Reproduce “SCOPED: Score-Curvature Out-of-distribution Proximity Evaluator for Diffusion” on OpenResearch

About this paper

Brett Barkley, Preston Culbertson, David Fridovich-Keil

The paper, SCOPED, introduces a computationally efficient method for out-of-distribution (OOD) detection in diffusion models by leveraging the information-theoretic relationship between the score function's squared norm and the local curvature of the log-probability density. It achieves competitive detection accuracy across vision and reinforcement learning tasks with an order of magnitude fewer model evaluations than prior diffusion-based approaches.

Key insights

• The ratio of the squared score norm to local curvature provides a theoretically grounded and reliable signal for OOD detection, expected to be near 1 for in-distribution samples and diverge for OOD inputs.
• Diffusion models, by directly learning the score function, enable efficient computation of both the score and its curvature, facilitating OOD detection with minimal additional computational overhead.
• Evaluating the diffusion model at just a few strategic noise levels, rather than along full denoising paths, drastically reduces computational cost while retaining sufficient information for accurate OOD detection.

Summary by alphaXiv.