GOLD: Guiding Contrastive Learning with Out-of-Distribution Detection for Universal Domain Adaptation

Universal domain adaptation seeks to extend knowledge from a labeled source domain to an unlabeled target domain, unconstrained by label space alignment. It is challenging to align shared categories and separate private ones without prior category overlap information. Existing methods often rely heavily on source domain information to learn a transferable classifier, neglecting the relationships between the manifold structures inherent within the two domains. This paper introduces a novel framework, called Guiding cOntrastive Learning with out-of-distribution Detection (GOLD) for universal domain adaptation. GOLD employs an instance-prototype hybrid contrastive learning approach with self-attention to reveal domain structures. Subsequently, it constructs a residual subspace from source prototypes to filter unknown categories and refines neighborhood structures through instance-level virtual adversarial training to reduce noise. Experiments on three datasets show GOLD surpasses current methods in various UniDA settings.