Bi-Classifier Adversarial Learning-Based Unsupervised Domain Adaptation

This chapter focuses on adversarial learning-based UDA techniques, especially the bi-classifier adversarial learning. Bi-classifier methods adopt two classifiers and feature alignment is achieved by minimizing their output discrepancy on target samples. However, previous bi-classifier methods have two limitations: (1) solely matching classifier outputs does not guarantee target accuracy, and (2) their evenly narrowing decision boundaries may degrade discriminability. To address these issues, this chapter proposes two techniques. First, a cross-domain gradient discrepancy minimization (CGDM) method is introduced to explicitly align the gradients of source and target samples, which provides supervision for improving target accuracy. Second, an uneven bi-classifier learning strategy is proposed, where one classifier is adversarially trained to generalize the feature extractor, while the other focuses on preserving target decision boundaries. Experiments on several domain adaptation benchmarks demonstrate the efficacy of both CGDM and uneven bi-classifier learning in boosting adaptation performance.