IEEE Information Theory Society Distinguished Lecture
IEEE Information Theory Society Distinguished Lecture
Title: Generative Modeling in Probability Space: From Wasserstein Optimization to Local Flow Matching
Speaker:Yao Xie (Official webpage) (CV) (2-page CV)
Coca-Cola Foundation Chair and Professor
H. Milton Stewart School of Industrial and Systems Engineering (ISyE)
Associate Director, Machine Learning Center
Adjunct Professor, School of Electrical and Computer Engineering (ECE)
Georgia Institute of Technology
Abstract: How can we learn complex probability distributions through a sequence of simpler transformations with theoretical guarantees? This talk presents flow-based generative modeling as iterative computation in probability space, connecting optimization, transport, and diffusion. I will first discuss JKO flow models, which implement proximal optimization in Wasserstein space. I will explain how convergence of the forward data-to-noise process yields guarantees for reverse generation, with bounds in Kullback-Leibler divergence that account for per-step optimization errors. I will then introduce Local Flow Matching, which learns short transformations between nearby distributions along a diffusion path. This construction enables smaller models and efficient training, with generation guarantees in chi-square divergence derived from diffusion contraction. Together, these results illustrate how optimization in probability space can guide the design and analysis of generative models. The talk will emphasize mathematical intuition and selected theoretical results, complemented by numerical examples. This is based on joint work with Xiuyuan Cheng, Jianfeng Lu, Yixin Tan, and Chen Xu.
Date: September 18, 2026, from 11:00am - 12:15pm Chicago time
Location: UIC east campus, lecture center C1, Chicago, IL 60607
Date and Time
Location
Hosts
Registration
-
Add Event to Calendar
- UIC east campus
- Chicago, Illinois
- United States 60607
- Building: Lecture Center C1
- Room Number: Lecture Center C1
- Starts 06 September 2026 05:00 AM UTC
- Ends 18 September 2026 05:00 AM UTC
- No Admission Charge
Speakers
Yao
Generative Modeling in Probability Space: From Wasserstein Optimization to Local Flow Matching
Abstract: How can we learn complex probability distributions through a sequence of simpler transformations with theoretical guarantees? This talk presents flow-based generative modeling as iterative computation in probability space, connecting optimization, transport, and diffusion. I will first discuss JKO flow models, which implement proximal optimization in Wasserstein space. I will explain how convergence of the forward data-to-noise process yields guarantees for reverse generation, with bounds in Kullback-Leibler divergence that account for per-step optimization errors. I will then introduce Local Flow Matching, which learns short transformations between nearby distributions along a diffusion path. This construction enables smaller models and efficient training, with generation guarantees in chi-square divergence derived from diffusion contraction. Together, these results illustrate how optimization in probability space can guide the design and analysis of generative models. The talk will emphasize mathematical intuition and selected theoretical results, complemented by numerical examples. This is based on joint work with Xiuyuan Cheng, Jianfeng Lu, Yixin Tan, and Chen Xu.