动。这与我们在3节中的经验"
] |
{
"arxiv_id": "2305.05392",
"language": "zh",
"page_idx": 3
}
|
||
2305.05392-zh-4
|
2305.05392-zh-4.jpg
|
Sharpness-Aware Minimization Alone can Improve Adversarial Robustness
理解是一致的，我们也在下一节中通过实验验证了这
$\underline{\text{些说法}}$
# 5. Experiment
在本节中，我们将介绍我们的实验结果以验证我们提
出的理解
# 5.1. Experiment set-up
为了证明 SAM 在提高对抗鲁棒性方面的有效性，我们将使用标准 SGD 优化器训练的模型与使用 SAM 训练的模型进行了比较。我们还讨论了对抗训练。但是，我们将 AT 获得的鲁棒性视为 SAM 的上限，而不是基线在我们的实验中，我们在具有交叉熵损失的 CIFAR-10和 CIAR-100 数据集$\boxed{\mathrm{E}}($Krizhevsky et al.,2009) 上训练PreActResNet-18 (PRN-18)$\boxed{\mathrm{F}}($He et al.,2016) 模型 100个时期。学习率初始化为0。1,并分别在第 75 和 90个时期除以 10。对于优化器，权重衰减设置为 5e-4 , 动量设置为 0。9
对于 SAM, 我们从范围$\{0.1,0.2,0.4\}$中选择扰动超参数$\rho$。对于 AT, 我们同时考虑了$\ell_2$和$\ell_\infty$的鲁棒性， 并分别为两种范数训练了4 个具有不同扰动界的模型至于鲁棒性评估，我们考虑$\ell_{\infty}$ -范数扰动界限$\epsilon$ $\in$ $\{ 1/ 255, 2/ 255\}$和$\ell_2$ -范数扰动界限$\epsilon$ $\in$ $\{16/255,32/255\}$下的鲁棒性。在 10 步 PGD 攻击下评估鲁棒性
对于所有模型，我们独立运行实验三次，并报告平均结果。我们省略了标准差，因为它们很小(小于0。5%), 不影响我们的索赔
5.2. Accuracy and robustness evaluation
在 CIFAR-100 和 CIFAR-10 数据集上进行的实验结果
分别显示在表$\mathbb{E}$1 和表$\mathbb{E}$2 中
我们首先讨论 SAM 的自然和稳健的准确性性能。从表中，我们可以看到，与使用标准训练(ST)训练的模型相比，使用 SAM 训练的所有模型都表现出明显更好的自然准确性和鲁棒性。特别是，通过将$\rho$的较大值与 SAM 一起使用 , 可以实现更高的鲁棒性。对于CIFAR-100 数据集，使用$\rho=0.4$训练的模型显示出比ST 甚至更高的鲁棒性准确性，并且其自然准确性仍高于 ST。与自然精度的提高(大约2个%)相比，鲁棒性的提高更为显着(平均超过 10 个%)。同样，对于CIFAR-10 数据集，使用 SAM 训练的模型在清洁精度方面也优于 ST, 并且比 ST 表现出更高的鲁棒性。因此，我们可以得出结论，具有相对较大权重扰动约束$\rho$的 SAM 是一种有前途的技术，可以在不牺牲自然精度的情况下提高模型在对抗鲁棒精度方面的性能
关于对抗训练的模型，尽管 SAM 和 AT 获得的鲁棒性之间仍然存在很大差距，但所有对抗训练的模型都表现出比标准训练更低的自然准确性，更不用说
SAM 了。特别是，对于$\ell_\infty$ -对抗训练，即使使用扰动约束$\epsilon=1/255$进行训练，自然精度也会降低 3。$8\%$forCIFAR-100 和$0。9\%$分别用于 CIFAR-10 数据集。并且还注意到，AT 中使用的扰动界$\epsilon$越大，通过相应的模型获得的自然精度就越差。因此，使用 SAM 而不是 AT 的一个关键好处是清洁精度不会降低。此外，请注意，在 AT 中解决 PGD 过程会导致大量的计算开销。具体来说，由于我们使用 10步 PGD,因此与 ST 相比， 所有 AT 实验都需要多 10 倍的计算成本，而 SAM 仅需要多1倍
基于上述讨论，我们得出的结论是，与标准训练方法相比，SAM 训练的模型在不降低任何自然准确性的情况下表现出明显更好的鲁棒性。此外，SAM 的另一个好处是它不需要大量的计算资源。因此，我们指出，SAM 可以作为 AT 的轻量级替代品 ,它可以提高鲁棒性，而不会降低自然准确性和显着的训练开销
## 6. Conclusion
在本文中，我们表明仅使用锐度感知最小化 (SAM) 可以提高对抗鲁棒性，并揭示 SAM 与对抗训练 (AT) 之间的基本关系。我们从经验和理论上证明，SAM 和 AT 都向特征添加了扰动，以实现更好的鲁棒泛化能力。但是，SAM 隐式地添加了中等扰动，而 AT 显式地添加了强扰动。因此，它们导致不同的准确性和鲁棒性权衡。我们进一步在基准数据集上进行实验，以验证我们提出的见解的有效性。最后，我们建议在某些要求下，SAM 可以作为 AT 的轻量级替代品
# References
Athalye, A., Carlini, N., and Wagner, D. Obfuscated gra-
dients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pp. 274-283. PMLR, 2018. Bahri, D., Mobahi, H., and Tay, Y. Sharpness-aware mini-
mization improves language model generalization. arXiv
preprint arXiv:2110.08529,2021.
Bai, Y., Feng, Y., Wang, Y., Dai, T., Xia, S.-T., and Jiang, Y. Hilbert-based generative defense for adversarial examples. In $ICCV,2019.$
Chaudhari, P. , Choromanska, A. , Soatto, S. , LeCun, Y. , Bal-
dassi, C. , Borgs, C. , Chayes, J. , Sagun, L. , andZecchina, R. Entropy-sgd: Biasing gradient descent into wide valleys. Journal of Statistical Mechanics: Theory and Ex- $periment,2019(12){:}124018,2019.$
Chen, H. , Dong, Y. , Wang, Z. , Yang, X. , Duan, C. , Su, H. ,
and Zhu, J. Robust classification via a single diffusion
${\mathrm{model,2023a.}}$
Chen, H., Zhang, Y., Dong, Y., and Zhu, J. Rethinking
|
[
[
0.25244140625,
0.058074951171875,
0.72216796875,
0.058074951171875,
0.72216796875,
0.07135009765625,
0.25244140625,
0.07135009765625
],
[
0.0882568359375,
0.08587646484375,
0.473876953125,
0.08587646484375,
0.473876953125,
0.0997314453125,
0.0882568359375,
0.0997314453125
],
[
0.48193359375,
0.091552734375,
0.494384765625,
0.091552734375,
0.494384765625,
0.0953369140625,
0.48193359375,
0.0953369140625
],
[
0.5,
0.08709716796875,
0.8857421875,
0.08709716796875,
0.8857421875,
0.10101318359375,
0.5,
0.10101318359375
],
[
0.08905029296875,
0.10162353515625,
0.1422119140625,
0.10162353515625,
0.1422119140625,
0.11614990234375,
0.08905029296875,
0.11614990234375
],
[
0.50244140625,
0.10162353515625,
0.8857421875,
0.10162353515625,
0.8857421875,
0.11553955078125,
0.50244140625,
0.11553955078125
],
[
0.5009765625,
0.117431640625,
0.88232421875,
0.117431640625,
0.88232421875,
0.1307373046875,
0.5009765625,
0.1307373046875
],
[
0.0882568359375,
0.136962890625,
0.214111328125,
0.136962890625,
0.214111328125,
0.1514892578125,
0.0882568359375,
0.1514892578125
],
[
0.50244140625,
0.1337890625,
0.8837890625,
0.1337890625,
0.8837890625,
0.14453125,
0.50244140625,
0.14453125
],
[
0.50244140625,
0.147705078125,
0.8837890625,
0.147705078125,
0.8837890625,
0.1610107421875,
0.50244140625,
0.1610107421875
],
[
0.08905029296875,
0.1622314453125,
0.473876953125,
0.1622314453125,
0.473876953125,
0.175537109375,
0.08905029296875,
0.175537109375
],
[
0.50341796875,
0.1641845703125,
0.8837890625,
0.1641845703125,
0.8837890625,
0.1749267578125,
0.50341796875,
0.1749267578125
],
[
0.0906982421875,
0.1787109375,
0.156005859375,
0.1787109375,
0.156005859375,
0.1900634765625,
0.0906982421875,
0.1900634765625
],
[
0.50244140625,
0.177978515625,
0.8857421875,
0.177978515625,
0.8857421875,
0.1912841796875,
0.50244140625,
0.1912841796875
],
[
0.50341796875,
0.1925048828125,
0.88232421875,
0.1925048828125,
0.88232421875,
0.205810546875,
0.50341796875,
0.205810546875
],
[
0.0882568359375,
0.2083740234375,
0.250732421875,
0.208984375,
0.250732421875,
0.2235107421875,
0.0882568359375,
0.222900390625
],
[
0.5009765625,
0.20703125,
0.8857421875,
0.205810546875,
0.8857421875,
0.2203369140625,
0.5009765625,
0.2215576171875
],
[
0.50244140625,
0.2235107421875,
0.58251953125,
0.2235107421875,
0.58251953125,
0.23486328125,
0.50244140625,
0.23486328125
],
[
0.0931396484375,
0.2342529296875,
0.471435546875,
0.2342529296875,
0.471435546875,
0.2442626953125,
0.0931396484375,
0.2442626953125
],
[
0.0906982421875,
0.2474365234375,
0.473876953125,
0.2474365234375,
0.473876953125,
0.261474609375,
0.0906982421875,
0.261474609375
],
[
0.5009765625,
0.24560546875,
0.8857421875,
0.24560546875,
0.8857421875,
0.2587890625,
0.5009765625,
0.2587890625
],
[
0.0906982421875,
0.26318359375,
0.473876953125,
0.26318359375,
0.473876953125,
0.276611328125,
0.0906982421875,
0.276611328125
],
[
0.50244140625,
0.260009765625,
0.88232421875,
0.260009765625,
0.88232421875,
0.2734375,
0.50244140625,
0.2734375
],
[
0.09234619140625,
0.27978515625,
0.471435546875,
0.27978515625,
0.471435546875,
0.290283203125,
0.09234619140625,
0.290283203125
],
[
0.50341796875,
0.27587890625,
0.8837890625,
0.27587890625,
0.8837890625,
0.2890625,
0.50341796875,
0.2890625
],
[
0.50244140625,
0.291748046875,
0.8857421875,
0.291748046875,
0.8857421875,
0.304931640625,
0.50244140625,
0.304931640625
],
[
0.0906982421875,
0.2998046875,
0.47216796875,
0.2998046875,
0.47216796875,
0.313232421875,
0.0906982421875,
0.313232421875
],
[
0.50244140625,
0.30615234375,
0.8857421875,
0.30615234375,
0.8857421875,
0.3193359375,
0.50244140625,
0.3193359375
],
[
0.0906982421875,
0.31640625,
0.473876953125,
0.31640625,
0.473876953125,
0.330078125,
0.0906982421875,
0.330078125
],
[
0.5009765625,
0.3212890625,
0.78125,
0.3212890625,
0.78125,
0.335205078125,
0.5009765625,
0.335205078125
],
[
0.0906982421875,
0.33203125,
0.47216796875,
0.33203125,
0.47216796875,
0.34228515625,
0.0906982421875,
0.34228515625
],
[
0.09234619140625,
0.347900390625,
0.47216796875,
0.347900390625,
0.47216796875,
0.35791015625,
0.09234619140625,
0.35791015625
],
[
0.0906982421875,
0.361083984375,
0.473876953125,
0.361083984375,
0.473876953125,
0.375,
0.0906982421875,
0.375
],
[
0.5,
0.35595703125,
0.62109375,
0.35595703125,
0.62109375,
0.369384765625,
0.5,
0.369384765625
],
[
0.0906982421875,
0.375732421875,
0.2115478515625,
0.375732421875,
0.2115478515625,
0.389404296875,
0.0906982421875,
0.389404296875
],
[
0.50244140625,
0.380615234375,
0.8857421875,
0.380615234375,
0.8857421875,
0.39453125,
0.50244140625,
0.39453125
],
[
0.0906982421875,
0.3984375,
0.473876953125,
0.3984375,
0.473876953125,
0.412353515625,
0.0906982421875,
0.412353515625
],
[
0.50341796875,
0.397705078125,
0.88134765625,
0.397705078125,
0.88134765625,
0.40771484375,
0.50341796875,
0.40771484375
],
[
0.08905029296875,
0.41357421875,
0.47802734375,
0.412353515625,
0.47802734375,
0.4267578125,
0.08905029296875,
0.427978515625
],
[
0.50244140625,
0.410888671875,
0.8837890625,
0.410888671875,
0.8837890625,
0.4248046875,
0.50244140625,
0.4248046875
],
[
0.09234619140625,
0.43115234375,
0.465576171875,
0.43115234375,
0.465576171875,
0.44140625,
0.09234619140625,
0.44140625
],
[
0.5,
0.4267578125,
0.8837890625,
0.4267578125,
0.8837890625,
0.43994140625,
0.5,
0.43994140625
],
[
0.50341796875,
0.443115234375,
0.88232421875,
0.443115234375,
0.88232421875,
0.453857421875,
0.50341796875,
0.453857421875
],
[
0.0906982421875,
0.451904296875,
0.473876953125,
0.451904296875,
0.473876953125,
0.46533203125,
0.0906982421875,
0.46533203125
],
[
0.50341796875,
0.45703125,
0.8857421875,
0.45703125,
0.8857421875,
0.47021484375,
0.50341796875,
0.47021484375
],
[
0.09234619140625,
0.468505859375,
0.471435546875,
0.468505859375,
0.471435546875,
0.479248046875,
0.09234619140625,
0.479248046875
],
[
0.5009765625,
0.47021484375,
0.8857421875,
0.4716796875,
0.8857421875,
0.486083984375,
0.5009765625,
0.48486328125
],
[
0.0906982421875,
0.481689453125,
0.47216796875,
0.481689453125,
0.47216796875,
0.49560546875,
0.0906982421875,
0.49560546875
],
[
0.50244140625,
0.4873046875,
0.8857421875,
0.4873046875,
0.8857421875,
0.50048828125,
0.50244140625,
0.50048828125
],
[
0.0906982421875,
0.498779296875,
0.17236328125,
0.498779296875,
0.17236328125,
0.51025390625,
0.0906982421875,
0.51025390625
],
[
0.50341796875,
0.50390625,
0.77783203125,
0.50390625,
0.77783203125,
0.5146484375,
0.50341796875,
0.5146484375
],
[
0.08905029296875,
0.52099609375,
0.47216796875,
0.52099609375,
0.47216796875,
0.5341796875,
0.08905029296875,
0.5341796875
],
[
0.0906982421875,
0.5341796875,
0.473876953125,
0.5341796875,
0.473876953125,
0.54736328125,
0.0906982421875,
0.54736328125
],
[
0.5,
0.53515625,
0.59326171875,
0.53857421875,
0.59326171875,
0.55322265625,
0.5,
0.5498046875
],
[
0.0906982421875,
0.5498046875,
0.221435546875,
0.5498046875,
0.221435546875,
0.56298828125,
0.0906982421875,
0.56298828125
],
[
0.50244140625,
0.56103515625,
0.8857421875,
0.56103515625,
0.8857421875,
0.57470703125,
0.50244140625,
0.57470703125
],
[
0.0882568359375,
0.58203125,
0.373291015625,
0.58203125,
0.373291015625,
0.59521484375,
0.0882568359375,
0.59521484375
],
[
0.51953125,
0.57763671875,
0.8837890625,
0.57763671875,
0.8837890625,
0.58837890625,
0.51953125,
0.58837890625
],
[
0.5185546875,
0.59130859375,
0.8857421875,
0.59130859375,
0.8857421875,
0.60498046875,
0.5185546875,
0.60498046875
],
[
0.0906982421875,
0.60498046875,
0.473876953125,
0.60498046875,
0.473876953125,
0.61865234375,
0.0906982421875,
0.61865234375
],
[
0.515625,
0.60693359375,
0.875,
0.60498046875,
0.875,
0.619140625,
0.515625,
0.6220703125
],
[
0.09234619140625,
0.6220703125,
0.293212890625,
0.6220703125,
0.293212890625,
0.6318359375,
0.09234619140625,
0.6318359375
],
[
0.50244140625,
0.63330078125,
0.88623046875,
0.63330078125,
0.88623046875,
0.646484375,
0.50244140625,
0.646484375
],
[
0.08905029296875,
0.642578125,
0.47216796875,
0.642578125,
0.47216796875,
0.65576171875,
0.08905029296875,
0.65576171875
],
[
0.51708984375,
0.6484375,
0.8837890625,
0.6484375,
0.8837890625,
0.662109375,
0.51708984375,
0.662109375
],
[
0.0906982421875,
0.65771484375,
0.47216796875,
0.65771484375,
0.47216796875,
0.671875,
0.0906982421875,
0.671875
],
[
0.51318359375,
0.66357421875,
0.740234375,
0.662109375,
0.740234375,
0.6767578125,
0.51318359375,
0.67822265625
],
[
0.0906982421875,
0.6748046875,
0.47216796875,
0.6748046875,
0.47216796875,
0.68505859375,
0.0906982421875,
0.68505859375
],
[
0.0906982421875,
0.68798828125,
0.473876953125,
0.68798828125,
0.473876953125,
0.7021484375,
0.0906982421875,
0.7021484375
],
[
0.5,
0.689453125,
0.8857421875,
0.69140625,
0.8857421875,
0.705078125,
0.5,
0.70263671875
],
[
0.0906982421875,
0.70263671875,
0.47216796875,
0.70263671875,
0.47216796875,
0.71630859375,
0.0906982421875,
0.71630859375
],
[
0.5146484375,
0.7041015625,
0.8857421875,
0.705078125,
0.8857421875,
0.7197265625,
0.5146484375,
0.71826171875
],
[
0.0906982421875,
0.71826171875,
0.47216796875,
0.71826171875,
0.47216796875,
0.73193359375,
0.0906982421875,
0.73193359375
],
[
0.515625,
0.72021484375,
0.67333984375,
0.7197265625,
0.67333984375,
0.734375,
0.515625,
0.73486328125
],
[
0.0906982421875,
0.73486328125,
0.471435546875,
0.73486328125,
0.471435546875,
0.74560546875,
0.0906982421875,
0.74560546875
],
[
0.08905029296875,
0.74755859375,
0.473876953125,
0.74853515625,
0.473876953125,
0.76220703125,
0.08905029296875,
0.76123046875
],
[
0.498291015625,
0.7470703125,
0.8857421875,
0.7470703125,
0.8857421875,
0.76025390625,
0.498291015625,
0.76025390625
],
[
0.08905029296875,
0.76318359375,
0.473876953125,
0.76318359375,
0.473876953125,
0.7763671875,
0.08905029296875,
0.7763671875
],
[
0.51708984375,
0.76318359375,
0.8857421875,
0.76318359375,
0.8857421875,
0.7763671875,
0.51708984375,
0.7763671875
],
[
0.08905029296875,
0.77880859375,
0.473876953125,
0.77880859375,
0.473876953125,
0.79248046875,
0.08905029296875,
0.79248046875
],
[
0.515625,
0.77783203125,
0.8857421875,
0.77783203125,
0.8857421875,
0.79248046875,
0.515625,
0.79248046875
],
[
0.09234619140625,
0.79541015625,
0.471435546875,
0.79541015625,
0.471435546875,
0.80615234375,
0.09234619140625,
0.80615234375
],
[
0.51708984375,
0.79248046875,
0.88623046875,
0.79248046875,
0.88623046875,
0.80615234375,
0.51708984375,
0.80615234375
],
[
0.09234619140625,
0.81103515625,
0.471435546875,
0.81103515625,
0.471435546875,
0.8212890625,
0.09234619140625,
0.8212890625
],
[
0.515625,
0.8095703125,
0.740234375,
0.8095703125,
0.740234375,
0.82275390625,
0.515625,
0.82275390625
],
[
0.0906982421875,
0.82470703125,
0.47216796875,
0.82470703125,
0.47216796875,
0.83837890625,
0.0906982421875,
0.83837890625
],
[
0.5009765625,
0.83544921875,
0.88232421875,
0.83544921875,
0.88232421875,
0.8486328125,
0.5009765625,
0.8486328125
],
[
0.09234619140625,
0.84130859375,
0.431396484375,
0.84130859375,
0.431396484375,
0.8515625,
0.09234619140625,
0.8515625
],
[
0.5185546875,
0.8515625,
0.8837890625,
0.8515625,
0.8837890625,
0.8623046875,
0.5185546875,
0.8623046875
],
[
0.09234619140625,
0.861328125,
0.47216796875,
0.861328125,
0.47216796875,
0.875,
0.09234619140625,
0.875
],
[
0.5185546875,
0.8662109375,
0.609375,
0.8662109375,
0.609375,
0.87744140625,
0.5185546875,
0.87744140625
],
[
0.0906982421875,
0.876953125,
0.47216796875,
0.876953125,
0.47216796875,
0.89013671875,
0.0906982421875,
0.89013671875
],
[
0.09234619140625,
0.89306640625,
0.47216796875,
0.89306640625,
0.47216796875,
0.90380859375,
0.09234619140625,
0.90380859375
],
[
0.5009765625,
0.892578125,
0.8857421875,
0.892578125,
0.8857421875,
0.90576171875,
0.5009765625,
0.90576171875
]
] |
[
"Sharpness-Aware Minimization Alone can Improve Adversarial Robustness",
"理解是一致的，我们也在下一节中通过实验验证了这",
"",
"SAM了。特别是，对于l-对抗训练，即使使用扰",
"些说法",
"动约束e=1/255进行训练，自然精度也会降低3。",
"8%forCIFAR-100和0。9%分别用于CIFAR-10数据集",
"5. Experiment",
"并且还注意到，AT中使用的扰动界E越大，通过相应",
"的模型获得的自然精度就越差。因此，使用SAM而不",
"在本节中，我们将介绍我们的实验结果以验证我们提",
"是AT的一个关键好处是清洁精度不会降低。此外，请",
"出的理解",
"注意，在AT中解决PGD过程会导致大量的计算开销。",
"具体来说，由于我们使用10步PGD，因此与ST相比",
"5.1.Experiment set-up",
"所有AT实验都需要多10倍的计算成本，而SAM仅",
"需要多1倍",
"为了证明SAM在提高对抗鲁棒性方面的有效性，我们",
"将使用标准SGD优化器训练的模型与使用SAM训练",
"基于上述讨论，我们得出的结论是，与标准训练方法相",
"的模型进行了比较。我们还讨论了对抗训练。但是，我",
"比，SAM训练的模型在不降低任何自然准确性的情况",
"们将AT获得的鲁棒性视为SAM的上限，而不是基线",
"下表现出明显更好的鲁棒性。此外，SAM的另一个好",
"处是它不需要大量的计算资源。因此，我们指出，SAM",
"在我们的实验中，我们在具有交叉摘损失的CIFAR-10",
"可以作为AT的轻量级替代品，它可以提高鲁棒性，而",
"和CIAR-100数据集E(Krizhevskyetal.，2009）上训练",
"不会降低自然准确性和显着的训练开销",
"PreActResNet-18（PRN-18)/F(Heetal..2016)模型100",
"个时期。学习率初始化为0。1，并分别在第75和90",
"个时期除以10。对于优化器，权重衰减设置为5e-4，",
"6. Conclusion",
"动量设置为0。9",
"在本文中，我们表明仅使用锐度感知最小化（SAM）可",
"对于SAM，我们从范围[0.1,0.2,0.4中选择扰动超参",
"以提高对抗鲁棒性，并揭示SAM与对抗训练（AT）之",
"数p。对于AT，我们同时考虑了l2和l。的鲁棒性，",
"间的基本关系。我们从经验和理论上证明，SAM和AT",
"并分别为两种范数训练4个具有不同扰动界的模型",
"都向特征添加了扰动，以实现更好的鲁棒泛化能力。但",
"是，SAM隐式地添加了中等扰动，而AT显式地添加",
"至于鲁棒性评估，我们考虑l-范数扰动界限",
"了强扰动。因此，它们导致不同的准确性和鲁棒性权",
"EE1/255.2/2551和l2-范数扰动界限E",
"衡。我们进一步在基准数据集上进行实验，以验证我",
"[16/255,32/255]下的鲁棒性。在10步PGD攻击下",
"们提出的见解的有效性。最后，我们建议在某些要求",
"评估鲁棒性",
"下，SAM可以作为AT的轻量级替代品",
"对于所有模型，我们独立运行实验三次，并报告平均结",
"果。我们省略了标准差，因为它们很小（小于0。5%），",
"References",
"不影响我们的索赔",
"Athalye, A., Carlini, N., and Wagner, D. Obfuscated gra-",
"5.2. Accuracy and robustness evaluation",
"dientsgive afalse sense of security:Circumventingde",
"fenses to adversarial examples. In International confer",
"在CIFAR-100和CIFAR-10数据集上进行的实验结果",
"ence on machine learning, pp.274-283.PMLR, 2018.",
"分别显示在表F1和表F2中",
"Bahri, D., Mobahi, H., and Tay, Y. Sharpness-aware mini-",
"我们首先讨论SAM的自然和稳健的准确性性能。从",
"mization improves language model generalization. arXiv",
"表中，我们可以看到，与使用标准训练（ST）训练的",
"preprint arXiv:2110.08529,2021.",
"模型相比，使用SAM训练的所有模型都表现出明显",
"更好的自然准确性和鲁棒性。特别是，通过将p的较",
"Bai, Y.,Feng,Y.,Wang, Y., Dai, T., Xia, S.T., and Jiang,",
"大值与SAM一起使用，可以实现更高的鲁棒性。对于",
"Y. Hilbert-based generative defense for adversarial ex-",
"CIFAR-100数据集，使用p=0.4训练的模型显示出比",
"amples. In ICCV, 2019",
"ST甚至更高的鲁棒性准确性，并且其自然准确性仍高",
"于ST。与自然精度的提高（大约2个%）相比，鲁棒",
"Chaudhari, P., Choromanska, A., Soatto, S., LeCun, Y., Bal-",
"性的提高更为显着（平均超过10个%）。同样，对于",
"dassi, C., Borgs, C., Chayes, J., Sagun, L., and Zecchina,",
"CIFAR-1O数据集，使用SAM训练的模型在清洁精度",
"R. Entropy-sgd: Biasing gradient descent into wide val-",
"方面也优于ST，并且比ST表现出更高的鲁棒性。因",
"leys.Journal of Statistical Mechanics:Theory and Ex-",
"此，我们可以得出结论，具有相对较天权重扰动约束",
"periment,2019(12):124018,2019",
"p的SAM是一种有前途的技术，可以在不牺牲自然精",
"Chen, H., Dong, Y., Wang, Z., Yang, X., Duan, C., Su, H.",
"度的情况下提高模型在对抗鲁棒精度方面的性能",
"and Zhu,J.Robust classification via a single diffusion",
"关于对抗训练的模型，尽管SAM和AT获得的鲁棒",
"model,2023a.",
"性之间仍然存在很大差距，但所有对抗训练的模型",
"都表现出比标准训练更低的自然准确性，更不用说",
"Chen, H., Zhang, Y., Dong, Y., and Zhu, J.Rethinking"
] |
{
"arxiv_id": "2305.05392",
"language": "zh",
"page_idx": 4
}
|
||
1905.01866-zh-1
|
1905.01866-zh-1.jpg
|
KDD'19, August 4-8, 2019, Anchorage, AK, USA
Wen Chen and Pipei Huang, et al.
# $\mathsf{POG:Personalized~Outfit~Generation~for~Fashion}$ $\textbf{Recommendation at Alibaba iFashion}$
Wen Chen, Pipei Huang*
Jiaming Xu, Xin Guo, Cheng Guo, Fei Sun, Chao Li, Andreas Pfadler, Huan Zhao, Binqiang Zhao
Alibaba Group, Beijing, China
$\{$ chenyu.cw, pipei.hpp, jiaming.xjm, gary.gx, mike.gc, ofey.sf, ruide.lc, andreaswernerrober, chuanfei.zh binqiang.zhao
$\}$ @alibaba-inc.com
$\mathbf{ABSTRACT}$
ACM Reference Format:
Wen Chen, Pipei Huang and Jiaming Xu, Xin Guo, Cheng Guo, Fei Sun, Chao Li, Andreas Pfadler, Huan Zhao, Binqiang Zhao. 2019. POG:Personalized Outfit Generation for Fashion Recommendation at Alibaba iFashion . In Proceedings of The 25th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD’19).ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/3292500.3330652
## 1 INTRODUCTION
Increasing demand for fashion recommendation raises a lot of challenges for online shopping platforms and fashion communities. In particular, there exist two requirements for fashion outfit recommendation: the Compatibility of the generated fashion outfits, and the Personalization in the recommendation process. In this paper, we demonstrate these two requirements can be satisfied via building a bridge between outfit generation and recommendation. Through large data analysis, we observe that people have similar tastes in individual items and outfits. Therefore, we propose a P ersonalized O utfit G eneration ( POG ) model, which connects user preferences regarding individual items and outfits with Transformer architecture. Extensive offline and online experiments provide strong quantitative evidence that our method outperforms alternative methods regarding both compatibility and personalization metrics. Furthermore, we deploy POG on a platform named Dida in Alibaba to generate personalized outfits for the users of the online application iFashion. This work represents a first step towards an industrial-scale fashion outfit generation and recommendation solution, which goes beyond generating outfits based on explicit queries, or merely recommending from existing outfit pools. As part of this work, we release a large-scale dataset consisting of 1.01 million outfits with rich context information, and 0.28 billion user click actions from 3.57 million users. To the best of our knowledge, this dataset is the largest, publicly available, fashion related dataset, and the first to provide user behaviors relating to both outfts and fashion items.
面对看似丰富的不断变化的款式选择，时尚服装推荐对于现代消费者而言变得越来越重要，因此引起了在线零售行业的兴趣。时装是一组时尚物品，看起来既视觉兼容又功能上不冗余的[3](例如，见图 1)。与传统的物品推荐相比，时尚服装推荐涉及一个非常有创意的服装生成过程，这需要创新和特色。因此，此任务通常由时尚专家执行，并在众多在线时尚社区中广受欢迎，例如 Lookbook $^{1}$和 Chictopia $^{2}$。在中国最大的在线消费者对消费者平台 Taobao 中，创建了一个新的应用程序iFashion 以支持时尚服装推荐。约 1。截至 2018 年 3月 31 日500 万内容创作者积极支持 Taobao $^{3}$。尽管如此，有限的人力与不断增长的市场之间的巨大差距要求补充甚至替代体力劳动。为了缓解这个问题，我们的目标是协助 iFashion 中的生成和推荐过程
在时装生成和推荐中存在两个要求：1)生成的时装的兼容性，2) 推荐过程中的个性化。兼容性是衡量一组项目的和谐程度。早期的研究主要集中在学习成对项目$\mathbb{E}[17,30]$之间的兼容性指标，或预测服装$\mathbb{E}[$11]的受欢迎程度。最近的一些作品试图通过将服装建模为有序的物品序列来生成服装[3,18]。但是，这是不合理的，因为在服装中洗牌物品对其兼容性没有影响。个性化代表建议如何满足用户的个人时尚品味。在最近的作品中，个性化是依靠显式输入(例如。g。图像或文本)由用户提供[4,15,18] 。这种类型的时尚生成更像搜索而不是推荐，因为它需要明确的用户查询。另一方面，经典的协同过滤(CF)方法[14,22]在推荐中大多侧重于推荐单个项目，而不是服装
在这项工作中，我们收集了1。21亿用户点击4次。6800 万件商品和 1.92 万件服装来自 5。5400 万用户在 iFashion 。如表 1 所示，81。3 点击服装的% 包含具有相同类别的项目，这些项目已出现在同一用户的最新 50 个点击项目中。我们观察到，用户点击服装中物品的品牌，样式和图案也很有可能出现在用户最新点击的物品中(我们为每个用户选择最新的 50 个物品)。图 2说明了三个示例用户在物品和服装上的行为。所有这些结果表明，用户倾向于在单个物品和服装中保持相似的口味。我们发现，尽管近年来对时装的生成和推荐进行了深人研究，但现有的作品通常会分别研究这两个要求
${\overline{^{1}\text{https://lookbook.nu/}}}$
$\csc\csc$coNCEPTS
$\cdot$Information systems$\to$Recommender systems; Multimedia
## content creation.
# $\mathbf{KEYWORDS}$
# Fashion Outfit Generation; Fashion Outfit Recommendation; Deep Learning; Transformer; Self-attention
$^{\cdot}$Pipei Huang is the Corresponding author.
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org
$KDD^{\prime }19$, August 4-8, 2019, $Anchorage$, $AK$, $USA$
$\textcircled{\circ}$ 2019 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-6201-6/19/08....815.00
https://doi.org/10.1145/3292500.3330652
${}_{a}^{2}$http://chictopia.com/
$^{3}$https://www.alibabagroup.com/en/news/press$\_$pdf/p180504.pdf
|
[
[
0.0849609375,
0.07574462890625,
0.32275390625,
0.07574462890625,
0.32275390625,
0.0889892578125,
0.0849609375,
0.0889892578125
],
[
0.74853515625,
0.07574462890625,
0.90869140625,
0.07574462890625,
0.90869140625,
0.0889892578125,
0.74853515625,
0.0889892578125
],
[
0.1846923828125,
0.10479736328125,
0.81640625,
0.10479736328125,
0.81640625,
0.12249755859375,
0.1846923828125,
0.12249755859375
],
[
0.2548828125,
0.1300048828125,
0.74609375,
0.1300048828125,
0.74609375,
0.147705078125,
0.2548828125,
0.147705078125
],
[
0.398681640625,
0.160400390625,
0.59619140625,
0.160400390625,
0.59619140625,
0.1749267578125,
0.398681640625,
0.1749267578125
],
[
0.11517333984375,
0.177978515625,
0.8837890625,
0.177978515625,
0.8837890625,
0.1925048828125,
0.11517333984375,
0.1925048828125
],
[
0.398681640625,
0.1925048828125,
0.6005859375,
0.1932373046875,
0.6005859375,
0.2083740234375,
0.398681640625,
0.20703125
],
[
0.108642578125,
0.208984375,
0.890625,
0.208984375,
0.890625,
0.2235107421875,
0.108642578125,
0.2235107421875
],
[
0.43310546875,
0.2247314453125,
0.56640625,
0.2247314453125,
0.56640625,
0.238037109375,
0.43310546875,
0.238037109375
],
[
0.08660888671875,
0.2474365234375,
0.1846923828125,
0.2474365234375,
0.1846923828125,
0.261962890625,
0.08660888671875,
0.261962890625
],
[
0.51953125,
0.250732421875,
0.6611328125,
0.250732421875,
0.6611328125,
0.261474609375,
0.51953125,
0.261474609375
],
[
0.0849609375,
0.266357421875,
0.483642578125,
0.266357421875,
0.483642578125,
0.27978515625,
0.0849609375,
0.27978515625
],
[
0.51708984375,
0.261962890625,
0.91259765625,
0.261962890625,
0.91259765625,
0.27587890625,
0.51708984375,
0.27587890625
],
[
0.0849609375,
0.281005859375,
0.48193359375,
0.281005859375,
0.48193359375,
0.294189453125,
0.0849609375,
0.294189453125
],
[
0.5185546875,
0.276611328125,
0.9111328125,
0.276611328125,
0.9111328125,
0.287353515625,
0.5185546875,
0.287353515625
],
[
0.5185546875,
0.2890625,
0.91259765625,
0.2890625,
0.91259765625,
0.2998046875,
0.5185546875,
0.2998046875
],
[
0.08660888671875,
0.296630859375,
0.48046875,
0.296630859375,
0.48046875,
0.306884765625,
0.08660888671875,
0.306884765625
],
[
0.5185546875,
0.300537109375,
0.9140625,
0.300537109375,
0.9140625,
0.314453125,
0.5185546875,
0.314453125
],
[
0.0849609375,
0.309326171875,
0.48046875,
0.309326171875,
0.48046875,
0.322509765625,
0.0849609375,
0.322509765625
],
[
0.51708984375,
0.313232421875,
0.91259765625,
0.313232421875,
0.91259765625,
0.326904296875,
0.51708984375,
0.326904296875
],
[
0.08416748046875,
0.32080078125,
0.48193359375,
0.3212890625,
0.48193359375,
0.3359375,
0.08416748046875,
0.335205078125
],
[
0.515625,
0.32568359375,
0.73388671875,
0.32568359375,
0.73388671875,
0.339111328125,
0.515625,
0.339111328125
],
[
0.0849609375,
0.3359375,
0.48193359375,
0.335205078125,
0.48193359375,
0.348388671875,
0.0849609375,
0.349853515625
],
[
0.0849609375,
0.349853515625,
0.48046875,
0.349853515625,
0.48046875,
0.363037109375,
0.0849609375,
0.363037109375
],
[
0.51953125,
0.354736328125,
0.53369140625,
0.354736328125,
0.53369140625,
0.363037109375,
0.51953125,
0.363037109375
],
[
0.54736328125,
0.3515625,
0.69287109375,
0.3515625,
0.69287109375,
0.365478515625,
0.54736328125,
0.365478515625
],
[
0.0849609375,
0.3642578125,
0.48046875,
0.3642578125,
0.48046875,
0.37744140625,
0.0849609375,
0.37744140625
],
[
0.5185546875,
0.36865234375,
0.91259765625,
0.36865234375,
0.91259765625,
0.3818359375,
0.5185546875,
0.3818359375
],
[
0.08660888671875,
0.37890625,
0.48046875,
0.37890625,
0.48046875,
0.389404296875,
0.08660888671875,
0.389404296875
],
[
0.5185546875,
0.3837890625,
0.91259765625,
0.3837890625,
0.91259765625,
0.397705078125,
0.5185546875,
0.397705078125
],
[
0.0849609375,
0.391357421875,
0.48193359375,
0.391357421875,
0.48193359375,
0.40478515625,
0.0849609375,
0.40478515625
],
[
0.5185546875,
0.397705078125,
0.91259765625,
0.397705078125,
0.91259765625,
0.410888671875,
0.5185546875,
0.410888671875
],
[
0.0849609375,
0.40478515625,
0.48193359375,
0.404052734375,
0.48193359375,
0.41845703125,
0.0849609375,
0.419189453125
],
[
0.5185546875,
0.410888671875,
0.91259765625,
0.410888671875,
0.91259765625,
0.4248046875,
0.5185546875,
0.4248046875
],
[
0.08660888671875,
0.42041015625,
0.48193359375,
0.42041015625,
0.48193359375,
0.43115234375,
0.08660888671875,
0.43115234375
],
[
0.5185546875,
0.425537109375,
0.91259765625,
0.425537109375,
0.91259765625,
0.439453125,
0.5185546875,
0.439453125
],
[
0.0849609375,
0.43310546875,
0.48193359375,
0.43310546875,
0.48193359375,
0.4462890625,
0.0849609375,
0.4462890625
],
[
0.5185546875,
0.439453125,
0.91259765625,
0.439453125,
0.91259765625,
0.45263671875,
0.5185546875,
0.45263671875
],
[
0.0849609375,
0.447509765625,
0.48193359375,
0.447509765625,
0.48193359375,
0.4609375,
0.0849609375,
0.4609375
],
[
0.51953125,
0.45263671875,
0.9140625,
0.45263671875,
0.9140625,
0.466552734375,
0.51953125,
0.466552734375
],
[
0.08660888671875,
0.462158203125,
0.48046875,
0.462158203125,
0.48046875,
0.472900390625,
0.08660888671875,
0.472900390625
],
[
0.51953125,
0.468505859375,
0.91259765625,
0.468505859375,
0.91259765625,
0.479248046875,
0.51953125,
0.479248046875
],
[
0.0849609375,
0.474853515625,
0.48193359375,
0.474853515625,
0.48193359375,
0.488037109375,
0.0849609375,
0.488037109375
],
[
0.51953125,
0.48095703125,
0.91259765625,
0.48095703125,
0.91259765625,
0.494384765625,
0.51953125,
0.494384765625
],
[
0.0849609375,
0.488037109375,
0.4794921875,
0.488037109375,
0.4794921875,
0.501953125,
0.0849609375,
0.501953125
],
[
0.5185546875,
0.494384765625,
0.91259765625,
0.494384765625,
0.91259765625,
0.50830078125,
0.5185546875,
0.50830078125
],
[
0.0849609375,
0.501953125,
0.48046875,
0.501953125,
0.48046875,
0.51513671875,
0.0849609375,
0.51513671875
],
[
0.51953125,
0.5087890625,
0.91259765625,
0.5087890625,
0.91259765625,
0.52294921875,
0.51953125,
0.52294921875
],
[
0.08660888671875,
0.517578125,
0.48046875,
0.517578125,
0.48046875,
0.52783203125,
0.08660888671875,
0.52783203125
],
[
0.5185546875,
0.521484375,
0.91259765625,
0.521484375,
0.91259765625,
0.53515625,
0.5185546875,
0.53515625
],
[
0.0849609375,
0.52978515625,
0.48193359375,
0.52978515625,
0.48193359375,
0.54345703125,
0.0849609375,
0.54345703125
],
[
0.51953125,
0.5361328125,
0.60693359375,
0.5361328125,
0.60693359375,
0.546875,
0.51953125,
0.546875
],
[
0.08660888671875,
0.54345703125,
0.48193359375,
0.54345703125,
0.48193359375,
0.556640625,
0.08660888671875,
0.556640625
],
[
0.53369140625,
0.5498046875,
0.9140625,
0.5498046875,
0.9140625,
0.56298828125,
0.53369140625,
0.56298828125
],
[
0.08660888671875,
0.5595703125,
0.4794921875,
0.5595703125,
0.4794921875,
0.5693359375,
0.08660888671875,
0.5693359375
],
[
0.5185546875,
0.56298828125,
0.91259765625,
0.56298828125,
0.91259765625,
0.5771484375,
0.5185546875,
0.5771484375
],
[
0.08660888671875,
0.57275390625,
0.4794921875,
0.57275390625,
0.4794921875,
0.58349609375,
0.08660888671875,
0.58349609375
],
[
0.51953125,
0.5791015625,
0.9111328125,
0.5791015625,
0.9111328125,
0.5888671875,
0.51953125,
0.5888671875
],
[
0.0849609375,
0.58544921875,
0.48046875,
0.58544921875,
0.48046875,
0.5986328125,
0.0849609375,
0.5986328125
],
[
0.5185546875,
0.59130859375,
0.91259765625,
0.59130859375,
0.91259765625,
0.60498046875,
0.5185546875,
0.60498046875
],
[
0.08660888671875,
0.599609375,
0.4794921875,
0.599609375,
0.4794921875,
0.60986328125,
0.08660888671875,
0.60986328125
],
[
0.51953125,
0.60693359375,
0.9111328125,
0.60693359375,
0.9111328125,
0.6171875,
0.51953125,
0.6171875
],
[
0.08660888671875,
0.61279296875,
0.271240234375,
0.61279296875,
0.271240234375,
0.62353515625,
0.08660888671875,
0.62353515625
],
[
0.521484375,
0.62060546875,
0.9111328125,
0.62060546875,
0.9111328125,
0.630859375,
0.521484375,
0.630859375
],
[
0.51953125,
0.63330078125,
0.91259765625,
0.63330078125,
0.91259765625,
0.646484375,
0.51953125,
0.646484375
],
[
0.08660888671875,
0.6396484375,
0.221435546875,
0.6396484375,
0.221435546875,
0.65283203125,
0.08660888671875,
0.65283203125
],
[
0.51708984375,
0.64599609375,
0.91259765625,
0.646484375,
0.91259765625,
0.6611328125,
0.51708984375,
0.66015625
],
[
0.08660888671875,
0.65771484375,
0.48046875,
0.65771484375,
0.48046875,
0.671875,
0.08660888671875,
0.671875
],
[
0.51953125,
0.66015625,
0.91259765625,
0.66015625,
0.91259765625,
0.673828125,
0.51953125,
0.673828125
],
[
0.08660888671875,
0.67236328125,
0.188720703125,
0.67236328125,
0.188720703125,
0.6845703125,
0.08660888671875,
0.6845703125
],
[
0.5185546875,
0.673828125,
0.91259765625,
0.673828125,
0.91259765625,
0.6875,
0.5185546875,
0.6875
],
[
0.5185546875,
0.68798828125,
0.9140625,
0.68798828125,
0.9140625,
0.7021484375,
0.5185546875,
0.7021484375
],
[
0.08660888671875,
0.69970703125,
0.188720703125,
0.69970703125,
0.188720703125,
0.7109375,
0.08660888671875,
0.7109375
],
[
0.5185546875,
0.7021484375,
0.54833984375,
0.7021484375,
0.54833984375,
0.7138671875,
0.5185546875,
0.7138671875
],
[
0.08416748046875,
0.71533203125,
0.48046875,
0.71630859375,
0.48046875,
0.73095703125,
0.08416748046875,
0.72998046875
],
[
0.53515625,
0.71533203125,
0.9140625,
0.71533203125,
0.9140625,
0.72900390625,
0.53515625,
0.72900390625
],
[
0.08416748046875,
0.73095703125,
0.313720703125,
0.73095703125,
0.313720703125,
0.744140625,
0.08416748046875,
0.744140625
],
[
0.51953125,
0.72998046875,
0.9140625,
0.72998046875,
0.9140625,
0.74365234375,
0.51953125,
0.74365234375
],
[
0.5185546875,
0.74365234375,
0.91259765625,
0.74365234375,
0.91259765625,
0.7568359375,
0.5185546875,
0.7568359375
],
[
0.0849609375,
0.75390625,
0.285888671875,
0.755859375,
0.285888671875,
0.76953125,
0.0849609375,
0.767578125
],
[
0.51708984375,
0.755859375,
0.91259765625,
0.7568359375,
0.91259765625,
0.77099609375,
0.51708984375,
0.76953125
],
[
0.5185546875,
0.771484375,
0.9140625,
0.771484375,
0.9140625,
0.78515625,
0.5185546875,
0.78515625
],
[
0.08660888671875,
0.78076171875,
0.4794921875,
0.78076171875,
0.4794921875,
0.79150390625,
0.08660888671875,
0.79150390625
],
[
0.5185546875,
0.78515625,
0.91259765625,
0.78515625,
0.91259765625,
0.798828125,
0.5185546875,
0.798828125
],
[
0.08660888671875,
0.79150390625,
0.48193359375,
0.79150390625,
0.48193359375,
0.8046875,
0.08660888671875,
0.8046875
],
[
0.08660888671875,
0.80322265625,
0.48046875,
0.80322265625,
0.48046875,
0.81298828125,
0.08660888671875,
0.81298828125
],
[
0.51953125,
0.80029296875,
0.9111328125,
0.80029296875,
0.9111328125,
0.81103515625,
0.51953125,
0.81103515625
],
[
0.08660888671875,
0.81298828125,
0.48046875,
0.81298828125,
0.48046875,
0.82373046875,
0.08660888671875,
0.82373046875
],
[
0.51953125,
0.814453125,
0.9111328125,
0.814453125,
0.9111328125,
0.82470703125,
0.51953125,
0.82470703125
],
[
0.08660888671875,
0.82373046875,
0.48046875,
0.82373046875,
0.48046875,
0.833984375,
0.08660888671875,
0.833984375
],
[
0.5185546875,
0.8271484375,
0.91259765625,
0.8271484375,
0.91259765625,
0.84033203125,
0.5185546875,
0.84033203125
],
[
0.0882568359375,
0.83349609375,
0.48046875,
0.83349609375,
0.48046875,
0.84326171875,
0.0882568359375,
0.84326171875
],
[
0.08660888671875,
0.84326171875,
0.43310546875,
0.84326171875,
0.43310546875,
0.85400390625,
0.08660888671875,
0.85400390625
],
[
0.5185546875,
0.84130859375,
0.84326171875,
0.84130859375,
0.84326171875,
0.85498046875,
0.5185546875,
0.85498046875
],
[
0.08660888671875,
0.85400390625,
0.314453125,
0.85400390625,
0.314453125,
0.8642578125,
0.08660888671875,
0.8642578125
],
[
0.0849609375,
0.86279296875,
0.473876953125,
0.8623046875,
0.473876953125,
0.87548828125,
0.0849609375,
0.876953125
],
[
0.5185546875,
0.8623046875,
0.62109375,
0.8623046875,
0.62109375,
0.87353515625,
0.5185546875,
0.87353515625
],
[
0.08660888671875,
0.875,
0.293212890625,
0.875,
0.293212890625,
0.88525390625,
0.08660888671875,
0.88525390625
],
[
0.5185546875,
0.87353515625,
0.62109375,
0.87353515625,
0.62109375,
0.88525390625,
0.5185546875,
0.88525390625
],
[
0.0849609375,
0.88525390625,
0.273681640625,
0.8837890625,
0.273681640625,
0.89599609375,
0.0849609375,
0.896484375
],
[
0.51708984375,
0.8818359375,
0.82177734375,
0.88330078125,
0.82177734375,
0.89794921875,
0.51708984375,
0.896484375
]
] |
[
"KDD '19, August 4-8, 2019, Anchorage, AK, USA",
"Wen Chen and Pipei Huang, et al.",
"POG :Personalized Outfit Generation for Fashion",
"Recommendation at Alibaba iFashion",
"Wen Chen, Pipei Huang\"",
"Jiaming Xu, Xin Guo, Cheng Guo, Fei Sun, Chao Li, Andreas Pfadler, Huan Zhao, Binqiang Zhao",
"Alibaba Group, Beijing,China",
"[ chenyu.cw, pipei.hpp, jiaming.xjm, gary.gx, mike.gc, ofey.sf, ruide.lc, andreaswernerrober, chuanfei.zh binqiang.zhao",
"1 @alibaba-inc.com",
"ABSTRACT",
"ACMReferenceFormat:",
"Increasing demand for fashion recommendation raises a lot of chal-",
"Wen Chen, Pipei Huang and Jiaming Xu, Xin Guo, Cheng Guo, Fei Sun,",
"lenges for online shopping platforms and fashion communities.",
"Chao Li, Andreas Pfadler,Huan Zhao, Binqiang Zhao.2019.POG:Person",
"alized OutfitGenerationforFashion RecommendationatAlibabaiFashion",
"In particular, there existtwo requirements for fashion outfitrec",
". In Proceedings of The 25th ACM SIGKDD Conference on Knowledge Dis-",
"ommendation: the Compatibility of thegenerated fashion outfits,",
"covery and Data Mining (KDD '19). ACM, New York, NY, USA, 9 pages.",
"and the Personalization in the recommendation process. In this pa-",
"https://doi.org/10.1145/3292500.3330652",
"per, we demonstrate these two requirements can be satisfied via",
"building a bridge between outfit generation and recommendation.",
"1",
"INTRODUCTION",
"Through large data analysis, we observe that people have similar",
"面对看似丰富的不断变化的款式选择，时尚服装推荐对于现代",
"tastes in individual items and outfits.Therefore,wepropose a P",
"消费者而言变得越来越重要，因此引起了在线零售行业的兴",
"ersonalized O utfit G eneration （POG)model, which connects",
"趣。时装是一组时尚物品，看起来既视觉兼容又功能上不允余",
"user preferences regarding individual items and outfits with Trans-",
"的[3]（例如，见图1）。与传统的物品推荐相比，时尚服装推",
"formerarchitecture.Extensiveofflineand online experimentspro-",
"荐涉及一个非常有创意的服装生成过程，这需要创新和特色。",
"vide strong quantitative evidence that our method outperforms al-",
"因此，此任务通常由时尚专家执行，并在众多在线时尚社区中",
"ternative methods regarding both compatibility and personaliza-",
"广受欢迎，例如Lookbook|和Chictopia？。在中国最大的在",
"tionmetrics.Furthermore, wedeployPOGon aplatformnamed",
"线消费者对消费者平台Taobao中，创建了一个新的应用程序",
"Dida in Alibaba to generate personalized outfits for the users of",
"iFashion以支持时尚服装推荐。约1。截至2018年3月31日，",
"the online application iFashion.This work represents a first step",
"500万内容创作者积极支持Taobao3。尽管如此，有限的人力",
"towards an industrial-scale fashion outfit generation and recom",
"与不断增长的市场之间的巨大差距要求补充甚至替代体力劳",
"mendationsolution,whichgoesbeyondgeneratingoutfitsbased",
"动。为了缓解这个问题，我们的目标是协助iFashion中的生",
"on explicit queries, or merely recommending from existing outfit",
"成和推荐过程",
"pools. As part of this work, we release a large-scale dataset consist-",
"在时装生成和推荐中存在两个要求：1）生成的时装的兼容",
"ing of1.01millionoutfitswithrich contextinformation,ando.28",
"性，2）推荐过程中的个性化。兼容性是衡量一组项目的和谐",
"billionuserclick actionsfrom3.57millionusers.Tothebestofour",
"程度。早期的研究主要集中在学习成对项目E[17，30之间的",
"knowledge, this dataset is the largest, publicly available, fashion",
"兼容性指标，或预测服装[11]的受欢迎程度。最近的一些作",
"related dataset,and thefirst to provideuser behaviors relating to",
"品试图通过将服装建模为有序的物品序列来生成服装[3，18]。",
"bothoutfitsand fashionitems.",
"但是，这是不合理的，因为在服装中洗牌物品对其兼容性没有",
"影响。个性化代表建议如何满足用户的个人时尚品味。在最近",
"CCS CONCEPTS",
"的作品中，个性化是依靠显式输入（例如。g。、图像或文本)由",
".Information systemsRecommender systems; Multimedia",
"用户提供[4,15,18]。这种类型的时尚生成更像搜索而不是推",
"content creation.",
"荐，因为它需要明确的用户查询。另一方面，经典的协同过滤",
"(CF)方法[14,22]在推荐中大多侧重于推荐单个项目，而不是",
"KEYWORDS",
"服装",
"Fashion Outfit Generation; Fashion Outfit Recommendation; Deep",
"在这项工作中，我们收集了1。21亿用户点击4次。6800万",
"Learning:Transformer; Self-attention",
"件商品和1.92万件服装来自5。5400万用户在iFashion。如",
"表1所示，81。3点击服装的%包含具有相同类别的项目，这",
"-Pipei Huang is the Corresponding author.",
"些项目已出现在同一用户的最新50个点击项目中。我们观察",
"到，用户点击服装中物品的品牌，样式和图案也很有可能出现",
"Permission to make digital or hard copies of all or part of this work for personal on",
"在用户最新点击的物品中（我们为每个用户选择最新的50个",
"classroom use is granted without fee provided that copies are not made or distributed",
"for profit or commercial advantage and that copies bear this notice and the full cita",
"物品）。图2说明了三个示例用户在物品和服装上的行为。所",
"tion on the first page. Copyrights for components of this work owned by others than",
"有这些结果表明，用户倾向于在单个物品和服装中保持相似",
"the author(s) must be honored. Abstracting with credit is permitted. To copy other",
"的口味。我们发现，尽管近年来对时装的生成和推荐进行了深",
"wise, or republish, to post on servers or to redistribute to lists,requires prior specific",
"permission and/or a fee. Request permissions from permissions@acm.org.",
"入研究，但现有的作品通常会分别研究这两个要求",
"KDD'19,August4-8,2019,Anchorage,AK,USA",
"@ 2019 Copyright held by the owner/author(s).Publication rights licensed to ACM.",
"Thttps:/lookbook.nu/",
"ACMISBN978-1-4503-6201-6/19/08..$15.00",
"2http:/chictopia.com/",
"https://doi.org/10.1145/3292500.3330652",
"3https://www.alibabagroup.com/en/news/press_pdf/p180504.pdf"
] |
{
"arxiv_id": "1905.01866",
"language": "zh",
"page_idx": 1
}
|
||
1905.01866-zh-2
|
1905.01866-zh-2.jpg
|
Table 1: The percentage of the same brand, category, style, and pattern of the items in users' clicked outfits appearing in users' latest clicked items.
<table>
<tbody>
<tr>
<th>$\operatorname{Property}$</th>
<th>brand</th>
<th>category</th>
<th>style</th>
<th>pattern</th>
</tr>
<tr>
<td>Percentage</td>
<td>$56.9\%$</td>
<td>81.3 $\%$</td>
<td>$73.0\%$</td>
<td>$53.9\%$</td>
</tr>
</tbody>
</table>
因此，我们试图在拥有数百万用户的现实应用程序中建立时装生成和推荐之间的桥梁。更具体地说，我们通过从用户对时尚物品的历史互动中捕捉用户的兴趣和品味来生成个性化服装。对于兼容性要求，我们通过学习服装中每个项目与所有其他项目之间的兼容性来提出 FashionOutfitM 模型 (FOM)。每个项目都应该与服装中的其他项目具有不同的加权交互。因此，我们根据自我注意机制$\overline{\mathbb{E}}[29]$ 设置了一个蒙版项目预测任务，该任务在服装中一次蒙版一个项目，并根据上下文从服装中的其他项目中预测蒙版项目。针对个性化需求，通过将用户偏好集成到预先训练的 FOM 中，我们提出了一种 P 个性化的 OutfitGeneration (POG) 模型，该模型可以根据用户的近期行为生成兼容且个性化的服装。具体来说，POG 使用Transformer 编码器-解码器架构$\mathbb{E}[29]$ 对来自用户偏好和装备兼容性的信号进行建模。据我们所知，这是第一个使用编码器-解码器框架根据用户的历史行为生成个性化服装的研究。最后，我们开发了一个名为 Dida 的平台，其中已部署了 POG, 以帮助大规模在线应用程序 iFashion 中的服装生成和推荐
这项工作的贡献总结如下：
(1) We propose POG: an encoder-decoder model to generate
personalized fashion outfits, which takes into account both outfit compatibility and user personalization. It makes fashion recommendations by generating personalized outfits based
on users' recent behaviors.
(2) We demonstrate that our model significantly outperforms
other alternative methods through outfit compatibility experiments, including pushing the FITB (Fill In The Blanks) benchmark to 68.79%(5.98% relative improvement) and CP (Compatibility Prediction) benchmark to 86.32 %(25.81 % relative improvement).
POG: Personalized Outfit Generation for Fashion Recommendation at Alibaba iFashion
<FigureHere>
Figure 1: A sample of iFashion application in Taobao .We recommend fashion outfits (sets of fashion items which interact with each other) to users.
KDD'19, August 4-8, 2019, Anchorage, AK, USA
<FigureHere>
Figure 2: IIlustration of user clicked items and outfits. User (a) is a young girl who likes clothes and outfits in light colors and sweet styles. User (b) is a college boy who clicks several winter outfits after clicking a lot of winter clothes. User (c) is probably an office lady who prefers OL style items as well as outfits.
(3) We deploy POG on the real-world outfit generation platform
Dida. Through extensive online experiments, we show that POG clearly outperforms the CF method by 70% increase in CTR (Click-Through-Rate) metric.
$(4)\text{ We release a dataset }^4$ of 1.01 million outfits, 583 thousand
fashion items associated with rich context information, and
0.28 billion user click actions from 3.57 million users.
# 2 $\textbf{RELATED WORK}$
时尚是计算机视觉和多媒体的重要应用领域。在该领域已经进行了大量的研究工作，重点是时尚图像检索[2,32],服装识别[16],服装解析[12,33],属性学习[5,15],服装兼容性[3,17,27,28,30]和时尚推荐[4,8,15] 。这项工作的目标是根据用户行为自动组成个性化的时尚服装，因此我们专注于服装生成和推荐的研究领域
# 2.1 Fashion Outfit Generation
时尚服装生成方法通常属于以下两类之一。第一类方法侧重于计算成对兼容性度量[17,23,30] 。McAuley 等人。提取视觉特征以模拟人类对亚马逊共同购买数据集 [17] 的一对物品的视觉偏好。Siameses 网络 [30] 基于大规模用户行为数据中的共现估计成对兼容性。属于第二类的方法，例如[11]和[3]中提出的方法，是基于将时装建模为一组或有序序列。Li 等人。部署端到端深度学习系统，该系统可以将给定的装备分类为流行或不受欢迎的[11] 。Han 等人。训练双向 LSTM 模型以顺序生成服装[3] 。这些方法通常使用简单的项目向量池来表示服装，或者严重依赖服装项目的顺序。我们注意到，属于这两个类别的方法几乎不考虑服装中物品之间的所有相互作用。此外，将服装视为有序序列是不合理的，因为在服装中洗牌项目对其兼容性没有影响。我们尝试通过要求每个项目相对于一个装备中的其他项目应具有不同的交互权重来将其明确地纳人我们的建模体系结构中。例如，一件“红色衬衫”应该与“蓝色牛仔裤”具有较高的交互权重，而与一双“白手套”具有较小的权重
2.2 Fashion Outfit Recommendation
早期的推荐工作通常使用协作过滤来根据用户的行为历史对用户的偏好进行建模[14,22] 。但是 ,以前的工作大多将注意
$\overline{{^4}\text{https://github.c}}$om/wenyuer/POG
|
[
[
0.08660888671875,
0.0770263671875,
0.5048828125,
0.0770263671875,
0.5048828125,
0.08709716796875,
0.08660888671875,
0.08709716796875
],
[
0.67236328125,
0.07574462890625,
0.9111328125,
0.07574462890625,
0.9111328125,
0.0889892578125,
0.67236328125,
0.0889892578125
],
[
0.09478759765625,
0.12249755859375,
0.148681640625,
0.12371826171875,
0.148681640625,
0.1351318359375,
0.09478759765625,
0.1337890625
],
[
0.146240234375,
0.126953125,
0.2115478515625,
0.126953125,
0.2115478515625,
0.1307373046875,
0.146240234375,
0.1307373046875
],
[
0.21728515625,
0.126953125,
0.226318359375,
0.126953125,
0.226318359375,
0.1307373046875,
0.21728515625,
0.1307373046875
],
[
0.51953125,
0.1287841796875,
0.5439453125,
0.1287841796875,
0.5439453125,
0.147705078125,
0.51953125,
0.147705078125
],
[
0.52392578125,
0.14453125,
0.5390625,
0.14453125,
0.5390625,
0.154052734375,
0.52392578125,
0.154052734375
],
[
0.8095703125,
0.1483154296875,
0.81787109375,
0.1483154296875,
0.81787109375,
0.1514892578125,
0.8095703125,
0.1514892578125
],
[
0.423095703125,
0.1673583984375,
0.4599609375,
0.1673583984375,
0.4599609375,
0.1787109375,
0.423095703125,
0.1787109375
],
[
0.52392578125,
0.1673583984375,
0.54150390625,
0.1673583984375,
0.54150390625,
0.18310546875,
0.52392578125,
0.18310546875
],
[
0.191162109375,
0.1767578125,
0.19775390625,
0.1787109375,
0.193603515625,
0.186279296875,
0.1871337890625,
0.18310546875
],
[
0.283447265625,
0.18310546875,
0.346435546875,
0.18310546875,
0.346435546875,
0.1932373046875,
0.283447265625,
0.1932373046875
],
[
0.525390625,
0.1817626953125,
0.53759765625,
0.1817626953125,
0.53759765625,
0.1925048828125,
0.525390625,
0.1925048828125
],
[
0.1690673828125,
0.1900634765625,
0.2197265625,
0.1900634765625,
0.2197265625,
0.1944580078125,
0.1690673828125,
0.1944580078125
],
[
0.52392578125,
0.203857421875,
0.54150390625,
0.203857421875,
0.54150390625,
0.2171630859375,
0.52392578125,
0.2171630859375
],
[
0.529296875,
0.2215576171875,
0.5361328125,
0.2215576171875,
0.5361328125,
0.2247314453125,
0.529296875,
0.2247314453125
],
[
0.87353515625,
0.2203369140625,
0.8798828125,
0.2203369140625,
0.8798828125,
0.2247314453125,
0.87353515625,
0.2247314453125
],
[
0.26953125,
0.238037109375,
0.406982421875,
0.238037109375,
0.406982421875,
0.248779296875,
0.26953125,
0.248779296875
],
[
0.419189453125,
0.238037109375,
0.450927734375,
0.240478515625,
0.449462890625,
0.250732421875,
0.418212890625,
0.248779296875
],
[
0.51953125,
0.238037109375,
0.91015625,
0.238037109375,
0.91015625,
0.248779296875,
0.51953125,
0.248779296875
],
[
0.51708984375,
0.250732421875,
0.91259765625,
0.250732421875,
0.91259765625,
0.264404296875,
0.51708984375,
0.264404296875
],
[
0.26708984375,
0.264404296875,
0.346435546875,
0.264404296875,
0.346435546875,
0.274658203125,
0.26708984375,
0.274658203125
],
[
0.370849609375,
0.266357421875,
0.404296875,
0.266357421875,
0.404296875,
0.27587890625,
0.370849609375,
0.27587890625
],
[
0.51708984375,
0.264404296875,
0.91259765625,
0.264404296875,
0.91259765625,
0.27783203125,
0.51708984375,
0.27783203125
],
[
0.12335205078125,
0.272705078125,
0.1502685546875,
0.272705078125,
0.1502685546875,
0.276611328125,
0.12335205078125,
0.276611328125
],
[
0.1871337890625,
0.272705078125,
0.214111328125,
0.271484375,
0.214111328125,
0.27587890625,
0.1871337890625,
0.276611328125
],
[
0.1690673828125,
0.274658203125,
0.178955078125,
0.274658203125,
0.178955078125,
0.27978515625,
0.1690673828125,
0.27978515625
],
[
0.1004638671875,
0.2734375,
0.12335205078125,
0.2734375,
0.12335205078125,
0.2822265625,
0.1004638671875,
0.2822265625
],
[
0.281005859375,
0.279052734375,
0.294189453125,
0.27978515625,
0.293212890625,
0.290283203125,
0.279296875,
0.2890625
],
[
0.294189453125,
0.27978515625,
0.3447265625,
0.27978515625,
0.3447265625,
0.290283203125,
0.294189453125,
0.290283203125
],
[
0.36767578125,
0.27978515625,
0.414306640625,
0.27978515625,
0.414306640625,
0.290283203125,
0.36767578125,
0.290283203125
],
[
0.5185546875,
0.279052734375,
0.9111328125,
0.279052734375,
0.9111328125,
0.2890625,
0.5185546875,
0.2890625
],
[
0.098876953125,
0.287353515625,
0.1258544921875,
0.287353515625,
0.1258544921875,
0.294189453125,
0.098876953125,
0.294189453125
],
[
0.1683349609375,
0.28857421875,
0.1748046875,
0.28857421875,
0.1748046875,
0.29345703125,
0.1683349609375,
0.29345703125
],
[
0.51708984375,
0.292236328125,
0.91259765625,
0.292236328125,
0.91259765625,
0.30615234375,
0.51708984375,
0.30615234375
],
[
0.08416748046875,
0.30810546875,
0.48046875,
0.306884765625,
0.48046875,
0.32080078125,
0.08416748046875,
0.3212890625
],
[
0.51708984375,
0.306884765625,
0.583984375,
0.306884765625,
0.583984375,
0.318115234375,
0.51708984375,
0.318115234375
],
[
0.08660888671875,
0.323974609375,
0.48193359375,
0.323974609375,
0.48193359375,
0.333984375,
0.08660888671875,
0.333984375
],
[
0.0849609375,
0.337158203125,
0.294189453125,
0.337158203125,
0.294189453125,
0.347900390625,
0.0849609375,
0.347900390625
],
[
0.53759765625,
0.335205078125,
0.9111328125,
0.335205078125,
0.9111328125,
0.34521484375,
0.53759765625,
0.34521484375
],
[
0.55615234375,
0.347900390625,
0.9140625,
0.347900390625,
0.9140625,
0.361083984375,
0.55615234375,
0.361083984375
],
[
0.5556640625,
0.35986328125,
0.9140625,
0.361083984375,
0.9140625,
0.375,
0.5556640625,
0.373779296875
],
[
0.5595703125,
0.376953125,
0.76318359375,
0.376953125,
0.76318359375,
0.386962890625,
0.5595703125,
0.386962890625
],
[
0.08416748046875,
0.38818359375,
0.4794921875,
0.389404296875,
0.4794921875,
0.404052734375,
0.08416748046875,
0.40283203125
],
[
0.53515625,
0.389404296875,
0.9140625,
0.389404296875,
0.9140625,
0.40283203125,
0.53515625,
0.40283203125
],
[
0.08416748046875,
0.40283203125,
0.48193359375,
0.404052734375,
0.48193359375,
0.41845703125,
0.08416748046875,
0.417236328125
],
[
0.5556640625,
0.401611328125,
0.9140625,
0.40283203125,
0.9140625,
0.416748046875,
0.5556640625,
0.415283203125
],
[
0.0849609375,
0.417236328125,
0.27783203125,
0.417236328125,
0.27783203125,
0.43115234375,
0.0849609375,
0.43115234375
],
[
0.55615234375,
0.416748046875,
0.8759765625,
0.417236328125,
0.8759765625,
0.43115234375,
0.55615234375,
0.429931640625
],
[
0.2548828125,
0.4375,
0.324462890625,
0.43994140625,
0.32275390625,
0.455078125,
0.253173828125,
0.451904296875
],
[
0.386474609375,
0.439453125,
0.438720703125,
0.439453125,
0.438720703125,
0.453857421875,
0.386474609375,
0.453857421875
],
[
0.2091064453125,
0.43994140625,
0.256591796875,
0.43994140625,
0.256591796875,
0.451904296875,
0.2091064453125,
0.451904296875
],
[
0.1380615234375,
0.43994140625,
0.193603515625,
0.43994140625,
0.193603515625,
0.451904296875,
0.1380615234375,
0.451904296875
],
[
0.340576171875,
0.43994140625,
0.373291015625,
0.43994140625,
0.373291015625,
0.45263671875,
0.340576171875,
0.45263671875
],
[
0.5185546875,
0.4462890625,
0.53173828125,
0.4462890625,
0.53173828125,
0.455078125,
0.5185546875,
0.455078125
],
[
0.5361328125,
0.447509765625,
0.54345703125,
0.447509765625,
0.54345703125,
0.45263671875,
0.5361328125,
0.45263671875
],
[
0.54736328125,
0.443115234375,
0.6953125,
0.443115234375,
0.6953125,
0.45703125,
0.54736328125,
0.45703125
],
[
0.1282958984375,
0.45703125,
0.205078125,
0.458984375,
0.205078125,
0.473388671875,
0.1282958984375,
0.4716796875
],
[
0.334228515625,
0.458251953125,
0.378173828125,
0.458251953125,
0.378173828125,
0.472900390625,
0.334228515625,
0.472900390625
],
[
0.3896484375,
0.458251953125,
0.43701171875,
0.458251953125,
0.43701171875,
0.472900390625,
0.3896484375,
0.472900390625
],
[
0.273681640625,
0.458984375,
0.31787109375,
0.458984375,
0.31787109375,
0.47021484375,
0.273681640625,
0.47021484375
],
[
0.2017822265625,
0.458984375,
0.2548828125,
0.458984375,
0.2548828125,
0.47021484375,
0.2017822265625,
0.47021484375
],
[
0.5185546875,
0.462158203125,
0.91259765625,
0.462158203125,
0.91259765625,
0.47607421875,
0.5185546875,
0.47607421875
],
[
0.51708984375,
0.474853515625,
0.91259765625,
0.47607421875,
0.91259765625,
0.490478515625,
0.51708984375,
0.4892578125
],
[
0.10296630859375,
0.488037109375,
0.48046875,
0.488037109375,
0.48046875,
0.501953125,
0.10296630859375,
0.501953125
],
[
0.51953125,
0.490478515625,
0.91259765625,
0.490478515625,
0.91259765625,
0.50390625,
0.51953125,
0.50390625
],
[
0.08660888671875,
0.501953125,
0.48046875,
0.501953125,
0.48046875,
0.51513671875,
0.08660888671875,
0.51513671875
],
[
0.51708984375,
0.50244140625,
0.9140625,
0.50390625,
0.9140625,
0.517578125,
0.51708984375,
0.5166015625
],
[
0.08660888671875,
0.5166015625,
0.48046875,
0.5166015625,
0.48046875,
0.52978515625,
0.08660888671875,
0.52978515625
],
[
0.51953125,
0.517578125,
0.9140625,
0.517578125,
0.9140625,
0.53076171875,
0.51953125,
0.53076171875
],
[
0.0882568359375,
0.53076171875,
0.48046875,
0.53076171875,
0.48046875,
0.54150390625,
0.0882568359375,
0.54150390625
],
[
0.5185546875,
0.52978515625,
0.66650390625,
0.53076171875,
0.66650390625,
0.54443359375,
0.5185546875,
0.54345703125
],
[
0.0882568359375,
0.54345703125,
0.48046875,
0.54345703125,
0.48046875,
0.556640625,
0.0882568359375,
0.556640625
],
[
0.08660888671875,
0.55615234375,
0.4794921875,
0.55615234375,
0.4794921875,
0.5693359375,
0.08660888671875,
0.5693359375
],
[
0.5185546875,
0.56005859375,
0.56201171875,
0.56005859375,
0.56201171875,
0.57080078125,
0.5185546875,
0.57080078125
],
[
0.5556640625,
0.5595703125,
0.77783203125,
0.5595703125,
0.77783203125,
0.57275390625,
0.5556640625,
0.57275390625
],
[
0.0882568359375,
0.57275390625,
0.47802734375,
0.57275390625,
0.47802734375,
0.58349609375,
0.0882568359375,
0.58349609375
],
[
0.5185546875,
0.5771484375,
0.91259765625,
0.5771484375,
0.91259765625,
0.59033203125,
0.5185546875,
0.59033203125
],
[
0.08660888671875,
0.58544921875,
0.48046875,
0.58544921875,
0.48046875,
0.5986328125,
0.08660888671875,
0.5986328125
],
[
0.51953125,
0.59033203125,
0.91259765625,
0.59033203125,
0.91259765625,
0.60400390625,
0.51953125,
0.60400390625
],
[
0.08660888671875,
0.5986328125,
0.48046875,
0.5986328125,
0.48046875,
0.6123046875,
0.08660888671875,
0.6123046875
],
[
0.5185546875,
0.60498046875,
0.91259765625,
0.60498046875,
0.91259765625,
0.61572265625,
0.5185546875,
0.61572265625
],
[
0.0849609375,
0.6123046875,
0.4794921875,
0.611328125,
0.4794921875,
0.62548828125,
0.0849609375,
0.626953125
],
[
0.5185546875,
0.61865234375,
0.91259765625,
0.61865234375,
0.91259765625,
0.6318359375,
0.5185546875,
0.6318359375
],
[
0.08660888671875,
0.626953125,
0.4794921875,
0.626953125,
0.4794921875,
0.64013671875,
0.08660888671875,
0.64013671875
],
[
0.5185546875,
0.6318359375,
0.91259765625,
0.6318359375,
0.91259765625,
0.64599609375,
0.5185546875,
0.64599609375
],
[
0.08660888671875,
0.64013671875,
0.48046875,
0.64013671875,
0.48046875,
0.65380859375,
0.08660888671875,
0.65380859375
],
[
0.51953125,
0.64794921875,
0.9111328125,
0.64794921875,
0.9111328125,
0.65771484375,
0.51953125,
0.65771484375
],
[
0.0882568359375,
0.65576171875,
0.47802734375,
0.65576171875,
0.47802734375,
0.66650390625,
0.0882568359375,
0.66650390625
],
[
0.5185546875,
0.66015625,
0.91259765625,
0.66015625,
0.91259765625,
0.673828125,
0.5185546875,
0.673828125
],
[
0.0882568359375,
0.66748046875,
0.48046875,
0.66748046875,
0.48046875,
0.68115234375,
0.0882568359375,
0.68115234375
],
[
0.51708984375,
0.673828125,
0.91259765625,
0.67236328125,
0.91259765625,
0.68603515625,
0.51708984375,
0.6875
],
[
0.08660888671875,
0.681640625,
0.48046875,
0.681640625,
0.48046875,
0.69580078125,
0.08660888671875,
0.69580078125
],
[
0.51953125,
0.6875,
0.9140625,
0.6875,
0.9140625,
0.70068359375,
0.51953125,
0.70068359375
],
[
0.08660888671875,
0.69580078125,
0.48046875,
0.69580078125,
0.48046875,
0.708984375,
0.08660888671875,
0.708984375
],
[
0.5185546875,
0.70068359375,
0.91259765625,
0.70068359375,
0.91259765625,
0.7138671875,
0.5185546875,
0.7138671875
],
[
0.08660888671875,
0.708984375,
0.461669921875,
0.708984375,
0.461669921875,
0.72265625,
0.08660888671875,
0.72265625
],
[
0.5185546875,
0.71533203125,
0.91259765625,
0.71533203125,
0.91259765625,
0.72900390625,
0.5185546875,
0.72900390625
],
[
0.10455322265625,
0.724609375,
0.271240234375,
0.724609375,
0.271240234375,
0.73486328125,
0.10455322265625,
0.73486328125
],
[
0.5185546875,
0.72802734375,
0.91259765625,
0.72900390625,
0.91259765625,
0.7421875,
0.5185546875,
0.7412109375
],
[
0.10455322265625,
0.744140625,
0.48046875,
0.744140625,
0.48046875,
0.75830078125,
0.10455322265625,
0.75830078125
],
[
0.51953125,
0.7421875,
0.91259765625,
0.7421875,
0.91259765625,
0.755859375,
0.51953125,
0.755859375
],
[
0.125,
0.75830078125,
0.48046875,
0.755859375,
0.48046875,
0.77099609375,
0.125,
0.77294921875
],
[
0.52685546875,
0.7548828125,
0.91259765625,
0.7568359375,
0.91259765625,
0.77099609375,
0.52685546875,
0.76806640625
],
[
0.125,
0.771484375,
0.483642578125,
0.77099609375,
0.483642578125,
0.78515625,
0.125,
0.7861328125
],
[
0.5185546875,
0.77099609375,
0.91259765625,
0.77099609375,
0.91259765625,
0.7841796875,
0.5185546875,
0.7841796875
],
[
0.1258544921875,
0.7861328125,
0.496826171875,
0.7861328125,
0.496826171875,
0.7998046875,
0.1258544921875,
0.7998046875
],
[
0.51708984375,
0.78271484375,
0.9140625,
0.78271484375,
0.9140625,
0.796875,
0.51708984375,
0.796875
],
[
0.125,
0.7998046875,
0.289306640625,
0.7998046875,
0.289306640625,
0.81298828125,
0.125,
0.81298828125
],
[
0.51953125,
0.7998046875,
0.60693359375,
0.7998046875,
0.60693359375,
0.81103515625,
0.51953125,
0.81103515625
],
[
0.10296630859375,
0.81298828125,
0.48046875,
0.81298828125,
0.48046875,
0.8271484375,
0.10296630859375,
0.8271484375
],
[
0.51953125,
0.82763671875,
0.54345703125,
0.82763671875,
0.54345703125,
0.8359375,
0.51953125,
0.8359375
],
[
0.5458984375,
0.82763671875,
0.560546875,
0.82763671875,
0.560546875,
0.83544921875,
0.5458984375,
0.83544921875
],
[
0.55615234375,
0.82470703125,
0.830078125,
0.82568359375,
0.830078125,
0.8388671875,
0.55615234375,
0.83837890625
],
[
0.1258544921875,
0.84130859375,
0.48193359375,
0.84130859375,
0.48193359375,
0.85498046875,
0.1258544921875,
0.85498046875
],
[
0.12744140625,
0.8291015625,
0.48046875,
0.8291015625,
0.48046875,
0.8388671875,
0.12744140625,
0.8388671875
],
[
0.521484375,
0.8447265625,
0.9111328125,
0.8447265625,
0.9111328125,
0.85498046875,
0.521484375,
0.85498046875
],
[
0.1258544921875,
0.85498046875,
0.4794921875,
0.85498046875,
0.4794921875,
0.86865234375,
0.1258544921875,
0.86865234375
],
[
0.51953125,
0.85693359375,
0.91259765625,
0.85693359375,
0.91259765625,
0.87060546875,
0.51953125,
0.87060546875
],
[
0.12744140625,
0.87060546875,
0.48193359375,
0.87060546875,
0.48193359375,
0.880859375,
0.12744140625,
0.880859375
],
[
0.12335205078125,
0.88330078125,
0.26318359375,
0.88330078125,
0.26318359375,
0.896484375,
0.12335205078125,
0.896484375
],
[
0.515625,
0.8818359375,
0.6806640625,
0.88330078125,
0.6806640625,
0.896484375,
0.515625,
0.89599609375
]
] |
[
"POG:Personalized Outfit Generation for Fashion Recommendation at Alibaba iFashion",
"KDD'19, August 4-8, 2019,Anchorage, AK, USA",
"请的女",
"",
"t",
"",
"(a)",
"",
"outfit",
"®",
"9",
"ASHION",
"(b)",
"rS小",
"国",
"",
"",
"演绎甜美乖巧的小腔调",
"title",
"Figure 2:Illustration of user clicked items and outfits.User",
"(a) is a young girl who likes clothes and outfits in light colors",
"fashion expert",
"views",
"and sweet styles. User (b) is a college boy who clicks several",
"收",
"3nm",
"",
"本连衣租",
"9",
"喵喵小禅",
"@6081",
"winter outfitsafterclickingalot of winter clothes.User (c)",
"d",
"3",
"is probably an office lady who prefers OL style items as well",
"Figure1:A sampleof iFashionapplicationinTaobao.We",
"as outfits.",
"recommend fashionoutfits (sets offashionitems which in-",
"teract with each other)to users",
"(3)WedeployPOGonthereal-worldoutfitgenerationplatform",
"Dida . Through extensive online experiments, we show that",
"POG clearly outperforms the CF method by 70 % increase in",
"CTR (Click-Through-Rate)metric.",
"Table 1: The percentage of the same brand, category, style,",
"(4) We release a dataset 4 of 1.01 million outfits, 583 thousand",
"and pattern of the items in users' clicked outfits appearing",
"fashion items associated with rich context information, and",
"in users'latest clicked items.",
"0.28billion user click actions from3.57million users.",
"category",
"pattern",
"brand",
"Property",
"style",
"2",
"",
"RELATED WORK",
"Percentage",
"73.0 %",
"53.9%",
"81.3%",
"156.9%",
"时尚是计算机视觉和多媒体的重要应用领域。在该领域已经",
"进行了大量的研究工作，重点是时尚图像检索[2，32]，服装",
"因此，我们试图在拥有数百万用户的现实应用程序中建立",
"识别[16]，服装解析[12,33]，属性学习[5,15]，服装兼容性",
"时装生成和推荐之间的桥梁。更具体地说，我们通过从用户对",
"[3,17,27,28,30]和时尚推荐[4,8,15]。这项工作的目标是根据",
"时尚物品的历史互动中捕提用户的兴趣和品味来生成个性化",
"用户行为自动组成个性化的时尚服装，因此我们专注于服装",
"服装。对于兼容性要求，我们通过学习服装中每个项目与所",
"生成和推荐的研究领域",
"有其他项目之间的兼容性来提出FashionOutfitM模型（FOM)。",
"每个项目都应该与服装中的其他项目具有不同的加权交互。因",
"2.1",
"Fashion Outfit Generation",
"此，我们根据自我注意机制E[29]设置了一个蒙版项目预测",
"时尚服装生成方法通常属于以下两类之一。第一类方法侧重",
"任务，该任务在服装中一次蒙版一个项目，并根据上下文从",
"于计算成对兼容性度量[17,23,30]。McAuley等人。提取视觉",
"服装中的其他项目中预测蒙版项目。针对个性化需求，通过",
"特征以模拟人类对亚马逊共同购买数据集【17】的一对物品的",
"将用户偏好集成到预先训练的FOM中，我们提出了一种P个",
"视觉偏好。Siameses网络[30]基于大规模用户行为数据中的",
"性化的OutfitGeneration（POG）模型，该模型可以根据用户",
"共现估计成对兼容性。属于第二类的方法，例如[11]和[3]中",
"的近期行为生成兼容且个性化的服装。具体来说，POG使用",
"提出的方法，是基于将时装建模为一组或有序序列。Li等人。",
"Transformer编码器-解码器架构E[29]对来自用户偏好和装备",
"部署端到端深度学习系统，该系统可以将给定的装备分类为",
"兼容性的信号进行建模。据我们所知，这是第一个使用编码",
"流行或不受欢迎的[11]。Han等人。训练双向LSTM模型以",
"器-解码器框架根据用户的历史行为生成个性化服装的研究。",
"顺序生成服装[3]。这些方法通常使用简单的项目向量池来表",
"最后，我们开发了一个名为Dida的平台，其中已部署了POG，",
"示服装，或者严重依赖服装项目的顺序。我们注意到，属于这",
"以帮助大规模在线应用程序iFashion中的服装生成和推荐",
"两个类别的方法几乎不考虑服装中物品之间的所有相互作用。",
"这项工作的贡献总结如下：",
"此外，将服装视为有序序列是不合理的，因为在服装中洗牌项",
"(1) We propose POG: an encoder-decoder model to generate",
"目对其兼容性没有影响。我们尝试通过要求每个项目相对于",
"personalized fashion outfits, which takes into account both",
"一个装备中的其他项目应具有不同的交互权重来将其明确地",
"outfit compatibility and user personalization. It makes fash-",
"纳人我们的建模体系结构中。例如，一件“红色衬衫”应该与",
"ion recommendations by generating personalized outfits based",
"“蓝色牛仔裤”具有较高的交互权重，而与一双“白手套”具",
"on users' recent behaviors.",
"有较小的权重",
"(2) We demonstrate that our model significantly outperforms",
"2.2",
"",
"FashionOutfitRecommendation",
"periments, including pushing the FITB (Fill In The Blanks)",
"otheralternativemethodsthroughoutfitcompatibilityex",
"早期的推荐工作通常使用协作过滤来根据用户的行为历史对",
"benchmark to 68.79 % (5.98 % relative improvement) and CP",
"用户的偏好进行建模[14,22]。但是，以前的工作大多将注意",
"(CompatibilityPrediction)benchmarkto86.32%(25.81%",
"relative improvement),",
"https://github.com/wenyuer/POG"
] |
{
"arxiv_id": "1905.01866",
"language": "zh",
"page_idx": 2
}
|
||
1905.01866-zh-3
|
1905.01866-zh-3.jpg
|
KDD'19, August 4-8, 2019, Anchorage, AK, USA
Table 2: Statistics of the datasets.
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>Outfts 井 1</th>
<th>并 Users</th>
<th>升 Items</th>
</tr>
<tr>
<td>Outfit data</td>
<td>1,013,136</td>
<td> </td>
<td>583,464</td>
</tr>
<tr>
<td>Item data</td>
<td> </td>
<td> </td>
<td>4,747,039</td>
</tr>
<tr>
<td>User data</td>
<td>127,169</td>
<td>3,569,112</td>
<td>4,463,302</td>
</tr>
</tbody>
</table>
力限制在推荐单个项目上。。有几种方法可以推荐整个时尚服装。刘思等人。[15]提出一种基于属性和类别的面向场合的服装推荐方法。[17]和[3]中的工作需要图像或文本查询才能找到互补的衣服。使用功能张量分解方法向[4]中的用户建议项集。由于这些方法都需要用户查询或用户上传的数据作为输人，因此典型用户可能会认为它们对用户不那么友好。此外， 我们注意到，这些方法在大规模在线推荐系统中以有效的方式实现是相当不切实际的
2.3 $\textbf{Self- Attention and Transformer}$
自我注意是一种将单个序列 [29] 的不同位置相关联的注意机制，已成功用于各种任务[13,19]。Transformer[29]是一种依靠自我关注的转换模型，已被广泛使用，并大大提高了语言处理任务的性能$\mathbb{E}[1,21]$ 。在最近的语言表示研究中，与从左到右的 Transformer 解码器 (OpenAIGPT [21] )和双向 LSTM (ELMo[20])相比，双向 Transformer 编码器表现出了最佳性能 (BERT[1])。最近还引人了多头自我关注，对用户的行为序列进行建模，进行序贯推荐[9,25] 。与他们不同的是，在本文中，我们采用自我注意来建模时装一代中的兼容性
# 3 DATASET
Taobao 的时尚专家每天创造成千上万的服装。所有这些手工制作的时装在网上展出之前都会经过审查。关于 1。到目前为止，已经以这种方式创建和审查了 4300 万套服装。我们收集了 80 个最常见的叶子类别 (例如。g。毛衣、外套、t 恤、靴子和戒指)。不常见类别中的所有物品都将从服装中删除。在本文的其余部分中，我们仅考虑这 80 个叶子类别中的项目。之后，我们过滤包含少于 4 个项目的服装。总计 1。剩下 01 万件服装，由 58.3 万件单独物品组成。此外，我们收集了最近三个月 iFashi-on 用户对项目和服装的点击。我们从 3 中选择点击行为。5700 万活跃用户，总共观看了40 多套服装。只有当同一用户之前点击超过 10 个项目时，才会记录服装上的点击次数。然后，我们通过将服装点击与来自同一用户的最新 50项点击配对来构建训练样本。最后，我们得到 19。200 万个训练样本，其中包括 4个。4600万件商品和 12.7 万件服装。我们数据集中的每个项目都与白色背景图像，标题和叶子类别相关联
据我们所知，与现有数据集相比，我们的数据集是最大的公开可用的时尚物品数据集，具有丰富的信息。此外，我们是第一个提供服装数据和相关用户行为数据的人，这些数据可用于未来的时尚推荐研究。我们提供三个数据集，分别描述服装、物品和用户行为。数据集的统计信息如表 2 所示
# 4 мЕТНОрогобҮ
在本节中，我们将详细介绍我们的方法。POG 在三个步骤的过程中构建：我们首先嵌人项目。其次，我们构建 FOM,它学习服装内物品的兼容性。最后，一旦训练完成，我们使用生成的预训练 FOM 在 Transformer 架构上初始化 POG
Wen Chen and Pipei Huang, et al.
