Hedwig Nora Nordlinder: Generator Matching - The Mathematical Foundations of (Many) Generative Machine Learning Methods

Abstract: In this talk an introduction to recent work in the field of generative machine learning called Generator Matching will be given. The Generator Matching framework subsumes and puts on strong theoretical footing a large class of generative machine learning models, including flow matching and diffusion models, and elegantly outlines which loss functions and training methods are permissible.