HYPO: Hyperspherical Out-of-Distribution Generalization

• DOI: 10.48550/arxiv.2402.07785
• Publication Date: 2024-02-12
• Authors: Haoyue Bai; Yifei Ming; Julian Katz-Samuels; Yixuan Li
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Machine Learning (cs.LG)

Out-of-distribution (OOD) generalization is critical for machine learning models deployed in the real world. However, achieving this can be fundamentally challenging, as it requires ability to learn invariant features across different domains or environments. In paper, we propose a novel framework HYPO (HYPerspherical OOD generalization) that provably learns domain-invariant representations hyperspherical space. particular, our algorithm guided by intra-class variation and inter-class separation principles -- ensuring from same class (across training domains) are closely aligned with their prototypes, while prototypes maximally separated. We further provide theoretical justifications on how prototypical objective improves bound. Through extensive experiments challenging benchmarks, demonstrate approach outperforms competitive baselines achieves superior performance. Code available at https://github.com/deeplearning-wisc/hypo.