作为第一步，我们使用多模式嵌人模型表示所有项目。然后我们详细介绍了 FOM 和 POG。最后，我们在本节中介绍我们的 Dida 平台，这有助于确保这项工作在大规模在线应用中的效率和质量要求
# 4.1 Multi-modal Embedding
对于每个时尚商品$f$ , 我们计算嵌入$f$的非线性特征。时尚的概念主要依赖于视觉和文本信息。大多数以前的作品建议利用图像和文本来学习多模式嵌人[3,11]。在我们的工作中，我们使用多模式嵌人模型，该模型为每个项目生成以下输人：(1) 密集向量编码来自 CNN 模型的项目的白色背景图片，(2)密集向量编码从 TextCNN 网络获得的项目的标题，该网络已经过预先训练，可以根据其标题预测项目的叶子类别，(3)密集向量编码基于 Alibaba 专有的 MAXXXTHX 的项目嵌人 5
我们的目标是获得一个嵌人空间，其中相似的项目嵌入在附近，不同的项目位于不同的区域。我们连接从图像，文本和CF 派生的嵌入作为最终完全连接层的输入。输出为$d_e$维向量$f$。对于每个项目$f$ ,我们将正样本$f^+$定义为与$f$属于同一叶子类别的项目，因此负样本$f^-$是不属于同一叶子类别的项目。然后使用三重态损失通过以下方式训练整个网络：
$$\mathcal{L}_{E}=\sum_{f}\max(d(f,f^{+})-d(f,f^{-})+\alpha,0)$$
(1)
,其中距离度量$d$表示欧几里得距离$,\alpha$表示余量。通过最小化$\mathcal{L}_E$,同一类别中的$f$和$f^+$之间的距离被迫小于与不同类别中的$f^-$之间的距离$\alpha$
# 4.2 FOM: Fashion Outfit Model
Outfit 是一组项目，其中每个项目应与服装中的其他项目具有不同的加权交互。为了捕获服装中的物品交互，我们设计了一个基于双向 Transformer 编码器架构的屏蔽物品预测任务。在服装中一次掩盖一个项目，我们要求模型根据上下文用正确的项目填空。由于服装中的每个物品都被遮盖以融合其左右上下文，因此可以从自我注意机制中学习每个物品与服装中所有其他物品之间的兼容性。因此，可以通过组合所有这些项目的兼容性来学习服装中的兼容性
令$\mathcal{F}$ 表示所有服装的集合。给定装备$F\in\mathcal{F},F=\{f_1,\ldots,f_t,\ldots,f_n\}$
,其中$f_t$是第$t$项。令$f_t$为从具有$d_e$维的多模态嵌人派生的项$f_t$的表示形式。我们为蒙版项目使用特定的嵌入[MASK]。非屏蔽项由其多模态嵌入表示。然后，我们将输人嵌入集表示为$F_{mask}$。给定 $F_{mask}$ ,任务是预测屏蔽项，而不是重建整个装备。更正式地说，我们最小化 FOM 的以下损失函数：
(2)
$$\mathcal{L}_{F}=-\frac{1}{n}\sum_{mask=1}^{n}\log Pr(f_{mask}|F_{mask};\Theta_{F})$$
其中$\Theta_F$表示模型参数，$Pr(\cdot)$是在非屏蔽项的条件下选择正
确项的概率
模型架构如图 3 所示。我们不像 Transformer [29] 那样使用位置嵌人，因为我们将装备中的项目作为一组，而不是具有位置信息的序列。让$F_{mask}$ 通过两个完全连接的层，其间有一个整流线性单元(ReLU)激活，以将所有输人嵌人从单个项目空间传输到装备空间。我们将两个完全连接的层称为过渡层，并将输出表示为$H^0\in\mathbb{R}^{n\times d_m}:$
$$H^0=\text{ReLU}(\begin{bmatrix}f_1^T\:;\ldots;f_n^T\end{bmatrix}^TW_0^F+b_0^F)W_1^F+b_1^F$$
(3)
|
[
[
0.08660888671875,
0.07574462890625,
0.32177734375,
0.0770263671875,
0.32177734375,
0.08837890625,
0.08660888671875,
0.08709716796875
],
[
0.74853515625,
0.07574462890625,
0.90673828125,
0.0770263671875,
0.90673828125,
0.08837890625,
0.74853515625,
0.08709716796875
],
[
0.17236328125,
0.10986328125,
0.39208984375,
0.111083984375,
0.39208984375,
0.125,
0.17236328125,
0.12371826171875
],
[
0.53369140625,
0.106689453125,
0.91259765625,
0.10797119140625,
0.91259765625,
0.121826171875,
0.53369140625,
0.12060546875
],
[
0.51708984375,
0.121826171875,
0.91259765625,
0.121826171875,
0.91259765625,
0.1351318359375,
0.51708984375,
0.1351318359375
],
[
0.1422119140625,
0.1300048828125,
0.19287109375,
0.1319580078125,
0.191162109375,
0.146484375,
0.1405029296875,
0.14453125
],
[
0.303955078125,
0.1319580078125,
0.3544921875,
0.1319580078125,
0.3544921875,
0.146484375,
0.303955078125,
0.146484375
],
[
0.374267578125,
0.1300048828125,
0.4248046875,
0.1319580078125,
0.423095703125,
0.146484375,
0.373291015625,
0.14453125
],
[
0.5185546875,
0.1351318359375,
0.91259765625,
0.1351318359375,
0.91259765625,
0.1483154296875,
0.5185546875,
0.1483154296875
],
[
0.5185546875,
0.15087890625,
0.62353515625,
0.15087890625,
0.62353515625,
0.1634521484375,
0.5185546875,
0.1634521484375
],
[
0.2279052734375,
0.1331787109375,
0.281982421875,
0.1331787109375,
0.281982421875,
0.14453125,
0.2279052734375,
0.14453125
],
[
0.1446533203125,
0.1514892578125,
0.2197265625,
0.1514892578125,
0.2197265625,
0.1634521484375,
0.1446533203125,
0.1634521484375
],
[
0.2254638671875,
0.15283203125,
0.281982421875,
0.15283203125,
0.281982421875,
0.1634521484375,
0.2254638671875,
0.1634521484375
],
[
0.374267578125,
0.1514892578125,
0.423095703125,
0.1514892578125,
0.423095703125,
0.1634521484375,
0.374267578125,
0.1634521484375
],
[
0.1405029296875,
0.1634521484375,
0.2034912109375,
0.1654052734375,
0.2034912109375,
0.179931640625,
0.1405029296875,
0.177978515625
],
[
0.343017578125,
0.1654052734375,
0.420654296875,
0.1654052734375,
0.420654296875,
0.1767578125,
0.343017578125,
0.1767578125
],
[
0.51708984375,
0.173583984375,
0.76318359375,
0.175537109375,
0.76318359375,
0.1900634765625,
0.51708984375,
0.1881103515625
],
[
0.143798828125,
0.18115234375,
0.2034912109375,
0.18115234375,
0.2034912109375,
0.1925048828125,
0.143798828125,
0.1925048828125
],
[
0.236083984375,
0.179931640625,
0.4248046875,
0.179931640625,
0.4248046875,
0.1932373046875,
0.236083984375,
0.1932373046875
],
[
0.5185546875,
0.1932373046875,
0.91259765625,
0.1932373046875,
0.91259765625,
0.20703125,
0.5185546875,
0.20703125
],
[
0.0882568359375,
0.210205078125,
0.4794921875,
0.210205078125,
0.4794921875,
0.2203369140625,
0.0882568359375,
0.2203369140625
],
[
0.51953125,
0.208984375,
0.9111328125,
0.208984375,
0.9111328125,
0.2197265625,
0.51953125,
0.2197265625
],
[
0.08660888671875,
0.222900390625,
0.48046875,
0.222900390625,
0.48046875,
0.236083984375,
0.08660888671875,
0.236083984375
],
[
0.5185546875,
0.2215576171875,
0.91259765625,
0.2215576171875,
0.91259765625,
0.23486328125,
0.5185546875,
0.23486328125
],
[
0.0882568359375,
0.238037109375,
0.48046875,
0.238037109375,
0.48046875,
0.248779296875,
0.0882568359375,
0.248779296875
],
[
0.5185546875,
0.23486328125,
0.91259765625,
0.23486328125,
0.91259765625,
0.248779296875,
0.5185546875,
0.248779296875
],
[
0.08660888671875,
0.250732421875,
0.48046875,
0.250732421875,
0.48046875,
0.264404296875,
0.08660888671875,
0.264404296875
],
[
0.5185546875,
0.248779296875,
0.91259765625,
0.248779296875,
0.91259765625,
0.261962890625,
0.5185546875,
0.261962890625
],
[
0.0882568359375,
0.266357421875,
0.48046875,
0.266357421875,
0.48046875,
0.276611328125,
0.0882568359375,
0.276611328125
],
[
0.51953125,
0.264404296875,
0.9111328125,
0.264404296875,
0.9111328125,
0.274658203125,
0.51953125,
0.274658203125
],
[
0.08905029296875,
0.27978515625,
0.4794921875,
0.27978515625,
0.4794921875,
0.290283203125,
0.08905029296875,
0.290283203125
],
[
0.51953125,
0.27783203125,
0.9111328125,
0.27783203125,
0.9111328125,
0.28857421875,
0.51953125,
0.28857421875
],
[
0.08660888671875,
0.292236328125,
0.47802734375,
0.29345703125,
0.47802734375,
0.304931640625,
0.08660888671875,
0.3037109375
],
[
0.5185546875,
0.290283203125,
0.88232421875,
0.290283203125,
0.88232421875,
0.3037109375,
0.5185546875,
0.3037109375
],
[
0.0849609375,
0.30615234375,
0.25,
0.30615234375,
0.25,
0.3193359375,
0.0849609375,
0.3193359375
],
[
0.53515625,
0.30615234375,
0.9111328125,
0.30615234375,
0.9111328125,
0.31640625,
0.53515625,
0.31640625
],
[
0.521484375,
0.3193359375,
0.9111328125,
0.3193359375,
0.9111328125,
0.330078125,
0.521484375,
0.330078125
],
[
0.125,
0.3359375,
0.390625,
0.337158203125,
0.390625,
0.35107421875,
0.125,
0.349853515625
],
[
0.515625,
0.330810546875,
0.91259765625,
0.33203125,
0.91259765625,
0.34521484375,
0.515625,
0.3447265625
],
[
0.08660888671875,
0.33837890625,
0.13232421875,
0.33837890625,
0.13232421875,
0.348388671875,
0.08660888671875,
0.348388671875
],
[
0.51953125,
0.347900390625,
0.89794921875,
0.347900390625,
0.89794921875,
0.35791015625,
0.51953125,
0.35791015625
],
[
0.0882568359375,
0.354736328125,
0.48046875,
0.354736328125,
0.48046875,
0.36865234375,
0.0882568359375,
0.36865234375
],
[
0.900390625,
0.35107421875,
0.90673828125,
0.35107421875,
0.90673828125,
0.354736328125,
0.900390625,
0.354736328125
],
[
0.51953125,
0.361083984375,
0.9111328125,
0.361083984375,
0.9111328125,
0.371826171875,
0.51953125,
0.371826171875
],
[
0.08905029296875,
0.37060546875,
0.4794921875,
0.37060546875,
0.4794921875,
0.380615234375,
0.08905029296875,
0.380615234375
],
[
0.51708984375,
0.37255859375,
0.861328125,
0.373779296875,
0.861328125,
0.386962890625,
0.51708984375,
0.386474609375
],
[
0.08660888671875,
0.38330078125,
0.4794921875,
0.38330078125,
0.4794921875,
0.396484375,
0.08660888671875,
0.396484375
],
[
0.0882568359375,
0.397705078125,
0.48046875,
0.397705078125,
0.48046875,
0.410888671875,
0.0882568359375,
0.410888671875
],
[
0.5908203125,
0.40087890625,
0.8408203125,
0.40087890625,
0.8408203125,
0.421630859375,
0.5908203125,
0.421630859375
],
[
0.0882568359375,
0.410888671875,
0.48046875,
0.410888671875,
0.48046875,
0.4248046875,
0.0882568359375,
0.4248046875
],
[
0.8935546875,
0.406005859375,
0.9140625,
0.406005859375,
0.9140625,
0.41845703125,
0.8935546875,
0.41845703125
],
[
0.62841796875,
0.41845703125,
0.646484375,
0.41845703125,
0.646484375,
0.431884765625,
0.62841796875,
0.431884765625
],
[
0.0882568359375,
0.425537109375,
0.47802734375,
0.425537109375,
0.47802734375,
0.436279296875,
0.0882568359375,
0.436279296875
],
[
0.08660888671875,
0.438232421875,
0.48046875,
0.438232421875,
0.48046875,
0.451904296875,
0.08660888671875,
0.451904296875
],
[
0.52392578125,
0.43994140625,
0.53173828125,
0.43994140625,
0.53173828125,
0.4443359375,
0.52392578125,
0.4443359375
],
[
0.52685546875,
0.43505859375,
0.91259765625,
0.436279296875,
0.91259765625,
0.449462890625,
0.52685546875,
0.448974609375
],
[
0.08660888671875,
0.451904296875,
0.48193359375,
0.451904296875,
0.48193359375,
0.46533203125,
0.08660888671875,
0.46533203125
],
[
0.51953125,
0.45068359375,
0.91259765625,
0.45068359375,
0.91259765625,
0.464111328125,
0.51953125,
0.464111328125
],
[
0.08905029296875,
0.46728515625,
0.408447265625,
0.46728515625,
0.408447265625,
0.477783203125,
0.08905029296875,
0.477783203125
],
[
0.5185546875,
0.464111328125,
0.6748046875,
0.464111328125,
0.6748046875,
0.477783203125,
0.5185546875,
0.477783203125
],
[
0.51708984375,
0.4892578125,
0.78369140625,
0.4892578125,
0.78369140625,
0.50244140625,
0.51708984375,
0.50244140625
],
[
0.0849609375,
0.498779296875,
0.1993408203125,
0.498779296875,
0.1993408203125,
0.5087890625,
0.0849609375,
0.5087890625
],
[
0.0849609375,
0.51318359375,
0.48193359375,
0.5146484375,
0.48193359375,
0.52783203125,
0.0849609375,
0.5263671875
],
[
0.5185546875,
0.5087890625,
0.91015625,
0.5087890625,
0.91015625,
0.51953125,
0.5185546875,
0.51953125
],
[
0.5185546875,
0.521484375,
0.90185546875,
0.521484375,
0.90185546875,
0.53515625,
0.5185546875,
0.53515625
],
[
0.08905029296875,
0.53076171875,
0.48046875,
0.53076171875,
0.48046875,
0.54150390625,
0.08905029296875,
0.54150390625
],
[
0.5185546875,
0.53515625,
0.91259765625,
0.53515625,
0.91259765625,
0.548828125,
0.5185546875,
0.548828125
],
[
0.08905029296875,
0.54443359375,
0.4794921875,
0.54443359375,
0.4794921875,
0.5546875,
0.08905029296875,
0.5546875
],
[
0.5185546875,
0.55029296875,
0.9111328125,
0.55029296875,
0.9111328125,
0.56103515625,
0.5185546875,
0.56103515625
],
[
0.0882568359375,
0.556640625,
0.48046875,
0.556640625,
0.48046875,
0.57080078125,
0.0882568359375,
0.57080078125
],
[
0.51708984375,
0.56298828125,
0.91259765625,
0.56298828125,
0.91259765625,
0.5771484375,
0.51708984375,
0.5771484375
],
[
0.0849609375,
0.5693359375,
0.48193359375,
0.57080078125,
0.48193359375,
0.583984375,
0.0849609375,
0.58349609375
],
[
0.521484375,
0.5791015625,
0.9111328125,
0.5791015625,
0.9111328125,
0.5888671875,
0.521484375,
0.5888671875
],
[
0.08905029296875,
0.5859375,
0.4794921875,
0.5859375,
0.4794921875,
0.5966796875,
0.08905029296875,
0.5966796875
],
[
0.51953125,
0.59228515625,
0.9111328125,
0.59228515625,
0.9111328125,
0.60302734375,
0.51953125,
0.60302734375
],
[
0.0882568359375,
0.5986328125,
0.48046875,
0.5986328125,
0.48046875,
0.6123046875,
0.0882568359375,
0.6123046875
],
[
0.5185546875,
0.60498046875,
0.740234375,
0.60498046875,
0.740234375,
0.61865234375,
0.5185546875,
0.61865234375
],
[
0.0849609375,
0.6123046875,
0.48046875,
0.611328125,
0.48046875,
0.625,
0.0849609375,
0.62548828125
],
[
0.53369140625,
0.6171875,
0.9453125,
0.61865234375,
0.9453125,
0.63330078125,
0.53369140625,
0.6318359375
],
[
0.08905029296875,
0.62744140625,
0.4794921875,
0.62744140625,
0.4794921875,
0.63818359375,
0.08905029296875,
0.63818359375
],
[
0.529296875,
0.630859375,
0.91259765625,
0.6318359375,
0.91259765625,
0.646484375,
0.529296875,
0.64599609375
],
[
0.08660888671875,
0.64013671875,
0.48046875,
0.64013671875,
0.48046875,
0.65380859375,
0.08660888671875,
0.65380859375
],
[
0.5185546875,
0.646484375,
0.91259765625,
0.646484375,
0.91259765625,
0.66015625,
0.5185546875,
0.66015625
],
[
0.0882568359375,
0.65576171875,
0.4794921875,
0.65576171875,
0.4794921875,
0.66650390625,
0.0882568359375,
0.66650390625
],
[
0.5185546875,
0.6591796875,
0.91259765625,
0.66015625,
0.91259765625,
0.673828125,
0.5185546875,
0.67236328125
],
[
0.08660888671875,
0.66748046875,
0.48193359375,
0.66748046875,
0.48193359375,
0.68115234375,
0.08660888671875,
0.68115234375
],
[
0.5185546875,
0.673828125,
0.9111328125,
0.673828125,
0.9111328125,
0.68798828125,
0.5185546875,
0.68798828125
],
[
0.08660888671875,
0.68115234375,
0.48193359375,
0.68115234375,
0.48193359375,
0.6943359375,
0.08660888671875,
0.6943359375
],
[
0.5185546875,
0.68798828125,
0.865234375,
0.68798828125,
0.865234375,
0.69873046875,
0.5185546875,
0.69873046875
],
[
0.0882568359375,
0.6962890625,
0.4794921875,
0.6962890625,
0.4794921875,
0.70703125,
0.0882568359375,
0.70703125
],
[
0.08660888671875,
0.70849609375,
0.4794921875,
0.70849609375,
0.4794921875,
0.7216796875,
0.08660888671875,
0.7216796875
],
[
0.62109375,
0.7197265625,
0.66015625,
0.7197265625,
0.66015625,
0.73095703125,
0.62109375,
0.73095703125
],
[
0.65673828125,
0.7197265625,
0.67626953125,
0.7138671875,
0.68310546875,
0.72900390625,
0.6640625,
0.734375
],
[
0.68896484375,
0.72607421875,
0.84228515625,
0.72607421875,
0.84228515625,
0.7392578125,
0.68896484375,
0.7392578125
],
[
0.8935546875,
0.72607421875,
0.91259765625,
0.72607421875,
0.91259765625,
0.73779296875,
0.8935546875,
0.73779296875
],
[
0.59228515625,
0.7265625,
0.62744140625,
0.7265625,
0.62744140625,
0.73779296875,
0.59228515625,
0.73779296875
],
[
0.0882568359375,
0.7236328125,
0.1168212890625,
0.7236328125,
0.1168212890625,
0.73486328125,
0.0882568359375,
0.73486328125
],
[
0.62158203125,
0.72998046875,
0.66845703125,
0.734375,
0.6650390625,
0.74755859375,
0.619140625,
0.74365234375
],
[
0.66845703125,
0.73193359375,
0.6748046875,
0.7373046875,
0.6689453125,
0.7421875,
0.66259765625,
0.7373046875
],
[
0.68310546875,
0.72998046875,
0.689453125,
0.72998046875,
0.689453125,
0.734375,
0.68310546875,
0.734375
],
[
0.10455322265625,
0.73779296875,
0.4794921875,
0.73779296875,
0.4794921875,
0.74853515625,
0.10455322265625,
0.74853515625
],
[
0.64794921875,
0.74072265625,
0.69287109375,
0.74072265625,
0.69287109375,
0.751953125,
0.64794921875,
0.751953125
],
[
0.08660888671875,
0.75048828125,
0.48046875,
0.75048828125,
0.48046875,
0.7646484375,
0.08660888671875,
0.7646484375
],
[
0.51953125,
0.755859375,
0.9111328125,
0.755859375,
0.9111328125,
0.7666015625,
0.51953125,
0.7666015625
],
[
0.08660888671875,
0.7646484375,
0.4794921875,
0.7646484375,
0.4794921875,
0.77783203125,
0.08660888671875,
0.77783203125
],
[
0.5185546875,
0.76953125,
0.59228515625,
0.76953125,
0.59228515625,
0.77978515625,
0.5185546875,
0.77978515625
],
[
0.08905029296875,
0.77978515625,
0.4794921875,
0.77978515625,
0.4794921875,
0.79052734375,
0.08905029296875,
0.79052734375
],
[
0.53515625,
0.7822265625,
0.91259765625,
0.7822265625,
0.91259765625,
0.79541015625,
0.53515625,
0.79541015625
],
[
0.08660888671875,
0.79248046875,
0.4248046875,
0.79248046875,
0.4248046875,
0.80615234375,
0.08660888671875,
0.80615234375
],
[
0.5185546875,
0.79443359375,
0.91259765625,
0.796875,
0.91259765625,
0.81005859375,
0.5185546875,
0.80810546875
],
[
0.5185546875,
0.81005859375,
0.91259765625,
0.81005859375,
0.91259765625,
0.82470703125,
0.5185546875,
0.82470703125
],
[
0.08660888671875,
0.82470703125,
0.1168212890625,
0.82470703125,
0.1168212890625,
0.8359375,
0.08660888671875,
0.8359375
],
[
0.1168212890625,
0.82470703125,
0.25732421875,
0.82470703125,
0.25732421875,
0.83544921875,
0.1168212890625,
0.83544921875
],
[
0.5185546875,
0.82373046875,
0.9140625,
0.82373046875,
0.9140625,
0.8369140625,
0.5185546875,
0.8369140625
],
[
0.08660888671875,
0.84130859375,
0.48046875,
0.84130859375,
0.48046875,
0.85498046875,
0.08660888671875,
0.85498046875
],
[
0.5185546875,
0.83837890625,
0.9140625,
0.83837890625,
0.9140625,
0.8515625,
0.5185546875,
0.8515625
],
[
0.08660888671875,
0.85498046875,
0.48046875,
0.85498046875,
0.48046875,
0.86865234375,
0.08660888671875,
0.86865234375
],
[
0.515625,
0.85107421875,
0.701171875,
0.8486328125,
0.701171875,
0.86279296875,
0.515625,
0.86572265625
],
[
0.0882568359375,
0.87060546875,
0.4794921875,
0.87060546875,
0.4794921875,
0.880859375,
0.0882568359375,
0.880859375
],
[
0.5810546875,
0.8818359375,
0.8486328125,
0.8818359375,
0.8486328125,
0.896484375,
0.5810546875,
0.896484375
],
[
0.8935546875,
0.8837890625,
0.91259765625,
0.8837890625,
0.91259765625,
0.896484375,
0.8935546875,
0.896484375
],
[
0.0882568359375,
0.88525390625,
0.40185546875,
0.88525390625,
0.40185546875,
0.89599609375,
0.0882568359375,
0.89599609375
]
] |
[
"KDD*19,August 4-8,2019,Anchorage,AK,USA",
"Wen Chen and Pipei Huang, et al.",
"Table 2: Statistics of the datasets.",
"作为第一步，我们使用多模式嵌人模型表示所有项目。然后",
"我们详细介绍了FOM和POG。最后，我们在本节中介绍我们",
"Dataset",
"#Users",
"# Items",
"的Dida平台，这有助于确保这项工作在大规模在线应用中的",
"效率和质量要求",
"#Outfits",
"Outfit data",
"1,013,136",
"583,464",
"Item data",
"-4,747,039",
"4.1Multi-modalEmbedding",
"User data",
"127,1693,569,1124,463,302",
"对于每个时尚商品，我们计算嵌入于的非线性特征。时尚的",
"力限制在推荐单个项目上。。有几种方法可以推荐整个时尚服",
"概念主要依赖于视觉和文本信息。大多数以前的作品建议利用",
"装。刘思等人。[15]提出一种基于属性和类别的面向场合的服",
"图像和文本来学习多模式嵌人[3,11]。在我们的工作中，我们",
"装推荐方法。[17]和[3]中的工作需要图像或文本查询才能找",
"使用多模式嵌人模型，该模型为每个项目生成以下输人：（1）",
"到互补的衣服。使用功能张量分解方法向[4]中的用户建议项",
"密集向量编码来自CNN模型的项目的白色背景图片，（2）密",
"集。由于这些方法都需要用户查询或用户上传的数据作为输",
"集向量编码从TextCNN网络获得的项目的标题，该网络已经",
"入，因此典型用户可能会认为它们对用户不那么友好。此外",
"过预先训练，可以根据其标题预测项目的叶子类别，（3）密集",
"我们注意到，这些方法在大规模在线推荐系统中以有效的方",
"向量编码基于Alibaba专有的MAXXXTHX的项目嵌人5",
"式实现是相当不切实际的",
"我们的目标是获得一个嵌人空间，其中相似的项目嵌入在",
"附近，不同的项目位于不同的区域。我们连接从图像，文本和",
"Self-Attention and Transformer",
"CF派生的嵌入作为最终完全连接层的输人。输出为de维向量",
"2.3",
"f。对于每个项目f：我们将正样本f+定义为与f属于同",
"自我注意是一种将单个序列[29]的不同位置相关联的注意机",
"",
"叶子类别的项目，因此负样本f-是不属于同一叶子类别的项",
"制，已成功用于各种任务「13，19]。Transformer[29]是一种依",
"目。然后使用三重态损失通过以下方式训练整个网络：",
"靠自我关注的转换模型，已被广泛使用，并大大提高了语言",
"处理任务的性能E[1,21]。在最近的语言表示研究中，与从左",
"Le= Emax(d(f. ft) - d(f.f-)+α, 0)",
"到右的Transformer解码器（OpenAIGPT[21]）和双向LSTM",
"(1)",
"f",
"（ELMo[2o]）相比，双向Transformer编码器表现出了最佳性",
"能（BERT[1]）。最近还引人了多头自我关注,对用户的行为序",
"",
"其中距离度量d表示欧几里得距离，α表示余量。通过最小",
"列进行建模，进行序贯推荐[9,25]。与他们不同的是，在本文",
"化LE，同一类别中的f和f之间的距离被迫小于与不同类",
"中，我们采用自我注意来建模时装一代中的兼容性",
"别中的f-之间的距离α",
"4.2FOM:FashionOutfitModel",
"3DATASET",
"Taobao的时尚专家每天创造成千上万的服装。所有这些手工",
"Outft是一组项目，其中每个项目应与服装中的其他项目具有",
"不同的加权交互。为了捕获服装中的物品交互，我们设计了",
"制作的时装在网上展出之前都会经过审查。关于1。到目前为",
"个基于双向Transformer编码器架构的屏蔽物品预测任务。在",
"止，已经以这种方式创建和审查了4300万套服装。我们收集",
"服装中一次掩盖一个项目，我们要求模型根据上下文用正确",
"了80个最常见的叶子类别（例如。g。毛衣、外套、t恤、靴子",
"的项目填空。由于服装中的每个物品都被遮盖以融合其左右",
"和戒指)。不常见类别中的所有物品都将从服装中删除。在本",
"上下文，因此可以从自我注意机制中学习每个物品与服装中",
"文的其余部分中，我们仅考虑这80个叶子类别中的项目。之",
"所有其他物品之间的兼容性。因此，可以通过组合所有这些项",
"后，我们过滤包含少于4个项目的服装。总计1。剩下01万件",
"目的兼容性来学习服装中的兼容性",
"服装，由58.3万件单独物品组成。此外，我们收集了最近三个",
"令F表示所有服装的集合。给定装备FEF，F=(fi...ftfn)",
"月iFashi-on用户对项目和服装的点击。我们从3中选择点",
"其中f是第t项。令ft为从具有de维的多模态嵌人派生的",
"击行为。5700万活跃用户，总共观看了40多套服装。只有当",
"项ft的表示形式。我们为蒙版项目使用特定的嵌入[MASK]。",
"同一用户之前点击超过10个项目时，才会记录服装上的点击",
"非屏蔽项由其多模态嵌入表示。然后，我们将输人嵌人集表示",
"次数。然后，我们通过将服装点击与来自同一用户的最新50",
"为Fmask。给定Fmask，任务是预测屏蔽项，而不是重建整个",
"项点击配对来构建训练样本。最后，我们得到19。200万个训",
"装备。更正式地说，我们最小化FOM的以下损失函数",
"练样本，其中包括4个。4600万件商品和12.7万件服装。我们",
"数据集中的每个项目都与白色背景图像，标题和叶子类别相",
"倍1",
"P",
"logPr(fmask/Fmask;0F)",
"(2)",
"LP=",
"关联",
"-n",
"L",
"",
"据我们所知，与现有数据集相比，我们的数据集是最大的",
"mask=1",
"公开可用的时尚物品数据集，具有丰富的信息。此外，我们是",
"其中ep表示模型参数，Pr(）是在非屏蔽项的条件下选择正",
"第一个提供服装数据和相关用户行为数据的人，这些数据可",
"确项的概率",
"用于未来的时尚推荐研究。我们提供三个数据集，分别描述服",
"模型架构如图3所示。我们不像Transformer[29]那样使用",
"装、物品和用户行为。数据集的统计信息如表2所示",
"位置嵌人，因为我们将装备中的项目作为一组，而不是具有位",
"置信息的序列。让Fmask通过两个完全连接的层，其间有一个",
"4",
"METHODOLOGY",
"整流线性单元（ReLU）激活，以将所有输人嵌人从单个项目空",
"在本节中，我们将详细介绍我们的方法。POG在三个步骤的",
"间传输到装备空间。我们将两个完全连接的层称为过渡层，并",
"过程中构建：我们首先嵌入项目。其次，我们构建FOM，它学",
"将输出表示为HoeRnxdm：",
"习服装内物品的兼容性。最后，一日训练完成，我们使用生成",
"H° - ReLU([fT ....; fh i'w +bo)wf +bf",
"(3)",
"的预训练FOM在Transformer架构上初始化POG"
] |
{
"arxiv_id": "1905.01866",
"language": "zh",
"page_idx": 3
}
|
||
1905.01866-zh-4
|
1905.01866-zh-4.jpg
|
KDD'19, August 4-8, 2019 , Anchorage, AK, USA
POG:Personalized Outfit Generation for Fashion Recommendation at Alibaba iFashion
<FigureHere>
Figure 3: The architecture of FOM. We mask the items in the outfit one at a time. For example, we mask a pair of jeans in the
outfit. The model is learned to choose the correct jeans from a candidate pool, to complement other items in the outfit.
<FigureHere>
一组不同的样本来学习服装的兼容性信息。请注意，$h_{mask}$和
$h$不对应于原始项嵌人，而是对应于过渡层的输出
Figure 4:The architecture of POG, which is an encoder-decoder architecture with a Per network and a Gen network. The outfit item is generated step by step according to the user preference signal from the Per network and the compatibility signal from the Gen network.
$$\text{其中}W_0^F\:\in\:\mathbb{R}^{d_e\times d_m}\:,\:W_1^F\:\in\:\mathbb{R}^{d_m\times d_m}\:,\:b_0^F\:\in\:\mathbb{R}^{n\times d_m}\:\text{和}\:b_1^F\:\in $$
$\mathbb{R}^{n\times d_m\text{ 是可学习的参数}}$
以下 Transformer 编码器包含多个层。每个层包含一个多头自关注(MH)子层和一个位置前馈网络(PFFN)子层，其中在两个子层中的每个子层周围都采用剩余连接，然后是层归一化(LN)。MH 和 PFFN 的定义与论文[29] 相同。我们将$i$层中第一个子层的输出定义为$H_1^i$。因此，每个层 $H^i$ 可以迭代计算：
4.3 POG: Personalized Outfit Generation
# Model
(4) (5)
$$\begin{aligned}&H^{i}=\mathrm{Transformer}_{\mathrm{e}}(H^{i-1}),\forall i=1,\ldots,l\\&\mathrm{Transformer}_{\mathrm{e}}(H^{i-1})=\mathrm{LN}(H_{1}^{i-1}+\mathrm{PFFN}(H_{1}^{i-1}))\\&H_{1}^{i-1}=\mathrm{LN}(H^{i-1}+\mathrm{MH}(H^{i-1},H^{i-1},H^{i-1}))\end{aligned}$$
在$\iota$层之后，我们获得输出$G=H^l$。让$g_{mask}$表示输
在对服装兼容性进行建模之后，我们现在考虑一种生成模型，该模型通过引人用户偏好信号来生成个性化和兼容的服装。利用编码器-解码器结构的优势，我们的目标是将用户的历史行为转化为个性化的服装。令$u$ 表示所有用户的集合，$\mathcal{F}$表示所有装备的集合。我们使用一系列用户行为$U=\{u_{1},\ldots,u_{i},\ldots,u_{m}\}$ 来表征用户，其中$u_i$ 是用户点击的项目。$F=\{f_1,\ldots,f_t,\ldots,f_n\}$是来自同一用户的点击装备，其中$f_t$是装备中的项目。在每个时间步，我们预测下一个装备项目给定以前的装备项目和用户的点击顺序上的项目$U$。因此， 对于 $(U,F)$ 对，POG 的目标函数可以写为：
(6)
人[MASK]的相应输出，然后我们在$g_mask$的顶部附加一个
softmax 层来计算屏蔽项的概率：
$$Pr(f_{mask}|F_{mask};\Theta_{F})=\frac{\exp(g_{mask}h_{mask})}{\sum_{\boldsymbol{h}\in\mathcal{H}}\exp(g_{mask}\boldsymbol{h})}$$
(7)
$$\mathcal{L}_{(U,F)}=-\frac{1}{n}\sum_{t=1}^{n}\log Pr\big(f_{t+1}|f_{1},\ldots,f_{t},U;\Theta_{(U,F)}\big)$$
(8)
, 其中$h_mask$ 是屏蔽项的基本事实过渡嵌人，而$\mathcal{H}$包含$\mathcal{F}$ 中所有项的过渡嵌人。可以选择$\mathcal{H}$ 作为所有项目的整个项目转换嵌入集，但是，由于大量的高维嵌人，这是不实际的。因此， 我们通过从整个过渡集中随机抽取 3 个项目来获得$h$ ,这些项目未出现在装备$F$中，并与$h_{mask}$一起。这允许模型通过查看
其中$\Theta_{(U,F)}$表示模型参数。$Pr(\cdot)$是看到$f_t+1$以以前的装备项
目和用户单击的项目为条件的概率
模型架构如图 4 所示。在 POG 中，编码器将用户单击的项目作为输人。给定一个特殊的令牌 ,解码器然后一次
|
[
[
0.08660888671875,
0.0770263671875,
0.50634765625,
0.0770263671875,
0.50634765625,
0.08709716796875,
0.08660888671875,
0.08709716796875
],
[
0.6708984375,
0.07574462890625,
0.9111328125,
0.07574462890625,
0.9111328125,
0.0889892578125,
0.6708984375,
0.0889892578125
],
[
0.66015625,
0.12371826171875,
0.69287109375,
0.12371826171875,
0.69287109375,
0.1351318359375,
0.66015625,
0.1351318359375
],
[
0.5087890625,
0.149658203125,
0.5390625,
0.15087890625,
0.5390625,
0.1578369140625,
0.5087890625,
0.1572265625
],
[
0.65869140625,
0.147705078125,
0.69140625,
0.147705078125,
0.69140625,
0.160400390625,
0.65869140625,
0.160400390625
],
[
0.5849609375,
0.15087890625,
0.615234375,
0.15087890625,
0.615234375,
0.1578369140625,
0.5849609375,
0.1578369140625
],
[
0.455078125,
0.15087890625,
0.463134765625,
0.15087890625,
0.463134765625,
0.154052734375,
0.455078125,
0.154052734375
],
[
0.325927734375,
0.173583984375,
0.348876953125,
0.173583984375,
0.348876953125,
0.1817626953125,
0.325927734375,
0.1817626953125
],
[
0.5146484375,
0.173583984375,
0.5390625,
0.173583984375,
0.5390625,
0.18115234375,
0.5146484375,
0.18115234375
],
[
0.58837890625,
0.173583984375,
0.615234375,
0.173583984375,
0.615234375,
0.18115234375,
0.58837890625,
0.18115234375
],
[
0.6640625,
0.17236328125,
0.689453125,
0.17236328125,
0.689453125,
0.184326171875,
0.6640625,
0.184326171875
],
[
0.6650390625,
0.199462890625,
0.689453125,
0.199462890625,
0.689453125,
0.2083740234375,
0.6650390625,
0.2083740234375
],
[
0.455078125,
0.1749267578125,
0.463134765625,
0.1749267578125,
0.463134765625,
0.179931640625,
0.455078125,
0.179931640625
],
[
0.51953125,
0.2020263671875,
0.53759765625,
0.2020263671875,
0.53759765625,
0.205810546875,
0.51953125,
0.205810546875
],
[
0.6650390625,
0.2247314453125,
0.689453125,
0.2247314453125,
0.689453125,
0.231689453125,
0.6650390625,
0.231689453125
],
[
0.0849609375,
0.250732421875,
0.9140625,
0.250732421875,
0.9140625,
0.264404296875,
0.0849609375,
0.264404296875
],
[
0.0849609375,
0.264404296875,
0.87158203125,
0.264404296875,
0.87158203125,
0.27783203125,
0.0849609375,
0.27783203125
],
[
0.32177734375,
0.31640625,
0.361083984375,
0.31640625,
0.361083984375,
0.32763671875,
0.32177734375,
0.32763671875
],
[
0.70166015625,
0.31640625,
0.71337890625,
0.31640625,
0.71337890625,
0.32080078125,
0.70166015625,
0.32080078125
],
[
0.3505859375,
0.347900390625,
0.36279296875,
0.348388671875,
0.36279296875,
0.352783203125,
0.3505859375,
0.3515625
],
[
0.56396484375,
0.347900390625,
0.57421875,
0.347900390625,
0.57421875,
0.35107421875,
0.56396484375,
0.35107421875
],
[
0.5947265625,
0.349853515625,
0.615234375,
0.349853515625,
0.615234375,
0.35791015625,
0.5947265625,
0.35791015625
],
[
0.64892578125,
0.35107421875,
0.6640625,
0.35107421875,
0.6640625,
0.35791015625,
0.64892578125,
0.35791015625
],
[
0.69921875,
0.35107421875,
0.71826171875,
0.35107421875,
0.71826171875,
0.359130859375,
0.69921875,
0.359130859375
],
[
0.41015625,
0.35595703125,
0.435546875,
0.35595703125,
0.435546875,
0.363037109375,
0.41015625,
0.363037109375
],
[
0.361083984375,
0.35595703125,
0.381591796875,
0.35595703125,
0.381591796875,
0.363037109375,
0.361083984375,
0.363037109375
],
[
0.48779296875,
0.356689453125,
0.5048828125,
0.356689453125,
0.5048828125,
0.361083984375,
0.48779296875,
0.361083984375
],
[
0.2099609375,
0.363037109375,
0.256591796875,
0.363037109375,
0.256591796875,
0.37060546875,
0.2099609375,
0.37060546875
],
[
0.2059326171875,
0.375,
0.301513671875,
0.375,
0.301513671875,
0.385009765625,
0.2059326171875,
0.385009765625
],
[
0.35693359375,
0.380126953125,
0.386474609375,
0.3818359375,
0.3857421875,
0.39013671875,
0.35546875,
0.38818359375
],
[
0.58837890625,
0.380126953125,
0.625,
0.380126953125,
0.625,
0.391357421875,
0.58837890625,
0.391357421875
],
[
0.63623046875,
0.380126953125,
0.67333984375,
0.380126953125,
0.67333984375,
0.391357421875,
0.63623046875,
0.391357421875
],
[
0.689453125,
0.380126953125,
0.72802734375,
0.380126953125,
0.72802734375,
0.391357421875,
0.689453125,
0.391357421875
],
[
0.74267578125,
0.380126953125,
0.78173828125,
0.380126953125,
0.78173828125,
0.391357421875,
0.74267578125,
0.391357421875
],
[
0.496826171875,
0.38330078125,
0.51318359375,
0.38330078125,
0.51318359375,
0.39013671875,
0.496826171875,
0.39013671875
],
[
0.41259765625,
0.3818359375,
0.43701171875,
0.3818359375,
0.43701171875,
0.389404296875,
0.41259765625,
0.389404296875
],
[
0.404296875,
0.4072265625,
0.4306640625,
0.4091796875,
0.428955078125,
0.416748046875,
0.40283203125,
0.4140625
],
[
0.47802734375,
0.4072265625,
0.51318359375,
0.4091796875,
0.51318359375,
0.417236328125,
0.47802734375,
0.415283203125
],
[
0.5888671875,
0.4072265625,
0.6259765625,
0.4072265625,
0.6259765625,
0.417236328125,
0.5888671875,
0.417236328125
],
[
0.63134765625,
0.404052734375,
0.6787109375,
0.4072265625,
0.67724609375,
0.41845703125,
0.6298828125,
0.415283203125
],
[
0.689453125,
0.4072265625,
0.72802734375,
0.4072265625,
0.72802734375,
0.417236328125,
0.689453125,
0.417236328125
],
[
0.74609375,
0.4072265625,
0.78173828125,
0.4072265625,
0.78173828125,
0.417236328125,
0.74609375,
0.417236328125
],
[
0.2418212890625,
0.425537109375,
0.261474609375,
0.425537109375,
0.261474609375,
0.434326171875,
0.2418212890625,
0.434326171875
],
[
0.35546875,
0.4091796875,
0.3896484375,
0.4091796875,
0.3896484375,
0.416748046875,
0.35546875,
0.416748046875
],
[
0.486083984375,
0.43505859375,
0.51123046875,
0.43505859375,
0.51123046875,
0.443115234375,
0.486083984375,
0.443115234375
],
[
0.5947265625,
0.43505859375,
0.615234375,
0.43505859375,
0.615234375,
0.442626953125,
0.5947265625,
0.442626953125
],
[
0.646484375,
0.43505859375,
0.6640625,
0.43505859375,
0.6640625,
0.442626953125,
0.646484375,
0.442626953125
],
[
0.416748046875,
0.436279296875,
0.423095703125,
0.436279296875,
0.423095703125,
0.44140625,
0.416748046875,
0.44140625
],
[
0.36767578125,
0.436279296875,
0.374267578125,
0.436279296875,
0.374267578125,
0.44140625,
0.36767578125,
0.44140625
],
[
0.7587890625,
0.436279296875,
0.771484375,
0.436279296875,
0.771484375,
0.43994140625,
0.7587890625,
0.43994140625
],
[
0.58642578125,
0.455810546875,
0.619140625,
0.455810546875,
0.619140625,
0.46728515625,
0.58642578125,
0.46728515625
],
[
0.6435546875,
0.455078125,
0.6640625,
0.45703125,
0.66259765625,
0.46533203125,
0.64208984375,
0.46337890625
],
[
0.486083984375,
0.458251953125,
0.51123046875,
0.458251953125,
0.51123046875,
0.466552734375,
0.486083984375,
0.466552734375
],
[
0.36279296875,
0.458984375,
0.3798828125,
0.458984375,
0.3798828125,
0.46337890625,
0.36279296875,
0.46337890625
],
[
0.41259765625,
0.458984375,
0.428955078125,
0.458984375,
0.428955078125,
0.46337890625,
0.41259765625,
0.46337890625
],
[
0.69921875,
0.458984375,
0.71728515625,
0.458984375,
0.71728515625,
0.46337890625,
0.69921875,
0.46337890625
],
[
0.75341796875,
0.458984375,
0.771484375,
0.458984375,
0.771484375,
0.46337890625,
0.75341796875,
0.46337890625
],
[
0.48779296875,
0.4716796875,
0.5087890625,
0.4716796875,
0.5087890625,
0.474853515625,
0.48779296875,
0.474853515625
],
[
0.64453125,
0.48095703125,
0.654296875,
0.48095703125,
0.654296875,
0.484130859375,
0.64453125,
0.484130859375
],
[
0.2017822265625,
0.496826171875,
0.2900390625,
0.496826171875,
0.2900390625,
0.50390625,
0.2017822265625,
0.50390625
],
[
0.325927734375,
0.496826171875,
0.35302734375,
0.496826171875,
0.35302734375,
0.5048828125,
0.325927734375,
0.5048828125
],
[
0.5849609375,
0.49365234375,
0.62109375,
0.49365234375,
0.62109375,
0.505859375,
0.5849609375,
0.505859375
],
[
0.0849609375,
0.5146484375,
0.91259765625,
0.5146484375,
0.91259765625,
0.52783203125,
0.0849609375,
0.52783203125
],
[
0.0849609375,
0.52880859375,
0.9140625,
0.52880859375,
0.9140625,
0.54248046875,
0.0849609375,
0.54248046875
],
[
0.08416748046875,
0.54150390625,
0.19775390625,
0.54150390625,
0.19775390625,
0.5546875,
0.08416748046875,
0.5546875
],
[
0.0849609375,
0.564453125,
0.48193359375,
0.564453125,
0.48193359375,
0.58203125,
0.0849609375,
0.58203125
],
[
0.52783203125,
0.56640625,
0.9140625,
0.5673828125,
0.9140625,
0.58203125,
0.52783203125,
0.58056640625
],
[
0.08416748046875,
0.58056640625,
0.236083984375,
0.58203125,
0.236083984375,
0.5966796875,
0.08416748046875,
0.59521484375
],
[
0.5146484375,
0.580078125,
0.8408203125,
0.58203125,
0.8408203125,
0.59521484375,
0.5146484375,
0.59326171875
],
[
0.10296630859375,
0.59765625,
0.48046875,
0.59765625,
0.48046875,
0.611328125,
0.10296630859375,
0.611328125
],
[
0.51708984375,
0.611328125,
0.5537109375,
0.611328125,
0.5537109375,
0.62255859375,
0.51708984375,
0.62255859375
],
[
0.560546875,
0.60986328125,
0.86279296875,
0.60986328125,
0.86279296875,
0.62353515625,
0.560546875,
0.62353515625
],
[
0.08660888671875,
0.62548828125,
0.47216796875,
0.62548828125,
0.47216796875,
0.6396484375,
0.08660888671875,
0.6396484375
],
[
0.08905029296875,
0.61279296875,
0.4794921875,
0.61279296875,
0.4794921875,
0.62353515625,
0.08905029296875,
0.62353515625
],
[
0.55615234375,
0.62548828125,
0.61279296875,
0.62548828125,
0.61279296875,
0.64013671875,
0.55615234375,
0.64013671875
],
[
0.08660888671875,
0.6396484375,
0.48193359375,
0.6396484375,
0.48193359375,
0.65283203125,
0.08660888671875,
0.65283203125
],
[
0.5185546875,
0.64453125,
0.91259765625,
0.64453125,
0.91259765625,
0.65771484375,
0.5185546875,
0.65771484375
],
[
0.09722900390625,
0.65380859375,
0.48046875,
0.65380859375,
0.48046875,
0.66748046875,
0.09722900390625,
0.66748046875
],
[
0.521484375,
0.66015625,
0.9111328125,
0.66015625,
0.9111328125,
0.67041015625,
0.521484375,
0.67041015625
],
[
0.5185546875,
0.67236328125,
0.91259765625,
0.67236328125,
0.91259765625,
0.68603515625,
0.5185546875,
0.68603515625
],
[
0.13232421875,
0.681640625,
0.381591796875,
0.68310546875,
0.381591796875,
0.69775390625,
0.13232421875,
0.6962890625
],
[
0.461669921875,
0.6845703125,
0.483642578125,
0.6845703125,
0.483642578125,
0.69775390625,
0.461669921875,
0.69775390625
],
[
0.51953125,
0.68603515625,
0.91259765625,
0.68603515625,
0.91259765625,
0.69970703125,
0.51953125,
0.69970703125
],
[
0.131591796875,
0.69970703125,
0.43310546875,
0.70068359375,
0.43310546875,
0.71826171875,
0.131591796875,
0.71728515625
],
[
0.463134765625,
0.7021484375,
0.483642578125,
0.705078125,
0.4794921875,
0.71826171875,
0.459228515625,
0.71533203125
],
[
0.521484375,
0.7021484375,
0.9111328125,
0.7021484375,
0.9111328125,
0.7119140625,
0.521484375,
0.7119140625
],
[
0.51708984375,
0.71337890625,
0.91259765625,
0.71337890625,
0.91259765625,
0.7265625,
0.51708984375,
0.7265625
],
[
0.1356201171875,
0.7216796875,
0.38818359375,
0.7216796875,
0.38818359375,
0.736328125,
0.1356201171875,
0.736328125
],
[
0.461669921875,
0.7236328125,
0.48193359375,
0.7236328125,
0.48193359375,
0.736328125,
0.461669921875,
0.736328125
],
[
0.51953125,
0.72802734375,
0.91259765625,
0.72802734375,
0.91259765625,
0.7412109375,
0.51953125,
0.7412109375
],
[
0.101318359375,
0.7412109375,
0.48193359375,
0.7412109375,
0.48193359375,
0.7587890625,
0.101318359375,
0.7587890625
],
[
0.5185546875,
0.7412109375,
0.91259765625,
0.7412109375,
0.91259765625,
0.7548828125,
0.5185546875,
0.7548828125
],
[
0.0849609375,
0.755859375,
0.48046875,
0.7548828125,
0.48046875,
0.76953125,
0.0849609375,
0.77099609375
],
[
0.51953125,
0.7548828125,
0.91259765625,
0.7548828125,
0.91259765625,
0.76806640625,
0.51953125,
0.76806640625
],
[
0.0849609375,
0.76953125,
0.295654296875,
0.76953125,
0.295654296875,
0.78271484375,
0.0849609375,
0.78271484375
],
[
0.51953125,
0.77099609375,
0.7900390625,
0.77099609375,
0.7900390625,
0.78076171875,
0.51953125,
0.78076171875
],
[
0.29736328125,
0.796875,
0.410888671875,
0.796875,
0.410888671875,
0.81005859375,
0.29736328125,
0.81005859375
],
[
0.1478271484375,
0.80322265625,
0.2900390625,
0.8046875,
0.2900390625,
0.8193359375,
0.1478271484375,
0.8173828125
],
[
0.461669921875,
0.8046875,
0.483642578125,
0.8046875,
0.483642578125,
0.8173828125,
0.461669921875,
0.8173828125
],
[
0.6337890625,
0.8017578125,
0.66845703125,
0.79736328125,
0.6708984375,
0.81103515625,
0.63623046875,
0.81494140625
],
[
0.56884765625,
0.8095703125,
0.625,
0.8125,
0.62353515625,
0.8271484375,
0.56787109375,
0.82373046875
],
[
0.65625,
0.80810546875,
0.86279296875,
0.8095703125,
0.86279296875,
0.82373046875,
0.65625,
0.82275390625
],
[
0.8935546875,
0.81005859375,
0.9140625,
0.81005859375,
0.9140625,
0.82275390625,
0.8935546875,
0.82275390625
],
[
0.63818359375,
0.81640625,
0.67236328125,
0.82470703125,
0.66650390625,
0.8359375,
0.63232421875,
0.82763671875
],
[
0.09234619140625,
0.8271484375,
0.48046875,
0.82568359375,
0.48046875,
0.84033203125,
0.09234619140625,
0.84130859375
],
[
0.294189453125,
0.8125,
0.4150390625,
0.8125,
0.4150390625,
0.8271484375,
0.294189453125,
0.8271484375
],
[
0.0882568359375,
0.84130859375,
0.48046875,
0.84130859375,
0.48046875,
0.85498046875,
0.0882568359375,
0.85498046875
],
[
0.5185546875,
0.84033203125,
0.9140625,
0.84033203125,
0.9140625,
0.85498046875,
0.5185546875,
0.85498046875
],
[
0.08660888671875,
0.85498046875,
0.4794921875,
0.85498046875,
0.4794921875,
0.86865234375,
0.08660888671875,
0.86865234375
],
[
0.51953125,
0.85498046875,
0.740234375,
0.85498046875,
0.740234375,
0.86865234375,
0.51953125,
0.86865234375
],
[
0.08660888671875,
0.86865234375,
0.48046875,
0.86865234375,
0.48046875,
0.8818359375,
0.08660888671875,
0.8818359375
],
[
0.53369140625,
0.86865234375,
0.9140625,
0.86865234375,
0.9140625,
0.8818359375,
0.53369140625,
0.8818359375
],
[
0.0882568359375,
0.88330078125,
0.4794921875,
0.88330078125,
0.4794921875,
0.89794921875,
0.0882568359375,
0.89794921875
],
[
0.51953125,
0.88330078125,
0.91259765625,
0.88330078125,
0.91259765625,
0.896484375,
0.51953125,
0.896484375
]
] |
[
"POG:PersonalizedOutfit.GenerationforFashionRecommendation atAlibabaiFashion",
"KDD '19, August 4-8, 2019, Anchorage, AK, USA",
"91",
"Trm",
"g2",
"Trm",
"a",
"MAS",
"Trm",
"Trm",
"g3",
"94",
"3",
"Trm",
"an",
"Figure 3: The architecture of FOM. We mask the items in the outfit one at a time. For example, we mask a pair of jeans in the",
"outfit. The model is learned to choose the correct jeans from a candidate pool, to complement other items in the outfit.",
"Output",
"",
"Pe",
"",
"V1",
"V2",
"v3",
"C2",
"C1",
"cm",
"",
"personalizedoutfit",
"Trm",
"Trm",
"Trm",
"Trm",
"Trm",
"m",
"Trm",
"Trm",
"Trm",
"Trm",
"Trm",
"Trm",
"Trm",
"POG",
"Trm",
"bm",
"h1",
"h2",
"b",
"h",
"an",
"f1",
"f2",
"um",
"11",
"2",
"",
"fn",
"",
"301",
"userclickeditems",
"Input",
"START",
"Figure 4: The architecture of POG, which is an encoder-decoder architecture with a Per network and a Gen network. The outfit",
"item is generated step by step according to the user preference signal from the Per network and the compatibility signal from",
"the Gen network.",
"其中WFERdexdm、WEERdmxdm、bF ERnxdm和bFe",
"一组不同的样本来学习服装的兼容性信息。请注意，hmask和",
"Rnxdm是可学习的参数",
"h不对应于原始项嵌入，而是对应于过渡层的输出",
"以下Transformer编码器包含多个层。每个层包含一个多头",
"4.3",
"POG:Personalized Outfit Generation",
"两个子层中的每个子层周围都采用剩余连接，然后是层归一",
"自关注（MH）子层和一个位置前馈网络（PFFN）子层，其中在",
"Model",
"化（LN)。MH和PFFN的定义与论文[29]相同。我们将i层中第",
"在对服装兼容性进行建模之后，我们现在考虑一种生成模",
"个子层的输出定义为H。因此，每个层H可以迭代计算：",
"型，该模型通过引入用户偏好信号来生成个性化和兼容的",
"服装。利用编码器-解码器结构的优势，我们的目标是将用",
"H' - Transformere(Hi-l), Vi = 1,...,",
"(4)",
"户的历史行为转化为个性化的服装。令U表示所有用户的",
"Transformere(Hi-1) =LN(Hi-1 + PFFN(Hi-1)",
"(5)",
"集合，乐表示所有装备的集合。我们使用一系列用户行为",
"U=（ui....ui.....um）来表征用户，其中ui是用户点击的项",
"Hi-1 = LN(Hi-1 +MH(Hi-1, Hi-1, Hi-1))",
"(6)",
"目。F=fi,,ft..fn）是来自同一用户的点击装备，其中",
"在I层之后，我们获得输出G=H。让9mask表示输",
"f是装备中的项目。在每个时间步，我们预测下一个装备项",
"入[MASK]的相应输出，然后我们在9mask的顶部附加一个",
"目给定以前的装备项目和用户的点击顺序上的项目U。因此，",
"softmax层来计算屏蔽项的概率：",
"对于（U,F)对，POG的目标函数可以写为：",
"exp(gmaskhmask)",
"Pr(fmask|Fmask:OF)=",
"(7)",
"1n",
"L(U,F) =",
"logPr(ft+ilfi,.*fr,U;e(u.F))",
"(8)",
"nE",
"其中hmask是屏蔽项的基本事实过渡嵌入，而H包含F中",
"ZheH exp(gmaskh)",
"所有项的过渡嵌人。可以选择H作为所有项目的整个项目转",
"其中①(U,F）表示模型参数。Pr()是看到fi+1以以前的装备项",
"换嵌入集，但是，由于大量的高维嵌入，这是不实际的。因此，",
"目和用户单击的项目为条件的概率",
"我们通过从整个过渡集中随机抽取3个项目来获得h，这些项",
"模型架构如图4所示。在POG中，编码器将用户单击的项",
"目未出现在装备F中，并与hmask一起。这允许模型通过查看",
"目作为输人。给定一个特殊的令牌[START]，解码器然后一次"
] |
{
"arxiv_id": "1905.01866",
"language": "zh",
"page_idx": 4
}
|
||
1801.10496-zh-3
|
1801.10496-zh-3.jpg
|
DRAFT
cluster these results to obtain a target template pool 适应特定的要求。(3) 不可靠跟踪结果的校正可以通过各种
for subsequent verification.
在 PTAV 中，$\mathcal{T}$和$\mathcal{V}$在具有必要交互的两个不同线程上并行运行，如图$\color{red}{\mathsf{E}}$3 所示。跟踪器$\mathcal{T}$和验证器$\mathcal{V}$在第一帧中初始化。之后，$\mathcal{T}$开始处理每个到达的帧并生成结果(图 3因为跟踪比验证快得多，所以$\tau$和$\nu$异步工作。这种机制允许 PTAV 容忍临时跟踪漂移 (例如。g。,在图 3 中的第 380帧处),稍后将由$\mathcal{V}$进行校正。当$\mathcal{V}$发现跟踪结果不可靠时， 它会从本地区域搜索正确答案并将其发送到$\mathcal{T}$。收到此类反馈后，$\mathcal{T}$停止当前跟踪作业并回溯以使用 $\mathcal{V}$ 提供的校正恢复跟踪
<table>
<tbody>
<tr>
<th>$A$</th>
<th>1: $\operatorname{lgorithm}$ PTAV)</th>
<th>$\operatorname{Verifying}$ and</th>
</tr>
<tr>
<td>1</td>
<td>Initialize ;the</td>
<td>$T$ acker</td>
</tr>
<tr>
<td>2</td>
<td>;the Initialize</td>
<td>erifier $\nu$</td>
</tr>
<tr>
<td>3</td>
<td>Initialize Cur</td>
<td>nd frame; </td>
</tr>
<tr>
<td>4</td>
<td>Run $T$ (Alg.</td>
<td>he end of tracking</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td>$\Delta$ $\bot$</td>
<td>$\mathrm{Algorithm}^{\circ}$ 2. Tracking Thread 7</td>
</tr>
<tr>
<td>1</td>
<td>frame is valid do while current</td>
</tr>
<tr>
<td>2</td>
<td>if received from $\nu$ then</td>
</tr>
<tr>
<td>3</td>
<td>if verification passed then</td>
</tr>
<tr>
<td>4</td>
<td>Undate trs 10del 11 (optional):</td>
</tr>
<tr>
<td>5</td>
<td>else</td>
</tr>
<tr>
<td>6</td>
<td>Correct tracking;</td>
</tr>
<tr>
<td>7</td>
<td>Trace back and reset current frame ;</td>
</tr>
<tr>
<td>8</td>
<td>end</td>
</tr>
<tr>
<td>9</td>
<td>end</td>
</tr>
<tr>
<td>10 </td>
<td>Tracking 1the current frame ;</td>
</tr>
<tr>
<td>11</td>
<td>if time ,1 for ) $N_{\mathrm{int}}$ then</td>
</tr>
<tr>
<td>12</td>
<td>Send the current result to $,V_{1}$ to verify</td>
</tr>
<tr>
<td>13 $a_{1}=a_{1}$</td>
<td>end</td>
</tr>
<tr>
<td>14</td>
<td>current frame next frame;</td>
</tr>
<tr>
<td>5 1.5 </td>
<td>end</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<th>$\Delta$ $I$</th>
<th>Algorithm 3: Verifving ${\mathrm{g~Thread~}}{\mathcal{V}}$</th>
</tr>
<tr>
<td>1</td>
<td>while not t ended do</td>
</tr>
<tr>
<td>2</td>
<td>if received reque est. from ${\mathcal{T}}$ then</td>
</tr>
<tr>
<td>3</td>
<td>3 Verifving the tracking result: </td>
</tr>
<tr>
<td>4</td>
<td>4 Collect tracking result and perform $k$-means clustering if needed;</td>
</tr>
<tr>
<td>5</td>
<td>if verification failed then</td>
</tr>
<tr>
<td>6</td>
<td>6 Provide $\mathbf{m}_{\mathrm{for}}$ 1 in s $\frac{1}{1}$</td>
</tr>
<tr>
<td>7</td>
<td>Adjust $N_{\mathrm{int~if~needed}};$ 7</td>
</tr>
<tr>
<td>8</td>
<td>end </td>
</tr>
<tr>
<td>9</td>
<td>Send 1verif $t_{0}7$ 9</td>
</tr>
<tr>
<td>10</td>
<td>end 0</td>
</tr>
<tr>
<td>11 </td>
<td>end</td>
</tr>
</tbody>
</table>
值得注意的是，PTAV 提供了一个非常灵活的框架，一些重要的设计选择如下。(1) $\mathcal{T}$和$\mathcal{V}$的基本算法可能取决于特定的应用程序和可用的计算资源。此外，在实践中，可以使用多个验证器甚至基础跟踪器。(2) $\mathcal{T}$ 对 $\mathcal{V}$反馈的响应，无论是正面的还是负面的，都可以在很大程度上进行设计，以
4
方式实现，甚至可以纯粹通过$\mathcal{T}$ (i。e。,包括目标检测)。(4) $\mathcal{T}$有许多方法可以使用预先计算和存档的信息来加速。算法$\color{red}{1-3}$总结了一般的 PTAV 框架。
中的粉红色实心点)。同时，$\mathcal{V}$每隔几帧验证一次跟踪结果。3.2 PTAV Implementation
3.2.1 Tracking
我们为 PTAV 中的 $\mathcal{T}$ 选择 Staple 跟踪器$\mathbb{E}[?]$ 。Staple 的主要思想是结合两个互补的线索，即。e。,模板和直方图，用于跟踪。为此，在给定图像补丁 z 的情况下，提出了来自模板和直方图的跟踪分数的线性组合
$$y(\mathrm{z})=(1-\alpha)y_{\mathrm{tmpl}}(\mathrm{z})+\alpha y_{\mathrm{hist}}(\mathrm{z})$$
(1)
,其中$\alpha$表示权衡参数，$y_\mathrm{tmpl}($z) 和$y_\mathrm{hist}($z) 分别表示基于模
板和直方图信息的跟踪响应
通过学习最佳相关滤波器模型 w 得出模板上的跟踪响应， 该模型通过快速傅立叶变换(FFT)在频域中有效求解。在时间$t$ ,首先使用 w 计算滤波器响应的 FFT,然后进行逆FFT 以得出最终响应$y_\mathrm{tmpl}($z) 。模型 w 在每个帧中在线更新
直方图上的跟踪响应基于学习的颜色统计模型 h ,该模型在抵抗变形方面具有鲁棒性。在时间$t$ , h 用于计算$y_\mathrm{hist}($z) , 然后动态更新。为了使跟踪器适应尺度变化，采用尺度滤波器来估计目标尺度。有关 Staple 跟踪器的更多详细信息，请参见[?]
为了有效地利用 Staple 作为 PTAV 中的$\mathcal{T}$ ,除了原始的Staple 算法之外，$T$还存储所有中间结果(e。g。,w 和 h)在发出最后一次验证请求后的每帧。令$\mathcal{W} = \{ \mathrm{w} _{\xi - \triangle }, \cdots , \mathrm{w} _{\xi }\}$ 和$\mathcal{H} = \{ \mathrm{h} _{\xi - \triangle }, \cdots , \mathrm{h} _{\xi }\}$表示 w 和 h 的集合，其中$\xi$是$\mathcal{T}$ 处理的最后一帧的索引，$\triangle$表示用于存储跟踪模型的时间滑动窗口的固定大小。$\mathcal{W}$和$\mathcal{H}$中的这些中间结果允许$\mathcal{T}$快速回溯
特别是，当$\mathcal{V}$在帧$k$ 中检测到不可靠的跟踪结果，而$\mathcal{T}$ 开始在帧$j(j>k$ )上工作时，由正确的目标位置和帧索引信息组成的反馈将发送到$\mathcal{T}$。一旦从$\mathcal{V}$收到此反馈，$\mathcal{T}$首先停止处理帧$j$,然后利用存档的目标位置和跟踪模型(i。e。, $\mathrm{w}_{k-1}$和 h$_k-1$)从$\mathcal{W}$和$\mathcal{H}$检索，以从帧$k$恢复后续跟踪。同时，无用的中间结果在$\mathcal{W}($i。e。、w$_k$至 w$_j-1)$ 和$\mathcal{H}($i。e。, h$_k\text{到 h}_{j-1}$)将被丢弃
请注意，我们不假设上述策略中 w$_k-1$和 h$_k-1$的正确性。实际上，一种方法是从$k-1$向后跟踪，以找到可靠的帧来恢复跟踪，但要付出更多验证操作的额外代价。然而，在实践中，我们发现 w$_{k-1}$和 h$_{k-1}$通常为帧$k$提供足够的初始猜测，并依靠检测部分来纠正不正确的跟踪结果。有关验证和检测的更多详细信息，请参见以下部分
为了验证跟踪结果，$\mathcal{T}$每$N_\mathrm{int}$帧发送一次验证结果，其
中$N_\mathrm{int}$表示动态可调的验证间隔，如下所述
## 3.2.2 Verifying
验证的目的是测量给定样本与目标对象之间的相似性。受$\mathbb{E}$ [?]的启发，我们使用暹罗网络在 PTAV 中开发验证器$\mathcal{V}$(类似于$\mathbb{E}[?]$ ),如图 4 所示。Siamese 网络包含 CNN 的两个分支，并分别处理两个输人。在这项工作中，我们从 VGGNet$\boxed{\mathrm{E}}$ [?] 借用了 CNN 的体系结构 , 但增加了一个感兴趣区域 (RoI) 池化层$\mathbb{E}[?]$。这是因为，为了进行检测，$\mathcal{V}$需要处理图像中的多个区域，从中选择与目标对象最相似的候选对象作为最终结果。为了提高效率，RoI 池化用于同时处理一组区域
在 Siamese 网络中，两个 CNN 分支与单个对比损失层
$$\begin{aligned}\mathcal{L}(\mathrm{x}_i,\mathrm{x}_j,r_{ij})=\frac{1}{2}r_{ij}D^2+\frac{1}{2}(1-r_{ij})&\mathrm{max}(0,\varepsilon-D^2)\end{aligned}$$
|
[
[
0.0767822265625,
0.0328369140625,
0.125,
0.0328369140625,
0.125,
0.044189453125,
0.0767822265625,
0.044189453125
],
[
0.91259765625,
0.034088134765625,
0.92333984375,
0.034088134765625,
0.92333984375,
0.045440673828125,
0.91259765625,
0.045440673828125
],
[
0.12744140625,
0.054931640625,
0.496826171875,
0.054931640625,
0.496826171875,
0.06817626953125,
0.12744140625,
0.06817626953125
],
[
0.491943359375,
0.058074951171875,
0.5048828125,
0.059356689453125,
0.5048828125,
0.0631103515625,
0.491943359375,
0.0625
],
[
0.50634765625,
0.054931640625,
0.9208984375,
0.054931640625,
0.9208984375,
0.06817626953125,
0.50634765625,
0.06817626953125
],
[
0.125,
0.0694580078125,
0.3095703125,
0.0694580078125,
0.3095703125,
0.08331298828125,
0.125,
0.08331298828125
],
[
0.5087890625,
0.0694580078125,
0.9189453125,
0.0694580078125,
0.9189453125,
0.08331298828125,
0.5087890625,
0.08331298828125
],
[
0.50732421875,
0.08331298828125,
0.9189453125,
0.08331298828125,
0.9189453125,
0.09661865234375,
0.50732421875,
0.09661865234375
],
[
0.101318359375,
0.0921630859375,
0.490234375,
0.0921630859375,
0.490234375,
0.1029052734375,
0.101318359375,
0.1029052734375
],
[
0.50634765625,
0.0985107421875,
0.72607421875,
0.0985107421875,
0.72607421875,
0.11236572265625,
0.50634765625,
0.11236572265625
],
[
0.07843017578125,
0.10479736328125,
0.490234375,
0.10479736328125,
0.490234375,
0.11932373046875,
0.07843017578125,
0.11932373046875
],
[
0.080078125,
0.121826171875,
0.491943359375,
0.121826171875,
0.491943359375,
0.1319580078125,
0.080078125,
0.1319580078125
],
[
0.07843017578125,
0.1337890625,
0.494384765625,
0.1337890625,
0.494384765625,
0.147705078125,
0.07843017578125,
0.147705078125
],
[
0.5009765625,
0.1337890625,
0.705078125,
0.1337890625,
0.705078125,
0.147705078125,
0.5009765625,
0.147705078125
],
[
0.07843017578125,
0.149658203125,
0.491943359375,
0.149658203125,
0.491943359375,
0.1634521484375,
0.07843017578125,
0.1634521484375
],
[
0.5048828125,
0.15087890625,
0.62353515625,
0.154052734375,
0.62353515625,
0.1685791015625,
0.5048828125,
0.1654052734375
],
[
0.07843017578125,
0.1641845703125,
0.490234375,
0.1641845703125,
0.490234375,
0.1787109375,
0.07843017578125,
0.1787109375
],
[
0.50732421875,
0.1717529296875,
0.92236328125,
0.17236328125,
0.92236328125,
0.1875,
0.50732421875,
0.186279296875
],
[
0.07843017578125,
0.1787109375,
0.491943359375,
0.1787109375,
0.491943359375,
0.1925048828125,
0.07843017578125,
0.1925048828125
],
[
0.5087890625,
0.1875,
0.9208984375,
0.1875,
0.9208984375,
0.2008056640625,
0.5087890625,
0.2008056640625
],
[
0.080078125,
0.1944580078125,
0.488525390625,
0.1944580078125,
0.488525390625,
0.2052001953125,
0.080078125,
0.2052001953125
],
[
0.5087890625,
0.2020263671875,
0.92236328125,
0.2020263671875,
0.92236328125,
0.21533203125,
0.5087890625,
0.21533203125
],
[
0.0767822265625,
0.2083740234375,
0.490234375,
0.2083740234375,
0.490234375,
0.2215576171875,
0.0767822265625,
0.2215576171875
],
[
0.07598876953125,
0.2203369140625,
0.11273193359375,
0.222900390625,
0.111083984375,
0.2374267578125,
0.0743408203125,
0.23486328125
],
[
0.51123046875,
0.2183837890625,
0.72607421875,
0.2183837890625,
0.72607421875,
0.2291259765625,
0.51123046875,
0.2291259765625
],
[
0.59619140625,
0.2412109375,
0.83251953125,
0.240478515625,
0.83251953125,
0.255126953125,
0.59619140625,
0.255615234375
],
[
0.900390625,
0.2392578125,
0.92236328125,
0.2412109375,
0.9208984375,
0.255615234375,
0.89892578125,
0.25390625
],
[
0.0882568359375,
0.25244140625,
0.1854248046875,
0.25390625,
0.1854248046875,
0.26513671875,
0.0882568359375,
0.264404296875
],
[
0.205078125,
0.250732421875,
0.443603515625,
0.25244140625,
0.443603515625,
0.267578125,
0.205078125,
0.26513671875
],
[
0.0882568359375,
0.268310546875,
0.143798828125,
0.268310546875,
0.143798828125,
0.27978515625,
0.0882568359375,
0.27978515625
],
[
0.51953125,
0.26513671875,
0.92236328125,
0.26513671875,
0.92236328125,
0.27978515625,
0.51953125,
0.27978515625
],
[
0.50732421875,
0.279052734375,
0.6953125,
0.279052734375,
0.6953125,
0.292236328125,
0.50732421875,
0.292236328125
],
[
0.08251953125,
0.285400390625,
0.398681640625,
0.285888671875,
0.398681640625,
0.2998046875,
0.08251953125,
0.298583984375
],
[
0.53173828125,
0.294189453125,
0.9208984375,
0.294189453125,
0.9208984375,
0.30810546875,
0.53173828125,
0.30810546875
],
[
0.0849609375,
0.300537109375,
0.400390625,
0.300537109375,
0.400390625,
0.314453125,
0.0849609375,
0.314453125
],
[
0.08416748046875,
0.313232421875,
0.406005859375,
0.31494140625,
0.406005859375,
0.328857421875,
0.08416748046875,
0.326904296875
],
[
0.5087890625,
0.309326171875,
0.9208984375,
0.309326171875,
0.9208984375,
0.322509765625,
0.5087890625,
0.322509765625
],
[
0.08416748046875,
0.328857421875,
0.46728515625,
0.330810546875,
0.46728515625,
0.3447265625,
0.08416748046875,
0.34228515625
],
[
0.5087890625,
0.323974609375,
0.9208984375,
0.323974609375,
0.9208984375,
0.337158203125,
0.5087890625,
0.337158203125
],
[
0.50732421875,
0.33837890625,
0.9208984375,
0.33837890625,
0.9208984375,
0.352783203125,
0.50732421875,
0.352783203125
],
[
0.5361328125,
0.354736328125,
0.9189453125,
0.354736328125,
0.9189453125,
0.365478515625,
0.5361328125,
0.365478515625
],
[
0.5087890625,
0.36865234375,
0.92333984375,
0.36865234375,
0.92333984375,
0.3818359375,
0.5087890625,
0.3818359375
],
[
0.08660888671875,
0.3818359375,
0.314453125,
0.3818359375,
0.314453125,
0.395751953125,
0.08660888671875,
0.395751953125
],
[
0.50732421875,
0.3818359375,
0.9189453125,
0.3818359375,
0.9189453125,
0.395751953125,
0.50732421875,
0.395751953125
],
[
0.08251953125,
0.399658203125,
0.3203125,
0.399658203125,
0.3203125,
0.41357421875,
0.08251953125,
0.41357421875
],
[
0.5087890625,
0.395751953125,
0.92236328125,
0.396484375,
0.92236328125,
0.410888671875,
0.5087890625,
0.410400390625
],
[
0.5107421875,
0.4091796875,
0.564453125,
0.410888671875,
0.56396484375,
0.425537109375,
0.5087890625,
0.423583984375
],
[
0.0849609375,
0.41845703125,
0.09722900390625,
0.41845703125,
0.09722900390625,
0.4267578125,
0.0849609375,
0.4267578125
],
[
0.1217041015625,
0.415283203125,
0.35302734375,
0.415283203125,
0.35302734375,
0.4287109375,
0.1217041015625,
0.4287109375
],
[
0.1478271484375,
0.429931640625,
0.324462890625,
0.429931640625,
0.324462890625,
0.443115234375,
0.1478271484375,
0.443115234375
],
[
0.53173828125,
0.4267578125,
0.9208984375,
0.4267578125,
0.9208984375,
0.43994140625,
0.53173828125,
0.43994140625
],
[
0.50732421875,
0.44140625,
0.9189453125,
0.44140625,
0.9189453125,
0.455078125,
0.50732421875,
0.455078125
],
[
0.0849609375,
0.43115234375,
0.09722900390625,
0.43115234375,
0.09722900390625,
0.442626953125,
0.0849609375,
0.442626953125
],
[
0.08660888671875,
0.44580078125,
0.098876953125,
0.44580078125,
0.098876953125,
0.45703125,
0.08660888671875,
0.45703125
],
[
0.1690673828125,
0.443115234375,
0.404296875,
0.443115234375,
0.404296875,
0.458251953125,
0.1690673828125,
0.458251953125
],
[
0.1478271484375,
0.458984375,
0.1748046875,
0.458984375,
0.1748046875,
0.47021484375,
0.1478271484375,
0.47021484375
],
[
0.50732421875,
0.453857421875,
0.9208984375,
0.455810546875,
0.9208984375,
0.47021484375,
0.50732421875,
0.468505859375
],
[
0.0849609375,
0.460205078125,
0.09722900390625,
0.460205078125,
0.09722900390625,
0.4716796875,
0.0849609375,
0.4716796875
],
[
0.1707763671875,
0.4716796875,
0.289306640625,
0.472900390625,
0.289306640625,
0.4873046875,
0.1707763671875,
0.486083984375
],
[
0.5087890625,
0.47021484375,
0.9189453125,
0.47021484375,
0.9189453125,
0.48486328125,
0.5087890625,
0.48486328125
],
[
0.5087890625,
0.48486328125,
0.9208984375,
0.48486328125,
0.9208984375,
0.498779296875,
0.5087890625,
0.498779296875
],
[
0.08660888671875,
0.474853515625,
0.098876953125,
0.474853515625,
0.098876953125,
0.486083984375,
0.08660888671875,
0.486083984375
],
[
0.08660888671875,
0.4892578125,
0.098876953125,
0.4892578125,
0.098876953125,
0.5,
0.08660888671875,
0.5
],
[
0.17236328125,
0.4873046875,
0.42724609375,
0.4873046875,
0.42724609375,
0.50048828125,
0.17236328125,
0.50048828125
],
[
0.1478271484375,
0.50244140625,
0.176513671875,
0.50244140625,
0.176513671875,
0.51513671875,
0.1478271484375,
0.51513671875
],
[
0.50732421875,
0.498779296875,
0.92236328125,
0.5,
0.92236328125,
0.51318359375,
0.50732421875,
0.51220703125
],
[
0.08660888671875,
0.5048828125,
0.09722900390625,
0.5048828125,
0.09722900390625,
0.51513671875,
0.08660888671875,
0.51513671875
],
[
0.12335205078125,
0.517578125,
0.1519775390625,
0.517578125,
0.1519775390625,
0.52880859375,
0.12335205078125,
0.52880859375
],
[
0.5087890625,
0.5146484375,
0.54150390625,
0.5146484375,
0.54150390625,
0.5263671875,
0.5087890625,
0.5263671875
],
[
0.53125,
0.52880859375,
0.9208984375,
0.52880859375,
0.9208984375,
0.54248046875,
0.53125,
0.54248046875
],
[
0.0849609375,
0.51953125,
0.09722900390625,
0.51953125,
0.09722900390625,
0.52978515625,
0.0849609375,
0.52978515625
],
[
0.08087158203125,
0.5341796875,
0.096435546875,
0.5341796875,
0.096435546875,
0.54443359375,
0.08087158203125,
0.54443359375
],
[
0.1217041015625,
0.52978515625,
0.3447265625,
0.5322265625,
0.3447265625,
0.54541015625,
0.1217041015625,
0.54345703125
],
[
0.12335205078125,
0.546875,
0.4306640625,
0.546875,
0.4306640625,
0.56005859375,
0.12335205078125,
0.56005859375
],
[
0.5087890625,
0.54345703125,
0.9189453125,
0.54345703125,
0.9189453125,
0.556640625,
0.5087890625,
0.556640625
],
[
0.08087158203125,
0.5498046875,
0.09478759765625,
0.5498046875,
0.09478759765625,
0.5595703125,
0.08087158203125,
0.5595703125
],
[
0.1446533203125,
0.5595703125,
0.404296875,
0.56005859375,
0.404296875,
0.57470703125,
0.1446533203125,
0.57373046875
],
[
0.5107421875,
0.55810546875,
0.92236328125,
0.55810546875,
0.92236328125,
0.5712890625,
0.5107421875,
0.5712890625
],
[
0.50732421875,
0.5712890625,
0.9189453125,
0.5712890625,
0.9189453125,
0.58544921875,
0.50732421875,
0.58544921875
],
[
0.080078125,
0.56103515625,
0.09722900390625,
0.56103515625,
0.09722900390625,
0.57568359375,
0.080078125,
0.57568359375
],
[
0.08087158203125,
0.57763671875,
0.096435546875,
0.57763671875,
0.096435546875,
0.58837890625,
0.08087158203125,
0.58837890625
],
[
0.125,
0.57568359375,
0.1502685546875,
0.57568359375,
0.1502685546875,
0.58837890625,
0.125,
0.58837890625
],
[
0.12091064453125,
0.58837890625,
0.328369140625,
0.5888671875,
0.328369140625,
0.60400390625,
0.12091064453125,
0.60302734375
],
[
0.50732421875,
0.5869140625,
0.9208984375,
0.5859375,
0.9208984375,
0.60107421875,
0.50732421875,
0.6015625
],
[
0.5087890625,
0.6015625,
0.92333984375,
0.6015625,
0.92333984375,
0.6162109375,
0.5087890625,
0.6162109375
],
[
0.08087158203125,
0.59130859375,
0.096435546875,
0.59130859375,
0.096435546875,
0.6015625,
0.08087158203125,
0.6015625
],
[
0.080078125,
0.60400390625,
0.12744140625,
0.60400390625,
0.12744140625,
0.61865234375,
0.080078125,
0.61865234375
],
[
0.50732421875,
0.6162109375,
0.69140625,
0.6162109375,
0.69140625,
0.6298828125,
0.50732421875,
0.6298828125
],
[
0.53125,
0.6298828125,
0.9208984375,
0.630859375,
0.9208984375,
0.64599609375,
0.53125,
0.64453125
],
[
0.50732421875,
0.64453125,
0.9208984375,
0.64599609375,
0.9208984375,
0.6591796875,
0.50732421875,
0.65771484375
],
[
0.0882568359375,
0.65576171875,
0.31787109375,
0.65576171875,
0.31787109375,
0.66943359375,
0.0882568359375,
0.66943359375
],
[
0.5087890625,
0.66015625,
0.9208984375,
0.66015625,
0.9208984375,
0.673828125,
0.5087890625,
0.673828125
],
[
0.08416748046875,
0.6748046875,
0.2279052734375,
0.6748046875,
0.2279052734375,
0.68505859375,
0.08416748046875,
0.68505859375
],
[
0.5087890625,
0.6748046875,
0.9208984375,
0.6748046875,
0.9208984375,
0.68798828125,
0.5087890625,
0.68798828125
],
[
0.12335205078125,
0.689453125,
0.338134765625,
0.689453125,
0.338134765625,
0.70263671875,
0.12335205078125,
0.70263671875
],
[
0.5087890625,
0.689453125,
0.92236328125,
0.689453125,
0.92236328125,
0.70263671875,
0.5087890625,
0.70263671875
],
[
0.148681640625,
0.7041015625,
0.343017578125,
0.7041015625,
0.343017578125,
0.71728515625,
0.148681640625,
0.71728515625
],
[
0.08660888671875,
0.6923828125,
0.09722900390625,
0.6923828125,
0.09722900390625,
0.7021484375,
0.08660888671875,
0.7021484375
],
[
0.0849609375,
0.70556640625,
0.09722900390625,
0.70556640625,
0.09722900390625,
0.71728515625,
0.0849609375,
0.71728515625
],
[
0.5087890625,
0.705078125,
0.7900390625,
0.705078125,
0.7900390625,
0.71826171875,
0.5087890625,
0.71826171875
],
[
0.1478271484375,
0.71826171875,
0.455078125,
0.71826171875,
0.455078125,
0.73193359375,
0.1478271484375,
0.73193359375
],
[
0.53173828125,
0.7197265625,
0.92333984375,
0.7197265625,
0.92333984375,
0.73291015625,
0.53173828125,
0.73291015625
],
[
0.50732421875,
0.73193359375,
0.81396484375,
0.73291015625,
0.81396484375,
0.74755859375,
0.50732421875,
0.7470703125
],
[
0.08660888671875,
0.72021484375,
0.098876953125,
0.72021484375,
0.098876953125,
0.73193359375,
0.08660888671875,
0.73193359375
],
[
0.156005859375,
0.73291015625,
0.293212890625,
0.73291015625,
0.293212890625,
0.7470703125,
0.156005859375,
0.7470703125
],
[
0.08660888671875,
0.75,
0.09722900390625,
0.75,
0.09722900390625,
0.76123046875,
0.08660888671875,
0.76123046875
],
[
0.1478271484375,
0.74755859375,
0.316162109375,
0.74755859375,
0.316162109375,
0.76123046875,
0.1478271484375,
0.76123046875
],
[
0.08660888671875,
0.7646484375,
0.098876953125,
0.7646484375,
0.098876953125,
0.77587890625,
0.08660888671875,
0.77587890625
],
[
0.17236328125,
0.76220703125,
0.42236328125,
0.76220703125,
0.42236328125,
0.77587890625,
0.17236328125,
0.77587890625
],
[
0.50732421875,
0.7646484375,
0.6259765625,
0.7646484375,
0.6259765625,
0.77783203125,
0.50732421875,
0.77783203125
],
[
0.08660888671875,
0.77880859375,
0.098876953125,
0.77880859375,
0.098876953125,
0.79052734375,
0.08660888671875,
0.79052734375
],
[
0.1707763671875,
0.7763671875,
0.32275390625,
0.7763671875,
0.32275390625,
0.79052734375,
0.1707763671875,
0.79052734375
],
[
0.5087890625,
0.7822265625,
0.9208984375,
0.7822265625,
0.9208984375,
0.79541015625,
0.5087890625,
0.79541015625
],
[
0.0849609375,
0.79345703125,
0.098876953125,
0.79345703125,
0.098876953125,
0.8037109375,
0.0849609375,
0.8037109375
],
[
0.148681640625,
0.79248046875,
0.176513671875,
0.79248046875,
0.176513671875,
0.8037109375,
0.148681640625,
0.8037109375
],
[
0.5107421875,
0.796875,
0.9208984375,
0.796875,
0.9208984375,
0.81005859375,
0.5107421875,
0.81005859375
],
[
0.0849609375,
0.806640625,
0.09722900390625,
0.806640625,
0.09722900390625,
0.818359375,
0.0849609375,
0.818359375
],
[
0.1446533203125,
0.8037109375,
0.363525390625,
0.8046875,
0.363525390625,
0.8193359375,
0.1446533203125,
0.818359375
],
[
0.5107421875,
0.81103515625,
0.9208984375,
0.81103515625,
0.9208984375,
0.82470703125,
0.5107421875,
0.82470703125
],
[
0.080078125,
0.8212890625,
0.098876953125,
0.8212890625,
0.098876953125,
0.833984375,
0.080078125,
0.833984375
],
[
0.12335205078125,
0.82080078125,
0.1519775390625,
0.82080078125,
0.1519775390625,
0.83349609375,
0.12335205078125,
0.83349609375
],
[
0.5107421875,
0.82568359375,
0.9208984375,
0.82568359375,
0.9208984375,
0.8388671875,
0.5107421875,
0.8388671875
],
[
0.07843017578125,
0.8359375,
0.125,
0.833984375,
0.1258544921875,
0.8466796875,
0.080078125,
0.8486328125
],
[
0.5107421875,
0.84033203125,
0.92333984375,
0.84033203125,
0.92333984375,
0.85400390625,
0.5107421875,
0.85400390625
],
[
0.50732421875,
0.85400390625,
0.9208984375,
0.85498046875,
0.9208984375,
0.869140625,
0.50732421875,
0.86865234375
],
[
0.101318359375,
0.87060546875,
0.490234375,
0.87060546875,
0.490234375,
0.8837890625,
0.101318359375,
0.8837890625
],
[
0.50634765625,
0.869140625,
0.9208984375,
0.869140625,
0.9208984375,
0.88330078125,
0.50634765625,
0.88330078125
],
[
0.07843017578125,
0.88525390625,
0.490234375,
0.88525390625,
0.490234375,
0.8984375,
0.07843017578125,
0.8984375
],
[
0.50634765625,
0.8837890625,
0.89794921875,
0.8837890625,
0.89794921875,
0.89794921875,
0.50634765625,
0.89794921875
],
[
0.080078125,
0.9013671875,
0.488525390625,
0.9013671875,
0.488525390625,
0.912109375,
0.080078125,
0.912109375
],
[
0.53125,
0.8984375,
0.90283203125,
0.8984375,
0.90283203125,
0.912109375,
0.53125,
0.912109375
],
[
0.0767822265625,
0.9140625,
0.490234375,
0.91552734375,
0.490234375,
0.9287109375,
0.0767822265625,
0.92822265625
],
[
0.51953125,
0.921875,
0.92724609375,
0.921875,
0.92724609375,
0.95263671875,
0.51953125,
0.95263671875
],
[
0.63623046875,
0.91748046875,
0.650390625,
0.92236328125,
0.64453125,
0.93115234375,
0.63134765625,
0.92529296875
],
[
0.07843017578125,
0.9296875,
0.491943359375,
0.9296875,
0.491943359375,
0.943359375,
0.07843017578125,
0.943359375
],
[
0.6435546875,
0.93798828125,
0.65283203125,
0.93798828125,
0.65283203125,
0.9423828125,
0.6435546875,
0.9423828125
]
] |
[
"DRAFT",
"4",
"cluster these results to obtain a target template pool",
"",
"适应特定的要求。（3）不可靠跟踪结果的校正可以通过各种",
"for subsequent verification.",
"方式实现，甚至可以纯粹通过T（i。e。，包括目标检测)。（4)",
"丁有许多方法可以使用预先计算和存档的信息来加速。算法",
"在PTAV中，T和V在具有必要交互的两个不同线程上",
"1-3总结了一般的PTAV框架。",
"并行运行，如图3所示。跟踪器T和验证器V在第一顿中",
"初始化。之后，T开始处理每个到达的顿并生成结果（图3",
"中的粉红色实心点）。同时，√每隔几顿验证一次跟踪结果。",
"3.2PTAVImplementation",
"因为跟踪比验证快得多，所以T和V异步工作。这种机制允",
"3.2.1Tracking",
"许PTAV容忍临时跟踪漂移（例如。g。，在图3中的第380",
"我们为PTAV中的T选择Staple跟踪器E[?]。Staple的主",
"顿处），稍后将由V进行校正。当V发现跟踪结果不可靠时，",
"要思想是结合两个互补的线索，即。e。，模板和直方图，用于",
"它会从本地区域搜索正确答案并将其发送到丁。收到此类反",
"跟踪。为此，在给定图像补丁的情况下，提出了来自模板",
"馈后，丁停止当前跟踪作业并回溯以使用V提供的校正恢复",
"跟踪",
"和直方图的跟踪分数的线性组合",
"y(z) = (1 -Q)ytmpl(z) +Qyhist(z)",
"(1)",
"Algorithm 1:",
"Parallel Tracking and Verifying",
"(PTAV)",
"其中α表示权衡参数，Utmpl(z)和Yhist(z)分别表示基于模",
"板和直方图信息的跟踪响应",
"1 Initialize the tracking thread for tracker T ;",
"通过学习最佳相关滤波器模型W得出模板上的跟踪响应",
"2 Initialize the verifying thread for verifier V ;",
"3 Initialize current frame as the second frame;",
"该模型通过快速傅立叶变换（FFT）在频域中有效求解。在",
"4 Run T (Alg.2) and V (Alg.3) till the end of tracking;",
"时间t，首先使用w计算滤波器响应的FFT，然后进行逆",
"FFT以得出最终响应ytmpl(z）。模型w在每个顿中在线更新",
"直方图上的跟踪响应基于学习的颜色统计模型h，该模型",
"在抵抗变形方面具有鲁棒性。在时间t，h用于计算yhist(z），",
"Algorithm 2: Tracking Thread T",
"然后动态更新。为了使跟踪器适应尺度变化，采用尺度滤波",
"1 while current_frame is valid do",
"器来估计目标尺度。有关Staple跟踪器的更多详细信息，请",
"参见[?]",
"2",
"if received a message from V then",
"if verification passed then",
"为了有效地利用Staple作为PTAV中的T，除了原始的",
"Staple算法之外，T还存储所有中间结果（e。g。，w和h）在",
"3",
"4",
"Update tracking model (optional):",
"else",
"发出最后一次验证请求后的每顿。令W=【w-A，，we}",
"5",
"Correct tracking;",
"和H={he-△,.,he}表示w和h的集合，其中ε是T",
"处理的最后一顿的索引，表示用于存储跟踪模型的时间滑",
"6",
"7",
"Trace back and reset current_frame ;",
"end",
"动窗口的固定大小。W和H中的这些中间结果允许T快速",
"8",
"end",
"回溯",
"特别是，当V在顿中检测到不可靠的跟踪结果，而T",
"9",
"10",
"Tracking on the current_frame;",
"if time for verification according to Nint then",
"开始在顿i（j>k）上工作时，由正确的目标位置和顿索引",
"11",
"Send the current result to V to verify:",
"信息组成的反馈将发送到T。一旦从V收到此反馈，T首先",
"停止处理顿i，然后利用存档的目标位置和跟踪模型（i。e。",
"12",
"13",
"end",
"current_frame next frame;",
"Wk-1和hk-1）从W和H检索，以从顿k恢复后续跟踪。",
"同时，无用的中间结果在W(i。e。、wk至wj-1）和H(i。",
"14",
"15 end",
"e。，hk到hj-1）将被丢弃",
"请注意，我们不假设上述策略中wk-1和hk-1的正确性。",
"实际上，一种方法是从k－1向后跟踪，以找到可靠的顿来",
"Algorithm 3: Verifying Thread V",
"恢复跟踪，但要付出更多验证操作的额外代价。然而，在实",
"iwhile not ended dc",
"践中，我们发现wk-1和hk-1通常为顿k提供足够的初始",
"if received request from T then",
"猜测，并依靠检测部分来纠正不正确的跟踪结果。有关验证",
"Verifying the tracking result:",
"2",
"3",
"和检测的更多详细信息，请参见以下部分",
"Collect tracking result and perform k -means",
"为了验证跟踪结果，T每Nint顿发送一次验证结果，其",
"中Nint表示动态可调的验证间隔，如下所述",
"4",
"clustering if needed;",
"5",
"if verification failed then",
"6",
"Provide correction information in s ;",
"3.2.2Verifying",
"7.",
"Adjust Nint if needed;",
"验证的目的是测量给定样本与目标对象之间的相似性。受E",
"8",
"end",
"[?]的启发，我们使用暹罗网络在PTAV中开发验证器V（类",
"9",
"Send verification result s to T ;",
"似于E[?]），如图4所示。Siamese网络包含CNN的两个分",
"10",
"end",
"支，并分别处理两个输人。在这项工作中，我们从VGGNetF",
"11end",
"[?]借用了CNN的体系结构，但增加了一个感兴趣区域（RoI)",
"池化层目[?]。这是因为，为了进行检测，√需要处理图像中",
"值得注意的是，PTAV提供了二个非常灵活的框架，一些",
"的多个区域，从中选择与目标对象最相似的候选对象作为最",
"重要的设计选择如下。（1）T和V的基本算法可能取决于特",
"终结果。为了提高效率，RoI池化用于同时处理一组区域",
"定的应用程序和可用的计算资源。此外，在实践中，可以使",
"在Siamese网络中，两个CNN分支与单个对比损失层",
"用多个验证器甚至基础跟踪器。（2）T对V反馈的响应，无",
"C(x,x,s)=rD+(1s)max(0,eD) (2)",
"1",
"论是正面的还是负面的，都可以在很大程度上进行设计，以",
""
] |
{
"arxiv_id": "1801.10496",
"language": "zh",
"page_idx": 3
}
|
||
2306.02349-zh-1
|
2306.02349-zh-1.jpg
|
# $bgGLUE{:}$ A Bulgarian General Language Understanding $\textbf{Evaluation Benchmark}$
$\mathbf{Momchil~Hardalov}^{*}$
$\textbf{Todor Mihaylov}$
Pepa Atanasova
AWS AI Labs
Meta AI
University of Copenhagen, DIKU
$\mathbf{Kiril}$ Simov
Petya Osenova
Galia Angelova
IICT, Bulgarian Academy of Sciences
$\mathbf{Ves}$ Stoyanov
$\textbf{Ivan Koychev}$
Preslav Nakov
Dragomir Radev
Sofia University
$MBZUAI$
Meta AI
$\bar{\text{Yale University}}$
$\mathbf{Abstract}$
We present bgGLUE (Bulgarian General Language Understanding Evaluation), a benchmark for evaluating language models on Natural Language Understanding (NLU) tasks in Bulgarian.
Our benchmark includes
NLU tasks targeting a variety of NLP problems (e.g., natural language inference, factchecking, named entity recognition, sentiment analysis, question answering, etc.) and machine learning tasks (sequence labeling, document-level classification, and regression). We run the first systematic evaluation of pretrained language models for Bulgarian, comparing and contrasting results across the nine tasks in the benchmark. The evaluation results show strong performance on sequence labeling tasks, but there is a lot of room for improvement for tasks that require more complex reasoning. We make bgGLUE publicly available together with the fine-tuning and the evaluation code, as well as a public leaderboard at https://bgglue.github.io, and we hope that it will enable further advancements in developing NLU models for Bulgarian.
2020)或多种语言(Liang et al.,2020; Hu et al.,
2020) 上测试 NLU 任务
在这里，我们旨在通过提出 bgGLUE 来改善 NLU 基准中表示的语言的多样性，bgGLUE 是保加利亚语的基准，由九个 NLU 任务组成， 包括令牌分类，回归和分类。到目前为止，只有保加利亚语的单个数据集被用于模型开发和评估。一些多语言基准测试中还包含了保加利亚语中最多三个下游任务的小子集 (Liang et al., 2020; Hu et al., 2020)。此外，还有针对其他 Balto-Slavic 语言的现有基准测试，例如俄语 SuperGLUEE(Shavrina et al.,2020) 和斯洛文尼亚语 SuperGLUE$\overline {\mathrm{F} }( \text{\v{Z} agar and Robnik- })$ Šikonja, 2022)。但是，没有为东南斯拉夫语分组的代表，特别是保加利亚语的代表制定全面的基准。我们的目标是解决这些限制与bgGLUE
bgGLUE 统一并促进了对保加利亚语现有数据集和任务的访问。通过包括更具挑战性的任务，如自然语言推理，事实检查和问题回答，我们确保它包括为保加利亚语开发的 NLP 模型的严格测试集。我们还通过 HuggingFaceHub (Lhoest et al., 2021)提供对基准的访问，以便于使用，我们鼓励保加利亚语的模型共享。此外，我们进行了微调，并对保加利亚语的现有语言模型进行了首次系统评估，比较和对比了基准测试中所有任务的结果。我们的评估结果表明，更大，更健壮的预训练模型在所有任务上都能产生更好的性能，而且高效的蒸馏模型是大型模型的强大竞争对手
这些模型在词性标记和命名实体识别方面表现出很强的性能，但在需要更复杂推理的任务上却很挣扎，例如解决入学考试或评估新闻文章的可信度和准确性。我们的贡献如下：
# 1 Introduction
自然语言理解(NLU)基准测试，如
GLUE
(Wang et al., 2019b) 和 Super-
GLUE (Wang et al., 2019a) , 旨在对各种自然语言理解(NLU)任务的语言模型进行严格评估。此类基准的广泛采用推动了模型的快速开发，这些模型在这些基准的一部分任务上表现良好，但也超出了$\operatorname{E}($Devlin et al.,2019; Liu et al.,2019)。但是，直到最近，此类基准测试的重点一直放在英语上，对其他语言的兴趣不大$\overline{\mathrm{E}}$(Bender, 2011; Ponti et al., 2019)
为了解决这个问题，最近的工作设计了基准测试模型，用于在非英语语言(Le et al.,2020; Rodriguez-Penagos et al., 2021; Shavrina et al.,
$^{*}$Work done while Momchil was in the Sofia University,
prior to joining Amazon.
