26 July 2026 to 1 August 2026
University of Maryland, College Park
US/Eastern timezone

Unbiased Diffusion-Based Generation via Scalable Non-Equilibrium Transport

30 Jul 2026, 15:00
20m
Benjamin Banneker B (Adele H. Stamp Student Union)

Benjamin Banneker B

Adele H. Stamp Student Union

3972 Campus Dr, College Park, MD 20742
Contributed talk Algorithms and artificial intelligence Algorithms and artificial intelligence

Speaker

Octavio Vega (University of Illinois Urbana-Champaign)

Description

Sampling with flows and diffusion models has emerged as a promising alternative to MCMC in lattice field theory. A central obstacle to their practical adoption is the degradation of sample quality as the lattice volume grows, and identifying a scalable prescription for applying deep generative models to the lattice setting is still an open problem. We adapt and extend a framework based on non-equilibrium transport sampling (NETS) as a continuous-time generalization of annealed importance sampling. Our approach enables diffusion-based generation that remains asymptotically unbiased by accumulating dynamical reweighting factors along the interpolating trajectory at inference time, with the effective sample size systematically refinable through tunable post-training parameters and no retraining. We illustrate our method by training energy-based diffusion models in the context of scalar field theory and the ${\rm O}(3)$ sigma model, demonstrating monotonic improvement of the ESS and reduction in the variance of the importance weights with the number of integration steps and strength of the stochastic transport term.

Authors

Gurtej Kanwar (University of Edinburgh) Octavio Vega (University of Illinois Urbana-Champaign)

Presentation materials

There are no materials yet.