Online Program Home
  My Program

All Times EDT

Abstract Details

Activity Number: 32 - Statistical Answers to Astrophysical Questions: A Vital Chapter in the Chase for New Discoveries
Type: Invited
Date/Time: Sunday, August 8, 2021 : 3:30 PM to 5:20 PM
Sponsor: Section on Statistical Computing
Abstract #316794
Title: Advances in statistical learning for new physics searches when either the signal or the background is unknown
Author(s): Mikael Kuusela* and Purvasha Chakravarti and Larry Wasserman and Jing Lei and Tudor Manole and Patrick Bryant and John Alison
Companies: Carnegie Mellon University and Imperial College London and Carnegie Mellon University and Carnegie Mellon University and Carnegie Mellon University and Carnegie Mellon University and Carnegie Mellon University
Keywords: particle physics; Large Hadron Collider; mixture model; anomaly detection; optimal transport; semi-supervised learning

Searches of new phenomena in high-energy physics involve testing for the presence of a signal in a mixture model consisting of a background component and a signal component in a high-dimensional space of particle collision events. Modern particle physics experiments, such as those at the Large Hadron Collider at CERN, rely heavily on classifiers to perform this test. When a reliable model for both the background and the signal is available, the output of a supervised classifier can be used to obtain a powerful test for the presence of a signal. However, many open problems arise when either the signal or the background distributions are not well-known in advance. In this talk, I will present recent progress in these situations. In the first case, where the signal distribution is unknown, a semi-supervised classifier can be used as the basis of the test and I will describe methods for calibration, interpretation and signal strength estimation in this context. In the second case, where the background distribution is unknown, I will describe an approach based on optimal transport, which produces a data-driven background estimate by morphing a closely related signal-free data sample.

Authors who are presenting talks have a * after their name.

Back to the full JSM 2021 program