$\bullet$ We propose the first benchmark for evaluating the capabilities of language models on NLU in Bulgarian, bgGLUE, which includes nine diverse and challenging down-
|
[
[
0.2032470703125,
0.08074951171875,
0.79248046875,
0.08197021484375,
0.79248046875,
0.09979248046875,
0.2032470703125,
0.09857177734375
],
[
0.378662109375,
0.10272216796875,
0.62060546875,
0.10272216796875,
0.62060546875,
0.11639404296875,
0.378662109375,
0.11639404296875
],
[
0.16796875,
0.127685546875,
0.339111328125,
0.127685546875,
0.339111328125,
0.142578125,
0.16796875,
0.142578125
],
[
0.42822265625,
0.1270751953125,
0.57080078125,
0.1270751953125,
0.57080078125,
0.144287109375,
0.42822265625,
0.144287109375
],
[
0.68115234375,
0.127685546875,
0.8193359375,
0.127685546875,
0.8193359375,
0.142578125,
0.68115234375,
0.142578125
],
[
0.19482421875,
0.144287109375,
0.310546875,
0.144287109375,
0.310546875,
0.1591796875,
0.19482421875,
0.1591796875
],
[
0.359375,
0.1436767578125,
0.63818359375,
0.142578125,
0.63818359375,
0.1602783203125,
0.359375,
0.160888671875
],
[
0.7138671875,
0.144287109375,
0.78662109375,
0.144287109375,
0.78662109375,
0.1591796875,
0.7138671875,
0.1591796875
],
[
0.214111328125,
0.1728515625,
0.349365234375,
0.17333984375,
0.349365234375,
0.188232421875,
0.214111328125,
0.18701171875
],
[
0.45166015625,
0.1728515625,
0.5517578125,
0.1728515625,
0.5517578125,
0.18701171875,
0.45166015625,
0.18701171875
],
[
0.65576171875,
0.17333984375,
0.78515625,
0.17333984375,
0.78515625,
0.188232421875,
0.65576171875,
0.188232421875
],
[
0.343505859375,
0.189453125,
0.654296875,
0.190673828125,
0.654296875,
0.204833984375,
0.343505859375,
0.2037353515625
],
[
0.14111328125,
0.218505859375,
0.255126953125,
0.218505859375,
0.255126953125,
0.2327880859375,
0.14111328125,
0.2327880859375
],
[
0.339111328125,
0.218505859375,
0.459228515625,
0.218505859375,
0.459228515625,
0.2327880859375,
0.339111328125,
0.2327880859375
],
[
0.53564453125,
0.21728515625,
0.66162109375,
0.218505859375,
0.66162109375,
0.2327880859375,
0.53564453125,
0.232177734375
],
[
0.72705078125,
0.218505859375,
0.873046875,
0.218505859375,
0.873046875,
0.232177734375,
0.72705078125,
0.232177734375
],
[
0.33154296875,
0.2340087890625,
0.465087890625,
0.235107421875,
0.465087890625,
0.2493896484375,
0.33154296875,
0.248779296875
],
[
0.736328125,
0.2340087890625,
0.8623046875,
0.235107421875,
0.8623046875,
0.2493896484375,
0.736328125,
0.248779296875
],
[
0.55908203125,
0.236328125,
0.6396484375,
0.236328125,
0.6396484375,
0.2476806640625,
0.55908203125,
0.2476806640625
],
[
0.162841796875,
0.236328125,
0.2333984375,
0.236328125,
0.2333984375,
0.2476806640625,
0.162841796875,
0.2476806640625
],
[
0.2626953125,
0.26953125,
0.341796875,
0.26953125,
0.341796875,
0.283935546875,
0.2626953125,
0.283935546875
],
[
0.51318359375,
0.270751953125,
0.880859375,
0.270751953125,
0.880859375,
0.28515625,
0.51318359375,
0.28515625
],
[
0.51318359375,
0.287353515625,
0.7001953125,
0.287353515625,
0.7001953125,
0.3017578125,
0.51318359375,
0.3017578125
],
[
0.145263671875,
0.297607421875,
0.45751953125,
0.297607421875,
0.45751953125,
0.311279296875,
0.145263671875,
0.311279296875
],
[
0.1435546875,
0.312255859375,
0.459228515625,
0.31005859375,
0.459228515625,
0.32470703125,
0.1435546875,
0.32666015625
],
[
0.53076171875,
0.30712890625,
0.880859375,
0.30712890625,
0.880859375,
0.320556640625,
0.53076171875,
0.320556640625
],
[
0.145263671875,
0.325927734375,
0.45751953125,
0.325927734375,
0.45751953125,
0.339111328125,
0.145263671875,
0.339111328125
],
[
0.5146484375,
0.322509765625,
0.880859375,
0.322509765625,
0.880859375,
0.33740234375,
0.5146484375,
0.33740234375
],
[
0.1435546875,
0.34033203125,
0.45751953125,
0.339111328125,
0.45751953125,
0.35400390625,
0.1435546875,
0.355224609375
],
[
0.51318359375,
0.33837890625,
0.87890625,
0.33837890625,
0.87890625,
0.3515625,
0.51318359375,
0.3515625
],
[
0.14697265625,
0.355712890625,
0.252685546875,
0.355712890625,
0.252685546875,
0.36767578125,
0.14697265625,
0.36767578125
],
[
0.268798828125,
0.355224609375,
0.45751953125,
0.355224609375,
0.45751953125,
0.3681640625,
0.268798828125,
0.3681640625
],
[
0.5146484375,
0.355712890625,
0.8818359375,
0.355712890625,
0.8818359375,
0.369384765625,
0.5146484375,
0.369384765625
],
[
0.145263671875,
0.3681640625,
0.459228515625,
0.3681640625,
0.459228515625,
0.3818359375,
0.145263671875,
0.3818359375
],
[
0.5146484375,
0.371826171875,
0.880859375,
0.371826171875,
0.880859375,
0.385498046875,
0.5146484375,
0.385498046875
],
[
0.1435546875,
0.383056640625,
0.459228515625,
0.3818359375,
0.459228515625,
0.396728515625,
0.1435546875,
0.39794921875
],
[
0.51318359375,
0.385986328125,
0.880859375,
0.385986328125,
0.880859375,
0.40087890625,
0.51318359375,
0.40087890625
],
[
0.145263671875,
0.39794921875,
0.459228515625,
0.39794921875,
0.459228515625,
0.410888671875,
0.145263671875,
0.410888671875
],
[
0.5146484375,
0.402587890625,
0.8818359375,
0.402587890625,
0.8818359375,
0.41748046875,
0.5146484375,
0.41748046875
],
[
0.14697265625,
0.412109375,
0.459228515625,
0.412109375,
0.459228515625,
0.42578125,
0.14697265625,
0.42578125
],
[
0.1435546875,
0.42333984375,
0.456787109375,
0.42578125,
0.456787109375,
0.440673828125,
0.1435546875,
0.438232421875
],
[
0.51318359375,
0.419189453125,
0.87744140625,
0.419189453125,
0.87744140625,
0.432861328125,
0.51318359375,
0.432861328125
],
[
0.51318359375,
0.435302734375,
0.880859375,
0.435302734375,
0.880859375,
0.448974609375,
0.51318359375,
0.448974609375
],
[
0.14697265625,
0.441162109375,
0.455078125,
0.441162109375,
0.455078125,
0.451904296875,
0.14697265625,
0.451904296875
],
[
0.145263671875,
0.454833984375,
0.45751953125,
0.454833984375,
0.45751953125,
0.468505859375,
0.145263671875,
0.468505859375
],
[
0.51318359375,
0.45068359375,
0.880859375,
0.45068359375,
0.880859375,
0.465576171875,
0.51318359375,
0.465576171875
],
[
0.145263671875,
0.468505859375,
0.45751953125,
0.468505859375,
0.45751953125,
0.482177734375,
0.145263671875,
0.482177734375
],
[
0.51318359375,
0.46728515625,
0.880859375,
0.466064453125,
0.880859375,
0.48095703125,
0.51318359375,
0.482177734375
],
[
0.14697265625,
0.482666015625,
0.45751953125,
0.482666015625,
0.45751953125,
0.496337890625,
0.14697265625,
0.496337890625
],
[
0.51318359375,
0.48388671875,
0.880859375,
0.48388671875,
0.880859375,
0.49755859375,
0.51318359375,
0.49755859375
],
[
0.145263671875,
0.496337890625,
0.45751953125,
0.496337890625,
0.45751953125,
0.51025390625,
0.145263671875,
0.51025390625
],
[
0.51318359375,
0.50048828125,
0.87744140625,
0.50048828125,
0.87744140625,
0.51416015625,
0.51318359375,
0.51416015625
],
[
0.14697265625,
0.51123046875,
0.459228515625,
0.51025390625,
0.459228515625,
0.52392578125,
0.14697265625,
0.52490234375
],
[
0.51318359375,
0.5166015625,
0.87890625,
0.5166015625,
0.87890625,
0.52978515625,
0.51318359375,
0.52978515625
],
[
0.1485595703125,
0.52685546875,
0.456787109375,
0.52685546875,
0.456787109375,
0.53759765625,
0.1485595703125,
0.53759765625
],
[
0.51123046875,
0.5322265625,
0.58447265625,
0.5322265625,
0.58447265625,
0.54638671875,
0.51123046875,
0.54638671875
],
[
0.145263671875,
0.5390625,
0.45751953125,
0.5390625,
0.45751953125,
0.552734375,
0.145263671875,
0.552734375
],
[
0.145263671875,
0.55419921875,
0.459228515625,
0.552734375,
0.459228515625,
0.56640625,
0.145263671875,
0.56787109375
],
[
0.52880859375,
0.5517578125,
0.8818359375,
0.5517578125,
0.8818359375,
0.5654296875,
0.52880859375,
0.5654296875
],
[
0.145263671875,
0.56787109375,
0.459228515625,
0.56787109375,
0.459228515625,
0.58056640625,
0.145263671875,
0.58056640625
],
[
0.51123046875,
0.56640625,
0.880859375,
0.56787109375,
0.880859375,
0.58203125,
0.51123046875,
0.58056640625
],
[
0.14697265625,
0.58203125,
0.459228515625,
0.58056640625,
0.459228515625,
0.59423828125,
0.14697265625,
0.595703125
],
[
0.5146484375,
0.58447265625,
0.880859375,
0.58447265625,
0.880859375,
0.59716796875,
0.5146484375,
0.59716796875
],
[
0.145263671875,
0.5966796875,
0.459228515625,
0.5966796875,
0.459228515625,
0.60986328125,
0.145263671875,
0.60986328125
],
[
0.5146484375,
0.60107421875,
0.880859375,
0.60107421875,
0.880859375,
0.6142578125,
0.5146484375,
0.6142578125
],
[
0.1427001953125,
0.609375,
0.45751953125,
0.60986328125,
0.45751953125,
0.62451171875,
0.1427001953125,
0.62353515625
],
[
0.51318359375,
0.6142578125,
0.880859375,
0.6162109375,
0.880859375,
0.63134765625,
0.51318359375,
0.62890625
],
[
0.145263671875,
0.62451171875,
0.361083984375,
0.62451171875,
0.361083984375,
0.63818359375,
0.145263671875,
0.63818359375
],
[
0.51318359375,
0.6318359375,
0.880859375,
0.6318359375,
0.880859375,
0.6455078125,
0.51318359375,
0.6455078125
],
[
0.51318359375,
0.646484375,
0.880859375,
0.64794921875,
0.880859375,
0.662109375,
0.51318359375,
0.6611328125
],
[
0.1175537109375,
0.65380859375,
0.133544921875,
0.65380859375,
0.133544921875,
0.6630859375,
0.1175537109375,
0.6630859375
],
[
0.15283203125,
0.65283203125,
0.2626953125,
0.65283203125,
0.2626953125,
0.66455078125,
0.15283203125,
0.66455078125
],
[
0.51318359375,
0.6630859375,
0.880859375,
0.66455078125,
0.880859375,
0.6787109375,
0.51318359375,
0.677734375
],
[
0.118408203125,
0.6767578125,
0.48779296875,
0.6767578125,
0.48779296875,
0.69140625,
0.118408203125,
0.69140625
],
[
0.51318359375,
0.68115234375,
0.8818359375,
0.68115234375,
0.8818359375,
0.6953125,
0.51318359375,
0.6953125
],
[
0.198974609375,
0.693359375,
0.399658203125,
0.693359375,
0.399658203125,
0.7080078125,
0.198974609375,
0.7080078125
],
[
0.391357421875,
0.69140625,
0.48779296875,
0.693359375,
0.48779296875,
0.7080078125,
0.391357421875,
0.7060546875
],
[
0.515625,
0.6953125,
0.880859375,
0.69677734375,
0.880859375,
0.71142578125,
0.515625,
0.71044921875
],
[
0.1175537109375,
0.708984375,
0.486083984375,
0.708984375,
0.486083984375,
0.72265625,
0.1175537109375,
0.72265625
],
[
0.118408203125,
0.69482421875,
0.17041015625,
0.69482421875,
0.17041015625,
0.7060546875,
0.118408203125,
0.7060546875
],
[
0.5146484375,
0.71337890625,
0.880859375,
0.71337890625,
0.880859375,
0.72705078125,
0.5146484375,
0.72705078125
],
[
0.118408203125,
0.7255859375,
0.483642578125,
0.7255859375,
0.483642578125,
0.7392578125,
0.118408203125,
0.7392578125
],
[
0.5146484375,
0.72802734375,
0.8818359375,
0.72802734375,
0.8818359375,
0.7421875,
0.5146484375,
0.7421875
],
[
0.118408203125,
0.7421875,
0.48779296875,
0.7421875,
0.48779296875,
0.75732421875,
0.118408203125,
0.75732421875
],
[
0.51123046875,
0.74365234375,
0.75146484375,
0.74462890625,
0.75146484375,
0.7587890625,
0.51123046875,
0.75830078125
],
[
0.1175537109375,
0.75830078125,
0.483642578125,
0.75830078125,
0.483642578125,
0.771484375,
0.1175537109375,
0.771484375
],
[
0.52880859375,
0.76318359375,
0.880859375,
0.76416015625,
0.880859375,
0.779296875,
0.52880859375,
0.77783203125
],
[
0.1175537109375,
0.77392578125,
0.484375,
0.7724609375,
0.484375,
0.78759765625,
0.1175537109375,
0.7880859375
],
[
0.51318359375,
0.78076171875,
0.87890625,
0.78076171875,
0.87890625,
0.79443359375,
0.51318359375,
0.79443359375
],
[
0.118408203125,
0.79052734375,
0.486083984375,
0.79052734375,
0.486083984375,
0.80419921875,
0.118408203125,
0.80419921875
],
[
0.51708984375,
0.79638671875,
0.880859375,
0.79638671875,
0.880859375,
0.81005859375,
0.51708984375,
0.81005859375
],
[
0.118408203125,
0.80712890625,
0.486083984375,
0.80712890625,
0.486083984375,
0.82080078125,
0.118408203125,
0.82080078125
],
[
0.51318359375,
0.81103515625,
0.8330078125,
0.8125,
0.8330078125,
0.82666015625,
0.51318359375,
0.826171875
],
[
0.118408203125,
0.822265625,
0.438232421875,
0.822265625,
0.438232421875,
0.8359375,
0.118408203125,
0.8359375
],
[
0.1368408203125,
0.8388671875,
0.484375,
0.8388671875,
0.484375,
0.8525390625,
0.1368408203125,
0.8525390625
],
[
0.1175537109375,
0.85498046875,
0.486083984375,
0.85400390625,
0.486083984375,
0.8681640625,
0.1175537109375,
0.869140625
],
[
0.5322265625,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.5322265625,
0.87060546875
],
[
0.1175537109375,
0.87158203125,
0.48779296875,
0.87158203125,
0.48779296875,
0.88525390625,
0.1175537109375,
0.88525390625
],
[
0.54833984375,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.54833984375,
0.8857421875
],
[
0.54736328125,
0.88720703125,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.54736328125,
0.90185546875
],
[
0.1435546875,
0.8935546875,
0.486083984375,
0.8935546875,
0.486083984375,
0.90673828125,
0.1435546875,
0.90673828125
],
[
0.1175537109375,
0.90625,
0.270263671875,
0.90625,
0.270263671875,
0.91943359375,
0.1175537109375,
0.91943359375
],
[
0.54833984375,
0.90478515625,
0.8818359375,
0.90478515625,
0.8818359375,
0.91943359375,
0.54833984375,
0.91943359375
]
] |
[
"bgGLUE: A Bulgarian General Language Understanding",
"Evaluation Benchmark",
"Momchil Hardalov*",
"Pepa Atanasova",
"Todor Mihaylov",
"AWS AILabs",
"University of Copenhagen, DIKU",
"Meta AI",
"Galia Angelova",
"Kiril Simov",
"Petya Osenova",
"IICT, Bulgarian Academy of Sciences",
"Ves Stoyanov",
"Ivan Koychev",
"Preslav Nakov",
"Dragomir Radev",
"Sofia University",
"Yale University",
"MBZUAI",
"Meta AI",
"Abstract",
"2020)或多种语言（Liang et al.,2020;Hu et al.,",
"2020）上测试NLU任务",
"We present bgGLUE (Bulgarian General Lan-",
"guage Understanding Evaluation), a bench-",
"在这里，我们旨在通过提出bgGLUE来改",
"mark for evaluating language models on Nat",
"善NLU基准中表示的语言的多样性，bgGLUE",
"ural Language Understanding (NLU) tasks",
"是保加利亚语的基准，由九个NLU任务组成，",
"in Bulgarian.",
"Our benchmark includes",
"包括令牌分类，回归和分类。到目前为止，只",
"NLU tasks targeting a variety of NLP prob-",
"有保加利亚语的单个数据集被用于模型开发",
"lems (e.g., natural language inference, fact-",
"和评估。一些多语言基准测试中还包含了保加",
"checking, named entity recognition, senti-",
"利亚语中最多三个下游任务的小子集（Liang",
"ment analysis, question answering, etc.) and",
"machine learning tasks (sequence labeling,",
"etal.，2020；Huetal.,2020）。此外，还有针",
"对其他Balto-Slavic语言的现有基准测试，例",
"document-level classification,andregression)",
"We run the first systematic evaluation of pre-",
"如俄语SuperGLUEE(Shavrina et al.，2020）和",
"trained language models for Bulgarian, com-",
"斯洛文尼亚语SuperGLUEE(之agarandRobnik",
"paring and contrasting results across the nine",
"Sikonja，2022）。但是，没有为东南斯拉夫语",
"tasks in the benchmark. The evaluation results",
"分组的代表，特别是保加利亚语的代表制定",
"show strong performance on sequence label-",
"全面的基准。我们的目标是解决这些限制与",
"ing tasks,but there is a lot of room for im",
"bgGLUE",
"provement for tasks that require more complex",
"reasoning. We make bgGLUE publicly avail-",
"bgGLUE统一并促进了对保加利亚语现有数",
"able together with the fine-tuning and the eval-",
"据集和任务的访问。通过包括更具挑战性的",
"uation code, as well as a public leaderboard",
"任务，如自然语言推理，事实检查和问题回",
"at https:/bgglue.github.io, and we hope that it",
"答，我们确保它包括为保加利亚语开发的NLP",
"will enable further advancements in develop-",
"模型的严格测试集。我们还通过HuggingFace",
"ing NLU models for Bulgarian.",
"Hub(Lhoestetal.,2021)提供对基准的访问，以",
"便于使用，我们鼓励保加利亚语的模型共享。",
"1",
"Introduction",
"此外，我们进行了微调，并对保加利亚语的现",
"自然语言理解(NLU)基准测试，如",
"有语言模型进行了首次系统评估，比较和对比",
"(Wang et al., 2019b)",
"和Super-",
"了基准测试中所有任务的结果。我们的评估结",
"GLUE（Wangetal..2019a），旨在对各种",
"GLUE",
"果表明，更大，更健壮的预训练模型在所有任",
"自然语言理解(NLU）任务的语言模型进行严",
"务上都能产生更好的性能，而且高效的蒸馏模",
"格评估。此类基准的广泛采用推动了模型的快",
"型是大型模型的强大竞争对手",
"速开发，这些模型在这些基准的一部分任务上",
"这些模型在词性标记和命名实体识别方面表",
"表现良好，但也超出了E(Devlinetal.，2019",
"现出很强的性能，但在需要更复杂推理的任务",
"Liuetal.，2019）。但是，直到最近，此类基准",
"上却很挣扎，例如解决入学考试或评估新闻文",
"测试的重点一直放在英语上，对其他语言的兴",
"章的可信度和准确性。我们的贡献如下：",
"趣不大E(Bender,2011;Ponti etal.,2019)",
"为了解决这个问题，最近的工作设计了基准",
"测试模型，用于在非英语语言（Leetal.,2020;",
": We propose the first benchmark for eval-",
"Rodriguez-Penagos et al., 202l; Shavrina et al.,",
"uating the capabilities of language models",
"on NLU in Bulgarian, bgGLUE, which in-",
"*Work done while Momchil was in the Sofia University,",
"prior to joining Amazon.",
"cludes nine diverse and challenging down-"
] |
{
"arxiv_id": "2306.02349",
"language": "zh",
"page_idx": 1
}
|
||
2306.02349-zh-4
|
2306.02349-zh-4.jpg
|
## 2.5 Question Answering
High School Examinations (EXAMS) EXAMS (Hardalov et al., 2019, 2020) 是高中考试跨语言和多语言问答的基准数据集。它包含 26种语言的 24,000 多个高质量考试问题，涵盖自然科学和社会科学等八个语言家庭和 24 个学校科目。EXAMS 提供跨多种语言和主题的细粒度评估框架，允许对各种模型进行精确分析和比较
# 2.6 Scoring
在 bgGLUE 中，我们选择对每个任务进行平均加权的简单方法，对于具有多个指标的任务， 首先对这些分数进行平均以获得任务分数。
# 3 Data Preparation
在这里，我们描述了在将数据集包含在 bgGLUE 基准测试之前准备数据集所采取的预处理步骤。我们的主要目标是确保该设置以原则性的方式并在不同的领域中评估模型的语言理解能力。由于所有数据集都是公开可用的，因此我们尽可能保留了原始设置。然而，我们发现一些数据集包含在他们的训练/开发/测试分裂重复的例子，或者所有的分裂来自同一个域，这可能会高估模型的性能。因此，我们消除了数据泄漏，并提出了新的基于主题或基于时间的(i。e。, 基于时间戳)在需要的地方进行数据拆分。我们根据两对规范化文本中的完整单词重叠对示例进行了重复数据删除，即。e。, 小写，不包括所有停用词
对于 BSNLP 任务，我们组合了来自 NER 共享任务的三个连续版本的数据。我们选择了所有保加利亚的例子，其中包括六个不同的主题。我们使用最新的两个主题进行测试，并以4:1的比例随机拆分其余部分以进行训练和验证
可信新闻数据集包含来自六个来源和各种新闻主题的数据。我们按主题和来源拆分数据集。特别是，我们将来自同一主题的所有新闻文章包含在同一拆分中。对于培训和验证，我们使用了来自两个类别中最大来源的文档。此外， 我们从两者中提取了相同的主题，并在拆分中将它们分组在一起，将班级比例保持在 1:10。在测试数据集中，我们使用了来自所有六个数据源的数据点。$^{2}$我们注意到，我们的测试集包含来自两个新来源的所有数据点，这些数据点是最近的，这使得它们更具挑战性。最后我们清除了重复的文本，并删除了指示来源的所有关键字：作者姓名，媒体来源，URL 等。${^{2}}$We did not have overlapping topics from the same source
in the different splits.
有关新闻文章收集过程的更多详细信息，请参
见附录$\mathbb{E}$D
我们为 Cinexio 准备了一个全新的拆分，因为没有标准的。首先，我们删除了所有重复的评论，并按发布时间对剩余的评论进行了排序。接下来，我们将注释分成一半用于训练， 其余的用于验证和测试。最后，我们从训练集中删除了出现在其他两个拆分中的同一部电影的评论，这通过稍微增加训练集的比例来改变分布
对于 EXAMS, 我们保留了原始的验证和测试拆分 (Hardalov et al.,2020) 。我们将 Hardalov et al.(2019) 中的所有其他问题添加到训练集中，即。e。,类别历史在线测验
假新闻数据集作为共享任务的一部分发布， 它已经被拆分为训练集和测试集。经过人工检查，我们发现有些文章仅对其标题进行了改写，并且内容相同。因此，我们仅根据文章正文中的精确匹配删除了重复项。此外，我们指定了 10 个% 的训练示例进行验证
PAN-X 包含维基百科的简短句子，根据可用的 Wiki 实体自动注释。数据集由 20K 个用于训练的示例和 10K 个用于验证和测试的示例组成。我们保留了建议的拆分，但是我们通过将每个句子转换为小写并删除标点符号来检查重复项。这导致了 10K 句子的整体删除
我们的分析没有发现 CT21 的任何问题。T1、 U。Dep 和 XNLI, 因此我们保留了最初提供的拆分。
## 4 $\textbf{Experiments}$
在本节中，我们首先描述我们实验的基准系
统，然后介绍评估结果
## 4.1 Baselines
Majority and Random Baselines 根据每个测试集中标签的分布计算多数类基线。在随机基线中，每个测试实例以相等的概率随机分配一个目标标签
Fine-tuned models 我们的基线包括几个突
出的仅多语言编码器预训练的 Transformer 模型。我们根据他们的预培训目标将他们划分如下：
I Masked language modeling :
$\bullet$ mBERT (Devlin et al., 2019) We use the base cased version, trained on 104 languages, including Bulgarian. The pre-training task is done on a Wikipedia dump for each language. $\bullet$ XLM-R (Conneau et al., 2020) We evaluate the Base and the Large versions of the model.
|
[
[
0.1158447265625,
0.08489990234375,
0.32421875,
0.0872802734375,
0.32421875,
0.1015625,
0.1158447265625,
0.09979248046875
],
[
0.5146484375,
0.0872802734375,
0.87890625,
0.0872802734375,
0.87890625,
0.1009521484375,
0.5146484375,
0.1009521484375
],
[
0.1259765625,
0.10809326171875,
0.49365234375,
0.10809326171875,
0.49365234375,
0.1217041015625,
0.1259765625,
0.1217041015625
],
[
0.509765625,
0.10272216796875,
0.60205078125,
0.1009521484375,
0.60205078125,
0.11517333984375,
0.509765625,
0.1175537109375
],
[
0.118408203125,
0.12353515625,
0.48779296875,
0.12353515625,
0.48779296875,
0.13720703125,
0.118408203125,
0.13720703125
],
[
0.53076171875,
0.119384765625,
0.880859375,
0.119384765625,
0.880859375,
0.133056640625,
0.53076171875,
0.133056640625
],
[
0.1175537109375,
0.1395263671875,
0.48779296875,
0.1383056640625,
0.48779296875,
0.152587890625,
0.1175537109375,
0.15380859375
],
[
0.51123046875,
0.1353759765625,
0.880859375,
0.1353759765625,
0.880859375,
0.1484375,
0.51123046875,
0.1484375
],
[
0.118408203125,
0.1561279296875,
0.484375,
0.1561279296875,
0.484375,
0.169189453125,
0.118408203125,
0.169189453125
],
[
0.51318359375,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.1650390625,
0.51318359375,
0.1650390625
],
[
0.118408203125,
0.1728515625,
0.48779296875,
0.1728515625,
0.48779296875,
0.1859130859375,
0.118408203125,
0.1859130859375
],
[
0.51318359375,
0.168701171875,
0.8818359375,
0.168701171875,
0.8818359375,
0.1817626953125,
0.51318359375,
0.1817626953125
],
[
0.118408203125,
0.188232421875,
0.48779296875,
0.188232421875,
0.48779296875,
0.201904296875,
0.118408203125,
0.201904296875
],
[
0.51318359375,
0.18408203125,
0.880859375,
0.18408203125,
0.880859375,
0.19775390625,
0.51318359375,
0.19775390625
],
[
0.118408203125,
0.204833984375,
0.486083984375,
0.204833984375,
0.486083984375,
0.218505859375,
0.118408203125,
0.218505859375
],
[
0.51318359375,
0.1995849609375,
0.880859375,
0.1995849609375,
0.880859375,
0.213134765625,
0.51318359375,
0.213134765625
],
[
0.118408203125,
0.218505859375,
0.175537109375,
0.2203369140625,
0.173828125,
0.235107421875,
0.1175537109375,
0.2327880859375
],
[
0.5146484375,
0.21728515625,
0.87744140625,
0.21728515625,
0.87744140625,
0.22802734375,
0.5146484375,
0.22802734375
],
[
0.51318359375,
0.2327880859375,
0.5498046875,
0.2327880859375,
0.5498046875,
0.2452392578125,
0.51318359375,
0.2452392578125
],
[
0.11419677734375,
0.2464599609375,
0.2275390625,
0.2476806640625,
0.2275390625,
0.262451171875,
0.11419677734375,
0.26123046875
],
[
0.53076171875,
0.248779296875,
0.87890625,
0.248779296875,
0.87890625,
0.262451171875,
0.53076171875,
0.262451171875
],
[
0.118408203125,
0.268310546875,
0.48779296875,
0.268310546875,
0.48779296875,
0.281982421875,
0.118408203125,
0.281982421875
],
[
0.51318359375,
0.265380859375,
0.87890625,
0.265380859375,
0.87890625,
0.279052734375,
0.51318359375,
0.279052734375
],
[
0.118408203125,
0.28515625,
0.490234375,
0.28515625,
0.490234375,
0.298583984375,
0.118408203125,
0.298583984375
],
[
0.509765625,
0.27978515625,
0.87890625,
0.27978515625,
0.87890625,
0.29345703125,
0.509765625,
0.29345703125
],
[
0.1175537109375,
0.2998046875,
0.45166015625,
0.2998046875,
0.45166015625,
0.3134765625,
0.1175537109375,
0.3134765625
],
[
0.51318359375,
0.297607421875,
0.77490234375,
0.297607421875,
0.77490234375,
0.311279296875,
0.51318359375,
0.311279296875
],
[
0.5322265625,
0.3134765625,
0.8818359375,
0.3134765625,
0.8818359375,
0.32666015625,
0.5322265625,
0.32666015625
],
[
0.11419677734375,
0.327880859375,
0.30322265625,
0.328857421875,
0.30322265625,
0.34326171875,
0.11419677734375,
0.342529296875
],
[
0.5146484375,
0.330078125,
0.880859375,
0.330078125,
0.880859375,
0.34326171875,
0.5146484375,
0.34326171875
],
[
0.51318359375,
0.345703125,
0.880859375,
0.345703125,
0.880859375,
0.359375,
0.51318359375,
0.359375
],
[
0.118408203125,
0.355224609375,
0.489501953125,
0.355224609375,
0.489501953125,
0.3681640625,
0.118408203125,
0.3681640625
],
[
0.51318359375,
0.3623046875,
0.87744140625,
0.3623046875,
0.87744140625,
0.372314453125,
0.51318359375,
0.372314453125
],
[
0.1175537109375,
0.369384765625,
0.484375,
0.37060546875,
0.484375,
0.38427734375,
0.1175537109375,
0.383056640625
],
[
0.5146484375,
0.377685546875,
0.880859375,
0.377685546875,
0.880859375,
0.391357421875,
0.5146484375,
0.391357421875
],
[
0.118408203125,
0.38720703125,
0.486083984375,
0.38720703125,
0.486083984375,
0.40087890625,
0.118408203125,
0.40087890625
],
[
0.51318359375,
0.393798828125,
0.78271484375,
0.393798828125,
0.78271484375,
0.40673828125,
0.51318359375,
0.40673828125
],
[
0.118408203125,
0.405029296875,
0.48193359375,
0.405029296875,
0.48193359375,
0.4150390625,
0.118408203125,
0.4150390625
],
[
0.52880859375,
0.410400390625,
0.880859375,
0.410400390625,
0.880859375,
0.42333984375,
0.52880859375,
0.42333984375
],
[
0.118408203125,
0.419189453125,
0.484375,
0.419189453125,
0.484375,
0.432861328125,
0.118408203125,
0.432861328125
],
[
0.51123046875,
0.427001953125,
0.87890625,
0.427001953125,
0.87890625,
0.440673828125,
0.51123046875,
0.440673828125
],
[
0.118408203125,
0.435302734375,
0.484375,
0.435302734375,
0.484375,
0.448974609375,
0.118408203125,
0.448974609375
],
[
0.5146484375,
0.443603515625,
0.880859375,
0.443603515625,
0.880859375,
0.457275390625,
0.5146484375,
0.457275390625
],
[
0.118408203125,
0.451904296875,
0.484375,
0.451904296875,
0.484375,
0.465576171875,
0.118408203125,
0.465576171875
],
[
0.51123046875,
0.457763671875,
0.880859375,
0.458984375,
0.880859375,
0.473876953125,
0.51123046875,
0.47265625
],
[
0.118408203125,
0.468505859375,
0.484375,
0.468505859375,
0.484375,
0.482177734375,
0.118408203125,
0.482177734375
],
[
0.51318359375,
0.4755859375,
0.880859375,
0.4755859375,
0.880859375,
0.4892578125,
0.51318359375,
0.4892578125
],
[
0.118408203125,
0.48388671875,
0.484375,
0.48388671875,
0.484375,
0.49755859375,
0.118408203125,
0.49755859375
],
[
0.51318359375,
0.4912109375,
0.82958984375,
0.4912109375,
0.82958984375,
0.5048828125,
0.51318359375,
0.5048828125
],
[
0.118408203125,
0.49951171875,
0.484375,
0.49951171875,
0.484375,
0.51318359375,
0.118408203125,
0.51318359375
],
[
0.52880859375,
0.50830078125,
0.87890625,
0.50830078125,
0.87890625,
0.521484375,
0.52880859375,
0.521484375
],
[
0.118408203125,
0.5166015625,
0.486083984375,
0.5166015625,
0.486083984375,
0.52978515625,
0.118408203125,
0.52978515625
],
[
0.509765625,
0.52392578125,
0.87890625,
0.52392578125,
0.87890625,
0.53759765625,
0.509765625,
0.53759765625
],
[
0.118408203125,
0.5322265625,
0.486083984375,
0.5322265625,
0.486083984375,
0.5458984375,
0.118408203125,
0.5458984375
],
[
0.509765625,
0.54052734375,
0.56103515625,
0.54052734375,
0.56103515625,
0.5546875,
0.509765625,
0.5546875
],
[
0.118408203125,
0.548828125,
0.48779296875,
0.548828125,
0.48779296875,
0.5625,
0.118408203125,
0.5625
],
[
0.1158447265625,
0.56396484375,
0.361083984375,
0.5634765625,
0.361083984375,
0.57763671875,
0.1158447265625,
0.5791015625
],
[
0.51123046875,
0.57080078125,
0.6591796875,
0.57080078125,
0.6591796875,
0.58447265625,
0.51123046875,
0.58447265625
],
[
0.1351318359375,
0.580078125,
0.486083984375,
0.580078125,
0.486083984375,
0.59423828125,
0.1351318359375,
0.59423828125
],
[
0.118408203125,
0.5966796875,
0.486083984375,
0.5966796875,
0.486083984375,
0.60986328125,
0.118408203125,
0.60986328125
],
[
0.509765625,
0.59423828125,
0.880859375,
0.595703125,
0.880859375,
0.60986328125,
0.509765625,
0.609375
],
[
0.118408203125,
0.61083984375,
0.486083984375,
0.61083984375,
0.486083984375,
0.62451171875,
0.118408203125,
0.62451171875
],
[
0.509765625,
0.61328125,
0.6943359375,
0.61328125,
0.6943359375,
0.62353515625,
0.509765625,
0.62353515625
],
[
0.118408203125,
0.62890625,
0.484375,
0.62890625,
0.484375,
0.642578125,
0.118408203125,
0.642578125
],
[
0.51123046875,
0.64111328125,
0.5556640625,
0.64111328125,
0.5556640625,
0.65283203125,
0.51123046875,
0.65283203125
],
[
0.5556640625,
0.64013671875,
0.63232421875,
0.64013671875,
0.63232421875,
0.65478515625,
0.5556640625,
0.65478515625
],
[
0.12005615234375,
0.6591796875,
0.14111328125,
0.6630859375,
0.1351318359375,
0.6767578125,
0.11419677734375,
0.6728515625
],
[
0.1158447265625,
0.64453125,
0.486083984375,
0.64453125,
0.486083984375,
0.65771484375,
0.1158447265625,
0.65771484375
],
[
0.51708984375,
0.662109375,
0.880859375,
0.6611328125,
0.880859375,
0.67578125,
0.51708984375,
0.6767578125
],
[
0.1376953125,
0.6767578125,
0.486083984375,
0.6767578125,
0.486083984375,
0.6904296875,
0.1376953125,
0.6904296875
],
[
0.51318359375,
0.6787109375,
0.880859375,
0.6787109375,
0.880859375,
0.6923828125,
0.51318359375,
0.6923828125
],
[
0.118408203125,
0.693359375,
0.489501953125,
0.693359375,
0.489501953125,
0.70703125,
0.118408203125,
0.70703125
],
[
0.51123046875,
0.69482421875,
0.880859375,
0.69482421875,
0.880859375,
0.7080078125,
0.51123046875,
0.7080078125
],
[
0.12005615234375,
0.71142578125,
0.484375,
0.71142578125,
0.484375,
0.7216796875,
0.12005615234375,
0.7216796875
],
[
0.5146484375,
0.71533203125,
0.52880859375,
0.71533203125,
0.52880859375,
0.7197265625,
0.5146484375,
0.7197265625
],
[
0.5283203125,
0.7119140625,
0.623046875,
0.7119140625,
0.623046875,
0.7236328125,
0.5283203125,
0.7236328125
],
[
0.12005615234375,
0.7255859375,
0.486083984375,
0.7255859375,
0.486083984375,
0.7392578125,
0.12005615234375,
0.7392578125
],
[
0.1175537109375,
0.74072265625,
0.486083984375,
0.74072265625,
0.486083984375,
0.75439453125,
0.1175537109375,
0.75439453125
],
[
0.51708984375,
0.736328125,
0.87890625,
0.736328125,
0.87890625,
0.75,
0.51708984375,
0.75
],
[
0.12005615234375,
0.7587890625,
0.483642578125,
0.7587890625,
0.483642578125,
0.76953125,
0.12005615234375,
0.76953125
],
[
0.51318359375,
0.75439453125,
0.880859375,
0.75439453125,
0.880859375,
0.76708984375,
0.51318359375,
0.76708984375
],
[
0.51123046875,
0.7685546875,
0.880859375,
0.76953125,
0.880859375,
0.783203125,
0.51123046875,
0.7822265625
],
[
0.118408203125,
0.77490234375,
0.48193359375,
0.77490234375,
0.48193359375,
0.78515625,
0.118408203125,
0.78515625
],
[
0.118408203125,
0.79052734375,
0.48779296875,
0.79052734375,
0.48779296875,
0.80419921875,
0.118408203125,
0.80419921875
],
[
0.5146484375,
0.78759765625,
0.5400390625,
0.78759765625,
0.5400390625,
0.798828125,
0.5146484375,
0.798828125
],
[
0.118408203125,
0.8056640625,
0.484375,
0.8056640625,
0.484375,
0.8193359375,
0.118408203125,
0.8193359375
],
[
0.50732421875,
0.80419921875,
0.7421875,
0.80419921875,
0.7421875,
0.81787109375,
0.50732421875,
0.81787109375
],
[
0.118408203125,
0.82373046875,
0.483642578125,
0.82373046875,
0.483642578125,
0.83447265625,
0.118408203125,
0.83447265625
],
[
0.53076171875,
0.82177734375,
0.87890625,
0.82177734375,
0.87890625,
0.8349609375,
0.53076171875,
0.8349609375
],
[
0.118408203125,
0.83837890625,
0.486083984375,
0.83837890625,
0.486083984375,
0.8515625,
0.118408203125,
0.8515625
],
[
0.54736328125,
0.83740234375,
0.880859375,
0.83837890625,
0.880859375,
0.8525390625,
0.54736328125,
0.8515625
],
[
0.118408203125,
0.85400390625,
0.486083984375,
0.85400390625,
0.486083984375,
0.86767578125,
0.118408203125,
0.86767578125
],
[
0.54736328125,
0.8525390625,
0.880859375,
0.85400390625,
0.880859375,
0.8681640625,
0.54736328125,
0.86767578125
],
[
0.118408203125,
0.87060546875,
0.484375,
0.87060546875,
0.484375,
0.88427734375,
0.118408203125,
0.88427734375
],
[
0.54833984375,
0.869140625,
0.880859375,
0.87060546875,
0.880859375,
0.88525390625,
0.54833984375,
0.88427734375
],
[
0.1368408203125,
0.89111328125,
0.486083984375,
0.8955078125,
0.486083984375,
0.91015625,
0.1368408203125,
0.90625
],
[
0.5341796875,
0.8896484375,
0.8818359375,
0.8896484375,
0.8818359375,
0.90283203125,
0.5341796875,
0.90283203125
],
[
0.1175537109375,
0.90625,
0.2451171875,
0.90673828125,
0.2451171875,
0.91845703125,
0.1175537109375,
0.91748046875
],
[
0.54833984375,
0.90625,
0.880859375,
0.90625,
0.880859375,
0.91943359375,
0.54833984375,
0.91943359375
]
] |
[
"2.5QuestionAnswering",
"有关新闻文章收集过程的更多详细信息，请参",
"High School Examinations (EXAMS)EX-",
"见附录ED",
"AMS(Hardalovetal.,2019,2020）是高中考试",
"我们为Cinexio准备了一个全新的拆分，因",
"跨语言和多语言问答的基准数据集。它包含26",
"为没有标准的。首先，我们删除了所有重复的",
"种语言的24,000多个高质量考试问题，涵盖自",
"评论，并按发布时间对剩余的评论进行了排",
"然科学和社会科学等八个语言家庭和24个学",
"序。接下来，我们将注释分成一半用于训练，",
"校科目。EXAMS提供跨多种语言和主题的细",
"其余的用于验证和测试。最后，我们从训练集",
"粒度评估框架，允许对各种模型进行精确分析",
"中删除了出现在其他两个拆分中的同一部电影",
"和比较",
"的评论，这通过稍微增加训练集的比例来改变",
"分布",
"2.6Scoring",
"对于EXAMS，我们保留了原始的验证和测",
"在bgGLUE中，我们选择对每个任务进行平均",
"试拆分（Hardalovetal.,2020）。我们将Hardalov",
"加权的简单方法，对于具有多个指标的任务，",
"etal.(2019）中的所有其他问题添加到训练集",
"首先对这些分数进行平均以获得任务分数",
"中，即。e。，类别历史在线测验",
"假新闻数据集作为共享任务的一部分发布，",
"3DataPreparation",
"它已经被拆分为训练集和测试集。经过人工",
"检查，我们发现有些文章仅对其标题进行了改",
"在这里，我们描述了在将数据集包含在bg-",
"写，并目内容相同。因此，我们仅根据文竟正",
"GLUE基准测试之前准备数据集所采取的预处",
"文中的精确匹配删除了重复项。此外，我们指",
"理步骤。我们的主要目标是确保该设置以原则",
"定了10个%的训练示例进行验证",
"性的方式并在不同的领域中评估模型的语言理",
"PAN-X包含维基百科的简短句子，根据可用",
"解能力。由于所有数据集都是公开可用的，因",
"的Wiki实体自动注释。数据集由20K个用于",
"此我们尽可能保留了原始设置。然而，我们发",
"训练的示例和10K个用于验证和测试的示例",
"现一些数据集包含在他们的训练/开发/测试分",
"组成。我们保留了建议的拆分，但是我们通过",
"裂重复的例子，或者所有的分裂来自同一个",
"将每个句子转换为小写并删除标点符号来检查",
"域，这可能会高估模型的性能。因此，我们消",
"重复项。这导致了10K句子的整体删除",
"除了数据泄漏，并提出了新的基于主题或基于",
"我们的分析没有发现CT21的任何问题。T1、",
"时间的(i。e。，基于时间戳)在需要的地方进",
"U。Dep和XNLI，因此我们保留了最初提供的",
"行数据拆分。我们根据两对规范化文本中的完",
"拆分。",
"整单词重叠对示例进行了重复数据删除，即。",
"e。，小写，不包括所有停用词",
"4Experiments",
"对于BSNLP任务，我们组合了来自NER共",
"享任务的三个连续版本的数据。我们选择了",
"在本节中，我们首先描述我们实验的基准系",
"所有保加利亚的例子，其中包括六个不同的主",
"统，然后介绍评估结果",
"题。我们使用最新的两个主题进行测试，并以",
"4.1",
"Baselines",
"证",
"4：1的比例随机拆分其余部分以进行训练和验",
"MajorityandRandomBaselines根据每个",
"可信新闻数据集包含来自六个来源和各种新",
"测试集中标签的分布计算多数类基线。在随机",
"闻主题的数据。我们按主题和来源拆分数据集。",
"基线中，每个测试实例以相等的概率随机分配",
"特别是，我们将来自同一主题的所有新闻文章",
"",
"个目标标签",
"包含在同一拆分中。对于培训和验证，我们使",
"用了来自两个类别中最大来源的文档。此外，",
"Fine-tunedmodels我们的基线包括几个突",
"我们从两者中提取了相同的主题，并在折分中",
"出的仅多语言编码器预训练的Transformer模",
"型。我们根据他们的预培训目标将他们划分如",
"将它们分组在一起，将班级比例保持在1:10",
"在测试数据集中，我们使用了来自所有六个数",
"下:",
"据源的数据点。2我们注意到，我们的测试集",
"I Masked language modeling :",
"包含来自两个新来源的所有数据点，这些数据",
": mBERT (Devlin et al., 2019) We use the",
"点是最近的，这使得它们更具挑战性。最后，",
"base cased version, trained on 104 languages,",
"我们清除了重复的文本，并删除了指示来源的",
"including Bulgarian. The pre-training task is",
"所有关键字：作者姓名，媒体来源，URL等。",
"done on a Wikipedia dump for each language.",
"2we did not have overlapping topics from the same source",
": XLM-R (Conneau et al., 2020) We evaluate",
"inthedifferentsplits",
"the Base and the Large versions of the model."
] |
{
"arxiv_id": "2306.02349",
"language": "zh",
"page_idx": 4
}
|
||
2204.02906-zh-1
|
2204.02906-zh-1.jpg
|
# 通过维度和精度降低进行知识库索引压缩
Vilém Zouhar, Marius Mosbach, Miaoran Zhang and Dietrich Klakow Department of Language Science & Technology, Saarland Informatics Campus,
Saarland University, Germany
$\{$ vzouhar,mmosbach,mzhang,dklakow $\}$ @lsv.uni-saarland.de
## Abstract
Recently neural network based approaches to knowledge-intensive NLP tasks, such as question answering, started to rely heavily on the combination of neural retrievers and readers. Retrieval is typically performed over a large textual knowledge base (KB) which requires significant memory and compute resources, especially when scaled up. On HotpotQA we systematically investigate reducing the size of the KB index by means of dimensionality (sparse random projections, PCA, autoencoders) and numerical precision reduction. Our results show that PCA is an easy solution that requires very little data and is only slightly worse than autoencoders, which are less stable. All methods are sensitive to pre- and postprocessing and data should always be centered and normalized both before and after dimension reduction. Finally, we show that it is possible to combine PCA with using 1bit per dimension. Overall we achieve (1) 100×compression with 75 % , and (2) 24$\times$compression with 92 % original retrieval performance.
### 1 $\textbf{Introduction}$
知识密集型 NLP 任务的最新方法将基于神经网络的模型与利用密集向量表示 (Guu et al., 2020; Lewis et al., 2020; Petroni et al., 2021) 的检索组件相结合。最简单的例子是问题回答其中检索器接收一个问题作为输人，并返回供阅读器 (编码器和解码器) 使用的相关文档，该文档输出答案(Chen,2020)。相同的方法也可以应用于其他上下文，例如事实检查(Tchechmedjiev et al., 2019) 或知识对话 (Dinan et al.,2018)。此外，这种范式也可以应用于利用 e 的系统。。g。从训练语料库中缓存上下文以提供更好的输出，例如 Khandelwal et al. (2019)提出的 k 最近邻语言模型或 Yogatama et al.(2021) 提出的动态门控语言模型机制。所有这些管道都被概括为从知识库 (Zouhar et al., 2021)中检索人工制品，在该知识库中，读者与查询一起受到制约
至关重要的是，前面的所有示例都依赖于检索组件和知识库的质量。知识库通常由密集向量表示$^{1}$索引，检索组件通常使用内积或 $L^2$ 距离执行最大相似性搜索，以从知识库中检索文档$^{2}$。仅索引就占用了大量的知识库，使得部署和实验非常困难。检索速度还取决于索引向量的维数。大型知识库的一个示例是 Borgeaud et al.(2021)的工作，该工作通过 1 的数据库执行检索。80 亿份文件
本文重点研究了通过降维和精度来压缩索引
的问题，并做出了以下贡献：
$\bullet$ Comparison of various unsupervised index
compression methods for retrieval, including random projections, PCA, autoencoder, precision reduction and their combination.
$\bullet$Examination of effective pre- and postprocessing transformations, showing that centering and normalization are necessary for boosting the performance.
· Analysis on the impact of adding irrelevant
documents and retrieval errors. Recommen-
dations for use by practicioners.
在 Section 3 中，我们描述了问题场景和实验
设置。我们讨论了 Section 4 中不同压缩方法的结果。我们在 Section 5 中提供了进一步的分析，并在 Section 6 中给出了使用建议。此项目的存储库是开源的。$^{3}$
### 2 $\textbf{Related Work}$
Reducing index size. Luan et al.(2021)对双编码器上下文中信息检索的降维问题进行了全面的概述。尽管深人并以形式论据为基础，但他们的研究重点是总体上降维的局限性和性质(即使使用稀疏表示)以及文档长度对性能的
Sparse representations via BM25 (Robertson et al. $1995)\ddot{\text{are also commonly used but not the focus of this work}}$
$^{2}$We refer to the retrieved objects as documents though they commonly range from spans of text (e.g. 100 tokens) to the full documents.
github.com/zouharvi/kb-shrink
|
[
[
0.450927734375,
0.09320068359375,
0.54833984375,
0.09320068359375,
0.54833984375,
0.11102294921875,
0.450927734375,
0.11102294921875
],
[
0.334228515625,
0.114013671875,
0.66748046875,
0.114013671875,
0.66748046875,
0.12890625,
0.334228515625,
0.12890625
],
[
0.189697265625,
0.15087890625,
0.81689453125,
0.1519775390625,
0.81689453125,
0.166259765625,
0.189697265625,
0.1650390625
],
[
0.181396484375,
0.169189453125,
0.82275390625,
0.169189453125,
0.82275390625,
0.182861328125,
0.181396484375,
0.182861328125
],
[
0.378662109375,
0.1859130859375,
0.623046875,
0.1859130859375,
0.623046875,
0.1995849609375,
0.378662109375,
0.1995849609375
],
[
0.17041015625,
0.203125,
0.84619140625,
0.203125,
0.84619140625,
0.2161865234375,
0.17041015625,
0.2161865234375
],
[
0.2626953125,
0.2529296875,
0.341796875,
0.2529296875,
0.341796875,
0.267333984375,
0.2626953125,
0.267333984375
],
[
0.5322265625,
0.254638671875,
0.880859375,
0.254638671875,
0.880859375,
0.268310546875,
0.5322265625,
0.268310546875
],
[
0.51123046875,
0.26953125,
0.87890625,
0.26953125,
0.87890625,
0.283203125,
0.51123046875,
0.283203125
],
[
0.145263671875,
0.281982421875,
0.45751953125,
0.281982421875,
0.45751953125,
0.295654296875,
0.145263671875,
0.295654296875
],
[
0.5146484375,
0.2861328125,
0.880859375,
0.2861328125,
0.880859375,
0.2998046875,
0.5146484375,
0.2998046875
],
[
0.1485595703125,
0.297607421875,
0.456787109375,
0.297607421875,
0.456787109375,
0.30810546875,
0.1485595703125,
0.30810546875
],
[
0.5146484375,
0.302734375,
0.880859375,
0.302734375,
0.880859375,
0.31640625,
0.5146484375,
0.31640625
],
[
0.1435546875,
0.308837890625,
0.45751953125,
0.30810546875,
0.45751953125,
0.322509765625,
0.1435546875,
0.32373046875
],
[
0.51318359375,
0.317626953125,
0.8818359375,
0.318359375,
0.8818359375,
0.333251953125,
0.51318359375,
0.33203125
],
[
0.14697265625,
0.325927734375,
0.455078125,
0.325927734375,
0.455078125,
0.336181640625,
0.14697265625,
0.336181640625
],
[
0.145263671875,
0.33837890625,
0.45751953125,
0.33837890625,
0.45751953125,
0.3515625,
0.145263671875,
0.3515625
],
[
0.51318359375,
0.3349609375,
0.880859375,
0.3349609375,
0.880859375,
0.3486328125,
0.51318359375,
0.3486328125
],
[
0.14697265625,
0.35400390625,
0.456787109375,
0.35400390625,
0.456787109375,
0.364013671875,
0.14697265625,
0.364013671875
],
[
0.5146484375,
0.3515625,
0.87744140625,
0.3515625,
0.87744140625,
0.3623046875,
0.5146484375,
0.3623046875
],
[
0.145263671875,
0.36767578125,
0.456787109375,
0.36767578125,
0.456787109375,
0.38134765625,
0.145263671875,
0.38134765625
],
[
0.51123046875,
0.366455078125,
0.880859375,
0.366455078125,
0.880859375,
0.380126953125,
0.51123046875,
0.380126953125
],
[
0.145263671875,
0.38134765625,
0.45751953125,
0.38134765625,
0.45751953125,
0.394287109375,
0.145263671875,
0.394287109375
],
[
0.51318359375,
0.3818359375,
0.68408203125,
0.383056640625,
0.68408203125,
0.39794921875,
0.51318359375,
0.396728515625
],
[
0.145263671875,
0.3955078125,
0.45751953125,
0.3955078125,
0.45751953125,
0.4091796875,
0.145263671875,
0.4091796875
],
[
0.5322265625,
0.399658203125,
0.880859375,
0.399658203125,
0.880859375,
0.413330078125,
0.5322265625,
0.413330078125
],
[
0.145263671875,
0.4091796875,
0.459228515625,
0.4091796875,
0.459228515625,
0.4228515625,
0.145263671875,
0.4228515625
],
[
0.51318359375,
0.4150390625,
0.7421875,
0.4150390625,
0.7421875,
0.4287109375,
0.51318359375,
0.4287109375
],
[
0.145263671875,
0.42333984375,
0.459228515625,
0.42333984375,
0.459228515625,
0.43701171875,
0.145263671875,
0.43701171875
],
[
0.145263671875,
0.438232421875,
0.45751953125,
0.438232421875,
0.45751953125,
0.451904296875,
0.145263671875,
0.451904296875
],
[
0.53564453125,
0.4423828125,
0.880859375,
0.4423828125,
0.880859375,
0.4560546875,
0.53564453125,
0.4560546875
],
[
0.145263671875,
0.451904296875,
0.456787109375,
0.451904296875,
0.456787109375,
0.465576171875,
0.145263671875,
0.465576171875
],
[
0.54736328125,
0.457763671875,
0.87890625,
0.457763671875,
0.87890625,
0.471435546875,
0.54736328125,
0.471435546875
],
[
0.145263671875,
0.466064453125,
0.45751953125,
0.466064453125,
0.45751953125,
0.479736328125,
0.145263671875,
0.479736328125
],
[
0.54736328125,
0.473876953125,
0.87890625,
0.473876953125,
0.87890625,
0.48681640625,
0.54736328125,
0.48681640625
],
[
0.145263671875,
0.479736328125,
0.459228515625,
0.479736328125,
0.459228515625,
0.493408203125,
0.145263671875,
0.493408203125
],
[
0.145263671875,
0.49462890625,
0.45751953125,
0.495361328125,
0.45751953125,
0.50732421875,
0.145263671875,
0.505859375
],
[
0.54833984375,
0.490478515625,
0.8369140625,
0.490478515625,
0.8369140625,
0.50439453125,
0.54833984375,
0.50439453125
],
[
0.5322265625,
0.50439453125,
0.88330078125,
0.5048828125,
0.88330078125,
0.51953125,
0.5322265625,
0.5185546875
],
[
0.14697265625,
0.51025390625,
0.45751953125,
0.51025390625,
0.45751953125,
0.52099609375,
0.14697265625,
0.52099609375
],
[
0.145263671875,
0.5224609375,
0.459228515625,
0.5224609375,
0.459228515625,
0.5361328125,
0.145263671875,
0.5361328125
],
[
0.54736328125,
0.5224609375,
0.88330078125,
0.521484375,
0.88330078125,
0.5361328125,
0.54736328125,
0.53759765625
],
[
0.14697265625,
0.53759765625,
0.459228515625,
0.53759765625,
0.459228515625,
0.55029296875,
0.14697265625,
0.55029296875
],
[
0.54833984375,
0.5380859375,
0.8818359375,
0.5380859375,
0.8818359375,
0.552734375,
0.54833984375,
0.552734375
],
[
0.145263671875,
0.5517578125,
0.459228515625,
0.5517578125,
0.459228515625,
0.5654296875,
0.145263671875,
0.5654296875
],
[
0.54736328125,
0.5546875,
0.74462890625,
0.5546875,
0.74462890625,
0.568359375,
0.54736328125,
0.568359375
],
[
0.14697265625,
0.5654296875,
0.459228515625,
0.5654296875,
0.459228515625,
0.5791015625,
0.14697265625,
0.5791015625
],
[
0.5322265625,
0.57080078125,
0.8818359375,
0.57080078125,
0.8818359375,
0.58447265625,
0.5322265625,
0.58447265625
],
[
0.145263671875,
0.580078125,
0.45751953125,
0.580078125,
0.45751953125,
0.59326171875,
0.145263671875,
0.59326171875
],
[
0.54833984375,
0.5859375,
0.88330078125,
0.58740234375,
0.88330078125,
0.60107421875,
0.54833984375,
0.599609375
],
[
0.1435546875,
0.59423828125,
0.424072265625,
0.59423828125,
0.424072265625,
0.60791015625,
0.1435546875,
0.60791015625
],
[
0.54736328125,
0.6015625,
0.78271484375,
0.6015625,
0.78271484375,
0.615234375,
0.54736328125,
0.615234375
],
[
0.1175537109375,
0.62353515625,
0.1318359375,
0.62353515625,
0.1318359375,
0.6328125,
0.1175537109375,
0.6328125
],
[
0.1485595703125,
0.62060546875,
0.264404296875,
0.62158203125,
0.264404296875,
0.6357421875,
0.1485595703125,
0.63525390625
],
[
0.53076171875,
0.626953125,
0.880859375,
0.62744140625,
0.880859375,
0.642578125,
0.53076171875,
0.64111328125
],
[
0.118408203125,
0.646484375,
0.48779296875,
0.646484375,
0.48779296875,
0.6611328125,
0.118408203125,
0.6611328125
],
[
0.51318359375,
0.6435546875,
0.87744140625,
0.6435546875,
0.87744140625,
0.65673828125,
0.51318359375,
0.65673828125
],
[
0.51123046875,
0.6591796875,
0.87890625,
0.66015625,
0.87890625,
0.6748046875,
0.51123046875,
0.67333984375
],
[
0.12005615234375,
0.6650390625,
0.486083984375,
0.6650390625,
0.486083984375,
0.67578125,
0.12005615234375,
0.67578125
],
[
0.1175537109375,
0.6796875,
0.486083984375,
0.6796875,
0.486083984375,
0.693359375,
0.1175537109375,
0.693359375
],
[
0.51123046875,
0.6748046875,
0.87744140625,
0.6748046875,
0.87744140625,
0.689453125,
0.51123046875,
0.689453125
],
[
0.1175537109375,
0.69482421875,
0.486083984375,
0.69482421875,
0.486083984375,
0.7080078125,
0.1175537109375,
0.7080078125
],
[
0.509765625,
0.6923828125,
0.6884765625,
0.6904296875,
0.6884765625,
0.705078125,
0.509765625,
0.70703125
],
[
0.1175537109375,
0.71044921875,
0.483642578125,
0.71044921875,
0.483642578125,
0.7236328125,
0.1175537109375,
0.7236328125
],
[
0.5263671875,
0.72265625,
0.54248046875,
0.72265625,
0.54248046875,
0.72998046875,
0.5263671875,
0.72998046875
],
[
0.54248046875,
0.7197265625,
0.6708984375,
0.71875,
0.6708984375,
0.73291015625,
0.54248046875,
0.73388671875
],
[
0.118408203125,
0.72802734375,
0.486083984375,
0.72802734375,
0.486083984375,
0.7412109375,
0.118408203125,
0.7412109375
],
[
0.51708984375,
0.7236328125,
0.5263671875,
0.7236328125,
0.5263671875,
0.728515625,
0.51708984375,
0.728515625
],
[
0.118408203125,
0.74365234375,
0.484375,
0.74365234375,
0.484375,
0.75732421875,
0.118408203125,
0.75732421875
],
[
0.51708984375,
0.7451171875,
0.68994140625,
0.7451171875,
0.68994140625,
0.7587890625,
0.51708984375,
0.7587890625
],
[
0.6884765625,
0.75,
0.6962890625,
0.75,
0.6962890625,
0.75439453125,
0.6884765625,
0.75439453125
],
[
0.7001953125,
0.7451171875,
0.8818359375,
0.7451171875,
0.8818359375,
0.7587890625,
0.7001953125,
0.7587890625
],
[
0.118408203125,
0.76025390625,
0.486083984375,
0.76025390625,
0.486083984375,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.51318359375,
0.76123046875,
0.880859375,
0.76123046875,
0.880859375,
0.77490234375,
0.51318359375,
0.77490234375
],
[
0.1158447265625,
0.77490234375,
0.484375,
0.77490234375,
0.484375,
0.7880859375,
0.1158447265625,
0.7880859375
],
[
0.5146484375,
0.77783203125,
0.8818359375,
0.77783203125,
0.8818359375,
0.79150390625,
0.5146484375,
0.79150390625
],
[
0.118408203125,
0.7919921875,
0.486083984375,
0.7919921875,
0.486083984375,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.51123046875,
0.7919921875,
0.880859375,
0.7919921875,
0.880859375,
0.8056640625,
0.51123046875,
0.8056640625
],
[
0.118408203125,
0.80810546875,
0.484375,
0.80810546875,
0.484375,
0.822265625,
0.118408203125,
0.822265625
],
[
0.51123046875,
0.81005859375,
0.880859375,
0.81005859375,
0.880859375,
0.82373046875,
0.51123046875,
0.82373046875
],
[
0.118408203125,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.5341796875,
0.83203125,
0.87890625,
0.83203125,
0.87890625,
0.845703125,
0.5341796875,
0.845703125
],
[
0.1175537109375,
0.8388671875,
0.48779296875,
0.84033203125,
0.48779296875,
0.85498046875,
0.1175537109375,
0.85400390625
],
[
0.509765625,
0.8427734375,
0.880859375,
0.84423828125,
0.880859375,
0.85791015625,
0.509765625,
0.85546875
],
[
0.1175537109375,
0.85693359375,
0.486083984375,
0.85693359375,
0.486083984375,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.5341796875,
0.85546875,
0.8818359375,
0.85791015625,
0.8818359375,
0.87158203125,
0.5341796875,
0.869140625
],
[
0.1175537109375,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.509765625,
0.86767578125,
0.8818359375,
0.8681640625,
0.8818359375,
0.8828125,
0.509765625,
0.8818359375
],
[
0.51123046875,
0.8798828125,
0.63037109375,
0.8798828125,
0.63037109375,
0.8935546875,
0.51123046875,
0.8935546875
],
[
0.1175537109375,
0.88818359375,
0.484375,
0.88818359375,
0.484375,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.12005615234375,
0.90478515625,
0.285400390625,
0.90478515625,
0.285400390625,
0.91845703125,
0.12005615234375,
0.91845703125
],
[
0.5283203125,
0.90478515625,
0.71728515625,
0.90478515625,
0.71728515625,
0.91845703125,
0.5283203125,
0.91845703125
]
] |
[
"通过维度",
"和精度降低进行知识库索引压缩",
"Vilem Zouhar, Marius Mosbach, Miaoran Zhang and Dietrich Klakow",
"Department of Language Science & Technology, Saarland Informatics Campus,",
"Saarland University, Germany",
"( vzouhar,mmosbach,mzhang,dklakow } @lsv.uni-saarland.de",
"Abstract",
"至关重要的是，前面的所有示例都依赖于检",
"索组件和知识库的质量。知识库通常由密集向",
"Recently neural network based approaches to",
"量表示1索引，检索组件通常使用内积或L?距",
"knowledge-intensive NLP tasks, such as ques",
"离执行最大相似性搜索，以从知识库中检索文",
"tion answering, started to rely heavily on the",
"档2。仅索引就占用了大量的知识库，使得部",
"combinationofneuralretrieversandreaders",
"Retrieval is typically performed over a large",
"署和实验非常困难。检索速度还取决于索引向",
"textualknowledgebase(KB)whichrequires",
"量的维数。大型知识库的一个示例是Borgeauc",
"significant memory and compute resources,",
"etal.(2021)的工作，该工作通过1的数据库执",
"especially when scaled up.On HotpotQA",
"行检索。80亿份文件",
"we systematically investigate reducing the size",
"本文重点研究了通过降维和精度来压缩索引",
"of the KB index by means of dimensional-",
"的问题，并做出了以下贡献：",
"ity (sparse random projections, PCA, autoen-",
"coders) and numerical precision reduction.",
": Comparison of various unsupervised index",
"Our results show that PCA is an easy solution",
"compression methods for retrieval, including",
"that requires very little data and is only slightly",
"random projections, PCA, autoencoder, pre",
"worse than autoencoders, which are less sta-",
"ble.Allmethodsare sensitivetopre-andpost",
"cision reduction and their combination.",
"·Examination of effective pre- and post-",
"processinganddatashouldalwaysbecentered",
"and normalized both before and after dimen-",
"processing transformations, showing that cen-",
"sion reduction. Finally, we show that it is pos-",
"tering and normalization are necessary for",
"sible to combine PCA with using Ibit per di-",
"boosting the performance.",
"mension. Overall we achieve (1) 100 × com-",
": Analysis on the impact of adding irrelevant",
"pression with 75 % , and (2) 24 × compression",
"documents and retrieval errors. Recommen-",
"with 92 % original retrieval performance.",
"dations for use by practicioners.",
"1",
"Introduction",
"在Section3中，我们描述了问题场景和实验",
"知识密集型NLP任务的最新方法将基于神经",
"设置。我们讨论了Section4中不同压缩方法的",
"结果。我们在Section5中提供了进一步的分",
"网络的模型与利用密集向量表示（Guuetal.,",
"2020; Lewis et al., 2020; Petroni et al., 2021) 的",
"析，并在Section6中给出了使用建议。此项目",
"检索组件相结合。最简单的例子是问题回答，",
"的存储库是开源的。3",
"其中检索器接收一个问题作为输人，并返回",
"",
" Related Work",
"供阅读器（编码器和解码器）使用的相关文",
"",
"档，该文档输出答案（Chen，2020）。相同的方",
"Reducing index size.",
"",
"Luan et al. (2021) 对双",
"法也可以应用于其他上下文，例如事实检查",
"编码器上下文中信息检索的降维问题进行了全",
"(Tchechmedjievetal..2019）或知识对话（Dinan",
"面的概述。尽管深人并以形式论据为基础，但",
"etal.,2018）。此外，这种范式也可以应用于利",
"他们的研究重点是总体上降维的局限性和性质",
"用e的系统。。g。从训练语料库中缓存上下",
"（即使使用稀疏表示）以及文档长度对性能的",
"文以提供更好的输出，例如Khandelwal etal.",
"'Sparse representations via BM25 (Robertson et al.",
"(2019）提出的k最近邻语言模型或Yogatama",
"1995) are also commonly used but not the focus of this work.",
"etal.(2021)提出的动态门控语言模型机制。所",
"2We refer to the retrieved objects as documents though",
"有这些管道都被概括为从知识库(Zouharetal.",
"they commonly range from spans of text (e.g. 100 tokens) to",
"the full documents.",
"2021）中检索人工制品，在该知识库中，读者",
"与查询一起受到制约",
"github.com/zouharvi/kb-shrink"
] |
{
"arxiv_id": "2204.02906",
"language": "zh",
"page_idx": 1
}
|
||
2204.02906-zh-2
|
2204.02906-zh-2.jpg
|
的方法，并通过实验证据给出实用的建议
Izacard et al. (2020) 最近提出了降维的基线， 其中他们在通过添加低维线性投影层作为最终输出层来训练文档(和查询)编码器的同时执行降维。与我们的工作相比，他们的方法受到监督
在 Ma et al. (2021) 的并发工作中，还使用PCA 来减小文档索引的大小。与我们的工作相比，他们使用所有问题和文档向量的组合来执行 PCA。我们在 Figures 4 and 6 中表明这是不需要的，并且可以更有效地估计 PCA 变换矩阵。此外，我们使用不同的无监督压缩方法进行比较，并对我们的发现进行额外的分析
Yamada et al. (2021) 提出了一种解决内存成本问题的正交方法。他们提出的方法不是移动到另一个连续的向量表示，而是将原始向量映射到二进制值的向量，这些向量使用来自下游任务的信号进行训练。然而，管道仍然依赖于使用未压缩向量的重新排序。这种方法与我们的不同，在 Section 4.4 中，我们证明了它们可以组合在一起
最后，He et al.(2021)研究了 kNN 语言建模任务的过滤和 k 均值修剪。通过确定给定输入是否实际需要检索，这项工作还避免了必须始终执行大型数据存储的昂贵检索的问题
Effect of normalization. Timkey and van Schijndel (2021) 检查主导嵌入维度如何降低检索性能。他们研究了各个维度的贡献，发现当使用基于 BERT 的嵌入时，规范化是基于密集向量表示的文档检索的关键。与我们的工作相比，他们直接学习预先训练的 BERT,而我们专注于 DPR 3 Setup
3.1 Problem Statement and Evaluation
给定一个查询$q$ ,以下一组方程式总结了从大量文档$\mathcal{D}$中检索顶级$k$相关文档$Z=$ $\{d_{1},d_{2},\ldots,d_{k}\}$的概念进展，从而使$d$与$q$的相关性最大化。为此，查询和文档嵌入函数$f_Q:\mathcal{Q}\to\mathbb{R}^d$和$f_D:\mathcal{D}\to\mathbb{R}^d$用于将查询和所有文档映射到共享嵌人空间，相似性函数sim$:\mathbb{R}^d\times\mathbb{R}^d\to\mathbb{R}$近似查询和文档之间的相关性。在这里，我们将内积或$L^2$距离视为 sim 。$^{4}$最后，为了加快对大量文档的相似性计算并减少内存使用($f_D$通常是预先计算的),我们将降维函数$r_Q:\mathbb{R}^d\to\mathbb{R}^{d^{\prime}}$和$r_D:\mathbb{R}^d\to\mathbb{R}^{d^{\prime}}$
$^{4}$Cosine similarity could also be used but for computation reasons we skip it. Results are the same as for inner product and $L^2$ distance when the vectors are normalized.
影响。与他们的工作相反，本文旨在比较更多 分别应用于查询和文档嵌人。正式地，我们正
在解决以下问题：
$$\begin{aligned}Z=\underset{d\in\mathcal{D}}{\operatorname*{\arg top-k~rel.}}(q,d)\:,\mathrm{with}&\text{(1}\\\mathrm{rel.}(q,d)&\approx\sin(f_{\mathrm{Q}}(q),f_{\mathrm{D}}(d))&\text{(2}\\&\approx\sin(r_{\mathrm{Q}}(f_{\mathrm{Q}}(q)),r_{\mathrm{D}}(f_{\mathrm{D}}(d)))&\text{(3}\end{aligned}$$
(2) 中的近似值在实践中对于内积和$L^{2}$距离(Lin,2021)表现良好。在这种情况下，$f_Q$通常针对特定的下游任务进行微调。因此，在(3) 中希望函数$r_Q$和$r_D$是可微分的，以便它们可以传播信号。这些降维函数不需要相同，因为即使它们投影到共享向量空间，输入分布可能仍然不同。与查询和文档嵌入功能类似，可以对它们进行微调
Task Agnostic Representation. 当处理共享单个(大型)知识库的多个下游任务时，通常
只有$f_Q$针对特定任务进行微调，而$f_D$保持固定 (Lewis et al., 2020; Petroni et al., 2021)。这假设文档向量空间的组织在任务之间是足够的， 并且只需要训练查询到该空间的映射。$^5$因此， 这项工作的动机主要是找到一个好的$r_D$ (因为文档索引的主要大小),尽管我们注意到$r_Q$ 同样重要和必要，因为即使没有任何向量语义，密钥和文档嵌人必须具有相同的维度
$\textbf{R- Precision. }$ 为了评估检索性能，我们计算查询的 R-Precision 平均值：($Z$中的前$k$段落中的相关文档)$/ r$ , $k=$相关文档中的段落数， 与 Petroni et al.(2021)相同。在前面的工作之后，我们考虑内积(IP)和$L^2$距离作为相似度$\overline{\text{函数}}$
## 3.2 Data
作为知识库，我们使用英语维基百科中的文档， 并遵循 Petroni et al.(2021) 描述的设置。我们将跨度 (原始文章分为 100 个令牌，总计 5000万个)标记为与查询相关，如果它们来自与来源之一相同的维基百科文章。$^{6}$为了使我们的实验在计算上可行且易于重现，我们使用此知识库的修改版本进行实验，在该版本中，我们仅保留与下游任务的训练或验证集中的至少一个查询相关的文档跨度。作为下游任务，我们使用 HotpotQA (Yang et al., 2018) 进行所有主要实验，并使用自然问题(Kwiatkowski et al. 2019) 验证结果是否也传输到其他数据集。这导致 HotpotQA 的编码跨度超过 200 万个 (有
$^{5}$Guu et al.(2020) provide evidence that this assumption
can lead to worse results in some cases.
Spans of the original text which help in answering the
$query.$
|
[
[
0.118408203125,
0.0872802734375,
0.483642578125,
0.0872802734375,
0.483642578125,
0.097412109375,
0.118408203125,
0.097412109375
],
[
0.51318359375,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.51318359375,
0.1009521484375
],
[
0.118408203125,
0.10272216796875,
0.447509765625,
0.10272216796875,
0.447509765625,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.51318359375,
0.10272216796875,
0.64990234375,
0.10272216796875,
0.64990234375,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.1368408203125,
0.119384765625,
0.483642578125,
0.119384765625,
0.483642578125,
0.133056640625,
0.1368408203125,
0.133056640625
],
[
0.591796875,
0.133056640625,
0.82275390625,
0.133056640625,
0.82275390625,
0.146728515625,
0.591796875,
0.146728515625
],
[
0.85400390625,
0.133056640625,
0.8818359375,
0.133056640625,
0.8818359375,
0.146728515625,
0.85400390625,
0.146728515625
],
[
0.8310546875,
0.1383056640625,
0.8505859375,
0.1383056640625,
0.8505859375,
0.142578125,
0.8310546875,
0.142578125
],
[
0.118408203125,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.118408203125,
0.1484375
],
[
0.64990234375,
0.1455078125,
0.68115234375,
0.1455078125,
0.68115234375,
0.15673828125,
0.64990234375,
0.15673828125
],
[
0.118408203125,
0.1519775390625,
0.484375,
0.1519775390625,
0.484375,
0.1650390625,
0.118408203125,
0.1650390625
],
[
0.5400390625,
0.162109375,
0.765625,
0.162109375,
0.765625,
0.177001953125,
0.5400390625,
0.177001953125
],
[
0.8564453125,
0.157958984375,
0.884765625,
0.162109375,
0.880859375,
0.1787109375,
0.85205078125,
0.174560546875
],
[
0.12005615234375,
0.169189453125,
0.483642578125,
0.169189453125,
0.483642578125,
0.179931640625,
0.12005615234375,
0.179931640625
],
[
0.609375,
0.18115234375,
0.83447265625,
0.1817626953125,
0.83447265625,
0.196533203125,
0.609375,
0.1953125
],
[
0.85546875,
0.182861328125,
0.880859375,
0.182861328125,
0.880859375,
0.1953125,
0.85546875,
0.1953125
],
[
0.8505859375,
0.1859130859375,
0.8564453125,
0.1859130859375,
0.8564453125,
0.191162109375,
0.8505859375,
0.191162109375
],
[
0.118408203125,
0.185302734375,
0.1561279296875,
0.185302734375,
0.1561279296875,
0.196533203125,
0.118408203125,
0.196533203125
],
[
0.133544921875,
0.1995849609375,
0.484375,
0.20068359375,
0.484375,
0.215576171875,
0.133544921875,
0.21435546875
],
[
0.1175537109375,
0.21728515625,
0.484375,
0.21728515625,
0.484375,
0.23095703125,
0.1175537109375,
0.23095703125
],
[
0.5283203125,
0.2120361328125,
0.87890625,
0.2120361328125,
0.87890625,
0.2257080078125,
0.5283203125,
0.2257080078125
],
[
0.12005615234375,
0.2327880859375,
0.484375,
0.2327880859375,
0.484375,
0.2464599609375,
0.12005615234375,
0.2464599609375
],
[
0.51123046875,
0.2286376953125,
0.880859375,
0.2286376953125,
0.880859375,
0.2423095703125,
0.51123046875,
0.2423095703125
],
[
0.118408203125,
0.248779296875,
0.48779296875,
0.248779296875,
0.48779296875,
0.262451171875,
0.118408203125,
0.262451171875
],
[
0.51318359375,
0.2452392578125,
0.87744140625,
0.2452392578125,
0.87744140625,
0.255859375,
0.51318359375,
0.255859375
],
[
0.51123046875,
0.2587890625,
0.87744140625,
0.260009765625,
0.87744140625,
0.27490234375,
0.51123046875,
0.273681640625
],
[
0.118408203125,
0.265380859375,
0.48779296875,
0.265380859375,
0.48779296875,
0.279052734375,
0.118408203125,
0.279052734375
],
[
0.51123046875,
0.275634765625,
0.87890625,
0.275634765625,
0.87890625,
0.289306640625,
0.51123046875,
0.289306640625
],
[
0.118408203125,
0.281982421875,
0.484375,
0.281982421875,
0.484375,
0.292236328125,
0.118408203125,
0.292236328125
],
[
0.118408203125,
0.297607421875,
0.44921875,
0.297607421875,
0.44921875,
0.311279296875,
0.118408203125,
0.311279296875
],
[
0.5146484375,
0.29345703125,
0.87890625,
0.29345703125,
0.87890625,
0.30712890625,
0.5146484375,
0.30712890625
],
[
0.1351318359375,
0.312255859375,
0.48779296875,
0.3134765625,
0.48779296875,
0.327880859375,
0.1351318359375,
0.32666015625
],
[
0.51123046875,
0.30810546875,
0.87890625,
0.30810546875,
0.87890625,
0.32177734375,
0.51123046875,
0.32177734375
],
[
0.12005615234375,
0.330078125,
0.486083984375,
0.330078125,
0.486083984375,
0.34326171875,
0.12005615234375,
0.34326171875
],
[
0.51318359375,
0.325927734375,
0.64208984375,
0.325927734375,
0.64208984375,
0.339111328125,
0.51318359375,
0.339111328125
],
[
0.118408203125,
0.3466796875,
0.486083984375,
0.3466796875,
0.486083984375,
0.35986328125,
0.118408203125,
0.35986328125
],
[
0.51708984375,
0.350830078125,
0.7900390625,
0.350830078125,
0.7900390625,
0.364013671875,
0.51708984375,
0.364013671875
],
[
0.78271484375,
0.352783203125,
0.880859375,
0.352783203125,
0.880859375,
0.364013671875,
0.78271484375,
0.364013671875
],
[
0.12005615234375,
0.3623046875,
0.486083984375,
0.3623046875,
0.486083984375,
0.3759765625,
0.12005615234375,
0.3759765625
],
[
0.51123046875,
0.366455078125,
0.87744140625,
0.366455078125,
0.87744140625,
0.380126953125,
0.51123046875,
0.380126953125
],
[
0.118408203125,
0.377685546875,
0.486083984375,
0.377685546875,
0.486083984375,
0.391357421875,
0.118408203125,
0.391357421875
],
[
0.51318359375,
0.3818359375,
0.87890625,
0.383056640625,
0.87890625,
0.396728515625,
0.51318359375,
0.3955078125
],
[
0.118408203125,
0.394287109375,
0.48779296875,
0.394287109375,
0.48779296875,
0.407958984375,
0.118408203125,
0.407958984375
],
[
0.51123046875,
0.40087890625,
0.87744140625,
0.40087890625,
0.87744140625,
0.410888671875,
0.51123046875,
0.410888671875
],
[
0.118408203125,
0.410400390625,
0.484375,
0.410400390625,
0.484375,
0.42333984375,
0.118408203125,
0.42333984375
],
[
0.51318359375,
0.4150390625,
0.880859375,
0.4150390625,
0.880859375,
0.4287109375,
0.51318359375,
0.4287109375
],
[
0.12005615234375,
0.427490234375,
0.2275390625,
0.427490234375,
0.2275390625,
0.439453125,
0.12005615234375,
0.439453125
],
[
0.51318359375,
0.43115234375,
0.880859375,
0.43115234375,
0.880859375,
0.44482421875,
0.51318359375,
0.44482421875
],
[
0.1368408203125,
0.443603515625,
0.48779296875,
0.443603515625,
0.48779296875,
0.457275390625,
0.1368408203125,
0.457275390625
],
[
0.509765625,
0.44775390625,
0.87890625,
0.44775390625,
0.87890625,
0.46142578125,
0.509765625,
0.46142578125
],
[
0.1175537109375,
0.457763671875,
0.486083984375,
0.458984375,
0.486083984375,
0.473876953125,
0.1175537109375,
0.47265625
],
[
0.509765625,
0.460205078125,
0.880859375,
0.4619140625,
0.880859375,
0.479736328125,
0.509765625,
0.47802734375
],
[
0.1175537109375,
0.474365234375,
0.486083984375,
0.474365234375,
0.486083984375,
0.488037109375,
0.1175537109375,
0.488037109375
],
[
0.5146484375,
0.478515625,
0.880859375,
0.478515625,
0.880859375,
0.4921875,
0.5146484375,
0.4921875
],
[
0.118408203125,
0.4912109375,
0.4306640625,
0.4912109375,
0.4306640625,
0.5048828125,
0.118408203125,
0.5048828125
],
[
0.515625,
0.49755859375,
0.84228515625,
0.49755859375,
0.84228515625,
0.50830078125,
0.515625,
0.50830078125
],
[
0.12176513671875,
0.5185546875,
0.48779296875,
0.5185546875,
0.48779296875,
0.5322265625,
0.12176513671875,
0.5322265625
],
[
0.51708984375,
0.521484375,
0.87890625,
0.521484375,
0.87890625,
0.53515625,
0.51708984375,
0.53515625
],
[
0.118408203125,
0.5361328125,
0.484375,
0.5361328125,
0.484375,
0.54638671875,
0.118408203125,
0.54638671875
],
[
0.51318359375,
0.53759765625,
0.880859375,
0.53759765625,
0.880859375,
0.55029296875,
0.51318359375,
0.55029296875
],
[
0.118408203125,
0.55029296875,
0.486083984375,
0.55029296875,
0.486083984375,
0.56396484375,
0.118408203125,
0.56396484375
],
[
0.5146484375,
0.55419921875,
0.880859375,
0.55419921875,
0.880859375,
0.56787109375,
0.5146484375,
0.56787109375
],
[
0.12176513671875,
0.568359375,
0.483642578125,
0.568359375,
0.483642578125,
0.5791015625,
0.12176513671875,
0.5791015625
],
[
0.51318359375,
0.57080078125,
0.87890625,
0.57080078125,
0.87890625,
0.58447265625,
0.51318359375,
0.58447265625
],
[
0.12005615234375,
0.5830078125,
0.484375,
0.5830078125,
0.484375,
0.5966796875,
0.12005615234375,
0.5966796875
],
[
0.51123046875,
0.5849609375,
0.87744140625,
0.5849609375,
0.87744140625,
0.5986328125,
0.51123046875,
0.5986328125
],
[
0.118408203125,
0.599609375,
0.48779296875,
0.599609375,
0.48779296875,
0.61328125,
0.118408203125,
0.61328125
],
[
0.51123046875,
0.6025390625,
0.55322265625,
0.6025390625,
0.55322265625,
0.6162109375,
0.51123046875,
0.6162109375
],
[
0.12176513671875,
0.6162109375,
0.214111328125,
0.6162109375,
0.214111328125,
0.62744140625,
0.12176513671875,
0.62744140625
],
[
0.51123046875,
0.6298828125,
0.599609375,
0.6298828125,
0.599609375,
0.64453125,
0.51123046875,
0.64453125
],
[
0.11419677734375,
0.642578125,
0.2032470703125,
0.6455078125,
0.202392578125,
0.6611328125,
0.11248779296875,
0.65771484375
],
[
0.51318359375,
0.65087890625,
0.87890625,
0.65087890625,
0.87890625,
0.66455078125,
0.51318359375,
0.66455078125
],
[
0.51318359375,
0.66748046875,
0.880859375,
0.66748046875,
0.880859375,
0.68115234375,
0.51318359375,
0.68115234375
],
[
0.118408203125,
0.67333984375,
0.438232421875,
0.67333984375,
0.438232421875,
0.68408203125,
0.118408203125,
0.68408203125
],
[
0.5146484375,
0.68408203125,
0.880859375,
0.68408203125,
0.880859375,
0.69775390625,
0.5146484375,
0.69775390625
],
[
0.118408203125,
0.69482421875,
0.484375,
0.69482421875,
0.484375,
0.7080078125,
0.118408203125,
0.7080078125
],
[
0.51318359375,
0.69970703125,
0.880859375,
0.69970703125,
0.880859375,
0.71337890625,
0.51318359375,
0.71337890625
],
[
0.118408203125,
0.71044921875,
0.466796875,
0.71044921875,
0.466796875,
0.7236328125,
0.118408203125,
0.7236328125
],
[
0.468505859375,
0.71435546875,
0.478515625,
0.71435546875,
0.478515625,
0.7197265625,
0.468505859375,
0.7197265625
],
[
0.51318359375,
0.71533203125,
0.87890625,
0.71533203125,
0.87890625,
0.728515625,
0.51318359375,
0.728515625
],
[
0.118408203125,
0.72705078125,
0.486083984375,
0.72705078125,
0.486083984375,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.51123046875,
0.732421875,
0.880859375,
0.732421875,
0.880859375,
0.7451171875,
0.51123046875,
0.7451171875
],
[
0.118408203125,
0.74365234375,
0.483642578125,
0.74365234375,
0.483642578125,
0.75439453125,
0.118408203125,
0.75439453125
],
[
0.51318359375,
0.7490234375,
0.880859375,
0.7490234375,
0.880859375,
0.7626953125,
0.51318359375,
0.7626953125
],
[
0.118408203125,
0.75732421875,
0.484375,
0.75732421875,
0.484375,
0.771484375,
0.118408203125,
0.771484375
],
[
0.51318359375,
0.76416015625,
0.869140625,
0.76416015625,
0.869140625,
0.77783203125,
0.51318359375,
0.77783203125
],
[
0.118408203125,
0.77490234375,
0.484375,
0.77490234375,
0.484375,
0.7880859375,
0.118408203125,
0.7880859375
],
[
0.509765625,
0.779296875,
0.87890625,
0.779296875,
0.87890625,
0.7919921875,
0.509765625,
0.7919921875
],
[
0.1175537109375,
0.7890625,
0.483642578125,
0.7890625,
0.483642578125,
0.802734375,
0.1175537109375,
0.802734375
],
[
0.51318359375,
0.79638671875,
0.8818359375,
0.79638671875,
0.8818359375,
0.81005859375,
0.51318359375,
0.81005859375
],
[
0.1175537109375,
0.80712890625,
0.483642578125,
0.80712890625,
0.483642578125,
0.82080078125,
0.1175537109375,
0.82080078125
],
[
0.51123046875,
0.8125,
0.87890625,
0.8125,
0.87890625,
0.826171875,
0.51123046875,
0.826171875
],
[
0.133544921875,
0.82470703125,
0.48193359375,
0.82470703125,
0.48193359375,
0.8349609375,
0.133544921875,
0.8349609375
],
[
0.51318359375,
0.82763671875,
0.880859375,
0.82763671875,
0.880859375,
0.84130859375,
0.51318359375,
0.84130859375
],
[
0.1175537109375,
0.8388671875,
0.484375,
0.8388671875,
0.484375,
0.8525390625,
0.1175537109375,
0.8525390625
],
[
0.51318359375,
0.84423828125,
0.880859375,
0.84423828125,
0.880859375,
0.85791015625,
0.51318359375,
0.85791015625
],
[
0.1175537109375,
0.85498046875,
0.483642578125,
0.8525390625,
0.483642578125,
0.86767578125,
0.1175537109375,
0.869140625
],
[
0.53076171875,
0.8662109375,
0.8818359375,
0.869140625,
0.8818359375,
0.88427734375,
0.53076171875,
0.88134765625
],
[
0.1351318359375,
0.8798828125,
0.48779296875,
0.8818359375,
0.48779296875,
0.8955078125,
0.1351318359375,
0.8935546875
],
[
0.51318359375,
0.8818359375,
0.75,
0.8818359375,
0.75,
0.892578125,
0.51318359375,
0.892578125
],
[
0.1175537109375,
0.89453125,
0.486083984375,
0.8955078125,
0.486083984375,
0.90673828125,
0.1175537109375,
0.90625
],
[
0.53564453125,
0.89453125,
0.8818359375,
0.89453125,
0.8818359375,
0.90771484375,
0.53564453125,
0.90771484375
],
[
0.1175537109375,
0.90380859375,
0.4189453125,
0.90625,
0.4189453125,
0.91943359375,
0.1175537109375,
0.91748046875
],
[
0.5146484375,
0.91015625,
0.54736328125,
0.91015625,
0.54736328125,
0.91845703125,
0.5146484375,
0.91845703125
]
] |
[
"影响。与他们的工作相反，本文旨在比较更多",
"分别应用于查询和文档嵌入。正式地，我们正",
"的方法，并通过实验证据给出实用的建议",
"在解决以下问题：",
"Izacard etal.(2020)最近提出了降维的基线",
"Z = arg top-k rel.(q, d) , with",
"(1)",
"",
"其中他们在通过添加低维线性投影层作为最终",
"deD",
"输出层来训练文档（和查询）编码器的同时执",
"rel.(q, d) ~ sim(fo(q), fp(d))",
"(2)",
"行降维。与我们的工作相比，他们的方法受到",
"~ sim(rQ(fo(q)), rp(fp(d))",
"(3)",
"",
"监督",
"在Maetal.（2021)的并发工作中，还使用",
"PCA来减小文档索引的大小。与我们的工作相",
"（2）中的近似值在实践中对于内积和L2距",
"比，他们使用所有问题和文档向量的组合来执",
"离（Lin，2021)表现良好。在这种情况下，fQ通",
"行PCA。我们在Figures4and6中表明这是不",
"常针对特定的下游任务进行微调。因此，在（3",
"中希望函数rQ和rD是可微分的，以便它们可",
"需要的，并且可以更有效地估计PCA变换矩",
"以传播信号。这些降维函数不需要相同，因为",
"阵。此外，我们使用不同的无监督压缩方法进",
"行比较，并对我们的发现进行额外的分析",
"即使它们投影到共享向量空间，输入分布可能",
"Yamadaetal.（2021）提出了一种解决内存成",
"仍然不同。与查询和文档嵌入功能类似，可以",
"本问题的正交方法。他们提出的方法不是移动",
"对它们进行微调",
"到另一个连续的向量表示，而是将原始向量映",
"Task Agnostic Representation.",
"当处理共享",
"射到二进制值的向量，这些向量使用来自下游",
"单个（大型）知识库的多个下游任务时，通常",
"任务的信号进行训练。然而，管道仍然依赖于",
"只有fo针对特定任务进行微调，而fD保持固",
"使用未压缩向量的重新排序。这种方法与我们",
"定（Lewisetal.,2020:Petronietal.,2021）。这假",
"的不同，在Section4.4中，我们证明了它们可",
"设文档向量空间的组织在任务之间是足够的，",
"以组合在一起",
"并且只需要训练查询到该空间的映射。5因此，",
"最后，Heetal.（2021)研究了kNN语言建模",
"这项工作的动机主要是找到一个好的rD（因",
"任务的过滤和k均值修剪。通过确定给定输入",
"为文档索引的主要大小），尽管我们注意到rQ",
"是否实际需要检索，这项工作还避免了必须始",
"同样重要和必要，因为即使没有任何向量语",
"终执行大型数据存储的昂贵检索的问题",
"义，密钥和文档嵌入必须具有相同的维度",
"Effect of normalization.Timkey and van Schi-",
"R-Precision。为了评估检索性能，我们计算",
"indel（2021）检香查主导嵌入维度如何降低检索",
"查询的R-Precision平均值：（Z中的前k段落",
"性能。他们研究了各个维度的贡献，发现当使",
"中的相关文档）/r，k=相关文档中的段落数，",
"用基于BERT的嵌入时，规范化是基于密集向",
"与Petronietal.（2021）相同。在前面的工作之",
"量表示的文档检索的关键。与我们的工作相",
"后，我们考虑内积(IP)和L2距离作为相似度",
"比，他们直接学习预先训练的BERT，而我们",
"函数",
"专注于DPR",
"3.2Data",
"3Setup",
"作为知识库，我们使用英语维基百科中的文档，",
"并遵循Petronietal.(2021）描述的设置。我们",
"3.1ProblemStatementandEvaluation",
"将跨度（原始文章分为100个令牌，总计5000",
"给定一个查询9，以下一组方程式总结了",
"万个）标记为与查询相关，如果它们来自与来",
"从大量文档D中检索顶级k相关文档Z",
"",
"源之一相同的维基百科文章。6为了使我们的",
"[di,d2,.,d}的概念进展，从而使d与q的",
"实验在计算上可行且易于重现，我们使用此知",
"相关性最大化。为此，查询和文档嵌入函数",
"识库的修改版本进行实验，在该版本中，我们",
"fo：Q→Rd和fD：D→Rd用于将查询和",
"仅保留与下游任务的训练或验证集中的至少-",
"所有文档映射到共享嵌人空间，相似性函数",
"个查询相关的文档跨度。作为下游任务，我们",
"sim：Rd×Rd→R近似查询和文档之间的相",
"使用HotpotQA（Yangetal.，2018）进行所有主",
"关性。在这里，我们将内积或L²距离视为sim",
"要实验，并使用自然问题（Kwiatkowskietal.",
"4最后，为了加快对大量文档的相似性计算",
"2019)验证结果是否也传输到其他数据集。这",
"并减少内存使用（fD通常是预先计算的），我",
"导致HotpotQA的编码跨度超过200万个（有",
"们将降维函数rQ：Rd→Rd和rD：Rd→Rd",
"°Guu et al. (2020) provide evidence that this assumption",
"4Cosine similarity could also be used but for computation",
"can lead to worse results in some cases",
"reasons we skip it.Results are the same as for inner product",
"6Spans of the original text which help in answering the",
"and L? distance when the vectors are normalized.",
"query"
] |
{
"arxiv_id": "2204.02906",
"language": "zh",
"page_idx": 2
}
|
||
2204.02906-zh-3
|
2204.02906-zh-3.jpg
|
<FigureHere>
Figure 1: Comparison of different BERT-based embedding models and versions when using faster but slightly inaccurate nearest neighbour search.[CLS] is the specific token embedding from the last layer while (Avg) is all token average.
关数据集大小，请参阅 Table 6)。此数据集(查询和文档)的 768 维嵌入(32 位浮点)加起来为 7GB (整个未修剪的数据集为 146GB)
3.3 Uncompressed Retrieval Peformance
为了建立未压缩性能的基线，我们使用基于BERT (Devlin et al., 2019) 的模型。我们考虑(1)香草 BERT, (2) SentenceBERT (Reimers and Gurevych,2019)和(3)DPR(Karpukhin et al., 2020),它经过专门的文档检索培训。为了获得文档嵌入，我们使用[CLS] 中的最后一个隐藏状态表示形式或最后一层的标记之间的平均值
我们的第一个实验比较了 HotpotQA 上不同模型的检索性能。结果显示在 Figure 1。与以前的作品 (Reimers and Gurevych, 2019)一致， 一个立即明显的结论是香草 BERT 的性能很差，特别是在为[CLS]令牌采用隐藏状态表示时。接下来，为了使计算易于处理，我们使用 FAISS (Johnson et al., 2019) 重复实验。$^{7}$我们发现模型之间的性能损失是系统性的，这保证了使用这种近似最近邻搜索进行比较，并且我们以下所有实验都将在 DPR-CLS 模型上使用 FAISS
Pre-processing Transformations. Figure 1 还表明，模型性能，尤其是对于DPR，在很大程度上取决于用于检索的相似性度量。这是因为没有一个模型在默认情况下产生归一化向量
Figure 2 表明，仅执行归一化$(\frac x{||x||}$)有时会损害性能，但是当预先加入居中($\frac{x-\bar{x}}{||x-\bar{x}||})时$,
$^{7}$IndexIVFFlat, nlist=200, nprobe=100
<FigureHere>
Figure 2: Effect of data centering and normalization on
performance (evaluated with FAISS).
