Simon Coste, researcher at Université Paris Cité and specialist of generative modelling, will be giving two lectures on generative modeling on september 20 and 27, from 17h30 to 19h.
Generative modelling
Abstract : Generative modelling consists in (1) learning a probability distribution directly from data, and (2) sampling new realizations from this distribution. Deep learning methods recently gave a new impetus to the domain and led to spectacular successes, especially in image generation. In this mini-course I’ll survey four methods in generative modelling : normalizing flows, energy-based models, score matching, and diffusion-like models. I will show their mathematical and algorithmic formulation, explain the theoretical guarantees if any, and detail their possible drawbacks.