Maximum Classifier Discrepancy for Unsupervised Domain Adaptation

Maximum Classifier Discrepancy for Unsupervised Domain Adaptation

无监督域自适应的最大分类器差异

Abstract

In this work, we present a method for unsupervised domain adaptation. Many adversarial learning methods train domain classifier networks to distinguish the features as either a source or target and train a feature generator network to mimic the discriminator. Two problems exist with these methods. First, the domain classifier only tries to distinguish the features as a source or target and thus does not consider task-specific decision boundaries between classes.Therefore, a trained generator can generate ambiguous features near class boundaries. Second, these methods aim to completely match the feature distributions between different domains, which is difficult because of each domain’s characteristics.
在这项工作中，我们提出了一种无监督域自适应的方法。 许多对抗性学习方法训练域分类器网络以将特征区分为源还是目标，并训练特征生成器网络以模仿鉴别器。 这些方法存在两个问题。 首先，领域分类器仅尝试将特征区分为源或目标，因此不考虑类之间的特定于任务的决策边界，因此，训练有素的生成器可以在类边界附近生成模糊特征。 其次，这些方法旨在完全匹配不同域之间的特征分布，由于每个域的特性，这很困难。

To solve these problems, we introduce a new approach that attempts to align distributions of source and target by utilizing the task-specific decision boundaries. We propose to maximize the discrepancy between two classifiers’ outputs to detect target samples that are far from the support of the source. A feature generator learns to generate target features near the support to minimize the discrepancy. Our method outperforms other methods on several datasets of image classification and semantic segmentation. The codes are available at https://github.com/mil-tokyo/MCD_DA
为了解决这些问题，我们引入了一种新方法，该方法试图通过利用特定于任务的决策边界来对齐源和目标的分布。 我们建议最大程度地提高两个分类器输出之间的差异，以检测远离来源支持的目标样本。 特征生成器学习如何在支撑附近生成目标特征，以最大程度地减少差异。 我们的方法在图像分类和语义分割的几个数据集上优于其他方法。 可以在https://github.com/mil-tokyo/MCD_DA上找到这些代码。

Introduction

The classification accuracy of images has improved substantially with the advent of deep convolutional neural networks (CNN) which utilize numerous labeled samples [16].However, collecting numerous labeled samples in various domains is expensive and time-consuming.
深度度卷积神经网络（CNN）的出现极大地提高了图像的分类精度，该网络利用了许多标记的样本[16]。然而，在各个领域收集大量的标记样本既昂贵又费时。

Domain adaptation (DA) tackles this problem by transferring knowledge from a label-rich domain (i.e., source domain) to a label-scarce domain (i.e., target domain). DA aims to train a classifier using source samples that generalize well to the target domain. However, each domain’s samples have different characteristics, which makes the problem difficult to solve. Consider neural networks trained on labeled source images collected from the Web. Although such neural networks perform well on the source images,correctly recognizing target images collected from a real camera is difficult for them. This is because the target images can have different characteristics from the source images, such as change of light, noise, and angle in which the image is captured. Furthermore, regarding unsupervised DA (UDA), we have access to labeled source samples and only unlabeled target samples. We must construct a model that works well on target samples despite the absence of their labels during training. UDA is the most challenging situation, and we propose a method for UDA in this study.
域适应（DA）通过将知识从标签丰富的域（即源域）转移到标签稀缺的域（即目标域）来解决此问题。 DA的目标是使用能够很好地推广到目标领域的源样本来训练分类器。但是，每个域的样本具有不同的特征，这使得问题难以解决。考虑对从网络收集的带标签源图像进行训练的神经网络。尽管这样的神经网络在源图像上表现良好，但是对于它们而言，正确识别从真实相机收集的目标图像却很困难。这是因为目标图像可能具有与源图像不同的特性，例如光，噪声和捕获图像的角度的变化。此外，关于无监督DA（UDA），我们可以访问标记的源样本，而只能访问未标记的目标样本。尽管训练过程中没有标签，但我们必须构建一个对目标样本运行良好的模型。 UDA是最具挑战性的情况，我们在本研究中提出了一种UDA方法。

