Discriminative Feature Alignment: Improving Transferability of Unsupervised Domain Adaptation by Gaussian-guided Latent Alignment

In this study, we focus on the unsupervised domain adaptation problem where\nan approximate inference model is to be learned from a labeled data domain and\nexpected to generalize well to an unlabeled data domain. The success of\nunsupervised domain adaptation largely relies on the cross-domain feature\nalignment. Previous work has attempted to directly align latent features by the\nclassifier-induced discrepancies. Nevertheless, a common feature space cannot\nalways be learned via this direct feature alignment especially when a large\ndomain gap exists. To solve this problem, we introduce a Gaussian-guided latent\nalignment approach to align the latent feature distributions of the two domains\nunder the guidance of the prior distribution. In such an indirect way, the\ndistributions over the samples from the two domains will be constructed on a\ncommon feature space, i.e., the space of the prior, which promotes better\nfeature alignment. To effectively align the target latent distribution with\nthis prior distribution, we also propose a novel unpaired L1-distance by taking\nadvantage of the formulation of the encoder-decoder. The extensive evaluations\non nine benchmark datasets validate the superior knowledge transferability\nthrough outperforming state-of-the-art methods and the versatility of the\nproposed method by improving the existing work significantly.\n