Action Matching: Learning Stochastic Dynamics from Samples

Explore a comprehensive lecture on Action Matching, a method for learning continuous dynamics from snapshot data. Delve into the challenges of modeling systems in natural sciences and machine learning using cross-sectional samples. Discover how Action Matching enables the simulation of individual trajectories without relying on full trajectory data. Learn about the tractable training objective, extensions to stochastic differential equations, and applications in biology, physics, and generative modeling. Follow along as the speaker, Kirill Neklyudov, covers topics such as minimal vector fields, unbalanced action matching, quantum system simulation, and comparisons with other methods. Engage in a detailed discussion and Q&A session to deepen your understanding of this innovative approach to learning stochastic dynamics from samples.

- Intro
- Motivation
- Minimal vector field
- Action matching algorithm - learning the vector field from the samples
- Discussion
- Unbalanced action matching
- Simulation of a quantum system from observations
- Pairwise comparison with VP-SDE
- Generative modeling
- Discussion + Q&A