MapFlow: latent transition via normalizing flow for unsupervised domain adaptation

Abstract Unsupervised domain adaptation (UDA) aims at enhancing the generalizability of classification model learned from labeled source to an unlabeled target domain. An established approach UDA is constrain classifier on intermediate representation that distributionally invariant across domains. However, recent theoretical and empirical research has revealed relying only invariance fails guarantee a small error, thus making equality in distribution representations unnecessary. In this paper, we propose relax learning by finding general relationship between representations, which allows interchange more discriminative information. To end, formalize MapFlow framework, explicitly constructs invertible mapping encoded variationally induced representation. Empirical results public benchmark datasets show desirable performance our proposed algorithm compared state-of-the-art methods.