它在所有情况下都会改善结果(与没有预处理相比)。对查询和文档分别进行规范化和居中。此外，如果向量是归一化的，那么对于$L^2$和内积，检索到的文档是相同的。$^{8}$
尽管如此，我们认为单独研究$L^2$的压缩能力和内积仍然是有意义的，因为归一化向量压缩的输出不需要归一化
### 4 Compression Methods
在建立了未压缩基线的检索性能之后，我们现
在转向压缩密集文档索引和查询的方法
请注意，我们考虑在已经训练的索引上使用无监督方法，以最大程度地简化使用和适用性。这与监督方法相反，监督方法可以访问查询-文档相关性映射，或者访问训练中的降维(i。e。较低的最终层尺寸)
### 4.1 Random Projection
对给定索引$x\in\mathbb{R}^d$执行降维的最简单方法是仅随机保留某些$d^{\prime}$维度并删除所有其他维度：
$$f_{\mathrm{drop.}}(\boldsymbol{x})=(x_{m_1},x_{m_2},\ldots,x_{m_{d'}})$$
另一种方法是贪婪地搜索要删除的维度(那些在省略时可以提高性能或减少性能的维度):
$$\begin{aligned}&p_{i}(\boldsymbol{x})=(x_{0},x_{1},\ldots,x_{i-1},x_{i+1},\ldots,x_{768})\\&\mathcal{L}_{i}=\mathrm{R-Prec}(p_{i}(Q),p_{i}(D))\\&m=\mathrm{sort}_{\mathcal{L}}^{\mathrm{desc.}}([1\ldots768])\\&f_{\mathrm{greedy~drop.}}(\boldsymbol{x})=(x_{m_{1}},x_{m_{2}},\ldots,x_{m_{d^{\prime}}})\\&^{8}\arg\max_{k}-||\boldsymbol{a}-\boldsymbol{b}||^{2}=\arg\max_{k}-\langle\boldsymbol{a},\boldsymbol{a}\rangle^{2}-\langle\boldsymbol{b},\boldsymbol{b}\rangle^{2}+\\&(a,b)=a^{\mathrm{arrmax}}-(a,b)\end{aligned}$$
$$^{8}\arg\operatorname*{max}_{k}-||\boldsymbol{a}-\boldsymbol{b}||^{2}=\arg\operatorname*{max}_{k}-\langle\boldsymbol{a},\boldsymbol{a}\rangle^{2}-\langle\boldsymbol{b},\boldsymbol{b}\rangle^{2}+\\2\cdot\langle\boldsymbol{a},\boldsymbol{b}\rangle=\arg\operatorname*{max}_{k}\:2\cdot\langle\boldsymbol{a},\boldsymbol{b}\rangle-2=\arg\operatorname*{max}_{k}\:\langle\boldsymbol{a},\boldsymbol{b}\rangle $$
|
[
[
0.19482421875,
0.09442138671875,
0.2275390625,
0.09442138671875,
0.2275390625,
0.1015625,
0.19482421875,
0.1015625
],
[
0.2376708984375,
0.09442138671875,
0.310546875,
0.09442138671875,
0.310546875,
0.1015625,
0.2376708984375,
0.1015625
],
[
0.391357421875,
0.09442138671875,
0.4609375,
0.09442138671875,
0.4609375,
0.1015625,
0.391357421875,
0.1015625
],
[
0.59423828125,
0.09442138671875,
0.61865234375,
0.09442138671875,
0.61865234375,
0.10272216796875,
0.59423828125,
0.10272216796875
],
[
0.71533203125,
0.09320068359375,
0.76318359375,
0.09320068359375,
0.76318359375,
0.10394287109375,
0.71533203125,
0.10394287109375
],
[
0.57666015625,
0.097412109375,
0.59423828125,
0.097412109375,
0.59423828125,
0.1009521484375,
0.57666015625,
0.1009521484375
],
[
0.337646484375,
0.0968017578125,
0.37451171875,
0.0968017578125,
0.37451171875,
0.1009521484375,
0.337646484375,
0.1009521484375
],
[
0.57080078125,
0.10394287109375,
0.66162109375,
0.1051025390625,
0.66162109375,
0.11639404296875,
0.57080078125,
0.11517333984375
],
[
0.71533203125,
0.1051025390625,
0.80615234375,
0.1051025390625,
0.80615234375,
0.11639404296875,
0.71533203125,
0.11639404296875
],
[
0.145263671875,
0.1181640625,
0.16796875,
0.1181640625,
0.16796875,
0.12890625,
0.145263671875,
0.12890625
],
[
0.57080078125,
0.11639404296875,
0.6767578125,
0.1175537109375,
0.6767578125,
0.12890625,
0.57080078125,
0.127685546875
],
[
0.71533203125,
0.1181640625,
0.85791015625,
0.1181640625,
0.85791015625,
0.12890625,
0.71533203125,
0.12890625
],
[
0.3359375,
0.127685546875,
0.340087890625,
0.127685546875,
0.340087890625,
0.1341552734375,
0.3359375,
0.1341552734375
],
[
0.1435546875,
0.13720703125,
0.166259765625,
0.13720703125,
0.166259765625,
0.149658203125,
0.1435546875,
0.149658203125
],
[
0.5400390625,
0.1455078125,
0.56103515625,
0.1455078125,
0.56103515625,
0.15380859375,
0.5400390625,
0.15380859375
],
[
0.130126953125,
0.155029296875,
0.16455078125,
0.155029296875,
0.16455078125,
0.1728515625,
0.130126953125,
0.1728515625
],
[
0.12176513671875,
0.162109375,
0.172119140625,
0.162109375,
0.172119140625,
0.204833984375,
0.12176513671875,
0.204833984375
],
[
0.54150390625,
0.166259765625,
0.55908203125,
0.166259765625,
0.55908203125,
0.17578125,
0.54150390625,
0.17578125
],
[
0.1435546875,
0.196533203125,
0.16455078125,
0.196533203125,
0.16455078125,
0.210205078125,
0.1435546875,
0.210205078125
],
[
0.52880859375,
0.196533203125,
0.53564453125,
0.196533203125,
0.53564453125,
0.207275390625,
0.52880859375,
0.207275390625
],
[
0.52880859375,
0.207275390625,
0.53564453125,
0.207275390625,
0.53564453125,
0.2161865234375,
0.52880859375,
0.2161865234375
],
[
0.1435546875,
0.21728515625,
0.162841796875,
0.21728515625,
0.162841796875,
0.23095703125,
0.1435546875,
0.23095703125
],
[
0.52880859375,
0.2197265625,
0.53564453125,
0.2197265625,
0.53564453125,
0.226806640625,
0.52880859375,
0.226806640625
],
[
0.5439453125,
0.2327880859375,
0.5517578125,
0.2327880859375,
0.5517578125,
0.2369384765625,
0.5439453125,
0.2369384765625
],
[
0.189697265625,
0.2476806640625,
0.2509765625,
0.2476806640625,
0.2509765625,
0.25830078125,
0.189697265625,
0.25830078125
],
[
0.2509765625,
0.2476806640625,
0.31494140625,
0.2476806640625,
0.31494140625,
0.25830078125,
0.2509765625,
0.25830078125
],
[
0.33154296875,
0.248779296875,
0.359375,
0.248779296875,
0.359375,
0.257080078125,
0.33154296875,
0.257080078125
],
[
0.355224609375,
0.2476806640625,
0.455078125,
0.2464599609375,
0.455078125,
0.257080078125,
0.355224609375,
0.25830078125
],
[
0.185546875,
0.257080078125,
0.2216796875,
0.257080078125,
0.2216796875,
0.268310546875,
0.185546875,
0.268310546875
],
[
0.424072265625,
0.25830078125,
0.453369140625,
0.25830078125,
0.453369140625,
0.2666015625,
0.424072265625,
0.2666015625
],
[
0.5439453125,
0.254638671875,
0.55517578125,
0.254638671875,
0.55517578125,
0.262451171875,
0.5439453125,
0.262451171875
],
[
0.27783203125,
0.2587890625,
0.312255859375,
0.2587890625,
0.312255859375,
0.2666015625,
0.27783203125,
0.2666015625
],
[
0.225830078125,
0.2587890625,
0.267822265625,
0.2587890625,
0.267822265625,
0.26318359375,
0.225830078125,
0.26318359375
],
[
0.330078125,
0.2587890625,
0.358642578125,
0.2587890625,
0.358642578125,
0.2666015625,
0.330078125,
0.2666015625
],
[
0.372802734375,
0.2587890625,
0.408935546875,
0.2587890625,
0.408935546875,
0.26318359375,
0.372802734375,
0.26318359375
],
[
0.231689453125,
0.265380859375,
0.267822265625,
0.265380859375,
0.267822265625,
0.276611328125,
0.231689453125,
0.276611328125
],
[
0.3779296875,
0.2666015625,
0.4072265625,
0.2666015625,
0.4072265625,
0.27490234375,
0.3779296875,
0.27490234375
],
[
0.5859375,
0.281982421875,
0.6171875,
0.281982421875,
0.6171875,
0.29345703125,
0.5859375,
0.29345703125
],
[
0.615234375,
0.283935546875,
0.6708984375,
0.283935546875,
0.6708984375,
0.29150390625,
0.615234375,
0.29150390625
],
[
0.66650390625,
0.283203125,
0.70947265625,
0.283203125,
0.70947265625,
0.29345703125,
0.66650390625,
0.29345703125
],
[
0.7255859375,
0.283935546875,
0.75390625,
0.283935546875,
0.75390625,
0.292236328125,
0.7255859375,
0.292236328125
],
[
0.75244140625,
0.283935546875,
0.8095703125,
0.283935546875,
0.8095703125,
0.29150390625,
0.75244140625,
0.29150390625
],
[
0.80615234375,
0.283935546875,
0.8427734375,
0.283935546875,
0.8427734375,
0.29150390625,
0.80615234375,
0.29150390625
],
[
0.58447265625,
0.292236328125,
0.615234375,
0.292236328125,
0.615234375,
0.303955078125,
0.58447265625,
0.303955078125
],
[
0.62744140625,
0.290283203125,
0.6630859375,
0.290283203125,
0.6630859375,
0.3017578125,
0.62744140625,
0.3017578125
],
[
0.6767578125,
0.29345703125,
0.70947265625,
0.29345703125,
0.70947265625,
0.3017578125,
0.6767578125,
0.3017578125
],
[
0.7255859375,
0.29345703125,
0.75390625,
0.29345703125,
0.75390625,
0.3017578125,
0.7255859375,
0.3017578125
],
[
0.77099609375,
0.292236328125,
0.8017578125,
0.292236328125,
0.8017578125,
0.2998046875,
0.77099609375,
0.2998046875
],
[
0.81689453125,
0.29345703125,
0.8447265625,
0.29345703125,
0.8447265625,
0.3017578125,
0.81689453125,
0.3017578125
],
[
0.77099609375,
0.2998046875,
0.8017578125,
0.2998046875,
0.8017578125,
0.311279296875,
0.77099609375,
0.311279296875
],
[
0.62890625,
0.300537109375,
0.66162109375,
0.300537109375,
0.66162109375,
0.312255859375,
0.62890625,
0.312255859375
],
[
0.1175537109375,
0.295654296875,
0.48779296875,
0.295654296875,
0.48779296875,
0.308837890625,
0.1175537109375,
0.308837890625
],
[
0.1175537109375,
0.308837890625,
0.483642578125,
0.308837890625,
0.483642578125,
0.322509765625,
0.1175537109375,
0.322509765625
],
[
0.1158447265625,
0.32373046875,
0.484375,
0.32470703125,
0.484375,
0.33837890625,
0.1158447265625,
0.33740234375
],
[
0.51123046875,
0.330810546875,
0.8818359375,
0.330810546875,
0.8818359375,
0.344482421875,
0.51123046875,
0.344482421875
],
[
0.118408203125,
0.33837890625,
0.486083984375,
0.33837890625,
0.486083984375,
0.3515625,
0.118408203125,
0.3515625
],
[
0.11419677734375,
0.350830078125,
0.252685546875,
0.3515625,
0.252685546875,
0.366455078125,
0.11419677734375,
0.365234375
],
[
0.51318359375,
0.3466796875,
0.76171875,
0.345703125,
0.76171875,
0.35693359375,
0.51318359375,
0.358154296875
],
[
0.51318359375,
0.385498046875,
0.87890625,
0.385498046875,
0.87890625,
0.3984375,
0.51318359375,
0.3984375
],
[
0.118408203125,
0.392578125,
0.484375,
0.392578125,
0.484375,
0.40625,
0.118408203125,
0.40625
],
[
0.5146484375,
0.402099609375,
0.880859375,
0.402099609375,
0.880859375,
0.4150390625,
0.5146484375,
0.4150390625
],
[
0.118408203125,
0.410400390625,
0.484375,
0.410400390625,
0.484375,
0.42333984375,
0.118408203125,
0.42333984375
],
[
0.51318359375,
0.41748046875,
0.880859375,
0.416259765625,
0.880859375,
0.43115234375,
0.51318359375,
0.431640625
],
[
0.1175537109375,
0.42578125,
0.48193359375,
0.42578125,
0.48193359375,
0.439453125,
0.1175537109375,
0.439453125
],
[
0.5146484375,
0.43408203125,
0.771484375,
0.43408203125,
0.771484375,
0.44775390625,
0.5146484375,
0.44775390625
],
[
0.1158447265625,
0.45361328125,
0.453369140625,
0.45361328125,
0.453369140625,
0.46728515625,
0.1158447265625,
0.46728515625
],
[
0.5322265625,
0.45068359375,
0.880859375,
0.45068359375,
0.880859375,
0.46435546875,
0.5322265625,
0.46435546875
],
[
0.5146484375,
0.46728515625,
0.880859375,
0.46728515625,
0.880859375,
0.48095703125,
0.5146484375,
0.48095703125
],
[
0.118408203125,
0.4755859375,
0.484375,
0.4755859375,
0.484375,
0.4892578125,
0.118408203125,
0.4892578125
],
[
0.51318359375,
0.482666015625,
0.69775390625,
0.482666015625,
0.69775390625,
0.496337890625,
0.51318359375,
0.496337890625
],
[
0.1175537109375,
0.4921875,
0.486083984375,
0.4921875,
0.486083984375,
0.505859375,
0.1175537109375,
0.505859375
],
[
0.1175537109375,
0.50732421875,
0.486083984375,
0.50732421875,
0.486083984375,
0.521484375,
0.1175537109375,
0.521484375
],
[
0.51708984375,
0.51416015625,
0.52294921875,
0.51416015625,
0.52294921875,
0.52099609375,
0.51708984375,
0.52099609375
],
[
0.5458984375,
0.5126953125,
0.7421875,
0.5126953125,
0.7421875,
0.525390625,
0.5458984375,
0.525390625
],
[
0.1158447265625,
0.52392578125,
0.483642578125,
0.52392578125,
0.483642578125,
0.53759765625,
0.1158447265625,
0.53759765625
],
[
0.1158447265625,
0.5380859375,
0.48779296875,
0.5390625,
0.48779296875,
0.55419921875,
0.1158447265625,
0.552734375
],
[
0.51318359375,
0.5380859375,
0.8818359375,
0.5380859375,
0.8818359375,
0.5517578125,
0.51318359375,
0.5517578125
],
[
0.1175537109375,
0.5546875,
0.486083984375,
0.5546875,
0.486083984375,
0.568359375,
0.1175537109375,
0.568359375
],
[
0.51318359375,
0.5546875,
0.8251953125,
0.5546875,
0.8251953125,
0.568359375,
0.51318359375,
0.568359375
],
[
0.118408203125,
0.57177734375,
0.486083984375,
0.57177734375,
0.486083984375,
0.5859375,
0.118408203125,
0.5859375
],
[
0.53076171875,
0.57080078125,
0.87890625,
0.57080078125,
0.87890625,
0.58447265625,
0.53076171875,
0.58447265625
],
[
0.118408203125,
0.5888671875,
0.1561279296875,
0.5888671875,
0.1561279296875,
0.6015625,
0.118408203125,
0.6015625
],
[
0.5146484375,
0.5859375,
0.87890625,
0.5859375,
0.87890625,
0.599609375,
0.5146484375,
0.599609375
],
[
0.1351318359375,
0.60400390625,
0.486083984375,
0.60400390625,
0.486083984375,
0.6181640625,
0.1351318359375,
0.6181640625
],
[
0.51318359375,
0.6025390625,
0.87744140625,
0.6025390625,
0.87744140625,
0.61328125,
0.51318359375,
0.61328125
],
[
0.1175537109375,
0.619140625,
0.484375,
0.619140625,
0.484375,
0.6328125,
0.1175537109375,
0.6328125
],
[
0.51318359375,
0.6181640625,
0.880859375,
0.6181640625,
0.880859375,
0.6318359375,
0.51318359375,
0.6318359375
],
[
0.118408203125,
0.63720703125,
0.484375,
0.63720703125,
0.484375,
0.65087890625,
0.118408203125,
0.65087890625
],
[
0.509765625,
0.63525390625,
0.72314453125,
0.63525390625,
0.72314453125,
0.6484375,
0.509765625,
0.6484375
],
[
0.1368408203125,
0.65283203125,
0.484375,
0.65283203125,
0.484375,
0.66650390625,
0.1368408203125,
0.66650390625
],
[
0.51318359375,
0.662109375,
0.5556640625,
0.662109375,
0.5556640625,
0.67333984375,
0.51318359375,
0.67333984375
],
[
0.55517578125,
0.662109375,
0.7119140625,
0.662109375,
0.7119140625,
0.67578125,
0.55517578125,
0.67578125
],
[
0.118408203125,
0.66943359375,
0.484375,
0.66943359375,
0.484375,
0.68310546875,
0.118408203125,
0.68310546875
],
[
0.118408203125,
0.68505859375,
0.484375,
0.68505859375,
0.484375,
0.69873046875,
0.118408203125,
0.69873046875
],
[
0.51123046875,
0.681640625,
0.87890625,
0.681640625,
0.87890625,
0.69677734375,
0.51123046875,
0.69677734375
],
[
0.1175537109375,
0.69970703125,
0.486083984375,
0.70068359375,
0.486083984375,
0.71533203125,
0.1175537109375,
0.71435546875
],
[
0.51318359375,
0.69873046875,
0.875,
0.69873046875,
0.875,
0.71337890625,
0.51318359375,
0.71337890625
],
[
0.118408203125,
0.71728515625,
0.484375,
0.71728515625,
0.484375,
0.73095703125,
0.118408203125,
0.73095703125
],
[
0.118408203125,
0.73388671875,
0.484375,
0.73388671875,
0.484375,
0.74755859375,
0.118408203125,
0.74755859375
],
[
0.57080078125,
0.72998046875,
0.8212890625,
0.73095703125,
0.8212890625,
0.7451171875,
0.57080078125,
0.74462890625
],
[
0.118408203125,
0.75,
0.486083984375,
0.75,
0.486083984375,
0.76318359375,
0.118408203125,
0.76318359375
],
[
0.53076171875,
0.7587890625,
0.87890625,
0.7587890625,
0.87890625,
0.7724609375,
0.53076171875,
0.7724609375
],
[
0.118408203125,
0.765625,
0.18896484375,
0.765625,
0.18896484375,
0.77685546875,
0.118408203125,
0.77685546875
],
[
0.5146484375,
0.775390625,
0.87890625,
0.775390625,
0.87890625,
0.7890625,
0.5146484375,
0.7890625
],
[
0.12176513671875,
0.7919921875,
0.382080078125,
0.7919921875,
0.382080078125,
0.8056640625,
0.12176513671875,
0.8056640625
],
[
0.385498046875,
0.79736328125,
0.39208984375,
0.79736328125,
0.39208984375,
0.80029296875,
0.385498046875,
0.80029296875
],
[
0.399658203125,
0.79345703125,
0.484375,
0.79345703125,
0.484375,
0.8046875,
0.399658203125,
0.8046875
],
[
0.12005615234375,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82177734375,
0.12005615234375,
0.82177734375
],
[
0.5380859375,
0.8046875,
0.8505859375,
0.80712890625,
0.8505859375,
0.82177734375,
0.5380859375,
0.8193359375
],
[
0.12005615234375,
0.82470703125,
0.486083984375,
0.82470703125,
0.486083984375,
0.83837890625,
0.12005615234375,
0.83837890625
],
[
0.54150390625,
0.82470703125,
0.748046875,
0.826171875,
0.748046875,
0.84033203125,
0.54150390625,
0.8388671875
],
[
0.118408203125,
0.84033203125,
0.468505859375,
0.84033203125,
0.468505859375,
0.85400390625,
0.118408203125,
0.85400390625
],
[
0.5380859375,
0.84423828125,
0.72900390625,
0.84423828125,
0.72900390625,
0.8623046875,
0.5380859375,
0.8623046875
],
[
0.133544921875,
0.85498046875,
0.48779296875,
0.85498046875,
0.48779296875,
0.876953125,
0.133544921875,
0.876953125
],
[
0.5439453125,
0.86767578125,
0.8310546875,
0.86767578125,
0.8310546875,
0.8818359375,
0.5439453125,
0.8818359375
],
[
0.118408203125,
0.8759765625,
0.3779296875,
0.8759765625,
0.3779296875,
0.8896484375,
0.118408203125,
0.8896484375
],
[
0.4013671875,
0.87451171875,
0.435791015625,
0.87451171875,
0.435791015625,
0.8818359375,
0.4013671875,
0.8818359375
],
[
0.4375,
0.876953125,
0.486083984375,
0.876953125,
0.486083984375,
0.88818359375,
0.4375,
0.88818359375
],
[
0.3837890625,
0.88525390625,
0.441650390625,
0.88525390625,
0.441650390625,
0.892578125,
0.3837890625,
0.892578125
],
[
0.5341796875,
0.892578125,
0.8818359375,
0.892578125,
0.8818359375,
0.90673828125,
0.5341796875,
0.90673828125
],
[
0.139404296875,
0.90478515625,
0.376220703125,
0.90478515625,
0.376220703125,
0.91845703125,
0.139404296875,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.8623046875,
0.90478515625,
0.8623046875,
0.91943359375,
0.51123046875,
0.91943359375
]
] |
[
"IP",
"IPfast",
"12fast",
"IP",
"L2",
"",
"",
"IP (center)",
"L? (center)",
"0.6-",
"IP,L (norm)",
"IP,L? (center, norm)",
".",
"0.5 ",
"0.6",
" 0.4",
"",
"0.5",
"0.2 ",
"ISIS",
"o",
"0.1 ",
"",
"",
"DPRSent",
"ence BERT",
"DPR",
"Sentence BERT",
"(Avg)",
"[CLS",
"0.1",
"(Avg)",
"CBERE",
"[CLS]",
"RERE",
"(Avg)",
"[CLS]",
"DPR",
"Sentence",
"BERT",
"DPR",
"Sentence",
"BER",
"(Avg)",
"BERT",
"(Avg)",
"[CLS]",
"BERT",
"[CLS",
"[CLS]",
"(Avg)",
"Figure l: Comparison of different BERT-based embed-",
"ding models and versions when using faster but slightly",
"inaccurate nearest neighbour search. [CLS] is the spe",
"Figure 2: Effect of data centering and normalization on",
"cific token embedding from the last layer while (Avg)",
"is all token average.",
"performance (evaluatedwithFAISS)",
"它在所有情况下都会改善结果（与没有预处理",
"关数据集大小，请参阅Table6）。此数据集",
"相比）。对查询和文档分别进行规范化和居中。",
"（查询和文档）的768维嵌入（32位浮点）加",
"此外，如果向量是归一化的，那么对于L?和",
"起来为7GB（整个未修剪的数据集为146GB）",
"内积，检索到的文档是相同的。",
"3.3Uncompressed Retrieval Peformance",
"尽管如此，我们认为单独研究L²的压缩能",
"力和内积仍然是有意义的，因为归一化向量压",
"为了建立未压缩性能的基线，我们使用基于",
"缩的输出不需要归一化",
"BERT(Devlinetal.，2019）的模型。我们考虑",
"(1)香草BERT,(2)SentenceBERT(Reimers and",
"1",
"Compression Methods",
"Gurevych, 2019)和 (3)DPR (Karpukhin et al.",
"2020），它经过专门的文档检索培训。为了获",
"在建立了未压缩基线的检索性能之后，我们现",
"得文档嵌人，我们使用「CLS】中的最后一个",
"在转向压缩密集文档索引和查询的方法",
"隐藏状态表示形式或最后一层的标记之间的平",
"请注意，我们考虑在已经训练的索引上使用",
"均值",
"无监督方法，以最大程度地简化使用和适用",
"我们的第一个实验比较了HotpotQA上不同",
"性。这与监督方法相反，监督方法可以访问查",
"模型的检索性能。结果显示在Figure1。与以",
"询-文档相关性映射，或者访问训练中的降维",
"前的作品（ReimersandGurevych，2019）一致，",
"(i。e。较低的最终层尺寸)",
"个立即明显的结论是香草BERT的性能很",
"4.1",
"Random Projection",
"差，特别是在为［CLS】令牌采用隐藏状态表",
"示时。接下来，为了使计算易于处理，我们使",
"对给定索引αERd执行降维的最简单方法是",
"用FAISS（Johnsonetal.,2019）重复实验。7我",
"仅随机保留某些d'维度并删除所有其他维度：",
"们发现模型之间的性能损失是系统性的，这保",
"证了使用这种近似最近邻搜索进行比较，并且",
"fdrop.(α)= (αm1, am2,...,md)",
"我们以下所有实验都将在DPR-CLS模型上使",
"另一种方法是贪婪地搜索要删除的维度（那",
"用FAISS",
"些在省略时可以提高性能或减少性能的维度）：",
"Pre-processing Transformations.",
"",
"Figure 1 还",
"表明，模型性能，尤其是对于DPR，在很大程",
"pi(a)= (o,1,...,&i-1,&i+1,..,768)",
"度上取决于用于检索的相似性度量。这是因为",
"L; = R-Prec(p;(Q), Pi(D))",
"没有一个模型在默认情况下产生归一化向量",
"m = sortesc-([1...768])",
"Figure2表明，仅执行归一化（高）有时会",
"fgreedy drop.(α) = (ami, Cm2.., mdt)",
"损害性能，但是当预先加入居中",
"C-T",
"）时，",
"a-",
"\"arg max -a-bll? = arg maxk-(a, a)?-(b, b)2 +",
"7IndexIVFFlat, nlist=200, nprobe=100",
"2 - (a,b) = argmax, 2. (a, b) - 2 = arg max (a,b)"
] |
{
"arxiv_id": "2204.02906",
"language": "zh",
"page_idx": 3
}
|
||
2204.02906-zh-4
|
2204.02906-zh-4.jpg
|
这两种方法的优点是它们可以很容易地用单个$\mathbb{R}^{768\times d}$矩阵表示。我们考虑了另外两种标准的随机投影方法：高斯随机投影和稀疏随机投影 (Fodor, 2002)。这种随机投影主要适用于内积 (Kaski, 1998) , 尽管通过归一化向量消除了差异(这也提高了性能)
<FigureHere>
Figure 3: Dimension reduction using different random projections methods. Presented values are the max of 3 runs (except for greedy dimension dropping, which is deterministic), semi-transparent lines correspond to the minimum. Embeddings are provided by centered and normalized DPR-CLS. Final vectors are also postprocessed by centering and normalization.
$\textbf{Results. }$ 所有随机投影方法的结果都显示
在 Figure 3 中。高斯随机投影的性能似乎与稀疏随机投影相同。这两种方法的性能没有完全恢复。有趣的是，简单地丢弃随机维度导致比稀疏或高斯随机投影更好的性能。贪婪的维度下降甚至在饱和之前的某些情况下比随机维度下降略微提高了性能，并且是确定性的。如Table 2 所示 , 带有后处理的贪婪降维在所有随机投影方法中实现了最佳性能。没有后处理， 与内积相比，$L_2$距离效果更好
# 4.2 Principal Component Analysis
降维的另一个自然候选是主成分分析(PCA) ( F. R. S. , 1901) 。PCA 考 虑 方 差 最 大 的 维 度 , 忽 略其余的维度。这导致了一个投影矩阵，该矩阵使用正交基$T$将原始数据投影到主分量上。以下损失最小化$\mathcal{L}=$MSE$(T^\prime Tx,x)$ 。请注意， 我们在文档索引，查询嵌入或两者的协方差矩阵上拟合 PCA, 然后将训练的降维投影应用于文档和查询嵌人
Results. 执行 PCA 的结果如 Figure 4 所示。首先，我们发现未压缩的性能以及压缩的效果高度依赖于数据预处理。这不应该令人惊讶， 因为 PCA 算法假设居中和预处理的数据。然而，我们强调并证明了这一步的重要性。这是
由输入向量的归一化给出的，并且 PCA 的列
向量是正交的
其次，当数据未居中时，PCA 对其训练的内容敏感。Figure 4 系统地显示，对可用查询集的训练比对文档或两者的组合进行训练提供了更好的性能。随后，在将数据居中之后，用于拟合的内容不再重要：查询和文档都提供了对数据方差的良好估计，并且在 Section 5.1 中明确探索了 PCA 对训练数据大小的依赖性。查询提供更好的结果而不居中的原因是它们首先更加居中，如 Table 1 所示
Avg. $L^1$ (std) Avg. $L^2$ (std)
Documents 243.0(20.1)
$12.3\pmod{1.6}$
137.0 (7.5)
9.3 (0.2)
Queries
Table 1: Average $L^1$ and $L^2$ norms of document and query embeddings from DPR-CLS without preprocessing.
在所有情况下，PCA 性能都在 128 维度附近开始平稳，并且在未压缩性能的 95% 范围内。最后，我们注意到，尽管 PCA 关注的是最小化重建损失，但 Figure 4 表明，即使在大大降低了重建损失之后，检索性能也没有显着提高。我们在 Section 5.4 中进一步讨论了这一发现
Component Scaling. PCA 的一个潜在问题是可能存在主导向量空间的维度。Mu et al. (2017)建议简单地删除与最高特征值相对应的维度，尽管我们发现简单地按比例缩小前 k 个特征向量系统地优于标准 PCA。为简单起见，我们专注于前 5 个特征向量，并对缩放因子进行了小规模的网格搜索。表现最好的是(0.5,0.8,0.8,0.9,0.8)和 Table 2 表明它在检索性能方面提供了一个小的额外提升
## 4.3 Autoencoder
PCA 用于降维的直接扩展是使用自动编码器该编码器已被广泛探索 (Hu et al., 2014; Wang et al.,2016)。通常，该模型由编码器$e:\mathbb{R}^d\to$ $\mathbb{R}^b$ ,从较高维度到目标 (瓶颈)维度的函数以及解码器$r:\mathbb{R}^b\to\mathbb{R}^d$描述，该解码器从目标维度映射回原始向量空间。然后通常将最终(重建)损失计算为$\mathcal{L}=\mathsf{MSE}((r\circ e)(x),x)$ 。为了减少数据集的维度，仅将函数$e$应用于查询和文档嵌入。我们考虑三种具有瓶颈的模型：
1. A linear projection similar to PCA but with-
out the restriction of orthonormal columns:
$$\begin{aligned}e_1(\boldsymbol x)&=L_{128}^{768}\\r_1(\boldsymbol x)&=L_{768}^{128}\end{aligned}$$
|
[
[
0.1376953125,
0.0885009765625,
0.486083984375,
0.0885009765625,
0.486083984375,
0.09857177734375,
0.1376953125,
0.09857177734375
],
[
0.51318359375,
0.0872802734375,
0.87744140625,
0.0872802734375,
0.87744140625,
0.097412109375,
0.51318359375,
0.097412109375
],
[
0.1175537109375,
0.09979248046875,
0.48779296875,
0.1015625,
0.48779296875,
0.11639404296875,
0.1175537109375,
0.114013671875
],
[
0.5146484375,
0.10394287109375,
0.623046875,
0.10394287109375,
0.623046875,
0.11517333984375,
0.5146484375,
0.11517333984375
],
[
0.118408203125,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.118408203125,
0.1318359375
],
[
0.53076171875,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.53076171875,
0.1318359375
],
[
0.118408203125,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.118408203125,
0.1484375
],
[
0.51318359375,
0.1353759765625,
0.880859375,
0.1353759765625,
0.880859375,
0.1484375,
0.51318359375,
0.1484375
],
[
0.118408203125,
0.149658203125,
0.483642578125,
0.149658203125,
0.483642578125,
0.163330078125,
0.118408203125,
0.163330078125
],
[
0.51318359375,
0.15087890625,
0.87890625,
0.15087890625,
0.87890625,
0.1644287109375,
0.51318359375,
0.1644287109375
],
[
0.118408203125,
0.166259765625,
0.32568359375,
0.166259765625,
0.32568359375,
0.179931640625,
0.118408203125,
0.179931640625
],
[
0.515625,
0.16748046875,
0.87890625,
0.16748046875,
0.87890625,
0.18115234375,
0.515625,
0.18115234375
],
[
0.51318359375,
0.1817626953125,
0.87890625,
0.1817626953125,
0.87890625,
0.1953125,
0.51318359375,
0.1953125
],
[
0.51123046875,
0.198974609375,
0.87890625,
0.198974609375,
0.87890625,
0.2120361328125,
0.51123046875,
0.2120361328125
],
[
0.145263671875,
0.210205078125,
0.162109375,
0.210205078125,
0.162109375,
0.218505859375,
0.145263671875,
0.218505859375
],
[
0.301513671875,
0.210205078125,
0.330078125,
0.210205078125,
0.330078125,
0.21435546875,
0.301513671875,
0.21435546875
],
[
0.51318359375,
0.21435546875,
0.880859375,
0.213134765625,
0.880859375,
0.22802734375,
0.51318359375,
0.2286376953125
],
[
0.1435546875,
0.235107421875,
0.16796875,
0.235107421875,
0.16796875,
0.248779296875,
0.1435546875,
0.248779296875
],
[
0.5146484375,
0.232177734375,
0.8818359375,
0.232177734375,
0.8818359375,
0.2452392578125,
0.5146484375,
0.2452392578125
],
[
0.12176513671875,
0.2410888671875,
0.181396484375,
0.24462890625,
0.173828125,
0.303955078125,
0.11419677734375,
0.300537109375
],
[
0.51123046875,
0.2464599609375,
0.72314453125,
0.2476806640625,
0.72314453125,
0.262451171875,
0.51123046875,
0.26123046875
],
[
0.64208984375,
0.273681640625,
0.74072265625,
0.273681640625,
0.74072265625,
0.2880859375,
0.64208984375,
0.2880859375
],
[
0.73046875,
0.271484375,
0.84228515625,
0.271484375,
0.84228515625,
0.289306640625,
0.73046875,
0.289306640625
],
[
0.14111328125,
0.289306640625,
0.166259765625,
0.289306640625,
0.166259765625,
0.297607421875,
0.14111328125,
0.297607421875
],
[
0.337646484375,
0.2880859375,
0.42236328125,
0.2880859375,
0.42236328125,
0.298583984375,
0.337646484375,
0.298583984375
],
[
0.64404296875,
0.292236328125,
0.7314453125,
0.292236328125,
0.7314453125,
0.30712890625,
0.64404296875,
0.30712890625
],
[
0.759765625,
0.292236328125,
0.8271484375,
0.292236328125,
0.8271484375,
0.30712890625,
0.759765625,
0.30712890625
],
[
0.62744140625,
0.295654296875,
0.64990234375,
0.295654296875,
0.64990234375,
0.302734375,
0.62744140625,
0.302734375
],
[
0.32666015625,
0.297607421875,
0.46337890625,
0.2998046875,
0.46337890625,
0.311279296875,
0.32666015625,
0.308837890625
],
[
0.5556640625,
0.29443359375,
0.63232421875,
0.29443359375,
0.63232421875,
0.305908203125,
0.5556640625,
0.305908203125
],
[
0.55517578125,
0.30810546875,
0.60791015625,
0.30810546875,
0.60791015625,
0.319580078125,
0.55517578125,
0.319580078125
],
[
0.6533203125,
0.30810546875,
0.7255859375,
0.30810546875,
0.7255859375,
0.319580078125,
0.6533203125,
0.319580078125
],
[
0.76318359375,
0.30810546875,
0.8212890625,
0.30810546875,
0.8212890625,
0.319580078125,
0.76318359375,
0.319580078125
],
[
0.343505859375,
0.308837890625,
0.372802734375,
0.312255859375,
0.370361328125,
0.322509765625,
0.340087890625,
0.319580078125
],
[
0.145263671875,
0.312255859375,
0.1695556640625,
0.31640625,
0.16455078125,
0.330078125,
0.14111328125,
0.325927734375
],
[
0.32568359375,
0.32177734375,
0.391357421875,
0.32177734375,
0.391357421875,
0.33203125,
0.32568359375,
0.33203125
],
[
0.509765625,
0.330810546875,
0.87890625,
0.330810546875,
0.87890625,
0.344482421875,
0.509765625,
0.344482421875
],
[
0.175537109375,
0.34033203125,
0.189697265625,
0.34033203125,
0.189697265625,
0.3515625,
0.175537109375,
0.3515625
],
[
0.255126953125,
0.34033203125,
0.279541015625,
0.34033203125,
0.279541015625,
0.352783203125,
0.255126953125,
0.352783203125
],
[
0.3525390625,
0.342529296875,
0.37451171875,
0.342529296875,
0.37451171875,
0.352783203125,
0.3525390625,
0.352783203125
],
[
0.447509765625,
0.34033203125,
0.47021484375,
0.34033203125,
0.47021484375,
0.352783203125,
0.447509765625,
0.352783203125
],
[
0.51318359375,
0.3466796875,
0.880859375,
0.3466796875,
0.880859375,
0.35986328125,
0.51318359375,
0.35986328125
],
[
0.28955078125,
0.352783203125,
0.35107421875,
0.352783203125,
0.35107421875,
0.363525390625,
0.28955078125,
0.363525390625
],
[
0.51318359375,
0.363525390625,
0.58837890625,
0.363525390625,
0.58837890625,
0.374755859375,
0.51318359375,
0.374755859375
],
[
0.1175537109375,
0.383056640625,
0.484375,
0.383056640625,
0.484375,
0.396728515625,
0.1175537109375,
0.396728515625
],
[
0.1158447265625,
0.396728515625,
0.486083984375,
0.3955078125,
0.486083984375,
0.410400390625,
0.1158447265625,
0.410888671875
],
[
0.53076171875,
0.391357421875,
0.880859375,
0.391357421875,
0.880859375,
0.405029296875,
0.53076171875,
0.405029296875
],
[
0.1158447265625,
0.410888671875,
0.486083984375,
0.410400390625,
0.486083984375,
0.424560546875,
0.1158447265625,
0.42578125
],
[
0.51318359375,
0.407958984375,
0.880859375,
0.407958984375,
0.880859375,
0.421630859375,
0.51318359375,
0.421630859375
],
[
0.1175537109375,
0.42578125,
0.48779296875,
0.42578125,
0.48779296875,
0.439453125,
0.1175537109375,
0.439453125
],
[
0.51318359375,
0.42333984375,
0.880859375,
0.42333984375,
0.880859375,
0.43701171875,
0.51318359375,
0.43701171875
],
[
0.11419677734375,
0.439453125,
0.486083984375,
0.439453125,
0.486083984375,
0.453125,
0.11419677734375,
0.453125
],
[
0.5146484375,
0.440673828125,
0.8818359375,
0.440673828125,
0.8818359375,
0.45361328125,
0.5146484375,
0.45361328125
],
[
0.1175537109375,
0.45361328125,
0.489501953125,
0.454833984375,
0.489501953125,
0.468505859375,
0.1175537109375,
0.46728515625
],
[
0.51318359375,
0.457275390625,
0.880859375,
0.457275390625,
0.880859375,
0.47021484375,
0.51318359375,
0.47021484375
],
[
0.1158447265625,
0.468505859375,
0.4013671875,
0.468505859375,
0.4013671875,
0.482177734375,
0.1158447265625,
0.482177734375
],
[
0.51318359375,
0.47265625,
0.86376953125,
0.47265625,
0.86376953125,
0.486328125,
0.51318359375,
0.486328125
],
[
0.51904296875,
0.496337890625,
0.7060546875,
0.496337890625,
0.7060546875,
0.51123046875,
0.51904296875,
0.51123046875
],
[
0.7001953125,
0.496337890625,
0.880859375,
0.496337890625,
0.880859375,
0.51025390625,
0.7001953125,
0.51025390625
],
[
0.12176513671875,
0.51025390625,
0.2149658203125,
0.51025390625,
0.2149658203125,
0.521484375,
0.12176513671875,
0.521484375
],
[
0.21240234375,
0.5087890625,
0.484375,
0.5087890625,
0.484375,
0.5224609375,
0.21240234375,
0.5224609375
],
[
0.5146484375,
0.51318359375,
0.87890625,
0.51318359375,
0.87890625,
0.52685546875,
0.5146484375,
0.52685546875
],
[
0.118408203125,
0.525390625,
0.486083984375,
0.525390625,
0.486083984375,
0.5390625,
0.118408203125,
0.5390625
],
[
0.51318359375,
0.529296875,
0.880859375,
0.529296875,
0.880859375,
0.5419921875,
0.51318359375,
0.5419921875
],
[
0.118408203125,
0.5419921875,
0.483642578125,
0.5419921875,
0.483642578125,
0.552734375,
0.118408203125,
0.552734375
],
[
0.51123046875,
0.54345703125,
0.880859375,
0.54345703125,
0.880859375,
0.55712890625,
0.51123046875,
0.55712890625
],
[
0.1175537109375,
0.5556640625,
0.486083984375,
0.55712890625,
0.486083984375,
0.57177734375,
0.1175537109375,
0.57080078125
],
[
0.5146484375,
0.56103515625,
0.880859375,
0.56103515625,
0.880859375,
0.57470703125,
0.5146484375,
0.57470703125
],
[
0.118408203125,
0.57373046875,
0.486083984375,
0.57373046875,
0.486083984375,
0.58740234375,
0.118408203125,
0.58740234375
],
[
0.515625,
0.5791015625,
0.880859375,
0.5791015625,
0.880859375,
0.5888671875,
0.515625,
0.5888671875
],
[
0.118408203125,
0.58837890625,
0.486083984375,
0.58837890625,
0.486083984375,
0.6015625,
0.118408203125,
0.6015625
],
[
0.51318359375,
0.59326171875,
0.880859375,
0.59326171875,
0.880859375,
0.60693359375,
0.51318359375,
0.60693359375
],
[
0.118408203125,
0.60546875,
0.48779296875,
0.60546875,
0.48779296875,
0.619140625,
0.118408203125,
0.619140625
],
[
0.51123046875,
0.60986328125,
0.880859375,
0.60986328125,
0.880859375,
0.62353515625,
0.51123046875,
0.62353515625
],
[
0.1158447265625,
0.62060546875,
0.486083984375,
0.62255859375,
0.486083984375,
0.6357421875,
0.1158447265625,
0.63427734375
],
[
0.51318359375,
0.6259765625,
0.7900390625,
0.6259765625,
0.7900390625,
0.6396484375,
0.51318359375,
0.6396484375
],
[
0.1175537109375,
0.63720703125,
0.483642578125,
0.63720703125,
0.483642578125,
0.65087890625,
0.1175537109375,
0.65087890625
],
[
0.51318359375,
0.65283203125,
0.54248046875,
0.65283203125,
0.54248046875,
0.66455078125,
0.51318359375,
0.66455078125
],
[
0.53564453125,
0.65185546875,
0.66015625,
0.65185546875,
0.66015625,
0.6650390625,
0.53564453125,
0.6650390625
],
[
0.51318359375,
0.6728515625,
0.87744140625,
0.6728515625,
0.87744140625,
0.68603515625,
0.51318359375,
0.68603515625
],
[
0.12005615234375,
0.65625,
0.364501953125,
0.65625,
0.364501953125,
0.66650390625,
0.12005615234375,
0.66650390625
],
[
0.1175537109375,
0.68115234375,
0.40478515625,
0.68115234375,
0.40478515625,
0.69482421875,
0.1175537109375,
0.69482421875
],
[
0.5146484375,
0.689453125,
0.880859375,
0.689453125,
0.880859375,
0.703125,
0.5146484375,
0.703125
],
[
0.118408203125,
0.70361328125,
0.48779296875,
0.70361328125,
0.48779296875,
0.71435546875,
0.118408203125,
0.71435546875
],
[
0.51123046875,
0.70361328125,
0.87744140625,
0.703125,
0.87744140625,
0.71728515625,
0.51123046875,
0.71875
],
[
0.1158447265625,
0.71728515625,
0.483642578125,
0.71728515625,
0.483642578125,
0.73095703125,
0.1158447265625,
0.73095703125
],
[
0.509765625,
0.71875,
0.8818359375,
0.7197265625,
0.8818359375,
0.73388671875,
0.509765625,
0.73291015625
],
[
0.118408203125,
0.7353515625,
0.483642578125,
0.7353515625,
0.483642578125,
0.7451171875,
0.118408203125,
0.7451171875
],
[
0.51318359375,
0.736328125,
0.880859375,
0.736328125,
0.880859375,
0.75,
0.51318359375,
0.75
],
[
0.118408203125,
0.75048828125,
0.484375,
0.75048828125,
0.484375,
0.76416015625,
0.118408203125,
0.76416015625
],
[
0.5146484375,
0.7529296875,
0.880859375,
0.7529296875,
0.880859375,
0.7666015625,
0.5146484375,
0.7666015625
],
[
0.12005615234375,
0.76708984375,
0.484375,
0.76708984375,
0.484375,
0.77783203125,
0.12005615234375,
0.77783203125
],
[
0.51318359375,
0.76953125,
0.8818359375,
0.76953125,
0.8818359375,
0.783203125,
0.51318359375,
0.783203125
],
[
0.1175537109375,
0.7822265625,
0.483642578125,
0.7822265625,
0.483642578125,
0.7958984375,
0.1175537109375,
0.7958984375
],
[
0.53076171875,
0.78515625,
0.87890625,
0.78515625,
0.87890625,
0.798828125,
0.53076171875,
0.798828125
],
[
0.1175537109375,
0.798828125,
0.484375,
0.798828125,
0.484375,
0.8125,
0.1175537109375,
0.8125
],
[
0.51123046875,
0.8017578125,
0.87890625,
0.8017578125,
0.87890625,
0.8154296875,
0.51123046875,
0.8154296875
],
[
0.118408203125,
0.8154296875,
0.2476806640625,
0.8154296875,
0.2476806640625,
0.8291015625,
0.118408203125,
0.8291015625
],
[
0.51318359375,
0.81787109375,
0.5400390625,
0.81787109375,
0.5400390625,
0.83056640625,
0.51318359375,
0.83056640625
],
[
0.1234130859375,
0.84130859375,
0.483642578125,
0.84130859375,
0.483642578125,
0.85498046875,
0.1234130859375,
0.85498046875
],
[
0.5263671875,
0.84423828125,
0.8818359375,
0.84423828125,
0.8818359375,
0.85791015625,
0.5263671875,
0.85791015625
],
[
0.12005615234375,
0.85693359375,
0.486083984375,
0.85693359375,
0.486083984375,
0.87060546875,
0.12005615234375,
0.87060546875
],
[
0.54833984375,
0.85986328125,
0.86572265625,
0.85986328125,
0.86572265625,
0.87353515625,
0.54833984375,
0.87353515625
],
[
0.12005615234375,
0.87451171875,
0.483642578125,
0.87451171875,
0.483642578125,
0.88525390625,
0.12005615234375,
0.88525390625
],
[
0.6572265625,
0.8818359375,
0.767578125,
0.87890625,
0.76904296875,
0.90087890625,
0.6591796875,
0.90380859375
],
[
0.118408203125,
0.8896484375,
0.486083984375,
0.8896484375,
0.486083984375,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.12005615234375,
0.90478515625,
0.486083984375,
0.90478515625,
0.486083984375,
0.91845703125,
0.12005615234375,
0.91845703125
],
[
0.66015625,
0.90283203125,
0.767578125,
0.90087890625,
0.767578125,
0.91845703125,
0.66015625,
0.92041015625
]
] |
[
"这两种方法的优点是它们可以很容易地用单",
"由输入向量的归一化给出的，并且PCA的列",
"个R768xd矩阵表示。我们考虑了另外两种标",
"向量是正交的",
"准的随机投影方法：高斯随机投影和稀疏随机",
"其次，当数据未居中时，PCA对其训练的内",
"投影（Fodor,2002）。这种随机投影主要适用于",
"容敏感。Figure4系统地显示，对可用查询集",
"内积（Kaski,1998），尽管通过归一化向量消除",
"的训练比对文档或两者的组合进行训练提供了",
"了差异（这也提高了性能）",
"更好的性能。随后，在将数据居中之后，用于",
"拟合的内容不再重要：查询和文档都提供了对",
"数据方差的良好估计，并且在Section5.1中明",
"0.6",
"..",
"确探索了PCA对训练数据大小的依赖性。查",
"0.5-",
"询提供更好的结果而不居中的原因是它们首先",
"",
"更加居中，如Table1所示",
"Avg. LI (std)",
" Avg. L? (std)",
"0.3",
"Dim.Dropping",
"243.0 (20.1)",
"12.3 (0.6)",
"",
"Greedy Dim.Dropping",
"Documents",
"Queries",
"137.0 (7.5)",
"9.3 (0.2)",
"Spars",
"0.2",
"Gaussian",
"Table 1: Average Ll and L? norms of document",
"32",
"256",
"512",
"768",
"and query embeddings from DPR-CLS without pre-",
"Dimensions",
"processing",
"Figure 3: Dimension reduction using different random",
"projections methods. Presented values are the max of",
"在所有情况下，PCA性能都在128维度附近",
"3 runs (except for greedy dimension dropping, which",
"开始平稳，并且在未压缩性能的95%范围内。",
"is deterministic), semi-transparent lines correspond to",
"最后，我们注意到，尽管PCA关注的是最小化",
"the minimum. Embeddings are provided by centered",
"重建损失，但Figure4表明，即使在大大降低",
"and normalized DPR-CLS. Final vectors are also post-",
"了重建损失之后，检索性能也没有显着提高。",
"processed by centering and normalization.",
"我们在Section5.4中进一步讨论了这一发现",
"Component Scaling.",
"PCA的一个潜在问题",
"Results.",
"所有随机投影方法的结果都显示",
"是可能存在主导向量空间的维度。Mu et al.",
"在Figure3中。高斯随机投影的性能似乎与稀",
"(2017)建议简单地删除与最高特征值相对应",
"疏随机投影相同。这两种方法的性能没有完全",
"的维度，尽管我们发现简单地按比例缩小前k",
"恢复。有趣的是，简单地丢弃随机维度导致比",
"个特征向量系统地优于标准PCA。为简单起",
"稀疏或高斯随机投影更好的性能。贪婪的维度",
"见，我们专注于前5个特征向量，并对缩放",
"下降甚至在饱和之前的某些情况下比随机维",
"因子进行了小规模的网格搜索。表现最好的是",
"度下降略微提高了性能，并且是确定性的。如",
"(0.5,0.8,0.8,0.9,0.8）和Table2表明它在检索",
"Table2所示，带有后处理的贪婪降维在所有随",
"性能方面提供了一个小的额外提升",
"机投影方法中实现了最佳性能。没有后处理",
"4.3",
"Autoencoder",
"PCA用于降维的直接扩展是使用自动编码器",
"与内积相比，L2距离效果更好",
"4.2Principal Component Analysis",
"该编码器已被广泛探索（Huetal.，2014；Wang",
"降维的另一个自然候选是主成分分析（PCA）",
"etal.,2016）。通常，该模型由编码器e：Rd一",
"(F.R.S.，1901）。PCA考虑方差最大的维度，忽",
"Rb，从较高维度到目标（瓶颈）维度的函数",
"略其余的维度。这导致了一个投影矩阵，该矩",
"以及解码器r：R→Rd描述，该解码器从",
"阵使用正交基T将原始数据投影到主分量上。",
"目标维度映射回原始向量空间。然后通常将最",
"以下损失最小化L=MSE(TT&.&）。请注意",
"终（重建）损失计算为L=MSE((roe)(α),α)",
"我们在文档索引，查询嵌入或两者的协方差矩",
"为了减少数据集的维度，仅将函数e应用于",
"阵上拟合PCA，然后将训练的降维投影应用于",
"查询和文档嵌人。我们考虑三种具有瓶颈的模",
"文档和查询嵌入",
"型：",
"Results。执行PCA的结果如Figure4所示。",
"1. A linear projection similar to PCA but with-",
"首先，我们发现未压缩的性能以及压缩的效果",
"out the restriction of orthonormal columns",
"高度依赖于数据预处理。这不应该令人惊边",
"e1(ac) = L728",
"因为PCA算法假设居中和预处理的数据。然",
"而，我们强调并证明了这一步的重要性。这是",
"ri(α) = L8"
] |
{
"arxiv_id": "2204.02906",
"language": "zh",
"page_idx": 4
}
|
||
2212.03597-zh-1
|
2212.03597-zh-1.jpg
|
DeepSpeed Data Efficiency: 通过高效的数据采样和路由提高深度学习模型质量
# 和训练效率
Conglong Li$^*$,Zhewei Yao$^*$,Xiaoxia Wu$^*$,Minjia Zhang, Connor Holmes, Cheng Li, Yuxiong He
Microsoft
$\operatorname{Abstract}$
Recent advances on deep learning models come at the price of formidable training cost. The increasing model size is one of the root causes, but another lessemphasized fact is that data scale is actually increasing at a similar speed as model scale, and the training cost is proportional to both of them. Compared to the rapidly evolving model architecture, how to efficiently use the training data (especially for the expensive foundation model pretraining) is both less explored and difficult to realize due to the lack of a convenient framework that focuses on data efficiency capabilities. To this end, we present DeepSpeed Data Efficiency, a framework that makes better use of data, increases training efficiency, and improves model quality. Specifically, we propose and combine two data efficiency techniques: ef- $\hat{\text{ficient data sampling via a general curriculum learning}}$ library, and efficient data routing via a novel random layerwise token dropping technique. For GPT-3 1.3B language model pretraining, our work achieves 12.5x less data/ time/ cost ( $\$$ 3. $\bar{7}$K if rent on Azure) , while still maintaining $95\%$ of model quality compared to baseline with full data and cost $(\bar{\$}46.\dot{3}$K$).\bar{\mathrm{For~GPT-}}$ 31.3B and BERT-large pretraining, our work can also achieve the same model quality with up to 2x less data/- time/cost, or achieve better model quality under same data/time/cost. DeepSpeed Data Efficiency is easy to use and tune, enabling us to easily apply it and verify its benefit on additional tasks including GPT-3 MoE model pretraining and small-scale GPT-2/ViT finetun- $\operatorname{ing}.$
## 1 Introduction
最近，大规模深度学习模型使我们能够以多种方式实现更多目标，例如代码生成$\overline{\mathrm{E}}($GitHub 2021)和文本到图像生成$\boxed{\mathrm{E}}($Ramesh et al. 2022; Rombach et al. 2022) 。为了不断提高服务质量，深度学习模型架构发展迅速， 模型规模也在以惊人的速度增长。不断增加的模型大小导致前所未有的训练成本 (特别是基础模型预训练),最近在数千个 GPU/TPU 上增长到 2 个月 (Smith et al. 2022; Chowdhery et al. 2022) 。另一方面，一个不太强调的观点是，数据规模实际上以与模型规模相似的速度
$^*$These authors contributed equally.
Copyright $\odot2024$, Association for the Advancement of Ar-
tificial Intelligence (www.aaai.org). All rights reserved.
<FigureHere>
Figure 1: Model scale (number of parameters) and data scale (number of consumed training tokens ) of representative language models in the last 5 years (Devlin et al. 2019; Shoeybi et al. 2019; Brown et al. 2020; Scao et al. 2022; Chowdhery et al. 2022) .
增长，并且训练成本与两者成正比。正如 Fig.1 中绘制的那样，对于过去 5 年中的几个代表性语言模型，模型和数据规模都以类似的速度增长。包括 Chinchilla$\underline{\mathrm{E}}$ (Hoffmann et al. 2022) 和 PaLM2$\underline{\mathrm{F}}($Google 2023) 在内的最新作品强调了以更快的速度增加数据规模的需求。这表明了提高数据效率的重要性：以更少的数据和降低的训练成本实现相同的模型质量，或者以相同的数据量和相似的训练成本实现更好的模型质量
现有的数据效率技术中有两个流行的研究方向：数据采样技术旨在通过从整个数据池中采样最合适的下一批数据来提高收敛速度；数据路由技术旨在通过将每个数据仅路由到模型组件的子集来减少计算量。这些技术提高了数据和训练效率，但现有解决方案有几个局限性：
$\bullet$ Techniques like curriculum learning (CL) improve
data efficiency by indexing and sampling training data based on certain difficulty metric (Bengio et al. 2009) , and it has recently proved effective on large-scale pretraining tasks (Li, Zhang, and He 2022).However, implementing different CL strategies for different user tasks can require a lot of code-refactoring, which is time-consuming and errorprone. In addition, existing implementations have less consideration on scalability, which makes it difficult to analyze and index large-scale training data based on different difficulty metrics.
Existing data routing techniques such as token drop/bypass/pruning were mostly designed for inference and inapplicable to training. TokenByExtended version of this paper (including appendix) can
be found on arxiv (Li et al. $\hat{2022}).$
|
[
[
0.09234619140625,
0.125,
0.90673828125,
0.125,
0.90673828125,
0.1427001953125,
0.09234619140625,
0.1427001953125
],
[
0.441162109375,
0.14453125,
0.5595703125,
0.1427001953125,
0.5595703125,
0.1610107421875,
0.441162109375,
0.1634521484375
],
[
0.098876953125,
0.184326171875,
0.90625,
0.1849365234375,
0.90625,
0.2008056640625,
0.098876953125,
0.199462890625
],
[
0.464111328125,
0.2083740234375,
0.53515625,
0.2083740234375,
0.53515625,
0.222900390625,
0.464111328125,
0.222900390625
],
[
0.25,
0.2734375,
0.316162109375,
0.2734375,
0.316162109375,
0.28857421875,
0.25,
0.28857421875
],
[
0.5908203125,
0.27783203125,
0.60302734375,
0.27783203125,
0.60302734375,
0.2822265625,
0.5908203125,
0.2822265625
],
[
0.6337890625,
0.27783203125,
0.71826171875,
0.27783203125,
0.71826171875,
0.28857421875,
0.6337890625,
0.28857421875
],
[
0.576171875,
0.2841796875,
0.60302734375,
0.2841796875,
0.60302734375,
0.296630859375,
0.576171875,
0.296630859375
],
[
0.583984375,
0.290283203125,
0.60693359375,
0.294189453125,
0.60302734375,
0.306884765625,
0.580078125,
0.302978515625
],
[
0.6337890625,
0.28857421875,
0.7138671875,
0.28857421875,
0.7138671875,
0.29541015625,
0.6337890625,
0.29541015625
],
[
0.10296630859375,
0.296630859375,
0.461669921875,
0.296630859375,
0.461669921875,
0.31005859375,
0.10296630859375,
0.31005859375
],
[
0.576171875,
0.296630859375,
0.5849609375,
0.296630859375,
0.5849609375,
0.2998046875,
0.576171875,
0.2998046875
],
[
0.57861328125,
0.292236328125,
0.611328125,
0.3037109375,
0.6005859375,
0.32080078125,
0.56884765625,
0.309326171875
],
[
0.101318359375,
0.30810546875,
0.463134765625,
0.31005859375,
0.463134765625,
0.324462890625,
0.101318359375,
0.322509765625
],
[
0.59716796875,
0.314453125,
0.60107421875,
0.3193359375,
0.5986328125,
0.3212890625,
0.59326171875,
0.31640625
],
[
0.615234375,
0.3193359375,
0.6611328125,
0.3193359375,
0.6611328125,
0.333984375,
0.615234375,
0.333984375
],
[
0.6640625,
0.32080078125,
0.69677734375,
0.32080078125,
0.69677734375,
0.33203125,
0.6640625,
0.33203125
],
[
0.7158203125,
0.32080078125,
0.7509765625,
0.32080078125,
0.7509765625,
0.333251953125,
0.7158203125,
0.333251953125
],
[
0.7646484375,
0.32080078125,
0.79736328125,
0.32080078125,
0.79736328125,
0.333251953125,
0.7646484375,
0.333251953125
],
[
0.81787109375,
0.32080078125,
0.84619140625,
0.32080078125,
0.84619140625,
0.333251953125,
0.81787109375,
0.333251953125
],
[
0.69775390625,
0.322509765625,
0.71142578125,
0.323974609375,
0.71142578125,
0.32763671875,
0.69775390625,
0.326904296875
],
[
0.10296630859375,
0.322509765625,
0.463134765625,
0.322509765625,
0.463134765625,
0.3359375,
0.10296630859375,
0.3359375
],
[
0.74658203125,
0.322509765625,
0.76123046875,
0.322509765625,
0.76123046875,
0.330810546875,
0.74658203125,
0.330810546875
],
[
0.7607421875,
0.323974609375,
0.765625,
0.323974609375,
0.765625,
0.330078125,
0.7607421875,
0.330078125
],
[
0.10296630859375,
0.335205078125,
0.463134765625,
0.335205078125,
0.463134765625,
0.348388671875,
0.10296630859375,
0.348388671875
],
[
0.10296630859375,
0.347900390625,
0.463134765625,
0.347900390625,
0.463134765625,
0.361083984375,
0.10296630859375,
0.361083984375
],
[
0.51953125,
0.34521484375,
0.9140625,
0.3466796875,
0.9140625,
0.361083984375,
0.51953125,
0.35986328125
],
[
0.10296630859375,
0.35986328125,
0.463134765625,
0.35986328125,
0.463134765625,
0.373779296875,
0.10296630859375,
0.373779296875
],
[
0.521484375,
0.363037109375,
0.9111328125,
0.363037109375,
0.9111328125,
0.373779296875,
0.521484375,
0.373779296875
],
[
0.10455322265625,
0.375,
0.4599609375,
0.375,
0.4599609375,
0.385009765625,
0.10455322265625,
0.385009765625
],
[
0.51953125,
0.375732421875,
0.8544921875,
0.375732421875,
0.8544921875,
0.389404296875,
0.51953125,
0.389404296875
],
[
0.8486328125,
0.376953125,
0.9111328125,
0.376953125,
0.9111328125,
0.386962890625,
0.8486328125,
0.386962890625
],
[
0.10455322265625,
0.386962890625,
0.4599609375,
0.386962890625,
0.4599609375,
0.397705078125,
0.10455322265625,
0.397705078125
],
[
0.51953125,
0.389404296875,
0.9140625,
0.389404296875,
0.9140625,
0.40283203125,
0.51953125,
0.40283203125
],
[
0.10296630859375,
0.397705078125,
0.463134765625,
0.397705078125,
0.463134765625,
0.410888671875,
0.10296630859375,
0.410888671875
],
[
0.51953125,
0.40283203125,
0.779296875,
0.40283203125,
0.779296875,
0.416748046875,
0.51953125,
0.416748046875
],
[
0.10455322265625,
0.412353515625,
0.461669921875,
0.412353515625,
0.461669921875,
0.42236328125,
0.10455322265625,
0.42236328125
],
[
0.10296630859375,
0.423583984375,
0.463134765625,
0.423583984375,
0.463134765625,
0.4375,
0.10296630859375,
0.4375
],
[
0.10296630859375,
0.436279296875,
0.463134765625,
0.436279296875,
0.463134765625,
0.449462890625,
0.10296630859375,
0.449462890625
],
[
0.101318359375,
0.448974609375,
0.463134765625,
0.449462890625,
0.463134765625,
0.46337890625,
0.101318359375,
0.462158203125
],
[
0.521484375,
0.4462890625,
0.9111328125,
0.4462890625,
0.9111328125,
0.460205078125,
0.521484375,
0.460205078125
],
[
0.10455322265625,
0.46337890625,
0.4599609375,
0.46337890625,
0.4599609375,
0.473388671875,
0.10455322265625,
0.473388671875
],
[
0.521484375,
0.458984375,
0.9111328125,
0.458251953125,
0.9111328125,
0.472900390625,
0.521484375,
0.473388671875
],
[
0.10296630859375,
0.473388671875,
0.461669921875,
0.472900390625,
0.461669921875,
0.486083984375,
0.10296630859375,
0.4873046875
],
[
0.521484375,
0.473388671875,
0.91259765625,
0.473388671875,
0.91259765625,
0.4873046875,
0.521484375,
0.4873046875
],
[
0.10296630859375,
0.486083984375,
0.461669921875,
0.486083984375,
0.461669921875,
0.5,
0.10296630859375,
0.5
],
[
0.52294921875,
0.488037109375,
0.9111328125,
0.488037109375,
0.9111328125,
0.501953125,
0.52294921875,
0.501953125
],
[
0.101318359375,
0.498779296875,
0.463134765625,
0.498779296875,
0.463134765625,
0.51220703125,
0.101318359375,
0.51220703125
],
[
0.521484375,
0.501953125,
0.9111328125,
0.501953125,
0.9111328125,
0.51513671875,
0.521484375,
0.51513671875
],
[
0.101318359375,
0.51123046875,
0.463134765625,
0.51123046875,
0.463134765625,
0.5244140625,
0.101318359375,
0.5244140625
],
[
0.521484375,
0.5166015625,
0.9111328125,
0.5166015625,
0.9111328125,
0.52978515625,
0.521484375,
0.52978515625
],
[
0.101318359375,
0.5234375,
0.463134765625,
0.5234375,
0.463134765625,
0.537109375,
0.101318359375,
0.537109375
],
[
0.521484375,
0.52978515625,
0.91259765625,
0.52978515625,
0.91259765625,
0.54345703125,
0.521484375,
0.54345703125
],
[
0.10296630859375,
0.53857421875,
0.461669921875,
0.53857421875,
0.461669921875,
0.548828125,
0.10296630859375,
0.548828125
],
[
0.101318359375,
0.548828125,
0.463134765625,
0.548828125,
0.463134765625,
0.5625,
0.101318359375,
0.5625
],
[
0.52294921875,
0.54345703125,
0.79833984375,
0.54345703125,
0.79833984375,
0.556640625,
0.52294921875,
0.556640625
],
[
0.10296630859375,
0.56298828125,
0.461669921875,
0.56298828125,
0.461669921875,
0.57373046875,
0.10296630859375,
0.57373046875
],
[
0.53759765625,
0.5595703125,
0.9111328125,
0.5595703125,
0.9111328125,
0.57275390625,
0.53759765625,
0.57275390625
],
[
0.101318359375,
0.57470703125,
0.463134765625,
0.57470703125,
0.463134765625,
0.58837890625,
0.101318359375,
0.58837890625
],
[
0.52294921875,
0.57275390625,
0.90185546875,
0.57275390625,
0.90185546875,
0.5859375,
0.52294921875,
0.5859375
],
[
0.10296630859375,
0.5869140625,
0.463134765625,
0.5869140625,
0.463134765625,
0.60107421875,
0.10296630859375,
0.60107421875
],
[
0.521484375,
0.5859375,
0.9111328125,
0.5859375,
0.9111328125,
0.599609375,
0.521484375,
0.599609375
],
[
0.10455322265625,
0.6015625,
0.461669921875,
0.6015625,
0.461669921875,
0.6123046875,
0.10455322265625,
0.6123046875
],
[
0.52294921875,
0.5986328125,
0.9111328125,
0.5986328125,
0.9111328125,
0.61279296875,
0.52294921875,
0.61279296875
],
[
0.10455322265625,
0.6142578125,
0.461669921875,
0.6142578125,
0.461669921875,
0.625,
0.10455322265625,
0.625
],
[
0.52294921875,
0.61279296875,
0.91259765625,
0.61279296875,
0.91259765625,
0.62744140625,
0.52294921875,
0.62744140625
],
[
0.10455322265625,
0.626953125,
0.461669921875,
0.626953125,
0.461669921875,
0.63720703125,
0.10455322265625,
0.63720703125
],
[
0.52294921875,
0.62890625,
0.5458984375,
0.62890625,
0.5458984375,
0.64013671875,
0.52294921875,
0.64013671875
],
[
0.1004638671875,
0.63720703125,
0.463134765625,
0.63623046875,
0.463134765625,
0.65087890625,
0.1004638671875,
0.6513671875
],
[
0.52392578125,
0.642578125,
0.91259765625,
0.642578125,
0.91259765625,
0.65576171875,
0.52392578125,
0.65576171875
],
[
0.10296630859375,
0.6494140625,
0.463134765625,
0.6494140625,
0.463134765625,
0.66357421875,
0.10296630859375,
0.66357421875
],
[
0.54150390625,
0.6572265625,
0.91259765625,
0.6572265625,
0.91259765625,
0.67041015625,
0.54150390625,
0.67041015625
],
[
0.10296630859375,
0.6640625,
0.131591796875,
0.6640625,
0.131591796875,
0.6767578125,
0.10296630859375,
0.6767578125
],
[
0.54150390625,
0.67041015625,
0.85693359375,
0.67041015625,
0.85693359375,
0.6845703125,
0.54150390625,
0.6845703125
],
[
0.8505859375,
0.67236328125,
0.91259765625,
0.67236328125,
0.91259765625,
0.68310546875,
0.8505859375,
0.68310546875
],
[
0.5400390625,
0.68310546875,
0.91259765625,
0.6845703125,
0.91259765625,
0.69873046875,
0.5400390625,
0.69775390625
],
[
0.2115478515625,
0.69775390625,
0.2467041015625,
0.69775390625,
0.2467041015625,
0.708984375,
0.2115478515625,
0.708984375
],
[
0.238525390625,
0.69580078125,
0.3544921875,
0.69775390625,
0.3544921875,
0.7119140625,
0.238525390625,
0.71044921875
],
[
0.54150390625,
0.69873046875,
0.91259765625,
0.69873046875,
0.91259765625,
0.7119140625,
0.54150390625,
0.7119140625
],
[
0.0882568359375,
0.71630859375,
0.47802734375,
0.71630859375,
0.47802734375,
0.73095703125,
0.0882568359375,
0.73095703125
],
[
0.54150390625,
0.7119140625,
0.91259765625,
0.7119140625,
0.91259765625,
0.72607421875,
0.54150390625,
0.72607421875
],
[
0.0882568359375,
0.73095703125,
0.47802734375,
0.73095703125,
0.47802734375,
0.744140625,
0.0882568359375,
0.744140625
],
[
0.54150390625,
0.72607421875,
0.9140625,
0.72607421875,
0.9140625,
0.7392578125,
0.54150390625,
0.7392578125
],
[
0.08660888671875,
0.74365234375,
0.4794921875,
0.744140625,
0.4794921875,
0.75830078125,
0.08660888671875,
0.7568359375
],
[
0.54150390625,
0.7392578125,
0.9140625,
0.7392578125,
0.9140625,
0.75244140625,
0.54150390625,
0.75244140625
],
[
0.08905029296875,
0.75830078125,
0.47802734375,
0.75830078125,
0.47802734375,
0.771484375,
0.08905029296875,
0.771484375
],
[
0.5400390625,
0.75390625,
0.9140625,
0.75244140625,
0.9140625,
0.7666015625,
0.5400390625,
0.767578125
],
[
0.08905029296875,
0.7734375,
0.476318359375,
0.7734375,
0.476318359375,
0.7841796875,
0.08905029296875,
0.7841796875
],
[
0.54345703125,
0.767578125,
0.9140625,
0.767578125,
0.9140625,
0.78076171875,
0.54345703125,
0.78076171875
],
[
0.0882568359375,
0.7861328125,
0.47802734375,
0.7861328125,
0.47802734375,
0.7998046875,
0.0882568359375,
0.7998046875
],
[
0.54150390625,
0.78076171875,
0.9140625,
0.78076171875,
0.9140625,
0.79443359375,
0.54150390625,
0.79443359375
],
[
0.0882568359375,
0.7998046875,
0.47802734375,
0.7998046875,
0.47802734375,
0.81298828125,
0.0882568359375,
0.81298828125
],
[
0.54345703125,
0.796875,
0.79833984375,
0.796875,
0.79833984375,
0.806640625,
0.54345703125,
0.806640625
],
[
0.40185546875,
0.81103515625,
0.47802734375,
0.81298828125,
0.47802734375,
0.82763671875,
0.40185546875,
0.82568359375
],
[
0.5400390625,
0.8095703125,
0.91259765625,
0.8095703125,
0.91259765625,
0.82275390625,
0.5400390625,
0.82275390625
],
[
0.52783203125,
0.81298828125,
0.54345703125,
0.81298828125,
0.54345703125,
0.8173828125,
0.52783203125,
0.8173828125
],
[
0.08660888671875,
0.8125,
0.3896484375,
0.814453125,
0.3896484375,
0.82763671875,
0.08660888671875,
0.82568359375
],
[
0.0882568359375,
0.82763671875,
0.47802734375,
0.82763671875,
0.47802734375,
0.84130859375,
0.0882568359375,
0.84130859375
],
[
0.54150390625,
0.82275390625,
0.9140625,
0.82275390625,
0.9140625,
0.8369140625,
0.54150390625,
0.8369140625
],
[
0.54150390625,
0.8369140625,
0.9140625,
0.8369140625,
0.9140625,
0.85107421875,
0.54150390625,
0.85107421875
],
[
0.108642578125,
0.85107421875,
0.34228515625,
0.85107421875,
0.34228515625,
0.8642578125,
0.108642578125,
0.8642578125
],
[
0.08660888671875,
0.86279296875,
0.4794921875,
0.86279296875,
0.4794921875,
0.876953125,
0.08660888671875,
0.876953125
],
[
0.5458984375,
0.8623046875,
0.9140625,
0.86279296875,
0.9140625,
0.87744140625,
0.5458984375,
0.876953125
],
[
0.0849609375,
0.875,
0.449462890625,
0.87548828125,
0.449462890625,
0.89013671875,
0.0849609375,
0.8896484375
],
[
0.5185546875,
0.875,
0.74658203125,
0.87548828125,
0.74658203125,
0.89013671875,
0.5185546875,
0.8896484375
]
] |
[
"DeepSpeedDataEfficiency：通过高效的数据采样和路由提高深度学习模型质量",
"和训练效率",
"Conglong Li*, Zhewei Yao*, Xiaoxia Wu*, Minjia Zhang, Connor Holmes, Cheng Li, Yuxiong He",
"Microsoft",
"Abstract",
"00",
"Model Scale (Billion)",
"600",
"400",
"DataScale(Billion)",
"Recent advances on deep learning models come at",
"-",
" 00",
"the price of formidable training cost. The increasing",
"",
"(2ERT)",
"2019)",
"(2020)",
"(2022)",
"(2022)",
"",
"model size is one of the root causes, but another less-",
"",
"",
"emphasized fact is that data scale is actually increas-",
"ing at a similar speed as model scale, and the training",
"Figure 1: Model scale (number of parameters) and data",
"cost is proportional to both of them. Compared to the",
"scale(numberof consumedtrainingtokens)of repre",
"rapidlyevolvingmodel architecture,howtoefficiently",
"sentative language models in the last 5 years",
"(Devlin",
"usethe training data (especially for theexpensivefoun",
"et al. 2019; Shoeybi et al. 2019; Brown et al. 2020; Scao",
"dation model pretraining) is both less explored and dif-",
"et al. 2022; Chowdhery et al. 2022)",
"ficult to realize due tothe lack ofaconvenient frame",
"work that focuses on data efficiency capabilities. To this",
"end, we present DeepSpeed Data Efficiency, a frame-",
"work that makes better use of data, increases training",
"增长，并且训练成本与两者成正比。正如Fig.1中绘",
"efficiency,and improvesmodelquality.Specifically,we",
"制的那样，对于过去5年中的几个代表性语言模型，模",
"propose and combine two data efficiency techniques: ef",
"型和数据规模都以类似的速度增长。包括ChinchillaE",
"ficient data sampling via a general curriculum learning",
"(Hoffmann etal.2022)和PaLM2E(Google2023)在内",
"library, and efficient data routing via a novel random",
"的最新作品强调了以更快的速度增加数据规模的需求。",
"layerwise token dropping technique. For GPT-3 1.3B",
"这表明了提高数据效率的重要性：以更少的数据和降低",
"language model pretraining, our work achieves 12.5x",
"的训练成本实现相同的模型质量，或者以相同的数据量",
"lessdata/time/cost ($3.7KifrentonAzure),while",
"still maintaining 95 % of model quality compared to",
"和相似的训练成本实现更好的模型质量",
"baselinewithfulldataandcost ($ 46.3K).ForGPT",
"现有的数据效率技术中有两个流行的研究方向：数据",
"3 1.3B and BERT-large pretraining, our work can also",
"采样技术旨在通过从整个数据池中采样最合适的下",
"achieve the same model quality with up to 2x less data/-",
"批数据来提高收敛速度；数据路由技术旨在通过将每",
"time/cost, or achieve better model quality under same",
"个数据仅路由到模型组件的子集来减少计算量。这些技",
"data/time/cost.DeepSpeed DataEficiencyis easy to",
"术提高了数据和训练效率，但现有解决方案有几个局限",
"use and tune, enabling us to easily apply it and verify",
"性：",
"its benefit on additional tasks including GPT-3 MoE",
" Techniques like curriculum learning (CL) improve",
"model pretraining and small-scale GPT-2/ViT finetun-",
"data efficiency by indexing and sampling training",
"ing.",
"data based on certain difficulty metric",
"(Bengio",
"et al. 2009), and it has recently proved effective",
"1",
"Introduction",
"on large-scale pretraining tasks (Li, Zhang, and He",
"最近，大规模深度学习模型使我们能够以多种方式实现",
"2022):However, implementing different CL strate-",
"更多目标，例如代码生成E（GitHub2021）和文本到图",
"gies for different user tasks can require a lot of",
"像生成E(Rameshetal.2022;Rombachetal.2022)。",
"code-refactoring, which is time-consuming and error-",
"为了不断提高服务质量，深度学习模型架构发展迅速，",
"prone. In addition, existing implementations have",
"模型规模也在以惊人的速度增长。不断增加的模型大小",
"less consideration on scalability, which makes it dif-",
"导致前所未有的训练成本（特别是基础模型预训练），最",
"ficult to analyze and index large-scale training data",
"近在数千个GPU/TPU上增长到2个月（Smith etal.",
"based on different difficulty metrics.",
"一个不太强",
"Existing data routing techniques such as token",
"",
"2022：Chowdheryetal.2022）。另一方面，",
"调的观点是，数据规模实际上以与模型规模相似的速度",
"drop/bypass/pruning were mostly designed for in-",
"ference and inapplicable to training.TokenBy-",
"*These authors contributed equally.",
"Copyright 2024, Association for the Advancement of Ar-",
"Extended version of this paper (including appendix) can",
"tificial Intelligence (www.aaai.org). All rights reserved.",
"be found on arxiv (Li et al. 2022)."
] |
{
"arxiv_id": "2212.03597",
"language": "zh",
"page_idx": 1
}
|
||
1612.00104-zh-1
|
1612.00104-zh-1.jpg
|
Robust Optimization for Tree-Structured Stochastic Network Design
$\begin{array}{cc}{\mathrm{Xiaojian~Wu^{1}}}&{\mathrm{Akshat~Kumar^{2}~Daniel~Sheldon^{3,4}~Shlomo~Zilberstein^{3}}}\\{\mathrm{Department~of~Computer~Science,~Cornell~University,~USA}}\\{\mathrm{School~of~Information~Systems,~Singapore~Management~University,~Singapore}}\end{array}$ $^3$ College of Information and Computer Sciences, University of Massachusetts Amherst, USA
$^{4}$Department of Computer Science, Mount Holyoke College, USA
xw458@cornell.edu
$\bar{\text{akshatkumar@smu.edu.sg}}$
$\{{\mathrm{~sheldon,shlomo~}}\}{\mathrm{~@cs.umass.edu}}$
Abstract
Stochastic network design is a general framework for optimizing network connectivity. It has several applications in computational sustainability including spatial conservation planning, pre-disaster network preparation, and river network optimization. A common assumption in previous work has been made that network parameters (e.g., probability of species colonization) are precisely known, which is unrealistic in real-world settings. We therefore address the robust river network design problem where the goal is to optimize river connectivity for fish movement by removing barriers. We assume that fsh passability probabilities are known only imprecisely, but are within some interval bounds. We then develop a planning approach that computes the policies with either high robust ratio or low regret .Empirically, our approach scales well to large river networks. We also provide insights into the solutions generated by our robust approach, which has significantly higher robust ratio than the baseline solution with mean parameter estimates.
## 1 Introduction
题，其中物种的移动概率和栖息地的大小没有准确指定。我们解决的健壮网络设计问题与以前的工作不同， 以前的工作不允许管理操作修改间隔参数(e。g。边缘概率)。他们只修改网络结构，例如，通过添加源或节点。相反，我们允许可以修改间隔边界和网络结构的管理操作。由于更丰富的动作空间，目前尚不清楚以前设置中使用的样本平均近似(SAA)方法$\mathbb{E}($Kumar et al. 2016)是否适用于我们的问题。为了解决这些挑战，我们开发了一种基于动态规划和混合整数规划的方法，可以在不使用 SAA 的情况下优化连接
我们研究了树状结构河流网络的鲁棒 SNDP。激励应用程序是障碍清除问题$\overline{\mathrm{E}}($Neeson et al.2015),其目标是决定要清除或修复哪些人流障碍，以帮助鱼类向上游移动并进入其历史栖息地。在这个域中，屏障的通过概率只能被不准确地估计，修复后的屏障的新通过概率更难估计。因此，我们使用众所周知的区间界限E (Boutilier et al. 2003) 对通过概率的不确定性进行建模。然后，我们开发了一种可扩展的算法，以找到强大的障碍消除策略
策略的健壮性可以通过两个相关指标来量化：健壮性比率$\underline{\mathrm{E}}($He and Kempe 2014; Chen et al. 2016) 和遗憾$\boxed{\mathrm{E}}($Boutilier et al. 2003; Kumar et al. 2016)。直觉上， 假设给定一个策略，自然选择一个在其区间范围内选择参数的对抗策略，以便最小化给定策略和对抗策略的值之间的比率 (称为稳健比率), 或者最大化它们之间的价值差异 (称为后悔)。我们开发了一种可扩展的算法，以找到一种健壮的策略，该策略通过解决双层优化问题来最大化健壮比率。我们还表明，只要稍加修改，我们的方法就可以用来最大程度地减少遗憾
该算法基于约束生成过程E$( Boutilier et al. 2003)$ 该过程在两个优化步骤之间进行交织。决策优化步骤找到一个决策策略，当自然可以从给定的有限数量的选择中选择策略和概率时，该决策策略可以最大化稳健比率。在第二个比率最小化步骤中，为选定的决策策略找到最佳对抗策略和概率，并将其添加到自然的选择集中。我们为决策优化问题提供了混合整数线性规划公式。比率最小化问题要困难得多；我们通过结合动态规划算法和舍人方法来开发一种称为舍人动态规划(RDP)的算法，并证明它是一种完全多项式时间近似模式 (FPTAS)。在实验中，我们表明 RDP 在选择对抗性策略和概率时表现几乎最佳。我们的算法可以找到比基线方法在两个健壮性指标上都更健壮的策略。我们还通过可视化解决方案提供有关健壮性指标的见解
许多问题，例如影响最大化$\mathbb{E}($Kempe, Kleinberg, and Tardos 2003) ,空间和鱼类保护计划$\hat{\mathrm{E}}($Sheldon et al. 2010; O'Hanley and Tomberlin 2005)以及灾前准备$\overline{\mathrm{E}}$ (Schichl and Sellmann 2015) ,都可以表述为随机网络设计问题的变体。随机网络设计问题(SNDP)由有向图定义，其中每个边以一定的概率存在或不存在。可以采取管理措施来更改边缘存在的概率。目标是根据预算确定采取哪些措施，以优化随机网络在一段时间内的某些结果。几种解决 SNDP 的方法已被证明可以扩展到大型网络 (Chen, Wang, and Wang 2010; Kumar, Wu, and Zilberstein 2012; Wu, Sheldon, and Zilberstein 2014b; 2016)
在$^{\prime}$SNDP 中做出的一个重要假设是，网络参数(e。g。, 边缘存在的概率)被准确地估计，由于嘈杂的观测、 模型漂移、气候变化和物种的多样性，这在现实世界的生态领域是不可行的。为了处理参数不确定性，研究人员制定了稳健的网络设计问题，其中包括不确定的网络概率 (He and Kempe 2014; Chen et al. 2016)。最近， Kumar et al. (2016) 还研究了一个稳健的保护规划问
Copyright $\odot2017$, Association for the Advancement of Ar-
tificial Intelligence (www.aaai.org). All rights reserved.
|
[
[
0.1544189453125,
0.12060546875,
0.84326171875,
0.121826171875,
0.84326171875,
0.1363525390625,
0.1544189453125,
0.1351318359375
],
[
0.1641845703125,
0.154052734375,
0.2900390625,
0.15283203125,
0.2900390625,
0.1673583984375,
0.1641845703125,
0.1685791015625
],
[
0.3095703125,
0.154052734375,
0.453369140625,
0.15283203125,
0.453369140625,
0.1673583984375,
0.3095703125,
0.1685791015625
],
[
0.48046875,
0.154052734375,
0.66650390625,
0.1514892578125,
0.66650390625,
0.1666259765625,
0.48046875,
0.1685791015625
],
[
0.6640625,
0.1546630859375,
0.83251953125,
0.1546630859375,
0.83251953125,
0.1685791015625,
0.6640625,
0.1685791015625
],
[
0.453369140625,
0.1572265625,
0.484375,
0.1572265625,
0.484375,
0.1641845703125,
0.453369140625,
0.1641845703125
],
[
0.283447265625,
0.1558837890625,
0.310546875,
0.1572265625,
0.310546875,
0.1641845703125,
0.283447265625,
0.1634521484375
],
[
0.3095703125,
0.1578369140625,
0.314453125,
0.1578369140625,
0.314453125,
0.1634521484375,
0.3095703125,
0.1634521484375
],
[
0.27783203125,
0.17236328125,
0.71826171875,
0.1749267578125,
0.71826171875,
0.189453125,
0.27783203125,
0.1875
],
[
0.214111328125,
0.1875,
0.78173828125,
0.1900634765625,
0.78173828125,
0.2052001953125,
0.214111328125,
0.2020263671875
],
[
0.16259765625,
0.2008056640625,
0.83251953125,
0.20263671875,
0.83251953125,
0.2215576171875,
0.16259765625,
0.2197265625
],
[
0.261474609375,
0.2171630859375,
0.734375,
0.2183837890625,
0.734375,
0.23291015625,
0.261474609375,
0.231689453125
],
[
0.178955078125,
0.2342529296875,
0.3203125,
0.2342529296875,
0.3203125,
0.2474365234375,
0.178955078125,
0.2474365234375
],
[
0.34716796875,
0.23291015625,
0.5400390625,
0.23486328125,
0.5400390625,
0.248779296875,
0.34716796875,
0.246826171875
],
[
0.5703125,
0.2342529296875,
0.81787109375,
0.2342529296875,
0.81787109375,
0.2474365234375,
0.5703125,
0.2474365234375
],
[
0.25,
0.2734375,
0.314453125,
0.2734375,
0.314453125,
0.28857421875,
0.25,
0.28857421875
],
[
0.521484375,
0.274658203125,
0.9111328125,
0.274658203125,
0.9111328125,
0.28857421875,
0.521484375,
0.28857421875
],
[
0.521484375,
0.28857421875,
0.9111328125,
0.28857421875,
0.9111328125,
0.3017578125,
0.521484375,
0.3017578125
],
[
0.10296630859375,
0.296630859375,
0.461669921875,
0.296630859375,
0.461669921875,
0.311279296875,
0.10296630859375,
0.311279296875
],
[
0.51953125,
0.3017578125,
0.9111328125,
0.300537109375,
0.9111328125,
0.31494140625,
0.51953125,
0.31640625
],
[
0.10455322265625,
0.311279296875,
0.459228515625,
0.311279296875,
0.459228515625,
0.3212890625,
0.10455322265625,
0.3212890625
],
[
0.101318359375,
0.32080078125,
0.4599609375,
0.322509765625,
0.4599609375,
0.337158203125,
0.101318359375,
0.335205078125
],
[
0.521484375,
0.31640625,
0.9111328125,
0.31640625,
0.9111328125,
0.330078125,
0.521484375,
0.330078125
],
[
0.10296630859375,
0.335205078125,
0.4599609375,
0.335205078125,
0.4599609375,
0.348388671875,
0.10296630859375,
0.348388671875
],
[
0.52294921875,
0.330078125,
0.9111328125,
0.330078125,
0.9111328125,
0.343505859375,
0.52294921875,
0.343505859375
],
[
0.10455322265625,
0.348388671875,
0.455078125,
0.348388671875,
0.455078125,
0.359130859375,
0.10455322265625,
0.359130859375
],
[
0.52294921875,
0.343505859375,
0.9111328125,
0.343505859375,
0.9111328125,
0.356689453125,
0.52294921875,
0.356689453125
],
[
0.10296630859375,
0.361083984375,
0.459228515625,
0.361083984375,
0.459228515625,
0.371826171875,
0.10296630859375,
0.371826171875
],
[
0.52294921875,
0.356689453125,
0.9111328125,
0.356689453125,
0.9111328125,
0.37060546875,
0.52294921875,
0.37060546875
],
[
0.10296630859375,
0.373779296875,
0.461669921875,
0.373779296875,
0.461669921875,
0.386962890625,
0.10296630859375,
0.386962890625
],
[
0.52294921875,
0.371826171875,
0.900390625,
0.371826171875,
0.900390625,
0.385009765625,
0.52294921875,
0.385009765625
],
[
0.89453125,
0.373779296875,
0.91015625,
0.373779296875,
0.91015625,
0.38330078125,
0.89453125,
0.38330078125
],
[
0.10296630859375,
0.3837890625,
0.461669921875,
0.3837890625,
0.461669921875,
0.397705078125,
0.10296630859375,
0.397705078125
],
[
0.52294921875,
0.3837890625,
0.91259765625,
0.3837890625,
0.91259765625,
0.3984375,
0.52294921875,
0.3984375
],
[
0.101318359375,
0.397705078125,
0.463134765625,
0.396484375,
0.463134765625,
0.410888671875,
0.101318359375,
0.412353515625
],
[
0.52294921875,
0.40087890625,
0.77685546875,
0.40087890625,
0.77685546875,
0.410888671875,
0.52294921875,
0.410888671875
],
[
0.10296630859375,
0.410400390625,
0.459228515625,
0.410400390625,
0.459228515625,
0.423583984375,
0.10296630859375,
0.423583984375
],
[
0.10296630859375,
0.42236328125,
0.461669921875,
0.42236328125,
0.461669921875,
0.436279296875,
0.10296630859375,
0.436279296875
],
[
0.5361328125,
0.41845703125,
0.91259765625,
0.419189453125,
0.91259765625,
0.434326171875,
0.5361328125,
0.43310546875
],
[
0.10296630859375,
0.43505859375,
0.461669921875,
0.43505859375,
0.461669921875,
0.448974609375,
0.10296630859375,
0.448974609375
],
[
0.521484375,
0.431884765625,
0.91259765625,
0.43310546875,
0.91259765625,
0.447509765625,
0.521484375,
0.4462890625
],
[
0.10455322265625,
0.449462890625,
0.459228515625,
0.449462890625,
0.459228515625,
0.460205078125,
0.10455322265625,
0.460205078125
],
[
0.52294921875,
0.4462890625,
0.91259765625,
0.4462890625,
0.91259765625,
0.460205078125,
0.52294921875,
0.460205078125
],
[
0.10296630859375,
0.4609375,
0.461669921875,
0.4609375,
0.461669921875,
0.474853515625,
0.10296630859375,
0.474853515625
],
[
0.52392578125,
0.462158203125,
0.91015625,
0.462158203125,
0.91015625,
0.472900390625,
0.52392578125,
0.472900390625
],
[
0.101318359375,
0.472900390625,
0.461669921875,
0.473388671875,
0.461669921875,
0.4873046875,
0.101318359375,
0.486083984375
],
[
0.52294921875,
0.474853515625,
0.9111328125,
0.474853515625,
0.9111328125,
0.488037109375,
0.52294921875,
0.488037109375
],
[
0.10540771484375,
0.486083984375,
0.461669921875,
0.486083984375,
0.461669921875,
0.5,
0.10540771484375,
0.5
],
[
0.52294921875,
0.488037109375,
0.91259765625,
0.488037109375,
0.91259765625,
0.501953125,
0.52294921875,
0.501953125
],
[
0.10296630859375,
0.498779296875,
0.4599609375,
0.498779296875,
0.4599609375,
0.51220703125,
0.10296630859375,
0.51220703125
],
[
0.51953125,
0.501953125,
0.91015625,
0.50244140625,
0.91015625,
0.5166015625,
0.51953125,
0.51513671875
],
[
0.10296630859375,
0.51318359375,
0.455810546875,
0.51318359375,
0.455810546875,
0.5234375,
0.10296630859375,
0.5234375
],
[
0.521484375,
0.5166015625,
0.9111328125,
0.5166015625,
0.9111328125,
0.52978515625,
0.521484375,
0.52978515625
],
[
0.10455322265625,
0.52587890625,
0.4599609375,
0.52587890625,
0.4599609375,
0.5361328125,
0.10455322265625,
0.5361328125
],
[
0.52294921875,
0.5322265625,
0.60302734375,
0.5322265625,
0.60302734375,
0.54345703125,
0.52294921875,
0.54345703125
],
[
0.10455322265625,
0.53857421875,
0.314453125,
0.53857421875,
0.314453125,
0.548828125,
0.10455322265625,
0.548828125
],
[
0.53759765625,
0.5498046875,
0.91259765625,
0.5498046875,
0.91259765625,
0.56298828125,
0.53759765625,
0.56298828125
],
[
0.240234375,
0.5693359375,
0.351318359375,
0.5712890625,
0.351318359375,
0.58544921875,
0.240234375,
0.58349609375
],
[
0.52294921875,
0.564453125,
0.91259765625,
0.564453125,
0.91259765625,
0.57763671875,
0.52294921875,
0.57763671875
],
[
0.2099609375,
0.5712890625,
0.2467041015625,
0.5712890625,
0.2467041015625,
0.58203125,
0.2099609375,
0.58203125
],
[
0.52392578125,
0.580078125,
0.91015625,
0.580078125,
0.91015625,
0.59033203125,
0.52392578125,
0.59033203125
],
[
0.0882568359375,
0.5888671875,
0.47802734375,
0.5888671875,
0.47802734375,
0.60400390625,
0.0882568359375,
0.60400390625
],
[
0.52294921875,
0.59228515625,
0.91259765625,
0.59228515625,
0.91259765625,
0.60595703125,
0.52294921875,
0.60595703125
],
[
0.0882568359375,
0.60400390625,
0.476318359375,
0.60400390625,
0.476318359375,
0.6171875,
0.0882568359375,
0.6171875
],
[
0.521484375,
0.60595703125,
0.91259765625,
0.60595703125,
0.91259765625,
0.619140625,
0.521484375,
0.619140625
],
[
0.08660888671875,
0.6171875,
0.47802734375,
0.6171875,
0.47802734375,
0.630859375,
0.08660888671875,
0.630859375
],
[
0.52294921875,
0.6220703125,
0.91015625,
0.6220703125,
0.91015625,
0.6318359375,
0.52294921875,
0.6318359375
],
[
0.08905029296875,
0.6318359375,
0.47802734375,
0.6318359375,
0.47802734375,
0.64599609375,
0.08905029296875,
0.64599609375
],
[
0.521484375,
0.6337890625,
0.9111328125,
0.6337890625,
0.9111328125,
0.64794921875,
0.521484375,
0.64794921875
],
[
0.08905029296875,
0.64599609375,
0.47802734375,
0.64599609375,
0.47802734375,
0.6591796875,
0.08905029296875,
0.6591796875
],
[
0.521484375,
0.64794921875,
0.9111328125,
0.64794921875,
0.9111328125,
0.6611328125,
0.521484375,
0.6611328125
],
[
0.0882568359375,
0.6591796875,
0.476318359375,
0.6591796875,
0.476318359375,
0.67236328125,
0.0882568359375,
0.67236328125
],
[
0.521484375,
0.662109375,
0.9111328125,
0.662109375,
0.9111328125,
0.67529296875,
0.521484375,
0.67529296875
],
[
0.0882568359375,
0.67236328125,
0.476318359375,
0.67236328125,
0.476318359375,
0.68603515625,
0.0882568359375,
0.68603515625
],
[
0.521484375,
0.67529296875,
0.78369140625,
0.67529296875,
0.78369140625,
0.689453125,
0.521484375,
0.689453125
],
[
0.0882568359375,
0.6875,
0.476318359375,
0.6875,
0.476318359375,
0.70068359375,
0.0882568359375,
0.70068359375
],
[
0.5390625,
0.6943359375,
0.904296875,
0.6943359375,
0.904296875,
0.70849609375,
0.5390625,
0.70849609375
],
[
0.0882568359375,
0.70068359375,
0.476318359375,
0.70068359375,
0.476318359375,
0.7138671875,
0.0882568359375,
0.7138671875
],
[
0.0882568359375,
0.7138671875,
0.47802734375,
0.7138671875,
0.47802734375,
0.72900390625,
0.0882568359375,
0.72900390625
],
[
0.521484375,
0.708984375,
0.91259765625,
0.708984375,
0.91259765625,
0.72265625,
0.521484375,
0.72265625
],
[
0.521484375,
0.7236328125,
0.9111328125,
0.7236328125,
0.9111328125,
0.7373046875,
0.521484375,
0.7373046875
],
[
0.08905029296875,
0.72998046875,
0.476318359375,
0.72998046875,
0.476318359375,
0.74072265625,
0.08905029296875,
0.74072265625
],
[
0.52294921875,
0.7373046875,
0.9111328125,
0.7373046875,
0.9111328125,
0.75048828125,
0.52294921875,
0.75048828125
],
[
0.0882568359375,
0.74365234375,
0.1258544921875,
0.74365234375,
0.1258544921875,
0.75390625,
0.0882568359375,
0.75390625
],
[
0.10296630859375,
0.755859375,
0.47802734375,
0.755859375,
0.47802734375,
0.76953125,
0.10296630859375,
0.76953125
],
[
0.52294921875,
0.75048828125,
0.9111328125,
0.75048828125,
0.9111328125,
0.7646484375,
0.52294921875,
0.7646484375
],
[
0.08660888671875,
0.77099609375,
0.47802734375,
0.77099609375,
0.47802734375,
0.7841796875,
0.08660888671875,
0.7841796875
],
[
0.521484375,
0.76513671875,
0.91259765625,
0.76513671875,
0.91259765625,
0.77880859375,
0.521484375,
0.77880859375
],
[
0.0882568359375,
0.7841796875,
0.4794921875,
0.7841796875,
0.4794921875,
0.79736328125,
0.0882568359375,
0.79736328125
],
[
0.521484375,
0.77880859375,
0.9111328125,
0.77880859375,
0.9111328125,
0.79248046875,
0.521484375,
0.79248046875
],
[
0.0882568359375,
0.79736328125,
0.476318359375,
0.79736328125,
0.476318359375,
0.81103515625,
0.0882568359375,
0.81103515625
],
[
0.521484375,
0.79248046875,
0.9111328125,
0.79248046875,
0.9111328125,
0.80615234375,
0.521484375,
0.80615234375
],
[
0.521484375,
0.80615234375,
0.9111328125,
0.80615234375,
0.9111328125,
0.8193359375,
0.521484375,
0.8193359375
],
[
0.08905029296875,
0.8125,
0.47802734375,
0.8125,
0.47802734375,
0.82568359375,
0.08905029296875,
0.82568359375
],
[
0.51953125,
0.818359375,
0.91259765625,
0.8193359375,
0.91259765625,
0.83349609375,
0.51953125,
0.83203125
],
[
0.0882568359375,
0.82568359375,
0.47802734375,
0.82568359375,
0.47802734375,
0.8388671875,
0.0882568359375,
0.8388671875
],
[
0.08660888671875,
0.8388671875,
0.47802734375,
0.8388671875,
0.47802734375,
0.85302734375,
0.08660888671875,
0.85302734375
],
[
0.521484375,
0.833984375,
0.9111328125,
0.833984375,
0.9111328125,
0.84765625,
0.521484375,
0.84765625
],
[
0.521484375,
0.84765625,
0.91015625,
0.84765625,
0.91015625,
0.861328125,
0.521484375,
0.861328125
],
[
0.08660888671875,
0.861328125,
0.4755859375,
0.8623046875,
0.4755859375,
0.876953125,
0.08660888671875,
0.87548828125
],
[
0.52294921875,
0.8623046875,
0.9111328125,
0.8623046875,
0.9111328125,
0.87548828125,
0.52294921875,
0.87548828125
],
[
0.08660888671875,
0.87548828125,
0.44775390625,
0.87548828125,
0.44775390625,
0.8896484375,
0.08660888671875,
0.8896484375
],
[
0.521484375,
0.87548828125,
0.8798828125,
0.87548828125,
0.8798828125,
0.8896484375,
0.521484375,
0.8896484375
]
] |
[
"Robust Optimization for Tree-Structured Stochastic Network Design",
"Xiaojian Wul",
"Akshat Kumar2",
"Daniel Sheldon3,4",
" Shlomo Zilberstein?",
"",
"",
"",
"1 Department of Computer Science, Cornell University, USA",
"2 School of Information Systems, Singapore Management University, Singapore",
"3 College of Information and Computer Sciences, University of Massachusetts Amherst, USA",
"4 Department of Computer Science, Mount Holyoke College, USA",
"xw458@cornell.edu",
"akshatkumar@smu.edu.sg",
"{ sheldon,shlomo } @cs.umass.edu",
"Abstract",
"题，其中物种的移动概率和栖息地的大小没有准确指",
"定。我们解决的健壮网络设计问题与以前的工作不同",
"Stochastic network design is a general framework for",
"以前的工作不允许管理操作修改间隔参数（e。g。、边缘",
"optimizing network connectivity.It hasseveral appli-",
"cations in computational sustainability including spa-",
"概率）。他们只修改网络结构，例如，通过添加源或节",
"tial conservation planning, pre-disaster network prepa",
"点。相反，我们允许可以修改间隔边界和网络结构的管",
"ration, and river network optimization. A common as",
"理操作。由于更丰富的动作空间，目前尚不清楚以前设",
"sumptioninprevious work has been madethatnetwork",
"置中使用的样本平均近似（SAA）方法E(Kumaretal.",
"parameters (e.g., probability of species colonization)",
"2016）是否适用于我们的问题。为了解决这些挑战，",
"我",
"are precisely known, which is unrealistic in real-world",
"们开发了一种基于动态规划和混合整数规划的方法，可",
"settings. We therefore address the robust river network",
"以在不使用SAA的情况下优化连接",
"design problem where the goal is to optimize river con",
"nectivity for fish movement by removing barriers. We",
"我们研究了树状结构河流网络的鲁棒SNDP。激励",
"assume that fish passability probabilities are known",
"应用程序是障碍清除问题E(Neeson etal.2015），其",
"onlyimprecisely,but are within some interval bounds",
"目标是决定要清除或修复哪些入流障碍，以帮助鱼类",
"We then develop a planning approach that computes",
"向上游移动并进入其历史栖息地。在这个域中，屏障的",
"the policies with either high robust ratio or low regret",
"通过概率只能被不准确地估计，修复后的屏障的新通过",
"Empirically, our approach scales well to large river",
"概率更难估计。因此，我们使用众所周知的区间界限E",
"networks. We also provide insights into the solutions",
"(Boutilieretal.2003）对通过概率的不确定性进行建模。",
"generated by our robust approach,whichhas signifi",
"然后，我们开发了一种可扩展的算法，以找到强大的障",
"cantlyhigher robust ratio than the baseline solution",
"碍消除策略",
"with mean parameter estimates",
"策略的健壮性可以通过两个相关指标来量化：健壮",
"Introduction",
"性比率E(He and Kempe 2014;Chen etal.2016)和遗",
"1",
"感F(Boutilieretal.2003:Kumaretal.2016)。直觉上",
"许多问题，例如影响最大化E(Kempe，Kleinberg，and",
"假设给定一个策略，自然选择一个在其区间范围内选择",
"Tardos2003），空间和鱼类保护计划E(Sheldonetal.",
"参数的对抗策略，以便最小化给定策略和对抗策略的值",
"2010：O'HanleyandTomberlin2005）以及灾前准备E",
"之间的比率（称为稳健比率），或者最大化它们之间的价",
"(SchichlandSellmann2015），都可以表述为随机网络",
"值差异（称为后悔)。我们开发了一种可扩展的算法，以",
"设计问题的变体。随机网络设计问题（SNDP）由有向图",
"找到一种健壮的策略，该策略通过解决双层优化问题来",
"定义，其中每个边以一定的概率存在或不存在。可以采",
"最大化健壮比率。我们还表明，只要稍加修改，我们的",
"取管理措施来更改边缘存在的概率。目标是根据预算确",
"方法就可以用来最大程度地减少遗憾",
"定采取哪些措施，以优化随机网络在一段时间内的某些",
"该算法基于约束生成过程E(Boutilier etal.2003)",
"结果。儿种解决SNDP的方法已被证明可以扩展到大型",
"网络 (Chen, Wang, and Wang 2010; Kumar, Wu, and",
"该过程在两个优化步骤之间进行交织。决策优化步骤",
"找到一个决策策略，当自然可以从给定的有限数量的",
"Zilberstein2012:Wu,Sheldon,andZilberstein2014b",
"选择中选择策略和概率时，该决策策略可以最大化稳",
"2016)",
"在SNDP中做出的一个重要假设是，网络参数（e。",
"健比率。在第二个比率最小化步骤中，为选定的决策策",
"g。，边缘存在的概率）被准确地估计，由于噜杂的观测、",
"略找到最佳对抗策略和概率，并将其添加到自然的选",
"模型漂移、气候变化和物种的多样性，这在现实世界的",
"择集中。我们为决策优化问题提供了混合整数线性规",
"生态领域是不可行的。为了处理参数不确定性，研究人",
"划公式。比率最小化问题要困难得多；我们通过结合动",
"态规划算法和舍人方法来开发一种称为舍入动态规划",
"员制定了稳健的网络设计问题，其中包括不确定的网络",
"（RDP）的算法，并证明它是一种完全多项式时间近似",
"概率 (He and Kempe 2014;Chen et al.2016)。最近",
"Kumaretal.(2016）还研究了一个稳健的保护规划问",
"模式（FPTAS）。在实验中，我们表明RDP在选择对",
"抗性策略和概率时表现几乎最佳。我们的算法可以找到",
"Copyright 2017, Association for the Advancement of Ar",
"比基线方法在两个健壮性指标上都更健壮的策略。我们",
"tificial Intelligence (www.aaai.org). All rights reserved.",
"还通过可视化解决方案提供有关健壮性指标的见解"
] |
{
"arxiv_id": "1612.00104",
"language": "zh",
"page_idx": 1
}
|
||
2311.15980-zh-1
|
2311.15980-zh-1.jpg
|
方面，模型适应于 2。5D 信息。另一方面，借助向自关注层注人多视图信息的结构修改来捕获联合多视图分布。在推理过程中，由 DDIME[35] 等常见调度器同步生成多视图图像，然后通过可微分光栅化将其直接融合到网格中。整个生成过程在几秒钟内完成，这比通常需要 30 分钟的基于 SDS 的方法要快得多。该系统通过复杂的文本提示进行了广泛的评估，并与基于 SDS 和直接 3D 生成方法进行了比较，证明了生成具有复杂几何形状，多样性和高保真度的 3D 纹理网格的能力
总而言之，本文的主要贡献包括：
$\bullet$ We propose to approach the 3D generation task
by training a multi-view 2.5D diffusion model, which explicitly models the 3D geometry distribution while inheriting a strong generalization ability of the large-scale pre-trained 2D image diffusion.
$\bullet$ We introduce an efficient differentiable rasterization
scheme to optimize a textured mesh directly from
the multi-view normal maps and RGB images.
$\bullet$ We carefully design a generation pipeline that achieves diverse, mode-seeking-free, and highfidelity 3D content generation in only 10 seconds.
