What's Anomalous in LHC Jets?

Abstract

Searches for anomalies are the main motivation for the LHC and define key analysis steps, including triggers. We discuss how LHC anomalies can be defined through probability density estimates, evaluated in a physics space or in an appropriate neural network latent space. We illustrate this for classical k-means clustering, a Dirichlet variational autoencoder, and invertible neural networks. For two especially challenging scenarios of jets from a dark sector we evaluate the strengths and limitations of each method.

Jets and Jet Substructure at Future Colliders

“…A promising path to improve this method is to extend the discriminative power from the physics phase space to include the latent space of the neural networks. This can be achieved, for example, using rapidity-mass matrices for standard autoencoders [264] (Dirichlet) variational autoencoders [265,266] or invertible normalizing flow network [267], benchmarked for dark-matter-inspired jet signatures. For any kind of neural network application to jet physics, self-supervised learning of symmetries, fundamental invariances, and detector effects is an exciting new direction which is expected to significantly improve the understanding and the experimental stability of neural networks applied to subjet physics [268].…”

Machine Learning and LHC Event Generation

“…In addition, normalizing flows come with significant advantages in controlling their performance and quantifying uncertainties, as discussed in the next section. Their invertible structure is useful for many LHC-applications, including anomaly detection or related density estimation tasks [87][88][89][90].…”

A Normalized Autoencoder for LHC Triggers

“…Motivated by their initial success, ML-methods for anomaly detection at the LHC were developed for anomalous jets [7][8][9][10][11][12][13][14][15][16], anomalous events [17][18][19][20][21][22][23][24][25][26][27][28][29][30][31][32][33][34][35], or to enhance search strategies [36][37][38][39][40][41][42][43][44]. They include a first ATLAS analysis [45], experimental validation [46,47], quantum machine learning [48], self-supervised learning [49,50], applications to heavy-ion collisions [51], the DarkMachines community challenge [52], and the LHC Olympics 2020 community challenge [53,…”