Machine Learning and PDE Seminar | An overview of score-based generative modeling and the associated PDEs

Abstract: Score-based generative models and diffusion probabilistic models have demonstrated remarkable capabilities across various machine learning applications, such as DALL-E and Stable Diffusion. This talk aims to understand the theoretical foundations of score-based generative modeling, specifically focusing on the Fokker-Planck equations that underlie both the forward and reverse processes of the model. Additionally, we will introduce recent advancements in the form of a convergence result in the Wasserstein distance, which theoretically validates the exceptional performance of these models.