Unsupervised domain adaptation via distilled discriminative clustering

Unsupervised domain adaptation addresses the problem of classifying data in an unlabeled target domain, given labeled source that share a common label space but follow different distribution. Most recent methods take approach explicitly aligning feature distributions between two domains. Differently, motivated by fundamental assumption for adaptability, we re-cast as discriminative clustering data, strong privileged information provided closely related, data. Technically, use objectives based on robust variant entropy minimization adaptively filters soft Fisher-like criterion, and additionally cluster ordering via centroid classification. To distill clustering, propose to jointly train network using parallel, supervised learning over We term our method distilled DisClusterDA. also give geometric intuition illustrates how constituent DisClusterDA help learn class-wisely pure, compact distributions. conduct careful ablation studies extensive experiments five popular benchmark datasets, including multi-source one. Based commonly used backbone networks, outperforms existing these benchmarks. It is interesting observe framework, adding additional loss learns align class-level across domains does harm performance, though more algorithmic frameworks are be conducted.