Title: “From Entropy to Artistry: on Thermodynamics and Generative AI”

Speaker: Stephan Mandt

Format: Hybrid

Virtual: Zoom link provided upon registration

*Please note that in-person space is limited.*

__________________________________________________________________________________

Abstract: Denoising diffusion models have led to a series of breakthroughs in image and video generation. In this talk, I will explore the deep connections between diffusion models and physics. Rooted in non-equilibrium thermodynamics, these models enable a variety of extensions by lifting diffusions into augmented spaces, encompassing position, momentum, and potentially additional variables. Augmented diffusions provide a “complete recipe” for constructing invertible diffusion processes, as well as new samplers that significantly enhance their inference efficiency. Additionally, thermodynamic processes offer a natural playground for generative AI. I will demonstrate how video diffusion models can effectively downscale precipitation patterns to finer scales, precisely capturing extreme event statistics and local geographical patterns.

Bio: Stephan Mandt is an Associate Professor of Computer Science and Statistics at the University of California, Irvine. His research centers on deep generative modeling, uncertainty quantification, neural data compression, and AI for science. Previously, he led the machine learning group at Disney Research in Pittsburgh and Los Angeles and held postdoctoral positions at Princeton and Columbia University. Stephan holds a Ph.D. in Theoretical Physics from the University of Cologne, where he received the German National Merit Scholarship. He is furthermore a recipient of the NSF CAREER Award, the UCI ICS Mid-Career Excellence in Research Award, the German Research Foundation’s Mercator Fellowship, a Kavli Fellow of the U.S. National Academy of Sciences, a member of the ELLIS Society, and a former visiting researcher at Google Brain. His research is currently supported by NSF, DARPA, IARPA, DOE, Disney, Intel, and Qualcomm. Stephan is an Action Editor of the Journal of Machine Learning Research and Transaction on Machine Learning Research, held tutorials at NeurIPS, AAAI, and UAI, and regularly serves as (Senior) Area Chair for NeurIPS, ICML, AAAI, and ICLR. He currently serves as Program Chair for AISTATS 2024 and General Chair for AISTATS 2025.

Time: 12:00 pm EST

Free!

Registration