Domain adaptation based on incremental adversarial learning

Abstract

Domain adaptation is a method of transfer learning. Domain adaptation has a source domain and target domain with related but different distributions. Unsupervised domain adaptation could be a scenario wherever we've labeled unlabeled target data and source data. In this paper, an incremental adversarial learning method is proposed for unsupervised domain adaptation. In this work, the unknown target labels are predicted and according to these estimated labels, some target data with more similarity to the source data are added to the source data to improve the adaptation between two domains. We use the adversarial discriminative approach as the base unsupervised domain adaptation technique. We do this to handle the large domain shift between the source and target domain distributions. Experimental reports prove that our approach performs much better on several visual domain adaptation tasks.