Enhancing transferability and discriminability simultaneously for unsupervised domain adaptation

摘要

Unsupervised domain adaptation aims to transfer knowledge from the labeled source domain to the unlabeled target domain. In recent studies, deep learning based unsupervised domain adaptation methods have achieved promising progress. However, most methods use a shared feature extractor in the source and target domains. Due to the large domain gap, the generated domain-invariant features may still contain some domain-specific information, which affects the transferability. In addition, the generated features change the geometric structure of the original feature space, causing the loss of discriminability. To this end, we propose an Enhancing Transferability and Discriminability Simultaneously (ETDS) method. Specifically, ETDS contains two domain-specific modules to explicitly capture domain-specific information, which contributes to preserving common characteristics in domain-invariant features, thereby enhancing transferability. Besides, ETDS enhances discriminability by forcing the features and their corresponding prototypes closer. In particular, we propose a new method for constructing prototypes without using ground-truth labels or pseudo-labels, and propose a balance strategy to control the relative contribution of source prototypes and target prototypes. Experiments are conducted on four widely used datasets, and the results show that our method outperforms recent domain adaptation methods, especially on DomainNet, the hardest domain adaptation dataset by far. And we employ spectral analysis to intuitively show the enhanced transferability and discriminability of generated features.