Student Probability Seminar

Theory for diffusion models and some open problems

Speaker: Santi Aranguri, CIMS

Location: Warren Weaver Hall 202

Date: Wednesday, April 3, 2024, 12:15 p.m.


Abstract: Diffusion models are the state-of-the-art algorithm for image generation and other ML applications. Given samples from a distribution, a diffusion model is able to generate new samples by learning a flow from a gaussian to the data, using the old samples. In this talk, we will give the setup and some theoretical foundations casting the problem as an SDE, and list some open problems that people care about in practice as well as theoretical problems. In particular, given the amount of different versions of diffusion models out there, we will show how we can use the relative entropy to benchmark different models.