Dual Adversity Training for Domain Generalization

Youjia Shao, Longfei Chen, Changshuo Wang, Wenna Liu, Wencang Zhao

Abstract: Domain generalization (DG) aims to learning a robust model across different source domains that can successfully generalize generic knowledge to unseen target domains. Inspired by human-oriented psychology, we explore the innovative direction that post-traumatic growth theory brings to DG. Accordingly, we propose dual adversity training for DG, designed to push the model beyond its comfort zone and enable it to fight against adversity stress, leading to self-transcendence and positive progress. Our method consists of the first image-level extrinsic adversity training and the second feature-level intrinsic adversity training, both of which align with human learning processes in society. The former constructs a global and dynamic loss bank to define the challenge score of images, allowing greater utilization of highly challenging images during model training. This encourage the model to explore its potential under adversity. The latter leverages the feature channel as a carrier for channel masking across multiple layers, guided by the gradient signal to noise ratio (GSNR). Features with high GSNR are removed, forcing the model to challenge itself by actively learning more domain-invariant features. These two are closely interconnected, working together to facilitate comprehensive domain-invariant representation learning with sufficient semantics, ultimately enhancing robust generalization performance. Extensive experimental results on several widely used datasets verify the feasibility and effectiveness of our proposed method.

External IDs:doi:10.1016/j.knosys.2025.114808