Speaker
Description
The Large Hadron Collider has recently started Run 3, which means protons are being collided at an astonishing energy of 13.6 TeV. Within the LHC is the CMS detector. Moreover, the Level one trigger is an integral part of the CMS detector and in many ways is considered the first responder of the High Level Trigger system. Its job is to make important initial cuts to the massive amounts of incoming data; therefore, it stores the data it was designed to flag as interesting for offline analysis and what is believed to be "uninteresting" data is thrown out forever and is never analyzed. However, what if within this "uninteresting" data lies some new physics unbeknownst to us, but the Level one trigger was able to pick out by deploying unsupervised machine learning algorithms onto the trigger? This is where autoencoders (AE) and variational autoencoders (VAE) enter the picture. The idea is that if the autoencoder is trained on standard collision events, then when the AE encounters events vastly different than the training data, the AE will produce a high reconstruction loss, which signifies an anomalous event that can be saved for offline analysis.
Currently, we are looking at two different AE and VAE architectures. One of the architectures is a deep neural network (DNN) VAE, meaning it is a fully connected network with dense Keras layers that make up the encoder and decoder of the AE and VAE, respectively. The other AE and VAE is an interaction network model, which is a kind of graphical neural network (GNN) AE that takes advantage of the natural graphical representation of particle collision data. The initial goal of the research is to determine which model, the DNN or GNN, is better for anomaly detection at Level one. Depending on the conclusion, a method to decrease the the computational resources used must be established in order to deploy the model on the Level one trigger. If the GNN proves to be the viable model, then the plan is to look to the method of knowledge distillation to train a student network, from a larger teacher network, which will meet the resource requirements without loss of generalization. Furthermore, one of these models could have the potential to be deployed on the Level one trigger to perform anomaly detection in the future.
