Giovanni Conforti (Ecole Polytechnique, Palaiseau)
The aim of this talk is to provide with an introduction to score-based diffusion models, that are a special class of generative models that has recently provided with impressive experimental results on various learning tasks. In the second part of the talk, I will illustrate a novel approach to show theoretical guarantees of convergence for these models that relies on the connection between time-reversal of diffusion processes and stochastic control. Despite its simplicity, the proposed methodology allows to considerably relax several assumptions made in the literature in order obtain explicit error bounds.
Joint work with A.Durmus and M.Gentiloni Silveri.
Zoom access data are available under www.math.uni-potsdam.de/fileadmin/user_upload/Prof-Wahr/Roelly/FS_23.pdf
