Multi-Source Open-Set Deep Adversarial Domain Adaptation

Abstract

We introduce a novel learning paradigm based on multi-source open-set unsupervised domain adaptation (MS-OSDA). Recently, the notion of single-source open-set domain adaptation (OSDA) has drawn much attention which considers the presence of previously unseen open-set (unknown) classes in the target-domain in addition to the source-domain closed-set (known) classes. It is reasonable to assume that labeled samples may be distributed over multiple source-domains while the target-domain is equipped with both the closed-set and open-set data. However, the existing single-source OSDA techniques cannot be directly extended to such a multi-source scenario considering the inhomogeneities present among the different source-domains. As a remedy, we propose a novel adversarial learning-driven approach to deal with the MS-OSDA setup. Precisely, we model a shared feature space for all the domains while encouraging fine-grained alignment among the known-class samples. Besides, an adversarial learning strategy is followed to model the discriminator between the target-domain known and unknown classes. We validate our method on the Office-31, Office-Home, Office-CalTech, and Digits datasets and find our model to consistently outperform the relevant literature.