Unsupervised Variational Domain Adaptation

Abstract

Unsupervised domain adaptation (UDA) aims at boosting learning tasks of the target domain (TD) via transferring learned knowledge from the source domain (SD). Feature alignment, as a key point of UDA, is often pursued by adversarial training or minimizing discrepancy of the marginal distributions of the two domains. However, global feature alignment is not sufficient to eliminate the gap between the domains. Most existing approaches often ignore category-level features during the feature alignment, which might lead to mode collapse. To deal with this issue, we propose a cross-domain probabilistic generative model (CPGM), which formulates the category-level feature adaptation as an issue of probabilistic approximation (i.e., the posterior probability of the TD is forced to approximate the prior probability of the SD). We further present theoretical analysis of evidence lower bound (ELBO) based on variational inference to solve the issue of probabilistic approximation. Consequently, we build an unsupervised variational domain adaptation (UVDA) method for classification tasks, which mitigates the mode collapse issue of the traditional global feature alignment method by constructing an ELBO loss, based on the CPGM. Our UVDA adopts an alternative training strategy that adapts category-level and global features via CPGM and adversarial training, respectively. In particular, we propose an effective sample screening module (SSM) to progressively select target samples with high confidence to facilitate the calculation of ELBO for maximizing the capability of CPGM. Experimental results on four popular datasets, namely, Digits, Office31, VisDA-2017 and DomainNet, demonstrate that our UVDA is effective, and outperforms the state-of-the-art methods.