Event description
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. This method, based on simple principles of stochastic processes, has become the state of the art for image generation.
However, the reasons for this spectacular technological success are not fully understood, and neither are its limitations. After an introduction to this topic, the talk will explain how statistical physics concepts allow to analyze generative diffusion in the limit where data live in large dimensions. We will show the importance of dynamic phase transitions occurring during the generation process, and how to avoid the mere memorization of the database.