Moment matching of joint distributions for unsupervised domain adaptation

Bo Zhang, Xiaoming Zhang, Zhibo Zhou, Yun Liu, Yancong Li, Feiran Huang

Abstract: Highlights•Introduced three novel UDA criteria for aligning cross-domain joint distributions.•Proposed a source-free solution leveraging gradient features for improved adaptation.•Enhanced the efficacy of conventional KD under non-aligned training conditions.•Experiments demonstrate the superiority in multiple UDA settings.