id
string | image
string | input
string | system
string | output
string | boxes
list | box_texts
list | detail
dict |
|---|---|---|---|---|---|---|---|
2209.13513-zh-4
|
2209.13513-zh-4.jpg
|
Title Suppressed Due to Excessive Size
函数是
$$\begin{aligned}\mathcal{L}(y,\{\hat{y},G_{1:T}\})&=\mathcal{L}_{\mathrm{CE}}(y,\hat{y})+\lambda_{\mathrm{FS}}\mathcal{L}_{\mathrm{FS}}(\boldsymbol{F}_{1:T},\boldsymbol{A}_{1:T})\\&+\lambda_{\mathrm{TS}}\mathcal{L}_{\mathrm{TS}}(\boldsymbol{A}_{1:T})+\lambda_{\mathrm{SP}}\mathcal{L}_{\mathrm{SP}}(\boldsymbol{A}_{1:T})\end{aligned}$$
(19)
,其中$\lambda_\mathrm{FS},\lambda_\mathrm{TS},\lambda_\mathrm{SP}\geq0$是加权每个正则化项的相对
贡献的超参数
## 5. Experiments
我们评估了 DynDepNet 在生物性别分类任务上的性能 , 该任务是为 fMRI 数据设计的监督式深度学习模型的广泛使用的基准$\overline{\mathrm{E}}($Kim et al., 2021; Azevedo et al., 2022) 。基于脑成像数据的生物性别分类在神经科学中具有重要意义，因为已经表明，男性和女性在各种大脑特征和功能连接模式方面存在可观察到的差异 (Bell et al., 2006; Mao et al., 2017) 。通过评估 DynDepNet 在此任务上的性能，我们可以深入了解其捕获与性别相关的差异并利用它们进行准确分类的能力
5.1. Datasets
我们使用来自人类连接组计划(HCP)S1200 版本$^{2}$ $\boxed{\mathrm{E}}($Van Essen et al., 2013) 的公开可用的 fMRI 数据构建了两个数据集。这两个数据集根据受试者在功能磁共振成像采集期间是处于静息状态(HCP-Rest)还是执行特定任务 (HCP-Task)而不同
HCP-Rest 对于 HCP-Rest,我们考虑了静息态 fMRI 扫描，该扫描按照$\boxed{\mathrm{F}}$Glasser et al.(2013) 中描述的管道进行了最少的预处理。我们从使用左右(LR)相位编码的第一个扫描会话(四个)中选择了总共$N=1,095$ 个对象。在图像采集期间，受试者被指示休息 15 分钟。表示连续图像采集之间的时间的重复时间(TR) 设置为0.72秒，从而导致每个对象总共有$T^\prime=1,200$ 个图像。我们使用每个受试者的生物学性别作为标签， 总共产生了$C=2$类。女性受试者占数据集的$54.4\%$ HCP-Task 对于 HCP 任务，我们考虑了来自情感任务的任务 fMRI 扫描，该任务使用与 HCP-Rest 相同的管道进行了最少的预处理。我们从具有 LR 相位编码的第一个扫描会话(两个中的一个)中选择了总共$N=926$个对象。在任务期间，要求受试者指出屏幕底部呈现的两个面中的哪个或两个形状中的哪个与屏幕顶部的面或形状匹配。扫描会话持续了大约 2.11 分钟，TR 为 0.72 秒，导致每个对象的$T^\prime=176$图像。与 HCP-Rest 数据集类似，这些类别是根据受试者的生物学性别定义的。女性受试者占数据集的 51.2%
Further preprocessing 由于来自两个数据集的 fMRI 扫描都代表了 3D 脑体积的时间序列，因此我们将它们分割为$V=243$ 平均脑区域 (210 皮层，36 皮层下)
${^{2}\mathrm{https:}//\mathrm{db}}.$humanconnectome.org
BOLD 使用脑小体图谱$\mathbb{E}^3$ $\mathbb{E}($Fan et al.,2016)的长度为$T^{\prime}$时间点的信号。然后通过按区域标准化将每个时间序列转换为$z$分数，以消除振幅影响。为了平衡两个数据集中的班级比例，我们随机对少数班级的男性进行了过抽样
# 5.2. Baselines
我们将 DynDepNet 与一系列不同的基线模型进行了比较，这些模型以前曾用于对 fMRI 数据进行分类，并且其代码是公开可用的。基线是根据它们是作为输入静态 FC、动态 FC 还是 BOLD 信号进行广泛分组的。对于静态(线性)基线，我们包括核岭回归(KRR) E(He et al.,2020) 和支持向量机(SVM)$\boxtimes$E(Abraham et al., 2017)。对于静态深度学习基线，我们包括多层感知(MLP)和 BrainnetCNN (BNCNN) (Kawahara et al.,2017)。对于动态基线，我们包括 ST-GCN(STGCN) (Gadgil et al., 2020) 和 STAGIN (Kim et al., 2021)。对于 GSL 基线， 我们包括 FBNetGen(FBNG) (Kan et al., 2022) 和DeepfMRI(DFMRI) (Riaz et al.,2020) 。最后，我们包括一个双向 LSTM(BLSTM) (Hebling Vieira et al., 2021),它直接从 BOLD 信号中学习。有关每个基线模型的更多详细信息，请参见附录 A
## 5.3.Evaluation metrics
使用精度(ACC)和接收器操作员曲线下面积(AUROC) 在测试数据上评估模型性能，因为类别是平衡的。为了比较模型的性能，我们使用了由 Ulmer et al. (2022)实现的具有统计意义的几乎随机顺序 ( ASO) 检验$\boxed{\mathrm{F}}($Del Barrio et al., 2018; Dror et al., 2019)。对于所有测试，显著性水平设置为$\alpha=0.05$,并在进行多次比较时使用 Bonferroni 校正 (Bonferroni,1936) 进行调整
# 5.4. Implementation
这两个数据集分为训练、验证和测试数据集，保持$80/10/10\%$比例的类比例。为了确保分类性能的差异可以主要归因于模型体系结构的差异，删除了基线的特定培训和测试策略。相反，所有模型都是使用Adam 优化器$\overline{\mathbb{E}}($Kingma &Ba, 2014) 和去耦权重衰减$\boxed{\mathrm{E}}($Loshchilov & Hutter, 2017) 进行训练的。我们对所有模型进行了最多5，000个历元的训练，并根据在验证数据集上观察到的最低准确性，在 15 的耐心下提前停止。此策略有助于防止过度拟合，并使我们能够选择性能最佳的模型。为了减轻计算负荷并引人随机增强，对训练样本的时间维度进行了随机抽样。具体来说，对于 HCP-Rest 数据集，时间维度设置为$T^\prime=600$ ,而对于 HCP-Task 数据集，则按照 Kim et al.(2021) 使用的方法将其设置为$T^{\prime}=150$ 。为了解决训练过程中的可变性，每个模型都使用不同的随机种子和数据
$^3$https://atlas.brainnetome.org
5
|
[
[
0.358642578125,
0.058074951171875,
0.615234375,
0.058074951171875,
0.615234375,
0.07135009765625,
0.358642578125,
0.07135009765625
],
[
0.09234619140625,
0.08837890625,
0.1380615234375,
0.08837890625,
0.1380615234375,
0.0997314453125,
0.09234619140625,
0.0997314453125
],
[
0.50244140625,
0.08587646484375,
0.8837890625,
0.08587646484375,
0.8837890625,
0.0997314453125,
0.50244140625,
0.0997314453125
],
[
0.50341796875,
0.10162353515625,
0.88232421875,
0.10162353515625,
0.88232421875,
0.11553955078125,
0.50341796875,
0.11553955078125
],
[
0.10540771484375,
0.11553955078125,
0.455810546875,
0.11553955078125,
0.455810546875,
0.1287841796875,
0.10540771484375,
0.1287841796875
],
[
0.50244140625,
0.117431640625,
0.8857421875,
0.117431640625,
0.8857421875,
0.1307373046875,
0.50244140625,
0.1307373046875
],
[
0.214111328125,
0.1337890625,
0.453369140625,
0.1337890625,
0.453369140625,
0.147705078125,
0.214111328125,
0.147705078125
],
[
0.50244140625,
0.1319580078125,
0.8837890625,
0.1319580078125,
0.8837890625,
0.1458740234375,
0.50244140625,
0.1458740234375
],
[
0.443603515625,
0.149658203125,
0.47216796875,
0.149658203125,
0.47216796875,
0.1622314453125,
0.443603515625,
0.1622314453125
],
[
0.5009765625,
0.147705078125,
0.61669921875,
0.147705078125,
0.61669921875,
0.1610107421875,
0.5009765625,
0.1610107421875
],
[
0.10540771484375,
0.1787109375,
0.47216796875,
0.1787109375,
0.47216796875,
0.189453125,
0.10540771484375,
0.189453125
],
[
0.50244140625,
0.179931640625,
0.60302734375,
0.179931640625,
0.60302734375,
0.1912841796875,
0.50244140625,
0.1912841796875
],
[
0.09234619140625,
0.1932373046875,
0.188720703125,
0.1932373046875,
0.188720703125,
0.2052001953125,
0.09234619140625,
0.2052001953125
],
[
0.50341796875,
0.203857421875,
0.88232421875,
0.203857421875,
0.88232421875,
0.2139892578125,
0.50341796875,
0.2139892578125
],
[
0.5048828125,
0.2171630859375,
0.88232421875,
0.2171630859375,
0.88232421875,
0.2310791015625,
0.5048828125,
0.2310791015625
],
[
0.08660888671875,
0.2259521484375,
0.223876953125,
0.2266845703125,
0.223876953125,
0.2412109375,
0.08660888671875,
0.240478515625
],
[
0.5009765625,
0.2310791015625,
0.8837890625,
0.231689453125,
0.8837890625,
0.246826171875,
0.5009765625,
0.24560546875
],
[
0.0906982421875,
0.25244140625,
0.47216796875,
0.25244140625,
0.47216796875,
0.266357421875,
0.0906982421875,
0.266357421875
],
[
0.50244140625,
0.2474365234375,
0.8837890625,
0.2474365234375,
0.8837890625,
0.261474609375,
0.50244140625,
0.261474609375
],
[
0.0906982421875,
0.268310546875,
0.47216796875,
0.268310546875,
0.47216796875,
0.2822265625,
0.0906982421875,
0.2822265625
],
[
0.50244140625,
0.26318359375,
0.8857421875,
0.26318359375,
0.8857421875,
0.276611328125,
0.50244140625,
0.276611328125
],
[
0.0906982421875,
0.28271484375,
0.47216796875,
0.28271484375,
0.47216796875,
0.296630859375,
0.0906982421875,
0.296630859375
],
[
0.5009765625,
0.27783203125,
0.8857421875,
0.27783203125,
0.8857421875,
0.291748046875,
0.5009765625,
0.291748046875
],
[
0.08905029296875,
0.29736328125,
0.47216796875,
0.29736328125,
0.47216796875,
0.311279296875,
0.08905029296875,
0.311279296875
],
[
0.5009765625,
0.29345703125,
0.8857421875,
0.29345703125,
0.8857421875,
0.306884765625,
0.5009765625,
0.306884765625
],
[
0.08905029296875,
0.311279296875,
0.473876953125,
0.313232421875,
0.473876953125,
0.326904296875,
0.08905029296875,
0.324462890625
],
[
0.50244140625,
0.30810546875,
0.8857421875,
0.30810546875,
0.8857421875,
0.3212890625,
0.50244140625,
0.3212890625
],
[
0.08905029296875,
0.326904296875,
0.471435546875,
0.326904296875,
0.471435546875,
0.34033203125,
0.08905029296875,
0.34033203125
],
[
0.5,
0.3212890625,
0.8857421875,
0.323974609375,
0.8857421875,
0.33837890625,
0.5,
0.3359375
],
[
0.5009765625,
0.337158203125,
0.88232421875,
0.337158203125,
0.88232421875,
0.35107421875,
0.5009765625,
0.35107421875
],
[
0.08905029296875,
0.343505859375,
0.47216796875,
0.343505859375,
0.47216796875,
0.356689453125,
0.08905029296875,
0.356689453125
],
[
0.08905029296875,
0.35791015625,
0.473876953125,
0.35791015625,
0.473876953125,
0.371826171875,
0.08905029296875,
0.371826171875
],
[
0.5009765625,
0.352783203125,
0.88134765625,
0.352783203125,
0.88134765625,
0.3662109375,
0.5009765625,
0.3662109375
],
[
0.08905029296875,
0.373779296875,
0.398681640625,
0.373779296875,
0.398681640625,
0.386962890625,
0.08905029296875,
0.386962890625
],
[
0.50244140625,
0.36865234375,
0.8837890625,
0.36865234375,
0.8837890625,
0.3818359375,
0.50244140625,
0.3818359375
],
[
0.50244140625,
0.3837890625,
0.8837890625,
0.3837890625,
0.8837890625,
0.397705078125,
0.50244140625,
0.397705078125
],
[
0.50341796875,
0.3984375,
0.88232421875,
0.3984375,
0.88232421875,
0.412353515625,
0.50341796875,
0.412353515625
],
[
0.0906982421875,
0.406005859375,
0.1871337890625,
0.406005859375,
0.1871337890625,
0.417236328125,
0.0906982421875,
0.417236328125
],
[
0.5009765625,
0.412353515625,
0.8857421875,
0.41357421875,
0.8857421875,
0.427978515625,
0.5009765625,
0.4267578125
],
[
0.08905029296875,
0.427978515625,
0.473876953125,
0.4267578125,
0.473876953125,
0.44140625,
0.08905029296875,
0.442626953125
],
[
0.5009765625,
0.4287109375,
0.771484375,
0.4287109375,
0.771484375,
0.442626953125,
0.5009765625,
0.442626953125
],
[
0.09234619140625,
0.44580078125,
0.471435546875,
0.44580078125,
0.471435546875,
0.455810546875,
0.09234619140625,
0.455810546875
],
[
0.08905029296875,
0.458251953125,
0.473876953125,
0.458984375,
0.473876953125,
0.472900390625,
0.08905029296875,
0.4716796875
],
[
0.5009765625,
0.460205078125,
0.67236328125,
0.460205078125,
0.67236328125,
0.473388671875,
0.5009765625,
0.473388671875
],
[
0.0906982421875,
0.473388671875,
0.47216796875,
0.473388671875,
0.47216796875,
0.4873046875,
0.0906982421875,
0.4873046875
],
[
0.08905029296875,
0.4892578125,
0.3349609375,
0.4892578125,
0.3349609375,
0.50244140625,
0.08905029296875,
0.50244140625
],
[
0.50341796875,
0.484130859375,
0.88623046875,
0.484130859375,
0.88623046875,
0.49755859375,
0.50341796875,
0.49755859375
],
[
0.5,
0.49755859375,
0.8857421875,
0.5,
0.8857421875,
0.51318359375,
0.5,
0.51123046875
],
[
0.08905029296875,
0.51953125,
0.473876953125,
0.51953125,
0.473876953125,
0.53271484375,
0.08905029296875,
0.53271484375
],
[
0.5009765625,
0.5146484375,
0.88232421875,
0.5146484375,
0.88232421875,
0.52783203125,
0.5009765625,
0.52783203125
],
[
0.5,
0.52880859375,
0.8857421875,
0.52880859375,
0.8857421875,
0.54248046875,
0.5,
0.54248046875
],
[
0.0906982421875,
0.53515625,
0.473876953125,
0.53515625,
0.473876953125,
0.548828125,
0.0906982421875,
0.548828125
],
[
0.50244140625,
0.54443359375,
0.88232421875,
0.54443359375,
0.88232421875,
0.55810546875,
0.50244140625,
0.55810546875
],
[
0.0906982421875,
0.55029296875,
0.473876953125,
0.55029296875,
0.473876953125,
0.564453125,
0.0906982421875,
0.564453125
],
[
0.08905029296875,
0.56494140625,
0.47216796875,
0.564453125,
0.47216796875,
0.57763671875,
0.08905029296875,
0.5791015625
],
[
0.5009765625,
0.56005859375,
0.8857421875,
0.56005859375,
0.8857421875,
0.57373046875,
0.5009765625,
0.57373046875
],
[
0.0931396484375,
0.58203125,
0.471435546875,
0.58203125,
0.471435546875,
0.59228515625,
0.0931396484375,
0.59228515625
],
[
0.50244140625,
0.5771484375,
0.8837890625,
0.5771484375,
0.8837890625,
0.5869140625,
0.50244140625,
0.5869140625
],
[
0.0906982421875,
0.59521484375,
0.47216796875,
0.59521484375,
0.47216796875,
0.609375,
0.0906982421875,
0.609375
],
[
0.5009765625,
0.59130859375,
0.55126953125,
0.59130859375,
0.55126953125,
0.60302734375,
0.5009765625,
0.60302734375
],
[
0.0906982421875,
0.60986328125,
0.47216796875,
0.60986328125,
0.47216796875,
0.62353515625,
0.0906982421875,
0.62353515625
],
[
0.0906982421875,
0.62548828125,
0.47216796875,
0.62548828125,
0.47216796875,
0.6396484375,
0.0906982421875,
0.6396484375
],
[
0.5009765625,
0.6220703125,
0.650390625,
0.6220703125,
0.650390625,
0.63525390625,
0.5009765625,
0.63525390625
],
[
0.0906982421875,
0.6416015625,
0.468017578125,
0.6416015625,
0.468017578125,
0.65478515625,
0.0906982421875,
0.65478515625
],
[
0.5,
0.64306640625,
0.8837890625,
0.64453125,
0.8837890625,
0.6591796875,
0.5,
0.65771484375
],
[
0.50244140625,
0.66015625,
0.8837890625,
0.66015625,
0.8837890625,
0.673828125,
0.50244140625,
0.673828125
],
[
0.08905029296875,
0.67041015625,
0.473876953125,
0.67236328125,
0.473876953125,
0.68603515625,
0.08905029296875,
0.6845703125
],
[
0.5048828125,
0.6767578125,
0.8837890625,
0.6767578125,
0.8837890625,
0.6875,
0.5048828125,
0.6875
],
[
0.09234619140625,
0.6875,
0.47216796875,
0.6875,
0.47216796875,
0.70068359375,
0.09234619140625,
0.70068359375
],
[
0.50244140625,
0.6923828125,
0.88134765625,
0.6923828125,
0.88134765625,
0.70263671875,
0.50244140625,
0.70263671875
],
[
0.0906982421875,
0.70068359375,
0.471435546875,
0.70068359375,
0.471435546875,
0.7138671875,
0.0906982421875,
0.7138671875
],
[
0.498291015625,
0.7041015625,
0.8837890625,
0.7041015625,
0.8837890625,
0.71826171875,
0.498291015625,
0.71826171875
],
[
0.09234619140625,
0.71826171875,
0.47216796875,
0.71826171875,
0.47216796875,
0.72900390625,
0.09234619140625,
0.72900390625
],
[
0.50244140625,
0.72021484375,
0.8857421875,
0.72021484375,
0.8857421875,
0.734375,
0.50244140625,
0.734375
],
[
0.08905029296875,
0.73193359375,
0.473876953125,
0.73193359375,
0.473876953125,
0.74560546875,
0.08905029296875,
0.74560546875
],
[
0.5,
0.73486328125,
0.8837890625,
0.736328125,
0.8837890625,
0.75,
0.5,
0.74853515625
],
[
0.09234619140625,
0.74853515625,
0.47216796875,
0.74853515625,
0.47216796875,
0.7587890625,
0.09234619140625,
0.7587890625
],
[
0.50244140625,
0.75244140625,
0.88232421875,
0.75244140625,
0.88232421875,
0.76318359375,
0.50244140625,
0.76318359375
],
[
0.0906982421875,
0.76220703125,
0.473876953125,
0.76220703125,
0.473876953125,
0.77587890625,
0.0906982421875,
0.77587890625
],
[
0.50244140625,
0.7666015625,
0.8857421875,
0.7666015625,
0.8857421875,
0.77978515625,
0.50244140625,
0.77978515625
],
[
0.0906982421875,
0.7763671875,
0.4755859375,
0.7763671875,
0.4755859375,
0.79052734375,
0.0906982421875,
0.79052734375
],
[
0.50244140625,
0.78271484375,
0.88232421875,
0.78271484375,
0.88232421875,
0.79345703125,
0.50244140625,
0.79345703125
],
[
0.09234619140625,
0.79248046875,
0.473876953125,
0.79248046875,
0.473876953125,
0.80615234375,
0.09234619140625,
0.80615234375
],
[
0.5009765625,
0.796875,
0.8837890625,
0.796875,
0.8837890625,
0.81005859375,
0.5009765625,
0.81005859375
],
[
0.0906982421875,
0.80810546875,
0.44775390625,
0.80810546875,
0.44775390625,
0.8212890625,
0.0906982421875,
0.8212890625
],
[
0.50244140625,
0.8125,
0.8837890625,
0.8125,
0.8837890625,
0.82275390625,
0.50244140625,
0.82275390625
],
[
0.5146484375,
0.82568359375,
0.8837890625,
0.82568359375,
0.8837890625,
0.8388671875,
0.5146484375,
0.8388671875
],
[
0.0906982421875,
0.83837890625,
0.473876953125,
0.83837890625,
0.473876953125,
0.85302734375,
0.0906982421875,
0.85302734375
],
[
0.50244140625,
0.84130859375,
0.8857421875,
0.84130859375,
0.8857421875,
0.85498046875,
0.50244140625,
0.85498046875
],
[
0.0906982421875,
0.85400390625,
0.47216796875,
0.85400390625,
0.47216796875,
0.8671875,
0.0906982421875,
0.8671875
],
[
0.50244140625,
0.85791015625,
0.88232421875,
0.85791015625,
0.88232421875,
0.86865234375,
0.50244140625,
0.86865234375
],
[
0.09234619140625,
0.87060546875,
0.471435546875,
0.87060546875,
0.471435546875,
0.880859375,
0.09234619140625,
0.880859375
],
[
0.51953125,
0.87890625,
0.7255859375,
0.880859375,
0.7255859375,
0.89599609375,
0.51953125,
0.89306640625
],
[
0.108642578125,
0.8916015625,
0.339111328125,
0.89306640625,
0.339111328125,
0.90771484375,
0.108642578125,
0.90576171875
],
[
0.4794921875,
0.92236328125,
0.495849609375,
0.92236328125,
0.495849609375,
0.9375,
0.4794921875,
0.9375
]
] |
[
"Title Suppressed Due to Excessive Size",
"函数是",
"BOLD使用脑小体图谱E3E(Fanetal.2016）的长",
"度为T时间点的信号。然后通过按区域标准化将每个",
"C(y, {9,G1:T)) =LcE(y,9)+AFsLFs(Fi:T,A1:T)",
"时间序列转换为分数，以消除振幅影响。为了平衡",
"+TsLTs(A1:T) +AsPLsP(Ai:T)",
"两个数据集中的班级比例，我们随机对少数班级的男",
"(19)",
"性进行了过抽样",
"其中^FS，入TS，ASP≥0是加权每个正则化项的相对",
"5.2.Baselines",
"贡献的超参数",
"我们将DynDepNet与一系列不同的基线模型进行",
"了比较，这些模型以前曾用于对fMRI数据进行分",
"5.Experiments",
"类，并且其代码是公开可用的。基线是根据它们是",
"我们评估了DynDepNet在生物性别分类任务上的性",
"作为输人静态FC、动态FC还是BOLD信号进行",
"能，该任务是为fMRI数据设计的监督式深度学习模型",
"广泛分组的。对于静态（线性）基线，我们包括核",
"的广泛使用的基准E(Kimetal.，2021；Azevedoetal.",
"岭回归（KRR）E(Heetal.，2020）和支持向量机",
"2022）。基于脑成像数据的生物性别分类在神经科学中",
"(SVM）E(Abrahametal.，2017）。对于静态深度学",
"具有重要意义，因为已经表明，男性和女性在各种大脑",
"习基线，我们包括多层感知（MLP）和BrainmetCNN",
"特征和功能连接模式方面存在可观察到的差异（Bell",
"(BNCNN）（Kawahara etal.，2017）。对于动态基",
"线，我们包括ST-GCN(STGCN）（Gadgiletal.,2020)",
"etal.,2006;Maoetal.,2017）。通过评估DynDepNet",
"在此任务上的性能，我们可以深人了解其捕获与性别",
"和STAGIN（Kimetal.,2021）。对于GSL基线",
"相关的差异并利用它们进行准确分类的能力",
"我们包括FBNetGen(FBNG）（Kanetal.,2022）和",
"DeepfMRI(DFMRI）（Riazetal.，2020）。最后，我们",
"包括一个双向LSTM(BLSTM）（HeblingVieiraetal.",
"5.1. Datasets",
"2021），它直接从BOLD信号中学习。有关每个基线",
"我们使用来自人类连接组计划（HCP）S1200版本2",
"模型的更多详细信息，请参见附录A",
"F（VanEssenetal.，2013）的公开可用的fMRI数据构",
"建了两个数据集。这两个数据集根据受试者在功能磁",
"5.3. Evaluation metrics",
"共振成像采集期间是处于静息状态（HCP-Rest）还是",
"执行特定任务（HCP-Task）而不同",
"使用精度（ACC）和接收器操作员曲线下面积（AU-",
"ROC）在测试数据上评估模型性能，因为类别是平衡",
"HCP-Rest对于HCP-Rest，我们考虑了静息态fMRI",
"的。为了比较模型的性能，我们使用了由Ulmeretal",
"（2022）实现的具有统计意义的几乎随机顺序（ASO）检",
"扫描，该扫描按照EGlasseretal.（2013）中描述的管",
"验E(Del Barrio et al.,2018;Dror etal.,2019)。对于",
"道进行了最少的预处理。我们从使用左右（LR）相位编",
"码的第一个扫描会话（四个）中选择了总共N=1,095",
"所有测试，显著性水平设置为α=0.05，并在进行多",
"个对象。在图像采集期间：受试者被指示休息15分",
"次比较时使用Bonferroni校正（Bonferroni，1936）进",
"钟。表示连续图像采集之间的时间的重复时间（TR）",
"行调整",
"设置为0.72秒，从而导致每个对象总共有T=1,200",
"个图像。我们使用每个受试者的生物学性别作为标签",
"5.4.Implementation",
"总共产生了C=2类。女性受试者占数据集的54.4%",
"这两个数据集分为训练、验证和测试数据集，保持",
"80/10/10%比例的类比例。为了确保分类性能的差异",
"HCP-Task对于HCP任务，我们考虑了来自情感任",
"可以主要归因于模型体系结构的差异，删除了基线",
"务的任务fMRI扫描，该任务使用与HCP-Rest相同",
"的特定培训和测试策略。相反，所有模型都是使用",
"的管道进行了最少的预处理。我们从具有LR相位编",
"Adam优化器E(Kingma&Ba，2014）和去耦权重衰",
"码的第一个扫描会话（两个中的一个）中选择了总共",
"减ELoshchilov&Hutter，2017）进行训练的。我们对",
"N=926个对象。在任务期间，要求受试者指出屏幕",
"所有模型进行了最多5,000个历元的训练，并根据在",
"底部呈现的两个面中的哪个或两个形状中的哪个与屏",
"验证数据集上观察到的最低准确性，在15的耐心下提",
"幕顶部的面或形状匹配。扫描会话持续了大约2.11分",
"前停止。此策略有助于防止过度拟合，并使我们能够",
"钟，TR为0.72秒，导致每个对象的T'=176图像。",
"选择性能最佳的模型。为了减轻计算负荷并引人随机",
"与HCP-Rest数据集类似，这些类别是根据受试者的",
"增强，对训练样本的时间维度进行了随机抽样。具体来",
"生物学性别定义的。女性受试者占数据集的51.2%",
"说，对于HCP-Rest数据集，时间维度设置为T/=600",
"而对于HCP-Task数据集，则按照Kimetal.（2021)",
"Furtherpreprocessing由于来自两个数据集的fMRI",
"使用的方法将其设置为T’=150。为了解决训练过程",
"扫描都代表了3D脑体积的时间序列，因此我们将它",
"中的可变性，每个模型都使用不同的随机种子和数据",
"们分割为V=243平均脑区域（210皮层，36皮层下",
"3https://atlas.brainnetome.org",
"2https://db.humanconnectome.org",
"5"
] |
{
"arxiv_id": "2209.13513",
"language": "zh",
"page_idx": 4
}
|
||
2210.13439-zh-1
|
2210.13439-zh-1.jpg
|
级联偏差：研究启发式注释策略对数据和模型的影响
Chaitanya Malaviya and Sudeep Bhatia and Mark Yatskar
University of Pennsylvania
$\{$ cmalaviy, bhatiasu, myatskar $\}$ @upenn.edu
## $\mathbf{Abstract}$
Cognitive psychologists have documented that humans use cognitive heuristics, or mental shortcuts, to make quick decisions while expending less effort. While performing annotation work on crowdsourcing platforms, we hypothesize that such heuristic use among annotators cascades on to data quality and model robustness. In this work, we study cognitive heuristic use in the context of annotating multiple-choice reading comprehension datasets. We propose tracking annotator heuristic traces , where we tangibly measure low-effort annotation strategies that could indicate usage of various cognitive heuristics. We find evidence that annotators might be using multiple such heuristics, based on correlations with a battery of psychological tests. Importantly, heuristic use among annotators determines data quality along several dimensions: (1) known biased models, such as partial input models, more easily solve examples authored by annotators that rate highly on heuristic use, (2) models trained on annotators scoring highly on heuristic use don't generalize as well, and (3) heuristic-seeking annotators tend to create qualitatively less challenging examples. Our fndings suggest that tracking heuristic usage among annotators can potentially help with collecting challenging datasets and diagnosing model biases.
### 1 $\textbf{Introduction}$
虽然众包是 NLP 中一种有效且广泛使用的数据收集方法，但它带有警告。已发现众包数据集包含来自注释过程的工件，并且在此类数据上训练的模型可能很脆弱，无法推广到分布偏移 (Gururangan et al., 2018; Kaushik and Lipton, 2018; McCoy et al.,2019)。在这项工作中，我们询问注释者行为中的系统模式是否会影响所收集数据的质量
我们假设使用认知启发式方法，这是人类在日常生活中使用的心理捷径，可以级联到数据质量和模型鲁棒性。例如，要求根据段落编
写问题的注释者可能不会阅读整个段落，或者可能只使用一个句子来构建问题。注释者可能会寻求捷径，以节省他们投入任务的时间和精力。注释器中的这种行为，其特征是可以接受但质量不高的示例，可能会有问题
我们通过以注释者启发式 traces 的形式从注释数据中跟踪指示性特征，来分析注释者参与各种低工作量策略的程度，类似于认知启发式。首先，我们众包阅读理解问题，我们指导工人写硬问题。受人类认知(Simon, 1956; Tversky and Kahneman, 1974)研究的启发，我们确定了注释者可以用于我们任务的几种启发式方法，例如满足(Simon,1956),可用性(Tversky and Kahneman, 1973) 和代表性 (Kahneman and Tversky,1972)。我们通过对收集的数据和注释元数据进行特征化来衡量它们的潜在使用情况(例如。g。,花费的时间和输入的击键)( §4 )。此外，我们确定了这些启发式的实例，这些实例与测量人类启发式思维倾向的心理测试密切相关，例如认知反思测试E|Frederick,2005; Toplak et al., 2014; Sirota et al., 2021)。在注释过程中，我们在心理上合理的启发式使用措施可以汇总到每个注释者，形成他们产生的数据的整体摘要
根据这些统计数据，我们分析了从事不同层次的启发式使用的注释者创建的示例之间的差异。我们的第一个发现是，由强启发式寻求注释器创建的示例也更容易使用启发式方法(§5 ) 求解模型。我们评估了利用一些已知偏见的模型，发现使用认知启发式的注释者的例子更容易被有偏见的模型解决。我们还研究了启发式算法对训练模型的影响。以前的工作 (Geva et al.,2019) 表明，当数据集被注释器随机拆分时，模型的泛化能力很差，这可能是由于存在工件。我们复制了这个结果，发现当在启发式寻求注释者的例子上训练时，模型的泛化程度甚至更糟
为了了解注释管道的哪些部分有助于注释器中的启发式搜索行为，我们还梳理了任务固有组件的影响(例如。g。,通过难度),而不是注释 者 本 身 ( e。g。, 注 释 器 疲 劳 ) ( § 6 ) 。不 幸 的
|
[
[
0.22412109375,
0.0968017578125,
0.77490234375,
0.0968017578125,
0.77490234375,
0.10986328125,
0.22412109375,
0.10986328125
],
[
0.2435302734375,
0.1455078125,
0.76171875,
0.1455078125,
0.76171875,
0.1602783203125,
0.2435302734375,
0.1602783203125
],
[
0.391357421875,
0.160888671875,
0.61279296875,
0.162109375,
0.61279296875,
0.177001953125,
0.391357421875,
0.17578125
],
[
0.2476806640625,
0.179931640625,
0.767578125,
0.179931640625,
0.767578125,
0.19482421875,
0.2476806640625,
0.19482421875
],
[
0.2626953125,
0.254150390625,
0.341796875,
0.254150390625,
0.341796875,
0.268310546875,
0.2626953125,
0.268310546875
],
[
0.51123046875,
0.254638671875,
0.87890625,
0.254638671875,
0.87890625,
0.268310546875,
0.51123046875,
0.268310546875
],
[
0.5146484375,
0.270751953125,
0.880859375,
0.270751953125,
0.880859375,
0.283935546875,
0.5146484375,
0.283935546875
],
[
0.145263671875,
0.27978515625,
0.456787109375,
0.27978515625,
0.456787109375,
0.29345703125,
0.145263671875,
0.29345703125
],
[
0.5146484375,
0.287353515625,
0.87890625,
0.287353515625,
0.87890625,
0.300537109375,
0.5146484375,
0.300537109375
],
[
0.145263671875,
0.295654296875,
0.45751953125,
0.295654296875,
0.45751953125,
0.308837890625,
0.145263671875,
0.308837890625
],
[
0.51318359375,
0.3017578125,
0.87890625,
0.3017578125,
0.87890625,
0.3154296875,
0.51318359375,
0.3154296875
],
[
0.14697265625,
0.308837890625,
0.45751953125,
0.308837890625,
0.45751953125,
0.322509765625,
0.14697265625,
0.322509765625
],
[
0.145263671875,
0.32373046875,
0.45751953125,
0.32373046875,
0.45751953125,
0.33740234375,
0.145263671875,
0.33740234375
],
[
0.51318359375,
0.319580078125,
0.78857421875,
0.319580078125,
0.78857421875,
0.333251953125,
0.51318359375,
0.333251953125
],
[
0.14697265625,
0.33837890625,
0.456787109375,
0.33837890625,
0.456787109375,
0.3515625,
0.14697265625,
0.3515625
],
[
0.53076171875,
0.33740234375,
0.880859375,
0.33740234375,
0.880859375,
0.350830078125,
0.53076171875,
0.350830078125
],
[
0.145263671875,
0.350830078125,
0.456787109375,
0.3515625,
0.456787109375,
0.365234375,
0.145263671875,
0.364013671875
],
[
0.5146484375,
0.355224609375,
0.87890625,
0.355224609375,
0.87890625,
0.365234375,
0.5146484375,
0.365234375
],
[
0.1485595703125,
0.36767578125,
0.456787109375,
0.36767578125,
0.456787109375,
0.377685546875,
0.1485595703125,
0.377685546875
],
[
0.5146484375,
0.369384765625,
0.880859375,
0.369384765625,
0.880859375,
0.383056640625,
0.5146484375,
0.383056640625
],
[
0.145263671875,
0.380126953125,
0.45751953125,
0.380126953125,
0.45751953125,
0.393798828125,
0.145263671875,
0.393798828125
],
[
0.51318359375,
0.385498046875,
0.87890625,
0.385498046875,
0.87890625,
0.3984375,
0.51318359375,
0.3984375
],
[
0.145263671875,
0.394287109375,
0.459228515625,
0.394287109375,
0.459228515625,
0.407958984375,
0.145263671875,
0.407958984375
],
[
0.51123046875,
0.402099609375,
0.87890625,
0.402099609375,
0.87890625,
0.4150390625,
0.51123046875,
0.4150390625
],
[
0.14697265625,
0.4091796875,
0.45751953125,
0.4091796875,
0.45751953125,
0.4228515625,
0.14697265625,
0.4228515625
],
[
0.145263671875,
0.42333984375,
0.45751953125,
0.42333984375,
0.45751953125,
0.43701171875,
0.145263671875,
0.43701171875
],
[
0.51123046875,
0.418701171875,
0.87744140625,
0.418701171875,
0.87744140625,
0.431640625,
0.51123046875,
0.431640625
],
[
0.1485595703125,
0.439453125,
0.455078125,
0.439453125,
0.455078125,
0.449462890625,
0.1485595703125,
0.449462890625
],
[
0.51708984375,
0.435791015625,
0.87744140625,
0.435791015625,
0.87744140625,
0.446533203125,
0.51708984375,
0.446533203125
],
[
0.145263671875,
0.45068359375,
0.45751953125,
0.45068359375,
0.45751953125,
0.46435546875,
0.145263671875,
0.46435546875
],
[
0.51123046875,
0.45068359375,
0.8759765625,
0.45068359375,
0.8759765625,
0.46435546875,
0.51123046875,
0.46435546875
],
[
0.14697265625,
0.465576171875,
0.45751953125,
0.465576171875,
0.45751953125,
0.478515625,
0.14697265625,
0.478515625
],
[
0.51318359375,
0.466064453125,
0.880859375,
0.466064453125,
0.880859375,
0.479736328125,
0.51318359375,
0.479736328125
],
[
0.14697265625,
0.479736328125,
0.45751953125,
0.479736328125,
0.45751953125,
0.493408203125,
0.14697265625,
0.493408203125
],
[
0.51123046875,
0.482177734375,
0.880859375,
0.482177734375,
0.880859375,
0.495361328125,
0.51123046875,
0.495361328125
],
[
0.14697265625,
0.493408203125,
0.456787109375,
0.493408203125,
0.456787109375,
0.50732421875,
0.14697265625,
0.50732421875
],
[
0.51123046875,
0.498779296875,
0.87890625,
0.498779296875,
0.87890625,
0.5126953125,
0.51123046875,
0.5126953125
],
[
0.14697265625,
0.50830078125,
0.456787109375,
0.50830078125,
0.456787109375,
0.521484375,
0.14697265625,
0.521484375
],
[
0.51123046875,
0.51416015625,
0.8818359375,
0.51416015625,
0.8818359375,
0.529296875,
0.51123046875,
0.529296875
],
[
0.14697265625,
0.521484375,
0.45751953125,
0.521484375,
0.45751953125,
0.5361328125,
0.14697265625,
0.5361328125
],
[
0.509765625,
0.52978515625,
0.8818359375,
0.529296875,
0.8818359375,
0.54345703125,
0.509765625,
0.54443359375
],
[
0.14697265625,
0.5361328125,
0.45751953125,
0.5361328125,
0.45751953125,
0.5498046875,
0.14697265625,
0.5498046875
],
[
0.145263671875,
0.55029296875,
0.459228515625,
0.55029296875,
0.459228515625,
0.56396484375,
0.145263671875,
0.56396484375
],
[
0.51318359375,
0.54736328125,
0.880859375,
0.54736328125,
0.880859375,
0.56103515625,
0.51318359375,
0.56103515625
],
[
0.14697265625,
0.5654296875,
0.45751953125,
0.5654296875,
0.45751953125,
0.5791015625,
0.14697265625,
0.5791015625
],
[
0.5146484375,
0.5634765625,
0.880859375,
0.5634765625,
0.880859375,
0.57666015625,
0.5146484375,
0.57666015625
],
[
0.1485595703125,
0.58056640625,
0.456787109375,
0.58056640625,
0.456787109375,
0.59130859375,
0.1485595703125,
0.59130859375
],
[
0.51123046875,
0.57763671875,
0.880859375,
0.57763671875,
0.880859375,
0.59130859375,
0.51123046875,
0.59130859375
],
[
0.1485595703125,
0.59423828125,
0.456787109375,
0.59423828125,
0.456787109375,
0.60498046875,
0.1485595703125,
0.60498046875
],
[
0.51318359375,
0.59423828125,
0.8818359375,
0.59423828125,
0.8818359375,
0.60791015625,
0.51318359375,
0.60791015625
],
[
0.14697265625,
0.609375,
0.455078125,
0.609375,
0.455078125,
0.619140625,
0.14697265625,
0.619140625
],
[
0.51318359375,
0.60986328125,
0.87890625,
0.60986328125,
0.87890625,
0.62353515625,
0.51318359375,
0.62353515625
],
[
0.14697265625,
0.62158203125,
0.456787109375,
0.62158203125,
0.456787109375,
0.63525390625,
0.14697265625,
0.63525390625
],
[
0.5146484375,
0.62744140625,
0.603515625,
0.62744140625,
0.603515625,
0.6396484375,
0.5146484375,
0.6396484375
],
[
0.145263671875,
0.63525390625,
0.45751953125,
0.63525390625,
0.45751953125,
0.6484375,
0.145263671875,
0.6484375
],
[
0.1435546875,
0.6484375,
0.45751953125,
0.6494140625,
0.45751953125,
0.66455078125,
0.1435546875,
0.6630859375
],
[
0.53076171875,
0.64453125,
0.880859375,
0.64453125,
0.880859375,
0.65771484375,
0.53076171875,
0.65771484375
],
[
0.145263671875,
0.66455078125,
0.45751953125,
0.66455078125,
0.45751953125,
0.677734375,
0.145263671875,
0.677734375
],
[
0.51318359375,
0.6611328125,
0.880859375,
0.6611328125,
0.880859375,
0.6748046875,
0.51318359375,
0.6748046875
],
[
0.145263671875,
0.6787109375,
0.45751953125,
0.6787109375,
0.45751953125,
0.6923828125,
0.145263671875,
0.6923828125
],
[
0.51318359375,
0.677734375,
0.880859375,
0.677734375,
0.880859375,
0.69140625,
0.51318359375,
0.69140625
],
[
0.1435546875,
0.69140625,
0.344970703125,
0.6923828125,
0.344970703125,
0.70703125,
0.1435546875,
0.7060546875
],
[
0.51123046875,
0.6923828125,
0.87890625,
0.6923828125,
0.87890625,
0.7060546875,
0.51123046875,
0.7060546875
],
[
0.51318359375,
0.71044921875,
0.880859375,
0.71044921875,
0.880859375,
0.7236328125,
0.51318359375,
0.7236328125
],
[
0.1158447265625,
0.7197265625,
0.145263671875,
0.7197265625,
0.145263671875,
0.73095703125,
0.1158447265625,
0.73095703125
],
[
0.151123046875,
0.7197265625,
0.26123046875,
0.7197265625,
0.26123046875,
0.73095703125,
0.151123046875,
0.73095703125
],
[
0.51318359375,
0.7255859375,
0.880859375,
0.7255859375,
0.880859375,
0.7392578125,
0.51318359375,
0.7392578125
],
[
0.1175537109375,
0.7421875,
0.483642578125,
0.7421875,
0.483642578125,
0.755859375,
0.1175537109375,
0.755859375
],
[
0.51318359375,
0.7421875,
0.880859375,
0.7421875,
0.880859375,
0.755859375,
0.51318359375,
0.755859375
],
[
0.118408203125,
0.76025390625,
0.486083984375,
0.76025390625,
0.486083984375,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.51318359375,
0.75830078125,
0.880859375,
0.75830078125,
0.880859375,
0.771484375,
0.51318359375,
0.771484375
],
[
0.118408203125,
0.77685546875,
0.483642578125,
0.77685546875,
0.483642578125,
0.78759765625,
0.118408203125,
0.78759765625
],
[
0.51123046875,
0.77392578125,
0.87890625,
0.77392578125,
0.87890625,
0.78759765625,
0.51123046875,
0.78759765625
],
[
0.118408203125,
0.7919921875,
0.486083984375,
0.7919921875,
0.486083984375,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.51123046875,
0.7890625,
0.880859375,
0.7890625,
0.880859375,
0.802734375,
0.51123046875,
0.802734375
],
[
0.118408203125,
0.80810546875,
0.484375,
0.80810546875,
0.484375,
0.822265625,
0.118408203125,
0.822265625
],
[
0.51318359375,
0.80810546875,
0.87890625,
0.80810546875,
0.87890625,
0.818359375,
0.51318359375,
0.818359375
],
[
0.118408203125,
0.82470703125,
0.486083984375,
0.82470703125,
0.486083984375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.51123046875,
0.822265625,
0.880859375,
0.822265625,
0.880859375,
0.8359375,
0.51123046875,
0.8359375
],
[
0.118408203125,
0.84130859375,
0.484375,
0.84130859375,
0.484375,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.51318359375,
0.84033203125,
0.58837890625,
0.84033203125,
0.58837890625,
0.8515625,
0.51318359375,
0.8515625
],
[
0.1175537109375,
0.85693359375,
0.2476806640625,
0.85693359375,
0.2476806640625,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.53076171875,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.53076171875,
0.87060546875
],
[
0.1368408203125,
0.87255859375,
0.48779296875,
0.87255859375,
0.48779296875,
0.8857421875,
0.1368408203125,
0.8857421875
],
[
0.51318359375,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.88720703125,
0.51318359375,
0.88720703125
],
[
0.118408203125,
0.8896484375,
0.486083984375,
0.8896484375,
0.486083984375,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51123046875,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90380859375,
0.51123046875,
0.90380859375
],
[
0.1175537109375,
0.90380859375,
0.48779296875,
0.90478515625,
0.48779296875,
0.91943359375,
0.1175537109375,
0.91845703125
],
[
0.51318359375,
0.90625,
0.880859375,
0.90625,
0.880859375,
0.91943359375,
0.51318359375,
0.91943359375
]
] |
[
"级联偏：研究启发式注释策略对数据和模型的影响",
"Chaitanya Malaviya and Sudeep Bhatia and Mark Yatskar",
"University of Pennsylvania",
"( cmalaviy, bhatiasu, myatskar @upenn.edu",
"Abstract",
"写问题的注释者可能不会阅读整个段落，或者",
"可能只使用一个句子来构建问题。注释者可能",
"Cognitive psychologists have documented that",
"会寻求捷径，以节省他们投入任务的时间和精",
"humans use cognitive heuristics, or mental",
"力。注释器中的这种行为，其特征是可以接受",
"shortcuts, to make quick decisions while ex",
"pending less effort.While performing an-",
"但质量不高的示例，可能会有问题",
"notation work on crowdsourcing platforms,",
"我们通过以注释者启发式traces的形式从注",
"we hypothesize that such heuristic use among",
"释数据中跟踪指示性特征，来分析注释者参与",
"annotators cascadeson todata quality and",
"各种低工作量策略的程度，类似于认知启发式。",
"model robustness. In this work, we study",
"首先，我们众包阅读理解问题，我们指导工人",
"cognitive heuristic use in the context of an-",
"写硬问题。受人类认知（Simon，1956；Tversky",
"notating multiple-choice reading comprehen",
"sion datasets. We propose tracking annotator",
"andKahneman，1974）研究的启发，我们确定",
"heuristictraces,wherewetangiblymeasure",
"了注释者可以用于我们任务的几种启发式方",
"low-effort annotation strategies that could indi-",
"法，例如满足（Simon，1956），可用性（Tversk",
"cate usage of various cognitive heuristics. We",
"andKahneman，1973）和代表性（Kahneman and",
"find evidence that annotators might be using",
"Tversky，1972）。我们通过对收集的数据和注",
"multiple such heuristics, based on correlations",
"释元数据进行特征化来衡量它们的潜在使用情",
"with a battery of psychological tests. Impor",
"况（例如。g。，花费的时间和输入的击键)s4",
"tantly, heuristic use among annotators deter-",
"）。此外，我们确定了这些启发式的实例，这些",
"mines data quality along several dimensions:",
"(1) known biased models, such as partial in-",
"实例与测量人类启发式思维倾向的心理测试密",
"put models, more easily solve examples au",
"切相关，例如认知反思测试E(Frederick，2005；",
"thoredbyannotatorsthatratehighlyonheuris-",
"Toplak et al., 2014; Sirota et al., 2021)。在注释",
"ticuse,(2)models trained on annotators scor",
"过程中，我们在心理上合理的启发式使用措施",
"ing highly on heuristic use don't generalize",
"可以汇总到每个注释者，形成他们产生的数据",
"as well, and (3) heuristic-seeking annotators",
"的整体摘要",
"tend to create qualitatively less challenging ex-",
"amples. Our findings suggest that tracking",
"根据这些统计数据，我们分析了从事不同层",
"heuristic usage among annotators can poten-",
"次的启发式使用的注释者创建的示例之间的差",
"tially help with collecting challenging datasets",
"异。我们的第一个发现是，由强启发式寻求注",
"and diagnosing model biases.",
"释器创建的示例也更容易使用启发式方法（S5",
"）求解模型。我们评估了利用一些已知偏见的",
"1",
"Introduction",
"模型，发现使用认知启发式的注释者的例子更",
"虽然众包是NLP中一种有效且广泛使用的数",
"容易被有偏见的模型解决。我们还研究了启发",
"据收集方法，但它带有警告。已发现众包数据",
"式算法对训练模型的影响。以前的工作（Geva",
"集包含来自注释过程的工件，并月在此类数据",
"etal.,2019)表明，当数据集被注释器随机拆分",
"上训练的模型可能很脆弱，无法推广到分布偏",
"时，模型的泛化能力很差，这可能是由于存在",
"移 (Gururangan et al., 2018; Kaushik and Lipton,",
"工件。我们复制了这个结果，发现当在启发式",
"2018；McCoyetal，2019）。在这项工作中，我",
"寻求注释者的例子上训练时，模型的泛化程度",
"们询问注释者行为中的系统模式是否会影响所",
"甚至更糟",
"收集数据的质量",
"为了了解注释管道的哪些部分有助于注释器",
"我们假设使用认知启发式方法，这是人类在",
"中的启发式搜索行为，我们还梳理了任务固有",
"日常生活中使用的心理捷径，可以级联到数",
"组件的影响（例如。g。，通过难度)，而不是注",
"据质量和模型鲁棒性。例如，要求根据段落编",
"释者本身(e。g。，注释器疲劳)（86)。不幸的"
] |
{
"arxiv_id": "2210.13439",
"language": "zh",
"page_idx": 1
}
|
||
2210.13439-zh-2
|
2210.13439-zh-2.jpg
|
件或偏差在 NLP 中得到了很好的记录。这些工件允许模型使用不可靠的快捷方式 (Geirhos et al., 2020) 解决 NLP 问题。几位研究人员提出了针对已知偏差实现鲁棒性的方法。我们请读者参考 Wang et al.(2022),以全面回顾这些方法。定位未知的偏见仍然是一个挑战，我们的工作可以通过识别启发式寻求注释者来帮助找到可能包含工件的示例
先前的工作提出了使用模型预测(Lertvit-) tayakumjorn and Toni, 2021)的解释来发现快捷方式的方法，包括基于样本的解释(Han et al., 2020) 和输入特征属性 (Bastings et al., 2021; Pezeshkpour et al., 2022)。其他有助于诊断模型偏差的技术包括构建测试用例 (Ribeiro et al., 2020; Ribeiro and Lundberg, 2022)的清单，构建对比(Gardner et al.,2020)或反事实(Wu et al., 2021)示例以及统计测试(Gururangan et al., 2018; Gardner et al., 2021)。我们的工作是对这些方法的补充，因为我们提供了一种与注释器相关的偏见发现的替代方法
是，我们没有发现简单的预测因子(i。e。,段
落难度) 注释者何时可能使用启发式
对收集的数据进行定性分析后发现，寻求启发式的注释器更有可能创建无效的示例，并且需要在明确声明的信息上进行更简单的单词匹配(§7 )。至关重要的是，这表明对启发式使用的测量，例如本文中研究的那些，可以提供一种通用方法来查找众包数据中不可靠的示例， 并指导我们在数据中发现工件。由于我们在健壮性和数据质量方面暗示了启发式使用，因此我们建议未来的数据集创建者跟踪类似的功能并评估模型对注释器启发式使用的敏感性。
# 2 Background and Related Work
Cognitive Heuristics. 人类判断，决策和推理中的启发式研究是 (Simon, 1956; Tversky and Kahneman, 1974) 研究的热门且有影响力的主题。启发式可以定义为心理捷径，我们在日常任务中用于快速决策。例如，Tversky and Kahneman (1974) 询问参与者是否有更多的英语单词以字母 K 开头或包含 K 作为$3^{rd}$字母， 超过 70 名% 参与者选择了前者，因为以 K 开头的单词更容易被召回，尽管这是不正确的。这是可用性启发式的一个示例。系统地使用这种启发式方法会导致认知偏差，这是我们思维中的非理性模式
乍一看，似乎启发式算法总是次优的，但是以前的工作认为，与优化(Gigerenzer and Gaissmaier, 2011)相比，启发式算法可以在不确定的情况下导致准确的推论。我们假设启发式方法在确定数据质量方面可以发挥相当大的作用，其影响取决于启发式方法的确切性质。先前的工作表明，在相关性判断任务(Eickhoff,2018)中，众筹人员容易受到认知偏差的影响，并提供了一个清单来对抗这些偏差(Draws et al.,2021)。相比之下，我们的工作重点是如何在写作任务中测量这种启发式方法的潜在使用，并提供了启发式使用与模型脆性相关的证据
注释器行为的特征以前在估计注释器任务准确性 (Rzeszotarski and Kittur, 2011;Goyal et al., 2018) 时很有用。还发现注释者身份会影响其注释 (Hube et al., 2019; Sap et al., 2022)。我们的工作建立在这些结果的基础上，并估计启发式使用功能来捕获有关数据质量的隐含线索
Mitigating and discovering biases. 在自然语
言推理，问题回答和参数理解(Gururangan et al., 2018; McCoy et al., 2019; Niven and Kao 2019, inter alia )等任务中，数据集中存在工
$^{1}$Our code and collected data is available at https://github.com/chaitanyamalaviya/annotator-heuristics .
Improved crowdsourcing. 相关工作研究了对众包协议的修改，以提高数据质量(Bowman et al., 2020; Nangia et al., 2021)。此外，诸如对抗性数据收集 (Nie et al., 2020) 和使用生成模型 (Bartolo et al., 2022; Liu et al., 2022) 之类的循环模型众包方法已被证明有助于创建更具挑战性的示例。我们相信跟踪注释器启发式使用可以帮助对众包协议进行明智的调整
## 3 $\textbf{Annotation Protocol}$
我们将多选题阅读理解视为我们的众包任务， 因为我们可以从注释者那里获得丰富的响应和交互，这使我们能够探索一系列假设的启发式方法。我们在这里描述数据收集的方法
我们在 AmazonMechanicalTurk 上提供了带有段落的注释器，并要求他们编写带有四个选项的多项选择题。我们使用了英文维基百科$^{2}$ 中“重要文章”的第一段，并确保段落至少有50个单词，最多有 250个单词。段落跨越 11 种类型，包括艺术，历史，物理科学等，并从10K 段落中随机抽取段落。Annotatorswereaskedtowritechallengingquestionsthatcannotbean-
sweredbyreadingjustthequestionorpassagealone,andhaveasingleco
此外，他们被要求确保段落提供足够的信息来回答问题，同时允许问题需要使用常识或因果关系进行基本推断
注释器首先有资格避免垃圾邮件行为。此资格以无效问题的形式检查垃圾邮件行为，而不
$^{2}$Wikipedia Level 4 vital articles: https:
//en.wikipedia.org/wiki/wikipedia:
Vital\_articles/Level/4
|
[
[
0.12005615234375,
0.0872802734375,
0.489501953125,
0.0872802734375,
0.489501953125,
0.1009521484375,
0.12005615234375,
0.1009521484375
],
[
0.486083984375,
0.0914306640625,
0.50537109375,
0.0914306640625,
0.50537109375,
0.0955810546875,
0.486083984375,
0.0955810546875
],
[
0.51123046875,
0.0885009765625,
0.87744140625,
0.0885009765625,
0.87744140625,
0.09857177734375,
0.51123046875,
0.09857177734375
],
[
0.118408203125,
0.10272216796875,
0.405517578125,
0.10272216796875,
0.405517578125,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.5146484375,
0.10272216796875,
0.87890625,
0.10272216796875,
0.87890625,
0.11639404296875,
0.5146484375,
0.11639404296875
],
[
0.1368408203125,
0.1181640625,
0.486083984375,
0.1181640625,
0.486083984375,
0.1318359375,
0.1368408203125,
0.1318359375
],
[
0.5146484375,
0.12054443359375,
0.87744140625,
0.12054443359375,
0.87744140625,
0.1312255859375,
0.5146484375,
0.1312255859375
],
[
0.118408203125,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.118408203125,
0.1484375
],
[
0.51318359375,
0.1353759765625,
0.87890625,
0.1353759765625,
0.87890625,
0.1484375,
0.51318359375,
0.1484375
],
[
0.12005615234375,
0.1519775390625,
0.483642578125,
0.1519775390625,
0.483642578125,
0.162109375,
0.12005615234375,
0.162109375
],
[
0.51318359375,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.1650390625,
0.51318359375,
0.1650390625
],
[
0.118408203125,
0.16748046875,
0.484375,
0.16748046875,
0.484375,
0.18115234375,
0.118408203125,
0.18115234375
],
[
0.51318359375,
0.16748046875,
0.87890625,
0.16748046875,
0.87890625,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.1175537109375,
0.1817626953125,
0.47021484375,
0.1817626953125,
0.47021484375,
0.1953125,
0.1175537109375,
0.1953125
],
[
0.468505859375,
0.18701171875,
0.474365234375,
0.18701171875,
0.474365234375,
0.1923828125,
0.468505859375,
0.1923828125
],
[
0.51318359375,
0.182861328125,
0.87890625,
0.182861328125,
0.87890625,
0.196533203125,
0.51318359375,
0.196533203125
],
[
0.12005615234375,
0.20068359375,
0.48779296875,
0.20068359375,
0.48779296875,
0.21142578125,
0.12005615234375,
0.21142578125
],
[
0.51123046875,
0.1995849609375,
0.71533203125,
0.1995849609375,
0.71533203125,
0.213134765625,
0.51123046875,
0.213134765625
],
[
0.118408203125,
0.215576171875,
0.48779296875,
0.215576171875,
0.48779296875,
0.2286376953125,
0.118408203125,
0.2286376953125
],
[
0.52880859375,
0.215576171875,
0.880859375,
0.215576171875,
0.880859375,
0.2286376953125,
0.52880859375,
0.2286376953125
],
[
0.1175537109375,
0.232177734375,
0.486083984375,
0.232177734375,
0.486083984375,
0.2452392578125,
0.1175537109375,
0.2452392578125
],
[
0.509765625,
0.232177734375,
0.8818359375,
0.23095703125,
0.8818359375,
0.24462890625,
0.509765625,
0.2452392578125
],
[
0.1175537109375,
0.2476806640625,
0.486083984375,
0.2476806640625,
0.486083984375,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.509765625,
0.2476806640625,
0.880859375,
0.2464599609375,
0.880859375,
0.26123046875,
0.509765625,
0.262451171875
],
[
0.1175537109375,
0.26416015625,
0.474365234375,
0.26416015625,
0.474365234375,
0.27783203125,
0.1175537109375,
0.27783203125
],
[
0.51318359375,
0.26416015625,
0.880859375,
0.26416015625,
0.880859375,
0.27783203125,
0.51318359375,
0.27783203125
],
[
0.51318359375,
0.27978515625,
0.8818359375,
0.27978515625,
0.8818359375,
0.29345703125,
0.51318359375,
0.29345703125
],
[
0.1158447265625,
0.290283203125,
0.420654296875,
0.290283203125,
0.420654296875,
0.303955078125,
0.1158447265625,
0.303955078125
],
[
0.5146484375,
0.295654296875,
0.880859375,
0.295654296875,
0.880859375,
0.308837890625,
0.5146484375,
0.308837890625
],
[
0.12176513671875,
0.31640625,
0.486083984375,
0.31640625,
0.486083984375,
0.330078125,
0.12176513671875,
0.330078125
],
[
0.51318359375,
0.311279296875,
0.880859375,
0.311279296875,
0.880859375,
0.325927734375,
0.51318359375,
0.325927734375
],
[
0.12005615234375,
0.33203125,
0.48779296875,
0.33203125,
0.48779296875,
0.345703125,
0.12005615234375,
0.345703125
],
[
0.51318359375,
0.327880859375,
0.880859375,
0.327880859375,
0.880859375,
0.341552734375,
0.51318359375,
0.341552734375
],
[
0.1158447265625,
0.3466796875,
0.48779296875,
0.347412109375,
0.48779296875,
0.3623046875,
0.1158447265625,
0.361083984375
],
[
0.5087890625,
0.34326171875,
0.8818359375,
0.344482421875,
0.8818359375,
0.358154296875,
0.5087890625,
0.35693359375
],
[
0.118408203125,
0.364013671875,
0.483642578125,
0.364013671875,
0.483642578125,
0.377685546875,
0.118408203125,
0.377685546875
],
[
0.51318359375,
0.361083984375,
0.880859375,
0.361083984375,
0.880859375,
0.374755859375,
0.51318359375,
0.374755859375
],
[
0.118408203125,
0.380126953125,
0.486083984375,
0.380126953125,
0.486083984375,
0.393798828125,
0.118408203125,
0.393798828125
],
[
0.51318359375,
0.377197265625,
0.880859375,
0.377197265625,
0.880859375,
0.39013671875,
0.51318359375,
0.39013671875
],
[
0.1158447265625,
0.396728515625,
0.486083984375,
0.394287109375,
0.486083984375,
0.407958984375,
0.1158447265625,
0.410400390625
],
[
0.51318359375,
0.393798828125,
0.8076171875,
0.393798828125,
0.8076171875,
0.40673828125,
0.51318359375,
0.40673828125
],
[
0.118408203125,
0.412109375,
0.486083984375,
0.412109375,
0.486083984375,
0.42578125,
0.118408203125,
0.42578125
],
[
0.51904296875,
0.418701171875,
0.87890625,
0.418701171875,
0.87890625,
0.431640625,
0.51904296875,
0.431640625
],
[
0.118408203125,
0.4287109375,
0.486083984375,
0.4287109375,
0.486083984375,
0.4423828125,
0.118408203125,
0.4423828125
],
[
0.5146484375,
0.435791015625,
0.87890625,
0.435791015625,
0.87890625,
0.446533203125,
0.5146484375,
0.446533203125
],
[
0.1175537109375,
0.44482421875,
0.486083984375,
0.44482421875,
0.486083984375,
0.457763671875,
0.1175537109375,
0.457763671875
],
[
0.51318359375,
0.45068359375,
0.880859375,
0.45068359375,
0.880859375,
0.46435546875,
0.51318359375,
0.46435546875
],
[
0.12005615234375,
0.4619140625,
0.486083984375,
0.4619140625,
0.486083984375,
0.47265625,
0.12005615234375,
0.47265625
],
[
0.51123046875,
0.466064453125,
0.880859375,
0.466064453125,
0.880859375,
0.479736328125,
0.51123046875,
0.479736328125
],
[
0.118408203125,
0.478515625,
0.484375,
0.478515625,
0.484375,
0.4892578125,
0.118408203125,
0.4892578125
],
[
0.51318359375,
0.482177734375,
0.87890625,
0.482177734375,
0.87890625,
0.495361328125,
0.51318359375,
0.495361328125
],
[
0.12005615234375,
0.49462890625,
0.246826171875,
0.49462890625,
0.246826171875,
0.5048828125,
0.12005615234375,
0.5048828125
],
[
0.515625,
0.49951171875,
0.87890625,
0.49951171875,
0.87890625,
0.51025390625,
0.515625,
0.51025390625
],
[
0.1368408203125,
0.5087890625,
0.48779296875,
0.5087890625,
0.48779296875,
0.52392578125,
0.1368408203125,
0.52392578125
],
[
0.51318359375,
0.51416015625,
0.880859375,
0.51416015625,
0.880859375,
0.52783203125,
0.51318359375,
0.52783203125
],
[
0.118408203125,
0.525390625,
0.486083984375,
0.525390625,
0.486083984375,
0.5390625,
0.118408203125,
0.5390625
],
[
0.51318359375,
0.52978515625,
0.80615234375,
0.52978515625,
0.80615234375,
0.54345703125,
0.51318359375,
0.54345703125
],
[
0.118408203125,
0.54150390625,
0.486083984375,
0.54150390625,
0.486083984375,
0.5546875,
0.118408203125,
0.5546875
],
[
0.118408203125,
0.55810546875,
0.483642578125,
0.55810546875,
0.483642578125,
0.568359375,
0.118408203125,
0.568359375
],
[
0.509765625,
0.56005859375,
0.5498046875,
0.56005859375,
0.5498046875,
0.57080078125,
0.509765625,
0.57080078125
],
[
0.54248046875,
0.55859375,
0.724609375,
0.55859375,
0.724609375,
0.572265625,
0.54248046875,
0.572265625
],
[
0.1175537109375,
0.572265625,
0.483642578125,
0.572265625,
0.483642578125,
0.5859375,
0.1175537109375,
0.5859375
],
[
0.51318359375,
0.5849609375,
0.87744140625,
0.5849609375,
0.87744140625,
0.5986328125,
0.51318359375,
0.5986328125
],
[
0.12005615234375,
0.59130859375,
0.484375,
0.59130859375,
0.484375,
0.6015625,
0.12005615234375,
0.6015625
],
[
0.1175537109375,
0.60498046875,
0.483642578125,
0.60498046875,
0.483642578125,
0.6181640625,
0.1175537109375,
0.6181640625
],
[
0.51123046875,
0.599609375,
0.87890625,
0.599609375,
0.87890625,
0.61328125,
0.51123046875,
0.61328125
],
[
0.1175537109375,
0.62060546875,
0.484375,
0.62060546875,
0.484375,
0.63427734375,
0.1175537109375,
0.63427734375
],
[
0.51318359375,
0.6162109375,
0.880859375,
0.6162109375,
0.880859375,
0.6298828125,
0.51318359375,
0.6298828125
],
[
0.1175537109375,
0.6357421875,
0.484375,
0.6357421875,
0.484375,
0.6494140625,
0.1175537109375,
0.6494140625
],
[
0.51318359375,
0.6328125,
0.82373046875,
0.6328125,
0.82373046875,
0.646484375,
0.51318359375,
0.646484375
],
[
0.118408203125,
0.65380859375,
0.483642578125,
0.65283203125,
0.483642578125,
0.66650390625,
0.118408203125,
0.66748046875
],
[
0.53076171875,
0.6494140625,
0.880859375,
0.6494140625,
0.880859375,
0.6630859375,
0.53076171875,
0.6630859375
],
[
0.118408203125,
0.66845703125,
0.483642578125,
0.66845703125,
0.483642578125,
0.681640625,
0.118408203125,
0.681640625
],
[
0.51318359375,
0.6650390625,
0.880859375,
0.6650390625,
0.880859375,
0.6787109375,
0.51318359375,
0.6787109375
],
[
0.118408203125,
0.68603515625,
0.484375,
0.68603515625,
0.484375,
0.69970703125,
0.118408203125,
0.69970703125
],
[
0.51123046875,
0.68115234375,
0.880859375,
0.6796875,
0.880859375,
0.693359375,
0.51123046875,
0.69482421875
],
[
0.51123046875,
0.6953125,
0.880859375,
0.69677734375,
0.880859375,
0.71142578125,
0.51123046875,
0.71044921875
],
[
0.1175537109375,
0.70166015625,
0.193115234375,
0.69970703125,
0.193115234375,
0.71435546875,
0.1175537109375,
0.71630859375
],
[
0.1368408203125,
0.71875,
0.48779296875,
0.71875,
0.48779296875,
0.732421875,
0.1368408203125,
0.732421875
],
[
0.51318359375,
0.71337890625,
0.880859375,
0.71337890625,
0.880859375,
0.72705078125,
0.51318359375,
0.72705078125
],
[
0.118408203125,
0.73388671875,
0.486083984375,
0.73388671875,
0.486083984375,
0.74755859375,
0.118408203125,
0.74755859375
],
[
0.5146484375,
0.728515625,
0.8818359375,
0.728515625,
0.8818359375,
0.7421875,
0.5146484375,
0.7421875
],
[
0.118408203125,
0.75048828125,
0.486083984375,
0.75048828125,
0.486083984375,
0.76416015625,
0.118408203125,
0.76416015625
],
[
0.51318359375,
0.7451171875,
0.8818359375,
0.7451171875,
0.8818359375,
0.7587890625,
0.51318359375,
0.7587890625
],
[
0.509765625,
0.76025390625,
0.83447265625,
0.76123046875,
0.83447265625,
0.775390625,
0.509765625,
0.77490234375
],
[
0.118408203125,
0.76708984375,
0.486083984375,
0.76708984375,
0.486083984375,
0.78076171875,
0.118408203125,
0.78076171875
],
[
0.118408203125,
0.783203125,
0.484375,
0.783203125,
0.484375,
0.79638671875,
0.118408203125,
0.79638671875
],
[
0.51318359375,
0.77783203125,
0.99853515625,
0.77783203125,
0.99853515625,
0.79150390625,
0.51318359375,
0.79150390625
],
[
0.1175537109375,
0.798828125,
0.466796875,
0.798828125,
0.466796875,
0.8125,
0.1175537109375,
0.8125
],
[
0.51318359375,
0.79443359375,
0.8759765625,
0.79443359375,
0.8759765625,
0.8046875,
0.51318359375,
0.8046875
],
[
0.5146484375,
0.81005859375,
0.880859375,
0.81005859375,
0.880859375,
0.82373046875,
0.5146484375,
0.82373046875
],
[
0.1234130859375,
0.82373046875,
0.486083984375,
0.82373046875,
0.486083984375,
0.83740234375,
0.1234130859375,
0.83740234375
],
[
0.5146484375,
0.82763671875,
0.66015625,
0.82763671875,
0.66015625,
0.83837890625,
0.5146484375,
0.83837890625
],
[
0.12005615234375,
0.8388671875,
0.486083984375,
0.8388671875,
0.486083984375,
0.8525390625,
0.12005615234375,
0.8525390625
],
[
0.53076171875,
0.8427734375,
0.880859375,
0.8427734375,
0.880859375,
0.85546875,
0.53076171875,
0.85546875
],
[
0.118408203125,
0.85498046875,
0.486083984375,
0.85498046875,
0.486083984375,
0.8681640625,
0.118408203125,
0.8681640625
],
[
0.51123046875,
0.85693359375,
0.880859375,
0.85791015625,
0.880859375,
0.87255859375,
0.51123046875,
0.87158203125
],
[
0.118408203125,
0.87158203125,
0.484375,
0.87158203125,
0.484375,
0.88525390625,
0.118408203125,
0.88525390625
],
[
0.5341796875,
0.88134765625,
0.7958984375,
0.88134765625,
0.7958984375,
0.89453125,
0.5341796875,
0.89453125
],
[
0.8251953125,
0.8818359375,
0.88330078125,
0.88427734375,
0.88330078125,
0.8955078125,
0.8251953125,
0.8935546875
],
[
0.1376953125,
0.892578125,
0.489501953125,
0.89453125,
0.489501953125,
0.90771484375,
0.1376953125,
0.90625
],
[
0.51318359375,
0.8935546875,
0.8193359375,
0.89453125,
0.8193359375,
0.90771484375,
0.51318359375,
0.90673828125
],
[
0.1175537109375,
0.90625,
0.47607421875,
0.90625,
0.47607421875,
0.91943359375,
0.1175537109375,
0.91943359375
],
[
0.51123046875,
0.90625,
0.7138671875,
0.90625,
0.7138671875,
0.91943359375,
0.51123046875,
0.91943359375
]
] |
[
"是，我们没有发现简单的预测因子(i。e。，段",
"",
"件或偏差在NLP中得到了很好的记录。这些",
"落难度)注释者何时可能使用启发式",
"工件允许模型使用不可靠的快捷方式（Geirhos",
"对收集的数据进行定性分析后发现，寻求启",
"etal.，2020）解决NLP问题。几位研究人员提",
"发式的注释器更有可能创建无效的示例，并且",
"出了针对已知偏差实现鲁棒性的方法。我们请",
"需要在明确声明的信息上进行更简单的单词匹",
"读者参考Wangetal.（2022），以全面回顾这些",
"配（s7）。至关重要的是，这表明对启发式使用",
"方法。定位未知的偏见仍然是一个挑战，我们",
"的测量，例如本文中研究的那些，可以提供",
"",
"的工作可以通过识别启发式寻求注释者来帮助",
"种通用方法来查找众包数据中不可靠的示例，",
"找到可能包含工件的示例",
"并指导我们在数据中发现工件。由于我们在健",
"先前的工作提出了使用模型预测（Lertvit+",
"壮性和数据质量方面暗示了启发式使用，因此",
"tayakumjornandToni，2021）的解释来发现快",
"我们建议未来的数据集创建者跟踪类似的功能",
"捷方式的方法，包括基于样本的解释（Han",
"并评估模型对注释器启发式使用的敏感性。1",
"etal.，2020）和输入特征属性(Bastings etal.,",
"2021;Pezeshkpouretal，2022）。其他有助于诊",
"2Background and Related Work",
"断模型偏差的技术包括构建测试用例(Ribeiro",
"CognitiveHeuristics。人类判断，决策和推理",
"et al., 2020; Ribeiro and Lundberg, 2022)的清",
"中的启发式研究是（Simon，1956；Tverskyand",
"单，构建对比（Gardneretal.，2020）或反事实",
"Kahneman，1974）研究的热门且有影响力的主",
"（Wuetal.，2021）示例以及统计测试（Gururan-",
"题。启发式可以定义为心理捷径，我们在日",
"gan et al.,2018;Gardner et al.,2021)。我们的工",
"常任务中用于快速决策。例如，Tverskyand",
"作是对这些方法的补充，因为我们提供了一种",
"Kahneman（1974）询问参与者是否有更多的英",
"与注释器相关的偏见发现的替代方法",
"语单词以字母K开头或包含K作为3rd字母，",
"Improvedcrowdsourcing.相关工作研究了",
"超过70名%参与者选择了前者，因为以K开",
"对众包协议的修改，以提高数据质量（Bowman",
"头的单词更容易被召回，尽管这是不正确的。",
"etal.,2020;Nangiaetal.,2021）。此外，诸如对",
"这是可用性启发式的一个示例。系统地使用这",
"抗性数据收集(Nieetal.,2020）和使用生成模",
"种启发式方法会导致认知偏差，这是我们思维",
"型(Bartoloetal.,2022Liu et al.,2022)之类的",
"中的非理性模式",
"循环模型众包方法已被证明有助于创建更具挑",
"乍一看，似乎启发式算法总是次优的，但",
"战性的示例。我们相信跟踪注释器启发式使用",
"是以前的工作认为，与优化（Gigerenzer and",
"可以帮助对众包协议进行明智的调整",
"Gaissmaier，2011）相比，启发式算法可以在不",
"确定的情况下导致准确的推论。我们假设启",
"3",
"Annotation Protocol",
"发式方法在确定数据质量方面可以发挥相当",
"我们将多选题阅读理解视为我们的众包任务",
"大的作用，其影响取决于启发式方法的确切",
"性质。先前的工作表明，在相关性判断任务",
"因为我们可以从注释者那里获得丰富的响应和",
"(Eickhoff,2018）中，众筹人员容易受到认知偏",
"交互，这使我们能够探索一系列假设的启发式",
"差的影响，并提供了一个清单来对抗这些偏差",
"方法。我们在这里描述数据收集的方法",
"(Drawsetal.，2021）。相比之下，我们的工作重",
"我们在AmazonMechanicalTurk上提供了带",
"点是如何在写作任务中测量这种启发式方法的",
"有段落的注释器，并要求他们编写带有四个选",
"潜在使用，并提供了启发式使用与模型脆性相",
"项的多项选择题。我们使用了英文维基百科2",
"中“重要文章”的第一段，并确保段落至少有",
"关的证据",
"注释器行为的特征以前在估计注释器任务准",
"50个单词，最多有250个单词。段落跨越11种",
"确性 (Rzeszotarski and Kittur, 201l; Goyal et al.,",
"类型，包括艺术，历史，物理科学等，并从10K",
"2018）时很有用。还发现注释者身份会影响其",
"段落中随机抽取段落。Annotatorswereasked-",
"towritechallengingquestionsthatcannotbean",
"注释(Hubeetal.,2019;Sapet al.,2022）。我们",
"的工作建立在这些结果的基础上，并估计启发",
"sweredbyreadingjustthequestionorpassagealone,andhaveasingleco",
"式使用功能来捕获有关数据质量的隐含线索",
"此外，他们被要求确保段落提供足够的信息来",
"回答问题，同时允许问题需要使用常识或因果",
"Mitigatingand discoveringbiases.在自然语",
"关系进行基本推断",
"言推理，问题回答和参数理解（Gururangan",
"注释器首先有资格避免垃圾邮件行为。此资",
"et al., 2018; McCoy et al., 2019; Niven and Kao,",
"格以无效问题的形式检查垃圾邮件行为，而不",
"2019，interalia）等任务中，数据集中存在工",
"2Wikipedia Level 4vital articles:",
"https:",
"'Our code and collected data is available at",
"//en.wikipedia.org/wiki/wikipedia:",
"https://github.com/chaitanyamalaviya/annotator-heuristics",
"Vital_articles/Level/4"
] |
{
"arxiv_id": "2210.13439",
"language": "zh",
"page_idx": 2
}
|
||
2210.13439-zh-3
|
2210.13439-zh-3.jpg
|
是示例质量。然后，要求注释者在 MTurk 上 查者可能会按顺序对几个问题选择相同的答的一个 HIT 中为 4 个段落写一个选择题。An- 案，随机选择一个答案，或者使用默许偏差notatorswereaskedtonotworkonmorethan8HIT。我 (他们总是选择同意给定的陈述)。。满足我们们从 73 个注释者中收集了 1225 个多项选择 任务的潜在结果是花在任务上的时间少，形成题-答案对。此外，我们还记录了他们的击键和 问题的努力少。假设工作时间为$t$ ,通道 $d$ 中完成一个示例所花费的时间 (确保不计算离开 的令牌数为$l_d$ ,我们考虑以下低时间特征：(1) 屏幕的时间)。我们的注释界面是基于 Nangia $t,(2)\log t,(3)t/l_d,(4)\log(t/l_d)。^3$
et al.(2021) 构建的。有关我们的注释协议的其
他详细信息，请参阅附录 A
4 Cognitive Heuristics in Crowdsourcing
认知启发式是心理捷径，人类在解决问题的任务中使用它来做出快速判断(Simon,1956; Tversky and Kahneman, 1974)。注释器的任务是编写自然语言示例，使用这种启发式方法并非绝对可靠。我们假设，在编写任务时，对启发式的依赖是数据质量差的可追溯指标。在本节中，我们确定了几种启发式方法，它们在注释者行为中的后果以及跟踪它们的功能。后来，我们还表明它们是数据质量重要维度的预测因子
# 4.1 Methodology
2008; Draws et al.,2021)。为了切实地测量启一个可测量的数量，该数量可以自动计算为一个示例(请参见表$\overline{\mathrm{E}}$1)。虽然我们没有最终确定注释者正在使用启发式方法，但我们探索了与每个启发式方法背后的直觉一致的各种特征。这些特征有时可以映射到一起交互的多个启发式，但为了便于演示，我们将它们列在最相关的认知启发式下。这些帮助我们创建注释器启发式 traces ,其中包含注释器所有示例的平均启发式值
为了验证我们的启发式实例是否与注释者中的启发式寻求趋势一致，我们在一系列心理测试 (Frederick, 2005; Toplak et al., 2014; Sirota et al., 2021) 上测量了启发式值与注释者表现的相关性，如 §4.4 所述
## 4.2 Heuristics Studied
Satisficing: 满意是一种认知启发式方法，涉及做出令人满意的选择，而不是最佳的选择(Simon,1956)。就心理过程而言，强烈的满足在社会认知中，Krosnick (1991) 描述了满意度如何体现在调查回复的各种模式中。例如，调
我们通过注释者的响应来估计他们的工作量。一个不断编辑他们的作品或写长篇问题的注释者可能会试图深思熟虑地起草他们的问题。虽然这可能并不总是正确的(例如，工人可能会花时间思考他们的问题，然后才开始写作),但我们假设通常情况下，简短的回答可能是满意的指标。给定在击键流中找到的单词数$k$ , 问题$q$以及所有选项$o_i$都是$l_k,l_q$和$l_o$ , 我们考虑这些不同的特征：(1) $l_q$ , (2) $l_k$ , (3) XMATH8
Availability heuristic: 依赖更容易从我们的
内存中检索的信息的趋势是可用性启发式(Tversky and Kahneman,1973)。例如，在新闻中听到飞机坠毁的消息后，人们可能会夸大飞行的危险。对于我们的任务，一旦注释者阅读了一段文章并提出了一个问题，这个问题和正
为了检验上述假设，我们考虑了几种可能与 确的答案很可能在他们的脑海中随时可用。这我们的任务相关的已知认知启发式方法。这个 可能会导致他们在任何干扰因素选项之前写人列表并不全面，我们请读者参考之前的工作，该信息。因此，我们检查为示例指定的第一个以全面概述认知偏差 (Shah and Oppenheimer, 选项是否也是正确的答案 (第一个选项偏差)
此启发式方法的另一个结果是串行位置效
发式的潜在使用，我们将每个启发式特征化为 应。当出现一系列项目时，例如杂货清单中的
单词或项目列表，人们会比中间的 (Murdock Jr, 1962; Ebbinghaus, 1964) 更好地回忆起该系列中的第一个和最后几个项目，因为它们更容易获得。这种效应也可以解释为首要效应和新近效应的结合。为了测试由于这种启发式方法， 注释者是否将他们的问题错定在段落的第一个或最后一个句子上，我们检查为示例标记的正确答案是否与段落的第一个或最后一个句子中的跨度匹配(串行位置)
Representativeness heuristic: 代表性启发式是我们倾向于使用项目的相似性来做出决策(Kahneman and Tversky, 1972)。例如，如果一个人正在挑选一部电影观看，他们可能会想到他们以前喜欢的电影，并在一部新电影中寻找这些属性。同样，注释者可以在他们的问题中重复相同的结构，以简化决策(e。g。“以下哪项是正确的？”或“[事件]发生在哪一年？”)。这可能意味着他们没有完全投人，或者，他们
感可能涉及对信息的不关注和缺乏信息综合。 找到了一个行之有效的写作策略，他们选择坚
Previous work shows that taking the logarithm normal-
izes the response time distribution (Whelan, $\tilde{2} 008)$ .
|
[
[
0.118408203125,
0.0872802734375,
0.489501953125,
0.0872802734375,
0.489501953125,
0.1009521484375,
0.118408203125,
0.1009521484375
],
[
0.51123046875,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.51123046875,
0.1009521484375
],
[
0.118408203125,
0.10272216796875,
0.484375,
0.10272216796875,
0.484375,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.515625,
0.105712890625,
0.54833984375,
0.105712890625,
0.54833984375,
0.114013671875,
0.515625,
0.114013671875
],
[
0.5439453125,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.5439453125,
0.11639404296875
],
[
0.118408203125,
0.12054443359375,
0.48193359375,
0.12054443359375,
0.48193359375,
0.1312255859375,
0.118408203125,
0.1312255859375
],
[
0.5146484375,
0.12054443359375,
0.87890625,
0.12054443359375,
0.87890625,
0.1312255859375,
0.5146484375,
0.1312255859375
],
[
0.118408203125,
0.1353759765625,
0.483642578125,
0.1353759765625,
0.483642578125,
0.1484375,
0.118408203125,
0.1484375
],
[
0.51318359375,
0.1353759765625,
0.880859375,
0.1353759765625,
0.880859375,
0.1484375,
0.51318359375,
0.1484375
],
[
0.118408203125,
0.1519775390625,
0.483642578125,
0.1519775390625,
0.483642578125,
0.1650390625,
0.118408203125,
0.1650390625
],
[
0.5146484375,
0.1519775390625,
0.87890625,
0.1519775390625,
0.87890625,
0.1650390625,
0.5146484375,
0.1650390625
],
[
0.118408203125,
0.16748046875,
0.484375,
0.16748046875,
0.484375,
0.18115234375,
0.118408203125,
0.18115234375
],
[
0.51318359375,
0.16748046875,
0.8818359375,
0.16748046875,
0.8818359375,
0.1817626953125,
0.51318359375,
0.1817626953125
],
[
0.12005615234375,
0.18408203125,
0.48193359375,
0.18408203125,
0.48193359375,
0.19482421875,
0.12005615234375,
0.19482421875
],
[
0.51123046875,
0.1817626953125,
0.85791015625,
0.179931640625,
0.85791015625,
0.19775390625,
0.51123046875,
0.1995849609375
],
[
0.118408203125,
0.20068359375,
0.483642578125,
0.20068359375,
0.483642578125,
0.21142578125,
0.118408203125,
0.21142578125
],
[
0.53076171875,
0.20068359375,
0.880859375,
0.20068359375,
0.880859375,
0.21435546875,
0.53076171875,
0.21435546875
],
[
0.118408203125,
0.215576171875,
0.343505859375,
0.215576171875,
0.343505859375,
0.2286376953125,
0.118408203125,
0.2286376953125
],
[
0.5146484375,
0.2161865234375,
0.87890625,
0.2161865234375,
0.87890625,
0.2298583984375,
0.5146484375,
0.2298583984375
],
[
0.51123046875,
0.23095703125,
0.880859375,
0.232177734375,
0.880859375,
0.2464599609375,
0.51123046875,
0.2452392578125
],
[
0.1158447265625,
0.2452392578125,
0.48779296875,
0.2452392578125,
0.48779296875,
0.260009765625,
0.1158447265625,
0.260009765625
],
[
0.51318359375,
0.248779296875,
0.880859375,
0.248779296875,
0.880859375,
0.262451171875,
0.51318359375,
0.262451171875
],
[
0.51708984375,
0.2666015625,
0.87890625,
0.2666015625,
0.87890625,
0.276611328125,
0.51708984375,
0.276611328125
],
[
0.118408203125,
0.2724609375,
0.486083984375,
0.2724609375,
0.486083984375,
0.2861328125,
0.118408203125,
0.2861328125
],
[
0.5146484375,
0.28076171875,
0.87890625,
0.28076171875,
0.87890625,
0.29443359375,
0.5146484375,
0.29443359375
],
[
0.118408203125,
0.2880859375,
0.48779296875,
0.2880859375,
0.48779296875,
0.3017578125,
0.118408203125,
0.3017578125
],
[
0.5146484375,
0.29638671875,
0.880859375,
0.29638671875,
0.880859375,
0.31005859375,
0.5146484375,
0.31005859375
],
[
0.118408203125,
0.303955078125,
0.484375,
0.303955078125,
0.484375,
0.317626953125,
0.118408203125,
0.317626953125
],
[
0.51123046875,
0.311279296875,
0.880859375,
0.312255859375,
0.880859375,
0.32666015625,
0.51123046875,
0.325927734375
],
[
0.118408203125,
0.320556640625,
0.483642578125,
0.320556640625,
0.483642578125,
0.330810546875,
0.118408203125,
0.330810546875
],
[
0.5087890625,
0.325927734375,
0.8818359375,
0.32666015625,
0.8818359375,
0.344482421875,
0.5087890625,
0.34326171875
],
[
0.1175537109375,
0.3349609375,
0.483642578125,
0.3349609375,
0.483642578125,
0.3486328125,
0.1175537109375,
0.3486328125
],
[
0.509765625,
0.345703125,
0.64404296875,
0.34326171875,
0.64404296875,
0.35693359375,
0.509765625,
0.359375
],
[
0.1175537109375,
0.3515625,
0.48779296875,
0.350830078125,
0.48779296875,
0.364013671875,
0.1175537109375,
0.365234375
],
[
0.118408203125,
0.369384765625,
0.484375,
0.369384765625,
0.484375,
0.383056640625,
0.118408203125,
0.383056640625
],
[
0.51904296875,
0.372314453125,
0.880859375,
0.372314453125,
0.880859375,
0.385986328125,
0.51904296875,
0.385986328125
],
[
0.118408203125,
0.385498046875,
0.484375,
0.385498046875,
0.484375,
0.3984375,
0.118408203125,
0.3984375
],
[
0.51318359375,
0.3896484375,
0.8818359375,
0.3896484375,
0.8818359375,
0.402587890625,
0.51318359375,
0.402587890625
],
[
0.118408203125,
0.402099609375,
0.48779296875,
0.402099609375,
0.48779296875,
0.4150390625,
0.118408203125,
0.4150390625
],
[
0.51318359375,
0.405029296875,
0.880859375,
0.405029296875,
0.880859375,
0.418701171875,
0.51318359375,
0.418701171875
],
[
0.12005615234375,
0.418701171875,
0.172119140625,
0.418701171875,
0.172119140625,
0.429931640625,
0.12005615234375,
0.429931640625
],
[
0.51318359375,
0.42041015625,
0.87890625,
0.42041015625,
0.87890625,
0.43408203125,
0.51318359375,
0.43408203125
],
[
0.1158447265625,
0.441162109375,
0.267822265625,
0.44482421875,
0.267822265625,
0.4619140625,
0.1158447265625,
0.458984375
],
[
0.51318359375,
0.435791015625,
0.880859375,
0.435791015625,
0.880859375,
0.449462890625,
0.51318359375,
0.449462890625
],
[
0.53564453125,
0.457763671875,
0.54736328125,
0.457763671875,
0.54736328125,
0.46142578125,
0.53564453125,
0.46142578125
],
[
0.5439453125,
0.454833984375,
0.87890625,
0.454833984375,
0.87890625,
0.465576171875,
0.5439453125,
0.465576171875
],
[
0.118408203125,
0.4697265625,
0.483642578125,
0.4697265625,
0.483642578125,
0.479736328125,
0.118408203125,
0.479736328125
],
[
0.51123046875,
0.4697265625,
0.880859375,
0.4697265625,
0.880859375,
0.482666015625,
0.51123046875,
0.482666015625
],
[
0.118408203125,
0.48388671875,
0.484375,
0.48388671875,
0.484375,
0.49755859375,
0.118408203125,
0.49755859375
],
[
0.5146484375,
0.485107421875,
0.880859375,
0.485107421875,
0.880859375,
0.498779296875,
0.5146484375,
0.498779296875
],
[
0.118408203125,
0.49951171875,
0.489501953125,
0.49951171875,
0.489501953125,
0.51318359375,
0.118408203125,
0.51318359375
],
[
0.51318359375,
0.501953125,
0.87158203125,
0.501953125,
0.87158203125,
0.515625,
0.51318359375,
0.515625
],
[
0.8671875,
0.5048828125,
0.873046875,
0.5048828125,
0.873046875,
0.51123046875,
0.8671875,
0.51123046875
],
[
0.12005615234375,
0.5166015625,
0.484375,
0.5166015625,
0.484375,
0.52978515625,
0.12005615234375,
0.52978515625
],
[
0.5146484375,
0.51953125,
0.869140625,
0.51953125,
0.869140625,
0.52978515625,
0.5146484375,
0.52978515625
],
[
0.1175537109375,
0.53076171875,
0.484375,
0.5322265625,
0.484375,
0.54638671875,
0.1175537109375,
0.5458984375
],
[
0.53076171875,
0.53369140625,
0.880859375,
0.53369140625,
0.880859375,
0.54736328125,
0.53076171875,
0.54736328125
],
[
0.12005615234375,
0.548828125,
0.48779296875,
0.548828125,
0.48779296875,
0.5625,
0.12005615234375,
0.5625
],
[
0.51318359375,
0.55029296875,
0.87890625,
0.55029296875,
0.87890625,
0.56396484375,
0.51318359375,
0.56396484375
],
[
0.1234130859375,
0.56396484375,
0.484375,
0.56396484375,
0.484375,
0.57763671875,
0.1234130859375,
0.57763671875
],
[
0.5146484375,
0.56787109375,
0.87890625,
0.56787109375,
0.87890625,
0.58056640625,
0.5146484375,
0.58056640625
],
[
0.118408203125,
0.580078125,
0.486083984375,
0.580078125,
0.486083984375,
0.59326171875,
0.118408203125,
0.59326171875
],
[
0.51318359375,
0.5830078125,
0.87890625,
0.5830078125,
0.87890625,
0.5966796875,
0.51318359375,
0.5966796875
],
[
0.1175537109375,
0.595703125,
0.486083984375,
0.595703125,
0.486083984375,
0.609375,
0.1175537109375,
0.609375
],
[
0.51123046875,
0.59716796875,
0.87744140625,
0.59716796875,
0.87744140625,
0.61083984375,
0.51123046875,
0.61083984375
],
[
0.12005615234375,
0.6123046875,
0.48779296875,
0.6123046875,
0.48779296875,
0.6259765625,
0.12005615234375,
0.6259765625
],
[
0.51318359375,
0.615234375,
0.87744140625,
0.615234375,
0.87744140625,
0.6259765625,
0.51318359375,
0.6259765625
],
[
0.118408203125,
0.62890625,
0.484375,
0.62890625,
0.484375,
0.6396484375,
0.118408203125,
0.6396484375
],
[
0.51318359375,
0.63134765625,
0.8818359375,
0.63134765625,
0.8818359375,
0.64453125,
0.51318359375,
0.64453125
],
[
0.1175537109375,
0.6435546875,
0.486083984375,
0.6435546875,
0.486083984375,
0.65673828125,
0.1175537109375,
0.65673828125
],
[
0.51318359375,
0.6484375,
0.87890625,
0.6484375,
0.87890625,
0.6591796875,
0.51318359375,
0.6591796875
],
[
0.1175537109375,
0.6591796875,
0.486083984375,
0.6591796875,
0.486083984375,
0.6728515625,
0.1175537109375,
0.6728515625
],
[
0.51318359375,
0.6630859375,
0.880859375,
0.6630859375,
0.880859375,
0.6767578125,
0.51318359375,
0.6767578125
],
[
0.118408203125,
0.6767578125,
0.48779296875,
0.6767578125,
0.48779296875,
0.69140625,
0.118408203125,
0.69140625
],
[
0.5146484375,
0.6787109375,
0.87890625,
0.6787109375,
0.87890625,
0.6923828125,
0.5146484375,
0.6923828125
],
[
0.12005615234375,
0.69482421875,
0.2275390625,
0.69482421875,
0.2275390625,
0.7060546875,
0.12005615234375,
0.7060546875
],
[
0.51318359375,
0.6953125,
0.705078125,
0.6953125,
0.705078125,
0.708984375,
0.51318359375,
0.708984375
],
[
0.1368408203125,
0.71044921875,
0.484375,
0.71044921875,
0.484375,
0.7236328125,
0.1368408203125,
0.7236328125
],
[
0.118408203125,
0.724609375,
0.486083984375,
0.7255859375,
0.486083984375,
0.74072265625,
0.118408203125,
0.7392578125
],
[
0.515625,
0.7216796875,
0.880859375,
0.7216796875,
0.880859375,
0.7353515625,
0.515625,
0.7353515625
],
[
0.118408203125,
0.7421875,
0.486083984375,
0.7421875,
0.486083984375,
0.755859375,
0.118408203125,
0.755859375
],
[
0.51318359375,
0.7392578125,
0.8818359375,
0.7392578125,
0.8818359375,
0.7529296875,
0.51318359375,
0.7529296875
],
[
0.118408203125,
0.75830078125,
0.484375,
0.75830078125,
0.484375,
0.771484375,
0.118408203125,
0.771484375
],
[
0.51123046875,
0.7548828125,
0.87158203125,
0.7548828125,
0.87158203125,
0.7685546875,
0.51123046875,
0.7685546875
],
[
0.118408203125,
0.77392578125,
0.2880859375,
0.77392578125,
0.2880859375,
0.78759765625,
0.118408203125,
0.78759765625
],
[
0.51123046875,
0.76953125,
0.87890625,
0.76953125,
0.87890625,
0.783203125,
0.51123046875,
0.783203125
],
[
0.5146484375,
0.78759765625,
0.87744140625,
0.78759765625,
0.87744140625,
0.79736328125,
0.5146484375,
0.79736328125
],
[
0.1175537109375,
0.802734375,
0.30908203125,
0.802734375,
0.30908203125,
0.81640625,
0.1175537109375,
0.81640625
],
[
0.51318359375,
0.802734375,
0.87890625,
0.802734375,
0.87890625,
0.81640625,
0.51318359375,
0.81640625
],
[
0.12176513671875,
0.82470703125,
0.48779296875,
0.82470703125,
0.48779296875,
0.83837890625,
0.12176513671875,
0.83837890625
],
[
0.5146484375,
0.8193359375,
0.880859375,
0.8193359375,
0.880859375,
0.83447265625,
0.5146484375,
0.83447265625
],
[
0.51123046875,
0.8349609375,
0.8818359375,
0.8349609375,
0.8818359375,
0.8486328125,
0.51123046875,
0.8486328125
],
[
0.12005615234375,
0.84130859375,
0.48779296875,
0.84130859375,
0.48779296875,
0.85498046875,
0.12005615234375,
0.85498046875
],
[
0.51123046875,
0.8515625,
0.880859375,
0.8515625,
0.880859375,
0.865234375,
0.51123046875,
0.865234375
],
[
0.12005615234375,
0.859375,
0.483642578125,
0.859375,
0.483642578125,
0.869140625,
0.12005615234375,
0.869140625
],
[
0.1175537109375,
0.87255859375,
0.489501953125,
0.87353515625,
0.489501953125,
0.88720703125,
0.1175537109375,
0.8857421875
],
[
0.51318359375,
0.8681640625,
0.880859375,
0.8681640625,
0.880859375,
0.8818359375,
0.51318359375,
0.8818359375
],
[
0.118408203125,
0.8896484375,
0.486083984375,
0.8896484375,
0.486083984375,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.53076171875,
0.89111328125,
0.8818359375,
0.89453125,
0.8818359375,
0.9091796875,
0.53076171875,
0.90625
],
[
0.12176513671875,
0.90673828125,
0.486083984375,
0.90673828125,
0.486083984375,
0.91748046875,
0.12176513671875,
0.91748046875
],
[
0.509765625,
0.90625,
0.82373046875,
0.90625,
0.82373046875,
0.91943359375,
0.509765625,
0.91943359375
]
] |
[
"是示例质量。然后，要求注释者在MTurk上",
"查者可能会按顺序对几个问题选择相同的答",
"的一个HIT中为4个段落写一个选择题。An",
"案",
"随机选择一个答案，或者使用默许偏差",
"notatorswereaskedtonotworkonmorethan8HIT。我",
"（他们总是选择同意给定的陈述）。。满足我们",
"们从73个注释者中收集了1225个多项选择",
"任务的潜在结果是花在任务上的时间少，形成",
"题-答案对。此外，我们还记录了他们的击键和",
"问题的努力少。假设工作时间为t，通道d中",
"完成一个示例所花费的时间（确保不计算离开",
"的令牌数为ld，我们考虑以下低时间特征：（1）",
"屏幕的时间）。我们的注释界面是基于Nangia",
"t,(2)logt,(3)t/ld，(4)log(t/la)。3",
"etal.（2021）构建的。有关我们的注释协议的其",
"我们通过注释者的响应来估计他们的工作",
"他详细信息，请参阅附录A",
"量。一个不断编辑他们的作品或写长篇问题的",
"注释者可能会试图深思熟虑地起草他们的问",
"4Cognitive Heuristics in Crowdsourcing",
"题。虽然这可能并不总是正确的（例如，工人",
"可能会花时间思考他们的问题，然后才开始写",
"认知启发式是心理捷径，人类在解决问题的",
"作），但我们假设通常情况下，简短的回答可",
"任务中使用它来做出快速判断（Simon，1956;",
"能是满意的指标。给定在击键流中找到的单词",
"TverskyandKahneman，1974）。注释器的任务",
"数k，问题q以及所有选项o,都是lk，lg和",
"是编写自然语言示例，使用这种启发式方法",
"l。，我们考虑这些不同的特征：（1）lg，（2）lk",
"并非绝对可靠。我们假设，在编写任务时，对",
"，（3）XMATH8",
"启发式的依赖是数据质量差的可追溯指标。在",
"本节中，我们确定了几种启发式方法，它们在",
"Availabilityheuristic：依赖更容易从我们的",
"注释者行为中的后果以及跟踪它们的功能。后",
"内存中检索的信息的趋势是可用性启发式",
"来，我们还表明它们是数据质量重要维度的预",
"(TverskyandKahneman，1973）。例如，在新闻",
"测因子",
"中听到飞机坠毁的消息后，人们可能会夸大飞",
"4.1 Methodology",
"行的危险。对于我们的任务，一旦注释者阅读",
"",
"一段文章并提出了一个问题，这个问题和正",
"为了检验上述假设，我们考虑了儿种可能与",
"确的答案很可能在他们的脑海中随时可用。这",
"我们的任务相关的已知认知启发式方法。这个",
"可能会导致他们在任何干扰因素选项之前写入",
"列表并不全面，我们请读者参考之前的工作，",
"该信息。因此，我们检查为示例指定的第一个",
"",
"以全面概述认知偏差（ShahandOppenheimer,",
"选项是否也是正确的答案（第一个选项偏差）",
"2008;Drawsetal.，2021）。为了切实地测量启",
"此启发式方法的另一个结果是串行位置效",
"发式的潜在使用，我们将每个启发式特征化为",
"应。当出现一系列项目时，例如杂货清单中的",
"一个可测量的数量，该数量可以自动计算为一",
"单词或项目列表，人们会比中间的（MurdockJr,",
"个示例（请参见表E1）。虽然我们没有最终",
"1962；Ebbinghaus，1964）更好地回忆起该系列",
"确定注释者正在使用启发式方法，但我们探索",
"中的第一个和最后几个项目，因为它们更容易",
"了与每个启发式方法背后的直觉一致的各种特",
"获得。这种效应也可以解释为首要效应和新近",
"征。这些特征有时可以映射到一起交互的多个",
"效应的结合。为了测试由于这种启发式方法，",
"启发式，但为了便于演示，我们将它们列在最",
"注释者是否将他们的问题锚定在段落的第一个",
"相关的认知启发式下。这些帮助我们创建注释",
"或最后一个句子上，我们检查为示例标记的正",
"器启发式traces，其中包含注释器所有示例的",
"确答案是否与段落的第一个或最后一个句子中",
"平均启发式值",
"的跨度匹配（串行位置）",
"为了验证我们的启发式实例是否与注释者",
"中的启发式寻求趋势一致，我们在一系列心理",
"Representativenessheuristic：代表性启发式",
"测试 (Frederick, 2005; Toplak et al., 2014; Sirota",
"是我们倾向于使用项目的相似性来做出决策",
"etal.,2021)上测量了启发式值与注释者表现的",
"(KahnemanandTversky，1972）。例如，如果",
"相关性，如84.4所述",
"个人正在挑选一部电影观看，他们可能会想到",
"他们以前喜欢的电影，并在一部新电影中寻找",
"4.2Heuristics Studied",
"这些属性。同样，注释者可以在他们的问题中",
"Satisficing：满意是一种认知启发式方法，涉",
"重复相同的结构，以简化决策（e。g。，“以下哪",
"项是正确的？”或“[事件】发生在哪一年？”）。",
"及做出令人满意的选择，而不是最佳的选择",
"这可能意味着他们没有完全投入，或者，他们",
"Simon，1956）。就心理过程而言，强烈的满足",
"感可能涉及对信息的不关注和缺乏信息综合。",
"找到了一个行之有效的写作策略，他们选择坚",
"在社会认知中，Krosnick（1991）描述了满意度",
"3Previous work shows that taking the logarithm normal-",
"如何体现在调查回复的各种模式中。例如，调",
"izes the response time distribution (Whelan, 2008)"
] |
{
"arxiv_id": "2210.13439",
"language": "zh",
"page_idx": 3
}
|
||
2002.07933-zh-4
|
2002.07933-zh-4.jpg
|
Title Suppressed Due to Excessive Size
3.3. Predicting Gradients without Label Information 另一种方法是使用$q_\phi(\cdot\mid\boldsymbol{x},g_{<t})$预测的梯度来更新分类器权重，即。e。,示例$g_t\sim q_\phi(\cdot\mid\boldsymbol{x},g_{<t})$ 。这是一(再次假设$\phi$是一个常数或一个仅取决于$x$的随机变量)。请注意，最小化$H_{p,q}$会使预测梯度$g_t$很好地替代交叉熵梯度$\tilde{g}_t^\mathcal{L}$。因此，我们写下以下目标函数：
$$\min_{w,\phi}\tilde{\mathcal{L}}(w_{t-1};\phi,x_{t})-\lambda\mathbb{E}_{\tilde{g}_{t}^{L}}\left[\log q_{\phi}(\tilde{g}_{t}^{\mathcal{L}}\mid\boldsymbol{x},g_{<t})\right],$$
其中$\tilde{\mathcal{L}}(w;\phi,x)$ 是隐式定义的概率函数，使得$\nabla _{w}\tilde{\mathcal{L} } ( w; \phi , x)$ $\sim$ $q_{\phi }( \cdot \mid x, w)$ 。我们发现这种方法的性能明显优于( 10 )的惩罚方法。为了仅使用预测梯度更新$w$,我们在实现中禁用了(11)的第二项对 $w$ 的依赖性。此外，可以将上面的$\lambda=1$设置为第一项仅取决于$w$,而第二项仅取决于 $\phi$
我们选择仅预测相对于对数的梯度，并使用反向传播计算剩余的梯度。我们考虑$q_\phi$的两个不同的参数化-高斯：$q_\phi(\cdot\mid x,w)=\mathcal{N}\left(\mu,\sigma_q^2I\right)$和拉普拉斯：$q_\phi(\cdot\mid$ $x,w)=\prod_j\operatorname{Lap}\left(\mu_j,\sigma_q/\sqrt2\right)$ ,其中$\mu=s(a)-s(r_\phi(x))$ 和$r_\phi(\cdot)$与以前一样是辅助神经网络。在这些高斯和拉普拉斯参数化下，$H_p,q$分别与$\mathbb{E}||\mu_t-\tilde{g}_t^\mathcal{L}||_2^2$和$\mathbb{E}||\mu_t-\tilde{g}_t^{\mathcal{L}}||_1$成比例。在高斯情况下，$\phi$使用均方误差损失(MSE)函数进行更新，而在拉普拉斯情况下，则使用平均绝对误差损失(MAE)进行更新。前者预计学习速度更快，但对噪声的鲁棒性较差 (Ghosh et al., 2017)
3.4. Reducing Overftting in Gradient Prediction
在(10)和(11)两种方法中，如果$q_\phi(\cdot\mid x,w)$过拟合，分类器仍然可能过拟合。有多种方法可以防止这种情况。可以选择使用小型网络对$q_\phi$进行参数化，或者以无监督的方式预训练并冻结其某些层。在这项工作中，我们选择控制预测梯度均值$\|\mu\|_2^2$的 L2 范数， 同时保持方差$\sigma_q^2$固定。这可以被视为限制梯度$g_t$的容量。
${\mathrm{Proposition~3.1.~If~}g_{t}=\mu_{t}+\epsilon_{t}\:,{\mathrm{where~}\epsilon_{t}}\sim{\mathcal N}(0,\sigma_{q}^{2}I_{d})}$ is independent noise, and $\mathbb{E}\left[\mu_{t}^{T}\mu_{t}\right]\leq L^{2}$, then the following inequality holds:
$$I(g_t;\boldsymbol{y}\mid\boldsymbol{x},g_{<t})\leq\frac{d}{2}\log\left(1+\frac{L^2}{d\sigma_q^2}\right).$$
证明在补充部分 A.2 中提供。当从$d$单变量零均值拉普拉斯分布与方差$\sigma_q^2$的乘积中采样$\epsilon_t$时，同样的界限成立，因为证明仅依赖于$\epsilon_t$为零均值并具有方差$\sigma_q^2$。我们的主要方法的最终目标是：
$$\min_{w,\phi}\tilde{\mathcal{L}}(w;\phi,x_t)-\lambda\mathbb{E}_{\tilde{g}_t^{\mathcal{L}}}\left[\log q_\phi(\tilde{g}_t^{\mathcal{L}}\mid\boldsymbol{x},g_{<t})\right]+\beta\|\mu_t\|_2^2.$$
如前所述，为了仅使用预测的梯度更新$w$ ,我们在实现中禁用了上述第二项和第三项对$w$的依赖性。我们
将此最终方法命名为 LIMIT-\_ 模仿标签\_i 信息\_m 消减 i n t 下雨。我们将高斯分布和拉普拉斯分布的变量分别表示为 LIMIT $_{\mathcal{G}}$和 LIMIT $_{\mathcal{L}}$。LIMIT 的伪代码在补充材料中给出 (Alg。1)。请注意，与以前
个更严格的条件，因为它意味着$\begin{array}{cc}I(g_t;\boldsymbol{y}\mid\boldsymbol{x},g_{<t})=0&\text{例代码在朴儿物料干结山(Alg。1)。唱仕怎，与以前}\\\text{可次但识 }\phi_\text{ 是一个学数式 }{\text{众们和冲王}n\text{ 的阵却 亦}}&\text{的(8)方法相比，这遵循了(6)的精神，从某种意义}\end{array}$
上说，对 $\phi$ 的优化可以看作是对训练算法的优化；即， 通过梯度隐式学习损失函数$\mathcal{L}^{\prime}$ 。通过这种解释，梯度范数惩罚可以被视为平滑学习损失的一种方式，这是一种很好的归纳偏差，有利于学习
# 4. Experiments
我们对有噪声的数据集进行了实验，以查看所提出的方法对不同类型和数量的标签噪声的表现如何。我们比较中最简单的基线是标准交叉熵(CE)和平均绝对误差(MAE)损失函数。下一个基线是$\boxed{\mathrm{F}}($Patrini et al., 2017) 提出的前向校正方法 (FW),其中估计了标签噪声过渡矩阵并将其用于校正损失函数。最后， 我们包括最近提出的行列式互信息(DMI)损失，它是预测标签和给定标签之间混淆矩阵的对数行列式E (Xu et al.,2019)。FW 和 DMI 基线都需要使用 CE 基线的最佳结果进行初始化。为了避免小的实验差异，我们实现了所有基线，严格遵循 FW 和 DMI 的原始实现。我们使用学习率$\alpha=10^-3$和$\beta_1=0.9$的ADAM 优化器$\overline{\mathrm{E}}($Kingma &Ba,2014)训练除 DMI 以外的所有基线。由于 DMI 对学习速率非常敏感，因此我们通过从以下值$\{10^{-3},10^{-4},10^{-5},10^{-6}\}$网格中选择最佳值来对其进行调整。对于所有基线，模型选择是通过在遵循相应火车集的噪声模型的验证集上选择精度最高的模型来完成的。所有分数都报告在干净的测试集上。其他实验细节，包括超参数网格， 在补充部分 B 中介绍。所提出的方法的实现和用于复制实验的代码可在 https://github.com/hrayrhar/ limit-label-memorization 上获得
4.1. MNIST with Uniform Label Corruption
为了比较前面讨论的方法的变体，并查看哪些方法效果很好，我们在带有损坏标签的 MNIST 数据集上进行了实验。在此实验中，我们使用一个简单的均匀标签噪声模型，其中每个标签都以概率$p$随机均匀地设置为不正确的值。在我们的实验中，我们尝试了$p$ 的4 个值-0%, 50%, 80%, 89%。我们将 MNIST 的 60K 图像分为训练集和验证集，分别包含 48K 和 12K 样本。对于每个噪声量，我们尝试 3 种不同的训练集大小- $10^3,10^4$和$4.8\cdot10^4$ 。所有分类器和辅助网络都是 4 层 CNN, 其共享架构在补充 (Sec。B )。对于此实验，我们包括两个额外的基线，其中将加性噪声(高斯或拉普拉斯)添加到相对于对数的梯度中。我们用名称“CE+GN”和“CE+LN”表示这些基线。与这两个基线的比较表明，所提出的方法不仅仅是通过噪声减少梯度中的信息。我们还考虑了 LIMIT 的一种变体，其中不是从$q$中采样$g_t$ ,而是使用预测的平均值$\mu_t$
|
[
[
0.358642578125,
0.058074951171875,
0.61669921875,
0.058074951171875,
0.61669921875,
0.07135009765625,
0.358642578125,
0.07135009765625
],
[
0.08905029296875,
0.08837890625,
0.471435546875,
0.08837890625,
0.471435546875,
0.0985107421875,
0.08905029296875,
0.0985107421875
],
[
0.5,
0.085205078125,
0.88134765625,
0.08587646484375,
0.88134765625,
0.10101318359375,
0.5,
0.0997314453125
],
[
0.5009765625,
0.10162353515625,
0.8837890625,
0.10162353515625,
0.8837890625,
0.11553955078125,
0.5009765625,
0.11553955078125
],
[
0.0906982421875,
0.10986328125,
0.47216796875,
0.10986328125,
0.47216796875,
0.125,
0.0906982421875,
0.125
],
[
0.50244140625,
0.117431640625,
0.8837890625,
0.117431640625,
0.8837890625,
0.1307373046875,
0.50244140625,
0.1307373046875
],
[
0.0906982421875,
0.1256103515625,
0.473876953125,
0.1256103515625,
0.473876953125,
0.1395263671875,
0.0906982421875,
0.1395263671875
],
[
0.50341796875,
0.1337890625,
0.88232421875,
0.1337890625,
0.88232421875,
0.14453125,
0.50341796875,
0.14453125
],
[
0.0906982421875,
0.14013671875,
0.473876953125,
0.14013671875,
0.473876953125,
0.1546630859375,
0.0906982421875,
0.1546630859375
],
[
0.5009765625,
0.147705078125,
0.8837890625,
0.147705078125,
0.8837890625,
0.1610107421875,
0.5009765625,
0.1610107421875
],
[
0.08905029296875,
0.1546630859375,
0.47216796875,
0.1546630859375,
0.47216796875,
0.1685791015625,
0.08905029296875,
0.1685791015625
],
[
0.50244140625,
0.1622314453125,
0.88134765625,
0.1622314453125,
0.88134765625,
0.175537109375,
0.50244140625,
0.175537109375
],
[
0.0906982421875,
0.17041015625,
0.473876953125,
0.17041015625,
0.473876953125,
0.1849365234375,
0.0906982421875,
0.1849365234375
],
[
0.5009765625,
0.177978515625,
0.8837890625,
0.177978515625,
0.8837890625,
0.1912841796875,
0.5009765625,
0.1912841796875
],
[
0.0906982421875,
0.186279296875,
0.44775390625,
0.186279296875,
0.44775390625,
0.199462890625,
0.0906982421875,
0.199462890625
],
[
0.50244140625,
0.1925048828125,
0.8837890625,
0.1925048828125,
0.8837890625,
0.205810546875,
0.50244140625,
0.205810546875
],
[
0.1004638671875,
0.210205078125,
0.47216796875,
0.210205078125,
0.47216796875,
0.2279052734375,
0.1004638671875,
0.2279052734375
],
[
0.50244140625,
0.208984375,
0.74267578125,
0.208984375,
0.74267578125,
0.2197265625,
0.50244140625,
0.2197265625
],
[
0.108642578125,
0.2279052734375,
0.1217041015625,
0.2279052734375,
0.1217041015625,
0.2310791015625,
0.108642578125,
0.2310791015625
],
[
0.08905029296875,
0.2442626953125,
0.473876953125,
0.24560546875,
0.473876953125,
0.260009765625,
0.08905029296875,
0.2587890625
],
[
0.5,
0.24365234375,
0.6357421875,
0.24365234375,
0.6357421875,
0.2568359375,
0.5,
0.2568359375
],
[
0.08905029296875,
0.260009765625,
0.47216796875,
0.261474609375,
0.47216796875,
0.27587890625,
0.08905029296875,
0.274658203125
],
[
0.5,
0.267578125,
0.88232421875,
0.267578125,
0.88232421875,
0.281005859375,
0.5,
0.281005859375
],
[
0.08905029296875,
0.276611328125,
0.473876953125,
0.276611328125,
0.473876953125,
0.290283203125,
0.08905029296875,
0.290283203125
],
[
0.5009765625,
0.28271484375,
0.8837890625,
0.28271484375,
0.8837890625,
0.296630859375,
0.5009765625,
0.296630859375
],
[
0.09234619140625,
0.292236328125,
0.473876953125,
0.292236328125,
0.473876953125,
0.30615234375,
0.09234619140625,
0.30615234375
],
[
0.50341796875,
0.2998046875,
0.88232421875,
0.2998046875,
0.88232421875,
0.31005859375,
0.50341796875,
0.31005859375
],
[
0.0906982421875,
0.306884765625,
0.473876953125,
0.306884765625,
0.473876953125,
0.32080078125,
0.0906982421875,
0.32080078125
],
[
0.5009765625,
0.313232421875,
0.8837890625,
0.313232421875,
0.8837890625,
0.326904296875,
0.5009765625,
0.326904296875
],
[
0.0906982421875,
0.322509765625,
0.34228515625,
0.322509765625,
0.34228515625,
0.3359375,
0.0906982421875,
0.3359375
],
[
0.5009765625,
0.328857421875,
0.88232421875,
0.328857421875,
0.88232421875,
0.34228515625,
0.5009765625,
0.34228515625
],
[
0.0906982421875,
0.34521484375,
0.47216796875,
0.34521484375,
0.47216796875,
0.359130859375,
0.0906982421875,
0.359130859375
],
[
0.50244140625,
0.343505859375,
0.88232421875,
0.3447265625,
0.88232421875,
0.35791015625,
0.50244140625,
0.356689453125
],
[
0.08905029296875,
0.359130859375,
0.47216796875,
0.359130859375,
0.47216796875,
0.37255859375,
0.08905029296875,
0.37255859375
],
[
0.5009765625,
0.359130859375,
0.8837890625,
0.359130859375,
0.8837890625,
0.37255859375,
0.5009765625,
0.37255859375
],
[
0.0906982421875,
0.375,
0.47216796875,
0.375,
0.47216796875,
0.38818359375,
0.0906982421875,
0.38818359375
],
[
0.5009765625,
0.373779296875,
0.8857421875,
0.37255859375,
0.8857421875,
0.386474609375,
0.5009765625,
0.386962890625
],
[
0.0882568359375,
0.392578125,
0.473876953125,
0.391357421875,
0.473876953125,
0.406005859375,
0.0882568359375,
0.4072265625
],
[
0.50244140625,
0.38818359375,
0.8857421875,
0.38818359375,
0.8857421875,
0.401611328125,
0.50244140625,
0.401611328125
],
[
0.50244140625,
0.404052734375,
0.8857421875,
0.404052734375,
0.8857421875,
0.417236328125,
0.50244140625,
0.417236328125
],
[
0.09234619140625,
0.410400390625,
0.47216796875,
0.410400390625,
0.47216796875,
0.42041015625,
0.09234619140625,
0.42041015625
],
[
0.5009765625,
0.41845703125,
0.8837890625,
0.41845703125,
0.8837890625,
0.431884765625,
0.5009765625,
0.431884765625
],
[
0.0906982421875,
0.4248046875,
0.47216796875,
0.4248046875,
0.47216796875,
0.438232421875,
0.0906982421875,
0.438232421875
],
[
0.0906982421875,
0.439453125,
0.473876953125,
0.439453125,
0.473876953125,
0.45263671875,
0.0906982421875,
0.45263671875
],
[
0.50244140625,
0.434326171875,
0.8837890625,
0.434326171875,
0.8837890625,
0.447509765625,
0.50244140625,
0.447509765625
],
[
0.0906982421875,
0.453857421875,
0.473876953125,
0.453857421875,
0.473876953125,
0.46728515625,
0.0906982421875,
0.46728515625
],
[
0.5009765625,
0.448974609375,
0.8837890625,
0.448974609375,
0.8837890625,
0.462158203125,
0.5009765625,
0.462158203125
],
[
0.0906982421875,
0.4697265625,
0.471435546875,
0.4697265625,
0.471435546875,
0.48291015625,
0.0906982421875,
0.48291015625
],
[
0.50244140625,
0.464111328125,
0.8837890625,
0.464111328125,
0.8837890625,
0.477783203125,
0.50244140625,
0.477783203125
],
[
0.08905029296875,
0.484130859375,
0.47216796875,
0.484130859375,
0.47216796875,
0.49755859375,
0.08905029296875,
0.49755859375
],
[
0.50244140625,
0.479248046875,
0.8837890625,
0.479248046875,
0.8837890625,
0.492431640625,
0.50244140625,
0.492431640625
],
[
0.0882568359375,
0.5,
0.131591796875,
0.5,
0.131591796875,
0.5146484375,
0.0882568359375,
0.5146484375
],
[
0.50244140625,
0.494384765625,
0.8857421875,
0.494384765625,
0.8857421875,
0.50830078125,
0.50244140625,
0.50830078125
],
[
0.50341796875,
0.51123046875,
0.8837890625,
0.51123046875,
0.8837890625,
0.521484375,
0.50341796875,
0.521484375
],
[
0.0882568359375,
0.53076171875,
0.44287109375,
0.53076171875,
0.44287109375,
0.54443359375,
0.0882568359375,
0.54443359375
],
[
0.50341796875,
0.5263671875,
0.88232421875,
0.5263671875,
0.88232421875,
0.537109375,
0.50341796875,
0.537109375
],
[
0.50244140625,
0.54052734375,
0.88232421875,
0.54052734375,
0.88232421875,
0.5537109375,
0.50244140625,
0.5537109375
],
[
0.0906982421875,
0.5546875,
0.473876953125,
0.5546875,
0.473876953125,
0.568359375,
0.0906982421875,
0.568359375
],
[
0.5,
0.5537109375,
0.8837890625,
0.5546875,
0.8837890625,
0.5693359375,
0.5,
0.568359375
],
[
0.09234619140625,
0.5693359375,
0.47216796875,
0.5693359375,
0.47216796875,
0.58349609375,
0.09234619140625,
0.58349609375
],
[
0.50244140625,
0.5712890625,
0.8837890625,
0.5712890625,
0.8837890625,
0.58203125,
0.50244140625,
0.58203125
],
[
0.0906982421875,
0.58544921875,
0.473876953125,
0.58544921875,
0.473876953125,
0.5986328125,
0.0906982421875,
0.5986328125
],
[
0.498291015625,
0.58349609375,
0.73388671875,
0.58544921875,
0.73388671875,
0.5986328125,
0.498291015625,
0.5966796875
],
[
0.0906982421875,
0.599609375,
0.47216796875,
0.599609375,
0.47216796875,
0.61279296875,
0.0906982421875,
0.61279296875
],
[
0.0906982421875,
0.61572265625,
0.47216796875,
0.61572265625,
0.47216796875,
0.62890625,
0.0906982421875,
0.62890625
],
[
0.5009765625,
0.6171875,
0.82177734375,
0.6171875,
0.82177734375,
0.630859375,
0.5009765625,
0.630859375
],
[
0.09234619140625,
0.630859375,
0.47216796875,
0.630859375,
0.47216796875,
0.64453125,
0.09234619140625,
0.64453125
],
[
0.09234619140625,
0.64453125,
0.1282958984375,
0.646484375,
0.12744140625,
0.6591796875,
0.0906982421875,
0.6572265625
],
[
0.50341796875,
0.642578125,
0.88232421875,
0.642578125,
0.88232421875,
0.65283203125,
0.50341796875,
0.65283203125
],
[
0.5009765625,
0.65478515625,
0.8857421875,
0.65380859375,
0.8857421875,
0.66845703125,
0.5009765625,
0.66943359375
],
[
0.08660888671875,
0.66357421875,
0.473876953125,
0.662109375,
0.473876953125,
0.6796875,
0.08660888671875,
0.68115234375
],
[
0.5009765625,
0.67041015625,
0.8857421875,
0.67041015625,
0.8857421875,
0.6845703125,
0.5009765625,
0.6845703125
],
[
0.08905029296875,
0.68115234375,
0.47216796875,
0.68115234375,
0.47216796875,
0.69580078125,
0.08905029296875,
0.69580078125
],
[
0.50244140625,
0.68505859375,
0.8837890625,
0.68505859375,
0.8837890625,
0.69873046875,
0.50244140625,
0.69873046875
],
[
0.0882568359375,
0.6962890625,
0.27783203125,
0.6962890625,
0.27783203125,
0.71044921875,
0.0882568359375,
0.71044921875
],
[
0.50341796875,
0.70263671875,
0.88232421875,
0.70263671875,
0.88232421875,
0.71337890625,
0.50341796875,
0.71337890625
],
[
0.148681640625,
0.7197265625,
0.306396484375,
0.72998046875,
0.303955078125,
0.755859375,
0.146240234375,
0.74560546875
],
[
0.5009765625,
0.71533203125,
0.8857421875,
0.71533203125,
0.8857421875,
0.72900390625,
0.5009765625,
0.72900390625
],
[
0.28515625,
0.72607421875,
0.3056640625,
0.72607421875,
0.3056640625,
0.734375,
0.28515625,
0.734375
],
[
0.36767578125,
0.724609375,
0.3798828125,
0.724609375,
0.3798828125,
0.73095703125,
0.36767578125,
0.73095703125
],
[
0.28759765625,
0.73291015625,
0.35546875,
0.73095703125,
0.35546875,
0.7421875,
0.28759765625,
0.744140625
],
[
0.3505859375,
0.734375,
0.396240234375,
0.7392578125,
0.393798828125,
0.7548828125,
0.346435546875,
0.75
],
[
0.50244140625,
0.73095703125,
0.8857421875,
0.73095703125,
0.8857421875,
0.744140625,
0.50244140625,
0.744140625
],
[
0.50244140625,
0.74560546875,
0.8837890625,
0.74560546875,
0.8837890625,
0.7587890625,
0.50244140625,
0.7587890625
],
[
0.5009765625,
0.7587890625,
0.8857421875,
0.76025390625,
0.8857421875,
0.7744140625,
0.5009765625,
0.7734375
],
[
0.0906982421875,
0.77099609375,
0.473876953125,
0.77099609375,
0.473876953125,
0.7841796875,
0.0906982421875,
0.7841796875
],
[
0.50244140625,
0.77587890625,
0.88232421875,
0.77587890625,
0.88232421875,
0.7890625,
0.50244140625,
0.7890625
],
[
0.0906982421875,
0.787109375,
0.471435546875,
0.787109375,
0.471435546875,
0.79736328125,
0.0906982421875,
0.79736328125
],
[
0.50244140625,
0.79150390625,
0.8857421875,
0.79150390625,
0.8857421875,
0.8046875,
0.50244140625,
0.8046875
],
[
0.0906982421875,
0.80029296875,
0.473876953125,
0.80029296875,
0.473876953125,
0.814453125,
0.0906982421875,
0.814453125
],
[
0.50341796875,
0.80810546875,
0.8837890625,
0.80810546875,
0.8837890625,
0.818359375,
0.50341796875,
0.818359375
],
[
0.0906982421875,
0.81640625,
0.34716796875,
0.81640625,
0.34716796875,
0.830078125,
0.0906982421875,
0.830078125
],
[
0.50244140625,
0.8212890625,
0.8857421875,
0.8212890625,
0.8857421875,
0.83544921875,
0.50244140625,
0.83544921875
],
[
0.09234619140625,
0.84326171875,
0.468017578125,
0.84326171875,
0.468017578125,
0.861328125,
0.09234619140625,
0.861328125
],
[
0.50244140625,
0.83837890625,
0.88232421875,
0.83837890625,
0.88232421875,
0.8486328125,
0.50244140625,
0.8486328125
],
[
0.09722900390625,
0.85595703125,
0.11273193359375,
0.85595703125,
0.11273193359375,
0.86279296875,
0.09722900390625,
0.86279296875
],
[
0.50244140625,
0.8515625,
0.8837890625,
0.8515625,
0.8837890625,
0.86572265625,
0.50244140625,
0.86572265625
],
[
0.5009765625,
0.8671875,
0.88232421875,
0.8671875,
0.88232421875,
0.880859375,
0.5009765625,
0.880859375
],
[
0.09234619140625,
0.876953125,
0.473876953125,
0.876953125,
0.473876953125,
0.89013671875,
0.09234619140625,
0.89013671875
],
[
0.50244140625,
0.8798828125,
0.5390625,
0.8818359375,
0.53759765625,
0.896484375,
0.5009765625,
0.89453125
],
[
0.09234619140625,
0.892578125,
0.47216796875,
0.892578125,
0.47216796875,
0.90576171875,
0.09234619140625,
0.90576171875
]
] |
[
"Title Suppressed Due to Excessive Size",
"3.3.PredictingGradientswithout LabelInformation",
"将此最终方法命名为LIMIT-1模仿标签i信息m",
"消减int下雨。我们将高斯分布和拉普拉斯分布",
"另一种方法是使用qs（:「c,9<t）预测的梯度来更新分",
"的变量分别表示为LIMITg和LIMIT℃。LIMIT的",
"类器权重，即。e。，示例gt～q（|a,g<t）。这是一",
"伪代码在补充材料中给出（AIg。1）。请注意，与以前",
"个更严格的条件，因为它意味着I(gt;ya,9<t）=0",
"的（8）方法相比，这遵循了（6）的精神，从某种意义",
"（再次假设Φ是一个常数或一个仅取决于的随机变",
"上说，对Φ的优化可以看作是对训练算法的优化；即",
"量）。请注意，最小化Hp.g会使预测梯度gt很好地替",
"通过梯度隐式学习损失函数L’。通过这种解释，梯度",
"代交叉摘梯度9。因此，我们写下以下目标函数",
"范数惩罚可以被视为平滑学习损失的一种方式，这是",
"min L(wt-1;Φ,at) - AEg [1ogq(gt I ,g<t)], (11)",
"一种很好的归纳偏差，有利于学习",
"U4",
"其中(w;Φ,a)是隐式定义的概率函数，使得",
"4.Experiments",
"Vc(w;Φ,a）～q（r,w）。我们发现这种方法的",
"我们对有噪声的数据集进行了实验，以查看所提出的",
"性能明显优于（10）的惩罚方法。为了仅使用预测梯",
"方法对不同类型和数量的标签噪声的表现如何。我们",
"度更新w，我们在实现中禁用了（11）的第二项对w",
"比较中最简单的基线是标准交叉摘（CE）和平均绝",
"的依赖性。此外，可以将上面的入=1设置为第一项",
"对误差（MAE）损失函数。下一个基线是E(Patrini",
"仅取决于w，而第二项仅取决于Φ",
"etal，2017）提出的前向校正方法（FW），其中估计",
"我们选择仅预测相对于对数的梯度，并使用反向传",
"了标签噪声过渡矩阵并将其用于校正损失函数。最后",
"播计算剩余的梯度。我们考虑9的两个不同的参数",
"我们包括最近提出的行列式互信息（DMI）损失，它",
"化-高斯：（,w）=（,o）和拉普拉斯：q（",
"是预测标签和给定标签之间混淆矩阵的对数行列式E",
"c,w)=II,Lap(μj,ag/V2)),其中μ= s(a)-s(r())",
"(Xuetal.,2019）。FW和DMI基线都需要使用CE",
"基线的最佳结果进行初始化。为了避免小的实验差",
"和T（）与以前一样是辅助神经网络。在这些高斯",
"异，我们实现了所有基线，严格遵循FW和DMI的",
"和拉普拉斯参数化下，Hp,g分别与Elt-和",
"Elut一gtl1成比例。在高斯情况下，Φ使用均方误差",
"原始实现。我们使用学习率α=10-3和β1=0.9的",
"损失（MSE）函数进行更新，而在拉普拉斯情况下，则",
"ADAM优化器E(Kingma&Ba，2014）训练除DMI",
"使用平均绝对误差损失（MAE）进行更新。前者预计",
"以外的所有基线。由于DMI 对学习速率非常敏感，因",
"学习速度更快，但对噪声的鲁棒性较差（Ghoshetal.,",
"此我们通过从以下值[10-3,10-4,10-5,10-61网格中",
"2017)",
"选择最佳值来对其进行调整。对于所有基线，模型选",
"择是通过在遵循相应火车集的噪声模型的验证集上",
"3.4. Reducing Overfitting in Gradient Prediction",
"选择精度最高的模型来完成的。所有分数都报告在",
"干净的测试集上。其他实验细节，包括超参数网格",
"在（10）和（11）两种方法中，如果q（「a,w）过拟",
"在补充部分B中介绍。所提出的方法的实现和用",
"合，分类器仍然可能过拟合。有多种方法可以防止这",
"于复制实验的代码可在https://github.com/hrayrhar/",
"种情况。可以选择使用小型网络对9Φ进行参数化，或",
"limit-label-memorization上获得",
"者以无监督的方式预训练并冻结其某些层。在这项工",
"作中，我们选择控制预测梯度均值2的L2范数",
"4.1. MNIST with Uniform Label Corruption",
"同时保持方差。固定。这可以被视为限制梯度9t的",
"容量。",
"为广比较前面讨论的方法的变体，并查看哪些方法效",
"果很好，我们在带有损坏标签的MNIST数据集上进",
"Proposition 3.1. If gt = μt +et, where et ~N(0,ogIa)",
"行了实验。在此实验中，我们使用一个简单的均匀标",
"is independent noise, and E μ μt] ≤ L? , then the",
"签噪声模型，其中每个标签都以概率β随机均匀地设",
"following inequality holds:",
"置为不正确的值。在我们的实验中，我们尝试了P的",
"I(gt:y / a,g<t)≤2l",
"4个值-0%，50%，80%，89%。我们将MNIST的60K",
"d",
"",
"5log（1+",
"doa",
"图像分为训练集和验证集，分别包含48K和12K样",
"本。对于每个噪声量，我们尝试3种不同的训练集大",
"小-103，104和4.8·104。所有分类器和辅助网络都",
"证明在补充部分A.2中提供。当从d单变量零均值",
"是4层CNN，其共享架构在补充（Sec。B）。对于",
"拉普拉斯分布与方差的乘积中采样Et时，同样的",
"此实验，我们包括两个额外的基线，其中将加性噪声",
"界限成立，因为证明仅依赖于Et为零均值并具有方差",
"（高斯或拉普拉斯）添加到相对于对数的梯度中。我们",
"o。。我们的主要方法的最终目标是：",
"用名称“CE+GN”和“CE+LN”表示这些基线。与",
"min L(w; Φ, t) -Ege [1og q(gt [, 9<t)] + Bllμtl2",
"这两个基线的比较表明：所提出的方法不仅仅是通过",
"w.",
"噪声减少梯度中的信息。我们还考虑了LIMIT的一种",
"变体，其中不是从q中采样9t，而是使用预测的平均",
"如前所述，为了仅使用预测的梯度更新w，我们在实",
"值μ",
"现中禁用了上述第二项和第三项对w的依赖性。我们"
] |
{
"arxiv_id": "2002.07933",
"language": "zh",
"page_idx": 4
}
|
||
1402.2641-zh-1
|
1402.2641-zh-1.jpg
|
# 在具有模型的圆弧图上，没有三个弧覆盖圆$^{*\dagger}$
Luciano N. Grippo Martín D. Safe
Instituto de Ciencias, Universidad Nacional de General Sarmiento
Los Polvorines, Buenos Aires, Argentina
## Abstract
An interval graph is the intersection graph of a finite set of intervals on a line and a circular-arc graph is the intersection graph of a finite set of arcs on a circle. While a forbidden induced subgraph characterization of interval graphs was found fifty years ago, finding an analogous characterization for circular-arc graphs is a long-standing open problem. In this work, we study the intersection graphs of finite sets of arcs on a circle no three of which cover the circle, known as normal Helly circular-arc graphs. Those circular-arc graphs which are minimal forbidden induced subgraphs for the class of normal Helly circular-arc graphs were identified by Lin, Soulignac, and Szwarcfiter, who also posed the problem of determining the remaining minimal forbidden induced subgraphs. In this work, we solve their problem, obtaining the complete list of minimal forbidden induced subgraphs for the class of normal Helly circular-arc graphs.
关键字。禁止子图，区间图，正常 Helly 圆弧图
# $\textbf{1}$ $\textbf{Introduction}$
有限集合族的交图有一个顶点表示该族的每个成员，当且仅当它们表示的族的成员具有非空交集时，两个顶点才相邻。区间图是一条线上一组有限区间的交集图。50 年前，Lekkerkerker and Boland (1962) 通过最小禁止诱导子图发现了他们著名的区间图特征。区间图的一个有趣的特例是单位区间图，它是具有相同长度的所有封闭 (或所有开放) 区间的有限集的交图，并且由于 Roberts (1969)(另请参见 ?)
圆弧图是圆上一组有限圆弧的相交图。尽管它们在定义上与区间图相似， 但用禁止的诱导子图表征圆弧图是一个长期存在的开放问题(见？,p。54 ; Klee, 1969 )。Tucker (1971) 率先研究了圆弧图和一些重要的子类，例如单位圆弧图(类似于单位间隔图定义)和适当的圆弧图，它们是有限的一组圆弧的相交图在一个圆上，没有一个弧包含在另一个弧中。此外，Tucker (1974) 发现了
This paper originally appeared in proceedings of the XVI Congreso Latino-Iberoamericano de Investigacin Operativa and the XLIV Simpósio Brasileiro de Pesquisa Operacional, September 24-28, 2012, Rio de Janeiro, Brazil. Anais do XLIV Simpósio Brasileiro de Pesquisa Operacional , SOBRAPO, Rio de Janeiro, Brazil, 2012, pages $4093–4104.$ URL: http: //www. din. uem. br/sbpo/sbpo2012/pdf/arq0518. pdf
$^{\dagger}$E-mail addresses: L.N. Grippo (lgrippo@ungs.edu.ar) and M.D. Safe (msafe@ungs.edu.ar)
2
|
[
[
0.408935546875,
0.1644287109375,
0.5859375,
0.162109375,
0.5859375,
0.18701171875,
0.408935546875,
0.189453125
],
[
0.2476806640625,
0.1923828125,
0.76416015625,
0.190673828125,
0.76416015625,
0.215576171875,
0.2476806640625,
0.21728515625
],
[
0.297119140625,
0.240478515625,
0.484375,
0.2423095703125,
0.484375,
0.260009765625,
0.297119140625,
0.25830078125
],
[
0.55322265625,
0.243408203125,
0.69775390625,
0.243408203125,
0.69775390625,
0.257080078125,
0.55322265625,
0.257080078125
],
[
0.2568359375,
0.27490234375,
0.74072265625,
0.27490234375,
0.74072265625,
0.2880859375,
0.2568359375,
0.2880859375
],
[
0.34765625,
0.292236328125,
0.64990234375,
0.292236328125,
0.64990234375,
0.305908203125,
0.34765625,
0.305908203125
],
[
0.46337890625,
0.35986328125,
0.53662109375,
0.35986328125,
0.53662109375,
0.374755859375,
0.46337890625,
0.374755859375
],
[
0.214111328125,
0.385498046875,
0.81005859375,
0.38427734375,
0.81005859375,
0.3984375,
0.214111328125,
0.399658203125
],
[
0.187255859375,
0.402099609375,
0.81201171875,
0.402099609375,
0.81201171875,
0.4150390625,
0.187255859375,
0.4150390625
],
[
0.187255859375,
0.41748046875,
0.81005859375,
0.41748046875,
0.81005859375,
0.43115234375,
0.187255859375,
0.43115234375
],
[
0.185546875,
0.432861328125,
0.81005859375,
0.43408203125,
0.81005859375,
0.448974609375,
0.185546875,
0.44775390625
],
[
0.187255859375,
0.45068359375,
0.81005859375,
0.45068359375,
0.81005859375,
0.46435546875,
0.187255859375,
0.46435546875
],
[
0.187255859375,
0.465576171875,
0.81005859375,
0.465576171875,
0.81005859375,
0.478515625,
0.187255859375,
0.478515625
],
[
0.187255859375,
0.482666015625,
0.81005859375,
0.482666015625,
0.81005859375,
0.496337890625,
0.187255859375,
0.496337890625
],
[
0.187255859375,
0.498779296875,
0.8095703125,
0.498779296875,
0.8095703125,
0.51318359375,
0.187255859375,
0.51318359375
],
[
0.187255859375,
0.51416015625,
0.81201171875,
0.51416015625,
0.81201171875,
0.52783203125,
0.187255859375,
0.52783203125
],
[
0.185546875,
0.53076171875,
0.81201171875,
0.53076171875,
0.81201171875,
0.54443359375,
0.185546875,
0.54443359375
],
[
0.187255859375,
0.54638671875,
0.81005859375,
0.54638671875,
0.81005859375,
0.56103515625,
0.187255859375,
0.56103515625
],
[
0.185546875,
0.5625,
0.328369140625,
0.5625,
0.328369140625,
0.576171875,
0.185546875,
0.576171875
],
[
0.187255859375,
0.58056640625,
0.56494140625,
0.58056640625,
0.56494140625,
0.59423828125,
0.187255859375,
0.59423828125
],
[
0.183837890625,
0.62353515625,
0.3955078125,
0.62353515625,
0.3955078125,
0.64111328125,
0.183837890625,
0.64111328125
],
[
0.18896484375,
0.65771484375,
0.81005859375,
0.65771484375,
0.81005859375,
0.67138671875,
0.18896484375,
0.67138671875
],
[
0.18896484375,
0.6748046875,
0.81005859375,
0.6748046875,
0.81005859375,
0.6884765625,
0.18896484375,
0.6884765625
],
[
0.187255859375,
0.6904296875,
0.8095703125,
0.6904296875,
0.8095703125,
0.70361328125,
0.187255859375,
0.70361328125
],
[
0.18896484375,
0.7060546875,
0.81005859375,
0.7060546875,
0.81005859375,
0.7197265625,
0.18896484375,
0.7197265625
],
[
0.187255859375,
0.72265625,
0.81201171875,
0.72265625,
0.81201171875,
0.736328125,
0.187255859375,
0.736328125
],
[
0.185546875,
0.73681640625,
0.343505859375,
0.73681640625,
0.343505859375,
0.75048828125,
0.185546875,
0.75048828125
],
[
0.2166748046875,
0.7548828125,
0.81201171875,
0.7548828125,
0.81201171875,
0.7685546875,
0.2166748046875,
0.7685546875
],
[
0.18896484375,
0.77099609375,
0.81201171875,
0.77099609375,
0.81201171875,
0.78369140625,
0.18896484375,
0.78369140625
],
[
0.18896484375,
0.7880859375,
0.8095703125,
0.7880859375,
0.8095703125,
0.798828125,
0.18896484375,
0.798828125
],
[
0.187255859375,
0.802734375,
0.81005859375,
0.802734375,
0.81005859375,
0.81640625,
0.187255859375,
0.81640625
],
[
0.187255859375,
0.81787109375,
0.8095703125,
0.818359375,
0.8095703125,
0.8330078125,
0.187255859375,
0.83203125
],
[
0.162109375,
0.84423828125,
0.85986328125,
0.84423828125,
0.85986328125,
0.85791015625,
0.162109375,
0.85791015625
],
[
0.1368408203125,
0.859375,
0.861328125,
0.85791015625,
0.861328125,
0.87158203125,
0.1368408203125,
0.87255859375
],
[
0.1368408203125,
0.87255859375,
0.85986328125,
0.87255859375,
0.85986328125,
0.8857421875,
0.1368408203125,
0.8857421875
],
[
0.1368408203125,
0.8857421875,
0.84814453125,
0.88720703125,
0.84814453125,
0.90185546875,
0.1368408203125,
0.90087890625
],
[
0.162109375,
0.90185546875,
0.78076171875,
0.90185546875,
0.78076171875,
0.9150390625,
0.162109375,
0.9150390625
],
[
0.490234375,
0.939453125,
0.5087890625,
0.939453125,
0.5087890625,
0.955078125,
0.490234375,
0.955078125
]
] |
[
"在具有模型",
"的圆弧图上，没有三个弧覆盖圆",
"Luciano N. Grippo",
"Martin D. Safe",
"Instituto de Ciencias, Universidad Nacional de General Sarmiento",
"Los Polvorines, Buenos Aires, Argentina",
"Abstract",
"An interval graph is the intersection graph of a finite set of intervals on a line and",
"a circular-arc graph is the intersection graph of a finite set of arcs on a circle. While a",
"forbidden induced subgraph characterization of interval graphs was found fifty years",
"ago, finding an analogous characterization for circular-arc graphs is a long-standing",
"open problem. In this work, we study the intersection graphs of finite sets of arcs",
"on a circle no three of which cover the circle, known as normal Helly circular-arc",
"graphs . Those circular-arc graphs which are minimal forbidden induced subgraphs",
"for the class of normal Helly circular-arc graphs were identified by Lin, Soulignac",
"and Szwarcfiter, who also posed the problem of determining the remaining minimal",
"forbidden induced subgraphs. In this work, we solve their problem, obtaining the",
"complete list of minimal forbidden induced subgraphs for the class of normal Helly",
"circular-arc graphs.",
"关键字。禁止子图，区间图，正常Helly圆弧图",
"1Introduction",
"有限集合族的交图有一个顶点表示该族的每个成员，当且仅当它们表示的族的",
"成员具有非空交集时，两个顶点才相邻。区间图是一条线上一组有限区间的交",
"集图。50年前，LekkerkerkerandBoland（1962）通过最小禁止诱导子图发现了",
"他们著名的区间图特征。区间图的一个有趣的特例是单位区间图，它是具有",
"相同长度的所有封闭（或所有开放）区间的有限集的交图，并且由于Roberts",
"(1969)（另请参见?)",
"圆弧图是圆上一组有限圆弧的相交图。尽管它们在定义上与区间图相似，",
"但用禁止的诱导子图表征圆弧图是一个长期存在的开放问题（见？，P。54；",
"Klee，1969）。Tucker（1971）率先研究了圆弧图和一些重要的子类，例如单位圆",
"弧图（类似于单位间隔图定义）和适当的圆弧图，它们是有限的一组圆弧的相",
"交图在一个圆上，没有一个弧包含在另一个弧中。此外，Tucker(1974）发现了",
"*This paper originally appeared in proceedings of the XVI Congreso Latino-Iberoamericano de Investi-",
"gacin Operativa and the XLIV Simposio Brasileiro de Pesquisa Operacional, September 24-28, 2012, Rio de",
"Janeiro, Brazil. Anais do XLIV Simposio Brasileiro de Pesquisa Operacional , SOBRAPO, Rio de Janeiro",
"Brazil, 2012, pages 4093-4104. URL: http://www.din.uem.br/sbpo/sbpo2012/pdf/arq0518.pdf",
"E-mail addresses: L.N. Grippo (Ilgrippo@ungs.edu.ar) and M.D. Safe (msafe @ungs.edu.ar)",
"2"
] |
{
"arxiv_id": "1402.2641",
"language": "zh",
"page_idx": 1
}
|
||
1402.2641-zh-2
|
1402.2641-zh-2.jpg
|
单位圆弧图和适当圆弧图的最小禁止诱导子图特征。从那时起，通过禁止的诱导子图或其他类型的障碍物来表征圆弧图及其某些子类的问题引起了广泛的关注 (Bang-Jensen and Hell, 1994; Bonomo et al., 2009; Feder et al., 1999; Hell and Huang, 2004; Joeris et al., 2011; Lin et al., 2007, 2013; Trotter and Moore, 1976)
我们说一个圆上的一组弧覆盖了圆，如果该组弧共同覆盖了圆的每个点。很容易看出，每个圆弧图是一组有限的圆弧的相交图，没有一个圆弧覆盖圆。圆上有限组圆弧的交集图，其中没有两个圆弧覆盖该圆，称为普通圆弧图。正常圆弧图的类别适当地包含 Tucker (1974) 所示的适当圆弧图的类别，并在 Hell and Huang (2004) 的共二部图的背景下进行了研究。。已知 (参见 ?,尤其是第 5节) 是正常圆弧图类的最小禁止诱导子图的一些部分表征
在这项工作中，我们研究了一个圆上没有三个弧覆盖该圆的有限弧集的交图，称为正常的 Helly 圆弧图 (Lin et al.,2007)。请注意，对于具有至少三个孤的圆上的任何一组弧，没有三个弧覆盖圆的属性也排除了少于三个弧覆盖圆的存在。如果$\mathscr{A}$ 是圆上的一组弧，则：(i)如果$\mathscr{A}$ 没有覆盖圆的两个弧，则称$\mathscr{A}$为正常；(ii) 如果$\mathscr{A}$的每个非空子集由成对相交的弧组成，则称$\mathscr{A}$为Helly; (iii) $\mathscr{A}$均为正常 Helly。事实证明，正常的 Helly 圆弧图可以定义为圆上有限的正常 Helly 弧集的交图。实际上，从 Lin and Szwarcfiter (2006) 的定理$\boxed{\mathrm{E}1\text{ 可以得出，正常的 Helly 圆弧图的定义等效于我们在这项工作中使用的}}$ 定义 (i。e。, 圆上没有三个弧覆盖圆的有限弧集的交图)
以前与普通 Helly 圆弧图有关的一些工作如下。Tucker (1975) 给出了一种算法，该算法最多使用 3$\omega/2$颜色输出任何给定的正常 Helly 圆弧图的适当着色，其中 $\omega$ 表示一组成对相邻顶点的最大大小。在$\overline{\mathrm{ELin~et~al.(2010)}}$中，在研究集团算子下圆弧图的收敛性时，自然会出现正常的 Helly 圆弧图。图$G$的盒性是$k$的最小值，使得$G$是$k$维盒族的交图(i。e。, $k$封闭区间的笛卡尔积); Bhowmick and Sunil Chandran (2011) 表明，正常的 Helly 圆弧图最多具有 3
最近，Lin, Soulignac, and Szwarcfiter (2013) 对正常的 Helly 圆弧图进行了深人研究，在这些图和间隔图之间绘制了许多相似之处。在这项工作中，他们确定了所有这些圆弧图，这些圆弧图是正常 Helly 圆弧图类别的最小禁止诱导子图，并提出了为正常 Helly 圆弧图类别找到剩余的最小禁止子图的问题 (i。e。, 那些不是圆弧图的)。在这项工作中，我们解决了他们的问题，为正常的 Helly 圆弧图提供了最小禁止诱导子图的完整列表
# $\textbf{2}$ $\textbf{Preliminaries}$
这项工作中的所有图都是有限的，无向的，并且没有循环，也没有平行边。图$G$的顶点和边集将分别用$V(G)$和$E(G)$表示。我们用$\overline G$表示$G$的补码，用$N_G(\nu)$表示$G$中顶点$\nu$的邻域，用$N_G[\nu]$表示其封闭邻域$N_G(\nu)\cup\{\nu\}$ 。如果$N_{G}(\nu)=\emptyset$,则顶点$\nu$是孤立的，如果$N_G[\nu]=V(G)$,则是通用的。由一组顶点 S 诱导的子图用$G[S]$表示。如果$H$是$G$的诱导子图，我们说$G$包含诱导的$H$ 。这项工作中的路径意味着至少有一个顶点。$a,b$ -路径是其端点为$a$ 和$b$ 的路径；路径的其余顶点是内部顶点。路径或循环$Z$的弦是不在$Z$中连接$Z$的两个顶点的任何边。无弦路径是没有和弦的路径，无弦循环是四个或更多没有和弦的顶点上的循环。我们用$P_n$ (resp。$C_n$)无弦路径(RESP。循环)在$n$ 顶点上。如果图形没有弦循环，则它是弦的。集团是一组成对相邻的顶点。如果在$G$中存在连接两个顶点的路径，则它们位于图$G$的同一组件中。如果$S$是一个集合，我们用$|S|$表示它的基数。对于此处未定义的标准符号和术语，我们参考 West (2001)
让$\mathscr{G}$成为图形类。如果$\mathscr{G}$的任何成员都不包含诱导的$H$,则图$H$是$\mathscr{G}$的禁止诱导子图。如果$\mathscr{G}$ 的每个成员的每个诱导子图也是$\mathscr{G}$ 的成员，则图的类
3
|
[
[
0.19140625,
0.11639404296875,
0.8095703125,
0.11639404296875,
0.8095703125,
0.1270751953125,
0.19140625,
0.1270751953125
],
[
0.189697265625,
0.1312255859375,
0.81005859375,
0.1312255859375,
0.81005859375,
0.144287109375,
0.189697265625,
0.144287109375
],
[
0.18896484375,
0.146728515625,
0.81201171875,
0.1455078125,
0.81201171875,
0.1602783203125,
0.18896484375,
0.160888671875
],
[
0.185546875,
0.162109375,
0.7841796875,
0.160888671875,
0.7841796875,
0.17578125,
0.185546875,
0.177001953125
],
[
0.2149658203125,
0.1787109375,
0.81005859375,
0.1787109375,
0.81005859375,
0.1923828125,
0.2149658203125,
0.1923828125
],
[
0.189697265625,
0.196533203125,
0.81005859375,
0.196533203125,
0.81005859375,
0.207275390625,
0.189697265625,
0.207275390625
],
[
0.19140625,
0.213134765625,
0.8095703125,
0.213134765625,
0.8095703125,
0.223876953125,
0.19140625,
0.223876953125
],
[
0.18896484375,
0.22802734375,
0.81201171875,
0.22802734375,
0.81201171875,
0.2410888671875,
0.18896484375,
0.2410888671875
],
[
0.187255859375,
0.2423095703125,
0.81201171875,
0.243408203125,
0.81201171875,
0.25830078125,
0.187255859375,
0.257080078125
],
[
0.189697265625,
0.260009765625,
0.63818359375,
0.260009765625,
0.63818359375,
0.273681640625,
0.189697265625,
0.273681640625
],
[
0.2149658203125,
0.275634765625,
0.81201171875,
0.275634765625,
0.81201171875,
0.290283203125,
0.2149658203125,
0.290283203125
],
[
0.18896484375,
0.290283203125,
0.81005859375,
0.29150390625,
0.81005859375,
0.305908203125,
0.18896484375,
0.3046875
],
[
0.18896484375,
0.30810546875,
0.81005859375,
0.30810546875,
0.81005859375,
0.32177734375,
0.18896484375,
0.32177734375
],
[
0.189697265625,
0.325927734375,
0.8095703125,
0.325927734375,
0.8095703125,
0.336181640625,
0.189697265625,
0.336181640625
],
[
0.187255859375,
0.34033203125,
0.81005859375,
0.34033203125,
0.81005859375,
0.35400390625,
0.187255859375,
0.35400390625
],
[
0.187255859375,
0.355712890625,
0.81005859375,
0.355712890625,
0.81005859375,
0.369384765625,
0.187255859375,
0.369384765625
],
[
0.189697265625,
0.372314453125,
0.81005859375,
0.372314453125,
0.81005859375,
0.385986328125,
0.189697265625,
0.385986328125
],
[
0.18896484375,
0.388427734375,
0.81201171875,
0.388427734375,
0.81201171875,
0.402099609375,
0.18896484375,
0.402099609375
],
[
0.187255859375,
0.405029296875,
0.63818359375,
0.405029296875,
0.63818359375,
0.418701171875,
0.187255859375,
0.418701171875
],
[
0.2149658203125,
0.42041015625,
0.486083984375,
0.42041015625,
0.486083984375,
0.43408203125,
0.2149658203125,
0.43408203125
],
[
0.496337890625,
0.42041015625,
0.81201171875,
0.42041015625,
0.81201171875,
0.43408203125,
0.496337890625,
0.43408203125
],
[
0.18896484375,
0.438232421875,
0.453369140625,
0.438232421875,
0.453369140625,
0.448974609375,
0.18896484375,
0.448974609375
],
[
0.5634765625,
0.438232421875,
0.8095703125,
0.438232421875,
0.8095703125,
0.448974609375,
0.5634765625,
0.448974609375
],
[
0.18896484375,
0.453125,
0.426513671875,
0.453125,
0.426513671875,
0.466064453125,
0.18896484375,
0.466064453125
],
[
0.57861328125,
0.457275390625,
0.60693359375,
0.457275390625,
0.60693359375,
0.46142578125,
0.57861328125,
0.46142578125
],
[
0.60693359375,
0.453125,
0.81005859375,
0.453125,
0.81005859375,
0.466064453125,
0.60693359375,
0.466064453125
],
[
0.189697265625,
0.47021484375,
0.4375,
0.47021484375,
0.4375,
0.48095703125,
0.189697265625,
0.48095703125
],
[
0.56103515625,
0.47021484375,
0.8076171875,
0.47021484375,
0.8076171875,
0.48095703125,
0.56103515625,
0.48095703125
],
[
0.18896484375,
0.485107421875,
0.426513671875,
0.485107421875,
0.426513671875,
0.498779296875,
0.18896484375,
0.498779296875
],
[
0.432373046875,
0.4892578125,
0.44580078125,
0.4892578125,
0.44580078125,
0.4921875,
0.432373046875,
0.4921875
],
[
0.61279296875,
0.485107421875,
0.8154296875,
0.485107421875,
0.8154296875,
0.498779296875,
0.61279296875,
0.498779296875
],
[
0.18896484375,
0.5029296875,
0.382080078125,
0.5029296875,
0.382080078125,
0.51318359375,
0.18896484375,
0.51318359375
],
[
0.391357421875,
0.505859375,
0.408935546875,
0.505859375,
0.408935546875,
0.51025390625,
0.391357421875,
0.51025390625
],
[
0.54248046875,
0.5029296875,
0.77734375,
0.5029296875,
0.77734375,
0.51318359375,
0.54248046875,
0.51318359375
],
[
0.2166748046875,
0.51708984375,
0.370361328125,
0.51708984375,
0.370361328125,
0.53076171875,
0.2166748046875,
0.53076171875
],
[
0.52880859375,
0.51708984375,
0.81005859375,
0.51708984375,
0.81005859375,
0.53076171875,
0.52880859375,
0.53076171875
],
[
0.189697265625,
0.53369140625,
0.81005859375,
0.53369140625,
0.81005859375,
0.54736328125,
0.189697265625,
0.54736328125
],
[
0.18896484375,
0.5498046875,
0.385498046875,
0.5498046875,
0.385498046875,
0.5634765625,
0.18896484375,
0.5634765625
],
[
0.4013671875,
0.5498046875,
0.8095703125,
0.5498046875,
0.8095703125,
0.5634765625,
0.4013671875,
0.5634765625
],
[
0.18896484375,
0.56396484375,
0.8095703125,
0.5654296875,
0.8095703125,
0.580078125,
0.18896484375,
0.5791015625
],
[
0.185546875,
0.580078125,
0.8095703125,
0.58056640625,
0.8095703125,
0.595703125,
0.185546875,
0.59423828125
],
[
0.18896484375,
0.59716796875,
0.53662109375,
0.59716796875,
0.53662109375,
0.61083984375,
0.18896484375,
0.61083984375
],
[
0.1822509765625,
0.63720703125,
0.405517578125,
0.6396484375,
0.405517578125,
0.65673828125,
0.1822509765625,
0.65478515625
],
[
0.187255859375,
0.67333984375,
0.81005859375,
0.67333984375,
0.81005859375,
0.68701171875,
0.187255859375,
0.68701171875
],
[
0.187255859375,
0.689453125,
0.81005859375,
0.689453125,
0.81005859375,
0.70361328125,
0.187255859375,
0.70361328125
],
[
0.187255859375,
0.7060546875,
0.81005859375,
0.7060546875,
0.81005859375,
0.7197265625,
0.187255859375,
0.7197265625
],
[
0.187255859375,
0.7216796875,
0.81005859375,
0.7216796875,
0.81005859375,
0.7353515625,
0.187255859375,
0.7353515625
],
[
0.18896484375,
0.73828125,
0.81005859375,
0.73828125,
0.81005859375,
0.751953125,
0.18896484375,
0.751953125
],
[
0.185546875,
0.7529296875,
0.8076171875,
0.7529296875,
0.8076171875,
0.7666015625,
0.185546875,
0.7666015625
],
[
0.189697265625,
0.771484375,
0.8095703125,
0.771484375,
0.8095703125,
0.7822265625,
0.189697265625,
0.7822265625
],
[
0.19140625,
0.7880859375,
0.8076171875,
0.7880859375,
0.8076171875,
0.798828125,
0.19140625,
0.798828125
],
[
0.18896484375,
0.802734375,
0.81005859375,
0.802734375,
0.81005859375,
0.81640625,
0.18896484375,
0.81640625
],
[
0.18896484375,
0.818359375,
0.81005859375,
0.818359375,
0.81005859375,
0.83203125,
0.18896484375,
0.83203125
],
[
0.187255859375,
0.8349609375,
0.81005859375,
0.8349609375,
0.81005859375,
0.8486328125,
0.187255859375,
0.8486328125
],
[
0.187255859375,
0.85107421875,
0.81005859375,
0.85107421875,
0.81005859375,
0.8642578125,
0.187255859375,
0.8642578125
],
[
0.18896484375,
0.86767578125,
0.300537109375,
0.86767578125,
0.300537109375,
0.87890625,
0.18896484375,
0.87890625
],
[
0.21240234375,
0.88134765625,
0.81005859375,
0.8818359375,
0.81005859375,
0.896484375,
0.21240234375,
0.8955078125
],
[
0.18896484375,
0.8984375,
0.81201171875,
0.8984375,
0.81201171875,
0.912109375,
0.18896484375,
0.912109375
],
[
0.489501953125,
0.9404296875,
0.5087890625,
0.9404296875,
0.5087890625,
0.955078125,
0.489501953125,
0.955078125
]
] |
[
"单位圆弧图和适当圆弧图的最小禁止诱导子图特征。从那时起，通过禁正的诱",
"导子图或其他类型的障碍物来表征圆弧图及其某些子类的问题引起了广泛的",
"关注(Bang-Jensen and Hell, 1994; Bonomo et al., 2009; Feder et al., 1999; Hell and",
"Huang, 2004; Joeris et al., 2011; Lin et al., 2007, 2013; Trotter and Moore, 1976)",
"我们说一个圆上的一组弧覆盖了圆，如果该组弧共同覆盖了圆的每个点。",
"很容易看出，每个圆弧图是一组有限的圆弧的相交图，没有一个圆弧覆盖圆。",
"圆上有限组圆弧的交集图，其中没有两个圆弧覆盖该圆，称为普通圆弧图。正",
"常圆弧图的类别适当地包含Tucker（1974）所示的适当圆弧图的类别，并在Hell",
"andHuang（2004）的共二部图的背景下进行了研究。。已知（参见？,尤其是第5",
"节）是正常圆弧图类的最小禁止诱导子图的一些部分表征",
"在这项工作中，我们研究了一个圆上没有三个弧覆盖该圆的有限弧集的交",
"图，称为正常的Helly圆弧图（Linetal.,2007）。请注意，对于具有至少三个弧",
"的圆上的任何一组弧，没有三个弧覆盖圆的属性也排除了少于三个弧覆盖圆",
"的存在。如果α是圆上的一组弧，则：（）如果α没有覆盖圆的两个弧，则",
"称为正常；（ii）如果α的每个非空子集由成对相交的弧组成，则称为",
"Helly；（ii）4均为正常Helly。事实证明，正常的Helly圆弧图可以定义为",
"圆上有限的正常Helly弧集的交图。实际上，从LinandSzwarcfiter（2006）的定",
"理E1可以得出，正常的Helly圆弧图的定义等效于我们在这项工作中使用的",
"定义（i。e。，圆上没有三个弧覆盖圆的有限弧集的交图）",
"以前与普通Helly圆弧图有关的一",
"些工作如下。Tucker（1975）给出了一种",
"算法，该算法最多使用3@/2颜色",
"定的正常Helly圆弧图的适当着",
"色，其中の表示一组成对相邻",
"H",
"Linetal.（2010）中，在研",
"究集团算子下圆弧图的收敛性时",
"正常的Helly圆弧图。图G的盒",
"性是k的最小值，使得G是k",
"",
"k封闭区间的笛卡尔积)；",
"BhowmickandSunilChar",
"",
"常的Helly圆弧图最多具有3",
"最近，Lin,Souligna",
"13)对正常的Helly圆弧图进行了深",
"人研究，在这些图和间隔图之间绘制了许多相似之处。在这项工作中，他们确",
"定了所有这些圆弧图，这",
"华圆弧图是正常Helly圆弧图类别的最小禁止诱导子",
"图，并提出了为正常Helly圆弧图类别找到剩余的最小禁止子图的问题（i。e。",
"那些不是圆弧图的)。在这项工作中，我们解决了他们的问题，为正常的Helly",
"圆弧图提供了最小禁止诱导子图的完整列表",
"2Preliminaries",
"这项工作中的所有图都是有限的，无向的，并且没有循环，也没有平行边。图",
"G的顶点和边集将分别用V(G)和E(G)表示。我们用G表示G的补码，用",
"NG(v)表示G中顶点的邻域，用NG[v]表示其封闭邻域NG(v)U[v)。如果",
"Nc(v)=の，则顶点v是孤立的，如果N[v)=V(G)，则是通用的。由一组顶",
"点S诱导的子图用G[S表示。如果H是G的诱导子图，我们说G包含诱导的",
"H。这项工作中的路径意味着至少有一个顶点。a,b-路径是其端点为a和b的",
"路径；路径的其余顶点是内部顶点。路径或循环Z的弦是不在Z中连接乙的",
"两个顶点的任何边。无弦路径是没有和弦的路径，无弦循环是四个或更多没有",
"和弦的顶点上的循环。我们用Pn(resp。Cn）无弦路径(RESP。循环）在n顶点",
"上。如果图形没有弦循环，则它是弦的。集团是一组成对相邻的顶点。如果在",
"G中存在连接两个顶点的路径，则它们位于图G的同一组件中。如果S是一个",
"集合，我们用S表示它的基数。对于此处未定义的标准符号和术语，我们参",
"考West(2001)",
"让9成为图形类。如果9的任何成员都不包含诱导的H，则图H是9的",
"禁止诱导子图。如果9的每个成员的每个诱导子图也是9的成员，则图的类",
"3"
] |
{
"arxiv_id": "1402.2641",
"language": "zh",
"page_idx": 2
}
|
||
1402.2641-zh-3
|
1402.2641-zh-3.jpg
|
$\mathscr{G}$是遗传的。如果$H$是遗传图类$\mathscr{G}$的禁止诱导子图，则$H$是$\mathscr{G}$的最小禁止诱导子图，如果$H$ 的每个诱导子图都不同于 $H$ 是$\mathscr{G}$ 的成员。显然，遗传图类完全由其最小禁止诱导子图决定：图类完全由那些不包含诱导的最小禁止诱导子图的图组成。以下著名的结果给出了间隔图类的最小禁止诱导子图的完整列表
Theorem 1 (Lekkerkerker and Boland, 1962). The minimal forbidden induced subgraphs for the class of interval graphs are: bipartite claw, umbrella, k -net for every $k\geq2,k$ -tent for every $k\geq3,andC_kfor$ every $k\geq4(see$ Figure $l).$
<FigureHere>
Figure 1: All minimal forbidden induced subgraphs for the class of interval graphs
与间隔图的情况相反，禁止诱导子图表征圆弧图的问题仍然存在。圆弧图类的一些最小禁止诱导子图也是$G_1,G_2,G_3,G_4$,多米诺骨牌，$\overline{G_6}$ 和$C_k^*$, 对于每个$k\geq4$,其中$C_k^*$表示由$C_k$诱导的图。。对于每个$k\geq4$,由$C_k$通过添加通用顶点而产生的图形$k$轮(请参见E图E2)是一个圆弧图，但也是正常类别的最小禁止诱导子图 Helly 圆弧图。在下面的内容中，我们分别使用 net 和 tent 作为 2 -net 和 3 -tent 的简写。
如果图$G$是有限族$\mathscr{F}$的交图，则$\mathscr{F}$称为$\mathscr{F}$的交模型。此外，如果$\mathscr{F}$由线上的间隔组成，则$\mathscr{F}$称为$G$ 的间隔模型，而如果 $\mathscr{F}$ 由圆上的弧组成，则$\mathscr{F}$称为$G$的圆弧模型。令$G$为包含顶点集$A$和$B$的区间图。如果$G$具有一个区间模型，其中与$A$ 的顶点相对应的所有区间都具有相同的左端点，并且没有其他端点在左侧，并且$B$ 的所有顶点都具有相同的右端点，并且没有其他端点在右侧 (de Figueiredo et al.,1997)。如果对$\{\nu\},\emptyset$是左右，则$G$的顶点$\nu$是$G$ 的结尾。Gimbel (1988) 给出了末端顶点的以下特征
Theorem 2 (Gimbel, 1988). Let G be an interval graph. If $\nu$ is $a$ vertex of $G$,
then vis an end vertex of G if and only if G contains none of the graphs in Figure 3
as an induced subgraph where the filled vertex represents v.
左右对由 de Figueiredo et al.(1997) 表征，如下所示。
<FigureHere>
Figure 2: Some minimal forbidden induced subgraphs for the class of normal Helly
circular-arc graphs
4
|
[
[
0.187255859375,
0.114013671875,
0.81201171875,
0.114013671875,
0.81201171875,
0.127685546875,
0.187255859375,
0.127685546875
],
[
0.189697265625,
0.1312255859375,
0.81201171875,
0.1312255859375,
0.81201171875,
0.144287109375,
0.189697265625,
0.144287109375
],
[
0.189697265625,
0.146728515625,
0.81005859375,
0.146728515625,
0.81005859375,
0.1602783203125,
0.189697265625,
0.1602783203125
],
[
0.18896484375,
0.163330078125,
0.81201171875,
0.163330078125,
0.81201171875,
0.177001953125,
0.18896484375,
0.177001953125
],
[
0.2149658203125,
0.189453125,
0.8134765625,
0.189453125,
0.8134765625,
0.203125,
0.2149658203125,
0.203125
],
[
0.18896484375,
0.2060546875,
0.81005859375,
0.2060546875,
0.81005859375,
0.2197265625,
0.18896484375,
0.2197265625
],
[
0.187255859375,
0.22265625,
0.74462890625,
0.22265625,
0.74462890625,
0.236328125,
0.187255859375,
0.236328125
],
[
0.438232421875,
0.325927734375,
0.54150390625,
0.325927734375,
0.54150390625,
0.333251953125,
0.438232421875,
0.333251953125
],
[
0.5517578125,
0.32666015625,
0.55908203125,
0.32666015625,
0.55908203125,
0.330078125,
0.5517578125,
0.330078125
],
[
0.5576171875,
0.322509765625,
0.646484375,
0.322509765625,
0.646484375,
0.334228515625,
0.5576171875,
0.334228515625
],
[
0.68994140625,
0.336181640625,
0.75732421875,
0.336181640625,
0.75732421875,
0.350830078125,
0.68994140625,
0.350830078125
],
[
0.366943359375,
0.327880859375,
0.385498046875,
0.327880859375,
0.385498046875,
0.330810546875,
0.366943359375,
0.330810546875
],
[
0.21826171875,
0.33740234375,
0.42236328125,
0.33740234375,
0.42236328125,
0.350830078125,
0.21826171875,
0.350830078125
],
[
0.4375,
0.33740234375,
0.53564453125,
0.33837890625,
0.53564453125,
0.349853515625,
0.4375,
0.3486328125
],
[
0.56103515625,
0.33837890625,
0.646484375,
0.33837890625,
0.646484375,
0.349853515625,
0.56103515625,
0.349853515625
],
[
0.19482421875,
0.3623046875,
0.80419921875,
0.3623046875,
0.80419921875,
0.3759765625,
0.19482421875,
0.3759765625
],
[
0.426513671875,
0.341552734375,
0.43994140625,
0.341552734375,
0.43994140625,
0.345703125,
0.426513671875,
0.345703125
],
[
0.21826171875,
0.396728515625,
0.81005859375,
0.396728515625,
0.81005859375,
0.410400390625,
0.21826171875,
0.410400390625
],
[
0.187255859375,
0.410888671875,
0.80615234375,
0.412109375,
0.80615234375,
0.427001953125,
0.187255859375,
0.42578125
],
[
0.187255859375,
0.427001953125,
0.81005859375,
0.427490234375,
0.81005859375,
0.4423828125,
0.187255859375,
0.441162109375
],
[
0.18896484375,
0.44482421875,
0.81005859375,
0.44482421875,
0.81005859375,
0.457763671875,
0.18896484375,
0.457763671875
],
[
0.18896484375,
0.460205078125,
0.81201171875,
0.460205078125,
0.81201171875,
0.473876953125,
0.18896484375,
0.473876953125
],
[
0.18896484375,
0.476806640625,
0.468505859375,
0.476806640625,
0.468505859375,
0.490478515625,
0.18896484375,
0.490478515625
],
[
0.2149658203125,
0.493408203125,
0.81005859375,
0.493408203125,
0.81005859375,
0.50732421875,
0.2149658203125,
0.50732421875
],
[
0.187255859375,
0.5087890625,
0.81005859375,
0.5087890625,
0.81005859375,
0.5224609375,
0.187255859375,
0.5224609375
],
[
0.189697265625,
0.525390625,
0.81005859375,
0.525390625,
0.81005859375,
0.5390625,
0.189697265625,
0.5390625
],
[
0.189697265625,
0.54150390625,
0.81005859375,
0.54150390625,
0.81005859375,
0.5546875,
0.189697265625,
0.5546875
],
[
0.18896484375,
0.55810546875,
0.8095703125,
0.55810546875,
0.8095703125,
0.57177734375,
0.18896484375,
0.57177734375
],
[
0.18896484375,
0.572265625,
0.81005859375,
0.572265625,
0.81005859375,
0.5859375,
0.18896484375,
0.5859375
],
[
0.18896484375,
0.5888671875,
0.59619140625,
0.5888671875,
0.59619140625,
0.6025390625,
0.18896484375,
0.6025390625
],
[
0.214111328125,
0.6142578125,
0.8095703125,
0.615234375,
0.8095703125,
0.6298828125,
0.214111328125,
0.62890625
],
[
0.187255859375,
0.6318359375,
0.81201171875,
0.6318359375,
0.81201171875,
0.646484375,
0.187255859375,
0.646484375
],
[
0.187255859375,
0.6484375,
0.63232421875,
0.6484375,
0.63232421875,
0.662109375,
0.187255859375,
0.662109375
],
[
0.2166748046875,
0.67578125,
0.62646484375,
0.67578125,
0.62646484375,
0.689453125,
0.2166748046875,
0.689453125
],
[
0.334228515625,
0.7236328125,
0.478515625,
0.71435546875,
0.483642578125,
0.74755859375,
0.339111328125,
0.75732421875
],
[
0.5146484375,
0.72265625,
0.6337890625,
0.72265625,
0.6337890625,
0.7490234375,
0.5146484375,
0.7490234375
],
[
0.623046875,
0.732421875,
0.63818359375,
0.732421875,
0.63818359375,
0.7412109375,
0.623046875,
0.7412109375
],
[
0.322509765625,
0.765625,
0.39208984375,
0.76708984375,
0.39208984375,
0.779296875,
0.322509765625,
0.77685546875
],
[
0.4375,
0.7685546875,
0.456787109375,
0.7685546875,
0.456787109375,
0.77783203125,
0.4375,
0.77783203125
],
[
0.5322265625,
0.7685546875,
0.5498046875,
0.7685546875,
0.5498046875,
0.77783203125,
0.5322265625,
0.77783203125
],
[
0.62890625,
0.76708984375,
0.64404296875,
0.76708984375,
0.64404296875,
0.77685546875,
0.62890625,
0.77685546875
],
[
0.6767578125,
0.8154296875,
0.68115234375,
0.8154296875,
0.68115234375,
0.82080078125,
0.6767578125,
0.82080078125
],
[
0.307373046875,
0.8525390625,
0.36279296875,
0.8525390625,
0.36279296875,
0.8642578125,
0.307373046875,
0.8642578125
],
[
0.399658203125,
0.8515625,
0.42236328125,
0.8515625,
0.42236328125,
0.865234375,
0.399658203125,
0.865234375
],
[
0.47607421875,
0.8525390625,
0.5458984375,
0.8525390625,
0.5458984375,
0.8642578125,
0.47607421875,
0.8642578125
],
[
0.58447265625,
0.85498046875,
0.68701171875,
0.85498046875,
0.68701171875,
0.8662109375,
0.58447265625,
0.8662109375
],
[
0.189697265625,
0.876953125,
0.80615234375,
0.87744140625,
0.80615234375,
0.892578125,
0.189697265625,
0.89111328125
],
[
0.189697265625,
0.89453125,
0.330078125,
0.89453125,
0.330078125,
0.90771484375,
0.189697265625,
0.90771484375
],
[
0.491943359375,
0.9404296875,
0.50732421875,
0.9404296875,
0.50732421875,
0.95166015625,
0.491943359375,
0.95166015625
]
] |
[
"9是遗传的。如果H是遗传图类9的禁止诱导子图，则H是9的最小禁止诱",
"导子图，如果H的每个诱导子图都不同于H是9的成员。显然，遗传图类完",
"全由其最小禁止诱导子图决定：图类完全由那些不包含诱导的最小禁止诱导子",
"图的图组成。以下著名的结果给出了间隔图类的最小禁止诱导子图的完整列表",
"Theorem 1 (Lekkerkerker and Boland, 1962). The minimal forbidden induced",
"subgraphs for the class of interval graphs are: bipartite claw, umbrella, k -net for",
"every k ≥2, k -tent for every k ≥3, and Ck for every k ≥ 4 (see Figure 1 ).",
"12...k",
"",
"123...k",
"Ck, k≥ 4",
"",
"bipartiteclawumbrella",
"k-net,k ≥2",
"k-tent, k≥3",
"Figure l: All minimal forbidden induced subgraphs for the class of interval graphs",
"",
"与间隔图的情况相反，禁止诱导子图表征圆弧图的问题仍然存在。圆弧图",
"类的一些最小禁止诱导子图也是G1，G2，G3，G4，多米诺骨牌，G6和Ck",
"对于每个k≥4，其中C*表示由C诱导的图。。对于每个k≥4，由Ck通过",
"添加通用顶点而产生的图形k轮（请参见E图E2）是一个圆弧图，但也是正",
"常类别的最小禁止诱导子图Helly圆弧图。在下面的内容中，我们分别使用net",
"和 tent 作为2-net和3-tent的简写。",
"如果图G是有限族的交图，则称为的交模型。此外，如果由",
"线上的间隔组成，则?称为G的间隔模型，而如果由圆上的弧组成，则",
"称为G的圆弧模型。令G为包含顶点集A和B的区间图。如果G具有一个",
"区间模型，其中与A的顶点相对应的所有区间都具有相同的左端点，并且没有",
"其他端点在左侧，并且B的所有顶点都具有相同的右端点，并且没有其他端点",
"在右侧（de Figueiredoetal.，1997）。如果对{vl,の是左右，则G的顶点v是G",
"的结尾。Gimbel（1988）给出了末端顶点的以下特征",
"Theorem 2 (Gimbel, 1988). Let G be an interval graph. If v is a vertex of G ,",
"then v is an end vertex of G if and only if G contains none of the graphs in Figure 3",
"as an induced subgraph where the filled vertex represents y .",
"左右对由deFigueiredoetal.（1997）表征，如下所示。",
"Λ",
"←",
"",
"G=K2.3",
"G2",
"G3",
"G",
"",
"domino",
"C6",
"C*, k≥4",
"k-wheel, k ≥ 4",
"Figure 2: Some minimal forbidden induced subgraphs for the class of normal Helly",
"circular-arc graphs",
"4"
] |
{
"arxiv_id": "1402.2641",
"language": "zh",
"page_idx": 3
}
|
||
1402.2641-zh-4
|
1402.2641-zh-4.jpg
|
<FigureHere>
Figure 3: Forbidden induced subgraphs characterizing end vertices. Filled vertices
are not end vertices
Theorem 3 (de Figueiredo et al., 1997). Let G be an interval graph and A,B C
$V( G) . Then, A, Bis\textit{left- right if and only if each of the following assertions holds: }$
$( i) \textit{A and B both are cliques. }$
$(ii)EachvertexofA\cup Bisanendvertex.$
$(iiii)$ Given any pair of vertices u and v, both in A or both in B, there is no any
chordless path on four vertices in G , with u and v as internal vertices of the
$path;$
$( iv) \textit{ For each a}\in A$ and $b\in B$, there is no chordless $a,b$ -path in G, together with
a vertex v which is adjacent in G with no vertex on the path.
在下一节中，我们将利用下面的定理，其证明是 Bonomo et al.(2009)中定
理$\boxed{\mathrm{E}}$6 的证明的简单改编。我们给出了经过修改的完整性证明
Theorem 4 (adapted from Theorem 6 of Bonomo et al., 2009). Let G be a graph containing $no$ induced $G_{1}$ $, G_{2}$ $, G_{3}$ $, G_{4}$ $, domino$ $, k$ $- wheel$ for any $k\geq 4$ , $or$ $C_{k}^{* }$ for any k $\geq4.IfC$ is a chordless cycle of G and v $\in V(G)\setminus V(C),then$ the neighbors of $\nu$ $in$ $V( C)$ induce a chordless path in G\:.
$Proof.$ First, notice that $N_G(\nu)\cap V(C)\neq\emptyset$ and $N_G(\nu)\cap V(C)\neq V(C)$ because $G$ contains no induced $k$ -wheel and no induced $C_k^*$ for any $k\geq4$ .Suppose, by the way of contradiction, that the neighbors of $\nu$ in $V(C)$ do not induce a chordless path in $G.$ Then, the nonneighbors of $\nu$ in $V(C)$ do not induce a connected subgraph in $G$ and let $P^1$ and $P^2$ be two components of the subgraph of $G$ induced by the nonneighbors of $\nu$ in $V(C).$ By construction, $P^1$ and $P^2$ are chordless paths and, by symmetry, assume that $|V(P^1)|\geq|V(P^2)|.$ Let $x_1$ and $x_2$ (resp. $y_1$ and $y_2$ )be the neighbors of the endpoints of $P^1$ (resp. $P^2$ )in $V(C).$ Without loss of generality, assume that $x_1,y_1,x_2,y_2$ are labeled in such a way that, in the subgraph of $G$ induced by $V(C)\setminus\left(V(P^{1})\cup V(P^{2})\right),x_{1}$ and $y_2$ are in the same component and also $x_2$ and $y_1$ are in the same component. Let $k=\left|V(P^2)\right|+3$ ; clearly, $k\geq4.$
Suppose first that $x_1=y_2$ and $x_2\neq y_1.$ On the one hand, if $|V(P^1)|=1$ ,then also $|V(P^2)|=1$ and $V(P^1)\cup V(P^2)\cup\{x_1,x_2,y_1,\nu\}$ would induce domino or $G_3$, depending on whether $|V(C)|\geq6$ or not, respectively, a contradiction. On the other hand, if $|V(P^1)|\geq2$ , then $V(P^2)\cup\{y_1,y_2,\nu,x\}$ would induce $C_k^*$ for any vertex $x$ of $P^1$ nonadjacent to $x_1$, a contradiction. These contradictions show, by symmetry, that either $x_1=y_2$ and $x_2=y_1$, or $x_1\neq y_2$ and $x_2\neq y_1.$ Notice that we can assume that $|V(P^1)|\leq2$ because otherwise $V(P^2)\cup\{y_1,y_2,\nu,x\}$ would induce $C_k^*$ in $G$ for any vertex $x$ in $P^1$ nonadjacent to both $x_1$ and $x_2.$ Therefore, if $x_1=y_2$ and $x_2=y_1$ ,then $V(C)\cup\{\nu\}$ would induce $G_1$, or $G_2$, or $G_4$ in $G$, a contradiction. Finally, if $x_1\neq y_2$ and $x_2\neq y_1$ ,then $V(P^2)\cup\{y_1,y_2,v,x\}$ would induce $C_k^*$ for any vertex $x$ in $P^1$, a contradiction. These contradictions arose from assuming that the neighbors of $\nu$ in
口
$V(C)$ do not induce a chordless path in $V(C).$
5
|
[
[
0.341796875,
0.1817626953125,
0.358642578125,
0.1817626953125,
0.358642578125,
0.190673828125,
0.341796875,
0.190673828125
],
[
0.61376953125,
0.1787109375,
0.6767578125,
0.1787109375,
0.6767578125,
0.190673828125,
0.61376953125,
0.190673828125
],
[
0.189697265625,
0.203125,
0.8095703125,
0.203125,
0.8095703125,
0.2161865234375,
0.189697265625,
0.2161865234375
],
[
0.189697265625,
0.2197265625,
0.337646484375,
0.2197265625,
0.337646484375,
0.2327880859375,
0.189697265625,
0.2327880859375
],
[
0.21240234375,
0.25830078125,
0.81005859375,
0.2587890625,
0.81005859375,
0.273681640625,
0.21240234375,
0.2724609375
],
[
0.187255859375,
0.275634765625,
0.791015625,
0.275634765625,
0.791015625,
0.289306640625,
0.187255859375,
0.289306640625
],
[
0.2032470703125,
0.29638671875,
0.418212890625,
0.29638671875,
0.418212890625,
0.311279296875,
0.2032470703125,
0.311279296875
],
[
0.197265625,
0.318359375,
0.50537109375,
0.318359375,
0.50537109375,
0.33203125,
0.197265625,
0.33203125
],
[
0.193115234375,
0.339111328125,
0.81005859375,
0.339111328125,
0.81005859375,
0.352783203125,
0.193115234375,
0.352783203125
],
[
0.2283935546875,
0.35400390625,
0.81201171875,
0.355224609375,
0.81201171875,
0.369384765625,
0.2283935546875,
0.3681640625
],
[
0.2283935546875,
0.371826171875,
0.273681640625,
0.371826171875,
0.273681640625,
0.385986328125,
0.2283935546875,
0.385986328125
],
[
0.19482421875,
0.392578125,
0.81201171875,
0.392578125,
0.81201171875,
0.40673828125,
0.19482421875,
0.40673828125
],
[
0.2283935546875,
0.4091796875,
0.6689453125,
0.4091796875,
0.6689453125,
0.4228515625,
0.2283935546875,
0.4228515625
],
[
0.214111328125,
0.432861328125,
0.81005859375,
0.432861328125,
0.81005859375,
0.446533203125,
0.214111328125,
0.446533203125
],
[
0.187255859375,
0.448974609375,
0.6767578125,
0.448974609375,
0.6767578125,
0.463134765625,
0.187255859375,
0.463134765625
],
[
0.214111328125,
0.471435546875,
0.81005859375,
0.47265625,
0.81005859375,
0.48681640625,
0.214111328125,
0.486328125
],
[
0.187255859375,
0.4892578125,
0.81005859375,
0.4892578125,
0.81005859375,
0.50439453125,
0.187255859375,
0.50439453125
],
[
0.183837890625,
0.50439453125,
0.81201171875,
0.5048828125,
0.81201171875,
0.51953125,
0.183837890625,
0.5185546875
],
[
0.185546875,
0.521484375,
0.48193359375,
0.521484375,
0.48193359375,
0.53515625,
0.185546875,
0.53515625
],
[
0.214111328125,
0.5458984375,
0.81201171875,
0.5458984375,
0.81201171875,
0.56005859375,
0.214111328125,
0.56005859375
],
[
0.185546875,
0.56103515625,
0.81201171875,
0.5625,
0.81201171875,
0.57666015625,
0.185546875,
0.576171875
],
[
0.187255859375,
0.5791015625,
0.81005859375,
0.5791015625,
0.81005859375,
0.5927734375,
0.187255859375,
0.5927734375
],
[
0.185546875,
0.59423828125,
0.81201171875,
0.59423828125,
0.81201171875,
0.60791015625,
0.185546875,
0.60791015625
],
[
0.185546875,
0.60986328125,
0.81201171875,
0.60986328125,
0.81201171875,
0.62353515625,
0.185546875,
0.62353515625
],
[
0.185546875,
0.6259765625,
0.81005859375,
0.6259765625,
0.81005859375,
0.64013671875,
0.185546875,
0.64013671875
],
[
0.187255859375,
0.642578125,
0.81201171875,
0.642578125,
0.81201171875,
0.65673828125,
0.187255859375,
0.65673828125
],
[
0.187255859375,
0.65771484375,
0.81005859375,
0.65771484375,
0.81005859375,
0.6728515625,
0.187255859375,
0.6728515625
],
[
0.185546875,
0.6748046875,
0.81201171875,
0.67333984375,
0.81201171875,
0.6884765625,
0.185546875,
0.689453125
],
[
0.187255859375,
0.6904296875,
0.81201171875,
0.6904296875,
0.81201171875,
0.705078125,
0.187255859375,
0.705078125
],
[
0.187255859375,
0.70703125,
0.6591796875,
0.70703125,
0.6591796875,
0.72021484375,
0.187255859375,
0.72021484375
],
[
0.214111328125,
0.72265625,
0.81201171875,
0.72265625,
0.81201171875,
0.73681640625,
0.214111328125,
0.73681640625
],
[
0.187255859375,
0.7392578125,
0.81005859375,
0.7392578125,
0.81005859375,
0.75439453125,
0.187255859375,
0.75439453125
],
[
0.185546875,
0.7548828125,
0.81201171875,
0.7548828125,
0.81201171875,
0.76953125,
0.185546875,
0.76953125
],
[
0.187255859375,
0.771484375,
0.8134765625,
0.771484375,
0.8134765625,
0.7861328125,
0.187255859375,
0.7861328125
],
[
0.187255859375,
0.78759765625,
0.8134765625,
0.78759765625,
0.8134765625,
0.8017578125,
0.187255859375,
0.8017578125
],
[
0.18896484375,
0.80419921875,
0.8134765625,
0.80419921875,
0.8134765625,
0.81787109375,
0.18896484375,
0.81787109375
],
[
0.18896484375,
0.8193359375,
0.81201171875,
0.8193359375,
0.81201171875,
0.83447265625,
0.18896484375,
0.83447265625
],
[
0.185546875,
0.83447265625,
0.81201171875,
0.8349609375,
0.81201171875,
0.849609375,
0.185546875,
0.8486328125
],
[
0.183837890625,
0.85107421875,
0.81005859375,
0.8515625,
0.81005859375,
0.8662109375,
0.183837890625,
0.865234375
],
[
0.187255859375,
0.8681640625,
0.78857421875,
0.8681640625,
0.78857421875,
0.8818359375,
0.187255859375,
0.8818359375
],
[
0.78857421875,
0.87158203125,
0.8095703125,
0.87158203125,
0.8095703125,
0.8798828125,
0.78857421875,
0.8798828125
],
[
0.18896484375,
0.8857421875,
0.81005859375,
0.8857421875,
0.81005859375,
0.896484375,
0.18896484375,
0.896484375
],
[
0.187255859375,
0.90087890625,
0.5205078125,
0.90087890625,
0.5205078125,
0.91455078125,
0.187255859375,
0.91455078125
],
[
0.80029296875,
0.90185546875,
0.81005859375,
0.90185546875,
0.81005859375,
0.9091796875,
0.80029296875,
0.9091796875
],
[
0.490234375,
0.939453125,
0.5087890625,
0.939453125,
0.5087890625,
0.95361328125,
0.490234375,
0.95361328125
]
] |
[
"Fo",
"Fs,S≥2",
"Figure 3: Forbidden induced subgraphs characterizing end vertices. Filled vertices",
"are not end vertices",
"Theorem 3 (de Figueiredo et al., 1997). Let G be an interval graph and A,B C",
"V(G) . Then, A, B is left-right if and only if each of the following assertions holds:",
"(i) A and B both are cliques.",
"(ii) Each vertex of AUB is an end vertex.",
"(ii) Given any pair of vertices u and v, both in A or both in B, there is no any",
"chordless path on four vertices in G, with u and v as internal vertices of the",
"path;",
"(iv) For each a EA and b E B, there is no chordless a,b -path in G, together with",
"a vertex v which is adjacent in G with no vertex on the path.",
"在下二节中，我们将利用下面的定理，其证明是Bonomoetal.（2009）中定",
"理E6的证明的简单改编。我们给出了经过修改的完整性证明",
"Theorem 4 (adapted from Theorem 6 of Bonomo et al., 2009). Let G be a graph",
"containing no induced Gi , G2, G3, G4, domino, k -wheel for any k ≥ 4, or C for",
"any k ≥ 4. IfC is a chordless cycle of G and vE V(G) /V(C), then the neighbors of",
"v in V(C) induce a chordless path in G.",
"Proof. First, notice that Nc(v) nV(C) + 0 and Nc(v) nV(C) + V(C) because",
"G contains no induced k -wheel and no induced C* for any k ≥4. Suppose, by the",
"way of contradiction, that the neighbors of v in V(C) do not induce a chordless path",
"in G . Then, the nonneighbors of v in V(C) do not induce a connected subgraph",
"in G and let pl and p2 be two components of the subgraph of G induced by the",
"nonneighbors of v in V(C) . By construction, pl and p? are chordless paths and, by",
"symmetry, assume that |V(pl)/ ≥ [V(p2)/ . Let xi and x2 (resp. yi and y2 ) be the",
"neighbors of the endpoints of pl (resp.p? ) in V(C) : Without loss of generality,",
"assume that x1, y1,x2, y2 are labeled in such a way that, in the subgraph of G induced",
"by V(C) / (V(Pl)UV(p2)) , xi and y2 are in the same component and also x2 and y1",
"are in the same component. Let k = [V(p2)/ +3 ; clearly, k ≥ 4 .",
"Suppose first that xi = y2 and x2 ± y1 . On the one hand, if JV(pl)l = 1 , then",
"also |V(P2)| = 1 and V(Pl)UV(P2)U (x1,x2,y1,v) would induce domino or G3 ",
"depending on whether v(C)] ≥ 6 or not, respectively, a contradiction. On the other",
"hand, if |V(pl)l ≥ 2 , then V(p2) U [y1, y2, v,x) would induce C* for any vertex x of",
"pl nonadjacent to x1 , a contradiction. These contradictions show, by symmetry, that",
"either xi = y2 and x2 = y1 , or xi ≠ y2 and x2 ≠ y1 . Notice that we can assume that",
"IV(Pl)l ≤ 2 because otherwise V(P2) U (y1,y2, v,x) would induce C* in G for any",
"vertex x in pl nonadjacent to both xi and x2 : Therefore, if xi = y2 and x2 = y1, then",
"V(C)U{v) would induce G1 , or G2, or G4 in G, a contradiction. Finally, if x1 + y2",
"and x2 y1 , then V(P2)U(y1,y2, v,x) would induce C* for any vertex x in pl",
",a",
"contradiction.These contradictions arosefrom assuming thattheneighbors of y in",
"V(C) do not induce a chordless path in V(C)",
"",
"5"
] |
{
"arxiv_id": "1402.2641",
"language": "zh",
"page_idx": 4
}
|
||
2104.06903-zh-2
|
2104.06903-zh-2.jpg
|
<FigureHere>
Figure 3: In a line set, a pair of lines can be harmonious and draw visual attention to meaningful regions as in (b). In contrast, they can be redundant or inharmonious as in (c).
输出。在$\mathbb{E}[13]$ 中，构建了区域间亲和力图，以将车道之间的结构关系从教师网络转移到学生网络。在$\mathbb{E}[25]$ 中，为了实现较高的处理速度，开发了一个网络来仅在预定义的行集上识别每个通道的位置
## $3.{\mathrm{Proposed~Algorithm}}$
图$\boxed{\mathrm{E}}2$ 是所提出算法的概述，其中包含 S-Net 和 HNet。首先，给定一幅图像和一组候选线，S-Net 计算每个候选的线概率和回归偏移。其次，通过选择和删除过程过滤掉不相关的候选人。第三，构建一个完整的图，其节点集由选定的线组成，并通过 H-Net 计算其边权重。最后，确定代表和谐语义线的最大权重集团。
## 3.1. Problem formulation
如果图像中的语义线和谐地传达图像的组成，则可以将其视为最佳。换句话说，在一个最佳集合中，每对语义线都应该相互协调。如图 3(b)所示，一对语义线应该将视觉注意力引向有意义的区域。相反，在图 3(c)中，两条线是冗余的或不和谐的。基于这一观察，我们将语义线检测公式化为在完整图中找到最大权重集团$\mathbb{E}[3,8]$ 。在完整的图中，检测到的线形成节点集，每个边权重表示关联的两条线的和谐程度。因此，通过找到最大权重集团，我们找到了一组最佳的和谐语义线
3.2. Node selection: filtering line candidates
为所有候选线构建一个完整的图在计算上是不可行
的。因此，我们仅通过过滤候选线来选择可靠节点
候选线生成：候选线是图像中的端到端直线，可以通过 Hough 空间中的极坐标来参数化$\mathbb{E}[10,17,22]$。令$\boldsymbol{l}=(\rho,\varphi)$表示一条线，其中$\dot{\rho }$ 是 它 与 图 像 中 心 的 距 离 , $\varphi$是它与$x$轴的角度。然后，我们通过统一量化$\rho$和$\varphi$来生成$N$候选行，用$l_n= ( \rho _n, \varphi _n)$ , $1\leq n\leq N$表示
S-Net:对于每个候选行，我们计算其分类概率和回归偏移。为此，我们开发了基于传统线探测器[10,16,19] 的 S-Net。图 4 (a)显示了 S-Net 的体系结构。S-Net
<FigureHere>
Figure 4: Architecture of S-Net and H-Net: (a) S-Net takes an image and line candidates and extracts the line feature map. The classification probabilities and regression offsets of the line candidates are then computed by two fully connected layers. (b) H-Net takes an image and a line pair $(i,j)$ to extract local and interregion features. Two types of regression layers are used to compute the harmony score $h_{ij}$ .
从图像中提取卷积特征图$X=[X^1,X^2,\ldots,X^C]\in$ $\mathbb{R} ^{H\times W\times C}$ ,其中$H$ $, W$和$C$表示特征高度，特征宽度和通道数。然后，通过对$l_n$上的像素特征进行平均来获得线特征图$Y=[Y^1,Y^2,\ldots,Y^C]\in\mathbb{R}^{N\times C};$对于$1\leq n\leq N$和$1\leq c\leq C$ ,
(1)
$$Y^c(n)=\frac{1}{|\boldsymbol{l}_n|}\sum_{\boldsymbol{p}\in\boldsymbol{l}_n}X^c(\boldsymbol{p})$$
,其中$|l_n|$表示沿$l_n$的像素数。然后，我们通过
$$P=\sigma(f_1(Y))\quad\text{and}\quad O=f_2(Y)$$
(2)
程
获得概率向量$P$和行偏移矩阵$O$,其中$f_1$和$f_2$是大小为$C\times1$和$C\times2$的全连接层，分别用于分类和回归，并且$H$。对于第$n$行候选$l_n= ( \rho _n, \varphi _n)$ , $P_n$ 表示它是语义的概率，而$O_n=\Delta\boldsymbol{l}_n=(\Delta\rho_n,\Delta\varphi_n)$ 是第$\mathbb{E}3.4$ 节中的行细化的偏移向量
补充文档中详细描述了 S-Net 的体系结构和培训过选择和删除：在传统算法[10,16,19]中，为了检测语义行，仅选择概率高于阈值的候选行，然后进行后处理(e.g.非最大抑制)。然而，这可能会导致假阴性，因为它是隐式的，但仍然是语义的，所以概率很低。为了减少这种假阴性，而不是阈值化，我们在图$\color{red}{\mathrm{E2}}$(b)中执行选择和删除过程。我们通过
$$i^\star=\arg\max_iP_i$$
(3)
选择最可靠的行$l_{i^\star}$ ,然后删除与所选行重叠的行。具体来说，我们在 Hough 空间$\mathbb{E}[10,22]$中以$l_i^*$为中心
|
[
[
0.8447265625,
0.10919189453125,
0.859375,
0.10919189453125,
0.859375,
0.11614990234375,
0.8447265625,
0.11614990234375
],
[
0.525390625,
0.1307373046875,
0.564453125,
0.1307373046875,
0.564453125,
0.1351318359375,
0.525390625,
0.1351318359375
],
[
0.84228515625,
0.1383056640625,
0.859375,
0.14013671875,
0.85693359375,
0.146484375,
0.8388671875,
0.14453125
],
[
0.53515625,
0.149658203125,
0.560546875,
0.149658203125,
0.560546875,
0.1578369140625,
0.53515625,
0.1578369140625
],
[
0.650390625,
0.1558837890625,
0.70166015625,
0.1558837890625,
0.70166015625,
0.1634521484375,
0.650390625,
0.1634521484375
],
[
0.74267578125,
0.154052734375,
0.814453125,
0.154052734375,
0.814453125,
0.1654052734375,
0.74267578125,
0.1654052734375
],
[
0.6689453125,
0.17236328125,
0.72216796875,
0.17041015625,
0.7236328125,
0.1849365234375,
0.6708984375,
0.1875
],
[
0.51123046875,
0.1590576171875,
0.57861328125,
0.1590576171875,
0.57861328125,
0.169189453125,
0.51123046875,
0.169189453125
],
[
0.096435546875,
0.1767578125,
0.111083984375,
0.1767578125,
0.111083984375,
0.179931640625,
0.096435546875,
0.179931640625
],
[
0.193603515625,
0.175537109375,
0.2115478515625,
0.175537109375,
0.2115478515625,
0.179931640625,
0.193603515625,
0.179931640625
],
[
0.10296630859375,
0.1976318359375,
0.158447265625,
0.1976318359375,
0.158447265625,
0.2083740234375,
0.10296630859375,
0.2083740234375
],
[
0.236083984375,
0.1976318359375,
0.313720703125,
0.1976318359375,
0.313720703125,
0.2083740234375,
0.236083984375,
0.2083740234375
],
[
0.37744140625,
0.1976318359375,
0.4599609375,
0.1976318359375,
0.4599609375,
0.2083740234375,
0.37744140625,
0.2083740234375
],
[
0.74267578125,
0.2008056640625,
0.7890625,
0.2008056640625,
0.7890625,
0.2083740234375,
0.74267578125,
0.2083740234375
],
[
0.73046875,
0.2083740234375,
0.7890625,
0.2083740234375,
0.7890625,
0.21533203125,
0.73046875,
0.21533203125
],
[
0.80224609375,
0.20263671875,
0.833984375,
0.20263671875,
0.833984375,
0.2139892578125,
0.80224609375,
0.2139892578125
],
[
0.08087158203125,
0.2203369140625,
0.46728515625,
0.2203369140625,
0.46728515625,
0.23486328125,
0.08087158203125,
0.23486328125
],
[
0.736328125,
0.222900390625,
0.79833984375,
0.222900390625,
0.79833984375,
0.2342529296875,
0.736328125,
0.2342529296875
],
[
0.7998046875,
0.222900390625,
0.833984375,
0.222900390625,
0.833984375,
0.2342529296875,
0.7998046875,
0.2342529296875
],
[
0.5146484375,
0.2310791015625,
0.5703125,
0.2310791015625,
0.5703125,
0.238037109375,
0.5146484375,
0.238037109375
],
[
0.64794921875,
0.23291015625,
0.66259765625,
0.23291015625,
0.66259765625,
0.2374267578125,
0.64794921875,
0.2374267578125
],
[
0.8388671875,
0.2298583984375,
0.85107421875,
0.2298583984375,
0.85107421875,
0.238037109375,
0.8388671875,
0.238037109375
],
[
0.08087158203125,
0.236083984375,
0.468017578125,
0.236083984375,
0.468017578125,
0.25,
0.08087158203125,
0.25
],
[
0.73193359375,
0.238037109375,
0.8037109375,
0.238037109375,
0.8037109375,
0.25,
0.73193359375,
0.25
],
[
0.79833984375,
0.2392578125,
0.83251953125,
0.2374267578125,
0.833984375,
0.248779296875,
0.7998046875,
0.250732421875
],
[
0.08087158203125,
0.251953125,
0.465576171875,
0.251953125,
0.465576171875,
0.26513671875,
0.08087158203125,
0.26513671875
],
[
0.64892578125,
0.25390625,
0.70361328125,
0.255615234375,
0.70361328125,
0.267578125,
0.64892578125,
0.26513671875
],
[
0.51123046875,
0.260009765625,
0.5810546875,
0.260009765625,
0.5810546875,
0.271484375,
0.51123046875,
0.271484375
],
[
0.080078125,
0.267578125,
0.2091064453125,
0.267578125,
0.2091064453125,
0.281005859375,
0.080078125,
0.281005859375
],
[
0.6689453125,
0.276611328125,
0.7236328125,
0.276611328125,
0.7236328125,
0.28857421875,
0.6689453125,
0.28857421875
],
[
0.080078125,
0.309326171875,
0.46728515625,
0.30810546875,
0.46728515625,
0.322509765625,
0.080078125,
0.323974609375
],
[
0.50341796875,
0.3037109375,
0.890625,
0.3037109375,
0.890625,
0.318115234375,
0.50341796875,
0.318115234375
],
[
0.08251953125,
0.324462890625,
0.46728515625,
0.324462890625,
0.46728515625,
0.33837890625,
0.08251953125,
0.33837890625
],
[
0.50244140625,
0.3193359375,
0.890625,
0.318115234375,
0.890625,
0.333251953125,
0.50244140625,
0.333984375
],
[
0.08251953125,
0.34033203125,
0.468017578125,
0.34033203125,
0.468017578125,
0.354248046875,
0.08251953125,
0.354248046875
],
[
0.50341796875,
0.335205078125,
0.890625,
0.335205078125,
0.890625,
0.348388671875,
0.50341796875,
0.348388671875
],
[
0.08251953125,
0.356689453125,
0.34228515625,
0.356689453125,
0.34228515625,
0.367431640625,
0.08251953125,
0.367431640625
],
[
0.50244140625,
0.35107421875,
0.88818359375,
0.349853515625,
0.88818359375,
0.363037109375,
0.50244140625,
0.3642578125
],
[
0.50341796875,
0.365478515625,
0.8896484375,
0.365478515625,
0.8896484375,
0.37890625,
0.50341796875,
0.37890625
],
[
0.50341796875,
0.380126953125,
0.88623046875,
0.380126953125,
0.88623046875,
0.39453125,
0.50341796875,
0.39453125
],
[
0.07843017578125,
0.386962890625,
0.281005859375,
0.386962890625,
0.281005859375,
0.40087890625,
0.07843017578125,
0.40087890625
],
[
0.50341796875,
0.395751953125,
0.890625,
0.395751953125,
0.890625,
0.410400390625,
0.50341796875,
0.410400390625
],
[
0.10296630859375,
0.410888671875,
0.46728515625,
0.410888671875,
0.46728515625,
0.425537109375,
0.10296630859375,
0.425537109375
],
[
0.50341796875,
0.410400390625,
0.76318359375,
0.410400390625,
0.76318359375,
0.4248046875,
0.50341796875,
0.4248046875
],
[
0.08087158203125,
0.4267578125,
0.46728515625,
0.4267578125,
0.46728515625,
0.43994140625,
0.08087158203125,
0.43994140625
],
[
0.08087158203125,
0.442626953125,
0.46728515625,
0.442626953125,
0.46728515625,
0.455810546875,
0.08087158203125,
0.455810546875
],
[
0.08087158203125,
0.458251953125,
0.468017578125,
0.458251953125,
0.468017578125,
0.4716796875,
0.08087158203125,
0.4716796875
],
[
0.50341796875,
0.45263671875,
0.88818359375,
0.45263671875,
0.88818359375,
0.466552734375,
0.50341796875,
0.466552734375
],
[
0.5,
0.462158203125,
0.8935546875,
0.46533203125,
0.8935546875,
0.486083984375,
0.5,
0.48291015625
],
[
0.080078125,
0.4716796875,
0.468017578125,
0.47021484375,
0.468017578125,
0.48486328125,
0.080078125,
0.486083984375
],
[
0.08087158203125,
0.4873046875,
0.464111328125,
0.4873046875,
0.464111328125,
0.50048828125,
0.08087158203125,
0.50048828125
],
[
0.5048828125,
0.484130859375,
0.8896484375,
0.484130859375,
0.8896484375,
0.49755859375,
0.5048828125,
0.49755859375
],
[
0.5009765625,
0.49365234375,
0.890625,
0.49365234375,
0.890625,
0.5146484375,
0.5009765625,
0.5146484375
],
[
0.07843017578125,
0.51220703125,
0.281005859375,
0.51318359375,
0.281005859375,
0.52783203125,
0.07843017578125,
0.5263671875
],
[
0.50341796875,
0.5146484375,
0.6953125,
0.5146484375,
0.6953125,
0.52783203125,
0.50341796875,
0.52783203125
],
[
0.101318359375,
0.537109375,
0.46728515625,
0.537109375,
0.46728515625,
0.55029296875,
0.101318359375,
0.55029296875
],
[
0.60107421875,
0.5390625,
0.7900390625,
0.54052734375,
0.7900390625,
0.5595703125,
0.60107421875,
0.55810546875
],
[
0.86669921875,
0.54052734375,
0.8935546875,
0.54052734375,
0.8935546875,
0.556640625,
0.86669921875,
0.556640625
],
[
0.08251953125,
0.55322265625,
0.46728515625,
0.55322265625,
0.46728515625,
0.56640625,
0.08251953125,
0.56640625
],
[
0.08087158203125,
0.568359375,
0.468017578125,
0.568359375,
0.468017578125,
0.58203125,
0.08087158203125,
0.58203125
],
[
0.50244140625,
0.5712890625,
0.85546875,
0.5712890625,
0.85546875,
0.58544921875,
0.50244140625,
0.58544921875
],
[
0.08416748046875,
0.58349609375,
0.468017578125,
0.58349609375,
0.468017578125,
0.5966796875,
0.08416748046875,
0.5966796875
],
[
0.08087158203125,
0.5986328125,
0.453369140625,
0.5986328125,
0.453369140625,
0.6123046875,
0.08087158203125,
0.6123046875
],
[
0.5810546875,
0.599609375,
0.81640625,
0.599609375,
0.81640625,
0.6142578125,
0.5810546875,
0.6142578125
],
[
0.86767578125,
0.60107421875,
0.890625,
0.60107421875,
0.890625,
0.6142578125,
0.86767578125,
0.6142578125
],
[
0.08087158203125,
0.61279296875,
0.46728515625,
0.61279296875,
0.46728515625,
0.626953125,
0.08087158203125,
0.626953125
],
[
0.08251953125,
0.62744140625,
0.46728515625,
0.62744140625,
0.46728515625,
0.6416015625,
0.08251953125,
0.6416015625
],
[
0.5048828125,
0.62744140625,
0.8896484375,
0.62744140625,
0.8896484375,
0.642578125,
0.5048828125,
0.642578125
],
[
0.08251953125,
0.64306640625,
0.46728515625,
0.64306640625,
0.46728515625,
0.6572265625,
0.08251953125,
0.6572265625
],
[
0.5048828125,
0.64306640625,
0.890625,
0.64306640625,
0.890625,
0.6572265625,
0.5048828125,
0.6572265625
],
[
0.08087158203125,
0.65771484375,
0.468017578125,
0.65771484375,
0.468017578125,
0.671875,
0.08087158203125,
0.671875
],
[
0.5048828125,
0.6591796875,
0.8896484375,
0.6591796875,
0.8896484375,
0.673828125,
0.5048828125,
0.673828125
],
[
0.08087158203125,
0.6748046875,
0.1478271484375,
0.6748046875,
0.1478271484375,
0.68603515625,
0.08087158203125,
0.68603515625
],
[
0.5048828125,
0.67236328125,
0.890625,
0.673828125,
0.890625,
0.68798828125,
0.5048828125,
0.6875
],
[
0.50341796875,
0.6875,
0.744140625,
0.6875,
0.744140625,
0.70068359375,
0.50341796875,
0.70068359375
],
[
0.080078125,
0.69970703125,
0.420654296875,
0.69970703125,
0.420654296875,
0.71337890625,
0.080078125,
0.71337890625
],
[
0.52294921875,
0.705078125,
0.89208984375,
0.705078125,
0.89208984375,
0.71826171875,
0.52294921875,
0.71826171875
],
[
0.101318359375,
0.7236328125,
0.46728515625,
0.7236328125,
0.46728515625,
0.7373046875,
0.101318359375,
0.7373046875
],
[
0.5048828125,
0.7216796875,
0.521484375,
0.7216796875,
0.521484375,
0.73095703125,
0.5048828125,
0.73095703125
],
[
0.08087158203125,
0.7392578125,
0.441162109375,
0.7392578125,
0.441162109375,
0.75244140625,
0.08087158203125,
0.75244140625
],
[
0.50634765625,
0.7412109375,
0.88818359375,
0.7412109375,
0.88818359375,
0.751953125,
0.50634765625,
0.751953125
],
[
0.08087158203125,
0.76025390625,
0.468017578125,
0.76025390625,
0.468017578125,
0.7734375,
0.08087158203125,
0.7734375
],
[
0.50244140625,
0.7548828125,
0.8896484375,
0.75390625,
0.8896484375,
0.76806640625,
0.50244140625,
0.76953125
],
[
0.08251953125,
0.7763671875,
0.46728515625,
0.7763671875,
0.46728515625,
0.787109375,
0.08251953125,
0.787109375
],
[
0.50244140625,
0.77099609375,
0.8896484375,
0.77099609375,
0.8896484375,
0.7841796875,
0.50244140625,
0.7841796875
],
[
0.080078125,
0.79052734375,
0.46728515625,
0.79052734375,
0.46728515625,
0.8037109375,
0.080078125,
0.8037109375
],
[
0.5048828125,
0.78515625,
0.890625,
0.78515625,
0.890625,
0.798828125,
0.5048828125,
0.798828125
],
[
0.50341796875,
0.7998046875,
0.890625,
0.798828125,
0.890625,
0.81298828125,
0.50341796875,
0.814453125
],
[
0.080078125,
0.80615234375,
0.468017578125,
0.80615234375,
0.468017578125,
0.8193359375,
0.080078125,
0.8193359375
],
[
0.08087158203125,
0.82080078125,
0.468017578125,
0.82080078125,
0.468017578125,
0.83544921875,
0.08087158203125,
0.83544921875
],
[
0.50341796875,
0.81494140625,
0.74169921875,
0.81494140625,
0.74169921875,
0.8291015625,
0.50341796875,
0.8291015625
],
[
0.080078125,
0.8359375,
0.1004638671875,
0.8359375,
0.1004638671875,
0.8486328125,
0.080078125,
0.8486328125
],
[
0.63818359375,
0.8447265625,
0.7548828125,
0.8447265625,
0.7548828125,
0.85791015625,
0.63818359375,
0.85791015625
],
[
0.86767578125,
0.8447265625,
0.89208984375,
0.8447265625,
0.89208984375,
0.85791015625,
0.86767578125,
0.85791015625
],
[
0.080078125,
0.85693359375,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.87060546875
],
[
0.08087158203125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88525390625,
0.08087158203125,
0.88525390625
],
[
0.50341796875,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.080078125,
0.88623046875,
0.468017578125,
0.88525390625,
0.468017578125,
0.8994140625,
0.080078125,
0.90087890625
],
[
0.50341796875,
0.88720703125,
0.89208984375,
0.88720703125,
0.89208984375,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"ls",
"",
"Reg",
"Image",
"Feature map",
"Line feature map",
"(a) S-Net",
"Line candidates",
"",
"",
"(a) Line set",
"(b) Positive pair",
"() Negative pair",
"Inter-region",
"correlation",
"Regl",
"Figure 3: In a line set, a pair of lines can be harmo-",
"Line pooling",
"Reg2",
"E",
"",
"+",
"nious and draw visual attention to meaningful regions",
"Line pooling",
"Reg2",
"as in (b). In contrast, they can be redundant or inhar",
"Feature map",
"Line pair (t,)",
"monious as in (c).",
"(b) H-Net",
"输出。在E[13]中，构建了区域间亲和力图，以将车道",
"Figure 4: Architecture of S-Net and H-Net: (a) S-Net",
"之间的结构关系从教师网络转移到学生网络。在E25",
"takes an image and line candidates and extracts the",
"中，为了实现较高的处理速度，开发了一个网络来仅在",
"line feature map.The classification probabilities and",
"预定义的行集上识别每个通道的位置",
"regression offsets of the line candidates are then com",
"puted by two fully connected layers. (b) H-Net takes",
"an image and a line pair (i, j) to extract local and inter",
"3. Proposed Algorithm",
"region features. Two types of regression layers are used",
"图E2是所提出算法的概述，其中包含S-Net和H-",
"to compute the harmony score hij :",
"Net。首先，给定一幅图像和一组候选线，S-Net计算",
"每个候选的线概率和回归偏移。其次，通过选择和删除",
"过程过滤掉不相关的候选人。第三，构建一个完整的",
"从图像中提取卷积特征图X=[X1,X?,.,XC1e",
"RHxWxC，其中H，W和C表示特征高度，特征宽",
"图，其节点集由选定的线组成，并通过H-Net计算其",
"边权重。最后，确定代表和谐语义线的最大权重集团。",
"度和通道数。然后，通过对ln上的像素特征进行平均",
"来获得线特征图Y-[Y,Y?,YCjERNxC，对于",
"3.1. Problem formulation",
"1≤n≤N和1≤c≤C,",
"如果图像中的语义线和谐地传达图像的组成，则可",
"Yc(n)=peln Xc(p)",
"(1)",
"以将其视为最佳。换句话说，在一个最佳集合中，每",
"对语义线都应该相互协调。如图3（b）所示，一对语",
"，其中nl表示沿l的像素数。然后，我们通过",
"义线应该将视觉注意力引向有意义的区域。相反，在",
"图3（c）中，两条线是穴余的或不和谐的。基于这",
"P= α(fi(Y)) and O= f2(Y)",
"(2)",
"观察，我们将语义线检测公式化为在完整图中找到最",
"大权重集团E3.81。在完整的图中，检测到的线形成",
"获得概率向量P和行偏移矩阵O，其中fi1和f2是",
"节点集，每个边权重表示关联的两条线的和谐程度。因",
"大小为C×1和C×2的全连接层，分别用于分类和",
"此，通过找到最大权重集团，我们找到了一组最佳的和",
"回归，并且H。对于第n行候选ln=（pn，Pn），Pn",
"谐语义线",
"表示它是语义的概率，而On=△n=（△pn，△pn）是",
"第E3.4节中的行细化的偏移向量",
"3.2. Node selection: filtering line candidates",
"补充文档中详细描述了S-Net的体系结构和培训过",
"为所有候选线构建一个完整的图在计算上是不可行",
"程",
"的。因此，我们仅通过过滤候选线来选择可靠节点",
"选择和删除：在传统算法「10.16，191中，为了检测语义",
"候选线生成：候选线是图像中的端到端直线，可以通",
"行，仅选择概率高于阈值的候选行，然后进行后处理",
"过Hough空间中的极坐标来参数化F10,17,22。令",
"(e.g.非最大抑制）。然而，这可能会导致假阴性，因为",
"l=（e,）表示一条线，其中p是它与图像中心的距离",
"它是隐式的，但仍然是语义的，所以概率很低。为了减",
"少这种假阴性，而不是阈值化，我们在图E2（b）中",
"是它与a轴的角度。然后，我们通过统一量化p和",
"来生成N候选行，用ln=（pn，Pn），1≤n≤N表",
"执行选择和删除过程。。我们通过",
"示",
"i* =argmax, Pi",
"(3)",
"S-Net：对于每个候选行，我们计算其分类概率和回归",
"偏移。为此，我们开发了基于传统线探测器[10,16,19]",
"选择最可靠的行L，然后删除与所选行重叠的行。具",
"的 S-Net。图 4（a）显示了 S-Net的体系结构。S-Net",
"体来说，我们在Hough空间E[10,22]中以li*为中心"
] |
{
"arxiv_id": "2104.06903",
"language": "zh",
"page_idx": 2
}
|
||
2104.06903-zh-3
|
2104.06903-zh-3.jpg
|
的$5\times5$网格中删除了 24 行。我们执行此过程$K$次以组成$K$选定行的节点集。图 5(b)和(e)分别在图像和 Hough 空间上显示了这样的选定行
3.3. Edge weighting: harmony score estimation
区域间相关性：为了区分图$\color{red}{\mathrm{E3}}$(b)中的正对与图$\color{red}{\mathrm{E3}}$ (c)中的负对，我们设计了区域间相关性(IRC)模块， 该模块分析由一对线分隔的区域
令$\mathcal{R} _i$ , $1\leq i\leq M$表示由两行分隔的区域。可以有
三个或四个区域，i.e. $\operatorname{E}M=3$或 4 。我们通过
(4)
$$r_i=\frac{1}{|\mathcal{R}_i|}\sum_{\boldsymbol{p}\in\mathcal{R}_i}X(\boldsymbol{p}).$$
提取$\mathcal{R}_i$的区域特征向量$r_i$我们计算区域$|\mathcal{R}_i|$的softmax 概率$a_i$来缩放区域特征向量，然后将缩放后的向量连接成大小为$C\times4$的
$$R=[a_1\mathrm r_1,a_2\mathrm r_2,a_3\mathrm r_3,a_4\mathrm r_4]$$
(5)
。如果$M=3$ ,我们用零填充最右边的向量。然后，$R$
被馈送到完全连接的层以产生 IRC 特征
H-Net:我们使用 IRC 模块开发 H-Net。它需要一个图像和一对线，由$i$和$j$索引，以产生范围从 0 到 1 的和声分数$h_{ij}$ 。图 4(b)显示了 H-Net 架构。VGG16 [26] 的卷积层被用作特征提取器，其后是 IRC 模块和线池化层的三个并行分支。我们使用行池化层在(1)中分别对行$i$和$j$执行池化。我们使用两种类型的回归层： 一种用于产生两行的 IRC 分数 (Reg1), 另一种用于计算每行的一元可靠性 (Reg2)。最后，我们通过将 IRC 得分与一元可靠性水平的平均值相乘来计算和声得分$h_{ij}$
我们为 H-Net 配置训练数据如下。假设图像中的每对基本事实语义线彼此协调。因此，我们将这样的对声明为正，而将其他对声明为负。换句话说，只有当$i$和$j$两行都是语义时，行对$(i,j)$才是正的。然后，根据对$(i,j)$是否为正，将和声得分$\bar{h}_{ij}$注释为 1或 0。但是，正对的这种严格定义会导致类不平衡：正对太少。因此，我们干扰每个正对的行位置，并注释相应的和声分数$\bar{h}_{ij}$与$e^-(d_i^2+d_j^2)$成比例，其中$d_i$和$d_j$表示行$i$ 和$j$的干扰。此外，用于训练 H-Net 的损失函数被定义为$\ell _{\mathrm{H} }= ( h_{ij}- \bar{\bar{h} } _{ij}) ^{2}$ ,其中$\bar{h}_{ij}$是地面实况和声得分，$h_{ij}$是其估计值。补充文档更详细地描述了 H-Net 的培训过程和架构
3.4. Graph optimization: finding harmonious lines
图构造：我们构造了一个完整的图$G=(\mathcal{V},\mathcal{E})$ ,其中节点集$\mathcal{V}=\{v_1,v_2,\ldots,v_K\}$表示在$\mathbb{E}\color{red}{3.2}$节中使用 S-Net 选择的$K$行。每对线由边缘集$\mathcal{E}=\{(v_i,v_j):i\neq j\}$ 中的一个边缘连接。在 3.3节中，H-Net 为每个边分配一个权重$w(v_i,v_j)=h_{ij}$ 。图 5(f)可视化了一个完整的加权图
MWCS:如前所述，一组语义行是最佳的，如果该组中的任何两行彼此和谐。因此，找到这样的最优集等同于
找到一组节点$\mathbb{E}[8]$ ,它们相互连接并具有最大的权重
和 (i.e. 和声分数)
令$\theta$表示一个集团，由成员节点的索引集表示。然
后，我们将集团$\theta$的协调能量$E_\text{harmony}(\theta)$定义为
$$E_{\text{harmony}}(\theta)=\sum_{i\in\theta}\sum_{j\in\theta,j>i}w(v_i,v_j)$$
(6)
,它是$\theta$中所有边权重的总和。找到最大化这个能量的集团是 NP-hard [6]。然而，在这项工作中，$K$被设置为一个小数字。默认的$K$是 8。大约有$2^K$个可能的集团，这些集团也是可以管理的。因此，采用穷举搜索来找到最大权重集团。首先，我们在图$G$中生成可能的集团$\Theta=\{\theta_t\}$的集合，其中每个集团$\theta_t$由两个以上的节点组成。然后，我们选择最大化协调能量的最大权重集团$\theta^\star:$
(7)
$$\theta^\star=\arg\max_{\theta_t\in\Theta}E_{\mathrm{harmony}}(\theta_t)$$
受约束
$$\min_{i,j\in\theta}w(v_i,v_j)>\kappa $$
(8)
,其中$\kappa$是阈值。如果没有满足约束的集团，则通过
$$i^\star=\arg\max_ih_{ii}.$$
(9)
选择最大的单节点集团$\theta^\star=\{i^\star\}$ 。通过将与重复输入
相同的行应用于 H-Net 来获得自和得分$h_{ii}$
在获得一组和谐的语义线之后，我们通过
(10)
$$l_{v_i}+\Delta l_{v_i}$$
来细化每一行，其中$\Delta l_{v_i}$表示偏移向量，由 S-Net 的回归层生成。图 5(c)和(g)显示了图像和 Hough 空间上的一组和谐语义线
4. Experimental Results
## $4. 1. \textbf{Datasets}$
SEL$\underline{\mathrm{E}}[19]:$它是第一个语义线数据集，包含 1,750 个室外图像，这些图像被拆分为 1,575 个训练图像和 175个测试图像。每条语义线由图像边界上两个端点的坐标注释
SEL\_Hard [16]:它是用于测试语义行检测器的更具挑战性的数据集。它包含 300 个测试图像，选自 ADE20K 分割数据集[34] 。它的语义线不太明显，在更混乱的场景中被更严重地遮挡
SL5K$\underline{\mathrm{E}}[33]:$就行数和场景类别而言，它是一个丰富多样的数据集。它由 4,000 个训练和 1,000 个测试图像组成
Culane$\boxed{\mathbb{E}}[24]:$它是用于道路车道检测的数据集，包含88,000个训练图像。它的 34,680 张测试图像分为 9 类。对于每个图像，提供最多 4 个道路车道的像素级遮罩。所提出的算法在“无车道”类别的 3,911 张测试图像上进行了测试，其中每条车道都是高度隐含的，甚至是不可见的
|
[
[
0.08251953125,
0.0947265625,
0.468017578125,
0.0947265625,
0.468017578125,
0.10479736328125,
0.08251953125,
0.10479736328125
],
[
0.50244140625,
0.0921630859375,
0.890625,
0.0921630859375,
0.890625,
0.1060791015625,
0.50244140625,
0.1060791015625
],
[
0.08251953125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.08251953125,
0.121826171875
],
[
0.50244140625,
0.10797119140625,
0.6357421875,
0.10797119140625,
0.6357421875,
0.121826171875,
0.50244140625,
0.121826171875
],
[
0.08087158203125,
0.12371826171875,
0.36767578125,
0.12371826171875,
0.36767578125,
0.136962890625,
0.08087158203125,
0.136962890625
],
[
0.52392578125,
0.125,
0.890625,
0.125,
0.890625,
0.1351318359375,
0.52392578125,
0.1351318359375
],
[
0.5048828125,
0.1395263671875,
0.86376953125,
0.1395263671875,
0.86376953125,
0.15283203125,
0.5048828125,
0.15283203125
],
[
0.080078125,
0.1483154296875,
0.44677734375,
0.1483154296875,
0.44677734375,
0.1622314453125,
0.080078125,
0.1622314453125
],
[
0.55810546875,
0.1622314453125,
0.8935546875,
0.1622314453125,
0.8935546875,
0.18310546875,
0.55810546875,
0.18310546875
],
[
0.08087158203125,
0.169189453125,
0.468017578125,
0.1685791015625,
0.468017578125,
0.18310546875,
0.08087158203125,
0.184326171875
],
[
0.08251953125,
0.1875,
0.46728515625,
0.1875,
0.46728515625,
0.1976318359375,
0.08251953125,
0.1976318359375
],
[
0.51953125,
0.1912841796875,
0.8896484375,
0.1912841796875,
0.8896484375,
0.2052001953125,
0.51953125,
0.2052001953125
],
[
0.08416748046875,
0.20263671875,
0.3095703125,
0.20263671875,
0.3095703125,
0.21337890625,
0.08416748046875,
0.21337890625
],
[
0.50244140625,
0.203857421875,
0.89208984375,
0.2052001953125,
0.89208984375,
0.2197265625,
0.50244140625,
0.2183837890625
],
[
0.5087890625,
0.1988525390625,
0.521484375,
0.1988525390625,
0.521484375,
0.2020263671875,
0.5087890625,
0.2020263671875
],
[
0.1004638671875,
0.216552734375,
0.46728515625,
0.216552734375,
0.46728515625,
0.2298583984375,
0.1004638671875,
0.2298583984375
],
[
0.50341796875,
0.2197265625,
0.890625,
0.2197265625,
0.890625,
0.2342529296875,
0.50341796875,
0.2342529296875
],
[
0.08087158203125,
0.2310791015625,
0.4248046875,
0.2310791015625,
0.4248046875,
0.2442626953125,
0.08087158203125,
0.2442626953125
],
[
0.5048828125,
0.236083984375,
0.890625,
0.236083984375,
0.890625,
0.25,
0.5048828125,
0.25
],
[
0.50341796875,
0.251953125,
0.89208984375,
0.251953125,
0.89208984375,
0.26513671875,
0.50341796875,
0.26513671875
],
[
0.188720703125,
0.25830078125,
0.35693359375,
0.25830078125,
0.35693359375,
0.279052734375,
0.188720703125,
0.279052734375
],
[
0.44677734375,
0.260009765625,
0.468017578125,
0.260009765625,
0.468017578125,
0.272705078125,
0.44677734375,
0.272705078125
],
[
0.50341796875,
0.26513671875,
0.8896484375,
0.267578125,
0.8896484375,
0.281005859375,
0.50341796875,
0.279052734375
],
[
0.50341796875,
0.2822265625,
0.8896484375,
0.2822265625,
0.8896484375,
0.29541015625,
0.50341796875,
0.29541015625
],
[
0.08087158203125,
0.2890625,
0.468017578125,
0.2890625,
0.468017578125,
0.302978515625,
0.08087158203125,
0.302978515625
],
[
0.08087158203125,
0.3037109375,
0.468017578125,
0.3037109375,
0.468017578125,
0.317626953125,
0.08087158203125,
0.317626953125
],
[
0.5048828125,
0.298583984375,
0.63623046875,
0.298583984375,
0.63623046875,
0.309326171875,
0.5048828125,
0.309326171875
],
[
0.08251953125,
0.32080078125,
0.295654296875,
0.32080078125,
0.295654296875,
0.330810546875,
0.08251953125,
0.330810546875
],
[
0.58642578125,
0.3212890625,
0.8037109375,
0.322509765625,
0.8037109375,
0.337158203125,
0.58642578125,
0.3359375
],
[
0.86767578125,
0.323974609375,
0.890625,
0.323974609375,
0.890625,
0.33837890625,
0.86767578125,
0.33837890625
],
[
0.178955078125,
0.3447265625,
0.36767578125,
0.34521484375,
0.36767578125,
0.363037109375,
0.178955078125,
0.3623046875
],
[
0.4453125,
0.347900390625,
0.4697265625,
0.347900390625,
0.4697265625,
0.361083984375,
0.4453125,
0.361083984375
],
[
0.50244140625,
0.347900390625,
0.5537109375,
0.347900390625,
0.5537109375,
0.3623046875,
0.50244140625,
0.3623046875
],
[
0.619140625,
0.3623046875,
0.77392578125,
0.363037109375,
0.77392578125,
0.37744140625,
0.619140625,
0.376953125
],
[
0.869140625,
0.363037109375,
0.890625,
0.363037109375,
0.890625,
0.37744140625,
0.869140625,
0.37744140625
],
[
0.0849609375,
0.375,
0.4697265625,
0.375,
0.4697265625,
0.38818359375,
0.0849609375,
0.38818359375
],
[
0.08087158203125,
0.389404296875,
0.36767578125,
0.389404296875,
0.36767578125,
0.40283203125,
0.08087158203125,
0.40283203125
],
[
0.5009765625,
0.385009765625,
0.8837890625,
0.385009765625,
0.8837890625,
0.3984375,
0.5009765625,
0.3984375
],
[
0.080078125,
0.410400390625,
0.46728515625,
0.410400390625,
0.46728515625,
0.423583984375,
0.080078125,
0.423583984375
],
[
0.63623046875,
0.410888671875,
0.7548828125,
0.410888671875,
0.7548828125,
0.425537109375,
0.63623046875,
0.425537109375
],
[
0.869140625,
0.410400390625,
0.890625,
0.410400390625,
0.890625,
0.4248046875,
0.869140625,
0.4248046875
],
[
0.08087158203125,
0.425537109375,
0.46728515625,
0.425537109375,
0.46728515625,
0.439453125,
0.08087158203125,
0.439453125
],
[
0.08087158203125,
0.43994140625,
0.465576171875,
0.43994140625,
0.465576171875,
0.453857421875,
0.08087158203125,
0.453857421875
],
[
0.5048828125,
0.443115234375,
0.88818359375,
0.443115234375,
0.88818359375,
0.453857421875,
0.5048828125,
0.453857421875
],
[
0.08087158203125,
0.455078125,
0.468017578125,
0.455078125,
0.468017578125,
0.468505859375,
0.08087158203125,
0.468505859375
],
[
0.50244140625,
0.45703125,
0.810546875,
0.45703125,
0.810546875,
0.47021484375,
0.50244140625,
0.47021484375
],
[
0.08251953125,
0.47021484375,
0.46728515625,
0.47021484375,
0.46728515625,
0.484130859375,
0.08251953125,
0.484130859375
],
[
0.52294921875,
0.472900390625,
0.81640625,
0.472900390625,
0.81640625,
0.486083984375,
0.52294921875,
0.486083984375
],
[
0.08251953125,
0.48486328125,
0.468017578125,
0.48486328125,
0.468017578125,
0.498779296875,
0.08251953125,
0.498779296875
],
[
0.66015625,
0.49755859375,
0.73193359375,
0.49755859375,
0.73193359375,
0.51220703125,
0.66015625,
0.51220703125
],
[
0.859375,
0.496826171875,
0.89208984375,
0.496826171875,
0.89208984375,
0.51220703125,
0.859375,
0.51220703125
],
[
0.08087158203125,
0.51513671875,
0.468017578125,
0.51513671875,
0.468017578125,
0.52880859375,
0.08087158203125,
0.52880859375
],
[
0.0882568359375,
0.50048828125,
0.468017578125,
0.50048828125,
0.468017578125,
0.5146484375,
0.0882568359375,
0.5146484375
],
[
0.50341796875,
0.5234375,
0.890625,
0.5234375,
0.890625,
0.537109375,
0.50341796875,
0.537109375
],
[
0.08087158203125,
0.53076171875,
0.468017578125,
0.53076171875,
0.468017578125,
0.54443359375,
0.08087158203125,
0.54443359375
],
[
0.50341796875,
0.537109375,
0.89208984375,
0.53857421875,
0.89208984375,
0.55322265625,
0.50341796875,
0.5517578125
],
[
0.08087158203125,
0.54443359375,
0.10296630859375,
0.548828125,
0.098876953125,
0.56103515625,
0.0767822265625,
0.556640625
],
[
0.5048828125,
0.5537109375,
0.66845703125,
0.5537109375,
0.66845703125,
0.5673828125,
0.5048828125,
0.5673828125
],
[
0.101318359375,
0.5625,
0.465576171875,
0.5625,
0.465576171875,
0.57275390625,
0.101318359375,
0.57275390625
],
[
0.08251953125,
0.57763671875,
0.46728515625,
0.57763671875,
0.46728515625,
0.58837890625,
0.08251953125,
0.58837890625
],
[
0.50244140625,
0.58203125,
0.7138671875,
0.58203125,
0.7138671875,
0.5966796875,
0.50244140625,
0.5966796875
],
[
0.08087158203125,
0.59130859375,
0.468017578125,
0.59130859375,
0.468017578125,
0.60498046875,
0.08087158203125,
0.60498046875
],
[
0.08251953125,
0.60791015625,
0.46728515625,
0.60791015625,
0.46728515625,
0.61865234375,
0.08251953125,
0.61865234375
],
[
0.50244140625,
0.60498046875,
0.60693359375,
0.60595703125,
0.60693359375,
0.62060546875,
0.50244140625,
0.619140625
],
[
0.08087158203125,
0.6220703125,
0.468017578125,
0.6220703125,
0.468017578125,
0.63525390625,
0.08087158203125,
0.63525390625
],
[
0.50341796875,
0.62890625,
0.8896484375,
0.62890625,
0.8896484375,
0.642578125,
0.50341796875,
0.642578125
],
[
0.08087158203125,
0.63623046875,
0.468017578125,
0.63623046875,
0.468017578125,
0.6494140625,
0.08087158203125,
0.6494140625
],
[
0.50341796875,
0.64453125,
0.890625,
0.64453125,
0.890625,
0.65771484375,
0.50341796875,
0.65771484375
],
[
0.08087158203125,
0.6513671875,
0.468017578125,
0.6513671875,
0.468017578125,
0.66552734375,
0.08087158203125,
0.66552734375
],
[
0.080078125,
0.66357421875,
0.471435546875,
0.66357421875,
0.471435546875,
0.6845703125,
0.080078125,
0.6845703125
],
[
0.5048828125,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.5048828125,
0.673828125
],
[
0.5048828125,
0.6767578125,
0.5537109375,
0.6767578125,
0.5537109375,
0.68798828125,
0.5048828125,
0.68798828125
],
[
0.08251953125,
0.68505859375,
0.46728515625,
0.68505859375,
0.46728515625,
0.69580078125,
0.08251953125,
0.69580078125
],
[
0.08251953125,
0.69873046875,
0.468017578125,
0.69873046875,
0.468017578125,
0.7119140625,
0.08251953125,
0.7119140625
],
[
0.50244140625,
0.6943359375,
0.890625,
0.69580078125,
0.890625,
0.708984375,
0.50244140625,
0.70849609375
],
[
0.08087158203125,
0.71337890625,
0.468017578125,
0.71337890625,
0.468017578125,
0.72802734375,
0.08087158203125,
0.72802734375
],
[
0.5048828125,
0.71044921875,
0.890625,
0.71044921875,
0.890625,
0.7236328125,
0.5048828125,
0.7236328125
],
[
0.080078125,
0.72802734375,
0.2132568359375,
0.72802734375,
0.2132568359375,
0.7412109375,
0.080078125,
0.7412109375
],
[
0.5048828125,
0.72607421875,
0.890625,
0.72607421875,
0.890625,
0.7392578125,
0.5048828125,
0.7392578125
],
[
0.50244140625,
0.7392578125,
0.66650390625,
0.7392578125,
0.66650390625,
0.75244140625,
0.50244140625,
0.75244140625
],
[
0.080078125,
0.75244140625,
0.468017578125,
0.75244140625,
0.468017578125,
0.7666015625,
0.080078125,
0.7666015625
],
[
0.50244140625,
0.7587890625,
0.8896484375,
0.76025390625,
0.8896484375,
0.7744140625,
0.50244140625,
0.7734375
],
[
0.08087158203125,
0.77587890625,
0.46728515625,
0.77587890625,
0.46728515625,
0.7890625,
0.08087158203125,
0.7890625
],
[
0.50341796875,
0.7763671875,
0.890625,
0.7763671875,
0.890625,
0.79052734375,
0.50341796875,
0.79052734375
],
[
0.08087158203125,
0.79150390625,
0.468017578125,
0.79052734375,
0.468017578125,
0.8046875,
0.08087158203125,
0.80615234375
],
[
0.5048828125,
0.79248046875,
0.5390625,
0.79248046875,
0.5390625,
0.8037109375,
0.5048828125,
0.8037109375
],
[
0.080078125,
0.80615234375,
0.468017578125,
0.806640625,
0.468017578125,
0.82080078125,
0.080078125,
0.8193359375
],
[
0.50341796875,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.08087158203125,
0.82080078125,
0.46728515625,
0.82080078125,
0.46728515625,
0.83544921875,
0.08087158203125,
0.83544921875
],
[
0.50341796875,
0.8271484375,
0.89208984375,
0.8271484375,
0.89208984375,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.07843017578125,
0.833984375,
0.4697265625,
0.83544921875,
0.4697265625,
0.85302734375,
0.07843017578125,
0.8515625
],
[
0.50634765625,
0.84326171875,
0.890625,
0.84326171875,
0.890625,
0.85400390625,
0.50634765625,
0.85400390625
],
[
0.08087158203125,
0.85302734375,
0.160888671875,
0.85302734375,
0.160888671875,
0.8642578125,
0.08087158203125,
0.8642578125
],
[
0.5048828125,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.080078125,
0.87060546875,
0.468017578125,
0.87255859375,
0.468017578125,
0.88623046875,
0.080078125,
0.8837890625
],
[
0.50244140625,
0.87255859375,
0.8896484375,
0.87255859375,
0.8896484375,
0.88623046875,
0.50244140625,
0.88623046875
],
[
0.08087158203125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.5048828125,
0.88818359375,
0.5537109375,
0.88818359375,
0.5537109375,
0.8994140625,
0.5048828125,
0.8994140625
]
] |
[
"的5×5网格中删除了24行。我们执行此过程K次，",
"找到一组节点E81它们相互连接并具有最大的权重",
"以组成K选定行的节点集。图5(b)和(e)分别在图",
"和(i.e.和声分数)",
"像和Hough空间上显示了这样的选定行",
"今令0表示一个集团：由成员节点的索集表示。然",
"后，我们将集团的协调能量Eharmony（の）定义为",
"3.3. Edge weighting: harmony score estimation",
"Eharmony(0) =Lie Zjeo,j> w(ui, uj) (6)",
"区域间相关性：为了区分图3（b）中的正对与图E3",
"c）中的负对，我们设计了区域间相关性（IRC）模块",
"它是中所有边权重的总和。找到最大化这个能量",
"该模块分析由一对线分隔的区域",
"的集团是NP-hard「6］。然而，在这项工作中，K被",
"",
"令R，1≤i≤M表示由两行分隔的区域。可以有",
"设置为一个小数字。默认的K是8。大约有2K个可",
"三个或四个区域，i.e.EM=3或4。我们通过",
"能的集团，这些集团也是可以管理的。因此，采用穷举",
"搜索来找到最大权重集团。首先，我们在图G中生成",
"ri=ZpeR, X(p).",
"(4)",
"可能的集团?={0t)的集合，其中每个集团0t由两",
"个以上的节点组成。然后，我们选择最大化协调能量的",
"提取R的区域特征向量ri我们计算区域Ril的",
"softmax概率ai来缩放区域特征向量，然后将缩放后",
"最大权重集团9*：",
"的向量连接成大小为C×4的",
"0*=argmaxeteeEharmony(Ot)",
"(7)",
"R= [airi,a2r2,a3r3, a4r4]",
"(5)",
"受约束",
"mini,jeo w(vi, v,) > K",
"(8)",
"。如果M=3，我们用零填充最右边的向量。然后，R",
"被馈送到完全连接的层以产生IRC特征",
"，其中是阈值。如果没有满足约束的集团，则通过",
"H-Net：我们使用IRC模块开发H-Net。它需要一个图",
"i* = argmax hii.",
"(9)",
"像和一对线，由i和j索引，以产生范围从0到1的和",
"声分数hii。图4(b）显示了H-Net架构。VGG16[26]",
"选择最大的单节点集团*=*！。通过将与重复输入",
"的卷积层被用作特征提取器，其后是IRC模块和线池",
"相同的行应用于H-Net来获得自和得分hii",
"化层的三个并行分支。我们使用行池化层在（1）中分",
"在获得一组和谐的语义线之后，我们通过",
"别对行i和3执行池化。我们使用两种类型的回归层：",
"lu+lui",
"(10)",
"算每行的一元可靠性（Reg2）。最后，我们通过将IRC",
"一种用于产生两行的IRC分数（Reg1），另一种用于计",
"来细化每一行，其中△lu表示偏移向量，由S-Net的",
"得分与一元可靠性水平的平均值相乘来计算和声得分",
"回归层生成。图5（c）和（g）显示了图像和Hough空",
"hij",
"间上的一组和谐语义线",
"我们为H-Net配置训练数据如下。假设图像中的每",
"对基本事实语义线彼此协调。因此，我们将这样的对声",
"4.Experimental Results",
"明为正，而将其他对声明为负。换句话说，只有当和",
"两行都是语义时，行对（，）才是正的。然后，根据",
"4.1. Datasets",
"对（i,）是否为正，将和声得分hii注释为1或0。但",
"SELE[19]：它是第一个语义线数据集，包含1,750个",
"是，正对的这种严格定义会导致类不平衡：正对太少。",
"室外图像，这些图像被拆分为1,575个训练图像和175",
"因此，我们干扰每个正对的行位置，并注释相应的和声",
"分数 hi,与e-(d+d）成比例，其中 d;和 d,表示行i",
"个测试图像。每条语义线由图像边界上两个端点的坐",
"标注释",
"和i的于扰。此外，用于训练H-Net的损失函数被定",
"义为lH=（hi一hii）²，其中hi是地面实况和声得",
"SEL_Hard【16]：它是用于测试语义行检测器的更具挑",
"分，hii是其估计值。补充文档更详细地描述了H-Net",
"战性的数据集。它包含300个测试图像，选自ADE20K",
"的培训过程和架构",
"分割数据集[34]。它的语义线不太明显，在更混乱的",
"场景中被更严重地遮挡",
"3.4. Graph optimization: finding harmonious lines",
"SL5KE[331：就行数和场景类别而言，它是一个丰富",
"图构造：我们构造了一个完整的图G=（V,），其中节",
"多样的数据集。它由4,000个训练和1,000个测试图像",
"点集V=[u1,2,..，Uk}表示在E3.2节中使用S-Net",
"组成",
"选择的K行。每对线由边缘集&=(（ui，U）：i≠)",
"CulaneE[24]：它是用于道路车道检测的数据集，包含",
"中的一个边缘连接。在3.3节中，H-Net为每个边分",
"88,000个训练图像。它的34,680张测试图像分为9类。",
"配一个权重w(ui,）= hi。图 5(f)可视化了一个完",
"对于每个图像，提供最多4个道路车道的像素级遮罩。",
"整的加权图",
"所提出的算法在“无车道”类别的3,911张测试图像上",
"MWCS：如前所述，一组语义行是最佳的，如果该组中",
"进行了测试，其中每条车道都是高度隐含的，甚至是不",
"的任何两行彼此和谐。因此，找到这样的最优集等同于",
"可见的"
] |
{
"arxiv_id": "2104.06903",
"language": "zh",
"page_idx": 3
}
|
||
2104.06903-zh-4
|
2104.06903-zh-4.jpg
|
<FigureHere>
Figure 5: Illustration of the proposed algorithm: (a) input image, (b) selected lines through the selection-andremoval process, (c) semantic lines, (d) probabilities of line candidates in the Hough space, (e) node set corresponding to the selected lines, ( f) complete graph, ( g) maximal clique. In ( f) , an edge is depicted in green or red depending on whether its weight is above the threshold $\kappa$ in (8) or not. As the weight approaches zero, the transparency increases.
<FigureHere>
Figure 6: IIlustration of two existing metrics of mIoU [19] and EA-score [10] and the proposed HIoU metric.
之间的欧几里得距离和线的角距离。在这两个度量中
精度和召回率都是由
(11)
$$\text{Precision}=\frac{N_l}{N_l+N_e},\quad\text{Recall}=\frac{N_l}{N_l+N_m}$$
<FigureHere>
计算的，其中$N_l$是正确检测到的语义行的数量，$N_e$ 是假阳性的数量，$N_m$是假阴性的数量。然后，通过
$$\mathrm{F-measure}=\frac{2\times\mathrm{Precision}\times\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}.$$
(12)
Figure 7: Comparison of mIoU[19] ,EA-score[10] ,and the proposed HIoU metric: There are two detection results for the same ground-truth. In result I , the position of each detected line is different from the ground-truth, but the detected lines convey the composition of the image relatively well. In result II , two detected lines match the ground-truth exactly, but they are not harmonious with the remaining one. As a group, they are inferior to result I. Since mIoU and EA-score consider only the accuracy of each individual line, they do not tell the difference between these two results and provide only marginally different scores. In contrast, HIoU quantifies the superiority of result I correctly.
计算 F 度量在阈值$\overline{\mathrm{E}}\tau$ 的整个范围内测量精度，召回率和 F 度量曲线的曲线下面积(AUC)性能，分别用${\mathrm{AUC}}$ ${\mathrm{P} , \mathrm{~AUC}}$ ${\mathrm{R} }$和 AUC ${\mathrm{F} }[ {\mathrm{EMAX6}}]$
但是，这些度量仅测量每条检测线的位置精度。他们不考虑场景中多条检测到的线之间的和谐程度。因此，它们可能会产生误导性的分数，如图 7 所示
HIoU 度量：我们提出了基于和声的相交联合 (HIoU) 度量来评估检测到的线的整体和声。当检测到的线条对图像的划分与地面真相的划分一致时，它们往往会传达出关于图像组成的和谐印象。。假设检测到的线的集合和地面实况线的集合将图像分别划分为区域$S=\left\{s_{1},s_{2},\ldots,s_{N}\right\}$和$T=\left\{t_{1},t_{2},\ldots,t_{M}\right\}$。然后，我们将 HIoU 定义为
$$\mathrm{HIoU}=\frac{\sum_{i=1}^{N}\max_{k}\mathrm{IoU}(s_{i},t_{k})+\sum_{j=1}^{M}\max_{k}\mathrm{IoU}(t_{j},s_{k})}{N+M}.$$
$\therefore(13)$
## $4. 2. \textbf{ Metrics}$
常规指标：有两个现有指标可用于评估语义行检测结果：mIoU$\underline{\mathrm{E}}[19]$ 和 EA-score [10]。在 mIoU 度量中， 如果检测到的行与地面实况语义行的 mIoU 得分大于阈值$\underline {\mathrm{E} }\tau$ , 则将其视为正确，如图$\underline {\mathrm{F} }6$ (a)所示。在EA 评分中，如果检测到的线与地面实况的相似度大于阈值，则将其视为正确，如图$\color{red}{\mathrm{E6(b)}}$所示。相似性由两个因素$S_d$和$S_\theta$组成，这两个因素分别基于线中点
换句话说，对于每个$s_i$ ,我们找到匹配的$t_k$并测量它们的 IoU。同样，对于每个$t_j$ ,我们找到它的 IoU 与匹配的$s_k$ 。然后，这些双向匹配 IoU 的平均值成为 HIoU 分数。图 6 (c)说明了如何计算 HIoU 分数。Figure 7 显示，HIoU 通过考虑检测到的线之间的和谐，比现有指标更合理地评估检测到的线
|
[
[
0.65673828125,
0.09661865234375,
0.86767578125,
0.09661865234375,
0.86767578125,
0.14013671875,
0.65673828125,
0.14013671875
],
[
0.1282958984375,
0.15283203125,
0.1502685546875,
0.15283203125,
0.1502685546875,
0.1666259765625,
0.1282958984375,
0.1666259765625
],
[
0.24267578125,
0.154052734375,
0.265625,
0.154052734375,
0.265625,
0.1666259765625,
0.24267578125,
0.1666259765625
],
[
0.359375,
0.154052734375,
0.381591796875,
0.154052734375,
0.381591796875,
0.1673583984375,
0.359375,
0.1673583984375
],
[
0.476318359375,
0.154052734375,
0.496826171875,
0.154052734375,
0.496826171875,
0.1666259765625,
0.476318359375,
0.1666259765625
],
[
0.70751953125,
0.1546630859375,
0.72607421875,
0.1546630859375,
0.72607421875,
0.1673583984375,
0.70751953125,
0.1673583984375
],
[
0.82275390625,
0.1546630859375,
0.84326171875,
0.1546630859375,
0.84326171875,
0.1673583984375,
0.82275390625,
0.1673583984375
],
[
0.08087158203125,
0.1787109375,
0.890625,
0.1787109375,
0.890625,
0.1925048828125,
0.08087158203125,
0.1925048828125
],
[
0.5947265625,
0.1558837890625,
0.609375,
0.1558837890625,
0.609375,
0.1654052734375,
0.5947265625,
0.1654052734375
],
[
0.080078125,
0.1932373046875,
0.8896484375,
0.1932373046875,
0.8896484375,
0.2083740234375,
0.080078125,
0.2083740234375
],
[
0.08087158203125,
0.208984375,
0.890625,
0.208984375,
0.890625,
0.2235107421875,
0.08087158203125,
0.2235107421875
],
[
0.08087158203125,
0.2247314453125,
0.89208984375,
0.2247314453125,
0.89208984375,
0.238037109375,
0.08087158203125,
0.238037109375
],
[
0.080078125,
0.2392578125,
0.2459716796875,
0.240478515625,
0.2459716796875,
0.25390625,
0.080078125,
0.25244140625
],
[
0.576171875,
0.268310546875,
0.5888671875,
0.279052734375,
0.57861328125,
0.28857421875,
0.564453125,
0.27783203125
],
[
0.491943359375,
0.27978515625,
0.498291015625,
0.285400390625,
0.49267578125,
0.290283203125,
0.484375,
0.2841796875
],
[
0.5888671875,
0.285400390625,
0.59619140625,
0.290283203125,
0.5888671875,
0.29541015625,
0.58251953125,
0.290283203125
],
[
0.833984375,
0.281005859375,
0.84619140625,
0.281005859375,
0.84619140625,
0.290283203125,
0.833984375,
0.290283203125
],
[
0.55615234375,
0.298583984375,
0.564453125,
0.298583984375,
0.564453125,
0.302978515625,
0.55615234375,
0.302978515625
],
[
0.438720703125,
0.306884765625,
0.4453125,
0.306884765625,
0.4453125,
0.3125,
0.438720703125,
0.3125
],
[
0.83642578125,
0.306884765625,
0.8447265625,
0.306884765625,
0.8447265625,
0.3125,
0.83642578125,
0.3125
],
[
0.75244140625,
0.309326171875,
0.779296875,
0.309326171875,
0.779296875,
0.317626953125,
0.75244140625,
0.317626953125
],
[
0.33251953125,
0.30810546875,
0.339111328125,
0.304931640625,
0.343017578125,
0.31005859375,
0.338134765625,
0.313232421875
],
[
0.619140625,
0.311279296875,
0.6298828125,
0.311279296875,
0.6298828125,
0.314453125,
0.619140625,
0.314453125
],
[
0.0931396484375,
0.32568359375,
0.1748046875,
0.32568359375,
0.1748046875,
0.3359375,
0.0931396484375,
0.3359375
],
[
0.205078125,
0.326904296875,
0.26953125,
0.326904296875,
0.26953125,
0.333984375,
0.205078125,
0.333984375
],
[
0.381591796875,
0.326904296875,
0.483642578125,
0.326904296875,
0.483642578125,
0.33837890625,
0.381591796875,
0.33837890625
],
[
0.7138671875,
0.326904296875,
0.8388671875,
0.326904296875,
0.8388671875,
0.337158203125,
0.7138671875,
0.337158203125
],
[
0.8349609375,
0.323974609375,
0.88818359375,
0.326904296875,
0.88623046875,
0.33837890625,
0.833984375,
0.335205078125
],
[
0.54345703125,
0.330810546875,
0.65673828125,
0.330810546875,
0.65673828125,
0.341552734375,
0.54345703125,
0.341552734375
],
[
0.6845703125,
0.333251953125,
0.71826171875,
0.333251953125,
0.71826171875,
0.34228515625,
0.6845703125,
0.34228515625
],
[
0.56884765625,
0.3466796875,
0.6337890625,
0.3466796875,
0.6337890625,
0.356689453125,
0.56884765625,
0.356689453125
],
[
0.3505859375,
0.333984375,
0.381591796875,
0.333984375,
0.381591796875,
0.341552734375,
0.3505859375,
0.341552734375
],
[
0.375732421875,
0.347900390625,
0.45751953125,
0.347900390625,
0.45751953125,
0.35791015625,
0.375732421875,
0.35791015625
],
[
0.74609375,
0.347900390625,
0.82666015625,
0.347900390625,
0.82666015625,
0.35791015625,
0.74609375,
0.35791015625
],
[
0.098876953125,
0.3642578125,
0.87109375,
0.3642578125,
0.87109375,
0.37890625,
0.098876953125,
0.37890625
],
[
0.50634765625,
0.4072265625,
0.890625,
0.4072265625,
0.890625,
0.42041015625,
0.50634765625,
0.42041015625
],
[
0.50341796875,
0.42041015625,
0.65185546875,
0.421630859375,
0.65185546875,
0.436279296875,
0.50341796875,
0.43505859375
],
[
0.2442626953125,
0.443115234375,
0.289306640625,
0.443115234375,
0.289306640625,
0.45068359375,
0.2442626953125,
0.45068359375
],
[
0.66259765625,
0.44580078125,
0.68896484375,
0.449462890625,
0.685546875,
0.4609375,
0.65869140625,
0.45703125
],
[
0.705078125,
0.442626953125,
0.8349609375,
0.45263671875,
0.83251953125,
0.474853515625,
0.70166015625,
0.464111328125
],
[
0.7958984375,
0.44580078125,
0.81884765625,
0.448974609375,
0.814453125,
0.460205078125,
0.79150390625,
0.45703125
],
[
0.859375,
0.451904296875,
0.890625,
0.451904296875,
0.890625,
0.464111328125,
0.859375,
0.464111328125
],
[
0.111083984375,
0.47021484375,
0.178955078125,
0.47021484375,
0.178955078125,
0.477783203125,
0.111083984375,
0.477783203125
],
[
0.5595703125,
0.4462890625,
0.70751953125,
0.451904296875,
0.7060546875,
0.472900390625,
0.55810546875,
0.46728515625
],
[
0.2303466796875,
0.47021484375,
0.318603515625,
0.47021484375,
0.318603515625,
0.477783203125,
0.2303466796875,
0.477783203125
],
[
0.359375,
0.47021484375,
0.449462890625,
0.47021484375,
0.449462890625,
0.477783203125,
0.359375,
0.477783203125
],
[
0.261474609375,
0.48095703125,
0.28759765625,
0.48095703125,
0.28759765625,
0.49365234375,
0.261474609375,
0.49365234375
],
[
0.390625,
0.48095703125,
0.418212890625,
0.48095703125,
0.418212890625,
0.49365234375,
0.390625,
0.49365234375
],
[
0.5048828125,
0.481689453125,
0.8896484375,
0.481689453125,
0.8896484375,
0.496826171875,
0.5048828125,
0.496826171875
],
[
0.334228515625,
0.48291015625,
0.3447265625,
0.48291015625,
0.3447265625,
0.49365234375,
0.334228515625,
0.49365234375
],
[
0.1854248046875,
0.481689453125,
0.2197265625,
0.481689453125,
0.2197265625,
0.49365234375,
0.1854248046875,
0.49365234375
],
[
0.2607421875,
0.492431640625,
0.289306640625,
0.492431640625,
0.289306640625,
0.5048828125,
0.2607421875,
0.5048828125
],
[
0.390625,
0.492431640625,
0.418212890625,
0.492431640625,
0.418212890625,
0.5048828125,
0.390625,
0.5048828125
],
[
0.334228515625,
0.49365234375,
0.3447265625,
0.49365234375,
0.3447265625,
0.50390625,
0.334228515625,
0.50390625
],
[
0.1666259765625,
0.49365234375,
0.221435546875,
0.49365234375,
0.221435546875,
0.5048828125,
0.1666259765625,
0.5048828125
],
[
0.50341796875,
0.496826171875,
0.87841796875,
0.496826171875,
0.87841796875,
0.51123046875,
0.50341796875,
0.51123046875
],
[
0.1854248046875,
0.50390625,
0.2197265625,
0.50390625,
0.2197265625,
0.51513671875,
0.1854248046875,
0.51513671875
],
[
0.2607421875,
0.50390625,
0.28759765625,
0.50390625,
0.28759765625,
0.51513671875,
0.2607421875,
0.51513671875
],
[
0.33251953125,
0.5048828125,
0.3447265625,
0.5048828125,
0.3447265625,
0.51318359375,
0.33251953125,
0.51318359375
],
[
0.39208984375,
0.50390625,
0.418212890625,
0.50390625,
0.418212890625,
0.51513671875,
0.39208984375,
0.51513671875
],
[
0.080078125,
0.5234375,
0.468017578125,
0.5234375,
0.468017578125,
0.537109375,
0.080078125,
0.537109375
],
[
0.685546875,
0.5244140625,
0.80126953125,
0.5244140625,
0.80126953125,
0.53515625,
0.685546875,
0.53515625
],
[
0.85888671875,
0.52587890625,
0.89208984375,
0.52587890625,
0.89208984375,
0.54150390625,
0.85888671875,
0.54150390625
],
[
0.69287109375,
0.53515625,
0.79736328125,
0.53515625,
0.79736328125,
0.54248046875,
0.69287109375,
0.54248046875
],
[
0.58837890625,
0.52783203125,
0.6806640625,
0.52783203125,
0.6806640625,
0.5390625,
0.58837890625,
0.5390625
],
[
0.080078125,
0.53857421875,
0.46728515625,
0.537109375,
0.46728515625,
0.5517578125,
0.080078125,
0.55322265625
],
[
0.080078125,
0.5537109375,
0.4697265625,
0.5537109375,
0.4697265625,
0.5673828125,
0.080078125,
0.5673828125
],
[
0.50341796875,
0.556640625,
0.890625,
0.55810546875,
0.890625,
0.57275390625,
0.50341796875,
0.5712890625
],
[
0.08087158203125,
0.568359375,
0.4697265625,
0.5673828125,
0.4697265625,
0.58203125,
0.08087158203125,
0.58349609375
],
[
0.50341796875,
0.57373046875,
0.890625,
0.57373046875,
0.890625,
0.5869140625,
0.50341796875,
0.5869140625
],
[
0.080078125,
0.583984375,
0.468017578125,
0.583984375,
0.468017578125,
0.59765625,
0.080078125,
0.59765625
],
[
0.50341796875,
0.5869140625,
0.82421875,
0.5869140625,
0.82421875,
0.60107421875,
0.50341796875,
0.60107421875
],
[
0.07843017578125,
0.5986328125,
0.468017578125,
0.59765625,
0.468017578125,
0.6123046875,
0.07843017578125,
0.61279296875
],
[
0.52294921875,
0.60302734375,
0.89208984375,
0.60400390625,
0.89208984375,
0.61865234375,
0.52294921875,
0.6171875
],
[
0.08251953125,
0.61279296875,
0.468017578125,
0.6142578125,
0.468017578125,
0.62890625,
0.08251953125,
0.62744140625
],
[
0.5048828125,
0.619140625,
0.890625,
0.619140625,
0.890625,
0.63330078125,
0.5048828125,
0.63330078125
],
[
0.08087158203125,
0.6298828125,
0.46728515625,
0.6298828125,
0.46728515625,
0.64306640625,
0.08087158203125,
0.64306640625
],
[
0.50244140625,
0.63330078125,
0.8544921875,
0.63330078125,
0.8544921875,
0.64794921875,
0.50244140625,
0.64794921875
],
[
0.080078125,
0.64599609375,
0.468017578125,
0.64306640625,
0.468017578125,
0.6572265625,
0.080078125,
0.6591796875
],
[
0.080078125,
0.66015625,
0.46728515625,
0.66015625,
0.46728515625,
0.673828125,
0.080078125,
0.673828125
],
[
0.50341796875,
0.65478515625,
0.89208984375,
0.65478515625,
0.89208984375,
0.66943359375,
0.50341796875,
0.66943359375
],
[
0.080078125,
0.6748046875,
0.468017578125,
0.6748046875,
0.468017578125,
0.68798828125,
0.080078125,
0.68798828125
],
[
0.5048828125,
0.67041015625,
0.8896484375,
0.67041015625,
0.8896484375,
0.6845703125,
0.5048828125,
0.6845703125
],
[
0.080078125,
0.689453125,
0.468017578125,
0.689453125,
0.468017578125,
0.70263671875,
0.080078125,
0.70263671875
],
[
0.50634765625,
0.6875,
0.88818359375,
0.6875,
0.88818359375,
0.69775390625,
0.50634765625,
0.69775390625
],
[
0.07843017578125,
0.705078125,
0.471435546875,
0.705078125,
0.471435546875,
0.7197265625,
0.07843017578125,
0.7197265625
],
[
0.5048828125,
0.70068359375,
0.890625,
0.70068359375,
0.890625,
0.7138671875,
0.5048828125,
0.7138671875
],
[
0.07843017578125,
0.71826171875,
0.205078125,
0.7197265625,
0.205078125,
0.734375,
0.07843017578125,
0.73291015625
],
[
0.50341796875,
0.71630859375,
0.890625,
0.71630859375,
0.890625,
0.72998046875,
0.50341796875,
0.72998046875
],
[
0.50244140625,
0.72998046875,
0.89208984375,
0.73095703125,
0.89208984375,
0.74560546875,
0.50244140625,
0.744140625
],
[
0.50341796875,
0.7470703125,
0.63623046875,
0.7470703125,
0.63623046875,
0.76025390625,
0.50341796875,
0.76025390625
],
[
0.07843017578125,
0.767578125,
0.1805419921875,
0.76953125,
0.1805419921875,
0.7841796875,
0.07843017578125,
0.7822265625
],
[
0.576171875,
0.7734375,
0.84716796875,
0.7734375,
0.84716796875,
0.787109375,
0.576171875,
0.787109375
],
[
0.515625,
0.78076171875,
0.57177734375,
0.78076171875,
0.57177734375,
0.79248046875,
0.515625,
0.79248046875
],
[
0.859375,
0.78076171875,
0.8896484375,
0.78076171875,
0.8896484375,
0.79345703125,
0.859375,
0.79345703125
],
[
0.69140625,
0.787109375,
0.73046875,
0.787109375,
0.73046875,
0.79736328125,
0.69140625,
0.79736328125
],
[
0.08251953125,
0.796875,
0.468017578125,
0.796875,
0.468017578125,
0.81005859375,
0.08251953125,
0.81005859375
],
[
0.08087158203125,
0.8095703125,
0.468017578125,
0.81005859375,
0.468017578125,
0.82470703125,
0.08087158203125,
0.82373046875
],
[
0.50341796875,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.08087158203125,
0.82470703125,
0.46728515625,
0.82568359375,
0.46728515625,
0.84033203125,
0.08087158203125,
0.8388671875
],
[
0.5048828125,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.5048828125,
0.84033203125
],
[
0.08251953125,
0.84130859375,
0.468017578125,
0.84130859375,
0.468017578125,
0.85498046875,
0.08251953125,
0.85498046875
],
[
0.50341796875,
0.8388671875,
0.89208984375,
0.8388671875,
0.89208984375,
0.85693359375,
0.50341796875,
0.85693359375
],
[
0.080078125,
0.85498046875,
0.468017578125,
0.85595703125,
0.468017578125,
0.869140625,
0.080078125,
0.86865234375
],
[
0.50341796875,
0.85498046875,
0.8896484375,
0.85595703125,
0.8896484375,
0.87060546875,
0.50341796875,
0.869140625
],
[
0.08251953125,
0.8720703125,
0.46728515625,
0.8720703125,
0.46728515625,
0.88525390625,
0.08251953125,
0.88525390625
],
[
0.50341796875,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.08087158203125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.08087158203125,
0.90087890625
],
[
0.50341796875,
0.88623046875,
0.79833984375,
0.88623046875,
0.79833984375,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"AA",
"(a)",
"(b)",
"(c)",
"(d)",
"(f)",
"(g)",
"Figure 5: Illustration of the proposed algorithm: (a) input image, (b) selected lines through the selection-and-",
"(e)",
"removal process, (c) semantic lines, (d) probabilities of line candidates in the Hough space, (e) node set corre-",
"sponding to the selected lines, (f) complete graph, (g) maximal clique. In (f), an edge is depicted in green or",
"red depending on whether its weight is above the threshold k in (8) or not. As the weight approaches zero, the",
"transparency increases.",
"",
"0",
"C",
"ti",
"",
"R",
"t2",
"S4",
"P",
"",
"Detection result",
"Ground-truth",
"IoU(R,R') + loU(Q.Q')",
"loU(si.t)+lou(s2.t) +..",
"+lou(tg.ss)",
"EA = (So(1,12)-Sa(11,12)2",
"Hlou=",
"(b) EA-score",
"mlou=",
"(a) mloU metric",
"(c) HloU metric",
"Figure 6: Illustration of two existing metrics of mloU [19] and EA-score [10] and the proposed HIoU metric.",
"之间的欧几里得距离和线的角距离。在这两个度量中，",
"精度和召回率都是由",
"",
"Ni",
"Recall = N+m",
"Ni",
"(11)",
"Ground-truth",
"Precision = N+,",
"Detection resultI",
"DetectionresultIl",
"0.90",
"0.88",
"计算的，其中Ni是正确检测到的语义行的数量，N",
"~",
"mloU",
"0.79",
"0.81",
"~",
"EA-score",
"是假阳性的数量，Nm是假阴性的数量。然后，通过",
"HloU",
"0.68",
">",
"0.59",
"Figure 7: Comparison of mloU [19] , EA-score [10]",
"2xPrecisionxRecall",
"(12)",
"Precision+Recall",
"F-measure =",
", and the proposed HloU metric: There are two de-",
"tection results for the same ground-truth. In result I",
"计算F度量在阈值E的整个范围内测量精度，召回",
", the position of each detected line is different from",
"率和F度量曲线的曲线下面积（AUC）性能，分别用",
"the ground-truth, but the detected lines convey the",
"AUCP,AUCR和AUCFEMAX6",
"composition of the image relatively well. In result II",
"但是，这些度量仅测量每条检测线的位置精度。他",
", two detected lines match the ground-truth exactly,",
"们不考虑场景中多条检测到的线之间的和谐程度。因",
"but they are not harmonious with the remaining one.",
"此，它们可能会产生误导性的分数，如图7所示",
"As a group, they are inferior to result I. Since mloU",
"and EA-score consider only the accuracy of each in-",
"HIoU度量：我们提出了基于和声的相交联合（HIoU)",
"dividual line, they do not tell the difference between",
"度量来评估检测到的线的整体和声。当检测到的线条",
"these two results and provide only marginally different",
"对图像的划分与地面真相的划分一致时：它们往往会",
"scores. In contrast, HloU quantifies the superiority of",
"传达出关于图像组成的和谐印象。。假设检测到的线",
"result I correctly.",
"的集合和地面实况线的集合将图像分别划分为区域",
"S={s1,82,...,Sn}和T={ti,t2,...,tM}。然后,我",
"们将HIoU定义为",
"4.2. Metrics",
"EN1 max IoU($t,tk)+EM1 max IoU(t,st)",
"HIoU=",
"(13)",
"N+M",
"常规指标：有两个现有指标可用于评估语义行检测结",
"果：mloUE[19]和EA-score[10]。在mloU度量中，",
"换句话说，对于每个si，我们找到匹配的t并测量",
"如果检测到的行与地面实况语义行的mIoU得分大于",
"它们的IoU。同样，对于每个t，，我们找到它的IoU",
"阈值ET，则将其视为正确，如图E6（a）所示。在",
"与匹配的sk。然后，这些双向匹配IoU的平均值成",
"EA评分中，如果检测到的线与地面实况的相似度大于",
"为HIoU分数。图 6(c）说明了如何计算HIoU分数。",
"阈值，则将其视为正确，如图E6（b）所示。相似性由",
"Figure7显示，HloU通过考虑检测到的线之间的和",
"两个因素Sa和S组成，这两个因素分别基于线中点",
"谐，比现有指标更合理地评估检测到的线"
] |
{
"arxiv_id": "2104.06903",
"language": "zh",
"page_idx": 4
}
|
||
2301.08440-zh-0
|
2301.08440-zh-0.jpg
|
非脆弱超边缘的超核心分解：
概念、算法、观察和应用
Fanchen Bu; Geon Lee; and Kijung Shin$^{\#}$
Abstract
Hypergraphs are a powerful abstraction for modeling high-order relations, which are ubiquitous in many fields. A hypergraph consists of nodes and hyperedges (i.e, subsets of nodes); and there have been a number of attempts to extend the notion of $k$ -cores, which proved useful with numerous applications for pairwise graphs, to hypergraphs However, the previous extensions are based on an unrealistic assumption that hyperedges are fragile , i.e., a highorder relation becomes obsolete as soon as a single member leaves it. In this work, we propose a new substructure model, called $(k,t)$ -hypercore, based on the assumption that high-order relations remain as long as at least $t$ fraction of the members remains. Specifically, it is defined as the maximal subhypergraph where(1) every node is contained in at least $k$ hyperedges in it and (2) at least $t$ fraction of the nodes remain in every hyperedge. We first prove that, given $t$ (or $k$ ), finding the $(k,t)$ -hypercore for every possible $k$ (or $t$ )can be computed in time linear w.r.t the sum of the sizes of hyperedges. Then, we demonstrate that real-world hypergraphs from the same domain share similar $(k,t)$ -hypercore structures, which capture different perspectives depending on $t.$ Lastly, we show the successful applications of our model in identifying influential nodes, dense substructures, and vulnerability in hypergraphs
# $\textbf{1}$ $\textbf{Introduction}$
图是表示成对关系的强大模型，它们已用于推荐系统 (Silva et al., 2010; Debnath et al., 2008) , 信息检索 (Blanco and Lioma, 2012; Mihalcea and Radev, 2011) , 知识表示 (Chein and Mugnier, 2008) 等。然而，图仅限于成对关系，因此无法精确描述高阶(i。e。,按组)两个以上节点之间的关系
超图，其中每个超边缘由任意数量的节点组成，通过精确描述高阶关系 (Benson et al.,2018a; Yin et al., 2017)来打破限制，并包含图作为特殊情况。超图已经成功地在不同领域对现实生活过程进行建模，包括化学反应 (Konstantinova and Skorobogatov, 2001)、流行病传播 (Bodó et al.,2016)和区块链经济 (Qu et al., 2018)
对于给定的成对图，$k$ -core (Seidman, 1983) 是一个内聚子结构，被定义为最大子图，其中每个节点的度至少为$k$ (i.e., each node is incident to at least $k$ edges)。已经进行了广泛的研究，以显示其线性时间可计算性(Batagelj and Zaversnik, 2003)和对$k$核的成功应用，包括图形可视化 (Alvarez-Hamelin et al.,2006),社区检测 (Corominas-Murtra et al., 2014) , 异常检测 (Shin et al., 2018a) 和生物过程建模 (Luo et al., 2009)
已经尝试将$k$核的概念推广到超图 (Hua et al., 2023; Luo et al., 2021, 2022; Gabert et al., 2021a,b; Sun et al. 2020),并且被称为超核的广义符号通常假设超边缘是脆弱的。也就是说，过度(i。e。,组关系)一旦任何组成节点选择退出，就会变得过时。具体来说，在超核计算期间，一旦删除其中的任何节点，就会忽略整个超边缘。然而，这样的假设是不现实的，并可能导致大量的信息丢失。例如，即使有人离开了在线群聊，它也可能保持活动状态；以及食谱(i。e。,一组成分)即使某些成分不可用，仍可能产生美味的结果。作为另一个示例，图中所示的超图。$\operatorname{El}$(d)不能分解为具有不同$k$的$k$ -超核，尽管子超图的内聚性有所不同，因为假定了脆弱的超边
为了更好地揭示超图中的结构信息，我们提出了$(k,t)$ -超核的概念。除了节点度阈值$k$之外，我们还引人了超比例阈值$t$ ,该阈值确定有多少个组成节点足以维持超比例。具体来说，给定一个超图和阈值$k$和$t$ , $(k,t)$ -hypercore 被定义为最大子超图，其中 (1)其中的每个节点 is contained in at least $k$ hyperedges 和 (2) 组成节点(i.e., the nodes constituting the original hyperedge) 的至少$t$部分保留在每个剩余的超边缘中。$t$的值越大，超级边缘越脆弱。基于该概念，我们将节点的$t$ -hypercoreness 定义为该节点在$(k,t)$ -hypercore 中的最大$k$ such ,并将节点的$k$ -分数定义为该节点在 MAXMTHX\_4 中的最大？??。InFig。1,我们展示了一个例子，其中$(k,t)$ -超核心结构随$t$而变化。值得注意的是，已经考虑了超核心的一些其他变体。Limnios
$^*$School of Electrical Engineering, KAIST, Daejeon, South Korea, boqvezen97@kaist.ac.kr $^{\dagger }$Kim Jaechul Graduate School of AI, KAIST, Seoul, South Korea, geonlee0325@ kaist. ac. kr $^{\ddagger}$Kim Jaechul Graduate School of AI and School of Electrical Engineering, KAIST, Seoul, South Korea, kijungs@ kaist.ac.kr
1
|
[
[
0.330078125,
0.14013671875,
0.6748046875,
0.14013671875,
0.6748046875,
0.1578369140625,
0.330078125,
0.1578369140625
],
[
0.343017578125,
0.1654052734375,
0.654296875,
0.1654052734375,
0.654296875,
0.1900634765625,
0.343017578125,
0.1900634765625
],
[
0.3349609375,
0.208984375,
0.66845703125,
0.208984375,
0.66845703125,
0.2235107421875,
0.3349609375,
0.2235107421875
],
[
0.473876953125,
0.281005859375,
0.52783203125,
0.281005859375,
0.52783203125,
0.292236328125,
0.473876953125,
0.292236328125
],
[
0.178955078125,
0.2998046875,
0.8388671875,
0.2998046875,
0.8388671875,
0.313232421875,
0.178955078125,
0.313232421875
],
[
0.1568603515625,
0.3125,
0.8408203125,
0.3125,
0.8408203125,
0.32568359375,
0.1568603515625,
0.32568359375
],
[
0.156005859375,
0.326904296875,
0.84228515625,
0.32763671875,
0.84228515625,
0.341552734375,
0.156005859375,
0.34033203125
],
[
0.156005859375,
0.34033203125,
0.8408203125,
0.341552734375,
0.8408203125,
0.354736328125,
0.156005859375,
0.354248046875
],
[
0.156005859375,
0.354248046875,
0.84228515625,
0.35595703125,
0.84228515625,
0.369384765625,
0.156005859375,
0.367431640625
],
[
0.158447265625,
0.37060546875,
0.8408203125,
0.37060546875,
0.8408203125,
0.380615234375,
0.158447265625,
0.380615234375
],
[
0.16015625,
0.3837890625,
0.47216796875,
0.3837890625,
0.47216796875,
0.39453125,
0.16015625,
0.39453125
],
[
0.46728515625,
0.386474609375,
0.50634765625,
0.386474609375,
0.50634765625,
0.393310546875,
0.46728515625,
0.393310546875
],
[
0.5087890625,
0.385009765625,
0.8408203125,
0.385009765625,
0.8408203125,
0.395751953125,
0.5087890625,
0.395751953125
],
[
0.1568603515625,
0.397705078125,
0.84228515625,
0.397705078125,
0.84228515625,
0.410888671875,
0.1568603515625,
0.410888671875
],
[
0.158447265625,
0.412353515625,
0.8388671875,
0.412353515625,
0.8388671875,
0.42236328125,
0.158447265625,
0.42236328125
],
[
0.16015625,
0.4267578125,
0.8388671875,
0.4267578125,
0.8388671875,
0.4375,
0.16015625,
0.4375
],
[
0.158447265625,
0.43994140625,
0.84228515625,
0.43994140625,
0.84228515625,
0.45068359375,
0.158447265625,
0.45068359375
],
[
0.156005859375,
0.45263671875,
0.814453125,
0.45263671875,
0.814453125,
0.466552734375,
0.156005859375,
0.466552734375
],
[
0.11273193359375,
0.477783203125,
0.28515625,
0.479248046875,
0.28515625,
0.49365234375,
0.11273193359375,
0.492431640625
],
[
0.11767578125,
0.5048828125,
0.88134765625,
0.5048828125,
0.88134765625,
0.51806640625,
0.11767578125,
0.51806640625
],
[
0.11767578125,
0.51953125,
0.88134765625,
0.51953125,
0.88134765625,
0.53271484375,
0.11767578125,
0.53271484375
],
[
0.11767578125,
0.53515625,
0.5947265625,
0.53515625,
0.5947265625,
0.548828125,
0.11767578125,
0.548828125
],
[
0.1405029296875,
0.5498046875,
0.88134765625,
0.5498046875,
0.88134765625,
0.56298828125,
0.1405029296875,
0.56298828125
],
[
0.11767578125,
0.56640625,
0.88134765625,
0.56640625,
0.88134765625,
0.5771484375,
0.11767578125,
0.5771484375
],
[
0.1168212890625,
0.580078125,
0.88134765625,
0.580078125,
0.88134765625,
0.59326171875,
0.1168212890625,
0.59326171875
],
[
0.143798828125,
0.5966796875,
0.8798828125,
0.5966796875,
0.8798828125,
0.60693359375,
0.143798828125,
0.60693359375
],
[
0.11767578125,
0.60986328125,
0.88134765625,
0.60986328125,
0.88134765625,
0.62353515625,
0.11767578125,
0.62353515625
],
[
0.1168212890625,
0.62548828125,
0.88232421875,
0.62548828125,
0.88232421875,
0.6396484375,
0.1168212890625,
0.6396484375
],
[
0.11767578125,
0.642578125,
0.79736328125,
0.642578125,
0.79736328125,
0.65283203125,
0.11767578125,
0.65283203125
],
[
0.1422119140625,
0.65576171875,
0.88134765625,
0.65576171875,
0.88134765625,
0.66943359375,
0.1422119140625,
0.66943359375
],
[
0.1168212890625,
0.67041015625,
0.88134765625,
0.67041015625,
0.88134765625,
0.6845703125,
0.1168212890625,
0.6845703125
],
[
0.1192626953125,
0.6875,
0.8798828125,
0.6875,
0.8798828125,
0.69775390625,
0.1192626953125,
0.69775390625
],
[
0.1168212890625,
0.70068359375,
0.8798828125,
0.70068359375,
0.8798828125,
0.7138671875,
0.1168212890625,
0.7138671875
],
[
0.1192626953125,
0.71728515625,
0.37744140625,
0.71728515625,
0.37744140625,
0.72802734375,
0.1192626953125,
0.72802734375
],
[
0.390625,
0.71533203125,
0.8798828125,
0.71533203125,
0.8798828125,
0.72900390625,
0.390625,
0.72900390625
],
[
0.11767578125,
0.73095703125,
0.88232421875,
0.73095703125,
0.88232421875,
0.744140625,
0.11767578125,
0.744140625
],
[
0.11767578125,
0.74755859375,
0.2296142578125,
0.74755859375,
0.2296142578125,
0.75830078125,
0.11767578125,
0.75830078125
],
[
0.1422119140625,
0.76123046875,
0.8798828125,
0.76123046875,
0.8798828125,
0.7744140625,
0.1422119140625,
0.7744140625
],
[
0.11767578125,
0.7763671875,
0.88232421875,
0.7763671875,
0.88232421875,
0.79052734375,
0.11767578125,
0.79052734375
],
[
0.1217041015625,
0.79150390625,
0.88232421875,
0.79150390625,
0.88232421875,
0.8046875,
0.1217041015625,
0.8046875
],
[
0.1168212890625,
0.806640625,
0.88134765625,
0.806640625,
0.88134765625,
0.82080078125,
0.1168212890625,
0.82080078125
],
[
0.1168212890625,
0.8212890625,
0.88134765625,
0.8212890625,
0.88134765625,
0.83544921875,
0.1168212890625,
0.83544921875
],
[
0.1168212890625,
0.8369140625,
0.875,
0.8369140625,
0.875,
0.85107421875,
0.1168212890625,
0.85107421875
],
[
0.11767578125,
0.8515625,
0.88232421875,
0.8515625,
0.88232421875,
0.86572265625,
0.11767578125,
0.86572265625
],
[
0.134033203125,
0.875,
0.611328125,
0.875,
0.611328125,
0.88818359375,
0.134033203125,
0.88818359375
],
[
0.13232421875,
0.88623046875,
0.61767578125,
0.88818359375,
0.61767578125,
0.9013671875,
0.13232421875,
0.8994140625
],
[
0.134033203125,
0.8994140625,
0.78515625,
0.8994140625,
0.78515625,
0.91357421875,
0.134033203125,
0.91357421875
],
[
0.49267578125,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94873046875,
0.49267578125,
0.94873046875
]
] |
[
"非脆弱超边缘的超核心分解：",
"概念、算法、观察和应用",
"Fanchen Bu, Geon Lee, and Kijung Shin",
"Abstract",
"Hypergraphs are a powerful abstraction for modeling high-order relations, which are ubiquitous in many fields",
"A hypergraph consists of nodes and hyperedges (i.e., subsets of nodes); and there have been a number of attempts to",
"extend the notion of k -cores,which proved useful with numerous applications for pairwise graphs, to hypergraphs.",
"However, the previous extensions are based on an unrealistic assumption that hyperedges are fragile,ie., a high",
"order relation becomes obsolete as soon as a single member leaves it. In this work, we propose a new substructure",
"model,called (k,t)-hypercore,based ontheassumptionthathigh-orderrelationsremain aslong as atleasttfraction",
"ofthemembersremains.Specifically,itisdefinedas",
"sthem",
"aximalsubhypergraphwhere(1)everynodeiscontained",
"in at least k hyperedges in it and (2) at least t fraction of the nodes remain in every hyperedge. We first prove that,",
"givent(ork),findingthe(k,t)-hypercoreforeverypossiblek(ort)canbecomputedintimelinearw.r.tthesum",
"of thesizes of hyperedges.Then,wedemonstratethatreal-worldhypergraphsfromthesamedomainsharesimilal",
"(k,t)-hypercore structures, which capture different perspectives depending on t:Lastly,we show the successful",
"applications of our model in identifying influential nodes, dense substructures, and vulnerability in hypergraphs.",
"1 Introduction",
"图是表示成对关系的强大模型，它们已用于推荐系统（Silvaetal.,2010;Debnathetal.,2008），信息检索（Blanco",
"andLioma，2012;MihalceaandRadev，2011），知识表示（CheinandMugnier2008）等。然而，图仅限于成对关",
"系，因此无法精确描述高阶（i。e。，按组）两个以上节点之间的关系",
"超图，其中每个超边缘由任意数量的节点组成，通过精确描述高阶关系（Bensonetal.,2018a；Yinetal.",
"2017）来打破限制，并包含图作为特殊情况。超图已经成功地在不同领域对现实生活过程进行建模，包括化",
"学反应（KonstantinovaandSkorobogatov，2001）、流行病传播（Bodoetal.，2016）和区块链经济（Quetal.，2018)",
"对于给定的成对图，k-core（Seidman，1983）是一个内聚子结构：被定义为最大子图，其中每个节点的度",
"至少为k（i.e.,eachnodeis incidenttoatleastkedges）。已经进行了广泛的研究，以显示其线性时间可计算性",
"(BatageljandZaversnik，2003）和对k核的成功应用，包括图形可视化（Alvarez-Hamelinetal.，2006），社区检",
"测（Corominas-Murtraetal.，2014），异常检测（Shinetal.，2018a）和生物过程建模（Luoetal.，2009）",
"已经尝试将k核的概念推广到超图（Huaetal.,2023;Luoetal,2021,2022;Gabertetal.,2021a,bSunetal.",
"2020），并且被称为超核的广义符号通常假设超边缘是脆弱的。也就是说，过度（i。e。，组关系）一旦任何组",
"成节点选择退出，就会变得过时。具体来说，在超核计算期间，一日删除其中的任何节点，就会忽略整个超",
"边缘。然而，这样的假设是不现实的，并可能导致大量的信息丢失。例如，即使有人离开了在线群聊，它也",
"可能保持活动状态；以及食谱（i。e。",
"一组成分）即使某些成分不可用，仍可能产生美味的结果。作为另一个",
"示例，图中所示的超图。E1（d）不能分解为具有不同k的k-超核，尽管子超图的内聚性有所不同，因为假",
"定了脆弱的超边",
"为了更好地揭示超图中的结构信息，我们提出了（k,t）-超核的概念。除了节点度阈值k之外，我们还引",
"人了超比例阈值t，该阈值确定有多少个组成节点足以维持超比例。具体来说，给定一个超图和阈值k和t",
"（k，t）-hypercore被定义为最大子超图，其中（1）其中的每个节点iscontainedinatleastkhyperedges和（2）",
"组成节点（i.e.，thenodesconstitutingtheoriginalhyperedge）的至少t部分保留在每个剩余的超边缘中。t的值",
"越大，超级边缘越脆弱。基于该概念，我们将节点的t-hypercoreness定义为该节点在（k,t）-hypercore中的",
"最大ksuch，并将节点的k-分数定义为该节点在MAXMTHX_4中的最大???。InFig。1，我们展示了",
"个例子，其中（k，t）-超核心结构随t而变化。值得注意的是，已经考虑了超核心的一些其他变体。Limnios",
"*School of Electrical Engineering, KAIST, Daejeon, South Korea, boqvezen97@kaist.ac.kr",
"Kim Jaechul Graduate School of AI,KAIST, Seoul, SouthKorea,geonlee0325@kaist.ac.kr",
"+Kim Jaechul Graduate School of AI and School of Electrical Engineering, KAIST, Seoul, South Korea, kijungs@kaist.ac.kr",
"1"
] |
{
"arxiv_id": "2301.08440",
"language": "zh",
"page_idx": 0
}
|
||
2301.08440-zh-1
|
2301.08440-zh-1.jpg
|
<FigureHere>
(a) $t\leq \frac 25$
(c) ${\frac 47}< t\leq {\frac 57}$
(d$)t>\frac57$
$( \mathbf{b} )$ ${\frac 25}< t\leq {\frac 47}$
Figure 1: An example of $(k$, fraction threshold $t$ ) rev when fragile hyperedges are
et al.(2021) 已经考虑了$(k;\ell)$ - 保留《组成节点 (而不是$(k,t)$ -hyperc 概念，它侧重于邻域的数量(i。e。,在每个节点的至少一个超边缘中共存的节点),并且该概念进一步扩展为(邻居，度)-($k,d)$-超核。与现有概念相比，我们经验证明的那样。
我们首先证明所提出的概念是定义明确的，并且具有包含属性 w。r。t $k$和$t$ ,然后提出类似剥离的计算算法来计算所有提出的概念，并显示它们的正确性和时间复杂度。特别是，我们表明，每个节点的 given $k$ 的$t$ -hypercoreness 和 given $t$的$k$ -fraction 都可以在时间上与超边的大小之和成比例地计算
为了证明所提出概念的有用性，我们在改变$t$的同时，研究了六个不同域 (Sinha et al., 2015; Mastrandrea et al., 2015; Leskovec et al., 2007) 中的十四个真实世界超图的$(k,t)$ -超核结构。检查从不同的角度得出以下观察结果：(1)(k,t)-超核心大小的基于域的模式 1 一致地遵循 MATHX_3-超核心大小的不同模式。(2) $t$的重尾分布-超超图：在大多数研究的真实节点-XcotMX 超图中，MAX
, 我们还通过以下三种应用来利用所提出的概念的某些属性来检测漏洞：(1)提出了有效的 MAX-3 节点识别：我们在超图中推广了 SIR 模型，并使用该模型表明$t$ -hypercoreness 是节点影响的可靠指标；(2)密集子结构发现：我们表明$(k,t)$ -hyperc
简而言之，我们的贡献有三个方面：
· New concepts. We propose the $(k,t)$ -hypercore, a new substructure model for hypergraphs, together with
$t$ -hypercoreness and $k$ -fraction (Defs. 5 to 7). In addition to the node-degree threshold $k$, the proposed
concepts incorporate the hyperedge-fraction threshold $t$ to provide more comprehensive information.
