Seminar on Foundations of Flow Matching for Generative Modelling

Seminar on Foundations of Flow Matching for Generative Modelling

Description: Flow Matching has recently emerged as a powerful framework for training continuous generative models, unifying ideas from normalizing flows, diffusion models, and optimal transport. At its core, Flow Matching replaces stochastic simulation with a direct regression objective, allowing continuous normalizing flows to be trained efficiently and with strong theoretical guarantees. This seminar will introduce the conceptual foundations of Flow Matching, beginning with the evolution from normalizing flows and neural ODEs to stochastic interpolants and rectified flows. We will examine how different probability paths give rise to diverse model families, how diffusion emerges as a special case, and why Flow Matching provides a principled and flexible alternative. Along the way, we will highlight key theoretical insights, practical design choices, and emerging applications in scientific simulation and high-dimensional generative modeling. The session is designed to equip participants with both an intuitive and formal understanding of Flow Matching, preparing them to explore advanced topics and research directions in this rapidly developing area.

Here is a nice overview of the concepts behind Flow Matching that will be explored in this Seminar: https://www.youtube.com/watch?v=DDq_pIfHqLs&ab_channel=Jia-BinHuang

Time slot: Tuesdays 2 - 4 pm.

Requirements: Strong background in Mathematics, Computer Science and the foundations of Machine Learning.

First Meeting: 21.10.2025