From Entropy to Artistry: Thermodynamics and Generative AI
Explore the fascinating connections between diffusion models and physics in this 55-minute lecture by Stephan Mandt from UC Irvine. Delve into the world of non-equilibrium thermodynamics and its application to denoising diffusion models, which have revolutionized image and video generation. Discover how augmenting these models with additional variables like position and momentum opens up new possibilities for invertible diffusion processes and efficient sampling techniques. Learn about the potential of thermodynamic processes in generative AI, including their application in downscaling precipitation patterns to capture extreme events and local geographical features. Gain insights into the intersection of artificial intelligence and scientific principles, strengthening the bond between these fields.