$•$ Properties and algorithms. We show some theoretical properties of the proposed concepts, and computation
algorithms (Algs. 1 to 3 ) f
$({\mathrm{Thms.~1~to~3~}}).$
· Observations and applicati
tions (Sec. 5), including a
also show successful applica
detection, and vulnerability
Reproducibility. 代码和数据集可在线获得
# 2 $\textbf{Preliminaries}$
在本节中，我们提供本文使用 throughout 的数学
Hypergraphs. 超图$H=(V,E)$由节每个超边缘与$\mathbb{N}$中的一个不同的正整数相关联， $v$的度$d(v;H)$是包含$v$, i 的
Ihttps://github.com/bokveize
$^{2}$A multiset is a set al
2
|
[
[
0.384033203125,
0.09344482421875,
0.414306640625,
0.0921630859375,
0.414306640625,
0.0997314453125,
0.384033203125,
0.10101318359375
],
[
0.340576171875,
0.11932373046875,
0.34716796875,
0.125,
0.340576171875,
0.1300048828125,
0.334228515625,
0.125
],
[
0.148681640625,
0.2008056640625,
0.16015625,
0.2083740234375,
0.1519775390625,
0.21533203125,
0.1405029296875,
0.2083740234375
],
[
0.55615234375,
0.20263671875,
0.564453125,
0.2083740234375,
0.55810546875,
0.2139892578125,
0.5498046875,
0.208984375
],
[
0.7587890625,
0.2052001953125,
0.7685546875,
0.2052001953125,
0.7685546875,
0.212158203125,
0.7587890625,
0.212158203125
],
[
0.1641845703125,
0.2310791015625,
0.223876953125,
0.2310791015625,
0.223876953125,
0.24560546875,
0.1641845703125,
0.24560546875
],
[
0.35302734375,
0.2310791015625,
0.441162109375,
0.2310791015625,
0.441162109375,
0.24560546875,
0.35302734375,
0.24560546875
],
[
0.55810546875,
0.2310791015625,
0.64453125,
0.2310791015625,
0.64453125,
0.24560546875,
0.55810546875,
0.24560546875
],
[
0.77685546875,
0.2298583984375,
0.83642578125,
0.2298583984375,
0.83642578125,
0.24560546875,
0.77685546875,
0.24560546875
],
[
0.1168212890625,
0.25830078125,
0.8798828125,
0.25830078125,
0.8798828125,
0.272705078125,
0.1168212890625,
0.272705078125
],
[
0.1168212890625,
0.272705078125,
0.88134765625,
0.272705078125,
0.88134765625,
0.287353515625,
0.1168212890625,
0.287353515625
],
[
0.11517333984375,
0.287353515625,
0.875,
0.28857421875,
0.875,
0.302978515625,
0.11517333984375,
0.3017578125
],
[
0.1168212890625,
0.330078125,
0.88134765625,
0.330078125,
0.88134765625,
0.3447265625,
0.1168212890625,
0.3447265625
],
[
0.11767578125,
0.34521484375,
0.88134765625,
0.34521484375,
0.88134765625,
0.359130859375,
0.11767578125,
0.359130859375
],
[
0.11517333984375,
0.35986328125,
0.88134765625,
0.359130859375,
0.88134765625,
0.373779296875,
0.11517333984375,
0.375
],
[
0.1168212890625,
0.375732421875,
0.88134765625,
0.375732421875,
0.88134765625,
0.39013671875,
0.1168212890625,
0.39013671875
],
[
0.11767578125,
0.392578125,
0.23779296875,
0.392578125,
0.23779296875,
0.40283203125,
0.11767578125,
0.40283203125
],
[
0.1405029296875,
0.406005859375,
0.88134765625,
0.406005859375,
0.88134765625,
0.419189453125,
0.1405029296875,
0.419189453125
],
[
0.11767578125,
0.42041015625,
0.88134765625,
0.42041015625,
0.88134765625,
0.434326171875,
0.11767578125,
0.434326171875
],
[
0.11767578125,
0.436279296875,
0.7587890625,
0.436279296875,
0.7587890625,
0.449462890625,
0.11767578125,
0.449462890625
],
[
0.1422119140625,
0.45068359375,
0.88134765625,
0.45068359375,
0.88134765625,
0.46533203125,
0.1422119140625,
0.46533203125
],
[
0.11767578125,
0.466552734375,
0.8798828125,
0.466552734375,
0.8798828125,
0.479736328125,
0.11767578125,
0.479736328125
],
[
0.11767578125,
0.48095703125,
0.88134765625,
0.48095703125,
0.88134765625,
0.494384765625,
0.11767578125,
0.494384765625
],
[
0.11767578125,
0.496826171875,
0.60302734375,
0.496826171875,
0.60302734375,
0.51025390625,
0.11767578125,
0.51025390625
],
[
0.1544189453125,
0.51123046875,
0.88134765625,
0.51123046875,
0.88134765625,
0.5244140625,
0.1544189453125,
0.5244140625
],
[
0.11767578125,
0.5263671875,
0.88134765625,
0.5263671875,
0.88134765625,
0.54052734375,
0.11767578125,
0.54052734375
],
[
0.11767578125,
0.54052734375,
0.374267578125,
0.54248046875,
0.374267578125,
0.55615234375,
0.11767578125,
0.5537109375
],
[
0.1422119140625,
0.5595703125,
0.390625,
0.5595703125,
0.390625,
0.5693359375,
0.1422119140625,
0.5693359375
],
[
0.1422119140625,
0.58203125,
0.88134765625,
0.58203125,
0.88134765625,
0.5966796875,
0.1422119140625,
0.5966796875
],
[
0.156005859375,
0.5966796875,
0.88232421875,
0.5966796875,
0.88232421875,
0.611328125,
0.156005859375,
0.611328125
],
[
0.156005859375,
0.6123046875,
0.814453125,
0.611328125,
0.814453125,
0.62548828125,
0.156005859375,
0.626953125
],
[
0.143798828125,
0.63720703125,
0.88134765625,
0.63720703125,
0.88134765625,
0.65087890625,
0.143798828125,
0.65087890625
],
[
0.1568603515625,
0.65283203125,
0.8798828125,
0.65283203125,
0.8798828125,
0.66650390625,
0.1568603515625,
0.66650390625
],
[
0.1568603515625,
0.66748046875,
0.265625,
0.66748046875,
0.265625,
0.68115234375,
0.1568603515625,
0.68115234375
],
[
0.1422119140625,
0.6923828125,
0.87841796875,
0.6923828125,
0.87841796875,
0.70703125,
0.1422119140625,
0.70703125
],
[
0.1568603515625,
0.70849609375,
0.8798828125,
0.70849609375,
0.8798828125,
0.72265625,
0.1568603515625,
0.72265625
],
[
0.1568603515625,
0.7236328125,
0.8798828125,
0.7236328125,
0.8798828125,
0.7373046875,
0.1568603515625,
0.7373046875
],
[
0.1568603515625,
0.73779296875,
0.406005859375,
0.73779296875,
0.406005859375,
0.751953125,
0.1568603515625,
0.751953125
],
[
0.125,
0.76220703125,
0.58837890625,
0.76220703125,
0.58837890625,
0.7763671875,
0.125,
0.7763671875
],
[
0.11358642578125,
0.787109375,
0.293212890625,
0.7890625,
0.293212890625,
0.8037109375,
0.11358642578125,
0.8017578125
],
[
0.11517333984375,
0.814453125,
0.57666015625,
0.814453125,
0.57666015625,
0.82763671875,
0.11517333984375,
0.82763671875
],
[
0.12744140625,
0.8291015625,
0.88134765625,
0.82763671875,
0.88134765625,
0.84228515625,
0.12744140625,
0.84326171875
],
[
0.11767578125,
0.8447265625,
0.8798828125,
0.8447265625,
0.8798828125,
0.85791015625,
0.11767578125,
0.85791015625
],
[
0.11767578125,
0.85986328125,
0.88134765625,
0.85986328125,
0.88134765625,
0.875,
0.11767578125,
0.875
],
[
0.1380615234375,
0.8837890625,
0.53369140625,
0.8837890625,
0.53369140625,
0.89794921875,
0.1380615234375,
0.89794921875
],
[
0.134033203125,
0.896484375,
0.384033203125,
0.89794921875,
0.384033203125,
0.9091796875,
0.134033203125,
0.90771484375
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"t-hyi",
"2",
"e",
"2",
"-",
"(a) t ≤",
"(b)<t≤4",
"(c)4<t≤",
"(d)t>号",
"Figure l: An example of (k,t )-hypercores. Assuming more robust hyperedges (i.e., decreasing the hyperedge-",
"fraction threshold t ) reveals cohesive substructures that are overlooked when fragile hyperedges are assumed. Notably,",
"when fragile hyperedges are assumed (i.e., when t = 1 ), every node has the same t -hypercoreness, as shown in (d).",
"etal.（2021）已经考虑了（k；l）-hypercore的概念，其中（k；）-hypercore要求在每个剩余的Hyperedge中至少",
"保留l组成节点（而不是（k,t）-hypercore中所需的t部分）。Arafatetal.（2023）已经考虑了邻域-k-超核的",
"概念，它侧重于邻域的数量(i。e。，在每个节点的至少一个超边缘中共存的节点），并且该概念进一步扩展为",
"（邻居，度）（k，d）-超核。与现有概念相比，我们提出的概念提供了关于超图的独特信息，正如理论证明和",
"经验证明的那样。",
"我们首先证明所提出的概念是定义明确的，并且具有包含属性w。r。tk和t，然后提出类似剥离的计算",
"算法来计算所有提出的概念，并显示它们的正确性和时间复杂度。特别是，我们表明，每个节点的givenk",
"的t-hypercoreness和givent的k-fraction都可以在时间上与超边的大小之和成比例地计算",
"为了证明所提出概念的有用性，我们在改变t的同时，研究了六个不同域（Sinhaetal.，2015；Mastrandrea",
"etal.，2015;Leskovecetal.,2007）中的十四个真实世界超图的（k,t）-超核结构。检查从不同的角度得出以下观",
"察结果：（1）（k,t）-超核心大小的基于域的模式1一致地遵循MATHX_3-超核心大小的不同模式。（2）t的",
"重尾分布-超超图：在大多数研究的真实节点-XcotMX超图中，MAX",
"我们还通过以下三种应用来利用所提出的概念的某些属性来检测漏洞：（1）提出了有效的MAX-3节点",
"识别：我们在超图中推广了SIR模型，并使用该模型表明t-hypercoreness是节点影响的可靠指标；（2）密集",
"子结构发现：我们表明（k,t)-hyperc",
"简而言之，我们的贡献有三个方面：",
": New concepts . We propose the (k,t) -hypercore, a new substructure model for hypergraphs, together with",
"t -hypercoreness and k -fraction (Defs. 5 to 7 ). In addition to the node-degree threshold k , the proposed",
"concepts incorporate the hyperedge-fraction threshold t to provide more comprehensive information.",
":Properties and algorithms .We show some theoretical properties of the proposed concepts, and computation",
"algorithms (Algs.1 to 3 ) for the proposed concepts with analyses of the correctness and time complexity",
"(Thms. 1 to 3).",
": Observations and applications . We investigate 14 real-world hypergraphs, which leads to interesting observa",
"tions (Sec. 5), including a surprising similarity in the (k, t) -hypercores of hypergraphs in the same domain. We",
"also show successful applications (Sec. 6)of the proposed concepts to influence estimation, dense-substructure",
"detection, and vulnerability detection.",
"Reproducibility代码和数据集可在线获得E(Buetal,2023）。1",
"2Preliminaries",
"在本节中，我们提供本文使用throughout的数学背景和预备知识",
"Hypergraphs.超图H=（V,E）由节点集V和超边缘多集E组成。2给定一个超图H=（V,E），我们将",
"每个超边缘与N中的一个不同的正整数相关联，i。e。，E={ei：iEIE，其中IE称为E的索引集。节点",
"u的度d(u;H)是包含u，i的超限数。e。，d(u;H)=iEI：UEei)l。节点u的邻居的集合N(u;H)是",
"https://github.com/bokveizen/non-fragile-hypercore",
"2Amultisetis a set allowing duplicate elements",
"2"
] |
{
"arxiv_id": "2301.08440",
"language": "zh",
"page_idx": 1
}
|
||
2301.08440-zh-2
|
2301.08440-zh-2.jpg
|
<table>
<tbody>
<tr>
<th>Notation</th>
<th>Definition</th>
</tr>
<tr>
<td>$I_{E}$</td>
<td>the index set of $E$</td>
</tr>
<tr>
<td>$k,t$</td>
<td>$11d$</td>
</tr>
<tr>
<td>$C_{k,t}$ $(H)$</td>
<td>the $(k,t)$-hypercore of $H$</td>
</tr>
</tbody>
</table>
在至少一个超边缘中与$v$共存的节点数，i。e。,$N(v;H)=\{u\in V:u\neq v,\exists e\in E$ s.t. $u,v\in e\}$ 。Hyperedge $e\in E$的组成节点是$e$中的节点。用$|E|$表示的超边缘$e\in E$的大小是$E$的基数 (i。e。, $e$的组成节点数)。$H$ 的大小，用$|H|$ 表示，是$H$ 中的节点数，i。e。$,|H|=|V|$。$H$ 的总大小，用$TS(H)$ 表示，是$H$ 中每个超边缘的大小之和(i。e。,$TS(H)=\sum_{i\in I_E}|e_i|$)。本文中的所有超图都是有限的，无向的，未加权的；并且在其中，每个节点的度至少为1 , i。e。$\bar{d}(\bar{v};H)\geq1,\forall v\in V$ ,每个超边的基数至少为2，i。e。,$|e_i|\geq2,\forall i\in I_E$ 。如果在超图$H=(V,E)$中，每个超边缘的基数恰好为 2,则 i。e。, $|e_i|=2,\forall i\in I_E$ ,则$H$也称为 (成对) 图。
Definition 1 (Subhypergraph). A hypergraph $H^\prime=(V^{\prime},E^{\prime})$ is a subhypergraph of $H=(V,E)$ if each hyperedge in $H^{\prime}$ is a subset of the hyperedge with the same index in $H$ ,i.e., $e_i^{\prime}\subseteq e_i,\forall i\in I_{E^{\prime}}\subseteq I_E$. If $e_i^{\prime}=e_i,\forall i\in I_{E^{\prime}}$, we call $H^{\prime}$ a complete subhypergraph of $H.$
请注意，子超图应该是超图，因此子超图中的每个超边缘也应具有至少两个基数。
我们总结了Tbl 中的符号。1。In the notations, 当上下文清楚时，可以省略输人超图$H$
Hypercores. 在成对图形中，$k$ -核 (Seidman,1983) is widely used 的概念。给定成对图$G$和$k\in\mathbb{N},G$的
$k$ 核心是最大子图，其中每个节点的度至少为$k$。
$\textbf{Definition 2( k - core) . Given a pairwise graph }G= ( V, E)$ and $k\in\mathbb{N}$, the $k$ -core of $H$ ,denoted by $C_k(G)=$ $(V^{\prime},E^{\prime})$ ,is the maximal subgraph of $G$ where each node has degree at least $k$ (i.e, is incident to at least $k$ edges) within $C_k$ $. \:^3$
它自然被推广为超图$\boxed{\mathbb{E}}($Hua et al., 2023; Luo et al., 2021, 2022; Gabert et al., 2021a,b; Sun et al., 2020), 如
follows 。
Definition 3( $k$ -hypercore). Given a hypergraph $H=(V,E)$ and $k\in\mathbb{N}$, the $k\textbf{- hypercore of}H$ ,denoted by $C_k(H)=(V^{\prime},E^{\prime})$ ,is the maximal complete subhypergraph of $H$ where each node has degree at least $k$ (i.e.,is contained in at least $k$ hyperedges) within $C_k.$
已经考虑了超核的一些变体。参见 Sec。3 的一些相关讨论。
Clique expansion. 将超图转换为成对图的最常见方法之一是集团扩展，其中每个超边缘$e\in E$都转换为由$e$中的节点组成的集团。给定一个超图$H=(V,E)$, its 未加权团扩展为$G_\mathrm{uc}( H) = ( V, \mathcal{E} ) $, its 加权团扩展为$G_\mathrm{wc}( H) = ( V, \mathcal{E} , \omega )$ ,其中边集$\mathcal{E} = \{ ( u, v) \in \binom V2: \exists e\in E$ $s. t. \{ u, v\} \subseteq e\}$ ,权函数$\omega\left((u,v)\right)=\left|\{i\in I_E:\{u,v\}\subseteq e_i\}\right|$。Clique 扩展提供了一种使超图更易于分析的方法，但是有关高阶交互的信息丢失了，这很自然，因为对于一组节点$V,V$中有$O(|V|^2)$可能对，而有$O(2^|V|)$可能的子集。结构明显不同的两个超图可能具有相同的集团扩展。\_ Star expansion. 每个超图都可以表示为二部图，称为其星展开$\overline{\mathbb{E}}($Zien et al.,1999)。超图$H$的星展开是二部图，其节点集是$V$和$E$的并集，其边缘集由$H$中的关联关系组成。
$\textbf{Definition 4 ( Star expansion) . Given a hypergraph }H= \left ( V, E\right ) $,its star expansion (i.e., bipartite-graph representation)
is $G_{se}(H)=(V\cup\hat{E},E_{se}(H))$ , where $\hat{E} _{se}( H) = \{ ( v, e) : v\in V, e\in \hat{E} , v\in e\} \subseteq V\times E$ $. ^{4}$
正如$\boxed{\mathrm{E}}$Yang et al.(2022)所指出的那样，尽管星型扩展包含超图中的所有发生率信息，但剩余的异构结构
在节点之间没有显式边，并且不适用于为简单齐次图设计的许多经过深入研究的图算法。
$\frac{3\text{In this work, the maximal subgraph(subhypergraph)}}$ satisfying some conditions means that every other graph (hypergraph) satisf $^3$In this work, the maximal subgraph (subhypergraph) satisfing such
conditions is a subgraph (subhypergraph) of the maximal one.
$^{4}$Similar to clique expansion, we can also h
3
|
[
[
0.435546875,
0.09027099609375,
0.560546875,
0.09027099609375,
0.560546875,
0.10479736328125,
0.435546875,
0.10479736328125
],
[
0.281005859375,
0.12060546875,
0.3447265625,
0.12249755859375,
0.343017578125,
0.1383056640625,
0.279296875,
0.1363525390625
],
[
0.414306640625,
0.125,
0.4794921875,
0.125,
0.4794921875,
0.1363525390625,
0.414306640625,
0.1363525390625
],
[
0.281005859375,
0.1427001953125,
0.365234375,
0.14453125,
0.365234375,
0.1590576171875,
0.281005859375,
0.1572265625
],
[
0.414306640625,
0.14453125,
0.71728515625,
0.14453125,
0.71728515625,
0.1590576171875,
0.414306640625,
0.1590576171875
],
[
0.281982421875,
0.1590576171875,
0.338134765625,
0.1590576171875,
0.338134765625,
0.173583984375,
0.281982421875,
0.173583984375
],
[
0.41259765625,
0.160400390625,
0.54833984375,
0.160400390625,
0.54833984375,
0.173583984375,
0.41259765625,
0.173583984375
],
[
0.281005859375,
0.17236328125,
0.306396484375,
0.17236328125,
0.306396484375,
0.1900634765625,
0.281005859375,
0.1900634765625
],
[
0.414306640625,
0.1749267578125,
0.53369140625,
0.1749267578125,
0.53369140625,
0.1881103515625,
0.414306640625,
0.1881103515625
],
[
0.281005859375,
0.1881103515625,
0.30810546875,
0.1881103515625,
0.30810546875,
0.2052001953125,
0.281005859375,
0.2052001953125
],
[
0.41259765625,
0.189453125,
0.708984375,
0.189453125,
0.708984375,
0.203857421875,
0.41259765625,
0.203857421875
],
[
0.281982421875,
0.203857421875,
0.34228515625,
0.203857421875,
0.34228515625,
0.2197265625,
0.281982421875,
0.2197265625
],
[
0.414306640625,
0.2052001953125,
0.58642578125,
0.2052001953125,
0.58642578125,
0.2197265625,
0.414306640625,
0.2197265625
],
[
0.281005859375,
0.2197265625,
0.3896484375,
0.2183837890625,
0.3896484375,
0.23291015625,
0.281005859375,
0.2342529296875
],
[
0.414306640625,
0.2197265625,
0.71337890625,
0.2197265625,
0.71337890625,
0.23291015625,
0.414306640625,
0.23291015625
],
[
0.281982421875,
0.23486328125,
0.39453125,
0.23486328125,
0.39453125,
0.25,
0.281982421875,
0.25
],
[
0.41259765625,
0.23486328125,
0.6748046875,
0.23486328125,
0.6748046875,
0.248779296875,
0.41259765625,
0.248779296875
],
[
0.1168212890625,
0.2890625,
0.8798828125,
0.2890625,
0.8798828125,
0.302978515625,
0.1168212890625,
0.302978515625
],
[
0.1168212890625,
0.3037109375,
0.88134765625,
0.3037109375,
0.88134765625,
0.317626953125,
0.1168212890625,
0.317626953125
],
[
0.11767578125,
0.3193359375,
0.88134765625,
0.3193359375,
0.88134765625,
0.333251953125,
0.11767578125,
0.333251953125
],
[
0.1168212890625,
0.335205078125,
0.88134765625,
0.335205078125,
0.88134765625,
0.349853515625,
0.1168212890625,
0.349853515625
],
[
0.1168212890625,
0.348388671875,
0.88134765625,
0.349853515625,
0.88134765625,
0.3642578125,
0.1168212890625,
0.363037109375
],
[
0.125,
0.363037109375,
0.88232421875,
0.3642578125,
0.88232421875,
0.37890625,
0.125,
0.37744140625
],
[
0.1168212890625,
0.380126953125,
0.1422119140625,
0.380126953125,
0.1422119140625,
0.393310546875,
0.1168212890625,
0.393310546875
],
[
0.1168212890625,
0.401611328125,
0.88134765625,
0.401611328125,
0.88134765625,
0.415283203125,
0.1168212890625,
0.415283203125
],
[
0.11767578125,
0.416748046875,
0.88134765625,
0.416748046875,
0.88134765625,
0.43115234375,
0.11767578125,
0.43115234375
],
[
0.1168212890625,
0.431884765625,
0.38232421875,
0.431884765625,
0.38232421875,
0.4462890625,
0.1168212890625,
0.4462890625
],
[
0.143798828125,
0.455078125,
0.71728515625,
0.455078125,
0.71728515625,
0.46533203125,
0.143798828125,
0.46533203125
],
[
0.1405029296875,
0.46728515625,
0.7685546875,
0.46728515625,
0.7685546875,
0.481689453125,
0.1405029296875,
0.481689453125
],
[
0.1258544921875,
0.48291015625,
0.88232421875,
0.48291015625,
0.88232421875,
0.496826171875,
0.1258544921875,
0.496826171875
],
[
0.11517333984375,
0.498779296875,
0.463134765625,
0.498779296875,
0.463134765625,
0.51220703125,
0.11517333984375,
0.51220703125
],
[
0.11517333984375,
0.51953125,
0.8798828125,
0.51953125,
0.8798828125,
0.53271484375,
0.11517333984375,
0.53271484375
],
[
0.1168212890625,
0.53515625,
0.88134765625,
0.53515625,
0.88134765625,
0.548828125,
0.1168212890625,
0.548828125
],
[
0.1168212890625,
0.55029296875,
0.200927734375,
0.55029296875,
0.200927734375,
0.5625,
0.1168212890625,
0.5625
],
[
0.1422119140625,
0.5712890625,
0.88134765625,
0.5712890625,
0.88134765625,
0.58544921875,
0.1422119140625,
0.58544921875
],
[
0.11767578125,
0.5869140625,
0.1732177734375,
0.5869140625,
0.1732177734375,
0.5986328125,
0.11767578125,
0.5986328125
],
[
0.11517333984375,
0.60693359375,
0.88134765625,
0.60791015625,
0.88134765625,
0.62255859375,
0.11517333984375,
0.6220703125
],
[
0.11517333984375,
0.62353515625,
0.88134765625,
0.62353515625,
0.88134765625,
0.63818359375,
0.11517333984375,
0.63818359375
],
[
0.1168212890625,
0.63818359375,
0.420654296875,
0.63818359375,
0.420654296875,
0.65283203125,
0.1168212890625,
0.65283203125
],
[
0.1405029296875,
0.6591796875,
0.57421875,
0.6591796875,
0.57421875,
0.673828125,
0.1405029296875,
0.673828125
],
[
0.1298828125,
0.6748046875,
0.88134765625,
0.6748046875,
0.88134765625,
0.689453125,
0.1298828125,
0.689453125
],
[
0.1168212890625,
0.68994140625,
0.88134765625,
0.68994140625,
0.88134765625,
0.7041015625,
0.1168212890625,
0.7041015625
],
[
0.11767578125,
0.705078125,
0.88134765625,
0.705078125,
0.88134765625,
0.7197265625,
0.11767578125,
0.7197265625
],
[
0.11767578125,
0.72021484375,
0.88134765625,
0.72021484375,
0.88134765625,
0.734375,
0.11767578125,
0.734375
],
[
0.11767578125,
0.73486328125,
0.88134765625,
0.73486328125,
0.88134765625,
0.74853515625,
0.11767578125,
0.74853515625
],
[
0.1192626953125,
0.75,
0.88134765625,
0.75,
0.88134765625,
0.76318359375,
0.1192626953125,
0.76318359375
],
[
0.11767578125,
0.76513671875,
0.88134765625,
0.76513671875,
0.88134765625,
0.77880859375,
0.11767578125,
0.77880859375
],
[
0.11767578125,
0.7822265625,
0.188720703125,
0.7822265625,
0.188720703125,
0.79345703125,
0.11767578125,
0.79345703125
],
[
0.11517333984375,
0.80029296875,
0.88134765625,
0.8017578125,
0.88134765625,
0.81640625,
0.11517333984375,
0.81494140625
],
[
0.11517333984375,
0.8173828125,
0.7294921875,
0.8173828125,
0.7294921875,
0.83056640625,
0.11517333984375,
0.83056640625
],
[
0.1405029296875,
0.83837890625,
0.88134765625,
0.83837890625,
0.88134765625,
0.8515625,
0.1405029296875,
0.8515625
],
[
0.11767578125,
0.85400390625,
0.74267578125,
0.85400390625,
0.74267578125,
0.8671875,
0.11767578125,
0.8671875
],
[
0.131591796875,
0.87353515625,
0.88232421875,
0.875,
0.88232421875,
0.8896484375,
0.131591796875,
0.88818359375
],
[
0.1192626953125,
0.8896484375,
0.4248046875,
0.8896484375,
0.4248046875,
0.8994140625,
0.1192626953125,
0.8994140625
],
[
0.1356201171875,
0.8994140625,
0.72607421875,
0.8994140625,
0.72607421875,
0.91357421875,
0.1356201171875,
0.91357421875
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"Table 1: Notations",
"Notation",
"Definition",
"H= (V,E)",
"a hypergraph with nodes V and hyperedges E",
"d(v; H)",
"the degree of u in H",
"Ie",
"the index set of E",
"k,t",
"the degree and hyperedge-fraction thresholds",
"Ck,t(H)",
"the (k,t) -hypercore of H",
"ct(v; H),c*(H)",
"the t -hypercoreness of u in H , and that of H",
"fh(v; H), f(H)",
"the k -fraction of w in H , and that of H",
"在至少一个超边缘中与u共存的节点数，i。e。，N(u;H)=（uEV：u≠u,3eEEs.t.uuEe}。Hyperedge",
"eEE的组成节点是e中的节点。用E表示的超边缘eEE的大小是E的基数（i。e。e的组成节点数)。H",
"的大小，用H表示，是H中的节点数，i。e。，H=VI。H的总大小，用TS(H)表示，是H中每个超边",
"缘的大小之和（i。e。，TS(H）=ZieIleil)。本文中的所有超图都是有限的，无向的，未加权的；并且在其",
"中，每个节点的度至少为1，i。e。，d(u;H)≥1,VueV，每个超边的基数至少为2，i。e。，leil≥2,ViEIe",
"如果在超图H=（V,E）中，每个超边缘的基数恰好为2，则i。e。，le=2,ViEIE，则H也称为（成对）",
"图。",
"Definition 1 (Subhypergraph). A hypergraph H' = (V', E') is a subhypergraph of H = (V, E) if each hyperedge in",
"H' is a subset of the hyperedge with the same index in H , ie., e' C ei, Vi e Ie? C Ie . If e, = ei, Vi e Ie', we",
"call H' a complete subhypergraph of H",
"请注意，子超图应该是超图，因此子超图中的每个超边缘也应具有至少两个基数。",
"我们总结了Tbl中的符号。1。In thenotations，当上下文清楚时，可以省略输人超图H",
"Hypercores.在成对图形中，k-核(Seidman，1983）iswidelyused的概念。给定成对图G和kEN，G的",
"k核心是最大子图，其中每个节点的度至少为k。",
"Definition 2 ( k -core). Given a pairwise graph G = (V, E) and k E N , the k -core of H , denoted by Ck(G) =",
"(V', E') , is the maximal subgraph of G where each node has degree at least k (i.e., is incident to at least k edges)",
"within Ck 3",
"它自然被推广为超图E(Huaetal.,2023;Lu0etal.,2021,2022;Gabertetal.,2021a,b;Sun etal.,2020)，如",
"follows",
"Definition 3(k -hypercore). Given a hypergraph H = (V, E) and k E N, the k -hypercore of H , denoted by",
"Ck(H) = (V', E'), is the maximal complete subhypergraph of H where each node has degree at least k (i.e., is",
"contained in at least k hyperedges) within Ck .",
"已经考虑了超核的一些变体。参见Sec。3的一些相关讨论。",
"Cliqueexpansion。将超图转换为成对图的最常见方法之一是集团扩展，其中每个超边缘eEE都转",
"换为由e中的节点组成的集团。给定一个超图H=（V,E），its未加权团扩展为Guc(H）=（V,），its",
"加权团扩展为Gwe(H）=（V,&,w），其中边集&={(u,u)E（）：eEEs.t.{u,)e)，权函数",
"w（（u,u)）=liEIE：【u,u)Cei}l。Clique扩展提供了一种使超图更易于分析的方法，但是有关高阶交互的",
"信息丢失了，这很自然，因为对于一组节点V，V中有O(IV2)可能对，而有O(2IVI)可能的子集。结构明",
"显不同的两个超图可能具有相同的集团扩展。Starexpansion.每个超图都可以表示为二部图，称为其星",
"展开E(Zienetal.，1999）。超图H的星展开是二部图，其节点集是V和E的并集，其边缘集由H中的关联",
"关系组成。",
"Definition 4 (Star expansion). Given a hypergraph H = (V, E), its star expansion (i.e., bipartite-graph representation)",
"isGse(H) = (VUE, Ese(H), where Ese(H) = {(u,e) : uEV,e E E, u Ee) C V x E.",
"正如EYangetal.(2022）所指出的那样，尽管星型扩展包含超图中的所有发生率信息，但剩余的异构结构",
"在节点之间没有显式边，并且不适用于为简单齐次图设计的许多经过深人研究的图算法。",
"3In this work, the maximal subgraph (subhypergraph) satisfying some conditions means that every other graph (hypergraph) satisfying such",
"conditionsisasubgraph(subhypergraph)ofthemaximalo",
"ASimilar to clique expansion, we can also have weighted star expansion, which is, however, not used in this work.",
"3"
] |
{
"arxiv_id": "2301.08440",
"language": "zh",
"page_idx": 2
}
|
||
2301.08440-zh-3
|
2301.08440-zh-3.jpg
|
# $\textbf{3}$ $\textbf{Concepts}$
在本节中，我们介绍所提出的概念并显示它们的一些理论性质。此外，我们还讨论了所提出的概念与一些现
有相关概念之间的联系和区别。
3.1 Proposed concepts
在成对图形中，每个边表示两个节点之间的连接，因此删除任何一个节点自然会导致边的完全无效。相反即使删除了某些组成节点，具有三个或更多节点的 Hyperedge 仍然表示其余节点之间的交互。正如我们在图中讨论和显示的那样。1 , Def 中的直接概括。3 毫无根据地假设了脆弱的超级边缘并遭受信息丢失。我们试图通过考虑非脆弱的超图来更好地揭示超图的结构
因此，我们引人了超边缘分数阈值$t$ ,该阈值确定了维持超边缘所需的组成节点的最小比例，从而导致
Def。5。
$\textbf{Definition 5( }\left ( k, t\right )$-hypercore). Given $H=\left(V,E\right),k\in\mathbb{N}$,and $t\in[0,1]$ ,the $\left ( k, t\right ) \textbf{- hypercore of }H$ ,denoted by $C_{k,t}(H)=(V^{\prime},E^{\prime})$ ,is the maximal (in terms of total size) subhypergraph of $H$ where (1) every node in has degree at least $k$ (i.e, is contained in at least $k$ hyperedges) within $C_k,t$ and (2) at least $t$ proportion of the constituent nodes remain in every hyperedge of $C_{k,t}(H)$ .Formally, $d(v;C_{k,t}(H))\geq k,\forall v\in V^{\prime}$ and $| e_i^{\prime }\cap e_i| \geq t| e_i| , \forall i\in I_{E^{\prime }}\subseteq I_E$ .
请注意，由于子超图的定义，$(k,t)$ -hypercore 的定义要求每个超边缘中至少保留两个节点(请参见 Sec。
2 )。另请参见 Alg 中的 Line 6。1。
$\textbf{Definition 6}( t$ -hypercoreness). Given $H=(V,E)$ and $t\in[0,1]$ ,the $t\textbf{ - hypercoreness of }v\in V$ ,denoted by $c_{t}( v; H)$ ,is the maximum positive integer such that $v$ is in the $(c_t(v),t)$ -hypercore, i.e., $c_t(v)=\max\{k\in\mathbb{N}:v\in$ $V(C_{k,t})\}.$ We call $c_t^*(H):=\max\{c_t(v):v\in V\}$ the $t$ -hypercoreness of $H.$
$\textbf{Definition 7}( k$ -fraction). Given $H=(V,E)$ and $k\in\mathbb{N}$ ,the $k\textbf{- fraction of }v\in V$ ,denoted by $f_k(v;H)$ ,is the maximum real number in [0,1] such that $v$ is in the $(k,f_k(v))$ -hypercore, i.e., $f_k(v)=\max\{t\in[0,1]:v\in V(C_{k,t})\}$ .For the completeness of definition, if $\{t\in[0,1]:v\in V(C_{k,t})\}=\varnothing$,we let $f_k(v;H)=-1.$ We call $f_k^*(H):=\max\{f_k(v):v\in V\}$ the $k$-fraction of $H.$
请注意，建议的概念可扩展到加权超图。具体来说，只要我们对加权超图上的节点度和超边缘分数有严
格的定义，扩展就很简单。
Example. 在图中。1 ,当$t$改变时，每个节点的$t$ -hypercoreness 也会改变。具体来说，当$t\leq\frac{4}{7}$时，四个节点具有$t$ -hypercoreness 3。它们在$\frac47<t\leq\frac57$时具有$t$ -hypercoreness 2 , 在$t>\frac{5}{7}$时具有$t$ -hypercoreness 1 ,这意味着它们的 3 -分数为$\frac{4}{7},2$
以下命题表明$(k,t)$ -hypercore 定义良好，并且具有双向包含属性。
Proposition 1 (Existence and uniqueness). Given any hypergraph $H$ ,$k\in\mathbb{N}$,and $t\in[0,1]$ ,$C_k,t$ uniquely exists
and is possibly empty .
$\square$
$Proof.$ See Appendix A.1.
$\textbf{Proposition 2 ( Two- way containment) . Let }H$ be any hypergraph. Fix any $k\in\mathbb{N}$, for any $0\leq t_1<t_2\leq1$, $C_{k,t_2}(H)$ is a subhypergraph of $C_k,t_1(H).$ Similarly, fix any $t\in[0,1]$ ,for any $k_1<k_2\in\mathbb{N},C_{k_2,t}(H)$ is a subhypergraph of $C_{k_1, t}( H)$ .
$Proof.$ See Appendix A.2 .
$$\text{口}$$
# 3.2 Related concepts
下面，我们讨论一些现有的相关概念，尤其是它们之间的联系和区别以及我们提出的概念
$\underline{\textbf{Existing variants of hypercores.}}$如 Sec 中所述。2(参见 Def。3), 大多数以前的作品 (Hua et al., 2023; Luo et al., 2021, 2022; Gabert et al., 2021a,b; Sun et al., 2020) 都是基于假设脆弱的超边缘 (i。e。, 当任何节点离开它时，一个超边缘就会被移除),这相当于$(k,t)$ -hypercore 与$t=1$ (i。e。, $(k,t)$ -超核心的特例)。Limnios et al. (2021) 将给定超图$H$的$(k;\ell)$ -超核定义为$H$的最大子超图，其中每个节点在子超图中至少具有$k$(i.e.,is contained in at least $k$ hyperedges)度，并且每个超图至少包含 $\ell$节点。根据这个概念，我们可以定义 l -hypercoreness
4
|
[
[
0.111083984375,
0.091552734375,
0.25244140625,
0.0947265625,
0.25244140625,
0.11236572265625,
0.111083984375,
0.10919189453125
],
[
0.1168212890625,
0.11932373046875,
0.88134765625,
0.11932373046875,
0.88134765625,
0.1331787109375,
0.1168212890625,
0.1331787109375
],
[
0.1168212890625,
0.1351318359375,
0.3447265625,
0.1351318359375,
0.3447265625,
0.1483154296875,
0.1168212890625,
0.1483154296875
],
[
0.11358642578125,
0.1546630859375,
0.3203125,
0.1572265625,
0.3203125,
0.1717529296875,
0.11358642578125,
0.169189453125
],
[
0.11517333984375,
0.177978515625,
0.8798828125,
0.1787109375,
0.8798828125,
0.1932373046875,
0.11517333984375,
0.1925048828125
],
[
0.11767578125,
0.1944580078125,
0.88134765625,
0.1944580078125,
0.88134765625,
0.2083740234375,
0.11767578125,
0.2083740234375
],
[
0.1192626953125,
0.210205078125,
0.88134765625,
0.210205078125,
0.88134765625,
0.2235107421875,
0.1192626953125,
0.2235107421875
],
[
0.1168212890625,
0.2259521484375,
0.49267578125,
0.2259521484375,
0.49267578125,
0.2392578125,
0.1168212890625,
0.2392578125
],
[
0.1422119140625,
0.240478515625,
0.88232421875,
0.240478515625,
0.88232421875,
0.25390625,
0.1422119140625,
0.25390625
],
[
0.11517333984375,
0.25390625,
0.1854248046875,
0.255615234375,
0.1854248046875,
0.26953125,
0.11517333984375,
0.267578125
],
[
0.1168212890625,
0.27978515625,
0.88134765625,
0.27978515625,
0.88134765625,
0.294189453125,
0.1168212890625,
0.294189453125
],
[
0.1168212890625,
0.29541015625,
0.88134765625,
0.29541015625,
0.88134765625,
0.309326171875,
0.1168212890625,
0.309326171875
],
[
0.1168212890625,
0.31005859375,
0.88134765625,
0.31005859375,
0.88134765625,
0.323974609375,
0.1168212890625,
0.323974609375
],
[
0.1168212890625,
0.32568359375,
0.88134765625,
0.32568359375,
0.88134765625,
0.339111328125,
0.1168212890625,
0.339111328125
],
[
0.1422119140625,
0.35107421875,
0.88232421875,
0.35107421875,
0.88232421875,
0.3642578125,
0.1422119140625,
0.3642578125
],
[
0.11517333984375,
0.3662109375,
0.384033203125,
0.3662109375,
0.384033203125,
0.380126953125,
0.11517333984375,
0.380126953125
],
[
0.11517333984375,
0.39013671875,
0.88134765625,
0.391357421875,
0.88134765625,
0.406005859375,
0.11517333984375,
0.40478515625
],
[
0.1168212890625,
0.406005859375,
0.88134765625,
0.406005859375,
0.88134765625,
0.42041015625,
0.1168212890625,
0.42041015625
],
[
0.1168212890625,
0.421630859375,
0.6337890625,
0.421630859375,
0.6337890625,
0.436279296875,
0.1168212890625,
0.436279296875
],
[
0.11517333984375,
0.4443359375,
0.88134765625,
0.44580078125,
0.88134765625,
0.460205078125,
0.11517333984375,
0.458984375
],
[
0.1168212890625,
0.462158203125,
0.8798828125,
0.462158203125,
0.8798828125,
0.47607421875,
0.1168212890625,
0.47607421875
],
[
0.1282958984375,
0.474853515625,
0.88134765625,
0.47607421875,
0.88134765625,
0.490478515625,
0.1282958984375,
0.4892578125
],
[
0.11517333984375,
0.4912109375,
0.471435546875,
0.490478515625,
0.471435546875,
0.5048828125,
0.11517333984375,
0.505859375
],
[
0.1405029296875,
0.51513671875,
0.8798828125,
0.5166015625,
0.8798828125,
0.53076171875,
0.1405029296875,
0.52978515625
],
[
0.1168212890625,
0.5322265625,
0.303955078125,
0.5322265625,
0.303955078125,
0.54541015625,
0.1168212890625,
0.54541015625
],
[
0.125,
0.546875,
0.88134765625,
0.546875,
0.88134765625,
0.56103515625,
0.125,
0.56103515625
],
[
0.1168212890625,
0.56103515625,
0.88134765625,
0.5625,
0.88134765625,
0.5771484375,
0.1168212890625,
0.57568359375
],
[
0.11358642578125,
0.5771484375,
0.365234375,
0.57568359375,
0.365234375,
0.59033203125,
0.11358642578125,
0.59130859375
],
[
0.1405029296875,
0.59228515625,
0.60888671875,
0.59130859375,
0.60888671875,
0.60595703125,
0.1405029296875,
0.60693359375
],
[
0.1168212890625,
0.6171875,
0.88134765625,
0.6171875,
0.88134765625,
0.630859375,
0.1168212890625,
0.630859375
],
[
0.11517333984375,
0.6318359375,
0.268798828125,
0.63330078125,
0.268798828125,
0.64794921875,
0.11517333984375,
0.646484375
],
[
0.11517333984375,
0.6572265625,
0.298095703125,
0.6572265625,
0.298095703125,
0.67041015625,
0.11517333984375,
0.67041015625
],
[
0.86767578125,
0.6572265625,
0.8857421875,
0.6572265625,
0.8857421875,
0.671875,
0.86767578125,
0.671875
],
[
0.1168212890625,
0.68310546875,
0.88232421875,
0.68310546875,
0.88232421875,
0.69775390625,
0.1168212890625,
0.69775390625
],
[
0.11517333984375,
0.69873046875,
0.88232421875,
0.69873046875,
0.88232421875,
0.7119140625,
0.11517333984375,
0.7119140625
],
[
0.11517333984375,
0.71337890625,
0.310546875,
0.71337890625,
0.310546875,
0.72802734375,
0.11517333984375,
0.72802734375
],
[
0.1168212890625,
0.73779296875,
0.29736328125,
0.73779296875,
0.29736328125,
0.75244140625,
0.1168212890625,
0.75244140625
],
[
0.86767578125,
0.7373046875,
0.88623046875,
0.7373046875,
0.88623046875,
0.751953125,
0.86767578125,
0.751953125
],
[
0.11273193359375,
0.76123046875,
0.30810546875,
0.76318359375,
0.30810546875,
0.78076171875,
0.11273193359375,
0.77880859375
],
[
0.11767578125,
0.78515625,
0.7509765625,
0.78515625,
0.7509765625,
0.798828125,
0.11767578125,
0.798828125
],
[
0.1258544921875,
0.8017578125,
0.88134765625,
0.8017578125,
0.88134765625,
0.81494140625,
0.1258544921875,
0.81494140625
],
[
0.11767578125,
0.81640625,
0.8798828125,
0.81640625,
0.8798828125,
0.830078125,
0.11767578125,
0.830078125
],
[
0.11767578125,
0.83203125,
0.88134765625,
0.83203125,
0.88134765625,
0.84521484375,
0.11767578125,
0.84521484375
],
[
0.1168212890625,
0.8466796875,
0.88134765625,
0.8466796875,
0.88134765625,
0.85986328125,
0.1168212890625,
0.85986328125
],
[
0.11767578125,
0.8623046875,
0.88134765625,
0.8623046875,
0.88134765625,
0.876953125,
0.11767578125,
0.876953125
],
[
0.1168212890625,
0.876953125,
0.2459716796875,
0.876953125,
0.2459716796875,
0.89013671875,
0.1168212890625,
0.89013671875
],
[
0.490234375,
0.93603515625,
0.50732421875,
0.93603515625,
0.50732421875,
0.95068359375,
0.490234375,
0.95068359375
]
] |
[
"3Concepts",
"在本节中，我们介绍所提出的概念并显示它们的一些理论性质。此外，我们还讨论了所提出的概念与一些现",
"有相关概念之间的联系和区别。。",
"3.1Proposed concepts",
"在成对图形中，每个边表示两个节点之间的连接，因此删除任何一个节点自然会导致边的完全无效。相反",
"即使删除了某些组成节点，具有三个或更多节点的Hyperedge仍然表示其余节点之间的交互。正如我们在图",
"中讨论和显示的那样。1，Def中的直接概括。3毫无根据地假设了脆弱的超级边缘并遭受信息丢失。我们",
"试图通过考虑非脆弱的超图来更好地揭示超图的结构",
"因此，我们引人了超边缘分数阈值t，该阈值确定了维持超边缘所需的组成节点的最小比例，从而导致",
"Def。5。",
"Definition 5((k,t) -hypercore). Given H = (V,E), k N, and t E [0, 1], the (k, t) -hypercore of H , denoted by",
"Ck,t(H) = (V', E') , is the maximal (in terms of total size) subhypergraph of H where (1) every node in has degree",
"at least k (i.e., is contained in at least k hyperedges) within Ck,t and (2) at least t proportion of the constituent nodes",
"remain in every hyperedge of Ck,t(H) . Formally, d(v; Ck,t(H)) ≥ k, Vu E V' and e, neil ≥ tleil, Vi E Ie C Ie .",
"请注意，由于子超图的定义，（k，t）-hypercore的定义要求每个超边缘中至少保留两个节点（请参见Sec。",
"2)。另请参见Alg中的Line6。1",
"Definition 6(t -hypercoreness). Given H = (V,E) and t E [o, 1], the t -hypercoreness of u E V, denoted by",
"Ct(v; H) , is the maximum positive integer such that u is in the (ct(u),t) -hypercore, i.e., Ct(u) = max[k N : u E",
"V(Ck,t)) . We call ct(H) := max[ct(u) : v E V) the t -hypercoreness of H .",
"Definition 7(k -fraction). Given H = (V,E) and k E N, the k -fraction of u e V, denoted by fk(u; H), is the",
"maximum real number in [0, 1] such that is in the (k, fa(v)) -hypercore, i.e., fk(o) = max(t E [0, 1] : u E V(Ck,t))",
"For the completeness of definition, if {t E [0,1] : E V(Ch.t)} = , we let fu(v; H) =-1. We call",
"f(H) :=max(f(v) : e V) the k -fraction of H.",
"请注意，建议的概念可扩展到加权超图。具体来说，只要我们对加权超图上的节点度和超边缘分数有严",
"格的定义，扩展就很简单。",
"Example.在图中。1，当t改变时，每个节点的t-hypercoreness也会改变。具体来说，当t≤时，四个",
"节点具有t-hypercoreness3。它们在<t≤号时具有t-hypercoreness2，在t>号时具有t-hypercoreness1",
"，这意味着它们的3-分数为，2",
"以下命题表明（h，t）-hypercore定义良好，并且具有双向包含属性。",
"Proposition 1 (Existence and uniqueness). Given any hypergraph H , k E N, and t E [o, 1] , Ch,t uniquely exists",
"and is possibly empty :",
"Proof. See Appendix A.1.",
"口",
"Proposition 2 (Two-way containment). Let H be any hypergraph. Fix any k E N, for any O ≤ ti < t2 ≤ 1 ,",
"Ck,t2(H) is a subhypergraph of Ck,ti(H) . Similarly, fix any t E [0,1], for any ki < k2 E N, Ck2,t(H) is a",
"subhypergraph of Ck1,t(H) .",
"Proof. See Appendix A.2.",
"口",
"3.2Related concepts",
"下面，我们讨论一些现有的相关概念，尤其是它们之间的联系和区别以及我们提出的概念",
"Existingvariantsofhypercores.如Sec中所述。2（参见Def。3），大多数以前的作品（Huaetal.,2023",
"Luoetal.,2021,2022;Gabertetal.，2021a,b;Sunetal.2020）都是基于假设脆弱的超边缘(i。e。，当任何节",
"点离开它时，一个超边缘就会被移除)，这相当于（k,t）-hypercore与t=1(i。e。，（k,t）-超核心的特例)。",
"Limniosetal.（2021)将给定超图H的（k;）-超核定义为H的最大子超图，其中每个节点在子超图中至少具",
"有k（i.e.,iscontainedinatleastkhyperedges）度，并且每个超图至少包含l节点。根据这个概念，我们可以定",
"义-hypercoreness",
"4"
] |
{
"arxiv_id": "2301.08440",
"language": "zh",
"page_idx": 3
}
|
||
2205.00302-zh-1
|
2205.00302-zh-1.jpg
|
retically guaranteed to quantify the marginal contribution of individual modalities and the crossmodal cooperation.
2. With the SHAPE scores, we systematically evalu-
ate the usage of multi-modal information by models using different fusion methods on the different multi-modal datasets for different tasks.
3. We extend the perceptual score to a more infor-
mative fne-grained version. Moreover, we demonstrate that the perceptual score implicitly measures the correlation among modalities within the dataset.
# 2. Related Work
2.1. Multi-modal Fusion Method
融合策略大致可分为早期融合、晚期融合和混合融合 (Baltrušaitis et al.,2018)。在早期的融合中，人们直接连接来自各个模态的输人，以构建统一的输入嵌入。在这种情况下，单路径模型足以进行预测，从而使训练管道更容易$\operatorname{E}($Kiela et al.,2019) 。后期融合通常以不同的途径处理单个模态。融合操作通常在每个路径的输出上实现。后一种融合策略忽略了低层特征交互，使得模态的联合表示难以学习。然而，在实践中，后期融合使模型使用不同的预训练模型为不同的路径，使其比早期融合更灵活 (Rodrigues Makiuchi et al., 2019) 。混合融合没有特定的配方。不同的模式可以在不同的阶段，不同的功能级别以及不同的模块之间进行集成。它可以有效地促进各个模式之间的合作。一般来说，最优的混合融合策略取决于底层任务的类型。例如，在 VQA 任务中，基于共同注意力的融合方法被证明比其他策略更好 (Liu et al.,2019)
2.2. Modality Contribution Attribution
目前，大多数评估工作都集中在功能重要性上。例如， Lundberg & Lee (2017) 提出了 SHapley 加法表达式(SHAP)值，该值可以使用 Shapley 值的特征近似值来估计特征重要性。为了评估模态重要性，Hessel $\&$ Lee (2020)采用了间接归因法。他们设计了一种称为经验多模态加法函数投影 (EMAP)的工具，该工具是一种有效的模态，可以消除参考多模态模型中的跨模态合作。通过比较参考多模态模型及其 EMAP 的性能，可以估计跨模态合作的贡献
与我们最相关的工作是感知得分，它是由 Gat et al. (2021)提出的，用于评估给定模型的输入数据中模态的重要性。他们采用了一种基于置换的方法来消除模态影响。具体来说，让$x^M$表示输入数据点中目标模态$M$的特征。为了消除$M$的影响，将$x^M$替换为从$\bar{\text{另一个数据点随机选择的数据，即。e。}},x^{\prime M}$。对置换数据产生的分类器称为置换分类器。$M$的感知得分定义为正常分类器和相应的置换分类器的精度之间的
归一化差
值得注意的是，感知得分并未考虑目标模态$M$的边际贡献。相反，它以数据集中的所有其他模态为条件来衡量$M$的贡献。两种类型的贡献仅在双模式情况下是相同的。在存在更多模态的情况下，它需要一种更复杂的方法来计算边际贡献。此外，仍然缺乏一种定量方法来隔离不同模式之间的合作
## $2.3.{\mathrm{~Shapley~Value}}$
Shapley 值$\overline{\mathrm{E}}($Shapley,1953) 最初是在联盟博弈论中提出的，目的是根据玩家对总支出的贡献来公平分配玩家之间的支出。多人游戏通常被定义为其参与者的实值效用函数$V$。让 $\mathcal{N}:=\{1,\ldots,n\}$ 表示游戏中的玩家集。玩家可以组成不同的联盟(i。e。, $\mathcal{N}$的子集), 其净效应可以是合作的，也可以是竞争的。考虑玩家$i$ 和子集$S\subseteq \mathcal{N} / \{ i\}$ , $i$对$S$的边际贡献可以通过
$$m_i(S;V)=V(S\cup\{i\})-V(S).$$
进行评估。由于玩家可以与任何其他玩家进行交互，因此为了评估$i$在游戏$V(\mathcal{N})$中的边际贡献，需要对所有可能的$m_i(S;V)$进行平均。这正是 Shapley 值$\phi_i$ 的定义方式，
$$\begin{aligned}\phi_i(\mathcal{N};V):=\sum_{S\subseteq\mathcal{N}/\{i\}}\frac{|S|!(n-|S|-1)!}{n!}m_i(S;V).\end{aligned}$$
认为支出的公平归属满足效率，对称性，假性和可加性的属性，而 Shapley 值是唯一具有这四个属性E (Molnar,2020) 的归属方法。更准确地说，
Efficiency 所有个人奖励的总和为总净支出，i。e。, $\sum _{i}\phi _{i}= V( \mathcal{N} ) - V( \emptyset )$ ,其中$V(\emptyset)$表示没有任何玩家的基线奖励。
Symmetry 如果两个成员$i,j$对所有可能的联盟做出同等贡献，则 i。e。,$V(S\cup\{i\})=V(S\cup\{j\}),\forall S\subseteq$ $\mathcal{N} / \{ i, j\}$ ,那么他们有同样的奖励$, \phi _i= \phi _j$ 。
Dummy 一个虚拟播放器$i$ ,它对任何联盟都没有贡献，i。e。$,V(S\cup\{i\})=V(S)+V(\{i\}),\forall S\subseteq\mathcal{N}/\{i\}$ ,有如独奏的奖励$, \phi _i= V( \{ i\} )$ 。
Additivity 考虑一个组合游戏$U(S)=V(S)+$ $W( S) , \forall S\subseteq \mathcal{N}$ ,它的奖励是单个游戏的奖励之和， i。e。$\phi _{i}( U) = \phi _{i}( V) + \phi _{i}( W)$
这四个属性又可以被视为定义公平归因方法的公理， 并且已经证明它们唯一地导致了 Shapley 值$\overline{\mathrm{E}}($Weber, 1988) 的定义
|
[
[
0.12335205078125,
0.08709716796875,
0.4755859375,
0.08709716796875,
0.4755859375,
0.10101318359375,
0.12335205078125,
0.10101318359375
],
[
0.5009765625,
0.08837890625,
0.56787109375,
0.08837890625,
0.56787109375,
0.0997314453125,
0.5009765625,
0.0997314453125
],
[
0.1217041015625,
0.10162353515625,
0.473876953125,
0.10162353515625,
0.473876953125,
0.11553955078125,
0.1217041015625,
0.11553955078125
],
[
0.50244140625,
0.10919189453125,
0.8837890625,
0.10919189453125,
0.8837890625,
0.12249755859375,
0.50244140625,
0.12249755859375
],
[
0.1217041015625,
0.117431640625,
0.2607421875,
0.11871337890625,
0.2607421875,
0.1300048828125,
0.1217041015625,
0.1287841796875
],
[
0.50244140625,
0.125,
0.8837890625,
0.125,
0.8837890625,
0.1383056640625,
0.50244140625,
0.1383056640625
],
[
0.1004638671875,
0.14453125,
0.473876953125,
0.14453125,
0.473876953125,
0.1578369140625,
0.1004638671875,
0.1578369140625
],
[
0.5009765625,
0.1395263671875,
0.8837890625,
0.1395263671875,
0.8837890625,
0.15283203125,
0.5009765625,
0.15283203125
],
[
0.12091064453125,
0.1590576171875,
0.473876953125,
0.1590576171875,
0.473876953125,
0.17236328125,
0.12091064453125,
0.17236328125
],
[
0.5009765625,
0.1546630859375,
0.8837890625,
0.1546630859375,
0.8837890625,
0.1685791015625,
0.5009765625,
0.1685791015625
],
[
0.12335205078125,
0.175537109375,
0.471435546875,
0.175537109375,
0.471435546875,
0.186279296875,
0.12335205078125,
0.186279296875
],
[
0.50244140625,
0.17041015625,
0.8857421875,
0.17041015625,
0.8857421875,
0.184326171875,
0.50244140625,
0.184326171875
],
[
0.1217041015625,
0.189453125,
0.406982421875,
0.189453125,
0.406982421875,
0.20263671875,
0.1217041015625,
0.20263671875
],
[
0.50244140625,
0.186279296875,
0.7294921875,
0.186279296875,
0.7294921875,
0.1962890625,
0.50244140625,
0.1962890625
],
[
0.1004638671875,
0.2171630859375,
0.473876953125,
0.2171630859375,
0.473876953125,
0.2310791015625,
0.1004638671875,
0.2310791015625
],
[
0.5009765625,
0.2171630859375,
0.640625,
0.2171630859375,
0.640625,
0.2310791015625,
0.5009765625,
0.2310791015625
],
[
0.1217041015625,
0.231689453125,
0.47216796875,
0.231689453125,
0.47216796875,
0.24560546875,
0.1217041015625,
0.24560546875
],
[
0.5009765625,
0.240478515625,
0.8837890625,
0.240478515625,
0.8837890625,
0.25390625,
0.5009765625,
0.25390625
],
[
0.12091064453125,
0.2474365234375,
0.47216796875,
0.2474365234375,
0.47216796875,
0.261474609375,
0.12091064453125,
0.261474609375
],
[
0.50244140625,
0.2568359375,
0.8837890625,
0.2568359375,
0.8837890625,
0.267578125,
0.50244140625,
0.267578125
],
[
0.1217041015625,
0.26318359375,
0.47216796875,
0.26318359375,
0.47216796875,
0.276611328125,
0.1217041015625,
0.276611328125
],
[
0.5009765625,
0.26953125,
0.88232421875,
0.26953125,
0.88232421875,
0.28271484375,
0.5009765625,
0.28271484375
],
[
0.12091064453125,
0.279052734375,
0.1805419921875,
0.279052734375,
0.1805419921875,
0.290283203125,
0.12091064453125,
0.290283203125
],
[
0.50341796875,
0.287353515625,
0.88232421875,
0.287353515625,
0.88232421875,
0.29736328125,
0.50341796875,
0.29736328125
],
[
0.50244140625,
0.300537109375,
0.88232421875,
0.300537109375,
0.88232421875,
0.314453125,
0.50244140625,
0.314453125
],
[
0.08660888671875,
0.311279296875,
0.233642578125,
0.311279296875,
0.233642578125,
0.324462890625,
0.08660888671875,
0.324462890625
],
[
0.50341796875,
0.317626953125,
0.8857421875,
0.317626953125,
0.8857421875,
0.32763671875,
0.50341796875,
0.32763671875
],
[
0.50244140625,
0.330810546875,
0.85107421875,
0.330810546875,
0.85107421875,
0.3447265625,
0.50244140625,
0.3447265625
],
[
0.0882568359375,
0.33837890625,
0.328369140625,
0.33837890625,
0.328369140625,
0.3515625,
0.0882568359375,
0.3515625
],
[
0.0906982421875,
0.361083984375,
0.473876953125,
0.361083984375,
0.473876953125,
0.375,
0.0906982421875,
0.375
],
[
0.5810546875,
0.365478515625,
0.8037109375,
0.365478515625,
0.8037109375,
0.37890625,
0.5810546875,
0.37890625
],
[
0.0906982421875,
0.376953125,
0.473876953125,
0.376953125,
0.473876953125,
0.39013671875,
0.0906982421875,
0.39013671875
],
[
0.0906982421875,
0.391357421875,
0.473876953125,
0.391357421875,
0.473876953125,
0.40478515625,
0.0906982421875,
0.40478515625
],
[
0.50244140625,
0.397705078125,
0.8837890625,
0.397705078125,
0.8837890625,
0.410888671875,
0.50244140625,
0.410888671875
],
[
0.08905029296875,
0.4072265625,
0.473876953125,
0.4072265625,
0.473876953125,
0.42041015625,
0.08905029296875,
0.42041015625
],
[
0.5009765625,
0.41357421875,
0.8857421875,
0.41357421875,
0.8857421875,
0.4267578125,
0.5009765625,
0.4267578125
],
[
0.08905029296875,
0.42041015625,
0.47216796875,
0.421630859375,
0.47216796875,
0.436279296875,
0.08905029296875,
0.43505859375
],
[
0.50244140625,
0.4287109375,
0.8837890625,
0.4287109375,
0.8837890625,
0.442626953125,
0.50244140625,
0.442626953125
],
[
0.0906982421875,
0.4375,
0.473876953125,
0.4375,
0.473876953125,
0.45068359375,
0.0906982421875,
0.45068359375
],
[
0.50244140625,
0.4443359375,
0.58837890625,
0.4443359375,
0.58837890625,
0.455810546875,
0.50244140625,
0.455810546875
],
[
0.0931396484375,
0.453857421875,
0.471435546875,
0.453857421875,
0.471435546875,
0.464111328125,
0.0931396484375,
0.464111328125
],
[
0.0906982421875,
0.46728515625,
0.473876953125,
0.46728515625,
0.473876953125,
0.48095703125,
0.0906982421875,
0.48095703125
],
[
0.66845703125,
0.466552734375,
0.86767578125,
0.4765625,
0.865234375,
0.501953125,
0.6650390625,
0.4912109375
],
[
0.0906982421875,
0.48291015625,
0.47216796875,
0.48291015625,
0.47216796875,
0.496826171875,
0.0906982421875,
0.496826171875
],
[
0.521484375,
0.48291015625,
0.62744140625,
0.48095703125,
0.62744140625,
0.49560546875,
0.521484375,
0.49755859375
],
[
0.62353515625,
0.48095703125,
0.67626953125,
0.47607421875,
0.6787109375,
0.4912109375,
0.6259765625,
0.496826171875
],
[
0.0906982421875,
0.49755859375,
0.47216796875,
0.49755859375,
0.47216796875,
0.51123046875,
0.0906982421875,
0.51123046875
],
[
0.7236328125,
0.492431640625,
0.74267578125,
0.492431640625,
0.74267578125,
0.505859375,
0.7236328125,
0.505859375
],
[
0.61279296875,
0.5,
0.6689453125,
0.5,
0.6689453125,
0.5146484375,
0.61279296875,
0.5146484375
],
[
0.08905029296875,
0.51220703125,
0.473876953125,
0.51220703125,
0.473876953125,
0.52587890625,
0.08905029296875,
0.52587890625
],
[
0.09234619140625,
0.52880859375,
0.471435546875,
0.52880859375,
0.471435546875,
0.5390625,
0.09234619140625,
0.5390625
],
[
0.0906982421875,
0.54443359375,
0.47216796875,
0.54443359375,
0.47216796875,
0.5546875,
0.0906982421875,
0.5546875
],
[
0.5009765625,
0.54052734375,
0.8837890625,
0.54052734375,
0.8837890625,
0.5537109375,
0.5009765625,
0.5537109375
],
[
0.0906982421875,
0.55810546875,
0.47216796875,
0.55810546875,
0.47216796875,
0.5712890625,
0.0906982421875,
0.5712890625
],
[
0.5009765625,
0.5546875,
0.8857421875,
0.5546875,
0.8857421875,
0.568359375,
0.5009765625,
0.568359375
],
[
0.0906982421875,
0.57275390625,
0.473876953125,
0.57275390625,
0.473876953125,
0.5859375,
0.0906982421875,
0.5859375
],
[
0.5,
0.568359375,
0.7958984375,
0.5693359375,
0.7958984375,
0.583984375,
0.5,
0.58349609375
],
[
0.0906982421875,
0.58837890625,
0.43310546875,
0.58837890625,
0.43310546875,
0.6015625,
0.0906982421875,
0.6015625
],
[
0.5009765625,
0.60595703125,
0.8837890625,
0.60693359375,
0.8837890625,
0.6220703125,
0.5009765625,
0.62060546875
],
[
0.0882568359375,
0.619140625,
0.374267578125,
0.619140625,
0.374267578125,
0.63330078125,
0.0882568359375,
0.63330078125
],
[
0.50244140625,
0.6220703125,
0.8857421875,
0.6220703125,
0.8857421875,
0.63623046875,
0.50244140625,
0.63623046875
],
[
0.50244140625,
0.63720703125,
0.58837890625,
0.63720703125,
0.58837890625,
0.6484375,
0.50244140625,
0.6484375
],
[
0.0906982421875,
0.64306640625,
0.473876953125,
0.64306640625,
0.473876953125,
0.6572265625,
0.0906982421875,
0.6572265625
],
[
0.0906982421875,
0.6591796875,
0.473876953125,
0.6591796875,
0.473876953125,
0.67236328125,
0.0906982421875,
0.67236328125
],
[
0.08905029296875,
0.67236328125,
0.47216796875,
0.67236328125,
0.47216796875,
0.68603515625,
0.08905029296875,
0.68603515625
],
[
0.5009765625,
0.6748046875,
0.8837890625,
0.6748046875,
0.8837890625,
0.68798828125,
0.5009765625,
0.68798828125
],
[
0.08905029296875,
0.68798828125,
0.473876953125,
0.6875,
0.473876953125,
0.70068359375,
0.08905029296875,
0.7021484375
],
[
0.5,
0.68798828125,
0.8857421875,
0.689453125,
0.8857421875,
0.7041015625,
0.5,
0.70263671875
],
[
0.09234619140625,
0.705078125,
0.471435546875,
0.705078125,
0.471435546875,
0.71533203125,
0.09234619140625,
0.71533203125
],
[
0.496826171875,
0.7021484375,
0.8203125,
0.7041015625,
0.8203125,
0.7216796875,
0.496826171875,
0.7197265625
],
[
0.0906982421875,
0.71826171875,
0.47216796875,
0.71826171875,
0.47216796875,
0.73193359375,
0.0906982421875,
0.73193359375
],
[
0.0906982421875,
0.734375,
0.47216796875,
0.734375,
0.47216796875,
0.74755859375,
0.0906982421875,
0.74755859375
],
[
0.5009765625,
0.7421875,
0.8837890625,
0.7421875,
0.8837890625,
0.755859375,
0.5009765625,
0.755859375
],
[
0.0906982421875,
0.74853515625,
0.473876953125,
0.74853515625,
0.473876953125,
0.76220703125,
0.0906982421875,
0.76220703125
],
[
0.50244140625,
0.7568359375,
0.8837890625,
0.7568359375,
0.8837890625,
0.77099609375,
0.50244140625,
0.77099609375
],
[
0.08905029296875,
0.7646484375,
0.3203125,
0.7646484375,
0.3203125,
0.77783203125,
0.08905029296875,
0.77783203125
],
[
0.5,
0.77099609375,
0.736328125,
0.77294921875,
0.736328125,
0.787109375,
0.5,
0.78515625
],
[
0.09234619140625,
0.787109375,
0.47216796875,
0.787109375,
0.47216796875,
0.80029296875,
0.09234619140625,
0.80029296875
],
[
0.0906982421875,
0.8017578125,
0.473876953125,
0.8017578125,
0.473876953125,
0.81494140625,
0.0906982421875,
0.81494140625
],
[
0.5,
0.80810546875,
0.88818359375,
0.8095703125,
0.88818359375,
0.82373046875,
0.5,
0.82275390625
],
[
0.09234619140625,
0.8173828125,
0.47216796875,
0.8173828125,
0.47216796875,
0.83056640625,
0.09234619140625,
0.83056640625
],
[
0.5009765625,
0.82470703125,
0.88623046875,
0.82470703125,
0.88623046875,
0.83837890625,
0.5009765625,
0.83837890625
],
[
0.0906982421875,
0.83056640625,
0.473876953125,
0.83056640625,
0.473876953125,
0.8447265625,
0.0906982421875,
0.8447265625
],
[
0.0906982421875,
0.8466796875,
0.47216796875,
0.8466796875,
0.47216796875,
0.861328125,
0.0906982421875,
0.861328125
],
[
0.5,
0.84130859375,
0.7158203125,
0.84130859375,
0.7158203125,
0.8515625,
0.5,
0.8515625
],
[
0.0906982421875,
0.85986328125,
0.473876953125,
0.85986328125,
0.473876953125,
0.875,
0.0906982421875,
0.875
],
[
0.5009765625,
0.8623046875,
0.88232421875,
0.8623046875,
0.88232421875,
0.87548828125,
0.5009765625,
0.87548828125
],
[
0.08905029296875,
0.876953125,
0.47216796875,
0.876953125,
0.47216796875,
0.89013671875,
0.08905029296875,
0.89013671875
],
[
0.5,
0.876953125,
0.8837890625,
0.87548828125,
0.8837890625,
0.89013671875,
0.5,
0.8916015625
],
[
0.09234619140625,
0.89453125,
0.471435546875,
0.89453125,
0.471435546875,
0.90478515625,
0.09234619140625,
0.90478515625
],
[
0.50244140625,
0.89306640625,
0.5947265625,
0.89306640625,
0.5947265625,
0.90478515625,
0.50244140625,
0.90478515625
]
] |
[
"retically guaranteed to quantify the marginal con-",
"归一化差",
"tribution of individual modalities and the cross-",
"值得注意的是，感知得分并未考虑目标模态M的边际",
"modal cooperation.",
"贡献。相反，它以数据集中的所有其他模态为条件来",
"2. With the SHAPE scores, we systematically evalu-",
"衡量M的贡献。两种类型的贡献仅在双模式情况下是",
"ate the usage of multi-modal information by mod-",
"相同的。在存在更多模态的情况下，它需要一种更复",
"elsusingdifferent fusionmethods on the different",
"杂的方法来计算边际贡献。此外，仍然缺乏一种定量",
"multi-modal datasets for different tasks.",
"方法来隔离不同模式之间的合作",
"3. We extend the perceptual score to a more infor-",
"2.3. Shapley Value",
"mative fine-grained version. Moreover, we demon",
"Shapley值E(Shapley，1953）最初是在联盟博奔论中",
"strate that the perceptual score implicitly mea",
"提出的，自的是根据玩家对总支出的贡献来公平分配",
"sures the correlation among modalities within the",
"玩家之间的支出。多人游戏通常被定义为其参与者的",
"dataset.",
"实值效用函数V。让N：=1...：.n表示游戏中的",
"玩家集。玩家可以组成不同的联盟(i。e。，N的子集）",
"2. Related Work",
"其净效应可以是合作的：也可以是竞争的。考虑玩家2",
"和子集SCN/i，i对S的边际贡献可以通过",
"2.1. Multi-modal Fusion Method",
"融合策略大致可分为早期融合、晚期融合和混合融",
"mi(S; V) = V(su(i)) - V(S).",
"合（Baltrusaitisetal.，2018）。在早期的融合中，人",
"们直接连接来自各个模态的输入，以构建统一的输人",
"进行评估。由于玩家可以与任何其他玩家进行交互，因",
"嵌入。在这种情况下，单路径模型足以进行预测，从",
"此为了评估i在游戏V(V）中的边际贡献，需要对所",
"而使训练管道更容易E(Kielaetal.，2019）。后期融合",
"有可能的m（S;V）进行平均。这正是Shapley值Φi",
"通常以不同的途径处理单个模态。融合操作通常在每",
"的定义方式，",
"个路径的输出上实现。后一种融合策略忽略了低层特",
"征交互，使得模态的联合表示难以学习。然而，在实",
"[SI(n - [S/ - 1) m;(S; ).",
"践中，后期融合使模型使用不同的预训练模型为不同",
"pi(N;V) :=",
"门",
"的路径，使其比早期融合更灵活（RodriguesMakiuchi",
"n!",
"SCN/(i)",
"etal.，2019）。混合融合没有特定的配方。不同的模式",
"可以在不同的阶段，不同的功能级别以及不同的模块",
"之间进行集成。它可以有效地促进各个模式之间的合",
"认为支出的公平归属满足效率，对称性，假性和可",
"作。一般来说，最优的混合融合策略取决于底层任务",
"加性的属性，而Shapley值是唯一具有这四个属性E",
"的类型。例如，在VQA任务中，基于共同注意力的融",
"(Molnar，2020）的归属方法。更准确地说",
"合方法被证明比其他策略更好（Liuetal.，2019)",
"Eficiency所有个人奖励的总和为总净支出，i。e。",
"2.2.Modality Contribution Attribution",
"Z,Φi=V(W)-V(の），其中V(の）表示没有任何玩家",
"的基线奖励。",
"目前，大多数评估工作都集中在功能重要性上。例如，",
"Lundberg&Lee（2017）提出了SHapley加法表达式",
"（SHAP）值，该值可以使用Shapley值的特征近似值",
"Symmetry如果两个成员i,j对所有可能的联盟做出",
"来估计特征重要性。为了评估模态重要性，Hessel&",
"同等贡献，则i。e。,V(Su(i)=V(S(i)),VS C",
"Lee（2020）采用了间接归因法。他们设计了一种称为",
"N/i,i），那么他们有同样的奖励，：=Φi。",
"经验多模态加法函数投影（EMAP）的工具，该工具",
"是一种有效的模态，可以消除参考多模态模型中的跨",
"Dummy一个虚拟播放器i，它对任何联盟都没有贡",
"模态合作。通过比较参考多模态模型及其EMAP的性",
"献, i。 e。, V(sU{i)) = V(S)+V((i)), vs CN/i)",
"能，可以估计跨模态合作的贡献",
"，有如独奏的奖励，中i=V（{i）。",
"与我们最相关的工作是感知得分，它是由Gatetal.",
"(2021）提出的，用于评估给定模型的输入数据中模态",
"Additivity考虑一个组合游戏U(S）=V(S）+",
"的重要性。他们采用了一种基于置换的方法来消除模",
"W(S),VSN，它的奖励是单个游戏的奖励之和，",
"态影响。具体来说，让M表示输人数据点中目标模",
"态M的特征。为了消除M的影响，将aM替换为从",
"i。e。@iU)=Φi(V)+ΦiW)",
"另一个数据点随机选择的数据，即。e。，aM。对置",
"这四个属性又可以被视为定义公平归因方法的公理",
"换数据产生的分类器称为置换分类器。M的感知得分",
"并且已经证明它们唯一地导致了Shapley值E（Weber",
"定义为正常分类器和相应的置换分类器的精度之间的",
"1988）的定义"
] |
{
"arxiv_id": "2205.00302",
"language": "zh",
"page_idx": 1
}
|
||
2205.00302-zh-3
|
2205.00302-zh-3.jpg
|
对于一个具体的示例，让我们考虑两种模式的情况。在 4. Experiments
简单代数的几个步骤之后，我们到达
$$\begin{aligned}\mathcal{C}_{\{M_{1},M_{2}\}}&=V(M_{1},M_{2})-V(M_{1},0_{2})-V(0_{1},M_{2})\\&+V(0_1,0_2),\end{aligned}$$
,其中$0_1$和$0_2$分别表示模态$M_1,M_2$的基线值
3.3. Fine-grained Perceptual Scores
在本小节中，我们继续仔细研究感知得分$\operatorname{E}($Gat et al., 2021)。首先，我们将显示此分数隐式地衡量数据集$\mathcal{D}$ 中不同模态之间的相关性。然后，我们提出了感知分数的细粒度版本。下面我们只保留关键步骤，把详细的扣除留给补充材料
令$(x,y)$表示$\mathcal{D}$中的数据样本，其中$x\in\mathcal{X}$是输人特征$,y\in\{0,\ldots,k-1\}$ 是相应的标签。为了明确引用模态，我们将$x$分解为$( x_1^\mathcal{M} , x^{\mathcal{M} _2})$ ,其中$\mathcal{M}_1\subseteq\mathcal{M}$ 和$\mathcal{M}_2=\mathcal{M}/\mathcal{M}_1$是模态的子集。回想一下，模型精度定义为$V_f( \mathcal{D} ) = \mathbb{E} _{\mathcal{D} }[ \mathbb{I} _{y^* = y}]$ ,其中$\mathbb{I}$是恒等函数$,y^*$ 是指输入$x$处的模型预测。简单代数将引导我们到
$$V_f(\mathcal{D})=\int_{\mathcal{X}}dx^{\mathcal{M}_1}dx^{\mathcal{M}_2}\mathrm{Pr}(y*,x^{\mathcal{M}_1},x^{\mathcal{M}_2}),$$
,其中 Pr$(y*,x^{\mathcal{M}_1},x^{\mathcal{M}_2})$是$\mathcal{D}$上的概率分布
感知分数中基于置换的步骤将改变数据分布，从而产 预测此标签。我们注意到，文本数据通常与他们的情生分布为 Pr$(y,x^{\mathcal{M}_1})$Pr$(x^\mathcal{M}_2)$的新有效数据集。我们 绪标签对齐，而大量的图像似乎与标签无关。B-T4SA 将此有效数据集称为$\mathcal{D}^p$ 。现在，$\mathcal{M}_2$的感知分数是 代表了不同模态互补且相互弱关联的情况
$$\begin{aligned}P_{\mathcal{M}_{2};f;\mathcal{D}}&=\frac{1}{Z_{f}}(V_{f}(\mathcal{D})-V_{f}(\mathcal{D}^{p}))\\&=\frac{1}{Z_{f}}\int_{\mathcal{X}}dx^{\mathcal{M}_{1}}dx^{\mathcal{M}_{2}}P(x^{\mathcal{M}_{2}})\Delta_{\mathcal{M}_{1},\mathcal{M}_{2};f},\end{aligned}$$
,其中$\Delta_{\mathcal{M}_1,\mathcal{M}_2;f}=P(y^*,x^{\mathcal{M}_1}|x^{\mathcal{M}_2})-P(y^*,x^{\mathcal{M}_1})$。显然，除了对模型$f$到$y^*$的依赖之外，术语$\Delta_{\mathcal{M}_1,\mathcal{M}_2;f}$ 有效地衡量了数据集$\mathcal{D}$中$\mathcal{M}_1$和$\mathcal{M}_2$之间的相关性由于$P(x^\mathcal{M}_2)$是非负的，因此如果$x^\mathcal{M}_2$部分与$x^\mathcal{M}_1$ 来自同一类，那么我们将期望$\Delta_{M_1,M_2;f}\geq0$ 。我们将这种情况称为类内排列。同样，在类外排列中，$x^\mathcal{M}_2$ 来自与$x^\mathcal{M}_1$不同的类，我们期望$\Delta_M_1,M_2;f\leq0$
来反映。基于这一观察，我们建议使用课堂内和课堂
外的感知分数来代替原始的。它们的定义如下：
$$P_{\mathcal{M}_{i};f;\mathcal{D}}^{in/out}:=\frac{1}{Z_{f}}\begin{bmatrix}V_{f}(\mathcal{M})-V_{f}([\mathcal{M}/\mathcal{M}_{i}]^{\pm})\end{bmatrix},$$
我们这里只考虑模型归一化。最后，为了简单起见，我
们通常省略下标$f;\mathcal{D}$ ,只要它不引起混淆。
4.1. Setup
对于所有型号，我们使用 4 层变压器编码器作为骨干。注意头数为 8, 隐藏维度为 1024。所有模型都使用 Adam 优化器进行训练，并在第一个时期进行学习率预热。我们用 GloVe (Pennington et al., 2014) 编码文本数据，用 MFCC (Tiwari, 2010) 编码音频数据。对于图像数据，我们采用补丁嵌人技术，遵循 ViT 模型 (Dosovitskiy et al.,2020) 。值得注意的是，我们的模型实现了与基于监督的 SOTA 模型相当的性能
# 4.2. Datasets
现实生活中的应用程序对多模式数据提出了不同的需求。根据特定的任务，有些必须通过组合所有涉及的方式来处理，而在其他方式中，其他方式仅提供补充信息。此外，多模态数据集中的不同模态可能彼此之间只有微弱的相关性，或者它们可能是提供完全相同信息的不同表示。为了解决这些复杂的情况，我们故意选择三个数据集，如下所述
B-T4SA 数据集$\boxed{\mathrm{E}}($Vadicamo et al.,2017) 是从 Twitter 收集的，包括发布的文本(文本模态 T) 和相关的图像(视觉模态 V)。每个数据对都有一个情绪标签(正面、负面和中性)。任务是使用文本和图像输入来CMU-MOSEI 此数据集 (Zadeh et al.,2018) 是用于句子级情感分析和情感识别的最大数据集。它包含超过 65 小时的注释视频数据，包括音频，视频和文本模式。在我们的实验中，我们只使用音频(A)和文本(T) 模态。文本模态包括从音频数据生成的脚本。这样，CMU-MOSEI 数据集代表了不同模态互补且高度相关的情况
SNLI SNLI 数据集$\boxed{\mathrm{E}}($Bowman et al.,2015)是 570k 人为写的英语句子对的集合。每一对都包含一个前提和一个假设，其关系被标记为蕴涵，矛盾或中立。我们将把前提和假设视为两种不同的模式，即。e。、$T_p$
班级内和班级外的情况提供了相互冲突的驱动力，这 和$T_h$ 。基本任务是推断前提和假设之间的关系。显些驱动力会相互抵消，并且无法通过原始的感知得分 然，这项任务本质上依赖于两种模式，因此代表了不
同模式不可或缺的情况
4.3. Contribution and Cooperation
在本小节中，我们研究了三个多模态数据集上不同模型的 SHAPE 分数，证明了它们在捕获单个模态和跨
,其中$[\mathcal{M}/\mathcal{M}_i]^\pm$运算符通过类内置换 (+ )或类外 模态合作的边际贡献方面的能力。Table 1 总结了 B置换(- )删除$\mathcal{M}_i$模态。$Z_f$是通常的归一化因子。T4SA 数据集的结果。我们发现 T 的单峰模型的性能
与多模态模型一样好，而 V 的单峰模型仅略微抑制了随机猜测。SHAPE 分数与这些观察结果一致。对于
|
[
[
0.09234619140625,
0.08837890625,
0.473876953125,
0.08837890625,
0.473876953125,
0.0985107421875,
0.09234619140625,
0.0985107421875
],
[
0.4755859375,
0.0889892578125,
0.5009765625,
0.0889892578125,
0.5009765625,
0.09661865234375,
0.4755859375,
0.09661865234375
],
[
0.495849609375,
0.08587646484375,
0.6337890625,
0.08587646484375,
0.6337890625,
0.0997314453125,
0.495849609375,
0.0997314453125
],
[
0.09234619140625,
0.10162353515625,
0.351318359375,
0.10162353515625,
0.351318359375,
0.11553955078125,
0.09234619140625,
0.11553955078125
],
[
0.5,
0.10919189453125,
0.580078125,
0.111083984375,
0.580078125,
0.1256103515625,
0.5,
0.12371826171875
],
[
0.1070556640625,
0.1287841796875,
0.455078125,
0.1287841796875,
0.455078125,
0.143310546875,
0.1070556640625,
0.143310546875
],
[
0.50341796875,
0.1363525390625,
0.88134765625,
0.1363525390625,
0.88134765625,
0.146484375,
0.50341796875,
0.146484375
],
[
0.1748046875,
0.147705078125,
0.259033203125,
0.147705078125,
0.259033203125,
0.1590576171875,
0.1748046875,
0.1590576171875
],
[
0.5009765625,
0.1483154296875,
0.8837890625,
0.149658203125,
0.8837890625,
0.1634521484375,
0.5009765625,
0.1622314453125
],
[
0.50341796875,
0.1666259765625,
0.88232421875,
0.1666259765625,
0.88232421875,
0.1767578125,
0.50341796875,
0.1767578125
],
[
0.09234619140625,
0.175537109375,
0.43310546875,
0.175537109375,
0.43310546875,
0.186279296875,
0.09234619140625,
0.186279296875
],
[
0.50244140625,
0.179931640625,
0.8857421875,
0.179931640625,
0.8857421875,
0.1932373046875,
0.50244140625,
0.1932373046875
],
[
0.50341796875,
0.1944580078125,
0.88232421875,
0.1944580078125,
0.88232421875,
0.208984375,
0.50341796875,
0.208984375
],
[
0.0882568359375,
0.20703125,
0.34716796875,
0.20703125,
0.34716796875,
0.2203369140625,
0.0882568359375,
0.2203369140625
],
[
0.50341796875,
0.212158203125,
0.88232421875,
0.212158203125,
0.88232421875,
0.222900390625,
0.50341796875,
0.222900390625
],
[
0.0906982421875,
0.2298583984375,
0.473876953125,
0.2298583984375,
0.473876953125,
0.24365234375,
0.0906982421875,
0.24365234375
],
[
0.5009765625,
0.2259521484375,
0.8857421875,
0.2259521484375,
0.8857421875,
0.2392578125,
0.5009765625,
0.2392578125
],
[
0.0882568359375,
0.2442626953125,
0.473876953125,
0.24365234375,
0.473876953125,
0.25830078125,
0.0882568359375,
0.2587890625
],
[
0.5009765625,
0.240478515625,
0.86767578125,
0.240478515625,
0.86767578125,
0.25390625,
0.5009765625,
0.25390625
],
[
0.0906982421875,
0.260009765625,
0.471435546875,
0.260009765625,
0.471435546875,
0.2734375,
0.0906982421875,
0.2734375
],
[
0.0906982421875,
0.274658203125,
0.471435546875,
0.274658203125,
0.471435546875,
0.28857421875,
0.0906982421875,
0.28857421875
],
[
0.50244140625,
0.2734375,
0.6005859375,
0.2734375,
0.6005859375,
0.285400390625,
0.50244140625,
0.285400390625
],
[
0.09234619140625,
0.292236328125,
0.221435546875,
0.292236328125,
0.221435546875,
0.302978515625,
0.09234619140625,
0.302978515625
],
[
0.50244140625,
0.29541015625,
0.8837890625,
0.29541015625,
0.8837890625,
0.309326171875,
0.50244140625,
0.309326171875
],
[
0.0906982421875,
0.313232421875,
0.473876953125,
0.313232421875,
0.473876953125,
0.326904296875,
0.0906982421875,
0.326904296875
],
[
0.5009765625,
0.311279296875,
0.8837890625,
0.311279296875,
0.8837890625,
0.324462890625,
0.5009765625,
0.324462890625
],
[
0.0906982421875,
0.328857421875,
0.47216796875,
0.328857421875,
0.47216796875,
0.34228515625,
0.0906982421875,
0.34228515625
],
[
0.50244140625,
0.32763671875,
0.8837890625,
0.32763671875,
0.8837890625,
0.33837890625,
0.50244140625,
0.33837890625
],
[
0.08905029296875,
0.34228515625,
0.473876953125,
0.34228515625,
0.473876953125,
0.356689453125,
0.08905029296875,
0.356689453125
],
[
0.50244140625,
0.341552734375,
0.88232421875,
0.341552734375,
0.88232421875,
0.354736328125,
0.50244140625,
0.354736328125
],
[
0.0906982421875,
0.359130859375,
0.473876953125,
0.359130859375,
0.473876953125,
0.37255859375,
0.0906982421875,
0.37255859375
],
[
0.50244140625,
0.35791015625,
0.88232421875,
0.35791015625,
0.88232421875,
0.36865234375,
0.50244140625,
0.36865234375
],
[
0.0906982421875,
0.373779296875,
0.4697265625,
0.373779296875,
0.4697265625,
0.38818359375,
0.0906982421875,
0.38818359375
],
[
0.50244140625,
0.371826171875,
0.8857421875,
0.371826171875,
0.8857421875,
0.385009765625,
0.50244140625,
0.385009765625
],
[
0.08905029296875,
0.38818359375,
0.450927734375,
0.386962890625,
0.450927734375,
0.401611328125,
0.08905029296875,
0.40283203125
],
[
0.50341796875,
0.38818359375,
0.71337890625,
0.38818359375,
0.71337890625,
0.3984375,
0.50341796875,
0.3984375
],
[
0.200927734375,
0.417236328125,
0.4306640625,
0.41845703125,
0.4306640625,
0.436279296875,
0.200927734375,
0.43505859375
],
[
0.5009765625,
0.42041015625,
0.8857421875,
0.42041015625,
0.8857421875,
0.434326171875,
0.5009765625,
0.434326171875
],
[
0.50341796875,
0.4375,
0.88232421875,
0.4375,
0.88232421875,
0.447509765625,
0.50341796875,
0.447509765625
],
[
0.134033203125,
0.42236328125,
0.2034912109375,
0.42236328125,
0.2034912109375,
0.434326171875,
0.134033203125,
0.434326171875
],
[
0.0906982421875,
0.455810546875,
0.410888671875,
0.455810546875,
0.410888671875,
0.47021484375,
0.0906982421875,
0.47021484375
],
[
0.5048828125,
0.45263671875,
0.88232421875,
0.45263671875,
0.88232421875,
0.46337890625,
0.5048828125,
0.46337890625
],
[
0.50244140625,
0.466552734375,
0.8857421875,
0.466552734375,
0.8857421875,
0.479736328125,
0.50244140625,
0.479736328125
],
[
0.09234619140625,
0.479248046875,
0.473876953125,
0.479248046875,
0.473876953125,
0.492431640625,
0.09234619140625,
0.492431640625
],
[
0.50244140625,
0.48291015625,
0.88232421875,
0.48291015625,
0.88232421875,
0.49365234375,
0.50244140625,
0.49365234375
],
[
0.0906982421875,
0.49365234375,
0.473876953125,
0.49365234375,
0.473876953125,
0.50830078125,
0.0906982421875,
0.50830078125
],
[
0.5009765625,
0.496826171875,
0.8857421875,
0.496826171875,
0.8857421875,
0.51025390625,
0.5009765625,
0.51025390625
],
[
0.0882568359375,
0.50830078125,
0.463134765625,
0.5087890625,
0.463134765625,
0.5234375,
0.0882568359375,
0.52294921875
],
[
0.5009765625,
0.51025390625,
0.7939453125,
0.51123046875,
0.7939453125,
0.52587890625,
0.5009765625,
0.5244140625
],
[
0.1094970703125,
0.54150390625,
0.193603515625,
0.546875,
0.19287109375,
0.56103515625,
0.108642578125,
0.55615234375
],
[
0.200927734375,
0.5361328125,
0.2091064453125,
0.54248046875,
0.2017822265625,
0.54736328125,
0.193603515625,
0.54150390625
],
[
0.2115478515625,
0.54248046875,
0.338134765625,
0.54248046875,
0.338134765625,
0.556640625,
0.2115478515625,
0.556640625
],
[
0.1871337890625,
0.55029296875,
0.21728515625,
0.55029296875,
0.21728515625,
0.56298828125,
0.1871337890625,
0.56298828125
],
[
0.50244140625,
0.54541015625,
0.8837890625,
0.54541015625,
0.8837890625,
0.5595703125,
0.50244140625,
0.5595703125
],
[
0.5048828125,
0.5625,
0.8837890625,
0.5625,
0.8837890625,
0.57275390625,
0.5048828125,
0.57275390625
],
[
0.19287109375,
0.5693359375,
0.221435546875,
0.5712890625,
0.2181396484375,
0.58349609375,
0.191162109375,
0.580078125
],
[
0.226318359375,
0.5712890625,
0.455810546875,
0.5771484375,
0.455810546875,
0.59326171875,
0.226318359375,
0.58837890625
],
[
0.17236328125,
0.580078125,
0.236083984375,
0.5859375,
0.233642578125,
0.6015625,
0.1690673828125,
0.59521484375
],
[
0.50244140625,
0.57568359375,
0.8857421875,
0.57568359375,
0.8857421875,
0.5888671875,
0.50244140625,
0.5888671875
],
[
0.50244140625,
0.59130859375,
0.8857421875,
0.59130859375,
0.8857421875,
0.60498046875,
0.50244140625,
0.60498046875
],
[
0.0906982421875,
0.60986328125,
0.4755859375,
0.60791015625,
0.4755859375,
0.62890625,
0.0906982421875,
0.630859375
],
[
0.50341796875,
0.60791015625,
0.88232421875,
0.60791015625,
0.88232421875,
0.61865234375,
0.50341796875,
0.61865234375
],
[
0.0882568359375,
0.625,
0.47216796875,
0.626953125,
0.47216796875,
0.64453125,
0.0882568359375,
0.642578125
],
[
0.5,
0.619140625,
0.8837890625,
0.6220703125,
0.8837890625,
0.63623046875,
0.5,
0.6337890625
],
[
0.50244140625,
0.63818359375,
0.58251953125,
0.63818359375,
0.58251953125,
0.6494140625,
0.50244140625,
0.6494140625
],
[
0.09234619140625,
0.64453125,
0.465576171875,
0.64453125,
0.465576171875,
0.65478515625,
0.09234619140625,
0.65478515625
],
[
0.08905029296875,
0.66357421875,
0.47216796875,
0.6611328125,
0.47216796875,
0.6787109375,
0.08905029296875,
0.68115234375
],
[
0.5009765625,
0.66943359375,
0.8857421875,
0.67041015625,
0.8857421875,
0.68505859375,
0.5009765625,
0.6845703125
],
[
0.08905029296875,
0.6796875,
0.473876953125,
0.6796875,
0.473876953125,
0.69775390625,
0.08905029296875,
0.69775390625
],
[
0.50341796875,
0.6875,
0.88232421875,
0.6875,
0.88232421875,
0.69775390625,
0.50341796875,
0.69775390625
],
[
0.08905029296875,
0.69580078125,
0.47216796875,
0.693359375,
0.47216796875,
0.70849609375,
0.08905029296875,
0.71044921875
],
[
0.50244140625,
0.70068359375,
0.8837890625,
0.70068359375,
0.8837890625,
0.7138671875,
0.50244140625,
0.7138671875
],
[
0.0882568359375,
0.70849609375,
0.4345703125,
0.71044921875,
0.4345703125,
0.72802734375,
0.0882568359375,
0.72607421875
],
[
0.50341796875,
0.71630859375,
0.88232421875,
0.71630859375,
0.88232421875,
0.72998046875,
0.50341796875,
0.72998046875
],
[
0.09234619140625,
0.73486328125,
0.47216796875,
0.73486328125,
0.47216796875,
0.74560546875,
0.09234619140625,
0.74560546875
],
[
0.50244140625,
0.73291015625,
0.88232421875,
0.73291015625,
0.88232421875,
0.74365234375,
0.50244140625,
0.74365234375
],
[
0.08905029296875,
0.74755859375,
0.473876953125,
0.74853515625,
0.473876953125,
0.76220703125,
0.08905029296875,
0.76123046875
],
[
0.5009765625,
0.7470703125,
0.8857421875,
0.7470703125,
0.8857421875,
0.76025390625,
0.5009765625,
0.76025390625
],
[
0.0906982421875,
0.76318359375,
0.47216796875,
0.76318359375,
0.47216796875,
0.7763671875,
0.0906982421875,
0.7763671875
],
[
0.5009765625,
0.76220703125,
0.6650390625,
0.76220703125,
0.6650390625,
0.77587890625,
0.5009765625,
0.77587890625
],
[
0.08905029296875,
0.77783203125,
0.4248046875,
0.77978515625,
0.4248046875,
0.79345703125,
0.08905029296875,
0.79150390625
],
[
0.12744140625,
0.798828125,
0.24267578125,
0.80615234375,
0.238525390625,
0.8388671875,
0.12335205078125,
0.83203125
],
[
0.5009765625,
0.79345703125,
0.75634765625,
0.79345703125,
0.75634765625,
0.806640625,
0.5009765625,
0.806640625
],
[
0.2197265625,
0.80615234375,
0.233642578125,
0.81103515625,
0.2296142578125,
0.81640625,
0.2156982421875,
0.81103515625
],
[
0.2320556640625,
0.81103515625,
0.238525390625,
0.81103515625,
0.238525390625,
0.81494140625,
0.2320556640625,
0.81494140625
],
[
0.236083984375,
0.8125,
0.4345703125,
0.8125,
0.4345703125,
0.82568359375,
0.236083984375,
0.82568359375
],
[
0.50244140625,
0.8173828125,
0.8837890625,
0.8173828125,
0.8837890625,
0.83056640625,
0.50244140625,
0.83056640625
],
[
0.50341796875,
0.83349609375,
0.88232421875,
0.83349609375,
0.88232421875,
0.84326171875,
0.50341796875,
0.84326171875
],
[
0.10455322265625,
0.8466796875,
0.47216796875,
0.8466796875,
0.47216796875,
0.85986328125,
0.10455322265625,
0.85986328125
],
[
0.5009765625,
0.8466796875,
0.8857421875,
0.8466796875,
0.8857421875,
0.85986328125,
0.5009765625,
0.85986328125
],
[
0.0906982421875,
0.8623046875,
0.4755859375,
0.8623046875,
0.4755859375,
0.87548828125,
0.0906982421875,
0.87548828125
],
[
0.5009765625,
0.8623046875,
0.8857421875,
0.8623046875,
0.8857421875,
0.87548828125,
0.5009765625,
0.87548828125
],
[
0.0906982421875,
0.876953125,
0.473876953125,
0.876953125,
0.473876953125,
0.89013671875,
0.0906982421875,
0.89013671875
],
[
0.50244140625,
0.876953125,
0.88232421875,
0.876953125,
0.88232421875,
0.89013671875,
0.50244140625,
0.89013671875
],
[
0.09234619140625,
0.89453125,
0.398681640625,
0.89453125,
0.398681640625,
0.90478515625,
0.09234619140625,
0.90478515625
],
[
0.50244140625,
0.892578125,
0.8837890625,
0.892578125,
0.8837890625,
0.90576171875,
0.50244140625,
0.90576171875
]
] |
[
"对于一个具体的示例，让我们考虑两种模式的情况。在",
"",
"4.Experiments",
"简单代数的几个步骤之后，我们到达",
"4.1.Setup",
"C(M1,M2) = V(M1, M2) - V(M1, 02) - V(01, M2)",
"对于所有型号，：我们使用4层变压器编码器作为骨",
"+ V(01.02)",
"干。注意头数为8，隐藏维度为1024。所有模型都使",
"用Adam优化器进行训练：并在第一入时期进行学习",
"：其中01和02分别表示模态M1.M2的基线值",
"率预热。我们用GloVe（Penningtonetal.2014）编码",
"文本数据，用MFCC（Tiwari,2010）编码音频数据",
"3.3. Fine-grained Perceptual Scores",
"对于图像数据，我们采用补丁嵌入技术，遵循V模",
"在本小节中，我们继续仔细研究感知得分EGatetal.，",
"型（Dosovitskiyetal.，2020）。值得注意的是，我们",
"2021）。首先，我们将显示此分数隐式地衡量数据集D",
"的模型实现了与基于监督的SOTA模型相当的性能",
"中不同模态之间的相关性。然后，我们提出了感知分",
"数的细粒度版本。下面我们只保留关键步骤，把详细的",
"4.2.Datasets",
"扣除留给补充材料",
"现实生活中的应用程序对多模式数据提出了不同的需",
"令（a,y)表示D中的数据样本，其中aE是输人特",
"求。根据特定的任务，有些必须通过组合所有涉及的",
"征，yE[0,...,k-1}是相应的标签。为了明确引用",
"方式来处理，而在其他方式中，其他方式仅提供补充",
"模态，我们将a分解为（aMi,aM2），其中MiCM",
"信息。此外，多模态数据集中的不同模态可能彼此之",
"和M2=M/Mi是模态的子集。回想一下，模型精",
"间只有微弱的相关性，或者它们可能是提供完全相同",
"度定义为Vf(D)=Ep[*=y]，其中I是恒等函数，y*",
"信息的不同表示。为了解决这些复杂的情况，我们故",
"是指输人处的模型预测。简单代数将引导我们到",
"意选择三个数据集，如下所述",
" darMdaM2Pr(y*,aMi, rM2),",
"B-T4SA数据集E(Vadicamoetal.，2017）是从Twit",
"ter收集的，包括发布的文本（文本模态T）和相关的",
"Vt(D) =",
"，其中Pr(y*,aM1,rM2）是D上的概率分布",
"图像（视觉模态V）。每个数据对都有一个情绪标签",
"（正面、负面和中性）。任务是使用文本和图像输入来",
"感知分数中基于置换的步骤将改变数据分布，从而产",
"预测此标签。我们注意到，文本数据通常与他们的情",
"生分布为Pr(y,cMi)Pr(αM2）的新有效数据集。我们",
"绪标签对齐，而大量的图像似乎与标签无关。B-T4SA",
"将此有效数据集称为DP。现在，M2的感知分数是",
"代表了不同模态互补且相互弱关联的情况",
"PM2;f;D =",
"Y",
"(Vf(D) -Vf(DP))",
"Zf",
"CMU-MOSEI此数据集（Zadehetal.，2018）是用于",
"句子级情感分析和情感识别的最大数据集。它包含超",
"1",
"daMidaM2P(rM2)AM1,M2f,",
"=ZfJx",
"过65小时的注释视频数据，包括音频，视频和文本",
"模式。在我们的实验中，我们只使用音频（A）和文本",
",其中AMi:Ma;f=P(y*,aM[rM2)-P(y*,aMi)。显",
"（T）模态。文本模态包括从音频数据生成的脚本。这",
"然，除了对模型f到y*的依赖之外，术语AM1,Ma:f",
"样，CMU-MOSEI数据集代表了不同模态互补且高度",
"相关的情况",
"有效地衡量了数据集D中Mi和M2之间的相关性",
"由于P(aM2）是非负的，因此如果M2部分与aM1",
"SNLISNLI数据集E(Bowmanetal.，2015）是570k",
"来自同一类，那么我们将期望△M1,M2过≥0。我们将",
"人为写的英语句子对的集合。每一对都包含一个前提",
"这种情况称为类内排列。同样，在类外排列中，rM2",
"和一个假设，其关系被标记为蕴涵，矛盾或中立。我",
"来自与aMi不同的类，我们期望△Mi,M2;f≤0",
"们将把前提和假设视为两种不同的模式，即。e。、T",
"班级内和班级外的情况提供了相互冲突的驱动力，这",
"和Th。基本任务是推断前提和假设之间的关系。显",
"些驱动力会相互抵消，并且无法通过原始的感知得分",
"然，这项任务本质上依赖于两种模式，因此代表了不",
"来反映。基于这一观察，我们建议使用课堂内和课堂",
"同模式不可或缺的情况",
"外的感知分数来代替原始的。它们的定义如下：",
"Pu/gf := ",
"4.3. Contribution and Cooperation",
"Y",
"",
"[Vf(M) - Vr([M/M)])/,",
"在本小节中，我们研究了三个多模态数据集上不同模",
"型的SHAPE分数，证明了它们在捕获单个模态和跨",
"其中[M/Mil±运算符通过类内置换（+）或类外",
"模态合作的边际贡献方面的能力。Table1总结了B",
"置换（一）删除M，模态。Zf是通常的归一化因子。",
"T4SA数据集的结果。我们发现T的单峰模型的性能",
"我们这里只考虑模型归一化。最后，为了简单起见，我",
"与多模态模型一样好，而V的单峰模型仅略微抑制了",
"们通常省略下标：D：只要它不起混清。",
"随机猜测。SHAPE分数与这些观察结果一致。对于"
] |
{
"arxiv_id": "2205.00302",
"language": "zh",
"page_idx": 3
}
|
||
1806.10474-zh-0
|
1806.10474-zh-0.jpg
|
# 逼真的音乐生成的挑战： 大规模建模原始音频
Sander Dieleman Aäron van den Oord Karen Simonyan
DeepMind ${\mathrm{London,UK}}$
$\{$ sedielem,avdnoord,simonyan $\}$ @google.com
# Abstract
Realistic music generation is a challenging task. When building generative models of music that are learnt from data, typically high-level representations such as scores or MIDI are used that abstract away the idiosyncrasies of a particular performance. But these nuances are very important for our perception of musicality and realism, so in this work we embark on modelling music in the raw audio domain. It has been shown that autoregressive models excel at generating raw audio waveforms of speech, but when applied to music, we find them biased towards capturing local signal structure at the expense of modelling long-range correlations. This is problematic because music exhibits structure at many different timescales. In this work, we explore autoregressive discrete autoencoders (ADAs) as a means to enable autoregressive models to capture long-range correlations in waveforms. We find that they allow us to unconditionally generate piano music directly in the raw audio domain, which shows stylistic consistency across tens of seconds.
## 1 Introduction
音乐是一种复杂的、高度结构化的顺序数据形式。当呈现为音频信号时，这种结构在不同的时间尺度上表现出来，从毫秒级的波形周期性，一直到一段音乐的音乐形式，通常跨越几分钟。对由此结构产生的序列中的所有时间相关性进行建模是具有挑战性的，因为它们跨越了许多不同的数量级
几十年来，人们对计算音乐生成产生了浓厚的兴趣[11,20]。最近，深度学习和现代生成建模技术已应用于此任务[5,7]。尽管音乐可以表示为波形，但我们可以通过抽象出特定表演的特质来更简洁地表示它。到目前为止，音乐生成中的几乎所有工作都集中在此类符号表示上：乐谱，MIDI $^{1}$序列和其他表示形式，它们在不同程度上消除了音乐的某些方面
符号表示的使用带有局限性：这种表示所抽象出的细微差别通常在音乐上非常重要，并且极大地影响了我们对音乐的享受。例如，音乐家演奏的音符的精确时间，音色和音量与乐谱中的音符并不完全一致。虽然对于某些乐器来说，这些变化可能是象征性的(例如。g。钢琴，在这里我们可以记录每个按键[41]的确切时间和强度),这对于大多数乐器来说通常是非常困难和不切实际的。此外，符号表示通常是针对特定仪器量身定制的，这降低了它们的通用性，并意味着需要大量工作才能将现有的建模技术应用于新仪器
## 1.1 Raw audio signals
为了克服这些限制，我们可以在原始音频域中对音乐进行建模。虽然音频波形的数字表示在某种程度上仍然是有损的，但它们保留了所有与音乐相关的信息。音频波形的模型也更加通用，可以应用于任何一组乐器的录音，或非音乐音频信号，如语音。也就是说，对音乐音频信号进行建模比对符号表示进行建模更具挑战性，因此，该领域受到的关注相对较少
$^{1}$Musical Instrument Digital Interface
Preprint. Work in progress.
|
[
[
0.3544921875,
0.1275634765625,
0.63818359375,
0.1275634765625,
0.63818359375,
0.1458740234375,
0.3544921875,
0.1458740234375
],
[
0.6337890625,
0.1337890625,
0.640625,
0.1337890625,
0.640625,
0.141357421875,
0.6337890625,
0.141357421875
],
[
0.375732421875,
0.1546630859375,
0.625,
0.1546630859375,
0.625,
0.17236328125,
0.375732421875,
0.17236328125
],
[
0.289306640625,
0.2291259765625,
0.57666015625,
0.2291259765625,
0.57666015625,
0.242431640625,
0.289306640625,
0.242431640625
],
[
0.56884765625,
0.2279052734375,
0.71337890625,
0.2291259765625,
0.71337890625,
0.24365234375,
0.56884765625,
0.242431640625
],
[
0.464111328125,
0.242431640625,
0.53515625,
0.242431640625,
0.53515625,
0.25390625,
0.464111328125,
0.25390625
],
[
0.459228515625,
0.255126953125,
0.5390625,
0.255126953125,
0.5390625,
0.266357421875,
0.459228515625,
0.266357421875
],
[
0.32177734375,
0.26513671875,
0.6845703125,
0.266357421875,
0.6845703125,
0.281005859375,
0.32177734375,
0.27978515625
],
[
0.463134765625,
0.31640625,
0.53759765625,
0.31640625,
0.53759765625,
0.330810546875,
0.463134765625,
0.330810546875
],
[
0.233642578125,
0.3466796875,
0.76708984375,
0.3466796875,
0.76708984375,
0.361083984375,
0.233642578125,
0.361083984375
],
[
0.2344970703125,
0.35986328125,
0.7646484375,
0.35986328125,
0.7646484375,
0.375,
0.2344970703125,
0.375
],
[
0.233642578125,
0.373779296875,
0.765625,
0.375,
0.765625,
0.38818359375,
0.233642578125,
0.386962890625
],
[
0.2320556640625,
0.38818359375,
0.7646484375,
0.386962890625,
0.7646484375,
0.401611328125,
0.2320556640625,
0.40283203125
],
[
0.2344970703125,
0.401611328125,
0.765625,
0.401611328125,
0.765625,
0.415283203125,
0.2344970703125,
0.415283203125
],
[
0.2344970703125,
0.416748046875,
0.765625,
0.416748046875,
0.765625,
0.429931640625,
0.2344970703125,
0.429931640625
],
[
0.2320556640625,
0.4287109375,
0.76708984375,
0.429931640625,
0.76708984375,
0.4443359375,
0.2320556640625,
0.443115234375
],
[
0.233642578125,
0.443115234375,
0.76123046875,
0.443115234375,
0.76123046875,
0.45703125,
0.233642578125,
0.45703125
],
[
0.233642578125,
0.45703125,
0.76318359375,
0.45703125,
0.76318359375,
0.47021484375,
0.233642578125,
0.47021484375
],
[
0.233642578125,
0.4716796875,
0.7646484375,
0.4716796875,
0.7646484375,
0.48486328125,
0.233642578125,
0.48486328125
],
[
0.233642578125,
0.48486328125,
0.76123046875,
0.48486328125,
0.76123046875,
0.498779296875,
0.233642578125,
0.498779296875
],
[
0.2320556640625,
0.49755859375,
0.7646484375,
0.498779296875,
0.7646484375,
0.51318359375,
0.2320556640625,
0.51220703125
],
[
0.2344970703125,
0.51220703125,
0.7294921875,
0.51220703125,
0.7294921875,
0.52587890625,
0.2344970703125,
0.52587890625
],
[
0.17236328125,
0.54443359375,
0.31787109375,
0.546875,
0.31787109375,
0.56103515625,
0.17236328125,
0.5595703125
],
[
0.176513671875,
0.5771484375,
0.82275390625,
0.5771484375,
0.82275390625,
0.59033203125,
0.176513671875,
0.59033203125
],
[
0.176513671875,
0.59228515625,
0.8203125,
0.59228515625,
0.8203125,
0.60302734375,
0.176513671875,
0.60302734375
],
[
0.176513671875,
0.60498046875,
0.82275390625,
0.60498046875,
0.82275390625,
0.61865234375,
0.176513671875,
0.61865234375
],
[
0.1748046875,
0.61865234375,
0.324462890625,
0.6162109375,
0.324462890625,
0.6298828125,
0.1748046875,
0.6318359375
],
[
0.176513671875,
0.63720703125,
0.82177734375,
0.63720703125,
0.82177734375,
0.65087890625,
0.176513671875,
0.65087890625
],
[
0.176513671875,
0.6513671875,
0.82275390625,
0.6513671875,
0.82275390625,
0.66650390625,
0.176513671875,
0.66650390625
],
[
0.176513671875,
0.66650390625,
0.82421875,
0.66650390625,
0.82421875,
0.6796875,
0.176513671875,
0.6796875
],
[
0.176513671875,
0.6796875,
0.78173828125,
0.6796875,
0.78173828125,
0.6943359375,
0.176513671875,
0.6943359375
],
[
0.176513671875,
0.70068359375,
0.82177734375,
0.70068359375,
0.82177734375,
0.7138671875,
0.176513671875,
0.7138671875
],
[
0.176513671875,
0.7138671875,
0.82275390625,
0.7138671875,
0.82275390625,
0.72802734375,
0.176513671875,
0.72802734375
],
[
0.176513671875,
0.7265625,
0.82275390625,
0.72802734375,
0.82275390625,
0.7421875,
0.176513671875,
0.7412109375
],
[
0.176513671875,
0.7421875,
0.82177734375,
0.7421875,
0.82177734375,
0.755859375,
0.176513671875,
0.755859375
],
[
0.176513671875,
0.755859375,
0.82275390625,
0.755859375,
0.82275390625,
0.76953125,
0.176513671875,
0.76953125
],
[
0.176513671875,
0.771484375,
0.66259765625,
0.771484375,
0.66259765625,
0.7822265625,
0.176513671875,
0.7822265625
],
[
0.1748046875,
0.79541015625,
0.343017578125,
0.79541015625,
0.343017578125,
0.8095703125,
0.1748046875,
0.8095703125
],
[
0.176513671875,
0.82373046875,
0.82275390625,
0.82373046875,
0.82275390625,
0.8369140625,
0.176513671875,
0.8369140625
],
[
0.176513671875,
0.8369140625,
0.82177734375,
0.8369140625,
0.82177734375,
0.8515625,
0.176513671875,
0.8515625
],
[
0.17724609375,
0.85302734375,
0.82177734375,
0.85302734375,
0.82177734375,
0.86279296875,
0.17724609375,
0.86279296875
],
[
0.17724609375,
0.86572265625,
0.810546875,
0.86572265625,
0.810546875,
0.87890625,
0.17724609375,
0.87890625
],
[
0.19775390625,
0.88818359375,
0.416748046875,
0.88818359375,
0.416748046875,
0.9013671875,
0.19775390625,
0.9013671875
],
[
0.1748046875,
0.9248046875,
0.334228515625,
0.9248046875,
0.334228515625,
0.93798828125,
0.1748046875,
0.93798828125
]
] |
[
"逼真的音乐生成的挑战",
"",
"大规模建模原始音频",
"SanderDielemanAaron van den Oord",
"dKaren Simonyan",
"DeepMind",
"London,UK",
"(sedielem,avdnoord,simonyan} @google.com",
"Abstract",
"Realistic music generation is a challenging task. When building generative mod-",
"els of music that are learnt from data, typically high-level representations such as",
"scores or MIDI are used that abstract away the idiosyncrasies of a particular per",
"formance. But these nuances are very important for our perception of musicality",
"and realism, so in this work we embark on modelling music in the raw audio do-",
"main. It has been shown that autoregressive models excel at generating raw audio",
"waveforms of speech, butwhen applied to music, wefind thembiased towards cap-",
"turing local signal structure at the expense of modelling long-range correlations",
"This is problematic because music exhibits structure at many different timescales",
"In this work, we explore autoregressive discrete autoencoders (ADAs) as a means",
"to enable autoregressive models to capture long-range correlations in waveforms",
"We find that they allow us to unconditionally generate piano music directly in the",
"raw audio domain, which shows stylistic consistency across tens of seconds.",
"1Introduction",
"音乐是一种复杂的、高度结构化的顺序数据形式。当呈现为音频信号时，这种结构在不同的",
"时间尺度上表现出来，从毫秒级的波形周期性，一直到一段音乐的音乐形式，通常跨越儿分",
"钟。对由此结构产生的序列中的所有时间相关性进行建模是具有挑战性的，因为它们跨越",
"了许多不同的数量级",
"几十年来，人们对计算音乐生成产生了浓厚的兴趣[11,20]。最近，深度学习和现代生成建",
"模技术已应用于此任务[5，7]。尽管音乐可以表示为波形，但我们可以通过抽象出特定表演",
"的特质来更简洁地表示它。到目前为止，音乐生成中的几乎所有工作都集中在此类符号表",
"示上：乐谱，MIDII序列和其他表示形式，它们在不同程度上消除了音乐的某些方面",
"符号表示的使用带有局限性：这种表示所抽象出的细微差别通常在音乐上非常重要，并且",
"极大地影响了我们对音乐的享受。例如，音乐家演奏的音符的精确时间，音色和音量与乐谱",
"中的音符并不完全一致。虽然对于某些乐器来说，这些变化可能是象征性的（例如。g。钢",
"琴，在这里我们可以记录每个按键[41]的确切时间和强度），这对于大多数乐器来说通常是",
"非常困难和不切实际的。此外，符号表示通常是针对特定仪器量身定制的，这降低了它们的",
"通用性，并意味着需要大量工作才能将现有的建模技术应用于新仪器",
"1.1Rawaudio signals",
"为了克服这些限制，我们可以在原始音频域中对音乐进行建模。虽然音频波形的数字表示",
"在某种程度上仍然是有损的，但它们保留了所有与音乐相关的信息。音频波形的模型也更",
"加通用，可以应用于任何一组乐器的录音，或非音乐音频信号，如语音。也就是说，对音乐",
"音频信号进行建模比对符号表示进行建模更具挑战性，因此，该领域受到的关注相对较少",
"'Musical Instrument Digital Interface",
"Preprint. Work in progress."
] |
{
"arxiv_id": "1806.10474",
"language": "zh",
"page_idx": 0
}
|
||
1806.10474-zh-1
|
1806.10474-zh-1.jpg
|
构建波形的生成模型，以捕获许多时间尺度的音乐结构，需要很高的表示能力，并有效地分布在各种音乐相关的时间尺度上。以前在原始音频域[10,13,31,43] 中进行音乐建模的工作表明，捕获局部结构(例如音色)是可行的，但是捕获更高级别的结构已被证明是困难的即使对于理论上应该能够这样做的模型，因为它们的接受领域足够大
1.2 Generative models of raw audio signals
能够直接生成音频波形的模型(与之后可以转换为音频的其他表示形式相反，例如频谱图或钢琴卷)只是最近才开始探索。长期以来，由于问题的规模，这被认为是不可行的，因为音频信号通常以 16kHz 或更高的速率采样
最近成功的尝试依赖于自回归 (AR)模型：WaveNet[43], VRNN[10], WaveRNN[23] 和SampleRNN[31]一次预测一个时间步长的数字波形。WaveNet 是具有扩展卷积[47]的卷积神经网络(CNN), WaveRNN 和 VRNN 是递归神经网络(RNN),而 SampleRNN 使用在不同时间尺度上运行的 RNN 层次结构。对于具有$t=1,\ldots,T$的序列$x_t$ ,他们将分布建模为条件的乘积：$p(x_1,x_2,\ldots,x_T)=p(x_1)\cdot p(x_2|x_1)\cdot p(x_3|x_1,x_2)\cdot\ldots=\prod_tp(x_t|x_{<t})$ 。AR 模型可以生成逼真的语音信号，尽管采样成本可能很高(每个时间步长都是按顺序产生的) , 但这些模型现在在实践中用于文本到语音 (TTS)[45]。开始探索的另一种方法是使用生成对抗网络(GAN)[16]产生音频波形[13]
文本到语音模型 [23, 43] 使用强条件信号使生成的波形对应于提供的文本输人，从而使它们在超过几百毫秒的时间尺度上模拟信号结构。生成音乐时无法使用这种条件，因此此任务需要具有更高容量的模型。SampleRNN 和 WaveNet 已应用于音乐生成，但在这两种情况下。样本都不会在几秒钟及以后的时间尺度上表现出有趣的结构。这些时间尺度对于我们对音乐的诠释和享受至关重要
# 1.3 Additional related work
除了原始音频生成之外，还针对图像[2,3,12,27],对话[40],3D形状[30]和符号音乐[37] 提出了几种捕获大规模结构的生成模型。这些模型倾向于使用层次结构作为归纳偏差。使神经网络能够学习已研究数据中的长期依赖关系的其他方法包括替代损失函数[42] 和模型体系结构[1,8,14,19,28,29,32]。这些工作大多集中在递归模型上。
## 1.4 0verview
我们研究如何使用自回归模型有效地对音乐音频信号中的远程结构进行建模。我们表明，可以在大约 400,000 个时间步长或以 16kHz 采样的大约 25 秒的音频中对结构进行建模。这使我们能够生成风格一致的钢琴音乐样本。我们通过将学习问题分层划分为单独的阶段来实现这一目标，每个阶段以不同的比例对信号结构进行建模。这些阶段是单独训练的，减轻了硬件限制。我们的贡献有三个方面：
$\bullet$ We address music generation in the raw audio domain, a task which has received little attention
in literature so far, and establish it as a useful benchmark to determine the ability of a model to
capture long-range structure in data.
$\bullet$ We investigate the capabilities of autoregressive models for this task, and demonstrate a computationally efficient method to enlarge their receptive fields using autoregressive discrete autoen- ${\mathrm{coders~(ADAs)}}.$
$\bullet$ We introduce the argmax autoencoder (AMAE) as an alternative to vector quantisation variational
autoencoders (VQ-VAE) [46] that converges more reliably when trained on a challenging dataset
and compare both models in this context.
2 Scaling up autoregressive models for music
要对音乐音频信号中的远程结构进行建模，我们需要扩大 AR 模型的感受野 (RF)。我们可以通过增加膨胀的卷积层来增加 WaveNet 的 RF。对于 SampleRNN, 这需要添加更多的层。在这两种情况下，所需的模型大小随 RF 长度呈对数增长。这似乎意味着将这些模型扩展到大型 RF 相对容易。但是，这些模型需要在至少与其 RF 一样长的音频摘录上进行训练，否则它们无法在此时间范围内捕获任何结构。这意味着训练的内存需求线性增加，而不是对数增加。因为每秒的音频对应着成千上万的时间步，我们很快就达到了硬件的极限
2
|
[
[
0.178955078125,
0.0947265625,
0.8203125,
0.0947265625,
0.8203125,
0.10479736328125,
0.178955078125,
0.10479736328125
],
[
0.1748046875,
0.106689453125,
0.82275390625,
0.1060791015625,
0.82275390625,
0.11932373046875,
0.1748046875,
0.12060546875
],
[
0.178955078125,
0.12249755859375,
0.82177734375,
0.12249755859375,
0.82177734375,
0.1331787109375,
0.178955078125,
0.1331787109375
],
[
0.1748046875,
0.1331787109375,
0.6650390625,
0.1351318359375,
0.6650390625,
0.1483154296875,
0.1748046875,
0.146484375
],
[
0.1732177734375,
0.160400390625,
0.490234375,
0.1610107421875,
0.490234375,
0.175537109375,
0.1732177734375,
0.1749267578125
],
[
0.17724609375,
0.1900634765625,
0.82177734375,
0.1900634765625,
0.82177734375,
0.203857421875,
0.17724609375,
0.203857421875
],
[
0.176513671875,
0.203857421875,
0.82275390625,
0.203857421875,
0.82275390625,
0.2171630859375,
0.176513671875,
0.2171630859375
],
[
0.176513671875,
0.2171630859375,
0.473876953125,
0.2171630859375,
0.473876953125,
0.2310791015625,
0.176513671875,
0.2310791015625
],
[
0.17724609375,
0.238037109375,
0.82275390625,
0.238037109375,
0.82275390625,
0.251953125,
0.17724609375,
0.251953125
],
[
0.176513671875,
0.251953125,
0.82177734375,
0.251953125,
0.82177734375,
0.26513671875,
0.176513671875,
0.26513671875
],
[
0.17724609375,
0.266357421875,
0.82275390625,
0.266357421875,
0.82275390625,
0.27978515625,
0.17724609375,
0.27978515625
],
[
0.17724609375,
0.27978515625,
0.82177734375,
0.27978515625,
0.82177734375,
0.29345703125,
0.17724609375,
0.29345703125
],
[
0.17724609375,
0.29345703125,
0.82275390625,
0.29345703125,
0.82275390625,
0.306884765625,
0.17724609375,
0.306884765625
],
[
0.176513671875,
0.306884765625,
0.8203125,
0.306884765625,
0.8203125,
0.32080078125,
0.176513671875,
0.32080078125
],
[
0.176513671875,
0.322509765625,
0.82275390625,
0.322509765625,
0.82275390625,
0.3359375,
0.176513671875,
0.3359375
],
[
0.176513671875,
0.337158203125,
0.468017578125,
0.337158203125,
0.468017578125,
0.347900390625,
0.176513671875,
0.347900390625
],
[
0.176513671875,
0.35595703125,
0.82177734375,
0.35595703125,
0.82177734375,
0.369384765625,
0.176513671875,
0.369384765625
],
[
0.176513671875,
0.37060546875,
0.82177734375,
0.37060546875,
0.82177734375,
0.3837890625,
0.176513671875,
0.3837890625
],
[
0.176513671875,
0.3837890625,
0.82177734375,
0.3837890625,
0.82177734375,
0.397705078125,
0.176513671875,
0.397705078125
],
[
0.17724609375,
0.399658203125,
0.8203125,
0.399658203125,
0.8203125,
0.410400390625,
0.17724609375,
0.410400390625
],
[
0.176513671875,
0.412353515625,
0.35546875,
0.412353515625,
0.35546875,
0.425537109375,
0.176513671875,
0.425537109375
],
[
0.1748046875,
0.439453125,
0.384033203125,
0.439453125,
0.384033203125,
0.45263671875,
0.1748046875,
0.45263671875
],
[
0.176513671875,
0.46728515625,
0.82177734375,
0.46728515625,
0.82177734375,
0.48095703125,
0.176513671875,
0.48095703125
],
[
0.176513671875,
0.48095703125,
0.82177734375,
0.48095703125,
0.82177734375,
0.494384765625,
0.176513671875,
0.494384765625
],
[
0.1748046875,
0.49365234375,
0.82275390625,
0.494384765625,
0.82275390625,
0.5087890625,
0.1748046875,
0.50830078125
],
[
0.176513671875,
0.5087890625,
0.66015625,
0.5087890625,
0.66015625,
0.52294921875,
0.176513671875,
0.52294921875
],
[
0.1732177734375,
0.5361328125,
0.283447265625,
0.5361328125,
0.283447265625,
0.5498046875,
0.1732177734375,
0.5498046875
],
[
0.176513671875,
0.564453125,
0.82275390625,
0.564453125,
0.82275390625,
0.57763671875,
0.176513671875,
0.57763671875
],
[
0.176513671875,
0.5771484375,
0.8203125,
0.5771484375,
0.8203125,
0.59033203125,
0.176513671875,
0.59033203125
],
[
0.17724609375,
0.59326171875,
0.8203125,
0.59326171875,
0.8203125,
0.60400390625,
0.17724609375,
0.60400390625
],
[
0.176513671875,
0.60595703125,
0.8203125,
0.60595703125,
0.8203125,
0.619140625,
0.176513671875,
0.619140625
],
[
0.17724609375,
0.6220703125,
0.42724609375,
0.6220703125,
0.42724609375,
0.6318359375,
0.17724609375,
0.6318359375
],
[
0.178955078125,
0.646484375,
0.82177734375,
0.646484375,
0.82177734375,
0.66015625,
0.178955078125,
0.66015625
],
[
0.191162109375,
0.66015625,
0.8203125,
0.66015625,
0.8203125,
0.67041015625,
0.191162109375,
0.67041015625
],
[
0.193603515625,
0.67529296875,
0.42724609375,
0.67529296875,
0.42724609375,
0.68603515625,
0.193603515625,
0.68603515625
],
[
0.176513671875,
0.6923828125,
0.82275390625,
0.693359375,
0.82275390625,
0.70849609375,
0.176513671875,
0.70703125
],
[
0.19287109375,
0.70849609375,
0.826171875,
0.70849609375,
0.826171875,
0.7216796875,
0.19287109375,
0.7216796875
],
[
0.19287109375,
0.7216796875,
0.301513671875,
0.7216796875,
0.301513671875,
0.73486328125,
0.19287109375,
0.73486328125
],
[
0.17724609375,
0.74072265625,
0.82275390625,
0.7412109375,
0.82275390625,
0.7548828125,
0.17724609375,
0.75390625
],
[
0.191162109375,
0.75390625,
0.82275390625,
0.7548828125,
0.82275390625,
0.76953125,
0.191162109375,
0.76806640625
],
[
0.19287109375,
0.76953125,
0.463134765625,
0.76953125,
0.463134765625,
0.78271484375,
0.19287109375,
0.78271484375
],
[
0.1732177734375,
0.7998046875,
0.56884765625,
0.7998046875,
0.56884765625,
0.814453125,
0.1732177734375,
0.814453125
],
[
0.1748046875,
0.830078125,
0.8203125,
0.830078125,
0.8203125,
0.84326171875,
0.1748046875,
0.84326171875
],
[
0.176513671875,
0.84326171875,
0.8203125,
0.84326171875,
0.8203125,
0.85693359375,
0.176513671875,
0.85693359375
],
[
0.176513671875,
0.85791015625,
0.82177734375,
0.85791015625,
0.82177734375,
0.8720703125,
0.176513671875,
0.8720703125
],
[
0.176513671875,
0.869140625,
0.82177734375,
0.87060546875,
0.82177734375,
0.88525390625,
0.176513671875,
0.8837890625
],
[
0.17724609375,
0.88525390625,
0.82275390625,
0.88525390625,
0.82275390625,
0.8984375,
0.17724609375,
0.8984375
],
[
0.1748046875,
0.89794921875,
0.76318359375,
0.8984375,
0.76318359375,
0.912109375,
0.1748046875,
0.9111328125
],
[
0.5,
0.93310546875,
0.51123046875,
0.943359375,
0.5009765625,
0.9521484375,
0.48779296875,
0.9423828125
]
] |
[
"构建波形的生成模型，以捕获许多时间尺度的音乐结构，需要很高的表示能力，并有效地分",
"布在各种音乐相关的时间尺度上。以前在原始音频域[10,13,31，43]中进行音乐建模的工作",
"表明，捕获局部结构（例如音色）是可行的，但是捕获更高级别的结构已被证明是困难的",
"即使对于理论上应该能够这样做的模型，因为它们的接受领域足够大",
"1.2Generative models of raw audio signals",
"能够直接生成音频波形的模型（与之后可以转换为音频的其他表示形式相反，例如频谱图",
"或钢琴卷）只是最近才开始探索。长期以来，由于问题的规模，这被认为是不可行的，因为",
"音频信号通常以16kHz或更高的速率采样",
"最近成功的尝试依赖于自回归（AR）模型：WaveNet[43]】，VRNN[10]，WaveRNN[23]和",
"SampleRNN[31]一次预测一个时间步长的数字波形。WaveNet是具有扩展卷积[47]的卷积",
"神经网络（CNN），WaveRNN和VRNN是递归神经网络（RNN），而SampleRNN使用在不",
"同时间尺度上运行的RNN层次结构。对于具有t=1....,T的序列at，他们将分布建模为",
"条件的乘积：p(1,2,..,a)=pai)p(21)p(31,2).：=Itp(t<t)。AR模",
"型可以生成逼真的语音信号，尽管采样成本可能很高（每个时间步长都是按顺序产生的）",
"但这些模型现在在实践中用于文本到语音（TTS）45】。开始探索的另一种方法是使用生成",
"对抗网络（GAN）[16]产生音频波形「13]",
"文本到语音模型[23，431使用强条件信号使生成的波形对应于提供的文本输入：从而使它们",
"在超过几百毫秒的时间尺度上模拟信号结构。生成音乐时无法使用这种条件，因此此任务",
"需要具有更高容量的模型。SampleRNN和WaveNet已应用于音乐生成，但在这两种情况下",
"样本都不会在儿秒钟及以后的时间尺度上表现出有趣的结构。这些时间尺度对于我们对音",
"乐的诠释和享受至关重要",
"1.3Additional related work",
"除了原始音频生成之外，还针对图像[2,3,12,27]，对话[40]，3D形状[30]和符号音乐[37]",
"提出了几种捕获大规模结构的生成模型。这些模型倾向于使用层次结构作为归纳偏差。使",
"神经网络能够学习已研究数据中的长期依赖关系的其他方法包括替代损失函数[42]和模型",
"体系结构[1,8,14,19,28,29,32]。这些工作大多集中在递归模型上。",
"1.4Overview",
"我们研究如何使用自回归模型有效地对音乐音频信号中的远程结构进行建模。我们表明，可",
"以在大约400,000个时间步长或以16kHz采样的大约25秒的音频中对结构进行建模。这使",
"我们能够生成风格一致的钢琴音乐样本。我们通过将学习问题分层划分为单独的阶段来实",
"现这一目标，每个阶段以不同的比例对信号结构进行建模。这些阶段是单独训练的，减轻了",
"硬件限制。我们的贡献有三个方面：",
". We address music generation in the raw audio domain, a task which has received little attention",
"in literature sofar,and establishitas auseful benchmark to determinethe ability of a model tc",
"capturelong-rangestructureindata",
".We investigate the capabilities of autoregressivemodels for this task,and demonstrate a compu",
"tationally efficient method to enlarge their receptive fields using autoregressive discrete autoen-",
"coders (ADAs).",
".We introduce the argmax autoencoder (AMAE)as an alternative to vector quantisation variational",
"autoencoders (VQ-VAE) [46] that converges more reliably when trained on a challenging dataset,",
"and compare both models in this context.",
"2Scaling up autoregressive models for music",
"要对音乐音频信号中的远程结构进行建模，我们需要扩大AR模型的感受野（RF）。我们可",
"以通过增加膨胀的卷积层来增加WaveNet的RF。对于SampleRNN，这需要添加更多的层",
"在这两种情况下，所需的模型大小随RF长度呈对数增长。这似乎意味着将这些模型扩展到",
"大型RF相对容易。但是，这些模型需要在至少与其RF一样长的音频摘录上进行训练，否",
"则它们无法在此时间范围内捕获任何结构。这意味着训练的内存需求线性增加，而不是对",
"数增加。因为每秒的音频对应着成千上万的时间步，我们很快就达到了硬件的极限",
"2"
] |
{
"arxiv_id": "1806.10474",
"language": "zh",
"page_idx": 1
}
|
||
1806.10474-zh-2
|
1806.10474-zh-2.jpg
|
<FigureHere>
Figure 1: Schematic overview of an autoregressive discrete autoencoder. The encoder, modulator
and local model are neural networks.
此外，这些模型强烈偏向于对局部结构进行建模。这可以更容易地在图像模型中看到，其中e 的 RF。g。在对象图像上训练的 PixelCNN [44] 可以很容易地多次包含整个图像，但它可能仍然无法很好地对数据中的大规模结构进行建模，使样本看起来像可识别的对象。由于音频信号往往以低频分量为主，因此这是一个合理的偏差：要预测给定的时间步长，信号的最近过去将比更早之前发生的情况提供更多信息。然而，这只是在某种程度上是正确的，我们需要重新分配一些模型容量来捕获长期相关性。正如我们将看到的，这种权衡表现为信号保真度的降低，但音乐性的提高
2.1 Stacking autoregressive models
使 AR 模型产生具有远程结构的样本的一种方法是提供丰富的调节信号。这个概念构成了我们方法的基础：我们通过附加编码器将 AR 模型转换为自动编码器，以直接从数据中学习高级调节信号。我们可以将时间下采样操作插人编码器，使该信号比原始波形更粗粒度。然后，生成的自动编码器使用其 AR 解码器对该压缩信号无法捕获的任何局部结构进行建模。我们将调节信号的采样率与输人之间的比率称为跳数(h)
因为调节信号又是一个序列，所以我们也可以用 AR 模型对此进行建模。它的采样率是$h$ 倍，因此在此表示上使用$r$时间步长的 RF 训练模型会在音频域中产生相应的$h\cdot r\textbf{ RF。这 }$ 个两步训练过程使我们能够使用比以前大$h$倍的 RF 构建模型。当然，$h$的值不能任意选择： 较大的跳大小意味着更大的压缩和更多的信息丢失
由于我们使用的 AR 模型是概率性的，因此可以考虑将编码器拟合到此框架中，并将学习的条件序列解释为一系列概率性潜在变量。然后，我们可以将此模型视为变分自编码器(VAE) [25,36]。不幸的是，具有强大解码器的 VAE 会遭受后验崩溃[6,9,18,46]的困扰：它们通常会完全忽略使用潜在变量，因为损失函数中的正则化项明确地阻止了这一点。相反，我们选择从我们的模型中删除与潜在变量相关的任何惩罚项，并在此过程中使其编码器具有确定性。或者，我们可以限制解码器[9,18]的 RF,但我们发现这会导致非常差的音频保真度。其他可能性包括使用关联压缩网络[17]或“自由位”方法[26]
3 Autoregressive discrete autoencoders
在典型的确定性自动编码器中，学习表示的信息内容仅受编码器和解码器的容量限制，因为该表示是实值的(如果它们非常非线性，则可以将大量信息压缩为甚至单个标量值)[15] 。相反，我们使该表示离散，以便我们可以直接控制其信息内容。这样做的另一个好处是可以更轻松地训练另一个 AR 模型：迄今为止，最成功的图像和音频 AR 模型也将它们视为离散对象[39,43,44]。此外，我们在初步实验中发现，带有 AR 解码器的确定性连续自动编码器可能无法学习正确使用自回归连接，因为更容易通过编码器传递信息(这与 VAE 中的后验崩溃相反)
图 1 显示了用于音频波形的自回归离散自动编码器 (ADA) 的示意图。它具有一个编码器该编码器计算输人$x$的更高级表示，以及一个解码器，该解码器尝试在给定此表示的情况下重建原始波形。此外，它还具有一个量化模块，该模块采用连续编码器输出，我们将其称为查询向量($q$),并对其进行量化。解码器接收量化查询向量($q^\prime$)作为输人，从而创建离散瓶颈。通过使用量化质心的索引，量化查询也可以表示为整数序列。这是代码序列($y$
3
|
[
[
0.57275390625,
0.09027099609375,
0.619140625,
0.09027099609375,
0.619140625,
0.10101318359375,
0.57275390625,
0.10101318359375
],
[
0.39794921875,
0.10101318359375,
0.4345703125,
0.10101318359375,
0.4345703125,
0.111083984375,
0.39794921875,
0.111083984375
],
[
0.371826171875,
0.11614990234375,
0.4599609375,
0.11614990234375,
0.4599609375,
0.1275634765625,
0.371826171875,
0.1275634765625
],
[
0.51708984375,
0.12060546875,
0.57666015625,
0.12060546875,
0.57666015625,
0.1307373046875,
0.51708984375,
0.1307373046875
],
[
0.609375,
0.12060546875,
0.67626953125,
0.12060546875,
0.67626953125,
0.1307373046875,
0.609375,
0.1307373046875
],
[
0.48779296875,
0.1558837890625,
0.50634765625,
0.1558837890625,
0.50634765625,
0.1590576171875,
0.48779296875,
0.1590576171875
],
[
0.3056640625,
0.169189453125,
0.374267578125,
0.169189453125,
0.374267578125,
0.179931640625,
0.3056640625,
0.179931640625
],
[
0.443603515625,
0.1634521484375,
0.5146484375,
0.1634521484375,
0.5146484375,
0.173583984375,
0.443603515625,
0.173583984375
],
[
0.418212890625,
0.184326171875,
0.483642578125,
0.186279296875,
0.483642578125,
0.1976318359375,
0.418212890625,
0.1956787109375
],
[
0.328369140625,
0.20703125,
0.373291015625,
0.20703125,
0.373291015625,
0.2183837890625,
0.328369140625,
0.2183837890625
],
[
0.60693359375,
0.2108154296875,
0.67626953125,
0.2108154296875,
0.67626953125,
0.2215576171875,
0.60693359375,
0.2215576171875
],
[
0.3798828125,
0.2412109375,
0.455078125,
0.2412109375,
0.455078125,
0.25244140625,
0.3798828125,
0.25244140625
],
[
0.176513671875,
0.270751953125,
0.82421875,
0.270751953125,
0.82421875,
0.2841796875,
0.176513671875,
0.2841796875
],
[
0.1748046875,
0.2841796875,
0.414306640625,
0.285400390625,
0.414306640625,
0.296630859375,
0.1748046875,
0.29541015625
],
[
0.176513671875,
0.31640625,
0.82275390625,
0.31640625,
0.82275390625,
0.330078125,
0.176513671875,
0.330078125
],
[
0.176513671875,
0.330078125,
0.82177734375,
0.330078125,
0.82177734375,
0.3447265625,
0.176513671875,
0.3447265625
],
[
0.1748046875,
0.34228515625,
0.8203125,
0.34228515625,
0.8203125,
0.35595703125,
0.1748046875,
0.35595703125
],
[
0.17724609375,
0.35791015625,
0.82177734375,
0.35791015625,
0.82177734375,
0.371826171875,
0.17724609375,
0.371826171875
],
[
0.17724609375,
0.371826171875,
0.82275390625,
0.371826171875,
0.82275390625,
0.385009765625,
0.17724609375,
0.385009765625
],
[
0.17724609375,
0.385009765625,
0.82275390625,
0.385009765625,
0.82275390625,
0.3984375,
0.17724609375,
0.3984375
],
[
0.176513671875,
0.399658203125,
0.420654296875,
0.399658203125,
0.420654296875,
0.41357421875,
0.176513671875,
0.41357421875
],
[
0.1732177734375,
0.4267578125,
0.4345703125,
0.427978515625,
0.4345703125,
0.442626953125,
0.1732177734375,
0.44140625
],
[
0.176513671875,
0.455078125,
0.82275390625,
0.45703125,
0.82275390625,
0.47021484375,
0.176513671875,
0.468505859375
],
[
0.17724609375,
0.47021484375,
0.82177734375,
0.47021484375,
0.82177734375,
0.484130859375,
0.17724609375,
0.484130859375
],
[
0.1748046875,
0.484130859375,
0.82275390625,
0.484130859375,
0.82275390625,
0.49755859375,
0.1748046875,
0.49755859375
],
[
0.1748046875,
0.49755859375,
0.82275390625,
0.496826171875,
0.82275390625,
0.51123046875,
0.1748046875,
0.51220703125
],
[
0.1748046875,
0.51123046875,
0.5908203125,
0.51220703125,
0.5908203125,
0.52587890625,
0.1748046875,
0.5244140625
],
[
0.17724609375,
0.5341796875,
0.82275390625,
0.5341796875,
0.82275390625,
0.54443359375,
0.17724609375,
0.54443359375
],
[
0.1748046875,
0.54541015625,
0.82275390625,
0.546875,
0.82275390625,
0.56005859375,
0.1748046875,
0.5595703125
],
[
0.176513671875,
0.56005859375,
0.82275390625,
0.56005859375,
0.82275390625,
0.57373046875,
0.176513671875,
0.57373046875
],
[
0.17724609375,
0.57568359375,
0.53369140625,
0.57568359375,
0.53369140625,
0.5859375,
0.17724609375,
0.5859375
],
[
0.178955078125,
0.5966796875,
0.8203125,
0.5966796875,
0.8203125,
0.60693359375,
0.178955078125,
0.60693359375
],
[
0.17724609375,
0.60791015625,
0.82275390625,
0.60791015625,
0.82275390625,
0.6220703125,
0.17724609375,
0.6220703125
],
[
0.17724609375,
0.6220703125,
0.82177734375,
0.6220703125,
0.82177734375,
0.63525390625,
0.17724609375,
0.63525390625
],
[
0.17724609375,
0.63623046875,
0.82177734375,
0.63623046875,
0.82177734375,
0.6494140625,
0.17724609375,
0.6494140625
],
[
0.17724609375,
0.6494140625,
0.82177734375,
0.6494140625,
0.82177734375,
0.66357421875,
0.17724609375,
0.66357421875
],
[
0.17724609375,
0.66357421875,
0.82275390625,
0.66357421875,
0.82275390625,
0.6767578125,
0.17724609375,
0.6767578125
],
[
0.176513671875,
0.67822265625,
0.61669921875,
0.67822265625,
0.61669921875,
0.69140625,
0.176513671875,
0.69140625
],
[
0.17236328125,
0.70849609375,
0.5185546875,
0.708984375,
0.5185546875,
0.7236328125,
0.17236328125,
0.72265625
],
[
0.17724609375,
0.7412109375,
0.8203125,
0.7412109375,
0.8203125,
0.751953125,
0.17724609375,
0.751953125
],
[
0.17724609375,
0.75390625,
0.82177734375,
0.75390625,
0.82177734375,
0.767578125,
0.17724609375,
0.767578125
],
[
0.178955078125,
0.767578125,
0.82177734375,
0.767578125,
0.82177734375,
0.78076171875,
0.178955078125,
0.78076171875
],
[
0.176513671875,
0.77978515625,
0.8203125,
0.77978515625,
0.8203125,
0.79345703125,
0.176513671875,
0.79345703125
],
[
0.176513671875,
0.79345703125,
0.82177734375,
0.79541015625,
0.82177734375,
0.8095703125,
0.176513671875,
0.806640625
],
[
0.176513671875,
0.8095703125,
0.82177734375,
0.8095703125,
0.82177734375,
0.82275390625,
0.176513671875,
0.82275390625
],
[
0.17724609375,
0.82373046875,
0.26708984375,
0.82373046875,
0.26708984375,
0.83544921875,
0.17724609375,
0.83544921875
],
[
0.176513671875,
0.84326171875,
0.8203125,
0.84326171875,
0.8203125,
0.85693359375,
0.176513671875,
0.85693359375
],
[
0.17724609375,
0.85791015625,
0.82177734375,
0.85791015625,
0.82177734375,
0.8720703125,
0.17724609375,
0.8720703125
],
[
0.178955078125,
0.87255859375,
0.8203125,
0.87255859375,
0.8203125,
0.88330078125,
0.178955078125,
0.88330078125
],
[
0.176513671875,
0.8837890625,
0.82275390625,
0.8837890625,
0.82275390625,
0.8984375,
0.176513671875,
0.8984375
],
[
0.1748046875,
0.89794921875,
0.82275390625,
0.8984375,
0.82275390625,
0.91357421875,
0.1748046875,
0.912109375
],
[
0.49267578125,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94775390625,
0.49267578125,
0.94775390625
]
] |
[
"decoder",
"y code",
"..5207 13118168..",
"modulator",
"localmodel",
"SU",
"quantisation",
"quantised query",
"q-f(x) query",
"encoder",
"ldiieio",
"",
"Figure l: Schematic overview of an autoregressive discrete autoencoder. The encoder, modulator",
"and localmodelareneuralnetworks",
"此外，这些模型强烈偏向于对局部结构进行建模。这可以更容易地在图像模型中看到，其中",
"e的RF。g。在对象图像上训练的PixelCNN[44]可以很容易地多次包含整个图像，但它可",
"能仍然无法很好地对数据中的大规模结构进行建模，使样本看起来像可识别的对象。由于",
"音频信号往往以低频分量为主，因此这是一个合理的偏差：要预测给定的时间步长，信号的",
"最近过去将比更早之前发生的情况提供更多信息。然而，这只是在某种程度上是正确的，我",
"们需要重新分配一些模型容量来捕获长期相关性。正如我们将看到的，这种权衡表现为信",
"号保真度的降低，但音乐性的提高",
"2.1Stacking autoregressive models",
"使AR模型产生具有远程结构的样本的一种方法是提供丰富的调节信号。这个概念构成了我",
"们方法的基础：我们通过附加编码器将AR模型转换为自动编码器，以直接从数据中学习高",
"级调节信号。我们可以将时间下采样操作插入编码器，使该信号比原始波形更粗粒度。然",
"后，生成的自动编码器使用其AR解码器对该压缩信号无法捕获的任何局部结构进行建模。",
"我们将调节信号的采样率与输人之间的比率称为跳数（h）",
"因为调节信号又是一个序列：所以我们也可以用AR模型对此进行建模。它的采样率是h",
"倍，因此在此表示上使用r时间步长的RF训练模型会在音频域中产生相应的h·rRF。这",
"个两步训练过程使我们能够使用比以前大h倍的RF构建模型。当然，h的值不能任意选择",
"较大的跳大小意味着更大的压缩和更多的信息丢失",
"由于我们使用的AR模型是概率性的：因此可以考虑将编码器拟合到此框架中，并将学寸的",
"条件序列解释为一系列概率性潜在变量。然后，我们可以将此模型视为变分自编码器（VAE",
"[25,36]。不幸的是，具有强大解码器的VAE会遭受后验崩溃[6,9.18,46]的困扰：它们通",
"常会完全忽略使用潜在变量，因为损失函数中的正则化项明确地阻止了这一点。相反，我们",
"选择从我们的模型中删除与潜在变量相关的任何惩罚项，并在此过程中使其编码器具有确",
"定性。或者，我们可以限制解码器[9,18]的RF，但我们发现这会导致非常差的音频保真度。",
"其他可能性包括使用关联压缩网络[17]或“自由位”方法[26]",
"3Autoregressive discrete autoencoders",
"在典型的确定性自动编码器中，学习表示的信息内容仅受编码器和解码器的容量限制，因",
"为该表示是实值的（如果它们非常非线性，则可以将大量信息压缩为甚至单个标量值）【15]",
"。相反，我们使该表示离散，以便我们可以直接控制其信息内容。这样做的另一个好处是可",
"以更轻松地训练另一个AR模型：迄今为止，最成功的图像和音频AR模型也将它们视为离",
"散对象[39,43，44]。此外，我们在初步实验中发现，带有AR解码器的确定性连续自动编码",
"器可能无法学习正确使用自回归连接，因为更容易通过编码器传递信息（这与VAE中的后",
"验崩溃相反）",
"图1显示了用于音频波形的自回归离散自动编码器（ADA）的示意图。它具有一个编码器",
"该编码器计算输人的更高级表示，以及一个解码器，该解码器尝试在给定此表示的情况",
"下重建原始波形。此外，它还具有一个量化模块，该模块采用连续编码器输出，我们将其称",
"为查询向量（q），并对其进行量化。解码器接收量化查询向量（q'）作为输人，从而创建",
"离散瓶颈。通过使用量化质心的索引，量化查询也可以表示为整数序列。这是代码序列（y",
"3"
] |
{
"arxiv_id": "1806.10474",
"language": "zh",
"page_idx": 2
}
|
||
1806.10474-zh-3
|
1806.10474-zh-3.jpg
|
)。解码器由自回归局部模型和调制器组成，调制器使用量化查询来引导局部模型忠实地再现原始波形。后者类似于 WaveNet 中用于文本到语音[43] 的条件堆栈。现在，我们将讨论ADA 的两个实例化
# 3.1 vQ-VAE
矢量量化变分自动编码器[46]使用矢量量化(VQ): 查询是$d$维空间中的矢量，并且$k$这
样的矢量的码本是即时学习的，以及其余的模型参数。损失函数采用以下形式：
$$\mathcal{L}_{VQ-VAE}=-\log p(\boldsymbol{x}|\boldsymbol{q'})+(\boldsymbol{q'}-[\boldsymbol{q}])^{2}+\beta\cdot([\boldsymbol{q'}]-\boldsymbol{q})^{2}.$$
(1)
方括号表示所包含的表达式被视为常量 w。r。t。分化 2。这三个术语分别是给定量化查询$q^{\prime}$的$x$的负对数似然 (NLL), 码本丢失和承诺丢失。代替通过梯度下降来最小化组合损失还可以使用 K-means 算法的指数平滑版本来最小化码本损失。这往往会加快收敛速度，所以我们在这里采用这种做法。我们将码本向量的指数移动平均更新系数表示为 α。尽管它的名字(包括“变异”),没有成本与在 VQ-VAE 中使用编码器途径相关
我们已经观察到，VQ-VAE 接受了具有挑战性的训练(i。e。高熵)数据集经常遭受码本崩溃： 在训练过程中的某个时刻，码本的某些部分可能会停止使用，模型将不再使用离散瓶颈的全部容量，从而导致更糟糕的可能性和较差的重建。这种现象的原因尚不清楚，但请注意K-means 和高斯混合模型训练算法可能存在类似的问题。我们发现，我们可以通过使用基于人群的训练 (PBT)[22]在训练期间在线调整超参数$\alpha$和$\beta$来在一定程度上减轻这种情况(请参阅4节)
# 3.2 AMAE
由于 PBT 在计算上很昂贵，因此我们还尝试通过引人不涉及学习码本的替代量化策略来解决码本崩溃问题。我们将此模型命名为 argmax 自动编码器 ( AMAE)。它的编码器产生$k$维查询，并具有非线性特性，可确保所有输出都在$(k-1)$ -单纯形上。然后，量化操作仅是$\operatorname{argmax}$操作，这等效于在欧几里得意义上取最近的 $k$ 维一热向量
到单纯形上的投影限制了最大量化误差，这使得通过它的梯度(使用直通估计[4])更加准确。为了确保使用全部容量，我们必须添加一个额外的多样性损失项，以鼓励模型以相等的度量使用所有输出。可以使用批处理统计信息来计算此损失，方法是对批处理和时间轴上的所有查询$q$ (量化之前)进行平均，并鼓励生成的向量$\bar{q}$类似于均匀分布
将神经网络的输出限制为单纯形的一种可能方法是使用 softmax 非线性。这可以与使每个批次的平均分布熵最大化的损失配对$:\mathcal{L}_{diversity}=-H(\bar{\boldsymbol{q}})=\sum_i\bar{\boldsymbol{q}}_i\log\bar{\boldsymbol{q}}_i$ 。但是，我们发现使用$\mathrm{ReLU~非线性~[33]~,~然后进行除法归一化~}^{3}$ ,再加上$L_2\textit{多 样 性 损 失 }\mathcal{L} _diversity= \sum ( k\cdot \bar{\boldsymbol{q}} - 1) ^2$ ,往往会更稳健地收敛。我们认为这是因为它使模型能够输出精确的零值，并且独热向量大多为零。那么，全部 AMAE 损失为：
$$\mathcal{L}_{AMAE}=-\log p(\boldsymbol{x}|\boldsymbol{q'})+\nu\cdot\mathcal{L}_{diversity}.$$
(2)
我们还尝试采用来自 VQ-VAE 的承诺损失项，以进一步提高直通估计器的准确性，但发现它在实践中没有明显的区别。正如我们将展示的那样，AMAE 通常略逊于具有相同架构的VQ-VAE,但在 VQ-VAE 遭受码本崩溃的情况下，它更可靠地收敛
# 3.3 Architecture
对于 ADA 的三个子网 (请参见图 1),我们采用 WaveNet[43]架构，因为它允许我们精确地指定它们的 RF。编码器必须以较低的采样率生成查询序列，因此它必须包含下采样操作。计算效率最高的方法是在编码器的前几层中降低此速率。然而，我们发现在输出端执行均值池化是有帮助的，以鼓励编码器学习对时间偏移更不变的内部表示。这使得对输入中存在的本地噪声进行编码变得更加困难；虽然无条件 AR 模型将简单地忽略任何噪声，因为它是不可预测的，但 ADA 可能会尝试通过将此信息合并到代码序列中来“复制”噪声
$^2[x]{\mathrm{~is~implemented~as~tf.stop\_gradient(x)~in~TensorFlow.}}$
$^{3}f(x_{i})=\frac{\mathrm{ReLU}(x_{i})}{\sum_{j}\mathrm{ReLU}(x_{j})+\epsilon}$
4
|
[
[
0.178955078125,
0.09661865234375,
0.1968994140625,
0.09661865234375,
0.1968994140625,
0.10418701171875,
0.178955078125,
0.10418701171875
],
[
0.1968994140625,
0.0947265625,
0.82177734375,
0.0947265625,
0.82177734375,
0.10479736328125,
0.1968994140625,
0.10479736328125
],
[
0.176513671875,
0.106689453125,
0.82275390625,
0.106689453125,
0.82275390625,
0.121826171875,
0.176513671875,
0.121826171875
],
[
0.1748046875,
0.11932373046875,
0.313720703125,
0.12060546875,
0.313720703125,
0.1351318359375,
0.1748046875,
0.1337890625
],
[
0.176513671875,
0.1483154296875,
0.27783203125,
0.1483154296875,
0.27783203125,
0.160400390625,
0.176513671875,
0.160400390625
],
[
0.17724609375,
0.177978515625,
0.82177734375,
0.177978515625,
0.82177734375,
0.1881103515625,
0.17724609375,
0.1881103515625
],
[
0.176513671875,
0.1900634765625,
0.73779296875,
0.1900634765625,
0.73779296875,
0.203857421875,
0.176513671875,
0.203857421875
],
[
0.2998046875,
0.2266845703125,
0.701171875,
0.2247314453125,
0.701171875,
0.2412109375,
0.2998046875,
0.24365234375
],
[
0.80224609375,
0.2279052734375,
0.826171875,
0.2279052734375,
0.826171875,
0.240478515625,
0.80224609375,
0.240478515625
],
[
0.17724609375,
0.2568359375,
0.82275390625,
0.2568359375,
0.82275390625,
0.270751953125,
0.17724609375,
0.270751953125
],
[
0.176513671875,
0.270751953125,
0.82177734375,
0.270751953125,
0.82177734375,
0.2841796875,
0.176513671875,
0.2841796875
],
[
0.1748046875,
0.28271484375,
0.82177734375,
0.2841796875,
0.82177734375,
0.29736328125,
0.1748046875,
0.296630859375
],
[
0.17724609375,
0.298583984375,
0.82177734375,
0.298583984375,
0.82177734375,
0.3125,
0.17724609375,
0.3125
],
[
0.17724609375,
0.314453125,
0.6845703125,
0.314453125,
0.6845703125,
0.324462890625,
0.17724609375,
0.324462890625
],
[
0.176513671875,
0.333251953125,
0.82421875,
0.333251953125,
0.82421875,
0.3466796875,
0.176513671875,
0.3466796875
],
[
0.17724609375,
0.347900390625,
0.8203125,
0.347900390625,
0.8203125,
0.35791015625,
0.17724609375,
0.35791015625
],
[
0.17724609375,
0.35986328125,
0.82177734375,
0.35986328125,
0.82177734375,
0.373779296875,
0.17724609375,
0.373779296875
],
[
0.1748046875,
0.375,
0.82177734375,
0.375,
0.82177734375,
0.38818359375,
0.1748046875,
0.38818359375
],
[
0.178955078125,
0.38818359375,
0.82275390625,
0.38818359375,
0.82275390625,
0.401611328125,
0.178955078125,
0.401611328125
],
[
0.176513671875,
0.40283203125,
0.26953125,
0.40283203125,
0.26953125,
0.4140625,
0.176513671875,
0.4140625
],
[
0.1748046875,
0.427978515625,
0.26708984375,
0.427978515625,
0.26708984375,
0.442626953125,
0.1748046875,
0.442626953125
],
[
0.176513671875,
0.458251953125,
0.82275390625,
0.458251953125,
0.82275390625,
0.4716796875,
0.176513671875,
0.4716796875
],
[
0.176513671875,
0.4716796875,
0.82275390625,
0.4716796875,
0.82275390625,
0.48486328125,
0.176513671875,
0.48486328125
],
[
0.17724609375,
0.484130859375,
0.82275390625,
0.484130859375,
0.82275390625,
0.498779296875,
0.17724609375,
0.498779296875
],
[
0.176513671875,
0.498779296875,
0.64208984375,
0.498779296875,
0.64208984375,
0.51220703125,
0.176513671875,
0.51220703125
],
[
0.17724609375,
0.52099609375,
0.82177734375,
0.52099609375,
0.82177734375,
0.53076171875,
0.17724609375,
0.53076171875
],
[
0.176513671875,
0.53271484375,
0.82177734375,
0.53271484375,
0.82177734375,
0.546875,
0.176513671875,
0.546875
],
[
0.17724609375,
0.546875,
0.82421875,
0.546875,
0.82421875,
0.56005859375,
0.17724609375,
0.56005859375
],
[
0.176513671875,
0.56103515625,
0.71728515625,
0.56103515625,
0.71728515625,
0.57470703125,
0.176513671875,
0.57470703125
],
[
0.176513671875,
0.58056640625,
0.82275390625,
0.58056640625,
0.82275390625,
0.5947265625,
0.176513671875,
0.5947265625
],
[
0.176513671875,
0.59521484375,
0.82177734375,
0.59521484375,
0.82177734375,
0.60986328125,
0.176513671875,
0.60986328125
],
[
0.17236328125,
0.60791015625,
0.826171875,
0.60791015625,
0.826171875,
0.62890625,
0.17236328125,
0.62890625
],
[
0.188720703125,
0.62353515625,
0.82275390625,
0.62353515625,
0.82275390625,
0.63720703125,
0.188720703125,
0.63720703125
],
[
0.17724609375,
0.63818359375,
0.4345703125,
0.63818359375,
0.4345703125,
0.6513671875,
0.17724609375,
0.6513671875
],
[
0.358642578125,
0.673828125,
0.63623046875,
0.6748046875,
0.63623046875,
0.6923828125,
0.358642578125,
0.69140625
],
[
0.8037109375,
0.67529296875,
0.82421875,
0.67529296875,
0.82421875,
0.689453125,
0.8037109375,
0.689453125
],
[
0.176513671875,
0.7041015625,
0.82275390625,
0.7041015625,
0.82275390625,
0.71728515625,
0.176513671875,
0.71728515625
],
[
0.178955078125,
0.71826171875,
0.8203125,
0.71826171875,
0.8203125,
0.72900390625,
0.178955078125,
0.72900390625
],
[
0.176513671875,
0.73095703125,
0.646484375,
0.73095703125,
0.646484375,
0.744140625,
0.176513671875,
0.744140625
],
[
0.1748046875,
0.75830078125,
0.3056640625,
0.75830078125,
0.3056640625,
0.771484375,
0.1748046875,
0.771484375
],
[
0.176513671875,
0.787109375,
0.82177734375,
0.787109375,
0.82177734375,
0.80029296875,
0.176513671875,
0.80029296875
],
[
0.1748046875,
0.798828125,
0.82177734375,
0.80029296875,
0.82177734375,
0.814453125,
0.1748046875,
0.8125
],
[
0.178955078125,
0.81494140625,
0.8203125,
0.81494140625,
0.8203125,
0.82568359375,
0.178955078125,
0.82568359375
],
[
0.176513671875,
0.82763671875,
0.82177734375,
0.82763671875,
0.82177734375,
0.84130859375,
0.176513671875,
0.84130859375
],
[
0.176513671875,
0.84130859375,
0.82177734375,
0.84228515625,
0.82177734375,
0.85595703125,
0.176513671875,
0.85498046875
],
[
0.178955078125,
0.85693359375,
0.771484375,
0.85693359375,
0.771484375,
0.8671875,
0.178955078125,
0.8671875
],
[
0.193603515625,
0.880859375,
0.56787109375,
0.88330078125,
0.56787109375,
0.89794921875,
0.193603515625,
0.89599609375
],
[
0.1953125,
0.89599609375,
0.346435546875,
0.9013671875,
0.3447265625,
0.91943359375,
0.193603515625,
0.91357421875
],
[
0.490234375,
0.93603515625,
0.50732421875,
0.93603515625,
0.50732421875,
0.94970703125,
0.490234375,
0.94970703125
]
] |
[
"0。",
"解码器由自回归局部模型和调制器组成，调制器使用量化查询来引导局高部模型忠实地再",
"现原始波形。后者类似于WaveNet中用于文本到语音[43]的条件堆栈。现在，我们将讨论",
"ADA的两个实例化",
"3.1VQ-VAE",
"天量量化变分自动编码器「461使用矢量量化（VQ）：查询是d维空间中的天量，并且k这",
"样的矢量的码本是即时学习的，以及其余的模型参数。损失函数采用以下形式：",
"CvQ-VAE = -log p(a|q') + (q' - [gl)2 + β . ([q'] - q)2.",
"(1)",
"方括号表示所包含的表达式被视为常量W。I。t。分化2。这三个术语分别是给定量化查询",
"q'的&的负对数似然（NLL），码本丢失和承诺丢失。代替通过梯度下降来最小化组合损失",
"还可以使用K-means算法的指数平滑版本来最小化码本损失。这往往会加快收敛速度，所",
"以我们在这里采用这种做法。我们将码本向量的指数移动平均更新系数表示为α。尽管它",
"的名（包括“变异）：没有成本与在VQ-VAE中使用编码器途径相关",
"我们已经观察到，VQ-VAE接受了具有挑战性的训练（i。e。高）数据集经常遭受码本崩溃：",
"在训练过程中的某个时刻，码本的某些部分可能会停止使用，模型将不再使用离散瓶颈的",
"全部容量，从而导致更糟糕的可能性和较差的重建。这种现象的原因尚不清楚，但请注意",
"K-means和高斯混合模型训练算法可能存在类似的问题。我们发现，我们可以通过使用基于",
"人群的训练（PBT）[22]在训练期间在线调整超参数α和β来在一定程度上减轻这种情况",
"（请参阅4节）",
"3.2AMAE",
"由于PBT在计算上很昂贵，因此我们还尝试通过引入不涉及学习码本的替代量化策略来解",
"决码本崩溃问题。我们将此模型命名为argmax自动编码器（AMAE）。它的编码器产生k维",
"查询，并具有非线性特性，可确保所有输出都在（k一1）-单纯形上。然后，量化操作仅是",
"argmax操作，这等效于在欧几里得意义上取最近的k维一热向量",
"到单纯形上的投影限制了最大量化误差，这使得通过它的梯度（使用直通估计4）更加准",
"确。为了确保使用全部容量，我们必须添加一个额外的多样性损失项，以鼓励模型以相等的",
"度量使用所有输出。可以使用批处理统计信息来计算此损失，方法是对批处理和时间轴上",
"的所有查询q（量化之前）进行平均，并鼓励生成的向量类似于均匀分布",
"将神经网络的输出限制为单纯形的一种可能方法是使用softmax非线性。这可以与使每个批",
"次的平均分布炳最大化的损失配对：Ldiversity=-H(q)=,q,logqi。但是，我们发现使用",
"ReLU非线性[33]，然后进行除法归一化，再加上L2多样性损失Caiuersity=2(k:-1)?",
"往往会更稳健地收敛。我们认为这是因为它使模型能够输出精确的零值，并且独热向量大",
"多为零。那么，全部AMAE损失为：",
"LAMAE=-logp(alq)+V.Ldiuersity",
"(2)",
"我们还尝试采用来自VQ-VAE的承诺损失项，以进一步提高直通估计器的准确性，但发现",
"它在实践中没有明显的区别。正如我们将展示的那样，AMAE通常略逊于具有相同架构的",
"VQ-VAE，但在VQ-VAE遭受码本崩溃的情况下，它更可靠地收敛",
"3.3Architecture",
"对于ADA的三个子网（请参见图1），我们采用WaveNet[43]架构，因为它允许我们精确",
"地指定它们的RF。编码器必须以较低的采样率生成查询序列，因此它必须包含下采样操作。",
"计算效率最高的方法是在编码器的前几层中降低此速率。然而，我们发现在输出端执行均",
"值池化是有帮助的，以鼓励编码器学习对时间偏移更不变的内部表示。这使得对输入中存",
"在的本地噪声进行编码变得更加困难；虽然无条件AR模型将简单地忽略任何噪声，因为它",
"是不可预测的，但ADA可能会尝试通过将此信息合并到代码序列中来“复制”噪声",
"2[] is implementedastf.stop_gradient(x)inTensorFlow",
"3f(ri)=Z,ReLU(a,)+e",
"4"
] |
{
"arxiv_id": "1806.10474",
"language": "zh",
"page_idx": 3
}
|
||
1806.10474-zh-4
|
1806.10474-zh-4.jpg
|
Table 1: Results for ADAs trained on audio waveforms, for different input representations and hop sizes. NLLs are reported in nats per timestep. The NLL of a large unconditional WaveNet model is included for comparison purposes. The asterisks indicate the architectures which we selected for further experiments. $\dagger$ The perplexity for model#1.7 is underestimated because the length of the evaluation sequences was limited to one second.
<table>
<tbody>
<tr>
<th rowspan="3">井</th>
<th rowspan="3">$\mathbf{MoDEL}$</th>
<th rowspan="3">INPUT FORMAT</th>
<th>$HOP$ SIZE</th>
<th>RECONSTF $TRAIN$</th>
<th>$NLL$ RUCTION I $EVAL$</th>
<th rowspan="3">CODEBOOK PERPLEXITY</th>
</tr>
<tr>
<th>$HOP$</th>
<th colspan="2">RECONSTRUCTION NLL</th>
</tr>
<tr>
<th>SIZE</th>
<th>$TRAIN$</th>
<th>$EVAL$</th>
</tr>
<tr>
<td>1.1</td>
<td>WaveNet</td>
<td>continuous</td>
<td>$N/A$</td>
<td>1.192</td>
<td>1.151</td>
<td>$N/A$</td>
</tr>
<tr>
<td>*1.2 1.3</td>
<td>VQ- VAE VQ- VAE</td>
<td>continuous one-hot</td>
<td>$\begin{array}{l}8\\8\end{array}$</td>
<td>0.766 0.711</td>
<td>0.734 0.682</td>
<td>227.3 199.6</td>
</tr>
<tr>
<td>1.4</td>
<td>AMAE</td>
<td>continuous</td>
<td>8</td>
<td>0.786</td>
<td>0.759</td>
<td>228.9</td>
</tr>
<tr>
<td>1.4 1.5</td>
<td>AMAE AMAE</td>
<td>continuous one-hot</td>
<td>8 1 8 </td>
<td>0.786 0.721</td>
<td>0.759 0.695</td>
<td>228.9 225.6</td>
</tr>
<tr>
<td>1.6</td>
<td>AMAE with softmax</td>
<td>one-hot</td>
<td>8</td>
<td>0.833</td>
<td>0.806</td>
<td>183.2</td>
</tr>
<tr>
<td>*1.7</td>
<td>VO- VAE</td>
<td>continuous</td>
<td>64</td>
<td>1.203</td>
<td>1.172</td>
<td>84.04+</td>
</tr>
</tbody>
</table>
### 4 $\textbf{Experiments}$
要评估不同的体系结构，我们使用由 413 小时录制的独奏钢琴音乐组成的数据集 (有关详细信息，请参见附录 B)。我们将自己限制在单乐器设置上，因为它减少了数据中音色的多样性。我们之所以选择钢琴，是因为它是一种复调乐器，可以使用许多高质量的录音
因为人类可以直观地识别出听起来不错的音乐，而无需研究它，所以我们可以有效地进行定性评估。不幸的是，定量评估更加困难。对于生成建模问题，这通常是正确的，但是在某些领域已经提出了度量标准，例如 InceptionScore[38]和 FrechetInceptionDistance[21]来衡量生成图像的真实感，或者 BLEU 分数[34] 来评估机器翻译结果。到目前为止，还没有为音乐开发出这样的指标。我们试图提供一些指标，但最终我们发现，倾听样本对于有意义地比较模型至关重要。因此，我们正在分享样本，我们敦促读者自己倾听和比较。它们可以在https://bit.ly/2IPXoDu 上找到。大多数样本长 10 秒，但我们还包括了一些分钟长的样本，用于最佳无条件模型(#3。表3 中的 6),以展示其捕获远程结构的能力。还包括一些来自真实录音的摘录，这些摘录被用作样本重建的输入
为了将音频波形表示为离散序列，我们采用了原始 WaveNet 论文[43]的设置：它们以 16kHz 的频率采样，并使用对数($\mu-law)变换将其量化为8位(256级)。这引人了一些量化噪声$ 但我们发现增加信号的位深度以减少这种噪声会大大加剧对局部结构建模的偏见。因为我们的目标正是捕获远程结构，所以这是不可取的
### 4.1 $\mathbf{ADA}$ architectures for audio
首先，我们比较了在波形上训练的几种 ADA 体系结构。我们训练几个具有不同输人表示的VQ-VAE 和 AMAE 模型。具有连续输人表示的模型以-1 和 1 之间的实值标量的形式接收音频输人。具有独热输入的模型改为采用 256 维独热向量 (在编码器和解码器端都是如此， 尽管我们发现它实际上只对前者有所不同)。除非另有说明，否则 AMAE 模型使用 ReLU, 然后进行除法归一化。有关模型体系结构和训练的详细信息，请参见附录 A
在表 1 中，我们报告了几种模型的 NLL (以每时间步的 nat 为单位) 和码本困惑。困惑是代码熵的指数，衡量代码的使用效率(值越高越好；值为 256 意味着所有代码的使用频率相同)。我们报告了有关训练和保留数据的 NLL,以表明模型不会过度拟合。作为基准，我们还训练了一个强大的无条件 WaveNet, 其 RF 为 384ms, 并报告 NLL 以进行比较。由于 ADA 能够对代码序列中的大量信息进行编码，因此我们获得了跳大小为8的较低的 NLL (#1。2-#1。5)-但是这些模型是有条件的，因此它们不能与无条件的 NLL 进行公平的比较 (#1 。1)。根据经验，我们还发现，具有更好的 NLL 的模型在感知重建质量方面不一定表现得更好
Input representation 比较#1。2和#1。3,我们看到这显著影响了结果。为编码器提供独热输入使得可以更准确地编码精确的信号值，这得到了多项式 NLL 的奖励。然而，从感知上看，#1的重建质量。2 优越。事实证明，#1。3 遭受部分码本崩溃
$\textbf{VQ- VAE vs. AMAE}$ AMAE 的性能稍差(比较#1。2和#1。4,#1。3 和#1。5),但码
本崩溃不是问题，如困惑所示。重建质量较差，重建的体积有时不一致
5
|
[
[
0.1748046875,
0.08587646484375,
0.82275390625,
0.08587646484375,
0.82275390625,
0.0997314453125,
0.1748046875,
0.0997314453125
],
[
0.1748046875,
0.0985107421875,
0.82275390625,
0.0985107421875,
0.82275390625,
0.11236572265625,
0.1748046875,
0.11236572265625
],
[
0.1748046875,
0.11236572265625,
0.82275390625,
0.11236572265625,
0.82275390625,
0.1256103515625,
0.1748046875,
0.1256103515625
],
[
0.1732177734375,
0.125,
0.82275390625,
0.125,
0.82275390625,
0.1383056640625,
0.1732177734375,
0.1383056640625
],
[
0.176513671875,
0.1395263671875,
0.393798828125,
0.1395263671875,
0.393798828125,
0.149658203125,
0.176513671875,
0.149658203125
],
[
0.53173828125,
0.1590576171875,
0.689453125,
0.1590576171875,
0.689453125,
0.17236328125,
0.53173828125,
0.17236328125
],
[
0.7060546875,
0.1590576171875,
0.78125,
0.1590576171875,
0.78125,
0.17041015625,
0.7060546875,
0.17041015625
],
[
0.486083984375,
0.160400390625,
0.5185546875,
0.160400390625,
0.5185546875,
0.1717529296875,
0.486083984375,
0.1717529296875
],
[
0.4150390625,
0.160400390625,
0.45751953125,
0.160400390625,
0.45751953125,
0.1717529296875,
0.4150390625,
0.1717529296875
],
[
0.256591796875,
0.169189453125,
0.310546875,
0.1717529296875,
0.3095703125,
0.186279296875,
0.2548828125,
0.184326171875
],
[
0.41015625,
0.17236328125,
0.464111328125,
0.17236328125,
0.464111328125,
0.184326171875,
0.41015625,
0.184326171875
],
[
0.486083984375,
0.17236328125,
0.5185546875,
0.17236328125,
0.5185546875,
0.184326171875,
0.486083984375,
0.184326171875
],
[
0.53369140625,
0.17236328125,
0.57861328125,
0.17236328125,
0.57861328125,
0.184326171875,
0.53369140625,
0.184326171875
],
[
0.62109375,
0.17041015625,
0.6611328125,
0.17041015625,
0.6611328125,
0.1849365234375,
0.62109375,
0.1849365234375
],
[
0.7060546875,
0.17236328125,
0.78369140625,
0.17236328125,
0.78369140625,
0.184326171875,
0.7060546875,
0.184326171875
],
[
0.2442626953125,
0.1767578125,
0.25244140625,
0.173583984375,
0.2548828125,
0.1767578125,
0.2467041015625,
0.179931640625
],
[
0.2279052734375,
0.1749267578125,
0.24267578125,
0.1685791015625,
0.25,
0.177978515625,
0.2344970703125,
0.184326171875
],
[
0.62158203125,
0.1849365234375,
0.66015625,
0.1875,
0.65869140625,
0.20263671875,
0.62109375,
0.2008056640625
],
[
0.7294921875,
0.1875,
0.7607421875,
0.1875,
0.7607421875,
0.20263671875,
0.7294921875,
0.20263671875
],
[
0.259033203125,
0.189453125,
0.316162109375,
0.189453125,
0.316162109375,
0.2008056640625,
0.259033203125,
0.2008056640625
],
[
0.221435546875,
0.1881103515625,
0.2459716796875,
0.1881103515625,
0.2459716796875,
0.2008056640625,
0.221435546875,
0.2008056640625
],
[
0.486083984375,
0.189453125,
0.515625,
0.189453125,
0.515625,
0.2020263671875,
0.486083984375,
0.2020263671875
],
[
0.5361328125,
0.189453125,
0.57421875,
0.189453125,
0.57421875,
0.2008056640625,
0.5361328125,
0.2008056640625
],
[
0.214111328125,
0.205810546875,
0.2442626953125,
0.205810546875,
0.2442626953125,
0.2183837890625,
0.214111328125,
0.2183837890625
],
[
0.40283203125,
0.1900634765625,
0.47216796875,
0.1900634765625,
0.47216796875,
0.2008056640625,
0.40283203125,
0.2008056640625
],
[
0.259033203125,
0.20703125,
0.316162109375,
0.20703125,
0.316162109375,
0.2183837890625,
0.259033203125,
0.2183837890625
],
[
0.53759765625,
0.20703125,
0.57275390625,
0.20703125,
0.57275390625,
0.2183837890625,
0.53759765625,
0.2183837890625
],
[
0.62353515625,
0.20703125,
0.65869140625,
0.20703125,
0.65869140625,
0.2183837890625,
0.62353515625,
0.2183837890625
],
[
0.72607421875,
0.205810546875,
0.7646484375,
0.205810546875,
0.7646484375,
0.2197265625,
0.72607421875,
0.2197265625
],
[
0.495849609375,
0.2083740234375,
0.50732421875,
0.2083740234375,
0.50732421875,
0.2197265625,
0.495849609375,
0.2197265625
],
[
0.40283203125,
0.2083740234375,
0.4697265625,
0.2083740234375,
0.4697265625,
0.2197265625,
0.40283203125,
0.2197265625
],
[
0.259033203125,
0.2183837890625,
0.314453125,
0.2183837890625,
0.314453125,
0.2298583984375,
0.259033203125,
0.2298583984375
],
[
0.5361328125,
0.2183837890625,
0.57275390625,
0.2183837890625,
0.57275390625,
0.2298583984375,
0.5361328125,
0.2298583984375
],
[
0.62353515625,
0.2183837890625,
0.65673828125,
0.2183837890625,
0.65673828125,
0.2298583984375,
0.62353515625,
0.2298583984375
],
[
0.72802734375,
0.2171630859375,
0.76318359375,
0.2171630859375,
0.76318359375,
0.2298583984375,
0.72802734375,
0.2298583984375
],
[
0.41259765625,
0.2197265625,
0.461669921875,
0.2197265625,
0.461669921875,
0.2310791015625,
0.41259765625,
0.2310791015625
],
[
0.22216796875,
0.2197265625,
0.24267578125,
0.2197265625,
0.24267578125,
0.2279052734375,
0.22216796875,
0.2279052734375
],
[
0.495849609375,
0.2197265625,
0.5087890625,
0.2197265625,
0.5087890625,
0.2291259765625,
0.495849609375,
0.2291259765625
],
[
0.221435546875,
0.236083984375,
0.2442626953125,
0.236083984375,
0.2442626953125,
0.2474365234375,
0.221435546875,
0.2474365234375
],
[
0.5361328125,
0.236083984375,
0.57275390625,
0.236083984375,
0.57275390625,
0.2474365234375,
0.5361328125,
0.2474365234375
],
[
0.62353515625,
0.236083984375,
0.65869140625,
0.236083984375,
0.65869140625,
0.2474365234375,
0.62353515625,
0.2474365234375
],
[
0.72802734375,
0.236083984375,
0.76318359375,
0.236083984375,
0.76318359375,
0.2474365234375,
0.72802734375,
0.2474365234375
],
[
0.40283203125,
0.2374267578125,
0.4697265625,
0.2374267578125,
0.4697265625,
0.248779296875,
0.40283203125,
0.248779296875
],
[
0.2607421875,
0.2374267578125,
0.303955078125,
0.2374267578125,
0.303955078125,
0.248779296875,
0.2607421875,
0.248779296875
],
[
0.495849609375,
0.2374267578125,
0.50732421875,
0.2374267578125,
0.50732421875,
0.2474365234375,
0.495849609375,
0.2474365234375
],
[
0.410888671875,
0.2474365234375,
0.463134765625,
0.2474365234375,
0.463134765625,
0.261962890625,
0.410888671875,
0.261962890625
],
[
0.5361328125,
0.24560546875,
0.57275390625,
0.248779296875,
0.57177734375,
0.261474609375,
0.53369140625,
0.2568359375
],
[
0.62158203125,
0.2474365234375,
0.66015625,
0.2474365234375,
0.66015625,
0.261474609375,
0.62158203125,
0.261474609375
],
[
0.72607421875,
0.2474365234375,
0.7646484375,
0.2474365234375,
0.7646484375,
0.261474609375,
0.72607421875,
0.261474609375
],
[
0.259033203125,
0.248779296875,
0.303955078125,
0.248779296875,
0.303955078125,
0.260009765625,
0.259033203125,
0.260009765625
],
[
0.223876953125,
0.248779296875,
0.2442626953125,
0.248779296875,
0.2442626953125,
0.2587890625,
0.223876953125,
0.2587890625
],
[
0.490234375,
0.25390625,
0.50244140625,
0.24560546875,
0.51318359375,
0.25390625,
0.5009765625,
0.261962890625
],
[
0.221435546875,
0.26513671875,
0.2442626953125,
0.26513671875,
0.2442626953125,
0.27783203125,
0.221435546875,
0.27783203125
],
[
0.259033203125,
0.26513671875,
0.386474609375,
0.26513671875,
0.386474609375,
0.279052734375,
0.259033203125,
0.279052734375
],
[
0.5361328125,
0.26513671875,
0.57421875,
0.26513671875,
0.57421875,
0.279052734375,
0.5361328125,
0.279052734375
],
[
0.62158203125,
0.264404296875,
0.66015625,
0.264404296875,
0.66015625,
0.27978515625,
0.62158203125,
0.27978515625
],
[
0.72802734375,
0.26513671875,
0.76318359375,
0.26513671875,
0.76318359375,
0.27783203125,
0.72802734375,
0.27783203125
],
[
0.495849609375,
0.266357421875,
0.5087890625,
0.266357421875,
0.5087890625,
0.27783203125,
0.495849609375,
0.27783203125
],
[
0.41259765625,
0.266357421875,
0.4599609375,
0.266357421875,
0.4599609375,
0.27783203125,
0.41259765625,
0.27783203125
],
[
0.25732421875,
0.2822265625,
0.316162109375,
0.2822265625,
0.316162109375,
0.296630859375,
0.25732421875,
0.296630859375
],
[
0.7197265625,
0.2822265625,
0.76953125,
0.2822265625,
0.76953125,
0.296630859375,
0.7197265625,
0.296630859375
],
[
0.40185546875,
0.28271484375,
0.468017578125,
0.28271484375,
0.468017578125,
0.294189453125,
0.40185546875,
0.294189453125
],
[
0.2132568359375,
0.28271484375,
0.24267578125,
0.28271484375,
0.24267578125,
0.294189453125,
0.2132568359375,
0.294189453125
],
[
0.491943359375,
0.28271484375,
0.51123046875,
0.28271484375,
0.51123046875,
0.29541015625,
0.491943359375,
0.29541015625
],
[
0.53759765625,
0.28271484375,
0.57275390625,
0.28271484375,
0.57275390625,
0.29541015625,
0.53759765625,
0.29541015625
],
[
0.62353515625,
0.28271484375,
0.65869140625,
0.28271484375,
0.65869140625,
0.29541015625,
0.62353515625,
0.29541015625
],
[
0.17236328125,
0.328857421875,
0.31787109375,
0.330810546875,
0.31787109375,
0.34521484375,
0.17236328125,
0.343505859375
],
[
0.176513671875,
0.3623046875,
0.82275390625,
0.3623046875,
0.82275390625,
0.375732421875,
0.176513671875,
0.375732421875
],
[
0.17724609375,
0.37744140625,
0.82177734375,
0.37744140625,
0.82177734375,
0.38818359375,
0.17724609375,
0.38818359375
],
[
0.176513671875,
0.389404296875,
0.76318359375,
0.389404296875,
0.76318359375,
0.40283203125,
0.176513671875,
0.40283203125
],
[
0.176513671875,
0.410400390625,
0.82275390625,
0.410400390625,
0.82275390625,
0.423583984375,
0.176513671875,
0.423583984375
],
[
0.176513671875,
0.4248046875,
0.82275390625,
0.4248046875,
0.82275390625,
0.438232421875,
0.176513671875,
0.438232421875
],
[
0.176513671875,
0.438232421875,
0.82275390625,
0.438232421875,
0.82275390625,
0.451904296875,
0.176513671875,
0.451904296875
],
[
0.176513671875,
0.451904296875,
0.82275390625,
0.451904296875,
0.82275390625,
0.46533203125,
0.176513671875,
0.46533203125
],
[
0.176513671875,
0.46533203125,
0.82421875,
0.46533203125,
0.82421875,
0.479248046875,
0.176513671875,
0.479248046875
],
[
0.17724609375,
0.479736328125,
0.82275390625,
0.479736328125,
0.82275390625,
0.49365234375,
0.17724609375,
0.49365234375
],
[
0.1732177734375,
0.49365234375,
0.82275390625,
0.492431640625,
0.82275390625,
0.505859375,
0.1732177734375,
0.5068359375
],
[
0.1748046875,
0.5068359375,
0.82421875,
0.5048828125,
0.82421875,
0.51953125,
0.1748046875,
0.521484375
],
[
0.1732177734375,
0.52099609375,
0.56884765625,
0.521484375,
0.56884765625,
0.53515625,
0.1732177734375,
0.5341796875
],
[
0.17724609375,
0.54052734375,
0.82275390625,
0.54052734375,
0.82275390625,
0.5537109375,
0.17724609375,
0.5537109375
],
[
0.176513671875,
0.5537109375,
0.82177734375,
0.5537109375,
0.82177734375,
0.568359375,
0.176513671875,
0.568359375
],
[
0.176513671875,
0.568359375,
0.82421875,
0.568359375,
0.82421875,
0.58203125,
0.176513671875,
0.58203125
],
[
0.176513671875,
0.58349609375,
0.51953125,
0.58349609375,
0.51953125,
0.5966796875,
0.176513671875,
0.5966796875
],
[
0.1732177734375,
0.61279296875,
0.414306640625,
0.61279296875,
0.414306640625,
0.626953125,
0.1732177734375,
0.626953125
],
[
0.176513671875,
0.64306640625,
0.82275390625,
0.64306640625,
0.82275390625,
0.6572265625,
0.176513671875,
0.6572265625
],
[
0.1748046875,
0.6572265625,
0.82275390625,
0.6572265625,
0.82275390625,
0.67041015625,
0.1748046875,
0.67041015625
],
[
0.17724609375,
0.671875,
0.82177734375,
0.671875,
0.82177734375,
0.68505859375,
0.17724609375,
0.68505859375
],
[
0.176513671875,
0.68505859375,
0.82421875,
0.68505859375,
0.82421875,
0.69873046875,
0.176513671875,
0.69873046875
],
[
0.1748046875,
0.69873046875,
0.7138671875,
0.69873046875,
0.7138671875,
0.7119140625,
0.1748046875,
0.7119140625
],
[
0.17724609375,
0.72021484375,
0.82177734375,
0.72021484375,
0.82177734375,
0.73095703125,
0.17724609375,
0.73095703125
],
[
0.176513671875,
0.73291015625,
0.82177734375,
0.73291015625,
0.82177734375,
0.7470703125,
0.176513671875,
0.7470703125
],
[
0.1748046875,
0.74560546875,
0.82275390625,
0.7470703125,
0.82275390625,
0.76123046875,
0.1748046875,
0.76025390625
],
[
0.176513671875,
0.76123046875,
0.82421875,
0.76123046875,
0.82421875,
0.7744140625,
0.176513671875,
0.7744140625
],
[
0.176513671875,
0.7734375,
0.82275390625,
0.7734375,
0.82275390625,
0.787109375,
0.176513671875,
0.787109375
],
[
0.1748046875,
0.78857421875,
0.82421875,
0.78857421875,
0.82421875,
0.8017578125,
0.1748046875,
0.8017578125
],
[
0.191162109375,
0.8017578125,
0.82275390625,
0.8017578125,
0.82275390625,
0.81494140625,
0.191162109375,
0.81494140625
],
[
0.17724609375,
0.81640625,
0.2091064453125,
0.81640625,
0.2091064453125,
0.82763671875,
0.17724609375,
0.82763671875
],
[
0.17724609375,
0.8369140625,
0.82275390625,
0.8359375,
0.82275390625,
0.849609375,
0.17724609375,
0.85107421875
],
[
0.176513671875,
0.85107421875,
0.82421875,
0.85107421875,
0.82421875,
0.8642578125,
0.176513671875,
0.8642578125
],
[
0.17724609375,
0.86572265625,
0.68896484375,
0.86572265625,
0.68896484375,
0.87548828125,
0.17724609375,
0.87548828125
],
[
0.1805419921875,
0.88623046875,
0.82275390625,
0.88623046875,
0.82275390625,
0.896484375,
0.1805419921875,
0.896484375
],
[
0.176513671875,
0.8984375,
0.68115234375,
0.8984375,
0.68115234375,
0.912109375,
0.176513671875,
0.912109375
],
[
0.49267578125,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94775390625,
0.49267578125,
0.94775390625
]
] |
[
"Table l: Results for ADAs trained on audio waveforms, for different input representations and hop",
"sizes. NLLs are reported in nats per timestep. The NLL of a large unconditional WaveNet model",
"is included for comparison purposes.The asterisks indicate the architectures which we selected for",
"further experiments. The perplexity for model # 1.7 is underestimated because the length of the evaluation",
"sequenceswaslimitedtoonesecond",
"RECONSTRUCTIONNLL",
"CODEBOOK",
"Hop",
"INPUT",
"MODEL",
"FORMAT",
"SIZE",
"TRAIN",
"EVAL",
"PERPLEXITY",
"",
"来",
"1.151",
"N/A",
"WaveNet",
"1.1",
"N/A",
"1.192",
"*1.2",
"continuous",
"VO-VAE",
"0.766",
"0.734",
"227.3",
"8",
"continuous",
"VQ-VAE",
"0.711",
"0.682",
"199.6",
"one-hot",
"1.3",
"8",
"1.4",
"0.786",
"0.759",
"228.9",
"continuous",
"AMAE",
"8",
"one-hot",
"0.721",
"0.695",
"225.6",
"AMAE",
"1.5",
"8",
"1.6",
"AMAEwithsoftmax",
"0.833",
"0.806",
"183.2",
"8",
"one-hot",
"VQ-VAE",
"84.04 ↑",
"continuous",
"*1.7",
"64",
"1.203",
"1.172",
"4Experiments",
"要评估不同的体系结构，我们使用由413小时录制的独奏钢琴音乐组成的数据集（有关详细",
"信息，请参见附录B）。我们将自已限制在单乐器设置上，因为它减少了数据中音色的多样",
"性。我们之所以选择钢琴，是因为它是一种复调乐器，可以使用许多高质量的录音",
"因为人类可以直观地识别出听起来不错的音乐，而无需研究它，所以我们可以有效地进行",
"定性评估。不幸的是，定量评估更加困难。对于生成建模问题，这通常是正确的，但是在某",
"些领域已经提出了度量标准，例如InceptionScore[38]和FrechetInceptionDistance[21]来衡",
"量生成图像的真实感，或者BLEU分数[34]来评估机器翻译结果。到目前为止，还没有为",
"音乐开发出这样的指标。我们试图提供一些指标，但最终我们发现，倾听样本对于有意义地",
"比较模型至关重要。因此，我们正在分享样本，我们敦促读者自已倾听和比较。它们可以在",
"https：//bit.1y/2IPxoDu上找到。大多数样本长10秒，但我们还包括了一些分钟长的样",
"本，用于最佳无条件模型（#3。表3中的6)，以展示其捕获远程结构的能力。还包括一些",
"来自真实录音的摘录，这些摘录被用作样本重建的输入",
"为了将音频波形表示为离散序列，我们采用了原始WaveNet论文[43]的设置：它们以16kHz",
"的频率采样，并使用对数（μ-law）变换将其量化为8位（256级）。这引人了一些量化噪声",
"但我们发现增加信号的位深度以减少这种噪声会大大加剧对局部结构建模的偏见。因为我",
"们的目标正是捕获远程结构，所以这是不可取的",
"4.1ADAarchitectures for audio",
"首先，我们比较了在波形上训练的几种ADA体系结构。我们训练几个具有不同输人表示的",
"VQ-VAE和AMAE模型。具有连续输人表示的模型以-1和1之间的实值标量的形式接收",
"音频输人。具有独热输人的模型改为采用256维独热向量（在编码器和解码器端都是如此",
"尽管我们发现它实际上只对前者有所不同）。除非另有说明，否则AMAE模型使用ReLU，",
"然后进行除法归一化。有关模型体系结构和训练的详细信息，请参见附录A",
"在表1中，我们报告了儿种模型的NLL（以每时间步的nat为单位）和码本困惑。困惑是代",
"码摘的指数，衡量代码的使用效率（值越高越好；值为256意味着所有代码的使用频率相",
"同）。我们报告了有关训练和保留数据的NLL，以表明模型不会过度拟合。作为基准，我们",
"还训练了一个强大的无条件WaveNet，其RF为384ms，并报告NLL以进行比较。由于ADA",
"能够对代码序列中的大量信息进行编码，因此我们获得了跳大小为8的较低的NLL（#1",
"2-#1。5）-但是这些模型是有条件的，因此它们不能与无条件的NLL进行公平的比较（#1",
"1)。根据经验，我们还发现，具有更好的NLL的模型在感知重建质量方面不一定表现得",
"更好",
"Inputrepresentation比较#1。2和#1。3，我们看到这显著影响了结果。为编码器提供",
"独热输人使得可以更准确地编码精确的信号值，这得到了多项式NLL的奖励。然而，从感",
"知上看，#1的重建质量。2优越。事实证明，#1。3遭受部分码本崩溃",
"VQ-VAEVS.AMAEAMAE的性能稍差（比较#1。2和#1。4,#1。3和#1。5），但码",
"本崩溃不是问题，如困惑所示。重建质量较差，重建的体积有时不一致",
"5"
] |
{
"arxiv_id": "1806.10474",
"language": "zh",
"page_idx": 4
}
|
||
2302.00871-zh-2
|
2302.00871-zh-2.jpg
|
安全方法。虽然目前的对话系统中存在许多安全问题，但我们特别关注降低反应毒性。1遵循 Dinan et al. (2021) 引人的分类法，我们的工作研究了减少对话系统中 INSTIGATOR 和YEA- SAYER 的影响。鉴于不安全的对话上下文，我们建议检索对类似对话上下文的示例性安全响应的演示。例如(参见图1),给定一个包含性别歧视的对话上下文，我们从其他包含性别歧视的对话上下文中检索安全响应的演示。然后可以在上下文中使用这些检索到的演示来引导模型生成理想的响应
具体来说，我们的工作旨在回答以下研究问
题：
Q1 Do in-context safety demonstrations improve
response safeness from dialogue systems?
Q2 How does in-context learning compare to
popular methods for safe response genera-
tion?
为了回答Q1 (Q15 ), 我们在三个模型家族中评估了我们的方法：OPT (Zhang et al.,2022) , LLaMA (Touvron et al., 2023) 和 Vicuna (Chiang et al., 2023) 。我们将精力集中在公开可用的OPT 模型上。使用自动(§5.1)和人类(§5.3) 评估，我们发现我们的方法在不降低一般响应质量的情况下降低了毒性。为了回答 Q2( §6 ),我们将我们的方法与三个流行的安全响应生成基线进行了比较。我们发现我们的方法在这些基线上具有竞争力，不需要任何培训。除了上述研究问题，我们还提出了一个广泛的消融 Appendix A 。例如 , 我们通过有限的安全演示来调查我们的方法的有效性
## 2 $\textbf{Related Work}$
我们的工作扩展了上下文学习和对话安全性的研究。下面，我们讨论为创建更安全的对话系统而提出的方法，并将它们与我们自己的方法进行对比。。我们还描述了上下文学习的相关工作
$\textbf{Safety Filters. 创 建 更 安 全 的 对 话 系 统 的 一 }$ 种流行方法涉及使用安全过滤器(Xu et al. 2021b; Shuster et al.,2022)。这些过滤器通常以三种方式使用：1)从模型的训练语料库(Solaiman and Dennison, 2021; Ngo et al., 2021) 中过滤不安全的内容；2) 阻止对模型 (Shuster et al., 2022) 的不安全输人；3) 从模型 (Xu et al. 2021b) 中过滤不安全的代。这些过滤器需要
${^{1}}$While our approach can be used to mitigate a range of safety issues, we focus on reducing toxicity as a wealth of datasets and tools exist for quantifying it.
大量对话，并将其标记为安全或不安全以训练 (Dinan et al., 2019a; Xu et al., 2021a; Barikeri et al., 2021; Sun et al., 2022)。与我们的方法相反，这些过滤器不容易适应新的不安全输人或新的不安全响应-您希望减轻的每种不良行为都必须反映在安全过滤器的训练语料库中
Safe Response Fine-Tuning. 创建更安全的对
话系统的另一种方法涉及对示例性安全响应(Ung et al., 2022; Kim et al., 2022) 的培训。已经发布了几个包含亲社会或安全反应的数据集。Ung et al. (2022) 收集了 SaFeR 对话，这是 Bot 对抗对话 (Xu et al.,2021a) 的增强变体，包含安全反馈和恢复响应。Kim et al.(2022)引人了ProsocialDialog, 这是一个对话数据集，其中包含基于社会经验法则的亲社会反应。最近的工作表明，对自己的反应进行改进的训练模型可以减少 (Sun et al., 2023; Bai et al., 2022b) 的危害。Zhou et al.(2023) 最近表明，即使对少量高质量响应进行微调也可以带来很大的安全性改进
Reinforcement Learning from Human Feedback. 从人类反馈中强化学习(RLHF)已成为创建更安全的语言模型 (Christiano et al., 2017; Ziegler et al., 2020; Bai et al., 2022a; Glaese et al., 2022; Ouyang et al., 2022; Bai et al., 2022b; OpenAI, 2022) 的有效方法。一般来说，RLHF 利用人类偏好数据来使语言模型与人类价值观保持一致。我们的方法是免费的 RLHF。在我们的工作中，我们证明了 Vicuna (Chiang et al., 2023), 一种源自 ChatGPT (OpenAI, 2022) 的模型，可以通过检索和上下文学习获得降低的毒性
Safe Decoding Procedures. 已经提出了几种解码程序，用于从语言模型安全生成。Schick et al.(2021) 建议使用语言模型对有毒内容的隐式知识来解毒生成。使用控制信号从语言模型生成条件来研究 Keskar et al.(2019) 和 Dinan et al. (2020) 。其他工作已经研究了使用分类器来指导 (Dathathri et al., 2020; Arora et al., 2022) 的生成。最后，Liu et al.(2021)提出了一种基于专家产品的解毒生成程序。与我们的方法一样，大多数这些程序不需要培训，但在推理时涉及额外的计算
In-Context Learning. 上下文学习 (Brown et al., 2020; Du et al., 2021; Rae et al., 2022) 已被证明在许多 NLP 任务 (Hu et al., 2022; Lampinen et al., 2022; Qiu et al., 2022) 中有效。据我们所知，我们对对话安全的情境学习进
行了第一次大规模评估。Askell et al.(2021)的
|
[
[
0.118408203125,
0.0872802734375,
0.484375,
0.0872802734375,
0.484375,
0.1009521484375,
0.118408203125,
0.1009521484375
],
[
0.51123046875,
0.0872802734375,
0.87744140625,
0.0872802734375,
0.87744140625,
0.097412109375,
0.51123046875,
0.097412109375
],
[
0.118408203125,
0.10272216796875,
0.47021484375,
0.10272216796875,
0.47021484375,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.51318359375,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.1175537109375,
0.1181640625,
0.486083984375,
0.1181640625,
0.486083984375,
0.1318359375,
0.1175537109375,
0.1318359375
],
[
0.51318359375,
0.12054443359375,
0.87744140625,
0.12054443359375,
0.87744140625,
0.1312255859375,
0.51318359375,
0.1312255859375
],
[
0.118408203125,
0.1353759765625,
0.486083984375,
0.1353759765625,
0.486083984375,
0.1484375,
0.118408203125,
0.1484375
],
[
0.51318359375,
0.1353759765625,
0.880859375,
0.1353759765625,
0.880859375,
0.1484375,
0.51318359375,
0.1484375
],
[
0.1175537109375,
0.149658203125,
0.48779296875,
0.15087890625,
0.48779296875,
0.1650390625,
0.1175537109375,
0.1644287109375
],
[
0.51318359375,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.1650390625,
0.51318359375,
0.1650390625
],
[
0.12005615234375,
0.16748046875,
0.48779296875,
0.16748046875,
0.48779296875,
0.18115234375,
0.12005615234375,
0.18115234375
],
[
0.51318359375,
0.16748046875,
0.84228515625,
0.16748046875,
0.84228515625,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.1175537109375,
0.1817626953125,
0.465087890625,
0.1817626953125,
0.465087890625,
0.1953125,
0.1175537109375,
0.1953125
],
[
0.118408203125,
0.1995849609375,
0.484375,
0.1995849609375,
0.484375,
0.213134765625,
0.118408203125,
0.213134765625
],
[
0.515625,
0.19482421875,
0.880859375,
0.19482421875,
0.880859375,
0.2078857421875,
0.515625,
0.2078857421875
],
[
0.12005615234375,
0.215576171875,
0.48779296875,
0.215576171875,
0.48779296875,
0.2286376953125,
0.12005615234375,
0.2286376953125
],
[
0.51123046875,
0.21142578125,
0.87890625,
0.21142578125,
0.87890625,
0.2244873046875,
0.51123046875,
0.2244873046875
],
[
0.118408203125,
0.232177734375,
0.48779296875,
0.232177734375,
0.48779296875,
0.2452392578125,
0.118408203125,
0.2452392578125
],
[
0.509765625,
0.226806640625,
0.87890625,
0.226806640625,
0.87890625,
0.240478515625,
0.509765625,
0.240478515625
],
[
0.118408203125,
0.2476806640625,
0.35693359375,
0.2476806640625,
0.35693359375,
0.26123046875,
0.118408203125,
0.26123046875
],
[
0.5146484375,
0.243408203125,
0.8818359375,
0.243408203125,
0.8818359375,
0.257080078125,
0.5146484375,
0.257080078125
],
[
0.1368408203125,
0.26416015625,
0.484375,
0.26416015625,
0.484375,
0.27783203125,
0.1368408203125,
0.27783203125
],
[
0.51318359375,
0.2587890625,
0.8818359375,
0.2587890625,
0.8818359375,
0.273681640625,
0.51318359375,
0.273681640625
],
[
0.1175537109375,
0.279052734375,
0.14697265625,
0.28076171875,
0.145263671875,
0.29443359375,
0.1158447265625,
0.292236328125
],
[
0.51318359375,
0.273681640625,
0.87744140625,
0.273681640625,
0.87744140625,
0.287353515625,
0.51318359375,
0.287353515625
],
[
0.5146484375,
0.29150390625,
0.87744140625,
0.29150390625,
0.87744140625,
0.3046875,
0.5146484375,
0.3046875
],
[
0.1234130859375,
0.30810546875,
0.48779296875,
0.30712890625,
0.48779296875,
0.32177734375,
0.1234130859375,
0.322509765625
],
[
0.51123046875,
0.30712890625,
0.880859375,
0.30712890625,
0.880859375,
0.32177734375,
0.51123046875,
0.32177734375
],
[
0.154541015625,
0.32470703125,
0.465087890625,
0.32470703125,
0.465087890625,
0.339111328125,
0.154541015625,
0.339111328125
],
[
0.5146484375,
0.32373046875,
0.87744140625,
0.32373046875,
0.87744140625,
0.334228515625,
0.5146484375,
0.334228515625
],
[
0.51318359375,
0.34033203125,
0.87744140625,
0.34033203125,
0.87744140625,
0.350830078125,
0.51318359375,
0.350830078125
],
[
0.124267578125,
0.3466796875,
0.486083984375,
0.3466796875,
0.486083984375,
0.35986328125,
0.124267578125,
0.35986328125
],
[
0.51123046875,
0.355224609375,
0.87890625,
0.355224609375,
0.87890625,
0.3681640625,
0.51123046875,
0.3681640625
],
[
0.154541015625,
0.3623046875,
0.484375,
0.3623046875,
0.484375,
0.377197265625,
0.154541015625,
0.377197265625
],
[
0.51123046875,
0.37060546875,
0.880859375,
0.371826171875,
0.880859375,
0.385986328125,
0.51123046875,
0.385498046875
],
[
0.154541015625,
0.37890625,
0.1956787109375,
0.37890625,
0.1956787109375,
0.392578125,
0.154541015625,
0.392578125
],
[
0.51318359375,
0.385986328125,
0.880859375,
0.38720703125,
0.880859375,
0.402099609375,
0.51318359375,
0.40087890625
],
[
0.12005615234375,
0.40673828125,
0.484375,
0.40673828125,
0.484375,
0.42041015625,
0.12005615234375,
0.42041015625
],
[
0.5146484375,
0.40625,
0.5322265625,
0.40625,
0.5322265625,
0.4150390625,
0.5146484375,
0.4150390625
],
[
0.118408203125,
0.4228515625,
0.48779296875,
0.4228515625,
0.48779296875,
0.43701171875,
0.118408203125,
0.43701171875
],
[
0.1158447265625,
0.435791015625,
0.486083984375,
0.438232421875,
0.486083984375,
0.45361328125,
0.1158447265625,
0.451904296875
],
[
0.5205078125,
0.43115234375,
0.8818359375,
0.43115234375,
0.8818359375,
0.44482421875,
0.5205078125,
0.44482421875
],
[
0.51123046875,
0.446533203125,
0.880859375,
0.44775390625,
0.880859375,
0.4619140625,
0.51123046875,
0.46142578125
],
[
0.118408203125,
0.454833984375,
0.486083984375,
0.454833984375,
0.486083984375,
0.468505859375,
0.118408203125,
0.468505859375
],
[
0.51318359375,
0.46435546875,
0.880859375,
0.46435546875,
0.880859375,
0.47802734375,
0.51318359375,
0.47802734375
],
[
0.118408203125,
0.47021484375,
0.486083984375,
0.47021484375,
0.486083984375,
0.48388671875,
0.118408203125,
0.48388671875
],
[
0.51318359375,
0.479736328125,
0.8818359375,
0.479736328125,
0.8818359375,
0.493408203125,
0.51318359375,
0.493408203125
],
[
0.118408203125,
0.48681640625,
0.484375,
0.48681640625,
0.484375,
0.50048828125,
0.118408203125,
0.50048828125
],
[
0.51318359375,
0.495361328125,
0.880859375,
0.495361328125,
0.880859375,
0.51025390625,
0.51318359375,
0.51025390625
],
[
0.118408203125,
0.5029296875,
0.48779296875,
0.5029296875,
0.48779296875,
0.5166015625,
0.118408203125,
0.5166015625
],
[
0.5146484375,
0.51123046875,
0.880859375,
0.51123046875,
0.880859375,
0.52490234375,
0.5146484375,
0.52490234375
],
[
0.1158447265625,
0.51953125,
0.48779296875,
0.5185546875,
0.48779296875,
0.5322265625,
0.1158447265625,
0.533203125
],
[
0.5146484375,
0.52783203125,
0.880859375,
0.52783203125,
0.880859375,
0.54150390625,
0.5146484375,
0.54150390625
],
[
0.118408203125,
0.53515625,
0.484375,
0.53515625,
0.484375,
0.548828125,
0.118408203125,
0.548828125
],
[
0.5146484375,
0.54443359375,
0.8818359375,
0.54443359375,
0.8818359375,
0.55810546875,
0.5146484375,
0.55810546875
],
[
0.118408203125,
0.5517578125,
0.484375,
0.5517578125,
0.484375,
0.5654296875,
0.118408203125,
0.5654296875
],
[
0.51318359375,
0.56005859375,
0.880859375,
0.56005859375,
0.880859375,
0.57373046875,
0.51318359375,
0.57373046875
],
[
0.12176513671875,
0.568359375,
0.484375,
0.568359375,
0.484375,
0.58203125,
0.12176513671875,
0.58203125
],
[
0.51318359375,
0.576171875,
0.880859375,
0.576171875,
0.880859375,
0.59033203125,
0.51318359375,
0.59033203125
],
[
0.118408203125,
0.58447265625,
0.48779296875,
0.58447265625,
0.48779296875,
0.59716796875,
0.118408203125,
0.59716796875
],
[
0.51123046875,
0.59033203125,
0.880859375,
0.5927734375,
0.880859375,
0.60693359375,
0.51123046875,
0.60498046875
],
[
0.1175537109375,
0.5986328125,
0.35693359375,
0.599609375,
0.35693359375,
0.6142578125,
0.1175537109375,
0.61328125
],
[
0.51123046875,
0.60693359375,
0.55322265625,
0.609375,
0.5517578125,
0.62353515625,
0.509765625,
0.62158203125
],
[
0.11248779296875,
0.626953125,
0.274658203125,
0.62744140625,
0.274658203125,
0.642578125,
0.11248779296875,
0.64111328125
],
[
0.51708984375,
0.6357421875,
0.880859375,
0.6357421875,
0.880859375,
0.6494140625,
0.51708984375,
0.6494140625
],
[
0.118408203125,
0.65478515625,
0.484375,
0.65478515625,
0.484375,
0.66845703125,
0.118408203125,
0.66845703125
],
[
0.51318359375,
0.65185546875,
0.8818359375,
0.65185546875,
0.8818359375,
0.6650390625,
0.51318359375,
0.6650390625
],
[
0.12005615234375,
0.6728515625,
0.483642578125,
0.6728515625,
0.483642578125,
0.68310546875,
0.12005615234375,
0.68310546875
],
[
0.51318359375,
0.66845703125,
0.880859375,
0.66845703125,
0.880859375,
0.681640625,
0.51318359375,
0.681640625
],
[
0.118408203125,
0.68701171875,
0.48193359375,
0.68701171875,
0.48193359375,
0.69775390625,
0.118408203125,
0.69775390625
],
[
0.5146484375,
0.68408203125,
0.880859375,
0.68408203125,
0.880859375,
0.69775390625,
0.5146484375,
0.69775390625
],
[
0.1175537109375,
0.70166015625,
0.484375,
0.703125,
0.484375,
0.71728515625,
0.1175537109375,
0.71630859375
],
[
0.51123046875,
0.69970703125,
0.880859375,
0.70068359375,
0.880859375,
0.71435546875,
0.51123046875,
0.71337890625
],
[
0.1175537109375,
0.7197265625,
0.15869140625,
0.7197265625,
0.15869140625,
0.73291015625,
0.1175537109375,
0.73291015625
],
[
0.5146484375,
0.71728515625,
0.880859375,
0.71728515625,
0.880859375,
0.73095703125,
0.5146484375,
0.73095703125
],
[
0.5146484375,
0.73291015625,
0.87890625,
0.73291015625,
0.87890625,
0.74658203125,
0.5146484375,
0.74658203125
],
[
0.12176513671875,
0.74365234375,
0.489501953125,
0.7451171875,
0.489501953125,
0.76025390625,
0.12176513671875,
0.75830078125
],
[
0.51123046875,
0.74755859375,
0.87890625,
0.74755859375,
0.87890625,
0.76123046875,
0.51123046875,
0.76123046875
],
[
0.118408203125,
0.76123046875,
0.489501953125,
0.76123046875,
0.489501953125,
0.77490234375,
0.118408203125,
0.77490234375
],
[
0.51318359375,
0.765625,
0.86376953125,
0.765625,
0.86376953125,
0.779296875,
0.51318359375,
0.779296875
],
[
0.1175537109375,
0.77783203125,
0.48779296875,
0.77783203125,
0.48779296875,
0.79150390625,
0.1175537109375,
0.79150390625
],
[
0.51123046875,
0.78076171875,
0.880859375,
0.78076171875,
0.880859375,
0.79443359375,
0.51123046875,
0.79443359375
],
[
0.118408203125,
0.7919921875,
0.48779296875,
0.7919921875,
0.48779296875,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.51123046875,
0.79736328125,
0.64208984375,
0.79736328125,
0.64208984375,
0.81103515625,
0.51123046875,
0.81103515625
],
[
0.118408203125,
0.81103515625,
0.483642578125,
0.81103515625,
0.483642578125,
0.82177734375,
0.118408203125,
0.82177734375
],
[
0.1175537109375,
0.826171875,
0.483642578125,
0.826171875,
0.483642578125,
0.8388671875,
0.1175537109375,
0.8388671875
],
[
0.5205078125,
0.82373046875,
0.880859375,
0.82470703125,
0.880859375,
0.8388671875,
0.5205078125,
0.83837890625
],
[
0.118408203125,
0.84130859375,
0.484375,
0.84130859375,
0.484375,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.5146484375,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.5146484375,
0.85498046875
],
[
0.1175537109375,
0.85546875,
0.484375,
0.85693359375,
0.484375,
0.87158203125,
0.1175537109375,
0.87060546875
],
[
0.51318359375,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.51318359375,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.88720703125,
0.51318359375,
0.88720703125
],
[
0.139404296875,
0.8798828125,
0.48779296875,
0.8818359375,
0.48779296875,
0.896484375,
0.139404296875,
0.89453125
],
[
0.118408203125,
0.89453125,
0.48779296875,
0.8935546875,
0.48779296875,
0.90673828125,
0.118408203125,
0.90771484375
],
[
0.51123046875,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1158447265625,
0.90380859375,
0.366943359375,
0.90673828125,
0.366943359375,
0.92041015625,
0.1158447265625,
0.91748046875
],
[
0.51123046875,
0.90478515625,
0.87890625,
0.90478515625,
0.87890625,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"安全方法。虽然目前的对话系统中存在许多",
"大量对话，并将其标记为安全或不安全以",
"安全问题，但我们特别关注降低反应毒性。",
"练 (Dinan et al., 2019a; Xu et al., 2021a; Barikeri",
"遵循Dinanetal.（2021）引人的分类法，我们",
"etal.2021；Sunetal.，2022）。与我们的方法相",
"的工作研究了减少对话系统中INSTIGATOR和",
"反，这些过滤器不容易适应新的不安全输入或",
"YEA-SAYER的影响。鉴于不安全的对话上下",
"新的不安全响应-您希望减轻的每种不良行为",
"文，我们建议检索对类似对话上下文的示例性",
"都必须反映在安全过滤器的训练语料库中",
"安全响应的演示。例如（参见图1），给定",
"个包含性别歧视的对话上下文，我们从其他包",
"SafeResponseFine-Tuning.创建更安全的对",
"含性别歧视的对话上下文中检索安全响应的演",
"话系统的另一种方法涉及对示例性安全响应",
"示。然后可以在上下文中使用这些检索到的演",
"(Ung et al.,2022;Kimet al.,2022)的培训l。已经",
"示来引导模型生成理想的响应",
"发布了几个包含亲社会或安全反应的数据集。",
"具体来说，我们的工作旨在回答以下研究问",
"Ungetal.（2022）收集了SaFeR对话，这是Bot",
"题：",
"对抗对话(Xuetal.,202la）的增强变体，包含",
"安全反馈和恢复响应。Kimetal.(2022）引入了",
"Q1 Do in-context safety demonstrations improve",
"ProsocialDialog，这是一个对话数据集，其中包",
"response safeness from dialogue systems?",
"含基于社会经验法则的亲社会反应。最近的工",
"作表明：对自已的反应进行改进的训练模型可",
"Q2 How does in-context learning compare to",
"以减少(Sun et al.,2023Bai et al.,2022b)的危",
"popular methods for safe response genera",
"害。Zhouetal.（2023）最近表明，即使对少量高",
"tion?",
"质量响应进行微调也可以带来很大的安全性改",
"为了回答Q1（Q15），我们在三个模型家族中",
"进",
"评估了我们的方法：OPT（Zhangetal.，2022），",
"LLaMA (Touvron et al., 2023)和 Vicuna (Chiang",
"Reinforcement Learning from Human Feed-",
"back．从人类反馈中强化学习（RLHF)已成为",
"etal.,2023）。我们将精力集中在公开可用的",
"创建更安全的语言模型（Christianoetal.，2017;",
"OPT模型上。使用自动（s5.1）和人类（85.3）",
"Ziegler et al., 2020; Bai et al., 2022a; Glaese et al.",
"评估，我们发现我们的方法在不降低一般响应",
"2022; Ouyang et al., 2022; Bai et al., 2022b; Ope",
"质量的情况下降低了毒性。为了回答Q2（$6",
"nAI,2022)的有效方法。一般来说，RLHF利",
"），我们将我们的方法与三个流行的安全响应",
"用人类偏好数据来使语言模型与人类价值观",
"生成基线进行了比较。我们发现我们的方法在",
"保持一致。我们的方法是免费的RLHF。在我",
"这些基线上具有竞争力，不需要任何培训。除",
"们的工作中，我们证明了Vicuna（Chiangetal.,",
"了上述研究问题，我们还提出了一个广泛的消",
"2023），一种源自ChatGPT（OpenAI，2022）的",
"融AppendixA。例如，我们通过有限的安全演",
"模型，可以通过检索和上下文学习获得降低的",
"示来调查我们的方法的有效性",
"毒性",
"2Related Work",
"SafeDecodingProcedures.已经提出了几种",
"我们的工作扩展了上下文学习和对话安全性的",
"解码程序，用于从语言模型安全生成。Schick",
"研究。下面，我们讨论为创建更安全的对话系",
"etal.（2021）建议使用语言模型对有毒内容的",
"统而提出的方法，并将它们与我们自已的方法",
"隐式知识来解毒生成。使用控制信号从语言模",
"进行对比。。我们还描述了上下文学习的相关",
"型生成条件来研究Keskaretal.（2019）和Dinan",
"工作",
"etal.（2020）。其他工作已经研究了使用分类器",
"来指导(Dathathri et al.,2020;Arora et al.,2022)",
"SafetyFilters。创建更安全的对话系统的一",
"的生成。最后，Liuetal.(2021)提出了一种基",
"种流行方法涉及使用安全过滤器（Xuetal.",
"于专家产品的解毒生成程序。与我们的方法",
"2021b；Shusteretal.，2022）。这些过滤器通常",
"样，大多数这些程序不需要培训，但在推理时",
"以三种方式使用：1）从模型的训练语料库（So-",
"涉及额外的计算",
"laiman and Dennison,2021;Ngo etal.,2021)中",
"过滤不安全的内容；2）阻止对模型（Shuster",
"In-ContextLearning.上下文学习(Brown",
"etal.,2022)的不安全输人；3）从模型（Xuetal.，",
"et al., 2020; Du et al., 2021; Rae et al., 2022)",
"2021b）中过滤不安全的代。这些过滤器需要",
"已被证明在许多NLP任务(Huetal.，2022;",
"Lampinen et al., 2022; Qiu et al., 2022) 中有效。",
"IWhile our approach can be used to mitigate a range of",
"safety issues, we focus on reducing toxicity as a wealth of",
"据我们所知，我们对对话安全的情境学习进",
"datasets and tools exist for quantifying it.",
"行了第一次大规模评估。Askelletal.(2021)的"
] |
{
"arxiv_id": "2302.00871",
"language": "zh",
"page_idx": 2
}
|
||
2302.00871-zh-3
|
2302.00871-zh-3.jpg
|
Figure 2: Prompt for response generation. Each prompt consists of the retrieved demonstrations and the target context. Each safety demonstration is separated by an empty line and the target context is separated from the safety demonstrations by an empty line.
工作与我们自己的工作最相关。虽然他们调查上下文学习以进行对齐，但他们不调查检索相关演示。最近的工作还研究了有关上下文学习的基本问题。Lu et al.(2022b)调查了上下文演示顺序对性能的影响。我们发现上下文演示的顺序不会影响响应质量或安全性。Liu et al. $(2022b)$演示了基于与测试查询的语义相似性检索上下文演示可提高 NLU 基准的性能。我们发现检索与对话上下文高度相似的演示对于减少反应毒性很有用。最后，Rubin et al.(2022) 和 Agrawal et al. (2023) 研究了选择上下文演示的方法。我们还研究了为对话安全选择上下文演示的不同方法。
# 3 Methodology
我们研究了一种基于检索的方法，用于从仅解码器的 Transformer (Vaswani et al., 2017) 模型中安全生成响应。具体而言，我们尝试使用不同尺寸的 OPT (Zhang et al., 2022) , LLaMA (Touvron et al.,2023) 和 Vicuna (Chiang et al., 2023) 模型。我们主要使用 OPT 模型进行实验，因为模型代码和权重是公开可用的，但是，我们还在整个工作中重点介绍了相关的LLaMA 和 Vicuna 结果 (有关完整结果，请参见 Appendix E )
此后，我们将要生成响应的对话上下文称为目标上下文，将安全模型行为的演示称为安全演示。在高层次上，我们的方法包括两个步骤： 1) 我们根据与目标上下文的相似性来检索安全演示；2)我们在上下文中使用检索到的安全演示来生成条件。我们在下面详细描述这些步骤
$1) \textbf{ Retrieving Safety Demonstrations. }$我们研究了三种为目标上下文选择安全演示的方法：
1) 随机选择演示；2) 使用 BM25 (Robertson and Zaragoza, 2009)选择演示；3) 使用句子转换器 (Reimers and Gurevych, 2019)。$^{2}$对于每个检索器，我们使用目标上下文作为查询来选择演示。这些安全演示是由不安全的话语和亲社会反应组成的整个对话。在我们的整个工
$^{2}$We also investigated using a Dense Passage Retriever (DPR; Karpukhin et al. 2020 ) to select demonstrations and defer readers to Appendix D for results and additional retriever details.
作中，我们将我们的 SentenceTransformer 检索
器称为“密集”检索器
$2) \textbf{ Response Generation. }$ 一旦选择了安全演示，我们将在上下文中使用它们来生成条件。具体来说，给定$K$安全演示和目标上下文，我们使用 Figure 2 中显示的提示格式。我们在输人中的每个对话前加上“两个人之间的对话”, 以进行对话。演示根据其检索分数按降序放置在提示中。更简单地说，排名最高的演示位于输入的开头。目标上下文放置在输入的末尾。我们标记每个话语的说话者(人1或人2),并在提示末尾为我们要为其生成响应的说话者提供尾随注释(在Figure 2中，这是人 2)
### 4 $\textbf{Experimental Setup}$
下面，我们描述了这项工作中使用的对话数据集。此外，我们还讨论了如何评估反应的安全性和相关性(i。e。质量)
## 4.1 Dialogue Datasets
我们在这项工作中尝试了三个对话数据集。来自这些数据集的对话被用作响应生成的输入或安全演示。在目标上下文和安全演示中，我们最多使用两个对话回合。我们在下面描述每个数据集，并将读者推迟到 Appendix H 以获取更多详细信息
$\textbf{ProsocialDialog( Kim et al. , 2022) . }$ ProsocialDialog 包含带有亲社会响应的不安全话语。我们使用来自 ProsocialDialog 培训部门的 42K 对话作为我们所有实验的安全演示来源。我们还尝试从 ProsocialDialog 的验证拆分中生成对 7 K 对话的响应。3
DiaSafety (Sun et al.,2022) . DiaSafety 是对抗性话语的集合，可以从会话模型中非法获取不安全的响应。我们尝试从 DiaSafety 的验证集生成对 1K 对话的响应。我们使用 DiaSafety 来评估对不安全输入的响应生成。我们注意到DiaSafety 的每个目标上下文都由一个单一的话语组成
Commonsense-Dialogues (Zhou et al., 2021) . 常识-对话是基于社会背景的对话的集合。我们尝试从“常识对话”的验证集中生成对 1K 对话的响应。我们使用常识对话来评估对安全输入的响应生成
We consider a conversation turn to be an exchange be-
tween two speakers in a conversation.
|
[
[
0.5146484375,
0.0872802734375,
0.87890625,
0.0872802734375,
0.87890625,
0.1009521484375,
0.5146484375,
0.1009521484375
],
[
0.1175537109375,
0.10272216796875,
0.489501953125,
0.10272216796875,
0.489501953125,
0.11639404296875,
0.1175537109375,
0.11639404296875
],
[
0.51318359375,
0.10272216796875,
0.69775390625,
0.10272216796875,
0.69775390625,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.1175537109375,
0.1175537109375,
0.483642578125,
0.11639404296875,
0.483642578125,
0.1300048828125,
0.1175537109375,
0.1312255859375
],
[
0.118408203125,
0.1318359375,
0.486083984375,
0.1318359375,
0.486083984375,
0.1455078125,
0.118408203125,
0.1455078125
],
[
0.515625,
0.133056640625,
0.87890625,
0.133056640625,
0.87890625,
0.146728515625,
0.515625,
0.146728515625
],
[
0.1175537109375,
0.146728515625,
0.484375,
0.146728515625,
0.484375,
0.15673828125,
0.1175537109375,
0.15673828125
],
[
0.51123046875,
0.1484375,
0.87890625,
0.1484375,
0.87890625,
0.162109375,
0.51123046875,
0.162109375
],
[
0.1158447265625,
0.1591796875,
0.44580078125,
0.1602783203125,
0.44580078125,
0.17333984375,
0.1158447265625,
0.1728515625
],
[
0.51318359375,
0.166259765625,
0.8818359375,
0.166259765625,
0.8818359375,
0.179931640625,
0.51318359375,
0.179931640625
],
[
0.5146484375,
0.182861328125,
0.880859375,
0.182861328125,
0.880859375,
0.196533203125,
0.5146484375,
0.196533203125
],
[
0.1175537109375,
0.20068359375,
0.486083984375,
0.20068359375,
0.486083984375,
0.21435546875,
0.1175537109375,
0.21435546875
],
[
0.51318359375,
0.198974609375,
0.87890625,
0.198974609375,
0.87890625,
0.2120361328125,
0.51318359375,
0.2120361328125
],
[
0.118408203125,
0.21728515625,
0.48779296875,
0.21728515625,
0.48779296875,
0.23095703125,
0.118408203125,
0.23095703125
],
[
0.51318359375,
0.21435546875,
0.880859375,
0.21435546875,
0.880859375,
0.22802734375,
0.51318359375,
0.22802734375
],
[
0.118408203125,
0.2327880859375,
0.486083984375,
0.2327880859375,
0.486083984375,
0.2464599609375,
0.118408203125,
0.2464599609375
],
[
0.51318359375,
0.23095703125,
0.880859375,
0.23095703125,
0.880859375,
0.24462890625,
0.51318359375,
0.24462890625
],
[
0.12005615234375,
0.248779296875,
0.48779296875,
0.248779296875,
0.48779296875,
0.262451171875,
0.12005615234375,
0.262451171875
],
[
0.51123046875,
0.2464599609375,
0.87890625,
0.2464599609375,
0.87890625,
0.260009765625,
0.51123046875,
0.260009765625
],
[
0.118408203125,
0.265380859375,
0.48193359375,
0.265380859375,
0.48193359375,
0.275634765625,
0.118408203125,
0.275634765625
],
[
0.51123046875,
0.262451171875,
0.87890625,
0.262451171875,
0.87890625,
0.275634765625,
0.51123046875,
0.275634765625
],
[
0.118408203125,
0.28076171875,
0.48779296875,
0.28076171875,
0.48779296875,
0.29443359375,
0.118408203125,
0.29443359375
],
[
0.51123046875,
0.279052734375,
0.880859375,
0.279052734375,
0.880859375,
0.292236328125,
0.51123046875,
0.292236328125
],
[
0.118408203125,
0.297607421875,
0.48779296875,
0.297607421875,
0.48779296875,
0.311279296875,
0.118408203125,
0.311279296875
],
[
0.51123046875,
0.29345703125,
0.82373046875,
0.29345703125,
0.82373046875,
0.30712890625,
0.51123046875,
0.30712890625
],
[
0.118408203125,
0.312255859375,
0.484375,
0.312255859375,
0.484375,
0.325927734375,
0.118408203125,
0.325927734375
],
[
0.118408203125,
0.330078125,
0.48779296875,
0.330078125,
0.48779296875,
0.34326171875,
0.118408203125,
0.34326171875
],
[
0.50537109375,
0.32666015625,
0.7197265625,
0.328857421875,
0.7197265625,
0.34326171875,
0.50537109375,
0.341552734375
],
[
0.118408203125,
0.345703125,
0.484375,
0.345703125,
0.484375,
0.359375,
0.118408203125,
0.359375
],
[
0.1175537109375,
0.361083984375,
0.486083984375,
0.35986328125,
0.486083984375,
0.374755859375,
0.1175537109375,
0.3759765625
],
[
0.5146484375,
0.359375,
0.87890625,
0.359375,
0.87890625,
0.372314453125,
0.5146484375,
0.372314453125
],
[
0.1175537109375,
0.377197265625,
0.486083984375,
0.377197265625,
0.486083984375,
0.39013671875,
0.1175537109375,
0.39013671875
],
[
0.51318359375,
0.374755859375,
0.87890625,
0.374755859375,
0.87890625,
0.388427734375,
0.51318359375,
0.388427734375
],
[
0.1175537109375,
0.392578125,
0.255126953125,
0.393798828125,
0.255126953125,
0.407958984375,
0.1175537109375,
0.40673828125
],
[
0.51318359375,
0.391357421875,
0.7197265625,
0.391357421875,
0.7197265625,
0.405029296875,
0.51318359375,
0.405029296875
],
[
0.11419677734375,
0.419189453125,
0.26611328125,
0.4228515625,
0.26611328125,
0.438232421875,
0.11419677734375,
0.435302734375
],
[
0.51123046875,
0.42333984375,
0.7001953125,
0.42333984375,
0.7001953125,
0.43701171875,
0.51123046875,
0.43701171875
],
[
0.118408203125,
0.448974609375,
0.48779296875,
0.448974609375,
0.48779296875,
0.4619140625,
0.118408203125,
0.4619140625
],
[
0.51123046875,
0.448974609375,
0.87890625,
0.448974609375,
0.87890625,
0.4619140625,
0.51123046875,
0.4619140625
],
[
0.118408203125,
0.465576171875,
0.489501953125,
0.465576171875,
0.489501953125,
0.478515625,
0.118408203125,
0.478515625
],
[
0.51318359375,
0.46435546875,
0.87890625,
0.46435546875,
0.87890625,
0.47802734375,
0.51318359375,
0.47802734375
],
[
0.118408203125,
0.48095703125,
0.48779296875,
0.48095703125,
0.48779296875,
0.49462890625,
0.118408203125,
0.49462890625
],
[
0.51318359375,
0.48095703125,
0.880859375,
0.48095703125,
0.880859375,
0.49462890625,
0.51318359375,
0.49462890625
],
[
0.118408203125,
0.49755859375,
0.489501953125,
0.49755859375,
0.489501953125,
0.51123046875,
0.118408203125,
0.51123046875
],
[
0.51318359375,
0.496337890625,
0.87890625,
0.496337890625,
0.87890625,
0.51025390625,
0.51318359375,
0.51025390625
],
[
0.1175537109375,
0.5126953125,
0.489501953125,
0.51318359375,
0.489501953125,
0.52783203125,
0.1175537109375,
0.52685546875
],
[
0.51123046875,
0.51123046875,
0.880859375,
0.5126953125,
0.880859375,
0.52685546875,
0.51123046875,
0.525390625
],
[
0.1175537109375,
0.529296875,
0.48779296875,
0.52783203125,
0.48779296875,
0.5419921875,
0.1175537109375,
0.54345703125
],
[
0.51318359375,
0.529296875,
0.623046875,
0.529296875,
0.623046875,
0.54345703125,
0.51318359375,
0.54345703125
],
[
0.118408203125,
0.5458984375,
0.484375,
0.5458984375,
0.484375,
0.55859375,
0.118408203125,
0.55859375
],
[
0.12005615234375,
0.5634765625,
0.483642578125,
0.5634765625,
0.483642578125,
0.57373046875,
0.12005615234375,
0.57373046875
],
[
0.51708984375,
0.56005859375,
0.8818359375,
0.56005859375,
0.8818359375,
0.57373046875,
0.51708984375,
0.57373046875
],
[
0.1175537109375,
0.57763671875,
0.48779296875,
0.57763671875,
0.48779296875,
0.59130859375,
0.1175537109375,
0.59130859375
],
[
0.51318359375,
0.576171875,
0.8818359375,
0.576171875,
0.8818359375,
0.5888671875,
0.51318359375,
0.5888671875
],
[
0.118408203125,
0.59423828125,
0.2435302734375,
0.59423828125,
0.2435302734375,
0.60791015625,
0.118408203125,
0.60791015625
],
[
0.5146484375,
0.5927734375,
0.880859375,
0.5927734375,
0.880859375,
0.60546875,
0.5146484375,
0.60546875
],
[
0.1351318359375,
0.60986328125,
0.486083984375,
0.609375,
0.486083984375,
0.62353515625,
0.1351318359375,
0.62451171875
],
[
0.51318359375,
0.60791015625,
0.880859375,
0.60791015625,
0.880859375,
0.62158203125,
0.51318359375,
0.62158203125
],
[
0.12005615234375,
0.626953125,
0.486083984375,
0.626953125,
0.486083984375,
0.64013671875,
0.12005615234375,
0.64013671875
],
[
0.51318359375,
0.62451171875,
0.8818359375,
0.62451171875,
0.8818359375,
0.63818359375,
0.51318359375,
0.63818359375
],
[
0.118408203125,
0.6435546875,
0.489501953125,
0.6435546875,
0.489501953125,
0.65673828125,
0.118408203125,
0.65673828125
],
[
0.509765625,
0.64013671875,
0.64990234375,
0.64013671875,
0.64990234375,
0.65380859375,
0.509765625,
0.65380859375
],
[
0.1175537109375,
0.6591796875,
0.484375,
0.6591796875,
0.484375,
0.6728515625,
0.1175537109375,
0.6728515625
],
[
0.118408203125,
0.6748046875,
0.484375,
0.6748046875,
0.484375,
0.6884765625,
0.118408203125,
0.6884765625
],
[
0.515625,
0.66943359375,
0.880859375,
0.67138671875,
0.880859375,
0.68603515625,
0.515625,
0.68408203125
],
[
0.1175537109375,
0.6904296875,
0.48779296875,
0.689453125,
0.48779296875,
0.70361328125,
0.1175537109375,
0.705078125
],
[
0.51318359375,
0.68701171875,
0.880859375,
0.68701171875,
0.880859375,
0.70068359375,
0.51318359375,
0.70068359375
],
[
0.12005615234375,
0.70703125,
0.1561279296875,
0.70703125,
0.1561279296875,
0.7197265625,
0.12005615234375,
0.7197265625
],
[
0.51318359375,
0.703125,
0.880859375,
0.70361328125,
0.880859375,
0.71728515625,
0.51318359375,
0.71630859375
],
[
0.51123046875,
0.71728515625,
0.880859375,
0.7197265625,
0.880859375,
0.73388671875,
0.51123046875,
0.732421875
],
[
0.12176513671875,
0.73388671875,
0.48779296875,
0.73388671875,
0.48779296875,
0.74755859375,
0.12176513671875,
0.74755859375
],
[
0.51123046875,
0.736328125,
0.87890625,
0.736328125,
0.87890625,
0.75,
0.51123046875,
0.75
],
[
0.1175537109375,
0.7490234375,
0.489501953125,
0.7490234375,
0.489501953125,
0.7626953125,
0.1175537109375,
0.7626953125
],
[
0.51318359375,
0.751953125,
0.880859375,
0.751953125,
0.880859375,
0.765625,
0.51318359375,
0.765625
],
[
0.1175537109375,
0.765625,
0.484375,
0.765625,
0.484375,
0.779296875,
0.1175537109375,
0.779296875
],
[
0.515625,
0.76953125,
0.587890625,
0.76953125,
0.587890625,
0.78076171875,
0.515625,
0.78076171875
],
[
0.1158447265625,
0.78076171875,
0.483642578125,
0.78076171875,
0.483642578125,
0.79443359375,
0.1158447265625,
0.79443359375
],
[
0.118408203125,
0.79736328125,
0.486083984375,
0.79736328125,
0.486083984375,
0.81103515625,
0.118408203125,
0.81103515625
],
[
0.51708984375,
0.79736328125,
0.880859375,
0.798828125,
0.880859375,
0.8134765625,
0.51708984375,
0.8125
],
[
0.118408203125,
0.8134765625,
0.48779296875,
0.8134765625,
0.48779296875,
0.82666015625,
0.118408203125,
0.82666015625
],
[
0.51123046875,
0.8154296875,
0.880859375,
0.8154296875,
0.880859375,
0.8291015625,
0.51123046875,
0.8291015625
],
[
0.1175537109375,
0.830078125,
0.486083984375,
0.830078125,
0.486083984375,
0.84326171875,
0.1175537109375,
0.84326171875
],
[
0.51318359375,
0.8330078125,
0.87890625,
0.8330078125,
0.87890625,
0.84326171875,
0.51318359375,
0.84326171875
],
[
0.118408203125,
0.84716796875,
0.484375,
0.84716796875,
0.484375,
0.85791015625,
0.118408203125,
0.85791015625
],
[
0.51123046875,
0.84716796875,
0.880859375,
0.84716796875,
0.880859375,
0.86083984375,
0.51123046875,
0.86083984375
],
[
0.1368408203125,
0.869140625,
0.484375,
0.87060546875,
0.484375,
0.88427734375,
0.1368408203125,
0.8828125
],
[
0.51318359375,
0.8642578125,
0.64404296875,
0.8642578125,
0.64404296875,
0.87744140625,
0.51318359375,
0.87744140625
],
[
0.118408203125,
0.8828125,
0.484375,
0.8818359375,
0.484375,
0.8935546875,
0.118408203125,
0.89453125
],
[
0.118408203125,
0.89453125,
0.486083984375,
0.89453125,
0.486083984375,
0.90771484375,
0.118408203125,
0.90771484375
],
[
0.5322265625,
0.89111328125,
0.880859375,
0.89453125,
0.880859375,
0.9091796875,
0.5322265625,
0.90625
],
[
0.1175537109375,
0.90625,
0.20068359375,
0.90625,
0.20068359375,
0.91748046875,
0.1175537109375,
0.91748046875
],
[
0.51318359375,
0.90625,
0.74072265625,
0.90625,
0.74072265625,
0.91943359375,
0.51318359375,
0.91943359375
]
] |
[
"作中，我们将我们的SentenceTransformer检索",
"Figure 2: Prompt for response generation.Each",
"器称为“密集”检索器",
"prompt consists of the retrieved demonstrations and the",
"target context. Each safety demonstration is separated",
"2)ResponseGeneration。一且选择了安全演",
"by an empty line and the target context is separated",
"示，我们将在上下文中使用它们来生成条件。",
"from the safety demonstrations by an empty line",
"具体来说，给定K安全演示和目标上下文，我",
"们使用Figure2中显示的提示格式。我们在输",
"工作与我们自已的工作最相关。虽然他们调",
"入中的每个对话前加上“两个人之间的对话”",
"查上下文学习以进行对齐，但他们不调查检索",
"以进行对话。演示根据其检索分数按降序放置",
"相关演示。最近的工作还研究了有关上下文学",
"在提示中。更简单地说，排名最高的演示位于",
"习的基本问题。Luetal.(2022b）调查了上下文",
"输入的开头。目标上下文放置在输人的末尾。",
"演示顺序对性能的影响。我们发现上下文演示",
"我们标记每个话语的说话者（人1或人2），并",
"的顺序不会影响响应质量或安全性。Liuetal.",
"在提示末尾为我们要为其生成响应的说话者提",
"(2022b）演示了基于与测试查询的语义相似性",
"供尾随注释（在Figure2中，这是人2）",
"检索上下文演示可提高NLU基准的性能。我",
"们发现检索与对话上下文高度相似的演示对于",
"4Experimental Setup",
"减少反应毒性很有用。最后，Rubinetal.(2022)",
"和Agrawaletal.（2023）研究了选择上下文演示",
"下面，我们描述了这项工作中使用的对话数据",
"的方法。我们还研究了为对话安全选择上下文",
"集。此外，我们还讨论了如何评估反应的安全",
"演示的不同方法。",
"性和相关性(i。e。、质量)",
"3Methodology",
"4.1Dialogue Datasets",
"我们研究了一种基于检索的方法，用于从仅",
"我们在这项工作中尝试了三个对话数据集。来",
"解码器的Transformer（Vaswanietal.，2017）模",
"自这些数据集的对话被用作响应生成的输入或",
"型中安全生成响应。具体而言，我们尝试使用",
"安全演示。在目标上下文和安全演示中，我们",
"不同尺寸的OPT(Zhangetal.，2022），LLaMA",
"最多使用两个对话回合。我们在下面描述每个",
"(Touvron et al., 2023) 和 Vicuna (Chiang et al.,",
"数据集，并将读者推迟到AppendixH以获取",
"2023）模型。我们主要使用OPT模型进行实",
"更多详细信息",
"验，因为模型代码和权重是公开可用的，但",
"是，我们还在整个工作中重点介绍了相关的",
"ProsocialDialog (Kim et al., 2022) :Prosocial-",
"LLaMA和Vicuna结果（有关完整结果，请参",
"Dialog包含带有亲社会响应的不安全话语。我",
"见AppendixE）",
"们使用来自ProsocialDialog培训部门的42K对",
"此后，我们将要生成响应的对话上下文称为",
"话作为我们所有实验的安全演示来源。我们还",
"目标上下文，将安全模型行为的演示称为安全",
"尝试从ProsocialDialog的验证拆分中生成对7",
"演示。在高层次上，我们的方法包括两个步骤：",
"K对话的响应。3",
"1）我们根据与目标上下文的相似性来检索安",
"全演示；2）我们在上下文中使用检索到的安",
"DiaSafety (Sun et al., 2022).DiaSafety 是对",
"全演示来生成条件。我们在下面详细描述这些",
"抗性话语的集合，可以从会话模型中非法获取",
"步骤",
"不安全的响应。我们尝试从DiaSafety的验证",
"集生成对1K对话的响应。我们使用DiaSafety",
"1)RetrievingSafetyDemonstrations.我们研",
"来评估对不安全输人的响应生成。我们注意到",
"究了三种为目标上下文选择安全演示的方法：",
"DiaSafety的每个目标上下文都由一个单一的",
"1）随机选择演示；2）使用BM25（Robertson",
"话语组成",
"andZaragoza,2009）选择演示；3）使用句子转",
"换器（Reimers andGurevych，2019）。2对于每",
"Commonsense-Dialogues (Zhou et al., 2021) .",
"个检索器，我们使用目标上下文作为查询来选",
"常识-对话是基于社会背景的对话的集合。我",
"择演示。这些安全演示是由不安全的话语和亲",
"们尝试从“常识对话”的验证集中生成对1K",
"社会反应组成的整个对话。。在我们的整个工",
"对话的响应。我们使用常识对话来评估对安全",
"2we also investigated using a Dense Passage Retriever",
"输入的响应生成",
"(DPR;Karpukhinetal.2020)to select demonstrations and",
"defer readers to Appendix D for results and additional re",
"3we consider a conversation turn to be an exchange be-",
"triever details",
"tween two speakers in a conversation."
] |
{
"arxiv_id": "2302.00871",
"language": "zh",
"page_idx": 3
}
|
||
2303.09752-zh-2
|
2303.09752-zh-2.jpg
|
<FigureHere>
Figure 2: CoLT5 achieves stronger performance than LoNGT5 at any speed. Average performance on all datasets as a function of inference and fine-tuning time per sample (ms) for LoNGT5 and CoLT5 Base, Large, and XL models. LONGT5 does not use MQA, but we report speed as though it had for a conservative baseline.
<table>
<tbody>
<tr>
<th>Encoder Layer Component</th>
<th>Flops</th>
</tr>
<tr>
<td>Vanilla self-attention computation</td>
<td>$n^{2}$ $2n^{:}$ $d$ 1</td>
</tr>
<tr>
<td>Attention QKV and output projections</td>
<td>$4nd^{2}$</td>
</tr>
<tr>
<td>Feedforward layer</td>
<td>$8nd^{2}$</td>
</tr>
<tr>
<td>LONGT5 local attention computation</td>
<td>$2nwd$</td>
</tr>
<tr>
<td>LONGT5 global attention computation</td>
<td>2 $m$ $d$ 8</td>
</tr>
<tr>
<td> </td>
<td>$\boldsymbol{U}$ 8</td>
</tr>
</tbody>
</table>
的另外两个修改。CoLT5 增加了多查询交叉注意力 (Shazeer, 2019) ,显著加快了推理速度。CoLT5 还采用了 UL2$\overline{\mathrm{E}}($Tay et al., 2022)预训练目标，我们证明了该目标允许在长输入上进行上下文学习
我们展示了 CoLT5 以相似或更好的模型质量执行更快的微调和推理，在 arXiv 摘要$\overline{\mathrm{E}}($Cohan et al., 2018) 和 TriviaQA 问题回答$\overline{\mathrm{E}}$(Joshi et al., 2017) 数据集上比 LoNGT5 有所改进，并在 SCROLLS 基准测试$\overline{\mathrm{E}}($Shaham et al., 2022) 上实现了 SOTA。此外，CoLT5 实现了具有极长输入(64k 令牌)的任务的质量和速度的进一步提升，“焦点”令牌的线性扩展不足
Table 1: Computational cost of encoder layer transformer components measured in FLOPs. $n$ is the input length, $d$ is the model dimensionality, and $w$ is the size of the local attention window.
## 2 $\textbf{Background}$
在本地窗口中，并且会出现在输入中每个 16个令牌块的均值汇总汇总表示中。LONGT5 的关注导致 FLOP 急剧减少(尽管仍然不可忽略)(表1)
Transformer FLOPs CoLT5 遵循广泛的工作路线，试图降低变压器模型的计算成本，特别是在长输入时。变压器模型的计算负担有几个不同的元素，不同的方法侧重于降低不同组件的成本。因此，从提供变压器组件计算成本的细分开始是有帮助的。表 1 显示了变压器编码器层 (Kaplan et al., 2020) 的每个组件的$FLOPs^{1}$
Sparse attention 将 Transformer 应用于长输人的第一个挑战是，自注意机制的 FLOP 在输人长度上呈二次方缩放，对于长输入变得棘手。大量的工作集中在减少自我注意力成本，将注意力限制在输入的子集$\overline{\mathrm{E}}$(Child et al., 2019; Ainslie et al., 2020; Beltagy et al., 2020; Zaheer et al., 2020; Wang et al., 2020; Guo et al., 2022) 之间或限制在层的子集$\overline{\mathrm{E}}$(Zemlyanskiy et al., 2021)之间。在与 CoLT5 最密切相关的模型LoNGT5 $\boxtimes$E(Guo et al.,2022)中，令牌会出现
Conditional computation 应用稀疏注意机制后，前馈和注意投影层占 FLOP 的大部分。这些成本与输入的长度成比例，因此处理长输入仍然非常昂贵。减少剩余成本的一种常见方法是采用某种形式的条件计算，避免将所有模型参数应用于整个输入。CALM (Schuster et al., 2022) 对每个解码令牌应用不同数量的解码器层，如果模型对其预测有信心，则提前输出令牌。专家混合模型 (Shazeer et al.,2017; Fedus et al., 2021; Zoph et al., 2022) 通过一小部分专家子模块路由输入，仅带来与输入最相关的参数。在检索增强模型的上下文中，许多作品根据检索到的段落与查询的相关性对检索到的段落进行重新排序，并仅处理得分最高的段落$\overline{\mathrm{E}}($Mao et al., 2021; Wang et al., 2018; Yu et al., 2022), 并根据模型置信度$\overline{\mathbb{E}}($Kratzwald and Feuerriegel, 2018; Varshney et al., 2022) 改
$^{1}$Each multiply-add is counted as a single FLOP
|
[
[
0.5810546875,
0.0872802734375,
0.67333984375,
0.09027099609375,
0.67236328125,
0.105712890625,
0.580078125,
0.10272216796875
],
[
0.1956787109375,
0.09320068359375,
0.268798828125,
0.09320068359375,
0.268798828125,
0.1051025390625,
0.1956787109375,
0.1051025390625
],
[
0.313232421875,
0.09442138671875,
0.31982421875,
0.09442138671875,
0.31982421875,
0.1009521484375,
0.313232421875,
0.1009521484375
],
[
0.310546875,
0.10394287109375,
0.32666015625,
0.10394287109375,
0.32666015625,
0.1134033203125,
0.310546875,
0.1134033203125
],
[
0.75390625,
0.1009521484375,
0.7734375,
0.1009521484375,
0.7734375,
0.10986328125,
0.75390625,
0.10986328125
],
[
0.5400390625,
0.1175537109375,
0.56494140625,
0.1175537109375,
0.56494140625,
0.1312255859375,
0.5400390625,
0.1312255859375
],
[
0.11419677734375,
0.105712890625,
0.1956787109375,
0.105712890625,
0.1956787109375,
0.196533203125,
0.11419677734375,
0.196533203125
],
[
0.43994140625,
0.12054443359375,
0.459228515625,
0.12054443359375,
0.459228515625,
0.1300048828125,
0.43994140625,
0.1300048828125
],
[
0.8310546875,
0.119384765625,
0.8369140625,
0.119384765625,
0.8369140625,
0.12469482421875,
0.8310546875,
0.12469482421875
],
[
0.156982421875,
0.163330078125,
0.1796875,
0.163330078125,
0.1796875,
0.177001953125,
0.156982421875,
0.177001953125
],
[
0.5400390625,
0.1602783203125,
0.5634765625,
0.1602783203125,
0.5634765625,
0.17333984375,
0.5400390625,
0.17333984375
],
[
0.2301025390625,
0.190673828125,
0.2359619140625,
0.190673828125,
0.2359619140625,
0.19775390625,
0.2301025390625,
0.19775390625
],
[
0.15869140625,
0.2060546875,
0.1796875,
0.2060546875,
0.1796875,
0.218505859375,
0.15869140625,
0.218505859375
],
[
0.54248046875,
0.204833984375,
0.5615234375,
0.204833984375,
0.5615234375,
0.21435546875,
0.54248046875,
0.21435546875
],
[
0.35107421875,
0.2161865234375,
0.385498046875,
0.2161865234375,
0.385498046875,
0.232177734375,
0.35107421875,
0.232177734375
],
[
0.4130859375,
0.2161865234375,
0.447509765625,
0.2161865234375,
0.447509765625,
0.232177734375,
0.4130859375,
0.232177734375
],
[
0.57080078125,
0.21728515625,
0.58447265625,
0.21728515625,
0.58447265625,
0.226806640625,
0.57080078125,
0.226806640625
],
[
0.60791015625,
0.215576171875,
0.8447265625,
0.215576171875,
0.8447265625,
0.2286376953125,
0.60791015625,
0.2286376953125
],
[
0.28955078125,
0.218505859375,
0.32080078125,
0.218505859375,
0.32080078125,
0.23095703125,
0.28955078125,
0.23095703125
],
[
0.2275390625,
0.218505859375,
0.2568359375,
0.218505859375,
0.2568359375,
0.23095703125,
0.2275390625,
0.23095703125
],
[
0.2476806640625,
0.2340087890625,
0.411376953125,
0.235107421875,
0.411376953125,
0.2493896484375,
0.2476806640625,
0.248779296875
],
[
0.62890625,
0.2340087890625,
0.79443359375,
0.235107421875,
0.79443359375,
0.2493896484375,
0.62890625,
0.248779296875
],
[
0.4072265625,
0.257080078125,
0.48779296875,
0.257080078125,
0.48779296875,
0.271484375,
0.4072265625,
0.271484375
],
[
0.52880859375,
0.260009765625,
0.54248046875,
0.260009765625,
0.54248046875,
0.268310546875,
0.52880859375,
0.268310546875
],
[
0.5859375,
0.257080078125,
0.65576171875,
0.257080078125,
0.65576171875,
0.2724609375,
0.5859375,
0.2724609375
],
[
0.1175537109375,
0.287353515625,
0.4072265625,
0.2880859375,
0.4072265625,
0.2998046875,
0.1175537109375,
0.298583984375
],
[
0.41064453125,
0.287353515625,
0.8818359375,
0.287353515625,
0.8818359375,
0.300537109375,
0.41064453125,
0.300537109375
],
[
0.1175537109375,
0.300537109375,
0.880859375,
0.300537109375,
0.880859375,
0.3154296875,
0.1175537109375,
0.3154296875
],
[
0.1175537109375,
0.317626953125,
0.80615234375,
0.317626953125,
0.80615234375,
0.327880859375,
0.1175537109375,
0.327880859375
],
[
0.1175537109375,
0.355224609375,
0.483642578125,
0.355224609375,
0.483642578125,
0.3681640625,
0.1175537109375,
0.3681640625
],
[
0.5205078125,
0.358154296875,
0.724609375,
0.359375,
0.724609375,
0.37353515625,
0.5205078125,
0.372314453125
],
[
0.8251953125,
0.35693359375,
0.875,
0.359375,
0.873046875,
0.374755859375,
0.82373046875,
0.372314453125
],
[
0.118408203125,
0.371826171875,
0.48779296875,
0.371826171875,
0.48779296875,
0.385986328125,
0.118408203125,
0.385986328125
],
[
0.52294921875,
0.3818359375,
0.77490234375,
0.3818359375,
0.77490234375,
0.3955078125,
0.52294921875,
0.3955078125
],
[
0.8271484375,
0.380126953125,
0.8759765625,
0.380126953125,
0.8759765625,
0.39794921875,
0.8271484375,
0.39794921875
],
[
0.118408203125,
0.38720703125,
0.486083984375,
0.38720703125,
0.486083984375,
0.402099609375,
0.118408203125,
0.402099609375
],
[
0.52294921875,
0.3984375,
0.80615234375,
0.3984375,
0.80615234375,
0.412109375,
0.52294921875,
0.412109375
],
[
0.8251953125,
0.396728515625,
0.86962890625,
0.394287109375,
0.87158203125,
0.410888671875,
0.8271484375,
0.413330078125
],
[
0.118408203125,
0.40380859375,
0.486083984375,
0.40380859375,
0.486083984375,
0.418701171875,
0.118408203125,
0.418701171875
],
[
0.5205078125,
0.413330078125,
0.66015625,
0.41455078125,
0.66015625,
0.4287109375,
0.5205078125,
0.427490234375
],
[
0.8251953125,
0.413330078125,
0.86962890625,
0.410888671875,
0.87158203125,
0.427490234375,
0.8271484375,
0.429931640625
],
[
0.12005615234375,
0.421630859375,
0.2275390625,
0.421630859375,
0.2275390625,
0.432861328125,
0.12005615234375,
0.432861328125
],
[
0.5224609375,
0.43115234375,
0.80029296875,
0.43115234375,
0.80029296875,
0.44482421875,
0.5224609375,
0.44482421875
],
[
0.8251953125,
0.429931640625,
0.8759765625,
0.429931640625,
0.8759765625,
0.44482421875,
0.8251953125,
0.44482421875
],
[
0.1368408203125,
0.435791015625,
0.486083984375,
0.435791015625,
0.486083984375,
0.45068359375,
0.1368408203125,
0.45068359375
],
[
0.5205078125,
0.446533203125,
0.81005859375,
0.44775390625,
0.81005859375,
0.4619140625,
0.5205078125,
0.46142578125
],
[
0.8310546875,
0.443603515625,
0.86962890625,
0.443603515625,
0.86962890625,
0.465576171875,
0.8310546875,
0.465576171875
],
[
0.118408203125,
0.453125,
0.48779296875,
0.453125,
0.48779296875,
0.46728515625,
0.118408203125,
0.46728515625
],
[
0.1175537109375,
0.468505859375,
0.486083984375,
0.468505859375,
0.486083984375,
0.482177734375,
0.1175537109375,
0.482177734375
],
[
0.51123046875,
0.476806640625,
0.8818359375,
0.478515625,
0.8818359375,
0.493408203125,
0.51123046875,
0.4912109375
],
[
0.1175537109375,
0.485107421875,
0.484375,
0.485107421875,
0.484375,
0.498779296875,
0.1175537109375,
0.498779296875
],
[
0.51123046875,
0.4921875,
0.8818359375,
0.4912109375,
0.8818359375,
0.505859375,
0.51123046875,
0.50732421875
],
[
0.118408203125,
0.50048828125,
0.484375,
0.50048828125,
0.484375,
0.51416015625,
0.118408203125,
0.51416015625
],
[
0.51123046875,
0.50732421875,
0.8818359375,
0.50732421875,
0.8818359375,
0.52099609375,
0.51123046875,
0.52099609375
],
[
0.12005615234375,
0.51708984375,
0.484375,
0.51708984375,
0.484375,
0.53076171875,
0.12005615234375,
0.53076171875
],
[
0.509765625,
0.521484375,
0.71533203125,
0.521484375,
0.71533203125,
0.53515625,
0.509765625,
0.53515625
],
[
0.1175537109375,
0.5322265625,
0.484375,
0.5322265625,
0.484375,
0.5458984375,
0.1175537109375,
0.5458984375
],
[
0.1259765625,
0.5498046875,
0.438232421875,
0.5498046875,
0.438232421875,
0.5634765625,
0.1259765625,
0.5634765625
],
[
0.5146484375,
0.56640625,
0.8818359375,
0.56640625,
0.8818359375,
0.580078125,
0.5146484375,
0.580078125
],
[
0.118408203125,
0.57763671875,
0.133544921875,
0.57763671875,
0.133544921875,
0.58837890625,
0.118408203125,
0.58837890625
],
[
0.1502685546875,
0.576171875,
0.26025390625,
0.576171875,
0.26025390625,
0.59033203125,
0.1502685546875,
0.59033203125
],
[
0.51123046875,
0.5791015625,
0.8818359375,
0.580078125,
0.8818359375,
0.59716796875,
0.51123046875,
0.5966796875
],
[
0.124267578125,
0.6015625,
0.484375,
0.6015625,
0.484375,
0.615234375,
0.124267578125,
0.615234375
],
[
0.5146484375,
0.5986328125,
0.880859375,
0.5986328125,
0.880859375,
0.6123046875,
0.5146484375,
0.6123046875
],
[
0.118408203125,
0.6162109375,
0.483642578125,
0.6162109375,
0.483642578125,
0.6298828125,
0.118408203125,
0.6298828125
],
[
0.51123046875,
0.6142578125,
0.60693359375,
0.6142578125,
0.60693359375,
0.62890625,
0.51123046875,
0.62890625
],
[
0.12005615234375,
0.63427734375,
0.483642578125,
0.63427734375,
0.483642578125,
0.64794921875,
0.12005615234375,
0.64794921875
],
[
0.118408203125,
0.6494140625,
0.484375,
0.6494140625,
0.484375,
0.6630859375,
0.118408203125,
0.6630859375
],
[
0.51708984375,
0.646484375,
0.880859375,
0.646484375,
0.880859375,
0.6611328125,
0.51708984375,
0.6611328125
],
[
0.118408203125,
0.6650390625,
0.486083984375,
0.6650390625,
0.486083984375,
0.6787109375,
0.118408203125,
0.6787109375
],
[
0.51123046875,
0.662109375,
0.880859375,
0.6630859375,
0.880859375,
0.677734375,
0.51123046875,
0.6767578125
],
[
0.118408203125,
0.681640625,
0.486083984375,
0.681640625,
0.486083984375,
0.6953125,
0.118408203125,
0.6953125
],
[
0.51123046875,
0.677734375,
0.880859375,
0.6787109375,
0.880859375,
0.693359375,
0.51123046875,
0.6923828125
],
[
0.1175537109375,
0.69677734375,
0.483642578125,
0.69677734375,
0.483642578125,
0.71044921875,
0.1175537109375,
0.71044921875
],
[
0.5146484375,
0.69482421875,
0.880859375,
0.69482421875,
0.880859375,
0.708984375,
0.5146484375,
0.708984375
],
[
0.11419677734375,
0.71337890625,
0.1822509765625,
0.71142578125,
0.1822509765625,
0.7255859375,
0.11419677734375,
0.72802734375
],
[
0.515625,
0.71142578125,
0.87890625,
0.71142578125,
0.87890625,
0.724609375,
0.515625,
0.724609375
],
[
0.5146484375,
0.72705078125,
0.87744140625,
0.72705078125,
0.87744140625,
0.74072265625,
0.5146484375,
0.74072265625
],
[
0.1234130859375,
0.73828125,
0.484375,
0.73828125,
0.484375,
0.751953125,
0.1234130859375,
0.751953125
],
[
0.51318359375,
0.74365234375,
0.880859375,
0.74365234375,
0.880859375,
0.75732421875,
0.51318359375,
0.75732421875
],
[
0.1175537109375,
0.7529296875,
0.486083984375,
0.75439453125,
0.486083984375,
0.7685546875,
0.1175537109375,
0.76708984375
],
[
0.5146484375,
0.7587890625,
0.87890625,
0.7587890625,
0.87890625,
0.77392578125,
0.5146484375,
0.77392578125
],
[
0.12005615234375,
0.76953125,
0.486083984375,
0.76953125,
0.486083984375,
0.78369140625,
0.12005615234375,
0.78369140625
],
[
0.51318359375,
0.775390625,
0.880859375,
0.775390625,
0.880859375,
0.7890625,
0.51318359375,
0.7890625
],
[
0.1175537109375,
0.78515625,
0.486083984375,
0.7861328125,
0.486083984375,
0.80029296875,
0.1175537109375,
0.7998046875
],
[
0.51123046875,
0.79052734375,
0.87890625,
0.79052734375,
0.87890625,
0.80419921875,
0.51123046875,
0.80419921875
],
[
0.1175537109375,
0.802734375,
0.484375,
0.802734375,
0.484375,
0.81640625,
0.1175537109375,
0.81640625
],
[
0.5146484375,
0.80810546875,
0.880859375,
0.80810546875,
0.880859375,
0.82177734375,
0.5146484375,
0.82177734375
],
[
0.1175537109375,
0.818359375,
0.484375,
0.818359375,
0.484375,
0.83203125,
0.1175537109375,
0.83203125
],
[
0.51318359375,
0.822265625,
0.880859375,
0.82373046875,
0.880859375,
0.83837890625,
0.51318359375,
0.83740234375
],
[
0.1158447265625,
0.8349609375,
0.483642578125,
0.8349609375,
0.483642578125,
0.8486328125,
0.1158447265625,
0.8486328125
],
[
0.51318359375,
0.84033203125,
0.880859375,
0.84033203125,
0.880859375,
0.85400390625,
0.51318359375,
0.85400390625
],
[
0.118408203125,
0.85107421875,
0.483642578125,
0.85107421875,
0.483642578125,
0.8642578125,
0.118408203125,
0.8642578125
],
[
0.51318359375,
0.85693359375,
0.880859375,
0.85693359375,
0.880859375,
0.87060546875,
0.51318359375,
0.87060546875
],
[
0.1158447265625,
0.865234375,
0.484375,
0.8662109375,
0.484375,
0.88134765625,
0.1158447265625,
0.8798828125
],
[
0.51318359375,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.1175537109375,
0.8828125,
0.484375,
0.8828125,
0.484375,
0.896484375,
0.1175537109375,
0.896484375
],
[
0.51123046875,
0.88818359375,
0.880859375,
0.88818359375,
0.880859375,
0.90185546875,
0.51123046875,
0.90185546875
],
[
0.14111328125,
0.90478515625,
0.432373046875,
0.90478515625,
0.432373046875,
0.91845703125,
0.14111328125,
0.91845703125
],
[
0.51123046875,
0.90380859375,
0.8818359375,
0.90380859375,
0.8818359375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"Fine-tuning",
"Inference",
"伽",
"XI",
"XI",
"46",
"",
"XI",
"T",
"44",
"44",
"a",
"42",
"42",
"600",
"800",
"0",
"5001,0001,5002.0002,500",
"400",
"200",
"Time per sample (ms)",
"Time per sample (ms)",
"LONGT5",
"量",
"CoLT5",
"Figure2:CoLT5achieves strongerperforn",
"nancethan LoNGT5at any speed.Averageperformanceon all datasets",
"as a function of inference and fine-tuning time per sample (ms) for LoNGT5 and CoLT5 Base, Large, and XL",
"models.LoNGT5 does not use MQA, but we report speed as though it had fora conservative baseline.",
"的另外两个修改。CoLT5增加了多查询交叉",
"Encoder Layer Component",
"Flops",
"注意力（Shazeer,2019），显著加快了推理速度。",
"Vanilla self-attention computation",
"2n?d",
"CoLT5 还采用了 UL2E(Tay et al.,2022)预训",
"Attention QKV and output projections",
"4nd2",
"练目标，我们证明了该目标允许在长输入上进",
"Feedforward layer",
"8nd2",
"行上下文学习",
"LONGT5 local attention computation",
"2nwd",
"我们展示了CoLT5以相似或更好的模型质",
"LONGT5 global attention computation",
"d",
"量执行更快的微调和推理，在arXiv摘要E(Co-",
"han etal.,2018）和TriviaQA问题回答Joshi",
"Table 1: Computational cost of encoder layer trans-",
"etal.,2017)数据集上比LONGT5有所改进，并",
"former components measured in FLOPs. n is the input",
"在SCROLLS基准测试E(Shahametal,2022)",
"length, d is the model dimensionality, and w is the size",
"上实现了SOTA。此外，CoLT5实现了具有极",
" of the local attention window.",
"长输入（64k令牌）的任务的质量和速度的进",
"一步提升，“焦点”令牌的线性扩展不足",
"在本地窗口中，并且会出现在输入中每个16",
"2",
"Background",
"个令牌块的均值汇总汇总表示中。LONGT5的",
"TransformerFLOPsCoLT5遵循广泛的工",
"关注导致FLOP急剧减少（尽管仍然不可忽",
"作路线，试图降低变压器模型的计算成本，特",
"略）(表1)",
"别是在长输入时。变压器模型的计算负担有",
"几个不同的元素，不同的方法侧重于降低不同",
"Conditionalcomputation应用稀疏注意机制",
"组件的成本。因此，从提供变压器组件计算成",
"后，前馈和注意投影层占FLOP的大部分。这",
"本的细分开始是有帮助的。表1显示了变压",
"些成本与输入的长度成比例，因此处理长输入",
"器编码器层(Kaplanetal.,2020)的每个组件的",
"仍然非常昂贵。减少剩余成本的一种常见方法",
"FLOPs1",
"是采用某种形式的条件计算，避免将所有模型",
"参数应用于整个输入。CALM（Schusteretal.",
"Sparseattention将Transformer应用于长输",
"2022)对每个解码令牌应用不同数量的解码器",
"人的第一个挑战是，自注意机制的FLOP在输",
"层，如果模型对其预测有信心，则提前输出令",
"入长度上呈二次方缩放，对于长输人变得棘手。",
"牌。专家混合模型（Shazeeretal.，2017；Fedus",
"大量的工作集中在减少自我注意力成本，将",
"etal.,2021:Zophetal.,2022）通过一小部分专",
"注意力限制在输入的子集E(Childetal.,2019",
"家子模块路由输入，仅带来与输人最相关的",
"Ainslie et al., 2020; Beltagy et al., 2020; Zaheer",
"参数。在检索增强模型的上下文中，许多作",
"et al., 2020; Wang et al., 2020; Guo et al., 2022)",
"品根据检索到的段落与查询的相关性对检索",
"之间或限制在层的子集E(Zemlyanskiyetal.",
"到的段落进行重新排序，并仅处理得分最高",
"2021)之间。。在与CoLT5最密切相关的模型",
"的段落E(Mao et al.,2021;Wang et al.,2018;Yu",
"LoNGT5E(Guoetal.,2022）中，令牌会出现",
"etal.，2022），并根据模型置信度E(Kratzwald",
"'Each multiply-add is counted as a single FLOP.",
"and Feuerriegel, 2018; Varshney et al., 2022) 改"
] |
{
"arxiv_id": "2303.09752",
"language": "zh",
"page_idx": 2
}
|
||
2303.09752-zh-3
|
2303.09752-zh-3.jpg
|
变处理的段落数。并发工作 CoDA$\underline{\mathrm{E}}($Lei et al., 2023) 采用了相关的条件计算机制，旨在有效地适应而不是对长文档进行建模
Device utilization FLOP 并不能说明全部情况，因为建模选择会影响加速器实现的有效运行速度。对于长文本输入，由于重复加载键和值的长序列的内存带宽限制，自回归解码器推理非常慢 (Shazeer, 2019; de Jong et al., 2022) 。Shazeer (2019)引入了多查询注意(MQA), 共享键和值的头部以减少内存带宽开销。Pope et al. (2022) 研究了如何对大型模型进行分片， 尤其是在 MQA 的情况下，以获得最佳的设备利用率，从而获得最佳的速度
Training objectives T5 引人了 span 损坏目标$\overline{\mathrm{E}}$(Raffel et al.,2020) , 这是掩码语言建模$\overline{\mathrm{E}}$ (Devlin et al., 2019) 的修改。LONGT5 利用 PEGASUS (Zhang et al., 2020) 句子重构目标来提高摘要性能。Tay et al.(2022) 提出了 UL2, 这是 span 损坏，前缀和因果语言建模的混合物， 并表明它在短输出和生成任务上都具有出色的性能
# 3 $\mathbf{CoLT5}$
3.1 Conditional computation
如上一节所述，很大一部分 TransformerFLOP 来自前馈和投影层，这些层随输入序列的长度而缩放。因此，LONGT5 在长文档上的训练和推理仍然很昂贵
CoLT5 进一步降低了通过条件计算处理长文档的成本，遵循一些令牌更重要的直觉，因此从繁重的计算中受益更多。首先，某些类型的令牌可能固有地需要较少的计算，例如填充词和标点符号。其次，尤其是在长文档中，输人的大部分可能与当前问题，任务或处理阶段无关
CoLT5 条件计算机制由三个组件组成：路由模块，条件前馈层和条件关注层。所有令牌都由标准，轻量级的关注和前馈层处理。路由模块还从每个关注或前馈层的输入中选择重要的令牌，并且一个繁重的条件层将额外的计算应用于路由令牌。本节详细介绍每个组件。图 1提供了 CoLT5 条件计算机制的概述，表2 比较了 CoLT5 和 LONGT5 FLOP
$\textbf{Routing}$ 为了在每个层中分别为每个组件选择重要的令牌，我们需要一个可学习且易于处理的路由功能。我们遵循 Lei et al.(2023) 的简单三步机制：(1)将输人与学习的嵌入相乘以获得路由分数，(2)归一化，以及(3)选择最高得分的$k$输入
<table>
<tbody>
<tr>
<th>Model</th>
<th>Encoder Layer Flops</th>
</tr>
<tr>
<td>$T5$</td>
<td>$+2n^{2}d$ $12nd^{2}$</td>
</tr>
<tr>
<td>LONGT5</td>
<td>$12nd^{2}$ $\eta_{i}$ 2 $d$ +</td>
</tr>
<tr>
<td>CoLT5</td>
<td>$7\frac{1}{4}nd^{2}$ $\eta_{i}$ $d$ + $\overline{84}$</td>
</tr>
</tbody>
</table>
Table 2: COLT5 uses significantly fewer FLOPs than LONGT5. Comparison of approximate encoder layer total FLOPs between T5 , LONGT5, and CoLT5. CoLT5 FLOPs rounded to readable fractions.
令$X_i$为令牌$i$的表示形式，而$u$为$d$维的
可学习嵌入。那么令牌$i$的路由分数是
$$s_i=X_i\cdot u$$
我们选择顶部- $k$得分最高的输人。为了向评分嵌入提供学习信号，我们确保路由令牌对层更新的贡献根据路由分数进行缩放，稍后将看到。为了向所有令牌提供更好的分布式信号， 我们还使用广义的 softmax 对路由分数进行全局归一化，以汇总到所需的路由令牌数，从而得出归一化分数$\tilde{s}_i$ 。每个 CoLT5 层都有三个独立的路由器，每个路由器用于前馈层，注意查询和注意键值
$\textbf{Conditional Feedforward}$ 直观上，一些令牌表示可能比其他令牌表示受益于更多的处理。CoLT5 条件前馈层将额外的高容量前馈层应用于所选令牌。特别是，让$X_i$是第$i$个令牌的模型状态，$\tilde{s}_i$表示标准化的路由分数(对于非路由令牌，设置为0)。然后，CoLT5 的前馈更新由
$$X_i=X_i+\mathrm{FFd}_{\mathrm{Light}}(X_i)+\tilde{s}_i\cdot\mathrm{FFd}_{\mathrm{Heavy}}(X_i)$$
给出。轻前馈分支和重前馈分支仅在隐藏维度上有所不同，轻分支的隐藏维度小于标准 T5前馈层，重分支更大。令$n$表示输入令牌的数量，$m$表示所选令牌的数量，$r_L$和$r_H$表示轻型和重型隐藏尺寸与标准 T5 隐藏尺寸的比率。然后，CoLT5 层的 FLOP 由
$$\text{FLOPs}_{\text{FFd}}=\underbrace{8nr_Ld^2}_{\text{Light branch}}+\underbrace{8mr_Hd^2}_{\text{Heavy branch}}$$
给出。我们将轻和重比设置为$r_L=\frac{1}{2}$和$r_H=$ 4 ,分别是标准 T5 隐藏尺寸的一半和四倍。对于我们的主要实验，令牌的一小部分$\frac1{16}$被路由到重分支。结果，来自 CoLT5 前馈层的近似 FLOP 等于
$$\text{FLOPs}_{\text{FFd}}=\underbrace{4nd^2}_{\text{Light branch}}+\underbrace{2nd^2}_{\text{Heavy branch}}$$
消耗了标准 T5 前馈层的 FLOP 的 75 个%。
|
[
[
0.12005615234375,
0.08612060546875,
0.483642578125,
0.08612060546875,
0.483642578125,
0.09979248046875,
0.12005615234375,
0.09979248046875
],
[
0.57421875,
0.09027099609375,
0.623046875,
0.09027099609375,
0.623046875,
0.1015625,
0.57421875,
0.1015625
],
[
0.6630859375,
0.0885009765625,
0.82275390625,
0.09027099609375,
0.82275390625,
0.1051025390625,
0.6630859375,
0.10272216796875
],
[
0.1158447265625,
0.1009521484375,
0.486083984375,
0.1015625,
0.486083984375,
0.11639404296875,
0.1158447265625,
0.11517333984375
],
[
0.57080078125,
0.10986328125,
0.60107421875,
0.10986328125,
0.60107421875,
0.1270751953125,
0.57080078125,
0.1270751953125
],
[
0.6650390625,
0.10986328125,
0.7734375,
0.10986328125,
0.7734375,
0.12469482421875,
0.6650390625,
0.12469482421875
],
[
0.1175537109375,
0.1175537109375,
0.376220703125,
0.1181640625,
0.376220703125,
0.1318359375,
0.1175537109375,
0.1312255859375
],
[
0.57421875,
0.1312255859375,
0.6630859375,
0.1312255859375,
0.6630859375,
0.142578125,
0.57421875,
0.142578125
],
[
0.66015625,
0.1270751953125,
0.765625,
0.1270751953125,
0.765625,
0.144287109375,
0.66015625,
0.144287109375
],
[
0.1234130859375,
0.1455078125,
0.484375,
0.1455078125,
0.484375,
0.1591796875,
0.1234130859375,
0.1591796875
],
[
0.5751953125,
0.1484375,
0.6337890625,
0.1484375,
0.6337890625,
0.1602783203125,
0.5751953125,
0.1602783203125
],
[
0.6591796875,
0.1436767578125,
0.767578125,
0.1455078125,
0.767578125,
0.1644287109375,
0.6591796875,
0.162109375
],
[
0.118408203125,
0.162109375,
0.48193359375,
0.162109375,
0.48193359375,
0.1728515625,
0.118408203125,
0.1728515625
],
[
0.118408203125,
0.1776123046875,
0.484375,
0.1776123046875,
0.484375,
0.191162109375,
0.118408203125,
0.191162109375
],
[
0.51123046875,
0.1787109375,
0.87890625,
0.1787109375,
0.87890625,
0.1923828125,
0.51123046875,
0.1923828125
],
[
0.118408203125,
0.193603515625,
0.48779296875,
0.193603515625,
0.48779296875,
0.207275390625,
0.118408203125,
0.207275390625
],
[
0.51318359375,
0.193603515625,
0.87890625,
0.193603515625,
0.87890625,
0.207275390625,
0.51318359375,
0.207275390625
],
[
0.1175537109375,
0.210205078125,
0.48779296875,
0.210205078125,
0.48779296875,
0.223876953125,
0.1175537109375,
0.223876953125
],
[
0.51123046875,
0.207275390625,
0.87890625,
0.207275390625,
0.87890625,
0.2203369140625,
0.51123046875,
0.2203369140625
],
[
0.1318359375,
0.2257080078125,
0.48779296875,
0.2257080078125,
0.48779296875,
0.2392578125,
0.1318359375,
0.2392578125
],
[
0.51123046875,
0.2203369140625,
0.8193359375,
0.221435546875,
0.8193359375,
0.235107421875,
0.51123046875,
0.2340087890625
],
[
0.1234130859375,
0.23095703125,
0.1368408203125,
0.23095703125,
0.1368408203125,
0.235107421875,
0.1234130859375,
0.235107421875
],
[
0.1175537109375,
0.240478515625,
0.484375,
0.2410888671875,
0.484375,
0.255859375,
0.1175537109375,
0.254638671875
],
[
0.118408203125,
0.25830078125,
0.48779296875,
0.25830078125,
0.48779296875,
0.271484375,
0.118408203125,
0.271484375
],
[
0.5322265625,
0.2587890625,
0.880859375,
0.2587890625,
0.880859375,
0.2724609375,
0.5322265625,
0.2724609375
],
[
0.1175537109375,
0.2724609375,
0.48779296875,
0.2724609375,
0.48779296875,
0.2861328125,
0.1175537109375,
0.2861328125
],
[
0.490234375,
0.276611328125,
0.515625,
0.276611328125,
0.515625,
0.283935546875,
0.490234375,
0.283935546875
],
[
0.51123046875,
0.273681640625,
0.82275390625,
0.27490234375,
0.82275390625,
0.2880859375,
0.51123046875,
0.287353515625
],
[
0.118408203125,
0.290283203125,
0.35693359375,
0.290283203125,
0.35693359375,
0.303955078125,
0.118408203125,
0.303955078125
],
[
0.6513671875,
0.2998046875,
0.74072265625,
0.297607421875,
0.74072265625,
0.3134765625,
0.6513671875,
0.3154296875
],
[
0.1234130859375,
0.3154296875,
0.48193359375,
0.3154296875,
0.48193359375,
0.328857421875,
0.1234130859375,
0.328857421875
],
[
0.51318359375,
0.325927734375,
0.87890625,
0.325927734375,
0.87890625,
0.339111328125,
0.51318359375,
0.339111328125
],
[
0.118408203125,
0.33203125,
0.486083984375,
0.33203125,
0.486083984375,
0.345703125,
0.118408203125,
0.345703125
],
[
0.51318359375,
0.341552734375,
0.880859375,
0.341552734375,
0.880859375,
0.355224609375,
0.51318359375,
0.355224609375
],
[
0.118408203125,
0.349853515625,
0.486083984375,
0.349853515625,
0.486083984375,
0.35986328125,
0.118408203125,
0.35986328125
],
[
0.51318359375,
0.358154296875,
0.87890625,
0.358154296875,
0.87890625,
0.371826171875,
0.51318359375,
0.371826171875
],
[
0.118408203125,
0.365234375,
0.486083984375,
0.365234375,
0.486083984375,
0.37890625,
0.118408203125,
0.37890625
],
[
0.509765625,
0.372314453125,
0.87890625,
0.372314453125,
0.87890625,
0.385986328125,
0.509765625,
0.385986328125
],
[
0.1175537109375,
0.38134765625,
0.48779296875,
0.380126953125,
0.48779296875,
0.394287109375,
0.1175537109375,
0.3955078125
],
[
0.51123046875,
0.388427734375,
0.87890625,
0.388427734375,
0.87890625,
0.402099609375,
0.51123046875,
0.402099609375
],
[
0.118408203125,
0.39794921875,
0.48779296875,
0.39794921875,
0.48779296875,
0.410888671875,
0.118408203125,
0.410888671875
],
[
0.51318359375,
0.40625,
0.880859375,
0.40625,
0.880859375,
0.419189453125,
0.51318359375,
0.419189453125
],
[
0.1175537109375,
0.413330078125,
0.484375,
0.413330078125,
0.484375,
0.427001953125,
0.1175537109375,
0.427001953125
],
[
0.51318359375,
0.421630859375,
0.87890625,
0.421630859375,
0.87890625,
0.435302734375,
0.51318359375,
0.435302734375
],
[
0.1175537109375,
0.4287109375,
0.156982421875,
0.4287109375,
0.156982421875,
0.44482421875,
0.1175537109375,
0.44482421875
],
[
0.5146484375,
0.438232421875,
0.880859375,
0.438232421875,
0.880859375,
0.451904296875,
0.5146484375,
0.451904296875
],
[
0.11419677734375,
0.457763671875,
0.2216796875,
0.457763671875,
0.2216796875,
0.47265625,
0.11419677734375,
0.47265625
],
[
0.51123046875,
0.453125,
0.64404296875,
0.45361328125,
0.64404296875,
0.468505859375,
0.51123046875,
0.46728515625
],
[
0.51708984375,
0.47802734375,
0.8818359375,
0.47802734375,
0.8818359375,
0.4912109375,
0.51708984375,
0.4912109375
],
[
0.11419677734375,
0.48388671875,
0.361083984375,
0.485107421875,
0.361083984375,
0.49951171875,
0.11419677734375,
0.498779296875
],
[
0.51318359375,
0.493408203125,
0.8818359375,
0.4921875,
0.8818359375,
0.50732421875,
0.51318359375,
0.50830078125
],
[
0.118408203125,
0.50732421875,
0.48779296875,
0.505859375,
0.48779296875,
0.51953125,
0.118408203125,
0.52099609375
],
[
0.51318359375,
0.51025390625,
0.880859375,
0.51025390625,
0.880859375,
0.52392578125,
0.51318359375,
0.52392578125
],
[
0.1158447265625,
0.5224609375,
0.48779296875,
0.521484375,
0.48779296875,
0.5361328125,
0.1158447265625,
0.53759765625
],
[
0.51318359375,
0.52685546875,
0.880859375,
0.52685546875,
0.880859375,
0.54052734375,
0.51318359375,
0.54052734375
],
[
0.118408203125,
0.5380859375,
0.486083984375,
0.5380859375,
0.486083984375,
0.5517578125,
0.118408203125,
0.5517578125
],
[
0.51123046875,
0.54150390625,
0.87744140625,
0.54150390625,
0.87744140625,
0.5546875,
0.51123046875,
0.5546875
],
[
0.1175537109375,
0.5556640625,
0.2476806640625,
0.5556640625,
0.2476806640625,
0.5693359375,
0.1175537109375,
0.5693359375
],
[
0.51318359375,
0.55810546875,
0.87890625,
0.55810546875,
0.87890625,
0.57177734375,
0.51318359375,
0.57177734375
],
[
0.133544921875,
0.57080078125,
0.483642578125,
0.57080078125,
0.483642578125,
0.58447265625,
0.133544921875,
0.58447265625
],
[
0.51123046875,
0.57373046875,
0.587890625,
0.576171875,
0.587890625,
0.59033203125,
0.51123046875,
0.58837890625
],
[
0.118408203125,
0.58837890625,
0.484375,
0.58837890625,
0.484375,
0.6015625,
0.118408203125,
0.6015625
],
[
0.1175537109375,
0.60400390625,
0.486083984375,
0.6025390625,
0.486083984375,
0.61767578125,
0.1175537109375,
0.6181640625
],
[
0.5283203125,
0.5986328125,
0.861328125,
0.5986328125,
0.861328125,
0.6123046875,
0.5283203125,
0.6123046875
],
[
0.118408203125,
0.62060546875,
0.48779296875,
0.62060546875,
0.48779296875,
0.63427734375,
0.118408203125,
0.63427734375
],
[
0.51318359375,
0.62451171875,
0.87890625,
0.62451171875,
0.87890625,
0.63818359375,
0.51318359375,
0.63818359375
],
[
0.12005615234375,
0.63818359375,
0.483642578125,
0.63818359375,
0.483642578125,
0.6484375,
0.12005615234375,
0.6484375
],
[
0.51318359375,
0.64111328125,
0.880859375,
0.64111328125,
0.880859375,
0.65478515625,
0.51318359375,
0.65478515625
],
[
0.118408203125,
0.65283203125,
0.48779296875,
0.65283203125,
0.48779296875,
0.66650390625,
0.118408203125,
0.66650390625
],
[
0.51123046875,
0.65625,
0.87890625,
0.65625,
0.87890625,
0.66943359375,
0.51123046875,
0.66943359375
],
[
0.118408203125,
0.66943359375,
0.1561279296875,
0.66943359375,
0.1561279296875,
0.681640625,
0.118408203125,
0.681640625
],
[
0.5146484375,
0.67333984375,
0.880859375,
0.67333984375,
0.880859375,
0.68701171875,
0.5146484375,
0.68701171875
],
[
0.1368408203125,
0.68505859375,
0.48779296875,
0.68505859375,
0.48779296875,
0.69873046875,
0.1368408203125,
0.69873046875
],
[
0.509765625,
0.6884765625,
0.880859375,
0.6884765625,
0.880859375,
0.70166015625,
0.509765625,
0.70166015625
],
[
0.12005615234375,
0.70068359375,
0.486083984375,
0.70068359375,
0.486083984375,
0.71435546875,
0.12005615234375,
0.71435546875
],
[
0.51318359375,
0.7060546875,
0.74072265625,
0.7060546875,
0.74072265625,
0.7197265625,
0.51318359375,
0.7197265625
],
[
0.118408203125,
0.71728515625,
0.483642578125,
0.71728515625,
0.483642578125,
0.73095703125,
0.118408203125,
0.73095703125
],
[
0.118408203125,
0.73388671875,
0.484375,
0.73388671875,
0.484375,
0.74755859375,
0.118408203125,
0.74755859375
],
[
0.5615234375,
0.728515625,
0.82275390625,
0.728515625,
0.82275390625,
0.74658203125,
0.5615234375,
0.74658203125
],
[
0.118408203125,
0.75,
0.484375,
0.75,
0.484375,
0.76318359375,
0.118408203125,
0.76318359375
],
[
0.6630859375,
0.75,
0.82861328125,
0.75,
0.82861328125,
0.76416015625,
0.6630859375,
0.76416015625
],
[
0.118408203125,
0.7666015625,
0.486083984375,
0.7666015625,
0.486083984375,
0.77978515625,
0.118408203125,
0.77978515625
],
[
0.51123046875,
0.77392578125,
0.8759765625,
0.77490234375,
0.8759765625,
0.7890625,
0.51123046875,
0.7880859375
],
[
0.118408203125,
0.7822265625,
0.48779296875,
0.7822265625,
0.48779296875,
0.7958984375,
0.118408203125,
0.7958984375
],
[
0.51318359375,
0.79052734375,
0.880859375,
0.79052734375,
0.880859375,
0.80419921875,
0.51318359375,
0.80419921875
],
[
0.1158447265625,
0.79736328125,
0.393798828125,
0.79736328125,
0.393798828125,
0.81103515625,
0.1158447265625,
0.81103515625
],
[
0.51318359375,
0.80712890625,
0.87890625,
0.80712890625,
0.87890625,
0.82080078125,
0.51318359375,
0.82080078125
],
[
0.12176513671875,
0.82470703125,
0.48779296875,
0.82470703125,
0.48779296875,
0.83837890625,
0.12176513671875,
0.83837890625
],
[
0.51318359375,
0.822265625,
0.880859375,
0.822265625,
0.880859375,
0.8359375,
0.51318359375,
0.8359375
],
[
0.1175537109375,
0.84033203125,
0.484375,
0.84033203125,
0.484375,
0.85400390625,
0.1175537109375,
0.85400390625
],
[
0.51318359375,
0.84033203125,
0.62109375,
0.84033203125,
0.62109375,
0.8515625,
0.51318359375,
0.8515625
],
[
0.118408203125,
0.85693359375,
0.484375,
0.85693359375,
0.484375,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.5615234375,
0.86328125,
0.6826171875,
0.8642578125,
0.6826171875,
0.87890625,
0.5615234375,
0.87744140625
],
[
0.6767578125,
0.8623046875,
0.81005859375,
0.8623046875,
0.81005859375,
0.8798828125,
0.6767578125,
0.8798828125
],
[
0.1175537109375,
0.87158203125,
0.484375,
0.87255859375,
0.484375,
0.8857421875,
0.1175537109375,
0.88525390625
],
[
0.6630859375,
0.88134765625,
0.75390625,
0.88134765625,
0.75390625,
0.8955078125,
0.6630859375,
0.8955078125
],
[
0.74658203125,
0.8828125,
0.8271484375,
0.8828125,
0.8271484375,
0.89453125,
0.74658203125,
0.89453125
],
[
0.1175537109375,
0.8896484375,
0.486083984375,
0.8896484375,
0.486083984375,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.12005615234375,
0.90673828125,
0.246826171875,
0.90673828125,
0.246826171875,
0.91748046875,
0.12005615234375,
0.91748046875
],
[
0.51318359375,
0.90478515625,
0.8671875,
0.90478515625,
0.8671875,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"变处理的段落数。并发工作CoDAE(Leietal..",
"Model",
"Encoder LayerFlops",
"2023）采用了相关的条件计算机制，旨在有效",
"T5",
"12nd² + 2n²d",
"地适应而不是对长文档进行建模",
"LONGT5",
"12nd?+d",
"DeviceutilizationFLOP并不能说明全部情",
"CoLT5",
"7nd2 +d",
"况，因为建模选择会影响加速器实现的有效运",
"行速度。对于长文本输入，由于重复加载键和",
"Table 2: CoLT5 uses significantly fewer FLOPs than",
"值的长序列的内存带宽限制，自回归解码器推",
"LoNGT5. Comparison of approximate encoder layer",
"理非常慢(Shazeer,2019;de Jongetal.，2022)",
"total FLOPs between T5, LONGT5, and CoLT5.",
"Shazeer（2019）引人了多查询注意（MQA），",
"CoLT5 FLOPs rounded to readable fractions",
"",
"共享键和值的头部以减少内存带宽开销。Pope",
"etal.（2022）研究了如何对大型模型进行分片，",
"令X为令牌i的表示形式，而u为d维的",
"尤其是在MQA的情况下，以获得最佳的设备",
"",
"可学习嵌入。那么令牌i的路由分数是",
"利用率，从而获得最佳的速度",
"Si= X,-u",
"TrainingobjectivesT5引人了span损坏目",
"我们选择顶部-k得分最高的输人。为了向评",
"标E(Raffeletal.,2020），这是掩码语言建模E",
"分嵌入提供学习信号，我们确保路由令牌对层",
"(Devlinetal..2019）的修改。LONGT5利用PE",
"更新的贡献根据路由分数进行缩放，稍后将看",
"GASUS（Zhangetal.2020）句子重构目标来提",
"到。为了向所有令牌提供更好的分布式信号，",
"高摘要性能。Tayetal.(2022）提出了UL2，这",
"我们还使用广义的softmax对路由分数进行全",
"是span损坏，前缀和因果语言建模的混合物，",
"局归一化，以汇总到所需的路由令牌数，从而",
"并表明它在短输出和生成任务上都具有出色的",
"得出归一化分数si。每个CoLT5层都有三个",
"性能",
"独立的路由器，每个路由器用于前馈层，注意",
"3CoLT5",
"查询和注意键值",
"ConditionalFeedforward直观上，一些令牌",
"3.1Conditional computation",
"表示可能比其他令牌表示受益于更多的处理。",
"如上一节所述，很大一部分TransformerFLOP",
"CoLT5条件前馈层将额外的高容量前馈层应",
"来自前馈和投影层，这些层随输人序列的长度",
"用于所选令牌。特别是，让X；是第i个令牌",
"而缩放。因此，LONGT5在长文档上的训练和",
"的模型状态，si表示标准化的路由分数（对于",
"推理仍然很昂贵",
"非路由令牌，设置为0）。然后，CoLT5的前",
"CoLT5进一步降低了通过条件计算处理长",
"馈更新由",
"文档的成本，遵循一些令牌更重要的直觉，因",
"此从繁重的计算中受益更多。首先，某些类型",
"X, = X, + FFdLight(X) + $, ·FFdHeavv(X))",
"的令牌可能固有地需要较少的计算，例如填充",
"给出。轻前馈分支和重前馈分支仅在隐藏维度",
"词和标点符号。其次，尤其是在长文档中，输",
"上有所不同，轻分支的隐藏维度小于标准T5",
"人的大部分可能与当前问题，任务或处理阶段",
"前馈层，重分支更大。令n表示输人令牌的数",
"无关",
"量，m表示所选令牌的数量，rL和rH表示轻",
"CoLT5条件计算机制由三个组件组成：路",
"型和重型隐藏尺寸与标准T5隐藏尺寸的比率。",
"由模块，条件前馈层和条件关注层。所有令牌",
"然后，CoLT5层的FLOP由",
"都由标准，轻量级的关注和前馈层处理。路由",
"模块还从每个关注或前馈层的输入中选择重",
"FLOPsFFd = 8nrLd² + 8mrHd²",
"要的令牌，并且一个繁重的条件层将额外的计",
"Light branch Heavy branch",
"算应用于路由令牌。本节详细介绍每个组件。",
"给出。我们将轻和重比设置为rL=号和rH=",
"图1提供了CoLT5条件计算机制的概述，表",
"4，分别是标准T5隐藏尺寸的一半和四倍。对",
"2比较了CoLT5和LONGT5FLOP",
"于我们的主要实验，令牌的一小部分示被路",
"Routing为了在每个层中分别为每个组件选",
"由到重分支。结果，来自CoLT5前馈层的近",
"择重要的令牌，我们需要一个可学习且易于处",
"似FLOP等于",
"理的路由功能。我们遵循Leietal.(2023）的简",
"FLOPSFFd=",
"4nd + 2nd",
"单三步机制：（1）将输入与学习的嵌入相乘以",
"Light branch",
"Heavy branch",
"获得路由分数，（2）归一化，以及（3）选择最",
"高得分的k输入",
"，消耗了标准T5前馈层的FLOP的75个%。"
] |
{
"arxiv_id": "2303.09752",
"language": "zh",
"page_idx": 3
}
|
||
2404.12541-zh-0
|
2404.12541-zh-0.jpg
|
GenVideo: 使用 T2I 扩散模型进行单次目标图像和形状感知视频编辑
Sai Sree Harsha
Adobe
ssree@adobe.com
Ambareesh Revanur
Dhwanit Agarwal
Shradha Agrawal
Adobe
Adobe
Adobe
arevanur@adobe.com
dhagarwa@adobe.com
shradagr@adobe.com
<FigureHere>
Figure 1. Editing the source car video based on a target-image and a text prompt Driving a“white bus”/“red cadillac"
down a mountain road with GenVideo method. GenVideo can accurately replace the object in the source video, even with
a target object of a different shape.