Many UDA algorithms, particularly those for training neural networks, attempt to match the distribution of the source features with that of the target without considering the category of the samples [8, 37, 4, 40]. In particular, domain classifier-based adaptation algorithms have been applied to many tasks [8, 4]. The methods utilize two players to align distributions in an adversarial manner: domain classifier (i.e., a discriminator) and feature generator. Source and target samples are input to the same feature generator.
许多UDA算法，尤其是那些用于训练神经网络的算法，都试图在不考虑样本类别的情况下将源特征的分布与目标的分布进行匹配[8，37，4，40]。 特别是，基于域分类器的自适应算法已应用于许多任务[8，4]。 该方法利用两个参与者以对抗的方式对齐分布：域分类器（即，鉴别器）和特征生成器。 源样本和目标样本被输入到同一特征生成器。
Features from the feature generator are shared by the discriminator and a task-specific classifier. The discriminator is trained to discriminate the domain labels of the features generated by the generator whereas the generator is trained to fool it. The generator aims to match distributions between the source and target because such distributions will mimic the discriminator. They assume that such target features are classified correctly by the task-specific classifier because they are aligned with the source samples.
来自特征生成器的特征由鉴别器和特定于任务的分类器共享。 鉴别器受过训练以区分由生成器生成的特征的域标签，而生成器则受过欺骗。 生成器旨在匹配源和目标之间的分布，因为这样的分布将模仿鉴别器。 他们认为此类目标特征已由任务特定的分类器正确分类，因为它们与源样本对齐。

However, this method should fail to extract discriminative features because it does not consider the relationship between target samples and the task-specific decision boundary when aligning distributions. As shown in the left side of Fig. 1, the generator can generate ambiguous features near the boundary because it simply tries to make the two distributions similar.
但是，此方法应该无法提取区分特征，因为在对齐分布时，它不会考虑目标样本与特定于任务的决策边界之间的关系。 如图1左侧所示，生成器可以在边界附近生成歧义特征，因为它只是试图使两个分布相似。
图1.（以最佳视图显示。）以前的和建议的分布匹配方法的比较。左：以前的方法试图通过模仿域分类器来匹配不同的分布。 他们不考虑决策边界。 正确：我们提出的方法尝试使用特定于任务的分类器来检测源分布范围之外的目标样本。
To overcome both problems, we propose to align distributions of features from source and target domain by using the classifier’s output for the target samples.
为了克服这两个问题，我们建议通过使用目标样本的分类器输出来调整源域和目标域中要素的分布。

We introduce a new adversarial learning method that utilizes two types of players: task-specific classifiers and a feature generator. task-specific classifiers denotes the classifiers trained for each task such as object classification or semantic segmentation. Two classifiers take features from the generator. Two classifiers try to classify source samples correctly and, simultaneously, are trained to detect the target samples that are far from the support of the source. The samples existing far from the support do not have discriminative features because they are not clearly categorized into some classes. Thus, our method utilizes the task-specific classifiers as a discriminator. Generator tries to fool the classifiers. In other words, it is trained to generate target features near the support while considering classifiers’ output for target samples. Thus, our method allows the generator to generate discriminative features for target samples because it considers the relationship between the decision boundary and target samples. This training is achieved in an adversarial manner. In addition, please note that we do not use domain labels in our method.
我们介绍一种新的对抗学习方法，该方法利用两种类型的参与者：特定于任务的分类器和特征生成器。特定于任务的分类器表示针对每个任务（例如对象分类或语义分段）训练的分类器。两个分类器从生成器中获取特征。两个分类器试图正确分类源样本，同时训练它们以检测远离源支持的目标样本。距离支持者较远的样本没有判别功能，因为它们没有明确分类为某些类别。因此，我们的方法利用特定于任务的分类器作为区分器。 Generator试图欺骗分类器。换句话说，在考虑分类器针对目标样本的输出的同时，训练它可以在支持附近生成目标特征。因此，我们的方法允许生成器生成目标样本的判别特征，因为它考虑了决策边界和目标样本之间的关系。该训练是对抗性的。此外，请注意，我们在方法中不使用域标签。