2. Related Work
2.1. 3D Generation by Score Distillation
评分蒸馏[28,39]是最近通过预训练的 2D 扩散模型进行 3D 生成的最流行方法之一。它将图像去噪的知识提炼到可微分渲染系统的优化过程中，以便随机渲染的视图逐渐细化以描述输入的文本提示。存在基本问题：1) 二维扩散模型不具有三维感知能力，生成的样本因此存在多面问题；2) 每个优化步骤都需要去噪 UNet 的单次转发，使得整个过程耗时；3) 为了更好的收敛，优选高引导比例的提示，这导致外观过饱和；4)优化是寻模的，失去了二维扩散模型的强多样性。提出了解决其中一些问题的后续工作，但不是全部。Zero-1-to-3[16]使用多视图数据集对 2D 扩散模型进行微调，以授予透视控制能力并减轻图像到 3D 任务中的问题 1。ProlificDreamer [43]通过利用 KL 发散损耗执行采样而不是模式搜索来缓解问题 3 和 4,但代价是时间复杂度更高。在这项工作中，我们不应用分数蒸馏和完全分离的扩散过程和 3D 模型优化。扩散可以正常调度和调节，使结果具有多样性和现实色彩。并且 3D 模型优化在显式表示上运行，因此可以快速完成
## 2.2. Direct 3D Diffusion
可以通过使用 3D 数据集训练直接 3D 扩散模型来实现快速 3D 生成。一个关键问题是选择 3D 表示并为其设计特殊的编码器/解码器。有一些早期尝试训练点云[22,26,46,49] ,网格[18]和隐式表示(如 NeRF 或SDF [5,10,12,34] )的直接 3D 模型。但是，它们是在有限的数据集上训练的，例如 ShapeNet[3],这些数据集具有相当小的数据大小，几何复杂性或类别多样性。最近的 3D 数据集，例如 Objaverse[7],极大
地改善了 3D 数据集的最新技术，但与用于训练 2D 扩散模型的 2D 图像字幕数据集相比，仍然受到限制。在这项工作中，我们仍然使用 2D 神经网络来处理 2。5D 映射，因此我们可以对现有的 2D 扩散模型进行微调， 以便继承它们的强泛化
# 2.3. Multi-view Diffusion
同时生成多视图图像是将 3D 感知引人 2D 扩散模型的另一种策略。提出了两个关键的修改来实现这一点： 1) 来自其他视图的信息与当前视图连接，作为自关注层中的键和查询。收集的信息可以来自单个投影[36] ,对极线[17,37]或所有像素[33] ; 2) 该模型在来自Objaverse [7]等 3D 数据集的多视图渲染上进行了微调。为了构建 3D 模型，以前的工作要么使用仍然耗时的 SDS [33] ,要么使用像 NeuS [17,19,41]这样的基于图像的重建系统，该系统至少需要 10 个视图才能产生合理的重建。类似于探索 2 的 JointNet [48] 。5D 域， 我们选择生成多视图 2。5D 像普通地图一样，这样我们就可以使用无 SDS 重建，同时仍然保持小视图数
## 3. Method
在本节中，我们将介绍我们的多视图 2。5D 扩散系统， 同步生成多视图 2。5D 几何图像，i。e。、法线贴图和相应的纹理贴图，并给出文本提示作为 3D 内容生成的输入(图。1)。我们的方法足够高效，可以在 10 秒内生成各种结果。在 SEC。3.1 , 我们首先简要回顾了2D 扩散模型，并制定了多视图 2。5D 适配。然后，我们说明了跨视图注意力，它增强了 Sec 中的多视图一致性。3.2 。在 SEC。3.3 , 我们描述如何从生成的 2生成最终的 3D 模型。5D 几何图像，最后在秒。3.4 , 我们演示了如何在给定生成的法线贴图的情况下合成纹理贴图，并构建高质量的最终纹理三角形网格
3.1. Diffusion Models and 2.5D Adaptation
扩散模型学习从各向同性高斯分布到目标分布的转换(e。g。图像空间) 通过迭代去噪操作。我们在潜在扩散模型 (LDM) 上构建我们的系统，该模型包含一个包含编码器和解码器的变分自动编码器 (VAE),一个去噪网络和一个条件输入编码器。与原始扩散模型相比， LDM 在潜像空间中进行整个扩散过程，大大提高了效率和质量。具体来说，在正向过程中，在时间$t$处的噪声潜在在潜在空间中被采样，并逐渐被噪声降级，这使得它与高斯噪声无法区分，而去噪过程则反转该过程， 迭代地预测和去除噪声以获得真实图像
在这项工作中，我们扩展了 2D 文本到图像扩散模型，以生成多视图几何图像。通过使用我们的 2 微调预先训练的 2D 扩散模型。5D 图像数据集，我们能够继承泛化，并获得多视图 2 的表达生成能力。5D 几何图像。令$(\boldsymbol{X},c)$为训练数据集中带有标题的 3D 数据， $x_i\in\mathcal{X}$为多视图渲染，$x_i,t$为在$t$时被独立噪声$\epsilon_i\in\mathcal{E}$
2
|
[
[
0.08251953125,
0.09344482421875,
0.468017578125,
0.09344482421875,
0.468017578125,
0.106689453125,
0.08251953125,
0.106689453125
],
[
0.50341796875,
0.0947265625,
0.8896484375,
0.0947265625,
0.8896484375,
0.10479736328125,
0.50341796875,
0.10479736328125
],
[
0.08251953125,
0.10986328125,
0.46728515625,
0.10986328125,
0.46728515625,
0.12060546875,
0.08251953125,
0.12060546875
],
[
0.5048828125,
0.10986328125,
0.8896484375,
0.10986328125,
0.8896484375,
0.12060546875,
0.5048828125,
0.12060546875
],
[
0.08087158203125,
0.12371826171875,
0.46728515625,
0.12371826171875,
0.46728515625,
0.136962890625,
0.08087158203125,
0.136962890625
],
[
0.5048828125,
0.125,
0.8896484375,
0.125,
0.8896484375,
0.1351318359375,
0.5048828125,
0.1351318359375
],
[
0.08087158203125,
0.1395263671875,
0.468017578125,
0.1395263671875,
0.468017578125,
0.15283203125,
0.08087158203125,
0.15283203125
],
[
0.50244140625,
0.136962890625,
0.89208984375,
0.1383056640625,
0.89208984375,
0.15283203125,
0.50244140625,
0.1514892578125
],
[
0.08251953125,
0.154052734375,
0.468017578125,
0.154052734375,
0.468017578125,
0.1673583984375,
0.08251953125,
0.1673583984375
],
[
0.50341796875,
0.154052734375,
0.66845703125,
0.154052734375,
0.66845703125,
0.1673583984375,
0.50341796875,
0.1673583984375
],
[
0.08251953125,
0.1685791015625,
0.46728515625,
0.1685791015625,
0.46728515625,
0.1817626953125,
0.08251953125,
0.1817626953125
],
[
0.08251953125,
0.184326171875,
0.46728515625,
0.184326171875,
0.46728515625,
0.1976318359375,
0.08251953125,
0.1976318359375
],
[
0.50244140625,
0.1875,
0.701171875,
0.1875,
0.701171875,
0.2008056640625,
0.50244140625,
0.2008056640625
],
[
0.08251953125,
0.1988525390625,
0.46728515625,
0.1988525390625,
0.46728515625,
0.212158203125,
0.08251953125,
0.212158203125
],
[
0.08087158203125,
0.2139892578125,
0.455810546875,
0.2139892578125,
0.455810546875,
0.2279052734375,
0.08087158203125,
0.2279052734375
],
[
0.5048828125,
0.21337890625,
0.890625,
0.21337890625,
0.890625,
0.2266845703125,
0.5048828125,
0.2266845703125
],
[
0.101318359375,
0.2310791015625,
0.336669921875,
0.2310791015625,
0.336669921875,
0.2412109375,
0.101318359375,
0.2412109375
],
[
0.5048828125,
0.2291259765625,
0.890625,
0.2291259765625,
0.890625,
0.242431640625,
0.5048828125,
0.242431640625
],
[
0.0849609375,
0.246826171875,
0.468017578125,
0.246826171875,
0.468017578125,
0.261474609375,
0.0849609375,
0.261474609375
],
[
0.5048828125,
0.24560546875,
0.8896484375,
0.24560546875,
0.8896484375,
0.255615234375,
0.5048828125,
0.255615234375
],
[
0.1004638671875,
0.261474609375,
0.468017578125,
0.260009765625,
0.468017578125,
0.274658203125,
0.1004638671875,
0.27587890625
],
[
0.5048828125,
0.261474609375,
0.8896484375,
0.261474609375,
0.8896484375,
0.271484375,
0.5048828125,
0.271484375
],
[
0.10296630859375,
0.27587890625,
0.465576171875,
0.276611328125,
0.465576171875,
0.290283203125,
0.10296630859375,
0.2890625
],
[
0.5185546875,
0.2734375,
0.88818359375,
0.2734375,
0.88818359375,
0.287353515625,
0.5185546875,
0.287353515625
],
[
0.10455322265625,
0.292236328125,
0.468017578125,
0.292236328125,
0.468017578125,
0.30615234375,
0.10455322265625,
0.30615234375
],
[
0.50341796875,
0.287353515625,
0.890625,
0.2890625,
0.890625,
0.3037109375,
0.50341796875,
0.3017578125
],
[
0.50341796875,
0.302978515625,
0.890625,
0.3037109375,
0.890625,
0.318115234375,
0.50341796875,
0.317626953125
],
[
0.10455322265625,
0.309326171875,
0.4453125,
0.309326171875,
0.4453125,
0.3193359375,
0.10455322265625,
0.3193359375
],
[
0.0849609375,
0.322509765625,
0.46728515625,
0.322509765625,
0.46728515625,
0.3359375,
0.0849609375,
0.3359375
],
[
0.50341796875,
0.3193359375,
0.890625,
0.3193359375,
0.890625,
0.333251953125,
0.50341796875,
0.333251953125
],
[
0.101318359375,
0.337158203125,
0.468017578125,
0.3359375,
0.468017578125,
0.35107421875,
0.101318359375,
0.3515625
],
[
0.50244140625,
0.333251953125,
0.8896484375,
0.333984375,
0.8896484375,
0.347900390625,
0.50244140625,
0.3466796875
],
[
0.101318359375,
0.3515625,
0.4345703125,
0.3515625,
0.4345703125,
0.365478515625,
0.101318359375,
0.365478515625
],
[
0.50341796875,
0.349853515625,
0.88818359375,
0.349853515625,
0.88818359375,
0.363037109375,
0.50341796875,
0.363037109375
],
[
0.0849609375,
0.367431640625,
0.468017578125,
0.367431640625,
0.468017578125,
0.3818359375,
0.0849609375,
0.3818359375
],
[
0.50634765625,
0.3662109375,
0.890625,
0.3662109375,
0.890625,
0.376953125,
0.50634765625,
0.376953125
],
[
0.10540771484375,
0.3837890625,
0.465576171875,
0.3837890625,
0.465576171875,
0.39453125,
0.10540771484375,
0.39453125
],
[
0.50341796875,
0.380126953125,
0.87158203125,
0.380126953125,
0.87158203125,
0.393310546875,
0.50341796875,
0.393310546875
],
[
0.10296630859375,
0.397705078125,
0.455810546875,
0.397705078125,
0.455810546875,
0.410888671875,
0.10296630859375,
0.410888671875
],
[
0.50341796875,
0.419189453125,
0.5986328125,
0.419189453125,
0.5986328125,
0.43115234375,
0.50341796875,
0.43115234375
],
[
0.080078125,
0.4287109375,
0.2279052734375,
0.4287109375,
0.2279052734375,
0.442626953125,
0.080078125,
0.442626953125
],
[
0.5048828125,
0.4462890625,
0.890625,
0.4462890625,
0.890625,
0.460205078125,
0.5048828125,
0.460205078125
],
[
0.080078125,
0.453857421875,
0.390625,
0.453857421875,
0.390625,
0.46728515625,
0.080078125,
0.46728515625
],
[
0.5048828125,
0.462158203125,
0.890625,
0.462158203125,
0.890625,
0.47607421875,
0.5048828125,
0.47607421875
],
[
0.08087158203125,
0.4765625,
0.468017578125,
0.4765625,
0.468017578125,
0.490478515625,
0.08087158203125,
0.490478515625
],
[
0.5048828125,
0.479248046875,
0.890625,
0.479248046875,
0.890625,
0.4892578125,
0.5048828125,
0.4892578125
],
[
0.08087158203125,
0.492431640625,
0.46728515625,
0.492431640625,
0.46728515625,
0.505859375,
0.08087158203125,
0.505859375
],
[
0.50341796875,
0.4912109375,
0.890625,
0.4912109375,
0.890625,
0.5048828125,
0.50341796875,
0.5048828125
],
[
0.08251953125,
0.5087890625,
0.46728515625,
0.5087890625,
0.46728515625,
0.51953125,
0.08251953125,
0.51953125
],
[
0.50341796875,
0.5068359375,
0.8896484375,
0.5068359375,
0.8896484375,
0.52099609375,
0.50341796875,
0.52099609375
],
[
0.08087158203125,
0.52294921875,
0.468017578125,
0.52294921875,
0.468017578125,
0.5361328125,
0.08087158203125,
0.5361328125
],
[
0.50341796875,
0.521484375,
0.89208984375,
0.521484375,
0.89208984375,
0.53515625,
0.50341796875,
0.53515625
],
[
0.0882568359375,
0.5390625,
0.46728515625,
0.5390625,
0.46728515625,
0.5498046875,
0.0882568359375,
0.5498046875
],
[
0.5048828125,
0.537109375,
0.88623046875,
0.537109375,
0.88623046875,
0.55029296875,
0.5048828125,
0.55029296875
],
[
0.08251953125,
0.5537109375,
0.46728515625,
0.5537109375,
0.46728515625,
0.564453125,
0.08251953125,
0.564453125
],
[
0.50341796875,
0.5517578125,
0.89208984375,
0.5517578125,
0.89208984375,
0.56494140625,
0.50341796875,
0.56494140625
],
[
0.08087158203125,
0.568359375,
0.46728515625,
0.568359375,
0.46728515625,
0.58203125,
0.08087158203125,
0.58203125
],
[
0.50341796875,
0.5673828125,
0.88818359375,
0.5673828125,
0.88818359375,
0.58056640625,
0.50341796875,
0.58056640625
],
[
0.08251953125,
0.58349609375,
0.468017578125,
0.58349609375,
0.468017578125,
0.5966796875,
0.08251953125,
0.5966796875
],
[
0.50341796875,
0.58349609375,
0.89208984375,
0.58349609375,
0.89208984375,
0.5966796875,
0.50341796875,
0.5966796875
],
[
0.08251953125,
0.5966796875,
0.465576171875,
0.5966796875,
0.465576171875,
0.60986328125,
0.08251953125,
0.60986328125
],
[
0.50244140625,
0.5966796875,
0.84619140625,
0.5986328125,
0.84619140625,
0.6123046875,
0.50244140625,
0.60986328125
],
[
0.08251953125,
0.6123046875,
0.468017578125,
0.6123046875,
0.468017578125,
0.62548828125,
0.08251953125,
0.62548828125
],
[
0.08251953125,
0.62744140625,
0.468017578125,
0.62744140625,
0.468017578125,
0.6416015625,
0.08251953125,
0.6416015625
],
[
0.50341796875,
0.6318359375,
0.83642578125,
0.6318359375,
0.83642578125,
0.64599609375,
0.50341796875,
0.64599609375
],
[
0.08251953125,
0.64306640625,
0.46728515625,
0.64306640625,
0.46728515625,
0.6572265625,
0.08251953125,
0.6572265625
],
[
0.08087158203125,
0.65771484375,
0.468017578125,
0.65771484375,
0.468017578125,
0.671875,
0.08087158203125,
0.671875
],
[
0.50341796875,
0.65771484375,
0.8896484375,
0.65771484375,
0.8896484375,
0.671875,
0.50341796875,
0.671875
],
[
0.08251953125,
0.6748046875,
0.46728515625,
0.6748046875,
0.46728515625,
0.68505859375,
0.08251953125,
0.68505859375
],
[
0.5048828125,
0.673828125,
0.890625,
0.673828125,
0.890625,
0.6875,
0.5048828125,
0.6875
],
[
0.08251953125,
0.68798828125,
0.468017578125,
0.68798828125,
0.468017578125,
0.7021484375,
0.08251953125,
0.7021484375
],
[
0.50244140625,
0.68603515625,
0.890625,
0.6875,
0.890625,
0.7021484375,
0.50244140625,
0.70068359375
],
[
0.08251953125,
0.705078125,
0.46728515625,
0.705078125,
0.46728515625,
0.71533203125,
0.08251953125,
0.71533203125
],
[
0.5048828125,
0.70263671875,
0.890625,
0.70263671875,
0.890625,
0.71630859375,
0.5048828125,
0.71630859375
],
[
0.08416748046875,
0.72021484375,
0.46728515625,
0.72021484375,
0.46728515625,
0.73095703125,
0.08416748046875,
0.73095703125
],
[
0.50341796875,
0.71826171875,
0.88818359375,
0.71826171875,
0.88818359375,
0.73193359375,
0.50341796875,
0.73193359375
],
[
0.08087158203125,
0.73486328125,
0.259033203125,
0.73486328125,
0.259033203125,
0.74560546875,
0.08087158203125,
0.74560546875
],
[
0.50341796875,
0.73193359375,
0.890625,
0.73291015625,
0.890625,
0.7470703125,
0.50341796875,
0.74560546875
],
[
0.5048828125,
0.74853515625,
0.890625,
0.74853515625,
0.890625,
0.76220703125,
0.5048828125,
0.76220703125
],
[
0.080078125,
0.7587890625,
0.268798828125,
0.7587890625,
0.268798828125,
0.77294921875,
0.080078125,
0.77294921875
],
[
0.50341796875,
0.76220703125,
0.89208984375,
0.76318359375,
0.89208984375,
0.77783203125,
0.50341796875,
0.7763671875
],
[
0.08251953125,
0.7822265625,
0.468017578125,
0.7822265625,
0.468017578125,
0.79541015625,
0.08251953125,
0.79541015625
],
[
0.5048828125,
0.77880859375,
0.89208984375,
0.77880859375,
0.89208984375,
0.79248046875,
0.5048828125,
0.79248046875
],
[
0.08087158203125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.08087158203125,
0.81005859375
],
[
0.50341796875,
0.79443359375,
0.77783203125,
0.79443359375,
0.77783203125,
0.8046875,
0.50341796875,
0.8046875
],
[
0.08416748046875,
0.81298828125,
0.465576171875,
0.81298828125,
0.465576171875,
0.82373046875,
0.08416748046875,
0.82373046875
],
[
0.52392578125,
0.8125,
0.890625,
0.8125,
0.890625,
0.82568359375,
0.52392578125,
0.82568359375
],
[
0.08087158203125,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.08087158203125,
0.84033203125
],
[
0.50341796875,
0.8271484375,
0.8896484375,
0.8271484375,
0.8896484375,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.080078125,
0.84033203125,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.080078125,
0.85400390625
],
[
0.5048828125,
0.84228515625,
0.890625,
0.84228515625,
0.890625,
0.85595703125,
0.5048828125,
0.85595703125
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85595703125,
0.468017578125,
0.869140625,
0.080078125,
0.87060546875
],
[
0.5048828125,
0.859375,
0.8896484375,
0.859375,
0.8896484375,
0.869140625,
0.5048828125,
0.869140625
],
[
0.08251953125,
0.87255859375,
0.468017578125,
0.87255859375,
0.468017578125,
0.88623046875,
0.08251953125,
0.88623046875
],
[
0.50341796875,
0.87255859375,
0.890625,
0.87255859375,
0.890625,
0.88623046875,
0.50341796875,
0.88623046875
],
[
0.080078125,
0.88623046875,
0.468017578125,
0.88720703125,
0.468017578125,
0.9013671875,
0.080078125,
0.90087890625
],
[
0.50244140625,
0.88623046875,
0.89208984375,
0.88720703125,
0.89208984375,
0.9013671875,
0.50244140625,
0.90087890625
],
[
0.48046875,
0.92529296875,
0.49267578125,
0.92529296875,
0.49267578125,
0.93798828125,
0.48046875,
0.93798828125
]
] |
[
"方面，模型适应于2。5D信息。另一方面，借助向自关",
"地改善了3D数据集的最新技术，但与用于训练2D扩",
"注层注入多视图信息的结构修改来捕获联合多视图分",
"散模型的2D图像字幕数据集相比，仍然受到限制。在",
"布。在推理过程中，由DDIME[35]等常见调度器同步",
"这项工作中，我们仍然使用2D神经网络来处理2。5D",
"生成多视图图像，然后通过可微分光栅化将其直接融",
"映射，因此我们可以对现有的2D扩散模型进行微调，",
"合到网格中。整个生成过程在几秒钟内完成，这比通常",
"以便继承它们的强泛化",
"需要30分钟的基于SDS的方法要快得多。该系统通",
"过复杂的文本提示进行了广泛的评估，并与基于SDS",
"2.3. Multi-view Diffusion",
"和直接3D生成方法进行了比较，证明了生成具有复杂",
"几何形状，多样性和高保真度的3D纹理网格的能力",
"同时生成多视图图像是将3D感知引入2D扩散模型",
"总而言之，本文的主要贡献包括：",
"的另一种策略。提出了两个关键的修改来实现这一点",
" We propose to approach the 3D generation task",
"1）来自其他视图的信息与当前视图连接，作为自关注",
"by training a multi-view 2.5D diffusion model,",
"层中的键和查询。收集的信息可以来自单个投影36",
"which explicitly models the 3D geometry distribu-",
"对极线【17，37]或所有像素【33]；2】该模型在来自",
"tion while inheriting a strong generalization ability",
"Objaverse[7等3D数据集的多视图谊染上进行了微",
"调。为了构建3D模型，以前的工作要么使用仍然耗时",
"ofthelarge-scalepre-trained2Dimagediffusion",
".We introduce an efficient differentiable rasterization",
"的SDS[33]，要么使用像NeuS[17，19，41]这样的基",
"scheme to optimize a textured mesh directly from",
"于图像的重建系统，该系统至少需要10个视图才能产",
"the multi-view normal maps and RGB images.",
"生合理的重建。类似于探索2的JointNet48]。5D域",
". We carefully design a generation pipeline that",
"我们选择生成多视图2。5D像普通地图一样，这样我",
"achievesdiverse,mode-seeking-free,andhigh-",
"们就可以使用无SDS重建，同时仍然保持小视图数",
"fidelity 3D content generation in only 10 seconds.",
"3.Method",
"2.Related Work",
"在本节中，我们将介绍我们的多视图2。5D扩散系统，",
"2.1. 3D Generation by Score Distillation",
"同步生成多视图2。5D几何图像，i。e。法线贴图和",
"评分蒸馏[28，39]是最近通过预训练的2D扩散模型进",
"相应的纹理贴图，并给出文本提示作为3D内容生成",
"行3D生成的最流行方法之一。它将图像去噪的知识提",
"的输人（图。1）。我们的方法足够高效，可以在10秒",
"炼到可微分渣染系统的优化过程中，以便随机遣染的视",
"内生成各种结果。在SEC。3.1，我们首先简要回顾了",
"图逐渐细化以描述输入的文本提示。存在基本问题：1)",
"2D扩散模型，并制定了多视图2。5D适配。然后，我",
"维扩散模型不具有三维感知能力，生成的样本因此存",
"们说明了跨视图注意力，它增强了Sec中的多视图一",
"在多面问题：2）每个优化步骤都需要去噪UNet的单次",
"致性。3.2。在SEC。3.3，我们描述如何从生成的2",
"转发，使得整个过程耗时；3）为了更好的收敛，优选高引",
"生成最终的3D模型。5D几何图像，最后在秒。3.4",
"导比例的提示，这导致外观过饱和；4）优化是寻模的，失",
"我们演示了如何在给定生成的法线贴图的情况下合成",
"去了二维扩散模型的强多样性。提出了解决其中一些问",
"纹理贴图，并构建高质量的最终纹理三角形网格",
"题的后续工作，但不是全部。Zero-1-to-3[16]使用多视",
"图数据集对2D扩散模型进行微调，以授予透视控制能",
"3.1. Diffusion Models and 2.5D Adaptation",
"力并减轻图像到3D任务中的问题1。ProlificDreamer",
"[43]通过利用KL发散损耗执行采样而不是模式搜索",
"扩散模型学习从各向同性高斯分布到目标分布的转换",
"来缓解问题3和4，但代价是时间复杂度更高。在这项",
"(e。g。图像空间）通过迭代去噪操作。我们在潜在扩散",
"工作中，我们不应用分数蒸馏和完全分离的扩散过程",
"模型（LDM）上构建我们的系统，该模型包含一个包",
"和3D模型优化。扩散可以正常调度和调节，使结果具",
"含编码器和解码器的变分自动编码器（VAE），一个去",
"有多样性和现实色彩。并且3D模型优化在显式表示上",
"噪网络和一个条件输入编码器。与原始扩散模型相比",
"运行，因此可以快速完成",
"LDM在潜像空间中进行整个扩散过程，大大提高了效",
"率和质量。具体来说，在正向过程中，在时间t处的噪",
"2.2. Direct 3D Diffusion",
"声潜在在潜在空间中被采样，并逐渐被噪声降级，这使",
"可以通过使用3D数据集训练直接3D扩散模型来实",
"得它与高斯噪声无法区分，而去噪过程则反转该过程，",
"现快速3D生成。一个关键问题是选择3D表示并为其",
"迭代地预测和去除噪声以获得真实图像",
"设计特殊的编码器/解码器。有一些早期尝试训练点云",
"在这项工作中，我们扩展了2D文本到图像扩散模",
"[22，26，46，49]，网格18]和隐式表示（如NeRF或",
"型，以生成多视图几何图像。通过使用我们的2微调",
"SDF[5，10，12，341）的直接3D模型。但是，它们是",
"预先训练的2D扩散模型。5D图像数据集，我们能够",
"在有限的数据集上训练的，例如ShapeNet[3]，这些",
"继承泛化，并获得多视图2的表达生成能力。5D几何",
"数据集具有相当小的数据大小，几何复杂性或类别多",
"图像。令（X,c）为训练数据集中带有标题的3D数据",
"样性。最近的3D数据集，例如Objaverse[7]，极大",
"aiEx为多视图煊染，ai,t为在t时被独立噪声eE8",
"2"
] |
{
"arxiv_id": "2311.15980",
"language": "zh",
"page_idx": 1
}
|
||
2311.15980-zh-2
|
2311.15980-zh-2.jpg
|
<FigureHere>
Figure 1. Overview of our text-to-3D content generation system. The generation is a two-stage process, frst generating
geoemtry and then appearance. Specifically, the system is composed of the following steps: 1) a single denoising process to simultaneously generate 4 normal maps; 2) fast mesh optimization by differentiable rasterization; 3) a single denoising process to generate 4 images conditioned on rendered normal maps; 4) texture construction from multi-view images. The whole generation process only takes 10 seconds.
损坏的视图。用
$$L=\mathbb{E}_{(\boldsymbol{X},c);\boldsymbol{E}\sim N(0,1);t}\sum_{x_{i}\in\boldsymbol{X};\epsilon_{i}\in\boldsymbol{E}}\|\epsilon_{i}-\epsilon_{\theta}(x_{i,t},c,t)\|_{2}^{2}.$$
(1)
训练去噪神经网络$\epsilon_\mathrm{\theta}$
# 3.2. Cross-view Attention
在实践中，我们从不同的视点(前，左，右，后)生成与世界坐标对齐的 4 个法线贴图。为了将这些观察结果融合到三角形网格中，我们通过可微分光栅化从初始化网格中探索几何优化的见解。这种独立于神经网络推理的优化在几秒钟内快速实现收敛(参见 Alg。1 (1)
空间雕刻初始化。一种简单直接的方法是使用基本的几何图元(如球体和立方体)初始化形状，并优化。然而，这通常会在后面的几何优化过程中引人重大挑战， 特别是当目标形状的拓扑与这些基本形式明显不同时。为了应对这一挑战，我们采用空间雕刻算法[13]进行形状拓扑初始化。此外，它还为后面的几何优化提供了一个很好的初始化。Fig。2(a)显示了空间雕刻结果。具体来说，这个过程开始通过一个简单的值阈值分离背景法线贴图。随后，在感兴趣的空间中创建体积，并使用相机参数将每个体素投影到图像上，从而确定相应的像素是对象的一部分还是背景。通过收集不同视图下的所有投影，我们构造了一个占用体积，其中如果体素的所有投影都属于背景，则将体素的占用设置为 0 (表示空置),否则设置为 1(表示占用)。最后，我们在占用体积上应用行进立方体[20]来提取零水平集表面，以形成初始化形状。该技术不仅有效地保留了拓扑结构，而且还提供了从多视图法线图像生成的粗略形状估计
通过可微分光栅化优化。一旦我们获得了初始化的几何形状，我们将根据观测数据进一步细化网格细节。
在进行微调之前，不能保证从基础模型为同一文本提示生成的多个图像描述同一对象，因为它们是从不同的噪声图发起的，并且是独立降噪的。我们使用类似于[33] 的解决方案：我们在扩散过程之间添加数据通信， 并在多视图图像数据集上微调模型以学习多视图条件。在实现方面，我们同步所有的扩散过程。当计算到达自关注层时，我们将所有中间结果收集为查询和值，而不是仅使用当前分支的结果。由于图像被视为顺序输入， 因此可以简单地将附加信息连接在一起，而无需引人更多可训练的参数。这种架构确保扩散过程是相互制约的，这是多视图一致生成的结构先决条件
# 3.3. Explicit Multi-view 2.5D Fusion
有多种方法可用于从多视图观察构建 3D 模型。其中， 基于图像的三维重建方法如多视角立体[9,44,45,47] 或 NeRF [1,8,24,25]需要至少 10 幅图像进行高保真重建，这在多视角扩散场景中带来了巨大的计算挑战。然而，通过从 2。5D 信息，可以有效地降低这一要求。
3
|
[
[
0.09722900390625,
0.10797119140625,
0.5,
0.10797119140625,
0.5,
0.121826171875,
0.09722900390625,
0.121826171875
],
[
0.491943359375,
0.10797119140625,
0.7490234375,
0.10797119140625,
0.7490234375,
0.121826171875,
0.491943359375,
0.121826171875
],
[
0.400390625,
0.143310546875,
0.414306640625,
0.143310546875,
0.414306640625,
0.1514892578125,
0.400390625,
0.1514892578125
],
[
0.10540771484375,
0.1514892578125,
0.25,
0.1514892578125,
0.25,
0.169189453125,
0.10540771484375,
0.169189453125
],
[
0.2646484375,
0.160400390625,
0.27685546875,
0.160400390625,
0.27685546875,
0.1666259765625,
0.2646484375,
0.1666259765625
],
[
0.50732421875,
0.1666259765625,
0.57421875,
0.1666259765625,
0.57421875,
0.1817626953125,
0.50732421875,
0.1817626953125
],
[
0.2181396484375,
0.18310546875,
0.32177734375,
0.18310546875,
0.32177734375,
0.2259521484375,
0.2181396484375,
0.2259521484375
],
[
0.431396484375,
0.18115234375,
0.439453125,
0.1817626953125,
0.438720703125,
0.186279296875,
0.4306640625,
0.1849365234375
],
[
0.5009765625,
0.1881103515625,
0.57666015625,
0.1881103515625,
0.57666015625,
0.199462890625,
0.5009765625,
0.199462890625
],
[
0.0931396484375,
0.1976318359375,
0.12091064453125,
0.1976318359375,
0.12091064453125,
0.2052001953125,
0.0931396484375,
0.2052001953125
],
[
0.10455322265625,
0.205810546875,
0.136474609375,
0.205810546875,
0.136474609375,
0.2171630859375,
0.10455322265625,
0.2171630859375
],
[
0.51708984375,
0.2008056640625,
0.55810546875,
0.2008056640625,
0.55810546875,
0.2108154296875,
0.51708984375,
0.2108154296875
],
[
0.37744140625,
0.21337890625,
0.398681640625,
0.21337890625,
0.398681640625,
0.2279052734375,
0.37744140625,
0.2279052734375
],
[
0.416748046875,
0.212158203125,
0.44287109375,
0.212158203125,
0.44287109375,
0.231689453125,
0.416748046875,
0.231689453125
],
[
0.1405029296875,
0.242431640625,
0.339111328125,
0.242431640625,
0.339111328125,
0.255615234375,
0.1405029296875,
0.255615234375
],
[
0.336669921875,
0.240478515625,
0.47802734375,
0.242431640625,
0.47802734375,
0.255615234375,
0.336669921875,
0.25390625
],
[
0.64794921875,
0.246826171875,
0.69921875,
0.246826171875,
0.69921875,
0.25390625,
0.64794921875,
0.25390625
],
[
0.49267578125,
0.2587890625,
0.5888671875,
0.260009765625,
0.5888671875,
0.271484375,
0.49267578125,
0.270751953125
],
[
0.11358642578125,
0.27978515625,
0.1568603515625,
0.27978515625,
0.1568603515625,
0.291748046875,
0.11358642578125,
0.291748046875
],
[
0.214111328125,
0.2841796875,
0.273681640625,
0.2841796875,
0.273681640625,
0.294189453125,
0.214111328125,
0.294189453125
],
[
0.483642578125,
0.285400390625,
0.576171875,
0.287353515625,
0.57421875,
0.326904296875,
0.48193359375,
0.324462890625
],
[
0.09478759765625,
0.294189453125,
0.1519775390625,
0.292236328125,
0.15283203125,
0.30810546875,
0.096435546875,
0.31005859375
],
[
0.60888671875,
0.311279296875,
0.62158203125,
0.311279296875,
0.62158203125,
0.322509765625,
0.60888671875,
0.322509765625
],
[
0.67724609375,
0.309326171875,
0.72119140625,
0.309326171875,
0.72119140625,
0.32080078125,
0.67724609375,
0.32080078125
],
[
0.3349609375,
0.313232421875,
0.38818359375,
0.313232421875,
0.38818359375,
0.34033203125,
0.3349609375,
0.34033203125
],
[
0.096435546875,
0.3125,
0.1732177734375,
0.31005859375,
0.1732177734375,
0.32568359375,
0.096435546875,
0.32763671875
],
[
0.44287109375,
0.314453125,
0.453369140625,
0.317626953125,
0.44775390625,
0.32763671875,
0.43701171875,
0.324462890625
],
[
0.60546875,
0.324462890625,
0.625,
0.324462890625,
0.625,
0.337158203125,
0.60546875,
0.337158203125
],
[
0.6806640625,
0.32080078125,
0.72119140625,
0.32080078125,
0.72119140625,
0.335205078125,
0.6806640625,
0.335205078125
],
[
0.11517333984375,
0.333251953125,
0.1748046875,
0.333251953125,
0.1748046875,
0.343505859375,
0.11517333984375,
0.343505859375
],
[
0.486083984375,
0.32763671875,
0.57421875,
0.32763671875,
0.57421875,
0.3642578125,
0.486083984375,
0.3642578125
],
[
0.1004638671875,
0.333984375,
0.11358642578125,
0.333984375,
0.11358642578125,
0.343505859375,
0.1004638671875,
0.343505859375
],
[
0.66259765625,
0.341552734375,
0.736328125,
0.34033203125,
0.736328125,
0.35595703125,
0.66259765625,
0.356689453125
],
[
0.09722900390625,
0.349853515625,
0.11767578125,
0.349853515625,
0.11767578125,
0.3623046875,
0.09722900390625,
0.3623046875
],
[
0.11273193359375,
0.349853515625,
0.16015625,
0.349853515625,
0.16015625,
0.35986328125,
0.11273193359375,
0.35986328125
],
[
0.223876953125,
0.354736328125,
0.250732421875,
0.356689453125,
0.25,
0.365478515625,
0.22216796875,
0.363037109375
],
[
0.32275390625,
0.352783203125,
0.398681640625,
0.352783203125,
0.398681640625,
0.367431640625,
0.32275390625,
0.367431640625
],
[
0.1168212890625,
0.36865234375,
0.1641845703125,
0.36865234375,
0.1641845703125,
0.375732421875,
0.1168212890625,
0.375732421875
],
[
0.473876953125,
0.367431640625,
0.60302734375,
0.367431640625,
0.60302734375,
0.380615234375,
0.473876953125,
0.380615234375
],
[
0.240234375,
0.375,
0.44287109375,
0.375,
0.44287109375,
0.38818359375,
0.240234375,
0.38818359375
],
[
0.78759765625,
0.371826171875,
0.8486328125,
0.371826171875,
0.8486328125,
0.38330078125,
0.78759765625,
0.38330078125
],
[
0.56640625,
0.3818359375,
0.640625,
0.3818359375,
0.640625,
0.399658203125,
0.56640625,
0.399658203125
],
[
0.29736328125,
0.38818359375,
0.3857421875,
0.38818359375,
0.3857421875,
0.3984375,
0.29736328125,
0.3984375
],
[
0.76953125,
0.386474609375,
0.86669921875,
0.386474609375,
0.86669921875,
0.40087890625,
0.76953125,
0.40087890625
],
[
0.07843017578125,
0.410888671875,
0.89208984375,
0.412353515625,
0.89208984375,
0.4267578125,
0.07843017578125,
0.425537109375
],
[
0.080078125,
0.425537109375,
0.89208984375,
0.425537109375,
0.89208984375,
0.43994140625,
0.080078125,
0.43994140625
],
[
0.080078125,
0.43994140625,
0.89208984375,
0.43994140625,
0.89208984375,
0.453857421875,
0.080078125,
0.453857421875
],
[
0.07843017578125,
0.45263671875,
0.89208984375,
0.453857421875,
0.89208984375,
0.468505859375,
0.07843017578125,
0.46728515625
],
[
0.080078125,
0.46728515625,
0.396240234375,
0.46728515625,
0.396240234375,
0.48095703125,
0.080078125,
0.48095703125
],
[
0.08087158203125,
0.494384765625,
0.1968994140625,
0.494384765625,
0.1968994140625,
0.50830078125,
0.08087158203125,
0.50830078125
],
[
0.50341796875,
0.494384765625,
0.890625,
0.494384765625,
0.890625,
0.50830078125,
0.50341796875,
0.50830078125
],
[
0.5048828125,
0.50830078125,
0.890625,
0.5087890625,
0.890625,
0.5234375,
0.5048828125,
0.52294921875
],
[
0.09478759765625,
0.521484375,
0.2467041015625,
0.5263671875,
0.2467041015625,
0.54150390625,
0.09478759765625,
0.5361328125
],
[
0.26953125,
0.52294921875,
0.283447265625,
0.52294921875,
0.283447265625,
0.5322265625,
0.26953125,
0.5322265625
],
[
0.3095703125,
0.521484375,
0.449462890625,
0.52294921875,
0.449462890625,
0.537109375,
0.3095703125,
0.5361328125
],
[
0.5048828125,
0.5244140625,
0.890625,
0.5244140625,
0.890625,
0.53857421875,
0.5048828125,
0.53857421875
],
[
0.2459716796875,
0.54052734375,
0.31201171875,
0.54052734375,
0.31201171875,
0.5517578125,
0.2459716796875,
0.5517578125
],
[
0.5048828125,
0.54052734375,
0.890625,
0.54052734375,
0.890625,
0.5537109375,
0.5048828125,
0.5537109375
],
[
0.44677734375,
0.5537109375,
0.468017578125,
0.5537109375,
0.468017578125,
0.5673828125,
0.44677734375,
0.5673828125
],
[
0.50244140625,
0.5537109375,
0.86669921875,
0.55615234375,
0.86669921875,
0.5693359375,
0.50244140625,
0.5673828125
],
[
0.87353515625,
0.56103515625,
0.8798828125,
0.56103515625,
0.8798828125,
0.564453125,
0.87353515625,
0.564453125
],
[
0.88134765625,
0.56005859375,
0.88818359375,
0.56005859375,
0.88818359375,
0.56494140625,
0.88134765625,
0.56494140625
],
[
0.08087158203125,
0.58056640625,
0.2320556640625,
0.58056640625,
0.2320556640625,
0.5947265625,
0.08087158203125,
0.5947265625
],
[
0.5048828125,
0.59228515625,
0.890625,
0.59228515625,
0.890625,
0.60595703125,
0.5048828125,
0.60595703125
],
[
0.07843017578125,
0.60302734375,
0.281982421875,
0.60400390625,
0.281982421875,
0.61865234375,
0.07843017578125,
0.6171875
],
[
0.5048828125,
0.60791015625,
0.890625,
0.60791015625,
0.890625,
0.6220703125,
0.5048828125,
0.6220703125
],
[
0.08087158203125,
0.62744140625,
0.468017578125,
0.62744140625,
0.468017578125,
0.6416015625,
0.08087158203125,
0.6416015625
],
[
0.50341796875,
0.62255859375,
0.890625,
0.62255859375,
0.890625,
0.63623046875,
0.50341796875,
0.63623046875
],
[
0.08087158203125,
0.64306640625,
0.468017578125,
0.64306640625,
0.468017578125,
0.6572265625,
0.08087158203125,
0.6572265625
],
[
0.50341796875,
0.63818359375,
0.89208984375,
0.63818359375,
0.89208984375,
0.6513671875,
0.50341796875,
0.6513671875
],
[
0.08087158203125,
0.65771484375,
0.46728515625,
0.65771484375,
0.46728515625,
0.671875,
0.08087158203125,
0.671875
],
[
0.5048828125,
0.65283203125,
0.890625,
0.65283203125,
0.890625,
0.66650390625,
0.5048828125,
0.66650390625
],
[
0.08087158203125,
0.673828125,
0.468017578125,
0.673828125,
0.468017578125,
0.6875,
0.08087158203125,
0.6875
],
[
0.50341796875,
0.66845703125,
0.88818359375,
0.66845703125,
0.88818359375,
0.681640625,
0.50341796875,
0.681640625
],
[
0.08087158203125,
0.68798828125,
0.4697265625,
0.68798828125,
0.4697265625,
0.7021484375,
0.08087158203125,
0.7021484375
],
[
0.51708984375,
0.6845703125,
0.88623046875,
0.6845703125,
0.88623046875,
0.69775390625,
0.51708984375,
0.69775390625
],
[
0.08087158203125,
0.7041015625,
0.46728515625,
0.7041015625,
0.46728515625,
0.71728515625,
0.08087158203125,
0.71728515625
],
[
0.5048828125,
0.69873046875,
0.8896484375,
0.69873046875,
0.8896484375,
0.7119140625,
0.5048828125,
0.7119140625
],
[
0.08251953125,
0.71826171875,
0.468017578125,
0.71826171875,
0.468017578125,
0.73193359375,
0.08251953125,
0.73193359375
],
[
0.5048828125,
0.71337890625,
0.890625,
0.71337890625,
0.890625,
0.7265625,
0.5048828125,
0.7265625
],
[
0.08087158203125,
0.73193359375,
0.468017578125,
0.73291015625,
0.468017578125,
0.74755859375,
0.08087158203125,
0.7470703125
],
[
0.5048828125,
0.72802734375,
0.890625,
0.72802734375,
0.890625,
0.7412109375,
0.5048828125,
0.7412109375
],
[
0.08087158203125,
0.74853515625,
0.465576171875,
0.74853515625,
0.465576171875,
0.76220703125,
0.08087158203125,
0.76220703125
],
[
0.5048828125,
0.74365234375,
0.890625,
0.74365234375,
0.890625,
0.7568359375,
0.5048828125,
0.7568359375
],
[
0.08087158203125,
0.76220703125,
0.468017578125,
0.76318359375,
0.468017578125,
0.77783203125,
0.08087158203125,
0.7763671875
],
[
0.5048828125,
0.7587890625,
0.890625,
0.7587890625,
0.890625,
0.77294921875,
0.5048828125,
0.77294921875
],
[
0.08087158203125,
0.77783203125,
0.3896484375,
0.77783203125,
0.3896484375,
0.79150390625,
0.08087158203125,
0.79150390625
],
[
0.5048828125,
0.7734375,
0.89208984375,
0.7734375,
0.89208984375,
0.787109375,
0.5048828125,
0.787109375
],
[
0.50341796875,
0.7890625,
0.890625,
0.7890625,
0.890625,
0.80322265625,
0.50341796875,
0.80322265625
],
[
0.080078125,
0.80322265625,
0.363525390625,
0.80322265625,
0.363525390625,
0.81640625,
0.080078125,
0.81640625
],
[
0.5048828125,
0.8037109375,
0.89208984375,
0.8037109375,
0.89208984375,
0.8173828125,
0.5048828125,
0.8173828125
],
[
0.5048828125,
0.8193359375,
0.8896484375,
0.8193359375,
0.8896484375,
0.83349609375,
0.5048828125,
0.83349609375
],
[
0.08087158203125,
0.82568359375,
0.468017578125,
0.82568359375,
0.468017578125,
0.8388671875,
0.08087158203125,
0.8388671875
],
[
0.50341796875,
0.833984375,
0.890625,
0.833984375,
0.890625,
0.84765625,
0.50341796875,
0.84765625
],
[
0.08087158203125,
0.84130859375,
0.46728515625,
0.84130859375,
0.46728515625,
0.85498046875,
0.08087158203125,
0.85498046875
],
[
0.08087158203125,
0.85693359375,
0.46728515625,
0.85693359375,
0.46728515625,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.5048828125,
0.85107421875,
0.55224609375,
0.85107421875,
0.55224609375,
0.8623046875,
0.5048828125,
0.8623046875
],
[
0.08251953125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.08251953125,
0.88525390625
],
[
0.52392578125,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.52392578125,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50341796875,
0.88525390625,
0.89208984375,
0.88623046875,
0.89208984375,
0.90087890625,
0.50341796875,
0.8994140625
],
[
0.47802734375,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
">-: Generate a high-quality mesh of \"a DSLR photo of ",
"a pirate collie dog, high resolution'",
"S",
"360-MLP",
"nce",
" 2~3 sec.",
"司",
"",
"Fast Meshing",
"Geon",
"Noise",
"(Sec. 3.3)",
"工",
"2",
"Multi-view Normal Diffusion Model",
"Multi-viewNormalMaps",
"Geometry",
"Normal Rendering",
"Legend",
"Texture Noise",
"1",
"①Concat.",
"4x",
"Texture",
"oD",
"VAE deeoder",
"D",
"SR",
"Fusion",
"4xSuperRes.",
"大n",
"面",
"~1.5 sec.",
"O",
"Pos. Enco",
"Promp",
"Text Encoder",
"Cam.Parg",
"Multi-view RGB Images",
"Multi-view Normal Conditioned RGB",
"Textured",
"~2.5 sec.",
"Diffusion Model",
"Triangle Mesh",
"Figure 1. Overview of our text-to-3D content generation system. The generation is a two-stage process, first generating",
"geoemtry and then appearance. Specifically, the system is composed of the following steps: 1) a single denoising process",
"to simultaneously generate 4 normal maps; 2) fast mesh optimization by differentiable rasterization; 3) a single denoising",
"process to generate 4 images conditioned on rendered normal maps; 4) texture construction from multi-view images: The",
"whole generation process only takes 10 seconds.",
"损坏的视图。用",
"在实践中，我们从不同的视点（前，左，右，后）生成",
"与世界坐标对齐的4个法线贴图。为了将这些观察结",
"L =E(X,c);E~N(0,1);t",
"1",
"e-e(it,c,t)",
"果融合到三角形网格中，我们通过可微分光栅化从初",
"TiEX;EE",
"始化网格中探索几何优化的见解。这种独立于神经网",
"(1)",
"络推理的优化在几秒钟内快速实现收敛（参见Alg。",
"",
"",
"训练去噪神经网络Eo",
"空间雕刻初始化。一种简单直接的方法是使用基本的",
"3.2. Cross-view Attention",
"几何图元（如球体和立方体）初始化形状，并优化。然",
"在进行微调之前，不能保证从基础模型为同一文本提",
"而，这通常会在后面的几何优化过程中引人重大挑战，",
"示生成的多个图像描述同一对象，因为它们是从不同",
"特别是当目标形状的拓扑与这些基本形式明显不同时。",
"的噪声图发起的，并且是独立降噪的。我们使用类似于",
"为了应对这一挑战，我们采用空间雕刻算法[13]进行",
"[33】的解决方案：我们在扩散过程之间添加数据通信，",
"形状拓扑初始化。此外，它还为后面的几何优化提供了",
"并在多视图图像数据集上微调模型以学习多视图条件。",
"个很好的初始化。Fig。2（a）显示了空间雕刻结果",
"在实现方面，我们同步所有的扩散过程。当计算到达自",
"具体来说，这个过程开始通过一个简单的值阈值分离",
"关注层时，我们将所有中间结果收集为查询和值，而不",
"背景法线贴图。随后，在感兴趣的空间中创建体积，并",
"是仅使用当前分支的结果。由于图像被视为顺序输入，",
"使用相机参数将每个体素投影到图像上，从而确定相",
"因此可以简单地将附加信息连接在一起，而无需引人",
"应的像素是对象的一部分还是背景。通过收集不同视",
"更多可训练的参数。这种架构确保扩散过程是相互制",
"图下的所有投影，我们构造了一个占用体积，其中如果",
"约的，这是多视图一致生成的结构先决条件",
"体素的所有投影都属于背景，则将体素的占用设置为0",
"（表示空置），否则设置为1（表示占用）。最后，我们",
"3.3.Explicit Multi-view 2.5DFusion",
"在占用体积上应用行进立方体【20】来提取零水平集表",
"面，以形成初始化形状。该技术不仅有效地保留了拓扑",
"有多种方法可用于从多视图观察构建3D模型。其中，",
"结构，而且还提供了从多视图法线图像生成的粗略形",
"基于图像的三维重建方法如多视角立体9，44，45，47]",
"或NeRF[1，8,24,25]需要至少10幅图像进行高保真",
"状估计",
"重建，这在多视角扩散场景中带来了巨大的计算挑战。",
"通过可微分光栅化优化。一旦我们获得了初始化的",
"然而，通过从2。5D信息，可以有效地降低这一要求。",
"几何形状，我们将根据观测数据进一步细化网格细节。",
"3"
] |
{
"arxiv_id": "2311.15980",
"language": "zh",
"page_idx": 2
}
|
||
1812.06280-zh-1
|
1812.06280-zh-1.jpg
|
本文的主要贡献总结如下：
$\cdot{\mathrm{~We~present~Wikipedia2Vec,~a~tool~for~learn-}}$ ing the embeddings of words and entities easily and efficiently from Wikipedia.
$\bullet$ Our tool achieved a state-of-the-art result
on the KORE entity relatedness dataset, and performed competitively on the various benchmark datasets.
$\bullet$ We present a web-based demonstration that
allows users to explore the learned embed-
$\dim$gs.
$\bullet$ We publicize the code, demonstration, and
the pretrained embeddings for 12 languages
at https://wikipedia2vec.github.io .
## 2 Related Work
许多研究最近提出了从 KB (Hu et al., 2015; Li et al., 2016; Tsai and Roth, 2016; Yamada et al., 2016, 2017, 2018a; Cao et al., 2017; Ganea and Hofmann, 2017) 学习实体嵌人的方法。这些嵌人通常基于传统的单词嵌入模型(例如。g。, skip-gram (Mikolov et al., 2013a) )用从 KB 检索的数据训练。例如，Ristoski et al.(2018)提出了RDF2Vec，它使用 skipgram 模型学习实体嵌入，并通过在 Wikidata 和 DBpedia 等大型知识图上随机游走生成输人。此外，在各种研究中广泛使用的一种简单方法 (Yaghoobzadeh and Schutze, 2015; Yamada et al. , 2017, 2018a; Al- Badrashiny et al. 2017; Suzuki et al., 2018) 通过将输入语料库中的实体注释替换为其引用实体的唯一标识符，并将该语料库输入到单词嵌人模型中来训练 实 体 嵌 入 ( e。g。, skip- gram) 。两 个 开 源 工 具 , 即 Wiki2Vec $^{1}$ 和 Wikipedia 实体向量 $^{2}$已经实现了此方法。我们提出的工具基于 Yamada et al.(2016),它通过使用由 Wikipedia 的内部超链接连接的相邻实体作为附加上下文来训练模型，从而扩展了这一思想。请注意，我们在实验中使用 RDF2Vec 和 Wiki2Vec 作为基线，并在 KORE 数据集上实现了比这些工具更高的经验性能。此外，已经提出了各种关系嵌入模型 (Bordes et al., 2013; Wang et al., 2014; Lin et al.,2015), 旨在学习对知识图完成任务特别有效的实体表示
### 3 Overview
Wikipedia2Vec 是一个易于使用的优化工具，用于从 Wikipedia 学习嵌人。可以使用Python 的 pip 工具( pip install wikipedia2vec
$^{1}$https://github.com/idio/wiki2vec
$^2$https://github.com/singletongue/WikiEntVec
<FigureHere>
Figure 1: Shell commands to train embeddings
from the latest English Wikipedia dump.
<FigureHere>
Figure 2: An example that uses the Wikipedia2Vec
embeddings on a Python interactive shell.
)安装此工具。通过使用 Wikipedia 转储文件$^{3}$作为参数运行 wikipedia2vec train 命令，可以轻松学习嵌入。图 1 显示了下载最新的英文Wikipedia 转储文件并使用默认超参数基于此转储运行嵌入训练的 shell 命令。$^{4}$此外，用户可以轻松使用学习的嵌入。图 2 显示了加载学习的嵌入文件的示例 Python 代码，并获得实体 ScarlettJohansson 和单词 tokyo 的嵌入， 以及实体 Python 最相似的单词和实体
### 3.1 Model
Wikipedia2Vec 实现了传统的 skip-gram 模型(Mikolov et al., 2013a,b) 及其在 Yamada et al. (2016)中提出的扩展，以将单词和实体映射到相同的$d$维向量空间。skip-gram 模型是一种神经网络模型，其训练目标是找到对预测上下文项有用的嵌入(i。e。, 本文中的单词或实体) 给定每个项目。模型的损失函数定义为：
(1)
$$\mathcal{L}_{s}=-\sum_{o_{i}\in O}\sum_{o_{c}\in C_{o_{i}}}\log P(o_{c}|o_{i}),$$
其中$O$ 是所有项的集合(i。e。,单词或实体), $C_o$是$o$的上下文项集，条件概率$\log P(o_c|o_i)$ 是使用以下 softmax 函数定义的：
(2)
$$P(o_c|o_i)=\frac{\exp({V_{o_i}}^\top U_{o_c})}{\sum_{o\in O}\exp({V_{o_i}}^\top U_o)},$$
$^{3}$The dump fle can be downloaded at Wikimedia
${\mathrm{Downloads:~https://dumps.wikimedia.org}}$
$^{4}$The train command has many optional hyper-
parameters that are described in detail in the docu-
mentation.
|
[
[
0.1376953125,
0.07659912109375,
0.34765625,
0.07659912109375,
0.34765625,
0.09027099609375,
0.1376953125,
0.09027099609375
],
[
0.568359375,
0.07659912109375,
0.869140625,
0.07659912109375,
0.869140625,
0.09027099609375,
0.568359375,
0.09027099609375
],
[
0.1175537109375,
0.0926513671875,
0.48779296875,
0.0926513671875,
0.48779296875,
0.10687255859375,
0.1175537109375,
0.10687255859375
],
[
0.603515625,
0.08905029296875,
0.880859375,
0.08905029296875,
0.880859375,
0.09979248046875,
0.603515625,
0.09979248046875
],
[
0.609375,
0.09979248046875,
0.63623046875,
0.09979248046875,
0.63623046875,
0.10809326171875,
0.609375,
0.10809326171875
],
[
0.1351318359375,
0.1092529296875,
0.489501953125,
0.1092529296875,
0.489501953125,
0.122314453125,
0.1351318359375,
0.122314453125
],
[
0.58251953125,
0.105712890625,
0.861328125,
0.10687255859375,
0.861328125,
0.12054443359375,
0.58251953125,
0.119384765625
],
[
0.5498046875,
0.1175537109375,
0.74072265625,
0.1175537109375,
0.74072265625,
0.127685546875,
0.5498046875,
0.127685546875
],
[
0.1351318359375,
0.12469482421875,
0.411376953125,
0.12469482421875,
0.411376953125,
0.1383056640625,
0.1351318359375,
0.1383056640625
],
[
0.1175537109375,
0.141357421875,
0.484375,
0.141357421875,
0.484375,
0.1519775390625,
0.1175537109375,
0.1519775390625
],
[
0.5146484375,
0.15087890625,
0.8818359375,
0.15087890625,
0.8818359375,
0.1644287109375,
0.5146484375,
0.1644287109375
],
[
0.1351318359375,
0.15673828125,
0.48779296875,
0.15673828125,
0.48779296875,
0.171630859375,
0.1351318359375,
0.171630859375
],
[
0.51123046875,
0.1644287109375,
0.81787109375,
0.1650390625,
0.81787109375,
0.1787109375,
0.51123046875,
0.1776123046875
],
[
0.1351318359375,
0.17333984375,
0.489501953125,
0.17333984375,
0.489501953125,
0.18701171875,
0.1351318359375,
0.18701171875
],
[
0.1351318359375,
0.189453125,
0.306396484375,
0.189453125,
0.306396484375,
0.203125,
0.1351318359375,
0.203125
],
[
0.5087890625,
0.198974609375,
0.81787109375,
0.198974609375,
0.81787109375,
0.2120361328125,
0.5087890625,
0.2120361328125
],
[
0.11419677734375,
0.204833984375,
0.48779296875,
0.204833984375,
0.48779296875,
0.218505859375,
0.11419677734375,
0.218505859375
],
[
0.5146484375,
0.210205078125,
0.81201171875,
0.210205078125,
0.81201171875,
0.2203369140625,
0.5146484375,
0.2203369140625
],
[
0.133544921875,
0.2197265625,
0.489501953125,
0.2203369140625,
0.489501953125,
0.235107421875,
0.133544921875,
0.2340087890625
],
[
0.51318359375,
0.2197265625,
0.85791015625,
0.2197265625,
0.85791015625,
0.2298583984375,
0.51318359375,
0.2298583984375
],
[
0.52294921875,
0.2286376953125,
0.62744140625,
0.2286376953125,
0.62744140625,
0.2392578125,
0.52294921875,
0.2392578125
],
[
0.63232421875,
0.2286376953125,
0.69189453125,
0.2286376953125,
0.69189453125,
0.2392578125,
0.63232421875,
0.2392578125
],
[
0.74462890625,
0.2298583984375,
0.85205078125,
0.2298583984375,
0.85205078125,
0.240478515625,
0.74462890625,
0.240478515625
],
[
0.1368408203125,
0.236328125,
0.187255859375,
0.2381591796875,
0.185546875,
0.2529296875,
0.1351318359375,
0.25048828125
],
[
0.515625,
0.2392578125,
0.56494140625,
0.2392578125,
0.56494140625,
0.2464599609375,
0.515625,
0.2464599609375
],
[
0.5634765625,
0.2410888671875,
0.60107421875,
0.2410888671875,
0.60107421875,
0.2452392578125,
0.5634765625,
0.2452392578125
],
[
0.60205078125,
0.2381591796875,
0.74462890625,
0.2381591796875,
0.74462890625,
0.248779296875,
0.60205078125,
0.248779296875
],
[
0.5146484375,
0.2476806640625,
0.69287109375,
0.2476806640625,
0.69287109375,
0.25830078125,
0.5146484375,
0.25830078125
],
[
0.71875,
0.2476806640625,
0.85205078125,
0.2476806640625,
0.85205078125,
0.25830078125,
0.71875,
0.25830078125
],
[
0.5146484375,
0.257080078125,
0.73486328125,
0.257080078125,
0.73486328125,
0.267333984375,
0.5146484375,
0.267333984375
],
[
0.73291015625,
0.25830078125,
0.85986328125,
0.25830078125,
0.85986328125,
0.268310546875,
0.73291015625,
0.268310546875
],
[
0.54833984375,
0.2666015625,
0.74462890625,
0.2666015625,
0.74462890625,
0.27978515625,
0.54833984375,
0.27978515625
],
[
0.1158447265625,
0.2529296875,
0.489501953125,
0.2529296875,
0.489501953125,
0.2666015625,
0.1158447265625,
0.2666015625
],
[
0.133544921875,
0.267333984375,
0.48779296875,
0.267333984375,
0.48779296875,
0.281982421875,
0.133544921875,
0.281982421875
],
[
0.51708984375,
0.276611328125,
0.6884765625,
0.276611328125,
0.6884765625,
0.287353515625,
0.51708984375,
0.287353515625
],
[
0.1351318359375,
0.2861328125,
0.385498046875,
0.2861328125,
0.385498046875,
0.2998046875,
0.1351318359375,
0.2998046875
],
[
0.5224609375,
0.28515625,
0.85400390625,
0.28515625,
0.85400390625,
0.298583984375,
0.5224609375,
0.298583984375
],
[
0.5205078125,
0.29345703125,
0.69189453125,
0.29443359375,
0.69189453125,
0.30810546875,
0.5205078125,
0.30712890625
],
[
0.12005615234375,
0.314208984375,
0.156982421875,
0.314208984375,
0.156982421875,
0.325927734375,
0.12005615234375,
0.325927734375
],
[
0.15283203125,
0.3134765625,
0.27783203125,
0.3134765625,
0.27783203125,
0.32666015625,
0.15283203125,
0.32666015625
],
[
0.5146484375,
0.328857421875,
0.88330078125,
0.328857421875,
0.88330078125,
0.342529296875,
0.5146484375,
0.342529296875
],
[
0.12176513671875,
0.33837890625,
0.489501953125,
0.33837890625,
0.489501953125,
0.3515625,
0.12176513671875,
0.3515625
],
[
0.51318359375,
0.34326171875,
0.82373046875,
0.34326171875,
0.82373046875,
0.35693359375,
0.51318359375,
0.35693359375
],
[
0.1175537109375,
0.352783203125,
0.48779296875,
0.352783203125,
0.48779296875,
0.366455078125,
0.1175537109375,
0.366455078125
],
[
0.1175537109375,
0.369384765625,
0.486083984375,
0.369384765625,
0.486083984375,
0.383056640625,
0.1175537109375,
0.383056640625
],
[
0.118408203125,
0.385498046875,
0.486083984375,
0.385498046875,
0.486083984375,
0.3984375,
0.118408203125,
0.3984375
],
[
0.515625,
0.385498046875,
0.8818359375,
0.385498046875,
0.8818359375,
0.3984375,
0.515625,
0.3984375
],
[
0.12176513671875,
0.402099609375,
0.48779296875,
0.40087890625,
0.48779296875,
0.4150390625,
0.12176513671875,
0.416259765625
],
[
0.51123046875,
0.3984375,
0.8818359375,
0.40087890625,
0.8818359375,
0.4150390625,
0.51123046875,
0.413330078125
],
[
0.12176513671875,
0.418701171875,
0.48779296875,
0.418701171875,
0.48779296875,
0.432861328125,
0.12176513671875,
0.432861328125
],
[
0.515625,
0.41748046875,
0.8818359375,
0.41748046875,
0.8818359375,
0.43115234375,
0.515625,
0.43115234375
],
[
0.12005615234375,
0.43408203125,
0.489501953125,
0.432861328125,
0.489501953125,
0.44775390625,
0.12005615234375,
0.448974609375
],
[
0.51318359375,
0.431640625,
0.8818359375,
0.432861328125,
0.8818359375,
0.44775390625,
0.51318359375,
0.446533203125
],
[
0.12005615234375,
0.45068359375,
0.489501953125,
0.45068359375,
0.489501953125,
0.46435546875,
0.12005615234375,
0.46435546875
],
[
0.51318359375,
0.448974609375,
0.880859375,
0.448974609375,
0.880859375,
0.463134765625,
0.51318359375,
0.463134765625
],
[
0.12005615234375,
0.466064453125,
0.489501953125,
0.465576171875,
0.489501953125,
0.479736328125,
0.12005615234375,
0.48095703125
],
[
0.5146484375,
0.46435546875,
0.8818359375,
0.465576171875,
0.8818359375,
0.479736328125,
0.5146484375,
0.478515625
],
[
0.12005615234375,
0.482177734375,
0.48779296875,
0.482666015625,
0.48779296875,
0.49755859375,
0.12005615234375,
0.496337890625
],
[
0.51318359375,
0.479736328125,
0.8818359375,
0.48095703125,
0.8818359375,
0.495361328125,
0.51318359375,
0.49462890625
],
[
0.12176513671875,
0.49951171875,
0.48779296875,
0.49951171875,
0.48779296875,
0.51318359375,
0.12176513671875,
0.51318359375
],
[
0.5146484375,
0.496337890625,
0.88330078125,
0.49755859375,
0.88330078125,
0.5126953125,
0.5146484375,
0.51123046875
],
[
0.12176513671875,
0.515625,
0.490234375,
0.515625,
0.490234375,
0.529296875,
0.12176513671875,
0.529296875
],
[
0.5146484375,
0.51416015625,
0.8251953125,
0.51416015625,
0.8251953125,
0.52783203125,
0.5146484375,
0.52783203125
],
[
0.1175537109375,
0.52978515625,
0.484375,
0.53076171875,
0.484375,
0.5458984375,
0.1175537109375,
0.54443359375
],
[
0.51318359375,
0.54345703125,
0.54150390625,
0.54345703125,
0.54150390625,
0.5546875,
0.51318359375,
0.5546875
],
[
0.56103515625,
0.54150390625,
0.615234375,
0.54150390625,
0.615234375,
0.5556640625,
0.56103515625,
0.5556640625
],
[
0.12005615234375,
0.54736328125,
0.48779296875,
0.54736328125,
0.48779296875,
0.56103515625,
0.12005615234375,
0.56103515625
],
[
0.12005615234375,
0.5625,
0.489501953125,
0.5634765625,
0.489501953125,
0.57763671875,
0.12005615234375,
0.57666015625
],
[
0.51318359375,
0.56396484375,
0.8818359375,
0.56396484375,
0.8818359375,
0.5791015625,
0.51318359375,
0.5791015625
],
[
0.12005615234375,
0.580078125,
0.48779296875,
0.580078125,
0.48779296875,
0.59326171875,
0.12005615234375,
0.59326171875
],
[
0.5146484375,
0.580078125,
0.880859375,
0.580078125,
0.880859375,
0.59326171875,
0.5146484375,
0.59326171875
],
[
0.12005615234375,
0.5966796875,
0.489501953125,
0.5966796875,
0.489501953125,
0.60986328125,
0.12005615234375,
0.60986328125
],
[
0.515625,
0.5966796875,
0.8818359375,
0.5966796875,
0.8818359375,
0.60986328125,
0.515625,
0.60986328125
],
[
0.12005615234375,
0.61083984375,
0.48779296875,
0.61083984375,
0.48779296875,
0.62451171875,
0.12005615234375,
0.62451171875
],
[
0.515625,
0.61328125,
0.8818359375,
0.61328125,
0.8818359375,
0.626953125,
0.515625,
0.626953125
],
[
0.12005615234375,
0.62744140625,
0.489501953125,
0.62744140625,
0.489501953125,
0.64111328125,
0.12005615234375,
0.64111328125
],
[
0.515625,
0.62890625,
0.8818359375,
0.62890625,
0.8818359375,
0.642578125,
0.515625,
0.642578125
],
[
0.118408203125,
0.64453125,
0.48779296875,
0.64453125,
0.48779296875,
0.65771484375,
0.118408203125,
0.65771484375
],
[
0.5146484375,
0.6435546875,
0.8818359375,
0.64453125,
0.8818359375,
0.6591796875,
0.5146484375,
0.65771484375
],
[
0.12176513671875,
0.66015625,
0.486083984375,
0.66015625,
0.486083984375,
0.67333984375,
0.12176513671875,
0.67333984375
],
[
0.5146484375,
0.6611328125,
0.83447265625,
0.6611328125,
0.83447265625,
0.6748046875,
0.5146484375,
0.6748046875
],
[
0.118408203125,
0.67578125,
0.486083984375,
0.67578125,
0.486083984375,
0.689453125,
0.118408203125,
0.689453125
],
[
0.12005615234375,
0.6923828125,
0.48779296875,
0.6923828125,
0.48779296875,
0.7060546875,
0.12005615234375,
0.7060546875
],
[
0.5751953125,
0.6884765625,
0.82373046875,
0.689453125,
0.82373046875,
0.7080078125,
0.5751953125,
0.70703125
],
[
0.6689453125,
0.69677734375,
0.67919921875,
0.69677734375,
0.67919921875,
0.70068359375,
0.6689453125,
0.70068359375
],
[
0.8564453125,
0.693359375,
0.8818359375,
0.693359375,
0.8818359375,
0.7060546875,
0.8564453125,
0.7060546875
],
[
0.12005615234375,
0.708984375,
0.48779296875,
0.708984375,
0.48779296875,
0.72265625,
0.12005615234375,
0.72265625
],
[
0.634765625,
0.70703125,
0.72314453125,
0.708984375,
0.72314453125,
0.7236328125,
0.634765625,
0.7216796875
],
[
0.12176513671875,
0.724609375,
0.48779296875,
0.724609375,
0.48779296875,
0.73828125,
0.12176513671875,
0.73828125
],
[
0.12005615234375,
0.7392578125,
0.490234375,
0.74072265625,
0.490234375,
0.7548828125,
0.12005615234375,
0.75439453125
],
[
0.515625,
0.7392578125,
0.88330078125,
0.7392578125,
0.88330078125,
0.7529296875,
0.515625,
0.7529296875
],
[
0.118408203125,
0.7548828125,
0.489501953125,
0.755859375,
0.489501953125,
0.77099609375,
0.118408203125,
0.76953125
],
[
0.5146484375,
0.755859375,
0.884765625,
0.755859375,
0.884765625,
0.76953125,
0.5146484375,
0.76953125
],
[
0.12176513671875,
0.7724609375,
0.341796875,
0.7724609375,
0.341796875,
0.7861328125,
0.12176513671875,
0.7861328125
],
[
0.5146484375,
0.77099609375,
0.78271484375,
0.77099609375,
0.78271484375,
0.78369140625,
0.5146484375,
0.78369140625
],
[
0.12005615234375,
0.8017578125,
0.133544921875,
0.8017578125,
0.133544921875,
0.8134765625,
0.12005615234375,
0.8134765625
],
[
0.154541015625,
0.8017578125,
0.2376708984375,
0.8017578125,
0.2376708984375,
0.8134765625,
0.154541015625,
0.8134765625
],
[
0.6826171875,
0.8017578125,
0.79833984375,
0.8017578125,
0.79833984375,
0.81640625,
0.6826171875,
0.81640625
],
[
0.57080078125,
0.81005859375,
0.6533203125,
0.81005859375,
0.6533203125,
0.82470703125,
0.57080078125,
0.82470703125
],
[
0.6591796875,
0.8154296875,
0.82373046875,
0.81787109375,
0.82373046875,
0.83837890625,
0.6591796875,
0.8359375
],
[
0.85791015625,
0.81103515625,
0.8818359375,
0.81103515625,
0.8818359375,
0.826171875,
0.85791015625,
0.826171875
],
[
0.1376953125,
0.826171875,
0.486083984375,
0.826171875,
0.486083984375,
0.8388671875,
0.1376953125,
0.8388671875
],
[
0.12005615234375,
0.8427734375,
0.48779296875,
0.8427734375,
0.48779296875,
0.85546875,
0.12005615234375,
0.85546875
],
[
0.5322265625,
0.84423828125,
0.88330078125,
0.845703125,
0.88330078125,
0.85986328125,
0.5322265625,
0.859375
],
[
0.12005615234375,
0.859375,
0.489501953125,
0.859375,
0.489501953125,
0.87255859375,
0.12005615234375,
0.87255859375
],
[
0.51318359375,
0.859375,
0.8017578125,
0.859375,
0.8017578125,
0.87255859375,
0.51318359375,
0.87255859375
],
[
0.5341796875,
0.869140625,
0.884765625,
0.87060546875,
0.884765625,
0.88525390625,
0.5341796875,
0.88427734375
],
[
0.139404296875,
0.88134765625,
0.3779296875,
0.8818359375,
0.3779296875,
0.896484375,
0.139404296875,
0.8955078125
],
[
0.51318359375,
0.88427734375,
0.88671875,
0.88427734375,
0.88671875,
0.89794921875,
0.51318359375,
0.89794921875
],
[
0.14111328125,
0.8955078125,
0.462646484375,
0.8955078125,
0.462646484375,
0.91015625,
0.14111328125,
0.91015625
],
[
0.5146484375,
0.89794921875,
0.58837890625,
0.89794921875,
0.58837890625,
0.90771484375,
0.5146484375,
0.90771484375
]
] |
[
"本文的主要贡献总结如下：",
"$wget https://dumps.wikimedia.org/enwiki/",
": We present Wikipedia2Vec, a tool for learn-",
"latest/enwiki-latest-pages-articles.xml",
"bz2",
"ing the embeddings of words and entities eas-",
"wikipedia2vec train enwiki-latest-pages",
"articles.xml.bz2 MODEL FILE",
"ily and efficiently from Wikipedia.",
". Our tool achieved a state-of-the-art result",
"Figure 1: Shell commands to train embeddings",
"on the KORE entity relatedness dataset,",
"from the latest English Wikipedia dump.",
"and performed competitively on the various",
"benchmark datasets.",
">>>from wikipedia2vec import Wikipedia2Vec",
": We present a web-based demonstration that",
">>> model =Wikipedia2Vec.load (MODEL FILE)",
"allows users to explore the learned embed-",
">>> model.get_entity_vector(\"Scarlett Johansson\")",
"emmap(1-0.1979,",
"0.3086,",
"dtype=float32)",
"dings.",
">>>mo",
"elne:",
"_word_vector(\"tokyo",
"memmap(0.0161,-0.0332,",
"1,dtype=float32)",
">>> model.most similar(model.ge",
"t_entity(\"Python(",
"programming language)\"))[:3]",
". We publicize the code, demonstration, and",
"the pretrained embeddings for 12 languages",
"[(<Word python>, 0.7265)",
"at https://wikipedia2vec.github.io.",
"(<Entity Ruby (programming language)>, 0.6856),",
"(<Entity Perl>, 0.6794)])",
"2",
"Related Work",
"Figure 2: An example that uses the Wikipedia2Vec",
"许多研究最近提出了从KB（Huetal.，2015;",
"embeddings on a Python interactive shell.",
"Li et al.. 2016: Tsai and Roth, 2016: Yamada",
"et al.. 2016, 2017, 2018a: Ca0 et al.. 2017:",
"Ganea andHofmann，2017）学习实体嵌入的",
"）安装此工具。通过使用Wikipedia转储文件",
"方法。这些嵌入通常基于传统的单词嵌入模型",
"3作为参数运行wikipedia2vectrain命令，可",
"(例如。g。, skip-gram (Mikolov et al., 2013a)",
"以轻松学习嵌入。图1显示了下载最新的英文",
"）用从KB检索的数据训练。例如，Ristoski",
"Wikipedia转储文件并使用默认超参数基于此",
"etal.（2018）提出了RDF2Vec，它使用skip-",
"转储运行嵌入训练的shell命令。4此外，用",
"gram模型学习实体嵌入，并通过在Wikidata",
"户可以轻松使用学习的嵌人。图2显示了加载",
"和DBpedia等大型知识图上随机游走生成输",
"学习的嵌人文件的示例Python代码，并获得",
"入。此外，在各种研究中广泛使用的一种简",
"实体ScarlettJohansson和单词tokyo的嵌人，",
"单方法（Yaghoobzadeh and Schutze, 2015;Ya-",
"以及实体Python最相似的单词和实体",
"mada et al., 2017, 2018a; A1-Badrashiny et al.",
"3.1",
"Model",
"2017；Suzukietal.，2018）通过将输人语料库",
"中的实体注释替换为其引用实体的唯一标识",
"Wikipedia2Vec实现了传统的skip-gram模型",
"符，并将该语料库输入到单词嵌入模型中来训",
"(Mikolovet al.,2013a,b)及其在Yamadaet al.",
"练实体嵌入（e。g。skip-gram)。两个开源工具，",
"(2016）中提出的扩展，以将单词和实体映射到",
"即Wiki2Vec1和Wikipedia实体向量2已经",
"相同的d维向量空间。skip-gram模型是一种",
"实现了此方法。我们提出的工具基于Yamada",
"神经网络模型，其训练目标是找到对预测上下",
"etal.（2016），它通过使用由Wikipedia的内",
"文项有用的嵌入(i。e。，本文中的单词或实体）",
"部超链接连接的相邻实体作为附加上下文来",
"给定每个项目。模型的损失函数定义为：",
"训练模型，从而扩展了这一思想。请注意，我",
"们在实验中使用RDF2Vec和Wiki2Vec作为",
"Ls = - log P(ocloi),",
"",
"(1)",
"基线，并在KORE数据集上实现了比这些工",
"O;EO0cECo",
"具更高的经验性能。此外，已经提出了各种关",
"系嵌模型(Bordesetal.,2013;Wangetal.,",
"其中O是所有项的集合（i。é。，单词或实体）",
"2014;Linetal.,2015），旨在学习对知识图完",
"C。是o的上下文项集，条件概率logP(ocoi）",
"成任务特别有效的实体表示",
"是使用以下softmax函数定义的：",
"3",
"Overview",
"exp(VoTUo.)",
"P(ocloi) =",
"Doeo exp(VoTU.)",
"(2)",
"Wikipedia2Vec是一个易于使用的优化工",
"具，用于从Wikipedia学习嵌人。可以使用",
"3The dump file can be downloaded at Wikimedia",
"Python的pip工具（pipinstall wikipedia2vec",
"Downloads: https://dumps.wikimedia.org",
"4The train command has many optional hyper-",
"1https://github.com/idio/wiki2vec",
"parameters that are described in detail in the docu-",
"2https://github.com/singletongue/WikiEntVec",
"mentation."
] |
{
"arxiv_id": "1812.06280",
"language": "zh",
"page_idx": 1
}
|
||
1812.06280-zh-2
|
1812.06280-zh-2.jpg
|
<FigureHere>
Figure 3: Wikipedia2Vec learns embeddings by jointly optimizing word-based skip-gram, anchor context,
and link graph models.
其中$E$是词汇表中所有实体的集合，$C_e$是链接图中实体$e$的相邻实体$,P(e_o|e_i)$由等式 (2) 计算
最后，我们通过线性组合上述三个损失函数
来定义我们模型的损失函数：
其中$V_o\in\mathbb{R}^d$和$U_o\in\mathbb{R}^d$分别表示项$o$在嵌
人矩阵$V$和 THMAXX 中的嵌入
我们的工具通过联合优化下面描述的三个基于 skip-gram 的子模型来学习嵌入(另请参见图 3)。请注意，矩阵$V$和$U$包含单词和实体的嵌入
Word-based Skip-gram Model 给定维基百科页面中的每个单词，该模型通过预测给定单词的相邻单词来学习单词嵌人。形式上，给定一个单词序列$w_1,w_2,...,w_N$,该模型的损失函数定义如下：
$$\mathcal{L}=\mathcal{L}_w+\mathcal{L}_a+\mathcal{L}_e$$
(6)
通过使用随机梯度下降最小化此损失函数来执行训练。我们使用负采样 (Mikolov et al., 2013b)将 softmax 函数(方程 (2))转换为计算可行的函数。得到的矩阵$V$被用作学习的嵌入
$$\mathcal{L}_w=-\sum_{i=1}^N\sum_{-c\leq j\leq c,j\neq0}\log P(w_{i+j}|w_i),$$
3.2 Automatic Generation of Hyperlinks
其中$c$ 是上下文单词的大小，而$P(w_i+j|w_i)$
是根据公式 (2) 计算得出的
Anchor Context Model 此模型旨在使用维基百科中的超链接及其相邻单词将相似的单词和实体彼此靠近放置在向量空间中。从给定的维基百科页面，该模型提取引用实体和周围的单词 (i。e。,上一个和下一个$c$单词)来自页面中的每个超链接，并通过预测给定每个实体的周围单词来学习嵌入。因此，该模型的损失$函数定义如下：$
由于 Wikipedia 指示其贡献者仅在页面上首次出现实体名称时才创建超链接，因此许多实体名称不会显示为超链接。这对于我们的锚上下文模型是有问题的，因为它使用超链接作为源来学习嵌人
要解决此问题，我们的工具会使用映射实体名称的提及实体字典自动生成超链接(例如。g。,“苹果”)到其可能的引用实体(e。g。、苹果公司。orApple_(food))(seeSection 4 fordetails)。我们的工具从维基百科页面中提取所有单词和短语，并将每个单词和短语转换为指向实体的超链接，如果该实体在同一页面上被超链接引用，或者只有一个引用实体与字典中的名称相关联
(4)
$$\mathcal{L}_{a}=-\sum_{(e_{i},Q)\in A}\sum_{w_{c}\in Q}\log P(w_{c}|e_{i}),$$
### 4 Implementatior
其中$A$表示 Wikipedia 中所有超链接的集合每个超链接包含一对引用实体$e_i$和一组周围的单词$Q$ ,并且$P(w_c|e_i)$是根据公式 (2)计算的
Link Graph Model 该模型旨在通过预测Wikipedia 链接图中每个实体的相邻实体来学习实体嵌人-一个无向图，其节点是实体，边表示实体之间存在超链接。如果一个实体的页面具有到另一个实体的超链接，或者如果两个页面彼此链接，则我们在一对实体之间创建边缘。该模型的损失函数定义为：
我们的工具是用 Python 实现的，其大部分代码都使用 Cython (Behnel et al.,2011) 编译为C++,以优化运行时性能
如 3.1 节所述，我们的链接图和锚上下文模型基于 Wikipedia 中的超链接。由于Wikipedia 包含许多超链接，因此有效地使用它们具有挑战性。为了解决这个问题，我们引入了两个优化的组件-链接图矩阵和提及实体字典-在训练过程中使用
$\begin{array}{ccc}{,\:(5)\:}&{\mathrm{Link~Graph~Matrix~在训练过程中，我们的}}\\&{\text{链接图模型需要在 Wikipedia 的大型链接图}}\end{array}$
$$\mathcal{L}_{e}=-\sum\limits_{e_{i}\in E}\sum\limits_{e_{o}\in C_{e_{i}}}\log P(e_{o}|e_{i}),$$
|
[
[
0.1376953125,
0.07244873046875,
0.32421875,
0.07244873046875,
0.32421875,
0.08612060546875,
0.1376953125,
0.08612060546875
],
[
0.426513671875,
0.07244873046875,
0.5615234375,
0.07244873046875,
0.5615234375,
0.0831298828125,
0.426513671875,
0.0831298828125
],
[
0.7109375,
0.0736083984375,
0.8193359375,
0.0736083984375,
0.8193359375,
0.08489990234375,
0.7109375,
0.08489990234375
],
[
0.6884765625,
0.08489990234375,
0.8017578125,
0.0914306640625,
0.80029296875,
0.105712890625,
0.68701171875,
0.09979248046875
],
[
0.12176513671875,
0.0968017578125,
0.340087890625,
0.097412109375,
0.340087890625,
0.11224365234375,
0.12176513671875,
0.11102294921875
],
[
0.385498046875,
0.0968017578125,
0.603515625,
0.097412109375,
0.603515625,
0.11224365234375,
0.385498046875,
0.11102294921875
],
[
0.8212890625,
0.0914306640625,
0.8388671875,
0.0914306640625,
0.8388671875,
0.09442138671875,
0.8212890625,
0.09442138671875
],
[
0.69287109375,
0.1015625,
0.779296875,
0.1015625,
0.779296875,
0.1134033203125,
0.69287109375,
0.1134033203125
],
[
0.78515625,
0.097412109375,
0.880859375,
0.0955810546875,
0.880859375,
0.10986328125,
0.78515625,
0.11224365234375
],
[
0.145263671875,
0.10986328125,
0.193115234375,
0.11517333984375,
0.189697265625,
0.1258544921875,
0.1427001953125,
0.12054443359375
],
[
0.361083984375,
0.10809326171875,
0.3720703125,
0.11517333984375,
0.364501953125,
0.12054443359375,
0.353515625,
0.1134033203125
],
[
0.420654296875,
0.11102294921875,
0.456787109375,
0.1175537109375,
0.45166015625,
0.12890625,
0.41650390625,
0.122314453125
],
[
0.73486328125,
0.11224365234375,
0.78515625,
0.11224365234375,
0.78515625,
0.12353515625,
0.73486328125,
0.12353515625
],
[
0.276123046875,
0.114013671875,
0.3046875,
0.114013671875,
0.3046875,
0.1217041015625,
0.276123046875,
0.1217041015625
],
[
0.189697265625,
0.1134033203125,
0.25439453125,
0.1134033203125,
0.25439453125,
0.127685546875,
0.189697265625,
0.127685546875
],
[
0.724609375,
0.127685546875,
0.79248046875,
0.127685546875,
0.79248046875,
0.1383056640625,
0.724609375,
0.1383056640625
],
[
0.118408203125,
0.1395263671875,
0.344970703125,
0.1395263671875,
0.344970703125,
0.152587890625,
0.118408203125,
0.152587890625
],
[
0.39208984375,
0.1395263671875,
0.599609375,
0.1383056640625,
0.599609375,
0.1519775390625,
0.39208984375,
0.152587890625
],
[
0.64990234375,
0.13720703125,
0.87890625,
0.1383056640625,
0.87890625,
0.152587890625,
0.64990234375,
0.1519775390625
],
[
0.185546875,
0.152587890625,
0.27783203125,
0.152587890625,
0.27783203125,
0.163330078125,
0.185546875,
0.163330078125
],
[
0.378662109375,
0.1519775390625,
0.611328125,
0.15087890625,
0.611328125,
0.1644287109375,
0.378662109375,
0.1650390625
],
[
0.6484375,
0.152587890625,
0.8818359375,
0.152587890625,
0.8818359375,
0.163330078125,
0.6484375,
0.163330078125
],
[
0.12005615234375,
0.17333984375,
0.880859375,
0.17333984375,
0.880859375,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.12005615234375,
0.189453125,
0.28955078125,
0.189453125,
0.28955078125,
0.203125,
0.12005615234375,
0.203125
],
[
0.118408203125,
0.22802734375,
0.490234375,
0.2286376953125,
0.490234375,
0.243408203125,
0.118408203125,
0.2423095703125
],
[
0.5146484375,
0.2298583984375,
0.8818359375,
0.2298583984375,
0.8818359375,
0.243408203125,
0.5146484375,
0.243408203125
],
[
0.12005615234375,
0.2452392578125,
0.39208984375,
0.2452392578125,
0.39208984375,
0.2587890625,
0.12005615234375,
0.2587890625
],
[
0.51318359375,
0.24462890625,
0.8818359375,
0.2452392578125,
0.8818359375,
0.260009765625,
0.51318359375,
0.2587890625
],
[
0.1376953125,
0.262451171875,
0.489501953125,
0.262451171875,
0.489501953125,
0.275634765625,
0.1376953125,
0.275634765625
],
[
0.51318359375,
0.262451171875,
0.55517578125,
0.262451171875,
0.55517578125,
0.275634765625,
0.51318359375,
0.275634765625
],
[
0.118408203125,
0.27783203125,
0.486083984375,
0.276611328125,
0.486083984375,
0.29150390625,
0.118408203125,
0.292236328125
],
[
0.5322265625,
0.279052734375,
0.8818359375,
0.279052734375,
0.8818359375,
0.292236328125,
0.5322265625,
0.292236328125
],
[
0.12005615234375,
0.29443359375,
0.489501953125,
0.29443359375,
0.489501953125,
0.30810546875,
0.12005615234375,
0.30810546875
],
[
0.5146484375,
0.29443359375,
0.744140625,
0.29443359375,
0.744140625,
0.30810546875,
0.5146484375,
0.30810546875
],
[
0.118408203125,
0.308837890625,
0.19482421875,
0.311279296875,
0.19482421875,
0.32470703125,
0.118408203125,
0.322509765625
],
[
0.62451171875,
0.322509765625,
0.7734375,
0.322509765625,
0.7734375,
0.336181640625,
0.62451171875,
0.336181640625
],
[
0.85791015625,
0.32373046875,
0.8818359375,
0.32373046875,
0.8818359375,
0.33837890625,
0.85791015625,
0.33837890625
],
[
0.1275634765625,
0.333251953125,
0.484375,
0.333251953125,
0.484375,
0.3466796875,
0.1275634765625,
0.3466796875
],
[
0.12176513671875,
0.3515625,
0.486083984375,
0.3515625,
0.486083984375,
0.3623046875,
0.12176513671875,
0.3623046875
],
[
0.5146484375,
0.352783203125,
0.8818359375,
0.352783203125,
0.8818359375,
0.366455078125,
0.5146484375,
0.366455078125
],
[
0.12005615234375,
0.366455078125,
0.48779296875,
0.366455078125,
0.48779296875,
0.380126953125,
0.12005615234375,
0.380126953125
],
[
0.5146484375,
0.3681640625,
0.8818359375,
0.3681640625,
0.8818359375,
0.3818359375,
0.5146484375,
0.3818359375
],
[
0.133544921875,
0.3818359375,
0.48779296875,
0.3818359375,
0.48779296875,
0.396728515625,
0.133544921875,
0.396728515625
],
[
0.5146484375,
0.385498046875,
0.880859375,
0.385498046875,
0.880859375,
0.3984375,
0.5146484375,
0.3984375
],
[
0.118408203125,
0.3984375,
0.2376708984375,
0.3984375,
0.2376708984375,
0.412109375,
0.118408203125,
0.412109375
],
[
0.515625,
0.402099609375,
0.8818359375,
0.402099609375,
0.8818359375,
0.4150390625,
0.515625,
0.4150390625
],
[
0.5146484375,
0.4150390625,
0.5576171875,
0.41748046875,
0.5556640625,
0.432861328125,
0.51318359375,
0.43115234375
],
[
0.189697265625,
0.42578125,
0.47265625,
0.427490234375,
0.47265625,
0.451904296875,
0.189697265625,
0.449462890625
],
[
0.214111328125,
0.421630859375,
0.231689453125,
0.421630859375,
0.231689453125,
0.429931640625,
0.214111328125,
0.429931640625
],
[
0.1435546875,
0.435302734375,
0.2032470703125,
0.435302734375,
0.2032470703125,
0.446533203125,
0.1435546875,
0.446533203125
],
[
0.4609375,
0.435791015625,
0.486083984375,
0.435791015625,
0.486083984375,
0.44775390625,
0.4609375,
0.44775390625
],
[
0.51123046875,
0.441162109375,
0.8564453125,
0.4423828125,
0.8564453125,
0.457275390625,
0.51123046875,
0.4560546875
],
[
0.20654296875,
0.448974609375,
0.32080078125,
0.45068359375,
0.32080078125,
0.465576171875,
0.20654296875,
0.463134765625
],
[
0.51708984375,
0.46435546875,
0.8818359375,
0.46435546875,
0.8818359375,
0.47802734375,
0.51708984375,
0.47802734375
],
[
0.118408203125,
0.471435546875,
0.489501953125,
0.47265625,
0.489501953125,
0.48681640625,
0.118408203125,
0.486328125
],
[
0.51708984375,
0.482177734375,
0.8818359375,
0.482177734375,
0.8818359375,
0.4921875,
0.51708984375,
0.4921875
],
[
0.12176513671875,
0.4892578125,
0.340087890625,
0.4892578125,
0.340087890625,
0.5029296875,
0.12176513671875,
0.5029296875
],
[
0.515625,
0.496337890625,
0.8818359375,
0.496337890625,
0.8818359375,
0.51025390625,
0.515625,
0.51025390625
],
[
0.1292724609375,
0.5126953125,
0.48779296875,
0.5126953125,
0.48779296875,
0.525390625,
0.1292724609375,
0.525390625
],
[
0.51904296875,
0.5126953125,
0.8818359375,
0.5126953125,
0.8818359375,
0.525390625,
0.51904296875,
0.525390625
],
[
0.12176513671875,
0.52978515625,
0.484375,
0.52978515625,
0.484375,
0.54052734375,
0.12176513671875,
0.54052734375
],
[
0.515625,
0.529296875,
0.623046875,
0.529296875,
0.623046875,
0.54052734375,
0.515625,
0.54052734375
],
[
0.12005615234375,
0.54443359375,
0.48779296875,
0.54443359375,
0.48779296875,
0.55810546875,
0.12005615234375,
0.55810546875
],
[
0.5341796875,
0.54443359375,
0.88330078125,
0.54443359375,
0.88330078125,
0.55810546875,
0.5341796875,
0.55810546875
],
[
0.118408203125,
0.55859375,
0.48779296875,
0.56005859375,
0.48779296875,
0.57373046875,
0.118408203125,
0.572265625
],
[
0.51318359375,
0.55859375,
0.880859375,
0.55859375,
0.880859375,
0.572265625,
0.51318359375,
0.572265625
],
[
0.12005615234375,
0.57666015625,
0.48779296875,
0.57666015625,
0.48779296875,
0.59033203125,
0.12005615234375,
0.59033203125
],
[
0.5146484375,
0.57666015625,
0.88330078125,
0.57666015625,
0.88330078125,
0.59130859375,
0.5146484375,
0.59130859375
],
[
0.12005615234375,
0.5927734375,
0.489501953125,
0.5927734375,
0.489501953125,
0.60546875,
0.12005615234375,
0.60546875
],
[
0.51318359375,
0.59326171875,
0.884765625,
0.5927734375,
0.884765625,
0.60693359375,
0.51318359375,
0.60791015625
],
[
0.12005615234375,
0.609375,
0.489501953125,
0.609375,
0.489501953125,
0.62255859375,
0.12005615234375,
0.62255859375
],
[
0.5146484375,
0.609375,
0.8818359375,
0.609375,
0.8818359375,
0.62255859375,
0.5146484375,
0.62255859375
],
[
0.12005615234375,
0.6259765625,
0.2418212890625,
0.6259765625,
0.2418212890625,
0.6396484375,
0.12005615234375,
0.6396484375
],
[
0.515625,
0.6259765625,
0.8818359375,
0.6259765625,
0.8818359375,
0.6396484375,
0.515625,
0.6396484375
],
[
0.1695556640625,
0.6455078125,
0.465087890625,
0.6455078125,
0.465087890625,
0.66650390625,
0.1695556640625,
0.66650390625
],
[
0.515625,
0.64111328125,
0.8818359375,
0.64111328125,
0.8818359375,
0.65478515625,
0.515625,
0.65478515625
],
[
0.45751953125,
0.65380859375,
0.46337890625,
0.65380859375,
0.46337890625,
0.6591796875,
0.45751953125,
0.6591796875
],
[
0.4609375,
0.65087890625,
0.486083984375,
0.65087890625,
0.486083984375,
0.662109375,
0.4609375,
0.662109375
],
[
0.51708984375,
0.65673828125,
0.8818359375,
0.65673828125,
0.8818359375,
0.67041015625,
0.51708984375,
0.67041015625
],
[
0.231689453125,
0.66650390625,
0.33154296875,
0.66650390625,
0.33154296875,
0.677734375,
0.231689453125,
0.677734375
],
[
0.5146484375,
0.67333984375,
0.6630859375,
0.67333984375,
0.6630859375,
0.68701171875,
0.5146484375,
0.68701171875
],
[
0.12005615234375,
0.6884765625,
0.484375,
0.6884765625,
0.484375,
0.703125,
0.12005615234375,
0.703125
],
[
0.51318359375,
0.703125,
0.5498046875,
0.70068359375,
0.5517578125,
0.7119140625,
0.5146484375,
0.71435546875
],
[
0.54248046875,
0.70068359375,
0.6884765625,
0.70068359375,
0.6884765625,
0.71435546875,
0.54248046875,
0.71435546875
],
[
0.118408203125,
0.7197265625,
0.48779296875,
0.72021484375,
0.48779296875,
0.7353515625,
0.118408203125,
0.73388671875
],
[
0.12005615234375,
0.705078125,
0.48779296875,
0.705078125,
0.48779296875,
0.71875,
0.12005615234375,
0.71875
],
[
0.5146484375,
0.7255859375,
0.8818359375,
0.7255859375,
0.8818359375,
0.7392578125,
0.5146484375,
0.7392578125
],
[
0.12005615234375,
0.73681640625,
0.156982421875,
0.73681640625,
0.156982421875,
0.75,
0.12005615234375,
0.75
],
[
0.515625,
0.7421875,
0.8818359375,
0.7421875,
0.8818359375,
0.755859375,
0.515625,
0.755859375
],
[
0.1275634765625,
0.76025390625,
0.48779296875,
0.76025390625,
0.48779296875,
0.77490234375,
0.1275634765625,
0.77490234375
],
[
0.515625,
0.76025390625,
0.72314453125,
0.76025390625,
0.72314453125,
0.77099609375,
0.515625,
0.77099609375
],
[
0.118408203125,
0.77685546875,
0.48779296875,
0.77685546875,
0.48779296875,
0.79052734375,
0.118408203125,
0.79052734375
],
[
0.5341796875,
0.77490234375,
0.8818359375,
0.77490234375,
0.8818359375,
0.7880859375,
0.5341796875,
0.7880859375
],
[
0.12176513671875,
0.79443359375,
0.486083984375,
0.79443359375,
0.486083984375,
0.8046875,
0.12176513671875,
0.8046875
],
[
0.5146484375,
0.7890625,
0.880859375,
0.7890625,
0.880859375,
0.802734375,
0.5146484375,
0.802734375
],
[
0.12176513671875,
0.80859375,
0.48779296875,
0.80859375,
0.48779296875,
0.822265625,
0.12176513671875,
0.822265625
],
[
0.5146484375,
0.8056640625,
0.88330078125,
0.8056640625,
0.88330078125,
0.8193359375,
0.5146484375,
0.8193359375
],
[
0.12176513671875,
0.82470703125,
0.48779296875,
0.82470703125,
0.48779296875,
0.83837890625,
0.12176513671875,
0.83837890625
],
[
0.515625,
0.822265625,
0.8818359375,
0.822265625,
0.8818359375,
0.8359375,
0.515625,
0.8359375
],
[
0.1234130859375,
0.8427734375,
0.486083984375,
0.8427734375,
0.486083984375,
0.8525390625,
0.1234130859375,
0.8525390625
],
[
0.515625,
0.84033203125,
0.880859375,
0.84033203125,
0.880859375,
0.85107421875,
0.515625,
0.85107421875
],
[
0.118408203125,
0.85546875,
0.366943359375,
0.85693359375,
0.366943359375,
0.87060546875,
0.118408203125,
0.869140625
],
[
0.515625,
0.85498046875,
0.724609375,
0.85498046875,
0.724609375,
0.8681640625,
0.515625,
0.8681640625
],
[
0.1796875,
0.87744140625,
0.432373046875,
0.876953125,
0.432373046875,
0.89794921875,
0.1796875,
0.8984375
],
[
0.462646484375,
0.8818359375,
0.48779296875,
0.8818359375,
0.48779296875,
0.89453125,
0.462646484375,
0.89453125
],
[
0.51904296875,
0.8798828125,
0.880859375,
0.8798828125,
0.880859375,
0.8935546875,
0.51904296875,
0.8935546875
],
[
0.2418212890625,
0.896484375,
0.32421875,
0.8984375,
0.32421875,
0.9130859375,
0.2418212890625,
0.9111328125
],
[
0.515625,
0.8955078125,
0.880859375,
0.8955078125,
0.880859375,
0.9091796875,
0.515625,
0.9091796875
]
] |
[
"Word-based skip-gram model",
"Anchorcontextmodel",
"Linkgraphmodel",
"Philosophy",
"Aristotle was a philosopher",
"Aristotle was a philosopher",
"4",
"Aristotle K",
"Avicenna",
"AE",
"?",
"##日",
"Plato",
"A",
"",
"Socrates",
"The neighboring words of each word are",
"The neighboring words of a hyperlink",
"The neighboring entities ofeach entity in",
"used as contexts",
"pointing to an entity are used as contexts",
"Wikipedia's linkgraphareusedas contexts",
"Figure 3: Wikipedia2Vec learns embeddings by jointly optimizing word-based skip-gram, anchor context",
"and link graph models.",
"其中V。ERd和U。ERd分别表示项o在嵌",
"其中E是词汇表中所有实体的集合，Ce是链",
"人矩阵V和THMAXX中的嵌入",
"接图中实体e的相邻实体，P(eolei）由等式（2)",
"我们的工具通过联合优化下面描述的三个基",
"计算",
"于skip-gram的子模型来学习嵌人（另请参见",
"最后，我们通过线性组合上述三个损失函数",
"图3）。请注意，矩阵V和U包含单词和实",
"来定义我们模型的损失函数：",
"体的嵌入",
"C = Cu + La+ Ce",
"(6)",
"Word-based Skip-gramModel给定维基百",
"科页面中的每个单词，该模型通过预测给定单",
"通过使用随机梯度下降最小化此损失函数来",
"词的相邻单词来学习单词嵌入。形式上，给定",
"执行训练。我们使用负采样(Mikolovetal.,",
"个单词序列wi,w2.WN，该模型的损失",
"2013b）将softmax函数（方程（2））转换为计",
"函数定义如下：",
"算可行的函数。得到的矩阵V被用作学习的",
"嵌入",
"-log P(wi+jlwi)，（",
"N",
"Lw=",
"(3)",
"3.2Automatic Generation of Hyperlinks",
"=1-cj<cj0",
"由于Wikipedia指示其贡献者仅在页面上首",
"其中c是上下文单词的大小，而P(wi+ilwi)",
"次出现实体名称时才创建超链接，因此许多实",
"是根据公式（2）计算得出的",
"体名称不会显示为超链接。这对于我们的锚上",
"AnchorContextModel此模型旨在使用维",
"下文模型是有问题的，因为它使用超链接作为",
"基百科中的超链接及其相邻单词将相似的单词",
"源来学习嵌入",
"和实体彼此靠近放置在向量空间中。从给定的",
"要解决此问题，我们的工具会使用映射实",
"维基百科页面，该模型提取引用实体和周围的",
"体名称的提及实体字典自动生成超链接（例",
"单词(i。e。，上一个和下一个c单词）来自页",
"如。g。，“苹果”）到其可能的引用实体（e。",
"面中的每个超链接，并通过预测给定每个实体",
"g。、苹果公司。orApple_(food))(seeSection4",
"的周围单词来学习嵌入。因此，该模型的损失",
"fordetails）。我们的工具从维基百科页面中提",
"函数定义如下：",
"取所有单词和短语，并将每个单词和短语转换",
"Ca=- Z log P(welei)，",
"为指向实体的超链接，如果该实体在同一页面",
"",
"(4)",
"上被超链接引用，或者只有一个引用实体与字",
"(ei,Q)EAweEQ",
"典中的名称相关联",
"其中A表示Wikipedia中所有超链接的集合",
"4",
" Implementation",
"的单词Q，并且P(wclei)是根据公式（2)计",
"每个超链接包含一对引用实体ei和一组周围",
"我们的工具是用Python实现的，其大部分代",
"算的",
"码都使用Cython（Behneletal.，2011）编译为",
"LinkGraphModel该模型旨在通过预测",
"C十十，以优化运行时性能",
"Wikipedia链接图中每个实体的相邻实体来",
"如3.1节所述，我们的链接图和锚上下",
"学习实体嵌人-一个无向图，其节点是实体，边",
"文模型基于Wikipedia中的超链接。由于",
"表示实体之间存在超链接。如果一个实体的页",
"Wikipedia包含许多超链接，因此有效地使",
"面具有到另一个实体的超链接，或者如果两个",
"用它们具有挑战性。为了解决这个问题，我们",
"页面彼此链接，则我们在一对实体之间创建边",
"引入了两个优化的组件-链接图矩阵和提及实",
"缘。该模型的损失函数定义为：",
"体字典-在训练过程中使用",
"Ce = - Z Z log P(eolei),",
"(5)",
"LinkGraphMatrix在训练过程中，我们的",
"eiEEoECe",
"链接图模型需要在Wikipedia的大型链接图"
] |
{
"arxiv_id": "1812.06280",
"language": "zh",
"page_idx": 2
}
|
||
1812.06280-zh-3
|
1812.06280-zh-3.jpg
|
中获取一个实体的众多相邻实体。为了减少延迟，此组件使用压缩稀疏行(CSR)格式的二进制稀疏矩阵将整个图存储在内存中，其中其行和列表示实体，其值表示相应实体对之间存在超链接。由于此矩阵的大小通常很小，因此可以轻松地将其存储在内存中。$^{5}$注意，给定CSR 矩阵中的行索引，获取其非零列索引(对应于对应于行索引的实体的相邻实体)的时间复杂度为$O(1)$
Mention-entity Dictionary 提及实体字典用于生成 3.2 节中描述的超链接。该字典将实体名称映射到其可能的引用实体，并根据从Wikipedia 中的所有超链接获得的名称及其引用实体创建。我们的工具从维基百科页面中提取包含大量实体名称的字典中包含的所有单词和短语。为了有效地实现这一点，我们使用Aho-Corasick 算法，这是一种有效的字符串搜索算法，使用从所有实体名称构造的有限状态机。在检测到字典中的单词和短语后，我们的工具根据 3.2 节中描述的启发式方法将它们转换为超链接
通过以随机顺序同时迭代维基百科中的页面和链接图中的实体来训练嵌入。使用 mwparserfromhellMediaWiki 解析器提取每个页面中的文本和超链接。$^{6}$我们不使用半结构化数据，例如表和信息框。我们还使用提及实体字典生成超链接。我们将嵌人作为浮动矩阵存储在共享内存中，并使用多个进程更新它。学习嵌入所需的线性代数运算使用基本线性代数子程序 (BLAS) 中的 C 函数实现
此外，我们的工具使用标记器来检测维基百科页面中的单词。当前在我们的工具中实现了以下四个标记器：(1)实现 Unicode 文本分割算法 (Davis, 2019) 的多语言 ICU 标记器$^{7}$ ,(2)使用空白字符拆分文本的简单基于规则的标记器，(3)用于中文的 Jieba 标记器$^{8}$,以及(4)用于日语的 MeCab 标记器 XMATHX35 Experiments
我们进行了实验，以比较我们的工具与现有工具的质量和效率。为了评估实体嵌入的质量，我们使用了 KORE 实体相关性数据集(Hoffart et al.,2012)。该数据集由 21 个实体组成，每个实体有 20 个相关实体，分数由人类评估。在过去的工作之后，我们报告了黄金分数和实体嵌人之间的余弦相似度之间的斯皮
$^{5}$The size of the matrix of English Wikipedia is less than 500 megabytes with our default hyper-parameter $\operatorname{settings.}$
$^{6}$https://github.com/earwig/mwparserfromhell
$^{7}$http: / / site. icu- project. org
$^{8}$https://github.com/fxsjy/jieba
<table>
<tbody>
<tr>
<th>$\overline{\mathrm{Name}}$</th>
<th>$\overline{\mathrm{Score}}$</th>
</tr>
<tr>
<td>$\overline{Ours}$</td>
<td>0.71</td>
</tr>
<tr>
<td>Ours (w/olink $\kappa$graph model</td>
<td>0.61</td>
</tr>
<tr>
<td>Ours (w/ o hvperlink generation)</td>
<td>0.69</td>
</tr>
<tr>
<td>RDF2Vec Ristoski et al. 2018</td>
<td>0.69</td>
</tr>
<tr>
<td>Wiki2vec</td>
<td>0.52</td>
</tr>
</tbody>
</table>
Table 1: The results of Wikipedia2Vec and the
baseline entity embeddings on the KORE dataset.
