A robust anomaly finder based on autoencoders

TL;DR

This paper introduces a robust autoencoder-based method for identifying anomalous jets in high-energy physics, which remains unaffected by phase space variations and can be trained on low m/p T data to detect anomalies in high m/p T regions.

Contribution

The paper presents a novel jet preprocessing technique combined with autoencoders that ensures robustness against phase space differences, enhancing anomaly detection in jet data.

Findings

The method maintains stable discriminating variable distribution across different phase spaces.

It performs comparably to existing top taggers using less physics information.

The approach is effective for anomaly detection in high-energy physics datasets.

Abstract

We propose a robust method to identify anomalous jets by vetoing QCD-jets. The robustness of this method ensures that the distribution of the proposed discriminating variable (which allows us to veto QCD-jets) remains unaffected by the phase space of QCD-jets, even if they were different from the region on which the model was trained. This suggests that our method can be used to look for anomalous jets in high m/p T bins by simply training on jets from low m/p T bins, where sufficient background-enriched data is available. The robustness follows from combining an autoencoder with a novel way of pre-processing jets. We use momentum rescaling followed by a Lorentz boost to find the frame of reference where any given jet is characterized by predetermined mass and energy. In this frame we generate jet images by constructing a set of orthonormal basis vectors using the Gram-Schmidt method to span the plane transverse to the jet axis. Due to our preprocessing, the autoencoder loss function does not depend on the initial jet mass, momentum or orientation while still offering remarkable performance. We also explore the application of this loss function combined (using supervised learning techniques like boosted decision trees) with few other jet observables like the mass and Nsubjettiness for the purpose of top tagging. This exercise shows that our method performs almost as well as existing top taggers which use a large amount of physics information associated with top decays while also reinforcing the fact that the loss function is mostly independent of the additional jet observables.