Re-weighted Adversarial Adaptation Network for Unsupervised Domain Adaptation

The development of deep networks has enabled state-of-the-art performance to be achieved in many application domains at the cost of a large amount of domain-specific annotations. Unsupervised Domain Adaptation (UDA), however, aims to transfer domain knowledge from existing well-defined tasks to new ones where labels are unavailable. Real-world domain discrepancies from different data modalities are caused by complex and uncontrollable factors, which motivates the desire to match the feature distributions when the discrepancies are disparate. In this chapter, we propose the Re-weighted Adversarial Adaptation Network (RAAN) that is composed of two procedures: one to reduce the feature distribution divergence and the other to adapt the classifier when domain discrepancies are disparate. Specifically, to alleviate the need for common supports in matching the feature distribution, we choose to minimize optimal transport (OT) based Earth-Mover (EM) distance and reformulate it as an adversarial training scheme. Utilizing this, RAAN can be trained in an end-to-end and adversarial manner. To adapt the classifier, we propose to match the label distribution by learning the cross-domain label distribution ratios and embed it into the adversarial training. Finally, RAAN was shown to outperform other various methods by a large margin when the domain shifts and discrepancies are disparate.