Generative AI represents a groundbreaking development within the broader “Machine Learning Revolution,” significantly influencing technology, science, and society. In this work researchers in particular of the team “Disordered systems and applications” of the LPENS focus on the state-of-the-art “diffusion models”, which are currently used to generate images, videos, and sounds. They are very fascinating algorithms for physicists, as they are very much connected to concepts from stochastic thermodynamics, particularly time-reversed Langevin dynamics. These diffusion models initiate from a simple white noise input and evolve it through a Langevin process to generate complex outputs such as images, videos, and sounds.
They show that statistical physics provides principles and methods to characterise this generation process. Specifically, they unveil how phenomena such as the transition from memorization to generalization and the emergence of features can be understood through the lens of symmetry breaking, phase transitions, and methods used to study disordered systems.
Illustration of the three dynamical regimes occurring during the generative stochastic dynamics
More:
https://www.nature.com/articles/s41467-024-54281-3
Affiliation author:
Laboratoire de physique de L’École normale supérieure (LPENS, ENS Paris/CNRS/Sorbonne Université/Université de Paris)
Corresponding author : Giulio Biroli
Communication contact: Communication team