1. Introduction
Abstract
使用扩散模型[7,8,12,18,25,26,30]生成图像和视频最近非常受欢迎。最近，已经提出了许多方法来使用文本提示编辑视觉内容，如图像[33,35,36] 和视频[5,37,43,47]。尽管有几部作品展示了图像编辑[1,6,11,28,41]任务的高质量结果，但扩散模型在视频编辑任务中的利用受到限制。[8,24]等作品，使用文本到视频(T2V)扩散模型进行视频编辑。然而，这些方法的实用性有限，因为它们需要大型计算基础设施来进行训练以及大型和多样化的视频数据集。由于这些限制，最近的方法已经开始探索膨胀的文本到图像(T2I)扩散模型，作为 T2V 模型的替代方案，以一次性或零拍方式[2,4,5,9,21,31,43,46]进行视频编辑。尽管如此，除了 Make-A-Protagoner [46] 可以另外以目标图像为条件外，基于 T2I 扩散模型的现有和并发视频编辑方法仅关注文本驱动的视频编辑。因此， 这些方法不适用于仅使用文本无法准确表达编辑性质的情况。换句话说，它们不是目标图像感知的。我们在这里强调，目标图像感知为所需的编辑提供了精确的视觉指导。这种方法可以精确复制形状、大小和纹理， 这对于创建服务于用例的内容至关重要，例如创建与
Video editing methods based on diffusion models that rely solely on a text prompt for the edit are hindered by the limited expressive power of text prompts. Thus, incorporating a reference target image as a visual guide becomes desirable for precise control over edit. Also, most existing methods struggle to accurately edit a video when the shape and size of the object in the target image differ from the source object. To address these challenges, we propose“GenVideo” for editing videos leveraging target-image aware T2I models. Our approach handles edits with target objects of varying shapes and sizes while maintaining the temporal consistency of the edit using our novel target and shape aware InvEdit masks. Further, we propose a novel target-image aware latent noise correction strategy during inference to improve the temporal consistency of the edits. Experimental analyses indicate that GenVideo can effectively handle edits with objects of varying shapes, where existing approaches fail.
|
[
[
0.1298828125,
0.1307373046875,
0.8408203125,
0.1331787109375,
0.8408203125,
0.15087890625,
0.1298828125,
0.1483154296875
],
[
0.5107421875,
0.1817626953125,
0.66259765625,
0.18310546875,
0.66259765625,
0.1988525390625,
0.5107421875,
0.1976318359375
],
[
0.70166015625,
0.1817626953125,
0.8486328125,
0.18310546875,
0.8486328125,
0.1988525390625,
0.70166015625,
0.1976318359375
],
[
0.2998046875,
0.18310546875,
0.4697265625,
0.18310546875,
0.4697265625,
0.1976318359375,
0.2998046875,
0.1976318359375
],
[
0.1217041015625,
0.18310546875,
0.256591796875,
0.18310546875,
0.256591796875,
0.1976318359375,
0.1217041015625,
0.1976318359375
],
[
0.16259765625,
0.1988525390625,
0.221435546875,
0.2008056640625,
0.2197265625,
0.216552734375,
0.160888671875,
0.2139892578125
],
[
0.35693359375,
0.199462890625,
0.414306640625,
0.199462890625,
0.414306640625,
0.21533203125,
0.35693359375,
0.21533203125
],
[
0.5595703125,
0.199462890625,
0.615234375,
0.199462890625,
0.615234375,
0.21533203125,
0.5595703125,
0.21533203125
],
[
0.7490234375,
0.199462890625,
0.8056640625,
0.199462890625,
0.8056640625,
0.21533203125,
0.7490234375,
0.21533203125
],
[
0.1298828125,
0.2197265625,
0.2484130859375,
0.2197265625,
0.2484130859375,
0.23291015625,
0.1298828125,
0.23291015625
],
[
0.313720703125,
0.2197265625,
0.455078125,
0.2197265625,
0.455078125,
0.23291015625,
0.313720703125,
0.23291015625
],
[
0.5107421875,
0.2197265625,
0.65869140625,
0.2197265625,
0.65869140625,
0.23291015625,
0.5107421875,
0.23291015625
],
[
0.70166015625,
0.2197265625,
0.84326171875,
0.2197265625,
0.84326171875,
0.23291015625,
0.70166015625,
0.23291015625
],
[
0.82421875,
0.3125,
0.85107421875,
0.318115234375,
0.84716796875,
0.330078125,
0.8203125,
0.323974609375
],
[
0.09234619140625,
0.3818359375,
0.096435546875,
0.3818359375,
0.096435546875,
0.386962890625,
0.09234619140625,
0.386962890625
],
[
0.08905029296875,
0.460205078125,
0.098876953125,
0.460205078125,
0.098876953125,
0.488037109375,
0.08905029296875,
0.488037109375
],
[
0.73388671875,
0.47607421875,
0.7509765625,
0.479248046875,
0.7490234375,
0.48486328125,
0.73046875,
0.481689453125
],
[
0.8203125,
0.4765625,
0.8486328125,
0.4765625,
0.8486328125,
0.48486328125,
0.8203125,
0.48486328125
],
[
0.08087158203125,
0.517578125,
0.88818359375,
0.517578125,
0.88818359375,
0.53076171875,
0.08087158203125,
0.53076171875
],
[
0.08087158203125,
0.53271484375,
0.890625,
0.53271484375,
0.890625,
0.54345703125,
0.08087158203125,
0.54345703125
],
[
0.08087158203125,
0.54736328125,
0.313720703125,
0.54736328125,
0.313720703125,
0.55810546875,
0.08087158203125,
0.55810546875
],
[
0.238525390625,
0.57275390625,
0.313720703125,
0.57275390625,
0.313720703125,
0.5869140625,
0.238525390625,
0.5869140625
],
[
0.50341796875,
0.57275390625,
0.6396484375,
0.57275390625,
0.6396484375,
0.5869140625,
0.50341796875,
0.5869140625
],
[
0.5048828125,
0.599609375,
0.890625,
0.599609375,
0.890625,
0.61279296875,
0.5048828125,
0.61279296875
],
[
0.1004638671875,
0.60986328125,
0.468017578125,
0.60986328125,
0.468017578125,
0.62353515625,
0.1004638671875,
0.62353515625
],
[
0.5048828125,
0.61572265625,
0.8896484375,
0.61572265625,
0.8896484375,
0.62890625,
0.5048828125,
0.62890625
],
[
0.08087158203125,
0.626953125,
0.468017578125,
0.626953125,
0.468017578125,
0.64013671875,
0.08087158203125,
0.64013671875
],
[
0.5048828125,
0.6298828125,
0.890625,
0.6298828125,
0.890625,
0.64306640625,
0.5048828125,
0.64306640625
],
[
0.08087158203125,
0.6416015625,
0.46728515625,
0.6416015625,
0.46728515625,
0.65478515625,
0.08087158203125,
0.65478515625
],
[
0.50341796875,
0.64453125,
0.8896484375,
0.64306640625,
0.8896484375,
0.65771484375,
0.50341796875,
0.6591796875
],
[
0.08087158203125,
0.6572265625,
0.468017578125,
0.6572265625,
0.468017578125,
0.67041015625,
0.08087158203125,
0.67041015625
],
[
0.5048828125,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.673828125,
0.5048828125,
0.673828125
],
[
0.080078125,
0.67236328125,
0.4697265625,
0.67236328125,
0.4697265625,
0.68603515625,
0.080078125,
0.68603515625
],
[
0.5048828125,
0.67529296875,
0.8896484375,
0.67529296875,
0.8896484375,
0.689453125,
0.5048828125,
0.689453125
],
[
0.07843017578125,
0.68505859375,
0.468017578125,
0.6875,
0.468017578125,
0.7021484375,
0.07843017578125,
0.69970703125
],
[
0.50634765625,
0.68994140625,
0.890625,
0.68994140625,
0.890625,
0.7041015625,
0.50634765625,
0.7041015625
],
[
0.080078125,
0.7021484375,
0.46728515625,
0.7021484375,
0.46728515625,
0.71533203125,
0.080078125,
0.71533203125
],
[
0.50341796875,
0.7041015625,
0.890625,
0.705078125,
0.890625,
0.7197265625,
0.50341796875,
0.71826171875
],
[
0.080078125,
0.71728515625,
0.46728515625,
0.71728515625,
0.46728515625,
0.73095703125,
0.080078125,
0.73095703125
],
[
0.5048828125,
0.72021484375,
0.8896484375,
0.72021484375,
0.8896484375,
0.734375,
0.5048828125,
0.734375
],
[
0.080078125,
0.73193359375,
0.465576171875,
0.73095703125,
0.465576171875,
0.74560546875,
0.080078125,
0.7470703125
],
[
0.50634765625,
0.7373046875,
0.88623046875,
0.7373046875,
0.88623046875,
0.74755859375,
0.50634765625,
0.74755859375
],
[
0.07843017578125,
0.7470703125,
0.468017578125,
0.74560546875,
0.468017578125,
0.76025390625,
0.07843017578125,
0.76123046875
],
[
0.5048828125,
0.75048828125,
0.890625,
0.75048828125,
0.890625,
0.76513671875,
0.5048828125,
0.76513671875
],
[
0.08087158203125,
0.76220703125,
0.46728515625,
0.76220703125,
0.46728515625,
0.7763671875,
0.08087158203125,
0.7763671875
],
[
0.5185546875,
0.7666015625,
0.88818359375,
0.7666015625,
0.88818359375,
0.77978515625,
0.5185546875,
0.77978515625
],
[
0.080078125,
0.77783203125,
0.468017578125,
0.77783203125,
0.468017578125,
0.79150390625,
0.080078125,
0.79150390625
],
[
0.5048828125,
0.78076171875,
0.8896484375,
0.78076171875,
0.8896484375,
0.79443359375,
0.5048828125,
0.79443359375
],
[
0.08087158203125,
0.79248046875,
0.468017578125,
0.79248046875,
0.468017578125,
0.80615234375,
0.08087158203125,
0.80615234375
],
[
0.5048828125,
0.796875,
0.8896484375,
0.796875,
0.8896484375,
0.81005859375,
0.5048828125,
0.81005859375
],
[
0.080078125,
0.806640625,
0.468017578125,
0.80810546875,
0.468017578125,
0.8212890625,
0.080078125,
0.82080078125
],
[
0.50634765625,
0.8125,
0.88623046875,
0.8125,
0.88623046875,
0.82275390625,
0.50634765625,
0.82275390625
],
[
0.080078125,
0.82373046875,
0.46728515625,
0.82373046875,
0.46728515625,
0.8369140625,
0.080078125,
0.8369140625
],
[
0.50341796875,
0.82568359375,
0.890625,
0.82568359375,
0.890625,
0.8388671875,
0.50341796875,
0.8388671875
],
[
0.080078125,
0.83837890625,
0.46728515625,
0.8369140625,
0.46728515625,
0.8515625,
0.080078125,
0.85302734375
],
[
0.5048828125,
0.84130859375,
0.8896484375,
0.84130859375,
0.8896484375,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.08251953125,
0.85498046875,
0.46728515625,
0.85498046875,
0.46728515625,
0.86572265625,
0.08251953125,
0.86572265625
],
[
0.5048828125,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.5048828125,
0.87060546875
],
[
0.08087158203125,
0.86865234375,
0.4697265625,
0.86865234375,
0.4697265625,
0.8818359375,
0.08087158203125,
0.8818359375
],
[
0.50341796875,
0.8720703125,
0.88623046875,
0.8720703125,
0.88623046875,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.08087158203125,
0.8837890625,
0.414306640625,
0.8837890625,
0.414306640625,
0.89794921875,
0.08087158203125,
0.89794921875
],
[
0.50341796875,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"GenVideo：使用T2I扩散模型进行单次目标图像和形状感知视频编辑",
"Dhwanit Agarwal",
"Shradha Agrawal",
"AmbareeshRevanur",
"Sai Sree Harsha",
"Adobe",
"Adobe",
"Adobe",
"Adobe",
"ssree@adobe.com",
"arevanur@adobe.com",
"dhagarwa@adobe.com",
"shradagr@adobe.com",
"",
"",
"",
"",
"",
"Figure 1. Editing the source car video based on a target-image and a text prompt Driving a “white bus\"/\"red cadillac'",
"down a mountain road with GenVideo method. GenVideo can accurately replace the object in the source video,even with",
"a target object of a different shape.",
"Abstract",
"1.Introduction",
"使用扩散模型[7，8，12,18，25，26，30]生成图像和视",
"Video editing methods based on diffusion models",
"频最近非常受欢迎。最近，已经提出了许多方法来使",
"that rely solely on a text prompt for the edit are hin-",
"用文本提示编辑视觉内容，如图像33，35，36】和视",
"dered by the limited expressive power of text prompts.",
"频[5，37，43，47]。尽管有几部作品展示了图像编辑",
"Thus, incorporating a reference target image as a vi-",
"1，611，28，41]任务的高质量结果，但扩散模型在视",
"sual guide becomes desirable for precise control over",
"频编辑任务中的利用受到限制。[8，24]等作品，使用",
"edit. Also, most existing methods struggle to accu-",
"文本到视频（T2V）扩散模型进行视频编辑。然而，这",
"rately edit a video when the shape and size of the ob-",
"些方法的实用性有限，因为它们需要大型计算基础设",
"ject in the target image differ from the source object.",
"施来进行训练以及大型和多样化的视频数据集。由于",
"To address these challenges, we propose \"GenVideo\"",
"这些限制，最近的方法已经开始探索膨胀的文本到图",
"for editing videos leveraging target-image aware T21",
"像（T2I）扩散模型，作为T2V模型的替代方案，以",
"models. Our approach handles edits with target ob-",
"次性或零拍方式[2，4,5,9，21，31，43，46]进行视频",
"jects of varying shapes and sizes while maintaining the",
"编辑。尽管如此，除了Make-A-Protagoner「46]可以另",
"temporal consistency of the edit using our novel target",
"外以目标图像为条件外，基于T2I扩散模型的现有和",
"and shape aware InvEdit masks.Further, we proposea",
"并发视频编辑方法仅关注文本驱动的视频编辑。因此",
"novel target-image aware latent noise correction strat-",
"这些方法不适用于仅使用文本无法准确表达编辑性质",
"egy during inference to improve the temporal consis-",
"的情况。换句话说，它们不是目标图像感知的。我们在",
"tencyoftheedits.Experimentalanalysesindicatethat",
"这里强调，目标图像感知为所需的编辑提供了精确的",
"GenVideo can effectively handle edits with objects of",
"视觉指导。这种方法可以精确复制形状、大小和纹理",
"varying shapes, where existing approaches fail.",
"这对于创建服务于用例的内容至关重要，例如创建与"
] |
{
"arxiv_id": "2404.12541",
"language": "zh",
"page_idx": 0
}
|
||
2404.12541-zh-1
|
2404.12541-zh-1.jpg
|
特定产品细节(如特定型号的汽车)对齐的内容。此外，以前的方法缺乏形状感知，因此无法在目标对象的形状和大小与源视频有很大不同的情况下进行编辑
在本文中，我们提出了一种新颖的目标图像和形状感知的 InvEdit 方法，以针对给定的文本提示和目标图像有效地识别视频中的感兴趣区域。这是使用类似于[6]的基础 T2II 扩散模型完成的，但对于视频和使用额外的图像调节。我们的方法的优势在于 InvEdit 及其简单性，它可以适应任何图像条件扩散模型，以获得用于视频编辑的目标图像和形状感知掩码。局部掩模引导对于精确编辑和避免编辑泄漏到未掩蔽区域通常是至关重要的，但是获得目标图像和形状感知掩模是一个具有挑战性的问题，因为以前的方法无法处理形状和大小明显不同的编辑
我们进一步解决了在编辑的视频中保持时间一致性的关键挑战。由于依赖于基于源的神经层图集(NLA) [5]或基于源的帧间特征传播 [9] ,因此该研究领域的某些先前工作未能有效地解决这一挑战。。因此，由这些方法生成的视频的时间一致性主要由源视频决定。当然，当目标对象与源对象有很大不同时，这些方法就会失败。。为了解决这些问题，我们引入了一种新颖的目标图像感知潜在校正策略，以在推理过程中即时混合扩散模型的帧间延迟，以改善编辑视频中目标对象的帧间时间一致性。我们的方法使用 InvEdit 掩码指导来实现时间一致性，即使目标对象的形状和大小与源视频中的对象不同。我们与最先进的视频编辑方法进行了广泛的定量和定性比较，并得出结论，我们的方法在不同的视频集上优于以前和并发的方法。我们的贡献可以概括如下：
$\bullet$ We introduce GenVideo, the first pipeline for
target-image and shape aware editing of videos using image diffusion models that can make temporally consistent edits for target objects of varied shapes and structures.
$\bullet$ We introduce InvEdit , a novel target-image and
shape aware, zero-shot approach for generating video masks using image diffusion models. This mask enables us to identify the region of interest for editing objects of varying shapes and sizes.
$\bullet$ We introduce a target-image and shape aware latent
correction strategy to enforce inter-frame temporal consistency in the edited video, where existing approaches fail.
## 2. Related Work
文本驱动的图像和视频编辑：在扩散模型 [12,38] 出现之前，生成对抗网络(GAN)已用于编辑图像[15,27,1] 29,34,42,44] 。最近，在互联网规模的图像和文本数据上训练的生成扩散模型[33,35,36]已经实现了比GAN 显著更高的图像生成质量和多样性。受这些 T2I 扩散模型的成功启发，已开发出用于视频生成的文本到视频扩散模型$\boxed{\mathrm{F} }[ 13, 14, 22, 37] $。[37] 中的作者使用额外的时间注意力层来扩展预训练的 T2I 扩散模型
以生成视频，而最近的作品如[13]使用 3D-UNet 进行视频生成。随着生成，现有图像和视频的编辑也引起了极大的兴趣。SDEdit [23] 仔细地将噪声添加到输入图像中，然后使用扩散模型对其进行降噪以实现所需的编辑。Prompt-to-prompt(P2P)[11] 允许通过控制交叉注意图进行文本引导的图像编辑。在视频编辑领域， Tune-A-Video (TAV) [43] 在单个源视频上微调膨胀的T2I 模型，以生成具有相似运动的编辑视频。Video-P2P [21] 和 FateZero [31] 通过更改交叉注意图实现视频中的可控编辑。Pix2Video [4]通过在推理过程中改变自我注意特征来实现视频编辑。Text2Live [2] 利用 NLA,然后使用文本指南独立编辑每个图层。Text2Video-Zero [16]使用膨胀的自我关注层以及 ControlNet [45] 进行文本引导视频编辑。尽管如此，这些使用 T2I 扩散模型的作品仅关注通过文本进行的图像和视频编辑，并且不能用于仅由文本无法描述的编辑。因此，它们不是目标图像感知的，这通常是精确视频编辑[46]的关键方面
目标图像感知视觉内容编辑：为了满足仅靠文本无法轻松或准确描述的视觉内容编辑，已开发了 DALLE-2 [33]和 Gen-1 [8] 等模型，以利用 CLIP [32] 图像嵌入对图像和视频进行目标图像感知编辑。虽然 DALLE-2编辑图像，但 Gen-1 可以编辑视频，但这里的缺点是整个图像，包括背景，由于缺乏遮罩指导而影响编辑。像Dreamix$\boxed{\mathrm{F}}[24]$这样的主题驱动解决方案可以编辑以目标图像数据为条件的视频，通常需要多个目标图像才能准确地学习概念/目标对象，并且还需要[13]中使用的大型视频扩散模型。虽然他们可以基于单个图像生成和编辑视频，但编辑的视频无法保持对目标图像中目标对象的编辑的保真度。Make-A-Protagoner [46] 通过使用称为 SD-unCLIP [33] 的图像引导 T2I 扩散模型来解决这些限制。他们的管道可以使用单个目标图像实现目标图像感知的视频对象编辑，并采用掩码引导推理进行本地化编辑。但是，编辑仍然缺乏时间一致性，并且当目标对象的形状和大小与源视频中存在的源对象不同时，该方法将失败。这是因为用于编辑视频的掩码仅使用源视频生成，并且不知道目标对象
具有时间一致性的视频编辑：上述方法在编辑的视频中缺乏帧间时间一致性。TokenFlow[9]是一项并发工作， 通过在源视频的特征空间中使用最近邻域来帧间传播自关注特征，从而提高了时间一致性。然而，TokenFlow 中的编辑和时间一致性都不是目标图像和形状感知的。因此，它不能在目标对象与源对象的形状和大小有很大差异的地方进行编辑。此外，由于缺乏掩码指导，它不提供本地化编辑的方法，并且还需要手动选择关键帧以实现时间一致性。StableVideo [5] 使用源视频的NLA 来增强时间一致性，因此也缺乏目标图像和形状感知以进行准确的形状更改编辑。另一项工作[19]通过制定将原始地图集映射到已编辑地图集的变形图来解决形状感知编辑问题。虽然该方法取得了良好的效果，但 NLA 映射需要花费数十小时的训练，并且不准确的 NLA 映射会导致对象丢失部分和大量不良背景变化。我们的工作 GenVideo 通过使用 T2I 模型提供第
|
[
[
0.08087158203125,
0.09344482421875,
0.4697265625,
0.09344482421875,
0.4697265625,
0.106689453125,
0.08087158203125,
0.106689453125
],
[
0.50341796875,
0.09344482421875,
0.89208984375,
0.09344482421875,
0.89208984375,
0.106689453125,
0.50341796875,
0.106689453125
],
[
0.08416748046875,
0.10986328125,
0.46728515625,
0.10986328125,
0.46728515625,
0.12060546875,
0.08416748046875,
0.12060546875
],
[
0.50341796875,
0.10797119140625,
0.890625,
0.10797119140625,
0.890625,
0.121826171875,
0.50341796875,
0.121826171875
],
[
0.08251953125,
0.125,
0.439453125,
0.125,
0.439453125,
0.1351318359375,
0.08251953125,
0.1351318359375
],
[
0.50634765625,
0.12371826171875,
0.88818359375,
0.125,
0.88818359375,
0.1363525390625,
0.50634765625,
0.1351318359375
],
[
0.1004638671875,
0.1395263671875,
0.468017578125,
0.1395263671875,
0.468017578125,
0.15283203125,
0.1004638671875,
0.15283203125
],
[
0.50341796875,
0.136962890625,
0.890625,
0.1383056640625,
0.890625,
0.15283203125,
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,
0.1685791015625,
0.46728515625,
0.1685791015625,
0.46728515625,
0.1817626953125,
0.08251953125,
0.1817626953125
],
[
0.5048828125,
0.1685791015625,
0.890625,
0.1685791015625,
0.890625,
0.1817626953125,
0.5048828125,
0.1817626953125
],
[
0.08416748046875,
0.1849365234375,
0.46728515625,
0.1849365234375,
0.46728515625,
0.1956787109375,
0.08416748046875,
0.1956787109375
],
[
0.50634765625,
0.1849365234375,
0.8896484375,
0.1849365234375,
0.8896484375,
0.1956787109375,
0.50634765625,
0.1956787109375
],
[
0.08251953125,
0.1988525390625,
0.46728515625,
0.1988525390625,
0.46728515625,
0.212158203125,
0.08251953125,
0.212158203125
],
[
0.5048828125,
0.199462890625,
0.88818359375,
0.199462890625,
0.88818359375,
0.210205078125,
0.5048828125,
0.210205078125
],
[
0.08416748046875,
0.21533203125,
0.465576171875,
0.21533203125,
0.465576171875,
0.2259521484375,
0.08416748046875,
0.2259521484375
],
[
0.5048828125,
0.21337890625,
0.8896484375,
0.21337890625,
0.8896484375,
0.2266845703125,
0.5048828125,
0.2266845703125
],
[
0.08251953125,
0.2291259765625,
0.468017578125,
0.2291259765625,
0.468017578125,
0.242431640625,
0.08251953125,
0.242431640625
],
[
0.50341796875,
0.2291259765625,
0.8896484375,
0.2291259765625,
0.8896484375,
0.242431640625,
0.50341796875,
0.242431640625
],
[
0.08251953125,
0.24560546875,
0.46728515625,
0.24560546875,
0.46728515625,
0.255615234375,
0.08251953125,
0.255615234375
],
[
0.5048828125,
0.24560546875,
0.8896484375,
0.24560546875,
0.8896484375,
0.255615234375,
0.5048828125,
0.255615234375
],
[
0.08087158203125,
0.2587890625,
0.468017578125,
0.25830078125,
0.468017578125,
0.271484375,
0.08087158203125,
0.272705078125
],
[
0.5048828125,
0.2587890625,
0.890625,
0.2587890625,
0.890625,
0.272705078125,
0.5048828125,
0.272705078125
],
[
0.08251953125,
0.27587890625,
0.46728515625,
0.27587890625,
0.46728515625,
0.285888671875,
0.08251953125,
0.285888671875
],
[
0.50341796875,
0.2734375,
0.8896484375,
0.274658203125,
0.8896484375,
0.28857421875,
0.50341796875,
0.287353515625
],
[
0.08087158203125,
0.2890625,
0.2303466796875,
0.2890625,
0.2303466796875,
0.302978515625,
0.08087158203125,
0.302978515625
],
[
0.5048828125,
0.2890625,
0.890625,
0.2890625,
0.890625,
0.302978515625,
0.5048828125,
0.302978515625
],
[
0.1004638671875,
0.3037109375,
0.468017578125,
0.302978515625,
0.468017578125,
0.317626953125,
0.1004638671875,
0.318115234375
],
[
0.50634765625,
0.30615234375,
0.8896484375,
0.30615234375,
0.8896484375,
0.31640625,
0.50634765625,
0.31640625
],
[
0.080078125,
0.3193359375,
0.4697265625,
0.318115234375,
0.4697265625,
0.33203125,
0.080078125,
0.333251953125
],
[
0.50634765625,
0.3212890625,
0.8896484375,
0.3212890625,
0.8896484375,
0.33203125,
0.50634765625,
0.33203125
],
[
0.08087158203125,
0.335205078125,
0.468017578125,
0.335205078125,
0.468017578125,
0.348388671875,
0.08087158203125,
0.348388671875
],
[
0.50634765625,
0.3359375,
0.8896484375,
0.3359375,
0.8896484375,
0.3466796875,
0.50634765625,
0.3466796875
],
[
0.08251953125,
0.349853515625,
0.46728515625,
0.349853515625,
0.46728515625,
0.363037109375,
0.08251953125,
0.363037109375
],
[
0.50341796875,
0.35107421875,
0.5361328125,
0.35107421875,
0.5361328125,
0.3623046875,
0.50341796875,
0.3623046875
],
[
0.08416748046875,
0.3662109375,
0.46728515625,
0.3662109375,
0.46728515625,
0.376953125,
0.08416748046875,
0.376953125
],
[
0.50634765625,
0.371826171875,
0.8896484375,
0.371826171875,
0.8896484375,
0.3818359375,
0.50634765625,
0.3818359375
],
[
0.08251953125,
0.380126953125,
0.468017578125,
0.380126953125,
0.468017578125,
0.393310546875,
0.08251953125,
0.393310546875
],
[
0.50341796875,
0.385009765625,
0.89208984375,
0.385009765625,
0.89208984375,
0.3984375,
0.50341796875,
0.3984375
],
[
0.08251953125,
0.396484375,
0.465576171875,
0.396484375,
0.465576171875,
0.4072265625,
0.08251953125,
0.4072265625
],
[
0.5048828125,
0.399658203125,
0.8896484375,
0.399658203125,
0.8896484375,
0.41357421875,
0.5048828125,
0.41357421875
],
[
0.08087158203125,
0.410400390625,
0.468017578125,
0.410400390625,
0.468017578125,
0.423583984375,
0.08087158203125,
0.423583984375
],
[
0.5048828125,
0.4140625,
0.89208984375,
0.4140625,
0.89208984375,
0.427978515625,
0.5048828125,
0.427978515625
],
[
0.08416748046875,
0.4267578125,
0.46728515625,
0.4267578125,
0.46728515625,
0.4375,
0.08416748046875,
0.4375
],
[
0.50341796875,
0.429931640625,
0.890625,
0.429931640625,
0.890625,
0.443115234375,
0.50341796875,
0.443115234375
],
[
0.08087158203125,
0.43994140625,
0.468017578125,
0.43994140625,
0.468017578125,
0.453857421875,
0.08087158203125,
0.453857421875
],
[
0.5048828125,
0.44580078125,
0.890625,
0.44580078125,
0.890625,
0.458984375,
0.5048828125,
0.458984375
],
[
0.08087158203125,
0.455810546875,
0.468017578125,
0.455810546875,
0.468017578125,
0.4697265625,
0.08087158203125,
0.4697265625
],
[
0.50341796875,
0.460205078125,
0.890625,
0.460205078125,
0.890625,
0.473388671875,
0.50341796875,
0.473388671875
],
[
0.08087158203125,
0.47021484375,
0.468017578125,
0.47021484375,
0.468017578125,
0.484130859375,
0.08087158203125,
0.484130859375
],
[
0.5048828125,
0.474853515625,
0.890625,
0.474853515625,
0.890625,
0.488037109375,
0.5048828125,
0.488037109375
],
[
0.0849609375,
0.4873046875,
0.46728515625,
0.4873046875,
0.46728515625,
0.49755859375,
0.0849609375,
0.49755859375
],
[
0.50341796875,
0.4892578125,
0.88818359375,
0.490478515625,
0.88818359375,
0.50390625,
0.50341796875,
0.50244140625
],
[
0.08087158203125,
0.5,
0.468017578125,
0.50048828125,
0.468017578125,
0.51513671875,
0.08087158203125,
0.5146484375
],
[
0.50341796875,
0.5048828125,
0.890625,
0.5048828125,
0.890625,
0.51806640625,
0.50341796875,
0.51806640625
],
[
0.08251953125,
0.517578125,
0.1846923828125,
0.517578125,
0.1846923828125,
0.52783203125,
0.08251953125,
0.52783203125
],
[
0.5048828125,
0.52099609375,
0.8896484375,
0.52099609375,
0.8896484375,
0.5341796875,
0.5048828125,
0.5341796875
],
[
0.08416748046875,
0.53076171875,
0.46728515625,
0.5322265625,
0.46728515625,
0.546875,
0.08416748046875,
0.54541015625
],
[
0.5048828125,
0.5361328125,
0.890625,
0.5361328125,
0.890625,
0.5498046875,
0.5048828125,
0.5498046875
],
[
0.101318359375,
0.54736328125,
0.46728515625,
0.54736328125,
0.46728515625,
0.56103515625,
0.101318359375,
0.56103515625
],
[
0.5048828125,
0.55029296875,
0.890625,
0.55029296875,
0.890625,
0.564453125,
0.5048828125,
0.564453125
],
[
0.10296630859375,
0.56298828125,
0.46728515625,
0.56298828125,
0.46728515625,
0.5771484375,
0.10296630859375,
0.5771484375
],
[
0.5048828125,
0.56640625,
0.88818359375,
0.56640625,
0.88818359375,
0.580078125,
0.5048828125,
0.580078125
],
[
0.101318359375,
0.57763671875,
0.468017578125,
0.5771484375,
0.468017578125,
0.59033203125,
0.101318359375,
0.59130859375
],
[
0.5048828125,
0.58056640625,
0.88818359375,
0.58056640625,
0.88818359375,
0.5947265625,
0.5048828125,
0.5947265625
],
[
0.101318359375,
0.59326171875,
0.2607421875,
0.59326171875,
0.2607421875,
0.60693359375,
0.101318359375,
0.60693359375
],
[
0.5048828125,
0.59521484375,
0.890625,
0.59521484375,
0.890625,
0.609375,
0.5048828125,
0.609375
],
[
0.08416748046875,
0.60791015625,
0.468017578125,
0.60791015625,
0.468017578125,
0.6220703125,
0.08416748046875,
0.6220703125
],
[
0.50634765625,
0.6123046875,
0.8896484375,
0.6123046875,
0.8896484375,
0.62255859375,
0.50634765625,
0.62255859375
],
[
0.1004638671875,
0.62353515625,
0.468017578125,
0.62353515625,
0.468017578125,
0.63720703125,
0.1004638671875,
0.63720703125
],
[
0.5048828125,
0.626953125,
0.890625,
0.626953125,
0.890625,
0.64013671875,
0.5048828125,
0.64013671875
],
[
0.101318359375,
0.63720703125,
0.46728515625,
0.63720703125,
0.46728515625,
0.6513671875,
0.101318359375,
0.6513671875
],
[
0.50341796875,
0.64013671875,
0.84326171875,
0.64013671875,
0.84326171875,
0.65380859375,
0.50341796875,
0.65380859375
],
[
0.10296630859375,
0.65283203125,
0.46728515625,
0.65283203125,
0.46728515625,
0.66650390625,
0.10296630859375,
0.66650390625
],
[
0.1004638671875,
0.66748046875,
0.431396484375,
0.66845703125,
0.431396484375,
0.681640625,
0.1004638671875,
0.68115234375
],
[
0.50634765625,
0.662109375,
0.88623046875,
0.662109375,
0.88623046875,
0.67236328125,
0.50634765625,
0.67236328125
],
[
0.5048828125,
0.67529296875,
0.890625,
0.67529296875,
0.890625,
0.689453125,
0.5048828125,
0.689453125
],
[
0.08251953125,
0.68310546875,
0.46728515625,
0.681640625,
0.46728515625,
0.6962890625,
0.08251953125,
0.69775390625
],
[
0.50634765625,
0.6923828125,
0.8896484375,
0.6923828125,
0.8896484375,
0.70263671875,
0.50634765625,
0.70263671875
],
[
0.10455322265625,
0.70068359375,
0.46728515625,
0.70068359375,
0.46728515625,
0.7109375,
0.10455322265625,
0.7109375
],
[
0.50341796875,
0.70556640625,
0.88818359375,
0.705078125,
0.88818359375,
0.71826171875,
0.50341796875,
0.7197265625
],
[
0.101318359375,
0.7119140625,
0.464111328125,
0.7138671875,
0.464111328125,
0.72900390625,
0.101318359375,
0.7265625
],
[
0.50341796875,
0.7197265625,
0.8896484375,
0.72021484375,
0.8896484375,
0.734375,
0.50341796875,
0.73291015625
],
[
0.098876953125,
0.72900390625,
0.19775390625,
0.7265625,
0.19775390625,
0.7412109375,
0.098876953125,
0.74365234375
],
[
0.50634765625,
0.73779296875,
0.8896484375,
0.73779296875,
0.8896484375,
0.74853515625,
0.50634765625,
0.74853515625
],
[
0.080078125,
0.7568359375,
0.2279052734375,
0.7568359375,
0.2279052734375,
0.77099609375,
0.080078125,
0.77099609375
],
[
0.5048828125,
0.751953125,
0.890625,
0.751953125,
0.890625,
0.76513671875,
0.5048828125,
0.76513671875
],
[
0.50634765625,
0.767578125,
0.890625,
0.767578125,
0.890625,
0.78076171875,
0.50634765625,
0.78076171875
],
[
0.08251953125,
0.78271484375,
0.46728515625,
0.78271484375,
0.46728515625,
0.79345703125,
0.08251953125,
0.79345703125
],
[
0.50341796875,
0.7822265625,
0.890625,
0.7822265625,
0.890625,
0.79541015625,
0.50341796875,
0.79541015625
],
[
0.080078125,
0.796875,
0.46728515625,
0.79541015625,
0.46728515625,
0.81005859375,
0.080078125,
0.81103515625
],
[
0.50244140625,
0.79541015625,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.50244140625,
0.8095703125
],
[
0.08251953125,
0.81298828125,
0.46728515625,
0.81298828125,
0.46728515625,
0.82373046875,
0.08251953125,
0.82373046875
],
[
0.50634765625,
0.81298828125,
0.8896484375,
0.81298828125,
0.8896484375,
0.82373046875,
0.50634765625,
0.82373046875
],
[
0.08251953125,
0.8271484375,
0.465576171875,
0.8271484375,
0.465576171875,
0.84033203125,
0.08251953125,
0.84033203125
],
[
0.5048828125,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.5048828125,
0.84033203125
],
[
0.08087158203125,
0.84130859375,
0.46728515625,
0.84033203125,
0.46728515625,
0.85400390625,
0.08087158203125,
0.85498046875
],
[
0.5048828125,
0.84228515625,
0.890625,
0.84228515625,
0.890625,
0.85595703125,
0.5048828125,
0.85595703125
],
[
0.08087158203125,
0.85693359375,
0.468017578125,
0.85595703125,
0.468017578125,
0.869140625,
0.08087158203125,
0.87060546875
],
[
0.50341796875,
0.85693359375,
0.890625,
0.85693359375,
0.890625,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.08087158203125,
0.87255859375,
0.46728515625,
0.87255859375,
0.46728515625,
0.88623046875,
0.08087158203125,
0.88623046875
],
[
0.50341796875,
0.87060546875,
0.890625,
0.8720703125,
0.890625,
0.88623046875,
0.50341796875,
0.88525390625
],
[
0.08251953125,
0.88720703125,
0.468017578125,
0.88720703125,
0.468017578125,
0.90087890625,
0.08251953125,
0.90087890625
],
[
0.5048828125,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.5048828125,
0.90087890625
]
] |
[
"特定产品细节（如特定型号的汽车）对齐的内容。。此",
"以生成视频，而最近的作品如【13]使用3D-UNet进",
"外，以前的方法缺乏形状感知，因此无法在目标对象的",
"行视频生成。随着生成，现有图像和视频的编辑也引起",
"形状和大小与源视频有很大不同的情况下进行编辑",
"了极大的兴趣。SDEdit23仔细地将噪声添加到输入",
"在本文中，我们提出了一种新颖的目标图像和形状",
"图像中，然后使用扩散模型对其进行降噪以实现所需",
"感知的InvEdit方法，以针对给定的文本提示和目标图",
"的编辑。Prompt-to-prompt(P2P）[11]允许通过控制交",
"像有效地识别视频中的感兴趣区域。这是使用类似于",
"叉注意图进行文本引导的图像编辑。在视频编辑领域，",
"6的基础T2I扩散模型完成的，但对于视频和使用额",
"Tune-A-Video（TAV）43在单个源视频上微调膨胀的",
"外的图像调节。我们的方法的优势在于InvEdit及其简",
"T2I模型，以生成具有相似运动的编辑视频。Video-P2F",
"单性，它可以适应任何图像条件扩散模型，以获得用于",
"[21]和FateZero[31]通过更改交叉注意图实现视频中",
"视频编辑的目标图像和形状感知掩码。局部掩模引导",
"的可控编辑。Pix2Video[4]通过在推理过程中改变自我",
"对于精确编辑和避免编辑泄漏到未掩蔽区域通常是至",
"注意特征来实现视频编辑。Text2Live2利用NLA，然",
"关重要的，但是获得目标图像和形状感知掩模是一个",
"后使用文本指南独立编辑每个图层。Text2Video-Zero",
"具有挑战性的问题，因为以前的方法无法处理形状和",
"[16】使用膨胀的自我关注层以及ControlNet【45]进行",
"大小明显不同的编辑",
"文本引导视频编辑。尽管如此，这些使用T2I扩散模",
"我们进一步解决了在编辑的视频中保持时间一致性",
"型的作品仅关注通过文本进行的图像和视频编辑，并",
"的关键挑战。由于依赖于基于源的神经层图集（NLA）",
"且不能用于仅由文本无法描述的编辑。因此，它们不是",
"[5]或基于源的顿间特征传播[9]，因此该研究领域的某",
"自标图像感知的，这通常是精确视频编辑46的关键",
"些先前工作未能有效地解决这一挑战。。因此，由这些",
"方面",
"方法生成的视频的时间一致性主要由源视频决定。当",
"自标图像感知视觉内容编辑：为广满足仅靠文本无法",
"然，当目标对象与源对象有很大不同时，这些方法就会",
"轻松或准确描述的视觉内容编辑，已开发了DALLE-2",
"失败。。为了解决这些问题，我们引入了一种新颖的目",
"[33]和Gen-1[8]等模型，以利用CLIP[32]图像嵌人",
"标图像感知潜在校正策略，以在推理过程中即时混合",
"对图像和视频进行目标图像感知编辑。虽然DALLE-2",
"扩散模型的间延迟，以改善编辑视频中目标对象的",
"编辑图像，但Gen-1可以编辑视频，但这里的缺点是整",
"顿间时间一致性。我们的方法使用InvEdit掩码指导来",
"个图像，包括背景，由于缺乏遮罩指导而影响编辑。像",
"实现时间一致性，即使目标对象的形状和大小与源视",
"DreamixE[24]这样的主题驱动解决方案可以编辑以目",
"频中的对象不同。我们与最先进的视频编辑方法进行",
"标图像数据为条件的视频，通常需要多个目标图像才",
"了广泛的定量和定性比较，并得出结论，我们的方法在",
"能准确地学习概念/目标对象，并且还需要【13]中使用",
"不同的视频集上优于以前和并发的方法。我们的贡献",
"的大型视频扩散模型。虽然他们可以基于单个图像生",
"可以概括如下：",
"成和编辑视频，但编辑的视频无法保持对目标图像中",
": We introduce GenVideo , the first pipeline for",
"目标对象的编辑的保真度。Make-A-Protagoner[46]】通",
"target-image and shape aware editing of videos us-",
"过使用称为SD-unCLIP[33]的图像引导T2I扩散模",
"ing image diffusion models that can make tempo-",
"型来解决这些限制。他们的管道可以使用单个目标图",
"rally consistent edits for target objects of varied",
"像实现目标图像感知的视频对象编辑，并采用掩码引",
"shapes and structures.",
"导推理进行本地化编辑。但是，编辑仍然缺乏时间一致",
": We introduce InvEdit , a novel target-image and",
"性，并且当自标对象的形状和大小与源视频中存在的",
"shape aware, zero-shot approach for generating",
"源对象不同时，该方法将失败。这是因为用于编辑视频",
"video masks using image diffusion models. This",
"的掩码仅使用源视频生成，并且不知道目标对象",
"mask enables us to identify the region of interest",
"for editing objects of varying shapes and sizes",
"具有时间一致性的视频编辑：上述方法在编辑的视频中",
"缺乏顿间时间一致性。TokenFlow[9]是一项并发工作，",
": We introduce a target-image and shape aware latent",
"通过在源视频的特征空间中使用最近邻域来顿间传播",
"correctionstrategyto enforceinter-frametemporal",
"自关注特征，从而提高了时间一致性。然而，TokenFlow",
"consistency in the edited video, where existing ap",
"中的编辑和时间一致性都不是目标图像和形状感知的。",
"proaches fail.",
"因此，它不能在自标对象与源对象的形状和大小有很",
"2. Related Work",
"大差异的地方进行编辑。此外，由于缺乏掩码指导，它",
"不提供本地化编辑的方法，并且还需要手动选择关键",
"文本驱动的图像和视频编辑：在扩散模型12，38出现",
"顿以实现时间一致性。StableVideo[5]使用源视频的",
"之前，生成对抗网络（GAN）已用于编辑图像[15,27，",
"NLA来增强时间一致性，因此也缺乏目标图像和形状",
"29，34，42，44。最近，在互联网规模的图像和文本数",
"感知以进行准确的形状更改编辑。另一项工作19通",
"据上训练的生成扩散模型[33，35，36]已经实现了比",
"过制定将原始地图集映射到已编辑地图集的变形图来",
"GAN显著更高的图像生成质量和多样性。受这些T2I",
"解决形状感知编辑问题。虽然该方法取得了良好的效",
"扩散模型的成功启发，已开发出用于视频生成的文本",
"果，但NLA映射需要花费数十小时的训练，并且不准",
"到视频扩散模型E[13，14，22，37]。[37]中的作者使",
"确的NLA映射会导致对象丢失部分和大量不良背景变",
"用额外的时间注意力层来扩展预训练的T2I扩散模型",
"化。我们的工作GenVideo通过使用T2I模型提供第"
] |
{
"arxiv_id": "2404.12541",
"language": "zh",
"page_idx": 1
}
|
||
2404.12541-zh-2
|
2404.12541-zh-2.jpg
|
<FigureHere>
Figure 2. Overview of GenVideo. Inflated attention layers are finetuned during source video fnetuning. During inference, InvEdit predicts a region to edit and latent correction uses that mask to improve the inter-frame temporal consistency. $\mathcal{M}_{\phi}-``$no mask"
一个“目标图像和形状感知”视频编辑来解决这些限制，仅需 30 分钟的训练和 2 分钟的推理，使用局部掩码指导来避免背景变化
# 3. Approach
GenVideo 旨在根据目标文本提示和包含任何形状对象的目标图像来编辑给定的源视频，同时保持时间一致性。更正式地说，给定一个输人源视频$\mathcal{V}^{src}=$ $[ I_{1}^{src}, \cdots , I_{N}^{src}]$ ,该输人源视频由包含源对象的$N$帧， 描述源视频的源文本提示$\mathcal{P}^src$,包含目标对象的目标图像$I^trg$和描述源视频所需编辑的目标文本提示$\mathcal{P}^trg$ 组成。整个训练和推理管道总结在图。2。首先，我们使用标准 LDMs [35] 的重建损失在源视频上微调膨胀的稳定扩散 unCLIP (SD-unCLIP) 模型[33,46] (秒。3.1 )。然后，我们采用了新颖的目标图像和形状感知遮罩生成方法，称为 InvEdit,在该方法中，我们使用微调模型来推断需要本地化编辑的感兴趣区域 (Sec。3.2 )。最后，我们介绍了一种新颖的潜在校正方法来提高帧间时间一致性 (Sec。3.3 )
3.1. Fine-tuning on the source video
膨胀的预训练 T2I 扩散模型的微调程序符合 TAV 方
法[43](图。2)。与[43]不同，我们使用 SD-unCLIP 模型 [33] , 它在目标图像和文本提示 [46] 上生成条件。该模型使用 CLIP-视觉分支获得输入参考图像的图像
嵌入$\mathcal{J}^*$,并使用 CLIP-文本分支获得文本嵌人$\mathcal{C}^*.$ 作为膨胀过程的一部分，空间自我注意被膨胀为时空注意(ST-attn),并且在 ST-attn 和交叉注意块之后引人了额外的时间注意(T-attn)。见补充。有关详细信息
# $3. 2. \textbf{ InvEdit mask generation}$
在本节中，我们将使用 Sec 的微调扩散模型描述InvEdit 的详细信息-我们新颖的零射，目标图像和形状感知遮罩生成策略。3.1 。现有方法 [31,46] 仅使用源视频计算蒙版，因此对目标对象的相对形状和大小没有形状感知 (例如，将汽车换成公共汽车)。InvEdit 是针对视频的 DiffEdit[6]的改编，包括目标图像和形状感知
InvEdit 步骤：首先，我们执行 DDIM 反转 [38] ,将源视频转换为相应的随机潜噪$\mathcal{Z}_T^{src}=[z_{T,1}^{src},z_{T,2}^{src},\cdots,z_{T,N}^{src}]$ 。然后，我们使用来自 Sec 的微调膨胀 SD-unCLIP 模型对潜在代码$\mathcal{Z}_T^{src}$进行去噪。3.1 使用确定性 DDIM 采样源和目标分支，如图所示。3。对于源分支，我们使用源文本提示$\mathcal{P}^src$和从$\mathcal{V}^src$中随机选择的帧$I_n^{src}$ 作为微调 SD-unCLIP 模型的条件输入来对$z_T^{src}$进行降噪。同样，并行地，我们使用目标文本提示$\bar{\mathcal{P}^{trg}}$和目标图像$I^{trg}$中的前景对象作为条件输人来重复 DDIM 采样。我们使用 GroundedSAM [17,20] 从目标图像中分割出前景对象
我们通过在源和目标分支中的每个去噪时间步长获得的去噪 UNet 模型(由$\varepsilon_{\theta}$表示)来计算预测$\begin{array}{c}\text{噪声的差。更正式地说，对于}&\mathcal{V}^{src}&\text{中的每个}&I_{n}^{src}\\\text{,我们计算 }\Delta\varepsilon_{t,n}&=&abs(\varepsilon_{t,n}^{src}-\varepsilon_{t,n}^{trg})&,&\text{其中}&\varepsilon_{t,n}^{u}=\end{array}$ $\varepsilon_{\theta}\left(z_{t,n}^{u},t,\mathcal{C}^{u},\mathcal{J}_{n}^{u}\right),u\in\left\{src,trg\right\}$。在每个 DDIM 去噪步骤中，我们获得
$$\mathcal{Z}_{t}^{src}=\mathrm{DDIM}(\mathcal{Z}_{t+1}^{src},\epsilon_{t+1}^{src}),\mathcal{Z}_{t}^{trg}=\mathrm{DDIM}(\mathcal{Z}_{t+1}^{trg},\epsilon_{t+1}^{trg}),$$
其中$t\in\{T-1,\cdots,1\}$表示时间步长
这些差异$\Delta\varepsilon_{t,n}$ ,在图中表示为热图。3 在多个去噪时间步长上进行平均，并进行二值化，以获得视频每帧的目标感知 InvEdit 掩码。InFig。3 , InvEdit 掩码能够确定将编辑放置在形状像卡车而不是汽车的区域中，因为卡车比汽车大得多。我们用$\mathcal{M}^{inv}=[M_1,M_2,\cdots,M_N]$表示$N$帧的掩码，其中$M_{n}= binarize( mean_{t\in [ t_{i}, t_{j}] }( \Delta \varepsilon _{t, n}) )$ 。InvEdit 于标识将目标图像嵌人与时间步长嵌人[33]一起注人到 UNet 的 ResNet 块的特征中的区域
InvEdit 直觉：我们的直觉遵循 DiffEdit[6]进行文本驱动的图像编辑。我们将其扩展为目标图像感知视频编辑用例。在 DDIM 去噪期间，SD-unCLIP 模型将在不同的文本和图像条件下产生不同的噪声估计。在潜在代码最终将根据条件解码不同形状，颜色和纹理的区域中，噪声估计将有所不同。另一方面，对于背景， 噪声估计几乎没有变化。因此，可以使用噪声估计之间的差异来推断掩码，该掩码标识需要编辑的每个视频帧的部分
|
[
[
0.1690673828125,
0.09344482421875,
0.375732421875,
0.0947265625,
0.375732421875,
0.10919189453125,
0.1690673828125,
0.10797119140625
],
[
0.5048828125,
0.0921630859375,
0.88134765625,
0.0921630859375,
0.88134765625,
0.1060791015625,
0.5048828125,
0.1060791015625
],
[
0.5048828125,
0.10797119140625,
0.8896484375,
0.10797119140625,
0.8896484375,
0.121826171875,
0.5048828125,
0.121826171875
],
[
0.365234375,
0.1142578125,
0.44287109375,
0.11614990234375,
0.44287109375,
0.1307373046875,
0.365234375,
0.1287841796875
],
[
0.50341796875,
0.121826171875,
0.88818359375,
0.12249755859375,
0.88818359375,
0.136962890625,
0.50341796875,
0.1363525390625
],
[
0.5048828125,
0.1383056640625,
0.890625,
0.1383056640625,
0.890625,
0.1514892578125,
0.5048828125,
0.1514892578125
],
[
0.400390625,
0.1483154296875,
0.41259765625,
0.1483154296875,
0.41259765625,
0.1514892578125,
0.400390625,
0.1514892578125
],
[
0.226318359375,
0.1558837890625,
0.2646484375,
0.1558837890625,
0.2646484375,
0.169189453125,
0.226318359375,
0.169189453125
],
[
0.50341796875,
0.154052734375,
0.52294921875,
0.154052734375,
0.52294921875,
0.1673583984375,
0.50341796875,
0.1673583984375
],
[
0.1854248046875,
0.1787109375,
0.200927734375,
0.1787109375,
0.200927734375,
0.18310546875,
0.1854248046875,
0.18310546875
],
[
0.50244140625,
0.1767578125,
0.73046875,
0.177978515625,
0.73046875,
0.1925048828125,
0.50244140625,
0.1912841796875
],
[
0.17724609375,
0.184326171875,
0.2132568359375,
0.186279296875,
0.2115478515625,
0.1962890625,
0.176513671875,
0.1944580078125
],
[
0.5048828125,
0.1944580078125,
0.890625,
0.1944580078125,
0.890625,
0.2083740234375,
0.5048828125,
0.2083740234375
],
[
0.160888671875,
0.2083740234375,
0.2279052734375,
0.210205078125,
0.2279052734375,
0.2247314453125,
0.160888671875,
0.222900390625
],
[
0.50244140625,
0.2083740234375,
0.8896484375,
0.208984375,
0.8896484375,
0.2235107421875,
0.50244140625,
0.222900390625
],
[
0.10455322265625,
0.2203369140625,
0.1446533203125,
0.2139892578125,
0.148681640625,
0.2298583984375,
0.108642578125,
0.236083984375
],
[
0.5048828125,
0.2259521484375,
0.8896484375,
0.2259521484375,
0.8896484375,
0.2392578125,
0.5048828125,
0.2392578125
],
[
0.1683349609375,
0.242431640625,
0.226318359375,
0.242431640625,
0.226318359375,
0.25390625,
0.1683349609375,
0.25390625
],
[
0.5048828125,
0.240478515625,
0.890625,
0.240478515625,
0.890625,
0.25390625,
0.5048828125,
0.25390625
],
[
0.50341796875,
0.255126953125,
0.890625,
0.25390625,
0.890625,
0.268310546875,
0.50341796875,
0.26953125
],
[
0.09478759765625,
0.24560546875,
0.15283203125,
0.24560546875,
0.15283203125,
0.2568359375,
0.09478759765625,
0.2568359375
],
[
0.2197265625,
0.271484375,
0.330078125,
0.271484375,
0.330078125,
0.285888671875,
0.2197265625,
0.285888671875
],
[
0.5048828125,
0.270751953125,
0.890625,
0.270751953125,
0.890625,
0.2841796875,
0.5048828125,
0.2841796875
],
[
0.134033203125,
0.291748046875,
0.1968994140625,
0.291748046875,
0.1968994140625,
0.30615234375,
0.134033203125,
0.30615234375
],
[
0.291748046875,
0.291748046875,
0.410888671875,
0.291748046875,
0.410888671875,
0.30615234375,
0.291748046875,
0.30615234375
],
[
0.5048828125,
0.285888671875,
0.5537109375,
0.285888671875,
0.5537109375,
0.29736328125,
0.5048828125,
0.29736328125
],
[
0.50341796875,
0.300537109375,
0.890625,
0.300537109375,
0.890625,
0.314453125,
0.50341796875,
0.314453125
],
[
0.0882568359375,
0.311279296875,
0.1168212890625,
0.311279296875,
0.1168212890625,
0.322509765625,
0.0882568359375,
0.322509765625
],
[
0.50244140625,
0.31494140625,
0.88818359375,
0.31640625,
0.88818359375,
0.330810546875,
0.50244140625,
0.330078125
],
[
0.5087890625,
0.33203125,
0.890625,
0.330810546875,
0.890625,
0.34521484375,
0.5087890625,
0.3466796875
],
[
0.5048828125,
0.3466796875,
0.890625,
0.3466796875,
0.890625,
0.361083984375,
0.5048828125,
0.361083984375
],
[
0.5048828125,
0.363037109375,
0.8896484375,
0.363037109375,
0.8896484375,
0.376953125,
0.5048828125,
0.376953125
],
[
0.5048828125,
0.376953125,
0.88818359375,
0.376953125,
0.88818359375,
0.391357421875,
0.5048828125,
0.391357421875
],
[
0.50341796875,
0.391357421875,
0.88818359375,
0.391357421875,
0.88818359375,
0.406005859375,
0.50341796875,
0.406005859375
],
[
0.134033203125,
0.412353515625,
0.2017822265625,
0.412353515625,
0.2017822265625,
0.423583984375,
0.134033203125,
0.423583984375
],
[
0.50244140625,
0.4072265625,
0.88623046875,
0.4072265625,
0.88623046875,
0.421630859375,
0.50244140625,
0.421630859375
],
[
0.0906982421875,
0.41357421875,
0.12744140625,
0.41357421875,
0.12744140625,
0.423583984375,
0.0906982421875,
0.423583984375
],
[
0.24267578125,
0.41357421875,
0.361083984375,
0.41357421875,
0.361083984375,
0.423583984375,
0.24267578125,
0.423583984375
],
[
0.410888671875,
0.41357421875,
0.455810546875,
0.41357421875,
0.455810546875,
0.423583984375,
0.410888671875,
0.423583984375
],
[
0.5048828125,
0.42236328125,
0.89208984375,
0.42236328125,
0.89208984375,
0.436279296875,
0.5048828125,
0.436279296875
],
[
0.5048828125,
0.4375,
0.890625,
0.4375,
0.890625,
0.451904296875,
0.5048828125,
0.451904296875
],
[
0.205078125,
0.41845703125,
0.226318359375,
0.41845703125,
0.226318359375,
0.42236328125,
0.205078125,
0.42236328125
],
[
0.08087158203125,
0.438232421875,
0.468017578125,
0.438232421875,
0.468017578125,
0.451904296875,
0.08087158203125,
0.451904296875
],
[
0.080078125,
0.451904296875,
0.465576171875,
0.451904296875,
0.465576171875,
0.46533203125,
0.080078125,
0.46533203125
],
[
0.5048828125,
0.45263671875,
0.61767578125,
0.45263671875,
0.61767578125,
0.466552734375,
0.5048828125,
0.466552734375
],
[
0.080078125,
0.466552734375,
0.46728515625,
0.466552734375,
0.46728515625,
0.479736328125,
0.080078125,
0.479736328125
],
[
0.52294921875,
0.46728515625,
0.8896484375,
0.468505859375,
0.8896484375,
0.48291015625,
0.52294921875,
0.481689453125
],
[
0.080078125,
0.479248046875,
0.46728515625,
0.479736328125,
0.46728515625,
0.49365234375,
0.080078125,
0.492431640625
],
[
0.5048828125,
0.484130859375,
0.8896484375,
0.484130859375,
0.8896484375,
0.49755859375,
0.5048828125,
0.49755859375
],
[
0.080078125,
0.49365234375,
0.2418212890625,
0.49365234375,
0.2418212890625,
0.5068359375,
0.080078125,
0.5068359375
],
[
0.50341796875,
0.49755859375,
0.88623046875,
0.49755859375,
0.88623046875,
0.51220703125,
0.50341796875,
0.51220703125
],
[
0.08087158203125,
0.51220703125,
0.468017578125,
0.51318359375,
0.468017578125,
0.52783203125,
0.08087158203125,
0.5263671875
],
[
0.50341796875,
0.51318359375,
0.8759765625,
0.51318359375,
0.8759765625,
0.52783203125,
0.50341796875,
0.52783203125
],
[
0.08251953125,
0.52880859375,
0.46728515625,
0.52880859375,
0.46728515625,
0.54248046875,
0.08251953125,
0.54248046875
],
[
0.50244140625,
0.53076171875,
0.890625,
0.52880859375,
0.890625,
0.54345703125,
0.50244140625,
0.54541015625
],
[
0.08087158203125,
0.54345703125,
0.2442626953125,
0.54345703125,
0.2442626953125,
0.556640625,
0.08087158203125,
0.556640625
],
[
0.50341796875,
0.54345703125,
0.65283203125,
0.54541015625,
0.65283203125,
0.56005859375,
0.50341796875,
0.55810546875
],
[
0.07843017578125,
0.57470703125,
0.19287109375,
0.57470703125,
0.19287109375,
0.5888671875,
0.07843017578125,
0.5888671875
],
[
0.5107421875,
0.57080078125,
0.88134765625,
0.57080078125,
0.88134765625,
0.58544921875,
0.5107421875,
0.58544921875
],
[
0.86767578125,
0.5859375,
0.8896484375,
0.5859375,
0.8896484375,
0.5986328125,
0.86767578125,
0.5986328125
],
[
0.08087158203125,
0.599609375,
0.46728515625,
0.599609375,
0.46728515625,
0.61279296875,
0.08087158203125,
0.61279296875
],
[
0.08251953125,
0.6162109375,
0.46728515625,
0.6162109375,
0.46728515625,
0.626953125,
0.08251953125,
0.626953125
],
[
0.52392578125,
0.6123046875,
0.7939453125,
0.6123046875,
0.7939453125,
0.626953125,
0.52392578125,
0.626953125
],
[
0.09234619140625,
0.62744140625,
0.451904296875,
0.62548828125,
0.451904296875,
0.64306640625,
0.09234619140625,
0.64599609375
],
[
0.521484375,
0.62744140625,
0.890625,
0.626953125,
0.890625,
0.64453125,
0.521484375,
0.64599609375
],
[
0.08251953125,
0.64453125,
0.46728515625,
0.64453125,
0.46728515625,
0.6591796875,
0.08251953125,
0.6591796875
],
[
0.5048828125,
0.64453125,
0.890625,
0.64453125,
0.890625,
0.65771484375,
0.5048828125,
0.65771484375
],
[
0.08087158203125,
0.6591796875,
0.468017578125,
0.6591796875,
0.468017578125,
0.673828125,
0.08087158203125,
0.673828125
],
[
0.5048828125,
0.66015625,
0.88232421875,
0.66015625,
0.88232421875,
0.673828125,
0.5048828125,
0.673828125
],
[
0.080078125,
0.67236328125,
0.468017578125,
0.671875,
0.468017578125,
0.689453125,
0.080078125,
0.68994140625
],
[
0.87841796875,
0.66650390625,
0.88623046875,
0.66650390625,
0.88623046875,
0.67041015625,
0.87841796875,
0.67041015625
],
[
0.50244140625,
0.673828125,
0.890625,
0.6748046875,
0.890625,
0.689453125,
0.50244140625,
0.68798828125
],
[
0.08087158203125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.08087158203125,
0.7041015625
],
[
0.5048828125,
0.68994140625,
0.8896484375,
0.68994140625,
0.8896484375,
0.7041015625,
0.5048828125,
0.7041015625
],
[
0.08251953125,
0.70556640625,
0.468017578125,
0.70556640625,
0.468017578125,
0.7197265625,
0.08251953125,
0.7197265625
],
[
0.50244140625,
0.7021484375,
0.890625,
0.7041015625,
0.890625,
0.7216796875,
0.50244140625,
0.7197265625
],
[
0.08087158203125,
0.72021484375,
0.46728515625,
0.72021484375,
0.46728515625,
0.734375,
0.08087158203125,
0.734375
],
[
0.5,
0.71630859375,
0.890625,
0.71728515625,
0.890625,
0.73779296875,
0.5,
0.7373046875
],
[
0.08087158203125,
0.736328125,
0.468017578125,
0.736328125,
0.468017578125,
0.75,
0.08087158203125,
0.75
],
[
0.50634765625,
0.73486328125,
0.890625,
0.73486328125,
0.890625,
0.74853515625,
0.50634765625,
0.74853515625
],
[
0.08251953125,
0.75,
0.46728515625,
0.75,
0.46728515625,
0.76318359375,
0.08251953125,
0.76318359375
],
[
0.50341796875,
0.75048828125,
0.779296875,
0.75048828125,
0.779296875,
0.7646484375,
0.50341796875,
0.7646484375
],
[
0.08251953125,
0.767578125,
0.46728515625,
0.767578125,
0.46728515625,
0.77783203125,
0.08251953125,
0.77783203125
],
[
0.50341796875,
0.7666015625,
0.890625,
0.7666015625,
0.890625,
0.77978515625,
0.50341796875,
0.77978515625
],
[
0.08087158203125,
0.78076171875,
0.468017578125,
0.78076171875,
0.468017578125,
0.79443359375,
0.08087158203125,
0.79443359375
],
[
0.5048828125,
0.78271484375,
0.8896484375,
0.78271484375,
0.8896484375,
0.79345703125,
0.5048828125,
0.79345703125
],
[
0.08087158203125,
0.79541015625,
0.281005859375,
0.79541015625,
0.281005859375,
0.8095703125,
0.08087158203125,
0.8095703125
],
[
0.5048828125,
0.79541015625,
0.890625,
0.79541015625,
0.890625,
0.81005859375,
0.5048828125,
0.81005859375
],
[
0.5048828125,
0.81103515625,
0.890625,
0.81103515625,
0.890625,
0.82470703125,
0.5048828125,
0.82470703125
],
[
0.080078125,
0.82568359375,
0.361083984375,
0.82568359375,
0.361083984375,
0.8388671875,
0.080078125,
0.8388671875
],
[
0.50341796875,
0.82470703125,
0.890625,
0.82568359375,
0.890625,
0.84033203125,
0.50341796875,
0.8388671875
],
[
0.08251953125,
0.84033203125,
0.46728515625,
0.84033203125,
0.46728515625,
0.85498046875,
0.08251953125,
0.85498046875
],
[
0.5048828125,
0.84130859375,
0.88818359375,
0.84130859375,
0.88818359375,
0.85498046875,
0.5048828125,
0.85498046875
],
[
0.080078125,
0.85595703125,
0.468017578125,
0.85498046875,
0.468017578125,
0.869140625,
0.080078125,
0.87060546875
],
[
0.50341796875,
0.85498046875,
0.8896484375,
0.85693359375,
0.8896484375,
0.87060546875,
0.50341796875,
0.86865234375
],
[
0.08251953125,
0.8720703125,
0.468017578125,
0.8720703125,
0.468017578125,
0.88525390625,
0.08251953125,
0.88525390625
],
[
0.50341796875,
0.8720703125,
0.890625,
0.8720703125,
0.890625,
0.88525390625,
0.50341796875,
0.88525390625
],
[
0.080078125,
0.88525390625,
0.468017578125,
0.88623046875,
0.468017578125,
0.90087890625,
0.080078125,
0.8994140625
],
[
0.50341796875,
0.88623046875,
0.56884765625,
0.88623046875,
0.56884765625,
0.90087890625,
0.50341796875,
0.90087890625
]
] |
[
"One-shot source video finetuning",
"嵌入J*，并使用CLIP-文本分支获得文本嵌人C*",
"作为膨胀过程的一部分，空间自我注意被膨胀为时空",
"Ilo()-e()Il",
"注意（ST-attn），并且在ST-attn和交叉注意块之后引",
"人了额外的时间注意（T-attn）。见补充。有关详细信",
"X",
"Mo-",
"息",
"Noi",
"3.2.InvEdit mask generation",
"addition",
"在本节中，我们将使用Sec的微调扩散模型描述",
"CLIP-Vision",
"InvEdit的详细信息-我们新颖的零射，目标图像和形",
"",
"状感知遮罩生成策略。3.1。现有方法[31，46]仅使用",
"CLIP-Text",
"源视频计算蒙版，因此对目标对象的相对形状和大小",
"没有形状感知（例如，将汽车换成公共汽车）。InvEdit",
"<prompt>",
"Inferencepipeline",
"是针对视频的DifEdit[6]的改编，包括目标图像和形",
"OInvEdit",
"?Latent correction",
"状感知",
"InvEdit步骤：首先，我们执行DDIM反转[38l，将源视",
"Mo",
"频转换为相应的随机潜噪T=[，，",
"。然后，我们使用来自Sec的微调膨胀SD-unCLIP模",
"型对潜在代码Zsrc进行去噪。3.1使用确定性DDIM",
"采样源和目标分支，如图所示。3。对于源分支，我们",
"使用源文本提示PSTC和从VSTC中随机选择的顿ISrc",
"作为微调SD-unCLIP模型的条件输人来对zsrc进行",
"VAE",
"降噪。同样，并行地，我们使用目标文本提示Ptrg和目",
"Legend",
"Inflated attention layers",
"S.D.Unet",
"标图像Itrg中的前景对象作为条件输人来重复DDIM",
"采样。我们使用GroundedSAM[17，20]从目标图像中",
"",
"Figure 2. Overview of GenVideo . Inflated attention layers",
"are finetuned during source video finetuning. During infer-",
"分割出前景对象",
"ence, InvEdit predicts a region to edit and latent correction",
"我们通过在源和目标分支中的每个去噪时间步长",
"uses that mask to improve the inter-frame temporal consis-",
"获得的去噪UNet模型（由Ee表示）来计算预测",
"tency. Ms - “\"no mask\"",
"噪声的差。更正式地说，对于Vsrc中的每个Isr",
"一个“目标图像和形状感知”视频编辑来解决这些限",
"，我们计算△et.n=abs(etre-etrg），其中etn",
"制，仅需30分钟的训练和2分钟的推理，使用局部掩",
"eo（zun，t,cu,Ju),u(src,trg)。在每个DDIM去",
"码指导来避免背景变化",
"噪步骤中，我们获得",
"3. Approach",
"z$rc =DDIM(zt+i, e+i), ztrg =DDIM(zt1,eg)",
"(1)",
"GenVideo旨在根据目标文本提示和包含任何形状对",
"象的目标图像来编辑给定的源视频，同时保持时间",
"其中teT-1,.,1]表示时间步长",
"一致性。更正式地说，给定一个输入源视频Vsre",
"这些差异△etn，在图中表示为热图。3在多个",
"[ISrc,..,Isrc]，该输人源视频由包含源对象的N顿",
"去噪时间步长上进行平均，并进行二值化，以获",
"描述源视频的源文本提示Psrc，包含目标对象的目标",
"得视频每顿的目标感知InvEdit掩码。InFig。3",
"图像Itrg和描述源视频所需编辑的目标文本提示Ptrg",
"",
"InvEdit掩码能够确定将编辑放置在形状像卡车而不",
"组成。整个训练和推理管道总结在图。2。首先，我们",
"是汽车的区域中，因为卡车比汽车大得多。我们用",
"使用标准LDMs[35]的重建损失在源视频上微调膨胀",
"Min=[Mi,M2，.,MN]表示N顿的掩码，其中",
"的稳定扩散unCLIP（SD-unCLIP）模型[33，46]］（秒。",
"Mn=binarize(meante[tt)(Aet.n)。InvEdit掩码用",
"3.1）。然后，我们采用了新颖的目标图像和形状感知遮",
"于标识将目标图像嵌人与时间步长嵌入[33]一起注入",
"罩生成方法，称为InvEdit，在该方法中，我们使用微",
"到UNet的ResNet块的特征中的区域",
"调模型来推断需要本地化编辑的感兴趣区域（Sec。3.2",
"InvEdit直觉：我们的直觉遵循DiffEdit[6]进行文本",
")。最后，我们介绍了一种新颖的潜在校正方法来提高",
"驱动的图像编辑。我们将其扩展为目标图像感知视频",
"顿间时间一致性（Sec。3.3）",
"编辑用例。在 DDIM 去噪期间，SD-unCLIP 模型将在",
"不同的文本和图像条件下产生不同的噪声估计。在潜",
"3.1. Fine-tuning on the source video",
"在代码最终将根据条件解码不同形状，颜色和纹理的",
"膨胀的预训练T2I扩散模型的微调程序符合TAV方",
"区域中，噪声估计将有所不同。另一方面，对于背景",
"法[43]】（图。2）。与[43]不同，我们使用SD-unCLIP",
"噪声估计几乎没有变化。因此，可以使用噪声估计之间",
"模型「33】，它在目标图像和文本提示「46]上生成条件。",
"的差异来推断掩码，该掩码标识需要编辑的每个视频",
"该模型使用CLIP-视觉分支获得输人参考图像的图像",
"顿的部分"
] |
{
"arxiv_id": "2404.12541",
"language": "zh",
"page_idx": 2
}
|
||
2404.12541-zh-4
|
2404.12541-zh-4.jpg
|
the masked area [39] . This is achieved by retaining the clean latent corresponding to the source video frame outside the masked region as given by $\tilde{z}_{t,i}=$ $( 1- M_{i}) \odot \mathcal{E} ( I_{i}^{src}) + ( M_{i}) \odot \tilde{z} _{t, i}$ ,where $\mathcal{E}$ is the encoder of VAE. We skip this step when the background of the target video is expected to be different from the source.
为什么要使用 Up-block-2? 我们使用 Up-block-2 的功能，因为它在 DDIM 步骤后获得的潜在噪声中显示出比 CE 低的对应误差 (CE)。继续图中汽车的示例。4 ,我们首先使用源视频帧上的 RAFT 光流[40]计算连续帧中的特征对应关系。这用作此示例的基本事实对应关系，因为编辑的对象与源对象具有相同的形状。然后，我们计算连续帧中 Up-block-2 特征的对应关系， 并发现这些特征具有较低的 CE 率，如图中的热图所示。4 D 。另一方面，对于在 DDIM 步骤之后计算的潜在噪声，CE 率更高 (图。4 E )。通过我们的潜在校正策略，CE 降低，因为所提出的混合策略提高了连续帧中潜在噪声特征的一致性，如图 3 所示。4 C 和图。4 F 。我们在实验中表明，此过程提高了已编辑目标视频的时间一致性
4. Experiments
4.1. Implementation details
对于我们所有的实验，我们使用稳定扩散 v2 的 SDunCLIP 版本。1T2I 型号 [33] 。我们像在 [43] 中那样用时间层来膨胀模型。有关模型通胀的更多详细信息，请参阅 SM。在源微调期间，我们使用 16 源视频帧训练膨胀层，400 迭代的学习速率为 1$\times10^-5$。我们在推理过程中使用 DDIM 采样器，通过计算$\Delta\epsilon_t$在$t\in[0.8\times T,T]$上的平均值来获得 DDIM 反转源视频上的 InvEdit 掩码，其中$T=50$并使用差异阈值0。6将热图二值化。在潜在校正期间，我们将最近邻搜索限制为像素的$4\times4$窗口，并在实验中设置$w_- 1= 0. 1$ , $w_0=0.8$和$w_1=0.1$ 。在训练和推理过程中，我们还将无分类器指导量表固定为 12.5
## 4.2. Applications
形状感知视频对象编辑：我们的方法可用于进行目标图像和形状感知视频对象编辑。Fig。5,以及图中的第二个系列。1 ,展示了使用 GenVideo 进行时间一致的视频对象编辑的结果。这个任务是困难的，因为掩模需要目标图像和形状感知，以有效地识别感兴趣的区域进行本地化编辑。例如，在男子$\longrightarrow$火影忍者，男子$\longrightarrow$熊猫和银色汽车$\longrightarrow$白色巴士的示例中，InvEdit 掩码能够分别准确识别火影忍者，熊猫和白色巴士的掩码。此外，我们的潜在校正策略能够在推理过程中即时校正 UNet 潜伏期。。
零拍图像编辑：此外，我们的方法还能够对不同形状和大小的对象进行零拍图像编辑。这可以在图中观察到。6,它演示了图像中对象形状的变化(如船$\longrightarrow$帆船)。我们提供的 SM 中提供了其他示例和结果
Table 1. Quantitative comparison of the state-of-the-art video editing methods. For the user study, Text , Image , and Visual refer to the average rank in target text alignment, target-image alignment, and visual quality of the edited video respectively.
<table>
<tbody>
<tr>
<th rowspan="2">Method</th>
<th colspan="3">Model scoring metrics $\uparrow$</th>
<th colspan="3">Study UserS $(\downarrow$,</th>
</tr>
<tr>
<th>$CLIP$ $\because-1$ 1 </th>
<th>$DINO$</th>
<th>$\operatorname{Temp}$</th>
<th>|Text</th>
<th>Image</th>
<th>Visual T</th>
</tr>
<tr>
<td>TAV 43</td>
<td>0.238</td>
<td>0.236</td>
<td>0.957</td>
<td>3.6</td>
<td>3.3</td>
<td>4.2</td>
</tr>
<tr>
<td>StableVideo 5</td>
<td>0.234</td>
<td>0.189</td>
<td>0.980</td>
<td>4.3</td>
<td>4.3</td>
<td>3.7</td>
</tr>
<tr>
<td>TokenFlow 0</td>
<td>0.231</td>
<td>0.216</td>
<td>0.985</td>
<td>3.3</td>
<td>3.8</td>
<td>2.1</td>
</tr>
<tr>
<td>FateZero $\begin{bmatrix}31&0\\0&0\end{bmatrix}$</td>
<td>0.235</td>
<td>0.262</td>
<td>0.951</td>
<td>3.9</td>
<td>3.6</td>
<td>3.4</td>
</tr>
<tr>
<td>Make- a- Pro 46 J</td>
<td>0.234</td>
<td>0.195</td>
<td>0.949</td>
<td>4.0</td>
<td>4.1</td>
<td>5.0</td>
</tr>
<tr>
<td>Ours GenVideo</td>
<td>0.241</td>
<td>0.374</td>
<td>0.967</td>
<td>1.7</td>
<td>1.8</td>
<td>2.3</td>
</tr>
</tbody>
</table>
# 4.3. Comparison with baselines
我们与以下五种现有方法进行定性和定量比较：1) Tune-A-Video (TAV)[43] , 2) StableVideo [5] , 3) TokenFlow [9] , 4)FateZero [31]和 5) Make-A-Protagoner (MAP)。我们仅将目标参考图像与我们的方法和[46] 一起使用
定性结果：图。7 显示了我们的 GenVideo 方法与上述基线的比较。TAV [43] 中的编辑不是目标图像感知的，并且由于缺少掩码而无法本地化。因此，它不能保证对目标图像的保真度，并且最终会大幅修改背景。StableVideo 无法识别目标图像，因此无法生成目标图像的难以描述的特征。虽然它使用 NLA 提供时间一致性，但这些图集仅使用源视频进行计算，并且没有目标对象的意识。因此，当目标对象具有不同的形状和大小时，它们会失败。TokenFlow 也不为不同形状和大小的目标对象提供时间一致性，因为它们使用的最近邻域仅基于源视频帧。此外，TokenFlow 和 StableVideo 都不了解目标图像，因此不能很好地与目标图像对齐。FateZero 可以在一定程度上改变形状，但由于他们的面具再次基于源视频，他们不能在形状上产生大的变化 (如汽车到公共汽车)。MAP 还提供基于源视频的蒙版指导，因此不能大幅改变形状。当目标对象具有完全不同的形状和大小，同时保持对目标图像的保真度时，我们的结果执行时间一致的视频编辑
定量结果：表 1 显示了我们的方法与最新 (SOTA) 方法之间的定量比较。我们使用经过训练的 CLIP [32] 模型来计算生成的视频帧与目标文本提示的平均相似度， 在表 1 中表示为 CLIP-T。为了评估时间一致性，我们使用 CLIP 模型来计算编辑视频的连续帧对的平均帧间相似性得分。我们在表 1 中将此分数表示为温度。为了测量与目标图像的对齐，我们使用经过训练的 DINO [3]来计算视频帧与目标图像之间的相似性。我们还进行了一项用户研究，以评估三个指标的编辑视频：目标文本对齐，目标图像对齐和编辑视频的视觉质量。我们要求 20 名受试者对六种方法的编辑视频进行排名 (i。e。,我们和上面提到的五个基线)在上述三个指标上， 通过询问每个受试者 9 组比较。结果列于表 1 。这些结果表明，我们提出的方法在目标文本和目标图像对齐方面都优于现有方法。这反映在更高的 CLIP-T 和 DINO 分数以及用户研究中，GenVideo 在文本和图像对齐方
|
[
[
0.1004638671875,
0.091552734375,
0.468017578125,
0.0921630859375,
0.468017578125,
0.106689453125,
0.1004638671875,
0.1060791015625
],
[
0.5048828125,
0.091552734375,
0.57275390625,
0.091552734375,
0.57275390625,
0.10162353515625,
0.5048828125,
0.10162353515625
],
[
0.56884765625,
0.091552734375,
0.8896484375,
0.091552734375,
0.8896484375,
0.10162353515625,
0.56884765625,
0.10162353515625
],
[
0.10296630859375,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.10296630859375,
0.121826171875
],
[
0.50244140625,
0.10418701171875,
0.890625,
0.10418701171875,
0.890625,
0.117431640625,
0.50244140625,
0.117431640625
],
[
0.101318359375,
0.12371826171875,
0.46728515625,
0.12371826171875,
0.46728515625,
0.136962890625,
0.101318359375,
0.136962890625
],
[
0.50244140625,
0.11871337890625,
0.890625,
0.117431640625,
0.890625,
0.1307373046875,
0.50244140625,
0.1319580078125
],
[
0.50341796875,
0.1319580078125,
0.89208984375,
0.1319580078125,
0.89208984375,
0.1458740234375,
0.50341796875,
0.1458740234375
],
[
0.10296630859375,
0.1383056640625,
0.468017578125,
0.1383056640625,
0.468017578125,
0.15283203125,
0.10296630859375,
0.15283203125
],
[
0.1004638671875,
0.15283203125,
0.4697265625,
0.1514892578125,
0.4697265625,
0.1666259765625,
0.1004638671875,
0.1673583984375
],
[
0.5048828125,
0.147705078125,
0.69384765625,
0.147705078125,
0.69384765625,
0.1578369140625,
0.5048828125,
0.1578369140625
],
[
0.609375,
0.1666259765625,
0.765625,
0.1673583984375,
0.765625,
0.1817626953125,
0.609375,
0.18115234375
],
[
0.779296875,
0.1654052734375,
0.875,
0.1673583984375,
0.875,
0.1817626953125,
0.779296875,
0.179931640625
],
[
0.5107421875,
0.1749267578125,
0.5595703125,
0.1749267578125,
0.5595703125,
0.186279296875,
0.5107421875,
0.186279296875
],
[
0.101318359375,
0.1685791015625,
0.468017578125,
0.1685791015625,
0.468017578125,
0.1817626953125,
0.101318359375,
0.1817626953125
],
[
0.72119140625,
0.184326171875,
0.7607421875,
0.186279296875,
0.7587890625,
0.1988525390625,
0.7197265625,
0.1962890625
],
[
0.76708984375,
0.184326171875,
0.8857421875,
0.1849365234375,
0.8857421875,
0.199462890625,
0.76708984375,
0.1988525390625
],
[
0.61376953125,
0.186279296875,
0.71337890625,
0.186279296875,
0.71337890625,
0.1976318359375,
0.61376953125,
0.1976318359375
],
[
0.101318359375,
0.1849365234375,
0.1502685546875,
0.1849365234375,
0.1502685546875,
0.1956787109375,
0.101318359375,
0.1956787109375
],
[
0.5087890625,
0.2020263671875,
0.56884765625,
0.203857421875,
0.56787109375,
0.2183837890625,
0.50732421875,
0.216552734375
],
[
0.62353515625,
0.2052001953125,
0.65673828125,
0.2052001953125,
0.65673828125,
0.216552734375,
0.62353515625,
0.216552734375
],
[
0.67333984375,
0.2052001953125,
0.7099609375,
0.2052001953125,
0.7099609375,
0.216552734375,
0.67333984375,
0.216552734375
],
[
0.72216796875,
0.2052001953125,
0.75732421875,
0.2052001953125,
0.75732421875,
0.216552734375,
0.72216796875,
0.216552734375
],
[
0.810546875,
0.2052001953125,
0.83251953125,
0.2052001953125,
0.83251953125,
0.2171630859375,
0.810546875,
0.2171630859375
],
[
0.8544921875,
0.2020263671875,
0.8759765625,
0.205810546875,
0.87158203125,
0.2183837890625,
0.8505859375,
0.2139892578125
],
[
0.08251953125,
0.208984375,
0.468017578125,
0.208984375,
0.468017578125,
0.222900390625,
0.08251953125,
0.222900390625
],
[
0.775390625,
0.205810546875,
0.7958984375,
0.205810546875,
0.7958984375,
0.216552734375,
0.775390625,
0.216552734375
],
[
0.6748046875,
0.21533203125,
0.7099609375,
0.2171630859375,
0.708984375,
0.2298583984375,
0.67333984375,
0.2279052734375
],
[
0.77392578125,
0.216552734375,
0.7958984375,
0.216552734375,
0.7958984375,
0.2291259765625,
0.77392578125,
0.2291259765625
],
[
0.810546875,
0.216552734375,
0.83251953125,
0.216552734375,
0.83251953125,
0.2291259765625,
0.810546875,
0.2291259765625
],
[
0.85302734375,
0.216552734375,
0.875,
0.216552734375,
0.875,
0.2298583984375,
0.85302734375,
0.2298583984375
],
[
0.62353515625,
0.2171630859375,
0.65673828125,
0.2171630859375,
0.65673828125,
0.2291259765625,
0.62353515625,
0.2291259765625
],
[
0.51123046875,
0.2171630859375,
0.60107421875,
0.2171630859375,
0.60107421875,
0.2291259765625,
0.51123046875,
0.2291259765625
],
[
0.7236328125,
0.2171630859375,
0.75732421875,
0.2171630859375,
0.75732421875,
0.2291259765625,
0.7236328125,
0.2291259765625
],
[
0.08251953125,
0.2259521484375,
0.465576171875,
0.2259521484375,
0.465576171875,
0.236083984375,
0.08251953125,
0.236083984375
],
[
0.8095703125,
0.2291259765625,
0.83251953125,
0.2291259765625,
0.83251953125,
0.242431640625,
0.8095703125,
0.242431640625
],
[
0.85302734375,
0.2291259765625,
0.875,
0.2291259765625,
0.875,
0.242431640625,
0.85302734375,
0.242431640625
],
[
0.62353515625,
0.2298583984375,
0.65625,
0.2298583984375,
0.65625,
0.2412109375,
0.62353515625,
0.2412109375
],
[
0.5107421875,
0.2298583984375,
0.59716796875,
0.2298583984375,
0.59716796875,
0.2412109375,
0.5107421875,
0.2412109375
],
[
0.6748046875,
0.2298583984375,
0.708984375,
0.2298583984375,
0.708984375,
0.2412109375,
0.6748046875,
0.2412109375
],
[
0.7236328125,
0.2298583984375,
0.75732421875,
0.2298583984375,
0.75732421875,
0.242431640625,
0.7236328125,
0.242431640625
],
[
0.77392578125,
0.2298583984375,
0.7958984375,
0.2298583984375,
0.7958984375,
0.242431640625,
0.77392578125,
0.242431640625
],
[
0.08087158203125,
0.2392578125,
0.4697265625,
0.2392578125,
0.4697265625,
0.25244140625,
0.08087158203125,
0.25244140625
],
[
0.5107421875,
0.242431640625,
0.5908203125,
0.242431640625,
0.5908203125,
0.25390625,
0.5107421875,
0.25390625
],
[
0.62353515625,
0.242431640625,
0.65673828125,
0.242431640625,
0.65673828125,
0.25390625,
0.62353515625,
0.25390625
],
[
0.6748046875,
0.242431640625,
0.708984375,
0.242431640625,
0.708984375,
0.25390625,
0.6748046875,
0.25390625
],
[
0.7236328125,
0.242431640625,
0.75634765625,
0.242431640625,
0.75634765625,
0.255126953125,
0.7236328125,
0.255126953125
],
[
0.810546875,
0.242431640625,
0.83251953125,
0.242431640625,
0.83251953125,
0.255126953125,
0.810546875,
0.255126953125
],
[
0.85302734375,
0.2412109375,
0.875,
0.2412109375,
0.875,
0.255126953125,
0.85302734375,
0.255126953125
],
[
0.09478759765625,
0.25390625,
0.468017578125,
0.25244140625,
0.468017578125,
0.267578125,
0.09478759765625,
0.268310546875
],
[
0.775390625,
0.24365234375,
0.7958984375,
0.24365234375,
0.7958984375,
0.25390625,
0.775390625,
0.25390625
],
[
0.5107421875,
0.25390625,
0.60693359375,
0.255126953125,
0.60693359375,
0.266357421875,
0.5107421875,
0.26513671875
],
[
0.62353515625,
0.255126953125,
0.65673828125,
0.255126953125,
0.65673828125,
0.266357421875,
0.62353515625,
0.266357421875
],
[
0.6748046875,
0.255126953125,
0.7099609375,
0.255126953125,
0.7099609375,
0.266357421875,
0.6748046875,
0.266357421875
],
[
0.7236328125,
0.255126953125,
0.75634765625,
0.255126953125,
0.75634765625,
0.267578125,
0.7236328125,
0.267578125
],
[
0.775390625,
0.255126953125,
0.7958984375,
0.255126953125,
0.7958984375,
0.267578125,
0.775390625,
0.267578125
],
[
0.81201171875,
0.255126953125,
0.8310546875,
0.255126953125,
0.8310546875,
0.267578125,
0.81201171875,
0.267578125
],
[
0.85302734375,
0.25390625,
0.875,
0.25390625,
0.875,
0.267578125,
0.85302734375,
0.267578125
],
[
0.0882568359375,
0.260009765625,
0.09722900390625,
0.260009765625,
0.09722900390625,
0.26318359375,
0.0882568359375,
0.26318359375
],
[
0.72216796875,
0.26513671875,
0.75732421875,
0.267578125,
0.75634765625,
0.27978515625,
0.72119140625,
0.27783203125
],
[
0.81201171875,
0.264404296875,
0.833984375,
0.267578125,
0.830078125,
0.2822265625,
0.80810546875,
0.27783203125
],
[
0.85302734375,
0.266357421875,
0.87744140625,
0.266357421875,
0.87744140625,
0.27978515625,
0.85302734375,
0.27978515625
],
[
0.62158203125,
0.267578125,
0.65673828125,
0.267578125,
0.65673828125,
0.279052734375,
0.62158203125,
0.279052734375
],
[
0.5087890625,
0.267578125,
0.6044921875,
0.267578125,
0.6044921875,
0.27783203125,
0.5087890625,
0.27783203125
],
[
0.67333984375,
0.267578125,
0.7099609375,
0.267578125,
0.7099609375,
0.279052734375,
0.67333984375,
0.279052734375
],
[
0.77294921875,
0.267578125,
0.7958984375,
0.267578125,
0.7958984375,
0.27978515625,
0.77294921875,
0.27978515625
],
[
0.08087158203125,
0.285400390625,
0.46728515625,
0.285400390625,
0.46728515625,
0.298583984375,
0.08087158203125,
0.298583984375
],
[
0.08087158203125,
0.26953125,
0.468017578125,
0.26953125,
0.468017578125,
0.28271484375,
0.08087158203125,
0.28271484375
],
[
0.50341796875,
0.294189453125,
0.7490234375,
0.294189453125,
0.7490234375,
0.30810546875,
0.50341796875,
0.30810546875
],
[
0.08251953125,
0.3017578125,
0.46728515625,
0.3017578125,
0.46728515625,
0.3125,
0.08251953125,
0.3125
],
[
0.08087158203125,
0.31494140625,
0.468017578125,
0.31494140625,
0.468017578125,
0.328857421875,
0.08087158203125,
0.328857421875
],
[
0.50341796875,
0.3125,
0.89208984375,
0.3125,
0.89208984375,
0.32568359375,
0.50341796875,
0.32568359375
],
[
0.08087158203125,
0.330078125,
0.468017578125,
0.330078125,
0.468017578125,
0.343505859375,
0.08087158203125,
0.343505859375
],
[
0.50341796875,
0.32568359375,
0.890625,
0.32568359375,
0.890625,
0.339111328125,
0.50341796875,
0.339111328125
],
[
0.080078125,
0.343505859375,
0.468017578125,
0.3447265625,
0.468017578125,
0.359130859375,
0.080078125,
0.35791015625
],
[
0.50341796875,
0.34228515625,
0.8896484375,
0.34228515625,
0.8896484375,
0.35595703125,
0.50341796875,
0.35595703125
],
[
0.08251953125,
0.35986328125,
0.46728515625,
0.35986328125,
0.46728515625,
0.373779296875,
0.08251953125,
0.373779296875
],
[
0.5009765625,
0.356689453125,
0.8896484375,
0.356689453125,
0.8896484375,
0.37060546875,
0.5009765625,
0.37060546875
],
[
0.080078125,
0.373779296875,
0.468017578125,
0.375,
0.468017578125,
0.389404296875,
0.080078125,
0.38818359375
],
[
0.5146484375,
0.37255859375,
0.5703125,
0.37255859375,
0.5703125,
0.386962890625,
0.5146484375,
0.386962890625
],
[
0.08087158203125,
0.39013671875,
0.468017578125,
0.39013671875,
0.468017578125,
0.404052734375,
0.08087158203125,
0.404052734375
],
[
0.5048828125,
0.392578125,
0.890625,
0.392578125,
0.890625,
0.406005859375,
0.5048828125,
0.406005859375
],
[
0.08087158203125,
0.406005859375,
0.2459716796875,
0.406005859375,
0.2459716796875,
0.419189453125,
0.08087158203125,
0.419189453125
],
[
0.50341796875,
0.4072265625,
0.8896484375,
0.4072265625,
0.8896484375,
0.42041015625,
0.50341796875,
0.42041015625
],
[
0.5048828125,
0.42236328125,
0.890625,
0.42236328125,
0.890625,
0.436279296875,
0.5048828125,
0.436279296875
],
[
0.080078125,
0.43505859375,
0.21728515625,
0.43505859375,
0.21728515625,
0.448974609375,
0.080078125,
0.448974609375
],
[
0.5048828125,
0.4375,
0.890625,
0.4375,
0.890625,
0.45068359375,
0.5048828125,
0.45068359375
],
[
0.50244140625,
0.45068359375,
0.8896484375,
0.451904296875,
0.8896484375,
0.46533203125,
0.50244140625,
0.464111328125
],
[
0.08087158203125,
0.458251953125,
0.29736328125,
0.458251953125,
0.29736328125,
0.4716796875,
0.08087158203125,
0.4716796875
],
[
0.5048828125,
0.46728515625,
0.890625,
0.46728515625,
0.890625,
0.48095703125,
0.5048828125,
0.48095703125
],
[
0.08087158203125,
0.47607421875,
0.46728515625,
0.47607421875,
0.46728515625,
0.4892578125,
0.08087158203125,
0.4892578125
],
[
0.50341796875,
0.481689453125,
0.890625,
0.481689453125,
0.890625,
0.49560546875,
0.50341796875,
0.49560546875
],
[
0.08251953125,
0.4912109375,
0.468017578125,
0.4912109375,
0.468017578125,
0.5048828125,
0.08251953125,
0.5048828125
],
[
0.50341796875,
0.49755859375,
0.890625,
0.49755859375,
0.890625,
0.51123046875,
0.50341796875,
0.51123046875
],
[
0.08087158203125,
0.505859375,
0.468017578125,
0.505859375,
0.468017578125,
0.51953125,
0.08087158203125,
0.51953125
],
[
0.50341796875,
0.51220703125,
0.890625,
0.51220703125,
0.890625,
0.52587890625,
0.50341796875,
0.52587890625
],
[
0.08251953125,
0.521484375,
0.468017578125,
0.521484375,
0.468017578125,
0.53515625,
0.08251953125,
0.53515625
],
[
0.50341796875,
0.52783203125,
0.890625,
0.52783203125,
0.890625,
0.54150390625,
0.50341796875,
0.54150390625
],
[
0.08087158203125,
0.53515625,
0.46728515625,
0.53515625,
0.46728515625,
0.548828125,
0.08087158203125,
0.548828125
],
[
0.50341796875,
0.54345703125,
0.890625,
0.54345703125,
0.890625,
0.556640625,
0.50341796875,
0.556640625
],
[
0.08251953125,
0.55029296875,
0.46728515625,
0.55029296875,
0.46728515625,
0.564453125,
0.08251953125,
0.564453125
],
[
0.50341796875,
0.55810546875,
0.890625,
0.55810546875,
0.890625,
0.5712890625,
0.50341796875,
0.5712890625
],
[
0.08087158203125,
0.56640625,
0.468017578125,
0.56640625,
0.468017578125,
0.580078125,
0.08087158203125,
0.580078125
],
[
0.50244140625,
0.5712890625,
0.890625,
0.57373046875,
0.890625,
0.5869140625,
0.50244140625,
0.58544921875
],
[
0.0849609375,
0.58056640625,
0.468017578125,
0.58056640625,
0.468017578125,
0.5947265625,
0.0849609375,
0.5947265625
],
[
0.5048828125,
0.58837890625,
0.890625,
0.58837890625,
0.890625,
0.6015625,
0.5048828125,
0.6015625
],
[
0.08251953125,
0.5966796875,
0.468017578125,
0.5966796875,
0.468017578125,
0.60986328125,
0.08251953125,
0.60986328125
],
[
0.5048828125,
0.60302734375,
0.890625,
0.60302734375,
0.890625,
0.6162109375,
0.5048828125,
0.6162109375
],
[
0.08251953125,
0.6123046875,
0.4697265625,
0.6123046875,
0.4697265625,
0.62548828125,
0.08251953125,
0.62548828125
],
[
0.50244140625,
0.6171875,
0.890625,
0.61865234375,
0.890625,
0.6318359375,
0.50244140625,
0.630859375
],
[
0.080078125,
0.626953125,
0.468017578125,
0.62548828125,
0.468017578125,
0.64013671875,
0.080078125,
0.6416015625
],
[
0.5048828125,
0.63525390625,
0.8896484375,
0.63525390625,
0.8896484375,
0.64599609375,
0.5048828125,
0.64599609375
],
[
0.08087158203125,
0.642578125,
0.31201171875,
0.642578125,
0.31201171875,
0.65576171875,
0.08087158203125,
0.65576171875
],
[
0.50341796875,
0.6484375,
0.79736328125,
0.6484375,
0.79736328125,
0.662109375,
0.50341796875,
0.662109375
],
[
0.50341796875,
0.66015625,
0.890625,
0.66015625,
0.890625,
0.6748046875,
0.50341796875,
0.6748046875
],
[
0.080078125,
0.67236328125,
0.2156982421875,
0.673828125,
0.2156982421875,
0.68798828125,
0.080078125,
0.6875
],
[
0.5048828125,
0.67529296875,
0.89208984375,
0.67529296875,
0.89208984375,
0.689453125,
0.5048828125,
0.689453125
],
[
0.08087158203125,
0.6923828125,
0.468017578125,
0.6923828125,
0.468017578125,
0.70556640625,
0.08087158203125,
0.70556640625
],
[
0.5048828125,
0.6923828125,
0.88623046875,
0.6923828125,
0.88623046875,
0.70263671875,
0.5048828125,
0.70263671875
],
[
0.08087158203125,
0.70556640625,
0.465576171875,
0.70556640625,
0.465576171875,
0.7197265625,
0.08087158203125,
0.7197265625
],
[
0.50341796875,
0.70556640625,
0.890625,
0.70556640625,
0.890625,
0.7197265625,
0.50341796875,
0.7197265625
],
[
0.080078125,
0.7216796875,
0.468017578125,
0.72021484375,
0.468017578125,
0.73486328125,
0.080078125,
0.736328125
],
[
0.5048828125,
0.72021484375,
0.890625,
0.72021484375,
0.890625,
0.734375,
0.5048828125,
0.734375
],
[
0.08087158203125,
0.7373046875,
0.468017578125,
0.7373046875,
0.468017578125,
0.75048828125,
0.08087158203125,
0.75048828125
],
[
0.5048828125,
0.736328125,
0.890625,
0.736328125,
0.890625,
0.75,
0.5048828125,
0.75
],
[
0.08251953125,
0.75390625,
0.465576171875,
0.75390625,
0.465576171875,
0.7646484375,
0.08251953125,
0.7646484375
],
[
0.50732421875,
0.75244140625,
0.8896484375,
0.75244140625,
0.8896484375,
0.76318359375,
0.50732421875,
0.76318359375
],
[
0.08251953125,
0.76953125,
0.46728515625,
0.76953125,
0.46728515625,
0.77978515625,
0.08251953125,
0.77978515625
],
[
0.50634765625,
0.76806640625,
0.88818359375,
0.76806640625,
0.88818359375,
0.77880859375,
0.50634765625,
0.77880859375
],
[
0.09722900390625,
0.78271484375,
0.468017578125,
0.78271484375,
0.468017578125,
0.796875,
0.09722900390625,
0.796875
],
[
0.50341796875,
0.7822265625,
0.890625,
0.7822265625,
0.890625,
0.79541015625,
0.50341796875,
0.79541015625
],
[
0.08087158203125,
0.79736328125,
0.46728515625,
0.79736328125,
0.46728515625,
0.81103515625,
0.08087158203125,
0.81103515625
],
[
0.5048828125,
0.796875,
0.890625,
0.796875,
0.890625,
0.81005859375,
0.5048828125,
0.81005859375
],
[
0.08087158203125,
0.81298828125,
0.468017578125,
0.81298828125,
0.468017578125,
0.8271484375,
0.08087158203125,
0.8271484375
],
[
0.50341796875,
0.81103515625,
0.89208984375,
0.81103515625,
0.89208984375,
0.82470703125,
0.50341796875,
0.82470703125
],
[
0.08087158203125,
0.82763671875,
0.2418212890625,
0.82763671875,
0.2418212890625,
0.84130859375,
0.08087158203125,
0.84130859375
],
[
0.50341796875,
0.8271484375,
0.890625,
0.8271484375,
0.890625,
0.84033203125,
0.50341796875,
0.84033203125
],
[
0.08251953125,
0.8447265625,
0.46728515625,
0.8447265625,
0.46728515625,
0.85498046875,
0.08251953125,
0.85498046875
],
[
0.50634765625,
0.84326171875,
0.8896484375,
0.84326171875,
0.8896484375,
0.85400390625,
0.50634765625,
0.85400390625
],
[
0.08251953125,
0.85791015625,
0.468017578125,
0.85791015625,
0.468017578125,
0.8720703125,
0.08251953125,
0.8720703125
],
[
0.50341796875,
0.85693359375,
0.8896484375,
0.85693359375,
0.8896484375,
0.87060546875,
0.50341796875,
0.87060546875
],
[
0.08087158203125,
0.87353515625,
0.38818359375,
0.87353515625,
0.38818359375,
0.88720703125,
0.08087158203125,
0.88720703125
],
[
0.381591796875,
0.875,
0.4697265625,
0.875,
0.4697265625,
0.88623046875,
0.381591796875,
0.88623046875
],
[
0.50244140625,
0.8720703125,
0.890625,
0.87060546875,
0.890625,
0.8837890625,
0.50244140625,
0.88525390625
],
[
0.08087158203125,
0.88818359375,
0.37744140625,
0.88818359375,
0.37744140625,
0.9013671875,
0.08087158203125,
0.9013671875
],
[
0.5048828125,
0.88720703125,
0.8896484375,
0.88720703125,
0.8896484375,
0.90087890625,
0.5048828125,
0.90087890625
]
] |
[
"the masked area [39] . This is achieved by retaining",
"Table 1.",
"Quantitative comparison of the state-of-the-art",
"the clean latent corresponding to the source video",
"video editing methods. For the user study, Text , Im-",
"frame outside the masked region as given by zt,i =",
"age , and Visual refer to the average rank in target text",
"alignment, target-image alignment, and visual quality of",
"(1-M,)o(Isrc)+(M,)ozt,i, where & is the encoder",
"of VAE.We skip this step when the background of",
"theedited videorespectively",
"Model scoring metrics (t)",
"User Study (t)",
"Method",
"the target video is expected to be different from the",
"Temp",
"Text Image Visual",
"CLIP-T DINO",
"source.",
"TAV [43]",
"0.238",
"0.236",
"0.957",
"3.3",
"4.2",
"为什么要使用Up-block-2？我们使用Up-block-2的功",
"3.6",
"0.189",
"4.3",
"4.3",
"3.7",
"0.234",
"StableVideo5]",
"0.980",
"能，因为它在DDIM步骤后获得的潜在噪声中显示出",
"3.8",
"2.1",
"0.231",
"TokenFlow9]",
"0.216",
"0.985",
"3.3",
"比CE低的对应误差（CE）。继续图中汽车的示例。4",
"FateZero [31]",
"0.235",
"0.262",
"0.951",
"3.6",
"3.4",
"我们首先使用源视频顿上的RAFT光流【40】计算",
"3.9",
"Make-a-Pro[46]",
"0.234",
"0.195",
"0.949",
"4.0",
"4.1",
"5.0",
"",
"0.967",
"1.8",
"2.3",
"0.241",
"Ours GenVideo",
"0.374",
"1.7",
"对应关系，因为编辑的对象与源对象具有相同的形状。",
"连续顿中的特征对应关系。这用作此示例的基本事实",
"4.3. Comparison with baselines",
"然后，我们计算连续顿中Up-block-2特征的对应关系，",
"并发现这些特征具有较低的CE率，如图中的热图所",
"我们与以下五种现有方法进行定性和定量比较：1",
"示。4D。另一方面，对于在DDIM步骤之后计算的",
"Tune-A-Video (TAV)[43],2) StableVideo [5] ,3) To-",
"潜在噪声，CE率更高（图。4E）。通过我们的潜在",
"kenFlow[9],4)FateZero[31] 和5]Make-A-Protagoner",
"校正策略，CE降低，因为所提出的混合策略提高了连",
"（MAP）。我们仅将目标参考图像与我们的方法和「46]",
"续帧中潜在噪声特征的一致性，如图3所示。4C和",
"一起使用",
"图。4F。我们在实验中表明，此过程提高了已编辑",
"定性结果：图。7显示了我们的GenVideo方法与上",
"目标视频的时间一致性",
"述基线的比较。TAV[43]中的编辑不是目标图像感知",
"的，并且由于缺少掩码而无法本地化。因此，它不能",
"4.Experiments",
"保证对目标图像的保真度，并且最终会大幅修改背景。",
"StableVideo无法识别目标图像，因此无法生成目标图",
"4.1. Implementation details",
"像的难以描述的特征。虽然它使用NLA提供时间一致",
"对于我们所有的实验，我们使用稳定扩散v2的SD-",
"性，但这些图集仅使用源视频进行计算，并且没有目标",
"unCLIP版本。1T2I型号[33]：我们像在【43】中那",
"对象的意识。因此，当目标对象具有不同的形状和大",
"样用时间层来膨胀模型。有关模型通胀的更多详细信",
"小时，它们会失败。TokenFlow也不为不同形状和大小",
"息，请参阅SM。在源微调期间，我们使用16源视频",
"的目标对象提供时间一致性，因为它们使用的最近邻",
"顿训练膨胀层，400迭代的学习速率为1×10-5。我",
"域仅基于源视频顿。此外，TokenFlow和StableVideo",
"们在推理过程中使用DDIM采样器，通过计算△et在",
"都不了解目标图像，因此不能很好地与目标图像对齐。",
"tE[0.8×T,TI上的平均值来获得DDIM反转源视频",
"FateZero可以在一定程度上改变形状，但由于他们的",
"上的InvEdit掩码，其中T=50并使用差异阈值0。6",
"面具再次基于源视频，他们不能在形状上产生大的变",
"将热图二值化。在潜在校正期间，我们将最近邻搜索限",
"化（如汽车到公共汽车）。MAP还提供基于源视频的",
"制为像素的4×4窗口，并在实验中设置w-1=0.1",
"蒙版指导，因此不能大幅改变形状。当目标对象具有完",
"Wo=0.8和W1=0.1。在训练和推理过程中，我们还",
"全不同的形状和大小，同时保持对自标图像的保真度",
"将无分类器指导量表固定为12.5",
"时，我们的结果执行时间一致的视频编辑",
"定量结果：表1显示了我们的方法与最新（SOTA）方",
"4.2.Applications",
"法之间的定量比较。我们使用经过训练的CLIP[32]模",
"形状感知视频对象编辑：我们的方法可用于进行目标",
"型来计算生成的视频顿与目标文本提示的平均相似度",
"图像和形状感知视频对象编辑。Fig。5，以及图中的",
"在表1中表示为CLIP-T。为了评估时间一致性，我们",
"第二个系列。1，展示了使用GenVideo进行时间一致",
"使用CLIP模型来计算编辑视频的连续顿对的平均",
"的视频对象编辑的结果。这个任务是困难的，因为掩模",
"间相似性得分。我们在表1中将此分数表示为温度。为",
"需要目标图像和形状感知：以有效地识别感兴趣的区",
"了测量与目标图像的对齐，我们使用经过训练的DINO",
"域进行本地化编辑。例如，在男子一一火影忍者，男子",
"3来计算视频顿与自标图像之间的相似性。我们还进",
"→熊猫和银色汽车一→白色巴士的示例中，InvEdit",
"行了一项用户研究，以评估三个指标的编辑视频：目标",
"掩码能够分别准确识别火影忍者，熊猫和白色巴士的",
"文本对齐，目标图像对齐和编辑视频的视觉质量。我们",
"掩码。此外，我们的潜在校正策略能够在推理过程中即",
"要求20名受试者对六种方法的编辑视频进行排名（i。",
"时校正UNet潜伏期。",
"e。，我们和上面提到的五个基线）在上述三个指标上",
"零拍图像编辑：此外，我们的方法还能够对不同形状和",
"通过询问每个受试者9组比较。结果列于表1。这些结",
"大小的对象进行零拍图像编辑。这可以在图中观察到。",
"果表明，我们提出的方法在目标文本和目标图像对齐方",
"6，它演示了图像中对象形状的变化（如船",
"→帆船）。",
"面都优于现有方法。这反映在更高的CLIP-T和DINO",
"我们提供的SM中提供了其他示例和结果",
"分数以及用户研究中，GenVideo在文本和图像对齐方"
] |
{
"arxiv_id": "2404.12541",
"language": "zh",
"page_idx": 4
}
|
||
2109.05782-zh-2
|
2109.05782-zh-2.jpg
|
$\bullet$TOD-BERT (Wu et al., 2020) further pre-trains BERT on a task-oriented dialogue corpus of 100,000 unlabeled samples with masked language modelling (MLM) and response contrastive objectives.
$\bullet$ USE-ConveRT (Henderson et al.,2020; Casanueva et al., 2020) investigates a dual encoder model trained with response selection tasks on 727 million input-response pairs.
$\bullet$ DNNC (Zhang et al.,2020a) pre-trains a language model with around 1 million annotated samples for natural language inference (NLI) and use the pre-trained model for intent detection.
$\cdot$ WikiHowRoBERTa (Zhang et al., 2020b)
constructs some pre-training tasks based on
the wikiHow database with 110,000 articles.
虽然这些方法取得了令人印象深刻的性能， 但它们严重依赖于大规模语料库 (Mehri et al., 2020)的存在，该语料库在语义上与目标域接近，或者由类似的任务组成以进行持续的预训本很高。更重要的是，他们完全忽略了“免费午餐”-公开可用的，高质量的，手动注释的意图检测基准。例如，数据集 OOS (Larson et al., 2019)提供跨 10 不同域的标记话语。因此，本文的研究围绕以下研究问题展开：
·Is it possible to utilize publicly available datasets to pre-train an intent detection model that can learn transferable taskspecific knowledge to generalize across different domains ?
在本文中，我们对此问题提供了肯定的答案。我们使用简单的标准监督训练对 BERT 进行微调，其中包含来自公共数据集的大约 1,000 个标记的话语，并获得一个预训练模型，称为IntentBERT。它可以直接应用于目标域上的少数意图分类，该目标域与预训练数据大不相同， 并且明显优于现有的预训练模型，而无需进一步微调目标数据(标记或未标记)。这个简单的“免费午餐”解决方案不仅证实了少拍意图检测的可行性和实用性，还提供了一个随时可用的性能良好的模型供实际使用，节省了算法设计和数据收集方面的精力。此外，IntentBERT 对跨域少射分类任务的高泛化能力，由于较大 用它通过同时优化 $\mathcal{D}_\mathrm{target}^\mathrm{unlabeled}$上的语言建模损
的域间隙和较少的数据约束，通常被认为非常困难，这表明大多数意图检测任务可能共享一个可以从一小部分数据中学习的通用底层结构
此外，为了利用目标域中的未标记数据，我们设计了一种联合预训练方案，该方案同时优化了源标记数据上的分类错误和目标未标记数据上的语言建模损失。这种联合训练方案可以学习更好的语义表示，并且明显优于现有的两阶段预训练方法 (Gururangan et al.,2020)。图中提供了由强基线和我们的方法产生的嵌入空间的可视化。1 , 它清楚地展示了我们预训练模型的优越性
## 2 $\textbf{Methodology}$
我们基于预训练的语言模型 BERT$\boxed{\mathrm{E}}$(Devlin et al.,2019) ,提出了一个用于意图分类的持续预训练框架
我们的预训练方法依赖于一个小标签数据集$\mathcal{D}_{\mathrm{source}}^{\mathrm{labeled}}=\left\{(x_{i},y_{i})\right\}$的存在，其中$y_i$ 是话语 $x_i$ 的标签。这样的数据样本可以很容易地从公共意图检测数据集中获得，如 OOS (Larson et al. 2019)和 HWU64 (Liu et al., 2021)。如实验所示，来自 OOS或 HWU64 的大致1，000示例足以使预训练的意图检测模型在完全不同的目标域(例如“Covid-19”)上实现卓越的性能
我们进一步考虑了目标域中未标记的话语
近 , 或 者 由 英 似 的 任 务 组 成 以 进 行 持 续 的 预 训 练 , 这 需 要 大 量 的 数 据 收 集 工 作 , 并 且 计 算 成 $\mathcal{D} _{\text{target}}^{\text{unlabeled}}= \{ x_i\}$可 用 的 场 景 , 并 提 出 了 一 种 联 会 硫 训 练 古 安 该 方 安 级 级 险 证 明 县 宫 审 右 效
合预训练方案，该方案经经验证明是高度有效
的
## 2.1 Supervised Pre-training
给定$\mathcal{D}_\text{source}^\mathrm{labeled}=\{(x_i,y_i)\}$与$N$不同的类，我们采用一种简单的方法来微调 BERT。具体来说， 在 BERT 的顶部附加一个线性层作为分类器， 即。e。,
$$p(y|h_i)=\mathrm{softmax}\left(\boldsymbol{W}h_i+\boldsymbol{b}\right)\in\mathbb{R}^N,$$
其中$h_i\in\mathbb{R}^d$是由$[CLS]$令牌给出的$x_i$的特征表示$,W\in\mathbb{R}^{N\times d}$和$b\in\mathbb{R}^N$是线性层的参数。模型参数$\theta=\{\phi,\boldsymbol{W},\boldsymbol{b}\}$ ,其中$\phi$是 BERT 的参数，在 $\mathcal{D}_\mathrm{source}^{\mathrm{labeled}}$上进行了交叉熵损失训练：
$$\theta^{*}=\underset{\theta}{\arg\min}\mathcal{L}_{ce}\left(\mathcal{D}_{\mathrm{source}}^{\mathrm{labeled}};\theta\right).$$
(2)
经过训练后，经过微调的 BERT 有望学习一般的意图检测技能，因此我们将其称为 Intent- $BERT$
# 2.2 Joint Pre-training 给定未标记的目标数据$\mathcal{D}_\mathrm{target}^\mathrm{unlabeled}$,我们可以利
失和方程$\mathbb{E}($2)中的监督损失来进一步增强IntentBERT。语言建模损失可以帮助学习目标域的语义表示，同时防止与源数据过度拟合
|
[
[
0.26123046875,
0.08489990234375,
0.48779296875,
0.08612060546875,
0.48779296875,
0.1009521484375,
0.26123046875,
0.09979248046875
],
[
0.53076171875,
0.0872802734375,
0.880859375,
0.0872802734375,
0.880859375,
0.1009521484375,
0.53076171875,
0.1009521484375
],
[
0.25439453125,
0.09027099609375,
0.267822265625,
0.09027099609375,
0.267822265625,
0.09442138671875,
0.25439453125,
0.09442138671875
],
[
0.139404296875,
0.0885009765625,
0.2568359375,
0.0885009765625,
0.2568359375,
0.09857177734375,
0.139404296875,
0.09857177734375
],
[
0.151123046875,
0.10272216796875,
0.48779296875,
0.1015625,
0.48779296875,
0.11639404296875,
0.151123046875,
0.1175537109375
],
[
0.51318359375,
0.10272216796875,
0.880859375,
0.10272216796875,
0.880859375,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.154541015625,
0.119384765625,
0.48779296875,
0.119384765625,
0.48779296875,
0.133056640625,
0.154541015625,
0.133056640625
],
[
0.515625,
0.12054443359375,
0.87890625,
0.12054443359375,
0.87890625,
0.1312255859375,
0.515625,
0.1312255859375
],
[
0.15283203125,
0.1353759765625,
0.48779296875,
0.133056640625,
0.48779296875,
0.1478271484375,
0.15283203125,
0.149658203125
],
[
0.51318359375,
0.1353759765625,
0.880859375,
0.1353759765625,
0.880859375,
0.1484375,
0.51318359375,
0.1484375
],
[
0.15283203125,
0.15087890625,
0.3720703125,
0.149658203125,
0.3720703125,
0.1644287109375,
0.15283203125,
0.1650390625
],
[
0.51318359375,
0.1519775390625,
0.87744140625,
0.1519775390625,
0.87744140625,
0.1650390625,
0.51318359375,
0.1650390625
],
[
0.51318359375,
0.16748046875,
0.87890625,
0.16748046875,
0.87890625,
0.18115234375,
0.51318359375,
0.18115234375
],
[
0.1376953125,
0.1776123046875,
0.295654296875,
0.1776123046875,
0.295654296875,
0.191162109375,
0.1376953125,
0.191162109375
],
[
0.28955078125,
0.1776123046875,
0.489501953125,
0.1776123046875,
0.489501953125,
0.191162109375,
0.28955078125,
0.191162109375
],
[
0.51123046875,
0.1817626953125,
0.87744140625,
0.1817626953125,
0.87744140625,
0.1953125,
0.51123046875,
0.1953125
],
[
0.154541015625,
0.193603515625,
0.489501953125,
0.193603515625,
0.489501953125,
0.207275390625,
0.154541015625,
0.207275390625
],
[
0.51318359375,
0.1995849609375,
0.880859375,
0.1995849609375,
0.880859375,
0.213134765625,
0.51318359375,
0.213134765625
],
[
0.154541015625,
0.210205078125,
0.48779296875,
0.210205078125,
0.48779296875,
0.223876953125,
0.154541015625,
0.223876953125
],
[
0.5146484375,
0.2161865234375,
0.62451171875,
0.2161865234375,
0.62451171875,
0.22802734375,
0.5146484375,
0.22802734375
],
[
0.151123046875,
0.2244873046875,
0.459228515625,
0.2257080078125,
0.459228515625,
0.240478515625,
0.151123046875,
0.2392578125
],
[
0.51123046875,
0.240478515625,
0.66162109375,
0.243408203125,
0.66162109375,
0.25830078125,
0.51123046875,
0.254638671875
],
[
0.1368408203125,
0.25048828125,
0.486083984375,
0.251708984375,
0.486083984375,
0.2666015625,
0.1368408203125,
0.265380859375
],
[
0.15283203125,
0.268310546875,
0.48779296875,
0.267333984375,
0.48779296875,
0.281982421875,
0.15283203125,
0.283203125
],
[
0.51123046875,
0.2666015625,
0.87890625,
0.2666015625,
0.87890625,
0.27978515625,
0.51123046875,
0.27978515625
],
[
0.154541015625,
0.28515625,
0.48779296875,
0.28515625,
0.48779296875,
0.298583984375,
0.154541015625,
0.298583984375
],
[
0.51123046875,
0.283203125,
0.87890625,
0.283203125,
0.87890625,
0.29638671875,
0.51123046875,
0.29638671875
],
[
0.154541015625,
0.300537109375,
0.490234375,
0.300537109375,
0.490234375,
0.314208984375,
0.154541015625,
0.314208984375
],
[
0.5146484375,
0.300537109375,
0.603515625,
0.300537109375,
0.603515625,
0.312255859375,
0.5146484375,
0.312255859375
],
[
0.154541015625,
0.318359375,
0.2091064453125,
0.318359375,
0.2091064453125,
0.330078125,
0.154541015625,
0.330078125
],
[
0.53076171875,
0.3154296875,
0.880859375,
0.3154296875,
0.880859375,
0.330078125,
0.53076171875,
0.330078125
],
[
0.509765625,
0.328857421875,
0.8818359375,
0.330810546875,
0.8818359375,
0.3486328125,
0.509765625,
0.3466796875
],
[
0.1376953125,
0.342529296875,
0.48779296875,
0.342529296875,
0.48779296875,
0.355712890625,
0.1376953125,
0.355712890625
],
[
0.51318359375,
0.3486328125,
0.880859375,
0.3486328125,
0.880859375,
0.3623046875,
0.51318359375,
0.3623046875
],
[
0.151123046875,
0.359375,
0.486083984375,
0.358154296875,
0.486083984375,
0.372314453125,
0.151123046875,
0.37353515625
],
[
0.51318359375,
0.365234375,
0.87890625,
0.365234375,
0.87890625,
0.37890625,
0.51318359375,
0.37890625
],
[
0.154541015625,
0.374755859375,
0.483642578125,
0.374755859375,
0.483642578125,
0.388427734375,
0.154541015625,
0.388427734375
],
[
0.51123046875,
0.37890625,
0.880859375,
0.38134765625,
0.880859375,
0.394287109375,
0.51123046875,
0.392578125
],
[
0.1351318359375,
0.402099609375,
0.483642578125,
0.402099609375,
0.483642578125,
0.4150390625,
0.1351318359375,
0.4150390625
],
[
0.51123046875,
0.396728515625,
0.87744140625,
0.396728515625,
0.87744140625,
0.410400390625,
0.51123046875,
0.410400390625
],
[
0.5146484375,
0.412109375,
0.880859375,
0.412109375,
0.880859375,
0.42578125,
0.5146484375,
0.42578125
],
[
0.118408203125,
0.418701171875,
0.484375,
0.418701171875,
0.484375,
0.431640625,
0.118408203125,
0.431640625
],
[
0.51318359375,
0.4287109375,
0.84814453125,
0.4287109375,
0.84814453125,
0.4423828125,
0.51318359375,
0.4423828125
],
[
0.118408203125,
0.435302734375,
0.486083984375,
0.435302734375,
0.486083984375,
0.448974609375,
0.118408203125,
0.448974609375
],
[
0.53076171875,
0.44482421875,
0.880859375,
0.44482421875,
0.880859375,
0.457763671875,
0.53076171875,
0.457763671875
],
[
0.118408203125,
0.453125,
0.48193359375,
0.453125,
0.48193359375,
0.463134765625,
0.118408203125,
0.463134765625
],
[
0.5087890625,
0.4560546875,
0.88330078125,
0.4560546875,
0.88330078125,
0.476806640625,
0.5087890625,
0.476806640625
],
[
0.118408203125,
0.46728515625,
0.486083984375,
0.46728515625,
0.486083984375,
0.48095703125,
0.118408203125,
0.48095703125
],
[
0.51318359375,
0.476806640625,
0.880859375,
0.476806640625,
0.880859375,
0.490478515625,
0.51318359375,
0.490478515625
],
[
0.118408203125,
0.482666015625,
0.486083984375,
0.482666015625,
0.486083984375,
0.496337890625,
0.118408203125,
0.496337890625
],
[
0.118408203125,
0.498779296875,
0.484375,
0.498779296875,
0.484375,
0.5126953125,
0.118408203125,
0.5126953125
],
[
0.51318359375,
0.493408203125,
0.5341796875,
0.493408203125,
0.5341796875,
0.505859375,
0.51318359375,
0.505859375
],
[
0.118408203125,
0.515625,
0.48779296875,
0.515625,
0.48779296875,
0.529296875,
0.118408203125,
0.529296875
],
[
0.509765625,
0.51708984375,
0.748046875,
0.51953125,
0.748046875,
0.53369140625,
0.509765625,
0.5322265625
],
[
0.118408203125,
0.53076171875,
0.486083984375,
0.53076171875,
0.486083984375,
0.54443359375,
0.118408203125,
0.54443359375
],
[
0.509765625,
0.53759765625,
0.8818359375,
0.5390625,
0.8818359375,
0.55712890625,
0.509765625,
0.5546875
],
[
0.118408203125,
0.54736328125,
0.385498046875,
0.54736328125,
0.385498046875,
0.56103515625,
0.118408203125,
0.56103515625
],
[
0.5146484375,
0.5556640625,
0.8818359375,
0.5556640625,
0.8818359375,
0.5693359375,
0.5146484375,
0.5693359375
],
[
0.1376953125,
0.57373046875,
0.490234375,
0.57373046875,
0.490234375,
0.58740234375,
0.1376953125,
0.58740234375
],
[
0.486083984375,
0.57666015625,
0.490234375,
0.57666015625,
0.490234375,
0.58203125,
0.486083984375,
0.58203125
],
[
0.509765625,
0.57177734375,
0.8818359375,
0.57177734375,
0.8818359375,
0.5849609375,
0.509765625,
0.5849609375
],
[
0.15283203125,
0.5888671875,
0.490234375,
0.5888671875,
0.490234375,
0.6025390625,
0.15283203125,
0.6025390625
],
[
0.51123046875,
0.58837890625,
0.57666015625,
0.59033203125,
0.57666015625,
0.60400390625,
0.51123046875,
0.6015625
],
[
0.15283203125,
0.60498046875,
0.48779296875,
0.60498046875,
0.48779296875,
0.6181640625,
0.15283203125,
0.6181640625
],
[
0.154541015625,
0.62158203125,
0.489501953125,
0.62158203125,
0.489501953125,
0.6357421875,
0.154541015625,
0.6357421875
],
[
0.55322265625,
0.6162109375,
0.880859375,
0.615234375,
0.880859375,
0.6328125,
0.55322265625,
0.63427734375
],
[
0.154541015625,
0.63818359375,
0.280517578125,
0.63818359375,
0.280517578125,
0.65185546875,
0.154541015625,
0.65185546875
],
[
0.51318359375,
0.64794921875,
0.880859375,
0.64794921875,
0.880859375,
0.6611328125,
0.51318359375,
0.6611328125
],
[
0.1368408203125,
0.66455078125,
0.489501953125,
0.66455078125,
0.489501953125,
0.677734375,
0.1368408203125,
0.677734375
],
[
0.51123046875,
0.6630859375,
0.880859375,
0.662109375,
0.880859375,
0.6767578125,
0.51123046875,
0.677734375
],
[
0.118408203125,
0.6796875,
0.489501953125,
0.6796875,
0.489501953125,
0.693359375,
0.118408203125,
0.693359375
],
[
0.51318359375,
0.6796875,
0.8818359375,
0.6796875,
0.8818359375,
0.693359375,
0.51318359375,
0.693359375
],
[
0.1175537109375,
0.69482421875,
0.48779296875,
0.69482421875,
0.48779296875,
0.7080078125,
0.1175537109375,
0.7080078125
],
[
0.51123046875,
0.693359375,
0.87890625,
0.693359375,
0.87890625,
0.71142578125,
0.51123046875,
0.71142578125
],
[
0.118408203125,
0.7119140625,
0.486083984375,
0.7119140625,
0.486083984375,
0.7255859375,
0.118408203125,
0.7255859375
],
[
0.5751953125,
0.71533203125,
0.86572265625,
0.71875,
0.86572265625,
0.74365234375,
0.5751953125,
0.74072265625
],
[
0.1175537109375,
0.72802734375,
0.484375,
0.72802734375,
0.484375,
0.7412109375,
0.1175537109375,
0.7412109375
],
[
0.85546875,
0.728515625,
0.880859375,
0.728515625,
0.880859375,
0.7412109375,
0.85546875,
0.7412109375
],
[
0.118408203125,
0.74462890625,
0.484375,
0.74462890625,
0.484375,
0.75830078125,
0.118408203125,
0.75830078125
],
[
0.64404296875,
0.74365234375,
0.654296875,
0.73828125,
0.66015625,
0.74365234375,
0.64990234375,
0.7490234375
],
[
0.118408203125,
0.76025390625,
0.486083984375,
0.76025390625,
0.486083984375,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.53076171875,
0.76025390625,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.53076171875,
0.77392578125
],
[
0.118408203125,
0.77685546875,
0.48779296875,
0.77685546875,
0.48779296875,
0.79052734375,
0.118408203125,
0.79052734375
],
[
0.51318359375,
0.77685546875,
0.880859375,
0.77685546875,
0.880859375,
0.79052734375,
0.51318359375,
0.79052734375
],
[
0.1175537109375,
0.79150390625,
0.48779296875,
0.79345703125,
0.48779296875,
0.80712890625,
0.1175537109375,
0.8046875
],
[
0.51318359375,
0.79345703125,
0.56103515625,
0.79345703125,
0.56103515625,
0.8046875,
0.51318359375,
0.8046875
],
[
0.1175537109375,
0.80810546875,
0.484375,
0.80859375,
0.484375,
0.822265625,
0.1175537109375,
0.82177734375
],
[
0.509765625,
0.81640625,
0.7021484375,
0.818359375,
0.7021484375,
0.8330078125,
0.509765625,
0.83056640625
],
[
0.118408203125,
0.82470703125,
0.484375,
0.82470703125,
0.484375,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.509765625,
0.83447265625,
0.88330078125,
0.83447265625,
0.88330078125,
0.85498046875,
0.509765625,
0.85498046875
],
[
0.118408203125,
0.84130859375,
0.48779296875,
0.84130859375,
0.48779296875,
0.85498046875,
0.118408203125,
0.85498046875
],
[
0.118408203125,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.509765625,
0.85400390625,
0.880859375,
0.85400390625,
0.880859375,
0.87158203125,
0.509765625,
0.87158203125
],
[
0.1175537109375,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.1175537109375,
0.8857421875
],
[
0.51318359375,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.118408203125,
0.8896484375,
0.486083984375,
0.8896484375,
0.486083984375,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51123046875,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.12005615234375,
0.90673828125,
0.484375,
0.90673828125,
0.484375,
0.91748046875,
0.12005615234375,
0.91748046875
],
[
0.509765625,
0.90625,
0.861328125,
0.90625,
0.861328125,
0.91943359375,
0.509765625,
0.91943359375
]
] |
[
"(Wu et al., 2020) further",
"此外，为了利用目标域中的未标记数据，我",
"",
"·TOD-BERT",
"pre-trains BERT on a task-oriented dialogue",
"们设计了一种联合预训练方案，该方案同时优",
"corpus of 100, 000 unlabeled samples with",
"化了源标记数据上的分类错误和自标未标记数",
"masked language modelling (MLM) and re-",
"据上的语言建模损失。这种联合训练方案可以",
"sponse contrastive objectives.",
"学习更好的语义表示，并且明显优于现有的两",
"阶段预训练方法（Gururanganetal.，2020）。图",
": USE-ConveRT",
"(Henderson et al., 2020;",
"中提供了由强基线和我们的方法产生的嵌入空",
"Casanueva et al., 2020) investigates a dual en-",
"间的可视化。1，它清楚地展示了我们预训练",
"coder model trained with response selection",
"模型的优越性",
"tasks on 727 million input-response pairs.",
"2Methodology",
": DNNC (Zhang et al., 2020a) pre-trains a lan-",
"guage model with around 1 million annotated",
"我们基于预训练的语言模型BERTE(Devlin",
"samples for natural language inference (NLI)",
"etal.,2019），提出了一个用于意图分类的持续",
"and use the pre-trained model for intent de-",
"预训练框架",
"tection.",
"我们的预训练方法依赖于一个小标签数据集",
"Dlabeled =【(ci,yi)】的存在，其中yi是话语 ci",
": WikiHowRoBERTa(Zhang et al., 2020b)",
"的标签。这样的数据样本可以很容易地从公共",
"constructs some pre-training tasks based on",
"意图检测数据集中获得，如OOs（Larsonetal.，",
"the wikiHow database with 110, 000 articles.",
"2019）和HWU64（Liuetal.2021）。如实验所",
"虽然这些方法取得了令人印象深刻的性能",
"示，来自OOS或HWU64的大致1,000示例足",
"以使预训练的意图检测模型在完全不同的目标",
"但它们严重依赖于大规模语料库（Mehrietal.",
"域（例如“Covid-19\"）上实现卓越的性能",
"2020）的存在，该语料库在语义上与目标域接",
"我们进一步考虑了目标域中未标记的话语",
"近，或者由类似的任务组成以进行持续的预",
"Dunlabeled－【ai)可用的场景，并提出了一种联",
"练，这需要大量的数据收集工作，并且计算成",
"合预训练方案，该方案经经验证明是高度有效",
"本很高。更重要的是，他们完全忽略了“免费",
"午餐”公开可用的，高质量的，手动注释的意",
"的",
"图检测基准。例如，数据集Oos（Larsonetal.，",
"2.1Supervised Pre-training",
"2019)提供跨10不同域的标记话语。因此，本",
"给定Dlabeled=[(ci，yi)与N不同的类，我们",
"文的研究围绕以下研究问题展开：",
"采用一种简单的方法来微调BERT。具体来说，",
": Is it possible to utilize publicly avail-",
".",
"在BERT的顶部附加一个线性层作为分类器，",
"able datasets to pre-train an intent detec-",
"即。e。,",
"tion model that can learn transferable task-",
"specific knowledge to generalize across dif-",
"p(ylhi) = softmax(Whi+b) e RN, (1)",
"ferent domains ?",
"其中hiERd是由[CLSI令牌给出的ai的特",
"在本文中，我们对此问题提供了肯定的答案。",
"征表示，WERNxd和bERN是线性层的参",
"我们使用简单的标准监督训练对BERT进行微",
"数。模型参数θ={Φ,W,b)，其中Φ是BERT",
"调，其中包含来自公共数据集的大约1.000个",
"的参数，在Dlobeled上进行了交叉摘损失训练：",
"标记的话语，并获得一个预训练模型，称为",
"* = arg min Ce (Dlabeld; ).(",
"IntentBERT。它可以直接应用于目标域上的少",
"(2)",
"数意图分类，该目标域与预训练数据大不相同，",
"8",
"并且明显优于现有的预训练模型，而无需进一",
"经过训练后，经过微调的BERT有望学习一",
"步微调目标数据（标记或未标记）。这个简单的",
"般的意图检测技能，因此我们将其称为Intent",
"“免费午餐”解决方案不仅证实了少拍意图检",
"BERT",
"测的可行性和实用性，还提供了一个随时可用",
"2.2Joint Pre-training",
"的性能良好的模型供实际使用，节省了算法设",
"给定未标记的目标数据Dunlabeled，我们可以利",
"计和数据收集方面的精力。此外，IntentBERT",
"对跨域少射分类任务的高泛化能力，由于较大",
"用它通过同时优化Dunlabeled上的语言建模损",
"的域间隙和较少的数据约束，通常被认为非常",
"失和方程E（2）中的监督损失来进一步增强",
"困难，这表明大多数意图检测任务可能共享一",
"IntentBERT。语言建模损失可以帮助学习目标",
"个可以从一小部分数据中学习的通用底层结构",
"域的语义表示，同时防止与源数据过度拟合"
] |
{
"arxiv_id": "2109.05782",
"language": "zh",
"page_idx": 2
}
|
||
2109.05782-zh-3
|
2109.05782-zh-3.jpg
|
具体来说，我们使用 MLM 作为语言建模损失，其中一部分输入令牌被特殊令牌$\left[MASK\right]$ 屏蔽，并且模型被训练以检索被屏蔽的令牌。联合训练损失的公式为：
$$\mathcal{L}_{\mathrm{joint}}=\mathcal{L}_{ce}(\mathcal{D}_{\mathrm{source}}^{\mathrm{labeled}};\theta)+\lambda\mathcal{L}_{mlm}(\mathcal{D}_{\mathrm{target}}^{\mathrm{unlabeled}};\theta),$$
(3)
,其中$\lambda$是平衡监督损失和非监督损失的超参
数
## 2.3 Few-shot Intent Classification
经过预训练后，IntentBERT 的参数是固定的可以立即用作新颖的少拍意图分类任务的特征提取器。分类器可以是参数的，例如逻辑回归，也可以是非参数的，例如最近邻。将使用任务中提供的几个带标签的示例来训练参数分类器，并对未标记的查询进行预测。如实验所示，由于 IntentBERT 产生的有效话语表示，简单的线性分类器足以实现非常好的性能
## 3 Experiments
3.1 Experimental Setup
<FigureHere>
Figure 2: Vocabulary overlap
数据集。为了训练我们的 IntentBERT,我们继续在两个数据集 OOS (Larson et al., 2019) 1和 HWU64 (Liu et al.,2021)上对 BERT 进行预训练，这两个数据集都包含多个域，提供了丰富的资源来学习$^2$。为了进行评估，我们使用了三个数据集：BANKING77$\boxed{\mathrm{E}}($Casanueva et al.,2020) 是专注于“银行”的细粒度意图检测数据集；MCID$\boxed{\mathrm{F}}($Arora et al.,2020a) 是“Covid- 19”聊天机器人的数据集；HINT3$\boxed{\mathrm{E}}($Arora et al., 2020b)包含 3 域，“床垫产品零售”,“健身用品零售”。数据集统计汇总在表 2 中
图。$\operatorname{E}$2 可视化了源训练数据和目标测试数据之间的词汇重叠，该重叠计算为删除停止词
$^{\mathrm{l}}$The domains“Banking”and“Credit Cards”are excluded
because they are semantically close to the evaluation data.
$^{2}$We have also experimented with the combination of both
datasets but observed no better results.
后共享词在任何两个数据集的组合词汇中所占的比例。据观察，重叠是相当小的，表明存在较大的语义缺口
评估。通过$C$ -way $K$ -shot 任务评估分类性能。对于每个任务，我们随机采样每个类的$C$ 类和$K$示例以训练分类器，然后我们采样每个类的额外 5 示例作为查询以进行评估。在500这样的任务上平均精度
基线。我们将 IntentBERT 与以下强基线进行比较。然后，BERT\_Freeze 将基于现成的 BERT 的 RoRoXkiX 分类模型进一步冻结； TOD-BERT$\overline{\mathrm{E}}($Wu et al., 2020) 进一步在具有MLM 和响应选择任务的大量面向任务的对话中对 BERT 进行预训练；CONVBERT$\boxed{\mathrm{E}}$(Mehri et al., 2020) 进一步在大型开放域 Howi 多轮对话语料库上对 BERT 进行预训练；USE-Conve
所有基线(BERT-Freeze 除外)都采用第二个预训练阶段，但目标不同，语料库不同。在我们的实验中，所有的基线(DNNC 除外)都使用逻辑回归作为分类器。对于 DNNC, 我们严格遵循原始实现$^{3}$来预先训练 BERT 风格的成对编码器，以估计查询话语的最佳匹配训练示例
培训详情。我们使用 BERT base $^{4}\mathbb{E}(d=768$ 的基本配置) 作为编码器，Adam (Kingma and Ba,2015)作为优化器，并使用 PyTorch 库来实现。该模型使用 NvidiaGeForceRTX2080TiGPU 进行训练。对于有监督的预训练，我们使用验证来控制早期停止，以防止过度拟合。具体来说，当使用 OOS 进行预训练时，我们使用HWU64 进行验证，反之亦然。如果在 3 纪元中没有观察到准确性的提高，则停止训练。对于联合预训练，将$\lambda$ 设置为1。训练时期的数量固定为 10 , 因为它不容易过度拟合
## 3.2 Main Results
主要结果见表 1。首先，在大多数情况下， IntentBERT (使用 OOS 或 HWU64 进行预训练)始终比所有基线表现出明显的优势。以MCID 上的 5 -way 2 -shot 分类结果为例，IntentBERT(OOS)以 9.4% 的绝对余量优于最强基线 CONVBERT, 证明了我们预训练方法的高效性。IntentBERT 的跨域可移植性表明，尽管存在语义域差距，但大多数意图检测任务可能共享类似的底层结构，可以通过一小组标记的话语来学习。其次，IntentBERT ( OOS)似乎比 IntentBERT (HWU64) 更有效，这可能是由于训练语料库的语义多样性所致。尽管如此， 它们之间的性能差异很小，这表明我们的预训
$^{3}$https://github.com/salesforce/DNNC-few-shot-intent
$^{4}$https://github.com/huggingface/transformers
|
[
[
0.139404296875,
0.0885009765625,
0.483642578125,
0.0885009765625,
0.483642578125,
0.09857177734375,
0.139404296875,
0.09857177734375
],
[
0.51318359375,
0.0872802734375,
0.87890625,
0.0872802734375,
0.87890625,
0.1009521484375,
0.51318359375,
0.1009521484375
],
[
0.118408203125,
0.10272216796875,
0.483642578125,
0.10272216796875,
0.483642578125,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.51318359375,
0.10272216796875,
0.8818359375,
0.10272216796875,
0.8818359375,
0.11639404296875,
0.51318359375,
0.11639404296875
],
[
0.118408203125,
0.1181640625,
0.490234375,
0.1181640625,
0.490234375,
0.1318359375,
0.118408203125,
0.1318359375
],
[
0.51318359375,
0.1181640625,
0.64208984375,
0.1181640625,
0.64208984375,
0.1318359375,
0.51318359375,
0.1318359375
],
[
0.118408203125,
0.1353759765625,
0.30908203125,
0.1353759765625,
0.30908203125,
0.1484375,
0.118408203125,
0.1484375
],
[
0.5322265625,
0.1353759765625,
0.8818359375,
0.1353759765625,
0.8818359375,
0.1484375,
0.5322265625,
0.1484375
],
[
0.1175537109375,
0.1561279296875,
0.484375,
0.1561279296875,
0.484375,
0.18701171875,
0.1175537109375,
0.18701171875
],
[
0.51318359375,
0.1519775390625,
0.880859375,
0.1519775390625,
0.880859375,
0.1650390625,
0.51318359375,
0.1650390625
],
[
0.51318359375,
0.168701171875,
0.880859375,
0.168701171875,
0.880859375,
0.1817626953125,
0.51318359375,
0.1817626953125
],
[
0.46337890625,
0.18115234375,
0.486083984375,
0.18115234375,
0.486083984375,
0.193603515625,
0.46337890625,
0.193603515625
],
[
0.51318359375,
0.18408203125,
0.8818359375,
0.18408203125,
0.8818359375,
0.19775390625,
0.51318359375,
0.19775390625
],
[
0.1318359375,
0.196533203125,
0.48779296875,
0.196533203125,
0.48779296875,
0.210205078125,
0.1318359375,
0.210205078125
],
[
0.51318359375,
0.20068359375,
0.72900390625,
0.20068359375,
0.72900390625,
0.21435546875,
0.51318359375,
0.21435546875
],
[
0.1158447265625,
0.2120361328125,
0.139404296875,
0.2120361328125,
0.139404296875,
0.226806640625,
0.1158447265625,
0.226806640625
],
[
0.124267578125,
0.201904296875,
0.1376953125,
0.201904296875,
0.1376953125,
0.2078857421875,
0.124267578125,
0.2078857421875
],
[
0.52880859375,
0.2161865234375,
0.880859375,
0.2161865234375,
0.880859375,
0.2298583984375,
0.52880859375,
0.2298583984375
],
[
0.51318359375,
0.2327880859375,
0.880859375,
0.2327880859375,
0.880859375,
0.2464599609375,
0.51318359375,
0.2464599609375
],
[
0.1158447265625,
0.240478515625,
0.397216796875,
0.240478515625,
0.397216796875,
0.254150390625,
0.1158447265625,
0.254150390625
],
[
0.51318359375,
0.2493896484375,
0.8818359375,
0.2493896484375,
0.8818359375,
0.26318359375,
0.51318359375,
0.26318359375
],
[
0.1175537109375,
0.26123046875,
0.483642578125,
0.26123046875,
0.483642578125,
0.27490234375,
0.1175537109375,
0.27490234375
],
[
0.51123046875,
0.265380859375,
0.880859375,
0.265380859375,
0.880859375,
0.279052734375,
0.51123046875,
0.279052734375
],
[
0.12005615234375,
0.27783203125,
0.484375,
0.27783203125,
0.484375,
0.2880859375,
0.12005615234375,
0.2880859375
],
[
0.509765625,
0.27978515625,
0.880859375,
0.28076171875,
0.880859375,
0.29443359375,
0.509765625,
0.29345703125
],
[
0.118408203125,
0.29345703125,
0.484375,
0.29345703125,
0.484375,
0.30712890625,
0.118408203125,
0.30712890625
],
[
0.51318359375,
0.297607421875,
0.880859375,
0.297607421875,
0.880859375,
0.311279296875,
0.51318359375,
0.311279296875
],
[
0.118408203125,
0.31005859375,
0.484375,
0.31005859375,
0.484375,
0.32373046875,
0.118408203125,
0.32373046875
],
[
0.51318359375,
0.3134765625,
0.880859375,
0.3134765625,
0.880859375,
0.32666015625,
0.51318359375,
0.32666015625
],
[
0.1175537109375,
0.32470703125,
0.483642578125,
0.32470703125,
0.483642578125,
0.33837890625,
0.1175537109375,
0.33837890625
],
[
0.51318359375,
0.330078125,
0.87744140625,
0.330078125,
0.87744140625,
0.34326171875,
0.51318359375,
0.34326171875
],
[
0.12005615234375,
0.342529296875,
0.483642578125,
0.342529296875,
0.483642578125,
0.352783203125,
0.12005615234375,
0.352783203125
],
[
0.5322265625,
0.3466796875,
0.880859375,
0.3466796875,
0.880859375,
0.35986328125,
0.5322265625,
0.35986328125
],
[
0.12005615234375,
0.358154296875,
0.48779296875,
0.358154296875,
0.48779296875,
0.371826171875,
0.12005615234375,
0.371826171875
],
[
0.51123046875,
0.3623046875,
0.880859375,
0.3623046875,
0.880859375,
0.377197265625,
0.51123046875,
0.377197265625
],
[
0.12005615234375,
0.37353515625,
0.4306640625,
0.37353515625,
0.4306640625,
0.38720703125,
0.12005615234375,
0.38720703125
],
[
0.51123046875,
0.377685546875,
0.880859375,
0.377685546875,
0.880859375,
0.391357421875,
0.51123046875,
0.391357421875
],
[
0.51318359375,
0.393798828125,
0.880859375,
0.394287109375,
0.880859375,
0.4091796875,
0.51318359375,
0.407958984375
],
[
0.1158447265625,
0.402099609375,
0.14697265625,
0.402099609375,
0.14697265625,
0.416259765625,
0.1158447265625,
0.416259765625
],
[
0.1427001953125,
0.40625,
0.1485595703125,
0.40625,
0.1485595703125,
0.410888671875,
0.1427001953125,
0.410888671875
],
[
0.151123046875,
0.402099609375,
0.264404296875,
0.402099609375,
0.264404296875,
0.416259765625,
0.151123046875,
0.416259765625
],
[
0.51123046875,
0.4091796875,
0.880859375,
0.410888671875,
0.880859375,
0.424560546875,
0.51123046875,
0.4228515625
],
[
0.11248779296875,
0.424560546875,
0.322509765625,
0.427001953125,
0.322509765625,
0.44482421875,
0.11248779296875,
0.4423828125
],
[
0.5146484375,
0.427001953125,
0.880859375,
0.427001953125,
0.880859375,
0.440673828125,
0.5146484375,
0.440673828125
],
[
0.5146484375,
0.443603515625,
0.5498046875,
0.443603515625,
0.5498046875,
0.4560546875,
0.5146484375,
0.4560546875
],
[
0.52880859375,
0.460205078125,
0.8818359375,
0.460205078125,
0.8818359375,
0.473876953125,
0.52880859375,
0.473876953125
],
[
0.51318359375,
0.476806640625,
0.880859375,
0.476806640625,
0.880859375,
0.490478515625,
0.51318359375,
0.490478515625
],
[
0.51318359375,
0.4921875,
0.880859375,
0.4921875,
0.880859375,
0.505859375,
0.51318359375,
0.505859375
],
[
0.246826171875,
0.50048828125,
0.279541015625,
0.50048828125,
0.279541015625,
0.5126953125,
0.246826171875,
0.5126953125
],
[
0.31494140625,
0.50048828125,
0.34765625,
0.50048828125,
0.34765625,
0.5126953125,
0.31494140625,
0.5126953125
],
[
0.382080078125,
0.50048828125,
0.414794921875,
0.50048828125,
0.414794921875,
0.5126953125,
0.382080078125,
0.5126953125
],
[
0.5146484375,
0.50830078125,
0.880859375,
0.50830078125,
0.880859375,
0.521484375,
0.5146484375,
0.521484375
],
[
0.193115234375,
0.501953125,
0.225830078125,
0.501953125,
0.225830078125,
0.51416015625,
0.193115234375,
0.51416015625
],
[
0.51318359375,
0.52490234375,
0.87890625,
0.52490234375,
0.87890625,
0.5380859375,
0.51318359375,
0.5380859375
],
[
0.51123046875,
0.5390625,
0.87890625,
0.5390625,
0.87890625,
0.552734375,
0.51123046875,
0.552734375
],
[
0.2476806640625,
0.54638671875,
0.279541015625,
0.54638671875,
0.279541015625,
0.55859375,
0.2476806640625,
0.55859375
],
[
0.31494140625,
0.54638671875,
0.34765625,
0.54638671875,
0.34765625,
0.55859375,
0.31494140625,
0.55859375
],
[
0.382080078125,
0.54638671875,
0.414794921875,
0.54638671875,
0.414794921875,
0.55859375,
0.382080078125,
0.55859375
],
[
0.51318359375,
0.55712890625,
0.880859375,
0.55712890625,
0.880859375,
0.57080078125,
0.51318359375,
0.57080078125
],
[
0.172119140625,
0.548828125,
0.2275390625,
0.548828125,
0.2275390625,
0.56005859375,
0.172119140625,
0.56005859375
],
[
0.51123046875,
0.57177734375,
0.880859375,
0.572265625,
0.880859375,
0.5859375,
0.51123046875,
0.5849609375
],
[
0.2216796875,
0.57763671875,
0.349365234375,
0.57763671875,
0.349365234375,
0.59130859375,
0.2216796875,
0.59130859375
],
[
0.372802734375,
0.5791015625,
0.418212890625,
0.5791015625,
0.418212890625,
0.59033203125,
0.372802734375,
0.59033203125
],
[
0.5146484375,
0.5888671875,
0.8818359375,
0.5888671875,
0.8818359375,
0.6025390625,
0.5146484375,
0.6025390625
],
[
0.5146484375,
0.60498046875,
0.8818359375,
0.60498046875,
0.8818359375,
0.6181640625,
0.5146484375,
0.6181640625
],
[
0.20068359375,
0.6162109375,
0.39794921875,
0.6162109375,
0.39794921875,
0.6298828125,
0.20068359375,
0.6298828125
],
[
0.51318359375,
0.62158203125,
0.81689453125,
0.62158203125,
0.81689453125,
0.63525390625,
0.51318359375,
0.63525390625
],
[
0.1175537109375,
0.65087890625,
0.486083984375,
0.65087890625,
0.486083984375,
0.66455078125,
0.1175537109375,
0.66455078125
],
[
0.509765625,
0.65087890625,
0.6650390625,
0.65087890625,
0.6650390625,
0.66455078125,
0.509765625,
0.66455078125
],
[
0.1175537109375,
0.66748046875,
0.483642578125,
0.66748046875,
0.483642578125,
0.68115234375,
0.1175537109375,
0.68115234375
],
[
0.51318359375,
0.6728515625,
0.87744140625,
0.6728515625,
0.87744140625,
0.68603515625,
0.51318359375,
0.68603515625
],
[
0.118408203125,
0.68408203125,
0.486083984375,
0.68408203125,
0.486083984375,
0.69775390625,
0.118408203125,
0.69775390625
],
[
0.51123046875,
0.689453125,
0.880859375,
0.689453125,
0.880859375,
0.703125,
0.51123046875,
0.703125
],
[
0.1175537109375,
0.69873046875,
0.483642578125,
0.69873046875,
0.483642578125,
0.7119140625,
0.1175537109375,
0.7119140625
],
[
0.51123046875,
0.7060546875,
0.8818359375,
0.7060546875,
0.8818359375,
0.7197265625,
0.51123046875,
0.7197265625
],
[
0.12005615234375,
0.71533203125,
0.486083984375,
0.71533203125,
0.486083984375,
0.72998046875,
0.12005615234375,
0.72998046875
],
[
0.51123046875,
0.7216796875,
0.8818359375,
0.7216796875,
0.8818359375,
0.7353515625,
0.51123046875,
0.7353515625
],
[
0.118408203125,
0.732421875,
0.486083984375,
0.732421875,
0.486083984375,
0.7451171875,
0.118408203125,
0.7451171875
],
[
0.51318359375,
0.73681640625,
0.880859375,
0.73681640625,
0.880859375,
0.75048828125,
0.51318359375,
0.75048828125
],
[
0.1175537109375,
0.7490234375,
0.484375,
0.74755859375,
0.484375,
0.76123046875,
0.1175537109375,
0.7626953125
],
[
0.51123046875,
0.751953125,
0.87744140625,
0.751953125,
0.87744140625,
0.765625,
0.51123046875,
0.765625
],
[
0.118408203125,
0.76416015625,
0.490234375,
0.76416015625,
0.490234375,
0.77783203125,
0.118408203125,
0.77783203125
],
[
0.51123046875,
0.7685546875,
0.87890625,
0.7685546875,
0.87890625,
0.7822265625,
0.51123046875,
0.7822265625
],
[
0.1158447265625,
0.77978515625,
0.486083984375,
0.77978515625,
0.486083984375,
0.79345703125,
0.1158447265625,
0.79345703125
],
[
0.5146484375,
0.78759765625,
0.87890625,
0.78759765625,
0.87890625,
0.79736328125,
0.5146484375,
0.79736328125
],
[
0.1158447265625,
0.79638671875,
0.486083984375,
0.79638671875,
0.486083984375,
0.81005859375,
0.1158447265625,
0.81005859375
],
[
0.51318359375,
0.8017578125,
0.880859375,
0.8017578125,
0.880859375,
0.8154296875,
0.51318359375,
0.8154296875
],
[
0.118408203125,
0.8125,
0.408935546875,
0.8125,
0.408935546875,
0.826171875,
0.118408203125,
0.826171875
],
[
0.51318359375,
0.81787109375,
0.880859375,
0.81787109375,
0.880859375,
0.83056640625,
0.51318359375,
0.83056640625
],
[
0.1376953125,
0.8291015625,
0.484375,
0.8291015625,
0.484375,
0.8427734375,
0.1376953125,
0.8427734375
],
[
0.51318359375,
0.83447265625,
0.880859375,
0.83447265625,
0.880859375,
0.84716796875,
0.51318359375,
0.84716796875
],
[
0.118408203125,
0.845703125,
0.486083984375,
0.845703125,
0.486083984375,
0.859375,
0.118408203125,
0.859375
],
[
0.51318359375,
0.85107421875,
0.87744140625,
0.85107421875,
0.87744140625,
0.8642578125,
0.51318359375,
0.8642578125
],
[
0.139404296875,
0.8681640625,
0.48779296875,
0.8681640625,
0.48779296875,
0.8818359375,
0.139404296875,
0.8818359375
],
[
0.5146484375,
0.8681640625,
0.87744140625,
0.8681640625,
0.87744140625,
0.87890625,
0.5146484375,
0.87890625
],
[
0.1158447265625,
0.88134765625,
0.466796875,
0.88134765625,
0.466796875,
0.89453125,
0.1158447265625,
0.89453125
],
[
0.1376953125,
0.892578125,
0.48779296875,
0.8935546875,
0.48779296875,
0.90673828125,
0.1376953125,
0.90625
],
[
0.5341796875,
0.892578125,
0.85791015625,
0.892578125,
0.85791015625,
0.90625,
0.5341796875,
0.90625
],
[
0.118408203125,
0.90673828125,
0.349365234375,
0.90673828125,
0.349365234375,
0.91748046875,
0.118408203125,
0.91748046875
],
[
0.53662109375,
0.90625,
0.8095703125,
0.90625,
0.8095703125,
0.91943359375,
0.53662109375,
0.91943359375
]
] |
[
"具体来说，我们使用MLM作为语言建模损",
"后共享词在任何两个数据集的组合词汇中所占",
"失，其中一部分输入令牌被特殊令牌[MASK]",
"的比例。据观察，重叠是相当小的，表明存在",
"屏蔽，并且模型被训练以检索被屏蔽的令牌。",
"较大的语义缺口",
"联合训练损失的公式为：",
"评估。通过C-wayK-shot任务评估分类性",
"Con e(tle; ) + XCmm (Dmltla,),",
"能。对于每个任务，我们随机采样每个类的C",
"类和K示例以训练分类器，然后我们采样每",
"(3)",
"个类的额外5示例作为查询以进行评估。在",
"其中入是平衡监督损失和非监督损失的超参",
"500这样的任务上平均精度",
"数",
"",
"基线。我们将IntentBERT与以下强基线",
"进行比较。然后，BERT_Freeze将基于现成",
"2.3Few-shot Intent Classification",
"的BERT的RoRoXkiX分类模型进一步冻结；",
"经过预训练后，IntentBERT的参数是固定的",
"TOD-BERTE(Wuetal.，2020）进一步在具有",
"可以立即用作新颖的少拍意图分类任务的特",
"MLM和响应选择任务的大量面向任务的对话",
"征提取器。分类器可以是参数的，例如逻辑回",
"中对BERT进行预训练；CONVBERTE(Mehri",
"归，也可以是非参数的，例如最近邻。将使用",
"etal.,2020)进一步在大型开放域Howi多轮对",
"任务中提供的几个带标签的示例来训练参数分",
"话语料库上对BERT进行预训练；USE-Conve",
"类器：并对未标记的香询进行预测。如实验所",
"所有基线（BERT-Freeze除外）都采用第二",
"示，由于IntentBERT产生的有效话语表示，简",
"个预训练阶段，但目标不同，语料库不同。在",
"单的线性分类器足以实现非常好的性能",
"我们的实验中，所有的基线（DNNC除外）都",
"使用逻辑回归作为分类器。对于DNNC，我们",
"3",
"",
"Experiments",
"严格遵循原始实现3来预先训练BERT风格的",
"3.1Experimental Setup",
"成对编码器，以估计查询话语的最佳匹配训练",
"示例",
"培训详情。我们使用BERTbase4E（d=768",
"的基本配置）作为编码器，Adam（Kingmaand",
"Ba,2015）作为优化器，并使用PyTorch库来实",
"0.19",
"0.12",
"0.10",
"现。该模型使用NvidiaGeForceRTX2080TiGPU",
"Oos",
"进行训练。对于有监督的预训练，我们使用",
"验证来控制早期停止，以防止过度拟合。具体",
"0.15",
"0.10",
"0.09",
"来说，当使用OOS进行预训练时，我们使用",
"HWU64",
"HWU64进行验证，反之亦然。如果在3纪元",
"BANKING77 MCID",
"HINT3",
"中没有观察到准确性的提高，则停止训练。对",
"于联合预训练，将入设置为1。训练时期的数",
"Figure 2: Vocabulary overlap",
"量固定为10，因为它不容易过度拟合",
"数据集。为了训练我们的IntentBERT，我们",
"3.2Main Results",
"继续在两个数据集O0S（Larsonetal.2019）1",
"主要结果见表1。首先，在大多数情况下",
"和HWU64（Liuetal.,2021）上对BERT进行",
"IntentBERT（使用OOS或HWU64进行预训",
"预训练，这两个数据集都包含多个域，提供了",
"练）始终比所有基线表现出明显的优势。以",
"丰富的资源来学习2。为了进行评估，我们",
"MCID上的5-way2-shot分类结果为例，In-",
"使用了三个数据集：BANKING77ECasanueva",
"tentBERT(OOS）以9.4%的绝对余量优于最强",
"etal.,2020)是专注于“银行”的细粒度意图检测",
"基线CONVBERT，证明了我们预训练方法的",
"数据集；MCIDE(Aroraetal.,2020a）是“Covid-",
"高效性。IntentBERT的跨域可移植性表明，尽",
"19”聊天机器人的数据集；HINT3E(Aroraetal.,",
"管存在语义域差距，但大多数意图检测任务可",
"2020b）包含3域，“床垫产品零售”，“健身用",
"能共享类似的底层结构，可以通过一小组标记",
"品零售”。数据集统计汇总在表2中",
"的话语来学习。其次，IntentBERT（OOS）似乎",
"图。E2可视化了源训练数据和目标测试数",
"比IntentBERT（HWU64）更有效，这可能是由",
"据之间的词汇重叠，该重叠计算为删除停止词",
"于训练语料库的语义多样性所致。尽管如此",
"'The domains “Banking\" and “Credit Cards\" are excluded",
"它们之间的性能差异很小，这表明我们的预训",
"because they are semantically close to the evaluation data",
"2we have also experimented with the combination of both",
"3https://github.com/salesforce/DNNC-few-shot-intent",
"datasets but observed no better results.",
"*https:/github.com/huggingface/transformers"
] |
{
"arxiv_id": "2109.05782",
"language": "zh",
"page_idx": 3
}
|
||
2109.05782-zh-4
|
2109.05782-zh-4.jpg
|
<table>
<tbody>
<tr>
<th rowspan="2">Method</th>
<th rowspan="2">unlabeled $\mathcal{D}.$ target</th>
<th colspan="2">BANKING77</th>
<th colspan="2">MCID</th>
<th colspan="2">$HINT3$</th>
</tr>
<tr>
<th>$2-shot$</th>
<th>10-shot</th>
<th>$2-shot$</th>
<th>10-shot</th>
<th>$2-shot$</th>
<th>10-shot</th>
</tr>
<tr>
<td>BERT- Freeze</td>
<td>$X$</td>
<td>52.6 $\pm12.4$</td>
<td>70.0 $\pm11.7$</td>
<td>57.8 $\pm11.7$</td>
<td>72.4 $\pm10.7$</td>
<td>47.3 $\pm12.1$</td>
<td>66.8 $\pm10.5$</td>
</tr>
<tr>
<td>CONVBERT</td>
<td>$X$</td>
<td>68.3 $\pm12.3$</td>
<td>86.6 $\pm8.2$</td>
<td>67.7 $\pm11.5$</td>
<td>83.5 $\pm7.9$</td>
<td>72.6 $\pm10.9$</td>
<td>87.2 $\pm7.9$</td>
</tr>
<tr>
<td>TOD- BERT</td>
<td>$X$</td>
<td>77.7 $\pm7.4$</td>
<td>89.4 $\pm5.1$</td>
<td>64.1 $\pm9.0$</td>
<td>77.7 $\pm11.0$</td>
<td>68.9 $\pm11.7$</td>
<td>83.5 $\pm8.6$</td>
</tr>
<tr>
<td>$\mathbf{USE-ConveRT}$</td>
<td>$X$</td>
<td> </td>
<td>85.2</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>DNNC</td>
<td>$X$</td>
<td>67.5 $\pm15.4$</td>
<td>89.8 $\pm7.5$</td>
<td>56.2 $\pm16.7$</td>
<td>80.0 $\pm9.9$</td>
<td>64.1 $\pm14.8$</td>
<td>87.9 $\pm8.1$</td>
</tr>
<tr>
<td>WikiHowRoBERTa</td>
<td>$X$</td>
<td>34.9 $\pm10.5$</td>
<td>41.6 $\pm10.1$</td>
<td>30.8 $\pm9.9$</td>
<td>36.4 $\pm9.7$</td>
<td>31.7 $\pm10.3$</td>
<td>39.0 $\pm9.9$</td>
</tr>
<tr>
<td>IntentBERT (HWU64) ( ours) </td>
<td>$X$</td>
<td>78.4 $\pm10.6$</td>
<td>90.0 $\pm7.5$</td>
<td>74.5 $\pm11.9$</td>
<td>85.9 $\pm8.8$</td>
<td>77.9 $\pm10.6$</td>
<td>89.4 $\pm7.9$</td>
</tr>
<tr>
<td>IntentBERT ( OOS) ( ours) </td>
<td>$X$</td>
<td>82.4 $\pm8.3$</td>
<td>91.8 $\pm4.2$</td>
<td>77.1 $\pm9.0$</td>
<td>88.1 $\pm5.9$</td>
<td>80.1 $\pm10.4$</td>
<td>90.2 $\pm7.4$</td>
</tr>
<tr>
<td>IntentBERT ( OOS) + MLM ( ours) </td>
<td>$√$</td>
<td>88.9 $\pm9.0$</td>
<td>95.2 $\pm5.1$</td>
<td>86.3 $\pm9.8$</td>
<td>92.4 $\pm6.2$</td>
<td>87.1 $\pm9.8$</td>
<td>94.0 $\pm6.0$</td>
</tr>
</tbody>
</table>
Table 1: Main results for 5 -way tasks. ¶ stands for results from the original paper.
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>$\#$domain</th>
<th>$\#$ intent</th>
<th>井 t utterances 、 </th>
</tr>
<tr>
<td>$OOS$</td>
<td>8</td>
<td>120</td>
<td>18000</td>
</tr>
<tr>
<td>HWU64</td>
<td>21</td>
<td>64</td>
<td>25716</td>
</tr>
<tr>
<td>$\mathbf{BANKING77}$</td>
<td>1</td>
<td>777</td>
<td>13083</td>
</tr>
<tr>
<td>MCID</td>
<td>1</td>
<td>16</td>
<td>1745</td>
</tr>
<tr>
<td>HINT3</td>
<td>3</td>
<td>51</td>
<td>2011</td>
</tr>
</tbody>
</table>
Table 2: Dataset statistics.
练方法对训练语料库不敏感。
<FigureHere>
Figure 3: Effect of the amount of labeled data used for pre-training in the source domain (OOS). The results are evaluated on 5 -way 2 -shot tasks on BANKING77.
最后，我们提出的联合预训练方案(第☑2.2节)比 IntentBERT(最多 9。2%绝对余量),显示了目标未标记数据可访问时联合预训练的高效性。我们的联合预训练方案也可以应用于其他语言模型，如 GPT-2 (Radford et al., 2019) 和ELMo (Peters et al.,2018),这将作为未来的工作
# 3.3 Analysis
预训练的标记数据量。我们在两个维度上减少了用于预训练的数据：域的数量和每个类的样本数。我们对1，2,4 和 8 域进行了多次随机采样，并在图中报告了平均结果。3。发现可以在不损害性能的情况下大大减少训练数
据。在 4 域和每个类的 20 样本上训练的模型的性能与在 8 域和每个类的 150 样本上的模型相当。一般来说，我们只需要围绕1，000注释的话语来训练 IntentBERT, 这可以很容易地在公共数据集中获得。这一发现表明，使用与任务相关的小数据进行预训练可能是一种更有效和高效的微调范例
联合预训练的未标记数据量。我们随机抽样一部分未标记的话语，并重新运行联合训练。如图所示。$\operatorname{E}$4 ,当未标记的样本数量从 10 增加到1，000并在达到1，000后趋于饱和时，精度不断提高。令人惊讶的是，BANKING77 中的 1,000 话语可以产生与完整数据集 (13,083话语)相当的性能。通常，它不需要太多未标记的数据就能达到很高的精度。消融联合预
<FigureHere>
Figure 4: Effect of the amount of unlabeled data used for joint pre-training in the target domain. The results are evaluated on 5 -way 2 -shot tasks with OOS as the source dataset.
训练研究。首先，我们研究一个两阶段预训练方案 (Gururangan et al.,2020),其中我们使用BERT 或 IntentBERT 作为初始化，并在目标域中执行 MLM(表 3 中的前两行)。可以看出， 它们的表现要比我们的联合预训练方案差得多(底排)。其次，我们在联合预训练(第三行)中
|
[
[
0.345947265625,
0.0872802734375,
0.411376953125,
0.0872802734375,
0.411376953125,
0.11102294921875,
0.345947265625,
0.11102294921875
],
[
0.456787109375,
0.0885009765625,
0.5439453125,
0.0885009765625,
0.5439453125,
0.09979248046875,
0.456787109375,
0.09979248046875
],
[
0.640625,
0.0885009765625,
0.68408203125,
0.0885009765625,
0.68408203125,
0.09979248046875,
0.640625,
0.09979248046875
],
[
0.8017578125,
0.0885009765625,
0.8486328125,
0.0885009765625,
0.8486328125,
0.09979248046875,
0.8017578125,
0.09979248046875
],
[
0.4375,
0.10394287109375,
0.483642578125,
0.10394287109375,
0.483642578125,
0.1181640625,
0.4375,
0.1181640625
],
[
0.1259765625,
0.09320068359375,
0.177978515625,
0.09320068359375,
0.177978515625,
0.1051025390625,
0.1259765625,
0.1051025390625
],
[
0.599609375,
0.10394287109375,
0.6455078125,
0.10394287109375,
0.6455078125,
0.1181640625,
0.599609375,
0.1181640625
],
[
0.76171875,
0.10394287109375,
0.8076171875,
0.10394287109375,
0.8076171875,
0.1181640625,
0.76171875,
0.1181640625
],
[
0.84228515625,
0.10394287109375,
0.892578125,
0.10394287109375,
0.892578125,
0.1181640625,
0.84228515625,
0.1181640625
],
[
0.67919921875,
0.105712890625,
0.73046875,
0.105712890625,
0.73046875,
0.1175537109375,
0.67919921875,
0.1175537109375
],
[
0.51708984375,
0.105712890625,
0.56494140625,
0.105712890625,
0.56494140625,
0.1175537109375,
0.51708984375,
0.1175537109375
],
[
0.1259765625,
0.12353515625,
0.210693359375,
0.12353515625,
0.210693359375,
0.1353759765625,
0.1259765625,
0.1353759765625
],
[
0.426513671875,
0.1217041015625,
0.495361328125,
0.1217041015625,
0.495361328125,
0.135986328125,
0.426513671875,
0.135986328125
],
[
0.509765625,
0.12353515625,
0.5751953125,
0.12353515625,
0.5751953125,
0.1353759765625,
0.509765625,
0.1353759765625
],
[
0.58837890625,
0.12353515625,
0.65576171875,
0.12353515625,
0.65576171875,
0.1353759765625,
0.58837890625,
0.1353759765625
],
[
0.6708984375,
0.12353515625,
0.736328125,
0.12353515625,
0.736328125,
0.1353759765625,
0.6708984375,
0.1353759765625
],
[
0.75146484375,
0.12353515625,
0.81787109375,
0.12353515625,
0.81787109375,
0.1353759765625,
0.75146484375,
0.1353759765625
],
[
0.8330078125,
0.12353515625,
0.89990234375,
0.12353515625,
0.89990234375,
0.1353759765625,
0.8330078125,
0.1353759765625
],
[
0.1259765625,
0.135986328125,
0.210693359375,
0.135986328125,
0.210693359375,
0.1478271484375,
0.1259765625,
0.1478271484375
],
[
0.372802734375,
0.12469482421875,
0.38623046875,
0.12469482421875,
0.38623046875,
0.135986328125,
0.372802734375,
0.135986328125
],
[
0.3720703125,
0.135986328125,
0.38623046875,
0.135986328125,
0.38623046875,
0.1478271484375,
0.3720703125,
0.1478271484375
],
[
0.426513671875,
0.135986328125,
0.49365234375,
0.135986328125,
0.49365234375,
0.1478271484375,
0.426513671875,
0.1478271484375
],
[
0.509765625,
0.135986328125,
0.57275390625,
0.135986328125,
0.57275390625,
0.1478271484375,
0.509765625,
0.1478271484375
],
[
0.58837890625,
0.135986328125,
0.654296875,
0.135986328125,
0.654296875,
0.1478271484375,
0.58837890625,
0.1478271484375
],
[
0.67236328125,
0.135986328125,
0.73486328125,
0.135986328125,
0.73486328125,
0.1478271484375,
0.67236328125,
0.1478271484375
],
[
0.75146484375,
0.135986328125,
0.8193359375,
0.135986328125,
0.8193359375,
0.1478271484375,
0.75146484375,
0.1478271484375
],
[
0.83447265625,
0.135986328125,
0.8984375,
0.135986328125,
0.8984375,
0.1478271484375,
0.83447265625,
0.1478271484375
],
[
0.75,
0.144287109375,
0.81787109375,
0.146728515625,
0.81787109375,
0.160888671875,
0.75,
0.1591796875
],
[
0.83447265625,
0.144287109375,
0.896484375,
0.146728515625,
0.896484375,
0.1602783203125,
0.8330078125,
0.157958984375
],
[
0.6689453125,
0.1455078125,
0.73779296875,
0.1478271484375,
0.73779296875,
0.1602783203125,
0.6689453125,
0.157958984375
],
[
0.1275634765625,
0.1478271484375,
0.202392578125,
0.1478271484375,
0.202392578125,
0.1591796875,
0.1275634765625,
0.1591796875
],
[
0.372802734375,
0.1478271484375,
0.38623046875,
0.1478271484375,
0.38623046875,
0.1591796875,
0.372802734375,
0.1591796875
],
[
0.429931640625,
0.1478271484375,
0.490234375,
0.1478271484375,
0.490234375,
0.1591796875,
0.429931640625,
0.1591796875
],
[
0.5087890625,
0.1478271484375,
0.57275390625,
0.1478271484375,
0.57275390625,
0.1591796875,
0.5087890625,
0.1591796875
],
[
0.59033203125,
0.1455078125,
0.654296875,
0.1478271484375,
0.654296875,
0.1602783203125,
0.59033203125,
0.157958984375
],
[
0.1275634765625,
0.1602783203125,
0.2301025390625,
0.1602783203125,
0.2301025390625,
0.171630859375,
0.1275634765625,
0.171630859375
],
[
0.3720703125,
0.160888671875,
0.38623046875,
0.160888671875,
0.38623046875,
0.1728515625,
0.3720703125,
0.1728515625
],
[
0.52490234375,
0.1602783203125,
0.5576171875,
0.1602783203125,
0.5576171875,
0.1728515625,
0.52490234375,
0.1728515625
],
[
0.426513671875,
0.17041015625,
0.495361328125,
0.17041015625,
0.495361328125,
0.185302734375,
0.426513671875,
0.185302734375
],
[
0.83642578125,
0.169189453125,
0.8984375,
0.171630859375,
0.896484375,
0.1859130859375,
0.83447265625,
0.18408203125
],
[
0.36865234375,
0.1728515625,
0.385498046875,
0.169189453125,
0.3896484375,
0.182861328125,
0.372802734375,
0.1859130859375
],
[
0.1259765625,
0.1728515625,
0.172119140625,
0.1728515625,
0.172119140625,
0.18408203125,
0.1259765625,
0.18408203125
],
[
0.509765625,
0.1728515625,
0.57275390625,
0.1728515625,
0.57275390625,
0.18408203125,
0.509765625,
0.18408203125
],
[
0.59033203125,
0.1728515625,
0.654296875,
0.1728515625,
0.654296875,
0.18408203125,
0.59033203125,
0.18408203125
],
[
0.67333984375,
0.1728515625,
0.73486328125,
0.1728515625,
0.73486328125,
0.18408203125,
0.67333984375,
0.18408203125
],
[
0.75244140625,
0.1728515625,
0.81787109375,
0.1728515625,
0.81787109375,
0.18408203125,
0.75244140625,
0.18408203125
],
[
0.1259765625,
0.182861328125,
0.2493896484375,
0.182861328125,
0.2493896484375,
0.196533203125,
0.1259765625,
0.196533203125
],
[
0.426513671875,
0.1817626953125,
0.495361328125,
0.1817626953125,
0.495361328125,
0.196533203125,
0.426513671875,
0.196533203125
],
[
0.509765625,
0.18408203125,
0.57275390625,
0.18408203125,
0.57275390625,
0.1953125,
0.509765625,
0.1953125
],
[
0.591796875,
0.18408203125,
0.6533203125,
0.18408203125,
0.6533203125,
0.1953125,
0.591796875,
0.1953125
],
[
0.67333984375,
0.18408203125,
0.73291015625,
0.18408203125,
0.73291015625,
0.1953125,
0.67333984375,
0.1953125
],
[
0.75244140625,
0.18408203125,
0.81689453125,
0.18408203125,
0.81689453125,
0.1953125,
0.75244140625,
0.1953125
],
[
0.83642578125,
0.18408203125,
0.896484375,
0.18408203125,
0.896484375,
0.1953125,
0.83642578125,
0.1953125
],
[
0.509765625,
0.198974609375,
0.57275390625,
0.20068359375,
0.57080078125,
0.215576171875,
0.5087890625,
0.213134765625
],
[
0.3720703125,
0.185302734375,
0.38623046875,
0.185302734375,
0.38623046875,
0.1953125,
0.3720703125,
0.1953125
],
[
0.587890625,
0.198974609375,
0.6572265625,
0.20068359375,
0.6572265625,
0.215576171875,
0.587890625,
0.213134765625
],
[
0.6708984375,
0.198974609375,
0.736328125,
0.20068359375,
0.73486328125,
0.215576171875,
0.6689453125,
0.213134765625
],
[
0.75,
0.198974609375,
0.8193359375,
0.20068359375,
0.8193359375,
0.215576171875,
0.75,
0.213134765625
],
[
0.8330078125,
0.1995849609375,
0.8984375,
0.201904296875,
0.8984375,
0.215576171875,
0.8330078125,
0.213134765625
],
[
0.4248046875,
0.1995849609375,
0.495361328125,
0.201904296875,
0.495361328125,
0.215576171875,
0.4248046875,
0.213134765625
],
[
0.124267578125,
0.201904296875,
0.30908203125,
0.201904296875,
0.30908203125,
0.215576171875,
0.124267578125,
0.215576171875
],
[
0.370361328125,
0.20068359375,
0.387939453125,
0.20068359375,
0.387939453125,
0.215576171875,
0.370361328125,
0.215576171875
],
[
0.42822265625,
0.21142578125,
0.49365234375,
0.213134765625,
0.491943359375,
0.22802734375,
0.426513671875,
0.2257080078125
],
[
0.5087890625,
0.21142578125,
0.57421875,
0.213134765625,
0.57275390625,
0.22802734375,
0.50732421875,
0.2257080078125
],
[
0.8330078125,
0.21142578125,
0.8984375,
0.213134765625,
0.896484375,
0.22802734375,
0.8310546875,
0.2257080078125
],
[
0.3720703125,
0.21435546875,
0.38623046875,
0.21435546875,
0.38623046875,
0.2257080078125,
0.3720703125,
0.2257080078125
],
[
0.124267578125,
0.213134765625,
0.287109375,
0.213134765625,
0.287109375,
0.226806640625,
0.124267578125,
0.226806640625
],
[
0.591796875,
0.21435546875,
0.654296875,
0.21435546875,
0.654296875,
0.2257080078125,
0.591796875,
0.2257080078125
],
[
0.67333984375,
0.21435546875,
0.73486328125,
0.21435546875,
0.73486328125,
0.2257080078125,
0.67333984375,
0.2257080078125
],
[
0.75146484375,
0.21435546875,
0.81787109375,
0.21435546875,
0.81787109375,
0.2257080078125,
0.75146484375,
0.2257080078125
],
[
0.509765625,
0.223876953125,
0.57080078125,
0.2257080078125,
0.57080078125,
0.2381591796875,
0.509765625,
0.236328125
],
[
0.67236328125,
0.2244873046875,
0.73486328125,
0.226806640625,
0.73486328125,
0.2381591796875,
0.67236328125,
0.236328125
],
[
0.83447265625,
0.22265625,
0.89990234375,
0.2257080078125,
0.8984375,
0.240478515625,
0.8330078125,
0.2369384765625
],
[
0.59375,
0.2257080078125,
0.6513671875,
0.2257080078125,
0.6513671875,
0.2369384765625,
0.59375,
0.2369384765625
],
[
0.429931640625,
0.2257080078125,
0.491943359375,
0.2257080078125,
0.491943359375,
0.2369384765625,
0.429931640625,
0.2369384765625
],
[
0.75390625,
0.2257080078125,
0.8154296875,
0.2257080078125,
0.8154296875,
0.2369384765625,
0.75390625,
0.2369384765625
],
[
0.3720703125,
0.226806640625,
0.385498046875,
0.226806640625,
0.385498046875,
0.2369384765625,
0.3720703125,
0.2369384765625
],
[
0.1259765625,
0.226806640625,
0.330078125,
0.226806640625,
0.330078125,
0.2369384765625,
0.1259765625,
0.2369384765625
],
[
0.219970703125,
0.2529296875,
0.77685546875,
0.254150390625,
0.77685546875,
0.268310546875,
0.219970703125,
0.267333984375
],
[
0.139404296875,
0.29345703125,
0.189697265625,
0.295654296875,
0.18896484375,
0.30810546875,
0.1376953125,
0.305908203125
],
[
0.2435302734375,
0.295654296875,
0.306396484375,
0.295654296875,
0.306396484375,
0.30712890625,
0.2435302734375,
0.30712890625
],
[
0.31982421875,
0.29638671875,
0.3720703125,
0.29638671875,
0.3720703125,
0.30810546875,
0.31982421875,
0.30810546875
],
[
0.385498046875,
0.29638671875,
0.46337890625,
0.29638671875,
0.46337890625,
0.30712890625,
0.385498046875,
0.30712890625
],
[
0.51318359375,
0.29345703125,
0.880859375,
0.29345703125,
0.880859375,
0.30712890625,
0.51318359375,
0.30712890625
],
[
0.1376953125,
0.3134765625,
0.17626953125,
0.3134765625,
0.17626953125,
0.325927734375,
0.1376953125,
0.325927734375
],
[
0.333251953125,
0.3134765625,
0.361083984375,
0.3134765625,
0.361083984375,
0.32666015625,
0.333251953125,
0.32666015625
],
[
0.405517578125,
0.3134765625,
0.447509765625,
0.3134765625,
0.447509765625,
0.32666015625,
0.405517578125,
0.32666015625
],
[
0.51123046875,
0.308837890625,
0.87890625,
0.308837890625,
0.87890625,
0.322509765625,
0.51123046875,
0.322509765625
],
[
0.264404296875,
0.32373046875,
0.285400390625,
0.32373046875,
0.285400390625,
0.33837890625,
0.264404296875,
0.33837890625
],
[
0.267822265625,
0.314208984375,
0.2822265625,
0.314208984375,
0.2822265625,
0.325927734375,
0.267822265625,
0.325927734375
],
[
0.334228515625,
0.32470703125,
0.35693359375,
0.32470703125,
0.35693359375,
0.33837890625,
0.334228515625,
0.33837890625
],
[
0.403076171875,
0.32373046875,
0.447509765625,
0.32373046875,
0.447509765625,
0.33837890625,
0.403076171875,
0.33837890625
],
[
0.51318359375,
0.32666015625,
0.880859375,
0.32666015625,
0.880859375,
0.34033203125,
0.51318359375,
0.34033203125
],
[
0.1376953125,
0.341552734375,
0.2275390625,
0.34326171875,
0.2275390625,
0.35693359375,
0.1376953125,
0.355224609375
],
[
0.14111328125,
0.325927734375,
0.1956787109375,
0.325927734375,
0.1956787109375,
0.33740234375,
0.14111328125,
0.33740234375
],
[
0.3359375,
0.342529296875,
0.35693359375,
0.342529296875,
0.35693359375,
0.355712890625,
0.3359375,
0.355712890625
],
[
0.403076171875,
0.342529296875,
0.447509765625,
0.342529296875,
0.447509765625,
0.35693359375,
0.403076171875,
0.35693359375
],
[
0.51123046875,
0.341552734375,
0.87890625,
0.341552734375,
0.87890625,
0.355224609375,
0.51123046875,
0.355224609375
],
[
0.3359375,
0.355224609375,
0.358642578125,
0.355224609375,
0.358642578125,
0.3681640625,
0.3359375,
0.3681640625
],
[
0.268798828125,
0.34326171875,
0.2822265625,
0.34326171875,
0.2822265625,
0.355712890625,
0.268798828125,
0.355712890625
],
[
0.14111328125,
0.355712890625,
0.185546875,
0.355712890625,
0.185546875,
0.36767578125,
0.14111328125,
0.36767578125
],
[
0.268798828125,
0.355712890625,
0.2822265625,
0.355712890625,
0.2822265625,
0.36767578125,
0.268798828125,
0.36767578125
],
[
0.4072265625,
0.355224609375,
0.443359375,
0.355224609375,
0.443359375,
0.36767578125,
0.4072265625,
0.36767578125
],
[
0.5146484375,
0.359375,
0.880859375,
0.359375,
0.880859375,
0.372314453125,
0.5146484375,
0.372314453125
],
[
0.334228515625,
0.36767578125,
0.35693359375,
0.36767578125,
0.35693359375,
0.380126953125,
0.334228515625,
0.380126953125
],
[
0.408935546875,
0.36767578125,
0.441650390625,
0.36767578125,
0.441650390625,
0.380126953125,
0.408935546875,
0.380126953125
],
[
0.268798828125,
0.3681640625,
0.2822265625,
0.3681640625,
0.2822265625,
0.37890625,
0.268798828125,
0.37890625
],
[
0.1376953125,
0.3681640625,
0.187255859375,
0.3681640625,
0.187255859375,
0.380126953125,
0.1376953125,
0.380126953125
],
[
0.51318359375,
0.37353515625,
0.880859375,
0.37353515625,
0.880859375,
0.38720703125,
0.51318359375,
0.38720703125
],
[
0.21240234375,
0.394287109375,
0.38623046875,
0.3955078125,
0.38623046875,
0.4091796875,
0.21240234375,
0.407958984375
],
[
0.5146484375,
0.39013671875,
0.66162109375,
0.39013671875,
0.66162109375,
0.405029296875,
0.5146484375,
0.405029296875
],
[
0.52880859375,
0.407958984375,
0.8818359375,
0.407958984375,
0.8818359375,
0.421630859375,
0.52880859375,
0.421630859375
],
[
0.52880859375,
0.424560546875,
0.8818359375,
0.424560546875,
0.8818359375,
0.438232421875,
0.52880859375,
0.438232421875
],
[
0.118408203125,
0.435302734375,
0.344970703125,
0.435302734375,
0.344970703125,
0.448974609375,
0.118408203125,
0.448974609375
],
[
0.5146484375,
0.440673828125,
0.880859375,
0.440673828125,
0.880859375,
0.454833984375,
0.5146484375,
0.454833984375
],
[
0.51318359375,
0.457275390625,
0.880859375,
0.457275390625,
0.880859375,
0.47021484375,
0.51318359375,
0.47021484375
],
[
0.16796875,
0.463134765625,
0.197265625,
0.463134765625,
0.197265625,
0.4755859375,
0.16796875,
0.4755859375
],
[
0.5146484375,
0.47265625,
0.8818359375,
0.47265625,
0.8818359375,
0.486328125,
0.5146484375,
0.486328125
],
[
0.166259765625,
0.479736328125,
0.197265625,
0.479736328125,
0.197265625,
0.4921875,
0.166259765625,
0.4921875
],
[
0.4375,
0.4892578125,
0.441650390625,
0.4892578125,
0.441650390625,
0.49462890625,
0.4375,
0.49462890625
],
[
0.51123046875,
0.488037109375,
0.880859375,
0.488037109375,
0.880859375,
0.5029296875,
0.51123046875,
0.5029296875
],
[
0.145263671875,
0.50048828125,
0.20654296875,
0.50048828125,
0.20654296875,
0.548828125,
0.145263671875,
0.548828125
],
[
0.16796875,
0.496337890625,
0.1956787109375,
0.496337890625,
0.1956787109375,
0.5087890625,
0.16796875,
0.5087890625
],
[
0.51318359375,
0.5048828125,
0.880859375,
0.5048828125,
0.880859375,
0.5185546875,
0.51318359375,
0.5185546875
],
[
0.5146484375,
0.521484375,
0.8818359375,
0.521484375,
0.8818359375,
0.53515625,
0.5146484375,
0.53515625
],
[
0.313232421875,
0.52978515625,
0.340087890625,
0.52978515625,
0.340087890625,
0.53759765625,
0.313232421875,
0.53759765625
],
[
0.34765625,
0.52783203125,
0.438232421875,
0.52783203125,
0.438232421875,
0.5390625,
0.34765625,
0.5390625
],
[
0.15283203125,
0.5361328125,
0.202392578125,
0.54638671875,
0.19482421875,
0.56396484375,
0.145263671875,
0.55419921875
],
[
0.34765625,
0.54150390625,
0.438232421875,
0.54150390625,
0.438232421875,
0.552734375,
0.34765625,
0.552734375
],
[
0.34765625,
0.5556640625,
0.438232421875,
0.5556640625,
0.438232421875,
0.56787109375,
0.34765625,
0.56787109375
],
[
0.16796875,
0.56396484375,
0.197265625,
0.56396484375,
0.197265625,
0.576171875,
0.16796875,
0.576171875
],
[
0.34765625,
0.57080078125,
0.438232421875,
0.57080078125,
0.438232421875,
0.58056640625,
0.34765625,
0.58056640625
],
[
0.5400390625,
0.56787109375,
0.57080078125,
0.56787109375,
0.57080078125,
0.5791015625,
0.5400390625,
0.5791015625
],
[
0.16796875,
0.58056640625,
0.202392578125,
0.58056640625,
0.202392578125,
0.5927734375,
0.16796875,
0.5927734375
],
[
0.193115234375,
0.58837890625,
0.447509765625,
0.5927734375,
0.447509765625,
0.60986328125,
0.193115234375,
0.60546875
],
[
0.276123046875,
0.60986328125,
0.372802734375,
0.60986328125,
0.372802734375,
0.62158203125,
0.276123046875,
0.62158203125
],
[
0.5380859375,
0.60986328125,
0.56689453125,
0.60986328125,
0.56689453125,
0.6181640625,
0.5380859375,
0.6181640625
],
[
0.5263671875,
0.6162109375,
0.53662109375,
0.6162109375,
0.53662109375,
0.646484375,
0.5263671875,
0.646484375
],
[
0.1175537109375,
0.63720703125,
0.486083984375,
0.63720703125,
0.486083984375,
0.65087890625,
0.1175537109375,
0.65087890625
],
[
0.5263671875,
0.646484375,
0.57275390625,
0.6494140625,
0.57080078125,
0.662109375,
0.52490234375,
0.6591796875
],
[
0.75,
0.6435546875,
0.8388671875,
0.6435546875,
0.8388671875,
0.65380859375,
0.75,
0.65380859375
],
[
0.1158447265625,
0.65185546875,
0.48779296875,
0.65185546875,
0.48779296875,
0.6650390625,
0.1158447265625,
0.6650390625
],
[
0.75,
0.65673828125,
0.79443359375,
0.65673828125,
0.79443359375,
0.66845703125,
0.75,
0.66845703125
],
[
0.1158447265625,
0.6650390625,
0.486083984375,
0.6650390625,
0.486083984375,
0.6787109375,
0.1158447265625,
0.6787109375
],
[
0.75244140625,
0.67041015625,
0.80029296875,
0.67041015625,
0.80029296875,
0.681640625,
0.75244140625,
0.681640625
],
[
0.1368408203125,
0.69677734375,
0.486083984375,
0.69677734375,
0.486083984375,
0.71044921875,
0.1368408203125,
0.71044921875
],
[
0.5380859375,
0.69140625,
0.5693359375,
0.69140625,
0.5693359375,
0.703125,
0.5380859375,
0.703125
],
[
0.580078125,
0.70068359375,
0.591796875,
0.70068359375,
0.591796875,
0.71044921875,
0.580078125,
0.71044921875
],
[
0.640625,
0.703125,
0.6572265625,
0.703125,
0.6572265625,
0.71337890625,
0.640625,
0.71337890625
],
[
0.7021484375,
0.703125,
0.724609375,
0.703125,
0.724609375,
0.71630859375,
0.7021484375,
0.71630859375
],
[
0.118408203125,
0.71142578125,
0.483642578125,
0.71142578125,
0.483642578125,
0.724609375,
0.118408203125,
0.724609375
],
[
0.118408203125,
0.72802734375,
0.484375,
0.72802734375,
0.484375,
0.7412109375,
0.118408203125,
0.7412109375
],
[
0.66650390625,
0.724609375,
0.759765625,
0.724609375,
0.759765625,
0.7353515625,
0.66650390625,
0.7353515625
],
[
0.1175537109375,
0.74365234375,
0.484375,
0.74365234375,
0.484375,
0.75732421875,
0.1175537109375,
0.75732421875
],
[
0.51123046875,
0.74658203125,
0.8818359375,
0.74658203125,
0.8818359375,
0.76025390625,
0.51123046875,
0.76025390625
],
[
0.1175537109375,
0.76025390625,
0.486083984375,
0.7587890625,
0.486083984375,
0.77392578125,
0.1175537109375,
0.77490234375
],
[
0.51123046875,
0.76123046875,
0.880859375,
0.76025390625,
0.880859375,
0.77392578125,
0.51123046875,
0.77490234375
],
[
0.1175537109375,
0.77685546875,
0.486083984375,
0.77685546875,
0.486083984375,
0.79052734375,
0.1175537109375,
0.79052734375
],
[
0.51123046875,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51123046875,
0.7890625
],
[
0.1175537109375,
0.7919921875,
0.139404296875,
0.7919921875,
0.139404296875,
0.80712890625,
0.1175537109375,
0.80712890625
],
[
0.51318359375,
0.79150390625,
0.61279296875,
0.79150390625,
0.61279296875,
0.8017578125,
0.51318359375,
0.8017578125
],
[
0.1175537109375,
0.82080078125,
0.162841796875,
0.82080078125,
0.162841796875,
0.83203125,
0.1175537109375,
0.83203125
],
[
0.15869140625,
0.82080078125,
0.231689453125,
0.82080078125,
0.231689453125,
0.83203125,
0.15869140625,
0.83203125
],
[
0.51318359375,
0.82470703125,
0.87890625,
0.82470703125,
0.87890625,
0.83837890625,
0.51318359375,
0.83837890625
],
[
0.1158447265625,
0.84033203125,
0.484375,
0.84033203125,
0.484375,
0.85400390625,
0.1158447265625,
0.85400390625
],
[
0.51318359375,
0.84033203125,
0.87890625,
0.84033203125,
0.87890625,
0.85400390625,
0.51318359375,
0.85400390625
],
[
0.118408203125,
0.85546875,
0.486083984375,
0.85498046875,
0.486083984375,
0.869140625,
0.118408203125,
0.87060546875
],
[
0.509765625,
0.85498046875,
0.8818359375,
0.85546875,
0.8818359375,
0.87060546875,
0.509765625,
0.869140625
],
[
0.118408203125,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.51318359375,
0.87255859375,
0.8818359375,
0.87255859375,
0.8818359375,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.118408203125,
0.88818359375,
0.486083984375,
0.88818359375,
0.486083984375,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51123046875,
0.88818359375,
0.8818359375,
0.88818359375,
0.8818359375,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.1175537109375,
0.90478515625,
0.486083984375,
0.90478515625,
0.486083984375,
0.91845703125,
0.1175537109375,
0.91845703125
],
[
0.51123046875,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"Dawltriad",
"BANKING77",
"MCID",
"HINT3",
"2-shot",
"Method",
"2-shot",
"2-shot",
"10-shot",
"10-shot",
"10-shot",
"BERT-Freeze",
"52.6 ± 12.4",
"70.0±11.7",
"57.8 ± 11.7",
"72.4 ± 10.7",
"47.3 ±12.1",
"66.8 ± 10.5",
"CONVBERT",
"x",
"x",
"68.3 ± 12.3",
"86.6 ± 8.2",
"67.7 ± 11.5",
"83.5 ± 7.9",
"72.6 ± 10.9",
"87.2 ± 7.9",
"68.9 ± 11.7",
"83.5 ±8.6",
"77.7 ± 11.0",
"TOD-BERT",
"x",
"77.7 ± 7.4",
"89.4 ±5.1",
"64.1 ± 9.0",
"USE-ConveRT ",
"x",
"85.2",
"67.5 ± 15.4",
"87.9 ± 8.1",
"X",
"DNNC",
"89.8 ±7.5",
"56.2 ± 16.7",
"80.0±9.9",
"64.1 ± 14.8",
"WikiHowRoBERTa",
"34.9 ± 10.5",
"41.6 ± 10.1",
"30.8 ± 9.9",
"36.4 ± 9.7",
"31.7 ± 10.3",
"39.0±9.9",
"90.0 ± 7.5",
"X",
"74.5 ± 11.9",
"85.9 ± 8.8",
"77.9 ± 10.6",
"89.4 ± 7.9",
"78.4 ± 10.6",
"IntentBERT (HWU64) (ours)",
"x",
"82.4 ±8.3",
"91.8 ± 4.2",
"90.2±74",
"x",
"IntentBERT (OOS) (ours)",
"77.1 ± 9.0",
"88.1 ±5.9",
"80.1 ±10.4",
"95.2 ± 5.1",
"92.4 ± 6.2",
"94.0 ± 6.0",
"86.3 ± 9.8",
"88.9 ± 9.0",
"87.1 ± 9.8",
"V",
"IntentBERT(OOS)+MLM (ours)",
"Table 1: Main results for 5 -way tasks. stands for results from the original paper.",
"Dataset",
"# domain",
"# intent",
"#utterances",
"据。在4域和每个类的20样本上训练的模型",
"oos",
"120",
"18000",
"的性能与在8域和每个类的150样本上的模型",
"21",
"8",
"64",
"25716",
"相当。一般来说，我们只需要围绕1,000注释",
"BANKING77",
"HWU64",
"77",
"13083",
"的话语来训练IntentBERT，这可以很容易地在",
"16",
"1",
"MCID",
"1",
"1745",
"公共数据集中获得。这一发现表明，使用与任",
"51",
"2011",
"3",
"HINT3",
"务相关的小数据进行预训练可能是一种更有效",
"Table 2: Dataset statistics",
"和高效的微调范例",
"联合预训练的未标记数据量。我们随机抽样",
"部分未标记的话语，并重新运行联合训练。",
"练方法对训练语料库不敏感。",
"如图所示。E4，当未标记的样本数量从10增",
"加到1,000并在达到1,000后趋于饱和时，精",
"0.85",
"度不断提高。令人惊的是，BANKING77中",
"0.80",
"",
"的1,000话语可以产生与完整数据集（13,083",
"",
"0.75",
"话语）相当的性能。通常，它不需要太多未标",
"记的数据就能达到很高的精度。消融联合预",
"中",
"#domains=1",
"00.60",
"#domains=2",
"#domains=4",
"0.55",
"#domains=8",
"0.90",
"0.50",
"boo",
"data per label",
"0.85",
"",
"Figure 3: Effect of the amount of labeled data used for",
"00.80",
"BANKING77",
"pre-training in the source domain (OOS). The results",
"MCID",
"are evaluated on 5 -way 2 -shot tasks on BANKING77.",
"HINT3",
"最后，我们提出的联合预训练方案（第E2.2",
"0.75",
"O",
"10",
"100",
"节）比IntentBERT（最多9。2%绝对余量），显",
"示了目标未标记数据可访问时联合预训练的高",
"amount of data",
"效性。我们的联合预训练方案也可以应用于其",
"Figure 4: Effect of the amount of unlabeled data used",
"他语言模型，如GPT-2（Radfordetal.2019和",
"for joint pre-training in the target domain. The results",
"ELMo（Petersetal.,2018），这将作为未来的工",
"are evaluated on 5 -way 2 -shot tasks with OOS as the",
"作",
"source dataset.",
"3.3",
"Analysis",
"训练研究。首先，我们研究一个两阶段预训练",
"预训练的标记数据量。我们在两个维度上减少",
"方案（Gururanganetal.，2020），其中我们使用",
"了用于预训练的数据：域的数量和每个类的样",
"BERT或IntentBERT作为初始化，并在目标域",
"本数。我们对1，2，4和8域进行了多次随",
"中执行MLM（表3中的前两行）。可以看出，",
"机采样，并在图中报告了平均结果。3。发",
"它们的表现要比我们的联合预训练方案差得多",
"现可以在不损害性能的情况下大大减少训练数",
"（底排）。其次，我们在联合预训练（第三行）中"
] |
{
"arxiv_id": "2109.05782",
"language": "zh",
"page_idx": 4
}
|
||
2305.14878-zh-1
|
2305.14878-zh-1.jpg
|
Arul Menezes
Vikas Raunak
利用 GPT-4 进行自动翻译后编辑
Amr Sharaf Yiren Wang Hany Hassan Awadalla
Microsoft Azure AI Redmond, Washington
$\{$ viraunak,amrsharaf,wangyiren,hanyh,arulm $\}$ @microsoft.com
# Abstract
While Neural Machine Translation (NMT) represents the leading approach to Machine Translation (MT), the outputs of NMT models still require translation post-editing to rectify errors and enhance quality under critical settings. In this work, we formalize the task of direct translation post-editing with Large Language Models (LLMs) and explore the use of GPT-4 to automatically post-edit NMT outputs across several language pairs. Our results demonstrate that GPT-4 is adept at translation post-editing, producing meaningful and trustworthy edits to translations that help improve its general quality as well as remove different classes of major errors in translations. In particular, human evaluations on assessing edit trustworthiness show that GPT-4 exhibits a large improvement over the prior state-of-the-art LLM. Notably, we improve upon state-of-the-art performance on WMT-22 English-Chinese, English-German, ChineseEnglish and German-English language pairs using GPT-4 based post-editing, as evaluated by state-of-the-art MT quality metrics. However, we also show that GPT-4 could produce hallucinated edits, thereby urging caution in its use as an expert translation post-editor.
### 1 $\textbf{Introduction}$
在 Web 挖掘的并行语料库上训练的最先进的神经机器翻译(NMT)模型，即使对于较高资源的语言对，也会遇到可靠性问题，尽管(He et al., 2020; Gupta et al., 2020; Sun et al., 2020; Wang et al., 2021; He et al., 2021; Raunak et al., 2022; Raunak and Menezes, 2022) 的平均案例性能很高。许多先前的工作表明，在极少数情况下，NMT 中的并行数据和模型训练工件可能会表现为灾难性的输出，并且检测这种恶劣的模型行为仍然是一项具有挑战性的任务 (Rau-1) nak et al., 2021; Tang et al., 2022; Guerreiro et al., 2023b; Xu et al., 2023; Guerreiro et al., 2023a) 。因此，编辑后神经机器翻译仍然是在翻译和本地化行业的关键环境中使用它们的重要练习。
因此，要问的一个相关问题是，是否可以利用大型语言模型(LLM),例如GPT-3,GPT-4和 PaLM, PaLM2 (Brown et al., 2020; OpenAI, 2023; Anil et al., 2023; Chowdhery et al., 2022) , 它们已经展示了广泛的通用推理和基于知识的功能用于翻译后编辑任务。基于 LLM 的自动翻译后编辑可以帮助检测和修复翻译错误，以确保 NMT 输出的更高可靠性。除了缓解 NMT 中的可靠性问题之外，利用 LLM 进行后期编辑可能是适当的，因为最新的 LLM 具有先进的多语言理解能力，特别是 GPT-4(OpenAI, 2023),并且可能具有将基于知识或特定文化的自定义应用于 (Baker,2009) 翻译的能力，这很难通过源文本确定
在这项工作中，我们探讨了GPT-4 等最先进的 LLM 在直接设置的翻译后编辑任务中的功效，即。e。,在后期编辑之前，没有对翻译应用任何质量估计或错误检测步骤。我们的贡献如下：(i)我们使用 LLM 正式确定直接翻译后编辑的任务，并提出一系列研究问题和措施， 以量化其效用，以改善从 NMT 模型获得的翻译。(ii)我们证明了GPT-4 对具有高度可信度的最新翻译产生了有意义的，与人类判断一致的编辑。(iii)我们证明了在 WMT-22 基准测试 (Koehn et al.,2022) 上，与 WMT-22 的最佳系统相比，在许多语言对上的翻译质量有所提高，从而提高了基于 GPT-4 的基于 GPT 的后期编辑的英汉，英德，汉英和德语-英语的最新翻译性能，如最新的 MT 质量指标所评估的
### 2 $\textbf{The Translation Post- Editing Task}$
Task Definition: 我们在生成设置中将后期编辑任务形式化如下：给定源($S$)和翻译($T$ ),提出对翻译($E$ )的改进，并使用建议的编辑($T^\prime$)生成翻译，i。e。:
$$\begin{matrix}(S,T)\to E+T'\end{matrix}$$
在此任务设置下，$E$表示由 LLM 描述的改进或编辑。在没有$E$的情况下，任务被简化为简单地生成改进的翻译，而无需任何中间推
|
[
[
0.330078125,
0.09442138671875,
0.6708984375,
0.09442138671875,
0.6708984375,
0.1092529296875,
0.330078125,
0.1092529296875
],
[
0.0889892578125,
0.1341552734375,
0.214111328125,
0.1341552734375,
0.214111328125,
0.1484375,
0.0889892578125,
0.1484375
],
[
0.25439453125,
0.1341552734375,
0.36279296875,
0.1341552734375,
0.36279296875,
0.1484375,
0.25439453125,
0.1484375
],
[
0.399658203125,
0.1318359375,
0.51123046875,
0.1341552734375,
0.51123046875,
0.149658203125,
0.399658203125,
0.1478271484375
],
[
0.54833984375,
0.1353759765625,
0.75244140625,
0.1353759765625,
0.75244140625,
0.1484375,
0.54833984375,
0.1484375
],
[
0.79248046875,
0.1341552734375,
0.916015625,
0.1353759765625,
0.916015625,
0.149658203125,
0.79248046875,
0.1484375
],
[
0.5224609375,
0.14013671875,
0.54150390625,
0.14013671875,
0.54150390625,
0.144287109375,
0.5224609375,
0.144287109375
],
[
0.420654296875,
0.16748046875,
0.58251953125,
0.16748046875,
0.58251953125,
0.1817626953125,
0.420654296875,
0.1817626953125
],
[
0.4072265625,
0.182861328125,
0.59423828125,
0.18408203125,
0.59423828125,
0.198974609375,
0.4072265625,
0.19775390625
],
[
0.210693359375,
0.201904296875,
0.8037109375,
0.201904296875,
0.8037109375,
0.215576171875,
0.210693359375,
0.215576171875
],
[
0.2626953125,
0.254150390625,
0.340087890625,
0.254150390625,
0.340087890625,
0.268310546875,
0.2626953125,
0.268310546875
],
[
0.5146484375,
0.254638671875,
0.87890625,
0.254638671875,
0.87890625,
0.268310546875,
0.5146484375,
0.268310546875
],
[
0.5146484375,
0.270751953125,
0.8818359375,
0.270751953125,
0.8818359375,
0.283935546875,
0.5146484375,
0.283935546875
],
[
0.145263671875,
0.276611328125,
0.45751953125,
0.276611328125,
0.45751953125,
0.290283203125,
0.145263671875,
0.290283203125
],
[
0.14697265625,
0.292236328125,
0.45751953125,
0.292236328125,
0.45751953125,
0.305908203125,
0.14697265625,
0.305908203125
],
[
0.51318359375,
0.287353515625,
0.880859375,
0.287353515625,
0.880859375,
0.300537109375,
0.51318359375,
0.300537109375
],
[
0.145263671875,
0.305908203125,
0.459228515625,
0.305908203125,
0.459228515625,
0.319580078125,
0.145263671875,
0.319580078125
],
[
0.51123046875,
0.302734375,
0.880859375,
0.302734375,
0.880859375,
0.31640625,
0.51123046875,
0.31640625
],
[
0.1485595703125,
0.322509765625,
0.456787109375,
0.322509765625,
0.456787109375,
0.333251953125,
0.1485595703125,
0.333251953125
],
[
0.51318359375,
0.318359375,
0.87890625,
0.318359375,
0.87890625,
0.33203125,
0.51318359375,
0.33203125
],
[
0.145263671875,
0.334228515625,
0.45751953125,
0.333251953125,
0.45751953125,
0.347412109375,
0.145263671875,
0.3486328125
],
[
0.51318359375,
0.3349609375,
0.880859375,
0.3349609375,
0.880859375,
0.3486328125,
0.51318359375,
0.3486328125
],
[
0.1485595703125,
0.350830078125,
0.456787109375,
0.350830078125,
0.456787109375,
0.361083984375,
0.1485595703125,
0.361083984375
],
[
0.51318359375,
0.3515625,
0.880859375,
0.3515625,
0.880859375,
0.365234375,
0.51318359375,
0.365234375
],
[
0.1485595703125,
0.365234375,
0.456787109375,
0.365234375,
0.456787109375,
0.3759765625,
0.1485595703125,
0.3759765625
],
[
0.51318359375,
0.36767578125,
0.880859375,
0.36767578125,
0.880859375,
0.3818359375,
0.51318359375,
0.3818359375
],
[
0.145263671875,
0.377685546875,
0.456787109375,
0.377197265625,
0.456787109375,
0.39013671875,
0.145263671875,
0.391357421875
],
[
0.5146484375,
0.38427734375,
0.880859375,
0.38427734375,
0.880859375,
0.39794921875,
0.5146484375,
0.39794921875
],
[
0.145263671875,
0.391357421875,
0.45751953125,
0.391357421875,
0.45751953125,
0.405029296875,
0.145263671875,
0.405029296875
],
[
0.51123046875,
0.399658203125,
0.87890625,
0.399658203125,
0.87890625,
0.413330078125,
0.51123046875,
0.413330078125
],
[
0.14697265625,
0.40625,
0.45751953125,
0.40625,
0.45751953125,
0.419189453125,
0.14697265625,
0.419189453125
],
[
0.51123046875,
0.4150390625,
0.880859375,
0.41455078125,
0.880859375,
0.4287109375,
0.51123046875,
0.429931640625
],
[
0.1485595703125,
0.421630859375,
0.453369140625,
0.421630859375,
0.453369140625,
0.431640625,
0.1485595703125,
0.431640625
],
[
0.145263671875,
0.43408203125,
0.45751953125,
0.432861328125,
0.45751953125,
0.44775390625,
0.145263671875,
0.448974609375
],
[
0.51318359375,
0.431640625,
0.880859375,
0.431640625,
0.880859375,
0.4453125,
0.51318359375,
0.4453125
],
[
0.1485595703125,
0.449462890625,
0.456787109375,
0.449462890625,
0.456787109375,
0.460205078125,
0.1485595703125,
0.460205078125
],
[
0.51123046875,
0.44775390625,
0.87890625,
0.44775390625,
0.87890625,
0.46142578125,
0.51123046875,
0.46142578125
],
[
0.14697265625,
0.463134765625,
0.456787109375,
0.463134765625,
0.456787109375,
0.476806640625,
0.14697265625,
0.476806640625
],
[
0.51318359375,
0.46435546875,
0.67822265625,
0.46435546875,
0.67822265625,
0.47802734375,
0.51318359375,
0.47802734375
],
[
0.145263671875,
0.476806640625,
0.456787109375,
0.476806640625,
0.456787109375,
0.490478515625,
0.145263671875,
0.490478515625
],
[
0.52880859375,
0.479736328125,
0.880859375,
0.479736328125,
0.880859375,
0.49462890625,
0.52880859375,
0.49462890625
],
[
0.1485595703125,
0.4921875,
0.456787109375,
0.4921875,
0.456787109375,
0.5029296875,
0.1485595703125,
0.5029296875
],
[
0.51123046875,
0.495361328125,
0.880859375,
0.496337890625,
0.880859375,
0.51123046875,
0.51123046875,
0.51025390625
],
[
0.14697265625,
0.5048828125,
0.45751953125,
0.5048828125,
0.45751953125,
0.5185546875,
0.14697265625,
0.5185546875
],
[
0.51123046875,
0.51123046875,
0.8818359375,
0.5126953125,
0.8818359375,
0.52685546875,
0.51123046875,
0.525390625
],
[
0.1435546875,
0.5185546875,
0.45751953125,
0.52099609375,
0.45751953125,
0.53369140625,
0.1435546875,
0.5322265625
],
[
0.14697265625,
0.53369140625,
0.456787109375,
0.53369140625,
0.456787109375,
0.54736328125,
0.14697265625,
0.54736328125
],
[
0.5146484375,
0.529296875,
0.880859375,
0.529296875,
0.880859375,
0.5419921875,
0.5146484375,
0.5419921875
],
[
0.145263671875,
0.54736328125,
0.45751953125,
0.54638671875,
0.45751953125,
0.56103515625,
0.145263671875,
0.5625
],
[
0.51318359375,
0.54443359375,
0.880859375,
0.5458984375,
0.880859375,
0.55859375,
0.51318359375,
0.55810546875
],
[
0.145263671875,
0.5625,
0.456787109375,
0.5625,
0.456787109375,
0.576171875,
0.145263671875,
0.576171875
],
[
0.51318359375,
0.56103515625,
0.87744140625,
0.56103515625,
0.87744140625,
0.57470703125,
0.51318359375,
0.57470703125
],
[
0.1435546875,
0.57470703125,
0.456787109375,
0.576171875,
0.456787109375,
0.59033203125,
0.1435546875,
0.5888671875
],
[
0.5146484375,
0.5791015625,
0.87890625,
0.5791015625,
0.87890625,
0.5888671875,
0.5146484375,
0.5888671875
],
[
0.1485595703125,
0.5927734375,
0.456787109375,
0.5927734375,
0.456787109375,
0.6025390625,
0.1485595703125,
0.6025390625
],
[
0.51318359375,
0.59326171875,
0.880859375,
0.59326171875,
0.880859375,
0.60693359375,
0.51318359375,
0.60693359375
],
[
0.145263671875,
0.60498046875,
0.45751953125,
0.60498046875,
0.45751953125,
0.6181640625,
0.145263671875,
0.6181640625
],
[
0.51318359375,
0.60986328125,
0.880859375,
0.60986328125,
0.880859375,
0.62353515625,
0.51318359375,
0.62353515625
],
[
0.1485595703125,
0.62060546875,
0.455078125,
0.62060546875,
0.455078125,
0.63134765625,
0.1485595703125,
0.63134765625
],
[
0.5146484375,
0.6259765625,
0.880859375,
0.6259765625,
0.880859375,
0.6396484375,
0.5146484375,
0.6396484375
],
[
0.145263671875,
0.6328125,
0.456787109375,
0.6328125,
0.456787109375,
0.646484375,
0.145263671875,
0.646484375
],
[
0.51318359375,
0.64013671875,
0.87890625,
0.64013671875,
0.87890625,
0.65380859375,
0.51318359375,
0.65380859375
],
[
0.145263671875,
0.64794921875,
0.4072265625,
0.64794921875,
0.4072265625,
0.6611328125,
0.145263671875,
0.6611328125
],
[
0.51318359375,
0.65673828125,
0.87890625,
0.65673828125,
0.87890625,
0.67041015625,
0.51318359375,
0.67041015625
],
[
0.1158447265625,
0.67138671875,
0.151123046875,
0.67138671875,
0.151123046875,
0.68310546875,
0.1158447265625,
0.68310546875
],
[
0.145263671875,
0.66943359375,
0.2626953125,
0.67041015625,
0.2626953125,
0.68505859375,
0.145263671875,
0.68408203125
],
[
0.5146484375,
0.6748046875,
0.880859375,
0.6748046875,
0.880859375,
0.6884765625,
0.5146484375,
0.6884765625
],
[
0.118408203125,
0.6953125,
0.484375,
0.6953125,
0.484375,
0.708984375,
0.118408203125,
0.708984375
],
[
0.51318359375,
0.6904296875,
0.880859375,
0.6904296875,
0.880859375,
0.70361328125,
0.51318359375,
0.70361328125
],
[
0.51318359375,
0.7060546875,
0.8623046875,
0.7060546875,
0.8623046875,
0.7197265625,
0.51318359375,
0.7197265625
],
[
0.118408203125,
0.7119140625,
0.48193359375,
0.7119140625,
0.48193359375,
0.72265625,
0.118408203125,
0.72265625
],
[
0.118408203125,
0.72802734375,
0.486083984375,
0.72802734375,
0.486083984375,
0.7412109375,
0.118408203125,
0.7412109375
],
[
0.509765625,
0.7353515625,
0.84228515625,
0.7353515625,
0.84228515625,
0.7490234375,
0.509765625,
0.7490234375
],
[
0.118408203125,
0.74462890625,
0.48779296875,
0.74462890625,
0.48779296875,
0.75830078125,
0.118408203125,
0.75830078125
],
[
0.118408203125,
0.76025390625,
0.48779296875,
0.76025390625,
0.48779296875,
0.77392578125,
0.118408203125,
0.77392578125
],
[
0.51708984375,
0.7626953125,
0.880859375,
0.7626953125,
0.880859375,
0.775390625,
0.51708984375,
0.775390625
],
[
0.1158447265625,
0.77490234375,
0.48779296875,
0.775390625,
0.48779296875,
0.79052734375,
0.1158447265625,
0.7890625
],
[
0.51318359375,
0.779296875,
0.87744140625,
0.779296875,
0.87744140625,
0.7919921875,
0.51318359375,
0.7919921875
],
[
0.118408203125,
0.7919921875,
0.486083984375,
0.7919921875,
0.486083984375,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.51123046875,
0.79443359375,
0.880859375,
0.79443359375,
0.880859375,
0.80810546875,
0.51123046875,
0.80810546875
],
[
0.12176513671875,
0.81005859375,
0.486083984375,
0.81005859375,
0.486083984375,
0.82080078125,
0.12176513671875,
0.82080078125
],
[
0.509765625,
0.80859375,
0.744140625,
0.81103515625,
0.744140625,
0.82470703125,
0.509765625,
0.822265625
],
[
0.118408203125,
0.82470703125,
0.483642578125,
0.82470703125,
0.483642578125,
0.83837890625,
0.118408203125,
0.83837890625
],
[
0.12005615234375,
0.84130859375,
0.48779296875,
0.84130859375,
0.48779296875,
0.85498046875,
0.12005615234375,
0.85498046875
],
[
0.63232421875,
0.84130859375,
0.76171875,
0.84130859375,
0.76171875,
0.85546875,
0.63232421875,
0.85546875
],
[
0.118408203125,
0.85693359375,
0.48779296875,
0.85693359375,
0.48779296875,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.1175537109375,
0.87158203125,
0.489501953125,
0.87255859375,
0.489501953125,
0.88720703125,
0.1175537109375,
0.8857421875
],
[
0.53076171875,
0.87255859375,
0.880859375,
0.87255859375,
0.880859375,
0.8857421875,
0.53076171875,
0.8857421875
],
[
0.118408203125,
0.8896484375,
0.489501953125,
0.8896484375,
0.489501953125,
0.90283203125,
0.118408203125,
0.90283203125
],
[
0.51123046875,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.51123046875,
0.90283203125
],
[
0.12005615234375,
0.90673828125,
0.489501953125,
0.90673828125,
0.489501953125,
0.91748046875,
0.12005615234375,
0.91748046875
],
[
0.51318359375,
0.90625,
0.880859375,
0.90625,
0.880859375,
0.91943359375,
0.51318359375,
0.91943359375
]
] |
[
"利用GPT-4进行自动翻译后编辑",
"Vikas Raunak",
"Amr Sharaf",
"Yiren Wang",
"Hany Hassan Awadalla",
"Arul Menezes",
"",
"Microsoft Azure Al",
"Redmond, Washington",
" viraunak,amrsharaf,wangyiren,hanyh,arulm @microsoft.com",
"Abstract",
"因此，要问的一个相关问题是，是否可以利用",
"大型语言模型（LLM），例如GPT-3，GPT-4",
"While Neural Machine Translation (NMT) rep",
"resents the leading approach to Machine Trans-",
"和 PaLM, PaLM2 (Brown et al., 2020; OpenAI,",
"lation (MT), the outputs of NMT models still",
"2023; Anil et al., 2023; Chowdhery et al., 2022),",
"require translation post-editingto rectify er",
"它们已经展示了广泛的通用推理和基于知识的",
"rors and enhance quality under critical set.",
"功能用于翻译后编辑任务。基于LLM的自动",
"tings.Inthiswork,weformalizethetaskof",
"翻译后编辑可以帮助检测和修复翻译错误，以",
"direct translation post-editing with Large Lan",
"确保NMT输出的更高可靠性。除了缓解NMT",
"guage Models (LLMs) and explore the use of",
"中的可靠性问题之外，利用LLM进行后期编",
"GPT-4 to automatically post-edit NMT out",
"辑可能是适当的，因为最新的LLM具有先进",
"puts across several language pairs.Our re",
"的多语言理解能力，特别是GPT-4(OpenAI,",
"sultsdemonstratethatGPT-4isadeptattrans",
"lation post-editing, producing meaningful and",
"2023），并且可能具有将基于知识或特定文化",
"trustworthy edits to translations that helpim",
"的自定义应用于（Baker，2009）翻译的能力，这",
"prove its general quality as well as remove",
"很难通过源文本确定",
"different classes of major errors in transla",
"在这项工作中，我们探讨了GPT-4等最先进",
"tions.In particular,human evaluations on as",
"的LLM在直接设置的翻译后编辑任务中的功",
"sessing edit trustworthiness show that GPT-4",
"效，即。e。，在后期编辑之前，没有对翻译应",
"exhibits a large improvement over the prior",
"state-of-the-art LLM. Notably, we improve",
"用任何质量估计或错误检测步骤。我们的贡献",
"upon state-of-the-art performance on WMT-22",
"如下：（i）我们使用LLM正式确定直接翻译后",
"English-Chinese, English-German, Chinese",
"编辑的任务，并提出一系列研究问题和措施",
"English and German-English language pairs",
"以量化其效用，以改善从NMT模型获得的翻",
"usingGPT-4basedpost-editing,asevaluated",
"译。（ii）我们证明了GPT-4对具有高度可信度",
"by state-of-the-art MT quality metrics. How",
"的最新翻译产生了有意义的，与人类判断一致",
"ever,wealsoshowthatGPT-4 couldproduce",
"的编辑。（ii）我们证明了在WMT-22基准测",
"hallucinated edits, thereby urging caution in its",
"试（Koehnetal.,2022）上，与WMT-22的最佳",
"use as an expert translation post-editor",
"系统相比，在许多语言对上的翻译质量有所提",
"1",
"Introduction",
"高，从而提高了基于GPT-4的基于GPT的后",
"在Web挖掘的并行语料库上训练的最先进的",
"期编辑的英汉，英德，汉英和德语-英语的最新",
"翻译性能，如最新的MT质量指标所评估的",
"神经机器翻译（NMT）模型，即使对干较高资",
"源的语言对，也会遇到可靠性问题，尽管（He",
"2The Translation Post-Editing Task",
"et al., 2020; Gupta et al., 2020; Sun et al., 2020;",
"Wang et al., 2021; He et al., 2021; Raunak et al.",
"TaskDefinition：我们在生成设置中将后期编",
"2022;RaunakandMenezes,2022）的平均案例性",
"辑任务形式化如下：给定源（S）和翻译（T",
"能很高。许多先前的工作表明，在极少数情况",
"），提出对翻译（E）的改进，并使用建议的编",
"下，NMT中的并行数据和模型训练工件可能",
"辑（T）生成翻译，i。e。：",
"会表现为灾难性的输出，并且检测这种恶劣的",
"模型行为仍然是一项具有挑战性的任务(Rau-",
"(S,T) →E+T",
"nak et al., 2021; Tang et al., 2022; Guerreiro et al.,",
"2023b; Xu et al., 2023; Guerreiro et al., 2023a) 。",
"在此任务设置下，E表示由LLM描述的改",
"因此，编辑后神经机器翻译仍然是在翻译和本",
"进或编辑。在没有E的情况下，任务被简化",
"地化行业的关键环境中使用它们的重要练习。",
"为简单地生成改进的翻译，而无需任何中间推"
] |
{
"arxiv_id": "2305.14878",
"language": "zh",
"page_idx": 1
}
|
||
2305.14878-zh-2
|
2305.14878-zh-2.jpg
|
<table>
<tbody>
<tr>
<th>$E$</th>
<th>of Translation Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting)</th>
</tr>
<tr>
<td>$S$ S1</td>
<td>an ihren Sohn gerichtet, der Autismus hat und in einer privaten Pflegeeinrichtung lebt, sagte anstelle des Namens ihres Sohnes im Inneren, als Sie sie öffneten, hieSS es in den Briefer</td>
</tr>
</tbody>
</table>
Table 1: Post-Editing Task Example: GPT-4 produces edits that result in improving the general quality of the
translation as well as in
ofNMT
Section 3
undesirable artifacts across
<table>
<tbody>
<tr>
<th>Research Question</th>
<th>Measurement</th>
<th>Datasets</th>
</tr>
<tr>
<td>Nature of the Post-Edited Translation</td>
<td>TER$(T^{\prime}$,Zero-Shot) vs TER$(T^{\prime},T)$</td>
<td>$\mathbf{WMT-22}$</td>
</tr>
<tr>
<td>General Quality Improvements</td>
<td>$\mathrm{COMET}^{*}(S,T)\mathrm{~vs~COMET}^{*}(S,T^{\prime})$</td>
<td>$\mathbf{WMT-22}$</td>
</tr>
<tr>
<td>Edits On Human Annotated Error Spans</td>
<td>Edit Efficacy over Error Spans (E3S)</td>
<td>$\mathbf{WMT-20}$ 21,22</td>
</tr>
<tr>
<td>Trustworthiness of the e Proposed Edits</td>
<td>Edit Realization Rate (ERR)</td>
<td>$\mathbf{WMT-22}$</td>
</tr>
</tbody>
</table>
Table 2: Measuring Post-Editing Efficacy: Given, the Source $S$ , Translation $T$ and the Post-Edited Output $T^{\prime}$, we explore the four research questions in section 2 through experiments on the corresponding datasets, using the proposed measurements described in detail in Section 4.COMET* representsanyoftheCOMETMTmetrics.
理链或思想链 (CoT) (Wei et al., 2022b; Kojima 是根据任务说明编辑提供的初始翻译$T$。一个et al.,2022)。在整个工作中，我们将上述零拍 相关的问题是描述 CoT 在确定编辑后翻译的
性质方面的作用，即。e。编辑后的翻译是否更
接近初始翻译或 LLM 产生的零翻译
<NEXT$\_$LINE
CoT 设置中的后期编辑任务称为使用 CoT 进行后期编辑，而将没有$E$的设置称为没有 CoT 的后期编辑。表1显示了CoT 设置下编辑后任务的输入输出示例。此外，在整个工作中，我们将$Z$ 称为从用于后期编辑任务的同一 LLM 获得的给定源(S)的零转换。通过这种形式化，我们提出并调查了以下研究问题(RQ):
(RQ1) Nature of Post-Edited Translations: LLM 已被证明可以在零拍设置中跨多种语言对生成高质量，最先进的翻译(Hendy et al., 2023)。因此，在后期编辑期间，LLM 可能会生成无法识别所提供的初始翻译的翻译。因此，在使用 CoT 设置的后期编辑中，作为 CoT 提我们调查在翻译后编辑期间，LLM 是否从头 出的编辑是否实际上出现在 LLM 产生的改进开始生成改进的翻译(i。e。仅基于源$S$ ),还 翻译中？可以想象，LLM 可能会提出编辑建议
(RQ3) Modifying Human Annotated Error Spans: LLM 是否能够在后期编辑步骤中修改人工注释的翻译错误跨度？对人工注释的错误跨度进行高频率的修改，特别是如果它伴随着一般的质量改进，将意味着在识别翻译错误时与人类判断有更大的相关性
$( \textbf{RQ4) Trustworthiness of the Proposed Edits: }$
|
[
[
0.124267578125,
0.0872802734375,
0.8095703125,
0.08905029296875,
0.8095703125,
0.10394287109375,
0.124267578125,
0.1015625
],
[
0.1275634765625,
0.11224365234375,
0.869140625,
0.11224365234375,
0.869140625,
0.1258544921875,
0.1275634765625,
0.1258544921875
],
[
0.1275634765625,
0.127685546875,
0.86962890625,
0.127685546875,
0.86962890625,
0.141357421875,
0.1275634765625,
0.141357421875
],
[
0.1259765625,
0.1436767578125,
0.63232421875,
0.1436767578125,
0.63232421875,
0.15673828125,
0.1259765625,
0.15673828125
],
[
0.1259765625,
0.157958984375,
0.87158203125,
0.1591796875,
0.87158203125,
0.17333984375,
0.1259765625,
0.1728515625
],
[
0.124267578125,
0.17578125,
0.87158203125,
0.174560546875,
0.87158203125,
0.189453125,
0.124267578125,
0.190673828125
],
[
0.1259765625,
0.190673828125,
0.86962890625,
0.190673828125,
0.86962890625,
0.2037353515625,
0.1259765625,
0.2037353515625
],
[
0.1259765625,
0.2078857421875,
0.432373046875,
0.2078857421875,
0.432373046875,
0.221435546875,
0.1259765625,
0.221435546875
],
[
0.124267578125,
0.2286376953125,
0.31982421875,
0.23095703125,
0.31982421875,
0.2452392578125,
0.124267578125,
0.243408203125
],
[
0.1259765625,
0.2452392578125,
0.63623046875,
0.2452392578125,
0.63623046875,
0.260009765625,
0.1259765625,
0.260009765625
],
[
0.1259765625,
0.262451171875,
0.87158203125,
0.262451171875,
0.87158203125,
0.275634765625,
0.1259765625,
0.275634765625
],
[
0.1259765625,
0.279052734375,
0.60546875,
0.27783203125,
0.60546875,
0.292236328125,
0.1259765625,
0.29345703125
],
[
0.1259765625,
0.29443359375,
0.478515625,
0.29345703125,
0.478515625,
0.30810546875,
0.1259765625,
0.308837890625
],
[
0.1259765625,
0.311279296875,
0.59619140625,
0.311279296875,
0.59619140625,
0.32470703125,
0.1259765625,
0.32470703125
],
[
0.1275634765625,
0.327880859375,
0.297119140625,
0.327880859375,
0.297119140625,
0.341552734375,
0.1275634765625,
0.341552734375
],
[
0.1275634765625,
0.34326171875,
0.86962890625,
0.34326171875,
0.86962890625,
0.35693359375,
0.1275634765625,
0.35693359375
],
[
0.1259765625,
0.359375,
0.87158203125,
0.359375,
0.87158203125,
0.372314453125,
0.1259765625,
0.372314453125
],
[
0.1292724609375,
0.374755859375,
0.87158203125,
0.374755859375,
0.87158203125,
0.388427734375,
0.1292724609375,
0.388427734375
],
[
0.124267578125,
0.391357421875,
0.2493896484375,
0.391357421875,
0.2493896484375,
0.405029296875,
0.124267578125,
0.405029296875
],
[
0.1158447265625,
0.421630859375,
0.8818359375,
0.421630859375,
0.8818359375,
0.435302734375,
0.1158447265625,
0.435302734375
],
[
0.1175537109375,
0.435791015625,
0.880859375,
0.435791015625,
0.880859375,
0.449462890625,
0.1175537109375,
0.449462890625
],
[
0.124267578125,
0.458984375,
0.268798828125,
0.460205078125,
0.268798828125,
0.474365234375,
0.124267578125,
0.473876953125
],
[
0.43994140625,
0.458984375,
0.5458984375,
0.46142578125,
0.5458984375,
0.4755859375,
0.43994140625,
0.473876953125
],
[
0.77490234375,
0.460205078125,
0.8427734375,
0.460205078125,
0.8427734375,
0.474365234375,
0.77490234375,
0.474365234375
],
[
0.124267578125,
0.482666015625,
0.403076171875,
0.482666015625,
0.403076171875,
0.496337890625,
0.124267578125,
0.496337890625
],
[
0.443359375,
0.482666015625,
0.72900390625,
0.482666015625,
0.72900390625,
0.496337890625,
0.443359375,
0.496337890625
],
[
0.77685546875,
0.482177734375,
0.8486328125,
0.482177734375,
0.8486328125,
0.496337890625,
0.77685546875,
0.496337890625
],
[
0.124267578125,
0.50439453125,
0.359375,
0.5048828125,
0.359375,
0.51953125,
0.124267578125,
0.5185546875
],
[
0.443359375,
0.5048828125,
0.759765625,
0.5048828125,
0.759765625,
0.51953125,
0.443359375,
0.51953125
],
[
0.77490234375,
0.5048828125,
0.8486328125,
0.5048828125,
0.8486328125,
0.51953125,
0.77490234375,
0.51953125
],
[
0.1259765625,
0.52783203125,
0.71533203125,
0.52783203125,
0.71533203125,
0.5419921875,
0.1259765625,
0.5419921875
],
[
0.77490234375,
0.52685546875,
0.8505859375,
0.52685546875,
0.8505859375,
0.54150390625,
0.77490234375,
0.54150390625
],
[
0.77490234375,
0.54345703125,
0.8271484375,
0.54345703125,
0.8271484375,
0.55859375,
0.77490234375,
0.55859375
],
[
0.1259765625,
0.56640625,
0.4130859375,
0.56640625,
0.4130859375,
0.58056640625,
0.1259765625,
0.58056640625
],
[
0.441650390625,
0.56640625,
0.654296875,
0.56640625,
0.654296875,
0.580078125,
0.441650390625,
0.580078125
],
[
0.77685546875,
0.56640625,
0.8486328125,
0.56640625,
0.8486328125,
0.58056640625,
0.77685546875,
0.58056640625
],
[
0.420654296875,
0.572265625,
0.43408203125,
0.57080078125,
0.435791015625,
0.57373046875,
0.42236328125,
0.576171875
],
[
0.1158447265625,
0.5966796875,
0.875,
0.5966796875,
0.875,
0.60986328125,
0.1158447265625,
0.60986328125
],
[
0.118408203125,
0.61328125,
0.880859375,
0.61328125,
0.880859375,
0.62353515625,
0.118408203125,
0.62353515625
],
[
0.1158447265625,
0.62451171875,
0.8671875,
0.62451171875,
0.8671875,
0.63818359375,
0.1158447265625,
0.63818359375
],
[
0.1158447265625,
0.65380859375,
0.490234375,
0.65283203125,
0.490234375,
0.66748046875,
0.1158447265625,
0.66845703125
],
[
0.51318359375,
0.65478515625,
0.8369140625,
0.65478515625,
0.8369140625,
0.66845703125,
0.51318359375,
0.66845703125
],
[
0.86376953125,
0.6591796875,
0.873046875,
0.6591796875,
0.873046875,
0.6630859375,
0.86376953125,
0.6630859375
],
[
0.1175537109375,
0.67041015625,
0.48779296875,
0.67041015625,
0.48779296875,
0.68408203125,
0.1175537109375,
0.68408203125
],
[
0.51318359375,
0.67041015625,
0.87890625,
0.67041015625,
0.87890625,
0.68408203125,
0.51318359375,
0.68408203125
],
[
0.1175537109375,
0.68603515625,
0.48779296875,
0.68603515625,
0.48779296875,
0.69970703125,
0.1175537109375,
0.69970703125
],
[
0.51123046875,
0.68505859375,
0.880859375,
0.68603515625,
0.880859375,
0.70068359375,
0.51123046875,
0.69970703125
],
[
0.1175537109375,
0.70166015625,
0.48779296875,
0.70068359375,
0.48779296875,
0.71533203125,
0.1175537109375,
0.71630859375
],
[
0.51318359375,
0.70166015625,
0.80029296875,
0.70166015625,
0.80029296875,
0.71630859375,
0.51318359375,
0.71630859375
],
[
0.1175537109375,
0.71728515625,
0.483642578125,
0.71728515625,
0.483642578125,
0.73095703125,
0.1175537109375,
0.73095703125
],
[
0.5322265625,
0.72021484375,
0.64208984375,
0.72021484375,
0.64208984375,
0.732421875,
0.5322265625,
0.732421875
],
[
0.12005615234375,
0.73388671875,
0.484375,
0.73388671875,
0.484375,
0.7490234375,
0.12005615234375,
0.7490234375
],
[
0.12005615234375,
0.75048828125,
0.484375,
0.75048828125,
0.484375,
0.76416015625,
0.12005615234375,
0.76416015625
],
[
0.52294921875,
0.74755859375,
0.8818359375,
0.74755859375,
0.8818359375,
0.76123046875,
0.52294921875,
0.76123046875
],
[
0.118408203125,
0.7666015625,
0.483642578125,
0.7666015625,
0.483642578125,
0.77978515625,
0.118408203125,
0.77978515625
],
[
0.51123046875,
0.76318359375,
0.880859375,
0.76318359375,
0.880859375,
0.77783203125,
0.51123046875,
0.77783203125
],
[
0.12005615234375,
0.78369140625,
0.4609375,
0.78369140625,
0.4609375,
0.79736328125,
0.12005615234375,
0.79736328125
],
[
0.51318359375,
0.77978515625,
0.880859375,
0.77978515625,
0.880859375,
0.79345703125,
0.51318359375,
0.79345703125
],
[
0.515625,
0.7958984375,
0.87890625,
0.7958984375,
0.87890625,
0.81005859375,
0.515625,
0.81005859375
],
[
0.1275634765625,
0.80810546875,
0.486083984375,
0.80810546875,
0.486083984375,
0.82177734375,
0.1275634765625,
0.82177734375
],
[
0.51318359375,
0.81005859375,
0.880859375,
0.81103515625,
0.880859375,
0.826171875,
0.51318359375,
0.82470703125
],
[
0.1158447265625,
0.822265625,
0.48779296875,
0.82373046875,
0.48779296875,
0.83837890625,
0.1158447265625,
0.83740234375
],
[
0.51318359375,
0.82763671875,
0.75244140625,
0.82763671875,
0.75244140625,
0.8427734375,
0.51318359375,
0.8427734375
],
[
0.118408203125,
0.84033203125,
0.484375,
0.84033203125,
0.484375,
0.85400390625,
0.118408203125,
0.85400390625
],
[
0.1175537109375,
0.85546875,
0.483642578125,
0.85546875,
0.483642578125,
0.87060546875,
0.1175537109375,
0.87060546875
],
[
0.51904296875,
0.85546875,
0.880859375,
0.85546875,
0.880859375,
0.869140625,
0.51904296875,
0.869140625
],
[
0.118408203125,
0.87255859375,
0.484375,
0.87255859375,
0.484375,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.51318359375,
0.87158203125,
0.880859375,
0.87158203125,
0.880859375,
0.8857421875,
0.51318359375,
0.8857421875
],
[
0.1175537109375,
0.88818359375,
0.486083984375,
0.88818359375,
0.486083984375,
0.90283203125,
0.1175537109375,
0.90283203125
],
[
0.51123046875,
0.88818359375,
0.8818359375,
0.88720703125,
0.8818359375,
0.90185546875,
0.51123046875,
0.90283203125
],
[
0.118408203125,
0.90478515625,
0.486083984375,
0.90478515625,
0.486083984375,
0.91845703125,
0.118408203125,
0.91845703125
],
[
0.51123046875,
0.90380859375,
0.880859375,
0.90380859375,
0.880859375,
0.91845703125,
0.51123046875,
0.91845703125
]
] |
[
"Example of Translation Post-Editing using GPT-4 (Verbatim Outputs under the CoT Setting)",
"Source : They were addressed to her son, who has autism and lives in a private care facility, she said.",
"But instead of her son's name inside when you opened them, the letters said Dear Maine's Department",
"of Health and Human Services -in Cincinnati, she told local media.",
"Translation : Sie waren an ihren Sohn gerichtet, der Autismus hat und in einer privaten Pflegeeinrich-",
"tung lebt, sagte sie. Aber anstelle des Namens ihres Sohnes im Inneren, als Sie sie offneten, hiess",
"es in den Briefen Dear Maine Dear Maine 's Department of Health and Human ServicesServices - in",
"Cincinnati, sagte sie den lokalen Medien.",
"Proposed Improvements :",
"1. Remove the repetition of \"Dear Maine \" in the German translation.",
"2. Correct the translation of \"Dear Maine's Department of Health and Human Services\" to \"Sehr",
"geehrtes Department of Health and Human Services von Maine\".",
"3. Replace\"ServicesServices\"with\"Services\".",
"4. Add a comma after \"Cincinnati\" for better sentence structure.",
"Improved Translation :",
"Sie waren an ihren Sohn gerichtet, der Autismus hat und in einer privaten Pflegeeinrichtung lebt, sagte",
"sie. Aber anstelle des Namens ihres Sohnes im Inneren, als Sie sie offneten, hiesS es in den Briefen",
"\"Sehr geehrtes Department of Health and Human Services von Maine\"- in Cincinnati, sagte sie den",
"lokalen Medien.",
"Table 1: Post-Editing Task Example : GPT-4 produces edits that result in improving the general quality of the",
"translation as well as in removing undesirable artifacts across a range of NMT systems, as quantified in Section 3 .",
"ResearchQuestion",
"Measurement",
"Datasets",
"Nature of the Post-Edited Translation",
"TER(T',Zero-Shot) vs TER(T', T)",
"WMT-22",
"GeneralQualityImprovements",
"COMET*(S,T)vs COMET*(S,T')",
"WMT-22",
"Edits OnHuman Annotated Error SpansEditEfficacyoverError Spans (E3S)",
"WMT-20,",
"21, 22",
"Trustworthiness of the Proposed Edits",
"Edit Realization Rate (ERR)",
"WMT-22",
"",
"Table 2: Measuring Post-Editing Efficacy : Given, the Source S , Translation T and the Post-Edited Output T'",
"we explorethefourresearchquestions in section2through experiments onthe corresponding datasets,using the",
"proposed measurements described in detail in Section 4 . COMET * represents any of the COMET MT metrics.",
"理链或思想链（CoT）（Weietal.，2022b;Kojima",
"是根据任务说明编辑提供的初始翻译T。",
"r",
"etal.,2022）。在整个工作中，我们将上述零拍",
"相关的问题是描述CoT在确定编辑后翻译的",
"CoT设置中的后期编辑任务称为使用CoT进",
"性质方面的作用，即。e。编辑后的翻译是否更",
"行后期编辑，而将没有E的设置称为没有CoT",
"接近初始翻译或LLM产生的零翻译",
"的后期编辑。表1显示了CoT设置下编辑后任",
"<NEXT LINE",
"务的输人输出示例。此外，在整个工作中，我",
"们将Z称为从用于后期编辑任务的同一LLM",
"(RQ3) Modifying Human Annotated Error",
"获得的给定源（S）的零转换。通过这种形式",
"Spans：LLM是否能够在后期编辑步骤中修",
"化，我们提出并调查了以下研究问题（RQ)：",
"改人工注释的翻译错误跨度？对人工注释的错",
"误跨度进行高频率的修改，特别是如果它伴随",
"(RQI) Nature of Post-Edited Translations:",
"着一般的质量改进，将意味着在识别翻译错误",
"LLM已被证明可以在零拍设置中跨多种语言",
"时与人类判断有更大的相关性",
"对生成高质量，最先进的翻译(Hendyetal.,",
"2023）。因此，在后期编辑期间，LLM可能会",
"(RQ4) Trustworthiness of the Proposed Edits:",
"生成无法识别所提供的初始翻译的翻译。因此",
"在使用CoT设置的后期编辑中，作为CoT提",
"我们调查在翻译后编辑期间，LLM是否从头",
"出的编辑是否实际上出现在LLM产生的改进",
"开始生成改进的翻译(i。e。仅基于源S），还",
"翻译中？可以想象，LLM可能会提出编辑建议"
] |
{
"arxiv_id": "2305.14878",
"language": "zh",
"page_idx": 2
}
|
||
2305.14878-zh-3
|
2305.14878-zh-3.jpg
|
或产生在同一模型(Ye and Durrett, 2022; Turpin et al., 2023) 产生的最终编辑后翻译中未实现的思想链。但是，如果编辑后解释或 CoT 是翻译后编辑过程的当务之急，那么除了最终的翻译质量之外，检查建议的编辑的保真度就变得至关重要。更高的拟议编辑实现率也将有助于建立 LLM 作为专家翻译后编辑的可信度
在接下来的部分中，我们将使用最先进的LLM 探讨上述四个研究问题。我们在 3 节中描述了我们的实验设置&在4节中给出了结果
# 3 Experimental Settings
Datasets: 我们对 WMT-22 通用 MT 翻译任务数据集 (Kocmi et al., 2022)以及 WMT-20 和WMT-21 进行了实验。带有 MQM(多维质量度量框架)错误$^1$Freitag et al.(2021)注释的新闻翻译任务提交。对于与 MQM 注释的 WMT-20和 WMT-21 系统输出相关的后期编辑实验，我们使用具有主要错误的样本作为注释进行实验，而我们在整个过程中使用完整的 WMT-22数据集进行实验。我们在大多数实验中使用最新的 WMT-22 测试集，其管理超出了GPT-4 和其他正在研究的 LLM 的训练截止日期$^{2}$
LLMs and Baselines: 我们用 GPT-4 和 gpt-3进行实验。5-turbo 在我们的实验中。这些模型代表了最有能力的公开 LLM(Liang et al.,2022) 。我们使用一个提示，将系统角色描述为翻译后编辑器，并在 CoT 设置下，指示 LLM 在生成最终的编辑后翻译($T^\prime$)之前，对给定源($S$ )的提供的翻译($T$ )提出改进。对于后期编辑， 我们在三种设置下进行实验：(i)使用 CoT 进行后期编辑，(ii)不使用 CoT 进行后期编辑， 以及(iii)使用结构化 CoT(SCoT)进行后期编辑。SCoT 基线使用 Freitag et al.(2021)中的MQM 注释指令在源-转换对上以 MQM 注释的形式生成中间 CoT。我们在附录 C 中描述了用于三个基线的提示。为了在 WMT-22 上生成初始翻译，我们使用 Microsoft-Translator, 这是最强的公开 MT 系统 (Raunak et al., 2022; Hendy et al., 2023) 之一。对于 WMT-20 和 WMT-21系统，我们将不同 NMT 系统提供并在 Freitag et al. (2021) 中注释的翻译作为应用后期编辑的初始翻译
Metrics and Evaluation: 对于提出的四个研
究问题中的每一个，我们使用表 2 中突出显
$^{1}$https://github.com/google/
wmt-mqm-human-evaluation
$^{2}$LLMs:
https: / / platform. openai . com/ docs/
models
示的指标。我们在相关章节中解释这些测量。对于一般质量测量，我们使用四个 COMET (Rei et al., 2020) 模型$^3:$无参考的 COMET-QE (wmt20-comet-qe-da),COMET-KIWI(wmt-22- cometkiwi-da)质量估计(QE)模型和基于参考的 COMET-20 (wmt20-mwtet-。为了衡量翻译的相似性，我们使用 SacreBLEU (Post,2018) 中的翻译编辑率 (TER) (Snover et al.,2006) 实现
4 Results and Measurements
4.1 Nature of Post-Edited Translations
为了衡量 LLM 产生的编辑后翻译是否符合编辑所提供的初始翻译，我们计算编辑后翻译的翻译编辑率(TER)(Snover et al.,2006)与使用相同 LLM 获得的零拍翻译，并将其与编辑后翻译的 TER 与初始翻译进行比较。TER( $T^{\prime},T$ )的值越高，意味着编辑后的翻译($T^{\prime}$)越接近初始翻译($T$),并且 LLM 坚持编辑初始翻译的任务
<table>
<tbody>
<tr>
<th>PE Setting</th>
<th>$TER$ ${\mathcal{R} }( T^{\prime }, Z$ )</th>
<th>$\operatorname{TER}\left(T^{\prime},T\right)$</th>
</tr>
<tr>
<td>With CoT</td>
<td>92.0</td>
<td>70.3</td>
</tr>
<tr>
<td>$\mathbf{Without}$ CoT</td>
<td>84.6</td>
<td>94.5</td>
</tr>
</tbody>
</table>
Table 3: WMT-22 En-Zh: The post-edited translations $(T^{\prime})$ are closer to the initial translations$(T)$ than the zero-shot translations $(Z)$ in the CoT setting.
<table>
<tbody>
<tr>
<th>PE Setting</th>
<th>$TER$ {}( $T^{\prime }$ , $Z$</th>
<th>$Z$ ) $T$ $TER$ $.(T^{\prime\prime}$</th>
</tr>
<tr>
<td>With CoT</td>
<td>42.9</td>
<td>9 22.0</td>
</tr>
<tr>
<td>$\mathbf{Without}$ CoT</td>
<td>38.1</td>
<td>1 34.9</td>
</tr>
</tbody>
</table>
Table 4: WMT-22 Zh-En: The post-edited translations $(T^{\prime})$ are closer to the initial translations $(T)$ than the zero-shot translations $(Z)$ in the CoT setting.
Impact of CoT: 表 3 描述了我们在 WMT- 22En-Zh 上的结果，表 4 描述了我们在 Zh-En 上使用 GPT-4 进行后期编辑的结果。我们发现CoT 限制了最终翻译更接近初始翻译。在没有CoT 的后期编辑设置中，即使 TER 差异比 CoT 设置中的差异小得多，最终的翻译也更接近于零拍翻译
Discussion 我们发现上述结果适用于不同的指标，例如编辑距离，BLEU (Post, 2018)或ChrF (Popović, 2015)以及 WMT-22 语言对和gpt-3。5-涡轮增压。此结果还显示了CoT 设
$^3$COMET: https: /gi thub. com/Unbabel/COMET
|
[
[
0.1175537109375,
0.08489990234375,
0.483642578125,
0.08489990234375,
0.483642578125,
0.09857177734375,
0.1175537109375,
0.09857177734375
],
[
0.5146484375,
0.08612060546875,
0.880859375,
0.08612060546875,
0.880859375,
0.09979248046875,
0.5146484375,
0.09979248046875
],
[
0.1175537109375,
0.1015625,
0.484375,
0.1015625,
0.484375,
0.11517333984375,
0.1175537109375,
0.11517333984375
],
[
0.51318359375,
0.1015625,
0.880859375,
0.1015625,
0.880859375,
0.11517333984375,
0.51318359375,
0.11517333984375
],
[
0.118408203125,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.118408203125,
0.1318359375
],
[
0.51123046875,
0.11639404296875,
0.8818359375,
0.1175537109375,
0.8818359375,
0.1318359375,
0.51123046875,
0.1312255859375
],
[
0.118408203125,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.118408203125,
0.1484375
],
[
0.51123046875,
0.1341552734375,
0.8818359375,
0.1341552734375,
0.8818359375,
0.1484375,
0.51123046875,
0.1484375
],
[
0.118408203125,
0.15087890625,
0.484375,
0.15087890625,
0.484375,
0.1644287109375,
0.118408203125,
0.1644287109375
],
[
0.51318359375,
0.15087890625,
0.87890625,
0.15087890625,
0.87890625,
0.1644287109375,
0.51318359375,
0.1644287109375
],
[
0.12005615234375,
0.168701171875,
0.484375,
0.168701171875,
0.484375,
0.1787109375,
0.12005615234375,
0.1787109375
],
[
0.51123046875,
0.1650390625,
0.87744140625,
0.1650390625,
0.87744140625,
0.1787109375,
0.51123046875,
0.1787109375
],
[
0.118408203125,
0.182861328125,
0.42236328125,
0.182861328125,
0.42236328125,
0.196533203125,
0.118408203125,
0.196533203125
],
[
0.51318359375,
0.182861328125,
0.880859375,
0.182861328125,
0.880859375,
0.196533203125,
0.51318359375,
0.196533203125
],
[
0.1368408203125,
0.1995849609375,
0.484375,
0.1995849609375,
0.484375,
0.213134765625,
0.1368408203125,
0.213134765625
],
[
0.51318359375,
0.198974609375,
0.880859375,
0.198974609375,
0.880859375,
0.2120361328125,
0.51318359375,
0.2120361328125
],
[
0.1175537109375,
0.215576171875,
0.484375,
0.215576171875,
0.484375,
0.2298583984375,
0.1175537109375,
0.2298583984375
],
[
0.5146484375,
0.21142578125,
0.53662109375,
0.2161865234375,
0.52880859375,
0.2298583984375,
0.50732421875,
0.2244873046875
],
[
0.118408203125,
0.232177734375,
0.484375,
0.232177734375,
0.484375,
0.2452392578125,
0.118408203125,
0.2452392578125
],
[
0.51123046875,
0.2410888671875,
0.78076171875,
0.2423095703125,
0.78076171875,
0.257080078125,
0.51123046875,
0.255859375
],
[
0.1175537109375,
0.2476806640625,
0.139404296875,
0.2476806640625,
0.139404296875,
0.26123046875,
0.1175537109375,
0.26123046875
],
[
0.51123046875,
0.268310546875,
0.8330078125,
0.268310546875,
0.8330078125,
0.281982421875,
0.51123046875,
0.281982421875
],
[
0.11419677734375,
0.276611328125,
0.344970703125,
0.279052734375,
0.344970703125,
0.29345703125,
0.11419677734375,
0.29150390625
],
[
0.51318359375,
0.289306640625,
0.87890625,
0.290283203125,
0.87890625,
0.3046875,
0.51318359375,
0.303955078125
],
[
0.12176513671875,
0.3046875,
0.48779296875,
0.303955078125,
0.48779296875,
0.318359375,
0.12176513671875,
0.319580078125
],
[
0.51318359375,
0.30712890625,
0.87890625,
0.30712890625,
0.87890625,
0.320556640625,
0.51318359375,
0.320556640625
],
[
0.1175537109375,
0.319580078125,
0.486083984375,
0.319580078125,
0.486083984375,
0.334228515625,
0.1175537109375,
0.334228515625
],
[
0.51123046875,
0.32177734375,
0.87744140625,
0.32177734375,
0.87744140625,
0.3349609375,
0.51123046875,
0.3349609375
],
[
0.1175537109375,
0.33740234375,
0.486083984375,
0.33740234375,
0.486083984375,
0.350830078125,
0.1175537109375,
0.350830078125
],
[
0.51318359375,
0.339111328125,
0.880859375,
0.339111328125,
0.880859375,
0.352783203125,
0.51318359375,
0.352783203125
],
[
0.12005615234375,
0.352783203125,
0.484375,
0.352783203125,
0.484375,
0.36767578125,
0.12005615234375,
0.36767578125
],
[
0.51123046875,
0.35400390625,
0.880859375,
0.352783203125,
0.880859375,
0.36767578125,
0.51123046875,
0.3681640625
],
[
0.118408203125,
0.369384765625,
0.48779296875,
0.369384765625,
0.48779296875,
0.383056640625,
0.118408203125,
0.383056640625
],
[
0.51318359375,
0.37060546875,
0.880859375,
0.37060546875,
0.880859375,
0.38427734375,
0.51318359375,
0.38427734375
],
[
0.118408203125,
0.385498046875,
0.486083984375,
0.385498046875,
0.486083984375,
0.3984375,
0.118408203125,
0.3984375
],
[
0.51123046875,
0.385986328125,
0.87890625,
0.385986328125,
0.87890625,
0.399658203125,
0.51123046875,
0.399658203125
],
[
0.118408203125,
0.402099609375,
0.484375,
0.402099609375,
0.484375,
0.4150390625,
0.118408203125,
0.4150390625
],
[
0.51123046875,
0.402587890625,
0.5693359375,
0.402587890625,
0.5693359375,
0.41748046875,
0.51123046875,
0.41748046875
],
[
0.1175537109375,
0.41748046875,
0.484375,
0.416259765625,
0.484375,
0.43115234375,
0.1175537109375,
0.431640625
],
[
0.118408203125,
0.43408203125,
0.484375,
0.43408203125,
0.484375,
0.44775390625,
0.118408203125,
0.44775390625
],
[
0.52490234375,
0.435791015625,
0.60693359375,
0.438232421875,
0.60693359375,
0.453125,
0.52490234375,
0.45068359375
],
[
0.63818359375,
0.438232421875,
0.87744140625,
0.438232421875,
0.87744140625,
0.451904296875,
0.63818359375,
0.451904296875
],
[
0.118408203125,
0.449462890625,
0.484375,
0.449462890625,
0.484375,
0.46435546875,
0.118408203125,
0.46435546875
],
[
0.52294921875,
0.457275390625,
0.603515625,
0.458984375,
0.603515625,
0.474365234375,
0.52294921875,
0.47265625
],
[
0.67822265625,
0.457763671875,
0.71728515625,
0.457763671875,
0.71728515625,
0.473876953125,
0.67822265625,
0.473876953125
],
[
0.8037109375,
0.4560546875,
0.8427734375,
0.458984375,
0.84033203125,
0.474365234375,
0.8017578125,
0.47265625
],
[
0.118408203125,
0.466064453125,
0.43408203125,
0.466064453125,
0.43408203125,
0.479736328125,
0.118408203125,
0.479736328125
],
[
0.52490234375,
0.473876953125,
0.62744140625,
0.4755859375,
0.62744140625,
0.4912109375,
0.52490234375,
0.4892578125
],
[
0.67822265625,
0.474365234375,
0.71728515625,
0.474365234375,
0.71728515625,
0.490478515625,
0.67822265625,
0.490478515625
],
[
0.8037109375,
0.474365234375,
0.84228515625,
0.474365234375,
0.84228515625,
0.490478515625,
0.8037109375,
0.490478515625
],
[
0.12176513671875,
0.493408203125,
0.486083984375,
0.493408203125,
0.486083984375,
0.50732421875,
0.12176513671875,
0.50732421875
],
[
0.118408203125,
0.51025390625,
0.484375,
0.51025390625,
0.484375,
0.52392578125,
0.118408203125,
0.52392578125
],
[
0.51123046875,
0.5087890625,
0.880859375,
0.5087890625,
0.880859375,
0.5224609375,
0.51123046875,
0.5224609375
],
[
0.118408203125,
0.525390625,
0.484375,
0.525390625,
0.484375,
0.5390625,
0.118408203125,
0.5390625
],
[
0.51318359375,
0.52490234375,
0.87890625,
0.52490234375,
0.87890625,
0.53515625,
0.51318359375,
0.53515625
],
[
0.51123046875,
0.5361328125,
0.82275390625,
0.5380859375,
0.82275390625,
0.552734375,
0.51123046875,
0.55029296875
],
[
0.118408203125,
0.5419921875,
0.484375,
0.5419921875,
0.484375,
0.5556640625,
0.118408203125,
0.5556640625
],
[
0.118408203125,
0.55810546875,
0.484375,
0.55810546875,
0.484375,
0.57177734375,
0.118408203125,
0.57177734375
],
[
0.12005615234375,
0.576171875,
0.484375,
0.576171875,
0.484375,
0.5859375,
0.12005615234375,
0.5859375
],
[
0.52294921875,
0.57666015625,
0.60693359375,
0.5791015625,
0.60693359375,
0.59326171875,
0.52294921875,
0.59130859375
],
[
0.6396484375,
0.5791015625,
0.87744140625,
0.5791015625,
0.87744140625,
0.5927734375,
0.6396484375,
0.5927734375
],
[
0.118408203125,
0.59033203125,
0.484375,
0.59033203125,
0.484375,
0.60400390625,
0.118408203125,
0.60400390625
],
[
0.67822265625,
0.5986328125,
0.71728515625,
0.5986328125,
0.71728515625,
0.615234375,
0.67822265625,
0.615234375
],
[
0.8037109375,
0.5986328125,
0.84228515625,
0.5986328125,
0.84228515625,
0.615234375,
0.8037109375,
0.615234375
],
[
0.118408203125,
0.60546875,
0.48779296875,
0.60546875,
0.48779296875,
0.619140625,
0.118408203125,
0.619140625
],
[
0.5263671875,
0.599609375,
0.603515625,
0.599609375,
0.603515625,
0.6142578125,
0.5263671875,
0.6142578125
],
[
0.67822265625,
0.615234375,
0.7138671875,
0.615234375,
0.7138671875,
0.63134765625,
0.67822265625,
0.63134765625
],
[
0.8037109375,
0.615234375,
0.84033203125,
0.615234375,
0.84033203125,
0.63134765625,
0.8037109375,
0.63134765625
],
[
0.118408203125,
0.62158203125,
0.484375,
0.62158203125,
0.484375,
0.6357421875,
0.118408203125,
0.6357421875
],
[
0.5283203125,
0.61767578125,
0.62451171875,
0.61767578125,
0.62451171875,
0.62890625,
0.5283203125,
0.62890625
],
[
0.1175537109375,
0.63720703125,
0.486083984375,
0.63818359375,
0.486083984375,
0.65283203125,
0.1175537109375,
0.65185546875
],
[
0.118408203125,
0.65380859375,
0.486083984375,
0.65380859375,
0.486083984375,
0.66845703125,
0.118408203125,
0.66845703125
],
[
0.509765625,
0.6484375,
0.880859375,
0.6494140625,
0.880859375,
0.66455078125,
0.509765625,
0.6630859375
],
[
0.509765625,
0.6630859375,
0.87890625,
0.6630859375,
0.87890625,
0.6767578125,
0.509765625,
0.6767578125
],
[
0.1175537109375,
0.67041015625,
0.484375,
0.67041015625,
0.484375,
0.68408203125,
0.1175537109375,
0.68408203125
],
[
0.51123046875,
0.677734375,
0.82373046875,
0.6796875,
0.82373046875,
0.693359375,
0.51123046875,
0.69140625
],
[
0.1175537109375,
0.68505859375,
0.483642578125,
0.68505859375,
0.483642578125,
0.69873046875,
0.1175537109375,
0.69873046875
],
[
0.12005615234375,
0.70166015625,
0.484375,
0.70166015625,
0.484375,
0.71630859375,
0.12005615234375,
0.71630859375
],
[
0.51904296875,
0.71044921875,
0.8818359375,
0.71044921875,
0.8818359375,
0.7236328125,
0.51904296875,
0.7236328125
],
[
0.118408203125,
0.71875,
0.484375,
0.71875,
0.484375,
0.732421875,
0.118408203125,
0.732421875
],
[
0.51318359375,
0.72705078125,
0.8818359375,
0.72705078125,
0.8818359375,
0.74072265625,
0.51318359375,
0.74072265625
],
[
0.1175537109375,
0.73388671875,
0.484375,
0.73291015625,
0.484375,
0.74755859375,
0.1175537109375,
0.7490234375
],
[
0.515625,
0.7421875,
0.880859375,
0.7421875,
0.880859375,
0.755859375,
0.515625,
0.755859375
],
[
0.1175537109375,
0.75048828125,
0.252685546875,
0.75048828125,
0.252685546875,
0.76416015625,
0.1175537109375,
0.76416015625
],
[
0.2476806640625,
0.75,
0.483642578125,
0.75,
0.483642578125,
0.76318359375,
0.2476806640625,
0.76318359375
],
[
0.51318359375,
0.7587890625,
0.880859375,
0.7587890625,
0.880859375,
0.7724609375,
0.51318359375,
0.7724609375
],
[
0.12005615234375,
0.7666015625,
0.484375,
0.7666015625,
0.484375,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.51318359375,
0.775390625,
0.8818359375,
0.775390625,
0.8818359375,
0.7890625,
0.51318359375,
0.7890625
],
[
0.118408203125,
0.783203125,
0.484375,
0.783203125,
0.484375,
0.79638671875,
0.118408203125,
0.79638671875
],
[
0.51318359375,
0.79052734375,
0.87744140625,
0.79052734375,
0.87744140625,
0.80419921875,
0.51318359375,
0.80419921875
],
[
0.12005615234375,
0.80029296875,
0.2091064453125,
0.80029296875,
0.2091064453125,
0.8125,
0.12005615234375,
0.8125
],
[
0.51318359375,
0.80712890625,
0.58837890625,
0.80712890625,
0.58837890625,
0.82177734375,
0.51318359375,
0.82177734375
],
[
0.1234130859375,
0.82666015625,
0.484375,
0.82666015625,
0.484375,
0.84033203125,
0.1234130859375,
0.84033203125
],
[
0.515625,
0.83056640625,
0.880859375,
0.83203125,
0.880859375,
0.8466796875,
0.515625,
0.845703125
],
[
0.118408203125,
0.84326171875,
0.484375,
0.84326171875,
0.484375,
0.85693359375,
0.118408203125,
0.85693359375
],
[
0.51318359375,
0.8486328125,
0.880859375,
0.8486328125,
0.880859375,
0.8623046875,
0.51318359375,
0.8623046875
],
[
0.139404296875,
0.8681640625,
0.341796875,
0.8681640625,
0.341796875,
0.8818359375,
0.139404296875,
0.8818359375
],
[
0.51318359375,
0.8642578125,
0.880859375,
0.8642578125,
0.880859375,
0.87890625,
0.51318359375,
0.87890625
],
[
0.1158447265625,
0.8798828125,
0.301513671875,
0.88134765625,
0.301513671875,
0.892578125,
0.1158447265625,
0.89111328125
],
[
0.51123046875,
0.88134765625,
0.880859375,
0.8798828125,
0.880859375,
0.89453125,
0.51123046875,
0.8955078125
],
[
0.1376953125,
0.89111328125,
0.18896484375,
0.8935546875,
0.187255859375,
0.90673828125,
0.1368408203125,
0.90478515625
],
[
0.2333984375,
0.8935546875,
0.484375,
0.8935546875,
0.484375,
0.90673828125,
0.2333984375,
0.90673828125
],
[
0.1175537109375,
0.90625,
0.166259765625,
0.90625,
0.166259765625,
0.91748046875,
0.1175537109375,
0.91748046875
],
[
0.5341796875,
0.90380859375,
0.83447265625,
0.90478515625,
0.83447265625,
0.91845703125,
0.5341796875,
0.91748046875
]
] |
[
"或产生在同一模型（YeandDurrett，2022：Turpin",
"示的指标。我们在相关章节中解释这些测量。",
"etal.,2023)产生的最终编辑后翻译中未实现的",
"对于一般质量测量，我们使用四个COMET",
"思想链。但是，如果编辑后解释或CoT是翻译",
"(Reietal.,2020)模型3：无参考的COMET-QE",
"后编辑过程的当务之急，那么除了最终的翻译",
"(wmt20-comet-qe-da), COMET-KIWI (wmt-22-",
"质量之外，检查建议的编辑的保真度就变得至",
"cometkiwi-da）质量估计（QE）模型和基于参",
"关重要。更高的拟议编辑实现率也将有助于建",
"考的COMET-20（wmt20-mwtet-。为了衡量翻",
"立LLM作为专家翻译后编辑的可信度",
"译的相似性，我们使用SacreBLEU(Post,2018)",
"在接下来的部分中，我们将使用最先进的",
"中的翻译编辑率（TER）（Snoveretal.,2006)实",
"LLM探讨上述四个研究问题。我们在3节中",
"现",
"描述了我们的实验设置&在4节中给出了结",
"4Results and Measurements",
"果",
"4.1 Nature of Post-Edited Translations",
"3Experimental Settings",
"为了衡量LLM产生的编辑后翻译是否符合编",
"Datasets：我们对WMT-22通用MT翻译任",
"辑所提供的初始翻译，我们计算编辑后翻译的",
"务数据集（Kocmietal.,2022）以及WMT-20和",
"翻译编辑率（TER）（Snoveretal.，2006）与使用",
"WMT-21进行了实验。带有MQM（多维质量度",
"相同LLM获得的零拍翻译，并将其与编辑后",
"量框架）错误Freitagetal.（2021）注释的新闻",
"翻译的TER与初始翻译进行比较。TER(T',T",
"翻译任务提交。对于与MQM注释的WMT-20",
"）的值越高，意味着编辑后的翻译（T）越接近",
"和WMT-21系统输出相关的后期编辑实验，我",
"初始翻译（T），并且LLM坚持编辑初始翻译",
"们使用具有主要错误的样本作为注释进行实",
"的任务",
"验，而我们在整个过程中使用完整的WMT-22",
"数据集进行实验。我们在大多数实验中使用最",
"PE Setting",
"TER(T',Z)TER(T',T)",
"新的WMT-22测试集，其管理超出了GPT-4和",
"With CoT",
"92.0",
"70.3",
"其他正在研究的LLM的训练截止日期2",
"Without CoT",
"84.6",
"94.5",
"LLMsandBaselines：我们用GPT-4和gpt-3",
"进行实验。5-turbo在我们的实验中。这些模型",
"Table 3: WMT-22 En-Zh : The post-edited translations",
"代表了最有能力的公开LLM(Liangetal.,2022)",
"CT)areclosertotheinitial translations（T)thanthe",
"zero-shot translations (Z)in the CoT setting.",
"。我们使用一个提示，将系统角色描述为翻译",
"后编辑器，并在CoT设置下，指示LLM在生",
"成最终的编辑后翻译（T）之前，对给定源（S",
"PE Setting",
"TER(T',Z)TER(T,T)",
")的提供的翻译（T）提出改进。对于后期编辑，",
"42.9",
"22.0",
"我们在三种设置下进行实验：（i）使用CoT进",
"With CoT",
"38.1",
"34.9",
"行后期编辑，（ii）不使用CoT进行后期编辑，",
"Without CoT",
"以及（ii）使用结构化CoT（SCoT）进行后期",
"编辑。SCoT基线使用Freitagetal.(2021)中的",
"Table 4: WMT-22 Zh-En : The post-edited translations",
"(T' ) are closer to the initial translations (T) than the",
"MQM注释指令在源-转换对上以MQM注释的",
"zero-shot translations (Z ) in the CoT setting.",
"形式生成中间CoT。我们在附录C中描述了用",
"于三个基线的提示。为了在WMT-22上生成初",
"ImpactofCoT：表3描述了我们在WMT",
"始翻译，我们使用Microsoft-Translator，这是最",
"22En-Zh上的结果，表4描述了我们在Zh-En",
"强的公开MT系统(Raunaketal.，2022;Hendy",
"上使用GPT-4进行后期编辑的结果。我们发现",
"et al., 2023)之一",
"一。对于 WMT-20 和 WMT-21",
"CoT限制了最终翻译更接近初始翻译。在没有",
"系统，我们将不同NMT系统提供并在Freitag",
"CoT的后期编辑设置中，即使TER差异比CoT",
"etal.（2021)中注释的翻译作为应用后期编辑",
"设置中的差异小得多，最终的翻译也更接近于",
"的初始翻译",
"零拍翻译",
"MetricsandEvaluation：对于提出的四个研",
"Discussion我们发现上述结果适用于不同的",
"究问题中的每一个，我们使用表2中突出显",
"指标，例如编辑距离，BLEU（Post，2018）或",
"Ihttps://github.com/google/",
"ChrF（Popovic，2015）以及WMT-22语言对和",
"wmt-mgm-human-evaluation",
"gpt-3。5-涡轮增压。此结果还显示了CoT设",
"2LLMs:",
"https://platform.openai.com/docs/",
"models",
"3COMET:https:/github.com/Unbabel/COMET"
] |
{
"arxiv_id": "2305.14878",
"language": "zh",
"page_idx": 3
}
|
||
2305.14878-zh-4
|
2305.14878-zh-4.jpg
|
置下的后期编辑任务的特殊副作用-如果初始翻译的质量低于所考虑的 LLM 的零翻译质量， 则后期编辑系统翻译可能最终导致最终翻译质量较低。在下一个小节中，我们将在一般质量改进方面评估不同编辑后设置下的 GPT-4
## 4.2 General Quality Improvements
我们使用无参考和基于参考的最新神经 MT 质量指标将编辑后翻译的翻译质量与初始翻译进行比较
Results: 表5，7,6和8提供了在 WMT-22测试集上进行的实验结果。在整个过程中，我们发现在 CoT 和直接设置下进行后期编辑可以改善通过 MS-Translator 获得的高质量初始翻译。此外，表5，7,6和8表明，使用GPT-4对 MS-Translator 输出进行直接后期编辑始终可以提高 WMT-22-Best 翻译系统的质量。我们还发现 gpt-3。5-turbo 的性能始终低于 GPT-4以及初始翻译的质量，这表明 GPT-4 的后期编辑效率有质的飞跃
4.3 Edits On Human Annotated Error Spans
我们使用 Freitag et al. (2021)提供的 MQM 注释的 WMT-22 系统输出，并测量编辑后的翻译是否修改了人工注释者注释的翻译错误范围。对于修改的每个主要 MQM 错误跨度，我们记录得分为 1, 否则得分为 0。如果在编辑后的翻译中修改了更多的错误跨度，则报告的最终分数称为“编辑错误跨度的效率”(E3S) 会更高。。E3S 度量报告为测试集的百分比。
结果：表9，10和11报告了从 WMT-22 输出的 14 种不同的 WMT-22NMT 系统输出在三种语言对上获得的结果：英语-德语，中文-英语和英语-俄语。我们发现 GPT-4 产生的 E3S 率高于 50%, 总体质量有相当大的提高 (通过COMET-KIWI 测量), 这意味着它能够消除翻译中存在的不良伪像(跨度)。我们在 WMT-20和 WMT-21MQM 注释的系统输出以及附录 A 中重复此实验，结果相似
4.4 Trustworthiness of the Proposed Edits
在实际设置中，编辑后任务中产生的编辑($E$ ) 可能有助于说明 LLM 在编辑后翻译中所做的更改。因此，建议编辑的保真度对于在基于LLM 的后期编辑过程中给予更多信任非常重要。因此，建议的编辑是否存在于最终改进的翻译中或被模型幻觉的问题具有重要的实际意义。WequanticthispropertyusingEditRealizationRate(ERR),whichmeasures:oftheproposededits(
$E$
)bytheLLMintheCoTpost-
editingsetting,howmanyoftheeditswereactuallyrealizedintheimpro
ERR Human Evaluation Protocol: 我们要求人类注释者(目标语言中的双语和母语)从WMT-22 测试集中标记 100个 En-De 和 De-En 的编辑后样本，由 gpt-3 生成。5-turbo 和 GPT- 4。注释器被要求确定是否所有建议的编辑都在最终翻译中实现 (全部),或者是否实现了部分建议的编辑 (部分),或者是否没有实现任何建议的编辑(无)。因此，人类注释器用三个标签之一标记每个编辑后样本$(S,T,E,T^{\prime})$
<FigureHere>
Figure 1: Edit Realization Rate (ERR) Human Evaluation on WMT-22 En-De and De-En: GPT-4 obtains higher ERR than gpt-3.5-turbo, imparting greater trustworthiness to the post-editing process under the CoT setting. In conjunction with this result, we also observe that GPT-4 obtains better ERR score distributions on En-Zh translation post-editing, vs Zh-En post-editing.
ERR Human Evaluation Results: 人体评估的结果如图 1 所示。一般来说，对于 En-De 和De-En, gpt-3 的 ERR 分布之间存在很大的差距。5-turbo 和 GPT-4。我们提供了一个典型的编辑后示例，说明了表 12 中的这种差异。我们的研究结果表明，GPT-4 为后期编辑任务产生了更多值得信赖的世代，并与第4小节中的结果相结合。1、4。2 和 4。3, 这表明 GPT-4可以帮助自动后期编辑，具有更大的可解释性。我们也观察到 En-Zh 的 ERR 同样高，尽管人类评估仍然报告了 GPT-4 编辑未完全实现的情况，特别是在源语言不是英语的语言对中 (Zh-En 和 De-En)。我们在附录 D 中提供了此类幻觉编辑建议的示例，表明 GPT-4 可能会像 NMT 在翻译中一样，在后期编辑中面临类似的可靠性挑战
### 5 $\textbf{Further Discussion}$
Post-Editing Across Language Pairs: 我们报告了 CoT 设置下的 GPT-4 后期编辑性能，其中 MS-Translator 作为附录 B 中其他几种语言
|
[
[
0.12176513671875,
0.0885009765625,
0.48779296875,
0.0885009765625,
0.48779296875,
0.09857177734375,
0.12176513671875,
0.09857177734375
],
[
0.496337890625,
0.08905029296875,
0.51318359375,
0.08905029296875,
0.51318359375,
0.0968017578125,
0.496337890625,
0.0968017578125
],
[
0.51318359375,
0.08905029296875,
0.5283203125,
0.08905029296875,
0.5283203125,
0.097412109375,
0.51318359375,
0.097412109375
],
[
0.68994140625,
0.08489990234375,
0.87744140625,
0.08612060546875,
0.87744140625,
0.1009521484375,
0.68994140625,
0.09979248046875
],
[
0.486083984375,
0.0914306640625,
0.497802734375,
0.0914306640625,
0.497802734375,
0.0955810546875,
0.486083984375,
0.0955810546875
],
[
0.118408203125,
0.10272216796875,
0.489501953125,
0.10272216796875,
0.489501953125,
0.11639404296875,
0.118408203125,
0.11639404296875
],
[
0.51123046875,
0.10272216796875,
0.99658203125,
0.10272216796875,
0.99658203125,
0.11639404296875,
0.51123046875,
0.11639404296875
],
[
0.12005615234375,
0.1181640625,
0.484375,
0.1181640625,
0.484375,
0.1318359375,
0.12005615234375,
0.1318359375
],
[
0.51708984375,
0.12890625,
0.880859375,
0.12890625,
0.880859375,
0.142578125,
0.51708984375,
0.142578125
],
[
0.12005615234375,
0.1353759765625,
0.484375,
0.1353759765625,
0.484375,
0.1484375,
0.12005615234375,
0.1484375
],
[
0.1175537109375,
0.15087890625,
0.447509765625,
0.149658203125,
0.447509765625,
0.163330078125,
0.1175537109375,
0.1644287109375
],
[
0.515625,
0.1455078125,
0.87744140625,
0.1455078125,
0.87744140625,
0.1561279296875,
0.515625,
0.1561279296875
],
[
0.5146484375,
0.162109375,
0.87890625,
0.162109375,
0.87890625,
0.1728515625,
0.5146484375,
0.1728515625
],
[
0.1175537109375,
0.1817626953125,
0.41064453125,
0.1817626953125,
0.41064453125,
0.1953125,
0.1175537109375,
0.1953125
],
[
0.51318359375,
0.1776123046875,
0.8818359375,
0.1776123046875,
0.8818359375,
0.191162109375,
0.51318359375,
0.191162109375
],
[
0.51318359375,
0.1953125,
0.87890625,
0.1953125,
0.87890625,
0.2060546875,
0.51318359375,
0.2060546875
],
[
0.118408203125,
0.2037353515625,
0.486083984375,
0.2037353515625,
0.486083984375,
0.21728515625,
0.118408203125,
0.21728515625
],
[
0.51123046875,
0.210205078125,
0.87890625,
0.210205078125,
0.87890625,
0.223876953125,
0.51123046875,
0.223876953125
],
[
0.12005615234375,
0.2203369140625,
0.48779296875,
0.2203369140625,
0.48779296875,
0.2340087890625,
0.12005615234375,
0.2340087890625
],
[
0.51318359375,
0.2257080078125,
0.880859375,
0.2257080078125,
0.880859375,
0.2392578125,
0.51318359375,
0.2392578125
],
[
0.1175537109375,
0.2369384765625,
0.175537109375,
0.2369384765625,
0.175537109375,
0.251708984375,
0.1175537109375,
0.251708984375
],
[
0.51318359375,
0.2423095703125,
0.880859375,
0.2423095703125,
0.880859375,
0.255859375,
0.51318359375,
0.255859375
],
[
0.51318359375,
0.25830078125,
0.8818359375,
0.25830078125,
0.8818359375,
0.271484375,
0.51318359375,
0.271484375
],
[
0.12176513671875,
0.265380859375,
0.486083984375,
0.265380859375,
0.486083984375,
0.279052734375,
0.12176513671875,
0.279052734375
],
[
0.118408203125,
0.281982421875,
0.483642578125,
0.281982421875,
0.483642578125,
0.292236328125,
0.118408203125,
0.292236328125
],
[
0.5859375,
0.2861328125,
0.8388671875,
0.2861328125,
0.8388671875,
0.2998046875,
0.5859375,
0.2998046875
],
[
0.118408203125,
0.297607421875,
0.484375,
0.297607421875,
0.484375,
0.311279296875,
0.118408203125,
0.311279296875
],
[
0.52880859375,
0.29345703125,
0.5458984375,
0.29345703125,
0.5458984375,
0.3017578125,
0.52880859375,
0.3017578125
],
[
0.587890625,
0.2998046875,
0.60791015625,
0.2998046875,
0.60791015625,
0.312255859375,
0.587890625,
0.312255859375
],
[
0.1175537109375,
0.312255859375,
0.48779296875,
0.3134765625,
0.48779296875,
0.327880859375,
0.1175537109375,
0.32666015625
],
[
0.587890625,
0.30810546875,
0.63232421875,
0.31005859375,
0.63037109375,
0.32470703125,
0.5859375,
0.322509765625
],
[
0.52490234375,
0.319580078125,
0.53076171875,
0.32373046875,
0.52490234375,
0.327880859375,
0.51904296875,
0.32373046875
],
[
0.56494140625,
0.319580078125,
0.62451171875,
0.319580078125,
0.62451171875,
0.334228515625,
0.56494140625,
0.334228515625
],
[
0.118408203125,
0.330078125,
0.486083984375,
0.330078125,
0.486083984375,
0.34326171875,
0.118408203125,
0.34326171875
],
[
0.118408203125,
0.3466796875,
0.484375,
0.3466796875,
0.484375,
0.35986328125,
0.118408203125,
0.35986328125
],
[
0.5205078125,
0.345703125,
0.53076171875,
0.345703125,
0.53076171875,
0.352783203125,
0.5205078125,
0.352783203125
],
[
0.5341796875,
0.3486328125,
0.54248046875,
0.3486328125,
0.54248046875,
0.352783203125,
0.5341796875,
0.352783203125
],
[
0.1175537109375,
0.361083984375,
0.483642578125,
0.361083984375,
0.483642578125,
0.374755859375,
0.1175537109375,
0.374755859375
],
[
0.118408203125,
0.377685546875,
0.486083984375,
0.377685546875,
0.486083984375,
0.391357421875,
0.118408203125,
0.391357421875
],
[
0.5205078125,
0.37353515625,
0.52880859375,
0.37353515625,
0.52880859375,
0.377685546875,
0.5205078125,
0.377685546875
],
[
0.118408203125,
0.394287109375,
0.484375,
0.394287109375,
0.484375,
0.407958984375,
0.118408203125,
0.407958984375
],
[
0.118408203125,
0.412109375,
0.26611328125,
0.412109375,
0.26611328125,
0.4228515625,
0.118408203125,
0.4228515625
],
[
0.568359375,
0.42578125,
0.8212890625,
0.4287109375,
0.8212890625,
0.45068359375,
0.568359375,
0.44775390625
],
[
0.8095703125,
0.427490234375,
0.8671875,
0.42578125,
0.8671875,
0.435791015625,
0.8095703125,
0.438232421875
],
[
0.1175537109375,
0.441162109375,
0.486083984375,
0.441162109375,
0.486083984375,
0.454833984375,
0.1175537109375,
0.454833984375
],
[
0.1175537109375,
0.463134765625,
0.48779296875,
0.4619140625,
0.48779296875,
0.476806640625,
0.1175537109375,
0.47802734375
],
[
0.51123046875,
0.458984375,
0.87890625,
0.458984375,
0.87890625,
0.47265625,
0.51123046875,
0.47265625
],
[
0.51318359375,
0.473876953125,
0.880859375,
0.473876953125,
0.880859375,
0.48681640625,
0.51318359375,
0.48681640625
],
[
0.1175537109375,
0.479736328125,
0.484375,
0.479736328125,
0.484375,
0.493408203125,
0.1175537109375,
0.493408203125
],
[
0.51123046875,
0.488037109375,
0.880859375,
0.488037109375,
0.880859375,
0.501953125,
0.51123046875,
0.501953125
],
[
0.118408203125,
0.496337890625,
0.484375,
0.496337890625,
0.484375,
0.51025390625,
0.118408203125,
0.51025390625
],
[
0.51123046875,
0.501953125,
0.880859375,
0.501953125,
0.880859375,
0.515625,
0.51123046875,
0.515625
],
[
0.118408203125,
0.5126953125,
0.486083984375,
0.5126953125,
0.486083984375,
0.525390625,
0.118408203125,
0.525390625
],
[
0.51318359375,
0.51708984375,
0.8759765625,
0.51708984375,
0.8759765625,
0.52783203125,
0.51318359375,
0.52783203125
],
[
0.12005615234375,
0.52978515625,
0.486083984375,
0.52978515625,
0.486083984375,
0.54052734375,
0.12005615234375,
0.54052734375
],
[
0.51123046875,
0.52978515625,
0.880859375,
0.52978515625,
0.880859375,
0.54345703125,
0.51123046875,
0.54345703125
],
[
0.118408203125,
0.54443359375,
0.484375,
0.54443359375,
0.484375,
0.55810546875,
0.118408203125,
0.55810546875
],
[
0.509765625,
0.54345703125,
0.875,
0.5458984375,
0.875,
0.55859375,
0.509765625,
0.55712890625
],
[
0.1175537109375,
0.55859375,
0.484375,
0.55859375,
0.484375,
0.572265625,
0.1175537109375,
0.572265625
],
[
0.118408203125,
0.57666015625,
0.46337890625,
0.57666015625,
0.46337890625,
0.59033203125,
0.118408203125,
0.59033203125
],
[
0.51904296875,
0.576171875,
0.880859375,
0.576171875,
0.880859375,
0.5888671875,
0.51904296875,
0.5888671875
],
[
0.5146484375,
0.5927734375,
0.880859375,
0.5927734375,
0.880859375,
0.60546875,
0.5146484375,
0.60546875
],
[
0.1351318359375,
0.60986328125,
0.486083984375,
0.60986328125,
0.486083984375,
0.62353515625,
0.1351318359375,
0.62353515625
],
[
0.51318359375,
0.60791015625,
0.880859375,
0.60791015625,
0.880859375,
0.62255859375,
0.51318359375,
0.62255859375
],
[
0.118408203125,
0.6259765625,
0.486083984375,
0.6259765625,
0.486083984375,
0.6396484375,
0.118408203125,
0.6396484375
],
[
0.51318359375,
0.6259765625,
0.87890625,
0.6259765625,
0.87890625,
0.6357421875,
0.51318359375,
0.6357421875
],
[
0.1175537109375,
0.64111328125,
0.484375,
0.64111328125,
0.484375,
0.65478515625,
0.1175537109375,
0.65478515625
],
[
0.51318359375,
0.64013671875,
0.8818359375,
0.64013671875,
0.8818359375,
0.65380859375,
0.51318359375,
0.65380859375
],
[
0.118408203125,
0.6591796875,
0.486083984375,
0.6591796875,
0.486083984375,
0.6728515625,
0.118408203125,
0.6728515625
],
[
0.51318359375,
0.65625,
0.880859375,
0.65625,
0.880859375,
0.66943359375,
0.51318359375,
0.66943359375
],
[
0.118408203125,
0.6748046875,
0.486083984375,
0.6748046875,
0.486083984375,
0.6884765625,
0.118408203125,
0.6884765625
],
[
0.51318359375,
0.6728515625,
0.880859375,
0.6728515625,
0.880859375,
0.68603515625,
0.51318359375,
0.68603515625
],
[
0.1175537109375,
0.6904296875,
0.486083984375,
0.6904296875,
0.486083984375,
0.70361328125,
0.1175537109375,
0.70361328125
],
[
0.51123046875,
0.6884765625,
0.8818359375,
0.6884765625,
0.8818359375,
0.70166015625,
0.51123046875,
0.70166015625
],
[
0.118408203125,
0.70703125,
0.48779296875,
0.70703125,
0.48779296875,
0.72021484375,
0.118408203125,
0.72021484375
],
[
0.51318359375,
0.703125,
0.880859375,
0.703125,
0.880859375,
0.71630859375,
0.51318359375,
0.71630859375
],
[
0.118408203125,
0.72265625,
0.48779296875,
0.72265625,
0.48779296875,
0.736328125,
0.118408203125,
0.736328125
],
[
0.51318359375,
0.72021484375,
0.880859375,
0.72021484375,
0.880859375,
0.73388671875,
0.51318359375,
0.73388671875
],
[
0.1175537109375,
0.73681640625,
0.32080078125,
0.73828125,
0.32080078125,
0.7529296875,
0.1175537109375,
0.751953125
],
[
0.5146484375,
0.736328125,
0.880859375,
0.736328125,
0.880859375,
0.75,
0.5146484375,
0.75
],
[
0.51318359375,
0.7529296875,
0.880859375,
0.7529296875,
0.880859375,
0.7666015625,
0.51318359375,
0.7666015625
],
[
0.1175537109375,
0.76953125,
0.46337890625,
0.76953125,
0.46337890625,
0.783203125,
0.1175537109375,
0.783203125
],
[
0.5146484375,
0.77099609375,
0.8759765625,
0.77099609375,
0.8759765625,
0.78076171875,
0.5146484375,
0.78076171875
],
[
0.51123046875,
0.78369140625,
0.880859375,
0.78515625,
0.880859375,
0.7998046875,
0.51123046875,
0.798828125
],
[
0.118408203125,
0.7919921875,
0.48779296875,
0.7919921875,
0.48779296875,
0.8056640625,
0.118408203125,
0.8056640625
],
[
0.51123046875,
0.7998046875,
0.880859375,
0.80029296875,
0.880859375,
0.8154296875,
0.51123046875,
0.81396484375
],
[
0.12005615234375,
0.81005859375,
0.486083984375,
0.81005859375,
0.486083984375,
0.82080078125,
0.12005615234375,
0.82080078125
],
[
0.51123046875,
0.81787109375,
0.64404296875,
0.81787109375,
0.64404296875,
0.83056640625,
0.51123046875,
0.83056640625
],
[
0.12005615234375,
0.82666015625,
0.484375,
0.82666015625,
0.484375,
0.83740234375,
0.12005615234375,
0.83740234375
],
[
0.1175537109375,
0.84130859375,
0.48779296875,
0.84130859375,
0.48779296875,
0.85498046875,
0.1175537109375,
0.85498046875
],
[
0.50732421875,
0.84423828125,
0.7119140625,
0.845703125,
0.7119140625,
0.85986328125,
0.50732421875,
0.859375
],
[
0.118408203125,
0.85693359375,
0.484375,
0.85693359375,
0.484375,
0.87060546875,
0.118408203125,
0.87060546875
],
[
0.118408203125,
0.87255859375,
0.486083984375,
0.87255859375,
0.486083984375,
0.8857421875,
0.118408203125,
0.8857421875
],
[
0.515625,
0.87158203125,
0.880859375,
0.87255859375,
0.880859375,
0.88720703125,
0.515625,
0.8857421875
],
[
0.12005615234375,
0.8896484375,
0.48193359375,
0.8896484375,
0.48193359375,
0.90283203125,
0.12005615234375,
0.90283203125
],
[
0.5146484375,
0.8896484375,
0.880859375,
0.8896484375,
0.880859375,
0.90283203125,
0.5146484375,
0.90283203125
],
[
0.1175537109375,
0.90380859375,
0.47021484375,
0.90478515625,
0.47021484375,
0.91943359375,
0.1175537109375,
0.91845703125
],
[
0.51318359375,
0.90478515625,
0.880859375,
0.90478515625,
0.880859375,
0.91845703125,
0.51318359375,
0.91845703125
]
] |
[
"置下的后期编辑任务的特殊副作用-如果初始",
"",
"E",
")bytheLLMintheCoTpost",
"",
"翻译的质量低于所考虑的LLM的零翻译质量，",
"editingsetting,howmanyoftheeditswereactuallyrealizedintheimproy",
"则后期编辑系统翻译可能最终导致最终翻译质",
"ERRHumanEvaluationProtocol：我们要求",
"量较低。在下一个小节中，我们将在一般质量",
"改进方面评估不同编辑后设置下的GPT-4",
"人类注释者（目标语言中的双语和母语）从",
"WMT-22测试集中标记100个En-De和De-En",
"4.2General Quality Improvements",
"的编辑后样本，由gpt-3生成。5-turbo和GPT",
"4。注释器被要求确定是否所有建议的编辑都",
"我们使用无参考和基于参考的最新神经MT质",
"在最终翻译中实现（全部），或者是否实现了部",
"量指标将编辑后翻译的翻译质量与初始翻译进",
"分建议的编辑（部分)，或者是否没有实现任何",
"行比较",
"建议的编辑(无)。因此，人类注释器用三个标",
"签之一标记每个编辑后样本（S，T，E，T）",
"Results：表5，7，6和8提供了在WMT-22",
"测试集上进行的实验结果。在整个过程中，我",
"Edit Realization Rate(ERR)Comparison",
"们发现在CoT和直接设置下进行后期编辑可",
"100",
"ALL",
"以改善通过MS-Translator获得的高质量初始",
"PARTIAL",
"L",
"NONE",
"翻译。此外，表5，7，6和8表明，使用GPT-4",
"对MS-Translator输出进行直接后期编辑始终",
"0",
"b",
"可以提高WMT-22-Best翻译系统的质量。我们",
"还发现gpt-3。5-turbo的性能始终低于GPT-4",
"I.",
"以及初始翻译的质量，这表明GPT-4的后期编",
"辑效率有质的飞跃",
"ENio",
"DE-EN,GPT-4",
"4.3Edits On Human Annotated Error Spans",
"我们使用Freitagetal.(2021）提供的MQM注",
"Figure 1: Edit Realization Rate (ERR) Human Evalu",
"ation on WMT-22 En-De and De-En : GPT-4 obtains",
"释的WMT-22系统输出，并测量编辑后的翻",
"higher ERR than gpt-3.5-turbo, imparting greater trust",
"译是否修改了人工注释者注释的翻译错误范",
"worthiness to the post-editing process under the CoT",
"围。对于修改的每个主要MQM错误跨度，我",
"setting. In conjunction with this result, we also observe",
"们记录得分为1，否则得分为0。如果在编辑",
"that GPT-4 obtains better ERR score distributions on",
"后的翻译中修改了更多的错误跨度，则报告的",
"En-Zh translation post-editing, vs Zh-En post-editing.",
"最终分数称为“编辑错误跨度的效率”（E3S）",
"会更高。。E3S度量报告为测试集的百分比。",
"ERRHumanEvaluationResults：人体评估",
"的结果如图1所示。一般来说，对于En-De和",
"结果：表9，10和11报告了从WMT-22输",
"De-En，gpt-3的ERR分布之间存在很大的差",
"出的14种不同的WMT-22NMT系统输出在三",
"距。5-turbo和GPT-4。我们提供了一个典型的",
"种语言对上获得的结果：英语-德语，中文-英",
"编辑后示例，说明了表12中的这种差异。我",
"语和英语-俄语。我们发现GPT-4产生的E3S",
"们的研究结果表明，GPT-4为后期编辑任务产",
"率高于50%，总体质量有相当大的提高（通过",
"生了更多值得信赖的世代，并与第4小节中的",
"COMET-KIWI测量），这意味着它能够消除翻",
"结果相结合。1、4。2和4。3，这表明GPT-4",
"译中存在的不良伪像（跨度）。我们在WMT-20",
"可以帮助自动后期编辑，具有更大的可解释",
"和WMT-21MQM注释的系统输出以及附录A",
"性。我们也观察到En-Zh的ERR同样高，尽",
"中重复此实验，结果相似",
"管人类评估仍然报告了GPT-4编辑未完全实",
"现的情况，特别是在源语言不是英语的语言对",
"4.4Trustworthiness of the Proposed Edits",
"中（Zh-En和De-En）。我们在附录D中提供了",
"此类幻觉编辑建议的示例，表明GPT-4可能会",
"在实际设置中，编辑后任务中产生的编辑（E",
"像NMT在翻译中一样，在后期编辑中面临类",
"）可能有助于说明LLM在编辑后翻译中所做",
"似的可靠性挑战",
"的更改。因此，建议编辑的保真度对于在基于",
"LLM的后期编辑过程中给予更多信任非常重",
"5 FurtherDiscussion",
"要。因此，建议的编辑是否存在于最终改进的",
"翻译中或被模型幻觉的问题具有重要的实际意",
"Post-EditingAcrossLanguagePairs：我们报",
"义。WequanticthispropertyusingEditRealization-",
"告了CoT设置下的GPT-4后期编辑性能，其",
"Rate(ERR),whichmeasures:oftheproposededits(",
"中MS-Translator作为附录B中其他几种语言"
] |
{
"arxiv_id": "2305.14878",
"language": "zh",
"page_idx": 4
}
|
||
2011.01143-zh-0
|
2011.01143-zh-0.jpg
|
Published as a conference paper at ICLR 2021
使用 AUDIOSCOPE 进入野外：屏幕上声音的无监督视
听分离
$\textbf{Efthymios Tzinis}^{*1},\textbf{Scott Wisdom}^2,\textbf{Aren Jansen}^2,\textbf{Shawn Hershey}^2,\textbf{Tal Remez}^2$,
$DanielP.W.Ellis^2,JohnR.Hershey^2$
$^{1}$University of Illinois at Urbana-Champaign$^2$Google Research
etzinis2@illinois.edu
scottwisdomegoogle.com
## ABSTRACT
Recent progress in deep learning has enabled many advances in sound separation and visual scene understanding. However, extracting sound sources which are apparent in natural videos remains an open problem. In this work, we present AudioScope , a novel audio-visual sound separation framework that can be trained without supervision to isolate on-screen sound sources from real in-the-wild videos. Prior audio-visual separation work assumed artificial limitations on the domain of sound classes (e.g., to speech or music), constrained the number of sources, and required strong sound separation or visual segmentation labels. AudioScope overcomes these limitations, operating on an open domain of sounds, with variable numbers of sources, and without labels or prior visual segmentation. The training procedure for AudioScope uses mixture invariant training (MixIT) to separate synthetic mixtures of mixtures (MoMs) into individual sources, where noisy labels for mixtures are provided by an unsupervised audio-visual coincidence model. Using the noisy labels, along with attention between video and audio features, AudioScope learns to identify audio-visual similarity and to suppress off-screen sounds. We demonstrate the effectiveness of our approach using a dataset of video clips extracted from open-domain YFCC100m video data. This dataset contains a wide diversity of sound classes recorded in unconstrained conditions, making the application of previous methods unsuitable. For evaluation and semi-supervised experiments, we collected human labels for presence of onscreen and off-screen sounds on a small subset of clips.
