[Paper Review] Anomaly Awareness

Charanjit K. Khosa, Verónica Sanz|arXiv (Cornell University)|Jul 20, 2020

Particle physics theoretical and experimental studies14 citations

TL;DR

This paper introduces Anomaly Awareness, a hybrid machine learning method that enhances anomaly detection by training a classifier on normal data while explicitly incorporating diverse anomalies into the loss function via a modified cost function. The approach improves robustness in detecting previously unseen anomalies across particle physics and computer vision tasks, demonstrating strong performance on LHC jet topologies and MNIST-based benchmarks with minimal prior knowledge of anomaly types.

ABSTRACT

We present a new algorithm for anomaly detection called Anomaly Awareness. The algorithm learns about normal events while being made aware of the anomalies through a modification of the cost function. We show how this method works in different Particle Physics situations and in standard Computer Vision tasks. For example, we apply the method to images from a Fat Jet topology generated by Standard Model Top and QCD events, and test it against an array of new physics scenarios, including Higgs production with EFT effects and resonances decaying into two, three or four subjets. We find that the algorithm is effective identifying anomalies not seen before, and becomes robust as we make it aware of a varied-enough set of anomalies.

Motivation & Objective

• To address the limitations of traditional supervised and unsupervised anomaly detection in high-energy physics, where known anomalies may not generalize to unknown or novel phenomena.
• To develop a model-agnostic, hybrid learning framework that combines elements of supervised and unsupervised learning to improve detection of unknown anomalies.
• To reduce dependence on prior assumptions about specific new physics scenarios in LHC searches, enabling more robust, model-independent anomaly detection.
• To validate the method across diverse domains, including LHC jet topologies and standard computer vision benchmarks like MNIST.
• To demonstrate that awareness of a varied set of anomalies during training leads to improved generalization and robustness in detecting novel anomalies.

Proposed method

• The method begins with a prior binary classification run on normal events only, training a neural network to distinguish between two normal processes (e.g., SM top vs. QCD jets).
• In a subsequent 'anomaly awareness' training run, the model is fine-tuned using a modified cost function that includes both normal and anomaly events, with a hyperparameter λAA controlling the influence of anomaly examples.
• The loss function is adjusted to penalize misclassification of anomalies more heavily, encouraging the model to push anomaly scores away from the normal class boundary and toward a distinct region in the latent space.
• The architecture is adapted to input types: convolutional neural networks (CNNs) for jet images and fully connected networks for kinematic feature vectors in Higgs boson decay scenarios.
• The method uses the output probability distribution of the classifier as a proxy for anomaly scores, with higher uncertainty or deviation from normal class confidence indicating potential anomalies.
• The approach is validated using both LHC-based simulations (fat jet topologies, Higgs + photons + jets) and standard benchmarks (MNIST, CIFAR-10, fMNIST) to assess generalization across domains.

Experimental results

Research questions

• RQ1 Can a machine learning model trained on normal events become robust to detecting previously unseen anomalies if explicitly made aware of a diverse set of anomaly types during training?
• RQ2 How does the inclusion of multiple anomaly types in the training process affect the generalization capability of the model in detecting novel anomalies?
• RQ3 To what extent does the Anomaly Awareness method outperform standard supervised and unsupervised anomaly detection techniques in model-independent searches at the LHC?
• RQ4 Does the method maintain effectiveness when applied to non-physical datasets such as MNIST and CIFAR-10, demonstrating transferability across domains?
• RQ5 How does the hyperparameter λAA influence the model’s sensitivity and robustness in anomaly detection across different scenarios?

Key findings

• The Anomaly Awareness algorithm successfully identifies anomalies in fat jet topologies from new physics scenarios—including Higgs production with EFT effects and resonances decaying into 2–4 subjets—outperforming baseline models.
• When trained with a single anomaly type (e.g., ZH production), the model shows a measurable shift in anomaly score distribution, indicating improved detection capability.
• Incorporating two or more anomaly types (e.g., ZH, axion, graviton) further enhances robustness, with the model becoming less sensitive to the specific origin of anomalies.
• The method generalizes effectively to non-physical datasets: on MNIST, it successfully detects CIFAR-10, fMNIST, and nMNIST as anomalies when trained on two MNIST digits.
• The algorithm’s performance improves with increasing diversity in the set of anomalies used during training, confirming that 'awareness' of varied anomalies leads to stronger generalization.
• The approach achieves effective anomaly detection without requiring prior knowledge of the anomaly’s structure, making it suitable for model-independent searches in high-energy physics.