Image Generation by Score Diffusion and the Renormalisation Group

Explore a 26-minute lecture from the CMSA Mathematics and Machine Learning Closing Workshop where Stéphane Mallat from Flatiron/College de France delves into score-based diffusions and their applications in generating images, sounds, and complex physical systems. Learn about the fascinating intersection of deep networks and high-dimensional score estimation, examining how these models achieve generalization without falling prey to the curse of dimensionality. Discover the crucial role of multiscale properties and renormalisation group decomposition from statistical physics in enabling effective model performance. Gain insights into practical applications, particularly in modeling turbulence, while understanding the theoretical foundations that make these impressive generative models possible.