We evaluate our method on image recognition and semantic segmentation. In many settings, our method outperforms other methods by a large margin. The contributions of our paper are summarized as follows:
我们评估图像识别和语义分割的方法。 在许多情况下，我们的方法在很大程度上优于其他方法。 本文的贡献总结如下：
• We propose a novel adversarial training method for domain adaptation that tries to align the distribution of a target domain by considering task-specific decision boundaries.
• We confirm the behavior of our method through a toy problem.
• We extensively evaluate our method on various tasks:digit classification, object classification, and semantic segmentation.
•我们提出了一种新颖的域适应对抗训练方法，该方法试图通过考虑特定于任务的决策边界来调整目标域的分布。
•我们通过玩具问题来确认我们方法的行为。
•我们在各种任务上广泛评估我们的方法：数字分类，对象分类和语义分割。

Related Work

Training CNN for DA can be realized through various strategies. Ghifary et al. proposed using an autoencoder for the target domain to obtain domain-invariant features [9]. Sener et al. proposed using clustering techniques and pseudo-labels to obtain discriminative features [33].Taigman et al. proposed cross-domain image translation methods [38]. Matching distributions of the middle features in CNN is considered to be effective in realizing an accurate adaptation. To this end, numerous methods have
been proposed [8, 37, 4, 29, 40, 36].
可以通过各种策略来实现针对DA的CNN培训。 Ghifary等人提出对目标域使用自动编码器以获得域不变特征[9]。 Sener等人提出使用聚类技术和伪标记来获得区分特征[33]。 提出了跨域图像翻译方法[38]。 CNN中的中间特征的匹配分布被认为对实现准确的匹配有效。 为此，已经提出了许多方法[8、37、4、29、40、36]。

The representative method of distribution matching involves training a domain classifier using the middle features and generating the features that deceive the domain classifier [8]. This method utilizes the techniques used in generative adversarial networks [10]. The domain classifier is trained to predict the domain of each input, and the category classifier is trained to predict the task-specific category labels. Feature extraction layers are shared by the two classifiers. The layers are trained to correctly predict the label of source samples as well as to deceive the domain classifier.Thus, the distributions of the middle features of the target and source samples are made similar. Some methods utilize maximum mean discrepancy (MMD) [22, 21], which can be applied to measure the divergence in high-dimensional space between different domains. This approach can train the CNN to simultaneously minimize both the divergence and category loss for the source domain. These methods are based on the theory proposed by [2], which states that the error on the target domain is bounded by the divergence of the distributions. To our understanding, these distribution aligning methods using GAN or MMD do not consider the relationship between target samples and decision boundaries. To tackle these problems, we propose a novel approach using task-specific classifiers as a discriminator.
分布匹配的代表性方法包括使用中间特征训练域分类器，并生成欺骗域分类器的特征[8]。这种方法利用了在生成对抗网络中使用的技术[10]。训练领域分类器以预测每个输入的域，并且训练类别分类器以预测特定于任务的类别标签。特征提取层由两个分类器共享。训练各层可以正确预测源样本的标签以及欺骗域分类器，从而使目标样本和源样本的中间特征的分布相似。一些方法利用最大平均差异（MMD）[22，21]，可以将其应用于测量不同域之间的高维空间中的差异。这种方法可以训练CNN，以同时最小化源域的发散和类别损失。这些方法基于[2]提出的理论，该理论指出目标域上的误差受分布散度的限制。据我们了解，这些使用GAN或MMD进行分布调整的方法没有考虑目标样本与决策边界之间的关系。为了解决这些问题，我们提出了一种使用特定于任务的分类器作为区分器的新颖方法。

Consensus regularization is a technique used in multisource domain adaptation and multi-view learning, in which multiple classifiers are trained to maximize the consensus of their outputs [23]. In our method, we address a training step that minimizes the consensus of two classifiers, which is totally different from consensus regularization. Consensus regularization utilizes samples of multi-source domains to construct different classifiers as in [23]. In order to construct different classifiers, it relies on the different characteristics of samples in different source domains. By contrast,our method can construct different classifiers from only one source domain.
共识正则化是一种用于多源域自适应和多视图学习的技术，其中训练了多个分类器以最大化其输出的共识[23]。 在我们的方法中，我们解决了使两个分类器的共识最小化的训练步骤，这与共识正则化完全不同。 共识正则化利用多源域的样本来构建不同的分类器，如[23]。 为了构造不同的分类器，它依赖于不同源域中样本的不同特征。 相比之下，我们的方法只能从一个源域构造不同的分类器。

