Science of Diffusion Models: Speciation, Creativity, New Testbeds

Beyond the practical aspects explored in the first workshop, diffusion models remain a very active field of theoretical study. Given their special footing at the cross-roads of several disciplines across mathematics, statistics and computer science, this workshop will be an opportunity to provide a common language, as well as a catalyst to transfer progress across these sub-communities. This workshop will gather researchers aiming to uncover foundational aspects of diffusion models from diverse angles of attack: from classical learning theory aspects of score learning (including approximation, estimation and optimization), to analysis of diffusion models on canonical models from statistical physics, to connections with multiscale renormalization group methods, as well as the dichotomy between generalization and memorization. There will also be a day centered around the important question of evaluation, namely how to quantitatively assess the quality of a trained diffusion model.