Abstract: Domain adversarial approaches have been at the core of many recent unsupervised domain adaptation algorithms. However, each new algorithm is presented independently with limited or no connections mentioned across the works. Instead, in this work we propose a unified view of adversarial adaptation methods. We show how to describe a variety of state-of-the-art adaptation methods within our framework and furthermore use our generalized view in order to better understand the similarities and differences between these recent approaches. In turn, this framework facilitates the development of new adaptation methods through modeling choices that combine the desirable properties of multiple existing methods. In this way, we propose a novel adversarial adaptation method that is effective yet considerably simpler than other competing methods. We demonstrate the promise of our approach by achieving state-of-the-art unsupervised adaptation results on the standard Office dataset.
Conflicts: bu.edu, berkeley.edu, stanford.edu