Reproduce “NODI: Out-Of-Distribution Detection with Noise from Diffusion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jingqiu Zhou, Aojun Zhou, Hongsheng Li
NODI introduces a novel out-of-distribution detection method that leverages diffusion models to estimate noise from in-distribution data, integrating global dataset information for robust OOD scoring. The approach achieves significant AUROC gains, up to 3.7% on near-OOD detection, and demonstrates consistent superior performance across diverse image encoder architectures compared to existing methods.
Key insights
- Diffusion models can effectively capture the comprehensive global manifold of in-distribution data, addressing the limitation of prior methods using only local information.
- Normalizing in-distribution features onto a hypersphere and deriving a closed-form "stable point" for the noise vector provides a robust framework for OOD scoring.
- The performance of OOD detection methods is significantly influenced by the underlying image encoder, and true robustness requires consistent performance across diverse architectures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.