Domain Adaptation with Adversarial Training on Penultimate Activations

Enhancing model prediction confidence on target data is an important objective in Unsupervised Domain Adaptation (UDA). In this paper, we explore adversarial training penultimate activations, i.e., input features of the final linear classification layer. We show that strategy more efficient and better correlated with boosting than images or intermediate features, as used previous works. Furthermore, activation normalization commonly domain adaptation to reduce gap, derive two variants systematically analyze effects our training. This illustrated both theory through empirical analysis real tasks. Extensive experiments are conducted popular UDA benchmarks under standard setting source-data free setting. The results validate method achieves best scores against arts. Code available at https://github.com/tsun/APA.