Interpretable deep learning for two-prong jet classification with jet spectra

Abstract

Classification of jets with deep learning has gained significant attention in recent times. However, the performance of deep neural networks is often achieved at the cost of interpretability.Here we propose an interpretable network trained on the jet spectrum S 2 (R) which is a two-point correlation function of the jet constituents. The spectrum can be derived from a functional Taylor series of an arbitrary jet classifier function of energy flows. An interpretable network can be obtained by truncating the series. The intermediate feature of the network is an infrared and collinear safe C-correlator which allows us to estimate the importance of an S 2 (R) deposit at an angular scale R in the classification. The performance of the architecture is comparable to that of a convolutional neural network (CNN) trained on jet images, although the number of inputs and complexity of the architecture is significantly simpler than the CNN classifier. We consider two examples: one is the classification of two-prong jets which differ in color charge of the mother particle, and the other is a comparison between Pythia 8 and Herwig 7 generated jets.

Neural network-based top tagger with two-point energy correlations and geometry of soft emissions

“…Graph networks are flexible enough for analyzing multiple objects appears at the LHC, and have been studied in various contexts [16,20,25,41,[117][118][119][120][121][122][123][124]. The graph network in this paper has access to only IRC safe two-point energy correlations [19,21,[96][97][98][125][126][127][128]. It was shown that the network has comparable performance to the CNN in the Higgs jet vs. QCD jet classification [21].…”

“…The graph network in this paper has access to only IRC safe two-point energy correlations [19,21,[96][97][98][125][126][127][128]. It was shown that the network has comparable performance to the CNN in the Higgs jet vs. QCD jet classification [21]. We use this network for top jet vs. QCD jet classification, and it is a good starting point toward the network whose top tagging performance is comparable to the CNN.…”

“…Still, it is not straightforward to identify the key physical features that contributed to the improvement, other than looking for the hidden data representations of the CNN [7,11,28,146], or checking the response of the network after perturbing the inputs [109]. Note that organized networks whose hidden representations have physical interpretations [9,16,18,21,24] allow us to interpret the results in terms of physics. For this purpose, we consider flexible and interpretable quantities derived from the jet image and use them as inputs to a jet classifier modeled by a multilayer perceptron (MLP).…”

Exploring the space of jets with CMS open data

“…The stress-energy flow [21][22][23] is a particularly powerful probe of jets, since it in principle contains all the information about a jet that is infrared and collinear (IRC) safe [24][25][26]. A variety of observables have been built around the energy flow concept [27][28][29][30][31], including recent work on machine learning for jet substructure [32][33][34].…”

Study of energy deposition patterns in hadron calorimeter for prompt and displaced jets using convolutional neural network

“…Then one can apply sophisticated algorithms developed for image recognition. This and related neural network approaches were used for boosted W boson tagging [49,[51][52][53], quark-gluon discrimination [54][55][56], top tagging [50,[57][58][59], generic anti-QCD tagging [60,61], photon identification [62], heavy-flavor tagging [63] and search for BSM particles [64][65][66][67][68][69]. Modern deep learning algorithms trained on HCAL energy deposition images can be used for LLP studies.…”