Domain-wise Adversarial Training for Out-of-Distribution Generalization

Shiji Xin, Yifei Wang, Jingtong Su, Yisen Wang

Keywords: Domain Generalization, IRM, Adversarial Training

Abstract: Despite the impressive success on many tasks, deep learning models are shown to rely on spurious features, which will catastrophically fail when generalized to out-of-distribution (OOD) data. To alleviate this issue, Invariant Risk Minimization (IRM) is proposed to extract domain-invariant features for OOD generalization. Nevertheless, recent work shows that IRM is only effective for a certain type of distribution shift (e.g., correlation shift) while fails for other cases (e.g., diversity shift). Meanwhile, another thread of method, Adversarial Training (AT), has shown better domain transfer performance, suggesting that it is potential to be an effective candidate for extracting domain-invariant features. In this paper, we investigate this possibility by exploring the similarity between the IRM and AT objectives. Inspired by this connection, we propose Domain-wise Adversarial Training (DAT), an AT-inspired method for alleviating distribution shift by domain-specific perturbations. Extensive experiments show that our proposed DAT can effectively remove the domain-varying features and improve OOD generalization on both correlation shift and diversity shift tasks.

One-sentence Summary: We explore the similarity between Invariant Risk Minimization (IRM) and Adversarial Training (AT), based on which we propose a Domain-wise AT (DAT) with superior performance on benchmark OOD datasets.