Cross-Domain Error Minimization for Unsupervised Domain Adaptation

Unsupervised domain adaptation aims to transfer knowledge from a labeled source an unlabeled target domain. Previous methods focus on learning domain-invariant features decrease the discrepancy between feature distributions as well minimizing error and have made remarkable progress. However, recently proposed theory reveals that such strategy is not sufficient for successful adaptation. It shows besides small error, both labeling functions should be across domains. The essentially cross-domain errors which are ignored by existing methods. To overcome this issue, in paper, novel method integrate all objectives into unified optimization framework. Moreover, incorrect pseudo labels widely used previous can lead accumulation during learning. alleviate problem, obtained utilizing structural information of classifier we propose curriculum based select samples with more accurate pseudo-labels training. Comprehensive experiments conducted, results validate our approach outperforms state-of-the-art