Towards self-similarity consistency and feature discrimination for unsupervised domain adaptation

Recent advances in unsupervised domain adaptation mainly focus on learning shared representations by global statistics alignment, such as the Maximum Mean Discrepancy (MMD) which matches across domains. The lack of class information, however, may lead to partial alignment (or even misalignment) and poor generalization performance. For robust we argue that similarities different features source should be consistent with target domain. Based this assumption, propose a new discrepancy metric, i.e., Self-similarity Consistency (SSC), enforce pairwise relationship between being Gram matrix matching Correlation Alignment is proven special case, sub-optimal measure our proposed SSC. Furthermore, also mitigate side effect misalignment incorporating discriminative information deep representations. Specifically, simple yet effective feature norm constraint exploited enlarge inter-class samples. It relieves requirements strict when performing adaptation, therefore improving performance significantly. Extensive experiments visual tasks demonstrate effectiveness SSC metric discrimination approach.