A Bayesian Dual-Pathway Network for Unsupervised Domain Adaptation

Abstract

Unsupervised Domain Adaptation (UDA) endeavors to address the challenges presented by domain shifts between domains characterized by differing yet related distributions. Traditional adversarial approaches typically adopt a single-pathway adversarial paradigm, which relies on a singular pathway to align the marginal distributions at the domain level. Despite notable advancements, this paradigm is constrained by two major limitations that lead to sub-optimal performance in both source and target domains. First, naive domain-level alignment often results in class mismatches. Second, the single-pathway adversarial approach grapples with the conflicting demands of reducing domain shift while simultaneously learning comprehensive features. Drawing inspiration from cognitive neuroscience, we propose a Bayesian Dual-Pathway Network (BDNet) for UDA to compute a classification prior for each domain, comprising a domain-shared pathway and a domain-specific pathway, designed to enhance target domain performance while preserving source domain efficacy. Specifically, the domain-shared pathway is employed to learn classification prior features through an adversarial paradigm grounded in structural alignment. Concurrently, a domain-specific pathway is crafted to extract distinct features, incorporating domain likelihood and domain prior features. Comprehensive features are synthesized through the fusion of common and specific attributes via a lightweight fusion module. Extensive experiments across three publicly available datasets demonstrate the efficacy of our approach, evidencing superior performance in both source and target domains.

Keywords: Domain adaption, Bayesian theory, source domain preserving, domain-specific classification posterior