Project
EF-LI-Opt-2
Exploding Velocity Fields in Generative Neural Flows: Analysis and Circumvention
Project Heads
Gabriele Steidl, Peter K. Friz
Project Members
Richard Duong
Project Duration
01.07.2025 – 30.06.2028
Located at
TU Berlin
Description
Generative neural models based on flows in measure spaces have shown a large impact on applications. However, the velocity fields learned by neural networks may lack a certain continuity which leads to undesired effects in neural approximations. This project aims to examine the behavior of such velocity fields theoretically and to use our insights towards the modeling of other physics-inspired generative flow models. A unifying framework of certain physics-inspired approaches are operator semigroups.
Related Publications
Richard Duong, Jannis Chemseddine, Peter K. Friz, Gabriele Steidl.
Telegrapher’s Generative Model via Kac Flows.
To appear in: SIAM Review, Research Spotlights.
Jannis Chemseddine, Gregor Kornhardt, Richard Duong, Gabriele Steidl.
Adapting Noise to Data: Generative Flows from 1D Processes.
To appear in: International Conference on Machine Learning (ICML).
Moritz Piening, Richard Duong, Gabriele Steidl.
Generalized Wasserstein Flow Matching: Transport Plans, Everywhere, All at Once