Delving Deep into the Generalization of Vision Transformers under Distribution Shifts

Recently, Vision Transformers (ViTs) have achieved impressive results on various vision tasks. Yet, their generalization ability under different distribution shifts is rarely understood. In this work, we provide a comprehensive study on the out-of-distribution generalization of ViTs. To support a systematic investigation, we first present a taxonomy of distribution shifts by categorizing them into five conceptual groups: corruption shift, background shift, texture shift, destruction shift, and style shift. Then we perform extensive evaluations of ViT variants under different groups of distribution shifts and compare their generalization ability with CNNs. Several important observations are obtained: 1) ViTs generalize better than CNNs under multiple distribution shifts. With the same or fewer parameters, ViTs are ahead of corresponding CNNs by more than 5% in top-1 accuracy under most distribution shifts. 2) Larger ViTs gradually narrow the in-distribution and out-of-distribution performance gap. To further improve the generalization of ViTs, we design the Generalization-Enhanced ViTs by integrating adversarial learning, information theory, and self-supervised learning. By investigating three types of generalization-enhanced ViTs, we observe their gradient-sensitivity and design a smoother learning strategy to achieve a stable training process. With modified training schemes, we achieve improvements on performance towards out-of-distribution data by 4% from vanilla ViTs. We comprehensively compare three generalization-enhanced ViTs with their corresponding CNNs, and observe that: 1) For the enhanced model, larger ViTs still benefit more for the out-of-distribution generalization. 2) generalization-enhanced ViTs are more sensitive to the hyper-parameters than corresponding CNNs. We hope our comprehensive study could shed light on the design of more generalizable learning architectures.