Bi-Classifier Determinacy Maximization for Unsupervised Domain Adaptation

Unsupervised domain adaptation challenges the problem of transferring knowledge from a well-labelled source to an unlabelled target domain. Recently, adversarial learning with bi-classifier has been proven effective in pushing cross-domain distributions close. Prior approaches typically leverage disagreement between learn transferable representations, however, they often neglect classifier determinacy domain, which could result lack feature discriminability. In this paper, we present simple yet method, namely Bi-Classifier Determinacy Maximization (BCDM), tackle problem. Motivated by observation that samples cannot always be separated distinctly decision boundary, here proposed BCDM, design novel disparity (CDD) metric, formulates discrepancy as class relevance distinct predictions and implicitly introduces constraint on To end, BCDM can generate discriminative representations encouraging predictive outputs consistent determined, meanwhile, preserve diversity manner. Furthermore, properties CDD well theoretical guarantees BCDM's generalization bound are both elaborated. Extensive experiments show compares favorably against existing state-of-the-art methods.