PHYSTAT Seminar: "Robust and provably monotonic networks for Particle Physics and beyond"

by J Michael Williams (Massachusetts Inst. of Technology (US))

Wednesday May 10, 2023, 4:00 PM → 5:30 PM Europe/Zurich

Machine-learning algorithms deployed by particle-physics experiments, especially those run during real-time data processing, must be designed to minimize the impact of effects like experimental instabilities that occur during data taking and deficiencies in simulated training samples. (If we knew all of the physics required to produce perfect training samples, there would be no point in performing the experiment.) Therefore, what is needed are highly expressive models that are both robust and interpretable. This talk presents a neural-network architecture that achieves both of these requirements, and provides formal guarantees of robustness. In addition, formally guaranteed monotonic behavior can be specified in any feature direction(s), which allows the adoption of "outliers are better" using a priori knowledge, e.g., to ensure that particle decays with longer lifetimes or higher transverse momenta are inherently understood to be more interesting, even if they have larger lifetime and/or pT values than anything seen in the training data. This neural-network architecture has been adopted for use in the primary classification algorithms in the LHCb trigger for Run 3, which was the initial motivation for its development. This same algorithm, largely just out of the box, has also been shown to beat the state-of-the-art models for applications in criminal justice, finance, medicine, etc. More broadly, it has ubiquitous potential uses anywhere that interpretability and fairness are important. Finally, another application is estimation of the Wasserstein metric (Earth Mover?s Distance, EMD) in optimal transport by employing the Kantorovich-Rubinstein duality to enable use of the EMD in geometric fitting applications. In particle physics, a specific example involves determining the Energy Mover's Distance in an exact and differentiable way.

Mike Williams is the founder and leader of the LHCb group at MIT, and the inaugural Deputy Director of the Institute for Artificial Intelligence and Fundamental Interactions (IAIFI), one of the US NSF national AI institutes. At MIT, he is a professor in the Physics Department and the Statistics & Data Science Center, and co-creator and co-chair of the PhD in Physics, Statistics, and Data Science program. Mike was also co-convener of the Dark Sectors topical group for the recent Snowmass Community study. 

 

GMT20230510-140145_Recording_1228x684.mp4

PhyStat-Williams.pdf