Comparing Manifold and Deep Learning Strategies for Anomaly Detection in the CMS Experiment

At the Compact Muon Solenoid (CMS) Experiment at the LHC, the Level-1 Trigger reduces proton-proton collision event rates by over 99% to satisfy strict resource constraints. Anomaly detection trigger algorithms search for new physics in data that would otherwise be discarded. We benchmark unsupervised machine learning techniques for model-independent anomaly detection using the public Anomaly Detection Challenge (ADC) 2021 [1] and LHC Olympics [2] datasets. Traditional autoencoders are compared to manifold learning-based anomaly detection strategies for their effectiveness in extracting rare standard model or beyond-the-standard model simulated samples from background data. We train, test, and evaluate neural networks to mimic manifold learning techniques. Input spaces are reconstructed from latent spaces, and receiver operating characteristic (ROC) curves are used to gauge performance across signals of benchmark particle decay scenarios. These studies aim to inform the development of future anomaly detection trigger and analysis strategies.

[1] https://arxiv.org/pdf/2107.02157.pdf

[2] https://arxiv.org/pdf/2101.08320.pdf