对抗式域适配迁移学习研究

Research on Adversarial Domain Adaptation

摘要: 对抗式域自适配以对抗学习的方式最小化源领域中的任务损失和最大化域混肴损失，学习到领域间的共享特征空间，从而进行特征适配(分布式配)，辅助目标领域学习任务，是当前域自适应研究，被广泛应用到在行人重识别、图像分类和情感分析等领域。本文梳理了当前对抗式域适配研究工作，按照源领域和目标领域标签类别空间异同将其分为同构对抗式域自适配和异构对抗式域自适配两种子类型。依次详细介绍边缘分布同构对抗式域自适配、条件分布对抗域自适配、联合分布对抗域自适配和动态分布对抗域自适配、开放集对抗式域自适配、局部对抗式域自适配和通用对抗式域适配六类子域自适配的研究问题、研究思路及主要研究工作。

Abstract: Adversarial domain adaptation can be used to learn a shared feature space across domains by minimizing the task loss in the source domain and maximizing the domain confusion loss. Then the task in the target domain can get assistance from feature adaptation (distributed adaptation). Now, adversarial domain adaptation is widely used in the fields of pedestrian re-recognition, image classification, sentiment analysis, and so on. This paper summarizes the current research work of adversarial domain adaptation and divides it into two subtypes: homogeneous adversarial domain adaptation, heterogeneous adversarial domain adaptation according to whether the label space of the source domain and the target domain is the same. This paper also introduces in detail the research problems, research ideas, and main research work of six kinds of subdomain adaptation: marginal distribution adversarial domain adaptation, conditional distribution adversarial domain adaptation, joint distribution adversarial domain adaptation, dynamic distributed adversarial domain adaptation, open set adversarial domain adaptation, partial adversarial domain adaptation, and universal adversarial domain adaptation.