Machine Learning in Montpellier, Theory & Practice – Eddie Aamari (CNRS)
The aim of this talk is to introduce generative models based on diffusions. After a brief reminder of the key concepts of stochastic calculus, we’ll detail how a time-reversed Ornstein-Uhlenbeck process can be used to transport distributions when starting from a Gaussian source. As this reversed process involves the so-called score function, we will then address the question of score learning via the minimization of an empirical contrast. Finally, we’ll discuss the stability of such a method, as well as minimax estimation speeds if time permits. Notes are available [ https://isdm.umontpellier.fr/www.math.ens.psl.eu/~eaamari/files/Diffusion-based%20generative%20modeling%20-%20For%20statisticians%20and%20probabilists.pdf | here ] .