Conditional Flow Divergence Matching

<<< All talks

Conditional Flow Divergence Matching

When	Friday, February 14, 2025, 1:30 PM – 2:30 PM (MT)
Where	WEB L112

Abstract

Conditional flow matching (CFM) stands out as an efficient simulation-free approach for training flow-based generative models, achieving remarkable performance for data generation. However, CFM is insufficient to ensure accuracy in learning probability paths, and the learned vector field significantly violates the continuity equation governing probability flows. In response, we establish a new total-variation bound between the learned and ground-truth probability paths, showing that the gap between probability paths is bounded above by a combination of CFM loss and an associated divergence loss. This theoretical bound informs us to design a new objective to match both flow and divergence accompanied by an efficient implementation. Our new training approach improves the performance of the flow-based generative model by a noticeable margin without significantly raising the computational cost.

Speaker

Tags: algorithms & theory statistics

---

Part of the Data Science & AI Lecture Series. Something wrong on this page? Edit _data/talks/2025-02-14-bao-wang.toml.