Training Continuous Normalizing Flows with Flow Matching for Generative AI

by Dr Tomasz Kacprzak (Senior Data Scientist SDSC)

Wednesday 14 May 2025, 11:45 → 13:45 Europe/Zurich

OHSA/E13

Abstract:
Flow matching approaches have gained a lot of attention recently in the field of generative AI. Flow Matching (FM) is a simulation-free approach for training Conditional Normalizing Flows (CNF) based on regressing vector fields of fixed conditional probability paths.  In this talk I will introduce the background of CNFs, the core idea of Flow Matching, and the connection of FM to generative Diffusion Models and to Optimal Transport. I will then show how FMs can be used for various generative AI tasks, as well as for Simulations-Based Inference in physics, and review recent results.

Papers:

1. Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., & Le, M. (2022). Flow matching for generative modeling. arXiv preprint arXiv:2210.02747.
2. Liu, X., Gong, C., & Liu, Q. (2022). Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003.
3. Albergo, Michael S., and Eric Vanden-Eijnden. "Building normalizing flows with stochastic interpolants." arXiv preprint arXiv:2209.15571 (2022).
4. Onken, D., Fung, S. W., Li, X., & Ruthotto, L. (2021, May). Ot-flow: Fast and accurate continuous normalizing flows via optimal transport. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 35, No. 10, pp. 9223-9232).
5. Wildberger, Jonas, Maximilian Dax, Simon Buchholz, Stephen Green, Jakob H. Macke, and Bernhard Schölkopf. "Flow matching for scalable simulation-based inference." Advances in Neural Information Processing Systems 36 (2024).

TEAMS link: https://teams.microsoft.com/l/meetup-join/19%3ameeting_NTE3NDdlMGEtODIyYy00MDY5LWIzMjctM2E2MjBjNjRlOTk2%40thread.v2/0?context=%7b%22Tid%22%3a%2250f89ee2-f910-47c5-9913-a6ea08928f11%22%2c%22Oid%22%3a%221942eab3-fd82-44ba-a638-5be253ed2d71%22%7d