Reproduce “Renormalizing Diffusion Models” on OpenResearch

About this paper

Jordan Cotler, Semon Rezchikov

Researchers at Harvard and Princeton developed a framework that rigorously connects diffusion models to the Renormalization Group, allowing for the creation of physically interpretable machine learning models. This framework provides a new method for efficiently simulating statistical and quantum field theories and for finding quantum ground states with inherent multiscale interpretability.

Key insights

• The intrinsic multiscale nature of diffusion models can be formally understood as learning inverse Renormalization Group flows, providing a theoretical foundation for their generative capabilities.
• By parameterizing diffusion models with RG principles, learned parameters directly correspond to physical quantities like renormalized mass and couplings, enabling highly interpretable scientific machine learning.
• This framework yields a novel variational ansatz for quantum field theory ground states, which naturally incorporates hierarchical correlations to address issues like 'exploding gradients' in higher dimensions.

Summary by alphaXiv.