Speaker
Description
Arrays of imaging atmospheric Cherenkov telescopes (IACTs) are superb instruments to probe the very-high-energy gamma-ray sky. Cosmic rays entering the atmosphere create air showers comprised of particles which produce Cherenkov light. It is detected by IACTs and contains spatial, temporal and calorimetric information of the event. Upon detection, IACT trigger systems determine, on real time, whether data is to be recorded. When trigger levels are surpassed, the signals from all pixels in the camera are digitised. Events can be classified as hadronic or electromagnetic, according to the particle that generated the shower. However, hadronic showers are tens of thousands of times more common than gamma-rays. Thus, computational resource needs could be reduced by performing a rough selection at trigger-level. We study the viability of deep neural network implementation in specialised hardware (FPGAs) at trigger level. The Deep Learning (DL) based Python package CTLearn performs full-event reconstruction: direction and energy estimation, as well as classification of the particle type which originated the shower. We implement non-structural pruning, a compression strategy, on a ResNet model using CTLearn and evaluate its performance and how it differs when varying the hyperparameters. This is part of the long-term goal of implementing DL models in the camera trigger for the CTAO telescopes.