尔曼秩相关系数。我们使用了两个流行的实体嵌人工具 RDF2Vec (Ristoski et al., 2018) 和Wiki2vec 作为基线
我们还通过使用两个标准任务来评估单词嵌入的质量：(1)使用 GoogleWord 类比数据集 (Mikolov et al., 2013a)的语义子集(SEM)和句法子集 (SYN)的单词类比任务， 以及(2) 使用两个标准数据集的单词相似性任务，即 SimLex-999(SL) (Hill et al.,2015) 和 WordSim-THMA353 (WS)。根据过去的工作，我们报告了单词类比任务的准确性，以及单词相似性任务的黄金得分与单词嵌人之间的余弦相似性之间的 Spearman 等级相关系数。作为这些任务的基线，我们使用了在 gensim 库 3 中实现的 skip-gram 模型 (Mikolov et al., 2013a) 。6。0 ($\text{\v {R}eh\r {u}\v {r}ekandSojka, 2010) 和 }$ 在 fastText 工具0中实现的扩展 skip-gram 模型。1。0 (Bojanowski et al., 2017) 。我们使用WikiExtractor$^{9}$创建基线的训练语料库。在可能的范围内，我们使用相同的超参数来训练我们的模型和基线。10
我们还报告了使用我们的工具和基线单词嵌入工具进行培训所需的时间。请注意， RDF2Vec 和 Wiki2vec 工具的训练是使用 gen- $\operatorname{sim}$实现的
我们使用 Python3 进行了实验。6 和 OpenBLAS0。3。3 安装在 c5d 上。在亚马逊云科技上部署了 36 个 CPU 核心的 9xlarge 实例。为了训练我们的模型和基准词嵌入模型，我们使用了 2018 年 4 月版的英文维基百科转储
### 5.1 Results
表 1 显示了我们的模型和 KORE 数据集的基线实体嵌入模型的结果。$^{11}$w/o 链接图模型和w/o 超链接生成分别是消融研究禁用链接图模型和自动生成超链接的结果
我们的模型成功地超越了 RDF2Vec 和Wiki2vec 模型，并在 KORE 数据集上取得了最先进的结果。结果还表明，链接图模型和
$^{9}$https://github.com/attardi/wikiextractor
$\begin{array}{c}{{^{10}\mathrm{We~used~the~following~settings:~}dim_{-}size=500~,}}\\{{window=5~,~negative=5~,~iteration=5}}\end{array}$ ${\mathrm{Vind}}^{11}$We${\mathrm{obtained~the~results~of~the~RDF2Vec~and}}$ Wiki2vec models from Ristoski et al. (2018) .
|
[
[
0.5693359375,
0.0748291015625,
0.611328125,
0.0748291015625,
0.611328125,
0.08612060546875,
0.5693359375,
0.08612060546875
],
[
0.78857421875,
0.07244873046875,
0.8271484375,
0.0748291015625,
0.8251953125,
0.0872802734375,
0.78662109375,
0.08489990234375
],
[
0.7900390625,
0.08612060546875,
0.8212890625,
0.0885009765625,
0.8193359375,
0.09979248046875,
0.78857421875,
0.097412109375
],
[
0.12176513671875,
0.07659912109375,
0.48779296875,
0.07659912109375,
0.48779296875,
0.09027099609375,
0.12176513671875,
0.09027099609375
],
[
0.57080078125,
0.0872802734375,
0.60693359375,
0.0872802734375,
0.60693359375,
0.09857177734375,
0.57080078125,
0.09857177734375
],
[
0.12005615234375,
0.09320068359375,
0.48779296875,
0.09320068359375,
0.48779296875,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.57421875,
0.09979248046875,
0.75146484375,
0.09979248046875,
0.75146484375,
0.10986328125,
0.57421875,
0.10986328125
],
[
0.791015625,
0.097412109375,
0.8212890625,
0.1009521484375,
0.81787109375,
0.11102294921875,
0.78857421875,
0.10809326171875
],
[
0.12005615234375,
0.10986328125,
0.48779296875,
0.10986328125,
0.48779296875,
0.12353515625,
0.12005615234375,
0.12353515625
],
[
0.5751953125,
0.1134033203125,
0.77490234375,
0.1134033203125,
0.77490234375,
0.12353515625,
0.5751953125,
0.12353515625
],
[
0.7841796875,
0.11639404296875,
0.791015625,
0.11639404296875,
0.791015625,
0.119384765625,
0.7841796875,
0.119384765625
],
[
0.7900390625,
0.11224365234375,
0.8193359375,
0.11224365234375,
0.8193359375,
0.12353515625,
0.7900390625,
0.12353515625
],
[
0.57275390625,
0.1217041015625,
0.771484375,
0.12353515625,
0.771484375,
0.13720703125,
0.57275390625,
0.1353759765625
],
[
0.791015625,
0.122314453125,
0.82275390625,
0.1258544921875,
0.8193359375,
0.135986328125,
0.7900390625,
0.1341552734375
],
[
0.791015625,
0.1341552734375,
0.8251953125,
0.135986328125,
0.82373046875,
0.1484375,
0.7900390625,
0.146728515625
],
[
0.12005615234375,
0.1258544921875,
0.486083984375,
0.1258544921875,
0.486083984375,
0.1395263671875,
0.12005615234375,
0.1395263671875
],
[
0.57275390625,
0.135986328125,
0.63232421875,
0.135986328125,
0.63232421875,
0.146728515625,
0.57275390625,
0.146728515625
],
[
0.12176513671875,
0.1436767578125,
0.486083984375,
0.1436767578125,
0.486083984375,
0.15380859375,
0.12176513671875,
0.15380859375
],
[
0.12005615234375,
0.1561279296875,
0.486083984375,
0.15673828125,
0.486083984375,
0.171630859375,
0.12005615234375,
0.17041015625
],
[
0.51318359375,
0.1602783203125,
0.8818359375,
0.1602783203125,
0.8818359375,
0.17333984375,
0.51318359375,
0.17333984375
],
[
0.12005615234375,
0.17333984375,
0.48779296875,
0.17333984375,
0.48779296875,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.5146484375,
0.17578125,
0.87890625,
0.17578125,
0.87890625,
0.1859130859375,
0.5146484375,
0.1859130859375
],
[
0.12176513671875,
0.190673828125,
0.484375,
0.190673828125,
0.484375,
0.20068359375,
0.12176513671875,
0.20068359375
],
[
0.12005615234375,
0.2060546875,
0.2392578125,
0.2060546875,
0.2392578125,
0.2197265625,
0.12005615234375,
0.2197265625
],
[
0.51708984375,
0.21435546875,
0.8818359375,
0.21435546875,
0.8818359375,
0.22802734375,
0.51708984375,
0.22802734375
],
[
0.124267578125,
0.2298583984375,
0.361083984375,
0.2298583984375,
0.361083984375,
0.243408203125,
0.124267578125,
0.243408203125
],
[
0.3525390625,
0.2298583984375,
0.48779296875,
0.2298583984375,
0.48779296875,
0.243408203125,
0.3525390625,
0.243408203125
],
[
0.5146484375,
0.2298583984375,
0.88330078125,
0.2298583984375,
0.88330078125,
0.243408203125,
0.5146484375,
0.243408203125
],
[
0.12005615234375,
0.2452392578125,
0.489501953125,
0.2452392578125,
0.489501953125,
0.2587890625,
0.12005615234375,
0.2587890625
],
[
0.51123046875,
0.2464599609375,
0.66650390625,
0.2464599609375,
0.66650390625,
0.260009765625,
0.51123046875,
0.260009765625
],
[
0.12176513671875,
0.26318359375,
0.484375,
0.26318359375,
0.484375,
0.273681640625,
0.12176513671875,
0.273681640625
],
[
0.5322265625,
0.262451171875,
0.88330078125,
0.262451171875,
0.88330078125,
0.275634765625,
0.5322265625,
0.275634765625
],
[
0.12005615234375,
0.27783203125,
0.486083984375,
0.27783203125,
0.486083984375,
0.29150390625,
0.12005615234375,
0.29150390625
],
[
0.51318359375,
0.27783203125,
0.880859375,
0.27783203125,
0.880859375,
0.29150390625,
0.51318359375,
0.29150390625
],
[
0.12176513671875,
0.295654296875,
0.486083984375,
0.295654296875,
0.486083984375,
0.305908203125,
0.12176513671875,
0.305908203125
],
[
0.5146484375,
0.29443359375,
0.8818359375,
0.29443359375,
0.8818359375,
0.30810546875,
0.5146484375,
0.30810546875
],
[
0.118408203125,
0.308837890625,
0.486083984375,
0.308837890625,
0.486083984375,
0.322509765625,
0.118408203125,
0.322509765625
],
[
0.51318359375,
0.311279296875,
0.8818359375,
0.311279296875,
0.8818359375,
0.32470703125,
0.51318359375,
0.32470703125
],
[
0.12176513671875,
0.32666015625,
0.48779296875,
0.32666015625,
0.48779296875,
0.34033203125,
0.12176513671875,
0.34033203125
],
[
0.515625,
0.32666015625,
0.8818359375,
0.32666015625,
0.8818359375,
0.34033203125,
0.515625,
0.34033203125
],
[
0.12005615234375,
0.342529296875,
0.48779296875,
0.342529296875,
0.48779296875,
0.355712890625,
0.12005615234375,
0.355712890625
],
[
0.5146484375,
0.342529296875,
0.880859375,
0.34326171875,
0.880859375,
0.35693359375,
0.5146484375,
0.355712890625
],
[
0.12005615234375,
0.359375,
0.48779296875,
0.359375,
0.48779296875,
0.372314453125,
0.12005615234375,
0.372314453125
],
[
0.5146484375,
0.359375,
0.8818359375,
0.359375,
0.8818359375,
0.372314453125,
0.5146484375,
0.372314453125
],
[
0.12005615234375,
0.374755859375,
0.48779296875,
0.374755859375,
0.48779296875,
0.388427734375,
0.12005615234375,
0.388427734375
],
[
0.51318359375,
0.374755859375,
0.880859375,
0.374755859375,
0.880859375,
0.388427734375,
0.51318359375,
0.388427734375
],
[
0.118408203125,
0.3896484375,
0.486083984375,
0.3896484375,
0.486083984375,
0.402587890625,
0.118408203125,
0.402587890625
],
[
0.515625,
0.391357421875,
0.8818359375,
0.391357421875,
0.8818359375,
0.405029296875,
0.515625,
0.405029296875
],
[
0.12176513671875,
0.407958984375,
0.210693359375,
0.407958984375,
0.210693359375,
0.419189453125,
0.12176513671875,
0.419189453125
],
[
0.515625,
0.407958984375,
0.8818359375,
0.407958984375,
0.8818359375,
0.421630859375,
0.515625,
0.421630859375
],
[
0.1376953125,
0.4228515625,
0.48779296875,
0.4228515625,
0.48779296875,
0.435791015625,
0.1376953125,
0.435791015625
],
[
0.515625,
0.42333984375,
0.8818359375,
0.42333984375,
0.8818359375,
0.43701171875,
0.515625,
0.43701171875
],
[
0.12005615234375,
0.439453125,
0.414794921875,
0.439453125,
0.414794921875,
0.453125,
0.12005615234375,
0.453125
],
[
0.4072265625,
0.440673828125,
0.48779296875,
0.440673828125,
0.48779296875,
0.451904296875,
0.4072265625,
0.451904296875
],
[
0.5146484375,
0.440673828125,
0.8818359375,
0.440673828125,
0.8818359375,
0.45361328125,
0.5146484375,
0.45361328125
],
[
0.12005615234375,
0.454833984375,
0.48779296875,
0.454833984375,
0.48779296875,
0.468505859375,
0.12005615234375,
0.468505859375
],
[
0.51318359375,
0.4560546875,
0.880859375,
0.4560546875,
0.880859375,
0.4697265625,
0.51318359375,
0.4697265625
],
[
0.12176513671875,
0.47021484375,
0.48779296875,
0.47021484375,
0.48779296875,
0.48388671875,
0.12176513671875,
0.48388671875
],
[
0.51318359375,
0.471435546875,
0.88330078125,
0.47265625,
0.88330078125,
0.486328125,
0.51318359375,
0.485107421875
],
[
0.12005615234375,
0.48681640625,
0.489501953125,
0.48681640625,
0.489501953125,
0.50048828125,
0.12005615234375,
0.50048828125
],
[
0.5146484375,
0.48681640625,
0.8818359375,
0.48681640625,
0.8818359375,
0.501953125,
0.5146484375,
0.501953125
],
[
0.12005615234375,
0.5029296875,
0.48779296875,
0.5029296875,
0.48779296875,
0.5166015625,
0.12005615234375,
0.5166015625
],
[
0.51318359375,
0.501953125,
0.8818359375,
0.50439453125,
0.8818359375,
0.5185546875,
0.51318359375,
0.5166015625
],
[
0.12005615234375,
0.51953125,
0.48779296875,
0.51953125,
0.48779296875,
0.533203125,
0.12005615234375,
0.533203125
],
[
0.51318359375,
0.52099609375,
0.880859375,
0.52099609375,
0.880859375,
0.53369140625,
0.51318359375,
0.53369140625
],
[
0.12176513671875,
0.53759765625,
0.48779296875,
0.53759765625,
0.48779296875,
0.54736328125,
0.12176513671875,
0.54736328125
],
[
0.51318359375,
0.53515625,
0.69775390625,
0.5322265625,
0.69775390625,
0.5498046875,
0.51318359375,
0.552734375
],
[
0.12005615234375,
0.5517578125,
0.397216796875,
0.5517578125,
0.397216796875,
0.5654296875,
0.12005615234375,
0.5654296875
],
[
0.5341796875,
0.5517578125,
0.88330078125,
0.5517578125,
0.88330078125,
0.5654296875,
0.5341796875,
0.5654296875
],
[
0.139404296875,
0.5693359375,
0.486083984375,
0.5693359375,
0.486083984375,
0.580078125,
0.139404296875,
0.580078125
],
[
0.5146484375,
0.568359375,
0.88330078125,
0.568359375,
0.88330078125,
0.58203125,
0.5146484375,
0.58203125
],
[
0.12005615234375,
0.58447265625,
0.484375,
0.58447265625,
0.484375,
0.59716796875,
0.12005615234375,
0.59716796875
],
[
0.51318359375,
0.5830078125,
0.88330078125,
0.5849609375,
0.88330078125,
0.5986328125,
0.51318359375,
0.5966796875
],
[
0.12005615234375,
0.599609375,
0.48779296875,
0.599609375,
0.48779296875,
0.61328125,
0.12005615234375,
0.61328125
],
[
0.51318359375,
0.5986328125,
0.60546875,
0.60107421875,
0.60546875,
0.615234375,
0.51318359375,
0.61328125
],
[
0.118408203125,
0.615234375,
0.489501953125,
0.61328125,
0.489501953125,
0.62744140625,
0.118408203125,
0.6298828125
],
[
0.53076171875,
0.6162109375,
0.8818359375,
0.6162109375,
0.8818359375,
0.6298828125,
0.53076171875,
0.6298828125
],
[
0.1351318359375,
0.6318359375,
0.48779296875,
0.6318359375,
0.48779296875,
0.6455078125,
0.1351318359375,
0.6455078125
],
[
0.51318359375,
0.6318359375,
0.87890625,
0.6318359375,
0.87890625,
0.6455078125,
0.51318359375,
0.6455078125
],
[
0.118408203125,
0.646484375,
0.48779296875,
0.646484375,
0.48779296875,
0.66015625,
0.118408203125,
0.66015625
],
[
0.51318359375,
0.64794921875,
0.8818359375,
0.6484375,
0.8818359375,
0.6630859375,
0.51318359375,
0.662109375
],
[
0.118408203125,
0.6630859375,
0.48779296875,
0.6630859375,
0.48779296875,
0.6767578125,
0.118408203125,
0.6767578125
],
[
0.515625,
0.6650390625,
0.8818359375,
0.6650390625,
0.8818359375,
0.6787109375,
0.515625,
0.6787109375
],
[
0.5146484375,
0.6796875,
0.85986328125,
0.6796875,
0.85986328125,
0.693359375,
0.5146484375,
0.693359375
],
[
0.12176513671875,
0.69140625,
0.156982421875,
0.69140625,
0.156982421875,
0.703125,
0.12176513671875,
0.703125
],
[
0.156982421875,
0.6923828125,
0.26611328125,
0.6923828125,
0.26611328125,
0.70361328125,
0.156982421875,
0.70361328125
],
[
0.51318359375,
0.70703125,
0.623046875,
0.70703125,
0.623046875,
0.7216796875,
0.51318359375,
0.7216796875
],
[
0.12005615234375,
0.71533203125,
0.48779296875,
0.71533203125,
0.48779296875,
0.728515625,
0.12005615234375,
0.728515625
],
[
0.12005615234375,
0.732421875,
0.489501953125,
0.732421875,
0.489501953125,
0.7451171875,
0.12005615234375,
0.7451171875
],
[
0.51318359375,
0.72705078125,
0.88330078125,
0.7255859375,
0.88330078125,
0.74072265625,
0.51318359375,
0.7412109375
],
[
0.118408203125,
0.74658203125,
0.486083984375,
0.74658203125,
0.486083984375,
0.76025390625,
0.118408203125,
0.76025390625
],
[
0.51318359375,
0.74365234375,
0.8818359375,
0.7421875,
0.8818359375,
0.755859375,
0.51318359375,
0.75732421875
],
[
0.118408203125,
0.7626953125,
0.489501953125,
0.76318359375,
0.489501953125,
0.77783203125,
0.118408203125,
0.77685546875
],
[
0.51318359375,
0.76025390625,
0.88330078125,
0.76025390625,
0.88330078125,
0.77392578125,
0.51318359375,
0.77392578125
],
[
0.12005615234375,
0.779296875,
0.48779296875,
0.779296875,
0.48779296875,
0.7919921875,
0.12005615234375,
0.7919921875
],
[
0.51318359375,
0.77490234375,
0.73486328125,
0.775390625,
0.73486328125,
0.7890625,
0.51318359375,
0.7880859375
],
[
0.12176513671875,
0.7958984375,
0.48779296875,
0.7958984375,
0.48779296875,
0.80859375,
0.12176513671875,
0.80859375
],
[
0.5322265625,
0.7919921875,
0.8818359375,
0.7919921875,
0.8818359375,
0.8056640625,
0.5322265625,
0.8056640625
],
[
0.12005615234375,
0.8125,
0.486083984375,
0.8125,
0.486083984375,
0.826171875,
0.12005615234375,
0.826171875
],
[
0.5146484375,
0.80810546875,
0.8818359375,
0.80810546875,
0.8818359375,
0.82177734375,
0.5146484375,
0.82177734375
],
[
0.51708984375,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.51708984375,
0.83837890625
],
[
0.1368408203125,
0.83056640625,
0.486083984375,
0.8330078125,
0.486083984375,
0.84716796875,
0.1368408203125,
0.845703125
],
[
0.12005615234375,
0.8466796875,
0.48779296875,
0.8466796875,
0.48779296875,
0.85986328125,
0.12005615234375,
0.85986328125
],
[
0.5322265625,
0.84423828125,
0.8251953125,
0.8466796875,
0.8251953125,
0.86083984375,
0.5322265625,
0.859375
],
[
0.12005615234375,
0.859375,
0.175537109375,
0.859375,
0.175537109375,
0.869140625,
0.12005615234375,
0.869140625
],
[
0.5283203125,
0.85546875,
0.884765625,
0.85791015625,
0.884765625,
0.8759765625,
0.5283203125,
0.87353515625
],
[
0.139404296875,
0.87060546875,
0.459228515625,
0.87060546875,
0.459228515625,
0.88525390625,
0.139404296875,
0.88525390625
],
[
0.515625,
0.87158203125,
0.80615234375,
0.87158203125,
0.80615234375,
0.88525390625,
0.515625,
0.88525390625
],
[
0.14111328125,
0.8818359375,
0.32421875,
0.88427734375,
0.32421875,
0.89794921875,
0.14111328125,
0.8955078125
],
[
0.5283203125,
0.8828125,
0.8818359375,
0.88427734375,
0.8818359375,
0.89794921875,
0.5283203125,
0.896484375
],
[
0.139404296875,
0.896484375,
0.358642578125,
0.896484375,
0.358642578125,
0.91015625,
0.139404296875,
0.91015625
],
[
0.51318359375,
0.896484375,
0.8193359375,
0.89794921875,
0.8193359375,
0.9091796875,
0.51318359375,
0.90771484375
]
] |
[
"Name",
"Score",
"0.71",
"中获取一个实体的众多相邻实体。为了减少延",
"Ours",
"迟，此组件使用压缩稀疏行（CSR）格式的二",
"Ours (w/o linkgraph model)",
"0.61",
"进制稀疏矩阵将整个图存储在内存中，其中其",
"Ours (w/o hyperlinkgeneration)",
"",
"0.69",
"RDF2Vec (Ristoski et al., 2018)",
"0.69",
"0.52",
"行和列表示实体，其值表示相应实体对之间存",
"Wiki2vec",
"在超链接。由于此矩阵的大小通常很小，因此",
"可以轻松地将其存储在内存中。5注意，给定",
"Table 1: The results of Wikipedia2Vec and the",
"CSR矩阵中的行索引，获取其非零列索引（对",
"baselineentityembeddings ontheKORE dataset",
"应于对应干行索引的实体的相邻实体）的时间",
"复杂度为0(1)",
"尔曼秩相关系数。我们使用了两个流行的实体",
"Mention-entityDictionary",
"提及实体字典用",
"嵌人工具RDF2Vec(Ristoskietal.,2018)和",
"于生成3.2节中描述的超链接。该字典将实",
"Wiki2vec作为基线",
"体名称映射到其可能的引用实体，并根据从",
"我们还通过使用两个标准任务来评估单词",
"Wikipedia中的所有超链接获得的名称及其引",
"嵌人的质量：（1）使用GoogleWord类比",
"用实体创建。我们的工具从维基百科贞面中提",
"数据集（Mikolovetal.，2013a）的语义子集",
"取包含大量实体名称的字典中包含的所有单",
"（SEM）和句法子集（SYN）的单词类比任务，",
"词和短语。为了有效地实现这一点，我们使用",
"以及（2）使用两个标准数据集的单词相似性",
"Aho-Corasick算法，这是一种有效的字符串搜",
"任务，即SimLex-999（SL）（Hilletal.，2015)",
"索算法，使用从所有实体名称构造的有限状态",
"和WordSim-THMA353（WS）。根据过去的工",
"机。在检测到字典中的单词和短语后，我们的",
"作，我们报告了单词类比任务的准确性，以及",
"工具根据3.2节中描述的启发式方法将它们转",
"单词相似性任务的黄金得分与单词嵌入之间的",
"换为超链接",
"余弦相似性之间的Spearman等级相关系数。",
"通过以随机顺序同时选代维基百科中的页",
"作为这些任务的基线，我们使用了在gensim",
"面和链接图中的实体来训练嵌入。1",
"使用mw-",
"库3中实现的skip-gram模型（Mikolovetal.",
"parserfromhellMediaWiki解析器提取每个页",
"2013a)。6。0 (Rehirek and Sojka,2010)和",
"面中的文本和超链接。6我们不使用半结构化",
"在fastText工具0中实现的扩展skip-gram模",
"数据，例如表和信息框。我们还使用提及实体",
"型。1。0(Bojanowskietal.,2017)。我们使用",
"字典生成超链接。我们将嵌人作为浮动矩阵存",
"WikiExtractor9创建基线的训练语料库。在",
"储在共享内存中，并使用多个进程更新它。学",
"可能的范围内，我们使用相同的超参数来训练",
"习款人所需的线性代数运算使用基本线性代数",
"我们的模型和基线。10",
"子程序（BLAS）中的C函数实现",
"我们还报告了使用我们的工具和基线单",
"此外，我们的工具使用标记器来检测维基百",
"词嵌人工具进行培训所需的时间。请注意，",
"科页面中的单词。当前在我们的工具中实现了",
"RDF2Vec和Wiki2vec工具的训练是使用gen-",
"以下四个标记器：（1）实现Unicode文本分",
"sim实现的",
"割算法（Davis，2019）的多语言ICU标记器7",
"我们使用Python3进行了实验。6和Open",
"（2）使用空白字符拆分文本的简单基于规则",
"BLASO。3。3安装在c5d上。在亚马逊云科",
"的标记器，（3）用于中文的Jieba标记器8，以",
"技上部署了36个CPU核心的9xlarge实例。",
"及（4）用于日语的MeCab标记器XMATHX3",
"为了训练我们的模型和基准词嵌入模型，我们",
"使用了2018年4月版的英文维基百科转储",
"5",
"Experiments",
"5.1Results",
"我们进行了实验，以比较我们的工具与现有",
"工具的质量和效率。为了评估实体嵌人的质",
"表1显示了我们的模型和KORE数据集的基",
"量，我们使用了KORE实体相关性数据集",
"线实体嵌入模型的结果。11w/o链接图模型和",
"(Hoffartetal.，2012）。该数据集由21个实体",
"w/o超链接生成分别是消融研究禁用链接图模",
"组成，每个实体有20个相关实体，分数由人",
"型和自动生成超链接的结果",
"类评估。在过去的工作之后，我们报告了黄金",
"我们的模型成功地超越了RDF2Vec和",
"分数和实体嵌人之间的余弦相似度之间的斯皮",
"Wiki2vec模型，并在KORE数据集上取得",
"了最先进的结果。结果还表明，链接图模型和",
"5The size of the matrix of English Wikipedia is less",
"than 500 megabytes with our default hyper-parameter",
"https://github.com/attardi/wikiextractor",
"settings.",
"10We used the following settings: dim_size = 500,",
"https://github.com/earwig/mwparserfromhell",
"window = 5, negative = 5, iteration = 5",
"7http://site.icu-project.org",
"1lWe obtained the results of the RDF2Vec and",
"8https://github.com/fxsjy/jieba",
"Wiki2vec models from Ristoski et al. (2018)"
] |
{
"arxiv_id": "1812.06280",
"language": "zh",
"page_idx": 3
}
|
||
1812.06280-zh-4
|
1812.06280-zh-4.jpg
|
<table>
<tbody>
<tr>
<th> </th>
<th>$\overline{\mathrm{SEM}}$</th>
<th>$\overline{\mathrm{SL}}$ $N$</th>
<th>$\overline{WS}$</th>
<th>Time</th>
</tr>
<tr>
<td>$\overline{Ours}$</td>
<td>0.79</td>
<td>$i8$ 0.40</td>
<td>0.71 1</td>
<td>276min</td>
</tr>
<tr>
<td>Ours ( w/ o link graph model) </td>
<td>0.77</td>
<td>:7 0.39</td>
<td>0.70 9</td>
<td>170min</td>
</tr>
<tr>
<td>Ours (w/o hyperlink ${\mathrm{.~generation}}$</td>
<td>0.79</td>
<td>:7 0.39</td>
<td>0.72 9</td>
<td>211min</td>
</tr>
<tr>
<td>Ours $^{\prime }$word- based skip- gram</td>
<td>0.75</td>
<td>:7 0.36</td>
<td>0.70 $\tilde{j}$</td>
<td>154min</td>
</tr>
<tr>
<td>( $\tilde{R}$eh$\" {u}$rekandSojka, 2010) gensim</td>
<td>0.75</td>
<td>:7 0.37</td>
<td>0.70</td>
<td>197min</td>
</tr>
<tr>
<td>fastText Boianowski et al. ,2017</td>
<td>0.63</td>
<td>$'0$ 0.37</td>
<td>0.69</td>
<td>243min</td>
</tr>
</tbody>
</table>
$197\min(2019)$最近证明，通过将 BERT 与我们的工具$243\min$训练的实体嵌人相结合，其性能优于 BERT 和
Table 2: The results of Wikipedia2Vec and the baseline word embeddings on the word analogy and word similarity datasets.
超链接的自动生成提高了 KORE 数据集的性
## 能
的工具中实现的基于单词的 skip-gram 模型的
# 性能
表 2 显示了我们的模型与基准词嵌入模型在词类比和词相似性数据集上的结果。我们还通过禁用链接图和锚上下文模型来测试在我们
我们的模型在 SEM 数据集以及两个单词相似度数据集上的表现均优于基线单词嵌入模型。这表明链接图和锚上下文模型提供的实体的语义信号有利于提高词嵌入的质量。此外自动生成超链接的功能通常不会对这些数据集的性能做出贡献
我们基于单词的 skip-gram 模型的实现比gensim 和 fastText 快得多。此外，我们的完整模型的训练时间与基线词嵌入模型的训练时间相当
### 6 Interactive Demonstration
我们开发了一个基于 Web 的交互式演示，该演示使用户能够探索通过我们提出的工具学习的单词和实体的嵌入(请参见图 4 )。此演示使用户能够使用三种降维算法将嵌入可视化到二维或三维空间中，即 t 分布随机邻居嵌人 (t-SNE) (Maaten and Hinton, 2008), 均匀流形近似和投影(UMAP)(McInnes et al., 2018)和主成分分析 (PCA)。用户可以通过使用鼠标拖动和缩放来移动可视化的嵌人空间。此外，该演示还允许用户通过查询任意项目的类似项目(单词或实体)来探索嵌入
我们使用本文发布的 12 种语言的预训练嵌人作为目标嵌入。此外，我们还提供了未经链接图模型训练的英语嵌入，以允许用户定性地研究链接图模型如何影响结果嵌入
我们的演示是通过扩展 TensorFlow 嵌入投影仪开发的。$^{12}$演示可在 https://wikipedia2vec. github.io/demo
${\frac{^{12}\mathrm{https:}//\mathrm{pr}}\mathrm{ojector.tensorflow.org}}$
Use Cases
使用我们提出的工具学习的嵌入已经在最近的各种研究中得到了有效的使用。Poerner et al 知识增强的上下文化单词嵌入(i。e。,ERNIE (Zhang et al., 2019) ) 在无监督的问题回答和关系分类任务上，没有任何计算上昂贵的额外的 BERT 预训练。Yamada et al. (2018b) 基于我们的工具开发了基于神经网络的问答系统，并赢得了 NIPS2017 大会举办的比赛。Sato et al. (2017)、Chen et al.(2019)和 Yamada and Shindo (2019)基于我们的工具学习的嵌入，分别在命名实体识别、实体链接和文本分类任务上取得了最先进的结果。此外，Papalampidi et al. (2019)提出了一种神经网络模型，该模型使用我们的工具学习的实体嵌入来分析电影的情节结构。其他示例包括实体链接(Yamada et al.,2016; Eshel et al.,2017), 命名实体识别 (Lara-Clares and Garcia-Serrano, 2019) , 释义检测 (Duong et al.,2019),假新闻检测 (Singh et al., 2019) 和知识图完成(Shah et al., 2019)
# 8 Conclusions
在本文中，我们介绍了 Wikipedia2Vec,这是一个开源工具，用于从 Wikipedia 轻松有效地学习单词和实体的嵌人。我们的实验证明了与现有工具相比，所提出的工具在嵌入质量和训练效率方面的优势。此外，我们的工具已有效地用于许多最新的最先进的模型，这表明我们的工具对下游任务的有效性。我们还引入了一个基于 Web 的交互式演示，使用户能够探索学习的嵌入。本文发布了 12 种语言的源代码和预训练嵌人
### $\operatorname{References}$
Mohamed Al-Badrashiny, Jason Bolton, Arun Te-
$\mathop{\text{jasvi Chaganty, Kevin Clark, Craig Harman}}$, Lifu Huang, Matthew Lamm, Jinhao Lei, Di Lu, Xiaoman $\overset{'}{\operatorname*{\operatorname*{Pan}}}$,andothers. 2017.TinkerBell: Cross-lingual Cold-Start Knowledge Base Construction. In Text Analysis Conference.
Stefan Behnel, Robert Bradshaw, Craig Citro, Lisandro Dalcin, Dag Sverre Seljebotn, and Kurt Smith. 2011. Cython: The Best of Both Worlds. Computing in Science &Engineering, 13(2){:}31{-}39.
Piotr Bojanowski, Edouard Grave, Armand Joulin,
and Tomas Mikolov. 2017. Enriching Word Vectors with Subword Information. Transactions of
|
[
[
0.340087890625,
0.07183837890625,
0.52490234375,
0.0736083984375,
0.52490234375,
0.0885009765625,
0.340087890625,
0.08612060546875
],
[
0.54736328125,
0.0748291015625,
0.6396484375,
0.07659912109375,
0.6396484375,
0.0914306640625,
0.54736328125,
0.08905029296875
],
[
0.1275634765625,
0.0872802734375,
0.1561279296875,
0.08905029296875,
0.154541015625,
0.097412109375,
0.1259765625,
0.0955810546875
],
[
0.343505859375,
0.0885009765625,
0.5146484375,
0.0885009765625,
0.5146484375,
0.0955810546875,
0.343505859375,
0.0955810546875
],
[
0.1259765625,
0.09857177734375,
0.306396484375,
0.09857177734375,
0.306396484375,
0.11224365234375,
0.1259765625,
0.11224365234375
],
[
0.339111328125,
0.0955810546875,
0.8818359375,
0.1015625,
0.8818359375,
0.119384765625,
0.339111328125,
0.1134033203125
],
[
0.1259765625,
0.11102294921875,
0.33154296875,
0.11102294921875,
0.33154296875,
0.12469482421875,
0.1259765625,
0.12469482421875
],
[
0.341796875,
0.11224365234375,
0.51123046875,
0.11224365234375,
0.51123046875,
0.122314453125,
0.341796875,
0.122314453125
],
[
0.5087890625,
0.1181640625,
0.880859375,
0.1181640625,
0.880859375,
0.1318359375,
0.5087890625,
0.1318359375
],
[
0.1275634765625,
0.12469482421875,
0.306396484375,
0.12469482421875,
0.306396484375,
0.1353759765625,
0.1275634765625,
0.1353759765625
],
[
0.343505859375,
0.12469482421875,
0.52490234375,
0.12469482421875,
0.52490234375,
0.1353759765625,
0.343505859375,
0.1353759765625
],
[
0.1259765625,
0.1341552734375,
0.8818359375,
0.133056640625,
0.8818359375,
0.1478271484375,
0.1259765625,
0.1484375
],
[
0.12176513671875,
0.144287109375,
0.88330078125,
0.149658203125,
0.88330078125,
0.166259765625,
0.12176513671875,
0.160888671875
],
[
0.515625,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.179931640625,
0.515625,
0.179931640625
],
[
0.118408203125,
0.1728515625,
0.48779296875,
0.1728515625,
0.48779296875,
0.1859130859375,
0.118408203125,
0.1859130859375
],
[
0.1175537109375,
0.1859130859375,
0.486083984375,
0.1859130859375,
0.486083984375,
0.1995849609375,
0.1175537109375,
0.1995849609375
],
[
0.5146484375,
0.182861328125,
0.8818359375,
0.182861328125,
0.8818359375,
0.196533203125,
0.5146484375,
0.196533203125
],
[
0.12005615234375,
0.201904296875,
0.298828125,
0.203125,
0.298828125,
0.21435546875,
0.12005615234375,
0.213134765625
],
[
0.515625,
0.198974609375,
0.880859375,
0.198974609375,
0.880859375,
0.2120361328125,
0.515625,
0.2120361328125
],
[
0.51708984375,
0.21435546875,
0.8818359375,
0.21435546875,
0.8818359375,
0.22802734375,
0.51708984375,
0.22802734375
],
[
0.515625,
0.23095703125,
0.8818359375,
0.23095703125,
0.8818359375,
0.24462890625,
0.515625,
0.24462890625
],
[
0.12176513671875,
0.2476806640625,
0.489501953125,
0.2476806640625,
0.489501953125,
0.26123046875,
0.12176513671875,
0.26123046875
],
[
0.515625,
0.2476806640625,
0.884765625,
0.2476806640625,
0.884765625,
0.26123046875,
0.515625,
0.26123046875
],
[
0.12005615234375,
0.26416015625,
0.14111328125,
0.26416015625,
0.14111328125,
0.27783203125,
0.12005615234375,
0.27783203125
],
[
0.51318359375,
0.26318359375,
0.88330078125,
0.26318359375,
0.88330078125,
0.276611328125,
0.51318359375,
0.276611328125
],
[
0.1376953125,
0.283203125,
0.48779296875,
0.283203125,
0.48779296875,
0.29638671875,
0.1376953125,
0.29638671875
],
[
0.5146484375,
0.279052734375,
0.880859375,
0.279052734375,
0.880859375,
0.292236328125,
0.5146484375,
0.292236328125
],
[
0.12005615234375,
0.298583984375,
0.489501953125,
0.298583984375,
0.489501953125,
0.312255859375,
0.12005615234375,
0.312255859375
],
[
0.51318359375,
0.29443359375,
0.8818359375,
0.295654296875,
0.8818359375,
0.308837890625,
0.51318359375,
0.30810546875
],
[
0.12005615234375,
0.314208984375,
0.489501953125,
0.314208984375,
0.489501953125,
0.327880859375,
0.12005615234375,
0.327880859375
],
[
0.5146484375,
0.311279296875,
0.8818359375,
0.311279296875,
0.8818359375,
0.32470703125,
0.5146484375,
0.32470703125
],
[
0.12005615234375,
0.330810546875,
0.489501953125,
0.330810546875,
0.489501953125,
0.345703125,
0.12005615234375,
0.345703125
],
[
0.5146484375,
0.327880859375,
0.8818359375,
0.327880859375,
0.8818359375,
0.341552734375,
0.5146484375,
0.341552734375
],
[
0.12176513671875,
0.345703125,
0.156982421875,
0.347412109375,
0.1561279296875,
0.361083984375,
0.12005615234375,
0.359375
],
[
0.515625,
0.34326171875,
0.8818359375,
0.34326171875,
0.8818359375,
0.35693359375,
0.515625,
0.35693359375
],
[
0.1376953125,
0.365234375,
0.48779296875,
0.365234375,
0.48779296875,
0.37890625,
0.1376953125,
0.37890625
],
[
0.515625,
0.35986328125,
0.8818359375,
0.35986328125,
0.8818359375,
0.37353515625,
0.515625,
0.37353515625
],
[
0.5146484375,
0.3759765625,
0.8818359375,
0.3759765625,
0.8818359375,
0.3896484375,
0.5146484375,
0.3896484375
],
[
0.12176513671875,
0.3818359375,
0.48779296875,
0.3818359375,
0.48779296875,
0.3955078125,
0.12176513671875,
0.3955078125
],
[
0.5146484375,
0.39013671875,
0.880859375,
0.391357421875,
0.880859375,
0.40625,
0.5146484375,
0.405029296875
],
[
0.12005615234375,
0.39794921875,
0.48779296875,
0.39794921875,
0.48779296875,
0.410888671875,
0.12005615234375,
0.410888671875
],
[
0.5146484375,
0.407958984375,
0.8818359375,
0.407958984375,
0.8818359375,
0.421630859375,
0.5146484375,
0.421630859375
],
[
0.12005615234375,
0.41455078125,
0.490234375,
0.41455078125,
0.490234375,
0.427490234375,
0.12005615234375,
0.427490234375
],
[
0.12176513671875,
0.429931640625,
0.48779296875,
0.429931640625,
0.48779296875,
0.443603515625,
0.12176513671875,
0.443603515625
],
[
0.515625,
0.42578125,
0.8818359375,
0.42578125,
0.8818359375,
0.435791015625,
0.515625,
0.435791015625
],
[
0.5146484375,
0.440673828125,
0.6630859375,
0.440673828125,
0.6630859375,
0.45361328125,
0.5146484375,
0.45361328125
],
[
0.12005615234375,
0.446533203125,
0.2476806640625,
0.446533203125,
0.2476806640625,
0.460205078125,
0.12005615234375,
0.460205078125
],
[
0.1376953125,
0.465576171875,
0.489501953125,
0.465576171875,
0.489501953125,
0.478515625,
0.1376953125,
0.478515625
],
[
0.5087890625,
0.4697265625,
0.5498046875,
0.4697265625,
0.5498046875,
0.479736328125,
0.5087890625,
0.479736328125
],
[
0.5498046875,
0.4697265625,
0.654296875,
0.4697265625,
0.654296875,
0.48095703125,
0.5498046875,
0.48095703125
],
[
0.118408203125,
0.48095703125,
0.48779296875,
0.479736328125,
0.48779296875,
0.49462890625,
0.118408203125,
0.495361328125
],
[
0.12005615234375,
0.49755859375,
0.489501953125,
0.49755859375,
0.489501953125,
0.51123046875,
0.12005615234375,
0.51123046875
],
[
0.51318359375,
0.493408203125,
0.8818359375,
0.493408203125,
0.8818359375,
0.50732421875,
0.51318359375,
0.50732421875
],
[
0.12005615234375,
0.515625,
0.175537109375,
0.515625,
0.175537109375,
0.52685546875,
0.12005615234375,
0.52685546875
],
[
0.5263671875,
0.51025390625,
0.8818359375,
0.51025390625,
0.8818359375,
0.52392578125,
0.5263671875,
0.52392578125
],
[
0.515625,
0.525390625,
0.8818359375,
0.525390625,
0.8818359375,
0.5390625,
0.515625,
0.5390625
],
[
0.515625,
0.54150390625,
0.8818359375,
0.54150390625,
0.8818359375,
0.5546875,
0.515625,
0.5546875
],
[
0.1175537109375,
0.55029296875,
0.382080078125,
0.55029296875,
0.382080078125,
0.56396484375,
0.1175537109375,
0.56396484375
],
[
0.515625,
0.55810546875,
0.8818359375,
0.55810546875,
0.8818359375,
0.57177734375,
0.515625,
0.57177734375
],
[
0.51318359375,
0.572265625,
0.880859375,
0.572265625,
0.880859375,
0.5859375,
0.51318359375,
0.5859375
],
[
0.12005615234375,
0.58203125,
0.489501953125,
0.58203125,
0.489501953125,
0.595703125,
0.12005615234375,
0.595703125
],
[
0.5146484375,
0.59033203125,
0.87158203125,
0.59033203125,
0.87158203125,
0.60400390625,
0.5146484375,
0.60400390625
],
[
0.118408203125,
0.59716796875,
0.48779296875,
0.5966796875,
0.48779296875,
0.61083984375,
0.118408203125,
0.6123046875
],
[
0.5146484375,
0.60498046875,
0.8818359375,
0.60546875,
0.8818359375,
0.62060546875,
0.5146484375,
0.619140625
],
[
0.12005615234375,
0.615234375,
0.484375,
0.615234375,
0.484375,
0.6259765625,
0.12005615234375,
0.6259765625
],
[
0.5146484375,
0.62255859375,
0.88330078125,
0.62255859375,
0.88330078125,
0.6357421875,
0.5146484375,
0.6357421875
],
[
0.12005615234375,
0.63134765625,
0.48779296875,
0.63134765625,
0.48779296875,
0.64453125,
0.12005615234375,
0.64453125
],
[
0.51318359375,
0.63720703125,
0.62646484375,
0.63818359375,
0.62646484375,
0.65283203125,
0.51318359375,
0.65185546875
],
[
0.12005615234375,
0.64794921875,
0.484375,
0.64794921875,
0.484375,
0.65771484375,
0.12005615234375,
0.65771484375
],
[
0.12176513671875,
0.6630859375,
0.48779296875,
0.6630859375,
0.48779296875,
0.6767578125,
0.12176513671875,
0.6767578125
],
[
0.12176513671875,
0.677734375,
0.489501953125,
0.677734375,
0.489501953125,
0.6923828125,
0.12176513671875,
0.6923828125
],
[
0.51123046875,
0.6796875,
0.60791015625,
0.681640625,
0.60791015625,
0.69677734375,
0.51123046875,
0.69482421875
],
[
0.12005615234375,
0.69482421875,
0.48779296875,
0.69482421875,
0.48779296875,
0.7080078125,
0.12005615234375,
0.7080078125
],
[
0.51318359375,
0.703125,
0.88330078125,
0.703125,
0.88330078125,
0.71630859375,
0.51318359375,
0.71630859375
],
[
0.12005615234375,
0.71044921875,
0.486083984375,
0.71044921875,
0.486083984375,
0.7236328125,
0.12005615234375,
0.7236328125
],
[
0.5341796875,
0.71728515625,
0.8818359375,
0.71728515625,
0.8818359375,
0.73095703125,
0.5341796875,
0.73095703125
],
[
0.12005615234375,
0.72705078125,
0.48779296875,
0.72705078125,
0.48779296875,
0.74072265625,
0.12005615234375,
0.74072265625
],
[
0.5341796875,
0.72998046875,
0.8818359375,
0.72998046875,
0.8818359375,
0.74365234375,
0.5341796875,
0.74365234375
],
[
0.12176513671875,
0.74462890625,
0.411376953125,
0.74462890625,
0.411376953125,
0.7548828125,
0.12176513671875,
0.7548828125
],
[
0.53076171875,
0.7421875,
0.8818359375,
0.7412109375,
0.8818359375,
0.755859375,
0.53076171875,
0.75732421875
],
[
0.5322265625,
0.755859375,
0.88330078125,
0.755859375,
0.88330078125,
0.76953125,
0.5322265625,
0.76953125
],
[
0.139404296875,
0.7626953125,
0.48779296875,
0.7626953125,
0.48779296875,
0.775390625,
0.139404296875,
0.775390625
],
[
0.53076171875,
0.76953125,
0.8212890625,
0.76953125,
0.8212890625,
0.77978515625,
0.53076171875,
0.77978515625
],
[
0.12176513671875,
0.77783203125,
0.48779296875,
0.77783203125,
0.48779296875,
0.79150390625,
0.12176513671875,
0.79150390625
],
[
0.12176513671875,
0.7958984375,
0.486083984375,
0.7958984375,
0.486083984375,
0.8056640625,
0.12176513671875,
0.8056640625
],
[
0.51123046875,
0.79150390625,
0.880859375,
0.7919921875,
0.880859375,
0.80712890625,
0.51123046875,
0.8056640625
],
[
0.12005615234375,
0.81103515625,
0.393798828125,
0.81103515625,
0.393798828125,
0.82470703125,
0.12005615234375,
0.82470703125
],
[
0.5322265625,
0.8056640625,
0.8818359375,
0.8056640625,
0.8818359375,
0.8193359375,
0.5322265625,
0.8193359375
],
[
0.53564453125,
0.82080078125,
0.880859375,
0.82080078125,
0.880859375,
0.83056640625,
0.53564453125,
0.83056640625
],
[
0.1376953125,
0.8291015625,
0.48779296875,
0.8291015625,
0.48779296875,
0.8427734375,
0.1376953125,
0.8427734375
],
[
0.5341796875,
0.83447265625,
0.603515625,
0.83447265625,
0.603515625,
0.84423828125,
0.5341796875,
0.84423828125
],
[
0.59423828125,
0.83203125,
0.8818359375,
0.8330078125,
0.8818359375,
0.84716796875,
0.59423828125,
0.8466796875
],
[
0.1175537109375,
0.8427734375,
0.490234375,
0.84423828125,
0.490234375,
0.86083984375,
0.1175537109375,
0.859375
],
[
0.5341796875,
0.84716796875,
0.62109375,
0.84716796875,
0.62109375,
0.85791015625,
0.5341796875,
0.85791015625
],
[
0.12005615234375,
0.8642578125,
0.22412109375,
0.8642578125,
0.22412109375,
0.8759765625,
0.12005615234375,
0.8759765625
],
[
0.5146484375,
0.869140625,
0.880859375,
0.869140625,
0.880859375,
0.8828125,
0.5146484375,
0.8828125
],
[
0.5341796875,
0.8828125,
0.8818359375,
0.8828125,
0.8818359375,
0.89794921875,
0.5341796875,
0.89794921875
],
[
0.1292724609375,
0.892578125,
0.365234375,
0.89453125,
0.365234375,
0.912109375,
0.1292724609375,
0.91015625
],
[
0.53564453125,
0.89794921875,
0.88330078125,
0.89794921875,
0.88330078125,
0.90771484375,
0.53564453125,
0.90771484375
]
] |
[
"SEMSYNSLWSTime7",
"Use Cases",
"Ours",
"0.79.0.68.0.40.0.71.276min",
"Ours (w/o link graph model)",
"0.770.670.390.70170min使用我们提出的工具学习的嵌入已经在最近的",
"Ours (w/o hyperlink generation)",
"0.790.670.390.72211mir",
"，各种研究中得到了有效的使用。Poerneretal.",
"Ours (word-based skip-gram)",
"0.750.670.360.70154min1",
"gensim（RehirekandSojka,2010）0.750.670.370.70197min(2019）最近证明，通过将BERT与我们的工具",
"fastText（Bojanowskietal.，2017）0.630.700.370.69243min训练的实体嵌人相结合，其性能优于BERT和",
"知识增强的上下文化单词嵌入（i。e。，ERNIE",
"Table 2:The results of Wikipedia2Vec and the",
"baseline word embeddings on the word analogy and",
"(Zhangetal.，2019））在无监督的问题回答和",
"word similarity datasets.",
"关系分类任务上，没有任何计算上昂贵的额",
"外的BERT预训练。Yamadaetal.(2018b)",
"基于我们的工具开发了基于神经网络的问答",
"超链接的自动生成提高了KORE数据集的性",
"系统，并赢得了NIPS2017大会举办的比赛。",
"能",
"Sato et al. (2017)、Chen et al. (2019)和 Ya",
"表2显示了我们的模型与基准词嵌人模型",
"madaandShindo（2019）基于我们的工具学习",
"在词类比和词相似性数据集上的结果。我们还",
"的嵌入，分别在命名实体识别、实体链接和文",
"通过禁用链接图和锚上下文模型来测试在我们",
"本分类任务上取得了最先进的结果。此外，Pa",
"的工具中实现的基于单词的 skip-gram模型的",
"palampidietal.（2019）提出了一种神经网络模",
"性能",
"型，该模型使用我们的工具学习的实体嵌人来",
"我们的模型在SEM数据集以及两个单词相",
"分析电影的情节结构。其他示例包括实体链接",
"(Yamada et al., 2016; Eshel et al., 2017),命",
"似度数据集上的表现均优于基线单词嵌人模",
"名实体识别（Lara-Clares and Garcia-Serrano,",
"型。这表明链接图和锚上下文模型提供的实体",
"2019），释义检测（Duongetal.，2019），假",
"的语义信号有利于提高词嵌入的质量。此外，",
"自动生成超链接的功能通常不会对这些数据集",
"新闻检测（Singhetal..2019）和知识图完成",
"(Shah et al., 2019)",
"的性能做出贡献",
"我们基于单词的skip-gram模型的实现比",
"8",
"Conclusions",
"gensim和fastText快得多。此外，我们的完",
"整模型的训练时间与基线词嵌入模型的训练时",
"在本文中，我们介绍了Wikipedia2Vec，这是",
"间相当",
"个开源工具，用于从Wikipedia轻松有效地",
"学习单词和实体的嵌入。我们的实验证明了与",
"现有工具相比，所提出的工具在嵌人质量和训",
"6Interactive Demonstration",
"练效率方面的优势。此外，我们的工具已有效",
"地用于许多最新的最先进的模型，这表明我们",
"我们开发了一个基于Web的交互式演示，该",
"的工具对下游任务的有效性。我们还引入了",
"演示使用户能够探索通过我们提出的工具学",
"个基于Web的交互式演示，使用户能够探索",
"习的单词和实体的嵌入（请参见图4）。此演",
"学习的嵌人。本文发布了12种语言的源代码",
"示使用户能够使用三种降维算法将嵌入可视化",
"和预训练嵌入",
"到二维或三维空间中，即t分布随机邻居嵌",
"人(t-SNE）(MaatenandHinton,2008)，均",
"匀流形近似和投影（UMAP）（McInnes etal.,",
"References",
"2018）和主成分分析（PCA）。用户可以通过使",
"Mohamed Al-Badrashiny, Jason Bolton, Arun Te-",
"用鼠标拖动和缩放来移动可视化的嵌入空间。",
"jasvi Chaganty, Kevin Clark, Craig Harman,",
"此外，该演示还允许用户通过查询任意项目的",
"Lifu Huang, Matthew Lamm, Jinhao Lei, Di Lu,",
"类似项目（单词或实体）来探索嵌入",
"Xiaoman Pan, and others. 2017. TinkerBell:",
"Cross-lingual Cold-Start Knowledge Base Con-",
"我们使用本文发布的12种语言的预训练嵌",
"struction. In Text Analvsis Conference",
"人作为目标嵌人。此外，我们还提供了未经链",
"接图模型训练的英语嵌入，以充许用户定性地",
"Stefan Behnel, Robert Bradshaw, Craig Citro,",
"研究链接图模型如何影响结果嵌入",
"Lisandro Dalcin, Dag Sverre Seljebotn, and",
"KurtSmith.20ll.Cython:TheBestofBoth",
"我们的演示是通过扩展TensorFlow嵌入投",
"Worlds.",
"Computing in Science & Engineering,",
"影仪开发的。12演示可在https://wikipedlia2vec.",
"13(2):31-39",
"github.io/demo",
"Piotr Bojanowski, Edouard Grave, Armand Joulin,",
"and Tomas Mikolov. 2017. Enriching Word Vec",
"12https:/projector.tensorflow.org",
"tors withSubwordInformation.Transactions of"
] |
{
"arxiv_id": "1812.06280",
"language": "zh",
"page_idx": 4
}
|
||
2303.08120-zh-0
|
2303.08120-zh-0.jpg
|
Blind Video Deflickering by Neural Filtering with a Flawed Atlas
$\begin{array}{cc}\text{Chenyang Lei }^{1,2*}\text{ Xuanchi Ren }^{3,4*}\text{ Zhaoxiang Zhang }^{1}&\text{Qifeng Chen}^{5}\\\text{CAIR, HKISI-CAS }^{2}\text{Princeton University }^{3}\text{ University of Toronto}&^{4}\text{Vector Institute}&^{5}\text{HKUST}\end{array}$
<FigureHere>
Figure 1. Blind deflickering performance of our approach on Betty Boop (1934). Many videos can have flickering artifacts for various reasons. Our approach takes only the input video and removes the flicker automatically without any extra guidance.
$\operatorname{Abstract}$
$\mathrm{able}_{\mathbf{r}}$ath$ttps: / / chenyanglei. github. io/ deflicker .$
$1.^{\mathrm{ore}}{\mathrm{Introduction}}$
Many videos contain flickering artifacts; common causes of flicker include video processing algorithms, video generation algorithms, and capturing videos under specific situations. Prior work usually requires specific guidance such as the flickering frequency, manual annotations, or extra consistent videos to remove the flicker. In this work, we propose a general flicker removal framework that only receives a single flickering video as input without additional guidance. Since it is blind to a specific flickering type or guidance, we name this “blind deflickering.”The core of our approach is utilizing the neural atlas in cooperation with a neural filtering strategy. The neural atlas is a unified representation for all frames in a video that provides temporal consistency guidance but is flawed in many cases. To this end, a neural network is trained to mimic a flter to learn the consistent features (e.g., color, brightness) and avoid introducing the artifacts in the atlas. To validate our method, we construct a dataset that contains diverse real-world flickering videos. Extensive experiments show that our method achieves satisfying deflickering performance and even outperforms baselines that use extra guidance on a public benchmark. The source code is publicly avail- $^*$ Equal contribution.
一个高质量的视频通常在时间上是一致的，但是许多视频由于各种原因而遭受闪烁的伪像，如图$\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\operatorname*{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}\color{reed}{\mathrm{E}}\color{red}{\mathrm{E}}\color{red}{\mathrm{E}}$ 示。例如，旧电影的亮度可能非常不稳定，因为一些具有低质量硬件的旧相机无法将每个帧的曝光时间设置为相同$\mathbb{E}[16]$ 。此外，曝光时间非常短的高速相机可以捕获高频 (e。g。,60Hz)室内照明变化 [25]。有效的处理算法，例如增强$\mathbb{E}[38,45]$,着色$\mathbb{E}[29,60]$ 和样式转换$\mathbb{E}[33]$ ,在应用于时间一致的视频时可能会带来闪烁。来自视频生成方法的视频$\boxed{\mathrm{F}}[46,50,61]$ 也可能包含闪烁的伪像。由于时间上一致的视频通常在视觉上更令人愉悦，因此在视频处理$\boxed{\mathrm{E}}[9,13,14,54,59]$ 和计算摄影中非常需要从视频中去除闪烁
在这项工作中，我们对一种用于偏转的通用方法感兴趣：(1)它与闪烁的模式或水平无关(e。g。,老电影，高速摄像机，处理工件),(2)它只拍摄单个闪烁的视频，不需要其他指导(e。g。、闪烁类型、额外一致的视频)。也就是说，这个模型对闪烁的类型和指导是盲目的，我们将这个任务命名为盲目的偏转。由于盲属性，盲偏转具有非常广泛的应用
盲偏转非常具有挑战性，因为在没有任何额外指导的情况下很难在整个视频中强制执行时间一致性。现有技术通常使用特定知识为每种闪烁类型设计特定策略。例如，对于高速摄像机拍摄的慢动作视频，先前的工作$\mathbb{E}[25]$ 可以分析照明频率。对于图像处理算法处理的视频，盲视频时间一致性[31,32]通过在时间一致的
1
|
[
[
0.1568603515625,
0.1319580078125,
0.81396484375,
0.1331787109375,
0.81396484375,
0.15087890625,
0.1568603515625,
0.149658203125
],
[
0.2115478515625,
0.177978515625,
0.71142578125,
0.177978515625,
0.71142578125,
0.1988525390625,
0.2115478515625,
0.1988525390625
],
[
0.73388671875,
0.1817626953125,
0.85693359375,
0.179931640625,
0.85693359375,
0.1976318359375,
0.73388671875,
0.199462890625
],
[
0.08660888671875,
0.1962890625,
0.268798828125,
0.1988525390625,
0.268798828125,
0.216552734375,
0.08660888671875,
0.2139892578125
],
[
0.7197265625,
0.1962890625,
0.8759765625,
0.199462890625,
0.8759765625,
0.2171630859375,
0.7197265625,
0.2139892578125
],
[
0.89208984375,
0.1976318359375,
0.9794921875,
0.2008056640625,
0.97802734375,
0.216552734375,
0.890625,
0.21337890625
],
[
0.495849609375,
0.1988525390625,
0.701171875,
0.199462890625,
0.701171875,
0.2139892578125,
0.495849609375,
0.21337890625
],
[
0.275390625,
0.1988525390625,
0.483642578125,
0.2008056640625,
0.483642578125,
0.21533203125,
0.275390625,
0.21337890625
],
[
0.2418212890625,
0.248779296875,
0.29736328125,
0.251953125,
0.295654296875,
0.267578125,
0.240234375,
0.264404296875
],
[
0.441162109375,
0.248779296875,
0.498291015625,
0.251953125,
0.496826171875,
0.267578125,
0.439453125,
0.264404296875
],
[
0.63818359375,
0.25,
0.689453125,
0.251953125,
0.68896484375,
0.266357421875,
0.63623046875,
0.264404296875
],
[
0.833984375,
0.250732421875,
0.8857421875,
0.250732421875,
0.8857421875,
0.266357421875,
0.833984375,
0.266357421875
],
[
0.0906982421875,
0.271484375,
0.10540771484375,
0.271484375,
0.10540771484375,
0.302978515625,
0.0906982421875,
0.302978515625
],
[
0.250732421875,
0.33837890625,
0.291748046875,
0.343505859375,
0.289306640625,
0.356689453125,
0.2484130859375,
0.3515625
],
[
0.6435546875,
0.3447265625,
0.68115234375,
0.3447265625,
0.68115234375,
0.3515625,
0.6435546875,
0.3515625
],
[
0.8310546875,
0.339111328125,
0.88232421875,
0.341552734375,
0.88134765625,
0.35595703125,
0.830078125,
0.354248046875
],
[
0.455078125,
0.34521484375,
0.483642578125,
0.34521484375,
0.483642578125,
0.349853515625,
0.455078125,
0.349853515625
],
[
0.08905029296875,
0.356689453125,
0.108642578125,
0.35791015625,
0.10540771484375,
0.401611328125,
0.08660888671875,
0.40087890625
],
[
0.080078125,
0.429931640625,
0.89208984375,
0.429931640625,
0.89208984375,
0.4443359375,
0.080078125,
0.4443359375
],
[
0.080078125,
0.44580078125,
0.89208984375,
0.44580078125,
0.89208984375,
0.458984375,
0.080078125,
0.458984375
],
[
0.080078125,
0.460205078125,
0.143798828125,
0.460205078125,
0.143798828125,
0.4716796875,
0.080078125,
0.4716796875
],
[
0.23779296875,
0.484130859375,
0.31201171875,
0.486083984375,
0.31201171875,
0.501953125,
0.23779296875,
0.5
],
[
0.5048828125,
0.488037109375,
0.8408203125,
0.488037109375,
0.8408203125,
0.501953125,
0.5048828125,
0.501953125
],
[
0.53369140625,
0.498779296875,
0.6396484375,
0.498779296875,
0.6396484375,
0.5087890625,
0.53369140625,
0.5087890625
],
[
0.09722900390625,
0.5166015625,
0.4697265625,
0.517578125,
0.4697265625,
0.5322265625,
0.09722900390625,
0.53076171875
],
[
0.5390625,
0.52294921875,
0.890625,
0.52294921875,
0.890625,
0.5361328125,
0.5390625,
0.5361328125
],
[
0.08087158203125,
0.53271484375,
0.46728515625,
0.53271484375,
0.46728515625,
0.54736328125,
0.08087158203125,
0.54736328125
],
[
0.5048828125,
0.53515625,
0.890625,
0.5361328125,
0.890625,
0.55029296875,
0.5048828125,
0.5498046875
],
[
0.08087158203125,
0.548828125,
0.46728515625,
0.548828125,
0.46728515625,
0.5625,
0.08087158203125,
0.5625
],
[
0.50341796875,
0.5517578125,
0.890625,
0.55029296875,
0.890625,
0.56494140625,
0.50341796875,
0.56640625
],
[
0.08087158203125,
0.56298828125,
0.46728515625,
0.56298828125,
0.46728515625,
0.5771484375,
0.08087158203125,
0.5771484375
],
[
0.50634765625,
0.5693359375,
0.88818359375,
0.5693359375,
0.88818359375,
0.580078125,
0.50634765625,
0.580078125
],
[
0.08087158203125,
0.5791015625,
0.468017578125,
0.5791015625,
0.468017578125,
0.59228515625,
0.08087158203125,
0.59228515625
],
[
0.5048828125,
0.58349609375,
0.88818359375,
0.58349609375,
0.88818359375,
0.5966796875,
0.5048828125,
0.5966796875
],
[
0.08087158203125,
0.59326171875,
0.4697265625,
0.59326171875,
0.4697265625,
0.60693359375,
0.08087158203125,
0.60693359375
],
[
0.5048828125,
0.5986328125,
0.8896484375,
0.5986328125,
0.8896484375,
0.6123046875,
0.5048828125,
0.6123046875
],
[
0.08087158203125,
0.609375,
0.46728515625,
0.609375,
0.46728515625,
0.62255859375,
0.08087158203125,
0.62255859375
],
[
0.5048828125,
0.61279296875,
0.890625,
0.61279296875,
0.890625,
0.626953125,
0.5048828125,
0.626953125
],
[
0.080078125,
0.62255859375,
0.468017578125,
0.62353515625,
0.468017578125,
0.63818359375,
0.080078125,
0.63720703125
],
[
0.50341796875,
0.626953125,
0.8896484375,
0.62744140625,
0.8896484375,
0.642578125,
0.50341796875,
0.6416015625
],
[
0.08087158203125,
0.6396484375,
0.468017578125,
0.6396484375,
0.468017578125,
0.65283203125,
0.08087158203125,
0.65283203125
],
[
0.5048828125,
0.64306640625,
0.8896484375,
0.64306640625,
0.8896484375,
0.6572265625,
0.5048828125,
0.6572265625
],
[
0.07843017578125,
0.65283203125,
0.468017578125,
0.65380859375,
0.468017578125,
0.66845703125,
0.07843017578125,
0.66748046875
],
[
0.5048828125,
0.6591796875,
0.890625,
0.6591796875,
0.890625,
0.67236328125,
0.5048828125,
0.67236328125
],
[
0.07843017578125,
0.66748046875,
0.464111328125,
0.66845703125,
0.464111328125,
0.68310546875,
0.07843017578125,
0.681640625
],
[
0.50634765625,
0.673828125,
0.890625,
0.673828125,
0.890625,
0.68798828125,
0.50634765625,
0.68798828125
],
[
0.08087158203125,
0.6845703125,
0.468017578125,
0.6845703125,
0.468017578125,
0.69775390625,
0.08087158203125,
0.69775390625
],
[
0.50341796875,
0.6875,
0.73388671875,
0.68798828125,
0.73388671875,
0.70263671875,
0.50341796875,
0.7021484375
],
[
0.08087158203125,
0.70068359375,
0.464111328125,
0.70068359375,
0.464111328125,
0.7109375,
0.08087158203125,
0.7109375
],
[
0.52294921875,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.52294921875,
0.71826171875
],
[
0.080078125,
0.7138671875,
0.465576171875,
0.7138671875,
0.465576171875,
0.72802734375,
0.080078125,
0.72802734375
],
[
0.50341796875,
0.71826171875,
0.890625,
0.7197265625,
0.890625,
0.734375,
0.50341796875,
0.73291015625
],
[
0.080078125,
0.72900390625,
0.468017578125,
0.72802734375,
0.468017578125,
0.7421875,
0.080078125,
0.74365234375
],
[
0.5048828125,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.5048828125,
0.74853515625
],
[
0.08087158203125,
0.744140625,
0.468017578125,
0.744140625,
0.468017578125,
0.75830078125,
0.08087158203125,
0.75830078125
],
[
0.5048828125,
0.75048828125,
0.8759765625,
0.75048828125,
0.8759765625,
0.7646484375,
0.5048828125,
0.7646484375
],
[
0.080078125,
0.76025390625,
0.46728515625,
0.76025390625,
0.46728515625,
0.7734375,
0.080078125,
0.7734375
],
[
0.50341796875,
0.76513671875,
0.8896484375,
0.76513671875,
0.8896484375,
0.77880859375,
0.50341796875,
0.77880859375
],
[
0.08087158203125,
0.7744140625,
0.468017578125,
0.7744140625,
0.468017578125,
0.78857421875,
0.08087158203125,
0.78857421875
],
[
0.5048828125,
0.77978515625,
0.890625,
0.77978515625,
0.890625,
0.79345703125,
0.5048828125,
0.79345703125
],
[
0.080078125,
0.79052734375,
0.468017578125,
0.79052734375,
0.468017578125,
0.8037109375,
0.080078125,
0.8037109375
],
[
0.5048828125,
0.79541015625,
0.74853515625,
0.79541015625,
0.74853515625,
0.8095703125,
0.5048828125,
0.8095703125
],
[
0.07843017578125,
0.8037109375,
0.4697265625,
0.8046875,
0.4697265625,
0.8193359375,
0.07843017578125,
0.818359375
],
[
0.52392578125,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.52392578125,
0.82470703125
],
[
0.08087158203125,
0.8212890625,
0.46728515625,
0.8212890625,
0.46728515625,
0.83203125,
0.08087158203125,
0.83203125
],
[
0.5048828125,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.8388671875,
0.5048828125,
0.8388671875
],
[
0.080078125,
0.83544921875,
0.46728515625,
0.83544921875,
0.46728515625,
0.8486328125,
0.080078125,
0.8486328125
],
[
0.5048828125,
0.84130859375,
0.890625,
0.84130859375,
0.890625,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.08251953125,
0.8515625,
0.46728515625,
0.8515625,
0.46728515625,
0.8623046875,
0.08251953125,
0.8623046875
],
[
0.50341796875,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.08087158203125,
0.86572265625,
0.468017578125,
0.86572265625,
0.468017578125,
0.87890625,
0.08087158203125,
0.87890625
],
[
0.5048828125,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.5048828125,
0.88525390625
],
[
0.09234619140625,
0.88720703125,
0.2254638671875,
0.88720703125,
0.2254638671875,
0.90087890625,
0.09234619140625,
0.90087890625
],
[
0.50244140625,
0.88525390625,
0.890625,
0.88623046875,
0.890625,
0.90087890625,
0.50244140625,
0.8994140625
],
[
0.483642578125,
0.92822265625,
0.490234375,
0.92822265625,
0.490234375,
0.93310546875,
0.483642578125,
0.93310546875
]
] |
[
"Blind Video Deflickering by Neural Filtering with a Flawed Atlas",
"Chenyang Lei 1,2 * Xuanchi Ren 3,4 * Zhaoxiang Zhang 1",
"Qifeng Chen 5",
"1 CAIR, HKISI-CAS",
"4 Vector Institute",
"5HKUST",
"3 University of Toronto",
"2PrincetonUniversity",
"出",
"时",
"手",
"",
"d",
"",
"厂",
"手",
"",
"anno",
"Figure 1. Blind deflickering performance of our approach on Betty Boop (1934). Many videos can have flickering artifacts",
"for various reasons. Our approach takes only the input video and removes the flicker automatically without any extra",
"guidance.",
"Abstract",
"able.athttps://chenyanglei.github.io/deflicker",
"Introduction",
"Many videos contain fickering artifacts; common",
"个高质量的视频通常在时间上是一致的，但是许",
"causes of flicker include video processing algorithms,",
"多视频由于各种原因而遭受闪烁的伪像，如图E2所",
"video generation algorithms, and capturing videos un-",
"示。例如，旧电影的亮度可能非常不稳定，因为一些具",
"der specific situations. Prior work usually requires spe-",
"有低质量硬件的旧相机无法将每个顿的曝光时间设置",
"cific guidance such as the flickering frequency, manual",
"为相同E[16]。此外，曝光时间非常短的高速相机可以",
"annotations, or extra consistent videos to remove the",
"捕获高频（e。g。，60Hz）室内照明变化[25]。有效的",
"ficker. In this work, we propose a general flicker re-",
"处理算法，例如增强E[38,45]，着色E[29,60]和样式",
"moval framework that only receives a single flickering",
"转换E[33]，在应用于时间一致的视频时可能会带来闪",
"video as input without additional guidance. Since it",
"烁。来自视频生成方法的视频E[46,50,61]也可能包含",
"is blind to a specific flickering type or guidance, we",
"闪烁的伪像。由于时间上一致的视频通常在视觉上更",
"name this“blind deflickering.\" The core of our ap",
"令人愉悦，因此在视频处理E[9,13,14,54,59]和计算摄",
"proach is utilizing the neural atlas in cooperation with",
"影中非常需要从视频中去除闪烁",
"a neural filtering strategy.The neuralatlas is a uni",
"在这项工作中，我们对一种用于偏转的通用方法感",
"fied representation for all frames in a video that pro-",
"兴趣：（1）它与闪烁的模式或水平无关（e。g。，老电",
"vides temporal consistency guidance but is flawed in",
"影，高速摄像机，处理工件），（2）它只拍摄单个闪烁",
"many cases. To this end, a neural network is trained",
"的视频，不需要其他指导（e。g。、闪烁类型、额外",
"to mimic a filter to learn the consistent features (e.g.",
"致的视频）。也就是说，这个模型对闪烁的类型和指导",
"color, brightness) and avoid introducing the artifacts",
"是盲目的，我们将这个任务命名为盲目的偏转。由于盲",
"in the atlas.To validate our method, we construct",
"属性，盲偏转具有非常广泛的应用",
"a dataset that contains diverse real-world flickering",
"盲偏转非常具有挑战性，因为在没有任何额外指导",
"videos. Extensive experiments show that our method",
"的情况下很难在整个视频中强制执行时间一致性。现",
"achieves satisfying deflickering performance and even",
"有技术通常使用特定知识为每种闪烁类型设计特定策",
"outperforms baselines that use extra guidance on a",
"略。例如，对于高速摄像机拍摄的慢动作视频，先前的",
"public benchmark. The source code is publicly avail-",
"工作E251可以分析照明频率。对于图像处理算法处理",
"* Equal contribution.",
"的视频，盲视频时间一致性[31,32]通过在时间一致的",
""
] |
{
"arxiv_id": "2303.08120",
"language": "zh",
"page_idx": 0
}
|
||
2303.08120-zh-1
|
2303.08120-zh-1.jpg
|
<FigureHere>
Old movies Old cartoons Time-lapse Slow-motion Colorization White balance Intrinsic
Text- to- video
Figure 2. Videos with flickering artifacts. Flickering artifacts exist in unprocessed videos, including old movies, old cartoons, time-lapse videos, and slow-motion videos. Besides, some processing algorithms [6,18,60] can introduce the flicker to temporally consistent unprocessed videos. Synthesized videos from video generations approaches [46,50,61] also might be temporal inconsistent. Blind deflickering aims to remove various types of flickering with only an unprocessed video as input.
2. Related Work
未处理视频上训练获得长期一致性。但是，闪烁类型或未处理的视频并不总是可用的，并且在这种情况下无法应用现有的闪烁特定算法。一种直观的解决方案是使用光流来跟踪对应关系。然而，来自闪烁视频的光流是不准确的，并且由于不准确的估计，光流的累积误差也随着帧数的增加而增加[7]
通过两个关键的观察和设计，我们成功地提出了第一种用于盲偏转的方法，该方法可以在没有额外指导或了解闪烁的情况下去除各种闪烁伪像。首先，我们利用一个名为神经图谱 [26] 的统一视频表示来解决解决长期不一致性的主要挑战。该神经图谱跟踪视频中的所有像素，并且不同帧中的对应关系在该图谱中共享相同的像素。因此，可以通过从共享图集中采样来获得时间上一致的帧序列。其次，尽管共享图集的帧是一致的，但图像的结构却存在缺陷：神经图集无法轻松地对具有大运动的动态对象进行建模；用于构建图集的光流不完善。因此，我们提出了一种神经过滤策略，从有缺陷的地图集中获取宝藏并扔掉垃圾。训练了一个神经网络来学习两种失真下的不变量，它分别模仿了图谱中的伪像和视频中的闪烁。在测试时，此网络可以很好地用作过滤器，以保留一致性属性并阻止有缺陷的图集中的工件。我们构建了包含各种类型的闪烁视频的第一个数据集，以忠实地评估盲偏转方法的性能。广泛的实验结果表明了我们的方法在处理不同闪烁方面的有效性。我们的方法还优于使用额外输入视频作为公共基准指导的盲视频时间一致性方法
我们的贡献可以总结如下：
特定于任务的 deflikcering。针对特定的闪烁类型设计了不同的策略。Kanj 等人。[25] 提出了一种高速摄像机的策略。Delon 等人。[16] 提出了一种局部对比度校正方法，可用于老电影和生物电影序列。Xu 等人。[56]关注来自基于 GAN 的编辑的时间闪烁伪像。通过特定的图像到图像翻译算法$\mathbb{E}[6,18,24,30,33,41,60,62]$ 处理的视频可能会出现闪烁伪影，并且盲目的视频时间一致性$\mathbb{E}[7,27,28,31,32,57]$ 旨在消除这些处理后的视频的闪烁。这些方法对特定的图像处理算法是盲目的。然而，生成帧的时间一致性由时间一致的未处理视频引导。Bonneel 等人。[7] 计算输入帧的梯度作为指导。Lai 等人。[27] 输入两个连续的输入帧作为指导。Lei 等人。[31] 直接学习输入和处理帧之间的映射函数。虽然这些方法在许多任务上都能达到令人满意的性能，但时间上一致的视频并不总是可用的。例如，对于许多闪烁的视频，例如老电影和来自视频生成方法的合成视频$\mathbb{E} [ 19, 50, 61]$ , 原始视频在时间上不一致。并发预印本$\mathbb{E}[2]$ 试图在推理过程中消除对未处理视频的需求。尽管如此，它仍然依赖于光流来实现局部时间一致性，并且不研究其他类型的闪烁视频。与盲时间一致性相比，我们的方法具有更广泛的应用
此外，一些商业软件$\mathbb{E}[3,47]$ 可以通过集成各种特定于任务的偏转方法来用于偏转。但是，这些方法要求用户具有闪烁类型的知识背景。我们的方法旨在消除此要求，以便可以为大多数用户有效地处理更多视频
视频马赛克和神经图谱。继承自全景图像拼接$\mathbb{E}[8]$ 视频马赛克是一种将视频数据组织成紧凑马赛克表示的技术，尤其是对于动态场景。它支持各种应用程序， 包括视频压缩$\mathbb{E}[23]$ ,视频索引$\mathbb{E}[22]$ ,视频纹理$\mathbb{E}[1]$ 2D 到 3D 转换$\mathrm\overline{\mathrm{E}}[48,49]$ 和视频编辑E[44]。基于单应性扭曲构建视频马赛克通常无法描绘运动。为了处理动态内容，研究人员通过时空蒙版$\mathbb{E}[15]$ 组成前景和背景区域，或者以基于补丁的方式将视频混合到多尺度挂毯中$\mathbb{E}[5]$ 。然而，这些方法严重依赖于图像外观信息，因此对照明变化和闪烁很敏感
最近，旨在实现一致的视频编辑，Kasten 等人。$\mathbb{E}[26]$ 提出了神经分层地图集(NLA),它通过学习地图集和视频帧之间的映射网络将视频分解为一组地图集。在图集上进行编辑，然后从图集中重建帧，可以实现一致
$\bullet$ We formulate the problem of blind deflickering and
construct a deflickering dataset containing diverse
flickering videos for further study.
$\bullet$ We propose the frst blind deflickering approach
to remove diverse flicker. We introduce the neural atlas to the deflickering problem and design a dedicated strategy to filter the flawed atlas for satisfying deflickering performance.
$\bullet$ Our method outperforms baselines on our dataset
and even outperforms methods that use extra in-
put videos on a public benchmark.
