Performance Comparison of Generic and Quantized Fully Connected and Convolutional Neural Networks for Real- Time Signal/Background Classification

Abstract

Since the beginning of the Large Hadron Collider (LHC) project, one of the biggest problems faced by scientists is dealing with the enormous amount of data produced by detectors. For the High Luminosity LHC phase, a new calorimeter endcap named High Granularity Calorimeter (HGCAL) has been developed for the upgrade of the Compact Muon Solenoid (CMS) detector. High granularity together with increased pile-up will result in a huge increase in data rate. Therefore, efficient real-time analysis methods are required to select data coming from events of interest from tremendous background production. The development of specialized libraries, like QKeras, enables the quantization of neural networks (NNs) so far used mostly in the offline analysis due to their high processing requirements. The reduction of NN size together with input quantization makes possible their usage in limited resources as a particle classification strategy. We present a comparison of fully connected and convolutional NNs used for the potential real-time signal/background classification method. Results show that convolutional models slightly outperform fully connected architectures in both generic and quantized cases.