Semi-supervised Adversarial Domain Adaptation Based on Maximum Mean Discrepancy

Abstract

Domain adaptation aims to transfer knowledge from a source domain to a new but related target domain. Most adversarial training methods align feature distributions thus both domains can share the same classifier. However, compared to unsupervised adversarial domain adaptation, supervised information from the labeled target domain can better guide the transfer process by learning transferable as well as discriminative features. In this paper, we propose a novel semi-supervised adversarial domain adaptation (SSADA) method that can align the feature distributions across domains. In SSADA, labeled target samples are used to learn discriminative features while unlabeled target samples are used to learn transferable features based on Maximum Mean Discrepancy. The experiments on public datasets demonstrate the effectiveness and efficiency of our proposed method.