1
INTRODUCTION
<FigureHere>
Figure 1: AudioScope separating on-screen bird chirping from wind noise and off-screen sounds from fireworks and human laugh. More demos online at https://audioscope.github.io.
近年来，在大规模深度学习的推动下，视听机器感知正在经历复兴。激励观察是人类感知中听觉和视觉感知之间的相互作用。我们通过将世界解析为我们可以感知的音频和视频信号的来源来理解世界。然而，这些来源产生的声音和景象具有相当不同和互补的特性。对象可能会间歇性地发出声音，而它们的视觉外观通常是持久的。不同物体的视觉感知往往在空间上是不同的，而来自不同来源的声音可以混合在一起并在单个信号中重叠，因此很难分别感知各个来源。这表明通过调整我们的音频和视觉感知可以获得一些东西：如果我们可以识别哪些音频信号对应于哪些视觉对象，我们可以通过视觉选择对象来选择性地关注对象的音频信号
$^{*}$Work done during an internship at Google.
1
|
[
[
0.1732177734375,
0.0328369140625,
0.476318359375,
0.0328369140625,
0.476318359375,
0.046722412109375,
0.1732177734375,
0.046722412109375
],
[
0.178955078125,
0.10479736328125,
0.8203125,
0.10479736328125,
0.8203125,
0.12249755859375,
0.178955078125,
0.12249755859375
],
[
0.1748046875,
0.1275634765625,
0.261474609375,
0.1275634765625,
0.261474609375,
0.15087890625,
0.1748046875,
0.15087890625
],
[
0.1846923828125,
0.1767578125,
0.7509765625,
0.1767578125,
0.7509765625,
0.1912841796875,
0.1846923828125,
0.1912841796875
],
[
0.1854248046875,
0.1925048828125,
0.44287109375,
0.1925048828125,
0.44287109375,
0.205810546875,
0.1854248046875,
0.205810546875
],
[
0.1846923828125,
0.2052001953125,
0.619140625,
0.203857421875,
0.619140625,
0.2215576171875,
0.1846923828125,
0.222900390625
],
[
0.1871337890625,
0.222900390625,
0.39208984375,
0.222900390625,
0.39208984375,
0.23291015625,
0.1871337890625,
0.23291015625
],
[
0.495849609375,
0.2215576171875,
0.71337890625,
0.2215576171875,
0.71337890625,
0.23486328125,
0.495849609375,
0.23486328125
],
[
0.453369140625,
0.271484375,
0.5458984375,
0.271484375,
0.5458984375,
0.285888671875,
0.453369140625,
0.285888671875
],
[
0.233642578125,
0.3017578125,
0.765625,
0.3017578125,
0.765625,
0.31494140625,
0.233642578125,
0.31494140625
],
[
0.2344970703125,
0.31640625,
0.76318359375,
0.31640625,
0.76318359375,
0.326904296875,
0.2344970703125,
0.326904296875
],
[
0.236083984375,
0.330810546875,
0.765625,
0.330810546875,
0.765625,
0.341552734375,
0.236083984375,
0.341552734375
],
[
0.233642578125,
0.341552734375,
0.765625,
0.341552734375,
0.765625,
0.354736328125,
0.233642578125,
0.354736328125
],
[
0.236083984375,
0.35791015625,
0.76123046875,
0.35791015625,
0.76123046875,
0.36865234375,
0.236083984375,
0.36865234375
],
[
0.2344970703125,
0.37060546875,
0.765625,
0.37060546875,
0.765625,
0.3837890625,
0.2344970703125,
0.3837890625
],
[
0.2344970703125,
0.385009765625,
0.7646484375,
0.385009765625,
0.7646484375,
0.3984375,
0.2344970703125,
0.3984375
],
[
0.233642578125,
0.396484375,
0.7646484375,
0.397705078125,
0.7646484375,
0.412353515625,
0.233642578125,
0.410888671875
],
[
0.2344970703125,
0.41357421875,
0.765625,
0.412353515625,
0.765625,
0.423583984375,
0.2344970703125,
0.4248046875
],
[
0.233642578125,
0.425537109375,
0.765625,
0.425537109375,
0.765625,
0.439453125,
0.233642578125,
0.439453125
],
[
0.2320556640625,
0.439453125,
0.765625,
0.438232421875,
0.765625,
0.45263671875,
0.2320556640625,
0.453857421875
],
[
0.2344970703125,
0.453857421875,
0.765625,
0.453857421875,
0.765625,
0.46728515625,
0.2344970703125,
0.46728515625
],
[
0.233642578125,
0.46728515625,
0.76708984375,
0.46728515625,
0.76708984375,
0.48095703125,
0.233642578125,
0.48095703125
],
[
0.2344970703125,
0.48291015625,
0.7646484375,
0.48291015625,
0.7646484375,
0.49365234375,
0.2344970703125,
0.49365234375
],
[
0.236083984375,
0.496826171875,
0.76318359375,
0.496826171875,
0.76318359375,
0.5068359375,
0.236083984375,
0.5068359375
],
[
0.233642578125,
0.5087890625,
0.7646484375,
0.5087890625,
0.7646484375,
0.52294921875,
0.233642578125,
0.52294921875
],
[
0.233642578125,
0.52294921875,
0.7646484375,
0.52294921875,
0.7646484375,
0.5361328125,
0.233642578125,
0.5361328125
],
[
0.2344970703125,
0.53857421875,
0.7646484375,
0.53857421875,
0.7646484375,
0.548828125,
0.2344970703125,
0.548828125
],
[
0.233642578125,
0.55029296875,
0.765625,
0.55029296875,
0.765625,
0.564453125,
0.233642578125,
0.564453125
],
[
0.2344970703125,
0.564453125,
0.765625,
0.564453125,
0.765625,
0.57763671875,
0.2344970703125,
0.57763671875
],
[
0.233642578125,
0.57763671875,
0.5947265625,
0.5791015625,
0.5947265625,
0.59228515625,
0.233642578125,
0.59130859375
],
[
0.17724609375,
0.61865234375,
0.1895751953125,
0.61865234375,
0.1895751953125,
0.626953125,
0.17724609375,
0.626953125
],
[
0.2059326171875,
0.61572265625,
0.34228515625,
0.6162109375,
0.34228515625,
0.630859375,
0.2059326171875,
0.6298828125
],
[
0.2115478515625,
0.65087890625,
0.2607421875,
0.65087890625,
0.2607421875,
0.65771484375,
0.2115478515625,
0.65771484375
],
[
0.3095703125,
0.65283203125,
0.328369140625,
0.65283203125,
0.328369140625,
0.65576171875,
0.3095703125,
0.65576171875
],
[
0.351318359375,
0.65283203125,
0.378173828125,
0.65283203125,
0.378173828125,
0.65576171875,
0.351318359375,
0.65576171875
],
[
0.4150390625,
0.65283203125,
0.484375,
0.65283203125,
0.484375,
0.65576171875,
0.4150390625,
0.65576171875
],
[
0.55224609375,
0.65283203125,
0.5595703125,
0.65283203125,
0.5595703125,
0.65576171875,
0.55224609375,
0.65576171875
],
[
0.57421875,
0.65283203125,
0.5810546875,
0.65283203125,
0.5810546875,
0.65576171875,
0.57421875,
0.65576171875
],
[
0.176513671875,
0.751953125,
0.82421875,
0.751953125,
0.82421875,
0.76513671875,
0.176513671875,
0.76513671875
],
[
0.1732177734375,
0.76513671875,
0.82177734375,
0.76513671875,
0.82177734375,
0.77880859375,
0.1732177734375,
0.77880859375
],
[
0.176513671875,
0.79345703125,
0.82275390625,
0.79345703125,
0.82275390625,
0.806640625,
0.176513671875,
0.806640625
],
[
0.1748046875,
0.80810546875,
0.8203125,
0.80810546875,
0.8203125,
0.8212890625,
0.1748046875,
0.8212890625
],
[
0.176513671875,
0.8212890625,
0.82177734375,
0.8212890625,
0.82177734375,
0.83544921875,
0.176513671875,
0.83544921875
],
[
0.17724609375,
0.83544921875,
0.82275390625,
0.83544921875,
0.82275390625,
0.8486328125,
0.17724609375,
0.8486328125
],
[
0.17724609375,
0.849609375,
0.82177734375,
0.849609375,
0.82177734375,
0.86279296875,
0.17724609375,
0.86279296875
],
[
0.17724609375,
0.86279296875,
0.82177734375,
0.86279296875,
0.82177734375,
0.876953125,
0.17724609375,
0.876953125
],
[
0.176513671875,
0.876953125,
0.82275390625,
0.876953125,
0.82275390625,
0.89013671875,
0.176513671875,
0.89013671875
],
[
0.17724609375,
0.892578125,
0.27294921875,
0.892578125,
0.27294921875,
0.90380859375,
0.17724609375,
0.90380859375
],
[
0.1953125,
0.91015625,
0.451904296875,
0.9111328125,
0.451904296875,
0.92529296875,
0.1953125,
0.9248046875
],
[
0.49267578125,
0.94970703125,
0.5048828125,
0.94970703125,
0.5048828125,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2021",
"使用AUDIOSCOPE进入野外：屏幕上声音的无监督视",
"听分离",
"Efthymios Tzinis*1, Scott Wisdom?, Aren Jansen?, Shawn Hershey?, Tal Remez?",
"Daniel P.W. Ellis?, John R. Hershey?",
"IUniversity of Illinois at Urbana-Champaign 2Google Research",
"etzinis2@illinois.edu",
"scottwisdom@google.com",
"ABSTRACT",
"Recent progress in deep learning has enabled many advances in sound separation",
"and visual scene understanding.However, extracting sound sources which are ap",
"parentinnaturalvideosremainsanopenproblem.Inthiswork,wepresentAudio-",
"Scope , a novel audio-visual sound separation framework that can be trained with-",
"outsupervision toisolate on-screen sound sourcesfromreal in-the-wild videos",
"Prior audio-visual separation work assumed artificial limitations on the domain",
"of sound classes (e.g., to speech or music), constrained the number of sources,",
"and required strong sound separation or visual segmentation labels. AudioScope",
"overcomes these limitations,operating on an open domain of sounds,with vari-",
"able numbers of sources, and without labels or prior visual segmentation. The",
"training procedure for AudioScope uses mixture invariant training (MixIT) to",
"separate synthetic mixtures of mixtures (MoMs) into individual sources, where",
"noisy labels for mixtures are provided by an unsupervised audio-visual coinci-",
"dencemodel.Using the noisylabels,along with attention between video and",
"audiofeatures,AudioScopelearnstoidentifyaudio-visualsimilarityandtosup",
"press off-screen sounds. We demonstrate the effectiveness of our approach using",
"a dataset of video clips extracted from open-domain YFCC10Om video data. This",
"datasetcontainsawidediversityof soundclassesrecordedinunconstrainedcon",
"ditions, making the application of previous methods unsuitable. For evaluation",
"and semi-supervised experiments, we collected human labels for presence of on-",
"screen and off-screen sounds ona small subsetof clips.",
"1",
"INTRODUCTION",
"Videofram",
"n-c",
"a",
"tccreenain",
"",
"1",
"Figure l: AudioScope separating on-screen bird chirping from wind noise and off-screen sounds",
"from fireworks and human laugh. More demos online at https : //audioscope.github.io.",
"近年来，在大规模深度学习的推动下，视听机器感知正在经历复兴。激励观察是人类感知中",
"听觉和视觉感知之间的相互作用。我们通过将世界解析为我们可以感知的音频和视频信号",
"的来源来理解世界。然而，这些来源产生的声音和景象具有相当不同和互补的特性。对象可",
"能会间歇性地发出声音，而它们的视觉外观通常是持久的。不同物体的视觉感知往往在空",
"间上是不同的，而来自不同来源的声音可以混合在一起并在单个信号中重叠，因此很难分",
"别感知各个来源。这表明通过调整我们的音频和视觉感知可以获得一些东西：如果我们可",
"以识别哪些音频信号对应于哪些视觉对象，我们可以通过视觉选择对象来选择性地关注对",
"象的音频信号",
"*Work done during an internship at Google",
"1"
] |
{
"arxiv_id": "2011.01143",
"language": "zh",
"page_idx": 0
}
|
||
2011.01143-zh-1
|
2011.01143-zh-1.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2021}}$
这种直觉促使使用视觉作为音频处理的接口，其中一个主要问题是选择性地保留所需的声音，同时去除不需要的声音。在某些任务中，例如语音增强，可以按其类别选择所需的声音： 在这种情况下，语音与非语音。在开放域设置中，所需声音的选择由用户自行决定。这带来了一个用户界面问题：使用音频以有效的方式选择源具有挑战性。如果我们使用视频选择作为音频选择的代理，例如，通过从屏幕上的对象中选择声音并删除屏幕外的声音，则在视听情况下可以大大简化此问题。最近的工作使用视频选择和分离语音(Ephrat et al., 2018; Afouras et al., 2020) 或音乐 (Zhao et al., 2018; Gao & Grauman, 2019; Gan et al., 2020)。但是， 针对任意声音(Gao et al., 2018; Rouditchenko et al., 2019; Owens & Efros, 2018) 解决此问题的系统可能在更一般的情况下有用，例如视频录制，其中无法预先定义感兴趣的声音
将任意声音与其视觉对象相关联的问题在开放域中具有挑战性。出现了一些以前的工作尚末完全解决的复杂性。首先，为了覆盖可能的声音空间，需要大量的训练数据。受监督的方法需要标记的示例，其中已知孤立的屏幕声音。由此产生的数据收集和标签负担限制了可用数据的数量和质量。为了克服这个问题，我们提出了一种使用混合不变训练(MixIT) (Wisdom et al,2020)的无监督方法，该方法可以学习将单个源与野外视频分开，其中屏幕上和屏幕外的声音是未知的。另一个问题是，不同的音频源可能对应于任意空间位置的一组动态屏幕上对象。我们通过使用注意机制来适应这一点，该机制将每个假设的音频源与视频中相应对象的不同空间和时间位置对齐。最后，我们需要确定哪些音频源出现在屏幕上在没有强标签的情况下。这是使用基于注意力机制产生的音频和视频嵌人的源的弱训练分类器来处理的
2 RELATION TO PREVIOUS WORK
从混合物中分离任意声音，称为“通用声音分离”,最近被证明可以使用固定数量的声音(Kavalerov et al., 2019)。有关存在哪些声音类别的条件信息可以提高分离性能(Tzinis et al. 2020)。FUSS 数据集 (Wisdom et al.,2021)扩展了范围，以分离可变数量的声音，以处理更真实的数据。还提出了一个框架，可以从输人声音混合物(Ochiai et al., 2020) 中提取特定的声音类别。这些方法需要包含孤立声音的精选数据进行训练，这阻止了它们应用于真正的开放域数据，并引人了诸如注释成本，真实声学混合的准确模拟以及有偏差的数据集之类的困难。
为了避免这些问题，最近的许多工作都建议用来自相关模态的弱监督标签代替对参考源信号的强监督，例如声音类别 (Pishdadian et al.,2020; Kong et al.,2020), 视觉输入 (Gao & Grauman, 2019) 或来自多麦克风录音的空间位置 (Tzinis et al., 2019; Seetharaman et al., 2019; Drude et al., 2019)。最近，Wisdom et al.(2020) 提出了混合不变训练 ( MixIT), 它为可变数量的潜在源提供了纯无监督的源分离框架
各种研究为解决视听屏幕上的源分离 (Michelsanti et al.,2020) 奠定了基础。通常，两种主要方法是使用视听定位 (Hershey & Movellan, 2000; Senocak et al., 2018; Wu et al., 2019; Afouras et al., 2020)或对象检测网络(有监督的 (Ephrat et al., 2018; Gao& Grauman, 2019; Gan et al., 2020) 或无监督的 (Zhao et al.,2018)) 来预测视觉调节信息。但是，这些作品仅考虑受限制的域，例如语音(Hershey& < Casey, 2002; Ephrat et al., 2018; Afouras et al., 2020) 或音乐 (Zhao et al., 2018; Gao & Grauman, 2019; Gan et al., 2020) 。Gao et al. (2018) 报告了来自广泛领域的视频的结果，但依赖于有监督的视觉对象检测器，这排除了对检测器定义的封闭类集之外的声源外观的了解。Rouditchenko et al.(2019) 提出了一个用于广泛声音领域的系统，但需要声音类标签以及来自这些类的孤立声音。我们的方法避免了对类标签和孤立源的监督，以处理多源数据中出现的末知视觉和声音类
为了直接从限制较少的野外视频数据开放域中学习，Tian et al.(2018) 学会了在不受约束的视频中本地化视听事件，并提供了一个临时数据集。Korbar et al.(2018) 预训练模型来识别音频-视频对的时间同步，并在动作识别和音频分类方面展示了有希望的结果。Arandjelovic & Zisserman (2017) 通过对一个视频帧和一秒钟音频对的视听对应关系进行分类，采用了类似的方法。Hu et al.(2020) 提出了一种课程学习方法，其中模型逐渐学习更难的示例以分离最接近我们工作的是 Owens & Efros (2018) 的方法，这是一种基于时间视听对齐的自监督视听屏幕语音分离系统。但是，Owens&Efros (2018)假定训练视频仅包含屏幕上的源，目前尚不清楚如何使其适应训练视频包含屏幕外的源的情况
我们的方法与这些先前的作品显着不同，因为我们不将我们的领域限制为乐器或人类扬声器，并且我们使用包含任意数量的对象的真实野外视频进行训练和测试，没有对象类限制。
2
|
[
[
0.1748046875,
0.034088134765625,
0.476318359375,
0.034088134765625,
0.476318359375,
0.048614501953125,
0.1748046875,
0.048614501953125
],
[
0.176513671875,
0.1060791015625,
0.82275390625,
0.1060791015625,
0.82275390625,
0.11932373046875,
0.176513671875,
0.11932373046875
],
[
0.17724609375,
0.11932373046875,
0.82421875,
0.11932373046875,
0.82421875,
0.1331787109375,
0.17724609375,
0.1331787109375
],
[
0.176513671875,
0.1331787109375,
0.82275390625,
0.1331787109375,
0.82275390625,
0.146484375,
0.176513671875,
0.146484375
],
[
0.1829833984375,
0.1483154296875,
0.82177734375,
0.1483154296875,
0.82177734375,
0.1590576171875,
0.1829833984375,
0.1590576171875
],
[
0.178955078125,
0.1622314453125,
0.82177734375,
0.1622314453125,
0.82177734375,
0.17236328125,
0.178955078125,
0.17236328125
],
[
0.17724609375,
0.1749267578125,
0.82177734375,
0.1749267578125,
0.82177734375,
0.1881103515625,
0.17724609375,
0.1881103515625
],
[
0.176513671875,
0.1881103515625,
0.82177734375,
0.1881103515625,
0.82177734375,
0.2020263671875,
0.176513671875,
0.2020263671875
],
[
0.176513671875,
0.2020263671875,
0.82177734375,
0.2020263671875,
0.82177734375,
0.21533203125,
0.176513671875,
0.21533203125
],
[
0.176513671875,
0.216552734375,
0.76123046875,
0.216552734375,
0.76123046875,
0.2298583984375,
0.176513671875,
0.2298583984375
],
[
0.176513671875,
0.2374267578125,
0.82177734375,
0.2374267578125,
0.82177734375,
0.250732421875,
0.176513671875,
0.250732421875
],
[
0.176513671875,
0.251953125,
0.82177734375,
0.251953125,
0.82177734375,
0.26513671875,
0.176513671875,
0.26513671875
],
[
0.176513671875,
0.26513671875,
0.8203125,
0.26513671875,
0.8203125,
0.279052734375,
0.176513671875,
0.279052734375
],
[
0.178955078125,
0.281005859375,
0.82421875,
0.281005859375,
0.82421875,
0.291748046875,
0.178955078125,
0.291748046875
],
[
0.176513671875,
0.29345703125,
0.82275390625,
0.29345703125,
0.82275390625,
0.306884765625,
0.176513671875,
0.306884765625
],
[
0.1748046875,
0.30615234375,
0.8203125,
0.30615234375,
0.8203125,
0.3193359375,
0.1748046875,
0.3193359375
],
[
0.176513671875,
0.3193359375,
0.82177734375,
0.3193359375,
0.82177734375,
0.333251953125,
0.176513671875,
0.333251953125
],
[
0.176513671875,
0.333984375,
0.82177734375,
0.333984375,
0.82177734375,
0.347900390625,
0.176513671875,
0.347900390625
],
[
0.1748046875,
0.347900390625,
0.82177734375,
0.3466796875,
0.82177734375,
0.361083984375,
0.1748046875,
0.3623046875
],
[
0.17724609375,
0.363037109375,
0.27294921875,
0.363037109375,
0.27294921875,
0.375,
0.17724609375,
0.375
],
[
0.176513671875,
0.40771484375,
0.214111328125,
0.40771484375,
0.214111328125,
0.419189453125,
0.176513671875,
0.419189453125
],
[
0.2059326171875,
0.4072265625,
0.4697265625,
0.40771484375,
0.4697265625,
0.421630859375,
0.2059326171875,
0.42041015625
],
[
0.178955078125,
0.44580078125,
0.82177734375,
0.44580078125,
0.82177734375,
0.455810546875,
0.178955078125,
0.455810546875
],
[
0.176513671875,
0.458251953125,
0.8203125,
0.458251953125,
0.8203125,
0.4716796875,
0.176513671875,
0.4716796875
],
[
0.176513671875,
0.4716796875,
0.82177734375,
0.4716796875,
0.82177734375,
0.48486328125,
0.176513671875,
0.48486328125
],
[
0.178955078125,
0.4873046875,
0.8203125,
0.4873046875,
0.8203125,
0.49755859375,
0.178955078125,
0.49755859375
],
[
0.176513671875,
0.5,
0.82177734375,
0.5,
0.82177734375,
0.51318359375,
0.176513671875,
0.51318359375
],
[
0.17724609375,
0.5146484375,
0.82275390625,
0.5146484375,
0.82275390625,
0.52783203125,
0.17724609375,
0.52783203125
],
[
0.17724609375,
0.52880859375,
0.2303466796875,
0.52880859375,
0.2303466796875,
0.5390625,
0.17724609375,
0.5390625
],
[
0.176513671875,
0.548828125,
0.82275390625,
0.548828125,
0.82275390625,
0.5625,
0.176513671875,
0.5625
],
[
0.176513671875,
0.56298828125,
0.82275390625,
0.56298828125,
0.82275390625,
0.5771484375,
0.176513671875,
0.5771484375
],
[
0.176513671875,
0.5771484375,
0.82275390625,
0.5771484375,
0.82275390625,
0.59033203125,
0.176513671875,
0.59033203125
],
[
0.1748046875,
0.59033203125,
0.82275390625,
0.59033203125,
0.82275390625,
0.60498046875,
0.1748046875,
0.60498046875
],
[
0.176513671875,
0.60498046875,
0.4697265625,
0.60498046875,
0.4697265625,
0.61865234375,
0.176513671875,
0.61865234375
],
[
0.17724609375,
0.625,
0.82275390625,
0.625,
0.82275390625,
0.63818359375,
0.17724609375,
0.63818359375
],
[
0.176513671875,
0.6396484375,
0.82275390625,
0.6396484375,
0.82275390625,
0.65283203125,
0.176513671875,
0.65283203125
],
[
0.176513671875,
0.65283203125,
0.8203125,
0.65283203125,
0.8203125,
0.66650390625,
0.176513671875,
0.66650390625
],
[
0.17724609375,
0.66845703125,
0.8203125,
0.66845703125,
0.8203125,
0.6787109375,
0.17724609375,
0.6787109375
],
[
0.1748046875,
0.6796875,
0.8203125,
0.6796875,
0.8203125,
0.693359375,
0.1748046875,
0.693359375
],
[
0.176513671875,
0.6943359375,
0.82177734375,
0.6943359375,
0.82177734375,
0.70849609375,
0.176513671875,
0.70849609375
],
[
0.178955078125,
0.71044921875,
0.8203125,
0.71044921875,
0.8203125,
0.72021484375,
0.178955078125,
0.72021484375
],
[
0.176513671875,
0.72265625,
0.82177734375,
0.72265625,
0.82177734375,
0.736328125,
0.176513671875,
0.736328125
],
[
0.176513671875,
0.736328125,
0.82275390625,
0.736328125,
0.82275390625,
0.75,
0.176513671875,
0.75
],
[
0.17724609375,
0.75048828125,
0.46728515625,
0.75048828125,
0.46728515625,
0.76123046875,
0.17724609375,
0.76123046875
],
[
0.176513671875,
0.771484375,
0.82177734375,
0.77099609375,
0.82177734375,
0.78515625,
0.176513671875,
0.7861328125
],
[
0.17724609375,
0.7841796875,
0.82177734375,
0.7841796875,
0.82177734375,
0.79736328125,
0.17724609375,
0.79736328125
],
[
0.1748046875,
0.798828125,
0.82177734375,
0.7998046875,
0.82177734375,
0.81298828125,
0.1748046875,
0.8125
],
[
0.1748046875,
0.81298828125,
0.82275390625,
0.81298828125,
0.82275390625,
0.8271484375,
0.1748046875,
0.8271484375
],
[
0.176513671875,
0.8271484375,
0.82177734375,
0.8271484375,
0.82177734375,
0.84033203125,
0.176513671875,
0.84033203125
],
[
0.1748046875,
0.8466796875,
0.82275390625,
0.84765625,
0.82275390625,
0.8623046875,
0.1748046875,
0.861328125
],
[
0.176513671875,
0.8623046875,
0.82275390625,
0.8623046875,
0.82275390625,
0.87548828125,
0.176513671875,
0.87548828125
],
[
0.17724609375,
0.87744140625,
0.564453125,
0.87744140625,
0.564453125,
0.88818359375,
0.17724609375,
0.88818359375
],
[
0.176513671875,
0.89794921875,
0.82275390625,
0.89794921875,
0.82275390625,
0.9111328125,
0.176513671875,
0.9111328125
],
[
0.176513671875,
0.9111328125,
0.82275390625,
0.9111328125,
0.82275390625,
0.9248046875,
0.176513671875,
0.9248046875
],
[
0.491943359375,
0.94970703125,
0.5048828125,
0.94970703125,
0.5048828125,
0.96044921875,
0.491943359375,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2021",
"这种直觉促使使用视觉作为音频处理的接口，其中一个主要问题是选择性地保留所需的声",
"音，同时去除不需要的声音。在某些任务中，例如语音增强，可以按其类别选择所需的声音：",
"在这种情况下，语音与非语音。在开放域设置中，所需声音的选择由用户自行决定。这带来",
"了一个用户界面问题：使用音频以有效的方式选择源具有挑战性。如果我们使用视频选择",
"作为音频选择的代理，例如，通过从屏幕上的对象中选择声音并册删除屏幕外的声音，则在",
"视听情况下可以大大简化此问题。最近的工作使用视频选择和分离语音(Ephratetal.,2018",
"Afouras et al.,2020)或音乐(Zhao et al.,2018;Gao &Grauman,2019;Gan et al.,2020)。但是",
"针对任意声音（Gaoetal.，2018；Rouditchenkoetal.,2019;Owens&Efros，2018）解决此问题的",
"系统可能在更一般的情况下有用，例如视频录制，其中无法预先定义感兴趣的声音",
"将任意声音与其视觉对象相关联的问题在开放域中具有挑战性。出现了一些以前的工作尚",
"未完全解决的复杂性。首先，为了覆盖可能的声音空间，需要大量的训练数据。受监督的",
"方法需要标记的示例，其中已知孤立的屏幕声音。由此产生的数据收集和标签负担限制了",
"可用数据的数量和质量。为了克服这个问题，我们提出了一种使用混合不变训练（MixIT）",
"(Wisdometal.,2020)的无监督方法，该方法可以学习将单个源与野外视频分开，其中屏幕上",
"和屏幕外的声音是未知的。另一个问题是，不同的音频源可能对应于任意空间位置的一组",
"动态屏幕上对象。我们通过使用注意机制来适应这一点，该机制将每个假设的音频源与视",
"频中相应对象的不同空间和时间位置对齐。最后，我们需要确定哪些音频源出现在屏幕上，",
"在没有强标签的情况下。这是使用基于注意力机制产生的音频和视频嵌入的源的弱训练分",
"类器来处理的",
"2",
"RELATIONTOPREVIOUSWORK",
"从混合物中分离任意声音，称为“通用声音分离”，最近被证明可以使用固定数量的声音",
"(Kavalerovetal.，2019）。有关存在哪些声音类别的条件信息可以提高分离性能（Tzinisetal.",
"2020）。FUSS数据集（Wisdometal.,2021）扩展了范围，以分离可变数量的声音，以处理更",
"真实的数据。还提出了一个框架，可以从输入声音混合物（Ochiaietal.2020）中提取特定的",
"声音类别。这些方法需要包含孤立声音的精选数据进行训练，这阻止了它们应用于真正的",
"开放域数据，并引人了诸如注释成本，真实声学混合的准确模拟以及有偏差的数据集之类",
"的困难。",
"为了避免这些问题，最近的许多工作都建议用来自相关模态的弱监督标签代替对参考源",
"信号的强监督，例如声音类别（Pishdadianetal.，2020;Kongetal.，2020），视觉输人（Gao&",
"Grauman，2019）或来自多麦克风录音的空间位置（Tzinisetal.，2019;Seetharamanetal.，2019：",
"Drudeetal.，2019）。最近，Wisdometal.（2020）提出了混合不变训练（MixIT），它为可变数",
"量的潜在源提供了纯无监督的源分离框架",
"各种研究为解决视听屏幕上的源分离（Michelsantietal.，2020）奠定了基础。通常，两种主要",
"方法是使用视听定位（Hershey&Movellan，2000;Senocaketal.,2018；Wuetal.,2019；Afouras",
"etal.,2020）或对象检测网络（有监督的(Ephratetal.，2018;Gao&Grauman，2019;Ganetal.",
"2020）或无监督的（Zha0etal.，2018））来预测视觉调节信息。但是，这些作品仅考虑受限制",
"的域，例如语音(Hershey&Casey2002;Ephratetal.,2018;Afouras etal.,2020)或音乐(Zhac",
"etal.,2018;Gao&Grauman,2019;Gan etal.,2020）。Gaoetal.（2018）报告了来自广泛领域的",
"视频的结果，但依赖于有监督的视觉对象检测器，这排除了对检测器定义的封闭类集之外",
"的声源外观的了解。Rouditchenkoetal.（2019）提出了一个用于广泛声音领域的系统，但需要",
"声音类标签以及来自这些类的孤立声音。我们的方法避免了对类标签和孤立源的监督，以",
"处理多源数据中出现的未知视觉和声音类",
"为了直接从限制较少的野外视频数据开放域中学习，Tian etal.(2018)学会了在不受约束的",
"视频中本地化视听事件，并提供了一个临时数据集。Korbaretal.(2018）预训练模型来识别",
"音频-视频对的时间同步，并在动作识别和音频分类方面展示了有希望的结果。Arandjelovic",
"&Zisserman(2017）通过对一个视频顿和一秒钟音频对的视听对应关系进行分类，采用了类",
"似的方法。Huetal.(2020)提出了一种课程学习方法，其中模型逐渐学习更难的示例以分离",
"最接近我们工作的是Owens&Efros（2018)的方法，这是一种基于时间视听对齐的自监督视",
"听屏幕语音分离系统。但是，Owens&Efros（2018）假定训练视频仅包含屏幕上的源，目前",
"尚不清楚如何使其适应训练视频包含屏幕外的源的情况",
"我们的方法与这些先前的作品显着不同，因为我们不将我们的领域限制为乐器或人类扬声",
"器，并且我们使用包含任意数量的对象的真实野外视频进行训练和测试，没有对象类限制。",
"2"
] |
{
"arxiv_id": "2011.01143",
"language": "zh",
"page_idx": 1
}
|
||
2011.01143-zh-2
|
2011.01143-zh-2.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2021}}$
我们提出的框架可以处理嘈杂的标签(例如。g。没有屏幕声音的视频),在完全开放的野外
视频域上操作，并有效隔离来自屏幕对象的声音
我们解决以下任务，该任务扩展了屏幕上语音分离问题 (Owens&Efros,2018) 的表述。给定输人视频，目标是分离构成输入混合的所有源，然后估计每个分离源的视听对应得分。对于屏幕上明显的分离源，这些概率分数应该很高，否则应该很低。可以将由其估计的屏幕上概率加权的分离的音频源相加，以重建屏幕上的混合。我们强调，我们的方法比以前的建议更普遍适用，因为现实世界的视频可能包含未知数量的屏幕上和屏幕外来源，属于未定义的类本体
我们在本文中做出以下贡献：
1. We provide the first solution for training an unsupervised, open-domain, audio-visual on-
screen separation system from scratch on real in-the-wild video data, with no requirement
on modules such as object detectors that require supervised data.
2. We develop a new dataset for the on-screen audio-visual separation task, drawn from 2,500
hours of unlabeled videos from YFCC100m, and 55 hours of videos that are human-labeled
for presence of on-screen and off-screen sounds.
# 3 MODEL ARCHITECTURE
AudioScope 的整体架构由以下块构建：图像嵌人网络，音频分离网络，音频嵌入网络，视听注意力机制和屏幕分类器 (请参见图 2)。分离和嵌人网络基于先前的工作，并在以下小节中进行描述。然而，这项工作的主要重点是整体架构，以及训练框架和损失函数
使用图像嵌人网络对视频进行分析，该网络为每个帧内的 64 个位置中的每个位置生成局部嵌入，以及整个帧的嵌人。这些嵌入既用作音频分离网络的调节输入，也用作屏幕声音分类的输入。音频分离网络将混合的输人波形作为输入，并生成固定数量的输出波形，其中可变数量的非零取决于混合中估计的源数量。对视频的调节使分离能够在执行分离时利用关于存在的源的线索。将音频嵌人网络应用于每个估计源，以获得每个源的每帧一个嵌人。然后，这些音频嵌人会随着时间的推移而汇集，并在视听时空关注网络中使用，以针对每个源检索与音频最匹配的视觉活动的表示，类似于从内部网络中提取的关联图 Harwath et al. (2018)提出的表示
该体系结构旨在解决野外开放域数据上的无监督学习问题。首先，由于目标训练视频可以包含屏幕上和屏幕外的声音，因此训练系统直接产生目标视频的音频将鼓励包含屏幕外声音以及屏幕上的声音$^{1}$。相反，我们提出的多源分离网络使用无监督的 MixIT 目标产生潜在的源估计，该目标已被证明在一般声音分离(Wisdom et al., 2020)上表现良好。。通过将分离与屏幕上的分类解耦，我们的架构促进了强大的目标的使用，这些目标允许将某些源视为屏幕外的源，即使它们出现在目标视频的配乐中
视听注意力体系结构是由音频和视频之间的对齐问题激发的：视频中的声源对象可能会被本地化，可能会随时间移动，并且可能在相应的音频活动之前和之后出现。由于开放域，我们不能像在某些先前作品 (Ephrat et al., 2018; Gao & Grauman, 2019; Gan et al., 2020) 中所做的那样，依靠一组预定义的对象检测器来锚定屏幕上源的视频表示。相反，我们建议注意以更灵活的方式找到与来源相对应的视频表示
所提出的音频嵌入的时间池化策略，在将其用于时空关注之前，允许网络导出表示源音频的活动片段的嵌入，并忽略模糊的静音区域。在本模型中，以低帧速率分析视频，因此视听对应关系可能基于对象的相对静态属性，而不是其运动与音频的同步。在这种情况下，音频的单个时不变表示可能足以作为概念证明。然而，在未来的工作中，随着更高的视频帧速率，可能值得考虑使用注意力来对齐音频和视频嵌人序列，以便检测其活动模式中的同步屏幕分类器使用密集网络对一个估计源的音频嵌人以及由时空注意机制检索的视频嵌人进行操作。这大概允许检测嵌人之间的一致性。为了提供此决策的附加上下文，提供了通过时间池化产生的全局视频嵌人作为附加输入。对于这种分类器设计，许多替代选择是可能的， 我们将其留给未来的工作，例如使用更复杂的分类架构，或提供额外的音频嵌人作为输入
$^{\text{l}}$We train such a system in Appendix A.3.5, and find that it is not an effective approach.
3
|
[
[
0.1732177734375,
0.034088134765625,
0.47802734375,
0.0328369140625,
0.47802734375,
0.04736328125,
0.1732177734375,
0.048614501953125
],
[
0.176513671875,
0.1060791015625,
0.82177734375,
0.1060791015625,
0.82177734375,
0.12060546875,
0.176513671875,
0.12060546875
],
[
0.176513671875,
0.11871337890625,
0.515625,
0.11932373046875,
0.515625,
0.1331787109375,
0.176513671875,
0.1319580078125
],
[
0.176513671875,
0.141357421875,
0.82275390625,
0.141357421875,
0.82275390625,
0.1546630859375,
0.176513671875,
0.1546630859375
],
[
0.1748046875,
0.154052734375,
0.8203125,
0.154052734375,
0.8203125,
0.1673583984375,
0.1748046875,
0.1673583984375
],
[
0.176513671875,
0.1685791015625,
0.82275390625,
0.1685791015625,
0.82275390625,
0.1817626953125,
0.176513671875,
0.1817626953125
],
[
0.176513671875,
0.18310546875,
0.82177734375,
0.18310546875,
0.82177734375,
0.1962890625,
0.176513671875,
0.1962890625
],
[
0.176513671875,
0.1962890625,
0.82177734375,
0.1962890625,
0.82177734375,
0.210205078125,
0.176513671875,
0.210205078125
],
[
0.176513671875,
0.2108154296875,
0.223876953125,
0.2108154296875,
0.223876953125,
0.222900390625,
0.176513671875,
0.222900390625
],
[
0.17724609375,
0.23291015625,
0.374267578125,
0.23291015625,
0.374267578125,
0.24365234375,
0.17724609375,
0.24365234375
],
[
0.214111328125,
0.26318359375,
0.82275390625,
0.26318359375,
0.82275390625,
0.27783203125,
0.214111328125,
0.27783203125
],
[
0.233642578125,
0.27783203125,
0.82275390625,
0.276611328125,
0.82275390625,
0.291748046875,
0.233642578125,
0.292236328125
],
[
0.2344970703125,
0.292236328125,
0.65673828125,
0.292236328125,
0.65673828125,
0.30615234375,
0.2344970703125,
0.30615234375
],
[
0.2115478515625,
0.31494140625,
0.82275390625,
0.31640625,
0.82275390625,
0.330810546875,
0.2115478515625,
0.330078125
],
[
0.2344970703125,
0.330810546875,
0.82421875,
0.330810546875,
0.82421875,
0.3447265625,
0.2344970703125,
0.3447265625
],
[
0.2344970703125,
0.3466796875,
0.54833984375,
0.3466796875,
0.54833984375,
0.356689453125,
0.2344970703125,
0.356689453125
],
[
0.17236328125,
0.386474609375,
0.410888671875,
0.386962890625,
0.410888671875,
0.401611328125,
0.17236328125,
0.40087890625
],
[
0.176513671875,
0.421630859375,
0.82177734375,
0.421630859375,
0.82177734375,
0.43505859375,
0.176513671875,
0.43505859375
],
[
0.17724609375,
0.4375,
0.8203125,
0.4375,
0.8203125,
0.447509765625,
0.17724609375,
0.447509765625
],
[
0.178955078125,
0.451904296875,
0.744140625,
0.451904296875,
0.744140625,
0.462158203125,
0.178955078125,
0.462158203125
],
[
0.17724609375,
0.4716796875,
0.82421875,
0.4716796875,
0.82421875,
0.48486328125,
0.17724609375,
0.48486328125
],
[
0.176513671875,
0.48486328125,
0.8203125,
0.48486328125,
0.8203125,
0.498779296875,
0.176513671875,
0.498779296875
],
[
0.1748046875,
0.498779296875,
0.8203125,
0.498779296875,
0.8203125,
0.51220703125,
0.1748046875,
0.51220703125
],
[
0.176513671875,
0.51123046875,
0.82177734375,
0.51220703125,
0.82177734375,
0.5263671875,
0.176513671875,
0.52587890625
],
[
0.17724609375,
0.5263671875,
0.82177734375,
0.5263671875,
0.82177734375,
0.54052734375,
0.17724609375,
0.54052734375
],
[
0.176513671875,
0.54052734375,
0.82275390625,
0.54052734375,
0.82275390625,
0.5537109375,
0.176513671875,
0.5537109375
],
[
0.1748046875,
0.55322265625,
0.82275390625,
0.5537109375,
0.82275390625,
0.568359375,
0.1748046875,
0.5673828125
],
[
0.1748046875,
0.5673828125,
0.303955078125,
0.5673828125,
0.303955078125,
0.58056640625,
0.1748046875,
0.58056640625
],
[
0.176513671875,
0.5888671875,
0.82275390625,
0.5888671875,
0.82275390625,
0.60302734375,
0.176513671875,
0.60302734375
],
[
0.17724609375,
0.60400390625,
0.82275390625,
0.60400390625,
0.82275390625,
0.6171875,
0.17724609375,
0.6171875
],
[
0.17724609375,
0.6171875,
0.82275390625,
0.6171875,
0.82275390625,
0.630859375,
0.17724609375,
0.630859375
],
[
0.176513671875,
0.630859375,
0.82177734375,
0.630859375,
0.82177734375,
0.64453125,
0.176513671875,
0.64453125
],
[
0.17724609375,
0.64599609375,
0.82177734375,
0.64599609375,
0.82177734375,
0.6591796875,
0.17724609375,
0.6591796875
],
[
0.17724609375,
0.66015625,
0.51708984375,
0.66015625,
0.51708984375,
0.67041015625,
0.17724609375,
0.67041015625
],
[
0.178955078125,
0.68115234375,
0.82177734375,
0.68115234375,
0.82177734375,
0.69140625,
0.178955078125,
0.69140625
],
[
0.17724609375,
0.693359375,
0.82275390625,
0.693359375,
0.82275390625,
0.70703125,
0.17724609375,
0.70703125
],
[
0.176513671875,
0.70556640625,
0.82275390625,
0.70703125,
0.82275390625,
0.7216796875,
0.176513671875,
0.72021484375
],
[
0.176513671875,
0.7216796875,
0.82275390625,
0.7216796875,
0.82275390625,
0.73486328125,
0.176513671875,
0.73486328125
],
[
0.1748046875,
0.734375,
0.484375,
0.73486328125,
0.484375,
0.74853515625,
0.1748046875,
0.74755859375
],
[
0.1748046875,
0.755859375,
0.82275390625,
0.7568359375,
0.82275390625,
0.77099609375,
0.1748046875,
0.76953125
],
[
0.17724609375,
0.77294921875,
0.82177734375,
0.77294921875,
0.82177734375,
0.78271484375,
0.17724609375,
0.78271484375
],
[
0.178955078125,
0.7861328125,
0.8203125,
0.7861328125,
0.8203125,
0.796875,
0.178955078125,
0.796875
],
[
0.176513671875,
0.798828125,
0.82275390625,
0.798828125,
0.82275390625,
0.8125,
0.176513671875,
0.8125
],
[
0.17724609375,
0.814453125,
0.8095703125,
0.814453125,
0.8095703125,
0.82470703125,
0.17724609375,
0.82470703125
],
[
0.176513671875,
0.83349609375,
0.82421875,
0.83349609375,
0.82421875,
0.8466796875,
0.176513671875,
0.8466796875
],
[
0.17724609375,
0.8486328125,
0.82177734375,
0.8486328125,
0.82177734375,
0.859375,
0.17724609375,
0.859375
],
[
0.1748046875,
0.85986328125,
0.82275390625,
0.861328125,
0.82275390625,
0.875,
0.1748046875,
0.87353515625
],
[
0.17724609375,
0.876953125,
0.810546875,
0.876953125,
0.810546875,
0.88720703125,
0.17724609375,
0.88720703125
],
[
0.1968994140625,
0.91015625,
0.71337890625,
0.912109375,
0.71337890625,
0.9267578125,
0.1968994140625,
0.9248046875
],
[
0.5,
0.9462890625,
0.5107421875,
0.955078125,
0.5,
0.9638671875,
0.488525390625,
0.95556640625
]
] |
[
"Published as a conference paper at ICLR 2021",
"我们提出的框架可以处理噜杂的标签（例如。g。没有屏幕声音的视频)，在完全开放的野外",
"视频域上操作，并有效隔离来自屏幕对象的声音",
"我们解决以下任务，该任务扩展了屏幕上语音分离问题（Owens&Efros2018）的表述。给定",
"输人视频，目标是分离构成输入混合的所有源，然后估计每个分离源的视听对应得分。对于",
"屏幕上明显的分离源，这些概率分数应该很高，否则应该很低。可以将由其估计的屏幕上概",
"率加权的分离的音频源相加，以重建屏幕上的混合。我们强调，我们的方法比以前的建议更",
"普遍适用，因为现实世界的视频可能包含未知数量的屏幕上和屏幕外来源，属于未定义的",
"类本体",
"我们在本文中做出以下贡献",
"1. We provide the first solution for training an unsupervised, open-domain, audio-visual on",
"screen separation system from scratch on real in-the-wild video data, with no requirement",
"on modules such as object detectors that require supervised data.",
"2. We develop a new dataset for the on-screen audio-visual separation task, drawn from 2,500",
"hours of unlabeled videos from YFCC10Om, and 55 hours of videos that are human-labeled",
"forpresenceofon-screenandoff-screensounds",
"3MODELARCHITECTURE",
"AudioScope的整体架构由以下块构建：图像嵌人网络，音频分离网络，音频嵌入网络，视听",
"注意力机制和屏幕分类器（请参见图2）。分离和嵌入网络基于先前的工作，并在以下小节",
"中进行描述。然而，这项工作的主要重点是整体架构，以及训练框架和损失函数",
"使用图像嵌入网络对视频进行分析，该网络为每个顿内的64个位置中的每个位置生成局部",
"嵌入，以及整个帧的嵌入。这些嵌人既用作音频分离网络的调节输入，也用作屏幕声音分",
"类的输入。音频分离网络将混合的输入波形作为输入，并生成固定数量的输出波形，其中",
"可变数量的非零取决于混合中估计的源数量。对视频的调节使分离能够在执行分离时利用",
"关于存在的源的线索。将音频嵌人网络应用于每个估计源，以获得每个源的每顿一个嵌人。",
"然后，这些音频嵌入会随着时间的推移而汇集，并在视听时空关注网络中使用，以针对每",
"个源检索与音频最匹配的视觉活动的表示，类似于从内部网络中提取的关联图Harwathetal.",
"(2018）提出的表示",
"该体系结构旨在解决野外开放域数据上的无监督学习问题。首先，由于目标训练视频可以",
"包含屏幕上和屏幕外的声音，因此训练系统直接产生目标视频的音频将鼓励包含屏幕外声",
"音以及屏幕上的声音1。相反，我们提出的多源分离网络使用无监督的MixIT目标产生潜在",
"的源估计，该目标已被证明在一般声音分离（Wisdometal.,2020）上表现良好。。通过将分离",
"与屏幕上的分类解耦，我们的架构促进了强大的目标的使用，这些目标允许将某些源视为",
"屏幕外的源，即使它们出现在自标视频的配乐中",
"视听注意力体系结构是由音频和视频之间的对齐问题激发的：视频中的声源对象可能会被",
"本地化，可能会随时间移动，并且可能在相应的音频活动之前和之后出现。由于开放域，我",
"们不能像在某些先前作品（Ephratetal.，2018;Gao&Grauman，2019;Ganetal，2020）中所做",
"的那样，依靠一组预定义的对象检测器来锚定屏幕上源的视频表示。相反，我们建议注意以",
"更灵活的方式找到与来源相对应的视频表示",
"所提出的音频嵌人的时间池化策略，在将其用于时空关注之前，允许网络导出表示源音频",
"的活动片段的嵌入，并忽略模糊的静音区域。在本模型中，以低顿速率分析视频，因此视听",
"对应关系可能基于对象的相对静态属性，而不是其运动与音频的同步。在这种情况下，音频",
"的单个时不变表示可能足以作为概念证明。然而，在未来的工作中，随着更高的视频顿速",
"率，可能值得考虑使用注意力来对齐音频和视频嵌入序列，以便检测其活动模式中的同步",
"屏幕分类器使用密集网络对一个估计源的音频嵌入以及由时空注意机制检索的视频嵌人进",
"行操作。这大概充许检测嵌人之间的一致性。为了提供此决策的附加上下文，提供了通过时",
"间池化产生的全局视频嵌人作为附加输人。对于这种分类器设计，许多替代选择是可能的，",
"我们将其留给未来的工作，例如使用更复杂的分类架构，或提供额外的音频嵌入作为输入",
"'We train such a system in Appendix A.3.5,and find that it is not an effective approach.",
"3"
] |
{
"arxiv_id": "2011.01143",
"language": "zh",
"page_idx": 2
}
|
||
2011.01143-zh-3
|
2011.01143-zh-3.jpg
|
Published as a conference paper at ICLR 2021
<FigureHere>
Figure 2: AudioScope system diagram with 2 input mixtures and 4 output sources
3.1 AUDIO SEPARATION NETWORK
分离网络$\mathcal{M}^\mathrm{s}$体系结构由可学习的卷积编码器和解码器层以及改进的时域卷积网络(TDCN++) 掩码网络(Wisdom et al., 2020) 组成。应用混合一致性投影 (Wisdom et al., 2019) 来约束分离的源，以将其添加到输人混合物中。分离网络处理$T$ -样本输人混合波形并输出不 约 术 刀 向 时 $l\text{标 , 以 小 共 你 加 到 输 入 批 日 彻 }^{\text{个 。万 两 网 }\text{扫 处 理 }1\text{ 一 个 本 轴 人 化 日 放 尼 介 拥 山 }}$ $M$ 估 计 源 $\hat{s} \in \mathbb{R} ^{M\times T}$。在 内 部 , 网 络 估 计 $M$ 掩 码 , 这 些 掩 码 与 编 码 的 输 人 混 合 的 激 活 相 手乘。时域信号$\hat{s}$是通过将解码器，转置卷积层应用于屏蔽系数来计算的
3.2 AUDIO EMBEDDING NETWORK
对于每个分离的源 $\hat{s}_m$ , 我们使用 MobileNetv1 架构$\boxed{\mathrm{E}}($Howard et al., 2017) 提取相应的全局音频嵌人，该架构由堆叠的 2D 可分离的扩展卷积块组成，最后有一个致密层。该网络$\mathcal{M}^\mathrm{a}$ 首先使用来自时域分离源的$F_\mathrm{a}$音频帧计算对数梅尔级频谱图，然后应用深度可分离卷积堆栈来生成$F_\mathrm{a}\times N$嵌人矩阵$Z_m^\mathrm{a}$ ,其中包含每个帧的$N$维行嵌人。对于每个源$m$ ,使用注意力池操作(Girdhar &Ramanan, 2017) 来形成静态音频嵌人向量$z_m^\mathrm{a}=$attend$(\bar{Z}_m^\mathrm{a},Z_m^\mathrm{a},Z_m^\mathrm{a})$ ,其中平均嵌人$\bar{Z}_m^\mathrm{a}=\frac1{F_\mathrm{a}}\sum_iZ_{m,i}^\mathrm{a}$是源$m$的查询向量。注意机制 (Bahdanau et al.,2015) 定义如下：
$$\mathrm{attend}(q,K,V)=\alpha^Tf_\mathrm{V}(V),\:\alpha=\mathrm{softmax}(\mathrm{tanh}\left(f_\mathrm{K}(K)\right)\mathrm{tanh}\left(f_\mathrm{q}(q)\right)^T),$$
(1)
带查询行向量$q$ ,注意权重分布列向量$\alpha$ ,密钥矩阵$K$ ,值矩阵$V$ ,可训练的行式密集层$f_\mathrm{V}$ ,
3.3 IMAGE EMBEDDING NETWORK
要从视频帧中提取视觉特征，我们再次使用 MobileNetv1 体系结构。此视觉嵌人模型$\mathcal{M}^\mathrm{v}$独
立应用于每个$F_\mathrm{v}$输入视频帧，并为每个图像$Z_j^\mathrm{v},~j\in\{1,\ldots,F_\mathrm{v}\}$提取静态长度嵌人
具有时间视频嵌人的调节分离网络：视频输入$Z_i^\mathrm{v}$的嵌人可用于调节分离网络 (Tzinis et al., 2020)。具体来说，图像嵌人通过密集层进行馈送，并且简单的最近邻上采样将时间维度与中间分离网络激活的时间维度相匹配。这些上采样和变换的图像嵌人与中间 TDCN++ 激活连接，并作为输入馈送到分离网络层
全局视频嵌人：使用所有视频帧的注意力池来提取视频输入的全局嵌入，由$z^\mathrm{vg}=$
attend$(\bar{Z}^{\mathrm{v}},Z^{\mathrm{v}},Z^{\mathrm{v}})$给出，其中平均嵌入$\bar{Z}^{\mathrm{v}}=\frac1{F_{\mathrm{v}}}\sum_{j}Z_{j}^{\mathrm{v}}$是查询向量
局部时空视频嵌入：我们还使用从视觉卷积网络中具有$8\times8$空间位置的中间层提取的局部特征。这些表示为$Z_k^\mathrm{vl}$,其中$k=(j,n)$索引视频帧$j$和空间位置索引$n$ 。这些提供了空间特征，用于识别具有视觉对象的源，以用于视听时空关注
4
|
[
[
0.1732177734375,
0.0328369140625,
0.47802734375,
0.032196044921875,
0.47802734375,
0.046722412109375,
0.1732177734375,
0.04736328125
],
[
0.328369140625,
0.10479736328125,
0.366943359375,
0.10479736328125,
0.366943359375,
0.11614990234375,
0.328369140625,
0.11614990234375
],
[
0.39208984375,
0.106689453125,
0.406005859375,
0.106689453125,
0.406005859375,
0.111083984375,
0.39208984375,
0.111083984375
],
[
0.483642578125,
0.1060791015625,
0.50341796875,
0.1060791015625,
0.50341796875,
0.11871337890625,
0.483642578125,
0.11871337890625
],
[
0.564453125,
0.10919189453125,
0.62841796875,
0.111083984375,
0.62841796875,
0.12249755859375,
0.564453125,
0.12060546875
],
[
0.316162109375,
0.117431640625,
0.378173828125,
0.117431640625,
0.378173828125,
0.1275634765625,
0.316162109375,
0.1275634765625
],
[
0.455810546875,
0.1256103515625,
0.525390625,
0.1256103515625,
0.525390625,
0.136962890625,
0.455810546875,
0.136962890625
],
[
0.56787109375,
0.121826171875,
0.62744140625,
0.121826171875,
0.62744140625,
0.1331787109375,
0.56787109375,
0.1331787109375
],
[
0.324462890625,
0.1287841796875,
0.370849609375,
0.1287841796875,
0.370849609375,
0.1395263671875,
0.324462890625,
0.1395263671875
],
[
0.646484375,
0.1275634765625,
0.66259765625,
0.1275634765625,
0.66259765625,
0.136962890625,
0.646484375,
0.136962890625
],
[
0.334228515625,
0.1383056640625,
0.361083984375,
0.1383056640625,
0.361083984375,
0.15087890625,
0.334228515625,
0.15087890625
],
[
0.45751953125,
0.1383056640625,
0.52294921875,
0.1383056640625,
0.52294921875,
0.149658203125,
0.45751953125,
0.149658203125
],
[
0.57421875,
0.1383056640625,
0.59716796875,
0.1383056640625,
0.59716796875,
0.15283203125,
0.57421875,
0.15283203125
],
[
0.55224609375,
0.1546630859375,
0.64208984375,
0.1572265625,
0.64208984375,
0.1717529296875,
0.55224609375,
0.169189453125
],
[
0.46728515625,
0.1610107421875,
0.5087890625,
0.1610107421875,
0.5087890625,
0.1767578125,
0.46728515625,
0.1767578125
],
[
0.57421875,
0.17041015625,
0.625,
0.17041015625,
0.625,
0.18115234375,
0.57421875,
0.18115234375
],
[
0.47216796875,
0.1787109375,
0.5048828125,
0.1787109375,
0.5048828125,
0.1900634765625,
0.47216796875,
0.1900634765625
],
[
0.701171875,
0.177978515625,
0.765625,
0.177978515625,
0.765625,
0.1881103515625,
0.701171875,
0.1881103515625
],
[
0.57666015625,
0.1849365234375,
0.5947265625,
0.1849365234375,
0.5947265625,
0.199462890625,
0.57666015625,
0.199462890625
],
[
0.69775390625,
0.1849365234375,
0.76318359375,
0.1875,
0.76318359375,
0.1988525390625,
0.69775390625,
0.1962890625
],
[
0.2099609375,
0.1925048828125,
0.285888671875,
0.1932373046875,
0.285888671875,
0.2052001953125,
0.2099609375,
0.203857421875
],
[
0.4453125,
0.1900634765625,
0.53125,
0.1900634765625,
0.53125,
0.2008056640625,
0.4453125,
0.2008056640625
],
[
0.441162109375,
0.2020263671875,
0.53759765625,
0.20263671875,
0.53759765625,
0.2139892578125,
0.441162109375,
0.21337890625
],
[
0.57861328125,
0.20263671875,
0.62744140625,
0.20263671875,
0.62744140625,
0.21337890625,
0.57861328125,
0.21337890625
],
[
0.685546875,
0.205810546875,
0.76708984375,
0.205810546875,
0.76708984375,
0.2171630859375,
0.685546875,
0.2171630859375
],
[
0.2279052734375,
0.2139892578125,
0.2646484375,
0.2139892578125,
0.2646484375,
0.2247314453125,
0.2279052734375,
0.2247314453125
],
[
0.583984375,
0.21337890625,
0.615234375,
0.21337890625,
0.615234375,
0.2203369140625,
0.583984375,
0.2203369140625
],
[
0.340576171875,
0.2215576171875,
0.37744140625,
0.2215576171875,
0.37744140625,
0.23291015625,
0.340576171875,
0.23291015625
],
[
0.619140625,
0.222900390625,
0.66650390625,
0.2298583984375,
0.66259765625,
0.246826171875,
0.615234375,
0.2392578125
],
[
0.71142578125,
0.2171630859375,
0.740234375,
0.2171630859375,
0.740234375,
0.2298583984375,
0.71142578125,
0.2298583984375
],
[
0.2115478515625,
0.2266845703125,
0.27783203125,
0.2266845703125,
0.27783203125,
0.2374267578125,
0.2115478515625,
0.2374267578125
],
[
0.324462890625,
0.23291015625,
0.39208984375,
0.2342529296875,
0.39208984375,
0.24560546875,
0.324462890625,
0.2442626953125
],
[
0.58251953125,
0.24560546875,
0.61669921875,
0.24560546875,
0.61669921875,
0.2568359375,
0.58251953125,
0.2568359375
],
[
0.689453125,
0.242431640625,
0.76953125,
0.242431640625,
0.76953125,
0.25244140625,
0.689453125,
0.25244140625
],
[
0.564453125,
0.255615234375,
0.6298828125,
0.25830078125,
0.6298828125,
0.26953125,
0.564453125,
0.267578125
],
[
0.2296142578125,
0.2474365234375,
0.2484130859375,
0.2474365234375,
0.2484130859375,
0.255126953125,
0.2296142578125,
0.255126953125
],
[
0.39453125,
0.25830078125,
0.43310546875,
0.25830078125,
0.43310546875,
0.26953125,
0.39453125,
0.26953125
],
[
0.384033203125,
0.270751953125,
0.443603515625,
0.270751953125,
0.443603515625,
0.281005859375,
0.384033203125,
0.281005859375
],
[
0.57666015625,
0.26953125,
0.625,
0.26953125,
0.625,
0.27978515625,
0.57666015625,
0.27978515625
],
[
0.240234375,
0.27978515625,
0.2484130859375,
0.27978515625,
0.2484130859375,
0.2841796875,
0.240234375,
0.2841796875
],
[
0.3896484375,
0.281005859375,
0.435546875,
0.281005859375,
0.435546875,
0.292236328125,
0.3896484375,
0.292236328125
],
[
0.58837890625,
0.27978515625,
0.61376953125,
0.27978515625,
0.61376953125,
0.29345703125,
0.58837890625,
0.29345703125
],
[
0.303955078125,
0.296630859375,
0.36279296875,
0.298583984375,
0.36279296875,
0.31005859375,
0.303955078125,
0.30810546875
],
[
0.398681640625,
0.294189453125,
0.420654296875,
0.290283203125,
0.4248046875,
0.302978515625,
0.40283203125,
0.306884765625
],
[
0.71728515625,
0.296630859375,
0.76708984375,
0.296630859375,
0.76708984375,
0.311279296875,
0.71728515625,
0.311279296875
],
[
0.57177734375,
0.304931640625,
0.58837890625,
0.304931640625,
0.58837890625,
0.314453125,
0.57177734375,
0.314453125
],
[
0.5986328125,
0.30615234375,
0.67236328125,
0.30615234375,
0.67236328125,
0.317626953125,
0.5986328125,
0.317626953125
],
[
0.736328125,
0.309326171875,
0.75341796875,
0.309326171875,
0.75341796875,
0.317626953125,
0.736328125,
0.317626953125
],
[
0.30810546875,
0.311279296875,
0.35693359375,
0.311279296875,
0.35693359375,
0.3212890625,
0.30810546875,
0.3212890625
],
[
0.259033203125,
0.311279296875,
0.26708984375,
0.311279296875,
0.26708984375,
0.314453125,
0.259033203125,
0.314453125
],
[
0.314453125,
0.3212890625,
0.35302734375,
0.3212890625,
0.35302734375,
0.333251953125,
0.314453125,
0.333251953125
],
[
0.7158203125,
0.3212890625,
0.74853515625,
0.3212890625,
0.74853515625,
0.333251953125,
0.7158203125,
0.333251953125
],
[
0.2197265625,
0.32763671875,
0.27294921875,
0.32763671875,
0.27294921875,
0.33837890625,
0.2197265625,
0.33837890625
],
[
0.59326171875,
0.339111328125,
0.67333984375,
0.341552734375,
0.67333984375,
0.35595703125,
0.59326171875,
0.354248046875
],
[
0.69677734375,
0.333984375,
0.7685546875,
0.333984375,
0.7685546875,
0.3447265625,
0.69677734375,
0.3447265625
],
[
0.2059326171875,
0.34033203125,
0.281005859375,
0.34033203125,
0.281005859375,
0.3515625,
0.2059326171875,
0.3515625
],
[
0.438720703125,
0.3466796875,
0.54345703125,
0.3466796875,
0.54345703125,
0.356689453125,
0.438720703125,
0.356689453125
],
[
0.2303466796875,
0.375,
0.76708984375,
0.375,
0.76708984375,
0.38818359375,
0.2303466796875,
0.38818359375
],
[
0.1732177734375,
0.4091796875,
0.221435546875,
0.4091796875,
0.221435546875,
0.423583984375,
0.1732177734375,
0.423583984375
],
[
0.21728515625,
0.410400390625,
0.43310546875,
0.410400390625,
0.43310546875,
0.423583984375,
0.21728515625,
0.423583984375
],
[
0.178955078125,
0.438232421875,
0.8203125,
0.438232421875,
0.8203125,
0.448974609375,
0.178955078125,
0.448974609375
],
[
0.176513671875,
0.448974609375,
0.82177734375,
0.449462890625,
0.82177734375,
0.464111328125,
0.176513671875,
0.46337890625
],
[
0.17724609375,
0.46337890625,
0.82177734375,
0.46337890625,
0.82177734375,
0.4765625,
0.17724609375,
0.4765625
],
[
0.1748046875,
0.477783203125,
0.330810546875,
0.473388671875,
0.330810546875,
0.488037109375,
0.1748046875,
0.492431640625
],
[
0.325927734375,
0.477783203125,
0.82275390625,
0.477783203125,
0.82275390625,
0.4912109375,
0.325927734375,
0.4912109375
],
[
0.176513671875,
0.4912109375,
0.68115234375,
0.4912109375,
0.68115234375,
0.5048828125,
0.176513671875,
0.5048828125
],
[
0.1748046875,
0.5234375,
0.431396484375,
0.5234375,
0.431396484375,
0.537109375,
0.1748046875,
0.537109375
],
[
0.176513671875,
0.5498046875,
0.82275390625,
0.5498046875,
0.82275390625,
0.56298828125,
0.176513671875,
0.56298828125
],
[
0.17724609375,
0.56298828125,
0.82177734375,
0.56298828125,
0.82177734375,
0.57763671875,
0.17724609375,
0.57763671875
],
[
0.17724609375,
0.57763671875,
0.82275390625,
0.57763671875,
0.82275390625,
0.59130859375,
0.17724609375,
0.59130859375
],
[
0.176513671875,
0.59130859375,
0.82275390625,
0.59130859375,
0.82275390625,
0.60498046875,
0.176513671875,
0.60498046875
],
[
0.1748046875,
0.60400390625,
0.82275390625,
0.60498046875,
0.82275390625,
0.619140625,
0.1748046875,
0.61865234375
],
[
0.1895751953125,
0.61865234375,
0.82421875,
0.61865234375,
0.82421875,
0.63623046875,
0.1895751953125,
0.63623046875
],
[
0.17724609375,
0.63623046875,
0.2303466796875,
0.63623046875,
0.2303466796875,
0.64794921875,
0.17724609375,
0.64794921875
],
[
0.2442626953125,
0.6591796875,
0.75634765625,
0.6591796875,
0.75634765625,
0.673828125,
0.2442626953125,
0.673828125
],
[
0.80126953125,
0.6591796875,
0.82421875,
0.6591796875,
0.82421875,
0.671875,
0.80126953125,
0.671875
],
[
0.1748046875,
0.68310546875,
0.82421875,
0.6845703125,
0.82421875,
0.69873046875,
0.1748046875,
0.69775390625
],
[
0.1732177734375,
0.71533203125,
0.4306640625,
0.71533203125,
0.4306640625,
0.72900390625,
0.1732177734375,
0.72900390625
],
[
0.176513671875,
0.7412109375,
0.82275390625,
0.7412109375,
0.82275390625,
0.7548828125,
0.176513671875,
0.7548828125
],
[
0.176513671875,
0.755859375,
0.775390625,
0.755859375,
0.775390625,
0.76953125,
0.176513671875,
0.76953125
],
[
0.1748046875,
0.77880859375,
0.82177734375,
0.77783203125,
0.82177734375,
0.79248046875,
0.1748046875,
0.79345703125
],
[
0.1748046875,
0.79443359375,
0.8203125,
0.79443359375,
0.8203125,
0.80810546875,
0.1748046875,
0.80810546875
],
[
0.17724609375,
0.81005859375,
0.82275390625,
0.81005859375,
0.82275390625,
0.82373046875,
0.17724609375,
0.82373046875
],
[
0.17724609375,
0.82470703125,
0.43701171875,
0.82470703125,
0.43701171875,
0.83544921875,
0.17724609375,
0.83544921875
],
[
0.1748046875,
0.8447265625,
0.81884765625,
0.8447265625,
0.81884765625,
0.85791015625,
0.1748046875,
0.85791015625
],
[
0.1707763671875,
0.85498046875,
0.66259765625,
0.85693359375,
0.66259765625,
0.876953125,
0.1707763671875,
0.875
],
[
0.176513671875,
0.88330078125,
0.82421875,
0.88330078125,
0.82421875,
0.896484375,
0.176513671875,
0.896484375
],
[
0.1748046875,
0.89599609375,
0.82275390625,
0.896484375,
0.82275390625,
0.9111328125,
0.1748046875,
0.91015625
],
[
0.1748046875,
0.9111328125,
0.5849609375,
0.9111328125,
0.5849609375,
0.9248046875,
0.1748046875,
0.9248046875
],
[
0.491943359375,
0.94970703125,
0.50634765625,
0.94970703125,
0.50634765625,
0.96044921875,
0.491943359375,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2021",
"Image",
"",
"zvg",
"On-screen",
"embedding",
"Globalvideo",
"classifier",
"network",
"y",
"M",
"embedding",
"zay",
"Spatiotemporal",
"Zvl",
"attention",
"Local",
"1234",
"zm",
"Predictions",
"Videoframes",
"spatiotemporal",
"video embedding",
"Temporal",
"Classification",
"Video",
"pooling",
"Visual",
"Zm.1:F.",
"loss",
"soundtrack",
"conditioning",
"Audio",
"Pseudo-labels",
"embedding",
"",
"Audio",
"separation",
"network",
"",
"network",
"Ma",
"Mixture of",
"M",
"234",
"3m",
"m= 1,2,3,4",
"y",
"mixtures",
"",
"(MoM)",
"Mix/T",
"Random",
"SNRloss",
"assignment",
"soundtrack",
"Estimatedsources",
"Figure 2: AudioScope system diagram with 2 input mixtures and 4 output sources.",
"3.1",
"AUDIOSEPARATIONNETWORK",
"分离网络M体系结构由可学习的卷积编码器和解码器层以及改进的时域卷积网络",
"（TDCN++）掩码网络（Wisdometal.，2020）组成。应用混合一致性投影（Wisdometal.，2019）",
"来约束分离的源，以将其添加到输入混合物中。分离网络处理T-样本输人混合波形并输出",
"M估计源SERMXT",
"。在内部，网络估计M掩码，这些掩码与编码的输入混合的激活相",
"乘。时域信号3是通过将解码器，转置卷积层应用于屏蔽系数来计算的",
"3.2AUDIOEMBEDDINGNETWORK",
"对于每个分离的源sm，我们使用MobileNetv1架构E(Howardetal.，2017）提取相应的全局",
"音频嵌人，该架构由堆叠的2D可分离的扩展卷积块组成，最后有一个致密层。该网络Ma",
"首先使用来自时域分离源的Fa音频顿计算对数梅尔级频谱图，然后应用深度可分离卷积堆",
"栈来生成Fa×N嵌人矩阵Za，其中包含每个顿的N维行嵌人。对于每个源m，使用注",
"意力池操作（Girdhar&Ramanan，2017）来形成静态音频嵌人向量=attend(Za，Za，Za）",
"其中平均嵌人Zm=青Z，Zm，是源m的查询向量。注意机制（Bahdanauetal.,20i5）定",
"义如下：",
"attend(q, K,V) = αT fv(V), α = softmax(tanh (fk(K))tanh(f.(q)\"),",
"(1)",
"带查询行向量9，注意权重分布列向量α，密钥矩阵K，值矩阵V，可训练的行式密集层fvV，",
"3.3IMAGEEMBEDDINGNETWORK",
"要从视频顿中提取视觉特征，我们再次使用MobileNetv1体系结构。此视觉嵌入模型MV独",
"立应用于每个Fv输人视频顿，并为每个图像Zy，jE[1,...,F}提取静态长度嵌入",
"具有时间视频嵌入的调节分离网络：视频输入Zy的嵌人可用于调节分离网络（Tzinisetal.",
"2020）。具体来说，图像嵌入通过密集层进行馈送，并且简单的最近邻上采样将时间维度与",
"中间分离网络激活的时间维度相匹配。这些上采样和变换的图像嵌人与中间TDCN++激活",
"连接，并作为输入馈送到分离网络层",
"全局视频嵌人：使用所有视频帧的注意力池来提取视频输入的全局嵌入，由之vg",
"attend(会，zz）给出，其中平均嵌人v=，z是查询向量",
"局部时空视频嵌入：我们还使用从视觉卷积网络中具有8×8空间位置的中间层提取的局部",
"特征。这些表示为z，其中k=（i,n）索引视频顿和空间位置索引n。这些提供了空间",
"特征，用于识别具有视觉对象的源，以用于视听时空关注",
"4"
] |
{
"arxiv_id": "2011.01143",
"language": "zh",
"page_idx": 3
}
|
||
2011.01143-zh-4
|
2011.01143-zh-4.jpg
|
$\underline{\text{Published as a conference paper at ICLR 2021}}$
3.4 AUDIO-VISUAL SPATIO-TEMPORAL ATTENTION
这项工作的一个重要方面是将音频和视觉信息结合起来，以推断每个分离的源与视频中相关对象之间的对应关系。这反过来将用于识别哪些源在屏幕上可见。为此，我们采用视听时空关注方案，让网络关注每个分离源的视觉嵌人的局部特征。在这种机制中，我们使用音频嵌人$z_m^\mathrm{a}$作为源$m$的查询输入，键和值输入由时空视频嵌人$Z_\:^\mathrm{vl}$给出。结果，与第$m$个源相对应的输出时空嵌人的扁平版本为$z_m^\mathrm{av}=$attend$(z_m^\mathrm{a},Z^\mathrm{vl},Z^\mathrm{vl})$
## 3.5 ON-SCREEN CLASSIFIER
为了推断每个分离的源的视觉存在，我们连接了全局视频嵌人$z^\mathrm{vg}$ ,每个源的全局音频嵌人$z_m^\mathrm{a}$以及相应的本地时空视听嵌人$z_m^\mathrm{av}$。级联向量通过具有逻辑激活的密集层$f_\mathrm{C}$馈送： $\hat{y}_{m}=$logistic$\left(f_{\mathrm{C}}\left([z^{\mathrm{vg}},z_{m}^{\mathrm{a}},z_{m}^{\mathrm{av}}]\right)\right)$
## 3.6 SEPARATION Loss
我们使用 MixIT 分离损失 (Wisdom et al, 2020), 它优化了$M$估计源$\hat{s}=\mathcal{M}^\mathrm{s}\left(x_1+x_2\right)$到两个参考混合物$x_1,x_2$的分配，如下所示：。由于$A$的限制，每个源$\hat{s}_m$只能分配给一个参考混合物。估计信号$\hat{t}\in\mathbb{R}^T$和目标信号$t\in\mathbb{R}^T$的 SNR 损耗定义为：
$$\mathcal{L}_{\mathrm{SNR}}(t,\hat{t})=10\log_{10}\left(\|t-\hat{t}\|^{2}+10^{-3}\|t\|^{2}\right).$$
(2)
3.7 CLASSIFICATION LOSS
为了训练屏幕上的分类器，我们考虑以下分类损失。这些损失使用二进制标签$y_m$ , 其 中 给 出了监督示例，在无监督情况下，每个源$m$使用$y_m=A_{1,m}^*$,其中$A^*$是通过(??)中的最小化找到的最佳混合矩阵。我们还使用符号$\mathcal{R} = \{ m| y_m= 1$, $m\in \{ 1, \ldots , M\} \}$来表示一组正标签
精确二进制交叉熵：
(3)
$$\mathcal{L}_{\text{exact}}\left(y,\hat{y}\right)=\sum_{m=1}^{M}\Big(-y_{m}\log\left(\hat{y}_{m}\right)+\left(y_{m}-1\right)\log\left(1-\hat{y}_{m}\right)\Big).$$
多实例交叉熵：由于分配给屏幕上混合的一些单独的源不在屏幕上，因此多实例(MI)(Maron
& Lozano-Pérez, 1998) 损失，可最大程度地减少正标签集$\mathcal{R}$可能更健壮：
(4)
$$\mathcal{L}_{\mathrm{MI}}\left(y,\hat{y}\right)=\min_{m\in\mathcal{R}}\Big(-\log\left(\hat{y}_{m}\right)-\sum_{m'\notin\mathcal{R}}\log\left(1-\hat{y}_{m'}\right)\Big).$$
$\begin{array}{c}\text{活动组合交叉熵：MI 损失的替代方法，活动组合(AC),对应于标签 s 的所有设置 }{\mathcal{P}}_{\geq1}\left(\mathcal{R}\right)\\\text{上的最小损失。t。至少一个标签是 }1:\end{array}$
(5)
$$\mathcal{L}_{\mathrm{AC}}\left(y,\hat{y}\right)=\min_{\mathcal{S}\in\mathcal{S}_{\geq1}(\mathcal{R})}\Big(-\sum_{m\in\mathcal{S}}\log\left(\hat{y}_{m}\right)-\sum_{m'\notin\mathcal{S}}\log\left(1-\hat{y}_{m'}\right)\Big).$$
, 其中 $\mathcal{O}_{\geq1}\left(\mathcal{R}\right)$表示标签为 1 的索引的幂集合
4 EXPERIMENTAL FRAMEWORK
4.1 DATA PREPARATION
为了训练我们的开放域系统的真实视听记录环境，我们使用 YahooFlickrCreativeCommons100百万数据集(YFCC100m)(Thomee et al.,2016)。该数据集来自各种长度的 200,0000 视频(2,500总小时数), 涵盖了各种语义声音类别。通过在视频上传器上进行拆分，我们选择1,600视频进行训练，并使用剩余的视频进行验证和测试。我们以 1秒的跳跃大小提取 5 秒的剪辑，从而产生大约 7.2 百万个剪辑。剪辑由以 16kHz 采样的 5 秒音频波形和 5 个视频帧$x^{(f)}$组成，其中每个帧是 128$\times128\times3$ RGB 图像
5
|
[
[
0.1748046875,
0.034088134765625,
0.47802734375,
0.034088134765625,
0.47802734375,
0.048614501953125,
0.1748046875,
0.048614501953125
],
[
0.1748046875,
0.1060791015625,
0.54345703125,
0.1060791015625,
0.54345703125,
0.11932373046875,
0.1748046875,
0.11932373046875
],
[
0.176513671875,
0.1307373046875,
0.82177734375,
0.1307373046875,
0.82177734375,
0.14453125,
0.176513671875,
0.14453125
],
[
0.176513671875,
0.14453125,
0.8203125,
0.14453125,
0.8203125,
0.1578369140625,
0.176513671875,
0.1578369140625
],
[
0.17724609375,
0.1590576171875,
0.82275390625,
0.1590576171875,
0.82275390625,
0.17236328125,
0.17724609375,
0.17236328125
],
[
0.176513671875,
0.17236328125,
0.82275390625,
0.17236328125,
0.82275390625,
0.186279296875,
0.176513671875,
0.186279296875
],
[
0.1732177734375,
0.1849365234375,
0.6259765625,
0.184326171875,
0.6259765625,
0.2020263671875,
0.1732177734375,
0.20263671875
],
[
0.1748046875,
0.2183837890625,
0.390625,
0.2183837890625,
0.390625,
0.231689453125,
0.1748046875,
0.231689453125
],
[
0.176513671875,
0.2442626953125,
0.82275390625,
0.2442626953125,
0.82275390625,
0.25830078125,
0.176513671875,
0.25830078125
],
[
0.1748046875,
0.2587890625,
0.82275390625,
0.2587890625,
0.82275390625,
0.272705078125,
0.1748046875,
0.272705078125
],
[
0.1748046875,
0.272705078125,
0.406005859375,
0.272705078125,
0.406005859375,
0.285888671875,
0.1748046875,
0.285888671875
],
[
0.1748046875,
0.3037109375,
0.3505859375,
0.3037109375,
0.3505859375,
0.317626953125,
0.1748046875,
0.317626953125
],
[
0.176513671875,
0.330078125,
0.82275390625,
0.330078125,
0.82275390625,
0.343505859375,
0.176513671875,
0.343505859375
],
[
0.17724609375,
0.3447265625,
0.82275390625,
0.3447265625,
0.82275390625,
0.359130859375,
0.17724609375,
0.359130859375
],
[
0.176513671875,
0.35791015625,
0.6611328125,
0.35791015625,
0.6611328125,
0.371826171875,
0.176513671875,
0.371826171875
],
[
0.346435546875,
0.380615234375,
0.65283203125,
0.380615234375,
0.65283203125,
0.39453125,
0.346435546875,
0.39453125
],
[
0.8037109375,
0.380615234375,
0.82421875,
0.380615234375,
0.82421875,
0.39453125,
0.8037109375,
0.39453125
],
[
0.1748046875,
0.41357421875,
0.37744140625,
0.41357421875,
0.37744140625,
0.4267578125,
0.1748046875,
0.4267578125
],
[
0.17724609375,
0.439453125,
0.82275390625,
0.439453125,
0.82275390625,
0.453857421875,
0.17724609375,
0.453857421875
],
[
0.176513671875,
0.45263671875,
0.82177734375,
0.45263671875,
0.82177734375,
0.46728515625,
0.176513671875,
0.46728515625
],
[
0.176513671875,
0.468505859375,
0.810546875,
0.468505859375,
0.810546875,
0.48291015625,
0.176513671875,
0.48291015625
],
[
0.176513671875,
0.484130859375,
0.240234375,
0.484130859375,
0.240234375,
0.49560546875,
0.176513671875,
0.49560546875
],
[
0.176513671875,
0.50390625,
0.31201171875,
0.50390625,
0.31201171875,
0.517578125,
0.176513671875,
0.517578125
],
[
0.40185546875,
0.5234375,
0.414306640625,
0.53271484375,
0.40283203125,
0.54150390625,
0.390625,
0.5322265625
],
[
0.38818359375,
0.5390625,
0.416748046875,
0.5390625,
0.416748046875,
0.5537109375,
0.38818359375,
0.5537109375
],
[
0.416748046875,
0.53857421875,
0.708984375,
0.53271484375,
0.708984375,
0.5546875,
0.416748046875,
0.56005859375
],
[
0.80224609375,
0.54052734375,
0.82421875,
0.54052734375,
0.82421875,
0.5546875,
0.80224609375,
0.5546875
],
[
0.390625,
0.556640625,
0.414306640625,
0.556640625,
0.414306640625,
0.56494140625,
0.390625,
0.56494140625
],
[
0.285888671875,
0.54052734375,
0.39453125,
0.54052734375,
0.39453125,
0.5537109375,
0.285888671875,
0.5537109375
],
[
0.17724609375,
0.58349609375,
0.82275390625,
0.58349609375,
0.82275390625,
0.5966796875,
0.17724609375,
0.5966796875
],
[
0.1748046875,
0.5966796875,
0.68701171875,
0.5966796875,
0.68701171875,
0.60986328125,
0.1748046875,
0.60986328125
],
[
0.31201171875,
0.62353515625,
0.438720703125,
0.625,
0.438720703125,
0.6396484375,
0.31201171875,
0.63818359375
],
[
0.443603515625,
0.62255859375,
0.68115234375,
0.62255859375,
0.68115234375,
0.64013671875,
0.443603515625,
0.64013671875
],
[
0.8037109375,
0.625,
0.826171875,
0.625,
0.826171875,
0.6396484375,
0.8037109375,
0.6396484375
],
[
0.416748046875,
0.63525390625,
0.44677734375,
0.63525390625,
0.44677734375,
0.642578125,
0.416748046875,
0.642578125
],
[
0.54150390625,
0.642578125,
0.580078125,
0.642578125,
0.580078125,
0.65576171875,
0.54150390625,
0.65576171875
],
[
0.1748046875,
0.66845703125,
0.82275390625,
0.66943359375,
0.82275390625,
0.6845703125,
0.1748046875,
0.68310546875
],
[
0.176513671875,
0.6845703125,
0.44677734375,
0.6845703125,
0.44677734375,
0.69775390625,
0.176513671875,
0.69775390625
],
[
0.27685546875,
0.71044921875,
0.386474609375,
0.7119140625,
0.386474609375,
0.7265625,
0.27685546875,
0.724609375
],
[
0.439453125,
0.7109375,
0.47216796875,
0.705078125,
0.476318359375,
0.71728515625,
0.443603515625,
0.7236328125
],
[
0.465576171875,
0.70556640625,
0.71826171875,
0.70556640625,
0.71826171875,
0.72802734375,
0.465576171875,
0.72802734375
],
[
0.38818359375,
0.7119140625,
0.42236328125,
0.7119140625,
0.42236328125,
0.7236328125,
0.38818359375,
0.7236328125
],
[
0.8037109375,
0.7109375,
0.82421875,
0.7109375,
0.82421875,
0.72607421875,
0.8037109375,
0.72607421875
],
[
0.36767578125,
0.7236328125,
0.449462890625,
0.7236328125,
0.449462890625,
0.73486328125,
0.36767578125,
0.73486328125
],
[
0.4697265625,
0.7265625,
0.5009765625,
0.7265625,
0.5009765625,
0.7392578125,
0.4697265625,
0.7392578125
],
[
0.57861328125,
0.72900390625,
0.61376953125,
0.72900390625,
0.61376953125,
0.74072265625,
0.57861328125,
0.74072265625
],
[
0.17724609375,
0.75048828125,
0.50341796875,
0.75048828125,
0.50341796875,
0.76123046875,
0.17724609375,
0.76123046875
],
[
0.17236328125,
0.78271484375,
0.461669921875,
0.7841796875,
0.461669921875,
0.798828125,
0.17236328125,
0.79736328125
],
[
0.1732177734375,
0.81494140625,
0.358642578125,
0.81494140625,
0.358642578125,
0.8291015625,
0.1732177734375,
0.8291015625
],
[
0.178955078125,
0.84326171875,
0.82177734375,
0.84326171875,
0.82177734375,
0.85400390625,
0.178955078125,
0.85400390625
],
[
0.17724609375,
0.85498046875,
0.82275390625,
0.85498046875,
0.82275390625,
0.86865234375,
0.17724609375,
0.86865234375
],
[
0.176513671875,
0.869140625,
0.82177734375,
0.869140625,
0.82177734375,
0.88330078125,
0.176513671875,
0.88330078125
],
[
0.1748046875,
0.88330078125,
0.8203125,
0.88330078125,
0.8203125,
0.896484375,
0.1748046875,
0.896484375
],
[
0.176513671875,
0.89794921875,
0.82275390625,
0.89794921875,
0.82275390625,
0.9111328125,
0.176513671875,
0.9111328125
],
[
0.176513671875,
0.9091796875,
0.529296875,
0.91015625,
0.529296875,
0.9248046875,
0.176513671875,
0.923828125
],
[
0.49267578125,
0.94873046875,
0.50634765625,
0.94873046875,
0.50634765625,
0.96044921875,
0.49267578125,
0.96044921875
]
] |
[
"Published as a conference paper at ICLR 2021",
"3.4AUDIO-VISUAL SPATIO-TEMPORAL ATTENTION",
"这项工作的一个重要方面是将音频和视觉信息结合起来，以推断每个分离的源与视频中相",
"关对象之间的对应关系。这反过来将用于识别哪些源在屏幕上可见。为此，我们采用视听时",
"空关注方案，让网络关注每个分离源的视觉嵌入的局部特征。在这种机制中，我们使用音频",
"嵌人作为源m的查询输入，键和值输人由时空视频嵌人zvl给出。结果，与第m个源",
"相对应的输出时空嵌人的扁平版本为zay=attend(zm，z,zvl)",
"3.5ON-SCREENCLASSIFIER",
"为了推断每个分离的源的视觉存在，我们连接了全局视频嵌人zvg，每个源的全局音频嵌",
"人zm以及相应的本地时空视听嵌人zav。级联向量通过具有逻辑激活的密集层fc馈送",
"ym = logistic (fc ([zvg, za, zavl))",
"3.6 SEPARATIONLOSS",
"我们使用MixIT分离损失（Wisdometal.,2020），它优化了M估计源=Ms（ai+c2）到两",
"个参考混合物C1，2的分配，如下所示：。由于A的限制，每个源sm只能分配给一个参",
"考混合物。估计信号tERT和目标信号tERT的SNR损耗定义为：",
"CsNR(t,t)=10log10(llt-t2+10-3t2)",
"(2)",
"3.7CLASSIFICATIONLOSS",
"为了训练屏幕上的分类器，我们考虑以下分类损失。这些损失使用二进制标签9m，其中给",
"出了监督示例，在无监督情况下，每个源m使用ym=A1.m，其中A*是通过（?？）中的",
"最小化找到的最佳混合矩阵。我们还使用符号R=mlym=1，mE[1,...,M))来表示",
"组正标签",
"精确二进制交叉：",
"M",
"",
"(- 9m log (9m) + (ym - 1) log (1 - 9m))",
"(3)",
"m=",
"Lexact (y,9) =",
"多实例交叉摘：由于分配给屏幕上混合的一些单独的源不在屏幕上，因此多实例（MI）(Maron",
"&Lozano-Pérez，1998）损失，可最大程度地减少正标签集R可能更健壮：",
"CMI (y,9)= min",
"-log(9m)-Z log(1-9m))",
"(4)",
"ERL",
"m'ER",
"活动组合交叉摘：MI损失的替代方法，活动组合（AC)，对应于标签s的所有设置8>1（R）",
"上的最小损失。t。至少一个标签是1：",
"CAc (y,9) =",
"",
"Z log (9m) -Z log(1 - 9m))",
"min",
"(5)",
"SEO(R)",
"mES",
"m'ts",
"，其中80>1（R）表示标签为1的索引的幕集合",
"4EXPERIMENTALFRAMEWORK",
"4.1DATAPREPARATION",
"为了训练我们的开放域系统的真实视听记录环境，我们使用YahooFlickrCreativeCommons1oo",
"百万数据集（YFCC100m）（Thomeeetal.，2016）。该数据集来自各种长度的200,000视频",
"（2,500总小时数），涵盖了各种语义声音类别。通过在视频上传器上进行拆分，我们选择",
"1,600视频进行训练，并使用剩余的视频进行验证和测试。我们以1秒的跳跃大小提取5秒",
"的剪辑，从而产生大约7.2百万个剪辑。剪辑由以16kHz采样的5秒音频波形和5个视频顿",
"(）组成，其中每个顿是128×128×3RGB图像",
"5"
] |
{
"arxiv_id": "2011.01143",
"language": "zh",
"page_idx": 4
}
|
||
2206.14649-zh-0
|
2206.14649-zh-0.jpg
|
Cooperative Retriever and Ranker in Deep Recommenders
Xu Huang $^{1}$, Defu Lian $^{1,2,3}$, Jin Chen $^{4},\text{Zheng Liu}^{5},\text{Xing Xie}^{5}$, Enhong Chen$^{1,2,3}$
$^{1}$School of Computer Science, University of Science and Technology of China, Hefei, China
$^2$School of Data Science, University of Science and Technology of China, Hefei, China
$^{3}$ State Key Laboratory of Cognitive Intelligence, Hefei, China
$^{4}$University of Electronic Science and Technology of China, Chengdu, China
$^{5}$ Microsoft Research Asia, Beijing, China
xuhuangcs@mail.ustc.edu.cn,{liandefu,cheneh}@ustc.edu.cn,chenjin@std.uestc.edu.cn,{zhengliu,xingxie{@}microsoft.com}
# $\mathbf{ABSTRACT}$
and the existing joint training methods with notable advantages.
Our code will be open-sourced to facilitate future research
$\csc\cos$coNCEPTS
$\cdot$Information systems$\to$Information retrieval.
$\mathbf{KEYWORDS}$
Recommender Systems, Retriever and Ranker, Cooperative Train-
ing, Knowledge Distillation, Negative Sampling
ACM Reference Format:
Xu Huang $^1$,Defu Lian $^1,2,3$,Jin Chen $^4,\text{Zheng Liu}^5$,Xing Xie$^5$,Enhong Chen $^1,2,3.2023.$ Cooperative Retriever and Ranker in Deep Recommenders . In Proceedings of the ACM Web Conference 2023 (WWW'23), April 30-May $4,2023,Austin,TX,USA.$ ACM, New York, NY, USA, 11 pages. https://doi $\operatorname{org}/10.1145/3543507.3583422$
Deep recommender systems (DRS) are intensively applied in modern web services. To deal with the massive web contents, DRS employs a two-stage workflow: retrieval and ranking, to generate its recommendation results. The retriever aims to select a small set of relevant candidates from the entire items with high efficiency; while the ranker, usually more precise but time-consuming, is supposed to further refine the best items from the retrieved candidates. Traditionally, the two components are trained either independently or within a simple cascading pipeline, which is prone to poor collaboration effect. Though some latest works suggested to train retriever and ranker jointly, there still exist many severe limitations: item distribution shift between training and inference, false negative, and misalignment of ranking order. As such, it remains to explore effective collaborations between retriever and ranker.
In this work, we present a novel framework for the joint training of retriever and ranker, named CoRR ( Co operative R etriever and R anker). With CoRR, the retriever is improved by deriving highquality training signals from the ranker, while the ranker is improved by learning to discriminate hard negatives sampled by the retriever. We introduce two critical techniques. Firstly, we develop an adaptive and scalable sampler based on the retriever, to generate hard negative samples for the ranker's training. Compared with the widely-used exact top-k sampling, our method effectively alleviates the issues of false negative and item distribution shift, and thus improves the ranker's discriminability. Secondly, we propose a novel asymptotic-unbiased estimation of KL divergence, which serves as the objective for knowledge distillation. The new objective can be efficiently optimized with commonly-used optimizers. More importantly, it leads to better alignment of ranking order between retriever and ranker, which helps to improve the retrieval quality. We conduct comprehensive experiments over four largescale datasets, where CoRR outperforms both conventional DRS
# 1 INTRODUCTION
推荐系统在现代 Web 服务(如电子商务和在线广告)中起着重要作用，因为它通过根据用户自己的兴趣向用户推荐个性化商品，在很大程度上缓解了信息过载问题。由于深度学习的显着进步，深度推荐系统(DRS)在实践中变得越来越流行[7,9,10,15,21,31,46,49]。鉴于在线项目的庞大规模，深度推荐系统需要两个阶段的工作流程：检索和排名[7]。特别是， 检索器的目标是在特定上下文下选择一小部分候选项(例如。g。、用户配置文件和历史交互)从整个项目中高效。通常，检索器模型学习将上下文和项目表示为密集嵌人，以便可以基于嵌人相似性(如内积或余弦相似性)有效地估计用户对项目的偏好。相反，排名器用于从检索结果中优化最喜欢的项目。为了获得最佳精度，它通常利用高度表达但耗时的网络，特别是那些在上下文和项目之间建立深度交互的网络(例如。g。DIN[52]、DIEN[51] 和 DeepFM[12] )。
## 1.1 Existing Problems
$^{*}$Defu Lian is the Corresponding Author
尽管检索器和排名器具有协作性质，但两个模型的典型训练工作流程缺乏有效的协作，这严重损害了整体推荐质量。在许多情况下，这两个模型是独立训练的，并直接应用于推荐系统$\left[4,32,51-53\right]$。在其他时候，可以根据检索结果$\overline{\mathrm{E}}[11,17]$ 连续训练排名器；然而，检索器仍然是独立训练的$\mathbb{E}[7]$。由于以下原因，此类培训工作流程较差
$\bullet$ The independent training of the retriever only leverages the historical user-item interactions, which can be limited in reality As a result, it may suffer from the sparsity of training data, which severely restricts the retrieval quality. For another thing, the retriever is likely to generate candidate items that are not favored
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.
$WWW^{\prime}23,April$ 30-May 4, 2023, Austin, TX, USA
$\odot$ 2023 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-9416-1/23/04...815.00
https://doi.org/10.1145/3543507.3583422
|
[
[
0.12744140625,
0.10479736328125,
0.875,
0.10479736328125,
0.875,
0.12249755859375,
0.12744140625,
0.12249755859375
],
[
0.1519775390625,
0.1351318359375,
0.8447265625,
0.1331787109375,
0.8447265625,
0.1572265625,
0.1519775390625,
0.1590576171875
],
[
0.1953125,
0.1572265625,
0.8056640625,
0.1572265625,
0.8056640625,
0.17041015625,
0.1953125,
0.17041015625
],
[
0.2115478515625,
0.17041015625,
0.78759765625,
0.1717529296875,
0.78759765625,
0.186279296875,
0.2115478515625,
0.1849365234375
],
[
0.293212890625,
0.1849365234375,
0.7099609375,
0.1875,
0.7099609375,
0.2020263671875,
0.293212890625,
0.199462890625
],
[
0.2418212890625,
0.1988525390625,
0.75732421875,
0.199462890625,
0.75732421875,
0.2171630859375,
0.2418212890625,
0.216552734375
],
[
0.361083984375,
0.21533203125,
0.6396484375,
0.2171630859375,
0.6396484375,
0.231689453125,
0.361083984375,
0.2298583984375
],
[
0.10540771484375,
0.23291015625,
0.91650390625,
0.23291015625,
0.91650390625,
0.246826171875,
0.10540771484375,
0.246826171875
],
[
0.0882568359375,
0.255615234375,
0.1829833984375,
0.255615234375,
0.1829833984375,
0.270751953125,
0.0882568359375,
0.270751953125
],
[
0.51708984375,
0.25830078125,
0.9111328125,
0.25830078125,
0.9111328125,
0.271484375,
0.51708984375,
0.271484375
],
[
0.0849609375,
0.27587890625,
0.483642578125,
0.27587890625,
0.483642578125,
0.2890625,
0.0849609375,
0.2890625
],
[
0.5146484375,
0.270751953125,
0.87158203125,
0.270751953125,
0.87158203125,
0.2841796875,
0.5146484375,
0.2841796875
],
[
0.0849609375,
0.2890625,
0.483642578125,
0.2890625,
0.483642578125,
0.302978515625,
0.0849609375,
0.302978515625
],
[
0.08416748046875,
0.3037109375,
0.48193359375,
0.302978515625,
0.48193359375,
0.31640625,
0.08416748046875,
0.317626953125
],
[
0.51708984375,
0.2998046875,
0.65283203125,
0.298583984375,
0.65283203125,
0.313232421875,
0.51708984375,
0.314453125
],
[
0.0849609375,
0.317626953125,
0.48193359375,
0.317626953125,
0.48193359375,
0.330810546875,
0.0849609375,
0.330810546875
],
[
0.51708984375,
0.318115234375,
0.833984375,
0.318115234375,
0.833984375,
0.33203125,
0.51708984375,
0.33203125
],
[
0.0849609375,
0.330810546875,
0.48046875,
0.330810546875,
0.48046875,
0.3447265625,
0.0849609375,
0.3447265625
],
[
0.0849609375,
0.343505859375,
0.48046875,
0.3447265625,
0.48046875,
0.35791015625,
0.0849609375,
0.356689453125
],
[
0.5185546875,
0.3466796875,
0.62353515625,
0.3466796875,
0.62353515625,
0.361083984375,
0.5185546875,
0.361083984375
],
[
0.08660888671875,
0.35986328125,
0.48193359375,
0.35986328125,
0.48193359375,
0.37060546875,
0.08660888671875,
0.37060546875
],
[
0.51708984375,
0.365478515625,
0.9140625,
0.365478515625,
0.9140625,
0.37890625,
0.51708984375,
0.37890625
],
[
0.08660888671875,
0.373779296875,
0.48193359375,
0.373779296875,
0.48193359375,
0.3837890625,
0.08660888671875,
0.3837890625
],
[
0.51708984375,
0.380126953125,
0.8056640625,
0.380126953125,
0.8056640625,
0.393310546875,
0.51708984375,
0.393310546875
],
[
0.0849609375,
0.386474609375,
0.48046875,
0.386474609375,
0.48046875,
0.399658203125,
0.0849609375,
0.399658203125
],
[
0.08416748046875,
0.3984375,
0.48193359375,
0.399658203125,
0.48193359375,
0.4140625,
0.08416748046875,
0.41357421875
],
[
0.5185546875,
0.401611328125,
0.6640625,
0.401611328125,
0.6640625,
0.415283203125,
0.5185546875,
0.415283203125
],
[
0.08416748046875,
0.41357421875,
0.48193359375,
0.4140625,
0.48193359375,
0.427978515625,
0.08416748046875,
0.4267578125
],
[
0.5146484375,
0.412353515625,
0.9140625,
0.41357421875,
0.9140625,
0.427978515625,
0.5146484375,
0.4267578125
],
[
0.0849609375,
0.4267578125,
0.48046875,
0.427978515625,
0.48046875,
0.44140625,
0.0849609375,
0.43994140625
],
[
0.515625,
0.4248046875,
0.91259765625,
0.425537109375,
0.91259765625,
0.43994140625,
0.515625,
0.439453125
],
[
0.08660888671875,
0.443115234375,
0.48046875,
0.443115234375,
0.48046875,
0.453857421875,
0.08660888671875,
0.453857421875
],
[
0.51953125,
0.439453125,
0.9140625,
0.439453125,
0.9140625,
0.45263671875,
0.51953125,
0.45263671875
],
[
0.0849609375,
0.455810546875,
0.483642578125,
0.455810546875,
0.483642578125,
0.4697265625,
0.0849609375,
0.4697265625
],
[
0.51708984375,
0.45263671875,
0.9150390625,
0.45263671875,
0.9150390625,
0.466552734375,
0.51708984375,
0.466552734375
],
[
0.515625,
0.46533203125,
0.67333984375,
0.464111328125,
0.67333984375,
0.477783203125,
0.515625,
0.479248046875
],
[
0.0849609375,
0.4716796875,
0.1282958984375,
0.4716796875,
0.1282958984375,
0.481689453125,
0.0849609375,
0.481689453125
],
[
0.10296630859375,
0.48486328125,
0.4794921875,
0.48486328125,
0.4794921875,
0.49560546875,
0.10296630859375,
0.49560546875
],
[
0.08416748046875,
0.49560546875,
0.48193359375,
0.496826171875,
0.48193359375,
0.51123046875,
0.08416748046875,
0.51025390625
],
[
0.5439453125,
0.50244140625,
0.54833984375,
0.50244140625,
0.54833984375,
0.5087890625,
0.5439453125,
0.5087890625
],
[
0.54736328125,
0.5,
0.69287109375,
0.5,
0.69287109375,
0.51318359375,
0.54736328125,
0.51318359375
],
[
0.5361328125,
0.50390625,
0.54345703125,
0.50390625,
0.54345703125,
0.5087890625,
0.5361328125,
0.5087890625
],
[
0.52294921875,
0.50390625,
0.52685546875,
0.50390625,
0.52685546875,
0.5087890625,
0.52294921875,
0.5087890625
],
[
0.0849609375,
0.51025390625,
0.483642578125,
0.51025390625,
0.483642578125,
0.5234375,
0.0849609375,
0.5234375
],
[
0.5185546875,
0.517578125,
0.91259765625,
0.517578125,
0.91259765625,
0.53076171875,
0.5185546875,
0.53076171875
],
[
0.0849609375,
0.5244140625,
0.483642578125,
0.5244140625,
0.483642578125,
0.53857421875,
0.0849609375,
0.53857421875
],
[
0.51953125,
0.53076171875,
0.91259765625,
0.53076171875,
0.91259765625,
0.54443359375,
0.51953125,
0.54443359375
],
[
0.0849609375,
0.537109375,
0.48193359375,
0.537109375,
0.48193359375,
0.55029296875,
0.0849609375,
0.55029296875
],
[
0.521484375,
0.54736328125,
0.9111328125,
0.54736328125,
0.9111328125,
0.55810546875,
0.521484375,
0.55810546875
],
[
0.08660888671875,
0.5537109375,
0.48046875,
0.5537109375,
0.48046875,
0.564453125,
0.08660888671875,
0.564453125
],
[
0.51953125,
0.5595703125,
0.91259765625,
0.5595703125,
0.91259765625,
0.57275390625,
0.51953125,
0.57275390625
],
[
0.08660888671875,
0.5673828125,
0.48046875,
0.5673828125,
0.48046875,
0.57763671875,
0.08660888671875,
0.57763671875
],
[
0.51953125,
0.57275390625,
0.91259765625,
0.57275390625,
0.91259765625,
0.5859375,
0.51953125,
0.5859375
],
[
0.0849609375,
0.580078125,
0.48193359375,
0.580078125,
0.48193359375,
0.59326171875,
0.0849609375,
0.59326171875
],
[
0.0849609375,
0.59326171875,
0.48046875,
0.59326171875,
0.48046875,
0.60693359375,
0.0849609375,
0.60693359375
],
[
0.521484375,
0.58837890625,
0.9111328125,
0.58837890625,
0.9111328125,
0.5986328125,
0.521484375,
0.5986328125
],
[
0.51953125,
0.60107421875,
0.91259765625,
0.60107421875,
0.91259765625,
0.6142578125,
0.51953125,
0.6142578125
],
[
0.08660888671875,
0.609375,
0.48046875,
0.609375,
0.48046875,
0.619140625,
0.08660888671875,
0.619140625
],
[
0.08416748046875,
0.619140625,
0.48193359375,
0.6220703125,
0.48193359375,
0.63623046875,
0.08416748046875,
0.6337890625
],
[
0.51708984375,
0.6142578125,
0.9140625,
0.61279296875,
0.9140625,
0.62744140625,
0.51708984375,
0.62890625
],
[
0.0849609375,
0.63525390625,
0.48193359375,
0.63525390625,
0.48193359375,
0.6484375,
0.0849609375,
0.6484375
],
[
0.51953125,
0.6298828125,
0.9111328125,
0.6298828125,
0.9111328125,
0.64013671875,
0.51953125,
0.64013671875
],
[
0.08416748046875,
0.64794921875,
0.48193359375,
0.6494140625,
0.48193359375,
0.66357421875,
0.08416748046875,
0.6611328125
],
[
0.51953125,
0.642578125,
0.9111328125,
0.642578125,
0.9111328125,
0.65576171875,
0.51953125,
0.65576171875
],
[
0.0849609375,
0.66357421875,
0.4794921875,
0.66357421875,
0.4794921875,
0.6767578125,
0.0849609375,
0.6767578125
],
[
0.521484375,
0.65771484375,
0.9111328125,
0.65771484375,
0.9111328125,
0.66845703125,
0.521484375,
0.66845703125
],
[
0.0849609375,
0.6767578125,
0.48046875,
0.6767578125,
0.48046875,
0.68994140625,
0.0849609375,
0.68994140625
],
[
0.521484375,
0.671875,
0.91259765625,
0.671875,
0.91259765625,
0.681640625,
0.521484375,
0.681640625
],
[
0.5185546875,
0.68310546875,
0.91259765625,
0.6845703125,
0.91259765625,
0.69873046875,
0.5185546875,
0.69775390625
],
[
0.08416748046875,
0.68994140625,
0.48193359375,
0.689453125,
0.48193359375,
0.70263671875,
0.08416748046875,
0.7041015625
],
[
0.515625,
0.6962890625,
0.765625,
0.69775390625,
0.765625,
0.7119140625,
0.515625,
0.7109375
],
[
0.08660888671875,
0.70556640625,
0.48046875,
0.70556640625,
0.48046875,
0.71630859375,
0.08660888671875,
0.71630859375
],
[
0.0849609375,
0.71728515625,
0.48046875,
0.71728515625,
0.48046875,
0.73095703125,
0.0849609375,
0.73095703125
],
[
0.521484375,
0.7265625,
0.54833984375,
0.7265625,
0.54833984375,
0.73779296875,
0.521484375,
0.73779296875
],
[
0.55810546875,
0.72607421875,
0.71142578125,
0.72607421875,
0.71142578125,
0.7392578125,
0.55810546875,
0.7392578125
],
[
0.51953125,
0.74365234375,
0.91259765625,
0.74365234375,
0.91259765625,
0.7568359375,
0.51953125,
0.7568359375
],
[
0.0849609375,
0.75,
0.275390625,
0.75,
0.275390625,
0.76318359375,
0.0849609375,
0.76318359375
],
[
0.51953125,
0.7568359375,
0.9111328125,
0.7568359375,
0.9111328125,
0.77099609375,
0.51953125,
0.77099609375
],
[
0.51953125,
0.771484375,
0.91259765625,
0.771484375,
0.91259765625,
0.78515625,
0.51953125,
0.78515625
],
[
0.08660888671875,
0.7822265625,
0.48046875,
0.7822265625,
0.48046875,
0.79248046875,
0.08660888671875,
0.79248046875
],
[
0.51708984375,
0.78515625,
0.91259765625,
0.78515625,
0.91259765625,
0.798828125,
0.51708984375,
0.798828125
],
[
0.0882568359375,
0.79150390625,
0.4794921875,
0.79150390625,
0.4794921875,
0.8017578125,
0.0882568359375,
0.8017578125
],
[
0.08660888671875,
0.80322265625,
0.48046875,
0.80322265625,
0.48046875,
0.81298828125,
0.08660888671875,
0.81298828125
],
[
0.5185546875,
0.7998046875,
0.91259765625,
0.7998046875,
0.91259765625,
0.81298828125,
0.5185546875,
0.81298828125
],
[
0.0849609375,
0.81103515625,
0.48193359375,
0.81103515625,
0.48193359375,
0.82470703125,
0.0849609375,
0.82470703125
],
[
0.5185546875,
0.81298828125,
0.740234375,
0.81298828125,
0.740234375,
0.8271484375,
0.5185546875,
0.8271484375
],
[
0.08416748046875,
0.8212890625,
0.48193359375,
0.82275390625,
0.48193359375,
0.833984375,
0.08416748046875,
0.83349609375
],
[
0.098876953125,
0.83349609375,
0.48046875,
0.83349609375,
0.48046875,
0.84326171875,
0.098876953125,
0.84326171875
],
[
0.53369140625,
0.8291015625,
0.9111328125,
0.8291015625,
0.9111328125,
0.8388671875,
0.53369140625,
0.8388671875
],
[
0.1258544921875,
0.84326171875,
0.43310546875,
0.84326171875,
0.43310546875,
0.85400390625,
0.1258544921875,
0.85400390625
],
[
0.51708984375,
0.84130859375,
0.9111328125,
0.84130859375,
0.9111328125,
0.85498046875,
0.51708984375,
0.85498046875
],
[
0.08660888671875,
0.8515625,
0.310546875,
0.85302734375,
0.310546875,
0.86279296875,
0.08660888671875,
0.8623046875
],
[
0.0882568359375,
0.8447265625,
0.1298828125,
0.8447265625,
0.1298828125,
0.8515625,
0.0882568359375,
0.8515625
],
[
0.51708984375,
0.85498046875,
0.9140625,
0.85498046875,
0.9140625,
0.86865234375,
0.51708984375,
0.86865234375
],
[
0.0882568359375,
0.8642578125,
0.47216796875,
0.86279296875,
0.47216796875,
0.875,
0.0882568359375,
0.87548828125
],
[
0.0882568359375,
0.87548828125,
0.2900390625,
0.87548828125,
0.2900390625,
0.88330078125,
0.0882568359375,
0.88330078125
],
[
0.5185546875,
0.87060546875,
0.91259765625,
0.87060546875,
0.91259765625,
0.880859375,
0.5185546875,
0.880859375
],
[
0.0849609375,
0.88525390625,
0.273681640625,
0.88525390625,
0.273681640625,
0.89599609375,
0.0849609375,
0.89599609375
],
[
0.515625,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.515625,
0.896484375
]
] |
[
"CooperativeRetrieverandRankerinDeepRecommenders",
"Xu Huang ', Defu Lian 1.2.3 , Jin Chen 4 , Zheng Liu 5, Xing Xie 5 , Enhong Chen 123 *",
"1 School of Computer Science, University of Scienceand Technology of China, Hefei,China",
"2 School of Data Science, University of Science and Technology of China, Hefei, China",
"3 State Key Laboratory of Cognitive Intelligence, Hefei, China",
"4 University of Electronic Science and Technology of China, Chengdu, China",
"5 Microsoft Research Asia, Beijing, China",
"xuhuangcs@mail.ustc.edu.cn,fliandefu,chenehj@ustc.edu.cn,chenjin@std.uestc.edu.cn,(zhengliu,xingxiej@microsoft.com",
"ABSTRACT",
"and the existing joint training methods with notable advantages.",
"Deep recommender systems (DRS)are intensively applied in mod-",
" Our code will be open-sourced to facilitate future research.",
"ern web services. To deal with the massive web contents, DRS em-",
"ploys a two-stage workflow:retrieval and ranking,to generate its",
"CCS CONCEPTS",
"recommendation results. The retriever aims to select a small set",
". Information systemsInformation retrieval.",
"of relevant candidates from the entire items with high efficiency:",
"while theranker,usuallymore precise but time-consuming, is sup",
"KEYWORDS",
"posed to furtherrefinethebest items from theretrieved candi-",
"Recommender Systems, Retriever and Ranker, Cooperative Train-",
"dates.Traditionally,the two components are trained either inde-",
"ing,Knowledge Distillation, Negative Sampling",
"pendently or within a simple cascading pipeline, which is prone",
"to poor collaboration effect. Though some latest works suggested",
"ACM Reference Format:",
"to train retriever and ranker jointly, there still exist many severe",
"Xu Huang ', Defu Lian 1.23 , Jin Chen 4, Zheng Liu 5, Xing Xie 5, Enhong",
"limitations: item distribution shift between training and inference,",
"Chen 12.3.2023. Cooperative Retriever and Ranker in Deep Recommenders",
"false negative, and misalignment of ranking order. As such, it re",
". In Proceedings of the ACM Web Conference 2023 (WWW '23), April 30-May",
"mains to explore effective collaborations between retriever and",
"4, 2023, Austin, TX, USA. ACM, New York, NY, USA, 11 pages. https://doi.",
"0rg/10.1145/3543507.3583422",
"ranker",
"In this work,wepresentanovel framework for the joint training",
"of retriever and ranker, named CoRR (Co operative R etriever and",
"",
"INTRODUCTION",
"",
"-",
"R anker). With CoRR, the retriever is improved by deriving high-",
"推荐系统在现代Web服务（如电子商务和在线广告）中起着",
"quality training signals from the ranker, while the ranker is im-",
"重要作用，因为它通过根据用户自已的兴趣向用户推荐个性",
"proved by learning to discriminate hard negatives sampled by the",
"化商品，在很大程度上缓解了信息过载问题。由于深度学习",
"retriever.We introduce two critical techniques.Firstly,we develop",
"的显着进步，深度推荐系统（DRS）在实践中变得越来越流行",
"anadaptiveand scalablesampler based on theretriever,togener-",
"[7,9,10,15,21,31,46,49]。鉴于在线项目的庞大规模，深度推",
"ate hard negative samples for the ranker's training. Compared with",
"the widely-used exact top-k sampling, our method effectively alle",
"荐系统需要两个阶段的工作流程：检索和排名[7]。特别是",
"检索器的目标是在特定上下文下选择一小部分候选项（例如。",
"viates theissues offalse negativeand item distribution shift,and",
"thus improves the ranker's discriminability. Secondly, we propose",
"g。、用户配置文件和历史交互)从整个项目中高效。通常，检",
"a novel asymptotic-unbiased estimation of KL divergence, which",
"索器模型学习将上下文和项目表示为密集嵌入，以便可以基",
"serves as the objective for knowledge distillation.The new objec-",
"于嵌入相似性（如内积或余弦相似性）有效地估计用户对项目",
"tive can be efficiently optimized with commonly-used optimizers.",
"的偏好。相反，排名器用于从检索结果中优化最喜欢的项目。",
"More importantly, it leads to better alignment of ranking order be",
"为了获得最佳精度，它通常利用高度表达但耗时的网络，特",
"别是那些在上下文和项目之间建立深度交互的网络（例如。g。、",
"tween retriever and ranker, which helps to improve the retrieval",
"DIN [52] 、DIEN [51] 和 DeepFM [12] )。",
"quality.We conduct comprehensive experiments over four large",
"scale datasets, where CoRR outperforms both conventional DRS",
"1.1",
"ExistingProblems",
"尽管检索器和排名器具有协作性质，但两个模型的典型训练",
"'Defu Lian is the Corresponding Author",
"工作流程缺乏有效的协作，这严重损害了整体推荐质量。在许",
"多情况下，这两个模型是独立训练的，并直接应用于推荐系",
"Permission to make digital or hard copies of all or part of this work for personal or",
"统[4,32,51-53]。在其他时候，可以根据检索结果E[11,17]连",
"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",
"续训练排名器；然而，检索器仍然是独立训练的国[7]。由于",
"tion on the first page. Copyrights for components of this work owned by others than",
"以下原因，此类培训工作流程较差",
"the author(s) must be honored. Abstracting with credit is permitted. To copy other-",
"se, or republish, to post on servers or to redistribute to lists,requires prior specific",
".The independent training of theretriever only leverages the",
"ion and/or a fee.Request permissions from permissions@acm.org.",
"historical user-item interactions, which can be limited in reality",
"www23.April30-May4,2023,Austin,Tx,USA",
"permissic",
"As a result, it may suffer from the sparsity of training data, which",
" 2023 Copyright held by the owner/author(s).Publication rights licensed to ACM",
"ACMISBN978-1-4503-9416-1/23/04..$15.00",
"severelyrestricts the retrieval quality.For another thing, the re-",
"https://doi.org/10.1145/3543507.3583422",
"triever is likely to generate candidate items that are not favored"
] |
{
"arxiv_id": "2206.14649",
"language": "zh",
"page_idx": 0
}
|
||
2206.14649-zh-1
|
2206.14649-zh-1.jpg
|
WWW'23, April 30-May 4, 2023, Austin, TX, USA
by the ranker; thus, it may harm the downstream performance as
## well.
$\bullet$ The independently-trained ranker is learned with randomly or heuristically collected training samples. Such training samples can be too easy to be distinguished, making the ranker converge to a limited discriminative capacity. Besides, the item distribution will also be highly differentiated between the training and inference stages; as a result, the ranker may not effectively recognize the high-quality candidates generated by the retriever.
Recent studies on multi-stage ranking models are closely related to the problem of retriever-ranker collaboration; e.g., in [34] ,a two-pass training workflow is proposed. However, the two-pass workflow is still limited from several critical perspectives. In the forward pass, the rankers are trained by the retrieval results at the exact top-k cutoff, which is prone to false-negatives. Besides, when the retrieval cutoffs are changed during the inference stage, the rankers will face highly shifted item distributions from the training stage, which may severely harm their prediction accuracy. In the backward pass, the retrievers are trained to preserve the consistency of absolute ranking scores and to distinguish ranking results from retrieval ones; while for the sake of high-quality retrieval, it is the consistency of relative ranking order that really matters. In all, it remains a challenging problem to explore more effective collaboration mechanisms between the retriever and ranker.
## 1.2 Our Solution
在这项工作中，我们提出了一种新颖的框架，用于寻回者和排名者的合作训练，a。k。a。CoRR(合作 Retriever 和 Ranker)。在我们的框架中，检索器和排名器在统一的工作流程中同时训练，其中两个模型可以相互加强
$\bullet$ Training of retriever. On one hand, the retriever is learned
from both user-item interactions via sampled softmax and the ranker's
predictions via knowledge distillation [16]. In a specific context, a few items are sampled firstly. Then, the ranker is required to predict the fne-grained preferences towards the sampled items. Rather than preserving the absolute preference scores, the retriever is required to generate the same ranking order for the sampled items as the ranker. To realize this goal, the KL-divergence is minimized for the softmax-normalized predictions between the retriever and ranker [16] .In this case, user's preferred items, whether interacted or not, will probably get highly-rated by the ranker, while real non-interested items get lower-rated (thanks to the highlyprecise ranker). As a result, such items will virtually become labeled samples , which substantially augment the training data
$\bullet$ Training of ranker. The ranker is trained by sampled soft- $\max\left[1\right]$on top of the hard negative items sampled by the retriever. Particularly, instead of working with the“easy negatives”which are randomly or heuristically sampled from the whole item set [12, 36,52] , the ranker is iteratively trained to discriminate the true positive from the increasingly harder negatives as the retriever improves. Therefore, it prevents the ranker from converging too early to a limited discriminative capability. Besides, unlike the widelyused exact top-k sampling, we collect informative negative samples from the entire itemset based on the retriever; by doing so, it alleviates the false negative issue and closes the gap between the training and inference stages .
Xu Huang $^1,\text{Defu Lian}^{1,2,3},\text{Jin Chen}^{4},\text{Zheng Liu}^{5},\text{Xing Xie}^{5},\text{Enhong Chen}^{1,2,3}$
It's worth noting that the realization of the above training framework is non-trivial. Particularly, both retriever and ranker need to learn from the sampled items; however, the sampling operation on the retriever can be inefficient and biased. To overcome such challenges, a couple of technical designs are introduced. 首先，我们知道在处理大型项目集时，直接从检索器进行采样可能非常耗时，因此我们开发了一种可扩展且自适应的采样策略，在该策略中，可以根据项目大小在次线性时间内对检索器偏爱的项目进行有效采样。其次，对抽样项目的直接知识提炼是有偏见的，容易出现劣质表现。为了缓解这个问题，我们提出了一种新的 KL 散度的渐近无偏估计，以补偿项目抽样引起的偏差。在此操作之上，可以更好地在排名器和检索器之间对齐项目的排名顺序
我们对四个基准数据集进行了全面的实验研究。根据实验结果，与现有的训练方法相比，CoRR 可以大大提高整体推荐质量。更详细的分析进一步验证了 CoRR 对寻回者和排名者的独立有效性，以及其作为模型不可知训练框架的有效性。我们工作的 contributions 总结如下
$\bullet$ We present a novel training framework for deep recommender systems, where the retriever and ranker can be mutually reinforced for the effective cooperation.
$\bullet$ Two critical techniques are introduced for the optimized conduct of CoRR: 1) the scalable and adaptive sampling strategy, which enables the efficient sampling from the retriever; 2) the asymptotic-unbiased estimation of KL divergence, as the objective of knowledge distillation, which better aligns the ranking order of items between the ranker and retriever, and contributes to the retrieval of high-quality items
$\bullet$ We perform comprehensive experimental studies on four benchmark datasets, whose results verify CoRR's advantage against the existing training algorithms, and its standalone effectiveness to the retriever and ranker. 2 $\textbf{RELATED WORK}$
本文研究了深度推荐系统中排名器和检索器之间的合作，以提高多级系统的推荐质量。我们首先回顾了密切相关的级联排序技术，然后介绍了负采样和知识提炼。
## 2.1 Cascade Ranking
关于多阶段级联排名的先前工作通常为每个阶段分配不同的排名，以在效率和有效性之间共同实现所需的权衡$\overline{\mathbb{E}}[43]$。它们在对每个排名者的成本进行建模时彼此不同[6,47,48]。最近的工作转向通过梯度下降[11]直接优化级联排序模型作为一个整体，或者用级联排序模型识别一些坏的情况来增加训练数据[8] 。观察这些级联排名系统不考虑排名者之间的合作， 工作$\mathbb{E}[34]$ 建议通过让级联排名者在级联系统中为彼此提供监督信号来共同优化它们。然而，如上所述，从几个关键的角度来看，这项工作仍然有限。这项工作旨在探索级联排名者之间更好的协作机制
## 2.2 $\textbf{Negative Sampling in RecSys}$
许多方法从静态分布中对负面项目进行采样，例如均匀分布$\mathbb{E}$ [14,36] 和基于流行度的分布$\overline{\mathbb{E}}[35]$ 。为了适应推荐模型，已经提出了许多先进的抽样方法[2,5,18,30,35,39,45,50]。例如，AOBPR$\boxed{\mathbb{F}}[35]$将自适应采样转换为以随机抽取的等级搜索项目。CML[18] 和 WARP [45] 通过首先从均匀分布中抽取候
|
[
[
0.0849609375,
0.0738525390625,
0.318603515625,
0.07452392578125,
0.318603515625,
0.08837890625,
0.0849609375,
0.08709716796875
],
[
0.5087890625,
0.07452392578125,
0.91015625,
0.07452392578125,
0.91015625,
0.08837890625,
0.5087890625,
0.08837890625
],
[
0.0849609375,
0.10797119140625,
0.48193359375,
0.10797119140625,
0.48193359375,
0.121826171875,
0.0849609375,
0.121826171875
],
[
0.53173828125,
0.10797119140625,
0.9140625,
0.10797119140625,
0.9140625,
0.121826171875,
0.53173828125,
0.121826171875
],
[
0.08660888671875,
0.12249755859375,
0.1168212890625,
0.12249755859375,
0.1168212890625,
0.1337890625,
0.08660888671875,
0.1337890625
],
[
0.51708984375,
0.121826171875,
0.9140625,
0.121826171875,
0.9140625,
0.1351318359375,
0.51708984375,
0.1351318359375
],
[
0.101318359375,
0.1351318359375,
0.48046875,
0.1351318359375,
0.48046875,
0.1483154296875,
0.101318359375,
0.1483154296875
],
[
0.5185546875,
0.1363525390625,
0.91259765625,
0.1363525390625,
0.91259765625,
0.146484375,
0.5185546875,
0.146484375
],
[
0.0849609375,
0.149658203125,
0.48193359375,
0.149658203125,
0.48193359375,
0.1634521484375,
0.0849609375,
0.1634521484375
],
[
0.51708984375,
0.149658203125,
0.9140625,
0.149658203125,
0.9140625,
0.1634521484375,
0.51708984375,
0.1634521484375
],
[
0.0849609375,
0.1622314453125,
0.48046875,
0.1622314453125,
0.48046875,
0.175537109375,
0.0849609375,
0.175537109375
],
[
0.51708984375,
0.1634521484375,
0.91259765625,
0.1634521484375,
0.91259765625,
0.1767578125,
0.51708984375,
0.1767578125
],
[
0.08660888671875,
0.1787109375,
0.48046875,
0.1787109375,
0.48046875,
0.189453125,
0.08660888671875,
0.189453125
],
[
0.5185546875,
0.1767578125,
0.91259765625,
0.1767578125,
0.91259765625,
0.1900634765625,
0.5185546875,
0.1900634765625
],
[
0.0849609375,
0.1912841796875,
0.48193359375,
0.1912841796875,
0.48193359375,
0.2052001953125,
0.0849609375,
0.2052001953125
],
[
0.51953125,
0.1912841796875,
0.91259765625,
0.1912841796875,
0.91259765625,
0.2052001953125,
0.51953125,
0.2052001953125
],
[
0.0849609375,
0.2052001953125,
0.48193359375,
0.2052001953125,
0.48193359375,
0.2183837890625,
0.0849609375,
0.2183837890625
],
[
0.51953125,
0.203857421875,
0.91259765625,
0.2052001953125,
0.91259765625,
0.2183837890625,
0.51953125,
0.2171630859375
],
[
0.0849609375,
0.2203369140625,
0.38818359375,
0.2203369140625,
0.38818359375,
0.2310791015625,
0.0849609375,
0.2310791015625
],
[
0.51953125,
0.2183837890625,
0.91259765625,
0.2183837890625,
0.91259765625,
0.231689453125,
0.51953125,
0.231689453125
],
[
0.101318359375,
0.23291015625,
0.48046875,
0.23291015625,
0.48046875,
0.246826171875,
0.101318359375,
0.246826171875
],
[
0.5185546875,
0.23291015625,
0.9111328125,
0.23291015625,
0.9111328125,
0.246826171875,
0.5185546875,
0.246826171875
],
[
0.0849609375,
0.246826171875,
0.48193359375,
0.246826171875,
0.48193359375,
0.260009765625,
0.0849609375,
0.260009765625
],
[
0.53173828125,
0.2474365234375,
0.9111328125,
0.2474365234375,
0.9111328125,
0.25830078125,
0.53173828125,
0.25830078125
],
[
0.0849609375,
0.260009765625,
0.48193359375,
0.260009765625,
0.48193359375,
0.2734375,
0.0849609375,
0.2734375
],
[
0.5185546875,
0.260009765625,
0.9140625,
0.260009765625,
0.9140625,
0.2734375,
0.5185546875,
0.2734375
],
[
0.0849609375,
0.274658203125,
0.48046875,
0.274658203125,
0.48046875,
0.28857421875,
0.0849609375,
0.28857421875
],
[
0.521484375,
0.274658203125,
0.60693359375,
0.274658203125,
0.60693359375,
0.285888671875,
0.521484375,
0.285888671875
],
[
0.08660888671875,
0.2890625,
0.4794921875,
0.2890625,
0.4794921875,
0.2998046875,
0.08660888671875,
0.2998046875
],
[
0.53515625,
0.287353515625,
0.9140625,
0.287353515625,
0.9140625,
0.300537109375,
0.53515625,
0.300537109375
],
[
0.0849609375,
0.3017578125,
0.48193359375,
0.3017578125,
0.48193359375,
0.31494140625,
0.0849609375,
0.31494140625
],
[
0.51708984375,
0.3017578125,
0.91259765625,
0.3017578125,
0.91259765625,
0.31494140625,
0.51708984375,
0.31494140625
],
[
0.08660888671875,
0.317626953125,
0.48046875,
0.317626953125,
0.48046875,
0.32763671875,
0.08660888671875,
0.32763671875
],
[
0.51708984375,
0.314453125,
0.91259765625,
0.31494140625,
0.91259765625,
0.330078125,
0.51708984375,
0.328857421875
],
[
0.0849609375,
0.330078125,
0.483642578125,
0.330078125,
0.483642578125,
0.343505859375,
0.0849609375,
0.343505859375
],
[
0.5185546875,
0.328857421875,
0.91259765625,
0.328857421875,
0.91259765625,
0.34228515625,
0.5185546875,
0.34228515625
],
[
0.0849609375,
0.343505859375,
0.483642578125,
0.343505859375,
0.483642578125,
0.356689453125,
0.0849609375,
0.356689453125
],
[
0.5185546875,
0.34228515625,
0.71337890625,
0.34228515625,
0.71337890625,
0.35595703125,
0.5185546875,
0.35595703125
],
[
0.08660888671875,
0.359130859375,
0.48046875,
0.359130859375,
0.48046875,
0.369384765625,
0.08660888671875,
0.369384765625
],
[
0.51708984375,
0.35986328125,
0.91259765625,
0.35986328125,
0.91259765625,
0.373779296875,
0.51708984375,
0.373779296875
],
[
0.08660888671875,
0.37255859375,
0.48046875,
0.37255859375,
0.48046875,
0.38330078125,
0.08660888671875,
0.38330078125
],
[
0.529296875,
0.375,
0.9140625,
0.373779296875,
0.9140625,
0.386962890625,
0.529296875,
0.38818359375
],
[
0.08660888671875,
0.385009765625,
0.48193359375,
0.385009765625,
0.48193359375,
0.3984375,
0.08660888671875,
0.3984375
],
[
0.53125,
0.386962890625,
0.74658203125,
0.38818359375,
0.74658203125,
0.401611328125,
0.53125,
0.40087890625
],
[
0.08660888671875,
0.40087890625,
0.4794921875,
0.40087890625,
0.4794921875,
0.410888671875,
0.08660888671875,
0.410888671875
],
[
0.51708984375,
0.401611328125,
0.91259765625,
0.401611328125,
0.91259765625,
0.415283203125,
0.51708984375,
0.415283203125
],
[
0.08660888671875,
0.4140625,
0.48193359375,
0.4140625,
0.48193359375,
0.4248046875,
0.08660888671875,
0.4248046875
],
[
0.53125,
0.416748046875,
0.91259765625,
0.416748046875,
0.91259765625,
0.429931640625,
0.53125,
0.429931640625
],
[
0.08416748046875,
0.425537109375,
0.4306640625,
0.425537109375,
0.4306640625,
0.439453125,
0.08416748046875,
0.439453125
],
[
0.53125,
0.429931640625,
0.91259765625,
0.429931640625,
0.91259765625,
0.443115234375,
0.53125,
0.443115234375
],
[
0.53125,
0.443115234375,
0.91259765625,
0.443115234375,
0.91259765625,
0.45703125,
0.53125,
0.45703125
],
[
0.12744140625,
0.462158203125,
0.23779296875,
0.46337890625,
0.23779296875,
0.477783203125,
0.12744140625,
0.4765625
],
[
0.53125,
0.458251953125,
0.9140625,
0.458251953125,
0.9140625,
0.4716796875,
0.53125,
0.4716796875
],
[
0.1168212890625,
0.466552734375,
0.131591796875,
0.466552734375,
0.131591796875,
0.473388671875,
0.1168212890625,
0.473388671875
],
[
0.53173828125,
0.472900390625,
0.91259765625,
0.472900390625,
0.91259765625,
0.48291015625,
0.53173828125,
0.48291015625
],
[
0.0882568359375,
0.464111328125,
0.11767578125,
0.464111328125,
0.11767578125,
0.47607421875,
0.0882568359375,
0.47607421875
],
[
0.08660888671875,
0.48095703125,
0.48046875,
0.48095703125,
0.48046875,
0.494384765625,
0.08660888671875,
0.494384765625
],
[
0.53125,
0.48486328125,
0.75732421875,
0.48486328125,
0.75732421875,
0.498779296875,
0.53125,
0.498779296875
],
[
0.0849609375,
0.494384765625,
0.48193359375,
0.49560546875,
0.48193359375,
0.5087890625,
0.0849609375,
0.50830078125
],
[
0.51708984375,
0.498779296875,
0.9140625,
0.498779296875,
0.9140625,
0.51220703125,
0.51708984375,
0.51220703125
],
[
0.08660888671875,
0.5087890625,
0.48193359375,
0.5087890625,
0.48193359375,
0.52294921875,
0.08660888671875,
0.52294921875
],
[
0.53125,
0.51318359375,
0.9140625,
0.51318359375,
0.9140625,
0.5263671875,
0.53125,
0.5263671875
],
[
0.0882568359375,
0.5244140625,
0.30810546875,
0.5244140625,
0.30810546875,
0.53515625,
0.0882568359375,
0.53515625
],
[
0.53125,
0.5263671875,
0.9140625,
0.5263671875,
0.9140625,
0.54052734375,
0.53125,
0.54052734375
],
[
0.10296630859375,
0.5361328125,
0.48046875,
0.5361328125,
0.48046875,
0.5498046875,
0.10296630859375,
0.5498046875
],
[
0.53125,
0.54052734375,
0.6953125,
0.54052734375,
0.6953125,
0.5537109375,
0.53125,
0.5537109375
],
[
0.0849609375,
0.55029296875,
0.5009765625,
0.55029296875,
0.5009765625,
0.564453125,
0.0849609375,
0.564453125
],
[
0.0849609375,
0.56494140625,
0.48046875,
0.56494140625,
0.48046875,
0.5791015625,
0.0849609375,
0.5791015625
],
[
0.51953125,
0.5693359375,
0.53173828125,
0.5693359375,
0.53173828125,
0.57763671875,
0.51953125,
0.57763671875
],
[
0.54736328125,
0.56494140625,
0.69384765625,
0.56494140625,
0.69384765625,
0.5791015625,
0.54736328125,
0.5791015625
],
[
0.0849609375,
0.5791015625,
0.48193359375,
0.5791015625,
0.48193359375,
0.59228515625,
0.0849609375,
0.59228515625
],
[
0.5185546875,
0.58544921875,
0.91259765625,
0.58544921875,
0.91259765625,
0.5986328125,
0.5185546875,
0.5986328125
],
[
0.0849609375,
0.59130859375,
0.48046875,
0.59228515625,
0.48046875,
0.60595703125,
0.0849609375,
0.60498046875
],
[
0.5185546875,
0.5986328125,
0.9140625,
0.5986328125,
0.9140625,
0.6123046875,
0.5185546875,
0.6123046875
],
[
0.0849609375,
0.60595703125,
0.48193359375,
0.60595703125,
0.48193359375,
0.619140625,
0.0849609375,
0.619140625
],
[
0.51708984375,
0.61279296875,
0.79150390625,
0.61279296875,
0.79150390625,
0.626953125,
0.51708984375,
0.626953125
],
[
0.08660888671875,
0.6220703125,
0.48046875,
0.6220703125,
0.48046875,
0.6318359375,
0.08660888671875,
0.6318359375
],
[
0.08416748046875,
0.6337890625,
0.483642578125,
0.6337890625,
0.483642578125,
0.64794921875,
0.08416748046875,
0.64794921875
],
[
0.5185546875,
0.6416015625,
0.54736328125,
0.6416015625,
0.54736328125,
0.65283203125,
0.5185546875,
0.65283203125
],
[
0.560546875,
0.6396484375,
0.7060546875,
0.64013671875,
0.7060546875,
0.65478515625,
0.560546875,
0.65380859375
],
[
0.0849609375,
0.64794921875,
0.484375,
0.64794921875,
0.484375,
0.6611328125,
0.0849609375,
0.6611328125
],
[
0.08660888671875,
0.66357421875,
0.48046875,
0.66357421875,
0.48046875,
0.673828125,
0.08660888671875,
0.673828125
],
[
0.51953125,
0.65771484375,
0.91259765625,
0.65771484375,
0.91259765625,
0.671875,
0.51953125,
0.671875
],
[
0.08416748046875,
0.67529296875,
0.48193359375,
0.67529296875,
0.48193359375,
0.689453125,
0.08416748046875,
0.689453125
],
[
0.5185546875,
0.671875,
0.91259765625,
0.671875,
0.91259765625,
0.68505859375,
0.5185546875,
0.68505859375
],
[
0.51953125,
0.68505859375,
0.900390625,
0.68505859375,
0.900390625,
0.69873046875,
0.51953125,
0.69873046875
],
[
0.89453125,
0.68798828125,
0.9111328125,
0.68798828125,
0.9111328125,
0.6962890625,
0.89453125,
0.6962890625
],
[
0.5185546875,
0.69873046875,
0.91259765625,
0.69873046875,
0.91259765625,
0.7119140625,
0.5185546875,
0.7119140625
],
[
0.0849609375,
0.689453125,
0.48193359375,
0.689453125,
0.48193359375,
0.70263671875,
0.0849609375,
0.70263671875
],
[
0.08416748046875,
0.7041015625,
0.48193359375,
0.70263671875,
0.48193359375,
0.71630859375,
0.08416748046875,
0.71728515625
],
[
0.0849609375,
0.71728515625,
0.47216796875,
0.71728515625,
0.47216796875,
0.73095703125,
0.0849609375,
0.73095703125
],
[
0.53125,
0.71337890625,
0.91259765625,
0.71337890625,
0.91259765625,
0.7265625,
0.53125,
0.7265625
],
[
0.101318359375,
0.73095703125,
0.48193359375,
0.73095703125,
0.48193359375,
0.744140625,
0.101318359375,
0.744140625
],
[
0.5185546875,
0.7265625,
0.91259765625,
0.7265625,
0.91259765625,
0.74072265625,
0.5185546875,
0.74072265625
],
[
0.08660888671875,
0.7470703125,
0.4794921875,
0.7470703125,
0.4794921875,
0.7568359375,
0.08660888671875,
0.7568359375
],
[
0.51953125,
0.7412109375,
0.9140625,
0.7412109375,
0.9140625,
0.7548828125,
0.51953125,
0.7548828125
],
[
0.0849609375,
0.75830078125,
0.48193359375,
0.75830078125,
0.48193359375,
0.771484375,
0.0849609375,
0.771484375
],
[
0.5185546875,
0.7548828125,
0.91259765625,
0.7548828125,
0.91259765625,
0.76806640625,
0.5185546875,
0.76806640625
],
[
0.08416748046875,
0.77099609375,
0.48046875,
0.771484375,
0.48046875,
0.7861328125,
0.08416748046875,
0.78515625
],
[
0.5185546875,
0.76806640625,
0.9111328125,
0.76806640625,
0.9111328125,
0.7822265625,
0.5185546875,
0.7822265625
],
[
0.08660888671875,
0.787109375,
0.48046875,
0.787109375,
0.48046875,
0.79736328125,
0.08660888671875,
0.79736328125
],
[
0.5185546875,
0.78271484375,
0.63818359375,
0.78271484375,
0.63818359375,
0.796875,
0.5185546875,
0.796875
],
[
0.0849609375,
0.7998046875,
0.48046875,
0.7998046875,
0.48046875,
0.81298828125,
0.0849609375,
0.81298828125
],
[
0.08416748046875,
0.8125,
0.48193359375,
0.81298828125,
0.48193359375,
0.82763671875,
0.08416748046875,
0.8271484375
],
[
0.515625,
0.80810546875,
0.79736328125,
0.8095703125,
0.79736328125,
0.82373046875,
0.515625,
0.82275390625
],
[
0.08660888671875,
0.8291015625,
0.48193359375,
0.8291015625,
0.48193359375,
0.8388671875,
0.08660888671875,
0.8388671875
],
[
0.5185546875,
0.8271484375,
0.91259765625,
0.8271484375,
0.91259765625,
0.84033203125,
0.5185546875,
0.84033203125
],
[
0.0849609375,
0.84130859375,
0.483642578125,
0.84130859375,
0.483642578125,
0.85498046875,
0.0849609375,
0.85498046875
],
[
0.51953125,
0.84130859375,
0.91259765625,
0.84130859375,
0.91259765625,
0.85498046875,
0.51953125,
0.85498046875
],
[
0.0849609375,
0.85595703125,
0.48193359375,
0.85595703125,
0.48193359375,
0.869140625,
0.0849609375,
0.869140625
],
[
0.5185546875,
0.85595703125,
0.89892578125,
0.85498046875,
0.89892578125,
0.8662109375,
0.5185546875,
0.8671875
],
[
0.89453125,
0.85791015625,
0.91015625,
0.85791015625,
0.91015625,
0.86572265625,
0.89453125,
0.86572265625
],
[
0.0849609375,
0.869140625,
0.48193359375,
0.869140625,
0.48193359375,
0.88330078125,
0.0849609375,
0.88330078125
],
[
0.51953125,
0.87060546875,
0.9111328125,
0.87060546875,
0.9111328125,
0.880859375,
0.51953125,
0.880859375
],
[
0.0849609375,
0.8837890625,
0.26953125,
0.8837890625,
0.26953125,
0.89794921875,
0.0849609375,
0.89794921875
],
[
0.5185546875,
0.88330078125,
0.9140625,
0.88330078125,
0.9140625,
0.896484375,
0.5185546875,
0.896484375
]
] |
[
"www23, April 30-May 4, 2023, Austin, Tx, usA",
"Xu Huang, Defu Lian 12.3, Jin Chen , Zheng Liu 5, Xing Xie', Enhong Chen 1.2.3",
"by the ranker; thus, it may harm the downstream performance as",
"It's worth noting that the realization of the above training frame-",
"well.",
"work is non-trivial. Particularly, both retriever and ranker need to",
". The independently-trained ranker is learned with randomly",
"learnfromthesampleditems;however,thesamplingoperationon",
"or heuristically collected training samples. Such training samples",
"the retriever can be inefficient and biased. To overcome such chal",
"can be too easy to be distinguished, making the ranker converge to",
"lengesa coupleof technicaldesignsareintroduced.首先，我们",
"a limited discriminativecapacity.Besides,theitem distributionwill",
"知道在处理大型项目集时，直接从检索器进行采样可能非常",
"also be highly differentiated between the training and inference",
"耗时，因此我们开发了一种可扩展且自适应的采样策略，在该",
"stages; as a result, the ranker may not effectively recognize the",
"策略中，可以根据项目大小在次线性时间内对检索器偏爱的",
"high-quality candidatesgenerated by the retriever",
"项目进行有效采样。其次，对抽样项目的直接知识提炼是有",
"Recent studies on multi-stage ranking models are closely related",
"偏见的，容易出现劣质表现。为了缓解这个问题，我们提出了",
"to the problem of retriever-ranker collaboration; e.g., in [34] , a",
"种新的KL散度的渐近无偏估计，以补偿项目抽样引起的偏",
"two-pass training workflow is proposed. However, the two-pass",
"差。在此操作之上，可以更好地在排名器和检索器之间对齐项",
"workflow is still limited from several critical perspectives. In the",
"目的排名顺序",
"forward pass,the rankers aretrained by the retrieval results at the",
"我们对四个基准数据集进行了全面的实验研究。根据实验",
"exact top-k cutoff, which is prone to false-negatives. Besides, when",
"结果，与现有的训练方法相比，CoRR可以大大提高整体推荐",
"the retrieval cutoffs are changed during the inference stage, the",
"质量。更详细的分析进一步验证了CoRR对寻回者和排名者的",
"rankers will face highly shifted item distributions from the train-",
"独立有效性，以及其作为模型不可知训练框架的有效性。我们",
"ing stage, which may severely harm their prediction accuracy. In",
"工作的contributions总结如下",
"thebackwardpass,theretrieversaretrainedtopreservetheconsis-",
". We present a novel training framework for deep recommender",
"tency of absoluteranking scores and todistinguishranking results",
"systems, where the retriever and ranker can be mutually rein-",
"from retrieval ones; while for the sake of high-quality retrieval, it",
"forced for the effective cooperation.",
"is the consistency of relativeranking order that really matters.In",
". Two critical techniques are introduced for the optimized con-",
"all,it remains achallengingproblem toexploremore effectivecol-",
"duct of CoRR: 1) the scalable and adaptive sampling strategy,",
"laboration mechanisms between the retriever and ranker.",
"which enables the efficient sampling from the retriever; 2) the",
"asymptotic-unbiased estimation of KL divergence, as the objec-",
"Our Solution",
"tive of knowledge distillation, which better aligns the ranking",
"",
"orderofitemsbetweentherankerand retriever,and contributes",
"1.2",
"在这项工作中，我们提出了一种新颖的框架，用于寻回者和",
"to the retrieval of high-quality items",
"排名者的合作训练，a。k。a。CoRR(合作Retriever和Ranker)。",
".We perform comprehensive experimental studies on four bench-",
"在我们的框架中，检索器和排名器在统一的工作流程中同时",
"mark datasets, whose results verify CoRR's advantage against",
"训练，其中两个模型可以相互加强",
"the existing training algorithms, and its standalone effectiveness",
". Training of retriever . On one hand, the retriever is learned",
"to the retriever and ranker.",
"from both user-item interactions via sampled softmax and the ranker's",
"predictions via knowledge distillation [16] . In a specific context,",
"2.",
"RELATED WORK",
"a few items are sampled firstly. Then, the ranker is required to",
"本文研究了深度推荐系统中排名器和检索器之间的合作，以",
"predict the fine-grained preferences towards the sampled items.",
"提高多级系统的推荐质量。我们首先回顾了密切相关的级联",
"Rather than preservingthe absolute preference scores, theretriever",
"排序技术，然后介绍了负采样和知识提炼。",
"is required togenerate the same ranking orderfor the sampled",
"items as the ranker. To realize this goal, the KL-divergence is mini-",
"2.1",
"Cascade Ranking",
"mizedforthe softmax-normalized predictions betweentheretriever",
"and ranker[16].In this case,user'spreferreditems,whetherinter-",
"关于多阶段级联排名的先前工作通常为每个阶段分配不同的",
"acted or not, will probably get highly-rated by the ranker, while",
"排名，以在效率和有效性之间共同实现所需的权衡E[43]。它",
"们在对每个排名者的成本进行建模时彼此不同[6,47,48]。",
"最",
"近的工作转向通过梯度下降[11]直接优化级联排序模型作为",
"real non-interested items get lower-rated (thanks to the highly-",
"precise ranker). As a result, such items will virtually become la-",
"beled samples, which substantially augment the training data.",
"个整体，或者用级联排序模型识别一些坏的情况来增加训",
". Training of ranker . The ranker is trained by sampled soft-",
"练数据[8]。观察这些级联排名系统不考虑排名者之间的合作，",
"max1]ontopofthehardnegativeitemssampledbytheretriever",
"工作E[34]建议通过让级联排名者在级联系统中为彼此提供",
"Particularly, instead of working with the\"easy negatives\"which",
"监督信号来共同优化它们。然而，如上所述，从几个关键的角",
"are randomly or heuristically sampled from the whole item set [12,",
"度来看，这项工作仍然有限。这项工作旨在探索级联排名者之",
"36, 52l,the ranker is iteratively trained to discriminate the true",
"间更好的协作机制",
"positive from the increasingly harder negatives as the retriever im",
"proves. Therefore, it prevents the rankerfrom converging too early",
"2.2NegativeSamplinginRecSys",
"to a limited discriminative capability.Besides, unlikethe widely-",
"许多方法从静态分布中对负面项目进行采样，例如均匀分布E",
"used exact top-k sampling, we collect informative negative sam-",
"[14,36]和基于流行度的分布[35]。为了适应推荐模型，已",
"ples from the entire itemset based on the retriever; by doing so, it",
"经提出了许多先进的抽样方法[2,5,18,30,35.39.45.50]。",
"例",
"alleviates the false negative issue and closes the gap between the",
"如，AOBPRE[35]将自适应采样转换为以随机抽取的等级搜索",
"training and inference stages .",
"项目。CML[18]和WARP[45]通过首先从均匀分布中抽取候"
] |
{
"arxiv_id": "2206.14649",
"language": "zh",
"page_idx": 1
}
|
||
2206.14649-zh-2
|
2206.14649-zh-2.jpg
|
Cooperative Retriever and Ranker in Deep Recommenders
选项目，然后选择得分高于正负 1 的项目作为负来为每个正数抽取负样本。动态负抽样 (DNS)[50] 从均匀分布中挑选一组负样本，然后选择得分最高的项目作为负。自我对抗负抽样[39] 从均匀分布中抽取负样本，但使用 softmax-normalized 分数对其进行加权。基于内核的采样$\mathbb{E}[2]$ 以分而治之的方式按比例选择与二次内核成比例的样本。随机扰动数据库上的局部敏感哈希 (LSH) 可以进行次线性时间采样，并且 LSH 本身可以生成相关且未归一化的样本$\mathbb{E}[37]$。基于量化的采样E[5] 通过乘积量化将 softmax 归一化概率分解为簇上的概率乘积，从而使采样项目处于亚线性时间
# 2.3 Knowledge Distillation in RecSys
KD [16] 提供了一个强大的模型无关框架，用于通过学习模仿教师的预测来将教师模型压缩为学生模型。因此，KD 已在RecSys 中用于压缩深度推荐器。一个开创性的工作是排名蒸馏(RD)$\mathbb{E}[41]$,其中来自教师模型的 top- $k$ 推荐被认为是位置加权伪阳性。随后的蒸馏方法改进了 top- $k$建议的使用，例如通过基于排名的分布$\operatorname{E}[29]$ ,其具有均匀分布$\operatorname{E}[24]$ 和排名差异感知分布$\overline{\mathbb{E}}[28]$ 的混合物，从前 k 个项目中抽取样本。除了教师的预测外，还可以通过提示回归$\mathbb{E}[24]$ 和拓扑蒸馏$\mathbb{E}$ [25]将教师中的潜在知识提炼到学生中。
3 COOPERATIVE RETRIEVER AND RANKER
在本节中，我们首先概述了在训练数据集$\mathcal{D}=\{(c,k)\}$上的CoRR 训练，该数据集由上下文$c$和正项$k$的对组成。上下文表示除项目以外的所有信息，例如用户、交互历史记录、时间和位置。之后，我们详细阐述了如何基于硬负采样来训练排名者，以及如何基于知识提炼来训练排名者。
# 3.1 0verview
为了可扩展的推荐，深度推荐器依赖于两个模型的协作：检索模型 (检索器) 和排名模型 (排名器)。检索器的目标是从整个项目中高效地选择一小组潜在的积极项目。通常，检索器由$M_{\theta}(i,c)=\operatorname{sim}(E_{{\theta_{1}}}(i),E_{{\theta_{2}}}(c))$ , i 表示。e。, 相似性 (e。g。项嵌入$E_{\theta_1}(i)$ 和上下文嵌入$E_\theta_2(c)$之间的欧氏距离、内积和余弦相似度)。因此，我们可以使用现成的人工神经网络，如FAISS[23] 和 SCANN[13] ,在亚线性时间检索最相似的项目作为候选。排名器旨在从检索结果中高精度地识别最佳项目， 通常由富有表现力但耗时的网络$R_\phi(i,c)$直接表示，以项目$i$ 和上下文$c$作为输入
由于排名器中的耗时计算，在大多数情况下，传统的训练方法学会区分阳性和随机挑选的阴性。尽管简单，但排名者是独立于猎犬的。在某些情况下，通过从检索者的前 k 个检索结果中区分阳性来训练排名者。尽管与检索者建立了联系，但排名者很容易遭受假阴性问题，因为潜在的阳性肯定包含在前 k 个检索结果中。此外，这些负面因素可能会给梯度带来很大的偏差，使得优化可能会缓慢收敛，甚至很容易陷入局部最优。为此，我们建议优化排名器 w。r。t 抽样的 log-softmax [1] , 基于重要性抽样的 log-softmax 的渐近无偏估计。抽样的$\log$- softmax 依赖于从提案分布中随机抽取的样本。为了将排名器与检索器联系起来，我们建议根据自适应和可扩展的策略对检索器的底片进行采样
除了来自数据的监督之外，检索器的训练还可以由排名器基于知识提炼来指导，因为排名器被认为在排名项目上更精确。关于来自推荐数据的监督，还利用采样的 log-softmax 目标来优化检索器，并在排名器优化中使用相同的建议。关于从
WWW'23, April 30-May 4, 2023, Austin, TX, USA
<FigureHere>
Figure 1: The framework of CoRR. The context c indicates all information except items. $Q(\cdot|c)$ indicates the proposal distribution given context $c. \textbf{The dotted line means the gra- }$
排eptris中提取知d识，大多数现有方法从排名者E[24,29,41] 中提取前 k 个结果中的知识，但是前 k 个排名结果对于排名者来说是耗时的。在本文中，我们通过基于 KL 散度直接对齐排名者和检索者之间的 softmax 归一化预测，从排名者的预测中提取排名顺序信息。为了减少优化的时间成本，我们提出了 KL 散度的渐近无偏估计，称为采样 KL 散度
检索器和排名器通过统一的工作流程同时进行训练，其中两个模型可以相互增强。随着训练的进行，排名器变得越来越精确，这反过来为检索器提供了更多的信息监督信号；同时， 随着检索器的改进，负样本将变得越来越困难，这有助于排名器的更高判别能力。整体框架如图 1 所示。整个过程可以参考算法 1
## 3.2 Training Ranker with Retrieven
3.2.1 $LossFunction.$在本文中，我们优化了排名器 w。r。t 抽样 log-softmax, 它是 log-softmax 的渐近无偏估计。log-softmax 的使用是由于其与平均倒数排名(MRR)和归一化贴现累积收益 (NDCG)$\operatorname{\overline{\mathrm{E}}[3]}$的对数的密切关系以及出色的推荐性能$\operatorname{\overline{\mathrm{E}}}$ [33] 。假设上下文$c$处的正项为$k$,则 log-softmax 目标 w。r。t$k$的公式如下：
$$\begin{aligned}\ell(k,c)&=R_{\phi}(k,c)-\log\left(\sum_{i}\exp\left(R_{\phi}(i,c)\right)\right)\\&=-\log\left(\sum_{i}\exp\left(R_{\phi}(i,c)-R_{\phi}(k,c)\right)\right)\\&\leq-\log\left(\sum_{i}\mathbb{I}_{\geq0}\left(R_{\phi}(i,c)-R_{\phi}(k,c)\right)\right)\\&=\log\left(\frac{1}{\mathrm{rank}_{R}(k,c)}\right)=\log\mathrm{MRR}\end{aligned}$$
,其中不等式由于 exp$(\cdot)\geq\mathbb{I}_\geq0(\cdot)$而成立
但是，优化 log-softmax 在计算上具有挑战性，因为梯度计
算与项目数呈线性比例，即。e。,
$$\nabla\ell_{\phi}(k,c)=\nabla R_{\phi}(k,c)-\mathbb{E}_{i\sim P_{\phi}(\cdot|c)}[\nabla R_{\phi}(i,c)]$$
其中$P_\phi(\cdot|c)$表示具有参数 softmax$(R_\phi(\cdot,c))$ , i 的整个项目$I$ 的分类概率分布。e。,$P_\phi(j|c)=\frac{\exp^{\prime}(R_\phi(j,c)}{\sum_{i\in I}\exp(R_\phi(i,c))}$。为了提高效率，通过从建议分布$Q(\cdot|c)$中绘制$L$样本$\mathcal{S}=\{o_1,\cdots,o_L\}$ ,使
|
[
[
0.0882568359375,
0.07763671875,
0.361083984375,
0.07763671875,
0.361083984375,
0.08837890625,
0.0882568359375,
0.08837890625
],
[
0.68115234375,
0.0770263671875,
0.9111328125,
0.0770263671875,
0.9111328125,
0.08709716796875,
0.68115234375,
0.08709716796875
],
[
0.08660888671875,
0.106689453125,
0.48046875,
0.10797119140625,
0.48046875,
0.121826171875,
0.08660888671875,
0.12060546875
],
[
0.57421875,
0.10797119140625,
0.615234375,
0.10797119140625,
0.615234375,
0.12249755859375,
0.57421875,
0.12249755859375
],
[
0.65869140625,
0.10919189453125,
0.76953125,
0.10919189453125,
0.76953125,
0.12249755859375,
0.65869140625,
0.12249755859375
],
[
0.81201171875,
0.10919189453125,
0.85107421875,
0.10919189453125,
0.85107421875,
0.12060546875,
0.81201171875,
0.12060546875
],
[
0.0882568359375,
0.121826171875,
0.48046875,
0.121826171875,
0.48046875,
0.1351318359375,
0.0882568359375,
0.1351318359375
],
[
0.576171875,
0.12249755859375,
0.61767578125,
0.12249755859375,
0.61767578125,
0.14453125,
0.576171875,
0.14453125
],
[
0.685546875,
0.1256103515625,
0.74658203125,
0.1256103515625,
0.74658203125,
0.1363525390625,
0.685546875,
0.1363525390625
],
[
0.810546875,
0.121826171875,
0.85888671875,
0.121826171875,
0.85888671875,
0.1458740234375,
0.810546875,
0.1458740234375
],
[
0.0882568359375,
0.136962890625,
0.4794921875,
0.136962890625,
0.4794921875,
0.147705078125,
0.0882568359375,
0.147705078125
],
[
0.08905029296875,
0.15087890625,
0.47802734375,
0.15087890625,
0.47802734375,
0.1610107421875,
0.08905029296875,
0.1610107421875
],
[
0.0849609375,
0.1622314453125,
0.4794921875,
0.1634521484375,
0.4794921875,
0.1767578125,
0.0849609375,
0.175537109375
],
[
0.08905029296875,
0.1787109375,
0.47802734375,
0.1787109375,
0.47802734375,
0.189453125,
0.08905029296875,
0.189453125
],
[
0.82861328125,
0.1787109375,
0.8408203125,
0.1787109375,
0.8408203125,
0.1875,
0.82861328125,
0.1875
],
[
0.0882568359375,
0.1925048828125,
0.47802734375,
0.1925048828125,
0.47802734375,
0.20263671875,
0.0882568359375,
0.20263671875
],
[
0.59228515625,
0.1932373046875,
0.60302734375,
0.1932373046875,
0.60302734375,
0.210205078125,
0.59228515625,
0.210205078125
],
[
0.830078125,
0.1932373046875,
0.8408203125,
0.1932373046875,
0.8408203125,
0.205810546875,
0.830078125,
0.205810546875
],
[
0.8310546875,
0.1900634765625,
0.83837890625,
0.1900634765625,
0.83837890625,
0.1944580078125,
0.8310546875,
0.1944580078125
],
[
0.08660888671875,
0.2052001953125,
0.48046875,
0.2052001953125,
0.48046875,
0.2183837890625,
0.08660888671875,
0.2183837890625
],
[
0.56201171875,
0.208984375,
0.583984375,
0.21337890625,
0.580078125,
0.2259521484375,
0.55810546875,
0.2215576171875
],
[
0.7099609375,
0.2052001953125,
0.7158203125,
0.2052001953125,
0.7158203125,
0.21337890625,
0.7099609375,
0.21337890625
],
[
0.583984375,
0.2171630859375,
0.59326171875,
0.21337890625,
0.59619140625,
0.216552734375,
0.58642578125,
0.2203369140625
],
[
0.79833984375,
0.21337890625,
0.82421875,
0.21337890625,
0.82421875,
0.2247314453125,
0.79833984375,
0.2247314453125
],
[
0.6845703125,
0.2197265625,
0.74267578125,
0.2197265625,
0.74267578125,
0.23486328125,
0.6845703125,
0.23486328125
],
[
0.08660888671875,
0.23291015625,
0.27783203125,
0.23291015625,
0.27783203125,
0.246826171875,
0.08660888671875,
0.246826171875
],
[
0.0882568359375,
0.2183837890625,
0.48046875,
0.2183837890625,
0.48046875,
0.231689453125,
0.0882568359375,
0.231689453125
],
[
0.7099609375,
0.2442626953125,
0.71728515625,
0.2442626953125,
0.71728515625,
0.251953125,
0.7099609375,
0.251953125
],
[
0.57421875,
0.25390625,
0.61669921875,
0.25390625,
0.61669921875,
0.26513671875,
0.57421875,
0.26513671875
],
[
0.814453125,
0.25390625,
0.85693359375,
0.25390625,
0.85693359375,
0.264404296875,
0.814453125,
0.264404296875
],
[
0.0849609375,
0.260009765625,
0.40185546875,
0.260009765625,
0.40185546875,
0.2734375,
0.0849609375,
0.2734375
],
[
0.576171875,
0.26318359375,
0.61669921875,
0.26318359375,
0.61669921875,
0.2734375,
0.576171875,
0.2734375
],
[
0.72802734375,
0.261962890625,
0.7685546875,
0.261962890625,
0.7685546875,
0.276611328125,
0.72802734375,
0.276611328125
],
[
0.814453125,
0.26318359375,
0.85546875,
0.26318359375,
0.85546875,
0.2734375,
0.814453125,
0.2734375
],
[
0.08660888671875,
0.27783203125,
0.4794921875,
0.27783203125,
0.4794921875,
0.291748046875,
0.08660888671875,
0.291748046875
],
[
0.72119140625,
0.28271484375,
0.72802734375,
0.28271484375,
0.72802734375,
0.290283203125,
0.72119140625,
0.290283203125
],
[
0.7294921875,
0.27978515625,
0.76318359375,
0.285400390625,
0.7587890625,
0.2998046875,
0.7255859375,
0.294189453125
],
[
0.0882568359375,
0.29345703125,
0.4794921875,
0.29345703125,
0.4794921875,
0.3037109375,
0.0882568359375,
0.3037109375
],
[
0.0849609375,
0.30615234375,
0.48046875,
0.30615234375,
0.48046875,
0.3193359375,
0.0849609375,
0.3193359375
],
[
0.5185546875,
0.30615234375,
0.9111328125,
0.30615234375,
0.9111328125,
0.3193359375,
0.5185546875,
0.3193359375
],
[
0.08660888671875,
0.3193359375,
0.48046875,
0.3193359375,
0.48046875,
0.333251953125,
0.08660888671875,
0.333251953125
],
[
0.5185546875,
0.318115234375,
0.9111328125,
0.318115234375,
0.9111328125,
0.33203125,
0.5185546875,
0.33203125
],
[
0.08905029296875,
0.335205078125,
0.4794921875,
0.335205078125,
0.4794921875,
0.34521484375,
0.08905029296875,
0.34521484375
],
[
0.5185546875,
0.333984375,
0.91259765625,
0.333984375,
0.91259765625,
0.347900390625,
0.5185546875,
0.347900390625
],
[
0.08660888671875,
0.347900390625,
0.48046875,
0.347900390625,
0.48046875,
0.361083984375,
0.08660888671875,
0.361083984375
],
[
0.51708984375,
0.348388671875,
0.9111328125,
0.349853515625,
0.9111328125,
0.363037109375,
0.51708984375,
0.3623046875
],
[
0.08660888671875,
0.361083984375,
0.48046875,
0.361083984375,
0.48046875,
0.375,
0.08660888671875,
0.375
],
[
0.5185546875,
0.363037109375,
0.91259765625,
0.363037109375,
0.91259765625,
0.376953125,
0.5185546875,
0.376953125
],
[
0.08905029296875,
0.375,
0.48193359375,
0.375,
0.48193359375,
0.38818359375,
0.08905029296875,
0.38818359375
],
[
0.5185546875,
0.37744140625,
0.9111328125,
0.37890625,
0.9111328125,
0.39013671875,
0.5185546875,
0.389404296875
],
[
0.0882568359375,
0.39013671875,
0.34228515625,
0.39013671875,
0.34228515625,
0.40087890625,
0.0882568359375,
0.40087890625
],
[
0.51953125,
0.393310546875,
0.9111328125,
0.393310546875,
0.9111328125,
0.404052734375,
0.51953125,
0.404052734375
],
[
0.51953125,
0.4072265625,
0.91259765625,
0.4072265625,
0.91259765625,
0.417236328125,
0.51953125,
0.417236328125
],
[
0.08660888671875,
0.41845703125,
0.11517333984375,
0.41845703125,
0.11517333984375,
0.4267578125,
0.08660888671875,
0.4267578125
],
[
0.111083984375,
0.41845703125,
0.4755859375,
0.41845703125,
0.4755859375,
0.4287109375,
0.111083984375,
0.4287109375
],
[
0.5185546875,
0.419189453125,
0.77783203125,
0.419189453125,
0.77783203125,
0.43310546875,
0.5185546875,
0.43310546875
],
[
0.0882568359375,
0.436279296875,
0.47802734375,
0.436279296875,
0.47802734375,
0.4462890625,
0.0882568359375,
0.4462890625
],
[
0.53515625,
0.43505859375,
0.9111328125,
0.43505859375,
0.9111328125,
0.44580078125,
0.53515625,
0.44580078125
],
[
0.0849609375,
0.448974609375,
0.48046875,
0.448974609375,
0.48046875,
0.462158203125,
0.0849609375,
0.462158203125
],
[
0.5185546875,
0.4462890625,
0.91259765625,
0.4462890625,
0.91259765625,
0.460205078125,
0.5185546875,
0.460205078125
],
[
0.0882568359375,
0.464111328125,
0.47802734375,
0.464111328125,
0.47802734375,
0.474853515625,
0.0882568359375,
0.474853515625
],
[
0.51953125,
0.462158203125,
0.9111328125,
0.462158203125,
0.9111328125,
0.472900390625,
0.51953125,
0.472900390625
],
[
0.0849609375,
0.474853515625,
0.48046875,
0.47607421875,
0.48046875,
0.4892578125,
0.0849609375,
0.488037109375
],
[
0.5185546875,
0.474853515625,
0.91259765625,
0.474853515625,
0.91259765625,
0.488037109375,
0.5185546875,
0.488037109375
],
[
0.08660888671875,
0.4892578125,
0.358642578125,
0.4892578125,
0.358642578125,
0.50244140625,
0.08660888671875,
0.50244140625
],
[
0.5185546875,
0.488037109375,
0.9111328125,
0.488037109375,
0.9111328125,
0.501953125,
0.5185546875,
0.501953125
],
[
0.5185546875,
0.50390625,
0.56201171875,
0.50390625,
0.56201171875,
0.51513671875,
0.5185546875,
0.51513671875
],
[
0.0849609375,
0.517578125,
0.2091064453125,
0.517578125,
0.2091064453125,
0.53076171875,
0.0849609375,
0.53076171875
],
[
0.51708984375,
0.53076171875,
0.82177734375,
0.53076171875,
0.82177734375,
0.54443359375,
0.51708984375,
0.54443359375
],
[
0.08905029296875,
0.537109375,
0.4794921875,
0.537109375,
0.4794921875,
0.54736328125,
0.08905029296875,
0.54736328125
],
[
0.0849609375,
0.548828125,
0.4794921875,
0.5498046875,
0.4794921875,
0.56298828125,
0.0849609375,
0.5625
],
[
0.5185546875,
0.548828125,
0.91259765625,
0.548828125,
0.91259765625,
0.5625,
0.5185546875,
0.5625
],
[
0.08905029296875,
0.56494140625,
0.4794921875,
0.56494140625,
0.4794921875,
0.57568359375,
0.08905029296875,
0.57568359375
],
[
0.51708984375,
0.56103515625,
0.91259765625,
0.56298828125,
0.91259765625,
0.5771484375,
0.51708984375,
0.57470703125
],
[
0.08660888671875,
0.5771484375,
0.4794921875,
0.57763671875,
0.4794921875,
0.59130859375,
0.08660888671875,
0.59033203125
],
[
0.5185546875,
0.57568359375,
0.91259765625,
0.57568359375,
0.91259765625,
0.5888671875,
0.5185546875,
0.5888671875
],
[
0.08660888671875,
0.59130859375,
0.48046875,
0.59033203125,
0.48046875,
0.60400390625,
0.08660888671875,
0.60498046875
],
[
0.51953125,
0.59130859375,
0.9111328125,
0.59130859375,
0.9111328125,
0.6015625,
0.51953125,
0.6015625
],
[
0.0882568359375,
0.60498046875,
0.48193359375,
0.60498046875,
0.48193359375,
0.61865234375,
0.0882568359375,
0.61865234375
],
[
0.5185546875,
0.60400390625,
0.9140625,
0.60400390625,
0.9140625,
0.6171875,
0.5185546875,
0.6171875
],
[
0.0849609375,
0.61865234375,
0.4794921875,
0.61865234375,
0.4794921875,
0.6318359375,
0.0849609375,
0.6318359375
],
[
0.51708984375,
0.619140625,
0.61767578125,
0.619140625,
0.61767578125,
0.6298828125,
0.51708984375,
0.6298828125
],
[
0.0882568359375,
0.63330078125,
0.4794921875,
0.63330078125,
0.4794921875,
0.646484375,
0.0882568359375,
0.646484375
],
[
0.57861328125,
0.64013671875,
0.83837890625,
0.64013671875,
0.83837890625,
0.6611328125,
0.57861328125,
0.6611328125
],
[
0.0882568359375,
0.6484375,
0.48193359375,
0.6484375,
0.48193359375,
0.6591796875,
0.0882568359375,
0.6591796875
],
[
0.08660888671875,
0.6611328125,
0.2181396484375,
0.6611328125,
0.2181396484375,
0.6748046875,
0.08660888671875,
0.6748046875
],
[
0.101318359375,
0.6748046875,
0.48046875,
0.67529296875,
0.48046875,
0.689453125,
0.101318359375,
0.68798828125
],
[
0.62353515625,
0.673828125,
0.84716796875,
0.673828125,
0.84716796875,
0.69140625,
0.62353515625,
0.69140625
],
[
0.08660888671875,
0.689453125,
0.48046875,
0.689453125,
0.48046875,
0.70263671875,
0.08660888671875,
0.70263671875
],
[
0.08660888671875,
0.70263671875,
0.48046875,
0.70263671875,
0.48046875,
0.71630859375,
0.08660888671875,
0.71630859375
],
[
0.62158203125,
0.705078125,
0.8447265625,
0.705078125,
0.8447265625,
0.72265625,
0.62158203125,
0.72265625
],
[
0.08905029296875,
0.71826171875,
0.4794921875,
0.71826171875,
0.4794921875,
0.72900390625,
0.08905029296875,
0.72900390625
],
[
0.0882568359375,
0.73095703125,
0.48193359375,
0.73095703125,
0.48193359375,
0.744140625,
0.0882568359375,
0.744140625
],
[
0.62353515625,
0.7373046875,
0.74169921875,
0.75,
0.73779296875,
0.76806640625,
0.619140625,
0.755859375
],
[
0.69384765625,
0.73779296875,
0.70166015625,
0.73779296875,
0.70166015625,
0.74365234375,
0.69384765625,
0.74365234375
],
[
0.0882568359375,
0.744140625,
0.48046875,
0.744140625,
0.48046875,
0.75830078125,
0.0882568359375,
0.75830078125
],
[
0.73779296875,
0.7421875,
0.80810546875,
0.7421875,
0.80810546875,
0.75390625,
0.73779296875,
0.75390625
],
[
0.08660888671875,
0.75830078125,
0.48046875,
0.75830078125,
0.48046875,
0.771484375,
0.08660888671875,
0.771484375
],
[
0.08905029296875,
0.7734375,
0.48046875,
0.7734375,
0.48046875,
0.7841796875,
0.08905029296875,
0.7841796875
],
[
0.51953125,
0.76953125,
0.77783203125,
0.76953125,
0.77783203125,
0.78271484375,
0.51953125,
0.78271484375
],
[
0.0882568359375,
0.78515625,
0.48193359375,
0.7861328125,
0.48193359375,
0.7998046875,
0.0882568359375,
0.798828125
],
[
0.53369140625,
0.78271484375,
0.91015625,
0.78271484375,
0.91015625,
0.796875,
0.53369140625,
0.796875
],
[
0.08660888671875,
0.798828125,
0.48046875,
0.7998046875,
0.48046875,
0.81298828125,
0.08660888671875,
0.8125
],
[
0.51708984375,
0.79736328125,
0.72802734375,
0.798828125,
0.72802734375,
0.8125,
0.51708984375,
0.81103515625
],
[
0.08660888671875,
0.8125,
0.48046875,
0.81298828125,
0.48046875,
0.8271484375,
0.08660888671875,
0.82568359375
],
[
0.57861328125,
0.82080078125,
0.85107421875,
0.8193359375,
0.85107421875,
0.8369140625,
0.57861328125,
0.83837890625
],
[
0.0882568359375,
0.8291015625,
0.233642578125,
0.8291015625,
0.233642578125,
0.8388671875,
0.0882568359375,
0.8388671875
],
[
0.101318359375,
0.84130859375,
0.48193359375,
0.84130859375,
0.48193359375,
0.85498046875,
0.101318359375,
0.85498046875
],
[
0.5185546875,
0.8466796875,
0.9111328125,
0.8466796875,
0.9111328125,
0.85986328125,
0.5185546875,
0.85986328125
],
[
0.08660888671875,
0.85498046875,
0.48046875,
0.85498046875,
0.48046875,
0.86865234375,
0.08660888671875,
0.86865234375
],
[
0.51318359375,
0.8623046875,
0.91650390625,
0.8623046875,
0.91650390625,
0.88720703125,
0.51318359375,
0.88720703125
],
[
0.73779296875,
0.85986328125,
0.80615234375,
0.859375,
0.80615234375,
0.87060546875,
0.73779296875,
0.8720703125
],
[
0.0882568359375,
0.87060546875,
0.47802734375,
0.87060546875,
0.47802734375,
0.880859375,
0.0882568359375,
0.880859375
],
[
0.0849609375,
0.8818359375,
0.48046875,
0.88330078125,
0.48046875,
0.896484375,
0.0849609375,
0.89599609375
],
[
0.5185546875,
0.88330078125,
0.91259765625,
0.88330078125,
0.91259765625,
0.896484375,
0.5185546875,
0.896484375
]
] |
[
"Cooperative Retriever and Ranker in Deep Recommenders",
"WWW23,April 30-May4,2023,Austin,TX,USA",
"选项目，然后选择得分高于正负1的项目作为负来为每个正数",
"(k,c)",
"Dkz (Ps( 1c)I Q(-c)",
"g(k,c)",
"抽取负样本。动态负抽样(DNS)[50]从均匀分布中挑选一组负",
"sompled",
"samplsd kl-diy",
"orpid",
"样本，然后选择得分最高的项目作为负。自我对抗负抽样[39]",
"从均匀分布中抽取负样本，但使用softmax-normalized分数对",
"其进行加权。基于内核的采样国[2]以分而治之的方式按比例",
"选择与二次内核成比例的样本。随机扰动数据库上的局部敏",
"0",
"感哈希（LSH）可以进行次线性时间采样，并且LSH本身可以",
".",
"IBI",
"-",
"生成相关且未归一化的样本E[37]。基于量化的采样E[5]通",
"Me",
"+",
"",
"Ro",
"sampling",
"而使采样项目处于亚线性时间",
"过乘积量化将softmax归一化概率分解为簇上的概率乘积，从",
"丰",
"context",
"context",
"2.3KnowledgeDistillation inRecSys",
"",
"Q(- Ic)",
"C",
"KD[16]提供了一个强大的模型无关框架，用于通过学习模仿",
"",
"",
"教师的预测来将教师模型压缩为学生模型。因此，KD已在",
"RecSys中用于压缩深度推荐器。一个开创性的工作是排名蒸",
"Figure I: The framework of CoRR. The context c indicates",
"馏（RD)国[41]，其中来自教师模型的top-k推荐被认为是位",
"all information except items. Q(lc) indicates the proposal",
"置加权伪阳性。随后的蒸馏方法改进了top-k建议的使用，例",
"distribution given context c. The dotted line means the gra-",
"如通过基于排名的分布国[29]，其具有均匀分布E[24]和排名",
"ss提取识，大多数现有方法从排名者国[24,29.41]中",
"差异感知分布E[28]的混合物，从前k个项目中抽取样本。除",
"提取前k个结果中的知识，但是前k个排名结果对于排名者来",
"了教师的预测外，还可以通过提示回归E[24]和拓扑蒸馏E",
"说是耗时的。在本文中，我们通过基于KL散度直接对齐排名",
"[25]将教师中的潜在知识提炼到学生中。",
"者和检索者之间的softmax归一化预测，从排名者的预测中提",
"取排名顺序信息。为了减少优化的时间成本，我们提出了KL",
"3",
"COOPERATIVERETRIEVERANDRANKER",
"散度的渐近无偏估计，称为采样KL散度",
"在本节中，我们首先概述了在训练数据集①=（c.k上的",
"检索器和排名器通过统一的工作流程同时进行训练，其中",
"CoRR训练，该数据集由上下文c和正项k的对组成。上下文",
"两个模型可以相互增强。随着训练的进行，排名器变得越来越",
"表示除项目以外的所有信息，例如用户、交互历史记录、时间",
"精确，这反过来为检索器提供了更多的信息监督信号；同时，",
"和位置。之后，我们详细阐述了如何基于硬负采样来训练排名",
"随着检索器的改进，负样本将变得越来越困难，这有助于排名",
"者，以及如何基于知识提炼来训练排名者。",
"器的更高判别能力。整体框架如图1所示。整个过程可以参考",
"算法1",
"3.1Overview",
"3.2TrainingRanker withRetriever",
"为了可扩展的推荐，深度推荐器依赖于两个模型的协作：检",
"索模型（检索器）和排名模型（排名器）。检索器的目标是从",
"3.2.1LossFunction.在本文中，我们优化了排名器w。r。t抽",
"整个项目中高效地选择一小组潜在的积极项目。通常，检索",
"样log-softmax，它是log-softmax的渐近无偏估计。log-softmax",
"器由Me(i,c)=sim(Eo,(i),Eo,(c))，i表示。e。，相似性(e。g。",
"的使用是由于其与平均倒数排名（MRR）和归一化贴现累积",
"项嵌人Eo.(i）和上下文嵌入Ee,(c)之间的欧氏距离、内积和",
"收益（NDCG)E[3】的对数的密切关系以及出色的推荐性能F",
"余弦相似度)。因此，我们可以使用现成的人工神经网络，如",
"[33]。假设上下文c处的正项为k，则log-softmax目标w。r。",
"FAISS[23]和SCANN[13]，在亚线性时间检索最相似的项目",
"tk的公式如下：",
"作为候选。排名器旨在从检索结果中高精度地识别最佳项目，",
"(k,c) = Rg(k,c) -log (Zexp (Rg(i,c)",
"通常由富有表现力但耗时的网络R（i，c直接表示，以项目i",
"和上下文c作为输入",
"由于排名器中的耗时计算，在大多数情况下，传统的训练",
"= -log (Zexp (Rg(i,c) - Rg(k,c))",
"方法学会区分阳性和随机挑选的阴性。尽管简单，但排名者是",
"独立于猎犬的。在某些情况下，通过从检索者的前k个检索结",
"≤-log (ZIz0(Rg(i,c) - Rg(k,c))",
"果中区分阳性来训练排名者。尽管与检索者建立了联系，但排",
"名者很容易遭受假阴性问题，因为潜在的阳性肯定包含在前k",
"=log(rankr(k,))",
"",
"个检索结果中。此外，这些负面因素可能会给梯度带来很大",
"=logMRR",
"的偏差，使得优化可能会缓慢收敛，甚至很容易陷入局部最",
"优。为此，我们建议优化排名器w。r。t抽样的log-softmax[1]",
"，其中不等式由于exp()≥I>o()而成立",
"，基于重要性抽样的log-softmax的渐近无偏估计。抽样的log-",
"但是，优化log-softmax在计算上具有挑战性，因为梯度计",
"softmax依赖于从提案分布中随机抽取的样本。为了将排名器",
"算与项目数呈线性比例，即。e。",
"与检索器联系起来，我们建议根据自适应和可扩展的策略对",
"Vlg(k,c) = VRg(k,c) -Bi~Pa(-le) [VRg(i,c)]",
"检索器的底片进行采样",
"除了来自数据的监督之外，检索器的训练还可以由排名器",
"其中Ps（lc)表示具有参数softmax(R（,c)），i的整个项目I",
"基于知识提炼来指导，因为排名器被认为在排名项目上更精",
"的分类概率分布。e.Pg(lc)=Zerep(kaa）·为了提高效",
"exp(Rg(j.c)",
"确。关于来自推荐数据的监督，还利用采样的log-softmax目",
"标来优化检索器，并在排名器优化中使用相同的建议。关于从",
"率，通过从建议分布Q(lc)中绘制L样本S={o1,*+，oL)，使"
] |
{
"arxiv_id": "2206.14649",
"language": "zh",
"page_idx": 2
}
|
||
2206.14649-zh-4
|
2206.14649-zh-4.jpg
|
Cooperative Retriever and Ranker in Deep Recommenders
THEOREM 3.1. $D_{KL}\left(P_{\phi}^{\mathcal{S}}(\cdot|c)\parallel P_{\theta}^{\mathcal{S}}(\cdot|c)\right)convergestoD_{KL}\left(P_{\phi}(\cdot|c)\parallel\right.$
$P_{\theta }( \cdot | c) \right ) with$ probability 1 when $L= | S| \to \infty$ .
当用于学习检索器时，采样的 KL 散度基于$\overline{\mathrm{E}}$3.2.2 节中的采样器。下面我们讨论采样器的一些特殊情况，以进一步了解其通用性
3.3.3 Special Cases.在本部分中，我们研究了两个特殊的建
议$:Q(\cdot|c)=P_{\theta}(\cdot|c)=$softmax$(M_\theta(\cdot,c))$和$Q(\cdot|c)$是统一的。
$\bullet Q(\cdot|c)=P_{\theta}(\cdot|c):$在这种情况下，排名器通过采样的$\log$-
oftmax $\tilde{\ell}(k,c)=\log\frac{\exp(\Delta_{k}^{c})}{\sum_{i\in S}\exp(\Delta_{i}^{c})}$进行优化，其中$\Delta_k^{c}=R_{\phi}(k,c)$
$M_\theta(k,c)$。蒸馏损失可以简化如下：
CoROLLARY 3.2. Ifeach sample in S is drawn according to P $P_{\theta}(\cdot|c)$ $,concatenating\left\{\Delta_{i}^{c}|i\in\mathcal{S}\right\}asavector\Delta^{c}oflengthL,thenD_{KL}(P_{\phi}\parallel$ $P_{\theta})$ can be asymptotic-unbiasedly estimated by$\log L-H($softmax$(\Delta^{c}))$ ,where H(p) is the Shannon entropy of categorical distribution parameterized by p .
附录中提供了证明。推论表明，当从$P_\theta(\cdot|c)$精确绘制样本时，最小化 KL 散度等效于最大化分类分布的熵。通过排名器和检索器之间的 softmax 归一化预测差异来参数化分布
$\bullet Q(\cdot|c)$是统一的：在这种情况下，排名器通过采样的$\operatorname { og- softmax} \tilde{\ell } ( k, c)$ = $\log \frac {\exp ( R_{\phi }( k, c) ) }{\sum _{i\in S}\exp ( R_{\phi }( i, c) ) }$ 进行了优化。$P_{S}(\cdot|c|)$和$O_{\mathrm{s} }( \cdot | c| c)$ 可以简化：$P^{S}( i| c) \equiv \frac {\exp ( R_{\phi }( j, c) ) }{\exp ( i| c) }$ 和$P^{S}(i|c)=$ 和 $Q_{S}( \cdot | c)$ 可 以 简 化 : $P_{\phi }^{S}( j| c) = \frac {\exp ( R_{\phi }( j, c) ) }{\sum _{i\in S}\exp ( R_{\phi }( i, c) ) }$ 和 $P_{\theta }^{S}( j| c) =$ $\exp(M_{e}(i,c))$ $\frac{\exp(M_{\theta}(j,c))}{\sum_{i\in S}\exp(M_{\theta}(i,c))}$。最小化它们之间的 KL 差异等效于在排名者和检索者之间匹配随机样本集$\mathcal{S}$上的排名顺序
# 4 EXPERIMENTS
通过回答以下问题进行实验以验证所提出的 CoRR 的有效性： RQ1: Does CoRR outperforms conventional DRS and the existing
joint training methods?
RQ2: Could the adaptive sampler generate higher-quality nega-
tive items than the exact top-k sampling to help training?
RQ3: Does the ranking-order preserving distillation loss improve
the retriever?
由于 CoRR 与模型无关，因此为了证明 CoRR 的有效性，我们
在本文中将框架应用于一般推荐和顺序推荐
# 4.1 Experimental Settings
4.1.1 Datasets.如表$\mathbb{E}$1 所示，我们在四个现实世界的数据集上评估了我们的方法。数据集来自不同的域和平台，它们的大小和稀疏性差异很大。Gowalla 数据集包含用户在不同时间的位置的签到数据。淘宝数据集是阿里巴巴集团收集的一个大型工业数据集，其中包含用户行为，包括点击，购买，将商品添加到购物车以及商品偏爱。我们选择包含点击行为的最大子集。Amazon 数据集是 AmazonElectronics 产品评论的子集。MovieLens 数据集是一个经典的电影评级数据集，其中评级范围从 0。5 到 5。我们选择一个具有 10M 交互的子集进行实验。然后我们过滤掉所有数据集少于 10 次交互的用户和项目(位置/产品/电影)
对于一般推荐任务，每个用户的行为历史记录按比例 0 拆分为 train/valid/test。8/0。1/0。1。对于顺序推荐任务，给定用户的行为历史记录为$(i_1,i_2,\cdots,i_k,\cdots,i_n)$ ,目标是使用第一个$k$项目预测第$(k+1)$个项目。在所有实验中，我们使用
WWW'23, April 30-May 4, 2023, Austin, TX, USA
Table 1: Dataset Statistics.
<table>
<tbody>
<tr>
<th> </th>
<th>并 users 1</th>
<th>并 items</th>
<th>井 interactions</th>
</tr>
<tr>
<td>Amazon</td>
<td>9,280</td>
<td>6,066</td>
<td>158,979</td>
</tr>
<tr>
<td>Gowalla</td>
<td>29,859</td>
<td>40,989</td>
<td>1,027,464</td>
</tr>
<tr>
<td>MovieLens</td>
<td>66,958</td>
<td>10,682</td>
<td>5,857,041</td>
</tr>
<tr>
<td>Taobao</td>
<td>941,853</td>
<td>1101,236</td>
<td>63,721,355</td>
</tr>
</tbody>
</table>
$k=1,2,\cdots,n-3$为所有用户生成训练集，并分别在有效集和测试集中给出第一个$n-2$和$n-1$项的情况下预测下一个。此外，我们将所有数据集中用户行为序列的最大序列长度设置为20
4.1.2 Metric.评估中使用了三个常见的 top-k 指标，NDCGE [44] , Recall$\boxed{\mathrm{F}}[19,42]$和 MRR。Recall@k 表示目标项目在前 k 个项目中的案例比例。NDCG@k 在更高的级别上赋予更高的权重。MRR@k 表示目标项目倒数排名的平均值。这些指标的值越大，表示性能越好
4.1.3 Implementation Details.在本文中，一般推荐分别选择MF(矩阵分解)$\operatorname{E}[27,36]$和 DeepFM$\boxed{\mathbb{E}}[12]$作为检索器和排名器。在顺序推荐中，我们将 SASRec[26]设置为检索器，将DIN[52] 设置为排名器。在预测中，我们首先使用检索器从所有项目中检索前 100 个项目，然后按排名器对这些项目进行排名以获得最终输出。对于模型不可知论的广泛研究，还考虑了检索器和排名器的其他选择，如$\overline{\mathrm{E}}$4.3 节所示。源代码在 github $^{1}$中发布。
4.2 Comparison with Baselines
4.2.1 $Baselines Methods. 我 们 将 建 议 的 CoRR与 四 种 检 索 器$ 方法和三种现有的训练算法进行比较(i。e。ICC、RankFlow 和独立)。基线的详细信息可参考附录 B.1
4.2.2 Results.所有数据集上的实验结果如表$\color{red}{\mathrm{E}}$2 所示。结果表明，CoRR 优于传统的 DRS 和现有的联合训练方法，具有显著的优势。研究结果可总结如下：
发现 1: 独立训练方法的性能优于所有检索器，尽管检索器和排名器都是在独立方法中独立训练的。在一般建议中， Independent 方法在 Recall@10、NDCG@10 和 MRR@10 方面优于 5.39%, 5.16 % , 3.50 % 总体数据集的最佳检索器。类似地在顺序建议中，改进最多可达 6.23 % $,4.11\%,6.18\%$ 。这些结果表明，排名者可以补偿检索者的有限表现力，并且检索者和排名者可以捕获训练数据中的不同信息
发现 2: 所有三种联合训练方法 (ICC, RankFlow 和 CoRR) 几乎都优于 Independent 方法。其中最好的-CoRR在 NDCG@10方面对独立方法的平均相对改进分别是两个任务中的 25.62 % 和 18.57% 。这表明，对检索器和排名器之间的协作进行建模可以显着改善级联排名系统。它们之间更好的协作机制可以带来更多的改进
发现 3: 在三种联合训练方法中，CoRR 在项目推荐方面表现最佳。这表明了拟议的联合培训框架的优越性。与仅通过训练数据学习检索器和排名器的 ICC 相比，CoRR 在一般推荐和顺序推荐的所有数据集上分别在 NDCG@10 方面实现了 21.65 % 和 33.35% 的平均相对改进。与 RankFlow 相比，其中检索器和排名器相互联合训练和增强，CoRR 在这两个任务中的所有数据集上，在 NDCG@10 方面仍然实现了 15.72% and 11.62 % 的平均相对改进。结果表明，CoRR 的合作训练比 RankFlow
$\overline{^1\text{https://github.c}}$om/AngusHuang17/CoRR\_www
|
[
[
0.08660888671875,
0.0770263671875,
0.359375,
0.0770263671875,
0.359375,
0.08709716796875,
0.08660888671875,
0.08709716796875
],
[
0.6787109375,
0.07452392578125,
0.9111328125,
0.07574462890625,
0.9111328125,
0.0889892578125,
0.6787109375,
0.08837890625
],
[
0.10296630859375,
0.106689453125,
0.495849609375,
0.1060791015625,
0.495849609375,
0.12371826171875,
0.10296630859375,
0.125
],
[
0.62841796875,
0.10479736328125,
0.7998046875,
0.10479736328125,
0.7998046875,
0.11871337890625,
0.62841796875,
0.11871337890625
],
[
0.0882568359375,
0.1300048828125,
0.366943359375,
0.1300048828125,
0.366943359375,
0.14453125,
0.0882568359375,
0.14453125
],
[
0.64453125,
0.1395263671875,
0.69775390625,
0.1395263671875,
0.69775390625,
0.149658203125,
0.64453125,
0.149658203125
],
[
0.7158203125,
0.1395263671875,
0.76708984375,
0.1395263671875,
0.76708984375,
0.149658203125,
0.7158203125,
0.149658203125
],
[
0.76953125,
0.1427001953125,
0.77685546875,
0.1427001953125,
0.77685546875,
0.1458740234375,
0.76953125,
0.1458740234375
],
[
0.77392578125,
0.1395263671875,
0.86279296875,
0.1395263671875,
0.86279296875,
0.149658203125,
0.77392578125,
0.149658203125
],
[
0.10455322265625,
0.1546630859375,
0.48046875,
0.1546630859375,
0.48046875,
0.1685791015625,
0.10455322265625,
0.1685791015625
],
[
0.576171875,
0.154052734375,
0.62744140625,
0.1558837890625,
0.6259765625,
0.1685791015625,
0.57421875,
0.1666259765625
],
[
0.6611328125,
0.1558837890625,
0.6953125,
0.1558837890625,
0.6953125,
0.1685791015625,
0.6611328125,
0.1685791015625
],
[
0.72802734375,
0.1558837890625,
0.76318359375,
0.1558837890625,
0.76318359375,
0.1685791015625,
0.72802734375,
0.1685791015625
],
[
0.814453125,
0.1558837890625,
0.861328125,
0.1558837890625,
0.861328125,
0.1673583984375,
0.814453125,
0.1673583984375
],
[
0.57421875,
0.1673583984375,
0.6298828125,
0.1673583984375,
0.6298828125,
0.1817626953125,
0.57421875,
0.1817626953125
],
[
0.65283203125,
0.1673583984375,
0.69677734375,
0.1673583984375,
0.69677734375,
0.1817626953125,
0.65283203125,
0.1817626953125
],
[
0.7197265625,
0.1673583984375,
0.7646484375,
0.1673583984375,
0.7646484375,
0.1817626953125,
0.7197265625,
0.1817626953125
],
[
0.8037109375,
0.169189453125,
0.861328125,
0.169189453125,
0.861328125,
0.18115234375,
0.8037109375,
0.18115234375
],
[
0.564453125,
0.177978515625,
0.6357421875,
0.179931640625,
0.6357421875,
0.1944580078125,
0.564453125,
0.1925048828125
],
[
0.08660888671875,
0.1685791015625,
0.48193359375,
0.1685791015625,
0.48193359375,
0.1817626953125,
0.08660888671875,
0.1817626953125
],
[
0.65283203125,
0.179931640625,
0.69677734375,
0.179931640625,
0.69677734375,
0.1944580078125,
0.65283203125,
0.1944580078125
],
[
0.72119140625,
0.179931640625,
0.7646484375,
0.179931640625,
0.7646484375,
0.1944580078125,
0.72119140625,
0.1944580078125
],
[
0.80224609375,
0.177978515625,
0.861328125,
0.179931640625,
0.859375,
0.1944580078125,
0.80126953125,
0.1925048828125
],
[
0.08660888671875,
0.184326171875,
0.13232421875,
0.184326171875,
0.13232421875,
0.1956787109375,
0.08660888671875,
0.1956787109375
],
[
0.57666015625,
0.1932373046875,
0.625,
0.1932373046875,
0.625,
0.2052001953125,
0.57666015625,
0.2052001953125
],
[
0.646484375,
0.1944580078125,
0.69677734375,
0.1944580078125,
0.69677734375,
0.2052001953125,
0.646484375,
0.2052001953125
],
[
0.7060546875,
0.1944580078125,
0.76123046875,
0.1944580078125,
0.76123046875,
0.2052001953125,
0.7060546875,
0.2052001953125
],
[
0.79736328125,
0.1944580078125,
0.86279296875,
0.1944580078125,
0.86279296875,
0.205810546875,
0.79736328125,
0.205810546875
],
[
0.0882568359375,
0.205810546875,
0.4794921875,
0.205810546875,
0.4794921875,
0.216552734375,
0.0882568359375,
0.216552734375
],
[
0.515625,
0.21337890625,
0.91259765625,
0.2139892578125,
0.91259765625,
0.2291259765625,
0.515625,
0.2279052734375
],
[
0.0882568359375,
0.2197265625,
0.459228515625,
0.2197265625,
0.459228515625,
0.2298583984375,
0.0882568359375,
0.2298583984375
],
[
0.10296630859375,
0.231689453125,
0.483642578125,
0.231689453125,
0.483642578125,
0.24560546875,
0.10296630859375,
0.24560546875
],
[
0.5185546875,
0.2291259765625,
0.9140625,
0.2291259765625,
0.9140625,
0.242431640625,
0.5185546875,
0.242431640625
],
[
0.2279052734375,
0.24560546875,
0.273681640625,
0.24560546875,
0.273681640625,
0.255615234375,
0.2279052734375,
0.255615234375
],
[
0.51953125,
0.2442626953125,
0.9111328125,
0.2442626953125,
0.9111328125,
0.255126953125,
0.51953125,
0.255126953125
],
[
0.5185546875,
0.25830078125,
0.55224609375,
0.25830078125,
0.55224609375,
0.26953125,
0.5185546875,
0.26953125
],
[
0.08087158203125,
0.2474365234375,
0.484375,
0.2474365234375,
0.484375,
0.271484375,
0.08087158203125,
0.271484375
],
[
0.08660888671875,
0.267578125,
0.3095703125,
0.267578125,
0.3095703125,
0.281005859375,
0.08660888671875,
0.281005859375
],
[
0.51708984375,
0.27783203125,
0.9140625,
0.276611328125,
0.9140625,
0.290283203125,
0.51708984375,
0.291748046875
],
[
0.10296630859375,
0.28857421875,
0.48046875,
0.28857421875,
0.48046875,
0.3017578125,
0.10296630859375,
0.3017578125
],
[
0.51953125,
0.29345703125,
0.91259765625,
0.29345703125,
0.91259765625,
0.3037109375,
0.51953125,
0.3037109375
],
[
0.0882568359375,
0.302978515625,
0.484375,
0.302978515625,
0.484375,
0.31640625,
0.0882568359375,
0.31640625
],
[
0.5185546875,
0.30615234375,
0.91259765625,
0.30615234375,
0.91259765625,
0.3193359375,
0.5185546875,
0.3193359375
],
[
0.08660888671875,
0.31640625,
0.495849609375,
0.31640625,
0.495849609375,
0.330078125,
0.08660888671875,
0.330078125
],
[
0.515625,
0.318115234375,
0.91259765625,
0.3193359375,
0.91259765625,
0.333251953125,
0.515625,
0.33203125
],
[
0.08660888671875,
0.330810546875,
0.48193359375,
0.330810546875,
0.48193359375,
0.3447265625,
0.08660888671875,
0.3447265625
],
[
0.51953125,
0.335205078125,
0.6650390625,
0.335205078125,
0.6650390625,
0.34521484375,
0.51953125,
0.34521484375
],
[
0.0849609375,
0.3447265625,
0.19287109375,
0.34521484375,
0.19287109375,
0.356689453125,
0.0849609375,
0.35595703125
],
[
0.51953125,
0.354736328125,
0.91259765625,
0.354736328125,
0.91259765625,
0.36865234375,
0.51953125,
0.36865234375
],
[
0.10296630859375,
0.3642578125,
0.4794921875,
0.3642578125,
0.4794921875,
0.37744140625,
0.10296630859375,
0.37744140625
],
[
0.51708984375,
0.36865234375,
0.91259765625,
0.36865234375,
0.91259765625,
0.3818359375,
0.51708984375,
0.3818359375
],
[
0.08660888671875,
0.37890625,
0.48046875,
0.37890625,
0.48046875,
0.392578125,
0.08660888671875,
0.392578125
],
[
0.5185546875,
0.3818359375,
0.91259765625,
0.3818359375,
0.91259765625,
0.395751953125,
0.5185546875,
0.395751953125
],
[
0.0882568359375,
0.39453125,
0.4345703125,
0.39453125,
0.4345703125,
0.40478515625,
0.0882568359375,
0.40478515625
],
[
0.51953125,
0.397705078125,
0.91259765625,
0.397705078125,
0.91259765625,
0.40771484375,
0.51953125,
0.40771484375
],
[
0.101318359375,
0.406005859375,
0.48046875,
0.40478515625,
0.48046875,
0.419189453125,
0.101318359375,
0.42041015625
],
[
0.51953125,
0.410888671875,
0.9111328125,
0.410888671875,
0.9111328125,
0.421630859375,
0.51953125,
0.421630859375
],
[
0.08251953125,
0.421630859375,
0.484375,
0.421630859375,
0.484375,
0.4462890625,
0.08251953125,
0.4462890625
],
[
0.256591796875,
0.419189453125,
0.325927734375,
0.419189453125,
0.325927734375,
0.429931640625,
0.256591796875,
0.429931640625
],
[
0.5185546875,
0.423583984375,
0.9111328125,
0.423583984375,
0.9111328125,
0.4375,
0.5185546875,
0.4375
],
[
0.080078125,
0.43994140625,
0.48779296875,
0.442626953125,
0.48779296875,
0.47021484375,
0.080078125,
0.468505859375
],
[
0.30810546875,
0.44140625,
0.3857421875,
0.439453125,
0.3857421875,
0.45263671875,
0.30810546875,
0.455078125
],
[
0.5185546875,
0.438232421875,
0.9140625,
0.438232421875,
0.9140625,
0.451904296875,
0.5185546875,
0.451904296875
],
[
0.5185546875,
0.448974609375,
0.58251953125,
0.451904296875,
0.5810546875,
0.466552734375,
0.51708984375,
0.46337890625
],
[
0.08416748046875,
0.46728515625,
0.483642578125,
0.4609375,
0.483642578125,
0.48291015625,
0.08416748046875,
0.4892578125
],
[
0.1004638671875,
0.464111328125,
0.176513671875,
0.46337890625,
0.176513671875,
0.474853515625,
0.1004638671875,
0.47607421875
],
[
0.5185546875,
0.479248046875,
0.5439453125,
0.479248046875,
0.5439453125,
0.490478515625,
0.5185546875,
0.490478515625
],
[
0.5400390625,
0.479736328125,
0.56640625,
0.479736328125,
0.56640625,
0.488037109375,
0.5400390625,
0.488037109375
],
[
0.560546875,
0.477783203125,
0.7861328125,
0.477783203125,
0.7861328125,
0.4912109375,
0.560546875,
0.4912109375
],
[
0.08660888671875,
0.48486328125,
0.384033203125,
0.48486328125,
0.384033203125,
0.498779296875,
0.08660888671875,
0.498779296875
],
[
0.51953125,
0.49755859375,
0.560546875,
0.496826171875,
0.560546875,
0.5048828125,
0.51953125,
0.505859375
],
[
0.560546875,
0.494384765625,
0.91259765625,
0.49560546875,
0.91259765625,
0.5087890625,
0.560546875,
0.50830078125
],
[
0.11358642578125,
0.51025390625,
0.2442626953125,
0.51025390625,
0.2442626953125,
0.5234375,
0.11358642578125,
0.5234375
],
[
0.51953125,
0.51123046875,
0.9111328125,
0.51123046875,
0.9111328125,
0.521484375,
0.51953125,
0.521484375
],
[
0.51953125,
0.5244140625,
0.78173828125,
0.5244140625,
0.78173828125,
0.53515625,
0.51953125,
0.53515625
],
[
0.08660888671875,
0.51318359375,
0.11517333984375,
0.51318359375,
0.11517333984375,
0.521484375,
0.08660888671875,
0.521484375
],
[
0.0882568359375,
0.53076171875,
0.4697265625,
0.53076171875,
0.4697265625,
0.54150390625,
0.0882568359375,
0.54150390625
],
[
0.0849609375,
0.54443359375,
0.48193359375,
0.54541015625,
0.48193359375,
0.56005859375,
0.0849609375,
0.5595703125
],
[
0.5185546875,
0.54541015625,
0.56201171875,
0.54541015625,
0.56201171875,
0.55615234375,
0.5185546875,
0.55615234375
],
[
0.560546875,
0.54443359375,
0.91259765625,
0.54443359375,
0.91259765625,
0.55810546875,
0.560546875,
0.55810546875
],
[
0.125,
0.5595703125,
0.26953125,
0.5595703125,
0.26953125,
0.57275390625,
0.125,
0.57275390625
],
[
0.5185546875,
0.55810546875,
0.91259765625,
0.55810546875,
0.91259765625,
0.5712890625,
0.5185546875,
0.5712890625
],
[
0.12335205078125,
0.57373046875,
0.48046875,
0.57373046875,
0.48046875,
0.5869140625,
0.12335205078125,
0.5869140625
],
[
0.51953125,
0.57373046875,
0.73046875,
0.57373046875,
0.73046875,
0.583984375,
0.51953125,
0.583984375
],
[
0.53369140625,
0.583984375,
0.9140625,
0.58544921875,
0.9140625,
0.599609375,
0.53369140625,
0.5986328125
],
[
0.08660888671875,
0.57470703125,
0.12744140625,
0.57470703125,
0.12744140625,
0.5859375,
0.08660888671875,
0.5859375
],
[
0.1258544921875,
0.58837890625,
0.47216796875,
0.58837890625,
0.47216796875,
0.6015625,
0.1258544921875,
0.6015625
],
[
0.0849609375,
0.60107421875,
0.48046875,
0.60107421875,
0.48046875,
0.6142578125,
0.0849609375,
0.6142578125
],
[
0.51708984375,
0.5986328125,
0.91259765625,
0.60107421875,
0.91259765625,
0.6142578125,
0.51708984375,
0.6123046875
],
[
0.1258544921875,
0.6162109375,
0.2091064453125,
0.6162109375,
0.2091064453125,
0.626953125,
0.1258544921875,
0.626953125
],
[
0.51708984375,
0.61279296875,
0.9140625,
0.61279296875,
0.9140625,
0.626953125,
0.51708984375,
0.626953125
],
[
0.08660888671875,
0.630859375,
0.48193359375,
0.630859375,
0.48193359375,
0.64453125,
0.08660888671875,
0.64453125
],
[
0.51708984375,
0.62548828125,
0.9140625,
0.62744140625,
0.9140625,
0.6416015625,
0.51708984375,
0.6396484375
],
[
0.08660888671875,
0.646484375,
0.365234375,
0.646484375,
0.365234375,
0.6572265625,
0.08660888671875,
0.6572265625
],
[
0.5185546875,
0.6416015625,
0.9140625,
0.6416015625,
0.9140625,
0.65478515625,
0.5185546875,
0.65478515625
],
[
0.51953125,
0.65478515625,
0.9140625,
0.65478515625,
0.9140625,
0.66845703125,
0.51953125,
0.66845703125
],
[
0.1258544921875,
0.67041015625,
0.313720703125,
0.671875,
0.313720703125,
0.68603515625,
0.1258544921875,
0.68505859375
],
[
0.515625,
0.66845703125,
0.76708984375,
0.66943359375,
0.76708984375,
0.68310546875,
0.515625,
0.681640625
],
[
0.53515625,
0.68310546875,
0.9140625,
0.68310546875,
0.9140625,
0.6962890625,
0.53515625,
0.6962890625
],
[
0.08660888671875,
0.67236328125,
0.13232421875,
0.67236328125,
0.13232421875,
0.68310546875,
0.08660888671875,
0.68310546875
],
[
0.08660888671875,
0.68798828125,
0.4794921875,
0.68798828125,
0.4794921875,
0.7021484375,
0.08660888671875,
0.7021484375
],
[
0.5185546875,
0.6962890625,
0.91259765625,
0.6962890625,
0.91259765625,
0.71044921875,
0.5185546875,
0.71044921875
],
[
0.08905029296875,
0.705078125,
0.4794921875,
0.705078125,
0.4794921875,
0.71533203125,
0.08905029296875,
0.71533203125
],
[
0.5185546875,
0.71044921875,
0.9140625,
0.71044921875,
0.9140625,
0.7236328125,
0.5185546875,
0.7236328125
],
[
0.08660888671875,
0.71630859375,
0.4794921875,
0.71630859375,
0.4794921875,
0.72998046875,
0.08660888671875,
0.72998046875
],
[
0.08660888671875,
0.72998046875,
0.48046875,
0.72998046875,
0.48046875,
0.74365234375,
0.08660888671875,
0.74365234375
],
[
0.51953125,
0.72607421875,
0.9111328125,
0.72607421875,
0.9111328125,
0.736328125,
0.51953125,
0.736328125
],
[
0.0849609375,
0.74365234375,
0.48046875,
0.744140625,
0.48046875,
0.75830078125,
0.0849609375,
0.7568359375
],
[
0.51953125,
0.73779296875,
0.9140625,
0.73779296875,
0.9140625,
0.751953125,
0.51953125,
0.751953125
],
[
0.51953125,
0.75244140625,
0.62109375,
0.75244140625,
0.62109375,
0.76318359375,
0.51953125,
0.76318359375
],
[
0.0882568359375,
0.76025390625,
0.4794921875,
0.76025390625,
0.4794921875,
0.77099609375,
0.0882568359375,
0.77099609375
],
[
0.53369140625,
0.7646484375,
0.9140625,
0.76513671875,
0.9140625,
0.77978515625,
0.53369140625,
0.77880859375
],
[
0.08905029296875,
0.7734375,
0.47802734375,
0.7734375,
0.47802734375,
0.7841796875,
0.08905029296875,
0.7841796875
],
[
0.51708984375,
0.77783203125,
0.9140625,
0.77880859375,
0.9140625,
0.79345703125,
0.51708984375,
0.79248046875
],
[
0.08660888671875,
0.787109375,
0.48046875,
0.787109375,
0.48046875,
0.79736328125,
0.08660888671875,
0.79736328125
],
[
0.5185546875,
0.79345703125,
0.91259765625,
0.79345703125,
0.91259765625,
0.806640625,
0.5185546875,
0.806640625
],
[
0.08660888671875,
0.7998046875,
0.4794921875,
0.7998046875,
0.4794921875,
0.81298828125,
0.08660888671875,
0.81298828125
],
[
0.0882568359375,
0.81494140625,
0.4794921875,
0.81494140625,
0.4794921875,
0.82568359375,
0.0882568359375,
0.82568359375
],
[
0.51953125,
0.8095703125,
0.91259765625,
0.8095703125,
0.91259765625,
0.8193359375,
0.51953125,
0.8193359375
],
[
0.51708984375,
0.8212890625,
0.91259765625,
0.8212890625,
0.91259765625,
0.83544921875,
0.51708984375,
0.83544921875
],
[
0.0882568359375,
0.8291015625,
0.176513671875,
0.8291015625,
0.176513671875,
0.84033203125,
0.0882568359375,
0.84033203125
],
[
0.5185546875,
0.83544921875,
0.91259765625,
0.83544921875,
0.91259765625,
0.8486328125,
0.5185546875,
0.8486328125
],
[
0.10296630859375,
0.84130859375,
0.48193359375,
0.84130859375,
0.48193359375,
0.85498046875,
0.10296630859375,
0.85498046875
],
[
0.08660888671875,
0.85498046875,
0.48046875,
0.85498046875,
0.48046875,
0.86865234375,
0.08660888671875,
0.86865234375
],
[
0.5185546875,
0.85107421875,
0.9111328125,
0.85107421875,
0.9111328125,
0.861328125,
0.5185546875,
0.861328125
],
[
0.5146484375,
0.86279296875,
0.9111328125,
0.86279296875,
0.9111328125,
0.876953125,
0.5146484375,
0.876953125
],
[
0.08660888671875,
0.869140625,
0.48046875,
0.869140625,
0.48046875,
0.88330078125,
0.08660888671875,
0.88330078125
],
[
0.0931396484375,
0.8818359375,
0.48046875,
0.88330078125,
0.48046875,
0.896484375,
0.0931396484375,
0.89599609375
],
[
0.51708984375,
0.88330078125,
0.74853515625,
0.88330078125,
0.74853515625,
0.896484375,
0.51708984375,
0.896484375
]
] |
[
"Cooperative Retriever and Ranker in Deep Recommenders",
"www'23, April 30-May 4, 2023, Austin, TX, USA",
"THEOREM 3.1. DkL(PS(lc) II PS(lc) converges to DkL(Pg(lc) ",
"Table 1: Dataset Statistics.",
"Pe(-lc)) with probability 1 when L = [S[ - c0",
"#users",
"#items",
"",
"#interactions",
"当用于学习检索器时，采样的KL散度基于E3.2.2节中的采",
"Amazon",
"9,280",
"6,066",
"158,979",
"Gowalla",
"29,859",
"40,989",
"1,027,464",
"MovieLens",
"样器。下面我们讨论采样器的一些特殊情况，以进一步了解其",
"66,958",
"10,682",
"5,857,041",
"通用性",
"Taobao",
"941,853",
"1101,236",
"63,721,355",
"3.3.3SpecialCases，在本部分中，我们研究了两个特殊的建",
"k=1,2,.…，n-3为所有用户生成训练集，并分别在有效集和",
"议：Q(c）=Pe（c）=softmax(Mo（）和Q（c）是统一一的。",
"Q(-lc）=Po(-lc）：在这种情况下，排名器通过采样的log-",
"测试集中给出第一个n-2和n-1项的情况下预测下一个。此",
"exp(A)",
"外，我们将所有数据集中用户行为序列的最大序列长度设置",
"为20",
"softmax(k,c)=log resep(A进行优化，其中A=Rg(k,c)-",
"M(k,c）。蒸馏损失可以简化如下：",
"4.1.2Metric：评估中使用了三个常见的top-k指标，NDCGE",
"COROLLARY 3.2. If each sample in S is drawn according to Po(lc)",
"[44]】，RecallE[19,42]和MRR。Recall@k表示目标项目在前k",
" concatenating (Ali S) asa vectoroflengthL,thenDkL(P ll",
"个项目中的案例比例。NDCG@k在更高的级别上赋予更高的",
"Po) can be asymptotic-unbiasedly estimated bylog L-H(softmax(°))",
"权重。MRR@k表示目标项目倒数排名的平均值。这些指标的",
", where H(p) is the Shannon entropy of categorical distribution pa-",
"值越大，表示性能越好",
"rameterizedbyp.",
"4.1.3ImplementationDetails.在本文中，一般推荐分别选择",
"附录中提供了证明。推论表明，当从Pe(lc)精确绘制样本",
"MF（矩阵分解)E[27,36]和DeepFME[12]作为检索器和排",
"时，最小化KL散度等效于最大化分类分布的嫡。通过排名器",
"名器。在顺序推荐中，我们将SASRec[26]设置为检索器，将",
"和检索器之间的softmax归一化预测差异来参数化分布",
"DIN[52]设置为排名器。在预测中，我们首先使用检索器从所",
"·Q(-Ic）是统一的：在这种情况下，排名器通过采样的",
"有项目中检索前100个项目，然后按排名器对这些项目进行排",
"log-softmax i(k,c) = log esep(ke()进行了优化。 Ps(lc)",
"exp(Rg(k,c))",
"名以获得最终输出。对于模型不可知论的广泛研究，还考虑了",
"和 Qs(le) 可以简化：: P%(l)= rese((k() 和 Pg(l) =",
"exp(Rg(j,c))",
"检索器和排名器的其他选择，如E4.3节所示。源代码在github",
"1中发布。",
"M。最小化它们之间的KL差异等效于在排名者",
"exp(Mg(j,c))",
"4.2",
"",
"ComparisonwithBaselines",
"和检索者之间匹配随机样本集S上的排名顺序",
"4.2.1",
"BaselinesMethods.我们将建议的CoRR与四种检索器",
"EXPERIMENTS",
"方法和三种现有的训练算法进行比较（i。e。、ICC、RankFlow",
"和独立)。基线的详细信息可参考附录B.1",
"4",
"通过回答以下问题进行实验以验证所提出的CoRR的有效性：",
"RQ1: Does CoRR outperforms conventional DRS and the existing",
"4.2.2",
"Results.所有数据集上的实验结果如表E2所示。结果",
"joint training methods?",
"表明，CoRR优于传统的DRS和现有的联合训练方法，具有显",
"Could the adaptive sampler generate higher-quality nega",
"著的优势。研究结果可总结如下：",
"发现1：独立训练方法的性能优于所有检索器，尽管检索",
"RQ2:",
"tive items than the exact top-k sampling to help training?",
"RQ3:Does the ranking-order preserving distillation loss improve",
"器和排名器都是在独立方法中独立训练的。在一般建议中，",
"the retriever?",
"Independent方法在Recall@10、NDCG@10和MRR@10方面优",
"由于CoRR与模型无关，因此为了证明CoRR的有效性，我们",
"于5.39%，5.16%，3.50%总体数据集的最佳检索器。类似地，",
"在本文中将框架应用于一般推荐和顺序推荐",
"在顺序建议中，改进最多可达6.23%，4.11%，6.18%。这些结",
"果表明，排名者可以补偿检索者的有限表现力，并且检索者和",
"Experimental Settings",
"排名者可以捕获训练数据中的不同信息",
"发现2：所有三种联合训练方法（ICC，RankFlow和CoRR）",
"4.1",
"4.1.1Datasets.如表E1所示，我们在四个现实世界的数据集",
"几乎都优于Independent方法。其中最好的-CoRR在NDCG@10",
"上评估了我们的方法。数据集来自不同的域和平台，它们的大",
"方面对独立方法的平均相对改进分别是两个任务中的25.62%",
"小和稀疏性差异很大。Gowalla数据集包含用户在不同时间的",
"位置的签到数据。淘宝数据集是阿里巴巴集团收集的一个大",
"和18.57%。这表明，对检索器和排名器之间的协作进行建模",
"型工业数据集，其中包含用户行为，包括点击，购买，将商品",
"可以显着改善级联排名系统。它们之间更好的协作机制可以",
"带来更多的改进",
"添加到购物车以及商品偏爱。我们选择包含点击行为的最大",
"发现3：在三种联合训练方法中，CoRR在项目推荐方面表",
"子集。Amazon数据集是AmazonElectronics产品评论的子集，",
"现最佳。这表明了拟议的联合培训框架的优越性。与仅通过训",
"MovieLens数据集是一个经典的电影评级数据集，其中评级范",
"练数据学习检索器和排名器的ICC相比，CoRR在一般推荐和",
"围从0。5到5。我们选择一个具有10M交互的子集进行实验。",
"然后我们过滤掉所有数据集少于10次交互的用户和项目（位",
"顺序推荐的所有数据集上分别在NDCG@10方面实现了21.65",
"%和33.35%的平均相对改进。与RankFlow相比，其中检索",
"置/产品/电影）",
"器和排名器相互联合训练和增强，CoRR在这两个任务中的所",
"对于一般推荐任务，每个用户的行为历史记录按比例0拆",
"分为train/valid/test。8/0。1/0。1。对于顺序推荐任务，给定",
"有数据集上，在NDCG@10方面仍然实现了15.72%and11.62",
"%的平均相对改进。结果表明，CoRR的合作训练比RankFlow",
"用户的行为历史记录为（it,i2ik，，in），目标是使用第",
"一个k项目预测第（k+1）个项目。在所有实验中，我们使用",
"1https://github.com/AngusHuang17/CoRR_wwW"
] |
{
"arxiv_id": "2206.14649",
"language": "zh",
"page_idx": 4
}
|
||
1805.11729-zh-1
|
1805.11729-zh-1.jpg
|
$HeadOn{:}$ Real-time Reenactment of Human Portrait Videos
JUSTUS THIES, Technical University of Munich MICHAEL ZOLLHOFER, Stanford University CHRISTIAN THEOBALT, Max-Planck-Institute for Informatics
MARC STAMMINGER, University of Erlangen-Nuremberg
MATTHIAS NIESSNER, Technical University of Munich
<FigureHere>
Fig. 1. Our novel HeadOn approach enables real-time reenactment of upper body motion, head pose, face expression, and eye gaze in human portrait videos. For synthesis of new photo-realistic video content, we employ a novel video based rendering approach that builds on top of a fully controllable 3D actor model. The person-specific model is constructed from a short RGB-D calibration sequence and is driven by a real-time torso and face tracker.
We propose HeadOn, the first real-time source-to-target reenactment approach for complete human portrait videos that enables transfer of torso and $\hat{\text{head motion, face expression, and eye gaze. Given a short RGB-D video of}}$ the target actor, we automatically construct a personalized geometry proxy that embeds a parametric head, eye, and kinematic torso model. A novel realtime reenactment algorithm employs this proxy to photo-realistically map the captured motion from the source actor to the target actor. On top of the coarse geometric proxy, we propose a video-based rendering technique that composites the modified target portrait video via view- and pose-dependent texturing, and creates photo-realistic imagery of the target actor under novel torso and head poses, facial expressions, and gaze directions. To this end, we propose a robust tracking of the face and torso of the source actor. We extensively evaluate our approach and show significant improvements in enabling much greater flexibility in creating realistic reenacted output videos.
# CCS Concepts: · Computing methodologies $\to$Computer vision; Computer graphics;
Additional Key Words and Phrases: Reenactment, Face tracking, Video-based
Rendering, Real-time
1 $INTRODUCTION$
Reenactment approaches aim to transfer the motion of a source actor to an image or video of a target actor. Very recently, facial reenactment methods have been successfully employed to achieve highly-realistic manipulations of facial expressions based on commodity video data[ Averbuch-Elor et al. 2017; Suwajanakorn et al. 2017; Thies et al. 2015,2016,2018; Vlasic et al. 2005]. Rather than animating a virtual, stylized avatar (e.g., as used in video games), these algorithms replace the face region of a person with a synthetic re-rendering, or modify the target image/video under guidance of a 3D face model. This enables changing the expression of a target person and creating a manipulated output video that suggests different content; e.g., a person who is sitting still could appear as if he/she is talking. Modern reenactment approaches achieve highly believable results, even in real-time, and have enjoyed wide media coverage due to the interest in general movie and video editing [Vlasic et al. 2005], teleconferencing [Thies et al. 2018], reactive profile pictures [Averbuch-Elor et al. 2017],or visual dubbing of foreign language movies [Garrido et al. 2015].
Even though current facial reenactment results are impressive, they are still fundamentally limited in the type of manipulations they enable. For instance, these approaches are only able to modify facial expressions, whereas the rigid pose of the head, including its orientation, remains unchanged and does not follow the input video. Thus, only subtle changes, such as opening the mouth or adding wrinkles on the forehead are realized, which severely limits the applicability to video editing, where the control of the pose of the target person is also required. Furthermore, without joint modification of the head pose, the modified facial expressions often seem out-of-place, since they do not well align with visual pauses in the body and head motion; as noted by Suwajanakorn et al. [2017] this significantly restricts the applicability in teleconferencing scenarios.
ACM Reference Format:
Justus Thies, Michael Zollhöfer, Christian Theobalt, Marc Stamminger,
and Matthias Nießner. 2018. HeadOn: Real-time Reenactment of Human Portrait Videos. $ACM$ Trans. Graph. 37, 4, Article 164 (August 2018), 13 pages. https://doi.org/10.1145/3197517.3201350
Authors' addresses: Justus Thies, Technical University of Munich, justus.thies@tum.de; Michael Zollhöfer, Stanford University, zollhoefer@cs.stanford.edu; Christian Theobalt, Max-Planck-Institute for Informatics, theobalt@mpi-inf.mpg.de; Marc Stamminger, University of Erlangen-Nuremberg, marc.stamminger@fau.de; Matthias Nießner, Technical University of Munich, niessner@tum.de.
$\textcircled{\circ}2018$ Association for Computing Machinery.
This is the author's version of the work. It is posted here for your personal use. Not for redistribution. The definitive Version of Record was published in ACM Transactions on $Graphics$, https://doi.org/10.1145/3197517.3201350.
ACM Trans. Graph., Vol. 37, No. 4, Article 164. Publication date: August 2018
|
[
[
0.08563232421875,
0.1318359375,
0.767578125,
0.1318359375,
0.767578125,
0.149658203125,
0.08563232421875,
0.149658203125
],
[
0.0831298828125,
0.1650390625,
0.424072265625,
0.1650390625,
0.424072265625,
0.1787109375,
0.0831298828125,
0.1787109375
],
[
0.0831298828125,
0.1817626953125,
0.420654296875,
0.1817626953125,
0.420654296875,
0.1953125,
0.0831298828125,
0.1953125
],
[
0.0814208984375,
0.196533203125,
0.53662109375,
0.19775390625,
0.53662109375,
0.2120361328125,
0.0814208984375,
0.21142578125
],
[
0.0814208984375,
0.2120361328125,
0.5087890625,
0.215576171875,
0.5087890625,
0.2298583984375,
0.0814208984375,
0.226806640625
],
[
0.0814208984375,
0.2286376953125,
0.48779296875,
0.2298583984375,
0.48779296875,
0.24462890625,
0.0814208984375,
0.243408203125
],
[
0.0814208984375,
0.39013671875,
0.91748046875,
0.39013671875,
0.91748046875,
0.40380859375,
0.0814208984375,
0.40380859375
],
[
0.07977294921875,
0.402099609375,
0.91748046875,
0.402099609375,
0.91748046875,
0.4150390625,
0.07977294921875,
0.4150390625
],
[
0.07977294921875,
0.41455078125,
0.84228515625,
0.41455078125,
0.84228515625,
0.427490234375,
0.07977294921875,
0.427490234375
],
[
0.0814208984375,
0.432861328125,
0.483642578125,
0.432861328125,
0.483642578125,
0.446533203125,
0.0814208984375,
0.446533203125
],
[
0.5380859375,
0.431640625,
0.6630859375,
0.431640625,
0.6630859375,
0.4453125,
0.5380859375,
0.4453125
],
[
0.0814208984375,
0.44482421875,
0.483642578125,
0.44482421875,
0.483642578125,
0.457763671875,
0.0814208984375,
0.457763671875
],
[
0.5146484375,
0.448974609375,
0.91748046875,
0.448974609375,
0.91748046875,
0.4619140625,
0.5146484375,
0.4619140625
],
[
0.0814208984375,
0.4560546875,
0.484375,
0.4560546875,
0.484375,
0.4697265625,
0.0814208984375,
0.4697265625
],
[
0.51708984375,
0.463134765625,
0.91455078125,
0.463134765625,
0.91455078125,
0.473876953125,
0.51708984375,
0.473876953125
],
[
0.0814208984375,
0.4697265625,
0.480224609375,
0.4697265625,
0.480224609375,
0.479736328125,
0.0814208984375,
0.479736328125
],
[
0.0831298828125,
0.48095703125,
0.48193359375,
0.48095703125,
0.48193359375,
0.4912109375,
0.0831298828125,
0.4912109375
],
[
0.51708984375,
0.4755859375,
0.916015625,
0.4755859375,
0.916015625,
0.486328125,
0.51708984375,
0.486328125
],
[
0.0831298828125,
0.4921875,
0.480224609375,
0.4921875,
0.480224609375,
0.5029296875,
0.0831298828125,
0.5029296875
],
[
0.51708984375,
0.488037109375,
0.91455078125,
0.488037109375,
0.91455078125,
0.498779296875,
0.51708984375,
0.498779296875
],
[
0.0831298828125,
0.50439453125,
0.478515625,
0.50439453125,
0.478515625,
0.51416015625,
0.0831298828125,
0.51416015625
],
[
0.5146484375,
0.498779296875,
0.91748046875,
0.49951171875,
0.91748046875,
0.51318359375,
0.5146484375,
0.5126953125
],
[
0.0814208984375,
0.515625,
0.48193359375,
0.515625,
0.48193359375,
0.525390625,
0.0814208984375,
0.525390625
],
[
0.51708984375,
0.51318359375,
0.916015625,
0.51318359375,
0.916015625,
0.52392578125,
0.51708984375,
0.52392578125
],
[
0.0814208984375,
0.525390625,
0.483642578125,
0.525390625,
0.483642578125,
0.5390625,
0.0814208984375,
0.5390625
],
[
0.515625,
0.52490234375,
0.91748046875,
0.52490234375,
0.91748046875,
0.5380859375,
0.515625,
0.5380859375
],
[
0.0814208984375,
0.5380859375,
0.48193359375,
0.5380859375,
0.48193359375,
0.548828125,
0.0814208984375,
0.548828125
],
[
0.51708984375,
0.5390625,
0.916015625,
0.5390625,
0.916015625,
0.5498046875,
0.51708984375,
0.5498046875
],
[
0.07977294921875,
0.548828125,
0.48193359375,
0.548828125,
0.48193359375,
0.5625,
0.07977294921875,
0.5625
],
[
0.515625,
0.55029296875,
0.91748046875,
0.55029296875,
0.91748046875,
0.56396484375,
0.515625,
0.56396484375
],
[
0.07977294921875,
0.56005859375,
0.483642578125,
0.56005859375,
0.483642578125,
0.57373046875,
0.07977294921875,
0.57373046875
],
[
0.5146484375,
0.5625,
0.91748046875,
0.5634765625,
0.91748046875,
0.57666015625,
0.5146484375,
0.576171875
],
[
0.07977294921875,
0.57177734375,
0.483642578125,
0.57177734375,
0.483642578125,
0.5849609375,
0.07977294921875,
0.5849609375
],
[
0.515625,
0.576171875,
0.91748046875,
0.576171875,
0.91748046875,
0.5888671875,
0.515625,
0.5888671875
],
[
0.0814208984375,
0.5830078125,
0.450927734375,
0.5830078125,
0.450927734375,
0.5966796875,
0.0814208984375,
0.5966796875
],
[
0.515625,
0.58837890625,
0.91748046875,
0.58837890625,
0.91748046875,
0.6015625,
0.515625,
0.6015625
],
[
0.0814208984375,
0.60107421875,
0.483642578125,
0.60107421875,
0.483642578125,
0.6142578125,
0.0814208984375,
0.6142578125
],
[
0.515625,
0.60107421875,
0.916015625,
0.60107421875,
0.916015625,
0.6142578125,
0.515625,
0.6142578125
],
[
0.0831298828125,
0.6142578125,
0.175537109375,
0.6142578125,
0.175537109375,
0.6259765625,
0.0831298828125,
0.6259765625
],
[
0.51904296875,
0.615234375,
0.91455078125,
0.615234375,
0.91455078125,
0.6259765625,
0.51904296875,
0.6259765625
],
[
0.515625,
0.6259765625,
0.91748046875,
0.6259765625,
0.91748046875,
0.6396484375,
0.515625,
0.6396484375
],
[
0.0814208984375,
0.6318359375,
0.483642578125,
0.6318359375,
0.483642578125,
0.6455078125,
0.0814208984375,
0.6455078125
],
[
0.07977294921875,
0.64453125,
0.1956787109375,
0.6455078125,
0.1956787109375,
0.65673828125,
0.07977294921875,
0.65625
],
[
0.515625,
0.6396484375,
0.91748046875,
0.6396484375,
0.91748046875,
0.65283203125,
0.515625,
0.65283203125
],
[
0.51708984375,
0.65185546875,
0.916015625,
0.65185546875,
0.916015625,
0.6650390625,
0.51708984375,
0.6650390625
],
[
0.0814208984375,
0.6630859375,
0.2275390625,
0.6630859375,
0.2275390625,
0.6767578125,
0.0814208984375,
0.6767578125
],
[
0.51904296875,
0.6650390625,
0.68701171875,
0.6650390625,
0.68701171875,
0.67578125,
0.51904296875,
0.67578125
],
[
0.0831298828125,
0.67578125,
0.48193359375,
0.67578125,
0.48193359375,
0.68603515625,
0.0831298828125,
0.68603515625
],
[
0.5322265625,
0.677734375,
0.916015625,
0.677734375,
0.916015625,
0.6884765625,
0.5322265625,
0.6884765625
],
[
0.0831298828125,
0.68701171875,
0.3779296875,
0.68701171875,
0.3779296875,
0.69775390625,
0.0831298828125,
0.69775390625
],
[
0.3779296875,
0.68701171875,
0.480224609375,
0.68701171875,
0.480224609375,
0.69775390625,
0.3779296875,
0.69775390625
],
[
0.515625,
0.689453125,
0.91748046875,
0.689453125,
0.91748046875,
0.703125,
0.515625,
0.703125
],
[
0.07977294921875,
0.69775390625,
0.483642578125,
0.69775390625,
0.483642578125,
0.71142578125,
0.07977294921875,
0.71142578125
],
[
0.51708984375,
0.70361328125,
0.91357421875,
0.70361328125,
0.91357421875,
0.71435546875,
0.51708984375,
0.71435546875
],
[
0.07977294921875,
0.71044921875,
0.295654296875,
0.708984375,
0.295654296875,
0.72021484375,
0.07977294921875,
0.7216796875
],
[
0.515625,
0.71435546875,
0.916015625,
0.71435546875,
0.916015625,
0.72802734375,
0.515625,
0.72802734375
],
[
0.0814208984375,
0.73291015625,
0.48193359375,
0.73388671875,
0.48193359375,
0.7451171875,
0.0814208984375,
0.74462890625
],
[
0.51708984375,
0.728515625,
0.916015625,
0.728515625,
0.916015625,
0.7392578125,
0.51708984375,
0.7392578125
],
[
0.0831298828125,
0.74462890625,
0.48193359375,
0.74462890625,
0.48193359375,
0.7548828125,
0.0831298828125,
0.7548828125
],
[
0.515625,
0.74072265625,
0.91748046875,
0.74072265625,
0.91748046875,
0.75439453125,
0.515625,
0.75439453125
],
[
0.0814208984375,
0.7529296875,
0.48193359375,
0.7529296875,
0.48193359375,
0.76318359375,
0.0814208984375,
0.76318359375
],
[
0.51708984375,
0.75439453125,
0.91455078125,
0.75439453125,
0.91455078125,
0.76416015625,
0.51708984375,
0.76416015625
],
[
0.0814208984375,
0.76123046875,
0.480224609375,
0.76123046875,
0.480224609375,
0.771484375,
0.0814208984375,
0.771484375
],
[
0.515625,
0.765625,
0.91943359375,
0.765625,
0.91943359375,
0.779296875,
0.515625,
0.779296875
],
[
0.0814208984375,
0.771484375,
0.298828125,
0.771484375,
0.298828125,
0.7822265625,
0.0814208984375,
0.7822265625
],
[
0.51708984375,
0.779296875,
0.91357421875,
0.779296875,
0.91357421875,
0.7890625,
0.51708984375,
0.7890625
],
[
0.07977294921875,
0.79443359375,
0.30322265625,
0.79443359375,
0.30322265625,
0.80810546875,
0.07977294921875,
0.80810546875
],
[
0.515625,
0.79052734375,
0.91748046875,
0.79052734375,
0.91748046875,
0.80419921875,
0.515625,
0.80419921875
],
[
0.07977294921875,
0.80419921875,
0.48193359375,
0.80419921875,
0.48193359375,
0.81787109375,
0.07977294921875,
0.81787109375
],
[
0.51708984375,
0.8046875,
0.916015625,
0.8046875,
0.916015625,
0.8154296875,
0.51708984375,
0.8154296875
],
[
0.0814208984375,
0.8154296875,
0.48193359375,
0.8154296875,
0.48193359375,
0.826171875,
0.0814208984375,
0.826171875
],
[
0.515625,
0.81640625,
0.91748046875,
0.81640625,
0.91748046875,
0.830078125,
0.515625,
0.830078125
],
[
0.0831298828125,
0.82373046875,
0.318115234375,
0.82373046875,
0.318115234375,
0.83447265625,
0.0831298828125,
0.83447265625
],
[
0.51708984375,
0.830078125,
0.916015625,
0.830078125,
0.916015625,
0.84033203125,
0.51708984375,
0.84033203125
],
[
0.5498046875,
0.85791015625,
0.91455078125,
0.859375,
0.91455078125,
0.87255859375,
0.5498046875,
0.87158203125
]
] |
[
"HeadOn: Real-time Reenactment of Human Portrait Videos",
"JUSTUS THIES, Technical University of Munich",
"MICHAEL ZOLLHOFER, Stanford University",
"CHRISTIAN THEOBALT,Max-Planck-Institute for Informatics",
"MARC STAMMINGER, University of Erlangen-Nuremberg",
"MATTHIAS NIESSNER, Technical University of Munich",
"Fig. 1.Our novel HeadOn approach enables real-time reenactment ofupper body motion, head pose, face expression, and eye gaze in human portrait videos.",
"For synthesis of new photo-realistic video content, we employ a novel video-based rendering approach that builds on top of a fully controllable 3D actor",
"model. The person-specific model is constructed from a short RGB-D calibration sequence and is driven by a real-time torso and face tracker.",
"We propose HeadOn, the first real-time source-to-target reenactment ap-",
"INTRODUCTION",
"proach for complete human portrait videos that enables transfer of torso and",
"Reenactment approaches aim to transfer the motion of a source",
"head motion, face expression, and eye gaze. Given a short RGB-D video of",
"actor to an image or video of a target actor.Very recently, facial",
"the target actor, we automatically construct a personalized geometry proxy",
"that embeds a parametric head, eye, and kinematic torso model. A novel real-",
"reenactment methods havebeen successfully employed to achieve",
"time reenactment algorithm employs thisproxy to photo-realisticallymap",
"highly-realistic manipulations offacial expressions based on com",
"the captured motion from the source actor to the target actor.On top of the",
"modity video data [Averbuch-Elor et al. 2017; Suwajanakorn et al.",
"coarse geometric proxy, we propose a video-based rendering technique that",
"2017; Thies et al.2015,2016,2018; Vlasic et al.2005].Rather than",
"composites the modified target portrait video via view-and pose-dependent",
"animating a virtual, stylized avatar (e.g., as used in video games),",
"texturing, and creates photo-realistic imagery of the target actor under novel",
"these algorithms replacethe face region of a person with a synthetic",
"toro andheadposes,facialexpressions,andgaze directions.To this end,we",
"re-rendering, or modify the target image/video under guidance of a",
"propose a robust tracking of the face and torso of the source actor. We exten-",
"3D facemodel.This enables changing theexpression ofa target per-",
"sively evaluate our approach and show significant improvements in enabling",
"son and creating a manipulated output video that suggests different",
"much greater flexibility in creating realistic reenacted output videos.",
"content; e.g., a person who is sitting still could appear as if he/she is",
"CCS Concepts: .Computing methodologies- Computervision; Com-",
"talking. Modern reenactment approaches achieve highly believable",
"putergraphics:",
"results, even in real-time,and have enjoyed widemedia coverage",
"due to the interest in general movie and video editing [Vlasic et al.",
"Additional Key Words and Phrases: Reenactment, Face tracking, Video-based",
"Rendering,Real-time",
"2005], teleconferencing [Thies et al.2018], reactive profile pictures",
"[Averbuch-Elor et al. 2017], or visual dubbing of foreign language",
"ACM Reference Format:",
"movies [Garrido et al. 2015]",
"Justus Thies,Michael Zollhofer, Christian Theobalt, Marc Stamminger,",
"Even though current facial reenactment results are impressive",
"and MatthiasNieBner.2018.HeadOn:Real-timeReen",
"actmentofHuman",
"they are still fundamentally limited in the type of manipulations",
"Portrait Videos. ACM Trans. Graph.37, 4, Article 164 (August 2018), 13 pages.",
"they enable.For instance, these approaches are only able to modify",
"https:/doi.org/10.1145/3197517.3201350",
"facial expressions, whereas the rigid pose of the head, including",
"Authors addresses: Justus Thies, Technical University of Munich justus.thies@tum.de:",
"its orientation, remains unchanged and does not follow the input",
"Michael Zollhofer, Stanford University,zollhoefer@cs.stanford.edu; Christian Theobalt,",
"video. Thus, only subtle changes, such as opening the mouth or",
"Max-Planck-Institute for Informatics, theobalt@mpi-inf.mpg.de; Marc Stamminger,",
"adding wrinkles on theforehead are realized, which severely limits",
"University of Erlangen-Nuremberg,marc.stamminger@fau.de; Matthias Niener,Tech",
"the applicability to video editing, where the control of the pose of",
"nical University of Munich, niessner@tum.de.",
"the target person is also required. Furthermore, without joint modi",
" 2018 Association for Computing Machinery.",
"fication of the head pose, the modified facial expressions often seem",
"This isthe author's versionofthe workItisposted hereforyourpersonaluse.Notfor",
"out-of-place, since theydo not well align withvisual pauses in the",
"redistribution. The definitive Version of Record was published in ACM Transactions on",
"body and head motion; as noted by Suwajanakorn et al. [2017] this",
"Graphics, https:/doi.org/10.1145/3197517.3201350.",
"significantlyrestrictstheapplicabilityinteleconferencing scenarios.",
"ACM Trans. Graph, Vol. 37, No.4, Article 164. Publication date: August 2018"
] |
{
"arxiv_id": "1805.11729",
"language": "zh",
"page_idx": 1
}
|
||
1805.11729-zh-2
|
1805.11729-zh-2.jpg
|
164:2 $\bullet$ Justus Thies, Michael Zollhöfer, Christian Theobalt, Marc Stamminger, and Matthias Nießner
et al. 2005] jointly represent face identity and expression in a single multi-linear model. Joint shape and motion models [Li et al. 2017] have also been learned from a large collection of 4D scan data. Other approaches [Garrido et al. 2016] reconstruct personalized face rigs, including reflectance and fine-scale detail from monocular video. Liang et al.[Liang et al. 2014] reconstruct the identity of a face from monocular Kinect data using a part-based matching algorithm. They select face parts (eyes,nose,mouth,cheeks) from a database of faces that best match the input data. To get an improved and personalized output they fuse these parts with the Kinect depth data. Ichim et al.[2015] propose to reconstruct 3D avatars from multi-view images recorded by a mobile phone and personalize the expression space using a calibration sequence.
In this work, we thus go one step further by introducing HeadOn, a reenactment system for portrait videos recorded with a commodity RGB-D camera. We overcome the limitations of current facial reenactment methods by not only controlling changes in facial expression, but also reenacting the rigid position of the head, of the upper body, and the eye gaze- i.e., the entire person-related content in a portrait video.
At the core of our approach is the combination of robust and accurate tracking of a deformation proxy with view-dependent texturing for video-based re-rendering. To achieve this, we propose a new method to swiftly and automatically construct a personalized head and torso geometry proxy of a human from a brief RGB-D initialization sequence. The shape proxy features a personalized parametric $3\dot{\mathrm{D~model~of~the~complete~head~that~is~rigged~with~blendshapes~for}}$ expression control and is integrated with a personalized upper torso model. A new real-time reenactment algorithm employs this proxy to photo-realistically map face expression and eye gaze, as well as head and torso motion of a captured source actor to a target actor. To this end, we contribute a new photo-realistic video-based rendering approach that composites the reenacted target portrait video via view- and pose-dependent texturing and video compositing.
In summary, we contribute the following:
$\bullet$ rapid automatic construction of a personalized geometry proxy that embeds a parametric human face, eye, full head, ${\mathrm{and~upper~body~model}}$,
$\bullet$ a photo-realistic, view-, and pose-dependent texturing and
$\dot{\mathrm{compositing~approach}}$,
$\bullet$ a robust tracking approach of the source actor,
$\bullet$ and real-time source-to-target reenactment of complete hu-
man portrait videos.
Commodity Face Reconstruction and Tracking. The first commodity face reconstruction approaches that employed lightweight capture setups, i.e., stereo [Valgaerts et al. 2012], RGB [Fyffe et al. 2014; Garrido et al. 2013; Shi et al. 2014a; Suwajanakorn et al. 2014, 2015], or RGB-D [Chen et al. 2013] cameras had slow off-line frame rates and required up to several minutes to process a single input frame. These methods either deform a personalized template mesh [Suwajanakorn et al. 2014, 2015; Valgaerts et al. 2012], use a 3D template and expression blendshapes [Fyffe et al. 2014; Garrido et al. 2013], a template and an underlying generic deformation graph [Chen et al. 2013], or additionally solve for the parameters of a multilinear face model [Shi et al. 2014a]. Suwajanakorn et al. [2014; 2015] build a modifiable mesh model from internet photo collections. Shi et al.[2014b] key-frame based bundle adjustment to fit the multilinear model. Recently, first methods have appeared that reconstruct facial performances in real-time from a single commodity RGB-D camera [Bouaziz et al. 2013; Hsieh et al. 2015; Li et al. 2013; Thies et al. 2015; Weise et al. 2011; Zollhöfer et al. 2014]. Dense real-time face reconstruction has also been demonstrated based on monocular RGB data using trained regressors [Cao et al. 2014a, 2013] or analysis-by-synthesis [Thies et al. 2016]. Even fine-scale detail can be recovered at real-time frame rates [Cao et al. 2015].
## 2 RELATED WORK
Face reconstruction and reenactment have a long history in computer graphics and vision. We focus on recent approaches based on lightweight commodity sensors. For an overview of high-quality techniques that use controlled acquisition setups, we refer to Klehm et al.[2015]. Recently, a state-of-the-art report on monocular 3d face reconstruction, tracking and applications has been published that gives a comprehensive overview of current methods [Zollhöfer et al. 2018]. In the following we concentrate on the most related techniques.
Performance Driven Facial Animation. Face tracking has been applied to control virtual avatars in many contexts. First approaches were based on sparse detected feature points [Chai et al.2003; Chuang and Bregler 2002]. Current methods for character animation [Cao et al. 2015, 2014a, 2013; Weise et al. 2009], teleconferences [Weise et al. 2011], games [Ichim et al. 2015], and virtual reality [Li et al. 2015; Olszewski et al. 2016] are based on dense alignment energies. Olszewski et al. [2016] proposed an approach to control a digital avatar in real-time based on an HMD-mounted RGB camera. Recently, Hu et al. [2017] reconstructed a stylized 3D avatar, including hair, from a single image that can be animated and displayed in virtual environments. General image-based modeling and rendering techniques [Gortler et al. 1996; Isaksen et al. 2000; Kang et al. 2006; Kopf et al. 2013; Wood et al. 2000] enable the creation of photo-realistic imagery for many real-world effects that are hard to render and reconstruct at a sufficiently high quality using current approaches. In the context of portrait videos, especially fine details, e.g., single strands of hair or high-quality apparel, are hard to reconstruct. Cao et al. [2016] drive dynamic image-based 3D avatars
Parametric Face Representations. Current state-of-the-art monocular face tracking and reconstruction approaches heavily rely on 3D parametric identity [Blanz et al. 2003; Blanz and Vetter 1999] and expression models [Tena et al. 2011] that generalize active appearance models [Cootes et al. 2001] from 2D to 3D space. Even combinations of the two have been proposed [Xiao et al. 2004]. Recently, large-scale models in terms of geometry [Booth et al.2016] and texture [Zafeiriou et al. 2017] have been constructed based on an immense amount of training data (10,000 scans). For modeling facial expressions, the de facto standard in the industry are blendshapes [Lewis et al. 2014; Pighin et al. 1998]. Physics-based models [Ichim et al. 2017; Sifakis et al. 2005] have been proposed in research, but fitting such complex models to commodity video at real-time rates is still challenging. Some approaches [Shi et al. 2014a; Vlasic
ACM Trans. Graph., Vol. 37, No. 4, Article 164. Publication date: August 2018.
|
[
[
0.08477783203125,
0.1051025390625,
0.7109375,
0.1051025390625,
0.7109375,
0.11517333984375,
0.08477783203125,
0.11517333984375
],
[
0.0948486328125,
0.1353759765625,
0.483642578125,
0.1353759765625,
0.483642578125,
0.1484375,
0.0948486328125,
0.1484375
],
[
0.51708984375,
0.135986328125,
0.916015625,
0.135986328125,
0.916015625,
0.146728515625,
0.51708984375,
0.146728515625
],
[
0.0814208984375,
0.1478271484375,
0.483642578125,
0.1478271484375,
0.483642578125,
0.160888671875,
0.0814208984375,
0.160888671875
],
[
0.515625,
0.1478271484375,
0.916015625,
0.1478271484375,
0.916015625,
0.160888671875,
0.515625,
0.160888671875
],
[
0.0814208984375,
0.1602783203125,
0.483642578125,
0.1602783203125,
0.483642578125,
0.17333984375,
0.0814208984375,
0.17333984375
],
[
0.515625,
0.1602783203125,
0.91748046875,
0.1602783203125,
0.91748046875,
0.17333984375,
0.515625,
0.17333984375
],
[
0.0814208984375,
0.17333984375,
0.48193359375,
0.17333984375,
0.48193359375,
0.18701171875,
0.0814208984375,
0.18701171875
],
[
0.515625,
0.17333984375,
0.91748046875,
0.17333984375,
0.91748046875,
0.18701171875,
0.515625,
0.18701171875
],
[
0.0814208984375,
0.1859130859375,
0.48193359375,
0.1859130859375,
0.48193359375,
0.1995849609375,
0.0814208984375,
0.1995849609375
],
[
0.5146484375,
0.185302734375,
0.916015625,
0.185302734375,
0.916015625,
0.198974609375,
0.5146484375,
0.198974609375
],
[
0.0814208984375,
0.198974609375,
0.483642578125,
0.198974609375,
0.483642578125,
0.2120361328125,
0.0814208984375,
0.2120361328125
],
[
0.515625,
0.198974609375,
0.91748046875,
0.198974609375,
0.91748046875,
0.2120361328125,
0.515625,
0.2120361328125
],
[
0.0831298828125,
0.2120361328125,
0.197265625,
0.2120361328125,
0.197265625,
0.2244873046875,
0.0831298828125,
0.2244873046875
],
[
0.515625,
0.21142578125,
0.916015625,
0.21142578125,
0.916015625,
0.2244873046875,
0.515625,
0.2244873046875
],
[
0.0948486328125,
0.223876953125,
0.48193359375,
0.223876953125,
0.48193359375,
0.2369384765625,
0.0948486328125,
0.2369384765625
],
[
0.515625,
0.223876953125,
0.916015625,
0.223876953125,
0.916015625,
0.2369384765625,
0.515625,
0.2369384765625
],
[
0.0814208984375,
0.236328125,
0.483642578125,
0.236328125,
0.483642578125,
0.2493896484375,
0.0814208984375,
0.2493896484375
],
[
0.5146484375,
0.235107421875,
0.91748046875,
0.236328125,
0.91748046875,
0.2493896484375,
0.5146484375,
0.248779296875
],
[
0.0814208984375,
0.2493896484375,
0.48193359375,
0.2493896484375,
0.48193359375,
0.26318359375,
0.0814208984375,
0.26318359375
],
[
0.5146484375,
0.2493896484375,
0.916015625,
0.2493896484375,
0.916015625,
0.26318359375,
0.5146484375,
0.26318359375
],
[
0.07977294921875,
0.262451171875,
0.483642578125,
0.262451171875,
0.483642578125,
0.275634765625,
0.07977294921875,
0.275634765625
],
[
0.515625,
0.262451171875,
0.91748046875,
0.262451171875,
0.91748046875,
0.275634765625,
0.515625,
0.275634765625
],
[
0.07977294921875,
0.27490234375,
0.48193359375,
0.27490234375,
0.48193359375,
0.2880859375,
0.07977294921875,
0.2880859375
],
[
0.515625,
0.27490234375,
0.91748046875,
0.27490234375,
0.91748046875,
0.2880859375,
0.515625,
0.2880859375
],
[
0.0814208984375,
0.289306640625,
0.48193359375,
0.289306640625,
0.48193359375,
0.2998046875,
0.0814208984375,
0.2998046875
],
[
0.51904296875,
0.289306640625,
0.69287109375,
0.289306640625,
0.69287109375,
0.2998046875,
0.51904296875,
0.2998046875
],
[
0.0814208984375,
0.2998046875,
0.483642578125,
0.2998046875,
0.483642578125,
0.314208984375,
0.0814208984375,
0.314208984375
],
[
0.0831298828125,
0.314208984375,
0.480224609375,
0.314208984375,
0.480224609375,
0.32470703125,
0.0831298828125,
0.32470703125
],
[
0.5341796875,
0.312255859375,
0.916015625,
0.312255859375,
0.916015625,
0.322509765625,
0.5341796875,
0.322509765625
],
[
0.0814208984375,
0.325927734375,
0.48193359375,
0.325927734375,
0.48193359375,
0.339111328125,
0.0814208984375,
0.339111328125
],
[
0.51708984375,
0.325927734375,
0.91455078125,
0.325927734375,
0.91455078125,
0.336181640625,
0.51708984375,
0.336181640625
],
[
0.0814208984375,
0.33837890625,
0.48193359375,
0.33837890625,
0.48193359375,
0.3515625,
0.0814208984375,
0.3515625
],
[
0.51708984375,
0.33837890625,
0.916015625,
0.33837890625,
0.916015625,
0.3486328125,
0.51708984375,
0.3486328125
],
[
0.07977294921875,
0.349853515625,
0.48193359375,
0.349853515625,
0.48193359375,
0.363525390625,
0.07977294921875,
0.363525390625
],
[
0.5146484375,
0.3486328125,
0.91748046875,
0.349853515625,
0.91748046875,
0.363525390625,
0.5146484375,
0.3623046875
],
[
0.0831298828125,
0.365234375,
0.48193359375,
0.365234375,
0.48193359375,
0.3759765625,
0.0831298828125,
0.3759765625
],
[
0.5146484375,
0.361083984375,
0.91748046875,
0.3623046875,
0.91748046875,
0.3759765625,
0.5146484375,
0.374755859375
],
[
0.0814208984375,
0.3759765625,
0.483642578125,
0.3759765625,
0.483642578125,
0.3896484375,
0.0814208984375,
0.3896484375
],
[
0.515625,
0.374755859375,
0.91748046875,
0.374755859375,
0.91748046875,
0.388427734375,
0.515625,
0.388427734375
],
[
0.0814208984375,
0.3896484375,
0.444091796875,
0.3896484375,
0.444091796875,
0.402587890625,
0.0814208984375,
0.402587890625
],
[
0.515625,
0.388427734375,
0.91455078125,
0.388427734375,
0.91455078125,
0.3984375,
0.515625,
0.3984375
],
[
0.09320068359375,
0.402099609375,
0.34765625,
0.402099609375,
0.34765625,
0.4150390625,
0.09320068359375,
0.4150390625
],
[
0.51708984375,
0.402099609375,
0.916015625,
0.402099609375,
0.916015625,
0.412109375,
0.51708984375,
0.412109375
],
[
0.10662841796875,
0.416259765625,
0.48193359375,
0.416259765625,
0.48193359375,
0.429931640625,
0.10662841796875,
0.429931640625
],
[
0.51708984375,
0.41455078125,
0.916015625,
0.41455078125,
0.916015625,
0.424560546875,
0.51708984375,
0.424560546875
],
[
0.1175537109375,
0.429931640625,
0.483642578125,
0.4287109375,
0.483642578125,
0.4423828125,
0.1175537109375,
0.443603515625
],
[
0.515625,
0.42578125,
0.91748046875,
0.42578125,
0.91748046875,
0.439453125,
0.515625,
0.439453125
],
[
0.118408203125,
0.4423828125,
0.26123046875,
0.4423828125,
0.26123046875,
0.4560546875,
0.118408203125,
0.4560546875
],
[
0.51708984375,
0.439453125,
0.916015625,
0.439453125,
0.916015625,
0.449462890625,
0.51708984375,
0.449462890625
],
[
0.10662841796875,
0.454833984375,
0.483642578125,
0.454833984375,
0.483642578125,
0.468505859375,
0.10662841796875,
0.468505859375
],
[
0.515625,
0.45068359375,
0.91748046875,
0.45068359375,
0.91748046875,
0.46435546875,
0.515625,
0.46435546875
],
[
0.1175537109375,
0.466064453125,
0.26025390625,
0.46728515625,
0.26025390625,
0.482177734375,
0.1175537109375,
0.48095703125
],
[
0.515625,
0.463134765625,
0.91748046875,
0.463134765625,
0.91748046875,
0.476806640625,
0.515625,
0.476806640625
],
[
0.10498046875,
0.479736328125,
0.403076171875,
0.478515625,
0.403076171875,
0.493408203125,
0.10498046875,
0.49462890625
],
[
0.515625,
0.4755859375,
0.916015625,
0.4755859375,
0.916015625,
0.4892578125,
0.515625,
0.4892578125
],
[
0.10498046875,
0.4921875,
0.483642578125,
0.4912109375,
0.483642578125,
0.505859375,
0.10498046875,
0.50732421875
],
[
0.51904296875,
0.490478515625,
0.916015625,
0.490478515625,
0.916015625,
0.50048828125,
0.51904296875,
0.50048828125
],
[
0.118408203125,
0.505859375,
0.2451171875,
0.505859375,
0.2451171875,
0.51953125,
0.118408203125,
0.51953125
],
[
0.515625,
0.50048828125,
0.91748046875,
0.49951171875,
0.91748046875,
0.51416015625,
0.515625,
0.515625
],
[
0.51904296875,
0.515625,
0.916015625,
0.515625,
0.916015625,
0.525390625,
0.51904296875,
0.525390625
],
[
0.10498046875,
0.529296875,
0.2275390625,
0.529296875,
0.2275390625,
0.5419921875,
0.10498046875,
0.5419921875
],
[
0.515625,
0.52685546875,
0.916015625,
0.52685546875,
0.916015625,
0.54052734375,
0.515625,
0.54052734375
],
[
0.08563232421875,
0.533203125,
0.0948486328125,
0.533203125,
0.0948486328125,
0.5380859375,
0.08563232421875,
0.5380859375
],
[
0.515625,
0.5390625,
0.91748046875,
0.5390625,
0.91748046875,
0.552734375,
0.515625,
0.552734375
],
[
0.0831298828125,
0.54638671875,
0.48193359375,
0.54638671875,
0.48193359375,
0.55712890625,
0.0831298828125,
0.55712890625
],
[
0.515625,
0.5517578125,
0.916015625,
0.5517578125,
0.916015625,
0.5654296875,
0.515625,
0.5654296875
],
[
0.0814208984375,
0.55810546875,
0.483642578125,
0.55810546875,
0.483642578125,
0.57177734375,
0.0814208984375,
0.57177734375
],
[
0.515625,
0.56396484375,
0.91748046875,
0.56396484375,
0.91748046875,
0.57763671875,
0.515625,
0.57763671875
],
[
0.0814208984375,
0.57080078125,
0.48193359375,
0.57080078125,
0.48193359375,
0.58447265625,
0.0814208984375,
0.58447265625
],
[
0.51708984375,
0.5791015625,
0.84228515625,
0.5791015625,
0.84228515625,
0.5888671875,
0.51708984375,
0.5888671875
],
[
0.0831298828125,
0.5849609375,
0.48193359375,
0.5849609375,
0.48193359375,
0.595703125,
0.0831298828125,
0.595703125
],
[
0.0831298828125,
0.59716796875,
0.48193359375,
0.59716796875,
0.48193359375,
0.60791015625,
0.0831298828125,
0.60791015625
],
[
0.53076171875,
0.599609375,
0.916015625,
0.6015625,
0.916015625,
0.615234375,
0.53076171875,
0.61328125
],
[
0.0814208984375,
0.609375,
0.483642578125,
0.609375,
0.483642578125,
0.62255859375,
0.0814208984375,
0.62255859375
],
[
0.515625,
0.6142578125,
0.91748046875,
0.6142578125,
0.91748046875,
0.62744140625,
0.515625,
0.62744140625
],
[
0.0831298828125,
0.62353515625,
0.480224609375,
0.62353515625,
0.480224609375,
0.63427734375,
0.0831298828125,
0.63427734375
],
[
0.515625,
0.626953125,
0.91748046875,
0.626953125,
0.91748046875,
0.64013671875,
0.515625,
0.64013671875
],
[
0.0814208984375,
0.63427734375,
0.483642578125,
0.63427734375,
0.483642578125,
0.64794921875,
0.0814208984375,
0.64794921875
],
[
0.515625,
0.63818359375,
0.916015625,
0.63818359375,
0.916015625,
0.65185546875,
0.515625,
0.65185546875
],
[
0.0831298828125,
0.6484375,
0.15283203125,
0.6484375,
0.15283203125,
0.66015625,
0.0831298828125,
0.66015625
],
[
0.51708984375,
0.65185546875,
0.91748046875,
0.65185546875,
0.91748046875,
0.6650390625,
0.51708984375,
0.6650390625
],
[
0.0965576171875,
0.6650390625,
0.48193359375,
0.6650390625,
0.48193359375,
0.6787109375,
0.0965576171875,
0.6787109375
],
[
0.51708984375,
0.66455078125,
0.916015625,
0.66455078125,
0.916015625,
0.677734375,
0.51708984375,
0.677734375
],
[
0.0831298828125,
0.68115234375,
0.48193359375,
0.68115234375,
0.48193359375,
0.69140625,
0.0831298828125,
0.69140625
],
[
0.515625,
0.6767578125,
0.91748046875,
0.677734375,
0.91748046875,
0.69140625,
0.515625,
0.6904296875
],
[
0.0831298828125,
0.693359375,
0.480224609375,
0.693359375,
0.480224609375,
0.70361328125,
0.0831298828125,
0.70361328125
],
[
0.515625,
0.6904296875,
0.91748046875,
0.6904296875,
0.91748046875,
0.70361328125,
0.515625,
0.70361328125
],
[
0.0814208984375,
0.70361328125,
0.48193359375,
0.70361328125,
0.48193359375,
0.71728515625,
0.0814208984375,
0.71728515625
],
[
0.5146484375,
0.703125,
0.916015625,
0.70166015625,
0.916015625,
0.71533203125,
0.5146484375,
0.71630859375
],
[
0.0814208984375,
0.71728515625,
0.483642578125,
0.71728515625,
0.483642578125,
0.73095703125,
0.0814208984375,
0.73095703125
],
[
0.515625,
0.71533203125,
0.91748046875,
0.71533203125,
0.91748046875,
0.728515625,
0.515625,
0.728515625
],
[
0.0814208984375,
0.728515625,
0.48193359375,
0.72998046875,
0.48193359375,
0.74365234375,
0.0814208984375,
0.7421875
],
[
0.515625,
0.72802734375,
0.91748046875,
0.72802734375,
0.91748046875,
0.7412109375,
0.515625,
0.7412109375
],
[
0.0814208984375,
0.7421875,
0.483642578125,
0.7421875,
0.483642578125,
0.755859375,
0.0814208984375,
0.755859375
],
[
0.51708984375,
0.7421875,
0.916015625,
0.7421875,
0.916015625,
0.7529296875,
0.51708984375,
0.7529296875
],
[
0.0831298828125,
0.75732421875,
0.48193359375,
0.75732421875,
0.48193359375,
0.76708984375,
0.0831298828125,
0.76708984375
],
[
0.515625,
0.75439453125,
0.91748046875,
0.75439453125,
0.91748046875,
0.76708984375,
0.515625,
0.76708984375
],
[
0.07977294921875,
0.7666015625,
0.483642578125,
0.76708984375,
0.483642578125,
0.7822265625,
0.07977294921875,
0.78076171875
],
[
0.515625,
0.7666015625,
0.91748046875,
0.7666015625,
0.91748046875,
0.77978515625,
0.515625,
0.77978515625
],
[
0.0814208984375,
0.78076171875,
0.483642578125,
0.78076171875,
0.483642578125,
0.79443359375,
0.0814208984375,
0.79443359375
],
[
0.51708984375,
0.78076171875,
0.916015625,
0.78076171875,
0.916015625,
0.79150390625,
0.51708984375,
0.79150390625
],
[
0.0814208984375,
0.7919921875,
0.48193359375,
0.7919921875,
0.48193359375,
0.8056640625,
0.0814208984375,
0.8056640625
],
[
0.515625,
0.79150390625,
0.91748046875,
0.79150390625,
0.91748046875,
0.8046875,
0.515625,
0.8046875
],
[
0.0814208984375,
0.8056640625,
0.48193359375,
0.8056640625,
0.48193359375,
0.8193359375,
0.0814208984375,
0.8193359375
],
[
0.51708984375,
0.8056640625,
0.916015625,
0.8056640625,
0.916015625,
0.81640625,
0.51708984375,
0.81640625
],
[
0.0814208984375,
0.818359375,
0.483642578125,
0.818359375,
0.483642578125,
0.83203125,
0.0814208984375,
0.83203125
],
[
0.5146484375,
0.81787109375,
0.91748046875,
0.81640625,
0.91748046875,
0.830078125,
0.5146484375,
0.83056640625
],
[
0.07977294921875,
0.830078125,
0.48193359375,
0.83056640625,
0.48193359375,
0.845703125,
0.07977294921875,
0.84423828125
],
[
0.515625,
0.830078125,
0.91748046875,
0.830078125,
0.91748046875,
0.84326171875,
0.515625,
0.84326171875
],
[
0.0831298828125,
0.86083984375,
0.44580078125,
0.86083984375,
0.44580078125,
0.87158203125,
0.0831298828125,
0.87158203125
]
] |
[
"164:2.JustusThies,MichaelZollhofer,ChristianTheobalt,MarcStamminger,andMatthiasNieBner",
"In this work, we thus go one step further by introducing HeadOn,",
"et al.2005ljointly represent face identity and expression in a single",
"a reenactment system for portrait videos recorded with a commod-",
"multi-linear model. Joint shape and motion models [Li et al. 2017]",
"ity RGB-D camera. We overcome the limitations of current facial",
"have also been learned from a large collection of 4D scan data. Other",
"reenactment methods by not only controlling changes in facial ex",
"approaches [Garrido et al. 2016] reconstruct personalized face rigs,",
"pression, but also reenacting the rigid position of the head, of the",
"including reflectance and fine-scale detail from monocular video.",
"upper body,and the eye gaze-i.e., the entire person-related content",
"Liang et al. [Liang et al. 2014] reconstruct the identity of a face from",
"in a portrait video.",
"monocular Kinect data using a part-based matching algorithm. They",
"At the core of our approach is the combination of robust and accu-",
"select face parts (eyes,nose,mouth,cheeks) from a database of faces",
"rate tracking of a deformation proxy with view-dependent texturing",
"that best match the input data. To get an improved and personalized",
"for video-based re-rendering. To achieve this, we propose a new",
"output they fuse these parts with the Kinect depth data. Ichim et",
"method to swiftly and automatically construct a personalized head",
"al. [2015] propose to reconstruct 3D avatars from multi-view images",
"and torso geometry proxy of a human from a brief RGB-D initializa-",
"recorded by a mobile phone and personalize the expression space",
"tion sequence.The shape proxy features a personalized parametric",
"usinga calibration sequence.",
"3D model of the complete head that is rigged with blendshapes for",
"expression controland is integrated witha personalized upper torso",
"CommodityFaceReconstruction and Tracking.Thefirst commod-",
"model.A new real-time reenactment algorithm employs this proxy",
"ityfacereconstruction approachesthat employed lightweightcap",
"to photo-realistically map face expression and eye gaze, as well as",
"ture setups, i.e., stereo [Valgaerts et al. 2012], RGB [Fyffe et al.2014;",
"head and torso motion of a captured source actor to a target actor. To",
"Garrido et al. 2013; Shi et al. 2014a; Suwajanakorn et al. 2014,2015].",
"this end, we contribute a new photo-realistic video-based rendering",
"or RGB-D [Chen et al. 2013] cameras had slow off-line frame rates",
"approach that composites the reenacted target portrait video via",
"and required up to several minutes to process a single input frame.",
"view- and pose-dependent texturing and video compositing.",
"Thesemethods eitherdeform a personalized template mesh [Suwa",
"In summary,we contribute thefollowing:",
"janakorn et al.2014,2015;Valgaerts etal.2012],use a3D template",
". rapid automatic construction of a personalized geometry",
"andexpressionblendshapes[Fyffeetal.2014;Garridoetal.2013],",
"proxy that embeds a parametric human face, eye, full head,",
"a template and an underlying generic deformation graph [Chen",
"and upper body model,",
"etal.2013],oradditionallysolvefortheparametersofamulti-",
". a photo-realistic, view-, and pose-dependent texturing and",
"linear face model [Shi et al. 2014a]. Suwajanakorn et al. [2014; 2015]",
"compositing approach,",
"build a modifiable mesh model from internet photo collections. Shi",
"· a robust tracking approach of the source actor,",
"et al. [2014b] key-frame based bundle adjustment to fit the multi",
". and real-time source-to-target reenactment of complete hu-",
"linearmodel.Recently,firstmethodshaveappearedthatreconstruct",
"man portrait videos.",
"facial performances in real-time from a single commodity RGB-D",
"camera[Bouaziz et al.2013;Hsieh et al.2015;Li et al.2013; Thies",
"RELATEDWORK",
"et al. 2015; Weise et al. 2011; Zollhofer et al. 2014]. Dense real-time",
"",
"face reconstruction has also been demonstrated based on monoc-",
"Face reconstruction and reenactment have a long history in com-",
"ular RGB data using trained regressors [Cao et al. 2014a, 2013] or",
"puter graphics and vision. We focus on recent approaches based on",
"analysis-by-synthesis [Thies et al. 2016]. Even fine-scale detail can",
"lightweight commodity sensors. For an overview of high-quality",
"be recovered at real-time frame rates [Cao et al.2015]",
"techniques that use controlled acquisition setups, we refer to Klehm",
"et al. [2015]. Recently, a state-of-the-art report on monocular 3d",
"Performance Driven Facial Animation. Face tracking has been ap-",
"face reconstruction, tracking and applications has been published",
"plied to control virtual avatars in many contexts. First approaches",
"that gives a comprehensive overview of current methods [Zollhofer",
"were based on sparse detected feature points [Chai et al. 2003;",
"et al. 2018]. In the following we concentrate on the most related",
"Chuang and Bregler 2002]. Current methods for character anima-",
"techniques.",
"tion [Cao et al. 2015, 2014a, 2013; Weise et al. 2009], teleconferences",
"Parametric Face Representations. Current state-of-the-art monoc-",
"[Weise et al. 2011], games [Ichim et al. 2015], and virtual reality",
"ularfacetracking and reconstructionapproachesheavilyrelyon",
"[Li et al. 2015; Olszewski et al. 2016] are based on dense alignment",
"3D parametric identity[Blanz etal.2003;Blanz and Vetter1999]",
"energies. Olszewski et al. [2016] proposed an approach to control",
"and expression models [Tena et al. 2011] that generalize active ap-",
"a digital avatar in real-time based on an HMD-mounted RGB cam-",
"pearance models [Cootes et al. 2001] from 2D to 3D space. Even",
"era. Recently, Hu et al. [2017] reconstructed a stylized 3D avatar,",
"combinations of the two have been proposed [Xiao et al. 2004]. Re-",
"including hair, from a single image that can be animated and dis-",
"cently, large-scale models in terms of geometry [Booth et al. 2016]",
"played in virtual environments.General image-based modeling and",
"and texture[Zafeiriouet al.2017] havebeen constructedbased on",
"rendering techniques [Gortler et al. 1996; Isaksen et al. 2000; Kang",
"an immense amount of training data (10,000 scans).For modeling",
"et al. 2006; Kopf et al. 2013; Wood et al. 2000] enable the creation of",
"facial expressions, the de facto standard in the industry are blend-",
"photo-realistic imageryfor many real-world effects that are hard to",
"shapes [Lewis et al. 2014; Pighin et al. 1998]. Physics-based models",
"render and reconstruct at a sufficiently high quality using current",
"[Ichim et al. 2017; Sifakis et al. 2005] have been proposed in research,",
"approaches. In the context of portrait videos, especiallyfine details",
"but fitting such complex models to commodity video at real-time",
"e.g, single strands of hair or high-quality apparel, are hard to re-",
"rates is still challenging. Some approaches [Shi et al. 2014a; Vlasic",
"construct. Cao et al. [2016] drive dynamic image-based 3D avatars",
"ACM Trans. Graph., Vol 37, No. 4, Article 164. Publication date: August 2018."
] |
{
"arxiv_id": "1805.11729",
"language": "zh",
"page_idx": 2
}
|
||
2207.07180-zh-1
|
2207.07180-zh-1.jpg
|
# 用于基础模型组健壮性的对比适配器
Michael Zhang* and Christopher Ré
Stanford University
$\{$ mzhang,chrismre $\}$ @cs.stanford.edu
Abstract
While large pretrained foundation models (FMs) have shown remarkable zero-shot classification robustness to dataset-level distribution shifts, their robustness to subpopulation or group shifts is relatively underexplored. We study this problem, and find that FMs such as CLIP may not be robust to various group shifts. Across 9 robustness benchmarks, zero-shot classification with their embeddings results in gaps of up to 80.7 percentage points (pp) between average and worst-group accuracy. Unfortunately, existing methods to improve robustness require retraining, which can be prohibitively expensive on large foundation models. We also find that efficient ways to improve model inference $(e.g.$, via adapters, lightweight networks with FM embeddings as inputs) do not consistently improve and can sometimes hurt group robustness compared to zero-shot $(e.g.$, increasing the accuracy gap by 50.1 pp on CelebA). We thus develop an adapter training strategy to effectively and efficiently improve FM group robustness. Our motivating observation is that while poor robustness results from groups in the same class being embedded far apart in the foundation model“embedding space,” standard adapter training may not bring these points closer together. We thus propose contrastive adapting, which trains adapters with contrastive learning to bring sample embeddings close to both their ground-truth class embeddings and other sample embeddings in the same class. Across the 9 benchmarks, our approach consistently improves group robustness, raising worst-group accuracy by 8.5 to 56.0 pp over zero-shot. Our approach is also efficient, doing so without any FM finetuning and only a fixed set of frozen FM embeddings. On benchmarks such as Waterbirds and CelebA, this leads to worst-group accuracy comparable to state-of-the-art methods that retrain entire models, while only training $\leq1\%$ of the model parameters
# 1 $\textbf{Introduction}$
基础模型 (FMs) -在海量数据集上训练的大型预训练模型-为深度学习提供了令人兴奋的新范例。最近的工作表明，在没有任何微调的情况下，基础模型可以很好地推广到各种数据集[10,30,53,61] ,并对某些分布偏移[36,67]表现出令人印象深刻的鲁棒性。在这种零拍范式下，从业者可以避免训练任务特定的模型，而是使用 FM 嵌人进行高效和有效的推理。然而，一个未被充分开发的问题是，这种零拍推断对“群体转移” 有多稳健，数据中的亚群或有意义的群体之间的分布转移。先前的工作已经确定，组鲁棒性- $i.e.$在所有组上都表现良好-是现代深度学习[4,11,34,45,49,58,63]的基本和现实挑战。然而 , 大多数先前的基础模型评估都集中在整体或平均性能[36,53,67]上；很少有作品考虑其在小组中的准确性
因此，这项工作进一步研究了用于零拍分类的基础模型的组鲁棒性。我们首先通过证明基础模型可能具有较差的零拍组鲁棒性来激发此问题。在 9 健壮性基准测试中评估 11 基础模型时，我们发现它们达到了 80。平均和最差组精度之间的差距为 7 个百分点 (pp), 其中 CLIPResNet-50$\overline{\mathrm{E}}[53]$仅分类为 6。0 在 CelebA 数据集上正确的最差组样本的% [41,57] 。鉴于这一观察到的弱点，我们的目标是提高基础模型组的鲁棒性。这带来了几个挑战和有待回答的问题。首先，尽管对机器学习中的组鲁棒性进行了深人研究，但现有的鲁棒性方法需要重新训练一个 (通常是多个) 整个模型[1,15,33,40,45,57,63,64,70] 。由于基础模型的大小和规模，这对于基础模型来说可能是非常昂贵的，这引发了一个问题，即我们是否可以在不进行任何重新训练或微调的情况下使这些模型更加健壮。其次，对于零拍分类，许多从业者也可能只访问基础模型输出或嵌人$(e.g.,\text{ 通过 API }^1)$。为了提高鲁棒性，理想的解决方案应该只需要预先训练的 FM 嵌人。但是，这些相同
$^*$Corresponding author
$^{1}$https://beta.openai.com/docs/introduction. ,https://studio.ai21.com/docs/ ,https://docs.cohere.ai/
2
|
[
[
0.27685546875,
0.1395263671875,
0.7255859375,
0.1395263671875,
0.7255859375,
0.1634521484375,
0.27685546875,
0.1634521484375
],
[
0.3505859375,
0.1881103515625,
0.650390625,
0.1881103515625,
0.650390625,
0.20263671875,
0.3505859375,
0.20263671875
],
[
0.414306640625,
0.203857421875,
0.58251953125,
0.2052001953125,
0.58251953125,
0.222900390625,
0.414306640625,
0.2215576171875
],
[
0.32275390625,
0.2247314453125,
0.685546875,
0.2247314453125,
0.685546875,
0.2392578125,
0.32275390625,
0.2392578125
],
[
0.46728515625,
0.279052734375,
0.53369140625,
0.279052734375,
0.53369140625,
0.29345703125,
0.46728515625,
0.29345703125
],
[
0.1568603515625,
0.298583984375,
0.84326171875,
0.298583984375,
0.84326171875,
0.3125,
0.1568603515625,
0.3125
],
[
0.1568603515625,
0.313232421875,
0.84228515625,
0.313232421875,
0.84228515625,
0.326904296875,
0.1568603515625,
0.326904296875
],
[
0.1568603515625,
0.32568359375,
0.8408203125,
0.32568359375,
0.8408203125,
0.339111328125,
0.1568603515625,
0.339111328125
],
[
0.156005859375,
0.341552734375,
0.84228515625,
0.34033203125,
0.84228515625,
0.354248046875,
0.156005859375,
0.354736328125
],
[
0.1568603515625,
0.354736328125,
0.84228515625,
0.354736328125,
0.84228515625,
0.36865234375,
0.1568603515625,
0.36865234375
],
[
0.1568603515625,
0.36865234375,
0.84228515625,
0.36865234375,
0.84228515625,
0.3818359375,
0.1568603515625,
0.3818359375
],
[
0.1568603515625,
0.38330078125,
0.84228515625,
0.38330078125,
0.84228515625,
0.396484375,
0.1568603515625,
0.396484375
],
[
0.158447265625,
0.397705078125,
0.8408203125,
0.397705078125,
0.8408203125,
0.40771484375,
0.158447265625,
0.40771484375
],
[
0.156005859375,
0.410400390625,
0.84228515625,
0.410400390625,
0.84228515625,
0.423583984375,
0.156005859375,
0.423583984375
],
[
0.1568603515625,
0.423583984375,
0.8408203125,
0.423583984375,
0.8408203125,
0.4375,
0.1568603515625,
0.4375
],
[
0.158447265625,
0.439453125,
0.8408203125,
0.439453125,
0.8408203125,
0.449462890625,
0.158447265625,
0.449462890625
],
[
0.1568603515625,
0.451904296875,
0.84228515625,
0.451904296875,
0.84228515625,
0.46533203125,
0.1568603515625,
0.46533203125
],
[
0.156005859375,
0.46533203125,
0.84228515625,
0.46533203125,
0.84228515625,
0.479248046875,
0.156005859375,
0.479248046875
],
[
0.1568603515625,
0.479248046875,
0.84326171875,
0.479248046875,
0.84326171875,
0.492431640625,
0.1568603515625,
0.492431640625
],
[
0.1568603515625,
0.49365234375,
0.84228515625,
0.49365234375,
0.84228515625,
0.5068359375,
0.1568603515625,
0.5068359375
],
[
0.1568603515625,
0.5068359375,
0.84228515625,
0.5068359375,
0.84228515625,
0.52099609375,
0.1568603515625,
0.52099609375
],
[
0.1568603515625,
0.52099609375,
0.84326171875,
0.52099609375,
0.84326171875,
0.5341796875,
0.1568603515625,
0.5341796875
],
[
0.1568603515625,
0.53515625,
0.84228515625,
0.53515625,
0.84228515625,
0.548828125,
0.1568603515625,
0.548828125
],
[
0.1568603515625,
0.548828125,
0.734375,
0.548828125,
0.734375,
0.5625,
0.1568603515625,
0.5625
],
[
0.11517333984375,
0.5869140625,
0.1650390625,
0.5869140625,
0.1650390625,
0.5986328125,
0.11517333984375,
0.5986328125
],
[
0.160888671875,
0.5869140625,
0.3095703125,
0.5869140625,
0.3095703125,
0.60107421875,
0.160888671875,
0.60107421875
],
[
0.11767578125,
0.6162109375,
0.88232421875,
0.6162109375,
0.88232421875,
0.6298828125,
0.11767578125,
0.6298828125
],
[
0.1192626953125,
0.63330078125,
0.8798828125,
0.63330078125,
0.8798828125,
0.64306640625,
0.1192626953125,
0.64306640625
],
[
0.11767578125,
0.646484375,
0.88134765625,
0.646484375,
0.88134765625,
0.66015625,
0.11767578125,
0.66015625
],
[
0.1192626953125,
0.66357421875,
0.8798828125,
0.66357421875,
0.8798828125,
0.673828125,
0.1192626953125,
0.673828125
],
[
0.1192626953125,
0.6787109375,
0.8798828125,
0.6787109375,
0.8798828125,
0.689453125,
0.1192626953125,
0.689453125
],
[
0.11767578125,
0.6923828125,
0.88232421875,
0.6923828125,
0.88232421875,
0.70556640625,
0.11767578125,
0.70556640625
],
[
0.11767578125,
0.708984375,
0.7099609375,
0.708984375,
0.7099609375,
0.7197265625,
0.11767578125,
0.7197265625
],
[
0.11767578125,
0.72900390625,
0.88134765625,
0.72900390625,
0.88134765625,
0.7421875,
0.11767578125,
0.7421875
],
[
0.11767578125,
0.7470703125,
0.8798828125,
0.7470703125,
0.8798828125,
0.7568359375,
0.11767578125,
0.7568359375
],
[
0.1168212890625,
0.76025390625,
0.88134765625,
0.76025390625,
0.88134765625,
0.7734375,
0.1168212890625,
0.7734375
],
[
0.11517333984375,
0.7744140625,
0.88134765625,
0.77587890625,
0.88134765625,
0.7890625,
0.11517333984375,
0.78857421875
],
[
0.11517333984375,
0.79052734375,
0.88232421875,
0.79052734375,
0.88232421875,
0.8037109375,
0.11517333984375,
0.8037109375
],
[
0.1168212890625,
0.8046875,
0.88134765625,
0.8046875,
0.88134765625,
0.818359375,
0.1168212890625,
0.818359375
],
[
0.1168212890625,
0.82080078125,
0.88134765625,
0.82080078125,
0.88134765625,
0.833984375,
0.1168212890625,
0.833984375
],
[
0.11517333984375,
0.833984375,
0.88134765625,
0.83544921875,
0.88134765625,
0.849609375,
0.11517333984375,
0.8486328125
],
[
0.11358642578125,
0.85107421875,
0.88134765625,
0.849609375,
0.88134765625,
0.8642578125,
0.11358642578125,
0.86572265625
],
[
0.136474609375,
0.87548828125,
0.273681640625,
0.87548828125,
0.273681640625,
0.88623046875,
0.136474609375,
0.88623046875
],
[
0.1356201171875,
0.88623046875,
0.7685546875,
0.88623046875,
0.7685546875,
0.8994140625,
0.1356201171875,
0.8994140625
],
[
0.49267578125,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94873046875,
0.49267578125,
0.94873046875
]
] |
[
"用于基础模型组健壮性的对比适配器",
"Michael Zhang*and Christopher Ré",
"Stanford University",
"fmzhang,chrismre@cs.stanford.edu",
"Abstract",
"While large pretrained foundation models (FMs) have shown remarkable zero-shot classification robust-",
"ness to dataset-level distribution shifts, their robustness to subpopulation or group shifts is relatively",
"underexplored. We study this problem, and find that FMs such as CLIP may not be robust to various",
"group shifts. Across 9 robustness benchmarks, zero-shot classification with their embeddings results in",
"gaps of up to 80.7 percentage points (pp) between average and worst-group accuracy. Unfortunately,",
"existing methods to improve robustness require retraining, which can be prohibitively expensive on large",
"foundation models. We also find that efficient ways to improve model inference ( e.g., via adapters,",
"lightweightnetworkswithFMembeddingsasinputs)donotconsistentlyimproveandcansometimes",
"hurt group robustness compared to zero-shot ( e.g., increasing the accuracy gap by 50.1 pp on CelebA).",
"We thus develop an adapter training strategy to effectively and efficiently improve FM group robustness.",
"Our motivating observation isthat whilepoorrobustnessresults from groups inthe same classbeing",
"embedded far apart in the foundation model “\"embedding space,\" standard adapter training may not",
"bring these points closer together. We thus propose contrastive adapting, which trains adapters with",
"contrastive learning to bring sample embeddings close to both their ground-truth class embeddings and",
"other sample embeddings in the same class. Across the 9 benchmarks, our approach consistently improves",
"group robustness, raising worst-group accuracy by 8.5 to 56.0 pp over zero-shot. Our approach is also",
"efficient, doing so without any FM finetuning and only a fixed set of frozen FM embeddings. On bench-",
"marks such as Waterbirds and CelebA, this leads to worst-group accuracy comparable to state-of-the-art",
"methods that retrain entire models, while only training ≤ 1 % of the model parameters",
"1",
"Introduction",
"基础模型（FMs）-在海量数据集上训练的大型预训练模型-为深度学习提供了令人兴奋的新范例。最近的工",
"作表明，在没有任何微调的情况下，基础模型可以很好地推厂广到各种数据集10，30，53，61，并对某些分布",
"偏移[36，67]表现出令人印象深刻的鲁棒性。在这种零拍范式下，从业者可以避免训练任务特定的模型，而",
"是使用FM嵌人进行高效和有效的推理。然而，一个未被充分开发的问题是，这种零拍推断对“群体转移”",
"有多稳健，数据中的亚群或有意义的群体之间的分布转移。先前的工作已经确定，组鲁棒性-i.e.在所有组",
"上都表现良好-是现代深度学习4，11，34，45，49，58，63的基本和现实挑战。然而，大多数先前的基础模型",
"评估都集中在整体或平均性能136，53.671上：很少有作品考虑其在小组中的准确性",
"因此，这项工作进一步研究了用于零拍分类的基础模型的组鲁棒性。我们首先通过证明基础模型可能具有较",
"差的零拍组鲁棒性来激发此问题。在9健狂性基准测试中评估11基础模型时，我们发现它们达到了80。平",
"均和最差组精度之间的差距为7个百分点（pp），其中CLIPResNet-50E[53]仅分类为6。0在CelebA数据",
"集上正确的最差组样本的%[41，57]。鉴于这一观察到的弱点，我们的目标是提高基础模型组的鲁棒性。这",
"带来了几个挑战和有待回答的问题。首先，尽管对机器学习中的组鲁棒性进行了深入研究，但现有的鲁棒性",
"方法需要重新训练一个（通常是多个）整个模型[1，15，33，40，45，57，63，64，70]。由于基础模型的大小和",
"规模，这对于基础模型来说可能是非常昂贵的，这引发了一个问题，即我们是否可以在不进行任何重新训练",
"或微调的情况下使这些模型更加健壮。其次，对于零拍分类，许多从业者也可能只访问基础模型输出或嵌入",
"（e.9.，通过API1）。为了提高鲁棒性，理想的解决方案应该只需要预先训练的FM嵌入。但是，这些相同",
"*Corresponding author",
"\"https://beta.openai.com/docs/introduction.,https://studio.ai21.com/docs/,https://docs.cohere.ai/",
"2"
] |
{
"arxiv_id": "2207.07180",
"language": "zh",
"page_idx": 1
}
|
||
2207.07180-zh-2
|
2207.07180-zh-2.jpg
|
<FigureHere>
Figure 1: (a) Adapter classification with FM embeddings. Adapters learn transformations to align sample embeddings to ground-truth class embeddings. (b) Cross-entropy loss encourages alignment between class embeddings [21] . (c) Contrastive adapting adds alignment between sample embeddings.
的嵌人会导致零拍鲁棒性较差，从而引发了一个问题，即它们是否甚至对正确分类所有组所需的信息进行了
编码
受这些挑战和问题的启发，我们研究了有效且高效的解决方案，以提高 FM 组的鲁棒性。作为基线，我们首先发现，虽然改进 FM 推理的有效方法一-例如在 FM 嵌人之上训练线性探针 (线性分类器) [36,53] 和适配器 (小瓶颈 MLP) [21,27]一可以提高零触发上的组鲁棒性 (将差距减少多达 50。2pponrepresentativebenchments),theyfailtodosoconsistently,andcanhurtrobustness。他们将最差组的准确性降低了多达 37。9pp,精度差距增加高达 74。9pp。对于这种不一致的原因，我们注意到，当FM 在嵌人空间中“相距遥远” 的不同组中嵌人相同类别的样本时，会导致零拍鲁棒性差。虽然适配器训练比线性探测实现了更高的鲁棒性，但我们发现如果训练数据是组不平衡的，它仍然无法关闭这个距离$e.g.$,的设置
然后，为了处理基线损害鲁棒性的设置，并在零拍上始终如一地提高组的鲁棒性，我们提出了 contrastive adapting ,这是一种简单的适配器训练方法，它更加强调将这些最初“相距遥远”的点放在一起。对于每个任务，我们首先使用基础模型来计算每个训练样本和类的嵌人。然后我们在这些嵌人上训练适配器。与先前的工作$\mathbb{E}[21]$ 一样，这些适配器将示例嵌人作为输入，并输出与基本事实类嵌入具有更大余弦相似性的转换嵌人。然而，关键的区别是 contrastive adapting 也适用于其他样本嵌入的监督对比损失。具体来说，我们提供了一种在同一类中“拉在一起”相距遥远的样本嵌人，并在不同类中“推开”附近的样本嵌人的方法
在我们的实验中，我们验证了 contrastive adapting 有效且高效地提高了 FM 组的鲁棒性。首先，在所有 9个健壮性基准测试中，我们发现 contrastive adapting 相对于零拍始终提高了最差组的准确性(提高了8。5至 56。0pp), 不使用训练组标签，只使用0个训练 MLP。将 1 个% 转换为0。图 3% 的原始 FM 参数。然后，在具有各种组移位和训练数据组大小的代表性基准集上，我们发现 contrastive adapting 可以比以前的适配器训练策略大大提高组的鲁棒性，并且优于仅使用固定 FM 嵌人的其他方法 (将最差组精度提高 12。4pp, 平均精度为0。比 CLIPResNet-50 嵌人的下一个最佳鲁棒性方法平均高出 6pp)。最后，除了提高调频鲁棒性，我们发现 contrastive adapting 也实现了有效和高效的群体鲁棒分类。我们在流行的鲁棒性基准上实现了接近最先进的 (SoTA) 或 SoTA 最差组精度，只有 1。0SoTA 模型可训练参数的%( e.g.,改进 0。通过训练单个适配器而不是 ResNet-50, 在 CelebAE[41]上比之前的 SoTAE[46]高出 2pp)
我们总结了我们的贡献如下：
· We demonstrate that zero-shot classification with existing foundation models may not be group robust
across a variety of popular group robustness benchmarks.
· We propose a contrastive adapter training strategy that effciently improves foundation model group
robustness without any finetuning of the original model weights.
$\bullet$ We find that contrastive adapting more effectively improves foundation model group robustness over sim-
ilarly efficient alternatives that do not require finetuning, and more efficiently achieves comparable group robustness to state-of-the-art approaches outside the foundation model regime on standard benchmarks.
