Reading group · Nov 2026 – Feb 2027

Mathematical foundations of diffusion models

Sign up by Friday, Oct 30 First meeting will be the week of Nov 2 in Gateway (room/day/time TBD)
Organizers
Curtis McDonald Simons Institute · UC Berkeley
Dhruv Rohatgi Statistics / Miller Institute · UC Berkeley
About

In this ``reading group'', we (the organizers) will give a self-contained introduction to the theory behind diffusion models, loosely following the text A Mathematical Introduction to Diffusion Models. We will introduce the necessary mathematical background as we go (e.g. stochastic processes and PDEs), but will emphasize intuition and algorithmic concepts over full rigor. Later sessions will cover (a selection of) recent developments, including discrete diffusion models and inference-time steering.

The meeting time will be set by a poll sent to everyone who signs up.

Tentative syllabus

Part I · Foundations

Nov – Dec
  1. 1
    Week of Nov 2
    Langevin dynamics, continuous processes, MCMC, mixing
  2. 2
    Week of Nov 9
    Forward and reverse SDEs, Fokker–Planck, Anderson’s equation
  3. 3
    Week of Nov 16
    Discretization and error control: step size, Girsanov, time schedules
  4. 4
    Week of Dec 1
    Score matching and other objectives, Tweedie’s formula, score estimators
— winter break —

Part II · Assorted topics

Jan – Feb
  1. 5
    Week of Jan 25
    Discrete diffusion models
  2. 6
    Week of Feb 1
    Inference time I: oracle guidance methods and empirical approximations
  3. 7
    Week of Feb 8
    Inference time II: SMC, closed-form guidance, provable guarantees
  4. 8
    Week of Feb 15
    Applications: text, images, molecules