engleski

Input Data Quantization For Real Time Signal Vs Background Classification Using Quantized Neural Network

The volume of data generated as a result of proton-proton collisions at the Large Hadron Collider (LHC) project represents the real challenge for processing and storage. A huge increase in data rate expected in the High Luminosity phase of the project requires more efficient real-time analysis methods that will enable a fast and accurate selection of events of interest from the enormous background production. Machine learning methods have been used in high energy physics successfully for different ranges of tasks. Quantized neural networks together with specialized library hls4ml enable the deployment of neural network models in Field Programmable Gate Arrays (FPGAs). Input quantization can, even more, reduce the model size while maintaining accuracy. A model with saturated data quantized with 2 bit retains accuracy higher than 90% representing a possible mechanism for real-time signal/background classification.

LHC, CMS, neural network, quantization, classification

nije evidentirano