Invariance Principle Meets Vicinal Risk Minimization

TL;DR

This paper introduces a novel Semantic Data Augmentation module based on Variance Risk Minimization, improving out-of-distribution generalization in deep learning models for computer vision by enhancing dataset diversity and providing theoretical guarantees.

Contribution

It proposes a domain-shared SDA module utilizing VRM, along with a Rademacher complexity analysis, to improve OOD generalization and outperform existing methods.

Findings

Consistent performance improvements on PACS, VLCS, OfficeHome, and TerraIncognita.

Tighter generalization error bounds via Rademacher complexity analysis.

Enhanced dataset diversity without label instability.

Abstract

Deep learning models excel in computer vision tasks but often fail to generalize to out-of-distribution (OOD) domains. Invariant Risk Minimization (IRM) aims to address OOD generalization by learning domain-invariant features. However, IRM struggles with datasets exhibiting significant diversity shifts. While data augmentation methods like Mixup and Semantic Data Augmentation (SDA) enhance diversity, they risk over-augmentation and label instability. To address these challenges, we propose a domain-shared Semantic Data Augmentation (SDA) module, a novel implementation of Variance Risk Minimization (VRM) designed to enhance dataset diversity while maintaining label consistency. We further provide a Rademacher complexity analysis, establishing a tighter generalization error bound compared to baseline methods. Extensive evaluations on OOD benchmarks, including PACS, VLCS, OfficeHome, and TerraIncognita, demonstrate consistent performance improvements over state-of-the-art domain generalization methods.