Adversarial Transform Networks for Unsupervised Transfer Learning

Transfer learning, especially unsupervised domain adaptation, is a crucial technology for sample-efficient learning. Recently, deep adversarial domain adaptation methods perform remarkably well in various tasks, which introduce a domain classifier to promote domain-invariant representation. However, previous methods either constrain the representative ability with an identical feature extractor for both domains or ignore the relationship between domains with separate extractors. In this paper, we propose a novel adversarial domain adaptation method named Adversarial Transform Network (ATN) to both enhance the representative ability and transfer general information between domains. Residual connections are used to share features in the bottom layers, which deliver transferrable features to boost generalization performance. Moreover, a regularizer is proposed to alleviate a vanishing gradient problem, thus stabilizing the optimization procedure. Extensive experiments are conducted to show that the proposed ATN is comparable with the methods of the state-of-the-art and effectively deals with the vanishing gradient problem.