Overview
Explore a comprehensive talk on Metric Flow Matching for smooth interpolations on data manifolds. Delve into the innovative framework that challenges traditional Euclidean-based conditional paths in generative models. Learn how this simulation-free approach minimizes kinetic energy of data-induced Riemannian metrics to create more meaningful interpolations. Discover the application of Metric Flow Matching in various challenging domains, including LiDAR navigation, unpaired image translation, and cellular dynamics modeling. Gain insights into the methodology's superiority over Euclidean baselines, particularly in single-cell trajectory prediction. Follow the speaker's journey from background and motivation through the intricacies of the algorithm, including geodesic interpolants training and pseudocode implementation. Conclude with a thorough examination of experimental results, key takeaways, and an engaging Q&A session.
Syllabus
- Intro + Background
- Motivation
- Metric Flow Matching
- Geodesic Interpolants Training
- Pseudocode
- Experiments
- Conclusions
- Q+A
Taught by
Valence Labs