2
|
[
[
0.08660888671875,
0.106689453125,
0.09234619140625,
0.106689453125,
0.09234619140625,
0.12060546875,
0.08660888671875,
0.12060546875
],
[
0.08660888671875,
0.1572265625,
0.0906982421875,
0.1572265625,
0.0906982421875,
0.17041015625,
0.08660888671875,
0.17041015625
],
[
0.85302734375,
0.175537109375,
0.88134765625,
0.175537109375,
0.88134765625,
0.1787109375,
0.85302734375,
0.1787109375
],
[
0.09234619140625,
0.1944580078125,
0.69677734375,
0.1932373046875,
0.69677734375,
0.2083740234375,
0.09234619140625,
0.208984375
],
[
0.71826171875,
0.1962890625,
0.78125,
0.1962890625,
0.78125,
0.20703125,
0.71826171875,
0.20703125
],
[
0.80224609375,
0.1962890625,
0.896484375,
0.1962890625,
0.896484375,
0.20703125,
0.80224609375,
0.20703125
],
[
0.08416748046875,
0.2183837890625,
0.88818359375,
0.2183837890625,
0.88818359375,
0.2291259765625,
0.08416748046875,
0.2291259765625
],
[
0.08087158203125,
0.2310791015625,
0.890625,
0.2310791015625,
0.890625,
0.2442626953125,
0.08087158203125,
0.2442626953125
],
[
0.08416748046875,
0.246826171875,
0.8896484375,
0.246826171875,
0.8896484375,
0.2568359375,
0.08416748046875,
0.2568359375
],
[
0.080078125,
0.2587890625,
0.88818359375,
0.2587890625,
0.88818359375,
0.272705078125,
0.080078125,
0.272705078125
],
[
0.08251953125,
0.287353515625,
0.4697265625,
0.287353515625,
0.4697265625,
0.300537109375,
0.08251953125,
0.300537109375
],
[
0.5009765625,
0.285888671875,
0.650390625,
0.285888671875,
0.650390625,
0.2998046875,
0.5009765625,
0.2998046875
],
[
0.08251953125,
0.302978515625,
0.468017578125,
0.302978515625,
0.468017578125,
0.31640625,
0.08251953125,
0.31640625
],
[
0.08251953125,
0.317626953125,
0.468017578125,
0.317626953125,
0.468017578125,
0.330810546875,
0.08251953125,
0.330810546875
],
[
0.50341796875,
0.3125,
0.8896484375,
0.3125,
0.8896484375,
0.32568359375,
0.50341796875,
0.32568359375
],
[
0.50634765625,
0.326904296875,
0.890625,
0.326904296875,
0.890625,
0.34033203125,
0.50634765625,
0.34033203125
],
[
0.08416748046875,
0.333251953125,
0.468017578125,
0.333251953125,
0.468017578125,
0.3466796875,
0.08416748046875,
0.3466796875
],
[
0.5048828125,
0.34228515625,
0.890625,
0.34228515625,
0.890625,
0.35595703125,
0.5048828125,
0.35595703125
],
[
0.08416748046875,
0.349853515625,
0.46728515625,
0.349853515625,
0.46728515625,
0.35986328125,
0.08416748046875,
0.35986328125
],
[
0.5048828125,
0.356689453125,
0.8896484375,
0.356689453125,
0.8896484375,
0.37060546875,
0.5048828125,
0.37060546875
],
[
0.08251953125,
0.3642578125,
0.2900390625,
0.3642578125,
0.2900390625,
0.37744140625,
0.08251953125,
0.37744140625
],
[
0.50634765625,
0.37255859375,
0.890625,
0.37255859375,
0.890625,
0.386474609375,
0.50634765625,
0.386474609375
],
[
0.101318359375,
0.385009765625,
0.46728515625,
0.385009765625,
0.46728515625,
0.3984375,
0.101318359375,
0.3984375
],
[
0.5048828125,
0.386962890625,
0.88818359375,
0.386962890625,
0.88818359375,
0.40087890625,
0.5048828125,
0.40087890625
],
[
0.09234619140625,
0.3984375,
0.46728515625,
0.399658203125,
0.46728515625,
0.41357421875,
0.09234619140625,
0.412353515625
],
[
0.5048828125,
0.40283203125,
0.890625,
0.40283203125,
0.890625,
0.416748046875,
0.5048828125,
0.416748046875
],
[
0.08251953125,
0.416748046875,
0.465576171875,
0.416748046875,
0.465576171875,
0.4267578125,
0.08251953125,
0.4267578125
],
[
0.5048828125,
0.41845703125,
0.88623046875,
0.41845703125,
0.88623046875,
0.431884765625,
0.5048828125,
0.431884765625
],
[
0.08251953125,
0.429931640625,
0.468017578125,
0.429931640625,
0.468017578125,
0.443115234375,
0.08251953125,
0.443115234375
],
[
0.5048828125,
0.43310546875,
0.88623046875,
0.43310546875,
0.88623046875,
0.4462890625,
0.5048828125,
0.4462890625
],
[
0.08416748046875,
0.44580078125,
0.46728515625,
0.44580078125,
0.46728515625,
0.455810546875,
0.08416748046875,
0.455810546875
],
[
0.50341796875,
0.447509765625,
0.890625,
0.447509765625,
0.890625,
0.4609375,
0.50341796875,
0.4609375
],
[
0.08251953125,
0.460205078125,
0.468017578125,
0.460205078125,
0.468017578125,
0.473388671875,
0.08251953125,
0.473388671875
],
[
0.5048828125,
0.462158203125,
0.890625,
0.462158203125,
0.890625,
0.47607421875,
0.5048828125,
0.47607421875
],
[
0.08416748046875,
0.474853515625,
0.46728515625,
0.474853515625,
0.46728515625,
0.488037109375,
0.08416748046875,
0.488037109375
],
[
0.5048828125,
0.477783203125,
0.88818359375,
0.477783203125,
0.88818359375,
0.4912109375,
0.5048828125,
0.4912109375
],
[
0.08251953125,
0.4912109375,
0.46728515625,
0.4912109375,
0.46728515625,
0.501953125,
0.08251953125,
0.501953125
],
[
0.50341796875,
0.49365234375,
0.88818359375,
0.49365234375,
0.88818359375,
0.5068359375,
0.50341796875,
0.5068359375
],
[
0.08251953125,
0.5068359375,
0.46728515625,
0.5068359375,
0.46728515625,
0.517578125,
0.08251953125,
0.517578125
],
[
0.5048828125,
0.50830078125,
0.890625,
0.50830078125,
0.890625,
0.521484375,
0.5048828125,
0.521484375
],
[
0.08251953125,
0.52099609375,
0.468017578125,
0.52099609375,
0.468017578125,
0.5341796875,
0.08251953125,
0.5341796875
],
[
0.50341796875,
0.5234375,
0.8896484375,
0.5244140625,
0.8896484375,
0.5361328125,
0.50341796875,
0.53515625
],
[
0.08251953125,
0.537109375,
0.465576171875,
0.537109375,
0.465576171875,
0.54736328125,
0.08251953125,
0.54736328125
],
[
0.50341796875,
0.537109375,
0.890625,
0.53857421875,
0.890625,
0.55322265625,
0.50341796875,
0.5517578125
],
[
0.08251953125,
0.55029296875,
0.46728515625,
0.55029296875,
0.46728515625,
0.564453125,
0.08251953125,
0.564453125
],
[
0.5048828125,
0.5546875,
0.88623046875,
0.5546875,
0.88623046875,
0.56494140625,
0.5048828125,
0.56494140625
],
[
0.08251953125,
0.5673828125,
0.465576171875,
0.5673828125,
0.465576171875,
0.57763671875,
0.08251953125,
0.57763671875
],
[
0.5048828125,
0.568359375,
0.890625,
0.568359375,
0.890625,
0.58203125,
0.5048828125,
0.58203125
],
[
0.08087158203125,
0.580078125,
0.465576171875,
0.580078125,
0.465576171875,
0.59326171875,
0.08087158203125,
0.59326171875
],
[
0.50341796875,
0.58349609375,
0.88818359375,
0.583984375,
0.88818359375,
0.5986328125,
0.50341796875,
0.59765625
],
[
0.08416748046875,
0.59765625,
0.465576171875,
0.59765625,
0.465576171875,
0.60791015625,
0.08416748046875,
0.60791015625
],
[
0.51953125,
0.60107421875,
0.8759765625,
0.60107421875,
0.8759765625,
0.611328125,
0.51953125,
0.611328125
],
[
0.08251953125,
0.60986328125,
0.46728515625,
0.60986328125,
0.46728515625,
0.62353515625,
0.08251953125,
0.62353515625
],
[
0.50244140625,
0.6123046875,
0.79736328125,
0.61279296875,
0.79736328125,
0.62744140625,
0.50244140625,
0.626953125
],
[
0.08416748046875,
0.626953125,
0.455810546875,
0.626953125,
0.455810546875,
0.63720703125,
0.08416748046875,
0.63720703125
],
[
0.525390625,
0.62890625,
0.890625,
0.62890625,
0.890625,
0.642578125,
0.525390625,
0.642578125
],
[
0.08251953125,
0.6416015625,
0.46728515625,
0.6416015625,
0.46728515625,
0.65478515625,
0.08251953125,
0.65478515625
],
[
0.5048828125,
0.64453125,
0.88818359375,
0.64453125,
0.88818359375,
0.65771484375,
0.5048828125,
0.65771484375
],
[
0.08251953125,
0.65771484375,
0.465576171875,
0.65771484375,
0.465576171875,
0.66845703125,
0.08251953125,
0.66845703125
],
[
0.5048828125,
0.66015625,
0.89208984375,
0.66015625,
0.89208984375,
0.673828125,
0.5048828125,
0.673828125
],
[
0.08087158203125,
0.67041015625,
0.358642578125,
0.67041015625,
0.358642578125,
0.6845703125,
0.08087158203125,
0.6845703125
],
[
0.50634765625,
0.6748046875,
0.86376953125,
0.6748046875,
0.86376953125,
0.68798828125,
0.50634765625,
0.68798828125
],
[
0.101318359375,
0.6923828125,
0.289306640625,
0.6923828125,
0.289306640625,
0.70556640625,
0.101318359375,
0.70556640625
],
[
0.5048828125,
0.689453125,
0.8896484375,
0.689453125,
0.8896484375,
0.70263671875,
0.5048828125,
0.70263671875
],
[
0.50341796875,
0.7041015625,
0.890625,
0.705078125,
0.890625,
0.71826171875,
0.50341796875,
0.71728515625
],
[
0.5048828125,
0.72021484375,
0.8896484375,
0.72021484375,
0.8896484375,
0.734375,
0.5048828125,
0.734375
],
[
0.096435546875,
0.7373046875,
0.465576171875,
0.7373046875,
0.465576171875,
0.74755859375,
0.096435546875,
0.74755859375
],
[
0.50341796875,
0.73486328125,
0.88818359375,
0.734375,
0.88818359375,
0.74755859375,
0.50341796875,
0.74853515625
],
[
0.11358642578125,
0.75244140625,
0.465576171875,
0.75244140625,
0.465576171875,
0.76318359375,
0.11358642578125,
0.76318359375
],
[
0.51708984375,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.7646484375,
0.51708984375,
0.7646484375
],
[
0.11273193359375,
0.7666015625,
0.3544921875,
0.7666015625,
0.3544921875,
0.77978515625,
0.11273193359375,
0.77978515625
],
[
0.5048828125,
0.7666015625,
0.88818359375,
0.7666015625,
0.88818359375,
0.77978515625,
0.5048828125,
0.77978515625
],
[
0.096435546875,
0.7822265625,
0.468017578125,
0.7822265625,
0.468017578125,
0.79541015625,
0.096435546875,
0.79541015625
],
[
0.50341796875,
0.78076171875,
0.890625,
0.78076171875,
0.890625,
0.79443359375,
0.50341796875,
0.79443359375
],
[
0.111083984375,
0.796875,
0.46728515625,
0.79541015625,
0.46728515625,
0.8095703125,
0.111083984375,
0.81005859375
],
[
0.5048828125,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.5048828125,
0.81005859375
],
[
0.111083984375,
0.81103515625,
0.465576171875,
0.81103515625,
0.465576171875,
0.82470703125,
0.111083984375,
0.82470703125
],
[
0.5048828125,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.5048828125,
0.82470703125
],
[
0.11273193359375,
0.8271484375,
0.468017578125,
0.8271484375,
0.468017578125,
0.84033203125,
0.11273193359375,
0.84033203125
],
[
0.50341796875,
0.8271484375,
0.765625,
0.8271484375,
0.765625,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.111083984375,
0.84033203125,
0.36279296875,
0.84228515625,
0.36279296875,
0.85595703125,
0.111083984375,
0.85400390625
],
[
0.52294921875,
0.84130859375,
0.8896484375,
0.84130859375,
0.8896484375,
0.85498046875,
0.52294921875,
0.85498046875
],
[
0.096435546875,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.096435546875,
0.87060546875
],
[
0.5048828125,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.11273193359375,
0.87255859375,
0.46728515625,
0.87255859375,
0.46728515625,
0.88623046875,
0.11273193359375,
0.88623046875
],
[
0.50341796875,
0.87060546875,
0.890625,
0.87255859375,
0.890625,
0.88623046875,
0.50341796875,
0.8837890625
],
[
0.111083984375,
0.88720703125,
0.359375,
0.88623046875,
0.359375,
0.90087890625,
0.111083984375,
0.9013671875
],
[
0.5048828125,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.5048828125,
0.90087890625
],
[
0.48046875,
0.9248046875,
0.49267578125,
0.9248046875,
0.49267578125,
0.93798828125,
0.48046875,
0.93798828125
]
] |
[
"UTD",
"STTTY",
"",
"OldmoviesOld cartoonsTime-lapseSlow-motionColorization Whitebalance",
"Intrinsic",
"Text-to-videc",
"Figure2.Videoswithflickering artifacts.Flickeringartifactsexistinunprocessed videos,including old movies,old cartoons",
"time-lapse videos, and slow-motion videos. Besides, some processing algorithms [6, 18, 60] can introduce the flicker to",
"temporally consistentunprocessedvideos.Synthesized videosfromvideogenerationsapproaches46,50,6lalsomightbe",
"temporal inconsistent. Blind deflickering aims to remove various types of flickering with only an unprocessed video as input",
"未处理视频上训练获得长期一致性。但是，闪烁类型或",
"2. Related Work",
"未处理的视频并不总是可用的，并且在这种情况下无",
"法应用现有的闪烁特定算法。一种直观的解决方案是",
"特定于任务的deflikcering。针对特定的闪烁类型设计",
"了不同的策略。Kanj等人。[25]提出了一种高速摄",
"使用光流来跟踪对应关系。然而，来自闪烁视频的光流",
"像机的策略。Delon等人。[16]提出了一种局部对比",
"是不准确的，并且由于不准确的估计，光流的累积误差",
"度校正方法，可用于老电影和生物电影序列。Xu等人。",
"也随着顿数的增加而增加[7]",
"56】关注来自基于GAN的编辑的时间闪烁伪像。通过",
"通过两个关键的观察和设计，我们成功地提出了第",
"特定的图像到图像翻译算法[6,18,24,30,33,41,60,62",
"一种用于盲偏转的方法，该方法可以在没有额外指导",
"处理的视频可能会出现闪烁伪影，并且盲目的视频时",
"或了解闪烁的情况下去除各种闪烁伪像。首先，我们利",
"间一致性国[7,27,28,31,32,57]旨在消除这些处理后的",
"用一个名为神经图谱[26]的统一视频表示来解决解决",
"视频的闪烁。这些方法对特定的图像处理算法是盲目",
"长期不一致性的主要挑战。该神经图谱跟踪视频中的",
"的。然而，生成顿的时间一致性由时间一致的未处理",
"所有像素，并且不同顿中的对应关系在该图谱中共享",
"视频引导。Bonneel等人。[7]计算输人顿的梯度作为",
"相同的像素。因此，可以通过从共享图集中采样来获得",
"指导。Lai等人。[27]输人两个连续的输人顿作为指导",
"时间上一致的顿序列。其次，尽管共享图集的顿是一致",
"Lei等人。[31]直接学习输人和处理顿之间的映射函",
"的，但图像的结构却存在缺陷：神经图集无法轻松地对",
"数。虽然这些方法在许多任务上都能达到令人满意的",
"具有大运动的动态对象进行建模；用于构建图集的光",
"性能，但时间上一致的视频并不总是可用的。例如，对",
"流不完善。因此，我们提出了一种神经过滤策略，从有",
"于许多闪烁的视频，例如老电影和来自视频生成方法",
"缺陷的地图集中获取宝藏并扔掉垃圾。训练了一个神",
"的合成视频F19.50.611：原始视频在时间上不一致",
"经网络来学习两种失真下的不变量，它分别模仿了图",
"并发预印本E[2]试图在推理过程中消除对未处理视频",
"谱中的伪像和视频中的闪烁。在测试时，此网络可以很",
"的需求。尽管如此，它仍然依赖于光流来实现局部时间",
"好地用作过滤器，以保留一致性属性并阻止有缺陷的",
"致性，并且不研究其他类型的闪炼视频。与盲时间",
"图集中的工件。我们构建了包含各种类型的闪烁视频",
"致性相比，我们的方法具有更广泛的应用",
"的第一个数据集，以忠实地评估盲偏转方法的性能。了",
"此外，一些商业软件E[3,47]可以通过集成各种特定",
"泛的实验结果表明了我们的方法在处理不同闪烁方面",
"于任务的偏转方法来用于偏转。但是，这些方法要求用",
"的有效性。我们的方法还优于使用额外输入视频作为",
"户具有闪烁类型的知识背景。我们的方法旨在消除此",
"公共基准指导的盲视频时间一致性方法",
"要求，以便可以为大多数用户有效地处理更多视频",
"我们的贡献可以总结如下：",
"视频马赛克和神经图谱。继承自全景图像拼接E81",
"视频马赛克是一种将视频数据组织成紧凑马赛克表示",
"的技术，尤其是对于动态场景。它支持各种应用程序，",
".Weformulatetheproblemofblind deflickering and",
"包括视频压缩E23]，视频索引E[22]，视频纹理E[1",
"construct a deflickering dataset containing diverse",
"2D到3D转换E[48,49]和视频编辑E[44]。基于单",
"fickering videos for further study",
"应性扭曲构建视频马赛克通常无法描绘运动。为了处",
". We propose the first blind deflickering approach",
"理动态内容，研究人员通过时空蒙版E[15]组成前景和",
"to remove diverse flicker. We introduce the neu-",
"背景区域，或者以基于补丁的方式将视频混合到多尺",
"ral atlas to the deflickering problem and design",
"度挂毯中E51。然而，这些方法严重依赖于图像外观",
"a dedicated strategy to filter the flawed atlas for",
"信息，因此对照明变化和闪烁很敏感",
"satisfying deflickering performance",
"最近，旨在实现一致的视频编辑，Kasten等人。E[26]",
": Our method outperforms baselines on our dataset",
"提出了神经分层地图集（NLA），它通过学习地图集和",
"and even outperforms methods that use extra in-",
"视频顿之间的映射网络将视频分解为一组地图集。在",
"put videos on a public benchmark.",
"图集上进行编辑，然后从图集中重建顿，可以实现一致",
"2"
] |
{
"arxiv_id": "2303.08120",
"language": "zh",
"page_idx": 1
}
|
||
2303.08120-zh-2
|
2303.08120-zh-2.jpg
|
的视频编辑。后续工作验证了其在文本驱动视频风格化$\mathbb{E}[4,37]$ 和面部视频编辑$\mathbb{E}[36]$ 上的强大功能。直接采用 NLA 进行盲目性任务并不简单，主要受到两方面的限制：(i) 其在复杂场景中的性能仍然不能满足显着的工件。(ii) 它需要分割掩码作为分解动态对象的指导，并且每个动态对象都需要一个额外的映射网络。为了便于自动偏转，我们通过为有缺陷的图集设计有效的神经过滤策略，使用单层图集，而不需要分割蒙版。我们提出的策略也与其他地图集生成技术兼容
隐式图像/视频表示。随着使用多层感知器(MLP)作为 3D 几何[E[39,40,42] 的连续隐式表示的成功，这种表示在表示图像和视频$\boxed{\mathrm{F}}[10,34,51,52]$ 方面越来越受欢迎。后续工作将这些模型扩展到各种任务，例如图像超分辨率$\mathbb{E}[12]$ ,视频分解$\mathbb{E}[58]$ 和语义分割$\mathbb{E}[21]$ 。在我们的工作中，我们遵循[26]使用基于坐标的 MLP 来表示神经图谱
# 3. Method
# 3.1.0verview
令$\{I_t\}_{t=1}^T$为具有闪烁伪像的输入视频序列，其中$T$为视频帧数。我们的方法旨在消除闪烁的伪影，以生成时间上一致的视频$\{O_t\}_{t=1}^T$。闪烁表示时间不一致的一种类型，即不同帧中的对应共享不同的特征(e。g。、颜色、亮度)。闪烁在空间维度上可以是全局的或局部的，在时间维度上可以是长期的或短期的
图 3 显示了我们方法的框架。我们通过以下关键设计来解决盲目偏转的问题：(i)我们首先建议使用单个神经图谱 ($\mathbb{E}\color{red}3.2$节)进行偏转任务。$(ii)$我们为神经图谱设计了一种神经过滤(第$\boxed{\mathrm{E}}3.3$ 节)策略，因为单个图谱不可避免地存在缺陷
## 3.2. Flawed Atlas Generation
动机。一个好的盲偏转模型应该有能力跟踪所有视频帧的对应关系。视频处理中的大多数体系结构只能将少量帧作为输入，从而导致有限的接收场不足以实现长期一致性。为此，我们将神经图谱 [26]引入到偏转任务中，因为我们观察到它完全适合这个任务。神经图谱是视频中所有像素的统一而简洁的表示。如图$\color{red}{\mathrm{E}}\color{blue}{\mathrm{E}}\color{red}{\mathrm{3}}$ (a)所示，令$p=(x,y,t)\in\mathbb{R}^3$为位于帧$I_t$中$(x,y)$ 处的像素。每个像素$p$都被馈送到映射网络 M 中，该网络预测表示图集中像素相应位置的 2D 坐标$(u^p,v^p)$ 。理想情况下，即使输入像素的颜色不同，不同帧中的对应关系也应共享图集中的一个像素。也就是说，确保了时间上的一致性。培训。图 3(a)显示了生成图集的管道。对于每个像素$p$ ,我们有：
$$\begin{aligned}(u^p,v^p)&=\mathbb{M}(p),\\c^p&=\mathbb{A}(\phi(u^p),\phi(v^p)).\end{aligned}$$
(1)
(2)
将 2D 坐标$(u^p,v^p)$馈送到具有位置编码功能$\phi(\cdot)$的图集网络 A 中，以估计像素的 RGB 颜色$c^p$ 。通过优化原始 RGB 颜色和预测颜色$c^p$之间的损失，对映射
网络 M 和图集网络 A 进行联合训练。除了重建项之外，还采用一致性损失来鼓励将不同帧中的相应像素映射到相同的图集位置。我们遵循[26]中损失函数的实现。在训练网络 M 和 A 之后，我们可以通过提供整个视频中所有像素的坐标来重建视频$\{A_t\}_{t=1}^T$。重建的基于图集的视频$\{A_t\}_{t=1}^T$在时间上是一致的，因为所有像素都是从单个图集映射的。在这项工作中，我们利用该视频的时间一致性属性进行盲偏转
考虑到性能和效率之间的权衡，我们仅使用单层地图集来表示整个视频，尽管使用两层(背景层和前景层)或多层地图集可能会稍微提高性能。首先，在实践中，我们注意到层的数量是非常不同的，从单层到多层 (超过两层),这使得将它们自动应用于不同的视频具有挑战性。此外，我们注意到许多场景中不可避免地会出现伪影和失真，如[26] 中所述。我们介绍了如何在 Secion$\boxed{\mathrm{F}}3.3$中使用以下网络设计来处理有缺陷的地图集中的工件
# 3.3. Neural Filtering and Refinement
动机。神经图谱不仅包含宝藏，还包含垃圾。在$\color{red}{\mathrm{E3.2}}$ 节中，我们认为地图集是盲目偏转的有力提示，因为它可以在整个视频中提供一致的指导。然而，从地图集重建的帧是有缺陷的。首先，正如在 NLA$\mathbb{E}[26]$中分析的那样，当对象快速移动时，它不能很好地执行，或者多个对象可能需要多个层。对于盲偏转，我们需要去除闪烁并避免引入新的伪影。其次，闪烁的视频获得的光流不准确 , 这导致图谱中存在更多缺陷。最后，仍然存在一些自然变化，例如阴影变化，我们希望保留这些模式
因此，我们设计了一种神经过滤策略，该策略利用基于图集的视频的有前途的时间一致性属性，并防止伪像破坏我们的输出视频
训练策略。图$\color{red}3(\mathrm{b})$显示了使用图集的框架。给定一个输入视频$\{I_t\}_{t=1}^T$和一个图集$A$,在每次迭代中，我们从视频中获得一帧$I_t$ ,并将其输人到滤波器网络 $\mathbb{F}:$
$$O_t^f=\mathbb{F}(I_t,A_t),$$
(3)
,其中$A_t$是通过从共享图集$A$中获取像素而获得的
我们为有缺陷的地图集设计了专用的训练策略，如图$\boxed{\mathrm{E}}$4 所示。我们仅使用单个图像 $X$ 而不是连续帧来训练网络。在训练时间内，我们应用变换$\tau_a(\cdot)$来扭曲图像的外观，包括颜色，饱和度和亮度，这模仿了$I_t$ 中的闪烁模式。我们应用另一个变换$\tau_s(\cdot)$来扭曲图像的结构，模仿$A_t$中基于图集的有缺陷帧的失真。最后，通过最小化预测$\mathbb{F}(\tau_a(X),\tau_s(X))$和干净地面实况$X$之间的 L2 损失函数$\mathcal{L}$来训练网络 (i。e。,增强前的图像):
(4)
$$\mathcal{L}=||\mathbb{F}(\tau_a(X),\tau_s(X);\theta_\mathbb{F})-X||_2^2,$$
其中$\theta_\mathbb{F}$是滤波网络$\mathbb{F}$的参数
网络倾向于分别从两个扭曲的视图中学习不变部分。具体来说，$\mathbb{F}$从输入帧$I_t$和外观 (e。g。,亮度，颜色)
3
|
[
[
0.08087158203125,
0.09344482421875,
0.468017578125,
0.09344482421875,
0.468017578125,
0.106689453125,
0.08087158203125,
0.106689453125
],
[
0.5048828125,
0.0947265625,
0.88623046875,
0.0947265625,
0.88623046875,
0.10479736328125,
0.5048828125,
0.10479736328125
],
[
0.0849609375,
0.10986328125,
0.46728515625,
0.10986328125,
0.46728515625,
0.12060546875,
0.0849609375,
0.12060546875
],
[
0.5048828125,
0.10986328125,
0.88818359375,
0.10986328125,
0.88818359375,
0.12060546875,
0.5048828125,
0.12060546875
],
[
0.08251953125,
0.12371826171875,
0.46728515625,
0.12371826171875,
0.46728515625,
0.136962890625,
0.08251953125,
0.136962890625
],
[
0.5048828125,
0.125,
0.88818359375,
0.125,
0.88818359375,
0.1351318359375,
0.5048828125,
0.1351318359375
],
[
0.08251953125,
0.14013671875,
0.46728515625,
0.14013671875,
0.46728515625,
0.15087890625,
0.08251953125,
0.15087890625
],
[
0.50341796875,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1514892578125,
0.50341796875,
0.1514892578125
],
[
0.08087158203125,
0.154052734375,
0.468017578125,
0.154052734375,
0.468017578125,
0.1673583984375,
0.08087158203125,
0.1673583984375
],
[
0.50341796875,
0.154052734375,
0.890625,
0.154052734375,
0.890625,
0.1673583984375,
0.50341796875,
0.1673583984375
],
[
0.08251953125,
0.1673583984375,
0.46728515625,
0.1673583984375,
0.46728515625,
0.18115234375,
0.08251953125,
0.18115234375
],
[
0.50341796875,
0.1685791015625,
0.890625,
0.1685791015625,
0.890625,
0.1817626953125,
0.50341796875,
0.1817626953125
],
[
0.08416748046875,
0.184326171875,
0.468017578125,
0.184326171875,
0.468017578125,
0.1976318359375,
0.08416748046875,
0.1976318359375
],
[
0.50341796875,
0.184326171875,
0.890625,
0.184326171875,
0.890625,
0.1976318359375,
0.50341796875,
0.1976318359375
],
[
0.08087158203125,
0.1988525390625,
0.4697265625,
0.1988525390625,
0.4697265625,
0.212158203125,
0.08087158203125,
0.212158203125
],
[
0.50341796875,
0.1988525390625,
0.79736328125,
0.1988525390625,
0.79736328125,
0.212158203125,
0.50341796875,
0.212158203125
],
[
0.08251953125,
0.21533203125,
0.40283203125,
0.21533203125,
0.40283203125,
0.2259521484375,
0.08251953125,
0.2259521484375
],
[
0.52685546875,
0.21533203125,
0.8896484375,
0.21533203125,
0.8896484375,
0.2259521484375,
0.52685546875,
0.2259521484375
],
[
0.08087158203125,
0.2291259765625,
0.468017578125,
0.2279052734375,
0.468017578125,
0.242431640625,
0.08087158203125,
0.24365234375
],
[
0.5048828125,
0.2291259765625,
0.8896484375,
0.2291259765625,
0.8896484375,
0.242431640625,
0.5048828125,
0.242431640625
],
[
0.08087158203125,
0.2442626953125,
0.468017578125,
0.2442626953125,
0.468017578125,
0.25830078125,
0.08087158203125,
0.25830078125
],
[
0.5048828125,
0.24560546875,
0.8896484375,
0.24560546875,
0.8896484375,
0.255615234375,
0.5048828125,
0.255615234375
],
[
0.08251953125,
0.260009765625,
0.468017578125,
0.260009765625,
0.468017578125,
0.2734375,
0.08251953125,
0.2734375
],
[
0.5048828125,
0.2587890625,
0.8896484375,
0.2587890625,
0.8896484375,
0.272705078125,
0.5048828125,
0.272705078125
],
[
0.08416748046875,
0.276611328125,
0.46728515625,
0.276611328125,
0.46728515625,
0.287353515625,
0.08416748046875,
0.287353515625
],
[
0.50341796875,
0.274658203125,
0.890625,
0.274658203125,
0.890625,
0.28857421875,
0.50341796875,
0.28857421875
],
[
0.080078125,
0.28857421875,
0.4697265625,
0.2890625,
0.4697265625,
0.3037109375,
0.080078125,
0.302978515625
],
[
0.50341796875,
0.2890625,
0.890625,
0.2890625,
0.890625,
0.302978515625,
0.50341796875,
0.302978515625
],
[
0.08087158203125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.318115234375,
0.08087158203125,
0.318115234375
],
[
0.50341796875,
0.304931640625,
0.890625,
0.304931640625,
0.890625,
0.318115234375,
0.50341796875,
0.318115234375
],
[
0.08087158203125,
0.3193359375,
0.1968994140625,
0.3193359375,
0.1968994140625,
0.333251953125,
0.08087158203125,
0.333251953125
],
[
0.5048828125,
0.3193359375,
0.890625,
0.3193359375,
0.890625,
0.333251953125,
0.5048828125,
0.333251953125
],
[
0.50634765625,
0.3359375,
0.60107421875,
0.3359375,
0.60107421875,
0.347900390625,
0.50634765625,
0.347900390625
],
[
0.07843017578125,
0.348388671875,
0.176513671875,
0.348388671875,
0.176513671875,
0.363037109375,
0.07843017578125,
0.363037109375
],
[
0.50244140625,
0.35791015625,
0.79150390625,
0.359130859375,
0.79150390625,
0.373779296875,
0.50244140625,
0.37255859375
],
[
0.080078125,
0.375732421875,
0.19287109375,
0.375732421875,
0.19287109375,
0.389404296875,
0.080078125,
0.389404296875
],
[
0.5048828125,
0.38330078125,
0.89208984375,
0.38330078125,
0.89208984375,
0.396484375,
0.5048828125,
0.396484375
],
[
0.098876953125,
0.3984375,
0.46728515625,
0.3984375,
0.46728515625,
0.412353515625,
0.098876953125,
0.412353515625
],
[
0.5048828125,
0.399658203125,
0.8896484375,
0.399658203125,
0.8896484375,
0.410400390625,
0.5048828125,
0.410400390625
],
[
0.080078125,
0.41357421875,
0.468017578125,
0.41357421875,
0.468017578125,
0.4267578125,
0.080078125,
0.4267578125
],
[
0.50341796875,
0.41357421875,
0.8896484375,
0.41357421875,
0.8896484375,
0.4267578125,
0.50341796875,
0.4267578125
],
[
0.08087158203125,
0.4287109375,
0.46728515625,
0.4287109375,
0.46728515625,
0.442626953125,
0.08087158203125,
0.442626953125
],
[
0.5048828125,
0.427978515625,
0.890625,
0.427978515625,
0.890625,
0.44140625,
0.5048828125,
0.44140625
],
[
0.080078125,
0.442626953125,
0.468017578125,
0.443115234375,
0.468017578125,
0.458251953125,
0.080078125,
0.45703125
],
[
0.50341796875,
0.443115234375,
0.890625,
0.443115234375,
0.890625,
0.45703125,
0.50341796875,
0.45703125
],
[
0.080078125,
0.460205078125,
0.468017578125,
0.458984375,
0.468017578125,
0.472900390625,
0.080078125,
0.473388671875
],
[
0.5048828125,
0.458984375,
0.8896484375,
0.458984375,
0.8896484375,
0.472900390625,
0.5048828125,
0.472900390625
],
[
0.08251953125,
0.47607421875,
0.404296875,
0.47607421875,
0.404296875,
0.486083984375,
0.08251953125,
0.486083984375
],
[
0.50634765625,
0.474853515625,
0.8896484375,
0.474853515625,
0.8896484375,
0.48486328125,
0.50634765625,
0.48486328125
],
[
0.101318359375,
0.490478515625,
0.468017578125,
0.490478515625,
0.468017578125,
0.50390625,
0.101318359375,
0.50390625
],
[
0.5048828125,
0.488037109375,
0.890625,
0.488037109375,
0.890625,
0.501953125,
0.5048828125,
0.501953125
],
[
0.08251953125,
0.505859375,
0.468017578125,
0.505859375,
0.468017578125,
0.51953125,
0.08251953125,
0.51953125
],
[
0.5048828125,
0.5048828125,
0.8896484375,
0.5048828125,
0.8896484375,
0.51513671875,
0.5048828125,
0.51513671875
],
[
0.08087158203125,
0.51953125,
0.468017578125,
0.51953125,
0.468017578125,
0.53271484375,
0.08087158203125,
0.53271484375
],
[
0.5048828125,
0.5166015625,
0.52392578125,
0.51953125,
0.51953125,
0.5322265625,
0.5009765625,
0.52880859375
],
[
0.08087158203125,
0.53515625,
0.46728515625,
0.53515625,
0.46728515625,
0.548828125,
0.08087158203125,
0.548828125
],
[
0.52294921875,
0.5341796875,
0.8896484375,
0.5341796875,
0.8896484375,
0.54736328125,
0.52294921875,
0.54736328125
],
[
0.08251953125,
0.5498046875,
0.261474609375,
0.55029296875,
0.261474609375,
0.564453125,
0.08251953125,
0.56298828125
],
[
0.50634765625,
0.55029296875,
0.8896484375,
0.55029296875,
0.8896484375,
0.56103515625,
0.50634765625,
0.56103515625
],
[
0.50341796875,
0.56298828125,
0.68310546875,
0.56298828125,
0.68310546875,
0.5771484375,
0.50341796875,
0.5771484375
],
[
0.07843017578125,
0.57470703125,
0.310546875,
0.57568359375,
0.310546875,
0.59033203125,
0.07843017578125,
0.5888671875
],
[
0.5048828125,
0.5791015625,
0.890625,
0.5791015625,
0.890625,
0.59228515625,
0.5048828125,
0.59228515625
],
[
0.08087158203125,
0.599609375,
0.468017578125,
0.599609375,
0.468017578125,
0.61279296875,
0.08087158203125,
0.61279296875
],
[
0.5048828125,
0.5947265625,
0.89208984375,
0.5947265625,
0.89208984375,
0.60791015625,
0.5048828125,
0.60791015625
],
[
0.08251953125,
0.6162109375,
0.46728515625,
0.6162109375,
0.46728515625,
0.626953125,
0.08251953125,
0.626953125
],
[
0.5048828125,
0.611328125,
0.89208984375,
0.611328125,
0.89208984375,
0.6220703125,
0.5048828125,
0.6220703125
],
[
0.08416748046875,
0.6318359375,
0.465576171875,
0.6318359375,
0.465576171875,
0.642578125,
0.08416748046875,
0.642578125
],
[
0.640625,
0.63525390625,
0.75244140625,
0.63720703125,
0.75244140625,
0.6513671875,
0.640625,
0.6494140625
],
[
0.869140625,
0.63818359375,
0.890625,
0.63818359375,
0.890625,
0.6513671875,
0.869140625,
0.6513671875
],
[
0.08087158203125,
0.64599609375,
0.46728515625,
0.64599609375,
0.46728515625,
0.6591796875,
0.08087158203125,
0.6591796875
],
[
0.08251953125,
0.66015625,
0.46728515625,
0.66015625,
0.46728515625,
0.673828125,
0.08251953125,
0.673828125
],
[
0.498291015625,
0.66552734375,
0.88134765625,
0.6640625,
0.88134765625,
0.67822265625,
0.498291015625,
0.6787109375
],
[
0.08416748046875,
0.6767578125,
0.46728515625,
0.6767578125,
0.46728515625,
0.6875,
0.08416748046875,
0.6875
],
[
0.52392578125,
0.681640625,
0.8896484375,
0.681640625,
0.8896484375,
0.6923828125,
0.52392578125,
0.6923828125
],
[
0.08087158203125,
0.69140625,
0.46728515625,
0.69140625,
0.46728515625,
0.705078125,
0.08087158203125,
0.705078125
],
[
0.50341796875,
0.693359375,
0.890625,
0.6943359375,
0.890625,
0.708984375,
0.50341796875,
0.70849609375
],
[
0.08087158203125,
0.705078125,
0.468017578125,
0.7041015625,
0.468017578125,
0.71826171875,
0.08087158203125,
0.7197265625
],
[
0.5048828125,
0.71044921875,
0.8896484375,
0.71044921875,
0.8896484375,
0.7236328125,
0.5048828125,
0.7236328125
],
[
0.08087158203125,
0.72021484375,
0.46728515625,
0.72021484375,
0.46728515625,
0.734375,
0.08087158203125,
0.734375
],
[
0.5048828125,
0.72607421875,
0.890625,
0.72607421875,
0.890625,
0.7392578125,
0.5048828125,
0.7392578125
],
[
0.08660888671875,
0.74072265625,
0.09722900390625,
0.74072265625,
0.09722900390625,
0.744140625,
0.08660888671875,
0.744140625
],
[
0.0931396484375,
0.73486328125,
0.46728515625,
0.73486328125,
0.46728515625,
0.74853515625,
0.0931396484375,
0.74853515625
],
[
0.5048828125,
0.7412109375,
0.890625,
0.7412109375,
0.890625,
0.7548828125,
0.5048828125,
0.7548828125
],
[
0.08251953125,
0.75244140625,
0.46728515625,
0.75244140625,
0.46728515625,
0.76318359375,
0.08251953125,
0.76318359375
],
[
0.5048828125,
0.75830078125,
0.8896484375,
0.75830078125,
0.8896484375,
0.76806640625,
0.5048828125,
0.76806640625
],
[
0.0849609375,
0.7666015625,
0.46728515625,
0.7666015625,
0.46728515625,
0.77978515625,
0.0849609375,
0.77978515625
],
[
0.5048828125,
0.771484375,
0.890625,
0.771484375,
0.890625,
0.78515625,
0.5048828125,
0.78515625
],
[
0.08087158203125,
0.77978515625,
0.318603515625,
0.78076171875,
0.318603515625,
0.79541015625,
0.08087158203125,
0.79443359375
],
[
0.5048828125,
0.7861328125,
0.8896484375,
0.7861328125,
0.8896484375,
0.7998046875,
0.5048828125,
0.7998046875
],
[
0.5048828125,
0.80322265625,
0.564453125,
0.80322265625,
0.564453125,
0.814453125,
0.5048828125,
0.814453125
],
[
0.1805419921875,
0.8095703125,
0.294189453125,
0.8095703125,
0.294189453125,
0.82373046875,
0.1805419921875,
0.82373046875
],
[
0.4453125,
0.81005859375,
0.468017578125,
0.81005859375,
0.468017578125,
0.82470703125,
0.4453125,
0.82470703125
],
[
0.2156982421875,
0.8271484375,
0.370849609375,
0.82763671875,
0.370849609375,
0.84228515625,
0.2156982421875,
0.84130859375
],
[
0.4453125,
0.8291015625,
0.468017578125,
0.8291015625,
0.468017578125,
0.84326171875,
0.4453125,
0.84326171875
],
[
0.57861328125,
0.82763671875,
0.81640625,
0.8291015625,
0.81640625,
0.84326171875,
0.57861328125,
0.84228515625
],
[
0.82421875,
0.8359375,
0.84716796875,
0.8359375,
0.84716796875,
0.8388671875,
0.82421875,
0.8388671875
],
[
0.86767578125,
0.830078125,
0.890625,
0.830078125,
0.890625,
0.84326171875,
0.86767578125,
0.84326171875
],
[
0.08087158203125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.08087158203125,
0.87060546875
],
[
0.50341796875,
0.85693359375,
0.7158203125,
0.85693359375,
0.7158203125,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.08251953125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88623046875,
0.08251953125,
0.88623046875
],
[
0.52392578125,
0.87255859375,
0.8896484375,
0.87255859375,
0.8896484375,
0.88623046875,
0.52392578125,
0.88623046875
],
[
0.08251953125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.08251953125,
0.90087890625
],
[
0.50244140625,
0.88623046875,
0.8896484375,
0.88720703125,
0.8896484375,
0.9013671875,
0.50244140625,
0.90087890625
],
[
0.4794921875,
0.92529296875,
0.49267578125,
0.92529296875,
0.49267578125,
0.9375,
0.4794921875,
0.9375
]
] |
[
"的视频编辑。后续工作验证了其在文本驱动视频风格",
"网络M和图集网络A进行联合训练。除了重建项之",
"化F4，37和面部视频编辑E36上的强大功能。直接",
"外，还采用一致性损失来鼓励将不同顿中的相应像素",
"采用NLA进行盲目性任务并不简单，主要受到两方面",
"映射到相同的图集位置。我们遵循26中损失函数的",
"的限制：（i）其在复杂场景中的性能仍然不能满足显着",
"实现。在训练网络M和A之后，我们可以通过提供整",
"的工件。（i）它需要分割掩码作为分解动态对象的指",
"个视频中所有像素的坐标来重建视频【At}t=1.。重建",
"导，并且每个动态对象都需要一个额外的映射网络。为",
"的基于图集的视频At}=1在时间上是一致的，因为",
"了便于自动偏转，我们通过为有缺陷的图集设计有效",
"所有像素都是从单个图集映射的。在这项工作中，我们",
"的神经过滤策略，使用单层图集，而不需要分割蒙版。",
"利用该视频的时间一致性属性进行盲偏转",
"我们提出的策略也与其他地图集生成技术兼容",
"考虑到性能和效率之间的权衡，我们仅使用单层地",
"隐式图像/视频表示。随着使用多层感知器（MLP）作",
"图集来表示整个视频，尽管使用两层（背景层和前景",
"为3D几何E[39,40,42]的连续隐式表示的成功，这种",
"层）或多层地图集可能会稍微提高性能。首先，在实践",
"表示在表示图像和视频E10,34,51,52]方面越来越受",
"中，我们注意到层的数量是非常不同的，从单层到多",
"欢迎。后续工作将这些模型扩展到各种任务，例如图像",
"层（超过两层），这使得将它们自动应用于不同的视频",
"超分辨率E[12]，视频分解E[58]和语义分割E[21]。",
"具有挑战性。此外，我们注意到许多场景中不可避免地",
"在我们的工作中，我们遵循[26]使用基于坐标的MLP",
"会出现伪影和失真，如[26]中所述。我们介绍了如何",
"来表示神经图谱",
"在SecionE3.3中使用以下网络设计来处理有缺陷的地",
"图集中的工件",
"3. Method",
"3.3. Neural Filtering and Refinement",
"3.1. Overview",
"动机。神经图谱不仅包含宝藏，还包含垃圾。在E3.2",
"令I}-1为具有闪烁伪像的输人视频序列，其中",
"节中，我们认为地图集是盲自偏转的有力提示，因为它",
"T为视频帧数。我们的方法在消除闪烁的伪影，以",
"可以在整个视频中提供一致的指导。然而，从地图集重",
"生成时间上一致的视频{Ot]=1。闪烁表示时间不一",
"建的顿是有缺陷的。首先，正如在NLAE[26]中分析",
"致的一种类型，即不同顿中的对应共享不同的特征e。",
"的那样，当对象快速移动时，它不能很好地执行，或者",
"g。、颜色、亮度)。闪烁在空间维度上可以是全局的或",
"多个对象可能需要多个层。对于盲偏转，我们需要去除",
"局部的，在时间维度上可以是长期的或短期的",
"闪烁并避免弓入新的伪影。其次，闪烁的视频获得的光",
"图3显示了我们方法的框架。我们通过以下关键设",
"流不准确，这导致图谱中存在更多缺陷。最后，仍然存",
"计来解决盲目偏转的问题：（）我们首先建议使用单个",
"在一些自然变化，例如阴影变化，我们希望保留这些模",
"神经图谱（E3.2节）进行偏转任务。（ii）我们为神经图",
"式",
"谱设计了一种神经过滤（第E3.3节）策略，因为单个",
"因此，我们设计了一种神经过滤策略，该策略利用",
"图谱不可避免地存在缺陷",
"基于图集的视频的有前途的时间一致性属性，并防止",
"伪像破坏我们的输出视频",
"3.2.Flawed Atlas Generation",
"训练策略。图3（b）显示了使用图集的框架。给定一个",
"动机。一个好的盲偏转模型应该有能力跟踪所有视频",
"输人视频I}-1和一个图集A，在每次送代中，我",
"顿的对应关系。视频处理中的大多数体系结构只能将",
"们从视频中获得一顿I：并将其输入到滤波器网络F：",
"少量顿作为输入，从而导致有限的接收场不足以实现",
"Of = F(It, At),",
"(3)",
"长期一致性。为此，我们将神经图谱[26]引入到偏转",
"任务中，因为我们观察到它完全适合这个任务。神经",
"，其中At是通过从共享图集A中获取像素而获得的",
"图谱是视频中所有像素的统一而简洁的表示。如图F3",
"我们为有缺陷的地图集设计了专用的训练策略，如",
"(a）所示，令p=(a,y,t)ER3为位于顿I中(c,y)",
"图E4所示。我们仅使用单个图像X而不是连续顿来",
"处的像素。每个像素p都被馈送到映射网络M中，该",
"训练网络。在训练时间内，我们应用变换Ta（）来扭曲",
"网络预测表示图集中像素相应位置的2D坐标（uP,uP）",
"图像的外观，包括颜色，饱和度和亮度，这模仿了It",
"",
"理想情况下，即使输人像素的颜色不同，不同顿中的",
"中的闪烁模式。我们应用另一个变换T（）来扭曲图像",
"对应关系也应共享图集中的一个像素。也就是说，确保",
"的结构，模仿At中基于图集的有缺陷顿的失真。最",
"了时间上的一致性。培训。图3（a）显示了生成图集的",
"后，通过最小化预测F(Ta(X),Ts(X))和干净地面实况",
"管道。对于每个像素p，我们有：",
"X之间的L2损失函数C来训练网络(i。e。，增强前",
"的图像）：",
"(uP, uP) = M(p):",
"(1)",
"cP = A(Φ(uP), (vP)):",
"(2)",
"C = F(Ta(X), Ts(X); OF) - XIl2,",
"",
"(4)",
"将2D坐标（uP,uP）馈送到具有位置编码功能Φ(）的",
"其中是滤波网络F的参数",
"图集网络A中，以估计像素的RGB颜色cP。通过优",
"网络倾向于分别从两个扭曲的视图中学习不变部分。",
"化原始RGB颜色和预测颜色cP之间的损失，对映射",
"具体来说，F从输人顿It和外观（e。g。，亮度，颜色）",
"3"
] |
{
"arxiv_id": "2303.08120",
"language": "zh",
"page_idx": 2
}
|
||
2303.08120-zh-3
|
2303.08120-zh-3.jpg
|
<FigureHere>
Consistency Loss
(a) Flawed Atlas Generation (Sec. 3.2)
$A_t$
<FigureHere>
(b) Neural Filtering and Refinement (Sec. 3.3)
Figure 3. The framework of our approach. We first generate an atlas as a unified representation of the whole video,
providing consistent guidance for deflickering. Since the atlas is flawed, we then propose a neural filtering strategy to filter
${\tilde{\text{the flaws.}}}$
信息是在基于图集的帧$A_t$中编码的，因此可以强制执
行长期一致性
局部细化。视频帧$\{O_t^f\}_{t=1}^T$在短期和长期内都是全局一致的。但是，由于输人和基于图集的帧之间的未对准，它可能包含局部闪烁。因此，我们使用一个额外的局部偏转网络来进一步完善结果。先前的工作表明， 基于流的正则化可以很好地解决局部闪烁问题。因此， 我们选择了一个经过修改的轻量级管道 [27]。如图$\overline{\mathrm{F}}$ 3 ( b) 所 示 , 我 们 通 过 将 两 个 连 续 帧 $O_{t}^{f}, O_{t- 1}^{f}$和先前的输出$O_{t-1}$提供给我们的本地细化网络 正 来预测输出帧$O_t$ 。两个连续的帧首先跟随几个卷积层，然后与$O_{t-1}$融合。使用简单的时间一致性损失$\mathcal{L}_local$训练本地闪烁网络，以消除本地闪烁伪像：
<FigureHere>
$$\mathcal{L}_{local}(O_t,O_{t-1})=||M_{t,t-1}\odot(O_t-\hat{O}_{t-1})||_1,$$
(5)
Figure 4. Training pipeline of our neural filtering strategy. We apply two transformations to a clean image $X$ for mimicking the flickering input frame and the flawed atlas frame.
其中$\hat{O}_{t-1}$是通过使用从帧$t$到帧$t-1$的光流对$O_{t-1}$进行扭曲而获得的。$M_t,t-1$是对应的遮挡遮罩。对于没有局部伪影的帧，输出$O_t$应与$O_t^f$相同。因此， 我们还通过最小化$O_t$和$O_t^f$之间的距离来提供重建损失，以正则化质量
实现细节。网络$\mathbb{F}$是在 MS-COCO 数据集 [35]上训练的，因为我们只需要图像进行训练。我们为 20 纪元训练它，批量大小为 8。对于网络 L ,我们在已处理
来自图集框架$A_t$ ,因为它们对结构失真$\tau_\mathrm{s}$是不变的。同时，$\tau_s(X)$的失真不会通过网络$\mathbb{F}$传递。通过这种策略，我们用有缺陷的图谱实现了神经滤波的目标
请注意，虽然此网络$\mathbb{F}$仅接收一个帧，但由于时间
4
|
[
[
0.775390625,
0.11553955078125,
0.830078125,
0.11553955078125,
0.830078125,
0.1300048828125,
0.775390625,
0.1300048828125
],
[
0.31201171875,
0.1275634765625,
0.40283203125,
0.1256103515625,
0.40283203125,
0.14013671875,
0.31201171875,
0.1427001953125
],
[
0.51123046875,
0.1287841796875,
0.57861328125,
0.1287841796875,
0.57861328125,
0.14013671875,
0.51123046875,
0.14013671875
],
[
0.529296875,
0.143310546875,
0.56201171875,
0.143310546875,
0.56201171875,
0.177978515625,
0.529296875,
0.177978515625
],
[
0.41259765625,
0.154052734375,
0.47802734375,
0.154052734375,
0.47802734375,
0.1685791015625,
0.41259765625,
0.1685791015625
],
[
0.611328125,
0.1558837890625,
0.625,
0.1558837890625,
0.625,
0.1641845703125,
0.611328125,
0.1641845703125
],
[
0.65869140625,
0.17236328125,
0.744140625,
0.17236328125,
0.744140625,
0.184326171875,
0.65869140625,
0.184326171875
],
[
0.111083984375,
0.189453125,
0.1805419921875,
0.189453125,
0.1805419921875,
0.2008056640625,
0.111083984375,
0.2008056640625
],
[
0.398681640625,
0.2197265625,
0.498291015625,
0.2197265625,
0.498291015625,
0.2310791015625,
0.398681640625,
0.2310791015625
],
[
0.826171875,
0.2183837890625,
0.85107421875,
0.2183837890625,
0.85107421875,
0.2266845703125,
0.826171875,
0.2266845703125
],
[
0.74609375,
0.23291015625,
0.85107421875,
0.2342529296875,
0.85107421875,
0.248779296875,
0.74609375,
0.2474365234375
],
[
0.363525390625,
0.2412109375,
0.62353515625,
0.2412109375,
0.62353515625,
0.255126953125,
0.363525390625,
0.255126953125
],
[
0.68310546875,
0.24560546875,
0.70361328125,
0.24560546875,
0.70361328125,
0.2587890625,
0.68310546875,
0.2587890625
],
[
0.188720703125,
0.26953125,
0.461669921875,
0.270751953125,
0.461669921875,
0.2841796875,
0.188720703125,
0.28271484375
],
[
0.66259765625,
0.271484375,
0.82177734375,
0.271484375,
0.82177734375,
0.2822265625,
0.66259765625,
0.2822265625
],
[
0.7099609375,
0.3193359375,
0.7158203125,
0.3193359375,
0.7158203125,
0.324462890625,
0.7099609375,
0.324462890625
],
[
0.7255859375,
0.335205078125,
0.78125,
0.335205078125,
0.78125,
0.37744140625,
0.7255859375,
0.37744140625
],
[
0.26318359375,
0.34228515625,
0.27294921875,
0.34228515625,
0.27294921875,
0.349853515625,
0.26318359375,
0.349853515625
],
[
0.62158203125,
0.341552734375,
0.646484375,
0.341552734375,
0.646484375,
0.349853515625,
0.62158203125,
0.349853515625
],
[
0.11358642578125,
0.348388671875,
0.1829833984375,
0.348388671875,
0.1829833984375,
0.35986328125,
0.11358642578125,
0.35986328125
],
[
0.330078125,
0.37890625,
0.371826171875,
0.380615234375,
0.370849609375,
0.393310546875,
0.328369140625,
0.391357421875
],
[
0.73046875,
0.37890625,
0.77783203125,
0.37890625,
0.77783203125,
0.393310546875,
0.73046875,
0.393310546875
],
[
0.82275390625,
0.375,
0.83837890625,
0.375,
0.83837890625,
0.385009765625,
0.82275390625,
0.385009765625
],
[
0.410888671875,
0.3818359375,
0.4345703125,
0.3818359375,
0.4345703125,
0.40087890625,
0.410888671875,
0.40087890625
],
[
0.48779296875,
0.380126953125,
0.5400390625,
0.3818359375,
0.5390625,
0.401611328125,
0.486083984375,
0.399658203125
],
[
0.326904296875,
0.392578125,
0.374267578125,
0.39453125,
0.373291015625,
0.4072265625,
0.325927734375,
0.40478515625
],
[
0.72607421875,
0.39453125,
0.78515625,
0.39453125,
0.78515625,
0.406005859375,
0.72607421875,
0.406005859375
],
[
0.6044921875,
0.41845703125,
0.65185546875,
0.42041015625,
0.650390625,
0.436279296875,
0.60302734375,
0.434326171875
],
[
0.11358642578125,
0.429931640625,
0.191162109375,
0.429931640625,
0.191162109375,
0.43994140625,
0.11358642578125,
0.43994140625
],
[
0.34228515625,
0.451904296875,
0.65673828125,
0.45068359375,
0.65673828125,
0.46533203125,
0.34228515625,
0.466552734375
],
[
0.08087158203125,
0.48486328125,
0.88818359375,
0.48486328125,
0.88818359375,
0.498779296875,
0.08087158203125,
0.498779296875
],
[
0.080078125,
0.5,
0.89208984375,
0.5,
0.89208984375,
0.51318359375,
0.080078125,
0.51318359375
],
[
0.080078125,
0.5146484375,
0.143798828125,
0.5146484375,
0.143798828125,
0.52587890625,
0.080078125,
0.52587890625
],
[
0.1953125,
0.53515625,
0.298095703125,
0.5341796875,
0.298095703125,
0.54541015625,
0.1953125,
0.546875
],
[
0.5048828125,
0.5390625,
0.890625,
0.5390625,
0.890625,
0.55322265625,
0.5048828125,
0.55322265625
],
[
0.50244140625,
0.5537109375,
0.60302734375,
0.5517578125,
0.60302734375,
0.56640625,
0.50244140625,
0.568359375
],
[
0.50341796875,
0.568359375,
0.890625,
0.568359375,
0.890625,
0.58349609375,
0.50341796875,
0.58349609375
],
[
0.51708984375,
0.58544921875,
0.8896484375,
0.58544921875,
0.8896484375,
0.5986328125,
0.51708984375,
0.5986328125
],
[
0.50341796875,
0.60107421875,
0.88818359375,
0.60107421875,
0.88818359375,
0.6142578125,
0.50341796875,
0.6142578125
],
[
0.5048828125,
0.61572265625,
0.88623046875,
0.61572265625,
0.88623046875,
0.62890625,
0.5048828125,
0.62890625
],
[
0.5048828125,
0.6298828125,
0.88818359375,
0.6298828125,
0.88818359375,
0.64453125,
0.5048828125,
0.64453125
],
[
0.50341796875,
0.64599609375,
0.890625,
0.64599609375,
0.890625,
0.6591796875,
0.50341796875,
0.6591796875
],
[
0.50244140625,
0.66015625,
0.88818359375,
0.66015625,
0.88818359375,
0.6748046875,
0.50244140625,
0.6748046875
],
[
0.431396484375,
0.68115234375,
0.438720703125,
0.67822265625,
0.441162109375,
0.68115234375,
0.435546875,
0.6845703125
],
[
0.5048828125,
0.6767578125,
0.8896484375,
0.6767578125,
0.8896484375,
0.69140625,
0.5048828125,
0.69140625
],
[
0.5048828125,
0.6923828125,
0.890625,
0.6923828125,
0.890625,
0.70556640625,
0.5048828125,
0.70556640625
],
[
0.50341796875,
0.70703125,
0.890625,
0.70703125,
0.890625,
0.7216796875,
0.50341796875,
0.7216796875
],
[
0.50341796875,
0.72265625,
0.7548828125,
0.72265625,
0.7548828125,
0.736328125,
0.50341796875,
0.736328125
],
[
0.529296875,
0.74755859375,
0.890625,
0.74755859375,
0.890625,
0.76220703125,
0.529296875,
0.76220703125
],
[
0.205078125,
0.76025390625,
0.238525390625,
0.76025390625,
0.238525390625,
0.767578125,
0.205078125,
0.767578125
],
[
0.236083984375,
0.76025390625,
0.279296875,
0.76025390625,
0.279296875,
0.767578125,
0.236083984375,
0.767578125
],
[
0.08087158203125,
0.77880859375,
0.468017578125,
0.77880859375,
0.468017578125,
0.79248046875,
0.08087158203125,
0.79248046875
],
[
0.5185546875,
0.77587890625,
0.890625,
0.7763671875,
0.890625,
0.79150390625,
0.5185546875,
0.79052734375
],
[
0.08087158203125,
0.79248046875,
0.468017578125,
0.79248046875,
0.468017578125,
0.80615234375,
0.08087158203125,
0.80615234375
],
[
0.50341796875,
0.79248046875,
0.89208984375,
0.79248046875,
0.89208984375,
0.806640625,
0.50341796875,
0.806640625
],
[
0.080078125,
0.80615234375,
0.468017578125,
0.80615234375,
0.468017578125,
0.8193359375,
0.080078125,
0.8193359375
],
[
0.50341796875,
0.8095703125,
0.890625,
0.80810546875,
0.890625,
0.82275390625,
0.50341796875,
0.82373046875
],
[
0.5048828125,
0.82568359375,
0.89208984375,
0.82568359375,
0.89208984375,
0.84033203125,
0.5048828125,
0.84033203125
],
[
0.08251953125,
0.8388671875,
0.468017578125,
0.8388671875,
0.468017578125,
0.85302734375,
0.08251953125,
0.85302734375
],
[
0.50341796875,
0.84130859375,
0.6357421875,
0.84130859375,
0.6357421875,
0.85498046875,
0.50341796875,
0.85498046875
],
[
0.08251953125,
0.85400390625,
0.468017578125,
0.85400390625,
0.468017578125,
0.86865234375,
0.08251953125,
0.86865234375
],
[
0.50341796875,
0.85595703125,
0.890625,
0.85595703125,
0.890625,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.08251953125,
0.869140625,
0.439453125,
0.869140625,
0.439453125,
0.88330078125,
0.08251953125,
0.88330078125
],
[
0.50341796875,
0.8720703125,
0.89208984375,
0.8720703125,
0.89208984375,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.101318359375,
0.88623046875,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.101318359375,
0.90087890625
],
[
0.50244140625,
0.88623046875,
0.8896484375,
0.88623046875,
0.8896484375,
0.90087890625,
0.50244140625,
0.90087890625
],
[
0.4794921875,
0.923828125,
0.49267578125,
0.9248046875,
0.491943359375,
0.93603515625,
0.47802734375,
0.93505859375
]
] |
[
"(up,vp)",
"Mapping MLP",
"Atlas MLP",
"川",
"(up,vp)",
"cp",
"FlawedAtlas",
"Input video",
"UVmap,framet",
"At+i",
"Consistency Loss",
"(a) Flawed Atlas Generation (Sec. 3.2)",
"At",
"Phase l:Neural filtering with a flawed atlas",
"Phase2:Localrefinement",
"5",
"川",
"I",
"Ot-1",
"Input video",
"Repeat",
"Repeat",
"0.",
"of",
"(0f T-1",
"Ttimes",
"T-1 times",
"of-,o!",
"FlawedAtlas",
"(b) Neural Filtering and Refinement (Sec. 3.3)",
"Figure 3.The framework of our approach :We first generate an atlas as a unified representation of the whole video",
"providing consistent guidance for deflickering. Since the atlas is flawed, we then propose a neural filtering strategy to flter",
"the flaws.",
"AppearanceDistortion",
"信息是在基于图集的顿At中编码的，因此可以强制执",
"行长期一致性",
"局部细化。视频顿O/T-在短期和长期内都是全局",
"致的。但是，由于输人和基于图集的顿之间的未对",
"准，它可能包含局部闪烁。因此，我们使用一个额外",
"的局部偏转网络来进一步完善结果。先前的工作表明",
"基于流的正则化可以很好地解决局部闪烁问题。因此",
"我们选择了一个经过修改的轻量级管道27。如图E",
"3（b）所示，我们通过将两个连续顿Of,Of-1和先前",
"",
"的输出Ot-1提供给我们的本地细化网络L来预测输",
"出顿Ot。两个连续的顿首先跟随几个卷积层，然后与",
"Ot-1融合。使用简单的时间一致性损失Clocal训练本",
"地闪烁网络，以消除本地闪烁伪像：",
"Clocal(Ot,Ot-1) = |/Mt.t-1 @ (Ot -Ot-1)ll1,(5)",
"Structu",
"re Distorti",
"Figure 4. Training pipeline of our neural filtering strategy.",
"其中Ot-1是通过使用从顿t到顿t-1的光流对",
"We apply two transformations to a clean image X for mim-",
"Ot-1进行扭曲而获得的。Mt,t-1是对应的遮挡遮罩。",
"icking the flickering input frame and the flawed atlas frame.",
"对于没有局部伪影的顿，输出Ot应与Of相同。因此，",
"我们还通过最小化Ot和O之间的距离来提供重建损",
"来自图集框架At，因为它们对结构失真Ts是不变的。",
"失，以正则化质量",
"同时，Ts（X）的失真不会通过网络F传递。通过这种",
"实现细节。网络F是在MS-COCO数据集35]上训",
"策略，我们用有缺陷的图谱实现了神经滤波的目标",
"练的，因为我们只需要图像进行训练。我们为20纪元",
"请注意，虽然此网络F仅接收一个顿，但由于时间",
"训练它，批量大小为8。对于网络L，我们在已处理",
"4"
] |
{
"arxiv_id": "2303.08120",
"language": "zh",
"page_idx": 3
}
|
||
2303.08120-zh-4
|
2303.08120-zh-4.jpg
|
的 DAVIS 数据集$\mathbb{E}[27,43]$ 上对 50 纪元进行训练，批处理大小为 8。我们采用 Adam 优化器，并将学习率设置为0。0001
# 4. Blind Deflickering Dataset
我们构建了第一个公开可用的盲偏转数据集
真实世界数据。我们首先收集包含各种类型的闪烁工件的真实视频。具体来说，我们收集了五种类型的现实世界闪烁的视频：
$\bullet$ Old movies contain complicated flickering pat-
terns. The flicker is caused by multiple reasons, such as unstable exposure time and aging of film materials. Hence, the flickering can be highfrequency or low-frequency, globally and locally.
$\bullet$ Old cartoons are similar to old movies, but the
structures differ from natural videos.
$\bullet$ Time-lapse videos capture a scene for a long time,
and the environment illumination usually changes
a lot.
$\bullet$ Slow-motion videos can capture high-frequency
changes in lighting.
$\bullet$ Processed videos denote the videos processed by
various processing algorithms. The patterns of flicker are decided by the specific algorithm. We follow the setting in [27] .
合成数据。虽然现实世界的视频有利于评估感知性能，但它们没有定量评估的基本事实。因此，我们创建了一个合成数据集，为定量分析提供了基本事实。令$\{G_t\}_{t=1}^T$为干净的视频帧，可以通过在时间 $t$ 为每个帧添加闪烁的伪像$F_t$来获得闪烁的视频$\{G_t\}_t=1^T:$
$$I_t=G_t+F_t,$$
(6)
,其中$\{F_t\}_{t=1}^T$是合成的闪烁伪像。对于时间维度，我们综合了短期和长期闪烁。具体来说，我们设置一个窗口大小$W$ ,它表示共享相同闪烁伪像的帧数。我们将窗口大小$W$ 分别设置为1、3、10
摘要。我们提供 20 , 10 , 10 , 10 , 157 和 90 , 用于老电影，旧卡通，慢动作视频，延时视频，处理视频和合成视频
5. Experiments
# 5.1. Evaluation Setup
数据集。我们主要使用我们构建的盲闪烁数据集进行
评估。详情见 4 节
评估度量。我们基于 DVP$\mathbb{E}[31]$中使用的翘曲误差来测量时间不一致性，该翘曲误差同时考虑了短期和长期翘曲误差以进行定量评估。给定一对帧$O_t$ 和$O_s$, 可以通过以下方式计算翘曲误差$E_{pair}:$
(7)
$$\begin{aligned}&E_{pair}(O_{t},O_{s})=||M_{t,s}\odot(O_{t}-\hat{O}_{s})||_{1},\\&E_{warp}^{t}=E_{pair}(O_{t},O_{t-1})+E_{pair}(O_{t},O_{1}),\end{aligned}$$
(8)
其中$\hat{O}_s$是通过使用从帧$t$到帧$s$的光流对$O_s$进行翘曲而获得的。$M_t,s$是来自帧$t$和帧$s$的遮挡掩码。对于每个帧$t$ ,使用视频中的前一帧和前一帧计算翘曲误差$E_{warp}^t$ 。我们计算视频中所有帧的翘曲误差
## 5.2. Comparisons to Baselines
基线。由于我们的方法是第一种用于盲目变调的方法，因此无法使用现有的公共方法进行比较。因此，我们设计了一个受盲视频时间一致性方法启发的基线： ConvLSTM,修改自 Lai 等人。[27]。具体来说，我们用闪烁的成对帧替换一致的输入帧对，并在其训练数据集$\mathbb{E}[27]$ 上重新训练 ConvLSTM
结果。表$\color{red}3$(a)提供了两种视频的定量结果，我们可以从一致的视频中估计高质量的光流，以计算扭曲误差。我们的结果始终优于基线。在表 3(b)中，我们按照A/B 测试协议 [11] 对 AmazonMechanicalTurk 进行了用户研究，以评估主基线 ConvLSTM 和我们的方法之间的感知偏好。每个用户需要从我们的方法处理的视频中选择一个具有更好感知质量的视频，并且基线。我们使用所有无法获得高质量光流的真实世界视频进行定量评估。总的来说，我们有 20 用户和 50 对比较。我们的方法在所有任务中都明显优于基线。图 5 显示了我们的方法和基线之间的定性比较。我们的方法去除了各种类型的闪烁，我们的结果比基线更符合时间与盲时间一致性 methods 的比较。我们的方法和盲视频时间一致性方法之间的比较是不公平的，因为我们的方法不需要额外的视频，并且基线使用额外的输入视频作为指导。但是，我们仍然提供比较结果以供参考。具体来说，我们使用三个最先进的基线，包括 Bonneel 等人。[7],赖等人。[27]和DVP[31]。表 4显示了我们的方法和基线之间的定量结果。我们的方法在各种处理视频上的扭曲误差低于所有基线，这表明即使没有输入视频引导，我们的方法也更加时间一致。
## 5.3. Ablation Study
我们在 Figure 5 的第三列中介绍了有缺陷的图集。虽然帧之间的时间一致性是完美的，但基于图集的帧包含许多伪影和失真。因此，设计专门的策略对于盲目欺骗至关重要。在这一部分中，我们分别分析了我们两个模块的重要性：(1)神经过滤表示图集生成和神经过滤，因为它们不能被拆分。(2)局部细化表示局部细化模块
神经滤波。我们首先通过去除这部分来分析神经滤波的重要性。我们直接在闪烁的输入视频上应用我们的本地细化模块。如表 5 所示，删除神经滤波模块会显著增加时间不一致性
局部细化。如表 5 所示，去除局部细化网络会使定量性能略有下降。在图 6 中，我们显示了帧中的一些局部区域是时间不一致的。虽然这些区域很小，翘曲误差仅略有减少，但这种局部闪烁确实会损害感知性能，如我们的补充材料所示
5
|
[
[
0.08251953125,
0.0947265625,
0.46728515625,
0.0947265625,
0.46728515625,
0.10479736328125,
0.08251953125,
0.10479736328125
],
[
0.5009765625,
0.091552734375,
0.890625,
0.0921630859375,
0.890625,
0.106689453125,
0.5009765625,
0.1060791015625
],
[
0.476318359375,
0.0985107421875,
0.494384765625,
0.0985107421875,
0.494384765625,
0.10162353515625,
0.476318359375,
0.10162353515625
],
[
0.08416748046875,
0.10986328125,
0.46728515625,
0.10986328125,
0.46728515625,
0.12060546875,
0.08416748046875,
0.12060546875
],
[
0.5048828125,
0.10986328125,
0.8896484375,
0.10986328125,
0.8896484375,
0.12060546875,
0.5048828125,
0.12060546875
],
[
0.08087158203125,
0.12371826171875,
0.19287109375,
0.12371826171875,
0.19287109375,
0.136962890625,
0.08087158203125,
0.136962890625
],
[
0.5048828125,
0.12371826171875,
0.8896484375,
0.12371826171875,
0.8896484375,
0.136962890625,
0.5048828125,
0.136962890625
],
[
0.50341796875,
0.1395263671875,
0.85107421875,
0.1395263671875,
0.85107421875,
0.154052734375,
0.50341796875,
0.154052734375
],
[
0.080078125,
0.149658203125,
0.33251953125,
0.149658203125,
0.33251953125,
0.1641845703125,
0.080078125,
0.1641845703125
],
[
0.1004638671875,
0.1749267578125,
0.41015625,
0.1749267578125,
0.41015625,
0.1881103515625,
0.1004638671875,
0.1881103515625
],
[
0.50341796875,
0.1717529296875,
0.73876953125,
0.1717529296875,
0.73876953125,
0.186279296875,
0.50341796875,
0.186279296875
],
[
0.08251953125,
0.1900634765625,
0.468017578125,
0.1900634765625,
0.468017578125,
0.203857421875,
0.08251953125,
0.203857421875
],
[
0.50634765625,
0.199462890625,
0.88818359375,
0.199462890625,
0.88818359375,
0.210205078125,
0.50634765625,
0.210205078125
],
[
0.08087158203125,
0.205810546875,
0.468017578125,
0.205810546875,
0.468017578125,
0.2197265625,
0.08087158203125,
0.2197265625
],
[
0.50244140625,
0.2108154296875,
0.890625,
0.21337890625,
0.890625,
0.2266845703125,
0.50244140625,
0.2247314453125
],
[
0.08251953125,
0.222900390625,
0.205078125,
0.222900390625,
0.205078125,
0.23291015625,
0.08251953125,
0.23291015625
],
[
0.5048828125,
0.2279052734375,
0.89208984375,
0.2279052734375,
0.89208984375,
0.2412109375,
0.5048828125,
0.2412109375
],
[
0.0931396484375,
0.242431640625,
0.46728515625,
0.242431640625,
0.46728515625,
0.255615234375,
0.0931396484375,
0.255615234375
],
[
0.5048828125,
0.242431640625,
0.89208984375,
0.242431640625,
0.89208984375,
0.255615234375,
0.5048828125,
0.255615234375
],
[
0.111083984375,
0.2568359375,
0.46728515625,
0.2568359375,
0.46728515625,
0.270751953125,
0.111083984375,
0.270751953125
],
[
0.5048828125,
0.25830078125,
0.89208984375,
0.25830078125,
0.89208984375,
0.271484375,
0.5048828125,
0.271484375
],
[
0.11273193359375,
0.272705078125,
0.468017578125,
0.272705078125,
0.468017578125,
0.285888671875,
0.11273193359375,
0.285888671875
],
[
0.50244140625,
0.272705078125,
0.77392578125,
0.271484375,
0.77392578125,
0.285400390625,
0.50244140625,
0.285888671875
],
[
0.11358642578125,
0.287353515625,
0.468017578125,
0.287353515625,
0.468017578125,
0.300537109375,
0.11358642578125,
0.300537109375
],
[
0.50341796875,
0.291748046875,
0.890625,
0.291748046875,
0.890625,
0.304931640625,
0.50341796875,
0.304931640625
],
[
0.11273193359375,
0.300537109375,
0.453369140625,
0.300537109375,
0.453369140625,
0.314453125,
0.11273193359375,
0.314453125
],
[
0.50341796875,
0.30615234375,
0.890625,
0.30615234375,
0.890625,
0.3193359375,
0.50341796875,
0.3193359375
],
[
0.0931396484375,
0.317626953125,
0.468017578125,
0.317626953125,
0.468017578125,
0.330810546875,
0.0931396484375,
0.330810546875
],
[
0.50341796875,
0.3212890625,
0.890625,
0.3212890625,
0.890625,
0.335205078125,
0.50341796875,
0.335205078125
],
[
0.11273193359375,
0.333984375,
0.365234375,
0.333984375,
0.365234375,
0.3447265625,
0.11273193359375,
0.3447265625
],
[
0.50341796875,
0.3359375,
0.890625,
0.3359375,
0.890625,
0.349853515625,
0.50341796875,
0.349853515625
],
[
0.09478759765625,
0.347900390625,
0.46728515625,
0.347900390625,
0.46728515625,
0.361083984375,
0.09478759765625,
0.361083984375
],
[
0.50634765625,
0.3515625,
0.890625,
0.3515625,
0.890625,
0.365478515625,
0.50634765625,
0.365478515625
],
[
0.1094970703125,
0.3623046875,
0.468017578125,
0.363037109375,
0.468017578125,
0.376953125,
0.1094970703125,
0.375732421875
],
[
0.50244140625,
0.365478515625,
0.890625,
0.3662109375,
0.890625,
0.380126953125,
0.50244140625,
0.37890625
],
[
0.11273193359375,
0.37890625,
0.1502685546875,
0.37890625,
0.1502685546875,
0.39013671875,
0.11273193359375,
0.39013671875
],
[
0.50634765625,
0.38330078125,
0.8896484375,
0.38330078125,
0.8896484375,
0.393310546875,
0.50634765625,
0.393310546875
],
[
0.0931396484375,
0.393310546875,
0.46728515625,
0.393310546875,
0.46728515625,
0.4072265625,
0.0931396484375,
0.4072265625
],
[
0.5048828125,
0.396484375,
0.89208984375,
0.396484375,
0.89208984375,
0.410400390625,
0.5048828125,
0.410400390625
],
[
0.111083984375,
0.4072265625,
0.2484130859375,
0.4091796875,
0.2484130859375,
0.42236328125,
0.111083984375,
0.42041015625
],
[
0.50341796875,
0.412353515625,
0.8896484375,
0.412353515625,
0.8896484375,
0.425537109375,
0.50341796875,
0.425537109375
],
[
0.0931396484375,
0.423583984375,
0.46728515625,
0.423583984375,
0.46728515625,
0.4375,
0.0931396484375,
0.4375
],
[
0.50634765625,
0.4287109375,
0.88818359375,
0.4287109375,
0.88818359375,
0.439453125,
0.50634765625,
0.439453125
],
[
0.111083984375,
0.438232421875,
0.4697265625,
0.438232421875,
0.4697265625,
0.451904296875,
0.111083984375,
0.451904296875
],
[
0.5048828125,
0.442626953125,
0.890625,
0.442626953125,
0.890625,
0.455810546875,
0.5048828125,
0.455810546875
],
[
0.11358642578125,
0.453857421875,
0.4697265625,
0.453857421875,
0.4697265625,
0.46728515625,
0.11358642578125,
0.46728515625
],
[
0.5048828125,
0.458984375,
0.87841796875,
0.458984375,
0.87841796875,
0.4697265625,
0.5048828125,
0.4697265625
],
[
0.11273193359375,
0.46728515625,
0.29736328125,
0.468505859375,
0.29736328125,
0.481689453125,
0.11273193359375,
0.48095703125
],
[
0.50341796875,
0.473388671875,
0.890625,
0.474853515625,
0.890625,
0.488037109375,
0.50341796875,
0.4873046875
],
[
0.08416748046875,
0.4912109375,
0.46728515625,
0.4912109375,
0.46728515625,
0.501953125,
0.08416748046875,
0.501953125
],
[
0.50341796875,
0.490478515625,
0.890625,
0.490478515625,
0.890625,
0.50390625,
0.50341796875,
0.50390625
],
[
0.08087158203125,
0.5048828125,
0.468017578125,
0.5048828125,
0.468017578125,
0.51806640625,
0.08087158203125,
0.51806640625
],
[
0.5048828125,
0.5068359375,
0.8896484375,
0.5068359375,
0.8896484375,
0.517578125,
0.5048828125,
0.517578125
],
[
0.08416748046875,
0.51806640625,
0.468017578125,
0.51953125,
0.468017578125,
0.53271484375,
0.08416748046875,
0.5322265625
],
[
0.50341796875,
0.52099609375,
0.890625,
0.52099609375,
0.890625,
0.5341796875,
0.50341796875,
0.5341796875
],
[
0.08251953125,
0.53515625,
0.468017578125,
0.53515625,
0.468017578125,
0.548828125,
0.08251953125,
0.548828125
],
[
0.5048828125,
0.537109375,
0.8896484375,
0.537109375,
0.8896484375,
0.54736328125,
0.5048828125,
0.54736328125
],
[
0.08087158203125,
0.55029296875,
0.428955078125,
0.55029296875,
0.428955078125,
0.564453125,
0.08087158203125,
0.564453125
],
[
0.50341796875,
0.55029296875,
0.8896484375,
0.55029296875,
0.8896484375,
0.564453125,
0.50341796875,
0.564453125
],
[
0.223876953125,
0.57080078125,
0.32275390625,
0.57275390625,
0.32275390625,
0.5869140625,
0.223876953125,
0.58544921875
],
[
0.50732421875,
0.5673828125,
0.8896484375,
0.5673828125,
0.8896484375,
0.57763671875,
0.50732421875,
0.57763671875
],
[
0.44677734375,
0.57373046875,
0.468017578125,
0.57373046875,
0.468017578125,
0.5859375,
0.44677734375,
0.5859375
],
[
0.50341796875,
0.5791015625,
0.8896484375,
0.580078125,
0.8896484375,
0.5947265625,
0.50341796875,
0.59326171875
],
[
0.0931396484375,
0.5966796875,
0.468017578125,
0.5966796875,
0.468017578125,
0.60986328125,
0.0931396484375,
0.60986328125
],
[
0.5048828125,
0.59765625,
0.86669921875,
0.59765625,
0.86669921875,
0.60791015625,
0.5048828125,
0.60791015625
],
[
0.08251953125,
0.6123046875,
0.468017578125,
0.6123046875,
0.468017578125,
0.62548828125,
0.08251953125,
0.62548828125
],
[
0.08251953125,
0.626953125,
0.468017578125,
0.626953125,
0.468017578125,
0.64013671875,
0.08251953125,
0.64013671875
],
[
0.5009765625,
0.62744140625,
0.65673828125,
0.62890625,
0.65673828125,
0.64306640625,
0.5009765625,
0.642578125
],
[
0.08087158203125,
0.642578125,
0.328369140625,
0.642578125,
0.328369140625,
0.65576171875,
0.08087158203125,
0.65576171875
],
[
0.08087158203125,
0.65771484375,
0.468017578125,
0.65771484375,
0.468017578125,
0.671875,
0.08087158203125,
0.671875
],
[
0.52294921875,
0.65478515625,
0.890625,
0.65478515625,
0.890625,
0.66845703125,
0.52294921875,
0.66845703125
],
[
0.08087158203125,
0.67236328125,
0.468017578125,
0.67236328125,
0.468017578125,
0.68603515625,
0.08087158203125,
0.68603515625
],
[
0.50341796875,
0.66943359375,
0.890625,
0.66943359375,
0.890625,
0.68310546875,
0.50341796875,
0.68310546875
],
[
0.08251953125,
0.689453125,
0.1478271484375,
0.689453125,
0.1478271484375,
0.70068359375,
0.08251953125,
0.70068359375
],
[
0.5048828125,
0.6845703125,
0.88623046875,
0.68505859375,
0.88623046875,
0.69873046875,
0.5048828125,
0.69775390625
],
[
0.50341796875,
0.69970703125,
0.88818359375,
0.70068359375,
0.88818359375,
0.7138671875,
0.50341796875,
0.71337890625
],
[
0.07843017578125,
0.7138671875,
0.21728515625,
0.7138671875,
0.21728515625,
0.72900390625,
0.07843017578125,
0.72900390625
],
[
0.5048828125,
0.71533203125,
0.890625,
0.71533203125,
0.890625,
0.72900390625,
0.5048828125,
0.72900390625
],
[
0.5048828125,
0.73095703125,
0.89208984375,
0.73095703125,
0.89208984375,
0.744140625,
0.5048828125,
0.744140625
],
[
0.07843017578125,
0.7373046875,
0.250732421875,
0.7392578125,
0.250732421875,
0.75390625,
0.07843017578125,
0.751953125
],
[
0.5048828125,
0.7470703125,
0.53759765625,
0.7470703125,
0.53759765625,
0.7587890625,
0.5048828125,
0.7587890625
],
[
0.08087158203125,
0.76220703125,
0.468017578125,
0.76220703125,
0.468017578125,
0.77587890625,
0.08087158203125,
0.77587890625
],
[
0.50341796875,
0.76318359375,
0.890625,
0.76318359375,
0.890625,
0.7763671875,
0.50341796875,
0.7763671875
],
[
0.08251953125,
0.77880859375,
0.2197265625,
0.77880859375,
0.2197265625,
0.7890625,
0.08251953125,
0.7890625
],
[
0.5048828125,
0.77880859375,
0.8896484375,
0.77880859375,
0.8896484375,
0.79248046875,
0.5048828125,
0.79248046875
],
[
0.08087158203125,
0.79248046875,
0.468017578125,
0.79248046875,
0.468017578125,
0.80615234375,
0.08087158203125,
0.80615234375
],
[
0.50341796875,
0.79345703125,
0.89208984375,
0.79345703125,
0.89208984375,
0.806640625,
0.50341796875,
0.806640625
],
[
0.08087158203125,
0.80810546875,
0.468017578125,
0.80810546875,
0.468017578125,
0.8212890625,
0.08087158203125,
0.8212890625
],
[
0.50341796875,
0.8095703125,
0.654296875,
0.8095703125,
0.654296875,
0.82275390625,
0.50341796875,
0.82275390625
],
[
0.08251953125,
0.82470703125,
0.4697265625,
0.82470703125,
0.4697265625,
0.83544921875,
0.08251953125,
0.83544921875
],
[
0.5048828125,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.5048828125,
0.84033203125
],
[
0.08416748046875,
0.84033203125,
0.363525390625,
0.84033203125,
0.363525390625,
0.85107421875,
0.08416748046875,
0.85107421875
],
[
0.50341796875,
0.84228515625,
0.890625,
0.84228515625,
0.890625,
0.85595703125,
0.50341796875,
0.85595703125
],
[
0.1258544921875,
0.8623046875,
0.398681640625,
0.861328125,
0.398681640625,
0.87548828125,
0.1258544921875,
0.876953125
],
[
0.4453125,
0.8623046875,
0.468017578125,
0.8623046875,
0.468017578125,
0.87548828125,
0.4453125,
0.87548828125
],
[
0.5048828125,
0.85693359375,
0.89208984375,
0.85693359375,
0.89208984375,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.406005859375,
0.8671875,
0.428955078125,
0.8671875,
0.428955078125,
0.8720703125,
0.406005859375,
0.8720703125
],
[
0.50341796875,
0.87060546875,
0.89208984375,
0.8720703125,
0.89208984375,
0.88623046875,
0.50341796875,
0.88525390625
],
[
0.1258544921875,
0.880859375,
0.4697265625,
0.88330078125,
0.4697265625,
0.89794921875,
0.1258544921875,
0.89599609375
],
[
0.50341796875,
0.88720703125,
0.65185546875,
0.88720703125,
0.65185546875,
0.90087890625,
0.50341796875,
0.90087890625
],
[
0.47802734375,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.93798828125,
0.47802734375,
0.93798828125
]
] |
[
"的DAVIS数据集E27.43上对50纪元进行训练，批",
"其中O。是通过使用从顿t到顿s的光流对O。进行翘",
"",
"处理大小为8。我们采用Adam优化器，并将学习率",
"曲而获得的。Mt.s是来自顿t和顿s的遮挡掩码。对",
"设置为0。0001",
"于每个顿t，使用视频中的前一顿和前一顿计算翘曲",
"误差Ewarp。我们计算视频中所有顿的翘曲误差",
"4.Blind Deflickering Dataset",
"我们构建了第一个公开可用的盲偏转数据集",
"5.2. Comparisons to Baselines",
"真实世界数据。我们首先收集包含各种类型的闪烁工",
"基线。由于我们的方法是第一种用于盲自变调的方",
"件的真实视频。具体来说，我们收集了五种类型的现实",
"法，因此无法使用现有的公共方法进行比较。因此，我",
"世界闪烁的视频：",
"们设计了一个受盲视频时间一致性方法启发的基线：",
": Old movies contain complicated flickering pat-",
"ConvLSTM，修改自Lai等人。[27]。具体来说，我",
"terns. The flicker is caused by multiple reasons,",
"们用闪烁的成对顿替换一致的输入顿对，并在其训练",
"such as unstable exposure time and aging of film",
"数据集E[27]上重新训练ConvLSTM",
"materials. Hence, the flickering can be high-",
"结果。表3（a）提供了两种视频的定量结果，我们可以",
"frequency or low-frequency, globally and locally",
"从一致的视频中估计高质量的光流，以计算扭曲误差。",
". Old cartoons are similar to old movies, but the",
"我们的结果始终优于基线。在表3（b）中，我们按照",
"structures differ from natural video",
"A/B测试协议11]对AmazonMechanicalTurk进行",
". Time-lapse videos capture a scene for a long time,",
"了用户研究，以评估主基线ConvLSTM和我们的方法",
"and the environment illumination usually changes",
"之间的感知偏好。每个用户需要从我们的方法处理的",
"a lot.",
"视频中选择一个具有更好感知质量的视频，并且基线。",
". Slow-motion videos can capture high-frequency",
"我们使用所有无法获得高质量光流的真实世界视频进",
"changes in lighting",
"行定量评估。总的来说，我们有20用户和50对比较。",
". Processed videos denote the videos processed by",
"我们的方法在所有任务中都明显优于基线。图5显示",
"various processing algorithms. The patterns of",
"了我们的方法和基线之间的定性比较。我们的方法去",
"ficker are decided by the specific algorithm. We",
"除了各种类型的闪烁，我们的结果比基线更符合时间",
"follow the setting in [27] .",
"与盲时间一致性methods的比较。我们的方法和盲视",
"合成数据。虽然现实世界的视频有利于评估感知性",
"频时间一致性方法之间的比较是不公平的，因为我们",
"能，但它们没有定量评估的基本事实。因此，我们创建",
"的方法不需要额外的视频，并且基线使用额外的输入视",
"了一个合成数据集，为定量分析提供了基本事实。令",
"频作为指导。但是，我们仍然提供比较结果以供参考。",
"[Gt]=1为干净的视频顿，可以通过在时间t为每个顿",
"具体来说，我们使用三个最先进的基线，包括Bonneel",
"添加闪烁的伪像Ft来获得闪烁的视频[Gt}=1：",
"等人。[7]，赖等人。[27]和DVP[31]。表4显示",
"It=Gt+Ft,",
"了我们的方法和基线之间的定量结果。我们的方法在",
"(6)",
"各种处理视频上的扭曲误差低于所有基线，这表明即",
"其中F}1是合成的闪烁伪像。对于时间维度，我",
"使没有输入视频引导，我们的方法也更加时间一致。",
"们综合了短期和长期闪烁。具体来说，我们设置一个窗",
"口大小W，它表示共享相同闪烁伪像的顿数。我们将",
"5.3. Ablation Study",
"窗口大小W分别设置为1、3、10",
"摘要。我们提供20，10，10，10，157和90，用于",
"我们在Figure5的第三列中介绍了有缺陷的图集。",
"老电影，旧卡通，慢动作视频，延时视频，处理视频和",
"虽然顿之间的时间一致性是完美的，但基于图集的顿",
"合成视频",
"包含许多伪影和失真。因此，设计专门的策略对于盲目",
"欺骗至关重要。在这一部分中，我们分别分析了我们两",
"5.Experiments",
"个模块的重要性：（1）神经过滤表示图集生成和神经过",
"滤，因为它们不能被拆分。（2）局部细化表示局部细化",
"5.1. Evaluation Setup",
"模块",
"数据集。我们主要使用我们构建的盲闪烁数据集进行",
"神经滤波。我们首先通过去除这部分来分析神经滤波",
"评估。详情见4节",
"的重要性。我们直接在闪烁的输入视频上应用我们的",
"评估度量。我们基于DVPE[31]中使用的翘曲误差来",
"本地细化模块。如表5所示，删除神经滤波模块会显",
"测量时间不一致性，该翘曲误差同时考虑了短期和长",
"著增加时间不一致性",
"期翘曲误差以进行定量评估。给定一对顿Ot和Os，",
"局部细化。如表5所示，去除局部细化网络会使定量",
"可以通过以下方式计算翘曲误差Epair：",
"性能略有下降。在图6中，我们显示了顿中的一些局",
"Epair(Ot, Os) = IMt,s O(Ot -O.)l1,",
"(7)",
"部区域是时间不一致的。虽然这些区域很小，翘曲误差",
"",
"仅略有减少，但这种局部闪烁确实会损害感知性能，如",
"Ewarp=Epair(Ot,Ot-1)+Epair(Ot,O1)，(8)",
"我们的补充材料所示",
"5"
] |
{
"arxiv_id": "2303.08120",
"language": "zh",
"page_idx": 4
}
|
||
2306.06024-zh-1
|
2306.06024-zh-1.jpg
|
Title Suppressed Due to Excessive Size
while keeping $u$ unchanged.
在本文中，我们提出了一种可自我解释的时间序列预测模型，称为反事实时间序列 (CountfactualTimeSeries, Counts),该模型既可以(1)进行时间序列预测，又可以(2)为其预测提供可操作和可行的反事实解释。在常见的因果结构假设下，我们的方法可以保证在存在外生(混杂)变量的情况下识别输入和输出之间的因果效应，从而提高生成的反事实解释的可行性。我们的贡献总结如下：
$\bullet$ We identify the actionability and feasibility re-
quirements for generating counterfactual explanations for time series models and develop the first general self-interpretable method, dubbed CounTS, that satisfes such requirements.
$\bullet$ We provide theoretical guarantees that CounTS
can identify the causal effect between the time series input and output in the presence of exogenous (confounding) variables, thereby improving feasibility in the generated explanations.
$\bullet$ Experiments on both synthetic and real-world
datasets show that compared to state-of-the-art methods, CounTS significantly improves performance for generating counterfactual explanations while still maintaining comparable prediction ac- $\operatorname{curacy}.$
## 2. Related Work
神经网络的解释方法。近年来提出了各种基于归因的解释方法。一些方法专注于局部解释 (Ribeiro et al., 2016; Lundberg & Lee, 2017b; Plumb et al., 2018; Chen et al.,2018; Wang et al.,2019b) ,而另一些方法则设计用于全局解释 (Ghorbani et al., 2019; Natesan Ramamurthy et al.,2020)。主要思想是根据输人特征对预测(输出)的影响将归因或重要性分数分配给输入特征。例如，可以使用相对于输人(Selvaraju et al., 2017; Lundberg & Lee, 2017a; Shrikumar et al., 2017; Sundararajan et al., 2017) 的预测梯度来计算这种重要性分数。一些解释方法专门用于时间序列数据； 这些方法包括基于扰动的 (Pan et al.,2021) ,基于规则的 (Rajapaksha & Bergmeir, 2022) 和基于注意力的方法 (Heo et al., 2018; Lim et al., 2021) 。一种典型的方法，即时间特征重要性(FIT),基于时间分布偏移和无法解释的分布偏移 (Tonekaboni et al.,2020) 评估输入数据的重要性。但是，这些方法只能产生当前预测的输入特征的重要性得分，因此无法生成反事实解释 (有关经验结果，请参见E$Sec. 5 和 ▣ EAppendix D )$ 时间序列模型的反事实解释。也有为时间序列模型生成反事实解释的作品。(Dhaou et al.,2021) 提出了一种关联规则算法，通过查找频繁的时间戳对并生成反事实示例来解释时间序列预测。(Nemirovsky et al.,
2022) 提出了一个通用的解释框架，该框架使用残差生成对抗网络(RGAN)生成反事实示例；它可以适用于时间序列模型。但是，这些工作要么无法生成现实的反事实解释(由于离散化错误),要么无法为时间序列模型生成可行的反事实解释。相反，我们的 CounTS 作为原则性变分因果方法E(Wang et al., 2020; Mao et al., 2021b; Gupta et al., 2021) 可以自然地产生现实可行的反事实解释。这种优势在 Sec.5 中得到了实证验证
贝叶斯深度学习和变分自编码器。我们的工作还与变分自动编码器(VAEs)$\operatorname{E}($Kingma&Welling,2013) (使用推理网络来近似后验分布)和贝叶斯深度学习(BDL)$\boxtimes$E(Wang et al., 2015; Wang &Yeung, 2016; Wang, 2017; Huang et al., 2019; Wang et al., 2019a; Wang & Yeung, 2020; Ding et al., 2022) 模型(使用深层组件来处理高维信号和特定于任务/图形组件来处理条件/因果依赖性)的广泛类别有关。(Lin et al., 2022) 提出了第一个基于 VAE 的模型，用于生成图神经网络的因果解释。(Louizos et al., 2017; Pawlowski et al., 2020) 提出了第一个基于 VAE 的模型，用于执行因果推断和估计治疗效果。但是，它们都没有解决反事实解释的问题，该问题涉及解决反问题以获得最佳的反事实输入。相比之下，我们的 CounTS 是第一个基于 VAE 的模型来应对这一挑战，具有理论保证和有希望的实证结果。从 BDL (Wang $\&$ Yeung, 2016; 2020) 的角度来看，CounTS 使用深度神经网络处理高维信号( $i。e。, ( Wang \& Yeung, 2016) 中 的 深 层 组$ 件),并使用贝叶斯网络处理变量之间的条件/因果依赖关系(i。e。, (Wang &Yeung,2016)中的任务特定或图形组件)。因此，CounTS 也是第一个生成反事实解释的 BDL 模型
## 3. Preliminaries
因果模型。根据(Pearl,2009)中的定义，因果模型由 3 元组$M=\langle U,V,F\rangle$描述。$U$是一组外生变量$\{ u_1, \ldots , u_m\}$ , 它不是由该因果模型中的任何其他变量决定的。$V$是由$U\cup V$中的变量决定的一组内生变量$\{v_1,\ldots,v_n\}$ 。我们假设因果模型可以根据有向图分解，其中每个节点代表一个变量。$F$是一组描述$V$ 生成过程的函数$\{f_1,\ldots,f_n\}:$
$$v_i=f_i(pa_i,u_i),i=1,\ldots,n,$$
,其中$pa_i$表示$v_i$的直接父节点
反事实推理。反事实推理对诸如“观察$X=x$和$Y=y$ ,如果输入$X$是$x^cf,Y=y^{cf}$的概率是多少？”。形式上，给定一个因果模型$\left\langle U,V,F\right\rangle$,其中$Y, X\in V$ , 反事实推理分三个步骤进行$\operatorname{E}($Pearl, 2009) :
1. Abduction. Calculate the posterior distribution
of$u$ given the observation $X=x$ and $Y=y$ , i.e.,
$P( u| X= x, Y= y)$ .
2. Action. Perform causal intervention on the vari-
able $X$ , i.e., $do( X= x^cf)$ .
2
|
[
[
0.359375,
0.059356689453125,
0.61376953125,
0.059356689453125,
0.61376953125,
0.0694580078125,
0.359375,
0.0694580078125
],
[
0.0882568359375,
0.085205078125,
0.283447265625,
0.08587646484375,
0.283447265625,
0.10101318359375,
0.0882568359375,
0.0997314453125
],
[
0.5009765625,
0.08587646484375,
0.88232421875,
0.08587646484375,
0.88232421875,
0.0997314453125,
0.5009765625,
0.0997314453125
],
[
0.50244140625,
0.10162353515625,
0.88232421875,
0.10162353515625,
0.88232421875,
0.11553955078125,
0.50244140625,
0.11553955078125
],
[
0.0906982421875,
0.11297607421875,
0.47216796875,
0.11297607421875,
0.47216796875,
0.126953125,
0.0906982421875,
0.126953125
],
[
0.50244140625,
0.117431640625,
0.8837890625,
0.117431640625,
0.8837890625,
0.1307373046875,
0.50244140625,
0.1307373046875
],
[
0.09234619140625,
0.1300048828125,
0.47216796875,
0.1300048828125,
0.47216796875,
0.14013671875,
0.09234619140625,
0.14013671875
],
[
0.50341796875,
0.1337890625,
0.88232421875,
0.1337890625,
0.88232421875,
0.14453125,
0.50341796875,
0.14453125
],
[
0.08905029296875,
0.1427001953125,
0.473876953125,
0.1427001953125,
0.473876953125,
0.1558837890625,
0.08905029296875,
0.1558837890625
],
[
0.5009765625,
0.147705078125,
0.8857421875,
0.147705078125,
0.8857421875,
0.1610107421875,
0.5009765625,
0.1610107421875
],
[
0.09234619140625,
0.1578369140625,
0.473876953125,
0.1578369140625,
0.473876953125,
0.1717529296875,
0.09234619140625,
0.1717529296875
],
[
0.5009765625,
0.1622314453125,
0.88232421875,
0.1610107421875,
0.88232421875,
0.1749267578125,
0.5009765625,
0.175537109375
],
[
0.0906982421875,
0.1749267578125,
0.471435546875,
0.1749267578125,
0.471435546875,
0.1849365234375,
0.0906982421875,
0.1849365234375
],
[
0.5009765625,
0.177978515625,
0.8857421875,
0.177978515625,
0.8857421875,
0.1912841796875,
0.5009765625,
0.1912841796875
],
[
0.0906982421875,
0.1881103515625,
0.47216796875,
0.1881103515625,
0.47216796875,
0.2020263671875,
0.0906982421875,
0.2020263671875
],
[
0.50341796875,
0.1925048828125,
0.8857421875,
0.1925048828125,
0.8857421875,
0.205810546875,
0.50341796875,
0.205810546875
],
[
0.0906982421875,
0.203857421875,
0.47216796875,
0.203857421875,
0.47216796875,
0.2171630859375,
0.0906982421875,
0.2171630859375
],
[
0.50244140625,
0.20703125,
0.53515625,
0.208984375,
0.53369140625,
0.2215576171875,
0.5009765625,
0.2197265625
],
[
0.08905029296875,
0.2183837890625,
0.27783203125,
0.2183837890625,
0.27783203125,
0.231689453125,
0.08905029296875,
0.231689453125
],
[
0.50341796875,
0.2298583984375,
0.8837890625,
0.2298583984375,
0.8837890625,
0.24365234375,
0.50341796875,
0.24365234375
],
[
0.101318359375,
0.248779296875,
0.468017578125,
0.248779296875,
0.468017578125,
0.261962890625,
0.101318359375,
0.261962890625
],
[
0.50244140625,
0.24560546875,
0.8837890625,
0.24560546875,
0.8837890625,
0.2587890625,
0.50244140625,
0.2587890625
],
[
0.12091064453125,
0.26318359375,
0.473876953125,
0.26318359375,
0.473876953125,
0.27783203125,
0.12091064453125,
0.27783203125
],
[
0.5,
0.260009765625,
0.88134765625,
0.260009765625,
0.88134765625,
0.2734375,
0.5,
0.2734375
],
[
0.1217041015625,
0.27783203125,
0.47216796875,
0.27783203125,
0.47216796875,
0.291748046875,
0.1217041015625,
0.291748046875
],
[
0.50244140625,
0.27587890625,
0.8857421875,
0.27587890625,
0.8857421875,
0.2890625,
0.50244140625,
0.2890625
],
[
0.12091064453125,
0.29345703125,
0.473876953125,
0.292236328125,
0.473876953125,
0.306884765625,
0.12091064453125,
0.30810546875
],
[
0.5009765625,
0.290283203125,
0.88232421875,
0.290283203125,
0.88232421875,
0.3037109375,
0.5009765625,
0.3037109375
],
[
0.1217041015625,
0.309326171875,
0.419189453125,
0.309326171875,
0.419189453125,
0.322509765625,
0.1217041015625,
0.322509765625
],
[
0.50244140625,
0.30615234375,
0.8837890625,
0.30615234375,
0.8837890625,
0.3193359375,
0.50244140625,
0.3193359375
],
[
0.50244140625,
0.32080078125,
0.8857421875,
0.32080078125,
0.8857421875,
0.333984375,
0.50244140625,
0.333984375
],
[
0.10296630859375,
0.33203125,
0.473876953125,
0.33203125,
0.473876953125,
0.34521484375,
0.10296630859375,
0.34521484375
],
[
0.50244140625,
0.3359375,
0.88232421875,
0.3359375,
0.88232421875,
0.349853515625,
0.50244140625,
0.349853515625
],
[
0.12091064453125,
0.347900390625,
0.473876953125,
0.347900390625,
0.473876953125,
0.361083984375,
0.12091064453125,
0.361083984375
],
[
0.5009765625,
0.349853515625,
0.88232421875,
0.349853515625,
0.88232421875,
0.363037109375,
0.5009765625,
0.363037109375
],
[
0.12091064453125,
0.361083984375,
0.473876953125,
0.363037109375,
0.473876953125,
0.376953125,
0.12091064453125,
0.375
],
[
0.50244140625,
0.367431640625,
0.8837890625,
0.367431640625,
0.8837890625,
0.37744140625,
0.50244140625,
0.37744140625
],
[
0.1217041015625,
0.376953125,
0.47216796875,
0.376953125,
0.47216796875,
0.39013671875,
0.1217041015625,
0.39013671875
],
[
0.50244140625,
0.380615234375,
0.8857421875,
0.380615234375,
0.8857421875,
0.39453125,
0.50244140625,
0.39453125
],
[
0.12091064453125,
0.393310546875,
0.410888671875,
0.393310546875,
0.410888671875,
0.4072265625,
0.12091064453125,
0.4072265625
],
[
0.5009765625,
0.39453125,
0.8837890625,
0.395751953125,
0.8837890625,
0.410400390625,
0.5009765625,
0.4091796875
],
[
0.50244140625,
0.410400390625,
0.8837890625,
0.410888671875,
0.8837890625,
0.4248046875,
0.50244140625,
0.423583984375
],
[
0.10296630859375,
0.416748046875,
0.47216796875,
0.415283203125,
0.47216796875,
0.429931640625,
0.10296630859375,
0.43115234375
],
[
0.1192626953125,
0.43115234375,
0.47216796875,
0.431884765625,
0.47216796875,
0.44580078125,
0.1192626953125,
0.4443359375
],
[
0.50244140625,
0.4267578125,
0.88134765625,
0.4267578125,
0.88134765625,
0.43994140625,
0.50244140625,
0.43994140625
],
[
0.5009765625,
0.44140625,
0.8857421875,
0.44140625,
0.8857421875,
0.455078125,
0.5009765625,
0.455078125
],
[
0.1217041015625,
0.447509765625,
0.473876953125,
0.447509765625,
0.473876953125,
0.4609375,
0.1217041015625,
0.4609375
],
[
0.5,
0.455810546875,
0.8857421875,
0.45703125,
0.8857421875,
0.47021484375,
0.5,
0.4697265625
],
[
0.12335205078125,
0.462158203125,
0.47216796875,
0.462158203125,
0.47216796875,
0.47607421875,
0.12335205078125,
0.47607421875
],
[
0.12091064453125,
0.47607421875,
0.47216796875,
0.477783203125,
0.47216796875,
0.4912109375,
0.12091064453125,
0.4892578125
],
[
0.5,
0.47021484375,
0.8857421875,
0.4716796875,
0.8857421875,
0.486083984375,
0.5,
0.48486328125
],
[
0.5009765625,
0.48486328125,
0.61669921875,
0.488037109375,
0.61669921875,
0.50244140625,
0.5009765625,
0.5
],
[
0.625,
0.48486328125,
0.8837890625,
0.4873046875,
0.8837890625,
0.50048828125,
0.625,
0.498779296875
],
[
0.12091064453125,
0.4912109375,
0.1732177734375,
0.49365234375,
0.17236328125,
0.50830078125,
0.1192626953125,
0.505859375
],
[
0.62353515625,
0.4912109375,
0.6298828125,
0.4912109375,
0.6298828125,
0.49560546875,
0.62353515625,
0.49560546875
],
[
0.50244140625,
0.501953125,
0.8837890625,
0.501953125,
0.8837890625,
0.51513671875,
0.50244140625,
0.51513671875
],
[
0.50244140625,
0.517578125,
0.611328125,
0.51806640625,
0.611328125,
0.52978515625,
0.50244140625,
0.52880859375
],
[
0.611328125,
0.51513671875,
0.8837890625,
0.517578125,
0.8837890625,
0.53076171875,
0.611328125,
0.52880859375
],
[
0.08660888671875,
0.5263671875,
0.233642578125,
0.5263671875,
0.233642578125,
0.54052734375,
0.08660888671875,
0.54052734375
],
[
0.50244140625,
0.5322265625,
0.8837890625,
0.5322265625,
0.8837890625,
0.54541015625,
0.50244140625,
0.54541015625
],
[
0.5009765625,
0.546875,
0.646484375,
0.546875,
0.646484375,
0.56103515625,
0.5009765625,
0.56103515625
],
[
0.0906982421875,
0.55322265625,
0.47216796875,
0.55322265625,
0.47216796875,
0.56640625,
0.0906982421875,
0.56640625
],
[
0.0906982421875,
0.5673828125,
0.468017578125,
0.5673828125,
0.468017578125,
0.58056640625,
0.0906982421875,
0.58056640625
],
[
0.0906982421875,
0.58349609375,
0.47216796875,
0.58349609375,
0.47216796875,
0.5966796875,
0.0906982421875,
0.5966796875
],
[
0.498291015625,
0.58056640625,
0.6396484375,
0.58203125,
0.6396484375,
0.5966796875,
0.498291015625,
0.59521484375
],
[
0.08905029296875,
0.5966796875,
0.47216796875,
0.59765625,
0.47216796875,
0.6123046875,
0.08905029296875,
0.611328125
],
[
0.50244140625,
0.60693359375,
0.8837890625,
0.60693359375,
0.8837890625,
0.62060546875,
0.50244140625,
0.62060546875
],
[
0.08905029296875,
0.61279296875,
0.473876953125,
0.61279296875,
0.473876953125,
0.626953125,
0.08905029296875,
0.626953125
],
[
0.0906982421875,
0.62744140625,
0.47216796875,
0.62744140625,
0.47216796875,
0.6416015625,
0.0906982421875,
0.6416015625
],
[
0.5009765625,
0.62255859375,
0.8857421875,
0.62255859375,
0.8857421875,
0.63623046875,
0.5009765625,
0.63623046875
],
[
0.5,
0.63720703125,
0.8857421875,
0.63623046875,
0.8857421875,
0.65087890625,
0.5,
0.6513671875
],
[
0.08905029296875,
0.64306640625,
0.473876953125,
0.64306640625,
0.473876953125,
0.6572265625,
0.08905029296875,
0.6572265625
],
[
0.5009765625,
0.6513671875,
0.88232421875,
0.6513671875,
0.88232421875,
0.66552734375,
0.5009765625,
0.66552734375
],
[
0.08905029296875,
0.6591796875,
0.47216796875,
0.6591796875,
0.47216796875,
0.67236328125,
0.08905029296875,
0.67236328125
],
[
0.5009765625,
0.66845703125,
0.88232421875,
0.66845703125,
0.88232421875,
0.681640625,
0.5009765625,
0.681640625
],
[
0.0906982421875,
0.6748046875,
0.47216796875,
0.6748046875,
0.47216796875,
0.68505859375,
0.0906982421875,
0.68505859375
],
[
0.08905029296875,
0.68798828125,
0.473876953125,
0.68798828125,
0.473876953125,
0.7021484375,
0.08905029296875,
0.7021484375
],
[
0.5009765625,
0.68310546875,
0.88134765625,
0.68310546875,
0.88134765625,
0.6962890625,
0.5009765625,
0.6962890625
],
[
0.08905029296875,
0.7021484375,
0.473876953125,
0.70263671875,
0.473876953125,
0.71728515625,
0.08905029296875,
0.71630859375
],
[
0.5009765625,
0.69775390625,
0.72216796875,
0.69970703125,
0.72216796875,
0.71337890625,
0.5009765625,
0.7109375
],
[
0.0906982421875,
0.71826171875,
0.473876953125,
0.71826171875,
0.473876953125,
0.73193359375,
0.0906982421875,
0.73193359375
],
[
0.59228515625,
0.7197265625,
0.79150390625,
0.72021484375,
0.79150390625,
0.73486328125,
0.59228515625,
0.734375
],
[
0.09234619140625,
0.734375,
0.47216796875,
0.734375,
0.47216796875,
0.74755859375,
0.09234619140625,
0.74755859375
],
[
0.5009765625,
0.74072265625,
0.73876953125,
0.7412109375,
0.73876953125,
0.755859375,
0.5009765625,
0.7548828125
],
[
0.0906982421875,
0.74853515625,
0.47216796875,
0.74853515625,
0.47216796875,
0.76220703125,
0.0906982421875,
0.76220703125
],
[
0.0906982421875,
0.7646484375,
0.473876953125,
0.7646484375,
0.473876953125,
0.77783203125,
0.0906982421875,
0.77783203125
],
[
0.50244140625,
0.7646484375,
0.8857421875,
0.7646484375,
0.8857421875,
0.77783203125,
0.50244140625,
0.77783203125
],
[
0.0906982421875,
0.77880859375,
0.473876953125,
0.77880859375,
0.473876953125,
0.79248046875,
0.0906982421875,
0.79248046875
],
[
0.5087890625,
0.7841796875,
0.515625,
0.7841796875,
0.515625,
0.78857421875,
0.5087890625,
0.78857421875
],
[
0.51123046875,
0.77880859375,
0.8857421875,
0.77880859375,
0.8857421875,
0.79248046875,
0.51123046875,
0.79248046875
],
[
0.0906982421875,
0.79443359375,
0.47216796875,
0.79443359375,
0.47216796875,
0.80810546875,
0.0906982421875,
0.80810546875
],
[
0.50244140625,
0.79443359375,
0.8857421875,
0.79443359375,
0.8857421875,
0.80810546875,
0.50244140625,
0.80810546875
],
[
0.0906982421875,
0.81005859375,
0.47216796875,
0.81005859375,
0.47216796875,
0.82373046875,
0.0906982421875,
0.82373046875
],
[
0.50244140625,
0.8095703125,
0.8203125,
0.8095703125,
0.8203125,
0.82275390625,
0.50244140625,
0.82275390625
],
[
0.0906982421875,
0.82470703125,
0.473876953125,
0.82470703125,
0.473876953125,
0.83837890625,
0.0906982421875,
0.83837890625
],
[
0.5087890625,
0.830078125,
0.8837890625,
0.83056640625,
0.8837890625,
0.84521484375,
0.5087890625,
0.8447265625
],
[
0.0906982421875,
0.8466796875,
0.473876953125,
0.8466796875,
0.473876953125,
0.85986328125,
0.0906982421875,
0.85986328125
],
[
0.53369140625,
0.8466796875,
0.8857421875,
0.8466796875,
0.8857421875,
0.861328125,
0.53369140625,
0.861328125
],
[
0.0906982421875,
0.8623046875,
0.461669921875,
0.8623046875,
0.461669921875,
0.87548828125,
0.0906982421875,
0.87548828125
],
[
0.53369140625,
0.861328125,
0.6845703125,
0.8623046875,
0.6845703125,
0.876953125,
0.53369140625,
0.87548828125
],
[
0.0906982421875,
0.876953125,
0.47216796875,
0.876953125,
0.47216796875,
0.89013671875,
0.0906982421875,
0.89013671875
],
[
0.5107421875,
0.876953125,
0.88623046875,
0.876953125,
0.88623046875,
0.89013671875,
0.5107421875,
0.89013671875
],
[
0.0906982421875,
0.892578125,
0.47216796875,
0.892578125,
0.47216796875,
0.90576171875,
0.0906982421875,
0.90576171875
],
[
0.53369140625,
0.892578125,
0.73193359375,
0.892578125,
0.73193359375,
0.90576171875,
0.53369140625,
0.90576171875
],
[
0.48193359375,
0.923828125,
0.494384765625,
0.923828125,
0.494384765625,
0.93603515625,
0.48193359375,
0.93603515625
]
] |
[
"Title Suppressed Due to ExcessiveSize",
"while keeping u unchanged",
"2022）提出了一个通用的解释框架，该框架使用残差生",
"成对抗网络（RGAN）生成反事实示例；它可以适用于",
"在本文中，我们提出了一种可自我解释的时间序列",
"时间序列模型。但是，这些工作要么无法生成现实的",
"预测模型，称为反事实时间序列（CountfactualTime",
"反事实解释（由于离散化错误），要么无法为时间序列",
"Series，Counts），该模型既可以（1）进行时间序列预",
"模型生成可行的反事实解释。相反，我们的CounTS",
"测，又可以（2）为其预测提供可操作和可行的反事实",
"作为原则性变分因果方法E（Wangetal.，2020Mac",
"解释。在常见的因果结构假设下，我们的方法可以保",
"etal.，2021b;Guptaetal.,2021）可以自然地产生现实",
"证在存在外生（混杂）变量的情况下识别输入和输出",
"可行的反事实解释。这种优势在Sec.5中得到了实证",
"之间的因果效应，从而提高生成的反事实解释的可行",
"验证",
"性。我们的贡献总结如下：",
"贝叶斯深度学习和变分自编码器。我们的工作还与变",
": We identify the actionability and feasibility re",
"分自动编码器（VAEs）E（Kingma&Welling，2013)",
"quirements for generating counterfactual expla-",
"（使用推理网络来近似后验分布）和贝叶斯深度学习",
"nations for time series models and develop the",
"(BDL)E(Wang etal.,2015;Wang &Yeung,2016;",
"first general self-interpretable method, dubbed",
"Wang, 2017; Huang et al., 2019; Wang et al., 2019a.",
"CounTS, that satisfies such requirements.",
"Wang&Yeung，2020;Dingetal.，2022）模型（使用",
"深层组件来处理高维信号和特定于任务/图形组件来",
". We provide theoretical guarantees that CounTS",
"处理条件/因果依赖性）的广泛类别有关。。（Linetal.",
"can identify the causal effect between the time",
"2022）提出了第一个基于VAE的模型，用于生成图神",
"series input and output in the presence of exoge-",
"经网络的因果解释。（Louizosetal.，2017：Pawlowski",
"nous (confounding) variables, thereby improving",
"etal.，2020）提出了第一个基于VAE的模型，用于执",
"feasibility in the generated explanations.",
"行因果推断和估计治疗效果。但是，它们都没有解决",
"反事实解释的问题，该问题涉及解决反问题以获得最",
": Experiments on both synthetic and real-world",
"datasets show that compared to state-of-the-art",
"佳的反事实输人。相比之下，我们的CounTS是第一",
"个基于VAE的模型来应对这一挑战，具有理论保证和",
"methods, CounTS significantly improves perfor-",
"有希望的实证结果。从BDL（Wang&Yeung，2016;",
"mance for generating counterfactual explanations",
"while still maintaining comparable prediction ac",
"2020）的角度来看，CounTS使用深度神经网络处理高",
"维信号(i。e。，",
"（Wang&Yeung，2016）中的深层组",
"curacy.",
"",
"件），并使用贝叶斯网络处理变量之间的条件/因果依",
"赖关系i。e。，",
"（Wang&Yeung，2016）中的任务特",
"2. Related Work",
"定或图形组件)。因此，CounTS也是第一个生成反事",
"实解释的BDL模型",
"神经网络的解释方法。近年来提出了各种基于归因的",
"解释方法。一些方法专注于局部解释（Ribeiroetal.",
"2016;Lundberg & Lee, 2017b;Plumb et al.,2018;",
"3.Preliminaries",
"Chenetal.，2018;Wangetal.,2019b），而另一些方",
"因果模型。根据（Pearl，2009）中的定义，因果模型",
"法则设计用于全局解释（Ghorbanietal.，2019；Nate-",
"sanRamamurthyetal.，2020）。主要思想是根据输人",
"由3元组M=(U,V,F）描述。U是一组外生变量",
"【ui,...,um}，它不是由该因果模型中的任何其他变",
"特征对预测（输出）的影响将归因或重要性分数分配",
"量决定的。V是由UUV中的变量决定的一组内生变",
"给输人特征。例如，可以使用相对于输人（Selvaraju",
"量u1，...，Un}。我们假设因果模型可以根据有向图",
"et al..2017:Lundberg & Lee,2017a:Shrikumar et al..",
"2017；Sundararajanetal.，2017）的预测梯度来计算这",
"分解，其中每个节点代表一个变量。F是一组描述V",
"种重要性分数。一些解释方法专门用于时间序列数据；",
"生成过程的函数（f1，...，fn】：",
"这些方法包括基于扰动的（Panetal.,2021），基于规",
"Vi=fi(pai,ui),i=l,...,n,",
"则的（Rajapaksha&Bergmeir，2022）和基于注意力的",
"，其中pai表示v的直接父节点",
"方法（Heoetal.,2018;Limetal.,2021）。一种典型的",
"方法，即时间特征重要性（FIT），基于时间分布偏移",
"反事实推理。反事实推理对诸如“观察X=和Y=y",
"和无法解释的分布偏移（Tonekabonietal.，2020）评估",
"",
"如果输人X是rcf，Y=ycf的概率是多少?”。形",
"输入数据的重要性。但是，这些方法只能产生当前预",
"式上，给定一个因果模型(U,V,F)，其中Y,XEV，",
"测的输人特征的重要性得分，因此无法生成反事实解",
"反事实推理分三个步骤进行E(Pearl，2009）：",
"释（有关经验结果，请参见ESec.5和EAppendixD）",
"1. Abduction. Calculate the posterior distribution",
"时间序列模型的反事实解释。也有为时间序列模型生",
"of u given the observation X = r and Y = y , i.e.,",
"成反事实解释的作品。（Dhaouetal.，2021）提出了",
"P(uX =r,Y =y):",
"种关联规则算法，通过查找频繁的时间戳对并生成反",
"2. Action. Perform causal intervention on the vari-",
"事实示例来解释时间序列预测。（Nemirovsky etal.",
"able X , i.e., do(X = rcf) :",
"2"
] |
{
"arxiv_id": "2306.06024",
"language": "zh",
"page_idx": 1
}
|
||
2306.06024-zh-3
|
2306.06024-zh-3.jpg
|
Title Suppressed Due to Excessive Size
<FigureHere>
Figure 2. Network structure. We omit subscripts of $p_\theta$ and $q_\phi$ for clarity. Left: CounTS makes predictions on y using the yellow branch. Right: Given the current x and y$^{pre\hat{d}}$, CounTS generates counterfactual explanation x$^cf$ for any target $\mathrm{label~y}^{cf}$ using the blue branch.
