We current an accessible first course on the arithmetic of diffusion fashions and circulate matching for machine studying. We goal to show diffusion as merely as potential, with minimal mathematical and machine studying stipulations, however sufficient technical element to cause about its correctness. Not like most tutorials on this topic, we take neither a Variational Auto Encoder (VAE) nor a Stochastic Differential Equations (SDE) strategy. The truth is, for the core concepts we is not going to want any SDEs, Proof-Primarily based-Decrease-Bounds (ELBOs), Langevin dynamics, and even the notion of a rating. The reader want solely be accustomed to primary chance, calculus, linear algebra, and multivariate Gaussians.
This tutorial has 5 elements, every comparatively self-contained. Part 1 presents the basics of diffusion: the issue we are attempting to unravel and an outline of the fundamental strategy. Sections 2 and three present assemble a stochastic and deterministic diffusion sampler, respectively, and provides intuitive derivations for why these samplers accurately reverse the ahead diffusion course of. Part 4 covers the closely-related matter of Move Matching, which may be considered a generalization of diffusion that provides further flexibility (together with what are referred to as rectified flows or linear flows). Lastly, in Part 5 we return to diffusion and join this tutorial to the broader literature whereas highlighting a few of the design selections that matter most in apply, together with samplers, noise schedules, and parametrizations.
† Mila, Université de Montréal
