Physics ∩ ML Seminars |
Events on Wednesday, February 21st, 2024
- Diffusion Models - A Thermodynamic Perspective
- Time: 12:00 pm - 1:00 pm
- Place: Chamberlin 2104
- Speaker: Akhil Premkumar, KICP Chicago, University of Chicago
- Abstract: Diffusion models have found immense success in modeling complex, high dimensional data distributions. Most recently, the text-to-video generation tool, Sora by OpenAI, uses such models to produce extremely high fidelity videos from simple text prompts. In this talk I will introduce a physicist-friendly intuition for diffusion models. Starting with first principles I will demonstrate how diffusion models can be understood as a variational problem, like the ones we come across in physics. I will give a thermodynamic interpretation to the these models, connecting them back to the fluctuation theorems that originally inspired their invention. Based on: arXiv:2310.04490
- Host: Moritz Muenchmeyer