Cycle label-consistent networks for unsupervised domain adaptation

Domain adaptation aims to leverage a labeled source domain learn classifier for the unlabeled target with different distribution. Previous methods mostly match distribution between two domains by global or class alignment. However, alignment cannot achieve fine-grained class-to-class overlap; supervised pseudo-labels guarantee their reliability. In this paper, we propose simple yet efficient method, i.e. Cycle Label-Consistent Network (CLCN), exploiting cycle consistency of classification label, which applies dual cross-domain nearest centroid procedures generate reliable self-supervised signal discrimination in domain. The label-consistent loss reinforces ground-truth labels and samples leading statistically similar latent representations domains. This new can easily be added any existing network almost no computational overhead. We demonstrate effectiveness our approach on MNIST-USPS-SVHN, Office-31, Office-Home Image CLEF-DA benchmarks. Results validate that proposed method alleviate negative influence falsely-labeled more discriminative features, absolute improvement over source-only model 9.4% Office-31 6.3% CLEF-DA.