Comparing and improving hybrid deep learning algorithms for identifying and locating primary vertices

Abstract

Using deep neural networks to identify and locate proton-proton collision points, or primary vertices, in LHCb has been studied for several years. Preliminary results demonstrated the ability for a hybrid deep learning algorithm to achieve similar or better physics performances compared to standard heuristic approaches. The previously studied architectures relied directly on hand-calculated Kernel Density Estimators (KDEs) as input features. Calculating these KDEs was slow, making use of the DNN inference engines in the experiment's real-time analysis (trigger) system problematic. Here we present recent results from a high-performance hybrid deep learning algorithm that uses track parameters as input features rather than KDEs, opening the path to deployment in the real-time trigger system.