Computer vision and jets
Image based jet physics classifier with a convolutional neural network
The Large Hadron Collider (LHC) at CERN collides two beams of high-energy protons and produces many new (unstable) particles. Some of these new particles (quarks and gluons) will undergo a showering process, where they radiate many other quarks and gluons in successive binary splittings. These splittings can be represented with a binary tree, where the energy of the particles decreases after each step. When the energy is below a given threshold, the showering terminates, resulting in a spray of particles that is called a jet. The particle detectors only observe the leaves of this binary tree (the jet constituents).
It is important to classify different types of jets, e.g. originated from W, Z and Higgs bosons, as well as top quarks, with respect to the main QCD background. In this project we use computer vision with deep learning to build a classifier for boosted top jets at the LHC, that could also be straightforwardly extended to other types of jets. In particular, we implemented a convolutional neural network (CNN) to identify jet substructure in signal and background events. Our CNN inputs are jet images that consist of five channels, where each of them represents a color. The first three colors are given by the transverse momentum (pT) of neutral particles from the Hadronic Calorimeter (HCAL) towers, the pT of charged particles from the tracking system and the charged particles multiplicity. The last two colors specify the muon multiplicity and b quark tagging information.
Publications:
Pulling Out All the Tops with Computer Vision and Deep Learning
S. Macaluso and D. Shih
Journal of High Energy Physics volume 2018, Article number: 121 (2018) [ journal paper ] [ arXiv ]