Application of deep learning to the reconstruction of electron tracks in the CMS experiment

Abstract

This work proposes an implementation of a Deep Regression model for the purpose of predicting electron track parameters from collision events at the CMS Experiment at CERN. It is entirely written in Python, one of the most popular programming languages in the field of machine learning, and makes use of PyTorch, a cutting-edge Deep Learning framework, widely known for its dynamic graph structure. Several architectures were used, including the base architecture proposed by Bernkopf (Masters Thesis in preparation) and convolutional neural networks. They were trained and tested using a variety of algorithms and hyper-parameters to assess their performance. A training error reduction of a factor 2 to 3.5 was achieved with respect to the baseline, depending on the parameter. With under 5k network parameters the model offers a light-weight and precise tool to predict electron tracks. With automated predictions of streaming particle track data in mind, a possible bottleneck regarding data imports using different data formats was tested.