Generative models for noise-robust training in unsupervised domain adaptation

Abstract

Recent unsupervised domain adaptation (UDA) methods show the effectiveness of pseudo-labels for unlabeled target domain. However, pseudo-labels inevitably contain noise, which can degrade adaptation performance. This paper thus propose a Generative models for Noise-Robust Training (GeNRT), a method designed to mitigate label noise while reducing domain shift. The key idea is leveraging the class-wise distributions of the target domain, modeled by generative models, provide more reliable pseudo-labels than individual pseudo-labeled instances. This is because the distributions statistically better represent class-wise information than a single instance. Based on this observation, GeNRT incorporates a Distribution-based Class-wise Feature Augmentation (D-CFA), which enhances feature representations by sampling features from target class distributions modeled by generative models. These augmented features serve a dual purpose: (1) providing class-level knowledge from generative models to train a noise-robust discriminative classifier, and (2) acting as intermediate features to bridge the domain gap at the class level. Furthermore, GeNRT leverages Generative and Discriminative Consistency (GDC), enforcing consistency regularization between a generative classifier (formed by all class-wise generative models) and the learned discriminative classifier. By aggregating knowledge across target class distributions, GeNRT improves pseudo-label reliability and enhances robustness against label noise. Extensive experiments on Office-Home, VisDA-2017, PACS, and Digit-Five show that our GeNRT achieves comparable performance to state-of-the-art methods under both single-source and multi-source UDA settings.