“Theoretical Guarantees for Score-Based Generative Models under Minimal Data Assumptions”
Martedì 26 Maggio 2026, ore 17:00 11:00 - Aula 2AB45 2BC30- Marta Gentiloni Silveri (École polytechnique)
Abstract
Score-based generative models (SGMs) are a class of generative models based on learning the score function associated with a stochastic differential equation (SDE). Once the score is estimated, it is used to simulate the corresponding time-reversed dynamics, enabling the generation of approximate data samples. Despite their strong empirical performance, a rigorous theoretical understanding of SGMs remains incomplete. In particular, recent work has focused on characterizing how different sources of error—namely score approximation, time discretization, and initialization—propagate through the generative process and affect sample quality. In this talk, I will present the mathematical framework underlying SGMs and discuss recent results on their convergence under minimal data assumptions, obtained during my PhD in collaboration with G. Conforti, A. Durmus, and A. Ocello. (https://epubs.siam.org/doi/10.1137/23M1613670, https://openreview.net/forum?id=gVenpttGec)
