Robust Anomaly Detection for Particle Physics Using Multi-Background Representation Learning

TL;DR

This work tackles anomaly detection in high-energy physics by introducing robust multi-background representation learning. By training representations that distinguish multiple background processes and enforcing decorrelation with respect to a search variable $z$ via a NuRD-based objective, the method yields anomaly scores that are more robust to background-specific biases. The authors implement two scores, max logit (ML) and Mahalanobis distance (MD), and demonstrate improvements over a single-background VAE baseline on LHC jet data, including higher AUROC and reduced mass sculpting. The approach offers increased discovery potential for new physics by leveraging richer background information and stronger decorrelation guarantees, with practical implications for bump-hunt analyses in particle experiments.

Abstract

Anomaly, or out-of-distribution, detection is a promising tool for aiding discoveries of new particles or processes in particle physics. In this work, we identify and address two overlooked opportunities to improve anomaly detection for high-energy physics. First, rather than train a generative model on the single most dominant background process, we build detection algorithms using representation learning from multiple background types, thus taking advantage of more information to improve estimation of what is relevant for detection. Second, we generalize decorrelation to the multi-background setting, thus directly enforcing a more complete definition of robustness for anomaly detection. We demonstrate the benefit of the proposed robust multi-background anomaly detection algorithms on a high-dimensional dataset of particle decays at the Large Hadron Collider.

Figures (4)

• Figure 1: Overview of robust anomaly detection with multi-background representation learning. During training, jets (depicted by cones) as well as their mass $m$ and label $l$ are used to learn robust multi-background representations $r$. The data used for training includes jets of different background processes, not just QCD. Then, these learned representations are used to derive an anomaly score $\phi = f \circ r$ used at test time.
• Figure 2: An overview of the Nuisance-Randomized Distillation algorithm puli2021out used in this work for learning robust multi-background representations for anomaly detection in high-energy physics. First, we calculate weights $w$ for each input based on its label $y$ and jet mass $z$. These weights are used to approximate the distribution ${p_{\scaleto{\perp \!\!\! \perp}{4pt}}}$ such that the $y$ and $z$ are marginally independent. Then, we train our classifier to optimize a reweighted objective. We additionally include a mutual information-based penalty term to the loss by training a critic model $n$ gradient steps for every gradient step of the classifier.
• Figure 3: Our proposed robust multi-background detection methods outperform the baseline VAE implementation of Ref. qcdorwhat in both the overall detection performance (AUROC, left) as well as decorrelation with jet mass (less sculpting, right).
• Figure 4: (a) PCA on the learned representations show that they yield good separation of the out-of-distribution top process and the in-distribution QCD and W/Z processes. This leads to good separation of the downstream anomaly detection scores, Mahalanobis distance (b) and Max Logit (c).