DGSAM: Domain Generalization via Individual Sharpness-Aware Minimization
This work addresses domain generalization for machine learning models that need to perform robustly on unseen data distributions, offering an incremental improvement over existing sharpness-aware methods.
The paper tackles the problem of domain generalization (DG) by identifying that Sharpness-Aware Minimization (SAM) can converge to fake flat minima, which appear globally flat but remain sharp for individual source domains, limiting generalization to unseen domains. They propose DGSAM, a method that minimizes individual sharpness efficiently, improving average accuracy and reducing performance variance across domains with less computational overhead than SAM.
Domain generalization (DG) aims to learn models that perform well on unseen target domains by training on multiple source domains. Sharpness-Aware Minimization (SAM), known for finding flat minima that improve generalization, has therefore been widely adopted in DG. However, our analysis reveals that SAM in DG may converge to \textit{fake flat minima}, where the total loss surface appears flat in terms of global sharpness but remains sharp with respect to individual source domains. To understand this phenomenon more precisely, we formalize the average worst-case domain risk as the maximum loss under domain distribution shifts within a bounded divergence, and derive a generalization bound that reveals the limitations of global sharpness-aware minimization. In contrast, we show that individual sharpness provides a valid upper bound on this risk, making it a more suitable proxy for robust domain generalization. Motivated by these insights, we shift the DG paradigm toward minimizing individual sharpness across source domains. We propose \textit{Decreased-overhead Gradual SAM (DGSAM)}, which applies gradual domain-wise perturbations in a computationally efficient manner to consistently reduce individual sharpness. Extensive experiments demonstrate that DGSAM not only improves average accuracy but also reduces performance variance across domains, while incurring less computational overhead than SAM.