Stochastic Thermodynamics and SDE-based Generative Models
A new paper explores the application of stochastic thermodynamics to SDE-based generative models, such as diffusion models and the Schrödinger bridge. The research introduces trajectory-level definitions for work, heat, and entropy production, along with a generalized Jarzynski identity and a second-law-like inequality. This thermodynamic framework extends existing setups to handle time-dependent bath temperatures and nonconservative forces, offering a new perspective on these generative models from nonequilibrium statistical mechanics. AI
IMPACT This research offers a novel theoretical lens for understanding generative models, potentially leading to new insights and model development.