The Use of Neural Networks in High-Energy Physics

Abstract

In the past few years a wide variety of applications of neural networks to pattern recognition in experimental high energy physics has appeared. The neural network salulions are in general of high quality, and. in a number of cases, are superior to those obtained using 'uaditionar methods. But neural networks are of particular interest in high energy physics for another rason as well: much of the pattern recognition must be performed online, i.e., in a few microseconds or less. The inherent parallelism of neural network algorithms. and the ability to implement them as very fast hardware devices, may make them an ideal technology for this application.

Search for B+→K+τ+τ− at the BaBar Experiment

“…The purity of a B tag decay mode is determined from MC studies and is defined as the fraction of B tag candidates with m ES > 5. 27 GeV/c 2 that are properly reconstructed within the given mode. If more than one B tag candidate with the same purity exists, the one with the smallest |∆E| is chosen.…”

“…At this point in the selection, remaining backgrounds are primarily BB events in which a properly reconstructed B tag is accompanied by B sig → D ( * ) ν , with D ( * ) → K ν and thus have the same detected finalstate particles as signal events. A multi-layer perceptron (MLP) neural network [27], with seven input variables and one hidden layer, is employed to suppress this background. The input variables are: the angle between the kaon and the oppositely charged lepton, the angle between the two leptons, and the momentum of the lepton with charge opposite to the K, all in the τ + τ − rest frame, which is calculated as p Bsig − p K ; the angle between the B sig and the oppositely charged lepton, the angle between the K and the low-momentum lepton, and the invariant mass of the K + − pair, all in the CM frame.…”

B- > K tau tau, B - >K pi pi gamma and B -> K* l l at BaBar

“…At this stage, remaining backgrounds are primarily BB events in which a properly reconstructed B tag accompanied by a B sig → D ( * ) lν l , with D ( * ) → Kl ν l which have the same detected final state particles as signal events. A multi-layer perceptron (MLP) neural network [11], with eight input variables and one hidden layer, is employed to suppress this background. The MLP is trained and tested using randomly split dedicated signal MC and B + B − background events, for each of the three channels.…”

EW and radiative penguin transitions from B-factories

“…The main remaining background, with a final state identical to that of signal, originates from BB events in which a properly reconstructed B tag is accompanied by B sig → D ( * ) ν , with D ( * ) → K ν . These contributions are suppressed using a multi-layer perceptron (MLP) neural network [24] based on kinematic and event shape variables. Figure 4 shows the MLP distribution, which output is required to be > 0.70 for the e + e − and µ + µ − channels and > 0.75 for the e + µ − channel.…”