3
|
[
[
0.5107421875,
0.085205078125,
0.58837890625,
0.08587646484375,
0.58837890625,
0.09783935546875,
0.5107421875,
0.09661865234375
],
[
0.2548828125,
0.0947265625,
0.374267578125,
0.0947265625,
0.374267578125,
0.10797119140625,
0.2548828125,
0.10797119140625
],
[
0.52685546875,
0.09661865234375,
0.57177734375,
0.0985107421875,
0.5703125,
0.10986328125,
0.525390625,
0.10797119140625
],
[
0.1422119140625,
0.10418701171875,
0.1690673828125,
0.10418701171875,
0.1690673828125,
0.121826171875,
0.1422119140625,
0.121826171875
],
[
0.654296875,
0.11236572265625,
0.71337890625,
0.11236572265625,
0.71337890625,
0.12249755859375,
0.654296875,
0.12249755859375
],
[
0.6611328125,
0.12371826171875,
0.70361328125,
0.12371826171875,
0.70361328125,
0.1337890625,
0.6611328125,
0.1337890625
],
[
0.23779296875,
0.1610107421875,
0.28515625,
0.1610107421875,
0.28515625,
0.175537109375,
0.23779296875,
0.175537109375
],
[
0.1405029296875,
0.184326171875,
0.1641845703125,
0.184326171875,
0.1641845703125,
0.1956787109375,
0.1405029296875,
0.1956787109375
],
[
0.29736328125,
0.1900634765625,
0.310546875,
0.1900634765625,
0.310546875,
0.1944580078125,
0.29736328125,
0.1944580078125
],
[
0.400390625,
0.1900634765625,
0.426513671875,
0.1900634765625,
0.426513671875,
0.1988525390625,
0.400390625,
0.1988525390625
],
[
0.365234375,
0.208984375,
0.39794921875,
0.208984375,
0.39794921875,
0.2215576171875,
0.365234375,
0.2215576171875
],
[
0.273681640625,
0.23486328125,
0.2998046875,
0.23486328125,
0.2998046875,
0.25244140625,
0.273681640625,
0.25244140625
],
[
0.5390625,
0.2374267578125,
0.56201171875,
0.2374267578125,
0.56201171875,
0.251953125,
0.5390625,
0.251953125
],
[
0.7509765625,
0.2374267578125,
0.77392578125,
0.2374267578125,
0.77392578125,
0.251953125,
0.7509765625,
0.251953125
],
[
0.1168212890625,
0.26953125,
0.88134765625,
0.26953125,
0.88134765625,
0.28271484375,
0.1168212890625,
0.28271484375
],
[
0.1168212890625,
0.2841796875,
0.88134765625,
0.2841796875,
0.88134765625,
0.29736328125,
0.1168212890625,
0.29736328125
],
[
0.11517333984375,
0.2998046875,
0.7548828125,
0.2998046875,
0.7548828125,
0.313232421875,
0.11517333984375,
0.313232421875
],
[
0.1192626953125,
0.34033203125,
0.88134765625,
0.34033203125,
0.88134765625,
0.354248046875,
0.1192626953125,
0.354248046875
],
[
0.1168212890625,
0.35595703125,
0.1502685546875,
0.35595703125,
0.1502685546875,
0.367431640625,
0.1168212890625,
0.367431640625
],
[
0.11767578125,
0.37744140625,
0.88134765625,
0.37744140625,
0.88134765625,
0.391357421875,
0.11767578125,
0.391357421875
],
[
0.11767578125,
0.393310546875,
0.88134765625,
0.393310546875,
0.88134765625,
0.4072265625,
0.11767578125,
0.4072265625
],
[
0.11767578125,
0.40771484375,
0.88134765625,
0.40771484375,
0.88134765625,
0.421630859375,
0.11767578125,
0.421630859375
],
[
0.1168212890625,
0.423583984375,
0.88232421875,
0.423583984375,
0.88232421875,
0.4375,
0.1168212890625,
0.4375
],
[
0.11517333984375,
0.438232421875,
0.8798828125,
0.4375,
0.8798828125,
0.451904296875,
0.11517333984375,
0.45263671875
],
[
0.1168212890625,
0.451904296875,
0.88232421875,
0.45263671875,
0.88232421875,
0.46728515625,
0.1168212890625,
0.466552734375
],
[
0.1168212890625,
0.468505859375,
0.71337890625,
0.468505859375,
0.71337890625,
0.48291015625,
0.1168212890625,
0.48291015625
],
[
0.11767578125,
0.4912109375,
0.88134765625,
0.4912109375,
0.88134765625,
0.5048828125,
0.11767578125,
0.5048828125
],
[
0.11517333984375,
0.505859375,
0.88134765625,
0.5048828125,
0.88134765625,
0.51953125,
0.11517333984375,
0.52099609375
],
[
0.11767578125,
0.521484375,
0.88134765625,
0.521484375,
0.88134765625,
0.53515625,
0.11767578125,
0.53515625
],
[
0.11767578125,
0.5361328125,
0.88134765625,
0.5361328125,
0.88134765625,
0.5498046875,
0.11767578125,
0.5498046875
],
[
0.1168212890625,
0.5517578125,
0.88134765625,
0.5517578125,
0.88134765625,
0.56494140625,
0.1168212890625,
0.56494140625
],
[
0.11767578125,
0.5673828125,
0.88134765625,
0.5673828125,
0.88134765625,
0.58056640625,
0.11767578125,
0.58056640625
],
[
0.11517333984375,
0.58349609375,
0.1356201171875,
0.58349609375,
0.1356201171875,
0.59521484375,
0.11517333984375,
0.59521484375
],
[
0.11767578125,
0.60400390625,
0.88232421875,
0.60400390625,
0.88232421875,
0.61865234375,
0.11767578125,
0.61865234375
],
[
0.1168212890625,
0.61865234375,
0.88232421875,
0.619140625,
0.88232421875,
0.63330078125,
0.1168212890625,
0.6318359375
],
[
0.1168212890625,
0.6337890625,
0.88134765625,
0.63330078125,
0.88134765625,
0.64794921875,
0.1168212890625,
0.6484375
],
[
0.11767578125,
0.6494140625,
0.88134765625,
0.6494140625,
0.88134765625,
0.66357421875,
0.11767578125,
0.66357421875
],
[
0.11767578125,
0.66650390625,
0.87841796875,
0.66650390625,
0.87841796875,
0.6767578125,
0.11767578125,
0.6767578125
],
[
0.11767578125,
0.6796875,
0.88134765625,
0.6796875,
0.88134765625,
0.693359375,
0.11767578125,
0.693359375
],
[
0.1192626953125,
0.6962890625,
0.8798828125,
0.6962890625,
0.8798828125,
0.70703125,
0.1192626953125,
0.70703125
],
[
0.11767578125,
0.71044921875,
0.88134765625,
0.71044921875,
0.88134765625,
0.724609375,
0.11767578125,
0.724609375
],
[
0.1168212890625,
0.724609375,
0.75732421875,
0.724609375,
0.75732421875,
0.73779296875,
0.1168212890625,
0.73779296875
],
[
0.11767578125,
0.75,
0.318603515625,
0.75,
0.318603515625,
0.76025390625,
0.11767578125,
0.76025390625
],
[
0.1192626953125,
0.76806640625,
0.88134765625,
0.77099609375,
0.88134765625,
0.78515625,
0.1192626953125,
0.78271484375
],
[
0.134033203125,
0.7861328125,
0.5498046875,
0.7841796875,
0.5498046875,
0.798828125,
0.134033203125,
0.80029296875
],
[
0.1192626953125,
0.80810546875,
0.88134765625,
0.80810546875,
0.88134765625,
0.82275390625,
0.1192626953125,
0.82275390625
],
[
0.1356201171875,
0.82373046875,
0.6005859375,
0.82373046875,
0.6005859375,
0.8369140625,
0.1356201171875,
0.8369140625
],
[
0.1192626953125,
0.84521484375,
0.88232421875,
0.84521484375,
0.88232421875,
0.85986328125,
0.1192626953125,
0.85986328125
],
[
0.134033203125,
0.85986328125,
0.88134765625,
0.861328125,
0.88134765625,
0.87548828125,
0.134033203125,
0.875
],
[
0.1356201171875,
0.876953125,
0.88134765625,
0.876953125,
0.88134765625,
0.89013671875,
0.1356201171875,
0.89013671875
],
[
0.491943359375,
0.93505859375,
0.5087890625,
0.93505859375,
0.5087890625,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"Cross-entropy",
"AdapterClassification",
"Adapter",
"Clase",
"Contrastive",
"Adapte",
"Adapter",
"Eamo",
"T",
"f(u)",
"y=1",
"(a)",
"(b)",
"(c)",
"Figure l: (a) Adapter classification with FM embeddings. Adapters learn transformations to align sample",
"embeddings to ground-truth class embeddings. (b) Cross-entropy loss encourages alignment between class",
"embeddings [21] : (c) Contrastive adapting adds alignment between sample embeddings.",
"的嵌人会导致零拍鲁棒性较差，从而引发了一个问题，即它们是否甚至对正确分类所有组所需的信息进行了",
"编码",
"受这些挑战和问题的启发，我们研究了有效且高效的解决方案，以提高FM组的鲁棒性。作为基线，我",
"们首先发现，虽然改进FM推理的有效方法一-例如在FM嵌人之上训练线性探针（线性分类器）[36，53]",
"和适配器（小瓶颈MLP）[21，27]一可以提高零触发上的组鲁棒性（将差距减少多达50。2pponrepresen",
"tativebenchments),theyfailtodosoconsistently,andcanhurtrobustness。他们将最差组的准确性降低了多达37",
"9pP，精度差距增加高达74。9pP。对于这种不一致的原因，我们注意到，当FM在嵌人空间中“相距遥远”",
"的不同组中嵌入相同类别的样本时，会导致零拍鲁棒性差。虽然适配器训练比线性探测实现了更高的鲁棒",
"性，但我们发现如果训练数据是组不平衡的，它仍然无法关闭这个距离e.9.，的设置",
"然后，为了处理基线损害鲁棒性的设置，并在零拍上始终如一地提高组的鲁棒性，我们提出了contrastive",
"adapting，这是一种简单的适配器训练方法，它更加强调将这些最初“相距遥远”的点放在一起。对于每个",
"任务，我们首先使用基础模型来计算每个训练样本和类的嵌入。然后我们在这些嵌人上训练适配器。与先",
"前的工作E21]一样，这些适配器将示例嵌入作为输入，并输出与基本事实类嵌人具有更大余弦相似性的转",
"换嵌人。然而，关键的区别是contrastiveadapting也适用于其他样本嵌入的监督对比损失。具体来说，我",
"们提供了一种在同一类中“拉在一起”相距遥远的样本嵌入，并在不同类中“推开”附近的样本嵌入的方",
"法",
"在我们的实验中，我们验证了contrastive adapting有效且高效地提高了FM组的鲁棒性。首先，在所有9",
"个健壮性基准测试中，我们发现contrastiveadapting相对于零拍始终提高了最差组的准确性（提高了8。5",
"至56。0pP)，不使用训练组标签，只使用0个训练MLP。将1个%转换为0。图3%的原始FM参数。然",
"后，在具有各种组移位和训练数据组大小的代表性基准集上，我们发现contrastiveadapting可以比以前的",
"适配器训练策略大大提高组的鲁棒性，并且优于仅使用固定FM嵌入的其他方法（将最差组精度提高12",
"4pP，平均精度为0。比CLIPResNet-50嵌人的下一个最佳鲁棒性方法平均高出6pp)。最后，除了提高调频",
"鲁棒性：我们发现contrastiveadapting也实现了有效和高效的群体鲁棒分类。我们在流行的鲁棒性基准上",
"实现了接近最先进的（SoTA）或SoTA最差组精度，只有1。0SoTA模型可训练参数的%（e.g，改进0。通",
"过训练单个适配器而不是ResNet-50，在CelebAE[41]上比之前的SoTAE[46]高出2pp)",
"我们总结了我们的贡献如下：",
". We demonstrate that zero-shot classification with existing foundation models may not be group robust",
"across a variety of popular group robustness benchmarks.",
". We propose a contrastive adapter training strategy that efficiently improves foundation model group",
"robustness without any finetuning of the original model weights.",
". We find that contrastive adapting more effectively improves foundation model group robustness over sim-",
"ilarly efficient alternatives that do not require finetuning, and more efficiently achieves comparable group",
"robustness to state-of-the-art approaches outside the foundation model regime on standard benchmarks.",
"3"
] |
{
"arxiv_id": "2207.07180",
"language": "zh",
"page_idx": 2
}
|
||
2403.11752-zh-0
|
2403.11752-zh-0.jpg
|
重温经典：识别和纠正押韵和诗歌中的性别刻板印象的研究
# Aditya Narayan Sankaran$^{*\dagger}$,Vigneshwaran Shankaran$^{*\S}$ $\dot{\text{Sampath Lonka}^{+}}$,Rajesh Sharma$^{\|\perp}$ $^{\dagger}$Sri Sathya Sai Institute of Higher Learning,
$^{8}$GESIS- Leibniz Institute for the Social Sciences,
$^{\ddagger}$Citridot Solutions, "University of Tartu, $^\perp$CARDS, IIT Ropar
$\{$ adityanarayans@sssihl.edu.in, vigneshwaran.shankaran@gesis.org,
Isampath@citridot.com, rajesh.sharma@ut.ee $\}$
Abstract
Rhymes and poems are a powerful medium for transmitting cultural norms and societal roles. However, the pervasive existence of gender stereotypes in these works perpetuates biased perceptions and limits the scope of individuals'identities. Past works have shown that stereotyping and prejudice emerge in early childhood, and developmental research on causal mechanisms is critical for understanding and controlling stereotyping and prejudice. This work contributes by gathering a dataset of rhymes and poems to identify gender stereotypes and propose a model with 97 % accuracy to identify gender bias. Gender stereotypes were rectiffed using a Large Language Model (LLM) and its effectiveness was evaluated in $a$ comparative survey against human educator rectifications. To summarize, this work highlights the pervasive nature of gender stereotypes in literary works and reveals the potential of LLMs to rectify gender stereotypes. This study raises awareness and promotes inclusivity within artistic expressions, making a significant contribution to the discourse on gender equality.
Keywords: Gender Stereotypes, Bias, Large-Language Models
## 1. Introduction
<FigureHere>
$(\mathbf{b})$ Stereotypical
Figure 1: Example of poems with and without ex-
plicit gender stereotype
通过经验获取知识是身份发展的基本过程。人类倾向于占据与其性别 (Blackstone, 2003) 相关的角色，我们称之为性别角色。性别角色是社会建构的位置或行为，由个人根据其性别认同和流行的文化规范学习和执行 (Bigler and Liben,2007)。这种学习始于幼儿接触故事和押韵，这是语言习得的基本学习实践，也有助于他们理解社会是如何运作的。。图 1a 是教给孩子们的押韵，简短而引人人胜，同时清楚地传达了其固有的信息，即食用豆类等健康食品的好处
虽然许多童谣传达了有价值的信息并教授语言技能，但有些童谣包含的过时信息不再与当代价值观保持一致。图 1b , 一首看似幽默的关于婚姻问题的诗，通过描绘丈夫对妻子的控制，延续了父权制的价值观。尽管社会取得了进步，但这些押韵仍继续在世界各地的托儿所中教给幼儿
到孩子成年时，无论有意识还是无意识，他们都会内化许多性别刻板印象。这些刻板印象可以以各种方式塑造我们的思想，感受和行为。过去，在政治，军事和执法等领域，男性占主导地位，而在儿童保育，医疗保健和社会工作等与护理相关的职业中，女性占主导地位。这些专业角色是典型的男性和女性行为的例证，其根源在于我们的文化(Diamond,2002)。这些发现使我们在研究中将性别刻板印象定义为基于性别对男性和女性的典型特征，角色和行为做出的概括或假设。这些刻板印象往往被个人从小就内化，可以塑造他们对自己和他人的信仰，态度和行为 (Haines et al.,2016) 。性
别刻板印象可能是有害的，因为它们限制了个人的潜力或仅基于其性别而产生不公平的期望，从而促进了毒性和厌女症(Heilman,2001)
知识获取的主要模式源于教材，如教科书和口头教学。值得注意的是，在学校教科书中观察到男性倾向于女性，特别是在东南亚国家(Islam and Asadullah,2018)。调查显示，文字和图形指标的综合显示，女性的总体代表性明显较低，仅占 40。$4\%$。此外，值得注意的是，源自著名的“鹅妈妈” 系列的流行儿童押韵延续了性别歧视意识形态，并明确认可针对女性的性别陈规定型观念 (Nadesan, 1974)
可以说，18$^\mathrm{th}$和20$^\mathrm{th}$世纪之间的押韵和诗歌反映了这些时代的社会规范。但是，此后社会发生了重大变化，这些押韵和诗歌中存在的一些概念不再被认为是适当的。因此，教育工作者需要注意这些
$^{*}$Equal contribution
|
[
[
0.189697265625,
0.097412109375,
0.8134765625,
0.09857177734375,
0.8134765625,
0.1134033203125,
0.189697265625,
0.11224365234375
],
[
0.2392578125,
0.1383056640625,
0.76318359375,
0.1383056640625,
0.76318359375,
0.1519775390625,
0.2392578125,
0.1519775390625
],
[
0.3359375,
0.15380859375,
0.66650390625,
0.152587890625,
0.66650390625,
0.16748046875,
0.3359375,
0.168701171875
],
[
0.340087890625,
0.168701171875,
0.66162109375,
0.169189453125,
0.66162109375,
0.182861328125,
0.340087890625,
0.1817626953125
],
[
0.31494140625,
0.1817626953125,
0.68408203125,
0.18408203125,
0.68408203125,
0.198974609375,
0.31494140625,
0.196533203125
],
[
0.276123046875,
0.196533203125,
0.724609375,
0.19775390625,
0.724609375,
0.2120361328125,
0.276123046875,
0.21142578125
],
[
0.2418212890625,
0.213134765625,
0.75732421875,
0.213134765625,
0.75732421875,
0.226806640625,
0.2418212890625,
0.226806640625
],
[
0.32568359375,
0.22802734375,
0.67822265625,
0.22802734375,
0.67822265625,
0.2410888671875,
0.32568359375,
0.2410888671875
],
[
0.465087890625,
0.2452392578125,
0.53662109375,
0.2452392578125,
0.53662109375,
0.260009765625,
0.465087890625,
0.260009765625
],
[
0.12005615234375,
0.26123046875,
0.8818359375,
0.26123046875,
0.8818359375,
0.27490234375,
0.12005615234375,
0.27490234375
],
[
0.12005615234375,
0.273681640625,
0.880859375,
0.273681640625,
0.880859375,
0.287353515625,
0.12005615234375,
0.287353515625
],
[
0.12176513671875,
0.289306640625,
0.8818359375,
0.289306640625,
0.8818359375,
0.2998046875,
0.12176513671875,
0.2998046875
],
[
0.118408203125,
0.2998046875,
0.88330078125,
0.300537109375,
0.88330078125,
0.314208984375,
0.118408203125,
0.3134765625
],
[
0.12005615234375,
0.312255859375,
0.8818359375,
0.312255859375,
0.8818359375,
0.325927734375,
0.12005615234375,
0.325927734375
],
[
0.12005615234375,
0.325927734375,
0.880859375,
0.325927734375,
0.880859375,
0.339111328125,
0.12005615234375,
0.339111328125
],
[
0.118408203125,
0.339111328125,
0.880859375,
0.339111328125,
0.880859375,
0.352783203125,
0.118408203125,
0.352783203125
],
[
0.12005615234375,
0.352783203125,
0.88330078125,
0.352783203125,
0.88330078125,
0.366455078125,
0.12005615234375,
0.366455078125
],
[
0.12005615234375,
0.365234375,
0.8818359375,
0.365234375,
0.8818359375,
0.37890625,
0.12005615234375,
0.37890625
],
[
0.118408203125,
0.37890625,
0.75390625,
0.37890625,
0.75390625,
0.392578125,
0.118408203125,
0.392578125
],
[
0.12005615234375,
0.405029296875,
0.54248046875,
0.405029296875,
0.54248046875,
0.418701171875,
0.12005615234375,
0.418701171875
],
[
0.5634765625,
0.431640625,
0.779296875,
0.431640625,
0.779296875,
0.4453125,
0.5634765625,
0.4453125
],
[
0.5634765625,
0.44775390625,
0.81689453125,
0.44775390625,
0.81689453125,
0.46142578125,
0.5634765625,
0.46142578125
],
[
0.2166748046875,
0.46142578125,
0.382080078125,
0.4619140625,
0.382080078125,
0.476806640625,
0.2166748046875,
0.4755859375
],
[
0.56494140625,
0.46435546875,
0.8271484375,
0.46435546875,
0.8271484375,
0.47802734375,
0.56494140625,
0.47802734375
],
[
0.5634765625,
0.479736328125,
0.8037109375,
0.479736328125,
0.8037109375,
0.493408203125,
0.5634765625,
0.493408203125
],
[
0.12005615234375,
0.4912109375,
0.486083984375,
0.4912109375,
0.486083984375,
0.5048828125,
0.12005615234375,
0.5048828125
],
[
0.12176513671875,
0.5048828125,
0.48779296875,
0.5048828125,
0.48779296875,
0.51953125,
0.12176513671875,
0.51953125
],
[
0.62890625,
0.50439453125,
0.767578125,
0.50439453125,
0.767578125,
0.51416015625,
0.62890625,
0.51416015625
],
[
0.12005615234375,
0.5185546875,
0.48779296875,
0.51953125,
0.48779296875,
0.533203125,
0.12005615234375,
0.5322265625
],
[
0.5615234375,
0.5185546875,
0.75146484375,
0.52099609375,
0.75146484375,
0.53515625,
0.5615234375,
0.533203125
],
[
0.12176513671875,
0.53369140625,
0.48779296875,
0.53369140625,
0.48779296875,
0.54736328125,
0.12176513671875,
0.54736328125
],
[
0.5634765625,
0.5361328125,
0.796875,
0.5361328125,
0.796875,
0.5498046875,
0.5634765625,
0.5498046875
],
[
0.12005615234375,
0.54736328125,
0.48779296875,
0.54736328125,
0.48779296875,
0.56103515625,
0.12005615234375,
0.56103515625
],
[
0.5634765625,
0.5517578125,
0.77099609375,
0.5517578125,
0.77099609375,
0.5654296875,
0.5634765625,
0.5654296875
],
[
0.118408203125,
0.56103515625,
0.486083984375,
0.56103515625,
0.486083984375,
0.57470703125,
0.118408203125,
0.57470703125
],
[
0.5634765625,
0.56787109375,
0.7841796875,
0.56787109375,
0.7841796875,
0.58056640625,
0.5634765625,
0.58056640625
],
[
0.118408203125,
0.576171875,
0.48779296875,
0.57666015625,
0.48779296875,
0.59033203125,
0.118408203125,
0.5888671875
],
[
0.12005615234375,
0.59130859375,
0.489501953125,
0.59130859375,
0.489501953125,
0.60498046875,
0.12005615234375,
0.60498046875
],
[
0.6455078125,
0.59130859375,
0.75244140625,
0.59130859375,
0.75244140625,
0.6025390625,
0.6455078125,
0.6025390625
],
[
0.12005615234375,
0.60498046875,
0.48779296875,
0.60498046875,
0.48779296875,
0.6181640625,
0.12005615234375,
0.6181640625
],
[
0.12005615234375,
0.619140625,
0.202392578125,
0.619140625,
0.202392578125,
0.6298828125,
0.12005615234375,
0.6298828125
],
[
0.51318359375,
0.6142578125,
0.8818359375,
0.6142578125,
0.8818359375,
0.62744140625,
0.51318359375,
0.62744140625
],
[
0.51123046875,
0.62890625,
0.6884765625,
0.6298828125,
0.6884765625,
0.64453125,
0.51123046875,
0.6435546875
],
[
0.1376953125,
0.63525390625,
0.48779296875,
0.63525390625,
0.48779296875,
0.6455078125,
0.1376953125,
0.6455078125
],
[
0.118408203125,
0.646484375,
0.48779296875,
0.64794921875,
0.48779296875,
0.662109375,
0.118408203125,
0.6611328125
],
[
0.118408203125,
0.6611328125,
0.486083984375,
0.6611328125,
0.486083984375,
0.6748046875,
0.118408203125,
0.6748046875
],
[
0.12005615234375,
0.67578125,
0.48779296875,
0.67578125,
0.48779296875,
0.689453125,
0.12005615234375,
0.689453125
],
[
0.5146484375,
0.67578125,
0.8818359375,
0.67578125,
0.8818359375,
0.689453125,
0.5146484375,
0.689453125
],
[
0.12005615234375,
0.689453125,
0.489501953125,
0.689453125,
0.489501953125,
0.703125,
0.12005615234375,
0.703125
],
[
0.5146484375,
0.6904296875,
0.88330078125,
0.6904296875,
0.88330078125,
0.70361328125,
0.5146484375,
0.70361328125
],
[
0.12005615234375,
0.70361328125,
0.3896484375,
0.70361328125,
0.3896484375,
0.71728515625,
0.12005615234375,
0.71728515625
],
[
0.51318359375,
0.70361328125,
0.7734375,
0.70361328125,
0.7734375,
0.71728515625,
0.51318359375,
0.71728515625
],
[
0.1368408203125,
0.71875,
0.48779296875,
0.71875,
0.48779296875,
0.732421875,
0.1368408203125,
0.732421875
],
[
0.52880859375,
0.72021484375,
0.8818359375,
0.7216796875,
0.8818359375,
0.736328125,
0.52880859375,
0.7353515625
],
[
0.12005615234375,
0.73291015625,
0.48779296875,
0.73291015625,
0.48779296875,
0.74658203125,
0.12005615234375,
0.74658203125
],
[
0.5146484375,
0.736328125,
0.87890625,
0.736328125,
0.87890625,
0.74658203125,
0.5146484375,
0.74658203125
],
[
0.12005615234375,
0.74755859375,
0.486083984375,
0.74755859375,
0.486083984375,
0.75830078125,
0.12005615234375,
0.75830078125
],
[
0.5146484375,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76416015625,
0.5146484375,
0.76416015625
],
[
0.118408203125,
0.76123046875,
0.48779296875,
0.76025390625,
0.48779296875,
0.77392578125,
0.118408203125,
0.77490234375
],
[
0.51123046875,
0.76318359375,
0.8818359375,
0.76416015625,
0.8818359375,
0.779296875,
0.51123046875,
0.77783203125
],
[
0.12176513671875,
0.775390625,
0.489501953125,
0.775390625,
0.489501953125,
0.7890625,
0.12176513671875,
0.7890625
],
[
0.51318359375,
0.77783203125,
0.880859375,
0.77783203125,
0.880859375,
0.79150390625,
0.51318359375,
0.79150390625
],
[
0.12176513671875,
0.79150390625,
0.48779296875,
0.79150390625,
0.48779296875,
0.8017578125,
0.12176513671875,
0.8017578125
],
[
0.51318359375,
0.79150390625,
0.8818359375,
0.7919921875,
0.8818359375,
0.80712890625,
0.51318359375,
0.8056640625
],
[
0.118408203125,
0.802734375,
0.489501953125,
0.80419921875,
0.489501953125,
0.81787109375,
0.118408203125,
0.81640625
],
[
0.515625,
0.80859375,
0.87890625,
0.80859375,
0.87890625,
0.8193359375,
0.515625,
0.8193359375
],
[
0.118408203125,
0.81640625,
0.48779296875,
0.818359375,
0.48779296875,
0.83203125,
0.118408203125,
0.830078125
],
[
0.5146484375,
0.82177734375,
0.8818359375,
0.82177734375,
0.8818359375,
0.8349609375,
0.5146484375,
0.8349609375
],
[
0.12176513671875,
0.83447265625,
0.484375,
0.83447265625,
0.484375,
0.84423828125,
0.12176513671875,
0.84423828125
],
[
0.515625,
0.83447265625,
0.55908203125,
0.8359375,
0.5576171875,
0.8486328125,
0.5146484375,
0.8466796875
],
[
0.12005615234375,
0.8466796875,
0.48779296875,
0.8466796875,
0.48779296875,
0.85986328125,
0.12005615234375,
0.85986328125
],
[
0.53076171875,
0.8525390625,
0.880859375,
0.8525390625,
0.880859375,
0.8662109375,
0.53076171875,
0.8662109375
],
[
0.118408203125,
0.85986328125,
0.486083984375,
0.85986328125,
0.486083984375,
0.87353515625,
0.118408203125,
0.87353515625
],
[
0.5146484375,
0.86767578125,
0.880859375,
0.86767578125,
0.880859375,
0.88134765625,
0.5146484375,
0.88134765625
],
[
0.12005615234375,
0.87451171875,
0.48779296875,
0.87451171875,
0.48779296875,
0.88818359375,
0.12005615234375,
0.88818359375
],
[
0.5146484375,
0.8798828125,
0.8818359375,
0.88134765625,
0.8818359375,
0.8955078125,
0.5146484375,
0.89453125
],
[
0.1502685546875,
0.90087890625,
0.279541015625,
0.90087890625,
0.279541015625,
0.91455078125,
0.1502685546875,
0.91455078125
],
[
0.5146484375,
0.896484375,
0.88330078125,
0.896484375,
0.88330078125,
0.91015625,
0.5146484375,
0.91015625
]
] |
[
"重温经典：识别和纠正押韵和诗歌中的性别刻板印象的研究",
"AdityaNarayanSankaran*t,VigneshwaranShankaran*",
"SampathLonkat,Rajesh Sharmall",
"tSri Sathya Sai Instituteof Higher Learning,",
"GESIS - Leibniz Institute for the Social Sciences,",
"#Citridot Solutions, lIUniversity of Tartu,+ CARDS, IIT Ropar",
"{adityanarayans@sssihl.edu.in,vigneshwaran.shankaran@gesis.org",
"Isampath@citridot.com,rajesh.sharma@ut.ee)",
"Abstract",
"Rhymes and poems are a powerful medium for transmitting cultural norms and societal roles. However, the",
"pervasive existence of gender stereotypes in these works perpetuates biased perceptions and limits the scope",
"of individualsidentities.Pastworks have shown that stereotyping and prejudice emerge in early childhood, and",
"developmental research on causal mechanisms is critical for understanding and controlling stereotyping and",
"prejudice. This work contributes by gathering a dataset of rhymes and poems to identify gender stereotypes and",
"propose a model with 97 % accuracy to identify gender bias. Gender stereotypes were rectified using a Large",
"Language Model (LLM)and its effectiveness was evaluated in a comparative survey against human educator",
"rectifications. To summarize, this work highlights the pervasive nature of gender stereotypes in literary works and",
"reveals the potential of LLMs to rectify gender stereotypes. This study raises awareness and promotes inclusivity",
"within artistic expressions, making a significant contribution to the discourse on gender equality.",
"Keywords: Gender Stereotypes, Bias, Large-Language Models",
"Beans, beans, the magical fruit.",
"The more you eat, the more you toot,",
"1.Introduction",
"The more you toot, the better you feel.",
"So let's have beans for every meal!",
"通过经验获取知识是身份发展的基本过程。人类",
"倾向于占据与其性别（Blackstone，2003）相关的角",
"(a)Non-stereotypical",
"色，我们称之为性别角色。性别角色是社会建构的",
"Peter,Peterpumpkineater",
"位置或行为，由个人根据其性别认同和流行的文化",
"Had a wife but couldn't keep her;",
"规范学习和执行（BiglerandLiben，2007）。这种学",
"He put her in a pumpkin shell",
"习始于幼儿接触故事和押韵，这是语言习得的基本",
"And there he kept her very well.",
"学习实践，也有助于他们理解社会是如何运作的。。",
"图1a是教给孩子们的押韵，简短而引人人胜，同",
"(b)Stereotypical",
"时清楚地传达了其固有的信息，即食用豆类等健康",
"食品的好处",
"Figure 1: Example of poems with and without ex",
"plicit gender stereotype",
"虽然许多童谣传达了有价值的信息并教授语言",
"技能，但有些童谣包含的过时信息不再与当代价值",
"观保持一致。图1b，一首看似幽默的关于婚姻问",
"题的诗，通过描绘丈夫对妻子的控制，延续了父权",
"别刻板印象可能是有害的，因为它们限制了个人的",
"制的价值观。尽管社会取得了进步，但这些押韵仍",
"潜力或仅基于其性别而产生不公平的期望，从而促",
"继续在世界各地的托儿所中教给幼儿",
"进了毒性和厌女症(Heilman,2001)",
"到孩子成年时，无论有意识还是无意识，他们都",
"知识获取的主要模式源于教材，如教科书和口",
"会内化许多性别刻板印象。这些刻板印象可以以各",
"头教学。值得注意的是在学校教科书中观察至",
"种方式塑造我们的思想，感受和行为。过去，在政",
"男性倾向于女性，特别是在东南亚国家（lslamand",
"治，军事和执法等领域，男性占主导地位，而在儿",
"Asadullah，2018）。调查显示，文字和图形指标的",
"童保育，医疗保健和社会工作等与护理相关的职",
"综合显示，女性的总体代表性明显较低，仅占40",
"业中，女性占主导地位。。这些专业角色是典型的",
"4%。此外，值得注意的是，源自著名的“鹅妈妈”",
"男性和女性行为的例证，其根源在于我们的文化",
"系列的流行儿童押韵延续了性别歧视意识形态，并",
"(Diamond，2002）。这些发现使我们在研究中将性",
"明确认可针对女性的性别陈规定型观念（Nadesan,",
"别刻板印象定义为基于性别对男性和女性的典型",
"1974)",
"特征，角色和行为做出的概括或假设。这些刻板印",
"可以说，18th和20th世纪之间的押韵和诗歌反",
"象往往被个人从小就内化，可以塑造他们对自己和",
"映了这些时代的社会规范。但是，此后社会发生了",
"他人的信仰，态度和行为（Hainesetal.,2016）。性",
"重大变化，这些押韵和诗歌中存在的一些概念不再",
"Equalcontribution",
"被认为是适当的。因此，教育工作者需要注意这些"
] |
{
"arxiv_id": "2403.11752",
"language": "zh",
"page_idx": 0
}
|
||
2403.11752-zh-1
|
2403.11752-zh-1.jpg
|
作品中可能存在的刻板印象，并在将其教给学生之前对其内容进行批判性评估。这项研究试图通过使用各种机器学习技术来填补这一空白，以减少人为干预来纠正这种陈规定型观念。这项研究的亮点有三个：
1. Dataset : The identification of explicit gen-
der stereotypes in rhymes and poems necessitates the availability of datasets. To the best of our knowledge, there are no publicly accessible repositories specifically dedicated to documenting poems and rhymes that contain explicit gender stereotypes. We created a publicly available dataset $^{1}$ of rhymes and poems collected from various sources, with each line manually annotated for gender stereotypes. This contribution will significantly advance research in this area, leading to a better understanding of the harms perpetuated by these rhymes and poems.
2. Classification : Different Al models were
trained and tested to classify whether a poem or a rhyme contains an explicit gender stereotype. Additionally, we propose a heuristic encoder, that utilizes annotator-learned features and sentiment analysis. The best model achieved an accuracy of 97 % while maintaining the recall at 0.81, signifying a better overall performance given that the dataset is highly imbalanced.
3. Rectification: We test the efficacy of Large
Language Models (LLMs) in rectifying gender stereotypes in rhymes and poems by conducting a survey that compares human and LLM rectifications.
## 2. Related Work
发展研究强调了理解和控制儿童早期 (Bigler and Liben, 2007) 中的陈规定型观念和偏见的重要性。Kane (1996) 的跨文化研究表明，两三岁的孩子可以标记性别并相应地对对象进行分类，而到了四岁，他们就符合社会性别规范。Fast et al.(2016) 发现，在在线写作社区中通常会观察到男性过度代表和传统的性别刻板印象
使用监督任务的早期工作在性别刻板印象分析中显示出可喜的结果，利用大量未标记的数据来减少性别分类 (Bergsma et al.,2009) 中的错误。Park et al. (2018) 开发了在滥用语言数据集上训练的语言模型中测量性别偏见的方法，分析了预先训练的单词嵌人和模型体系结构的影响
已建立 Word 嵌人以包含重要的性别刻板印象， 这些刻板印象与现实世界的偏见 (Bolukbasi et al., 2016; Du et al., 2019) 相关。Gonen et al. (2019) 紧随其后，纠正了单词嵌入中的语法性别偏见。使
$^{1}$https://github.com/s-vigneshwaran/
Revisiting-The-Classics
用单词嵌入量化语言语料库中的性别刻板印象，揭示了它们的一致和强大的存在，包括理论上选择的刻板印象(Charlesworth et al.,2021)。但是，用于消除嵌人的传统技术可以通过提供用于传达性别信息 (Prost et al., 2019) 的噪声较小的通道来恶化下游分类器的偏差。此外，实验表明，当神经模型利用迁移学习时，它们的性别偏见总体上有所增加，特别是当使用已经偏向(Rekabsaz and Schedl, 2020) 的预训练嵌人时。Lu et al. (2020) 利用反事实数据增强来减轻性别偏见，优于单词嵌入消除偏见方法
除了性别偏见外，与性别认同有关的在线毒性， 如厌女症和性别歧视，也是研究的主要焦点。许多研究从社交媒体平台收集了有关性别歧视和厌女症的数据，这些数据通常包含语言复杂性，例如 leet-speech 和代码转换。机器学习技术已成功用于对这些数据进行分类 (Rodríguez-Sánchez et al., 2020; Frenda et al. ,2019; Chiril et al., 2020)
尽管这些作品强调了理解基于性别的威胁，性别陈规定型观念及其在社交媒体平台等各种与语言相关的领域中的存在的重要性，但仍有必要解决文学作品中的性别偏见，例如押韵和诗歌，它们是文化规范和社会角色的有影响力的来源。这项研究为该领域的未来工作奠定了基础，以确定儿童押韵和诗歌中是否存在偏见，以及确定押韵和诗歌中存在的性别陈规定型观念的策略。我们还提出了有效的策略，通过利用大型语言模型纠正符合现代价值观的押韵和诗歌，减轻性别偏见的押韵和诗歌对儿童的负面影响
## 3. Dataset
数据收集过程涉及从各种来源获取儿童押韵和适合青少年的诗歌。这些资源涵盖了广泛的内容，包括莎士比亚和弗罗斯特等著名诗人的作品，以及诸如 MotherGoose (Grover and Richardson, 1915)
创建儿童押韵和诗歌综合数据集的选择过程旨在确保风格，内容和文化背景方面的多样性。具体来说，为了确保我们的数据集多样化和可访问性， 在与文学和教育领域的教育工作者进行广泛协商后，我们从各种出版来源收集了押韵和诗歌。此过程确保数据集代表了广泛的文化背景和观点，并且押韵和诗歌没有语言错误和复杂性。除了最初用英语写的押韵&诗歌，we used 20 publicly available translated poems from 11 different languages (Refer to Table 7 in Appendix) 。目标是为分析和研究创造丰富的资源，代表广泛的押韵和诗歌。本研究中使用的数据集包含明显的班级不平衡。这可以归因于文化资源的不完整翻译和英语资源的数字化。此外，并非所有已出版的英语诗歌和押韵都已在线提供
注释采用了两阶段注释策略。初始阶段涉及利用数据集的子集建立注释准则。第二阶段采用已建立的准则来标记剩余的数据集。一对分别为 22 岁和 23岁的注释者分别注释了50 首押韵和诗歌。两位注释者都是非英语母语人士，并且已经接受了17 年的英语培训，强调了他们理解语言的复杂性和微妙
|
[
[
0.1234130859375,
0.07781982421875,
0.486083984375,
0.07781982421875,
0.486083984375,
0.0885009765625,
0.1234130859375,
0.0885009765625
],
[
0.51318359375,
0.0748291015625,
0.8818359375,
0.07598876953125,
0.8818359375,
0.09027099609375,
0.51318359375,
0.08905029296875
],
[
0.12176513671875,
0.0914306640625,
0.48779296875,
0.09027099609375,
0.48779296875,
0.1015625,
0.12176513671875,
0.10272216796875
],
[
0.515625,
0.0926513671875,
0.880859375,
0.0926513671875,
0.880859375,
0.10272216796875,
0.515625,
0.10272216796875
],
[
0.12176513671875,
0.105712890625,
0.48779296875,
0.105712890625,
0.48779296875,
0.11639404296875,
0.12176513671875,
0.11639404296875
],
[
0.515625,
0.105712890625,
0.880859375,
0.105712890625,
0.880859375,
0.11639404296875,
0.515625,
0.11639404296875
],
[
0.12176513671875,
0.1181640625,
0.48779296875,
0.1181640625,
0.48779296875,
0.1318359375,
0.12176513671875,
0.1318359375
],
[
0.5146484375,
0.1181640625,
0.8818359375,
0.1181640625,
0.8818359375,
0.1318359375,
0.5146484375,
0.1318359375
],
[
0.12176513671875,
0.1341552734375,
0.15869140625,
0.1341552734375,
0.15869140625,
0.1455078125,
0.12176513671875,
0.1455078125
],
[
0.515625,
0.133056640625,
0.8818359375,
0.133056640625,
0.8818359375,
0.146728515625,
0.515625,
0.146728515625
],
[
0.1318359375,
0.15380859375,
0.490234375,
0.15380859375,
0.490234375,
0.16748046875,
0.1318359375,
0.16748046875
],
[
0.51708984375,
0.1484375,
0.880859375,
0.1484375,
0.880859375,
0.1591796875,
0.51708984375,
0.1591796875
],
[
0.51318359375,
0.1602783203125,
0.8818359375,
0.160888671875,
0.8818359375,
0.174560546875,
0.51318359375,
0.17333984375
],
[
0.151123046875,
0.16748046875,
0.48779296875,
0.168701171875,
0.48779296875,
0.1817626953125,
0.151123046875,
0.18115234375
],
[
0.5146484375,
0.17578125,
0.8818359375,
0.17578125,
0.8818359375,
0.189453125,
0.5146484375,
0.189453125
],
[
0.154541015625,
0.18408203125,
0.48779296875,
0.18408203125,
0.48779296875,
0.19482421875,
0.154541015625,
0.19482421875
],
[
0.5146484375,
0.189453125,
0.88330078125,
0.189453125,
0.88330078125,
0.203125,
0.5146484375,
0.203125
],
[
0.15283203125,
0.196533203125,
0.48779296875,
0.196533203125,
0.48779296875,
0.210205078125,
0.15283203125,
0.210205078125
],
[
0.51708984375,
0.2060546875,
0.8818359375,
0.2060546875,
0.8818359375,
0.2161865234375,
0.51708984375,
0.2161865234375
],
[
0.154541015625,
0.2120361328125,
0.486083984375,
0.2120361328125,
0.486083984375,
0.22265625,
0.154541015625,
0.22265625
],
[
0.515625,
0.2197265625,
0.61279296875,
0.2197265625,
0.61279296875,
0.23095703125,
0.515625,
0.23095703125
],
[
0.1561279296875,
0.226806640625,
0.486083984375,
0.226806640625,
0.486083984375,
0.2369384765625,
0.1561279296875,
0.2369384765625
],
[
0.53076171875,
0.2327880859375,
0.8818359375,
0.2327880859375,
0.8818359375,
0.2464599609375,
0.53076171875,
0.2464599609375
],
[
0.15283203125,
0.2392578125,
0.48779296875,
0.2392578125,
0.48779296875,
0.2529296875,
0.15283203125,
0.2529296875
],
[
0.151123046875,
0.251708984375,
0.489501953125,
0.2529296875,
0.489501953125,
0.267333984375,
0.151123046875,
0.2666015625
],
[
0.515625,
0.2464599609375,
0.88330078125,
0.2464599609375,
0.88330078125,
0.260009765625,
0.515625,
0.260009765625
],
[
0.515625,
0.26123046875,
0.88330078125,
0.26123046875,
0.88330078125,
0.27490234375,
0.515625,
0.27490234375
],
[
0.1561279296875,
0.26953125,
0.486083984375,
0.26953125,
0.486083984375,
0.27978515625,
0.1561279296875,
0.27978515625
],
[
0.515625,
0.275634765625,
0.8818359375,
0.275634765625,
0.8818359375,
0.289306640625,
0.515625,
0.289306640625
],
[
0.154541015625,
0.283935546875,
0.48779296875,
0.283935546875,
0.48779296875,
0.29443359375,
0.154541015625,
0.29443359375
],
[
0.51318359375,
0.2880859375,
0.88330078125,
0.290283203125,
0.88330078125,
0.303955078125,
0.51318359375,
0.3017578125
],
[
0.154541015625,
0.29638671875,
0.48779296875,
0.29638671875,
0.48779296875,
0.31005859375,
0.154541015625,
0.31005859375
],
[
0.515625,
0.303955078125,
0.88330078125,
0.303955078125,
0.88330078125,
0.317626953125,
0.515625,
0.317626953125
],
[
0.15283203125,
0.31005859375,
0.48779296875,
0.31005859375,
0.48779296875,
0.32373046875,
0.15283203125,
0.32373046875
],
[
0.5146484375,
0.317626953125,
0.8818359375,
0.317626953125,
0.8818359375,
0.330810546875,
0.5146484375,
0.330810546875
],
[
0.15283203125,
0.32470703125,
0.48779296875,
0.32470703125,
0.48779296875,
0.33837890625,
0.15283203125,
0.33837890625
],
[
0.5322265625,
0.330810546875,
0.880859375,
0.33203125,
0.880859375,
0.345703125,
0.5322265625,
0.344482421875
],
[
0.15283203125,
0.33837890625,
0.300537109375,
0.33837890625,
0.300537109375,
0.3515625,
0.15283203125,
0.3515625
],
[
0.5146484375,
0.3466796875,
0.8818359375,
0.3466796875,
0.8818359375,
0.35986328125,
0.5146484375,
0.35986328125
],
[
0.1502685546875,
0.35986328125,
0.48779296875,
0.35986328125,
0.48779296875,
0.37353515625,
0.1502685546875,
0.37353515625
],
[
0.5146484375,
0.361083984375,
0.87890625,
0.361083984375,
0.87890625,
0.371826171875,
0.5146484375,
0.371826171875
],
[
0.15283203125,
0.374755859375,
0.489501953125,
0.3759765625,
0.489501953125,
0.3896484375,
0.15283203125,
0.388427734375
],
[
0.130126953125,
0.363525390625,
0.1561279296875,
0.363525390625,
0.1561279296875,
0.371826171875,
0.130126953125,
0.371826171875
],
[
0.5146484375,
0.377197265625,
0.880859375,
0.377197265625,
0.880859375,
0.38720703125,
0.5146484375,
0.38720703125
],
[
0.15283203125,
0.38720703125,
0.48779296875,
0.38720703125,
0.48779296875,
0.402099609375,
0.15283203125,
0.402099609375
],
[
0.515625,
0.391357421875,
0.880859375,
0.391357421875,
0.880859375,
0.402099609375,
0.515625,
0.402099609375
],
[
0.15283203125,
0.402587890625,
0.489501953125,
0.402587890625,
0.489501953125,
0.416259765625,
0.15283203125,
0.416259765625
],
[
0.5146484375,
0.40380859375,
0.8818359375,
0.40380859375,
0.8818359375,
0.41748046875,
0.5146484375,
0.41748046875
],
[
0.154541015625,
0.41748046875,
0.489501953125,
0.41748046875,
0.489501953125,
0.43115234375,
0.154541015625,
0.43115234375
],
[
0.5146484375,
0.419189453125,
0.87890625,
0.419189453125,
0.87890625,
0.429931640625,
0.5146484375,
0.429931640625
],
[
0.154541015625,
0.432861328125,
0.48779296875,
0.432861328125,
0.48779296875,
0.443603515625,
0.154541015625,
0.443603515625
],
[
0.515625,
0.43408203125,
0.880859375,
0.43408203125,
0.880859375,
0.44482421875,
0.515625,
0.44482421875
],
[
0.154541015625,
0.44775390625,
0.489501953125,
0.44775390625,
0.489501953125,
0.457763671875,
0.154541015625,
0.457763671875
],
[
0.5146484375,
0.446533203125,
0.8818359375,
0.446533203125,
0.8818359375,
0.460205078125,
0.5146484375,
0.460205078125
],
[
0.15283203125,
0.460205078125,
0.48779296875,
0.460205078125,
0.48779296875,
0.473876953125,
0.15283203125,
0.473876953125
],
[
0.5146484375,
0.4619140625,
0.880859375,
0.4619140625,
0.880859375,
0.47265625,
0.5146484375,
0.47265625
],
[
0.15283203125,
0.474365234375,
0.486083984375,
0.474365234375,
0.486083984375,
0.488037109375,
0.15283203125,
0.488037109375
],
[
0.515625,
0.476806640625,
0.60107421875,
0.476806640625,
0.60107421875,
0.48681640625,
0.515625,
0.48681640625
],
[
0.154541015625,
0.4892578125,
0.2451171875,
0.4892578125,
0.2451171875,
0.50048828125,
0.154541015625,
0.50048828125
],
[
0.1275634765625,
0.5087890625,
0.48779296875,
0.51025390625,
0.48779296875,
0.52490234375,
0.1275634765625,
0.52392578125
],
[
0.6337890625,
0.50830078125,
0.755859375,
0.51025390625,
0.755859375,
0.52490234375,
0.6337890625,
0.5224609375
],
[
0.15283203125,
0.52490234375,
0.48779296875,
0.52490234375,
0.48779296875,
0.5390625,
0.15283203125,
0.5390625
],
[
0.15283203125,
0.5390625,
0.489501953125,
0.5390625,
0.489501953125,
0.552734375,
0.15283203125,
0.552734375
],
[
0.5146484375,
0.5390625,
0.8818359375,
0.5390625,
0.8818359375,
0.552734375,
0.5146484375,
0.552734375
],
[
0.151123046875,
0.55419921875,
0.489501953125,
0.5517578125,
0.489501953125,
0.5654296875,
0.151123046875,
0.56787109375
],
[
0.5146484375,
0.55419921875,
0.8818359375,
0.55419921875,
0.8818359375,
0.56787109375,
0.5146484375,
0.56787109375
],
[
0.154541015625,
0.568359375,
0.2493896484375,
0.568359375,
0.2493896484375,
0.580078125,
0.154541015625,
0.580078125
],
[
0.5146484375,
0.568359375,
0.880859375,
0.568359375,
0.880859375,
0.5791015625,
0.5146484375,
0.5791015625
],
[
0.515625,
0.5830078125,
0.869140625,
0.5830078125,
0.869140625,
0.59326171875,
0.515625,
0.59326171875
],
[
0.214111328125,
0.599609375,
0.387939453125,
0.599609375,
0.387939453125,
0.61328125,
0.214111328125,
0.61328125
],
[
0.5341796875,
0.5986328125,
0.880859375,
0.5986328125,
0.880859375,
0.609375,
0.5341796875,
0.609375
],
[
0.5146484375,
0.61083984375,
0.8818359375,
0.61083984375,
0.8818359375,
0.62451171875,
0.5146484375,
0.62451171875
],
[
0.12005615234375,
0.62890625,
0.48779296875,
0.62744140625,
0.48779296875,
0.6396484375,
0.12005615234375,
0.64013671875
],
[
0.51708984375,
0.626953125,
0.880859375,
0.626953125,
0.880859375,
0.63720703125,
0.51708984375,
0.63720703125
],
[
0.118408203125,
0.64111328125,
0.48779296875,
0.64111328125,
0.48779296875,
0.65478515625,
0.118408203125,
0.65478515625
],
[
0.515625,
0.64111328125,
0.880859375,
0.64111328125,
0.880859375,
0.65185546875,
0.515625,
0.65185546875
],
[
0.118408203125,
0.65478515625,
0.486083984375,
0.65478515625,
0.486083984375,
0.66845703125,
0.118408203125,
0.66845703125
],
[
0.515625,
0.65380859375,
0.8818359375,
0.65380859375,
0.8818359375,
0.66748046875,
0.515625,
0.66748046875
],
[
0.12176513671875,
0.6728515625,
0.48779296875,
0.6728515625,
0.48779296875,
0.68310546875,
0.12176513671875,
0.68310546875
],
[
0.5146484375,
0.66943359375,
0.880859375,
0.66943359375,
0.880859375,
0.6796875,
0.5146484375,
0.6796875
],
[
0.12005615234375,
0.68505859375,
0.48779296875,
0.68505859375,
0.48779296875,
0.69873046875,
0.12005615234375,
0.69873046875
],
[
0.5146484375,
0.68408203125,
0.8818359375,
0.68408203125,
0.8818359375,
0.69482421875,
0.5146484375,
0.69482421875
],
[
0.12005615234375,
0.69970703125,
0.489501953125,
0.69970703125,
0.489501953125,
0.71337890625,
0.12005615234375,
0.71337890625
],
[
0.5146484375,
0.69677734375,
0.88330078125,
0.69677734375,
0.88330078125,
0.71044921875,
0.5146484375,
0.71044921875
],
[
0.12005615234375,
0.71337890625,
0.3046875,
0.71337890625,
0.3046875,
0.72705078125,
0.12005615234375,
0.72705078125
],
[
0.5146484375,
0.71142578125,
0.88330078125,
0.71142578125,
0.88330078125,
0.724609375,
0.5146484375,
0.724609375
],
[
0.1376953125,
0.728515625,
0.48779296875,
0.728515625,
0.48779296875,
0.7392578125,
0.1376953125,
0.7392578125
],
[
0.51318359375,
0.724609375,
0.8818359375,
0.724609375,
0.8818359375,
0.73828125,
0.51318359375,
0.73828125
],
[
0.12005615234375,
0.74072265625,
0.489501953125,
0.7412109375,
0.489501953125,
0.7548828125,
0.12005615234375,
0.75439453125
],
[
0.51318359375,
0.73681640625,
0.8818359375,
0.73681640625,
0.8818359375,
0.75048828125,
0.51318359375,
0.75048828125
],
[
0.12005615234375,
0.755859375,
0.48779296875,
0.755859375,
0.48779296875,
0.76953125,
0.12005615234375,
0.76953125
],
[
0.5146484375,
0.751953125,
0.88330078125,
0.7529296875,
0.88330078125,
0.7666015625,
0.5146484375,
0.765625
],
[
0.12176513671875,
0.77099609375,
0.486083984375,
0.77099609375,
0.486083984375,
0.78076171875,
0.12176513671875,
0.78076171875
],
[
0.515625,
0.76708984375,
0.8818359375,
0.76708984375,
0.8818359375,
0.78076171875,
0.515625,
0.78076171875
],
[
0.12005615234375,
0.783203125,
0.486083984375,
0.783203125,
0.486083984375,
0.79638671875,
0.12005615234375,
0.79638671875
],
[
0.51318359375,
0.77978515625,
0.8818359375,
0.77978515625,
0.8818359375,
0.79345703125,
0.51318359375,
0.79345703125
],
[
0.12005615234375,
0.798828125,
0.353515625,
0.7998046875,
0.353515625,
0.81103515625,
0.12005615234375,
0.81005859375
],
[
0.5146484375,
0.79638671875,
0.55322265625,
0.79638671875,
0.55322265625,
0.80859375,
0.5146484375,
0.80859375
],
[
0.1368408203125,
0.8134765625,
0.484375,
0.8134765625,
0.484375,
0.82666015625,
0.1368408203125,
0.82666015625
],
[
0.12176513671875,
0.82666015625,
0.48779296875,
0.82666015625,
0.48779296875,
0.84033203125,
0.12176513671875,
0.84033203125
],
[
0.515625,
0.82470703125,
0.8818359375,
0.82470703125,
0.8818359375,
0.83837890625,
0.515625,
0.83837890625
],
[
0.12005615234375,
0.84130859375,
0.48779296875,
0.84130859375,
0.48779296875,
0.85498046875,
0.12005615234375,
0.85498046875
],
[
0.5146484375,
0.8388671875,
0.8818359375,
0.8388671875,
0.8818359375,
0.8525390625,
0.5146484375,
0.8525390625
],
[
0.12005615234375,
0.85546875,
0.489501953125,
0.85546875,
0.489501953125,
0.869140625,
0.12005615234375,
0.869140625
],
[
0.515625,
0.85498046875,
0.8818359375,
0.85498046875,
0.8818359375,
0.865234375,
0.515625,
0.865234375
],
[
0.5146484375,
0.8662109375,
0.8818359375,
0.86767578125,
0.8818359375,
0.88134765625,
0.5146484375,
0.8798828125
],
[
0.14111328125,
0.8828125,
0.453369140625,
0.8828125,
0.453369140625,
0.896484375,
0.14111328125,
0.896484375
],
[
0.5146484375,
0.8818359375,
0.8818359375,
0.8818359375,
0.8818359375,
0.8955078125,
0.5146484375,
0.8955078125
],
[
0.12005615234375,
0.896484375,
0.328369140625,
0.89794921875,
0.328369140625,
0.9091796875,
0.12005615234375,
0.90771484375
],
[
0.5146484375,
0.896484375,
0.8818359375,
0.896484375,
0.8818359375,
0.91015625,
0.5146484375,
0.91015625
]
] |
[
"作品中可能存在的刻板印象，并在将其教给学生之",
"用单词嵌入量化语言语料库中的性别刻板印象，揭",
"前对其内容进行批判性评估。这项研究试图通过使",
"示了它们的一致和强大的存在，包括理论上选择的",
"用各种机器学习技术来填补这一空白，以减少人为",
"刻板印象（Charlesworthetal.，2021）。但是，用",
"干预来纠正这种陈规定型观念。这项研究的亮点有",
"于消除嵌人的传统技术可以通过提供用于传达性",
"三个",
"别信息（Prostetal.2019)的噪声较小的通道来恶",
"1. Dataset : The identification of explicit gen-",
"化下游分类器的偏差。此外，实验表明，当神经",
"模型利用迁移学习时，它们的性别偏见总体上有",
"derstereotypes inrhymes and poems neces",
"所增加，特别是当使用已经偏向（Rekabsazand",
"sitatestheavailabilityofdatasets.Tothebest",
"Schedl,2020）的预训练嵌人时。Luetal.(2020）利",
"of our knowledge, there are no publicly acces",
"用反事实数据增强来减轻性别偏见，优于单词嵌人",
"siblerepositoriesspecificallydedicatedtodoc",
"消除偏见方法",
"umentingpoemsandrhymesthatcontainex",
"除了性别偏见外，与性别认同有关的在线毒性，",
"plicit gender stereotypes. We created a pub-",
"licly available dataset 1 of rhymes and poems",
"如厌女症和性别歧视，也是研究的主要焦点。许",
"多研究从社交媒体平台收集了有关性别歧视和厌",
"collectedfromvarioussources,witheachline",
"女症的数据，这些数据通常包含语言复杂性，例",
"manuallyannotatedforgenderstereotypes.",
"如leet-speech和代码转换。机器学习技术已成",
"This contribution will significantly advance re-",
"功用于对这些数据进行分类（Rodriguez-Sanchez",
"search in this area, leading to a better under",
"et al., 2020; Frenda et al., 2019; Chiril et al., 2020)",
"standing of the harms perpetuated by these",
"尽管这些作品强调了理解基于性别的威胁，性别",
"rhymesand poems.",
"陈规定型观念及其在社交媒体平台等各种与语言",
"Classification : Different Al models were",
"相关的领域中的存在的重要性，但仍有必要解决文",
"trained and tested to classify whether a poem",
"2.",
"学作品中的性别偏见，例如押韵和诗歌，它们是文",
"or a rhyme contains an explicit gender stereo-",
"化规范和社会角色的有影响力的来源。这项研究为",
"type. Additionally, we propose a heuristic",
"该领域的未来工作奠定了基础，以确定儿童押韵和",
"encoder, that utilizes annotator-learned fea-",
"诗歌中是否存在偏见，以及确定押韵和诗歌中存在",
"turesandsentimentanalysis.Thebestmodel",
"的性别陈规定型观念的策略。我们还提出了有效的",
"achievedanaccuracyof97%whilemaintain-",
"策略，通过利用大型语言模型纠正符合现代价值观",
"ing the recall at 0.81, signifying a better overall",
"的押韵和诗歌，减轻性别偏见的押韵和诗歌对儿童",
"performance given that the dataset is highly",
"的负面影响",
"imbalanced.",
"3. Rectification : We test the efficacy of Large",
"3.Dataset",
"Language Models (LLMs) in rectifying gender",
"stereotypes in rhymes and poems by conduct",
"数据收集过程涉及从各种来源获取儿童押韵和适",
"ing a survey that compares human and LLM",
"合青少年的诗歌。这些资源涵盖了广泛的内容，包",
"rectifications",
"括莎士比亚和弗罗斯特等著名诗人的作品，以及诸",
"如MotherGoose(GroverandRichardson.1915",
"2.Related Work",
"创建儿童押韵和诗歌综合数据集的选择过程旨",
"在确保风格，内容和文化背景方面的多样性。具体",
"发展研究强调了理解和控制儿童早期（Biglerand",
"来说，为了确保我们的数据集多样化和可访问性",
"Liben,2007）中的陈规定型观念和偏见的重要性。",
"在与文学和教育领域的教育工作者进行广泛协商",
"Kane（1996)的跨文化研究表明，两三岁的孩子可",
"后，我们从各种出版来源收集了押韵和诗歌。此过",
"以标记性别并相应地对对象进行分类，而到了四",
"程确保数据集代表了广泛的文化背景和观点，并且",
"岁，他们就符合社会性别规范。Fastetal.(2016)",
"押韵和诗歌没有语言错误和复杂性。除了最初用英",
"发现，在在线写作社区中通常会观察到男性过度代",
"语写的押韵&诗歌，weused20publiclyavailable",
"表和传统的性别刻板印象",
"translatedpoems from 11differentlanguages (Re-",
"使用监督任务的早期工作在性别刻板印象分析",
"fertoTable7inAppendix）。目标是为分析和研究",
"中显示出可喜的结果，利用大量未标记的数据来减",
"创造丰富的资源，代表广泛的押韵和诗歌。本研究",
"少性别分类（Bergsmaetal.，2009）中的错误。Park",
"中使用的数据集包含明显的班级不平衡。这可以归",
"etal.（2018）开发了在滥用语言数据集上训练的语",
"因于文化资源的不完整翻译和英语资源的数字化。",
"言模型中测量性别偏见的方法，分析了预先训练的",
"此外，并非所有已出版的英语诗歌和押韵都已在线",
"单词嵌人和模型体系结构的影响",
"提供",
"已建立Word嵌入以包含重要的性别刻板印象",
"这些刻板印象与现实世界的偏见（Bolukbasietal.，",
"注释采用了两阶段注释策略。初始阶段涉及利用数",
"2016; Du et al., 2019) 相关。 Gonen et al. (2019)",
"据集的子集建立注释准则。第二阶段采用已建立的",
"紧随其后，纠正了单词嵌入中的语法性别偏见。使",
"准则来标记剩余的数据集。一对分别为22岁和23",
"岁的注释者分别注释了50首押韵和诗歌。两位注",
"https://github.com/s-vigneshwaran/",
"释者都是非英语母语人士，并且已经接受了17年",
"Revisiting-The-classics",
"的英语培训，强调了他们理解语言的复杂性和微妙"
] |
{
"arxiv_id": "2403.11752",
"language": "zh",
"page_idx": 1
}
|
||
2403.11752-zh-2
|
2403.11752-zh-2.jpg
|
<table>
<tbody>
<tr>
<th>Label</th>
<th>$\#$Verses</th>
<th>$\#$Lines</th>
</tr>
<tr>
<td>Stereotypical Rhymes</td>
<td>65</td>
<td>151</td>
</tr>
<tr>
<td>Non-Stereotypical Rhymes</td>
<td>274</td>
<td>5,157</td>
</tr>
<tr>
<td>Stereotypical Poems</td>
<td>80</td>
<td>359</td>
</tr>
<tr>
<td>Non-Stereotypical $|Poems$</td>
<td>242</td>
<td>7,647</td>
</tr>
<tr>
<td>Augmented I Stereotypes</td>
<td>290</td>
<td>1,347</td>
</tr>
<tr>
<td>Total (Without Augmentation)</td>
<td>661</td>
<td>13,314</td>
</tr>
<tr>
<td>Total (With Augmentation)</td>
<td>951</td>
<td>14,661</td>
</tr>
</tbody>
</table>
Table 1: Dataset Statistic
中的名词或主题/对象 ”]。有力地保持诗韵计划和句子结构的完整”。我们专门针对名词和同义词设计了这个提示，以便在不影响偏见的情况下增加文本。如表1 所示，这种增加导致定型押韵和诗歌的数量增加了一倍，使定型诗歌和押韵的总数增加到290。。表 2 包含原始版本和增强版本的示例
<table>
<tbody>
<tr>
<th>Initial Sentence</th>
<th>Augmented Version</th>
</tr>
</tbody>
</table>
Table 2: Examples of augmented text
## 之处的资格在第一阶段，注释者进行了注释过程，他们不知
道所注释的诗歌或押韵的身份。这样做是为了防止任何无意识的偏见影响他们的注释。换句话说，注释者对诗歌或押韵的身份视而不见，这意味着他们不知道哪些诗歌或押韵包含明确的性别刻板印象。这样做是为了确保注释者的注释尽可能客观。无论注释者的背景如何，两个注释者都足以完成标记押韵和诗歌的任务，因为该任务在语言上相对简单。无论地理位置如何，押韵和诗歌都会教给小学生， 因此来自不同背景的注释者可能对它们有相似的了解
第一阶段继续进行多次迭代，直到它们达到令人满意的注释者间可靠性分数。在迭代之间，注释者开会讨论和裁定任何分歧。分歧主要围绕单词的选择和给定行的解释。例如，在诸如“我的主人一个” 和“你愿意成为我的吗？”之类的行中，特别注意所有权的概念。应当指出的是，“主人”和“我的”这两个术语在所有权方面具有不同的含义，因为“我的”可以暗示为异性的占有，而“主人”并不意味着特定性别的所有权。像漂亮这样与唯美主义有关的词，当它们的用法与特定的性别联系在一起时，也会有分歧，但由于美的主观性质，它们被认为是非刻板印象。为了测量注释过程的编码器间可靠性， 使用了 Krippendorff 的$\alpha$ (De Swert,2012)。经过四次迭代的不一致分析，我们得到了 Krippendorff 的$\alpha$为0。96
第二阶段涉及通过审议和评估在迭代之间进行
的讨论来建立准则
$\cdot$ Lines containing words or phrases that could
be interpreted as toxic or hateful towards a particular gender are classified as containing stereotypes.
·Stereotypes are identified in lines containing
sexist words or phrases toward women.
· Lines are considered stereotypes if they con-
tain gender-specific behaviour stereotypes
· Sentences implying ownership over the other
gender are classified as stereotypical.
·Stereotypical sentences include those sug-
gesting that certain behaviour is exclusive to
one gender.
$\cdot$ Sentences with terms that were historically
honorific but now have negative meanings are
classified as stereotypical.
在指南的帮助下，两个注释者之一标记了其余数据。表1显示了在线条和整体方面定型的诗歌和押韵的数量
文本增强由于数据集中存在类别不平衡，如表1所示，我们使用增强来增加训练集中非定型诗歌的数量。通过使用 GPT-3 在训练集中合成诗歌和押韵的同义词版本来执行数据增强。5 (OpenAl,2022)
。使用了以下提示：“用同义词替换「$^*$诗歌或押韵
# 4. Methodology
本节阐明了通过使用多个模型来检测性别陈规定型观念的方法。此外，它还深人研究了所提出的启发式编码器的功能和机制。
# 4.1. Heuristic Encoder
通常，分类模型仅使用基本事实来优化预测。然而，研究表明，知识注人模型在特定任务 (Logan et al.,2019) 中表现更好。受这一发现的启发，我们提出了启发式编码器的概念。启发式编码器背后的核心思想涉及注释器学习功能的使用。在启发式编码器中，我们不使用外部知识库，而是使用注释器学习的特征来补充输入特征，以增强模型的预测能力。为了实现这一目标，要求注释者从注释的诗歌和押韵列表中汇编他们认为是刻板印象的单词， 短语和特定于性别的名称的完整列表，使功能列表仅限于数据集，以便更好地理解上下文。此列表充当文本标签以外的附加源。此外，我们还将情感分析作为编码过程的一部分进行，为此我们采用了现成的 TextBlob (Loria et al.,2018) 。这使得能够识别和表征文本中的情感模式和倾向，有助于更深入地理解所传达的情感方面
使用在编码器设计过程中收集的注释器学习特征作为注释过程的一部分，为给定文本生成二进制值特征向量。每个载体的结构定义如下：[男性姓名，女性姓名，刻板印象，阴性，阳性]。此向量旨在包含与特定性别名称，已识别的刻板印象和情感极性相关的相关信息。例如，“有一个妻子，但不
|
[
[
0.52294921875,
0.0748291015625,
0.56689453125,
0.0748291015625,
0.56689453125,
0.08612060546875,
0.52294921875,
0.08612060546875
],
[
0.744140625,
0.07598876953125,
0.8095703125,
0.07598876953125,
0.8095703125,
0.08612060546875,
0.744140625,
0.08612060546875
],
[
0.8193359375,
0.07183837890625,
0.87744140625,
0.0736083984375,
0.8759765625,
0.0885009765625,
0.81787109375,
0.08612060546875
],
[
0.5263671875,
0.0872802734375,
0.67822265625,
0.0872802734375,
0.67822265625,
0.1009521484375,
0.5263671875,
0.1009521484375
],
[
0.12005615234375,
0.07659912109375,
0.2032470703125,
0.07659912109375,
0.2032470703125,
0.0885009765625,
0.12005615234375,
0.0885009765625
],
[
0.7841796875,
0.0885009765625,
0.80615234375,
0.0885009765625,
0.80615234375,
0.1009521484375,
0.7841796875,
0.1009521484375
],
[
0.84814453125,
0.0885009765625,
0.875,
0.0885009765625,
0.875,
0.1009521484375,
0.84814453125,
0.1009521484375
],
[
0.8310546875,
0.09857177734375,
0.875,
0.1009521484375,
0.873046875,
0.11517333984375,
0.82958984375,
0.1134033203125
],
[
0.1368408203125,
0.09027099609375,
0.489501953125,
0.09027099609375,
0.489501953125,
0.10394287109375,
0.1368408203125,
0.10394287109375
],
[
0.5263671875,
0.1009521484375,
0.7119140625,
0.1009521484375,
0.7119140625,
0.114013671875,
0.5263671875,
0.114013671875
],
[
0.77734375,
0.1009521484375,
0.8076171875,
0.1009521484375,
0.8076171875,
0.1134033203125,
0.77734375,
0.1134033203125
],
[
0.5263671875,
0.1134033203125,
0.6708984375,
0.1134033203125,
0.6708984375,
0.127685546875,
0.5263671875,
0.127685546875
],
[
0.118408203125,
0.10394287109375,
0.489501953125,
0.1051025390625,
0.489501953125,
0.1181640625,
0.118408203125,
0.1175537109375
],
[
0.7841796875,
0.1134033203125,
0.8076171875,
0.1134033203125,
0.8076171875,
0.1270751953125,
0.7841796875,
0.1270751953125
],
[
0.8447265625,
0.114013671875,
0.875,
0.114013671875,
0.875,
0.1270751953125,
0.8447265625,
0.1270751953125
],
[
0.8310546875,
0.1258544921875,
0.8759765625,
0.1258544921875,
0.8759765625,
0.14013671875,
0.8310546875,
0.14013671875
],
[
0.12005615234375,
0.1181640625,
0.489501953125,
0.1181640625,
0.489501953125,
0.1318359375,
0.12005615234375,
0.1318359375
],
[
0.5263671875,
0.1270751953125,
0.70361328125,
0.1270751953125,
0.70361328125,
0.14013671875,
0.5263671875,
0.14013671875
],
[
0.77734375,
0.1270751953125,
0.80615234375,
0.1270751953125,
0.80615234375,
0.14013671875,
0.77734375,
0.14013671875
],
[
0.52490234375,
0.1383056640625,
0.6943359375,
0.1395263671875,
0.6943359375,
0.15380859375,
0.52490234375,
0.152587890625
],
[
0.118408203125,
0.1318359375,
0.489501953125,
0.133056640625,
0.489501953125,
0.146728515625,
0.118408203125,
0.1455078125
],
[
0.77734375,
0.1383056640625,
0.8076171875,
0.141357421875,
0.80419921875,
0.152587890625,
0.77490234375,
0.149658203125
],
[
0.8310546875,
0.1383056640625,
0.8759765625,
0.1383056640625,
0.8759765625,
0.152587890625,
0.8310546875,
0.152587890625
],
[
0.118408203125,
0.1478271484375,
0.486083984375,
0.1478271484375,
0.486083984375,
0.157958984375,
0.118408203125,
0.157958984375
],
[
0.5263671875,
0.15380859375,
0.72705078125,
0.15380859375,
0.72705078125,
0.16748046875,
0.5263671875,
0.16748046875
],
[
0.77734375,
0.15380859375,
0.80615234375,
0.15380859375,
0.80615234375,
0.166259765625,
0.77734375,
0.166259765625
],
[
0.82275390625,
0.152587890625,
0.87744140625,
0.152587890625,
0.87744140625,
0.16748046875,
0.82275390625,
0.16748046875
],
[
0.12176513671875,
0.163330078125,
0.486083984375,
0.163330078125,
0.486083984375,
0.17333984375,
0.12176513671875,
0.17333984375
],
[
0.5283203125,
0.1650390625,
0.7060546875,
0.166259765625,
0.7060546875,
0.179931640625,
0.5283203125,
0.1787109375
],
[
0.77734375,
0.166259765625,
0.80615234375,
0.166259765625,
0.80615234375,
0.1787109375,
0.77734375,
0.1787109375
],
[
0.82275390625,
0.1644287109375,
0.8759765625,
0.166259765625,
0.875,
0.18115234375,
0.8212890625,
0.1787109375
],
[
0.12005615234375,
0.17578125,
0.486083984375,
0.17578125,
0.486083984375,
0.189453125,
0.12005615234375,
0.189453125
],
[
0.12005615234375,
0.190673828125,
0.489501953125,
0.190673828125,
0.489501953125,
0.2037353515625,
0.12005615234375,
0.2037353515625
],
[
0.60546875,
0.1923828125,
0.79248046875,
0.1923828125,
0.79248046875,
0.2060546875,
0.60546875,
0.2060546875
],
[
0.12005615234375,
0.2060546875,
0.48193359375,
0.2060546875,
0.48193359375,
0.2161865234375,
0.12005615234375,
0.2161865234375
],
[
0.12005615234375,
0.218505859375,
0.48779296875,
0.218505859375,
0.48779296875,
0.232177734375,
0.12005615234375,
0.232177734375
],
[
0.12176513671875,
0.2327880859375,
0.154541015625,
0.2327880859375,
0.154541015625,
0.2452392578125,
0.12176513671875,
0.2452392578125
],
[
0.51708984375,
0.2327880859375,
0.88330078125,
0.2327880859375,
0.88330078125,
0.2464599609375,
0.51708984375,
0.2464599609375
],
[
0.1368408203125,
0.2464599609375,
0.48779296875,
0.2464599609375,
0.48779296875,
0.260009765625,
0.1368408203125,
0.260009765625
],
[
0.515625,
0.248779296875,
0.880859375,
0.248779296875,
0.880859375,
0.2587890625,
0.515625,
0.2587890625
],
[
0.12176513671875,
0.262451171875,
0.486083984375,
0.262451171875,
0.486083984375,
0.2724609375,
0.12176513671875,
0.2724609375
],
[
0.5146484375,
0.26123046875,
0.8818359375,
0.26123046875,
0.8818359375,
0.27490234375,
0.5146484375,
0.27490234375
],
[
0.12005615234375,
0.27490234375,
0.489501953125,
0.27490234375,
0.489501953125,
0.2880859375,
0.12005615234375,
0.2880859375
],
[
0.5146484375,
0.275634765625,
0.8818359375,
0.275634765625,
0.8818359375,
0.289306640625,
0.5146484375,
0.289306640625
],
[
0.118408203125,
0.289306640625,
0.489501953125,
0.289306640625,
0.489501953125,
0.302734375,
0.118408203125,
0.302734375
],
[
0.5146484375,
0.289306640625,
0.8818359375,
0.289306640625,
0.8818359375,
0.302734375,
0.5146484375,
0.302734375
],
[
0.12005615234375,
0.303955078125,
0.484375,
0.303955078125,
0.484375,
0.314208984375,
0.12005615234375,
0.314208984375
],
[
0.51123046875,
0.302734375,
0.8388671875,
0.303955078125,
0.8388671875,
0.317626953125,
0.51123046875,
0.31640625
],
[
0.12005615234375,
0.317626953125,
0.48779296875,
0.317626953125,
0.48779296875,
0.330810546875,
0.12005615234375,
0.330810546875
],
[
0.51123046875,
0.330810546875,
0.64208984375,
0.330810546875,
0.64208984375,
0.344482421875,
0.51123046875,
0.344482421875
],
[
0.7001953125,
0.330810546875,
0.85205078125,
0.330810546875,
0.85205078125,
0.344482421875,
0.7001953125,
0.344482421875
],
[
0.6962890625,
0.344482421875,
0.86376953125,
0.345703125,
0.86376953125,
0.35986328125,
0.6962890625,
0.359375
],
[
0.12005615234375,
0.33203125,
0.489501953125,
0.33203125,
0.489501953125,
0.345703125,
0.12005615234375,
0.345703125
],
[
0.12005615234375,
0.3466796875,
0.48779296875,
0.3466796875,
0.48779296875,
0.35986328125,
0.12005615234375,
0.35986328125
],
[
0.5205078125,
0.345703125,
0.68701171875,
0.345703125,
0.68701171875,
0.359375,
0.5205078125,
0.359375
],
[
0.6826171875,
0.349853515625,
0.7021484375,
0.349853515625,
0.7021484375,
0.35400390625,
0.6826171875,
0.35400390625
],
[
0.12005615234375,
0.35986328125,
0.48779296875,
0.35986328125,
0.48779296875,
0.37353515625,
0.12005615234375,
0.37353515625
],
[
0.52294921875,
0.35986328125,
0.69189453125,
0.35986328125,
0.69189453125,
0.37353515625,
0.52294921875,
0.37353515625
],
[
0.69873046875,
0.359375,
0.8623046875,
0.359375,
0.8623046875,
0.372314453125,
0.69873046875,
0.372314453125
],
[
0.52490234375,
0.3759765625,
0.58837890625,
0.3759765625,
0.58837890625,
0.38720703125,
0.52490234375,
0.38720703125
],
[
0.7021484375,
0.374755859375,
0.8623046875,
0.374755859375,
0.8623046875,
0.388427734375,
0.7021484375,
0.388427734375
],
[
0.52294921875,
0.388427734375,
0.68701171875,
0.388427734375,
0.68701171875,
0.402099609375,
0.52294921875,
0.402099609375
],
[
0.12176513671875,
0.377197265625,
0.48779296875,
0.377197265625,
0.48779296875,
0.38720703125,
0.12176513671875,
0.38720703125
],
[
0.12176513671875,
0.39013671875,
0.48779296875,
0.39013671875,
0.48779296875,
0.40087890625,
0.12176513671875,
0.40087890625
],
[
0.6826171875,
0.392578125,
0.68701171875,
0.392578125,
0.68701171875,
0.39794921875,
0.6826171875,
0.39794921875
],
[
0.6884765625,
0.392578125,
0.6962890625,
0.392578125,
0.6962890625,
0.396728515625,
0.6884765625,
0.396728515625
],
[
0.7001953125,
0.3896484375,
0.86376953125,
0.3896484375,
0.86376953125,
0.402587890625,
0.7001953125,
0.402587890625
],
[
0.12176513671875,
0.40380859375,
0.48779296875,
0.40380859375,
0.48779296875,
0.41455078125,
0.12176513671875,
0.41455078125
],
[
0.52294921875,
0.40380859375,
0.646484375,
0.40380859375,
0.646484375,
0.41748046875,
0.52294921875,
0.41748046875
],
[
0.7021484375,
0.405029296875,
0.8095703125,
0.405029296875,
0.8095703125,
0.416259765625,
0.7021484375,
0.416259765625
],
[
0.12176513671875,
0.418701171875,
0.486083984375,
0.418701171875,
0.486083984375,
0.4287109375,
0.12176513671875,
0.4287109375
],
[
0.118408203125,
0.43115234375,
0.48779296875,
0.43115234375,
0.48779296875,
0.44482421875,
0.118408203125,
0.44482421875
],
[
0.55908203125,
0.427490234375,
0.8388671875,
0.4287109375,
0.8388671875,
0.443603515625,
0.55908203125,
0.4423828125
],
[
0.12176513671875,
0.446533203125,
0.22412109375,
0.446533203125,
0.22412109375,
0.457763671875,
0.12176513671875,
0.457763671875
],
[
0.1368408203125,
0.46142578125,
0.484375,
0.46142578125,
0.484375,
0.471435546875,
0.1368408203125,
0.471435546875
],
[
0.12005615234375,
0.473876953125,
0.255126953125,
0.473876953125,
0.255126953125,
0.48681640625,
0.12005615234375,
0.48681640625
],
[
0.60791015625,
0.479736328125,
0.779296875,
0.482666015625,
0.779296875,
0.49755859375,
0.60791015625,
0.49462890625
],
[
0.1368408203125,
0.49755859375,
0.489501953125,
0.49755859375,
0.489501953125,
0.51123046875,
0.1368408203125,
0.51123046875
],
[
0.154541015625,
0.51416015625,
0.48779296875,
0.51416015625,
0.48779296875,
0.52490234375,
0.154541015625,
0.52490234375
],
[
0.515625,
0.51318359375,
0.87890625,
0.51318359375,
0.87890625,
0.52392578125,
0.515625,
0.52392578125
],
[
0.15283203125,
0.52685546875,
0.48779296875,
0.52685546875,
0.48779296875,
0.54052734375,
0.15283203125,
0.54052734375
],
[
0.515625,
0.52783203125,
0.880859375,
0.52783203125,
0.880859375,
0.5380859375,
0.515625,
0.5380859375
],
[
0.154541015625,
0.5419921875,
0.2435302734375,
0.5419921875,
0.2435302734375,
0.55419921875,
0.154541015625,
0.55419921875
],
[
0.515625,
0.54052734375,
0.7080078125,
0.54052734375,
0.7080078125,
0.55419921875,
0.515625,
0.55419921875
],
[
0.139404296875,
0.5634765625,
0.156982421875,
0.56787109375,
0.154541015625,
0.576171875,
0.1368408203125,
0.57177734375
],
[
0.1502685546875,
0.56396484375,
0.48779296875,
0.56396484375,
0.48779296875,
0.5791015625,
0.1502685546875,
0.5791015625
],
[
0.5146484375,
0.57373046875,
0.72900390625,
0.57373046875,
0.72900390625,
0.58740234375,
0.5146484375,
0.58740234375
],
[
0.15283203125,
0.58056640625,
0.44580078125,
0.58056640625,
0.44580078125,
0.59130859375,
0.15283203125,
0.59130859375
],
[
0.1351318359375,
0.60107421875,
0.490234375,
0.6015625,
0.490234375,
0.615234375,
0.1351318359375,
0.6142578125
],
[
0.5146484375,
0.59716796875,
0.88330078125,
0.59716796875,
0.88330078125,
0.61083984375,
0.5146484375,
0.61083984375
],
[
0.15283203125,
0.6162109375,
0.470947265625,
0.6162109375,
0.470947265625,
0.6298828125,
0.15283203125,
0.6298828125
],
[
0.515625,
0.6123046875,
0.880859375,
0.6123046875,
0.880859375,
0.62255859375,
0.515625,
0.62255859375
],
[
0.5146484375,
0.62451171875,
0.880859375,
0.6259765625,
0.880859375,
0.6396484375,
0.5146484375,
0.63818359375
],
[
0.1351318359375,
0.63818359375,
0.489501953125,
0.6396484375,
0.489501953125,
0.65380859375,
0.1351318359375,
0.65283203125
],
[
0.5146484375,
0.6396484375,
0.8818359375,
0.6396484375,
0.8818359375,
0.65283203125,
0.5146484375,
0.65283203125
],
[
0.151123046875,
0.65380859375,
0.43408203125,
0.65380859375,
0.43408203125,
0.66748046875,
0.151123046875,
0.66748046875
],
[
0.5146484375,
0.65478515625,
0.88330078125,
0.65478515625,
0.88330078125,
0.66845703125,
0.5146484375,
0.66845703125
],
[
0.5146484375,
0.66845703125,
0.88330078125,
0.66845703125,
0.88330078125,
0.681640625,
0.5146484375,
0.681640625
],
[
0.1351318359375,
0.67578125,
0.489501953125,
0.6767578125,
0.489501953125,
0.69140625,
0.1351318359375,
0.6904296875
],
[
0.5146484375,
0.68310546875,
0.87890625,
0.68310546875,
0.87890625,
0.693359375,
0.5146484375,
0.693359375
],
[
0.151123046875,
0.6923828125,
0.489501953125,
0.6904296875,
0.489501953125,
0.70361328125,
0.151123046875,
0.7060546875
],
[
0.51318359375,
0.6953125,
0.88330078125,
0.69677734375,
0.88330078125,
0.71044921875,
0.51318359375,
0.708984375
],
[
0.15283203125,
0.7080078125,
0.2418212890625,
0.7080078125,
0.2418212890625,
0.71875,
0.15283203125,
0.71875
],
[
0.51318359375,
0.71044921875,
0.8818359375,
0.71142578125,
0.8818359375,
0.724609375,
0.51318359375,
0.7236328125
],
[
0.14111328125,
0.72998046875,
0.156982421875,
0.73095703125,
0.1561279296875,
0.7392578125,
0.139404296875,
0.73828125
],
[
0.1502685546875,
0.72998046875,
0.48779296875,
0.72998046875,
0.48779296875,
0.74365234375,
0.1502685546875,
0.74365234375
],
[
0.5146484375,
0.7255859375,
0.880859375,
0.7255859375,
0.880859375,
0.736328125,
0.5146484375,
0.736328125
],
[
0.151123046875,
0.7421875,
0.48779296875,
0.74365234375,
0.48779296875,
0.75732421875,
0.151123046875,
0.755859375
],
[
0.5146484375,
0.7392578125,
0.88330078125,
0.7392578125,
0.88330078125,
0.7529296875,
0.5146484375,
0.7529296875
],
[
0.5146484375,
0.751953125,
0.8818359375,
0.75439453125,
0.8818359375,
0.76708984375,
0.5146484375,
0.765625
],
[
0.154541015625,
0.7587890625,
0.345947265625,
0.7587890625,
0.345947265625,
0.76953125,
0.154541015625,
0.76953125
],
[
0.51318359375,
0.7666015625,
0.8818359375,
0.7666015625,
0.8818359375,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.1351318359375,
0.783203125,
0.48779296875,
0.783203125,
0.48779296875,
0.79638671875,
0.1351318359375,
0.79638671875
],
[
0.51318359375,
0.78076171875,
0.8818359375,
0.78076171875,
0.8818359375,
0.79443359375,
0.51318359375,
0.79443359375
],
[
0.118408203125,
0.79638671875,
0.48779296875,
0.79638671875,
0.48779296875,
0.81005859375,
0.118408203125,
0.81005859375
],
[
0.5146484375,
0.79638671875,
0.88330078125,
0.79638671875,
0.88330078125,
0.81005859375,
0.5146484375,
0.81005859375
],
[
0.12005615234375,
0.8125,
0.187255859375,
0.8125,
0.187255859375,
0.82373046875,
0.12005615234375,
0.82373046875
],
[
0.5146484375,
0.81005859375,
0.69873046875,
0.81103515625,
0.69873046875,
0.822265625,
0.5146484375,
0.82177734375
],
[
0.53076171875,
0.82470703125,
0.880859375,
0.82470703125,
0.880859375,
0.83837890625,
0.53076171875,
0.83837890625
],
[
0.12176513671875,
0.84033203125,
0.484375,
0.84033203125,
0.484375,
0.85107421875,
0.12176513671875,
0.85107421875
],
[
0.5146484375,
0.8388671875,
0.8818359375,
0.8388671875,
0.8818359375,
0.8525390625,
0.5146484375,
0.8525390625
],
[
0.12176513671875,
0.85400390625,
0.489501953125,
0.85400390625,
0.489501953125,
0.86767578125,
0.12176513671875,
0.86767578125
],
[
0.5146484375,
0.85498046875,
0.87890625,
0.85498046875,
0.87890625,
0.865234375,
0.5146484375,
0.865234375
],
[
0.12005615234375,
0.86767578125,
0.48779296875,
0.86767578125,
0.48779296875,
0.88134765625,
0.12005615234375,
0.88134765625
],
[
0.5146484375,
0.869140625,
0.87890625,
0.869140625,
0.87890625,
0.8798828125,
0.5146484375,
0.8798828125
],
[
0.12005615234375,
0.8818359375,
0.48779296875,
0.8818359375,
0.48779296875,
0.8955078125,
0.12005615234375,
0.8955078125
],
[
0.5146484375,
0.8818359375,
0.8818359375,
0.8818359375,
0.8818359375,
0.8955078125,
0.5146484375,
0.8955078125
],
[
0.118408203125,
0.896484375,
0.48779296875,
0.896484375,
0.48779296875,
0.91015625,
0.118408203125,
0.91015625
],
[
0.51318359375,
0.896484375,
0.88330078125,
0.896484375,
0.88330078125,
0.91015625,
0.51318359375,
0.91015625
]
] |
[
"Label",
"# Verses",
"# Lines",
"Stereotypical Rhymes",
"之处的资格",
"65",
"151",
"5,157",
"在第一阶段，注释者进行了注释过程，他们不知",
"Non-Stereotypical Rhymes",
"274",
"Stereotypical Poems",
"道所注释的诗歌或押韵的身份。这样做是为了防止",
"80",
"359",
"7,647",
"任何无意识的偏见影响他们的注释。换句话说，注",
"Non-Stereotypical Poems",
"242",
"Augmented Stereotypes",
"释者对诗歌或押韵的身份视而不见，这意味着他们",
"290",
"1,347",
"不知道哪些诗歌或押韵包含明确的性别刻板印象。",
"Total (Without Augmentation)",
"661",
"13,314",
"这样做是为了确保注释者的注释尽可能客观。无论",
"Total (With Augmentation)",
"951",
"14,661",
"注释者的背景如何，两个注释者都足以完成标记押",
"韵和诗歌的任务，因为该任务在语言上相对简单。",
"Table 1: Dataset Statistic",
"无论地理位置如何，押韵和诗歌都会教给小学生",
"因此来自不同背景的注释者可能对它们有相似的",
"了解",
"中的名词或主题/对象。有力地保持诗韵计划和",
"第一阶段继续进行多次送代，直到它们达到令人",
"句子结构的完整”。我们专门针对名词和同义词设",
"满意的注释者间可靠性分数。在迭代之间，注释者",
"计了这个提示，以便在不影响偏见的情况下增加文",
"开会讨论和裁定任何分歧。分歧主要围绕单词的选",
"本。如表1所示，这种增加导致定型押韵和诗歌的",
"择和给定行的解释。例如，在诸如“我的主人一个”",
"数量增加了一倍，使定型诗歌和押韵的总数增加到",
"和“你愿意成为我的吗？”之类的行中，特别注意所",
"290。。表2包含原始版本和增强版本的示例",
"有权的概念。应当指出的是，“主人”和“我的”这",
"Initial Sentence",
"Augmented Version",
"Jack and Jill wentup",
"两个术语在所有权方面具有不同的含义，因为“我",
"的”可以暗示为异性的占有，而“主人”并不意味着",
"Jack and Jill went up",
"",
"特定性别的所有权。像漂亮这样与唯美主义有关的",
"the hill, To fetch a pail ",
"the mountain, To ob-",
"of water.",
"tain a bucket of water.",
"And when I bake, I'lI",
"词，当它们的用法与特定的性别联系在一起时，也",
"会有分歧，但由于美的主观性质，它们被认为是非",
"",
"",
"And when I fry, I'll",
"刻板印象。为了测量注释过程的编码器间可靠性，",
"give you a cake,",
"give you apie,",
"使用了Krippendorff的aα（DeSwert，2012）。经过",
"四次送代的不一致分析，我们得到了Krippendorff",
"Table 2: Examples of augmented text",
"的α为0。96",
"第二阶段涉及通过审议和评估在代之间进行",
"的讨论来建立准则",
"4.Methodology",
": Lines containing words or phrases that could",
"beinterpreted astoxic orhatefultowardsa",
"本节阐明了通过使用多个模型来检测性别陈规定",
"particular gender are classified as containing",
"型观念的方法。此外，它还深入研究了所提出的启",
"stereotypes.",
"发式编码器的功能和机制。",
".",
"Stereotypes are identified in lines containing",
"4.1.Heuristic Encoder",
"sexistwordsorphrasestowardwomen.",
":Lines are considered stereotypes if they con-",
"通常，分类模型仅使用基本事实来优化预测。然",
"tain gender-specific behaviour stereotypes.",
"而，研究表明：知识注入模型在特定任务（Logan",
"etal.,2019）中表现更好。受这一发现的启发，我",
": Sentences implying ownership over the other",
"们提出了启发式编码器的概念。启发式编码器背后",
"gender are classified as stereotypical.",
"的核心思想涉及注释器学习功能的使用。在启发式",
"编码器中，我们不使用外部知识库，而是使用注释",
": Stereotypical sentences include those sug",
"器学习的特征来补充输入特征，以增强模型的预测",
"gestingthat certain behaviour is exclusiveto",
"能力。为了实现这一目标，要求注释者从注释的诗",
"one gender.",
"歌和押韵列表中汇编他们认为是刻板印象的单词",
".:",
"Sentences with terms that were historically",
"短语和特定于性别的名称的完整列表，使功能列表",
"honorific but now have negative meanings are",
"仅限于数据集，以便更好地理解上下文。此列表充",
"当文本标签以外的附加源。此外，我们还将情感分",
"classifiedasstereotypical.",
"析作为编码过程的一部分进行，为此我们采用了现",
"在指南的帮助下，两个注释者之一标记了其余数",
"成的TextBlob（Loria etal.,2018）。这使得能够识",
"据。表1显示了在线条和整体方面定型的诗歌和押",
"别和表征文本中的情感模式和倾向，有助于更深人",
"韵的数量",
"地理解所传达的情感方面",
"使用在编码器设计过程中收集的注释器学习特",
"文本增强由于数据集中存在类别不平衡，如表1所",
"征作为注释过程的一部分，为给定文本生成二进制",
"示，我们使用增强来增加训练集中非定型诗歌的数",
"值特征向量。每个载体的结构定义如下：男性姓",
"量。通过使用GPT-3在训练集中合成诗歌和押韵",
"名，女性姓名，刻板印象，阴性，阳性。此向量旨",
"的同义词版本来执行数据增强。5(OpenAl,2022)",
"在包含与特定性别名称，已识别的刻板印象和情感",
"。使用了以下提示：“用同义词替换［*诗歌或押韵",
"极性相关的相关信息。例如，“有一个妻子，但不"
] |
{
"arxiv_id": "2403.11752",
"language": "zh",
"page_idx": 2
}
|
||
2012.12509-zh-3
|
2012.12509-zh-3.jpg
|
<FigureHere>
Figure 2: Illustration of Deep Semantic Dictionary Learning. The semantic dictionary learning stage (training stage) aims at training CNN for feature extraction and auto-encoder for semantic dictionary generation. In the multi-label image classification stage (testing stage), the label occurrence probabilities of given image can be estimated by encoding its visual features with the learned semantic dictionary.
其中$\theta$表示 CNN 和$d=2048$的可学习参数
Semantic Dictionary Generation Module 语义字典的质量是我们模型成功的关键。我们打算从类别的单词嵌人中得出视觉空间中的语义字典，因为语义信息已被证明在多标签图像分类中具有判别特性$\operatorname{E}($Xing et al. 2019; Murthy, Maji, and Manmatha 2015; Chen et al. 2019b,a; Lin et al. 2015)。详细地说，我们采用自动编码器以非线性方式完成字典生成。许多 NLP 研究表明自动编码器可以很好地利用语义空间，并通过双向转换来缓解模型的讨度拟合
在我们的模型中，利用在 Wikipedia 数据集上训练的GloVe$\underline{\mathrm{E}}($Pennington, Socher, and Manning 2014) 来提取每个类的 $k$ 维单词嵌入，
Overview
我们提出了一种新颖的多标签图像分类方法，称为深度语义字典学习 (DSDL),该方法从字典学习的角度为多标签图像分类提供了一种创新的解决方案。DSDL 是一个端到端可训练的深度学习模型，它结合了类别的语义信息，并提供了一种天真而优雅的方式来利用和协调标签，语义和视觉空间。所提出的模型将多标签图像分类视为特征字典表示问题，其中类别的单词嵌入用于导出语义字典，而标签嵌人被视为表示给定样本的视觉特征的系数。DSDL 由三个模块组成：特征学习、语义字典生成和视觉特征表示模块。图 2 说明了 DSDL 的详细流程，包括语义字典学习和多标签图像分类阶段。由于传统的字典学习无法直接求解表示系数和语义字典，因此我们设计了一种新颖的 DSDL 替代参数更新策略 (APUS),以实现表示系数和语义字典的联合学习。APUS 可以通过在字典学习中建立连接，在网络的前向和后向传播过程中交替优化表示系数和语义字典
(4)
$$s_i=f_g(w_i)\in\mathcal{R}^k$$
其中$w_i$是$i$的语义单词，$s_i$是相应的单词嵌人。然后编码器$f_en$将单词嵌人$S=[s_1,s_2,\ldots,s_c]\in\mathcal{R}^{k\times c}$投影义字典投影回语义空间以重建$S$ 。这样的过程可以在数学上表示如下，
Deep Semantic Dictionary Learning
Feature Learning Module CNN 目前是主要的特征学习方法。在这里，我们遵循以前的作品(Zhu et al.2017; Chen et al. 2019b,a) , 它采用在 ImageNet 上预先训练的 ResNet-101 作为特征学习的骨干。给一个输入图像$x$ ,我们将其馈送到 ResNet-101 中，以从最终的卷积层获得 2048$\times14\times14$维特征图。然后，我们应用全局最大池化来获得 2048 维全局特征$f$,如下所示，
$$f=\mathcal{F}(x)=f_{pool}(f_{cnn}(x;\theta))\in\mathcal{R}^d,$$
(5) (6)
$$\begin{array}{rcl}D&=&f_{en}(S;\psi),\\\hat{S}&=&f_{de}(D;\psi^T),\end{array}$$
其中$f_{en}(\cdot)$和$f_de(\cdot)$共享相同的参数。$\psi$表示可学习的参数。为了确保学习的语义词典尽可能保持原始语义信
(3)
|
[
[
0.396240234375,
0.0706787109375,
0.61767578125,
0.0706787109375,
0.61767578125,
0.083984375,
0.396240234375,
0.083984375
],
[
0.301513671875,
0.08587646484375,
0.406982421875,
0.08709716796875,
0.406982421875,
0.0985107421875,
0.301513671875,
0.09783935546875
],
[
0.6611328125,
0.0889892578125,
0.75634765625,
0.0889892578125,
0.75634765625,
0.0997314453125,
0.6611328125,
0.0997314453125
],
[
0.66015625,
0.10101318359375,
0.75244140625,
0.10101318359375,
0.75244140625,
0.11236572265625,
0.66015625,
0.11236572265625
],
[
0.240234375,
0.12371826171875,
0.29736328125,
0.12371826171875,
0.29736328125,
0.1351318359375,
0.240234375,
0.1351318359375
],
[
0.4697265625,
0.11871337890625,
0.54150390625,
0.11871337890625,
0.54150390625,
0.1300048828125,
0.4697265625,
0.1300048828125
],
[
0.35302734375,
0.126953125,
0.398681640625,
0.126953125,
0.398681640625,
0.1383056640625,
0.35302734375,
0.1383056640625
],
[
0.48046875,
0.1287841796875,
0.53515625,
0.1307373046875,
0.53369140625,
0.1458740234375,
0.4794921875,
0.143310546875
],
[
0.69384765625,
0.1307373046875,
0.71142578125,
0.1307373046875,
0.71142578125,
0.14013671875,
0.69384765625,
0.14013671875
],
[
0.359375,
0.1395263671875,
0.393798828125,
0.1395263671875,
0.393798828125,
0.1514892578125,
0.359375,
0.1514892578125
],
[
0.1502685546875,
0.1483154296875,
0.1568603515625,
0.1483154296875,
0.1568603515625,
0.1514892578125,
0.1502685546875,
0.1514892578125
],
[
0.1568603515625,
0.1483154296875,
0.17724609375,
0.1483154296875,
0.17724609375,
0.15283203125,
0.1568603515625,
0.15283203125
],
[
0.80224609375,
0.149658203125,
0.81787109375,
0.149658203125,
0.81884765625,
0.2392578125,
0.8037109375,
0.2392578125
],
[
0.42236328125,
0.1572265625,
0.42724609375,
0.1572265625,
0.42724609375,
0.1622314453125,
0.42236328125,
0.1622314453125
],
[
0.88232421875,
0.1558837890625,
0.89892578125,
0.1558837890625,
0.89892578125,
0.1641845703125,
0.88232421875,
0.1641845703125
],
[
0.289306640625,
0.175537109375,
0.34716796875,
0.177978515625,
0.34716796875,
0.1849365234375,
0.289306640625,
0.18310546875
],
[
0.371826171875,
0.177978515625,
0.40185546875,
0.177978515625,
0.40185546875,
0.1849365234375,
0.371826171875,
0.1849365234375
],
[
0.75732421875,
0.1749267578125,
0.77392578125,
0.1749267578125,
0.77392578125,
0.184326171875,
0.75732421875,
0.184326171875
],
[
0.705078125,
0.1900634765625,
0.71142578125,
0.1900634765625,
0.71142578125,
0.1932373046875,
0.705078125,
0.1932373046875
],
[
0.1405029296875,
0.1976318359375,
0.1568603515625,
0.1976318359375,
0.1568603515625,
0.2020263671875,
0.1405029296875,
0.2020263671875
],
[
0.223876953125,
0.199462890625,
0.361083984375,
0.208984375,
0.35546875,
0.260009765625,
0.2181396484375,
0.250732421875
],
[
0.8447265625,
0.20263671875,
0.8505859375,
0.20263671875,
0.8505859375,
0.212158203125,
0.8447265625,
0.212158203125
],
[
0.1405029296875,
0.210205078125,
0.1568603515625,
0.210205078125,
0.1568603515625,
0.2183837890625,
0.1405029296875,
0.2183837890625
],
[
0.84326171875,
0.21533203125,
0.84716796875,
0.21533203125,
0.84716796875,
0.2203369140625,
0.84326171875,
0.2203369140625
],
[
0.1446533203125,
0.2247314453125,
0.1641845703125,
0.2247314453125,
0.1641845703125,
0.23291015625,
0.1446533203125,
0.23291015625
],
[
0.69287109375,
0.2215576171875,
0.71337890625,
0.2215576171875,
0.71337890625,
0.236083984375,
0.69287109375,
0.236083984375
],
[
0.1641845703125,
0.2291259765625,
0.2034912109375,
0.2291259765625,
0.2034912109375,
0.2392578125,
0.1641845703125,
0.2392578125
],
[
0.261474609375,
0.23291015625,
0.265625,
0.23291015625,
0.265625,
0.238037109375,
0.261474609375,
0.238037109375
],
[
0.5107421875,
0.231689453125,
0.52783203125,
0.231689453125,
0.52783203125,
0.240478515625,
0.5107421875,
0.240478515625
],
[
0.88232421875,
0.23291015625,
0.90283203125,
0.23291015625,
0.90283203125,
0.24365234375,
0.88232421875,
0.24365234375
],
[
0.19775390625,
0.2392578125,
0.2156982421875,
0.2392578125,
0.2156982421875,
0.246826171875,
0.19775390625,
0.246826171875
],
[
0.26953125,
0.246826171875,
0.294189453125,
0.246826171875,
0.294189453125,
0.25390625,
0.26953125,
0.25390625
],
[
0.30810546875,
0.246826171875,
0.314453125,
0.246826171875,
0.314453125,
0.25,
0.30810546875,
0.25
],
[
0.7587890625,
0.2442626953125,
0.76953125,
0.2442626953125,
0.76953125,
0.2474365234375,
0.7587890625,
0.2474365234375
],
[
0.136474609375,
0.25244140625,
0.160888671875,
0.25244140625,
0.160888671875,
0.261474609375,
0.136474609375,
0.261474609375
],
[
0.39453125,
0.25390625,
0.406982421875,
0.25390625,
0.406982421875,
0.25830078125,
0.39453125,
0.25830078125
],
[
0.69384765625,
0.2568359375,
0.71142578125,
0.2568359375,
0.71142578125,
0.272705078125,
0.69384765625,
0.272705078125
],
[
0.131591796875,
0.26318359375,
0.1666259765625,
0.26513671875,
0.1650390625,
0.27783203125,
0.1298828125,
0.27587890625
],
[
0.2181396484375,
0.279052734375,
0.31201171875,
0.27978515625,
0.31201171875,
0.291748046875,
0.2181396484375,
0.290283203125
],
[
0.483642578125,
0.279052734375,
0.60107421875,
0.281005859375,
0.60107421875,
0.292236328125,
0.483642578125,
0.290283203125
],
[
0.685546875,
0.279052734375,
0.72802734375,
0.279052734375,
0.72802734375,
0.2890625,
0.685546875,
0.2890625
],
[
0.36767578125,
0.28271484375,
0.396240234375,
0.28271484375,
0.396240234375,
0.287353515625,
0.36767578125,
0.287353515625
],
[
0.11767578125,
0.27978515625,
0.1544189453125,
0.27978515625,
0.1544189453125,
0.290283203125,
0.11767578125,
0.290283203125
],
[
0.406005859375,
0.2841796875,
0.435546875,
0.2841796875,
0.435546875,
0.287353515625,
0.406005859375,
0.287353515625
],
[
0.38232421875,
0.300537109375,
0.61669921875,
0.302978515625,
0.61669921875,
0.31640625,
0.38232421875,
0.314453125
],
[
0.64892578125,
0.31640625,
0.66845703125,
0.31640625,
0.66845703125,
0.416748046875,
0.64892578125,
0.416748046875
],
[
0.80810546875,
0.3193359375,
0.81396484375,
0.3193359375,
0.81396484375,
0.32763671875,
0.80810546875,
0.32763671875
],
[
0.71728515625,
0.32763671875,
0.7236328125,
0.32763671875,
0.7236328125,
0.33203125,
0.71728515625,
0.33203125
],
[
0.775390625,
0.328857421875,
0.7998046875,
0.328857421875,
0.7998046875,
0.34033203125,
0.775390625,
0.34033203125
],
[
0.71142578125,
0.34228515625,
0.7236328125,
0.33837890625,
0.73046875,
0.348388671875,
0.71826171875,
0.352783203125
],
[
0.71337890625,
0.35107421875,
0.771484375,
0.352783203125,
0.76953125,
0.367431640625,
0.71142578125,
0.365478515625
],
[
0.775390625,
0.347900390625,
0.7939453125,
0.347900390625,
0.7939453125,
0.354736328125,
0.775390625,
0.354736328125
],
[
0.4794921875,
0.3515625,
0.51123046875,
0.3515625,
0.51123046875,
0.359130859375,
0.4794921875,
0.359130859375
],
[
0.236083984375,
0.352783203125,
0.281982421875,
0.352783203125,
0.281982421875,
0.3642578125,
0.236083984375,
0.3642578125
],
[
0.416748046875,
0.3515625,
0.4599609375,
0.3515625,
0.4599609375,
0.359130859375,
0.416748046875,
0.359130859375
],
[
0.50244140625,
0.359130859375,
0.5185546875,
0.3623046875,
0.515625,
0.36865234375,
0.5,
0.365478515625
],
[
0.52294921875,
0.35986328125,
0.5361328125,
0.35986328125,
0.5361328125,
0.36865234375,
0.52294921875,
0.36865234375
],
[
0.5498046875,
0.35595703125,
0.5537109375,
0.354248046875,
0.55810546875,
0.359130859375,
0.55224609375,
0.361083984375
],
[
0.2303466796875,
0.3623046875,
0.289306640625,
0.365478515625,
0.28759765625,
0.380126953125,
0.2296142578125,
0.376953125
],
[
0.57421875,
0.35986328125,
0.57861328125,
0.35986328125,
0.57861328125,
0.367431640625,
0.57421875,
0.367431640625
],
[
0.465576171875,
0.367431640625,
0.483642578125,
0.37255859375,
0.4794921875,
0.380615234375,
0.461669921875,
0.375732421875
],
[
0.78369140625,
0.371826171875,
0.7958984375,
0.371826171875,
0.7958984375,
0.37744140625,
0.78369140625,
0.37744140625
],
[
0.1004638671875,
0.365478515625,
0.13232421875,
0.367431640625,
0.131591796875,
0.37890625,
0.098876953125,
0.376953125
],
[
0.71728515625,
0.37890625,
0.7236328125,
0.37890625,
0.7236328125,
0.38330078125,
0.71728515625,
0.38330078125
],
[
0.7158203125,
0.39013671875,
0.72802734375,
0.39013671875,
0.72802734375,
0.40087890625,
0.7158203125,
0.40087890625
],
[
0.77783203125,
0.386962890625,
0.81787109375,
0.392578125,
0.814453125,
0.40478515625,
0.775390625,
0.399658203125
],
[
0.10296630859375,
0.39453125,
0.111083984375,
0.39453125,
0.111083984375,
0.3984375,
0.10296630859375,
0.3984375
],
[
0.396240234375,
0.401611328125,
0.59326171875,
0.401611328125,
0.59326171875,
0.415283203125,
0.396240234375,
0.415283203125
],
[
0.0849609375,
0.431884765625,
0.9111328125,
0.43310546875,
0.9111328125,
0.447509765625,
0.0849609375,
0.4462890625
],
[
0.08660888671875,
0.4462890625,
0.91259765625,
0.4462890625,
0.91259765625,
0.4609375,
0.08660888671875,
0.4609375
],
[
0.0849609375,
0.458984375,
0.91259765625,
0.460205078125,
0.91259765625,
0.474853515625,
0.0849609375,
0.473388671875
],
[
0.0849609375,
0.474853515625,
0.55224609375,
0.474853515625,
0.55224609375,
0.488037109375,
0.0849609375,
0.488037109375
],
[
0.0931396484375,
0.51025390625,
0.1666259765625,
0.51025390625,
0.1666259765625,
0.521484375,
0.0931396484375,
0.521484375
],
[
0.521484375,
0.5087890625,
0.8408203125,
0.5087890625,
0.8408203125,
0.52294921875,
0.521484375,
0.52294921875
],
[
0.08905029296875,
0.5263671875,
0.476318359375,
0.5263671875,
0.476318359375,
0.537109375,
0.08905029296875,
0.537109375
],
[
0.525390625,
0.52978515625,
0.9111328125,
0.52978515625,
0.9111328125,
0.54345703125,
0.525390625,
0.54345703125
],
[
0.08905029296875,
0.5390625,
0.47802734375,
0.5390625,
0.47802734375,
0.55322265625,
0.08905029296875,
0.55322265625
],
[
0.52294921875,
0.54443359375,
0.9111328125,
0.54443359375,
0.9111328125,
0.55810546875,
0.52294921875,
0.55810546875
],
[
0.0882568359375,
0.55322265625,
0.47802734375,
0.55029296875,
0.47802734375,
0.56494140625,
0.0882568359375,
0.5673828125
],
[
0.521484375,
0.55810546875,
0.91259765625,
0.55810546875,
0.91259765625,
0.5712890625,
0.521484375,
0.5712890625
],
[
0.08660888671875,
0.56494140625,
0.47802734375,
0.56640625,
0.47802734375,
0.58056640625,
0.08660888671875,
0.580078125
],
[
0.521484375,
0.57275390625,
0.9111328125,
0.57275390625,
0.9111328125,
0.5859375,
0.521484375,
0.5859375
],
[
0.0882568359375,
0.58056640625,
0.47802734375,
0.58056640625,
0.47802734375,
0.5947265625,
0.0882568359375,
0.5947265625
],
[
0.521484375,
0.5859375,
0.9111328125,
0.5859375,
0.9111328125,
0.599609375,
0.521484375,
0.599609375
],
[
0.08905029296875,
0.59521484375,
0.476318359375,
0.59521484375,
0.476318359375,
0.609375,
0.08905029296875,
0.609375
],
[
0.521484375,
0.599609375,
0.91259765625,
0.599609375,
0.91259765625,
0.61279296875,
0.521484375,
0.61279296875
],
[
0.0882568359375,
0.609375,
0.476318359375,
0.609375,
0.476318359375,
0.62255859375,
0.0882568359375,
0.62255859375
],
[
0.51953125,
0.6123046875,
0.91015625,
0.61279296875,
0.91015625,
0.62744140625,
0.51953125,
0.626953125
],
[
0.0882568359375,
0.62255859375,
0.476318359375,
0.62255859375,
0.476318359375,
0.63623046875,
0.0882568359375,
0.63623046875
],
[
0.52392578125,
0.62744140625,
0.91259765625,
0.62744140625,
0.91259765625,
0.6416015625,
0.52392578125,
0.6416015625
],
[
0.0849609375,
0.63623046875,
0.476318359375,
0.63525390625,
0.476318359375,
0.6494140625,
0.0849609375,
0.65087890625
],
[
0.52294921875,
0.6416015625,
0.685546875,
0.6416015625,
0.685546875,
0.65478515625,
0.52294921875,
0.65478515625
],
[
0.08660888671875,
0.6494140625,
0.47802734375,
0.6484375,
0.47802734375,
0.66357421875,
0.08660888671875,
0.6640625
],
[
0.5361328125,
0.65478515625,
0.91259765625,
0.65478515625,
0.91259765625,
0.66845703125,
0.5361328125,
0.66845703125
],
[
0.0882568359375,
0.6640625,
0.4755859375,
0.6640625,
0.4755859375,
0.67822265625,
0.0882568359375,
0.67822265625
],
[
0.521484375,
0.66845703125,
0.91259765625,
0.66845703125,
0.91259765625,
0.681640625,
0.521484375,
0.681640625
],
[
0.0882568359375,
0.67822265625,
0.476318359375,
0.6767578125,
0.476318359375,
0.69140625,
0.0882568359375,
0.6923828125
],
[
0.521484375,
0.681640625,
0.71142578125,
0.681640625,
0.71142578125,
0.69580078125,
0.521484375,
0.69580078125
],
[
0.08905029296875,
0.69140625,
0.47802734375,
0.69140625,
0.47802734375,
0.705078125,
0.08905029296875,
0.705078125
],
[
0.0882568359375,
0.705078125,
0.476318359375,
0.705078125,
0.476318359375,
0.71826171875,
0.0882568359375,
0.71826171875
],
[
0.650390625,
0.70556640625,
0.775390625,
0.70068359375,
0.77685546875,
0.71826171875,
0.65185546875,
0.7236328125
],
[
0.8896484375,
0.70556640625,
0.9140625,
0.70556640625,
0.9140625,
0.7197265625,
0.8896484375,
0.7197265625
],
[
0.08660888671875,
0.71728515625,
0.476318359375,
0.7197265625,
0.476318359375,
0.73291015625,
0.08660888671875,
0.73095703125
],
[
0.51953125,
0.724609375,
0.9111328125,
0.72607421875,
0.9111328125,
0.74365234375,
0.51953125,
0.7421875
],
[
0.0882568359375,
0.73291015625,
0.463134765625,
0.73291015625,
0.463134765625,
0.7470703125,
0.0882568359375,
0.7470703125
],
[
0.521484375,
0.74072265625,
0.78125,
0.7412109375,
0.78125,
0.755859375,
0.521484375,
0.7548828125
],
[
0.76953125,
0.73779296875,
0.91259765625,
0.7373046875,
0.91259765625,
0.755859375,
0.76953125,
0.7568359375
],
[
0.09234619140625,
0.7568359375,
0.374267578125,
0.75830078125,
0.374267578125,
0.77294921875,
0.09234619140625,
0.771484375
],
[
0.52294921875,
0.7548828125,
0.85546875,
0.7548828125,
0.85546875,
0.76953125,
0.52294921875,
0.76953125
],
[
0.8544921875,
0.7568359375,
0.9111328125,
0.7568359375,
0.9111328125,
0.767578125,
0.8544921875,
0.767578125
],
[
0.51953125,
0.77099609375,
0.5556640625,
0.7646484375,
0.5595703125,
0.77587890625,
0.52392578125,
0.7822265625
],
[
0.56396484375,
0.76806640625,
0.830078125,
0.767578125,
0.830078125,
0.7822265625,
0.56396484375,
0.78271484375
],
[
0.09234619140625,
0.7744140625,
0.476318359375,
0.7744140625,
0.476318359375,
0.78857421875,
0.09234619140625,
0.78857421875
],
[
0.82861328125,
0.77099609375,
0.91015625,
0.77099609375,
0.91015625,
0.7822265625,
0.82861328125,
0.7822265625
],
[
0.525390625,
0.7822265625,
0.9140625,
0.7822265625,
0.9140625,
0.796875,
0.525390625,
0.796875
],
[
0.08660888671875,
0.78857421875,
0.476318359375,
0.787109375,
0.476318359375,
0.8017578125,
0.08660888671875,
0.80322265625
],
[
0.521484375,
0.796875,
0.6435546875,
0.796875,
0.6435546875,
0.81005859375,
0.521484375,
0.81005859375
],
[
0.08660888671875,
0.80322265625,
0.476318359375,
0.80322265625,
0.476318359375,
0.81640625,
0.08660888671875,
0.81640625
],
[
0.0882568359375,
0.81640625,
0.47802734375,
0.81640625,
0.47802734375,
0.830078125,
0.0882568359375,
0.830078125
],
[
0.6435546875,
0.81640625,
0.77392578125,
0.818359375,
0.77392578125,
0.8359375,
0.6435546875,
0.833984375
],
[
0.88818359375,
0.818359375,
0.9150390625,
0.818359375,
0.9150390625,
0.83544921875,
0.88818359375,
0.83544921875
],
[
0.08660888671875,
0.830078125,
0.47802734375,
0.8291015625,
0.47802734375,
0.84326171875,
0.08660888671875,
0.8447265625
],
[
0.64453125,
0.83544921875,
0.78759765625,
0.8369140625,
0.78759765625,
0.85498046875,
0.64453125,
0.85302734375
],
[
0.08905029296875,
0.8447265625,
0.4794921875,
0.8447265625,
0.4794921875,
0.85791015625,
0.08905029296875,
0.85791015625
],
[
0.8896484375,
0.8388671875,
0.91259765625,
0.8388671875,
0.91259765625,
0.85302734375,
0.8896484375,
0.85302734375
],
[
0.08660888671875,
0.85693359375,
0.4345703125,
0.85791015625,
0.4345703125,
0.87255859375,
0.08660888671875,
0.8720703125
],
[
0.52294921875,
0.861328125,
0.9111328125,
0.861328125,
0.9111328125,
0.875,
0.52294921875,
0.875
],
[
0.1568603515625,
0.875,
0.406005859375,
0.87353515625,
0.406005859375,
0.88818359375,
0.1568603515625,
0.8896484375
],
[
0.455810546875,
0.87548828125,
0.47802734375,
0.87548828125,
0.47802734375,
0.8896484375,
0.455810546875,
0.8896484375
],
[
0.521484375,
0.87548828125,
0.9111328125,
0.87548828125,
0.9111328125,
0.8896484375,
0.521484375,
0.8896484375
]
] |
[
"Semantic Dictionary Learning",
"Feature Learning",
"Visual Features",
"Representation",
"ResNet",
"Global Max",
"Feature",
"Pooling",
"~",
"Maps",
"",
"",
"LeastSquaresMetho",
"C",
"dic",
"Dictionar",
"neral",
"D",
"-",
"ooa",
"",
"..",
"dog",
"",
"car",
"X",
"iGloVe",
"",
"",
"sim",
"Ak",
"",
"",
"",
"able",
"00",
"+",
"Cbird",
"Word Embedding",
"Semantic Dictionary",
"",
"Auto",
"Words",
"",
"Multi-label ImageClassification",
"Least Squares Method",
"..",
"",
"bird",
"",
"Esigmoid",
"boa",
"::",
"Feature",
":：:",
"：",
"X",
"-",
"Learning",
"",
"...",
"car",
"福",
"",
"二",
"dog:",
"F",
"Visual Features Representation",
"Figure 2: Illustration of Deep Semantic Dictionary Learning. The semantic dictionary learning stage (training stage)",
"aims at training CNN for feature extraction and auto-encoder for semantic dictionary generation. In the multi-label",
"image classification stage (testing stage), the label occurrence probabilities of given image can be estimated by",
"encoding its visual features with the learned semantic dictionary.",
"Overview",
"其中θ表示CNN和d=2048的可学习参数",
"我们提出了一种新颖的多标签图像分类方法，称为深",
"SemanticDictionaryGenerationModule语义字典的",
"度语义字典学习（DSDL），该方法从字典学习的角度",
"质量是我们模型成功的关键。我们打算从类别的单词",
"为多标签图像分类提供了一种创新的解决方案。DSDL",
"嵌人中得出视觉空间中的语义字典，因为语义信息已被",
"是一个端到端可训练的深度学习模型，它结合了类别的",
"证明在多标签图像分类中具有判别特性E(Xingetal.",
"语义信息，并提供了一种天真而优雅的方式来利用和协",
"2019; Murthy, Maji, and Manmatha 2015; Chen et al.",
"调标签，语义和视觉空间。所提出的模型将多标签图像",
"2019b,a;Linetal.2015）。详细地说，我们采用自动编",
"分类视为特征字典表示问题，其中类别的单词嵌入用于",
"码器以非线性方式完成字典生成。许多NLP研究表明",
"导出语义字典，而标签嵌入被视为表示给定样本的视觉",
"自动编码器可以很好地利用语义空间，并通过双向转换",
"特征的系数。DSDL由三个模块组成：特征学习、语义",
"来缓解模型的过度拟合",
"字典生成和视觉特征表示模块。图2说明了DSDL的",
"在我们的模型中，利用在Wikipedia数据集上训练的",
"详细流程，包括语义字典学习和多标签图像分类阶段",
"GloVeE(Pennington, Socher,and Manning 2014)来提",
"由于传统的字典学习无法直接求解表示系数和语义字",
"取每个类的k维单词嵌入，",
"典，因此我们设计了一种新颖的DSDL替代参数更新",
"策略（APUS），以实现表示系数和语义字典的联合学",
"Si = fg(w)E Rh",
"(4)",
"习。APUS可以通过在字典学习中建立连接，在网络的",
"其中wi是讠的语义单词，si是相应的单词嵌人。然后",
"前向和后向传播过程中交替优化表示系数和语义字典",
"编码器fen将单词嵌人S=[s1,S2,..",
",sJERkx投影",
"Deep Semantic Dictionary Learning",
"到视觉空间中，以生成语义字典D=[a1,a2，",
"...aclE",
"Rdxe",
"，其中aiERd是第i个字典原子，",
"FeatureLearningModuleCNN目前是主要的特征学",
"解码器将语",
"义字典投影回语义空间以重建S。这样的过程可以在",
"习方法。在这里，我们遵循以前的作品（Zhuetal.2017",
"数学上表示如下，",
"Chenetal.2019b,a），它采用在ImageNet上预先训练",
"的ResNet-101作为特征学习的骨干。给一个输入图像",
"D = fen(S;3),",
"(5)",
"a，我们将其馈送到ResNet-101中，以从最终的卷积",
"$ = fde(D;T),",
"层获得2048×14×14维特征图。然后，我们应用全局",
"(6)",
"最大池化来获得2048维全局特征，如下所示，",
"其中fen(）和fde（）共享相同的参数。表示可学习的",
"f = F(r) = fpool(fenn(r; 0)) E Rd",
"(3)",
"参数。为了确保学习的语义词典尽可能保持原始语义信"
] |
{
"arxiv_id": "2012.12509",
"language": "zh",
"page_idx": 3
}
|
||
2110.02001-zh-1
|
2110.02001-zh-1.jpg
|
掌握显式意见-角色交互：统一意见角色标记的语法辅助神经转换系统
Shengqiong Wu,$^1$Hao Fei,$^1*$ Fei Li,$^1$ Donghong Ji,$^1$ Meishan Zhang,$^2$Yijiang Liu,$^1$Chong Teng$^1$
$^{1}$ Key Laboratory of Aerospace Information Security and Trusted Computing, Ministry
of Education, School of Cyber Science and Engineering, Wuhan University, China
$^{2}$Institute of Computing and Intelligence, Harbin Institute of Technology (Shenzhen), China $\{$ whuwsq, hao.fei, lifei\_csnlp, cslyj, tengchong, dhji $\}$ @whu.edu.cn, mason.zms@gmail.com
$\operatorname{Abstract}$
<FigureHere>
Figure 1: IIlustration of the ORL structure (d) based on an example sentence (b) with its corresponding dependency structure (a) and POS tags (c).
Unified opinion role labeling (ORL) aims to detect all possible opinion structures of `opinion-holder-target' in one shot, given a text. The existing transition-based unified method, unfortunately, is subject to longer opinion terms and fails to solve the term overlap issue. Current top performance has been achieved by employing the span-based graph model, which however still suffers from both high model complexity and insufficient interaction among opinions and roles. In this work, we investigate a novel solution by revisiting the transition architecture, and augmenting it with a pointer network (PointNet). The framework parses out all opinion structures in linear-time complexity, meanwhile breaks through the limitation of any length of terms with PointNet. To achieve the explicit opinion-role interactions, we further propose a unified dependency-opinion graph (UDOG), co-modeling the syntactic dependency structure and the partial opinion-role structure. We then devise a relation-centered graph aggregator (RCGA) to encode the multi-relational UDOG, where the resulting highorder representations are used to promote the predictions in the vanilla transition system. Our model achieves new state-of-the-art results on the MPQA benchmark. Analyses further demonstrate the superi- $\underset{1}{\operatorname*{\text{ority of our methods on both effcacy and efficiency.}}}$
忽略意见和角色术语之间的相互作用。于是，后续研究设计统一 (又名。端到端)ORL 的方法，i。e。, 通过基于LSTM 的模型将任务建模为“意见持有者-目标”结构提取 (Katiyar 等人。2016 )或基于过渡的模型 (Zhang 等人。2019 )。统一的方法有效地规避噪声传播，从而实现更好的性能。尽管如此，这些联合模型要么无法解决术语重叠问题$^2$,这是相关数据中普遍存在的一种情况，要么仍然容易受到冗长提及的检测
希望最近的工作 (Xia et al. 2021) 提出了一种基于跨度图模型的统一系统。它通过详尽地列举所有可能的观点和角色术语来工作，这些术语成功地帮助解决了术语重叠挑战 (Fuetal。2021 )。不幸的是，它仍然有两个关键的弱点。第一个是高模型复杂性，由于其跨度枚举性质。我们的数据统计显示，14。1意见和角色术语的% 跨越超过 8 个令牌，而短于 3 个令牌的术语占 54。8%。基于 span 的图模型试图以巨大的计算和时间消耗为代价搜索出尽可能多的术语。第二个问题在于意见提取和角色术语之间的相互作用不足，即。e。,它需要基于共享编码器的两个单独的子任务解码器。从本质上讲，意见和角色术语的检测应该相互依赖。例如在图。1,通过捕获意见和角色之间的相互作用，将“他”识别为意见“说”的持有者角色，可以同时为检测“代理”作为“说”的目标角色提供信息，因为意见控制持有者和目标。同样，作为“代理”目标的角色可以进一步导致预测相反的角色(i。e。,持有者)的意见术语“严重需要”
在这项工作中，我们试图通过研究一种新颖的统一
## 1 Introduction
在现实世界的应用程序中，对意见角色标签(ORL)的兴趣一直在增加，因为它可以精细地分析用户的意见(如图。1),i。e。,什么(持有者)对什么(目标)(Kim and Hovy 2006; Kannangara 2018; Tang et al. 2019) 表达了什么意见。对于 ORL 任务，传统工作采用流水线方法，即。e。,要么先提取所有术语(意见和角色术语),然后检测角色类型(Breck等人。2007;Yang等人。2012 )或首先提取意见术语，然后找到它们相应的角色术语 (Zhang, Liang, and Fu 2019; Zhang et al. 2020)。然而，这种级联方案可能会导致错误传播，同时
$^{*}$Corresponding author
Copyright $\odot2022$, Association for the Advancement of Ar-
tificial Intelligence (www.aaai.org). All rights reserved.
Codes at https://github.com/ChocoWu/SyPtrTrans-
$ORL$
$^{2}$For example in Fig. 1, the role term“the agency’involves with two opinion structure simultaneously. More generally, overlap issue refers to one word participating in multiple different span terms (Zeng et al. 2018; Fei, Ren, and Ji 2020; Li et al. 2021) .
|
[
[
0.1446533203125,
0.12249755859375,
0.8544921875,
0.12371826171875,
0.8544921875,
0.141357421875,
0.1446533203125,
0.14013671875
],
[
0.1004638671875,
0.1610107421875,
0.90185546875,
0.1610107421875,
0.90185546875,
0.175537109375,
0.1004638671875,
0.175537109375
],
[
0.1846923828125,
0.177978515625,
0.81201171875,
0.179931640625,
0.81201171875,
0.1976318359375,
0.1846923828125,
0.1956787109375
],
[
0.2059326171875,
0.1956787109375,
0.7939453125,
0.1956787109375,
0.7939453125,
0.210205078125,
0.2059326171875,
0.210205078125
],
[
0.1650390625,
0.205810546875,
0.83251953125,
0.2083740234375,
0.83251953125,
0.2259521484375,
0.1650390625,
0.2235107421875
],
[
0.17236328125,
0.2235107421875,
0.82861328125,
0.2235107421875,
0.82861328125,
0.2374267578125,
0.17236328125,
0.2374267578125
],
[
0.25,
0.2734375,
0.316162109375,
0.2734375,
0.316162109375,
0.28857421875,
0.25,
0.28857421875
],
[
0.580078125,
0.272705078125,
0.59716796875,
0.272705078125,
0.59716796875,
0.2822265625,
0.580078125,
0.2822265625
],
[
0.521484375,
0.279052734375,
0.5400390625,
0.279052734375,
0.5400390625,
0.292236328125,
0.521484375,
0.292236328125
],
[
0.6650390625,
0.2734375,
0.67333984375,
0.2734375,
0.67333984375,
0.276611328125,
0.6650390625,
0.276611328125
],
[
0.56201171875,
0.27978515625,
0.6005859375,
0.287353515625,
0.59619140625,
0.298583984375,
0.55810546875,
0.291748046875
],
[
0.7900390625,
0.285400390625,
0.81640625,
0.290283203125,
0.81396484375,
0.296630859375,
0.78759765625,
0.291748046875
],
[
0.8486328125,
0.291748046875,
0.861328125,
0.300537109375,
0.8486328125,
0.31005859375,
0.83642578125,
0.300537109375
],
[
0.86279296875,
0.292236328125,
0.9111328125,
0.294189453125,
0.91015625,
0.309326171875,
0.861328125,
0.306884765625
],
[
0.56884765625,
0.29541015625,
0.6005859375,
0.29541015625,
0.6005859375,
0.306884765625,
0.56884765625,
0.306884765625
],
[
0.521484375,
0.294189453125,
0.56201171875,
0.294189453125,
0.56201171875,
0.30615234375,
0.521484375,
0.30615234375
],
[
0.685546875,
0.29541015625,
0.740234375,
0.29541015625,
0.740234375,
0.306884765625,
0.685546875,
0.306884765625
],
[
0.74658203125,
0.29541015625,
0.78515625,
0.29541015625,
0.78515625,
0.306884765625,
0.74658203125,
0.306884765625
],
[
0.7900390625,
0.29541015625,
0.833984375,
0.29541015625,
0.833984375,
0.306884765625,
0.7900390625,
0.306884765625
],
[
0.60888671875,
0.296630859375,
0.6259765625,
0.296630859375,
0.6259765625,
0.30615234375,
0.60888671875,
0.30615234375
],
[
0.10296630859375,
0.296630859375,
0.463134765625,
0.296630859375,
0.463134765625,
0.31005859375,
0.10296630859375,
0.31005859375
],
[
0.6357421875,
0.296630859375,
0.67724609375,
0.296630859375,
0.67724609375,
0.306884765625,
0.6357421875,
0.306884765625
],
[
0.521484375,
0.306884765625,
0.564453125,
0.306884765625,
0.564453125,
0.318115234375,
0.521484375,
0.318115234375
],
[
0.56787109375,
0.306884765625,
0.59716796875,
0.306884765625,
0.59716796875,
0.318115234375,
0.56787109375,
0.318115234375
],
[
0.64208984375,
0.306884765625,
0.66845703125,
0.306884765625,
0.66845703125,
0.318115234375,
0.64208984375,
0.318115234375
],
[
0.7509765625,
0.306884765625,
0.78369140625,
0.306884765625,
0.78369140625,
0.318115234375,
0.7509765625,
0.318115234375
],
[
0.7998046875,
0.306884765625,
0.82666015625,
0.306884765625,
0.82666015625,
0.318115234375,
0.7998046875,
0.318115234375
],
[
0.70361328125,
0.30810546875,
0.7236328125,
0.30810546875,
0.7236328125,
0.318115234375,
0.70361328125,
0.318115234375
],
[
0.60888671875,
0.30810546875,
0.6259765625,
0.30810546875,
0.6259765625,
0.317626953125,
0.60888671875,
0.317626953125
],
[
0.84228515625,
0.30810546875,
0.85888671875,
0.30810546875,
0.85888671875,
0.317626953125,
0.84228515625,
0.317626953125
],
[
0.8798828125,
0.30810546875,
0.900390625,
0.30810546875,
0.900390625,
0.318115234375,
0.8798828125,
0.318115234375
],
[
0.5390625,
0.318115234375,
0.60107421875,
0.32080078125,
0.60107421875,
0.33203125,
0.5390625,
0.330078125
],
[
0.101318359375,
0.309326171875,
0.463134765625,
0.30810546875,
0.463134765625,
0.322509765625,
0.101318359375,
0.323974609375
],
[
0.10296630859375,
0.322509765625,
0.463134765625,
0.322509765625,
0.463134765625,
0.3359375,
0.10296630859375,
0.3359375
],
[
0.52392578125,
0.32568359375,
0.5400390625,
0.32568359375,
0.5400390625,
0.335205078125,
0.52392578125,
0.335205078125
],
[
0.6298828125,
0.32080078125,
0.65869140625,
0.32080078125,
0.65869140625,
0.33203125,
0.6298828125,
0.33203125
],
[
0.55126953125,
0.333251953125,
0.580078125,
0.333251953125,
0.580078125,
0.34033203125,
0.55126953125,
0.34033203125
],
[
0.583984375,
0.333984375,
0.60546875,
0.333984375,
0.60546875,
0.341552734375,
0.583984375,
0.341552734375
],
[
0.62744140625,
0.339111328125,
0.6708984375,
0.339111328125,
0.6708984375,
0.35107421875,
0.62744140625,
0.35107421875
],
[
0.71728515625,
0.339111328125,
0.75732421875,
0.339111328125,
0.75732421875,
0.35107421875,
0.71728515625,
0.35107421875
],
[
0.8388671875,
0.339111328125,
0.865234375,
0.339111328125,
0.865234375,
0.35107421875,
0.8388671875,
0.35107421875
],
[
0.101318359375,
0.3466796875,
0.4599609375,
0.3466796875,
0.4599609375,
0.35986328125,
0.101318359375,
0.35986328125
],
[
0.10296630859375,
0.335205078125,
0.463134765625,
0.335205078125,
0.463134765625,
0.348388671875,
0.10296630859375,
0.348388671875
],
[
0.6611328125,
0.354248046875,
0.68115234375,
0.354248046875,
0.68115234375,
0.35791015625,
0.6611328125,
0.35791015625
],
[
0.7900390625,
0.354248046875,
0.82177734375,
0.354248046875,
0.82177734375,
0.361083984375,
0.7900390625,
0.361083984375
],
[
0.101318359375,
0.371826171875,
0.45751953125,
0.371826171875,
0.45751953125,
0.385009765625,
0.101318359375,
0.385009765625
],
[
0.10296630859375,
0.35986328125,
0.461669921875,
0.35986328125,
0.461669921875,
0.373779296875,
0.10296630859375,
0.373779296875
],
[
0.521484375,
0.373779296875,
0.91259765625,
0.373779296875,
0.91259765625,
0.386962890625,
0.521484375,
0.386962890625
],
[
0.10296630859375,
0.385009765625,
0.461669921875,
0.385009765625,
0.461669921875,
0.3984375,
0.10296630859375,
0.3984375
],
[
0.521484375,
0.38818359375,
0.9111328125,
0.38818359375,
0.9111328125,
0.401611328125,
0.521484375,
0.401611328125
],
[
0.10296630859375,
0.397705078125,
0.4599609375,
0.397705078125,
0.4599609375,
0.410888671875,
0.10296630859375,
0.410888671875
],
[
0.521484375,
0.401611328125,
0.81884765625,
0.401611328125,
0.81884765625,
0.416748046875,
0.521484375,
0.416748046875
],
[
0.10296630859375,
0.410400390625,
0.4599609375,
0.410400390625,
0.4599609375,
0.423583984375,
0.10296630859375,
0.423583984375
],
[
0.101318359375,
0.421630859375,
0.463134765625,
0.423583984375,
0.463134765625,
0.4375,
0.101318359375,
0.43505859375
],
[
0.1004638671875,
0.43505859375,
0.459228515625,
0.43505859375,
0.459228515625,
0.448974609375,
0.1004638671875,
0.448974609375
],
[
0.10455322265625,
0.449462890625,
0.461669921875,
0.449462890625,
0.461669921875,
0.460205078125,
0.10455322265625,
0.460205078125
],
[
0.521484375,
0.44580078125,
0.91259765625,
0.4462890625,
0.91259765625,
0.4609375,
0.521484375,
0.460205078125
],
[
0.10296630859375,
0.460205078125,
0.461669921875,
0.460205078125,
0.461669921875,
0.473388671875,
0.10296630859375,
0.473388671875
],
[
0.52294921875,
0.4609375,
0.9111328125,
0.4609375,
0.9111328125,
0.474853515625,
0.52294921875,
0.474853515625
],
[
0.10455322265625,
0.474853515625,
0.461669921875,
0.474853515625,
0.461669921875,
0.48486328125,
0.10455322265625,
0.48486328125
],
[
0.521484375,
0.474853515625,
0.9111328125,
0.474853515625,
0.9111328125,
0.488037109375,
0.521484375,
0.488037109375
],
[
0.101318359375,
0.48486328125,
0.4599609375,
0.486083984375,
0.4599609375,
0.5,
0.101318359375,
0.498779296875
],
[
0.51953125,
0.4873046875,
0.91259765625,
0.488037109375,
0.91259765625,
0.50244140625,
0.51953125,
0.501953125
],
[
0.10296630859375,
0.498779296875,
0.4599609375,
0.498779296875,
0.4599609375,
0.51220703125,
0.10296630859375,
0.51220703125
],
[
0.51953125,
0.50048828125,
0.9111328125,
0.501953125,
0.9111328125,
0.5166015625,
0.51953125,
0.51513671875
],
[
0.10455322265625,
0.51318359375,
0.461669921875,
0.51318359375,
0.461669921875,
0.5234375,
0.10455322265625,
0.5234375
],
[
0.521484375,
0.5166015625,
0.9111328125,
0.5166015625,
0.9111328125,
0.52978515625,
0.521484375,
0.52978515625
],
[
0.10296630859375,
0.5234375,
0.463134765625,
0.5234375,
0.463134765625,
0.537109375,
0.10296630859375,
0.537109375
],
[
0.521484375,
0.52978515625,
0.91259765625,
0.52978515625,
0.91259765625,
0.54345703125,
0.521484375,
0.54345703125
],
[
0.10296630859375,
0.537109375,
0.463134765625,
0.537109375,
0.463134765625,
0.55029296875,
0.10296630859375,
0.55029296875
],
[
0.521484375,
0.54248046875,
0.79736328125,
0.54248046875,
0.79736328125,
0.55615234375,
0.521484375,
0.55615234375
],
[
0.101318359375,
0.548828125,
0.461669921875,
0.548828125,
0.461669921875,
0.5625,
0.101318359375,
0.5625
],
[
0.10296630859375,
0.5625,
0.455078125,
0.5625,
0.455078125,
0.57275390625,
0.10296630859375,
0.57275390625
],
[
0.53759765625,
0.5595703125,
0.91259765625,
0.5595703125,
0.91259765625,
0.57275390625,
0.53759765625,
0.57275390625
],
[
0.10296630859375,
0.57275390625,
0.461669921875,
0.57275390625,
0.461669921875,
0.5859375,
0.10296630859375,
0.5859375
],
[
0.52294921875,
0.57275390625,
0.9111328125,
0.57275390625,
0.9111328125,
0.5859375,
0.52294921875,
0.5859375
],
[
0.101318359375,
0.5859375,
0.463134765625,
0.5859375,
0.463134765625,
0.599609375,
0.101318359375,
0.599609375
],
[
0.52294921875,
0.5869140625,
0.9111328125,
0.5869140625,
0.9111328125,
0.60107421875,
0.52294921875,
0.60107421875
],
[
0.10455322265625,
0.60107421875,
0.4599609375,
0.60107421875,
0.4599609375,
0.611328125,
0.10455322265625,
0.611328125
],
[
0.52294921875,
0.60107421875,
0.9111328125,
0.60107421875,
0.9111328125,
0.6142578125,
0.52294921875,
0.6142578125
],
[
0.10296630859375,
0.6123046875,
0.461669921875,
0.6123046875,
0.461669921875,
0.62548828125,
0.10296630859375,
0.62548828125
],
[
0.52294921875,
0.6142578125,
0.9111328125,
0.6142578125,
0.9111328125,
0.62744140625,
0.52294921875,
0.62744140625
],
[
0.52294921875,
0.62744140625,
0.9111328125,
0.62744140625,
0.9111328125,
0.6416015625,
0.52294921875,
0.6416015625
],
[
0.521484375,
0.6416015625,
0.9111328125,
0.6416015625,
0.9111328125,
0.65478515625,
0.521484375,
0.65478515625
],
[
0.2418212890625,
0.6572265625,
0.35302734375,
0.65771484375,
0.35302734375,
0.67236328125,
0.2418212890625,
0.671875
],
[
0.52294921875,
0.65576171875,
0.91259765625,
0.65576171875,
0.91259765625,
0.66943359375,
0.52294921875,
0.66943359375
],
[
0.52294921875,
0.66943359375,
0.91259765625,
0.66943359375,
0.91259765625,
0.68310546875,
0.52294921875,
0.68310546875
],
[
0.2132568359375,
0.66015625,
0.2467041015625,
0.66015625,
0.2467041015625,
0.66943359375,
0.2132568359375,
0.66943359375
],
[
0.08905029296875,
0.6787109375,
0.476318359375,
0.6787109375,
0.476318359375,
0.689453125,
0.08905029296875,
0.689453125
],
[
0.521484375,
0.6845703125,
0.91259765625,
0.6845703125,
0.91259765625,
0.69775390625,
0.521484375,
0.69775390625
],
[
0.0882568359375,
0.69140625,
0.476318359375,
0.69140625,
0.476318359375,
0.705078125,
0.0882568359375,
0.705078125
],
[
0.52294921875,
0.69775390625,
0.91259765625,
0.69775390625,
0.91259765625,
0.7109375,
0.52294921875,
0.7109375
],
[
0.08660888671875,
0.705078125,
0.47802734375,
0.705078125,
0.47802734375,
0.7197265625,
0.08660888671875,
0.7197265625
],
[
0.52294921875,
0.7109375,
0.91259765625,
0.7109375,
0.91259765625,
0.724609375,
0.52294921875,
0.724609375
],
[
0.08660888671875,
0.71826171875,
0.47802734375,
0.71826171875,
0.47802734375,
0.73291015625,
0.08660888671875,
0.73291015625
],
[
0.0882568359375,
0.73193359375,
0.47802734375,
0.73193359375,
0.47802734375,
0.7470703125,
0.0882568359375,
0.7470703125
],
[
0.52294921875,
0.72607421875,
0.9111328125,
0.72607421875,
0.9111328125,
0.7392578125,
0.52294921875,
0.7392578125
],
[
0.521484375,
0.7392578125,
0.9111328125,
0.7392578125,
0.9111328125,
0.75244140625,
0.521484375,
0.75244140625
],
[
0.0906982421875,
0.74755859375,
0.191162109375,
0.74755859375,
0.191162109375,
0.7587890625,
0.0906982421875,
0.7587890625
],
[
0.1895751953125,
0.7470703125,
0.47802734375,
0.7470703125,
0.47802734375,
0.76025390625,
0.1895751953125,
0.76025390625
],
[
0.521484375,
0.75244140625,
0.91015625,
0.75244140625,
0.91015625,
0.7666015625,
0.521484375,
0.7666015625
],
[
0.08660888671875,
0.76025390625,
0.476318359375,
0.7587890625,
0.476318359375,
0.7734375,
0.08660888671875,
0.7744140625
],
[
0.521484375,
0.767578125,
0.91259765625,
0.767578125,
0.91259765625,
0.78076171875,
0.521484375,
0.78076171875
],
[
0.08905029296875,
0.7744140625,
0.47802734375,
0.7744140625,
0.47802734375,
0.78857421875,
0.08905029296875,
0.78857421875
],
[
0.521484375,
0.77978515625,
0.90869140625,
0.77978515625,
0.90869140625,
0.79345703125,
0.521484375,
0.79345703125
],
[
0.08660888671875,
0.78857421875,
0.476318359375,
0.78857421875,
0.476318359375,
0.80322265625,
0.08660888671875,
0.80322265625
],
[
0.0849609375,
0.8017578125,
0.47802734375,
0.8017578125,
0.47802734375,
0.81640625,
0.0849609375,
0.81640625
],
[
0.53759765625,
0.796875,
0.9111328125,
0.796875,
0.9111328125,
0.81005859375,
0.53759765625,
0.81005859375
],
[
0.1070556640625,
0.82568359375,
0.2607421875,
0.82568359375,
0.2607421875,
0.8388671875,
0.1070556640625,
0.8388671875
],
[
0.5390625,
0.82373046875,
0.9111328125,
0.82568359375,
0.9111328125,
0.84033203125,
0.5390625,
0.83837890625
],
[
0.08660888671875,
0.83837890625,
0.476318359375,
0.83837890625,
0.476318359375,
0.8515625,
0.08660888671875,
0.8515625
],
[
0.51953125,
0.8359375,
0.9111328125,
0.83837890625,
0.9111328125,
0.85302734375,
0.51953125,
0.85107421875
],
[
0.08660888671875,
0.85107421875,
0.44677734375,
0.85107421875,
0.44677734375,
0.8642578125,
0.08660888671875,
0.8642578125
],
[
0.521484375,
0.85107421875,
0.9111328125,
0.85107421875,
0.9111328125,
0.8642578125,
0.521484375,
0.8642578125
],
[
0.108642578125,
0.86279296875,
0.476318359375,
0.86279296875,
0.476318359375,
0.87744140625,
0.108642578125,
0.87744140625
],
[
0.51953125,
0.8623046875,
0.9140625,
0.86279296875,
0.9140625,
0.87744140625,
0.51953125,
0.876953125
],
[
0.0882568359375,
0.876953125,
0.1217041015625,
0.876953125,
0.1217041015625,
0.88818359375,
0.0882568359375,
0.88818359375
],
[
0.51953125,
0.87548828125,
0.68115234375,
0.87548828125,
0.68115234375,
0.8896484375,
0.51953125,
0.8896484375
]
] |
[
"掌握显式意见-角色交互：统一意见角色标记的语法辅助神经转换系统",
"Shengqiong Wu,l Hao Fei,1* Fei Li,l Donghong Ji,1 Meishan Zhang,2 Yijiang Liu,1 Chong Teng",
"1 Key Laboratory of Aerospace Information Security and Trusted Computing, Ministry",
"of Education, School of Cyber Science and Engineering, Wuhan University, China",
"2 Institute of Computing and Intelligence, Harbin Institute of Technology (Shenzhen), China",
"[ whuwsq, hao.fei, lifei_csnlp, cslyj, tengchong, dhji ] @whu.edu.cn, mason.zms@gmail.com",
"Abstract",
"ROO",
"(a)",
"",
"subin",
"ob",
"1",
"develop",
"says",
"(b) He",
"seriously",
"needs",
"money",
"the",
"Unified opinion role labeling (ORL) aims to detect all",
"agency",
"(c) PRP",
"VBZ",
"NN",
"VBZ",
"NN",
"RB",
"DT",
"TO",
"VB",
"RoleOpinior",
"possible opinion structures of ‘opinion-holder-target'in",
"one shot, given a text. The existing transition-based",
"(d)",
"Role",
"oldey",
"Varg",
"Role",
"Opinion",
"Role",
"opinion terms and fails to solve the term overlap is-",
"unified method, unfortunately, is subject to longer",
"",
"arget",
"employing the span-based graph model, which how",
"sue. Current top performance has been achieved by",
"Figure l:Illustration of the ORL structure (d) based",
"ever still suffers from both high model complexity",
"on an example sentence (b) with its corresponding de-",
"and insufficient interaction among opinions and roles.",
"pendency structure (a) and POS tags (c).",
"In this work, we investigate a novel solution by re",
"visiting the transition architecture, and augmenting",
"it with a pointer network (PointNet). The frame",
"work parses out all opinion structures in linear-time",
"忽略意见和角色术语之间的相互作用。于是，后续研究",
"complexity, meanwhile breaks through the limitation",
"设计统一（又名。端到端)ORL的方法，i。e。，通过基于",
"of any length of terms with PointNet.To achieve",
"LSTM的模型将任务建模为“意见持有者-目标”结构",
"the explicit opinion-role interactions, we further pro-",
"提取（Katiyar等人。2016）或基于过渡的模型（Zhang",
"pose a unified dependency-opinion graph (UDOG):",
"等人。2019）。统一的方法有效地规避噪声传播，从而",
"co-modelingthe syntactic dependency structure and",
"实现更好的性能。尽管如此，这些联合模型要么无法解",
"the partial opinion-role structure. We then devise a",
"决术语重叠问题2，这是相关数据中普遍存在的一种情",
"relation-centered graph aggregator (RCGA) to encode",
"况，要么仍然容易受到长提及的检测",
"the multi-relational UDOG, where the resulting high-",
"order representations are used to promote the pre",
"希望最近的工作（Xiaet al.2021)提出了一种基于跨",
"dictions in the vanilla transition system. Our model",
"度图模型的统一系统。它通过详尽地列举所有可能的观",
"achieves new state-of-the-art results on the MPQA",
"点和角色术语来工作，这些术语成功地帮助解决了术语",
"benchmark. Analyses further demonstrate the superi-",
"重叠挑战（Fuetal。2021）。不幸的是，它仍然有两个关",
"ority of our methods on both efficacy and efficiency.",
"键的弱点。第一个是高模型复杂性，由于其跨度枚举性",
"质。我们的数据统计显示，14。1意见和角色术语的%",
"跨越超过8个令牌，而短于3个令牌的术语占54。8%",
"Introduction",
"基于span的图模型试图以巨大的计算和时间消耗为代",
"价搜索出尽可能多的术语。第二个问题在于意见提取和",
"1",
"在现实世界的应用程序中，对意见角色标签（ORL）的",
"角色术语之间的相互作用不足，即。e。，它需要基于共",
"兴趣一直在增加，因为它可以精细地分析用户的意见",
"享编码器的两个单独的子任务解码器。从本质上讲，意",
"(如图。1),i。e。，什么（持有者）对什么（目标）(Kim",
"见和角色术语的检测应该相互依赖。例如在图。1，通",
"and Hovy 2006; Kannangara 2018; Tang et al. 2019)表",
"达了什么意见。对于ORL任务，传统工作采用流水线",
"过捕获意见和角色之间的相互作用，将“他”识别为意",
"见“说”的持有者角色，可以同时为检测“代理”作为",
"方法，即。e。",
"，要么先提取所有术语（意见和角色术",
"“说”的目标角色提供信息，因为意见控制持有者和目",
"语)，然后检测角色类型（Breck等人。2007；Yang等",
"标。同样，作为“代理”目标的角色可以进一步导致预",
"人。2012）或首先提取意见术语，然后找到它们相应",
"测相反的角色（i。e。，持有者）的意见术语“严重需要",
"的角色术语（Zhang，Liang，andFu2019;Zhangetal",
"2020）。然而，这种级联方案可能会导致错误传播，同时",
"在这项工作中，我们试图通过研究一种新颖的统一",
"*Corresponding author",
"2For example in Fig.1, the role term ‘the agency'in-",
"Copyright 2022, Association for the Advancement of Ar-",
"volves with two opinion structure simultaneously. More gen-",
"tificial Intelligence (www.aaai.org). All rights reserved.",
"erally, overlap issue refers to one word participating in mul-",
"'Codes at https://github.com/ChocoWu/SyPtrTrans-",
"tiple different span terms (Zeng et al. 2018; Fei, Ren, and",
"ORL",
"Ji 2020; Li et al.2021)."
] |
{
"arxiv_id": "2110.02001",
"language": "zh",
"page_idx": 1
}
|
||
2106.13863-zh-0
|
2106.13863-zh-0.jpg
|
# Steerable 3D Spherical Neurons
Pavlo Melnyk, Michael Felsberg, Mårten Wadenbäck
Computer Vision Laboratory, Department of Electrical Engineering, Linköping University
{ pavlo.melnyk, michael.felsberg, marten.wadenback} @liu.se
<FigureHere>
Figure 1: Our approach overview: First, we train a classifier ( Ancestor ) consisting of the spherical neurons (necessarily in the first layer) to classify 3D point clouds in a canonical orientation. We then fix all the learned parameters and construct rotation-equivariant spherical filter banks $B(S)(7)$ from the first layer parameters. Next, we compute the interpolation coefficients $v(R)$ (14) to fulfill the steerability constraint (11). The result is a Steerable classifier, where y is equivariant and h is invariant to rotations of the model input.
# $\mathbf{Abstract}$
Emerging from low-level vision theory, steerable filters found their counterpart in prior work on steerable convolutional neural networks equivariant to rigid transformations. In our work, we propose a steerable feed-forward learningbased approach that consists of neurons with spherical decision surfaces and operates on point clouds. Such spherical neurons are obtained by conformal embedding of Euclidean space and have recently been revisited in the context of learning representations of point sets. Focusing on 3D geometry, we exploit the isometry property of spherical neurons and derive a 3D steerability constraint. After training spherical neurons to classify point clouds in a canonical orientation, we use a tetrahedron basis to quadruplicate the neurons and construct rotation-equivariant spherical filter banks. We then apply the derived constraint to interpolate the filter bank outputs and, thus, obtain a rotationinvariant network. Finally, we use a synthetic point set and real-world 3D skeleton data to verify our theoretical findings. The code is available at https://github.com/ pavlo-melnyk/steerable-3d-neurons.
# 1 $\textbf{Introduction}$
我们提出了一个前馈模型，该模型由可转向的 3D 神经元组成，用于点云分类，这是许多应用中的重要且具有挑战性的问题，例如自动驾驶，人机交互和混合现实安装。构造旋转等变(可操纵)模型使我们能够使用给定点云的特征来合成具有不同方向的同一点云的特征。此外，旋转等变网络可以产生旋转不变的预测， 因此降低了学习的数据增强要求
在本文中，我们使用保形嵌入来实现可操纵性，以获得高阶决策面。遵循 Melnyk et al.(2021)最近工作的动机，我们专注于 3D 几何形状和球形决策面，论证了它们对欧几里得空间问题的自然适用性。我们展示了如何将球形神经元 i.e., 超球形神经元 (Banarer et al., 2003b) 或其对 3D 输入点集的概括---几何神经元(Melnyk et al., 2021)---变成可控制的神经元。使球形神经元可操纵增加了 Melnyk et al.(2021)先前工作的实用价值。
我们证明了上述任何维度的球形神经元仅需要多达一度的球形谐波即可适应旋转的影响。这使我们能够导出此类神经元的 3D 可转向性约束，并描述从预训练分类器创建可转向模型的方法(请参见图 1)。使
用合成的俄罗斯方块数据集 (Thomas et al., 2018) 和来自 UTKinect-Action3D 数据集 (Xia et al., 2012) 的骨架数据，我们验证了导出的约束，并检查了其相对于输人扰动的稳定性
重要的是，我们专注于我们的可操纵方法的理论方面，以产生旋转不变的预测，并表明它可以通过使用合成和真实的 3D 数据来实现 (第$\overline{\mathrm{E}}$5.3 节)。剩下要做的是设计一种切实可行的方法来实现它
我们贡献的核心是关于可操纵性和等方差的新颖理论结果，可以总结如下：(a)我们证明了球形神经元(Banarer et al., 2003b; Melnyk et al., 2021) 在旋转输人上的激活仅在旋转角度上变化多达一阶球面谐波(第回4.1节)
(b) 基于旋转到正四面体的相应顶点的四个球形神经元的最小集合，我们构造了旋转等变球形滤波器组(E4.2 节),并得出了我们论文的主要结果(E4.3 节) --球形神经元(11)和(15)的 3D 可操纵性约束
|
[
[
0.318603515625,
0.106689453125,
0.6806640625,
0.10797119140625,
0.6806640625,
0.126953125,
0.318603515625,
0.1256103515625
],
[
0.27685546875,
0.149658203125,
0.7236328125,
0.149658203125,
0.7236328125,
0.1634521484375,
0.27685546875,
0.1634521484375
],
[
0.2344970703125,
0.169189453125,
0.76318359375,
0.169189453125,
0.76318359375,
0.179931640625,
0.2344970703125,
0.179931640625
],
[
0.240234375,
0.186279296875,
0.7685546875,
0.186279296875,
0.7685546875,
0.199462890625,
0.240234375,
0.199462890625
],
[
0.0906982421875,
0.2310791015625,
0.1829833984375,
0.2310791015625,
0.1829833984375,
0.248779296875,
0.0906982421875,
0.248779296875
],
[
0.65625,
0.238037109375,
0.69677734375,
0.238037109375,
0.69677734375,
0.24560546875,
0.65625,
0.24560546875
],
[
0.736328125,
0.25830078125,
0.7509765625,
0.25830078125,
0.7509765625,
0.266357421875,
0.736328125,
0.266357421875
],
[
0.85888671875,
0.25390625,
0.8857421875,
0.25390625,
0.8857421875,
0.25830078125,
0.85888671875,
0.25830078125
],
[
0.0906982421875,
0.26318359375,
0.47802734375,
0.26318359375,
0.47802734375,
0.27783203125,
0.0906982421875,
0.27783203125
],
[
0.810546875,
0.261474609375,
0.88623046875,
0.261474609375,
0.88623046875,
0.271484375,
0.810546875,
0.271484375
],
[
0.0906982421875,
0.27978515625,
0.4794921875,
0.27978515625,
0.4794921875,
0.29345703125,
0.0906982421875,
0.29345703125
],
[
0.0906982421875,
0.294189453125,
0.47802734375,
0.294189453125,
0.47802734375,
0.30810546875,
0.0906982421875,
0.30810546875
],
[
0.0906982421875,
0.31005859375,
0.47802734375,
0.31005859375,
0.47802734375,
0.323974609375,
0.0906982421875,
0.323974609375
],
[
0.08905029296875,
0.324462890625,
0.4794921875,
0.324462890625,
0.4794921875,
0.33837890625,
0.08905029296875,
0.33837890625
],
[
0.65185546875,
0.330078125,
0.69677734375,
0.330078125,
0.69677734375,
0.34033203125,
0.65185546875,
0.34033203125
],
[
0.0906982421875,
0.34033203125,
0.47802734375,
0.34033203125,
0.47802734375,
0.354248046875,
0.0906982421875,
0.354248046875
],
[
0.0906982421875,
0.354736328125,
0.4794921875,
0.354736328125,
0.4794921875,
0.36865234375,
0.0906982421875,
0.36865234375
],
[
0.5361328125,
0.361083984375,
0.56201171875,
0.361083984375,
0.56201171875,
0.373779296875,
0.5361328125,
0.373779296875
],
[
0.82421875,
0.361083984375,
0.8837890625,
0.361083984375,
0.8837890625,
0.365478515625,
0.82421875,
0.365478515625
],
[
0.09234619140625,
0.37060546875,
0.4794921875,
0.37060546875,
0.4794921875,
0.3837890625,
0.09234619140625,
0.3837890625
],
[
0.810546875,
0.367431640625,
0.8857421875,
0.367431640625,
0.8857421875,
0.37744140625,
0.810546875,
0.37744140625
],
[
0.0906982421875,
0.385009765625,
0.47802734375,
0.385009765625,
0.47802734375,
0.3984375,
0.0906982421875,
0.3984375
],
[
0.0906982421875,
0.40087890625,
0.476318359375,
0.40087890625,
0.476318359375,
0.4140625,
0.0906982421875,
0.4140625
],
[
0.0931396484375,
0.416748046875,
0.476318359375,
0.416748046875,
0.476318359375,
0.4267578125,
0.0931396484375,
0.4267578125
],
[
0.521484375,
0.41845703125,
0.90869140625,
0.41845703125,
0.90869140625,
0.431884765625,
0.521484375,
0.431884765625
],
[
0.0906982421875,
0.43115234375,
0.4794921875,
0.43115234375,
0.4794921875,
0.4443359375,
0.0906982421875,
0.4443359375
],
[
0.521484375,
0.43310546875,
0.90869140625,
0.43310546875,
0.90869140625,
0.4462890625,
0.521484375,
0.4462890625
],
[
0.08905029296875,
0.4443359375,
0.47802734375,
0.44580078125,
0.47802734375,
0.458984375,
0.08905029296875,
0.458251953125
],
[
0.51953125,
0.448974609375,
0.90869140625,
0.447509765625,
0.90869140625,
0.4609375,
0.51953125,
0.462158203125
],
[
0.09234619140625,
0.462158203125,
0.476318359375,
0.462158203125,
0.476318359375,
0.472900390625,
0.09234619140625,
0.472900390625
],
[
0.51953125,
0.462158203125,
0.91015625,
0.46337890625,
0.91015625,
0.477783203125,
0.51953125,
0.4765625
],
[
0.0906982421875,
0.47607421875,
0.4794921875,
0.47607421875,
0.4794921875,
0.490478515625,
0.0906982421875,
0.490478515625
],
[
0.521484375,
0.479248046875,
0.90869140625,
0.479248046875,
0.90869140625,
0.492431640625,
0.521484375,
0.492431640625
],
[
0.08905029296875,
0.4912109375,
0.4794921875,
0.490478515625,
0.4794921875,
0.50390625,
0.08905029296875,
0.5048828125
],
[
0.52294921875,
0.494384765625,
0.90869140625,
0.494384765625,
0.90869140625,
0.50830078125,
0.52294921875,
0.50830078125
],
[
0.0906982421875,
0.505859375,
0.4794921875,
0.505859375,
0.4794921875,
0.51953125,
0.0906982421875,
0.51953125
],
[
0.51953125,
0.5087890625,
0.90869140625,
0.50830078125,
0.90869140625,
0.52294921875,
0.51953125,
0.5234375
],
[
0.0906982421875,
0.521484375,
0.47802734375,
0.521484375,
0.47802734375,
0.53515625,
0.0906982421875,
0.53515625
],
[
0.52294921875,
0.52587890625,
0.90673828125,
0.52587890625,
0.90673828125,
0.5361328125,
0.52294921875,
0.5361328125
],
[
0.0906982421875,
0.5361328125,
0.47802734375,
0.5361328125,
0.47802734375,
0.5498046875,
0.0906982421875,
0.5498046875
],
[
0.521484375,
0.5390625,
0.91015625,
0.5390625,
0.91015625,
0.55322265625,
0.521484375,
0.55322265625
],
[
0.08905029296875,
0.5517578125,
0.42724609375,
0.5517578125,
0.42724609375,
0.56494140625,
0.08905029296875,
0.56494140625
],
[
0.521484375,
0.5546875,
0.6298828125,
0.5546875,
0.6298828125,
0.568359375,
0.521484375,
0.568359375
],
[
0.09234619140625,
0.5966796875,
0.10540771484375,
0.5966796875,
0.10540771484375,
0.60791015625,
0.09234619140625,
0.60791015625
],
[
0.1298828125,
0.5966796875,
0.261474609375,
0.5966796875,
0.261474609375,
0.611328125,
0.1298828125,
0.611328125
],
[
0.52392578125,
0.62060546875,
0.90673828125,
0.62060546875,
0.90673828125,
0.6337890625,
0.52392578125,
0.6337890625
],
[
0.0906982421875,
0.62744140625,
0.476318359375,
0.62744140625,
0.476318359375,
0.6416015625,
0.0906982421875,
0.6416015625
],
[
0.52392578125,
0.63525390625,
0.90673828125,
0.63525390625,
0.90673828125,
0.6484375,
0.52392578125,
0.6484375
],
[
0.09234619140625,
0.64306640625,
0.4755859375,
0.64306640625,
0.4755859375,
0.6572265625,
0.09234619140625,
0.6572265625
],
[
0.52294921875,
0.65087890625,
0.90673828125,
0.65087890625,
0.90673828125,
0.6640625,
0.52294921875,
0.6640625
],
[
0.09234619140625,
0.6591796875,
0.476318359375,
0.6591796875,
0.476318359375,
0.67236328125,
0.09234619140625,
0.67236328125
],
[
0.52392578125,
0.66650390625,
0.6396484375,
0.66650390625,
0.6396484375,
0.6796875,
0.52392578125,
0.6796875
],
[
0.09234619140625,
0.673828125,
0.476318359375,
0.673828125,
0.476318359375,
0.6875,
0.09234619140625,
0.6875
],
[
0.0931396484375,
0.689453125,
0.476318359375,
0.689453125,
0.476318359375,
0.70263671875,
0.0931396484375,
0.70263671875
],
[
0.0906982421875,
0.70263671875,
0.47802734375,
0.7041015625,
0.47802734375,
0.71826171875,
0.0906982421875,
0.71728515625
],
[
0.5400390625,
0.70068359375,
0.90625,
0.70068359375,
0.90625,
0.7109375,
0.5400390625,
0.7109375
],
[
0.09234619140625,
0.7197265625,
0.3203125,
0.7197265625,
0.3203125,
0.73291015625,
0.09234619140625,
0.73291015625
],
[
0.52294921875,
0.7138671875,
0.90673828125,
0.7138671875,
0.90673828125,
0.72802734375,
0.52294921875,
0.72802734375
],
[
0.108642578125,
0.73486328125,
0.47802734375,
0.73486328125,
0.47802734375,
0.74853515625,
0.108642578125,
0.74853515625
],
[
0.52294921875,
0.72998046875,
0.90869140625,
0.72998046875,
0.90869140625,
0.74365234375,
0.52294921875,
0.74365234375
],
[
0.521484375,
0.74365234375,
0.80126953125,
0.744140625,
0.80126953125,
0.7587890625,
0.521484375,
0.75830078125
],
[
0.0931396484375,
0.75048828125,
0.476318359375,
0.75048828125,
0.476318359375,
0.7646484375,
0.0931396484375,
0.7646484375
],
[
0.09234619140625,
0.76513671875,
0.476318359375,
0.76513671875,
0.476318359375,
0.77880859375,
0.09234619140625,
0.77880859375
],
[
0.09234619140625,
0.78076171875,
0.476318359375,
0.78076171875,
0.476318359375,
0.79443359375,
0.09234619140625,
0.79443359375
],
[
0.5390625,
0.77783203125,
0.90673828125,
0.77783203125,
0.90673828125,
0.79150390625,
0.5390625,
0.79150390625
],
[
0.09234619140625,
0.79541015625,
0.476318359375,
0.79541015625,
0.476318359375,
0.8095703125,
0.09234619140625,
0.8095703125
],
[
0.521484375,
0.79345703125,
0.90673828125,
0.79345703125,
0.90673828125,
0.806640625,
0.521484375,
0.806640625
],
[
0.0906982421875,
0.81005859375,
0.47802734375,
0.81005859375,
0.47802734375,
0.82373046875,
0.0906982421875,
0.82373046875
],
[
0.521484375,
0.80810546875,
0.90673828125,
0.80810546875,
0.90673828125,
0.8212890625,
0.521484375,
0.8212890625
],
[
0.09234619140625,
0.82568359375,
0.476318359375,
0.82568359375,
0.476318359375,
0.8388671875,
0.09234619140625,
0.8388671875
],
[
0.52294921875,
0.82373046875,
0.90673828125,
0.82373046875,
0.90673828125,
0.8369140625,
0.52294921875,
0.8369140625
],
[
0.09234619140625,
0.84130859375,
0.476318359375,
0.84130859375,
0.476318359375,
0.85498046875,
0.09234619140625,
0.85498046875
],
[
0.521484375,
0.83837890625,
0.57177734375,
0.83837890625,
0.57177734375,
0.85302734375,
0.521484375,
0.85302734375
],
[
0.0931396484375,
0.85693359375,
0.1282958984375,
0.85693359375,
0.1282958984375,
0.869140625,
0.0931396484375,
0.869140625
],
[
0.108642578125,
0.87255859375,
0.4755859375,
0.87255859375,
0.4755859375,
0.88623046875,
0.108642578125,
0.88623046875
],
[
0.5361328125,
0.87060546875,
0.90673828125,
0.8720703125,
0.90673828125,
0.88623046875,
0.5361328125,
0.88525390625
],
[
0.10455322265625,
0.88720703125,
0.4755859375,
0.88720703125,
0.4755859375,
0.90087890625,
0.10455322265625,
0.90087890625
],
[
0.52294921875,
0.88720703125,
0.90673828125,
0.88720703125,
0.90673828125,
0.90087890625,
0.52294921875,
0.90087890625
],
[
0.09478759765625,
0.90380859375,
0.4755859375,
0.90380859375,
0.4755859375,
0.9140625,
0.09478759765625,
0.9140625
],
[
0.521484375,
0.9013671875,
0.90673828125,
0.9013671875,
0.90673828125,
0.91552734375,
0.521484375,
0.91552734375
],
[
0.0906982421875,
0.91748046875,
0.47802734375,
0.91748046875,
0.47802734375,
0.93115234375,
0.0906982421875,
0.93115234375
],
[
0.51953125,
0.91748046875,
0.84619140625,
0.91748046875,
0.84619140625,
0.93115234375,
0.51953125,
0.93115234375
],
[
0.494384765625,
0.95556640625,
0.5048828125,
0.95556640625,
0.5048828125,
0.96728515625,
0.494384765625,
0.96728515625
]
] |
[
"Steerable 3D Spherical Neurons",
"Pavlo Melnyk,MichaelFelsberg,Marten Wadenback",
"ComputerVisionLaboratory,DepartmentofElectrical Engineering,LinkopingUniversity",
"( pavlo.melnyk, michael.felsberg, marten.wadenback } @liu.se",
"Abstract",
"Ancestor",
"rest",
"nvas",
"Emerging from low-level vision theory, steerable filters",
"class prediction",
"found their counterpart in prior work on steerable convolu-",
"tional neural networks equivariant to rigid transformations.",
"In our work, we propose a steerable feed-forward learning",
"based approach that consists of neurons with spherical de-",
"Steerable",
"cision surfaces and operates on point clouds. Such spher-",
"ical neurons are obtained by conformal embedding of Eu-",
"RX",
"tion-invarian",
"clidean space and have recently been revisited in the con-",
"class prediction",
"text of learning representations of point sets. Focusing on",
"3D geometry, we exploit the isometry property of spherical",
"neuronsandderivea3Dsteerabilityconstraint.Aftertrain",
"Figure 1:Our approach overview: First, we train a clas-",
"ing spherical neurons to classify point clouds in a canoni-",
"sifier (Ancestor ) consisting of the spherical neurons (nec-",
"cal orientation, we use a tetrahedron basis to quadruplicate",
"essarily in the first layer) to classify 3D point clouds in a",
"theneurons and construct rotation-equivariant spherical fil",
"canonical orientation. We then fix all the learned parame-",
"ter banks. We then apply the derived constraint to inter-",
"ters and construct rotation-equivariant spherical filter banks",
"polate the filter bank outputs and, thus, obtain a rotation-",
"B(S) (7) from the first layer parameters. Next, we com-",
"invariant network. Finally, we use a synthetic point set and",
"pute the interpolation coefficients v(R) (14) to fulfill the",
"real-world 3D skeleton data to verify our theoretical find-",
"steerabilityconstraint(11)TheresultisaSteerableclassi",
"ings. The code is available at https: //github.com/",
"fier, where y is equivariant and h is invariant to rotations of",
"pavlo-melnyk/steerable-3d-neurons.",
"the model input.",
"1",
"Introduction",
"用合成的俄罗斯方块数据集（Thomasetal.,2018）和来",
"我们提出了一个前馈模型，该模型由可转向的3D神经",
"自UTKinect-Action3D数据集（Xiaetal.，2012）的骨架",
"元组成，用于点云分类，这是许多应用中的重要且具",
"数据，我们验证了导出的约束，并检查了其相对于输",
"有挑战性的问题，例如自动驾驶，人机交互和混合现",
"人扰动的稳定性",
"实安装。构造旋转等变（可操纵）模型使我们能够使",
"用给定点云的特征来合成具有不同方向的同一点云的",
"特征。此外，旋转等变网络可以产生旋转不变的预测，",
"重要的是，我们专注于我们的可操纵方法的理论方",
"因此降低了学习的数据增强要求",
"面，以产生旋转不变的预测，并表明它可以通过使用",
"在本文中，我们使用保形嵌入来实现可操纵性，以",
"合成和真实的3D数据来实现（第E5.3节）。剩下要做",
"的是设计一种切实可行的方法来实现它",
"获得高阶决策面。遵循Melnyketal.(2021)最近工作",
"的动机，我们专注于3D几何形状和球形决策面，论",
"证了它们对欧几里得空间问题的自然适用性。我们展",
"我们贡献的核心是关于可操纵性和等方差的新颖理",
"示了如何将球形神经元i.e.，，超球形神经元（Banarer",
"论结果，可以总结如下：（a）我们证明了球形神经元",
"etal.,2003b)或其对3D输入点集的概括---几何神经元",
"(Banareretal.,2003b;Melnyketal.,2021)在旋转输人上",
"(Melnyketal.,2021）---变成可控制的神经元。使球形神",
"的激活仅在旋转角度上变化多达一阶球面谐波（第E",
"经元可操纵增加了Melnyketal.(2021)先前工作的实用",
"4.1节)",
"价值，",
"我们证明了上述任何维度的球形神经元仅需要多达",
"（b）基于旋转到正四面体的相应顶点的四个球形神",
"度的球形谐波即可适应旋转的影响。这使我们能够",
"经元的最小集合，我们构造了旋转等变球形滤波器组",
"导出此类神经元的3D可转向性约束，并描述从预训",
"(E4.2节），并得出了我们论文的主要结果（E4.3节）",
"练分类器创建可转向模型的方法（请参见图1）。使",
"---球形神经元（11)和（15)的3D可操纵性约束",
"1"
] |
{
"arxiv_id": "2106.13863",
"language": "zh",
"page_idx": 0
}
|
||
2106.13863-zh-1
|
2106.13863-zh-1.jpg
|
# $\textbf{2}$ $\textbf{Related work}$
## 2.1 Steerability and equivariance
可控性是来自早期视觉和图像处理的强大概念(Freeman et al., 1991; Knutsson et al., 1992; Simoncelli et al., 1992; Perona, 1995; Simoncelli & Freeman, 1995; Teo & Hel-Or, 1998) 在深度学习时代引起共鸣。在 Freeman et al. (1991) 的开创性工作中介绍了可操纵滤波器的实用性及其构造的主要定理
在过去的几十年中，计算机视觉环境中的几何等方差一直是广泛研究的对象 (Van Gool et al., 1995)。等方差是可控制性的必要属性，因为可控制性需要根据特定组的操作更改函数输出。因此，输出需要包含作用于输入的组的表示。因此，操作员必须与组通勤(请注意，表示可能会从输入更改为输出)。因此，Lie 理论可用于研究等方差和不变性(Reisert, 2008) 的可操纵滤波器
如今，与等方差相关的研究已扩展到深度学习，e.g., , Fuchs et al. (2020) , Thomas et al. (2018)和 Zhao et al. (2020)的 SE(3)等变模型，以及 Anderson et al.(2019) 的 SO(3)等变网络，以及 Marcos et al.(2017)和 Kondor et al.(2018)。在当前的工作中也越来越多地使用可操纵的过滤器概念。例如，Cohen &Welling (2016) 的工作考虑了图像数据，并提出了一种 CNN 架构，以产生具有可操纵特征的等变表示，与传统 CNN 相比，它涉及的参数更少。他们考虑了二面角组(离散旋转，平移和反射),并提出了工作中的可操纵表示作为基本特征类型的形成。他们的方法的一个局限性是旋转被限制在四个方向上，i.e., 乘以$\pi/2$。最近，Weiler et al. (2018b) 利用组卷积并引人了对图像进行操作的可控滤波器卷积神经网络(SFCNN),以在平移和离散旋转下共同获得等方差。在他们的工作中，滤波器组是学习的，而不是固定的。此外，Weiler &Cesa(2019) 的工作提出了E (2) -等变可控 CNN 的统一框架，并介绍了它们的一般理论
Weiler et al. (2018a) 提出的用于 3D 数据的可操纵CNN 与我们的工作密切相关。作者采用标量，矢量和张量场的组合作为由 SO(3) 表示转换的特征，并提出了一个与 SE(3) 转换等变的模型。他们还考虑了适用于特征空间的非标量分量的不同类型的非线性。Fuchs et al.(2020)的新型 SE (3)-等变方法引人了一种自注意机制，该机制对其输入的全局旋转和平移不变，并解决了其他等变方法 e.g., Thomas et al.(2018) 中角度约束滤波器的局限性。值得注意的是，Jing et al.(2020) 的工作提出了几何矢量感知器 (GVP),该几何矢量感知器由输入标量和矢量特征的两个线性变换组成，然后是非线性。相对于 3D 欧几里得空间中旋转和反射的任意组合，GVP 标量和矢量输出分别是不变的和等变的。
# 2.2 $\textbf{Conformal modeling and the hyper}$ sphere neuron
Li et al. (2001a) 彻底讨论了欧几里得几何保形嵌人的实用性以及与 Minkowski 空间的紧密联系。一个重要的结果是，可以像 Perwass et al.(2003)的工作一样，使用
保形空间 (Li et al., 2001b) 中的表示来构造超球面决策面。Banarer et al.(2003b) 提出的超球面神经元就是这样一个球面分类器。值得注意的是，由于可以将超球体视为超平面的概括，因此可以将标准神经元视为超球体神经元的特殊情况。
在前馈网络中堆叠多个超球面神经元会产生多层超球面感知器 ( MLHP) , Banarer et al.(2003a) 显示该感知器在某些分类任务中优于标准 MLP。然而，它在点集上的应用没有被讨论。。这激发了 Melnyk et al. (2021) 在几何神经元上的工作，其中学习的参数显示为表示球形决策面的组合。此外，几何(和超球面)神经元激活被证明在 3D 欧几里得空间中是等距的。在我们的工作中，我们使用后一个观察结果作为推导可操纵性约束的必要条件。
# 2.3 Comparison to other methods
可操纵性要求滤波器要么是球谐函数(的组合)(参见e.g., Fuchs et al.(2020) ), 要么是基于学习的神经元构造的，并且表现得像这样(我们的工作)
将我们的方法与其他等变网络区分开来的关键点是， 我们不限制可学习参数的空间(与 e.g., Thomas et al. (2018) , Weiler et al. ( 2018a) 和 Fuchs et al.(2020)相反) 而是从自由训练的基础网络构建我们的可控模型，正如我们在$\mathbb{E}$4 节中详细讨论的那样。也就是说，相关工作使用球谐函数作为原子，i.e.,是手工设计的基础，并且仅学习约束下的线性系数。相反，我们从手部设计中继承的唯一东西是一阶谐波的约束(请参见定理$\overline{\mathrm{E}}$ 4.1)-所有其他自由度都是可学习的
此外，即使我们在当前工作中仅考虑标量场，也可
以将$\overline{\mathbb{E}}$4 节的理论结果应用于更广泛的特征场类别
# $\textbf{3}$ $\textbf{Background}$
# 3.1 Steerability
根据 Freeman et al.(1991), 如果可以将 2D 函数$f(x,y)$ 编写为自身旋转版本的线性组合，则可以将其转向。e。, 当它满足约束
$$f^\theta(x,y)=\sum_{j=1}^Mv_j(\theta)f^{\theta_j}(x,y)\:,$$
(1)
时，其中$v_j(\theta)$是插值函数，$\theta_j$是基函数方向，$M$是引导函数所需的基函数项的数量。在 Knutsson et al. (1992的工作中可以找到另一种公式。在 3D 中，转向方程变为
(2)
$$f^R(x,y,z)=\sum_{j=1}^Mv_j(R)f^{R_j}(x,y,z)\:,$$
其中$f^R(x,y,z)$是由$R\in\operatorname{SO}(3)$旋转的$f( x, y, z)$ ,
并且每个$R_j\in$SO(3)定向相应的$j$ th 基函数
Freeman et al.(1991)中的定理 1、2 和 4 描述了可操纵性约束(1)和(2)成立的条件，以及如何分别确定 2D 和 3D 情况下的基函数的最小数量
2
|
[
[
0.0906982421875,
0.0706787109375,
0.26708984375,
0.0706787109375,
0.26708984375,
0.085205078125,
0.0906982421875,
0.085205078125
],
[
0.525390625,
0.07452392578125,
0.90625,
0.07452392578125,
0.90625,
0.085205078125,
0.525390625,
0.085205078125
],
[
0.52392578125,
0.08837890625,
0.90673828125,
0.08837890625,
0.90673828125,
0.10162353515625,
0.52392578125,
0.10162353515625
],
[
0.1356201171875,
0.1029052734375,
0.3857421875,
0.10418701171875,
0.3857421875,
0.11871337890625,
0.1356201171875,
0.117431640625
],
[
0.521484375,
0.10162353515625,
0.90673828125,
0.1029052734375,
0.90673828125,
0.117431640625,
0.521484375,
0.11614990234375
],
[
0.52294921875,
0.11871337890625,
0.90673828125,
0.11871337890625,
0.90673828125,
0.1319580078125,
0.52294921875,
0.1319580078125
],
[
0.09234619140625,
0.10418701171875,
0.143798828125,
0.10418701171875,
0.143798828125,
0.11553955078125,
0.09234619140625,
0.11553955078125
],
[
0.0931396484375,
0.1307373046875,
0.476318359375,
0.1307373046875,
0.476318359375,
0.141357421875,
0.0931396484375,
0.141357421875
],
[
0.521484375,
0.1331787109375,
0.69287109375,
0.1331787109375,
0.69287109375,
0.146484375,
0.521484375,
0.146484375
],
[
0.0931396484375,
0.14453125,
0.4755859375,
0.14453125,
0.4755859375,
0.1578369140625,
0.0931396484375,
0.1578369140625
],
[
0.5390625,
0.149658203125,
0.90869140625,
0.149658203125,
0.90869140625,
0.1634521484375,
0.5390625,
0.1634521484375
],
[
0.09234619140625,
0.160400390625,
0.476318359375,
0.160400390625,
0.476318359375,
0.173583984375,
0.09234619140625,
0.173583984375
],
[
0.51953125,
0.1634521484375,
0.90869140625,
0.1641845703125,
0.90869140625,
0.1787109375,
0.51953125,
0.177978515625
],
[
0.08905029296875,
0.173583984375,
0.476318359375,
0.1749267578125,
0.476318359375,
0.1881103515625,
0.08905029296875,
0.1875
],
[
0.521484375,
0.1787109375,
0.90869140625,
0.179931640625,
0.90869140625,
0.1944580078125,
0.521484375,
0.1932373046875
],
[
0.09234619140625,
0.189453125,
0.47802734375,
0.189453125,
0.47802734375,
0.20263671875,
0.09234619140625,
0.20263671875
],
[
0.52392578125,
0.1944580078125,
0.90869140625,
0.1944580078125,
0.90869140625,
0.2083740234375,
0.52392578125,
0.2083740234375
],
[
0.09234619140625,
0.2052001953125,
0.27294921875,
0.2052001953125,
0.27294921875,
0.2183837890625,
0.09234619140625,
0.2183837890625
],
[
0.52294921875,
0.208984375,
0.90673828125,
0.210205078125,
0.90673828125,
0.2235107421875,
0.52294921875,
0.222900390625
],
[
0.1070556640625,
0.2203369140625,
0.476318359375,
0.2203369140625,
0.476318359375,
0.2342529296875,
0.1070556640625,
0.2342529296875
],
[
0.52294921875,
0.2259521484375,
0.90673828125,
0.2259521484375,
0.90673828125,
0.2392578125,
0.52294921875,
0.2392578125
],
[
0.09234619140625,
0.23486328125,
0.476318359375,
0.23486328125,
0.476318359375,
0.248779296875,
0.09234619140625,
0.248779296875
],
[
0.52294921875,
0.240478515625,
0.90673828125,
0.240478515625,
0.90673828125,
0.25390625,
0.52294921875,
0.25390625
],
[
0.09234619140625,
0.250732421875,
0.4755859375,
0.250732421875,
0.4755859375,
0.264404296875,
0.09234619140625,
0.264404296875
],
[
0.52392578125,
0.255126953125,
0.90869140625,
0.255126953125,
0.90869140625,
0.268310546875,
0.52392578125,
0.268310546875
],
[
0.0931396484375,
0.266357421875,
0.4755859375,
0.266357421875,
0.4755859375,
0.27978515625,
0.0931396484375,
0.27978515625
],
[
0.52294921875,
0.270751953125,
0.6435546875,
0.270751953125,
0.6435546875,
0.2841796875,
0.52294921875,
0.2841796875
],
[
0.096435546875,
0.2822265625,
0.4755859375,
0.2822265625,
0.4755859375,
0.292236328125,
0.096435546875,
0.292236328125
],
[
0.0931396484375,
0.29541015625,
0.4755859375,
0.29541015625,
0.4755859375,
0.309326171875,
0.0931396484375,
0.309326171875
],
[
0.52392578125,
0.306884765625,
0.57275390625,
0.306884765625,
0.57275390625,
0.318115234375,
0.52392578125,
0.318115234375
],
[
0.56640625,
0.30615234375,
0.8203125,
0.30615234375,
0.8203125,
0.32080078125,
0.56640625,
0.32080078125
],
[
0.09234619140625,
0.326904296875,
0.125,
0.326904296875,
0.125,
0.33837890625,
0.09234619140625,
0.33837890625
],
[
0.09234619140625,
0.311279296875,
0.476318359375,
0.311279296875,
0.476318359375,
0.324462890625,
0.09234619140625,
0.324462890625
],
[
0.52392578125,
0.33203125,
0.90673828125,
0.33203125,
0.90673828125,
0.34521484375,
0.52392578125,
0.34521484375
],
[
0.1070556640625,
0.34033203125,
0.4755859375,
0.341552734375,
0.4755859375,
0.35595703125,
0.1070556640625,
0.354736328125
],
[
0.521484375,
0.3466796875,
0.90673828125,
0.34521484375,
0.90673828125,
0.35986328125,
0.521484375,
0.361083984375
],
[
0.10540771484375,
0.35791015625,
0.473876953125,
0.35791015625,
0.473876953125,
0.36865234375,
0.10540771484375,
0.36865234375
],
[
0.52294921875,
0.361083984375,
0.82177734375,
0.361083984375,
0.82177734375,
0.375,
0.52294921875,
0.375
],
[
0.09234619140625,
0.371826171875,
0.476318359375,
0.371826171875,
0.476318359375,
0.385009765625,
0.09234619140625,
0.385009765625
],
[
0.5390625,
0.37744140625,
0.904296875,
0.37744140625,
0.904296875,
0.391357421875,
0.5390625,
0.391357421875
],
[
0.09234619140625,
0.386474609375,
0.4755859375,
0.386474609375,
0.4755859375,
0.399658203125,
0.09234619140625,
0.399658203125
],
[
0.52294921875,
0.393310546875,
0.90673828125,
0.393310546875,
0.90673828125,
0.4072265625,
0.52294921875,
0.4072265625
],
[
0.09234619140625,
0.401611328125,
0.476318359375,
0.401611328125,
0.476318359375,
0.415283203125,
0.09234619140625,
0.415283203125
],
[
0.52392578125,
0.40771484375,
0.90869140625,
0.40771484375,
0.90869140625,
0.421630859375,
0.52392578125,
0.421630859375
],
[
0.0906982421875,
0.415283203125,
0.4755859375,
0.416748046875,
0.4755859375,
0.429931640625,
0.0906982421875,
0.4287109375
],
[
0.52294921875,
0.42236328125,
0.90673828125,
0.423583984375,
0.90673828125,
0.4375,
0.52294921875,
0.436279296875
],
[
0.09234619140625,
0.431884765625,
0.476318359375,
0.431884765625,
0.476318359375,
0.44580078125,
0.09234619140625,
0.44580078125
],
[
0.52294921875,
0.4375,
0.90673828125,
0.4375,
0.90673828125,
0.451904296875,
0.52294921875,
0.451904296875
],
[
0.0931396484375,
0.4462890625,
0.4755859375,
0.4462890625,
0.4755859375,
0.4609375,
0.0931396484375,
0.4609375
],
[
0.521484375,
0.451904296875,
0.90673828125,
0.45263671875,
0.90673828125,
0.466552734375,
0.521484375,
0.46533203125
],
[
0.0931396484375,
0.462158203125,
0.476318359375,
0.462158203125,
0.476318359375,
0.47607421875,
0.0931396484375,
0.47607421875
],
[
0.52392578125,
0.468505859375,
0.90673828125,
0.468505859375,
0.90673828125,
0.481689453125,
0.52392578125,
0.481689453125
],
[
0.0931396484375,
0.479248046875,
0.4755859375,
0.479248046875,
0.4755859375,
0.4892578125,
0.0931396484375,
0.4892578125
],
[
0.52392578125,
0.48291015625,
0.90673828125,
0.48291015625,
0.90673828125,
0.496826171875,
0.52392578125,
0.496826171875
],
[
0.09234619140625,
0.492431640625,
0.476318359375,
0.492431640625,
0.476318359375,
0.505859375,
0.09234619140625,
0.505859375
],
[
0.52294921875,
0.5,
0.779296875,
0.5,
0.779296875,
0.51025390625,
0.52294921875,
0.51025390625
],
[
0.09234619140625,
0.50830078125,
0.473876953125,
0.50830078125,
0.473876953125,
0.521484375,
0.09234619140625,
0.521484375
],
[
0.5390625,
0.5146484375,
0.90673828125,
0.5146484375,
0.90673828125,
0.52783203125,
0.5390625,
0.52783203125
],
[
0.0906982421875,
0.521484375,
0.476318359375,
0.5234375,
0.476318359375,
0.537109375,
0.0906982421875,
0.53515625
],
[
0.52294921875,
0.52978515625,
0.87744140625,
0.52978515625,
0.87744140625,
0.54345703125,
0.52294921875,
0.54345703125
],
[
0.09234619140625,
0.53857421875,
0.476318359375,
0.53857421875,
0.476318359375,
0.5517578125,
0.09234619140625,
0.5517578125
],
[
0.0931396484375,
0.5537109375,
0.473876953125,
0.5537109375,
0.473876953125,
0.564453125,
0.0931396484375,
0.564453125
],
[
0.0906982421875,
0.5673828125,
0.4755859375,
0.56640625,
0.4755859375,
0.58056640625,
0.0906982421875,
0.58203125
],
[
0.529296875,
0.5693359375,
0.54150390625,
0.5791015625,
0.529296875,
0.58837890625,
0.51708984375,
0.5791015625
],
[
0.53515625,
0.57568359375,
0.560546875,
0.57568359375,
0.560546875,
0.58349609375,
0.53515625,
0.58349609375
],
[
0.5595703125,
0.5712890625,
0.68701171875,
0.5712890625,
0.68701171875,
0.5888671875,
0.5595703125,
0.5888671875
],
[
0.09234619140625,
0.58349609375,
0.473876953125,
0.58349609375,
0.473876953125,
0.5966796875,
0.09234619140625,
0.5966796875
],
[
0.09234619140625,
0.59765625,
0.20751953125,
0.59765625,
0.20751953125,
0.611328125,
0.09234619140625,
0.611328125
],
[
0.521484375,
0.60498046875,
0.56884765625,
0.60498046875,
0.56884765625,
0.619140625,
0.521484375,
0.619140625
],
[
0.56884765625,
0.60498046875,
0.66650390625,
0.60498046875,
0.66650390625,
0.619140625,
0.56884765625,
0.619140625
],
[
0.1070556640625,
0.61279296875,
0.4755859375,
0.61279296875,
0.4755859375,
0.626953125,
0.1070556640625,
0.626953125
],
[
0.09234619140625,
0.62890625,
0.476318359375,
0.62890625,
0.476318359375,
0.642578125,
0.09234619140625,
0.642578125
],
[
0.51953125,
0.62744140625,
0.90869140625,
0.62890625,
0.90869140625,
0.646484375,
0.51953125,
0.64599609375
],
[
0.0931396484375,
0.64599609375,
0.473876953125,
0.64599609375,
0.473876953125,
0.65576171875,
0.0931396484375,
0.65576171875
],
[
0.521484375,
0.64453125,
0.90625,
0.64599609375,
0.90625,
0.66015625,
0.521484375,
0.6591796875
],
[
0.09478759765625,
0.6591796875,
0.4755859375,
0.6591796875,
0.4755859375,
0.67236328125,
0.09478759765625,
0.67236328125
],
[
0.52294921875,
0.66357421875,
0.65185546875,
0.66357421875,
0.65185546875,
0.673828125,
0.52294921875,
0.673828125
],
[
0.09234619140625,
0.67236328125,
0.476318359375,
0.673828125,
0.476318359375,
0.6875,
0.09234619140625,
0.68603515625
],
[
0.0906982421875,
0.689453125,
0.47802734375,
0.689453125,
0.47802734375,
0.70263671875,
0.0906982421875,
0.70263671875
],
[
0.69287109375,
0.68798828125,
0.70751953125,
0.68798828125,
0.70751953125,
0.69873046875,
0.69287109375,
0.69873046875
],
[
0.611328125,
0.69970703125,
0.6953125,
0.69970703125,
0.6953125,
0.71728515625,
0.611328125,
0.71728515625
],
[
0.68310546875,
0.69873046875,
0.82275390625,
0.69873046875,
0.82275390625,
0.7197265625,
0.68310546875,
0.7197265625
],
[
0.8857421875,
0.7021484375,
0.90869140625,
0.7021484375,
0.90869140625,
0.71533203125,
0.8857421875,
0.71533203125
],
[
0.68701171875,
0.71728515625,
0.71337890625,
0.71728515625,
0.71337890625,
0.72998046875,
0.68701171875,
0.72998046875
],
[
0.09234619140625,
0.7041015625,
0.4755859375,
0.7041015625,
0.4755859375,
0.71728515625,
0.09234619140625,
0.71728515625
],
[
0.09234619140625,
0.7197265625,
0.4755859375,
0.7197265625,
0.4755859375,
0.73291015625,
0.09234619140625,
0.73291015625
],
[
0.09234619140625,
0.734375,
0.4755859375,
0.734375,
0.4755859375,
0.74755859375,
0.09234619140625,
0.74755859375
],
[
0.09234619140625,
0.75,
0.476318359375,
0.75,
0.476318359375,
0.76318359375,
0.09234619140625,
0.76318359375
],
[
0.52294921875,
0.744140625,
0.90625,
0.744140625,
0.90625,
0.75830078125,
0.52294921875,
0.75830078125
],
[
0.0931396484375,
0.7666015625,
0.476318359375,
0.7666015625,
0.476318359375,
0.7763671875,
0.0931396484375,
0.7763671875
],
[
0.52392578125,
0.76123046875,
0.90625,
0.76123046875,
0.90625,
0.771484375,
0.52392578125,
0.771484375
],
[
0.09234619140625,
0.77978515625,
0.476318359375,
0.77978515625,
0.476318359375,
0.79345703125,
0.09234619140625,
0.79345703125
],
[
0.52294921875,
0.7744140625,
0.90673828125,
0.7744140625,
0.90673828125,
0.78857421875,
0.52294921875,
0.78857421875
],
[
0.09234619140625,
0.79541015625,
0.476318359375,
0.79541015625,
0.476318359375,
0.8095703125,
0.09234619140625,
0.8095703125
],
[
0.52392578125,
0.79150390625,
0.53759765625,
0.79150390625,
0.53759765625,
0.8017578125,
0.52392578125,
0.8017578125
],
[
0.09234619140625,
0.80810546875,
0.1168212890625,
0.81103515625,
0.11358642578125,
0.82568359375,
0.08905029296875,
0.82275390625
],
[
0.68701171875,
0.80322265625,
0.7060546875,
0.80322265625,
0.7060546875,
0.81494140625,
0.68701171875,
0.81494140625
],
[
0.58837890625,
0.814453125,
0.84619140625,
0.814453125,
0.84619140625,
0.83544921875,
0.58837890625,
0.83544921875
],
[
0.88623046875,
0.818359375,
0.90869140625,
0.818359375,
0.90869140625,
0.83203125,
0.88623046875,
0.83203125
],
[
0.6845703125,
0.83349609375,
0.7099609375,
0.83349609375,
0.7099609375,
0.84521484375,
0.6845703125,
0.84521484375
],
[
0.08905029296875,
0.84326171875,
0.1217041015625,
0.84326171875,
0.1217041015625,
0.859375,
0.08905029296875,
0.859375
],
[
0.1168212890625,
0.8466796875,
0.1478271484375,
0.8466796875,
0.1478271484375,
0.85498046875,
0.1168212890625,
0.85498046875
],
[
0.1422119140625,
0.8447265625,
0.47802734375,
0.8447265625,
0.47802734375,
0.859375,
0.1422119140625,
0.859375
],
[
0.53515625,
0.85595703125,
0.90869140625,
0.85693359375,
0.90869140625,
0.8720703125,
0.53515625,
0.87060546875
],
[
0.1356201171875,
0.86279296875,
0.256591796875,
0.86279296875,
0.256591796875,
0.87744140625,
0.1356201171875,
0.87744140625
],
[
0.53369140625,
0.8623046875,
0.53759765625,
0.8623046875,
0.53759765625,
0.8671875,
0.53369140625,
0.8671875
],
[
0.52392578125,
0.87353515625,
0.8388671875,
0.87353515625,
0.8388671875,
0.8837890625,
0.52392578125,
0.8837890625
],
[
0.09234619140625,
0.88818359375,
0.476318359375,
0.88818359375,
0.476318359375,
0.9013671875,
0.09234619140625,
0.9013671875
],
[
0.53759765625,
0.88720703125,
0.90673828125,
0.88720703125,
0.90673828125,
0.90087890625,
0.53759765625,
0.90087890625
],
[
0.0931396484375,
0.90283203125,
0.4755859375,
0.90283203125,
0.4755859375,
0.91650390625,
0.0931396484375,
0.91650390625
],
[
0.52294921875,
0.90283203125,
0.90673828125,
0.90283203125,
0.90673828125,
0.91650390625,
0.52294921875,
0.91650390625
],
[
0.09234619140625,
0.91845703125,
0.4755859375,
0.91845703125,
0.4755859375,
0.931640625,
0.09234619140625,
0.931640625
],
[
0.52392578125,
0.91943359375,
0.82177734375,
0.91943359375,
0.82177734375,
0.9296875,
0.52392578125,
0.9296875
],
[
0.49267578125,
0.95556640625,
0.50634765625,
0.95556640625,
0.50634765625,
0.96728515625,
0.49267578125,
0.96728515625
]
] |
[
"2Related work",
"保形空间（Lietal.2001b）中的表示来构造超球面决策",
"面。Banareretal.（2003b）提出的超球面神经元就是这",
"Steerability and equivariance",
"样一个球面分类器。值得注意的是，由于可以将超球",
"体视为超平面的概括，因此可以将标准神经元视为超",
"2.1",
"可控性是来自早期视觉和图像处理的强大概念（Free",
"球体神经元的特殊情况。",
"man et al., 1991; Knutsson et al., 1992; Simoncelli et al..",
"在前馈网络中堆叠多个超球面神经元会产生多层超",
"1992;Perona, 1995; Simoncelli & Freeman, 1995; Teo &",
"球面感知器（MLHP），Banareretal.（2003a）显示该感",
"Hel-Or，1998）在深度学习时代引起共鸣。在Freeman",
"知器在某些分类任务中优于标准MLP。然而，它在点",
"etal.（1991)的开创性工作中介绍了可操纵滤波器的实",
"集上的应用没有被讨论。。这激发了Melnyketal.(2021)",
"用性及其构造的主要定理",
"在几何神经元上的工作，其中学习的参数显示为表示",
"在过去的几十年中，计算机视觉环境中的几何等方",
"球形决策面的组合。此外，几何（和超球面）神经元激",
"差一直是广泛研究的对象（VanGooletal.,1995）。等方",
"活被证明在3D欧几里得空间中是等距的。。在我们的",
"差是可控制性的必要属性，因为可控制性需要根据特",
"工作中，我们使用后一个观察结果作为推导可操纵性",
"定组的操作更改函数输出。因此，输出需要包含作用",
"约束的必要条件。",
"于输入的组的表示。因此，操作员必须与组通勤（请注",
"意，表示可能会从输入更改为输出）。因此，Lie理论",
"2.3",
"Comparison to other methods",
"波器",
"可用于研究等方差和不变性(Reisert，2008)的可操纵滤",
"可操纵性要求滤波器要么是球谐函数（的组合）（参见",
"如今，与等方差相关的研究已扩展到深度学习，e.g",
"e.g.，，Fuchsetal.（2020）），要么是基于学习的神经元",
"Fuchs et al.(2020)Thomas et al.(2018)和 Zhaoetal",
"构造的，并且表现得像这样（我们的工作）",
"(2020)的SE(3）等变模型，以及Andersonetal.(2019)",
"将我们的方法与其他等变网络区分开来的关键点是",
"的SO(3）等变网络，以及Marcosetal.(2017)和Kondor",
"我们不限制可学习参数的空间（与e.g，，Thomasetal.",
"etal.（2018）。在当前的工作中也越来越多地使用可操",
"(2018)，Weiler et al.(2018a)和Fuchs et al.(2020)相反）,",
"纵的过滤器概念。例如，Cohen&Welling（2016）的工",
"而是从自由训练的基础网络构建我们的可控模型，正",
"作考虑了图像数据，并提出了一种CNN架构，以产生",
"如我们在E4节中详细讨论的那样。也就是说，相关工",
"具有可操纵特征的等变表示，与传统CNN相比，它涉",
"作使用球谐函数作为原子，i.e.，是手工设计的基础，并",
"及的参数更少。他们考虑了二面角组（离散旋转，平",
"且仅学习约束下的线性系数。相反，我们从手部设计",
"移和反射），并提出了工作中的可操纵表示作为基本特",
"中继承的唯一东西是一阶谐波的约束（请参见定理E",
"征类型的形成。他们的方法的一个局限性是旋转被限",
"4.1）-所有其他自由度都是可学习的",
"制在四个方向上，i.e.，，乘以元/2。最近，Weileretal",
"此外，即使我们在当前工作中仅考虑标量场，也可",
"(2018b）利用组卷积并引人了对图像进行操作的可控滤",
"以将E4节的理论结果应用于更广泛的特征场类别",
"波器卷积神经网络（SFCNN），以在平移和离散旋转下",
"共同获得等方差。在他们的工作中，滤波器组是学习",
"的，而不是固定的。此外，Weiler&Cesa（2019)的工作",
"3",
"",
"Background",
"提出了E（2）-等变可控CNN的统一框架，并介绍了",
"它们的一般理论",
"3.1",
"Steerability",
"Weileretal.（2018a）提出的用于3D数据的可操纵",
"CNN与我们的工作密切相关。作者采用标量，矢量和",
"根据Freemanetal.(1991），如果可以将2D函数f(r，g)",
"张量场的组合作为由SO（3）表示转换的特征，并提出",
"编写为自身旋转版本的线性组合，则可以将其转向。",
"了一个与SE(3)转换等变的模型。他们还考虑了适用",
"e。，当它满足约束",
"于特征空间的非标量分量的不同类型的非线性。Fuchs",
"etal.（2020)的新型SE（3）-等变方法引人了一种自注",
"M",
"fe(c,y) =",
"Zu;(0)f (a, y),",
"(1)",
"5=1",
"意机制，该机制对其输人的全局旋转和平移不变，并",
"解决了其他等变方法e.g.，Thomasetal.(2018）中角度",
"约束滤波器的局限性。值得注意的是，Jingetal.(2020)",
"的工作提出了几何矢量感知器（GVP），该几何矢量感",
"时，其中U,(の)是插值函数，,是基函数方向，M是引",
"知器由输入标量和失量特征的两个线性变换组成：然",
"导函数所需的基函数项的数量。在Knutssonetal.（1992）",
"后是非线性。相对于3D欧几里得空间中旋转和反射的",
"的工作中可以找到另一种公式。在3D中，转向方程变",
"任意组合，GVP标量和失量输出分别是不变的和等变",
"为",
"的。",
"M",
"fR(a, y,2) =u;(R)fRi(a,y,2) ,",
"(2)",
"j=1",
"2.2",
"",
"Conformal modeling and the hyper-",
"其中fR(a,y,z)是由RESo(3)旋转的f(r,y,z)，",
"sphere neuron",
"",
"并且每个R，ESO(3）定向相应的ith基函数",
"Lietal.(2001a)彻底讨论了欧几里得几何保形嵌入的实",
"Freemanetal.(1991）中的定理1、2和4描述了可操",
"用性以及与Minkowski空间的紧密联系。一个重要的",
"纵性约束（1）和（2）成立的条件，以及如何分别",
"结果是，可以像Perwassetal.（2003）的工作一样，使用",
"确定2D和3D情况下的基函数的最小数量",
"2"
] |
{
"arxiv_id": "2106.13863",
"language": "zh",
"page_idx": 1
}
|
||
2106.13863-zh-2
|
2106.13863-zh-2.jpg
|
<FigureHere>
Figure 2: The geometric neuron $(\ell_k=-\frac{1}{2}\|x_k\|^2).$
## 3.2 $\textbf{Conformal embedding}$
我们请读者参考 Melnyk et al.(2021) 工作中的第 3 节以获取更多详细信息，并且仅在本节中简要介绍重要的符号。欧几里得$\mathbb{R}^n$对应物的共形空间可以形成为$\mathbb{ME}^n\equiv\mathbb{R}^{n+1,1}=\mathbb{R}^n\oplus\mathbb{R}^{1,1}$,其中$\mathbb{R}^1$,是 Minkowski 平面(Li et al., 2001a) , 正交基定义为$\{e_+,e_-\}$ ,空基$\left\{e_0,\:e_\infty\right\}$表示原点 $e_0= \frac 12( e_- - e_+ )$并指向无穷大$e_{\infty}=e_{-}+e_{+}$。因此$,\tilde\text{欧几里得向量 }x\in\mathbb{R}^n$可以作为
$$X=\mathcal C(x)=x+\frac{1}{2}\|x\|^2\:e_\infty+e_0\:,$$
(3)
嵌入到保形空间 M$\mathbb{E}^n$中，其中$X\in\mathbb{M}\mathbb{E}^n$被称为归一化。共形嵌入(3)表示$x$在 M$\mathbb{E}^n$中的投影球体上的立体投影，并且是齐次的，i.e., 等价类$[X]=\{Z\in$ $\mathbb{R} ^{n+ 1, 1}: Z= \gamma X$, $\gamma \in \mathbb{R} \setminus \{ 0\}$中的所有嵌入向量均表示相同的向量$x$ 。重要的是，给定保形嵌入$X$和一些$Y= y+ \frac 12\| y\| ^{2}e_{\infty }+ e_{0}$ ,它们在保形空间中的标量积对应于它们的欧几里得距离$X\cdot Y=-\frac12\|x-y\|^2$。对保形空间中标量积的这种解释是构建球形神经元的主要动机
## 3.3 $\textbf{Spherical neurons}$
由球形神经元，我们统称为超球形 (Banarer et al., 2003a) 和几何 (Melnyk et al.,2021)神经元，它们具有球形决策面
正如 Banarer et al. (2003a) 中的标准结果所讨论的那样，通过在$\mathbb{R}^n+2$中嵌人数据向量$x\in\mathbb{R}^n$和超球面$S\in\mathbb{M}\mathbb{E}^n$作为
$$X=\begin{pmatrix}x_1,\ldots,x_n,-1,-\frac12\|x\|^2\end{pmatrix}\in\mathbb{R}^{n+2},\\S=\begin{pmatrix}c_1,\ldots,c_n,\frac12(\|c\|^2-r^2),1\end{pmatrix}\in\mathbb{R}^{n+2},$$
(4)
其中$c=(c_1,\ldots,c_n)\in\mathbb{R}^n$ 是超球面中心，而 $r\in\mathbb{R}$ 是半径，它们的标量积 XMATH。在训练过程中，超球面矢量分量被视为独立的可学习参数。因此，球形分类器有效地学习$\tilde{S}=(s_1,\ldots,s_{n+2})\in\mathbb{R}^{n+2}$形式的非归一化超球面。由于表示的均匀性，归一化和非归一化的超球面都表示相同的决策面。更多细节可以在第 3节找到。2 在 Melnyk et al. (2021) 的工作
几何神经元是以点集为输入的超球面神经元的概括， 请参见图$\operatorname{E2}$。因此，单个几何神经元输出是 $K$ 输入
<FigureHere>
Figure 3: The effect of rotation on the spherical neuron activation in 2D; $t(\theta)$ denotes the tangent length as a function of rotation angle, i.e., $t( 0) ^2= \left \| x- c_{0}\right \| ^2- r^2= - 2X^{\top }S$ . 点到$K$学习的超球面
(5)
$$z=\sum\limits_{k=1}^K\gamma_k\:X_k^\top S_k\:,$$
的有符号距离之和，其中$z\in \mathbb{R}$ , $X_k\in \mathbb{R} ^5$是正确嵌人的 3D 输入点$,\gamma_k\in\mathbb{R}$是比例因子，i。e。,学习参数向量$\widetilde{S}_k$的最后一个元素，$S_k= \widetilde{S} _k/ \gamma _k\in \mathbb{R} ^5$是相应的归一化学习参数(球体)
此外，Melnyk et al.(2021)证明了超球面(和几何)神经元激活在 3D 中是等距的。也就是说，旋转输人等同于旋转决策球。此结果是考虑用球形神经元构建的模型的旋转和平移等方差的必要条件，并构成了我们方法的基础
在以下各节中，我们对欧几里得空间$\mathbb{R}^3$ ,齐次(投影)空间$P(\mathbb{R}^3)$和 M$\mathbb{E}^3\cong\mathbb{R}^5$中表示的 3D 旋转$R$使用相同的表示法，具体取决于上下文。这是可能的，因为我们可以在不改变变换表示的情况下将所需数量的1 附加到原始旋转矩阵的对角线上
### $\textbf{4}$ $\textbf{Method}$
在本节中，我们确定了可以根据其 3D 输入操纵球形神经元的条件。换句话说，我们使用学习的球形神经元参数的旋转版本，得出一个表达式，该表达式为我们提供了假设的球形神经元对某些输入的响应。。我们首先考虑单个球体的可操纵性条件，对相应的输入点$X$ i进行分类。e。$,f(X)=X^\top S$,其中$X$和$S$根据(4 )嵌入到 M$\mathbb{E}^3\cong\mathbb{R}^5$中
### 4.1 Basis construction
要为具有一个球体作为决策面的球形神经元制定可操纵性约束，我们首先需要确定基函数的最小数量 i.e.
(2)中的项数$M$ 。此数字仅取决于计算转向结果(Freeman et al., 1991) 所需的球谐度。因此，我们需要确定所需的度。
Theorem 4.1. Let S $\in\mathbb{R}^{n+2}$ be an nD spherical classifrer with center c$_0\in\mathbb{R}^n(R:=\|\mathbf{c}_0\|)and$ radius $r,and$
3
|
[
[
0.42724609375,
0.0706787109375,
0.450927734375,
0.07452392578125,
0.44775390625,
0.083984375,
0.4248046875,
0.0802001953125
],
[
0.11517333984375,
0.08837890625,
0.1871337890625,
0.09027099609375,
0.1871337890625,
0.10101318359375,
0.11517333984375,
0.0985107421875
],
[
0.2132568359375,
0.0921630859375,
0.2442626953125,
0.0921630859375,
0.2442626953125,
0.09661865234375,
0.2132568359375,
0.09661865234375
],
[
0.38818359375,
0.09027099609375,
0.451904296875,
0.09027099609375,
0.451904296875,
0.10101318359375,
0.38818359375,
0.10101318359375
],
[
0.289306640625,
0.1256103515625,
0.301513671875,
0.1256103515625,
0.301513671875,
0.1363525390625,
0.289306640625,
0.1363525390625
],
[
0.64453125,
0.1307373046875,
0.65869140625,
0.1307373046875,
0.65869140625,
0.141357421875,
0.64453125,
0.141357421875
],
[
0.15283203125,
0.141357421875,
0.1707763671875,
0.141357421875,
0.1707763671875,
0.1546630859375,
0.15283203125,
0.1546630859375
],
[
0.39794921875,
0.143310546875,
0.410888671875,
0.143310546875,
0.410888671875,
0.1546630859375,
0.39794921875,
0.1546630859375
],
[
0.271240234375,
0.1666259765625,
0.294189453125,
0.1666259765625,
0.294189453125,
0.186279296875,
0.271240234375,
0.186279296875
],
[
0.583984375,
0.17236328125,
0.60888671875,
0.17236328125,
0.60888671875,
0.1849365234375,
0.583984375,
0.1849365234375
],
[
0.71826171875,
0.1666259765625,
0.744140625,
0.1666259765625,
0.744140625,
0.179931640625,
0.71826171875,
0.179931640625
],
[
0.66015625,
0.18310546875,
0.6708984375,
0.179931640625,
0.67333984375,
0.1849365234375,
0.6611328125,
0.1875
],
[
0.640625,
0.1988525390625,
0.646484375,
0.1988525390625,
0.646484375,
0.203857421875,
0.640625,
0.203857421875
],
[
0.6689453125,
0.2203369140625,
0.69775390625,
0.2203369140625,
0.69775390625,
0.23291015625,
0.6689453125,
0.23291015625
],
[
0.1094970703125,
0.2298583984375,
0.450927734375,
0.2298583984375,
0.450927734375,
0.2442626953125,
0.1094970703125,
0.2442626953125
],
[
0.08905029296875,
0.2587890625,
0.330078125,
0.261474609375,
0.330078125,
0.27587890625,
0.08905029296875,
0.2734375
],
[
0.521484375,
0.261962890625,
0.90869140625,
0.261962890625,
0.90869140625,
0.276611328125,
0.521484375,
0.276611328125
],
[
0.521484375,
0.279052734375,
0.90673828125,
0.279052734375,
0.90673828125,
0.292236328125,
0.521484375,
0.292236328125
],
[
0.09234619140625,
0.285400390625,
0.473876953125,
0.285400390625,
0.473876953125,
0.298583984375,
0.09234619140625,
0.298583984375
],
[
0.51953125,
0.29345703125,
0.90673828125,
0.292236328125,
0.90673828125,
0.306884765625,
0.51953125,
0.30810546875
],
[
0.0931396484375,
0.300537109375,
0.476318359375,
0.300537109375,
0.476318359375,
0.314453125,
0.0931396484375,
0.314453125
],
[
0.09234619140625,
0.31640625,
0.476318359375,
0.31640625,
0.476318359375,
0.330078125,
0.09234619140625,
0.330078125
],
[
0.0882568359375,
0.326904296875,
0.47802734375,
0.328857421875,
0.47802734375,
0.34521484375,
0.0882568359375,
0.343505859375
],
[
0.52294921875,
0.32763671875,
0.67626953125,
0.32763671875,
0.67626953125,
0.341552734375,
0.52294921875,
0.341552734375
],
[
0.0931396484375,
0.34521484375,
0.47802734375,
0.34521484375,
0.47802734375,
0.35986328125,
0.0931396484375,
0.35986328125
],
[
0.68115234375,
0.35107421875,
0.70166015625,
0.35107421875,
0.70166015625,
0.3642578125,
0.68115234375,
0.3642578125
],
[
0.09234619140625,
0.361083984375,
0.476318359375,
0.361083984375,
0.476318359375,
0.375732421875,
0.09234619140625,
0.375732421875
],
[
0.67626953125,
0.3623046875,
0.78515625,
0.3623046875,
0.78515625,
0.380126953125,
0.67626953125,
0.380126953125
],
[
0.8857421875,
0.3642578125,
0.91015625,
0.3642578125,
0.91015625,
0.37744140625,
0.8857421875,
0.37744140625
],
[
0.67626953125,
0.375732421875,
0.70751953125,
0.375732421875,
0.70751953125,
0.392578125,
0.67626953125,
0.392578125
],
[
0.65185546875,
0.367431640625,
0.6806640625,
0.367431640625,
0.6806640625,
0.375,
0.65185546875,
0.375
],
[
0.08905029296875,
0.376953125,
0.476318359375,
0.375,
0.476318359375,
0.38818359375,
0.08905029296875,
0.39013671875
],
[
0.28759765625,
0.40771484375,
0.404296875,
0.4072265625,
0.404296875,
0.425537109375,
0.28759765625,
0.4267578125
],
[
0.52294921875,
0.40283203125,
0.90625,
0.40283203125,
0.90625,
0.416748046875,
0.52294921875,
0.416748046875
],
[
0.1650390625,
0.4091796875,
0.285888671875,
0.410400390625,
0.285888671875,
0.4248046875,
0.1650390625,
0.423583984375
],
[
0.455810546875,
0.410400390625,
0.47802734375,
0.410400390625,
0.47802734375,
0.423583984375,
0.455810546875,
0.423583984375
],
[
0.52294921875,
0.417236328125,
0.90673828125,
0.417236328125,
0.90673828125,
0.431884765625,
0.52294921875,
0.431884765625
],
[
0.52294921875,
0.43310546875,
0.90869140625,
0.43310546875,
0.90869140625,
0.4462890625,
0.52294921875,
0.4462890625
],
[
0.09234619140625,
0.44140625,
0.476318359375,
0.44140625,
0.476318359375,
0.455810546875,
0.09234619140625,
0.455810546875
],
[
0.53759765625,
0.447509765625,
0.67724609375,
0.447509765625,
0.67724609375,
0.4609375,
0.53759765625,
0.4609375
],
[
0.1070556640625,
0.45703125,
0.476318359375,
0.45703125,
0.476318359375,
0.47021484375,
0.1070556640625,
0.47021484375
],
[
0.5390625,
0.462158203125,
0.90673828125,
0.462158203125,
0.90673828125,
0.4765625,
0.5390625,
0.4765625
],
[
0.0931396484375,
0.472900390625,
0.47802734375,
0.472900390625,
0.47802734375,
0.486083984375,
0.0931396484375,
0.486083984375
],
[
0.521484375,
0.477783203125,
0.90673828125,
0.477783203125,
0.90673828125,
0.4912109375,
0.521484375,
0.4912109375
],
[
0.08905029296875,
0.48486328125,
0.4794921875,
0.4873046875,
0.4794921875,
0.5048828125,
0.08905029296875,
0.50244140625
],
[
0.52392578125,
0.492431640625,
0.90673828125,
0.492431640625,
0.90673828125,
0.505859375,
0.52392578125,
0.505859375
],
[
0.09234619140625,
0.50390625,
0.476318359375,
0.50390625,
0.476318359375,
0.517578125,
0.09234619140625,
0.517578125
],
[
0.521484375,
0.5068359375,
0.90673828125,
0.50830078125,
0.90673828125,
0.52294921875,
0.521484375,
0.521484375
],
[
0.0906982421875,
0.517578125,
0.47802734375,
0.51806640625,
0.47802734375,
0.53271484375,
0.0906982421875,
0.5322265625
],
[
0.52294921875,
0.5244140625,
0.5888671875,
0.5244140625,
0.5888671875,
0.5361328125,
0.52294921875,
0.5361328125
],
[
0.0906982421875,
0.53271484375,
0.476318359375,
0.5341796875,
0.476318359375,
0.548828125,
0.0906982421875,
0.54736328125
],
[
0.53759765625,
0.53857421875,
0.90673828125,
0.53857421875,
0.90673828125,
0.5517578125,
0.53759765625,
0.5517578125
],
[
0.09234619140625,
0.5498046875,
0.476318359375,
0.5498046875,
0.476318359375,
0.56298828125,
0.09234619140625,
0.56298828125
],
[
0.52294921875,
0.55322265625,
0.90673828125,
0.55322265625,
0.90673828125,
0.56640625,
0.52294921875,
0.56640625
],
[
0.09234619140625,
0.56640625,
0.1422119140625,
0.56640625,
0.1422119140625,
0.57763671875,
0.09234619140625,
0.57763671875
],
[
0.52392578125,
0.568359375,
0.90673828125,
0.568359375,
0.90673828125,
0.58203125,
0.52392578125,
0.58203125
],
[
0.52294921875,
0.583984375,
0.90673828125,
0.583984375,
0.90673828125,
0.59765625,
0.52294921875,
0.59765625
],
[
0.0906982421875,
0.599609375,
0.1396484375,
0.599609375,
0.1396484375,
0.6142578125,
0.0906982421875,
0.6142578125
],
[
0.136474609375,
0.60107421875,
0.293212890625,
0.60107421875,
0.293212890625,
0.6142578125,
0.136474609375,
0.6142578125
],
[
0.52294921875,
0.599609375,
0.76123046875,
0.599609375,
0.76123046875,
0.60986328125,
0.52294921875,
0.60986328125
],
[
0.0931396484375,
0.62548828125,
0.476318359375,
0.62548828125,
0.476318359375,
0.6396484375,
0.0931396484375,
0.6396484375
],
[
0.09234619140625,
0.64013671875,
0.476318359375,
0.64013671875,
0.476318359375,
0.65380859375,
0.09234619140625,
0.65380859375
],
[
0.5185546875,
0.63525390625,
0.64208984375,
0.63720703125,
0.64208984375,
0.65478515625,
0.5185546875,
0.65283203125
],
[
0.09234619140625,
0.65576171875,
0.125,
0.65576171875,
0.125,
0.66845703125,
0.09234619140625,
0.66845703125
],
[
0.108642578125,
0.67041015625,
0.47802734375,
0.67041015625,
0.47802734375,
0.6845703125,
0.108642578125,
0.6845703125
],
[
0.52392578125,
0.67041015625,
0.904296875,
0.67041015625,
0.904296875,
0.68115234375,
0.52392578125,
0.68115234375
],
[
0.09234619140625,
0.6845703125,
0.476318359375,
0.6845703125,
0.476318359375,
0.69873046875,
0.09234619140625,
0.69873046875
],
[
0.521484375,
0.6845703125,
0.90673828125,
0.6845703125,
0.90673828125,
0.69775390625,
0.521484375,
0.69775390625
],
[
0.08905029296875,
0.69873046875,
0.1968994140625,
0.69970703125,
0.1968994140625,
0.7138671875,
0.08905029296875,
0.71337890625
],
[
0.52392578125,
0.69970703125,
0.90625,
0.69970703125,
0.90625,
0.71337890625,
0.52392578125,
0.71337890625
],
[
0.52294921875,
0.7138671875,
0.90673828125,
0.7138671875,
0.90673828125,
0.72802734375,
0.52294921875,
0.72802734375
],
[
0.134033203125,
0.73095703125,
0.4150390625,
0.7265625,
0.4150390625,
0.751953125,
0.134033203125,
0.755859375
],
[
0.52392578125,
0.72998046875,
0.904296875,
0.72998046875,
0.904296875,
0.74365234375,
0.52392578125,
0.74365234375
],
[
0.455810546875,
0.74853515625,
0.47802734375,
0.74853515625,
0.47802734375,
0.76123046875,
0.455810546875,
0.76123046875
],
[
0.51953125,
0.74365234375,
0.90869140625,
0.7421875,
0.90869140625,
0.7568359375,
0.51953125,
0.75830078125
],
[
0.136474609375,
0.76123046875,
0.431396484375,
0.7568359375,
0.431396484375,
0.7822265625,
0.136474609375,
0.7861328125
],
[
0.521484375,
0.75830078125,
0.68896484375,
0.75830078125,
0.68896484375,
0.771484375,
0.521484375,
0.771484375
],
[
0.10540771484375,
0.79541015625,
0.47802734375,
0.79443359375,
0.47802734375,
0.8095703125,
0.10540771484375,
0.81005859375
],
[
0.521484375,
0.79345703125,
0.7255859375,
0.79345703125,
0.7255859375,
0.80810546875,
0.521484375,
0.80810546875
],
[
0.09234619140625,
0.81103515625,
0.47802734375,
0.81103515625,
0.47802734375,
0.82470703125,
0.09234619140625,
0.82470703125
],
[
0.52294921875,
0.818359375,
0.90673828125,
0.818359375,
0.90673828125,
0.83203125,
0.52294921875,
0.83203125
],
[
0.09234619140625,
0.8271484375,
0.476318359375,
0.8271484375,
0.476318359375,
0.84033203125,
0.09234619140625,
0.84033203125
],
[
0.52294921875,
0.833984375,
0.904296875,
0.833984375,
0.904296875,
0.84765625,
0.52294921875,
0.84765625
],
[
0.0906982421875,
0.84033203125,
0.294189453125,
0.84130859375,
0.294189453125,
0.85595703125,
0.0906982421875,
0.85498046875
],
[
0.28759765625,
0.84130859375,
0.47802734375,
0.84130859375,
0.47802734375,
0.85595703125,
0.28759765625,
0.85595703125
],
[
0.5390625,
0.84765625,
0.90625,
0.8486328125,
0.90625,
0.8623046875,
0.5390625,
0.861328125
],
[
0.09234619140625,
0.85693359375,
0.4697265625,
0.85693359375,
0.4697265625,
0.87060546875,
0.09234619140625,
0.87060546875
],
[
0.51953125,
0.8623046875,
0.90673828125,
0.86279296875,
0.90673828125,
0.87744140625,
0.51953125,
0.876953125
],
[
0.0931396484375,
0.87255859375,
0.4794921875,
0.87255859375,
0.4794921875,
0.88623046875,
0.0931396484375,
0.88623046875
],
[
0.52392578125,
0.880859375,
0.625,
0.880859375,
0.625,
0.892578125,
0.52392578125,
0.892578125
],
[
0.0931396484375,
0.88720703125,
0.37744140625,
0.88720703125,
0.37744140625,
0.90087890625,
0.0931396484375,
0.90087890625
],
[
0.108642578125,
0.9013671875,
0.473876953125,
0.90283203125,
0.473876953125,
0.91650390625,
0.108642578125,
0.91552734375
],
[
0.51953125,
0.90087890625,
0.90673828125,
0.9013671875,
0.90673828125,
0.91650390625,
0.51953125,
0.91552734375
],
[
0.09234619140625,
0.91748046875,
0.476318359375,
0.91748046875,
0.476318359375,
0.93115234375,
0.09234619140625,
0.93115234375
],
[
0.51953125,
0.91650390625,
0.90869140625,
0.91748046875,
0.90869140625,
0.931640625,
0.51953125,
0.93115234375
],
[
0.491943359375,
0.9541015625,
0.50732421875,
0.9541015625,
0.50732421875,
0.96826171875,
0.491943359375,
0.96826171875
]
] |
[
"1l",
"Q000",
"",
"000e8",
"s.",
"R",
"S1",
"Sh",
"Z",
"Te",
"t(0)",
"",
"X",
"t(0)",
"Figure 2: The geometric neuron (lk =-lkl2)",
"3.2Conformal embedding",
"Figure 3:The effect of rotation on the spherical neuron ac-",
"tivation in 2D; t(0) denotes the tangent length as a function",
"我们请读者参考Melnyketal.(2021）工作中的第3节",
"of rotation angle,i.e.,t(o)2 =- col2r2=-2xTs.",
"以获取更多详细信息，并且仅在本节中简要介绍重要",
"的符号。欧几里得R\"对应物的共形空间可以形成为",
"MEn=IRn+i,1=R\"Ri,1，其中RI,是Minkowski",
"点到K学习的超球面",
"平面（Lietal.,2001a)，正交基定义为{e+，e-}，空",
"K",
"基{eo，e表示原点eo=号(e_-e+)并指向无穷大",
"Eh Xt Sh.",
"(5)",
"",
"Z=",
"e。=e_+e+。因此，欧几单得向量ERn可以作为",
"l1 eo + e0 ,",
"的有符号距离之和，其中ER，XER5是正确嵌人",
"X=c()=+",
"(3)",
"的3D输人点，kER是比例因子，i。e。，学习参数向",
"量Sk的最后一个元素，Sk=Sk/kER5是相应的归",
"嵌人到保形空间MEn中，其中XEME\"被称为归",
"化学习参数（球体）",
"化。共形嵌人（3）表示&在MEn中的投影球体上的",
"此外，Melnyketal.(2021）证明了超球面（和几何）神",
"立体投影，并且是齐次的，i.e.，，等价类[X=ZE",
"经元激活在3D中是等距的。也就是说，旋转输人等同",
"Rn+1,1：Z=X，ER|[0]}中的所有嵌人向量均表",
"于旋转决策球。此结果是考虑用球形神经元构建的模",
"示相同的向量。重要的是，给定保形嵌人X和一些",
"型的旋转和平移等方差的必要条件，并构成了我们方",
"Y=y+号llyl2eα+eo，它们在保形空间中的标量积",
"法的基础",
"对应于它们的欧几里得距离X·Y=-la-y2。对",
"在以下各节中，我们对欧几里得空间R3，齐次（投",
"保形空间中标量积的这种解释是构建球形神经元的主",
"影）空间P（R3）和ME3=R5中表示的3D旋转R使",
"要动机",
"用相同的表示法，具体取决于上下文。这是可能的，因",
"为我们可以在不改变变换表示的情况下将所需数量的",
"3.3",
" Spherical neurons",
"1附加到原始旋转矩阵的对角线上",
"由球形神经元，我们统称为超球形(Banareretal.,2003a)",
"和几何（Melnyketal,2021)神经元，它们具有球形决",
"4 Method",
"策面",
"正如Banareretal.（2003a）中的标准结果所讨论的那",
"在本节中，我们确定了可以根据其3D输入操纵球形神",
"样，通过在Rn+2中嵌人数据向量Rn和超球面",
"经元的条件。换句话说，我们使用学习的球形神经元",
"SEME\"作为",
"参数的旋转版本，得出一个表达式，该表达式为我们",
"提供了假设的球形神经元对某些输入的响应。。我们首",
"X= (1,.,an, 1, --ll) e Rn+2,",
"先考虑单个球体的可操纵性条件，对相应的输入点X",
"(4)",
"i进行分类。e。，f(X）=XTS，其中X和S根据（4",
"S= (c1,..,Cn,(l? -- 2),1) e R\"+2",
"）嵌人到ME3=R5中",
"其中c=（c..,Cn）ERn是超球面中心，而rER",
"4.1Basis construction",
"是半径，它们的标量积XMATH。在训练过程中，超球",
"要为具有一个球体作为决策面的球形神经元制定可操",
"面矢量分量被视为独立的可学习参数。因此，球形分",
"纵性约束，我们首先需要确定基函数的最小数量i.e.",
"类器有效地学习S=（s1，..",
".,sn+2）ERn+2形式的非",
"（2）中的项数M。此数字仅取决于计算转向结果",
"归一化超球面。由于表示的均匀性，归一化和非归一",
"(Freemanetal.，1991）所需的球谐度。因此，我们需要",
"化的超球面都表示相同的决策面。更多细节可以在第3",
"确定所需的度。",
"节找到。2在Melnyketal.(2021)的工作",
"几何神经元是以点集为输入的超球面神经元的概括",
"Theorem 4.1. Let S e Rn+2 be an nD spherical classifier",
"请参见图E2。因此，单个几何神经元输出是K输人",
"with center co E Rn ( R := llcoll ) and radius r, and",
"3"
] |
{
"arxiv_id": "2106.13863",
"language": "zh",
"page_idx": 2
}
|
||
2106.13863-zh-3
|
2106.13863-zh-3.jpg
|
<FigureHere>
Figure 5: A regular tetrahedron as a basis. Without loss of
generality, assume$\| c_0\| = \sqrt {3}$ .
放),以得到正交矩阵
<FigureHere>
Figure 4: The effect of rotation on the spherical neuron activation in 3D; $t(\theta)$ denotes the tangent length as a function of rotation angle.
$\mathbf{x}\in\mathbb{R}^n$ be a point represented by $X\in \mathbb{R} ^n+ 2, \textit{see ( 4) }.$ Let further S' be the classifter that is obtained by rotating $S$ in $nD$ space( i.e, using an element of SO(n)), then $X^{\top}S^{\prime}$ and $X^\top S$ are related by spherical harmonics up to first degree.
Proof. Without loss of generality, the rotation is defined by the plane of rotation $\pi$ and the angle $\theta.$ Denote the projection of a vector $v\in\mathbb{R}^n$ on $\boldsymbol\pi$ by $v_\mathbf{\pi}$ and define $v_\perp\boldsymbol\pi=v-v_\mathbf{\pi}$ . From (??) we obtain
$$\begin{aligned}2X^\top S&=r^2-\left\|x-c_0\right\|^2\\&=r^2-\left\|(x-c_0)_{\perp\boldsymbol{\pi}}\right\|^2-\left\|(x-c_0)_{\boldsymbol{\pi}}\right\|^2\:.\end{aligned}$$
A rotation in $\pi$ only affects the rightmost term above and
there exists a $\phi\in[0,2\pi)$ such that
$$\begin{aligned}\left\|(x-c_{0})_{\boldsymbol{\pi}}\right\|^{2}&=\left\|x_{\boldsymbol{\pi}}-c_{0\:\boldsymbol{\pi}}\right\|^{2}\\&=\left\|x_{\boldsymbol{\pi}}\right\|^{2}+\left\|c_{0\:\boldsymbol{\pi}}\right\|^{2}-2\left\|x_{\boldsymbol{\pi}}\right\|\left\|c_{0\:\boldsymbol{\pi}}\right\|\cos\phi\:.\end{aligned}$$
With a similar argument, we obtain
$$\begin{aligned}2X^{\top}S^{\prime}&=r^{2}-\left\|(x-c_{0})_{\perp\boldsymbol{\pi}}\right\|^{2}\\&-\left\|x_{\pi}\right\|^{2}-\left\|c_{0\:\pi}\right\|^{2}+2\left\|x_{\pi}\right\|\left\|c_{0\:\pi}\right\|\:\cos(\phi+\theta).\end{aligned}$$
口
此结果在任何维度上都有效，但是我们主要对$n=2$ 感兴趣，如图$\boxed{\mathrm{E}}$3 所示的情况 $\phi=0$ 和 $n=3$, 如图$\boxed{\mathrm{E}}$ 4所示
根据 Freeman et al.(1991) 中定理 4 的结果并使用$N=1$ ,我们认为$M=(N+1)^2=4$基函数足以满足3D 情况(2)
## 4.2 Spherical filter banks in 3D
因此，在 3D 中，我们选择函数$f(X)$的四个旋转版本作为基函数。必须选择旋转$\{R_j\}_{j=1}^4$以满足定理 4 (Freeman et al., 1991)中的条件(b)。因此，我们变换$f( X)$ , 使得得到的四个球体在三个维度上等距，i.e., ,形成一个正四面体，其顶点为(1,1,1) ,(1,-1,-1) $、(-1,1,-1)$和(-1,-1,1),如图 5 所示。我们将四面体顶点$m_j$的齐次坐标按列堆叠在矩阵中(按 1/2 缩
$$M=\begin{bmatrix}m_1&m_2&m_3&m_4\end{bmatrix}=\dfrac{1}{2}\begin{bmatrix}1&1&-1&-1\\1&-1&1&-1\\1&-1&-1&1\\1&1&1&1\end{bmatrix}.$$
我们将使用此矩阵运算符$M$计算由正四面体
(Granlund& Knutsson, 1995) 的顶点生成的向量空间中的线性系数。这对于在 4.3 节中找到适当的插值函数并制定可操纵性约束是必要的
函数$f(X)$的四个旋转版本将构成我们称为球形滤波器组的基函数。要构造此过滤器组，我们选择以下约定。最初学习的球形分类器$f(X)=X^\top S$首先旋转到$\|c_0\|\cdot(1,1,1)$(见图 5),并将相应的(测地线)转换表示为$R_O$ 。接下来，我们将变换后的球体旋转到正四面体的其他三个顶点中，并变换回原始坐标系(对于(1,-1,-1)的情况，请参见图E6 )
因此，一个球形分类器的结果滤波器组组成为以下
$4\times5$ 矩阵：
(7)
$$B(S)=\begin{bmatrix}R_{O}^{\top}\:R_{T_{i}}\:R_{O}\:S\end{bmatrix}_{i=0...3}\:,$$
其中每个$\{R_{T_i}\}_{i=0}^3$是$\mathbb{R}^5$中的旋转同构，对应于从(1,1,1)到正四面体的顶点$i+1$的 3D 旋转。因此， $R_{T_0}=I_5$
正如我们通过定理E$4. 2 所 示 , 球 形 滤 波 器 组 B( S)$
在 3D 旋转下是等变的
Theorem 4.2. Let S ER 5 be a 3 D spherical classifter( i.e., the hypersphere neuron) and $\mathbf{x}\in\dot{\mathbb{R}}^3$ be an input point representedby $X$ $\in$ $\mathbb{R} ^{5}$ , see (4) . Let further $B( S)$ $\in$ $\mathbb{R}^{4\times5}$ be a flter bank obtained according to (7). Then the $filterbankoutput\mathbf{y}=B(S)X\in\mathbb{R}^4$ is equivariant to 3D rotations of x .
$Proof.$ Since rotations in $\mathbb{R}^3$ are embedded into $\mathbb{R}^5$ acting on the first three components (for details, see Melnyk et al. (2021) ), we need to show that the left sub-matrix $B_3(S)\in$ $\mathbb{R}^{4\times3}$ of $B(S)$ is of rank 3. By construction, the four rows of $B_3(S)$ form the vertices of a tetrahedron, thus span $\mathbb{R}^3$
口
请注意，$B(S)$的右子矩阵$B_2(S)\in\mathbb{R}^4\times2$会产生一个独立于旋转的恒定矢量$B_2(S)[x_4,x_5]^\top\in\mathbb{R}^4$,其中$x_{4}$和$x_5$是 MAXTHX 的最后两个分量。这个额外的常数向量需要使用四面体的第四个顶点
4
|
[
[
0.60107421875,
0.0738525390625,
0.6650390625,
0.0738525390625,
0.6650390625,
0.085205078125,
0.60107421875,
0.085205078125
],
[
0.75634765625,
0.0738525390625,
0.79833984375,
0.0738525390625,
0.79833984375,
0.085205078125,
0.75634765625,
0.085205078125
],
[
0.1953125,
0.0802001953125,
0.2197265625,
0.0802001953125,
0.2197265625,
0.09344482421875,
0.1953125,
0.09344482421875
],
[
0.378173828125,
0.0770263671875,
0.3857421875,
0.0770263671875,
0.3857421875,
0.083984375,
0.378173828125,
0.083984375
],
[
0.33251953125,
0.09027099609375,
0.361083984375,
0.09027099609375,
0.361083984375,
0.10479736328125,
0.33251953125,
0.10479736328125
],
[
0.1748046875,
0.10986328125,
0.1854248046875,
0.10986328125,
0.1854248046875,
0.125,
0.1748046875,
0.125
],
[
0.259033203125,
0.1256103515625,
0.26318359375,
0.1256103515625,
0.26318359375,
0.1319580078125,
0.259033203125,
0.1319580078125
],
[
0.371826171875,
0.1395263671875,
0.386474609375,
0.1395263671875,
0.386474609375,
0.149658203125,
0.371826171875,
0.149658203125
],
[
0.58837890625,
0.1427001953125,
0.650390625,
0.1427001953125,
0.650390625,
0.15283203125,
0.58837890625,
0.15283203125
],
[
0.740234375,
0.1634521484375,
0.8037109375,
0.1634521484375,
0.8037109375,
0.1749267578125,
0.740234375,
0.1749267578125
],
[
0.09234619140625,
0.1875,
0.47802734375,
0.1875,
0.47802734375,
0.2008056640625,
0.09234619140625,
0.2008056640625
],
[
0.521484375,
0.1944580078125,
0.90869140625,
0.1944580078125,
0.90869140625,
0.208984375,
0.521484375,
0.208984375
],
[
0.0906982421875,
0.20263671875,
0.47802734375,
0.20263671875,
0.47802734375,
0.216552734375,
0.0906982421875,
0.216552734375
],
[
0.08905029296875,
0.216552734375,
0.2034912109375,
0.2183837890625,
0.2034912109375,
0.231689453125,
0.08905029296875,
0.2298583984375
],
[
0.51953125,
0.2108154296875,
0.72802734375,
0.208984375,
0.72802734375,
0.2235107421875,
0.51953125,
0.2259521484375
],
[
0.52294921875,
0.2442626953125,
0.67724609375,
0.2442626953125,
0.67724609375,
0.2587890625,
0.52294921875,
0.2587890625
],
[
0.09234619140625,
0.25244140625,
0.4755859375,
0.25244140625,
0.4755859375,
0.267578125,
0.09234619140625,
0.267578125
],
[
0.08905029296875,
0.267578125,
0.47802734375,
0.268310546875,
0.47802734375,
0.28271484375,
0.08905029296875,
0.2822265625
],
[
0.7587890625,
0.272705078125,
0.771484375,
0.272705078125,
0.771484375,
0.28271484375,
0.7587890625,
0.28271484375
],
[
0.86669921875,
0.272705078125,
0.88623046875,
0.272705078125,
0.88623046875,
0.28271484375,
0.86669921875,
0.28271484375
],
[
0.0906982421875,
0.28271484375,
0.47802734375,
0.2841796875,
0.47802734375,
0.298583984375,
0.0906982421875,
0.29736328125
],
[
0.833984375,
0.285888671875,
0.84619140625,
0.285888671875,
0.84619140625,
0.296630859375,
0.833984375,
0.296630859375
],
[
0.859375,
0.285400390625,
0.88134765625,
0.285400390625,
0.88134765625,
0.29736328125,
0.859375,
0.29736328125
],
[
0.7900390625,
0.287353515625,
0.80810546875,
0.287353515625,
0.80810546875,
0.296630859375,
0.7900390625,
0.296630859375
],
[
0.525390625,
0.291748046875,
0.71728515625,
0.292236328125,
0.71728515625,
0.30810546875,
0.525390625,
0.306884765625
],
[
0.7607421875,
0.287353515625,
0.77294921875,
0.287353515625,
0.77294921875,
0.296630859375,
0.7607421875,
0.296630859375
],
[
0.0906982421875,
0.29736328125,
0.476318359375,
0.298583984375,
0.476318359375,
0.313232421875,
0.0906982421875,
0.3125
],
[
0.7939453125,
0.3037109375,
0.80615234375,
0.3037109375,
0.80615234375,
0.30810546875,
0.7939453125,
0.30810546875
],
[
0.82861328125,
0.300537109375,
0.84326171875,
0.300537109375,
0.84326171875,
0.31005859375,
0.82861328125,
0.31005859375
],
[
0.7548828125,
0.314453125,
0.77392578125,
0.314453125,
0.77392578125,
0.330078125,
0.7548828125,
0.330078125
],
[
0.79345703125,
0.31494140625,
0.81201171875,
0.31494140625,
0.81201171875,
0.32763671875,
0.79345703125,
0.32763671875
],
[
0.833984375,
0.31640625,
0.8447265625,
0.31640625,
0.8447265625,
0.32568359375,
0.833984375,
0.32568359375
],
[
0.87158203125,
0.31640625,
0.88232421875,
0.31640625,
0.88232421875,
0.326904296875,
0.87158203125,
0.326904296875
],
[
0.88623046875,
0.330078125,
0.90869140625,
0.330078125,
0.90869140625,
0.343505859375,
0.88623046875,
0.343505859375
],
[
0.08905029296875,
0.31640625,
0.1683349609375,
0.31640625,
0.1683349609375,
0.32763671875,
0.08905029296875,
0.32763671875
],
[
0.0906982421875,
0.343505859375,
0.476318359375,
0.343505859375,
0.476318359375,
0.356689453125,
0.0906982421875,
0.356689453125
],
[
0.52294921875,
0.3447265625,
0.90673828125,
0.3447265625,
0.90673828125,
0.35791015625,
0.52294921875,
0.35791015625
],
[
0.0906982421875,
0.35791015625,
0.476318359375,
0.35791015625,
0.476318359375,
0.371826171875,
0.0906982421875,
0.371826171875
],
[
0.521484375,
0.35986328125,
0.90673828125,
0.35986328125,
0.90673828125,
0.373779296875,
0.521484375,
0.373779296875
],
[
0.0906982421875,
0.37255859375,
0.4755859375,
0.37255859375,
0.4755859375,
0.386962890625,
0.0906982421875,
0.386962890625
],
[
0.52392578125,
0.375732421875,
0.90673828125,
0.375732421875,
0.90673828125,
0.389404296875,
0.52392578125,
0.389404296875
],
[
0.09234619140625,
0.38818359375,
0.238525390625,
0.38818359375,
0.238525390625,
0.401611328125,
0.09234619140625,
0.401611328125
],
[
0.52294921875,
0.39013671875,
0.7509765625,
0.39013671875,
0.7509765625,
0.404052734375,
0.52294921875,
0.404052734375
],
[
0.5390625,
0.406005859375,
0.90673828125,
0.406005859375,
0.90673828125,
0.419189453125,
0.5390625,
0.419189453125
],
[
0.12091064453125,
0.417236328125,
0.28759765625,
0.415283203125,
0.28759765625,
0.429931640625,
0.12091064453125,
0.431884765625
],
[
0.52392578125,
0.42041015625,
0.90673828125,
0.42041015625,
0.90673828125,
0.434326171875,
0.52392578125,
0.434326171875
],
[
0.17236328125,
0.438232421875,
0.42724609375,
0.438232421875,
0.42724609375,
0.455810546875,
0.17236328125,
0.455810546875
],
[
0.52392578125,
0.43505859375,
0.90673828125,
0.43505859375,
0.90673828125,
0.448974609375,
0.52392578125,
0.448974609375
],
[
0.52294921875,
0.449462890625,
0.90673828125,
0.45068359375,
0.90673828125,
0.46533203125,
0.52294921875,
0.464111328125
],
[
0.09234619140625,
0.4697265625,
0.47802734375,
0.4697265625,
0.47802734375,
0.48291015625,
0.09234619140625,
0.48291015625
],
[
0.52294921875,
0.466552734375,
0.90673828125,
0.466552734375,
0.90673828125,
0.479736328125,
0.52294921875,
0.479736328125
],
[
0.0906982421875,
0.48486328125,
0.318603515625,
0.48486328125,
0.318603515625,
0.498779296875,
0.0906982421875,
0.498779296875
],
[
0.52294921875,
0.48095703125,
0.90869140625,
0.48095703125,
0.90869140625,
0.494384765625,
0.52294921875,
0.494384765625
],
[
0.52294921875,
0.49560546875,
0.79150390625,
0.49560546875,
0.79150390625,
0.5087890625,
0.52294921875,
0.5087890625
],
[
0.5361328125,
0.5087890625,
0.90625,
0.51025390625,
0.90625,
0.5244140625,
0.5361328125,
0.5234375
],
[
0.09478759765625,
0.51513671875,
0.291748046875,
0.51318359375,
0.291748046875,
0.53076171875,
0.09478759765625,
0.53271484375
],
[
0.52294921875,
0.5263671875,
0.60107421875,
0.5263671875,
0.60107421875,
0.53857421875,
0.52294921875,
0.53857421875
],
[
0.1895751953125,
0.5361328125,
0.465576171875,
0.5361328125,
0.465576171875,
0.5537109375,
0.1895751953125,
0.5537109375
],
[
0.6044921875,
0.55322265625,
0.76123046875,
0.55029296875,
0.76123046875,
0.568359375,
0.6044921875,
0.57080078125
],
[
0.8857421875,
0.55615234375,
0.91015625,
0.55615234375,
0.91015625,
0.5693359375,
0.8857421875,
0.5693359375
],
[
0.09234619140625,
0.5673828125,
0.32275390625,
0.5673828125,
0.32275390625,
0.58056640625,
0.09234619140625,
0.58056640625
],
[
0.53759765625,
0.5859375,
0.90625,
0.5859375,
0.90625,
0.60107421875,
0.53759765625,
0.60107421875
],
[
0.0931396484375,
0.5966796875,
0.298095703125,
0.5966796875,
0.298095703125,
0.6142578125,
0.0931396484375,
0.6142578125
],
[
0.52294921875,
0.6015625,
0.90869140625,
0.6015625,
0.90869140625,
0.61572265625,
0.52294921875,
0.61572265625
],
[
0.1650390625,
0.61865234375,
0.473876953125,
0.61865234375,
0.473876953125,
0.63623046875,
0.1650390625,
0.63623046875
],
[
0.521484375,
0.6171875,
0.58837890625,
0.6171875,
0.58837890625,
0.6318359375,
0.521484375,
0.6318359375
],
[
0.53759765625,
0.6318359375,
0.90869140625,
0.6318359375,
0.90869140625,
0.64599609375,
0.53759765625,
0.64599609375
],
[
0.4599609375,
0.6494140625,
0.4794921875,
0.6494140625,
0.4794921875,
0.662109375,
0.4599609375,
0.662109375
],
[
0.51953125,
0.64599609375,
0.68310546875,
0.646484375,
0.68310546875,
0.6611328125,
0.51953125,
0.66015625
],
[
0.521484375,
0.671875,
0.90673828125,
0.671875,
0.90673828125,
0.68505859375,
0.521484375,
0.68505859375
],
[
0.108642578125,
0.68310546875,
0.47802734375,
0.68310546875,
0.47802734375,
0.6962890625,
0.108642578125,
0.6962890625
],
[
0.521484375,
0.6875,
0.90869140625,
0.6875,
0.90869140625,
0.70068359375,
0.521484375,
0.70068359375
],
[
0.0906982421875,
0.69775390625,
0.476318359375,
0.6962890625,
0.476318359375,
0.7109375,
0.0906982421875,
0.7119140625
],
[
0.52294921875,
0.7021484375,
0.90673828125,
0.7021484375,
0.90673828125,
0.71533203125,
0.52294921875,
0.71533203125
],
[
0.0906982421875,
0.71337890625,
0.1396484375,
0.71337890625,
0.1396484375,
0.72802734375,
0.0906982421875,
0.72802734375
],
[
0.5185546875,
0.71337890625,
0.90869140625,
0.71533203125,
0.90869140625,
0.73291015625,
0.5185546875,
0.73095703125
],
[
0.108642578125,
0.72900390625,
0.4755859375,
0.72900390625,
0.4755859375,
0.74365234375,
0.108642578125,
0.74365234375
],
[
0.521484375,
0.73095703125,
0.90869140625,
0.73095703125,
0.90869140625,
0.744140625,
0.521484375,
0.744140625
],
[
0.09234619140625,
0.74365234375,
0.473876953125,
0.74365234375,
0.473876953125,
0.7568359375,
0.09234619140625,
0.7568359375
],
[
0.521484375,
0.74853515625,
0.61669921875,
0.74853515625,
0.61669921875,
0.76025390625,
0.521484375,
0.76025390625
],
[
0.0906982421875,
0.7587890625,
0.1748046875,
0.7587890625,
0.1748046875,
0.7734375,
0.0906982421875,
0.7734375
],
[
0.5185546875,
0.76806640625,
0.90869140625,
0.76953125,
0.90869140625,
0.787109375,
0.5185546875,
0.7861328125
],
[
0.52294921875,
0.78857421875,
0.90625,
0.78857421875,
0.90625,
0.798828125,
0.52294921875,
0.798828125
],
[
0.0906982421875,
0.7998046875,
0.134033203125,
0.7998046875,
0.134033203125,
0.814453125,
0.0906982421875,
0.814453125
],
[
0.136474609375,
0.7998046875,
0.370849609375,
0.798828125,
0.370849609375,
0.81298828125,
0.136474609375,
0.814453125
],
[
0.521484375,
0.8017578125,
0.90673828125,
0.8017578125,
0.90673828125,
0.81494140625,
0.521484375,
0.81494140625
],
[
0.521484375,
0.814453125,
0.90673828125,
0.8173828125,
0.90673828125,
0.83203125,
0.521484375,
0.8291015625
],
[
0.09234619140625,
0.8271484375,
0.4755859375,
0.8271484375,
0.4755859375,
0.84033203125,
0.09234619140625,
0.84033203125
],
[
0.51953125,
0.83056640625,
0.90673828125,
0.830078125,
0.90673828125,
0.8447265625,
0.51953125,
0.84521484375
],
[
0.08905029296875,
0.84033203125,
0.47802734375,
0.8388671875,
0.47802734375,
0.85693359375,
0.08905029296875,
0.85791015625
],
[
0.8935546875,
0.8486328125,
0.90869140625,
0.8486328125,
0.90869140625,
0.859375,
0.8935546875,
0.859375
],
[
0.09234619140625,
0.85693359375,
0.473876953125,
0.85693359375,
0.473876953125,
0.87060546875,
0.09234619140625,
0.87060546875
],
[
0.09234619140625,
0.87255859375,
0.473876953125,
0.87255859375,
0.473876953125,
0.88623046875,
0.09234619140625,
0.88623046875
],
[
0.5361328125,
0.8720703125,
0.896484375,
0.869140625,
0.896484375,
0.8837890625,
0.5361328125,
0.88623046875
],
[
0.09722900390625,
0.88720703125,
0.4755859375,
0.88720703125,
0.4755859375,
0.90087890625,
0.09722900390625,
0.90087890625
],
[
0.521484375,
0.88623046875,
0.90673828125,
0.88525390625,
0.90673828125,
0.8994140625,
0.521484375,
0.90087890625
],
[
0.101318359375,
0.90087890625,
0.4755859375,
0.9013671875,
0.4755859375,
0.91650390625,
0.101318359375,
0.91552734375
],
[
0.52294921875,
0.9013671875,
0.90673828125,
0.9013671875,
0.90673828125,
0.91650390625,
0.52294921875,
0.91650390625
],
[
0.09234619140625,
0.91748046875,
0.476318359375,
0.91748046875,
0.476318359375,
0.93115234375,
0.09234619140625,
0.93115234375
],
[
0.52294921875,
0.91748046875,
0.78515625,
0.91748046875,
0.78515625,
0.93115234375,
0.52294921875,
0.93115234375
],
[
0.49267578125,
0.955078125,
0.50634765625,
0.955078125,
0.50634765625,
0.96630859375,
0.49267578125,
0.96630859375
]
] |
[
"(-1,1, -1)",
"(1,1,1)",
"t(0)",
"y",
"t(0)",
"0",
"",
"Co",
"(-1, -1,1)",
"(1, 1, 1)",
"Figure 4:The effect of rotation on the spherical neuron ac-",
"Figure 5: A regular tetrahedron as a basis. Without loss of",
"tivation in 3D; t(O) denotes the tangent length as a function",
"of rotation angle.",
"generality, assumecoll = V3.",
"放），以得到正交矩阵",
"x E IRn be a point represented by X e Rn+2 , see (4)",
"Let further S' be the classifier that is obtained by rotating",
"1",
"-1",
"S in nD space (i.e., using an element of SO(n) ), then",
"1",
"-1",
"-1",
"M=[m1 m2m3m4]",
"1",
"XT s' and xT S are related by spherical harmonics up to",
"",
"-1",
"[1 ",
"1",
"1",
"1",
"(6)",
"first degree.",
"Proof. Without loss of generality, the rotation is defined by",
"我们将使用此矩阵运算符M计算由正四面体",
"the plane of rotation and the angle . Denote the projec-",
"(Granlund&Knutsson，1995）的顶点生成的向量空间",
"tion of a vector E Rn on π by U元 and define v1元 = w- U元",
"中的线性系数。这对于在4.3节中找到适当的插值函",
". From (??) we obtain",
"数并制定可操纵性约束是必要的",
"函数f(X）的四个旋转版本将构成我们称为球形滤",
"2XTS=2--co2",
"波器组的基函数。要构造此过滤器组，我们选择以下",
"=2 _( - co)/12 ( co)/2",
"约定。最初学习的球形分类器f(X）=XTS首先旋转",
"到col·（1,1,1）（见图5），并将相应的（测地线）转",
"A rotation in only affects the rightmost term above and",
"换表示为Ro。接下来，我们将变换后的球体旋转到正",
"there exists a @Φ [0, 2) such that",
"四面体的其他三个顶点中，并变换回原始坐标系（对",
"于（1，-1,-1)的情况，请参见图E6）",
"因此，一个球形分类器的结果滤波器组组成为以下",
"( - c0)2 = [ — C0元2",
"4×5矩阵：",
"= [///2 + co //2 2元/ |co元l cos ",
"B(S) =[R Rr, Ro S",
"(7)",
"With a similar argument, we obtain",
"其中每个（RT1-。是R5中的旋转同构，对应于",
"2XTs = r2 -(- co)/?",
"从（1,1,1）到正四面体的顶点i+1的3D旋转。因此",
"1元/2 c 元/2 + 2元/ co元/ cos(+ 0),",
"RT。=Is",
"正如我们通过定理E4.2所示，球形滤波器组B(S)",
"口",
"在3D旋转下是等变的",
"Theorem 4.2. Let S R5 be a 3 D spherical classifier (",
"此结果在任何维度上都有效，但是我们主要对n=2",
"ie., the hypersphere neuron) and x E R3 be an input point",
"感兴趣，如图E3所示的情况Φ=0和n=3，如图F",
"represented by X E R5 ,see (4). Let further B(S) E",
"4所示",
"IR4x5 be a filter bank obtained according to (7) . Then the",
"根据Freemanetal.（1991）中定理4的结果并使用",
"filter bank output y = B(S)X E R4 is equivariant to 3D",
"N=1，我们认为M=（N+1)2=4基函数足以满足",
"rotations of x.",
"3D情况(2)",
"Proof. Since rotations in R3 are embedded into IR5 acting",
"onthefirstthreecomponents(fordetails,seeMelnyketal.",
"4.2",
"Spherical filter banks in 3D",
"(2021) ), we need to show that the left sub-matrix B3(S) E",
"IR4x3 of B(S) is of rank 3. By construction, the four rows",
"因此，在3D中，我们选择函数f(X）的四个旋转版",
"of B3(S) form the vertices of a tetrahedron, thus span R3",
"本作为基函数。必须选择旋转[R,}1-1 以满足定理4",
"口",
"(Freemanetal.,1991)中的条件（b）。因此，我们变换",
"f(X），使得得到的四个球体在三个维度上等距，i.e.",
"请注意，B(S)的右子矩阵B2（S）ER4×2会产生",
"，形成一个正四面体，其顶点为（1,1,1）、（1，-1，-1)",
"个独立于旋转的恒定矢量B2(S)[a4,a5]TER4其中",
"（-1,1,-1)和（-1,-1,1），如图5所示。我们将四",
"C4和5是MAXTHX的最后两个分量。这个额外的常",
"面体顶点m，的齐次坐标按列堆叠在矩阵中（按1/2缩",
"数向量需要使用四面体的第四个顶点",
"4"
] |
{
"arxiv_id": "2106.13863",
"language": "zh",
"page_idx": 3
}
|
||
2106.13863-zh-4
|
2106.13863-zh-4.jpg
|
$$\begin{array}{c|c}(1,-1,-1)&\boldsymbol{R}_O^\top\\\bullet&\boldsymbol{\bullet}^\mathbf{c}\\\boldsymbol{R}_{T_1}\uparrow&\boldsymbol{R}\\\bullet&\boldsymbol{R}_O&\bullet_{\mathbf{c}_0}\end{array}$$
Figure 6: A rotation from $c_0$ to $c$ described by a tetrahedron
rotation. Without loss of generality, assume$\| c_0\| = \sqrt {3}$ .
此外，我们可以明确显示如何在滤波器组$B(S)$输出
空间中获得 3D 旋转 R 的表示$V_R\in\mathbb{R}^{4\times4}$
Proposition 4.3. Let M be the 4×4 orthogonal matrix de- $fnedin(6),R_{O}^{k}\in\mathbb{R}^{4\times4}$ be the $P( \mathbb{R} ^{3}) \textit{ representation ( con- })$ structed by appending ones to the main diagonal- see the last paragraph in Section 3.3) of the 3D rotation ( geodesic) from the center c$_0^k\textit{of the learned k th spherical classifier S}_k$ $to\left \| \mathbf{c} _{0}^{k}\right \| \cdot ( 1, 1, 1)$ $, and$, slightly abusing $notation$, $R\in \mathbb{R} ^{4\times 4}$ be the P( R 3) representation of the 3D rotation R. Then $V_R^k\in\mathbb{R}^{4\times4}$ 'lefined below is the representation of the 3D rotation R in the flter bank B(Sk) output space:
$$V_{R}^{k}=\textbf{M}^{\top}R_{O}^{k}\:R\:R_{O}^{k\top}\textbf{M}\:.$$
(8)
$Proof.$ Since det M =1 ,we have that M $\in$SO(4),and thus, all terms on the right hand-side of (8) are 4×4 rotation matrices with only one variable $-R$ .Therefore, $V_{\boldsymbol{R}}^k$ is also a rotation matrix and is a unique map. The inverse transformation can be computed straightforwardly as
$$R=R_{O}^{k^{\top}}M\:V_{R}^{k}\:M^{\top}R_{O}^{k}$$
(9)
with subsequent extraction of the upper-left 3×3 sub-matrix as the original 3D rotation matrix $R.$ Hence, there is a one-
口
to-one mapping between $V_{R}^{k}$ and $R.$
因此，$V_{R}$和$R$都是 3D 旋转的表示，我们可以将球
形滤波器组$B(S)$的旋转等方差特性写为
(10)
$$V_{\boldsymbol{R}}\:B(S)\:X=B(S)\:RX.$$
### 4.3 3D steerability constraint
可操控性约束可表述如下。对于应用于函数$f(X)$的输人的任意旋转$R$ ,我们希望使用$v_j(R)$对 (7)中的球形滤波器组$B(S)$的输出进行插值，以使响应等于原始函数输出 i.e.,
$$\begin{aligned}f(X)=f^{R}(RX)&=\sum_{j=1}^{M}v_{j}(R)\:f^{R_{j}}(RX)\\&=v(R)^{\top}B(S)\:RX,\end{aligned}$$
(11)
,其中$X\in\mathbb{R}^5$是单个嵌入点，
系数$v(R)$应符合基函数构造 (Freeman et al.(1991) 中定理$\boxed{\mathrm{F}}4$ 中的条件 (b)),这就是为什么使用 (6)中
定义的 M 计算它们的原因。给定$X\in\mathbb{R}^5$作为输入，并且未知的旋转$R$作用在其上，转向方程(11)意味着
(12)
$$v(R)^\top B(S)\:RX=v(I)^\top B(S)\:X\:.$$
给定四面体旋转，e。g。$R_{T_1}$,如图E6中的图所示，我们可以相应地将未知旋转定义为$R=R_O^\top R_{T_1}R_O$。在这种情况下，很容易看出，要满足约束(12),v(R)必须是(0,1,0,0) , i。e。,则必须选择滤波器组$B(S)$中的第二个滤波器。这可以通过旋转$R_{T_1}$变换常数向量$m_1=(\frac12,\frac12,\frac12,\frac12)$并将其乘以基础矩阵$M$来实现，如下所示：
$$\begin{aligned}v(R)&=M^{\top}(R_{T_{1}}\:m_{1})\\&=M^{\top}\left(\frac{1}{2},-\frac{1}{2},-\frac{1}{2},\frac{1}{2}\right)=(0,1,0,0)\:.\end{aligned}$$
(13)
请注意，通常，几何神经元(5)将一组嵌人点作为输人。因此，通过上述设置，如果所有$k$都使用相同的$v$ , 则 对 于 每 个 输 入 形 状 点 $k$ , $m_1$ 的旋转将有所不同， 这与形状由刚体运动，i。e。, 所有$k$都使用相同的$R$ 。因此，我们需要为每个输入形状点$k$考虑一个合适的向量$v^k$,这样得到的$R$对于所有$k$都是相同的。这可以通过回顾我们如何在球形滤波器组(7)中构造基函数来实现：我们需要考虑相应的初始旋转 $R_O^k$。因此， 所需的插值系数$v^k$计算为
$$v^k(R)=M^\top\left(R_O^k\:R\:R_O^{k\top}m_1\right).$$
(14)
所得的$v^k(R)\in\mathbb{R}^4$对原始学习球体$S_k$的四面体副本
的响应进行插值，以替换旋转的球体
值得注意的是，插值向量$v^k(\boldsymbol{R})$ 是$V_R^k(8)$的第一列。请注意，$v^k(R)$只有三个自由度，因为正四面体的四个顶点总和为零(见图5)。因此，$v^k(R)$在 3D 旋转下是等变的。通过将(11)插人(5),我们现在可以为几何神经元$f_\mathrm{GN}$建立可操纵性约束，该约束将一组$K$嵌人点$\{X_1,\ldots,X_k\}$ 作为输人：
(15)
$$\begin{aligned}f_{\mathrm{GN}}^{R}(RX)&=\sum_{k=1}^{K}\gamma_{k}\:f^{k\:R}(RX_{k})\\&=\sum_{k=1}^{K}\gamma_{k}\:v^{k}(R)^{\top}B(S_{k})\:RX_{k}\:.\end{aligned}$$
4.4 Steerable model overview
要构建 3D 可操纵模型，我们执行以下步骤(请参见图$\textcircled{\mathrm{E}}1):$ 我们首先训练一个祖先模型，该模型由球形神经元组成。在训练模型参数以按规范方向对数据进行分类之后，我们将其冻结并根据 3D 可操纵性约束(11) (用于超球面神经元)或(15)(用于几何神经元) 转换第一层权重。最后，通过根据(7)组合球形滤波器组中的结果参数并添加插值系数(14)作为自由参数， 我们创建了一个可控制的模型。如果插值系数计算正确，模型将产生旋转不变的预测
## 5 $\textbf{Experiments}$
在本节中，我们将描述为确认E4 节中提出的发现而进
行的实验。
5
|
[
[
0.303955078125,
0.07452392578125,
0.326904296875,
0.0770263671875,
0.324462890625,
0.091552734375,
0.301513671875,
0.08837890625
],
[
0.52294921875,
0.07257080078125,
0.90673828125,
0.07257080078125,
0.90673828125,
0.08587646484375,
0.52294921875,
0.08587646484375
],
[
0.52392578125,
0.08709716796875,
0.90625,
0.08709716796875,
0.90625,
0.10101318359375,
0.52392578125,
0.10101318359375
],
[
0.193603515625,
0.07763671875,
0.268798828125,
0.07763671875,
0.268798828125,
0.0889892578125,
0.193603515625,
0.0889892578125
],
[
0.5947265625,
0.10919189453125,
0.8408203125,
0.10797119140625,
0.8408203125,
0.12249755859375,
0.5947265625,
0.12371826171875
],
[
0.24267578125,
0.11614990234375,
0.265625,
0.11871337890625,
0.26318359375,
0.1331787109375,
0.240234375,
0.1300048828125
],
[
0.87744140625,
0.111083984375,
0.90673828125,
0.111083984375,
0.90673828125,
0.12249755859375,
0.87744140625,
0.12249755859375
],
[
0.361083984375,
0.12060546875,
0.373291015625,
0.12060546875,
0.373291015625,
0.1319580078125,
0.361083984375,
0.1319580078125
],
[
0.52294921875,
0.1319580078125,
0.90869140625,
0.1319580078125,
0.90869140625,
0.1458740234375,
0.52294921875,
0.1458740234375
],
[
0.52294921875,
0.1458740234375,
0.90869140625,
0.1458740234375,
0.90869140625,
0.160400390625,
0.52294921875,
0.160400390625
],
[
0.30224609375,
0.1610107421875,
0.326904296875,
0.1610107421875,
0.326904296875,
0.17236328125,
0.30224609375,
0.17236328125
],
[
0.521484375,
0.1610107421875,
0.90869140625,
0.1610107421875,
0.90869140625,
0.1749267578125,
0.521484375,
0.1749267578125
],
[
0.35546875,
0.1634521484375,
0.36767578125,
0.1572265625,
0.375732421875,
0.1673583984375,
0.363525390625,
0.17236328125
],
[
0.2181396484375,
0.1622314453125,
0.26953125,
0.1622314453125,
0.26953125,
0.173583984375,
0.2181396484375,
0.173583984375
],
[
0.52392578125,
0.1767578125,
0.90673828125,
0.1767578125,
0.90673828125,
0.1900634765625,
0.52392578125,
0.1900634765625
],
[
0.09234619140625,
0.1932373046875,
0.4755859375,
0.1932373046875,
0.4755859375,
0.2083740234375,
0.09234619140625,
0.2083740234375
],
[
0.52294921875,
0.1912841796875,
0.90673828125,
0.1912841796875,
0.90673828125,
0.2052001953125,
0.52294921875,
0.2052001953125
],
[
0.08905029296875,
0.210205078125,
0.459228515625,
0.208984375,
0.459228515625,
0.2235107421875,
0.08905029296875,
0.2247314453125
],
[
0.521484375,
0.20703125,
0.90869140625,
0.205810546875,
0.90869140625,
0.2203369140625,
0.521484375,
0.2215576171875
],
[
0.52392578125,
0.2203369140625,
0.580078125,
0.222900390625,
0.57861328125,
0.2374267578125,
0.52294921875,
0.23486328125
],
[
0.108642578125,
0.2442626953125,
0.476318359375,
0.2442626953125,
0.476318359375,
0.25830078125,
0.108642578125,
0.25830078125
],
[
0.55224609375,
0.242431640625,
0.70166015625,
0.242431640625,
0.70166015625,
0.2568359375,
0.55224609375,
0.2568359375
],
[
0.08905029296875,
0.25830078125,
0.3857421875,
0.255126953125,
0.3857421875,
0.2734375,
0.08905029296875,
0.276611328125
],
[
0.6435546875,
0.261962890625,
0.74853515625,
0.261962890625,
0.74853515625,
0.274658203125,
0.6435546875,
0.274658203125
],
[
0.8759765625,
0.2568359375,
0.91015625,
0.2568359375,
0.91015625,
0.272705078125,
0.8759765625,
0.272705078125
],
[
0.5908203125,
0.267578125,
0.640625,
0.267578125,
0.640625,
0.2822265625,
0.5908203125,
0.2822265625
],
[
0.7509765625,
0.268310546875,
0.8447265625,
0.268310546875,
0.8447265625,
0.28271484375,
0.7509765625,
0.28271484375
],
[
0.6357421875,
0.27587890625,
0.7490234375,
0.27587890625,
0.7490234375,
0.2890625,
0.6357421875,
0.2890625
],
[
0.0906982421875,
0.2841796875,
0.47802734375,
0.2841796875,
0.47802734375,
0.29736328125,
0.0906982421875,
0.29736328125
],
[
0.0906982421875,
0.29736328125,
0.47802734375,
0.29736328125,
0.47802734375,
0.3125,
0.0906982421875,
0.3125
],
[
0.52294921875,
0.29541015625,
0.90625,
0.29541015625,
0.90625,
0.309326171875,
0.52294921875,
0.309326171875
],
[
0.0906982421875,
0.314453125,
0.47802734375,
0.313232421875,
0.47802734375,
0.32763671875,
0.0906982421875,
0.328857421875
],
[
0.52294921875,
0.311279296875,
0.90625,
0.311279296875,
0.90625,
0.324462890625,
0.52294921875,
0.324462890625
],
[
0.0906982421875,
0.32763671875,
0.4755859375,
0.32763671875,
0.4755859375,
0.34228515625,
0.0906982421875,
0.34228515625
],
[
0.521484375,
0.324462890625,
0.90673828125,
0.32568359375,
0.90673828125,
0.34033203125,
0.521484375,
0.339111328125
],
[
0.0906982421875,
0.3447265625,
0.473876953125,
0.3447265625,
0.473876953125,
0.35791015625,
0.0906982421875,
0.35791015625
],
[
0.52294921875,
0.34033203125,
0.90673828125,
0.34033203125,
0.90673828125,
0.354248046875,
0.52294921875,
0.354248046875
],
[
0.0882568359375,
0.35791015625,
0.47802734375,
0.35595703125,
0.47802734375,
0.373779296875,
0.0882568359375,
0.375732421875
],
[
0.52392578125,
0.35595703125,
0.90673828125,
0.35595703125,
0.90673828125,
0.369384765625,
0.52392578125,
0.369384765625
],
[
0.0906982421875,
0.375,
0.476318359375,
0.375,
0.476318359375,
0.38818359375,
0.0906982421875,
0.38818359375
],
[
0.52392578125,
0.37060546875,
0.90673828125,
0.37060546875,
0.90673828125,
0.3837890625,
0.52392578125,
0.3837890625
],
[
0.0906982421875,
0.386962890625,
0.47802734375,
0.386962890625,
0.47802734375,
0.40478515625,
0.0906982421875,
0.40478515625
],
[
0.521484375,
0.385009765625,
0.90673828125,
0.386474609375,
0.90673828125,
0.399658203125,
0.521484375,
0.3984375
],
[
0.08905029296875,
0.404052734375,
0.41015625,
0.40478515625,
0.41015625,
0.419189453125,
0.08905029296875,
0.41845703125
],
[
0.52294921875,
0.40087890625,
0.90673828125,
0.40087890625,
0.90673828125,
0.415283203125,
0.52294921875,
0.415283203125
],
[
0.52294921875,
0.416748046875,
0.7099609375,
0.416748046875,
0.7099609375,
0.429931640625,
0.52294921875,
0.429931640625
],
[
0.1953125,
0.43115234375,
0.374267578125,
0.43115234375,
0.374267578125,
0.44580078125,
0.1953125,
0.44580078125
],
[
0.455810546875,
0.43115234375,
0.47802734375,
0.43115234375,
0.47802734375,
0.44580078125,
0.455810546875,
0.44580078125
],
[
0.60888671875,
0.43505859375,
0.82275390625,
0.4375,
0.82275390625,
0.451904296875,
0.60888671875,
0.449462890625
],
[
0.87744140625,
0.438232421875,
0.90869140625,
0.438232421875,
0.90869140625,
0.45068359375,
0.87744140625,
0.45068359375
],
[
0.09234619140625,
0.458251953125,
0.47802734375,
0.458251953125,
0.47802734375,
0.4716796875,
0.09234619140625,
0.4716796875
],
[
0.52294921875,
0.458984375,
0.90869140625,
0.458984375,
0.90869140625,
0.472900390625,
0.52294921875,
0.472900390625
],
[
0.09234619140625,
0.474853515625,
0.476318359375,
0.474853515625,
0.476318359375,
0.48486328125,
0.09234619140625,
0.48486328125
],
[
0.52294921875,
0.474853515625,
0.78369140625,
0.474853515625,
0.78369140625,
0.488037109375,
0.52294921875,
0.488037109375
],
[
0.0906982421875,
0.4873046875,
0.47802734375,
0.488037109375,
0.47802734375,
0.50244140625,
0.0906982421875,
0.501953125
],
[
0.5400390625,
0.4912109375,
0.90625,
0.4912109375,
0.90625,
0.501953125,
0.5400390625,
0.501953125
],
[
0.0906982421875,
0.50390625,
0.476318359375,
0.50390625,
0.476318359375,
0.517578125,
0.0906982421875,
0.517578125
],
[
0.52294921875,
0.5048828125,
0.90673828125,
0.5048828125,
0.90673828125,
0.51806640625,
0.52294921875,
0.51806640625
],
[
0.0906982421875,
0.51953125,
0.438720703125,
0.51953125,
0.438720703125,
0.53271484375,
0.0906982421875,
0.53271484375
],
[
0.52294921875,
0.51953125,
0.90673828125,
0.51953125,
0.90673828125,
0.53271484375,
0.52294921875,
0.53271484375
],
[
0.52392578125,
0.5341796875,
0.90625,
0.5341796875,
0.90625,
0.54736328125,
0.52392578125,
0.54736328125
],
[
0.2034912109375,
0.546875,
0.363525390625,
0.546875,
0.363525390625,
0.56005859375,
0.2034912109375,
0.56005859375
],
[
0.455810546875,
0.546875,
0.47802734375,
0.546875,
0.47802734375,
0.56005859375,
0.455810546875,
0.56005859375
],
[
0.52294921875,
0.5498046875,
0.90673828125,
0.5498046875,
0.90673828125,
0.564453125,
0.52294921875,
0.564453125
],
[
0.52294921875,
0.56494140625,
0.69384765625,
0.56494140625,
0.69384765625,
0.5791015625,
0.52294921875,
0.5791015625
],
[
0.09234619140625,
0.57470703125,
0.476318359375,
0.57470703125,
0.476318359375,
0.58837890625,
0.09234619140625,
0.58837890625
],
[
0.0906982421875,
0.5888671875,
0.4794921875,
0.5888671875,
0.4794921875,
0.60302734375,
0.0906982421875,
0.60302734375
],
[
0.66650390625,
0.583984375,
0.68701171875,
0.583984375,
0.68701171875,
0.59765625,
0.66650390625,
0.59765625
],
[
0.57861328125,
0.5986328125,
0.64892578125,
0.5986328125,
0.64892578125,
0.61279296875,
0.57861328125,
0.61279296875
],
[
0.66650390625,
0.60107421875,
0.68115234375,
0.60107421875,
0.68115234375,
0.60986328125,
0.66650390625,
0.60986328125
],
[
0.6806640625,
0.60107421875,
0.7197265625,
0.60107421875,
0.7197265625,
0.611328125,
0.6806640625,
0.611328125
],
[
0.71826171875,
0.59521484375,
0.7861328125,
0.5986328125,
0.78515625,
0.61279296875,
0.71728515625,
0.60986328125
],
[
0.08905029296875,
0.60400390625,
0.328369140625,
0.60302734375,
0.328369140625,
0.6171875,
0.08905029296875,
0.61865234375
],
[
0.463134765625,
0.60498046875,
0.476318359375,
0.60498046875,
0.476318359375,
0.6162109375,
0.463134765625,
0.6162109375
],
[
0.6640625,
0.6142578125,
0.689453125,
0.6142578125,
0.689453125,
0.626953125,
0.6640625,
0.626953125
],
[
0.87744140625,
0.62060546875,
0.91015625,
0.62060546875,
0.91015625,
0.63623046875,
0.87744140625,
0.63623046875
],
[
0.108642578125,
0.62890625,
0.476318359375,
0.62890625,
0.476318359375,
0.642578125,
0.108642578125,
0.642578125
],
[
0.66845703125,
0.62890625,
0.685546875,
0.62890625,
0.685546875,
0.64306640625,
0.66845703125,
0.64306640625
],
[
0.09234619140625,
0.64453125,
0.381591796875,
0.64453125,
0.381591796875,
0.65771484375,
0.09234619140625,
0.65771484375
],
[
0.66259765625,
0.64306640625,
0.85107421875,
0.642578125,
0.85107421875,
0.6572265625,
0.66259765625,
0.65771484375
],
[
0.6640625,
0.66015625,
0.689453125,
0.66015625,
0.689453125,
0.671875,
0.6640625,
0.671875
],
[
0.200927734375,
0.67041015625,
0.366943359375,
0.67041015625,
0.366943359375,
0.68505859375,
0.200927734375,
0.68505859375
],
[
0.44677734375,
0.671875,
0.47802734375,
0.671875,
0.47802734375,
0.6845703125,
0.44677734375,
0.6845703125
],
[
0.52294921875,
0.68798828125,
0.57275390625,
0.68798828125,
0.57275390625,
0.69970703125,
0.52294921875,
0.69970703125
],
[
0.56640625,
0.68798828125,
0.78369140625,
0.68798828125,
0.78369140625,
0.7021484375,
0.56640625,
0.7021484375
],
[
0.09234619140625,
0.71728515625,
0.1380615234375,
0.71728515625,
0.1380615234375,
0.72900390625,
0.09234619140625,
0.72900390625
],
[
0.131591796875,
0.71728515625,
0.3544921875,
0.71728515625,
0.3544921875,
0.73095703125,
0.131591796875,
0.73095703125
],
[
0.52392578125,
0.7119140625,
0.90673828125,
0.7119140625,
0.90673828125,
0.72607421875,
0.52392578125,
0.72607421875
],
[
0.52392578125,
0.72802734375,
0.90673828125,
0.72802734375,
0.90673828125,
0.7412109375,
0.52392578125,
0.7412109375
],
[
0.0931396484375,
0.7421875,
0.4755859375,
0.7421875,
0.4755859375,
0.755859375,
0.0931396484375,
0.755859375
],
[
0.52392578125,
0.7421875,
0.90673828125,
0.7421875,
0.90673828125,
0.755859375,
0.52392578125,
0.755859375
],
[
0.096435546875,
0.7587890625,
0.4755859375,
0.7587890625,
0.4755859375,
0.76953125,
0.096435546875,
0.76953125
],
[
0.52294921875,
0.75830078125,
0.90673828125,
0.75830078125,
0.90673828125,
0.771484375,
0.52294921875,
0.771484375
],
[
0.0906982421875,
0.77099609375,
0.476318359375,
0.771484375,
0.476318359375,
0.7861328125,
0.0906982421875,
0.78515625
],
[
0.52294921875,
0.7744140625,
0.90625,
0.7744140625,
0.90625,
0.78515625,
0.52294921875,
0.78515625
],
[
0.0906982421875,
0.78515625,
0.17724609375,
0.787109375,
0.17724609375,
0.8017578125,
0.0906982421875,
0.7998046875
],
[
0.521484375,
0.7861328125,
0.90625,
0.787109375,
0.90625,
0.80029296875,
0.521484375,
0.7998046875
],
[
0.521484375,
0.8017578125,
0.904296875,
0.80322265625,
0.904296875,
0.81640625,
0.521484375,
0.81494140625
],
[
0.27294921875,
0.81103515625,
0.294189453125,
0.81103515625,
0.294189453125,
0.82373046875,
0.27294921875,
0.82373046875
],
[
0.1282958984375,
0.82080078125,
0.410888671875,
0.82080078125,
0.410888671875,
0.84228515625,
0.1282958984375,
0.84228515625
],
[
0.52294921875,
0.818359375,
0.90673828125,
0.818359375,
0.90673828125,
0.83203125,
0.52294921875,
0.83203125
],
[
0.44775390625,
0.8369140625,
0.4794921875,
0.8369140625,
0.4794921875,
0.849609375,
0.44775390625,
0.849609375
],
[
0.521484375,
0.83349609375,
0.7509765625,
0.83203125,
0.7509765625,
0.8466796875,
0.521484375,
0.84765625
],
[
0.268798828125,
0.84326171875,
0.293212890625,
0.84033203125,
0.295654296875,
0.8515625,
0.271240234375,
0.85498046875
],
[
0.253173828125,
0.85986328125,
0.386474609375,
0.85986328125,
0.386474609375,
0.87353515625,
0.253173828125,
0.87353515625
],
[
0.51708984375,
0.8671875,
0.69384765625,
0.87060546875,
0.69384765625,
0.8916015625,
0.51708984375,
0.88818359375
],
[
0.09234619140625,
0.88720703125,
0.3056640625,
0.88720703125,
0.3056640625,
0.90087890625,
0.09234619140625,
0.90087890625
],
[
0.1070556640625,
0.9013671875,
0.476318359375,
0.90087890625,
0.476318359375,
0.91552734375,
0.1070556640625,
0.91650390625
],
[
0.521484375,
0.90283203125,
0.90673828125,
0.90283203125,
0.90673828125,
0.91650390625,
0.521484375,
0.91650390625
],
[
0.0931396484375,
0.91748046875,
0.4755859375,
0.91748046875,
0.4755859375,
0.931640625,
0.0931396484375,
0.931640625
],
[
0.521484375,
0.91748046875,
0.59326171875,
0.91748046875,
0.59326171875,
0.931640625,
0.521484375,
0.931640625
],
[
0.491943359375,
0.95263671875,
0.50732421875,
0.95263671875,
0.50732421875,
0.96826171875,
0.491943359375,
0.96826171875
]
] |
[
"Ro",
"定义的M计算它们的原因。给定XER5作为输入，并",
"且未知的旋转R作用在其上，转向方程（11）意味着",
"(1, -1, -1)",
"U(R)B(S) RX = U(I)TB(S) X :",
"RT",
"(12)",
"R",
"给定四面体旋转，e。g。，RT，如图E6中的图所示，我",
"们可以相应地将未知旋转定义为R=RIRTRo。在",
"Ro",
"这种情况下，很容易看出，要满足约束（12），(R)必",
"C",
"(1, 1, 1)",
"须是(0,1,0,0)，i。e。，则必须选择滤波器组B(S)中",
"Figure 6: A rotation from co to c described by a tetrahedron",
"的第二个滤波器。这可以通过旋转RT变换常数向量",
"rotation. Without loss of generality, assume Icoll = V3 .",
"m1=（，号，，号）并将其乘以基础矩阵M来实现，如",
"下所示：",
"此外，我们可以明确显示如何在滤波器组B(S)输出",
"v(R) = M/(RT m1)",
"空间中获得3D旋转R的表示VREIR4x4",
"(111１",
"(13)",
"=MT",
"= (0, 1,0,0)",
"(2’2’-22)",
"Proposition 4.3. Let M be the 4 × 4 orthogonal matrix de",
"fined in (6), Rk e R4x4 be the P(R3) representation (con-",
"请注意，通常，几何神经元（5）将一组嵌人点作为输",
"structed by appending ones to the main diagonal - see the",
"人。因此，通过上述设置，如果所有k都使用相同的",
"last paragraph in Section 3.3)of the 3D rotation (geodesic)",
"u，则对于每个输人形状点k，m1的旋转将有所不同",
"from the center ch of the learned k th spherical classifier Sh",
"这与形状由刚体运动，i。e。，所有k都使用相同的R。",
"tolcill (1, 1, 1), and, slightly abusing notation, R e R4x4",
"因此，我们需要为每个输人形状点k考虑一个合适的",
"be the P(R3) representation of the 3D rotation R.Then",
"向量uk，这样得到的R对于所有k都是相同的。这可",
"Vh e IR4x4 defined below is the representation of the 3D",
"以通过回顾我们如何在球形滤波器组（7）中构造基函",
"rotation R in the filter bank B(Sk)output space:",
"数来实现：我们需要考虑相应的初始旋转R。因此",
"所需的插值系数计算为",
"VK = MR RROT M.",
"(8)",
"uk(R)= MT (RO RRTm1) :",
"(14)",
"Proof. Since det M = 1, we have that M E SO(4) , and",
"所得的k(R)ER4对原始学习球体S的四面体副本",
"thus.all terms on therighthand-sideof(8)are4x4rotation",
"的响应进行插值，以替换旋转的球体",
"matrices with only one variable R. Therefore, Vk is",
"值得注意的是，插值向量（R）是V（8）的第一列。",
"also a rotation matrix and is a unique map. The inverse",
"请注意，（R）只有三个自由度，因为正四面体的四个",
"transformation can be computed straightforwardly as",
"顶点总和为零（见图5）。因此，2k（R)在3D旋转下是",
"等变的。通过将（11)插入（5），我们现在可以为几何神",
"R= RO M VMR",
"(9)",
"经元fGN建立可操纵性约束，该约束将一组K嵌人点",
"[X1,..,X]作为输人·",
"with subsequent extraction of the upper-left 3x3 sub-matrix",
"as the original 3D rotation matrix R . Hence, there is a one-",
"K",
"fEN(RX)",
">",
"Ykf",
"kR(RX)",
"to-one mapping between Vk and R.",
"口",
"k=1",
"(15)",
"因此，VR和R都是3D旋转的表示，我们可以将球",
"K",
"形滤波器组B(S)的旋转等方差特性写为",
"ku(R)TB(Sk) RXk .",
"k=1",
"VR B(S) X = B(S) RX.",
"(10)",
"4.4",
"Steerablemodel overview",
"4.3",
"3D steerability constraint",
"要构建3D可操纵模型，我们执行以下步骤（请参见",
"图E1）：我们首先训练一个祖先模型，该模型由球形",
"可操控性约束可表述如下。对于应用于函数f(X）的输",
"神经元组成。在训练模型参数以按规范方向对数据进",
"人的任意旋转R，我们希望使用U（R）对（7）中的球形",
"行分类之后，我们将其冻结并根据3D可操纵性约束",
"滤波器组B(S)的输出进行插值，以使响应等于原始函",
"（11）（用于超球面神经元）或（15）（用于儿何神经元）",
"数输出i.e.，",
"转换第一层权重。最后，通过根据（7）组合球形滤波",
"器组中的结果参数并添加插值系数（14）作为自由参数",
"M",
"f(X) = fR(RX) =u;(R) JR(RX)",
"我们创建了一个可控制的模型。如果插值系数计算正",
"(11)",
"确，模型将产生旋转不变的预测",
"j=1",
"= (R) B(S) RX,",
"5 Experiments",
"，其中XER5是单个嵌人点，",
"系数(R)应符合基函数构造（Freemanetal.(1991)",
"在本节中，我们将描述为确认E4节中提出的发现而进",
"中定理E4中的条件（b)），这就是为什么使用（6）中",
"行的实验。",
"5"
] |
{
"arxiv_id": "2106.13863",
"language": "zh",
"page_idx": 4
}
|
||
2007.13138-zh-0
|
2007.13138-zh-0.jpg
|
# Virtual Multi-view Fusion for 3D Semantic Segmentation
Abhijit Kundu, Xiaoqi Yin, Alireza Fathi, David Ross,
Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru
Google Research
$\mathbf{Abstract.~Semantic~segmentation~of~3D~meshes~is~an~important~prob-}$ lem for 3D scene understanding. In this paper we revisit the classic multiview representation of 3D meshes and study several techniques that make them effective for 3D semantic segmentation of meshes. Given a 3D mesh reconstructed from RGBD sensors, our method effectively chooses different virtual views of the 3D mesh and renders multiple 2D channels for training an effective 2D semantic segmentation model. Features from multiple per view predictions are finally fused on 3D mesh vertices to predict mesh semantic segmentation labels. Using the large scale indoor 3D semantic segmentation benchmark of ScanNet, we show that our virtual views enable more effective training of 2D semantic segmentation networks than previous multiview approaches. When the 2D per pixel predictions are aggregated on 3D surfaces, our virtual multiview fusion method is able to achieve significantly better 3D semantic segmentation results compared to all prior multiview approaches and competitive with recent 3D convolution approaches.
