Flow Annealed Importance Sampling Bootstrap - Machine Learning for Complex Distributions

Explore a comprehensive lecture on Flow Annealed Importance Sampling Bootstrap presented by Laurence Midgley and Vincent Stimper from Valence Labs. Delve into the intricacies of normalizing flows as tractable density models for approximating complex target distributions, such as Boltzmann distributions in physical systems. Learn about the novel Flow AIS Bootstrap (FAB) method, which combines annealed importance sampling with flows to overcome mode-seeking behavior and high variance issues in current training approaches. Discover how FAB effectively approximates complex multimodal targets and successfully learns the Boltzmann distribution of the alanine dipeptide molecule without relying on Molecular Dynamics simulations. Gain insights into the method's superior performance in generating unbiased histograms of dihedral angles. The lecture covers topics including Boltzmann Generators, machine learning preliminaries, experimental results, and concludes with a summary, outlook, and interactive Q&A session.

- Intro
- Boltzmann Generator
- Machine Learning: Preliminaries
- Methods
- Experiments and Results
- Boltzmann Distribution of Alanine Dipeptide
- Summary and Outlook
- Discussion and Q+A