Deep domain adaptation via joint transfer networks

Abstract

Domain adaptation aims to transfer the enrich label knowledge from a large-scale labeled tasks to new ones with no labeled data. In the real-world scenario, the domain discrepancy of feature distributions between different tasks (domains) is usually uncontrollable, which is dramatically motivated to match the feature distributions in the face of the domain discrepancy is completely divergence. Under the condition that target task (domain) annotations are unknown, how to successfully adapt the trained classified from source to the target of interest still remains an open issue. In this paper, a unified domain adaptation method is proposed, Joint Transfer Networks (JTN), which jointly generates domain transferable features across two different domains and a double weighting scheme is able to learn more meaningful and transferable features. Additionally, we also exploit the pseudo-labeled target data to generate discriminative features in the target domain, which further boosts the adaptation performance. Finally, after a thorough evaluation of proposed method utilizing several benchmark datasets of varying difficulty. JTN yielded the state-of-the-art performance and outperformed baseline approaches for various domain adaptation tasks.