Keywords: 3D semantic segmentation, Scene Understanding
# 1 Introduction
3D 场景的语义分割是计算机视觉中的一个基本问题。给定场景的 3D 表示 (e. g。, 室内环境的纹理网格), 目标是为每个表面点输出一个语义标签。输出可用于语义映射，站点监控，训练自主导航和其他几个应用程序
用于 3D 语义分割的最新技术(SOTA) 方法目前使用 3D 稀疏体素卷积算子来处理输入数据。例如，MinkowskiNet [7] 和 SparseConvNet [11] 分别将输人数据加载到稀疏 3D 体素网格中，并使用稀疏 3D 卷积提取特征。这些“以地点为中心”的方法旨在识别 3D 图案，因此适用于具有独特 3D 形状的对象类型(例如。g。,椅子),对其他人来说不太好(e。g。,墙图片)。它们还占用大量内存，这限制了空间分辨率和/或批量大小
或者，当可以使用 RGB-D 图像时，一些研究人员尝试使用设计用于处理摄影 RGB 图像的 2D 网络来预测密集特征和/或语义标签，然后将它们聚集在可见的 3D 表面 [15,41] 上，其他人将特征投影到可见表面上并在 3D [10,40,18,19] 中进一步卷积它们。尽管这些“以视图为中心”的方法利用了在大型 RGB 图像数据集上预先训练的大规模图像处理网络，但由于 RGB-D 扫描数据集中的遮挡，光照变化和相机姿态未对准的困难，它们在标准 3D 分割基准上无法实现
|
[
[
0.256591796875,
0.1458740234375,
0.7490234375,
0.146484375,
0.7490234375,
0.1610107421875,
0.256591796875,
0.160400390625
],
[
0.42236328125,
0.1673583984375,
0.5810546875,
0.1685791015625,
0.5810546875,
0.186279296875,
0.42236328125,
0.1849365234375
],
[
0.30810546875,
0.21533203125,
0.69677734375,
0.21533203125,
0.69677734375,
0.2291259765625,
0.30810546875,
0.2291259765625
],
[
0.271240234375,
0.2298583984375,
0.73388671875,
0.2298583984375,
0.73388671875,
0.2442626953125,
0.271240234375,
0.2442626953125
],
[
0.44677734375,
0.2568359375,
0.55810546875,
0.2568359375,
0.55810546875,
0.270751953125,
0.44677734375,
0.270751953125
],
[
0.26708984375,
0.30810546875,
0.73779296875,
0.30810546875,
0.73779296875,
0.3212890625,
0.26708984375,
0.3212890625
],
[
0.2646484375,
0.3212890625,
0.73876953125,
0.3212890625,
0.73876953125,
0.335205078125,
0.2646484375,
0.335205078125
],
[
0.26708984375,
0.337158203125,
0.73779296875,
0.337158203125,
0.73779296875,
0.347900390625,
0.26708984375,
0.347900390625
],
[
0.26708984375,
0.35107421875,
0.73779296875,
0.35107421875,
0.73779296875,
0.361083984375,
0.26708984375,
0.361083984375
],
[
0.26708984375,
0.365478515625,
0.73779296875,
0.365478515625,
0.73779296875,
0.375732421875,
0.26708984375,
0.375732421875
],
[
0.26708984375,
0.37890625,
0.73876953125,
0.37890625,
0.73876953125,
0.389404296875,
0.26708984375,
0.389404296875
],
[
0.265625,
0.392578125,
0.736328125,
0.392578125,
0.736328125,
0.40283203125,
0.265625,
0.40283203125
],
[
0.265625,
0.40478515625,
0.740234375,
0.40478515625,
0.740234375,
0.41845703125,
0.265625,
0.41845703125
],
[
0.265625,
0.42041015625,
0.73779296875,
0.42041015625,
0.73779296875,
0.43115234375,
0.265625,
0.43115234375
],
[
0.2646484375,
0.43310546875,
0.736328125,
0.43310546875,
0.736328125,
0.4462890625,
0.2646484375,
0.4462890625
],
[
0.26708984375,
0.448974609375,
0.736328125,
0.448974609375,
0.736328125,
0.458984375,
0.26708984375,
0.458984375
],
[
0.265625,
0.462158203125,
0.73779296875,
0.462158203125,
0.73779296875,
0.472900390625,
0.265625,
0.472900390625
],
[
0.265625,
0.474853515625,
0.73779296875,
0.474853515625,
0.73779296875,
0.488037109375,
0.265625,
0.488037109375
],
[
0.265625,
0.488037109375,
0.73876953125,
0.488037109375,
0.73876953125,
0.501953125,
0.265625,
0.501953125
],
[
0.26318359375,
0.50048828125,
0.740234375,
0.501953125,
0.740234375,
0.5166015625,
0.26318359375,
0.51513671875
],
[
0.265625,
0.5166015625,
0.49267578125,
0.5166015625,
0.49267578125,
0.52978515625,
0.265625,
0.52978515625
],
[
0.2646484375,
0.54248046875,
0.67333984375,
0.54248046875,
0.67333984375,
0.55615234375,
0.2646484375,
0.55615234375
],
[
0.2181396484375,
0.58544921875,
0.25732421875,
0.58544921875,
0.25732421875,
0.5966796875,
0.2181396484375,
0.5966796875
],
[
0.250732421875,
0.58544921875,
0.378173828125,
0.58544921875,
0.378173828125,
0.5986328125,
0.250732421875,
0.5986328125
],
[
0.21728515625,
0.6142578125,
0.78515625,
0.61572265625,
0.78515625,
0.62890625,
0.21728515625,
0.62744140625
],
[
0.2181396484375,
0.630859375,
0.78515625,
0.62890625,
0.78515625,
0.64306640625,
0.2181396484375,
0.64599609375
],
[
0.2197265625,
0.64599609375,
0.64208984375,
0.64599609375,
0.64208984375,
0.6591796875,
0.2197265625,
0.6591796875
],
[
0.2459716796875,
0.662109375,
0.78173828125,
0.662109375,
0.78173828125,
0.67236328125,
0.2459716796875,
0.67236328125
],
[
0.221435546875,
0.67529296875,
0.78369140625,
0.67529296875,
0.78369140625,
0.689453125,
0.221435546875,
0.689453125
],
[
0.221435546875,
0.68994140625,
0.78515625,
0.68994140625,
0.78515625,
0.7041015625,
0.221435546875,
0.7041015625
],
[
0.2197265625,
0.70556640625,
0.78515625,
0.70556640625,
0.78515625,
0.7197265625,
0.2197265625,
0.7197265625
],
[
0.2181396484375,
0.7197265625,
0.7861328125,
0.72021484375,
0.7861328125,
0.73486328125,
0.2181396484375,
0.734375
],
[
0.221435546875,
0.7373046875,
0.50341796875,
0.7373046875,
0.50341796875,
0.74755859375,
0.221435546875,
0.74755859375
],
[
0.24267578125,
0.75,
0.78515625,
0.75048828125,
0.78515625,
0.76513671875,
0.24267578125,
0.7646484375
],
[
0.2197265625,
0.7666015625,
0.78369140625,
0.7666015625,
0.78369140625,
0.77978515625,
0.2197265625,
0.77978515625
],
[
0.221435546875,
0.78076171875,
0.78515625,
0.78076171875,
0.78515625,
0.79443359375,
0.221435546875,
0.79443359375
],
[
0.2197265625,
0.796875,
0.78515625,
0.796875,
0.78515625,
0.81005859375,
0.2197265625,
0.81005859375
],
[
0.221435546875,
0.81298828125,
0.78369140625,
0.81298828125,
0.78369140625,
0.82373046875,
0.221435546875,
0.82373046875
],
[
0.2181396484375,
0.8271484375,
0.78515625,
0.8271484375,
0.78515625,
0.84033203125,
0.2181396484375,
0.84033203125
]
] |
[
"Virtual Multi-view Fusion for 3D Semantic",
"Segmentation",
"Abhijit Kundu, Xiaoqi Yin, Alireza Fathi, David Ross,",
"Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru",
"Google Research",
"Abstract. Semantic segmentation of 3D meshes is an important prob-",
"lem for 3D scene understanding. In this paper we revisit the classic mul-",
"tiview representation of 3D meshes and study several techniques that",
"makethemeffectivefor3Dsemanticsegmentationofmeshes.Givena3D",
"meshreconstructedfromRGBDsensors,ourmethodeffectivelychooses",
"different virtual views ofthe 3D meshand rendersmultiple2D channels",
"fortraining aneffective2D semantic segmentationmodel.Featuresfrom",
"multiple per view predictions are finally fused on 3D mesh vertices to",
"predict mesh semantic segmentation labels.Using thelarge scale indoor",
"3D semantic segmentation benchmark of ScanNet, we show that our vir",
"tual views enable more effective training of 2D semantic segmentatior",
"networksthanpreviousmultiview approaches:Whenthe2Dper pixel",
"predictions are aggregated on 3D surfaces, our virtual multiview fusion",
"method is able to achieve significantly better 3D semantic segmentation",
"results compared to all prior multiview approaches and competitive with",
"recent 3D convolution approaches:",
"Keywords: 3D semantic segmentation, Scene Understanding",
"1",
"Introduction",
"3D场景的语义分割是计算机视觉中的一个基本问题。给定场景的3D表示（e。",
"g。，室内环境的纹理网格)，目标是为每个表面点输出一个语义标签。输出可用",
"于语义映射，站点监控，训练自主导航和其他几个应用程序",
"用于3D语义分割的最新技术（SOTA）方法自前使用3D稀疏体素卷积算",
"子来处理输人数据。例如，MinkowskiNet[7]和SparseConvNet[11]分别将输",
"入数据加载到稀疏3D体素网格中，并使用稀疏3D卷积提取特征。这些“以地",
"点为中心”的方法旨在识别3D图案，因此适用于具有独特3D形状的对象类型",
"（例如。g。，椅子），对其他人来说不太好（e。g。，墙图片)。它们还占用大量内",
"存，这限制了空间分辨率和/或批量大小",
"或者，当可以使用RGB-D图像时，一些研究人员尝试使用设计用于处理摄",
"影RGB图像的2D网络来预测密集特征和/或语义标签，然后将它们聚集在可",
"见的3D表面「15,411上，其他人将特征投影到可见表面上并在3D「10,40,18,19]",
"中进一步卷积它们。尽管这些“以视图为中心”的方法利用了在大型RGB图像",
"数据集上预先训练的大规模图像处理网络：但由于RGB-D扫描数据集中的遮",
"挡，光照变化和相机姿态未对准的困难，它们在标准3D分割基准上无法实现"
] |
{
"arxiv_id": "2007.13138",
"language": "zh",
"page_idx": 0
}
|
||
2007.13138-zh-1
|
2007.13138-zh-1.jpg
|
2 A. Kundu et al.
SOTA 性能。目前，在 ScanNet 基准测试的 3D 语义标签挑战赛中，没有一种
基于视图的方法位于当前排行榜的上半部分
在本文中，我们提出了一种新的基于视图的 3D 语义分割方法，该方法克服了以前方法的问题。关键思想是使用从 3D 场景的“虚拟视图” 渲染的合成图像，而不是将处理限制为物理相机获取的原始摄影图像。这种方法有几个优点， 可以解决以前以视图为中心的方法[3,21]遇到的关键问题。首先，我们为具有不自然的宽视场的虚拟视图选择相机内在函数，以增加在每个渲染图像中观察到的上下文。其次，我们在与场景表面的距离/角度变化较小，对象之间的遮挡相对较少以及表面覆盖冗余较大的位置选择虚拟视点。第三，我们渲染非真实感图像，而没有与视图相关的光照效果和背面表面的遮挡-i。e。,虚拟视图可以从墙壁、地板和天花板后面查看场景，以提供具有相对较大上下文和较小遮挡的视图。第四，我们根据虚拟视图的确切已知相机参数将像素级预测聚合到 3D 表面上，因此不会在遮挡轮廓上遇到语义标签的“出血”。第五，训练和推理期间的虚拟视图可以模仿多尺度训练和测试，并避免 2DCNN 的尺度方差问题。在训练和测试期间，我们可以生成任意数量的虚拟视图。在训练期间，由于数据增强，更多的虚拟视图提供了鲁棒性。在测试过程中，由于投票冗余，更多视图提供了稳健性。最后，我们的多视图融合方法中的 2D 分割模型可以从 ImageNet 和 COCO 等大型图像预训练数据中受益，这些数据对于纯 3D 卷积方法是不可用的
我们已经研究了使用各种消融研究将虚拟视图用于 3D 表面的语义分割的想
g。,墙后面)。例如，使用具有原始相机参数的虚拟视图将 3DmIoU 提高 3。1% 与使用原始摄影图像相比，使用额外的法线和坐标通道以及更高的视野可以进一步将 mIoU 提升 5。7 个%, 额外增益为 2。1% 可以通过仔细选择虚拟相机姿势来实现，以最好地捕获场景中的 3D 信息并优化训练 2DCNN
总体而言，我们的简单系统能够在 ScanNetBenchmark[F]9] 中实现 2D 和3D 语义标记任务的最新结果，并且明显优于以前性能最好的多视图方法，并且与基于 3D 点集和网格卷积的最新 3D 方法非常有竞争力。此外，我们表明，当训练场景较少时，我们提出的方法始终优于 3D 卷积和真实的多视图融合方法。最后，我们表明，在推理阶段，可以用明显更少的视图获得类似的性能。例如， 每个场景具有？?? 个虚拟视图的多视图融合将优于每个场景具有所有？?? 个原始视图的多视图融合
本文的其余部分组织如下。我们介绍了$\color{red}{82}$的研究概况和相关工作。我们在83-85中详细描述了拟议的虚拟多视图融合方法。§6 中给出了我们提出的方法的实验结果和消融研究。最后，我们通过讨论 §7 的未来方向来结束本文。
# 2 Related Work
之前已经有大量关于 3D 场景语义分割的工作。以下仅回顾最相关的工作
Multi-view labeling.受基于视图的对象分类方法[35]的成功启发，RGB-D 表面重建的语义分割的早期工作依赖于经过训练以预测 RGB 图像的密集语义标签的 2D 网络。通过加权平均[15,41] , CRFs [25] ,贝叶斯融合[24,41,46]或3D 卷积 [10,18,19] , 将像素级语义标签反向投影并聚合到 3D 重建表面上。这些方法仅对原始捕获的 RGB-D 摄影图像执行多视图聚合，这些图像具有有限的视野，有限的视点范围，与视图相关的照明效果以及与重建的表面几何形状
|
[
[
0.21728515625,
0.117431640625,
0.273681640625,
0.117431640625,
0.273681640625,
0.1287841796875,
0.21728515625,
0.1287841796875
],
[
0.26953125,
0.11614990234375,
0.37744140625,
0.11614990234375,
0.37744140625,
0.1300048828125,
0.26953125,
0.1300048828125
],
[
0.2181396484375,
0.1483154296875,
0.78515625,
0.1483154296875,
0.78515625,
0.1622314453125,
0.2181396484375,
0.1622314453125
],
[
0.2181396484375,
0.1641845703125,
0.52783203125,
0.1641845703125,
0.52783203125,
0.177978515625,
0.2181396484375,
0.177978515625
],
[
0.2442626953125,
0.1787109375,
0.78369140625,
0.1787109375,
0.78369140625,
0.1925048828125,
0.2442626953125,
0.1925048828125
],
[
0.221435546875,
0.1932373046875,
0.78515625,
0.1932373046875,
0.78515625,
0.20703125,
0.221435546875,
0.20703125
],
[
0.2197265625,
0.208984375,
0.78369140625,
0.208984375,
0.78369140625,
0.222900390625,
0.2197265625,
0.222900390625
],
[
0.22216796875,
0.2259521484375,
0.78173828125,
0.2259521484375,
0.78173828125,
0.236083984375,
0.22216796875,
0.236083984375
],
[
0.221435546875,
0.2392578125,
0.78515625,
0.2392578125,
0.78515625,
0.25244140625,
0.221435546875,
0.25244140625
],
[
0.2197265625,
0.25390625,
0.78369140625,
0.25390625,
0.78369140625,
0.267578125,
0.2197265625,
0.267578125
],
[
0.2197265625,
0.26953125,
0.78369140625,
0.26953125,
0.78369140625,
0.28271484375,
0.2197265625,
0.28271484375
],
[
0.221435546875,
0.2841796875,
0.78173828125,
0.2841796875,
0.78173828125,
0.29736328125,
0.221435546875,
0.29736328125
],
[
0.2197265625,
0.2998046875,
0.78369140625,
0.2998046875,
0.78369140625,
0.313232421875,
0.2197265625,
0.313232421875
],
[
0.221435546875,
0.31640625,
0.78369140625,
0.31640625,
0.78369140625,
0.326904296875,
0.221435546875,
0.326904296875
],
[
0.221435546875,
0.330078125,
0.78515625,
0.330078125,
0.78515625,
0.343505859375,
0.221435546875,
0.343505859375
],
[
0.2197265625,
0.3447265625,
0.78515625,
0.3447265625,
0.78515625,
0.35791015625,
0.2197265625,
0.35791015625
],
[
0.221435546875,
0.35986328125,
0.78369140625,
0.35986328125,
0.78369140625,
0.373779296875,
0.221435546875,
0.373779296875
],
[
0.2181396484375,
0.373779296875,
0.78369140625,
0.375732421875,
0.78369140625,
0.389404296875,
0.2181396484375,
0.386962890625
],
[
0.2197265625,
0.39013671875,
0.78369140625,
0.39013671875,
0.78369140625,
0.404052734375,
0.2197265625,
0.404052734375
],
[
0.2181396484375,
0.40478515625,
0.78369140625,
0.40478515625,
0.78369140625,
0.41845703125,
0.2181396484375,
0.41845703125
],
[
0.221435546875,
0.42041015625,
0.250732421875,
0.42041015625,
0.250732421875,
0.43310546875,
0.221435546875,
0.43310546875
],
[
0.24267578125,
0.43505859375,
0.78173828125,
0.43505859375,
0.78173828125,
0.448974609375,
0.24267578125,
0.448974609375
],
[
0.2197265625,
0.45068359375,
0.99609375,
0.45068359375,
0.99609375,
0.46533203125,
0.2197265625,
0.46533203125
],
[
0.2181396484375,
0.466552734375,
0.78369140625,
0.464111328125,
0.78369140625,
0.479248046875,
0.2181396484375,
0.48095703125
],
[
0.2197265625,
0.48095703125,
0.78515625,
0.48095703125,
0.78515625,
0.494384765625,
0.2197265625,
0.494384765625
],
[
0.233642578125,
0.496826171875,
0.78515625,
0.496826171875,
0.78515625,
0.51025390625,
0.233642578125,
0.51025390625
],
[
0.2197265625,
0.51220703125,
0.6689453125,
0.51220703125,
0.6689453125,
0.52294921875,
0.2197265625,
0.52294921875
],
[
0.2442626953125,
0.52587890625,
0.78515625,
0.52587890625,
0.78515625,
0.5390625,
0.2442626953125,
0.5390625
],
[
0.2181396484375,
0.54150390625,
0.78515625,
0.54150390625,
0.78515625,
0.5546875,
0.2181396484375,
0.5546875
],
[
0.2197265625,
0.5546875,
0.78369140625,
0.5546875,
0.78369140625,
0.568359375,
0.2197265625,
0.568359375
],
[
0.22216796875,
0.57275390625,
0.78369140625,
0.57275390625,
0.78369140625,
0.58349609375,
0.22216796875,
0.58349609375
],
[
0.2197265625,
0.5859375,
0.78369140625,
0.5859375,
0.78369140625,
0.599609375,
0.2197265625,
0.599609375
],
[
0.2197265625,
0.6015625,
0.78515625,
0.6015625,
0.78515625,
0.61572265625,
0.2197265625,
0.61572265625
],
[
0.2197265625,
0.61572265625,
0.348876953125,
0.6162109375,
0.348876953125,
0.6298828125,
0.2197265625,
0.62890625
],
[
0.2442626953125,
0.6318359375,
0.78369140625,
0.6318359375,
0.78369140625,
0.64599609375,
0.2442626953125,
0.64599609375
],
[
0.21728515625,
0.64599609375,
0.78515625,
0.646484375,
0.78515625,
0.6611328125,
0.21728515625,
0.66015625
],
[
0.2197265625,
0.662109375,
0.74169921875,
0.662109375,
0.74169921875,
0.67529296875,
0.2197265625,
0.67529296875
],
[
0.2181396484375,
0.69970703125,
0.39208984375,
0.69970703125,
0.39208984375,
0.71337890625,
0.2181396484375,
0.71337890625
],
[
0.221435546875,
0.73095703125,
0.73876953125,
0.73095703125,
0.73876953125,
0.7412109375,
0.221435546875,
0.7412109375
],
[
0.2181396484375,
0.75048828125,
0.78515625,
0.75048828125,
0.78515625,
0.7646484375,
0.2181396484375,
0.7646484375
],
[
0.2197265625,
0.7666015625,
0.78369140625,
0.7666015625,
0.78369140625,
0.77978515625,
0.2197265625,
0.77978515625
],
[
0.2181396484375,
0.78076171875,
0.78515625,
0.78076171875,
0.78515625,
0.79443359375,
0.2181396484375,
0.79443359375
],
[
0.2197265625,
0.796875,
0.78515625,
0.796875,
0.78515625,
0.81005859375,
0.2197265625,
0.81005859375
],
[
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
]
] |
[
"2",
"A. Kundu et al.",
"SOTA性能。目前，在ScanNet基准测试的3D语义标签挑战赛中，没有一种",
"基于视图的方法位于当前排行榜的上半部分",
"在本文中，我们提出了一种新的基于视图的3D语义分割方法，该方法克服",
"了以前方法的问题。关键思想是使用从3D场景的“虚拟视图”谊染的合成图",
"像，而不是将处理限制为物理相机获取的原始摄影图像。这种方法有几个优点",
"可以解决以前以视图为中心的方法3，21遇到的关键问题。首先，我们为具有不",
"自然的宽视场的虚拟视图选择相机内在函数，以增加在每个煊染图像中观察到",
"的上下文。其次，我们在与场景表面的距离/角度变化较小，对象之间的遮挡相",
"对较少以及表面覆盖几余较大的位置选择虚拟视点。第三，我们谊染非真实感",
"图像，而没有与视图相关的光照效果和背面表面的遮挡-i。e。，虚拟视图可以从",
"墙壁、地板和天花板后面查看场景，以提供具有相对较大上下文和较小遮挡的",
"视图。第四，我们根据虚拟视图的确切已知相机参数将像素级预测聚合到3D表",
"面上，因此不会在遮挡轮廓上遇到语义标签的“出血”。第五，训练和推理期间",
"的虚拟视图可以模仿多尺度训练和测试，并避免2DCNN的尺度方差问题。在",
"训练和测试期间，我们可以生成任意数量的虚拟视图。在训练期间，由于数据增",
"强，更多的虚拟视图提供了鲁棒性。在测试过程中，由于投票穴余，更多视图提",
"供了稳健性。最后，我们的多视图融合方法中的2D分割模型可以从ImageNet",
"和COCO等大型图像预训练数据中受益，这些数据对于纯3D卷积方法是不可",
"用的",
"我们已经研究了使用各种消融研究将虚拟视图用于3D表面的语义分割的想",
"法。Wefindthatthebroaderdesignspaceofviewselectionenabledbyvirtualcamerascansignificantlyboosttheperforma",
"g。，墙后面）。例如，使用具有原始相机参数的虚拟视图将3DmloU提高3。1%",
"与使用原始摄影图像相比，使用额外的法线和坐标通道以及更高的视野可以进",
"步将mIoU提升5。7个%，额外增益为2。1%可以通过仔细选择虚拟相机姿",
"势来实现，以最好地捕获场景中的3D信息并优化训练2DCNN",
"总体而言，我们的简单系统能够在ScanNetBenchmarkE[9]中实现2D和",
"3D语义标记任务的最新结果，并且明显优于以前性能最好的多视图方法，并且",
"与基于3D点集和网格卷积的最新3D方法非常有竞争力。此外，我们表明，当",
"训练场景较少时，我们提出的方法始终优于3D卷积和真实的多视图融合方法",
"最后，我们表明，在推理阶段，可以用明显更少的视图获得类似的性能。例如",
"每个场景具有？?？个虚拟视图的多视图融合将优于每个场景具有所有？?？个原始",
"视图的多视图融合",
"本文的其余部分组织如下。我们介绍了S2的研究概况和相关工作。我们在",
"S3-S5中详细描述了拟议的虚拟多视图融合方法。S6中给出了我们提出的方法",
"的实验结果和消融研究。最后，我们通过讨论S7的未来方向来结束本文。",
"2Related Work",
"之前已经有大量关于3D场景语义分割的工作。以下仅回顾最相关的工作",
"Multi-viewlabeling.受基于视图的对象分类方法[35]的成功启发，RGB-D",
"表面重建的语义分割的早期工作依赖于经过训练以预测RGB图像的密集语义",
"标签的2D网络。通过加权平均「15,41]，CRFs[25]】，贝叶斯融合[24.41,46]或",
"3D卷积[10,18,19]，将像素级语义标签反向投影并聚合到3D重建表面上。这",
"些方法仅对原始捕获的RGB-D摄影图像执行多视图聚合，这些图像具有有限",
"的视野，有限的视点范围，与视图相关的照明效果以及与重建的表面几何形状"
] |
{
"arxiv_id": "2007.13138",
"language": "zh",
"page_idx": 1
}
|
||
2007.13138-zh-2
|
2007.13138-zh-2.jpg
|
Virtual Multiview Fusion 3
的不对准，所有这些都会降低语义分割性能。为了克服这些问题，最近的一些工作提出了在多视图标记管道[3,21,12]中使用真实数据的合成图像，但它们仍然使用真实图像的典型摄像机参数 (e。g。,小视场),提出仅适用于室外环境 (城市的激光雷达点云)的方法，并且目前没有达到最先进的结果
3D convolution.最近有关 3D 语义分割的工作集中在直接使用 3D 卷积提取和分类特征的方法上。已经提出了网络架构来从 3D 点云[29,30,31,33,38,16] , 表面网格[14,17] ,体素网格 [34] 和八叉树 [32] 中提取特征。当前最先进的方法基于稀疏 3D 体素卷积[7,8,11] ,其中子流形稀疏卷积运算用于计算稀疏体素网格上的特征。这些方法比密集的体素网格更有效地利用内存，但与 2D 图像相比，空间分辨率仍然有限，并且只能在 3D 数据集上进行监督训练，与 2D 图像数据集相比，这些数据集通常非常小
Synthetic data.其他工作研究了使用 3D 合成数据 [47] 的计算机图形渲染来训练 2D 语义分割网络。这种方法的主要优点是可以通过渲染 3D 场景 [22,26] 的新颖视图来创建大小不受限制的图像数据集。然而，挑战通常是域自适应-在合成数据上训练并在真实数据上测试的网络通常表现不佳。我们的方法通过对使用相同过程渲染的合成图像进行训练和测试来避免此问题
<FigureHere>
Fig.1: Virtual multi-view fusion system overview.
3 Method Overview
所提出的多视图融合方法如图 1 所示。在高层次上，它由以下步骤组成
Training stage.在训练阶段，我们首先为每个 3D 场景选择虚拟视图，其中对于每个虚拟视图，我们选择摄像机内在函数，摄像机外在函数，要渲染的通道以及渲染参数 (e。g。、深度范围、背面剔除)。然后，我们通过渲染选定通道和地
|
[
[
0.56396484375,
0.11553955078125,
0.73388671875,
0.117431640625,
0.73388671875,
0.1275634765625,
0.56396484375,
0.1256103515625
],
[
0.77392578125,
0.117431640625,
0.7861328125,
0.117431640625,
0.7861328125,
0.126953125,
0.77392578125,
0.126953125
],
[
0.2197265625,
0.147705078125,
0.78515625,
0.147705078125,
0.78515625,
0.1610107421875,
0.2197265625,
0.1610107421875
],
[
0.221435546875,
0.1634521484375,
0.78515625,
0.1634521484375,
0.78515625,
0.1767578125,
0.221435546875,
0.1767578125
],
[
0.221435546875,
0.177978515625,
0.7861328125,
0.177978515625,
0.7861328125,
0.1925048828125,
0.221435546875,
0.1925048828125
],
[
0.221435546875,
0.1932373046875,
0.66015625,
0.1932373046875,
0.66015625,
0.20703125,
0.221435546875,
0.20703125
],
[
0.2181396484375,
0.2183837890625,
0.78515625,
0.2183837890625,
0.78515625,
0.231689453125,
0.2181396484375,
0.231689453125
],
[
0.2197265625,
0.2342529296875,
0.77392578125,
0.2342529296875,
0.77392578125,
0.2474365234375,
0.2197265625,
0.2474365234375
],
[
0.221435546875,
0.248779296875,
0.78515625,
0.248779296875,
0.78515625,
0.261962890625,
0.221435546875,
0.261962890625
],
[
0.2197265625,
0.26318359375,
0.78369140625,
0.26318359375,
0.78369140625,
0.276611328125,
0.2197265625,
0.276611328125
],
[
0.221435546875,
0.27783203125,
0.78369140625,
0.27783203125,
0.78369140625,
0.291748046875,
0.221435546875,
0.291748046875
],
[
0.221435546875,
0.29345703125,
0.78369140625,
0.29345703125,
0.78369140625,
0.306884765625,
0.221435546875,
0.306884765625
],
[
0.2197265625,
0.309326171875,
0.48193359375,
0.309326171875,
0.48193359375,
0.322509765625,
0.2197265625,
0.322509765625
],
[
0.2197265625,
0.333984375,
0.78369140625,
0.333984375,
0.78369140625,
0.347900390625,
0.2197265625,
0.347900390625
],
[
0.2197265625,
0.347900390625,
0.78173828125,
0.347900390625,
0.78173828125,
0.361083984375,
0.2197265625,
0.361083984375
],
[
0.2197265625,
0.363037109375,
0.78515625,
0.363037109375,
0.78515625,
0.376953125,
0.2197265625,
0.376953125
],
[
0.221435546875,
0.37890625,
0.78515625,
0.37890625,
0.78515625,
0.392578125,
0.221435546875,
0.392578125
],
[
0.221435546875,
0.393310546875,
0.6435546875,
0.393310546875,
0.6435546875,
0.4072265625,
0.221435546875,
0.4072265625
],
[
0.31201171875,
0.43994140625,
0.39794921875,
0.43994140625,
0.39794921875,
0.451904296875,
0.31201171875,
0.451904296875
],
[
0.59326171875,
0.43994140625,
0.68701171875,
0.43994140625,
0.68701171875,
0.451904296875,
0.59326171875,
0.451904296875
],
[
0.5400390625,
0.4609375,
0.56201171875,
0.4609375,
0.56201171875,
0.46533203125,
0.5400390625,
0.46533203125
],
[
0.26953125,
0.468505859375,
0.301513671875,
0.473388671875,
0.298095703125,
0.479736328125,
0.26708984375,
0.474853515625
],
[
0.54736328125,
0.4697265625,
0.58251953125,
0.472900390625,
0.5810546875,
0.48095703125,
0.5458984375,
0.477783203125
],
[
0.426513671875,
0.479248046875,
0.44287109375,
0.479248046875,
0.44287109375,
0.486083984375,
0.426513671875,
0.486083984375
],
[
0.259033203125,
0.496826171875,
0.289306640625,
0.496826171875,
0.289306640625,
0.50390625,
0.259033203125,
0.50390625
],
[
0.53759765625,
0.496826171875,
0.560546875,
0.496826171875,
0.560546875,
0.50390625,
0.53759765625,
0.50390625
],
[
0.303955078125,
0.5234375,
0.336669921875,
0.5234375,
0.336669921875,
0.52783203125,
0.303955078125,
0.52783203125
],
[
0.34716796875,
0.5234375,
0.386474609375,
0.5234375,
0.386474609375,
0.52783203125,
0.34716796875,
0.52783203125
],
[
0.26708984375,
0.60498046875,
0.303955078125,
0.60595703125,
0.303955078125,
0.609375,
0.26708984375,
0.60791015625
],
[
0.5361328125,
0.6015625,
0.59619140625,
0.60400390625,
0.59619140625,
0.61572265625,
0.5361328125,
0.61279296875
],
[
0.6787109375,
0.60693359375,
0.689453125,
0.60693359375,
0.689453125,
0.611328125,
0.6787109375,
0.611328125
],
[
0.6953125,
0.60498046875,
0.73876953125,
0.60498046875,
0.73876953125,
0.6123046875,
0.6953125,
0.6123046875
],
[
0.4599609375,
0.62548828125,
0.53515625,
0.62548828125,
0.53515625,
0.63623046875,
0.4599609375,
0.63623046875
],
[
0.455810546875,
0.6337890625,
0.5439453125,
0.63330078125,
0.5439453125,
0.64453125,
0.455810546875,
0.64599609375
],
[
0.338134765625,
0.66015625,
0.6650390625,
0.66015625,
0.6650390625,
0.673828125,
0.338134765625,
0.673828125
],
[
0.21728515625,
0.7265625,
0.4306640625,
0.7265625,
0.4306640625,
0.74072265625,
0.21728515625,
0.74072265625
],
[
0.2197265625,
0.771484375,
0.73388671875,
0.771484375,
0.73388671875,
0.78515625,
0.2197265625,
0.78515625
],
[
0.2197265625,
0.79541015625,
0.78515625,
0.79541015625,
0.78515625,
0.81005859375,
0.2197265625,
0.81005859375
],
[
0.221435546875,
0.81103515625,
0.78369140625,
0.81103515625,
0.78369140625,
0.82470703125,
0.221435546875,
0.82470703125
],
[
0.2197265625,
0.8271484375,
0.78515625,
0.8271484375,
0.78515625,
0.84033203125,
0.2197265625,
0.84033203125
]
] |
[
"Virtual MultiviewFusion",
"3",
"的不对准，所有这些都会降低语义分割性能。为了克服这些问题，最近的一些工",
"作提出了在多视图标记管道3,21,121中使用真实数据的合成图像，但它们仍然",
"使用真实图像的典型摄像机参数（e。g。，小视场），提出仅适用于室外环境（城",
"市的激光雷达点云）的方法，并且目前没有达到最先进的结果",
"3Dconvolution.最近有关3D语义分割的工作集中在直接使用3D卷积提取",
"和分类特征的方法上。已经提出了网络架构来从3D点云[29,30,31,33,38,16]",
"表面网格14,17]，体素网格[34]和八叉树[32]中提取特征。当前最先进的方法",
"基于稀疏3D体素卷积「7,8,111，其中子流形稀疏卷积运算用于计算稀疏体素网",
"格上的特征。这些方法比密集的体素网格更有效地利用内存，但与2D图像相",
"比，空间分辨率仍然有限，并且只能在3D数据集上进行监督训练，与2D图像",
"数据集相比，这些数据集通常非常小",
"Syntheticdata.其他工作研究了使用3D合成数据【47]的计算机图形染来",
"训练2D语义分割网络。这种方法的主要优点是可以通过染3D场景[2226",
"的新颖视图来创建大小不受限制的图像数据集。然而，挑战通常是域自适应-在",
"合成数据上训练并在真实数据上测试的网络通常表现不佳。我们的方法通过对",
"使用相同过程染的合成图像进行训练和测试来避免此问题",
"Training Phase",
"InferencePhase",
"",
"2",
"",
"",
"",
"",
"Viewsel",
"on&render",
"Taininninn",
"2D Inference inpu",
"20",
"ferenceoutpu",
"Train2D semanti",
"segmentationmodel",
"Fig.l: Virtual multi-view fusion system overview.",
"3Method Overview",
"所提出的多视图融合方法如图1所示。在高层次上，它由以下步骤组成",
"Trainingstage.在训练阶段，我们首先为每个3D场景选择虚拟视图，其中对",
"于每个虚拟视图，我们选择摄像机内在函数，摄像机外在函数，要染的通道以",
"及谊染参数（e。g。、深度范围、背面剔除）。然后，我们通过染选定通道和地"
] |
{
"arxiv_id": "2007.13138",
"language": "zh",
"page_idx": 2
}
|
||
2007.13138-zh-3
|
2007.13138-zh-3.jpg
|
4
A. Kundu et al.
<FigureHere>
Fig.2: Proposed virtual view selection approaches
面实况语义标签的选定虚拟视图来生成训练数据。我们使用渲染的训练数据训
练 2D 语义分割模型，并在推理阶段使用该模型
Inference stage.在推理阶段，我们使用与训练阶段类似的方法选择并渲染虚拟视图，但没有基本事实语义标签。我们使用训练好的模型对渲染的虚拟视图进行 2D 语义分割，将 2D 语义特征投影到 3D,然后通过融合多个投影的 2D 语义特征来导出 3D 中的语义类别
4 Virtual view selection
虚拟视图选择是所提出的多视图融合方法的核心，因为它带来了与原始图像视图的多视图融合相比的关键优势。首先，它允许我们自由选择最适合 2D 语义分割任务的相机参数，以及任何一组 2D 数据增强方法。其次，它通过放松真实相机的物理约束，并允许从不切实际但有用的相机位置查看，从而显著提高模型性能，从而显著拓宽了可供选择的视图集，例如。g。在墙后面。第三，它允许 2D 视图捕获难以用真实相机捕获的其他通道，例如。g。、法线和坐标。最后，通过选择和渲染虚拟视图，我们基本上消除了相机校准和姿态估计中的任何错误这些错误在 3D 重建过程中很常见。最后，在不同尺度上一致地采样视图解决了传统 2DCNN 的尺度方差问题
Camera intrinsics. 原始图像视图的一个重要约束是 FOV-例如，图像可能非常靠近物体或墙壁拍摄，并且缺乏准确分类所需的物体特征和背景。相反，我们使用具有比原始相机高得多的视场 (FOV)的针孔相机模型，提供更大的上下文，从而导致更准确的 2D 语义分割 [27]。图 3 显示了原始视图与高 FOV 虚拟视图的比较示例
Camera extrinsics. 我们混合使用以下采样策略来选择相机外在函数，如图E
2 和图$\mathbb{E}$4 所示。
- Uniform sampling. We want to uniformly sample camera extrinsics to gen-
erate many novel views, independently from the specific structure of the 3D scene. Specifically, we use top-down views from uniformly sampled positions
|
[
[
0.223876953125,
0.11297607421875,
0.2344970703125,
0.121826171875,
0.223876953125,
0.1300048828125,
0.2132568359375,
0.121826171875
],
[
0.27294921875,
0.117431640625,
0.375732421875,
0.117431640625,
0.375732421875,
0.1275634765625,
0.27294921875,
0.1275634765625
],
[
0.28759765625,
0.15283203125,
0.36767578125,
0.15283203125,
0.36767578125,
0.1634521484375,
0.28759765625,
0.1634521484375
],
[
0.57421875,
0.15087890625,
0.62109375,
0.15087890625,
0.62109375,
0.1610107421875,
0.57421875,
0.1610107421875
],
[
0.301513671875,
0.1634521484375,
0.35302734375,
0.1610107421875,
0.35302734375,
0.17236328125,
0.301513671875,
0.1749267578125
],
[
0.483642578125,
0.1634521484375,
0.495849609375,
0.1641845703125,
0.495849609375,
0.1685791015625,
0.483642578125,
0.1673583984375
],
[
0.54150390625,
0.1641845703125,
0.55615234375,
0.1641845703125,
0.55615234375,
0.1685791015625,
0.54150390625,
0.1685791015625
],
[
0.576171875,
0.1622314453125,
0.5986328125,
0.1622314453125,
0.5986328125,
0.17041015625,
0.576171875,
0.17041015625
],
[
0.576171875,
0.17236328125,
0.625,
0.17236328125,
0.625,
0.18310546875,
0.576171875,
0.18310546875
],
[
0.6787109375,
0.169189453125,
0.689453125,
0.169189453125,
0.689453125,
0.17236328125,
0.6787109375,
0.17236328125
],
[
0.416748046875,
0.1767578125,
0.42724609375,
0.1767578125,
0.42724609375,
0.1925048828125,
0.416748046875,
0.1925048828125
],
[
0.473876953125,
0.1767578125,
0.48779296875,
0.1767578125,
0.48779296875,
0.1925048828125,
0.473876953125,
0.1925048828125
],
[
0.53369140625,
0.1767578125,
0.54345703125,
0.1767578125,
0.54345703125,
0.1912841796875,
0.53369140625,
0.1912841796875
],
[
0.640625,
0.18115234375,
0.6806640625,
0.18115234375,
0.6806640625,
0.1912841796875,
0.640625,
0.1912841796875
],
[
0.64453125,
0.1900634765625,
0.66259765625,
0.1875,
0.6650390625,
0.1976318359375,
0.64794921875,
0.199462890625
],
[
0.705078125,
0.186279296875,
0.71728515625,
0.186279296875,
0.71728515625,
0.1900634765625,
0.705078125,
0.1900634765625
],
[
0.281982421875,
0.20263671875,
0.293212890625,
0.20263671875,
0.293212890625,
0.20703125,
0.281982421875,
0.20703125
],
[
0.66650390625,
0.20263671875,
0.68115234375,
0.20263671875,
0.68115234375,
0.205810546875,
0.66650390625,
0.205810546875
],
[
0.68701171875,
0.208984375,
0.76953125,
0.208984375,
0.76953125,
0.2197265625,
0.68701171875,
0.2197265625
],
[
0.68701171875,
0.2171630859375,
0.73876953125,
0.2197265625,
0.73876953125,
0.2310791015625,
0.68701171875,
0.2291259765625
],
[
0.236083984375,
0.248779296875,
0.3056640625,
0.248779296875,
0.3056640625,
0.260009765625,
0.236083984375,
0.260009765625
],
[
0.2344970703125,
0.260009765625,
0.285888671875,
0.260009765625,
0.285888671875,
0.270751953125,
0.2344970703125,
0.270751953125
],
[
0.34716796875,
0.27587890625,
0.37744140625,
0.27587890625,
0.37744140625,
0.2841796875,
0.34716796875,
0.2841796875
],
[
0.3349609375,
0.3359375,
0.6650390625,
0.337158203125,
0.6650390625,
0.35107421875,
0.3349609375,
0.349853515625
],
[
0.221435546875,
0.380615234375,
0.78515625,
0.380615234375,
0.78515625,
0.39453125,
0.221435546875,
0.39453125
],
[
0.2197265625,
0.395751953125,
0.56396484375,
0.395751953125,
0.56396484375,
0.4091796875,
0.2197265625,
0.4091796875
],
[
0.2181396484375,
0.41845703125,
0.78369140625,
0.41845703125,
0.78369140625,
0.431884765625,
0.2181396484375,
0.431884765625
],
[
0.2197265625,
0.43310546875,
0.78173828125,
0.43310546875,
0.78173828125,
0.4462890625,
0.2197265625,
0.4462890625
],
[
0.2197265625,
0.448974609375,
0.78515625,
0.448974609375,
0.78515625,
0.462158203125,
0.2197265625,
0.462158203125
],
[
0.2197265625,
0.46337890625,
0.464111328125,
0.46337890625,
0.464111328125,
0.4765625,
0.2197265625,
0.4765625
],
[
0.21728515625,
0.501953125,
0.46728515625,
0.501953125,
0.46728515625,
0.51513671875,
0.21728515625,
0.51513671875
],
[
0.221435546875,
0.53271484375,
0.78369140625,
0.53271484375,
0.78369140625,
0.54345703125,
0.221435546875,
0.54345703125
],
[
0.221435546875,
0.546875,
0.78369140625,
0.546875,
0.78369140625,
0.56005859375,
0.221435546875,
0.56005859375
],
[
0.22216796875,
0.56298828125,
0.78173828125,
0.56298828125,
0.78173828125,
0.57373046875,
0.22216796875,
0.57373046875
],
[
0.2197265625,
0.5771484375,
0.78369140625,
0.5771484375,
0.78369140625,
0.59033203125,
0.2197265625,
0.59033203125
],
[
0.2197265625,
0.59228515625,
0.78515625,
0.59228515625,
0.78515625,
0.60595703125,
0.2197265625,
0.60595703125
],
[
0.2197265625,
0.60791015625,
0.78515625,
0.60791015625,
0.78515625,
0.6220703125,
0.2197265625,
0.6220703125
],
[
0.2181396484375,
0.62060546875,
0.78515625,
0.6220703125,
0.78515625,
0.63623046875,
0.2181396484375,
0.63525390625
],
[
0.2181396484375,
0.63525390625,
0.78369140625,
0.63623046875,
0.78369140625,
0.65087890625,
0.2181396484375,
0.6494140625
],
[
0.2197265625,
0.6513671875,
0.4345703125,
0.6513671875,
0.4345703125,
0.66552734375,
0.2197265625,
0.66552734375
],
[
0.2197265625,
0.6748046875,
0.78515625,
0.6748046875,
0.78515625,
0.68798828125,
0.2197265625,
0.68798828125
],
[
0.221435546875,
0.689453125,
0.7861328125,
0.689453125,
0.7861328125,
0.70263671875,
0.221435546875,
0.70263671875
],
[
0.2197265625,
0.705078125,
0.78369140625,
0.705078125,
0.78369140625,
0.71826171875,
0.2197265625,
0.71826171875
],
[
0.221435546875,
0.7197265625,
0.78515625,
0.7197265625,
0.78515625,
0.73291015625,
0.221435546875,
0.73291015625
],
[
0.2181396484375,
0.734375,
0.351318359375,
0.734375,
0.351318359375,
0.74755859375,
0.2181396484375,
0.74755859375
],
[
0.2181396484375,
0.7568359375,
0.7861328125,
0.7568359375,
0.7861328125,
0.77099609375,
0.2181396484375,
0.77099609375
],
[
0.2181396484375,
0.77294921875,
0.338134765625,
0.77294921875,
0.338134765625,
0.7861328125,
0.2181396484375,
0.7861328125
],
[
0.2279052734375,
0.796875,
0.78173828125,
0.796875,
0.78173828125,
0.81005859375,
0.2279052734375,
0.81005859375
],
[
0.2459716796875,
0.81103515625,
0.78515625,
0.81103515625,
0.78515625,
0.82470703125,
0.2459716796875,
0.82470703125
],
[
0.2459716796875,
0.8271484375,
0.78515625,
0.8271484375,
0.78515625,
0.84033203125,
0.2459716796875,
0.84033203125
]
] |
[
"N",
"A.Kunduetal",
"Scale-invariant",
"Uniform",
"selection",
"",
"",
"view",
"selection",
"",
"OH",
"",
"OH",
"pullsink",
"any",
"",
"",
"",
"Class-balanced",
"selection",
"Originalview",
"selection",
"",
"Fig.2: Proposed virtual view selection approaches",
"面实况语义标签的选定虚拟视图来生成训练数据。我们使用染的训练数据训",
"练2D语义分割模型，并在推理阶段使用该模型",
"Inference stage.在推理阶段，我们使用与训练阶段类似的方法选择并谊染虚",
"拟视图，但没有基本事实语义标签。我们使用训练好的模型对染的虚拟视图",
"进行2D语义分割，将2D语义特征投影到3D，然后通过融合多个投影的2D",
"语义特征来导出3D中的语义类别",
"4Virtualview selection",
"虚拟视图选择是所提出的多视图融合方法的核心，因为它带来了与原始图像视",
"图的多视图融合相比的关键优势。首先，它允许我们自由选择最适合2D语义分",
"割任务的相机参数，以及任何一组2D数据增强方法。其次，它通过放松真实相",
"机的物理约束，并允许从不切实际但有用的相机位置查看，从而显著提高模型性",
"能，从而显著拓宽了可供选择的视图集，例如。g。在墙后面。第三，它允许2D",
"视图捕获难以用真实相机捕获的其他通道，例如。g。、法线和坐标。最后，通",
"过选择和谊染虚拟视图，我们基本上消除了相机校准和姿态估计中的任何错误，",
"这些错误在3D重建过程中很常见。最后，在不同尺度上一致地采样视图解决了",
"传统2DCNN的尺度方差问题",
"Cameraintrinsics.原始图像视图的一个重要约束是FOV-例如，图像可能非",
"常靠近物体或墙壁拍摄，并且缺乏准确分类所需的物体特征和背景。相反，我们",
"使用具有比原始相机高得多的视场（FOV）的针孔相机模型，提供更大的上下",
"文，从而导致更准确的2D语义分割[27]。图3显示了原始视图与高FOV虚",
"拟视图的比较示例",
"Cameraextrinsics.我们混合使用以下采样策略来选择相机外在函数，如图E",
"2 和图E4 所示。",
"- Uniform sampling. We want to uniformly sample camera extrinsics to gen",
"erate many novel views, independently from the specific structure of the 3D",
"scene. Specifically, we use top-down views from uniformly sampled positions"
] |
{
"arxiv_id": "2007.13138",
"language": "zh",
"page_idx": 3
}
|
||
2007.13138-zh-4
|
2007.13138-zh-4.jpg
|
5
Virtual Multiview Fusion
<FigureHere>
Fig.3: Original views vs. virtual views. High FOV provides larger context of the scene which helps 2D perception, e.g., the chair in the bottom right corner is partially represented in the original view but can easily segmented in the high FOV virtual view.
<FigureHere>
Fig.4: Example virtual view selection on two ScanNet scenes. Green curve is the trajectory of the original camera poses; Blue cameras are the selected views with proposed approaches. Note that we only show a random subset of all selected views for illustration purposes.
at the top of the 3D scene, as well as views that look through the center of
the scene but with uniformly sampled positions in the 3D scene.
- Scale- invariant sampling. As 2D convolutional neural networks are generally not scale invariant, the model performance may suffer if the scales of views do not match the 3D scene. To overcome this limitation, we propose sampling views at a range of scales with respect to segments in the 3D scene. Specifically, we do an over-segmentation of the 3D scene, and for each segment, we position the cameras to look at the segment by pulling back to a certain range of distances along the normal direction. We do a depth check to avoid occlusions by foreground objects. If backface culling is disabled in the rendering stage (discussed in more detail below), we do a ray tracing and drop any views blocked by the backfaces. Note the over-segmentation of the 3D scene is unsupervised and does not use the ground truth semantic
|
[
[
0.56396484375,
0.11553955078125,
0.73388671875,
0.11553955078125,
0.73388671875,
0.1287841796875,
0.56396484375,
0.1287841796875
],
[
0.77392578125,
0.11871337890625,
0.7861328125,
0.11871337890625,
0.7861328125,
0.126953125,
0.77392578125,
0.126953125
],
[
0.27294921875,
0.1458740234375,
0.343017578125,
0.1458740234375,
0.343017578125,
0.1572265625,
0.27294921875,
0.1572265625
],
[
0.43310546875,
0.1458740234375,
0.495849609375,
0.1458740234375,
0.495849609375,
0.1572265625,
0.43310546875,
0.1572265625
],
[
0.59228515625,
0.146484375,
0.7294921875,
0.146484375,
0.7294921875,
0.1572265625,
0.59228515625,
0.1572265625
],
[
0.259033203125,
0.1622314453125,
0.271240234375,
0.1622314453125,
0.271240234375,
0.1666259765625,
0.259033203125,
0.1666259765625
],
[
0.2181396484375,
0.260009765625,
0.7861328125,
0.260009765625,
0.7861328125,
0.2734375,
0.2181396484375,
0.2734375
],
[
0.2197265625,
0.27587890625,
0.78173828125,
0.27587890625,
0.78173828125,
0.285888671875,
0.2197265625,
0.285888671875
],
[
0.2156982421875,
0.287353515625,
0.77783203125,
0.285888671875,
0.77783203125,
0.2998046875,
0.2156982421875,
0.300537109375
],
[
0.2181396484375,
0.54150390625,
0.78369140625,
0.54150390625,
0.78369140625,
0.5546875,
0.2181396484375,
0.5546875
],
[
0.2181396484375,
0.55615234375,
0.78759765625,
0.55615234375,
0.78759765625,
0.5693359375,
0.2181396484375,
0.5693359375
],
[
0.21728515625,
0.57080078125,
0.78369140625,
0.568359375,
0.78369140625,
0.58203125,
0.21728515625,
0.583984375
],
[
0.2181396484375,
0.58544921875,
0.31201171875,
0.58544921875,
0.31201171875,
0.5966796875,
0.2181396484375,
0.5966796875
],
[
0.2467041015625,
0.63720703125,
0.78515625,
0.63720703125,
0.78515625,
0.64794921875,
0.2467041015625,
0.64794921875
],
[
0.2442626953125,
0.65087890625,
0.705078125,
0.65087890625,
0.705078125,
0.6640625,
0.2442626953125,
0.6640625
],
[
0.2320556640625,
0.6748046875,
0.78515625,
0.6748046875,
0.78515625,
0.689453125,
0.2320556640625,
0.689453125
],
[
0.2459716796875,
0.68994140625,
0.7861328125,
0.68994140625,
0.7861328125,
0.7041015625,
0.2459716796875,
0.7041015625
],
[
0.2459716796875,
0.70556640625,
0.78515625,
0.70556640625,
0.78515625,
0.7197265625,
0.2459716796875,
0.7197265625
],
[
0.2467041015625,
0.72021484375,
0.78515625,
0.72021484375,
0.78515625,
0.734375,
0.2467041015625,
0.734375
],
[
0.2442626953125,
0.73486328125,
0.78369140625,
0.736328125,
0.78369140625,
0.75,
0.2442626953125,
0.74853515625
],
[
0.2459716796875,
0.75048828125,
0.78759765625,
0.75048828125,
0.78759765625,
0.7646484375,
0.2459716796875,
0.7646484375
],
[
0.2467041015625,
0.767578125,
0.78515625,
0.767578125,
0.78515625,
0.77783203125,
0.2467041015625,
0.77783203125
],
[
0.2459716796875,
0.78076171875,
0.7861328125,
0.78076171875,
0.7861328125,
0.79443359375,
0.2459716796875,
0.79443359375
],
[
0.2467041015625,
0.796875,
0.78759765625,
0.796875,
0.78759765625,
0.81005859375,
0.2467041015625,
0.81005859375
],
[
0.2459716796875,
0.81103515625,
0.7861328125,
0.81103515625,
0.7861328125,
0.82470703125,
0.2459716796875,
0.82470703125
],
[
0.2459716796875,
0.8271484375,
0.7861328125,
0.8271484375,
0.7861328125,
0.84033203125,
0.2459716796875,
0.84033203125
]
] |
[
"Virtual MultiviewFusion",
"5",
"Originalview",
"Virtual view",
"Virtual viewwithhighFOV",
"",
"Fig. 3: Original views vs. virtual views. High FOV provides larger context of the scene",
"which helps 2D perception, e.g., the chair in the bottom right corner is partially rep",
"resented in the original viewbut can easily segmented in the high FOV virtual view.",
"Fig. 4: Example virtual view selection on two ScanNet scenes. Green curve is the tra-",
"jectory of the original camera poses; Blue cameras are the selected views with proposed",
"approaches. Note that we only show a random subset of all selected views for illustra-",
"tion purposes.",
"at the top of the 3D scene. as well as views that look through the center of",
"the scene but with uniformly sampled positions in the 3D scene.",
"- Scale-invariant sampling. As 2D convolutional neural networks are generally",
"not scale invariant, the model performance may suffer if the scales of views",
"do not match the 3D scene. To overcome this limitation, we propose sam-",
"pling views at a range of scales with respect to segments in the 3D scene.",
"Specifically, we do an over-segmentation of the 3D scene, and for each seg-",
"ment, we position the cameras to look at the segment by pulling back to a",
"certain range of distances along the normal direction.We do a depth check",
"to avoid occlusions by foreground objects. If backface culling is disabled in",
"the rendering stage (discussed in more detail below), we do a ray tracing",
"and drop any views blocked by the backfaces. Note the over-segmentation",
"of the 3D scene is unsupervised and does not use the ground truth semantic"
] |
{
"arxiv_id": "2007.13138",
"language": "zh",
"page_idx": 4
}
|
||
1804.05038-zh-1
|
1804.05038-zh-1.jpg
|
,其中$g$是嵌入前一个目标字$(y_t-1)$的两层前馈网络，解码器隐藏状态(XMATHLINX< XMAS
我们用两层 GRU 将$s_t$计算为
$$\begin{matrix}s'_t=r(s_{t-1},y_{t-1}^*)\end{matrix}$$
(3)
和
(4)
$$s_t=q(s_t',H_t),$$
,其中$s_t^\prime$是中间状态$,s_0=\overleftarrow{h}_0$ 。两个 GRU 单元$r$和$q$与注意力一起构成 Sennrich et al. (2017)的条件 GRU 层。$H_t$计算为
$$H_t=\begin{bmatrix}\sum_{i=1}^l(\alpha_{t,i}\cdot\stackrel{\leftarrow}{h}_i)\\\sum_{i=1}^l(\alpha_{t,i}\cdot\stackrel{\rightharpoonup}{h}_i)\end{bmatrix},$$
(5)
,其中$\alpha_{t,i}$ 是$\alpha_t$的元素$,\alpha_t$ 是注意力模型的
输出向量。这是通过两层前馈网络
$$\alpha_t'=v(\tanh(w(h_i)+u(s_{t-1}'))),$$
(6)
一个矩阵，每个编码器状态$h_i$产生一个实值。
$\alpha_t$是$\alpha_t^\prime$上的 softmax
计算的，其中$w$和$u$是权重矩阵，而$v$是另我们使用使用 Theano 框架 (Bastien et al.,
2012) 编写的程序来训练模型。通常，模型的批量大小范围为 64 到 128, 并且无偏 Adam 随机优化器 (Kingma and Ba, 2014)。我们使用 620 的嵌入大小和 1000 的隐藏层大小。我们根据超过 10 个时期的延迟开发集上的最佳BLEU 分数选择模型参数
## 3 8-bit Translatiom
我们的翻译引擎是一个 C++ 实现。该引擎使用 Eigen 矩阵库实现，该矩阵库提供高效的矩阵运算。每个 CPU 内核一次翻译一个句子。同一引擎支持批处理和交互式应用程序，后者使单句翻译延迟变得很重要。我们将速度数字报告为每秒字数(WPS)和每秒字数(WPCS), 即 WPS 除以运行的内核数。这为我们提供了跨许多内核和内存总线的总体扩展以及单句速度的度量
基于短语的 SMT 系统，例如(Tillmann, 2006),用于在 20 核 Xeon2690v2 系统上以每秒 170 个单词 (每秒 3400 个单词)的速度运行英语-德语。同样，基于语法的 SMT 系统， 如 (Zhao and Al-onaizan, 2008) , 对于相同的语言对运行在 21。每秒 5 个单词(每秒 430个单词)
相比之下，我们的具有 32 位解码的 NMT 系统(在 2 节中进行了描述)在 6。每秒 5 个单词 (每秒 131 个单词)。我们的目标是将 NMT
系统的解码速度提高到基于短语的系统可以实现的水平，同时保持 NMT 提供的流畅性和充分性
我们的 NMT 解码器的基准毫不奇怪地将矩阵乘法显示为计算周期的头号来源。在表 1中，我们看到在 Eigen 的矩阵和向量乘法例程(Eigen 矩阵向量乘积和 Eigen 矩阵乘法)中花费了超过 85 个% 的计算。它使超越函数计算的成本以及偏差相加相形见绌
Table 1: Profile before 8-bit conversion. More than 85 % is spent in Eigen matrix/vector multiply routines.
鉴于计算时间的这种分布，尝试尽可能加速矩阵运算是有意义的。提高速度的一种方法是量化矩阵运算。在神经网络中用 8 位整数逼近代替 32 位浮点数学运算已被证明可以提高速度和类似的精度 (Vanhoucke et al., 2011)。我们选择将类似的优化应用于翻译系统，以减少内存流量并提高 CPU 的并行性
我们的 8 位矩阵乘法例程使用无阻塞或复制的朴素实现。该代码使用英特尔 SSE4 向量指令实现，一次计算 4 行，类似于 (Devlin, 2017)。简单性导致实现 8 位矩阵乘法，结果被放入 32 位浮点结果。这样做的好处是不需要知道结果的规模。此外，输出是向量或窄矩阵，因此消耗的额外内存带宽很少
多层矩阵乘法算法的性能比朴素算法 (Goto and Geijn, 2008) 快得多。这是由于以下事实： 在乘以 NxN 矩阵时，$O(N^2)$元素上有$O(N^3)$ 数学运算，因此值得付出巨大的努力来最小化内存操作，同时最大化数学运算。但是，当将 NxN 矩阵乘以 NxP 矩阵时，其中 P 非常小(???),内存操作占主导地位，并且性能不会从复杂算法中受益。在对单个句子进行解码时，我们通常会按照这种系统(Koehn and Knowles, 2017) 中的标准做法将光束大小设置为小于 8 的值。我们实际上发现，在 P 值如此小的情况下，朴素算法要快一点
表 2 显示了将矩阵例程转换为 8 位整数计算后的配置文件。矩阵矩阵和矩阵向量乘法只有一个条目，因为它们由相同的例程处理。转换后，tanh 和 sigmoid 仍然消耗不到 7% 的CPU 时间。鉴于这一事实，我们决定不将这些操作转换为整数
|
[
[
0.1376953125,
0.07659912109375,
0.489501953125,
0.07659912109375,
0.489501953125,
0.09027099609375,
0.1376953125,
0.09027099609375
],
[
0.496337890625,
0.08013916015625,
0.51318359375,
0.08013916015625,
0.51318359375,
0.0872802734375,
0.496337890625,
0.0872802734375
],
[
0.51123046875,
0.07598876953125,
0.880859375,
0.07598876953125,
0.880859375,
0.08905029296875,
0.51123046875,
0.08905029296875
],
[
0.118408203125,
0.0926513671875,
0.489501953125,
0.0914306640625,
0.489501953125,
0.105712890625,
0.118408203125,
0.10687255859375
],
[
0.486083984375,
0.08074951171875,
0.490234375,
0.08074951171875,
0.490234375,
0.08612060546875,
0.486083984375,
0.08612060546875
],
[
0.5146484375,
0.09320068359375,
0.88330078125,
0.09320068359375,
0.88330078125,
0.10687255859375,
0.5146484375,
0.10687255859375
],
[
0.12005615234375,
0.10986328125,
0.235107421875,
0.10986328125,
0.235107421875,
0.12054443359375,
0.12005615234375,
0.12054443359375
],
[
0.515625,
0.10986328125,
0.5517578125,
0.10986328125,
0.5517578125,
0.1217041015625,
0.515625,
0.1217041015625
],
[
0.1376953125,
0.1258544921875,
0.3837890625,
0.1258544921875,
0.3837890625,
0.1395263671875,
0.1376953125,
0.1395263671875
],
[
0.5322265625,
0.1258544921875,
0.8818359375,
0.1258544921875,
0.8818359375,
0.1395263671875,
0.5322265625,
0.1395263671875
],
[
0.515625,
0.141357421875,
0.8818359375,
0.141357421875,
0.8818359375,
0.155029296875,
0.515625,
0.155029296875
],
[
0.235107421875,
0.152587890625,
0.372802734375,
0.15380859375,
0.372802734375,
0.168701171875,
0.235107421875,
0.16748046875
],
[
0.46337890625,
0.155029296875,
0.491943359375,
0.155029296875,
0.491943359375,
0.16748046875,
0.46337890625,
0.16748046875
],
[
0.515625,
0.1591796875,
0.880859375,
0.1591796875,
0.880859375,
0.169189453125,
0.515625,
0.169189453125
],
[
0.5146484375,
0.17333984375,
0.88330078125,
0.17333984375,
0.88330078125,
0.18701171875,
0.5146484375,
0.18701171875
],
[
0.12005615234375,
0.1817626953125,
0.1427001953125,
0.1817626953125,
0.1427001953125,
0.196533203125,
0.12005615234375,
0.196533203125
],
[
0.5146484375,
0.188232421875,
0.880859375,
0.188232421875,
0.880859375,
0.201904296875,
0.5146484375,
0.201904296875
],
[
0.2476806640625,
0.19775390625,
0.359375,
0.19775390625,
0.359375,
0.2120361328125,
0.2476806640625,
0.2120361328125
],
[
0.46337890625,
0.198974609375,
0.489501953125,
0.198974609375,
0.489501953125,
0.2120361328125,
0.46337890625,
0.2120361328125
],
[
0.51318359375,
0.2037353515625,
0.755859375,
0.204833984375,
0.755859375,
0.2197265625,
0.51318359375,
0.218505859375
],
[
0.1351318359375,
0.2203369140625,
0.489501953125,
0.2203369140625,
0.489501953125,
0.2381591796875,
0.1351318359375,
0.2381591796875
],
[
0.5517578125,
0.2340087890625,
0.59765625,
0.2340087890625,
0.59765625,
0.2452392578125,
0.5517578125,
0.2452392578125
],
[
0.7001953125,
0.232177734375,
0.77490234375,
0.2340087890625,
0.77490234375,
0.248779296875,
0.7001953125,
0.2464599609375
],
[
0.12005615234375,
0.2392578125,
0.486083984375,
0.2381591796875,
0.486083984375,
0.2529296875,
0.12005615234375,
0.254150390625
],
[
0.5439453125,
0.2493896484375,
0.62109375,
0.2493896484375,
0.62109375,
0.26123046875,
0.5439453125,
0.26123046875
],
[
0.6171875,
0.2493896484375,
0.85546875,
0.2493896484375,
0.85546875,
0.26318359375,
0.6171875,
0.26318359375
],
[
0.12005615234375,
0.255859375,
0.40478515625,
0.255859375,
0.40478515625,
0.26953125,
0.12005615234375,
0.26953125
],
[
0.54248046875,
0.26318359375,
0.60546875,
0.265380859375,
0.603515625,
0.27978515625,
0.54150390625,
0.27783203125
],
[
0.646484375,
0.265380859375,
0.8251953125,
0.265380859375,
0.8251953125,
0.279052734375,
0.646484375,
0.279052734375
],
[
0.5517578125,
0.279052734375,
0.60546875,
0.28076171875,
0.603515625,
0.29638671875,
0.5498046875,
0.29443359375
],
[
0.6572265625,
0.281982421875,
0.81689453125,
0.281982421875,
0.81689453125,
0.295654296875,
0.6572265625,
0.295654296875
],
[
0.208251953125,
0.295654296875,
0.246826171875,
0.295654296875,
0.246826171875,
0.30810546875,
0.208251953125,
0.30810546875
],
[
0.252685546875,
0.283935546875,
0.393798828125,
0.28076171875,
0.393798828125,
0.3017578125,
0.252685546875,
0.3046875
],
[
0.462646484375,
0.295654296875,
0.48779296875,
0.295654296875,
0.48779296875,
0.31005859375,
0.462646484375,
0.31005859375
],
[
0.5517578125,
0.297607421875,
0.60546875,
0.297607421875,
0.60546875,
0.312255859375,
0.5517578125,
0.312255859375
],
[
0.67236328125,
0.297607421875,
0.80029296875,
0.29638671875,
0.80029296875,
0.311279296875,
0.67236328125,
0.312255859375
],
[
0.25439453125,
0.303955078125,
0.391357421875,
0.302734375,
0.391357421875,
0.32177734375,
0.25439453125,
0.322509765625
],
[
0.55322265625,
0.3134765625,
0.60791015625,
0.3134765625,
0.60791015625,
0.327880859375,
0.55322265625,
0.327880859375
],
[
0.654296875,
0.312255859375,
0.81787109375,
0.314208984375,
0.81787109375,
0.328857421875,
0.654296875,
0.32666015625
],
[
0.1351318359375,
0.336181640625,
0.48779296875,
0.336181640625,
0.48779296875,
0.349853515625,
0.1351318359375,
0.349853515625
],
[
0.51318359375,
0.339111328125,
0.5810546875,
0.341552734375,
0.5810546875,
0.352783203125,
0.51318359375,
0.350830078125
],
[
0.59765625,
0.342529296875,
0.83642578125,
0.342529296875,
0.83642578125,
0.352783203125,
0.59765625,
0.352783203125
],
[
0.8369140625,
0.342529296875,
0.880859375,
0.342529296875,
0.880859375,
0.352783203125,
0.8369140625,
0.352783203125
],
[
0.12005615234375,
0.350830078125,
0.39208984375,
0.350830078125,
0.39208984375,
0.364013671875,
0.12005615234375,
0.364013671875
],
[
0.1275634765625,
0.34033203125,
0.1368408203125,
0.34033203125,
0.1368408203125,
0.344482421875,
0.1275634765625,
0.344482421875
],
[
0.51318359375,
0.35400390625,
0.8818359375,
0.355224609375,
0.8818359375,
0.369384765625,
0.51318359375,
0.3681640625
],
[
0.5146484375,
0.37060546875,
0.580078125,
0.37060546875,
0.580078125,
0.3818359375,
0.5146484375,
0.3818359375
],
[
0.1796875,
0.37890625,
0.462646484375,
0.38134765625,
0.462646484375,
0.3955078125,
0.1796875,
0.393798828125
],
[
0.45751953125,
0.385498046875,
0.46337890625,
0.385498046875,
0.46337890625,
0.391357421875,
0.45751953125,
0.391357421875
],
[
0.4609375,
0.3818359375,
0.486083984375,
0.3818359375,
0.486083984375,
0.393798828125,
0.4609375,
0.393798828125
],
[
0.5341796875,
0.396728515625,
0.8818359375,
0.396728515625,
0.8818359375,
0.410400390625,
0.5341796875,
0.410400390625
],
[
0.12005615234375,
0.4091796875,
0.486083984375,
0.4091796875,
0.486083984375,
0.4228515625,
0.12005615234375,
0.4228515625
],
[
0.5146484375,
0.412109375,
0.8818359375,
0.412109375,
0.8818359375,
0.42578125,
0.5146484375,
0.42578125
],
[
0.1351318359375,
0.427001953125,
0.491943359375,
0.427001953125,
0.491943359375,
0.43701171875,
0.1351318359375,
0.43701171875
],
[
0.5146484375,
0.427490234375,
0.8818359375,
0.4287109375,
0.8818359375,
0.4423828125,
0.5146484375,
0.441162109375
],
[
0.118408203125,
0.441162109375,
0.297119140625,
0.441162109375,
0.297119140625,
0.454833984375,
0.118408203125,
0.454833984375
],
[
0.5146484375,
0.44482421875,
0.880859375,
0.44482421875,
0.880859375,
0.457763671875,
0.5146484375,
0.457763671875
],
[
0.1368408203125,
0.457275390625,
0.48779296875,
0.457275390625,
0.48779296875,
0.47021484375,
0.1368408203125,
0.47021484375
],
[
0.5146484375,
0.460205078125,
0.8818359375,
0.458984375,
0.8818359375,
0.473876953125,
0.5146484375,
0.474365234375
],
[
0.12005615234375,
0.473876953125,
0.48779296875,
0.473876953125,
0.48779296875,
0.48681640625,
0.12005615234375,
0.48681640625
],
[
0.5146484375,
0.476806640625,
0.8818359375,
0.476806640625,
0.8818359375,
0.490478515625,
0.5146484375,
0.490478515625
],
[
0.12005615234375,
0.4892578125,
0.48779296875,
0.4892578125,
0.48779296875,
0.5029296875,
0.12005615234375,
0.5029296875
],
[
0.515625,
0.493408203125,
0.77099609375,
0.493408203125,
0.77099609375,
0.50732421875,
0.515625,
0.50732421875
],
[
0.118408203125,
0.50439453125,
0.48779296875,
0.5048828125,
0.48779296875,
0.51953125,
0.118408203125,
0.5185546875
],
[
0.53076171875,
0.51025390625,
0.8818359375,
0.51025390625,
0.8818359375,
0.52392578125,
0.53076171875,
0.52392578125
],
[
0.12005615234375,
0.52099609375,
0.489501953125,
0.51953125,
0.489501953125,
0.53369140625,
0.12005615234375,
0.53515625
],
[
0.515625,
0.525390625,
0.8818359375,
0.525390625,
0.8818359375,
0.5390625,
0.515625,
0.5390625
],
[
0.118408203125,
0.5361328125,
0.48779296875,
0.5361328125,
0.48779296875,
0.5498046875,
0.118408203125,
0.5498046875
],
[
0.515625,
0.5419921875,
0.8818359375,
0.5419921875,
0.8818359375,
0.5556640625,
0.515625,
0.5556640625
],
[
0.12005615234375,
0.55419921875,
0.32568359375,
0.55419921875,
0.32568359375,
0.56787109375,
0.12005615234375,
0.56787109375
],
[
0.515625,
0.55810546875,
0.8818359375,
0.55810546875,
0.8818359375,
0.57177734375,
0.515625,
0.57177734375
],
[
0.515625,
0.57373046875,
0.8818359375,
0.57373046875,
0.8818359375,
0.58740234375,
0.515625,
0.58740234375
],
[
0.118408203125,
0.58056640625,
0.301513671875,
0.58056640625,
0.301513671875,
0.59423828125,
0.118408203125,
0.59423828125
],
[
0.5146484375,
0.58837890625,
0.8818359375,
0.58837890625,
0.8818359375,
0.6015625,
0.5146484375,
0.6015625
],
[
0.12005615234375,
0.60546875,
0.489501953125,
0.60546875,
0.489501953125,
0.619140625,
0.12005615234375,
0.619140625
],
[
0.5146484375,
0.60546875,
0.791015625,
0.60546875,
0.791015625,
0.619140625,
0.5146484375,
0.619140625
],
[
0.118408203125,
0.62060546875,
0.48779296875,
0.62158203125,
0.48779296875,
0.6357421875,
0.118408203125,
0.63525390625
],
[
0.5341796875,
0.62353515625,
0.8818359375,
0.62353515625,
0.8818359375,
0.63427734375,
0.5341796875,
0.63427734375
],
[
0.12176513671875,
0.63818359375,
0.486083984375,
0.63818359375,
0.486083984375,
0.6484375,
0.12176513671875,
0.6484375
],
[
0.51318359375,
0.63818359375,
0.88330078125,
0.63818359375,
0.88330078125,
0.65185546875,
0.51318359375,
0.65185546875
],
[
0.130126953125,
0.65283203125,
0.48779296875,
0.65283203125,
0.48779296875,
0.66650390625,
0.130126953125,
0.66650390625
],
[
0.51318359375,
0.65185546875,
0.8818359375,
0.65283203125,
0.8818359375,
0.66748046875,
0.51318359375,
0.66650390625
],
[
0.1259765625,
0.65771484375,
0.1368408203125,
0.65771484375,
0.1368408203125,
0.662109375,
0.1259765625,
0.662109375
],
[
0.118408203125,
0.66943359375,
0.486083984375,
0.66943359375,
0.486083984375,
0.68310546875,
0.118408203125,
0.68310546875
],
[
0.5146484375,
0.67041015625,
0.8818359375,
0.67041015625,
0.8818359375,
0.68408203125,
0.5146484375,
0.68408203125
],
[
0.12005615234375,
0.68603515625,
0.490234375,
0.68603515625,
0.490234375,
0.69970703125,
0.12005615234375,
0.69970703125
],
[
0.515625,
0.68603515625,
0.8818359375,
0.68603515625,
0.8818359375,
0.69970703125,
0.515625,
0.69970703125
],
[
0.118408203125,
0.70068359375,
0.486083984375,
0.70068359375,
0.486083984375,
0.71435546875,
0.118408203125,
0.71435546875
],
[
0.515625,
0.70361328125,
0.880859375,
0.70361328125,
0.880859375,
0.71435546875,
0.515625,
0.71435546875
],
[
0.12005615234375,
0.71875,
0.489501953125,
0.71875,
0.489501953125,
0.732421875,
0.12005615234375,
0.732421875
],
[
0.51318359375,
0.71728515625,
0.880859375,
0.71728515625,
0.880859375,
0.73095703125,
0.51318359375,
0.73095703125
],
[
0.12176513671875,
0.736328125,
0.193115234375,
0.736328125,
0.193115234375,
0.74755859375,
0.12176513671875,
0.74755859375
],
[
0.51708984375,
0.7353515625,
0.8818359375,
0.7353515625,
0.8818359375,
0.7490234375,
0.51708984375,
0.7490234375
],
[
0.1376953125,
0.75048828125,
0.489501953125,
0.75048828125,
0.489501953125,
0.76416015625,
0.1376953125,
0.76416015625
],
[
0.515625,
0.75048828125,
0.8818359375,
0.75048828125,
0.8818359375,
0.76416015625,
0.515625,
0.76416015625
],
[
0.12176513671875,
0.76708984375,
0.486083984375,
0.76708984375,
0.486083984375,
0.77783203125,
0.12176513671875,
0.77783203125
],
[
0.5146484375,
0.76708984375,
0.87890625,
0.76708984375,
0.87890625,
0.77783203125,
0.5146484375,
0.77783203125
],
[
0.118408203125,
0.78076171875,
0.48779296875,
0.7822265625,
0.48779296875,
0.79638671875,
0.118408203125,
0.7958984375
],
[
0.515625,
0.783203125,
0.8818359375,
0.783203125,
0.8818359375,
0.79638671875,
0.515625,
0.79638671875
],
[
0.12005615234375,
0.798828125,
0.490234375,
0.798828125,
0.490234375,
0.8125,
0.12005615234375,
0.8125
],
[
0.51318359375,
0.798828125,
0.7900390625,
0.798828125,
0.7900390625,
0.8125,
0.51318359375,
0.8125
],
[
0.12176513671875,
0.81640625,
0.486083984375,
0.81640625,
0.486083984375,
0.82666015625,
0.12176513671875,
0.82666015625
],
[
0.5322265625,
0.81640625,
0.880859375,
0.81640625,
0.880859375,
0.82666015625,
0.5322265625,
0.82666015625
],
[
0.118408203125,
0.83203125,
0.489501953125,
0.83203125,
0.489501953125,
0.845703125,
0.118408203125,
0.845703125
],
[
0.515625,
0.8330078125,
0.880859375,
0.8330078125,
0.880859375,
0.84326171875,
0.515625,
0.84326171875
],
[
0.12005615234375,
0.8486328125,
0.183837890625,
0.8486328125,
0.183837890625,
0.85986328125,
0.12005615234375,
0.85986328125
],
[
0.5146484375,
0.84716796875,
0.8818359375,
0.84716796875,
0.8818359375,
0.86083984375,
0.5146484375,
0.86083984375
],
[
0.1376953125,
0.86328125,
0.489501953125,
0.86328125,
0.489501953125,
0.876953125,
0.1376953125,
0.876953125
],
[
0.51123046875,
0.86328125,
0.880859375,
0.86328125,
0.880859375,
0.876953125,
0.51123046875,
0.876953125
],
[
0.118408203125,
0.87890625,
0.489501953125,
0.87744140625,
0.489501953125,
0.892578125,
0.118408203125,
0.8935546875
],
[
0.51123046875,
0.87744140625,
0.8818359375,
0.87890625,
0.8818359375,
0.8935546875,
0.51123046875,
0.892578125
],
[
0.12005615234375,
0.8955078125,
0.489501953125,
0.8955078125,
0.489501953125,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.5146484375,
0.8955078125,
0.6630859375,
0.8955078125,
0.6630859375,
0.9091796875,
0.5146484375,
0.9091796875
]
] |
[
"其中g是嵌人前一个目标字（yt-1）的",
"",
"系统的解码速度提高到基于短语的系统可以实",
"两层前馈网络，解码器隐藏状态（XMATH-",
"",
"现的水平，同时保持NMT提供的流畅性和充",
"LINX<XMAS",
"分性",
"我们用两层GRU将St计算为",
"我们的NMT解码器的基准毫不奇怪地将",
"矩阵乘法显示为计算周期的头号来源。在表1",
"st = r(st-1, yt-1)",
"(3) ",
"中，我们看到在Eigen的矩阵和向量乘法例程",
"（Eigen矩阵向量乘积和Eigen矩阵乘法）中花",
"和",
"费了超过85个%的计算。它使超越函数计算",
"St = q(st, Ht),",
"(4)",
"的成本以及偏差相加相形见细",
"其中st 是中间状态，So=ho。两个GRU",
"Time",
"Function",
"单元r和q与注意力一起构成Sennrichetal.",
"50.77%",
" Eigen matrix vector product",
"(2017）的条件GRU层。Ht计算为",
"35.02%",
"Eigen matrix multiply",
"1.95%",
"NMT decoder layer",
"Ht =",
"[Ei-1(at,i. h)]",
"(5)",
"1.68%",
"Eigen fast tanh",
"[i1(at,i.hi)",
"1.35%",
"NMT tanh wrapper",
"其中αt,i是Qt的元素，αt是注意力模型的",
"Table 1:",
"Profile before 8-bit conversion.",
"More",
"输出向量。这是通过两层前馈网络",
"",
"than 85 % is spent in Eigen matrix/vector multiply",
"routines.",
"of = v(tanh(w(hi) + u(st-1))),",
"",
"(6)",
"鉴于计算时间的这种分布，尝试尽可能加速",
"计算的，其中w和u是权重矩阵，而u是另",
"矩阵运算是有意义的。提高速度的一种方法是",
"个矩阵：每个编码器状态h产生一个实值。",
"量化矩阵运算。在神经网络中用8位整数逼近",
"Qt是at上的softmax",
"代替32位浮点数学运算已被证明可以提高速",
"我们使用使用Theano框架（Bastienetal.",
"度和类似的精度（Vanhouckeetal.，2011）。我",
"2012）编写的程序来训练模型。通常，模型的",
"们选择将类似的优化应用于翻译系统，以减少",
"批量大小范围为64到128，并且无偏Adam",
"内存流量并提高CPU的并行性",
"随机优化器（KingmaandBa，2014）。我们使",
"我们的8位矩阵乘法例程使用无阻塞或复",
"用620的嵌人大小和1000的隐藏层大小。我",
"制的朴素实现。该代码使用英特尔SSE4向",
"们根据超过10个时期的延迟开发集上的最佳",
"量指令实现，一次计算4行，类似于（Devlin，",
"BLEU分数选择模型参数",
"2017）。简单性导致实现8位矩阵乘法，结果",
"被放人32位浮点结果。这样做的好处是不需",
"38-bit Translation",
"要知道结果的规模。此外，输出是向量或窄矩",
"我们的翻译引擎是一个C+十实现。该引擎使",
"阵，因此消耗的额外内存带宽很少",
"用Eigen矩阵库实现，该矩阵库提供高效的矩",
"多层矩阵乘法算法的性能比朴素算法（Goto",
"阵运算。每个CPU内核一次翻译一个句子。同",
"andGeijn，2008）快得多。这是由于以下事实：",
"引擎支持批处理和交互式应用程序，后者使",
"在乘以NxN矩阵时，O(N2)元素上有O(N3)",
"",
"单句翻译延迟变得很重要。我们将速度数字报",
"数学运算，因此值得付出巨大的努力来最小",
"告为每秒字数（WPS）和每秒字数（WPCS），",
"化内存操作，同时最大化数学运算。但是，当",
"即WPS除以运行的内核数。这为我们提供了",
"将NxN矩阵乘以NxP矩阵时，其中P非常",
"跨许多内核和内存总线的总体扩展以及单句速",
"小（？?？），内存操作占主导地位，并且性能",
"度的度量",
"不会从复杂算法中受益。在对单个句子进行解",
"基于短语的SMT系统，例如（Tillmann，",
"码时，我们通常会按照这种系统（Koehnand",
"2006）用于在20核Xe0n2690v2系统上以每",
"Knowles.2017）中的标准做法将光束大小设置",
"秒170个单词（每秒3400个单词）的速度运",
"为小于8的值。我们实际上发现，在P值如",
"行英语-德语。同样，基于语法的SMT系统，",
"此小的情况下，朴素算法要快一点",
"如（ZhaoandAl-onaizan，2008），对于相同的",
"表2显示了将矩阵例程转换为8位整数计",
"语言对运行在21。每秒5个单词（每秒430",
"算后的配置文件。矩阵矩阵和矩阵向量乘法只",
"个单词）",
"有一个条目，因为它们由相同的例程处理。转",
"相比之下，我们的具有32位解码的NMT",
"换后，tanh和 sigmoid仍然消耗不到7%的",
"系统（在2节中进行了描述）在6。每秒5个单",
"CPU时间。鉴于这一事实，我们决定不将这",
"词（每秒131个单词）。我们的目标是将NMT",
"些操作转换为整数"
] |
{
"arxiv_id": "1804.05038",
"language": "zh",
"page_idx": 1
}
|
||
1804.05038-zh-3
|
1804.05038-zh-3.jpg
|
总吞吐量只是 WPCS 和机器中内核数量的乘 表 8 显示了来自德语-英语系统的两个系统的积。报告的速度是 9 次运行的中位数，以确保 输出的几个示例。表 9 显示了来自荷兰语-英一致的数字。结果表明，与 32 位浮点解码相 语系统的另外 2 个示例
比，我们看到了 4-6 倍的加速比。德语-英语显示 8 位模式与 32 位模式的最大赤字。德语-英语测试集仅包含 168 个句子，因此这可能是虚假的差异
## 4.2 Human evaluation
这些自动结果表明，可以进行 8 位量化而不会出现可察觉的退化。为了证实这一点，我们进行了人体评估实验
在表 7 中，我们显示了在上一节中对某些相同语言对进行人工评估的结果。以英语为母语的独立母语人士(他也精通英语)对 100 个随机选择的句子进行了评分。在评估过程中对句子进行了改组，以避免评估者偏向不同的运行方式。我们采用从 0 到 5 的等级，0 是难以理解的，5 是完美的翻译。
Table 7: Human evaluation scores for 8-bit and
32-bit systems. All tests are news domain.
该表显示，上一节中显示的自动分数也由人类维持。根据人类评估者的说法，8 位解码与32 位解码一样好
### 5 Discussion
在不损失精度的情况下拥有更快的 NMT 引擎在商业上很有用。在我们的部署方案中，它是缓慢的交互式用户体验和不。即使在批处理模式下运行，也可以用1/4的硬件提供相同的吞吐量
此外，这种加速使部署小型模型集合变得实用。如上表 6 中的 En-De 模型所示，合奏可以以 2 倍的减速为代价提供更高的精度。这项工作使翻译质量更高，同时速度至少是以前基线的两倍。
正如 4 节中报告的数字所证明的那样，8 位和 32 位解码具有相似的平均质量。正如预期的那样，两个解码器产生的输出不相同。实际上，在 166 句 De-En 翻译中，两者之间只有51 句是相同的。此外，我们的人工评估结果和自动评分表明，与 32 位解码器相比，8 位解码器没有特定的降级。为了强调这些权利要求，
通常，由于 8 位解码，在充分性或流畅性方面存在细微差异，而不会造成任何损失。表 8中的句子 2 显示了由于重新组装不正确的子字单元而导致的 32 位输出中的拼写错误 (“预测”)。1
## 6 Related Work
减少一般解码神经网络，特别是神经机器翻译
所需的资源，近年来一直是人们关注的焦点
Vanhoucke et al. (2011) 探索了用于语音识别的具有 8 位整数解码的加速卷积神经网络。他们证明了可以使用低精度计算而不会造成明显的精度损失。Han et al.(2015) 研究了使用网络修剪，量化和霍夫曼编码的高度压缩图像分类神经网络，以便完全适合片上缓存，在保持准确性损失的同时，速度和能源效率显着提高
专注于机器翻译，Devlin (2017)实现了 16位定点整数数学运算，以加快矩阵乘法运算， 看到一个 2。59 倍的改进。他们在 WMT 英语-法语 NewsTest2014上显示出具有竞争力的 BLEU 分数，同时提供了显着的加速。同样，(Wu et al., 2016) 在翻译模型中应用 8 位端到端量化。他们还表明，自动指标不会因此而受到影响。在这项工作中，量化需要修改模型训练以限制矩阵输出的大小
7 Conclusions and Future Work
在本文中，我们证明了神经机器翻译的 8 位解码比类似的优化浮点实现快 4-6 倍。我们证明了这种近似的质量与 32 位版本的质量相似。我们还表明，没有必要修改训练过程以产生与8 位解码兼容的模型
总而言之，本文表明，在自动测量和从人类感知的角度来看，8 位解码与 32 位解码一样好，同时它大大改善了延迟
将来，我们计划实现多层矩阵乘法，该算法可以追溯到矩阵面板乘法的朴素算法。这将为可以利用它的应用程序提供批量解码的速度。我们还计划探索低精度的训练，以加快实验周转时间
我们的结果提供了提高准确性的提示，而不仅仅是奇偶校验。其他工作已将培训用作压缩过程的一部分。我们想看看训练量化模型是否会改变结果的好坏
$^{1}$In order to limit the vocabulary, we use BPE sub-
word units (Sennrich et al., 2016) in all models.
|
[
[
0.12176513671875,
0.07659912109375,
0.489501953125,
0.07659912109375,
0.489501953125,
0.09027099609375,
0.12176513671875,
0.09027099609375
],
[
0.490234375,
0.08013916015625,
0.515625,
0.08013916015625,
0.515625,
0.0872802734375,
0.490234375,
0.0872802734375
],
[
0.51123046875,
0.07598876953125,
0.880859375,
0.07598876953125,
0.880859375,
0.08905029296875,
0.51123046875,
0.08905029296875
],
[
0.12005615234375,
0.09320068359375,
0.489501953125,
0.09320068359375,
0.489501953125,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.5146484375,
0.09320068359375,
0.8818359375,
0.09320068359375,
0.8818359375,
0.10687255859375,
0.5146484375,
0.10687255859375
],
[
0.1318359375,
0.10986328125,
0.489501953125,
0.10986328125,
0.489501953125,
0.12353515625,
0.1318359375,
0.12353515625
],
[
0.515625,
0.11102294921875,
0.7021484375,
0.11102294921875,
0.7021484375,
0.1217041015625,
0.515625,
0.1217041015625
],
[
0.12176513671875,
0.1258544921875,
0.489501953125,
0.1258544921875,
0.489501953125,
0.1395263671875,
0.12176513671875,
0.1395263671875
],
[
0.1275634765625,
0.114013671875,
0.1368408203125,
0.114013671875,
0.1368408203125,
0.1175537109375,
0.1275634765625,
0.1175537109375
],
[
0.53076171875,
0.1258544921875,
0.8818359375,
0.1258544921875,
0.8818359375,
0.1395263671875,
0.53076171875,
0.1395263671875
],
[
0.12005615234375,
0.141357421875,
0.489501953125,
0.141357421875,
0.489501953125,
0.155029296875,
0.12005615234375,
0.155029296875
],
[
0.5146484375,
0.141357421875,
0.88330078125,
0.141357421875,
0.88330078125,
0.155029296875,
0.5146484375,
0.155029296875
],
[
0.12005615234375,
0.157958984375,
0.486083984375,
0.157958984375,
0.486083984375,
0.168701171875,
0.12005615234375,
0.168701171875
],
[
0.515625,
0.157958984375,
0.8818359375,
0.157958984375,
0.8818359375,
0.171630859375,
0.515625,
0.171630859375
],
[
0.118408203125,
0.171630859375,
0.19482421875,
0.17333984375,
0.19482421875,
0.188232421875,
0.118408203125,
0.1859130859375
],
[
0.5146484375,
0.174560546875,
0.880859375,
0.174560546875,
0.880859375,
0.185302734375,
0.5146484375,
0.185302734375
],
[
0.515625,
0.191162109375,
0.57666015625,
0.191162109375,
0.57666015625,
0.203125,
0.515625,
0.203125
],
[
0.118408203125,
0.20068359375,
0.313232421875,
0.20068359375,
0.313232421875,
0.21435546875,
0.118408203125,
0.21435546875
],
[
0.5146484375,
0.221435546875,
0.52880859375,
0.221435546875,
0.52880859375,
0.2298583984375,
0.5146484375,
0.2298583984375
],
[
0.52294921875,
0.218505859375,
0.6689453125,
0.21728515625,
0.6689453125,
0.23095703125,
0.52294921875,
0.232177734375
],
[
0.12005615234375,
0.2392578125,
0.489501953125,
0.2392578125,
0.489501953125,
0.2529296875,
0.12005615234375,
0.2529296875
],
[
0.12005615234375,
0.22265625,
0.48779296875,
0.22265625,
0.48779296875,
0.236328125,
0.12005615234375,
0.236328125
],
[
0.5146484375,
0.2423095703125,
0.88330078125,
0.24462890625,
0.88330078125,
0.25830078125,
0.5146484375,
0.255859375
],
[
0.12005615234375,
0.254638671875,
0.268798828125,
0.254638671875,
0.268798828125,
0.268310546875,
0.12005615234375,
0.268310546875
],
[
0.515625,
0.260009765625,
0.86376953125,
0.260009765625,
0.86376953125,
0.273681640625,
0.515625,
0.273681640625
],
[
0.1368408203125,
0.271484375,
0.48779296875,
0.271484375,
0.48779296875,
0.28515625,
0.1368408203125,
0.28515625
],
[
0.5322265625,
0.276611328125,
0.88330078125,
0.276611328125,
0.88330078125,
0.290283203125,
0.5322265625,
0.290283203125
],
[
0.118408203125,
0.2861328125,
0.48779296875,
0.2861328125,
0.48779296875,
0.2998046875,
0.118408203125,
0.2998046875
],
[
0.515625,
0.292236328125,
0.8818359375,
0.292236328125,
0.8818359375,
0.305908203125,
0.515625,
0.305908203125
],
[
0.12005615234375,
0.303955078125,
0.486083984375,
0.303955078125,
0.486083984375,
0.314208984375,
0.12005615234375,
0.314208984375
],
[
0.5146484375,
0.30810546875,
0.880859375,
0.30810546875,
0.880859375,
0.32177734375,
0.5146484375,
0.32177734375
],
[
0.12005615234375,
0.319580078125,
0.48779296875,
0.319580078125,
0.48779296875,
0.333251953125,
0.12005615234375,
0.333251953125
],
[
0.51708984375,
0.32666015625,
0.880859375,
0.32666015625,
0.880859375,
0.33740234375,
0.51708984375,
0.33740234375
],
[
0.12005615234375,
0.3349609375,
0.48779296875,
0.336181640625,
0.48779296875,
0.349853515625,
0.12005615234375,
0.3486328125
],
[
0.515625,
0.341552734375,
0.8818359375,
0.341552734375,
0.8818359375,
0.355224609375,
0.515625,
0.355224609375
],
[
0.12005615234375,
0.3515625,
0.489501953125,
0.3515625,
0.489501953125,
0.365234375,
0.12005615234375,
0.365234375
],
[
0.5146484375,
0.35693359375,
0.88330078125,
0.35693359375,
0.88330078125,
0.37060546875,
0.5146484375,
0.37060546875
],
[
0.12176513671875,
0.369384765625,
0.307373046875,
0.369384765625,
0.307373046875,
0.380126953125,
0.12176513671875,
0.380126953125
],
[
0.5146484375,
0.372314453125,
0.88330078125,
0.372314453125,
0.88330078125,
0.385986328125,
0.5146484375,
0.385986328125
],
[
0.197265625,
0.3955078125,
0.41064453125,
0.3955078125,
0.41064453125,
0.413330078125,
0.197265625,
0.413330078125
],
[
0.515625,
0.39013671875,
0.53662109375,
0.39013671875,
0.53662109375,
0.402587890625,
0.515625,
0.402587890625
],
[
0.5341796875,
0.405029296875,
0.88330078125,
0.405029296875,
0.88330078125,
0.419189453125,
0.5341796875,
0.419189453125
],
[
0.2149658203125,
0.412109375,
0.267822265625,
0.412109375,
0.267822265625,
0.427001953125,
0.2149658203125,
0.427001953125
],
[
0.306396484375,
0.413330078125,
0.341796875,
0.413330078125,
0.341796875,
0.42578125,
0.306396484375,
0.42578125
],
[
0.365234375,
0.413330078125,
0.4072265625,
0.413330078125,
0.4072265625,
0.427001953125,
0.365234375,
0.427001953125
],
[
0.515625,
0.421630859375,
0.88330078125,
0.421630859375,
0.88330078125,
0.435302734375,
0.515625,
0.435302734375
],
[
0.2149658203125,
0.4287109375,
0.267822265625,
0.4287109375,
0.267822265625,
0.443603515625,
0.2149658203125,
0.443603515625
],
[
0.306396484375,
0.4287109375,
0.341796875,
0.4287109375,
0.341796875,
0.441162109375,
0.306396484375,
0.441162109375
],
[
0.366943359375,
0.429931640625,
0.405517578125,
0.429931640625,
0.405517578125,
0.4423828125,
0.366943359375,
0.4423828125
],
[
0.21240234375,
0.44482421875,
0.268798828125,
0.446533203125,
0.267822265625,
0.46142578125,
0.210693359375,
0.458984375
],
[
0.515625,
0.439453125,
0.880859375,
0.439453125,
0.880859375,
0.449462890625,
0.515625,
0.449462890625
],
[
0.306396484375,
0.446533203125,
0.341796875,
0.446533203125,
0.341796875,
0.458984375,
0.306396484375,
0.458984375
],
[
0.364501953125,
0.4453125,
0.408935546875,
0.4453125,
0.408935546875,
0.460205078125,
0.364501953125,
0.460205078125
],
[
0.515625,
0.45361328125,
0.8818359375,
0.45361328125,
0.8818359375,
0.46728515625,
0.515625,
0.46728515625
],
[
0.3046875,
0.46142578125,
0.341796875,
0.46142578125,
0.341796875,
0.476806640625,
0.3046875,
0.476806640625
],
[
0.364501953125,
0.460205078125,
0.4072265625,
0.4619140625,
0.405517578125,
0.47802734375,
0.36279296875,
0.4755859375
],
[
0.515625,
0.4697265625,
0.8818359375,
0.4697265625,
0.8818359375,
0.482666015625,
0.515625,
0.482666015625
],
[
0.20068359375,
0.4619140625,
0.2822265625,
0.4619140625,
0.2822265625,
0.476806640625,
0.20068359375,
0.476806640625
],
[
0.20068359375,
0.47802734375,
0.2822265625,
0.47802734375,
0.2822265625,
0.4921875,
0.20068359375,
0.4921875
],
[
0.3046875,
0.478515625,
0.343505859375,
0.478515625,
0.343505859375,
0.4912109375,
0.3046875,
0.4912109375
],
[
0.364501953125,
0.478515625,
0.405517578125,
0.478515625,
0.405517578125,
0.4921875,
0.364501953125,
0.4921875
],
[
0.515625,
0.486328125,
0.8818359375,
0.486328125,
0.8818359375,
0.49951171875,
0.515625,
0.49951171875
],
[
0.19140625,
0.5048828125,
0.489501953125,
0.5048828125,
0.489501953125,
0.5185546875,
0.19140625,
0.5185546875
],
[
0.5146484375,
0.501953125,
0.880859375,
0.501953125,
0.880859375,
0.515625,
0.5146484375,
0.515625
],
[
0.515625,
0.5185546875,
0.8818359375,
0.5185546875,
0.8818359375,
0.5322265625,
0.515625,
0.5322265625
],
[
0.12176513671875,
0.505859375,
0.198974609375,
0.505859375,
0.198974609375,
0.51708984375,
0.12176513671875,
0.51708984375
],
[
0.118408203125,
0.51953125,
0.4306640625,
0.51953125,
0.4306640625,
0.533203125,
0.118408203125,
0.533203125
],
[
0.515625,
0.53369140625,
0.755859375,
0.53369140625,
0.755859375,
0.54736328125,
0.515625,
0.54736328125
],
[
0.1376953125,
0.55029296875,
0.48779296875,
0.55029296875,
0.48779296875,
0.56396484375,
0.1376953125,
0.56396484375
],
[
0.12005615234375,
0.56640625,
0.489501953125,
0.56640625,
0.489501953125,
0.580078125,
0.12005615234375,
0.580078125
],
[
0.51318359375,
0.5625,
0.8095703125,
0.5625,
0.8095703125,
0.576171875,
0.51318359375,
0.576171875
],
[
0.1175537109375,
0.58056640625,
0.255126953125,
0.58203125,
0.255126953125,
0.5966796875,
0.1175537109375,
0.595703125
],
[
0.5146484375,
0.58740234375,
0.880859375,
0.58740234375,
0.880859375,
0.60107421875,
0.5146484375,
0.60107421875
],
[
0.1158447265625,
0.609375,
0.246826171875,
0.6123046875,
0.246826171875,
0.626953125,
0.1158447265625,
0.62353515625
],
[
0.515625,
0.60400390625,
0.8818359375,
0.60400390625,
0.8818359375,
0.61767578125,
0.515625,
0.61767578125
],
[
0.515625,
0.619140625,
0.8818359375,
0.619140625,
0.8818359375,
0.6328125,
0.515625,
0.6328125
],
[
0.12005615234375,
0.63720703125,
0.489501953125,
0.63720703125,
0.489501953125,
0.65087890625,
0.12005615234375,
0.65087890625
],
[
0.5146484375,
0.63720703125,
0.87890625,
0.63720703125,
0.87890625,
0.64794921875,
0.5146484375,
0.64794921875
],
[
0.12176513671875,
0.65478515625,
0.486083984375,
0.65478515625,
0.486083984375,
0.6650390625,
0.12176513671875,
0.6650390625
],
[
0.51318359375,
0.65283203125,
0.67822265625,
0.65283203125,
0.67822265625,
0.66650390625,
0.51318359375,
0.66650390625
],
[
0.12005615234375,
0.66943359375,
0.48779296875,
0.66943359375,
0.48779296875,
0.68310546875,
0.12005615234375,
0.68310546875
],
[
0.53564453125,
0.66845703125,
0.88330078125,
0.66845703125,
0.88330078125,
0.681640625,
0.53564453125,
0.681640625
],
[
0.118408203125,
0.68408203125,
0.486083984375,
0.68408203125,
0.486083984375,
0.69775390625,
0.118408203125,
0.69775390625
],
[
0.5146484375,
0.68505859375,
0.88330078125,
0.68505859375,
0.88330078125,
0.69873046875,
0.5146484375,
0.69873046875
],
[
0.12005615234375,
0.70166015625,
0.156982421875,
0.70166015625,
0.156982421875,
0.71435546875,
0.12005615234375,
0.71435546875
],
[
0.51708984375,
0.703125,
0.73291015625,
0.703125,
0.73291015625,
0.71337890625,
0.51708984375,
0.71337890625
],
[
0.1376953125,
0.71875,
0.48779296875,
0.71875,
0.48779296875,
0.732421875,
0.1376953125,
0.732421875
],
[
0.53076171875,
0.71630859375,
0.88330078125,
0.71875,
0.88330078125,
0.732421875,
0.53076171875,
0.72998046875
],
[
0.12176513671875,
0.73388671875,
0.486083984375,
0.73388671875,
0.486083984375,
0.74755859375,
0.12176513671875,
0.74755859375
],
[
0.5146484375,
0.73291015625,
0.8818359375,
0.73388671875,
0.8818359375,
0.74755859375,
0.5146484375,
0.74658203125
],
[
0.12176513671875,
0.751953125,
0.486083984375,
0.751953125,
0.486083984375,
0.7626953125,
0.12176513671875,
0.7626953125
],
[
0.5146484375,
0.75,
0.8818359375,
0.75,
0.8818359375,
0.76318359375,
0.5146484375,
0.76318359375
],
[
0.12005615234375,
0.7666015625,
0.48779296875,
0.7666015625,
0.48779296875,
0.77978515625,
0.12005615234375,
0.77978515625
],
[
0.515625,
0.76708984375,
0.880859375,
0.76708984375,
0.880859375,
0.77783203125,
0.515625,
0.77783203125
],
[
0.12005615234375,
0.783203125,
0.198974609375,
0.783203125,
0.198974609375,
0.79443359375,
0.12005615234375,
0.79443359375
],
[
0.51318359375,
0.78076171875,
0.57080078125,
0.78076171875,
0.57080078125,
0.7958984375,
0.51318359375,
0.7958984375
],
[
0.1376953125,
0.798828125,
0.48779296875,
0.798828125,
0.48779296875,
0.8125,
0.1376953125,
0.8125
],
[
0.5322265625,
0.798828125,
0.8818359375,
0.798828125,
0.8818359375,
0.8125,
0.5322265625,
0.8125
],
[
0.12005615234375,
0.8154296875,
0.48779296875,
0.8154296875,
0.48779296875,
0.8291015625,
0.12005615234375,
0.8291015625
],
[
0.5146484375,
0.81396484375,
0.8818359375,
0.81396484375,
0.8818359375,
0.82763671875,
0.5146484375,
0.82763671875
],
[
0.12176513671875,
0.8330078125,
0.48779296875,
0.8330078125,
0.48779296875,
0.84326171875,
0.12176513671875,
0.84326171875
],
[
0.515625,
0.83203125,
0.880859375,
0.83203125,
0.880859375,
0.8427734375,
0.515625,
0.8427734375
],
[
0.118408203125,
0.84716796875,
0.484375,
0.84716796875,
0.484375,
0.86083984375,
0.118408203125,
0.86083984375
],
[
0.515625,
0.8466796875,
0.6630859375,
0.8466796875,
0.6630859375,
0.85986328125,
0.515625,
0.85986328125
],
[
0.12005615234375,
0.86328125,
0.48779296875,
0.86328125,
0.48779296875,
0.876953125,
0.12005615234375,
0.876953125
],
[
0.12176513671875,
0.8798828125,
0.489501953125,
0.8798828125,
0.489501953125,
0.8935546875,
0.12176513671875,
0.8935546875
],
[
0.515625,
0.88525390625,
0.52294921875,
0.88525390625,
0.52294921875,
0.8896484375,
0.515625,
0.8896484375
],
[
0.5322265625,
0.8818359375,
0.8818359375,
0.88427734375,
0.8818359375,
0.8984375,
0.5322265625,
0.896484375
],
[
0.12005615234375,
0.8955078125,
0.495361328125,
0.8955078125,
0.495361328125,
0.9091796875,
0.12005615234375,
0.9091796875
],
[
0.50390625,
0.90087890625,
0.51123046875,
0.90087890625,
0.51123046875,
0.90380859375,
0.50390625,
0.90380859375
],
[
0.51123046875,
0.896484375,
0.8369140625,
0.896484375,
0.8369140625,
0.91015625,
0.51123046875,
0.91015625
]
] |
[
"总吞吐量只是WPCS和机器中内核数量的乘",
"",
"表8显示了来自德语-英语系统的两个系统的",
"积。报告的速度是9次运行的中位数，以确保",
"输出的几个示例。表9显示了来自荷兰语-英",
"致的数字。结果表明，与32位浮点解码相",
"语系统的另外2个示例",
"比，我们看到了4-6倍的加速比。德语-英语显",
"",
"通常，由于8位解码，在充分性或流畅性方",
"示8位模式与32位模式的最大赤字。德语-英",
"面存在细微差异，而不会造成任何损失。表8",
"语测试集仅包含168个句子，因此这可能是虚",
"中的句子2显示了由于重新组装不正确的子",
"假的差异",
"字单元而导致的32位输出中的拼写错误（“预",
"测”）。1",
"4.2Human evaluation",
"6",
"RelatedWork",
"出现可察觉的退化。为了证实这一点，我们进",
"这些自动结果表明，可以进行8位量化而不会",
"减少一般解码神经网络，特别是神经机器翻译",
"行了人体评估实验",
"所需的资源，近年来一直是人们关注的焦点",
"在表7中，我们显示了在上一节中对某些相",
"Vanhoucke etal.（2011）探索了用于语音识",
"同语言对进行人工评估的结果。以英语为母语",
"别的具有8位整数解码的加速卷积神经网络。",
"的独立母语人十（他也精通英语）对100个随",
"他们证明了可以使用低精度计算而不会造成明",
"机选择的句子进行了评分。在评估过程中对句",
"显的精度损失。Hanetal.（2015）研究了使用",
"子进行了改组，以避免评估者偏向不同的运行",
"网络修剪，量化和霍夫曼编码的高度压缩图像",
"方式。我们采用从0到5的等级，0是难以理",
"分类神经网络，以便完全适合片上缓存，在保",
"解的，5是完美的翻译。",
"持准确性损失的同时，速度和能源效率显着提",
"Language32-bit8-bit",
"高",
"专注于机器翻译，Devlin（2017）实现了16",
"En-NI",
"4.02",
"4.08",
"位定点整数数学运算，以加快矩阵乘法运算，",
"NI-En",
"4.03",
"4.03",
"Ru-En",
"看到一个2。59倍的改进。他们在WMT英",
"4.10",
"4.06",
"语-法语NewsTest2014上显示出具有竞争力",
"4.05",
"4.16",
"的BLEU分数，同时提供了显着的加速。同",
"En-De 2x",
"En-De 1x",
"3.84",
"3.90",
"样，（Wuetal.,2016）在翻译模型中应用8位",
"Human evaluation scores for 8-bit and",
"端到端量化。他们还表明，自动指标不会因此",
"而受到影响。在这项工作中，量化需要修改模",
"Table 7:",
"32-bit systems. All tests are news domain.",
"型训练以限制矩阵输出的大小",
"该表显示，上一节中显示的自动分数也由人",
"类维持。根据人类评估者的说法，8位解码与",
"7Conclusions and Future Work",
"32位解码一样好",
"在本文中，我们证明了神经机器翻译的8位",
"5Discussion",
"解码比类似的优化浮点实现快4-6倍。我们证",
"明了这种近似的质量与32位版本的质量相似。",
"在不损失精度的情况下拥有更快的NMT引擎",
"我们还表明，没有必要修改训练过程以产生与",
"在商业上很有用。在我们的部署方案中，它是",
"8位解码兼容的模型",
"缓慢的交互式用户体验和不。即使在批处理模",
"总而言之，本文表明，在自动测量和从人类",
"式下运行，也可以用1/4的硬件提供相同的吞",
"感知的角度来看，8位解码与32位解码一样",
"吐量",
"好，同时它大天改善了延迟",
"此外，这种加速使部署小型模型集合变得实",
"将来，我们计划实现多层矩阵乘法，该算法",
"用。如上表6中的En-De模型所示，合奏可",
"可以追溯到矩阵面板乘法的朴素算法。这将为",
"以以2倍的减速为代价提供更高的精度。这项",
"可以利用它的应用程序提供批量解码的速度。",
"工作使翻译质量更高，同时速度至少是以前基",
"我们还计划探索低精度的训练，以加快实验周",
"线的两倍。",
"转时间",
"正如4节中报告的数字所证明的那样，8位",
"我们的结果提供了提高准确性的提示，而不",
"和32位解码具有相似的平均质量。正如预期",
"仅仅是奇偶校验。其他工作已将培训用作压缩",
"的那样，两个解码器产生的输出不相同。实际",
"过程的一部分。我们想看看训练量化模型是否",
"上，在166句De-En翻译中，两者之间只有",
"会改变结果的好坏",
"51句是相同的。此外，我们的人工评估结果和",
"自动评分表明，与32位解码器相比，8位解码",
"",
"'In order to limit the vocabulary,we use BPE sub-",
"器没有特定的降级。为了强调这些权利要求，",
"",
"word units (Sennrich et al., 2016) in all models."
] |
{
"arxiv_id": "1804.05038",
"language": "zh",
"page_idx": 3
}
|
||
1804.05038-zh-4
|
1804.05038-zh-4.jpg
|
Source
Time
Sie standen seit 1946 an der Parteispitze
32-bit
ms
had been at the
1946
<table>
<tbody>
<tr>
<th>$8-bit$</th>
<th>180 ms</th>
<th>Thev stood at the top of the partv since 1946</th>
</tr>
<tr>
<td>Source</td>
<td>Time</td>
<td>So erwarten die Experten für dieses Jahr lediglich einen Anstieg der Weltproduktion von 3,7 statt der im Juni prognostizierten 3,9 Prozent. Für 2009 sagt das Kielen</td>
</tr>
<tr>
<td>32-bit</td>
<td>4440 ms</td>
<td>For this year, the experts expect only an increase in world production of 3.7 instead of the 3.9 percent forecast in June. In 2009, the Kiel Institute</td>
</tr>
<tr>
<td>$8-bit$</td>
<td>750 $ms$</td>
<td>For this year, the experts expect only an increase in world production of 3.7 instead of the 3.9 percent forecast in June. In 2009, the Kiel Institute even forecast a slowdown to $3.3\%$instead of 3.7 per cent .</td>
</tr>
<tr>
<td>Source</td>
<td>Time</td>
<td>Heftige Regenfälle wegen“Ike”werden möglicherweise schwerere Schäden anrichten als seine Windböen. Besonders gefährdet sind dicht besiedelte Gebiete im Tal des R</td>
</tr>
<tr>
<td>32-bit</td>
<td>6150 ms</td>
<td>eldem Heavy rainfall due to“Ike”may cause more severe damage than its gusts of wind, particularly in densely populated areas in the Rio Grande valley, which are still</td>
</tr>
<tr>
<td>$8-bit$</td>
<td>1050 ms</td>
<td>Heavy rainfall due to“Ike” may cause heavier damage than its gusts of wind, particularly in densely populated areas in the Rio Grande valley, which still</td>
</tr>
</tbody>
</table>
Table 8: Examples of De-En news translation system comparing 32-bit and 8-bit decoding. Differences
are in boldface. Sentence times are average of 10 runs.
Source
Time
Het is tijd om de kloof te overbruggen.
730 ms It's time to bridge the gap.
32-bit 8-bit
180ms|It is time to bridge the gap.
$\overline{\text{Niet dat Barientos met zijn vader van plaats zou willen wisselen.}}$
$\overline{\mathrm{Source}}$
$\overline{\mathrm{Time}}$
32-bit | 1120 ms | Not that Barientos would want to change his father's place
290 ms | Not that Barientos would like to switch places with his father.
8-bit
Table 9: Examples of Nl-En news translation system comparing 32-bit and 8-bit decoding. Differences
are in boldface. Sentence times are average of 10 runs.
## References
the properties of neural machine transla-
CoRR
tion: Encoder-decoder approaches.
$\operatorname{abs/1409.1259}.$
Jacob Devlin. 2017. Sharp models on dull hard-
ware: Fast and accurate neural machine translation decoding on the cpu. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, pages 2820-2825.
Markus Freitag, Yaser Al-Onaizan, and Baskaran
Sankaran. 2017. Ensemble distillation for neural
machine translation. CoRR abs/1702.01802.
Kazushige Goto and Robert A. van de Geijn. 2008.
Anatomy of high-performance matrix multiplication. ACM Trans. Math. Softw. 34(3):12:1- 12:25.
Song Han, Huizi Mao, and William J. Dally. 2015.
Ahmad Abdelfattah, David Keyes, and Hatem
Ltaief. 2016. Kblas: An optimized library for dense matrix-vector multiplication on gpu accelerators. ACM Trans. Math. Softw. 42(3):18:1- 18{:}31. https: / / doi. org/ 10. 1145/ 2818311.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua
Bengio. 2014. Neural machine translationby jointly learning to align and translate. CoRR $\operatorname{abs}/1409.0473.$
Frédéric Bastien, Pascal Lamblin, Razvan Pas-
canu, James Bergstra, Ian J. Goodfellow, Arnaud Bergeron, $\bar{\text{Nicolas Bouchard, and Yoshua}}$ Bengio. 2012. Theano: new features and speed $\operatorname* { improvements. }$ $\operatorname* { Deep}$ $\operatorname* { Learning}$ and Unsuper- $\tilde{\text{vised Feature Learning NIPS 2012 Workshop}}.$ KyungHyun Cho, Bart van Merrienboer, Dzmitry
On
Bahdanau, and Yoshua Bengio. 2014.
Deep compression: Compressing deep neural
|
[
[
0.1275634765625,
0.07244873046875,
0.187255859375,
0.0748291015625,
0.185546875,
0.08905029296875,
0.1259765625,
0.0872802734375
],
[
0.214111328125,
0.07183837890625,
0.26025390625,
0.0736083984375,
0.258544921875,
0.08905029296875,
0.21240234375,
0.0872802734375
],
[
0.28955078125,
0.0736083984375,
0.6171875,
0.0736083984375,
0.6171875,
0.0885009765625,
0.28955078125,
0.0885009765625
],
[
0.130126953125,
0.08905029296875,
0.1822509765625,
0.08905029296875,
0.1822509765625,
0.1051025390625,
0.130126953125,
0.1051025390625
],
[
0.20654296875,
0.09027099609375,
0.270263671875,
0.09027099609375,
0.270263671875,
0.1051025390625,
0.20654296875,
0.1051025390625
],
[
0.28955078125,
0.09027099609375,
0.68994140625,
0.09027099609375,
0.68994140625,
0.1051025390625,
0.28955078125,
0.1051025390625
],
[
0.133544921875,
0.105712890625,
0.177978515625,
0.105712890625,
0.177978515625,
0.1217041015625,
0.133544921875,
0.1217041015625
],
[
0.208251953125,
0.105712890625,
0.270263671875,
0.105712890625,
0.270263671875,
0.12054443359375,
0.208251953125,
0.12054443359375
],
[
0.28955078125,
0.10687255859375,
0.66162109375,
0.10687255859375,
0.66162109375,
0.12054443359375,
0.28955078125,
0.12054443359375
],
[
0.1234130859375,
0.1217041015625,
0.18896484375,
0.12353515625,
0.187255859375,
0.1395263671875,
0.12176513671875,
0.13720703125
],
[
0.21240234375,
0.1217041015625,
0.26123046875,
0.12353515625,
0.26025390625,
0.1395263671875,
0.210693359375,
0.13720703125
],
[
0.2880859375,
0.12353515625,
0.9716796875,
0.12353515625,
0.9716796875,
0.13720703125,
0.2880859375,
0.13720703125
],
[
0.2880859375,
0.1395263671875,
0.9345703125,
0.1383056640625,
0.9345703125,
0.152587890625,
0.2880859375,
0.15380859375
],
[
0.2880859375,
0.15380859375,
0.8271484375,
0.155029296875,
0.8271484375,
0.169189453125,
0.2880859375,
0.168701171875
],
[
0.130126953125,
0.17041015625,
0.183837890625,
0.17041015625,
0.183837890625,
0.185302734375,
0.130126953125,
0.185302734375
],
[
0.202392578125,
0.17041015625,
0.273681640625,
0.17041015625,
0.273681640625,
0.185302734375,
0.202392578125,
0.185302734375
],
[
0.287109375,
0.171630859375,
0.90283203125,
0.171630859375,
0.90283203125,
0.185302734375,
0.287109375,
0.185302734375
],
[
0.2880859375,
0.18701171875,
0.85400390625,
0.18701171875,
0.85400390625,
0.20068359375,
0.2880859375,
0.20068359375
],
[
0.2880859375,
0.2037353515625,
0.7900390625,
0.2037353515625,
0.7900390625,
0.21728515625,
0.2880859375,
0.21728515625
],
[
0.1351318359375,
0.218505859375,
0.1796875,
0.218505859375,
0.1796875,
0.2340087890625,
0.1351318359375,
0.2340087890625
],
[
0.20654296875,
0.2197265625,
0.270263671875,
0.2197265625,
0.270263671875,
0.2340087890625,
0.20654296875,
0.2340087890625
],
[
0.2880859375,
0.2197265625,
0.90185546875,
0.2197265625,
0.90185546875,
0.2327880859375,
0.2880859375,
0.2327880859375
],
[
0.2880859375,
0.235107421875,
0.89404296875,
0.235107421875,
0.89404296875,
0.248779296875,
0.2880859375,
0.248779296875
],
[
0.2880859375,
0.25048828125,
0.7197265625,
0.251708984375,
0.7197265625,
0.2666015625,
0.2880859375,
0.265380859375
],
[
0.1234130859375,
0.2666015625,
0.187255859375,
0.268310546875,
0.185546875,
0.283203125,
0.12176513671875,
0.28076171875
],
[
0.2149658203125,
0.26953125,
0.258544921875,
0.26953125,
0.258544921875,
0.28076171875,
0.2149658203125,
0.28076171875
],
[
0.291259765625,
0.26953125,
0.495361328125,
0.26953125,
0.495361328125,
0.27978515625,
0.291259765625,
0.27978515625
],
[
0.495361328125,
0.268310546875,
0.95947265625,
0.268310546875,
0.95947265625,
0.281982421875,
0.495361328125,
0.281982421875
],
[
0.28955078125,
0.283203125,
0.97998046875,
0.283203125,
0.97998046875,
0.29638671875,
0.28955078125,
0.29638671875
],
[
0.291259765625,
0.300537109375,
0.927734375,
0.300537109375,
0.927734375,
0.314208984375,
0.291259765625,
0.314208984375
],
[
0.130126953125,
0.31640625,
0.183837890625,
0.31640625,
0.183837890625,
0.330810546875,
0.130126953125,
0.330810546875
],
[
0.202392578125,
0.31640625,
0.27197265625,
0.31640625,
0.27197265625,
0.330810546875,
0.202392578125,
0.330810546875
],
[
0.28955078125,
0.317626953125,
0.947265625,
0.317626953125,
0.947265625,
0.330810546875,
0.28955078125,
0.330810546875
],
[
0.287109375,
0.333251953125,
0.9296875,
0.33203125,
0.9296875,
0.3466796875,
0.287109375,
0.347412109375
],
[
0.28955078125,
0.349853515625,
0.80419921875,
0.349853515625,
0.80419921875,
0.363525390625,
0.28955078125,
0.363525390625
],
[
0.133544921875,
0.364013671875,
0.1796875,
0.364013671875,
0.1796875,
0.380126953125,
0.133544921875,
0.380126953125
],
[
0.202392578125,
0.364013671875,
0.27197265625,
0.364013671875,
0.27197265625,
0.37890625,
0.202392578125,
0.37890625
],
[
0.287109375,
0.365234375,
0.91162109375,
0.365234375,
0.91162109375,
0.380126953125,
0.287109375,
0.380126953125
],
[
0.2880859375,
0.38134765625,
0.89990234375,
0.38134765625,
0.89990234375,
0.3955078125,
0.2880859375,
0.3955078125
],
[
0.287109375,
0.3955078125,
0.7841796875,
0.396728515625,
0.7841796875,
0.410888671875,
0.287109375,
0.410400390625
],
[
0.1158447265625,
0.42041015625,
0.88671875,
0.421630859375,
0.88671875,
0.439453125,
0.1158447265625,
0.438232421875
],
[
0.118408203125,
0.438232421875,
0.5224609375,
0.438232421875,
0.5224609375,
0.451904296875,
0.118408203125,
0.451904296875
],
[
0.15869140625,
0.46435546875,
0.21826171875,
0.46435546875,
0.21826171875,
0.479736328125,
0.15869140625,
0.479736328125
],
[
0.2435302734375,
0.463134765625,
0.2939453125,
0.465576171875,
0.29296875,
0.48095703125,
0.2418212890625,
0.478515625
],
[
0.316650390625,
0.46435546875,
0.63232421875,
0.465576171875,
0.63232421875,
0.479736328125,
0.316650390625,
0.478515625
],
[
0.160400390625,
0.48095703125,
0.214111328125,
0.48095703125,
0.214111328125,
0.495361328125,
0.160400390625,
0.495361328125
],
[
0.2376708984375,
0.48095703125,
0.301513671875,
0.48095703125,
0.301513671875,
0.495361328125,
0.2376708984375,
0.495361328125
],
[
0.316650390625,
0.479736328125,
0.5380859375,
0.482177734375,
0.5380859375,
0.496337890625,
0.316650390625,
0.49462890625
],
[
0.16455078125,
0.496337890625,
0.210693359375,
0.496337890625,
0.210693359375,
0.5126953125,
0.16455078125,
0.5126953125
],
[
0.2376708984375,
0.495361328125,
0.300537109375,
0.49755859375,
0.298828125,
0.5126953125,
0.2359619140625,
0.51025390625
],
[
0.316650390625,
0.496337890625,
0.5439453125,
0.49755859375,
0.5439453125,
0.5126953125,
0.316650390625,
0.51123046875
],
[
0.154541015625,
0.5126953125,
0.219970703125,
0.51416015625,
0.21826171875,
0.529296875,
0.15283203125,
0.52685546875
],
[
0.2435302734375,
0.51416015625,
0.2939453125,
0.51416015625,
0.2939453125,
0.529296875,
0.2435302734375,
0.529296875
],
[
0.316650390625,
0.51318359375,
0.84228515625,
0.51416015625,
0.84228515625,
0.529296875,
0.316650390625,
0.52783203125
],
[
0.160400390625,
0.52783203125,
0.214111328125,
0.52978515625,
0.21240234375,
0.5458984375,
0.15869140625,
0.54345703125
],
[
0.235107421875,
0.52978515625,
0.301513671875,
0.52978515625,
0.301513671875,
0.54443359375,
0.235107421875,
0.54443359375
],
[
0.31982421875,
0.52978515625,
0.80419921875,
0.52978515625,
0.80419921875,
0.54443359375,
0.31982421875,
0.54443359375
],
[
0.16455078125,
0.5458984375,
0.2091064453125,
0.5458984375,
0.2091064453125,
0.56103515625,
0.16455078125,
0.56103515625
],
[
0.2376708984375,
0.5458984375,
0.301513671875,
0.5458984375,
0.301513671875,
0.56005859375,
0.2376708984375,
0.56005859375
],
[
0.31982421875,
0.54638671875,
0.8271484375,
0.54638671875,
0.8271484375,
0.56005859375,
0.31982421875,
0.56005859375
],
[
0.12005615234375,
0.57373046875,
0.18896484375,
0.57373046875,
0.18896484375,
0.5849609375,
0.12005615234375,
0.5849609375
],
[
0.19140625,
0.57373046875,
0.88330078125,
0.57373046875,
0.88330078125,
0.58740234375,
0.19140625,
0.58740234375
],
[
0.118408203125,
0.58740234375,
0.5205078125,
0.58740234375,
0.5205078125,
0.60107421875,
0.118408203125,
0.60107421875
],
[
0.12176513671875,
0.626953125,
0.21240234375,
0.626953125,
0.21240234375,
0.63818359375,
0.12176513671875,
0.63818359375
],
[
0.53076171875,
0.6259765625,
0.82275390625,
0.6259765625,
0.82275390625,
0.64013671875,
0.53076171875,
0.64013671875
],
[
0.8154296875,
0.626953125,
0.8818359375,
0.626953125,
0.8818359375,
0.63818359375,
0.8154296875,
0.63818359375
],
[
0.53076171875,
0.6396484375,
0.5751953125,
0.64111328125,
0.57421875,
0.65283203125,
0.52880859375,
0.65087890625
],
[
0.5693359375,
0.63818359375,
0.8134765625,
0.64013671875,
0.8134765625,
0.65380859375,
0.5693359375,
0.65185546875
],
[
0.8310546875,
0.64013671875,
0.88330078125,
0.64013671875,
0.88330078125,
0.65185546875,
0.8310546875,
0.65185546875
],
[
0.118408203125,
0.64794921875,
0.490234375,
0.64794921875,
0.490234375,
0.6611328125,
0.118408203125,
0.6611328125
],
[
0.52880859375,
0.65185546875,
0.6455078125,
0.65087890625,
0.6455078125,
0.6650390625,
0.52880859375,
0.66650390625
],
[
0.1351318359375,
0.66015625,
0.48779296875,
0.66015625,
0.48779296875,
0.67333984375,
0.1351318359375,
0.67333984375
],
[
0.1368408203125,
0.6748046875,
0.48779296875,
0.67578125,
0.48779296875,
0.68701171875,
0.1368408203125,
0.68603515625
],
[
0.51318359375,
0.6767578125,
0.8818359375,
0.6767578125,
0.8818359375,
0.6904296875,
0.51318359375,
0.6904296875
],
[
0.1351318359375,
0.68701171875,
0.486083984375,
0.68701171875,
0.486083984375,
0.70068359375,
0.1351318359375,
0.70068359375
],
[
0.5322265625,
0.6923828125,
0.8818359375,
0.6923828125,
0.8818359375,
0.703125,
0.5322265625,
0.703125
],
[
0.1376953125,
0.70068359375,
0.435791015625,
0.70068359375,
0.435791015625,
0.71435546875,
0.1376953125,
0.71435546875
],
[
0.5322265625,
0.70361328125,
0.88330078125,
0.70361328125,
0.88330078125,
0.71728515625,
0.5322265625,
0.71728515625
],
[
0.53076171875,
0.71533203125,
0.8818359375,
0.71630859375,
0.8818359375,
0.72998046875,
0.53076171875,
0.728515625
],
[
0.118408203125,
0.7255859375,
0.48779296875,
0.7255859375,
0.48779296875,
0.74072265625,
0.118408203125,
0.74072265625
],
[
0.5322265625,
0.72998046875,
0.88330078125,
0.72998046875,
0.88330078125,
0.74365234375,
0.5322265625,
0.74365234375
],
[
0.1368408203125,
0.7392578125,
0.489501953125,
0.7392578125,
0.489501953125,
0.7529296875,
0.1368408203125,
0.7529296875
],
[
0.52880859375,
0.7412109375,
0.82861328125,
0.7412109375,
0.82861328125,
0.755859375,
0.52880859375,
0.755859375
],
[
0.1368408203125,
0.751953125,
0.489501953125,
0.751953125,
0.489501953125,
0.7666015625,
0.1368408203125,
0.7666015625
],
[
0.1351318359375,
0.7666015625,
0.2476806640625,
0.765625,
0.2476806640625,
0.77685546875,
0.1351318359375,
0.77783203125
],
[
0.51318359375,
0.7666015625,
0.880859375,
0.7666015625,
0.880859375,
0.77978515625,
0.51318359375,
0.77978515625
],
[
0.53076171875,
0.77978515625,
0.88330078125,
0.77978515625,
0.88330078125,
0.79345703125,
0.53076171875,
0.79345703125
],
[
0.1175537109375,
0.7890625,
0.486083984375,
0.7890625,
0.486083984375,
0.802734375,
0.1175537109375,
0.802734375
],
[
0.5322265625,
0.79345703125,
0.85791015625,
0.79345703125,
0.85791015625,
0.80712890625,
0.5322265625,
0.80712890625
],
[
0.1376953125,
0.8056640625,
0.486083984375,
0.8056640625,
0.486083984375,
0.81640625,
0.1376953125,
0.81640625
],
[
0.1368408203125,
0.81787109375,
0.489501953125,
0.81787109375,
0.489501953125,
0.83056640625,
0.1368408203125,
0.83056640625
],
[
0.51123046875,
0.81787109375,
0.880859375,
0.81787109375,
0.880859375,
0.83056640625,
0.51123046875,
0.83056640625
],
[
0.1351318359375,
0.830078125,
0.48779296875,
0.830078125,
0.48779296875,
0.84423828125,
0.1351318359375,
0.84423828125
],
[
0.5341796875,
0.8330078125,
0.880859375,
0.8330078125,
0.880859375,
0.84326171875,
0.5341796875,
0.84326171875
],
[
0.2493896484375,
0.84326171875,
0.48779296875,
0.84326171875,
0.48779296875,
0.85693359375,
0.2493896484375,
0.85693359375
],
[
0.53076171875,
0.84423828125,
0.880859375,
0.84423828125,
0.880859375,
0.85791015625,
0.53076171875,
0.85791015625
],
[
0.1351318359375,
0.85546875,
0.470947265625,
0.85546875,
0.470947265625,
0.87060546875,
0.1351318359375,
0.87060546875
],
[
0.1376953125,
0.845703125,
0.252685546875,
0.845703125,
0.252685546875,
0.85546875,
0.1376953125,
0.85546875
],
[
0.5341796875,
0.859375,
0.580078125,
0.859375,
0.580078125,
0.87060546875,
0.5341796875,
0.87060546875
],
[
0.118408203125,
0.8818359375,
0.490234375,
0.8818359375,
0.490234375,
0.8955078125,
0.118408203125,
0.8955078125
],
[
0.51318359375,
0.8818359375,
0.8818359375,
0.8818359375,
0.8818359375,
0.8955078125,
0.51318359375,
0.8955078125
],
[
0.453369140625,
0.8935546875,
0.489501953125,
0.89794921875,
0.486083984375,
0.90771484375,
0.450927734375,
0.90380859375
],
[
0.52880859375,
0.89453125,
0.88330078125,
0.8955078125,
0.88330078125,
0.91015625,
0.52880859375,
0.9091796875
],
[
0.1368408203125,
0.8955078125,
0.44921875,
0.8955078125,
0.44921875,
0.9091796875,
0.1368408203125,
0.9091796875
]
] |
[
"Source",
"Time",
"Sie standen seit 1946 an der Parteispitze",
"32-bit",
"720 ms",
"They had been at the party leadership since 1946",
"8-bit",
"180 ms",
"They stood at the top of the party since 1946.",
"Source",
"Time",
"So erwarten die Experten fur dieses Jahr lediglich einen Anstieg der Weltproduktion",
"von 3,7 statt der im Juni prognostizierten 3,9 Prozent. Fur 2009 sagt das Kieler",
"Institut sogar eine Abschwachung auf 3,3 statt 3,7 Prozent voraus.",
"32-bit",
"4440 ms",
"For this year, the experts expect only an increase in world production of 3.7",
"instead of the 3.9 percent forecast in June. In 2009, the Kiel Institute",
"predictated a slowdown to 3.3 percent instead of 3.7 percent .",
"8-bit",
"750 ms",
"For this year, the experts expect only an increase in world production of 3.7",
"instead of the 3.9 percent forecast in June. In 2009, the Kiel Institute even",
"forecast a slowdown to 3.3 % instead of 3.7 per cent .",
"Source",
"Time",
"HeftigeRegenfallewegen",
"\"Ike\" werden moglicherweise schwerere Schaden anrichten",
"als seine Windboen. Besonders gefahrdet sind dicht besiedelte Gebiete im Tal des Rio",
"Grande, die noch immer unter den Folgen des Hurrikans“Dolly\"im Juli leiden",
"32-bit",
"6150 ms",
"Heavy rainfall due to “Ike\" may cause more severe damage than its gusts of wind",
"particularly in densely populated areas in the Rio Grande valley,which are still",
"suffering from the consequences of the \"dolly\" hurricane in July.",
"8-bit",
"1050 ms",
"Heavy rainfall due to “Ike\" may cause heavier damage than its gusts of wind,",
"particularly in densely populated areas in the Rio Grande valley, which still",
"suffer from the consequences of the “dolly\" hurricane in July.",
"Table 8: Examples of De-En news translation system comparing 32-bit and 8-bit decoding. Diferences",
"are in boldface. Sentence times are average of 10 runs.",
"Source",
"Time",
"Het is tijd om de kloof te overbruggen.",
"32-bit",
"730 ms",
"It's time to bridge the gap.",
"8-bit",
"180 ms",
"It is time to bridge the gap.",
"Source",
"Time",
"Niet dat Barientos met zijn vader van plaats zou willen wisselen.",
"32-bit",
"1120 ms",
"Not that Barientos would want to change his father's place",
"8-bit",
"290 ms",
"Not that Barientos would like to switch places with his father .",
"Table 9:",
"Examples of Nl-En news translation system comparing 32-bit and 8-bit decoding. Differences",
"are in boldface. Sentence times are average of 10 runs.",
"References",
"the properties of neural machine",
"transla-",
"tion:",
"Encoder-decoder approaches.",
"CoRR",
"Ahmad Abdelfattah, David Keyes, and Hatem",
"abs/1409.1259.",
"Ltaief. 2016. Kblas: An optimized library for",
"densematrix-vector multiplication ongpu accel-",
"Jacob Devlin. 2017. Sharp models on dull hard-",
"erators. ACM Trans. Math. Softw. 42(3):18:1-",
"ware: Fast and accurate neural machine transla-",
"18:31. https://doi.org/10.1145/2818311.",
"tion decoding on the cpu. In Proceedings of the",
"2017 Conference on Empirical Methods in Nat-",
"Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua",
"ural Language Processing. Association for Com-",
"Bengio. 2014.Neural machine translation by",
" putational Linguistics, pages 2820-2825.",
"jointly learning to align and translate. CoRR",
"abs/1409.0473.",
"Markus Freitag, Yaser Al-Onaizan, and Baskaran",
"Sankaran. 2017. Ensemble distillation for neural",
"Frederic Bastien. Pascal Lamblin, Razvan Pas-",
"machine translation. CoRR abs/1702.01802.",
"canu,James Bergstra,Ian J.Goodfellow,Ar-",
"naud Bergeron, Nicolas Bouchard, and Yoshua",
"Kazushige Goto and Robert A. van de Geijn. 2008.",
"Bengio. 2012. Theano: new features and speed",
"Anatomyof high-performance matrixmultipli-",
"Deep Learning and Unsuper-",
"cation. ACM Trans. Math. Softw. 34(3):12:1-",
"vised Feature Learning NIPS 2012 Workshop",
"improvements.",
"12:25.",
"KyungHyun Cho, Bart van Merrienboer, Dzmitry",
"Song Han, Huizi Mao, and William J. Dally. 2015",
"On",
"Deep compression: Compressing deep neural",
"Bahdanau, and Yoshua Bengio. 2014."
] |
{
"arxiv_id": "1804.05038",
"language": "zh",
"page_idx": 4
}
|
||
2008.00446-zh-0
|
2008.00446-zh-0.jpg
|
# Stochastic Bundle Adjustment for Efficient and Scalable 3D Reconstruction
Lei Zhou$^1$, Zixin Luo$^1$, Mingmin Zhen$^1$, Tianwei Shen$^1$, Shiwei Li$^2$,Zhuofei Huang$^1$,Tian Fang$^2$,and Long Quan$^1$
$^{1}$ Hong Kong University of Science and Technology
$\{$lzhouai,zluoag,mzhen,tshenaa,zhuangbr,quan$\}$@cse.ust.hk
$^{2}{\mathrm{~Everest~Innovation~Technology}}$ $\{$ sli,fangtian $\}$ @altizure.com
$\mathbf{Abstract.~Current~bundle~adjustment~solvers~such~as~the~Levenberg-}$ Marquardt (LM) algorithm are limited by the bottleneck in solving the Reduced Camera System (RCS) whose dimension is proportional to the camera number. When the problem is scaled up, this step is neither efficient in computation nor manageable for a single compute node.In this work, we propose a stochastic bundle adjustment algorithm which seeks to decompose the RCS approximately inside the LM iterations to improve the efficiency and scalability. It first reformulates the quadratic programming problem of an LM iteration based on the clustering of the visibility graph by introducing the equality constraints across clusters. Then, we propose to relax it into a chance constrained problem and solve it through sampled convex program. The relaxation is intended to eliminate the interdependence between clusters embodied by the constraints, so that a large RCS can be decomposed into independent linear sub-problems. Numerical experiments on unordered Internet image sets and sequential SLAM image sets, as well as distributed experiments on large-scale datasets, have demonstrated the high efficiency and scalability of the proposed approach. Codes are released at https://github.com/zlthinker/STBA
Keywords: Stochastic bundle adjustment, Clustering, 3D reconstruc-
tion
## 1 Introduction
束调整(BA)通常被公式化为非线性最小二乘问题，以优化相机和 3D 点的参数。通常由 Levenberg-Marquardt (LM) 算法解决，其中必须在每次迭代中求解称为简化相机系统(RCS)[26,16]的线性方程组。然而，当问题扩大时，解决RCS 一直是一个瓶颈，它占用了大部分计算时间(见图的第一条。1)。RCS 的尺寸与摄像机数量成正比，因此摄像机的增加会增加计算和内存消耗，尽管已经提出了使用有效的线性求解器[3,24,14,22,38]并节省矩阵操作[3,23,38]的方法。此外，与雅可比或梯度评估等其他操作不同，此步骤是不可分割的，因此很难将 BA 用于并行和分布式计算
|
[
[
0.093017578125,
0.052001953125,
0.90478515625,
0.05279541015625,
0.90478515625,
0.07391357421875,
0.093017578125,
0.07232666015625
],
[
0.26318359375,
0.08367919921875,
0.734375,
0.08367919921875,
0.734375,
0.1015625,
0.26318359375,
0.1015625
],
[
0.202880859375,
0.141357421875,
0.7900390625,
0.1422119140625,
0.7900390625,
0.1600341796875,
0.202880859375,
0.1591796875
],
[
0.1920166015625,
0.1624755859375,
0.80078125,
0.163330078125,
0.80078125,
0.18115234375,
0.1920166015625,
0.1795654296875
],
[
0.2427978515625,
0.193359375,
0.75244140625,
0.196533203125,
0.75244140625,
0.2169189453125,
0.2427978515625,
0.2144775390625
],
[
0.1702880859375,
0.2169189453125,
0.82861328125,
0.2169189453125,
0.82861328125,
0.2315673828125,
0.1702880859375,
0.2315673828125
],
[
0.333251953125,
0.2298583984375,
0.66162109375,
0.233154296875,
0.66162109375,
0.251708984375,
0.333251953125,
0.24853515625
],
[
0.324951171875,
0.2493896484375,
0.6728515625,
0.2509765625,
0.6728515625,
0.26806640625,
0.324951171875,
0.267333984375
],
[
0.1497802734375,
0.3193359375,
0.8466796875,
0.32177734375,
0.8466796875,
0.339599609375,
0.1497802734375,
0.337158203125
],
[
0.1497802734375,
0.34033203125,
0.8466796875,
0.34033203125,
0.8466796875,
0.353271484375,
0.1497802734375,
0.353271484375
],
[
0.1497802734375,
0.359130859375,
0.84912109375,
0.359130859375,
0.84912109375,
0.373779296875,
0.1497802734375,
0.373779296875
],
[
0.1497802734375,
0.375244140625,
0.8466796875,
0.375244140625,
0.8466796875,
0.389892578125,
0.1497802734375,
0.389892578125
],
[
0.1497802734375,
0.39404296875,
0.84765625,
0.39404296875,
0.84765625,
0.40869140625,
0.1497802734375,
0.40869140625
],
[
0.1497802734375,
0.411865234375,
0.84765625,
0.411865234375,
0.84765625,
0.426513671875,
0.1497802734375,
0.426513671875
],
[
0.1473388671875,
0.4296875,
0.84912109375,
0.4296875,
0.84912109375,
0.4443359375,
0.1473388671875,
0.4443359375
],
[
0.1473388671875,
0.447509765625,
0.84912109375,
0.447509765625,
0.84912109375,
0.462158203125,
0.1473388671875,
0.462158203125
],
[
0.1497802734375,
0.46630859375,
0.84912109375,
0.46630859375,
0.84912109375,
0.48095703125,
0.1497802734375,
0.48095703125
],
[
0.1497802734375,
0.484130859375,
0.8466796875,
0.484130859375,
0.8466796875,
0.498779296875,
0.1497802734375,
0.498779296875
],
[
0.1473388671875,
0.501953125,
0.84765625,
0.501953125,
0.84765625,
0.51513671875,
0.1473388671875,
0.51513671875
],
[
0.1473388671875,
0.51806640625,
0.84423828125,
0.51806640625,
0.84423828125,
0.53271484375,
0.1473388671875,
0.53271484375
],
[
0.1497802734375,
0.5361328125,
0.84765625,
0.5361328125,
0.84765625,
0.55078125,
0.1497802734375,
0.55078125
],
[
0.1473388671875,
0.55224609375,
0.84765625,
0.55224609375,
0.84765625,
0.5703125,
0.1473388671875,
0.5703125
],
[
0.1473388671875,
0.57275390625,
0.8466796875,
0.57275390625,
0.8466796875,
0.5859375,
0.1473388671875,
0.5859375
],
[
0.1473388671875,
0.5908203125,
0.84912109375,
0.5908203125,
0.84912109375,
0.60498046875,
0.1473388671875,
0.60498046875
],
[
0.1473388671875,
0.60693359375,
0.8466796875,
0.60693359375,
0.8466796875,
0.62158203125,
0.1473388671875,
0.62158203125
],
[
0.1497802734375,
0.62548828125,
0.84912109375,
0.62548828125,
0.84912109375,
0.64013671875,
0.1497802734375,
0.64013671875
],
[
0.1461181640625,
0.642578125,
0.52294921875,
0.642578125,
0.52294921875,
0.65966796875,
0.1461181640625,
0.65966796875
],
[
0.1461181640625,
0.6767578125,
0.8515625,
0.677734375,
0.8515625,
0.6953125,
0.1461181640625,
0.6943359375
],
[
0.1473388671875,
0.69873046875,
0.1920166015625,
0.69873046875,
0.1920166015625,
0.71240234375,
0.1473388671875,
0.71240234375
],
[
0.08331298828125,
0.75048828125,
0.1026611328125,
0.75732421875,
0.093017578125,
0.7685546875,
0.07489013671875,
0.76220703125
],
[
0.134033203125,
0.75048828125,
0.31640625,
0.75048828125,
0.31640625,
0.7685546875,
0.134033203125,
0.7685546875
],
[
0.08331298828125,
0.79541015625,
0.91552734375,
0.79541015625,
0.91552734375,
0.81005859375,
0.08331298828125,
0.81005859375
],
[
0.0821533203125,
0.8154296875,
0.9130859375,
0.8154296875,
0.9130859375,
0.82958984375,
0.0821533203125,
0.82958984375
],
[
0.0821533203125,
0.8349609375,
0.9130859375,
0.8349609375,
0.9130859375,
0.849609375,
0.0821533203125,
0.849609375
],
[
0.0809326171875,
0.85302734375,
0.91552734375,
0.85302734375,
0.91552734375,
0.8701171875,
0.0809326171875,
0.8701171875
],
[
0.08331298828125,
0.873046875,
0.9130859375,
0.873046875,
0.9130859375,
0.8876953125,
0.08331298828125,
0.8876953125
],
[
0.08331298828125,
0.89208984375,
0.9130859375,
0.89208984375,
0.9130859375,
0.90673828125,
0.08331298828125,
0.90673828125
],
[
0.08331298828125,
0.912109375,
0.91552734375,
0.912109375,
0.91552734375,
0.9267578125,
0.08331298828125,
0.9267578125
],
[
0.0821533203125,
0.931640625,
0.443115234375,
0.931640625,
0.443115234375,
0.9462890625,
0.0821533203125,
0.9462890625
]
] |
[
"StochasticBundleAdjustmentforEfficientand",
"Scalable 3D Reconstruction",
"Lei Zhoul,Zixin Luol, Mingmin Zhen', Tianwei Shen'",
"ShiweiLi?,ZhuofeiHuangl,TianFang?,and Long Quan",
"1 Hong Kong University of Science and Technology",
"{lzhouai,zluoag,mzhen,tshenaa,zhuangbr,quan } @cse.ust.hk",
"2 Everest Innovation Technology",
"{sli,fangtian@altizure.com",
"Abstract. Current bundle adjustment solvers such as the Levenberg-",
"Marquardt (LM)algorithmarelimited bythebottleneckinsolvingthe",
"Reduced Camera System (RCS)whose dimension is proportionalto the",
"camera number. When the problem is scaled up, this step is neither ef",
"ficient in computation nor manageable for a single compute node.In",
"this work, we propose a stochastic bundle adjustment algorithm which",
"seeks to decompose theRCS approximately inside the LM iterations to",
"improve the efficiency and scalability. It first reformulates the quadratic",
"programming problem of an LM iteration based on the clustering of",
"the visibility graph by introducing the equality constraints across clus-",
"ters.Then,we propose to relax it into a chance constrained problem",
"and solve it through sampled convex program. The relaxation is in.",
"tended to eliminate the interdependence between clusters embodied by",
"the constraints, so that a large RCS can be decomposed into indepen-",
"dent linear sub-problems. Numerical experiments on unordered Inter-",
"net image sets and sequential SLAM image sets, as well as distributed",
"experiments on large-scale datasets, have demonstrated the high eff-",
"ciency and scalability of the proposed approach. Codes are released at",
"https://github.com/zlthinker/STBA",
"Keywords: Stochastic bundle adjustment, Clustering, 3D reconstruc-",
"tion",
"1",
"Introduction",
"束调整（BA）通常被公式化为非线性最小二乘问题，以优化相机和3D点的参",
"数。通常由Levenberg-Marguardt（LM）算法解决，其中必须在每次迭代中求",
"解称为简化相机系统（RCS）[26，16的线性方程组。然而，当问题扩大时，解决",
"RCS一直是一个瓶颈，它占用了大部分计算时间（见图的第一条。1）。RCS的",
"尺寸与摄像机数量成正比，因此摄像机的增加会增加计算和内存消耗，尽管已",
"经提出了使用有效的线性求解器[3.24.14,22.38]并节省矩阵操作[3.23.38]的",
"方法。此外，与雅可比或梯度评估等其他操作不同，此步骤是不可分割的，因此",
"很难将BA用于并行和分布式计算"
] |
{
"arxiv_id": "2008.00446",
"language": "zh",
"page_idx": 0
}
|
||
2008.00446-zh-1
|
2008.00446-zh-1.jpg
|
L. Zhou, Z. Luo, M. Zhen, T. Shen, $et\textit{al. }$
2
<FigureHere>
Fig.1. Per-iteration time of bundle adjustment w.r.t. the compute node number. The Levenberg-Marquardt (LM) algorithm is limited by the bottleneck when solving the reduced camera system (RCS). Our STBA splits the RCS into independent sub-problems which achieves a speedup on a single-threaded compute node. Besides, STBA allows parallel and distributed computing with multiple compute nodes which further improves the effciency and scalability.
为了实现高效且可扩展的重建，已采用群集作为一种有用的做法，将大型问题分解为更小，更易于管理的问题。例如，已经以分而治之的方式开发了许多SfM 方法，它们首先独立地重建分区的子图，然后将部分重建合并在一起[32,1] 43,44,18] 。尽管这些方法能够以高效且可扩展的方式生成初始稀疏重建，但全束调整对于全局优化摄像机和点参数仍然是必不可少的。因此，在 BA 的背景下，[17,40]提出的方法将 BA 的目标分配给拆分的子模型，并在分布式优化框架[11,6]下优化目标的总和，然而，这涉及额外昂贵的内部迭代，从而使优化过于复杂
在这项工作中，我们遵循利用聚类方法的方向，并系统地推进了如何将聚类方案集成到 BA 问题中的研究。我们不是在预处理步骤中应用固定的一次性分区，而是在每个 LM 迭代中导出基于随机聚类的策略，以便分解 RCS 以提高效率和可伸缩性。
- First, we reformulate the quadratic programming problem of an LM iteration
based on the clustering of the visibility graph. Such a formulation splits the problem into the most elementary structures, but meanwhile introduces additional equality constraints and raises the computational cost.
- Second, in order to make the above problem efficiently solvable, we propose
to relax the constraints into chance constraints [25] and then solve it with sampled convex program [8].The approach helps to eliminate the interdependence between different clusters by randomized constraint reduction which hence decomposes the RCS into independent linear sub-problems related to the clusters. In this way, an approximate step can be achieved efficiently.
- Third, we present an add-on technique which helps to correct the approxi-
mate steps towards the steepest descent direction within a small trust region
to improve the convergence.
由于采样凸程序引起的随机过程，我们将我们的算法称为随机束调整(STBA) 它带来了以下明显的优势。首先，由于降低了复杂性，解决拆分 RCS 代替原始
|
[
[
0.0772705078125,
0.0162506103515625,
0.09906005859375,
0.0162506103515625,
0.09906005859375,
0.03167724609375,
0.0772705078125,
0.03167724609375
],
[
0.156982421875,
0.017059326171875,
0.5654296875,
0.017059326171875,
0.5654296875,
0.03167724609375,
0.156982421875,
0.03167724609375
],
[
0.384033203125,
0.052001953125,
0.41552734375,
0.052001953125,
0.41552734375,
0.0706787109375,
0.384033203125,
0.0706787109375
],
[
0.52880859375,
0.05279541015625,
0.576171875,
0.05279541015625,
0.576171875,
0.0682373046875,
0.52880859375,
0.0682373046875
],
[
0.30078125,
0.06500244140625,
0.368408203125,
0.061737060546875,
0.37451171875,
0.1129150390625,
0.306884765625,
0.11614990234375
],
[
0.388916015625,
0.08367919921875,
0.406982421875,
0.08123779296875,
0.411865234375,
0.09259033203125,
0.393798828125,
0.09503173828125
],
[
0.5458984375,
0.07879638671875,
0.63037109375,
0.07879638671875,
0.63037109375,
0.09344482421875,
0.5458984375,
0.09344482421875
],
[
0.52880859375,
0.09747314453125,
0.54931640625,
0.09747314453125,
0.54931640625,
0.1015625,
0.52880859375,
0.1015625
],
[
0.544921875,
0.09344482421875,
0.658203125,
0.09423828125,
0.658203125,
0.1055908203125,
0.544921875,
0.10479736328125
],
[
0.3056640625,
0.11126708984375,
0.379150390625,
0.1153564453125,
0.370849609375,
0.1787109375,
0.29833984375,
0.1746826171875
],
[
0.30908203125,
0.1080322265625,
0.35986328125,
0.1015625,
0.3671875,
0.1226806640625,
0.315185546875,
0.129150390625
],
[
0.4541015625,
0.11614990234375,
0.475830078125,
0.119384765625,
0.473388671875,
0.126708984375,
0.450439453125,
0.12347412109375
],
[
0.34912109375,
0.15673828125,
0.3623046875,
0.15673828125,
0.3623046875,
0.1673583984375,
0.34912109375,
0.1673583984375
],
[
0.5908203125,
0.1640625,
0.64501953125,
0.1640625,
0.64501953125,
0.1746826171875,
0.5908203125,
0.1746826171875
],
[
0.4541015625,
0.1746826171875,
0.5771484375,
0.17626953125,
0.5771484375,
0.19091796875,
0.4541015625,
0.1893310546875
],
[
0.0784912109375,
0.225830078125,
0.91650390625,
0.2266845703125,
0.91650390625,
0.2445068359375,
0.0784912109375,
0.24365234375
],
[
0.0809326171875,
0.24609375,
0.91162109375,
0.24609375,
0.91162109375,
0.2607421875,
0.0809326171875,
0.2607421875
],
[
0.0809326171875,
0.263916015625,
0.91552734375,
0.263916015625,
0.91552734375,
0.278564453125,
0.0809326171875,
0.278564453125
],
[
0.0821533203125,
0.28271484375,
0.91650390625,
0.28271484375,
0.91650390625,
0.29736328125,
0.0821533203125,
0.29736328125
],
[
0.0821533203125,
0.300537109375,
0.9130859375,
0.300537109375,
0.9130859375,
0.315185546875,
0.0821533203125,
0.315185546875
],
[
0.0821533203125,
0.318359375,
0.43115234375,
0.318359375,
0.43115234375,
0.3330078125,
0.0821533203125,
0.3330078125
],
[
0.11834716796875,
0.375244140625,
0.9130859375,
0.375244140625,
0.9130859375,
0.38916015625,
0.11834716796875,
0.38916015625
],
[
0.08331298828125,
0.39404296875,
0.9130859375,
0.39404296875,
0.9130859375,
0.40869140625,
0.08331298828125,
0.40869140625
],
[
0.0772705078125,
0.4111328125,
0.91552734375,
0.411865234375,
0.91552734375,
0.4296875,
0.0772705078125,
0.428955078125
],
[
0.0821533203125,
0.432861328125,
0.9130859375,
0.432861328125,
0.9130859375,
0.447509765625,
0.0821533203125,
0.447509765625
],
[
0.08331298828125,
0.452392578125,
0.91552734375,
0.452392578125,
0.91552734375,
0.467041015625,
0.08331298828125,
0.467041015625
],
[
0.08575439453125,
0.472900390625,
0.91552734375,
0.472900390625,
0.91552734375,
0.48583984375,
0.08575439453125,
0.48583984375
],
[
0.0809326171875,
0.491455078125,
0.91552734375,
0.491455078125,
0.91552734375,
0.505859375,
0.0809326171875,
0.505859375
],
[
0.0784912109375,
0.50927734375,
0.1787109375,
0.50927734375,
0.1787109375,
0.5263671875,
0.0784912109375,
0.5263671875
],
[
0.11956787109375,
0.52978515625,
0.9130859375,
0.52978515625,
0.9130859375,
0.54443359375,
0.11956787109375,
0.54443359375
],
[
0.08331298828125,
0.55078125,
0.9130859375,
0.55078125,
0.9130859375,
0.5654296875,
0.08331298828125,
0.5654296875
],
[
0.0809326171875,
0.56884765625,
0.91650390625,
0.56884765625,
0.91650390625,
0.5859375,
0.0809326171875,
0.5859375
],
[
0.0821533203125,
0.5888671875,
0.2403564453125,
0.5888671875,
0.2403564453125,
0.603515625,
0.0821533203125,
0.603515625
],
[
0.09783935546875,
0.62451171875,
0.9130859375,
0.62451171875,
0.9130859375,
0.63916015625,
0.09783935546875,
0.63916015625
],
[
0.11956787109375,
0.642578125,
0.91650390625,
0.642578125,
0.91650390625,
0.65966796875,
0.11956787109375,
0.65966796875
],
[
0.1231689453125,
0.6630859375,
0.91796875,
0.6630859375,
0.91796875,
0.677734375,
0.1231689453125,
0.677734375
],
[
0.12078857421875,
0.68310546875,
0.81396484375,
0.68310546875,
0.81396484375,
0.69775390625,
0.12078857421875,
0.69775390625
],
[
0.09783935546875,
0.70263671875,
0.91552734375,
0.70263671875,
0.91552734375,
0.71728515625,
0.09783935546875,
0.71728515625
],
[
0.12078857421875,
0.72216796875,
0.91552734375,
0.72216796875,
0.91552734375,
0.73583984375,
0.12078857421875,
0.73583984375
],
[
0.1231689453125,
0.74072265625,
0.91162109375,
0.74072265625,
0.91162109375,
0.75537109375,
0.1231689453125,
0.75537109375
],
[
0.11956787109375,
0.7587890625,
0.91650390625,
0.76025390625,
0.91650390625,
0.77734375,
0.11956787109375,
0.7763671875
],
[
0.1231689453125,
0.78076171875,
0.91162109375,
0.78076171875,
0.91162109375,
0.79541015625,
0.1231689453125,
0.79541015625
],
[
0.11712646484375,
0.7958984375,
0.91796875,
0.7978515625,
0.91796875,
0.81787109375,
0.11712646484375,
0.8173828125
],
[
0.12078857421875,
0.81787109375,
0.22705078125,
0.81787109375,
0.22705078125,
0.83251953125,
0.12078857421875,
0.83251953125
],
[
0.093017578125,
0.8349609375,
0.91552734375,
0.8359375,
0.91552734375,
0.85693359375,
0.093017578125,
0.8564453125
],
[
0.11956787109375,
0.85888671875,
0.91552734375,
0.85888671875,
0.91552734375,
0.873046875,
0.11956787109375,
0.873046875
],
[
0.11834716796875,
0.87548828125,
0.42138671875,
0.87744140625,
0.42138671875,
0.89501953125,
0.11834716796875,
0.8935546875
],
[
0.08575439453125,
0.912109375,
0.9130859375,
0.912109375,
0.9130859375,
0.9267578125,
0.08575439453125,
0.9267578125
],
[
0.08331298828125,
0.931640625,
0.91552734375,
0.931640625,
0.91552734375,
0.9462890625,
0.08331298828125,
0.9462890625
]
] |
[
"2",
"L.Zhou,Z.Luo,M.Zhen, T.Shen, et al.",
"LM",
"STBA",
"",
"",
"Timeof RCS",
"",
"Time of the others",
"H",
"0300",
"",
"0",
"1632",
"#computenodes",
"Fig.1.Per-iteration time of bundle adjustment w.r.t. the compute node number. The",
"Levenberg-Marquardt (LM) algorithm is limited by the bottleneck when solving the re",
"duced camera system (RCS).Our STBA splitstheRCSinto independent sub-problems,",
"which achieves a speedup on a single-threaded compute node.Besides, STBA allows",
"parallel and distributed computing with multiplecompute nodes which further im",
"provesthe efficiencyand scalability.",
"为了实现高效且可扩展的重建，已采用群集作为一种有用的做法，将大型问",
"题分解为更小，更易于管理的问题。例如，已经以分而治之的方式开发了许多",
"SfM方法，它们首先独立地重建分区的子图，然后将部分重建合并在一起[32，",
"43，44，18。尽管这些方法能够以高效且可扩展的方式生成初始稀疏重建，但全",
"束调整对于全局优化摄像机和点参数仍然是必不可少的。因此，在BA的背景",
"下，117，40提出的方法将BA的目标分配给拆分的子模型，并在分布式优化框",
"架11，6下优化目标的总和，然而，这涉及额外昂贵的内部迭代，从而使优化",
"过于复杂",
"在这项工作中，我们遵循利用聚类方法的方向，并系统地推进了如何将聚类",
"方案集成到BA问题中的研究。我们不是在预处理步骤中应用固定的一次性分",
"区，而是在每个LM迭代中导出基于随机聚类的策略，以便分解RCS以提高效",
"率和可伸缩性。",
"-First,wereformulatethequadraticprogrammingproblemof anLM iteration",
"based on the clustering of the visibility graph. Such a formulation splits",
"the problem into the most elementary structures, but meanwhile introduces",
"additional equality constraints and raises the computational cost.",
"-Second,inordertomaketheaboveproblemefficientlysolvable,wepropose",
"to relax the constraints into chance constraints25l and then solve it with",
"sampled convex program[8]:The approach helps to eliminate the inter-",
"dependence between different clusters by randomized constraint reduction,",
"whichhencedecomposestheRCS intoindependentlinear sub-problemsre",
"lated to the clusters. In this way, an approximate step can be achieved",
"efficiently.",
"- Third, we present an add-on technique which helps to correct the approxi-",
"mate steps towardsthesteepestdescent direction within a small trust region",
"to improve the convergence.",
"由于采样凸程序引起的随机过程，我们将我们的算法称为随机束调整（STBA）",
"它带来了以下明显的优势。首先，由于降低了复杂性，解决拆分RCS代替原始"
] |
{
"arxiv_id": "2008.00446",
"language": "zh",
"page_idx": 1
}
|
||
2008.00446-zh-2
|
2008.00446-zh-2.jpg
|
3
Stochastic Bundle Adjustment for Efficient and Scalable 3D Reconstruction
RCS 实现了极大的加速。其次，求解过程可以并行且可扩展，以适应摄像机数量的增长，因为 BA 迭代的所有子步骤都可以分解。InFig。1 ,我们可视化了如何通过在多个计算节点上分布 STBA 来减少运行时间
# 2 Related Works
束调整(BA)是 3D 重建[42,29,28,41,44,43]必不可少的步骤。通常由 LevenbergMarquardt (LM) 算法 [27] 求解，该算法将局部信任区域内的误差函数近似线性化，然后求解更新步骤的线性正规方程。SBA [26]首先利用特殊问题结构，通过 Schur 补码将范数方程简化为简化的相机系统(RCS)。在此之后，致力于以精确或不精确的方式更快地解决 RCS。精确求解器将 Cholesky 分解应用于简化的相机矩阵$S$ ,同时利用变量排序[4,13]和超节点方法[33,13]进行加速。不精确求解器基于共轭梯度(CG)方法[20]与各种预处理器[3,21,24]相结合，从而以更高的效率获得不精确解。除了算法改进外，[23,38]还提供了 LM 求解器的优化实现，以节省内存使用并利用 CPU 和 GPU 并行性。然而，尽管做出了上述努力，但当问题扩大时，解决一个大而不可分割的 RCS 将越来越成为 BA 求解器的瓶颈
为了使大规模重建易于处理，最初将聚类方法从运动(SfM)域引人结构中。基本上，应用了分而治之的策略，该策略首先将大型场景划分为多个子地图，然后将局部重建全局合并[32,43,44,18] 。在这些方法的表述中，解决了除原始BA 问题之外的简化优化问题，从而导致了次优结果。例如，[32,44]将内部变量分解在子映射中，而[43]将所有摄像机注册为运动平均[10],而不涉及点。在BA 领域，[24] 从摄像机的聚类中得出了 RCS 的块对角预处理器，但是聚类并没有像在 SfM 算法[32,43,44,18]中所做的那样帮助分解问题。相反，[17,40] 建议将分布式优化框架(如 Douglas-Rachford 方法[11]和 ADMM[6])应用于经验聚类的大规模 BA 问题。虽然建立在理论基础上，但这些方法需要昂贵的内部迭代，并在优化过程中引人了过多的潜在参数。
# 3 Bundle Adjustment Revisited
在本节中，我们首先重新讨论捆绑调整问题及其 LM 解决方案，以给出必要的
预备和术语。此后，向量和矩阵以粗体显示，$\|.\|$表示 L2 范数
一个包调整问题建立在二分可见性图$\mathcal{G}=(\mathcal{C}\cup\mathcal{P},\mathcal{E})$上。这里，$\mathcal{C}=\{c_i\in$ $\mathbb{R}^d\}_{i=1}^m$表示由$d$维向量参数化的$m$相机的集合，$\mathcal{P}=\{\boldsymbol p_i\in\mathbb{R}^3\}_i=1^n$表示？??D 点的集合$,\mathcal{E}=\{\boldsymbol{q}_i\in\mathbb{R}^2\}_{i=1}^q$表示$q$投影的集合。目标是最小化$F(x)=\|\boldsymbol{f}(x)\|^2$ ,其中$f$表示重投影误差的$2q$维向量，$x$连接相机参数$c\in\mathbb{R}^dm$和点参数$p\in \mathbb{R} ^{3n}$, $i. e.$ , $x= [ \boldsymbol{c}^T, \boldsymbol{p}^T] ^T$
LM 算法通过在$x$周围的信任区域中将$f(x)$线性化为$J(x)\Delta x+f(x)$ ,在每次迭代时实现更新步骤$\Delta x$ ,其中$J(x)=\nabla_xf(x)=[J(c),J(p)]$是雅可比矩阵。然后将$F(x)$的最小化转换为
$$\min_{\Delta\boldsymbol{x}}\|\boldsymbol{J}(\boldsymbol{x})\Delta\boldsymbol{x}+f(\boldsymbol{x})\|^2+\lambda\|\boldsymbol{D}\Delta\boldsymbol{x}\|^2,$$
,其解来自以下正规方程
$$\begin{bmatrix}J(\boldsymbol{x})\\\sqrt{\lambda}D\end{bmatrix}\Delta\boldsymbol{x}=\begin{bmatrix}-\boldsymbol{f}(\boldsymbol{x})\\\boldsymbol{0}\end{bmatrix},$$
(2)
|
[
[
0.09906005859375,
0.0186767578125,
0.8359375,
0.0186767578125,
0.8359375,
0.03167724609375,
0.09906005859375,
0.03167724609375
],
[
0.90087890625,
0.0186767578125,
0.92041015625,
0.0186767578125,
0.92041015625,
0.033294677734375,
0.90087890625,
0.033294677734375
],
[
0.08331298828125,
0.06011962890625,
0.91552734375,
0.06011962890625,
0.91552734375,
0.07391357421875,
0.08331298828125,
0.07391357421875
],
[
0.08331298828125,
0.07958984375,
0.91162109375,
0.07958984375,
0.91162109375,
0.09423828125,
0.08331298828125,
0.09423828125
],
[
0.0821533203125,
0.0982666015625,
0.66650390625,
0.0982666015625,
0.66650390625,
0.1129150390625,
0.0821533203125,
0.1129150390625
],
[
0.0809326171875,
0.1446533203125,
0.3466796875,
0.1446533203125,
0.3466796875,
0.1591796875,
0.0809326171875,
0.1591796875
],
[
0.0809326171875,
0.1754150390625,
0.91650390625,
0.1778564453125,
0.91650390625,
0.19580078125,
0.0809326171875,
0.193359375
],
[
0.0821533203125,
0.1973876953125,
0.9130859375,
0.1973876953125,
0.9130859375,
0.2120361328125,
0.0821533203125,
0.2120361328125
],
[
0.0821533203125,
0.2169189453125,
0.9130859375,
0.2169189453125,
0.9130859375,
0.2315673828125,
0.0821533203125,
0.2315673828125
],
[
0.0821533203125,
0.2371826171875,
0.9130859375,
0.2371826171875,
0.9130859375,
0.251708984375,
0.0821533203125,
0.251708984375
],
[
0.0821533203125,
0.256591796875,
0.91552734375,
0.256591796875,
0.91552734375,
0.271240234375,
0.0821533203125,
0.271240234375
],
[
0.08331298828125,
0.275390625,
0.91552734375,
0.275390625,
0.91552734375,
0.2900390625,
0.08331298828125,
0.2900390625
],
[
0.08331298828125,
0.294189453125,
0.91064453125,
0.294189453125,
0.91064453125,
0.30859375,
0.08331298828125,
0.30859375
],
[
0.08331298828125,
0.314453125,
0.9130859375,
0.314453125,
0.9130859375,
0.3291015625,
0.08331298828125,
0.3291015625
],
[
0.08331298828125,
0.333984375,
0.91162109375,
0.333984375,
0.91162109375,
0.348388671875,
0.08331298828125,
0.348388671875
],
[
0.0821533203125,
0.351806640625,
0.91552734375,
0.350830078125,
0.91552734375,
0.367919921875,
0.0821533203125,
0.36962890625
],
[
0.0821533203125,
0.371337890625,
0.224609375,
0.371337890625,
0.224609375,
0.385986328125,
0.0821533203125,
0.385986328125
],
[
0.11834716796875,
0.3916015625,
0.91162109375,
0.3916015625,
0.91162109375,
0.40625,
0.11834716796875,
0.40625
],
[
0.08331298828125,
0.411865234375,
0.91162109375,
0.411865234375,
0.91162109375,
0.426513671875,
0.08331298828125,
0.426513671875
],
[
0.08331298828125,
0.4306640625,
0.9130859375,
0.4306640625,
0.9130859375,
0.445068359375,
0.08331298828125,
0.445068359375
],
[
0.0821533203125,
0.450927734375,
0.91552734375,
0.450927734375,
0.91552734375,
0.465576171875,
0.0821533203125,
0.465576171875
],
[
0.0809326171875,
0.46875,
0.91650390625,
0.469482421875,
0.91650390625,
0.4873046875,
0.0809326171875,
0.48583984375
],
[
0.0809326171875,
0.48828125,
0.9130859375,
0.489013671875,
0.9130859375,
0.50341796875,
0.0809326171875,
0.5029296875
],
[
0.0821533203125,
0.50927734375,
0.9130859375,
0.50927734375,
0.9130859375,
0.52294921875,
0.0821533203125,
0.52294921875
],
[
0.08331298828125,
0.52783203125,
0.9130859375,
0.52783203125,
0.9130859375,
0.54248046875,
0.08331298828125,
0.54248046875
],
[
0.0821533203125,
0.54736328125,
0.9130859375,
0.54736328125,
0.9130859375,
0.56201171875,
0.0821533203125,
0.56201171875
],
[
0.08575439453125,
0.56689453125,
0.6181640625,
0.56689453125,
0.6181640625,
0.58154296875,
0.08575439453125,
0.58154296875
],
[
0.0821533203125,
0.61181640625,
0.5556640625,
0.61181640625,
0.5556640625,
0.62646484375,
0.0821533203125,
0.62646484375
],
[
0.08331298828125,
0.6474609375,
0.9130859375,
0.6474609375,
0.9130859375,
0.662109375,
0.08331298828125,
0.662109375
],
[
0.08331298828125,
0.666015625,
0.740234375,
0.666015625,
0.740234375,
0.6806640625,
0.08331298828125,
0.6806640625
],
[
0.11956787109375,
0.68310546875,
0.91650390625,
0.68408203125,
0.91650390625,
0.70166015625,
0.11956787109375,
0.7001953125
],
[
0.0821533203125,
0.70361328125,
0.91650390625,
0.70361328125,
0.91650390625,
0.72119140625,
0.0821533203125,
0.72119140625
],
[
0.0821533203125,
0.7236328125,
0.91552734375,
0.7236328125,
0.91552734375,
0.74072265625,
0.0821533203125,
0.74072265625
],
[
0.0821533203125,
0.74267578125,
0.91650390625,
0.74267578125,
0.91650390625,
0.76025390625,
0.0821533203125,
0.76025390625
],
[
0.0772705078125,
0.76025390625,
0.408203125,
0.76025390625,
0.408203125,
0.78076171875,
0.0772705078125,
0.78076171875
],
[
0.11712646484375,
0.77978515625,
0.91650390625,
0.78076171875,
0.91650390625,
0.79833984375,
0.11712646484375,
0.7978515625
],
[
0.08331298828125,
0.802734375,
0.9130859375,
0.802734375,
0.9130859375,
0.8173828125,
0.08331298828125,
0.8173828125
],
[
0.08331298828125,
0.8212890625,
0.46142578125,
0.8212890625,
0.46142578125,
0.8359375,
0.08331298828125,
0.8359375
],
[
0.302001953125,
0.84912109375,
0.69189453125,
0.84716796875,
0.69189453125,
0.865234375,
0.302001953125,
0.86572265625
],
[
0.88623046875,
0.84912109375,
0.91796875,
0.84912109375,
0.91796875,
0.8681640625,
0.88623046875,
0.8681640625
],
[
0.30908203125,
0.86376953125,
0.338134765625,
0.86572265625,
0.335693359375,
0.87646484375,
0.3056640625,
0.873046875
],
[
0.0772705078125,
0.88720703125,
0.34423828125,
0.884765625,
0.34423828125,
0.8994140625,
0.0772705078125,
0.90185546875
],
[
0.36474609375,
0.904296875,
0.64013671875,
0.904296875,
0.64013671875,
0.9599609375,
0.36474609375,
0.9599609375
],
[
0.88623046875,
0.923828125,
0.91650390625,
0.923828125,
0.91650390625,
0.9423828125,
0.88623046875,
0.9423828125
],
[
0.39013671875,
0.93505859375,
0.402099609375,
0.93505859375,
0.402099609375,
0.93896484375,
0.39013671875,
0.93896484375
]
] |
[
"StochasticBundleAdjustment forEfficientandScalable3DReconstruction",
"3",
"RCS实现了极大的加速。其次，求解过程可以并行且可扩展，以适应摄像机数",
"量的增长，因为BA迭代的所有子步骤都可以分解。InFig。1，我们可视化了",
"如何通过在多个计算节点上分布STBA来减少运行时间",
"2RelatedWorks",
"束调整(BA)是3D重建[42,29,28,41,44,43]必不可少的步骤。通常由Levenberg",
"Marquardt（LM）算法27]求解，该算法将局部信任区域内的误差函数近似线",
"性化，然后求解更新步骤的线性正规方程。SBA26]首先利用特殊问题结构，通",
"过Schur补码将范数方程简化为简化的相机系统（RCS）。在此之后，致力于以",
"精确或不精确的方式更快地解决RCS。精确求解器将Cholesky分解应用于简化",
"的相机矩阵S，同时利用变量排序413]和超节点方法[33.13]进行加速。不精",
"确求解器基于共轭梯度（CG）方法[20]与各种预处理器[3.21.24]相结合，从",
"而以更高的效率获得不精确解。除了算法改进外，23,38还提供了LM求解器",
"的优化实现，以节省内存使用并利用CPU和GPU并行性。然而，尽管做出了",
"上述努力，但当问题扩大时，解决一个大而不可分割的RCS将越来越成为BA",
"求解器的瓶颈",
"为了使天规模重建易于处理，最初将聚类方法从运动（SfM）域引人结构中",
"基本上，应用了分而治之的策略，该策略首先将大型场景划分为多个子地图，然",
"后将局部重建全局合并[32.43.44,18]。在这些方法的表述中，解决了除原始",
"BA问题之外的简化优化问题，从而导致了次优结果。例如，132，44将内部变量",
"分解在子映射中，而[43]将所有摄像机注册为运动平均[10]，而不涉及点。在",
"BA领域，[24I从摄像机的聚类中得出了RCS的块对角预处理器，但是聚类并",
"没有像在SfM算法32,43，44,18中所做的那样帮助分解问题。相反，17.40",
"建议将分布式优化框架（如Douglas-Rachford方法11和ADMM[6）应用于",
"经验聚类的大规模BA问题。虽然建立在理论基础上，但这些方法需要昂贵的",
"内部迭代，并在优化过程中引人了过多的潜在参数。",
"3BundleAdjustmentRevisited",
"在本节中，我们首先重新讨论捆绑调整问题及其LM解决方案，以给出必要的",
"预备和术语。此后，向量和矩阵以粗体显示，1表示L2范数",
"一个包调整问题建立在二分可见性图G=（CUP,）上。这里，C=【ciE",
"R)\"i表示由d维向量参数化的m相机的集合，P=(P,ER3=-1表示???D",
"点的集合，ε=(qER=表示q投影的集合。目标是最小化F(α)=f(aα)",
"，其中f表示重投影误差的2q维向量，&连接相机参数cERdm和点参数",
"pER3n,i.e.,=[cT,pTJT",
"LM算法通过在a周围的信任区域中将f(α）线性化为J(α)△a+f(α），在",
"每次迭代时实现更新步骤△a，其中J(a）=f(a）=J(c),J(p)l是雅可比",
"矩阵。然后将F(α）的最小化转换为",
"minlJ(α)Aa+f(α)2+DAa2",
"(1)",
"Ar",
"，其解来自以下正规方程",
"[] - [] ",
"(2)",
""
] |
{
"arxiv_id": "2008.00446",
"language": "zh",
"page_idx": 2
}
|
||
2008.00446-zh-3
|
2008.00446-zh-3.jpg
|
4 L. Zhou, Z. Luo, M. Zhen, T. Shen, $et\textit{al}$
$\begin{array}{c},\text{ 其中 }\lambda>0\text{ 是阻尼参数，通常是 }D=\operatorname{diag}(J^T(x)J(x))^{\frac12}\text{。为了简化符号，我}\\\text{们编写 }J\triangleq J(x)\mathrm{~,~}J_c\triangleq J(c)\mathrm{~,~}J_p\triangleq J(p)\text{ 和 }f\triangleq f(x)\text{ 。在将 }[J^T,\sqrt{\lambda}D^T]\end{array}$
乘以等式$\operatorname{E}$2 的两侧之后，我们得到了
$$(J^TJ+\lambda D^TD)\Delta\boldsymbol{x}=-\boldsymbol{J}^T\boldsymbol{f},$$
(3)
,它可以以
$$\begin{bmatrix}B&E\\E^T&C\end{bmatrix}\begin{bmatrix}\Delta c\\\Delta p\end{bmatrix}=\begin{bmatrix}v\\w\end{bmatrix},$$
(4)
的形式重写，其中$B=\boldsymbol J_c^T\boldsymbol{J}_c+\lambda$diag$( \boldsymbol{J}_c^T\boldsymbol{J}_c)$ , XMATHLIN5。可以通过 Schur
补码[26]简化方程E4 ,从而得出
(5)
$$\begin{aligned}S\Delta c&=v-EC^{-1}w,\\\Delta p&=C^{-1}(\boldsymbol{w}-\boldsymbol{E}^{T}\Delta\boldsymbol{c}),\end{aligned}$$
(6)
,其中$S=B-EC^{-1}E^T$是$C$的 Schur 补码。在这里，$S$ ,被称为简化的相机矩阵，是一个块结构的对称正定矩阵。块$S_{ij}\in\mathbb{R}^{d\times d}$是非零 iff 相机$c_i$和$c_j$ 至少观察到一个共同点。尽管已经开发了各种稀疏 Cholesky 分解技术[4,13,33] 和预处理共轭梯度方法 [3,21,24] 来求解方程E5 的简化相机系统(RCS),但是当相机数量$m$变大时，它仍然会令人望而却步
### 4 Stochastic Bundle Adjustment
在本节中，我们介绍了将 RCS 分解为 LM 迭代内的簇的随机束调整 (STBA) 方法。在 SEC。$\boxed{\mathrm{E}}4.1$,我们首先基于可见性图$\mathcal{G}$的聚类重新制定问题(1),但受到其他等式约束。接下来，在 Sec。4.2 , 我们将机会约束松弛应用于重构， 并通过采样凸程序[8,9]求解。它设法将 RCS 分解为与群集相关的线性子问题， 并有效地产生近似的 STBA 步骤。第三，提出了一个最陡的校正步骤来弥补 Sec 中 STBA 步骤的逼近误差。4.3 。最后，在 Sec。4.4 , 我们提出了机会约束松弛所需的随机约束采样器的实际实现
4.1 Clustering Based Reformulation
与先前在预处理阶段对问题进行分区的[32,43,44,18,17,40]方法相反，我们提供
出了问题的重新表述(1),以将 RCS 分解为 LM 迭代内的簇
特别是，我们考虑最常见的情况，即每个摄像机形成一个集群。为了保留所有投影$\mathcal{E}$ ,我们将点拆分应用于物理点，如图。2。对于$v_i$摄像机查看的物理点$p_i$ ,我们将其拆分为$v_i$虚拟点$\{\boldsymbol p_i^j\}_j=1^{v_i}$,每个虚拟点分配给一个群集。这样的聚类将把问题(1)等价地重新表述为新的约束二次规划 (QP) 问题，如下所示
(7)
$$\begin{aligned}&\min_{\Delta x^{\prime}}\:\|\boldsymbol{J}^{\prime}\Delta\boldsymbol{x}^{\prime}+\boldsymbol{f}\|^{2}+\lambda\|\boldsymbol{D}^{\prime}\Delta\boldsymbol{x}^{\prime}\|^{2},\\&\mathrm{s.t.}\:\boldsymbol{A}\Delta\boldsymbol{x}^{\prime}=\boldsymbol{0}.\end{aligned}$$
(8)
|
[
[
0.0784912109375,
0.017059326171875,
0.1014404296875,
0.017059326171875,
0.1014404296875,
0.03167724609375,
0.0784912109375,
0.03167724609375
],
[
0.156982421875,
0.017059326171875,
0.56396484375,
0.017059326171875,
0.56396484375,
0.03167724609375,
0.156982421875,
0.03167724609375
],
[
0.08575439453125,
0.056854248046875,
0.91552734375,
0.056854248046875,
0.91552734375,
0.07470703125,
0.08575439453125,
0.07470703125
],
[
0.0821533203125,
0.07635498046875,
0.91650390625,
0.07635498046875,
0.91650390625,
0.09423828125,
0.0821533203125,
0.09423828125
],
[
0.0784912109375,
0.0966796875,
0.485595703125,
0.09747314453125,
0.485595703125,
0.1153564453125,
0.0784912109375,
0.11456298828125
],
[
0.343017578125,
0.1307373046875,
0.65576171875,
0.1307373046875,
0.65576171875,
0.15185546875,
0.343017578125,
0.15185546875
],
[
0.88623046875,
0.1331787109375,
0.91796875,
0.1331787109375,
0.91796875,
0.1527099609375,
0.88623046875,
0.1527099609375
],
[
0.07489013671875,
0.1697998046875,
0.202880859375,
0.168212890625,
0.202880859375,
0.1868896484375,
0.07489013671875,
0.1884765625
],
[
0.37451171875,
0.1787109375,
0.6181640625,
0.1787109375,
0.6181640625,
0.2340087890625,
0.37451171875,
0.2340087890625
],
[
0.88623046875,
0.1998291015625,
0.91796875,
0.1998291015625,
0.91796875,
0.218505859375,
0.88623046875,
0.218505859375
],
[
0.5654296875,
0.211181640625,
0.59765625,
0.211181640625,
0.59765625,
0.223388671875,
0.5654296875,
0.223388671875
],
[
0.08331298828125,
0.2420654296875,
0.91552734375,
0.2420654296875,
0.91552734375,
0.260009765625,
0.08331298828125,
0.260009765625
],
[
0.08331298828125,
0.263916015625,
0.442138671875,
0.263916015625,
0.442138671875,
0.278564453125,
0.08331298828125,
0.278564453125
],
[
0.35986328125,
0.295654296875,
0.5859375,
0.29736328125,
0.5859375,
0.318359375,
0.35986328125,
0.31591796875
],
[
0.88525390625,
0.298828125,
0.91796875,
0.298828125,
0.91796875,
0.318359375,
0.88525390625,
0.318359375
],
[
0.3779296875,
0.324951171875,
0.638671875,
0.324951171875,
0.638671875,
0.34521484375,
0.3779296875,
0.34521484375
],
[
0.88525390625,
0.32568359375,
0.91796875,
0.32568359375,
0.91796875,
0.34521484375,
0.88525390625,
0.34521484375
],
[
0.08575439453125,
0.364013671875,
0.91650390625,
0.364013671875,
0.91650390625,
0.3818359375,
0.08575439453125,
0.3818359375
],
[
0.0784912109375,
0.3818359375,
0.91552734375,
0.382568359375,
0.91552734375,
0.40380859375,
0.0784912109375,
0.40283203125
],
[
0.08575439453125,
0.404541015625,
0.9130859375,
0.404541015625,
0.9130859375,
0.419189453125,
0.08575439453125,
0.419189453125
],
[
0.08331298828125,
0.4248046875,
0.9130859375,
0.4248046875,
0.9130859375,
0.439453125,
0.08331298828125,
0.439453125
],
[
0.08331298828125,
0.4443359375,
0.572265625,
0.4443359375,
0.572265625,
0.458984375,
0.08331298828125,
0.458984375
],
[
0.0809326171875,
0.4970703125,
0.56787109375,
0.4970703125,
0.56787109375,
0.51171875,
0.0809326171875,
0.51171875
],
[
0.0821533203125,
0.53955078125,
0.91552734375,
0.53955078125,
0.91552734375,
0.55419921875,
0.0821533203125,
0.55419921875
],
[
0.08331298828125,
0.55908203125,
0.9130859375,
0.55908203125,
0.9130859375,
0.57373046875,
0.08331298828125,
0.57373046875
],
[
0.08331298828125,
0.57763671875,
0.9130859375,
0.57763671875,
0.9130859375,
0.59228515625,
0.08331298828125,
0.59228515625
],
[
0.08331298828125,
0.59765625,
0.91552734375,
0.59765625,
0.91552734375,
0.61181640625,
0.08331298828125,
0.61181640625
],
[
0.0821533203125,
0.61669921875,
0.9130859375,
0.61669921875,
0.9130859375,
0.63134765625,
0.0821533203125,
0.63134765625
],
[
0.0809326171875,
0.6337890625,
0.91552734375,
0.63525390625,
0.91552734375,
0.65234375,
0.0809326171875,
0.6513671875
],
[
0.08331298828125,
0.65576171875,
0.490234375,
0.65576171875,
0.490234375,
0.67041015625,
0.08331298828125,
0.67041015625
],
[
0.0821533203125,
0.705078125,
0.53515625,
0.705078125,
0.53515625,
0.7197265625,
0.0821533203125,
0.7197265625
],
[
0.08575439453125,
0.74072265625,
0.9130859375,
0.74072265625,
0.9130859375,
0.75537109375,
0.08575439453125,
0.75537109375
],
[
0.0809326171875,
0.7587890625,
0.73681640625,
0.7578125,
0.73681640625,
0.77587890625,
0.0809326171875,
0.7763671875
],
[
0.11956787109375,
0.77978515625,
0.9130859375,
0.77978515625,
0.9130859375,
0.79443359375,
0.11956787109375,
0.79443359375
],
[
0.08331298828125,
0.79931640625,
0.91552734375,
0.79931640625,
0.91552734375,
0.81396484375,
0.08331298828125,
0.81396484375
],
[
0.0821533203125,
0.81787109375,
0.91552734375,
0.81787109375,
0.91552734375,
0.8359375,
0.0821533203125,
0.8359375
],
[
0.08331298828125,
0.83935546875,
0.9130859375,
0.83935546875,
0.9130859375,
0.85400390625,
0.08331298828125,
0.85400390625
],
[
0.0809326171875,
0.85791015625,
0.1304931640625,
0.85791015625,
0.1304931640625,
0.873046875,
0.0809326171875,
0.873046875
],
[
0.317626953125,
0.89208984375,
0.6787109375,
0.8876953125,
0.6787109375,
0.90966796875,
0.317626953125,
0.9140625
],
[
0.88525390625,
0.89208984375,
0.91796875,
0.89208984375,
0.91796875,
0.912109375,
0.88525390625,
0.912109375
],
[
0.333251953125,
0.90576171875,
0.341796875,
0.90966796875,
0.3369140625,
0.91552734375,
0.327392578125,
0.90966796875
],
[
0.322509765625,
0.92431640625,
0.488037109375,
0.9228515625,
0.488037109375,
0.93994140625,
0.322509765625,
0.9423828125
],
[
0.88525390625,
0.92431640625,
0.91796875,
0.92431640625,
0.91796875,
0.94482421875,
0.88525390625,
0.94482421875
]
] |
[
"4",
"L.Zhou, Z.Luo, M. Zhen, T.Shen, et al.",
"，其中>>0是阻尼参数，通常是D=diag(JT(a)J(α)。为了简化符号，我",
"们编写JJ(α），J兰J(c)，J≤J(p)和ff(a）。在将[JT,VADT",
"乘以等式E2的两侧之后，我们得到了",
"(JT J + ADTD)△&= -JT f",
"(3)",
"，它可以以",
"[8 [] [8]",
"(4)",
"w",
"的形式重写，其中B=JTJ。+入diag(JTJ），XMATHLIN5。可以通过Schur",
"补码26]简化方程E4，从而得出",
"S△c=U-EC-lw,",
"(5)",
"△p = C-I(w-ET△c),",
"(6)",
"，其中S=B-EC-1ET是C的Schur补码。在这里，S，被称为简化的相",
"机矩阵，是一个块结构的对称正定矩阵。块SiERdxd是非零if相机c,和ci",
"至少观察到一个共同点。尽管已经开发了各种稀疏Cholesky分解技术[4.13.33",
"和预处理共轭梯度方法[3.21,24来求解方程E5的简化相机系统（RCS），但是",
"当相机数量㎡变大时，它仍然会令人望而却步",
"4StochasticBundleAdjustment",
"在本节中，我们介绍了将RCS分解为LM送代内的簇的随机束调整（STBA",
"方法。在SEC。E4.1，我们首先基于可见性图9的聚类重新制定问题（1），但",
"受到其他等式约束。接下来，在Sec。4.2，我们将机会约束松弛应用于重构",
"并通过采样凸程序8，9求解。它设法将RCS分解为与群集相关的线性子问题",
"并有效地产生近似的STBA步骤。第三，提出了一个最陡的校正步骤来弥补Sec",
"中STBA步骤的逼近误差。4.3。最后，在Sec。4.4，我们提出了机会约束",
"松弛所需的随机约束采样器的实际实现",
"4.1Clustering Based Reformulatior",
"与先前在预处理阶段对问题进行分区的32，43.44,18，17，401方法相反，我们提",
"出了问题的重新表述（1），以将RCS分解为LM迭代内的簇",
"特别是，我们考虑最常见的情况，即每个摄像机形成一个集群。为了保留所",
"有投影&，我们将点拆分应用于物理点，如图。2。对于i摄像机查看的物理",
"点P,，我们将其拆分为虚拟点(p)=1，每个虚拟点分配给一个群集。这",
"样的聚类将把问题（1）等价地重新表述为新的约束二次规划（QP）问题，如下",
"所示",
"minIJ'Aa'+fl?+>D'a'12,",
"(7)",
"L",
"s.t.Ac'=0.",
"(8)"
] |
{
"arxiv_id": "2008.00446",
"language": "zh",
"page_idx": 3
}
|
||
2008.00446-zh-4
|
2008.00446-zh-4.jpg
|
5
Stochastic Bundle Adjustment for Efficient and Scalable 3D Reconstruction
<FigureHere>
Fig.2. Illustration of point splitting/binding over the visibility graph (top) and the corresponding structure of the reduced camera matrix (bottom). Point splitting helps to reshape the reduced camera matrix into the block-diagonal structure, while point binding does the inverse.
这里，$\Delta x^\prime=[\Delta c^T,\Delta p^{\prime T}]^T$是$\Delta x$的扩展，它考虑了所有虚拟点的更新步骤， 雅可比$J^\prime=[J_{\boldsymbol{c},\boldsymbol{J}_{p^{\prime}]}}$也是如此。因此，$D^\prime=\operatorname{diag}(\boldsymbol{J}^{\prime T}\boldsymbol{J}^{\prime})^{\frac12}$。问题(7)和(1) 之间值得注意的区别是，施加了等式$\overline{\mathrm{E}}$8 的新等式约束，以强制不同簇中相同点的步骤相同。例如，点$p_i$的$\Delta\boldsymbol{p}_i^s=\Delta\boldsymbol{p}_i^t\left(\forall s,t\in\{1,...,v_i\}\right)$和$A$的相应第 j 行以$\boldsymbol{a}_j=[0...,1,...,-1,...0]$的形式显示。由于一点$p_i$引人了$v_i-1$方程，$A$的行号为$r=\sum_{i=1}^n(v_i-1)$。此外，$A$是全行排名，因为每个定义唯一相等约束的行是线性独立的
上面的约束 QP 问题可以很容易地通过拉格朗日对偶来解决，这将问题转化
为
$$H_\lambda\Delta x^{\prime}=-(J^{\prime}{}^Tf+A^T\nu),$$
(9)
,其中$\boldsymbol H_\lambda=\boldsymbol J^{\prime}^T\boldsymbol{J}^{\prime}+\lambda\boldsymbol{D}^{\prime T}\boldsymbol{D}^{\prime}$和$\boldsymbol\nu=-(\boldsymbol AH_\lambda^{-1}\boldsymbol{A}^T)^{-1}\boldsymbol{A}\boldsymbol{H}_\lambda^{-1}\boldsymbol{J}^{\prime T}f$是拉格朗日乘数。方程式$\overline{\mathrm{E}}$9 的格式与方程式$\overline{\mathrm{E}}3$ 相似，但与方程式$\overline{\mathrm{E}}$3 相比，右侧有一个附加项$A^T\boldsymbol{\nu}$。当$J^{\prime T}f$包含梯度 w.r.t. 独立虚拟点时，$A^T\boldsymbol{\nu}$充当校正项，以确保解决方案符合方程式E8 的约束
基于聚类的重构的最终好处如下所示。通过同样将 Schur 补码应用于方程
式E9 , 我们得到了
$$S'\Delta c=v'-E'C'^{-1}w',$$
(10)
$$E^{\prime}=J_{c}^{T}J_{p}^{\prime},\:C^{\prime}=J_{p}^{\prime}{}^{T}J_{p}^{\prime}+\lambda\mathrm{diag}(J_{p}^{\prime}{}^{T}J_{p}^{\prime})\:,\:S^{\prime}=B-E^{\prime}C^{\prime-1}E^{\prime T}$$
$[\boldsymbol{v}^{\prime T},\boldsymbol{w}^{\prime T}]^T=-(\dot{\boldsymbol{J}}^T\boldsymbol{f}+A^T\boldsymbol{\nu})$。由于任何两个摄像机不共享任何公共虚拟点， 因此$S^{\prime}$现在变为块对角矩阵$i.e.,S_{ij}^{\prime}=\mathbf{0},\forall i\neq j$ 。然后，方程式$\overline{\mathrm{E}}$10 可以等效地分解为$m$个最基本的线性系统，每个系统对应一个摄像机
4.2 Chance Constrained Relaxation
上述基于聚类的重构的主要问题是评估拉格朗日乘数$\nu$的成本过高，因为它需要评估$H_\lambda^{-1}$ 。为了使问题实际上可以解决，我们希望消除通过问题(7)的松弛来评估方程$\mathbb{E}$9 的校正项 $A^T\nu$ 的需要。
我们将随机二进制变量$\theta_i$与方程$\overline{\mathrm{E}}$8 的第 i 个等式约束相乘，得出$\theta_ia_i\Delta x^{\prime}=$ 0。可以解释为，如果 $\theta _i= 1$ , 则必须满足约束 $\boldsymbol{a}_i\Delta x^{\prime }= 0$ ; 否则，允许违反
|
[
[
0.09783935546875,
0.0162506103515625,
0.8369140625,
0.017059326171875,
0.8369140625,
0.03167724609375,
0.09783935546875,
0.0308685302734375
],
[
0.90087890625,
0.017059326171875,
0.92138671875,
0.017059326171875,
0.92138671875,
0.033294677734375,
0.90087890625,
0.033294677734375
],
[
0.42138671875,
0.05767822265625,
0.52392578125,
0.05767822265625,
0.52392578125,
0.06903076171875,
0.42138671875,
0.06903076171875
],
[
0.452880859375,
0.06011962890625,
0.4638671875,
0.06011962890625,
0.4638671875,
0.06414794921875,
0.452880859375,
0.06414794921875
],
[
0.61962890625,
0.05767822265625,
0.72705078125,
0.05767822265625,
0.72705078125,
0.06903076171875,
0.61962890625,
0.06903076171875
],
[
0.7509765625,
0.08203125,
0.91552734375,
0.0828857421875,
0.91552734375,
0.09747314453125,
0.7509765625,
0.0966796875
],
[
0.333251953125,
0.1007080078125,
0.411865234375,
0.10235595703125,
0.411865234375,
0.11700439453125,
0.333251953125,
0.1153564453125
],
[
0.53369140625,
0.1007080078125,
0.6123046875,
0.10235595703125,
0.6123046875,
0.11700439453125,
0.53369140625,
0.1153564453125
],
[
0.1123046875,
0.10882568359375,
0.1666259765625,
0.11126708984375,
0.1666259765625,
0.121826171875,
0.1123046875,
0.119384765625
],
[
0.64990234375,
0.10400390625,
0.6982421875,
0.10400390625,
0.6982421875,
0.11859130859375,
0.64990234375,
0.11859130859375
],
[
0.10992431640625,
0.12347412109375,
0.187255859375,
0.12347412109375,
0.187255859375,
0.1348876953125,
0.10992431640625,
0.1348876953125
],
[
0.09783935546875,
0.10968017578125,
0.11712646484375,
0.10968017578125,
0.11712646484375,
0.119384765625,
0.09783935546875,
0.119384765625
],
[
0.1014404296875,
0.126708984375,
0.113525390625,
0.126708984375,
0.113525390625,
0.1307373046875,
0.1014404296875,
0.1307373046875
],
[
0.338134765625,
0.1307373046875,
0.408203125,
0.1307373046875,
0.408203125,
0.1422119140625,
0.338134765625,
0.1422119140625
],
[
0.53515625,
0.126708984375,
0.61474609375,
0.1300048828125,
0.61376953125,
0.1446533203125,
0.53369140625,
0.141357421875
],
[
0.1231689453125,
0.1373291015625,
0.1787109375,
0.1373291015625,
0.1787109375,
0.1478271484375,
0.1231689453125,
0.1478271484375
],
[
0.623046875,
0.134033203125,
0.6474609375,
0.134033203125,
0.6474609375,
0.1478271484375,
0.623046875,
0.1478271484375
],
[
0.64501953125,
0.1478271484375,
0.673828125,
0.1478271484375,
0.673828125,
0.16650390625,
0.64501953125,
0.16650390625
],
[
0.67138671875,
0.1649169921875,
0.7041015625,
0.1649169921875,
0.7041015625,
0.18359375,
0.67138671875,
0.18359375
],
[
0.75830078125,
0.160888671875,
0.9130859375,
0.160888671875,
0.9130859375,
0.1746826171875,
0.75830078125,
0.1746826171875
],
[
0.6943359375,
0.1851806640625,
0.72705078125,
0.1851806640625,
0.72705078125,
0.1998291015625,
0.6943359375,
0.1998291015625
],
[
0.0821533203125,
0.24365234375,
0.9130859375,
0.24365234375,
0.9130859375,
0.25830078125,
0.0821533203125,
0.25830078125
],
[
0.0821533203125,
0.262451171875,
0.91552734375,
0.262451171875,
0.91552734375,
0.275390625,
0.0821533203125,
0.275390625
],
[
0.0772705078125,
0.277099609375,
0.91650390625,
0.27783203125,
0.91650390625,
0.295654296875,
0.0772705078125,
0.294189453125
],
[
0.0809326171875,
0.295654296875,
0.3212890625,
0.29736328125,
0.3212890625,
0.31201171875,
0.0809326171875,
0.310302734375
],
[
0.0809326171875,
0.348388671875,
0.91552734375,
0.348388671875,
0.91552734375,
0.366455078125,
0.0809326171875,
0.366455078125
],
[
0.0784912109375,
0.36962890625,
0.92041015625,
0.3671875,
0.92041015625,
0.38818359375,
0.0784912109375,
0.389892578125
],
[
0.0821533203125,
0.39404296875,
0.9130859375,
0.39404296875,
0.9130859375,
0.40869140625,
0.0821533203125,
0.40869140625
],
[
0.0821533203125,
0.41357421875,
0.91650390625,
0.41357421875,
0.91650390625,
0.4306640625,
0.0821533203125,
0.4306640625
],
[
0.0821533203125,
0.432861328125,
0.91552734375,
0.432861328125,
0.91552734375,
0.450927734375,
0.0821533203125,
0.450927734375
],
[
0.0784912109375,
0.44921875,
0.91650390625,
0.450927734375,
0.91650390625,
0.47119140625,
0.0784912109375,
0.470458984375
],
[
0.0821533203125,
0.472900390625,
0.274169921875,
0.472900390625,
0.274169921875,
0.4873046875,
0.0821533203125,
0.4873046875
],
[
0.11956787109375,
0.4921875,
0.91552734375,
0.4921875,
0.91552734375,
0.5068359375,
0.11956787109375,
0.5068359375
],
[
0.0784912109375,
0.51025390625,
0.10748291015625,
0.51025390625,
0.10748291015625,
0.52783203125,
0.0784912109375,
0.52783203125
],
[
0.3515625,
0.5302734375,
0.64599609375,
0.5302734375,
0.64599609375,
0.54833984375,
0.3515625,
0.54833984375
],
[
0.88623046875,
0.5302734375,
0.91796875,
0.5302734375,
0.91796875,
0.55078125,
0.88623046875,
0.55078125
],
[
0.0809326171875,
0.5556640625,
0.91796875,
0.55712890625,
0.91796875,
0.5810546875,
0.0809326171875,
0.580078125
],
[
0.0821533203125,
0.580078125,
0.9130859375,
0.5810546875,
0.9130859375,
0.5986328125,
0.0821533203125,
0.59765625
],
[
0.0809326171875,
0.599609375,
0.91650390625,
0.60009765625,
0.91650390625,
0.6181640625,
0.0809326171875,
0.6171875
],
[
0.08331298828125,
0.62158203125,
0.4638671875,
0.62158203125,
0.4638671875,
0.63623046875,
0.08331298828125,
0.63623046875
],
[
0.11834716796875,
0.64013671875,
0.91552734375,
0.64013671875,
0.91552734375,
0.65478515625,
0.11834716796875,
0.65478515625
],
[
0.0784912109375,
0.658203125,
0.291015625,
0.65966796875,
0.291015625,
0.677734375,
0.0784912109375,
0.67578125
],
[
0.373291015625,
0.6767578125,
0.63037109375,
0.6767578125,
0.63037109375,
0.6943359375,
0.373291015625,
0.6943359375
],
[
0.873046875,
0.677734375,
0.91796875,
0.677734375,
0.91796875,
0.69775390625,
0.873046875,
0.69775390625
],
[
0.0784912109375,
0.70263671875,
0.91650390625,
0.70263671875,
0.91650390625,
0.72802734375,
0.0784912109375,
0.72802734375
],
[
0.0809326171875,
0.73095703125,
0.9130859375,
0.734375,
0.9130859375,
0.75146484375,
0.0809326171875,
0.748046875
],
[
0.08575439453125,
0.7548828125,
0.9130859375,
0.7548828125,
0.9130859375,
0.76953125,
0.08575439453125,
0.76953125
],
[
0.08331298828125,
0.7734375,
0.74267578125,
0.7734375,
0.74267578125,
0.7880859375,
0.08331298828125,
0.7880859375
],
[
0.0821533203125,
0.8203125,
0.53369140625,
0.8203125,
0.53369140625,
0.8349609375,
0.0821533203125,
0.8349609375
],
[
0.08575439453125,
0.85302734375,
0.9130859375,
0.85302734375,
0.9130859375,
0.86572265625,
0.08575439453125,
0.86572265625
],
[
0.08331298828125,
0.87158203125,
0.9130859375,
0.87158203125,
0.9130859375,
0.88623046875,
0.08331298828125,
0.88623046875
],
[
0.08331298828125,
0.89208984375,
0.5146484375,
0.89208984375,
0.5146484375,
0.90673828125,
0.08331298828125,
0.90673828125
],
[
0.11834716796875,
0.912109375,
0.91162109375,
0.912109375,
0.91162109375,
0.9267578125,
0.11834716796875,
0.9267578125
],
[
0.0809326171875,
0.931640625,
0.91162109375,
0.931640625,
0.91162109375,
0.9462890625,
0.0809326171875,
0.9462890625
]
] |
[
"StochasticBundleAdjustmentforEfficientandScalable3DReconstruction",
"5",
"",
"",
"AAAA",
"Bipartite visibility graph",
"Point spliting",
"Point splitting",
"Camera",
"0000",
"Physical po",
"",
"",
"Pointbinding",
"Point binding",
"Virtualpo",
"",
"",
"",
"Reduced camera matrix",
"",
"Fig.2.Illustration of point splitting/bindingover thevisibility graph(top)and the",
"corresponding structureofthereduced camera matrix (bottom).Point splittinghelps",
"to reshape the reduced camera matrix into the block-diagonal structure, while point",
"binding doesthe inverse.",
"这里，△a=[△cT△pTT是△的扩展，它考虑了所有虚拟点的更新步骤",
"雅可比J'=[Jc,J也是如此。因此，D'=diag(J\"TJ')。问题（7）和（1）",
"之间值得注意的区别是，施加了等式E8的新等式约束，以强制不同簇中相同点",
"的步骤相同。例如，点p,的△p=△p（Vs,tE{1,,u}）和A的相应第j行",
"以a,=[0..,1.……，-1,..]的形式显示。由于一点p,引人了-1方程，A的",
"行号为r==（ui－1）。此外，A是全行排名，因为每个定义唯一相等约束",
"的行是线性独立的",
"上面的约束QP问题可以很容易地通过拉格朗日对偶来解决，这将问题转化",
"为",
"HAr'=-(J\"Tf +ATv),",
"(9)",
"，其中H>=JTJ'+^D\"TD'和V=-(AH\"AT)-IAH\"J\"Tf是拉格朗",
"日乘数。方程式E9的格式与方程式E3相似，但与方程式E3相比，右侧有一",
"个附加项ATv。当JTf包含梯度w.r.t.独立虚拟点时，A\"v充当校正项，以",
"确保解决方案符合方程式E8的束",
"基于聚类的重构的最终好处如下所示。通过同样将Schur补码应用于方程",
"式E9，我们得到了",
"S'△c=v'-E'ci-lw',",
"(10)",
"，其中E'=JTJ,C=JTJ,+Adiag(J,TJ)，S'=B-E'Cl-IE'T和",
"[uT,w\"TT=-（JTf+ATv）。由于任何两个摄像机不共享任何公共虚拟点",
"因此S现在变为块对角矩阵i.e.，S=0,Vi手i。然后，方程式E10可以等",
"效地分解为㎡个最基本的线性系统，每个系统对应一个摄像机",
"4.2ChanceConstrainedRelaxation",
"上述基于聚类的重构的主要问题是评估拉格朗日乘数V的成本过高，因为它需",
"要评估H一1。为了使问题实际上可以解决，我们希望消除通过问题（7）的松",
"弛来评估方程E9的校正项Av的需要。",
"我们将随机二进制变量0，与方程8的第i个等式约束相乘，得出0;a△a=",
"0。可以解释为，如果；=1，则必须满足约束a△α=0；否则，充许违反"
] |
{
"arxiv_id": "2008.00446",
"language": "zh",
"page_idx": 4
}
|
||
1905.12277-zh-0
|
1905.12277-zh-0.jpg
|
# 学习序列标记中新词的特定任务表示
Minlong Peng, Qi Zhang, Xiaoyu Xing, Tao Gui, Jinlan Fu, Xuanjing Huang Shanghai Key Laboratory of Intelligent Information Processing, Fudan University
School of Computer Science, Fudan University
$825{\mathrm{~Zhangheng~Road,~Shanghai,~China}}$
$\{{\mathrm{~mlpeng16,~qz,~xyxing18,~tgui16,~fujl16,~xjhuang~}}\}{@}$fudan.edu.cn
Abstract
Word representation is a key component in ${\mathrm{neural-network-based}}{\mathrm{sequence}}{\mathrm{labeling}}{\mathrm{sys-}}$ tems. However, representations of unseen or rare words trained on the end task are usually poor for appreciable performance. This is commonly referred to as the out-of-vocabulary ( OOV) problem. In this work, we address the OOV problem in sequence labeling using only training data of the task. $\overset {\mathrm{To}}{\operatorname* { \mathrm{~this}} }$ end, we propose a novel method to predict representations for OOV words from their ${\mathrm{surface-forms~(e.g.,~character~sequence)~and}}$ contexts. The method is specifically designed to avoid the error propagation problem suffered by existing approaches in the same paradigm. To evaluate its effectiveness, we performed extensive empirical studies on four part-ofspeech tagging (POS) tasks and four named entity recognition (NER) tasks. Experimental results show that the proposed method can achieve better or competitive performance on the OOV problem compared with existing state-of-the-art methods.
# 1 Introduction
单词表示(或嵌人)是许多最先进的序列标记系统[Chen et al., 2015; Ma and Hovy, 2016; Zhang and Yang, 2018] 中的基础方面。但是，自然语言在单词上产生了 Zipfian 分布 [Zipf, 1949] 。这意味着大量的单词 (在长尾) 是罕见的。为了控制模型大小，序列标记系统通常将训练词汇表限制为仅覆盖训练集中的前$N$频繁单词。培训词汇中未涵盖的单词称为 OOV 单词。OOV 单词的学习表示受到挑战，因为标准的端到端监督学习方法需要每个单词多次出现才能更好地泛化。许多作品[Ma and Hovy, 2016; Madhyastha et al., 2016]已经证明，在遇到 OOV 单词时，序列标记的性能通常会下降很多。这通常被称为OOV 问题，我们将在这项工作中解决这个问题
在过去的几年中，已经提出了许多方法来处理 OOV 问题。这些方法可以大致分为两类：(1)在非常大的原始文本[Mikolov et al., 2013; Pennington et al., 2014; Peters et al., 2018] 语料库上训练单词表示；(2)进一步利用任务的训练数据。在数据丰富的领域中，第一类方
法通常可以为使用随机初始化单词向量[Devlin et al., 2018]训练的模型带来相当大的改进。然而，当遇到极度数据匮乏的情况时，这种方法可能会受到批评。对于低资源域，获取足够的数据可能很困难，例如。g。在技术领域和生物/医学领域 [Deleger et al., 2016]
因此，在这项工作中，我们强调了第二类方法。此类别的一种流行做法是使用单个共享嵌人表示所有 OOV 单词，该嵌人在训练集中的低频单词上进行训练，然后在测试时分配给所有 OOV 单词。然而 ,这种本质上的启发式解决方案是不雅的，因为它合并了许多单词，从而丢失了 OOV 单词的特定信息。另一种流行的做法是从其表面形式 (e。g。,字符序列) [Ling et al., 2015] 。这种做法成功地捕捉了形态学推导的语义(例如。g。“running” from“run”),但对编码器施加了巨大压力，以捕获语法相似但语义无关的单词之间的语义区别(例如。g。“运行”vs。“rung”)。因此，大多数最先进的序列标记系统将表面形式表示与独特的嵌人相结合，以表示单词。这再次将 OOV 问题引人系统
最近，正在研究第二方向的新范式，我们将其称为师生范式。此范例的方法分两个步骤解决 OOV 问题。在第一步中，他们训练一个监督模型(在这项工作中也称为教师网络),对那些词汇内的单词进行标签预测，以获得其特定于任务的单词表示。在第二步中，他们训练(或启发式地构建)一个预测模型(也称为学生网络),以从其表面形式[Pinter et al., 2017] , 上下文[Lazaridou et al.,2017] 或两者 [Schick and Schütze, 2018] 预测单词的表示。学生网络的训练对象通常是重建那些词汇内单词的表示。在测试时，当遇到句子中的 OOV 词时，他们首先使用学生网络来预测这些 OOV 词的表示形式。然后，基于生成的 OOV 表示，他们使用教师网络执行标签预测
从本质上讲，师生范式的方法可以看作是管道，教师和学生网络是管道的两个级联组件。尽管此范例的方法在多个任务上取得了显着成功，但它们遭受了管道范例[Caselli et al., 2015; Bojarski et al., 2016] 的典型错误传播问题，因为不能保证用于训练学生网络的辅助重建对象与用于训练教师网络的监督对象完全兼容
在这项工作中，我们提出了一种新颖的师生范式方法该方法是专门为解决错误传播问题而设计的。这种方法与现有方法的主要区别在于对学生网络的培训。我们训练学生网络预测可以在监督任务中取得良好表现的单词表示，而不是重建那些词汇内单词的表示。学生网络的训练信号由参数固定的教师网络直接反向传播。这样，我们将学生网络的学习与教师网络联系起来，从而避免了
|
[
[
0.313720703125,
0.11236572265625,
0.685546875,
0.11297607421875,
0.685546875,
0.1307373046875,
0.313720703125,
0.1300048828125
],
[
0.156005859375,
0.146484375,
0.83642578125,
0.147705078125,
0.83642578125,
0.1654052734375,
0.156005859375,
0.1641845703125
],
[
0.1568603515625,
0.17041015625,
0.8388671875,
0.1717529296875,
0.8388671875,
0.186279296875,
0.1568603515625,
0.1849365234375
],
[
0.2998046875,
0.186279296875,
0.6953125,
0.1881103515625,
0.6953125,
0.205810546875,
0.2998046875,
0.203857421875
],
[
0.3349609375,
0.20703125,
0.66259765625,
0.20703125,
0.66259765625,
0.2215576171875,
0.3349609375,
0.2215576171875
],
[
0.22216796875,
0.2266845703125,
0.77685546875,
0.2266845703125,
0.77685546875,
0.2412109375,
0.22216796875,
0.2412109375
],
[
0.25,
0.2734375,
0.324462890625,
0.2734375,
0.324462890625,
0.28857421875,
0.25,
0.28857421875
],
[
0.5146484375,
0.274658203125,
0.90869140625,
0.274658203125,
0.90869140625,
0.28857421875,
0.5146484375,
0.28857421875
],
[
0.5146484375,
0.28857421875,
0.9111328125,
0.28857421875,
0.9111328125,
0.3017578125,
0.5146484375,
0.3017578125
],
[
0.1192626953125,
0.296630859375,
0.453369140625,
0.296630859375,
0.453369140625,
0.31005859375,
0.1192626953125,
0.31005859375
],
[
0.515625,
0.3017578125,
0.91259765625,
0.3017578125,
0.91259765625,
0.31494140625,
0.515625,
0.31494140625
],
[
0.1192626953125,
0.309326171875,
0.451904296875,
0.31005859375,
0.451904296875,
0.324462890625,
0.1192626953125,
0.323974609375
],
[
0.5146484375,
0.31494140625,
0.91259765625,
0.31640625,
0.91259765625,
0.330810546875,
0.5146484375,
0.330078125
],
[
0.1217041015625,
0.32568359375,
0.4453125,
0.32568359375,
0.4453125,
0.3359375,
0.1217041015625,
0.3359375
],
[
0.11767578125,
0.3359375,
0.451904296875,
0.337158203125,
0.451904296875,
0.3515625,
0.11767578125,
0.35107421875
],
[
0.5146484375,
0.330078125,
0.82177734375,
0.330078125,
0.82177734375,
0.343505859375,
0.5146484375,
0.343505859375
],
[
0.53173828125,
0.34521484375,
0.91259765625,
0.34521484375,
0.91259765625,
0.359130859375,
0.53173828125,
0.359130859375
],
[
0.12091064453125,
0.3515625,
0.386474609375,
0.3515625,
0.386474609375,
0.365478515625,
0.12091064453125,
0.365478515625
],
[
0.384033203125,
0.35595703125,
0.396240234375,
0.35595703125,
0.396240234375,
0.35986328125,
0.384033203125,
0.35986328125
],
[
0.396240234375,
0.3515625,
0.451904296875,
0.3515625,
0.451904296875,
0.363037109375,
0.396240234375,
0.363037109375
],
[
0.515625,
0.359130859375,
0.9111328125,
0.359130859375,
0.9111328125,
0.37255859375,
0.515625,
0.37255859375
],
[
0.1217041015625,
0.3662109375,
0.44775390625,
0.3662109375,
0.44775390625,
0.376953125,
0.1217041015625,
0.376953125
],
[
0.51318359375,
0.371826171875,
0.9111328125,
0.371826171875,
0.9111328125,
0.385009765625,
0.51318359375,
0.385009765625
],
[
0.12091064453125,
0.380126953125,
0.451904296875,
0.380126953125,
0.451904296875,
0.393310546875,
0.12091064453125,
0.393310546875
],
[
0.515625,
0.386962890625,
0.91259765625,
0.386962890625,
0.91259765625,
0.40087890625,
0.515625,
0.40087890625
],
[
0.12091064453125,
0.393310546875,
0.451904296875,
0.393310546875,
0.451904296875,
0.4072265625,
0.12091064453125,
0.4072265625
],
[
0.515625,
0.40087890625,
0.9111328125,
0.40087890625,
0.9111328125,
0.4140625,
0.515625,
0.4140625
],
[
0.1217041015625,
0.4091796875,
0.375732421875,
0.4091796875,
0.375732421875,
0.419189453125,
0.1217041015625,
0.419189453125
],
[
0.381591796875,
0.40771484375,
0.450927734375,
0.40771484375,
0.450927734375,
0.41845703125,
0.381591796875,
0.41845703125
],
[
0.515625,
0.4140625,
0.91015625,
0.4140625,
0.91015625,
0.427978515625,
0.515625,
0.427978515625
],
[
0.1192626953125,
0.421630859375,
0.453369140625,
0.421630859375,
0.453369140625,
0.43505859375,
0.1192626953125,
0.43505859375
],
[
0.515625,
0.427978515625,
0.9111328125,
0.427978515625,
0.9111328125,
0.442626953125,
0.515625,
0.442626953125
],
[
0.12091064453125,
0.43505859375,
0.451904296875,
0.43505859375,
0.451904296875,
0.448974609375,
0.12091064453125,
0.448974609375
],
[
0.515625,
0.442626953125,
0.90869140625,
0.442626953125,
0.90869140625,
0.455810546875,
0.515625,
0.455810546875
],
[
0.12091064453125,
0.449462890625,
0.3447265625,
0.449462890625,
0.3447265625,
0.46337890625,
0.12091064453125,
0.46337890625
],
[
0.34228515625,
0.45068359375,
0.449462890625,
0.45068359375,
0.449462890625,
0.4609375,
0.34228515625,
0.4609375
],
[
0.51318359375,
0.453857421875,
0.9111328125,
0.455810546875,
0.9111328125,
0.4697265625,
0.51318359375,
0.46728515625
],
[
0.11767578125,
0.46337890625,
0.450927734375,
0.46337890625,
0.450927734375,
0.4765625,
0.11767578125,
0.4765625
],
[
0.5146484375,
0.468505859375,
0.91015625,
0.4697265625,
0.91015625,
0.484130859375,
0.5146484375,
0.48291015625
],
[
0.1192626953125,
0.4765625,
0.451904296875,
0.4765625,
0.451904296875,
0.490478515625,
0.1192626953125,
0.490478515625
],
[
0.1192626953125,
0.4892578125,
0.44677734375,
0.490478515625,
0.44677734375,
0.50390625,
0.1192626953125,
0.50244140625
],
[
0.515625,
0.484130859375,
0.9111328125,
0.484130859375,
0.9111328125,
0.49755859375,
0.515625,
0.49755859375
],
[
0.11767578125,
0.50244140625,
0.450927734375,
0.50390625,
0.450927734375,
0.517578125,
0.11767578125,
0.5166015625
],
[
0.5146484375,
0.49755859375,
0.91259765625,
0.49755859375,
0.91259765625,
0.51123046875,
0.5146484375,
0.51123046875
],
[
0.11767578125,
0.51513671875,
0.451904296875,
0.517578125,
0.451904296875,
0.5322265625,
0.11767578125,
0.52978515625
],
[
0.5146484375,
0.51123046875,
0.708984375,
0.51123046875,
0.708984375,
0.5244140625,
0.5146484375,
0.5244140625
],
[
0.11767578125,
0.5322265625,
0.451904296875,
0.53076171875,
0.451904296875,
0.54443359375,
0.11767578125,
0.54541015625
],
[
0.529296875,
0.5263671875,
0.91259765625,
0.5263671875,
0.91259765625,
0.54052734375,
0.529296875,
0.54052734375
],
[
0.1192626953125,
0.54541015625,
0.451904296875,
0.54541015625,
0.451904296875,
0.5595703125,
0.1192626953125,
0.5595703125
],
[
0.515625,
0.54052734375,
0.91259765625,
0.54052734375,
0.91259765625,
0.5537109375,
0.515625,
0.5537109375
],
[
0.5146484375,
0.55322265625,
0.91259765625,
0.5537109375,
0.91259765625,
0.5673828125,
0.5146484375,
0.56640625
],
[
0.1192626953125,
0.5595703125,
0.450927734375,
0.5595703125,
0.450927734375,
0.57373046875,
0.1192626953125,
0.57373046875
],
[
0.515625,
0.5673828125,
0.91259765625,
0.5673828125,
0.91259765625,
0.58056640625,
0.515625,
0.58056640625
],
[
0.1192626953125,
0.57373046875,
0.451904296875,
0.57373046875,
0.451904296875,
0.5869140625,
0.1192626953125,
0.5869140625
],
[
0.11767578125,
0.5859375,
0.453369140625,
0.58837890625,
0.453369140625,
0.6015625,
0.11767578125,
0.599609375
],
[
0.515625,
0.58203125,
0.9111328125,
0.58203125,
0.9111328125,
0.59521484375,
0.515625,
0.59521484375
],
[
0.5146484375,
0.5947265625,
0.91015625,
0.59521484375,
0.91015625,
0.609375,
0.5146484375,
0.60791015625
],
[
0.1192626953125,
0.60107421875,
0.298095703125,
0.60107421875,
0.298095703125,
0.6142578125,
0.1192626953125,
0.6142578125
],
[
0.515625,
0.60986328125,
0.91015625,
0.60986328125,
0.91015625,
0.62353515625,
0.515625,
0.62353515625
],
[
0.5146484375,
0.62353515625,
0.91015625,
0.62353515625,
0.91015625,
0.63720703125,
0.5146484375,
0.63720703125
],
[
0.08660888671875,
0.6337890625,
0.1217041015625,
0.63623046875,
0.12091064453125,
0.64794921875,
0.0849609375,
0.64599609375
],
[
0.1192626953125,
0.63623046875,
0.226318359375,
0.63623046875,
0.226318359375,
0.646484375,
0.1192626953125,
0.646484375
],
[
0.515625,
0.63720703125,
0.9111328125,
0.63720703125,
0.9111328125,
0.65087890625,
0.515625,
0.65087890625
],
[
0.0882568359375,
0.65380859375,
0.483642578125,
0.65380859375,
0.483642578125,
0.66748046875,
0.0882568359375,
0.66748046875
],
[
0.515625,
0.65087890625,
0.91015625,
0.65087890625,
0.91015625,
0.6640625,
0.515625,
0.6640625
],
[
0.0882568359375,
0.66650390625,
0.483642578125,
0.66650390625,
0.483642578125,
0.68115234375,
0.0882568359375,
0.68115234375
],
[
0.515625,
0.6640625,
0.9111328125,
0.6640625,
0.9111328125,
0.67822265625,
0.515625,
0.67822265625
],
[
0.0906982421875,
0.681640625,
0.483642578125,
0.681640625,
0.483642578125,
0.69580078125,
0.0906982421875,
0.69580078125
],
[
0.515625,
0.6787109375,
0.91259765625,
0.6787109375,
0.91259765625,
0.6923828125,
0.515625,
0.6923828125
],
[
0.08905029296875,
0.69580078125,
0.484375,
0.69580078125,
0.484375,
0.708984375,
0.08905029296875,
0.708984375
],
[
0.5146484375,
0.693359375,
0.56201171875,
0.693359375,
0.56201171875,
0.705078125,
0.5146484375,
0.705078125
],
[
0.08905029296875,
0.71044921875,
0.483642578125,
0.71044921875,
0.483642578125,
0.72021484375,
0.08905029296875,
0.72021484375
],
[
0.53125,
0.70849609375,
0.91259765625,
0.70849609375,
0.91259765625,
0.7216796875,
0.53125,
0.7216796875
],
[
0.08905029296875,
0.7236328125,
0.484375,
0.7236328125,
0.484375,
0.7373046875,
0.08905029296875,
0.7373046875
],
[
0.5146484375,
0.7216796875,
0.9111328125,
0.7216796875,
0.9111328125,
0.73486328125,
0.5146484375,
0.73486328125
],
[
0.0882568359375,
0.7373046875,
0.484375,
0.7373046875,
0.484375,
0.75048828125,
0.0882568359375,
0.75048828125
],
[
0.5146484375,
0.73486328125,
0.9111328125,
0.73486328125,
0.9111328125,
0.74853515625,
0.5146484375,
0.74853515625
],
[
0.0882568359375,
0.75048828125,
0.484375,
0.75048828125,
0.484375,
0.7646484375,
0.0882568359375,
0.7646484375
],
[
0.5146484375,
0.74853515625,
0.91015625,
0.75,
0.91015625,
0.76318359375,
0.5146484375,
0.76220703125
],
[
0.08905029296875,
0.76513671875,
0.483642578125,
0.76513671875,
0.483642578125,
0.77880859375,
0.08905029296875,
0.77880859375
],
[
0.5146484375,
0.76318359375,
0.9111328125,
0.76318359375,
0.9111328125,
0.7763671875,
0.5146484375,
0.7763671875
],
[
0.0882568359375,
0.77880859375,
0.484375,
0.77880859375,
0.484375,
0.79248046875,
0.0882568359375,
0.79248046875
],
[
0.51318359375,
0.7763671875,
0.82421875,
0.7763671875,
0.82421875,
0.79052734375,
0.51318359375,
0.79052734375
],
[
0.08905029296875,
0.79248046875,
0.484375,
0.79248046875,
0.484375,
0.80615234375,
0.08905029296875,
0.80615234375
],
[
0.529296875,
0.79248046875,
0.90869140625,
0.79248046875,
0.90869140625,
0.80615234375,
0.529296875,
0.80615234375
],
[
0.08660888671875,
0.80615234375,
0.4248046875,
0.80615234375,
0.4248046875,
0.8193359375,
0.08660888671875,
0.8193359375
],
[
0.51318359375,
0.8046875,
0.91015625,
0.806640625,
0.91015625,
0.82080078125,
0.51318359375,
0.818359375
],
[
0.10296630859375,
0.8193359375,
0.484375,
0.8173828125,
0.484375,
0.83203125,
0.10296630859375,
0.833984375
],
[
0.5185546875,
0.8212890625,
0.91015625,
0.8212890625,
0.91015625,
0.83203125,
0.5185546875,
0.83203125
],
[
0.08905029296875,
0.833984375,
0.484375,
0.833984375,
0.484375,
0.84765625,
0.08905029296875,
0.84765625
],
[
0.51318359375,
0.83203125,
0.91259765625,
0.833984375,
0.91259765625,
0.8486328125,
0.51318359375,
0.8466796875
],
[
0.0882568359375,
0.84765625,
0.484375,
0.84765625,
0.484375,
0.861328125,
0.0882568359375,
0.861328125
],
[
0.51318359375,
0.84765625,
0.91015625,
0.8466796875,
0.91015625,
0.861328125,
0.51318359375,
0.8623046875
],
[
0.08660888671875,
0.861328125,
0.484375,
0.85986328125,
0.484375,
0.875,
0.08660888671875,
0.87548828125
],
[
0.51318359375,
0.85986328125,
0.9111328125,
0.861328125,
0.9111328125,
0.87548828125,
0.51318359375,
0.875
],
[
0.0882568359375,
0.87548828125,
0.484375,
0.87548828125,
0.484375,
0.8896484375,
0.0882568359375,
0.8896484375
],
[
0.5146484375,
0.87548828125,
0.91015625,
0.87548828125,
0.91015625,
0.8896484375,
0.5146484375,
0.8896484375
]
] |
[
"学习序列标记中新词的特定任务表示",
"Minlong Peng, Qi Zhang, Xiaoyu Xing, Tao Gui, Jinlan Fu, Xuanjing Huang",
"Shanghai Key Laboratory of Intelligent Information Processing, Fudan University",
" School of Computer Science, Fudan University",
"825 Zhangheng Road, Shanghai, China",
"f mlpeng16, qz, xyxing18, tgui16, fujl16, xjhuang 1 @fudan.edu.cn",
"Abstract",
"法通常可以为使用随机初始化单词向量[Devlinetal.",
"2018]训练的模型带来相当大的改进。然而，当遇到极度",
"Word representation is a key component in",
"数据乏的情况时，这种方法可能会受到批评。对于低",
"neural-network-based sequence labeling sys-",
"资源域，获取足够的数据可能很困难，例如。g。在技术",
"tems.However,representationsof unseen o",
"rare words trained on the end task are usually",
"领域和生物/医学领域[Delegeretal.，2016]",
"因此，在这项工作中，我们强调了第二类方法。此类别",
"poor for appreciable performance.",
"",
"This is",
"的一种流行做法是使用单个共享嵌人表示所有OOV单",
"commonlyreferred toastheout-of-vocabulary",
"词，该嵌人在训练集中的低频单词上进行训练，然后在",
"(OOV)problem.In this work,we address",
"测试时分配给所有OOV单词。然而，这种本质上的启发",
"the OOV problem in sequence labeling using",
"式解决方案是不雅的，因为它合并了许多单词，从而丢",
"onlytraining dataof the task.",
"To this",
"失了OOV单词的特定信息。另一种流行的做法是从其",
"end, we propose a novel method to predict",
"表面形式(e。g。，字符序列）[Lingetal.,2015]。这种做",
"representations for OoV words from their",
"法成功地捕捉了形态学推导的语义（例如。g。“running",
"surface-forms (e.g., character",
"sequence)ano",
"from“run\"），但对编码器施加了巨大压力，以捕获语法",
"contexts. The method is specifically designed",
"相似但语义无关的单词之间的语义区别（例如。g。“运",
"to avoid the error propagation problem suffered",
"by existing approaches in the same paradigm",
"行”vs。“rung”)。因此，大多数最先进的序列标记系统",
"To evaluate its effectiveness, we performed",
"将表面形式表示与独特的嵌人相结合，以表示单词。这",
"extensive empirical studies on four part-of-",
"再次将OOV问题引人系统",
"speech tagging (POS)tasks and four named",
"最近，正在研究第二方向的新范式，我们将其称为师",
"entity recognition (NER) tasks: Experimental",
"生范式。此范例的方法分两个步骤解决OOV问题。在",
"第一步中，他们训练一个监督模型（在这项工作中也称",
"results show that the proposed method can",
"为教师网络），对那些词汇内的单词进行标签预测，以获",
"achieve better or competitive performance on",
"the Oovproblemcomparedwith existing",
"得其特定于任务的单词表示。在第二步中，他们训练（或",
"启发式地构建）一个预测模型（也称为学生网络），以从",
"state-of-the-art methods.",
"其表面形式[Pinteret al.,2017]，上下文[Lazaridouet",
"al.，2017]或两者[SchickandSchitze,2018]预测单词的",
"1",
"Introduction",
"表示。学生网络的训练对象通常是重建那些词汇内单词",
"单词表示（或嵌人）是许多最先进的序列标记系统【Chen",
"的表示。在测试时，当遇到句子中的OOV词时，他们",
"et al., 2015; Ma and Hovy, 2016; Zhang and Yang, 2018",
"首先使用学生网络来预测这些OOV词的表示形式。然",
"中的基础方面。但是，自然语言在单词上产生了Zipfian",
"后，基于生成的OOV表示，他们使用教师网络执行标",
"分布[Zipf,1949]。这意味着大量的单词（在长尾)是罕见",
"签预测",
"的。为了控制模型大小，序列标记系统通常将训练词汇表",
"从本质上讲，师生范式的方法可以看作是管道，教师",
"限制为仅覆盖训练集中的前N频繁单词。培训词汇中未",
"和学生网络是管道的两个级联组件。尽管此范例的方法",
"涵盖的单词称为OOV单词。OOV单词的学习表示受到",
"在多个任务上取得了显着成功，但它们遭受了管道范例",
"挑战，因为标准的端到端监督学习方法需要每个单词多",
"[Casellietal.,2015;Bojarskietal.，2016]的典型错误传",
"次出现才能更好地泛化。许多作品[MaandHovy,2016;",
"播问题，因为不能保证用于训练学生网络的辅助重建对",
"Madhyasthaetal.，2016]已经证明，在遇到OOV单词",
"象与用于训练教师网络的监督对象完全兼容",
"时，序列标记的性能通常会下降很多。这通常被称为",
"在这项工作中，我们提出了一种新颖的师生范式方法",
"OOV问题，我们将在这项工作中解决这个问题",
"该方法是专门为解决错误传播问题而设计的。这种方法",
"在过去的几年中，已经提出了许多方法来处理OOV",
"与现有方法的主要区别在于对学生网络的培训。我们训",
"问题。这些方法可以大致分为两类：（1）在非常大的原",
"练学生网络预测可以在监督任务中取得良好表现的单词",
"始文本[Mikolov et al.,2013;Pennington et al.,2014;",
"表示，而不是重建那些词汇内单词的表示。学生网络的",
"Petersetal.，2018)语料库上训练单词表示；（2)进一步",
"训练信号由参数固定的教师网络直接反向传播。这样，我",
"利用任务的训练数据。在数据丰富的领域中，第一类方",
"们将学生网络的学习与教师网络联系起来，从而避免了"
] |
{
"arxiv_id": "1905.12277",
"language": "zh",
"page_idx": 0
}
|
||
1905.12277-zh-1
|
1905.12277-zh-1.jpg
|
<FigureHere>
Figure 1: Schema of the proposed method when it trains the student network on $w_i.$ Boxes with a solid line denote learnable layers of the teacher network and that with a dot line denotes the trainable layer of the student network.
其中$d(\cdot,\cdot)$定义了两个嵌人之间的距离，例如。g。,欧
几里得距离函数
在这一行中，Schick 和 Schutze, [2018] 建议对 OOV 单词的上下文和表面形式(子词 n-gram)进行建模，以获取其表示形式：
(4)
$$\tilde{\boldsymbol{v}}_w(w)=\alpha\hat{\boldsymbol{v}}_w(w)+(1-\alpha)\boldsymbol{v}_c(w),$$
从使用学生网络的单词表示预测到使用教师网络的标签预测的错误传播。我们探索了该方法在词性标记(POS) 和命名实体识别(NER)中的八个序列标记任务中的适用性。对这些任务的实证研究表明，与此范式中的现有方法相比，我们提出的方法可以在 OOV 单词上实现更好或可比较的性能
这项工作的贡献可以总结如下：($i$ )我们提出了一种新颖的师生范式方法，仅使用任务的训练数据来解决序列标记中的 OOV 问题。它可以避免在同一范式中现有方法遭受的错误传播问题。($ii$ )我们对八个词性标记和命名实体识别任务进行了实验研究，并在 OOV 单词上取得了比现有的几种最先进的方法更好或比较的性能。这项工作的源代码可在 https://github.com/v-mipeng/ TaskOOV
,其中$v_c(w)$是从$w$的表面进行的单词表示建模，而
$\alpha\in[0,1]$是一个
通常，这些方法与我们提出的方法有两点不同。首先它们以一致的表示形式在不同的上下文中表示 OOV 单词$w$ ,而我们将同一单词的表示形式放宽为在不同的上下文中不同。第二，他们在辅助重建对象上训练学生网络，而我们直接在用于训练教师网络的监督对象上训练。当然，第一个差异可以很容易地通过调整方程(1)来解决：
$$\boldsymbol{v}_w(w|C(w))=\frac{1}{|C(w)|}\sum_{w'\in C(w)\cap\mathcal{V}}\boldsymbol{e}_w(w'),$$
$\begin{array}{ll}{2}&{{\mathrm{Typical~Methods~of~the~Teacher-Student}}}\\{{\mathrm{Framework}}}\end{array}$
(5)
在本节中，我们将重点介绍师生框架的一些典型方法。该框架的第一个代表性工作是由 Lazaridou 等人提出的。,[2017]。在这项工作中，他们建议通过对在其上下文$\mathcal{C}(w)$中出现的词汇内单词的表示进行求和来获得 OOV 单词$w$的表示：
其中$C(w)$表示$w$的单个上下文。根据我们的经验，这种调整通常可以提高序列标记方法的性能。因此，我们将这种调整应用于师生范式的所有比较方法。因此，我们提出的方法与现有工作之间的主要区别是学生网络的培训策略
(1)
$$\boldsymbol{v}_w(w)=\frac{1}{c(w)}\sum_{C\in\mathcal{C}(w)}\sum_{w'\in C\cap\mathcal{V}}\boldsymbol{e}_w(w').$$
## 3 Methodology
在这里，$\mathcal{C}(w)$表示包含$w$的上下文集，$\mathcal{V}$是训练词汇
表，XMATHX-5 中的单词总数
遵循这个想法，Khodak 等人。, [2018]进一步将线性
变换 A 应用于得到的嵌入：
$$\hat{\boldsymbol{v}}_w(w)=\mathrm{A}\boldsymbol{v}_w(w),$$
(2)
图 1 显示了我们提出的方法的模式。一般来说，该方法包含一个教师和一个学生网络。教师网络首先被训练来执行受监督的任务。然后，学生网络(i。e。, 表示预测层) 被训练为从单词的上下文和表面形式为单词生成适当的表示。生成的单词表示$v(w)$的质量由参数固定的教师网络(复制的序列建模和 CRF 损失层)测量，i。e。,学生网络的训练信号由教师网络反向传播。在测试时，它首先使用学生网络生成 OOV 单词的表示形式，然后使用教师网络执行标签预测
得到$w$的表示。为了确定 A 的形式，他们对这些词汇内单词进行了训练，训练对象是最小化随机选择的词汇内单词$w$的重建误差：
$$\mathcal{L}(w)=d\left(\boldsymbol{e}_w(w),\hat{\boldsymbol{v}}_w(w)\right),$$
(3)
|
[
[
0.348876953125,
0.0738525390625,
0.36767578125,
0.0738525390625,
0.36767578125,
0.08587646484375,
0.348876953125,
0.08587646484375
],
[
0.42724609375,
0.0738525390625,
0.44677734375,
0.0738525390625,
0.44677734375,
0.08587646484375,
0.42724609375,
0.08587646484375
],
[
0.5498046875,
0.07257080078125,
0.56884765625,
0.07257080078125,
0.56884765625,
0.085205078125,
0.5498046875,
0.085205078125
],
[
0.708984375,
0.0738525390625,
0.72802734375,
0.0738525390625,
0.72802734375,
0.08587646484375,
0.708984375,
0.08587646484375
],
[
0.62841796875,
0.07452392578125,
0.64453125,
0.07452392578125,
0.64453125,
0.083984375,
0.62841796875,
0.083984375
],
[
0.26953125,
0.07452392578125,
0.28515625,
0.07452392578125,
0.28515625,
0.085205078125,
0.26953125,
0.085205078125
],
[
0.3095703125,
0.07763671875,
0.3203125,
0.07763671875,
0.3203125,
0.080810546875,
0.3095703125,
0.080810546875
],
[
0.38818359375,
0.07763671875,
0.400390625,
0.07763671875,
0.400390625,
0.08209228515625,
0.38818359375,
0.08209228515625
],
[
0.67236328125,
0.07763671875,
0.6806640625,
0.07763671875,
0.6806640625,
0.08209228515625,
0.67236328125,
0.08209228515625
],
[
0.3505859375,
0.0889892578125,
0.36279296875,
0.0889892578125,
0.36279296875,
0.10101318359375,
0.3505859375,
0.10101318359375
],
[
0.43310546875,
0.0889892578125,
0.44287109375,
0.0889892578125,
0.44287109375,
0.10101318359375,
0.43310546875,
0.10101318359375
],
[
0.5556640625,
0.0889892578125,
0.56640625,
0.0889892578125,
0.56640625,
0.10101318359375,
0.5556640625,
0.10101318359375
],
[
0.63134765625,
0.0889892578125,
0.6435546875,
0.0889892578125,
0.6435546875,
0.10101318359375,
0.63134765625,
0.10101318359375
],
[
0.7158203125,
0.0889892578125,
0.7255859375,
0.0889892578125,
0.7255859375,
0.10101318359375,
0.7158203125,
0.10101318359375
],
[
0.3056640625,
0.1029052734375,
0.406982421875,
0.10479736328125,
0.406982421875,
0.11932373046875,
0.3056640625,
0.117431640625
],
[
0.275390625,
0.09027099609375,
0.281005859375,
0.09027099609375,
0.281005859375,
0.0997314453125,
0.275390625,
0.0997314453125
],
[
0.58837890625,
0.1029052734375,
0.689453125,
0.10479736328125,
0.689453125,
0.11932373046875,
0.58837890625,
0.117431640625
],
[
0.3505859375,
0.12249755859375,
0.359375,
0.12249755859375,
0.359375,
0.1337890625,
0.3505859375,
0.1337890625
],
[
0.4345703125,
0.12249755859375,
0.439453125,
0.12249755859375,
0.439453125,
0.1319580078125,
0.4345703125,
0.1319580078125
],
[
0.491943359375,
0.12371826171875,
0.5048828125,
0.12060546875,
0.5087890625,
0.1319580078125,
0.495849609375,
0.1351318359375
],
[
0.55615234375,
0.12371826171875,
0.56201171875,
0.12371826171875,
0.56201171875,
0.1319580078125,
0.55615234375,
0.1319580078125
],
[
0.48193359375,
0.1351318359375,
0.521484375,
0.136962890625,
0.51953125,
0.15283203125,
0.48046875,
0.15087890625
],
[
0.5400390625,
0.1351318359375,
0.58251953125,
0.136962890625,
0.5810546875,
0.1514892578125,
0.5390625,
0.149658203125
],
[
0.69677734375,
0.1337890625,
0.73876953125,
0.1383056640625,
0.736328125,
0.154052734375,
0.69384765625,
0.149658203125
],
[
0.418212890625,
0.1363525390625,
0.455810546875,
0.1383056640625,
0.455078125,
0.1514892578125,
0.416748046875,
0.149658203125
],
[
0.25732421875,
0.1363525390625,
0.298095703125,
0.1363525390625,
0.298095703125,
0.15087890625,
0.25732421875,
0.15087890625
],
[
0.338134765625,
0.136962890625,
0.373291015625,
0.136962890625,
0.373291015625,
0.149658203125,
0.338134765625,
0.149658203125
],
[
0.62109375,
0.136962890625,
0.65869140625,
0.136962890625,
0.65869140625,
0.15087890625,
0.62109375,
0.15087890625
],
[
0.57861328125,
0.1383056640625,
0.6044921875,
0.1383056640625,
0.6044921875,
0.146484375,
0.57861328125,
0.146484375
],
[
0.301513671875,
0.1383056640625,
0.32177734375,
0.1383056640625,
0.32177734375,
0.1458740234375,
0.301513671875,
0.1458740234375
],
[
0.386474609375,
0.14013671875,
0.398681640625,
0.14013671875,
0.398681640625,
0.146484375,
0.386474609375,
0.146484375
],
[
0.6689453125,
0.14013671875,
0.68115234375,
0.14013671875,
0.68115234375,
0.146484375,
0.6689453125,
0.146484375
],
[
0.771484375,
0.143310546875,
0.7958984375,
0.143310546875,
0.79345703125,
0.212158203125,
0.7685546875,
0.212158203125
],
[
0.27294921875,
0.15087890625,
0.283447265625,
0.15087890625,
0.283447265625,
0.1634521484375,
0.27294921875,
0.1634521484375
],
[
0.2059326171875,
0.14453125,
0.2254638671875,
0.14453125,
0.22216796875,
0.210205078125,
0.2034912109375,
0.210205078125
],
[
0.43310546875,
0.15087890625,
0.439453125,
0.15087890625,
0.439453125,
0.154052734375,
0.43310546875,
0.154052734375
],
[
0.6298828125,
0.1514892578125,
0.640625,
0.146484375,
0.646484375,
0.15283203125,
0.63623046875,
0.1578369140625
],
[
0.5537109375,
0.154052734375,
0.560546875,
0.1483154296875,
0.56640625,
0.15283203125,
0.5595703125,
0.1578369140625
],
[
0.34716796875,
0.15283203125,
0.35546875,
0.146484375,
0.36279296875,
0.1514892578125,
0.3544921875,
0.1578369140625
],
[
0.495849609375,
0.1546630859375,
0.50634765625,
0.1546630859375,
0.50634765625,
0.1666259765625,
0.495849609375,
0.1666259765625
],
[
0.273681640625,
0.17041015625,
0.43701171875,
0.17041015625,
0.43701171875,
0.1849365234375,
0.273681640625,
0.1849365234375
],
[
0.5556640625,
0.169189453125,
0.71826171875,
0.17041015625,
0.71826171875,
0.1849365234375,
0.5556640625,
0.184326171875
],
[
0.55810546875,
0.1881103515625,
0.56201171875,
0.1881103515625,
0.56201171875,
0.1944580078125,
0.55810546875,
0.1944580078125
],
[
0.25732421875,
0.2020263671875,
0.295654296875,
0.203857421875,
0.294189453125,
0.216552734375,
0.256591796875,
0.2139892578125
],
[
0.336669921875,
0.203857421875,
0.374267578125,
0.203857421875,
0.374267578125,
0.21533203125,
0.336669921875,
0.21533203125
],
[
0.416748046875,
0.203857421875,
0.453369140625,
0.203857421875,
0.453369140625,
0.21533203125,
0.416748046875,
0.21533203125
],
[
0.54150390625,
0.20263671875,
0.5849609375,
0.20263671875,
0.5849609375,
0.2139892578125,
0.54150390625,
0.2139892578125
],
[
0.62109375,
0.203857421875,
0.65673828125,
0.203857421875,
0.65673828125,
0.21533203125,
0.62109375,
0.21533203125
],
[
0.69921875,
0.203857421875,
0.734375,
0.203857421875,
0.734375,
0.21533203125,
0.69921875,
0.21533203125
],
[
0.384033203125,
0.2052001953125,
0.40185546875,
0.2052001953125,
0.40185546875,
0.212158203125,
0.384033203125,
0.212158203125
],
[
0.3095703125,
0.205810546875,
0.32177734375,
0.205810546875,
0.32177734375,
0.212158203125,
0.3095703125,
0.212158203125
],
[
0.3798828125,
0.205810546875,
0.3857421875,
0.210205078125,
0.38232421875,
0.212158203125,
0.37744140625,
0.2083740234375
],
[
0.67236328125,
0.20703125,
0.6787109375,
0.20703125,
0.6787109375,
0.210205078125,
0.67236328125,
0.210205078125
],
[
0.43310546875,
0.216552734375,
0.439453125,
0.216552734375,
0.439453125,
0.2203369140625,
0.43310546875,
0.2203369140625
],
[
0.6337890625,
0.216552734375,
0.64208984375,
0.216552734375,
0.64208984375,
0.2203369140625,
0.6337890625,
0.2203369140625
],
[
0.3505859375,
0.2171630859375,
0.36279296875,
0.2171630859375,
0.36279296875,
0.2310791015625,
0.3505859375,
0.2310791015625
],
[
0.27294921875,
0.2171630859375,
0.283447265625,
0.2171630859375,
0.283447265625,
0.2298583984375,
0.27294921875,
0.2298583984375
],
[
0.5556640625,
0.2171630859375,
0.56640625,
0.2171630859375,
0.56640625,
0.2291259765625,
0.5556640625,
0.2291259765625
],
[
0.7158203125,
0.2171630859375,
0.72216796875,
0.2171630859375,
0.72216796875,
0.2203369140625,
0.7158203125,
0.2203369140625
],
[
0.483642578125,
0.2266845703125,
0.5185546875,
0.2266845703125,
0.5185546875,
0.238037109375,
0.483642578125,
0.238037109375
],
[
0.26953125,
0.23291015625,
0.443603515625,
0.2342529296875,
0.443603515625,
0.248779296875,
0.26953125,
0.2474365234375
],
[
0.53515625,
0.23291015625,
0.73388671875,
0.23486328125,
0.73388671875,
0.25,
0.53515625,
0.2474365234375
],
[
0.55615234375,
0.248779296875,
0.564453125,
0.25,
0.56201171875,
0.264404296875,
0.55224609375,
0.261962890625
],
[
0.6337890625,
0.25,
0.640625,
0.25,
0.640625,
0.2568359375,
0.6337890625,
0.2568359375
],
[
0.431396484375,
0.250732421875,
0.441162109375,
0.250732421875,
0.441162109375,
0.264404296875,
0.431396484375,
0.264404296875
],
[
0.3505859375,
0.250732421875,
0.359375,
0.250732421875,
0.359375,
0.264404296875,
0.3505859375,
0.264404296875
],
[
0.268798828125,
0.264404296875,
0.289306640625,
0.268310546875,
0.28515625,
0.27978515625,
0.2646484375,
0.27587890625
],
[
0.346435546875,
0.266357421875,
0.366943359375,
0.266357421875,
0.366943359375,
0.279052734375,
0.346435546875,
0.279052734375
],
[
0.42724609375,
0.264404296875,
0.449462890625,
0.268310546875,
0.4453125,
0.281005859375,
0.423095703125,
0.276611328125
],
[
0.63134765625,
0.266357421875,
0.646484375,
0.267578125,
0.64453125,
0.27783203125,
0.6298828125,
0.276611328125
],
[
0.55126953125,
0.268310546875,
0.56787109375,
0.268310546875,
0.56787109375,
0.27783203125,
0.55126953125,
0.27783203125
],
[
0.306396484375,
0.268310546875,
0.32275390625,
0.268310546875,
0.32275390625,
0.276611328125,
0.306396484375,
0.276611328125
],
[
0.3896484375,
0.26953125,
0.398681640625,
0.26953125,
0.398681640625,
0.274658203125,
0.3896484375,
0.274658203125
],
[
0.71142578125,
0.268310546875,
0.72802734375,
0.268310546875,
0.72802734375,
0.27783203125,
0.71142578125,
0.27783203125
],
[
0.625,
0.279052734375,
0.65625,
0.279052734375,
0.65625,
0.287353515625,
0.625,
0.287353515625
],
[
0.30810546875,
0.291748046875,
0.404296875,
0.291748046875,
0.404296875,
0.302978515625,
0.30810546875,
0.302978515625
],
[
0.5888671875,
0.28857421875,
0.689453125,
0.290283203125,
0.689453125,
0.304931640625,
0.5888671875,
0.302978515625
],
[
0.08660888671875,
0.3125,
0.91259765625,
0.3125,
0.91259765625,
0.32568359375,
0.08660888671875,
0.32568359375
],
[
0.0882568359375,
0.326904296875,
0.904296875,
0.326904296875,
0.904296875,
0.337158203125,
0.0882568359375,
0.337158203125
],
[
0.0882568359375,
0.365478515625,
0.484375,
0.365478515625,
0.484375,
0.37890625,
0.0882568359375,
0.37890625
],
[
0.52392578125,
0.3642578125,
0.91259765625,
0.365478515625,
0.91259765625,
0.380126953125,
0.52392578125,
0.37890625
],
[
0.0882568359375,
0.37744140625,
0.484375,
0.37744140625,
0.484375,
0.391357421875,
0.0882568359375,
0.391357421875
],
[
0.51318359375,
0.37890625,
0.62841796875,
0.37890625,
0.62841796875,
0.392578125,
0.51318359375,
0.392578125
],
[
0.0882568359375,
0.391357421875,
0.484375,
0.391357421875,
0.484375,
0.40478515625,
0.0882568359375,
0.40478515625
],
[
0.529296875,
0.393310546875,
0.91015625,
0.393310546875,
0.91015625,
0.4072265625,
0.529296875,
0.4072265625
],
[
0.0882568359375,
0.4072265625,
0.484375,
0.4072265625,
0.484375,
0.42041015625,
0.0882568359375,
0.42041015625
],
[
0.51318359375,
0.4072265625,
0.91259765625,
0.40771484375,
0.91259765625,
0.42236328125,
0.51318359375,
0.421630859375
],
[
0.0882568359375,
0.42041015625,
0.484375,
0.42041015625,
0.484375,
0.434326171875,
0.0882568359375,
0.434326171875
],
[
0.51318359375,
0.42041015625,
0.63623046875,
0.421630859375,
0.63623046875,
0.436279296875,
0.51318359375,
0.43505859375
],
[
0.0849609375,
0.434326171875,
0.21728515625,
0.434326171875,
0.21728515625,
0.447509765625,
0.0849609375,
0.447509765625
],
[
0.59326171875,
0.447509765625,
0.833984375,
0.447509765625,
0.833984375,
0.4609375,
0.59326171875,
0.4609375
],
[
0.8896484375,
0.447509765625,
0.91259765625,
0.447509765625,
0.91259765625,
0.4609375,
0.8896484375,
0.4609375
],
[
0.0882568359375,
0.462158203125,
0.484375,
0.462158203125,
0.484375,
0.47607421875,
0.0882568359375,
0.47607421875
],
[
0.10296630859375,
0.448974609375,
0.484375,
0.448974609375,
0.484375,
0.462158203125,
0.10296630859375,
0.462158203125
],
[
0.0882568359375,
0.47607421875,
0.484375,
0.47607421875,
0.484375,
0.4892578125,
0.0882568359375,
0.4892578125
],
[
0.515625,
0.473388671875,
0.91259765625,
0.473388671875,
0.91259765625,
0.4873046875,
0.515625,
0.4873046875
],
[
0.08905029296875,
0.4892578125,
0.486083984375,
0.4892578125,
0.486083984375,
0.50244140625,
0.08905029296875,
0.50244140625
],
[
0.5087890625,
0.486083984375,
0.6337890625,
0.48486328125,
0.6337890625,
0.5,
0.5087890625,
0.50048828125
],
[
0.08660888671875,
0.50244140625,
0.486083984375,
0.501953125,
0.486083984375,
0.5166015625,
0.08660888671875,
0.517578125
],
[
0.52783203125,
0.50048828125,
0.91015625,
0.501953125,
0.91015625,
0.5166015625,
0.52783203125,
0.51513671875
],
[
0.09234619140625,
0.517578125,
0.484375,
0.517578125,
0.484375,
0.53076171875,
0.09234619140625,
0.53076171875
],
[
0.51318359375,
0.5166015625,
0.91015625,
0.5166015625,
0.91015625,
0.52978515625,
0.51318359375,
0.52978515625
],
[
0.0849609375,
0.52978515625,
0.483642578125,
0.53076171875,
0.483642578125,
0.54541015625,
0.0849609375,
0.54443359375
],
[
0.5146484375,
0.52978515625,
0.9111328125,
0.52978515625,
0.9111328125,
0.54345703125,
0.5146484375,
0.54345703125
],
[
0.0882568359375,
0.54443359375,
0.16015625,
0.54443359375,
0.16015625,
0.55615234375,
0.0882568359375,
0.55615234375
],
[
0.515625,
0.54345703125,
0.9111328125,
0.54345703125,
0.9111328125,
0.556640625,
0.515625,
0.556640625
],
[
0.5146484375,
0.55810546875,
0.91015625,
0.55810546875,
0.91015625,
0.5712890625,
0.5146484375,
0.5712890625
],
[
0.5146484375,
0.5693359375,
0.9111328125,
0.5712890625,
0.9111328125,
0.5859375,
0.5146484375,
0.583984375
],
[
0.08660888671875,
0.57568359375,
0.1217041015625,
0.57568359375,
0.1217041015625,
0.58837890625,
0.08660888671875,
0.58837890625
],
[
0.12091064453125,
0.57568359375,
0.463134765625,
0.57568359375,
0.463134765625,
0.5888671875,
0.12091064453125,
0.5888671875
],
[
0.5146484375,
0.5859375,
0.5400390625,
0.5859375,
0.5400390625,
0.5986328125,
0.5146484375,
0.5986328125
],
[
0.11517333984375,
0.59228515625,
0.2099609375,
0.59228515625,
0.2099609375,
0.60693359375,
0.11517333984375,
0.60693359375
],
[
0.6953125,
0.6015625,
0.72607421875,
0.6015625,
0.72607421875,
0.6142578125,
0.6953125,
0.6142578125
],
[
0.744140625,
0.60400390625,
0.865234375,
0.60400390625,
0.865234375,
0.6298828125,
0.744140625,
0.6298828125
],
[
0.67626953125,
0.611328125,
0.80615234375,
0.62548828125,
0.80224609375,
0.646484375,
0.67236328125,
0.6318359375
],
[
0.8837890625,
0.611328125,
0.9111328125,
0.611328125,
0.9111328125,
0.62353515625,
0.8837890625,
0.62353515625
],
[
0.56640625,
0.60693359375,
0.6845703125,
0.611328125,
0.68310546875,
0.62890625,
0.564453125,
0.625
],
[
0.0882568359375,
0.6123046875,
0.483642578125,
0.6123046875,
0.483642578125,
0.626953125,
0.0882568359375,
0.626953125
],
[
0.08660888671875,
0.62548828125,
0.486083984375,
0.625,
0.486083984375,
0.6396484375,
0.08660888671875,
0.64013671875
],
[
0.0906982421875,
0.6396484375,
0.483642578125,
0.6396484375,
0.483642578125,
0.65283203125,
0.0906982421875,
0.65283203125
],
[
0.0849609375,
0.65283203125,
0.484375,
0.65380859375,
0.484375,
0.66845703125,
0.0849609375,
0.66748046875
],
[
0.5146484375,
0.6494140625,
0.9111328125,
0.6494140625,
0.9111328125,
0.66357421875,
0.5146484375,
0.66357421875
],
[
0.51318359375,
0.6611328125,
0.9111328125,
0.662109375,
0.9111328125,
0.6767578125,
0.51318359375,
0.67529296875
],
[
0.0882568359375,
0.66845703125,
0.1993408203125,
0.66845703125,
0.1993408203125,
0.681640625,
0.0882568359375,
0.681640625
],
[
0.5146484375,
0.67529296875,
0.9111328125,
0.6767578125,
0.9111328125,
0.69140625,
0.5146484375,
0.68994140625
],
[
0.233642578125,
0.69140625,
0.2467041015625,
0.69140625,
0.2467041015625,
0.70068359375,
0.233642578125,
0.70068359375
],
[
0.268798828125,
0.6943359375,
0.324462890625,
0.6943359375,
0.324462890625,
0.71044921875,
0.268798828125,
0.71044921875
],
[
0.31787109375,
0.6962890625,
0.3505859375,
0.6962890625,
0.3505859375,
0.708984375,
0.31787109375,
0.708984375
],
[
0.515625,
0.69140625,
0.9111328125,
0.69140625,
0.9111328125,
0.705078125,
0.515625,
0.705078125
],
[
0.3447265625,
0.69775390625,
0.418212890625,
0.69775390625,
0.418212890625,
0.7119140625,
0.3447265625,
0.7119140625
],
[
0.463134765625,
0.69873046875,
0.486083984375,
0.69873046875,
0.486083984375,
0.7119140625,
0.463134765625,
0.7119140625
],
[
0.51318359375,
0.705078125,
0.5810546875,
0.705078125,
0.5810546875,
0.7197265625,
0.51318359375,
0.7197265625
],
[
0.2181396484375,
0.70556640625,
0.268798828125,
0.70556640625,
0.268798828125,
0.72021484375,
0.2181396484375,
0.72021484375
],
[
0.156005859375,
0.70068359375,
0.2197265625,
0.70068359375,
0.2197265625,
0.7109375,
0.156005859375,
0.7109375
],
[
0.273681640625,
0.70556640625,
0.294189453125,
0.70556640625,
0.294189453125,
0.71337890625,
0.273681640625,
0.71337890625
],
[
0.325927734375,
0.70703125,
0.34228515625,
0.70703125,
0.34228515625,
0.7109375,
0.325927734375,
0.7109375
],
[
0.253173828125,
0.71337890625,
0.36279296875,
0.7119140625,
0.36279296875,
0.7265625,
0.253173828125,
0.72802734375
],
[
0.08905029296875,
0.7373046875,
0.488525390625,
0.7373046875,
0.488525390625,
0.75048828125,
0.08905029296875,
0.75048828125
],
[
0.5087890625,
0.736328125,
0.65673828125,
0.7392578125,
0.65673828125,
0.7568359375,
0.5087890625,
0.75390625
],
[
0.0882568359375,
0.75048828125,
0.313720703125,
0.75048828125,
0.313720703125,
0.7646484375,
0.0882568359375,
0.7646484375
],
[
0.10455322265625,
0.7646484375,
0.486083984375,
0.7646484375,
0.486083984375,
0.77783203125,
0.10455322265625,
0.77783203125
],
[
0.515625,
0.76513671875,
0.9111328125,
0.76513671875,
0.9111328125,
0.77880859375,
0.515625,
0.77880859375
],
[
0.0882568359375,
0.77783203125,
0.281982421875,
0.77783203125,
0.281982421875,
0.79150390625,
0.0882568359375,
0.79150390625
],
[
0.5146484375,
0.77783203125,
0.91015625,
0.77783203125,
0.91015625,
0.79150390625,
0.5146484375,
0.79150390625
],
[
0.5146484375,
0.79248046875,
0.9111328125,
0.79248046875,
0.9111328125,
0.80615234375,
0.5146484375,
0.80615234375
],
[
0.221435546875,
0.8017578125,
0.348876953125,
0.8017578125,
0.348876953125,
0.81494140625,
0.221435546875,
0.81494140625
],
[
0.463134765625,
0.80322265625,
0.48779296875,
0.80322265625,
0.48779296875,
0.814453125,
0.463134765625,
0.814453125
],
[
0.51318359375,
0.80615234375,
0.9111328125,
0.80615234375,
0.9111328125,
0.8193359375,
0.51318359375,
0.8193359375
],
[
0.0882568359375,
0.82470703125,
0.486083984375,
0.82470703125,
0.486083984375,
0.83837890625,
0.0882568359375,
0.83837890625
],
[
0.5146484375,
0.8193359375,
0.9111328125,
0.8193359375,
0.9111328125,
0.83349609375,
0.5146484375,
0.83349609375
],
[
0.08660888671875,
0.8369140625,
0.486083984375,
0.83837890625,
0.486083984375,
0.85302734375,
0.08660888671875,
0.8515625
],
[
0.5146484375,
0.833984375,
0.9111328125,
0.833984375,
0.9111328125,
0.84765625,
0.5146484375,
0.84765625
],
[
0.51318359375,
0.8466796875,
0.9111328125,
0.84521484375,
0.9111328125,
0.85986328125,
0.51318359375,
0.861328125
],
[
0.0882568359375,
0.85302734375,
0.2484130859375,
0.85302734375,
0.2484130859375,
0.8662109375,
0.0882568359375,
0.8662109375
],
[
0.515625,
0.861328125,
0.9111328125,
0.861328125,
0.9111328125,
0.87548828125,
0.515625,
0.87548828125
],
[
0.193603515625,
0.87548828125,
0.37744140625,
0.87548828125,
0.37744140625,
0.8896484375,
0.193603515625,
0.8896484375
],
[
0.464111328125,
0.87548828125,
0.486083984375,
0.87548828125,
0.486083984375,
0.8896484375,
0.464111328125,
0.8896484375
],
[
0.5146484375,
0.87548828125,
0.69384765625,
0.87548828125,
0.69384765625,
0.8896484375,
0.5146484375,
0.8896484375
]
] |
[
"yi",
"yn",
"yi",
"Yn",
"yi",
"yi",
"..",
"...",
".",
"1",
"1",
"1",
"1",
"I",
"CRFLoss Layer",
"",
"CRF Loss Layer",
"",
"",
"",
"",
"Copy",
"h(wi)",
"h(wn)",
"h(wn)",
"h(wi)",
"h(w)",
"h(w)",
".",
"",
"..",
"+*",
"Backpropagate",
"",
"Backpropagate",
"",
"1",
"X",
"",
"+",
" Sequence Modeling Layer",
"Sequence Modeling Layer",
"",
"e(w))",
"e(w)",
"e(wn)",
"e(w)",
"v(wi)",
"e(wn)",
"..",
"...",
"",
"",
"4",
"4",
"t",
"",
"t",
"",
"Refer",
"Word Representation Layer",
"RepresentationPredictLayer",
"",
"",
"",
"↑",
"W1",
"Wi",
"Wn",
"Wi",
"Wi",
"..",
"++:",
"Wn",
"(OOV)",
"Teacher Network",
" Student Network",
"Figure l: Schema of the proposed method when it trains the student network on wi : Boxes with a solid line denote",
"learnablelavers of theteachernetwork andthat with a dot linedenotesthetrainable laver ofthestudent network",
"从使用学生网络的单词表示预测到使用教师网络的标签",
"其中d(，）定义了两个嵌人之间的距离，例如。g。，欧",
"预测的错误传播。我们探索了该方法在词性标记（POS）",
"几里得距离函数",
"和命名实体识别（NER）中的八个序列标记任务中的适",
"在这一行中，Schick和Schutze，[2018]建议对OOV",
"用性。对这些任务的实证研究表明，与此范式中的现有",
"单词的上下文和表面形式（子词n-gram）进行建模，以",
"方法相比，我们提出的方法可以在OOV单词上实现更",
"获取其表示形式：",
"好或可比较的性能",
"vw(w) = aow(w) + (1-α)v.(w),",
"(4)",
"新颖的师生范式方法，仅使用任务的训练数据来解决序",
"这项工作的贡献可以总结如下：（i）我们提出了一种",
"列标记中的OOV问题。它可以避免在同一范式中现有",
"，其中u(w）是从w的表面进行的单词表示建模，而",
"方法遭受的错误传播问题。（i）我们对八个词性标记",
"QE[0,1]是一个",
"和命名实体识别任务进行了实验研究，并在OOV单词",
"通常，这些方法与我们提出的方法有两点不同。首先",
"上取得了比现有的几种最先进的方法更好或比较的性能。",
"它们以一致的表示形式在不同的上下文中表示OOV单",
"这项工作的源代码可在https://github.com/v-mipeng/",
"词w，而我们将同一单词的表示形式放宽为在不同的上",
"Taskoov",
"下文中不同。第二，他们在辅助重建对象上训练学生网",
"络，而我们直接在用于训练教师网络的监督对象上训练",
"当然，第一个差异可以很容易地通过调整方程（1）来解",
"2",
"Typical Methods of the Teacher-Student",
"决：",
"Framework",
"1",
"ew(w),",
"IC(w)/w\"EC(w)nv",
"(5)",
"vw(w|C(w) = ",
"在本节中，我们将重点介绍师生框架的一些典型方法",
"该框架的第一个代表性工作是由Lazaridou等人提出的。",
"[2017]。在这项工作中，他们建议通过对在其上下文",
"c(w）中出现的词汇内单词的表示进行求和来获得OOV",
"其中C(w）表示w的单个上下文。根据我们的经验，这",
"种调整通常可以提高序列标记方法的性能。因此，我们",
"单词w的表示：",
"将这种调整应用于师生范式的所有比较方法。因此，我",
"1",
"门",
"门",
"们提出的方法与现有工作之间的主要区别是学生网络的",
"ew(w).",
"(1)",
"培训策略",
"c(w)",
"vw(w)=",
"",
"",
"CEC(w)W'ECnV",
"在这里，C(w）表示包含W的上下文集，V是训练词汇",
"3Methodology",
"表，XMATHX-5中的单词总数",
"遵循这个想法，Khodak等人。，[2018]进一步将线性",
"图1显示了我们提出的方法的模式。一般来说，该方法",
"变换A应用于得到的嵌入：",
"包含一个教师和一个学生网络。教师网络首先被训练来",
"执行受监督的任务。然后，学生网络(i。e。，表示预测层)",
"ow(w) = Avw(w)",
"(2)",
"被训练为从单词的上下文和表面形式为单词生成适当的",
"得到W的表示。为了确定A的形式，他们对这些词汇",
"表示。生成的单词表示(w）的质量由参数固定的教师",
"内单词进行了训练，训练对象是最小化随机选择的词汇",
"网络（复制的序列建模和CRF损失层）测量，i。e。，学",
"生网络的训练信号由教师网络反向传播。在测试时，它",
"内单词w的重建误差：",
"首先使用学生网络生成OOV单词的表示形式，然后使",
"C(w)=d(ew(w),ow(w))",
"(3)",
"用教师网络执行标签预测"
] |
{
"arxiv_id": "1905.12277",
"language": "zh",
"page_idx": 1
}
|
||
1905.12277-zh-2
|
1905.12277-zh-2.jpg
|
## 3.1 Notation
在整个工作中，我们将$\Omega$表示为单词空间，将$\mathcal{V}$表示为训练词汇。由单词序列组成的句子表示为 X = $\{ w_{1}, \cdots , w_{n}\}$ $\in$ $\Omega$ , 其相应的标签序列表示为 Y= $\{y_1,\cdots,y_n\}\in\mathcal{Y}$。给定其相应句子 X 的单词$w_i$的$\operatorname { \text{上 下 文 表 示 为 X}} _{\setminus i}= \{ w_{1}, \cdots , w_{i- 1}, w_{i+ 1}, \cdots , w_{n}\}$ ,而$\mathcal{X}_{<i,j>}$是指子序列$\{w_i,\cdots,w_j\}$ 。教师网络表示为$T$ 学生网络表示为$S$。此外，令$f$是在$\Omega$上定义的函数， 我们将 $f($X)表示为$\{f(w_1),\cdots,f(w_n)\}$ 的简写
3.2 Train the Teacher Network
教师网络用于在序列标记系统中进行标签预测。它可以是适合任务的任意架构。因为这项工作的主要重点是处理 OOV 问题，而不是设计一个优越的监督模型，所以在这项工作中，我们尝试了典型的基于 CNN 和基于LSTM 的体系结构[Huang et al.,2015] 来实现教师网络。简而言之，此体系结构表示一个单词$w$ ,其中一个唯一的密集向量(嵌人)$e_w(w)$和一个向量$e_c(w)$使用字符级卷积神经网络(CNN)[Kim,2014]从其字符序列建模：
$$\boldsymbol{e}(w)=[\boldsymbol{e}_w(w)\oplus\boldsymbol{e}_c(w)],$$
(6)
, 其中$\oplus$表示串联操作。将相应的句子表示 $e(\mathbb{X})$ 作为输入输人到双向长短期记忆网络(BiLSTM)[Hochreiter and Schmidhuber, 1997] 或内核大小设置为 3 的三层CNN 网络[Kim,2014]中，以对每个单词的上下文信息进行建模，从而获得给定 X 的每个单词的隐藏表示(不是单词嵌人)h$(w_i|$X)。在基于 BiLSTM 或基于 CNN 的序列建模层之上，它使用顺序条件随机场(CRF)[Lafferty] et al., 2001] 对整个句子的标签进行联合解码：
$$p(\mathrm Y|\mathrm X;\boldsymbol\theta_T)=\frac{\prod_{t=1}^m\phi_t(y_{t-1},y_t|\mathrm X)}{\sum_{Y'\in\mathcal Y}\prod_{t=1}^m\phi_t(y_{t-1}',y_t'|\mathrm X)}$$
(7)
其中$\phi_t(y^{\prime},y|\mathcal{X})=\exp(\boldsymbol{w}_{y^{\prime},y}^{T}$h$( w_t) + b_{y^{\prime }, y})$ $, \boldsymbol{w}_{y^{\prime }, y}$和$b_{y^{\prime},y}$ 是可训练的参数对应于整个标签对$(y^\prime,y)$ ,。然后，教师网络的培训损失定义为：
$$\mathcal{L}_T=\sum_{i=1}^N\log p(\mathrm{Y}_i|\mathrm{X}_i;\boldsymbol{\theta}_T),$$
(8)
,其中$N$是培训句子编号。该网络在训练集中出现的所有单词上进行训练。之后，我们在以下描述的学生网络的培训期间修复其参数
3.3 Train the Student Network
学生网络对单词的表面形式和上下文信息进行建模，以生成其表示形式。Schick and Schütze 、2018 证明了这两种信息资源是互补的
Model Surface-form. 单词$w$的表面形式表示是从其字符序列$w=\{c_1,c_2,\cdots,c_n\}\in\mathcal{V}_c$获得的，其中$\mathcal{V}_c$是字符词汇。具体来说，在[Schick and Schütze, 2018] 的工作之后，我们首先用特殊的开始和结束字符$c_0= \langle s\rangle$ , $c_{n+1}=\langle e\rangle$填充字符序列，并获得其向上的$k$ -gram 集：
$$G(w)=\cup_{m=1}^k\cup_{i=0}^{n+2-m}\{c_i,\cdots,c_{i+m-1}\}.$$
(9)
然后，我们将$w$的表面形式嵌人定义为其所有向上
XMATHX\_LINE 嵌入的平均值
Model Context. 为了对单词$w_i$的上下文进行建模我们在其上下文单词序列 X$_i$上应用了双向长短期记忆网络 (BiLSTM)。在这里，我们考虑在两种不同情况下应用这种做法。在第一种情况下，句子 X 中只有一个OOV 单词。而在第二种情况下，X 中有多个 OOV 单词
当一个句子 X 中只有一个 OOV 单词$w_i$时，其余单词的任务特定表示都是已知的，我们可以直接将 BiLSTM 应用于$\mathcal{X}_{\setminus i}$来预测$w_i$的表示。具体来说，我们使用正向 LSTM 来对从 X 的开头到$w_i-1$的序列进行建模：
$$\overrightarrow{\mathrm{h}}(w_i|\mathrm{X})=\overrightarrow{\mathrm{LSTM}}(e(\mathrm{X}_{<1,i-1>}))$$
(10)
,并使用反向 LSTM 以相反的顺序对从 X 的末尾到
$w_{i+1}$的序列进行建模：
$$\stackrel{\leftarrow}{\mathrm{h}}(w_{i}|\mathrm{X})=\stackrel{\leftarrow}{\mathrm{LSTM}}(e(\mathrm{X}_{<i+1,n>})).$$
(11)
当一个句子中有多个 OOV 单词时，对于给定的 OOV 单词$w_i\in\mathcal{X}$, X$_\setminus i$中其他一些单词的表示形式也是未知的。因此，我们不能直接将 BiLSTM 应用于$X_{\setminus i}$以获得其预测表示。为了解决这个问题，我们建议迭代预测句子中多个 OOV 单词的表示形式。令$v_\mathrm{context}^t( w_i| \mathcal{X} )$表示在$t^{th}$与$v_\mathrm{context}^0(w_i|$X)=0迭代时给定 X 的$w_i$的预测表示。在$(t+1)^{th}$迭代中，为了预测$w_i$的表示，我们将 BiLSTM 应用于其$t^th$迭代上下文$v_\mathrm{context}^t(\mathcal{X}_{\setminus i})$,获得$v_\mathrm{context}^{t+ 1}( w_i| \mathcal{X} )$ 。对所有 OOV 单词重复此过程以完成$(t+1)^{th}$迭代。迭代继续进行，直到$t$达到固定值$K$ , 从而获得每个 OOV 单词的最终预测表示。根据我们的经验，设置$K=2$是合适的
Combine Surface-form and Context. 我们最终将表
面形式和上下文的表示结合起来，以获得$w$的联合表示
$\boldsymbol{v}(w|\mathcal{X})$ 。在这项工作中，我们遵循[Schick and Schütze,
2018]的思想，并将这两个表示与一个门结合在一起：
(12)
$$\boldsymbol{v}_w(w|\mathrm{X})=\alpha\boldsymbol{v}_\mathrm{form}(w)+(1-\alpha)\boldsymbol{v}_\mathrm{context}(w|\mathrm{X})$$
其中
(13)
$$\alpha=\sigma(\boldsymbol{w}^T\left[\boldsymbol{v}_\mathrm{form}(w)\oplus\boldsymbol{v}_\mathrm{context}(w|\mathrm{X})\right]+b).$$
这里，$\sigma$表示 sigmoid 函数。对于表达式一致性，我们
$\widetilde{\text{表示}}$
$$v_w(w|\mathrm{X};\boldsymbol{\theta}_S)=\boldsymbol{v}_w(w|\mathrm{X}),$$
(14)
其中$\theta_S$是学生网络的那些可训练参数，实际上对应
于图 1 中描述的表示预测层
Training.
因为我们使用教师网络来执行标签预测
所以学生网络的培训对象应该学习为适合教师网络的OOV 单词生成表示。根据以前的作品 [Lazaridou et al., 2017; Khodak et al., 2018; Schick and Schütze, 2018] 的思想，我们可以训练学生网络以最小化具有由以下定义的损失函数的那些词汇内单词的重建错误：
$$\mathcal{L}_{recon}=\frac{1}{N}\sum_{i=1}^{N}\sum_{w\in\mathrm{X}_{i}\cap\mathcal{V}}d\left(\boldsymbol{v}_{w}(w|\mathrm{X}_{i};\boldsymbol{\theta}_{S}),\boldsymbol{e}_{w}(w)\right).$$
这里，$d(\cdot,\cdot)$定义了两个嵌入之间的距离，例如。g。,欧几里得距离函数。这种做法的最大问题是它可能会遇到错误传播问题，因为辅助重建训练标准$\mathcal{L}_{recon}$不能保证与教师网络的训练对象兼容
|
[
[
0.0849609375,
0.0694580078125,
0.205078125,
0.0694580078125,
0.205078125,
0.08331298828125,
0.0849609375,
0.08331298828125
],
[
0.51953125,
0.07135009765625,
0.64208984375,
0.07135009765625,
0.64208984375,
0.08209228515625,
0.51953125,
0.08209228515625
],
[
0.65673828125,
0.07135009765625,
0.90673828125,
0.07135009765625,
0.90673828125,
0.08209228515625,
0.65673828125,
0.08209228515625
],
[
0.0882568359375,
0.08709716796875,
0.486083984375,
0.08709716796875,
0.486083984375,
0.10101318359375,
0.0882568359375,
0.10101318359375
],
[
0.51318359375,
0.083984375,
0.9111328125,
0.083984375,
0.9111328125,
0.09783935546875,
0.51318359375,
0.09783935546875
],
[
0.0882568359375,
0.10162353515625,
0.48193359375,
0.10162353515625,
0.48193359375,
0.11553955078125,
0.0882568359375,
0.11553955078125
],
[
0.515625,
0.0997314453125,
0.91015625,
0.0997314453125,
0.91015625,
0.11297607421875,
0.515625,
0.11297607421875
],
[
0.08660888671875,
0.11553955078125,
0.48193359375,
0.11297607421875,
0.48193359375,
0.1275634765625,
0.08660888671875,
0.1300048828125
],
[
0.5146484375,
0.11297607421875,
0.91015625,
0.11297607421875,
0.91015625,
0.126953125,
0.5146484375,
0.126953125
],
[
0.08905029296875,
0.1287841796875,
0.484375,
0.1287841796875,
0.484375,
0.143310546875,
0.08905029296875,
0.143310546875
],
[
0.51318359375,
0.126953125,
0.91259765625,
0.126953125,
0.91259765625,
0.14013671875,
0.51318359375,
0.14013671875
],
[
0.08905029296875,
0.1427001953125,
0.486083984375,
0.143310546875,
0.486083984375,
0.1578369140625,
0.08905029296875,
0.1572265625
],
[
0.53125,
0.141357421875,
0.9111328125,
0.141357421875,
0.9111328125,
0.1546630859375,
0.53125,
0.1546630859375
],
[
0.0882568359375,
0.1590576171875,
0.4755859375,
0.1590576171875,
0.4755859375,
0.173583984375,
0.0882568359375,
0.173583984375
],
[
0.515625,
0.1546630859375,
0.91259765625,
0.1546630859375,
0.91259765625,
0.1685791015625,
0.515625,
0.1685791015625
],
[
0.08660888671875,
0.17041015625,
0.486083984375,
0.1717529296875,
0.486083984375,
0.186279296875,
0.08660888671875,
0.1849365234375
],
[
0.51318359375,
0.1673583984375,
0.9111328125,
0.1685791015625,
0.9111328125,
0.18310546875,
0.51318359375,
0.1817626953125
],
[
0.0882568359375,
0.1875,
0.431396484375,
0.1875,
0.431396484375,
0.2008056640625,
0.0882568359375,
0.2008056640625
],
[
0.515625,
0.18310546875,
0.88818359375,
0.18310546875,
0.88818359375,
0.1962890625,
0.515625,
0.1962890625
],
[
0.08416748046875,
0.210205078125,
0.346435546875,
0.2108154296875,
0.346435546875,
0.2259521484375,
0.08416748046875,
0.2247314453125
],
[
0.59716796875,
0.20703125,
0.83251953125,
0.2083740234375,
0.83251953125,
0.2259521484375,
0.59716796875,
0.2247314453125
],
[
0.8447265625,
0.21533203125,
0.859375,
0.21533203125,
0.859375,
0.222900390625,
0.8447265625,
0.222900390625
],
[
0.8798828125,
0.2108154296875,
0.9111328125,
0.2108154296875,
0.9111328125,
0.2235107421875,
0.8798828125,
0.2235107421875
],
[
0.0882568359375,
0.2298583984375,
0.484375,
0.2298583984375,
0.484375,
0.24365234375,
0.0882568359375,
0.24365234375
],
[
0.51318359375,
0.2342529296875,
0.9111328125,
0.2342529296875,
0.9111328125,
0.2474365234375,
0.51318359375,
0.2474365234375
],
[
0.08660888671875,
0.242431640625,
0.484375,
0.2442626953125,
0.484375,
0.25830078125,
0.08660888671875,
0.255615234375
],
[
0.51123046875,
0.2474365234375,
0.6748046875,
0.246826171875,
0.6748046875,
0.261474609375,
0.51123046875,
0.261962890625
],
[
0.0882568359375,
0.25830078125,
0.483642578125,
0.25830078125,
0.483642578125,
0.271484375,
0.0882568359375,
0.271484375
],
[
0.0882568359375,
0.271484375,
0.484375,
0.271484375,
0.484375,
0.285400390625,
0.0882568359375,
0.285400390625
],
[
0.59228515625,
0.268310546875,
0.83642578125,
0.270751953125,
0.83642578125,
0.291748046875,
0.59228515625,
0.2890625
],
[
0.8388671875,
0.27783203125,
0.87109375,
0.27783203125,
0.87109375,
0.285888671875,
0.8388671875,
0.285888671875
],
[
0.8798828125,
0.27587890625,
0.9111328125,
0.27587890625,
0.9111328125,
0.287353515625,
0.8798828125,
0.287353515625
],
[
0.0882568359375,
0.285888671875,
0.484375,
0.285888671875,
0.484375,
0.2998046875,
0.0882568359375,
0.2998046875
],
[
0.0849609375,
0.298583984375,
0.484375,
0.29736328125,
0.484375,
0.3125,
0.0849609375,
0.313232421875
],
[
0.53125,
0.296630859375,
0.9111328125,
0.296630859375,
0.9111328125,
0.31005859375,
0.53125,
0.31005859375
],
[
0.0882568359375,
0.313232421875,
0.484375,
0.313232421875,
0.484375,
0.326904296875,
0.0882568359375,
0.326904296875
],
[
0.51318359375,
0.309326171875,
0.91259765625,
0.31005859375,
0.91259765625,
0.324462890625,
0.51318359375,
0.323974609375
],
[
0.0882568359375,
0.326904296875,
0.484375,
0.326904296875,
0.484375,
0.341552734375,
0.0882568359375,
0.341552734375
],
[
0.515625,
0.326904296875,
0.91015625,
0.326904296875,
0.91015625,
0.34033203125,
0.515625,
0.34033203125
],
[
0.08660888671875,
0.341552734375,
0.1282958984375,
0.341552734375,
0.1282958984375,
0.35595703125,
0.08660888671875,
0.35595703125
],
[
0.515625,
0.34228515625,
0.9111328125,
0.34228515625,
0.9111328125,
0.35595703125,
0.515625,
0.35595703125
],
[
0.1968994140625,
0.354248046875,
0.374267578125,
0.354736328125,
0.374267578125,
0.369384765625,
0.1968994140625,
0.36865234375
],
[
0.463134765625,
0.35595703125,
0.484375,
0.35595703125,
0.484375,
0.36865234375,
0.463134765625,
0.36865234375
],
[
0.5146484375,
0.354248046875,
0.9111328125,
0.356689453125,
0.9111328125,
0.37060546875,
0.5146484375,
0.367431640625
],
[
0.5107421875,
0.3662109375,
0.9140625,
0.36865234375,
0.9140625,
0.39013671875,
0.5107421875,
0.38818359375
],
[
0.09234619140625,
0.375,
0.486083984375,
0.375,
0.486083984375,
0.38818359375,
0.09234619140625,
0.38818359375
],
[
0.08660888671875,
0.38818359375,
0.486083984375,
0.38818359375,
0.486083984375,
0.401611328125,
0.08660888671875,
0.401611328125
],
[
0.5146484375,
0.385009765625,
0.9111328125,
0.385009765625,
0.9111328125,
0.399658203125,
0.5146484375,
0.399658203125
],
[
0.0882568359375,
0.401611328125,
0.484375,
0.401611328125,
0.484375,
0.415283203125,
0.0882568359375,
0.415283203125
],
[
0.5107421875,
0.395751953125,
0.9140625,
0.3984375,
0.9140625,
0.416748046875,
0.5107421875,
0.41357421875
],
[
0.0882568359375,
0.416748046875,
0.483642578125,
0.416748046875,
0.483642578125,
0.429931640625,
0.0882568359375,
0.429931640625
],
[
0.51318359375,
0.417236328125,
0.9111328125,
0.417236328125,
0.9111328125,
0.431884765625,
0.51318359375,
0.431884765625
],
[
0.08905029296875,
0.429931640625,
0.484375,
0.429931640625,
0.484375,
0.443115234375,
0.08905029296875,
0.443115234375
],
[
0.515625,
0.43115234375,
0.90283203125,
0.43115234375,
0.90283203125,
0.44580078125,
0.515625,
0.44580078125
],
[
0.08905029296875,
0.443115234375,
0.484375,
0.443115234375,
0.484375,
0.458251953125,
0.08905029296875,
0.458251953125
],
[
0.5146484375,
0.44580078125,
0.9111328125,
0.44580078125,
0.9111328125,
0.458984375,
0.5146484375,
0.458984375
],
[
0.0882568359375,
0.458251953125,
0.484375,
0.458251953125,
0.484375,
0.4716796875,
0.0882568359375,
0.4716796875
],
[
0.51318359375,
0.458984375,
0.71826171875,
0.458984375,
0.71826171875,
0.472900390625,
0.51318359375,
0.472900390625
],
[
0.0882568359375,
0.473388671875,
0.408447265625,
0.473388671875,
0.408447265625,
0.484130859375,
0.0882568359375,
0.484130859375
],
[
0.521484375,
0.479248046875,
0.80810546875,
0.479248046875,
0.80810546875,
0.492431640625,
0.521484375,
0.492431640625
],
[
0.7998046875,
0.479248046875,
0.9111328125,
0.479248046875,
0.9111328125,
0.492431640625,
0.7998046875,
0.492431640625
],
[
0.265625,
0.49560546875,
0.40283203125,
0.49560546875,
0.40283203125,
0.51025390625,
0.265625,
0.51025390625
],
[
0.515625,
0.49365234375,
0.9111328125,
0.49365234375,
0.9111328125,
0.5068359375,
0.515625,
0.5068359375
],
[
0.1422119140625,
0.5048828125,
0.23779296875,
0.5048828125,
0.23779296875,
0.5166015625,
0.1422119140625,
0.5166015625
],
[
0.463134765625,
0.5048828125,
0.484375,
0.5048828125,
0.484375,
0.517578125,
0.463134765625,
0.517578125
],
[
0.51318359375,
0.505859375,
0.91015625,
0.505859375,
0.91015625,
0.51953125,
0.51318359375,
0.51953125
],
[
0.5146484375,
0.51953125,
0.8896484375,
0.51953125,
0.8896484375,
0.5341796875,
0.5146484375,
0.5341796875
],
[
0.236083984375,
0.50830078125,
0.43310546875,
0.51025390625,
0.43310546875,
0.53076171875,
0.236083984375,
0.52880859375
],
[
0.08660888671875,
0.53515625,
0.486083984375,
0.537109375,
0.486083984375,
0.5546875,
0.08660888671875,
0.55322265625
],
[
0.5361328125,
0.54443359375,
0.861328125,
0.54443359375,
0.861328125,
0.55810546875,
0.5361328125,
0.55810546875
],
[
0.8544921875,
0.54443359375,
0.91015625,
0.54443359375,
0.91015625,
0.55615234375,
0.8544921875,
0.55615234375
],
[
0.08905029296875,
0.5537109375,
0.484375,
0.5537109375,
0.484375,
0.568359375,
0.08905029296875,
0.568359375
],
[
0.0882568359375,
0.5673828125,
0.25732421875,
0.5673828125,
0.25732421875,
0.58056640625,
0.0882568359375,
0.58056640625
],
[
0.5146484375,
0.56640625,
0.5458984375,
0.56640625,
0.5458984375,
0.5791015625,
0.5146484375,
0.5791015625
],
[
0.2418212890625,
0.59033203125,
0.2548828125,
0.59033203125,
0.2548828125,
0.599609375,
0.2418212890625,
0.599609375
],
[
0.54736328125,
0.5869140625,
0.8505859375,
0.58837890625,
0.8505859375,
0.60302734375,
0.54736328125,
0.6015625
],
[
0.8447265625,
0.59228515625,
0.8759765625,
0.59228515625,
0.8759765625,
0.599609375,
0.8447265625,
0.599609375
],
[
0.8798828125,
0.59033203125,
0.91015625,
0.59033203125,
0.91015625,
0.6015625,
0.8798828125,
0.6015625
],
[
0.19287109375,
0.60107421875,
0.381591796875,
0.60107421875,
0.381591796875,
0.6220703125,
0.19287109375,
0.6220703125
],
[
0.463134765625,
0.60400390625,
0.486083984375,
0.60400390625,
0.486083984375,
0.6171875,
0.463134765625,
0.6171875
],
[
0.51318359375,
0.611328125,
0.9111328125,
0.611328125,
0.9111328125,
0.625,
0.51318359375,
0.625
],
[
0.236083984375,
0.619140625,
0.2607421875,
0.619140625,
0.2607421875,
0.630859375,
0.236083984375,
0.630859375
],
[
0.5146484375,
0.626953125,
0.54833984375,
0.626953125,
0.54833984375,
0.6396484375,
0.5146484375,
0.6396484375
],
[
0.101318359375,
0.6396484375,
0.483642578125,
0.6396484375,
0.483642578125,
0.65283203125,
0.101318359375,
0.65283203125
],
[
0.625,
0.6396484375,
0.80126953125,
0.6396484375,
0.80126953125,
0.65283203125,
0.625,
0.65283203125
],
[
0.8798828125,
0.63818359375,
0.9140625,
0.63818359375,
0.9140625,
0.65380859375,
0.8798828125,
0.65380859375
],
[
0.0882568359375,
0.65283203125,
0.483642578125,
0.65283203125,
0.483642578125,
0.66650390625,
0.0882568359375,
0.66650390625
],
[
0.52783203125,
0.6591796875,
0.91259765625,
0.6591796875,
0.91259765625,
0.67236328125,
0.52783203125,
0.67236328125
],
[
0.08660888671875,
0.66748046875,
0.25244140625,
0.66748046875,
0.25244140625,
0.68115234375,
0.08660888671875,
0.68115234375
],
[
0.515625,
0.671875,
0.71337890625,
0.671875,
0.71337890625,
0.68505859375,
0.515625,
0.68505859375
],
[
0.521484375,
0.69140625,
0.59228515625,
0.693359375,
0.59228515625,
0.705078125,
0.521484375,
0.70263671875
],
[
0.609375,
0.69140625,
0.91015625,
0.69140625,
0.91015625,
0.705078125,
0.609375,
0.705078125
],
[
0.5146484375,
0.70556640625,
0.9111328125,
0.70556640625,
0.9111328125,
0.7197265625,
0.5146484375,
0.7197265625
],
[
0.08660888671875,
0.6923828125,
0.346435546875,
0.6923828125,
0.346435546875,
0.70556640625,
0.08660888671875,
0.70556640625
],
[
0.0882568359375,
0.71044921875,
0.483642578125,
0.71044921875,
0.483642578125,
0.7236328125,
0.0882568359375,
0.7236328125
],
[
0.51123046875,
0.71728515625,
0.90869140625,
0.71826171875,
0.90869140625,
0.73291015625,
0.51123046875,
0.73193359375
],
[
0.08905029296875,
0.724609375,
0.484375,
0.724609375,
0.484375,
0.7392578125,
0.08905029296875,
0.7392578125
],
[
0.51318359375,
0.73291015625,
0.9111328125,
0.73291015625,
0.9111328125,
0.7470703125,
0.51318359375,
0.7470703125
],
[
0.0882568359375,
0.7392578125,
0.253173828125,
0.7392578125,
0.253173828125,
0.75244140625,
0.0882568359375,
0.75244140625
],
[
0.5146484375,
0.7470703125,
0.9111328125,
0.7470703125,
0.9111328125,
0.76025390625,
0.5146484375,
0.76025390625
],
[
0.0906982421875,
0.7568359375,
0.48779296875,
0.75830078125,
0.48779296875,
0.771484375,
0.0906982421875,
0.77099609375
],
[
0.5146484375,
0.76220703125,
0.81396484375,
0.76220703125,
0.81396484375,
0.77294921875,
0.5146484375,
0.77294921875
],
[
0.08660888671875,
0.771484375,
0.261474609375,
0.77294921875,
0.261474609375,
0.787109375,
0.08660888671875,
0.7861328125
],
[
0.2646484375,
0.771484375,
0.484375,
0.77099609375,
0.484375,
0.78515625,
0.2646484375,
0.7861328125
],
[
0.51123046875,
0.7744140625,
0.92626953125,
0.7744140625,
0.92626953125,
0.82763671875,
0.51123046875,
0.82763671875
],
[
0.0882568359375,
0.78515625,
0.486083984375,
0.78515625,
0.486083984375,
0.7998046875,
0.0882568359375,
0.7998046875
],
[
0.619140625,
0.7841796875,
0.6357421875,
0.7841796875,
0.6357421875,
0.79248046875,
0.619140625,
0.79248046875
],
[
0.08905029296875,
0.7998046875,
0.488525390625,
0.7998046875,
0.488525390625,
0.814453125,
0.08905029296875,
0.814453125
],
[
0.0849609375,
0.814453125,
0.486083984375,
0.8125,
0.486083984375,
0.8271484375,
0.0849609375,
0.8291015625
],
[
0.615234375,
0.8125,
0.69140625,
0.8125,
0.69140625,
0.82275390625,
0.615234375,
0.82275390625
],
[
0.1405029296875,
0.8359375,
0.4453125,
0.8359375,
0.4453125,
0.85400390625,
0.1405029296875,
0.85400390625
],
[
0.5146484375,
0.833984375,
0.91259765625,
0.833984375,
0.91259765625,
0.84765625,
0.5146484375,
0.84765625
],
[
0.44677734375,
0.84521484375,
0.455810546875,
0.84521484375,
0.455810546875,
0.8486328125,
0.44677734375,
0.8486328125
],
[
0.463134765625,
0.84130859375,
0.483642578125,
0.84130859375,
0.483642578125,
0.8515625,
0.463134765625,
0.8515625
],
[
0.5146484375,
0.84521484375,
0.91015625,
0.8466796875,
0.91015625,
0.8623046875,
0.5146484375,
0.861328125
],
[
0.0882568359375,
0.8623046875,
0.486083984375,
0.8623046875,
0.486083984375,
0.87548828125,
0.0882568359375,
0.87548828125
],
[
0.5146484375,
0.8623046875,
0.9111328125,
0.8623046875,
0.9111328125,
0.876953125,
0.5146484375,
0.876953125
],
[
0.0882568359375,
0.87548828125,
0.318603515625,
0.87548828125,
0.318603515625,
0.8896484375,
0.0882568359375,
0.8896484375
],
[
0.51708984375,
0.87548828125,
0.7099609375,
0.87548828125,
0.7099609375,
0.8896484375,
0.51708984375,
0.8896484375
]
] |
[
"3.1 Notation",
"Model Context..",
"为了对单词wi的上下文进行建模",
"在整个工作中，我们将表示为单词空间，将V表",
"我们在其上下文单词序列X上应用了双向长短期记忆",
"示为训练词汇。由单词序列组成的句子表示为X",
"网络（BiLSTM）。在这里，我们考虑在两种不同情况下",
"[W1,..,wn】E，其相应的标签序列表示为Y",
"应用这种做法。在第一种情况下，句子X中只有一个",
"{y1,..，yn]V。给定其相应句子X的单词wi的",
"OOV单词。而在第二种情况下，X中有多个OOV单词",
"上下文表示为X=【w1,..，wi-1,wi+1,..，wn}，而",
"当一个句子X中只有一个OOV单词Wi时，其余单词",
"X<ii>是指子序列【wi，.，w,}。教师网络表示为T",
"的任务特定表示都是已知的，我们可以直接将BLSTM",
"学生网络表示为S。此外，令f是在2上定义的函数，",
"应用于x来预测w的表示。具体来说，我们使用正",
"我们将f(X）表示为（f(wi)，，f(wn))的简写",
"向LSTM来对从X的开头到Wi-1的序列进行建模：",
"3.2TraintheTeacherNetwork",
"(w;|X) = LSTM(e(X<1,i-1>)",
"",
"(10)",
"教师网络用于在序列标记系统中进行标签预测。它可",
"，并使用反向LSTM以相反的顺序对从X的末尾到",
"以是适合任务的任意架构。因为这项工作的主要重点是",
"Wi+1的序列进行建模：",
"处理OOV问题，而不是设计一个优越的监督模型，所",
"以在这项工作中，我们尝试了典型的基于CNN和基于",
"(w;|X) = LSTM(e(X<i+1,n>).",
"",
"(11)",
"LSTM的体系结构[Huangetal.，2015]来实现教师网",
"络。简而言之，此体系结构表示一个单词W，其中二个",
"当一个句子中有多个OOV单词时，对于给定的OOV",
"唯一的密集向量（嵌人）ew（w）和一个向量ec（w）使用",
"单词wiex，x，中其他一些单词的表示形式也是未知",
"字符级卷积神经网络（CNN）[Kim，2014]从其字符序列",
"的。因此，我们不能直接将BLSTM应用于X以获得",
"建模：",
"其预测表示。为了解决这个问题，我们建议迭代预测句",
"e(w) = [ew(w) 田ec(w)],",
"(6)",
"子中多个OOV单词的表示形式。令utontext（wiX）表示",
"在 tth与vcontext(w:/X)=0选代时给定X的w;的预测",
"其中④表示串联操作。将相应的句子表示e(X）作为",
"输人输人到双向长短期记忆网络（BLSTM）[Hochreiter",
"表示。在（t+1)th选代中，为了预测wi的表示，我们",
"and Schmidhuber，1997]或内核大小设置为3的三层",
"将BiLSTM应用于其tth送代上下文utontext(X），获",
"CNN网络[Kim，2014]中，以对每个单词的上下文信息",
"得ucontext（w;X）。对所有OOV单词重复此过程以完成",
"进行建模，从而获得给定X的每个单词的隐藏表示（不是",
"(t+1)th选代。选代继续进行，直到t达到固定值K",
"单词嵌人）h(wiX）。在基于BLSTM或基于CNN的序",
"从而获得每个OOV单词的最终预测表示。根据我们的",
"列建模层之上，它使用顺序条件随机场（CRF）[Lafferty",
"经验，设置K=2是合适的",
"etal.，2001l对整个句子的标签进行联合解码",
"Combine Surface-form and Context.",
"我们最终将表",
"IIt=1 Φt(yt-1, yt[X)",
"面形式和上下文的表示结合起来，以获得W的联合表示",
"p(YX;0T) =",
"(7)",
"u(w|X）。在这项工作中，我们遵循[Schickand Schitze",
"2018]的思想，并将这两个表示与一个门结合在一起：",
"yey I t(yt-, rx)",
"其中pt(y,ylx)=exp(wf.h(wt)+by,y)，wy,和by',y",
"Vw(w|X) =QVform(w) + (1-Q)vcontext(w|X)",
"(12)",
"是可训练的参数对应于整个标签对（y',y），。然后，教师",
"网络的培训损失定义为：",
"其中",
"N",
"α = o(wT [Uform(w) @ context (w|X)l +b).",
"",
"(13)",
"CT = logp(Y;(X; OT),",
"(8)",
"这里，g表示sigmoid函数。对于表达式一致性，我们",
"i=1",
"表示",
"其中N是培训句子编号。该网络在训练集中出现的所",
"vw(w|X;0s) = ww(wX),",
"(14)",
"有单词上进行训练。之后，我们在以下描述的学生网络",
"其中θs是学生网络的那些可训练参数，实际上对应",
"的培训期间修复其参数",
"于图1中描述的表示预测层",
"Training.",
"因为我们使用教师网络来执行标签预测",
"所以学生网络的培训对象应该学习为适合教师网络的",
"3.3Train the Student Network",
"学生网络对单词的表面形式和上下文信息进行建模，以",
"OoV单词生成表示。根据以前的作品[Lazaridouetal.",
"生成其表示形式。Schick and Schitze、2018证明了这",
"2017; Khodak et al., 2018; Schick and Schitze, 2018] 的",
"两种信息资源是互补的",
"思想，我们可以训练学生网络以最小化具有由以下定义",
"ModelSurface-form.单词w的表面形式表示是从其",
"的损失函数的那些词汇内单词的重建错误：",
"字符序列W={c1,C2，·",
",CnlEV。获得的，其中V是",
"",
"字符词汇。具体来说，在[Schickand Schitze,2018]的",
"N",
"工作之后，我们首先用特殊的开始和结束字符co=《s），",
"Cn+1=《e）填充字符序列，并获得其向上的k-gram集：",
"i=lwex,nv",
"G(w) = Um-1 Un+?-m [ci,.- ,Ci+m-1].",
"这里，d(，）定义了两个嵌入之间的距离，例如。g。，欧",
"",
"(9)",
"几里得距离函数。这种做法的最大问题是它可能会遇到",
"然后，我们将W的表面形式嵌人定义为其所有向上",
"错误传播问题，因为辅助重建训练标准Crecon不能保证",
"XMATHX_LINE嵌人的平均值",
"与教师网络的训练对象兼容"
] |
{
"arxiv_id": "1905.12277",
"language": "zh",
"page_idx": 2
}
|
||
1905.12277-zh-3
|
1905.12277-zh-3.jpg
|
<table>
<tbody>
<tr>
<td>$A$</td>
<td>m 1 Training of the Student Network</td>
</tr>
<tr>
<td>$f$</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>Get $\mathcal{L} _S($X,Y$;\boldsymbol\theta_S,\boldsymbol\theta_T)$ based on $\boldsymbol{v}($X) accord- ${\mathrm{to~Eq.~}(16);}$</td>
</tr>
<tr>
<td>1 1 I</td>
<td>Update $\boldsymbol{\theta }_{S}\leftarrow \boldsymbol{\theta }_{S}- \alpha$ $\partial\mathcal{L}_{S}(\mathbf{X},\mathbf{Y};\boldsymbol{\theta}_{S},\boldsymbol{\theta}_{T}$ $\overline{\partial\boldsymbol{\theta}_S}$</td>
</tr>
</tbody>
</table>
在这项工作中，我们建议将学生网络的培训与教师网络的培训对象直接联系起来，从而避免设计适当的辅助培训标准。不失一般性，我们首先考虑在给定句子中只有一个 OOV 单词的情况下对学生网络进行培训。为了模拟这种情况，对于训练句子 X 及其相应的标签序列 Y ,我们从 X 中随机抽取一个单词$w_i$ 。结果对$( w_i, \mathrm{X} _\setminus i$,Y)构成了学生网络的训练示例。通过将$e(\mathcal{X})$的$i^th$列替换为$\boldsymbol{v}( w| \mathcal{X} ; \boldsymbol{\theta }_{S}) = [ \boldsymbol{v}_{w}( w| \mathcal{X} ; \boldsymbol{\theta }_{S}) \oplus \boldsymbol{e}_{c}( w) ]$ ,我们获得了教师网络 的 输 入 $\boldsymbol{v}( \dot{\mathcal{X} } ) = \{ \boldsymbol{e}( w_{1}) , \cdots , \boldsymbol{v}( w_{i}| \mathbf{X} ; \boldsymbol{\theta }_{S}) , \cdots , \boldsymbol{e}( w_{n}) \}$ . 基于$v($X),我们为每个单词$\widehat{\mathrm{h}}(w_i|$X) 获得新的隐藏表示，然后将任务损失定义为：
$$\mathcal{L}_{S}(\mathrm{X},\mathrm{Y};\boldsymbol{\theta}_{S},\boldsymbol{\theta}_{T})=\log\frac{\prod_{t=1}^{m}\phi_{t}(y_{t-1},y_{t}|\mathrm{X})}{\sum_{Y'\in\mathcal{Y}}\prod_{t=1}^{m}\phi_{t}(y_{t-1}^{\prime},y_{t}^{\prime}|\mathrm{X})},$$
(16)
,其中这次是$\phi _t( y^{\prime }, y| \mathcal{X} ) = \exp ( \boldsymbol{w}_{y^{\prime }, y^{\prime }}^T\widehat{\mathrm{h} } ( w_t) + b_{y^{\prime }, y})$。请
注意，学生网络的培训损失与教师网络的培训损失相同， 确保他们的培训兼容。为了训练学生网络，我们通过以下方式进行参数更新：
$$\theta_S\leftarrow\theta_S-\alpha\frac{\partial\mathcal{L}_S(\mathrm{X},\mathrm{Y};\boldsymbol{\theta}_S,\boldsymbol{\theta}_T)}{\partial\boldsymbol{\theta}_S}.$$
(17)
在训练学生网络期间，固定了教师网络$\theta_T$的参数。算法 1 显示了在句子中只有一个 OOV 单词的情况下，学生网络的一般训练过程
要将此过程扩展到多个 OOV 单词的情况，我们从 X , e 中采样多个单词。g。$,w_i$和$w_j$,生成学生网络的训练示例。在此示例中定义的损耗类似于在一个 OOV 情况下的损耗，并且在$\theta_S$上最小化
## 4 Experiments
为了评估我们提出的方法的有效性，我们在四个词性标记(POS)任务和四个命名实体识别(NER)任务上进行了实验。这些任务具有不同的 OOV 率，这是由训练集中出现少于五次的测试单词的百分比定义的。这些任务共享与 83 中所示相同的教师和学生网络架构
## 4.1 Datasets
POS: 对于 POS, 我们进行了以下实验：(1) PTBEnglish :英文 PennTreebank 数据集 [Marcus et al., 1993]的华尔街日报部分，(2) RIT-English :从英文Tweets 创建的数据集[Derczynski et al., 2013], (3)。对于 PTB-English,我们遵循标准拆分：第 2-21 节用
<table>
<tbody>
<tr>
<th rowspan="3">Dataset</th>
<th>$\#$OOV</th>
<th>Dev $OOV$ Rate</th>
<th>$\#$OOV</th>
<th>Test $OOV$ $Rat\epsilon$ </th>
</tr>
<tr>
<th colspan="2">Dev</th>
<th colspan="2">Test</th>
</tr>
<tr>
<th>$\#$OOV</th>
<th>$00V$ Rate</th>
<th>$\#$OOV</th>
<th>$OOV$ $Rat\epsilon$</th>
</tr>
<tr>
<td>$POS$</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>PTB-English</td>
<td>8,392</td>
<td>$6.37\%$</td>
<td>7,528</td>
<td>5.81 0</td>
</tr>
<tr>
<td>RIT-English</td>
<td>774</td>
<td>$34.52\%$</td>
<td>760</td>
<td>33.17</td>
</tr>
<tr>
<td>GSD-Russian</td>
<td>21.323</td>
<td>17.96 $\%$</td>
<td>21,523</td>
<td>18.31</td>
</tr>
<tr>
<td>RRT-Romanian</td>
<td>3,965</td>
<td>$23.22\%$</td>
<td>3,702</td>
<td>22.67 $\%$</td>
</tr>
<tr>
<td>$NER$</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>CoNLL02-Spanish</td>
<td>2,216</td>
<td>$50.91\%$</td>
<td>1,544</td>
<td>43.38 $0Z$</td>
</tr>
<tr>
<td>CoNLL02-Dutch</td>
<td>1,819</td>
<td>$69.53\%$</td>
<td>2,564</td>
<td></td>
</tr>
<tr>
<td>Twitter- English</td>
<td>1,266</td>
<td>$79.15\%$</td>
<td>4,131</td>
<td>79.13 07</td>
</tr>
<tr>
<td>CoNLL03-German</td>
<td>3,928</td>
<td>81.27 $\%$</td>
<td>2,685</td>
<td>73.10 $\%$</td>
</tr>
</tbody>
</table>
Table 1: Number of OOV words (for POS) and entities ( for NER) in the development and testing sets, when treating words occurring less than 5 times in the training set as OOV. An entity is treated as OOV if it contains at least one OOV word.