Method

In this section, we present the detail of our proposed method. First, we give the overall idea of our method in Section 3.1. Second, we explain about the loss function we used in experiments in Section 3.2. Finally, we explain the entire training procedure of our method in Section 3.3.
在本节中，我们介绍了我们提出的方法的细节。 首先，我们在3.1节中给出了我们方法的总体思路。 其次，我们将在3.2节中说明在实验中使用的损失函数。 最后，我们在3.3节中说明了我们方法的整个训练过程。

Overall Idea总体思路
We have access to a labeled source image xs and a corresponding label ys drawn from a set of labeled source images {Xs, Ys}, as well as an unlabeled target image xt drawn from unlabeled target images Xt. We train a feature generator network G, which takes inputs xs or xt, and classifier networks F1 and F2, which take features from G. F1 and F2 classify them into K classes, that is, they output a Kdimensional vector of logits. We obtain class probabilities by applying the softmax function for the vector. We use the notation p1(y|x), p2(y|x) to denote the K-dimensional probabilistic outputs for input x obtained by F1 and F2 respectively.
我们可以访问从一组标记的源图像{Xs，Ys}绘制的标记的源图像xs和相应的标签ys，以及从未标记的目标图像Xt绘制的未标记的目标图像xt。 我们训练一个特征生成器网络G，该网络接收输入xs或xt，以及分类器网络F1和F2，这些网络从G获取特征.F1和F2将它们分类为K类，即它们输出logits的K维向量。 我们通过对向量应用softmax函数来获得类概率。 我们使用符号p1（y | x），p2（y | x）表示分别由F1和F2获得的输入x的K维概率输出。
The goal of our method is to align source and target features by utilizing the task-specific classifiers as a discriminator in order to consider the relationship between class boundaries and target samples. For this objective, we have to detect target samples far from the support of the source.The question is how to detect target samples far from the support. These target samples are likely to be misclassified by the classifier learned from source samples because they are near the class boundaries. Then, in order to detect these target samples, we propose to utilize the disagreement of the two classifiers on the prediction for target samples. Consider two classifiers (F1 and F2) that have different characteristics in the leftmost side of Fig. 2. We assume that the two classifiers can classify source samples correctly. This assumption is realistic because we have access to labeled source samples in the setting of UDA. In addition, please note that F1 and F2 are initialized differently to obtain different classifiers from the beginning of training.Here, we have the key intuition that target samples outside the support of the source are likely to be classified differently by the two distinct classifiers. This region is denoted by black lines in the leftmost side of Fig. 2 (Discrepancy Region). Conversely, if we can measure the disagreement between the two classifiers and train the generator to minimize the disagreement, the generator will avoid generating target features outside the support of the source. Here, we consider measuring the difference for a target sample using the following equation, d(p1(y|xt); p2(y|xt)) where d denotes the function measuring divergence between two probabilistic outputs. This term indicates how the two classifiers disagree on their predictions and, hereafter, we call the term as discrepancy. Our goal is to obtain a feature generator that can minimize the discrepancy on target samples.
我们的方法的目标是通过使用特定于任务的分类器作为区分器来对齐源特征和目标特征，以便考虑类边界和目标样本之间的关系。为了这个目标，我们必须在远离源的支持下检测目标样本。问题是如何在远离支持的情况下检测目标样本。这些目标样本很可能会被从源样本中学习到的分类器误分类，因为它们接近类别边界。然后，为了检测这些目标样本，我们建议利用两个分类器在目标样本预测上的分歧。考虑图2最左侧具有不同特性的两个分类器（F1和F2）。我们假设这两个分类器可以正确分类源样本。这种假设是现实的，因为我们可以在UDA设置中访问带标签的源样本。另外，请注意，从训练开始就对F1和F2进行了不同的初始化，以获得不同的分类器。在此，我们有一个直觉，即两个不同的分类器可能会在源支持范围之外的目标样本进行不同的分类。该区域在图2的最左侧用黑线表示（差异区域）。相反，如果我们可以测量两个分类器之间的差异并训练生成器以最大程度地减少差异，则生成器将避免在源的支持范围之外生成目标特征。在这里，我们考虑使用以下等式d（p1（y | xt）; p2（y | xt））测量目标样本的差异，其中d表示测量两个概率输出之间的差异的函数。该术语表示两个分类器在其预测上存在分歧的方式，以下我们称该术语为差异。我们的目标是获得一个特征生成器，以最小化目标样本上的差异。
图2.（用彩色最佳视图。）两个分类器的示例，概述了所提出的方法。 差异是指两个分类器的预测之间的分歧。 首先，我们可以看到，源支持之外的目标样本可以通过两个不同的分类器（最左侧，两个不同的分类器）进行测量。 其次，关于训练过程，我们解决了一个极小极大问题，在其中找到两个分类器，使目标样本的差异最大化，然后生成使该差异最小的特征。

