D2AP: Double Debiasing with Adaptive Proxies for Domain Generalization in Noisy Environment

Abstract

Domain Generalization (DG) methods commonly rely on multiple source domains with correct annotations, while it is usually difficult to obtain a large amount of clean source samples in practice, which greatly limits their application in real noisy environments. Hence, we consider a more realistic setting of Noisy Domain Generalization (NDG), which learns with noisy data from multiple source domains. A simple solution is to introduce previous noise learning strategies into individual domains to detect and correct the noisy samples independently. However, knowledge from other domains can also help the noise detection and correction for the current domain. Moreover, there will be domain bias of domain imbalance in clean samples after noise detection, and confirmation bias of inaccurate pseudo-labels in noisy samples after noise correction, which greatly affect the learning performance. To this end, we propose a novel Double Debiasing method with Adaptive Proxies (D2AP) for NDG learning. In D2AP, an adaptive sample disentangling module is first developed with multi-scale Gaussian Mixture Model, so as to enable a mutual noise disentangling across domains. Further, double debiasing is proposed with the assistance of adaptive proxies, in order to make domain-invariant prototypes less sensitive to domain imbalance, and calibrate the pseudo-labels for noisy data, so as to address the both biases. Finally, empirical results over three benchmark datasets demonstrate the effectiveness of our D2AP.