Source-Constraint Adversarial Domain Adaptation

Abstract

Adversarial adaptation has made great contributions to transfer learning, while the adversarial training strategy lacks of stability when reducing the discrepancy between domains. In this paper, we propose a novel Source-constraint Adversarial Domain Adaptation (SADA) method, which jointly use adversarial adaptation and maximum mean discrepancy (MMD) so that the method can be easily optimized by gradient descent. Furthermore, motivated by metric learning, our method introduces metric loss to constrain the structure of source domain, which explicitly increases the inter-class distance and decreases the intra-class distance. As a result, SADA can not only reduce the domain discrepancy, but also make the extracted features become more domain-invariant and discriminative. We show that our model yields state of the art results on standard datasets.