于培训，第 22 节用于验证，第 23 节用于测试。对于RIT-English ,我们遵循 Gui 等人的拆分协议。,[2017] 。而对于 UD-俄语和 UD-罗马尼亚语，我们使用了它们给定的数据分割
NER: 对于 NER, 我们在以下方面进行了实验：(1) CoNLL02-Spanish : CoNLL2002 西班牙语 NER 共享任务数据集 [Sang, 2002] ; (2) CoNLL02-Dutch : 荷兰语的 CoNLL2002 数据集；(3) Twitter-English :从TweetsXMATHX\_THMA2 创建的英语 N。这些数据集由四种类型注释：PER, LOC, ORG 和 MISC。对于除Twitter-English 之外的数据集，我们使用官方的拆分训练集进行模型训练，使用 testa 进行验证，使用 testb 进行测试。而对于 Twitter-English,我们遵循了 Zhang 等人的数据拆分。,[2018]
表 1 报告了 OOV 问题在每个数据集的开发和测试集上的统计结果。从表中，我们可以看到 OOV 率在不同的数据集上有很大差异
### 4.2 Compared Methods
我们比较了以下基线：
$\bullet$RandomUNK:This baseline refers to the teacher network trained on all words occurring in the training set. At testing time, it represents words not occurring in the training set with a consistent random vector.
$\bullet$ SingleUNK: This baseline trains the teacher network on words that occur no less than 5 times in the training set. The other infrequent words and those words not occurring in the training set are all mapped to a single trainable embedding $\boldsymbol e_\mathrm{UNK}$, which is trained during model training.
$\dot{\text{RandomUNK as the teacher network. }}$ At testing time, it frst represents OOV words with their context word representations by mean-pooling as defined in equation (5) based on RandomUNK and then performs label prediction using RandomUNK. $\bullet$ [Khodak et al., 2018]: This baseline additionally trains a linear transformer A to transform the
|
[
[
0.0882568359375,
0.0694580078125,
0.414306640625,
0.0694580078125,
0.414306640625,
0.08331298828125,
0.0882568359375,
0.08331298828125
],
[
0.69140625,
0.07135009765625,
0.71826171875,
0.07135009765625,
0.71826171875,
0.083984375,
0.69140625,
0.083984375
],
[
0.826171875,
0.07135009765625,
0.8544921875,
0.07135009765625,
0.8544921875,
0.083984375,
0.826171875,
0.083984375
],
[
0.52392578125,
0.07891845703125,
0.57177734375,
0.07891845703125,
0.57177734375,
0.0889892578125,
0.52392578125,
0.0889892578125
],
[
0.64453125,
0.08331298828125,
0.69140625,
0.08331298828125,
0.69140625,
0.0947265625,
0.64453125,
0.0947265625
],
[
0.701171875,
0.08331298828125,
0.76708984375,
0.08331298828125,
0.76708984375,
0.0947265625,
0.701171875,
0.0947265625
],
[
0.77783203125,
0.08331298828125,
0.826171875,
0.08331298828125,
0.826171875,
0.0947265625,
0.77783203125,
0.0947265625
],
[
0.8349609375,
0.08209228515625,
0.900390625,
0.08209228515625,
0.900390625,
0.09344482421875,
0.8349609375,
0.09344482421875
],
[
0.0931396484375,
0.0889892578125,
0.473876953125,
0.0889892578125,
0.473876953125,
0.1029052734375,
0.0931396484375,
0.1029052734375
],
[
0.69384765625,
0.08709716796875,
0.701171875,
0.08709716796875,
0.701171875,
0.09027099609375,
0.69384765625,
0.09027099609375
],
[
0.09234619140625,
0.1029052734375,
0.334228515625,
0.1029052734375,
0.334228515625,
0.11614990234375,
0.09234619140625,
0.11614990234375
],
[
0.51953125,
0.0985107421875,
0.5537109375,
0.0985107421875,
0.5537109375,
0.111083984375,
0.51953125,
0.111083984375
],
[
0.52392578125,
0.11297607421875,
0.6005859375,
0.11297607421875,
0.6005859375,
0.125,
0.52392578125,
0.125
],
[
0.65625,
0.11297607421875,
0.69140625,
0.11297607421875,
0.69140625,
0.125,
0.65625,
0.125
],
[
0.7236328125,
0.111083984375,
0.765625,
0.11297607421875,
0.7646484375,
0.1256103515625,
0.72216796875,
0.12371826171875
],
[
0.79150390625,
0.11297607421875,
0.826171875,
0.11297607421875,
0.826171875,
0.1256103515625,
0.79150390625,
0.1256103515625
],
[
0.859375,
0.11297607421875,
0.90185546875,
0.11297607421875,
0.90185546875,
0.125,
0.859375,
0.125
],
[
0.52392578125,
0.12371826171875,
0.59716796875,
0.125,
0.59716796875,
0.1363525390625,
0.52392578125,
0.1351318359375
],
[
0.09234619140625,
0.117431640625,
0.330810546875,
0.117431640625,
0.330810546875,
0.1307373046875,
0.09234619140625,
0.1307373046875
],
[
0.66650390625,
0.125,
0.69140625,
0.125,
0.69140625,
0.1363525390625,
0.66650390625,
0.1363525390625
],
[
0.71826171875,
0.125,
0.765625,
0.125,
0.765625,
0.1363525390625,
0.71826171875,
0.1363525390625
],
[
0.80126953125,
0.12371826171875,
0.82666015625,
0.12371826171875,
0.82666015625,
0.1363525390625,
0.80126953125,
0.1363525390625
],
[
0.1356201171875,
0.1307373046875,
0.486083984375,
0.1307373046875,
0.486083984375,
0.14453125,
0.1356201171875,
0.14453125
],
[
0.8544921875,
0.125,
0.900390625,
0.125,
0.900390625,
0.1363525390625,
0.8544921875,
0.1363525390625
],
[
0.09478759765625,
0.1319580078125,
0.1094970703125,
0.1319580078125,
0.1094970703125,
0.141357421875,
0.09478759765625,
0.141357421875
],
[
0.525390625,
0.1363525390625,
0.60107421875,
0.1363525390625,
0.60107421875,
0.147705078125,
0.525390625,
0.147705078125
],
[
0.650390625,
0.1363525390625,
0.69140625,
0.1363525390625,
0.69140625,
0.147705078125,
0.650390625,
0.147705078125
],
[
0.71826171875,
0.1363525390625,
0.765625,
0.1363525390625,
0.765625,
0.147705078125,
0.71826171875,
0.147705078125
],
[
0.78515625,
0.1363525390625,
0.826171875,
0.1363525390625,
0.826171875,
0.147705078125,
0.78515625,
0.147705078125
],
[
0.8544921875,
0.1363525390625,
0.900390625,
0.1363525390625,
0.900390625,
0.147705078125,
0.8544921875,
0.147705078125
],
[
0.11358642578125,
0.14453125,
0.295654296875,
0.14453125,
0.295654296875,
0.1590576171875,
0.11358642578125,
0.1590576171875
],
[
0.525390625,
0.1458740234375,
0.61669921875,
0.1458740234375,
0.61669921875,
0.1572265625,
0.525390625,
0.1572265625
],
[
0.65625,
0.146484375,
0.69140625,
0.146484375,
0.69140625,
0.1590576171875,
0.65625,
0.1590576171875
],
[
0.71826171875,
0.146484375,
0.765625,
0.146484375,
0.765625,
0.1578369140625,
0.71826171875,
0.1578369140625
],
[
0.7900390625,
0.146484375,
0.826171875,
0.146484375,
0.826171875,
0.1590576171875,
0.7900390625,
0.1590576171875
],
[
0.85302734375,
0.146484375,
0.90185546875,
0.146484375,
0.90185546875,
0.1578369140625,
0.85302734375,
0.1578369140625
],
[
0.09478759765625,
0.160400390625,
0.1094970703125,
0.160400390625,
0.1094970703125,
0.17041015625,
0.09478759765625,
0.17041015625
],
[
0.136474609375,
0.1578369140625,
0.293212890625,
0.1578369140625,
0.293212890625,
0.1717529296875,
0.136474609375,
0.1717529296875
],
[
0.52294921875,
0.1610107421875,
0.5537109375,
0.1610107421875,
0.5537109375,
0.173583984375,
0.52294921875,
0.173583984375
],
[
0.09234619140625,
0.17236328125,
0.111083984375,
0.17236328125,
0.111083984375,
0.186279296875,
0.09234619140625,
0.186279296875
],
[
0.160888671875,
0.1717529296875,
0.38818359375,
0.1717529296875,
0.38818359375,
0.186279296875,
0.160888671875,
0.186279296875
],
[
0.52294921875,
0.175537109375,
0.63134765625,
0.175537109375,
0.63134765625,
0.189453125,
0.52294921875,
0.189453125
],
[
0.65673828125,
0.1767578125,
0.69140625,
0.1767578125,
0.69140625,
0.1881103515625,
0.65673828125,
0.1881103515625
],
[
0.71826171875,
0.1767578125,
0.765625,
0.1767578125,
0.765625,
0.1881103515625,
0.71826171875,
0.1881103515625
],
[
0.79345703125,
0.1767578125,
0.826171875,
0.1767578125,
0.826171875,
0.1881103515625,
0.79345703125,
0.1881103515625
],
[
0.8544921875,
0.1767578125,
0.90185546875,
0.1767578125,
0.90185546875,
0.1881103515625,
0.8544921875,
0.1881103515625
],
[
0.0931396484375,
0.186279296875,
0.111083984375,
0.186279296875,
0.111083984375,
0.199462890625,
0.0931396484375,
0.199462890625
],
[
0.16259765625,
0.186279296875,
0.463134765625,
0.186279296875,
0.463134765625,
0.2008056640625,
0.16259765625,
0.2008056640625
],
[
0.52392578125,
0.186279296875,
0.62109375,
0.1881103515625,
0.62109375,
0.199462890625,
0.52392578125,
0.1976318359375
],
[
0.65673828125,
0.1881103515625,
0.69140625,
0.1881103515625,
0.69140625,
0.199462890625,
0.65673828125,
0.199462890625
],
[
0.71826171875,
0.1881103515625,
0.765625,
0.1881103515625,
0.765625,
0.199462890625,
0.71826171875,
0.199462890625
],
[
0.79345703125,
0.1881103515625,
0.82666015625,
0.1881103515625,
0.82666015625,
0.2008056640625,
0.79345703125,
0.2008056640625
],
[
0.85302734375,
0.1881103515625,
0.900390625,
0.1881103515625,
0.900390625,
0.199462890625,
0.85302734375,
0.199462890625
],
[
0.525390625,
0.1976318359375,
0.61669921875,
0.1988525390625,
0.61669921875,
0.210205078125,
0.525390625,
0.208984375
],
[
0.65869140625,
0.1988525390625,
0.69140625,
0.1988525390625,
0.69140625,
0.210205078125,
0.65869140625,
0.210205078125
],
[
0.7197265625,
0.1988525390625,
0.765625,
0.1988525390625,
0.765625,
0.210205078125,
0.7197265625,
0.210205078125
],
[
0.79345703125,
0.1988525390625,
0.826171875,
0.1988525390625,
0.826171875,
0.210205078125,
0.79345703125,
0.210205078125
],
[
0.8544921875,
0.1988525390625,
0.900390625,
0.1988525390625,
0.900390625,
0.210205078125,
0.8544921875,
0.210205078125
],
[
0.16259765625,
0.199462890625,
0.484375,
0.199462890625,
0.484375,
0.21337890625,
0.16259765625,
0.21337890625
],
[
0.09234619140625,
0.199462890625,
0.111083984375,
0.199462890625,
0.111083984375,
0.21337890625,
0.09234619140625,
0.21337890625
],
[
0.52392578125,
0.2083740234375,
0.63134765625,
0.210205078125,
0.63134765625,
0.2215576171875,
0.52392578125,
0.2197265625
],
[
0.65673828125,
0.20703125,
0.69140625,
0.208984375,
0.689453125,
0.2215576171875,
0.65625,
0.2197265625
],
[
0.71826171875,
0.208984375,
0.765625,
0.208984375,
0.765625,
0.2203369140625,
0.71826171875,
0.2203369140625
],
[
0.79345703125,
0.208984375,
0.826171875,
0.208984375,
0.826171875,
0.2203369140625,
0.79345703125,
0.2203369140625
],
[
0.8544921875,
0.208984375,
0.90185546875,
0.208984375,
0.90185546875,
0.2203369140625,
0.8544921875,
0.2203369140625
],
[
0.158447265625,
0.2291259765625,
0.420654296875,
0.2203369140625,
0.42236328125,
0.242431640625,
0.16015625,
0.250732421875
],
[
0.11273193359375,
0.2139892578125,
0.2296142578125,
0.2139892578125,
0.2296142578125,
0.2279052734375,
0.11273193359375,
0.2279052734375
],
[
0.09234619140625,
0.2310791015625,
0.111083984375,
0.2310791015625,
0.111083984375,
0.2442626953125,
0.09234619140625,
0.2442626953125
],
[
0.351318359375,
0.23486328125,
0.378173828125,
0.2374267578125,
0.37744140625,
0.25,
0.348876953125,
0.246826171875
],
[
0.51123046875,
0.2374267578125,
0.9140625,
0.238037109375,
0.9140625,
0.25244140625,
0.51123046875,
0.251953125
],
[
0.51318359375,
0.251953125,
0.91259765625,
0.25244140625,
0.91259765625,
0.266357421875,
0.51318359375,
0.26513671875
],
[
0.51123046875,
0.266357421875,
0.91259765625,
0.267578125,
0.91259765625,
0.281005859375,
0.51123046875,
0.27978515625
],
[
0.10455322265625,
0.2734375,
0.484375,
0.2734375,
0.484375,
0.287353515625,
0.10455322265625,
0.287353515625
],
[
0.51318359375,
0.27978515625,
0.91259765625,
0.27978515625,
0.91259765625,
0.29345703125,
0.51318359375,
0.29345703125
],
[
0.08660888671875,
0.287353515625,
0.484375,
0.287353515625,
0.484375,
0.300537109375,
0.08660888671875,
0.300537109375
],
[
0.51318359375,
0.29345703125,
0.685546875,
0.29345703125,
0.685546875,
0.306884765625,
0.51318359375,
0.306884765625
],
[
0.08660888671875,
0.300537109375,
0.484375,
0.300537109375,
0.484375,
0.314453125,
0.08660888671875,
0.314453125
],
[
0.10296630859375,
0.31494140625,
0.484375,
0.31494140625,
0.484375,
0.328857421875,
0.10296630859375,
0.328857421875
],
[
0.08660888671875,
0.328857421875,
0.486083984375,
0.328857421875,
0.486083984375,
0.34228515625,
0.08660888671875,
0.34228515625
],
[
0.51318359375,
0.326904296875,
0.9111328125,
0.32763671875,
0.9111328125,
0.34228515625,
0.51318359375,
0.341552734375
],
[
0.0882568359375,
0.34228515625,
0.486083984375,
0.34228515625,
0.486083984375,
0.356689453125,
0.0882568359375,
0.356689453125
],
[
0.51318359375,
0.34228515625,
0.9111328125,
0.34228515625,
0.9111328125,
0.35595703125,
0.51318359375,
0.35595703125
],
[
0.0882568359375,
0.35791015625,
0.484375,
0.35791015625,
0.484375,
0.371826171875,
0.0882568359375,
0.371826171875
],
[
0.529296875,
0.35595703125,
0.91259765625,
0.35595703125,
0.91259765625,
0.37060546875,
0.529296875,
0.37060546875
],
[
0.08660888671875,
0.37255859375,
0.484375,
0.37255859375,
0.484375,
0.386474609375,
0.08660888671875,
0.386474609375
],
[
0.51318359375,
0.37060546875,
0.6298828125,
0.37060546875,
0.6298828125,
0.3837890625,
0.51318359375,
0.3837890625
],
[
0.0849609375,
0.386474609375,
0.490234375,
0.386962890625,
0.490234375,
0.401611328125,
0.0849609375,
0.40087890625
],
[
0.52294921875,
0.38818359375,
0.9140625,
0.38818359375,
0.9140625,
0.401611328125,
0.52294921875,
0.401611328125
],
[
0.0882568359375,
0.40283203125,
0.486083984375,
0.40283203125,
0.486083984375,
0.417236328125,
0.0882568359375,
0.417236328125
],
[
0.51318359375,
0.401611328125,
0.9111328125,
0.401611328125,
0.9111328125,
0.416748046875,
0.51318359375,
0.416748046875
],
[
0.0849609375,
0.417236328125,
0.291748046875,
0.416748046875,
0.291748046875,
0.43115234375,
0.0849609375,
0.431884765625
],
[
0.515625,
0.416748046875,
0.91259765625,
0.416748046875,
0.91259765625,
0.429931640625,
0.515625,
0.429931640625
],
[
0.515625,
0.4287109375,
0.9111328125,
0.4287109375,
0.9111328125,
0.443115234375,
0.515625,
0.443115234375
],
[
0.098876953125,
0.436279296875,
0.471435546875,
0.451904296875,
0.468017578125,
0.48486328125,
0.096435546875,
0.4697265625
],
[
0.295654296875,
0.439453125,
0.4345703125,
0.439453125,
0.4345703125,
0.453857421875,
0.295654296875,
0.453857421875
],
[
0.5146484375,
0.443115234375,
0.9111328125,
0.443115234375,
0.9111328125,
0.45703125,
0.5146484375,
0.45703125
],
[
0.5146484375,
0.45703125,
0.9111328125,
0.45703125,
0.9111328125,
0.47021484375,
0.5146484375,
0.47021484375
],
[
0.455078125,
0.472900390625,
0.483642578125,
0.472900390625,
0.483642578125,
0.484130859375,
0.455078125,
0.484130859375
],
[
0.51123046875,
0.4697265625,
0.91259765625,
0.47021484375,
0.91259765625,
0.48486328125,
0.51123046875,
0.484130859375
],
[
0.098876953125,
0.48486328125,
0.48779296875,
0.4873046875,
0.48779296875,
0.5048828125,
0.098876953125,
0.50244140625
],
[
0.51123046875,
0.48486328125,
0.9111328125,
0.48486328125,
0.9111328125,
0.498779296875,
0.51123046875,
0.498779296875
],
[
0.5146484375,
0.498779296875,
0.91259765625,
0.498779296875,
0.91259765625,
0.51318359375,
0.5146484375,
0.51318359375
],
[
0.0882568359375,
0.5048828125,
0.483642578125,
0.5048828125,
0.483642578125,
0.51806640625,
0.0882568359375,
0.51806640625
],
[
0.51318359375,
0.51220703125,
0.6806640625,
0.51123046875,
0.6806640625,
0.52587890625,
0.51318359375,
0.5263671875
],
[
0.0882568359375,
0.51806640625,
0.486083984375,
0.51806640625,
0.486083984375,
0.53271484375,
0.0882568359375,
0.53271484375
],
[
0.529296875,
0.52587890625,
0.9111328125,
0.52587890625,
0.9111328125,
0.54052734375,
0.529296875,
0.54052734375
],
[
0.0882568359375,
0.5322265625,
0.2418212890625,
0.5322265625,
0.2418212890625,
0.54541015625,
0.0882568359375,
0.54541015625
],
[
0.51708984375,
0.5390625,
0.9111328125,
0.5390625,
0.9111328125,
0.5537109375,
0.51708984375,
0.5537109375
],
[
0.26318359375,
0.5498046875,
0.39794921875,
0.5517578125,
0.39794921875,
0.5693359375,
0.26318359375,
0.5673828125
],
[
0.5146484375,
0.5537109375,
0.6787109375,
0.5537109375,
0.6787109375,
0.5673828125,
0.5146484375,
0.5673828125
],
[
0.17236328125,
0.56298828125,
0.273681640625,
0.56298828125,
0.273681640625,
0.57470703125,
0.17236328125,
0.57470703125
],
[
0.453369140625,
0.56103515625,
0.486083984375,
0.56103515625,
0.486083984375,
0.5771484375,
0.453369140625,
0.5771484375
],
[
0.313720703125,
0.5673828125,
0.348876953125,
0.57080078125,
0.346435546875,
0.58544921875,
0.310546875,
0.58203125
],
[
0.5146484375,
0.57763671875,
0.56201171875,
0.57763671875,
0.56201171875,
0.5888671875,
0.5146484375,
0.5888671875
],
[
0.5556640625,
0.57763671875,
0.71728515625,
0.57763671875,
0.71728515625,
0.59228515625,
0.5556640625,
0.59228515625
],
[
0.0882568359375,
0.59130859375,
0.486083984375,
0.59130859375,
0.486083984375,
0.60498046875,
0.0882568359375,
0.60498046875
],
[
0.51318359375,
0.59521484375,
0.66845703125,
0.59521484375,
0.66845703125,
0.609375,
0.51318359375,
0.609375
],
[
0.0882568359375,
0.60498046875,
0.486083984375,
0.60498046875,
0.486083984375,
0.61865234375,
0.0882568359375,
0.61865234375
],
[
0.0849609375,
0.6171875,
0.250732421875,
0.6171875,
0.250732421875,
0.630859375,
0.0849609375,
0.630859375
],
[
0.52685546875,
0.6123046875,
0.9111328125,
0.6123046875,
0.9111328125,
0.62548828125,
0.52685546875,
0.62548828125
],
[
0.10455322265625,
0.630859375,
0.486083984375,
0.6318359375,
0.486083984375,
0.646484375,
0.10455322265625,
0.64599609375
],
[
0.5439453125,
0.626953125,
0.91259765625,
0.626953125,
0.91259765625,
0.64013671875,
0.5439453125,
0.64013671875
],
[
0.5439453125,
0.64013671875,
0.91259765625,
0.64013671875,
0.91259765625,
0.65380859375,
0.5439453125,
0.65380859375
],
[
0.098876953125,
0.646484375,
0.486083984375,
0.646484375,
0.486083984375,
0.6611328125,
0.098876953125,
0.6611328125
],
[
0.0882568359375,
0.66015625,
0.486083984375,
0.66015625,
0.486083984375,
0.6748046875,
0.0882568359375,
0.6748046875
],
[
0.5439453125,
0.65478515625,
0.91259765625,
0.65478515625,
0.91259765625,
0.66845703125,
0.5439453125,
0.66845703125
],
[
0.08660888671875,
0.67236328125,
0.330810546875,
0.67236328125,
0.330810546875,
0.68603515625,
0.08660888671875,
0.68603515625
],
[
0.5439453125,
0.66845703125,
0.65625,
0.66845703125,
0.65625,
0.681640625,
0.5439453125,
0.681640625
],
[
0.52783203125,
0.68603515625,
0.91015625,
0.68603515625,
0.91015625,
0.69970703125,
0.52783203125,
0.69970703125
],
[
0.08660888671875,
0.7041015625,
0.1094970703125,
0.7041015625,
0.1094970703125,
0.71533203125,
0.08660888671875,
0.71533203125
],
[
0.1192626953125,
0.7041015625,
0.2303466796875,
0.7041015625,
0.2303466796875,
0.71728515625,
0.1192626953125,
0.71728515625
],
[
0.5439453125,
0.70068359375,
0.9140625,
0.70068359375,
0.9140625,
0.7138671875,
0.5439453125,
0.7138671875
],
[
0.11517333984375,
0.70703125,
0.1192626953125,
0.70703125,
0.1192626953125,
0.7119140625,
0.11517333984375,
0.7119140625
],
[
0.5458984375,
0.7138671875,
0.9140625,
0.7138671875,
0.9140625,
0.72802734375,
0.5458984375,
0.72802734375
],
[
0.0882568359375,
0.72265625,
0.486083984375,
0.72265625,
0.486083984375,
0.736328125,
0.0882568359375,
0.736328125
],
[
0.0882568359375,
0.73486328125,
0.483642578125,
0.73486328125,
0.483642578125,
0.74853515625,
0.0882568359375,
0.74853515625
],
[
0.5458984375,
0.72900390625,
0.9140625,
0.72900390625,
0.9140625,
0.7421875,
0.5458984375,
0.7421875
],
[
0.54345703125,
0.7392578125,
0.9150390625,
0.7412109375,
0.9150390625,
0.7587890625,
0.54345703125,
0.7568359375
],
[
0.08660888671875,
0.75,
0.486083984375,
0.75,
0.486083984375,
0.7646484375,
0.08660888671875,
0.7646484375
],
[
0.5458984375,
0.755859375,
0.82666015625,
0.755859375,
0.82666015625,
0.77099609375,
0.5458984375,
0.77099609375
],
[
0.0882568359375,
0.76318359375,
0.484375,
0.76318359375,
0.484375,
0.7763671875,
0.0882568359375,
0.7763671875
],
[
0.08660888671875,
0.7763671875,
0.4248046875,
0.77783203125,
0.4248046875,
0.79248046875,
0.08660888671875,
0.79150390625
],
[
0.54345703125,
0.77294921875,
0.9140625,
0.7734375,
0.9140625,
0.78857421875,
0.54345703125,
0.787109375
],
[
0.54345703125,
0.7861328125,
0.9140625,
0.787109375,
0.9140625,
0.8017578125,
0.54345703125,
0.80029296875
],
[
0.08660888671875,
0.80322265625,
0.12091064453125,
0.80322265625,
0.12091064453125,
0.814453125,
0.08660888671875,
0.814453125
],
[
0.11358642578125,
0.80322265625,
0.200927734375,
0.80322265625,
0.200927734375,
0.814453125,
0.11358642578125,
0.814453125
],
[
0.54345703125,
0.80029296875,
0.91259765625,
0.7998046875,
0.91259765625,
0.814453125,
0.54345703125,
0.81494140625
],
[
0.096435546875,
0.8193359375,
0.483642578125,
0.8193359375,
0.483642578125,
0.83349609375,
0.096435546875,
0.83349609375
],
[
0.5458984375,
0.8173828125,
0.9111328125,
0.8173828125,
0.9111328125,
0.82763671875,
0.5458984375,
0.82763671875
],
[
0.08660888671875,
0.83349609375,
0.484375,
0.83349609375,
0.484375,
0.8466796875,
0.08660888671875,
0.8466796875
],
[
0.5458984375,
0.830078125,
0.91259765625,
0.830078125,
0.91259765625,
0.84326171875,
0.5458984375,
0.84326171875
],
[
0.0882568359375,
0.84765625,
0.484375,
0.84765625,
0.484375,
0.861328125,
0.0882568359375,
0.861328125
],
[
0.5458984375,
0.84326171875,
0.91015625,
0.84326171875,
0.91015625,
0.85693359375,
0.5458984375,
0.85693359375
],
[
0.0849609375,
0.85986328125,
0.48779296875,
0.861328125,
0.48779296875,
0.87548828125,
0.0849609375,
0.875
],
[
0.52685546875,
0.861328125,
0.9111328125,
0.861328125,
0.9111328125,
0.875,
0.52685546875,
0.875
],
[
0.08660888671875,
0.87548828125,
0.484375,
0.87548828125,
0.484375,
0.8896484375,
0.08660888671875,
0.8896484375
],
[
0.54345703125,
0.87548828125,
0.91259765625,
0.87548828125,
0.91259765625,
0.8896484375,
0.54345703125,
0.8896484375
]
] |
[
"Algorithm 1 Training of the Student Network",
"Dev",
"Test",
"Dataset",
"# OOV",
"OOV Rate",
"#O0V",
"OOVRate",
"1: Input: the teacher network T , training dataset D",
"",
"2: Result: the student network S",
"POS",
"PTB-English",
"8,392",
"6.37%",
"7,528",
"5.81%",
"RIT-English",
"3: while S does not converge do",
"774",
"34.52%",
"760",
"sample X = {wi, ..., wn] and its corresponding",
"33.17%",
"4:",
"GSD-Russian",
"21,323",
"17.96%",
"21,523",
"18.31%",
"Y = (yi,.--, Yn) from D",
"RRT-Romanian",
"3,965",
"23.22%",
"3,702",
"22.67%",
"5:",
"for ie[1,... ,n] do",
"NER",
"6:",
"Generate v(wi; X) for wi e X ;",
"CoNLL02-Spanish",
"2,216",
"50.91%",
"1,544",
"43.38%",
"7:",
"v(X) = [e(wi), ..., v(w;|X),.. ,e(wn)) ;",
"CoNLL02-Dutch",
"1,819",
"69.53%",
"2,564",
"65.05%",
"Twitter-English",
"1,266",
"79.15 %",
"4,131",
"79.13 %",
"Get Ls(X, Y;Os, T) based on v(X) accord-",
"8:",
"CoNLL03-German",
"3.928",
"81.27 %",
"2.685",
"73.10 %",
"Update Os Os - a@Cs(XY:0s.0r)",
"ing to Eq. (16);",
"9:",
"00",
"Table l: Number of OOV words (for POS) and entities",
"(forNER)inthedevelopment and testing sets,when",
"treating words occurring less than 5 times in the training",
"在这项工作中，我们建议将学生网络的培训与教师网",
"set as OOV. An entity is treated as OOV if it contains",
"络的培训对象直接联系起来，从而避免设计适当的辅助",
"at least one OOV word.",
"培训标准。不失一般性，我们首先考虑在给定句子中只有",
"个OOV单词的情况下对学生网络进行培训。为了模拟",
"这种情况，对于训练句子X及其相应的标签序列Y，我",
"于培训，第22节用于验证，第23节用于测试。对于",
"们从X中随机抽取一个单词wi。结果对（wi,X/z,Y）构",
"RIT-English，我们遵循Gui等人的拆分协议。，[2017]",
"成了学生网络的训练示例。通过将e(X)的ith列替换为",
"而对于UD-俄语和UD-罗马尼亚语，我们使用了它们",
"v(wX;0s）=[vw(wX;9s)④ec(w)]，我们获得了教师网",
"给定的数据分割",
"络的输人 v(X) = e(wi), ..,v(wiX;0s),...,e(wn)) 。",
"NER：对于NER，我们在以下方面进行了实验：（1）",
"基于(X），我们为每个单词h(wiX)获得新的隐藏表",
"CoNLL02-Spanish：CoNLL2002西班牙语NER共享",
"示，然后将任务损失定义为：",
"任务数据集[Sang，2002]；（2）CoNLL02-Dutch：荷",
"兰语的CoNLL2002数据集；（3）Twitter-English：从",
"s(,Ys,r)=(u",
"IIf=i Φt(yt-1, yt/X)",
"TweetsXMATHXTHMA2创建的英语N。这些数据集",
"由四种类型注释：PER，LOC，ORG和MISC。对于除",
"(16)",
"Twitter-English之外的数据集，我们使用官方的拆分训",
"其中这次是Φt(yylx)=exp(w,h(wt)+by\"）。请",
"练集进行模型训练，使用testa进行验证，使用testb进",
"行测试。而对于Twitter-English，我们遵循了Zhang等",
"注意，学生网络的培训损失与教师网络的培训损失相同",
"人的数据拆分。，[2018]",
"确保他们的培训兼容。为了训练学生网络，我们通过以",
"表1报告了OOV问题在每个数据集的开发和测试集",
"下方式进行参数更新：",
"上的统计结果。从表中，我们可以看到OOV率在不同",
"OLs(X, Y;Os,OT)",
"的数据集上有很大差异",
"0s s-a-",
"(17)",
"aes",
"4.2",
"Compared Methods",
"在训练学生网络期间，固定了教师网络QT的参数。算",
"我们比较了以下基线：",
"法1显示了在句子中只有一个OOV单词的情况下，学",
"生网络的一般训练过程",
": RandomUNK : This baseline refers to the teacher",
"要将此过程扩展到多个OOV单词的情况，我们从X",
"network trained on all words occurring in the",
"training set. At testing time, it represents words",
"e中采样多个单词。g。，wi和wi，生成学生网络的训",
"练示例。在此示例中定义的损耗类似于在一个OOV情",
"not occurring in the training set with a consistent",
"况下的损耗，并且在s上最小化",
"random vector.",
". SingleUNK : This baseline trains the teacher net-",
"4",
"Experiments",
"work on words that occur no less than 5 times in",
"",
"the training set. The other infrequent words and",
"为了评估我们提出的方法的有效性，我们在四个词性标",
"记（POS）任务和四个命名实体识别（NER）任务上进",
"those words not occurring in the training set are",
"all mapped to a single trainable embedding euNk ,",
"行了实验。这些任务具有不同的OOV率，这是由训练",
"which is trained during model training.",
"集中出现少于五次的测试单词的百分比定义的。这些任",
"务共享与S3中所示相同的教师和学生网络架构",
"[Lazaridou et al.,2017] :This baseline uses",
"RandomUNK as the teacher network.At testing",
"4.1",
"Datasets",
"time, it first represents OOv words with their",
"POS：对于POS，我们进行了以下实验：（1）PTB-",
"context word representations by mean-pooling as",
"English：英文PennTreebank数据集[Marcusetal..",
"defined in equation (5) based on RandomUNK and",
"1993）的华尔街日报部分，（2）RIT-English：从英文",
"then performs label prediction using RandomUNK.",
"Tweets创建的数据集[Derczynskietal.，2013]，（3）。",
": [Khodak et al., 2018] : This baseline additionally",
"对于PTB-English，我们遵循标准拆分：第2-21节用",
"trains a linear transformer A to transform the"
] |
{
"arxiv_id": "1905.12277",
"language": "zh",
"page_idx": 3
}
|
||
1905.12277-zh-4
|
1905.12277-zh-4.jpg
|
mean-pooled representation of [Lazaridou et al., $2017]\bar{\text{ as illustrated by equation }}(2)$for predicting representations of OOV words. The training of A is performed on words that occur no less than 5 times ${\mathrm{in~the~training~set.}}$
$\bullet$[Schick and Schütze, 2018]:This baseline is a
variant of the proposed method, but its student network is trained on the reconstruction object as defined in equation ( 15 ). It is also related to the work of [Schick and Schütze, 2018] but using the same architecture of the proposed method to model word context and surface-form.
$\bullet$ [Akbik et al., 2018] : This baseline converts the
input sentence into a character sequence. Then, it applies a character language model to the character sequence to get the representation of every word within the sentence. Based on the obtained word representations, it applies a LSTM network to model the word sequence and performs fnal tag recommendation.
To make it compatible with
the setting of this work, we did not pre-train the language model on external data and not use pre- $\bar{\text{trained static word embeddings for this baseline.}}$
## 4.3 Implementation Detail
对于数据预处理，所有数字均替换为特殊令牌“<NUM $>”$,所有 URL 均替换为特殊令牌“<URL>”。教师和学生网络的单词嵌入、字符嵌人和 LSTM 的维度分别设置为 50、16 和 50。对于内核宽度 3 和 5, 字符 CNN 的内核大小设置为25。使用 Adam 步长规则[Kinga and Adam, 2015]进行优化，学习率设置为 1e-3
## 4.4 Evaluation
我们将测试(或开发)集分为两个子集：词汇内(WIV) 单词和词汇外(OOV) 单词。如果一个词在训练集中出现超过 5 次，则将其视为 WIV, 否则为 OOV。对于 NER 如果一个实体至少包含一个 OOV 词集的单词，则该实体被认为是 OOV 的。我们在 OOV 集上报告模型性能(POS 和 NER 的 F1 的准确性)。这是因为我们可以准确地说出一个词属于哪个子集，因此我们可以很容易地在每个子集上组合最佳模型，以在整个测试集上实现最佳的整体性能。例如，我们可以使用我们提出的模型对OOV 单词进行标签预测，并使用 WIV 子集上性能最好的模型对 WIV 单词进行标签预测
## 4.5 Main Results
表 2 和 3 报告了在使用 BiLSTM 和 CNN 实现教师网络时，分别为 POS 标记和 NER 设置的 OOV 上的模型性能。从该表，我们有以下观察：(1)在大多数任务中，处理 OOV 问题的方法优于 RandomUNK 基线。这验证了在序列标记中处理 OOV 问题的必要性。(2) 使用表面形式和上下文信息生成 OOV 单词表示的方法 [Schick and Schütze, 2018] 优于仅使用大多数数据集上的上下文信息的方法 [Khodak et al., 2018] 。这显示了表面形式和上下文信息的互补性；(3)师生范式[Schick and Schütze, 2018]中最具可比性的基线总体上优于 SingleUNK。这验证了师生范式下动机的有效性；(4)我们提出的方法始终优于[Schick and Schütze, 2018],这与提出的方法在
学生网络的训练中有所不同。这一方面表明了从学生网络到教师网络的错误传播的存在，另一方面也证实了我们提出的解决 OOV 问题的方法的有效性；最后 (5)我们提出的方法始终优于字符语言模型[Akbik et al., 2018] 。这个结果的一个可能的解释是，我们的方法可以使用单词级嵌人而不会遭受 OOV 问题，而字符语言模型不能
5 Conclusion
在这项工作中，我们提出了一种新颖的方法，仅使用任务的训练数据来解决序列标记系统中的词汇外问题。它旨在从 OOV 单词的表面形式和上下文中生成 OOV 单词的表示形式。此外，它的设计是为了避免在同一范式中现有方法遭受的错误传播问题。关于 POS 标记(POS) 和命名实体识别 (NER)的广泛实验研究表明，该方法可以在 OOV 问题上比现有方法取得更好或可比的性能。
## 6 Acknowledgements
作者希望感谢匿名审稿人的有益评论。ThisworkwaspartiallyfundedbyChinaNationalKeyR&DProgram(No。2018YFC0831105,2017YFB1002104,2018YFB1005104), 国家自然科学基金(No。61532011,61751201),上海市科技重大专项(No。2018SHZDZX01),STCSM(编号。16JC1420401,17JC1420200),ZJLab
## References
[Akbik et al., 2018] Alan Akbik, Duncan Blythe, and
Roland Vollgra$\dot{f}$.
Contextual_string embeddings
for sequence labeling. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1638-1649, 2018.
[Barbu Mititelu et al., 2016] V Barbu Mititelu, R Ion,
The
R Simionescu, E Irimia, and CA Perez.
romanian treebank annotated according to universal dependencies. In Proceedings of The Tenth International Conference on Natural Language Processing (HrTAL2016),2016.
[Bojarski et al., 2016] Mariusz
$\operatorname{Bojarski}$,
$\underline{\mathrm{Davide}}$
$\begin{array}{ll}{\mathrm{Del}}&{\mathrm{Testa,~Daniel~Dworakowski,~Bernhard~Firner,~}}\\{\mathrm{Del~Testa,~Daniel~Dworakowski,~Bernhard~Firner,~}}\\{\mathrm{Del~Testa,~Daniel~Dworakowski,~Bernhard~Firmer,~}}\end{array}$ Beat Flepp, Prasoon Goyal, Lawrence D Jackel, Mathew $\hat \text{Monfort, Urs Muller, Jiakai Zhang, et al}.$ End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316,2016.
[Caselli et al., 2015] Tommaso Caselli, Piek Vossen,
Marieke van Erp, Antske Fokkens, Filip Ilievski, Rubén Izquierdo, Minh Le, Roser Morante, and Marten Postma. When it's all piling up: investigating error propagation in an nlp pipeline. In WNÄCP@ NLD$\dot{B}$, 2$\dot{0}$15.
[Chen et al.,2015] Xinchi Chen, Xipeng Qiu, Chenxi Zhu, Pengfei Liu, and Xuanjing Huang. Long short-term memory neural networks for chinese word segmentation. In EMNLP, pages 1197-1206, 2015.
[Deleger et al., 2016] Louise Deleger, Robert Bossy, $\bar{\mathrm{Estelle}}$ Chaix, Mouhamadou Ba, Arnaud Ferrè, Philippe Bessieres, and Claire Nedellec. Overview of the bacteria biotope task at bionlp shared task 2016. In Proceedings $\dot\text{of the 4th BioNL\hat{P}}$ Shared Task Workshop, pages $1\bar{2}-22,2016.$
|
[
[
0.11767578125,
0.0694580078125,
0.486083984375,
0.06817626953125,
0.486083984375,
0.08331298828125,
0.11767578125,
0.083984375
],
[
0.51123046875,
0.06817626953125,
0.9111328125,
0.0694580078125,
0.9111328125,
0.083984375,
0.51123046875,
0.08331298828125
],
[
0.1192626953125,
0.083984375,
0.486083984375,
0.083984375,
0.486083984375,
0.09783935546875,
0.1192626953125,
0.09783935546875
],
[
0.51123046875,
0.08209228515625,
0.9111328125,
0.08331298828125,
0.9111328125,
0.09783935546875,
0.51123046875,
0.09661865234375
],
[
0.11767578125,
0.09783935546875,
0.48779296875,
0.09661865234375,
0.48779296875,
0.111083984375,
0.11767578125,
0.11236572265625
],
[
0.5146484375,
0.09783935546875,
0.91259765625,
0.09783935546875,
0.91259765625,
0.111083984375,
0.5146484375,
0.111083984375
],
[
0.1192626953125,
0.11236572265625,
0.488525390625,
0.11236572265625,
0.488525390625,
0.1256103515625,
0.1192626953125,
0.1256103515625
],
[
0.51318359375,
0.11236572265625,
0.9111328125,
0.11236572265625,
0.9111328125,
0.1256103515625,
0.51318359375,
0.1256103515625
],
[
0.11767578125,
0.1256103515625,
0.2548828125,
0.1256103515625,
0.2548828125,
0.1395263671875,
0.11767578125,
0.1395263671875
],
[
0.51318359375,
0.1256103515625,
0.9111328125,
0.1256103515625,
0.9111328125,
0.1395263671875,
0.51318359375,
0.1395263671875
],
[
0.11517333984375,
0.1427001953125,
0.484375,
0.1427001953125,
0.484375,
0.1558837890625,
0.11517333984375,
0.1558837890625
],
[
0.5146484375,
0.1395263671875,
0.904296875,
0.1395263671875,
0.904296875,
0.15283203125,
0.5146484375,
0.15283203125
],
[
0.10455322265625,
0.146484375,
0.12091064453125,
0.146484375,
0.12091064453125,
0.154052734375,
0.10455322265625,
0.154052734375
],
[
0.11767578125,
0.1578369140625,
0.486083984375,
0.1578369140625,
0.486083984375,
0.1717529296875,
0.11767578125,
0.1717529296875
],
[
0.5146484375,
0.17041015625,
0.52685546875,
0.17041015625,
0.52685546875,
0.1787109375,
0.5146484375,
0.1787109375
],
[
0.54833984375,
0.169189453125,
0.640625,
0.169189453125,
0.640625,
0.18115234375,
0.54833984375,
0.18115234375
],
[
0.53369140625,
0.17236328125,
0.54345703125,
0.17236328125,
0.54345703125,
0.1767578125,
0.53369140625,
0.1767578125
],
[
0.1192626953125,
0.1717529296875,
0.48779296875,
0.1717529296875,
0.48779296875,
0.1849365234375,
0.1192626953125,
0.1849365234375
],
[
0.12091064453125,
0.1875,
0.484375,
0.1875,
0.484375,
0.1976318359375,
0.12091064453125,
0.1976318359375
],
[
0.515625,
0.1881103515625,
0.9111328125,
0.1881103515625,
0.9111328125,
0.2020263671875,
0.515625,
0.2020263671875
],
[
0.1192626953125,
0.199462890625,
0.486083984375,
0.199462890625,
0.486083984375,
0.21337890625,
0.1192626953125,
0.21337890625
],
[
0.5146484375,
0.2008056640625,
0.900390625,
0.2008056640625,
0.900390625,
0.21533203125,
0.5146484375,
0.21533203125
],
[
0.90185546875,
0.20263671875,
0.90869140625,
0.2083740234375,
0.90185546875,
0.21337890625,
0.89453125,
0.2083740234375
],
[
0.1192626953125,
0.21337890625,
0.486083984375,
0.21337890625,
0.486083984375,
0.2279052734375,
0.1192626953125,
0.2279052734375
],
[
0.515625,
0.21533203125,
0.91259765625,
0.21533203125,
0.91259765625,
0.2291259765625,
0.515625,
0.2291259765625
],
[
0.1192626953125,
0.2279052734375,
0.343017578125,
0.2279052734375,
0.343017578125,
0.2412109375,
0.1192626953125,
0.2412109375
],
[
0.51318359375,
0.2279052734375,
0.9111328125,
0.2298583984375,
0.9111328125,
0.24365234375,
0.51318359375,
0.2412109375
],
[
0.1168212890625,
0.246826171875,
0.483642578125,
0.246826171875,
0.483642578125,
0.2568359375,
0.1168212890625,
0.2568359375
],
[
0.5146484375,
0.242431640625,
0.91015625,
0.242431640625,
0.91015625,
0.2568359375,
0.5146484375,
0.2568359375
],
[
0.5146484375,
0.2568359375,
0.91015625,
0.2568359375,
0.91015625,
0.270751953125,
0.5146484375,
0.270751953125
],
[
0.10540771484375,
0.2474365234375,
0.12091064453125,
0.2474365234375,
0.12091064453125,
0.255126953125,
0.10540771484375,
0.255126953125
],
[
0.11767578125,
0.2587890625,
0.486083984375,
0.2587890625,
0.486083984375,
0.272705078125,
0.11767578125,
0.272705078125
],
[
0.11767578125,
0.2734375,
0.484375,
0.2734375,
0.484375,
0.287353515625,
0.11767578125,
0.287353515625
],
[
0.5146484375,
0.271484375,
0.91015625,
0.271484375,
0.91015625,
0.285400390625,
0.5146484375,
0.285400390625
],
[
0.1192626953125,
0.2890625,
0.483642578125,
0.2890625,
0.483642578125,
0.2998046875,
0.1192626953125,
0.2998046875
],
[
0.1192626953125,
0.300537109375,
0.484375,
0.300537109375,
0.484375,
0.314453125,
0.1192626953125,
0.314453125
],
[
0.529296875,
0.3017578125,
0.5439453125,
0.3017578125,
0.5439453125,
0.31005859375,
0.529296875,
0.31005859375
],
[
0.5458984375,
0.298583984375,
0.70751953125,
0.2998046875,
0.70751953125,
0.314453125,
0.5458984375,
0.313232421875
],
[
0.11767578125,
0.31494140625,
0.486083984375,
0.31494140625,
0.486083984375,
0.328857421875,
0.11767578125,
0.328857421875
],
[
0.515625,
0.302978515625,
0.525390625,
0.302978515625,
0.525390625,
0.31005859375,
0.515625,
0.31005859375
],
[
0.5146484375,
0.3193359375,
0.90869140625,
0.3193359375,
0.90869140625,
0.333251953125,
0.5146484375,
0.333251953125
],
[
0.1168212890625,
0.32763671875,
0.486083984375,
0.328857421875,
0.486083984375,
0.343505859375,
0.1168212890625,
0.34228515625
],
[
0.51318359375,
0.33203125,
0.90625,
0.333251953125,
0.90625,
0.347900390625,
0.51318359375,
0.3466796875
],
[
0.1192626953125,
0.3447265625,
0.25,
0.3447265625,
0.25,
0.354736328125,
0.1192626953125,
0.354736328125
],
[
0.25,
0.34228515625,
0.486083984375,
0.34228515625,
0.486083984375,
0.35595703125,
0.25,
0.35595703125
],
[
0.51123046875,
0.34521484375,
0.904296875,
0.3466796875,
0.904296875,
0.361083984375,
0.51123046875,
0.35986328125
],
[
0.11767578125,
0.356689453125,
0.48779296875,
0.356689453125,
0.48779296875,
0.37060546875,
0.11767578125,
0.37060546875
],
[
0.515625,
0.361083984375,
0.9111328125,
0.361083984375,
0.9111328125,
0.375,
0.515625,
0.375
],
[
0.1192626953125,
0.37060546875,
0.486083984375,
0.37060546875,
0.486083984375,
0.385009765625,
0.1192626953125,
0.385009765625
],
[
0.5146484375,
0.375,
0.91259765625,
0.375,
0.91259765625,
0.38818359375,
0.5146484375,
0.38818359375
],
[
0.11767578125,
0.385009765625,
0.464111328125,
0.385009765625,
0.464111328125,
0.395751953125,
0.11767578125,
0.395751953125
],
[
0.5146484375,
0.38818359375,
0.7646484375,
0.38818359375,
0.7646484375,
0.401611328125,
0.5146484375,
0.401611328125
],
[
0.0882568359375,
0.41357421875,
0.12091064453125,
0.41357421875,
0.12091064453125,
0.4248046875,
0.0882568359375,
0.4248046875
],
[
0.1168212890625,
0.4140625,
0.131591796875,
0.4140625,
0.131591796875,
0.42236328125,
0.1168212890625,
0.42236328125
],
[
0.1258544921875,
0.41357421875,
0.310546875,
0.41357421875,
0.310546875,
0.4267578125,
0.1258544921875,
0.4267578125
],
[
0.51318359375,
0.415283203125,
0.6044921875,
0.417236328125,
0.6044921875,
0.431884765625,
0.51318359375,
0.429931640625
],
[
0.0882568359375,
0.429931640625,
0.484375,
0.429931640625,
0.484375,
0.4443359375,
0.0882568359375,
0.4443359375
],
[
0.51318359375,
0.434326171875,
0.9111328125,
0.43505859375,
0.9111328125,
0.449462890625,
0.51318359375,
0.448974609375
],
[
0.09722900390625,
0.4443359375,
0.484375,
0.4443359375,
0.484375,
0.458251953125,
0.09722900390625,
0.458251953125
],
[
0.529296875,
0.448974609375,
0.66015625,
0.448974609375,
0.66015625,
0.462158203125,
0.529296875,
0.462158203125
],
[
0.67626953125,
0.447509765625,
0.91259765625,
0.448974609375,
0.91259765625,
0.46337890625,
0.67626953125,
0.462158203125
],
[
0.0882568359375,
0.458251953125,
0.486083984375,
0.458251953125,
0.486083984375,
0.472900390625,
0.0882568359375,
0.472900390625
],
[
0.08905029296875,
0.451904296875,
0.09722900390625,
0.448974609375,
0.1004638671875,
0.451904296875,
0.09234619140625,
0.455078125
],
[
0.52783203125,
0.462158203125,
0.70751953125,
0.462158203125,
0.70751953125,
0.47607421875,
0.52783203125,
0.47607421875
],
[
0.7294921875,
0.462158203125,
0.91259765625,
0.462158203125,
0.91259765625,
0.47607421875,
0.7294921875,
0.47607421875
],
[
0.0882568359375,
0.472900390625,
0.484375,
0.472900390625,
0.484375,
0.486083984375,
0.0882568359375,
0.486083984375
],
[
0.529296875,
0.474853515625,
0.91259765625,
0.474853515625,
0.91259765625,
0.488037109375,
0.529296875,
0.488037109375
],
[
0.08660888671875,
0.48486328125,
0.486083984375,
0.486083984375,
0.486083984375,
0.50048828125,
0.08660888671875,
0.5
],
[
0.529296875,
0.4873046875,
0.78515625,
0.4873046875,
0.78515625,
0.50048828125,
0.529296875,
0.50048828125
],
[
0.08660888671875,
0.50048828125,
0.39794921875,
0.50048828125,
0.39794921875,
0.5146484375,
0.08660888671875,
0.5146484375
],
[
0.5146484375,
0.50390625,
0.9111328125,
0.505859375,
0.9111328125,
0.52099609375,
0.5146484375,
0.51806640625
],
[
0.52783203125,
0.51806640625,
0.64453125,
0.52099609375,
0.64453125,
0.5322265625,
0.52783203125,
0.52978515625
],
[
0.64892578125,
0.51806640625,
0.85888671875,
0.51806640625,
0.85888671875,
0.53271484375,
0.64892578125,
0.53271484375
],
[
0.08660888671875,
0.5244140625,
0.22216796875,
0.5244140625,
0.22216796875,
0.53857421875,
0.08660888671875,
0.53857421875
],
[
0.88134765625,
0.51953125,
0.9111328125,
0.51953125,
0.9111328125,
0.53076171875,
0.88134765625,
0.53076171875
],
[
0.529296875,
0.5322265625,
0.91259765625,
0.5322265625,
0.91259765625,
0.54541015625,
0.529296875,
0.54541015625
],
[
0.08660888671875,
0.54150390625,
0.484375,
0.54150390625,
0.484375,
0.5546875,
0.08660888671875,
0.5546875
],
[
0.529296875,
0.54443359375,
0.64453125,
0.54443359375,
0.64453125,
0.55810546875,
0.529296875,
0.55810546875
],
[
0.64208984375,
0.54443359375,
0.91015625,
0.54443359375,
0.91015625,
0.55810546875,
0.64208984375,
0.55810546875
],
[
0.0882568359375,
0.556640625,
0.484375,
0.556640625,
0.484375,
0.57080078125,
0.0882568359375,
0.57080078125
],
[
0.529296875,
0.556640625,
0.9140625,
0.556640625,
0.9140625,
0.57080078125,
0.529296875,
0.57080078125
],
[
0.08660888671875,
0.57080078125,
0.484375,
0.57080078125,
0.484375,
0.583984375,
0.08660888671875,
0.583984375
],
[
0.53125,
0.5693359375,
0.6748046875,
0.5693359375,
0.6748046875,
0.58349609375,
0.53125,
0.58349609375
],
[
0.0882568359375,
0.58349609375,
0.484375,
0.58349609375,
0.484375,
0.5966796875,
0.0882568359375,
0.5966796875
],
[
0.515625,
0.58837890625,
0.736328125,
0.58837890625,
0.736328125,
0.60302734375,
0.515625,
0.60302734375
],
[
0.7587890625,
0.59033203125,
0.826171875,
0.59033203125,
0.826171875,
0.6015625,
0.7587890625,
0.6015625
],
[
0.86279296875,
0.59033203125,
0.9111328125,
0.59033203125,
0.9111328125,
0.6015625,
0.86279296875,
0.6015625
],
[
0.0882568359375,
0.59765625,
0.486083984375,
0.59765625,
0.486083984375,
0.611328125,
0.0882568359375,
0.611328125
],
[
0.529296875,
0.6015625,
0.85546875,
0.6015625,
0.85546875,
0.61572265625,
0.529296875,
0.61572265625
],
[
0.85302734375,
0.6015625,
0.91015625,
0.60400390625,
0.91015625,
0.61572265625,
0.85302734375,
0.61279296875
],
[
0.0882568359375,
0.61279296875,
0.483642578125,
0.61279296875,
0.483642578125,
0.62353515625,
0.0882568359375,
0.62353515625
],
[
0.529296875,
0.61279296875,
0.859375,
0.6142578125,
0.859375,
0.62890625,
0.529296875,
0.62744140625
],
[
0.85302734375,
0.6162109375,
0.91015625,
0.6162109375,
0.91015625,
0.62744140625,
0.85302734375,
0.62744140625
],
[
0.0882568359375,
0.62548828125,
0.486083984375,
0.62548828125,
0.486083984375,
0.6396484375,
0.0882568359375,
0.6396484375
],
[
0.529296875,
0.62744140625,
0.91259765625,
0.62744140625,
0.91259765625,
0.642578125,
0.529296875,
0.642578125
],
[
0.0882568359375,
0.63818359375,
0.484375,
0.63818359375,
0.484375,
0.6513671875,
0.0882568359375,
0.6513671875
],
[
0.529296875,
0.64013671875,
0.85302734375,
0.64013671875,
0.85302734375,
0.65380859375,
0.529296875,
0.65380859375
],
[
0.869140625,
0.6416015625,
0.9111328125,
0.6416015625,
0.9111328125,
0.65283203125,
0.869140625,
0.65283203125
],
[
0.0882568359375,
0.65380859375,
0.484375,
0.65380859375,
0.484375,
0.66748046875,
0.0882568359375,
0.66748046875
],
[
0.529296875,
0.65380859375,
0.76318359375,
0.65283203125,
0.76318359375,
0.6640625,
0.529296875,
0.66552734375
],
[
0.08660888671875,
0.66748046875,
0.484375,
0.66748046875,
0.484375,
0.68115234375,
0.08660888671875,
0.68115234375
],
[
0.515625,
0.671875,
0.81396484375,
0.671875,
0.81396484375,
0.68505859375,
0.515625,
0.68505859375
],
[
0.8095703125,
0.67236328125,
0.85107421875,
0.67236328125,
0.85107421875,
0.6845703125,
0.8095703125,
0.6845703125
],
[
0.8486328125,
0.671875,
0.91015625,
0.6748046875,
0.90869140625,
0.68603515625,
0.84716796875,
0.68310546875
],
[
0.08660888671875,
0.68115234375,
0.328369140625,
0.68115234375,
0.328369140625,
0.6943359375,
0.08660888671875,
0.6943359375
],
[
0.529296875,
0.6845703125,
0.80224609375,
0.6845703125,
0.80224609375,
0.69873046875,
0.529296875,
0.69873046875
],
[
0.8056640625,
0.6845703125,
0.91015625,
0.6845703125,
0.91015625,
0.69873046875,
0.8056640625,
0.69873046875
],
[
0.53125,
0.6962890625,
0.91259765625,
0.6962890625,
0.91259765625,
0.7109375,
0.53125,
0.7109375
],
[
0.0882568359375,
0.705078125,
0.240234375,
0.705078125,
0.240234375,
0.71826171875,
0.0882568359375,
0.71826171875
],
[
0.529296875,
0.708984375,
0.91259765625,
0.71044921875,
0.91259765625,
0.724609375,
0.529296875,
0.7236328125
],
[
0.0882568359375,
0.72265625,
0.486083984375,
0.72265625,
0.486083984375,
0.7373046875,
0.0882568359375,
0.7373046875
],
[
0.529296875,
0.72265625,
0.91259765625,
0.7216796875,
0.91259765625,
0.736328125,
0.529296875,
0.7373046875
],
[
0.08660888671875,
0.736328125,
0.483642578125,
0.736328125,
0.483642578125,
0.75,
0.08660888671875,
0.75
],
[
0.53125,
0.73486328125,
0.62353515625,
0.736328125,
0.62353515625,
0.74755859375,
0.53125,
0.7470703125
],
[
0.0882568359375,
0.75048828125,
0.486083984375,
0.75048828125,
0.486083984375,
0.7646484375,
0.0882568359375,
0.7646484375
],
[
0.51318359375,
0.75244140625,
0.91259765625,
0.75390625,
0.91259765625,
0.76806640625,
0.51318359375,
0.767578125
],
[
0.0849609375,
0.76318359375,
0.48193359375,
0.76513671875,
0.48193359375,
0.77978515625,
0.0849609375,
0.77783203125
],
[
0.53125,
0.767578125,
0.85546875,
0.767578125,
0.85546875,
0.7822265625,
0.53125,
0.7822265625
],
[
0.87353515625,
0.76953125,
0.91015625,
0.76953125,
0.91015625,
0.78076171875,
0.87353515625,
0.78076171875
],
[
0.0882568359375,
0.77880859375,
0.484375,
0.77880859375,
0.484375,
0.79248046875,
0.0882568359375,
0.79248046875
],
[
0.529296875,
0.77978515625,
0.91259765625,
0.77978515625,
0.91259765625,
0.79345703125,
0.529296875,
0.79345703125
],
[
0.0882568359375,
0.79248046875,
0.486083984375,
0.79248046875,
0.486083984375,
0.80615234375,
0.0882568359375,
0.80615234375
],
[
0.529296875,
0.79345703125,
0.89208984375,
0.79345703125,
0.89208984375,
0.806640625,
0.529296875,
0.806640625
],
[
0.0849609375,
0.8046875,
0.486083984375,
0.80615234375,
0.486083984375,
0.82080078125,
0.0849609375,
0.8193359375
],
[
0.51318359375,
0.80810546875,
0.91259765625,
0.81005859375,
0.91259765625,
0.82763671875,
0.51318359375,
0.82568359375
],
[
0.0882568359375,
0.82080078125,
0.484375,
0.82080078125,
0.484375,
0.833984375,
0.0882568359375,
0.833984375
],
[
0.529296875,
0.82470703125,
0.91259765625,
0.82568359375,
0.91259765625,
0.8388671875,
0.529296875,
0.83837890625
],
[
0.08905029296875,
0.833984375,
0.484375,
0.833984375,
0.484375,
0.84765625,
0.08905029296875,
0.84765625
],
[
0.53125,
0.83837890625,
0.9111328125,
0.83837890625,
0.9111328125,
0.8515625,
0.53125,
0.8515625
],
[
0.08660888671875,
0.84765625,
0.484375,
0.84765625,
0.484375,
0.861328125,
0.08660888671875,
0.861328125
],
[
0.529296875,
0.85107421875,
0.9140625,
0.85107421875,
0.9140625,
0.8642578125,
0.529296875,
0.8642578125
],
[
0.0882568359375,
0.8623046875,
0.484375,
0.8623046875,
0.484375,
0.87548828125,
0.0882568359375,
0.87548828125
],
[
0.53125,
0.86279296875,
0.9140625,
0.86279296875,
0.9140625,
0.876953125,
0.53125,
0.876953125
],
[
0.08660888671875,
0.87548828125,
0.486083984375,
0.87548828125,
0.486083984375,
0.8896484375,
0.08660888671875,
0.8896484375
],
[
0.53125,
0.87548828125,
0.74609375,
0.87548828125,
0.74609375,
0.89013671875,
0.53125,
0.89013671875
]
] |
[
"mean-pooled representation of [Lazaridou et al.,",
"学生网络的训练中有所不同。这一方面表明了从学生网",
"2017] as illustrated by equation (2) for predicting",
"络到教师网络的错误传播的存在，另一方面也证实了我",
"representations of OOV words. The training of A is",
"们提出的解决OOV问题的方法的有效性；最后（5）我们",
"performed on words that occur no less than 5 times",
"提出的方法始终优于字符语言模型[Akbiketal.，2018]。",
"in the training set.",
"这个结果的一个可能的解释是，我们的方法可以使用单",
"[Schick and Schitze, 2018l :This baseline is a",
"词级嵌人而不会遭受OOV问题，而字符语言模型不能",
"",
"variant of the proposed method, but its student",
"5",
"Conclusion",
"",
"network is trained on the reconstruction object as",
"defined in equation (15).It is also related to the",
"在这项工作中，我们提出了一种新颖的方法，仅使用任",
"work of [Schick and Schutze, 2018] but using the",
"务的训练数据来解决序列标记系统中的词汇外问题。",
"V",
"same architecture of the proposed method to model",
"旨在从OOV单词的表面形式和上下文中生成OOV单",
"word context and surface-form.",
"词的表示形式。此外，它的设计是为了避免在同一范式",
"Akbik et al..2018:This baseline converts the",
"中现有方法遭受的错误传播问题。关于POS标记（POS",
"和命名实体识别（NER）的广泛实验研究表明，该方法",
"",
"input sentence into a character sequence. Then, it",
"applies a character language model to the character",
"可以在OOV问题上比现有方法取得更好或可比的性能。",
"sequence to get the representation of every word",
"within the sentence. Based on the obtained word",
"",
"Acknowledgements",
"representations, it applies a LSTM network to",
"b",
"作者希望感谢匿名审稿人的有益评论。Thisworkwas",
"model the word sequence and performs final tag",
"partiallyfundedbyChinaNationalKeyR&DProgram(No。",
"recommendation.",
"\"To make it compatible with",
"2018YFC0831105,2017YFB1002104,2018YFB1005104)，",
"the setting of this work, we did not pre-train the",
"国家自然科学基金（No。61532011,61751201)，上海市",
"language model on external data and not use pre-",
"科技重大专项（No。2018SHZDZX01),STCSM(编号。",
"trained staticword embeddings forthisbaseline",
"16JC1420401,17JC1420200),ZJLab",
"4.3",
"",
"Implementation Detail",
"References",
"对于数据预处理，所有数字均替换为特殊令牌“<NUM",
"[Akbik et al., 2018] Alan Akbik, Duncan Blythe, and",
"”，所有URL均替换为特殊令牌“<URL>”。教师",
"Roland Vollgraf.",
"Contextual string embeddings",
"和学生网络的单词嵌人、字符嵌人和LSTM的维度分别",
"",
"for sequence labeling.",
"In Proceedings of the",
"设置为50、16和50。对于内核宽度3和5，字符CNN",
"27th International Conference on Computational",
"的内核大小设置为25。使用Adam步长规则[Kingaand",
"Linguistics, pages 1638-1649, 2018.",
"Adam，2015]进行优化，学习率设置为1e-3",
"[Barbu Mititelu et al., 2016] V Barbu Mititelu, R Ion,",
"RSimionescu,",
"E Irimia, and CA Perez.",
"4.4Evaluation",
"The",
"romanian treebank annotated according to universal",
"我们将测试（或开发）集分为两个子集：词汇内（WIV）",
"dependencies.",
"In Proceedings of The Tenth Inter-",
"单词和词汇外（OOV）单词。如果一个词在训练集中出现",
"national Conference on Natural Language Processing",
"超过5次，则将其视为WIV，否则为OOV。对于NER",
"(HrTAL2016), 2016.",
"如果一个实体至少包含一个OOV词集的单词，则该实",
"[Bojarski et al., 2016] Mariusz",
"Bojarski,",
"Davide",
"体被认为是OOV的。我们在OOV集上报告模型性能",
"Del Testa, Daniel Dworakowski, Bernhard",
"Firner",
"（POS和NER的F1的准确性）。这是因为我们可以准",
"Beat Flepp, Prasoon Goyal, Lawrence D",
"Jackel",
"确地说出一个词属于哪个子集，因此我们可以很容易地",
"Mathew Monfort, Urs Muller, Jiakai Zhang, et al.",
"在每个子集上组合最佳模型，以在整个测试集上实现最",
"End to end learning for self-driving cars.",
"arXiv",
"佳的整体性能。例如，我们可以使用我们提出的模型对",
"preprintarXiv:1604.07316.2016.",
"OOV单词进行标签预测，并使用WIV子集上性能最好",
"[Caselli et al., 2015] Tommaso Caselli,",
"Piek",
"Vossen",
"的模型对WIV单词进行标签预测",
"Marieke van Erp, Antske Fokkens,",
"Filip Ilievski",
"Ruben Izquierdo, Minh Le, Roser Morante, and",
"4.5Main Results",
"Marten Postma. When it's all piling up: investigating",
"表2和3报告了在使用BiLSTM和CNN实现教师网络",
"error propagation in an nlp pipeline.In WNACP@",
"时，分别为POS标记和NER设置的OOV上的模型性",
"NLDB,2015",
"能。从该表，我们有以下观察：（1）在大多数任务中，处",
"[Chen et al., 2015] Xinchi Chen, Xipeng Qiu, Chenxi",
"理OOV问题的方法优于RandomUNK基线。这验证了",
"Zhu, Pengfei Liu, and Xuanjing Huang.",
"Long",
"在序列标记中处理OOV问题的必要性。（2）使用表面形",
"short-term memory neural networks for chinese word",
"式和上下文信息生成OOV单词表示的方法[Schickand",
"segmentation. In EMNLP, pages 1197-1206, 2015.",
"Schitze，2018]优于仅使用大多数数据集上的上下文信",
"[Deleger et al., 2016] Louise Deleger, Robert Bossy,",
"息的方法[Khodaketal.,2018]。这显示了表面形式和上",
"Estelle Chaix, Mouhamadou Ba,Arnaud Ferre,",
"下文信息的互补性；（3）师生范式[SchickandSchitze,",
"Philippe Bessieres, and Claire Nedellec.Overview",
"2018]中最具可比性的基线总体上优于SingleUNK。这",
"of the bacteria biotope task at bionlp shared task",
"验证了师生范式下动机的有效性；（4）我们提出的方法始",
"2016. In Proceedings of the 4th BioNLP Shared Task",
"终优于[SchickandSchitze，2018]，这与提出的方法在",
"Workshop, pages 12-22, 2016."
] |
{
"arxiv_id": "1905.12277",
"language": "zh",
"page_idx": 4
}
|
||
2207.09071-zh-1
|
2207.09071-zh-1.jpg
|
# 用于稀疏奖励任务上的元强化学习的学习动作翻译器
Yijie Guo$^1$,Qiucheng Wu$^1$,Honglak Lee$^1,2$
$^{1}$University of Michigan, $^2$LG AI Research
guoyijie@umich.edu, wuqiuche@umich.edu, honglak@eecs.umich.edu
<FigureHere>
Figure 1: IIlustration of our policy transfer. Size of arrows represents avg. episode reward of learned or transferred policy on target tasks. Different colors indicate different tasks.
$\operatorname{Abstract}$
Meta reinforcement learning (meta-RL) aims to learn a policy solving a set of training tasks simultaneously and quickly adapting to new tasks. It requires massive amounts of data drawn from training tasks to infer the common structure shared among tasks. Without heavy reward engineering, the sparse rewards in long-horizon tasks exacerbate the problem of sample efficiency in meta-RL. Another challenge in meta-RL is the discrepancy of difficulty level among tasks, which might cause one easy task dominating learning of the $\bar{\text{shared policy and thus preclude policy adaptation to}}$ new tasks. This work introduces a novel objective function to learn an action translator among training tasks. We theoretically verify that the value of the transferred policy with the action translator can be close to the value of the source policy and our objective function (approximately) upper bounds the value difference. We propose to combine the action translator with context-based meta-RL algorithms for better data collection and more efficient exploration during metatraining. Our approach empirically improves the sample efficiency and performance of meta-RL algorithms on sparse- reward tasks.
# 1 Introduction
深度强化学习(DRL)方法在解决复杂任务(Mnih et al. 2015;Silver et al. 2016; Schulman et al. 2017) 方面取得了显著成功。虽然传统的 DRL 方法为每个任务学习
个单独的策略，但元强化学习(meta-RL)算法(Finn, Abbeel, and Levine 2017;Duan et al. 2016; Mishra et al. 2017)可以跨任务分布学习共享结构，以便代理可以在测试阶段快速适应看不见的相关任务。与大多数现有的元 RL 方法处理具有密集奖励的任务不同，我们将重点放在稀疏奖励训练任务上，这些任务在现实世界中更常见，而无法在环境中访问精心设计的奖励功能。元 RL 的最新工作提出了非策略算法 (Rakelly et al. 2019; Fakoor et al. 2019) 和基于模型的算法 (Nagabandi, Finn, and Levine 2018; Nagabandi et al. 2018), 以提高元训练过程中的样本效率。然而，有效地解决需要长时间推理和稀疏奖励的多个任务仍然具有挑战性。在这些任务中， 正奖励的稀缺加剧了样本效率的问题，这困扰着元 RL
Copyright $\odot2022$, Association for the Advancement of Ar-
tificial Intelligence (www.aaai.org). All rights reserved.
算法，并且由于缺乏引导信号而使探索变得困难。直观上，我们希望解决一个任务有助于学习其他相关任务，因为培训任务共享一个共同的结构。然而，在实践中通常不是这样 (Rusu et al. 2015; Parisotto, Ba, and Salakhutdinov 2015)。先前的工作 (Teh et al. 2017; Yu et al. 2020a) 指出，有害的梯度干扰可能会导致多任务策略学习的不平衡。策略提炼(Teh et al.2017) 和梯度投影 (Yu et al. 2020a) 是在元 RL 算法中开发的，以缓解这个问题。但是，在稀疏奖励设置中，此问题可能会变得更加严重，因为很难探索每个任务以获取有意义的策略更新梯度信号。在一个任务中的良好性能不会自动帮助探索其他任务，因为代理在其他任务中缺乏积极的奖励来学习
在这项工作中，我们的目标是充分利用代理在探索中偶尔发现的高回报过渡。一项任务中的好经验不仅应该改进这项任务的政策，而且应该有利于其他任务的政府策，以推动更深入的探索。具体来说，一旦代理从一个训练任务中的成功轨迹中学习，我们就会将该任务中的良好策略转移到其他任务中，以在其他训练任务中获得更多的积极奖励。InFig。$\mathbb{E}1$,如果学习的策略 $\pi$ 在任务$\mathcal{T}^{(2)}$上的表现比其他任务更好，那么我们的目标是将良好的策略$\pi(\cdot,\mathcal{T}^{(2)})$转移到其他任务$\mathcal{T}^(1)$和$\mathcal{T}^(3)$ 。为了实现这种转移，我们建议在多个培训任务中学习动作翻译器。目标函数强制转换后的操作在目标任务上的行为与源任务上的源操作类似。我们考虑训练任务分
|
[
[
0.373291015625,
0.1256103515625,
0.62744140625,
0.1256103515625,
0.62744140625,
0.14013671875,
0.373291015625,
0.14013671875
],
[
0.359375,
0.14453125,
0.63623046875,
0.14453125,
0.63623046875,
0.1622314453125,
0.359375,
0.1622314453125
],
[
0.324462890625,
0.179931640625,
0.68115234375,
0.1767578125,
0.68115234375,
0.1976318359375,
0.324462890625,
0.2008056640625
],
[
0.34716796875,
0.2020263671875,
0.650390625,
0.203857421875,
0.650390625,
0.2183837890625,
0.34716796875,
0.216552734375
],
[
0.2548828125,
0.2183837890625,
0.74169921875,
0.216552734375,
0.74169921875,
0.2310791015625,
0.2548828125,
0.23291015625
],
[
0.64453125,
0.271484375,
0.72216796875,
0.2734375,
0.72216796875,
0.285400390625,
0.64453125,
0.28271484375
],
[
0.75634765625,
0.272705078125,
0.82421875,
0.2734375,
0.82421875,
0.285400390625,
0.75634765625,
0.2841796875
],
[
0.7548828125,
0.2841796875,
0.81201171875,
0.2841796875,
0.81201171875,
0.298583984375,
0.7548828125,
0.298583984375
],
[
0.250732421875,
0.2734375,
0.314453125,
0.2734375,
0.314453125,
0.28857421875,
0.250732421875,
0.28857421875
],
[
0.64453125,
0.285400390625,
0.69287109375,
0.285400390625,
0.69287109375,
0.296630859375,
0.64453125,
0.296630859375
],
[
0.10296630859375,
0.29736328125,
0.463134765625,
0.29736328125,
0.463134765625,
0.311279296875,
0.10296630859375,
0.311279296875
],
[
0.54150390625,
0.29345703125,
0.57275390625,
0.291748046875,
0.57421875,
0.3037109375,
0.54345703125,
0.30615234375
],
[
0.65673828125,
0.298583984375,
0.6689453125,
0.298583984375,
0.6689453125,
0.309326171875,
0.65673828125,
0.309326171875
],
[
0.10296630859375,
0.3125,
0.4599609375,
0.3125,
0.4599609375,
0.322509765625,
0.10296630859375,
0.322509765625
],
[
0.65283203125,
0.309326171875,
0.68310546875,
0.311279296875,
0.68115234375,
0.322509765625,
0.65185546875,
0.32080078125
],
[
0.66845703125,
0.318115234375,
0.68115234375,
0.326904296875,
0.6748046875,
0.335205078125,
0.66015625,
0.32763671875
],
[
0.10455322265625,
0.324462890625,
0.459228515625,
0.324462890625,
0.459228515625,
0.335205078125,
0.10455322265625,
0.335205078125
],
[
0.10296630859375,
0.3359375,
0.4599609375,
0.3359375,
0.4599609375,
0.349853515625,
0.10296630859375,
0.349853515625
],
[
0.55615234375,
0.347900390625,
0.57275390625,
0.347900390625,
0.57275390625,
0.356689453125,
0.55615234375,
0.356689453125
],
[
0.68115234375,
0.343505859375,
0.70751953125,
0.343505859375,
0.70751953125,
0.3515625,
0.68115234375,
0.3515625
],
[
0.6748046875,
0.352783203125,
0.7138671875,
0.352783203125,
0.7138671875,
0.363037109375,
0.6748046875,
0.363037109375
],
[
0.10296630859375,
0.361083984375,
0.463134765625,
0.361083984375,
0.463134765625,
0.375,
0.10296630859375,
0.375
],
[
0.10296630859375,
0.348388671875,
0.461669921875,
0.348388671875,
0.461669921875,
0.3623046875,
0.10296630859375,
0.3623046875
],
[
0.83642578125,
0.3662109375,
0.85546875,
0.3662109375,
0.85546875,
0.37060546875,
0.83642578125,
0.37060546875
],
[
0.10296630859375,
0.37255859375,
0.461669921875,
0.37255859375,
0.461669921875,
0.386474609375,
0.10296630859375,
0.386474609375
],
[
0.10296630859375,
0.385009765625,
0.461669921875,
0.385009765625,
0.461669921875,
0.3984375,
0.10296630859375,
0.3984375
],
[
0.10455322265625,
0.40087890625,
0.4599609375,
0.40087890625,
0.4599609375,
0.410888671875,
0.10455322265625,
0.410888671875
],
[
0.54345703125,
0.40087890625,
0.57275390625,
0.40087890625,
0.57275390625,
0.41357421875,
0.54345703125,
0.41357421875
],
[
0.72216796875,
0.397705078125,
0.73388671875,
0.397705078125,
0.73388671875,
0.401611328125,
0.72216796875,
0.401611328125
],
[
0.10455322265625,
0.41357421875,
0.461669921875,
0.41357421875,
0.461669921875,
0.423583984375,
0.10455322265625,
0.423583984375
],
[
0.640625,
0.42041015625,
0.71142578125,
0.42041015625,
0.71142578125,
0.43115234375,
0.640625,
0.43115234375
],
[
0.75732421875,
0.42041015625,
0.82275390625,
0.42041015625,
0.82275390625,
0.43115234375,
0.75732421875,
0.43115234375
],
[
0.10455322265625,
0.438232421875,
0.459228515625,
0.438232421875,
0.459228515625,
0.448974609375,
0.10455322265625,
0.448974609375
],
[
0.10455322265625,
0.425537109375,
0.461669921875,
0.425537109375,
0.461669921875,
0.436279296875,
0.10455322265625,
0.436279296875
],
[
0.10296630859375,
0.449462890625,
0.4599609375,
0.45068359375,
0.4599609375,
0.462158203125,
0.10296630859375,
0.4609375
],
[
0.521484375,
0.4462890625,
0.91015625,
0.4462890625,
0.91015625,
0.460205078125,
0.521484375,
0.460205078125
],
[
0.10296630859375,
0.46337890625,
0.45751953125,
0.46337890625,
0.45751953125,
0.473388671875,
0.10296630859375,
0.473388671875
],
[
0.521484375,
0.460205078125,
0.91015625,
0.460205078125,
0.91015625,
0.473388671875,
0.521484375,
0.473388671875
],
[
0.10455322265625,
0.47607421875,
0.461669921875,
0.47607421875,
0.461669921875,
0.486083984375,
0.10455322265625,
0.486083984375
],
[
0.521484375,
0.474853515625,
0.9111328125,
0.474853515625,
0.9111328125,
0.488037109375,
0.521484375,
0.488037109375
],
[
0.10455322265625,
0.488037109375,
0.461669921875,
0.488037109375,
0.461669921875,
0.498779296875,
0.10455322265625,
0.498779296875
],
[
0.52294921875,
0.4892578125,
0.62744140625,
0.4892578125,
0.62744140625,
0.501953125,
0.52294921875,
0.501953125
],
[
0.10296630859375,
0.5,
0.461669921875,
0.5,
0.461669921875,
0.51318359375,
0.10296630859375,
0.51318359375
],
[
0.10296630859375,
0.51220703125,
0.463134765625,
0.51220703125,
0.463134765625,
0.52587890625,
0.10296630859375,
0.52587890625
],
[
0.10296630859375,
0.5244140625,
0.461669921875,
0.5244140625,
0.461669921875,
0.53857421875,
0.10296630859375,
0.53857421875
],
[
0.10455322265625,
0.5390625,
0.4599609375,
0.5390625,
0.4599609375,
0.5498046875,
0.10455322265625,
0.5498046875
],
[
0.52294921875,
0.5390625,
0.91259765625,
0.5390625,
0.91259765625,
0.55322265625,
0.52294921875,
0.55322265625
],
[
0.10296630859375,
0.55029296875,
0.45751953125,
0.55029296875,
0.45751953125,
0.56103515625,
0.10296630859375,
0.56103515625
],
[
0.52294921875,
0.55322265625,
0.91259765625,
0.55322265625,
0.91259765625,
0.56640625,
0.52294921875,
0.56640625
],
[
0.10296630859375,
0.56298828125,
0.461669921875,
0.56298828125,
0.461669921875,
0.5771484375,
0.10296630859375,
0.5771484375
],
[
0.52294921875,
0.56640625,
0.9111328125,
0.56640625,
0.9111328125,
0.580078125,
0.52294921875,
0.580078125
],
[
0.10296630859375,
0.57568359375,
0.2607421875,
0.57568359375,
0.2607421875,
0.5888671875,
0.10296630859375,
0.5888671875
],
[
0.521484375,
0.58056640625,
0.91259765625,
0.58056640625,
0.91259765625,
0.5947265625,
0.521484375,
0.5947265625
],
[
0.52294921875,
0.59521484375,
0.9111328125,
0.59521484375,
0.9111328125,
0.60595703125,
0.52294921875,
0.60595703125
],
[
0.238525390625,
0.609375,
0.35302734375,
0.611328125,
0.35302734375,
0.62548828125,
0.238525390625,
0.62353515625
],
[
0.521484375,
0.60791015625,
0.91015625,
0.60791015625,
0.91015625,
0.6220703125,
0.521484375,
0.6220703125
],
[
0.52294921875,
0.6220703125,
0.9111328125,
0.6220703125,
0.9111328125,
0.63525390625,
0.52294921875,
0.63525390625
],
[
0.2099609375,
0.611328125,
0.2467041015625,
0.611328125,
0.2467041015625,
0.62255859375,
0.2099609375,
0.62255859375
],
[
0.0882568359375,
0.630859375,
0.476318359375,
0.630859375,
0.476318359375,
0.64453125,
0.0882568359375,
0.64453125
],
[
0.52294921875,
0.63525390625,
0.9111328125,
0.63525390625,
0.9111328125,
0.6484375,
0.52294921875,
0.6484375
],
[
0.08905029296875,
0.646484375,
0.476318359375,
0.646484375,
0.476318359375,
0.6572265625,
0.08905029296875,
0.6572265625
],
[
0.521484375,
0.6494140625,
0.9111328125,
0.6494140625,
0.9111328125,
0.66357421875,
0.521484375,
0.66357421875
],
[
0.0906982421875,
0.6591796875,
0.464111328125,
0.6591796875,
0.464111328125,
0.67236328125,
0.0906982421875,
0.67236328125
],
[
0.52294921875,
0.66357421875,
0.9111328125,
0.66357421875,
0.9111328125,
0.6767578125,
0.52294921875,
0.6767578125
],
[
0.0882568359375,
0.67236328125,
0.4755859375,
0.67236328125,
0.4755859375,
0.68603515625,
0.0882568359375,
0.68603515625
],
[
0.521484375,
0.6767578125,
0.9111328125,
0.6767578125,
0.9111328125,
0.68994140625,
0.521484375,
0.68994140625
],
[
0.08905029296875,
0.68798828125,
0.4755859375,
0.68798828125,
0.4755859375,
0.69873046875,
0.08905029296875,
0.69873046875
],
[
0.52294921875,
0.69140625,
0.91259765625,
0.69140625,
0.91259765625,
0.705078125,
0.52294921875,
0.705078125
],
[
0.08660888671875,
0.70068359375,
0.47802734375,
0.70068359375,
0.47802734375,
0.7138671875,
0.08660888671875,
0.7138671875
],
[
0.52294921875,
0.70556640625,
0.60302734375,
0.70556640625,
0.60302734375,
0.71728515625,
0.52294921875,
0.71728515625
],
[
0.0882568359375,
0.7138671875,
0.47802734375,
0.7138671875,
0.47802734375,
0.72802734375,
0.0882568359375,
0.72802734375
],
[
0.08905029296875,
0.72802734375,
0.47802734375,
0.72802734375,
0.47802734375,
0.7412109375,
0.08905029296875,
0.7412109375
],
[
0.53759765625,
0.7236328125,
0.9111328125,
0.7236328125,
0.9111328125,
0.7373046875,
0.53759765625,
0.7373046875
],
[
0.0882568359375,
0.7412109375,
0.476318359375,
0.7412109375,
0.476318359375,
0.7548828125,
0.0882568359375,
0.7548828125
],
[
0.52294921875,
0.7373046875,
0.91259765625,
0.7373046875,
0.91259765625,
0.75048828125,
0.52294921875,
0.75048828125
],
[
0.0882568359375,
0.755859375,
0.47802734375,
0.755859375,
0.47802734375,
0.76953125,
0.0882568359375,
0.76953125
],
[
0.52294921875,
0.75048828125,
0.91259765625,
0.75048828125,
0.91259765625,
0.7646484375,
0.52294921875,
0.7646484375
],
[
0.51953125,
0.7646484375,
0.8056640625,
0.7666015625,
0.8056640625,
0.77978515625,
0.51953125,
0.77783203125
],
[
0.79833984375,
0.7666015625,
0.91015625,
0.7666015625,
0.91015625,
0.7763671875,
0.79833984375,
0.7763671875
],
[
0.52294921875,
0.77880859375,
0.9111328125,
0.77880859375,
0.9111328125,
0.79248046875,
0.52294921875,
0.79248046875
],
[
0.08660888671875,
0.76953125,
0.476318359375,
0.76953125,
0.476318359375,
0.78271484375,
0.08660888671875,
0.78271484375
],
[
0.08660888671875,
0.7841796875,
0.47802734375,
0.7841796875,
0.47802734375,
0.79736328125,
0.08660888671875,
0.79736328125
],
[
0.08660888671875,
0.79736328125,
0.47802734375,
0.79736328125,
0.47802734375,
0.81103515625,
0.08660888671875,
0.81103515625
],
[
0.52392578125,
0.79248046875,
0.9111328125,
0.79248046875,
0.9111328125,
0.80615234375,
0.52392578125,
0.80615234375
],
[
0.0882568359375,
0.81103515625,
0.47802734375,
0.81103515625,
0.47802734375,
0.82470703125,
0.0882568359375,
0.82470703125
],
[
0.52294921875,
0.80615234375,
0.9111328125,
0.80615234375,
0.9111328125,
0.8193359375,
0.52294921875,
0.8193359375
],
[
0.51953125,
0.8173828125,
0.9111328125,
0.818359375,
0.9111328125,
0.833984375,
0.51953125,
0.83349609375
],
[
0.08660888671875,
0.82470703125,
0.47802734375,
0.82373046875,
0.47802734375,
0.83837890625,
0.08660888671875,
0.8388671875
],
[
0.5185546875,
0.83056640625,
0.91259765625,
0.830078125,
0.91259765625,
0.84765625,
0.5185546875,
0.8486328125
],
[
0.0882568359375,
0.8388671875,
0.47802734375,
0.8388671875,
0.47802734375,
0.85302734375,
0.0882568359375,
0.85302734375
],
[
0.53759765625,
0.84765625,
0.9111328125,
0.84765625,
0.9111328125,
0.861328125,
0.53759765625,
0.861328125
],
[
0.08660888671875,
0.86279296875,
0.4755859375,
0.86279296875,
0.4755859375,
0.876953125,
0.08660888671875,
0.876953125
],
[
0.52294921875,
0.8623046875,
0.9111328125,
0.8623046875,
0.9111328125,
0.87548828125,
0.52294921875,
0.87548828125
],
[
0.0849609375,
0.875,
0.44775390625,
0.87548828125,
0.44775390625,
0.89013671875,
0.0849609375,
0.8896484375
],
[
0.521484375,
0.87548828125,
0.9111328125,
0.87548828125,
0.9111328125,
0.8896484375,
0.521484375,
0.8896484375
]
] |
[
"用于稀疏奖励任务上的元",
"强化学习的学习动作翻译器",
"Yijie Guo', Qiucheng Wu', Honglak Leel,2",
"1University of Michigan, 2LG AI Research",
"guoyijie@umich.edu,wuqiuche@umich.edu,honglak@eecs.umich.edu",
"LearnedPolicy",
"Select bette",
"Policy",
"Abstract",
"元，T)",
"Meta reinforcement learning (meta-RL) aims to learn",
"T(1)",
"11",
"a policy solving a set of training tasks simultaneously",
"curce",
"IC",
"and quickly adapting to new tasks.It requires mas",
"sive amounts of data drawn from training tasks to in-",
"2",
"Actior",
"Translato",
"out heavy reward engineering, the sparse rewards in",
"fer the common structure shared among tasks. With-",
"",
"long-horizon tasks exacerbate the problem of sample",
"efficiency in meta-RL. Another challenge in meta-RL",
"isthediscrepancyof difficultylevelamongtasks,which",
"T(3)",
"O",
"might cause one easy task dominating learning of the",
"SourceTasks",
"Target Tasks",
"newtasks. This work introduces a novel objective func",
"shared policy and thus precludepolicy adaptationto",
"tionto learnan action translator among training tasks.",
"Figure l: Illustration of our policy transfer. Size of ar-",
"We theoretically verify that the value of the trans",
"rows represents avg. episode reward of learned or trans-",
"ferred policy with the action translator can be close",
"ferred policy on target tasks. Different colors indicate",
"to the value of the source policy and our objective",
"different tasks.",
"function (approximately) upper bounds the value dif-",
"ference. We propose to combine the action translator",
"with context-based meta-RL algorithms for better data",
"collection and more efficient exploration during meta",
"算法，并且由于缺乏引导信号而使探索变得困难。直",
"training.Our approach empirically improves the sam",
"观上，我们希望解决一个任务有助于学习其他相关任",
"ple efficiency and performance of meta-RL algorithms",
"务，因为培训任务共享一个共同的结构。然而，在实践",
"on sparse-reward tasks.",
"中通常不是这样（Rusuetal.2015；Parisotto，Ba，and",
"Salakhutdinov2015）。先前的工作（Tehetal.2017：Yu",
"Introduction",
"etal.2020a）指出，有害的梯度干扰可能会导致多任务",
"策略学习的不平衡。策略提炼（Tehetal.2017）和梯度",
"1",
"深度强化学习（DRL）方法在解决复杂任务（Mnihetal",
"投影（Yuetal.2020a）是在元RL算法中开发的，以缓",
"2015:Silveretal.2016:Schulmanetal.2017）方面取得",
"解这个问题。但是，在稀疏奖励设置中，此问题可能会",
"了显著成功。虽然传统的DRL方法为每个任务学习",
"变得更加严重，因为很难探索每个任务以获取有意义的",
"个单独的策略，但元强化学习（meta-RL）算法（Finn",
"策略更新梯度信号。在一个任务中的良好性能不会自动",
"Abbeel,andLevine2017:Duanet al.2016:Mishraet al",
"帮助探索其他任务，因为代理在其他任务中缺乏积极的",
"2017）可以跨任务分布学习共享结构，以便代理可以在",
"奖励来学习",
"测试阶段快速适应看不见的相关任务。与大多数现有的",
"元RL方法处理具有密集奖励的任务不同，我们将重点",
"在这项工作中，我们的目标是充分利用代理在探索中",
"放在稀疏奖励训练任务上，这些任务在现实世界中更常",
"偶尔发现的高回报过渡。一项任务中的好经验不仅应",
"见，而无法在环境中访问精心设计的奖励功能。元RL的",
"该改进这项任务的政策，而且应该有利于其他任务的政",
"策，以推动更深人的探索。具体来说，一",
"一日代理从一个",
"训练任务中的成功轨迹中学习，我们就会将该任务中的",
"最新工作提出了非策略算法（Rakellyetal.2019;Fakoor",
"etal.2019）和基于模型的算法（Nagabandi，Finn，and",
"Levine2018；Nagabandietal.2018），以提高元训练过",
"良好策略转移到其他任务中，以在其他训练任务中获得",
"程中的样本效率。然而，有效地解决需要长时间推理和",
"更多的积极奖励。InFig。E1，如果学习的策略元在任",
"务T(2）上的表现比其他任务更好，那么我们的目标是",
"稀疏奖励的多个任务仍然具有挑战性。在这些任务中",
"将良好的策略元（F(2)）转移到其他任务(i)和宁(③)",
"正奖励的稀缺加剧了样本效率的问题，这困扰着元RL",
"为了实现这种转移，我们建议在多个培训任务中学习",
"Copyright 2022, Association for the Advancement of Ar-",
"动作翻译器。目标函数强制转换后的操作在目标任务上",
"tificial Intelligence (www.aaai.org). All rights reserved.",
"的行为与源任务上的源操作类似。我们考虑训练任务分"
] |
{
"arxiv_id": "2207.09071",
"language": "zh",
"page_idx": 1
}
|
||
2207.09071-zh-2
|
2207.09071-zh-2.jpg
|
配中任何一对源和目标任务的策略转移(参见图。1)。如果在最近的事件中，转移的策略在目标任务上获得的奖励高于学习的策略，则代理将按照转移的策略执行操作。这种方法使代理能够利用来自多个训练任务的相关数据，鼓励学习的策略在多个训练任务上同样表现良好，从而在将训练有素的策略应用于测试任务时带来更好的性能
我们总结了贡献：(1)我们引入了一种新颖的目标函数，将任何策略从源马尔可夫决策过程(MDP)转移到目标 MDP。我们证明了一个理论保证，即转移的策略可以在源 MDP 上实现接近源策略的目标 MDP 的预期收益。预期收益的差异(近似)上界由我们的损失函数与常数乘法因子。(2)我们开发了一种名为Meta-RL 的非策略 RL 算法，该算法具有 Context 条件的 ActionTranslator (MCAT),在元 RL 中应用了策略传输机制，以帮助跨多个稀疏奖励任务进行探索。(3) 我们通过 MuJoCo 物理引擎 (Todorov, Erez, and Tassa 2012)实证证明了 MCAT 在各种模拟控制任务上的有效性，表明策略传输提高了基于上下文的元 RL 算法的性能
## 2 Related Work
基于上下文的元 RL 元强化学习已在文献 (Finn, Abbeel, and Levine 2017; Stadie et al. 2018; Sung et al. 2017;Xu, van Hasselt, and Silver 2018) 中进行了广泛的研究，其中许多作品开发了基于上下文的方法(Rakelly et al. 2019; Ren et al. 2020; Liu et al. 2020) 。Duan et al. (2016); Wang et al. (2016); Fakoor et al.(2019) 采用递归神经网络对上下文转换进行编码，并在上下文变量上制定策略条件。最大化预期收益的目标函数联合训练上下文编码器和策略。Rakelly et al.(2019) 利用置换不变编码器将经验聚合为概率上下文变量，并通过变分推理对其进行优化。后验采样有利于在适应阶段探索稀疏奖励任务，但在训练阶段可以获得密集奖励。Li, Pinto, and Abbeel (2020)考虑了奖励函数的任务族。Lee et al. (2020);Seo et al. (2020) 使用前向动态预测来训练上下文编码器。这些基于模型的元 RL 算法假设奖励函数可用于计划。在没有地面实况奖励函数的稀疏奖励设置中，他们可能难以发现非零奖励，并且准确地估计基于模型的计划的奖励也可能是有问题的
RL 中的策略转移策略转移研究给定一组源任务及其专家策略的目标任务中的知识转移。策略提炼(Rusu et al. 2015; Yin and Pan 2017; Parisotto, Ba, and Salakhutdinov 2015) 最小化目标任务上的源策略和学习策略之间的动作分布差异。沿着这条工作线，Teh et al. (2017) 在多任务强化学习中创建了一个质心策略， 并将知识从特定于任务的策略中提取到该质心策略。或者，源和目标任务(Zhu, Lin, and Zhou 2020) 之间的任务间映射可以帮助策略传输。这些作品 (Gupta et al. 2017; Konidaris and Barto 2006; Ammar and Taylor 2011) 中的大多数都假设在状态空间上存在对应关系， 并学习任务之间的状态映射。最近的工作 (Zhang et al. 2020c)学习了具有动态循环一致性损失的状态对应关系和动作对应关系。我们的方法与这种方法的不同之处在于，我们可以使用更简单的目标函数在多个任务之间进行动作转换。重要的是，我们的方法是新颖的，可以将策略转移用于元 RL 中的任何一对源和目标任务
MDP 中状态的双模拟最近关于状态表示学习的工作(Ferns, Panangaden, and Precup 2004; Zhang et al.
2020a; Agarwal et al. 2021)研究了多个 MDP 上状态的双性度量，并考虑如何学习导致不同 MDP 中同一动作下几乎相同行为的状态的表示。在多任务强化学习和元强化学习问题中，Zhang et al. (2020a,b) 基于任务和状态相似性导出迁移和泛化边界。我们还限制了跨任务的策略传输的价值，但我们的方法是建立动作等效而不是状态等效
## 3 Method
在本节中，我们首先描述我们的方法，以学习捕获任务特征的上下文编码器，并学习在给定任务上下文的情况下预测下一个状态分布的前向动力学模型(Sec。3.2 )。然后，我们引入一个目标函数来训练动作翻译器，以便目标任务上的翻译动作与源任务上的源动作等效。动作翻译器可以以任务上下文为条件，因此它可以将一个好的策略从任何任意源任务转移到训练集中的任何其他目标任务 (Sec。3.3 )。最后，我们建议将动作翻译器与基于上下文的元 RL 算法相结合，以将良好的策略从任何一个任务转移到其他任务。在元训练期间，这种策略传输方法有助于利用在任何一项任务上遇到的良好经验，并有利于数据收集和对其他稀疏奖励任务的进一步策略优化 (Sec。3.4 )。Fig。2 概述了我们的方法 MCAT
## 3.1 Problem Formulation
,初始分布 遵循先前工作 (Duan et al.2016; Mishra et al. 2017; Rakelly et al. 2019) 中的元 RL 公式，我们假设任务$p(\mathcal{T})$的分布，并且每个任务是定义为具有状态空间$\mathcal{S}$的元组$(\mathcal{S},\mathcal{A},p,r,\gamma,\rho_0)$的马尔可夫决策过程 (MDP), 动作空间 X。我们也可以将奖励函数定步长，$i$ 表示任务的索引，上下文变量 $z_t^{(i)}\in\mathcal{Z}$ 从任务MDP 的历史转换中捕获上下文信息，而$z$ 是上下文向量的空间。对共享策略进行了优化，以使其在每个训练$\begin{array}{l}{\text{任务 }\mathcal{T}^{(i)}\text{上的值 }V^{\pi}(\mathcal{T}^{(i)})=\mathbb{E}_{\rho_{\breve{0}}^{(i)}}^{(i)}[\sum_{t=0}^{\infty}\gamma^{t}r_{t}^{(i)}]\text{ 最}}\\{\text{大化。遵循 meta-RL (Yu et al.}^{\rho_{\breve{0}}^{\pi}p_{1}^{(i)}}\hat{\tau};\mathrm{~Nagabandi~et~al.}}\end{array}$ 2018; Nagabandi, Finn, and Levine 2018; Zhou, Pinto, and Gupta 2019; Lee et al. 2020) 中的先前工作，我们研究具有相同状态空间，动作空间，奖励函数但动态函数变化的任务。重要的是，我们专注于更具挑战性的稀疏奖励设置。我们的目标是学习一种对动态变化稳健的共享策略，并将其推广到看不见的任务
3.2 Learning Context &Forward Model
为了捕获有关任何任务$\mathcal{T}^{(i)}$的知识，我们利用上下文编码器$C: \mathcal{S} ^K\times \mathcal{A} ^K\to \mathcal{Z}$ ,其中$K$是用于推断上下文的过去步骤数。(Rakelly et al. 2019; Zhou, Pinto, and Gupta 2019; Lee et al. 2020)对相关思想进行了探索。InFig。E2 a,给$\begin{array}{ll}{\text{定}K\text{ 过去的转换}}&{(s_{t-K}^{(i)},a_{t-K}^{(i)},\cdots,s_{t-1}^{(i)},a_{t-1}^{(i)})}\\{,_{(i)}\text{上下文编码器}}&{C_{(i)}\text{产生潜在上下文}}\end{array}$ $\begin{array}{cc}{'_{(i)}^{\perp}=C(s_{t-K}^{(i)},a_{t-K}^{(i)},\cdots,s_{t-2}^{(i)},a_{t-2}^{(i)},s_{t-1}^{(i)},a_{t-1}^{(i)})}\\{\text{我们使用目标函数训练上下文编码器}C\text{ 和正向动力学}}\end{array}$ $F$ , 以预测 $M$ 未来步骤中未来转换 $s_{t+ m}^{( i) }$ $( 1\leq m\leq M$ ) 中的正向动力学。多个未来步骤中的状态预测驱动潜在上下文嵌人$z_t^{(i)}$在时间上是一致的。学习的上下
|
[
[
0.08905029296875,
0.07135009765625,
0.476318359375,
0.07135009765625,
0.476318359375,
0.08209228515625,
0.08905029296875,
0.08209228515625
],
[
0.521484375,
0.0694580078125,
0.9111328125,
0.0694580078125,
0.9111328125,
0.08331298828125,
0.521484375,
0.08331298828125
],
[
0.08905029296875,
0.08331298828125,
0.476318359375,
0.08331298828125,
0.476318359375,
0.09661865234375,
0.08905029296875,
0.09661865234375
],
[
0.521484375,
0.08331298828125,
0.9111328125,
0.08331298828125,
0.9111328125,
0.09661865234375,
0.521484375,
0.09661865234375
],
[
0.0882568359375,
0.09783935546875,
0.4755859375,
0.09783935546875,
0.4755859375,
0.111083984375,
0.0882568359375,
0.111083984375
],
[
0.52294921875,
0.09783935546875,
0.9111328125,
0.09783935546875,
0.9111328125,
0.111083984375,
0.52294921875,
0.111083984375
],
[
0.0882568359375,
0.11236572265625,
0.476318359375,
0.11236572265625,
0.476318359375,
0.1256103515625,
0.0882568359375,
0.1256103515625
],
[
0.52294921875,
0.11236572265625,
0.9111328125,
0.11236572265625,
0.9111328125,
0.1256103515625,
0.52294921875,
0.1256103515625
],
[
0.08905029296875,
0.1256103515625,
0.4755859375,
0.1256103515625,
0.4755859375,
0.1395263671875,
0.08905029296875,
0.1395263671875
],
[
0.521484375,
0.1256103515625,
0.91015625,
0.1256103515625,
0.91015625,
0.1395263671875,
0.521484375,
0.1395263671875
],
[
0.0882568359375,
0.1395263671875,
0.476318359375,
0.1395263671875,
0.476318359375,
0.15283203125,
0.0882568359375,
0.15283203125
],
[
0.521484375,
0.1395263671875,
0.9111328125,
0.1395263671875,
0.9111328125,
0.15283203125,
0.521484375,
0.15283203125
],
[
0.08905029296875,
0.154052734375,
0.15283203125,
0.154052734375,
0.15283203125,
0.1654052734375,
0.08905029296875,
0.1654052734375
],
[
0.52392578125,
0.15283203125,
0.60302734375,
0.15283203125,
0.60302734375,
0.1641845703125,
0.52392578125,
0.1641845703125
],
[
0.10455322265625,
0.1666259765625,
0.47802734375,
0.1666259765625,
0.47802734375,
0.179931640625,
0.10455322265625,
0.179931640625
],
[
0.0882568359375,
0.179931640625,
0.476318359375,
0.179931640625,
0.476318359375,
0.1932373046875,
0.0882568359375,
0.1932373046875
],
[
0.6640625,
0.18310546875,
0.7685546875,
0.18310546875,
0.7685546875,
0.1956787109375,
0.6640625,
0.1956787109375
],
[
0.0882568359375,
0.1932373046875,
0.4755859375,
0.1932373046875,
0.4755859375,
0.20703125,
0.0882568359375,
0.20703125
],
[
0.08660888671875,
0.20703125,
0.476318359375,
0.205810546875,
0.476318359375,
0.2203369140625,
0.08660888671875,
0.2215576171875
],
[
0.521484375,
0.2020263671875,
0.91015625,
0.2020263671875,
0.91015625,
0.21533203125,
0.521484375,
0.21533203125
],
[
0.521484375,
0.21533203125,
0.9111328125,
0.21533203125,
0.9111328125,
0.2291259765625,
0.521484375,
0.2291259765625
],
[
0.0882568359375,
0.2215576171875,
0.47802734375,
0.2215576171875,
0.47802734375,
0.23486328125,
0.0882568359375,
0.23486328125
],
[
0.52392578125,
0.2298583984375,
0.8896484375,
0.2298583984375,
0.8896484375,
0.24365234375,
0.52392578125,
0.24365234375
],
[
0.0882568359375,
0.236083984375,
0.476318359375,
0.236083984375,
0.476318359375,
0.25,
0.0882568359375,
0.25
],
[
0.8857421875,
0.231689453125,
0.91015625,
0.231689453125,
0.91015625,
0.240478515625,
0.8857421875,
0.240478515625
],
[
0.52294921875,
0.24365234375,
0.9111328125,
0.24365234375,
0.9111328125,
0.2568359375,
0.52294921875,
0.2568359375
],
[
0.0882568359375,
0.25,
0.476318359375,
0.25,
0.476318359375,
0.26318359375,
0.0882568359375,
0.26318359375
],
[
0.52294921875,
0.2568359375,
0.91259765625,
0.2568359375,
0.91259765625,
0.270751953125,
0.52294921875,
0.270751953125
],
[
0.08905029296875,
0.26318359375,
0.47802734375,
0.26318359375,
0.47802734375,
0.276611328125,
0.08905029296875,
0.276611328125
],
[
0.52294921875,
0.270751953125,
0.9111328125,
0.270751953125,
0.9111328125,
0.2841796875,
0.52294921875,
0.2841796875
],
[
0.0882568359375,
0.27783203125,
0.47802734375,
0.27783203125,
0.47802734375,
0.291748046875,
0.0882568359375,
0.291748046875
],
[
0.52294921875,
0.2841796875,
0.9111328125,
0.2841796875,
0.9111328125,
0.29736328125,
0.52294921875,
0.29736328125
],
[
0.0882568359375,
0.291748046875,
0.47802734375,
0.291748046875,
0.47802734375,
0.304931640625,
0.0882568359375,
0.304931640625
],
[
0.52294921875,
0.298583984375,
0.9111328125,
0.298583984375,
0.9111328125,
0.3125,
0.52294921875,
0.3125
],
[
0.0882568359375,
0.304931640625,
0.4755859375,
0.304931640625,
0.4755859375,
0.318115234375,
0.0882568359375,
0.318115234375
],
[
0.0882568359375,
0.318115234375,
0.476318359375,
0.318115234375,
0.476318359375,
0.33203125,
0.0882568359375,
0.33203125
],
[
0.52294921875,
0.3125,
0.9111328125,
0.3125,
0.9111328125,
0.32568359375,
0.52294921875,
0.32568359375
],
[
0.52294921875,
0.32568359375,
0.9111328125,
0.32568359375,
0.9111328125,
0.339111328125,
0.52294921875,
0.339111328125
],
[
0.0882568359375,
0.333251953125,
0.1217041015625,
0.333251953125,
0.1217041015625,
0.34521484375,
0.0882568359375,
0.34521484375
],
[
0.52294921875,
0.34033203125,
0.90869140625,
0.34033203125,
0.90869140625,
0.354248046875,
0.52294921875,
0.354248046875
],
[
0.2034912109375,
0.359130859375,
0.359375,
0.359130859375,
0.359375,
0.37255859375,
0.2034912109375,
0.37255859375
],
[
0.52294921875,
0.354248046875,
0.9111328125,
0.354248046875,
0.9111328125,
0.367431640625,
0.52294921875,
0.367431640625
],
[
0.53173828125,
0.3662109375,
0.9111328125,
0.367431640625,
0.9111328125,
0.3818359375,
0.53173828125,
0.380615234375
],
[
0.08660888671875,
0.376953125,
0.476318359375,
0.37744140625,
0.476318359375,
0.391357421875,
0.08660888671875,
0.39013671875
],
[
0.52294921875,
0.3818359375,
0.59228515625,
0.3818359375,
0.59228515625,
0.393310546875,
0.52294921875,
0.393310546875
],
[
0.0882568359375,
0.389404296875,
0.4755859375,
0.391357421875,
0.4755859375,
0.40478515625,
0.0882568359375,
0.40283203125
],
[
0.08660888671875,
0.40478515625,
0.476318359375,
0.40478515625,
0.476318359375,
0.41845703125,
0.08660888671875,
0.41845703125
],
[
0.51953125,
0.40771484375,
0.736328125,
0.40771484375,
0.736328125,
0.421630859375,
0.51953125,
0.421630859375
],
[
0.0882568359375,
0.41845703125,
0.476318359375,
0.41845703125,
0.476318359375,
0.431884765625,
0.0882568359375,
0.431884765625
],
[
0.5361328125,
0.427978515625,
0.91259765625,
0.427978515625,
0.91259765625,
0.44140625,
0.5361328125,
0.44140625
],
[
0.0882568359375,
0.434326171875,
0.476318359375,
0.434326171875,
0.476318359375,
0.4443359375,
0.0882568359375,
0.4443359375
],
[
0.08660888671875,
0.4462890625,
0.476318359375,
0.4462890625,
0.476318359375,
0.460205078125,
0.08660888671875,
0.460205078125
],
[
0.521484375,
0.44140625,
0.91259765625,
0.44140625,
0.91259765625,
0.455078125,
0.521484375,
0.455078125
],
[
0.08905029296875,
0.4609375,
0.4755859375,
0.4609375,
0.4755859375,
0.4716796875,
0.08905029296875,
0.4716796875
],
[
0.52294921875,
0.455810546875,
0.9111328125,
0.455810546875,
0.9111328125,
0.4697265625,
0.52294921875,
0.4697265625
],
[
0.0906982421875,
0.473388671875,
0.47802734375,
0.473388671875,
0.47802734375,
0.4873046875,
0.0906982421875,
0.4873046875
],
[
0.521484375,
0.4697265625,
0.9111328125,
0.4697265625,
0.9111328125,
0.48291015625,
0.521484375,
0.48291015625
],
[
0.521484375,
0.481689453125,
0.9111328125,
0.48291015625,
0.9111328125,
0.49755859375,
0.521484375,
0.496826171875
],
[
0.08905029296875,
0.488037109375,
0.476318359375,
0.488037109375,
0.476318359375,
0.501953125,
0.08905029296875,
0.501953125
],
[
0.52392578125,
0.496826171875,
0.64208984375,
0.496826171875,
0.64208984375,
0.51025390625,
0.52392578125,
0.51025390625
],
[
0.63232421875,
0.496826171875,
0.91259765625,
0.492431640625,
0.91259765625,
0.51123046875,
0.63232421875,
0.51513671875
],
[
0.5185546875,
0.50244140625,
0.6640625,
0.50830078125,
0.66259765625,
0.52978515625,
0.51708984375,
0.5244140625
],
[
0.08905029296875,
0.501953125,
0.465576171875,
0.501953125,
0.465576171875,
0.51513671875,
0.08905029296875,
0.51513671875
],
[
0.4599609375,
0.50390625,
0.4755859375,
0.50390625,
0.4755859375,
0.51220703125,
0.4599609375,
0.51220703125
],
[
0.08660888671875,
0.5146484375,
0.47802734375,
0.51513671875,
0.47802734375,
0.52978515625,
0.08660888671875,
0.52880859375
],
[
0.654296875,
0.51025390625,
0.9111328125,
0.51025390625,
0.9111328125,
0.5244140625,
0.654296875,
0.5244140625
],
[
0.52294921875,
0.5244140625,
0.68115234375,
0.5244140625,
0.68115234375,
0.53857421875,
0.52294921875,
0.53857421875
],
[
0.6953125,
0.5234375,
0.91015625,
0.5234375,
0.91015625,
0.53857421875,
0.6953125,
0.53857421875
],
[
0.0882568359375,
0.52978515625,
0.476318359375,
0.52978515625,
0.476318359375,
0.54345703125,
0.0882568359375,
0.54345703125
],
[
0.51953125,
0.5361328125,
0.81201171875,
0.537109375,
0.81201171875,
0.5517578125,
0.51953125,
0.55029296875
],
[
0.08905029296875,
0.54248046875,
0.4755859375,
0.54248046875,
0.4755859375,
0.55615234375,
0.08905029296875,
0.55615234375
],
[
0.814453125,
0.537109375,
0.91015625,
0.5390625,
0.91015625,
0.55029296875,
0.814453125,
0.548828125
],
[
0.08905029296875,
0.556640625,
0.47802734375,
0.556640625,
0.47802734375,
0.57080078125,
0.08905029296875,
0.57080078125
],
[
0.521484375,
0.5517578125,
0.9111328125,
0.5517578125,
0.9111328125,
0.56494140625,
0.521484375,
0.56494140625
],
[
0.52392578125,
0.56640625,
0.9111328125,
0.56640625,
0.9111328125,
0.580078125,
0.52392578125,
0.580078125
],
[
0.08905029296875,
0.57275390625,
0.4755859375,
0.57275390625,
0.4755859375,
0.58349609375,
0.08905029296875,
0.58349609375
],
[
0.52294921875,
0.5791015625,
0.75732421875,
0.5791015625,
0.75732421875,
0.59326171875,
0.52294921875,
0.59326171875
],
[
0.0906982421875,
0.5859375,
0.4755859375,
0.5859375,
0.4755859375,
0.5966796875,
0.0906982421875,
0.5966796875
],
[
0.7900390625,
0.58056640625,
0.91259765625,
0.5771484375,
0.9140625,
0.5947265625,
0.79150390625,
0.5986328125
],
[
0.52392578125,
0.59228515625,
0.9111328125,
0.59228515625,
0.9111328125,
0.60693359375,
0.52392578125,
0.60693359375
],
[
0.0882568359375,
0.5986328125,
0.476318359375,
0.5986328125,
0.476318359375,
0.6123046875,
0.0882568359375,
0.6123046875
],
[
0.521484375,
0.60595703125,
0.9111328125,
0.60693359375,
0.9111328125,
0.6220703125,
0.521484375,
0.62060546875
],
[
0.0882568359375,
0.61279296875,
0.4453125,
0.61279296875,
0.4453125,
0.626953125,
0.0882568359375,
0.626953125
],
[
0.10296630859375,
0.625,
0.4755859375,
0.625,
0.4755859375,
0.63818359375,
0.10296630859375,
0.63818359375
],
[
0.521484375,
0.62060546875,
0.9111328125,
0.62060546875,
0.9111328125,
0.6337890625,
0.521484375,
0.6337890625
],
[
0.52294921875,
0.63525390625,
0.9111328125,
0.63525390625,
0.9111328125,
0.6484375,
0.52294921875,
0.6484375
],
[
0.0906982421875,
0.6416015625,
0.4755859375,
0.6416015625,
0.4755859375,
0.6513671875,
0.0906982421875,
0.6513671875
],
[
0.08905029296875,
0.65478515625,
0.4755859375,
0.65478515625,
0.4755859375,
0.66552734375,
0.08905029296875,
0.66552734375
],
[
0.52294921875,
0.6494140625,
0.9111328125,
0.6494140625,
0.9111328125,
0.66357421875,
0.52294921875,
0.66357421875
],
[
0.0882568359375,
0.66748046875,
0.476318359375,
0.66748046875,
0.476318359375,
0.68115234375,
0.0882568359375,
0.68115234375
],
[
0.521484375,
0.66357421875,
0.9111328125,
0.66357421875,
0.9111328125,
0.6767578125,
0.521484375,
0.6767578125
],
[
0.08660888671875,
0.68115234375,
0.476318359375,
0.6796875,
0.476318359375,
0.6943359375,
0.08660888671875,
0.69580078125
],
[
0.521484375,
0.67529296875,
0.79736328125,
0.67529296875,
0.79736328125,
0.689453125,
0.521484375,
0.689453125
],
[
0.08660888671875,
0.6943359375,
0.476318359375,
0.6943359375,
0.476318359375,
0.70849609375,
0.08660888671875,
0.70849609375
],
[
0.08660888671875,
0.70849609375,
0.47802734375,
0.708984375,
0.47802734375,
0.72265625,
0.08660888671875,
0.7216796875
],
[
0.51953125,
0.7041015625,
0.8505859375,
0.7041015625,
0.8505859375,
0.71728515625,
0.51953125,
0.71728515625
],
[
0.11358642578125,
0.7236328125,
0.476318359375,
0.72265625,
0.476318359375,
0.736328125,
0.11358642578125,
0.7373046875
],
[
0.521484375,
0.72265625,
0.91015625,
0.7216796875,
0.91015625,
0.736328125,
0.521484375,
0.7373046875
],
[
0.52294921875,
0.734375,
0.79736328125,
0.73291015625,
0.79736328125,
0.75048828125,
0.52294921875,
0.751953125
],
[
0.09234619140625,
0.7265625,
0.12091064453125,
0.7265625,
0.12091064453125,
0.73486328125,
0.09234619140625,
0.73486328125
],
[
0.0882568359375,
0.7373046875,
0.476318359375,
0.7373046875,
0.476318359375,
0.75048828125,
0.0882568359375,
0.75048828125
],
[
0.78759765625,
0.736328125,
0.9111328125,
0.736328125,
0.9111328125,
0.75,
0.78759765625,
0.75
],
[
0.0882568359375,
0.75048828125,
0.47802734375,
0.75048828125,
0.47802734375,
0.7646484375,
0.0882568359375,
0.7646484375
],
[
0.52294921875,
0.75048828125,
0.9111328125,
0.75048828125,
0.9111328125,
0.7646484375,
0.52294921875,
0.7646484375
],
[
0.51953125,
0.76318359375,
0.68310546875,
0.76513671875,
0.68310546875,
0.77880859375,
0.51953125,
0.7763671875
],
[
0.6748046875,
0.7646484375,
0.91015625,
0.7646484375,
0.91015625,
0.77783203125,
0.6748046875,
0.77783203125
],
[
0.515625,
0.7734375,
0.9150390625,
0.77587890625,
0.9150390625,
0.80029296875,
0.515625,
0.798828125
],
[
0.08905029296875,
0.7666015625,
0.476318359375,
0.7666015625,
0.476318359375,
0.7763671875,
0.08905029296875,
0.7763671875
],
[
0.08905029296875,
0.77880859375,
0.476318359375,
0.77880859375,
0.476318359375,
0.79248046875,
0.08905029296875,
0.79248046875
],
[
0.08660888671875,
0.79052734375,
0.476318359375,
0.79150390625,
0.476318359375,
0.8046875,
0.08660888671875,
0.8037109375
],
[
0.51708984375,
0.7861328125,
0.91259765625,
0.7890625,
0.91259765625,
0.81103515625,
0.51708984375,
0.80810546875
],
[
0.08905029296875,
0.80615234375,
0.476318359375,
0.80615234375,
0.476318359375,
0.8193359375,
0.08905029296875,
0.8193359375
],
[
0.5458984375,
0.806640625,
0.7255859375,
0.8046875,
0.7255859375,
0.818359375,
0.5458984375,
0.82080078125
],
[
0.74169921875,
0.8017578125,
0.91259765625,
0.8037109375,
0.91259765625,
0.82275390625,
0.74169921875,
0.82080078125
],
[
0.525390625,
0.8125,
0.60888671875,
0.82275390625,
0.60546875,
0.8369140625,
0.52294921875,
0.8271484375
],
[
0.74658203125,
0.81640625,
0.76953125,
0.81640625,
0.76953125,
0.82373046875,
0.74658203125,
0.82373046875
],
[
0.79150390625,
0.8173828125,
0.79833984375,
0.8173828125,
0.79833984375,
0.82080078125,
0.79150390625,
0.82080078125
],
[
0.8349609375,
0.8193359375,
0.8388671875,
0.8193359375,
0.8388671875,
0.82470703125,
0.8349609375,
0.82470703125
],
[
0.8388671875,
0.82080078125,
0.85107421875,
0.82080078125,
0.85107421875,
0.8291015625,
0.8388671875,
0.8291015625
],
[
0.7158203125,
0.818359375,
0.7294921875,
0.818359375,
0.7294921875,
0.82275390625,
0.7158203125,
0.82275390625
],
[
0.08905029296875,
0.8212890625,
0.4755859375,
0.8212890625,
0.4755859375,
0.83203125,
0.08905029296875,
0.83203125
],
[
0.65283203125,
0.8212890625,
0.71728515625,
0.8212890625,
0.71728515625,
0.83203125,
0.65283203125,
0.83203125
],
[
0.76708984375,
0.8212890625,
0.83642578125,
0.8212890625,
0.83642578125,
0.83349609375,
0.76708984375,
0.83349609375
],
[
0.7294921875,
0.82275390625,
0.75634765625,
0.82470703125,
0.75634765625,
0.83056640625,
0.7294921875,
0.8291015625
],
[
0.609375,
0.82275390625,
0.654296875,
0.82275390625,
0.654296875,
0.83349609375,
0.609375,
0.83349609375
],
[
0.08905029296875,
0.833984375,
0.476318359375,
0.833984375,
0.476318359375,
0.84765625,
0.08905029296875,
0.84765625
],
[
0.52294921875,
0.83349609375,
0.9111328125,
0.83349609375,
0.9111328125,
0.84765625,
0.52294921875,
0.84765625
],
[
0.0882568359375,
0.84765625,
0.443603515625,
0.84765625,
0.443603515625,
0.861328125,
0.0882568359375,
0.861328125
],
[
0.521484375,
0.84521484375,
0.80810546875,
0.84765625,
0.80810546875,
0.8623046875,
0.521484375,
0.85986328125
],
[
0.81787109375,
0.8486328125,
0.91015625,
0.8486328125,
0.91015625,
0.859375,
0.81787109375,
0.859375
],
[
0.10296630859375,
0.861328125,
0.476318359375,
0.861328125,
0.476318359375,
0.87548828125,
0.10296630859375,
0.87548828125
],
[
0.52685546875,
0.86279296875,
0.5390625,
0.86279296875,
0.5390625,
0.8720703125,
0.52685546875,
0.8720703125
],
[
0.5361328125,
0.85986328125,
0.67236328125,
0.861328125,
0.67236328125,
0.87548828125,
0.5361328125,
0.875
],
[
0.66845703125,
0.86572265625,
0.67333984375,
0.86572265625,
0.67333984375,
0.87060546875,
0.66845703125,
0.87060546875
],
[
0.67333984375,
0.86279296875,
0.91015625,
0.86279296875,
0.91015625,
0.87353515625,
0.67333984375,
0.87353515625
],
[
0.0882568359375,
0.87548828125,
0.476318359375,
0.87548828125,
0.476318359375,
0.8896484375,
0.0882568359375,
0.8896484375
],
[
0.51953125,
0.875,
0.9111328125,
0.87353515625,
0.9111328125,
0.88818359375,
0.51953125,
0.8896484375
],
[
0.65869140625,
0.876953125,
0.6640625,
0.8798828125,
0.65869140625,
0.88623046875,
0.65283203125,
0.88330078125
]
] |
[
"配中任何一对源和目标任务的策略转移（参见图。1）。",
"2020a；Agarwaletal.2021）研究了多个MDP上状态",
"如果在最近的事件中，转移的策略在目标任务上获得",
"的双性度量，并考虑如何学习导致不同MDP中同一动",
"的奖励高于学习的策略，则代理将按照转移的策略执行",
"作下几乎相同行为的状态的表示。在多任务强化学习和",
"操作。这种方法使代理能够利用来自多个训练任务的相",
"元强化学习问题中，Zhangetal.（2020a,b）基于任务和",
"关数据，鼓励学习的策略在多个训练任务上同样表现良",
"状态相似性导出迁移和泛化边界。我们还限制了跨任务",
"好，从而在将训练有素的策略应用于测试任务时带来更",
"的策略传输的价值，但我们的方法是建立动作等效而不",
"好的性能",
"是状态等效",
"我们总结了贡献：（1）我们引人了一种新颖的目标",
"函数，将任何策略从源马尔可夫决策过程（MDP）转",
"3Method",
"移到目标MDP。我们证明了二个理论保证，即转移的",
"策略可以在源MDP上实现接近源策略的目标MDP",
"在本节中，我们首先描述我们的方法，以学习捕获任务",
"特征的上下文编码器，并学习在给定任务上下文的情况",
"的预期收益。预期收益的差异（近似）上界由我们的",
"下预测下一个状态分布的前向动力学模型（Sec。",
"损失函数与常数乘法因子。（2）我们开发了一种名为",
"3.2",
"）。然后，我们引人一个目标函数来训练动作翻译器，以",
"Meta-RL的非策略RL算法，该算法具有Context条",
"便目标任务上的翻译动作与源任务上的源动作等效。动",
"件的ActionTranslator（MCAT），在元RL中应用了策",
"作翻译器可以以任务上下文为条件，因此它可以将一个",
"略传输机制，以帮助跨多个稀疏奖励任务进行探索。（3）",
"好的策略从任何任意源任务转移到训练集中的任何其",
"我们通过MuJoCo物理引擎（Todorov,Erez，andTassa",
"他目标任务（Sec。3.3）。最后，我们建议将动作翻译",
"2012）实证证明了MCAT在各种模拟控制任务上的有",
"效性，表明策略传输提高了基于上下文的元RL算法的",
"器与基于上下文的元RL算法相结合，以将良好的策略",
"从任何一个任务转移到其他任务。在元训练期间，这种",
"性能",
"策略传输方法有助于利用在任何一项任务上遇到的良",
"2Related Work",
"好经验，并有利于数据收集和对其他稀疏奖励任务的进",
"步策略优化（Sec。3.4）。Fig。2概述了我们的方",
"基于上下文的元RL元强化学习已在文献（Finn",
"法MCAT",
"Abbeel, and Levine 2017:; Stadie et al. 2018; Sung et al",
"2017;Xu，vanHasselt，andSilver2018）中进行了广泛的",
"3.1Problem Formulation",
"研究，其中许多作品开发了基于上下文的方法（Rakelly",
"初始分布遵循先前工作（Duanetal.2016;Mishra",
"et al.2019:Ren et al.2020:Liu et al.2020).Duan",
"et al. (2016); Wang et al. (2016); Fako0r et al. (2019)",
"etal.2017;Rakellyetal.2019）中的元RL公式，我",
"采用递归神经网络对上下文转换进行编码，并在上下",
"们假设任务p(T）的分布，并且每个任务是定义为具有",
"文变量上制定策略条件。最大化预期收益的目标函数联",
"状态空间S的元组(S,A,P,r,%Po）的马尔可夫决策过",
"程（MDP），动作空间X。我们也可以将奖励函数定",
"合训练上下文编码器和策略。Rakellyetal.（2019）利",
"义为 r(s,a) =",
"Cssp(ssa)r(sa,s）。在基于上下文",
"的RL算法中",
"用置换不变编码器将经验聚合为概率上下文变量，",
"并",
"通过变分推理对其进行优化。后验采样有利于在适应",
"我们学习了为任何任务T(i）～p(T)",
"共享的策略(st),zt",
"），其中t表示情节中的时间",
"阶段探索稀疏奖励任务，但在训练阶段可以获得密集奖",
"步长，i表示任务的索引，上下文变量z",
"励。Li,Pinto,and Abbeel(2020)考虑了奖励函数的任",
"EZ从任务",
"务族。Leeetal.（2020)；Seoetal.（2020）使用前向动态",
"MDP的历史转换中捕获上下文信息，而2是上下文向",
"量的空间。对共享策略进行了优化，以使其在每个训练",
"预测来训练上下文编码器。这些基于模型的元RL算法",
"任务 T(i) 上的值 V(T(i))=E(",
"假设奖励函数可用于计划。在没有地面实况奖励函数的",
"=0最",
"大化。遵循meta-RL(Yu et alP2017;Nagabandi et al.",
"稀疏奖励设置中，他们可能难以发现非零奖励，并且准",
"2018;Nagabandi,Finn,and Levine 2018;Zhou,Pinto",
"确地估计基于模型的计划的奖励也可能是有问题的",
"RL中的策略转移策略转移研究给定一组源任务",
"andGupta2019:Leeetal.2020）中的先前工作，我们",
"研究具有相同状态空间，动作空间，奖励函数但动态函",
"及其专家策略的目标任务中的知识转移。策略提炼",
"(Rusu et al. 2015: Yin and Pan 2017: Parisotto. Ba",
"数变化的任务。重要的是，我们专注于更具挑战性的稀",
"andSalakhutdinov2015）最小化目标任务上的源策略",
"疏奖励设置。我们的目标是学习一种对动态变化稳健的",
"和学习策略之间的动作分布差异。沿着这条工作线，Teh",
"共享策略，并将其推广到看不见的任务",
"etal.（2017)在多任务强化学习中创建了一个质心策略",
"并将知识从特定于任务的策略中提取到该质心策略。或",
"3.2Learning Context & Forward Model",
"源和目标任务（Zhu，Lin，andZhou2020）之间的",
"为了捕获有关任何任务T(i）的知识，我们利用",
"上下文编码器C:SK×AK→2",
"者，",
"任务间映射可以帮助策略传输。这些作品（Guptaetal.",
"其中K是",
"2017; Konidaris and Barto 2006; Ammar and Taylor",
"用于推断上下文的过去步骤数。（Rakellyet al",
"2019:Zhou,Pinto,",
"and Gupta 20i9; Lee et al.",
"2020)）对相关思想进行了探索。igFig。E3a给",
"2011）中的大多数都假设在状态空间上存在对应关系，",
"并学习任务之间的状态映射。最近的工作（Zhangetal.",
"2020c）学习了具有动态循环一致性损失的状态对应关",
"定k过会的转换kaoka",
"系和动作对应关系。我们的方法与这种方法的不同之处",
"上下文编码器C",
"产星潜在工宇文",
"())=C(st)",
"（i",
"-",
"",
"",
"",
"在于，我们可以使用更简单的目标函数在多个任务之间",
"K.....s.",
"2, st1,at2",
"",
"K,a",
"进行动作转换。重要的是，我们的方法是新颖的，可以",
"我们使用自标函数训练上下文编码器C和正向动力学",
"将策略转移用于元RL中的任何一对源和目标任务",
"F，以预测M未来步骤中未来转换st+m",
"(1<m<M",
"MDP中状态的双模拟最近关于状态表示学习的工",
"",
"中的正向动力学。",
"",
"多个未来步骤中的状态预测驱动",
"作 (Ferns, Panangaden, and Precup 2004; Zhang et al.",
"潜在上下文嵌人2（）在时间上是一致的。学习的上下",
""
] |
{
"arxiv_id": "2207.09071",
"language": "zh",
"page_idx": 2
}
|
||
2309.01327-zh-0
|
2309.01327-zh-0.jpg
|
Can I Trust Your Answer? Visually Grounded Video Question Answering
Junbin Xiao Angela Yao* Yicong Li Tat-Seng Chua
Department of Computer Science, National University of Singapore
$\{$ junbin,ayao,chuats $\}$ @comp.nus.edu.sg, liyicong@u.nus.edu
<FigureHere>
Figure 1.
Top: Real predictions of VQA models (BlindQA, SigFQA and SoTA) on NExT-QA [55]. All the models
correctly answer Q1 but wrongly answer Q2 , albeit the two questions sharing visual evidence( the boy unwraps the present and throws the wrapping paper ). Bottom : Overlap in model predictions. BlindQA: A pure language model ( $\hat{\text{i.e. RoBERTa [36] )fne-tuned with question-answer text. SigFQA: An image-text model ( i.e. CLIP [44] ) using only the}}$ center video frame. SoTA: Temp[CLIP] (Sec. 5) model using 32 video frames. The analyses indicate that the models may not learn from causal visual content but more likely from language short-cut and irrelevant visual context.
## Abstract
a grounded-QA method via Gaussian mask optimization and cross-modal learning. Experiments with different backbones demonstrate that this grounding mechanism improves both grounding and QA. With these efforts, we aim to push towards trustworthy VLMs in VQA systems. Our dataset and code are available at https://github.com/doc-doc/NExT-GQA~.
## 1. Introduction
We study visually grounded VideoQA in response to the emerging trends of utilizing pretraining techniques for video-language understanding. Specifically, by forcing vision-language models (VLMs) to answer questions and simultaneously provide visual evidence, we seek to ascertain the extent to which the predictions of such techniques are genuinely anchored in relevant video content, versus spurious correlations from language or irrelevant visual context. Towards this, we construct NExT-GQA- an extension of NExT-QA with 10.5 $K$ temporal grounding (or location) labels tied to the original QA pairs. With NExT-GQA, we scrutinize a series of state-of-the-art VLMs. Through post-hoc attention analysis, we find that these models are extremely weak in substantiating the answers despite their strong QA performance. This exposes the limitation of current VLMs in making reliable predictions. As a remedy, we further explore and propose
视频问答(VideoQA)最近已成为开发视觉语言模型 (VLM) 的黄金测试平台，尤其是在多模态网络语料库[1,11,24,29,51,61,63,66]上大规模预训练的基础VLM。尽管 QA 性能有了长足的进步，但仍存在一个基本问题-此类技术的答案是否或在多大程度上基于相关的视觉内容？或者，他们是依靠语言捷径来使用强大的语言模型[20,34,42,52,62,64,69,70,73]还是通过跨模式预训练[45,60]捕获的虚假视觉语言相关性？
例如，图。1(上图)显示，现有的 VLM 倾向于
|
[
[
0.11517333984375,
0.1319580078125,
0.85693359375,
0.1331787109375,
0.85693359375,
0.15087890625,
0.11517333984375,
0.149658203125
],
[
0.24267578125,
0.1817626953125,
0.58251953125,
0.1817626953125,
0.58251953125,
0.1962890625,
0.24267578125,
0.1962890625
],
[
0.5908203125,
0.1875,
0.59716796875,
0.1875,
0.59716796875,
0.1925048828125,
0.5908203125,
0.1925048828125
],
[
0.59326171875,
0.184326171875,
0.7294921875,
0.184326171875,
0.7294921875,
0.1976318359375,
0.59326171875,
0.1976318359375
],
[
0.2017822265625,
0.2008056640625,
0.7685546875,
0.2008056640625,
0.7685546875,
0.21533203125,
0.2017822265625,
0.21533203125
],
[
0.281982421875,
0.2197265625,
0.68896484375,
0.2171630859375,
0.68896484375,
0.231689453125,
0.281982421875,
0.2342529296875
],
[
0.0882568359375,
0.292236328125,
0.3203125,
0.29345703125,
0.3203125,
0.304931640625,
0.0882568359375,
0.3037109375
],
[
0.6337890625,
0.29345703125,
0.69677734375,
0.29345703125,
0.69677734375,
0.3037109375,
0.6337890625,
0.3037109375
],
[
0.7998046875,
0.29345703125,
0.861328125,
0.29345703125,
0.861328125,
0.3037109375,
0.7998046875,
0.3037109375
],
[
0.7236328125,
0.29736328125,
0.73876953125,
0.29736328125,
0.73876953125,
0.30615234375,
0.7236328125,
0.30615234375
],
[
0.330810546875,
0.29736328125,
0.374267578125,
0.29736328125,
0.374267578125,
0.3017578125,
0.330810546875,
0.3017578125
],
[
0.0882568359375,
0.3037109375,
0.3447265625,
0.3037109375,
0.3447265625,
0.317626953125,
0.0882568359375,
0.317626953125
],
[
0.431396484375,
0.298583984375,
0.59716796875,
0.298583984375,
0.59716796875,
0.3359375,
0.431396484375,
0.3359375
],
[
0.63134765625,
0.304931640625,
0.70361328125,
0.304931640625,
0.70361328125,
0.31494140625,
0.63134765625,
0.31494140625
],
[
0.78369140625,
0.304931640625,
0.875,
0.304931640625,
0.875,
0.31494140625,
0.78369140625,
0.31494140625
],
[
0.7294921875,
0.317626953125,
0.74267578125,
0.323974609375,
0.736328125,
0.33203125,
0.72216796875,
0.32568359375
],
[
0.771484375,
0.31640625,
0.83251953125,
0.31640625,
0.83251953125,
0.326904296875,
0.771484375,
0.326904296875
],
[
0.8349609375,
0.3193359375,
0.85546875,
0.3193359375,
0.85546875,
0.323974609375,
0.8349609375,
0.323974609375
],
[
0.336669921875,
0.3212890625,
0.371826171875,
0.3212890625,
0.371826171875,
0.328857421875,
0.336669921875,
0.328857421875
],
[
0.0906982421875,
0.3212890625,
0.2296142578125,
0.3212890625,
0.2296142578125,
0.328857421875,
0.0906982421875,
0.328857421875
],
[
0.24267578125,
0.322509765625,
0.2998046875,
0.322509765625,
0.2998046875,
0.330078125,
0.24267578125,
0.330078125
],
[
0.6357421875,
0.330078125,
0.7060546875,
0.330078125,
0.7060546875,
0.34033203125,
0.6357421875,
0.34033203125
],
[
0.78515625,
0.32763671875,
0.87841796875,
0.32763671875,
0.87841796875,
0.339111328125,
0.78515625,
0.339111328125
],
[
0.1650390625,
0.35986328125,
0.250732421875,
0.35986328125,
0.250732421875,
0.37060546875,
0.1650390625,
0.37060546875
],
[
0.0882568359375,
0.33203125,
0.366943359375,
0.33203125,
0.366943359375,
0.34228515625,
0.0882568359375,
0.34228515625
],
[
0.361083984375,
0.3623046875,
0.455810546875,
0.3623046875,
0.455810546875,
0.369384765625,
0.361083984375,
0.369384765625
],
[
0.5595703125,
0.3623046875,
0.65869140625,
0.3623046875,
0.65869140625,
0.369384765625,
0.5595703125,
0.369384765625
],
[
0.7607421875,
0.3623046875,
0.8544921875,
0.3623046875,
0.8544921875,
0.369384765625,
0.7607421875,
0.369384765625
],
[
0.325927734375,
0.371826171875,
0.36767578125,
0.371826171875,
0.36767578125,
0.3818359375,
0.325927734375,
0.3818359375
],
[
0.09722900390625,
0.365478515625,
0.1217041015625,
0.365478515625,
0.1217041015625,
0.376953125,
0.09722900390625,
0.376953125
],
[
0.25244140625,
0.37255859375,
0.281982421875,
0.37255859375,
0.281982421875,
0.380126953125,
0.25244140625,
0.380126953125
],
[
0.4453125,
0.371826171875,
0.484375,
0.371826171875,
0.484375,
0.3818359375,
0.4453125,
0.3818359375
],
[
0.640625,
0.37255859375,
0.67724609375,
0.37255859375,
0.67724609375,
0.380126953125,
0.640625,
0.380126953125
],
[
0.73046875,
0.37255859375,
0.771484375,
0.37255859375,
0.771484375,
0.380126953125,
0.73046875,
0.380126953125
],
[
0.84619140625,
0.375732421875,
0.8837890625,
0.375732421875,
0.8837890625,
0.386474609375,
0.84619140625,
0.386474609375
],
[
0.52783203125,
0.375,
0.57177734375,
0.375,
0.57177734375,
0.385009765625,
0.52783203125,
0.385009765625
],
[
0.098876953125,
0.376953125,
0.1217041015625,
0.376953125,
0.1217041015625,
0.389404296875,
0.098876953125,
0.389404296875
],
[
0.1968994140625,
0.38330078125,
0.2181396484375,
0.38330078125,
0.2181396484375,
0.391357421875,
0.1968994140625,
0.391357421875
],
[
0.38232421875,
0.38818359375,
0.414306640625,
0.38818359375,
0.414306640625,
0.395751953125,
0.38232421875,
0.395751953125
],
[
0.6298828125,
0.386962890625,
0.64794921875,
0.386962890625,
0.64794921875,
0.39453125,
0.6298828125,
0.39453125
],
[
0.438720703125,
0.389404296875,
0.464111328125,
0.393310546875,
0.463134765625,
0.40087890625,
0.435546875,
0.397705078125
],
[
0.098876953125,
0.389404296875,
0.1217041015625,
0.389404296875,
0.1217041015625,
0.401611328125,
0.098876953125,
0.401611328125
],
[
0.7958984375,
0.391357421875,
0.81640625,
0.391357421875,
0.81640625,
0.3984375,
0.7958984375,
0.3984375
],
[
0.08660888671875,
0.399658203125,
0.1217041015625,
0.401611328125,
0.12091064453125,
0.4140625,
0.0849609375,
0.412353515625
],
[
0.3447265625,
0.399658203125,
0.370849609375,
0.399658203125,
0.370849609375,
0.40771484375,
0.3447265625,
0.40771484375
],
[
0.54833984375,
0.40283203125,
0.560546875,
0.40283203125,
0.560546875,
0.4091796875,
0.54833984375,
0.4091796875
],
[
0.08660888671875,
0.410888671875,
0.12335205078125,
0.410888671875,
0.12335205078125,
0.4267578125,
0.08660888671875,
0.4267578125
],
[
0.75341796875,
0.4072265625,
0.771484375,
0.4072265625,
0.771484375,
0.410400390625,
0.75341796875,
0.410400390625
],
[
0.39453125,
0.4140625,
0.419189453125,
0.4140625,
0.419189453125,
0.421630859375,
0.39453125,
0.421630859375
],
[
0.5947265625,
0.4140625,
0.62109375,
0.4140625,
0.62109375,
0.421630859375,
0.5947265625,
0.421630859375
],
[
0.6298828125,
0.42236328125,
0.65283203125,
0.42236328125,
0.65283203125,
0.429931640625,
0.6298828125,
0.429931640625
],
[
0.80224609375,
0.419189453125,
0.826171875,
0.419189453125,
0.826171875,
0.423583984375,
0.80224609375,
0.423583984375
],
[
0.0906982421875,
0.427978515625,
0.12091064453125,
0.427978515625,
0.12091064453125,
0.436279296875,
0.0906982421875,
0.436279296875
],
[
0.438720703125,
0.423583984375,
0.453369140625,
0.423583984375,
0.453369140625,
0.427978515625,
0.438720703125,
0.427978515625
],
[
0.56396484375,
0.431884765625,
0.57421875,
0.431884765625,
0.57421875,
0.43505859375,
0.56396484375,
0.43505859375
],
[
0.10296630859375,
0.442626953125,
0.11767578125,
0.442626953125,
0.11767578125,
0.4462890625,
0.10296630859375,
0.4462890625
],
[
0.1380615234375,
0.43994140625,
0.148681640625,
0.438232421875,
0.1502685546875,
0.442626953125,
0.1396484375,
0.4443359375
],
[
0.098876953125,
0.451904296875,
0.1217041015625,
0.451904296875,
0.1217041015625,
0.46337890625,
0.098876953125,
0.46337890625
],
[
0.40283203125,
0.462158203125,
0.431396484375,
0.462158203125,
0.431396484375,
0.473388671875,
0.40283203125,
0.473388671875
],
[
0.8037109375,
0.46337890625,
0.83251953125,
0.46337890625,
0.83251953125,
0.4716796875,
0.8037109375,
0.4716796875
],
[
0.59716796875,
0.464111328125,
0.6259765625,
0.464111328125,
0.6259765625,
0.472900390625,
0.59716796875,
0.472900390625
],
[
0.1004638671875,
0.464111328125,
0.12091064453125,
0.464111328125,
0.12091064453125,
0.472900390625,
0.1004638671875,
0.472900390625
],
[
0.1298828125,
0.472900390625,
0.1641845703125,
0.472900390625,
0.1641845703125,
0.479736328125,
0.1298828125,
0.479736328125
],
[
0.1895751953125,
0.47021484375,
0.226318359375,
0.47021484375,
0.226318359375,
0.48095703125,
0.1895751953125,
0.48095703125
],
[
0.256591796875,
0.472900390625,
0.281005859375,
0.472900390625,
0.281005859375,
0.479736328125,
0.256591796875,
0.479736328125
],
[
0.08087158203125,
0.494384765625,
0.143798828125,
0.494384765625,
0.143798828125,
0.5048828125,
0.08087158203125,
0.5048828125
],
[
0.1683349609375,
0.492431640625,
0.89208984375,
0.492431640625,
0.89208984375,
0.505859375,
0.1683349609375,
0.505859375
],
[
0.080078125,
0.5068359375,
0.89208984375,
0.5068359375,
0.89208984375,
0.52099609375,
0.080078125,
0.52099609375
],
[
0.080078125,
0.52099609375,
0.890625,
0.52099609375,
0.890625,
0.5341796875,
0.080078125,
0.5341796875
],
[
0.080078125,
0.5341796875,
0.890625,
0.5341796875,
0.890625,
0.54736328125,
0.080078125,
0.54736328125
],
[
0.07843017578125,
0.54736328125,
0.890625,
0.54736328125,
0.890625,
0.56103515625,
0.07843017578125,
0.56103515625
],
[
0.080078125,
0.56103515625,
0.78369140625,
0.5625,
0.78369140625,
0.57568359375,
0.080078125,
0.57470703125
],
[
0.238525390625,
0.5947265625,
0.31201171875,
0.5947265625,
0.31201171875,
0.609375,
0.238525390625,
0.609375
],
[
0.498291015625,
0.5947265625,
0.8935546875,
0.59326171875,
0.8935546875,
0.611328125,
0.498291015625,
0.6123046875
],
[
0.50244140625,
0.611328125,
0.8935546875,
0.611328125,
0.8935546875,
0.625,
0.50244140625,
0.625
],
[
0.09722900390625,
0.62890625,
0.4697265625,
0.6298828125,
0.4697265625,
0.64453125,
0.09722900390625,
0.64306640625
],
[
0.50244140625,
0.626953125,
0.89208984375,
0.626953125,
0.89208984375,
0.64013671875,
0.50244140625,
0.64013671875
],
[
0.080078125,
0.64599609375,
0.468017578125,
0.64599609375,
0.468017578125,
0.6591796875,
0.080078125,
0.6591796875
],
[
0.5009765625,
0.6416015625,
0.89208984375,
0.64013671875,
0.89208984375,
0.65478515625,
0.5009765625,
0.65576171875
],
[
0.080078125,
0.66015625,
0.468017578125,
0.66015625,
0.468017578125,
0.673828125,
0.080078125,
0.673828125
],
[
0.50244140625,
0.6572265625,
0.89208984375,
0.6572265625,
0.89208984375,
0.67041015625,
0.50244140625,
0.67041015625
],
[
0.07843017578125,
0.6748046875,
0.468017578125,
0.673828125,
0.468017578125,
0.68798828125,
0.07843017578125,
0.689453125
],
[
0.5009765625,
0.67041015625,
0.89208984375,
0.671875,
0.89208984375,
0.68603515625,
0.5009765625,
0.68505859375
],
[
0.080078125,
0.68994140625,
0.468017578125,
0.68994140625,
0.468017578125,
0.7041015625,
0.080078125,
0.7041015625
],
[
0.5009765625,
0.68603515625,
0.81396484375,
0.68505859375,
0.81396484375,
0.69970703125,
0.5009765625,
0.70068359375
],
[
0.08087158203125,
0.70703125,
0.468017578125,
0.70703125,
0.468017578125,
0.71728515625,
0.08087158203125,
0.71728515625
],
[
0.080078125,
0.72021484375,
0.4697265625,
0.72021484375,
0.4697265625,
0.73486328125,
0.080078125,
0.73486328125
],
[
0.08087158203125,
0.7373046875,
0.46728515625,
0.7373046875,
0.46728515625,
0.74755859375,
0.08087158203125,
0.74755859375
],
[
0.50341796875,
0.7373046875,
0.6396484375,
0.7373046875,
0.6396484375,
0.751953125,
0.50341796875,
0.751953125
],
[
0.07843017578125,
0.75048828125,
0.46728515625,
0.75,
0.46728515625,
0.7646484375,
0.07843017578125,
0.76513671875
],
[
0.080078125,
0.76513671875,
0.46728515625,
0.7646484375,
0.46728515625,
0.77880859375,
0.080078125,
0.77978515625
],
[
0.52392578125,
0.76318359375,
0.890625,
0.76318359375,
0.890625,
0.77783203125,
0.52392578125,
0.77783203125
],
[
0.080078125,
0.78076171875,
0.4697265625,
0.78076171875,
0.4697265625,
0.79443359375,
0.080078125,
0.79443359375
],
[
0.50244140625,
0.77783203125,
0.890625,
0.77880859375,
0.890625,
0.79345703125,
0.50244140625,
0.79248046875
],
[
0.080078125,
0.79541015625,
0.4697265625,
0.79541015625,
0.4697265625,
0.81005859375,
0.080078125,
0.81005859375
],
[
0.50341796875,
0.79345703125,
0.890625,
0.79345703125,
0.890625,
0.806640625,
0.50341796875,
0.806640625
],
[
0.07843017578125,
0.81103515625,
0.468017578125,
0.8095703125,
0.468017578125,
0.82275390625,
0.07843017578125,
0.82470703125
],
[
0.50341796875,
0.8095703125,
0.890625,
0.8095703125,
0.890625,
0.82275390625,
0.50341796875,
0.82275390625
],
[
0.07843017578125,
0.82568359375,
0.4697265625,
0.82470703125,
0.4697265625,
0.8388671875,
0.07843017578125,
0.84033203125
],
[
0.50341796875,
0.82470703125,
0.890625,
0.82470703125,
0.890625,
0.83837890625,
0.50341796875,
0.83837890625
],
[
0.08087158203125,
0.84228515625,
0.468017578125,
0.84228515625,
0.468017578125,
0.85302734375,
0.08087158203125,
0.85302734375
],
[
0.50341796875,
0.8388671875,
0.890625,
0.8388671875,
0.890625,
0.85302734375,
0.50341796875,
0.85302734375
],
[
0.080078125,
0.85595703125,
0.4697265625,
0.85595703125,
0.4697265625,
0.869140625,
0.080078125,
0.869140625
],
[
0.50341796875,
0.85400390625,
0.89208984375,
0.85400390625,
0.89208984375,
0.86865234375,
0.50341796875,
0.86865234375
],
[
0.080078125,
0.8720703125,
0.4697265625,
0.8720703125,
0.4697265625,
0.88525390625,
0.080078125,
0.88525390625
],
[
0.50341796875,
0.869140625,
0.8486328125,
0.869140625,
0.8486328125,
0.88330078125,
0.50341796875,
0.88330078125
],
[
0.080078125,
0.88720703125,
0.4697265625,
0.88720703125,
0.4697265625,
0.90087890625,
0.080078125,
0.90087890625
],
[
0.52392578125,
0.88623046875,
0.8896484375,
0.88623046875,
0.8896484375,
0.90087890625,
0.52392578125,
0.90087890625
]
] |
[
"Can I Trust Your Answer? Visually Grounded Video Question Answering",
"Junbin XiaoAngela Yao*Yicong Li",
"",
"Tat-Seng Chua",
"Department of Computer Science, National University of Singapore",
"[ junbin,ayao,chuats } @comp.nus.edu.sg, liyicong@u.nus.edu",
"Qi.Whydid thebabypickup onepre",
"Unwrapit",
"Unwrapit",
"V",
"pnttror",
"the group of them and move to the sofa?",
"VLM",
"(Prediction)",
"(Ground Truth)",
"+",
"Boy threv",
"",
"-ontb",
"O2.Why wastlheretor",
"wrapping",
"(Prediction)",
"(Ground Truth)",
"QAPerformance",
"grotnd near the sofa at the end of the video?",
"OverallPredictions",
"CorrectPredictions",
"Wrong Predictions",
"BlindQA",
"65.0",
"63.2",
"SigFQA",
"SigFQA",
"BlindQA",
"SigFQA",
"BlindQA",
"62.5",
"60.6",
"10.6%",
"4.2",
"13.49",
"60.0",
"6.8%",
"$57.5",
"26.9%",
".1",
" 55.0",
"4",
"53.89",
"37.8%",
"14.9%",
"24.8%",
"52.5",
"2.2",
"",
"50.0",
"1o",
"47.5",
"SoTA",
"SoTA",
"SoTA",
"45.0",
"BlindQA",
"SigFQA",
"SoTA",
"Figure 1.",
"Top : Real predictions of VQA models (BlindQA, SigFQA and SoTA) on NExT-QA [55] : All the models",
"correctly answer Q1 but wrongly answer Q2 , albeit the two questions sharing visual evidence ( the boy unwraps the",
"present and throws the wrapping paper ) Bottom : Overlap in model predictions. BlindQA: A pure language model (",
"i.e. RoBERTa [36] ) fine-tuned with question-answer text. SigFQA: An image-text model (i.e. CLIP [44] ) using only the",
"center video frame. SoTA: Temp[CLIP) (Sec. 5 ) model using 32 video frames. The analyses indicate that the models may",
"not learn from causal visual content but more likely from language short-cut and irrelevant visual context.",
"Abstract",
"a grounded-QA method via Gaussian mask optimiza-",
"tion and cross-modal learning. Experiments with differ-",
"We study visually grounded VideoQA in response",
"ent backbones demonstrate that this grounding mech-",
"to the emerging trends of utilizing pretraining tech-",
"anism improves both grounding and QA. With these",
"niques for video-language understanding. Specifically,",
"efforts, we aim to push towards trustworthy VLMs in",
"by forcing vision-language models (VLMs) to answer",
"VQA systems. Our dataset and code are available at",
"questions and simultaneously provide visual evidence,",
"https://github.com/doc-doc/NExT-GQA.",
"we seek to ascertain the extent to which the predic",
"tions of such techniques are genuinely anchored in rel-",
"evant video content.versus spurious correlations from",
"1.Introduction",
"language or irrelevant visual context..Towards this,",
"we construct NExT-GQA - an extension of NExT-QA",
"视频问答（VideoQA）最近已成为开发视觉语言模",
"with 10.5 K temporal grounding (or location) labels",
"型（VLM）的黄金测试平台，尤其是在多模态网络语",
"tied to the original QA pairs. With NExT-GQA, we",
"料库[1,11,24,29,51,61,63,66]上大规模预训练的基础",
"scrutinize a series of state-of-the-art VLMs. Through",
"VLM。尽管QA性能有了长足的进步，但仍存在一个",
"post-hoc attention analysis, we find that these models",
"基本问题-此类技术的答案是否或在多大程度上基于相",
"are extremely weak in substantiating the answers de-",
"关的视觉内容？或者，他们是依靠语言捷径来使用强大",
"spite their strong QA performance. This exposes the",
"的语言模型[20,34,42,52,62,64,69,70,73]还是通过跨",
"limitation of current VLMs in making reliable predic-",
"模式预训练[45,60]捕获的虚假视觉语言相关性？?",
"tions. As a remedy, we further explore and propose",
"例如，图。1（上图）显示，现有的VLM倾向于"
] |
{
"arxiv_id": "2309.01327",
"language": "zh",
"page_idx": 0
}
|
||
2309.01327-zh-3
|
2309.01327-zh-3.jpg
|
也可以由所有问题的全局答案集 [59] 给出 (开放式)。请注意，SoTA 变压器-方法$\mathbb{E}[11,57,61,62]$ 以统一的公式制定并解决了多项选择 QA 和开放式 QA:
$$a^*=\mathrm{argmax}_{a\in A}\:\Psi(a|v,q,A),$$
(1)
,其中映射$\Psi$通常实现为共享$\left[24,50\right]$,堆叠$\overline{\mathrm{E}}[11,29,62]$ 或双$\boxed{\mathrm{E}[44,57,61]\text{ 变压器。在这项工作中，我们主要研}}$ 究堆叠的行为-(图。3 a) 和双变压器(图。3 b)架构， 以获得相对更好的性能
弱接地 VideoQA。除了回答问题外，薄弱的VideoQA 还要求模型明确估计与 QA 相关的视频片段，以作为视觉证据。我们在下面介绍三个与模型无关的解决方案来实现这一目标：
事后 (PH)。直观地，可以通过对时间注意力 i.e. 进行事后分析来找到相关的时间片段，从而确定具有最大注意力值的片段或帧，然后对其进行阈值处理以获得时间间隔。为此，我们使用注意力池来总结双架构的时间转换器的输出。对于堆叠式架构，我们直接返回预测令牌对应的平均多头关注度值。
朴素高斯 (NG)。事后方法旨在分析模型，但不会影响其预测。更有利的是，我们建议将视频接地机制明确纳人 VideoQA。我们在图中说明了框架。4a , 并重新表述 Eqn。1作为
(2)
$$a^*,\:t^*=\operatorname{argmax}_{a\in A}\:\Psi(a|v_t,q,A)\Phi(t|v,q),$$
,其中接地模块 $\Phi$ 首先估计 $t$ 指定的关键时刻，然后 QA 模块$\Psi$获取更本地化的视频内容$v_t$进行答案预测。为了实现端到端学习$,t$由整个视频序列 i.e. $,t\sim N(\mu,\sigma^2)$ 上的可微分高斯权重表示，其中$\mu,\sigma\in[0,1]$ 是两个可学习的高斯参数，分别对应于均值和标准差。在推理过程中，可以通过置信区间$t=(\mu-\gamma\sigma,\mu+\gamma\sigma)*d$来实现接地，其中$\gamma$是控制置信区间宽度的超参数，$d$表示视频的持续时间
图。3 c 显示了此朴素解决方案的双变压器实例化。与原始 VideoQA 对应物的差异 (图。3 b) 位于高斯掩码预测头中，以及高斯加权令牌学习和聚合阶段(详细信息见附录 A.2 )。我们发现这种方法有效地学习和输出接地信息。然而，由于 QA 监管不力，对事后解决方案的改进有限
NG+。根据朴素的高斯结果，我们进一步设计了具有跨模态自监督功能的辅助目标，以将 VQA 目标正则化为更直观的 QA。具体来说，对于每个问题$q^+$ ,我们将相应的接地假设$v_t$视为锚点，并将其拉近$q^+$,同时将其推离特征空间中的其他问题$Q^-$。否定集$Q^-$ 包括：1) 在同一视频中定义的其他问题，因为每个间题的很大一部分 (将近一半)都会调用唯一的视频时刻来回答(图。$\mathsf{E}$2b (3));2)从其他视频中采样的问题， 以确保负样本的充分性和多样性。此外，我们通过重述每个问题(使用 GPT-4[41])来丰富 10 个积极问题的%,最多可增加 5 个问题以形成$Q^+$。值得注意的是，每次训练迭代只有一个正面问题，并且随机选择丰富的正面问题来代替原始问题以进行数据增强。因此，
通过将否定问题的数量固定为与干扰者答案的数量相同，可以将这种形式的对比实现为分类。因此，我们的最终解决方案是：
$$a^{*},\:t^{*}=\underbrace{\mathrm{argmax}_{a\in A}\Psi(a|v_{t},q^{+},A)\Phi(t|v,q^{+})}_{\mathrm{GroundedQA}}+\\\alpha\underbrace{\mathrm{argmax}_{q\in Q}\Psi(q^{+}|v_{t},Q)\Phi(t|v,q^{+})}_{\mathrm{Grounding}},$$
(3)
其中$Q=Q^{+}\cup Q^{-}$包含$v_t$和$\alpha$的正面和负面问题是一个权衡参数。请注意，Grounding-term 粗略地标识了与问题相关的视频时刻$t$ ,而 GroundedQA-term 不仅可以进行预测，还可以通过答案监督来细化时刻$t$ 。因此，总体目标要求基础视频内容与答案和问题都相关。
# 5. Experiments
5.1.0verview
我们的实验回答了三个研究问题：Q1:当前 VLM 的预测在多大程度上基于相关视频内容？Q2: 更好的QA 性能是否意味着更好的接地，反之亦然？Q3: 我们的高斯屏蔽机制有多有效？我们研究了各种各样的VLM, 涵盖了不同的体系结构(双变压器和堆叠变压器),视觉编码器(特定于任务并使用图像或视频文本数据进行了预训练的 XBERT。对于所有模型，我们从每个视频中均匀采样 32 帧，并冻结视觉编码器。在事后分析中，根据平均注意力值动态设置时间注意力阈值，以最大化基础 QA 准确性。Eqn 中否定问题的数量。3 与 MCQA 中的干扰者答案数保持不变，便于联合优化。权衡参数$\alpha$设置为 1 和0。1个，分别用于双变压器和堆叠变压器。在推理过程中，从$\{1,0~中选择$ 高斯置信区间的超参数$\gamma$ 。8} 取决于不同的型号。我们的最终结果是根据高斯和时间注意力的预测相结合而报告的。所有超参数都在验证集上进行了调整，除非另有说明，否则结果会在测试集上报告。其他细节见附录 A.2
评估。我们报告了 QA [76] 的准确性，它代表正确回答问题的百分比。对于视觉证据基础，我们使用预测交集 (IoP) 来测量预测的时间窗口是否位于地面实况之内。此外，我们还包括遵循视频接地基准的时间IoU。对于 IoP 和 IoU, 我们报告平均值和重叠阈值为0的值。3 和0。5。如果 QA 对涉及多个时间段，我们将根据与预测最大重叠的时间段报告结果。值得注意的是，我们定义了基础 QA 准确性(Acc@GQA),以检查正确回答的问题的百分比以及视觉上的基础(i.e. , IoP$\geq$。5)
# 5.2. Result and Analysis
5.2.1 Q1:Are the answers visually grounded?
我们专注于 Acc@QA, Acc@GQA 和 IoP@0。5 位于Tab 的 Post-hoc(PH) 块中。3 。通常，现有的 VLM
|
[
[
0.08251953125,
0.09344482421875,
0.47216796875,
0.09344482421875,
0.47216796875,
0.106689453125,
0.08251953125,
0.106689453125
],
[
0.50244140625,
0.0921630859375,
0.8896484375,
0.09344482421875,
0.8896484375,
0.106689453125,
0.50244140625,
0.1060791015625
],
[
0.08251953125,
0.10797119140625,
0.468017578125,
0.10797119140625,
0.468017578125,
0.121826171875,
0.08251953125,
0.121826171875
],
[
0.50341796875,
0.106689453125,
0.890625,
0.10919189453125,
0.890625,
0.12249755859375,
0.50341796875,
0.12060546875
],
[
0.08087158203125,
0.12249755859375,
0.4150390625,
0.12371826171875,
0.4150390625,
0.136962890625,
0.08087158203125,
0.1363525390625
],
[
0.50341796875,
0.121826171875,
0.625,
0.12249755859375,
0.625,
0.136962890625,
0.50341796875,
0.1363525390625
],
[
0.1690673828125,
0.1514892578125,
0.381591796875,
0.1514892578125,
0.381591796875,
0.1666259765625,
0.1690673828125,
0.1666259765625
],
[
0.4453125,
0.15283203125,
0.468017578125,
0.15283203125,
0.468017578125,
0.1654052734375,
0.4453125,
0.1654052734375
],
[
0.51953125,
0.149658203125,
0.8505859375,
0.149658203125,
0.8505859375,
0.1673583984375,
0.51953125,
0.1673583984375
],
[
0.396240234375,
0.1558837890625,
0.4306640625,
0.1558837890625,
0.4306640625,
0.1634521484375,
0.396240234375,
0.1634521484375
],
[
0.67236328125,
0.1717529296875,
0.74853515625,
0.1717529296875,
0.74853515625,
0.18310546875,
0.67236328125,
0.18310546875
],
[
0.08660888671875,
0.179931640625,
0.46728515625,
0.179931640625,
0.46728515625,
0.1932373046875,
0.08660888671875,
0.1932373046875
],
[
0.60107421875,
0.1900634765625,
0.8544921875,
0.1881103515625,
0.8544921875,
0.205810546875,
0.60107421875,
0.2083740234375
],
[
0.869140625,
0.184326171875,
0.88623046875,
0.184326171875,
0.88623046875,
0.1932373046875,
0.869140625,
0.1932373046875
],
[
0.08251953125,
0.1944580078125,
0.46728515625,
0.1944580078125,
0.46728515625,
0.2083740234375,
0.08251953125,
0.2083740234375
],
[
0.08087158203125,
0.210205078125,
0.4697265625,
0.210205078125,
0.4697265625,
0.2235107421875,
0.08087158203125,
0.2235107421875
],
[
0.701171875,
0.212158203125,
0.7646484375,
0.212158203125,
0.7646484375,
0.2279052734375,
0.701171875,
0.2279052734375
],
[
0.08251953125,
0.2266845703125,
0.2459716796875,
0.2266845703125,
0.2459716796875,
0.2374267578125,
0.08251953125,
0.2374267578125
],
[
0.101318359375,
0.242431640625,
0.46728515625,
0.242431640625,
0.46728515625,
0.25244140625,
0.101318359375,
0.25244140625
],
[
0.5048828125,
0.2412109375,
0.88818359375,
0.2412109375,
0.88818359375,
0.251953125,
0.5048828125,
0.251953125
],
[
0.080078125,
0.25390625,
0.46728515625,
0.255126953125,
0.46728515625,
0.26953125,
0.080078125,
0.268310546875
],
[
0.5107421875,
0.25244140625,
0.890625,
0.25390625,
0.890625,
0.268310546875,
0.5107421875,
0.267578125
],
[
0.08251953125,
0.26953125,
0.46728515625,
0.26953125,
0.46728515625,
0.28271484375,
0.08251953125,
0.28271484375
],
[
0.5048828125,
0.26953125,
0.890625,
0.26953125,
0.890625,
0.28271484375,
0.5048828125,
0.28271484375
],
[
0.08087158203125,
0.285400390625,
0.281982421875,
0.285400390625,
0.281982421875,
0.298583984375,
0.08087158203125,
0.298583984375
],
[
0.5048828125,
0.285400390625,
0.890625,
0.285400390625,
0.890625,
0.298583984375,
0.5048828125,
0.298583984375
],
[
0.1004638671875,
0.3017578125,
0.468017578125,
0.3017578125,
0.468017578125,
0.31494140625,
0.1004638671875,
0.31494140625
],
[
0.5048828125,
0.3017578125,
0.8896484375,
0.3017578125,
0.8896484375,
0.3125,
0.5048828125,
0.3125
],
[
0.08251953125,
0.318115234375,
0.46728515625,
0.318115234375,
0.46728515625,
0.328857421875,
0.08251953125,
0.328857421875
],
[
0.50244140625,
0.31640625,
0.53125,
0.31640625,
0.53125,
0.328857421875,
0.50244140625,
0.328857421875
],
[
0.08251953125,
0.333251953125,
0.46728515625,
0.333251953125,
0.46728515625,
0.343505859375,
0.08251953125,
0.343505859375
],
[
0.08087158203125,
0.3466796875,
0.468017578125,
0.3466796875,
0.468017578125,
0.35986328125,
0.08087158203125,
0.35986328125
],
[
0.5009765625,
0.343505859375,
0.6396484375,
0.3447265625,
0.6396484375,
0.359130859375,
0.5009765625,
0.35791015625
],
[
0.08251953125,
0.363037109375,
0.46728515625,
0.363037109375,
0.46728515625,
0.373779296875,
0.08251953125,
0.373779296875
],
[
0.50244140625,
0.367431640625,
0.615234375,
0.36865234375,
0.615234375,
0.38330078125,
0.50244140625,
0.3818359375
],
[
0.08087158203125,
0.376953125,
0.314453125,
0.376953125,
0.314453125,
0.39013671875,
0.08087158203125,
0.39013671875
],
[
0.1004638671875,
0.392578125,
0.468017578125,
0.392578125,
0.468017578125,
0.406005859375,
0.1004638671875,
0.406005859375
],
[
0.52294921875,
0.392578125,
0.890625,
0.392578125,
0.890625,
0.406005859375,
0.52294921875,
0.406005859375
],
[
0.08251953125,
0.4091796875,
0.465576171875,
0.4091796875,
0.465576171875,
0.419189453125,
0.08251953125,
0.419189453125
],
[
0.5048828125,
0.4091796875,
0.8896484375,
0.4091796875,
0.8896484375,
0.419189453125,
0.5048828125,
0.419189453125
],
[
0.08087158203125,
0.42236328125,
0.468017578125,
0.42236328125,
0.468017578125,
0.436279296875,
0.08087158203125,
0.436279296875
],
[
0.50341796875,
0.42236328125,
0.89208984375,
0.42236328125,
0.89208984375,
0.436279296875,
0.50341796875,
0.436279296875
],
[
0.08087158203125,
0.438232421875,
0.238525390625,
0.438232421875,
0.238525390625,
0.451904296875,
0.08087158203125,
0.451904296875
],
[
0.5048828125,
0.438232421875,
0.890625,
0.438232421875,
0.890625,
0.451904296875,
0.5048828125,
0.451904296875
],
[
0.50244140625,
0.45263671875,
0.8896484375,
0.45263671875,
0.8896484375,
0.466552734375,
0.50244140625,
0.466552734375
],
[
0.125,
0.466552734375,
0.468017578125,
0.466552734375,
0.468017578125,
0.479736328125,
0.125,
0.479736328125
],
[
0.5048828125,
0.468505859375,
0.890625,
0.468505859375,
0.890625,
0.481689453125,
0.5048828125,
0.481689453125
],
[
0.5048828125,
0.484130859375,
0.8896484375,
0.484130859375,
0.8896484375,
0.49755859375,
0.5048828125,
0.49755859375
],
[
0.0849609375,
0.49365234375,
0.468017578125,
0.49365234375,
0.468017578125,
0.5068359375,
0.0849609375,
0.5068359375
],
[
0.50341796875,
0.498779296875,
0.8896484375,
0.498779296875,
0.8896484375,
0.51220703125,
0.50341796875,
0.51220703125
],
[
0.08087158203125,
0.5087890625,
0.468017578125,
0.5087890625,
0.468017578125,
0.52294921875,
0.08087158203125,
0.52294921875
],
[
0.50634765625,
0.51513671875,
0.88818359375,
0.51513671875,
0.88818359375,
0.52587890625,
0.50634765625,
0.52587890625
],
[
0.08660888671875,
0.52587890625,
0.46728515625,
0.52587890625,
0.46728515625,
0.5361328125,
0.08660888671875,
0.5361328125
],
[
0.50341796875,
0.52880859375,
0.890625,
0.52880859375,
0.890625,
0.54248046875,
0.50341796875,
0.54248046875
],
[
0.08416748046875,
0.54150390625,
0.46728515625,
0.54150390625,
0.46728515625,
0.5517578125,
0.08416748046875,
0.5517578125
],
[
0.5048828125,
0.54345703125,
0.890625,
0.54345703125,
0.890625,
0.556640625,
0.5048828125,
0.556640625
],
[
0.08416748046875,
0.55615234375,
0.46728515625,
0.55615234375,
0.46728515625,
0.56640625,
0.08416748046875,
0.56640625
],
[
0.5048828125,
0.55810546875,
0.890625,
0.55810546875,
0.890625,
0.5712890625,
0.5048828125,
0.5712890625
],
[
0.08087158203125,
0.568359375,
0.46728515625,
0.568359375,
0.46728515625,
0.58203125,
0.08087158203125,
0.58203125
],
[
0.50244140625,
0.57275390625,
0.890625,
0.57373046875,
0.890625,
0.5869140625,
0.50244140625,
0.5859375
],
[
0.08251953125,
0.5859375,
0.46728515625,
0.5859375,
0.46728515625,
0.5966796875,
0.08251953125,
0.5966796875
],
[
0.5048828125,
0.58837890625,
0.89208984375,
0.58837890625,
0.89208984375,
0.6015625,
0.5048828125,
0.6015625
],
[
0.08251953125,
0.6015625,
0.2099609375,
0.6015625,
0.2099609375,
0.6123046875,
0.08251953125,
0.6123046875
],
[
0.5048828125,
0.60400390625,
0.890625,
0.60400390625,
0.890625,
0.6171875,
0.5048828125,
0.6171875
],
[
0.101318359375,
0.61572265625,
0.468017578125,
0.61572265625,
0.468017578125,
0.62890625,
0.101318359375,
0.62890625
],
[
0.50634765625,
0.62060546875,
0.8896484375,
0.62060546875,
0.8896484375,
0.630859375,
0.50634765625,
0.630859375
],
[
0.08251953125,
0.6298828125,
0.46728515625,
0.6298828125,
0.46728515625,
0.64306640625,
0.08251953125,
0.64306640625
],
[
0.50341796875,
0.6337890625,
0.8896484375,
0.6337890625,
0.8896484375,
0.64794921875,
0.50341796875,
0.64794921875
],
[
0.08087158203125,
0.64599609375,
0.468017578125,
0.64599609375,
0.468017578125,
0.6591796875,
0.08087158203125,
0.6591796875
],
[
0.50341796875,
0.65087890625,
0.55615234375,
0.65087890625,
0.55615234375,
0.6611328125,
0.50341796875,
0.6611328125
],
[
0.08251953125,
0.662109375,
0.465576171875,
0.662109375,
0.465576171875,
0.67236328125,
0.08251953125,
0.67236328125
],
[
0.52294921875,
0.66357421875,
0.8896484375,
0.6640625,
0.8896484375,
0.6787109375,
0.52294921875,
0.67822265625
],
[
0.08087158203125,
0.67529296875,
0.468017578125,
0.67529296875,
0.468017578125,
0.689453125,
0.08087158203125,
0.689453125
],
[
0.5048828125,
0.6787109375,
0.890625,
0.6787109375,
0.890625,
0.6923828125,
0.5048828125,
0.6923828125
],
[
0.08087158203125,
0.69140625,
0.1968994140625,
0.69140625,
0.1968994140625,
0.705078125,
0.08087158203125,
0.705078125
],
[
0.5048828125,
0.6943359375,
0.890625,
0.6943359375,
0.890625,
0.70849609375,
0.5048828125,
0.70849609375
],
[
0.09722900390625,
0.7041015625,
0.468017578125,
0.70556640625,
0.468017578125,
0.72021484375,
0.09722900390625,
0.71826171875
],
[
0.50341796875,
0.71044921875,
0.8896484375,
0.71044921875,
0.8896484375,
0.7236328125,
0.50341796875,
0.7236328125
],
[
0.080078125,
0.7197265625,
0.46728515625,
0.72021484375,
0.46728515625,
0.73486328125,
0.080078125,
0.734375
],
[
0.50244140625,
0.7236328125,
0.890625,
0.72607421875,
0.890625,
0.7392578125,
0.50244140625,
0.7373046875
],
[
0.08251953125,
0.736328125,
0.46728515625,
0.736328125,
0.46728515625,
0.75,
0.08251953125,
0.75
],
[
0.50244140625,
0.74072265625,
0.890625,
0.74072265625,
0.890625,
0.75390625,
0.50244140625,
0.75390625
],
[
0.08087158203125,
0.751953125,
0.468017578125,
0.751953125,
0.468017578125,
0.76513671875,
0.08087158203125,
0.76513671875
],
[
0.50341796875,
0.7548828125,
0.8896484375,
0.7548828125,
0.8896484375,
0.76806640625,
0.50341796875,
0.76806640625
],
[
0.08087158203125,
0.767578125,
0.459228515625,
0.767578125,
0.459228515625,
0.78076171875,
0.08087158203125,
0.78076171875
],
[
0.50341796875,
0.76953125,
0.890625,
0.76953125,
0.890625,
0.78271484375,
0.50341796875,
0.78271484375
],
[
0.08251953125,
0.7822265625,
0.46728515625,
0.7822265625,
0.46728515625,
0.79541015625,
0.08251953125,
0.79541015625
],
[
0.50341796875,
0.78515625,
0.88818359375,
0.78515625,
0.88818359375,
0.798828125,
0.50341796875,
0.798828125
],
[
0.08251953125,
0.798828125,
0.465576171875,
0.798828125,
0.465576171875,
0.8095703125,
0.08251953125,
0.8095703125
],
[
0.50244140625,
0.80029296875,
0.6005859375,
0.80029296875,
0.6005859375,
0.81494140625,
0.50244140625,
0.81494140625
],
[
0.08251953125,
0.81103515625,
0.46728515625,
0.81103515625,
0.46728515625,
0.82470703125,
0.08251953125,
0.82470703125
],
[
0.08087158203125,
0.8271484375,
0.46728515625,
0.8271484375,
0.46728515625,
0.84033203125,
0.08087158203125,
0.84033203125
],
[
0.50244140625,
0.82568359375,
0.6953125,
0.82568359375,
0.6953125,
0.8388671875,
0.50244140625,
0.8388671875
],
[
0.08251953125,
0.84228515625,
0.468017578125,
0.84228515625,
0.468017578125,
0.85595703125,
0.08251953125,
0.85595703125
],
[
0.50244140625,
0.8486328125,
0.84326171875,
0.8486328125,
0.84326171875,
0.86279296875,
0.50244140625,
0.86279296875
],
[
0.080078125,
0.85595703125,
0.468017578125,
0.85693359375,
0.468017578125,
0.87060546875,
0.080078125,
0.869140625
],
[
0.08416748046875,
0.87353515625,
0.46728515625,
0.87353515625,
0.46728515625,
0.8837890625,
0.08416748046875,
0.8837890625
],
[
0.50244140625,
0.8720703125,
0.8896484375,
0.87060546875,
0.8896484375,
0.88525390625,
0.50244140625,
0.88623046875
],
[
0.08251953125,
0.8896484375,
0.468017578125,
0.8896484375,
0.468017578125,
0.8994140625,
0.08251953125,
0.8994140625
],
[
0.50244140625,
0.88720703125,
0.890625,
0.88720703125,
0.890625,
0.90087890625,
0.50244140625,
0.90087890625
]
] |
[
"也可以由所有问题的全局答案集[59]给出（开放式）。",
"通过将否定问题的数量固定为与干扰者答案的数量相",
"请注意，SoTA变压器-方法E[11,57,61,62]以统一的",
"同，可以将这种形式的对比实现为分类。因此，我们的",
"公式制定并解决了多项选择QA和开放式QA：",
"最终解决方案是：",
"a* = argmaxaeA (a|u,q, A),",
"(1)",
"a*, t* = argmaxaeA (alut,qt, A)Φ(t/u,q+) +",
"",
"GroundedQA",
"其中映射亚通常实现为共享[24,50]，堆叠[11,29,62]",
"Q argmaxgeQ(q+lut,Q)(t/u,q+),",
"(3",
"或双E[44,57,61]变压器。在这项工作中，我们主要研",
"究堆叠的行为-(图。3a)和双变压器（图。3b)架构，",
"Grounding",
"以获得相对更好的性能",
"弱接地VideoQA。除了回答问题外，薄弱的",
"其中Q=Q+UQ-包含Ut和α的正面和负面问题是",
"VideoQA还要求模型明确估计与QA相关的视频片",
"一个权衡参数。请注意，Grounding-term粗略地标识",
"段，以作为视觉证据。我们在下面介绍三个与模型无关",
"了与问题相关的视频时刻t，而GroundedQA-term不",
"的解决方案来实现这一目标",
"仅可以进行预测，还可以通过答案监督来细化时刻t。",
"事后(PH)。直观地，可以通过对时间注意力i.e.进",
"因此，总体目标要求基础视频内容与答案和问题都相",
"行事后分析来找到相关的时间片段，从而确定具有最",
"关。",
"大注意力值的片段或顺，然后对其进行國值处理以获",
"得时间间隔。为此，我们使用注意力池来总结双架构的",
"5.Experiments",
"时间转换器的输出。对于堆叠式架构，我们直接返回预",
"5.1. Overview",
"测令牌对应的平均多头关注度值。",
"朴素高斯(NG)。事后方法旨在分析模型，但不会影",
"我们的实验回答了三个研究问题：Q1：当前VLM",
"响其预测。更有利的是，我们建议将视频接地机制明确",
"的预测在多大程度上基于相关视频内容？2：更好的",
"纳人VideoQA。我们在图中说明了框架。4a，并重",
"QA性能是否意味着更好的接地，反之亦然？Q3：我",
"新表述Eqm。1作为",
"们的高斯屏蔽机制有多有效？我们研究了各种各样的",
"VLM，涵盖了不同的体系结构（双变压器和堆叠变压",
"a*, t*= argmaxaeA (aut,q, A)d(tu,q),(2)",
"器），视觉编码器（特定于任务并使用图像或视频文本",
"数据进行了预训练的XBERT。对于所有模型，我们从",
"其中接地模块Φ首先估计t指定的关键时刻，然后QA",
"每个视频中均匀采样32顿，并冻结视觉编码器。在事",
"模块亚获取更本地化的视频内容Ut进行答案预测。为",
"后分析中，根据平均注意力值动态设置时间注意力域",
"了实现端到端学习.t由整个视频序列i.e.t~N（u.g2）",
"值，以最大化基础QA准确性。Eqm中否定问题的数",
"上的可微分高斯权重表示，其中u，aE0.1是两个",
"量。3与MCQA中的干扰者答案数保持不变，便于联",
"可学习的高斯参数，分别对应于均值和标准差。在推理",
"合优化。权衡参数α设置为1和0。1个，分别用于双",
"过程中，可以通过置信区间t=(μ-,μ+)*d来",
"变压器和堆叠变压器。在推理过程中，从（1，0中选择",
"实现接地，其中是控制置信区间宽度的超参数，d表",
"高斯置信区间的超参数。8}取决于不同的型号。我",
"示视频的持续时间",
"们的最终结果是根据高斯和时间注意力的预测相结合",
"图。3c显示了此朴素解决方案的双变压器实例化。",
"而报告的。所有超参数都在验证集上进行了调整，除非",
"与原始VideoQA对应物的差异（图。3b）位于高斯",
"另有说明，否则结果会在测试集上报告。其他细节见附",
"掩码预测头中，以及高斯加权令牌学习和聚合阶段（详",
"录A.2",
"细信息见附录A.2）。我们发现这种方法有效地学习和",
"评估。我们报告了QA[76]的准确性，它代表正确",
"输出接地信息。然而，由于QA监管不力，对事后解决",
"回答问题的百分比。对于视觉证据基础，我们使用预",
"方案的改进有限",
"测交集（IoP）来测量预测的时间窗口是否位于地面实",
"NG+。根据朴素的高斯结果，我们进一步设计了具",
"况之内。此外，我们还包括遵循视频接地基准的时间",
"有跨模态自监督功能的辅助目标，以将VQA目标正则",
"IoU。对于IoP和IoU，我们报告平均值和重叠阈值为",
"化为更直观的QA。具体来说，对于每个问题q+，我",
"0的值。3和0。5。如果QA对涉及多个时间段，我们",
"们将相应的接地假设ut视为锚点，并将其拉近q+，同",
"将根据与预测最大重叠的时间段报告结果。值得注意",
"时将其推离特征空间中的其他问题Q-。否定集Q",
"的是，我们定义了基础QA准确性（Acc@GQA），以",
"包括：1）在同一视频中定义的其他问题，因为每个问",
"检查正确回答的问题的百分比以及视觉上的基础（i.e",
"题的很天一部分（将近一半）都会调用唯一的视频时刻",
"，IoP ≥。5)",
"来回答（图。E2b（3)）；2）从其他视频中采样的问题",
"以确保负样本的充分性和多样性。此外，我们通过重述",
"5.2. Result and Analysis",
"每个问题（使用GPT-4[41]）来丰富10个积极问题",
"5.2.1 Ql :Are the answers visually grounded?",
"的%，最多可增加5个问题以形成Q+。值得注意的",
"是：每次训练迭代只有一个正面问题：并且随机选择手",
"我们专注于Acc@QA，Acc@GQA和IoP@0。5位于",
"富的正面问题来代替原始问题以进行数据增强。因此，",
"Tab的Post-hoc(PH)块中。3。通常，现有的VLM"
] |
{
"arxiv_id": "2309.01327",
"language": "zh",
"page_idx": 3
}
|
||
1408.4714-zh-0
|
1408.4714-zh-0.jpg
|
Conic Multi-Task Classification
Cong Li, Michael Georgiopoulos and Georgios C. Anagnostopoulos
congli@eecs.ucf.edu, michaelg@ucf.edu and georgio@ft.edu
? ? ? 。5in $\mathbf{Keywords: }$ Multiple Kernel Learning, Multi-task Learning, Statistical Learning Theory,
Generalization Bound, Multi-objective Optimization, Support Vector Machines
$\mathbf{Abstract}$
Traditionally, Multi-task Learning (MTL) models optimize the average of task-related objective functions, which is an intuitive approach and which we will be referring to as Average MTL. However, a more general framework, referred to as Conic MTL, can be formulated by considering conic combinations of the objective functions instead; in this framework, Average MTL arises as a special case, when all combination coefficients equal 1. Although the advantage of Conic MTL over Average MTL has been shown experimentally in previous works, no theoretical justification has been provided to date. In this paper, we derive a generalization bound for the Conic MTL method, and demonstrate that the tightest bound is not necessarily achieved, when all combination coefficients equal 1; hence, Average MTL may not always be the optimal choice, and it is important to consider Conic MTL. As a byproduct of the generalization bound, it also theoretically explains the good experimental results of previous relevant works. Finally, we propose a new Conic MTL model, whose conic combination coefficients minimize the generalization bound, instead of choosing them heuristically as has been done in previous methods The rationale and advantage of our model is demonstrated and verified via a series of experiments by comparing with several other methods.
# 1 Introduction
Multi-Task Learning (MTL)自[4] 诞生以来，十多年来一直是一个活跃的研究领域。通过使用共享信息同时训练多个任务，与分别训练每个任务相比，预计可以提高每个任务的泛化性能。以前，已经考虑了各种MTL 方案，其中许多方案通过权重为$\boldsymbol{w}_t,t=1,\cdots T$的线性函数对$t$任务进行建模，并假设任务之间存在某种潜在的关系。例如，[6] 中的作者假设所有$w_t$都是以$\bar{w}$为中心的集群的一部分，后者是与$w_t$共同学习的。此假设进一步扩展到以下情况：权重$\boldsymbol{w}_t$可以分组为不同的簇，而不是单个全局簇[31,32]。此外，广泛持有的 MTL 假设是任务共享一个共同的，可能稀疏的特征表示，如在[22,12,9,18,7,2,14]中所做的， 仅举几例。值得一提的是，其中许多作品仅允许在任务的子集之间共享功能，这些任务被认为是彼此“相似” 或“相关”的，其中在培训期间发现了任务之间的相关性。这种方法减少，有时，完全避免了“负转移”的影响，$i.e.$,无关任务之间的知识转移，这导致退化的泛化性能。其他几个最近的工作，专注于任务相关性的发现包括[30,29,26,24] 。此外，一些基于内核的 MTL 模型假设来自所有任务的数据都通过(部分)公共特征映射进行预处理，因此 (部分) 共享相同的内核函数；再次参见[28,25,27] ,仅举几例
大多数以前的 MTL 公式都考虑以下经典设置：为第$t$个任务$(t=1,\cdots,T$ )提供了一组训练数据$\{ \boldsymbol{x}_t^i, y_t^i\} \in \mathcal{X} \times \mathcal{Y} , i= 1, \cdots , N_t$ ,其中$\mathcal{X},\mathcal{Y}$是相应的输人和输出空间。假设来自第$t$个任务的每个数据都是从基础概率分布$P_t(X_t,Y_t)$中提取的，其中$X_t$和$Y_t$分别是输人和输出空间中的随机变量。然后，一个 MTL 问题被公式化如下
$$\min_{\boldsymbol{w}\in\Omega(\boldsymbol{w})}\:\sum_{t=1}^Tf(\boldsymbol{w}_t,\boldsymbol{x}_t,\boldsymbol{y}_t)$$
(1)
,其中$\boldsymbol w\triangleq(\boldsymbol w_1,\cdots,\boldsymbol w_T)$是所有$\boldsymbol w_t$的集合，类似地$, \boldsymbol{x}_t\triangleq ( x_t^1, \cdots , \boldsymbol{x}_t^{N_t})$ $, \boldsymbol{y}_t\triangleq ( y_t^1, \cdots , y_t^{N_t})$。$f$是所有任务通用的函数。重要的是要观察到，在没有约束$\boldsymbol{w}\in\Omega(\boldsymbol{w})$的情况下，Problem (1)会降级为$T$独立学习问题。因此，在大多数情况下，集$\Omega(\boldsymbol{w})$旨在捕获任务间的关系。例如，在[28]中，模型将 MTL 与Multi-Kernel Learning (MKL) 结合在一起，其公式如下
1
|
[
[
0.28515625,
0.1300048828125,
0.7138671875,
0.1300048828125,
0.7138671875,
0.14453125,
0.28515625,
0.14453125
],
[
0.226318359375,
0.175537109375,
0.77392578125,
0.175537109375,
0.77392578125,
0.189453125,
0.226318359375,
0.189453125
],
[
0.289306640625,
0.2020263671875,
0.708984375,
0.2020263671875,
0.708984375,
0.21533203125,
0.289306640625,
0.21533203125
],
[
0.1396484375,
0.242431640625,
0.88134765625,
0.24365234375,
0.88134765625,
0.25830078125,
0.1396484375,
0.2568359375
],
[
0.1168212890625,
0.2587890625,
0.68310546875,
0.2587890625,
0.68310546875,
0.272705078125,
0.1168212890625,
0.272705078125
],
[
0.468017578125,
0.287353515625,
0.53369140625,
0.287353515625,
0.53369140625,
0.298583984375,
0.468017578125,
0.298583984375
],
[
0.178955078125,
0.306884765625,
0.83837890625,
0.306884765625,
0.83837890625,
0.32080078125,
0.178955078125,
0.32080078125
],
[
0.158447265625,
0.322509765625,
0.8408203125,
0.322509765625,
0.8408203125,
0.333251953125,
0.158447265625,
0.333251953125
],
[
0.1568603515625,
0.335205078125,
0.8408203125,
0.335205078125,
0.8408203125,
0.348388671875,
0.1568603515625,
0.348388671875
],
[
0.1568603515625,
0.348388671875,
0.84228515625,
0.348388671875,
0.84228515625,
0.3623046875,
0.1568603515625,
0.3623046875
],
[
0.158447265625,
0.3642578125,
0.8388671875,
0.3642578125,
0.8388671875,
0.375,
0.158447265625,
0.375
],
[
0.156005859375,
0.376953125,
0.84228515625,
0.375732421875,
0.84228515625,
0.389404296875,
0.156005859375,
0.39013671875
],
[
0.156005859375,
0.39013671875,
0.84228515625,
0.389404296875,
0.84228515625,
0.404052734375,
0.156005859375,
0.40478515625
],
[
0.1568603515625,
0.404052734375,
0.8408203125,
0.404052734375,
0.8408203125,
0.417236328125,
0.1568603515625,
0.417236328125
],
[
0.156005859375,
0.417236328125,
0.84228515625,
0.416748046875,
0.84228515625,
0.429931640625,
0.156005859375,
0.43115234375
],
[
0.1568603515625,
0.431884765625,
0.84228515625,
0.431884765625,
0.84228515625,
0.44580078125,
0.1568603515625,
0.44580078125
],
[
0.158447265625,
0.44580078125,
0.8408203125,
0.44580078125,
0.8408203125,
0.458984375,
0.158447265625,
0.458984375
],
[
0.1568603515625,
0.458984375,
0.8388671875,
0.458984375,
0.8388671875,
0.472900390625,
0.1568603515625,
0.472900390625
],
[
0.156005859375,
0.472900390625,
0.8408203125,
0.473388671875,
0.8408203125,
0.4873046875,
0.156005859375,
0.486083984375
],
[
0.158447265625,
0.4892578125,
0.41259765625,
0.4892578125,
0.41259765625,
0.5,
0.158447265625,
0.5
],
[
0.11517333984375,
0.5244140625,
0.160888671875,
0.5244140625,
0.160888671875,
0.54052734375,
0.11517333984375,
0.54052734375
],
[
0.1544189453125,
0.5244140625,
0.30810546875,
0.52587890625,
0.30810546875,
0.54052734375,
0.1544189453125,
0.5390625
],
[
0.1168212890625,
0.55615234375,
0.88134765625,
0.55615234375,
0.88134765625,
0.5693359375,
0.1168212890625,
0.5693359375
],
[
0.1168212890625,
0.57080078125,
0.88134765625,
0.57080078125,
0.88134765625,
0.583984375,
0.1168212890625,
0.583984375
],
[
0.1168212890625,
0.58544921875,
0.88134765625,
0.58544921875,
0.88134765625,
0.5986328125,
0.1168212890625,
0.5986328125
],
[
0.1168212890625,
0.60107421875,
0.88134765625,
0.60107421875,
0.88134765625,
0.6142578125,
0.1168212890625,
0.6142578125
],
[
0.1192626953125,
0.6162109375,
0.8798828125,
0.6162109375,
0.8798828125,
0.6298828125,
0.1192626953125,
0.6298828125
],
[
0.11767578125,
0.630859375,
0.88134765625,
0.630859375,
0.88134765625,
0.64453125,
0.11767578125,
0.64453125
],
[
0.11767578125,
0.64599609375,
0.8798828125,
0.64599609375,
0.8798828125,
0.6591796875,
0.11767578125,
0.6591796875
],
[
0.11517333984375,
0.66015625,
0.88134765625,
0.6611328125,
0.88134765625,
0.6748046875,
0.11517333984375,
0.673828125
],
[
0.1168212890625,
0.67529296875,
0.8798828125,
0.67529296875,
0.8798828125,
0.689453125,
0.1168212890625,
0.689453125
],
[
0.1192626953125,
0.693359375,
0.88134765625,
0.693359375,
0.88134765625,
0.7041015625,
0.1192626953125,
0.7041015625
],
[
0.1168212890625,
0.70703125,
0.76953125,
0.70703125,
0.76953125,
0.72021484375,
0.1168212890625,
0.72021484375
],
[
0.1405029296875,
0.7216796875,
0.88134765625,
0.7216796875,
0.88134765625,
0.73486328125,
0.1405029296875,
0.73486328125
],
[
0.1192626953125,
0.736328125,
0.88134765625,
0.736328125,
0.88134765625,
0.75,
0.1192626953125,
0.75
],
[
0.1168212890625,
0.751953125,
0.86376953125,
0.751953125,
0.86376953125,
0.76513671875,
0.1168212890625,
0.76513671875
],
[
0.11767578125,
0.767578125,
0.310546875,
0.767578125,
0.310546875,
0.78076171875,
0.11767578125,
0.78076171875
],
[
0.47802734375,
0.79541015625,
0.486083984375,
0.79541015625,
0.486083984375,
0.80029296875,
0.47802734375,
0.80029296875
],
[
0.46728515625,
0.8017578125,
0.58642578125,
0.806640625,
0.5849609375,
0.82470703125,
0.465576171875,
0.8193359375
],
[
0.410888671875,
0.8095703125,
0.498291015625,
0.8173828125,
0.494384765625,
0.8369140625,
0.408447265625,
0.8291015625
],
[
0.42724609375,
0.8095703125,
0.449462890625,
0.8095703125,
0.449462890625,
0.8173828125,
0.42724609375,
0.8173828125
],
[
0.85888671875,
0.8046875,
0.8837890625,
0.8046875,
0.8837890625,
0.82275390625,
0.85888671875,
0.82275390625
],
[
0.131591796875,
0.84130859375,
0.88134765625,
0.84130859375,
0.88134765625,
0.85595703125,
0.131591796875,
0.85595703125
],
[
0.11767578125,
0.85791015625,
0.8798828125,
0.85791015625,
0.8798828125,
0.8720703125,
0.11767578125,
0.8720703125
],
[
0.1168212890625,
0.87255859375,
0.88134765625,
0.87255859375,
0.88134765625,
0.88623046875,
0.1168212890625,
0.88623046875
],
[
0.1168212890625,
0.88818359375,
0.5185546875,
0.88818359375,
0.5185546875,
0.9013671875,
0.1168212890625,
0.9013671875
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"Conic Multi-Task Classification",
"Cong Li, Michael Georgiopoulos and Georgios C. Anagnostopoulos",
"congli@eecs.ucf.edu, michaelg@ucf.edu and georgio@fit.edu",
"???。5in Keywords:Multiple Kernel Learning, Multi-task Learning, Statistical Learning Theory,",
"Generalization Bound, Multi-objective Optimization, Support Vector Machines",
"Abstract",
"Traditionally, Multi-task Learning (MTL) models optimize the average of task-related objective func.",
"tions, which is an intuitive approach and which we will be referring to as Average MTL.However,amore",
"general framework, referred to as Conic MTL, can be formulated by considering conic combinations",
"of the objective functions instead; in this framework, Average MTL arises as a special case, when all",
"combination coefficients equal l.Althoughthe advantage ofConic MTL over AverageMTLhasbeer",
"shown experimentally in previous works, no theoretical justification has been provided to date. In this",
"paper, we derive a generalization bound for the Conic MTL method, and demonstrate that the tightest",
"bound is not necessarily achieved, when all combination coefficients equal 1 ; hence, Average MTL may",
"not always be the optimal choice, and it is important to consider Conic MTL. As a byproduct of the",
"generalization bound, it also theoretically explains the good experimental results of previous relevant",
"works. Finally, we propose a new Conic MTL model, whose conic combination coefficients minimize",
"the generalization bound, instead of choosing them heuristically as has been done in previous methods",
"The rationale and advantage of our model is demonstrated and verified via a series of experiments by",
"comparing with several othermethods",
"1",
"Introduction",
"Multi-TaskLearning（MTL）自【4]诞生以来，十多年来一直是一个活跃的研究领域。通过使用共享信息同",
"时训练多个任务，与分别训练每个任务相比，预计可以提高每个任务的泛化性能。以前，已经考虑了各种",
"MTL方案，其中许多方案通过权重为Wt，t=1，·T的线性函数对t任务进行建模，并假设任务之间存在",
"某种潜在的关系。例如，[6]中的作者假设所有Wt都是以w为中心的集群的一部分，后者是与Wt共同学",
"习的。此假设进一步扩展到以下情况：权重Wt可以分组为不同的簇，而不是单个全局簇[31，32]。此外，广",
"泛持有的MTL假设是任务共享一个共同的，可能稀疏的特征表示，如在[22，12，9，18，7，2，14中所做的",
"仅举几例。值得一提的是，其中许多作品仅允许在任务的子集之间共享功能，这些任务被认为是彼此“相似”",
"或“相关”的，其中在培训期间发现了任务之间的相关性。这种方法减少，有时，完全避免了“负转移”的",
"影响，i.e.，无关任务之间的知识转移，这导致退化的泛化性能。其他几个最近的工作，专注于任务相关性的",
"发现包括30，29，26，24。此外，一些基于内核的MTL模型假设来自所有任务的数据都通过（部分）公共",
"特征映射进行预处理，因此（部分）共享相同的内核函数；再次参见[28，25，27]，仅举几例",
"大多数以前的MTL公式都考虑以下经典设置：为第t个任务（t=1，·，T）提供了一组训练数据",
"αct,yE×,i=1,Nt，其中，是相应的输人和输出空间。假设来自第t个任务的每个数据",
"都是从基础概率分布Pt(Xt，Yt）中提取的，其中Xt和Yt分别是输人和输出空间中的随机变量。然后，",
"个MTL问题被公式化如下",
"",
"f(wt,at,yt)",
"we2(w)台",
"mir",
"(1)",
"其中w会（wi,，wT）是所有wt的集合，类似地，at会（αt，,a），yt（yt，,yt）。f是所",
"有任务通用的函数。重要的是要观察到，在没有约束wE2(w）的情况下，Problem（1）会降级为T独立",
"学习问题。因此，在大多数情况下，集2(w）旨在捕获任务间的关系。例如，在[28]中，模型将MTL与",
"Multi-KernelLearning（MKL）结合在一起，其公式如下",
"1"
] |
{
"arxiv_id": "1408.4714",
"language": "zh",
"page_idx": 0
}
|
||
1408.4714-zh-1
|
1408.4714-zh-1.jpg
|
$$f(\boldsymbol{w}_t,\boldsymbol{x}_t,\boldsymbol{y}_t)\triangleq\frac{1}{2}\|\boldsymbol{w}_t\|^2+C\sum_{i=1}^{N_t}l(\boldsymbol{w}_t,\phi_t(\boldsymbol{x}_t^i),\boldsymbol{y}_t^i)$$
$$\Omega(\boldsymbol w)\triangleq\{\boldsymbol w=(\boldsymbol w_1,\cdots,\boldsymbol w_T):\boldsymbol w_t\in\mathcal H_{\boldsymbol\theta,\gamma_t},\boldsymbol\theta\in\Omega(\boldsymbol\theta),\gamma\in\Omega(\gamma)\}$$
(2)
这里，$l$是指定的损失函数，$\phi_t:\mathcal{X}\to\mathcal{H}_{\boldsymbol{\theta},\boldsymbol{\gamma}_t}$是第$t$个任务的特征映射，$\mathcal{H}_{\boldsymbol{\theta},\boldsymbol{\gamma}_t}$是具有再现内核函数$k_t\triangleq\sum_{m=1}^M(\theta_m+\gamma_t^m)k_m$的 Reproducing Kernel Hilbert Space (RKHS) ,其中$k_m:\mathcal{X}\times\mathcal{X}\to\mathbb{R},m=1,\cdots,M$ 是预先选择的内核函数。$\|\boldsymbol{w}_t\|\triangleq\sqrt{\langle\boldsymbol{w}_t,\boldsymbol{w}_t\rangle}$是$\mathcal{H}_\boldsymbol{\theta},\boldsymbol{\gamma}_t$中定义的规范。此外，$\Omega(\boldsymbol{\theta})$是$\boldsymbol{\theta}\triangleq(\theta_1,\cdots,\theta_M)$的可行集，同样，$\Omega(\gamma)$是$\gamma\triangleq(\gamma_1,\cdots,\gamma_T)$的可行集。不难看出，在此设置中，$\Omega(\boldsymbol{w})$被设计为所有任务以MKL 方式部分共享相同的内核函数，由公共系数$\theta$和特定于任务的系数$\gamma_t,t=1,\cdots,T$参数化
另一个示例 Sparse MTL [25] 具有以下公式：
$$\begin{aligned}&f(\boldsymbol{w}_{t},\boldsymbol{x}_{t},\boldsymbol{y}_{t})\triangleq\sum_{i=1}^{N_{t}}l(\boldsymbol{w}_{t},\phi_{t}(\boldsymbol{x}_{t}^{i}),\boldsymbol{y}_{t}^{i})\\&\Omega(\boldsymbol{w})\triangleq\{\boldsymbol{w}=(\boldsymbol{w}_{1},\cdots,\boldsymbol{w}_{T}):\boldsymbol{w}_{t}\triangleq(\boldsymbol{w}_{t}^{1},\cdots,\boldsymbol{w}_{t}^{M}),\sum_{m=1}^{M}(\sum_{t=1}^{T}\|\boldsymbol{w}_{t}^{m}\|^{q})^{p/q}\leq R\}\end{aligned}$$
(3)
其中$\boldsymbol w_t^m\in\mathcal{H}_m,\forall m=1,\cdots,M,t=1,\cdots,T$ 、$\boldsymbol w_t\in\mathcal{H}_1\times\cdots\times\mathcal{H}_M$ 、$0<p\leq1$ 、$1\leq q\leq2$ 。请注意，尽
管原始的 Sparse MTL 的公式如下
$$\min_{\boldsymbol{w}}\:\sum_{m=1}^{M}(\sum_{t=1}^{T}\|\boldsymbol{w}_{t}^{m}\|^{q})^{p/q}+C\sum_{t=1}^{T}\sum_{i=1}^{N_{t}}l(\boldsymbol{w}_{t},\phi_{t}(\boldsymbol{x}_{t}^{i}),\boldsymbol{y}_{t}^{i})$$
(4)
由于 [15] 中的命题 12 的第一部分，我们将其重述为$^{1}$以下的 Proposition 1,很明显，对于任何$C> 0$ ,
都存在$R>0$ ,因此 Problem (1)和 Problem (4)是等效的
$\textbf{Proposition 1. Let }\mathcal{D} \subseteq \mathcal{X} , and\textit{let f, g}: \mathcal{D} \mapsto \mathbb{R} \textit{be two functions. For any }\sigma > 0$ , $there\textit{must exist a}$
$\tau > 0$ , such that the following two problems are equivalent
(5)
$$\min\limits_{x\in\mathcal{D}}\:f(x)+\sigma g(x)$$
$$\min\limits_{x\in\mathcal{D},g(x)\leq\tau}f(x)$$
(6)
Problem (1) 中给出的公式，我们将其称为 Average MTL ,直观上很吸引人：可以合理地期望通过优化$T$ 目标函数的平均值来提高$T$任务的平均泛化性能。但是，正如 [16] 中所述，求解 Problem (1) 只会在以下Multi-Objective Optimization (MOO) 问题的帕累托前沿上产生特定的解决方案
(7)
$$\min_{w\in\Omega(w)}f(w,x,y)$$
其中$f(\boldsymbol{w},\boldsymbol{x},\boldsymbol{y})\triangleq[f(\boldsymbol{w}_1,\boldsymbol{x}_1,\boldsymbol{y}_1),\cdots,f(\boldsymbol{w}_T,\boldsymbol{x}_T,\boldsymbol{y}_T)]^{\prime}$。这是真的，因为通过优化目标函数的不同圆锥组合来缩放 MOO 问题，可以发现与间题的 ParetoFront [3, p. 178] 的凸部分上的点相对应的解决方案。换句话说，通过使用不同的$\boldsymbol{\lambda }\triangleq [ \lambda _1, \cdots , \lambda _T] ^{\prime }$, $\lambda _t> 0, \forall t= 1, \cdots , T$对 Problem (7)进行锥形缩放，优化问题
$$\min\limits_{\boldsymbol{w}\in\Omega(\boldsymbol{w})}\:\sum\limits_{t=1}^T\lambda_tf(\boldsymbol{w}_t,\boldsymbol{x}_t,\boldsymbol{y}_t)$$
(8)
在 Problem (7) 的帕累托前沿上产生不同的点。因此，几乎没有理由相信 Problem (8)的解决方案对于$\lambda _t= 1, \forall t= 1, \cdots , T$ , $i. e$.的特殊情况，即平均 MTL 的解决方案是可以实现的最佳解决方案。实际上，
$^{1}$Note that the difference between Proposition l here and Proposition 12 in [15] is that, Proposition 1 does not require convexity of $f,g$ and $\mathcal{D};$ these are requirements necessary for the second part of Proposition 12 in $\left[15\right]$, which we do not utilize here.
2
|
[
[
0.48193359375,
0.10479736328125,
0.496826171875,
0.10479736328125,
0.496826171875,
0.11297607421875,
0.48193359375,
0.11297607421875
],
[
0.27783203125,
0.117431640625,
0.384033203125,
0.117431640625,
0.384033203125,
0.1319580078125,
0.27783203125,
0.1319580078125
],
[
0.39208984375,
0.1142578125,
0.61767578125,
0.11553955078125,
0.61767578125,
0.1331787109375,
0.39208984375,
0.1319580078125
],
[
0.4794921875,
0.1319580078125,
0.50244140625,
0.1319580078125,
0.50244140625,
0.143310546875,
0.4794921875,
0.143310546875
],
[
0.859375,
0.1275634765625,
0.88232421875,
0.1275634765625,
0.88232421875,
0.1427001953125,
0.859375,
0.1427001953125
],
[
0.275390625,
0.1483154296875,
0.72216796875,
0.149658203125,
0.72216796875,
0.1641845703125,
0.275390625,
0.1634521484375
],
[
0.1168212890625,
0.17236328125,
0.88134765625,
0.17236328125,
0.88134765625,
0.1875,
0.1168212890625,
0.1875
],
[
0.11273193359375,
0.1849365234375,
0.8837890625,
0.1875,
0.8837890625,
0.2083740234375,
0.11273193359375,
0.205810546875
],
[
0.11767578125,
0.20703125,
0.88134765625,
0.20703125,
0.88134765625,
0.2215576171875,
0.11767578125,
0.2215576171875
],
[
0.11767578125,
0.2247314453125,
0.88134765625,
0.2247314453125,
0.88134765625,
0.238037109375,
0.11767578125,
0.238037109375
],
[
0.11358642578125,
0.238037109375,
0.80810546875,
0.2392578125,
0.80810546875,
0.25390625,
0.11358642578125,
0.25244140625
],
[
0.1422119140625,
0.255126953125,
0.4599609375,
0.255126953125,
0.4599609375,
0.268310546875,
0.1422119140625,
0.268310546875
],
[
0.339111328125,
0.27978515625,
0.359375,
0.27978515625,
0.359375,
0.29345703125,
0.339111328125,
0.29345703125
],
[
0.223876953125,
0.28857421875,
0.4794921875,
0.291748046875,
0.4794921875,
0.313232421875,
0.223876953125,
0.31005859375
],
[
0.339111328125,
0.30810546875,
0.363525390625,
0.311279296875,
0.361083984375,
0.32568359375,
0.336669921875,
0.322509765625
],
[
0.859375,
0.31494140625,
0.88232421875,
0.31494140625,
0.88232421875,
0.330078125,
0.859375,
0.330078125
],
[
0.60107421875,
0.32568359375,
0.63134765625,
0.32568359375,
0.63134765625,
0.333984375,
0.60107421875,
0.333984375
],
[
0.63134765625,
0.32763671875,
0.640625,
0.32763671875,
0.640625,
0.33203125,
0.63134765625,
0.33203125
],
[
0.2254638671875,
0.333984375,
0.771484375,
0.333984375,
0.771484375,
0.3515625,
0.2254638671875,
0.3515625
],
[
0.59228515625,
0.354248046875,
0.64892578125,
0.354248046875,
0.64892578125,
0.3642578125,
0.59228515625,
0.3642578125
],
[
0.11767578125,
0.373779296875,
0.88134765625,
0.373779296875,
0.88134765625,
0.386962890625,
0.11767578125,
0.386962890625
],
[
0.1168212890625,
0.38818359375,
0.363525390625,
0.38818359375,
0.363525390625,
0.401611328125,
0.1168212890625,
0.401611328125
],
[
0.3505859375,
0.412353515625,
0.40283203125,
0.4091796875,
0.404296875,
0.4248046875,
0.351318359375,
0.427978515625
],
[
0.52392578125,
0.412353515625,
0.56884765625,
0.412353515625,
0.56884765625,
0.427978515625,
0.52392578125,
0.427978515625
],
[
0.3095703125,
0.4248046875,
0.68701171875,
0.4248046875,
0.68701171875,
0.44580078125,
0.3095703125,
0.44580078125
],
[
0.859375,
0.427978515625,
0.88232421875,
0.427978515625,
0.88232421875,
0.442626953125,
0.859375,
0.442626953125
],
[
0.3505859375,
0.44580078125,
0.404296875,
0.44580078125,
0.404296875,
0.45263671875,
0.3505859375,
0.45263671875
],
[
0.51953125,
0.442626953125,
0.56884765625,
0.4443359375,
0.56884765625,
0.455810546875,
0.51953125,
0.453857421875
],
[
0.11767578125,
0.4609375,
0.8837890625,
0.462158203125,
0.8837890625,
0.4765625,
0.11767578125,
0.47607421875
],
[
0.1168212890625,
0.477783203125,
0.54345703125,
0.477783203125,
0.54345703125,
0.4912109375,
0.1168212890625,
0.4912109375
],
[
0.11517333984375,
0.501953125,
0.8837890625,
0.50244140625,
0.8837890625,
0.517578125,
0.11517333984375,
0.5166015625
],
[
0.11517333984375,
0.51806640625,
0.53759765625,
0.51806640625,
0.53759765625,
0.5322265625,
0.11517333984375,
0.5322265625
],
[
0.4345703125,
0.5498046875,
0.560546875,
0.548828125,
0.560546875,
0.56298828125,
0.4345703125,
0.564453125
],
[
0.859375,
0.548828125,
0.88232421875,
0.548828125,
0.88232421875,
0.56298828125,
0.859375,
0.56298828125
],
[
0.439453125,
0.56005859375,
0.45751953125,
0.56005859375,
0.45751953125,
0.5673828125,
0.439453125,
0.5673828125
],
[
0.5146484375,
0.5859375,
0.5556640625,
0.5859375,
0.5556640625,
0.60107421875,
0.5146484375,
0.60107421875
],
[
0.861328125,
0.5859375,
0.88232421875,
0.5859375,
0.88232421875,
0.60107421875,
0.861328125,
0.60107421875
],
[
0.4453125,
0.59521484375,
0.5107421875,
0.5986328125,
0.5087890625,
0.60986328125,
0.443603515625,
0.60693359375
],
[
0.46728515625,
0.58837890625,
0.495849609375,
0.58837890625,
0.495849609375,
0.5966796875,
0.46728515625,
0.5966796875
],
[
0.1168212890625,
0.619140625,
0.88232421875,
0.619140625,
0.88232421875,
0.63330078125,
0.1168212890625,
0.63330078125
],
[
0.11767578125,
0.6337890625,
0.88134765625,
0.6337890625,
0.88134765625,
0.6484375,
0.11767578125,
0.6484375
],
[
0.1168212890625,
0.6494140625,
0.68701171875,
0.6494140625,
0.68701171875,
0.66357421875,
0.1168212890625,
0.66357421875
],
[
0.441162109375,
0.6787109375,
0.564453125,
0.6796875,
0.564453125,
0.6943359375,
0.441162109375,
0.693359375
],
[
0.859375,
0.6796875,
0.88232421875,
0.6796875,
0.88232421875,
0.6943359375,
0.859375,
0.6943359375
],
[
0.435546875,
0.6923828125,
0.483642578125,
0.6923828125,
0.483642578125,
0.69970703125,
0.435546875,
0.69970703125
],
[
0.11767578125,
0.7119140625,
0.88134765625,
0.7119140625,
0.88134765625,
0.7265625,
0.11767578125,
0.7265625
],
[
0.1168212890625,
0.72607421875,
0.88134765625,
0.7265625,
0.88134765625,
0.7412109375,
0.1168212890625,
0.74072265625
],
[
0.1168212890625,
0.7421875,
0.859375,
0.7421875,
0.859375,
0.755859375,
0.1168212890625,
0.755859375
],
[
0.465576171875,
0.767578125,
0.484375,
0.767578125,
0.484375,
0.78076171875,
0.465576171875,
0.78076171875
],
[
0.461669921875,
0.77978515625,
0.59228515625,
0.7822265625,
0.59228515625,
0.796875,
0.461669921875,
0.79443359375
],
[
0.861328125,
0.7822265625,
0.88232421875,
0.7822265625,
0.88232421875,
0.79541015625,
0.861328125,
0.79541015625
],
[
0.406005859375,
0.78857421875,
0.490234375,
0.79541015625,
0.48779296875,
0.8095703125,
0.40283203125,
0.8017578125
],
[
0.418212890625,
0.7841796875,
0.44775390625,
0.7841796875,
0.44775390625,
0.79248046875,
0.418212890625,
0.79248046875
],
[
0.1168212890625,
0.81640625,
0.88232421875,
0.81640625,
0.88232421875,
0.830078125,
0.1168212890625,
0.830078125
],
[
0.1168212890625,
0.83203125,
0.88232421875,
0.83203125,
0.88232421875,
0.84521484375,
0.1168212890625,
0.84521484375
],
[
0.1356201171875,
0.85400390625,
0.88232421875,
0.85498046875,
0.88232421875,
0.869140625,
0.1356201171875,
0.86865234375
],
[
0.11517333984375,
0.86865234375,
0.88232421875,
0.869140625,
0.88232421875,
0.8837890625,
0.11517333984375,
0.88330078125
],
[
0.11767578125,
0.88330078125,
0.30810546875,
0.8837890625,
0.30810546875,
0.89599609375,
0.11767578125,
0.89453125
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"N.",
"f(Wt,&t,yt)≤",
"lwtll2 +ci(wt,pt(ai),yi)",
"=1",
"(2)",
"2(w) ≤ [w= (W1,**,wT) : Wt EHe,%,0 E2(0), E2())",
"这里，l是指定的损失函数，Φt：→He是第t个任务的特征映射，He是具有再现内核函数",
"kt m=(0m+)km的Reproducing Kernel Hilbert Space (RKHS),其中 km :XxX→ R,m = 1,:,M",
"是预先选择的内核函数。Iwtll会Vwtwt）是He中定义的规范。此外，2(0）是θ（01，·,0M）的",
"可行集，同样，2(~）是（1，，T）的可行集。不难看出，在此设置中，2(w）被设计为所有任务以",
"MKL方式部分共享相同的内核函数，由公共系数θ和特定于任务的系数t，t=1,..·，T参数化",
"另一个示例SparseMTL[25]具有以下公式：",
"N.",
"f(wt, at, yt) ≤(wt, t(at), yt)",
"i=1",
"(3)",
"M",
"",
"2(w) ≤ [w= (w1,... , wT) : wt ≤ (wi,...,wM), (lwm[9)p/q≤R)",
"m=1 t=1",
"其中wmEHm,Vm=1,..·,M,t=1,.·,T、wtEHi×...×HM0<p≤11≤q≤2。请注意,尽",
"管原始的SparseMTL的公式如下",
"MT",
"TNt",
"min Z(Zlwll9)p/q +cE(wt, pt(ri),yi)",
"(4)",
"m=1t=1",
"t=1 z=1",
"由于[15]中的命题12的第一部分，我们将其重述为1以下的Proposition1，很明显，对于任何C>0，",
"都存在R>0，因此Problem（1）和Problem（4）是等效的",
"Proposition 1.LetD X, and let f,g:D-R be two functions.For any >0,there must erist a",
"T > 0 , such that the following two problems are equialent",
"min f(r)+og(c)",
"(5)",
"TET",
"f(α)",
"(6)",
"TED,g()≤",
"min",
"Problem（1）中给出的公式，我们将其称为AverageMTL，直观上很吸引人：可以合理地期望通过优化T",
"目标函数的平均值来提高T任务的平均泛化性能。但是，正如[16]中所述，求解Problem（1）只会在以下",
"Multi-ObjectiveOptimization（MOO）问题的帕累托前沿上产生特定的解决方案",
"min, f(w,a,y)",
"(7)",
"WE(W)",
"其中f(w,a,y）全[f(w1,1,yi)，··，f(wT,aT,yT)'。这是真的，因为通过优化目标函数的不同圆锥组合",
"来缩放MOO问题，可以发现与问题的ParetoFront[3，p：178】的凸部分上的点相对应的解决方案。换句话",
"说，通过使用不同的入[入1，.·，入\"，入t>0,Vt=1，·，T对Problem（7）进行锥形缩放，优化问题",
"T",
"入tf(wt,at,yt)",
"(8)",
"WEn(w）=",
"min",
"在Problem（7）的帕累托前沿上产生不同的点。因此，几乎没有理由相信Problem（8）的解决方案对于",
"入t=1,Vt=1，：，T，i.e.的特殊情况，即平均MTL的解决方案是可以实现的最佳解决方案。实际上，",
"INote that the difference between Proposition 1 here and Proposition 12 in [15] is that, Proposition 1 does not",
"require convexity of f , g and D ; these are requirements necessary for the second part of Proposition 12 in [15] ,",
"which we do not utilize here.",
"2"
] |
{
"arxiv_id": "1408.4714",
"language": "zh",
"page_idx": 1
}
|
||
1408.4714-zh-2
|
1408.4714-zh-2.jpg
|
ParetoFront 上可能还有其他点可以为每个任务带来更好的泛化性能，从而为$T$ 任务带来更好的平均性能。
因此，不是求解 Problem (1),而是可以通过优化 Problem (8)来实现
在 [16] 中执行了这些方面的先前工作。作者考虑了以下 MTL 公式，名为 Pareto-Path MTL
(9)
$$\min_{\boldsymbol{w}\in\Omega(\boldsymbol{w})}\:[\sum_{t=1}^T(f(\boldsymbol{w}_t,\boldsymbol{x}_t,\boldsymbol{y}_t))^p]^{1/p}$$
,假设所有目标函数均为正，则在$p\geq1$时最小化目标的$L_p$范数，在$0<p<1$时最小化$L_p$伪范数。事实
证明，对于任何 $p> 0$ , Problem ( 9) 都等同于 Problem (8)
$$\lambda_t=\begin{cases}\frac{f(\boldsymbol{w}_t,\boldsymbol{x}_t,\boldsymbol{y}_t)^{p-1}}{\sum_{t=1}^T(f(\boldsymbol{w}_t,\boldsymbol{x}_t,\boldsymbol{y}_t))^p}&\text{if}p>1\\1&\text{if}p=1\\\frac{\sum_{t=1}^T(f(\boldsymbol{w}_t,\boldsymbol{x}_t,\boldsymbol{y}_t))^{\frac{1-p}{p}}}{f(\boldsymbol{w}_t,\boldsymbol{x}_t,\boldsymbol{y}_t)^{1-p}}&\text{if}0<p<1\end{cases},\forall t=1,\cdots,T$$
(10)
因此，通过改变$p> 0$ , Problem ( 9) 的解决方案在 Problem (7)的 ParetoFront 上跟踪路径。虽然平均MTL 等于 Problem (9),但当$p=1$时，已证明在基于 Support Vector Machine (SVM)的 MKL 设置中与$p=1$相比 , $p<1$的实验结果通常更好。尽管获得的优越结果与我们先前的论点密切相关，但作者并未提供严格的基础，说明考虑$T$任务目标平均值以外的目标函数的优势。因此，到目前为止，在论文的目标函数中使用$L_p$ -(伪)范数仍然是其方法的启发式元素
鉴于刚才提到的平均 MTL 的潜在缺点以及帕累托路径 MTL 缺乏支持理论，在本文中，我们分析性地证明了为什么值得考虑 Problem (8) ,我们称之为圆锥 MTL,以及为什么它是有利的。具体来说，本文的主要贡献是推导了 ConicMTL 的泛化界，这说明当所有$\lambda_t$都等于 1 时，确实不一定能实现最紧的界。因此，它回答了前面的问题，并证明了考虑圆锥 MTL 的重要性。此外，作为泛化边界的副产品，在 Section $\color{red}{2}$ 中，我们从理论上展示了 Pareto-PathMTL 的好处：与$p=1$相比，$p<1$时 Problem (9)的泛化边界通常更紧密。因此，它解释了帕累托路径 MTL 优于平均 MTL
关于圆锥 MTL, 一个自然的问题是如何选择$\lambda_t$的系数。我们没有像 Pareto-PathMTL 那样启发式地设置它们，而是提出了一个新的圆锥 MTL 模型，该模型通过最小化泛化边界来学习$\lambda_t$ 。与$\lambda_t$值的任何其他设置相比，它可确保我们的新模型实现最严格的泛化范围，并可能导致卓越的性能。新模型在 Section 3 中描述，并在 Section 4 中进行了实验评估。实验结果验证了我们的理论结论：在许多情况下，圆锥 MTL 确实可以优于平均 MTL 和 Pareto-PathMTL,因此，通过最小化泛化边界来学习系数$\lambda_t$是合理且有利的。最后，我们总结了我们在 Section 5 中的工作
在续集中，我们将使用以下符号约定：向量和矩阵以粗体表示。假设向量是列向量。如果$v$是一个向量， 则 $v^{\prime}$ 表示 $v$ 的转置。向量 0 和 1 分别是全零和全一向量。此外，向量之间的$\succeq,\succ,\preceq$和$\prec$将分别代表按组件的$\geq$ , > , $\leq$和<关系。类似地，对于任何 $\boldsymbol{v},\boldsymbol{v}^p$ 表示 $v$ 的按分量求幂
# 2 Generalization Bound
类似于先前对 MTL 方法[1,20,19,13,21,23]的理论分析，在本节中，我们推导了针对圆锥 MTL, $i.e.$, Problem (8)的基于 Rademacher 复杂性的泛化边界。具体来说，我们假设以下形式的$f$和$\Omega(\boldsymbol{w})$用于分类问题：
(11)
$$f(\boldsymbol{w}_{t},\boldsymbol{x}_{t},\boldsymbol{y}_{t})\triangleq\frac{1}{2}\|\boldsymbol{w}_{t}\|^{2}+C\sum_{i=1}^{N}l(y_{t}^{i}\langle\boldsymbol{w}_{t},\phi(\boldsymbol{x}_{t}^{i})\rangle)\\\Omega(\boldsymbol{w})\triangleq\{\boldsymbol{w}=(\boldsymbol{w}_{1},\cdots,\boldsymbol{w}_{T}):\boldsymbol{w}_{t}\in\mathcal{H}_{\boldsymbol{\theta}},\boldsymbol{\theta}\in\Omega(\boldsymbol{\theta})\}$$
其中$l$是保证金损失：
(12)
$$l(x)=\begin{cases}0&\text{if}\:\rho\leq x\\1-x/\rho&\text{if}\:0\leq x\leq\rho\\1&\text{if}\:x\leq0\end{cases}$$
$\phi:\mathcal{X}\to\mathcal{H}_{\boldsymbol{\theta}}$是所有任务的通用特征映射。$\mathcal{H}_{\boldsymbol{\theta}}$是由内核函数$k\triangleq\sum_{m=1}^M\theta_mk_m$定义的 RKHS ,其中$k_m:\mathcal{X}\times\mathcal{X}\to\mathbb{R},m=1,\cdots,M$是预先选择的内核函数。此外，我们假设训练数据$\{\boldsymbol x_t^i,y_t^i\}\in\mathcal{X}\times\mathcal{Y},t=$
3
|
[
[
0.1192626953125,
0.0947265625,
0.88134765625,
0.0947265625,
0.88134765625,
0.10479736328125,
0.1192626953125,
0.10479736328125
],
[
0.1192626953125,
0.10797119140625,
0.62158203125,
0.10797119140625,
0.62158203125,
0.121826171875,
0.1192626953125,
0.121826171875
],
[
0.1405029296875,
0.12371826171875,
0.7998046875,
0.12371826171875,
0.7998046875,
0.136962890625,
0.1405029296875,
0.136962890625
],
[
0.451904296875,
0.147705078125,
0.471435546875,
0.147705078125,
0.471435546875,
0.1610107421875,
0.451904296875,
0.1610107421875
],
[
0.400390625,
0.1622314453125,
0.451904296875,
0.1641845703125,
0.450927734375,
0.179931640625,
0.398681640625,
0.177978515625
],
[
0.441162109375,
0.160400390625,
0.609375,
0.1590576171875,
0.609375,
0.177978515625,
0.441162109375,
0.1787109375
],
[
0.859375,
0.1634521484375,
0.88232421875,
0.1634521484375,
0.88232421875,
0.1767578125,
0.859375,
0.1767578125
],
[
0.390625,
0.1717529296875,
0.4755859375,
0.1767578125,
0.473876953125,
0.1912841796875,
0.3896484375,
0.186279296875
],
[
0.1192626953125,
0.1962890625,
0.88134765625,
0.1962890625,
0.88134765625,
0.210205078125,
0.1192626953125,
0.210205078125
],
[
0.11767578125,
0.2108154296875,
0.5439453125,
0.2108154296875,
0.5439453125,
0.2247314453125,
0.11767578125,
0.2247314453125
],
[
0.36279296875,
0.238037109375,
0.455078125,
0.2342529296875,
0.455810546875,
0.25,
0.363525390625,
0.25390625
],
[
0.34716796875,
0.248779296875,
0.4697265625,
0.248779296875,
0.4697265625,
0.26318359375,
0.34716796875,
0.26318359375
],
[
0.5107421875,
0.24365234375,
0.56640625,
0.24365234375,
0.56640625,
0.255126953125,
0.5107421875,
0.255126953125
],
[
0.50732421875,
0.261962890625,
0.70166015625,
0.26513671875,
0.70166015625,
0.279052734375,
0.50732421875,
0.27587890625
],
[
0.8505859375,
0.264404296875,
0.88232421875,
0.264404296875,
0.88232421875,
0.27978515625,
0.8505859375,
0.27978515625
],
[
0.3447265625,
0.2822265625,
0.488525390625,
0.27783203125,
0.488525390625,
0.29541015625,
0.3447265625,
0.2998046875
],
[
0.295654296875,
0.26513671875,
0.35693359375,
0.261962890625,
0.358642578125,
0.276611328125,
0.29736328125,
0.27978515625
],
[
0.50634765625,
0.285888671875,
0.59619140625,
0.28857421875,
0.59619140625,
0.302978515625,
0.50634765625,
0.300537109375
],
[
0.370849609375,
0.294189453125,
0.464111328125,
0.29541015625,
0.464111328125,
0.306884765625,
0.370849609375,
0.30615234375
],
[
0.1192626953125,
0.3125,
0.88134765625,
0.3125,
0.88134765625,
0.32568359375,
0.1192626953125,
0.32568359375
],
[
0.1168212890625,
0.32763671875,
0.88134765625,
0.32763671875,
0.88134765625,
0.341552734375,
0.1168212890625,
0.341552734375
],
[
0.11767578125,
0.34228515625,
0.8798828125,
0.34228515625,
0.8798828125,
0.35595703125,
0.11767578125,
0.35595703125
],
[
0.11517333984375,
0.356689453125,
0.88134765625,
0.35791015625,
0.88134765625,
0.37255859375,
0.11517333984375,
0.371826171875
],
[
0.11767578125,
0.37255859375,
0.50732421875,
0.37255859375,
0.50732421875,
0.386474609375,
0.11767578125,
0.386474609375
],
[
0.1405029296875,
0.386962890625,
0.88134765625,
0.38818359375,
0.88134765625,
0.40283203125,
0.1405029296875,
0.401611328125
],
[
0.11767578125,
0.404052734375,
0.88134765625,
0.404052734375,
0.88134765625,
0.417236328125,
0.11767578125,
0.417236328125
],
[
0.11767578125,
0.41845703125,
0.88134765625,
0.41845703125,
0.88134765625,
0.431884765625,
0.11767578125,
0.431884765625
],
[
0.1168212890625,
0.434326171875,
0.88232421875,
0.434326171875,
0.88232421875,
0.447509765625,
0.1168212890625,
0.447509765625
],
[
0.11767578125,
0.448974609375,
0.88134765625,
0.448974609375,
0.88134765625,
0.462158203125,
0.11767578125,
0.462158203125
],
[
0.11767578125,
0.46337890625,
0.5361328125,
0.46337890625,
0.5361328125,
0.477783203125,
0.11767578125,
0.477783203125
],
[
0.1422119140625,
0.479248046875,
0.88134765625,
0.479248046875,
0.88134765625,
0.492431640625,
0.1422119140625,
0.492431640625
],
[
0.11767578125,
0.494384765625,
0.88134765625,
0.494384765625,
0.88134765625,
0.50830078125,
0.11767578125,
0.50830078125
],
[
0.11767578125,
0.5087890625,
0.88134765625,
0.5087890625,
0.88134765625,
0.52294921875,
0.11767578125,
0.52294921875
],
[
0.11767578125,
0.5244140625,
0.88232421875,
0.5244140625,
0.88232421875,
0.53857421875,
0.11767578125,
0.53857421875
],
[
0.11767578125,
0.5390625,
0.88232421875,
0.5390625,
0.88232421875,
0.55322265625,
0.11767578125,
0.55322265625
],
[
0.1168212890625,
0.5537109375,
0.420654296875,
0.55322265625,
0.420654296875,
0.5673828125,
0.1168212890625,
0.568359375
],
[
0.1422119140625,
0.57080078125,
0.88134765625,
0.57080078125,
0.88134765625,
0.583984375,
0.1422119140625,
0.583984375
],
[
0.11767578125,
0.58544921875,
0.88134765625,
0.58544921875,
0.88134765625,
0.599609375,
0.11767578125,
0.599609375
],
[
0.1168212890625,
0.60107421875,
0.705078125,
0.60107421875,
0.705078125,
0.61572265625,
0.1168212890625,
0.61572265625
],
[
0.11517333984375,
0.63818359375,
0.414306640625,
0.63818359375,
0.414306640625,
0.65283203125,
0.11517333984375,
0.65283203125
],
[
0.1168212890625,
0.66845703125,
0.88232421875,
0.66845703125,
0.88232421875,
0.68310546875,
0.1168212890625,
0.68310546875
],
[
0.1168212890625,
0.6845703125,
0.88134765625,
0.6845703125,
0.88134765625,
0.69775390625,
0.1168212890625,
0.69775390625
],
[
0.1192626953125,
0.69970703125,
0.156005859375,
0.69970703125,
0.156005859375,
0.7109375,
0.1192626953125,
0.7109375
],
[
0.529296875,
0.72607421875,
0.54150390625,
0.72607421875,
0.54150390625,
0.734375,
0.529296875,
0.734375
],
[
0.32177734375,
0.7373046875,
0.431396484375,
0.7373046875,
0.431396484375,
0.7548828125,
0.32177734375,
0.7548828125
],
[
0.43310546875,
0.736328125,
0.66015625,
0.736328125,
0.66015625,
0.75390625,
0.43310546875,
0.75390625
],
[
0.42724609375,
0.7470703125,
0.4345703125,
0.7470703125,
0.4345703125,
0.75,
0.42724609375,
0.75
],
[
0.43701171875,
0.74755859375,
0.44775390625,
0.75048828125,
0.44287109375,
0.76025390625,
0.431396484375,
0.7568359375
],
[
0.52294921875,
0.75244140625,
0.5458984375,
0.75244140625,
0.5458984375,
0.76513671875,
0.52294921875,
0.76513671875
],
[
0.8505859375,
0.74755859375,
0.8837890625,
0.74755859375,
0.8837890625,
0.7646484375,
0.8505859375,
0.7646484375
],
[
0.32177734375,
0.77099609375,
0.6748046875,
0.77099609375,
0.6748046875,
0.7841796875,
0.32177734375,
0.7841796875
],
[
0.11767578125,
0.79443359375,
0.26953125,
0.79443359375,
0.26953125,
0.8046875,
0.11767578125,
0.8046875
],
[
0.441162109375,
0.8212890625,
0.463134765625,
0.8212890625,
0.463134765625,
0.8359375,
0.441162109375,
0.8359375
],
[
0.52294921875,
0.8212890625,
0.58251953125,
0.8212890625,
0.58251953125,
0.8359375,
0.52294921875,
0.8359375
],
[
0.384033203125,
0.83837890625,
0.61279296875,
0.83837890625,
0.61279296875,
0.85595703125,
0.384033203125,
0.85595703125
],
[
0.8505859375,
0.8388671875,
0.88232421875,
0.8388671875,
0.88232421875,
0.85498046875,
0.8505859375,
0.85498046875
],
[
0.441162109375,
0.85693359375,
0.461669921875,
0.85693359375,
0.461669921875,
0.87255859375,
0.441162109375,
0.87255859375
],
[
0.52294921875,
0.85791015625,
0.58251953125,
0.85791015625,
0.58251953125,
0.87255859375,
0.52294921875,
0.87255859375
],
[
0.1168212890625,
0.8798828125,
0.88232421875,
0.8798828125,
0.88232421875,
0.89794921875,
0.1168212890625,
0.89794921875
],
[
0.1168212890625,
0.8984375,
0.8798828125,
0.8984375,
0.8798828125,
0.912109375,
0.1168212890625,
0.912109375
],
[
0.49267578125,
0.9375,
0.50634765625,
0.9375,
0.50634765625,
0.94873046875,
0.49267578125,
0.94873046875
]
] |
[
"ParetoFront上可能还有其他点可以为每个任务带来更好的泛化性能，从而为T任务带来更好的平均性能。",
"因此，不是求解Problem（1），而是可以通过优化Problem（8）来实现",
"在[16]中执行了这些方面的先前工作。作者考虑了以下MTL公式，名为Pareto-PathMTL",
"T",
"min.[",
"[(f(wt, at, yt)P]1/p",
"(9)",
"wEn(w)t=)",
"，假设所有目标函数均为正，则在p≥1时最小化目标的Lp范数，在0<p<1时最小化Lp伪范数。事实",
"证明，对于任何p>0，Problem(9）都等同于Problem（8)",
"f(wt,at.y.)p-1",
"Zt-i(f(wt,et.yt))p",
"ifp>1",
"ifp=1,t=1,,T",
"(10)",
"Z(f(wtEt.u)P",
"^t=31",
"if 0<p<1",
"(wt,ety)1-p",
"因此，通过改变p>0，Problem（9）的解决方案在Problem（7）的ParetoFront上跟踪路径。虽然平均",
"MTL等于Problem（9），但当p=1时，已证明在基于SupportVectorMachine（SVM）的MKL设置中",
"与P=1相比，P<1的实验结果通常更好。尽管获得的优越结果与我们先前的论点密切相关，但作者并未",
"提供严格的基础，说明考虑T任务目标平均值以外的目标函数的优势。因此，到目前为止，在论文的目标",
"函数中使用L-（伪）范数仍然是其方法的启发式元素",
"鉴于刚才提到的平均MTL的潜在缺点以及帕累托路径MTL缺乏支持理论，在本文中，我们分析性地",
"证明了为什么值得考虑Problem（8），我们称之为圆锥MTL，以及为什么它是有利的。具体来说，本文的",
"主要贡献是推导了ConicMTL的泛化界，这说明当所有入t都等于1时，确实不一定能实现最紧的界。。因",
"此，它回答了前面的问题，并证明了考虑圆锥MTL的重要性。此外，作为泛化边界的副产品，在Section2",
"中，我们从理论上展示了Pareto-PathMTL的好处：与p=1相比，p<1时Problem（9）的泛化边界通常",
"更紧密。因此，它解释了帕累托路径MTL优于平均MTL",
"关于圆锥MTL，一个自然的问题是如何选择入t的系数。我们没有像Pareto-PathMTL那样启发式地设",
"置它们，而是提出了一个新的圆锥MTL模型，该模型通过最小化泛化边界来学习入t。与入t值的任何其他",
"设置相比，它可确保我们的新模型实现最严格的泛化范围，并可能导致卓越的性能。新模型在Section3中",
"描述，并在Section4中进行了实验评估。实验结果验证了我们的理论结论：在许多情况下，圆锥MTL确实",
"可以优于平均MTL和Pareto-PathMTL，因此，通过最小化泛化边界来学习系数入t是合理且有利的。最",
"后，我们总结了我们在Section5中的工作",
"在续集中，我们将使用以下符号约定：向量和矩阵以粗体表示。假设向量是列向量。如果是一个向量",
"则'表示的转置。向量0和1分别是全零和全一向量。此外，向量之间的，，二和<将分别代表",
"按组件的≥，>，≤和<关系。类似地，对于任何，uP表示的按分量求幂",
"2Generalization Bound",
"类似于先前对MTL方法[1，20，19，13，21，23]的理论分析，在本节中，我们推导了针对圆锥MTL，i.e.",
"Problem（8）的基于Rademacher复杂性的泛化边界。具体来说，我们假设以下形式的f和2(w）用于分类",
"问题：",
"N",
"f(Wt, at, Yt) ≤",
"llwtll2 + ci(yt(wt, p(c))",
"",
"2",
"1=1",
"(11)",
"2(w) ≤ [w= (w1,., wT) : wt EHo,0 E2(0))",
"其中1是保证金损失：",
"(0",
"ifp≤r",
"l()=1-/pif0≤≤p",
"(12)",
"(1",
"if r≤0",
"Φ：→He是所有任务的通用特征映射。He是由内核函数k会Zm-0mkm定义的RKHS，其中",
"km×→R,m=1,,M是预先选择的内核函数。此外，我们假设训练数据[a,yE×y,t=",
"3"
] |
{
"arxiv_id": "1408.4714",
"language": "zh",
"page_idx": 2
}
|
||
1408.4714-zh-3
|
1408.4714-zh-3.jpg
|
$1,\cdots,T,i=1,\cdots,N$是从概率分布$P_t(X_t,Y_t)$中提取的，其中$X_t$和$Y_t$分别是输人和输出空间中的随机变量。请注意，在这里，我们假设所有任务具有相等数量的训练数据并共享一个公共内核函数。这两个假设是为了简化符号和说明，它们不会影响将我们的结果扩展到更一般的情况，即每个任务都有任意数量的训练样本，并且使用部分共享的核函数；在后一种情况下，只有相关任务可以共享公共核函数，因此，减少了“负转移”的影响
将(11)代人 Problem (8)并基于 Proposition 1,不难看出，对于 Equation (11) 中的任何$C$ ,都存在
一个$R>0$ ,使得 Problem (8)等效于以下问题
$$\min_{\boldsymbol{w}\in\Omega(\boldsymbol{w})}\:\sum_{t=1}^{T}\sum_{i=1}^{N_{t}}\lambda_{t}l(y_{t}^{i}\langle\boldsymbol{w}_{t},\phi(\boldsymbol{x}_{t}^{i})\rangle)\\s.t.\sum_{t=1}^{T}\lambda_{t}\|\boldsymbol{w}_{t}\|^{2}\leq R$$
(13)
显然，求解 Problem (13)是在假设空间$\mathcal{F}_\mathrm{\lambda}$中选择$\boldsymbol w$的过程，从而使经验损失 Problem (13)的目标函
数$i.e.$最小化。相关的假设空间定义如下：
(14)
$$\mathcal{F}_{\boldsymbol{\lambda}}\triangleq\{\boldsymbol{w}=(\boldsymbol{w}_1,\cdots,\boldsymbol{w}_T):\sum_{t=1}^T\lambda_t\|\boldsymbol{w}_t\|^2\leq R,\boldsymbol{w}_t\in\mathcal{H}_{\boldsymbol{\theta}},\boldsymbol{\theta}\in\Omega(\boldsymbol{\theta})\}$$
通过定义圆锥 MTL 预期误差$er(\boldsymbol{w})$和经验损失$\hat{er}_{\boldsymbol{\lambda}}(\boldsymbol{w})$如下
(15)
$$er(\boldsymbol{w})=\dfrac{1}{T}\sum_{t=1}^TE[\mathbf{1}_{(-\infty,0]}(Y_t\langle\boldsymbol{w}_t,\phi(X_t)\rangle)]$$
$$\hat{er}_{\boldsymbol{\lambda}}(\boldsymbol{w})=\frac{1}{TN}\sum_{t=1}^{T}\sum_{i=1}^{N}\lambda_{t}l(y_{t}^{i}\langle\boldsymbol{w}_{t},\phi(\boldsymbol{x}_{t}^{i})\rangle)$$
(16)
我们的主要贡献之一是以下定理，该定理给出了 Problem (13)在任何$\lambda_t\in(1,r_\lambda),\forall t=1,\cdots,T$的基于
MKL 的圆锥 MTL 的上下文中的泛化界，其中$r_\lambda$是$\lambda_t$的预先指定的上限
Theorem $1.Forfaxed\rho>0,r_\lambda\in\mathbb{N}withr_\lambda>1,andforany\boldsymbol{\lambda}=[\lambda_1,\cdots,\lambda_T]^{\prime},\lambda_t\in(1,r_\lambda),\forall t=1,\cdots,T$
, $\boldsymbol{w}\in \mathcal{F_{\boldsymbol{\lambda }}}$ , $0< \delta < 1$ , the following generalization bound holds with probability $at$ least $1- \delta :$
$$er(\boldsymbol{w})\leq\hat{er}_{\boldsymbol{\lambda}}(\boldsymbol{w})+\frac{\sqrt{2}r_{\lambda}}{\rho}R(\mathcal{F}_{\boldsymbol{\lambda}})+\sqrt{\frac{9}{TN}\ln\left(\frac{2r_{\lambda}}{T}\sum_{t=1}^{T}\frac{1}{\lambda_{t}}\right)}+\sqrt{\frac{9\ln\frac{1}{\delta}}{2TN}}$$
(17)
where $R( \mathcal{F} _{\mathbf{\lambda } })$ $is$ the empirical Rademacher complexity $of$ the hypothesis space $\mathcal{F} _{\mathbf{\lambda } }$ , which $is$ defined $as$
$$R(\mathcal{F}_{\boldsymbol{\lambda}})\triangleq\frac{2}{TN}E[\sup\limits_{\boldsymbol{w}\in\mathcal{F}_{\boldsymbol{\lambda}}}\sum\limits_{t=1}^{T}\sum\limits_{i=1}^{N}\sigma_{t}^{i}\langle\boldsymbol{w}_{t},\phi(\boldsymbol{x}_{t}^{i})\rangle]$$
(18)
and the $\sigma _{t}^{i}$ 's are $i. i. d.$ Rademacher-distributed $( i. e. $, Bernoulli(1/2) -distributed random variables with sample
space $\{ - 1, + 1\}$ ).
基于 Theorem 1 ,人们有动力选择最小化泛化边界的$\boldsymbol{\lambda}$,而不是像 Equation (10) 中那样启发式地选择
$\lambda$ ,这是 [16] 中建议的。实际上，这样做并不能保证获得最严格的泛化约束
但是，在提出我们的新的 ConicMTL 模型以最小化泛化边界之前，仍然有兴趣从理论上分析为什么Pareto-PathMTL, $i.e.$, Problem (9)在$0<p<1$时通常享有更好的泛化性能，而不是$p=1$时，如Section 1 中所述。虽然在 [16] 中没有给出分析，但幸运的是，我们可以通过利用 Theorem 1 并借助以下两个定理来提供模型在$0<p<1$中的良好性能的一些见解
Theorem $2. For\lambda \succ \mathbf{0} , the\textit{ empirical Rademacher complexity R( }\mathcal{F_{\boldsymbol{\lambda }}) }is\textit{ monotonically decreasing witheris }$
$respect\textit{ to each }\lambda _t, t= 1, \cdots , T$ .
4
|
[
[
0.1168212890625,
0.0921630859375,
0.88134765625,
0.0921630859375,
0.88134765625,
0.1060791015625,
0.1168212890625,
0.1060791015625
],
[
0.11767578125,
0.10797119140625,
0.88134765625,
0.10797119140625,
0.88134765625,
0.121826171875,
0.11767578125,
0.121826171875
],
[
0.1192626953125,
0.125,
0.8798828125,
0.125,
0.8798828125,
0.1351318359375,
0.1192626953125,
0.1351318359375
],
[
0.11767578125,
0.1383056640625,
0.88134765625,
0.1383056640625,
0.88134765625,
0.1514892578125,
0.11767578125,
0.1514892578125
],
[
0.1168212890625,
0.154052734375,
0.2156982421875,
0.154052734375,
0.2156982421875,
0.1654052734375,
0.1168212890625,
0.1654052734375
],
[
0.1380615234375,
0.1666259765625,
0.88232421875,
0.1673583984375,
0.88232421875,
0.1817626953125,
0.1380615234375,
0.18115234375
],
[
0.1258544921875,
0.18310546875,
0.464111328125,
0.18310546875,
0.464111328125,
0.1962890625,
0.1258544921875,
0.1962890625
],
[
0.44287109375,
0.2083740234375,
0.488525390625,
0.2083740234375,
0.488525390625,
0.222900390625,
0.44287109375,
0.222900390625
],
[
0.393798828125,
0.2247314453125,
0.42724609375,
0.2247314453125,
0.42724609375,
0.2342529296875,
0.393798828125,
0.2342529296875
],
[
0.43701171875,
0.2197265625,
0.61669921875,
0.222900390625,
0.61669921875,
0.2374267578125,
0.43701171875,
0.2342529296875
],
[
0.38232421875,
0.2342529296875,
0.435546875,
0.2342529296875,
0.435546875,
0.2442626953125,
0.38232421875,
0.2442626953125
],
[
0.428955078125,
0.236083984375,
0.488525390625,
0.238037109375,
0.488525390625,
0.250732421875,
0.428955078125,
0.248779296875
],
[
0.8505859375,
0.24365234375,
0.8837890625,
0.24365234375,
0.8837890625,
0.2587890625,
0.8505859375,
0.2587890625
],
[
0.439453125,
0.25830078125,
0.451904296875,
0.25,
0.463134765625,
0.25830078125,
0.450927734375,
0.266357421875
],
[
0.410888671875,
0.26513671875,
0.560546875,
0.26513671875,
0.560546875,
0.28271484375,
0.410888671875,
0.28271484375
],
[
0.438720703125,
0.281005859375,
0.463134765625,
0.281005859375,
0.463134765625,
0.294189453125,
0.438720703125,
0.294189453125
],
[
0.1422119140625,
0.3017578125,
0.88134765625,
0.3017578125,
0.88134765625,
0.31494140625,
0.1422119140625,
0.31494140625
],
[
0.1192626953125,
0.318115234375,
0.416748046875,
0.318115234375,
0.416748046875,
0.328857421875,
0.1192626953125,
0.328857421875
],
[
0.4697265625,
0.343505859375,
0.48193359375,
0.343505859375,
0.48193359375,
0.3515625,
0.4697265625,
0.3515625
],
[
0.2646484375,
0.3515625,
0.73046875,
0.3515625,
0.73046875,
0.369384765625,
0.2646484375,
0.369384765625
],
[
0.8505859375,
0.354736328125,
0.8837890625,
0.354736328125,
0.8837890625,
0.371826171875,
0.8505859375,
0.371826171875
],
[
0.465576171875,
0.37255859375,
0.48779296875,
0.37255859375,
0.48779296875,
0.380615234375,
0.465576171875,
0.380615234375
],
[
0.1405029296875,
0.391357421875,
0.580078125,
0.391357421875,
0.580078125,
0.40478515625,
0.1405029296875,
0.40478515625
],
[
0.419189453125,
0.42041015625,
0.443603515625,
0.42041015625,
0.443603515625,
0.43310546875,
0.419189453125,
0.43310546875
],
[
0.43701171875,
0.417236328125,
0.455078125,
0.417236328125,
0.455078125,
0.4267578125,
0.43701171875,
0.4267578125
],
[
0.35302734375,
0.4287109375,
0.4150390625,
0.4287109375,
0.4150390625,
0.443115234375,
0.35302734375,
0.443115234375
],
[
0.43310546875,
0.4267578125,
0.64794921875,
0.4267578125,
0.64794921875,
0.447509765625,
0.43310546875,
0.447509765625
],
[
0.85107421875,
0.4287109375,
0.8837890625,
0.4287109375,
0.8837890625,
0.4443359375,
0.85107421875,
0.4443359375
],
[
0.420654296875,
0.438232421875,
0.4345703125,
0.438232421875,
0.4345703125,
0.448974609375,
0.420654296875,
0.448974609375
],
[
0.435546875,
0.4443359375,
0.459228515625,
0.4443359375,
0.459228515625,
0.45703125,
0.435546875,
0.45703125
],
[
0.44287109375,
0.46728515625,
0.465576171875,
0.473388671875,
0.4599609375,
0.486083984375,
0.43701171875,
0.479736328125
],
[
0.465576171875,
0.46337890625,
0.5107421875,
0.46337890625,
0.5107421875,
0.477783203125,
0.465576171875,
0.477783203125
],
[
0.404296875,
0.473388671875,
0.64453125,
0.46533203125,
0.646484375,
0.49755859375,
0.406005859375,
0.505859375
],
[
0.359375,
0.48095703125,
0.4248046875,
0.48095703125,
0.4248046875,
0.492431640625,
0.359375,
0.492431640625
],
[
0.85107421875,
0.477783203125,
0.8837890625,
0.477783203125,
0.8837890625,
0.494384765625,
0.85107421875,
0.494384765625
],
[
0.463134765625,
0.492431640625,
0.51318359375,
0.494384765625,
0.51318359375,
0.505859375,
0.463134765625,
0.50390625
],
[
0.11767578125,
0.51318359375,
0.8798828125,
0.51318359375,
0.8798828125,
0.52783203125,
0.11767578125,
0.52783203125
],
[
0.1168212890625,
0.52880859375,
0.65185546875,
0.52880859375,
0.65185546875,
0.54248046875,
0.1168212890625,
0.54248046875
],
[
0.1168212890625,
0.55322265625,
0.8798828125,
0.55322265625,
0.8798828125,
0.56640625,
0.1168212890625,
0.56640625
],
[
0.11767578125,
0.568359375,
0.79345703125,
0.568359375,
0.79345703125,
0.58349609375,
0.11767578125,
0.58349609375
],
[
0.37744140625,
0.59521484375,
0.640625,
0.59521484375,
0.640625,
0.642578125,
0.37744140625,
0.642578125
],
[
0.69677734375,
0.60107421875,
0.740234375,
0.60302734375,
0.73876953125,
0.619140625,
0.6953125,
0.6171875
],
[
0.250732421875,
0.611328125,
0.398681640625,
0.60498046875,
0.400390625,
0.626953125,
0.25244140625,
0.63330078125
],
[
0.6357421875,
0.609375,
0.646484375,
0.609375,
0.646484375,
0.6123046875,
0.6357421875,
0.6123046875
],
[
0.689453125,
0.6162109375,
0.73876953125,
0.619140625,
0.73779296875,
0.63525390625,
0.68896484375,
0.6318359375
],
[
0.85107421875,
0.6123046875,
0.8837890625,
0.6123046875,
0.8837890625,
0.62744140625,
0.85107421875,
0.62744140625
],
[
0.404296875,
0.62548828125,
0.419189453125,
0.62548828125,
0.419189453125,
0.63525390625,
0.404296875,
0.63525390625
],
[
0.63818359375,
0.625,
0.65283203125,
0.625,
0.65283203125,
0.6318359375,
0.63818359375,
0.6318359375
],
[
0.1192626953125,
0.6484375,
0.8408203125,
0.6484375,
0.8408203125,
0.662109375,
0.1192626953125,
0.662109375
],
[
0.50732421875,
0.6748046875,
0.54736328125,
0.6748046875,
0.54736328125,
0.68505859375,
0.50732421875,
0.68505859375
],
[
0.34228515625,
0.6875,
0.455078125,
0.6767578125,
0.459228515625,
0.69873046875,
0.346435546875,
0.708984375
],
[
0.441162109375,
0.68505859375,
0.65283203125,
0.68603515625,
0.65283203125,
0.7041015625,
0.441162109375,
0.70263671875
],
[
0.8505859375,
0.68798828125,
0.8837890625,
0.68798828125,
0.8837890625,
0.7041015625,
0.8505859375,
0.7041015625
],
[
0.408447265625,
0.69580078125,
0.455810546875,
0.69775390625,
0.455078125,
0.71044921875,
0.406982421875,
0.70849609375
],
[
0.450927734375,
0.70068359375,
0.55126953125,
0.70556640625,
0.55126953125,
0.71630859375,
0.450927734375,
0.7109375
],
[
0.11767578125,
0.7236328125,
0.88134765625,
0.7236328125,
0.88134765625,
0.7373046875,
0.11767578125,
0.7373046875
],
[
0.11767578125,
0.7392578125,
0.2442626953125,
0.7392578125,
0.2442626953125,
0.75244140625,
0.11767578125,
0.75244140625
],
[
0.1405029296875,
0.76318359375,
0.88232421875,
0.76318359375,
0.88232421875,
0.77783203125,
0.1405029296875,
0.77783203125
],
[
0.11517333984375,
0.77978515625,
0.65869140625,
0.77978515625,
0.65869140625,
0.79345703125,
0.11517333984375,
0.79345703125
],
[
0.1422119140625,
0.79443359375,
0.8798828125,
0.79443359375,
0.8798828125,
0.80810546875,
0.1422119140625,
0.80810546875
],
[
0.1168212890625,
0.8095703125,
0.88134765625,
0.8095703125,
0.88134765625,
0.82373046875,
0.1168212890625,
0.82373046875
],
[
0.1168212890625,
0.82470703125,
0.88134765625,
0.82470703125,
0.88134765625,
0.83837890625,
0.1168212890625,
0.83837890625
],
[
0.11767578125,
0.8388671875,
0.52294921875,
0.8388671875,
0.52294921875,
0.85302734375,
0.11767578125,
0.85302734375
],
[
0.11517333984375,
0.86279296875,
0.88134765625,
0.8642578125,
0.88134765625,
0.87890625,
0.11517333984375,
0.87744140625
],
[
0.1168212890625,
0.8798828125,
0.3447265625,
0.87890625,
0.3447265625,
0.89306640625,
0.1168212890625,
0.89453125
],
[
0.494384765625,
0.93603515625,
0.50634765625,
0.93603515625,
0.50634765625,
0.94873046875,
0.494384765625,
0.94873046875
]
] |
[
"1，，T,i=1,，N是从概率分布Pt（Xt,Yt）中提取的，其中Xt和Yt分别是输人和输出空间中的随机变",
"量。请注意，在这里，我们假设所有任务具有相等数量的训练数据并共享一个公共内核函数。这两个假设是",
"为了简化符号和说明，它们不会影响将我们的结果扩展到更一般的情况，即每个任务都有任意数量的训练样",
"本，并且使用部分共享的核函数；在后一种情况下，只有相关任务可以共享公共核函数，因此，减少了“负",
"转移”的影响",
"将（11）代人Problem（8）并基于Proposition1，不难看出，对于Equation（11）中的任何C，都存在",
"个R>0，使得Problem（8）等效于以下问题",
"TN",
"min",
"Fatl(yi(wt, (a)))",
"WE2(W)",
"t=l=i",
"(13)",
"2",
"s.t. Z^tllwtl2 ≤R",
"t=i",
"显然，求解Problem（13）是在假设空间F>中选择w的过程，从而使经验损失Problem（13）的目标函",
"数i.e.最小化。相关的假设空间定义如下：",
"T",
"Fx ≤ [w = (wi,..., wWT) : >twt2 ≤ R, wt E He,0 E2(0))",
"(14)",
"t-1",
"通过定义圆锥MTL预期误差er(w）和经验损失er>（w）如下",
"1",
"T",
"er(w) =",
"ZE[1(-00,0)(Y(wt, d(X)]",
"(15)",
"T",
"t=1",
"1",
"TN",
"(oi(,(a)",
"er>(w)=",
"(16)",
"=11",
"我们的主要贡献之一是以下定理，该定理给出了Problem（13）在任何入tE（1,r>),Vt=1,,T的基于",
"MKL的圆锥MTL的上下文中的泛化界，其中r入是入t的预先指定的上限",
"Theorem 1. For fired p > 0,r> E N with r> > 1, and for any > =[Ai,..:,Arl', t E(1,r), Vt = 1,:,T",
", w E Fx , 0< d <1 , the following generalization bound holds with probability at least 1 - d :",
"+)",
"/9ln",
"er(w) ≤erx(w)+",
"",
"V2TN",
"(17)",
"p",
"t",
"where R(F) is the empirical Rademacher complerity of the hypothesis space F which is defined as",
"TN",
"R(A)≤2E",
"E[ sup o(wt, d(ct)]",
"(18)",
"TNE",
"wEFxt=li-l",
"and the o 's are i.i.d. Rademacher-distributed (i.e., Bernoulli(1/2) -distributed random variables with sample",
"space (-1, +1) ).",
"基于Theorem1，人们有动力选择最小化泛化边界的入，而不是像Equation（10）中那样启发式地选择",
"入，这是[16]中建议的。实际上，这样做并不能保证获得最严格的泛化约束",
"但是，在提出我们的新的ConicMTL模型以最小化泛化边界之前，仍然有兴趣从理论上分析为什么",
"Pareto-PathMTL，i.e.，Problem（9）在0<p<1时通常享有更好的泛化性能，而不是p=1时，如",
"Section1中所述。虽然在【16]中没有给出分析，但幸运的是，我们可以通过利用Theorem1并借助以下两",
"个定理来提供模型在0<P<1中的良好性能的一些见解",
"Theorem 2. For > o ,the empirical Rademacher complerity R(F) is monotonically decreasing with",
"respecttoeach^t,t=1,...,T",
"4"
] |
{
"arxiv_id": "1408.4714",
"language": "zh",
"page_idx": 3
}
|
||
1408.4714-zh-4
|
1408.4714-zh-4.jpg
|
$\textbf{Theorem 3. }Assume$ $f( \boldsymbol{w}_t, \boldsymbol{x}_t, \boldsymbol{y}_t)$ $> 0, \forall t$ = $1, \cdots , T$ . For $\boldsymbol{\lambda }$ that $is$ defined $in$ Equation (10), when
$0< p< 1$ , $we$ have $\lambda _{t}> 1$ and $\lambda _{t}$ $is$ monotonically decreasing with respect $to$ $p$ , $\forall t= 1, \cdots , T$ .
基于 Equation (10) ,如果$f(\boldsymbol w_t,\boldsymbol x_t,\boldsymbol y_t)>0,\forall t=1,\cdots,T$ ,则必须存在固定的$r_\lambda>0$ ,例如$\lambda_t\in(1,r_\lambda),\forall t=1,\cdots,T$。因此，我们可以分析基于 Theorem $\color{red}{1\text{ 的 Pareto-PathMTL 的泛化界，当}}$ $0<p<1$ 。尽管 Theorem 1 不适合$p=1$时的情况，但我们可以通过让$p$无限接近 1 来近似其边界
以上两个定理表明，当$0<p<1$。因此，泛化界中的第二项随着$p$的减小而减小。基于 Theorem 3 的绑定中的第三项也是如此。因此，毫不奇怪，当$0<p<1$时，泛化性能通常比$p=1$时更好，并且当$p$减小时，期望性能得到改善是合理的。实际上，在[16]的实验中报告了这种单调性：当$p$减小时，分类精度通常是单调增加的。值得一提的是，尽管很少观察到，但如果泛化边界中的第一项$i.e.($经验损失)随着$p$的减小而迅速增长，则我们可能不会在性能上具有这种单调性。但是，泛化界的单调行为(经验损失除外)仍然足以解释 Problem (9) 的实验结果，这证明了采用目标函数的任意加权圆锥组合而不是使用这些函数的平均值的基本原理
最后，我们提供了两个定理，它们不仅用于 Theorem 1 的证明，而且可能会自行引起人们的兴趣。随后，
在下一节中，我们将描述我们的新 MTL 模型
$\textbf{Theorem 4. Given }\gamma \triangleq [ \gamma _{1}, \cdots , \gamma _{T}] ^{\prime }with$ $\gamma \succ \mathbf{0}$ , defne
$$R(\mathcal{F}_{\boldsymbol{\lambda}},\boldsymbol{\gamma})=\dfrac{2}{TN}E[\sup\limits_{\boldsymbol{w}\in\mathcal{F}_{\boldsymbol{\lambda}}}\sum\limits_{t=1}^{T}\sum\limits_{i=1}^{N}\gamma_{t}\sigma_{t}^{i}\langle\boldsymbol{w}_{t},\phi(\boldsymbol{x}_{t}^{i})\rangle]$$
(19)
For faxed $\boldsymbol{\lambda }\succ \boldsymbol{0}$ , $R( \mathcal{F} _{\boldsymbol{\lambda }}, \boldsymbol{\gamma })$ $is$ monotonically increasing with respect $to$ each $\gamma _{t}$ .
$\textbf{Theorem 5. For fl}xed$ $r_{\lambda }\geq 1$ , $\rho > 0$ , $\boldsymbol{\lambda }= [ \lambda _{1}, \cdots , \lambda _{T}] ^{\prime }, \lambda _{t}\in [ 1, r_{\lambda }] , \forall t= 1, \cdots , T$ , and for any $\boldsymbol{w}\in \mathcal{F} _{{\boldsymbol{\lambda }}}$ ,
$0< \delta < 1$ , the following generalization bound holds with probability $at$ least $1- \delta$ :
$$er(\boldsymbol{w})\leq\hat{er}_{\boldsymbol{\lambda}}(\boldsymbol{w})+\frac{r_{\lambda}}{\rho}R(\mathcal{F}_{\boldsymbol{\lambda}})+\sqrt{\frac{9\ln\frac{1}{\delta}}{2TN}}$$
(20)
请注意，Theorem 5 和 Theorem 1之间的区别在于，Theorem 1 对任何$\lambda_t\in(1,r_\lambda)$都有效，而 Theorem 5仅对固定的$\lambda_t\in[1,r_\lambda]$有效。虽然 Theorem 1 中给出的界限更为笼统，但由于 (17)中的附加第三项以及$\sqrt{2}$乘以经验 Rademacher 复杂度而使其更加宽松
## 3 A New MTL Model
在本节中，我们提出了新的 MTL 模型。受 Theorem 1 中泛化约束的启发，我们的模型被公式化为通过最小
化约束来选择$w$和$\lambda$
$$\hat{e}r_{\boldsymbol{\lambda}}(\boldsymbol{w})+\frac{\sqrt{2}r_{\lambda}}{\rho}R(\mathcal{F}_{\boldsymbol{\lambda}})+\sqrt{\frac{9}{TN}\ln\left(\frac{2r_{\lambda}}{T}\sum_{t=1}^{T}\frac{1}{\lambda_{t}}\right)}+\sqrt{\frac{9\ln\frac{1}{\delta}}{2TN}}$$
(21)
而不是启发式地选择系数$\lambda$ ,例如通过 [16] 中的 Equation (10) 。请注意，边界的最后一项不依赖于任何模型参数，而第三项对边界只有很小的影响，当$\lambda_t\in(1,r_\lambda)$ 。因此，我们省略了这两个术语，并提出了以下模型：
$$\min_{\boldsymbol{w},\boldsymbol{\lambda}}\hat{er}_{\boldsymbol{\lambda}}(\boldsymbol{w})+\frac{\sqrt{2}r_{\lambda}}{\rho}R(\mathcal{F}_{\boldsymbol{\lambda}})$$
(22)
此外，由于$R(\mathcal{F}_{\mathcal{\lambda}})$的复杂性，很难直接优化 Problem (22) 。因此，在下面的定理中，我们证明了$R(\mathcal{F}_{\mathcal{\lambda}})$
的一个上限，它产生了一个更简单的表达式。我们提醒读者，假设空间$\mathcal{F}_\mathrm{\lambda}$被定义为
$$\mathcal{F}_{\boldsymbol{\lambda}}\triangleq\{\boldsymbol{w}=(\boldsymbol{w}_1,\cdots,\boldsymbol{w}_T):\sum_{t=1}^T\lambda_t\|\boldsymbol{w}_t\|^2\leq R,\boldsymbol{w}_t\in\mathcal{H}_{\boldsymbol{\theta}},\boldsymbol{\theta}\in\Omega(\boldsymbol{\theta})\}$$
(23)
其中$\mathcal{H}_{\theta}$是由内核函数$k\triangleq\sum_m=1^M\theta_mk_m$定义的 RKHS
5
|
[
[
0.1168212890625,
0.0921630859375,
0.8837890625,
0.0921630859375,
0.8837890625,
0.106689453125,
0.1168212890625,
0.106689453125
],
[
0.11517333984375,
0.10797119140625,
0.8095703125,
0.10797119140625,
0.8095703125,
0.121826171875,
0.11517333984375,
0.121826171875
],
[
0.1422119140625,
0.1307373046875,
0.88134765625,
0.1307373046875,
0.88134765625,
0.1458740234375,
0.1422119140625,
0.1458740234375
],
[
0.1168212890625,
0.1458740234375,
0.88134765625,
0.1458740234375,
0.88134765625,
0.1590576171875,
0.1168212890625,
0.1590576171875
],
[
0.11517333984375,
0.160400390625,
0.83642578125,
0.160400390625,
0.83642578125,
0.173583984375,
0.11517333984375,
0.173583984375
],
[
0.1405029296875,
0.1749267578125,
0.8798828125,
0.1749267578125,
0.8798828125,
0.1881103515625,
0.1405029296875,
0.1881103515625
],
[
0.11767578125,
0.1900634765625,
0.8798828125,
0.1900634765625,
0.8798828125,
0.203857421875,
0.11767578125,
0.203857421875
],
[
0.11767578125,
0.205810546875,
0.88134765625,
0.205810546875,
0.88134765625,
0.2197265625,
0.11767578125,
0.2197265625
],
[
0.1168212890625,
0.2197265625,
0.88134765625,
0.2203369140625,
0.88134765625,
0.23486328125,
0.1168212890625,
0.2342529296875
],
[
0.11767578125,
0.236083984375,
0.88232421875,
0.236083984375,
0.88232421875,
0.25,
0.11767578125,
0.25
],
[
0.11767578125,
0.250732421875,
0.88134765625,
0.250732421875,
0.88134765625,
0.26513671875,
0.11767578125,
0.26513671875
],
[
0.1168212890625,
0.266357421875,
0.2320556640625,
0.266357421875,
0.2320556640625,
0.27978515625,
0.1168212890625,
0.27978515625
],
[
0.1405029296875,
0.2822265625,
0.88232421875,
0.2822265625,
0.88232421875,
0.29541015625,
0.1405029296875,
0.29541015625
],
[
0.11517333984375,
0.296630859375,
0.44287109375,
0.29541015625,
0.44287109375,
0.31005859375,
0.11517333984375,
0.311279296875
],
[
0.11517333984375,
0.317626953125,
0.52783203125,
0.318115234375,
0.52783203125,
0.333251953125,
0.11517333984375,
0.33203125
],
[
0.42724609375,
0.347900390625,
0.450927734375,
0.354736328125,
0.443603515625,
0.367431640625,
0.420654296875,
0.35986328125
],
[
0.5087890625,
0.3447265625,
0.55126953125,
0.3447265625,
0.55126953125,
0.35595703125,
0.5087890625,
0.35595703125
],
[
0.326904296875,
0.3515625,
0.55615234375,
0.371826171875,
0.5537109375,
0.392578125,
0.324462890625,
0.37255859375
],
[
0.44287109375,
0.354248046875,
0.67236328125,
0.354248046875,
0.67236328125,
0.375,
0.44287109375,
0.375
],
[
0.8505859375,
0.35791015625,
0.8837890625,
0.35791015625,
0.8837890625,
0.373779296875,
0.8505859375,
0.373779296875
],
[
0.11767578125,
0.391357421875,
0.685546875,
0.391357421875,
0.685546875,
0.40478515625,
0.11767578125,
0.40478515625
],
[
0.11358642578125,
0.410400390625,
0.8837890625,
0.412353515625,
0.8837890625,
0.429931640625,
0.11358642578125,
0.427978515625
],
[
0.11517333984375,
0.4287109375,
0.71142578125,
0.4287109375,
0.71142578125,
0.443115234375,
0.11517333984375,
0.443115234375
],
[
0.35302734375,
0.45703125,
0.64208984375,
0.46533203125,
0.640625,
0.49755859375,
0.351318359375,
0.4892578125
],
[
0.59326171875,
0.455810546875,
0.63818359375,
0.460205078125,
0.6357421875,
0.477783203125,
0.5908203125,
0.473388671875
],
[
0.85107421875,
0.47021484375,
0.8837890625,
0.47021484375,
0.8837890625,
0.486083984375,
0.85107421875,
0.486083984375
],
[
0.5048828125,
0.48291015625,
0.51123046875,
0.488037109375,
0.5048828125,
0.49365234375,
0.498291015625,
0.488037109375
],
[
0.1422119140625,
0.501953125,
0.88232421875,
0.501953125,
0.88232421875,
0.5166015625,
0.1422119140625,
0.5166015625
],
[
0.1192626953125,
0.5166015625,
0.88134765625,
0.5166015625,
0.88134765625,
0.52978515625,
0.1192626953125,
0.52978515625
],
[
0.11767578125,
0.5322265625,
0.4697265625,
0.5322265625,
0.4697265625,
0.54541015625,
0.11767578125,
0.54541015625
],
[
0.11517333984375,
0.5693359375,
0.390625,
0.5693359375,
0.390625,
0.58349609375,
0.11517333984375,
0.58349609375
],
[
0.1168212890625,
0.599609375,
0.8798828125,
0.599609375,
0.8798828125,
0.61279296875,
0.1168212890625,
0.61279296875
],
[
0.1192626953125,
0.6171875,
0.27294921875,
0.6171875,
0.27294921875,
0.62744140625,
0.1192626953125,
0.62744140625
],
[
0.35693359375,
0.64599609375,
0.471435546875,
0.65771484375,
0.46728515625,
0.6796875,
0.35302734375,
0.66748046875
],
[
0.488525390625,
0.6484375,
0.53515625,
0.65380859375,
0.53125,
0.6748046875,
0.484375,
0.66943359375
],
[
0.5361328125,
0.65087890625,
0.6298828125,
0.65087890625,
0.6298828125,
0.66845703125,
0.5361328125,
0.66845703125
],
[
0.66650390625,
0.64599609375,
0.71142578125,
0.64794921875,
0.7099609375,
0.66748046875,
0.6650390625,
0.66552734375
],
[
0.28759765625,
0.6611328125,
0.3544921875,
0.6611328125,
0.3544921875,
0.67236328125,
0.28759765625,
0.67236328125
],
[
0.85107421875,
0.6591796875,
0.8837890625,
0.6591796875,
0.8837890625,
0.6748046875,
0.85107421875,
0.6748046875
],
[
0.4794921875,
0.66552734375,
0.51123046875,
0.66552734375,
0.51123046875,
0.68115234375,
0.4794921875,
0.68115234375
],
[
0.646484375,
0.66650390625,
0.6640625,
0.66650390625,
0.6640625,
0.673828125,
0.646484375,
0.673828125
],
[
0.6689453125,
0.66552734375,
0.708984375,
0.66552734375,
0.708984375,
0.68115234375,
0.6689453125,
0.68115234375
],
[
0.58837890625,
0.66845703125,
0.62744140625,
0.67041015625,
0.6259765625,
0.681640625,
0.58642578125,
0.6796875
],
[
0.55126953125,
0.66943359375,
0.56640625,
0.66943359375,
0.56640625,
0.6787109375,
0.55126953125,
0.6787109375
],
[
0.580078125,
0.66845703125,
0.58642578125,
0.66845703125,
0.58642578125,
0.673828125,
0.580078125,
0.673828125
],
[
0.66650390625,
0.66943359375,
0.6748046875,
0.66943359375,
0.6748046875,
0.6787109375,
0.66650390625,
0.6787109375
],
[
0.476318359375,
0.67041015625,
0.483642578125,
0.67041015625,
0.483642578125,
0.6767578125,
0.476318359375,
0.6767578125
],
[
0.371826171875,
0.67236328125,
0.3857421875,
0.67236328125,
0.3857421875,
0.68310546875,
0.371826171875,
0.68310546875
],
[
0.57861328125,
0.6767578125,
0.5986328125,
0.6767578125,
0.5986328125,
0.6845703125,
0.57861328125,
0.6845703125
],
[
0.1192626953125,
0.6943359375,
0.88134765625,
0.6943359375,
0.88134765625,
0.70849609375,
0.1192626953125,
0.70849609375
],
[
0.1168212890625,
0.71044921875,
0.88134765625,
0.71044921875,
0.88134765625,
0.7236328125,
0.1168212890625,
0.7236328125
],
[
0.1168212890625,
0.72607421875,
0.158447265625,
0.72607421875,
0.158447265625,
0.7373046875,
0.1168212890625,
0.7373046875
],
[
0.5087890625,
0.74755859375,
0.59619140625,
0.76123046875,
0.59228515625,
0.77880859375,
0.5048828125,
0.76513671875
],
[
0.40185546875,
0.76123046875,
0.5048828125,
0.7568359375,
0.50634765625,
0.7744140625,
0.40283203125,
0.77880859375
],
[
0.406982421875,
0.77294921875,
0.428955078125,
0.77294921875,
0.428955078125,
0.78076171875,
0.406982421875,
0.78076171875
],
[
0.525390625,
0.7734375,
0.53369140625,
0.7734375,
0.53369140625,
0.78271484375,
0.525390625,
0.78271484375
],
[
0.85107421875,
0.76806640625,
0.8837890625,
0.76806640625,
0.8837890625,
0.78515625,
0.85107421875,
0.78515625
],
[
0.408447265625,
0.787109375,
0.59228515625,
0.787109375,
0.59228515625,
0.80029296875,
0.408447265625,
0.80029296875
],
[
0.1422119140625,
0.8095703125,
0.8798828125,
0.8095703125,
0.8798828125,
0.82275390625,
0.1422119140625,
0.82275390625
],
[
0.1192626953125,
0.82568359375,
0.71826171875,
0.82568359375,
0.71826171875,
0.8359375,
0.1192626953125,
0.8359375
],
[
0.468017578125,
0.849609375,
0.484375,
0.849609375,
0.484375,
0.859375,
0.468017578125,
0.859375
],
[
0.26708984375,
0.8623046875,
0.7294921875,
0.8623046875,
0.7294921875,
0.876953125,
0.26708984375,
0.876953125
],
[
0.85107421875,
0.85986328125,
0.8837890625,
0.86279296875,
0.88134765625,
0.87890625,
0.8486328125,
0.87548828125
],
[
0.465576171875,
0.87744140625,
0.488525390625,
0.87744140625,
0.488525390625,
0.89013671875,
0.465576171875,
0.89013671875
],
[
0.11767578125,
0.89794921875,
0.51123046875,
0.89794921875,
0.51123046875,
0.912109375,
0.11767578125,
0.912109375
],
[
0.491943359375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"Theorem 3. Assume f(wt,&t,yt)> 0,Vt = 1,.,T.For ^ that is defined in Equation (10), when",
"O <p <1 , we have At > 1 and At is monotonically decreasing with respect to p, Vt = 1,..., T.",
"基于Equation（10），如果f(wt,at,yt）>0,Vt=1,,T，则必须存在固定的rx>0，例如",
"入tE（1,r>),Vt=1，,T。因此，我们可以分析基于Theorem1的Pareto-PathMTL的泛化界，当",
"0<p<1。尽管Theorem1不适合p=1时的情况，但我们可以通过让p无限接近1来近似其边界",
"以上两个定理表明，当0<p<1。因此，泛化界中的第二项随着p的减小而减小。基于Theorem3的",
"绑定中的第三项也是如此。因此，毫不奇怪，当0<p<1时，泛化性能通常比p=1时更好，并且当p减",
"小时，期望性能得到改善是合理的。实际上，在[16]的实验中报告了这种单调性：当p减小时，分类精度通",
"常是单调增加的。值得一提的是，尽管很少观察到，但如果泛化边界中的第一项i.e.（经验损失）随着p的",
"减小而迅速增长，则我们可能不会在性能上具有这种单调性。但是，泛化界的单调行为（经验损失除外）仍",
"然足以解释Problem（9）的实验结果，这证明了采用目标函数的任意加权圆锥组合而不是使用这些函数的平",
"均值的基本原理",
"最后，我们提供了两个定理，它们不仅用于Theorem1的证明，而且可能会自行引起人们的兴趣。随后，",
"在下一节中，我们将描述我们的新MTL模型",
"Theorem4.Given≤[1,...,T]'with0,define",
"2",
"TN",
"R(Fx,)=TNE[we台台",
"E[ sup toi(wt, d(ci)]",
"(19)",
"For fired 0, R(F,) is monotonically increasing with respect to each t :",
"Theorem 5. For fired r> ≥1 , p>0, >= [Ai,.. , 入r]', ^t E[1,ral, Vt = 1, .. ,T, and for any w e Fx ,",
"O < < 1 , the following generalization bound holds with probability at least 1-d :",
"er(u)≤(wu)+R(FA)+V",
"9ln",
"(20)",
"0",
"请注意，Theorem5和Theorem1之间的区别在于，Theorem1对任何入tE（1,r>）都有效，而Theorem5",
"仅对固定的入tE[1,r>]有效。虽然Theorem1中给出的界限更为笼统，但由于（17）中的附加第三项以及",
"V2乘以经验Rademacher复杂度而使其更加宽松",
"3 A New MTL Model",
"在本节中，我们提出了新的MTL模型。受Theorem1中泛化约束的启发，我们的模型被公式化为通过最小",
"化约束来选择W和入",
"2r> R(F>) +",
"9lm",
"(2r>1)",
"9n",
"er(w)+",
"(21)",
"TN",
"",
"2TN",
"一入t",
"T",
"/",
"",
"",
"p",
"t=1",
"而不是启发式地选择系数入，例如通过[16]中的Equation（10）。请注意，边界的最后一项不依赖于任何模",
"型参数，而第三项对边界只有很小的影响，当入E（1,r>）。因此，我们省略了这两个术语，并提出了以下",
"模型：",
"V2r>R(F>)",
"min erx(w) +",
"w,x",
"Q",
"(22)",
"s.t.we Fx,1><rl.",
"此外，由于R(F>）的复杂性，很难直接优化Problem（22）。因此，在下面的定理中，我们证明了R(F>）",
"的一个上限，它产生了一个更简单的表达式。我们提醒读者，假设空间被定义为",
"T",
"Fx ≤ [w= (wi,*..,wT) :入twtll?≤R,wt EHe,0 E2(0))",
"(23)",
"t=1",
"其中He是由内核函数km-,0mkm定义的RKHS",
"5"
] |
{
"arxiv_id": "1408.4714",
"language": "zh",
"page_idx": 4
}
|
||
2402.04035-zh-0
|
2402.04035-zh-0.jpg
|
# Low-Distortion Clustering with Ordinal and Limited $\operatorname{Cardinal}$ Information
$\begin{array}{ccc}\text{Jakob Burkhardt}^*&\text{Ioannis Caragiannis}^*&\text{Karl Fehrs}^*\\\text{Chris Schwiegelshohn}^*&\text{Sudarshan Shvam}^*\end{array}$Matteo Russo$^\dagger$
Sudarshan Shyam$^*$
$\mathbf{Abstract}$
Motivated by recent work in computational social choice, we extend the metric distortion framework to clustering problems. Given a set of $n$ agents located in an underlying metric space, our goal is to partition them into $k$ clusters, optimizing some social cost objective. The metric space is defined by a distance function $d$ between the agent locations. Information about $d$ is available only implicitly via $n$ rankings, through which each agent ranks all other agents in terms of their distance from her. Still, even though no cardinal information (i.e., the exact distance values) is available, we would like to evaluate clustering algorithms in terms of social cost objectives that are defined using $d.$ This is done using the notion of distortion, which measures how far from optimality a clustering can be, taking into account all underlying metrics that are consistent with the ordinal information available. Unfortunately, the most important clustering objectives (e.g., those used in the well-known $k$ -median and $k$ -center problems) do not admit algorithms with finite distortion. To sidestep this disappointing fact, we follow two alternative approaches: We first explore whether resource augmentation can be beneficial. We consider algorithms that use more than $k$ clusters but compare their social cost to that of the optimal $k$ -clusterings. We show that using exponentially (in terms of $k$ )many clusters, we can get low (constant or logarithmic) distortion for the $k$ -center and $k$ -median objectives. Interestingly, such an exponential blowup is shown to be necessary. More importantly, we explore whether limited cardinal information can be used to obtain better results. Somewhat surprisingly, for $k$ -median and $k$ -center, we show that a number of queries that is polynomial in $k$ and only logarithmic in $n$ ( i. e. , only sublinear in the number of agents for the most relevant scenarios in practice) is enough to get constant distortion.
## 1 $\textbf{Introduction}$
典型的计算社会选择问题包括优化替代方案的函数，每个方案都有不同的相关成本或价值。一个典型的例子是代表选举。每个选民对每个候选人都有不同的代表分数，我们假设这与某个基本指标中的距离相对应。理想情况下，代表将每个选民与其最接近的代表的距离之和最小化。在完全信息设置中，这对应于解决经典的$k$ -中值问题。但是这个例子已经说明了实施任何投票机制的困难：即使假设代表分数是距离，即使参与的选民也可能不知道。但是，我们可能很容易知道选民是否更喜欢$\mathbb{E}a$ 而不是$\mathbb{E}b$
这样的例子产生了序数算法。序数算法主要允许在基础度量中的距离之间进行比较。也就是说给定三个点$a,b,c$ ,我们可以自由地给出$d(a,b)\leq d(a,c)$的信息，但是我们没有给出$d(a,b)$和$d(a,c)$的确切数值。目标是解决一个给定的问题，主要依赖于序数信息，同时使用尽可能少的(理
$^{*}$Department of Computer Science, Aarhus University, Åbogade 34,8200 Aarhus N, Denmark. Email: $\{$ jakob,
iannis, karl, schwiegelshohn, shyam $\} @$cs.au.dk.
$^{\dagger}$Department of Computer, Control, Management Engineering, Sapienza University of Rome, Via Ariosto 25,
00185, Rome, Italy. Email: mrusso@diag.uniromal.it.
