Events

Date: 17 November 2026   Time: 10:00 - 11:00     

Location: G. O. Jones Building Room 516

Title: Using field theory to understand generative ML

Abstract:
Generative Machine Learning (ML) models that produce high-dimensional spatial data have become ubiquitous and vastly more powerful in recent years, from images to videos to chemicals and materials. I begin by presenting a brief overview of such models, focussing in particular on diffusion models, a particularly popular and powerful class of generative model that learn how to generate images by gradually denoising pure noise samples. I then present recent work in which such models, when using a neural network with certain (common) biases of locality and translation invariance, admit an Effective Field Theory (EFT) description in a defined regime. This description allows us to better understand the build-up of spatial correlations in the samples being denoised from noise into data. Next, I briefly present currently ongoing work following on from this, where conditional entropy is used to understand these spatial correlations in greater depth, including for samples drawn from canonical physical theories such as Landau-Ginzburg theories. If there is time, I shall finish with a quick introduction to other recent work on a clustering interpretation of a new class of related generative models, called drifting models.