Metric-learning-assisted domain adaptation

Domain alignment (DA) has been widely used in unsupervised domain adaptation. Many existing DA methods assume that a low source risk, together with the of distributions and target, means target risk. In this paper, we show does not always hold. We thus propose novel metric-learning-assisted adaptation (MLA-DA) method, which employs triplet loss for helping better feature alignment. explore relationship between second largest probability sample’s prediction its distance to decision boundary. Based on relationship, mechanism adaptively adjust margin according predictions. Experimental results use proposed can achieve clearly results. also demonstrate performance improvement MLA-DA all four standard benchmarks compared state-of-the-art methods. Furthermore, shows stable robust experiments.