In order to effectively detect target samples outside the support of the source, we propose to train discriminators (F1 and F2) to maximize the discrepancy given target features (Maximize Discrepancy in Fig. 2). Without this operation, the two classifiers can be very similar ones and cannot detect target samples outside the support of the source. We then train the generator to fool the discriminator, that is,by minimizing the discrepancy (Minimize Discrepancy in Fig. 2). This operation encourages the target samples to be generated inside the support of the source. This adversarial learning steps are repeated in our method. Our goal is to obtain the features, in which the support of the target is included by that of the source (Obtained Distributions in Fig.2). We show the loss function used for discrepancy loss in the next section. Then, we detail the training procedure.
为了在源头支持之外有效地检测目标样本，我们建议训练鉴别器（F1和F2）以在给定目标特征的情况下最大程度地增加差异（图2中的“最大化差异”）。 如果不执行此操作，则这两个分类器可能非常相似，并且无法在源支持范围之外检测目标样本。 然后，我们通过最小化差异（图2中的“最小化差异”）训练生成器来欺骗鉴别器。 此操作鼓励在源的支持内生成目标样本。 这种对抗性学习步骤在我们的方法中重复进行。 我们的目标是获得特征，其中目标的支持包括源的支持（图2中的获得分布）。 我们将在下一部分中显示用于差异损失的损失函数。 然后，我们详细介绍了培训过程。

Discrepancy Loss差异损失
In this study, we utilize the absolute values of the difference between the two classifiers’ probabilistic outputs as discrepancy loss:
在这项研究中，我们将两个分类器的概率输出之差的绝对值用作差异损失：
where the p1k and p2k denote probability output of p1 and p2 for class k respectively. The choice for L1-distance is based on the Theorem . Additionally, we experimentally found that L2-distance does not work well.
其中p1k和p2k分别表示类别k的p1和p2的概率输出。 L1距离的选择基于定理。 此外，我们通过实验发现L2距离效果不佳。
Training Steps
To sum up the previous discussion in Section 3.1, we need to train two classifiers, which take inputs from the generator and maximize d(p1(y|xt); p2(y|xt)), and the generator which tries to mimic the classifiers. Both the classifiers and generator must classify source samples correctly. We
will show the manner in which to achieve this. We solve this problem in three steps.
为了总结第3.1节中的先前讨论，我们需要训练两个分类器，它们从生成器获取输入并最大化d（p1（y | xt）; p2（y | xt）），而生成器则试图模拟 分类器。 分类器和生成器都必须正确分类源样本。 我们将展示实现此目标的方式。 我们分三步解决这个问题。
Step A First, we train both classifiers and generator to classify the source samples correctly. In order to make classifiers and generator obtain task-specific discriminative features, this step is crucial. We train the networks to minimize softmax cross entropy. The objective is as follows:
步骤A首先，我们训练分类器和生成器以正确分类源样本。 为了使分类器和生成器获得特定于任务的区分特征，此步骤至关重要。 我们训练网络以最小化softmax交叉熵。 目的如下：
Step B In this step, we train the classifiers (F1, F2) as a discriminator for a fixed generator (G). By training the classifiers to increase the discrepancy, they can detect the target samples excluded by the support of the source. This step corresponds to Step B in Fig. 3. We add a classification loss
on the source samples. Without this loss, we experimentally found that our algorithm’s performance drops significantly.We use the same number of source and target samples to update the model. The objective is as follows:
步骤B在此步骤中，我们将分类器（F1，F2）训练为固定生成器（G）的鉴别器。 通过训练分类器以增加差异，他们可以检测源支持下排除的目标样本。 该步骤对应于图3中的步骤B。我们在源样本上添加了分类损失。 在没有这种损失的情况下，我们通过实验发现算法的性能大大下降。我们使用相同数量的源样本和目标样本来更新模型。 目的如下：
Step C We train the generator to minimize the discrepancy for fixed classifiers. This step corresponds to Step C in Fig. 3. The number n indicates the number of times we repeat this for the same mini-batch. This number is a hyperparameter of our method. This term denotes the trade-off between the generator and the classifiers. The objective is as follows:
步骤C我们训练生成器以最小化固定分类器的差异。 该步骤对应于图3中的步骤C。数字n表示我们对相同的小批量重复此操作的次数。 此数字是我们方法的超参数。 该术语表示生成器和分类器之间的权衡。 目的如下：

