CSE DSI Machine Learning Seminar with Ricky Chen (Meta Fundamental AI Research (FAIR))

Flow Matching [1,2] provides a powerful and flexible framework for generative modeling by learning continuous flows that transport between probability distributions. In this talk, we introduce the fundamentals of Flow Matching, outline a very simple core principle that leads to a generalized recipe, and explore its generalizations to other stochastic processes [3]. In particular, a key focus is Discrete Flow Matching which enables principled modeling of discrete data using Continuous-Time Markov Chains [4]. Within this model class, we highlight the flexibility of using general discrete probability paths, offering new perspectives on scalable non- autoregressive discrete generative modeling [5].

[1] “Flow Matching Guide and Code” Lipman et al. 2024
[2] “Flow Matching for Generative Modeling” Lipman et al. 2022

[3] “Generator Matching: Generative modeling with arbitrary Markov processes” Holderrieth et al. 2024

[4] “Discrete Flow Matching” Gat et al. 2024

[5] “Flow Matching with General Discrete Paths: A Kinetic-Optimal Perspective” Shaul et al. 2024

Ricky Chen is a Research Scientist at Meta Fundamental AI Research (FAIR) team in New York. His research is on building simplified abstractions of the world through the lens of dynamical systems and flows. Lately, he has been exploring the use of stochastic control theory for large-scale generative modeling [Adjoint Matching], as well as constructing discrete generative models through continuous-time Markov chains [Discrete Flow Matching]. Their methods such as [Flow Matching] have been applied successfully for foundation models of video and audio [Movie Gen].