Partially-Shared Variational Auto-encoders for Unsupervised Domain Adaptation with Target Shift

Ryuhei Takahashi, Atsushi Hashimoto, Motoharu Sonogashira, Masaaki Iiyama

Abstract: Target shift, the different label distributions of source and target domains, is an important problem for practical use of unsupervised domain adaptation (UDA); as we do not know labels in target domain datasets, we cannot ensure an identical label distribution between the two domains. Despite this inaccessibility, modern UDA methods commonly try to match the shape of the feature distributions over the domains while projecting the features to labels by a common classifier. This implicitly assumes the identical label distribution. To overcome this problem, we propose a method that generates a pseudo pair by domain conversion where the label is preserved identically even trained with target-shifted datasets. A pair-wise metric learning enables to align feature over the domains without matching the shape of distributions. We conducted two experiments: one is a regression of pose-estimation, where label distribution is continuous and the target shift problem can seriously degrade the quality of UDA. The other is digit classification task where we can systematically control the distribution difference. The code and dataset are available at https://github.com/iiyama-lab/PS-VAEs .