Multiple Instance Learning for Beyond Standard Model Physics: Embedding Vectors as Summary Statistics

Synopsis

The Standard Model (SM) successfully describes fundamental interactions, yet physics Beyond the Standard Model (BSM) remains elusive. We utilize the Standard Model Effective Field Theory (SMEFT) to parameterize potential new physics via Wilson coefficients. However, constraining these coefficients at the Large Hadron Collider (LHC) is difficult for small Wilson coefficients, which is a practically "low-signal" regime where standard statistical methods lose sensitivity due to the performance degradation of conventional Machine Learning (ML) classifiers. This thesis introduces a Multiple Instance Learning (MIL) framework to address this challenge. Unlike traditional classifiers that process collision events individually, our approach aggregates embedding vectors from sets of events as learned summary statistics to extract latent information. Using simulated data for the H → WW decay channel, we demonstrate that this method allows ML models to mitigate this sensitivity loss. We show that, under specific conditions, MIL estimators yield tighter constraints on Wilson coefficients compared to standard single-instance neural networks, offering a new strategy for precision measurements in background-dominated environments.

For Zoom link info, please contact Lindiwe Coyne at physgrad@sfu.ca.