The physics behind diffusion models

Diffusion models build on the same mathematical framework as physical diffusion. In this video, we get to the core of the connection between the physics of motion and generative AI. Topics covered: • The intuition of probability landscapes (data as peaks, noise as valleys) • Forward diffusion: how real data is gradually noised into chaos • Brownian motion, Wiener processes, and the physics of particle motion • Stochastic differential equations (SDEs) and the noise schedule • Training a score function model (a “compass” in the probability landscape) • Reverse diffusion and Anderson’s reverse SDE (sampling from noise to data) • Probability flow ODEs for faster, deterministic sampling 🔗 Main resources: • Full reading list: https://www.patreon.com/posts/physics-behind-136741238 • DDPM: Denoising Diffusion Probabilistic Models (https://arxiv.org/abs/2006.11239) • Score-Based Generative Modeling through Stochastic Differential Equations (https://arxiv.org/abs/2011.13456) 00:00 Intro 01:06 Diffusion as a time-variant probability landscape 04:03 Where diffusion fits in the life of a model 04:34 Forward diffusion (training data generation) 06:25 The physics of diffusion 08:23 The forward SDE (Stochastic Differential Equation) 10:24 Case study: DDPM and noise schedules 13:17 The ML model as a local compass 14:43 Reverse diffusion and the reverse SDE 16:15 Samplers 17:27 Probability-flow ODE (Ordinary Differential Equation) 19:26 Outro