A Novel Multi-source Domain Learning Approach to Unsupervised Deep Domain Adaptation

Abstract

Even though it is anticipated that training and test data come from same distribution, but in many practical applications, they usually have different distributions, resulting in poor classification performance. To overcome this distribution difference, various Transfer Learning (TL) approaches such as Domain Adaptation (DA) have been proposed. Many existing DA approaches are capable of minimizing the distribution difference of single source and target domains, resulting in the transfer of useful information from only one source domain to the target domain. However, the useful information can also be transferred from multiple source domains to the target domain for further improving the performance. Therefore, in this paper, we proposed a novel Convolutional Neural Network (CNN) architecture that considers three (or more depending on a number of source domains) stream CNN networks and introduces domain confusion and discriminative losses, to learn a representation that is both semantically meaningful and domain invariant. Extensive experiments on different tasks of domain adaptation dataset such as PIE show that considering multiple sources can significantly boost the performance of deep domain adaptation model. As a result, our approach outperforms existing state-of-the-art methods in cross-domain recognition.