Improving Fast Inference Jet Detection Neural Networks in the CMS Trigger System

The Large Hadron Collider (LHC) is already one of the largest sources of data in the world. By redevelopments before the end of 2027, its upgrade named High-Luminosity Large Hadron Collider (HL-LHC) will gather data by a factor of ten beyond LHC's initial specifications. The objective is to allow observation of rare processes when increasing luminosity. Therefore, some enhancements must be made to current algorithms to satisfy further extremely low-latency requirements. The main aim of this project, carried out through a total of 8 weeks, is to contribute to the efforts for designing fast inference neural networks to be implemented in Compact Muon Solenoid's (CMS) trigger system. For this purpose, Residual Neural Network (ResNet) architectures were exploited for jet detection. Furthermore, MorphNet like schemes were developed to automate the design of Single Shot MultiBox Detector (SSD) neural networks with specific resource constraints.