Certain and Consistent Domain Adaptation

Bhadrinath Nagabandi, Andrew Dudley, Hemanth Venkateswara, Sethuraman Panchanathan

Abstract: Unsupervised domain adaptation algorithms seek to transfer knowledge from labeled source datasets in order to predict the labels for target datasets in the presence of domain-shift. In this paper we propose the Certain and Consistent Domain Adaptation (CCDA) model for unsupervised domain adaptation. The CCDA aligns the source and target domains using adversarial training and reduces the domain adaptation problem to a semi supervised learning (SSL) problem. We estimate the target labels using consistency regularization and entropy minimization on the domain-aligned target samples whose predictions are consistent across multiple stochastic perturbations. We evaluate the CCDA on benchmark datasets and demonstrate that it outperforms competitive baselines from domain adaptation literature.