Domain Adaptation Based on Domain-invariant and Class-distinguishable Feature Learning using Multiple Adversarial Networks

Abstract Adversarial networks have been used to learn transferable representations in many domain adaptation methods. However, there is no theoretical guarantee that two distributions are identical, even if the discriminator is fully confused. Therefore, a more elaborate domain adversarial method to better align distributions is desirable. In this paper, we propose two groups of multiple adversarial networks for domain-invariant and class-distinguishable feature learning: i) class-wise domain adversarial networks based on sample locations and ii) sample adversarial networks. We analyze the impact of the sample’s intra-class distribution on transfer learning and reveal that the distance between a sample and its cluster center affects the role of that sample in transfer learning. Domain adversarial learning is conducted separately on samples located near the cluster center (central samples) and samples located far away from cluster center (non-central samples) in order to align the distributions of the source domain and target domain better. However, separate domain adversarial learning on central and non-central samples ignores the relationship between them, because all these samples belong to the same class. We therefore propose a method called sample adversarial learning to convert the distribution of non-central samples to the distribution of central samples. In this way, the relationship between central samples and non-central samples is rebuilt. Sample adversarial learning solves the problem that arises in the sperate domain adversarial learning. Sample adversarial learning also enables us to obtain a class-distinguishable feature representation because of the reduction in intra-class distance. Experimental results show that our method extracts more transferable and class-distinguishable features than existing methods and achieves start-of-the-art results on several datasets. The significant contribution of this paper is to show how separate domain adversarial learning based on sample locations and sample adversarial learning together enhance positive transfer by maximally matching the multimodal structures underlying the data distributions across domains.