Generative models are typically based on explicit representations of probability distributions (e.g., autoregressive or VAEs) or implicit sampling procedures (e.g., GANs). I will present an alternative approach based on modeling directly the vector field of gradients of the data distribution (scores) which underlies recent score-based diffusion models. This framework allows flexible architectures, requires no sampling during training or the use of adversarial training methods. Additionally, score-based diffusion generative models enable exact likelihood evaluation through connections with neural ODEs, achieving state-of-the-art sample quality and excellent likelihoods on image datasets. I will discuss numerical and distillation methods to accelerate sampling and their application to inverse problem solving.

Bio: Stefano Ermon is an Associate Professor of Computer Science in the CS Department at Stanford University where he is affiliated with the Artificial Intelligence Laboratory. His research is centered on techniques for probabilistic modeling of data and is motivated by applications in the emerging field of computational sustainability. He has won several awards, including Best Paper Awards (ICLR, AAAI, UAI), a NSF Career Award, ONR and AFOSR Young Investigator Awards, Hellman Faculty Fellowship, Microsoft Research Fellowship, Sloan Fellowship, and the IJCAI Computers and Thought Award. Stefano earned his Ph.D. in Computer Science at Cornell University in 2015.