ENSEMBLE NEURAL CLASSIFIERS FED BY EXPERT INFORMATION FOR ONLINE FILTERING IN A PARTICLE DETECTOR
Neural Networks, online detection, instrumentation
ATLAS is one of the main experiments of the Large Hadron Collider (LHC),
which aims to investigate the building blocks of matter and their forms of inter-
action. At the LHC, particles are collided every 25 ns, reaching an energy of up
to 14 TeV, and during collisions, a large amount of data is generated (≈70 TB/s).
Electrons are particles of interest to the experiment, however they are masked by
an intense background noise composed of hadronic jets, wich may have a similar
deposition profile in the calorimeter (highly segmented energy meter). To handle
the large volume of information, ATLAS use an online event selection system, to
remove the non-relevant information and preserve the interesting signatures. The
NeuralRinger is the standard method for electron classification in the fast calorime-
try step, which describes the energy deposition profile of particles with concentric
rings, generated around the most energetic cell. The ring-shaped signals feed an
ensemble neural classifiers to make the electron/jet decision. An information com-
monly used in calorimetry, to electrons/jets discrimination is the lateral shower
width. Due to the different iteraction types of these particles with matter, jets have
a wider profile. Aiming at improving the NeuralRinger performance, in this work
an expert pre-processing step is proposed, which highlights the differences in the
lateral shower profile of rings-shaped signals, facilitating the discrimination. Neural
network training methods were developed to tuning the neural network weights and
the preprocessing step coefficients in an integrated way. The proposed methods were
evaluated using experimental data from ATLAS. Through the proposed approach,
it was possible to achieve superior performance to the NeuralRinger, reducing the
acceptance of false electrons.