Understanding the Effectiveness of Cross-Domain Contrastive Unsupervised Domain Adaptation

Viacheslav Sinii, Adín Ramírez Rivera, Adil Khan

Keywords: Domain Adaptation, Contrastive Learning

Abstract: Unsupervised domain adaptation helps to transfer learned tasks from a source to a target domain in the lack of labeled data. Recently, contrastive learning showed promising results on this setup. However, there are limitations on the performance due to unbalanced objectives between the self-representation and the adaptation tasks. We show that pre-training choices and hard negative mining provide consistent improvements to successfully pair contrastive learning and unsupervised domain adaptation.