Generative Diffusions and Minimax Estimation

par Eddie Aamari (CNRS ENS)

mardi 14 oct. 2025, 11:15 → 12:15 Europe/Paris

Salle K. Johnson (1R3, 1er étage)

The aim of this presentation is to introduce generative models based on diffusions. After a brief review of the key concepts of stochastic calculus and PDEs, we will detail how the time-reversed Ornstein-Uhlenbeck process can be used to transport distributions on R^d from a Gaussian source. Since this reversed process involves the so-called score function, we will then address the issue of learning the score by minimizing an empirical contrast. Finally, we will discuss the stability of such a method, as well as the minimax estimation rates obtained for the underlying distribution.