Marc Mézard - Bocconi University, Milano
Wednesday, April 8, 2026
Zoom opens: 10:30AM EDT
Seminar begins: 10:45AM EDT
How diffusion theory is used to produce fake data
Generative diffusion, in which one trains an algorithm to generate fake samples ‘similar’ to those of a data base, is a major new direction of machine learning which has become the state of the art for image and video generation. It is based on simple tools of diffusion processes: a noising process of the original data through a Langevin diffusion is followed by a time reversed denoising phase. However, a strict application of time reversal would lead back to the original data, and the whole difficulty of the generative process is to avoid this memorization effect, to build an algorithm that generalizes.
In this talk I will show how statistical physics can be used to analyze generative diffusion in the relevant regime where data live in large dimensions. I will underline the importance of dynamic phase transitions occurring during the generation process, and explain the mechanisms used to avoid the mere memorization of the database.