$4.2.{\mathrm{~Inference~Using~CounTS}}$
在通过最大化EEqn. 6中的$\mathcal{L}_CounTS$来学习生成模型(Eqn.??)和推理模型(Eqn.??)之后，我们的可自我解释的 CounTS 可以对(1)使用EXMTHX\_7 中的黄色分支对 Fig.2 进行预测(左),以及(2)使用任何 XMATHX
$4.2.1.{\mathrm{~Prediction}}$
于$\boxed{\mathrm{E}}$Defnition 4.1 中反事实解释的定义涉及对 x 进行干预的因果推断，因此我们需要首先使用观测概率来识别因果概率$p($y= $y^{cf}$| do( x= $x^{\prime }$) , u) , 即 。e。, 删 除等式中的'do'。下面的定理表明，这是可以实现的。Theorem 4.1( Identifiability ). Given the posterior $\operatorname*{distribution}_{\text{of exogenous variable}}p(\underset{c}{\operatorname*{u}},\underset{c}{\operatorname*{u}}|$x,y$),\operatorname*{the}$ effect of action $p($y = $y^{cf}$| do( x = $x^{\prime }$) , $u_{l}$, $u_{g}$) can be identified using $\mathbb{E}_{p(\mathrm{z}|\mathrm{x}^{\prime},\mathrm{u}_{l},\mathrm{u}_{g})}[p\left(\mathrm{y}^{cf}|\mathrm{z},\mathrm{u}_{l},\mathrm{u}_{g}\right)].$
有关证明，请参见 Appendix B.1 。使用 Theorem 4.1
,我们可以将 Eqn. 5 重写为
$$\mathcal{L}_{cf}=\mathbb{E}_{p(\mathrm{u|x=x,y=y}^{pred})}\mathbb{E}_{p(\mathrm{z|x',u)}}[p(\mathrm{y}^{cf}|\mathrm{z,u})],$$
(6)
,其中 u$=(\mathrm u_l,\mathrm u_g)$和$p($u$| x=$x,y$= \mathrm{y} ^pred)$由$q_\phi ($u$| x=$ x,y $= y^{pred}$) 近似。我们使用蒙特卡洛估计来计算Eqn. 5 和 Eqn. 6 中 的 期 望 值 , 迭 代 计 算 梯 度 $\frac {\partial \mathcal{L} _{cf}}{\partial x^{\prime }}$ $(\sqrt{3}+\sqrt{4}+\sqrt{4}+\sqrt{4}+\sqrt{4})$ (通过反向传播)以类似于|F|(Wang et al..2019a; Mac et al., 2021a) 的方式搜索最佳 x',并将其用作 x$^cf$ (请参阅 Appendix B.2 中的完整算法)
我们使用$\overline{\mathrm{EFig.2}}$(左)中的黄色分支来预测 y。具体来说，给定一个输入时间序列 x , 编码器将首先使用$q_\phi($y|x)推断初始标签 y ,然后使用$q_\phi($u$_l$,u$_g$,z|x,y) 从 x 和 y 推断$(\mathrm{u}_l,\mathrm{u}_g)$和 z。然后将$(\mathrm{u}_l,\mathrm{u}_g)$和 z 馈送到预测器$p_\theta($y|u$_l$,u$_g$,z)中，以推断最终标签 y。从形式上讲，CounTS 将标签预测为
(4)
$$\mathrm{y}^{pred}=\mathbb{E}_{q_\phi(\mathrm{y}|\mathrm{x})}\mathbb{E}_{q_\phi(\mathrm{u}_l,\mathrm{u}_g,\mathrm{z}|\mathrm{x},\mathrm{y})}[p_\theta(\mathrm{y}|\mathrm{u}_l,\mathrm{u}_g,\mathrm{z})].$$
。从经验上讲，我们发现直接使用$q_\phi ($y| x) 和$q_{\phi}(\mathrm{u}_{l},\mathrm{u}_{g}$,z|x,y) 作为$p_\theta($y|u$_l$,u$_g$,z|x,y)作为$p_\theta($y|u$_l$,u$_g$,z)的输入已经达到了令人满意的精度
## 5. Experiments
4.2.2. Generating Counterfactual Explanation
在本节中，我们将评估两个合成数据集和三个实际数据集上的 CounTS 和现有方法。对于每个数据集，我们根据三个指标评估不同的方法：(1)预测准确性，(2) 反事实准确性和 (3)反事实变化率，最后一个是最重要的指标。这些指标对不同的数据集采取不同的形式(详见 Sec. 5. 2- 5. 4 )
和
我们使用$\overline{\mathrm{EFig.2~(右)}}$中的蓝色分支通过反事实推理生成反事实解释。我们的目标是找到下面定义的最佳反事实解释 x$^cf$
Defnition 4.1 ( Optimal Counterfactual Explanation ).Given a factual observation x and prediction y$^{pred}$ , the optimal counterfactual explanation x$^cf$ for the $\bar{\text{counterfactual outcome for y}^{cf}}$is
$$\mathrm{x}^{cf}=\mathrm{argmax}_{\mathrm{x'}}\:p(Y_{\mathrm{x=x'}}(\mathrm{u})=\mathrm{y}^{cf}),$$
5.1. Baselines and Implementations
where u$=(\mathrm u_l,\mathrm u_g)$ and the counterfactual likelihood is
defined as mathfonts
$$p(Y_{\mathrm{x=x^{\prime}}}(\mathrm{u})=\mathrm{y}^{cf})\text{(5)}\\=\sum_{\mathrm{u}}p\left(\mathrm{y}=\mathrm{y}^{cf}|do\left(\mathrm{x}=\mathrm{x}^{\prime}\right),\mathrm{u}\right)p(\mathrm{u}|\mathrm{x}=\mathrm{x},\mathrm{y}=\mathrm{y}^{pred}).$$
我们将我们的计数与最先进的方法进行了比较，用于生成深度学习模型的解释，包括正则化梯度下降(RGD)E(Wachter et al.,2017a),梯度加权类激活映射 (GradTHCAM)$\operatorname{E}($Selvaraju et al.,2017),梯度SHapley 反生成扩展 (GradSHAP 模型)EXMATHX. 3,。请注意，在这些基线中，只有 RGD 和 CounteRGAN 可以生成可操作的解释。其他基线，包括 FIT(专为时间序列模型设计),仅提供重要性分数作为解释；因此，某些评估指标可能不适用于它们(在表中显示为
换句话说，我们搜索最佳的 x$^{cf}$,该$\mathrm{x}^{cf}$会将模型预测从 y$^pred$转移到 y$^cf$,同时保持$(\mathrm u_l,\mathrm u_g)$不变。由
4
|
[
[
0.358642578125,
0.05682373046875,
0.61669921875,
0.05682373046875,
0.61669921875,
0.0706787109375,
0.358642578125,
0.0706787109375
],
[
0.36279296875,
0.08331298828125,
0.43701171875,
0.08331298828125,
0.43701171875,
0.10101318359375,
0.36279296875,
0.10101318359375
],
[
0.77392578125,
0.0889892578125,
0.8447265625,
0.0889892578125,
0.8447265625,
0.10418701171875,
0.77392578125,
0.10418701171875
],
[
0.11517333984375,
0.11553955078125,
0.42724609375,
0.117431640625,
0.42724609375,
0.179931640625,
0.11517333984375,
0.177978515625
],
[
0.78125,
0.11932373046875,
0.7900390625,
0.11932373046875,
0.7900390625,
0.126953125,
0.78125,
0.126953125
],
[
0.6396484375,
0.1881103515625,
0.76953125,
0.1875,
0.76953125,
0.205810546875,
0.6396484375,
0.20703125
],
[
0.7587890625,
0.1881103515625,
0.87158203125,
0.1881103515625,
0.87158203125,
0.205810546875,
0.7587890625,
0.205810546875
],
[
0.5048828125,
0.1932373046875,
0.54345703125,
0.1932373046875,
0.54345703125,
0.20703125,
0.5048828125,
0.20703125
],
[
0.09478759765625,
0.1912841796875,
0.468017578125,
0.189453125,
0.468017578125,
0.20703125,
0.09478759765625,
0.208984375
],
[
0.08905029296875,
0.2197265625,
0.8857421875,
0.2197265625,
0.8857421875,
0.23291015625,
0.08905029296875,
0.23291015625
],
[
0.0906982421875,
0.23486328125,
0.4248046875,
0.23486328125,
0.4248046875,
0.24560546875,
0.0906982421875,
0.24560546875
],
[
0.42236328125,
0.236083984375,
0.43701171875,
0.236083984375,
0.43701171875,
0.2392578125,
0.42236328125,
0.2392578125
],
[
0.439453125,
0.2342529296875,
0.88623046875,
0.2342529296875,
0.88623046875,
0.2474365234375,
0.439453125,
0.2474365234375
],
[
0.0882568359375,
0.2474365234375,
0.301513671875,
0.2474365234375,
0.301513671875,
0.261474609375,
0.0882568359375,
0.261474609375
],
[
0.0882568359375,
0.272705078125,
0.301513671875,
0.271484375,
0.301513671875,
0.285888671875,
0.0882568359375,
0.287353515625
],
[
0.50244140625,
0.271484375,
0.8837890625,
0.272705078125,
0.8837890625,
0.287353515625,
0.50244140625,
0.285888671875
],
[
0.0882568359375,
0.29345703125,
0.473876953125,
0.29541015625,
0.473876953125,
0.3125,
0.0882568359375,
0.31005859375
],
[
0.50341796875,
0.28857421875,
0.88232421875,
0.28857421875,
0.88232421875,
0.3017578125,
0.50341796875,
0.3017578125
],
[
0.50244140625,
0.302978515625,
0.8857421875,
0.302978515625,
0.8857421875,
0.317626953125,
0.50244140625,
0.317626953125
],
[
0.08905029296875,
0.311279296875,
0.47216796875,
0.311279296875,
0.47216796875,
0.324462890625,
0.08905029296875,
0.324462890625
],
[
0.5009765625,
0.318115234375,
0.87744140625,
0.318115234375,
0.87744140625,
0.33203125,
0.5009765625,
0.33203125
],
[
0.0906982421875,
0.326904296875,
0.47216796875,
0.326904296875,
0.47216796875,
0.34033203125,
0.0906982421875,
0.34033203125
],
[
0.5009765625,
0.3359375,
0.88232421875,
0.3359375,
0.88232421875,
0.35107421875,
0.5009765625,
0.35107421875
],
[
0.09234619140625,
0.34228515625,
0.47216796875,
0.34228515625,
0.47216796875,
0.35595703125,
0.09234619140625,
0.35595703125
],
[
0.08905029296875,
0.356689453125,
0.2059326171875,
0.356689453125,
0.2059326171875,
0.37060546875,
0.08905029296875,
0.37060546875
],
[
0.5009765625,
0.352783203125,
0.8857421875,
0.352783203125,
0.8857421875,
0.367431640625,
0.5009765625,
0.367431640625
],
[
0.5009765625,
0.36865234375,
0.8857421875,
0.36865234375,
0.8857421875,
0.38330078125,
0.5009765625,
0.38330078125
],
[
0.496826171875,
0.380126953125,
0.82666015625,
0.380615234375,
0.82666015625,
0.401611328125,
0.496826171875,
0.40087890625
],
[
0.08905029296875,
0.38818359375,
0.214111328125,
0.38818359375,
0.214111328125,
0.3984375,
0.08905029296875,
0.3984375
],
[
0.0906982421875,
0.410400390625,
0.47216796875,
0.410400390625,
0.47216796875,
0.423583984375,
0.0906982421875,
0.423583984375
],
[
0.5009765625,
0.410400390625,
0.8857421875,
0.410400390625,
0.8857421875,
0.423583984375,
0.5009765625,
0.423583984375
],
[
0.0906982421875,
0.425537109375,
0.47216796875,
0.425537109375,
0.47216796875,
0.439453125,
0.0906982421875,
0.439453125
],
[
0.50244140625,
0.425537109375,
0.70751953125,
0.425537109375,
0.70751953125,
0.439453125,
0.50244140625,
0.439453125
],
[
0.08905029296875,
0.43994140625,
0.473876953125,
0.44140625,
0.473876953125,
0.455810546875,
0.08905029296875,
0.455078125
],
[
0.525390625,
0.449462890625,
0.88818359375,
0.449462890625,
0.88818359375,
0.47021484375,
0.525390625,
0.47021484375
],
[
0.08905029296875,
0.455810546875,
0.47216796875,
0.455810546875,
0.47216796875,
0.4697265625,
0.08905029296875,
0.4697265625
],
[
0.08905029296875,
0.47021484375,
0.473876953125,
0.47021484375,
0.473876953125,
0.48486328125,
0.08905029296875,
0.48486328125
],
[
0.50244140625,
0.4765625,
0.88232421875,
0.4765625,
0.88232421875,
0.494384765625,
0.50244140625,
0.494384765625
],
[
0.0906982421875,
0.48486328125,
0.3349609375,
0.48486328125,
0.3349609375,
0.5,
0.0906982421875,
0.5
],
[
0.5,
0.494384765625,
0.88232421875,
0.49365234375,
0.88232421875,
0.50830078125,
0.5,
0.5087890625
],
[
0.1070556640625,
0.505859375,
0.47802734375,
0.50830078125,
0.47802734375,
0.53271484375,
0.1070556640625,
0.53076171875
],
[
0.5009765625,
0.51025390625,
0.87744140625,
0.51123046875,
0.87744140625,
0.52587890625,
0.5009765625,
0.5244140625
],
[
0.5009765625,
0.5263671875,
0.8837890625,
0.5263671875,
0.8837890625,
0.54052734375,
0.5009765625,
0.54052734375
],
[
0.09234619140625,
0.5361328125,
0.471435546875,
0.5361328125,
0.471435546875,
0.55029296875,
0.09234619140625,
0.55029296875
],
[
0.5,
0.54052734375,
0.8837890625,
0.53857421875,
0.8837890625,
0.55615234375,
0.5,
0.55810546875
],
[
0.08905029296875,
0.55322265625,
0.47216796875,
0.55322265625,
0.47216796875,
0.5673828125,
0.08905029296875,
0.5673828125
],
[
0.498291015625,
0.55615234375,
0.77294921875,
0.556640625,
0.77294921875,
0.5712890625,
0.498291015625,
0.57080078125
],
[
0.0906982421875,
0.568359375,
0.22216796875,
0.568359375,
0.22216796875,
0.58203125,
0.0906982421875,
0.58203125
],
[
0.08905029296875,
0.59765625,
0.42236328125,
0.59765625,
0.42236328125,
0.611328125,
0.08905029296875,
0.611328125
],
[
0.5,
0.59228515625,
0.6357421875,
0.59228515625,
0.6357421875,
0.60693359375,
0.5,
0.60693359375
],
[
0.0906982421875,
0.62060546875,
0.473876953125,
0.62060546875,
0.473876953125,
0.6337890625,
0.0906982421875,
0.6337890625
],
[
0.50244140625,
0.6171875,
0.8857421875,
0.6171875,
0.8857421875,
0.6318359375,
0.50244140625,
0.6318359375
],
[
0.0906982421875,
0.63623046875,
0.473876953125,
0.63623046875,
0.473876953125,
0.6494140625,
0.0906982421875,
0.6494140625
],
[
0.50244140625,
0.6318359375,
0.88232421875,
0.6318359375,
0.88232421875,
0.64599609375,
0.50244140625,
0.64599609375
],
[
0.0906982421875,
0.65087890625,
0.1993408203125,
0.65087890625,
0.1993408203125,
0.6640625,
0.0906982421875,
0.6640625
],
[
0.50341796875,
0.6494140625,
0.88232421875,
0.6494140625,
0.88232421875,
0.66015625,
0.50341796875,
0.66015625
],
[
0.50244140625,
0.66357421875,
0.88232421875,
0.66357421875,
0.88232421875,
0.6767578125,
0.50244140625,
0.6767578125
],
[
0.08905029296875,
0.67041015625,
0.473876953125,
0.67041015625,
0.473876953125,
0.6845703125,
0.08905029296875,
0.6845703125
],
[
0.5009765625,
0.6767578125,
0.88232421875,
0.6767578125,
0.88232421875,
0.68994140625,
0.5009765625,
0.68994140625
],
[
0.08660888671875,
0.6845703125,
0.473876953125,
0.68115234375,
0.473876953125,
0.69873046875,
0.08660888671875,
0.7021484375
],
[
0.5,
0.6923828125,
0.65869140625,
0.6923828125,
0.65869140625,
0.70703125,
0.5,
0.70703125
],
[
0.09234619140625,
0.70068359375,
0.473876953125,
0.70068359375,
0.473876953125,
0.7138671875,
0.09234619140625,
0.7138671875
],
[
0.0882568359375,
0.7138671875,
0.326904296875,
0.7138671875,
0.326904296875,
0.72802734375,
0.0882568359375,
0.72802734375
],
[
0.5009765625,
0.71533203125,
0.521484375,
0.71533203125,
0.521484375,
0.72998046875,
0.5009765625,
0.72998046875
],
[
0.156005859375,
0.73779296875,
0.404296875,
0.7392578125,
0.404296875,
0.7568359375,
0.156005859375,
0.755859375
],
[
0.5009765625,
0.74755859375,
0.7607421875,
0.74755859375,
0.7607421875,
0.76123046875,
0.5009765625,
0.76123046875
],
[
0.08905029296875,
0.767578125,
0.473876953125,
0.767578125,
0.473876953125,
0.7822265625,
0.08905029296875,
0.7822265625
],
[
0.5048828125,
0.771484375,
0.88232421875,
0.771484375,
0.88232421875,
0.78515625,
0.5048828125,
0.78515625
],
[
0.0882568359375,
0.78076171875,
0.2467041015625,
0.78271484375,
0.2467041015625,
0.796875,
0.0882568359375,
0.79443359375
],
[
0.50341796875,
0.7861328125,
0.8837890625,
0.7861328125,
0.8837890625,
0.7998046875,
0.50341796875,
0.7998046875
],
[
0.5009765625,
0.8017578125,
0.8837890625,
0.8017578125,
0.8837890625,
0.81494140625,
0.5009765625,
0.81494140625
],
[
0.08905029296875,
0.80810546875,
0.2181396484375,
0.80615234375,
0.2181396484375,
0.82080078125,
0.08905029296875,
0.82275390625
],
[
0.451904296875,
0.80810546875,
0.473876953125,
0.80810546875,
0.473876953125,
0.8212890625,
0.451904296875,
0.8212890625
],
[
0.50341796875,
0.81640625,
0.88232421875,
0.81640625,
0.88232421875,
0.830078125,
0.50341796875,
0.830078125
],
[
0.10540771484375,
0.8291015625,
0.4697265625,
0.82568359375,
0.4697265625,
0.8466796875,
0.10540771484375,
0.849609375
],
[
0.101318359375,
0.8359375,
0.11358642578125,
0.8359375,
0.11358642578125,
0.84033203125,
0.101318359375,
0.84033203125
],
[
0.5009765625,
0.83056640625,
0.8837890625,
0.83056640625,
0.8837890625,
0.8447265625,
0.5009765625,
0.8447265625
],
[
0.1168212890625,
0.849609375,
0.131591796875,
0.849609375,
0.131591796875,
0.85791015625,
0.1168212890625,
0.85791015625
],
[
0.5009765625,
0.8466796875,
0.8857421875,
0.8466796875,
0.8857421875,
0.85986328125,
0.5009765625,
0.85986328125
],
[
0.50341796875,
0.861328125,
0.8837890625,
0.861328125,
0.8837890625,
0.875,
0.50341796875,
0.875
],
[
0.08905029296875,
0.87548828125,
0.473876953125,
0.87548828125,
0.473876953125,
0.89013671875,
0.08905029296875,
0.89013671875
],
[
0.5009765625,
0.876953125,
0.8837890625,
0.876953125,
0.8837890625,
0.89013671875,
0.5009765625,
0.89013671875
],
[
0.08905029296875,
0.8896484375,
0.4755859375,
0.8896484375,
0.4755859375,
0.9072265625,
0.08905029296875,
0.9072265625
],
[
0.5,
0.892578125,
0.8837890625,
0.892578125,
0.8837890625,
0.90576171875,
0.5,
0.90576171875
],
[
0.48193359375,
0.9248046875,
0.494384765625,
0.9248046875,
0.494384765625,
0.93505859375,
0.48193359375,
0.93505859375
]
] |
[
"Title Suppressed Due to Excessive Size",
"p(z|x, u)",
"p(zx, u)",
"",
"Z",
"g(u, zlypred, x)",
") p(ylz,u)",
"p(x)",
"p(x)q(ylx)q(u,zly,x)p(ylz,u)",
"Figure 2. Network structure. We omit subscripts of pe and qs for clarity. Left: CounTS makes predictions on y using the",
"yellow branch. Right: Given the current x and yp",
"",
", CounTS generates counterfactual explanation xef for any target",
"label yef using the blue branch.",
"4.2. Inference Using CounTS",
"于EDefinition4.1中反事实解释的定义涉及对x进行",
"在通过最大化EEqn.6中的CcounTs来学习生成模型",
"干预的因果推断，因此我们需要首先使用观测概率来",
"识别因果概率p(y=ycfdo(x=x'),u)，即。e。，删",
"（Eqn.？？）和推理模型（Eqn.？？）之后，我们的可自",
"除等式中的'do'。下面的定理表明，这是可以实现的。",
"我解释的CounTS可以对（1）使用EXMTHX_7中",
"Theorem 4.1 ( Identifiability ). Given the posterion",
"的黄色分支对Fig.2进行预测（左），以及（2）使用",
"任何XMATHX",
"distribution of exogenous variable p(ut, uglx,y) , the",
"effect of action p(y = ycf|do(x = x'), ul,ug) can be",
"identied using Ep(zx,u,u,)[p (yef z,ul,ug)]",
"4.2.1.Prediction",
"我们使用EFig.2（左）中的黄色分支来预测y。具",
"有关证明，请参见AppendixB.1。使用Theorem4.1",
"体来说，给定一个输入时间序列×，编码器将首先使",
"，我们可以将Eqn.5重写为",
"用qo(yx）推断初始标签y，然后使用q(u,ug,zxy)",
"Lef =Ep(u|x=x,y=ypread)Ep(z/x,u)[p(yef[z,u)], (6)",
"从x和y推断（ui,ug）和z。然后将（ut,ug）和z馈",
"送到预测器pe(ylut,ug,z)中，以推断最终标签y。从",
",其中u=(ut,ug)和p(ux=x,y=ypred)由 q(ux",
"形式上讲，CounTS将标签预测为",
"x,y=ypred）近似。我们使用蒙特卡洛估计来计算",
"yped=gs(yx)E(ur,ugi2xy)(yu,ug,z)](4)",
"Eqn.5和Eqn.6中的期望值，迭代计算梯度",
"（通过反向传播）以类似于E（Wangetal，2019a；Mao",
"。从经验上讲，我们发现直接使用qe(ylx）和",
"etal.，2021a）的方式搜索最佳x，并将其用作xcf",
"qo(ul,ug,zx,y）作为pe(ylut,ug,z）的输人已经达到",
"（请参阅AppendixB.2中的完整算法）",
"了令人满意的精度",
"4.2.2. Generating Counterfactual Explanation",
"5. Experiments",
"我们使用EFig.2（右）中的蓝色分支通过反事实推理",
"在本节中，我们将评估两个合成数据集和三个实际数",
"生成反事实解释。我们的目标是找到下面定义的最佳",
"据集上的CounTS和现有方法。对于每个数据集，我",
"反事实解释xcf",
"们根据三个指标评估不同的方法：（1）预测准确性，（2",
"反事实准确性和（3）反事实变化率，最后一个是最重",
"Definition 4.1 ( Optimal Counterfactual Explanation",
"要的指标。这些指标对不同的数据集采取不同的形式",
"). Given a factual observation x and prediction ypred",
"(详见Sec.5.2-5.4）",
"the optimal counterfactual explanation xcf for the",
"counterfactual outcome for ycf is",
"和",
"xcf = argmaxx p(Yx=x (u) = yef),",
"5.1. Baselines and Implementations",
"where u = (ul, ug) and the counterfactual likelihood is",
"，我们将我们的计数与最先进的方法进行了比较，用",
"defined as mathfonts",
"于生成深度学习模型的解释，包括正则化梯度下降",
"(RGD)E(Wachteretal.，2017a），梯度加权类激活映",
"p(Yx=x (u) = ycf)",
"(5)",
"射（GradTHCAM)E(Selvaraju et al.,2017)，梯度",
"Zp(y = y'f ]do(x= x),u) p(u|x = x, y = ypred)",
"-",
"SHapley反生成扩展（GradSHAP模型)EXMATHX",
"u",
"3，请注意，在这些基线中，只有RGD和CounteRGAN",
"可以生成可操作的解释。其他基线，包括FIT（专为",
"换句话说，我们搜索最佳的xcf，该xcf会将模型预",
"时间序列模型设计），仅提供重要性分数作为解释；因",
"测从ypred转移到yef，同时保持（ur,ug）不变。由",
"此，某些评估指标可能不适用于它们（在表中显示为",
"4"
] |
{
"arxiv_id": "2306.06024",
"language": "zh",
"page_idx": 3
}
|
||
2306.06024-zh-4
|
2306.06024-zh-4.jpg
|
Title Suppressed Due to Excessive Size
Table 1. Results on the
We mark the best
<table>
<tbody>
<tr>
<th rowspan="2">$\rho\partial uu$ $\mathbf{vv}1\boldsymbol{v}$L inderline .</th>
<th> </th>
<th> </th>
<th> </th>
<th> </th>
<th> </th>
<th> </th>
<th> </th>
<th> </th>
<th> </th>
</tr>
<tr>
<th>CounteRGA</th>
<th> RGD</th>
<th>GradCAM</th>
<th>GradSI</th>
<th>radSHAP</th>
<th>LIME</th>
<th>FIT</th>
<th>CAP</th>
<th>CouraTS $(0urs)$</th>
</tr>
<tr>
<td>Pred. Accuracy( %)中</td>
<td> </td>
<td> </td>
<td>85</td>
<td>19</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td>83.26</td>
</tr>
<tr>
<td>$CCR\uparrow$</td>
<td>1.25</td>
<td>1.21</td>
<td>1.09</td>
<td>1.1.</td>
<td>1.13</td>
<td>1.2</td>
<td>1.15 </td>
<td>0.97</td>
<td>1.33</td>
</tr>
<tr>
<td>Counterf. Accuracy 奶 个</td>
<td>78.75</td>
<td>77.96</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td>45.62</td>
<td> </td>
<td>70.28</td>
</tr>
</tbody>
</table>
$““_-”)$
以上所有方法都是用 PyTorch (Paszke et al., 2019) 实现的。为了公平比较，所有基线解释方法中的预测模型与我们的 CounTS 中的推理模块具有相同的神经网络架构。有关体系结构、训练和推理的更多详细信息， 请参见附录
5.2. Toy Dataset
数据集说明。我们设计了一个玩具数据集，其中标签仅受部分输入的影响。一个好的反事实解释应该只修改输人的这一部分，而保持另一部分不变。按照 Fig.1 (左)中的因果图，我们输入了 x$\in\mathbb{R}^12$,其中每个条目都从$\mathcal{N}(\mu_x,\sigma_x^2)$中独立采样，并且从$\mathcal{N}(\mu_u,\sigma_u^2)$中采样了一个外生变量 $u\in \mathbb{R}$ (作为混杂器)。为了指示 x 的哪个部分影响标签，我们引人了一个掩码向量m$\in\mathbb{R}^{12}$,其中第一个6条目 m$_1:6$设置为1，最后 6个条目 m$_7:12$设置为0。然后，我们生成$z\in\mathbb{R}^12$作为 z=u$\cdot\left(\mathrm{m}\odot\mathrm{x}\right)$和标签$y\sim Bern(\sigma(\mathrm{z}^{\top}1+u))$,其中 $\sigma$ 是 sigmoid 函数。这里 x 的第一个 6 条目 x$_{1:6}$与标签相关，最后一个6 条目 x$_{7:12}$与标签无关
评估度量。我们使用三个评估指标：
$\bullet$ Prediction Accuracy. This is the percentage of
time series correctly predicted $(y^{pred}=y)$ in the
test set.
$\bullet$ Counterfactual Accuracy. For a prediction model
$f$ , $\text{the generated counterfactual explanation x}^{cf}$ ,
and the target label $y^cf$, counterfactual accuracy
is the percentage of time series where x$^{cf}$ success-
fully change the model's prediction to $y^cf$ (i.e.,
$f( \bar{\mathrm{x} ^{cf}} ) = \bar{y^{cf}}$ ).
$\bullet$ Counterfactual Change Ratio (CCR). This mea-
sures how well the counterfactual explanation x$^{cf}$
changes the label-related input x$_1:6$ while keep-
ing the label-agnostic input x$_7:12$ unchanged. For-
mally we use the average ratio of $\frac{\|\mathrm{x}_{1:6}^{cf}-\mathrm{x}_{1:6}\|_{1}}{\|\mathrm{x}_{7:12}^{cf}-\mathrm{x}_{7:12}\|_{1}}$
across the test set.
请注意，CCR 是这三个指标中最重要的指标，因为我
们的主要重点是生成可操作和可行的反事实解释
定量结果。Table 1 根据三个指标将我们的计数与基线
进行比较。CounTS 在 CCR 方面优于所有基线，预
测精度损失最小。这表明我们的 CounTS 成功地识
别并修复了外生变量$u$和 m ,以生成可操作和可行
的反事实解释，同时仍然实现了与基线相当的预测精
度。请注意，可操作的方法(i。e。、CounTS、RGD
和 CounteRGAN) 优于重要性评分方法 (i。e。、FIT、
<table>
<tbody>
<tr>
<th>result $W1$ $\underline{\mathrm{underline}}$</th>
<th>th bol</th>
<th>$\operatorname{face}$ 1</th>
<th>and</th>
<th>the</th>
<th>$Se$ $\frac{3}{j}$</th>
<th>ccond</th>
<th>$be_{\mathrm{a}}^{\mathrm{c}}$</th>
<th>st 1</th>
<th>.esu</th>
<th>$1\mathbf{ts}$ Wlth</th>
</tr>
<tr>
<th> </th>
<th> </th>
<th>counteRGAN</th>
<th>$RGD$</th>
<th>GradCA</th>
<th>$lM$ $GI$</th>
<th>radSHAP</th>
<th>$LIME$</th>
<th>FIT</th>
<th>CAP</th>
<th>CounTS $(0urs$,</th>
</tr>
<tr>
<th>Pred. MSE L</th>
<th> </th>
<th> </th>
<th> </th>
<th> </th>
<th>0.128</th>
<th> </th>
<th> </th>
<th> </th>
<th> </th>
<th>0.117</th>
</tr>
<tr>
<td> </td>
<td>2 Active</td>
<td>2.206</td>
<td>2.217</td>
<td>2.043</td>
<td> </td>
<td>1.989</td>
<td>1.872</td>
<td>1.863</td>
<td>1.614</td>
<td>2.322</td>
</tr>
<tr>
<td>L,L,R</td>
<td>Active 1.</td>
<td>0.705</td>
<td>$\overline{0.683}$</td>
<td>0.654</td>
<td> </td>
<td>0.615</td>
<td>0.473</td>
<td>0.55</td>
<td>0.497</td>
<td>0.730</td>
</tr>
<tr>
<td colspan="2">Counterf. MSE</td>
<td>0.074</td>
<td>0.074</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td>0.394</td>
<td> </td>
<td>0.103</td>
</tr>
</tbody>
</table>
LIME 和 GradSHAP)。这是预料之中的，因为基于重要性分数的解释只能解释原始预测，因此无法推断 x 上的什么变化会将预测从$y^{pred}$转移到$y^cf$
我们的反事实准确性低于 CounteRGAN 和 RGD 方法。这是合理的，因为这些基线不会推断和固定后验分布$p($u|x,y),因此生成器(或反向传播)可以更灵活地修改其输入以将$y^{pred}$更接近目标$y^cf$ 。然而，这种灵活性是以低可行性为代价的，反映在它们糟糕的CCR 性能上
## 5.3. Spike Dataset
数据集说明。受 (Tonekaboni et al.,2020) 的启发，我们构建了 Spike 合成数据集。在数据集中，每个时间序列都包含 3 通道，即。e。, x $\in\mathbb{R}^{3\times T}$。每个通道都是具有随机分布尖峰的独立非线性自回归移动平均(NARMA)序列。标签序列 y$\in\mathbb{R}^T$从 0 开始；只要在任何 3 通道中观察到尖峰，它就可以切换到 1 , 并保持 1 直到最后一个时间戳。一个 3 维外生变量u$\in\{0,1\}^3$ 决定了通道中的尖峰是否会影响最终输出； 如果是，我们说通道是活动的。u 中的三个条目中的每一个，i。e。, 分别从参数为0.8、0.4和0的三个不同的伯努利分布中独立采样(有关 Appendix C.1 中数据集的更多详细信息)
评估度量。我们使用三个评估指标：
$\bullet$ Prediction MSE. We use the mean square error (MSE) $\frac 1N\sum _i^{N}\| \mathrm{y} _i^{pred}-$y$_i\|_2^2$to measure the prediction error in the test set with $N$ time series
$\bullet$ Counterfactual MSE. Similar to Sec. 5.2, for a pre-
diction model $f$ , the generated counterfactual explanation x$^cf$,and the target label y$^cf$,counter- ${\mathrm{factual~MSE~is~defined~as~\frac{1}{N}\sum_{i}^{N}\|f(x_{i}^{cf})^{\prime})-y_{i}^{cf}\|_{2}^{2}}}$ ; it measures how successfully x$^cf$ changes the model's prediction to $y^{cf}~.$
$\bullet$CCR. For a given input x$_i$,we set the target
counterfactual label y$_i^{cf}$ by shifting y$^pred$ by 20 timestamps to the right. If at timestamp $t$ , there is a spike in an active channel in x$_i$ triggering the output y$_i^{pred}$ to switch from 0 to 1, an ideal counterfactual explanation x$_i^{cf}$ should (1) suppress all spikes between $[t,t+20)^{\prime},(2)$ create a new spike at $t+20$ timestamp in all active channels of the original input x$_i$ , and (3) keep all inactive channels unchanged. Therefore the counterfactual change ratio can be defined as (with $N$ time series):$CCR=\frac1N\sum_{i=1}^{N}\frac{\|\mathrm{m}_{i}\odot(\mathrm{x}_{i}-\mathrm{x}_{i}^{cf})\|_{1}}{\|(1-\mathrm{m}_{i})\odot(\mathrm{x}_{i}-\mathrm{x}_{i}^{cf})\|_{1}}.$
5
|
[
[
0.358642578125,
0.05682373046875,
0.61767578125,
0.05682373046875,
0.61767578125,
0.0706787109375,
0.358642578125,
0.0706787109375
],
[
0.0882568359375,
0.085205078125,
0.4755859375,
0.085205078125,
0.4755859375,
0.0985107421875,
0.0882568359375,
0.0985107421875
],
[
0.50244140625,
0.08709716796875,
0.8837890625,
0.08709716796875,
0.8837890625,
0.09783935546875,
0.50244140625,
0.09783935546875
],
[
0.0906982421875,
0.10101318359375,
0.47216796875,
0.10101318359375,
0.47216796875,
0.111083984375,
0.0906982421875,
0.111083984375
],
[
0.50244140625,
0.10101318359375,
0.8857421875,
0.10101318359375,
0.8857421875,
0.111083984375,
0.50244140625,
0.111083984375
],
[
0.0906982421875,
0.11553955078125,
0.16015625,
0.11553955078125,
0.16015625,
0.126953125,
0.0906982421875,
0.126953125
],
[
0.50244140625,
0.1142578125,
0.57177734375,
0.1142578125,
0.57177734375,
0.1256103515625,
0.50244140625,
0.1256103515625
],
[
0.1895751953125,
0.1256103515625,
0.471435546875,
0.1256103515625,
0.471435546875,
0.1395263671875,
0.1895751953125,
0.1395263671875
],
[
0.5986328125,
0.125,
0.8896484375,
0.1256103515625,
0.8896484375,
0.1395263671875,
0.5986328125,
0.1383056640625
],
[
0.426513671875,
0.1363525390625,
0.455078125,
0.1363525390625,
0.455078125,
0.1578369140625,
0.426513671875,
0.1578369140625
],
[
0.701171875,
0.1319580078125,
0.765625,
0.146484375,
0.7607421875,
0.1610107421875,
0.6953125,
0.146484375
],
[
0.5087890625,
0.136962890625,
0.56396484375,
0.1383056640625,
0.56396484375,
0.1483154296875,
0.5087890625,
0.147705078125
],
[
0.098876953125,
0.136962890625,
0.1854248046875,
0.136962890625,
0.1854248046875,
0.147705078125,
0.098876953125,
0.147705078125
],
[
0.2900390625,
0.1383056640625,
0.31201171875,
0.1383056640625,
0.31201171875,
0.1458740234375,
0.2900390625,
0.1458740234375
],
[
0.8447265625,
0.136962890625,
0.87158203125,
0.136962890625,
0.87158203125,
0.1558837890625,
0.8447265625,
0.1558837890625
],
[
0.2059326171875,
0.14453125,
0.233642578125,
0.14453125,
0.233642578125,
0.1622314453125,
0.2059326171875,
0.1622314453125
],
[
0.31201171875,
0.14453125,
0.3349609375,
0.14453125,
0.3349609375,
0.1558837890625,
0.31201171875,
0.1558837890625
],
[
0.3505859375,
0.14453125,
0.41015625,
0.14453125,
0.41015625,
0.1558837890625,
0.3505859375,
0.1558837890625
],
[
0.56640625,
0.14453125,
0.6044921875,
0.14453125,
0.6044921875,
0.1546630859375,
0.56640625,
0.1546630859375
],
[
0.75634765625,
0.143310546875,
0.8310546875,
0.14453125,
0.8310546875,
0.1558837890625,
0.75634765625,
0.1546630859375
],
[
0.65625,
0.1458740234375,
0.71337890625,
0.1458740234375,
0.71337890625,
0.1558837890625,
0.65625,
0.1558837890625
],
[
0.1282958984375,
0.1458740234375,
0.156005859375,
0.1458740234375,
0.156005859375,
0.154052734375,
0.1282958984375,
0.154052734375
],
[
0.2459716796875,
0.146484375,
0.26318359375,
0.146484375,
0.26318359375,
0.1546630859375,
0.2459716796875,
0.1546630859375
],
[
0.271240234375,
0.146484375,
0.294189453125,
0.146484375,
0.294189453125,
0.1546630859375,
0.271240234375,
0.1546630859375
],
[
0.521484375,
0.149658203125,
0.55224609375,
0.149658203125,
0.55224609375,
0.1578369140625,
0.521484375,
0.1578369140625
],
[
0.61279296875,
0.146484375,
0.6396484375,
0.146484375,
0.6396484375,
0.1546630859375,
0.61279296875,
0.1546630859375
],
[
0.654296875,
0.1514892578125,
0.7099609375,
0.1514892578125,
0.7099609375,
0.1634521484375,
0.654296875,
0.1634521484375
],
[
0.1258544921875,
0.1514892578125,
0.193603515625,
0.1514892578125,
0.193603515625,
0.1622314453125,
0.1258544921875,
0.1622314453125
],
[
0.0931396484375,
0.154052734375,
0.125,
0.154052734375,
0.125,
0.1622314453125,
0.0931396484375,
0.1622314453125
],
[
0.371826171875,
0.154052734375,
0.39453125,
0.154052734375,
0.39453125,
0.1610107421875,
0.371826171875,
0.1610107421875
],
[
0.4306640625,
0.154052734375,
0.450927734375,
0.154052734375,
0.450927734375,
0.1622314453125,
0.4306640625,
0.1622314453125
],
[
0.56396484375,
0.1546630859375,
0.60546875,
0.1546630859375,
0.60546875,
0.1622314453125,
0.56396484375,
0.1622314453125
],
[
0.61767578125,
0.1546630859375,
0.63818359375,
0.1546630859375,
0.63818359375,
0.1634521484375,
0.61767578125,
0.1634521484375
],
[
0.7236328125,
0.1546630859375,
0.82861328125,
0.1546630859375,
0.82861328125,
0.1622314453125,
0.7236328125,
0.1622314453125
],
[
0.84716796875,
0.1546630859375,
0.86767578125,
0.1546630859375,
0.86767578125,
0.1634521484375,
0.84716796875,
0.1634521484375
],
[
0.5048828125,
0.160400390625,
0.56787109375,
0.160400390625,
0.56787109375,
0.17041015625,
0.5048828125,
0.17041015625
],
[
0.238525390625,
0.1546630859375,
0.2607421875,
0.15283203125,
0.261474609375,
0.1610107421875,
0.240234375,
0.1634521484375
],
[
0.619140625,
0.1610107421875,
0.6396484375,
0.1610107421875,
0.6396484375,
0.1685791015625,
0.619140625,
0.1685791015625
],
[
0.654296875,
0.1622314453125,
0.6787109375,
0.1622314453125,
0.6787109375,
0.169189453125,
0.654296875,
0.169189453125
],
[
0.78515625,
0.160400390625,
0.810546875,
0.1622314453125,
0.8095703125,
0.17041015625,
0.78369140625,
0.1685791015625
],
[
0.8505859375,
0.1622314453125,
0.86767578125,
0.1622314453125,
0.86767578125,
0.17041015625,
0.8505859375,
0.17041015625
],
[
0.5009765625,
0.177978515625,
0.8857421875,
0.179931640625,
0.8857421875,
0.1944580078125,
0.5009765625,
0.1925048828125
],
[
0.0882568359375,
0.1654052734375,
0.1217041015625,
0.17041015625,
0.11767578125,
0.186279296875,
0.08416748046875,
0.18115234375
],
[
0.08905029296875,
0.1932373046875,
0.473876953125,
0.1925048828125,
0.473876953125,
0.20703125,
0.08905029296875,
0.2083740234375
],
[
0.50244140625,
0.1956787109375,
0.88623046875,
0.1956787109375,
0.88623046875,
0.208984375,
0.50244140625,
0.208984375
],
[
0.0906982421875,
0.208984375,
0.47216796875,
0.208984375,
0.47216796875,
0.222900390625,
0.0906982421875,
0.222900390625
],
[
0.5009765625,
0.210205078125,
0.8037109375,
0.208984375,
0.8037109375,
0.2235107421875,
0.5009765625,
0.2247314453125
],
[
0.0906982421875,
0.2247314453125,
0.47216796875,
0.2247314453125,
0.47216796875,
0.238037109375,
0.0906982421875,
0.238037109375
],
[
0.0882568359375,
0.238037109375,
0.4755859375,
0.240478515625,
0.4755859375,
0.25390625,
0.0882568359375,
0.251953125
],
[
0.50244140625,
0.23291015625,
0.8837890625,
0.23291015625,
0.8837890625,
0.246826171875,
0.50244140625,
0.246826171875
],
[
0.5009765625,
0.248779296875,
0.88232421875,
0.248779296875,
0.88232421875,
0.261962890625,
0.5009765625,
0.261962890625
],
[
0.0906982421875,
0.255615234375,
0.17236328125,
0.255615234375,
0.17236328125,
0.267578125,
0.0906982421875,
0.267578125
],
[
0.50244140625,
0.264404296875,
0.8837890625,
0.264404296875,
0.8837890625,
0.27783203125,
0.50244140625,
0.27783203125
],
[
0.5009765625,
0.27783203125,
0.8837890625,
0.27783203125,
0.8837890625,
0.292236328125,
0.5009765625,
0.292236328125
],
[
0.0882568359375,
0.285888671875,
0.2156982421875,
0.285888671875,
0.2156982421875,
0.2998046875,
0.0882568359375,
0.2998046875
],
[
0.5009765625,
0.292236328125,
0.88232421875,
0.292236328125,
0.88232421875,
0.30615234375,
0.5009765625,
0.30615234375
],
[
0.0882568359375,
0.309326171875,
0.473876953125,
0.30810546875,
0.473876953125,
0.322509765625,
0.0882568359375,
0.323974609375
],
[
0.50244140625,
0.31005859375,
0.59326171875,
0.31005859375,
0.59326171875,
0.3212890625,
0.50244140625,
0.3212890625
],
[
0.0906982421875,
0.324462890625,
0.473876953125,
0.324462890625,
0.473876953125,
0.33837890625,
0.0906982421875,
0.33837890625
],
[
0.0906982421875,
0.34033203125,
0.473876953125,
0.34033203125,
0.473876953125,
0.354248046875,
0.0906982421875,
0.354248046875
],
[
0.5009765625,
0.34033203125,
0.63818359375,
0.34033203125,
0.63818359375,
0.354248046875,
0.5009765625,
0.354248046875
],
[
0.08905029296875,
0.354248046875,
0.473876953125,
0.354248046875,
0.473876953125,
0.36865234375,
0.08905029296875,
0.36865234375
],
[
0.0906982421875,
0.36865234375,
0.47216796875,
0.36865234375,
0.47216796875,
0.38330078125,
0.0906982421875,
0.38330078125
],
[
0.5009765625,
0.363037109375,
0.8857421875,
0.363037109375,
0.8857421875,
0.37744140625,
0.5009765625,
0.37744140625
],
[
0.5009765625,
0.37890625,
0.8857421875,
0.37890625,
0.8857421875,
0.392578125,
0.5009765625,
0.392578125
],
[
0.0906982421875,
0.385009765625,
0.473876953125,
0.385009765625,
0.473876953125,
0.3984375,
0.0906982421875,
0.3984375
],
[
0.5,
0.391357421875,
0.8857421875,
0.391357421875,
0.8857421875,
0.4091796875,
0.5,
0.4091796875
],
[
0.0906982421875,
0.40087890625,
0.473876953125,
0.40087890625,
0.473876953125,
0.4140625,
0.0906982421875,
0.4140625
],
[
0.08660888671875,
0.412353515625,
0.476318359375,
0.4140625,
0.476318359375,
0.43115234375,
0.08660888671875,
0.4287109375
],
[
0.5009765625,
0.4091796875,
0.8857421875,
0.4091796875,
0.8857421875,
0.42236328125,
0.5009765625,
0.42236328125
],
[
0.08905029296875,
0.4287109375,
0.4755859375,
0.427978515625,
0.4755859375,
0.44580078125,
0.08905029296875,
0.4462890625
],
[
0.5,
0.423583984375,
0.8857421875,
0.423583984375,
0.8857421875,
0.438232421875,
0.5,
0.438232421875
],
[
0.5009765625,
0.439453125,
0.87744140625,
0.439453125,
0.87744140625,
0.45263671875,
0.5009765625,
0.45263671875
],
[
0.0906982421875,
0.44580078125,
0.473876953125,
0.44580078125,
0.473876953125,
0.460205078125,
0.0906982421875,
0.460205078125
],
[
0.5009765625,
0.453857421875,
0.8857421875,
0.453857421875,
0.8857421875,
0.46728515625,
0.5009765625,
0.46728515625
],
[
0.08905029296875,
0.460205078125,
0.473876953125,
0.4609375,
0.473876953125,
0.47607421875,
0.08905029296875,
0.474853515625
],
[
0.5009765625,
0.4697265625,
0.8857421875,
0.4697265625,
0.8857421875,
0.48291015625,
0.5009765625,
0.48291015625
],
[
0.08905029296875,
0.47607421875,
0.41259765625,
0.47607421875,
0.41259765625,
0.490478515625,
0.08905029296875,
0.490478515625
],
[
0.50244140625,
0.48486328125,
0.8837890625,
0.48486328125,
0.8837890625,
0.498779296875,
0.50244140625,
0.498779296875
],
[
0.0906982421875,
0.498779296875,
0.340576171875,
0.498779296875,
0.340576171875,
0.51220703125,
0.0906982421875,
0.51220703125
],
[
0.50244140625,
0.5,
0.8857421875,
0.5,
0.8857421875,
0.5146484375,
0.50244140625,
0.5146484375
],
[
0.10296630859375,
0.52099609375,
0.4755859375,
0.52099609375,
0.4755859375,
0.53515625,
0.10296630859375,
0.53515625
],
[
0.50244140625,
0.51513671875,
0.8837890625,
0.51513671875,
0.8837890625,
0.52880859375,
0.50244140625,
0.52880859375
],
[
0.12091064453125,
0.5341796875,
0.4755859375,
0.5341796875,
0.4755859375,
0.5517578125,
0.12091064453125,
0.5517578125
],
[
0.5009765625,
0.52880859375,
0.67236328125,
0.52978515625,
0.67236328125,
0.54443359375,
0.5009765625,
0.54345703125
],
[
0.1217041015625,
0.55322265625,
0.1805419921875,
0.55322265625,
0.1805419921875,
0.564453125,
0.1217041015625,
0.564453125
],
[
0.50341796875,
0.55322265625,
0.75244140625,
0.55322265625,
0.75244140625,
0.56640625,
0.50341796875,
0.56640625
],
[
0.101318359375,
0.56640625,
0.473876953125,
0.56494140625,
0.473876953125,
0.580078125,
0.101318359375,
0.58056640625
],
[
0.51318359375,
0.57373046875,
0.8857421875,
0.57470703125,
0.8857421875,
0.5888671875,
0.51318359375,
0.58837890625
],
[
0.1217041015625,
0.58203125,
0.4697265625,
0.58203125,
0.4697265625,
0.59521484375,
0.1217041015625,
0.59521484375
],
[
0.52783203125,
0.583984375,
0.890625,
0.58349609375,
0.890625,
0.60791015625,
0.52783203125,
0.609375
],
[
0.12091064453125,
0.5966796875,
0.473876953125,
0.5966796875,
0.473876953125,
0.611328125,
0.12091064453125,
0.611328125
],
[
0.625,
0.5966796875,
0.63623046875,
0.5966796875,
0.63623046875,
0.60107421875,
0.625,
0.60107421875
],
[
0.53173828125,
0.60498046875,
0.87158203125,
0.60498046875,
0.87158203125,
0.61865234375,
0.53173828125,
0.61865234375
],
[
0.1217041015625,
0.611328125,
0.473876953125,
0.611328125,
0.473876953125,
0.62548828125,
0.1217041015625,
0.62548828125
],
[
0.5146484375,
0.61865234375,
0.88623046875,
0.619140625,
0.88623046875,
0.6337890625,
0.5146484375,
0.63330078125
],
[
0.1217041015625,
0.626953125,
0.473876953125,
0.626953125,
0.473876953125,
0.6416015625,
0.1217041015625,
0.6416015625
],
[
0.53173828125,
0.6337890625,
0.8857421875,
0.63623046875,
0.8857421875,
0.6494140625,
0.53173828125,
0.64794921875
],
[
0.1217041015625,
0.6416015625,
0.2320556640625,
0.6416015625,
0.2320556640625,
0.65478515625,
0.1217041015625,
0.65478515625
],
[
0.53369140625,
0.65087890625,
0.88623046875,
0.65087890625,
0.88623046875,
0.6640625,
0.53369140625,
0.6640625
],
[
0.10296630859375,
0.6572265625,
0.473876953125,
0.6572265625,
0.473876953125,
0.67041015625,
0.10296630859375,
0.67041015625
],
[
0.529296875,
0.66015625,
0.8896484375,
0.66015625,
0.8896484375,
0.6845703125,
0.529296875,
0.6845703125
],
[
0.1192626953125,
0.671875,
0.47216796875,
0.67041015625,
0.47216796875,
0.68505859375,
0.1192626953125,
0.68603515625
],
[
0.74658203125,
0.671875,
0.765625,
0.671875,
0.765625,
0.67529296875,
0.74658203125,
0.67529296875
],
[
0.53369140625,
0.6796875,
0.8857421875,
0.6796875,
0.8857421875,
0.693359375,
0.53369140625,
0.693359375
],
[
0.12091064453125,
0.68603515625,
0.473876953125,
0.6875,
0.473876953125,
0.7021484375,
0.12091064453125,
0.70068359375
],
[
0.53125,
0.69580078125,
0.73046875,
0.6943359375,
0.73046875,
0.708984375,
0.53125,
0.71044921875
],
[
0.12091064453125,
0.7021484375,
0.473876953125,
0.7021484375,
0.473876953125,
0.71630859375,
0.12091064453125,
0.71630859375
],
[
0.5146484375,
0.71044921875,
0.8837890625,
0.71044921875,
0.8837890625,
0.7236328125,
0.5146484375,
0.7236328125
],
[
0.1168212890625,
0.7197265625,
0.471435546875,
0.71337890625,
0.471435546875,
0.734375,
0.1168212890625,
0.74072265625
],
[
0.38232421875,
0.73095703125,
0.471435546875,
0.73095703125,
0.471435546875,
0.7421875,
0.38232421875,
0.7421875
],
[
0.53369140625,
0.72607421875,
0.88623046875,
0.72607421875,
0.88623046875,
0.74072265625,
0.53369140625,
0.74072265625
],
[
0.12091064453125,
0.74072265625,
0.2548828125,
0.74072265625,
0.2548828125,
0.75390625,
0.12091064453125,
0.75390625
],
[
0.53173828125,
0.7412109375,
0.8857421875,
0.74072265625,
0.8857421875,
0.7548828125,
0.53173828125,
0.755859375
],
[
0.0906982421875,
0.76220703125,
0.473876953125,
0.76220703125,
0.473876953125,
0.7763671875,
0.0906982421875,
0.7763671875
],
[
0.53369140625,
0.7568359375,
0.8857421875,
0.7568359375,
0.8857421875,
0.77099609375,
0.53369140625,
0.77099609375
],
[
0.52783203125,
0.76806640625,
0.88623046875,
0.76513671875,
0.88623046875,
0.7861328125,
0.52783203125,
0.7890625
],
[
0.0882568359375,
0.77783203125,
0.431396484375,
0.77783203125,
0.431396484375,
0.79150390625,
0.0882568359375,
0.79150390625
],
[
0.53173828125,
0.787109375,
0.8857421875,
0.787109375,
0.8857421875,
0.8017578125,
0.53173828125,
0.8017578125
],
[
0.0906982421875,
0.80029296875,
0.47216796875,
0.80029296875,
0.47216796875,
0.814453125,
0.0906982421875,
0.814453125
],
[
0.53173828125,
0.80322265625,
0.88623046875,
0.80322265625,
0.88623046875,
0.8173828125,
0.53173828125,
0.8173828125
],
[
0.0906982421875,
0.81640625,
0.47216796875,
0.81640625,
0.47216796875,
0.830078125,
0.0906982421875,
0.830078125
],
[
0.53369140625,
0.818359375,
0.8857421875,
0.818359375,
0.8857421875,
0.83349609375,
0.53369140625,
0.83349609375
],
[
0.0906982421875,
0.83056640625,
0.473876953125,
0.83056640625,
0.473876953125,
0.8447265625,
0.0906982421875,
0.8447265625
],
[
0.53173828125,
0.833984375,
0.8857421875,
0.833984375,
0.8857421875,
0.8486328125,
0.53173828125,
0.8486328125
],
[
0.0906982421875,
0.8466796875,
0.471435546875,
0.8466796875,
0.471435546875,
0.85986328125,
0.0906982421875,
0.85986328125
],
[
0.53369140625,
0.8486328125,
0.8857421875,
0.8486328125,
0.8857421875,
0.8623046875,
0.53369140625,
0.8623046875
],
[
0.0906982421875,
0.861328125,
0.47216796875,
0.861328125,
0.47216796875,
0.875,
0.0906982421875,
0.875
],
[
0.53173828125,
0.8642578125,
0.8837890625,
0.8642578125,
0.8837890625,
0.87744140625,
0.53173828125,
0.87744140625
],
[
0.0906982421875,
0.876953125,
0.473876953125,
0.876953125,
0.473876953125,
0.89013671875,
0.0906982421875,
0.89013671875
],
[
0.529296875,
0.875,
0.88623046875,
0.88623046875,
0.8857421875,
0.91015625,
0.52783203125,
0.8984375
],
[
0.74853515625,
0.87744140625,
0.86279296875,
0.87744140625,
0.86279296875,
0.8916015625,
0.74853515625,
0.8916015625
],
[
0.7197265625,
0.8837890625,
0.73779296875,
0.88330078125,
0.73779296875,
0.88720703125,
0.7197265625,
0.88818359375
],
[
0.08905029296875,
0.8916015625,
0.473876953125,
0.8916015625,
0.473876953125,
0.90478515625,
0.08905029296875,
0.90478515625
],
[
0.4794921875,
0.92236328125,
0.495849609375,
0.92236328125,
0.495849609375,
0.9375,
0.4794921875,
0.9375
]
] |
[
"Title Suppressed Due to Excessive Size",
"Table 1. Results on the toy dataset. We mark the best",
"Table2.Results on the Spike dataset.We mark the best",
"result with bold face and the second best results with",
"result with bold face and the second best results with",
"underline.",
"underline.",
"CounteRGAN RGD GradCAM GradSHAP LIME FIT CAP CounTS (Ous)",
"CounteRGAN RGD GradCAM GradSHAP LIME FIT CAP CounTS (Ous)",
"",
"01281.90",
"Pred. MSE ↓",
"Pred. Acuracy (%)+",
"85.19",
"9.417",
"",
"1.13",
"1.21.150.97",
"2Active",
"1.872 1.863 1.614",
"2.2172.043",
"CCRt",
"1.21",
"1.09",
"CCR+",
"2.206",
"0.6830.654",
"Accuracy (%)",
"Counterf",
"45.62",
"70.28",
"1Active",
"0.705",
"0.6150.4730.550.497",
"0.730",
"Counterf. MSE",
"77.9",
"0.074",
"0.074",
"0.394",
"0.103",
"LIME和GradSHAP)。这是预料之中的，因为基于重",
"\"_\")",
"以上所有方法都是用PyTorch（Paszkeetal.，2019）实",
"要性分数的解释只能解释原始预测，因此无法推断x",
"现的。为了公平比较，所有基线解释方法中的预测模",
"上的什么变化会将预测从ypred转移到ycf",
"型与我们的CounTS中的推理模块具有相同的神经网",
"络架构。有关体系结构、训练和推理的更多详细信息，",
"我们的反事实准确性低于CounteRGAN和RGD方",
"法。这是合理的，因为这些基线不会推断和固定后验",
"请参见附录",
"分布p(ux,y），因此生成器（或反向传播）可以更灵",
"活地修改其输人以将ypred更接近目标ycf。然而，这",
"5.2. Toy Dataset",
"种灵活性是以低可行性为代价的，反映在它们糟糕的",
"数据集说明。我们设计了一个玩具数据集，其中标签",
"CCR性能上",
"仅受部分输人的影响。一个好的反事实解释应该只修",
"改输人的这一部分，而保持另一部分不变。按照Fig.1",
"5.3. Spike Dataset",
"（左）中的因果图，我们输人了xER12，其中每个条",
"目都从N(μe,g2）中独立采样，并且从N（μuo2）中",
"数据集说明。受（Tonekabonietal.,2020）的启发，我",
"们构建了Spike合成数据集。在数据集中，每个时间",
"采样了一个外生变量uER（作为混杂器）。为了指",
"序列都包含3通道，即。e。，xER3xT。每个通道",
"示×的哪个部分影响标签，我们引人了一个掩码向量",
"mER12，其中第一个6条目m1:6设置为1，最后6",
"都是具有随机分布尖峰的独立非线性自回归移动平均",
"个条目m7:12设置为0。然后，我们生成zER12作",
"(NARMA）序列。标签序列yERT从O开始；只",
"要在任何3通道中观察到尖峰，它就可以切换到1",
"为z=u(mox)和标签y～Bern(o(z1+u))，其",
"并保持1直到最后一个时间戳。一个3维外生变量",
"中是sigmoid函数。这里x的第一个6条目xi:6与",
"uE{0,1]3决定了通道中的尖峰是否会影响最终输出；",
"标签相关，最后一个6条目x7:12与标签无关",
"如果是，我们说通道是活动的。u中的三个条目中的",
"评估度量。我们使用三个评估指标：",
"每一个，i。e。，分别从参数为0.8、0.4和0的三个不",
": Prediction Accuracy. This is the percentage of",
"同的伯努利分布中独立采样（有关AppendixC.1中",
"time series correctly predicted ( ypred = y ) in the",
"数据集的更多详细信息）",
"test set.",
"评估度量。我们使用三个评估指标：",
": Counterfactual Accuracy.For a prediction model",
". Prediction MSE. We use the mean square error",
"f , the generated counterfactual explanation xcf",
"(MSE) NIlpred - yll to measure the pre-",
"and the target label ycf , counterfactual accuracy",
"",
"diction error in the test set with N time series.",
"is the percentage of time series where xcf success-",
". Counterfactual MSE. Similar to Sec. 5.2, for a pre-",
"fully change the model's prediction to ycf (i.e.,",
"diction model f , the generated counterfactual ex-",
"f(xcf)=ycf ).",
"planation xcf , and the target label ycf , counter-",
": Counterfactual Change Ratio (CCR). This mea-",
"factual MSE is defined as N I/f(x)-y'",
"sures how well the counterfactual explanation xcf",
"",
": it measures how successfully xcf changes the",
"changes the label-related input xi:6 while keep-",
"model's prediction to yef.",
"ing the label-agnostic input x7:12 unchanged. For-",
". CCR. For a given input Xi , we set the target",
"mally we use the average ratio of llxja-xrell",
"x12—x7:12l1",
"counterfactual label ysf by shifting ypred by 20",
"across the test set.",
"timestamps to the right.If at timestamp t,",
"请注意，CCR是这三个指标中最重要的指标，因为我",
"there is a spike in an active channel in xi trig-",
"gering the output ypred to switch from O to 1,",
"们的主要重点是生成可操作和可行的反事实解释",
"an ideal counterfactual explanation xsf should (1)",
"定量结果。Table1根据三个指标将我们的计数与基线",
"suppress all spikes between [t,t + 20) , (2) create",
"进行比较。CounTS在CCR方面优于所有基线，预",
"a new spike at t+ 20 timestamp in all active chan-",
"测精度损失最小。这表明我们的CounTS成功地识",
"nels of the original input xi , and (3) keep all",
"别并修复了外生变量u和m，以生成可操作和可行",
"inactive channels unchanged. Therefore the coun-",
"的反事实解释，同时仍然实现了与基线相当的预测精",
"terfactual change ratio can be defined as (with N",
"度。请注意，可操作的方法i。e。、CounTS、RGD",
"time series): CCR = 1 (-m0)o(x1-,",
"mio(xi-x/)l",
"",
"和CounteRGAN）优于重要性评分方法（i。e。、FIT、",
"5"
] |
{
"arxiv_id": "2306.06024",
"language": "zh",
"page_idx": 4
}
|
||
2101.07988-zh-0
|
2101.07988-zh-0.jpg
|
$\underline{\text{Published as a conferenc}}$
# 半监督关键点定位
Olga Moskvyak, Frederic Mair School of Electrical Enginee Queensland University of Technology, Australia $\{$ olga.moskvyak,f.maire,fe
Mahsa Baktashmotlagh
School of Information Techno
The University of Queensland, Australia
# Abstract
Knowledge about the location sist in fine-grained classif case of objects that exhibit their visual appearance, suc of a keypoint detection netw for each animal species, whi for labeled data, we propose pose invariant keypoint representations in a semi-supervised manner using a small set of labeled image Keypoint representations are constraint that forces the k tures for the same keypoint by making keypoint represent closer together in feature s outperforms previous methods ma$\tilde{l}$ body landmark local
1 Introduction
检测关键点有助于细粒度分类 (Guo & Farrell, 2019) 和重新识别 (Zhu et al., 2020; Sarfraz et al., 2018) 。在野生动物 (Mathis et al., 2018; Moskvyak et al., 2020; Liu et al., 2019a;b) 领域，由于姿势变化大，需要领域专家进行注释，因此注释数据人类相比，对动物的关键点估计的商业利益较少，并且在收集和精力很少。
无监督地标的检测$^1\operatorname{E}($Jakab et al 但不能在没有监督的情况下检测可感知的地标。另一方面，如果仅在有限数量的标记示例上进行训练，则监督学习有过度拟合的风险。半监督学习在训练数据和大量的未标记数据。它主要是研究分类任务 (van 重要的关键点定位问题，因为每个图像注释多个关键点是一个耗是最重要的因素。伪标记 (Lee, 2013) 是一种常由在标记子集上训练的模型预测的标签(称为伪标签)。采用启程选择伪标记数据。最近，(Dong \& Yang, 2用于半监督面部标志检测的伪标记中的选择标准。但是，与人体志位置的变化较小，在这种情况下，将不准确的伪标记示例转移高，这对模型有害
(Honari et al., 2018) 在半监督些属性注释的数据集上进行测试。我们的工作重点是在常见的现实世界场景中的关键点定位任务，其中为来自大型未标记数据集的一小部分数据提供注释一种用于半监督关键点定位的方法，该方法学习每个图像的热图表 (图$\mathbb{E}$1 )。语义关键点表示是相对于图像大
$^{1}$We use terms keypoints generic than body joints (used in human pose estimation) because our method is applicable to a $\bar{\text{variety of categories}}.$
1
|
[
[
0.176513671875,
0.034088134765625,
0.50634765625,
0.034088134765625,
0.50634765625,
0.04736328125,
0.176513671875,
0.04736328125
],
[
0.1748046875,
0.10162353515625,
0.404296875,
0.10162353515625,
0.404296875,
0.126953125,
0.1748046875,
0.126953125
],
[
0.1854248046875,
0.15283203125,
0.51708984375,
0.15283203125,
0.51708984375,
0.1666259765625,
0.1854248046875,
0.1666259765625
],
[
0.5908203125,
0.1546630859375,
0.75732421875,
0.1546630859375,
0.75732421875,
0.1654052734375,
0.5908203125,
0.1654052734375
],
[
0.1846923828125,
0.1666259765625,
0.51318359375,
0.1666259765625,
0.51318359375,
0.179931640625,
0.1846923828125,
0.179931640625
],
[
0.5888671875,
0.1673583984375,
0.99853515625,
0.1673583984375,
0.99853515625,
0.18115234375,
0.5888671875,
0.18115234375
],
[
0.1829833984375,
0.179931640625,
0.52392578125,
0.18115234375,
0.52392578125,
0.1956787109375,
0.1829833984375,
0.1944580078125
],
[
0.5888671875,
0.18115234375,
0.8798828125,
0.18115234375,
0.8798828125,
0.1944580078125,
0.5888671875,
0.1944580078125
],
[
0.188720703125,
0.1944580078125,
0.59326171875,
0.1932373046875,
0.59326171875,
0.20703125,
0.188720703125,
0.2083740234375
],
[
0.5849609375,
0.1944580078125,
0.8056640625,
0.1944580078125,
0.8056640625,
0.2083740234375,
0.5849609375,
0.2083740234375
],
[
0.463134765625,
0.24560546875,
0.53759765625,
0.24560546875,
0.53759765625,
0.260009765625,
0.463134765625,
0.260009765625
],
[
0.233642578125,
0.27783203125,
0.76123046875,
0.279052734375,
0.76123046875,
0.29345703125,
0.233642578125,
0.292236328125
],
[
0.233642578125,
0.292236328125,
0.7646484375,
0.292236328125,
0.7646484375,
0.30615234375,
0.233642578125,
0.30615234375
],
[
0.2344970703125,
0.306884765625,
0.7646484375,
0.306884765625,
0.7646484375,
0.32080078125,
0.2344970703125,
0.32080078125
],
[
0.2344970703125,
0.3212890625,
0.765625,
0.3212890625,
0.765625,
0.335205078125,
0.2344970703125,
0.335205078125
],
[
0.2344970703125,
0.335205078125,
0.7646484375,
0.335205078125,
0.7646484375,
0.348388671875,
0.2344970703125,
0.348388671875
],
[
0.233642578125,
0.348388671875,
0.765625,
0.348388671875,
0.765625,
0.3623046875,
0.233642578125,
0.3623046875
],
[
0.2320556640625,
0.361083984375,
0.765625,
0.3623046875,
0.765625,
0.376953125,
0.2320556640625,
0.375732421875
],
[
0.2344970703125,
0.376953125,
0.765625,
0.376953125,
0.765625,
0.39013671875,
0.2344970703125,
0.39013671875
],
[
0.233642578125,
0.39013671875,
0.76318359375,
0.39013671875,
0.76318359375,
0.404052734375,
0.233642578125,
0.404052734375
],
[
0.2320556640625,
0.404052734375,
0.7646484375,
0.40478515625,
0.7646484375,
0.41845703125,
0.2320556640625,
0.417236328125
],
[
0.2320556640625,
0.41845703125,
0.76123046875,
0.417236328125,
0.76123046875,
0.43115234375,
0.2320556640625,
0.431884765625
],
[
0.233642578125,
0.43115234375,
0.765625,
0.43115234375,
0.765625,
0.4443359375,
0.233642578125,
0.4443359375
],
[
0.2344970703125,
0.44580078125,
0.765625,
0.44580078125,
0.765625,
0.458984375,
0.2344970703125,
0.458984375
],
[
0.2344970703125,
0.460205078125,
0.7646484375,
0.460205078125,
0.7646484375,
0.473388671875,
0.2344970703125,
0.473388671875
],
[
0.2320556640625,
0.472900390625,
0.7646484375,
0.4716796875,
0.7646484375,
0.486083984375,
0.2320556640625,
0.4873046875
],
[
0.236083984375,
0.488037109375,
0.46728515625,
0.488037109375,
0.46728515625,
0.498779296875,
0.236083984375,
0.498779296875
],
[
0.1707763671875,
0.53076171875,
0.316162109375,
0.53271484375,
0.316162109375,
0.54736328125,
0.1707763671875,
0.54541015625
],
[
0.1748046875,
0.56494140625,
0.82421875,
0.56494140625,
0.82421875,
0.5791015625,
0.1748046875,
0.5791015625
],
[
0.1748046875,
0.5791015625,
0.82275390625,
0.5791015625,
0.82275390625,
0.59228515625,
0.1748046875,
0.59228515625
],
[
0.176513671875,
0.59228515625,
0.82275390625,
0.59228515625,
0.82275390625,
0.60595703125,
0.176513671875,
0.60595703125
],
[
0.17724609375,
0.60595703125,
0.82275390625,
0.60595703125,
0.82275390625,
0.62060546875,
0.17724609375,
0.62060546875
],
[
0.176513671875,
0.6220703125,
0.25,
0.6220703125,
0.25,
0.63330078125,
0.176513671875,
0.63330078125
],
[
0.17724609375,
0.6416015625,
0.82177734375,
0.6416015625,
0.82177734375,
0.65478515625,
0.17724609375,
0.65478515625
],
[
0.1748046875,
0.65380859375,
0.82275390625,
0.65478515625,
0.82275390625,
0.66943359375,
0.1748046875,
0.66845703125
],
[
0.17724609375,
0.66943359375,
0.82421875,
0.66943359375,
0.82421875,
0.68310546875,
0.17724609375,
0.68310546875
],
[
0.176513671875,
0.68310546875,
0.82275390625,
0.68310546875,
0.82275390625,
0.6962890625,
0.176513671875,
0.6962890625
],
[
0.176513671875,
0.69775390625,
0.82275390625,
0.69775390625,
0.82275390625,
0.7109375,
0.176513671875,
0.7109375
],
[
0.176513671875,
0.7109375,
0.82275390625,
0.7109375,
0.82275390625,
0.724609375,
0.176513671875,
0.724609375
],
[
0.176513671875,
0.72607421875,
0.82275390625,
0.72607421875,
0.82275390625,
0.7392578125,
0.176513671875,
0.7392578125
],
[
0.176513671875,
0.73779296875,
0.82275390625,
0.73779296875,
0.82275390625,
0.751953125,
0.176513671875,
0.751953125
],
[
0.1748046875,
0.75048828125,
0.82275390625,
0.751953125,
0.82275390625,
0.7666015625,
0.1748046875,
0.76513671875
],
[
0.176513671875,
0.7666015625,
0.82275390625,
0.7666015625,
0.82275390625,
0.77978515625,
0.176513671875,
0.77978515625
],
[
0.176513671875,
0.7822265625,
0.306396484375,
0.7822265625,
0.306396484375,
0.79248046875,
0.176513671875,
0.79248046875
],
[
0.176513671875,
0.80029296875,
0.82275390625,
0.80029296875,
0.82275390625,
0.814453125,
0.176513671875,
0.814453125
],
[
0.1748046875,
0.81298828125,
0.82275390625,
0.814453125,
0.82275390625,
0.8291015625,
0.1748046875,
0.82763671875
],
[
0.17724609375,
0.83056640625,
0.8203125,
0.83056640625,
0.8203125,
0.84130859375,
0.17724609375,
0.84130859375
],
[
0.1854248046875,
0.84326171875,
0.82275390625,
0.84326171875,
0.82275390625,
0.85693359375,
0.1854248046875,
0.85693359375
],
[
0.1748046875,
0.85693359375,
0.82177734375,
0.85693359375,
0.82177734375,
0.87060546875,
0.1748046875,
0.87060546875
],
[
0.193603515625,
0.8837890625,
0.82421875,
0.88720703125,
0.82421875,
0.9013671875,
0.193603515625,
0.8984375
],
[
0.1748046875,
0.8994140625,
0.82421875,
0.8994140625,
0.82421875,
0.91357421875,
0.1748046875,
0.91357421875
],
[
0.17724609375,
0.9140625,
0.310546875,
0.9140625,
0.310546875,
0.9248046875,
0.17724609375,
0.9248046875
],
[
0.49267578125,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2021",
"半监督关键点定位",
"Olga Moskvyak, Frederic Maire, Feras Dayoub",
"MahsaBaktashmotlagh",
"School of Electrical Engineering and Robotics",
"School of Information Technology and Electrical Enginee",
"Queensland University of Technology, Australia",
"The University of Queensland, Australia",
"[olga.moskvyak,f.maire,feras.dayoub}@qut.edu.edu",
"m.baktashmotlagh@uq.edu.au",
"Abstract",
"Knowledge about the locations of keypoints of an object in an image can as",
"sist in fine-grained classification and identification tasks, particularly for the",
"case of objects that exhibit large variations in poses that greatly influence",
"their visual appearance, such as wild animals. However, supervised training",
"of a keypoint detection network requires annotating a large image dataset",
"for each animal species, which is a labor-intensive task. To reduce the need",
"for labeled data, we propose to learn simultaneously keypoint heatmaps and",
"pose invariant keypoint representations in a semi-supervised manner using",
"a small set of labeled images along with a larger set of unlabeled images.",
"Keypoint representations arelearnt with a semantic keypoint consistency",
"constraint that forces the keypoint detection network to learn similar fea",
"tures for the same keypoint across the dataset. Pose invariance is achieved",
"by making keypoint representations for the image and its augmented copies",
"closer together in feature space.Our semi-supervised approach significantly",
"outperforms previous methods on several benchmarks for human and ani-",
"malbodylandmarklocalization.",
"1Introduction",
"检测关键点有助于细粒度分类（Guo&Farrell，2019）和重新识别（Zhuetal.，2020;Sarfraz",
"et al.,2018)。在野生动物(Mathis et al.,2018;Moskvyak et al.,2020;Liu et al.,2019a;b)",
"领域，由于姿势变化大，需要领域专家进行注释，因此注释数据尤其具有挑战性。此外，与",
"人类相比，对动物的关键点估计的商业利益较少，并且在收集和注释公共数据集上投人的",
"精力很少。",
"无监督地标的检测1E(Jakabetal.，2018；Thewlisetal.,2017;2019）可以提取有用的特征",
"但不能在没有监督的情况下检测可感知的地标。另一方面，如果仅在有限数量的标记示例",
"上进行训练，则监督学习有过度拟合的风险。半监督学习在训练过程中结合了少量的标记",
"数据和大量的未标记数据。它主要是研究分类任务（vanEngelen&Hoos，2019），但它也是",
"重要的关键点定位问题，因为每个图像注释多个关键点是一个耗时的手工工作，其中精度",
"是最重要的因素。伪标记（Lee，2013）是一种常见的半监督方法，其中未标记的示例被分配",
"由在标记子集上训练的模型预测的标签（称为伪标签）。采用启发式无监督准则为再训练过",
"程选择伪标记数据。最近，（Dong&Yang，2019；Radosavovicetal.，2018）的工作将变体应",
"用于半监督面部标志检测的伪标记中的选择标准。但是，与人体或动物体关节相比，面部标",
"志位置的变化较小，在这种情况下，将不准确的伪标记示例转移到重新训练阶段的风险很",
"高，这对模型有害",
"(Honari etal.，2018）在半监督地标检测中的先前工作利用了其他类属性，并且仅在提供这",
"些属性注释的数据集上进行测试。我们的工作重点是在常见的现实世界场景中的关键点定",
"位任务，其中为来自大型未标记数据集的一小部分数据提供注释。更具体地说，我们提出了",
"一种用于半监督关键点定位的方法，该方法学习每个图像的热图列表和语义关键点表示列",
"表（图E1）。语义关键点表示是相对于图像大小的低维空间中的实数向量，不同图像中的",
"IWe use terms keypoints or landmarks interchangeably in our work.These terms are more",
"generic than body joints (used in human pose estimation) because our method is applicable to a",
"variety of categories",
"1"
] |
{
"arxiv_id": "2101.07988",
"language": "zh",
"page_idx": 0
}
|
||
2101.07988-zh-1
|
2101.07988-zh-1.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2021}}$
<FigureHere>
Figure 1: Our semi-supervised keypoint localization system learns a list of heatmaps and a list of semantic keypoint representations for each image. In addition to a supervised loss optimized on the labeled subset of the data, we propose several unsupervised constraints of transformation equivariance, transformation invariance, and semantic consistency.
相同关键点具有相似的表示。我们利用特定于地标本地化问题的属性来设计约束，以共同
优化两个表示
我们扩展了(Honari et al., 2018) 的转换一致性约束，以便能够将其不同地应用于每个表示(i。e。热图的变换等变约束和语义表示的变换不变约束)。此外，我们制定了一个语义一致性约束，该约束鼓励在同一地标的图像中检测相似特征，而与对象的姿态无关(例如。g。所有图像中的一只眼睛应该看起来相似)。同时学习这两种表示允许我们使用监督和无监督学习的力量
我们的工作是出于野生动物领域的数据稀缺，但不仅限于动物，而且还适用于人体地标检
测。我们工作的贡献有三个方面：
$\bullet$ We propose a technique for semi-supervised keypoint localization that jointly learns
keypoint heatmaps and semantic representations optimised with supervised and
unsupervised constraints;
$\bullet$ Our method can be easily added to any existing keypoint localization networks with
no structural and with minimal computational overhead;
$\bullet$ We evaluate the proposed method on annotated image datasets for both humans
and animals. As demonstrated by our results, our method significantly outperforms previously proposed supervised and unsupervised methods on several benchmarks, ${\mathrm{using~only~limited~labeled~data.}}$
论文组织如下。关于半监督学习和关键点定位的相关工作在 2节中进行了回顾。我们提出
的方法在 3 节中进行了描述。实验设置、数据集和结果在 4 节中讨论。
## 2 Related Work
关键点本地化。受监督的关键点定位研究由几个大型数据集驱动，这些数据集带有标记的关键点，跨越几个常见的研究领域，包括人类姿态估计 (Andriluka et al., 2014) 和面部关键点 (Sagonas et al., 2016) 。获得关键点注释的挑战导致了无监督地标定位研究的兴起。几种无监督方法利用等方差的概念，这意味着地标坐标在合成变换后或后续视频帧中保持一致。Thewlis et al. (2017)建议学习与视点无关的表示，这些表示对于不同的变换是等变的而 Dong et al.(2018)则利用光流的相干性作为监督源。Zhang et al.(2018) 通过强制约束来学习地标编码，这些约束反映了地标的必要属性，例如可分离性和集中度。Jakab et al. (2018)提出了一种生成方法，其中使用预测的热图从转换后的副本重建输人图像。最近的工作 (Thewlis et al., 2019) 通过交换描述符向量来强制同一对象的实例之间的一致性。这些方法主要是在运动和变换过程中的自由度小于人或动物身体关节的人脸上进行评估的。。我们将我们的方法与 4 节中有监督和上述无监督方法的组合进行了比较
半监督学习是分类任务中研究最多的。伪标记(Lee, 2013)是一种方法，它使用模型的类预测作为未标记示例的人工标签，然后训练模型来预测这些标签。另一种技术是一致性正则化，它指出来自未标记数据集的输人示例的实际扰动不应显著改变神经网络的输出。$\Pi$模
2
|
[
[
0.176513671875,
0.034088134765625,
0.50634765625,
0.034088134765625,
0.50634765625,
0.048614501953125,
0.176513671875,
0.048614501953125
],
[
0.55126953125,
0.10479736328125,
0.59228515625,
0.10479736328125,
0.59228515625,
0.11932373046875,
0.55126953125,
0.11932373046875
],
[
0.6806640625,
0.1060791015625,
0.79736328125,
0.106689453125,
0.79736328125,
0.11871337890625,
0.6806640625,
0.117431640625
],
[
0.68115234375,
0.11871337890625,
0.810546875,
0.11871337890625,
0.810546875,
0.1287841796875,
0.68115234375,
0.1287841796875
],
[
0.301513671875,
0.106689453125,
0.455078125,
0.106689453125,
0.455078125,
0.117431640625,
0.301513671875,
0.117431640625
],
[
0.1854248046875,
0.1300048828125,
0.25732421875,
0.1300048828125,
0.25732421875,
0.141357421875,
0.1854248046875,
0.141357421875
],
[
0.5390625,
0.1351318359375,
0.62158203125,
0.1351318359375,
0.62158203125,
0.146484375,
0.5390625,
0.146484375
],
[
0.5185546875,
0.1610107421875,
0.6337890625,
0.1634521484375,
0.6337890625,
0.1767578125,
0.5185546875,
0.1749267578125
],
[
0.398681640625,
0.1925048828125,
0.451904296875,
0.1925048828125,
0.451904296875,
0.203857421875,
0.398681640625,
0.203857421875
],
[
0.5361328125,
0.189453125,
0.6259765625,
0.189453125,
0.6259765625,
0.2008056640625,
0.5361328125,
0.2008056640625
],
[
0.398681640625,
0.2139892578125,
0.451904296875,
0.212158203125,
0.451904296875,
0.2235107421875,
0.398681640625,
0.2259521484375
],
[
0.515625,
0.216552734375,
0.63818359375,
0.216552734375,
0.63818359375,
0.2298583984375,
0.515625,
0.2298583984375
],
[
0.176513671875,
0.266357421875,
0.82421875,
0.266357421875,
0.82421875,
0.27978515625,
0.176513671875,
0.27978515625
],
[
0.1748046875,
0.281005859375,
0.82177734375,
0.281005859375,
0.82177734375,
0.294189453125,
0.1748046875,
0.294189453125
],
[
0.176513671875,
0.294189453125,
0.82421875,
0.294189453125,
0.82421875,
0.30810546875,
0.176513671875,
0.30810546875
],
[
0.1732177734375,
0.306884765625,
0.7587890625,
0.309326171875,
0.7587890625,
0.322509765625,
0.1732177734375,
0.32080078125
],
[
0.176513671875,
0.3466796875,
0.82275390625,
0.3466796875,
0.82275390625,
0.35986328125,
0.176513671875,
0.35986328125
],
[
0.17724609375,
0.3623046875,
0.27294921875,
0.3623046875,
0.27294921875,
0.37255859375,
0.17724609375,
0.37255859375
],
[
0.176513671875,
0.3818359375,
0.82275390625,
0.3818359375,
0.82275390625,
0.395751953125,
0.176513671875,
0.395751953125
],
[
0.176513671875,
0.395751953125,
0.82177734375,
0.395751953125,
0.82177734375,
0.410400390625,
0.176513671875,
0.410400390625
],
[
0.176513671875,
0.410400390625,
0.82275390625,
0.410400390625,
0.82275390625,
0.423583984375,
0.176513671875,
0.423583984375
],
[
0.17724609375,
0.425537109375,
0.8203125,
0.425537109375,
0.8203125,
0.436279296875,
0.17724609375,
0.436279296875
],
[
0.17724609375,
0.439453125,
0.2418212890625,
0.439453125,
0.2418212890625,
0.45068359375,
0.17724609375,
0.45068359375
],
[
0.176513671875,
0.458984375,
0.82275390625,
0.458984375,
0.82275390625,
0.472900390625,
0.176513671875,
0.472900390625
],
[
0.1748046875,
0.472900390625,
0.410888671875,
0.473388671875,
0.410888671875,
0.4873046875,
0.1748046875,
0.486083984375
],
[
0.21728515625,
0.5,
0.8203125,
0.5,
0.8203125,
0.51318359375,
0.21728515625,
0.51318359375
],
[
0.2320556640625,
0.51220703125,
0.82275390625,
0.51318359375,
0.82275390625,
0.52783203125,
0.2320556640625,
0.5263671875
],
[
0.233642578125,
0.52587890625,
0.416748046875,
0.5263671875,
0.416748046875,
0.54150390625,
0.233642578125,
0.54052734375
],
[
0.2156982421875,
0.54541015625,
0.82275390625,
0.54541015625,
0.82275390625,
0.5595703125,
0.2156982421875,
0.5595703125
],
[
0.2320556640625,
0.56005859375,
0.640625,
0.56005859375,
0.640625,
0.57373046875,
0.2320556640625,
0.57373046875
],
[
0.2156982421875,
0.5771484375,
0.82275390625,
0.5771484375,
0.82275390625,
0.59130859375,
0.2156982421875,
0.59130859375
],
[
0.236083984375,
0.59326171875,
0.8203125,
0.59326171875,
0.8203125,
0.60400390625,
0.236083984375,
0.60400390625
],
[
0.2344970703125,
0.60498046875,
0.82275390625,
0.60498046875,
0.82275390625,
0.619140625,
0.2344970703125,
0.619140625
],
[
0.233642578125,
0.619140625,
0.461669921875,
0.61865234375,
0.461669921875,
0.6318359375,
0.233642578125,
0.63330078125
],
[
0.17724609375,
0.64453125,
0.82275390625,
0.64453125,
0.82275390625,
0.65771484375,
0.17724609375,
0.65771484375
],
[
0.176513671875,
0.6591796875,
0.68896484375,
0.6591796875,
0.68896484375,
0.67236328125,
0.176513671875,
0.67236328125
],
[
0.1732177734375,
0.6923828125,
0.326904296875,
0.6923828125,
0.326904296875,
0.70703125,
0.1732177734375,
0.70703125
],
[
0.17724609375,
0.7236328125,
0.82177734375,
0.7236328125,
0.82177734375,
0.7373046875,
0.17724609375,
0.7373046875
],
[
0.178955078125,
0.7392578125,
0.8203125,
0.7392578125,
0.8203125,
0.75,
0.178955078125,
0.75
],
[
0.176513671875,
0.751953125,
0.82275390625,
0.751953125,
0.82275390625,
0.76513671875,
0.176513671875,
0.76513671875
],
[
0.17724609375,
0.767578125,
0.80810546875,
0.767578125,
0.80810546875,
0.77783203125,
0.17724609375,
0.77783203125
],
[
0.176513671875,
0.77978515625,
0.82177734375,
0.77978515625,
0.82177734375,
0.79345703125,
0.176513671875,
0.79345703125
],
[
0.176513671875,
0.79150390625,
0.8203125,
0.79150390625,
0.8203125,
0.8046875,
0.176513671875,
0.8046875
],
[
0.176513671875,
0.80615234375,
0.82177734375,
0.8046875,
0.82177734375,
0.818359375,
0.176513671875,
0.8193359375
],
[
0.176513671875,
0.82080078125,
0.8203125,
0.82080078125,
0.8203125,
0.833984375,
0.176513671875,
0.833984375
],
[
0.176513671875,
0.833984375,
0.8203125,
0.833984375,
0.8203125,
0.84765625,
0.176513671875,
0.84765625
],
[
0.178955078125,
0.849609375,
0.82177734375,
0.849609375,
0.82177734375,
0.85986328125,
0.178955078125,
0.85986328125
],
[
0.176513671875,
0.861328125,
0.67333984375,
0.8623046875,
0.67333984375,
0.87548828125,
0.176513671875,
0.875
],
[
0.178955078125,
0.88525390625,
0.82177734375,
0.88525390625,
0.82177734375,
0.89599609375,
0.178955078125,
0.89599609375
],
[
0.17724609375,
0.89794921875,
0.82177734375,
0.89794921875,
0.82177734375,
0.9111328125,
0.17724609375,
0.9111328125
],
[
0.17724609375,
0.91357421875,
0.82177734375,
0.91357421875,
0.82177734375,
0.923828125,
0.17724609375,
0.923828125
],
[
0.490234375,
0.94873046875,
0.50732421875,
0.94873046875,
0.50732421875,
0.9619140625,
0.490234375,
0.9619140625
]
] |
[
"Published as a conference paper at ICLR 2021",
"Losses",
"Ground truth heatmap:",
"fora subset ofthe datase",
"Keypointlocalizationnetwork",
"Image dataset",
"Supervised loss",
"Trans.equivariance",
"",
"Trans.invariance",
"",
"Semantic consistency",
"Figure l: Our semi-supervised keypoint localization system learns a list of heatmaps and",
"a list of semantic keypoint representations for each image. In addition to a supervised loss",
"optimized on the labeled subset of the data, we propose several unsupervised constraints of",
"transformation equivariance, transformation invariance, and semantic consistency.",
"相同关键点具有相似的表示。我们利用特定于地标本地化问题的属性来设计约束，以共同",
"优化两个表示",
"我们扩展了（Honarietal.，2018）的转换一致性约束，以便能够将其不同地应用于每个表示",
"(i。e。热图的变换等变约束和语义表示的变换不变约束)。此外，我们制定了一个语义一致",
"性约束，该约束鼓励在同一地标的图像中检测相似特征，而与对象的姿态无关（例如。g。所",
"有图像中的一只眼睛应该看起来相似。同时学习这两种表示充许我们使用监督和无监督学",
"习的力量",
"我们的工作是出于野生动物领域的数据稀缺，但不仅限于动物，而且还适用于人体地标检",
"测。我们工作的贡献有三个方面：",
".We propose a technique for semi-supervised keypoint localization that jointly learns",
"keypoint heatmaps and semantic representations optimised with supervised and",
"unsupervised constraints;",
".. Our method can be easily added to any existing keypoint localization networks with",
"no structural and with minimal computational overhead;",
": We evaluate the proposed method on annotated image datasets for both humans",
"andanimals.Asdemonstratedbyourresults,ourmethodsignificantlyoutperforms",
"previously proposed supervised and unsupervised methods on several benchmarks,",
"using only limited labeled data.",
"论文组织如下。关于半监督学习和关键点定位的相关工作在2节中进行了回顾。我们提出",
"的方法在3节中进行了描述。实验设置、数据集和结果在4节中讨论。",
"2Related Work",
"关键点本地化。受监督的关键点定位研究由几个大型数据集驱动，这些数据集带有标记的",
"关键点，跨越儿个常见的研究领域，包括人类姿态估计（Andrilukaetal.，2014）和面部关键",
"点（Sagonasetal.，2016）。获得关键点注释的挑战导致了无监督地标定位研究的兴起。几",
"种无监督方法利用等方差的概念：这意味看地标坐标在合成变换后或后续视频顿中保持",
"致。Thewlisetal.（2017）建议学习与视点无关的表示，这些表示对于不同的变换是等变的",
"而Dongetal.（2018）则利用光流的相干性作为监督源。Zhangetal.（2018）通过强制约束",
"来学习地标编码，这些约束反映了地标的必要属性，例如可分离性和集中度。Jakabetal.",
"(2018）提出了一种生成方法，其中使用预测的热图从转换后的副本重建输人图像。最近的",
"工作（Thewlisetal.，2019）通过交换描述符向量来强制同一对象的实例之间的一致性。这些",
"方法主要是在运动和变换过程中的自由度小于人或动物身体关节的人脸上进行评估的。。我",
"们将我们的方法与4节中有监督和上述无监督方法的组合进行了比较",
"半监督学习是分类任务中研究最多的。伪标记（Lee，2013）是一种方法，它使用模型的类预",
"测作为未标记示例的人工标签，然后训练模型来预测这些标签。另一种技术是一致性正则",
"化，它指出来自未标记数据集的输人示例的实际扰动不应显著改变神经网络的输出。I模",
"2"
] |
{
"arxiv_id": "2101.07988",
"language": "zh",
"page_idx": 1
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.