Reproduce “EigenScore: OOD Detection using Covariance in Diffusion Models” on OpenResearch

About this paper

Shirin Shoushtari, Yi Wang, Xiao Shi, M. Salman Asif, Ulugbek S. Kamilov

EigenScore introduces an unsupervised out-of-distribution (OOD) detection method that analyzes the eigenvalue spectrum of the posterior covariance in diffusion models during the denoising process. The approach provides a theoretically grounded and stable signal for identifying distribution shift, achieving state-of-the-art performance, especially in challenging near-OOD scenarios.

Key insights

• Distribution shift can be quantified as the accumulation of 'excess' denoising error incurred when an in-distribution (InD) trained denoiser processes out-of-distribution (OOD) data.
• This 'excess' denoising error corresponds to an inflation of the total posterior variance, which manifests as a distinct 'spectral signature'—specifically, an inflation of the eigenvalue spectrum of the posterior covariance—for OOD inputs.
• Focusing on the leading eigenvalues of the posterior covariance retains the most informative uncertainty directions, providing a more stable and discriminative OOD signal compared to scalar metrics, which often aggregate noise at high diffusion timesteps.

Summary by alphaXiv.