Karlsruhe 2024 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
T: Fachverband Teilchenphysik
T 16: Data, AI, Computing 1 (anomaly detection)
T 16.2: Talk
Monday, March 4, 2024, 16:15–16:30, Geb. 30.33: MTI
VAE-based anomaly detection in dijet events at √s=13 TeV — •Aritra Bal1, Benedikt Maier1, Thea Aarrestad2, Javier Duarte3, Markus Klute1, Jennifer Ngadiuba4, Maurizio Pierini5, Kinga Wozniak5, and Irene Zoi4 — 1Karlsruhe Institute of Technology — 2University of Zurich — 3University of California, San Diego — 4Fermi National Accelerator Laboratory — 5CERN
The reconstruction loss of an autoencoder can serve as a generic discriminator enabling anomaly searches. As part of the CMS Anomaly Search Effort (CASE), we present an approach for unsupervised anomaly detection in dijet events, by combining a variational autoencoder (VAE) with a novel technique for decorrelating the anomaly metric (i.e the autoencoder loss) from the dijet mass using a deep Quantile Regression. The resulting unsculpted spectra are then used to perform a bump hunt search that is sensitive to a range of narrow and broad signal resonances.
Keywords: Anomaly detection; Machine Learning; BSM physics