These three steps are repeated in our method. To our understanding, the order of the three steps is not important.Instead, our major concern is to train the classifiers and generator in an adversarial manner under the condition that they can classify source samples correctly.
这三个步骤在我们的方法中重复进行。 据我们了解，这三个步骤的顺序并不重要，相反，我们主要关注的是在能够正确分类源样本的条件下以对抗的方式训练分类器和生成器。
Theoretical Insight
Since our method is motivated by the theory proposed by Ben-David et al. [1], we want to show the relationshi between our method and the theory in this section.
由于我们的方法是由Ben-David等人[1]提出的理论所驱动的，因此我们希望在本节中展示我们的方法与理论之间的关系。
Ben-David等[1]提出了一种理论，该理论通过使用三个术语来限制目标样本的预期误差RT（h）：（i）源域的预期误差RS（h）； （ii）H∆H距离（dH∆H（S，T）），其度量为两个分类器之间的差异； （iii）理想联合假设λ的共同误差。 S和T分别表示源域和目标域。 另一种理论[2]将误差限制在目标域上，该误差引入了H距离（dH（S，T））用于域发散。 这两种理论及其关系可以解释如下。
在此，RT（h）是目标域上假设h的误差，RS（h）是源域上假设的误差。 I [a]是指标函数，如果谓词a为true，则为1，否则为0
H距离显示为凭经验通过域分类器的误差进行度量，该分类器经过训练可以区分特征域。 λ是一个常数（理想联合假设的共享误差），被认为足够低，无法实现精确的自适应。 早期的研究[8，37，4，29，40]试图测量Hdistance并将其最小化以实现这种适应。 正如这个不等式所暗示的那样，H距离在H∆H距离的上限。 我们将说明我们的方法与H∆Hdistance之间的关系。

This equation is very similar to the mini-max problem we solve in our method, in which classifiers are trained to maximize their discrepancy on target samples and generator tries to minimize it. Although we must train all networks to minimize the classification loss on source samples, we can see the connection to the theory proposed by [1].
该方程式与我们在方法中解决的极大极小问题非常相似，在该方法中，训练分类器以最大程度地减少目标样本上的差异，而生成器则尝试将其最小化。 尽管我们必须训练所有网络以最小化源样本的分类损失，但我们可以看到与[1]提出的理论的联系。

Conclusion

In this paper, we proposed a new approach for UDA,which utilizes task-specific classifiers to align distributions.We propose to utilize task-specific classifiers as discriminators that try to detect target samples that are far from the support of the source. A feature generator learns to generate target features near the support to fool the classifiers. Since the generator uses feedback from task-specific classifiers, it will avoid generating target features near class boundaries. We extensively evaluated our method on image classification and semantic segmentation datasets. In almost all experiments, our method outperformed state-of-the-art methods. We provide the results when applying gradient reversal layer [7] in the supplementary material, which enables to update parameters of the model in one step.
在本文中，我们提出了一种针对UDA的新方法，该方法利用特定于任务的分类器来对齐分布。我们建议利用特定于任务的分类器作为区分器，以尝试检测远离来源支持的目标样本。 特征生成器学习如何在支持附近生成目标特征，以欺骗分类器。 由于生成器使用来自特定任务分类器的反馈，因此它将避免在类边界附近生成目标特征。 我们在图像分类和语义分割数据集上广泛评估了我们的方法。 在几乎所有实验中，我们的方法都优于最新方法。 当在补充材料中应用梯度反转层[7]时，我们提供了结果，这使一步更新模型的参数成为可能。