Abstract
We describe the construction of novel end-to-end jet image classifiers to discriminate quark- versus gluon-initiated jets using the simulated CMS Open Data. These multi-detector images correspond to true maps of the low-level energy deposits in the detector, giving the classifiers direct access to the maximum recorded event information about the jet, differing fundamentally from conventional jet images constructed from reconstructed particle-level information. Using this approach, we achieve classification performance competitive with current state-of-the-art jet classifiers that are dominated by particle-based algorithms. We find the performance to be driven by the availability of precise spatial information, highlighting the importance of high-fidelity detector images. We then illustrate how end-to-end jet classification techniques can be incorporated into event classification workflows using Quantum Chromodynamics di-quark versus di-gluon events. We conclude with the end-to-end event classification of full detector images, which we find to be robust against the effects of underlying event and pileup outside the jet regions-of-interest.
Highlights
The study of jet substructure at the CERN Large Hadron Collider (LHC) has played an instrumental role in the understanding of the standard model (SM) of particle physics through the analysis of jets produced from Quantum Chromodynamics (QCD) [1,2] and from the decay of boosted heavy resonances or particles such as the top quark [3,4,5,6] or the Higgs boson [7,8]
As previous studies [12,26] have shown image-based approaches to underperform relative to direct particle-data based algorithms, our results suggest this discrepancy can be attributed to limitations in the jet image construction rather than to the use of convolutional neural networks (CNN) themselves
We demonstrated the application of end-to-end classification techniques to the tagging of light quark jets versus gluon jets using simulated Compact Muon Solenoid (CMS) Open Data
Summary
The study of jet substructure at the CERN Large Hadron Collider (LHC) has played an instrumental role in the understanding of the standard model (SM) of particle physics through the analysis of jets produced from Quantum Chromodynamics (QCD) [1,2] and from the decay of boosted heavy resonances or particles such as the top quark [3,4,5,6] or the Higgs boson [7,8]. Most of the breakthrough success in applying CNNs to computer vision have come from bypassing rule-based ‘‘feature engineering’’ altogether and instead allowing CNNs to learn relevant features directly from the raw camera data This fact motivates the idea of applying ‘‘end-to-end physics classification’’ whereby CNNs are used to directly train on maps of the true detector-level data (or their simulated equivalents), in all their richness, before any particle processing is performed. In theory, this gives the classifier full access to the maximum recorded event information at a level not achievable with processed particle- or jet-level data, while avoiding the particle ordering problem altogether.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
