A Unified Approach to Analysis and Design of Denoising Markov Models

  • Yinuo Ren | Stanford University

Diffusion and flow-based generative models can be viewed through the lens of measure transport via Markovian stochastic dynamics, where the choice of dynamics critically shapes both theory and algorithms. In this talk, I will present a unified framework for denoising Markov models, characterizing when a forward process admits an explicit reverse-time generator and a principled variational training objective. The framework connects to nonequilibrium statistical mechanics and Doob’s h-transform, and generalizes score-matching across both continuous and discrete settings. Besides its connections to the existing diffusion models, I will also discuss about how it enables new constructions driven by general Lévy-type processes, including jump and heavy-tailed dynamics.

Speaker bio

Yinuo Ren is a final-year Ph.D. candidate in Applied and Computational Mathematics at Stanford University, advised by Profs. Lexing Ying (Applied Mathematics) and Grant Rotskoff (Computational Chemistry). His research focuses on the mathematical foundations and scalable algorithms of generative models, particularly diffusion and flow-based methods, with applications to scientific modeling and sampling. He obtained his B.S. in Computational Mathematics from Peking University, and will join the Kempner Institute at Harvard as a Research Fellow this fall.

シリーズ: MSR New England Generative Modeling & Sampling Seminar