1
|
[
[
0.134033203125,
0.1427001953125,
0.865234375,
0.1427001953125,
0.865234375,
0.1634521484375,
0.134033203125,
0.1634521484375
],
[
0.35302734375,
0.17041015625,
0.64794921875,
0.1717529296875,
0.64794921875,
0.189453125,
0.35302734375,
0.1881103515625
],
[
0.1707763671875,
0.21337890625,
0.330078125,
0.21337890625,
0.330078125,
0.2279052734375,
0.1707763671875,
0.2279052734375
],
[
0.358642578125,
0.21337890625,
0.5400390625,
0.21337890625,
0.5400390625,
0.2279052734375,
0.358642578125,
0.2279052734375
],
[
0.56884765625,
0.21337890625,
0.705078125,
0.21337890625,
0.705078125,
0.2279052734375,
0.56884765625,
0.2279052734375
],
[
0.6953125,
0.21337890625,
0.830078125,
0.21337890625,
0.830078125,
0.2279052734375,
0.6953125,
0.2279052734375
],
[
0.301513671875,
0.231689453125,
0.48779296875,
0.231689453125,
0.48779296875,
0.246826171875,
0.301513671875,
0.246826171875
],
[
0.55810546875,
0.2171630859375,
0.564453125,
0.2171630859375,
0.564453125,
0.222900390625,
0.55810546875,
0.222900390625
],
[
0.54736328125,
0.231689453125,
0.70361328125,
0.231689453125,
0.70361328125,
0.246826171875,
0.54736328125,
0.246826171875
],
[
0.464111328125,
0.29541015625,
0.53759765625,
0.29541015625,
0.53759765625,
0.31005859375,
0.464111328125,
0.31005859375
],
[
0.1846923828125,
0.31640625,
0.83837890625,
0.317626953125,
0.83837890625,
0.33203125,
0.1846923828125,
0.330810546875
],
[
0.160888671875,
0.333251953125,
0.83837890625,
0.333251953125,
0.83837890625,
0.347900390625,
0.160888671875,
0.347900390625
],
[
0.16015625,
0.348388671875,
0.83837890625,
0.348388671875,
0.83837890625,
0.3623046875,
0.16015625,
0.3623046875
],
[
0.160888671875,
0.363037109375,
0.83837890625,
0.363037109375,
0.83837890625,
0.376953125,
0.160888671875,
0.376953125
],
[
0.160888671875,
0.37890625,
0.83837890625,
0.37890625,
0.83837890625,
0.392578125,
0.160888671875,
0.392578125
],
[
0.158447265625,
0.393310546875,
0.83837890625,
0.393310546875,
0.83837890625,
0.4072265625,
0.158447265625,
0.4072265625
],
[
0.160888671875,
0.4091796875,
0.83837890625,
0.4091796875,
0.83837890625,
0.42236328125,
0.160888671875,
0.42236328125
],
[
0.158447265625,
0.423583984375,
0.83837890625,
0.42236328125,
0.83837890625,
0.436279296875,
0.158447265625,
0.4375
],
[
0.158447265625,
0.439453125,
0.83837890625,
0.439453125,
0.83837890625,
0.45263671875,
0.158447265625,
0.45263671875
],
[
0.160888671875,
0.453857421875,
0.83837890625,
0.453857421875,
0.83837890625,
0.46728515625,
0.160888671875,
0.46728515625
],
[
0.160888671875,
0.47021484375,
0.8349609375,
0.47021484375,
0.8349609375,
0.48095703125,
0.160888671875,
0.48095703125
],
[
0.160888671875,
0.484130859375,
0.83642578125,
0.484130859375,
0.83642578125,
0.49755859375,
0.160888671875,
0.49755859375
],
[
0.16015625,
0.5,
0.83837890625,
0.5,
0.83837890625,
0.51318359375,
0.16015625,
0.51318359375
],
[
0.16015625,
0.5146484375,
0.8388671875,
0.5146484375,
0.8388671875,
0.52783203125,
0.16015625,
0.52783203125
],
[
0.16015625,
0.52978515625,
0.83837890625,
0.52978515625,
0.83837890625,
0.54345703125,
0.16015625,
0.54345703125
],
[
0.16015625,
0.54541015625,
0.83837890625,
0.54541015625,
0.83837890625,
0.5595703125,
0.16015625,
0.5595703125
],
[
0.16015625,
0.56005859375,
0.83837890625,
0.56005859375,
0.83837890625,
0.57373046875,
0.16015625,
0.57373046875
],
[
0.160888671875,
0.57568359375,
0.83837890625,
0.57568359375,
0.83837890625,
0.5888671875,
0.160888671875,
0.5888671875
],
[
0.160888671875,
0.59033203125,
0.83837890625,
0.59033203125,
0.83837890625,
0.60400390625,
0.160888671875,
0.60400390625
],
[
0.16015625,
0.60498046875,
0.8388671875,
0.60498046875,
0.8388671875,
0.619140625,
0.16015625,
0.619140625
],
[
0.158447265625,
0.62060546875,
0.49267578125,
0.62060546875,
0.49267578125,
0.6337890625,
0.158447265625,
0.6337890625
],
[
0.11358642578125,
0.66015625,
0.160888671875,
0.66015625,
0.160888671875,
0.67529296875,
0.11358642578125,
0.67529296875
],
[
0.15283203125,
0.66015625,
0.310546875,
0.66015625,
0.310546875,
0.6748046875,
0.15283203125,
0.6748046875
],
[
0.1356201171875,
0.66748046875,
0.1478271484375,
0.66748046875,
0.1478271484375,
0.671875,
0.1356201171875,
0.671875
],
[
0.11767578125,
0.6923828125,
0.88134765625,
0.6923828125,
0.88134765625,
0.70703125,
0.11767578125,
0.70703125
],
[
0.1168212890625,
0.708984375,
0.8798828125,
0.71044921875,
0.8798828125,
0.724609375,
0.1168212890625,
0.7236328125
],
[
0.1192626953125,
0.72802734375,
0.88134765625,
0.72802734375,
0.88134765625,
0.7412109375,
0.1192626953125,
0.7412109375
],
[
0.11767578125,
0.744140625,
0.88134765625,
0.744140625,
0.88134765625,
0.75830078125,
0.11767578125,
0.75830078125
],
[
0.11767578125,
0.76220703125,
0.88134765625,
0.76220703125,
0.88134765625,
0.77587890625,
0.11767578125,
0.77587890625
],
[
0.1168212890625,
0.77880859375,
0.2484130859375,
0.77880859375,
0.2484130859375,
0.79345703125,
0.1168212890625,
0.79345703125
],
[
0.143798828125,
0.796875,
0.8798828125,
0.796875,
0.8798828125,
0.81005859375,
0.143798828125,
0.81005859375
],
[
0.11767578125,
0.81298828125,
0.8798828125,
0.81298828125,
0.8798828125,
0.82763671875,
0.11767578125,
0.82763671875
],
[
0.11517333984375,
0.830078125,
0.88134765625,
0.830078125,
0.88134765625,
0.8447265625,
0.11517333984375,
0.8447265625
],
[
0.1356201171875,
0.85693359375,
0.88134765625,
0.85693359375,
0.88134765625,
0.87060546875,
0.1356201171875,
0.87060546875
],
[
0.11767578125,
0.87255859375,
0.435546875,
0.87255859375,
0.435546875,
0.88330078125,
0.11767578125,
0.88330078125
],
[
0.1356201171875,
0.88330078125,
0.88134765625,
0.8837890625,
0.88134765625,
0.8984375,
0.1356201171875,
0.89794921875
],
[
0.11358642578125,
0.89794921875,
0.4755859375,
0.8994140625,
0.4755859375,
0.91357421875,
0.11358642578125,
0.9111328125
],
[
0.490234375,
0.93505859375,
0.50732421875,
0.93505859375,
0.50732421875,
0.94873046875,
0.490234375,
0.94873046875
]
] |
[
"Low-Distortion Clustering with Ordinal and Limited",
"Cardinal Information",
"Jakob Burkhardt*",
"Ioannis Caragiannis*",
"Karl Fehrs*",
"Matteo Russot",
"Chris Schwiegelshohn*",
"",
"Sudarshan Shyam*",
"Abstract",
"Motivated by recent work in computational social choice, we extend the metric distortion",
"framework to clustering problems. Given a set of n agents located in an underlying metric",
"space, our goal is to partition them into k clusters, optimizing some social cost objective. The",
"metric space is defined by a distance function d between the agent locations. Information about",
"d is available only implicitly via n rankings, through which each agent ranks all other agents",
"in terms of their distance from her. Still, even though no cardinal information (i.e., the exact",
"distance values) is available, we would like to evaluate clustering algorithms in terms of social cost",
"objectives that are defined using d . This is done using the notion of distortion, which measures",
"how far from optimality a clustering can be, taking into account all underlying metrics that are",
"consistent with the ordinal information available. Unfortunately, the most important clustering",
"objectives(e.g.,thoseused in the well-knownk-median and k-center problems)do not admit",
"algorithms with finite distortion. To sidestep this disappointing fact, we follow two alternative",
"approaches: We first explore whether resource augmentation can be beneficial. We consider",
"algorithms that use more than k clusters but compare their social cost to that of the optimal k",
"-clusterings. We show that using exponentially (in terms of k ) many clusters, we can get low",
"(constant or logarithmic) distortion for the k -center and k -median objectives. Interestingly,",
"such an exponential blowup is shown to be necessary. More importantly, we explore whether",
"limited cardinal information can be used to obtain better results. Somewhat surprisingly, for",
"k -median and k -center, we show that a number of queries that is polynomial in k and only",
"logarithmic in n (i.e., only sublinear in the number of agents for the most relevant scenarios in",
"practice) is enough to get constant distortion.",
"1",
"Introduction",
"",
"典型的计算社会选择问题包括优化替代方案的函数，每个方案都有不同的相关成本或价值。一个典",
"型的例子是代表选举。每个选民对每个候选人都有不同的代表分数，我们假设这与某个基本指标中",
"的距离相对应。理想情况下，代表将每个选民与其最接近的代表的距离之和最小化。在完全信息设",
"置中，这对应于解决经典的k-中值问题。但是这个例子已经说明了实施任何投票机制的困难：即",
"使假设代表分数是距离，即使参与的选民也可能不知道。但是，我们可能很容易知道选民是否更喜",
"欢Ea而不是Eb",
"这样的例子产生了序数算法。序数算法主要允许在基础度量中的距离之间进行比较。也就是说",
"给定三个点a,b,c，我们可以自由地给出d(a,b)≤d(a,c)的信息，但是我们没有给出d(a,b)和",
"d(a,）的确切数值。目标是解决一个给定的问题，主要依赖于序数信息，同时使用尽可能少的（理",
"*Department of Computer Science, Aarhus University, Abogade 34, 8200 Aarhus N, Denmark. Email: jakob,",
"iannis,karl,schwiegelshohn,shyam @cs.au.dk.",
"tDepartment of Computer, Control, Management Engineering, Sapienza University of Rome, Via Ariosto 25,",
"00185,Rome,Italy.Email:mrusso@diag.uniromal.it.",
"1"
] |
{
"arxiv_id": "2402.04035",
"language": "zh",
"page_idx": 0
}
|
||
2402.04035-zh-1
|
2402.04035-zh-1.jpg
|
想情况下为零) 距离查询。与给定完整信息的最优解 OPT 的质量相比，这种算法的优度是根据计
算出的解$C$的质量来衡量的，通常称为度量失真
找到中位数可以说是该领域最重要的问题。给定一组点$X$和距离函数$d$ ,将中值$m$定义为使距离之和最小的点。经过长时间的工作[10,11,34,39,46,51],现在存在一种具有最佳度量失真的确定性算法 3 [38] , 它也是最佳的[8,9] 。使用随机化，Charikar et al. [23] 最近取得了重要突破，实现了2.753的度量失真。最著名的下界至少是 2.1126 [22]
扩展到更一般的聚类目标，如$(k,z)$ -聚类和设施位置相对更难，请参见 Anshelevich and Zhu
[6], Caragiannis et al.[20] 。在设施位置，我们要求一组中心$C$ ,使得
$$\begin{aligned}\sum_{x\in X}\min_{c\in C}d(x,c)+f\cdot|C|\end{aligned}$$
最小化，其中$f$是开设中心的成本。对于$(k,z)$聚类，我们改为考虑目标
$$\sqrt[z]{\sum\limits_{x\in X}\min\limits_{c\in C}d(x,c)^z},$$
i。e。,该算法不会产生打开中心的成本，而是最多可以放置$k$中心的预算。特殊情况包括$k$ -中位
数，其中$z=1$和$k$ -中心对应于$z\to\infty$ 。$^1$
不幸的是，对于纯序数算法，存在强烈的不可能性结果。即使对于 2 -中位数，也不可能获得具有有限度量失真的算法[6] 。因此，研究已经开始设计比纯粹的序数信息更强大的算法。实际上，最近在提供保证方面取得了一些成功，每点仅使用恒定数量的查询，请参见 Amanatidis et al. [4,3] 。对于聚类，Pulyassary [54]最近的工作表明，每点最多使用 polylog$(n)$距离查询，或者总体上使用$n\cdot$polylog$(n)$查询，可以实现常数因子近似。相同的工作还表明，每个点的$k$查询或总体上的$O(nk)$查询足以实现$k$的恒定因子近似-中位数。因此，我们问：
Question $1. 1. \textit{ What is the minimum number of queries necessary for an algorithm to achieve}$
$constant~metric~distortion~for~k~-median,~k~-center,~and~facility~location?$
虽然距离查询是向算法设计人员提供更多权力的自然方式，但如上所述，获取距离可能很昂贵。这导致了一个问题，即是否存在其他模型，允许算法设计者限制度量失真。聚类算法的一种非常自然的方法是允许算法返回$(\alpha,\beta)$ -bicriteria 近似。这样的算法将聚类成本限制在最佳$k$聚类成本的$\alpha$倍，同时使用$\beta$多个中心。我们问：
Question 1.2. What is the minimum value of β such that a bicriteria clustering algorithm using
$only\textit{ ordinal information has constant metric distortion? }$
# 1.1 Our Results
在本文中，我们在回答这两个问题方面取得了实质性进展。在低查询设置中，我们为$k$ -center 提供了两种确定性多项式时间算法，该算法最多使用$O(k^2)$总体距离评估，获得 2 失真，最多使用$O(k)$总体距离评估，获得 4 失真。我们还表明，就必要的查询数量而言，后一种结果是最佳的，而对于任何多项式时间算法，前者都是最佳的。对于$(k,z)$聚类，我们获得了一个随机多项式时间算法，该算法最多使用 poly$(k,\log n)$总体距离查询，并实现了恒定的度量失真。请注意，所有这些边界在输入大小上都是次线性的，即假设$k\ll n$ ,我们每点进行$o(1)$查询
最后，对于设施位置，存在对开创性的 Meyerson 算法 [50] 的简单调整，该算法使用每个点一个查询或整个$n$查询来实现恒定失真，另请参见第 4 节。1个 Pulyassary [54] 。我们证明，没有任何算法可以使用少于$\Omega(n)$的查询来实现常数因子近似，从而有效地解决了问题
$^{1}$Sometimes the$\underbrace{z/\text{ operation is omitted, as is the case for }k\text{ -means corresponds to }(k,2)\text{ -clustering. An }\alpha}$
approximation to $\sqrt[{z}]\sum_{x\in X}\operatorname*{min}_{c\in C}d(x,c)^{z}$ implies an $O(\alpha^{z})$ -approximation to $\sum _x\in X\min _{c\in C}d( x, c) ^{z}$ .
2
|
[
[
0.11767578125,
0.0947265625,
0.8798828125,
0.0947265625,
0.8798828125,
0.10797119140625,
0.11767578125,
0.10797119140625
],
[
0.11767578125,
0.111083984375,
0.5,
0.111083984375,
0.5,
0.125,
0.11767578125,
0.125
],
[
0.146240234375,
0.1275634765625,
0.88134765625,
0.1275634765625,
0.88134765625,
0.141357421875,
0.146240234375,
0.141357421875
],
[
0.1168212890625,
0.1458740234375,
0.88134765625,
0.1458740234375,
0.88134765625,
0.1590576171875,
0.1168212890625,
0.1590576171875
],
[
0.11767578125,
0.1622314453125,
0.88134765625,
0.1622314453125,
0.88134765625,
0.175537109375,
0.11767578125,
0.175537109375
],
[
0.1168212890625,
0.177978515625,
0.61279296875,
0.177978515625,
0.61279296875,
0.1912841796875,
0.1168212890625,
0.1912841796875
],
[
0.1422119140625,
0.1956787109375,
0.8798828125,
0.1962890625,
0.8798828125,
0.2108154296875,
0.1422119140625,
0.210205078125
],
[
0.11767578125,
0.2139892578125,
0.6708984375,
0.2139892578125,
0.6708984375,
0.2291259765625,
0.11767578125,
0.2291259765625
],
[
0.41015625,
0.242431640625,
0.59228515625,
0.242431640625,
0.59228515625,
0.260009765625,
0.41015625,
0.260009765625
],
[
0.40283203125,
0.261962890625,
0.463134765625,
0.255126953125,
0.465576171875,
0.26953125,
0.406005859375,
0.276611328125
],
[
0.11767578125,
0.287353515625,
0.69140625,
0.287353515625,
0.69140625,
0.300537109375,
0.11767578125,
0.300537109375
],
[
0.44287109375,
0.317626953125,
0.57275390625,
0.3212890625,
0.57177734375,
0.339111328125,
0.441162109375,
0.335205078125
],
[
0.473876953125,
0.337158203125,
0.50244140625,
0.337158203125,
0.50244140625,
0.34521484375,
0.473876953125,
0.34521484375
],
[
0.438720703125,
0.343505859375,
0.4755859375,
0.337158203125,
0.4794921875,
0.349853515625,
0.44287109375,
0.35595703125
],
[
0.11517333984375,
0.36865234375,
0.8798828125,
0.36865234375,
0.8798828125,
0.3818359375,
0.11517333984375,
0.3818359375
],
[
0.1168212890625,
0.385009765625,
0.451904296875,
0.385009765625,
0.451904296875,
0.3984375,
0.1168212890625,
0.3984375
],
[
0.1446533203125,
0.40283203125,
0.88134765625,
0.40283203125,
0.88134765625,
0.416748046875,
0.1446533203125,
0.416748046875
],
[
0.1168212890625,
0.419189453125,
0.88134765625,
0.419189453125,
0.88134765625,
0.43310546875,
0.1168212890625,
0.43310546875
],
[
0.11767578125,
0.436279296875,
0.8798828125,
0.436279296875,
0.8798828125,
0.45068359375,
0.11767578125,
0.45068359375
],
[
0.11767578125,
0.455078125,
0.8798828125,
0.455078125,
0.8798828125,
0.468505859375,
0.11767578125,
0.468505859375
],
[
0.11767578125,
0.4716796875,
0.8798828125,
0.4716796875,
0.8798828125,
0.486083984375,
0.11767578125,
0.486083984375
],
[
0.11767578125,
0.4892578125,
0.6787109375,
0.4892578125,
0.6787109375,
0.50244140625,
0.11767578125,
0.50244140625
],
[
0.11767578125,
0.51513671875,
0.8798828125,
0.51513671875,
0.8798828125,
0.52978515625,
0.11767578125,
0.52978515625
],
[
0.11517333984375,
0.5322265625,
0.68896484375,
0.53076171875,
0.68896484375,
0.54541015625,
0.11517333984375,
0.546875
],
[
0.1446533203125,
0.5595703125,
0.88134765625,
0.5595703125,
0.88134765625,
0.57275390625,
0.1446533203125,
0.57275390625
],
[
0.11767578125,
0.57568359375,
0.8798828125,
0.57568359375,
0.8798828125,
0.5888671875,
0.11767578125,
0.5888671875
],
[
0.11767578125,
0.59326171875,
0.88134765625,
0.59326171875,
0.88134765625,
0.60693359375,
0.11767578125,
0.60693359375
],
[
0.11517333984375,
0.60986328125,
0.414306640625,
0.60986328125,
0.414306640625,
0.62353515625,
0.11517333984375,
0.62353515625
],
[
0.11517333984375,
0.6337890625,
0.88232421875,
0.63525390625,
0.88232421875,
0.65283203125,
0.11517333984375,
0.6513671875
],
[
0.1168212890625,
0.65380859375,
0.5595703125,
0.65380859375,
0.5595703125,
0.66748046875,
0.1168212890625,
0.66748046875
],
[
0.1168212890625,
0.689453125,
0.17724609375,
0.689453125,
0.17724609375,
0.7041015625,
0.1168212890625,
0.7041015625
],
[
0.17236328125,
0.689453125,
0.2900390625,
0.689453125,
0.2900390625,
0.70263671875,
0.17236328125,
0.70263671875
],
[
0.11767578125,
0.71630859375,
0.88134765625,
0.71630859375,
0.88134765625,
0.72998046875,
0.11767578125,
0.72998046875
],
[
0.11767578125,
0.734375,
0.8798828125,
0.734375,
0.8798828125,
0.74755859375,
0.11767578125,
0.74755859375
],
[
0.11767578125,
0.75048828125,
0.88134765625,
0.75048828125,
0.88134765625,
0.76513671875,
0.11767578125,
0.76513671875
],
[
0.11767578125,
0.76806640625,
0.88134765625,
0.76806640625,
0.88134765625,
0.7822265625,
0.11767578125,
0.7822265625
],
[
0.1168212890625,
0.78515625,
0.88134765625,
0.78515625,
0.88134765625,
0.798828125,
0.1168212890625,
0.798828125
],
[
0.11767578125,
0.8037109375,
0.69921875,
0.8037109375,
0.69921875,
0.814453125,
0.11767578125,
0.814453125
],
[
0.1446533203125,
0.818359375,
0.88232421875,
0.818359375,
0.88232421875,
0.83349609375,
0.1446533203125,
0.83349609375
],
[
0.1192626953125,
0.8359375,
0.88134765625,
0.8359375,
0.88134765625,
0.849609375,
0.1192626953125,
0.849609375
],
[
0.11767578125,
0.85400390625,
0.7548828125,
0.85400390625,
0.7548828125,
0.8671875,
0.11767578125,
0.8671875
],
[
0.136474609375,
0.876953125,
0.8837890625,
0.87744140625,
0.8837890625,
0.89599609375,
0.136474609375,
0.89453125
],
[
0.1168212890625,
0.8994140625,
0.2467041015625,
0.8994140625,
0.2467041015625,
0.91357421875,
0.1168212890625,
0.91357421875
],
[
0.236083984375,
0.89794921875,
0.7998046875,
0.896484375,
0.7998046875,
0.9140625,
0.236083984375,
0.91552734375
],
[
0.491943359375,
0.93603515625,
0.50732421875,
0.93603515625,
0.50732421875,
0.95068359375,
0.491943359375,
0.95068359375
]
] |
[
"想情况下为零）距离查询。与给定完整信息的最优解OPT的质量相比，这种算法的优度是根据计",
"算出的解C的质量来衡量的，通常称为度量失真",
"找到中位数可以说是该领域最重要的问题。给定一组点X和距离函数d，将中值㎡定义为使",
"距离之和最小的点。经过长时间的工作10，11，34，39，46，51]，现在存在一种具有最佳度量失真",
"的确定性算法3[38】，它也是最佳的8，9]。使用随机化，Charikaretal.[23]最近取得了重要突",
"破，实现了2.753的度量失真。最著名的下界至少是2.1126[22",
"扩展到更一般的聚类目标，如（k，z）-聚类和设施位置相对更难，请参见AnshelevichandZhu",
"[6]Caragiannisetal.[20]。在设施位置，我们要求一组中心C，使得",
"mind(α,c) +f ·JCl",
"ZEXCEC",
"最小化，其中f是开设中心的成本。对于（k，之）聚类，我们改为考虑目标",
"min d(a,c)\",",
"CEO",
"ZEx",
"i。e。，该算法不会产生打开中心的成本，而是最多可以放置k中心的预算。特殊情况包括k-中位",
"数，其中之=1和k-中心对应于z→8。",
"不幸的是，对于纯序数算法，存在强烈的不可能性结果。即使对于2-中位数，也不可能获得",
"具有有限度量失真的算法[6]。因此，研究已经开始设计比纯粹的序数信息更强大的算法。实际",
"上，最近在提供保证方面取得了一些成功，每点仅使用恒定数量的查询，请参见Amanatidisetal.",
"[4，3]。对于聚类，Pulyassary54]最近的工作表明，每点最多使用polylog(n）距离查询，或者总",
"体上使用n·polylog(n）查询，可以实现常数因子近似。相同的工作还表明，每个点的k查询或总",
"体上的O（nk）查询足以实现k的恒定因子近似-中位数。因此，我们问：",
"Question 1.l. What is the minimum number of queries necessary for an algorithm to achieve",
"constant metric distortion for k -median, k -center, and facility location?",
"虽然距离查询是向算法设计人员提供更多权力的自然方式，但如上所述，获取距离可能很昂贵。",
"这导致了一个问题，即是否存在其他模型，允许算法设计者限制度量失真。聚类算法的一种非常自",
"然的方法是允许算法返回（α，β）-bicriteria近似。这样的算法将聚类成本限制在最佳k聚类成本的",
"Q倍，同时使用β多个中心。我们问：",
"Question 1.2. What is the minimum value of β such that a bicriteria clustering algorithm using",
"only ordinal information has constant metric distortion?",
"1.1",
"Our Results",
"在本文中，我们在回答这两个问题方面取得了实质性进展。在低查询设置中，我们为k-center提",
"供了两种确定性多项式时间算法，该算法最多使用O(k-）总体距离评估，获得2失真，最多使用",
"O（k）总体距离评估，获得4失真。我们还表明，就必要的查询数量而言，后一种结果是最佳的，而",
"对于任何多项式时间算法，前者都是最佳的。对于（k，2）聚类，我们获得了一个随机多项式时间算",
"法，该算法最多使用poly(k,logn）总体距离查询，并实现了恒定的度量失真。请注意，所有这些",
"边界在输入天小上都是次线性的，即假设k<n，我们每点进行o(1）查询",
"最后，对于设施位置，存在对开创性的Meyerson算法[50]的简单调整，该算法使用每个点一",
"个查询或整个n查询来实现恒定失真，另请参见第4节。1个Pulyassary54】。我们证明，没有",
"任何算法可以使用少于2(n）的查询来实现常数因子近似，从而有效地解决了问题",
"1Sometimes the / operation is omitted, as is the case for k -means corresponds to (k,2) -clustering. An α",
"-approximation to",
"/Zaex mineec d(a,c)\" implies an O(α\") -approximation to Zmex mineec d(a,c)\" .",
"2"
] |
{
"arxiv_id": "2402.04035",
"language": "zh",
"page_idx": 1
}
|
||
2402.04035-zh-2
|
2402.04035-zh-2.jpg
|
在零查询设置中，我们首先显示存在用于$k$ -center 的$(2,2^k-1)$ -bicriteria 算法。此外，该算法是最佳的，因为任何实现有限失真的算法都必须使用$\Omega(2^k)$中心。对于$(k,z)$聚类，我们获得了两种解决所有$(k,z)$聚类目标的算法。第一个以恒定概率成功，并以$(O(\log n)^k-1+o(1))$多个中心实现恒定失真。第二个需要$( O( \log n) ^{k+ o( 1) })$ , 并以期望和高概率实现$O(1)$失真。我们通过证明，对于$k$ -中位数的任何恒定因子失真，即使成功的可能性恒定，$\Omega((2^{\log^*n})^{k-1}+2^k\log n)$中心也是必要的，从而补充了这一结果。。对于 2 -中位数的特殊情况，我们的边界是最优的
# 1.2 Related Work
Ordinal Preferences and Distortion 第一篇考虑使用序数信息的优化问题的论文可能是Procaccia and Rosenschein [53] 。随后，建立了两个主要方向。继续使用 Procaccia and Rosenschein 引人的模型，一行主要关注在归一化假设下最大化福利，但不假设任何度量属性，请参阅 Amanatidis et al. [2,4,3], Caragiannis and Procaccia [19], Filos-Ratsikas et al. [35]。另一项工作是在没有归一化假设的情况下研究问题，但假设偏好是度量的，即。e。,它们满足三角不等式。除了引言中介绍的聚类论文之外，还研究了其他几个失真问题[17,24,25,52] 。虽然很少见，但也可以在不进行归一化或度量假设的情况下获得一些结果，请参见 Abramowitz and Anshelevich [1]
Clustering and Facility Location $( k, z)$ -clustering 在一般度量[29]中是 APX-hard 的， 尽管在对度量[37,30]或输入[5,14,27]进行假设时可以获得非常准确的算法。对于$k$ -center Gonzalez [40] 给出了一个最优的 2 -近似算法。对于$k$ -中位数，$k$ -均值和设施位置，经过大量的研究$\left[43,44,13,48,32,33\right]$,当前的技术水平是 $k$ -中位数 [41] 的 2.613 近似值，$k$ 的 9 近似值 2 的XMATHX\_。对于一般的$(k,z)$聚类，尽管$k$ -中位数和$k$ -均值的大多数证明在丢失$\exp(z)$近似因子的同时通过，但几乎没有声称的界限。可以在 Cohen-Addad et al.[28,31] 中找到明确的结果
# 2 Preliminaries
令$(X,d)$为度量空间，其中$X$为一组$n$点，$d:X\times X\to\mathbb{R}_\geq0$为度量。任何两个点$x,y\in X$之间的距离可以通过$d(x,y)$形式的query访问。我们假设这样的查询与成本相关联。为算法提供预算，并且该算法进行的每个查询都会消耗其预算的一个单位。虽然查询两点之间的确切距离是昂贵的，但我们的模型假设，对于每个点，有关其到其他点的相对距离的ordinal信息都是免费的。更具体地说，每个点$x\in X$提供了一个$ranking\pi_x:[n]\to X$ ,它是consistent和$d$ ,从这个意义上说，每个$i,j\in[n],i<j$都有$d(x,\pi_x(i))\leq d(x,\pi_x(j))$ 。也就是说，更接近$x$的点在$x$的排名中出现higher 。序数preference profle $P$只是积分排名的集合，即。e。,$P=\{\pi_x\}_{x\in X}$。我们为一组配置文件编写$\mathcal{P}(d)$,其中每个点的排名与距离$d$一致
将点的排名限制为$X$的某个子集通常很方便。设$S\subseteq X$和$m=|S|$ 。restriction of $\pi _x\textit{ to S}$ 是一个函数$\pi_x,S:[m]\to S$ ,因此，对于任何两个$y,y^\prime\in S$ ,当且仅当$y$在$\pi_x$中的排名高于$y^\prime$ 时，$y$在$\pi_x,S$中的排名高于$y^\prime$
序数首选项配置文件提供了基本距离度量$d$的非常粗略的草图。但是，配置文件表示的相对距离可以使算法以非常经济的方式分配其预算。考虑以下操作：对于一组点$S\subseteq X$和一个点$x\in X$ ,我们将$distance\textit{of }x\textit{ to S 定 义 为 }d( x, S) = \operatorname* { min} _{y\in S}d( x, y) .$
给定序数信息，可以很容易地将点$z=\arg\min_{y\in S}d(x,y)$识别为$x$在$S$中排名最高的点。因此一种算法可以通过单个查询$d(x,z)$确定$x$到$S\text{ \v{\text{的距离。显然，对于找到 }z=\arg\max}}_y\in Sd(x,y)$ 和距离$d(x,z)$可以进行相同的观察
如果我们将算法$\mathcal{A}$限制为序数信息和固定的查询预算，我们打算研究结果最优性的损失。我们考虑了各种聚类问题，其目标是找到使给定成本函数$\phi$最小化的解决方案。我们用$\mathcal{M}$表示所有度量空间的集合。对于度量空间$(X,d)\in\mathcal{M}$和轮廓$P\in\mathcal{P}(d)$ ,令$\mathcal{A}(P,d)$为由算法$\mathcal{A}$计算的解
3
|
[
[
0.146240234375,
0.09344482421875,
0.88134765625,
0.09344482421875,
0.88134765625,
0.10797119140625,
0.146240234375,
0.10797119140625
],
[
0.11767578125,
0.111083984375,
0.8798828125,
0.111083984375,
0.8798828125,
0.125,
0.11767578125,
0.125
],
[
0.11358642578125,
0.1256103515625,
0.8837890625,
0.12249755859375,
0.8837890625,
0.143310546875,
0.11358642578125,
0.146484375
],
[
0.1168212890625,
0.1427001953125,
0.88232421875,
0.1427001953125,
0.88232421875,
0.160400390625,
0.1168212890625,
0.160400390625
],
[
0.11517333984375,
0.1578369140625,
0.88232421875,
0.160400390625,
0.88232421875,
0.177978515625,
0.11517333984375,
0.175537109375
],
[
0.11767578125,
0.1787109375,
0.73876953125,
0.1787109375,
0.73876953125,
0.1925048828125,
0.11767578125,
0.1925048828125
],
[
0.1168212890625,
0.216552734375,
0.3095703125,
0.216552734375,
0.3095703125,
0.2298583984375,
0.1168212890625,
0.2298583984375
],
[
0.125,
0.2412109375,
0.88134765625,
0.242431640625,
0.88134765625,
0.2568359375,
0.125,
0.255615234375
],
[
0.1168212890625,
0.2587890625,
0.8798828125,
0.2587890625,
0.8798828125,
0.2734375,
0.1168212890625,
0.2734375
],
[
0.11767578125,
0.276611328125,
0.88134765625,
0.276611328125,
0.88134765625,
0.290283203125,
0.11767578125,
0.290283203125
],
[
0.11358642578125,
0.292236328125,
0.88134765625,
0.29345703125,
0.88134765625,
0.30810546875,
0.11358642578125,
0.306884765625
],
[
0.1192626953125,
0.311279296875,
0.8798828125,
0.311279296875,
0.8798828125,
0.324462890625,
0.1192626953125,
0.324462890625
],
[
0.11767578125,
0.328857421875,
0.88134765625,
0.328857421875,
0.88134765625,
0.34228515625,
0.11767578125,
0.34228515625
],
[
0.1168212890625,
0.343505859375,
0.7685546875,
0.3466796875,
0.7685546875,
0.361083984375,
0.1168212890625,
0.35791015625
],
[
0.12744140625,
0.380615234375,
0.88134765625,
0.380615234375,
0.88134765625,
0.395751953125,
0.12744140625,
0.395751953125
],
[
0.11767578125,
0.396484375,
0.8798828125,
0.3984375,
0.8798828125,
0.41357421875,
0.11767578125,
0.410888671875
],
[
0.1168212890625,
0.415283203125,
0.88134765625,
0.415283203125,
0.88134765625,
0.4287109375,
0.1168212890625,
0.4287109375
],
[
0.1168212890625,
0.43310546875,
0.8798828125,
0.431884765625,
0.8798828125,
0.4462890625,
0.1168212890625,
0.447509765625
],
[
0.11517333984375,
0.448974609375,
0.88134765625,
0.449462890625,
0.88134765625,
0.464111328125,
0.11517333984375,
0.46337890625
],
[
0.1192626953125,
0.46728515625,
0.87744140625,
0.46728515625,
0.87744140625,
0.48095703125,
0.1192626953125,
0.48095703125
],
[
0.11517333984375,
0.5087890625,
0.31787109375,
0.5087890625,
0.31787109375,
0.5234375,
0.11517333984375,
0.5234375
],
[
0.1192626953125,
0.54150390625,
0.8798828125,
0.54150390625,
0.8798828125,
0.55615234375,
0.1192626953125,
0.55615234375
],
[
0.1192626953125,
0.55810546875,
0.88134765625,
0.55810546875,
0.88134765625,
0.57275390625,
0.1192626953125,
0.57275390625
],
[
0.1168212890625,
0.57568359375,
0.88134765625,
0.57568359375,
0.88134765625,
0.5888671875,
0.1168212890625,
0.5888671875
],
[
0.11767578125,
0.59228515625,
0.88134765625,
0.59228515625,
0.88134765625,
0.60595703125,
0.11767578125,
0.60595703125
],
[
0.11767578125,
0.60986328125,
0.88134765625,
0.60986328125,
0.88134765625,
0.625,
0.11767578125,
0.625
],
[
0.11767578125,
0.62744140625,
0.88134765625,
0.62744140625,
0.88134765625,
0.6416015625,
0.11767578125,
0.6416015625
],
[
0.11517333984375,
0.6416015625,
0.88134765625,
0.642578125,
0.88134765625,
0.66015625,
0.11517333984375,
0.6591796875
],
[
0.13232421875,
0.6611328125,
0.5810546875,
0.6611328125,
0.5810546875,
0.6748046875,
0.13232421875,
0.6748046875
],
[
0.1446533203125,
0.67822265625,
0.88134765625,
0.67822265625,
0.88134765625,
0.6923828125,
0.1446533203125,
0.6923828125
],
[
0.11767578125,
0.69580078125,
0.8798828125,
0.69580078125,
0.8798828125,
0.71044921875,
0.11767578125,
0.71044921875
],
[
0.11767578125,
0.7119140625,
0.3544921875,
0.7119140625,
0.3544921875,
0.7265625,
0.11767578125,
0.7265625
],
[
0.1446533203125,
0.72802734375,
0.88134765625,
0.72900390625,
0.88134765625,
0.74365234375,
0.1446533203125,
0.7421875
],
[
0.11767578125,
0.7470703125,
0.88134765625,
0.7470703125,
0.88134765625,
0.76025390625,
0.11767578125,
0.76025390625
],
[
0.11358642578125,
0.76123046875,
0.59716796875,
0.76220703125,
0.59716796875,
0.77978515625,
0.11358642578125,
0.77880859375
],
[
0.143798828125,
0.78076171875,
0.8798828125,
0.77978515625,
0.8798828125,
0.79443359375,
0.143798828125,
0.79541015625
],
[
0.11517333984375,
0.79443359375,
0.8837890625,
0.79736328125,
0.8837890625,
0.81494140625,
0.11517333984375,
0.8125
],
[
0.1168212890625,
0.81494140625,
0.390625,
0.81494140625,
0.390625,
0.8291015625,
0.1168212890625,
0.8291015625
],
[
0.1446533203125,
0.83056640625,
0.8798828125,
0.83203125,
0.8798828125,
0.8466796875,
0.1446533203125,
0.84521484375
],
[
0.11767578125,
0.849609375,
0.8798828125,
0.849609375,
0.8798828125,
0.86279296875,
0.11767578125,
0.86279296875
],
[
0.1192626953125,
0.8662109375,
0.88134765625,
0.8662109375,
0.88134765625,
0.8798828125,
0.1192626953125,
0.8798828125
],
[
0.491943359375,
0.93603515625,
0.5087890625,
0.93603515625,
0.5087890625,
0.95068359375,
0.491943359375,
0.95068359375
]
] |
[
"在零查询设置中，我们首先显示存在用于k-center的（2,2h-1）-bicriteria算法。此外，该算法",
"是最佳的，因为任何实现有限失真的算法都必须使用2(2）中心。对于（k，2）聚类，我们获得了两",
"种解决所有（k,2）聚类目标的算法。第一个以恒定概率成功，并以（O(logn)k-1+o(1)）多个中心实",
"现恒定失真。第二个需要（O(logn)k+o(1)），并以期望和高概率实现O(1）失真。我们通过证明，对",
"于k-中位数的任何恒定因子失真，即使成功的可能性恒定，2((2logn)k-1+2hlogn）中心也是必",
"要的，从而补充了这一结果。。对于2-中位数的特殊情况，我们的边界是最优的",
"1.2Related Work",
"OrdinalPreferencesandDistortion第一篇考虑使用序数信息的优化问题的论文可能是",
"ProcacciaandRosenschein[53]。随后，建立了两个主要方向。继续使用ProcacciaandRosenschein",
"引人的模型，一行主要关注在归一化假设下最大化福利，但不假设任何度量属性，请参阅Amanatidis",
"etal.[2，4，3]，CaragiannisandProcaccia[19]，Filos-Ratsikasetal.[35]。另一项工作是在没有归",
"一化假设的情况下研究问题，但假设偏好是度量的，即。e。，它们满足三角不等式。除了引言中介",
"绍的聚类论文之外，还研究了其他几个失真问题17，24，25，52。虽然很少见，但也可以在不进行",
"归一化或度量假设的情况下获得一些结果，请参见AbramowitzandAnshelevich[1]",
"ClusteringandFacilityLocation（k,z）-clustering在一般度量[29]中是APX-hard的，",
"尽管在对度量[37，30]或输入[5，14，27]进行假设时可以获得非常准确的算法。对于k-center",
"Gonzalez[40给出了一个最优的2-近似算法。对于k-中位数，k-均值和设施位置，经过大量的研",
"究【43，44，13，48，32，33]，当前的技术水平是k-中位数[41]的2.613近似值，k的9近似值2的",
"XMATHX_。对于一般的（k，z）聚类，尽管k-中位数和k-均值的大多数证明在丢失exp(2）近似",
"因子的同时通过，但几乎没有声称的界限。可以在Cohen-Addadetal.[28，31]中找到明确的结果",
"2Preliminaries",
"令(X,d）为度量空间，其中X为一组n点，d：X×X→R≥0为度量。任何两个点a,yEX之",
"间的距离可以通过d(c,y）形式的query访问。我们假设这样的查询与成本相关联。为算法提供预",
"算，并且该算法进行的每个查询都会消耗其预算的一个单位。虽然查询两点之间的确切距离是昂贵",
"的，但我们的模型假设，对于每个点，有关其到其他点的相对距离的ordinal信息都是免费的。更",
"具体地说，每个点EX提供了一个ranking元a：[n]一X，它是consistent和d，从这个意义",
"上说，每个i,Enl,i<j都有d(,(）≤d(,()）。也就是说，更接近的点在的排名",
"中出现higher。序数preference profileP只是积分排名的集合，即。e。，P=【元}aex。我们为",
"组配置文件编写P(d），其中每个点的排名与距离d一致",
"将点的排名限制为X的某个子集通常很方便。设SCX和m=ISl。restrictionofπatoS",
"是一个函数元,s：[m]→S，因此，对于任何两个y,yES，当且仅当y在元中的排名高于y",
"时，y在元a,s中的排名高于y",
"序数首选项配置文件提供了基本距离度量d的非常粗略的草图。但是，配置文件表示的相对距",
"离可以使算法以非常经济的方式分配其预算。考虑以下操作：对于一组点SCX和一个点&EX",
"，我们将 distance of a to S定义为d(a,S)=minyesd(a,y)",
"给定序数信息，可以很容易地将点z=argminussd(c,y）识别为c在S中排名最高的点。因此",
"一种算法可以通过单个查询d(c,z)确定α到S的距离。显然，对于找到z=argmaxyesd(c,y)",
"和距离d(αc，2）可以进行相同的观察",
"如果我们将算法A限制为序数信息和固定的查询预算，我们打算研究结果最优性的损失。我们",
"考虑了各种聚类问题，其目标是找到使给定成本函数Φ最小化的解决方案。我们用M表示所有",
"度量空间的集合。对于度量空间（X,d)EM和轮廓PEP(d），令A(P,d）为由算法A计算的解",
"3"
] |
{
"arxiv_id": "2402.04035",
"language": "zh",
"page_idx": 2
}
|
||
2402.04035-zh-3
|
2402.04035-zh-3.jpg
|
(中心集),并使 $C^*(d)$ 为最小成本的解 (中心集)。如果
$$\sup\limits_{\substack{(X,d)\in\mathcal{M}\\P\in\mathcal{P}(d)}}\frac{\phi(\mathcal{A}(P,d))}{\phi(C^*(d))}\leq D$$
的概率至少为2/3 (分别概率至少为$1-1/n$ ),则我们说算法$\mathcal{A}$以恒定(分别为高)概率实现
$distortion~D~。{\mathcal{A}}~$的 expected distortion 由比率
$$\sup\limits_{\substack{(X,d)\in\mathcal{M}\\P\in\mathcal{P}(d)}}\frac{\mathbb{E}[\phi(\mathcal{A}(P,d))]}{\phi(C^*(d))}.$$
给出
现在，我们在序数设置中陈述$(k,z)$聚类问题的定义，并介绍一些在聚类问题上下文中常用的
标准术语
$\textbf{Definition 2. 1. In the ordinal }( k, z)$-clustering problem $, we\textit{are given positive integers k, z and}$ $a$ set $X$ $of$ $n$ points that form $a$ metric space $( X, d)$ under distances $d$ . Each point $x\in X$ reports the $points, we$ denote the cost $of$ $a$ given solution $C\subseteq X$ $by$
$$\phi_C(S,d)=\sqrt[z]{\sum_{x\in S}d(x,C)^z}.$$
$The\textit{ goal is to fnd a set C of k points such that the cost function }_{C}( X, d) \textit{ is minimized. For}$ $compactness, we$ drop the dependence $on$ $d$ and denote $by$ $\phi _{OPT}( S)$ the cost $of$ the optimal solution $on\textit{ an arbitrary set of points S X}$ .
给定一个解$C$到一个序数$(k,z)$聚类实例，我们通常调用$C\textit{ centers 的 元 素 。}C$自然地将$X$ 分成$k$ clusters $\{A_c\}_{c\in C}$ ,其中，对于每个$c\in C$ , $A_c=\{x\in X:\pi_{x,C}(1)=c\}$ 。我们将这些簇的集合称为$X$的clustering
最后，我们为所有$(k,z)$聚类目标定义采样概率。
$\textbf{Definition 2. 2. Let }z\textit{ be a positive integer, and let C X be a set of centers. The sampling}$
probability $of$ point $c\in X$ conditioned $on$ having already selected $a$ set $of$ centers $C$ $is$
$$p_z(c):=\mathbb{P}[c\:is\:added\:to\:C\mid C]=\frac{d(c,C)^z}{\sum_{x\in X}d(x,C)^z},$$
$and\textit{ denote the induced distribution by D}_{z}^{+ + }$ .
3 Algorithms for $k$ -Center
我们提出了三种解决有序 $k$ -center $((k,\infty)$ -clustering) 问题的算法。我们的算法基于$\overline{\mathrm{E}}$Gonzalez [40]的贪婪过程，已知该过程会产生$k$中心问题的 2 近似。此过程只需选择一个任意中心，然后在$k-1$迭代中，选择离已经选择的中心最远的中心(最远优先遍历)
4
|
[
[
0.11767578125,
0.09344482421875,
0.5537109375,
0.09344482421875,
0.5537109375,
0.106689453125,
0.11767578125,
0.106689453125
],
[
0.46728515625,
0.11297607421875,
0.5986328125,
0.12371826171875,
0.5947265625,
0.149658203125,
0.463134765625,
0.1395263671875
],
[
0.418212890625,
0.1307373046875,
0.45751953125,
0.1307373046875,
0.45751953125,
0.146484375,
0.418212890625,
0.146484375
],
[
0.40185546875,
0.14013671875,
0.5498046875,
0.1363525390625,
0.55126953125,
0.1572265625,
0.40283203125,
0.1610107421875
],
[
0.41015625,
0.1572265625,
0.463134765625,
0.1572265625,
0.463134765625,
0.1717529296875,
0.41015625,
0.1717529296875
],
[
0.1192626953125,
0.1849365234375,
0.88134765625,
0.1849365234375,
0.88134765625,
0.199462890625,
0.1192626953125,
0.199462890625
],
[
0.1168212890625,
0.20263671875,
0.495849609375,
0.20263671875,
0.495849609375,
0.2171630859375,
0.1168212890625,
0.2171630859375
],
[
0.473876953125,
0.2291259765625,
0.58251953125,
0.2291259765625,
0.58251953125,
0.246826171875,
0.473876953125,
0.246826171875
],
[
0.426513671875,
0.2412109375,
0.4599609375,
0.2412109375,
0.4599609375,
0.25390625,
0.426513671875,
0.25390625
],
[
0.41015625,
0.25390625,
0.473876953125,
0.251953125,
0.473876953125,
0.266357421875,
0.41015625,
0.268310546875
],
[
0.491943359375,
0.248779296875,
0.56640625,
0.248779296875,
0.56640625,
0.266357421875,
0.491943359375,
0.266357421875
],
[
0.4150390625,
0.26513671875,
0.468017578125,
0.26513671875,
0.468017578125,
0.27978515625,
0.4150390625,
0.27978515625
],
[
0.11517333984375,
0.292236328125,
0.1544189453125,
0.292236328125,
0.1544189453125,
0.30810546875,
0.11517333984375,
0.30810546875
],
[
0.1446533203125,
0.31005859375,
0.8798828125,
0.31005859375,
0.8798828125,
0.324462890625,
0.1446533203125,
0.324462890625
],
[
0.1168212890625,
0.32763671875,
0.1895751953125,
0.32763671875,
0.1895751953125,
0.34228515625,
0.1168212890625,
0.34228515625
],
[
0.1168212890625,
0.35595703125,
0.8837890625,
0.35595703125,
0.8837890625,
0.37060546875,
0.1168212890625,
0.37060546875
],
[
0.11517333984375,
0.373779296875,
0.88232421875,
0.373779296875,
0.88232421875,
0.38818359375,
0.11517333984375,
0.38818359375
],
[
0.11358642578125,
0.38818359375,
0.8857421875,
0.389404296875,
0.8857421875,
0.4072265625,
0.11358642578125,
0.406005859375
],
[
0.11358642578125,
0.4072265625,
0.58251953125,
0.40771484375,
0.58251953125,
0.42236328125,
0.11358642578125,
0.421630859375
],
[
0.396240234375,
0.4443359375,
0.486083984375,
0.4443359375,
0.486083984375,
0.462158203125,
0.396240234375,
0.462158203125
],
[
0.484375,
0.44140625,
0.6005859375,
0.438232421875,
0.60107421875,
0.4609375,
0.486083984375,
0.464111328125
],
[
0.495849609375,
0.46337890625,
0.53369140625,
0.46337890625,
0.53369140625,
0.4765625,
0.495849609375,
0.4765625
],
[
0.1168212890625,
0.492431640625,
0.88232421875,
0.492431640625,
0.88232421875,
0.5068359375,
0.1168212890625,
0.5068359375
],
[
0.1168212890625,
0.51025390625,
0.88134765625,
0.51025390625,
0.88134765625,
0.5244140625,
0.1168212890625,
0.5244140625
],
[
0.1168212890625,
0.52783203125,
0.404296875,
0.52783203125,
0.404296875,
0.54248046875,
0.1168212890625,
0.54248046875
],
[
0.1446533203125,
0.55615234375,
0.88134765625,
0.55615234375,
0.88134765625,
0.5693359375,
0.1446533203125,
0.5693359375
],
[
0.1192626953125,
0.5712890625,
0.8798828125,
0.5712890625,
0.8798828125,
0.5859375,
0.1192626953125,
0.5859375
],
[
0.11767578125,
0.58837890625,
0.336669921875,
0.59033203125,
0.336669921875,
0.60498046875,
0.11767578125,
0.60302734375
],
[
0.1446533203125,
0.60693359375,
0.52685546875,
0.60693359375,
0.52685546875,
0.62060546875,
0.1446533203125,
0.62060546875
],
[
0.11517333984375,
0.63330078125,
0.88134765625,
0.63525390625,
0.88134765625,
0.6494140625,
0.11517333984375,
0.64794921875
],
[
0.11767578125,
0.65283203125,
0.78369140625,
0.65283203125,
0.78369140625,
0.66748046875,
0.11767578125,
0.66748046875
],
[
0.5947265625,
0.6796875,
0.65673828125,
0.6796875,
0.65673828125,
0.69580078125,
0.5947265625,
0.69580078125
],
[
0.3056640625,
0.68798828125,
0.56787109375,
0.68798828125,
0.56787109375,
0.70556640625,
0.3056640625,
0.70556640625
],
[
0.56640625,
0.69775390625,
0.68701171875,
0.6962890625,
0.68701171875,
0.71533203125,
0.56640625,
0.71630859375
],
[
0.1168212890625,
0.72998046875,
0.4697265625,
0.72998046875,
0.4697265625,
0.744140625,
0.1168212890625,
0.744140625
],
[
0.11273193359375,
0.76953125,
0.453369140625,
0.77099609375,
0.453369140625,
0.78857421875,
0.11273193359375,
0.787109375
],
[
0.11767578125,
0.8037109375,
0.8798828125,
0.8037109375,
0.8798828125,
0.818359375,
0.11767578125,
0.818359375
],
[
0.1192626953125,
0.8212890625,
0.88134765625,
0.8212890625,
0.88134765625,
0.83544921875,
0.1192626953125,
0.83544921875
],
[
0.11767578125,
0.83837890625,
0.65869140625,
0.83837890625,
0.65869140625,
0.85302734375,
0.11767578125,
0.85302734375
],
[
0.491943359375,
0.93505859375,
0.5087890625,
0.93505859375,
0.5087890625,
0.94970703125,
0.491943359375,
0.94970703125
]
] |
[
"（中心集），并使C*（d）为最小成本的解（中心集）。如果",
"(A(P,)<D",
"sup",
"(X,a)eM (C*(d)",
"PeP(d)",
"的概率至少为2/3（分别概率至少为1－1/n），则我们说算法A以恒定（分别为高）概率实现",
"distortion D。A的 erpected distortion 由比率",
"E[∞(A(P, d)]",
"sup",
"(X,d)eM",
"Φ(C*(d)",
"PEP(d)",
"给出",
"现在，我们在序数设置中陈述（K，2）聚类问题的定义，并介绍一些在聚类问题上下文中常用的",
"标准术语",
"Definition 2.1. In the ordinal (k, z) -clustering problem , we are given positive integers k, z and",
"a set X of n points that form a metric space (X, d) under distances d . Each point E X reports",
"a ranking Ta that is consistent with the distances d . Let P= [nrjrex . For a subset S C X of",
"the points, we denote the cost of a given solution CCX by",
"pc(S,d) =",
"/Zd(r,C)",
"rEs",
"The goal is to find a set C of k points such that the cost function pc(X, d) is minimized. For",
"compactness, we drop the dependence on d and denote by ΦopT(S) the cost of the optimal solution",
"on an arbitrary set of points S C X",
"给定一个解C到一个序数（k,z）聚类实例，我们通常调用Ccenters的元素。C自然地将X",
"分成kclusters{Aclcec，其中，对于每个cEC，Ac=fcEX：πa.c(1)=cl。我们将这些簇",
"的集合称为X的clustering",
"最后，我们为所有（k，2）聚类目标定义采样概率。",
"Definition 2.2. Let z be a positive integer, and let C C X be a set of centers. The sampling",
"probability of point c E X conditioned on having already selected a set of centers C is",
"d(c, C)z",
"pz(c) := P[c is added to C [ C] =",
"Drex d(a,C)",
"and denote the induced distribution by Dt+ .",
"3Algorithms for k-Center",
"我们提出了三种解决有序k -center（（k，oo）-clustering）问题的算法。我们的算法基于EGonzalez",
"[40]的贪婪过程，已知该过程会产生k中心问题的2近似。此过程只需选择一个任意中心，然后",
"在k一1迭代中，选择离已经选择的中心最远的中心（最远优先遍历）",
"4"
] |
{
"arxiv_id": "2402.04035",
"language": "zh",
"page_idx": 3
}
|
||
2402.04035-zh-4
|
2402.04035-zh-4.jpg
|
3.1 2 -Distortion Algorithms
最远优先遍历方法很好地适应了序数设置。显然，给定一组群集，可以通过每个群集一个距离查询来确定距这些群集最远的点。为了完整性，我们在附录 A.1 中给出了该过程的伪代码实现。这会立即产生以下结果
$\textbf{Theorem 3. 1. There exists a deterministic 2 - }distortion\textit{ algorithm for k - center that makes }\frac {k^2- k}2$
$distance~queries.$
对于零查询状态，我们扩展了最远优先遍历方法，以便在每次迭代中选择every集群中最远的点。由于算法及其分析是 Gonzalez [40]的直接改编，我们只需在附录 A.2 中说明结果并给出细节$\textbf{Theorem 3. 2. There exists a deterministic algorithm that, using only ordinal preferences, returns}$ $a$ set $of$ centers $C$ $of$ size $| C| = 2^{k- 1}$ , such that $\max _{x\in X}$ $d( x, C) \leq 2\phi _{OPT}$ , where $\phi _{OPT}$ $is$ the cost $of$ $an$ optimal $k$ $- center$ $clustering.$
3.2 4 -Distortion Algorithm with $O(k)$ Queries
要通过线性(在$k$中) 数量的查询实现恒定失真，其思想是执行$\frac12$ -近似最远优先遍历。这样的最远优先遍历相对于失真是鲁棒的，仅损失 2 的一个因子。令人惊讶的是，使用序数信息，我们可以执行具有最佳查询边界的$\frac12$近似最远第一遍历。在非常高的水平上，我们跟踪整个算法中所有簇的 (中心，最远点) 对。但是，我们不会查询所有对。相反，我们跟踪要查询的一组独立的对，这有助于我们绑定创建的新对的数量，同时确保未查询的对的距离最多是查询距离的两倍。我们在这里给出了完整的分析，并在附录 A.3 中给出了伪代码
Theorem $3. 3. \textit{ There exists a deterministic 4 - }distortion\textit{algorithm to the optimal k - center clus- }$
$tering~that~makes~2k~queries.$
在算法的整个运行过程中，让$C$作为解决方案集，让$Q\subseteq C$作为所谓的查询集。$C$和$Q$都会随着时间的推移而改变，因此我们将$C_i$表示为解决方案，将$Q_i$表示为第$i$次迭代后的查询集，以清楚地说明。此外，对于$y\in C_i$ ,令$S_{y,i}$为点的集合，使得对于这些点中的每个点，$y$是$C_i$中最近的中心，并且令$z_i=\arg\max d(y,x)$。请注意，如果$y$属于查询集$Q_i$ ,则我们查询距离$d(y,z_i)$
在算法的迭代$i\in\{0\}\cup[k-1]$中，我们执行以下步骤：
$\begin{array}{l}1.&\text{Select the cluster }S_{y,i}\:,\mathrm{~for~}y\in Q_{i}\text{ such that }d(y,z_{i})\text{ is maximized and add }z_{i}\text{ to }C_{i}\:,\text{forming}\\C_{i+1}\:.\end{array}$ 2. Remove $y$ from $Q_i$ and let $R_i+1:=C_{i+1}\setminus Q_i$ (i.e. $R_i+1$ always consists at least of $y$ and $z_i).$ 3. Add centers from $R_i+1$ to $Q_i$ to obtain $Q_i+1$ as follows: Let $u\in R_i+1.$
$\bullet$ If there exists a center $p\in Q_i$ such that $d(p,q)\geq d(w,q)$ , where $w=\arg\max d(u,x)$
and $q=\operatorname*{arg}_{x\in\mathcal{S}}\operatorname*{max}_{x\in\mathcal{S}}d(p,x)$ , do not add $u$ to $Q_i.$
$\bullet$ If no such $p$ exists, add $u$ to $Q_i.$
Once all $u’$s have been discarded, we have obtained our new set $Q_i+1.$ All distances between centers in $Q_{i+1}$ and the respective furthest points are queried. Note that we only have to query novel pairs, i.e. already queried pairs do not require a new query.
我们现在证明了关于该算法的几个声明。前两个绑定了查询的数量。最后两个声明在失真上产生了所需的界限：特别是，我们表明，在每次迭代中，我们选择的点不比最远点的距离的一半更近， 并且这种近似的最远优先遍历也会产生对最佳$k$中心解的恒定失真
5
|
[
[
0.1168212890625,
0.09344482421875,
0.410888671875,
0.09344482421875,
0.410888671875,
0.10797119140625,
0.1168212890625,
0.10797119140625
],
[
0.11767578125,
0.12060546875,
0.8798828125,
0.12060546875,
0.8798828125,
0.1337890625,
0.11767578125,
0.1337890625
],
[
0.1168212890625,
0.1363525390625,
0.8798828125,
0.136962890625,
0.8798828125,
0.1514892578125,
0.1168212890625,
0.15087890625
],
[
0.11767578125,
0.154052734375,
0.2607421875,
0.154052734375,
0.2607421875,
0.1673583984375,
0.11767578125,
0.1673583984375
],
[
0.11358642578125,
0.175537109375,
0.875,
0.1749267578125,
0.875,
0.1956787109375,
0.11358642578125,
0.1962890625
],
[
0.11517333984375,
0.1962890625,
0.2442626953125,
0.1976318359375,
0.2442626953125,
0.212158203125,
0.11517333984375,
0.2108154296875
],
[
0.1446533203125,
0.222900390625,
0.88134765625,
0.222900390625,
0.88134765625,
0.2374267578125,
0.1446533203125,
0.2374267578125
],
[
0.1168212890625,
0.2392578125,
0.88134765625,
0.2392578125,
0.88134765625,
0.25390625,
0.1168212890625,
0.25390625
],
[
0.11358642578125,
0.261962890625,
0.88232421875,
0.26318359375,
0.88232421875,
0.281005859375,
0.11358642578125,
0.27978515625
],
[
0.11273193359375,
0.27783203125,
0.8857421875,
0.279052734375,
0.8857421875,
0.2998046875,
0.11273193359375,
0.298583984375
],
[
0.11517333984375,
0.298583984375,
0.381591796875,
0.2998046875,
0.381591796875,
0.314453125,
0.11517333984375,
0.313232421875
],
[
0.1168212890625,
0.335205078125,
0.580078125,
0.335205078125,
0.580078125,
0.349853515625,
0.1168212890625,
0.349853515625
],
[
0.11767578125,
0.361083984375,
0.88134765625,
0.361083984375,
0.88134765625,
0.375,
0.11767578125,
0.375
],
[
0.1168212890625,
0.376953125,
0.8798828125,
0.37744140625,
0.8798828125,
0.392578125,
0.1168212890625,
0.391357421875
],
[
0.11767578125,
0.395751953125,
0.88134765625,
0.395751953125,
0.88134765625,
0.4091796875,
0.11767578125,
0.4091796875
],
[
0.11767578125,
0.412353515625,
0.88134765625,
0.412353515625,
0.88134765625,
0.425537109375,
0.11767578125,
0.425537109375
],
[
0.11767578125,
0.429931640625,
0.88134765625,
0.429931640625,
0.88134765625,
0.443115234375,
0.11767578125,
0.443115234375
],
[
0.11767578125,
0.4462890625,
0.53759765625,
0.4462890625,
0.53759765625,
0.460205078125,
0.11767578125,
0.460205078125
],
[
0.1168212890625,
0.472900390625,
0.8798828125,
0.472900390625,
0.8798828125,
0.4873046875,
0.1168212890625,
0.4873046875
],
[
0.1168212890625,
0.4892578125,
0.346435546875,
0.4892578125,
0.346435546875,
0.50390625,
0.1168212890625,
0.50390625
],
[
0.146240234375,
0.51513671875,
0.88134765625,
0.51513671875,
0.88134765625,
0.52880859375,
0.146240234375,
0.52880859375
],
[
0.1192626953125,
0.5322265625,
0.88134765625,
0.5322265625,
0.88134765625,
0.54541015625,
0.1192626953125,
0.54541015625
],
[
0.11767578125,
0.548828125,
0.88134765625,
0.548828125,
0.88134765625,
0.56298828125,
0.11767578125,
0.56298828125
],
[
0.11767578125,
0.56640625,
0.88232421875,
0.56640625,
0.88232421875,
0.58056640625,
0.11767578125,
0.58056640625
],
[
0.31201171875,
0.5791015625,
0.3505859375,
0.58203125,
0.34716796875,
0.59326171875,
0.3095703125,
0.5888671875
],
[
0.146240234375,
0.5947265625,
0.57177734375,
0.5947265625,
0.57177734375,
0.60791015625,
0.146240234375,
0.60791015625
],
[
0.134033203125,
0.6171875,
0.88232421875,
0.61865234375,
0.88232421875,
0.63623046875,
0.134033203125,
0.63525390625
],
[
0.160888671875,
0.63720703125,
0.214111328125,
0.6396484375,
0.2132568359375,
0.65380859375,
0.16015625,
0.6513671875
],
[
0.1380615234375,
0.6640625,
0.8798828125,
0.6640625,
0.8798828125,
0.6787109375,
0.1380615234375,
0.6787109375
],
[
0.1356201171875,
0.68994140625,
0.7138671875,
0.6923828125,
0.7138671875,
0.70703125,
0.1356201171875,
0.705078125
],
[
0.178955078125,
0.71728515625,
0.8837890625,
0.71826171875,
0.8837890625,
0.73291015625,
0.178955078125,
0.73193359375
],
[
0.7685546875,
0.734375,
0.8203125,
0.734375,
0.8203125,
0.74560546875,
0.7685546875,
0.74560546875
],
[
0.1993408203125,
0.7470703125,
0.56201171875,
0.7470703125,
0.56201171875,
0.76123046875,
0.1993408203125,
0.76123046875
],
[
0.268798828125,
0.75830078125,
0.325927734375,
0.76123046875,
0.324462890625,
0.77587890625,
0.26708984375,
0.77294921875
],
[
0.17724609375,
0.7734375,
0.463134765625,
0.77587890625,
0.463134765625,
0.79345703125,
0.17724609375,
0.79150390625
],
[
0.16015625,
0.80029296875,
0.88134765625,
0.80322265625,
0.88134765625,
0.82080078125,
0.16015625,
0.818359375
],
[
0.16259765625,
0.82275390625,
0.88134765625,
0.82275390625,
0.88134765625,
0.8369140625,
0.16259765625,
0.8369140625
],
[
0.16015625,
0.8369140625,
0.72119140625,
0.83837890625,
0.72119140625,
0.85595703125,
0.16015625,
0.85498046875
],
[
0.1446533203125,
0.8642578125,
0.88134765625,
0.8642578125,
0.88134765625,
0.87744140625,
0.1446533203125,
0.87744140625
],
[
0.1192626953125,
0.8818359375,
0.88232421875,
0.8818359375,
0.88232421875,
0.89599609375,
0.1192626953125,
0.89599609375
],
[
0.11767578125,
0.8984375,
0.63818359375,
0.8984375,
0.63818359375,
0.912109375,
0.11767578125,
0.912109375
],
[
0.491943359375,
0.93603515625,
0.50732421875,
0.93603515625,
0.50732421875,
0.9521484375,
0.491943359375,
0.9521484375
]
] |
[
"3.12-Distortion Algorithms",
"最远优先遍历方法很好地适应了序数设置。显然，给定一组群集，可以通过每个群集一个距离查询",
"来确定距这些群集最远的点。为了完整性，我们在附录A.1中给出了该过程的伪代码实现。这会",
"立即产生以下结果",
"Theorem 3.1. There ecists a deterministic 2 -distortion algorithm for k -center that makes k",
"distance queries.",
"对于零查询状态，我们扩展了最远优先遍历方法，以便在每次选代中选择every集群中最远的",
"点。由于算法及其分析是Gonzalez【40的直接改编，我们只需在附录A.2中说明结果并给出细节",
"Theorem 3.2. There erists a deterministic algorithm that, using only ordinal preferences, returns",
"a set of centers C of size [Cl = 2k-1 , such that maxrex d(r,C) ≤ 2poPT , where ΦoPT is the cost",
"of an optimalk-center clustering",
"3.24-Distortion Algorithm with O(k) Queries",
"要通过线性（在k中）数量的查询实现恒定失真，其思想是执行-近似最远优先遍历。这样的最",
"远优先遍历相对于失真是鲁棒的，仅损失2的一个因子。令人惊的是，使用序数信息，我们可以",
"执行具有最佳查询边界的号近似最远第一遍历。在非常高的水平上，我们跟踪整个算法中所有簇",
"的（中心，最远点）对。但是，我们不会查询所有对。相反，我们跟踪要查询的一组独立的对，这",
"有助于我们绑定创建的新对的数量，同时确保未查询的对的距离最多是查询距离的两倍。我们在这",
"里给出了完整的分析，并在附录A.3中给出了伪代码",
"Theorem 3.3. There erists a deterministic 4 -distortion algorithm to the optimal k -center clus",
"tering that makes 2k queries.",
"在算法的整个运行过程中，让C作为解决方案集，让QC作为所谓的查询集。C和Q都会",
"随着时间的推移而改变，因此我们将Ci表示为解决方案，将Q；表示为第i次选代后的查询集，以",
"清楚地说明。此外，对于yECi，令Sy,为点的集合，使得对于这些点中的每个点，y是C；中最",
"近的中心，并且令zi=argmaxd(y,a）。请注意，如果y属于查询集Qi，则我们查询距离d(y，zi）",
"rESy.",
"在算法的选代iE{0]U[k-1]中，我们执行以下步骤",
"1. Select the cluster Sy,i , for y E Qi such that d(y, zi) is maximized and add zi to Ci , forming",
"Ci+1 :",
"2. Remove y from Qi and let Ri+1 := Ci+i / Qi (i.e. Ri+i always consists at least of y and zi)",
"3. Add centers from Ri+1 to Qi to obtain Qi+1 as follows: Let u ERi+1 :",
": If there exists a center p Qi such that d(p,q) ≥ d(w,q) , where w = argmax d(u,c)",
"rESu,i+",
"and q = argmax d(p, a) , do not add u to Qi -",
"TESp.i+1",
". If no such p exists, add u to Qi :",
"Once all u 's have been discarded, we have obtained our new set Qi+1 : All distances between",
"centers in Qi+i and the respective furthest points are queried. Note that we only have to",
"query novel pairs, i.e. already queried pairs do not require a new query.",
"我们现在证明了关于该算法的几个声明。前两个绑定了查询的数量。最后两个声明在失真上产",
"生了所需的界限：特别是，我们表明，在每次选代中，我们选择的点不比最远点的距离的一半更近，",
"并且这种近似的最远优先遍历也会产生对最佳k中心解的恒定失真",
"5"
] |
{
"arxiv_id": "2402.04035",
"language": "zh",
"page_idx": 4
}
|
||
1905.12304-zh-1
|
1905.12304-zh-1.jpg
|
# 连续空间中的修订：没有对抗性学习的无监督文本样式转换
Dayiheng Liu$^\dagger$,Jie Fu$^\dagger,\textbf{Yidan Zhang}^\dagger,\textbf{Chris Pal}^\dagger,\textbf{Jiancheng Lv}^\dagger{}^*$
$\uparrow$ College of Computer Science, Sichuan University
$\frac{\pi}{\text{Qu\'{e}becArtificialIntelligenceInstitute(Mila),PolytechniqueMontr\'{e}al}}$
losinuris@gmail.com
lvjianchenggescu.edu.cn
## $\mathbf{Abstract}$
Typical methods for unsupervised text style transfer often rely on two key ingredients: 1) seeking the explicit disentanglement of the content and the attributes, and 2) troublesome adversarial learning. In this paper, we show that neither of these components is indispensable. We propose a new framework that utilizes the gradients to revise the sentence in a continuous space during inference to achieve text style transfer. Our method consists of three key components: a variational auto-encoder (VAE), some attribute predictors (one for each attribute), and a content predictor. The VAE and the two types of predictors enable us to perform gradient-based optimization in the continuous space, which is mapped from sentences in a discrete space, to find the representation of a target sentence with the desired attributes and preserved content. Moreover, the proposed method naturally has the ability to simultaneously manipulate multiple fine-grained attributes, such as sentence length and the presence of specific words, when performing text style transfer tasks. Compared with previous adversarial learning based methods, the proposed method is more interpretable , controllable and easier to train. Extensive experimental studies on three popular text style transfer tasks show that the proposed method significantly outperforms five state-of-the-art methods.
# 1 Introduction
文本样式转换，这是文本生成领域中一项尚未探索的挑战性任务，旨在转换句子的某些属性(e。g。,负面情绪 ) 到 其 他 属 性 ( e。g。, 积 极 情 绪 ) , 同 时 保 留 与 属 性 无 关的内容。换句话说，文本样式转移可以以受控的方式生成具有所需属性的句子。由于难以获得具有相同内容和不同样式的训练句子对，因此该任务通常以无监督的方式工作，其中模型只能访问非平行但带有样式标签的句子
用于文本样式转移的大多数现有方法 (Hu et al. 2017; Shen et al. 2017; Fu et al. 2018; Li et al. 2018) 通常首先通过对抗性学习范式 (Goodfellow et al.2014)显式地解开内容和属性。然后将与属性无关的内容和所需的属性向量馈送到解码器中以生成目标句子。然而，最近的一些证据表明，使用对抗性学习可能无法学习解开 (Li et
$^*$ Correspondence to Jiancheng Lv.
Copyright $\textcircled{2}$020, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
al. 2018; Guillaume Lample 2019) 的表示。此外，香草对抗学习旨在生成实值和连续数据，但难以直接生成离散令牌序列。结果，使用了诸如 REINFORCE(Sutton et al. 2000)之类的算法，或者使用了具有温度-softmax 概率向量 (Kusner and Hernández-Lobato 2016; Zhang et al. 2017) 的离散令牌的算法。不幸的是，这些方法往往是不稳定的，缓慢的，和难以调整在实践中(Guillaume Lample 2019)
明确地解开内容和属性真的是必要的吗？另外，我们是否必须使用对抗性学习来实现文本样式转移？最近，将离散输入映射到连续空间中，然后使用预测器执行基于梯度的优化以找到具有所需属性的新离散输出的表示的想法已应用于句子修订 (Mueller, Gifford, and Jaakkola 2017) 和神经架构搜索 (Luo et al. 2018)。在这些作品成功的推动下，我们提出了一种新的解决方案来解决内容保留文本样式转移的任务
所提出的方法包含三个关键组件：(a)变分自动编码器 (VAE) (Kingma and Welling 2013) ,其编码器将句子映射到平滑的连续空间中，其解码器可以将连续表示映射回句子。(b) 一些以句子的连续表示为输入并预测其解码器输出句子的属性的属性预测器，分别。这些属性预测器使我们能够在连续空间中找到具有所需属性的目标句子。(c)内容预测器，将句子的连续表示作为输人，并预测其解码器输出句子的词袋(BoW)特征。组件(c)的目的有三个：首先，它可以增强样式转换过程中的内容保存；第二，它使目标句子包含一些特定的单词；第三，它可以解决 VAE (Zhao,Zhao, and Eskenazi 2017) 的消失的潜在变量问题。利用从这些预测器获得的梯度，我们可以通过基于梯度的优化来修改原始句子的连续表示，以找到具有所需细粒度属性的目标句子， 并实现保留内容的文本样式转移
与以前的方法相比，我们提出的方法具有三个主要优
点：
$\bullet$ The method can be easily trained on the non-parallel
dataset, avoiding the problem of training difficulties caused by adversarial learning and achieving higher performance.
$\bullet$ Unlike previous methods directly generate the target-
style sentence through once feed-forward in the inference stage, our method revises the original sentence with gradient information for several steps during inference, which explicitly presents the process of the style transfer and
|
[
[
0.30224609375,
0.1256103515625,
0.6953125,
0.1256103515625,
0.6953125,
0.14013671875,
0.30224609375,
0.14013671875
],
[
0.39453125,
0.14453125,
0.6044921875,
0.14453125,
0.6044921875,
0.1622314453125,
0.39453125,
0.1622314453125
],
[
0.22216796875,
0.1849365234375,
0.779296875,
0.18310546875,
0.779296875,
0.2008056640625,
0.22216796875,
0.20263671875
],
[
0.330078125,
0.199462890625,
0.66845703125,
0.2020263671875,
0.66845703125,
0.216552734375,
0.330078125,
0.2139892578125
],
[
0.2607421875,
0.216552734375,
0.73779296875,
0.216552734375,
0.73779296875,
0.2298583984375,
0.2607421875,
0.2298583984375
],
[
0.40185546875,
0.2310791015625,
0.5888671875,
0.2310791015625,
0.5888671875,
0.2442626953125,
0.40185546875,
0.2442626953125
],
[
0.390625,
0.24560546875,
0.60888671875,
0.24560546875,
0.60888671875,
0.2587890625,
0.390625,
0.2587890625
],
[
0.25244140625,
0.298583984375,
0.313720703125,
0.298583984375,
0.313720703125,
0.31005859375,
0.25244140625,
0.31005859375
],
[
0.5185546875,
0.296630859375,
0.9111328125,
0.29541015625,
0.9111328125,
0.31005859375,
0.5185546875,
0.311279296875
],
[
0.521484375,
0.311279296875,
0.9111328125,
0.311279296875,
0.9111328125,
0.324462890625,
0.521484375,
0.324462890625
],
[
0.10540771484375,
0.3212890625,
0.4599609375,
0.3212890625,
0.4599609375,
0.33203125,
0.10540771484375,
0.33203125
],
[
0.521484375,
0.323974609375,
0.91259765625,
0.324462890625,
0.91259765625,
0.339111328125,
0.521484375,
0.33837890625
],
[
0.10455322265625,
0.333984375,
0.4599609375,
0.333984375,
0.4599609375,
0.3447265625,
0.10455322265625,
0.3447265625
],
[
0.521484375,
0.339111328125,
0.9111328125,
0.339111328125,
0.9111328125,
0.352783203125,
0.521484375,
0.352783203125
],
[
0.10455322265625,
0.3466796875,
0.461669921875,
0.3466796875,
0.461669921875,
0.356689453125,
0.10455322265625,
0.356689453125
],
[
0.10296630859375,
0.35791015625,
0.463134765625,
0.35791015625,
0.463134765625,
0.371826171875,
0.10296630859375,
0.371826171875
],
[
0.521484375,
0.352783203125,
0.91259765625,
0.352783203125,
0.91259765625,
0.3662109375,
0.521484375,
0.3662109375
],
[
0.101318359375,
0.37060546875,
0.461669921875,
0.37060546875,
0.461669921875,
0.3837890625,
0.101318359375,
0.3837890625
],
[
0.521484375,
0.3662109375,
0.9111328125,
0.3662109375,
0.9111328125,
0.380126953125,
0.521484375,
0.380126953125
],
[
0.51953125,
0.37890625,
0.91259765625,
0.380126953125,
0.91259765625,
0.39453125,
0.51953125,
0.393310546875
],
[
0.10296630859375,
0.385009765625,
0.4599609375,
0.385009765625,
0.4599609375,
0.395751953125,
0.10296630859375,
0.395751953125
],
[
0.10296630859375,
0.397705078125,
0.459228515625,
0.397705078125,
0.459228515625,
0.40771484375,
0.10296630859375,
0.40771484375
],
[
0.52294921875,
0.39453125,
0.5595703125,
0.39453125,
0.5595703125,
0.406005859375,
0.52294921875,
0.406005859375
],
[
0.101318359375,
0.4072265625,
0.461669921875,
0.40771484375,
0.461669921875,
0.421630859375,
0.101318359375,
0.42041015625
],
[
0.53515625,
0.406005859375,
0.91259765625,
0.40771484375,
0.91259765625,
0.42236328125,
0.53515625,
0.42041015625
],
[
0.10455322265625,
0.42236328125,
0.4599609375,
0.42236328125,
0.4599609375,
0.43310546875,
0.10455322265625,
0.43310546875
],
[
0.521484375,
0.421630859375,
0.91259765625,
0.421630859375,
0.91259765625,
0.43505859375,
0.521484375,
0.43505859375
],
[
0.101318359375,
0.43310546875,
0.461669921875,
0.434326171875,
0.461669921875,
0.447509765625,
0.101318359375,
0.4462890625
],
[
0.521484375,
0.43505859375,
0.9111328125,
0.43505859375,
0.9111328125,
0.448974609375,
0.521484375,
0.448974609375
],
[
0.10296630859375,
0.4462890625,
0.461669921875,
0.4462890625,
0.461669921875,
0.460205078125,
0.10296630859375,
0.460205078125
],
[
0.521484375,
0.449462890625,
0.9111328125,
0.449462890625,
0.9111328125,
0.46337890625,
0.521484375,
0.46337890625
],
[
0.101318359375,
0.458984375,
0.461669921875,
0.458984375,
0.461669921875,
0.472900390625,
0.101318359375,
0.472900390625
],
[
0.521484375,
0.46337890625,
0.91259765625,
0.46337890625,
0.91259765625,
0.4765625,
0.521484375,
0.4765625
],
[
0.101318359375,
0.4716796875,
0.461669921875,
0.4716796875,
0.461669921875,
0.48486328125,
0.101318359375,
0.48486328125
],
[
0.51953125,
0.4765625,
0.9111328125,
0.4765625,
0.9111328125,
0.490478515625,
0.51953125,
0.490478515625
],
[
0.10455322265625,
0.48486328125,
0.4599609375,
0.48486328125,
0.4599609375,
0.49560546875,
0.10455322265625,
0.49560546875
],
[
0.10296630859375,
0.496826171875,
0.463134765625,
0.496826171875,
0.463134765625,
0.51025390625,
0.10296630859375,
0.51025390625
],
[
0.521484375,
0.4912109375,
0.91259765625,
0.4912109375,
0.91259765625,
0.5048828125,
0.521484375,
0.5048828125
],
[
0.10455322265625,
0.51025390625,
0.461669921875,
0.51025390625,
0.461669921875,
0.52099609375,
0.10455322265625,
0.52099609375
],
[
0.52294921875,
0.5068359375,
0.765625,
0.5068359375,
0.765625,
0.517578125,
0.52294921875,
0.517578125
],
[
0.10296630859375,
0.521484375,
0.461669921875,
0.521484375,
0.461669921875,
0.53515625,
0.10296630859375,
0.53515625
],
[
0.53759765625,
0.517578125,
0.91259765625,
0.51806640625,
0.91259765625,
0.53271484375,
0.53759765625,
0.5322265625
],
[
0.10296630859375,
0.5341796875,
0.463134765625,
0.5341796875,
0.463134765625,
0.54736328125,
0.10296630859375,
0.54736328125
],
[
0.521484375,
0.53271484375,
0.9111328125,
0.53271484375,
0.9111328125,
0.546875,
0.521484375,
0.546875
],
[
0.10296630859375,
0.546875,
0.463134765625,
0.546875,
0.463134765625,
0.56005859375,
0.10296630859375,
0.56005859375
],
[
0.521484375,
0.546875,
0.9111328125,
0.546875,
0.9111328125,
0.56005859375,
0.521484375,
0.56005859375
],
[
0.10455322265625,
0.56005859375,
0.461669921875,
0.56005859375,
0.461669921875,
0.57080078125,
0.10455322265625,
0.57080078125
],
[
0.52294921875,
0.56005859375,
0.91259765625,
0.56005859375,
0.91259765625,
0.57373046875,
0.52294921875,
0.57373046875
],
[
0.10296630859375,
0.5712890625,
0.463134765625,
0.5712890625,
0.463134765625,
0.58544921875,
0.10296630859375,
0.58544921875
],
[
0.521484375,
0.57373046875,
0.91259765625,
0.57373046875,
0.91259765625,
0.5869140625,
0.521484375,
0.5869140625
],
[
0.10296630859375,
0.583984375,
0.463134765625,
0.583984375,
0.463134765625,
0.59765625,
0.10296630859375,
0.59765625
],
[
0.52392578125,
0.59033203125,
0.9111328125,
0.59033203125,
0.9111328125,
0.60107421875,
0.52392578125,
0.60107421875
],
[
0.10455322265625,
0.5986328125,
0.25,
0.5986328125,
0.25,
0.609375,
0.10455322265625,
0.609375
],
[
0.52294921875,
0.6015625,
0.91015625,
0.6015625,
0.91015625,
0.61572265625,
0.52294921875,
0.61572265625
],
[
0.521484375,
0.61572265625,
0.91259765625,
0.6142578125,
0.91259765625,
0.62890625,
0.521484375,
0.6298828125
],
[
0.20751953125,
0.62890625,
0.35302734375,
0.6298828125,
0.35302734375,
0.64453125,
0.20751953125,
0.64306640625
],
[
0.51953125,
0.62744140625,
0.91259765625,
0.62890625,
0.91259765625,
0.64306640625,
0.51953125,
0.642578125
],
[
0.08905029296875,
0.6484375,
0.47802734375,
0.6484375,
0.47802734375,
0.662109375,
0.08905029296875,
0.662109375
],
[
0.521484375,
0.64306640625,
0.9111328125,
0.64306640625,
0.9111328125,
0.6572265625,
0.521484375,
0.6572265625
],
[
0.0882568359375,
0.6611328125,
0.476318359375,
0.662109375,
0.476318359375,
0.6767578125,
0.0882568359375,
0.67529296875
],
[
0.521484375,
0.6572265625,
0.91259765625,
0.6572265625,
0.91259765625,
0.67041015625,
0.521484375,
0.67041015625
],
[
0.0882568359375,
0.6767578125,
0.47802734375,
0.6767578125,
0.47802734375,
0.69140625,
0.0882568359375,
0.69140625
],
[
0.521484375,
0.671875,
0.9111328125,
0.671875,
0.9111328125,
0.68505859375,
0.521484375,
0.68505859375
],
[
0.521484375,
0.68505859375,
0.9111328125,
0.68505859375,
0.9111328125,
0.69873046875,
0.521484375,
0.69873046875
],
[
0.0906982421875,
0.6923828125,
0.476318359375,
0.6923828125,
0.476318359375,
0.70263671875,
0.0906982421875,
0.70263671875
],
[
0.51953125,
0.69775390625,
0.9140625,
0.69873046875,
0.9140625,
0.71337890625,
0.51953125,
0.7119140625
],
[
0.0882568359375,
0.705078125,
0.47802734375,
0.705078125,
0.47802734375,
0.71826171875,
0.0882568359375,
0.71826171875
],
[
0.0882568359375,
0.71826171875,
0.47802734375,
0.71826171875,
0.47802734375,
0.73193359375,
0.0882568359375,
0.73193359375
],
[
0.521484375,
0.71337890625,
0.7509765625,
0.71337890625,
0.7509765625,
0.7265625,
0.521484375,
0.7265625
],
[
0.0882568359375,
0.73095703125,
0.47802734375,
0.73193359375,
0.47802734375,
0.7470703125,
0.0882568359375,
0.74560546875
],
[
0.5390625,
0.7265625,
0.91259765625,
0.7265625,
0.91259765625,
0.74072265625,
0.5390625,
0.74072265625
],
[
0.52294921875,
0.7412109375,
0.5458984375,
0.7412109375,
0.5458984375,
0.75390625,
0.52294921875,
0.75390625
],
[
0.0882568359375,
0.74755859375,
0.1192626953125,
0.74755859375,
0.1192626953125,
0.7587890625,
0.0882568359375,
0.7587890625
],
[
0.10455322265625,
0.76025390625,
0.47802734375,
0.76025390625,
0.47802734375,
0.7734375,
0.10455322265625,
0.7734375
],
[
0.51953125,
0.75830078125,
0.91259765625,
0.7587890625,
0.91259765625,
0.7734375,
0.51953125,
0.77294921875
],
[
0.08660888671875,
0.7734375,
0.47802734375,
0.7734375,
0.47802734375,
0.787109375,
0.08660888671875,
0.787109375
],
[
0.5361328125,
0.7734375,
0.91259765625,
0.7734375,
0.91259765625,
0.787109375,
0.5361328125,
0.787109375
],
[
0.08660888671875,
0.78857421875,
0.47802734375,
0.78857421875,
0.47802734375,
0.8017578125,
0.08660888671875,
0.8017578125
],
[
0.5361328125,
0.78857421875,
0.9140625,
0.78857421875,
0.9140625,
0.8017578125,
0.5361328125,
0.8017578125
],
[
0.0882568359375,
0.8017578125,
0.47802734375,
0.8017578125,
0.47802734375,
0.81494140625,
0.0882568359375,
0.81494140625
],
[
0.5361328125,
0.80322265625,
0.60302734375,
0.80322265625,
0.60302734375,
0.814453125,
0.5361328125,
0.814453125
],
[
0.0882568359375,
0.81640625,
0.463134765625,
0.81640625,
0.463134765625,
0.830078125,
0.0882568359375,
0.830078125
],
[
0.521484375,
0.818359375,
0.9140625,
0.82080078125,
0.9140625,
0.83544921875,
0.521484375,
0.83349609375
],
[
0.08660888671875,
0.82763671875,
0.4794921875,
0.8291015625,
0.4794921875,
0.84228515625,
0.08660888671875,
0.84130859375
],
[
0.53759765625,
0.833984375,
0.91259765625,
0.833984375,
0.91259765625,
0.84765625,
0.53759765625,
0.84765625
],
[
0.1168212890625,
0.8515625,
0.31787109375,
0.8515625,
0.31787109375,
0.86572265625,
0.1168212890625,
0.86572265625
],
[
0.53759765625,
0.8486328125,
0.9140625,
0.8486328125,
0.9140625,
0.8623046875,
0.53759765625,
0.8623046875
],
[
0.08660888671875,
0.86279296875,
0.4794921875,
0.86279296875,
0.4794921875,
0.876953125,
0.08660888671875,
0.876953125
],
[
0.53759765625,
0.86279296875,
0.91259765625,
0.86279296875,
0.91259765625,
0.876953125,
0.53759765625,
0.876953125
],
[
0.08660888671875,
0.876953125,
0.370849609375,
0.876953125,
0.370849609375,
0.89013671875,
0.08660888671875,
0.89013671875
],
[
0.5390625,
0.87744140625,
0.9111328125,
0.87744140625,
0.9111328125,
0.88818359375,
0.5390625,
0.88818359375
]
] |
[
"连续空间中的修订：没有对抗性学习的",
"无监督文本样式转换",
"Dayiheng Liut, Jie Fut, Yidan Zhang', Chris Palt, Jiancheng Lvt *",
"t College of Computer Science, Sichuan University",
" Quebec Artificial Intelligence Institute (Mila), Polytechnique Montréal",
"losinuris@gmail.com",
"lvjiancheng@scu.edu.cn",
"Abstract",
"al.2018;GuillaumeLample2019）的表示。此外，香草对",
"抗学习旨在生成实值和连续数据，但难以直接生成离散",
"Typicalmethodsforunsupervisedtextstyletransferoftenrely",
"令牌序列。结果，使用了诸如REINFORCE(Suttonetal.",
"on two key ingredients:I) seeking the explicit disentangle",
"2000)之类的算法，或者使用了具有温度-softmax概率向",
"ment of the content and the attributes,and 2)troublesome",
"adversarial learning. In this paper, we show that neither of",
"量 (Kusner and Hernandez-Lobato 2016; Zhang et al. 2017)",
"these components is indispensable. We propose a new frame",
"的离散令牌的算法。不幸的是，这些方法往往是不稳",
"定的，缓慢的，和难以调整在实践中（GuillaumeLample",
"workthatutilizesthegradients torevise the sentence in a con",
"tinuous space during inferenceto achievetext style transfer",
"2019)",
"Our method consists of three key components: a variational",
"明确地解开内容和属性真的是必要的吗？另外，我",
"auto-encoder (VAE),someattributepredictors (onefor each",
"们是否必须使用对抗性学习来实现文本样式转移？最",
"attribute), and a content predictor.The VAE and the two types",
"近，将离散输人映射到连续空间中，然后使用预测器执",
"of predictors enable us to perform gradient-based optimiza-",
"行基于梯度的优化以找到具有所需属性的新离散输出",
"tion in the continuous space, which is mapped from sentences",
"的表示的想法已应用于句子修订（Mueller，Gifford，and",
"in a discrete space, to find the representation of a target sen.",
"Jaakkola2017)和神经架构搜索（Luoetal.2018）。在这",
"tence with the desired attributes and preserved content.More",
"over, the proposed method naturally has the ability to simul-",
"些作品成功的推动下，我们提出了一种新的解决方案来",
"taneously manipulate multiple fine-grained attributes, such",
"解决内容保留文本样式转移的任务",
"as sentence length and the presence of specific words, when",
"所提出的方法包含三个关键组件：（a）变分自动编码",
"performing text style transfer tasks. Compared with previous",
"器（VAE）（KingmaandWelling2013），其编码器将句子",
"adversarial learning based methods, the proposed method is",
"映射到平滑的连续空间中，其解码器可以将连续表示映",
"more interpretable,controllable andeasier to train.Extensive",
"射回句子。（b）一些以句子的连续表示为输人并预测其",
"experimental studies on three popular text style transfer tasks",
"解码器输出句子的属性的属性预测器，分别。这些属性",
"show that the proposed method significantly outperforms five",
"预测器使我们能够在连续空间中找到具有所需属性的",
"state-of-the-art methods.",
"目标句子。（）内容预测器，将句子的连续表示作为输",
"人，并预测其解码器输出句子的词袋（BoW）特征。组",
"1Introduction",
"件（c）的目的有三个：首先，它可以增强样式转换过",
"文本样式转换，这是文本生成领域中一项尚未探索的挑",
"程中的内容保存；第二，它使目标句子包含一些特定的",
"战性任务，旨在转换句子的某些属性（e。g。，负面情",
"单词；第三，它可以解决VAE（Zhao,Zhao，andEskenazi",
"绪)到其他属性(e。g。，积极情绪)，同时保留与属性无",
"2017）的消失的潜在变量问题。利用从这些预测器获得",
"的梯度，我们可以通过基于梯度的优化来修改原始句子",
"关的内容。换句话说，文本样式转移可以以受控的方式",
"的连续表示，以找到具有所需细粒度属性的目标句子，",
"生成具有所需属性的句子。由于难以获得具有相同内容",
"和不同样式的训练句子对，因此该任务通常以无监督的",
"并实现保留内容的文本样式转移",
"方式工作，其中模型只能访问非平行但带有样式标签的",
"与以前的方法相比，我们提出的方法具有三个主要优",
"点：",
"句子",
"用于文本样式转移的大多数现有方法(Huetal.2017;",
". The method can be easily trained on the non-parallel",
"Shen et al.2017;Fu et al.2018;Li etal.2018)通常首先",
"dataset, avoiding the problem of training difficulties",
"通过对抗性学习范式（Goodfellowetal.2014）显式地解",
"caused by adversarial learning and achieving higher per-",
"开内容和属性。然后将与属性无关的内容和所需的属性",
"formance.",
"向量馈送到解码器中以生成目标句子。然而，最近的",
" Unlike previous methods directly generate the target-",
"些证据表明，使用对抗性学习可能无法学习解开（Liet",
"style sentence through once feed-forward in the inference",
"Correspondence to Jiancheng Lv",
"stage, our method revises the original sentence with gradi-",
"Copyright 2020, Association for the Advancement of Artificial",
"ent information for several steps during inference, which",
"Intelligence (www.aaai.org).All rights reserved.",
"explicitlypresentstheprocessof thestyletransfer and"
] |
{
"arxiv_id": "1905.12304",
"language": "zh",
"page_idx": 1
}
|
||
1905.12304-zh-2
|
1905.12304-zh-2.jpg
|
can easily provide us multiple results with tuning the gradients. Therefore, the proposed method has higher interpretability and is more controllable .
$\bullet$ Most previous text style transfer methods that only control a single binary attribute (e.g., positive and negative sentiments). In contrast, our approach is more generic in the sense that it naturally has the ability to control multiple fine-grained attributes, such as sentence length and ${\hat{\text{the existence of specific words.}}}$
对三种流行的文本样式转移任务进行的广泛实验比较表明，所提出的方法明显优于五种最先进的方法。源代码可在 https://github.com/dayihengliu/Fine-GrainedStyle-Transfer
# 2 $\textbf{Methodology}$
令$\mathcal{D}=\{(x^1,s^1),...,(x^n,s^n)\}$表示包含$n$句子的数据集$x^i$与一组属性$s^i$配对。每个$s$都有感兴趣的$k$属性$s=\{s_{1},...,s_{k}\}$。与大多数以前的方法(Shen et al. 2017; Fu et al. 2018; Prabhumoye et al. 2018; Li et al. 2018; Yang et al. 2018) 不同 , 后者只考虑单个二进制属性(e。g。, 正可是名可使做》代好的文法中的一个时上在中中也故故也也们面或负面情绪),我们的方法自然具有在风格转移过程中控制多个细粒度属性的能力。在这里，我们采用两个细粒度属性，句子长度和特定单词的存在(e。g。,一个预定义的主语名词),作为案例研究。例如，给定一个原始句子 $x=“$沙拉新鲜美味。”,其属性集可以是 $s=\{$情绪 $= 正 $,长 度 =7, 主 题 $\_ 名 词 = 沙 拉 \} 。我 们 的 任 务 是$ 学习一个生成模型$G$,该模型可以生成具有所需属性$s^*$的新句子$x^*$,并尽可能保留$x$的与属性无关的内容
# 2.1 Model Structure
建议的模型由三个组件组成：变分自动编码器(VAE),
属性预测器和内容预测器
变分自动编码器$G$ 。VAE 将随机潜在表示$z$集成到自动编码器体系结构中。它的 RNN 编码器将句子$x$映射为连续的潜在表示$z:$
(1)
$$z\sim G_{\mathrm{enc}}(\theta_{\mathrm{enc}};x)=q_E(z|x),$$
,其 RNN 解码器将表示映射回以重建句子$x:$
(2)
$$x\sim G_{\mathrm{dec}}(\theta_{\mathrm{dec}};z)=p_{G}(x|z),$$
,其中$\theta_\mathrm{enc}$和$\theta_\mathrm{dec}$表示编码器和解码器的参数。然后优化 VAE 以最小化输入句子的重建误差 $\mathcal{L}_\mathrm{rec}$,同时最小化 KL 项$\mathcal{L}_\mathrm{KL}$以鼓励$q_E(z|x)$匹配先前的$p(z):$
$$\begin{aligned}&\mathcal{L}_{\mathrm{VAE}}(\theta_{\mathrm{enc}},\theta_{\mathrm{dec}})=\mathcal{L}_{\mathrm{rec}}+\mathcal{L}_{\mathrm{KL}}\\&=-\mathbb{E}_{q_{E}(z|x)}\left[\log p_{G}(x|z)\right]+\mathcal{D}_{\mathrm{KL}}(q_{E}(z|x)\|p(z)),\end{aligned}$$
(3)
,其中$\mathcal{D}_\mathrm{KL}(\cdot\|\cdot)$是 KL 发散度。与传统的确定性自动编码器相比，VAE 在我们的方法中提供了两个主要优势： (1)确定性自动编码器通常在其潜在空间中具有“孔”, 其中潜在表示可能无法生成任何逼真的(Roberts et al. 2018)。相反，通过在潜在表示上施加先前的标准化正态分布$\mathcal{N} ( z; 0, I)$ , VAE 学习潜在表示不是作为单个孤立点，而是作为连续潜在空间中的软密集区域，这使得它能够从潜在空间中的每个点生成合理的示例 (Bowman et al.2016)。此特性避免了由渐变修改(优化)的表示$z^*$无法生成合理句子的问题
(2) VAE学习的这种连续而平滑的潜在空间使相邻潜在表示生成的句子在内容和语义 (Bowman et al. 2016; Semeniuta, Severyn, and Barth 2017; Goyal et al. 2017; Yang et al. 2017; Shen et al. 2018) 上相似。因此，如果我们在合理的范围内修改表示$z$ (i。e。,足够小),则生成的新句子在内容上与原始句子不会有太大差异
属性预测器$f_1,...,f_k$ 。他们每个人都将表示$z$作为输人，并预测$z$生成的解码器输出句子$\hat{x}$的一个属性$s_j$ 。例如，属性预测器可以是用于正负情绪预测的二进制分类器或用于句子长度预测的回归模型。利用预测器提供的梯度，我们可以通过基于梯度的优化来修改原始句子$x$的连续表示$z$,以找到具有所需属性$s^*$的目标句子$x^*$
属性预测器$f_1,...,f_k$与 VAE 联合训练。对于 M 分类
预测因子，我们有
$$\mathcal{L}_{\mathrm{Attr},s_{j}}(\theta_{s_{j}},\theta_{\mathrm{enc}})=-\mathbb{E}_{q_{E}(z|x)}\log\left[f_{j}(z)\right],$$
(4)
其中$f_j(z)=\mathbf{MLP}_j(z)=p(s_j|z)\in\mathbb{R^M}$ 。对于回归预
测变量，我们有
$$\mathcal{L}_{\text{Attr},s_j}(\theta_{s_j},\theta_{\text{enc}})=\mathbb{E}_{q_E(z|x)}\left[(s_j-f_j(z))^2\right],$$
(5)
其中$f_j(z)=\mathbf{MLP}_j(z)\in\mathbb{R}^1$。在这个联合训练中，我们将输入句子$x$的属性作为预言者的标签。由于预测器是为了预测$z$生成的句子$\hat{x}$的属性而设计的，因此我们在联合训练后进一步单独训练每个预测器。我们从$\mathcal{N}(z;0,I)$中采样$z$并将其送入解码器以生成新句子$\hat{x}$ 。之后，我们将$\hat{x}$输入到 CNN 文本分类器(Kim 2014)中， 这些分类器在训练集上训练以预测其属性$^{1}$作为预测变量的标签：
$$\begin{aligned}&\mathcal{L}_{\mathrm{Attr},s_{j}}^{\prime}(\theta_{s_{j}})=-\mathbb{E}_{p(z)p_{G}(\hat{x}|z)}\log\left[p(\mathrm{CNN}(\hat{x})|z)\right],\\&\mathcal{L}_{\mathrm{Attr},s_{j}}^{\prime}(\theta_{s_{j}})=\mathbb{E}_{p(z)p_{G}(\hat{x}|z)}\left[(\hat{s}_{j}-f_{j}(z))^{2}\right].\end{aligned}$$
(6)
内容预测器$f_\mathrm{bow}$ 。它是一个多标签分类器，以$z$作为输
人，并预测其解码器输出句子的词袋特征$x_\mathrm{bow}:$
$$f_{\mathrm{bow}}(z)=\mathrm{MLP_{bow}}(z)=p(x_{\mathrm{bow}}|z).$$
(7)
我们假设$p(x_\mathrm{bow}|z)$为$|x|$ -试验多峰分布：
$$\log p(x_{\mathrm{bow}}|z)=\log\prod_{t=1}^{|x|}\frac{e^{f_{\mathrm{bow}}^{(x_{t})}}}{\sum_{j}^{\nu}e^{f_{\mathrm{bow}}^{(x_{j})}}},$$
(8)
的 XMATHX 值是
内容预测器$f_\mathrm{bow}$的训练类似于属性预测器。与 VAE
联合训练：
(9)
$$\mathcal{L}_{\mathrm{BOW}}(\theta_{\mathrm{bow}},\theta_{\mathrm{enc}})=-\mathbb{E}_{q_{E}(z|x)}\log\left[p(x_{\mathrm{bow}}|z)\right].$$
联合训练后，通过以下方式单独训练：
(10)
$$\mathcal{L}_{\mathrm{BOW}}^{\prime}(\theta_{\mathrm{bow}})=-\mathbb{E}_{p(z)p_{G}(\hat{x}|z)}\log\left[p(\hat{x}_{\mathrm{bow}}|z)\right].$$
在文本样式传输期间，我们可以类似地使用内容预测器$f_\mathrm{bow}$提供的渐变来修改表示$z$ ,以增强内容保存。在这里，我们考虑两种在样式传输过程中增强内容保存的方法。我们可以将$x_\mathrm{bow}$设置为包含原句$x$中的所有单
$^{1}$Some attributes can be obtained directly without using classi-
fiers, such as the length $\hat{s}_j$ of $\hat{x}.$
|
[
[
0.10296630859375,
0.0694580078125,
0.4794921875,
0.0706787109375,
0.4794921875,
0.083984375,
0.10296630859375,
0.08331298828125
],
[
0.53759765625,
0.07135009765625,
0.90869140625,
0.07135009765625,
0.90869140625,
0.08209228515625,
0.53759765625,
0.08209228515625
],
[
0.10455322265625,
0.083984375,
0.4794921875,
0.083984375,
0.4794921875,
0.09783935546875,
0.10455322265625,
0.09783935546875
],
[
0.51953125,
0.08331298828125,
0.91015625,
0.08209228515625,
0.91015625,
0.09661865234375,
0.51953125,
0.09783935546875
],
[
0.101318359375,
0.09783935546875,
0.339111328125,
0.09783935546875,
0.339111328125,
0.111083984375,
0.101318359375,
0.111083984375
],
[
0.521484375,
0.09783935546875,
0.91259765625,
0.09783935546875,
0.91259765625,
0.111083984375,
0.521484375,
0.111083984375
],
[
0.0882568359375,
0.11553955078125,
0.4794921875,
0.117431640625,
0.4794921875,
0.1307373046875,
0.0882568359375,
0.1287841796875
],
[
0.521484375,
0.111083984375,
0.9111328125,
0.111083984375,
0.9111328125,
0.125,
0.521484375,
0.125
],
[
0.521484375,
0.1256103515625,
0.91259765625,
0.1256103515625,
0.91259765625,
0.1395263671875,
0.521484375,
0.1395263671875
],
[
0.101318359375,
0.1319580078125,
0.47802734375,
0.1319580078125,
0.47802734375,
0.1458740234375,
0.101318359375,
0.1458740234375
],
[
0.51953125,
0.136962890625,
0.8798828125,
0.1383056640625,
0.8798828125,
0.15283203125,
0.51953125,
0.1514892578125
],
[
0.10455322265625,
0.1458740234375,
0.4794921875,
0.1458740234375,
0.4794921875,
0.1590576171875,
0.10455322265625,
0.1590576171875
],
[
0.521484375,
0.15283203125,
0.9111328125,
0.15283203125,
0.9111328125,
0.1666259765625,
0.521484375,
0.1666259765625
],
[
0.10455322265625,
0.160400390625,
0.47802734375,
0.160400390625,
0.47802734375,
0.17041015625,
0.10455322265625,
0.17041015625
],
[
0.10296630859375,
0.17236328125,
0.4794921875,
0.17236328125,
0.4794921875,
0.186279296875,
0.10296630859375,
0.186279296875
],
[
0.52294921875,
0.1666259765625,
0.90673828125,
0.1666259765625,
0.90673828125,
0.179931640625,
0.52294921875,
0.179931640625
],
[
0.521484375,
0.1787109375,
0.9111328125,
0.179931640625,
0.9111328125,
0.1932373046875,
0.521484375,
0.1925048828125
],
[
0.10296630859375,
0.186279296875,
0.306396484375,
0.186279296875,
0.306396484375,
0.199462890625,
0.10296630859375,
0.199462890625
],
[
0.521484375,
0.1925048828125,
0.9111328125,
0.1932373046875,
0.9111328125,
0.2083740234375,
0.521484375,
0.20703125
],
[
0.10296630859375,
0.205810546875,
0.47802734375,
0.205810546875,
0.47802734375,
0.2197265625,
0.10296630859375,
0.2197265625
],
[
0.521484375,
0.2083740234375,
0.9111328125,
0.2083740234375,
0.9111328125,
0.2215576171875,
0.521484375,
0.2215576171875
],
[
0.08660888671875,
0.2197265625,
0.4794921875,
0.2197265625,
0.4794921875,
0.23291015625,
0.08660888671875,
0.23291015625
],
[
0.521484375,
0.222900390625,
0.9111328125,
0.222900390625,
0.9111328125,
0.236083984375,
0.521484375,
0.236083984375
],
[
0.0882568359375,
0.23291015625,
0.4794921875,
0.23291015625,
0.4794921875,
0.246826171875,
0.0882568359375,
0.246826171875
],
[
0.52294921875,
0.2374267578125,
0.53759765625,
0.2374267578125,
0.53759765625,
0.24560546875,
0.52294921875,
0.24560546875
],
[
0.0882568359375,
0.248779296875,
0.1829833984375,
0.248779296875,
0.1829833984375,
0.260009765625,
0.0882568359375,
0.260009765625
],
[
0.53759765625,
0.25,
0.91259765625,
0.25,
0.91259765625,
0.264404296875,
0.53759765625,
0.264404296875
],
[
0.521484375,
0.264404296875,
0.65283203125,
0.264404296875,
0.65283203125,
0.27783203125,
0.521484375,
0.27783203125
],
[
0.2034912109375,
0.271484375,
0.3544921875,
0.274658203125,
0.3544921875,
0.292236328125,
0.2034912109375,
0.2890625
],
[
0.576171875,
0.285400390625,
0.9140625,
0.285888671875,
0.9140625,
0.3037109375,
0.576171875,
0.302978515625
],
[
0.0882568359375,
0.29541015625,
0.47802734375,
0.29541015625,
0.47802734375,
0.309326171875,
0.0882568359375,
0.309326171875
],
[
0.08660888671875,
0.30810546875,
0.476318359375,
0.309326171875,
0.476318359375,
0.323974609375,
0.08660888671875,
0.322509765625
],
[
0.52392578125,
0.31005859375,
0.91259765625,
0.309326171875,
0.91259765625,
0.326904296875,
0.52392578125,
0.32763671875
],
[
0.0882568359375,
0.323974609375,
0.47802734375,
0.323974609375,
0.47802734375,
0.337158203125,
0.0882568359375,
0.337158203125
],
[
0.521484375,
0.324462890625,
0.63623046875,
0.324462890625,
0.63623046875,
0.33837890625,
0.521484375,
0.33837890625
],
[
0.0849609375,
0.3359375,
0.47802734375,
0.337158203125,
0.47802734375,
0.3515625,
0.0849609375,
0.35107421875
],
[
0.0849609375,
0.349853515625,
0.476318359375,
0.35107421875,
0.476318359375,
0.365478515625,
0.0849609375,
0.3642578125
],
[
0.56640625,
0.348388671875,
0.91259765625,
0.348388671875,
0.91259765625,
0.363037109375,
0.56640625,
0.363037109375
],
[
0.0882568359375,
0.3642578125,
0.47802734375,
0.3642578125,
0.47802734375,
0.37744140625,
0.0882568359375,
0.37744140625
],
[
0.53369140625,
0.371826171875,
0.9111328125,
0.37255859375,
0.9111328125,
0.386962890625,
0.53369140625,
0.386474609375
],
[
0.08905029296875,
0.37890625,
0.476318359375,
0.37890625,
0.476318359375,
0.392578125,
0.08905029296875,
0.392578125
],
[
0.0882568359375,
0.392578125,
0.476318359375,
0.392578125,
0.476318359375,
0.4072265625,
0.0882568359375,
0.4072265625
],
[
0.52294921875,
0.386962890625,
0.9111328125,
0.386962890625,
0.9111328125,
0.40087890625,
0.52294921875,
0.40087890625
],
[
0.52294921875,
0.399658203125,
0.9111328125,
0.399658203125,
0.9111328125,
0.41357421875,
0.52294921875,
0.41357421875
],
[
0.08905029296875,
0.406005859375,
0.476318359375,
0.406005859375,
0.476318359375,
0.419189453125,
0.08905029296875,
0.419189453125
],
[
0.52294921875,
0.41357421875,
0.9111328125,
0.41357421875,
0.9111328125,
0.427978515625,
0.52294921875,
0.427978515625
],
[
0.0882568359375,
0.42041015625,
0.4755859375,
0.42041015625,
0.4755859375,
0.434326171875,
0.0882568359375,
0.434326171875
],
[
0.521484375,
0.427978515625,
0.91259765625,
0.427978515625,
0.91259765625,
0.44140625,
0.521484375,
0.44140625
],
[
0.0882568359375,
0.434326171875,
0.476318359375,
0.434326171875,
0.476318359375,
0.447509765625,
0.0882568359375,
0.447509765625
],
[
0.52294921875,
0.44140625,
0.90869140625,
0.44140625,
0.90869140625,
0.455078125,
0.52294921875,
0.455078125
],
[
0.08660888671875,
0.447509765625,
0.4755859375,
0.447509765625,
0.4755859375,
0.4609375,
0.08660888671875,
0.4609375
],
[
0.521484375,
0.453857421875,
0.9111328125,
0.455078125,
0.9111328125,
0.4697265625,
0.521484375,
0.468505859375
],
[
0.08660888671875,
0.462158203125,
0.4755859375,
0.462158203125,
0.4755859375,
0.47607421875,
0.08660888671875,
0.47607421875
],
[
0.52392578125,
0.47021484375,
0.59326171875,
0.47021484375,
0.59326171875,
0.481689453125,
0.52392578125,
0.481689453125
],
[
0.08660888671875,
0.488037109375,
0.259033203125,
0.488037109375,
0.259033203125,
0.501953125,
0.08660888671875,
0.501953125
],
[
0.54345703125,
0.4912109375,
0.8759765625,
0.4912109375,
0.8759765625,
0.5087890625,
0.54345703125,
0.5087890625
],
[
0.0882568359375,
0.50830078125,
0.4794921875,
0.50830078125,
0.4794921875,
0.521484375,
0.0882568359375,
0.521484375
],
[
0.54150390625,
0.51123046875,
0.8388671875,
0.51220703125,
0.8388671875,
0.52978515625,
0.54150390625,
0.52880859375
],
[
0.890625,
0.505859375,
0.9111328125,
0.505859375,
0.9111328125,
0.5166015625,
0.890625,
0.5166015625
],
[
0.0882568359375,
0.52294921875,
0.268798828125,
0.52294921875,
0.268798828125,
0.5361328125,
0.0882568359375,
0.5361328125
],
[
0.0882568359375,
0.5361328125,
0.4755859375,
0.5361328125,
0.4755859375,
0.5498046875,
0.0882568359375,
0.5498046875
],
[
0.52294921875,
0.53857421875,
0.9111328125,
0.53857421875,
0.9111328125,
0.5517578125,
0.52294921875,
0.5517578125
],
[
0.0882568359375,
0.5498046875,
0.47802734375,
0.5498046875,
0.47802734375,
0.56298828125,
0.0882568359375,
0.56298828125
],
[
0.521484375,
0.55029296875,
0.861328125,
0.5517578125,
0.861328125,
0.56640625,
0.521484375,
0.56494140625
],
[
0.0882568359375,
0.564453125,
0.24267578125,
0.564453125,
0.24267578125,
0.57763671875,
0.0882568359375,
0.57763671875
],
[
0.6005859375,
0.57470703125,
0.833984375,
0.57470703125,
0.833984375,
0.58837890625,
0.6005859375,
0.58837890625
],
[
0.890625,
0.57470703125,
0.91259765625,
0.57470703125,
0.91259765625,
0.58837890625,
0.890625,
0.58837890625
],
[
0.1846923828125,
0.5859375,
0.37744140625,
0.5859375,
0.37744140625,
0.599609375,
0.1846923828125,
0.599609375
],
[
0.455810546875,
0.5869140625,
0.4794921875,
0.5869140625,
0.4794921875,
0.60107421875,
0.455810546875,
0.60107421875
],
[
0.52294921875,
0.5966796875,
0.810546875,
0.5966796875,
0.810546875,
0.60986328125,
0.52294921875,
0.60986328125
],
[
0.0882568359375,
0.60986328125,
0.416748046875,
0.60986328125,
0.416748046875,
0.62353515625,
0.0882568359375,
0.62353515625
],
[
0.734375,
0.619140625,
0.75732421875,
0.619140625,
0.75732421875,
0.630859375,
0.734375,
0.630859375
],
[
0.7548828125,
0.62255859375,
0.80615234375,
0.62060546875,
0.80810546875,
0.63623046875,
0.75634765625,
0.63818359375
],
[
0.1846923828125,
0.6337890625,
0.378173828125,
0.6337890625,
0.378173828125,
0.64794921875,
0.1846923828125,
0.64794921875
],
[
0.45751953125,
0.6337890625,
0.47802734375,
0.6337890625,
0.47802734375,
0.64794921875,
0.45751953125,
0.64794921875
],
[
0.6005859375,
0.6337890625,
0.7548828125,
0.6337890625,
0.7548828125,
0.6484375,
0.6005859375,
0.6484375
],
[
0.890625,
0.63525390625,
0.91259765625,
0.63525390625,
0.91259765625,
0.6484375,
0.890625,
0.6484375
],
[
0.7255859375,
0.64599609375,
0.833984375,
0.63525390625,
0.83837890625,
0.66015625,
0.7294921875,
0.67041015625
],
[
0.10455322265625,
0.65771484375,
0.4755859375,
0.65771484375,
0.4755859375,
0.66845703125,
0.10455322265625,
0.66845703125
],
[
0.0882568359375,
0.67041015625,
0.47802734375,
0.67041015625,
0.47802734375,
0.68505859375,
0.0882568359375,
0.68505859375
],
[
0.521484375,
0.673828125,
0.64892578125,
0.673828125,
0.64892578125,
0.6875,
0.521484375,
0.6875
],
[
0.08660888671875,
0.68310546875,
0.42236328125,
0.6845703125,
0.42236328125,
0.69873046875,
0.08660888671875,
0.69775390625
],
[
0.5390625,
0.6875,
0.91259765625,
0.6875,
0.91259765625,
0.7021484375,
0.5390625,
0.7021484375
],
[
0.521484375,
0.70068359375,
0.59326171875,
0.70068359375,
0.59326171875,
0.71533203125,
0.521484375,
0.71533203125
],
[
0.10540771484375,
0.70703125,
0.2998046875,
0.70703125,
0.2998046875,
0.7216796875,
0.10540771484375,
0.7216796875
],
[
0.459228515625,
0.7197265625,
0.4755859375,
0.7197265625,
0.4755859375,
0.72900390625,
0.459228515625,
0.72900390625
],
[
0.56201171875,
0.72265625,
0.9140625,
0.7236328125,
0.9140625,
0.7412109375,
0.56201171875,
0.74072265625
],
[
0.521484375,
0.75,
0.78759765625,
0.75,
0.78759765625,
0.76318359375,
0.521484375,
0.76318359375
],
[
0.108642578125,
0.724609375,
0.443603515625,
0.7236328125,
0.443603515625,
0.7412109375,
0.108642578125,
0.7421875
],
[
0.10296630859375,
0.75048828125,
0.47802734375,
0.75048828125,
0.47802734375,
0.7646484375,
0.10296630859375,
0.7646484375
],
[
0.0882568359375,
0.7646484375,
0.471435546875,
0.7646484375,
0.471435546875,
0.77783203125,
0.0882568359375,
0.77783203125
],
[
0.5537109375,
0.771484375,
0.8857421875,
0.77294921875,
0.8857421875,
0.79052734375,
0.5537109375,
0.7890625
],
[
0.10455322265625,
0.77880859375,
0.476318359375,
0.77880859375,
0.476318359375,
0.79248046875,
0.10455322265625,
0.79248046875
],
[
0.8798828125,
0.7744140625,
0.9111328125,
0.7744140625,
0.9111328125,
0.787109375,
0.8798828125,
0.787109375
],
[
0.08905029296875,
0.79150390625,
0.476318359375,
0.79150390625,
0.476318359375,
0.8046875,
0.08905029296875,
0.8046875
],
[
0.53759765625,
0.79736328125,
0.9111328125,
0.79736328125,
0.9111328125,
0.81103515625,
0.53759765625,
0.81103515625
],
[
0.08660888671875,
0.8046875,
0.47802734375,
0.80615234375,
0.47802734375,
0.8193359375,
0.08660888671875,
0.818359375
],
[
0.521484375,
0.8125,
0.9111328125,
0.8125,
0.9111328125,
0.82568359375,
0.521484375,
0.82568359375
],
[
0.0882568359375,
0.8193359375,
0.476318359375,
0.8193359375,
0.476318359375,
0.83349609375,
0.0882568359375,
0.83349609375
],
[
0.521484375,
0.82568359375,
0.9111328125,
0.82568359375,
0.9111328125,
0.8388671875,
0.521484375,
0.8388671875
],
[
0.08905029296875,
0.833984375,
0.476318359375,
0.833984375,
0.476318359375,
0.84765625,
0.08905029296875,
0.84765625
],
[
0.52294921875,
0.8388671875,
0.9111328125,
0.8388671875,
0.9111328125,
0.85302734375,
0.52294921875,
0.85302734375
],
[
0.0882568359375,
0.84765625,
0.476318359375,
0.84765625,
0.476318359375,
0.861328125,
0.0882568359375,
0.861328125
],
[
0.0882568359375,
0.8623046875,
0.47802734375,
0.8623046875,
0.47802734375,
0.87548828125,
0.0882568359375,
0.87548828125
],
[
0.54150390625,
0.861328125,
0.9140625,
0.86279296875,
0.9140625,
0.87744140625,
0.54150390625,
0.87548828125
],
[
0.08660888671875,
0.87548828125,
0.28759765625,
0.87548828125,
0.28759765625,
0.8896484375,
0.08660888671875,
0.8896484375
],
[
0.521484375,
0.87548828125,
0.71142578125,
0.87548828125,
0.71142578125,
0.8896484375,
0.521484375,
0.8896484375
]
] |
[
"can easily provide us multiple results with tuning the gra-",
"（2）VAE学习的这种连续而平滑的潜在空间使相邻潜",
"dients. Therefore, the proposed method has higher inter-",
"在表示生成的句子在内容和语义（Bowmanetal.2016",
"pretability and is more controllable",
"Semeniuta, Severyn, and Barth 2017; Goyal et al. 2017;",
". Most previous text style transfer methods that only con-",
"Yangetal.2017:Shenetal.2018）上相似。因此，如果",
"我们在合理的范围内修改表示（i。e。，足够小），则",
"trol a single binary attribute (e.g., positive and negative",
"生成的新句子在内容上与原始句子不会有太大差异",
"sentiments). In contrast, our approach is more generic in",
"属性预测器f1..,fk。他们每个人都将表示之作为输",
"the sensethatitnaturallyhas the abilityto control mul",
"tiple fine-grained attributes, such as sentence length and",
"人，并预测z生成的解码器输出句子的一个属性Si",
"例如，属性预测器可以是用于正负情绪预测的二进制分",
"the existence of specific words.",
"类器或用于句子长度预测的回归模型。利用预测器提供",
"对三种流行的文本样式转移任务进行的广泛实验比",
"的梯度，我们可以通过基于梯度的优化来修改原始句子",
"较表明，所提出的方法明显优于五种最先进的方法。",
"的连续表示，以找到具有所需属性s*的目标句子",
"源代码可在https://github.com/dayihengliu/Fine-Grained-",
"r*",
"Style-Transfer",
"属性预测器fi..,fk与VAE联合训练。对于M分类",
"预测因子，我们有",
"2Methodology",
"CAtrs,(0s,, Dene) =-Eq(a/a) log [f;(2)],(4)",
"令D=(αl,sl),..,(α\",sn)}表示包含n句子的数据",
"集与一组属性s配对。每个s都有感兴趣的k属性",
"，其中fi(2）=MLP,(）=p(silz）ERM。对于回归预",
"s={s1,..sk)。与大多数以前的方法(Shenetal.2017;",
"测变量，我们有",
"Fu et al. 2018; Prabhumoye et al. 2018; Li et al. 2018; Yang",
"etal.2018)不同，后者只考虑单个二进制属性(e。g。，正",
"CAtr,s, (0s,, 0enc) = EgE(2|2) [(sj - f,(z))21,(5)",
"面或负面情绪），我们的方法自然具有在风格转移过程",
"其中f,(2)=MLP;(z)ER。在这个联合训练中，我",
"中控制多个细粒度属性的能力。在这里，我们采用两个",
"细粒度属性，句子长度和特定单词的存在（e。g。，一个",
"们将输入句子的属性作为预言者的标签。由于预测",
"器是为了预测之生成的句子的属性而设计的，因此",
"预定义的主语名词），作为案例研究。例如，给定一个原",
"我们在联合训练后进一步单独训练每个预测器。我们从",
"始句子=“沙拉新鲜美味。”，其属性集可以是S={情",
"Nz0,I)中采样z并将其送人解码器以生成新句子。",
"绪=正，长度=7，主题名词=沙拉)。我们的任务是",
"之后，我们将输人到CNN文本分类器（Kim2014)中",
"学习一个生成模型G，该模型可以生成具有所需属性",
"这些分类器在训练集上训练以预测其属性！作为预测变",
"s*的新句子α*，并尽可能保留的与属性无关的内容",
"量的标签：",
"2.1 Model Structure",
"CAlr,s, (0s,) = -Ep(z)Pc([2) log [p(CNN()[2)] , ",
"建议的模型由三个组件组成：变分自动编码器（VAE），",
"CAltr,s,(0,) = Ep(z)pe(a|2) [($j - fi(2)2] .",
"(6)",
"属性预测器和内容预测器",
"变分自动编码器G。VAE将随机潜在表示z集成到自",
"内容预测器fbow。它是一个多标签分类器，以z作为输",
"动编码器体系结构中。它的RNN编码器将句子&映射",
"人，并预测其解码器输出句子的词袋特征abow：",
"为连续的潜在表示之：",
"fbow(2) = MLPbow(2) = p(Cbow2).",
"(7)",
"z~ Genc(Oenci) =qE(z)",
"(1)",
"我们假设p(abowz）为-试验多峰分布：",
"，其RNN解码器将表示映射回以重建句子：",
"[r]",
"eflrt",
" ~ Gdec(Odec; z) = pG(rz),",
"(2)",
"log p(Cbow|z) = log T",
"(8)",
"Heyer",
"其中Oenc和6dec表示编码器和解码器的参数。然后优",
"化VAE以最小化输人句子的重建误差Crec，同时最小",
"的XMATHX值是",
"化KL项CkL以鼓励qE（zc）匹配先前的p(）：",
"内容预测器fbow的训练类似于属性预测器。与VAE",
"联合训练：",
"CVAE(Oenc,Odec)=Crec+CkI",
"(3)",
"CBow(Obow, Oenc) =-Ege(z/z) log [p(rbow|2)]. (9)",
"联合训练后，通过以下方式单独训练：",
"=-EgE(z) [log PG(a|2)] + DkL(qE(2[)p(2)),",
"其中DkL(-II-）是KL发散度。与传统的确定性自动编",
"码器相比，VAE在我们的方法中提供了两个主要优势：",
"CBow(Obow) =-Ep(2)pc(|z) log [p(abow/2)] ",
"（1）确定性自动编码器通常在其潜在空间中具有“孔”",
"(10)",
"其中潜在表示可能无法生成任何逼真的(Robertsetal.",
"在文本样式传输期间，我们可以类似地使用内容预测",
"2018）。相反，通过在潜在表示上施加先前的标准化正态",
"器fbow提供的渐变来修改表示z，以增强内容保存。在",
"分布N（z;0,I），VAE学习潜在表示不是作为单个孤立",
"这里，我们考虑两种在样式传输过程中增强内容保存的",
"点，而是作为连续潜在空间中的软密集区域，这使得它",
"方法。我们可以将Cbow设置为包含原句中的所有单",
"能够从潜在空间中的每个点生成合理的示例(Bowman",
"etal.2016）。此特性避免了由渐变修改（优化）的表示",
"ISome attributes can be obtained directly without using classi-",
"*无法生成合理句子的问题",
"fiers, such as the length s, of ?."
] |
{
"arxiv_id": "1905.12304",
"language": "zh",
"page_idx": 2
}
|
||
1905.12304-zh-3
|
1905.12304-zh-3.jpg
|
<FigureHere>
continuous latent space
Figure 1: There is an example of content-preserving text sentiment transfer, and we hope to further increase the length of the target sentence compared with the original sentence. The original sentence $x$ with negative sentiment is mapped to continuous representation $z$ via encoder. Then $z$ is revised into $z^*$ by minimizing the error $\mathcal{L}_\text{Attr}s_1(\theta_{s_1};s_1=\{$sentiment$=positive\})+$ $\mathcal{L} _{\mathrm{Aut, }s\_ 2}( \theta _{s_2}; s_2$ = $\{$length = $20\} ) +$ $\lambda _\mathrm{bow}\mathcal{L} _{\mathrm{BOW}}( \theta _{\mathrm{bow}}; x_{\mathrm{bow}}$ = $[ burgers, meat] )$ with the sentiment predictor $f_1$ , length predictor $f_2$ , and the content predictor $f_\mathrm{bow}\cdot$ Afterwards the target sentence $x^*$ is generated by decoding $z^*$ with beam search via decoder [best viewed in color].
文本情感传递任务进行了几次实验，并同时控制了其他
细粒度属性，例如长度和关键字存在
### 3.1 $\textbf{Text Sentiment Transfen}$
词，这意味着我们试图找到一个具有所需属性$s^*$的句子$x^*$,并尽可能保留原句的所有单词，以实现内容保存。但是，保留所有单词通常不是我们想要的。例如，$x^*$在文本情感传递任务中不应包含原始情感词。相反，原始句子中的名词应该保留在这样的任务 (Melnyk et al. 2017; Li et al. 2018; John et al. 2019) 中。因此，我们可以将$x_\mathrm{bow}$ 设置为仅包含$x$中的所有名词。此外，我们可以将$x_\mathrm{bow}$ 设置为包含一些所需的特定单词，以实现对目标句子的更细粒度控制
将它们放在一起，最终的联合训练损失$\mathcal{L}$如下：
Data 我们使用两个数据集，Yelp 餐厅评论和亚马逊产品评论 (He and McAuley 2016)$^{2}$,它们在以前的作品中也很常用 (Shen et al. 2017; Fu et al. 2018; Li et al. 2018; Prabhumoye et al. 2018)。按照他们的实验设置，我们使用相同的预处理步骤和类似的实验配置
(11)
$$\mathcal{L}=\mathcal{L}_{\mathrm{VAE}}+\lambda_{b}\mathcal{L}_{\mathrm{BOW}}+\lambda_{s}\sum_{j=1}^{k}\mathcal{L}_{\mathrm{Attr},s_{j}},$$
其中$\lambda_b$和$\lambda_s$是平衡超参数。应该注意的是，$\mathcal{L}_\mathrm{BOW}$和$\mathcal{L}_{\mathrm{Attr},s_j}$也充当正则化，防止编码器陷入 KL 消失状态Bowman et al. 2016: Kingma et al. 2016: Yang et al. 2017: Shen et al. 2018; Alemi et al. 2018)
## 2.2 $\textbf{Text Style Transfer}$
给定原始句子$x$ ,在连续空间中进行样式传递的推理过
程。我们通过基于梯度的优化将其表示$z$修改如下：
Metrics 良好的样式转换(Li et al. 2018; Prabhumoye et al.2018)有三个条件。具体而言，生成的句子应：1)具有所需的属性；2)流利；3)尽可能保留原始句子的与属性无关的内容。对于第一个和第二个标准，我们遵循以前的工作 (Shen et al. 2017; Fu et al. 2018; Li et al. 2018; Prabhumoye et al. 2018) 使用基于模型的评估。我们根据预先训练的双向 LSTM 分类器 (Schuster and Paliwal 1997) 的预测来衡量风格是否成功转移，并通过预先训练的语言模型生成的句子的困惑度(PPL)来衡量语言质量。在先前的工作之后，我们使用在相应数据集上训练的 TrigramKneser-Ney 平滑语言模型 (Kneser and Ney 1995)。由于难以衡量内容的保存性，因此我们遵循以前的工作并报告了两个指标：1)单词重叠， 它计算了原始句子$x$和生成的句子$\hat{x}$的单字单词重叠率，由$\frac{\mathrm{count}(w_x\cap w_x)}{\mathrm{count}(w_x\cup w_x)}$计算；2)正如(Melnyk et al.2017 Li et al. 2018) 中所指出的那样，句子中的几乎所有名词都是与属性无关的内容，应保留在样式转换任务中(。g。,由 POS 标记器检测到)出现在生成的句子中的原始句子中 (表示为名词%)。(Li et al.2018) 中有 1000 个带有人类注释的句子作为转移句子的基本事实。我们还将它们作为参考，并报告 bi-gramBLEU 分数 (Papineni et al.2002)
(12,
$$\hat{z}=z-\eta(\sum_{j=1}^{k}\nabla_{z}\mathcal{L}_{\mathrm{Attr},s_{j}}+\lambda_{c}\nabla_{z}\mathcal{L}_{\mathrm{BOW}}),$$
,其中$\eta$是步长$,\lambda_c$ 是权衡参数，以平衡内容保存和样式传输强度。我们迭代这种优化以找到$z^*$,直到属性预测器$p(s_j|z)$的输出置信度得分大于阈值$\beta$或达到最大回合数$T$。通过使用波束搜索(Och and Ney 2004)解码$z^*$来获得目标$x^*$ 。示例过程如图 1 所示
# 3 Experiments
在本节中，我们在情绪传递和性别风格传递任务的三个公开可用数据集上评估了所提出的方法。然后，我们对
$^{2}$These datasets can be download at http://bit.ly/2LHMUsl
|
[
[
0.3857421875,
0.08709716796875,
0.62841796875,
0.08709716796875,
0.62841796875,
0.09783935546875,
0.3857421875,
0.09783935546875
],
[
0.6298828125,
0.106689453125,
0.63623046875,
0.106689453125,
0.63623046875,
0.11236572265625,
0.6298828125,
0.11236572265625
],
[
0.62744140625,
0.1142578125,
0.6396484375,
0.1142578125,
0.6396484375,
0.1256103515625,
0.62744140625,
0.1256103515625
],
[
0.62744140625,
0.12371826171875,
0.640625,
0.126953125,
0.63623046875,
0.1383056640625,
0.62353515625,
0.1351318359375
],
[
0.62109375,
0.1331787109375,
0.640625,
0.1331787109375,
0.640625,
0.1590576171875,
0.62109375,
0.1590576171875
],
[
0.2132568359375,
0.14453125,
0.281005859375,
0.14453125,
0.281005859375,
0.1558837890625,
0.2132568359375,
0.1558837890625
],
[
0.72216796875,
0.1427001953125,
0.7939453125,
0.1427001953125,
0.7939453125,
0.1572265625,
0.72216796875,
0.1572265625
],
[
0.468017578125,
0.1610107421875,
0.47802734375,
0.1610107421875,
0.47802734375,
0.1666259765625,
0.468017578125,
0.1666259765625
],
[
0.62841796875,
0.160400390625,
0.640625,
0.160400390625,
0.640625,
0.1685791015625,
0.62841796875,
0.1685791015625
],
[
0.61767578125,
0.1685791015625,
0.6396484375,
0.1685791015625,
0.6396484375,
0.1767578125,
0.61767578125,
0.1767578125
],
[
0.62158203125,
0.1787109375,
0.63818359375,
0.1787109375,
0.63818359375,
0.1875,
0.62158203125,
0.1875
],
[
0.7099609375,
0.1787109375,
0.81396484375,
0.18115234375,
0.81396484375,
0.1956787109375,
0.7099609375,
0.1932373046875
],
[
0.7060546875,
0.1925048828125,
0.830078125,
0.1932373046875,
0.830078125,
0.2083740234375,
0.7060546875,
0.20703125
],
[
0.1895751953125,
0.1817626953125,
0.318603515625,
0.18310546875,
0.318603515625,
0.1962890625,
0.1895751953125,
0.1956787109375
],
[
0.1895751953125,
0.1976318359375,
0.306396484375,
0.1976318359375,
0.306396484375,
0.2083740234375,
0.1895751953125,
0.2083740234375
],
[
0.1895751953125,
0.210205078125,
0.265625,
0.210205078125,
0.265625,
0.2215576171875,
0.1895751953125,
0.2215576171875
],
[
0.7099609375,
0.20703125,
0.8095703125,
0.20703125,
0.8095703125,
0.2171630859375,
0.7099609375,
0.2171630859375
],
[
0.708984375,
0.2183837890625,
0.830078125,
0.2183837890625,
0.830078125,
0.231689453125,
0.708984375,
0.231689453125
],
[
0.708984375,
0.23291015625,
0.7958984375,
0.23291015625,
0.7958984375,
0.24365234375,
0.708984375,
0.24365234375
],
[
0.39453125,
0.25830078125,
0.59619140625,
0.25830078125,
0.59619140625,
0.271484375,
0.39453125,
0.271484375
],
[
0.08660888671875,
0.28857421875,
0.91259765625,
0.28857421875,
0.91259765625,
0.3017578125,
0.08660888671875,
0.3017578125
],
[
0.0849609375,
0.300537109375,
0.91259765625,
0.3017578125,
0.91259765625,
0.31640625,
0.0849609375,
0.31494140625
],
[
0.0849609375,
0.31494140625,
0.91259765625,
0.31640625,
0.91259765625,
0.330810546875,
0.0849609375,
0.330078125
],
[
0.0882568359375,
0.330078125,
0.91259765625,
0.330078125,
0.91259765625,
0.3447265625,
0.0882568359375,
0.3447265625
],
[
0.08660888671875,
0.3447265625,
0.91259765625,
0.3447265625,
0.91259765625,
0.35791015625,
0.08660888671875,
0.35791015625
],
[
0.08660888671875,
0.35791015625,
0.313720703125,
0.35791015625,
0.313720703125,
0.371826171875,
0.08660888671875,
0.371826171875
],
[
0.0882568359375,
0.397705078125,
0.47802734375,
0.397705078125,
0.47802734375,
0.410888671875,
0.0882568359375,
0.410888671875
],
[
0.52294921875,
0.397705078125,
0.91259765625,
0.397705078125,
0.91259765625,
0.410888671875,
0.52294921875,
0.410888671875
],
[
0.0882568359375,
0.410888671875,
0.47802734375,
0.410888671875,
0.47802734375,
0.4248046875,
0.0882568359375,
0.4248046875
],
[
0.51953125,
0.410888671875,
0.78369140625,
0.410888671875,
0.78369140625,
0.4248046875,
0.51953125,
0.4248046875
],
[
0.08905029296875,
0.4267578125,
0.476318359375,
0.4267578125,
0.476318359375,
0.4375,
0.08905029296875,
0.4375
],
[
0.08660888671875,
0.439453125,
0.47802734375,
0.439453125,
0.47802734375,
0.45263671875,
0.08660888671875,
0.45263671875
],
[
0.5185546875,
0.439453125,
0.75244140625,
0.439453125,
0.75244140625,
0.45263671875,
0.5185546875,
0.45263671875
],
[
0.08660888671875,
0.45263671875,
0.47802734375,
0.451904296875,
0.47802734375,
0.46533203125,
0.08660888671875,
0.466552734375
],
[
0.52392578125,
0.458984375,
0.9111328125,
0.458984375,
0.9111328125,
0.472900390625,
0.52392578125,
0.472900390625
],
[
0.0849609375,
0.46533203125,
0.476318359375,
0.46728515625,
0.476318359375,
0.481689453125,
0.0849609375,
0.479736328125
],
[
0.08660888671875,
0.479248046875,
0.473876953125,
0.479736328125,
0.473876953125,
0.494384765625,
0.08660888671875,
0.49365234375
],
[
0.521484375,
0.473388671875,
0.9111328125,
0.473388671875,
0.9111328125,
0.4873046875,
0.521484375,
0.4873046875
],
[
0.521484375,
0.486083984375,
0.91259765625,
0.4873046875,
0.91259765625,
0.501953125,
0.521484375,
0.50048828125
],
[
0.0882568359375,
0.494384765625,
0.47802734375,
0.494384765625,
0.47802734375,
0.50830078125,
0.0882568359375,
0.50830078125
],
[
0.521484375,
0.501953125,
0.91259765625,
0.501953125,
0.91259765625,
0.51513671875,
0.521484375,
0.51513671875
],
[
0.08905029296875,
0.5087890625,
0.1846923828125,
0.5087890625,
0.1846923828125,
0.52099609375,
0.08905029296875,
0.52099609375
],
[
0.52294921875,
0.51513671875,
0.7998046875,
0.51513671875,
0.7998046875,
0.52880859375,
0.52294921875,
0.52880859375
],
[
0.08660888671875,
0.521484375,
0.42236328125,
0.521484375,
0.42236328125,
0.5361328125,
0.08660888671875,
0.5361328125
],
[
0.339111328125,
0.54248046875,
0.35302734375,
0.54248046875,
0.35302734375,
0.55029296875,
0.339111328125,
0.55029296875
],
[
0.1502685546875,
0.548828125,
0.416748046875,
0.5517578125,
0.416748046875,
0.57373046875,
0.1502685546875,
0.57080078125
],
[
0.525390625,
0.55322265625,
0.58642578125,
0.55322265625,
0.58642578125,
0.564453125,
0.525390625,
0.564453125
],
[
0.59326171875,
0.5517578125,
0.91259765625,
0.5517578125,
0.91259765625,
0.56494140625,
0.59326171875,
0.56494140625
],
[
0.52294921875,
0.5673828125,
0.91015625,
0.5673828125,
0.91015625,
0.57763671875,
0.52294921875,
0.57763671875
],
[
0.449462890625,
0.55615234375,
0.47802734375,
0.55615234375,
0.47802734375,
0.568359375,
0.449462890625,
0.568359375
],
[
0.33251953125,
0.5712890625,
0.358642578125,
0.5712890625,
0.358642578125,
0.58349609375,
0.33251953125,
0.58349609375
],
[
0.521484375,
0.580078125,
0.9111328125,
0.580078125,
0.9111328125,
0.59326171875,
0.521484375,
0.59326171875
],
[
0.08905029296875,
0.59130859375,
0.47802734375,
0.59130859375,
0.47802734375,
0.60595703125,
0.08905029296875,
0.60595703125
],
[
0.52294921875,
0.59228515625,
0.9111328125,
0.59228515625,
0.9111328125,
0.60595703125,
0.52294921875,
0.60595703125
],
[
0.08660888671875,
0.60400390625,
0.4794921875,
0.60400390625,
0.4794921875,
0.6220703125,
0.08660888671875,
0.6220703125
],
[
0.521484375,
0.60693359375,
0.91259765625,
0.60693359375,
0.91259765625,
0.62060546875,
0.521484375,
0.62060546875
],
[
0.08660888671875,
0.6171875,
0.47802734375,
0.6171875,
0.47802734375,
0.630859375,
0.08660888671875,
0.630859375
],
[
0.51953125,
0.619140625,
0.9111328125,
0.62060546875,
0.9111328125,
0.63525390625,
0.51953125,
0.6337890625
],
[
0.0849609375,
0.6318359375,
0.32177734375,
0.6318359375,
0.32177734375,
0.64599609375,
0.0849609375,
0.64599609375
],
[
0.52294921875,
0.63623046875,
0.9111328125,
0.63623046875,
0.9111328125,
0.646484375,
0.52294921875,
0.646484375
],
[
0.521484375,
0.6484375,
0.91259765625,
0.6484375,
0.91259765625,
0.662109375,
0.521484375,
0.662109375
],
[
0.08660888671875,
0.65771484375,
0.27783203125,
0.65771484375,
0.27783203125,
0.671875,
0.08660888671875,
0.671875
],
[
0.52294921875,
0.662109375,
0.91259765625,
0.662109375,
0.91259765625,
0.67529296875,
0.52294921875,
0.67529296875
],
[
0.0882568359375,
0.6748046875,
0.4755859375,
0.6748046875,
0.4755859375,
0.68798828125,
0.0882568359375,
0.68798828125
],
[
0.521484375,
0.6767578125,
0.9111328125,
0.6767578125,
0.9111328125,
0.68994140625,
0.521484375,
0.68994140625
],
[
0.08905029296875,
0.69140625,
0.453369140625,
0.69140625,
0.453369140625,
0.7021484375,
0.08905029296875,
0.7021484375
],
[
0.521484375,
0.68994140625,
0.91259765625,
0.68994140625,
0.91259765625,
0.7041015625,
0.521484375,
0.7041015625
],
[
0.521484375,
0.705078125,
0.91259765625,
0.705078125,
0.91259765625,
0.71826171875,
0.521484375,
0.71826171875
],
[
0.226318359375,
0.7138671875,
0.23779296875,
0.7138671875,
0.23779296875,
0.72607421875,
0.226318359375,
0.72607421875
],
[
0.1405029296875,
0.72265625,
0.42724609375,
0.72607421875,
0.42724609375,
0.7470703125,
0.1405029296875,
0.74365234375
],
[
0.52294921875,
0.71826171875,
0.91015625,
0.71826171875,
0.91015625,
0.73193359375,
0.52294921875,
0.73193359375
],
[
0.44775390625,
0.72900390625,
0.47802734375,
0.72900390625,
0.47802734375,
0.74072265625,
0.44775390625,
0.74072265625
],
[
0.521484375,
0.73193359375,
0.91259765625,
0.73193359375,
0.91259765625,
0.74560546875,
0.521484375,
0.74560546875
],
[
0.2181396484375,
0.744140625,
0.24267578125,
0.744140625,
0.24267578125,
0.755859375,
0.2181396484375,
0.755859375
],
[
0.51953125,
0.74365234375,
0.65673828125,
0.75390625,
0.654296875,
0.77099609375,
0.51708984375,
0.76025390625
],
[
0.58837890625,
0.744140625,
0.646484375,
0.744140625,
0.646484375,
0.7548828125,
0.58837890625,
0.7548828125
],
[
0.64794921875,
0.74853515625,
0.91259765625,
0.74853515625,
0.91259765625,
0.76220703125,
0.64794921875,
0.76220703125
],
[
0.098876953125,
0.7587890625,
0.476318359375,
0.76025390625,
0.476318359375,
0.7744140625,
0.098876953125,
0.7734375
],
[
0.521484375,
0.76513671875,
0.91259765625,
0.76513671875,
0.91259765625,
0.77880859375,
0.521484375,
0.77880859375
],
[
0.08905029296875,
0.7744140625,
0.476318359375,
0.7744140625,
0.476318359375,
0.78857421875,
0.08905029296875,
0.78857421875
],
[
0.521484375,
0.77978515625,
0.91259765625,
0.77978515625,
0.91259765625,
0.79345703125,
0.521484375,
0.79345703125
],
[
0.0882568359375,
0.78857421875,
0.47802734375,
0.78857421875,
0.47802734375,
0.80322265625,
0.0882568359375,
0.80322265625
],
[
0.51953125,
0.79345703125,
0.91259765625,
0.79248046875,
0.91259765625,
0.806640625,
0.51953125,
0.80810546875
],
[
0.08660888671875,
0.80029296875,
0.47802734375,
0.8017578125,
0.47802734375,
0.81640625,
0.08660888671875,
0.81494140625
],
[
0.521484375,
0.806640625,
0.91259765625,
0.806640625,
0.91259765625,
0.82080078125,
0.521484375,
0.82080078125
],
[
0.0882568359375,
0.81640625,
0.398681640625,
0.81640625,
0.398681640625,
0.830078125,
0.0882568359375,
0.830078125
],
[
0.521484375,
0.8212890625,
0.91259765625,
0.8212890625,
0.91259765625,
0.83544921875,
0.521484375,
0.83544921875
],
[
0.521484375,
0.83544921875,
0.91259765625,
0.83544921875,
0.91259765625,
0.8486328125,
0.521484375,
0.8486328125
],
[
0.20751953125,
0.84228515625,
0.35302734375,
0.84228515625,
0.35302734375,
0.85693359375,
0.20751953125,
0.85693359375
],
[
0.521484375,
0.849609375,
0.580078125,
0.849609375,
0.580078125,
0.861328125,
0.521484375,
0.861328125
],
[
0.0882568359375,
0.861328125,
0.4755859375,
0.861328125,
0.4755859375,
0.875,
0.0882568359375,
0.875
],
[
0.0882568359375,
0.87548828125,
0.47802734375,
0.87548828125,
0.47802734375,
0.8896484375,
0.0882568359375,
0.8896484375
],
[
0.5390625,
0.87353515625,
0.890625,
0.875,
0.890625,
0.8896484375,
0.5390625,
0.88818359375
]
] |
[
"errorsurfaceofthepredictor",
"2.",
"1.5",
"1.0",
"9",
"Encoder",
"Decoder",
"",
"0.5",
"-1.0",
"1.5",
"\"The burgers were",
"cooked very perfectly,",
"\"The burgers were over",
"cooked and themeat",
"was crunchy.\"",
"and themeatwas",
"tasty and juicy, which",
"impressedme.\"",
"continuouslatent space",
"Figure l: There is an example of content-preserving text sentiment transfer, and we hope to further increase the length of the",
"target sentence compared with the original sentence. The original sentence r with negative sentiment is mapped to continuous",
"representation z via encoder. Then z is revised into z* by minimizing the error LAttr,s_i(0s; si = [sentiment = positive)) +",
"CAtr,s_2(0s2; $2 = [length = 20)) + bowLBow(0bow; Cbow =[burgers,meat]) with the sentiment predictor fi, length",
"predictor f2 , and the content predictor fbow . Afterwards the target sentence r* is generated by decoding z* with beam search",
"via decoder [best viewed in color].",
"词，这意味着我们试图找到一个具有所需属性s*的句子",
"文本情感传递任务进行了几次实验，并同时控制了其他",
"&*，并尽可能保留原句的所有单词，以实现内容保存。但",
"细粒度属性，例如长度和关键字存在",
"是，保留所有单词通常不是我们想要的。例如，*在文",
"本情感传递任务中不应包含原始情感词。相反，原始句",
"3.1Text Sentiment Transfer",
"子中的名词应该保留在这样的任务（Melnyketal.2017",
"Data我们使用两个数据集，Yelp餐厅评论和亚马逊",
"Lietal.2018;Johnetal.2019）中。因此，我们可以将abow",
"设置为仅包含&中的所有名词。此外，我们可以将bow",
"产品评论(HeandMcAuley2016)2，它们在以前的作品",
"中也很常用(Shen et al.2017;Fu et al.2018;Li et al.2018;",
"设置为包含一些所需的特定单词，以实现对目标句子的",
"Prabhumoyeetal.2018）。按照他们的实验设置，我们使",
"更细粒度控制",
"用相同的预处理步骤和类似的实验配置",
"将它们放在一起，最终的联合训练损失C如下：",
"k",
"C= CVAE + AbCBOW + >CAlur,sg,",
"Metrics",
"良好的样式转换（Lietal.2018;Prabhumoyeet",
"al.2018）有三个条件。具体而言，生成的句子应：1）具",
"(11)",
"j=1",
"有所需的属性；2）流利；3）尽可能保留原始句子的与",
"其中入和入。是平衡超参数。应该注意的是，CBOW和",
"属性无关的内容。对于第一个和第二个标准，我们遵循",
"CAttr,s，也充当正则化，防止编码器陷入KL消失状态",
"以前的工作(Shen et al.2017; Fu et al.2018; Li et al.2018;",
"(Bowman et al. 2016; Kingma et al. 2016: Yang et al. 2017:",
"Prabhumoyeet al.2018）使用基于模型的评估。我们根",
"Shen et al. 2018; Alemi et al. 2018)",
"据预先训练的双向LSTM分类器（SchusterandPaliwal",
"1997）的预测来衡量风格是否成功转移，并通过预先",
"2.2Text Style Transfer",
"训练的语言模型生成的句子的困惑度（PPL）来衡量",
"给定原始句子E，在连续空间中进行样式传递的推理过",
"语言质量。在先前的工作之后，我们使用在相应数据",
"程。我们通过基于梯度的优化将其表示之修改如下：",
"集上训练的TrigramKneser-Ney平滑语言模型（Kneser",
"and Ney1995）。由于难以衡量内容的保存性，因此我",
"k",
"= 2-n(VClur,s + 入V,CBow),",
"们遵循以前的工作并报告了两个指标：1）单词重叠",
"(12)",
"它计算了原始句子&和生成的句子的单字单词重叠",
"j=1",
"率，由count(waUw)",
"unt(wenw",
"计算；2）正如（Melnyketal.2017;",
"其中是步长，入。是权衡参数，以平衡内容保存和样",
"Lietal.2018)中所指出的那样，句子中的几乎所有名词",
"式传输强度。我们送代这种优化以找到*，直到属性",
"都是与属性无关的内容，应保留在样式转换任务中（。",
"预测器p(s;lz）的输出置信度得分大于阈值β或达到最",
"g。，由POS标记器检测到）出现在生成的句子中的原始",
"大回合数T。通过使用波束搜索(OchandNey2004)解",
"句子中（表示为名词%)。(Lietal.2018）中有1000个带",
"码*来获得目标*。示例过程如图1所示",
"有人类注释的句子作为转移句子的基本事实。我们还将",
"它们作为参考，并报告bi-gramBLEU分数（Papineniet",
"3Experiments",
"al.2002)",
"在本节中，我们在情绪传递和性别风格传递任务的三个",
"公开可用数据集上评估了所提出的方法。然后，我们对",
"2These datasets can be download at http://bit.ly/2LHMUsl."
] |
{
"arxiv_id": "1905.12304",
"language": "zh",
"page_idx": 3
}
|
||
1905.12304-zh-4
|
1905.12304-zh-4.jpg
|
<table>
<tbody>
<tr>
<th>$\overline{\mathrm{Methods}}$</th>
<th>Accuracy 个</th>
<th>$PPL\downarrow$</th>
<th>Overlap</th>
<th>Noun $\%\uparrow$</th>
<th>$\mathbf{BLEU}$</th>
</tr>
<tr>
<td>Original</td>
<td>$\overline{0.1}$</td>
<td>$\overline{22.9}$</td>
<td>100.0</td>
<td>100.0</td>
<td>$\overline{42.4}$</td>
</tr>
<tr>
<td>Human</td>
<td>91.8</td>
<td>76.9</td>
<td>47.2</td>
<td>78.5</td>
<td>100.0</td>
</tr>
<tr>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>TemplateBased</td>
<td>81.3</td>
<td>183.6</td>
<td>55.6</td>
<td>83.3</td>
<td>28.9</td>
</tr>
<tr>
<td>DeleteOnly</td>
<td>85.8</td>
<td>81.4</td>
<td>49.5</td>
<td>74.9</td>
<td>24.7</td>
</tr>
<tr>
<td>etrieve $||a|ata$ mr $1k$</td>
<td>89.5</td>
<td>96.1</td>
<td>49.4</td>
<td>74.0</td>
<td>24.9</td>
</tr>
<tr>
<td>RetrievalOnly</td>
<td>98.4</td>
<td>25.7</td>
<td>15.8</td>
<td>$\underline{39.6}$</td>
<td>$\underline{4.7}$</td>
</tr>
<tr>
<td>StyleEmbedding (Fu et al. 2018)</td>
<td>$\underline{7.2}$</td>
<td>93.9</td>
<td>75.4</td>
<td>74.2</td>
<td>31.9</td>
</tr>
<tr>
<td>MultiDe al. 2018 Ma lHi</td>
<td>48.8</td>
<td>166.5</td>
<td>51.5</td>
<td>52.2</td>
<td>23.1</td>
</tr>
<tr>
<td>BTS ( Prabhumove et al. 2018) </td>
<td>94.8</td>
<td>32.8</td>
<td>21.5</td>
<td>23.5</td>
<td>6.8</td>
</tr>
<tr>
<td>al. 2017 CrossAlio INh</td>
<td>$\underline{73.6}$</td>
<td>72.0</td>
<td>41.1</td>
<td>42.9</td>
<td>18.4</td>
</tr>
<tr>
<td>trengthen) $()_{\mathrm{urs}}$</td>
<td>88.2</td>
<td>26.5</td>
<td>46.6</td>
<td>777.4</td>
<td>21.8</td>
</tr>
<tr>
<td>0urs balance) 1cft</td>
<td>92.3</td>
<td>18.3</td>
<td>38.9</td>
<td>69.3</td>
<td>18.8</td>
</tr>
<tr>
<td>Ours (style-strengthen)</td>
<td>95.7</td>
<td>20.6</td>
<td>39.7</td>
<td>61.5</td>
<td>17.9</td>
</tr>
<tr>
<td>Methods</td>
<td>Accuracy 个</td>
<td>$PPL\downarrow$</td>
<td>$\overline{\mathrm{Overlap}}$</td>
<td>Noun $\%\uparrow$</td>
<td>$\mathbf{BLEU}$</td>
</tr>
<tr>
<td>${\mathrm{Methods}}$ $\lim_{n\to\infty}$</td>
<td>Accuracy 72</td>
<td>$PPL$</td>
<td>$f(x)$</td>
<td>$\mathbf{NOun}$ 70 10</td>
<td>BLEU</td>
</tr>
<tr>
<td>$\overline{{\mathrm{Original}}}$</td>
<td>$\overline{23.4}$</td>
<td>$\overline{24.4}$</td>
<td>100.0</td>
<td>100.0</td>
<td>$\overline{57.2}$</td>
</tr>
<tr>
<td>${\mathrm{Human}}$</td>
<td>88.1</td>
<td>62.9</td>
<td>60.5</td>
<td>85.0</td>
<td>100.0</td>
</tr>
<tr>
<td>Delete, Retrieve, & Generate (Li et al. 2018</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>TemplateBased</td>
<td>69.6</td>
<td>108.9</td>
<td>73.3</td>
<td>87.9</td>
<td>42.8</td>
</tr>
<tr>
<td>DeleteOnly</td>
<td>$\underline{51.6}$</td>
<td>49.3</td>
<td>74.4</td>
<td>95.1</td>
<td>44.7</td>
</tr>
<tr>
<td>DeleteA ${\mathrm{AndRetrieve}}$</td>
<td>$\underline{55.2}$</td>
<td>48.2</td>
<td>69.1</td>
<td>92.6</td>
<td>41.8</td>
</tr>
<tr>
<td>RetrievalOnly</td>
<td>87.2</td>
<td>28.7</td>
<td>21.0</td>
<td>44.5</td>
<td>$\underline{6.7}$</td>
</tr>
<tr>
<td>${\mathrm{t~al.2018}}$ $Nf$ vleEmhe $1di$ HIT</td>
<td>$\overline{40.5}$</td>
<td>$\overline{87.7}$</td>
<td>$\overline{42.2}$</td>
<td>$\overline{41.8}$</td>
<td>$\overline{22.1}$</td>
</tr>
<tr>
<td>MultiDe al. 2018 $m_{\mathrm{P}}$ lHi</td>
<td>$\underline{66.5}$</td>
<td>80.8</td>
<td>30.6</td>
<td>30.4</td>
<td>14.3</td>
</tr>
<tr>
<td>al. 2018 $\mathbf{R}^{\prime}$ $|Prahh$</td>
<td>82.6</td>
<td>25.3</td>
<td>24.7</td>
<td>22.5</td>
<td>9.2</td>
</tr>
<tr>
<td>CrossAligned (Shen et al. 2017)</td>
<td>$\underline{69.6}$</td>
<td>18.3</td>
<td>$\underline{19.3}$</td>
<td>$\underline{20.4}$</td>
<td>$\underline{5.0}$</td>
</tr>
<tr>
<td>Ours -strengthen)</td>
<td>81.9</td>
<td>35.0</td>
<td>37.7</td>
<td>76.0</td>
<td>11.5</td>
</tr>
<tr>
<td>0urs balance) I CTI</td>
<td>85.1</td>
<td>21.8</td>
<td>49.3</td>
<td>49.8</td>
<td>21.5</td>
</tr>
<tr>
<td>Ours (stvle -strengthen) $\bullet_{-1}$</td>
<td>90.0</td>
<td>15.9</td>
<td>39.5</td>
<td>41.4</td>
<td>16.3</td>
</tr>
</tbody>
</table>
Table 1:Evaluation results of the sentiment transfer tasks on Yelp (Top) and Amazon (Bottom). The notation $\uparrow$ means the higher the better, while $\downarrow$ means the lower the better. For our models, we report different results (denoted as Ours (content-strengthen), Ours (style-content balance), and Ours (style-strengthen)) corresponding to different choices of hyper-parameters( $\lambda_\mathrm{c}$ and $\beta$ ), which demonstrates our models’ ability to control the trade-off between attribute transfer and content preservation. For each evaluation criterion, we bold the best values (except for Human and Original). The accuracies of the classifier on the test set of Yelp and Amazon are $98.2\%$ and $84.0\%$ .Note that a good model should perform well on all metrics, we further highlight the metrics where the performances of the models are poor with underline.
Baselines 我们将我们的方法与以前的几种最先进的方法(Shen et al. 2017; Fu et al. 2018; Li et al. 2018; Prabhumoye et al. 2018) 进行了比较。我们报告的结果的人写的句子作为一个强大的基线。还报告了未对原始句子进行任何更改的结果(表示为原始句子)。附录 A 中分析了使用不同超参数的效果以及我们方法的消融研究。
Results 表 1 显示了两个数据集上的评估结果。应该注意的是，一个好的风格转移方法应该在我们上面讨论的所有指标上都表现良好。如果我们只使用原始句子作为输出而不进行任何修改，我们仍然可以在语言流利度(PPL)和内容保留(Overlap,名词%)上获得良好的性能，如表 1 的第一行所示。因此，我们用下划线突出显示模型性能较差的指标。我们发现 StyleEmbedding 和 MultiDecoder 实现了高内容保留(Overlap, BLEU 和 Noun%),但它们的流畅性(PPL)和传输准确性明显比我们的方法差。虽然 CrossAligned 的流畅性比 StyleEmbedding 和 MultiDecoder 更好，但它在内容保存和情感传递方面表现不佳。相反，BST 实现了高流畅性和传输准确性，而内容保存较差。在这两个任务的所
有指标上，我们的(style-enhance) 的表现都优于 BST 和
CrossAligned
由于(Liet al. 2018)( RetrievalOnly 除外) 中提出的方法是基于先验知识的，该方法直接在离散空间中修改原始句子中的少数单词，因此可以轻松实现高内容保留但不能保证流畅性和准确性。如结果所示，与我们的方法相比，它们的流畅性和传输精度都很差。RetrievalOnly 方法将人工书写的句子作为输出进行检索，因此该方法可以实现较高的传输准确性和流畅性，但其内容保留性比我们的方法差。我们的方法在连续的空间中修改原始句子，这在流畅性，内容保存和传输准确性方面表现良好。此外，我们的方法可以控制传输准确性和内容保存之间的权衡
Human Evaluation 我们进行人工评估，以进一步验证我们的方法在两个数据集上的性能。继以前的作品(Li et al. 2018; Fu et al. 2018)之后，我们随机选择 50 个原始句子，并请 7 个评估者 3 评估由不同方法生成的句
$^3$All evaluators have Bachelor or higher degree. They are inde-
pendent of the authors' research group.
|
[
[
0.60693359375,
0.06500244140625,
0.67236328125,
0.06756591796875,
0.67236328125,
0.08209228515625,
0.60693359375,
0.0802001953125
],
[
0.68310546875,
0.06500244140625,
0.7490234375,
0.06756591796875,
0.7490234375,
0.08209228515625,
0.68310546875,
0.0802001953125
],
[
0.7646484375,
0.06439208984375,
0.82421875,
0.0662841796875,
0.82275390625,
0.080810546875,
0.76318359375,
0.07891845703125
],
[
0.465576171875,
0.06756591796875,
0.53369140625,
0.06756591796875,
0.53369140625,
0.07891845703125,
0.465576171875,
0.07891845703125
],
[
0.1732177734375,
0.0662841796875,
0.23779296875,
0.0662841796875,
0.23779296875,
0.080810546875,
0.1732177734375,
0.080810546875
],
[
0.5537109375,
0.06756591796875,
0.59228515625,
0.06756591796875,
0.59228515625,
0.07891845703125,
0.5537109375,
0.07891845703125
],
[
0.1748046875,
0.07763671875,
0.2344970703125,
0.0802001953125,
0.233642578125,
0.0947265625,
0.1732177734375,
0.0921630859375
],
[
0.48779296875,
0.0802001953125,
0.515625,
0.0802001953125,
0.515625,
0.0947265625,
0.48779296875,
0.0947265625
],
[
0.5556640625,
0.07574462890625,
0.5908203125,
0.07891845703125,
0.58837890625,
0.09344482421875,
0.55224609375,
0.09027099609375
],
[
0.62158203125,
0.0802001953125,
0.66015625,
0.0802001953125,
0.66015625,
0.09344482421875,
0.62158203125,
0.09344482421875
],
[
0.701171875,
0.0802001953125,
0.73876953125,
0.0802001953125,
0.73876953125,
0.09344482421875,
0.701171875,
0.09344482421875
],
[
0.77685546875,
0.0802001953125,
0.80615234375,
0.0802001953125,
0.80615234375,
0.0921630859375,
0.77685546875,
0.0921630859375
],
[
0.17724609375,
0.09027099609375,
0.2279052734375,
0.09344482421875,
0.226318359375,
0.10797119140625,
0.176513671875,
0.10479736328125
],
[
0.484375,
0.0889892578125,
0.5185546875,
0.0921630859375,
0.515625,
0.106689453125,
0.48193359375,
0.10418701171875
],
[
0.55810546875,
0.0921630859375,
0.58837890625,
0.0921630859375,
0.58837890625,
0.10479736328125,
0.55810546875,
0.10479736328125
],
[
0.62353515625,
0.091552734375,
0.65625,
0.091552734375,
0.65625,
0.106689453125,
0.62353515625,
0.106689453125
],
[
0.70361328125,
0.0921630859375,
0.73388671875,
0.0921630859375,
0.73388671875,
0.10479736328125,
0.70361328125,
0.10479736328125
],
[
0.775390625,
0.0921630859375,
0.810546875,
0.0921630859375,
0.810546875,
0.10479736328125,
0.775390625,
0.10479736328125
],
[
0.1732177734375,
0.10418701171875,
0.44775390625,
0.10479736328125,
0.44775390625,
0.11932373046875,
0.1732177734375,
0.11871337890625
],
[
0.486083984375,
0.11871337890625,
0.51708984375,
0.11871337890625,
0.51708984375,
0.1307373046875,
0.486083984375,
0.1307373046875
],
[
0.5556640625,
0.11871337890625,
0.5908203125,
0.11871337890625,
0.5908203125,
0.1307373046875,
0.5556640625,
0.1307373046875
],
[
0.625,
0.11871337890625,
0.65625,
0.11871337890625,
0.65625,
0.1307373046875,
0.625,
0.1307373046875
],
[
0.70166015625,
0.11871337890625,
0.73388671875,
0.11871337890625,
0.73388671875,
0.1307373046875,
0.70166015625,
0.1307373046875
],
[
0.77685546875,
0.11871337890625,
0.80810546875,
0.11871337890625,
0.80810546875,
0.1307373046875,
0.77685546875,
0.1307373046875
],
[
0.701171875,
0.1287841796875,
0.734375,
0.1287841796875,
0.734375,
0.14453125,
0.701171875,
0.14453125
],
[
0.181396484375,
0.11932373046875,
0.273681640625,
0.11932373046875,
0.273681640625,
0.1307373046875,
0.181396484375,
0.1307373046875
],
[
0.5556640625,
0.1300048828125,
0.5888671875,
0.1300048828125,
0.5888671875,
0.1458740234375,
0.5556640625,
0.1458740234375
],
[
0.775390625,
0.1300048828125,
0.80810546875,
0.1300048828125,
0.80810546875,
0.1458740234375,
0.775390625,
0.1458740234375
],
[
0.625,
0.1307373046875,
0.654296875,
0.1307373046875,
0.654296875,
0.143310546875,
0.625,
0.143310546875
],
[
0.181396484375,
0.1319580078125,
0.250732421875,
0.1319580078125,
0.250732421875,
0.143310546875,
0.181396484375,
0.143310546875
],
[
0.486083984375,
0.1307373046875,
0.51708984375,
0.1307373046875,
0.51708984375,
0.143310546875,
0.486083984375,
0.143310546875
],
[
0.55615234375,
0.14013671875,
0.58642578125,
0.1427001953125,
0.5849609375,
0.1572265625,
0.5556640625,
0.1546630859375
],
[
0.62353515625,
0.14013671875,
0.65673828125,
0.143310546875,
0.654296875,
0.1578369140625,
0.62109375,
0.1546630859375
],
[
0.486083984375,
0.143310546875,
0.51708984375,
0.143310546875,
0.51708984375,
0.1558837890625,
0.486083984375,
0.1558837890625
],
[
0.178955078125,
0.141357421875,
0.29736328125,
0.143310546875,
0.29736328125,
0.1578369140625,
0.178955078125,
0.1558837890625
],
[
0.70361328125,
0.143310546875,
0.73388671875,
0.143310546875,
0.73388671875,
0.1558837890625,
0.70361328125,
0.1558837890625
],
[
0.77685546875,
0.143310546875,
0.80615234375,
0.143310546875,
0.80615234375,
0.1558837890625,
0.77685546875,
0.1558837890625
],
[
0.625,
0.1546630859375,
0.65625,
0.1546630859375,
0.65625,
0.17041015625,
0.625,
0.17041015625
],
[
0.70361328125,
0.154052734375,
0.73388671875,
0.1558837890625,
0.73193359375,
0.169189453125,
0.70166015625,
0.1673583984375
],
[
0.55810546875,
0.1558837890625,
0.58642578125,
0.1558837890625,
0.58642578125,
0.1685791015625,
0.55810546875,
0.1685791015625
],
[
0.486083984375,
0.1558837890625,
0.515625,
0.1558837890625,
0.515625,
0.1685791015625,
0.486083984375,
0.1685791015625
],
[
0.779296875,
0.1558837890625,
0.8037109375,
0.1558837890625,
0.8037109375,
0.169189453125,
0.779296875,
0.169189453125
],
[
0.705078125,
0.1673583984375,
0.73193359375,
0.169189453125,
0.73046875,
0.18310546875,
0.70361328125,
0.179931640625
],
[
0.181396484375,
0.1572265625,
0.2646484375,
0.1572265625,
0.2646484375,
0.1685791015625,
0.181396484375,
0.1685791015625
],
[
0.178955078125,
0.169189453125,
0.375732421875,
0.169189453125,
0.375732421875,
0.18310546875,
0.178955078125,
0.18310546875
],
[
0.490234375,
0.169189453125,
0.51318359375,
0.169189453125,
0.51318359375,
0.18310546875,
0.490234375,
0.18310546875
],
[
0.55810546875,
0.169189453125,
0.58837890625,
0.169189453125,
0.58837890625,
0.1817626953125,
0.55810546875,
0.1817626953125
],
[
0.625,
0.1685791015625,
0.65625,
0.1685791015625,
0.65625,
0.184326171875,
0.625,
0.184326171875
],
[
0.77685546875,
0.169189453125,
0.80810546875,
0.169189453125,
0.80810546875,
0.1817626953125,
0.77685546875,
0.1817626953125
],
[
0.178955078125,
0.179931640625,
0.359375,
0.179931640625,
0.359375,
0.1932373046875,
0.178955078125,
0.1932373046875
],
[
0.625,
0.1787109375,
0.65283203125,
0.18115234375,
0.65185546875,
0.1944580078125,
0.62353515625,
0.1925048828125
],
[
0.70166015625,
0.179931640625,
0.73388671875,
0.179931640625,
0.73388671875,
0.1956787109375,
0.70166015625,
0.1956787109375
],
[
0.77685546875,
0.179931640625,
0.80615234375,
0.1817626953125,
0.8056640625,
0.1956787109375,
0.775390625,
0.1932373046875
],
[
0.55615234375,
0.1817626953125,
0.5908203125,
0.1817626953125,
0.5908203125,
0.1944580078125,
0.55615234375,
0.1944580078125
],
[
0.48779296875,
0.1817626953125,
0.515625,
0.1817626953125,
0.515625,
0.1944580078125,
0.48779296875,
0.1944580078125
],
[
0.484375,
0.1912841796875,
0.51708984375,
0.1932373046875,
0.5146484375,
0.208984375,
0.48193359375,
0.205810546875
],
[
0.5556640625,
0.1925048828125,
0.5888671875,
0.1925048828125,
0.5888671875,
0.2083740234375,
0.5556640625,
0.2083740234375
],
[
0.62353515625,
0.1925048828125,
0.65625,
0.1925048828125,
0.65625,
0.2083740234375,
0.62353515625,
0.2083740234375
],
[
0.70166015625,
0.1925048828125,
0.73388671875,
0.1925048828125,
0.73388671875,
0.208984375,
0.70166015625,
0.208984375
],
[
0.779296875,
0.1944580078125,
0.8037109375,
0.1944580078125,
0.8037109375,
0.20703125,
0.779296875,
0.20703125
],
[
0.55615234375,
0.2052001953125,
0.5888671875,
0.2052001953125,
0.5888671875,
0.2203369140625,
0.55615234375,
0.2203369140625
],
[
0.1805419921875,
0.1932373046875,
0.363525390625,
0.1932373046875,
0.363525390625,
0.20703125,
0.1805419921875,
0.20703125
],
[
0.178955078125,
0.20703125,
0.371826171875,
0.20703125,
0.371826171875,
0.2203369140625,
0.178955078125,
0.2203369140625
],
[
0.48779296875,
0.20703125,
0.515625,
0.20703125,
0.515625,
0.2197265625,
0.48779296875,
0.2197265625
],
[
0.625,
0.20703125,
0.65283203125,
0.20703125,
0.65283203125,
0.2197265625,
0.625,
0.2197265625
],
[
0.70166015625,
0.2052001953125,
0.73388671875,
0.2052001953125,
0.73388671875,
0.2203369140625,
0.70166015625,
0.2203369140625
],
[
0.77783203125,
0.20703125,
0.80615234375,
0.20703125,
0.80615234375,
0.2197265625,
0.77783203125,
0.2197265625
],
[
0.486083984375,
0.2203369140625,
0.515625,
0.2203369140625,
0.515625,
0.23291015625,
0.486083984375,
0.23291015625
],
[
0.55810546875,
0.2203369140625,
0.58642578125,
0.2203369140625,
0.58642578125,
0.23291015625,
0.55810546875,
0.23291015625
],
[
0.625,
0.2203369140625,
0.654296875,
0.2203369140625,
0.654296875,
0.23291015625,
0.625,
0.23291015625
],
[
0.70361328125,
0.2203369140625,
0.73193359375,
0.2203369140625,
0.73193359375,
0.23291015625,
0.70361328125,
0.23291015625
],
[
0.77685546875,
0.2203369140625,
0.80810546875,
0.2203369140625,
0.80810546875,
0.23291015625,
0.77685546875,
0.23291015625
],
[
0.484375,
0.2291259765625,
0.51708984375,
0.231689453125,
0.5146484375,
0.2474365234375,
0.48193359375,
0.2442626953125
],
[
0.1805419921875,
0.2215576171875,
0.334228515625,
0.2215576171875,
0.334228515625,
0.231689453125,
0.1805419921875,
0.231689453125
],
[
0.1805419921875,
0.231689453125,
0.3505859375,
0.23291015625,
0.3505859375,
0.246826171875,
0.1805419921875,
0.24560546875
],
[
0.55810546875,
0.23291015625,
0.58837890625,
0.23291015625,
0.58837890625,
0.24560546875,
0.55810546875,
0.24560546875
],
[
0.62353515625,
0.2298583984375,
0.65673828125,
0.23291015625,
0.654296875,
0.2474365234375,
0.62109375,
0.2442626953125
],
[
0.70166015625,
0.23291015625,
0.73388671875,
0.23291015625,
0.73388671875,
0.24560546875,
0.70166015625,
0.24560546875
],
[
0.77783203125,
0.23291015625,
0.80810546875,
0.23291015625,
0.80810546875,
0.24560546875,
0.77783203125,
0.24560546875
],
[
0.1805419921875,
0.2442626953125,
0.318603515625,
0.24560546875,
0.318603515625,
0.2587890625,
0.1805419921875,
0.25830078125
],
[
0.484375,
0.2442626953125,
0.51708984375,
0.2442626953125,
0.51708984375,
0.260009765625,
0.484375,
0.260009765625
],
[
0.55615234375,
0.2442626953125,
0.5888671875,
0.2442626953125,
0.5888671875,
0.260009765625,
0.55615234375,
0.260009765625
],
[
0.62353515625,
0.2442626953125,
0.65625,
0.2442626953125,
0.65625,
0.260009765625,
0.62353515625,
0.260009765625
],
[
0.701171875,
0.2442626953125,
0.73388671875,
0.2442626953125,
0.73388671875,
0.260009765625,
0.701171875,
0.260009765625
],
[
0.779296875,
0.24560546875,
0.80810546875,
0.24560546875,
0.80810546875,
0.260009765625,
0.779296875,
0.260009765625
],
[
0.176513671875,
0.25830078125,
0.236083984375,
0.260009765625,
0.2344970703125,
0.274658203125,
0.1748046875,
0.272705078125
],
[
0.5537109375,
0.2587890625,
0.5908203125,
0.261474609375,
0.5888671875,
0.2734375,
0.55224609375,
0.271484375
],
[
0.7646484375,
0.25830078125,
0.82275390625,
0.260009765625,
0.82177734375,
0.274658203125,
0.76318359375,
0.272705078125
],
[
0.609375,
0.261962890625,
0.6689453125,
0.261962890625,
0.6689453125,
0.2734375,
0.609375,
0.2734375
],
[
0.464111328125,
0.260009765625,
0.5361328125,
0.261962890625,
0.5361328125,
0.276611328125,
0.464111328125,
0.274658203125
],
[
0.685546875,
0.261962890625,
0.74658203125,
0.261962890625,
0.74658203125,
0.2734375,
0.685546875,
0.2734375
],
[
0.486083984375,
0.272705078125,
0.51708984375,
0.272705078125,
0.51708984375,
0.28857421875,
0.486083984375,
0.28857421875
],
[
0.55615234375,
0.272705078125,
0.5888671875,
0.272705078125,
0.5888671875,
0.28857421875,
0.55615234375,
0.28857421875
],
[
0.62353515625,
0.2734375,
0.65673828125,
0.2734375,
0.65673828125,
0.285888671875,
0.62353515625,
0.285888671875
],
[
0.701171875,
0.2734375,
0.73876953125,
0.2734375,
0.73876953125,
0.287353515625,
0.701171875,
0.287353515625
],
[
0.775390625,
0.272705078125,
0.8095703125,
0.272705078125,
0.8095703125,
0.28857421875,
0.775390625,
0.28857421875
],
[
0.486083984375,
0.285400390625,
0.515625,
0.287353515625,
0.5146484375,
0.2998046875,
0.484375,
0.29736328125
],
[
0.17724609375,
0.274658203125,
0.233642578125,
0.274658203125,
0.233642578125,
0.285888671875,
0.17724609375,
0.285888671875
],
[
0.55615234375,
0.285888671875,
0.58837890625,
0.285888671875,
0.58837890625,
0.298583984375,
0.55615234375,
0.298583984375
],
[
0.62353515625,
0.285400390625,
0.65625,
0.285400390625,
0.65625,
0.300537109375,
0.62353515625,
0.300537109375
],
[
0.70166015625,
0.285888671875,
0.73388671875,
0.285888671875,
0.73388671875,
0.298583984375,
0.70166015625,
0.298583984375
],
[
0.775390625,
0.285888671875,
0.810546875,
0.285888671875,
0.810546875,
0.298583984375,
0.775390625,
0.298583984375
],
[
0.178955078125,
0.2998046875,
0.44775390625,
0.2998046875,
0.44775390625,
0.313232421875,
0.178955078125,
0.313232421875
],
[
0.1805419921875,
0.287353515625,
0.2279052734375,
0.287353515625,
0.2279052734375,
0.298583984375,
0.1805419921875,
0.298583984375
],
[
0.48779296875,
0.311279296875,
0.515625,
0.313232421875,
0.5146484375,
0.326904296875,
0.486083984375,
0.324462890625
],
[
0.625,
0.3125,
0.654296875,
0.3125,
0.654296875,
0.324462890625,
0.625,
0.324462890625
],
[
0.70361328125,
0.3125,
0.73388671875,
0.3125,
0.73388671875,
0.324462890625,
0.70361328125,
0.324462890625
],
[
0.77685546875,
0.3125,
0.80615234375,
0.3125,
0.80615234375,
0.324462890625,
0.77685546875,
0.324462890625
],
[
0.5556640625,
0.313232421875,
0.5908203125,
0.313232421875,
0.5908203125,
0.324462890625,
0.5556640625,
0.324462890625
],
[
0.1829833984375,
0.313232421875,
0.273681640625,
0.313232421875,
0.273681640625,
0.324462890625,
0.1829833984375,
0.324462890625
],
[
0.48779296875,
0.322509765625,
0.51953125,
0.32568359375,
0.51708984375,
0.341552734375,
0.484375,
0.33837890625
],
[
0.55615234375,
0.323974609375,
0.5888671875,
0.323974609375,
0.5888671875,
0.339111328125,
0.55615234375,
0.339111328125
],
[
0.70166015625,
0.3212890625,
0.734375,
0.324462890625,
0.73193359375,
0.339111328125,
0.69921875,
0.337158203125
],
[
0.775390625,
0.323974609375,
0.80810546875,
0.323974609375,
0.80810546875,
0.339111328125,
0.775390625,
0.339111328125
],
[
0.625,
0.324462890625,
0.654296875,
0.324462890625,
0.654296875,
0.337158203125,
0.625,
0.337158203125
],
[
0.181396484375,
0.32568359375,
0.250732421875,
0.32568359375,
0.250732421875,
0.337158203125,
0.181396484375,
0.337158203125
],
[
0.484375,
0.3359375,
0.51708984375,
0.3359375,
0.51708984375,
0.352783203125,
0.484375,
0.352783203125
],
[
0.55810546875,
0.335205078125,
0.5888671875,
0.337158203125,
0.58642578125,
0.3515625,
0.5556640625,
0.349853515625
],
[
0.70361328125,
0.3359375,
0.73388671875,
0.33837890625,
0.73193359375,
0.35107421875,
0.70166015625,
0.348388671875
],
[
0.77685546875,
0.3359375,
0.80615234375,
0.33837890625,
0.8056640625,
0.3515625,
0.775390625,
0.349853515625
],
[
0.625,
0.33837890625,
0.654296875,
0.33837890625,
0.654296875,
0.35107421875,
0.625,
0.35107421875
],
[
0.1805419921875,
0.337158203125,
0.29736328125,
0.337158203125,
0.29736328125,
0.35107421875,
0.1805419921875,
0.35107421875
],
[
0.486083984375,
0.348388671875,
0.51708984375,
0.348388671875,
0.51708984375,
0.3642578125,
0.486083984375,
0.3642578125
],
[
0.55615234375,
0.348388671875,
0.58837890625,
0.348388671875,
0.58837890625,
0.3642578125,
0.55615234375,
0.3642578125
],
[
0.625,
0.348388671875,
0.65625,
0.348388671875,
0.65625,
0.3642578125,
0.625,
0.3642578125
],
[
0.70361328125,
0.347900390625,
0.73193359375,
0.349853515625,
0.73046875,
0.363037109375,
0.70166015625,
0.361083984375
],
[
0.779296875,
0.349853515625,
0.8037109375,
0.349853515625,
0.8037109375,
0.363037109375,
0.779296875,
0.363037109375
],
[
0.176513671875,
0.3623046875,
0.375732421875,
0.3623046875,
0.375732421875,
0.375732421875,
0.176513671875,
0.375732421875
],
[
0.178955078125,
0.349853515625,
0.265625,
0.349853515625,
0.265625,
0.3642578125,
0.178955078125,
0.3642578125
],
[
0.486083984375,
0.363037109375,
0.515625,
0.363037109375,
0.515625,
0.375732421875,
0.486083984375,
0.375732421875
],
[
0.55615234375,
0.363037109375,
0.58837890625,
0.363037109375,
0.58837890625,
0.375732421875,
0.55615234375,
0.375732421875
],
[
0.625,
0.363037109375,
0.654296875,
0.363037109375,
0.654296875,
0.375,
0.625,
0.375
],
[
0.70166015625,
0.363037109375,
0.73388671875,
0.363037109375,
0.73388671875,
0.375732421875,
0.70166015625,
0.375732421875
],
[
0.77685546875,
0.3623046875,
0.80615234375,
0.3623046875,
0.80615234375,
0.375,
0.77685546875,
0.375
],
[
0.178955078125,
0.373779296875,
0.359375,
0.373779296875,
0.359375,
0.386962890625,
0.178955078125,
0.386962890625
],
[
0.484375,
0.37255859375,
0.5185546875,
0.375732421875,
0.515625,
0.39013671875,
0.48193359375,
0.386962890625
],
[
0.625,
0.373779296875,
0.65625,
0.373779296875,
0.65625,
0.389404296875,
0.625,
0.389404296875
],
[
0.77685546875,
0.373779296875,
0.80810546875,
0.373779296875,
0.80810546875,
0.389404296875,
0.77685546875,
0.389404296875
],
[
0.70361328125,
0.375732421875,
0.73388671875,
0.375732421875,
0.73388671875,
0.38818359375,
0.70361328125,
0.38818359375
],
[
0.55810546875,
0.375732421875,
0.58642578125,
0.375732421875,
0.58642578125,
0.38818359375,
0.55810546875,
0.38818359375
],
[
0.48779296875,
0.386474609375,
0.515625,
0.38818359375,
0.5146484375,
0.401611328125,
0.486083984375,
0.399658203125
],
[
0.55615234375,
0.386474609375,
0.5888671875,
0.386474609375,
0.5888671875,
0.401611328125,
0.55615234375,
0.401611328125
],
[
0.6259765625,
0.3837890625,
0.65673828125,
0.386962890625,
0.65283203125,
0.404052734375,
0.62158203125,
0.40087890625
],
[
0.70361328125,
0.386474609375,
0.734375,
0.386474609375,
0.734375,
0.40283203125,
0.70361328125,
0.40283203125
],
[
0.779296875,
0.386474609375,
0.8056640625,
0.386474609375,
0.8056640625,
0.404052734375,
0.779296875,
0.404052734375
],
[
0.70166015625,
0.396484375,
0.734375,
0.396484375,
0.734375,
0.4140625,
0.70166015625,
0.4140625
],
[
0.1805419921875,
0.38818359375,
0.365234375,
0.38818359375,
0.365234375,
0.401611328125,
0.1805419921875,
0.401611328125
],
[
0.486083984375,
0.399658203125,
0.51708984375,
0.399658203125,
0.51708984375,
0.415283203125,
0.486083984375,
0.415283203125
],
[
0.55810546875,
0.399658203125,
0.5888671875,
0.399658203125,
0.5888671875,
0.415283203125,
0.55810546875,
0.415283203125
],
[
0.625,
0.399658203125,
0.65625,
0.399658203125,
0.65625,
0.415283203125,
0.625,
0.415283203125
],
[
0.78125,
0.399658203125,
0.8037109375,
0.399658203125,
0.8037109375,
0.4140625,
0.78125,
0.4140625
],
[
0.17236328125,
0.412353515625,
0.336669921875,
0.4140625,
0.336669921875,
0.427978515625,
0.17236328125,
0.425537109375
],
[
0.178955078125,
0.40087890625,
0.371826171875,
0.40087890625,
0.371826171875,
0.4140625,
0.178955078125,
0.4140625
],
[
0.486083984375,
0.4140625,
0.515625,
0.4140625,
0.515625,
0.425537109375,
0.486083984375,
0.425537109375
],
[
0.55810546875,
0.4140625,
0.58837890625,
0.4140625,
0.58837890625,
0.4267578125,
0.55810546875,
0.4267578125
],
[
0.6259765625,
0.4140625,
0.65283203125,
0.4140625,
0.65283203125,
0.4267578125,
0.6259765625,
0.4267578125
],
[
0.70361328125,
0.4140625,
0.73388671875,
0.4140625,
0.73388671875,
0.4267578125,
0.70361328125,
0.4267578125
],
[
0.77783203125,
0.4140625,
0.80810546875,
0.4140625,
0.80810546875,
0.4267578125,
0.77783203125,
0.4267578125
],
[
0.484375,
0.4248046875,
0.51708984375,
0.4248046875,
0.51708984375,
0.43994140625,
0.484375,
0.43994140625
],
[
0.55615234375,
0.4248046875,
0.5888671875,
0.4248046875,
0.5888671875,
0.43994140625,
0.55615234375,
0.43994140625
],
[
0.62353515625,
0.423583984375,
0.65673828125,
0.4267578125,
0.654296875,
0.44140625,
0.62109375,
0.438232421875
],
[
0.70361328125,
0.4248046875,
0.73388671875,
0.4267578125,
0.73193359375,
0.44140625,
0.70166015625,
0.439453125
],
[
0.775390625,
0.4248046875,
0.80810546875,
0.4248046875,
0.80810546875,
0.43994140625,
0.775390625,
0.43994140625
],
[
0.486083984375,
0.4375,
0.515625,
0.439453125,
0.5146484375,
0.453857421875,
0.484375,
0.451904296875
],
[
0.1805419921875,
0.4267578125,
0.3505859375,
0.4267578125,
0.3505859375,
0.43994140625,
0.1805419921875,
0.43994140625
],
[
0.625,
0.4375,
0.654296875,
0.439453125,
0.65283203125,
0.45263671875,
0.62353515625,
0.45068359375
],
[
0.701171875,
0.4375,
0.734375,
0.4375,
0.734375,
0.45263671875,
0.701171875,
0.45263671875
],
[
0.77685546875,
0.4375,
0.80810546875,
0.4375,
0.80810546875,
0.45263671875,
0.77685546875,
0.45263671875
],
[
0.55810546875,
0.439453125,
0.58837890625,
0.439453125,
0.58837890625,
0.451904296875,
0.55810546875,
0.451904296875
],
[
0.178955078125,
0.439453125,
0.3203125,
0.439453125,
0.3203125,
0.45263671875,
0.178955078125,
0.45263671875
],
[
0.08660888671875,
0.4716796875,
0.9111328125,
0.4716796875,
0.9111328125,
0.48486328125,
0.08660888671875,
0.48486328125
],
[
0.0849609375,
0.48486328125,
0.9111328125,
0.48486328125,
0.9111328125,
0.498779296875,
0.0849609375,
0.498779296875
],
[
0.08660888671875,
0.5,
0.91259765625,
0.5,
0.91259765625,
0.51318359375,
0.08660888671875,
0.51318359375
],
[
0.08660888671875,
0.51318359375,
0.91259765625,
0.51318359375,
0.91259765625,
0.5263671875,
0.08660888671875,
0.5263671875
],
[
0.0882568359375,
0.52880859375,
0.9111328125,
0.52880859375,
0.9111328125,
0.5390625,
0.0882568359375,
0.5390625
],
[
0.08660888671875,
0.5390625,
0.9111328125,
0.5390625,
0.9111328125,
0.5537109375,
0.08660888671875,
0.5537109375
],
[
0.0849609375,
0.5537109375,
0.55126953125,
0.5537109375,
0.55126953125,
0.5673828125,
0.0849609375,
0.5673828125
],
[
0.09478759765625,
0.5966796875,
0.476318359375,
0.5966796875,
0.476318359375,
0.60693359375,
0.09478759765625,
0.60693359375
],
[
0.52294921875,
0.59521484375,
0.91259765625,
0.59521484375,
0.91259765625,
0.609375,
0.52294921875,
0.609375
],
[
0.08660888671875,
0.60791015625,
0.47802734375,
0.60791015625,
0.47802734375,
0.6220703125,
0.08660888671875,
0.6220703125
],
[
0.52294921875,
0.60986328125,
0.61376953125,
0.60986328125,
0.61376953125,
0.6220703125,
0.52294921875,
0.6220703125
],
[
0.08660888671875,
0.62255859375,
0.47802734375,
0.62255859375,
0.47802734375,
0.63623046875,
0.08660888671875,
0.63623046875
],
[
0.5361328125,
0.62255859375,
0.9111328125,
0.6220703125,
0.9111328125,
0.63623046875,
0.5361328125,
0.63720703125
],
[
0.0882568359375,
0.63525390625,
0.47802734375,
0.63623046875,
0.47802734375,
0.65087890625,
0.0882568359375,
0.6494140625
],
[
0.521484375,
0.63720703125,
0.91259765625,
0.63720703125,
0.91259765625,
0.65087890625,
0.521484375,
0.65087890625
],
[
0.0882568359375,
0.6494140625,
0.47802734375,
0.6494140625,
0.47802734375,
0.6640625,
0.0882568359375,
0.6640625
],
[
0.521484375,
0.65087890625,
0.91259765625,
0.65087890625,
0.91259765625,
0.6640625,
0.521484375,
0.6640625
],
[
0.0882568359375,
0.6640625,
0.47802734375,
0.6640625,
0.47802734375,
0.67822265625,
0.0882568359375,
0.67822265625
],
[
0.521484375,
0.6640625,
0.9111328125,
0.6640625,
0.9111328125,
0.67822265625,
0.521484375,
0.67822265625
],
[
0.08660888671875,
0.67822265625,
0.11273193359375,
0.67822265625,
0.11273193359375,
0.69140625,
0.08660888671875,
0.69140625
],
[
0.521484375,
0.6787109375,
0.91259765625,
0.6787109375,
0.91259765625,
0.6923828125,
0.521484375,
0.6923828125
],
[
0.52294921875,
0.6923828125,
0.91259765625,
0.6923828125,
0.91259765625,
0.70556640625,
0.52294921875,
0.70556640625
],
[
0.0906982421875,
0.708984375,
0.47802734375,
0.708984375,
0.47802734375,
0.72265625,
0.0906982421875,
0.72265625
],
[
0.521484375,
0.70556640625,
0.91259765625,
0.70556640625,
0.91259765625,
0.7197265625,
0.521484375,
0.7197265625
],
[
0.17236328125,
0.7216796875,
0.47802734375,
0.72265625,
0.47802734375,
0.7373046875,
0.17236328125,
0.736328125
],
[
0.52294921875,
0.72021484375,
0.91259765625,
0.72021484375,
0.91259765625,
0.734375,
0.52294921875,
0.734375
],
[
0.521484375,
0.734375,
0.91015625,
0.734375,
0.91015625,
0.74755859375,
0.521484375,
0.74755859375
],
[
0.0882568359375,
0.724609375,
0.1690673828125,
0.724609375,
0.1690673828125,
0.736328125,
0.0882568359375,
0.736328125
],
[
0.08660888671875,
0.7373046875,
0.47802734375,
0.7373046875,
0.47802734375,
0.75048828125,
0.08660888671875,
0.75048828125
],
[
0.0882568359375,
0.75048828125,
0.47802734375,
0.75048828125,
0.47802734375,
0.7646484375,
0.0882568359375,
0.7646484375
],
[
0.51953125,
0.7470703125,
0.9111328125,
0.74755859375,
0.9111328125,
0.76220703125,
0.51953125,
0.76123046875
],
[
0.0882568359375,
0.76513671875,
0.476318359375,
0.76513671875,
0.476318359375,
0.77880859375,
0.0882568359375,
0.77880859375
],
[
0.52294921875,
0.76318359375,
0.60302734375,
0.76318359375,
0.60302734375,
0.7744140625,
0.52294921875,
0.7744140625
],
[
0.08905029296875,
0.77880859375,
0.476318359375,
0.77880859375,
0.476318359375,
0.79248046875,
0.08905029296875,
0.79248046875
],
[
0.0882568359375,
0.79248046875,
0.4794921875,
0.79248046875,
0.4794921875,
0.80615234375,
0.0882568359375,
0.80615234375
],
[
0.525390625,
0.79541015625,
0.91259765625,
0.79541015625,
0.91259765625,
0.8095703125,
0.525390625,
0.8095703125
],
[
0.08660888671875,
0.80615234375,
0.47802734375,
0.80615234375,
0.47802734375,
0.8193359375,
0.08660888671875,
0.8193359375
],
[
0.521484375,
0.81005859375,
0.91259765625,
0.81005859375,
0.91259765625,
0.82373046875,
0.521484375,
0.82373046875
],
[
0.08660888671875,
0.8193359375,
0.4794921875,
0.8193359375,
0.4794921875,
0.83349609375,
0.08660888671875,
0.83349609375
],
[
0.521484375,
0.82373046875,
0.9111328125,
0.82373046875,
0.9111328125,
0.83837890625,
0.521484375,
0.83837890625
],
[
0.0882568359375,
0.833984375,
0.47802734375,
0.833984375,
0.47802734375,
0.84765625,
0.0882568359375,
0.84765625
],
[
0.521484375,
0.83837890625,
0.9140625,
0.83837890625,
0.9140625,
0.8515625,
0.521484375,
0.8515625
],
[
0.0882568359375,
0.84765625,
0.47802734375,
0.84765625,
0.47802734375,
0.861328125,
0.0882568359375,
0.861328125
],
[
0.0882568359375,
0.861328125,
0.47802734375,
0.861328125,
0.47802734375,
0.87548828125,
0.0882568359375,
0.87548828125
],
[
0.5400390625,
0.861328125,
0.9140625,
0.86279296875,
0.9140625,
0.87744140625,
0.5400390625,
0.87548828125
],
[
0.0882568359375,
0.87548828125,
0.47802734375,
0.87548828125,
0.47802734375,
0.8896484375,
0.0882568359375,
0.8896484375
],
[
0.5185546875,
0.87548828125,
0.75244140625,
0.876953125,
0.75244140625,
0.89013671875,
0.5185546875,
0.8896484375
]
] |
[
"Overlap",
"Noun%",
"BLEU个",
"Accuracy",
"Methods",
"PPL",
"Original",
"0.1",
"22.9",
"100.0",
"100.0",
"42.4",
"Human",
"91.8",
"76.9",
"47.2",
"78.5",
"100.0",
"Delete, Retrieve, & Generate (Li et al. 2018):",
"81.3",
"183.6",
"55.6",
"83.3",
"28.9",
"74.9",
"TemplateBased",
"81.4",
"24.7",
"49.5",
"DeleteOnly",
"85.8",
"96.1",
"49.4",
"89.5",
"DeleteAndRetrieve",
"74.0",
"24.9",
"15.8",
"39.6",
"25.7",
"98.4",
"4.7",
"74.2",
"RetrievalOnly",
"StyleEmbedding (Fu et al. 2018)",
"7.2",
"93.9",
"75.4",
"31.9",
"MultiDecoder (Fu et al. 2018)",
"51.5",
"52.2",
"23.1",
"166.5",
"48.8",
"94.8",
"32.8",
"21.5",
"23.5",
"6.8",
"72.0",
"BTS (Prabhumoye et al. 2018)",
"CrossAligned (Shen et al. 2017)",
"73.6",
"41.1",
"42.9",
"18.4",
"88.2",
"26.5",
"46.6",
"77.4",
"21.8",
"92.3",
"Ours (content-strengthen)",
"Ours (style-content balance)",
"18.3",
"38.9",
"69.3",
"18.8",
"Ours (style-strengthen)",
"95.7",
"20.6",
"39.7",
"61.5",
"17.9",
"Methods",
"PPLJ",
"BLEU个",
"Overlap+",
"Accuracy↑",
"Noun % ",
"23.4",
"24.4",
"100.0",
"100.0",
"57.2",
"88.1",
"Original",
"62.9",
"60.5",
"85.0",
"100.0",
"Delete, Retrieve, & Generate (Li et al. 2018):",
"Human",
"69.6",
"73.3",
"87.9",
"42.8",
"108.9",
"TemplateBased",
"51.6",
"49.3",
"95.1",
"44.7",
"74.4",
"DeleteOnly",
"55.2",
"48.2",
"92.6",
"41.8",
"69.1",
"DeleteAndRetrieve",
"87.2",
"28.7",
"21.0",
"44.5",
"6.7",
"StyleEmbedding (Fu et al.2018)",
"RetrievalOnly",
"40.5",
"87.7",
"42.2",
"41.8",
"22.1",
"MultiDecoder (Fu et al. 2018)",
"66.5",
"30.6",
"14.3",
"30.4",
"80.8",
"82.6",
"25.3",
"24.7",
"22.5",
"9.2",
"20.4",
"BTS (Prabhumoye et al. 2018)",
"69.6",
"18.3",
"19.3",
"5.0",
"Ours (content-strengthen)",
"CrossAligned (Shen et al. 2017)",
"81.9",
"35.0",
"37.7",
"76.0",
"11.5",
"85.1",
"21.8",
"49.3",
"49.8",
"21.5",
"90.0",
"Ours (style-content balance)",
"39.5",
"41.4",
"16.3",
"15.9",
"Ours (style-strengthen)",
"Table l: Evaluation results of the sentiment transfer tasks on Yelp (Top) and Amazon (Bottom). The notation means the higher",
"the better, while ↓ means the lower the better. For our models, we report different results (denoted as Ours (content-strengthen),",
"Ours (style-content balance), and Ours (style-strengthen) corresponding to different choices of hyper-parameters ( A。 and β",
"), which demonstrates our models' ability to control the trade-off between attribute transfer and content preservation. For each",
"evaluationcriterion,weboldthebestvalues(exceptforHumanandOriginal).Theaccuraciesoftheclassifieronthetestsetof",
"Yelp and Amazon are 98.2 % and 84.0 % . Note that a good model should perform well on all metrics, we further highlight the",
"metrics where the performances of the models are poor with underline.",
"Baselines我们将我们的方法与以前的几种最先进",
"有指标上，我们的（style-enhance）的表现都优于BST和",
"的方法 (Shen et al. 2017; Fu et al. 2018; Li et al. 2018;",
"CrossAligned",
"Prabhumoyeetal.2018）进行了比较。我们报告的结果的",
"由于（Lietal.2018）（RetrievalOnly除外）中提出的方",
"人写的句子作为一个强大的基线。还报告了未对原始句",
"法是基于先验知识的，该方法直接在离散空间中修改原",
"子进行任何更改的结果（表示为原始句子）。附录A中",
"始句子中的少数单词，因此可以轻松实现高内容保留，",
"分析了使用不同超参数的效果以及我们方法的消融研",
"但不能保证流畅性和准确性。如结果所示，与我们的方",
"究。",
"法相比，它们的流畅性和传输精度都很差。RetrievalOnly",
"方法将人工书写的句子作为输出进行检索，因此该方法",
"Results表1显示了两个数据集上的评估结果。应该",
"可以实现较高的传输准确性和流畅性，但其内容保留性",
"一个好的风格转移方法应该在我们上面讨",
"比我们的方法差。我们的方法在连续的空间中修改原始",
"句子，这在流畅性，内容保存和传输准确性方面表现良",
"注意的是，",
"论的所有指标上都表现良好。如果我们只使用原始句",
"子作为输出而不进行任何修改，我们仍然可以在语言",
"好。此外，我们的方法可以控制传输准确性和内容保存",
"流利度（PPL）和内容保留（Overlap，名词%）上获得",
"之间的权衡",
"良好的性能，如表1的第一行所示。因此，我们用下",
"划线突出显示模型性能较差的指标。我们发现StyleEm-",
"HumanEvaluation我们进行人工评估，以进一步验",
"bedding和MultiDecoder实现了高内容保留（Overlap，",
"证我们的方法在两个数据集上的性能。继以前的作品",
"BLEU和Noun%），但它们的流畅性（PPL）和传输准",
"(Lietal.2018;Fuetal.2018）之后，我们随机选择50个",
"确性明显比我们的方法差。虽然CrossAligned的流畅性",
"原始句子，并请7个评估者3评估由不同方法生成的句",
"比StyleEmbedding和MultiDecoder更好，但它在内容保",
"存和情感传递方面表现不佳。相反，BST实现了高流畅",
"3All evaluators have Bachelor or higher degree. They are inde-",
"性和传输准确性，而内容保存较差。在这两个任务的所",
"pendent of the authors' research group."
] |
{
"arxiv_id": "1905.12304",
"language": "zh",
"page_idx": 4
}
|
||
1906.03674-zh-0
|
1906.03674-zh-0.jpg
|
用于外部知识集成的基于注意力的调节方法
Katerina Margatina$^1,\textbf{Christos Baziotis}^2;$Alexandros Potamianos$^{1,3,4}$
$^{1}$School of ECE, National Technical University of Athens, Athens, Greece
$^{2}$School of Informatics, University of Edinburgh, UK
$^{3}$ Signal Analysis and Interpretation Laboratory (SAIL), USC, Los Angeles, USA
$^{4}$Behavioral Signal Technologies, Los Angeles, USA
el12108@central.ntua.gr, c.baziotis@sms.ed.ac.uk,
potamecentral.ntua.gr
## $\mathbf{Abstract}$
In this paper, we present a novel approach for incorporating external knowledge in Recurrent Neural Networks (RNNs). We propose the integration of lexicon features into the self-attention mechanism of RNN-based architectures. This form of conditioning on the attention distribution, enforces the contribution of the most salient words for the task at hand. We introduce three methods, namely attentional concatenation, feature-based gating and affine transformation. Experiments on six benchmark datasets show the effectiveness of our methods. Attentional feature-based gating yields consistent performance improvement across tasks. Our approach is implemented as a simple add-on module for RNN-based models with minimal computational overhead and can be adapted to any deep neural architecture.
容 (Rocktäschel et al.,2016)、问答(Xiong et al., 2017)和神经机器翻译 (Wu et al., 2016) 等任务中实现最先进性能的关键因素。通常在文献中提出了特定于任务的注意力体系结构，以进一步提高 DNN 性能 (Dhingra et al., 2017; Xu et al., 2015; Barrett et al., 2018)
在这项工作中，我们提出了一种利用以语言，情感和情感词典编码的单词级先验信息的新颖方法，以提高分类性能。通常，词典特征被连接到单词级表示$\overline{\mathrm{E}}($Wang et al., 2016; Yang et al., 2017; Trotzek et al., 2018) , 作为嵌入每个单词或模型隐藏状态的附加特征。相比之下，我们建议将它们纳人 RNNs 的自我注意机制。我们的目标是通过直接调节其附加的词典功能，使自我注意机制能够识别信息最丰富的单词
我们的贡献如下：(1)我们提出了一种将外部知识整合到基于 RNN 的体系结构中的替代方法，(2)我们提供了经验结果，即我们提出的方法始终优于强基线，以及(3)我们报告了两个数据集中的最新性能。我们公开我们的源代码 1
### 1 $\textbf{Introduction}$
### 2 $\textbf{Related Work}$
现代深度学习算法通常不需要特征工程，而是直接从作为深度神经网络(DNNs)输入的原始数据中学习潜在的表示形式。但是，已经表明，语言知识(手动或半自动编码为词典和知识库)可以显着提高自然语言处理(NLP)任务的 DNN 性能，例如自然语言推理$\overline{\mathrm{E}}($Mrkšić et al.,2017), 语言建模$\overline{\mathrm{E}}($Ahn et al.,2016),命名实体识别$\overline{\mathrm{E}}$(Ghaddar and Langlais, 2018) 和关系提取$\boxed{\mathrm{F}}($Vashishth et al., 2018)
对于 NLP 任务，外部信息源通常通过在此类外部语言知识的上下文中处理原始输入来合并到深度神经体系结构中。在机器学习中，这种上下文处理称为条件；模型执行的计算由从辅助输入中提取的信息进行条件或调制。最常用的调节方法是将外部信息的表示连接到输入或隐藏的网络层
注意机制 (Bahdanau et al., 2015; Vaswani et al., 2017; Lin et al., 2017) 是在文本内
在统计模型基于稀疏特征的传统机器学习文献中，情感词典已被证明对情感分析等任务非常有效，因为它们提供了原始训练数据(Hu and Liu, 2004; Kim and Hovy, 2004; Ding et al., 2008; Yu and Dredze, 2014; Taboada et al., 2011) 中未捕获的其他信息。在出现预训练的单词表示形式$\boxed{\mathrm{F}}($Mikolov et al., 2013; Pennington et al., 2014)之后，使用词典不再是常见的做法，因为单词嵌入也可以捕获这些单词的某些情感含义
最近，在将语言知识集成到 DNN 中以执行各种 NLP 任务方面做出了显着贡献。对于情感分析，Teng et al. (2016) 将词典功能集成到基
$^{\mathrm{l}}$https://github.com/mourga/
affective-attention
$^*$The research was conducted when the author was a re-
searcher at School of ECE, NTUA in Athens, Greece.
|
[
[
0.276123046875,
0.08197021484375,
0.72900390625,
0.08197021484375,
0.72900390625,
0.0955810546875,
0.276123046875,
0.0955810546875
],
[
0.2091064453125,
0.10986328125,
0.8017578125,
0.10809326171875,
0.8017578125,
0.122314453125,
0.2091064453125,
0.12469482421875
],
[
0.20654296875,
0.142578125,
0.8037109375,
0.1436767578125,
0.8037109375,
0.157958984375,
0.20654296875,
0.15673828125
],
[
0.28369140625,
0.1591796875,
0.7197265625,
0.1602783203125,
0.7197265625,
0.174560546875,
0.28369140625,
0.17333984375
],
[
0.173828125,
0.17578125,
0.8310546875,
0.177001953125,
0.8310546875,
0.191162109375,
0.173828125,
0.190673828125
],
[
0.285400390625,
0.191162109375,
0.7197265625,
0.193603515625,
0.7197265625,
0.2078857421875,
0.285400390625,
0.2060546875
],
[
0.2091064453125,
0.21142578125,
0.79443359375,
0.21142578125,
0.79443359375,
0.2257080078125,
0.2091064453125,
0.2257080078125
],
[
0.376220703125,
0.2286376953125,
0.62890625,
0.2286376953125,
0.62890625,
0.2423095703125,
0.376220703125,
0.2423095703125
],
[
0.264404296875,
0.26416015625,
0.343505859375,
0.2666015625,
0.343505859375,
0.281982421875,
0.264404296875,
0.27978515625
],
[
0.5146484375,
0.265380859375,
0.880859375,
0.2666015625,
0.880859375,
0.28076171875,
0.5146484375,
0.27978515625
],
[
0.5146484375,
0.283203125,
0.8818359375,
0.283203125,
0.8818359375,
0.29638671875,
0.5146484375,
0.29638671875
],
[
0.145263671875,
0.29443359375,
0.459228515625,
0.29443359375,
0.459228515625,
0.30810546875,
0.145263671875,
0.30810546875
],
[
0.515625,
0.297607421875,
0.880859375,
0.298583984375,
0.880859375,
0.312255859375,
0.515625,
0.311279296875
],
[
0.1485595703125,
0.30810546875,
0.45751953125,
0.30810546875,
0.45751953125,
0.32177734375,
0.1485595703125,
0.32177734375
],
[
0.515625,
0.3134765625,
0.8818359375,
0.314208984375,
0.8818359375,
0.328857421875,
0.515625,
0.327880859375
],
[
0.14697265625,
0.320556640625,
0.456787109375,
0.32177734375,
0.456787109375,
0.336181640625,
0.14697265625,
0.3349609375
],
[
0.515625,
0.330078125,
0.880859375,
0.330078125,
0.880859375,
0.34326171875,
0.515625,
0.34326171875
],
[
0.1485595703125,
0.336181640625,
0.45751953125,
0.3349609375,
0.45751953125,
0.3486328125,
0.1485595703125,
0.349853515625
],
[
0.1502685546875,
0.3515625,
0.45751953125,
0.3515625,
0.45751953125,
0.3623046875,
0.1502685546875,
0.3623046875
],
[
0.5146484375,
0.3466796875,
0.75146484375,
0.3466796875,
0.75146484375,
0.35986328125,
0.5146484375,
0.35986328125
],
[
0.14697265625,
0.363525390625,
0.459228515625,
0.364013671875,
0.459228515625,
0.37890625,
0.14697265625,
0.377685546875
],
[
0.5341796875,
0.364013671875,
0.88330078125,
0.364013671875,
0.88330078125,
0.377685546875,
0.5341796875,
0.377685546875
],
[
0.1502685546875,
0.38134765625,
0.456787109375,
0.38134765625,
0.456787109375,
0.391357421875,
0.1502685546875,
0.391357421875
],
[
0.51708984375,
0.380126953125,
0.8818359375,
0.380126953125,
0.8818359375,
0.393798828125,
0.51708984375,
0.393798828125
],
[
0.1485595703125,
0.394287109375,
0.45751953125,
0.394287109375,
0.45751953125,
0.405029296875,
0.1485595703125,
0.405029296875
],
[
0.51708984375,
0.396728515625,
0.8818359375,
0.396728515625,
0.8818359375,
0.410400390625,
0.51708984375,
0.410400390625
],
[
0.1502685546875,
0.4091796875,
0.456787109375,
0.4091796875,
0.456787109375,
0.419189453125,
0.1502685546875,
0.419189453125
],
[
0.51318359375,
0.4091796875,
0.88330078125,
0.410400390625,
0.88330078125,
0.427490234375,
0.51318359375,
0.427001953125
],
[
0.14697265625,
0.42041015625,
0.459228515625,
0.421630859375,
0.459228515625,
0.435791015625,
0.14697265625,
0.435302734375
],
[
0.5146484375,
0.427490234375,
0.8818359375,
0.427490234375,
0.8818359375,
0.441162109375,
0.5146484375,
0.441162109375
],
[
0.1502685546875,
0.43701171875,
0.456787109375,
0.43701171875,
0.456787109375,
0.44775390625,
0.1502685546875,
0.44775390625
],
[
0.5146484375,
0.4423828125,
0.8818359375,
0.443603515625,
0.8818359375,
0.457763671875,
0.5146484375,
0.457275390625
],
[
0.1502685546875,
0.451904296875,
0.45751953125,
0.451904296875,
0.45751953125,
0.4619140625,
0.1502685546875,
0.4619140625
],
[
0.1485595703125,
0.46435546875,
0.45751953125,
0.46435546875,
0.45751953125,
0.47802734375,
0.1485595703125,
0.47802734375
],
[
0.51708984375,
0.460205078125,
0.8818359375,
0.460205078125,
0.8818359375,
0.473876953125,
0.51708984375,
0.473876953125
],
[
0.1502685546875,
0.48095703125,
0.456787109375,
0.48095703125,
0.456787109375,
0.4912109375,
0.1502685546875,
0.4912109375
],
[
0.51708984375,
0.476806640625,
0.8818359375,
0.476806640625,
0.8818359375,
0.490478515625,
0.51708984375,
0.490478515625
],
[
0.1485595703125,
0.4921875,
0.459228515625,
0.4921875,
0.459228515625,
0.505859375,
0.1485595703125,
0.505859375
],
[
0.5146484375,
0.493408203125,
0.8818359375,
0.493408203125,
0.8818359375,
0.50732421875,
0.5146484375,
0.50732421875
],
[
0.1485595703125,
0.505859375,
0.45751953125,
0.505859375,
0.45751953125,
0.51953125,
0.1485595703125,
0.51953125
],
[
0.515625,
0.51025390625,
0.55322265625,
0.51025390625,
0.55322265625,
0.521484375,
0.515625,
0.521484375
],
[
0.1502685546875,
0.5224609375,
0.45751953125,
0.5224609375,
0.45751953125,
0.533203125,
0.1502685546875,
0.533203125
],
[
0.5322265625,
0.525390625,
0.8818359375,
0.525390625,
0.8818359375,
0.5390625,
0.5322265625,
0.5390625
],
[
0.1502685546875,
0.53759765625,
0.456787109375,
0.53759765625,
0.456787109375,
0.54736328125,
0.1502685546875,
0.54736328125
],
[
0.5146484375,
0.54052734375,
0.880859375,
0.54052734375,
0.880859375,
0.55419921875,
0.5146484375,
0.55419921875
],
[
0.1158447265625,
0.56103515625,
0.26611328125,
0.5625,
0.26611328125,
0.57666015625,
0.1158447265625,
0.576171875
],
[
0.515625,
0.55810546875,
0.8818359375,
0.55810546875,
0.8818359375,
0.57177734375,
0.515625,
0.57177734375
],
[
0.5146484375,
0.572265625,
0.880859375,
0.57373046875,
0.880859375,
0.58837890625,
0.5146484375,
0.58740234375
],
[
0.12176513671875,
0.5888671875,
0.486083984375,
0.5888671875,
0.486083984375,
0.599609375,
0.12176513671875,
0.599609375
],
[
0.515625,
0.59033203125,
0.8818359375,
0.59033203125,
0.8818359375,
0.60400390625,
0.515625,
0.60400390625
],
[
0.12176513671875,
0.60498046875,
0.489501953125,
0.60498046875,
0.489501953125,
0.6181640625,
0.12176513671875,
0.6181640625
],
[
0.51318359375,
0.60546875,
0.56494140625,
0.60546875,
0.56494140625,
0.62060546875,
0.51318359375,
0.62060546875
],
[
0.12176513671875,
0.62060546875,
0.489501953125,
0.62060546875,
0.489501953125,
0.63427734375,
0.12176513671875,
0.63427734375
],
[
0.12176513671875,
0.63818359375,
0.48779296875,
0.63818359375,
0.48779296875,
0.6484375,
0.12176513671875,
0.6484375
],
[
0.51318359375,
0.63720703125,
0.6708984375,
0.63720703125,
0.6708984375,
0.65087890625,
0.51318359375,
0.65087890625
],
[
0.12176513671875,
0.65380859375,
0.484375,
0.65380859375,
0.484375,
0.66455078125,
0.12176513671875,
0.66455078125
],
[
0.12005615234375,
0.66943359375,
0.48779296875,
0.66943359375,
0.48779296875,
0.68310546875,
0.12005615234375,
0.68310546875
],
[
0.515625,
0.66455078125,
0.88330078125,
0.66455078125,
0.88330078125,
0.677734375,
0.515625,
0.677734375
],
[
0.118408203125,
0.68408203125,
0.486083984375,
0.68408203125,
0.486083984375,
0.69775390625,
0.118408203125,
0.69775390625
],
[
0.515625,
0.6796875,
0.8818359375,
0.6796875,
0.8818359375,
0.693359375,
0.515625,
0.693359375
],
[
0.12176513671875,
0.70166015625,
0.489501953125,
0.70166015625,
0.489501953125,
0.71533203125,
0.12176513671875,
0.71533203125
],
[
0.515625,
0.69677734375,
0.8818359375,
0.69677734375,
0.8818359375,
0.71044921875,
0.515625,
0.71044921875
],
[
0.12005615234375,
0.71630859375,
0.382080078125,
0.71630859375,
0.382080078125,
0.72998046875,
0.12005615234375,
0.72998046875
],
[
0.515625,
0.7119140625,
0.8818359375,
0.7119140625,
0.8818359375,
0.7255859375,
0.515625,
0.7255859375
],
[
0.5146484375,
0.72802734375,
0.88330078125,
0.72802734375,
0.88330078125,
0.7412109375,
0.5146484375,
0.7412109375
],
[
0.139404296875,
0.7353515625,
0.48779296875,
0.7353515625,
0.48779296875,
0.7451171875,
0.139404296875,
0.7451171875
],
[
0.5146484375,
0.74365234375,
0.88330078125,
0.74462890625,
0.88330078125,
0.75830078125,
0.5146484375,
0.75732421875
],
[
0.12005615234375,
0.75,
0.48779296875,
0.75,
0.48779296875,
0.76318359375,
0.12005615234375,
0.76318359375
],
[
0.5146484375,
0.7587890625,
0.880859375,
0.7587890625,
0.880859375,
0.7724609375,
0.5146484375,
0.7724609375
],
[
0.12005615234375,
0.765625,
0.48779296875,
0.765625,
0.48779296875,
0.779296875,
0.12005615234375,
0.779296875
],
[
0.12005615234375,
0.78076171875,
0.489501953125,
0.78076171875,
0.489501953125,
0.79443359375,
0.12005615234375,
0.79443359375
],
[
0.5146484375,
0.77685546875,
0.8818359375,
0.77685546875,
0.8818359375,
0.79052734375,
0.5146484375,
0.79052734375
],
[
0.51318359375,
0.79150390625,
0.8818359375,
0.7919921875,
0.8818359375,
0.80712890625,
0.51318359375,
0.8056640625
],
[
0.12005615234375,
0.7998046875,
0.486083984375,
0.7998046875,
0.486083984375,
0.81005859375,
0.12005615234375,
0.81005859375
],
[
0.12005615234375,
0.81396484375,
0.490234375,
0.81396484375,
0.490234375,
0.82763671875,
0.12005615234375,
0.82763671875
],
[
0.51318359375,
0.81103515625,
0.53076171875,
0.80810546875,
0.53564453125,
0.818359375,
0.51708984375,
0.82177734375
],
[
0.118408203125,
0.830078125,
0.2493896484375,
0.830078125,
0.2493896484375,
0.84423828125,
0.118408203125,
0.84423828125
],
[
0.5341796875,
0.826171875,
0.8818359375,
0.826171875,
0.8818359375,
0.8388671875,
0.5341796875,
0.8388671875
],
[
0.1351318359375,
0.845703125,
0.48779296875,
0.84423828125,
0.48779296875,
0.859375,
0.1351318359375,
0.85986328125
],
[
0.5146484375,
0.84130859375,
0.880859375,
0.84130859375,
0.880859375,
0.85498046875,
0.5146484375,
0.85498046875
],
[
0.1175537109375,
0.8623046875,
0.486083984375,
0.8623046875,
0.486083984375,
0.8759765625,
0.1175537109375,
0.8759765625
],
[
0.5146484375,
0.85791015625,
0.8818359375,
0.85791015625,
0.8818359375,
0.87158203125,
0.5146484375,
0.87158203125
],
[
0.1485595703125,
0.88427734375,
0.484375,
0.88427734375,
0.484375,
0.89794921875,
0.1485595703125,
0.89794921875
],
[
0.5341796875,
0.8828125,
0.77685546875,
0.88427734375,
0.77685546875,
0.8984375,
0.5341796875,
0.89794921875
],
[
0.1234130859375,
0.89794921875,
0.443359375,
0.89794921875,
0.443359375,
0.90771484375,
0.1234130859375,
0.90771484375
],
[
0.5205078125,
0.8994140625,
0.68701171875,
0.8994140625,
0.68701171875,
0.90673828125,
0.5205078125,
0.90673828125
]
] |
[
"用于外部知识集成的基于注意力的调节方法",
"Katerina Margatina', Christos Baziotis? * Alexandros Potamianosl,3,4",
"1School of ECE, National Technical University of Athens, Athens, Greece",
"2 School of Informatics, University of Edinburgh, UK",
"3 Signal Analysis and Interpretation Laboratory (SAIL), USC, Los Angeles, USA",
"4 Behavioral Signal Technologies, Los Angeles, USA",
"el121o8@central.ntua.gr,c.baziotis@sms.ed.ac.uk,",
"potam@central.ntua.gr",
"Abstract",
"容(Rocktaschel et al.,2016)、问答(Xiong et al..",
"2017）和神经机器翻译（Wuetal.，2016）等任",
"In this paper, we present a novel approach",
"务中实现最先进性能的关键因素。通常在文献",
"for incorporating external knowledge in Re-",
"中提出了特定于任务的注意力体系结构，以进",
"current Neural Networks (RNNs).We pro",
"一步提高DNN性能（Dhingraetal.，2017；Xu",
"pose the integration of lexicon features into",
"the self-attention mechanismof RNN-based",
"et al., 2015; Barrett et al., 2018)",
"architectures.This form of conditioning on",
"在这项工作中，我们提出了一种利用以语",
"the attention distribution,enforcesthecontri-",
"言，情感和情感词典编码的单词级先验信息的",
"bution of the most salient words for the task",
"新颖方法，以提高分类性能。通常，词典特征",
"athand.We introducethree methods,namely",
"被连接到单词级表示(Wangetal.,2016;Yang",
"attentional concatenation, feature-based gating",
"etal.,2017；Trotzeketal,2018），作为嵌人每",
"and affinetransformation.Experiments onsix",
"个单词或模型隐藏状态的附加特征。相比之",
"benchmarkdatasetsshowtheeffectivenessof",
"our methods. Attentional feature-based gating",
"下，我们建议将它们纳人RNNs的自我注意机",
"yields consistentperformanceimprovement",
"制。我们的目标是通过直接调节其附加的词典",
"across tasks. Our approach is implemented as",
"功能，使自我注意机制能够识别信息最丰富的",
"a simple add-on module for RNN-based mod-",
"单词",
"elswithminimalcomputationaloverheadand",
"我们的贡献如下：（1）我们提出了一种将外",
"canbeadaptedtoanydeepneuralarchitecture",
"部知识整合到基于RNN的体系结构中的替代",
"1Introduction",
"方法，（2）我们提供了经验结果，即我们提出",
"的方法始终优于强基线，以及（3）我们报告了",
"现代深度学习算法通常不需要特征工程，而是",
"两个数据集中的最新性能。我们公开我们的源",
"直接从作为深度神经网络（DNNs）输入的原",
"代码1",
"始数据中学习潜在的表示形式。。但是，已经表",
"明，语言知识（手动或半自动编码为词典和知",
"2Related Work",
"识库）可以显着提高自然语言处理（NLP）任",
"务的DNN性能，例如自然语言推理E(Mrksic",
"在统计模型基于稀疏特征的传统机器学习文",
"etal.2017），语言建模E（Ahnetal,2016），命",
"献中，情感词典已被证明对情感分析等任务",
"名实体识别E（GhaddarandLanglais，2018）和",
"非常有效，因为它们提供了原始训练数据（Hu",
"关系提取E(Vashishthetal.,2018)",
"and Liu, 2004; Kim and Hovy, 2004; Ding et al..",
"2008; Yu and Dredze, 2014; Taboada et al., 2011)",
"对于NLP任务，外部信息源通常通过在此",
"中未捕获的其他信息。在出现预训练的单词表",
"类外部语言知识的上下文中处理原始输人来合",
"示形式E(Mikolov et al., 2013; Pennington et al.",
"并到深度神经体系结构中。在机器学习中，这",
"种上下文处理称为条件；模型执行的计算由从",
"2014）之后，使用词典不再是常见的做法，因",
"为单词嵌入也可以捕获这些单词的某些情感含",
"辅助输入中提取的信息进行条件或调制。最常",
"用的调节方法是将外部信息的表示连接到输入",
"义",
"或隐藏的网络层",
"最近，在将语言知识集成到DNN中以执行",
"注意机制（Bahdanauetal.，2015;Vaswani",
"各种NLP任务方面做出了显着贡献。对于情",
"etal.，2017;Lin et al.，2017）是在文本内",
"感分析，Tengetal.(2016)）将词典功能集成到基",
"* The research was conducted when the author was a re",
"'https://github.com/mourga/",
"searcherat SchoolofECE,NTUAinAthens,Greece.",
"affective-attention"
] |
{
"arxiv_id": "1906.03674",
"language": "zh",
"page_idx": 0
}
|
||
1906.03674-zh-1
|
1906.03674-zh-1.jpg
|
于 RNN 的模型中，并对单词功能进行自定义加权和计算。Shin et al.(2017)提出了三种卷积神经网络特定的词典集成方法，可在两个数据集上实现最先进的性能。Kumar et al.(2018) 将知识库中的功能连接到专注的双向 LSTM 体系结构中的单词表示，并报告最新的结果。对于讽刺检测，Yang et al.(2017) 通过将其连接到段落和文档级表示形式来整合心理语言学， 文体，结构和可读性功能
此外，关于深度神经网络中组合表示方法的开发和评估的文献有限。Peters et al.(2017) 声称级联，非线性映射和类似注意力的机制是将语言模型表示包含在其序列模型中的未开发方法。他们采用简单的连接，将其他方法的探索留给未来的工作。Dumoulin et al.(2018) 提供了特征转换的概述，例如基于级联的条件， 条件偏置和门控机制。他们审查了视觉问答$\overline{\mathrm{E}}$ (Strub et al., 2018), 风格转移$\overline{\mathrm{E}}($Dumoulin et al., 2017) 和语言建模$\mathbb{E}$(Dauphin et al., 2017) 等任务中条件方法的有效性。他们还扩展了$\overline{\mathrm{EPerez}}$ et al.(2017) 的工作，该工作提出了基于特征的线性调制 (FiLM) 框架，并研究了其在视觉推理任务中的应用。Balazs and Matsuo (2019) 提供了一项实证研究，显示了在单词级和句子级评估任务中组合字符和单词表示的不同方式的影响。报告的一些发现是，门控条件在各种单词相似性和相关性任务中始终表现更好
## 3 Proposed Model
## 3.1 Network Architecture
Word 嵌入层。单词的输人序列$w_1,w_2,...,w_T$ 被投影到低维向量空间$R^W$,其中$W$是嵌人层的大小，$T$是句子中的单词数。我们用预训练的单词嵌入初始化嵌入层的权重
LSTM层。长短期记忆单元(LSTM)(Hochreiter and Schmidhuber, 1997) 将句子的单词作为输人，并产生单词注释$h_1,h_2,...,h_T$,其中$h_i$ 是LSTM 在时间步长$i$处的隐藏状态，总结了直到$w_i$的所有句子信息
自关注层。我们使用自我注意机制 (Cheng et al., 2016) 来找到每个单词对手头任务的相对重要性。注意机制为每个单词注释$h_i$分配一个分数$a_i$ 。我们计算输入序列的固定表示$r$, 作为所有单词注释的加权和。形式上：
$$\begin{aligned}&a_{i}=softmax(v_{a}^{\mathsf{T}}f(h_{i}))\\&r=\sum_{i=1}^{T}a_{i}h_{i}\end{aligned}$$
(1)
(2)
其中$f(.)$对应于非线性变换$tanh(W_{a}h_{i}+b_{a})$
和$W_a,b_a,v_a$是关注层的参数
<table>
<tbody>
<tr>
<th>$\operatorname{Lexicons}$</th>
<th>Annotations</th>
<th>$\#$dim.</th>
<th>井 words</th>
</tr>
<tr>
<td>LIWC</td>
<td>psycho-linguistic</td>
<td>73</td>
<td>18,504</td>
</tr>
<tr>
<td>Bing Liu</td>
<td>valence</td>
<td>1</td>
<td>2,477</td>
</tr>
<tr>
<td>AFINN</td>
<td>sentiment</td>
<td>1</td>
<td>6,786</td>
</tr>
<tr>
<td>MPQA</td>
<td>sentiment</td>
<td>4</td>
<td>6,886</td>
</tr>
<tr>
<td>SemEval15</td>
<td>sentiment</td>
<td>1</td>
<td>1,515</td>
</tr>
<tr>
<td>Emolex</td>
<td>emotion</td>
<td>19</td>
<td>14,182</td>
</tr>
</tbody>
</table>
Table 1: The lexicons used as external knowledge.
### 3.2 $\textbf{External Knowledge}$
在这项工作中，我们使用来自人类专家的现有语言和情感知识来增强模型。具体来说，我们利用包含心理语言、情感和情感注释的词典。我们通过连接表$\overline{\mathbb{E}}$1 中所示的词典中的单词注释，为词汇表中的每个单词构造一个特征向量$c(w_i)$。对于缺失的单词，我们在$c(w_i)$的相应维度中附加零
3.3 Conditional Attention Mechanism
我们扩展了标准的自我注意机制(等式$\overline{\mathbb{E}}$1 ,回2 ), 以便在每个单词的先前词汇信息上调节给定句子的注意力分布。为此，我们将单词注释$h_i$以及每个单词的词典功能$c(w_i)$用作自我关注层的输入。因此，我们用$f(h_i,c(w_i))$代替方程 1 中的$f(h_i)$ 。具体来说，我们探索了三种调节方法，如图 1 所示。我们将条件函数称为$f_{i}( . )$ , 权重矩阵称为$W_i$ , 偏 差 称 为 $b_{i}$ ,其中$i$是每种方法的指示字母。我们在E5 节(表E 3 ) 中介绍了我们的结果，并将三种条件方法表示为“conc。“门”和“仿射”分别
连接注意事项。在这种方法中，如图。1(a),我们学习了每个单词注释$h_i$与其词典特征$c(w_i)$ 的串联函数。直觉是，通过向$h_i$添加额外的维度，学习的表示形式更具鉴别力。具体而言：
$$f_c(h_i,c(w_i))=tanh(W_c[h_i\parallel c(w_i)]+b_c)$$
其中$\parallel$表示连接操作$,W_c,b_c$是可学习的参数基于注意力特征的门控。第二种方法，如图所示。1(b),学习一个特征掩码，该掩码应用于每个单词注释$h_i$ 。具体来说，具有 S 形激活函数的门机制从每个$c(w_i)$生成一个掩码向量， 其值介于 0 和 1 之间 (图中的黑点和白点。1 (b))。直观上，这种门控机制选择了显著的维度(i。e。$h_i$的特征),以词汇信息为条件。形式上：
$$f_g(h_i,c(w_i))=\sigma(W_gc(w_i)+b_g)\odot h_i$$
(4)
,其中$\odot$表示逐元素乘法$,W_g,b_g$ 是可学习的
参数
|
[
[
0.5341796875,
0.07244873046875,
0.599609375,
0.0748291015625,
0.59765625,
0.08905029296875,
0.5322265625,
0.0872802734375
],
[
0.63818359375,
0.07244873046875,
0.72314453125,
0.0748291015625,
0.72314453125,
0.08905029296875,
0.63818359375,
0.0872802734375
],
[
0.75244140625,
0.0748291015625,
0.875,
0.0748291015625,
0.875,
0.08905029296875,
0.75244140625,
0.08905029296875
],
[
0.54248046875,
0.09027099609375,
0.591796875,
0.0926513671875,
0.59033203125,
0.10687255859375,
0.54150390625,
0.1051025390625
],
[
0.12005615234375,
0.07659912109375,
0.48779296875,
0.07659912109375,
0.48779296875,
0.09027099609375,
0.12005615234375,
0.09027099609375
],
[
0.62060546875,
0.0926513671875,
0.73876953125,
0.09027099609375,
0.73876953125,
0.105712890625,
0.62060546875,
0.10809326171875
],
[
0.765625,
0.0926513671875,
0.78662109375,
0.0926513671875,
0.78662109375,
0.10687255859375,
0.765625,
0.10687255859375
],
[
0.81787109375,
0.0926513671875,
0.873046875,
0.0926513671875,
0.873046875,
0.10687255859375,
0.81787109375,
0.10687255859375
],
[
0.6533203125,
0.105712890625,
0.7080078125,
0.10809326171875,
0.7060546875,
0.122314453125,
0.6513671875,
0.12054443359375
],
[
0.12005615234375,
0.09320068359375,
0.48779296875,
0.09320068359375,
0.48779296875,
0.10687255859375,
0.12005615234375,
0.10687255859375
],
[
0.53564453125,
0.10687255859375,
0.59765625,
0.10687255859375,
0.59765625,
0.1217041015625,
0.53564453125,
0.1217041015625
],
[
0.8212890625,
0.10687255859375,
0.86572265625,
0.10687255859375,
0.86572265625,
0.1217041015625,
0.8212890625,
0.1217041015625
],
[
0.77099609375,
0.1092529296875,
0.78271484375,
0.1092529296875,
0.78271484375,
0.12054443359375,
0.77099609375,
0.12054443359375
],
[
0.118408203125,
0.10809326171875,
0.48779296875,
0.1092529296875,
0.48779296875,
0.12353515625,
0.118408203125,
0.122314453125
],
[
0.54150390625,
0.12353515625,
0.59375,
0.12353515625,
0.59375,
0.1353759765625,
0.54150390625,
0.1353759765625
],
[
0.646484375,
0.12353515625,
0.7138671875,
0.12353515625,
0.7138671875,
0.1353759765625,
0.646484375,
0.1353759765625
],
[
0.8212890625,
0.12054443359375,
0.8671875,
0.122314453125,
0.86572265625,
0.1383056640625,
0.8193359375,
0.135986328125
],
[
0.76318359375,
0.12890625,
0.77685546875,
0.119384765625,
0.78857421875,
0.127685546875,
0.77490234375,
0.13720703125
],
[
0.8212890625,
0.13720703125,
0.86572265625,
0.13720703125,
0.86572265625,
0.1519775390625,
0.8212890625,
0.1519775390625
],
[
0.12005615234375,
0.1258544921875,
0.48779296875,
0.1258544921875,
0.48779296875,
0.1395263671875,
0.12005615234375,
0.1395263671875
],
[
0.54150390625,
0.1395263671875,
0.59033203125,
0.1395263671875,
0.59033203125,
0.15087890625,
0.54150390625,
0.15087890625
],
[
0.6455078125,
0.1383056640625,
0.71533203125,
0.1383056640625,
0.71533203125,
0.152587890625,
0.6455078125,
0.152587890625
],
[
0.77099609375,
0.1395263671875,
0.78271484375,
0.1395263671875,
0.78271484375,
0.149658203125,
0.77099609375,
0.149658203125
],
[
0.82373046875,
0.152587890625,
0.86572265625,
0.152587890625,
0.86572265625,
0.166259765625,
0.82373046875,
0.166259765625
],
[
0.12176513671875,
0.141357421875,
0.48779296875,
0.141357421875,
0.48779296875,
0.155029296875,
0.12176513671875,
0.155029296875
],
[
0.5263671875,
0.15380859375,
0.60693359375,
0.15380859375,
0.60693359375,
0.1650390625,
0.5263671875,
0.1650390625
],
[
0.646484375,
0.155029296875,
0.7138671875,
0.155029296875,
0.7138671875,
0.166259765625,
0.646484375,
0.166259765625
],
[
0.77099609375,
0.15380859375,
0.78271484375,
0.15380859375,
0.78271484375,
0.1650390625,
0.77099609375,
0.1650390625
],
[
0.5380859375,
0.16748046875,
0.59375,
0.16748046875,
0.59375,
0.1817626953125,
0.5380859375,
0.1817626953125
],
[
0.12176513671875,
0.157958984375,
0.48779296875,
0.157958984375,
0.48779296875,
0.171630859375,
0.12176513671875,
0.171630859375
],
[
0.767578125,
0.16748046875,
0.78662109375,
0.16748046875,
0.78662109375,
0.18115234375,
0.767578125,
0.18115234375
],
[
0.81787109375,
0.16748046875,
0.87158203125,
0.16748046875,
0.87158203125,
0.1817626953125,
0.81787109375,
0.1817626953125
],
[
0.12005615234375,
0.17333984375,
0.48779296875,
0.17333984375,
0.48779296875,
0.18701171875,
0.12005615234375,
0.18701171875
],
[
0.6513671875,
0.169189453125,
0.7080078125,
0.169189453125,
0.7080078125,
0.18115234375,
0.6513671875,
0.18115234375
],
[
0.118408203125,
0.188232421875,
0.489501953125,
0.189453125,
0.489501953125,
0.2037353515625,
0.118408203125,
0.203125
],
[
0.5283203125,
0.193603515625,
0.86572265625,
0.1953125,
0.86572265625,
0.208984375,
0.5283203125,
0.207275390625
],
[
0.118408203125,
0.2060546875,
0.32080078125,
0.2060546875,
0.32080078125,
0.2197265625,
0.118408203125,
0.2197265625
],
[
0.1368408203125,
0.221435546875,
0.48779296875,
0.221435546875,
0.48779296875,
0.235107421875,
0.1368408203125,
0.235107421875
],
[
0.118408203125,
0.2381591796875,
0.489501953125,
0.2381591796875,
0.489501953125,
0.251708984375,
0.118408203125,
0.251708984375
],
[
0.515625,
0.2381591796875,
0.54150390625,
0.2381591796875,
0.54150390625,
0.2464599609375,
0.515625,
0.2464599609375
],
[
0.53564453125,
0.236328125,
0.72119140625,
0.236328125,
0.72119140625,
0.2493896484375,
0.53564453125,
0.2493896484375
],
[
0.12005615234375,
0.254150390625,
0.489501953125,
0.254150390625,
0.489501953125,
0.267333984375,
0.12005615234375,
0.267333984375
],
[
0.515625,
0.257080078125,
0.8818359375,
0.257080078125,
0.8818359375,
0.270751953125,
0.515625,
0.270751953125
],
[
0.12005615234375,
0.270751953125,
0.48779296875,
0.270751953125,
0.48779296875,
0.283935546875,
0.12005615234375,
0.283935546875
],
[
0.51708984375,
0.27490234375,
0.880859375,
0.27490234375,
0.880859375,
0.28515625,
0.51708984375,
0.28515625
],
[
0.118408203125,
0.28515625,
0.48779296875,
0.28515625,
0.48779296875,
0.298583984375,
0.118408203125,
0.298583984375
],
[
0.515625,
0.289306640625,
0.8818359375,
0.289306640625,
0.8818359375,
0.302734375,
0.515625,
0.302734375
],
[
0.12005615234375,
0.302734375,
0.48779296875,
0.302734375,
0.48779296875,
0.31640625,
0.12005615234375,
0.31640625
],
[
0.5146484375,
0.3046875,
0.8818359375,
0.3046875,
0.8818359375,
0.318359375,
0.5146484375,
0.318359375
],
[
0.118408203125,
0.317626953125,
0.48779296875,
0.317626953125,
0.48779296875,
0.330810546875,
0.118408203125,
0.330810546875
],
[
0.5146484375,
0.320556640625,
0.8818359375,
0.320556640625,
0.8818359375,
0.334228515625,
0.5146484375,
0.334228515625
],
[
0.12005615234375,
0.3349609375,
0.489501953125,
0.3349609375,
0.489501953125,
0.3486328125,
0.12005615234375,
0.3486328125
],
[
0.51318359375,
0.33740234375,
0.8818359375,
0.33740234375,
0.8818359375,
0.350830078125,
0.51318359375,
0.350830078125
],
[
0.12005615234375,
0.350830078125,
0.489501953125,
0.350830078125,
0.489501953125,
0.364013671875,
0.12005615234375,
0.364013671875
],
[
0.515625,
0.355224609375,
0.62646484375,
0.355224609375,
0.62646484375,
0.366455078125,
0.515625,
0.366455078125
],
[
0.12005615234375,
0.365234375,
0.48779296875,
0.366455078125,
0.48779296875,
0.380126953125,
0.12005615234375,
0.37890625
],
[
0.12005615234375,
0.383056640625,
0.489501953125,
0.383056640625,
0.489501953125,
0.396728515625,
0.12005615234375,
0.396728515625
],
[
0.51318359375,
0.380126953125,
0.82861328125,
0.380126953125,
0.82861328125,
0.393798828125,
0.51318359375,
0.393798828125
],
[
0.12005615234375,
0.399658203125,
0.48779296875,
0.399658203125,
0.48779296875,
0.413330078125,
0.12005615234375,
0.413330078125
],
[
0.51318359375,
0.40087890625,
0.884765625,
0.399658203125,
0.884765625,
0.41455078125,
0.51318359375,
0.4150390625
],
[
0.12005615234375,
0.4150390625,
0.489501953125,
0.4150390625,
0.489501953125,
0.4287109375,
0.12005615234375,
0.4287109375
],
[
0.5146484375,
0.41748046875,
0.8818359375,
0.41748046875,
0.8818359375,
0.43115234375,
0.5146484375,
0.43115234375
],
[
0.118408203125,
0.43115234375,
0.48779296875,
0.43115234375,
0.48779296875,
0.44482421875,
0.118408203125,
0.44482421875
],
[
0.5146484375,
0.432861328125,
0.8818359375,
0.432861328125,
0.8818359375,
0.446533203125,
0.5146484375,
0.446533203125
],
[
0.12005615234375,
0.44775390625,
0.489501953125,
0.44775390625,
0.489501953125,
0.46142578125,
0.12005615234375,
0.46142578125
],
[
0.51123046875,
0.44775390625,
0.8818359375,
0.448974609375,
0.8818359375,
0.463134765625,
0.51123046875,
0.4619140625
],
[
0.12005615234375,
0.463134765625,
0.489501953125,
0.463134765625,
0.489501953125,
0.476806640625,
0.12005615234375,
0.476806640625
],
[
0.515625,
0.465576171875,
0.8818359375,
0.465576171875,
0.8818359375,
0.478515625,
0.515625,
0.478515625
],
[
0.12005615234375,
0.478515625,
0.489501953125,
0.478515625,
0.489501953125,
0.4921875,
0.12005615234375,
0.4921875
],
[
0.51318359375,
0.48095703125,
0.880859375,
0.48095703125,
0.880859375,
0.49462890625,
0.51318359375,
0.49462890625
],
[
0.118408203125,
0.495361328125,
0.429931640625,
0.495361328125,
0.429931640625,
0.5087890625,
0.118408203125,
0.5087890625
],
[
0.5146484375,
0.496337890625,
0.8818359375,
0.49755859375,
0.8818359375,
0.5126953125,
0.5146484375,
0.51123046875
],
[
0.515625,
0.51416015625,
0.8818359375,
0.51416015625,
0.8818359375,
0.52783203125,
0.515625,
0.52783203125
],
[
0.1158447265625,
0.5224609375,
0.2939453125,
0.5224609375,
0.2939453125,
0.5361328125,
0.1158447265625,
0.5361328125
],
[
0.5146484375,
0.52978515625,
0.88330078125,
0.52978515625,
0.88330078125,
0.54345703125,
0.5146484375,
0.54345703125
],
[
0.1175537109375,
0.5458984375,
0.339111328125,
0.54638671875,
0.339111328125,
0.56103515625,
0.1175537109375,
0.56005859375
],
[
0.51318359375,
0.54443359375,
0.880859375,
0.54443359375,
0.880859375,
0.55810546875,
0.51318359375,
0.55810546875
],
[
0.118408203125,
0.56640625,
0.48779296875,
0.568359375,
0.48779296875,
0.5830078125,
0.118408203125,
0.58056640625
],
[
0.515625,
0.5625,
0.8271484375,
0.5625,
0.8271484375,
0.576171875,
0.515625,
0.576171875
],
[
0.12005615234375,
0.5830078125,
0.489501953125,
0.5830078125,
0.489501953125,
0.59716796875,
0.12005615234375,
0.59716796875
],
[
0.5146484375,
0.580078125,
0.88330078125,
0.58056640625,
0.88330078125,
0.595703125,
0.5146484375,
0.59423828125
],
[
0.12005615234375,
0.5986328125,
0.486083984375,
0.5986328125,
0.486083984375,
0.6123046875,
0.12005615234375,
0.6123046875
],
[
0.515625,
0.59716796875,
0.880859375,
0.59716796875,
0.880859375,
0.61083984375,
0.515625,
0.61083984375
],
[
0.12005615234375,
0.6162109375,
0.393798828125,
0.6162109375,
0.393798828125,
0.6298828125,
0.12005615234375,
0.6298828125
],
[
0.5146484375,
0.6123046875,
0.880859375,
0.6123046875,
0.880859375,
0.6259765625,
0.5146484375,
0.6259765625
],
[
0.5146484375,
0.62890625,
0.87158203125,
0.62890625,
0.87158203125,
0.642578125,
0.5146484375,
0.642578125
],
[
0.12005615234375,
0.63525390625,
0.48779296875,
0.63525390625,
0.48779296875,
0.6484375,
0.12005615234375,
0.6484375
],
[
0.118408203125,
0.6494140625,
0.48779296875,
0.65185546875,
0.48779296875,
0.66650390625,
0.118408203125,
0.66455078125
],
[
0.5263671875,
0.6591796875,
0.880859375,
0.6591796875,
0.880859375,
0.6728515625,
0.5263671875,
0.6728515625
],
[
0.118408203125,
0.66748046875,
0.486083984375,
0.66748046875,
0.486083984375,
0.68115234375,
0.118408203125,
0.68115234375
],
[
0.118408203125,
0.68310546875,
0.489501953125,
0.68408203125,
0.489501953125,
0.69775390625,
0.118408203125,
0.69677734375
],
[
0.5146484375,
0.6884765625,
0.880859375,
0.6884765625,
0.880859375,
0.70166015625,
0.5146484375,
0.70166015625
],
[
0.12005615234375,
0.70068359375,
0.2939453125,
0.70068359375,
0.2939453125,
0.71435546875,
0.12005615234375,
0.71435546875
],
[
0.515625,
0.7080078125,
0.8818359375,
0.7080078125,
0.8818359375,
0.7216796875,
0.515625,
0.7216796875
],
[
0.12176513671875,
0.72021484375,
0.489501953125,
0.72021484375,
0.489501953125,
0.73388671875,
0.12176513671875,
0.73388671875
],
[
0.515625,
0.724609375,
0.87890625,
0.724609375,
0.87890625,
0.7353515625,
0.515625,
0.7353515625
],
[
0.118408203125,
0.7353515625,
0.48779296875,
0.73388671875,
0.48779296875,
0.7490234375,
0.118408203125,
0.75
],
[
0.5146484375,
0.74072265625,
0.8818359375,
0.74072265625,
0.8818359375,
0.75439453125,
0.5146484375,
0.75439453125
],
[
0.12005615234375,
0.751953125,
0.47265625,
0.751953125,
0.47265625,
0.765625,
0.12005615234375,
0.765625
],
[
0.466796875,
0.7548828125,
0.48779296875,
0.7548828125,
0.48779296875,
0.76318359375,
0.466796875,
0.76318359375
],
[
0.515625,
0.75732421875,
0.8818359375,
0.75732421875,
0.8818359375,
0.77099609375,
0.515625,
0.77099609375
],
[
0.12005615234375,
0.7685546875,
0.49365234375,
0.7685546875,
0.49365234375,
0.7822265625,
0.12005615234375,
0.7822265625
],
[
0.5146484375,
0.771484375,
0.880859375,
0.771484375,
0.880859375,
0.78515625,
0.5146484375,
0.78515625
],
[
0.12176513671875,
0.78369140625,
0.420654296875,
0.78369140625,
0.420654296875,
0.79736328125,
0.12176513671875,
0.79736328125
],
[
0.515625,
0.7890625,
0.8818359375,
0.7890625,
0.8818359375,
0.802734375,
0.515625,
0.802734375
],
[
0.208251953125,
0.80859375,
0.3955078125,
0.80810546875,
0.3955078125,
0.822265625,
0.208251953125,
0.82373046875
],
[
0.46337890625,
0.80859375,
0.489501953125,
0.80859375,
0.489501953125,
0.82177734375,
0.46337890625,
0.82177734375
],
[
0.515625,
0.8046875,
0.88330078125,
0.8046875,
0.88330078125,
0.818359375,
0.515625,
0.818359375
],
[
0.5146484375,
0.818359375,
0.56494140625,
0.822265625,
0.5615234375,
0.8359375,
0.51123046875,
0.83203125
],
[
0.252685546875,
0.82763671875,
0.274658203125,
0.82763671875,
0.274658203125,
0.84033203125,
0.252685546875,
0.84033203125
],
[
0.214111328125,
0.84033203125,
0.313232421875,
0.84033203125,
0.313232421875,
0.859375,
0.214111328125,
0.859375
],
[
0.46337890625,
0.8427734375,
0.489501953125,
0.8427734375,
0.489501953125,
0.85693359375,
0.46337890625,
0.85693359375
],
[
0.5439453125,
0.85107421875,
0.88330078125,
0.85107421875,
0.88330078125,
0.8642578125,
0.5439453125,
0.8642578125
],
[
0.2509765625,
0.85791015625,
0.27783203125,
0.85791015625,
0.27783203125,
0.87060546875,
0.2509765625,
0.87060546875
],
[
0.12176513671875,
0.8798828125,
0.48779296875,
0.8798828125,
0.48779296875,
0.8935546875,
0.12176513671875,
0.8935546875
],
[
0.51318359375,
0.8798828125,
0.880859375,
0.8798828125,
0.880859375,
0.8935546875,
0.51318359375,
0.8935546875
],
[
0.12176513671875,
0.8955078125,
0.3525390625,
0.8955078125,
0.3525390625,
0.9091796875,
0.12176513671875,
0.9091796875
],
[
0.5146484375,
0.892578125,
0.55517578125,
0.89453125,
0.55322265625,
0.9111328125,
0.51318359375,
0.9091796875
]
] |
[
"Lexicons",
"Annotations",
"#dim.#words",
"LIWC",
"于RNN的模型中，并对单词功能进行自定义",
"psycho-linguistic",
"73",
"18,504",
"valence",
"加权和计算。Shinetal.(2017)提出了三种卷积",
"Bing Liu",
"2,477",
"1",
"神经网络特定的词典集成方法，可在两个数据",
"AFINN",
"sentiment",
"6,786",
"I",
"6,886",
"集上实现最先进的性能。Kumaretal.(2018）将",
"MPQA",
"sentiment",
"4",
"1,515",
"知识库中的功能连接到专注的双向LSTM体",
"SemEval15",
"sentiment",
"1",
"Emolex",
"系结构中的单词表示，并报告最新的结果。对",
"19",
"14,182",
"于讽刺检测，Yangetal.(2017）通过将其连接",
"emotion",
"到段落和文档级表示形式来整合心理语言学，",
"Table 1: The lexicons used as external knowledge",
"文体，结构和可读性功能",
"此外，关于深度神经网络中组合表示方法的",
"开发和评估的文献有限。Petersetal.(2017)声",
"3.2",
"ExternalKnowledge",
"称级联，非线性映射和类似注意力的机制是将",
"在这项工作中，我们使用来自人类专家的现有",
"语言模型表示包含在其序列模型中的未开发",
"语言和情感知识来增强模型。具体来说，我们",
"方法。他们采用简单的连接，将其他方法的探",
"利用包含心理语言、情感和情感注释的词典。",
"索留给未来的工作。Dumoulinetal.（2018）提",
"我们通过连接表E1中所示的词典中的单词注",
"供了特征转换的概述，例如基于级联的条件，",
"释，为词汇表中的每个单词构造一个特征向量",
"条件偏置和门控机制。他们审查了视觉问答E",
"c(w;）。对于缺失的单词，我们在c(wi）的相应",
"(Strubetal.,2018)，风格转移E(Dumoulinetal.,",
"维度中附加零",
"2017）和语言建模E(Dauphinetal，2017）等任",
"务中条件方法的有效性。他们还扩展了EPerez",
"3.3Conditional Attention Mechanism",
"etal.(2017)的工作，该工作提出了基于特征的",
"我们扩展了标准的自我注意机制（等式E1，国",
"线性调制（FiLM）框架，并研究了其在视觉推",
"2），以便在每个单词的先前词汇信息上调节给",
"理任务中的应用。BalazsandMatsuo(2019）提",
"定句子的注意力分布。为此，我们将单词注释",
"供了一项实证研究，显示了在单词级和句子级",
"h,以及每个单词的词典功能cwi）用作自我关",
"评估任务中组合字符和单词表示的不同方式的",
"注层的输人。因此，我们用f(hi,c(wi)）代替方",
"影响。报告的一些发现是，门控条件在各种单",
"程1中的f（hi）。具体来说，我们探索了三种",
"词相似性和相关性任务中始终表现更好",
"调节方法，如图1所示。我们将条件函数称为",
"f（），权重矩阵称为Wi，偏差称为bi，其中",
"3ProposedModel",
"i是每种方法的指示字母。我们在E5节（表E",
"3.1NetworkArchitecture",
"3）中介绍了我们的结果，并将三种条件方法",
"Word嵌入层。单词的输入序列Wi,w2,wT",
"表示为“conc。、“门”和“仿射”分别",
"被投影到低维向量空间RW，其中W是嵌人",
"连接注意事项。在这种方法中，如图。1(a)，我",
"层的大小，T是句子中的单词数。我们用预训",
"们学习了每个单词注释hi与其词典特征cwi)",
"练的单词嵌人初始化嵌入层的权重",
"的串联函数。直觉是，通过向h；添加额外的维",
"度，学习的表示形式更具鉴别力。具体而言：",
"LSTM层。长短期记忆单元（LSTM)(Hochreiter",
"andSchmidhuber，1997）将句子的单词作为输",
"fe(hi, c(wi)) = tanh(Wc[hi Il c(wi)] +bc) (3)",
"人，并产生单词注释h1,h2，.,hT，其中hi是",
"LSTM在时间步长处的隐藏状态，总结了直",
"其中表示连接操作，Wc,bc是可学习的参数",
"到wi的所有句子信息",
"基于注意力特征的门控。第二种方法，如图所",
"自关注层。我们使用自我注意机制（Cheng",
"示。1（b），学习个特征掩码，该掩码应用于",
"etal.2016)来找到每个单词对手头任务的相对",
"每个单词注释hi。具体来说，具有S形激活",
"重要性。注意机制为每个单词注释hi分配",
"-",
"函数的门机制从每个c(wi）生成一个掩码向量，",
"个分数ai。我们计算输入序列的固定表示r，",
"其值介于0和1之间（图中的黑点和白点。1",
"作为所有单词注释的加权和。形式上：",
"(b))。直观上，这种门控机制选择了显著的维",
"a,= softmar(vif(hi))",
"(1)",
"度(i。e。hi的特征)，以词汇信息为条件。形",
"式上：",
"T",
"r=aihi",
"(2)",
"fg(hi,c(wi)) = α(Wgc(wi) + bg) o hi(4)",
"i=1",
"其中f(）对应于非线性变换tanh(Wahi+ba)",
"，其中表示逐元素乘法，Wg，bg是可学习的",
"和Wa,ba，Ua是关注层的参数",
"参数"
] |
{
"arxiv_id": "1906.03674",
"language": "zh",
"page_idx": 1
}
|
||
1906.03674-zh-3
|
1906.03674-zh-3.jpg
|
<table>
<tbody>
<tr>
<th>Dataset</th>
<th>Study</th>
<th>Task</th>
<th>Domain</th>
<th>Classes</th>
<th>$\underline{N_{train}}$</th>
<th>$\underline{N_{test}}$</th>
</tr>
<tr>
<td>$SST-5$</td>
<td>Socher et al. (2013)</td>
<td>Sentiment</td>
<td>Movie Reviews</td>
<td>5</td>
<td>9,645</td>
<td>2,210</td>
</tr>
<tr>
<td>Sent17</td>
<td>Rosenthal et al. (2017)</td>
<td>Sentiment</td>
<td>Twitter</td>
<td>3</td>
<td>49,570</td>
<td>12,284</td>
</tr>
<tr>
<td>PhychExp</td>
<td>Wallbott and Scherer (1986)</td>
<td>Emotion</td>
<td>$\operatorname{Experiences}$</td>
<td>7</td>
<td>1000</td>
<td>6480</td>
</tr>
<tr>
<td>$\operatorname{Irony18}$</td>
<td>Van Hee et al. (2018)</td>
<td>Irony</td>
<td>Twitter</td>
<td>4</td>
<td>3,834</td>
<td>784</td>
</tr>
<tr>
<td>$SCv1$</td>
<td>Lukin and Walker (2013)</td>
<td>Sarcasm</td>
<td>Debate Forums</td>
<td>2</td>
<td>1000</td>
<td>995</td>
</tr>
<tr>
<td>$SCv2$</td>
<td>Oraby et al. (2016)</td>
<td>Sarcasm</td>
<td>Debate ${\mathrm{e}}$Forums T</td>
<td>2</td>
<td>1000</td>
<td>2260</td>
</tr>
</tbody>
</table>
Table 2: Description of benchmark datasets. We split 10% of the train set to serve as the validation set.
<table>
<tbody>
<tr>
<th>Model</th>
<th>$SST-5$</th>
<th>Sent17</th>
<th>PhychExp</th>
<th>$\operatorname{Irony18}$</th>
<th>$SCv1$</th>
<th>$SCv2$</th>
</tr>
<tr>
<td>baseline</td>
<td>$43.5\pm0.5$</td>
<td>$68.3\pm0.2$</td>
<td>$53.2\pm0.8$</td>
<td>$46.3\pm1.4$</td>
<td>$64.1\pm0.5$</td>
<td>$74.0\pm0.7$</td>
</tr>
<tr>
<td>$emb.$ $\operatorname{conc}.$</td>
<td>$43.3\pm0.6$</td>
<td>$68.4\pm0.2$</td>
<td>$57.1\pm1.2$</td>
<td>$48.1\pm1.2$</td>
<td>$64.2\pm0.7$</td>
<td>$74.2\pm0.7$</td>
</tr>
<tr>
<td>conc.</td>
<td>$44.0\pm0.7$</td>
<td>$68.6\pm0.3$</td>
<td>$54.3\pm0.6$</td>
<td>$47.4\pm0.9$</td>
<td>$65.1\pm0.6$</td>
<td>$74.3\pm1.2$</td>
</tr>
<tr>
<td>gate</td>
<td>$44.2\pm0.4$</td>
<td>$68.7\pm0.3$</td>
<td>$53.4\pm1.0$</td>
<td>$48.5\pm0.7$</td>
<td>$64.7\pm0.7$</td>
<td>$74.3\pm1.2$</td>
</tr>
<tr>
<td>affine</td>
<td>$43.2\pm0.7$</td>
<td>$68.5\pm0.3$</td>
<td>$53.1\pm0.9$</td>
<td>$45.3\pm1.5$</td>
<td>$60.3\pm0.8$</td>
<td>$74.0\pm1.0$</td>
</tr>
<tr>
<td>gate+ emb. conc. </td>
<td>$46.2\pm0.5$</td>
<td>$68.9\pm0.3$</td>
<td>$57.2\pm1.1$</td>
<td>$48.4\pm1.0$</td>
<td>$64.9\pm0.6$</td>
<td>$74.4\pm0.9$</td>
</tr>
<tr>
<td>state-of-the-art</td>
<td>51.7 11 15</td>
<td>68.5</td>
<td>57.0 Yalk 17</td>
<td>53.6 Raziotis $n$</td>
<td>69.0 alko 24 $n$ /7017</td>
<td>76.0 1 $iA$ 21</td>
</tr>
<tr>
<td>Sldle UI-ulc-al</td>
<td>Shen et al. (2018)</td>
<td>Cliche (2017)</td>
<td>Felbo et al. (2017)</td>
<td>Baziotis et al. (2018</td>
<td>Felbo et al. (2017)</td>
<td>Ilić $et:$ al. .(2018</td>
</tr>
</tbody>
</table>
Table 3: Comparison across benchmark datasets. Reported values are averaged across ten runs. All reported
measures are $\hat{F}_1$ scores, apart from $SST-5$ which is evaluated with Accuracy .
杂的 NLP 任务上的性能，例如命名实体识别和序列标记，其中先验知识集成是一个活跃的研究领域
7
Acknowledgements
我们要感谢我们的
同事
AlexandraChronopoulou 和 Geor-
giosParaskevopoulos 的有益建议和意见。这项工作得到了国家 HPC 设施-ARIS 的希腊研究& 技术网络(GR-NET)授予的计算时间的部分支持。我们感谢 NVIDIA 通过捐赠TitanXGPU 来支持这项工作。
$Irony18$ , 基于特征的门控产生了最佳结果。最后，基于连接的条件是 SCv1 的顶级方法， 而SCv2 的组合方法
总的来说，基于注意力特征的门控是性能最好的调节方法，其次是串联。注意仿射变换表现不佳，特别是对于较小的数据集；这可能是由于该模型的容量较高。这是特别有趣的，因为门控(等式$\overline{\mathrm{E}}$4)是仿射变换方法(等式E5 )的特例，其中移位向量$\beta$为零，并且缩放向量 $\gamma$ 被限制在范围 [0,1] (等式$\overline{\mathrm{E}}6$ )。有趣的是，门控与传统嵌入层连接的结合为大多数任务提供了额外的性能提升，表明在各种调节方法中可以利用协同作用
除了性能改进之外，我们还可以直观地评估调节注意力分布对先验知识的影响，并提高我们方法的可解释性。正如我们在图 2 中看到的，词典特征引导模型关注更突出的单词，从而预测正确的类
# References
# 6 Conclusions &Future work
Sungjin Ahn, Heeyoul Choi, Tanel Pärnamaa, and
Yoshua Bengio.2016. A neural knowledge lan-
guage model. arXiv preprint arXiv:1608.00318 .
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben-
gio. 2015. Neural machine translation by jointly learning to align and translate. In Proceedings of the International Conference on Learning Representations .
Jorge Balazs and Yutaka Matsuo. 2019. Gating mecha-
nisms for combining character and word-level word representations: an empirical study. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Student Research Workshop , pages 110-
我们介绍了三种新颖的基于注意力的调节方法，并将其有效性与传统的基于串联的调节进行了比较。我们的方法简单而有效，可为所有数据集实现一致的性能改进。我们的方法可以作为额外的模块应用于任何基于 RNN 的架构， 以最小的计算开销进一步提高性能
将来，我们的目标是整合更精细的语言资源(例如。g。知识库),并研究我们的方法在更复
|
[
[
0.1259765625,
0.07598876953125,
0.177978515625,
0.07781982421875,
0.17626953125,
0.09027099609375,
0.124267578125,
0.0885009765625
],
[
0.76416015625,
0.07598876953125,
0.81201171875,
0.07781982421875,
0.81005859375,
0.0926513671875,
0.76318359375,
0.09027099609375
],
[
0.83447265625,
0.0736083984375,
0.87890625,
0.0789794921875,
0.875,
0.09442138671875,
0.8310546875,
0.08905029296875
],
[
0.50390625,
0.07659912109375,
0.5380859375,
0.07659912109375,
0.5380859375,
0.08905029296875,
0.50390625,
0.08905029296875
],
[
0.210693359375,
0.07781982421875,
0.2493896484375,
0.07781982421875,
0.2493896484375,
0.09027099609375,
0.210693359375,
0.09027099609375
],
[
0.5859375,
0.07781982421875,
0.63818359375,
0.07781982421875,
0.63818359375,
0.08905029296875,
0.5859375,
0.08905029296875
],
[
0.7001953125,
0.07781982421875,
0.75,
0.07781982421875,
0.75,
0.08905029296875,
0.7001953125,
0.08905029296875
],
[
0.124267578125,
0.0955810546875,
0.17041015625,
0.0955810546875,
0.17041015625,
0.10986328125,
0.124267578125,
0.10986328125
],
[
0.21240234375,
0.0968017578125,
0.337646484375,
0.0968017578125,
0.337646484375,
0.10986328125,
0.21240234375,
0.10986328125
],
[
0.771484375,
0.0968017578125,
0.8134765625,
0.0968017578125,
0.8134765625,
0.10986328125,
0.771484375,
0.10986328125
],
[
0.83642578125,
0.0968017578125,
0.87744140625,
0.0968017578125,
0.87744140625,
0.10986328125,
0.83642578125,
0.10986328125
],
[
0.58447265625,
0.097412109375,
0.68408203125,
0.097412109375,
0.68408203125,
0.1092529296875,
0.58447265625,
0.1092529296875
],
[
0.50390625,
0.097412109375,
0.57080078125,
0.097412109375,
0.57080078125,
0.1092529296875,
0.50390625,
0.1092529296875
],
[
0.73779296875,
0.097412109375,
0.75146484375,
0.097412109375,
0.75146484375,
0.1092529296875,
0.73779296875,
0.1092529296875
],
[
0.1259765625,
0.11102294921875,
0.17626953125,
0.1134033203125,
0.175537109375,
0.127685546875,
0.124267578125,
0.1258544921875
],
[
0.214111328125,
0.11224365234375,
0.355224609375,
0.1134033203125,
0.355224609375,
0.1270751953125,
0.214111328125,
0.1258544921875
],
[
0.76318359375,
0.11224365234375,
0.8154296875,
0.11224365234375,
0.8154296875,
0.1270751953125,
0.76318359375,
0.1270751953125
],
[
0.82861328125,
0.11224365234375,
0.87744140625,
0.11224365234375,
0.87744140625,
0.1270751953125,
0.82861328125,
0.1270751953125
],
[
0.58447265625,
0.114013671875,
0.6337890625,
0.114013671875,
0.6337890625,
0.1258544921875,
0.58447265625,
0.1258544921875
],
[
0.50390625,
0.114013671875,
0.57080078125,
0.114013671875,
0.57080078125,
0.1258544921875,
0.50390625,
0.1258544921875
],
[
0.73779296875,
0.114013671875,
0.75390625,
0.114013671875,
0.75390625,
0.12469482421875,
0.73779296875,
0.12469482421875
],
[
0.1259765625,
0.1312255859375,
0.1956787109375,
0.1312255859375,
0.1956787109375,
0.142578125,
0.1259765625,
0.142578125
],
[
0.210693359375,
0.12890625,
0.387939453125,
0.12890625,
0.387939453125,
0.142578125,
0.210693359375,
0.142578125
],
[
0.50390625,
0.1312255859375,
0.5615234375,
0.1312255859375,
0.5615234375,
0.142578125,
0.50390625,
0.142578125
],
[
0.5859375,
0.1312255859375,
0.6650390625,
0.1312255859375,
0.6650390625,
0.142578125,
0.5859375,
0.142578125
],
[
0.736328125,
0.1312255859375,
0.75146484375,
0.1312255859375,
0.75146484375,
0.142578125,
0.736328125,
0.142578125
],
[
0.77734375,
0.1312255859375,
0.8154296875,
0.1312255859375,
0.8154296875,
0.1436767578125,
0.77734375,
0.1436767578125
],
[
0.8388671875,
0.1312255859375,
0.880859375,
0.1312255859375,
0.880859375,
0.144287109375,
0.8388671875,
0.144287109375
],
[
0.58447265625,
0.144287109375,
0.63623046875,
0.146728515625,
0.634765625,
0.160888671875,
0.58251953125,
0.1591796875
],
[
0.8447265625,
0.144287109375,
0.880859375,
0.1478271484375,
0.87744140625,
0.162109375,
0.84228515625,
0.1591796875
],
[
0.50146484375,
0.1455078125,
0.54248046875,
0.1478271484375,
0.54150390625,
0.162109375,
0.49951171875,
0.1602783203125
],
[
0.1234130859375,
0.146728515625,
0.1822509765625,
0.146728515625,
0.1822509765625,
0.160888671875,
0.1234130859375,
0.160888671875
],
[
0.21240234375,
0.1455078125,
0.345947265625,
0.1455078125,
0.345947265625,
0.1591796875,
0.21240234375,
0.1591796875
],
[
0.73486328125,
0.146728515625,
0.75390625,
0.146728515625,
0.75390625,
0.1602783203125,
0.73486328125,
0.1602783203125
],
[
0.7734375,
0.1478271484375,
0.8154296875,
0.1478271484375,
0.8154296875,
0.160888671875,
0.7734375,
0.160888671875
],
[
0.1259765625,
0.1644287109375,
0.16796875,
0.1644287109375,
0.16796875,
0.1776123046875,
0.1259765625,
0.1776123046875
],
[
0.21240234375,
0.1644287109375,
0.370361328125,
0.1644287109375,
0.370361328125,
0.1776123046875,
0.21240234375,
0.1776123046875
],
[
0.73486328125,
0.163330078125,
0.75244140625,
0.163330078125,
0.75244140625,
0.1787109375,
0.73486328125,
0.1787109375
],
[
0.77734375,
0.1644287109375,
0.8154296875,
0.1644287109375,
0.8154296875,
0.177001953125,
0.77734375,
0.177001953125
],
[
0.84619140625,
0.1644287109375,
0.87744140625,
0.1644287109375,
0.87744140625,
0.177001953125,
0.84619140625,
0.177001953125
],
[
0.5859375,
0.1650390625,
0.6826171875,
0.1650390625,
0.6826171875,
0.177001953125,
0.5859375,
0.177001953125
],
[
0.50390625,
0.1650390625,
0.5615234375,
0.1650390625,
0.5615234375,
0.177001953125,
0.50390625,
0.177001953125
],
[
0.50390625,
0.1787109375,
0.56103515625,
0.1817626953125,
0.55908203125,
0.196533203125,
0.5029296875,
0.193603515625
],
[
0.5859375,
0.1787109375,
0.6826171875,
0.18115234375,
0.6826171875,
0.19482421875,
0.5859375,
0.1923828125
],
[
0.73486328125,
0.179931640625,
0.75244140625,
0.179931640625,
0.75244140625,
0.19482421875,
0.73486328125,
0.19482421875
],
[
0.214111328125,
0.179931640625,
0.334228515625,
0.179931640625,
0.334228515625,
0.193603515625,
0.214111328125,
0.193603515625
],
[
0.1275634765625,
0.18115234375,
0.166259765625,
0.18115234375,
0.166259765625,
0.193603515625,
0.1275634765625,
0.193603515625
],
[
0.77734375,
0.18115234375,
0.8154296875,
0.18115234375,
0.8154296875,
0.193603515625,
0.77734375,
0.193603515625
],
[
0.8388671875,
0.18115234375,
0.8759765625,
0.18115234375,
0.8759765625,
0.193603515625,
0.8388671875,
0.193603515625
],
[
0.15283203125,
0.2078857421875,
0.84814453125,
0.210205078125,
0.84814453125,
0.223876953125,
0.15283203125,
0.221435546875
],
[
0.1796875,
0.240478515625,
0.2275390625,
0.240478515625,
0.2275390625,
0.254638671875,
0.1796875,
0.254638671875
],
[
0.447509765625,
0.2381591796875,
0.51318359375,
0.240478515625,
0.51123046875,
0.254638671875,
0.44580078125,
0.2529296875
],
[
0.57275390625,
0.240478515625,
0.62744140625,
0.240478515625,
0.62744140625,
0.254638671875,
0.57275390625,
0.254638671875
],
[
0.35693359375,
0.2410888671875,
0.4013671875,
0.2410888671875,
0.4013671875,
0.2529296875,
0.35693359375,
0.2529296875
],
[
0.26123046875,
0.2410888671875,
0.30322265625,
0.2410888671875,
0.30322265625,
0.254638671875,
0.26123046875,
0.254638671875
],
[
0.7021484375,
0.2410888671875,
0.74072265625,
0.2410888671875,
0.74072265625,
0.254150390625,
0.7021484375,
0.254150390625
],
[
0.8095703125,
0.2410888671875,
0.84814453125,
0.2410888671875,
0.84814453125,
0.254150390625,
0.8095703125,
0.254150390625
],
[
0.172119140625,
0.260009765625,
0.22412109375,
0.260009765625,
0.22412109375,
0.271484375,
0.172119140625,
0.271484375
],
[
0.2476806640625,
0.260009765625,
0.31494140625,
0.260009765625,
0.31494140625,
0.271484375,
0.2476806640625,
0.271484375
],
[
0.345947265625,
0.260009765625,
0.4130859375,
0.260009765625,
0.4130859375,
0.271484375,
0.345947265625,
0.271484375
],
[
0.44580078125,
0.260009765625,
0.5146484375,
0.260009765625,
0.5146484375,
0.271484375,
0.44580078125,
0.271484375
],
[
0.568359375,
0.260009765625,
0.634765625,
0.260009765625,
0.634765625,
0.271484375,
0.568359375,
0.271484375
],
[
0.6884765625,
0.260009765625,
0.755859375,
0.260009765625,
0.755859375,
0.271484375,
0.6884765625,
0.271484375
],
[
0.7958984375,
0.260009765625,
0.86376953125,
0.260009765625,
0.86376953125,
0.271484375,
0.7958984375,
0.271484375
],
[
0.2476806640625,
0.275634765625,
0.316650390625,
0.275634765625,
0.316650390625,
0.287353515625,
0.2476806640625,
0.287353515625
],
[
0.344970703125,
0.275634765625,
0.4130859375,
0.275634765625,
0.4130859375,
0.287353515625,
0.344970703125,
0.287353515625
],
[
0.44580078125,
0.275634765625,
0.5146484375,
0.275634765625,
0.5146484375,
0.287353515625,
0.44580078125,
0.287353515625
],
[
0.56689453125,
0.275634765625,
0.634765625,
0.275634765625,
0.634765625,
0.287353515625,
0.56689453125,
0.287353515625
],
[
0.6884765625,
0.275634765625,
0.755859375,
0.275634765625,
0.755859375,
0.287353515625,
0.6884765625,
0.287353515625
],
[
0.791015625,
0.2724609375,
0.86376953125,
0.27490234375,
0.86376953125,
0.289306640625,
0.791015625,
0.287353515625
],
[
0.443359375,
0.289306640625,
0.515625,
0.29150390625,
0.515625,
0.305908203125,
0.443359375,
0.303955078125
],
[
0.1561279296875,
0.276611328125,
0.2225341796875,
0.276611328125,
0.2225341796875,
0.2880859375,
0.1561279296875,
0.2880859375
],
[
0.2476806640625,
0.292236328125,
0.31494140625,
0.292236328125,
0.31494140625,
0.303955078125,
0.2476806640625,
0.303955078125
],
[
0.344970703125,
0.292236328125,
0.414794921875,
0.292236328125,
0.414794921875,
0.303955078125,
0.344970703125,
0.303955078125
],
[
0.568359375,
0.292236328125,
0.634765625,
0.292236328125,
0.634765625,
0.303955078125,
0.568359375,
0.303955078125
],
[
0.6884765625,
0.292236328125,
0.755859375,
0.292236328125,
0.755859375,
0.303955078125,
0.6884765625,
0.303955078125
],
[
0.7958984375,
0.292236328125,
0.86376953125,
0.292236328125,
0.86376953125,
0.303955078125,
0.7958984375,
0.303955078125
],
[
0.444091796875,
0.30712890625,
0.5146484375,
0.30712890625,
0.5146484375,
0.32177734375,
0.444091796875,
0.32177734375
],
[
0.18896484375,
0.29345703125,
0.22412109375,
0.29345703125,
0.22412109375,
0.3046875,
0.18896484375,
0.3046875
],
[
0.19482421875,
0.308837890625,
0.225830078125,
0.308837890625,
0.225830078125,
0.32177734375,
0.19482421875,
0.32177734375
],
[
0.2476806640625,
0.308837890625,
0.31494140625,
0.308837890625,
0.31494140625,
0.320556640625,
0.2476806640625,
0.320556640625
],
[
0.345947265625,
0.308837890625,
0.414794921875,
0.308837890625,
0.414794921875,
0.320556640625,
0.345947265625,
0.320556640625
],
[
0.56689453125,
0.30712890625,
0.63623046875,
0.30712890625,
0.63623046875,
0.32177734375,
0.56689453125,
0.32177734375
],
[
0.6884765625,
0.308837890625,
0.755859375,
0.308837890625,
0.755859375,
0.320556640625,
0.6884765625,
0.320556640625
],
[
0.7958984375,
0.308837890625,
0.86376953125,
0.308837890625,
0.86376953125,
0.320556640625,
0.7958984375,
0.320556640625
],
[
0.18896484375,
0.32470703125,
0.22412109375,
0.32470703125,
0.22412109375,
0.33740234375,
0.18896484375,
0.33740234375
],
[
0.2493896484375,
0.32470703125,
0.31494140625,
0.32470703125,
0.31494140625,
0.336181640625,
0.2493896484375,
0.336181640625
],
[
0.344970703125,
0.32373046875,
0.414794921875,
0.32373046875,
0.414794921875,
0.33837890625,
0.344970703125,
0.33837890625
],
[
0.44580078125,
0.32470703125,
0.51318359375,
0.32470703125,
0.51318359375,
0.336181640625,
0.44580078125,
0.336181640625
],
[
0.568359375,
0.32470703125,
0.634765625,
0.32470703125,
0.634765625,
0.336181640625,
0.568359375,
0.336181640625
],
[
0.6884765625,
0.32470703125,
0.755859375,
0.32470703125,
0.755859375,
0.336181640625,
0.6884765625,
0.336181640625
],
[
0.7958984375,
0.32470703125,
0.86376953125,
0.32470703125,
0.86376953125,
0.336181640625,
0.7958984375,
0.336181640625
],
[
0.1259765625,
0.33837890625,
0.225830078125,
0.34033203125,
0.225830078125,
0.355224609375,
0.1259765625,
0.352783203125
],
[
0.2476806640625,
0.341552734375,
0.31494140625,
0.341552734375,
0.31494140625,
0.352783203125,
0.2476806640625,
0.352783203125
],
[
0.343505859375,
0.34033203125,
0.41650390625,
0.34033203125,
0.41650390625,
0.355224609375,
0.343505859375,
0.355224609375
],
[
0.44580078125,
0.341552734375,
0.51318359375,
0.341552734375,
0.51318359375,
0.352783203125,
0.44580078125,
0.352783203125
],
[
0.56689453125,
0.341552734375,
0.634765625,
0.341552734375,
0.634765625,
0.352783203125,
0.56689453125,
0.352783203125
],
[
0.68701171875,
0.341552734375,
0.755859375,
0.341552734375,
0.755859375,
0.352783203125,
0.68701171875,
0.352783203125
],
[
0.79443359375,
0.341552734375,
0.86376953125,
0.341552734375,
0.86376953125,
0.352783203125,
0.79443359375,
0.352783203125
],
[
0.264404296875,
0.35693359375,
0.297119140625,
0.35693359375,
0.297119140625,
0.369384765625,
0.264404296875,
0.369384765625
],
[
0.36279296875,
0.35693359375,
0.3955078125,
0.35693359375,
0.3955078125,
0.369384765625,
0.36279296875,
0.369384765625
],
[
0.46337890625,
0.35693359375,
0.496337890625,
0.35693359375,
0.496337890625,
0.369384765625,
0.46337890625,
0.369384765625
],
[
0.58447265625,
0.35693359375,
0.6171875,
0.35693359375,
0.6171875,
0.369384765625,
0.58447265625,
0.369384765625
],
[
0.7060546875,
0.35693359375,
0.73779296875,
0.35693359375,
0.73779296875,
0.369384765625,
0.7060546875,
0.369384765625
],
[
0.8134765625,
0.35693359375,
0.8447265625,
0.35693359375,
0.8447265625,
0.369384765625,
0.8134765625,
0.369384765625
],
[
0.133544921875,
0.364013671875,
0.225830078125,
0.364013671875,
0.225830078125,
0.3759765625,
0.133544921875,
0.3759765625
],
[
0.22412109375,
0.372314453125,
0.87890625,
0.372314453125,
0.87890625,
0.385986328125,
0.22412109375,
0.385986328125
],
[
0.118408203125,
0.399658203125,
0.8017578125,
0.399658203125,
0.8017578125,
0.413330078125,
0.118408203125,
0.413330078125
],
[
0.7958984375,
0.399658203125,
0.88330078125,
0.40087890625,
0.88330078125,
0.412109375,
0.7958984375,
0.410888671875
],
[
0.118408203125,
0.413330078125,
0.66015625,
0.413330078125,
0.66015625,
0.427490234375,
0.118408203125,
0.427490234375
],
[
0.1175537109375,
0.45361328125,
0.490234375,
0.453125,
0.490234375,
0.46728515625,
0.1175537109375,
0.468505859375
],
[
0.5146484375,
0.453125,
0.8818359375,
0.45361328125,
0.8818359375,
0.468505859375,
0.5146484375,
0.46728515625
],
[
0.12005615234375,
0.47021484375,
0.490234375,
0.47021484375,
0.490234375,
0.48388671875,
0.12005615234375,
0.48388671875
],
[
0.5146484375,
0.47021484375,
0.8818359375,
0.47021484375,
0.8818359375,
0.48388671875,
0.5146484375,
0.48388671875
],
[
0.12005615234375,
0.486328125,
0.287109375,
0.486328125,
0.287109375,
0.49951171875,
0.12005615234375,
0.49951171875
],
[
0.5146484375,
0.486328125,
0.591796875,
0.486328125,
0.591796875,
0.50048828125,
0.5146484375,
0.50048828125
],
[
0.139404296875,
0.5029296875,
0.48779296875,
0.5029296875,
0.48779296875,
0.5166015625,
0.139404296875,
0.5166015625
],
[
0.12005615234375,
0.5185546875,
0.489501953125,
0.5185546875,
0.489501953125,
0.533203125,
0.12005615234375,
0.533203125
],
[
0.5146484375,
0.515625,
0.5283203125,
0.515625,
0.5283203125,
0.525390625,
0.5146484375,
0.525390625
],
[
0.12005615234375,
0.53515625,
0.48779296875,
0.53515625,
0.48779296875,
0.548828125,
0.12005615234375,
0.548828125
],
[
0.515625,
0.54052734375,
0.6689453125,
0.54052734375,
0.6689453125,
0.55419921875,
0.515625,
0.55419921875
],
[
0.6630859375,
0.54443359375,
0.68603515625,
0.54443359375,
0.68603515625,
0.5517578125,
0.6630859375,
0.5517578125
],
[
0.68603515625,
0.54150390625,
0.8818359375,
0.54150390625,
0.8818359375,
0.5546875,
0.68603515625,
0.5546875
],
[
0.12176513671875,
0.5517578125,
0.48779296875,
0.5517578125,
0.48779296875,
0.5654296875,
0.12176513671875,
0.5654296875
],
[
0.5146484375,
0.5556640625,
0.8427734375,
0.55712890625,
0.8427734375,
0.57177734375,
0.5146484375,
0.57080078125
],
[
0.83642578125,
0.55712890625,
0.8818359375,
0.55859375,
0.880859375,
0.57080078125,
0.83447265625,
0.568359375
],
[
0.12005615234375,
0.56640625,
0.48779296875,
0.56640625,
0.48779296875,
0.580078125,
0.12005615234375,
0.580078125
],
[
0.51318359375,
0.572265625,
0.88330078125,
0.572265625,
0.88330078125,
0.5859375,
0.51318359375,
0.5859375
],
[
0.12176513671875,
0.58447265625,
0.48779296875,
0.58447265625,
0.48779296875,
0.59716796875,
0.12176513671875,
0.59716796875
],
[
0.5146484375,
0.5888671875,
0.88330078125,
0.5888671875,
0.88330078125,
0.6025390625,
0.5146484375,
0.6025390625
],
[
0.12176513671875,
0.599609375,
0.48779296875,
0.599609375,
0.48779296875,
0.61328125,
0.12176513671875,
0.61328125
],
[
0.51318359375,
0.60400390625,
0.8818359375,
0.60498046875,
0.8818359375,
0.619140625,
0.51318359375,
0.6181640625
],
[
0.118408203125,
0.6142578125,
0.486083984375,
0.6142578125,
0.486083984375,
0.62744140625,
0.118408203125,
0.62744140625
],
[
0.5146484375,
0.62060546875,
0.88330078125,
0.62060546875,
0.88330078125,
0.63525390625,
0.5146484375,
0.63525390625
],
[
0.118408203125,
0.63134765625,
0.484375,
0.63134765625,
0.484375,
0.64453125,
0.118408203125,
0.64453125
],
[
0.5146484375,
0.63720703125,
0.744140625,
0.63720703125,
0.744140625,
0.65087890625,
0.5146484375,
0.65087890625
],
[
0.118408203125,
0.646484375,
0.30322265625,
0.64794921875,
0.30322265625,
0.662109375,
0.118408203125,
0.6611328125
],
[
0.1376953125,
0.66455078125,
0.48779296875,
0.66455078125,
0.48779296875,
0.677734375,
0.1376953125,
0.677734375
],
[
0.12005615234375,
0.6796875,
0.48779296875,
0.6796875,
0.48779296875,
0.693359375,
0.12005615234375,
0.693359375
],
[
0.5146484375,
0.681640625,
0.609375,
0.681640625,
0.609375,
0.693359375,
0.5146484375,
0.693359375
],
[
0.12005615234375,
0.6953125,
0.486083984375,
0.6953125,
0.486083984375,
0.71044921875,
0.12005615234375,
0.71044921875
],
[
0.51318359375,
0.703125,
0.884765625,
0.703125,
0.884765625,
0.71630859375,
0.51318359375,
0.71630859375
],
[
0.12005615234375,
0.7119140625,
0.48779296875,
0.7119140625,
0.48779296875,
0.7255859375,
0.12005615234375,
0.7255859375
],
[
0.5322265625,
0.71630859375,
0.8818359375,
0.71630859375,
0.8818359375,
0.72998046875,
0.5322265625,
0.72998046875
],
[
0.12005615234375,
0.72802734375,
0.2493896484375,
0.72802734375,
0.2493896484375,
0.7421875,
0.12005615234375,
0.7421875
],
[
0.52880859375,
0.728515625,
0.85986328125,
0.72802734375,
0.85986328125,
0.7421875,
0.52880859375,
0.74365234375
],
[
0.1259765625,
0.7548828125,
0.1376953125,
0.76318359375,
0.1259765625,
0.771484375,
0.11419677734375,
0.76318359375
],
[
0.15283203125,
0.75732421875,
0.3955078125,
0.75732421875,
0.3955078125,
0.77099609375,
0.15283203125,
0.77099609375
],
[
0.5146484375,
0.75439453125,
0.87890625,
0.75439453125,
0.87890625,
0.76416015625,
0.5146484375,
0.76416015625
],
[
0.53076171875,
0.7666015625,
0.8818359375,
0.765625,
0.8818359375,
0.77978515625,
0.53076171875,
0.78076171875
],
[
0.12005615234375,
0.783203125,
0.484375,
0.783203125,
0.484375,
0.79345703125,
0.12005615234375,
0.79345703125
],
[
0.53076171875,
0.77978515625,
0.884765625,
0.77978515625,
0.884765625,
0.79345703125,
0.53076171875,
0.79345703125
],
[
0.118408203125,
0.798828125,
0.486083984375,
0.798828125,
0.486083984375,
0.8125,
0.118408203125,
0.8125
],
[
0.53076171875,
0.79345703125,
0.8818359375,
0.79345703125,
0.8818359375,
0.80712890625,
0.53076171875,
0.80712890625
],
[
0.5322265625,
0.80712890625,
0.587890625,
0.80712890625,
0.587890625,
0.818359375,
0.5322265625,
0.818359375
],
[
0.118408203125,
0.81396484375,
0.486083984375,
0.8134765625,
0.486083984375,
0.82763671875,
0.118408203125,
0.8291015625
],
[
0.12005615234375,
0.83056640625,
0.489501953125,
0.83056640625,
0.489501953125,
0.84423828125,
0.12005615234375,
0.84423828125
],
[
0.51123046875,
0.830078125,
0.8818359375,
0.830078125,
0.8818359375,
0.84326171875,
0.51123046875,
0.84326171875
],
[
0.12005615234375,
0.8466796875,
0.489501953125,
0.8466796875,
0.489501953125,
0.85986328125,
0.12005615234375,
0.85986328125
],
[
0.5322265625,
0.84423828125,
0.88330078125,
0.84423828125,
0.88330078125,
0.85498046875,
0.5322265625,
0.85498046875
],
[
0.53076171875,
0.85546875,
0.88330078125,
0.85498046875,
0.88330078125,
0.869140625,
0.53076171875,
0.87060546875
],
[
0.12005615234375,
0.86328125,
0.397216796875,
0.86328125,
0.397216796875,
0.876953125,
0.12005615234375,
0.876953125
],
[
0.53076171875,
0.869140625,
0.88330078125,
0.869140625,
0.88330078125,
0.8828125,
0.53076171875,
0.8828125
],
[
0.1368408203125,
0.87890625,
0.48779296875,
0.87890625,
0.48779296875,
0.892578125,
0.1368408203125,
0.892578125
],
[
0.53076171875,
0.8828125,
0.88330078125,
0.8828125,
0.88330078125,
0.896484375,
0.53076171875,
0.896484375
],
[
0.12005615234375,
0.8955078125,
0.486083984375,
0.8955078125,
0.486083984375,
0.91015625,
0.12005615234375,
0.91015625
],
[
0.52880859375,
0.8955078125,
0.88330078125,
0.89453125,
0.88330078125,
0.9091796875,
0.52880859375,
0.91015625
]
] |
[
"Dataset",
"Ntrain",
"Ntest",
"Task",
"Study",
"Domain",
"Classes",
"SST-5",
"Socher et al. (2013)",
"9,645",
"2,210",
"Movie Reviews",
"Sentiment",
"5",
"Sent17",
"Rosenthal et al. (2017)",
"49,570",
"12,284",
"Twitter",
"Sentiment",
"3",
"PhychExp",
"Wallbott and Scherer (1986)",
"Emotion",
"Experiences",
"7",
"1000",
"6480",
"Twitter",
"784",
"Irony",
"Irony18",
"Van Hee et al. (2018)",
"4",
"3,834",
" SCv1",
"Lukin and Walker (2013)",
"2",
"1000",
"995",
"Debate Forums",
"Sarcasm",
"Sarcasm",
"Debate Forums",
"2",
"Oraby et al. (2016)",
"SCv2",
"1000",
"2260",
"Table 2: Description of benchmark datasets. We split 10 % of the train set to serve as the validation set.",
"Model",
"PhychExp",
"Irony18",
"Sent17",
"SST-5",
"SCv1",
"SCv2",
"baseline",
"43.5±0.5",
"68.3±0.2",
"53.2 ±0.8",
"46.3±1.4",
"64.1 ± 0.5",
"74.0 ± 0.7",
"43.3±0.6",
"68.4 ± 0.2",
"57.1 ± 1.2",
"48.1 ± 1.2",
"64.2 ± 0.7",
"74.2±0.7",
"54.3 ± 0.6",
"emb.conc.",
"44.0±0.7",
"68.6±0.3",
"47.4± 0.9",
"65.1 ±0.6",
"74.3 ±1.2",
"53.4 ± 1.0",
"conc.",
"gate",
"44.2±0.4",
"68.7 ±0.3",
"48.5 ± 0.7",
"64.7 ± 0.7",
"74.3 ± 1.2",
"affine",
"43.2±0.7",
"68.5 ± 0.3",
"53.1±0.9",
"45.3±1.5",
"60.3±0.8",
"74.0±1.0",
"gate+emb.conc.",
"46.2±0.5",
"68.9 ± 0.3",
"57.2 ± 1.1",
"48.4±1.0",
"64.9 ± 0.6",
"74.4±0.9",
"51.7",
"68.5",
"57.0",
"53.6",
"69.0",
"76.0",
"state-of-the-art",
"Shen et al. (2018) Cliche (2017) Felbo et al. (2017) Baziotis et al. (2018) Felbo et al. (2017) Ilic et al. (2018)",
"Table 3: Comparison across benchmark datasets. Reported values are averaged across ten runs.",
"All reported",
"measures are Fi scores, apart from SST- 5 which is evaluated with Accuracy .",
"Irony18，基于特征的门控产生了最佳结果。",
"杂的NLP任务上的性能，例如命名实体识别",
"最后，基于连接的条件是SCu1的顶级方法，",
"和序列标记，其中先验知识集成是一个活跃的",
"而SC2的组合方法",
"研究领域",
"总的来说，基于注意力特征的门控是性能最",
"好的调节方法，其次是串联。注意仿射变换表",
"7",
"现不佳，特别是对于较小的数据集；这可能是",
"Acknowledgements",
"",
"我们要感谢我们的",
"由于该模型的容量较高。这是特别有趣的，因",
"同事AlexandraChronopoulou和",
"Geor-",
"为门控（等式E4）是仿射变换方法（等式E5",
"giosParaskevopoulos的有益建议和意见。",
"）的特例，其中移位向量β为零，并且缩放向",
"这项工作得到了国家HPC设施-ARIS的希腊",
"量被限制在范围[0,1]（等式E6）。有趣的",
"研究&技术网络（GR-NET）授予的计算时",
"是，门控与传统嵌入层连接的结合为大多数任",
"间的部分支持。我们感谢NVIDIA通过捐赠",
"务提供了额外的性能提升，表明在各种调节方",
"TitanXGPU来支持这项工作。",
"法中可以利用协同作用",
"除了性能改进之外，我们还可以直观地评估",
"调节注意力分布对先验知识的影响，并提高我",
"References",
"们方法的可解释性。正如我们在图2中看到",
"Sungjin Ahn, Heeyoul Choi, Tanel Parnamaa, and",
"的，词典特征引导模型关注更突出的单词，从",
"Yoshua Bengio. 2016. A neural knowledge lan-",
"而预测正确的类",
"guage model. arXiv preprint arXiv:1608.00318 .",
"6",
"Conclusions & Future work",
"DzmitryBahdanau.Kvunghyun Cho.and YoshuaBen",
"gio. 2015. Neural machine translation by jointly",
"我们介绍了三种新颖的基于注意力的调节方",
"learning to align and translate. In Proceedings of",
"法，并将其有效性与传统的基于串联的调节进",
"the International Conference on Learning Represen-",
"tations.",
"行了比较。我们的方法简单而有效，可为所有",
"数据集实现一致的性能改进。我们的方法可以",
" Jorge Balazs and Yutaka Matsuo. 2019. Gating mecha-",
"作为额外的模块应用于任何基于RNN的架构，",
"nisms for combining character and word-level word",
"representations: an empirical study.In Proceed-",
"以最小的计算开销进一步提高性能",
"ings of the 2019 Conference of the North American",
"将来，我们的目标是整合更精细的语言资源",
"Chapter of the Association for Computational Lin-",
"（例如。g。知识库），并研究我们的方法在更复",
" guistics: Student Research Workshop , pages 110-"
] |
{
"arxiv_id": "1906.03674",
"language": "zh",
"page_idx": 3
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.