Speaker
Description
Edge computing solutions can alleviate downstream hardware challenges that arise as a result of the massive streams of continuous data that high-energy physics experiments are capable of generating. The NEUROPix project proposes a novel approach for extracting particle information from pixel detectors on the edge with Spiking Neural Networks (SNNs)—an extremely low-latency and energy-efficient type of neural network that is inspired by biological neuron structure and behavior. Currently, our full data pipeline converts raw Timepix4 bitstreams into encoded spiking feature vectors, processes this with a SNN, and translates the decoded SNN output into signals that represent clustered particle IDs; however, extensions into other pixel detectors and edge-computing tasks are possible. This talk will focus on neuromorphic design strategies for detector applications and present current network training progress but will also briefly cover other aspects of the project, including realistic synthetic dataset development for algorithm training, and future hardware implementation onto FPGAs.