The research article presents a novel method for Out-of-distribution (OOD) detection in neural networks, inspired by the phenomenon of Neural Collapse (NC). Existing OOD detection methods have utilized the fact that In-distribution (ID) samples form a subspace in the feature space. The researchers propose a new OOD scoring function, Entropy-enhanced Principal Angle (EPA), that leverages insights from NC and integrates both the global characteristic of the ID subspace and its inner property. The article further compares the performance of EPA with other state-of-the-art approaches, concluding that EPA offers superior performance and robustness across different network architectures and OOD datasets.
Publication date: 4 Jan 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2401.01710
