SMuK 2023 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz
AKPIK 8: AI Topical Day – Normalizing Flows and Invertible Neural Networks (joint session AKPIK/T)
AKPIK 8.1: Talk
Thursday, March 23, 2023, 15:45–16:00, HSZ/0004
Efficient Sampling from Differentiable Matrix Elements with Normalizing Flows — •Annalena Kofler1,2, Vincent Stimper2,3, Mikhail Mikhasenko4, Michael Kagan5, and Lukas Heinrich1 — 1Technical University Munich — 2Max Planck Institute for Intelligent Systems, Tübingen — 3University of Cambridge, UK — 4ORIGINS Excellence Cluster, Munich — 5SLAC National Accelerator Laboratory, Menlo Park, USA
The large amount of data that will be produced by the high-luminosity LHC imposes a great challenge to current data analysis and sampling techniques. As a result, new approaches that allow for faster and more efficient sampling have to be developed. Machine Learning methods such as normalizing flows, have shown great promise in related fields. There, access to not only the density function but also its gradient has proven to be helpful for training. Recently, software for accessing differentiable amplitudes, which serve as densities in particle scattering, have become available that allow us to obtain the gradients and benchmark these new methods. The described approach is demonstrated by training rational-quadratic spline flows with differentiable matrix elements of the hadronic three-body decays, π(1800) → 3 π and Λc+ → p K− π+. To boost the ability to accurately learn and sample from complex densities whilst also reducing the number of training samples, we explore the use of the newly proposed method Flow Annealed Importance Sampling Bootstrap. Building on prior work, we plan to extend the approach to examples with more particles in the final state via the differentiable matrix elements provided by MadJax.