Adaptive prototype and consistency alignment for semi-supervised domain adaptation

Abstract

Unsupervised Domain Adaptation (UDA) aims to transfer knowledge from a label-rich source domain to an unlabeled target domain whose data distributions are different. There is a more realistic scenario where a few target labels are available, namely Semi-Supervised Domain Adaptation (SSDA). The existing methods reduce the inter-domain discrepancy by ignoring the class-level information, which may lead to cross-domain feature mismatch. Therefore, the model fails to learn discriminative feature representation for the target domain. In this paper, we propose a novel SSDA method, namely Adaptive Prototype and Consistency Alignment (APCA). To be specific, the Adaptive Prototype Alignment (APA) strategy employs a novel prototypical loss to realize the class-level alignment and further reduce the inter-domain discrepancy. Moreover, we apply Consistency Alignment (CA) to improve the robustness of the model and produce a robust cluster core which is beneficial to class-level alignment and thus facilitates the reduction of inter-domain discrepancy. We evaluate our approach on four domain adaptation datasets and the experimental results demonstrate the effectiveness of our proposed approach.