Contrastively Smoothed Class Alignment for Unsupervised Domain Adaptation

Recent unsupervised approaches to domain adaptation primarily focus on minimizing the gap between source and target domains through refining feature generator, in order learn a better alignment two domains. This minimization can be achieved via classifier detect target-domain features that are divergent from source-domain features. However, when optimizing such domain-classification discrepancy, ambiguous samples not smoothly distributed low-dimensional data manifold often missed. To solve this issue, we propose novel Contrastively Smoothed Class Alignment (CoSCA) model, explicitly incorporates both intra- inter-class discrepancy align with domain. CoSCA estimates underlying label hypothesis of samples, simultaneously adapts their representations by proposed contrastive loss. In addition, Maximum Mean Discrepancy (MMD) is utilized directly match for global alignment. Experiments several benchmark datasets demonstrate CoSCAoutperforms state-of-the-art producing more discriminative