Speaker
Description
Generative diffusion models — the engine behind much of modern image and video synthesis — turn out to admit a transparent physical reading: the forward process is a Langevin diffusion that washes out structure, while the reverse, score‑driven process reconstructs configurations from noise, realising a stochastic quantisation à la Parisi–Wu with the learned score playing the role of the drift derived from an effective action.
Building on this correspondence, I will present our recent results from the DM‑QFT collaboration on generative sampling for lattice field theory: scalar theories in two and three dimensions including their critical regions, group‑preserving diffusion for U(N)/SU(N) gauge fields, and ongoing progress. I will emphasise Expandability, Exactness and Efficiency, and close with the road ahead QCD.
| Affiliation | RIKEN/UTokyo |
|---|---|
| Career status | Senior |