Latent Outlier Exposure in Real-Time Anomaly Detection at the Large Hadron Collider

Abstract

We propose a novel approach to real-time anomaly detection at the Large Hadron Collider, aimed at enhancing the discovery potential for new fundamental phenomena in particle physics. Our method leverages the Latent Outlier Exposure technique and is evaluated using three distinct anomaly detection models. Among these is a novel adaptation of the variational autoencoder’s reparameterisation trick, specifically optimised for anomaly detection. The models are validated on simulated datasets representing collider processes from the Standard Model and hypothetical Beyond the Standard Model scenarios. The results demonstrate significant advantages, particularly in addressing the formidable challenge of developing a signal-agnostic, hardware-level anomaly detection trigger for experiments at the Large Hadron Collider.

Keywords

• anomaly detection
• unsupervised
• LHC
• particle physics
• autoencoder