DEEP LEARNING APPLIED IN THE CLASSIFICATION OF EVENTS GENERATED AT THE ATLAS EXPERIMENT
Resumen
A Toroidal LHC Apparatus (ATLAS) is one of the two general purpose detectors
at the Large Hadron Collider (LHC), at Conseil Européen pour la Recherche Nucléaire
(CERN). Inside this, bunches of protons collide with a frequency of 40 MHz, and each
collision, or event, can produce huge amounts of particles. The classification of LHC events
is one of the most important analysis tasks in HEP, and a fundamental work for searching
new phenomena. This work is focused in boosted di-Higgs decaying into b
¯bτ +τ
−, handled
as a classification task using deep learning techniques. Many models were trained, and
the best 9 of them are tested, evaluated and compared. The best model resulted from
taking an approach with Parameterized Neural Networks (PNN) and Cost-sensitive learning,
specifically increasing the background class weight. Scores with these techniques reached
above 0.9 F1-score on both background and signal classes. This work is a computer science
study in collaboration with the Physics Department and CCTVal.