Reproduce “Revisiting Likelihood-Based Out-of-Distribution Detection by Modeling Representations” on OpenResearch

About this paper

Yifan Ding, Arturas Aleksandraus, Amirhossein Ahmadian, Jonas Unger, Fredrik Lindsten, Gabriel Eilertsen

Researchers from Linköping University developed a method to re-evaluate likelihood-based out-of-distribution (OOD) detection by applying score-based diffusion models to learned representation spaces from pre-trained encoders. This approach demonstrates that these methods can achieve state-of-the-art performance, with the class-conditional variant reaching an average AUROC of 96.35% and FPR95 of 17.06% on the ImageNet benchmark, challenging the previous perception of their ineffectiveness.

Key insights

• The historical failures of likelihood-based OOD detection were not due to an inherent flaw in the likelihood principle but rather from applying it in raw pixel space, where low-level features obscure semantic shifts.
• Likelihood functions effectively as an OOD score when applied to semantically clustered representation spaces generated by modern pre-trained encoders.
• More accurate likelihood estimation in the representation space, as provided by score-based diffusion models, directly correlates with improved OOD detection performance.

Summary by alphaXiv.